diff --git a/.gitattributes b/.gitattributes index b841c1dda0705dd363e13d8e5fd710ecaffc9045..86e32eb695084884820a042c4c03f6dac4ff7c75 100644 --- a/.gitattributes +++ b/.gitattributes @@ -4892,3 +4892,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/Structure-Aware[[:space:]]Sparse-View[[:space:]]X-ray[[:space:]]3D[[:space:]]Reconstruction/e0ae7df8-d3f1-46c2-a34a-52862215b941_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Structure-Guided[[:space:]]Adversarial[[:space:]]Training[[:space:]]of[[:space:]]Diffusion[[:space:]]Models/b70a5045-dbc8-47eb-880e-4252d02e0e55_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Structured[[:space:]]Gradient-based[[:space:]]Interpretations[[:space:]]via[[:space:]]Norm-Regularized[[:space:]]Adversarial[[:space:]]Training/25726c0b-b33a-4e50-8b53-f013deb7f334_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Structured[[:space:]]Model[[:space:]]Probing_[[:space:]]Empowering[[:space:]]Efficient[[:space:]]Transfer[[:space:]]Learning[[:space:]]by[[:space:]]Structured[[:space:]]Regularization/fcd62d92-907d-4f7f-bc80-b88db83abf7f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/StyLitGAN_[[:space:]]Image-Based[[:space:]]Relighting[[:space:]]via[[:space:]]Latent[[:space:]]Control/37a6c7df-f82e-4da6-aa4d-10d6b8e2b178_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Style[[:space:]]Aligned[[:space:]]Image[[:space:]]Generation[[:space:]]via[[:space:]]Shared[[:space:]]Attention/0cae3168-e63f-4f71-908a-5bdeae8d8200_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Style[[:space:]]Blind[[:space:]]Domain[[:space:]]Generalized[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]via[[:space:]]Covariance[[:space:]]Alignment[[:space:]]and[[:space:]]Semantic[[:space:]]Consistence[[:space:]]Contrastive[[:space:]]Learning/1e340b0f-2934-49e8-9933-238d4fb891ff_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Style[[:space:]]Injection[[:space:]]in[[:space:]]Diffusion_[[:space:]]A[[:space:]]Training-free[[:space:]]Approach[[:space:]]for[[:space:]]Adapting[[:space:]]Large-scale[[:space:]]Diffusion[[:space:]]Models[[:space:]]for[[:space:]]Style[[:space:]]Transfer/43978830-73bd-4db5-8b8c-57888ce93c6e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/StyleCineGAN_[[:space:]]Landscape[[:space:]]Cinemagraph[[:space:]]Generation[[:space:]]using[[:space:]]a[[:space:]]Pre-trained[[:space:]]StyleGAN/51ff4f8c-4976-4d05-85df-cf04bbc75017_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SuGaR_[[:space:]]Surface-Aligned[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]for[[:space:]]Efficient[[:space:]]3D[[:space:]]Mesh[[:space:]]Reconstruction[[:space:]]and[[:space:]]High-Quality[[:space:]]Mesh[[:space:]]Rendering/8c37e125-7381-4a71-8f2b-015c3d8a066a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SubT-MRS[[:space:]]Dataset_[[:space:]]Pushing[[:space:]]SLAM[[:space:]]Towards[[:space:]]All-weather[[:space:]]Environments/80ad799c-9412-49c0-a322-3a5e2a9d1a8f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Summarize[[:space:]]the[[:space:]]Past[[:space:]]to[[:space:]]Predict[[:space:]]the[[:space:]]Future_[[:space:]]Natural[[:space:]]Language[[:space:]]Descriptions[[:space:]]of[[:space:]]Context[[:space:]]Boost[[:space:]]Multimodal[[:space:]]Object[[:space:]]Interaction[[:space:]]Anticipation/5f051557-0067-4cd2-89f8-ff27dddafa6f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Super-Resolution[[:space:]]Reconstruction[[:space:]]from[[:space:]]Bayer-Pattern[[:space:]]Spike[[:space:]]Streams/ac654108-cae7-444a-a7e8-a0faf89a682d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SuperNormal_[[:space:]]Neural[[:space:]]Surface[[:space:]]Reconstruction[[:space:]]via[[:space:]]Multi-View[[:space:]]Normal[[:space:]]Integration/a4b33f00-d1f9-45e2-a62c-bb25956b58ec_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SuperPrimitive_[[:space:]]Scene[[:space:]]Reconstruction[[:space:]]at[[:space:]]a[[:space:]]Primitive[[:space:]]Level/556fb01c-1445-4c08-b116-0426ba52ea17_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SuperSVG_[[:space:]]Superpixel-based[[:space:]]Scalable[[:space:]]Vector[[:space:]]Graphics[[:space:]]Synthesis/3fa1e5c3-5890-4208-976c-19acdb43c796_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Supervised[[:space:]]Anomaly[[:space:]]Detection[[:space:]]for[[:space:]]Complex[[:space:]]Industrial[[:space:]]Images/ef358e69-3d45-4c6c-aa99-f5a5ceb11e08_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Suppress[[:space:]]and[[:space:]]Rebalance_[[:space:]]Towards[[:space:]]Generalized[[:space:]]Multi-Modal[[:space:]]Face[[:space:]]Anti-Spoofing/91e475e8-d317-465c-893c-024b420e20a0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SurMo_[[:space:]]Surface-based[[:space:]]4D[[:space:]]Motion[[:space:]]Modeling[[:space:]]for[[:space:]]Dynamic[[:space:]]Human[[:space:]]Rendering/6283bde0-3818-458b-8029-7eb693822789_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SurroundSDF_[[:space:]]Implicit[[:space:]]3D[[:space:]]Scene[[:space:]]Understanding[[:space:]]Based[[:space:]]on[[:space:]]Signed[[:space:]]Distance[[:space:]]Field/213de3e8-6a97-4d16-96b2-ac03f2f0441a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SwiftBrush_[[:space:]]One-Step[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Model[[:space:]]with[[:space:]]Variational[[:space:]]Score[[:space:]]Distillation/13853f95-872e-49e0-9add-f414db12f2de_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SwitchLight_[[:space:]]Co-design[[:space:]]of[[:space:]]Physics-driven[[:space:]]Architecture[[:space:]]and[[:space:]]Pre-training[[:space:]]Framework[[:space:]]for[[:space:]]Human[[:space:]]Portrait[[:space:]]Relighting/ce409cac-fd5a-4614-9e6d-87e790ca611e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Symphonize[[:space:]]3D[[:space:]]Semantic[[:space:]]Scene[[:space:]]Completion[[:space:]]with[[:space:]]Contextual[[:space:]]Instance[[:space:]]Queries/ba02978b-283d-434b-af7a-d4b604064e5e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SynFog_[[:space:]]A[[:space:]]Photo-realistic[[:space:]]Synthetic[[:space:]]Fog[[:space:]]Dataset[[:space:]]based[[:space:]]on[[:space:]]End-to-end[[:space:]]Imaging[[:space:]]Simulation[[:space:]]for[[:space:]]Advancing[[:space:]]Real-World[[:space:]]Defogging[[:space:]]in[[:space:]]Autonomous[[:space:]]Driving/30117727-713c-4f8f-a215-de682288f1dd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SynSP_[[:space:]]Synergy[[:space:]]of[[:space:]]Smoothness[[:space:]]and[[:space:]]Precision[[:space:]]in[[:space:]]Pose[[:space:]]Sequences[[:space:]]Refinement/e165eedf-6d86-4e7b-90c3-ccc6312e093e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SyncMask_[[:space:]]Synchronized[[:space:]]Attentional[[:space:]]Masking[[:space:]]for[[:space:]]Fashion-centric[[:space:]]Vision-Language[[:space:]]Pretraining/92bc41a7-a993-4d23-a197-14e6dd1e0509_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/SyncTalk_[[:space:]]The[[:space:]]Devil[[:space:]]is[[:space:]]in[[:space:]]the[[:space:]]Synchronization[[:space:]]for[[:space:]]Talking[[:space:]]Head[[:space:]]Synthesis/89645a31-8829-4d8f-939e-0d17fcee95a5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Synergistic[[:space:]]Global-space[[:space:]]Camera[[:space:]]and[[:space:]]Human[[:space:]]Reconstruction[[:space:]]from[[:space:]]Videos/908ee619-1331-452b-9ccb-4e5b4d116ac0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Synthesize[[:space:]]Diagnose[[:space:]]and[[:space:]]Optimize_[[:space:]]Towards[[:space:]]Fine-Grained[[:space:]]Vision-Language[[:space:]]Understanding/eca2ceba-84d3-4751-b38a-1ba7e74c92ba_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Synthesize[[:space:]]Step-by-Step_[[:space:]]Tools[[:space:]]Templates[[:space:]]and[[:space:]]LLMs[[:space:]]as[[:space:]]Data[[:space:]]Generators[[:space:]]for[[:space:]]Reasoning-Based[[:space:]]Chart[[:space:]]VQA/58e8331e-f572-44c9-b724-02ca5cd604a2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Systematic[[:space:]]Comparison[[:space:]]of[[:space:]]Semi-supervised[[:space:]]and[[:space:]]Self-supervised[[:space:]]Learning[[:space:]]for[[:space:]]Medical[[:space:]]Image[[:space:]]Classification/a33fa80f-bfa4-45b2-878d-abb01f3d724e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/T-VSL_[[:space:]]Text-Guided[[:space:]]Visual[[:space:]]Sound[[:space:]]Source[[:space:]]Localization[[:space:]]in[[:space:]]Mixtures/2ac9042c-6511-4205-8b71-c79e7f00a075_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/T4P_[[:space:]]Test-Time[[:space:]]Training[[:space:]]of[[:space:]]Trajectory[[:space:]]Prediction[[:space:]]via[[:space:]]Masked[[:space:]]Autoencoder[[:space:]]and[[:space:]]Actor-specific[[:space:]]Token[[:space:]]Memory/11bec0da-9290-4827-bb16-e062c2aadc7c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/TACO_[[:space:]]Benchmarking[[:space:]]Generalizable[[:space:]]Bimanual[[:space:]]Tool-ACtion-Object[[:space:]]Understanding/b3378a53-d4f8-49d5-95a4-af3ae029dc3c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/TAMM_[[:space:]]TriAdapter[[:space:]]Multi-Modal[[:space:]]Learning[[:space:]]for[[:space:]]3D[[:space:]]Shape[[:space:]]Understanding/f474f9b7-e45c-4a08-b76f-0b118ae61d7f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/TASeg_[[:space:]]Temporal[[:space:]]Aggregation[[:space:]]Network[[:space:]]for[[:space:]]LiDAR[[:space:]]Semantic[[:space:]]Segmentation/745b72e4-2dd6-4811-ade0-f52413f14141_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/TCP_Textual-based[[:space:]]Class-aware[[:space:]]Prompt[[:space:]]tuning[[:space:]]for[[:space:]]Visual-Language[[:space:]]Model/137896e1-f0a6-4b80-849c-9a14cbb3250c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/TE-TAD_[[:space:]]Towards[[:space:]]Full[[:space:]]End-to-End[[:space:]]Temporal[[:space:]]Action[[:space:]]Detection[[:space:]]via[[:space:]]Time-Aligned[[:space:]]Coordinate[[:space:]]Expression/63fde73a-d557-47b6-aee9-074fa8fa7370_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/TEA_[[:space:]]Test-time[[:space:]]Energy[[:space:]]Adaptation/c8c44200-6095-44b9-b188-9639c747d9ec_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/TFMQ-DM_[[:space:]]Temporal[[:space:]]Feature[[:space:]]Maintenance[[:space:]]Quantization[[:space:]]for[[:space:]]Diffusion[[:space:]]Models/a91d56f7-042a-42e5-9c77-ba0dbaac2109_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/THRONE_[[:space:]]An[[:space:]]Object-based[[:space:]]Hallucination[[:space:]]Benchmark[[:space:]]for[[:space:]]the[[:space:]]Free-form[[:space:]]Generations[[:space:]]of[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Models/da9cf0d5-8d8b-4df8-b396-af5ff1ab4d0d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/TI2V-Zero_[[:space:]]Zero-Shot[[:space:]]Image[[:space:]]Conditioning[[:space:]]for[[:space:]]Text-to-Video[[:space:]]Diffusion[[:space:]]Models/17566a41-32ff-45d4-9eb1-0099d95d87b1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/TIGER_[[:space:]]Time-Varying[[:space:]]Denoising[[:space:]]Model[[:space:]]for[[:space:]]3D[[:space:]]Point[[:space:]]Cloud[[:space:]]Generation[[:space:]]with[[:space:]]Diffusion[[:space:]]Process/8fe5d2c1-4f78-435e-8bde-f7a4afed8604_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/TIM_[[:space:]]A[[:space:]]Time[[:space:]]Interval[[:space:]]Machine[[:space:]]for[[:space:]]Audio-Visual[[:space:]]Action[[:space:]]Recognition/03af8456-92c0-4a22-a6d9-94775506b39a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/TRINS_[[:space:]]Towards[[:space:]]Multimodal[[:space:]]Language[[:space:]]Models[[:space:]]that[[:space:]]Can[[:space:]]Read/527c7573-3fb6-42ad-a14f-b7b957e97243_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/TRIP_[[:space:]]Temporal[[:space:]]Residual[[:space:]]Learning[[:space:]]with[[:space:]]Image[[:space:]]Noise[[:space:]]Prior[[:space:]]for[[:space:]]Image-to-Video[[:space:]]Diffusion[[:space:]]Models/51572829-1348-4002-a40a-f8702f404317_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/TTA-EVF_[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]for[[:space:]]Event-based[[:space:]]Video[[:space:]]Frame[[:space:]]Interpolation[[:space:]]via[[:space:]]Reliable[[:space:]]Pixel[[:space:]]and[[:space:]]Sample[[:space:]]Estimation/fa811509-bdf9-4c56-97db-551fcc2b0a3d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/TULIP_[[:space:]]Multi-camera[[:space:]]3D[[:space:]]Precision[[:space:]]Assessment[[:space:]]of[[:space:]]Parkinson's[[:space:]]Disease/532b2907-27ca-48aa-bd4d-3e8291e221ab_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/TULIP_[[:space:]]Transformer[[:space:]]for[[:space:]]Upsampling[[:space:]]of[[:space:]]LiDAR[[:space:]]Point[[:space:]]Clouds/52ac8b22-207c-41ff-adcb-c9f4c82bca08_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/TUMTraf[[:space:]]V2X[[:space:]]Cooperative[[:space:]]Perception[[:space:]]Dataset/5f2a634b-45a2-4d92-81e8-cda538302a8f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Tackling[[:space:]]the[[:space:]]Singularities[[:space:]]at[[:space:]]the[[:space:]]Endpoints[[:space:]]of[[:space:]]Time[[:space:]]Intervals[[:space:]]in[[:space:]]Diffusion[[:space:]]Models/992f5837-40b9-47fe-b9bd-9d2e369d7fdb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Tactile-Augmented[[:space:]]Radiance[[:space:]]Fields/f9e9c348-b82c-415a-b07e-5b945a68c375_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Tailored[[:space:]]Visions_[[:space:]]Enhancing[[:space:]]Text-to-Image[[:space:]]Generation[[:space:]]with[[:space:]]Personalized[[:space:]]Prompt[[:space:]]Rewriting/7aad987a-39f2-4096-8aa1-a539bbaacb9c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Taming[[:space:]]Mode[[:space:]]Collapse[[:space:]]in[[:space:]]Score[[:space:]]Distillation[[:space:]]for[[:space:]]Text-to-3D[[:space:]]Generation/695a2b86-fc93-420e-a626-d23f402233d8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Taming[[:space:]]Self-Training[[:space:]]for[[:space:]]Open-Vocabulary[[:space:]]Object[[:space:]]Detection/08e6cd2d-6b02-4147-b946-b39f423452c4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Taming[[:space:]]Stable[[:space:]]Diffusion[[:space:]]for[[:space:]]Text[[:space:]]to[[:space:]]360[[:space:]]Panorama[[:space:]]Image[[:space:]]Generation/3259ff45-5fcb-4059-9736-190d9ce81cb7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Taming[[:space:]]the[[:space:]]Tail[[:space:]]in[[:space:]]Class-Conditional[[:space:]]GANs_[[:space:]]Knowledge[[:space:]]Sharing[[:space:]]via[[:space:]]Unconditional[[:space:]]Training[[:space:]]at[[:space:]]Lower[[:space:]]Resolutions/301b4dc8-6f1d-458c-a509-0fe38145fc73_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Targeted[[:space:]]Representation[[:space:]]Alignment[[:space:]]for[[:space:]]Open-World[[:space:]]Semi-Supervised[[:space:]]Learning/b33baf4c-8c46-4e13-a308-0177b1e2eee6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Task-Adaptive[[:space:]]Saliency[[:space:]]Guidance[[:space:]]for[[:space:]]Exemplar-free[[:space:]]Class[[:space:]]Incremental[[:space:]]Learning/ef2d4502-c3bc-43a6-a846-89e23e77cec3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Task-Aware[[:space:]]Encoder[[:space:]]Control[[:space:]]for[[:space:]]Deep[[:space:]]Video[[:space:]]Compression/b978be40-5174-4abc-9649-d7daa2d37c89_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Task-Conditioned[[:space:]]Adaptation[[:space:]]of[[:space:]]Visual[[:space:]]Features[[:space:]]in[[:space:]]Multi-Task[[:space:]]Policy[[:space:]]Learning/5ebc706d-8b34-4357-9c79-967e39bda2a5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Task-Customized[[:space:]]Mixture[[:space:]]of[[:space:]]Adapters[[:space:]]for[[:space:]]General[[:space:]]Image[[:space:]]Fusion/7c9be988-491f-4562-822c-11cfddbdaa7d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Task-Driven[[:space:]]Exploration_[[:space:]]Decoupling[[:space:]]and[[:space:]]Inter-Task[[:space:]]Feedback[[:space:]]for[[:space:]]Joint[[:space:]]Moment[[:space:]]Retrieval[[:space:]]and[[:space:]]Highlight[[:space:]]Detection/a6a833ff-f398-48fe-909b-70ff81bf1b94_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Task-Driven[[:space:]]Wavelets[[:space:]]using[[:space:]]Constrained[[:space:]]Empirical[[:space:]]Risk[[:space:]]Minimization/d6dc5c43-4c52-4779-8c6a-d513a47a71be_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Task-aligned[[:space:]]Part-aware[[:space:]]Panoptic[[:space:]]Segmentation[[:space:]]through[[:space:]]Joint[[:space:]]Object-Part[[:space:]]Representations/9dbd5f9f-08af-4d23-b3b6-162e326d318c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Task2Box_[[:space:]]Box[[:space:]]Embeddings[[:space:]]for[[:space:]]Modeling[[:space:]]Asymmetric[[:space:]]Task[[:space:]]Relationships/dd90eec4-a932-4045-a86b-cf06d264eadc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/TeMO_[[:space:]]Towards[[:space:]]Text-Driven[[:space:]]3D[[:space:]]Stylization[[:space:]]for[[:space:]]Multi-Object[[:space:]]Meshes/d9ed6877-0911-45bb-b575-e2ba93e2db7d_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/fcd62d92-907d-4f7f-bc80-b88db83abf7f_content_list.json b/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/fcd62d92-907d-4f7f-bc80-b88db83abf7f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..24e4f8f1402e69613d0f0b20f6860c78cfe7369a --- /dev/null +++ b/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/fcd62d92-907d-4f7f-bc80-b88db83abf7f_content_list.json @@ -0,0 +1,1671 @@ +[ + { + "type": "text", + "text": "Structured Model Probing: Empowering Efficient Transfer Learning by Structured Regularization", + "text_level": 1, + "bbox": [ + 120, + 106, + 838, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhi-Fan Wu Chaojie Mao Xue Wang Jianwen Jiang Yiliang Lv Rong Jin Alibaba Group", + "bbox": [ + 140, + 178, + 815, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 279, + 313, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite encouraging results from recent developments in transfer learning for adapting pre-trained model to downstream tasks, the performance of model probing is still lagging behind the state-of-the-art parameter efficient tuning methods. Our investigation reveals that existing model probing methods perform well for the easy case when the source domain (where models are pre-trained) and the adapted domain are similar, but fail for the difficult case when the two domains are significantly different. Simply incorporating features extracted from multiple layers and increasing complexity of the probing model can mitigate the gap in the difficult case, but degrades the performance in the easy case. To address this challenge, we propose structured model probing (SMP) that is able to deliver good performance for both cases through structured regularization. The regularization performs feature selection leveraging model structure as a prior, and controls the complexity of the probing model through the weights of selected structures. This enables us to construct a simple adaptation model, with a small number of selected features and a linear prediction model, for the easy case; and to automatically increase the complexity of adaptation model, with a large number of selected features and a non-linear model, for the difficult case. Our extensive empirical studies show that SMP significantly outperforms the state-of-the-art methods for parameter efficient tuning, and at the same time, still maintains the advantage of computational efficiency for probing-based methods.", + "bbox": [ + 76, + 313, + 473, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 752, + 209, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The parameters of models have grown drastically with the rapid development of deep learning [5, 37], posing challenges for adapting pre-trained models to downstream tasks. The widely used transfer strategy, fully fine-tuning, becomes infeasible due to excessive computation and storage overhead. Recently, parameter efficient tuning [19, 23] has gained attention for reducing storage costs and improving tuning performance with few trainable parameters when lim", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/63f688131516d3ed72daf25e836154d3ffd2d13587e9edb41098c34830f2f3e3.jpg", + "image_caption": [ + "Figure 1. Performance comparison on different dataset groups (left) and mean relative adaptation time (right) on VTAB-1k benchmark." + ], + "image_footnote": [], + "bbox": [ + 506, + 282, + 888, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ited data are available for downstream tasks. Despite the progress, parameter efficient tuning methods still face significant computational overhead in training due to the need to perform forward and backward passes through the pretrained model at each training step.", + "bbox": [ + 496, + 489, + 893, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Model probing, which leverages features extracted from the pre-trained model to perform adaptation, is promising due to its efficiency and simplicity: The training procedure is efficient since the pre-trained model is fixed and only performs one forward propagation to extract features, and probing model is decoupled from pre-trained model that makes it convenient to deploy small sized task-specific downstream models. However, it is a nuisance that probing methods are usually outperformed by state-of-the-art parameter efficient tuning methods. This is true even after several recent improvements made for the probing methods [16, 45]. In this paper, we aim to improve the performance of model probing significantly without sacrificing its simplicity and efficiency.", + "bbox": [ + 496, + 566, + 895, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We first investigate when model probing methods perform worse than the tuning-based methods on widely used visual task adaptation benchmark VTAB-1k [44] as shown in Figure 1. Target domains are categorized as easy and difficult cases based on their similarity to the source domain where the pre-trained model is trained on (see Section 3.1 for details). We find simple linear probing yields comparable performance as the two tuning-based approaches (finetuning and VPT [23]) for easy cases, whereas a significant", + "bbox": [ + 496, + 763, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16849", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8976b25f7d62a3b031921846e07f4362b5639759867d578f4903a604d17738bc.jpg", + "image_caption": [ + "Figure 2. Adaptation pipelines of Model Tuning (left) and the proposed Structured Model Probing (right)." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 401, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7160d55c197b3be1ab48dc688ed30909c72b2fb4a9022aedc0f7297b00ee4323.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 85, + 890, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "performance gap is observed between linear probing and tuning-based methods for difficult cases. This suggests that model probing exhibits different behavior for various target domains, and performance gap arises for target domains that dissimilar to the source domain.", + "bbox": [ + 75, + 286, + 470, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This is not surprising that we expect a more complicated model than a linear probing to compensate the difference between the source domain and the target domain. In general, we can increase model complexity in two ways, either by introducing more features extracted from the pre-trained model, or by introducing a non-linear model for prediction. A main challenge is how to develop a unified framework of model probing that is able to handle both the easy and difficult adaptations effectively. This is because, according to our study in Section 3, for easy cases, a complicated probing model (with more features and non-linear probing) performs significantly worse than a simple linear probing.", + "bbox": [ + 75, + 363, + 468, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose structured model probing, or SMP for short, that effectively address both easy and difficult cases of domain adaptation. The key to our unified framework is structured regularization. It performs feature selection leveraging the structure of the pre-trained model as a prior, and controls the complexity of the probing model through the weights assigned to selected features. Thus, for the easy case, this regularization constructs a simple linear model with a few selected features; and for the difficult case, it automatically introduces more features from the pre-trained model and more non-linearity in the prediction module based on the value of loss function, as shown in Figure 2. Our empirical studies show that the proposed structured model probing performs significantly better than state-of-the-art tuning-based approaches, while maintaining the advantage of computational efficiency.", + "bbox": [ + 75, + 547, + 470, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are as follows:", + "bbox": [ + 96, + 789, + 351, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. We observe and demonstrate the conflicting behaviors of model probing: a complicated probing model with more features and non-linear transformation can improve the performance for difficult cases of domain adaptation, but it yields worse performance than linear probing for easy cases of domain adaptation.", + "bbox": [ + 78, + 809, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. We propose structured model probing (SMP) based on structured regularization. This regularization enables us to construct simple probing model for easy tasks, and automatically increase model complexity for difficult tasks, to simultaneously address both the easy and difficult cases of domain adaptation.", + "3. We conduct extensive empirical studies on various adaptation tasks. The proposed SMP method shows superior experimental results compared to state-of-the-art probing and tuning methods, and requires significantly lower training costs compared to tuning methods." + ], + "bbox": [ + 500, + 286, + 893, + 453 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 470, + 640, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transfer learning has been extensively studied in vision domain [34, 48]. It aims to improve the performance on target domain by leveraging the related information contained in source domain. Transfer learning based on pre-trained model [8, 20, 21] has received significant attention due to its simplicity and good empirical performance [23, 45].", + "bbox": [ + 496, + 500, + 893, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model tuning adapts the pre-trained model to new tasks via updating model parameters. Fully fine-tuning is a widely used strategies due to its efficacy and simplicity [9, 25, 44]. However, it suffers from huge storage and computation overhead due to the ever-expanding size of large-scale pre-trained model [37, 40]. Heuristic approaches have been proposed to improve fine-tuning performance [3, 6, 17, 18, 27]. Parameter efficient tuning [19] aims to reduce the storage requirement of transferred model by updating a few parameters attached to the pre-trained model [23, 47], like Prompt Tuning [29, 30] or Adapter [22]. While originally studied in natural language processing tasks due to the dominance of large-scale pre-trained language models [5, 14], there is growing interest in applying these methods to computer vision tasks [7, 23, 31, 41, 46, 47]. Though parameter efficient tuning reduces storage cost significantly, the training cost remains high due to the forward and backward propagation need to pass through the entire pre-trained model [11].", + "bbox": [ + 496, + 594, + 893, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model probing utilizes frozen extracted features for adapting the pre-trained model to new tasks, thus achieving effi", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16850", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cient adaptation. The most popular method, linear probing, suffers from performance degradation [25], which limits its application in transfer learning [20, 21]. Recent methods [16, 36, 45] have been proposed to improve probing performance by leveraging side network [38, 45] or incorporating intermediate representations [2, 16, 32], but there is still a significant performance gap compared to tuning methods. This motivates us to investigate the limitation of probing methods, and come up with a better solution.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Understanding the Limitation of Model Probing for Transfer Learning", + "text_level": 1, + "bbox": [ + 76, + 242, + 472, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we analyze the performance of probing methods across various downstream domains to understand their limitations, motivated by the observation in Figure 1. We assess feature informativeness for different tasks and investigate the necessity of non-linear transformation.", + "bbox": [ + 75, + 285, + 470, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Analysis Setup", + "text_level": 1, + "bbox": [ + 76, + 371, + 227, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-trained model and source data. The ViT-B/16 [15] is used as the backbone because of the extensive usage of transformer-based models. The model is pre-trained on ImageNet-21k [12] in a supervised learning manner.", + "bbox": [ + 75, + 398, + 468, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Target data. We use the VTAB-1k benchmark, consisting of 19 datasets that cover natural images (natural), images captured by specialist equipment (specialized), and images generated from simulated environments (structured), as our target data for transfer learning. We follow the setting in [44] and use 1000 samples as the training set per task to evaluate adaptation with limited data as suggested in [44].", + "bbox": [ + 75, + 462, + 468, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain similarity. Several previous works have proposed metrics to measure the similarity between source domain and target domain [1, 10, 13, 16, 33]. Here, we adopt the domain similarity measure defined in [16]:", + "bbox": [ + 75, + 571, + 468, + 632 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD o m a i n \\quad \\text {S i m i l a r i t y} = \\mathrm {A c c} _ {\\text {L i n e a r}} - \\mathrm {A c c} _ {\\text {S c r a t c h}}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 645, + 468, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A higher domain similarity suggests that the features extracted from a pre-trained model can be utilized for the target domain, leading to performance gains via a linear classifier compared to training from scratch. Conversely, a lower domain similarity indicates that training from scratch on the target domain may be more effective than using extracted features. Thus, the domain similarity is evaluated based on the benefit gained by a linear classifier compared to training from scratch. We also report the label-feature correlation [13] in Suppl. C.8, which exhibits a high Spearman rank correlation (0.854) with the domain similarity scores.", + "bbox": [ + 75, + 672, + 470, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain similarity also reflects how difficult to adapt from source domain to target domain. As shown in Figure 4, tasks in natural and specialized groups have higher domain similarity to source domain where the pre-trained model is", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e6ce50144cb8482b59edc659595afb130625e9a8932639916a34f16206715741.jpg", + "image_caption": [ + "Figure 3. Performance of features extracted from different layers on Sun397 (easy) and Clevr-Dist (difficult)." + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 696, + 234 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a2274185ec5d76b15ee835772fee332d9d4e0b99a6ae64cf014e11b7db3753ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 92, + 885, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "trained on, so we mark them as easy cases. And tasks in structured groups have lower domain similarity to source domain, thus they are labeled as difficult cases in Figure 1.", + "bbox": [ + 498, + 297, + 890, + 342 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Informativeness of Extracted Features", + "text_level": 1, + "bbox": [ + 500, + 353, + 828, + 368 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To quantify the informativeness of features extracted from different layers for downstream tasks, we use leave-one-out $k$ -NN testing as the surrogate performance criterion:", + "bbox": [ + 498, + 377, + 890, + 422 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {I n f o r m a t i v e n e s s} = \\frac {\\sum_ {i = 1} ^ {m} \\sum_ {x _ {j} \\in N (x _ {i})} \\mathbb {1} \\left\\{y _ {i} = y _ {j} \\right\\}}{k m}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 446, + 890, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $m$ is the number of samples, $x$ is the extracted features of a sample, $N(x_{i})$ represents $k$ nearest neighbors of $x_{i}$ . We simply set $k = 1$ , which results in leave-one-out 1-NN testing. We choose $k$ -NN, instead of linear classifier, for informativeness measurement because $k$ -NN is a non-parametric classifier and thus avoids training additional parameters, making the measurement more robust.", + "bbox": [ + 498, + 491, + 893, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Figure 3, we plot the informativeness of features extracted from different layers on two tasks from different domains. For Sun397, an easy domain adaption task, we observe the best performance achieved by features extracted from deep layers. This explains the good performance of linear probing for easy domain adaption tasks, as observed in Figure 1. While for Clevr-Dist, a difficult task with target domain significantly different from the source domain, we observe that no layer has outstanding performance. The results echo the findings in [16].", + "bbox": [ + 496, + 598, + 893, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, we conduct experiments on gradually incorporating features extracted from different layers of the pre-trained model, demonstrating the different behaviors on easy and difficult tasks, which can be found in Section 5.4. The results show for easy tasks, more features may lead to overfitting, but for difficult tasks, performance improves with more features. Therefore, when the target domain is very different from the source domain, we need to provide probing model with diverse features extracted from multiple layers of the pre-trained model.", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16851", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3939f6c2b6b291cfdf9500bca9d202299da8bc6a024f8e3cf50e6af13a5703c3.jpg", + "image_caption": [ + "Figure 4. Performance of extracted features with linear/ non-linear transformation on VTAB-1k. Datasets are ordered from easy tasks (left) to difficult tasks (right) according to domain similarity." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 885, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Effects of Non-linear Transformation", + "text_level": 1, + "bbox": [ + 76, + 330, + 398, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since there is no significantly informative layer for difficult domain adaption tasks, a linear combination may not be enough for exploiting the information in extracted features. We further examine the efficacy of non-linear combination on extracted features, as neural networks achieve superior performance through non-linear feature transformation.", + "bbox": [ + 75, + 353, + 470, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ an multiple-layer perceptron (MLP) for nonlinear transformation and compare its performance with linear transformation. Features extracted from different structures of the pre-trained model are concatenated together as input (See Section 4.1 for details).", + "bbox": [ + 75, + 445, + 470, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Figure 4, we clearly observe that adding non-linearity to feature transformation can significantly boost the performance for the difficult domain adaption tasks (i.e. the datasets on right end of Figure 4). This phenomenon indicates that simply performing linear transformation on extracted features can not exploit the full potential of representations extracted from the pre-trained model. On the other hand, we notice that the performance on easy tasks is not improved, and is even hurt with the introduction of non-linearity. We speculate that this is because the linear transformation can already make full use of extracted information, while non-linearity increases model capacity, which eventually leads to overfitting to the training data.", + "bbox": [ + 75, + 520, + 470, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion. Our investigation reveals that the optimal probing model differs depending on the domain of downstream task. Incorporating diverse features from multiple layers of the pre-trained model and performing non-linear transformation on them can significantly improve the performance on difficult cases. However, this yields worse results on easy tasks. On the contrary, simply performing linear combination on features from deep layers can achieve satisfactory performance for easy adaptation tasks, but not for difficult ones. The conflicting behaviors of model probing motivates us to develop a unified framework that can solve the adaptation problems for both easy and difficult adaptation tasks.", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Structured Model Probing", + "text_level": 1, + "bbox": [ + 498, + 329, + 746, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we propose structured model probing, a unified probing method which can deal with both easy and difficult cases of adaptation tasks. We first present how to thoroughly extract features in a structured way, then introduce our probing model with structured regularization.", + "bbox": [ + 496, + 357, + 893, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Structured Feature Extraction", + "text_level": 1, + "bbox": [ + 498, + 449, + 769, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Incorporating diverse features extracted from the pre-trained model can improve the performance on difficult adaptation tasks for model probing, as shown in Section 3.2. To provide diverse features for the probing model, we extract features from all structures in a Vision Transformer (ViT). These structures include features before self-attention, features after self-attention, and features after Feed Forward Network (FFN), for each transformer layer. We also include tokenized image inputs and pre-logits. We use $V = \\{v_{i}\\}_{i}^{|V|}$ represents all structures in ViT, and $|V|$ is the total number of structures. For an input sample, we denote each set of tokens extracted from structure $v_{i}$ as $X_{i} \\in \\mathbb{R}^{t \\times c}$ , where $t$ is the dimension of tokens and $c$ is the dimension of channels.", + "bbox": [ + 496, + 474, + 893, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To reduce feature redundancy while preserving diversity, we apply 1D average pooling along different dimensions on tokens $X_{i}$ , resulting in two types of aggregation. To maintain channel information, we aggregate $X_{i}$ along the token dimension, which produces $x_{i}^{t}$ . And to preserve spatial information, we aggregate $X_{i}$ along the channel dimension, resulting in $x_{i}^{c}$ . Finally, we concatenate all structured features into a single vector, $x = [x_{1}^{t}, x_{1}^{c}, \\dots, x_{|V|}^{t}, x_{|V|}^{c}]$ . We denote $s_{i}$ as the index set of elements in $x$ , and $S = \\{s_{i} | i = 1\\dots |s|\\}$ .", + "bbox": [ + 496, + 675, + 895, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Probing with Structured Regularization", + "text_level": 1, + "bbox": [ + 498, + 829, + 843, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our structured regularization contains two components: structured sparsity regularizer and structured non-linearity regularizer. We introduce them as follows.", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16852", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1 Structured Feature Selection", + "text_level": 1, + "bbox": [ + 76, + 90, + 331, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Section 3.2, the informativeness of extracted features varies with tasks, and redundant features lead to performance degradation on some tasks. Thus, we propose to perform feature selection during the training stage of the probing model to prevent overfitting. A popular way is to use group lasso [16, 43] by constructing groups on feature dimension over the weight matrix of a linear classifier:", + "bbox": [ + 75, + 114, + 468, + 220 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} L (\\theta) + \\lambda \\sum_ {i = 1} ^ {d} | | \\theta_ {i} | | _ {2}, \\text {w h e r e} | | \\theta_ {i} | | _ {2} = \\sqrt {\\sum_ {j = 1} ^ {n} \\theta_ {i j} ^ {2}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 227, + 468, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use $\\theta \\in \\mathbb{R}^{d\\times n}$ to represent weight matrix, where $d$ is the feature dimension and $n$ is the number of classes. However, due to the high dimensionality of features, the final achieved sparsity is hardly the optimal result. It is worth noting that the structure of pre-trained model provides a natural way to divide features as groups, and features extracted from the same structure have higher relevance to each other. Thus, we leverage the structure of the pre-trained model, to achieve the structured sparsity with a structured sparsity regularizer:", + "bbox": [ + 75, + 301, + 472, + 439 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} L (\\theta) + \\lambda \\sum_ {s \\in S} \\frac {1}{| s |} \\| \\theta_ {s} \\| _ {2}, \\text {w h e r e} \\| \\theta_ {s} \\| _ {2} = \\sqrt {\\sum_ {i \\in s} \\sum_ {j = 1} ^ {n} \\theta_ {i j} ^ {2}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 458, + 473, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$s \\in S$ is the index set of a specific structure in feature as described in Section 4.1, and $|s|$ represents its dimension. Compared with constructing groups by features, in our regularizer, we construct groups by leveraging the structure of the pre-trained model. This formalization can significantly reduce the combinations of selection and thus prevent overfitting. With such an advantage, our method achieves superior performance with a small number of samples under high-dimensional features.", + "bbox": [ + 75, + 518, + 470, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2 Structured Non-linear Transformation", + "text_level": 1, + "bbox": [ + 76, + 672, + 400, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our discovery in Section 3.3 shows non-linear transformation can facilitate transfer performance for difficult tasks, while could reduce performance for easy adaptation tasks, mostly due to over-fitting. This means that we need a model that allows flexibly incorporating non-linearity to a linear model. We propose a structured non-linearity regularizer that nicely integrates feature selection (i.e. choosing the subset of informative features from the pre-trained model) with model selection (i.e. choosing either a linear model or nonlinear model for prediction) inspired by hierarchical sparse modeling [28, 42]. Specifically, we construct our model by the combination of a linear model and a neural network:", + "bbox": [ + 75, + 696, + 472, + 876 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf (x) = \\theta^ {\\top} x + f _ {W} (x), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 883, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta \\in \\mathbb{R}^{d\\times n}$ is a linear model, and $f_{W}$ represents an MLP with model weights $W$ . We use $W^{(1)}$ to represent the input layer of the MLP, and the parameters of rest layers are denoted as $W^{(2:)}$ .", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize the feature selection result $||\\theta ||_2$ , the $\\ell_2$ norm of weights assigned to the selected features, to regularize the complexity of the non-linear model:", + "bbox": [ + 496, + 152, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- When a small number of features are good enough for prediction, we expect a small value for $||\\theta||_2$ . By using $||\\theta||_2$ to control the complexity of non-linear model, we expect the resulting model to be mostly linear.", + "- With increasing $||\\theta||_2$ , more diverse structures from the pre-trained model are needed for prediction. By using $||\\theta||_2$ to regularize the size of non-linear model, we expect the resulting model to be more non-linear." + ], + "bbox": [ + 500, + 198, + 890, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A way to achieve such control is to add a constraint to the objective function in Eq.(4) as follows:", + "bbox": [ + 500, + 319, + 890, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta , W} L (\\theta , W) + \\lambda \\sum_ {s \\in S} | | \\theta_ {s} | | _ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 361, + 803, + 392 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\text {s u b j e c t} \\left. | | W _ {s} ^ {(1)} | | _ {2} \\leq M _ {1} | | \\theta_ {s} | | _ {2}, s = 1 \\dots | S |, \\right. \\tag {6} \\\\ \\left| \\left| W ^ {(2:)} \\right| \\right| _ {2} \\leq M _ {2} \\left| \\left| \\theta \\right| \\right| _ {2}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 392, + 890, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There are two constraints in Eq.(6): the first constraint is used to maintain the input feature sparsity of the neural network, and the second term is used to control the non-linearity through the linear model. To achieve efficient implementation, we convert the constraint in Eq.(6) into a regularization term, which we call structured non-linearity regularizer:", + "bbox": [ + 496, + 448, + 893, + 539 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Omega (\\theta , W) = \\sum_ {s \\in S} \\max \\left\\{\\left| \\left| W _ {s} ^ {(1)} \\right| \\right| _ {2} - M _ {1} \\left| \\left| \\theta_ {s} \\right| \\right| _ {2}, 0 \\right\\} \\tag {7} \\\\ + \\max \\left\\{\\left| \\left| W ^ {(2:)} \\right| \\right| _ {2} - M _ {2} \\left| \\left| \\theta \\right| \\right| _ {2}, 0 \\right\\}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 550, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This regularization term means that, when $||W||_2 > M||\\theta||_2$ , $W$ receives penalty until it achieves similar complexity as $M||\\theta||$ . Then our final objective function is:", + "bbox": [ + 500, + 616, + 890, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta , W} L (\\theta , W) + \\lambda_ {1} \\sum_ {s \\in S} | | \\theta_ {s} | | _ {2} + \\lambda_ {2} \\Omega (\\theta , W). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 675, + 890, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use $\\lambda$ to control the penalty strength, and simply set $\\lambda_{2} = 0.1$ works well in our experiments.", + "bbox": [ + 500, + 719, + 890, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss function. Since our model $f(x)$ consists of both linear and non-linear parts, we decouple these two components in loss function. This modification enables the structured sparsity regularizer to control the complexity of $\\| \\theta \\|$ without the interference from the non-linear part:", + "bbox": [ + 496, + 752, + 890, + 828 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL (\\theta , W) = \\operatorname {C E} \\left(\\theta^ {\\top} x, y\\right) + \\operatorname {C E} \\left(\\operatorname {s g} \\left(\\theta^ {\\top} x\\right) + f _ {W} (x), y\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 839, + 890, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y$ is the one-hot label vector of $x$ , CE is cross-entropy loss and sg means stop-gradient.", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16853", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/573bcd4ee05f2e73a3fd2ac396c5f700fde6e2bcc0248c0abd0cfb3fb1862da7.jpg", + "table_caption": [ + "Table 1. Median test accuracy (%) over 3 seeds on the VTAB-1k benchmark using ViT-B/16 pretrained on ImageNet-21k. * indicates results are obtained from [23] and † indicates results are obtained from [16]." + ], + "table_footnote": [], + "table_body": "
NaturalSpecializedStructured
CTFAR-100Caltech101DTDFlowers102PetsSVHNSun397Mean (Natural)CamelyonEuroSATResisc45RetinopathyMean (Specialized)Clevr-CountClevr-DistDMLabKITTI-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-ElevMean (Structured)Mean (All)
Tuning Methods
Scratch†7.619.113.129.66.719.42.314.071.071.029.372.060.831.652.527.239.166.129.711.724.135.332.8
Fine-tuning*68.987.764.397.286.987.438.875.979.795.784.273.983.756.358.641.765.557.546.725.729.147.665.6
Bias*72.887.059.297.585.359.951.473.378.791.672.969.878.361.555.632.455.966.640.015.725.144.162.0
VPT*78.890.865.898.088.378.149.678.581.896.183.468.482.468.560.046.572.873.647.932.937.855.069.4
Probing Methods
Linear77.286.866.699.188.835.755.172.778.489.672.273.778.531.029.935.254.911.727.714.423.728.655.4
Side-Tuning*60.760.853.695.566.734.935.358.258.587.765.261.068.127.622.631.351.78.214.49.821.823.445.6
Head2Toe75.390.975.099.586.183.550.980.284.495.784.474.384.751.759.444.065.647.140.332.541.147.767.5
LinearStruc.72.989.572.598.983.671.642.776.084.396.081.875.084.348.252.539.466.748.736.732.337.745.364.8
LinearStruc.w/FSR78.290.173.499.388.171.745.478.083.996.181.974.184.048.552.839.664.451.438.632.937.045.765.7
LinearStruc.w/SSR79.090.274.199.190.473.454.280.184.696.383.874.484.848.852.939.866.948.738.932.537.945.866.6
MLPStruc.66.888.770.397.482.171.133.372.881.695.977.073.582.077.257.739.366.572.344.132.747.754.767.1
SMP w/o SNR71.389.672.399.085.672.745.076.584.295.981.972.883.777.457.140.867.172.444.232.648.455.069.0
SMP79.390.974.999.390.475.055.380.784.896.383.175.084.877.558.040.867.572.544.533.049.055.470.9
", + "bbox": [ + 76, + 116, + 893, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 441, + 209, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate the proposed structured model probing method on multiple datasets across various tasks, and provide detailed analysis of the proposed method.", + "bbox": [ + 76, + 465, + 470, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We evaluate SMP on the popular vision backbone, ViT-B/16 [15] pre-trained on ImageNet-21k [12], which is widely used across relative research works [16, 23]. We provide details of probing model and hyperparameter selection in the supplementary material.", + "bbox": [ + 76, + 513, + 470, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared methods. We compare SMP with other commonly used and state-of-the-art tuning (Fine-tuning, Bias [6] and VPT [23]) and probing methods (Linear Probing, Side-Tuning [45] and Head2Toe [16]). All hyperparameters of baseline methods are carefully tuned, and further information is provided in supplementary material.", + "bbox": [ + 76, + 593, + 470, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiments on VTAB-1k", + "text_level": 1, + "bbox": [ + 76, + 693, + 313, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental results on the VTAB-1k benchmark (Table 1) indicate that SMP outperforms other methods, achieving a mean performance of $70.9\\%$ , $3.4\\%$ higher than Head2Toe and $1.5\\%$ higher than VPT. SMP also significantly outperforms Linear Probing, Bias, and Side-Tuning. SMP achieves optimal results in all three dataset groups. Especially in Structured group, SMP attains a mean accuracy of $55.4\\%$ , surpassing other probing methods by a significant margin. This demonstrates the incorporation of non-linear transformation is important for difficult adaptation tasks.", + "bbox": [ + 76, + 715, + 470, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation studies. We provide results of baseline methods that contribute to SMP, to validate the efficacy of structured", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "regularization and the contribution of each component.", + "bbox": [ + 500, + 443, + 861, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LinearStruc. represents extracting features by structured feature extraction proposed in Section 4.1, then training a linear classifier. Compare to Linear Probing, it gets $9\\%$ average performance gain, indicating that solely extracting features from the last layer is inadequate for model probing.", + "bbox": [ + 496, + 458, + 893, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further compare feature sparsity regularization in Eq.(3) (represents by $\\mathrm{Linear}_{\\mathrm{Struc.}} + \\mathrm{FSR}$ ) to the proposed structured sparsity regularizer (represents by $\\mathrm{Linear}_{\\mathrm{Struc.}} + \\mathrm{SSR}$ ). Structured sparsity regularizer outperforms feature sparsity regularization, indicating that feature selection can mitigate the overfitting caused by high dimensions, and leveraging model structure as a prior can bring extra benefits.", + "bbox": [ + 496, + 534, + 893, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We apply non-linear transformation on extracted features, represents by $\\mathrm{MLP}_{\\text{Struc.}}$ . We find the performance on structured group gets a significant improvement over $\\mathrm{Linear}_{\\text{Struc.}}$ , demonstrating the importance of non-linear transformation on difficult adaptation tasks. However, the performance on easy tasks, i.e., natural and specialized groups, is degraded, suggesting the risk of overfitting from non-linear transformation. We solve this problem by leveraging structured nonlinearity regularizer (SNR), which can automatically control the non-linearity through the norms of selected structures.", + "bbox": [ + 496, + 640, + 893, + 791 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Experiments on Few-shot Learning", + "text_level": 1, + "bbox": [ + 500, + 801, + 807, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate our method in few-shot learning scenario, where the number of training samples varies from 1 to 16 shots per class following existing studies. Five fine-grained recognition tasks are chosen, i.e., Food101 [4], CUB-200-2011 [39], Stanford Cars [26], Stanford Dogs [24]", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16854", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/817f5dd48fe878aaf1b3177c862d35bfe2ed6a9e9e8e3980f397eaecfecb0609.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 80, + 349, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2fe9524fb302c804246dd5f00e0babe0a03ccc14a1e9abb24d528eab516d278.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 80, + 617, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/484d8aefe8aaa968b7c51fd933bef7233ff1cddb14dede596e27d1856b3004b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 80, + 883, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1dcfc2b477ee64adeaf90836b3949d8250e549672b924d614cbda5c630efdc49.jpg", + "image_caption": [ + "Figure 5. Performance of few-shot learning on fine-grained visual recognition tasks. LinearStr. (F) indicates LinearStruc. w/ FSR." + ], + "image_footnote": [], + "bbox": [ + 83, + 244, + 349, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/47064e96137db55553faad763127478d043cbcae2ca099511d7e6dc3d2b89899.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 244, + 617, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e7681287e02199c4d9a68a98c6b192a4e43bb4f3e93194a7da447464654dafba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 244, + 883, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and Oxford Pets [35]. We compared SMP with two strong baselines, LinearStruc. and LinearStruc. w/ FSR, and two prominent methods, VPT (tuning) and Head2Toe (probing).", + "bbox": [ + 75, + 450, + 470, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental results presented in Figure 5 show that SMP consistently outperforms other methods in most cases, demonstrating its effectiveness in few-shot scenario, and its ability to improve performance with increasing training data. Notably, feature sparsity regularization performs worse than SMP, highlighting the importance of leveraging pre-trained model structure as prior knowledge in low-shot scenario.", + "bbox": [ + 75, + 498, + 468, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Experiments on More Transfer Scenarios", + "text_level": 1, + "bbox": [ + 76, + 619, + 429, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Larger downstream datasets. We provide average results on 9 full-size downstream datasets in Table 2. It can be observed even in the larger data regime, SMP still achieves competitive performance compared with tuning methods. Full results are presented in Suppl. B.1.", + "bbox": [ + 75, + 648, + 468, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different pre-trained models. We present VTAB-1k results on ViT-B/16 and ViT-L/14 pretrained by CLIP in Suppl. B.2. Despite the usage of stronger pre-trained models, SMP continuously outperforms baseline methods due to the incorporation of diverse features and non-linear transformation, as well as its flexible framework.", + "bbox": [ + 75, + 729, + 470, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different architectures. SMP is a versatile method that can be applied to convolutional neural networks as well. In Suppl. B.3, we present the VTAB-1k results obtained from ImageNet pre-trained ResNet-50 models, where SMP outperforms other methods and achieves superior performance.", + "bbox": [ + 75, + 824, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b028cc30c23ebe2fa7447b605a5e3411ab10a9d73667bb5385e9281e3bc16f10.jpg", + "table_caption": [ + "Table 2. Average accuracy (%) on 9 full-size downstream datasets." + ], + "table_footnote": [], + "table_body": "
Fine-tuningVPTLinearSMP
Average Accuracy88.488.781.788.7
", + "bbox": [ + 500, + 467, + 893, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Analysis and Discussion", + "text_level": 1, + "bbox": [ + 500, + 521, + 718, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct various experiments to analyze the components of the proposed SMP method. Full results of this section are presented in Suppl. C.", + "bbox": [ + 496, + 546, + 890, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization of group norms. We visualize the $\\ell_2$ norm of each structure on different tasks in Figure 6, which helps us to understand the behavior of the proposed structured regularization. For Sun397, an easy adaptation task, we can observe the structures from deep layers have larger norms in Figure 6a. This validates our analysis in Section 3.2, that for easy adaptation tasks, features extracted from deep layers contain more information. For Clevr-Dist, a difficult adaptation task, the structures have similar norms overall, which means for difficult tasks, information contained in intermediate layers is also important, and diverse features can improve the performance, as we find in Section 3.2. This motivates the proposed structured non-linearity regularization, utilizing $\\ell_2$ norm to control the complexity of non-linear model.", + "bbox": [ + 496, + 595, + 892, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Retraining probing model by gradually incorporating structures. We retrain a probing model by gradually incorporating structures selected based on the magnitude of their $\\ell_2$ norm, to validate their contribution to the prediction. Results are presented in Figure 7. For Sun397, an easy adaptation task, we can see in Figure 7a that there are redun", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16855", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e99f4b82417c31d3c5087dcde2007334e8fda3dd38a873613a6132ae66d91ab8.jpg", + "image_caption": [ + "(a) Sun397 (easy)" + ], + "image_footnote": [], + "bbox": [ + 84, + 84, + 477, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1fb3d64faad6bf3cad0400800a95c05e9b0fe2efb2efe6b342be0c3a1ad2b55b.jpg", + "image_caption": [ + "(b) Clevr-Dist (difficult)", + "Figure 6. $\\ell_2$ norm of structure groups on different tasks.", + "Figure 7. Retraining performance by gradually incorporating structures with larger $\\ell_{2}$ norm obtained by SMP." + ], + "image_footnote": [], + "bbox": [ + 84, + 218, + 475, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dant structures. And with a few selected structures, we can achieve comparable results compared with the original SMP, justifying the zero norm of some structured observed after performing structure regularization in Figure 6a. However, for Clevr-Dist, a difficult adaptation task, comparable performance is achieved until selecting 55 structures, suggesting intermediate structures also contribute to the final prediction. This demonstrates the necessity of proposed structured regularizer, which resolves the contradiction between easy and difficult domain adaptation tasks.", + "bbox": [ + 75, + 391, + 470, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of feature aggregation designs. The primary objective of performing token and channel aggregation is to minimize the redundancy in the extracted raw features, while maintaining feature diversity in the compact aggregated representation. To achieve this, two types of aggregation are designed to preserve channel and spatial information, respectively. We provide ablation studies in Suppl. C.5. Overall, channel information (aggregating along tokens) is essential for all datasets, while spatial information (aggregating along channels) is beneficial for tasks that are sensitive to spatial location, e.g., dSpr-Loc and sNORB-Azim.", + "bbox": [ + 75, + 545, + 470, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of different structures. To increase the feature diversity, we incorporate features extracted from various structures, coupled with a structured sparse regularizer to conduct model fitting and feature selection simultaneously. This flexible framework enables us to avoid the costly manual selection of candidate structures. We provide analysis of structures in Suppl. C.6. The results show that incorporating all candidate structured features yields best mean results, which demonstrates the efficacy of our method.", + "bbox": [ + 75, + 715, + 470, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cost of SMP. We compare the cost of SMP with baseline methods, including relative adaptation time, extra inference FLOPs, and extra parameters. The average results on VTAB-", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c4addc513272d803bc245c903c01bcb20ac3f895ce7323006e5714743781c417.jpg", + "image_caption": [ + "(a) Sun397 (easy)" + ], + "image_footnote": [], + "bbox": [ + 506, + 85, + 694, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/67152e2b9413ff984d8288c4e3f256e6f520b098e0b19564a1c7c9735b629693.jpg", + "image_caption": [ + "(b) Clevr-Dist (difficult)" + ], + "image_footnote": [], + "bbox": [ + 700, + 85, + 887, + 213 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e757e527962b43def038b34ac3b8d4cde7b8ad9fd8cb10b9f250d99508f9085e.jpg", + "table_caption": [ + "Table 3. Cost comparison (Average results on VTAB-1k). * indicates relative results compared with Fine-tuning." + ], + "table_footnote": [], + "table_body": "
LinearFTVPTSMP
Adaptation time (rel.)*0.0510.970.15
Extra Inference FLOPs (G)0011.010.01
Extra parameters (M)0.0485.830.646.91
", + "bbox": [ + 501, + 316, + 888, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "1k are shown in Table 3. SMP only needs $15\\%$ adaptation time on average compared with fine-tuning. On the contrary, the adaptation time of VPT is $97\\%$ of fine-tuning. This results in a significantly lower adaptation cost of SMP compared to tuning methods. SMP is also very efficient during inference time, and only introduces 0.01G inference FLOPs compared with fine-tuning. It can be noticed that VPT introduces extra 11.01G inference FLOPs, which is far more than SMP. SMP requires 6.91M extra parameters on average, which is the $8\\%$ of fine-tuning. Though the requirement of extra parameters is higher than VPT, we can leverage the retraining method proposed in Section 5.4, which can further reduce the extra parameters. Overall, SMP is a highly efficient method with low adaptation cost, fast inference, and fewer extra parameters, while maintaining superior performance compared to tuning-based methods.", + "bbox": [ + 496, + 405, + 893, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 662, + 617, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present structured model probing, an effective yet efficient probing method for transfer learning. Our investigation reveals that model probing behaves differently for easy and difficult adaptation tasks. To mitigate this conflicting behavior, we propose structured model probing, a method that is able to achieve good performance in both cases. With the proposed structured regularization, we can construct a simple adaptation model for easy adaptation cases, and increase the complexity of the adaptation model for difficult cases. Our method outperforms state-of-the-art tuning methods while maintaining computational efficiency. Our work highlights the potential of model probing, which motivates researchers to better exploit the use large-scale pre-trained model and apply model probing to diverse tasks.", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16856", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alessandro Achille, Michael Lam, Rahul Tewari, Avinash Ravichandran, Subhransu Maji, Charless C. Fowlkes, Stefano Soatto, and Pietro Perona. Task2vec: Task embedding for meta-learning. In IEEE International Conference on Computer Vision (ICCV), pages 6429-6438, 2019. 3", + "[2] Alessandro Achille, Aditya Golatkar, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. LQF: linear quadratic fine-tuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 15729-15739, 2021. 3", + "[3] Armen Aghajanyan, Akshit Shrivastava, Anchit Gupta, Naman Goyal, Luke Zettlemoyer, and Sonal Gupta. Better finetuning by reducing representational collapse. In 9th International Conference on Learning Representations (ICLR), 2021. 2", + "[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In European Conference on Computer Vision (ECCV), pages 446-461, 2014. 6", + "[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33 (NeurIPS), 2020. 1, 2", + "[6] Han Cai, Chuang Gan, Ligeng Zhu, and Song Han. Tinytl: Reduce memory, not parameters for efficient on-device learning. In Advances in Neural Information Processing Systems 33 (NeurIPS), 2020. 2, 6", + "[7] Shoufa Chen, Chongjian GE, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In Advances in Neural Information Processing Systems 35 (NeurIPS), 2022. 2", + "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 1597-1607, 2020. 2", + "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In IEEE International Conference on Computer Vision (ICCV), pages 9620-9629, 2021. 2", + "[10] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4109-4118, 2018. 3", + "[11] Mostafa Dehghani, Yi Tay, Anurag Arnab, Lucas Beyer, and Ashish Vaswani. The efficiency misnomer. In 10th International Conference on Learning Representations (ICLR), 2022. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 3, 6", + "[13] Aditya Deshpande, Alessandro Achille, Avinash Ravichandran, Hao Li, Luca Zancato, Charless C. Fowlkes, Rahul Bhotika, Stefano Soatto, and Pietro Perona. A linearized framework and a new benchmark for model selection for fine-tuning. arXiv:2102.00084, 2021. 3", + "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), pages 4171–4186, 2019. 2", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations (ICLR), 2021. 3, 6", + "[16] Utku Evci, Vincent Dumoulin, Hugo Larochelle, and Michael C. Mozer. Head2Toe: Utilizing intermediate representations for better transfer learning. In Proceedings of the 39th International Conference on Machine Learning (ICML), pages 6009-6033, 2022. 1, 3, 5, 6", + "[17] Weifeng Ge and Yizhou Yu. Borrowing treasures from the wealthy: Deep transfer learning through selective joint finetuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10-19, 2017. 2", + "[18] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogério Schmidt Feris. SpotTune: Transfer learning through adaptive fine-tuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4805-4814, 2019. 2", + "[19] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In 10th International Conference on Learning Representations (ICLR), 2022. 1, 2", + "[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 2, 3", + "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, 2022. 2, 3", + "[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning (ICML), pages 2790-2799, 2019. 2", + "[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge J. Belongie, Bharath Hariharan, and Ser-Nam Lim." + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16857", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6", + "[24] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition (CVPR Workshop), 2011. 6", + "[25] Simon Kornblith, Jonathon Shlens, and Quoc V. Le. Do better imagenet models transfer better? In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2661-2671, 2019. 2, 3", + "[26] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE International Conference on Computer Vision Workshop (ICCV Workshop), pages 554–561, 2013. 6", + "[27] Ananya Kumar, Aditi Raghunathan, Robbie Matthew Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In 10th International Conference on Learning Representations (ICLR), 2022. 2", + "[28] Ismael Lemhadri, Feng Ruan, Louis Abraham, and Robert Tibshirani. Lassonet: A neural network with feature sparsity. Journal of Machine Learning Research, 22:127:1-127:29, 2021. 5", + "[29] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3045-3059, 2021. 2", + "[30] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), pages 4582-4597, 2021. 2", + "[31] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5196-5205, 2022. 2", + "[32] Fangzhou Mu, Yingyu Liang, and Yin Li. Gradients as features for deep representation learning. In 8th International Conference on Learning Representations (ICLR), 2020. 3", + "[33] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. LEEP: A new measure to evaluate transferability of learned representations. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 7294-7305, 2020. 3", + "[34] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345-1359, 2010. 2", + "[35] Omkar M. Parkhi, Andreae Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3498-3505, 2012. 7", + "[36] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), pages 2227-2237, 2018. 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 1, 2", + "[38] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. LST: Ladder side-tuning for parameter and memory efficient transfer learning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3", + "[39] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical report, California Institute of Technology, 2011. 6", + "[40] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In 10th International Conference on Learning Representations (ICLR), 2022. 2", + "[41] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer G. Dy, and Tomas Pfister. Learning to prompt for continual learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 139-149, 2022. 2", + "[42] Xiaohan Yan and Jacob Bien. Hierarchical Sparse Modeling: A choice of two group lasso formulations. Statistical Science, 32(4):531 - 560. 5", + "[43] Ming Yuan and Yi Lin. Model selection and estimation in regression with grouped variables. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 68(1): 49-67, 2006. 5", + "[44] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv:1910.04867, 2019. 1, 2, 3", + "[45] Jeffrey O. Zhang, Alexander Sax, Amir Zamir, Leonidas J. Guibas, and Jitendra Malik. Side-tuning: A baseline for network adaptation via additive side networks. In European Conference on Computer Vision (ECCV), pages 698-714, 2020. 1, 2, 3, 6", + "[46] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 16795-16804, 2022. 2", + "[47] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2", + "[48] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2021. 2" + ], + "bbox": [ + 501, + 92, + 893, + 867 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16858", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/fcd62d92-907d-4f7f-bc80-b88db83abf7f_model.json b/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/fcd62d92-907d-4f7f-bc80-b88db83abf7f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..be38b1e02804357dca4caf0e576513050e3e447a --- /dev/null +++ b/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/fcd62d92-907d-4f7f-bc80-b88db83abf7f_model.json @@ -0,0 +1,2266 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.122, + 0.107, + 0.839, + 0.152 + ], + "angle": 0, + "content": "Structured Model Probing: Empowering Efficient Transfer Learning by Structured Regularization" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.179, + 0.816, + 0.222 + ], + "angle": 0, + "content": "Zhi-Fan Wu Chaojie Mao Xue Wang Jianwen Jiang Yiliang Lv Rong Jin Alibaba Group" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.28, + 0.314, + 0.297 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.314, + 0.474, + 0.722 + ], + "angle": 0, + "content": "Despite encouraging results from recent developments in transfer learning for adapting pre-trained model to downstream tasks, the performance of model probing is still lagging behind the state-of-the-art parameter efficient tuning methods. Our investigation reveals that existing model probing methods perform well for the easy case when the source domain (where models are pre-trained) and the adapted domain are similar, but fail for the difficult case when the two domains are significantly different. Simply incorporating features extracted from multiple layers and increasing complexity of the probing model can mitigate the gap in the difficult case, but degrades the performance in the easy case. To address this challenge, we propose structured model probing (SMP) that is able to deliver good performance for both cases through structured regularization. The regularization performs feature selection leveraging model structure as a prior, and controls the complexity of the probing model through the weights of selected structures. This enables us to construct a simple adaptation model, with a small number of selected features and a linear prediction model, for the easy case; and to automatically increase the complexity of adaptation model, with a large number of selected features and a non-linear model, for the difficult case. Our extensive empirical studies show that SMP significantly outperforms the state-of-the-art methods for parameter efficient tuning, and at the same time, still maintains the advantage of computational efficiency for probing-based methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.753, + 0.21, + 0.769 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.472, + 0.902 + ], + "angle": 0, + "content": "The parameters of models have grown drastically with the rapid development of deep learning [5, 37], posing challenges for adapting pre-trained models to downstream tasks. The widely used transfer strategy, fully fine-tuning, becomes infeasible due to excessive computation and storage overhead. Recently, parameter efficient tuning [19, 23] has gained attention for reducing storage costs and improving tuning performance with few trainable parameters when lim" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.284, + 0.889, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.442, + 0.895, + 0.47 + ], + "angle": 0, + "content": "Figure 1. Performance comparison on different dataset groups (left) and mean relative adaptation time (right) on VTAB-1k benchmark." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.895, + 0.566 + ], + "angle": 0, + "content": "ited data are available for downstream tasks. Despite the progress, parameter efficient tuning methods still face significant computational overhead in training due to the need to perform forward and backward passes through the pretrained model at each training step." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.896, + 0.764 + ], + "angle": 0, + "content": "Model probing, which leverages features extracted from the pre-trained model to perform adaptation, is promising due to its efficiency and simplicity: The training procedure is efficient since the pre-trained model is fixed and only performs one forward propagation to extract features, and probing model is decoupled from pre-trained model that makes it convenient to deploy small sized task-specific downstream models. However, it is a nuisance that probing methods are usually outperformed by state-of-the-art parameter efficient tuning methods. This is true even after several recent improvements made for the probing methods [16, 45]. In this paper, we aim to improve the performance of model probing significantly without sacrificing its simplicity and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.897, + 0.902 + ], + "angle": 0, + "content": "We first investigate when model probing methods perform worse than the tuning-based methods on widely used visual task adaptation benchmark VTAB-1k [44] as shown in Figure 1. Target domains are categorized as easy and difficult cases based on their similarity to the source domain where the pre-trained model is trained on (see Section 3.1 for details). We find simple linear probing yields comparable performance as the two tuning-based approaches (finetuning and VPT [23]) for easy cases, whereas a significant" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16849" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.403, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.087, + 0.891, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.17, + 0.256, + 0.796, + 0.272 + ], + "angle": 0, + "content": "Figure 2. Adaptation pipelines of Model Tuning (left) and the proposed Structured Model Probing (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.471, + 0.362 + ], + "angle": 0, + "content": "performance gap is observed between linear probing and tuning-based methods for difficult cases. This suggests that model probing exhibits different behavior for various target domains, and performance gap arises for target domains that dissimilar to the source domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.47, + 0.546 + ], + "angle": 0, + "content": "This is not surprising that we expect a more complicated model than a linear probing to compensate the difference between the source domain and the target domain. In general, we can increase model complexity in two ways, either by introducing more features extracted from the pre-trained model, or by introducing a non-linear model for prediction. A main challenge is how to develop a unified framework of model probing that is able to handle both the easy and difficult adaptations effectively. This is because, according to our study in Section 3, for easy cases, a complicated probing model (with more features and non-linear probing) performs significantly worse than a simple linear probing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.471, + 0.789 + ], + "angle": 0, + "content": "To this end, we propose structured model probing, or SMP for short, that effectively address both easy and difficult cases of domain adaptation. The key to our unified framework is structured regularization. It performs feature selection leveraging the structure of the pre-trained model as a prior, and controls the complexity of the probing model through the weights assigned to selected features. Thus, for the easy case, this regularization constructs a simple linear model with a few selected features; and for the difficult case, it automatically introduces more features from the pre-trained model and more non-linearity in the prediction module based on the value of loss function, as shown in Figure 2. Our empirical studies show that the proposed structured model probing performs significantly better than state-of-the-art tuning-based approaches, while maintaining the advantage of computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.79, + 0.352, + 0.804 + ], + "angle": 0, + "content": "Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.81, + 0.471, + 0.901 + ], + "angle": 0, + "content": "1. We observe and demonstrate the conflicting behaviors of model probing: a complicated probing model with more features and non-linear transformation can improve the performance for difficult cases of domain adaptation, but it yields worse performance than linear probing for easy cases of domain adaptation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.287, + 0.894, + 0.378 + ], + "angle": 0, + "content": "2. We propose structured model probing (SMP) based on structured regularization. This regularization enables us to construct simple probing model for easy tasks, and automatically increase model complexity for difficult tasks, to simultaneously address both the easy and difficult cases of domain adaptation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.378, + 0.894, + 0.454 + ], + "angle": 0, + "content": "3. We conduct extensive empirical studies on various adaptation tasks. The proposed SMP method shows superior experimental results compared to state-of-the-art probing and tuning methods, and requires significantly lower training costs compared to tuning methods." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.287, + 0.894, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.471, + 0.641, + 0.487 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.894, + 0.591 + ], + "angle": 0, + "content": "Transfer learning has been extensively studied in vision domain [34, 48]. It aims to improve the performance on target domain by leveraging the related information contained in source domain. Transfer learning based on pre-trained model [8, 20, 21] has received significant attention due to its simplicity and good empirical performance [23, 45]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.895, + 0.868 + ], + "angle": 0, + "content": "Model tuning adapts the pre-trained model to new tasks via updating model parameters. Fully fine-tuning is a widely used strategies due to its efficacy and simplicity [9, 25, 44]. However, it suffers from huge storage and computation overhead due to the ever-expanding size of large-scale pre-trained model [37, 40]. Heuristic approaches have been proposed to improve fine-tuning performance [3, 6, 17, 18, 27]. Parameter efficient tuning [19] aims to reduce the storage requirement of transferred model by updating a few parameters attached to the pre-trained model [23, 47], like Prompt Tuning [29, 30] or Adapter [22]. While originally studied in natural language processing tasks due to the dominance of large-scale pre-trained language models [5, 14], there is growing interest in applying these methods to computer vision tasks [7, 23, 31, 41, 46, 47]. Though parameter efficient tuning reduces storage cost significantly, the training cost remains high due to the forward and backward propagation need to pass through the entire pre-trained model [11]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Model probing utilizes frozen extracted features for adapting the pre-trained model to new tasks, thus achieving effi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16850" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "cient adaptation. The most popular method, linear probing, suffers from performance degradation [25], which limits its application in transfer learning [20, 21]. Recent methods [16, 36, 45] have been proposed to improve probing performance by leveraging side network [38, 45] or incorporating intermediate representations [2, 16, 32], but there is still a significant performance gap compared to tuning methods. This motivates us to investigate the limitation of probing methods, and come up with a better solution." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.243, + 0.473, + 0.279 + ], + "angle": 0, + "content": "3. Understanding the Limitation of Model Probing for Transfer Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.286, + 0.472, + 0.362 + ], + "angle": 0, + "content": "In this section, we analyze the performance of probing methods across various downstream domains to understand their limitations, motivated by the observation in Figure 1. We assess feature informativeness for different tasks and investigate the necessity of non-linear transformation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.372, + 0.228, + 0.388 + ], + "angle": 0, + "content": "3.1. Analysis Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.47, + 0.46 + ], + "angle": 0, + "content": "Pre-trained model and source data. The ViT-B/16 [15] is used as the backbone because of the extensive usage of transformer-based models. The model is pre-trained on ImageNet-21k [12] in a supervised learning manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.569 + ], + "angle": 0, + "content": "Target data. We use the VTAB-1k benchmark, consisting of 19 datasets that cover natural images (natural), images captured by specialist equipment (specialized), and images generated from simulated environments (structured), as our target data for transfer learning. We follow the setting in [44] and use 1000 samples as the training set per task to evaluate adaptation with limited data as suggested in [44]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.47, + 0.633 + ], + "angle": 0, + "content": "Domain similarity. Several previous works have proposed metrics to measure the similarity between source domain and target domain [1, 10, 13, 16, 33]. Here, we adopt the domain similarity measure defined in [16]:" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.646, + 0.469, + 0.661 + ], + "angle": 0, + "content": "\\[\nD o m a i n \\quad \\text {S i m i l a r i t y} = \\mathrm {A c c} _ {\\text {L i n e a r}} - \\mathrm {A c c} _ {\\text {S c r a t c h}}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.472, + 0.84 + ], + "angle": 0, + "content": "A higher domain similarity suggests that the features extracted from a pre-trained model can be utilized for the target domain, leading to performance gains via a linear classifier compared to training from scratch. Conversely, a lower domain similarity indicates that training from scratch on the target domain may be more effective than using extracted features. Thus, the domain similarity is evaluated based on the benefit gained by a linear classifier compared to training from scratch. We also report the label-feature correlation [13] in Suppl. C.8, which exhibits a high Spearman rank correlation (0.854) with the domain similarity scores." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Domain similarity also reflects how difficult to adapt from source domain to target domain. As shown in Figure 4, tasks in natural and specialized groups have higher domain similarity to source domain where the pre-trained model is" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.093, + 0.697, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.093, + 0.887, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.251, + 0.892, + 0.28 + ], + "angle": 0, + "content": "Figure 3. Performance of features extracted from different layers on Sun397 (easy) and Clevr-Dist (difficult)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.298, + 0.892, + 0.343 + ], + "angle": 0, + "content": "trained on, so we mark them as easy cases. And tasks in structured groups have lower domain similarity to source domain, thus they are labeled as difficult cases in Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.354, + 0.83, + 0.369 + ], + "angle": 0, + "content": "3.2. Informativeness of Extracted Features" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.378, + 0.892, + 0.424 + ], + "angle": 0, + "content": "To quantify the informativeness of features extracted from different layers for downstream tasks, we use leave-one-out \\( k \\)-NN testing as the surrogate performance criterion:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.447, + 0.892, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\text {I n f o r m a t i v e n e s s} = \\frac {\\sum_ {i = 1} ^ {m} \\sum_ {x _ {j} \\in N (x _ {i})} \\mathbb {1} \\left\\{y _ {i} = y _ {j} \\right\\}}{k m}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.492, + 0.895, + 0.598 + ], + "angle": 0, + "content": "where \\( m \\) is the number of samples, \\( x \\) is the extracted features of a sample, \\( N(x_{i}) \\) represents \\( k \\) nearest neighbors of \\( x_{i} \\). We simply set \\( k = 1 \\), which results in leave-one-out 1-NN testing. We choose \\( k \\)-NN, instead of linear classifier, for informativeness measurement because \\( k \\)-NN is a non-parametric classifier and thus avoids training additional parameters, making the measurement more robust." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.75 + ], + "angle": 0, + "content": "In Figure 3, we plot the informativeness of features extracted from different layers on two tasks from different domains. For Sun397, an easy domain adaption task, we observe the best performance achieved by features extracted from deep layers. This explains the good performance of linear probing for easy domain adaption tasks, as observed in Figure 1. While for Clevr-Dist, a difficult task with target domain significantly different from the source domain, we observe that no layer has outstanding performance. The results echo the findings in [16]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Furthermore, we conduct experiments on gradually incorporating features extracted from different layers of the pre-trained model, demonstrating the different behaviors on easy and difficult tasks, which can be found in Section 5.4. The results show for easy tasks, more features may lead to overfitting, but for difficult tasks, performance improves with more features. Therefore, when the target domain is very different from the source domain, we need to provide probing model with diverse features extracted from multiple layers of the pre-trained model." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16851" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.887, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.289, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Figure 4. Performance of extracted features with linear/ non-linear transformation on VTAB-1k. Datasets are ordered from easy tasks (left) to difficult tasks (right) according to domain similarity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.331, + 0.4, + 0.347 + ], + "angle": 0, + "content": "3.3. Effects of Non-linear Transformation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.471, + 0.445 + ], + "angle": 0, + "content": "Since there is no significantly informative layer for difficult domain adaption tasks, a linear combination may not be enough for exploiting the information in extracted features. We further examine the efficacy of non-linear combination on extracted features, as neural networks achieve superior performance through non-linear feature transformation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.472, + 0.52 + ], + "angle": 0, + "content": "We employ an multiple-layer perceptron (MLP) for nonlinear transformation and compare its performance with linear transformation. Features extracted from different structures of the pre-trained model are concatenated together as input (See Section 4.1 for details)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.472, + 0.718 + ], + "angle": 0, + "content": "In Figure 4, we clearly observe that adding non-linearity to feature transformation can significantly boost the performance for the difficult domain adaption tasks (i.e. the datasets on right end of Figure 4). This phenomenon indicates that simply performing linear transformation on extracted features can not exploit the full potential of representations extracted from the pre-trained model. On the other hand, we notice that the performance on easy tasks is not improved, and is even hurt with the introduction of non-linearity. We speculate that this is because the linear transformation can already make full use of extracted information, while non-linearity increases model capacity, which eventually leads to overfitting to the training data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Discussion. Our investigation reveals that the optimal probing model differs depending on the domain of downstream task. Incorporating diverse features from multiple layers of the pre-trained model and performing non-linear transformation on them can significantly improve the performance on difficult cases. However, this yields worse results on easy tasks. On the contrary, simply performing linear combination on features from deep layers can achieve satisfactory performance for easy adaptation tasks, but not for difficult ones. The conflicting behaviors of model probing motivates us to develop a unified framework that can solve the adaptation problems for both easy and difficult adaptation tasks." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.33, + 0.747, + 0.348 + ], + "angle": 0, + "content": "4. Structured Model Probing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.895, + 0.435 + ], + "angle": 0, + "content": "In this section, we propose structured model probing, a unified probing method which can deal with both easy and difficult cases of adaptation tasks. We first present how to thoroughly extract features in a structured way, then introduce our probing model with structured regularization." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.45, + 0.771, + 0.466 + ], + "angle": 0, + "content": "4.1. Structured Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.895, + 0.674 + ], + "angle": 0, + "content": "Incorporating diverse features extracted from the pre-trained model can improve the performance on difficult adaptation tasks for model probing, as shown in Section 3.2. To provide diverse features for the probing model, we extract features from all structures in a Vision Transformer (ViT). These structures include features before self-attention, features after self-attention, and features after Feed Forward Network (FFN), for each transformer layer. We also include tokenized image inputs and pre-logits. We use \\( V = \\{v_{i}\\}_{i}^{|V|} \\) represents all structures in ViT, and \\( |V| \\) is the total number of structures. For an input sample, we denote each set of tokens extracted from structure \\( v_{i} \\) as \\( X_{i} \\in \\mathbb{R}^{t \\times c} \\), where \\( t \\) is the dimension of tokens and \\( c \\) is the dimension of channels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.676, + 0.897, + 0.817 + ], + "angle": 0, + "content": "To reduce feature redundancy while preserving diversity, we apply 1D average pooling along different dimensions on tokens \\( X_{i} \\), resulting in two types of aggregation. To maintain channel information, we aggregate \\( X_{i} \\) along the token dimension, which produces \\( x_{i}^{t} \\). And to preserve spatial information, we aggregate \\( X_{i} \\) along the channel dimension, resulting in \\( x_{i}^{c} \\). Finally, we concatenate all structured features into a single vector, \\( x = [x_{1}^{t}, x_{1}^{c}, \\dots, x_{|V|}^{t}, x_{|V|}^{c}] \\). We denote \\( s_{i} \\) as the index set of elements in \\( x \\), and \\( S = \\{s_{i} | i = 1\\dots |s|\\} \\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.83, + 0.844, + 0.846 + ], + "angle": 0, + "content": "4.2. Probing with Structured Regularization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Our structured regularization contains two components: structured sparsity regularizer and structured non-linearity regularizer. We introduce them as follows." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16852" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.092, + 0.332, + 0.106 + ], + "angle": 0, + "content": "4.2.1 Structured Feature Selection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.221 + ], + "angle": 0, + "content": "As shown in Section 3.2, the informativeness of extracted features varies with tasks, and redundant features lead to performance degradation on some tasks. Thus, we propose to perform feature selection during the training stage of the probing model to prevent overfitting. A popular way is to use group lasso [16, 43] by constructing groups on feature dimension over the weight matrix of a linear classifier:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.228, + 0.469, + 0.276 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} L (\\theta) + \\lambda \\sum_ {i = 1} ^ {d} | | \\theta_ {i} | | _ {2}, \\text {w h e r e} | | \\theta_ {i} | | _ {2} = \\sqrt {\\sum_ {j = 1} ^ {n} \\theta_ {i j} ^ {2}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.473, + 0.44 + ], + "angle": 0, + "content": "We use \\(\\theta \\in \\mathbb{R}^{d\\times n}\\) to represent weight matrix, where \\(d\\) is the feature dimension and \\(n\\) is the number of classes. However, due to the high dimensionality of features, the final achieved sparsity is hardly the optimal result. It is worth noting that the structure of pre-trained model provides a natural way to divide features as groups, and features extracted from the same structure have higher relevance to each other. Thus, we leverage the structure of the pre-trained model, to achieve the structured sparsity with a structured sparsity regularizer:" + }, + { + "type": "equation", + "bbox": [ + 0.079, + 0.459, + 0.475, + 0.519 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} L (\\theta) + \\lambda \\sum_ {s \\in S} \\frac {1}{| s |} \\| \\theta_ {s} \\| _ {2}, \\text {w h e r e} \\| \\theta_ {s} \\| _ {2} = \\sqrt {\\sum_ {i \\in s} \\sum_ {j = 1} ^ {n} \\theta_ {i j} ^ {2}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.472, + 0.655 + ], + "angle": 0, + "content": "\\(s \\in S\\) is the index set of a specific structure in feature as described in Section 4.1, and \\(|s|\\) represents its dimension. Compared with constructing groups by features, in our regularizer, we construct groups by leveraging the structure of the pre-trained model. This formalization can significantly reduce the combinations of selection and thus prevent overfitting. With such an advantage, our method achieves superior performance with a small number of samples under high-dimensional features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.673, + 0.401, + 0.687 + ], + "angle": 0, + "content": "4.2.2 Structured Non-linear Transformation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.473, + 0.877 + ], + "angle": 0, + "content": "Our discovery in Section 3.3 shows non-linear transformation can facilitate transfer performance for difficult tasks, while could reduce performance for easy adaptation tasks, mostly due to over-fitting. This means that we need a model that allows flexibly incorporating non-linearity to a linear model. We propose a structured non-linearity regularizer that nicely integrates feature selection (i.e. choosing the subset of informative features from the pre-trained model) with model selection (i.e. choosing either a linear model or nonlinear model for prediction) inspired by hierarchical sparse modeling [28, 42]. Specifically, we construct our model by the combination of a linear model and a neural network:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.884, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\nf (x) = \\theta^ {\\top} x + f _ {W} (x), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "where \\(\\theta \\in \\mathbb{R}^{d\\times n}\\) is a linear model, and \\(f_{W}\\) represents an MLP with model weights \\(W\\). We use \\(W^{(1)}\\) to represent the input layer of the MLP, and the parameters of rest layers are denoted as \\(W^{(2:)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.892, + 0.198 + ], + "angle": 0, + "content": "We utilize the feature selection result \\(||\\theta ||_2\\), the \\(\\ell_2\\) norm of weights assigned to the selected features, to regularize the complexity of the non-linear model:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.199, + 0.892, + 0.258 + ], + "angle": 0, + "content": "- When a small number of features are good enough for prediction, we expect a small value for \\(||\\theta||_2\\). By using \\(||\\theta||_2\\) to control the complexity of non-linear model, we expect the resulting model to be mostly linear." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.259, + 0.892, + 0.319 + ], + "angle": 0, + "content": "- With increasing \\( ||\\theta||_2 \\), more diverse structures from the pre-trained model are needed for prediction. By using \\( ||\\theta||_2 \\) to regularize the size of non-linear model, we expect the resulting model to be more non-linear." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.199, + 0.892, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.32, + 0.892, + 0.35 + ], + "angle": 0, + "content": "A way to achieve such control is to add a constraint to the objective function in Eq.(4) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.362, + 0.804, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta , W} L (\\theta , W) + \\lambda \\sum_ {s \\in S} | | \\theta_ {s} | | _ {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.393, + 0.891, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\text {s u b j e c t} \\left. | | W _ {s} ^ {(1)} | | _ {2} \\leq M _ {1} | | \\theta_ {s} | | _ {2}, s = 1 \\dots | S |, \\right. \\tag {6} \\\\ \\left| \\left| W ^ {(2:)} \\right| \\right| _ {2} \\leq M _ {2} \\left| \\left| \\theta \\right| \\right| _ {2}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.894, + 0.54 + ], + "angle": 0, + "content": "There are two constraints in Eq.(6): the first constraint is used to maintain the input feature sparsity of the neural network, and the second term is used to control the non-linearity through the linear model. To achieve efficient implementation, we convert the constraint in Eq.(6) into a regularization term, which we call structured non-linearity regularizer:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.551, + 0.892, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Omega (\\theta , W) = \\sum_ {s \\in S} \\max \\left\\{\\left| \\left| W _ {s} ^ {(1)} \\right| \\right| _ {2} - M _ {1} \\left| \\left| \\theta_ {s} \\right| \\right| _ {2}, 0 \\right\\} \\tag {7} \\\\ + \\max \\left\\{\\left| \\left| W ^ {(2:)} \\right| \\right| _ {2} - M _ {2} \\left| \\left| \\theta \\right| \\right| _ {2}, 0 \\right\\}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.617, + 0.892, + 0.664 + ], + "angle": 0, + "content": "This regularization term means that, when \\( ||W||_2 > M||\\theta||_2 \\), \\( W \\) receives penalty until it achieves similar complexity as \\( M||\\theta|| \\). Then our final objective function is:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.676, + 0.891, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta , W} L (\\theta , W) + \\lambda_ {1} \\sum_ {s \\in S} | | \\theta_ {s} | | _ {2} + \\lambda_ {2} \\Omega (\\theta , W). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.72, + 0.892, + 0.75 + ], + "angle": 0, + "content": "We use \\(\\lambda\\) to control the penalty strength, and simply set \\(\\lambda_{2} = 0.1\\) works well in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.753, + 0.892, + 0.829 + ], + "angle": 0, + "content": "Loss function. Since our model \\( f(x) \\) consists of both linear and non-linear parts, we decouple these two components in loss function. This modification enables the structured sparsity regularizer to control the complexity of \\( \\| \\theta \\| \\) without the interference from the non-linear part:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.84, + 0.891, + 0.86 + ], + "angle": 0, + "content": "\\[\nL (\\theta , W) = \\operatorname {C E} \\left(\\theta^ {\\top} x, y\\right) + \\operatorname {C E} \\left(\\operatorname {s g} \\left(\\theta^ {\\top} x\\right) + f _ {W} (x), y\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(y\\) is the one-hot label vector of \\(x\\), CE is cross-entropy loss and sg means stop-gradient." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16853" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.077, + 0.892, + 0.107 + ], + "angle": 0, + "content": "Table 1. Median test accuracy (%) over 3 seeds on the VTAB-1k benchmark using ViT-B/16 pretrained on ImageNet-21k. * indicates results are obtained from [23] and † indicates results are obtained from [16]." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.117, + 0.895, + 0.433 + ], + "angle": 0, + "content": "
NaturalSpecializedStructured
CTFAR-100Caltech101DTDFlowers102PetsSVHNSun397Mean (Natural)CamelyonEuroSATResisc45RetinopathyMean (Specialized)Clevr-CountClevr-DistDMLabKITTI-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-ElevMean (Structured)Mean (All)
Tuning Methods
Scratch†7.619.113.129.66.719.42.314.071.071.029.372.060.831.652.527.239.166.129.711.724.135.332.8
Fine-tuning*68.987.764.397.286.987.438.875.979.795.784.273.983.756.358.641.765.557.546.725.729.147.665.6
Bias*72.887.059.297.585.359.951.473.378.791.672.969.878.361.555.632.455.966.640.015.725.144.162.0
VPT*78.890.865.898.088.378.149.678.581.896.183.468.482.468.560.046.572.873.647.932.937.855.069.4
Probing Methods
Linear77.286.866.699.188.835.755.172.778.489.672.273.778.531.029.935.254.911.727.714.423.728.655.4
Side-Tuning*60.760.853.695.566.734.935.358.258.587.765.261.068.127.622.631.351.78.214.49.821.823.445.6
Head2Toe75.390.975.099.586.183.550.980.284.495.784.474.384.751.759.444.065.647.140.332.541.147.767.5
LinearStruc.72.989.572.598.983.671.642.776.084.396.081.875.084.348.252.539.466.748.736.732.337.745.364.8
LinearStruc.w/FSR78.290.173.499.388.171.745.478.083.996.181.974.184.048.552.839.664.451.438.632.937.045.765.7
LinearStruc.w/SSR79.090.274.199.190.473.454.280.184.696.383.874.484.848.852.939.866.948.738.932.537.945.866.6
MLPStruc.66.888.770.397.482.171.133.372.881.695.977.073.582.077.257.739.366.572.344.132.747.754.767.1
SMP w/o SNR71.389.672.399.085.672.745.076.584.295.981.972.883.777.457.140.867.172.444.232.648.455.069.0
SMP79.390.974.999.390.475.055.380.784.896.383.175.084.877.558.040.867.572.544.533.049.055.470.9
" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.442, + 0.21, + 0.459 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.467, + 0.471, + 0.514 + ], + "angle": 0, + "content": "In this section, we evaluate the proposed structured model probing method on multiple datasets across various tasks, and provide detailed analysis of the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.515, + 0.471, + 0.592 + ], + "angle": 0, + "content": "Implementation details. We evaluate SMP on the popular vision backbone, ViT-B/16 [15] pre-trained on ImageNet-21k [12], which is widely used across relative research works [16, 23]. We provide details of probing model and hyperparameter selection in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.594, + 0.472, + 0.686 + ], + "angle": 0, + "content": "Compared methods. We compare SMP with other commonly used and state-of-the-art tuning (Fine-tuning, Bias [6] and VPT [23]) and probing methods (Linear Probing, Side-Tuning [45] and Head2Toe [16]). All hyperparameters of baseline methods are carefully tuned, and further information is provided in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.694, + 0.314, + 0.71 + ], + "angle": 0, + "content": "5.1. Experiments on VTAB-1k" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.716, + 0.472, + 0.868 + ], + "angle": 0, + "content": "Experimental results on the VTAB-1k benchmark (Table 1) indicate that SMP outperforms other methods, achieving a mean performance of \\(70.9\\%\\), \\(3.4\\%\\) higher than Head2Toe and \\(1.5\\%\\) higher than VPT. SMP also significantly outperforms Linear Probing, Bias, and Side-Tuning. SMP achieves optimal results in all three dataset groups. Especially in Structured group, SMP attains a mean accuracy of \\(55.4\\%\\), surpassing other probing methods by a significant margin. This demonstrates the incorporation of non-linear transformation is important for difficult adaptation tasks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Ablation studies. We provide results of baseline methods that contribute to SMP, to validate the efficacy of structured" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.444, + 0.862, + 0.459 + ], + "angle": 0, + "content": "regularization and the contribution of each component." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.894, + 0.535 + ], + "angle": 0, + "content": "LinearStruc. represents extracting features by structured feature extraction proposed in Section 4.1, then training a linear classifier. Compare to Linear Probing, it gets \\(9\\%\\) average performance gain, indicating that solely extracting features from the last layer is inadequate for model probing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.895, + 0.641 + ], + "angle": 0, + "content": "We further compare feature sparsity regularization in Eq.(3) (represents by \\(\\mathrm{Linear}_{\\mathrm{Struc.}} + \\mathrm{FSR}\\)) to the proposed structured sparsity regularizer (represents by \\(\\mathrm{Linear}_{\\mathrm{Struc.}} + \\mathrm{SSR}\\)). Structured sparsity regularizer outperforms feature sparsity regularization, indicating that feature selection can mitigate the overfitting caused by high dimensions, and leveraging model structure as a prior can bring extra benefits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.895, + 0.792 + ], + "angle": 0, + "content": "We apply non-linear transformation on extracted features, represents by \\(\\mathrm{MLP}_{\\text{Struc.}}\\). We find the performance on structured group gets a significant improvement over \\(\\mathrm{Linear}_{\\text{Struc.}}\\), demonstrating the importance of non-linear transformation on difficult adaptation tasks. However, the performance on easy tasks, i.e., natural and specialized groups, is degraded, suggesting the risk of overfitting from non-linear transformation. We solve this problem by leveraging structured nonlinearity regularizer (SNR), which can automatically control the non-linearity through the norms of selected structures." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.808, + 0.818 + ], + "angle": 0, + "content": "5.2. Experiments on Few-shot Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this section, we evaluate our method in few-shot learning scenario, where the number of training samples varies from 1 to 16 shots per class following existing studies. Five fine-grained recognition tasks are chosen, i.e., Food101 [4], CUB-200-2011 [39], Stanford Cars [26], Stanford Dogs [24]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16854" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.081, + 0.351, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.081, + 0.619, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.081, + 0.885, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.245, + 0.351, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.245, + 0.619, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.245, + 0.885, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.423, + 0.859, + 0.438 + ], + "angle": 0, + "content": "Figure 5. Performance of few-shot learning on fine-grained visual recognition tasks. LinearStr. (F) indicates LinearStruc. w/ FSR." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.451, + 0.471, + 0.497 + ], + "angle": 0, + "content": "and Oxford Pets [35]. We compared SMP with two strong baselines, LinearStruc. and LinearStruc. w/ FSR, and two prominent methods, VPT (tuning) and Head2Toe (probing)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.499, + 0.47, + 0.605 + ], + "angle": 0, + "content": "Experimental results presented in Figure 5 show that SMP consistently outperforms other methods in most cases, demonstrating its effectiveness in few-shot scenario, and its ability to improve performance with increasing training data. Notably, feature sparsity regularization performs worse than SMP, highlighting the importance of leveraging pre-trained model structure as prior knowledge in low-shot scenario." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.431, + 0.637 + ], + "angle": 0, + "content": "5.3. Experiments on More Transfer Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.649, + 0.47, + 0.725 + ], + "angle": 0, + "content": "Larger downstream datasets. We provide average results on 9 full-size downstream datasets in Table 2. It can be observed even in the larger data regime, SMP still achieves competitive performance compared with tuning methods. Full results are presented in Suppl. B.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.471, + 0.82 + ], + "angle": 0, + "content": "Different pre-trained models. We present VTAB-1k results on ViT-B/16 and ViT-L/14 pretrained by CLIP in Suppl. B.2. Despite the usage of stronger pre-trained models, SMP continuously outperforms baseline methods due to the incorporation of diverse features and non-linear transformation, as well as its flexible framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Different architectures. SMP is a versatile method that can be applied to convolutional neural networks as well. In Suppl. B.3, we present the VTAB-1k results obtained from ImageNet pre-trained ResNet-50 models, where SMP outperforms other methods and achieves superior performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.45, + 0.892, + 0.464 + ], + "angle": 0, + "content": "Table 2. Average accuracy (%) on 9 full-size downstream datasets." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.468, + 0.895, + 0.511 + ], + "angle": 0, + "content": "
Fine-tuningVPTLinearSMP
Average Accuracy88.488.781.788.7
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.522, + 0.72, + 0.538 + ], + "angle": 0, + "content": "5.4. Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.592 + ], + "angle": 0, + "content": "In this section, we conduct various experiments to analyze the components of the proposed SMP method. Full results of this section are presented in Suppl. C." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.893, + 0.806 + ], + "angle": 0, + "content": "Visualization of group norms. We visualize the \\(\\ell_2\\) norm of each structure on different tasks in Figure 6, which helps us to understand the behavior of the proposed structured regularization. For Sun397, an easy adaptation task, we can observe the structures from deep layers have larger norms in Figure 6a. This validates our analysis in Section 3.2, that for easy adaptation tasks, features extracted from deep layers contain more information. For Clevr-Dist, a difficult adaptation task, the structures have similar norms overall, which means for difficult tasks, information contained in intermediate layers is also important, and diverse features can improve the performance, as we find in Section 3.2. This motivates the proposed structured non-linearity regularization, utilizing \\(\\ell_2\\) norm to control the complexity of non-linear model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Retraining probing model by gradually incorporating structures. We retrain a probing model by gradually incorporating structures selected based on the magnitude of their \\(\\ell_2\\) norm, to validate their contribution to the prediction. Results are presented in Figure 7. For Sun397, an easy adaptation task, we can see in Figure 7a that there are redun" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16855" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.085, + 0.478, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.204, + 0.327, + 0.217 + ], + "angle": 0, + "content": "(a) Sun397 (easy)" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.219, + 0.476, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.337, + 0.344, + 0.35 + ], + "angle": 0, + "content": "(b) Clevr-Dist (difficult)" + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.362, + 0.441, + 0.376 + ], + "angle": 0, + "content": "Figure 6. \\(\\ell_2\\) norm of structure groups on different tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.471, + 0.543 + ], + "angle": 0, + "content": "dant structures. And with a few selected structures, we can achieve comparable results compared with the original SMP, justifying the zero norm of some structured observed after performing structure regularization in Figure 6a. However, for Clevr-Dist, a difficult adaptation task, comparable performance is achieved until selecting 55 structures, suggesting intermediate structures also contribute to the final prediction. This demonstrates the necessity of proposed structured regularizer, which resolves the contradiction between easy and difficult domain adaptation tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.471, + 0.713 + ], + "angle": 0, + "content": "Impact of feature aggregation designs. The primary objective of performing token and channel aggregation is to minimize the redundancy in the extracted raw features, while maintaining feature diversity in the compact aggregated representation. To achieve this, two types of aggregation are designed to preserve channel and spatial information, respectively. We provide ablation studies in Suppl. C.5. Overall, channel information (aggregating along tokens) is essential for all datasets, while spatial information (aggregating along channels) is beneficial for tasks that are sensitive to spatial location, e.g., dSpr-Loc and sNORB-Azim." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.471, + 0.852 + ], + "angle": 0, + "content": "Impact of different structures. To increase the feature diversity, we incorporate features extracted from various structures, coupled with a structured sparse regularizer to conduct model fitting and feature selection simultaneously. This flexible framework enables us to avoid the costly manual selection of candidate structures. We provide analysis of structures in Suppl. C.6. The results show that incorporating all candidate structured features yields best mean results, which demonstrates the efficacy of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Cost of SMP. We compare the cost of SMP with baseline methods, including relative adaptation time, extra inference FLOPs, and extra parameters. The average results on VTAB-" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.086, + 0.695, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.221, + 0.648, + 0.234 + ], + "angle": 0, + "content": "(a) Sun397 (easy)" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.086, + 0.888, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.221, + 0.858, + 0.234 + ], + "angle": 0, + "content": "(b) Clevr-Dist (difficult)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.245, + 0.894, + 0.273 + ], + "angle": 0, + "content": "Figure 7. Retraining performance by gradually incorporating structures with larger \\(\\ell_{2}\\) norm obtained by SMP." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.283, + 0.894, + 0.311 + ], + "angle": 0, + "content": "Table 3. Cost comparison (Average results on VTAB-1k). * indicates relative results compared with Fine-tuning." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.317, + 0.89, + 0.394 + ], + "angle": 0, + "content": "
LinearFTVPTSMP
Adaptation time (rel.)*0.0510.970.15
Extra Inference FLOPs (G)0011.010.01
Extra parameters (M)0.0485.830.646.91
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.895, + 0.648 + ], + "angle": 0, + "content": "1k are shown in Table 3. SMP only needs \\(15\\%\\) adaptation time on average compared with fine-tuning. On the contrary, the adaptation time of VPT is \\(97\\%\\) of fine-tuning. This results in a significantly lower adaptation cost of SMP compared to tuning methods. SMP is also very efficient during inference time, and only introduces 0.01G inference FLOPs compared with fine-tuning. It can be noticed that VPT introduces extra 11.01G inference FLOPs, which is far more than SMP. SMP requires 6.91M extra parameters on average, which is the \\(8\\%\\) of fine-tuning. Though the requirement of extra parameters is higher than VPT, we can leverage the retraining method proposed in Section 5.4, which can further reduce the extra parameters. Overall, SMP is a highly efficient method with low adaptation cost, fast inference, and fewer extra parameters, while maintaining superior performance compared to tuning-based methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.664, + 0.619, + 0.68 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present structured model probing, an effective yet efficient probing method for transfer learning. Our investigation reveals that model probing behaves differently for easy and difficult adaptation tasks. To mitigate this conflicting behavior, we propose structured model probing, a method that is able to achieve good performance in both cases. With the proposed structured regularization, we can construct a simple adaptation model for easy adaptation cases, and increase the complexity of the adaptation model for difficult cases. Our method outperforms state-of-the-art tuning methods while maintaining computational efficiency. Our work highlights the potential of model probing, which motivates researchers to better exploit the use large-scale pre-trained model and apply model probing to diverse tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16856" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.472, + 0.184 + ], + "angle": 0, + "content": "[1] Alessandro Achille, Michael Lam, Rahul Tewari, Avinash Ravichandran, Subhransu Maji, Charless C. Fowlkes, Stefano Soatto, and Pietro Perona. Task2vec: Task embedding for meta-learning. In IEEE International Conference on Computer Vision (ICCV), pages 6429-6438, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Alessandro Achille, Aditya Golatkar, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. LQF: linear quadratic fine-tuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 15729-15739, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.472, + 0.31 + ], + "angle": 0, + "content": "[3] Armen Aghajanyan, Akshit Shrivastava, Anchit Gupta, Naman Goyal, Luke Zettlemoyer, and Sonal Gupta. Better finetuning by reducing representational collapse. In 9th International Conference on Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.472, + 0.368 + ], + "angle": 0, + "content": "[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In European Conference on Computer Vision (ECCV), pages 446-461, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.472, + 0.521 + ], + "angle": 0, + "content": "[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33 (NeurIPS), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.472, + 0.577 + ], + "angle": 0, + "content": "[6] Han Cai, Chuang Gan, Ligeng Zhu, and Song Han. Tinytl: Reduce memory, not parameters for efficient on-device learning. In Advances in Neural Information Processing Systems 33 (NeurIPS), 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.472, + 0.648 + ], + "angle": 0, + "content": "[7] Shoufa Chen, Chongjian GE, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In Advances in Neural Information Processing Systems 35 (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.472, + 0.717 + ], + "angle": 0, + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 1597-1607, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.719, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In IEEE International Conference on Computer Vision (ICCV), pages 9620-9629, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[10] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4109-4118, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[11] Mostafa Dehghani, Yi Tay, Anurag Arnab, Lucas Beyer, and Ashish Vaswani. The efficiency misnomer. In 10th International Conference on Learning Representations (ICLR), 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[13] Aditya Deshpande, Alessandro Achille, Avinash Ravichandran, Hao Li, Luca Zancato, Charless C. Fowlkes, Rahul Bhotika, Stefano Soatto, and Pietro Perona. A linearized framework and a new benchmark for model selection for fine-tuning. arXiv:2102.00084, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.894, + 0.3 + ], + "angle": 0, + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), pages 4171–4186, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.894, + 0.398 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations (ICLR), 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.894, + 0.469 + ], + "angle": 0, + "content": "[16] Utku Evci, Vincent Dumoulin, Hugo Larochelle, and Michael C. Mozer. Head2Toe: Utilizing intermediate representations for better transfer learning. In Proceedings of the 39th International Conference on Machine Learning (ICML), pages 6009-6033, 2022. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.894, + 0.525 + ], + "angle": 0, + "content": "[17] Weifeng Ge and Yizhou Yu. Borrowing treasures from the wealthy: Deep transfer learning through selective joint finetuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10-19, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.894, + 0.594 + ], + "angle": 0, + "content": "[18] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogério Schmidt Feris. SpotTune: Transfer learning through adaptive fine-tuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4805-4814, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.894, + 0.651 + ], + "angle": 0, + "content": "[19] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In 10th International Conference on Learning Representations (ICLR), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.894, + 0.718 + ], + "angle": 0, + "content": "[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.894, + 0.788 + ], + "angle": 0, + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning (ICML), pages 2790-2799, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge J. Belongie, Bharath Hariharan, and Ser-Nam Lim." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16857" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.191 + ], + "angle": 0, + "content": "[24] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition (CVPR Workshop), 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[25] Simon Kornblith, Jonathon Shlens, and Quoc V. Le. Do better imagenet models transfer better? In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2661-2671, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.471, + 0.305 + ], + "angle": 0, + "content": "[26] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE International Conference on Computer Vision Workshop (ICCV Workshop), pages 554–561, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.306, + 0.471, + 0.375 + ], + "angle": 0, + "content": "[27] Ananya Kumar, Aditi Raghunathan, Robbie Matthew Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In 10th International Conference on Learning Representations (ICLR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.471, + 0.432 + ], + "angle": 0, + "content": "[28] Ismael Lemhadri, Feng Ruan, Louis Abraham, and Robert Tibshirani. Lassonet: A neural network with feature sparsity. Journal of Machine Learning Research, 22:127:1-127:29, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.471, + 0.489 + ], + "angle": 0, + "content": "[29] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3045-3059, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.471, + 0.546 + ], + "angle": 0, + "content": "[30] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), pages 4582-4597, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.471, + 0.603 + ], + "angle": 0, + "content": "[31] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5196-5205, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.471, + 0.646 + ], + "angle": 0, + "content": "[32] Fangzhou Mu, Yingyu Liang, and Yin Li. Gradients as features for deep representation learning. In 8th International Conference on Learning Representations (ICLR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.471, + 0.716 + ], + "angle": 0, + "content": "[33] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. LEEP: A new measure to evaluate transferability of learned representations. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 7294-7305, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.471, + 0.759 + ], + "angle": 0, + "content": "[34] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345-1359, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.471, + 0.815 + ], + "angle": 0, + "content": "[35] Omkar M. Parkhi, Andreae Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3498-3505, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[36] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), pages 2227-2237, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.189 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.894, + 0.245 + ], + "angle": 0, + "content": "[38] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. LST: Ladder side-tuning for parameter and memory efficient transfer learning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.894, + 0.302 + ], + "angle": 0, + "content": "[39] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical report, California Institute of Technology, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.894, + 0.372 + ], + "angle": 0, + "content": "[40] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In 10th International Conference on Learning Representations (ICLR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.374, + 0.894, + 0.444 + ], + "angle": 0, + "content": "[41] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer G. Dy, and Tomas Pfister. Learning to prompt for continual learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 139-149, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.445, + 0.894, + 0.486 + ], + "angle": 0, + "content": "[42] Xiaohan Yan and Jacob Bien. Hierarchical Sparse Modeling: A choice of two group lasso formulations. Statistical Science, 32(4):531 - 560. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.488, + 0.894, + 0.542 + ], + "angle": 0, + "content": "[43] Ming Yuan and Yi Lin. Model selection and estimation in regression with grouped variables. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 68(1): 49-67, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.544, + 0.894, + 0.626 + ], + "angle": 0, + "content": "[44] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv:1910.04867, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.629, + 0.894, + 0.697 + ], + "angle": 0, + "content": "[45] Jeffrey O. Zhang, Alexander Sax, Amir Zamir, Leonidas J. Guibas, and Jitendra Malik. Side-tuning: A baseline for network adaptation via additive side networks. In European Conference on Computer Vision (ECCV), pages 698-714, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.699, + 0.894, + 0.755 + ], + "angle": 0, + "content": "[46] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 16795-16804, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.756, + 0.894, + 0.81 + ], + "angle": 0, + "content": "[47] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.812, + 0.894, + 0.868 + ], + "angle": 0, + "content": "[48] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16858" + } + ] +] \ No newline at end of file diff --git a/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/fcd62d92-907d-4f7f-bc80-b88db83abf7f_origin.pdf b/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/fcd62d92-907d-4f7f-bc80-b88db83abf7f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d8d78615a984fb87b1e130702fa5e8b9ab61ba34 --- /dev/null +++ b/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/fcd62d92-907d-4f7f-bc80-b88db83abf7f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1056be9bec23bb7ed3b5c99c890486b01ea67bbde403174f828d535d8654198 +size 351536 diff --git a/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/full.md b/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4d863cedc0ff7a606215846332df5fb5dce8b0da --- /dev/null +++ b/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/full.md @@ -0,0 +1,333 @@ +# Structured Model Probing: Empowering Efficient Transfer Learning by Structured Regularization + +Zhi-Fan Wu Chaojie Mao Xue Wang Jianwen Jiang Yiliang Lv Rong Jin Alibaba Group + +# Abstract + +Despite encouraging results from recent developments in transfer learning for adapting pre-trained model to downstream tasks, the performance of model probing is still lagging behind the state-of-the-art parameter efficient tuning methods. Our investigation reveals that existing model probing methods perform well for the easy case when the source domain (where models are pre-trained) and the adapted domain are similar, but fail for the difficult case when the two domains are significantly different. Simply incorporating features extracted from multiple layers and increasing complexity of the probing model can mitigate the gap in the difficult case, but degrades the performance in the easy case. To address this challenge, we propose structured model probing (SMP) that is able to deliver good performance for both cases through structured regularization. The regularization performs feature selection leveraging model structure as a prior, and controls the complexity of the probing model through the weights of selected structures. This enables us to construct a simple adaptation model, with a small number of selected features and a linear prediction model, for the easy case; and to automatically increase the complexity of adaptation model, with a large number of selected features and a non-linear model, for the difficult case. Our extensive empirical studies show that SMP significantly outperforms the state-of-the-art methods for parameter efficient tuning, and at the same time, still maintains the advantage of computational efficiency for probing-based methods. + +# 1. Introduction + +The parameters of models have grown drastically with the rapid development of deep learning [5, 37], posing challenges for adapting pre-trained models to downstream tasks. The widely used transfer strategy, fully fine-tuning, becomes infeasible due to excessive computation and storage overhead. Recently, parameter efficient tuning [19, 23] has gained attention for reducing storage costs and improving tuning performance with few trainable parameters when lim + +![](images/63f688131516d3ed72daf25e836154d3ffd2d13587e9edb41098c34830f2f3e3.jpg) +Figure 1. Performance comparison on different dataset groups (left) and mean relative adaptation time (right) on VTAB-1k benchmark. + +ited data are available for downstream tasks. Despite the progress, parameter efficient tuning methods still face significant computational overhead in training due to the need to perform forward and backward passes through the pretrained model at each training step. + +Model probing, which leverages features extracted from the pre-trained model to perform adaptation, is promising due to its efficiency and simplicity: The training procedure is efficient since the pre-trained model is fixed and only performs one forward propagation to extract features, and probing model is decoupled from pre-trained model that makes it convenient to deploy small sized task-specific downstream models. However, it is a nuisance that probing methods are usually outperformed by state-of-the-art parameter efficient tuning methods. This is true even after several recent improvements made for the probing methods [16, 45]. In this paper, we aim to improve the performance of model probing significantly without sacrificing its simplicity and efficiency. + +We first investigate when model probing methods perform worse than the tuning-based methods on widely used visual task adaptation benchmark VTAB-1k [44] as shown in Figure 1. Target domains are categorized as easy and difficult cases based on their similarity to the source domain where the pre-trained model is trained on (see Section 3.1 for details). We find simple linear probing yields comparable performance as the two tuning-based approaches (finetuning and VPT [23]) for easy cases, whereas a significant + +![](images/8976b25f7d62a3b031921846e07f4362b5639759867d578f4903a604d17738bc.jpg) +Figure 2. Adaptation pipelines of Model Tuning (left) and the proposed Structured Model Probing (right). + +![](images/7160d55c197b3be1ab48dc688ed30909c72b2fb4a9022aedc0f7297b00ee4323.jpg) + +performance gap is observed between linear probing and tuning-based methods for difficult cases. This suggests that model probing exhibits different behavior for various target domains, and performance gap arises for target domains that dissimilar to the source domain. + +This is not surprising that we expect a more complicated model than a linear probing to compensate the difference between the source domain and the target domain. In general, we can increase model complexity in two ways, either by introducing more features extracted from the pre-trained model, or by introducing a non-linear model for prediction. A main challenge is how to develop a unified framework of model probing that is able to handle both the easy and difficult adaptations effectively. This is because, according to our study in Section 3, for easy cases, a complicated probing model (with more features and non-linear probing) performs significantly worse than a simple linear probing. + +To this end, we propose structured model probing, or SMP for short, that effectively address both easy and difficult cases of domain adaptation. The key to our unified framework is structured regularization. It performs feature selection leveraging the structure of the pre-trained model as a prior, and controls the complexity of the probing model through the weights assigned to selected features. Thus, for the easy case, this regularization constructs a simple linear model with a few selected features; and for the difficult case, it automatically introduces more features from the pre-trained model and more non-linearity in the prediction module based on the value of loss function, as shown in Figure 2. Our empirical studies show that the proposed structured model probing performs significantly better than state-of-the-art tuning-based approaches, while maintaining the advantage of computational efficiency. + +Our main contributions are as follows: + +1. We observe and demonstrate the conflicting behaviors of model probing: a complicated probing model with more features and non-linear transformation can improve the performance for difficult cases of domain adaptation, but it yields worse performance than linear probing for easy cases of domain adaptation. + +2. We propose structured model probing (SMP) based on structured regularization. This regularization enables us to construct simple probing model for easy tasks, and automatically increase model complexity for difficult tasks, to simultaneously address both the easy and difficult cases of domain adaptation. +3. We conduct extensive empirical studies on various adaptation tasks. The proposed SMP method shows superior experimental results compared to state-of-the-art probing and tuning methods, and requires significantly lower training costs compared to tuning methods. + +# 2. Related Work + +Transfer learning has been extensively studied in vision domain [34, 48]. It aims to improve the performance on target domain by leveraging the related information contained in source domain. Transfer learning based on pre-trained model [8, 20, 21] has received significant attention due to its simplicity and good empirical performance [23, 45]. + +Model tuning adapts the pre-trained model to new tasks via updating model parameters. Fully fine-tuning is a widely used strategies due to its efficacy and simplicity [9, 25, 44]. However, it suffers from huge storage and computation overhead due to the ever-expanding size of large-scale pre-trained model [37, 40]. Heuristic approaches have been proposed to improve fine-tuning performance [3, 6, 17, 18, 27]. Parameter efficient tuning [19] aims to reduce the storage requirement of transferred model by updating a few parameters attached to the pre-trained model [23, 47], like Prompt Tuning [29, 30] or Adapter [22]. While originally studied in natural language processing tasks due to the dominance of large-scale pre-trained language models [5, 14], there is growing interest in applying these methods to computer vision tasks [7, 23, 31, 41, 46, 47]. Though parameter efficient tuning reduces storage cost significantly, the training cost remains high due to the forward and backward propagation need to pass through the entire pre-trained model [11]. + +Model probing utilizes frozen extracted features for adapting the pre-trained model to new tasks, thus achieving effi + +cient adaptation. The most popular method, linear probing, suffers from performance degradation [25], which limits its application in transfer learning [20, 21]. Recent methods [16, 36, 45] have been proposed to improve probing performance by leveraging side network [38, 45] or incorporating intermediate representations [2, 16, 32], but there is still a significant performance gap compared to tuning methods. This motivates us to investigate the limitation of probing methods, and come up with a better solution. + +# 3. Understanding the Limitation of Model Probing for Transfer Learning + +In this section, we analyze the performance of probing methods across various downstream domains to understand their limitations, motivated by the observation in Figure 1. We assess feature informativeness for different tasks and investigate the necessity of non-linear transformation. + +# 3.1. Analysis Setup + +Pre-trained model and source data. The ViT-B/16 [15] is used as the backbone because of the extensive usage of transformer-based models. The model is pre-trained on ImageNet-21k [12] in a supervised learning manner. + +Target data. We use the VTAB-1k benchmark, consisting of 19 datasets that cover natural images (natural), images captured by specialist equipment (specialized), and images generated from simulated environments (structured), as our target data for transfer learning. We follow the setting in [44] and use 1000 samples as the training set per task to evaluate adaptation with limited data as suggested in [44]. + +Domain similarity. Several previous works have proposed metrics to measure the similarity between source domain and target domain [1, 10, 13, 16, 33]. Here, we adopt the domain similarity measure defined in [16]: + +$$ +D o m a i n \quad \text {S i m i l a r i t y} = \mathrm {A c c} _ {\text {L i n e a r}} - \mathrm {A c c} _ {\text {S c r a t c h}}. \tag {1} +$$ + +A higher domain similarity suggests that the features extracted from a pre-trained model can be utilized for the target domain, leading to performance gains via a linear classifier compared to training from scratch. Conversely, a lower domain similarity indicates that training from scratch on the target domain may be more effective than using extracted features. Thus, the domain similarity is evaluated based on the benefit gained by a linear classifier compared to training from scratch. We also report the label-feature correlation [13] in Suppl. C.8, which exhibits a high Spearman rank correlation (0.854) with the domain similarity scores. + +Domain similarity also reflects how difficult to adapt from source domain to target domain. As shown in Figure 4, tasks in natural and specialized groups have higher domain similarity to source domain where the pre-trained model is + +![](images/e6ce50144cb8482b59edc659595afb130625e9a8932639916a34f16206715741.jpg) +Figure 3. Performance of features extracted from different layers on Sun397 (easy) and Clevr-Dist (difficult). + +![](images/a2274185ec5d76b15ee835772fee332d9d4e0b99a6ae64cf014e11b7db3753ed.jpg) + +trained on, so we mark them as easy cases. And tasks in structured groups have lower domain similarity to source domain, thus they are labeled as difficult cases in Figure 1. + +# 3.2. Informativeness of Extracted Features + +To quantify the informativeness of features extracted from different layers for downstream tasks, we use leave-one-out $k$ -NN testing as the surrogate performance criterion: + +$$ +\text {I n f o r m a t i v e n e s s} = \frac {\sum_ {i = 1} ^ {m} \sum_ {x _ {j} \in N (x _ {i})} \mathbb {1} \left\{y _ {i} = y _ {j} \right\}}{k m}, \tag {2} +$$ + +where $m$ is the number of samples, $x$ is the extracted features of a sample, $N(x_{i})$ represents $k$ nearest neighbors of $x_{i}$ . We simply set $k = 1$ , which results in leave-one-out 1-NN testing. We choose $k$ -NN, instead of linear classifier, for informativeness measurement because $k$ -NN is a non-parametric classifier and thus avoids training additional parameters, making the measurement more robust. + +In Figure 3, we plot the informativeness of features extracted from different layers on two tasks from different domains. For Sun397, an easy domain adaption task, we observe the best performance achieved by features extracted from deep layers. This explains the good performance of linear probing for easy domain adaption tasks, as observed in Figure 1. While for Clevr-Dist, a difficult task with target domain significantly different from the source domain, we observe that no layer has outstanding performance. The results echo the findings in [16]. + +Furthermore, we conduct experiments on gradually incorporating features extracted from different layers of the pre-trained model, demonstrating the different behaviors on easy and difficult tasks, which can be found in Section 5.4. The results show for easy tasks, more features may lead to overfitting, but for difficult tasks, performance improves with more features. Therefore, when the target domain is very different from the source domain, we need to provide probing model with diverse features extracted from multiple layers of the pre-trained model. + +![](images/3939f6c2b6b291cfdf9500bca9d202299da8bc6a024f8e3cf50e6af13a5703c3.jpg) +Figure 4. Performance of extracted features with linear/ non-linear transformation on VTAB-1k. Datasets are ordered from easy tasks (left) to difficult tasks (right) according to domain similarity. + +# 3.3. Effects of Non-linear Transformation + +Since there is no significantly informative layer for difficult domain adaption tasks, a linear combination may not be enough for exploiting the information in extracted features. We further examine the efficacy of non-linear combination on extracted features, as neural networks achieve superior performance through non-linear feature transformation. + +We employ an multiple-layer perceptron (MLP) for nonlinear transformation and compare its performance with linear transformation. Features extracted from different structures of the pre-trained model are concatenated together as input (See Section 4.1 for details). + +In Figure 4, we clearly observe that adding non-linearity to feature transformation can significantly boost the performance for the difficult domain adaption tasks (i.e. the datasets on right end of Figure 4). This phenomenon indicates that simply performing linear transformation on extracted features can not exploit the full potential of representations extracted from the pre-trained model. On the other hand, we notice that the performance on easy tasks is not improved, and is even hurt with the introduction of non-linearity. We speculate that this is because the linear transformation can already make full use of extracted information, while non-linearity increases model capacity, which eventually leads to overfitting to the training data. + +Discussion. Our investigation reveals that the optimal probing model differs depending on the domain of downstream task. Incorporating diverse features from multiple layers of the pre-trained model and performing non-linear transformation on them can significantly improve the performance on difficult cases. However, this yields worse results on easy tasks. On the contrary, simply performing linear combination on features from deep layers can achieve satisfactory performance for easy adaptation tasks, but not for difficult ones. The conflicting behaviors of model probing motivates us to develop a unified framework that can solve the adaptation problems for both easy and difficult adaptation tasks. + +# 4. Structured Model Probing + +In this section, we propose structured model probing, a unified probing method which can deal with both easy and difficult cases of adaptation tasks. We first present how to thoroughly extract features in a structured way, then introduce our probing model with structured regularization. + +# 4.1. Structured Feature Extraction + +Incorporating diverse features extracted from the pre-trained model can improve the performance on difficult adaptation tasks for model probing, as shown in Section 3.2. To provide diverse features for the probing model, we extract features from all structures in a Vision Transformer (ViT). These structures include features before self-attention, features after self-attention, and features after Feed Forward Network (FFN), for each transformer layer. We also include tokenized image inputs and pre-logits. We use $V = \{v_{i}\}_{i}^{|V|}$ represents all structures in ViT, and $|V|$ is the total number of structures. For an input sample, we denote each set of tokens extracted from structure $v_{i}$ as $X_{i} \in \mathbb{R}^{t \times c}$ , where $t$ is the dimension of tokens and $c$ is the dimension of channels. + +To reduce feature redundancy while preserving diversity, we apply 1D average pooling along different dimensions on tokens $X_{i}$ , resulting in two types of aggregation. To maintain channel information, we aggregate $X_{i}$ along the token dimension, which produces $x_{i}^{t}$ . And to preserve spatial information, we aggregate $X_{i}$ along the channel dimension, resulting in $x_{i}^{c}$ . Finally, we concatenate all structured features into a single vector, $x = [x_{1}^{t}, x_{1}^{c}, \dots, x_{|V|}^{t}, x_{|V|}^{c}]$ . We denote $s_{i}$ as the index set of elements in $x$ , and $S = \{s_{i} | i = 1\dots |s|\}$ . + +# 4.2. Probing with Structured Regularization + +Our structured regularization contains two components: structured sparsity regularizer and structured non-linearity regularizer. We introduce them as follows. + +# 4.2.1 Structured Feature Selection + +As shown in Section 3.2, the informativeness of extracted features varies with tasks, and redundant features lead to performance degradation on some tasks. Thus, we propose to perform feature selection during the training stage of the probing model to prevent overfitting. A popular way is to use group lasso [16, 43] by constructing groups on feature dimension over the weight matrix of a linear classifier: + +$$ +\min _ {\theta} L (\theta) + \lambda \sum_ {i = 1} ^ {d} | | \theta_ {i} | | _ {2}, \text {w h e r e} | | \theta_ {i} | | _ {2} = \sqrt {\sum_ {j = 1} ^ {n} \theta_ {i j} ^ {2}}. \tag {3} +$$ + +We use $\theta \in \mathbb{R}^{d\times n}$ to represent weight matrix, where $d$ is the feature dimension and $n$ is the number of classes. However, due to the high dimensionality of features, the final achieved sparsity is hardly the optimal result. It is worth noting that the structure of pre-trained model provides a natural way to divide features as groups, and features extracted from the same structure have higher relevance to each other. Thus, we leverage the structure of the pre-trained model, to achieve the structured sparsity with a structured sparsity regularizer: + +$$ +\min _ {\theta} L (\theta) + \lambda \sum_ {s \in S} \frac {1}{| s |} \| \theta_ {s} \| _ {2}, \text {w h e r e} \| \theta_ {s} \| _ {2} = \sqrt {\sum_ {i \in s} \sum_ {j = 1} ^ {n} \theta_ {i j} ^ {2}}, \tag {4} +$$ + +$s \in S$ is the index set of a specific structure in feature as described in Section 4.1, and $|s|$ represents its dimension. Compared with constructing groups by features, in our regularizer, we construct groups by leveraging the structure of the pre-trained model. This formalization can significantly reduce the combinations of selection and thus prevent overfitting. With such an advantage, our method achieves superior performance with a small number of samples under high-dimensional features. + +# 4.2.2 Structured Non-linear Transformation + +Our discovery in Section 3.3 shows non-linear transformation can facilitate transfer performance for difficult tasks, while could reduce performance for easy adaptation tasks, mostly due to over-fitting. This means that we need a model that allows flexibly incorporating non-linearity to a linear model. We propose a structured non-linearity regularizer that nicely integrates feature selection (i.e. choosing the subset of informative features from the pre-trained model) with model selection (i.e. choosing either a linear model or nonlinear model for prediction) inspired by hierarchical sparse modeling [28, 42]. Specifically, we construct our model by the combination of a linear model and a neural network: + +$$ +f (x) = \theta^ {\top} x + f _ {W} (x), \tag {5} +$$ + +where $\theta \in \mathbb{R}^{d\times n}$ is a linear model, and $f_{W}$ represents an MLP with model weights $W$ . We use $W^{(1)}$ to represent the input layer of the MLP, and the parameters of rest layers are denoted as $W^{(2:)}$ . + +We utilize the feature selection result $||\theta ||_2$ , the $\ell_2$ norm of weights assigned to the selected features, to regularize the complexity of the non-linear model: + +- When a small number of features are good enough for prediction, we expect a small value for $||\theta||_2$ . By using $||\theta||_2$ to control the complexity of non-linear model, we expect the resulting model to be mostly linear. +- With increasing $||\theta||_2$ , more diverse structures from the pre-trained model are needed for prediction. By using $||\theta||_2$ to regularize the size of non-linear model, we expect the resulting model to be more non-linear. + +A way to achieve such control is to add a constraint to the objective function in Eq.(4) as follows: + +$$ +\min _ {\theta , W} L (\theta , W) + \lambda \sum_ {s \in S} | | \theta_ {s} | | _ {2} +$$ + +$$ +\begin{array}{l} \text {s u b j e c t} \left. | | W _ {s} ^ {(1)} | | _ {2} \leq M _ {1} | | \theta_ {s} | | _ {2}, s = 1 \dots | S |, \right. \tag {6} \\ \left| \left| W ^ {(2:)} \right| \right| _ {2} \leq M _ {2} \left| \left| \theta \right| \right| _ {2}. \\ \end{array} +$$ + +There are two constraints in Eq.(6): the first constraint is used to maintain the input feature sparsity of the neural network, and the second term is used to control the non-linearity through the linear model. To achieve efficient implementation, we convert the constraint in Eq.(6) into a regularization term, which we call structured non-linearity regularizer: + +$$ +\begin{array}{l} \Omega (\theta , W) = \sum_ {s \in S} \max \left\{\left| \left| W _ {s} ^ {(1)} \right| \right| _ {2} - M _ {1} \left| \left| \theta_ {s} \right| \right| _ {2}, 0 \right\} \tag {7} \\ + \max \left\{\left| \left| W ^ {(2:)} \right| \right| _ {2} - M _ {2} \left| \left| \theta \right| \right| _ {2}, 0 \right\}. \\ \end{array} +$$ + +This regularization term means that, when $||W||_2 > M||\theta||_2$ , $W$ receives penalty until it achieves similar complexity as $M||\theta||$ . Then our final objective function is: + +$$ +\min _ {\theta , W} L (\theta , W) + \lambda_ {1} \sum_ {s \in S} | | \theta_ {s} | | _ {2} + \lambda_ {2} \Omega (\theta , W). \tag {8} +$$ + +We use $\lambda$ to control the penalty strength, and simply set $\lambda_{2} = 0.1$ works well in our experiments. + +Loss function. Since our model $f(x)$ consists of both linear and non-linear parts, we decouple these two components in loss function. This modification enables the structured sparsity regularizer to control the complexity of $\| \theta \|$ without the interference from the non-linear part: + +$$ +L (\theta , W) = \operatorname {C E} \left(\theta^ {\top} x, y\right) + \operatorname {C E} \left(\operatorname {s g} \left(\theta^ {\top} x\right) + f _ {W} (x), y\right), \tag {9} +$$ + +where $y$ is the one-hot label vector of $x$ , CE is cross-entropy loss and sg means stop-gradient. + +Table 1. Median test accuracy (%) over 3 seeds on the VTAB-1k benchmark using ViT-B/16 pretrained on ImageNet-21k. * indicates results are obtained from [23] and † indicates results are obtained from [16]. + +
NaturalSpecializedStructured
CTFAR-100Caltech101DTDFlowers102PetsSVHNSun397Mean (Natural)CamelyonEuroSATResisc45RetinopathyMean (Specialized)Clevr-CountClevr-DistDMLabKITTI-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-ElevMean (Structured)Mean (All)
Tuning Methods
Scratch†7.619.113.129.66.719.42.314.071.071.029.372.060.831.652.527.239.166.129.711.724.135.332.8
Fine-tuning*68.987.764.397.286.987.438.875.979.795.784.273.983.756.358.641.765.557.546.725.729.147.665.6
Bias*72.887.059.297.585.359.951.473.378.791.672.969.878.361.555.632.455.966.640.015.725.144.162.0
VPT*78.890.865.898.088.378.149.678.581.896.183.468.482.468.560.046.572.873.647.932.937.855.069.4
Probing Methods
Linear77.286.866.699.188.835.755.172.778.489.672.273.778.531.029.935.254.911.727.714.423.728.655.4
Side-Tuning*60.760.853.695.566.734.935.358.258.587.765.261.068.127.622.631.351.78.214.49.821.823.445.6
Head2Toe75.390.975.099.586.183.550.980.284.495.784.474.384.751.759.444.065.647.140.332.541.147.767.5
LinearStruc.72.989.572.598.983.671.642.776.084.396.081.875.084.348.252.539.466.748.736.732.337.745.364.8
LinearStruc.w/FSR78.290.173.499.388.171.745.478.083.996.181.974.184.048.552.839.664.451.438.632.937.045.765.7
LinearStruc.w/SSR79.090.274.199.190.473.454.280.184.696.383.874.484.848.852.939.866.948.738.932.537.945.866.6
MLPStruc.66.888.770.397.482.171.133.372.881.695.977.073.582.077.257.739.366.572.344.132.747.754.767.1
SMP w/o SNR71.389.672.399.085.672.745.076.584.295.981.972.883.777.457.140.867.172.444.232.648.455.069.0
SMP79.390.974.999.390.475.055.380.784.896.383.175.084.877.558.040.867.572.544.533.049.055.470.9
+ +# 5. Experiments + +In this section, we evaluate the proposed structured model probing method on multiple datasets across various tasks, and provide detailed analysis of the proposed method. + +Implementation details. We evaluate SMP on the popular vision backbone, ViT-B/16 [15] pre-trained on ImageNet-21k [12], which is widely used across relative research works [16, 23]. We provide details of probing model and hyperparameter selection in the supplementary material. + +Compared methods. We compare SMP with other commonly used and state-of-the-art tuning (Fine-tuning, Bias [6] and VPT [23]) and probing methods (Linear Probing, Side-Tuning [45] and Head2Toe [16]). All hyperparameters of baseline methods are carefully tuned, and further information is provided in supplementary material. + +# 5.1. Experiments on VTAB-1k + +Experimental results on the VTAB-1k benchmark (Table 1) indicate that SMP outperforms other methods, achieving a mean performance of $70.9\%$ , $3.4\%$ higher than Head2Toe and $1.5\%$ higher than VPT. SMP also significantly outperforms Linear Probing, Bias, and Side-Tuning. SMP achieves optimal results in all three dataset groups. Especially in Structured group, SMP attains a mean accuracy of $55.4\%$ , surpassing other probing methods by a significant margin. This demonstrates the incorporation of non-linear transformation is important for difficult adaptation tasks. + +Ablation studies. We provide results of baseline methods that contribute to SMP, to validate the efficacy of structured + +regularization and the contribution of each component. + +LinearStruc. represents extracting features by structured feature extraction proposed in Section 4.1, then training a linear classifier. Compare to Linear Probing, it gets $9\%$ average performance gain, indicating that solely extracting features from the last layer is inadequate for model probing. + +We further compare feature sparsity regularization in Eq.(3) (represents by $\mathrm{Linear}_{\mathrm{Struc.}} + \mathrm{FSR}$ ) to the proposed structured sparsity regularizer (represents by $\mathrm{Linear}_{\mathrm{Struc.}} + \mathrm{SSR}$ ). Structured sparsity regularizer outperforms feature sparsity regularization, indicating that feature selection can mitigate the overfitting caused by high dimensions, and leveraging model structure as a prior can bring extra benefits. + +We apply non-linear transformation on extracted features, represents by $\mathrm{MLP}_{\text{Struc.}}$ . We find the performance on structured group gets a significant improvement over $\mathrm{Linear}_{\text{Struc.}}$ , demonstrating the importance of non-linear transformation on difficult adaptation tasks. However, the performance on easy tasks, i.e., natural and specialized groups, is degraded, suggesting the risk of overfitting from non-linear transformation. We solve this problem by leveraging structured nonlinearity regularizer (SNR), which can automatically control the non-linearity through the norms of selected structures. + +# 5.2. Experiments on Few-shot Learning + +In this section, we evaluate our method in few-shot learning scenario, where the number of training samples varies from 1 to 16 shots per class following existing studies. Five fine-grained recognition tasks are chosen, i.e., Food101 [4], CUB-200-2011 [39], Stanford Cars [26], Stanford Dogs [24] + +![](images/817f5dd48fe878aaf1b3177c862d35bfe2ed6a9e9e8e3980f397eaecfecb0609.jpg) + +![](images/e2fe9524fb302c804246dd5f00e0babe0a03ccc14a1e9abb24d528eab516d278.jpg) + +![](images/484d8aefe8aaa968b7c51fd933bef7233ff1cddb14dede596e27d1856b3004b2.jpg) + +![](images/1dcfc2b477ee64adeaf90836b3949d8250e549672b924d614cbda5c630efdc49.jpg) +Figure 5. Performance of few-shot learning on fine-grained visual recognition tasks. LinearStr. (F) indicates LinearStruc. w/ FSR. + +![](images/47064e96137db55553faad763127478d043cbcae2ca099511d7e6dc3d2b89899.jpg) + +![](images/e7681287e02199c4d9a68a98c6b192a4e43bb4f3e93194a7da447464654dafba.jpg) + +and Oxford Pets [35]. We compared SMP with two strong baselines, LinearStruc. and LinearStruc. w/ FSR, and two prominent methods, VPT (tuning) and Head2Toe (probing). + +Experimental results presented in Figure 5 show that SMP consistently outperforms other methods in most cases, demonstrating its effectiveness in few-shot scenario, and its ability to improve performance with increasing training data. Notably, feature sparsity regularization performs worse than SMP, highlighting the importance of leveraging pre-trained model structure as prior knowledge in low-shot scenario. + +# 5.3. Experiments on More Transfer Scenarios + +Larger downstream datasets. We provide average results on 9 full-size downstream datasets in Table 2. It can be observed even in the larger data regime, SMP still achieves competitive performance compared with tuning methods. Full results are presented in Suppl. B.1. + +Different pre-trained models. We present VTAB-1k results on ViT-B/16 and ViT-L/14 pretrained by CLIP in Suppl. B.2. Despite the usage of stronger pre-trained models, SMP continuously outperforms baseline methods due to the incorporation of diverse features and non-linear transformation, as well as its flexible framework. + +Different architectures. SMP is a versatile method that can be applied to convolutional neural networks as well. In Suppl. B.3, we present the VTAB-1k results obtained from ImageNet pre-trained ResNet-50 models, where SMP outperforms other methods and achieves superior performance. + +Table 2. Average accuracy (%) on 9 full-size downstream datasets. + +
Fine-tuningVPTLinearSMP
Average Accuracy88.488.781.788.7
+ +# 5.4. Analysis and Discussion + +In this section, we conduct various experiments to analyze the components of the proposed SMP method. Full results of this section are presented in Suppl. C. + +Visualization of group norms. We visualize the $\ell_2$ norm of each structure on different tasks in Figure 6, which helps us to understand the behavior of the proposed structured regularization. For Sun397, an easy adaptation task, we can observe the structures from deep layers have larger norms in Figure 6a. This validates our analysis in Section 3.2, that for easy adaptation tasks, features extracted from deep layers contain more information. For Clevr-Dist, a difficult adaptation task, the structures have similar norms overall, which means for difficult tasks, information contained in intermediate layers is also important, and diverse features can improve the performance, as we find in Section 3.2. This motivates the proposed structured non-linearity regularization, utilizing $\ell_2$ norm to control the complexity of non-linear model. + +Retraining probing model by gradually incorporating structures. We retrain a probing model by gradually incorporating structures selected based on the magnitude of their $\ell_2$ norm, to validate their contribution to the prediction. Results are presented in Figure 7. For Sun397, an easy adaptation task, we can see in Figure 7a that there are redun + +![](images/e99f4b82417c31d3c5087dcde2007334e8fda3dd38a873613a6132ae66d91ab8.jpg) +(a) Sun397 (easy) + +![](images/1fb3d64faad6bf3cad0400800a95c05e9b0fe2efb2efe6b342be0c3a1ad2b55b.jpg) +(b) Clevr-Dist (difficult) +Figure 6. $\ell_2$ norm of structure groups on different tasks. +Figure 7. Retraining performance by gradually incorporating structures with larger $\ell_{2}$ norm obtained by SMP. + +dant structures. And with a few selected structures, we can achieve comparable results compared with the original SMP, justifying the zero norm of some structured observed after performing structure regularization in Figure 6a. However, for Clevr-Dist, a difficult adaptation task, comparable performance is achieved until selecting 55 structures, suggesting intermediate structures also contribute to the final prediction. This demonstrates the necessity of proposed structured regularizer, which resolves the contradiction between easy and difficult domain adaptation tasks. + +Impact of feature aggregation designs. The primary objective of performing token and channel aggregation is to minimize the redundancy in the extracted raw features, while maintaining feature diversity in the compact aggregated representation. To achieve this, two types of aggregation are designed to preserve channel and spatial information, respectively. We provide ablation studies in Suppl. C.5. Overall, channel information (aggregating along tokens) is essential for all datasets, while spatial information (aggregating along channels) is beneficial for tasks that are sensitive to spatial location, e.g., dSpr-Loc and sNORB-Azim. + +Impact of different structures. To increase the feature diversity, we incorporate features extracted from various structures, coupled with a structured sparse regularizer to conduct model fitting and feature selection simultaneously. This flexible framework enables us to avoid the costly manual selection of candidate structures. We provide analysis of structures in Suppl. C.6. The results show that incorporating all candidate structured features yields best mean results, which demonstrates the efficacy of our method. + +Cost of SMP. We compare the cost of SMP with baseline methods, including relative adaptation time, extra inference FLOPs, and extra parameters. The average results on VTAB- + +![](images/c4addc513272d803bc245c903c01bcb20ac3f895ce7323006e5714743781c417.jpg) +(a) Sun397 (easy) + +![](images/67152e2b9413ff984d8288c4e3f256e6f520b098e0b19564a1c7c9735b629693.jpg) +(b) Clevr-Dist (difficult) + +Table 3. Cost comparison (Average results on VTAB-1k). * indicates relative results compared with Fine-tuning. + +
LinearFTVPTSMP
Adaptation time (rel.)*0.0510.970.15
Extra Inference FLOPs (G)0011.010.01
Extra parameters (M)0.0485.830.646.91
+ +1k are shown in Table 3. SMP only needs $15\%$ adaptation time on average compared with fine-tuning. On the contrary, the adaptation time of VPT is $97\%$ of fine-tuning. This results in a significantly lower adaptation cost of SMP compared to tuning methods. SMP is also very efficient during inference time, and only introduces 0.01G inference FLOPs compared with fine-tuning. It can be noticed that VPT introduces extra 11.01G inference FLOPs, which is far more than SMP. SMP requires 6.91M extra parameters on average, which is the $8\%$ of fine-tuning. Though the requirement of extra parameters is higher than VPT, we can leverage the retraining method proposed in Section 5.4, which can further reduce the extra parameters. Overall, SMP is a highly efficient method with low adaptation cost, fast inference, and fewer extra parameters, while maintaining superior performance compared to tuning-based methods. + +# 6. Conclusion + +In this paper, we present structured model probing, an effective yet efficient probing method for transfer learning. Our investigation reveals that model probing behaves differently for easy and difficult adaptation tasks. To mitigate this conflicting behavior, we propose structured model probing, a method that is able to achieve good performance in both cases. With the proposed structured regularization, we can construct a simple adaptation model for easy adaptation cases, and increase the complexity of the adaptation model for difficult cases. Our method outperforms state-of-the-art tuning methods while maintaining computational efficiency. Our work highlights the potential of model probing, which motivates researchers to better exploit the use large-scale pre-trained model and apply model probing to diverse tasks. + +# References + +[1] Alessandro Achille, Michael Lam, Rahul Tewari, Avinash Ravichandran, Subhransu Maji, Charless C. Fowlkes, Stefano Soatto, and Pietro Perona. Task2vec: Task embedding for meta-learning. In IEEE International Conference on Computer Vision (ICCV), pages 6429-6438, 2019. 3 +[2] Alessandro Achille, Aditya Golatkar, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. LQF: linear quadratic fine-tuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 15729-15739, 2021. 3 +[3] Armen Aghajanyan, Akshit Shrivastava, Anchit Gupta, Naman Goyal, Luke Zettlemoyer, and Sonal Gupta. Better finetuning by reducing representational collapse. In 9th International Conference on Learning Representations (ICLR), 2021. 2 +[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In European Conference on Computer Vision (ECCV), pages 446-461, 2014. 6 +[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33 (NeurIPS), 2020. 1, 2 +[6] Han Cai, Chuang Gan, Ligeng Zhu, and Song Han. Tinytl: Reduce memory, not parameters for efficient on-device learning. In Advances in Neural Information Processing Systems 33 (NeurIPS), 2020. 2, 6 +[7] Shoufa Chen, Chongjian GE, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In Advances in Neural Information Processing Systems 35 (NeurIPS), 2022. 2 +[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 1597-1607, 2020. 2 +[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In IEEE International Conference on Computer Vision (ICCV), pages 9620-9629, 2021. 2 +[10] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4109-4118, 2018. 3 +[11] Mostafa Dehghani, Yi Tay, Anurag Arnab, Lucas Beyer, and Ashish Vaswani. The efficiency misnomer. In 10th International Conference on Learning Representations (ICLR), 2022. 2 + +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 3, 6 +[13] Aditya Deshpande, Alessandro Achille, Avinash Ravichandran, Hao Li, Luca Zancato, Charless C. Fowlkes, Rahul Bhotika, Stefano Soatto, and Pietro Perona. A linearized framework and a new benchmark for model selection for fine-tuning. arXiv:2102.00084, 2021. 3 +[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), pages 4171–4186, 2019. 2 +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations (ICLR), 2021. 3, 6 +[16] Utku Evci, Vincent Dumoulin, Hugo Larochelle, and Michael C. Mozer. Head2Toe: Utilizing intermediate representations for better transfer learning. In Proceedings of the 39th International Conference on Machine Learning (ICML), pages 6009-6033, 2022. 1, 3, 5, 6 +[17] Weifeng Ge and Yizhou Yu. Borrowing treasures from the wealthy: Deep transfer learning through selective joint finetuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10-19, 2017. 2 +[18] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogério Schmidt Feris. SpotTune: Transfer learning through adaptive fine-tuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4805-4814, 2019. 2 +[19] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In 10th International Conference on Learning Representations (ICLR), 2022. 1, 2 +[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 2, 3 +[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, 2022. 2, 3 +[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning (ICML), pages 2790-2799, 2019. 2 +[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge J. Belongie, Bharath Hariharan, and Ser-Nam Lim. + +Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6 +[24] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition (CVPR Workshop), 2011. 6 +[25] Simon Kornblith, Jonathon Shlens, and Quoc V. Le. Do better imagenet models transfer better? In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2661-2671, 2019. 2, 3 +[26] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE International Conference on Computer Vision Workshop (ICCV Workshop), pages 554–561, 2013. 6 +[27] Ananya Kumar, Aditi Raghunathan, Robbie Matthew Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In 10th International Conference on Learning Representations (ICLR), 2022. 2 +[28] Ismael Lemhadri, Feng Ruan, Louis Abraham, and Robert Tibshirani. Lassonet: A neural network with feature sparsity. Journal of Machine Learning Research, 22:127:1-127:29, 2021. 5 +[29] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3045-3059, 2021. 2 +[30] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), pages 4582-4597, 2021. 2 +[31] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5196-5205, 2022. 2 +[32] Fangzhou Mu, Yingyu Liang, and Yin Li. Gradients as features for deep representation learning. In 8th International Conference on Learning Representations (ICLR), 2020. 3 +[33] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. LEEP: A new measure to evaluate transferability of learned representations. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 7294-7305, 2020. 3 +[34] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345-1359, 2010. 2 +[35] Omkar M. Parkhi, Andreae Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3498-3505, 2012. 7 +[36] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), pages 2227-2237, 2018. 3 + +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 1, 2 +[38] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. LST: Ladder side-tuning for parameter and memory efficient transfer learning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3 +[39] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical report, California Institute of Technology, 2011. 6 +[40] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In 10th International Conference on Learning Representations (ICLR), 2022. 2 +[41] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer G. Dy, and Tomas Pfister. Learning to prompt for continual learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 139-149, 2022. 2 +[42] Xiaohan Yan and Jacob Bien. Hierarchical Sparse Modeling: A choice of two group lasso formulations. Statistical Science, 32(4):531 - 560. 5 +[43] Ming Yuan and Yi Lin. Model selection and estimation in regression with grouped variables. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 68(1): 49-67, 2006. 5 +[44] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv:1910.04867, 2019. 1, 2, 3 +[45] Jeffrey O. Zhang, Alexander Sax, Amir Zamir, Leonidas J. Guibas, and Jitendra Malik. Side-tuning: A baseline for network adaptation via additive side networks. In European Conference on Computer Vision (ECCV), pages 698-714, 2020. 1, 2, 3, 6 +[46] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 16795-16804, 2022. 2 +[47] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2 +[48] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2021. 2 \ No newline at end of file diff --git a/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/images.zip b/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..31e3f9b743d00a5a223cd30667a140fbd2937d61 --- /dev/null +++ b/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd932ef59971783b170d567a689bd7dbfd0fad6b7014abd5ca8fd871f2bcfd8f +size 678822 diff --git a/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/layout.json b/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9e35a70937de6566449e898e7f95e6654ba1f0d7 --- /dev/null +++ b/2024/Structured Model Probing_ Empowering Efficient Transfer Learning by Structured Regularization/layout.json @@ -0,0 +1,8285 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 84, + 513, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 84, + 513, + 120 + ], + "spans": [ + { + "bbox": [ + 74, + 84, + 513, + 120 + ], + "type": "text", + "content": "Structured Model Probing: Empowering Efficient Transfer Learning by Structured Regularization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 141, + 499, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 141, + 499, + 175 + ], + "spans": [ + { + "bbox": [ + 86, + 141, + 499, + 175 + ], + "type": "text", + "content": "Zhi-Fan Wu Chaojie Mao Xue Wang Jianwen Jiang Yiliang Lv Rong Jin Alibaba Group" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 221, + 192, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 221, + 192, + 235 + ], + "spans": [ + { + "bbox": [ + 143, + 221, + 192, + 235 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 248, + 290, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 290, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 290, + 571 + ], + "type": "text", + "content": "Despite encouraging results from recent developments in transfer learning for adapting pre-trained model to downstream tasks, the performance of model probing is still lagging behind the state-of-the-art parameter efficient tuning methods. Our investigation reveals that existing model probing methods perform well for the easy case when the source domain (where models are pre-trained) and the adapted domain are similar, but fail for the difficult case when the two domains are significantly different. Simply incorporating features extracted from multiple layers and increasing complexity of the probing model can mitigate the gap in the difficult case, but degrades the performance in the easy case. To address this challenge, we propose structured model probing (SMP) that is able to deliver good performance for both cases through structured regularization. The regularization performs feature selection leveraging model structure as a prior, and controls the complexity of the probing model through the weights of selected structures. This enables us to construct a simple adaptation model, with a small number of selected features and a linear prediction model, for the easy case; and to automatically increase the complexity of adaptation model, with a large number of selected features and a non-linear model, for the difficult case. Our extensive empirical studies show that SMP significantly outperforms the state-of-the-art methods for parameter efficient tuning, and at the same time, still maintains the advantage of computational efficiency for probing-based methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 596, + 128, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 128, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 128, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "The parameters of models have grown drastically with the rapid development of deep learning [5, 37], posing challenges for adapting pre-trained models to downstream tasks. The widely used transfer strategy, fully fine-tuning, becomes infeasible due to excessive computation and storage overhead. Recently, parameter efficient tuning [19, 23] has gained attention for reducing storage costs and improving tuning performance with few trainable parameters when lim" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 224, + 544, + 338 + ], + "blocks": [ + { + "bbox": [ + 310, + 224, + 544, + 338 + ], + "lines": [ + { + "bbox": [ + 310, + 224, + 544, + 338 + ], + "spans": [ + { + "bbox": [ + 310, + 224, + 544, + 338 + ], + "type": "image", + "image_path": "63f688131516d3ed72daf25e836154d3ffd2d13587e9edb41098c34830f2f3e3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 350, + 547, + 372 + ], + "lines": [ + { + "bbox": [ + 306, + 350, + 547, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 547, + 372 + ], + "type": "text", + "content": "Figure 1. Performance comparison on different dataset groups (left) and mean relative adaptation time (right) on VTAB-1k benchmark." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 388, + 547, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 448 + ], + "type": "text", + "content": "ited data are available for downstream tasks. Despite the progress, parameter efficient tuning methods still face significant computational overhead in training due to the need to perform forward and backward passes through the pretrained model at each training step." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 449, + 548, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 548, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 548, + 605 + ], + "type": "text", + "content": "Model probing, which leverages features extracted from the pre-trained model to perform adaptation, is promising due to its efficiency and simplicity: The training procedure is efficient since the pre-trained model is fixed and only performs one forward propagation to extract features, and probing model is decoupled from pre-trained model that makes it convenient to deploy small sized task-specific downstream models. However, it is a nuisance that probing methods are usually outperformed by state-of-the-art parameter efficient tuning methods. This is true even after several recent improvements made for the probing methods [16, 45]. In this paper, we aim to improve the performance of model probing significantly without sacrificing its simplicity and efficiency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "content": "We first investigate when model probing methods perform worse than the tuning-based methods on widely used visual task adaptation benchmark VTAB-1k [44] as shown in Figure 1. Target domains are categorized as easy and difficult cases based on their similarity to the source domain where the pre-trained model is trained on (see Section 3.1 for details). We find simple linear probing yields comparable performance as the two tuning-based approaches (finetuning and VPT [23]) for easy cases, whereas a significant" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16849" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 246, + 194 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 246, + 194 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 246, + 194 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 246, + 194 + ], + "type": "image", + "image_path": "8976b25f7d62a3b031921846e07f4362b5639759867d578f4903a604d17738bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 202, + 487, + 215 + ], + "lines": [ + { + "bbox": [ + 104, + 202, + 487, + 215 + ], + "spans": [ + { + "bbox": [ + 104, + 202, + 487, + 215 + ], + "type": "text", + "content": "Figure 2. Adaptation pipelines of Model Tuning (left) and the proposed Structured Model Probing (right)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 253, + 68, + 545, + 194 + ], + "blocks": [ + { + "bbox": [ + 253, + 68, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 253, + 68, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 253, + 68, + 545, + 194 + ], + "type": "image", + "image_path": "7160d55c197b3be1ab48dc688ed30909c72b2fb4a9022aedc0f7297b00ee4323.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 227, + 288, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 288, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 288, + 286 + ], + "type": "text", + "content": "performance gap is observed between linear probing and tuning-based methods for difficult cases. This suggests that model probing exhibits different behavior for various target domains, and performance gap arises for target domains that dissimilar to the source domain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 288, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 287, + 432 + ], + "type": "text", + "content": "This is not surprising that we expect a more complicated model than a linear probing to compensate the difference between the source domain and the target domain. In general, we can increase model complexity in two ways, either by introducing more features extracted from the pre-trained model, or by introducing a non-linear model for prediction. A main challenge is how to develop a unified framework of model probing that is able to handle both the easy and difficult adaptations effectively. This is because, according to our study in Section 3, for easy cases, a complicated probing model (with more features and non-linear probing) performs significantly worse than a simple linear probing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 434, + 288, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 288, + 624 + ], + "type": "text", + "content": "To this end, we propose structured model probing, or SMP for short, that effectively address both easy and difficult cases of domain adaptation. The key to our unified framework is structured regularization. It performs feature selection leveraging the structure of the pre-trained model as a prior, and controls the complexity of the probing model through the weights assigned to selected features. Thus, for the easy case, this regularization constructs a simple linear model with a few selected features; and for the difficult case, it automatically introduces more features from the pre-trained model and more non-linearity in the prediction module based on the value of loss function, as shown in Figure 2. Our empirical studies show that the proposed structured model probing performs significantly better than state-of-the-art tuning-based approaches, while maintaining the advantage of computational efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 625, + 215, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 625, + 215, + 636 + ], + "spans": [ + { + "bbox": [ + 59, + 625, + 215, + 636 + ], + "type": "text", + "content": "Our main contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 641, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 641, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 641, + 288, + 713 + ], + "type": "text", + "content": "1. We observe and demonstrate the conflicting behaviors of model probing: a complicated probing model with more features and non-linear transformation can improve the performance for difficult cases of domain adaptation, but it yields worse performance than linear probing for easy cases of domain adaptation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 227, + 547, + 359 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 227, + 547, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 227, + 547, + 299 + ], + "spans": [ + { + "bbox": [ + 306, + 227, + 547, + 299 + ], + "type": "text", + "content": "2. We propose structured model probing (SMP) based on structured regularization. This regularization enables us to construct simple probing model for easy tasks, and automatically increase model complexity for difficult tasks, to simultaneously address both the easy and difficult cases of domain adaptation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 299, + 547, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 299, + 547, + 359 + ], + "spans": [ + { + "bbox": [ + 306, + 299, + 547, + 359 + ], + "type": "text", + "content": "3. We conduct extensive empirical studies on various adaptation tasks. The proposed SMP method shows superior experimental results compared to state-of-the-art probing and tuning methods, and requires significantly lower training costs compared to tuning methods." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 373, + 392, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 392, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 392, + 385 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 396, + 547, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 468 + ], + "type": "text", + "content": "Transfer learning has been extensively studied in vision domain [34, 48]. It aims to improve the performance on target domain by leveraging the related information contained in source domain. Transfer learning based on pre-trained model [8, 20, 21] has received significant attention due to its simplicity and good empirical performance [23, 45]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 471, + 547, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 547, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 547, + 687 + ], + "type": "text", + "content": "Model tuning adapts the pre-trained model to new tasks via updating model parameters. Fully fine-tuning is a widely used strategies due to its efficacy and simplicity [9, 25, 44]. However, it suffers from huge storage and computation overhead due to the ever-expanding size of large-scale pre-trained model [37, 40]. Heuristic approaches have been proposed to improve fine-tuning performance [3, 6, 17, 18, 27]. Parameter efficient tuning [19] aims to reduce the storage requirement of transferred model by updating a few parameters attached to the pre-trained model [23, 47], like Prompt Tuning [29, 30] or Adapter [22]. While originally studied in natural language processing tasks due to the dominance of large-scale pre-trained language models [5, 14], there is growing interest in applying these methods to computer vision tasks [7, 23, 31, 41, 46, 47]. Though parameter efficient tuning reduces storage cost significantly, the training cost remains high due to the forward and backward propagation need to pass through the entire pre-trained model [11]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "Model probing utilizes frozen extracted features for adapting the pre-trained model to new tasks, thus achieving effi" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16850" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "cient adaptation. The most popular method, linear probing, suffers from performance degradation [25], which limits its application in transfer learning [20, 21]. Recent methods [16, 36, 45] have been proposed to improve probing performance by leveraging side network [38, 45] or incorporating intermediate representations [2, 16, 32], but there is still a significant performance gap compared to tuning methods. This motivates us to investigate the limitation of probing methods, and come up with a better solution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 192, + 289, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 192, + 289, + 220 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 289, + 220 + ], + "type": "text", + "content": "3. Understanding the Limitation of Model Probing for Transfer Learning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 226, + 288, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 226, + 288, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 288, + 286 + ], + "type": "text", + "content": "In this section, we analyze the performance of probing methods across various downstream domains to understand their limitations, motivated by the observation in Figure 1. We assess feature informativeness for different tasks and investigate the necessity of non-linear transformation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 294, + 139, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 294, + 139, + 307 + ], + "spans": [ + { + "bbox": [ + 47, + 294, + 139, + 307 + ], + "type": "text", + "content": "3.1. Analysis Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 316, + 287, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 364 + ], + "type": "text", + "content": "Pre-trained model and source data. The ViT-B/16 [15] is used as the backbone because of the extensive usage of transformer-based models. The model is pre-trained on ImageNet-21k [12] in a supervised learning manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 366, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 450 + ], + "type": "text", + "content": "Target data. We use the VTAB-1k benchmark, consisting of 19 datasets that cover natural images (natural), images captured by specialist equipment (specialized), and images generated from simulated environments (structured), as our target data for transfer learning. We follow the setting in [44] and use 1000 samples as the training set per task to evaluate adaptation with limited data as suggested in [44]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 453, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 501 + ], + "type": "text", + "content": "Domain similarity. Several previous works have proposed metrics to measure the similarity between source domain and target domain [1, 10, 13, 16, 33]. Here, we adopt the domain similarity measure defined in [16]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 78, + 511, + 287, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 511, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 78, + 511, + 287, + 523 + ], + "type": "interline_equation", + "content": "D o m a i n \\quad \\text {S i m i l a r i t y} = \\mathrm {A c c} _ {\\text {L i n e a r}} - \\mathrm {A c c} _ {\\text {S c r a t c h}}. \\tag {1}", + "image_path": "b4a086e35ee24399681fc84fb06a4ed1502ff3b104522ed1ff8427222f7fbd9d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 665 + ], + "type": "text", + "content": "A higher domain similarity suggests that the features extracted from a pre-trained model can be utilized for the target domain, leading to performance gains via a linear classifier compared to training from scratch. Conversely, a lower domain similarity indicates that training from scratch on the target domain may be more effective than using extracted features. Thus, the domain similarity is evaluated based on the benefit gained by a linear classifier compared to training from scratch. We also report the label-feature correlation [13] in Suppl. C.8, which exhibits a high Spearman rank correlation (0.854) with the domain similarity scores." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "Domain similarity also reflects how difficult to adapt from source domain to target domain. As shown in Figure 4, tasks in natural and specialized groups have higher domain similarity to source domain where the pre-trained model is" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 426, + 186 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 426, + 186 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 426, + 186 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 426, + 186 + ], + "type": "image", + "image_path": "e6ce50144cb8482b59edc659595afb130625e9a8932639916a34f16206715741.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 198, + 545, + 221 + ], + "lines": [ + { + "bbox": [ + 305, + 198, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 545, + 221 + ], + "type": "text", + "content": "Figure 3. Performance of features extracted from different layers on Sun397 (easy) and Clevr-Dist (difficult)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 427, + 73, + 542, + 186 + ], + "blocks": [ + { + "bbox": [ + 427, + 73, + 542, + 186 + ], + "lines": [ + { + "bbox": [ + 427, + 73, + 542, + 186 + ], + "spans": [ + { + "bbox": [ + 427, + 73, + 542, + 186 + ], + "type": "image", + "image_path": "a2274185ec5d76b15ee835772fee332d9d4e0b99a6ae64cf014e11b7db3753ed.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 236, + 545, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 236, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 545, + 271 + ], + "type": "text", + "content": "trained on, so we mark them as easy cases. And tasks in structured groups have lower domain similarity to source domain, thus they are labeled as difficult cases in Figure 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 280, + 507, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 280, + 507, + 292 + ], + "spans": [ + { + "bbox": [ + 306, + 280, + 507, + 292 + ], + "type": "text", + "content": "3.2. Informativeness of Extracted Features" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 299, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 299, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 299, + 545, + 335 + ], + "type": "text", + "content": "To quantify the informativeness of features extracted from different layers for downstream tasks, we use leave-one-out " + }, + { + "bbox": [ + 305, + 299, + 545, + 335 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 299, + 545, + 335 + ], + "type": "text", + "content": "-NN testing as the surrogate performance criterion:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 354, + 545, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 545, + 381 + ], + "type": "interline_equation", + "content": "\\text {I n f o r m a t i v e n e s s} = \\frac {\\sum_ {i = 1} ^ {m} \\sum_ {x _ {j} \\in N (x _ {i})} \\mathbb {1} \\left\\{y _ {i} = y _ {j} \\right\\}}{k m}, \\tag {2}", + "image_path": "851b8cd6d8862454ac4b418110df1e294769f3a865f03105da683aad6784ec52.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "text", + "content": " is the number of samples, " + }, + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "text", + "content": " is the extracted features of a sample, " + }, + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "inline_equation", + "content": "N(x_{i})" + }, + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "text", + "content": " nearest neighbors of " + }, + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "text", + "content": ". We simply set " + }, + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "text", + "content": ", which results in leave-one-out 1-NN testing. We choose " + }, + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "text", + "content": "-NN, instead of linear classifier, for informativeness measurement because " + }, + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 389, + 547, + 473 + ], + "type": "text", + "content": "-NN is a non-parametric classifier and thus avoids training additional parameters, making the measurement more robust." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 594 + ], + "type": "text", + "content": "In Figure 3, we plot the informativeness of features extracted from different layers on two tasks from different domains. For Sun397, an easy domain adaption task, we observe the best performance achieved by features extracted from deep layers. This explains the good performance of linear probing for easy domain adaption tasks, as observed in Figure 1. While for Clevr-Dist, a difficult task with target domain significantly different from the source domain, we observe that no layer has outstanding performance. The results echo the findings in [16]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "Furthermore, we conduct experiments on gradually incorporating features extracted from different layers of the pre-trained model, demonstrating the different behaviors on easy and difficult tasks, which can be found in Section 5.4. The results show for easy tasks, more features may lead to overfitting, but for difficult tasks, performance improves with more features. Therefore, when the target domain is very different from the source domain, we need to provide probing model with diverse features extracted from multiple layers of the pre-trained model." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16851" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 542, + 216 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 542, + 216 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 542, + 216 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 542, + 216 + ], + "type": "image", + "image_path": "3939f6c2b6b291cfdf9500bca9d202299da8bc6a024f8e3cf50e6af13a5703c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 228, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 547, + 251 + ], + "type": "text", + "content": "Figure 4. Performance of extracted features with linear/ non-linear transformation on VTAB-1k. Datasets are ordered from easy tasks (left) to difficult tasks (right) according to domain similarity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 262, + 244, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 244, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 244, + 274 + ], + "type": "text", + "content": "3.3. Effects of Non-linear Transformation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 280, + 288, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 288, + 352 + ], + "type": "text", + "content": "Since there is no significantly informative layer for difficult domain adaption tasks, a linear combination may not be enough for exploiting the information in extracted features. We further examine the efficacy of non-linear combination on extracted features, as neural networks achieve superior performance through non-linear feature transformation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 353, + 288, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 288, + 411 + ], + "type": "text", + "content": "We employ an multiple-layer perceptron (MLP) for nonlinear transformation and compare its performance with linear transformation. Features extracted from different structures of the pre-trained model are concatenated together as input (See Section 4.1 for details)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 412, + 288, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 288, + 568 + ], + "type": "text", + "content": "In Figure 4, we clearly observe that adding non-linearity to feature transformation can significantly boost the performance for the difficult domain adaption tasks (i.e. the datasets on right end of Figure 4). This phenomenon indicates that simply performing linear transformation on extracted features can not exploit the full potential of representations extracted from the pre-trained model. On the other hand, we notice that the performance on easy tasks is not improved, and is even hurt with the introduction of non-linearity. We speculate that this is because the linear transformation can already make full use of extracted information, while non-linearity increases model capacity, which eventually leads to overfitting to the training data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Discussion. Our investigation reveals that the optimal probing model differs depending on the domain of downstream task. Incorporating diverse features from multiple layers of the pre-trained model and performing non-linear transformation on them can significantly improve the performance on difficult cases. However, this yields worse results on easy tasks. On the contrary, simply performing linear combination on features from deep layers can achieve satisfactory performance for easy adaptation tasks, but not for difficult ones. The conflicting behaviors of model probing motivates us to develop a unified framework that can solve the adaptation problems for both easy and difficult adaptation tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 261, + 457, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 261, + 457, + 275 + ], + "spans": [ + { + "bbox": [ + 305, + 261, + 457, + 275 + ], + "type": "text", + "content": "4. Structured Model Probing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 283, + 547, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 547, + 344 + ], + "type": "text", + "content": "In this section, we propose structured model probing, a unified probing method which can deal with both easy and difficult cases of adaptation tasks. We first present how to thoroughly extract features in a structured way, then introduce our probing model with structured regularization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 356, + 471, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 356, + 471, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 356, + 471, + 369 + ], + "type": "text", + "content": "4.1. Structured Feature Extraction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 376, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 547, + 533 + ], + "type": "text", + "content": "Incorporating diverse features extracted from the pre-trained model can improve the performance on difficult adaptation tasks for model probing, as shown in Section 3.2. To provide diverse features for the probing model, we extract features from all structures in a Vision Transformer (ViT). These structures include features before self-attention, features after self-attention, and features after Feed Forward Network (FFN), for each transformer layer. We also include tokenized image inputs and pre-logits. We use " + }, + { + "bbox": [ + 304, + 376, + 547, + 533 + ], + "type": "inline_equation", + "content": "V = \\{v_{i}\\}_{i}^{|V|}" + }, + { + "bbox": [ + 304, + 376, + 547, + 533 + ], + "type": "text", + "content": " represents all structures in ViT, and " + }, + { + "bbox": [ + 304, + 376, + 547, + 533 + ], + "type": "inline_equation", + "content": "|V|" + }, + { + "bbox": [ + 304, + 376, + 547, + 533 + ], + "type": "text", + "content": " is the total number of structures. For an input sample, we denote each set of tokens extracted from structure " + }, + { + "bbox": [ + 304, + 376, + 547, + 533 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 376, + 547, + 533 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 376, + 547, + 533 + ], + "type": "inline_equation", + "content": "X_{i} \\in \\mathbb{R}^{t \\times c}" + }, + { + "bbox": [ + 304, + 376, + 547, + 533 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 376, + 547, + 533 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 376, + 547, + 533 + ], + "type": "text", + "content": " is the dimension of tokens and " + }, + { + "bbox": [ + 304, + 376, + 547, + 533 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 376, + 547, + 533 + ], + "type": "text", + "content": " is the dimension of channels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "text", + "content": "To reduce feature redundancy while preserving diversity, we apply 1D average pooling along different dimensions on tokens " + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "text", + "content": ", resulting in two types of aggregation. To maintain channel information, we aggregate " + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "text", + "content": " along the token dimension, which produces " + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "inline_equation", + "content": "x_{i}^{t}" + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "text", + "content": ". And to preserve spatial information, we aggregate " + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "text", + "content": " along the channel dimension, resulting in " + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "inline_equation", + "content": "x_{i}^{c}" + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "text", + "content": ". Finally, we concatenate all structured features into a single vector, " + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "inline_equation", + "content": "x = [x_{1}^{t}, x_{1}^{c}, \\dots, x_{|V|}^{t}, x_{|V|}^{c}]" + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "text", + "content": ". We denote " + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "inline_equation", + "content": "s_{i}" + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "text", + "content": " as the index set of elements in " + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "inline_equation", + "content": "S = \\{s_{i} | i = 1\\dots |s|\\}" + }, + { + "bbox": [ + 304, + 535, + 548, + 647 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 657, + 516, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 657, + 516, + 670 + ], + "spans": [ + { + "bbox": [ + 305, + 657, + 516, + 670 + ], + "type": "text", + "content": "4.2. Probing with Structured Regularization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Our structured regularization contains two components: structured sparsity regularizer and structured non-linearity regularizer. We introduce them as follows." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16852" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 203, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 203, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 203, + 83 + ], + "type": "text", + "content": "4.2.1 Structured Feature Selection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 175 + ], + "type": "text", + "content": "As shown in Section 3.2, the informativeness of extracted features varies with tasks, and redundant features lead to performance degradation on some tasks. Thus, we propose to perform feature selection during the training stage of the probing model to prevent overfitting. A popular way is to use group lasso [16, 43] by constructing groups on feature dimension over the weight matrix of a linear classifier:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 180, + 287, + 218 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 287, + 218 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} L (\\theta) + \\lambda \\sum_ {i = 1} ^ {d} | | \\theta_ {i} | | _ {2}, \\text {w h e r e} | | \\theta_ {i} | | _ {2} = \\sqrt {\\sum_ {j = 1} ^ {n} \\theta_ {i j} ^ {2}}. \\tag {3}", + "image_path": "008ec398a001978468ae7835dc8dc2d93b2b0a032cc8754b0d29743d6a10f38c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 239, + 289, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 348 + ], + "type": "text", + "content": "We use " + }, + { + "bbox": [ + 46, + 239, + 289, + 348 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^{d\\times n}" + }, + { + "bbox": [ + 46, + 239, + 289, + 348 + ], + "type": "text", + "content": " to represent weight matrix, where " + }, + { + "bbox": [ + 46, + 239, + 289, + 348 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 239, + 289, + 348 + ], + "type": "text", + "content": " is the feature dimension and " + }, + { + "bbox": [ + 46, + 239, + 289, + 348 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 239, + 289, + 348 + ], + "type": "text", + "content": " is the number of classes. However, due to the high dimensionality of features, the final achieved sparsity is hardly the optimal result. It is worth noting that the structure of pre-trained model provides a natural way to divide features as groups, and features extracted from the same structure have higher relevance to each other. Thus, we leverage the structure of the pre-trained model, to achieve the structured sparsity with a structured sparsity regularizer:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 363, + 290, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 363, + 290, + 411 + ], + "spans": [ + { + "bbox": [ + 48, + 363, + 290, + 411 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} L (\\theta) + \\lambda \\sum_ {s \\in S} \\frac {1}{| s |} \\| \\theta_ {s} \\| _ {2}, \\text {w h e r e} \\| \\theta_ {s} \\| _ {2} = \\sqrt {\\sum_ {i \\in s} \\sum_ {j = 1} ^ {n} \\theta_ {i j} ^ {2}}, \\tag {4}", + "image_path": "07632c6400cf594fb67d84de9c4fa24588b6b2e6a0af3c6df3efa8dc2b88cf3f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 411, + 288, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 288, + 518 + ], + "type": "inline_equation", + "content": "s \\in S" + }, + { + "bbox": [ + 46, + 411, + 288, + 518 + ], + "type": "text", + "content": " is the index set of a specific structure in feature as described in Section 4.1, and " + }, + { + "bbox": [ + 46, + 411, + 288, + 518 + ], + "type": "inline_equation", + "content": "|s|" + }, + { + "bbox": [ + 46, + 411, + 288, + 518 + ], + "type": "text", + "content": " represents its dimension. Compared with constructing groups by features, in our regularizer, we construct groups by leveraging the structure of the pre-trained model. This formalization can significantly reduce the combinations of selection and thus prevent overfitting. With such an advantage, our method achieves superior performance with a small number of samples under high-dimensional features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 533, + 245, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 245, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 245, + 544 + ], + "type": "text", + "content": "4.2.2 Structured Non-linear Transformation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 552, + 289, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 289, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 289, + 694 + ], + "type": "text", + "content": "Our discovery in Section 3.3 shows non-linear transformation can facilitate transfer performance for difficult tasks, while could reduce performance for easy adaptation tasks, mostly due to over-fitting. This means that we need a model that allows flexibly incorporating non-linearity to a linear model. We propose a structured non-linearity regularizer that nicely integrates feature selection (i.e. choosing the subset of informative features from the pre-trained model) with model selection (i.e. choosing either a linear model or nonlinear model for prediction) inspired by hierarchical sparse modeling [28, 42]. Specifically, we construct our model by the combination of a linear model and a neural network:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 119, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 119, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "f (x) = \\theta^ {\\top} x + f _ {W} (x), \\tag {5}", + "image_path": "31e326bff25312203a34ac18e9af8ac8f9ea3aba12684c4ad833062bea815749.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^{d\\times n}" + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": " is a linear model, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "f_{W}" + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": " represents an MLP with model weights " + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "W^{(1)}" + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": " to represent the input layer of the MLP, and the parameters of rest layers are denoted as " + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "W^{(2:)}" + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 121, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 156 + ], + "type": "text", + "content": "We utilize the feature selection result " + }, + { + "bbox": [ + 304, + 121, + 545, + 156 + ], + "type": "inline_equation", + "content": "||\\theta ||_2" + }, + { + "bbox": [ + 304, + 121, + 545, + 156 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 304, + 121, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 121, + 545, + 156 + ], + "type": "text", + "content": " norm of weights assigned to the selected features, to regularize the complexity of the non-linear model:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 157, + 545, + 252 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 306, + 157, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 157, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 157, + 545, + 204 + ], + "type": "text", + "content": "- When a small number of features are good enough for prediction, we expect a small value for " + }, + { + "bbox": [ + 306, + 157, + 545, + 204 + ], + "type": "inline_equation", + "content": "||\\theta||_2" + }, + { + "bbox": [ + 306, + 157, + 545, + 204 + ], + "type": "text", + "content": ". By using " + }, + { + "bbox": [ + 306, + 157, + 545, + 204 + ], + "type": "inline_equation", + "content": "||\\theta||_2" + }, + { + "bbox": [ + 306, + 157, + 545, + 204 + ], + "type": "text", + "content": " to control the complexity of non-linear model, we expect the resulting model to be mostly linear." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 205, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 205, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 545, + 252 + ], + "type": "text", + "content": "- With increasing " + }, + { + "bbox": [ + 306, + 205, + 545, + 252 + ], + "type": "inline_equation", + "content": "||\\theta||_2" + }, + { + "bbox": [ + 306, + 205, + 545, + 252 + ], + "type": "text", + "content": ", more diverse structures from the pre-trained model are needed for prediction. By using " + }, + { + "bbox": [ + 306, + 205, + 545, + 252 + ], + "type": "inline_equation", + "content": "||\\theta||_2" + }, + { + "bbox": [ + 306, + 205, + 545, + 252 + ], + "type": "text", + "content": " to regularize the size of non-linear model, we expect the resulting model to be more non-linear." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 253, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 253, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 545, + 277 + ], + "type": "text", + "content": "A way to achieve such control is to add a constraint to the objective function in Eq.(4) as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 375, + 286, + 492, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 286, + 492, + 311 + ], + "spans": [ + { + "bbox": [ + 375, + 286, + 492, + 311 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta , W} L (\\theta , W) + \\lambda \\sum_ {s \\in S} | | \\theta_ {s} | | _ {2}", + "image_path": "cd84749c2450b9461cfe19cf58eaddf0466f66dd71f8aebcaa30aaebfb2e7338.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 333, + 311, + 545, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 311, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 333, + 311, + 545, + 346 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\text {s u b j e c t} \\left. | | W _ {s} ^ {(1)} | | _ {2} \\leq M _ {1} | | \\theta_ {s} | | _ {2}, s = 1 \\dots | S |, \\right. \\tag {6} \\\\ \\left| \\left| W ^ {(2:)} \\right| \\right| _ {2} \\leq M _ {2} \\left| \\left| \\theta \\right| \\right| _ {2}. \\\\ \\end{array}", + "image_path": "ba2602980fcdb4c1b014b9dc41fd796e3e54671a7eccd61a604b541aab522b30.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 355, + 547, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 547, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 547, + 427 + ], + "type": "text", + "content": "There are two constraints in Eq.(6): the first constraint is used to maintain the input feature sparsity of the neural network, and the second term is used to control the non-linearity through the linear model. To achieve efficient implementation, we convert the constraint in Eq.(6) into a regularization term, which we call structured non-linearity regularizer:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 322, + 436, + 545, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 436, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 322, + 436, + 545, + 479 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Omega (\\theta , W) = \\sum_ {s \\in S} \\max \\left\\{\\left| \\left| W _ {s} ^ {(1)} \\right| \\right| _ {2} - M _ {1} \\left| \\left| \\theta_ {s} \\right| \\right| _ {2}, 0 \\right\\} \\tag {7} \\\\ + \\max \\left\\{\\left| \\left| W ^ {(2:)} \\right| \\right| _ {2} - M _ {2} \\left| \\left| \\theta \\right| \\right| _ {2}, 0 \\right\\}. \\\\ \\end{array}", + "image_path": "ee28bfae3beb1dd390053c7102a79263428f1970454aa4d755c9a2e509f616ba.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 488, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 488, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 488, + 545, + 525 + ], + "type": "text", + "content": "This regularization term means that, when " + }, + { + "bbox": [ + 306, + 488, + 545, + 525 + ], + "type": "inline_equation", + "content": "||W||_2 > M||\\theta||_2" + }, + { + "bbox": [ + 306, + 488, + 545, + 525 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 488, + 545, + 525 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 306, + 488, + 545, + 525 + ], + "type": "text", + "content": " receives penalty until it achieves similar complexity as " + }, + { + "bbox": [ + 306, + 488, + 545, + 525 + ], + "type": "inline_equation", + "content": "M||\\theta||" + }, + { + "bbox": [ + 306, + 488, + 545, + 525 + ], + "type": "text", + "content": ". Then our final objective function is:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 329, + 535, + 545, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 535, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 329, + 535, + 545, + 559 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta , W} L (\\theta , W) + \\lambda_ {1} \\sum_ {s \\in S} | | \\theta_ {s} | | _ {2} + \\lambda_ {2} \\Omega (\\theta , W). \\tag {8}", + "image_path": "5468aa0a1ad5700d0a4af5c42f6b32266d1be57e107622e213e21c21f8bace8c.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 570, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 570, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 570, + 545, + 594 + ], + "type": "text", + "content": "We use " + }, + { + "bbox": [ + 306, + 570, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 306, + 570, + 545, + 594 + ], + "type": "text", + "content": " to control the penalty strength, and simply set " + }, + { + "bbox": [ + 306, + 570, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 0.1" + }, + { + "bbox": [ + 306, + 570, + 545, + 594 + ], + "type": "text", + "content": " works well in our experiments." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 596, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 596, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 596, + 545, + 656 + ], + "type": "text", + "content": "Loss function. Since our model " + }, + { + "bbox": [ + 304, + 596, + 545, + 656 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 304, + 596, + 545, + 656 + ], + "type": "text", + "content": " consists of both linear and non-linear parts, we decouple these two components in loss function. This modification enables the structured sparsity regularizer to control the complexity of " + }, + { + "bbox": [ + 304, + 596, + 545, + 656 + ], + "type": "inline_equation", + "content": "\\| \\theta \\|" + }, + { + "bbox": [ + 304, + 596, + 545, + 656 + ], + "type": "text", + "content": " without the interference from the non-linear part:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 311, + 665, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 665, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 311, + 665, + 545, + 681 + ], + "type": "interline_equation", + "content": "L (\\theta , W) = \\operatorname {C E} \\left(\\theta^ {\\top} x, y\\right) + \\operatorname {C E} \\left(\\operatorname {s g} \\left(\\theta^ {\\top} x\\right) + f _ {W} (x), y\\right), \\tag {9}", + "image_path": "54a5b6f71c9fa12ff4602792940d17d30d7efa39b0618a014767f768d08d9452.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " is the one-hot label vector of " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": ", CE is cross-entropy loss and sg means stop-gradient." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16853" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 92, + 547, + 342 + ], + "blocks": [ + { + "bbox": [ + 47, + 60, + 545, + 84 + ], + "lines": [ + { + "bbox": [ + 47, + 60, + 545, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 60, + 545, + 84 + ], + "type": "text", + "content": "Table 1. Median test accuracy (%) over 3 seeds on the VTAB-1k benchmark using ViT-B/16 pretrained on ImageNet-21k. * indicates results are obtained from [23] and † indicates results are obtained from [16]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 92, + 547, + 342 + ], + "lines": [ + { + "bbox": [ + 47, + 92, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 92, + 547, + 342 + ], + "type": "table", + "html": "
NaturalSpecializedStructured
CTFAR-100Caltech101DTDFlowers102PetsSVHNSun397Mean (Natural)CamelyonEuroSATResisc45RetinopathyMean (Specialized)Clevr-CountClevr-DistDMLabKITTI-DistdSpr-LocdSpr-OrisNORB-AzimsNORB-ElevMean (Structured)Mean (All)
Tuning Methods
Scratch†7.619.113.129.66.719.42.314.071.071.029.372.060.831.652.527.239.166.129.711.724.135.332.8
Fine-tuning*68.987.764.397.286.987.438.875.979.795.784.273.983.756.358.641.765.557.546.725.729.147.665.6
Bias*72.887.059.297.585.359.951.473.378.791.672.969.878.361.555.632.455.966.640.015.725.144.162.0
VPT*78.890.865.898.088.378.149.678.581.896.183.468.482.468.560.046.572.873.647.932.937.855.069.4
Probing Methods
Linear77.286.866.699.188.835.755.172.778.489.672.273.778.531.029.935.254.911.727.714.423.728.655.4
Side-Tuning*60.760.853.695.566.734.935.358.258.587.765.261.068.127.622.631.351.78.214.49.821.823.445.6
Head2Toe75.390.975.099.586.183.550.980.284.495.784.474.384.751.759.444.065.647.140.332.541.147.767.5
LinearStruc.72.989.572.598.983.671.642.776.084.396.081.875.084.348.252.539.466.748.736.732.337.745.364.8
LinearStruc.w/FSR78.290.173.499.388.171.745.478.083.996.181.974.184.048.552.839.664.451.438.632.937.045.765.7
LinearStruc.w/SSR79.090.274.199.190.473.454.280.184.696.383.874.484.848.852.939.866.948.738.932.537.945.866.6
MLPStruc.66.888.770.397.482.171.133.372.881.695.977.073.582.077.257.739.366.572.344.132.747.754.767.1
SMP w/o SNR71.389.672.399.085.672.745.076.584.295.981.972.883.777.457.140.867.172.444.232.648.455.069.0
SMP79.390.974.999.390.475.055.380.784.896.383.175.084.877.558.040.867.572.544.533.049.055.470.9
", + "image_path": "573bcd4ee05f2e73a3fd2ac396c5f700fde6e2bcc0248c0abd0cfb3fb1862da7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 350, + 128, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 128, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 128, + 363 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 369, + 288, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 288, + 407 + ], + "type": "text", + "content": "In this section, we evaluate the proposed structured model probing method on multiple datasets across various tasks, and provide detailed analysis of the proposed method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 407, + 288, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 288, + 468 + ], + "type": "text", + "content": "Implementation details. We evaluate SMP on the popular vision backbone, ViT-B/16 [15] pre-trained on ImageNet-21k [12], which is widely used across relative research works [16, 23]. We provide details of probing model and hyperparameter selection in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 470, + 288, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 470, + 288, + 543 + ], + "spans": [ + { + "bbox": [ + 47, + 470, + 288, + 543 + ], + "type": "text", + "content": "Compared methods. We compare SMP with other commonly used and state-of-the-art tuning (Fine-tuning, Bias [6] and VPT [23]) and probing methods (Linear Probing, Side-Tuning [45] and Head2Toe [16]). All hyperparameters of baseline methods are carefully tuned, and further information is provided in supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 549, + 192, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 192, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 192, + 562 + ], + "type": "text", + "content": "5.1. Experiments on VTAB-1k" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 567, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 288, + 687 + ], + "type": "text", + "content": "Experimental results on the VTAB-1k benchmark (Table 1) indicate that SMP outperforms other methods, achieving a mean performance of " + }, + { + "bbox": [ + 47, + 567, + 288, + 687 + ], + "type": "inline_equation", + "content": "70.9\\%" + }, + { + "bbox": [ + 47, + 567, + 288, + 687 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 567, + 288, + 687 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 47, + 567, + 288, + 687 + ], + "type": "text", + "content": " higher than Head2Toe and " + }, + { + "bbox": [ + 47, + 567, + 288, + 687 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 47, + 567, + 288, + 687 + ], + "type": "text", + "content": " higher than VPT. SMP also significantly outperforms Linear Probing, Bias, and Side-Tuning. SMP achieves optimal results in all three dataset groups. Especially in Structured group, SMP attains a mean accuracy of " + }, + { + "bbox": [ + 47, + 567, + 288, + 687 + ], + "type": "inline_equation", + "content": "55.4\\%" + }, + { + "bbox": [ + 47, + 567, + 288, + 687 + ], + "type": "text", + "content": ", surpassing other probing methods by a significant margin. This demonstrates the incorporation of non-linear transformation is important for difficult adaptation tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "Ablation studies. We provide results of baseline methods that contribute to SMP, to validate the efficacy of structured" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 351, + 527, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 351, + 527, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 351, + 527, + 363 + ], + "type": "text", + "content": "regularization and the contribution of each component." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 363, + 547, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 547, + 423 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 547, + 423 + ], + "type": "text", + "content": "LinearStruc. represents extracting features by structured feature extraction proposed in Section 4.1, then training a linear classifier. Compare to Linear Probing, it gets " + }, + { + "bbox": [ + 304, + 363, + 547, + 423 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 304, + 363, + 547, + 423 + ], + "type": "text", + "content": " average performance gain, indicating that solely extracting features from the last layer is inadequate for model probing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 423, + 547, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 547, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 547, + 507 + ], + "type": "text", + "content": "We further compare feature sparsity regularization in Eq.(3) (represents by " + }, + { + "bbox": [ + 304, + 423, + 547, + 507 + ], + "type": "inline_equation", + "content": "\\mathrm{Linear}_{\\mathrm{Struc.}} + \\mathrm{FSR}" + }, + { + "bbox": [ + 304, + 423, + 547, + 507 + ], + "type": "text", + "content": ") to the proposed structured sparsity regularizer (represents by " + }, + { + "bbox": [ + 304, + 423, + 547, + 507 + ], + "type": "inline_equation", + "content": "\\mathrm{Linear}_{\\mathrm{Struc.}} + \\mathrm{SSR}" + }, + { + "bbox": [ + 304, + 423, + 547, + 507 + ], + "type": "text", + "content": "). Structured sparsity regularizer outperforms feature sparsity regularization, indicating that feature selection can mitigate the overfitting caused by high dimensions, and leveraging model structure as a prior can bring extra benefits." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 507, + 547, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 547, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 547, + 627 + ], + "type": "text", + "content": "We apply non-linear transformation on extracted features, represents by " + }, + { + "bbox": [ + 304, + 507, + 547, + 627 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_{\\text{Struc.}}" + }, + { + "bbox": [ + 304, + 507, + 547, + 627 + ], + "type": "text", + "content": ". We find the performance on structured group gets a significant improvement over " + }, + { + "bbox": [ + 304, + 507, + 547, + 627 + ], + "type": "inline_equation", + "content": "\\mathrm{Linear}_{\\text{Struc.}}" + }, + { + "bbox": [ + 304, + 507, + 547, + 627 + ], + "type": "text", + "content": ", demonstrating the importance of non-linear transformation on difficult adaptation tasks. However, the performance on easy tasks, i.e., natural and specialized groups, is degraded, suggesting the risk of overfitting from non-linear transformation. We solve this problem by leveraging structured nonlinearity regularizer (SNR), which can automatically control the non-linearity through the norms of selected structures." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 635, + 494, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 494, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 494, + 647 + ], + "type": "text", + "content": "5.2. Experiments on Few-shot Learning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "In this section, we evaluate our method in few-shot learning scenario, where the number of training samples varies from 1 to 16 shots per class following existing studies. Five fine-grained recognition tasks are chosen, i.e., Food101 [4], CUB-200-2011 [39], Stanford Cars [26], Stanford Dogs [24]" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16854" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 64, + 214, + 192 + ], + "blocks": [ + { + "bbox": [ + 50, + 64, + 214, + 192 + ], + "lines": [ + { + "bbox": [ + 50, + 64, + 214, + 192 + ], + "spans": [ + { + "bbox": [ + 50, + 64, + 214, + 192 + ], + "type": "image", + "image_path": "817f5dd48fe878aaf1b3177c862d35bfe2ed6a9e9e8e3980f397eaecfecb0609.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 215, + 64, + 378, + 192 + ], + "blocks": [ + { + "bbox": [ + 215, + 64, + 378, + 192 + ], + "lines": [ + { + "bbox": [ + 215, + 64, + 378, + 192 + ], + "spans": [ + { + "bbox": [ + 215, + 64, + 378, + 192 + ], + "type": "image", + "image_path": "e2fe9524fb302c804246dd5f00e0babe0a03ccc14a1e9abb24d528eab516d278.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 379, + 64, + 541, + 192 + ], + "blocks": [ + { + "bbox": [ + 379, + 64, + 541, + 192 + ], + "lines": [ + { + "bbox": [ + 379, + 64, + 541, + 192 + ], + "spans": [ + { + "bbox": [ + 379, + 64, + 541, + 192 + ], + "type": "image", + "image_path": "484d8aefe8aaa968b7c51fd933bef7233ff1cddb14dede596e27d1856b3004b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 51, + 194, + 214, + 322 + ], + "blocks": [ + { + "bbox": [ + 51, + 194, + 214, + 322 + ], + "lines": [ + { + "bbox": [ + 51, + 194, + 214, + 322 + ], + "spans": [ + { + "bbox": [ + 51, + 194, + 214, + 322 + ], + "type": "image", + "image_path": "1dcfc2b477ee64adeaf90836b3949d8250e549672b924d614cbda5c630efdc49.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 335, + 525, + 346 + ], + "lines": [ + { + "bbox": [ + 66, + 335, + 525, + 346 + ], + "spans": [ + { + "bbox": [ + 66, + 335, + 525, + 346 + ], + "type": "text", + "content": "Figure 5. Performance of few-shot learning on fine-grained visual recognition tasks. LinearStr. (F) indicates LinearStruc. w/ FSR." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 215, + 194, + 378, + 322 + ], + "blocks": [ + { + "bbox": [ + 215, + 194, + 378, + 322 + ], + "lines": [ + { + "bbox": [ + 215, + 194, + 378, + 322 + ], + "spans": [ + { + "bbox": [ + 215, + 194, + 378, + 322 + ], + "type": "image", + "image_path": "47064e96137db55553faad763127478d043cbcae2ca099511d7e6dc3d2b89899.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 379, + 194, + 541, + 322 + ], + "blocks": [ + { + "bbox": [ + 379, + 194, + 541, + 322 + ], + "lines": [ + { + "bbox": [ + 379, + 194, + 541, + 322 + ], + "spans": [ + { + "bbox": [ + 379, + 194, + 541, + 322 + ], + "type": "image", + "image_path": "e7681287e02199c4d9a68a98c6b192a4e43bb4f3e93194a7da447464654dafba.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 357, + 288, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 288, + 393 + ], + "type": "text", + "content": "and Oxford Pets [35]. We compared SMP with two strong baselines, LinearStruc. and LinearStruc. w/ FSR, and two prominent methods, VPT (tuning) and Head2Toe (probing)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 395, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 395, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 287, + 479 + ], + "type": "text", + "content": "Experimental results presented in Figure 5 show that SMP consistently outperforms other methods in most cases, demonstrating its effectiveness in few-shot scenario, and its ability to improve performance with increasing training data. Notably, feature sparsity regularization performs worse than SMP, highlighting the importance of leveraging pre-trained model structure as prior knowledge in low-shot scenario." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 491, + 263, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 263, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 263, + 504 + ], + "type": "text", + "content": "5.3. Experiments on More Transfer Scenarios" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 514, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 574 + ], + "type": "text", + "content": "Larger downstream datasets. We provide average results on 9 full-size downstream datasets in Table 2. It can be observed even in the larger data regime, SMP still achieves competitive performance compared with tuning methods. Full results are presented in Suppl. B.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 578, + 288, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 288, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 288, + 649 + ], + "type": "text", + "content": "Different pre-trained models. We present VTAB-1k results on ViT-B/16 and ViT-L/14 pretrained by CLIP in Suppl. B.2. Despite the usage of stronger pre-trained models, SMP continuously outperforms baseline methods due to the incorporation of diverse features and non-linear transformation, as well as its flexible framework." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": "Different architectures. SMP is a versatile method that can be applied to convolutional neural networks as well. In Suppl. B.3, we present the VTAB-1k results obtained from ImageNet pre-trained ResNet-50 models, where SMP outperforms other methods and achieves superior performance." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 306, + 370, + 547, + 404 + ], + "blocks": [ + { + "bbox": [ + 305, + 356, + 545, + 367 + ], + "lines": [ + { + "bbox": [ + 305, + 356, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 305, + 356, + 545, + 367 + ], + "type": "text", + "content": "Table 2. Average accuracy (%) on 9 full-size downstream datasets." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 370, + 547, + 404 + ], + "lines": [ + { + "bbox": [ + 306, + 370, + 547, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 547, + 404 + ], + "type": "table", + "html": "
Fine-tuningVPTLinearSMP
Average Accuracy88.488.781.788.7
", + "image_path": "b028cc30c23ebe2fa7447b605a5e3411ab10a9d73667bb5385e9281e3bc16f10.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 413, + 440, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 413, + 440, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 413, + 440, + 426 + ], + "type": "text", + "content": "5.4. Analysis and Discussion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "text", + "content": "In this section, we conduct various experiments to analyze the components of the proposed SMP method. Full results of this section are presented in Suppl. C." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 472, + 546, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 546, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 546, + 638 + ], + "type": "text", + "content": "Visualization of group norms. We visualize the " + }, + { + "bbox": [ + 304, + 472, + 546, + 638 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 472, + 546, + 638 + ], + "type": "text", + "content": " norm of each structure on different tasks in Figure 6, which helps us to understand the behavior of the proposed structured regularization. For Sun397, an easy adaptation task, we can observe the structures from deep layers have larger norms in Figure 6a. This validates our analysis in Section 3.2, that for easy adaptation tasks, features extracted from deep layers contain more information. For Clevr-Dist, a difficult adaptation task, the structures have similar norms overall, which means for difficult tasks, information contained in intermediate layers is also important, and diverse features can improve the performance, as we find in Section 3.2. This motivates the proposed structured non-linearity regularization, utilizing " + }, + { + "bbox": [ + 304, + 472, + 546, + 638 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 472, + 546, + 638 + ], + "type": "text", + "content": " norm to control the complexity of non-linear model." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "Retraining probing model by gradually incorporating structures. We retrain a probing model by gradually incorporating structures selected based on the magnitude of their " + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": " norm, to validate their contribution to the prediction. Results are presented in Figure 7. For Sun397, an easy adaptation task, we can see in Figure 7a that there are redun" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16855" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 67, + 292, + 156 + ], + "blocks": [ + { + "bbox": [ + 52, + 67, + 292, + 156 + ], + "lines": [ + { + "bbox": [ + 52, + 67, + 292, + 156 + ], + "spans": [ + { + "bbox": [ + 52, + 67, + 292, + 156 + ], + "type": "image", + "image_path": "e99f4b82417c31d3c5087dcde2007334e8fda3dd38a873613a6132ae66d91ab8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 161, + 200, + 171 + ], + "lines": [ + { + "bbox": [ + 141, + 161, + 200, + 171 + ], + "spans": [ + { + "bbox": [ + 141, + 161, + 200, + 171 + ], + "type": "text", + "content": "(a) Sun397 (easy)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 173, + 291, + 262 + ], + "blocks": [ + { + "bbox": [ + 52, + 173, + 291, + 262 + ], + "lines": [ + { + "bbox": [ + 52, + 173, + 291, + 262 + ], + "spans": [ + { + "bbox": [ + 52, + 173, + 291, + 262 + ], + "type": "image", + "image_path": "1fb3d64faad6bf3cad0400800a95c05e9b0fe2efb2efe6b342be0c3a1ad2b55b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 266, + 210, + 277 + ], + "lines": [ + { + "bbox": [ + 131, + 266, + 210, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 266, + 210, + 277 + ], + "type": "text", + "content": "(b) Clevr-Dist (difficult)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 286, + 269, + 297 + ], + "lines": [ + { + "bbox": [ + 64, + 286, + 269, + 297 + ], + "spans": [ + { + "bbox": [ + 64, + 286, + 269, + 297 + ], + "type": "text", + "content": "Figure 6. " + }, + { + "bbox": [ + 64, + 286, + 269, + 297 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 64, + 286, + 269, + 297 + ], + "type": "text", + "content": " norm of structure groups on different tasks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 194, + 547, + 216 + ], + "lines": [ + { + "bbox": [ + 305, + 194, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 305, + 194, + 547, + 216 + ], + "type": "text", + "content": "Figure 7. Retraining performance by gradually incorporating structures with larger " + }, + { + "bbox": [ + 305, + 194, + 547, + 216 + ], + "type": "inline_equation", + "content": "\\ell_{2}" + }, + { + "bbox": [ + 305, + 194, + 547, + 216 + ], + "type": "text", + "content": " norm obtained by SMP." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 310, + 288, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 288, + 430 + ], + "type": "text", + "content": "dant structures. And with a few selected structures, we can achieve comparable results compared with the original SMP, justifying the zero norm of some structured observed after performing structure regularization in Figure 6a. However, for Clevr-Dist, a difficult adaptation task, comparable performance is achieved until selecting 55 structures, suggesting intermediate structures also contribute to the final prediction. This demonstrates the necessity of proposed structured regularizer, which resolves the contradiction between easy and difficult domain adaptation tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 432, + 288, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 288, + 564 + ], + "type": "text", + "content": "Impact of feature aggregation designs. The primary objective of performing token and channel aggregation is to minimize the redundancy in the extracted raw features, while maintaining feature diversity in the compact aggregated representation. To achieve this, two types of aggregation are designed to preserve channel and spatial information, respectively. We provide ablation studies in Suppl. C.5. Overall, channel information (aggregating along tokens) is essential for all datasets, while spatial information (aggregating along channels) is beneficial for tasks that are sensitive to spatial location, e.g., dSpr-Loc and sNORB-Azim." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 567, + 288, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 288, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 288, + 674 + ], + "type": "text", + "content": "Impact of different structures. To increase the feature diversity, we incorporate features extracted from various structures, coupled with a structured sparse regularizer to conduct model fitting and feature selection simultaneously. This flexible framework enables us to avoid the costly manual selection of candidate structures. We provide analysis of structures in Suppl. C.6. The results show that incorporating all candidate structured features yields best mean results, which demonstrates the efficacy of our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Cost of SMP. We compare the cost of SMP with baseline methods, including relative adaptation time, extra inference FLOPs, and extra parameters. The average results on VTAB-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 68, + 425, + 169 + ], + "blocks": [ + { + "bbox": [ + 310, + 68, + 425, + 169 + ], + "lines": [ + { + "bbox": [ + 310, + 68, + 425, + 169 + ], + "spans": [ + { + "bbox": [ + 310, + 68, + 425, + 169 + ], + "type": "image", + "image_path": "c4addc513272d803bc245c903c01bcb20ac3f895ce7323006e5714743781c417.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 175, + 396, + 185 + ], + "lines": [ + { + "bbox": [ + 337, + 175, + 396, + 185 + ], + "spans": [ + { + "bbox": [ + 337, + 175, + 396, + 185 + ], + "type": "text", + "content": "(a) Sun397 (easy)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 429, + 68, + 543, + 169 + ], + "blocks": [ + { + "bbox": [ + 429, + 68, + 543, + 169 + ], + "lines": [ + { + "bbox": [ + 429, + 68, + 543, + 169 + ], + "spans": [ + { + "bbox": [ + 429, + 68, + 543, + 169 + ], + "type": "image", + "image_path": "67152e2b9413ff984d8288c4e3f256e6f520b098e0b19564a1c7c9735b629693.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 175, + 525, + 185 + ], + "lines": [ + { + "bbox": [ + 446, + 175, + 525, + 185 + ], + "spans": [ + { + "bbox": [ + 446, + 175, + 525, + 185 + ], + "type": "text", + "content": "(b) Clevr-Dist (difficult)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 251, + 544, + 312 + ], + "blocks": [ + { + "bbox": [ + 305, + 224, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 305, + 224, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 547, + 246 + ], + "type": "text", + "content": "Table 3. Cost comparison (Average results on VTAB-1k). * indicates relative results compared with Fine-tuning." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 251, + 544, + 312 + ], + "lines": [ + { + "bbox": [ + 307, + 251, + 544, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 544, + 312 + ], + "type": "table", + "html": "
LinearFTVPTSMP
Adaptation time (rel.)*0.0510.970.15
Extra Inference FLOPs (G)0011.010.01
Extra parameters (M)0.0485.830.646.91
", + "image_path": "e757e527962b43def038b34ac3b8d4cde7b8ad9fd8cb10b9f250d99508f9085e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 321, + 547, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 547, + 513 + ], + "type": "text", + "content": "1k are shown in Table 3. SMP only needs " + }, + { + "bbox": [ + 304, + 321, + 547, + 513 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 321, + 547, + 513 + ], + "type": "text", + "content": " adaptation time on average compared with fine-tuning. On the contrary, the adaptation time of VPT is " + }, + { + "bbox": [ + 304, + 321, + 547, + 513 + ], + "type": "inline_equation", + "content": "97\\%" + }, + { + "bbox": [ + 304, + 321, + 547, + 513 + ], + "type": "text", + "content": " of fine-tuning. This results in a significantly lower adaptation cost of SMP compared to tuning methods. SMP is also very efficient during inference time, and only introduces 0.01G inference FLOPs compared with fine-tuning. It can be noticed that VPT introduces extra 11.01G inference FLOPs, which is far more than SMP. SMP requires 6.91M extra parameters on average, which is the " + }, + { + "bbox": [ + 304, + 321, + 547, + 513 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 304, + 321, + 547, + 513 + ], + "type": "text", + "content": " of fine-tuning. Though the requirement of extra parameters is higher than VPT, we can leverage the retraining method proposed in Section 5.4, which can further reduce the extra parameters. Overall, SMP is a highly efficient method with low adaptation cost, fast inference, and fewer extra parameters, while maintaining superior performance compared to tuning-based methods." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 525, + 378, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 378, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 378, + 538 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we present structured model probing, an effective yet efficient probing method for transfer learning. Our investigation reveals that model probing behaves differently for easy and difficult adaptation tasks. To mitigate this conflicting behavior, we propose structured model probing, a method that is able to achieve good performance in both cases. With the proposed structured regularization, we can construct a simple adaptation model for easy adaptation cases, and increase the complexity of the adaptation model for difficult cases. Our method outperforms state-of-the-art tuning methods while maintaining computational efficiency. Our work highlights the potential of model probing, which motivates researchers to better exploit the use large-scale pre-trained model and apply model probing to diverse tasks." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16856" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 145 + ], + "type": "text", + "content": "[1] Alessandro Achille, Michael Lam, Rahul Tewari, Avinash Ravichandran, Subhransu Maji, Charless C. Fowlkes, Stefano Soatto, and Pietro Perona. Task2vec: Task embedding for meta-learning. In IEEE International Conference on Computer Vision (ICCV), pages 6429-6438, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "type": "text", + "content": "[2] Alessandro Achille, Aditya Golatkar, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. LQF: linear quadratic fine-tuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 15729-15739, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 192, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 288, + 245 + ], + "type": "text", + "content": "[3] Armen Aghajanyan, Akshit Shrivastava, Anchit Gupta, Naman Goyal, Luke Zettlemoyer, and Sonal Gupta. Better finetuning by reducing representational collapse. In 9th International Conference on Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "type": "text", + "content": "[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In European Conference on Computer Vision (ECCV), pages 446-461, 2014. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 292, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 288, + 412 + ], + "type": "text", + "content": "[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33 (NeurIPS), 2020. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 413, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 456 + ], + "type": "text", + "content": "[6] Han Cai, Chuang Gan, Ligeng Zhu, and Song Han. Tinytl: Reduce memory, not parameters for efficient on-device learning. In Advances in Neural Information Processing Systems 33 (NeurIPS), 2020. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 458, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 288, + 513 + ], + "type": "text", + "content": "[7] Shoufa Chen, Chongjian GE, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In Advances in Neural Information Processing Systems 35 (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 514, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 288, + 567 + ], + "type": "text", + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 1597-1607, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 569, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 569, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 569, + 288, + 612 + ], + "type": "text", + "content": "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In IEEE International Conference on Computer Vision (ICCV), pages 9620-9629, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "text", + "content": "[10] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4109-4118, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "type": "text", + "content": "[11] Mostafa Dehghani, Yi Tay, Anurag Arnab, Lucas Beyer, and Ashish Vaswani. The efficiency misnomer. In 10th International Conference on Learning Representations (ICLR), 2022. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "text", + "content": "[13] Aditya Deshpande, Alessandro Achille, Avinash Ravichandran, Hao Li, Luca Zancato, Charless C. Fowlkes, Rahul Bhotika, Stefano Soatto, and Pietro Perona. A linearized framework and a new benchmark for model selection for fine-tuning. arXiv:2102.00084, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 173, + 547, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 547, + 237 + ], + "type": "text", + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), pages 4171–4186, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 239, + 547, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 547, + 315 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations (ICLR), 2021. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 316, + 547, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 547, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 547, + 371 + ], + "type": "text", + "content": "[16] Utku Evci, Vincent Dumoulin, Hugo Larochelle, and Michael C. Mozer. Head2Toe: Utilizing intermediate representations for better transfer learning. In Proceedings of the 39th International Conference on Machine Learning (ICML), pages 6009-6033, 2022. 1, 3, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 372, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 547, + 415 + ], + "type": "text", + "content": "[17] Weifeng Ge and Yizhou Yu. Borrowing treasures from the wealthy: Deep transfer learning through selective joint finetuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10-19, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 416, + 547, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 547, + 470 + ], + "type": "text", + "content": "[18] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogério Schmidt Feris. SpotTune: Transfer learning through adaptive fine-tuning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4805-4814, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 471, + 547, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 547, + 515 + ], + "type": "text", + "content": "[19] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In 10th International Conference on Learning Representations (ICLR), 2022. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 515, + 547, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 547, + 568 + ], + "type": "text", + "content": "[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 570, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 547, + 624 + ], + "type": "text", + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, 2022. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 625, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 547, + 689 + ], + "type": "text", + "content": "[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning (ICML), pages 2790-2799, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "text", + "content": "[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge J. Belongie, Bharath Hariharan, and Ser-Nam Lim." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16857" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "text", + "content": "[24] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition (CVPR Workshop), 2011. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "text", + "content": "[25] Simon Kornblith, Jonathon Shlens, and Quoc V. Le. Do better imagenet models transfer better? In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2661-2671, 2019. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 241 + ], + "type": "text", + "content": "[26] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE International Conference on Computer Vision Workshop (ICCV Workshop), pages 554–561, 2013. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 242, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 242, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 242, + 288, + 297 + ], + "type": "text", + "content": "[27] Ananya Kumar, Aditi Raghunathan, Robbie Matthew Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In 10th International Conference on Learning Representations (ICLR), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "type": "text", + "content": "[28] Ismael Lemhadri, Feng Ruan, Louis Abraham, and Robert Tibshirani. Lassonet: A neural network with feature sparsity. Journal of Machine Learning Research, 22:127:1-127:29, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 288, + 387 + ], + "type": "text", + "content": "[29] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3045-3059, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 288, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 432 + ], + "type": "text", + "content": "[30] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), pages 4582-4597, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 288, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 477 + ], + "type": "text", + "content": "[31] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5196-5205, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 478, + 288, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 288, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 288, + 511 + ], + "type": "text", + "content": "[32] Fangzhou Mu, Yingyu Liang, and Yin Li. Gradients as features for deep representation learning. In 8th International Conference on Learning Representations (ICLR), 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 512, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 288, + 567 + ], + "type": "text", + "content": "[33] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. LEEP: A new measure to evaluate transferability of learned representations. In Proceedings of the 37th International Conference on Machine Learning (ICML), pages 7294-7305, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 601 + ], + "type": "text", + "content": "[34] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345-1359, 2010. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 645 + ], + "type": "text", + "content": "[35] Omkar M. Parkhi, Andreae Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3498-3505, 2012. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "text", + "content": "[36] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), pages 2227-2237, 2018. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 687 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 149 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 151, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 547, + 194 + ], + "type": "text", + "content": "[38] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. LST: Ladder side-tuning for parameter and memory efficient transfer learning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 195, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 547, + 239 + ], + "type": "text", + "content": "[39] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical report, California Institute of Technology, 2011. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 240, + 547, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 547, + 294 + ], + "type": "text", + "content": "[40] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In 10th International Conference on Learning Representations (ICLR), 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 296, + 547, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 296, + 547, + 351 + ], + "spans": [ + { + "bbox": [ + 307, + 296, + 547, + 351 + ], + "type": "text", + "content": "[41] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer G. Dy, and Tomas Pfister. Learning to prompt for continual learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 139-149, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 352, + 547, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 352, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 352, + 547, + 384 + ], + "type": "text", + "content": "[42] Xiaohan Yan and Jacob Bien. Hierarchical Sparse Modeling: A choice of two group lasso formulations. Statistical Science, 32(4):531 - 560. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 386, + 547, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 386, + 547, + 429 + ], + "spans": [ + { + "bbox": [ + 307, + 386, + 547, + 429 + ], + "type": "text", + "content": "[43] Ming Yuan and Yi Lin. Model selection and estimation in regression with grouped variables. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 68(1): 49-67, 2006. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 430, + 547, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 430, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 430, + 547, + 495 + ], + "type": "text", + "content": "[44] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv:1910.04867, 2019. 1, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 498, + 547, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 547, + 552 + ], + "type": "text", + "content": "[45] Jeffrey O. Zhang, Alexander Sax, Amir Zamir, Leonidas J. Guibas, and Jitendra Malik. Side-tuning: A baseline for network adaptation via additive side networks. In European Conference on Computer Vision (ECCV), pages 698-714, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 553, + 547, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 553, + 547, + 597 + ], + "spans": [ + { + "bbox": [ + 307, + 553, + 547, + 597 + ], + "type": "text", + "content": "[46] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 16795-16804, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 598, + 547, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 598, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 307, + 598, + 547, + 641 + ], + "type": "text", + "content": "[47] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 643, + 547, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 643, + 547, + 687 + ], + "spans": [ + { + "bbox": [ + 307, + 643, + 547, + 687 + ], + "type": "text", + "content": "[48] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2021. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16858" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/StyLitGAN_ Image-Based Relighting via Latent Control/37a6c7df-f82e-4da6-aa4d-10d6b8e2b178_content_list.json b/2024/StyLitGAN_ Image-Based Relighting via Latent Control/37a6c7df-f82e-4da6-aa4d-10d6b8e2b178_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d6ac0e29a9457578c4a12730bfc00ab12b88748a --- /dev/null +++ b/2024/StyLitGAN_ Image-Based Relighting via Latent Control/37a6c7df-f82e-4da6-aa4d-10d6b8e2b178_content_list.json @@ -0,0 +1,1348 @@ +[ + { + "type": "text", + "text": "StyLitGAN: Image-based Relighting via Latent Control", + "text_level": 1, + "bbox": [ + 202, + 130, + 767, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anand Bhattad James Soole D.A. Forsyth", + "bbox": [ + 274, + 181, + 694, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Illinois Urbana-Champaign", + "bbox": [ + 321, + 199, + 648, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://anandbhattad.github.io/stylitgan/", + "bbox": [ + 241, + 217, + 725, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generated Image", + "bbox": [ + 133, + 255, + 227, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Relit-1 $(\\mathbf{w}^{+} + d_{1})$ Relit-2 $(\\mathbf{w}^{+} + d_{2})$ Relit-3 $(\\mathbf{w}^{+} + d_{3})$ Relit-4 $(\\mathbf{w}^{+} + d_{4})$ Relit-5 $(\\mathbf{w}^{+} + d_{5})$", + "bbox": [ + 245, + 255, + 841, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca50ceef1b4c4acb1b024de96b1a74ca80e39eb45597d829b2b3de7949c1e4ba.jpg", + "image_caption": [ + "Figure 1. StyLitGAN identifies directional vectors $(d_i)$ within StyleGAN's style space $(\\mathcal{W}^{+})$ which, when added to the $\\mathbf{w}^{+}$ style code, effectively modify the lighting of generated images while preserving their geometry and albedo. This process eliminates the need for per-image search or model fine-tuning. The first column displays images generated from StyleGAN2; subsequent columns illustrate the same scene, each relit using a specific direction. These relighting directions $(d_i)$ are derived through a forward selection method, ensuring diversity and avoiding cherry-picking. The directional effects are consistent across different scenes: for instance, $d_1$ activates an orange-tinged bedside lamp, $d_2$ a less intense white-tinged lamp, $d_3$ introduces strong directional light from the window, and so on, demonstrating diverse relighting capabilities of StyLitGAN." + ], + "image_footnote": [], + "bbox": [ + 119, + 267, + 846, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 661, + 313, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We describe a novel method, StyLitGAN, for relighting and resurfacing images in the absence of labeled data. StyLitGAN generates images with realistic lighting effects, including cast shadows, soft shadows, inter-reflections, and glossy effects, without the need for paired or CGI data. StyLitGAN uses an intrinsic image method to decompose an image, followed by a search of the latent space of a pretrained StyleGAN to identify a set of directions. By prompting the model to fix one component (e.g., albedo) and vary another (e.g., shading), we generate relighted images by adding the identified directions to the latent style codes. Quantitative metrics of change in albedo and lighting diversity allow us to choose effective directions using a forward selection process. Qualitative evaluation confirms the effectiveness of our method.", + "bbox": [ + 73, + 688, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 662, + 630, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene appearance shifts dramatically with varying lighting conditions - a sunlit room takes on a different character as daylight fades, and interior spaces transform with the flick of a switch. Similarly, surface changes, like a wall's paint color, change not only the wall's appearance but also the overall image due to light reflection. Despite the impressive realism achieved by current generative models like StyleGAN [22-24], they fall short in dynamically controlling scene lighting, a key aspect of realistic image generation.", + "bbox": [ + 496, + 688, + 893, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we present StyLitGAN, a novel approach that extends the editing capabilities of StyleGAN [38, 45, 46, 53]. StyLitGAN uniquely manipulates style codes to selectively change lighting while preserving other image attributes like albedo and geometry. This selective editing", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4231", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "addresses a critical gap in current generative methods, which typically lack the precision to control individual scene components independently.", + "bbox": [ + 75, + 90, + 470, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method, StyLitGAN, first uses StyleGAN to produce a set of images and then decomposes these generated images into albedo, diffuse shading, and glossy effects using an off-the-shelf, self-supervised network [14]. We then search for style code edits by prompting StyleGAN to produce images that (a) are diverse, but (b) have the same albedo (and so geometry and material) as the original generated images. Our search selects the most effective relighting directions in a data-driven manner.", + "bbox": [ + 75, + 136, + 468, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach generates images with realistic lighting effects, including cast shadows, soft shadows, inter-reflections, and glossy effects. Importantly, we observe that style code edits produce consistent effects across images. For instance, as seen in Fig. 1, adding the first and second directions tends to switch on bedside lamps (columns Relit-1 and Relit-2), while adding the fourth direction increases the light intensity from outside the window (column Relit-4). Since StyLitGAN can generate any image that a vanilla StyleGAN can, it also generate images that are out of distribution, one would expect FID scores to increase over StyleGAN; this happens.", + "bbox": [ + 75, + 272, + 470, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The recent method GAN-control [39] controls lighting on face images using an attribute procedure. We show, in contrast to StyLitGAN, GAN-control fails on indoor scenes, likely because the attribute vocabulary is too easily subverted by the complex lighting effects in indoor scenes.", + "bbox": [ + 75, + 438, + 468, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate applications of StyLitGAN to standard vision problems. Using the Multi-Ilum dataset of [29], we show that predictions from a SOTA surface normal predictor [21] vary significantly when lighting is changed. Finetuning this normal predictor using StyLitGAN images suppresses this effect. The improvement is comparable with that obtained by finetuning with true multi-illuminant images (which are very difficult to obtain in quantity).", + "bbox": [ + 75, + 513, + 470, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 648, + 218, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Manipulation: A significant literature deals with manipulating and editing images [3, 10, 11, 15, 17, 27, 33, 35, 44, 54]. Editing procedures for generative image models [16] are important, because they demand compact image representations with useful, disentangled interpretations. StyleGAN [22-24] is currently de facto state-of-the-art for editing generated images, likely because its mapping of initial noise vectors to style codes which control entire feature layers produces latent spaces that are heavily disentangled and so easy to manipulate. Recent editing methods include [8, 36, 38, 45, 46, 53], with a survey in [47]. The architecture can be adapted to incorporate spatial priors for authoring novel and edited images [13, 28, 43]. In contrast to this literature, we show how to fix one physically meaningful image factor while changing another. Doing so is difficult", + "bbox": [ + 75, + 672, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "because the latent spaces are not perfectly disentangled, and we must produce a diverse set of changes in only one factor. Relighting using StyleGAN: Relighting faces using StyleGAN can be achieved with Stylerig [43], but this method requires a 3D morphable face model. In contrast, StyLitGAN does not require a 3D model and can be extended to complex indoor scenes, which is not possible with Stylerig. Yang et al. [48] uses semantic label attributes to train a binary classifier to find latent space directions that represent indoor and natural lighting, but this method cannot produce diverse relighting effects. We also find Yang et al's relighting to change color or albedo. In contrast, StyLitGAN generates diverse realistic relighting effects without changing albedo and without requiring any labeled attributes.", + "bbox": [ + 496, + 90, + 893, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "StyleFlow [1] and GAN-control [39] require a parametric model to express lighting, such as spherical harmonics. These methods are limited to relighting faces and do not result in realistic relighting of rooms. Our experiments using GAN-control for rooms result in large geometry and albedo change. In contrast, StyLitGAN can produce relighted images without changing geometry or albedo. We also note that rooms are more challenging to relight than faces due to significant long-scale inter-reflection effects, diverse shadow patterns, stylized luminaires, stylized surface albedos, and surface brightnesses that are not a function of surface normal alone. These factors make it difficult to apply GAN control directly to rooms. Also, none of these methods can resurface or recolor rooms, though StyLitGAN can also edit color or materials while preserving the scene's lighting.", + "bbox": [ + 496, + 301, + 893, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other Face Relighting methods use carefully collected supervisory data from light-stages or parametric spherical harmonics [30, 32, 37, 41, 52]. ShadeGAN [31], Rendering with Style [6], and Volux-GAN [42] use a volumetric rendering approach to learn the 3D structure of the face and the illumination encoding. Volux-GAN [42] also requires image decomposition from [32] that is trained using carefully curated light-stage data. In comparison, we neither require any explicit 3D modeling of the scene nor labeled and curated data for training the image decomposition model.", + "bbox": [ + 496, + 529, + 893, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 500, + 694, + 607, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We follow convention and manipulate StyleGAN [24] by adjusting the $\\mathbf{w}^{+}$ latent variables. We do not modify StyleGAN weights, but instead, seek a set of lighting directions $\\mathbf{d}_i$ (same shape as $\\mathbf{w}^{+}$ ) which are independent of $\\mathbf{w}^{+}$ and have desired effects on the generated image. We obtain these directions by constructing losses that capture the desired outcomes, and then search for directions that minimize these losses. We find all directions only once and use 2000 randomly generated images for this search. Once found, these lighting directions apply to all other generated images. Our search procedure only sees each image once.", + "bbox": [ + 496, + 719, + 893, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fig. 2 summarizes our procedure and we call our model", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4232", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ef00eae8814e93b30e2a78df87c02ad12b5ad1ac95088f9a84c9e6b3b31bbaa6.jpg", + "image_caption": [ + "Figure 2. How StyLitGAN works: We generate an image from random Gaussian noise using a pretrained StyleGAN. We also generate novel relighted versions (16 in our case) of the same image using randomly initialized latent directions (d) that are added to $\\mathbf{w}^{+}$ latent style codes. We train a classifier (F) that takes in all the pairs of relighted and original images and predicts the relighting direction applied to them. We apply a distinction loss and jointly update the latent directions and the classifier. Next, we generate the decomposition of these images from a pretrained decomposition model (D). We then apply losses that force StyLitGAN to find latent directions such that the albedo does not change (consistency loss), but the image does (diversity loss)." + ], + "image_footnote": [], + "bbox": [ + 117, + 87, + 844, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "StyLitGAN. Our method consists of two stages. The first stage involves decomposing images using a pretrained model. The second stage jointly searches for directions and trains a classifier $\\mathrm{F}$ . The classifier $\\mathrm{F}$ predicts the latent direction applied to image pairs. It's a classification task with a fixed number of jointly learned latent directions. While classifier $\\mathrm{F}$ doesn't directly know whether latent directions relate to lighting, our use of image decomposition losses ensures these directions are lighting-related. The consistency loss maintains albedo, and the diversity loss ensures diverse shading changes, both properties expected when changing lighting conditions. Thus, while jointly updating the classifier $\\mathrm{F}$ and latent directions, these losses ensure the discovered directions pertain to relighting. We now elaborate on our search for directions and losses in detail.", + "bbox": [ + 75, + 349, + 472, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Base StyleGAN Models: We use baseline pretrained models from [49] that use a dual-contrastive loss to train StyleGAN for bedrooms, faces, and churches. We also use baseline pretrained StyleGAN2 models from [13] for conference rooms, kitchens, dining rooms, and living rooms.", + "bbox": [ + 75, + 579, + 470, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Decomposition: We decompose images into albedo, shading, and gloss maps (gloss only when available) as $A \\times S + G$ , where $A$ models albedo effects and $S$ and $G$ model shading and gloss effects respectively. We use the method of Forsyth and Rock [14] and its variant from Bhattad and Forsyth [3], which is easily adapted because it is self-supervised and uses only samples from statistical models derived from Land's Retinex theory [25]. By changing the statistical spatial model parameters, we can construct many decompositions using their approach. We evaluate many such decomposition models under several hyperparameter settings and create a large pool of relighting directions. We finalize our directions using a forward selection process that provides minimal albedo and geometry shift with a large relighting diversity (Section 4).", + "bbox": [ + 75, + 656, + 470, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Relighting a scene should produce a new, realistic image where the shading has changed but the albedo has", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "not. Write $I(\\mathbf{w}^{+})$ for the image produced by StyleGAN given style codes $\\mathbf{w}^{+}$ , and $A(I)$ , $S(I)$ , and $G(I)$ for the albedo, shading, and gloss respectively recovered from image $I$ . We search for multiple directions $\\mathbf{d}_i$ such that: (a) $A(I(\\mathbf{w}^{+} + \\mathbf{d}_i))$ is very close to $A(I(\\mathbf{w}^{+}))$ – so the image is a relighted version of $I(\\mathbf{w}^{+})$ , a property we call persistent consistency; (b) the images produced by the different directions are linearly independent – relighting diversity; (c) which direction was used can be determined from the image, so that different directions have visibly distinct effects – distinctive relighting; and (d) the new shading field (map) is not strongly correlated to the albedo – independent relighting. Not every shading field can be paired with a given albedo, otherwise there would be nothing to do. We assume that edited $\\mathbf{w}^{+}$ will result in realistic images [7].", + "bbox": [ + 496, + 349, + 893, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recoloring: Alternatively, we may wish to edit scenes where the colors or materials of objects have changed, but the lighting hasn't. Because shading conveys a great deal of information about shape, we can find these edits using modified losses by seeking consistency in the shading field.", + "bbox": [ + 496, + 577, + 893, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Persistent Consistency: The albedo decomposition of both the relighted scene: $A_{R} = A(I(\\mathbf{w}^{+} + \\mathbf{d}_{i}))$ and the original: $A_{O} = A(I(\\mathbf{w}^{+}))$ must be the same; where R refers to relighted images and O refers to StyleGAN generated images. We use a Huber loss and a perceptual feature loss [20, 51] from a VGG feature extractor (Φ) [40] at various feature layers (j) to preserve persistent effects (geometry, appearance, and texture) in the scene.", + "bbox": [ + 496, + 657, + 893, + 777 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n s t}} \\left(A _ {O}, A _ {R}\\right) = \\left\\{ \\begin{array}{c} \\frac {1}{2} \\left[ A _ {O} - A _ {R} \\right] ^ {2} \\quad \\text {f o r} | A _ {O} - A _ {R} | \\leq \\delta , \\\\ \\delta \\left(| A _ {O} - A _ {R} | - \\delta / 2\\right) \\quad \\text {o t h e r w i s e .} \\end{array} \\right. (1) \\\\ \\mathcal {L} _ {p e r} \\left(A _ {O}, A _ {R}\\right) = \\left| \\left| \\Phi_ {j} \\left(A _ {O}\\right) - \\Phi_ {j} \\left(A _ {R}\\right) \\right| \\right| _ {2}. (2) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 787, + 890, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Relighting Diversity: We want the set of relighted images produced by the directions to be diverse on a long scale so that regions that were in shadow in one image might be", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4233", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "bright in another. For each $S(\\mathbf{w}^{+} + \\mathbf{d}_{i})$ , we stack the two shading and gloss: $S$ and $G$ , and compute a smoothed and downsampled vector $\\mathbf{t}_{i}$ from these maps. We then compute $\\mathcal{L}_{div}(S,G)$ (diversity loss) which compels these $\\mathbf{t}_{i}$ to be linearly independent and encourages diversity in relighting.", + "bbox": [ + 76, + 90, + 468, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i v}} (S, G) = - \\log \\det N \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 176, + 468, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i^{th}$ & $j^{th}$ component of $\\mathbf{N}$ is $t_i^\\top t_j$", + "bbox": [ + 96, + 200, + 370, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distinctive Relighting: A network might try to cheat by making minimal changes to the image. Directions $\\mathbf{d}_i$ should have the property that $\\mathbf{d}_i$ is easy to impute from $I(\\mathbf{w}^{+} + \\mathbf{d}_{i})$ . Joint with the search for directions, we train a classifier to categorize the applied direction. This classifier accepts $I(\\mathbf{w}^{+})$ and $I(\\mathbf{w}^{+} + \\mathbf{d}_{i})$ and must predict $i$ . Its cross-entropy supplies our loss:", + "bbox": [ + 76, + 220, + 468, + 328 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {l, F} \\mathcal {L} _ {\\text {d i s t}} (I (\\mathbf {w} ^ {+}), I (\\mathbf {w} ^ {+} + d _ {i})) \\\\ = - \\sum_ {i = 1} ^ {M} y _ {i} \\log F \\left(I \\left(\\mathbf {w} ^ {+}\\right), I \\left(\\mathbf {w} ^ {+} + d _ {i}\\right)\\right) \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 337, + 468, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Saturation Penalty: Our diversity loss might cheat and obtain high diversity by generating blocks of over-saturated or under-saturated pixels. So, we apply a saturation penalty over several pixels within a certain threshold.", + "bbox": [ + 76, + 404, + 468, + 465 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {s a t} = \\lambda_ {\\text {o v e r s a t}} \\left[ \\frac {1}{H \\times W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\max (0, I _ {i, j} - s) ^ {2} \\right] \\tag {5} \\\\ + \\lambda_ {\\text {u n d e r s a t}} \\left[ \\frac {1}{H \\times W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\max (0, s - I _ {i, j}) ^ {2} \\right] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 476, + 468, + 563 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{oversat}$ and $\\lambda_{undersat}$ are the penalty weights for over-saturation and under-saturation respectively, $H$ and $W$ are the height and width of the images, $I_{i,j}$ is the pixel intensity at pixel location $(i,j)$ , and $s$ is the saturation threshold (i.e., the maximum allowed pixel intensity). The penalty is computed as the mean squared difference between the pixel intensity and the saturation threshold.", + "bbox": [ + 75, + 575, + 468, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recoloring requires swapping albedo and shading components in all losses, except we do not use decorrelation loss while recoloring. Obtaining good results requires quite a careful choice of loss weights ( $\\lambda$ coefficients). We experiment with several $\\lambda$ coefficients for both these edits (Section 4 and Supplementary).", + "bbox": [ + 75, + 680, + 470, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Model and Directions Selection", + "text_level": 1, + "bbox": [ + 76, + 784, + 362, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We prompt StyleGAN to find style code directions that: (a) do not change the albedo, and (b) strongly change the image. We use many image decomposition models to obtain directions across multiple hyperparameter settings. We have no particular reason to believe that a single model will give only good directions or all good directions. We then find", + "bbox": [ + 75, + 809, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a subset of admissible models. We must choose admissible models using a plot of albedo change versus diversity because there is no way to weigh these effects against one another. However, relatively few methods are admissible – see Figure 3. We then pool all directions from all of the admissible models and use forward selection to find a small set of polished directions in this pool.", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scoring Albedo Change: We use SuperPoints [19] to find 100 interest points in the original StyleGAN-generated image. Around each interest point, we form a $8 \\times 8$ patch. We then compare these patches with patches in the same locations for multiple different relightings of that image. If the albedo in the image does not change, then each patch will have the same albedo but different lighting.", + "bbox": [ + 496, + 204, + 893, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given two color image patches $\\mathbf{p}$ and $\\mathbf{q}$ , viewed under different lights, we must measure the difference between their albedos $d_{a}(\\mathbf{p},\\mathbf{q})$ . Write $p_{ij}$ for the RGB vector at the $i$ , $j$ 'th location ( $1 \\leq i \\leq M$ , $1 \\leq j \\leq N$ ) and write $p_{ij,k}$ for the $k$ 'th RGB component at that location. The intensity of the light may change without the albedo changing, so this problem is homogeneous (i.e. for $\\lambda, \\mu > 0$ , $d_{a}(\\mathbf{p},\\mathbf{q}) = d_{a}(\\lambda \\mathbf{p},\\mu \\mathbf{q}))$ . Assume that the illumination intensity changes, but not color. The patches are small, so the illumination field on a patch can be modeled as a linear function, so there are albedos $\\mathbf{a}$ , $\\mathbf{b}$ such that $p_{ij} = (p_{x}i + p_{y}j + p_{c})a_{ij}$ and $q_{ij} = (q_{x}i + q_{y}j + q_{c})b_{ij}$ . If the two patches have similar albedo, there will be $p_{x}$ etc. such that $p_{ij}' = (q_{x}i + q_{y}j + q_{c})p_{ij}$ is the same as $q_{ij}' = (p_{x}i + p_{y}j + p_{c})q_{ij}$ . We measure the cosine distance", + "bbox": [ + 496, + 311, + 893, + 539 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {a} (\\mathbf {p}, \\mathbf {q}) = 1 - \\max _ {p _ {x}, \\dots q _ {c}} \\frac {\\sum_ {i j k} p _ {i j k} ^ {\\prime} q _ {i j k} ^ {\\prime}}{\\sqrt {\\sum_ {i j k} \\left(p _ {i j k} ^ {\\prime}\\right) ^ {2}} \\sqrt {\\sum_ {i j k} \\left(p _ {i j k} ^ {\\prime}\\right) ^ {2}}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 551, + 890, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The relevant maximum can be calculated by analogy with canonical correlation analysis (Supplementary).", + "bbox": [ + 498, + 608, + 890, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scoring Lighting Diversity: Illumination cone theory [2] yields that any non-negative linear combination of $k$ shadings is a physically plausible shading. To determine if an image is new, we relax the non-negativity constraint and so must ensure that it cannot be expressed as a linear combination of existing images. In turn, we seek a measure of the linear independence of a set of images. This measure should: be large when there is a strong linear dependency; and not grow too fast when the images are scaled. Write $\\mathbf{x}_i$ for the $i$ 'th image, and $\\mathcal{X}$ for the matrix whose $i$ , $j$ 'th component is $\\mathbf{x}_i\\mathbf{x}_j$ . Then $-\\log \\det \\mathcal{X}$ is very large when the $\\mathbf{x}_i$ is close to linearly dependent, but does not scale too fast when the images are scaled.", + "bbox": [ + 496, + 648, + 890, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decomposition Models Investigated: We searched 25 instances in total obtained with different hyperparameter settings from three families of decomposition. The first family", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4234", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b5252d7045a48fc04cdc2dbabcad7e250e0ce1d2b8033904dcbac07430850133.jpg", + "image_caption": [ + "Figure 3. Each model from StyLitGAN produces 16 directions. Models differ by choice of hyperparameters and intrinsic image decomposition. We evaluate models by albedo change and by diversity, averaged across a small fixed validation set of test scenes. As the figure shows, there is typically a payoff, but some models are not admissible. Figure 4 shows examples from some of the models considered here. We exclude inadmissible models, then pool all directions from all other models, and apply a forward selection procedure (section 4). This yields 16 strong relighting directions (the gold star). In our comparison with Yang et al. [48], we focus on the changes in albedo since their method identifies only a single relighting direction. This limits the scope for evaluating relighting diversity. Yang et al. [48] aggressively change scene albedo, while our StyLitGAN ensures only lighting changes. Additionally, we compare with GAN-control [39], which, while attempting relighting, often changes the scene layout, leading to large albedo change and increased diversity score due to layout variations." + ], + "image_footnote": [], + "bbox": [ + 120, + 90, + 421, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is the SOTA unsupervised model of [14], which decomposes images into albedo and shading using example images drawn from statistical models. The second is a variant of that family that decomposes into albedo, shading, and gloss decomposition [3]. The third is an albedo, shading, and gloss decomposition that models fine edges in the albedo rather than the shading field. These models were chosen to represent a range of possible decompositions, but others could yield better results. The key point is that we can choose a model from a collection by a rational process.", + "bbox": [ + 75, + 571, + 470, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Selecting Directions: Our approach for selecting directions involves creating a scatter plot of 25 instances with various hyperparameters and image decomposition models. We find 16 directions for each instance in our final experiments, and the search for 16 directions takes about 14 minutes on an A40 GPU. We experimented with different numbers of directions of order $2^{n}$ for $n = 2, 3, 4, 5, 6, 7$ and found that 16 directions $(n = 4)$ strike a better balance between relighting diversity and albedo change. However, finding multiple directions is challenging because the search space is complex and high-dimensional, and we lack ground truth to supervise", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7f5e3de54231b4937aaace146c50da9cea0c22c9600791c8d69cdae838a7e777.jpg", + "image_caption": [ + "Figure 4. The bottom row shows scene relightings obtained using our final, forward selected, set of directions (star of Figure 3). For comparison, we also show scene relightings from different models obtained from StyLitGAN shown in Figure 3 (model numbers correspond to numbers on that figure). Note how most models are capable of producing some good directions, but not all directions from a given model may be good." + ], + "image_footnote": [], + "bbox": [ + 509, + 88, + 879, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the search. Therefore, we apply a two-step process to find effective directions and filter out any bad directions.", + "bbox": [ + 496, + 372, + 890, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first identify and discard inadmissible models that are located behind the Pareto frontier. We then select the top 10 admissible models based on their average albedo change when applied over a large set of fixed validation images. Our goal is to select the best relighting directions from these admissible models. To achieve this, a forward selection process is employed, which involves selecting a subset of directions from the set of admissible directions.", + "bbox": [ + 496, + 402, + 892, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Forward Selection Process: To select the best directions, we begin with all directions from the admissible models, resulting in 160 directions from 10 models. These directions are then filtered to remove \"bad\" directions that produce relighting similar to the original image or shading that does not vary across pixels, resulting in 108 directions.", + "bbox": [ + 496, + 529, + 893, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we use a greedy process to select the best 16 directions from the remaining 108 directions. We evaluate each direction one at a time and add it to the pool if it provides a large diversity score while incurring a small penalty for large albedo change. This process continues until the desired number of directions are selected. The forward selection process is fast and efficient, taking less than a minute.", + "bbox": [ + 496, + 619, + 893, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The resulting scores from the forward select 16 directions are marked with a star in golden color in Fig. 3. The directions obtained with this process are significantly better than individual models alone. A qualitative ablation is in Fig. 4.", + "bbox": [ + 496, + 726, + 893, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 800, + 632, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative Evaluation: StyLitGAN produces realistic images that are out of distribution but known to exist for straightforward physical reasons. Because they're out of distribution, current quantitative evaluation tools do not apply. We evaluate realism qualitatively. Further, there is no direct", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4235", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7166d8d77f767b023bc10ba80437020a9bbcd67714d16b616b5c6ea92a4fd88f.jpg", + "image_caption": [ + "Generated Image", + "Relit-1 $(+d_{1})$", + "Relit-2 $(+d_{2})$", + "Relit-3 $(+d_{3})$", + "Relit-4 $(+d_{4})$", + "Relit-5 $(+d_{5})$", + "Relit-6 $(+d_{6})$", + "Relit-7 $(+d_{7})$", + "Figure 5. First column images generated by the original StyleGAN. Other columns show images obtained from $\\mathbf{w}^{+} + \\mathbf{d}_{i}$ , our relighting directions added to the style codes $(\\mathbf{w}^{+})$ of the images in the first column. These directions have been chosen to fix albedo, but change shading. Note: each row shows the same scenes but with different illumination. Lighting varies aggressively, and the individual latent direction $\\mathbf{d}_{i}$ has persistent semantics - each column corresponds to a type of illumination. For example, the second $(d_{1})$ and third $(d_{2})$ column switches on bedside lamps. It is worth noting the presence of soft shadows, cast shadows, inter-reflections, and glossy changes across relights. Another important observation to note is that all relightings are with respect to world coordinates, not camera coordinates." + ], + "image_footnote": [], + "bbox": [ + 80, + 99, + 893, + 489 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d1bc21dc64fd1c6702e3e66077b3aaa869db1ebc164fc3a4c97c19c1a94e7b06.jpg", + "image_caption": [ + "Figure 6. Resurfacing Generated Images. Instead of relighting images, we can generate resurfaced images by swapping our consistency and diversity loss. We apply diversity loss to change the albedo and consistency loss to maintain the shading and global illumination. The first column shows images generated by the original StyleGAN, and the other columns show images obtained from $\\mathbf{w}^{+} + \\mathbf{d}_{i}$ for our resurfacing or recoloring directions. Each column shows the same scene as in the first column, but with varying surface colors and materials, while the individual latent direction $\\mathbf{d}_i$ retains its semantics." + ], + "image_footnote": [], + "bbox": [ + 94, + 580, + 870, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "comparable method. However, we show relighting comparisons to a recent SOTA method that is physically motivated and trained with CGI data [26] by using a SOTA inversion method [5] (Fig. 14 in Supplementary). For relighting, our", + "bbox": [ + 75, + 829, + 472, + 891 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "method should generate images that: are clearly relightings of a scene; fix geometry and albedo but visibly change shading; and display complicated illumination effects, including soft shadows, cast shadows and gloss. For resurfacing, our", + "bbox": [ + 498, + 829, + 893, + 891 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4236", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2fd4b9a12e44e0c1910ee4e27fdfba0f92dd8940b5ac0a140ab0b22371286c64.jpg", + "image_caption": [ + "Figure 7. Controllability: The first and last columns of the figure show relighted images generated using our Relit-1 and Relit-5 directions. The bottom section of the figure features a user-controllable slider that enables adjusting the weight of the relighting effects produced by these two directions. Moving the slider from left to right results in a seamless interpolation between the two lighting directions and provides precise control over the relighting of the generated images." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 870, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea7eef6570ef1ec9bd1a3fc1856d73f98a7de1090b9b33827b3fb51fa5f55fae.jpg", + "image_caption": [ + "Figure 8. Scaling Directions: The figure depicts the persistent and smooth effects of applying the direction at different scalar coefficients. We use our Relit-2 direction. A slider at the bottom allows the user to adjust the weight of the relighting direction, producing a seamless interpolation when increasing or decreasing the intensity of the chosen direction. The relighting effects range from a well-lit room with the bedside lamp off to weak external lighting with a bedside lamp on." + ], + "image_footnote": [], + "bbox": [ + 94, + 319, + 875, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "method should generate images that: are clearly images of the original layout, but with different materials or colors or changes in furniture; and display illumination effects that are consistent with these color changes. As Figure 6 shows, our method meets these goals. Figure 7 and Figure 8 show interpolation sequences for a relighting between two directions and scaling only one direction. Note that the lighting changes smoothly, as one would expect. Figure 15 in Supplementary shows that our relighting and recoloring directions are largely disentangled. An ablation for individual losses is also given in Supplementary Figure 13.", + "bbox": [ + 75, + 560, + 472, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Evaluation: Figure 3 shows how we can evaluate albedo and lighting shifts. In Table 1 we show we can generate image datasets with increased FID [18, 34] (clean-FID) from the base comparison set. This is strong evidence our method can produce a set of images that is a strict superset of those that the vanilla StyleGAN can produce.", + "bbox": [ + 75, + 727, + 470, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generality: We have applied our method to StyleGAN trained on Conference Room, Kitchen, Living Room, Dining Room, Church, and Face datasets (results in Figure 10).", + "bbox": [ + 75, + 816, + 470, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison to GAN-control: GAN-control (GC) [39] represents lighting with a spherical harmonic predictor pre", + "bbox": [ + 75, + 869, + 472, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/53cbc0b29954b52114849d5ddae759d90d3ec9e264e16a1f6efdee545ce62eb0.jpg", + "table_caption": [ + "Table 1. FID measures distribution shift and not realism. Our generated images are realistic and are out-of-distribution because of large illumination and color changes in the images. This results in large FID scores. $KDL$ in the table is for kitchen, dining and living room which are jointly trained [13]." + ], + "table_footnote": [], + "table_body": "
TypeBedroomKDLConferenceChurchFaces
StyleGAN (SG)5.015.869.353.805.02
SG + Relighting (RL)14.236.8710.4812.1237.87
SG + Resurfacing (RS)17.039.4110.6318.6034.06
SG + RL + RS21.3911.6812.7121.0837.40
", + "bbox": [ + 501, + 618, + 888, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "trained on a parameterized 3D face reconstruction model [9]. This model does not apply to indoor lighting (among other problems, it predicts all points with the same normal have the same shading). We trained a GC model on the LSUN Bedroom dataset with 2 subspaces $z^k$ - illumination corresponding to spherical harmonic coefficients, and other to represent all other structural information. The model was trained for 800 epochs. GC produces images whose structure varies wildly with any lighting changes, resulting in large albedo changes (GC in Figures 3, 9). The difficulty appears to be that the attribute predictor is easily subverted; if the lighting representation cannot produce an image that is (say) dark on the left side, the albedo is adjusted instead.", + "bbox": [ + 496, + 704, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4237", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/616e562904dbfc4791c0582ec40a42c8d2ed3c14537c46285ea6b0cee9bcadd9.jpg", + "image_caption": [ + "Figure 9. GAN-control (GC) [39] cannot relight complex scenes like bedrooms; it completely changes the scene's layout." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 450, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/136a1a252d87a1153f2c7be876b8d88a7fcd64fcff843b5203840378e6914d27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 194, + 267, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4ee88bae8989c9c8a895b5d15a0e25021316baa5246f7afedb1db4702bcb771b.jpg", + "image_caption": [ + "(c) Living Room Relighting" + ], + "image_footnote": [], + "bbox": [ + 88, + 333, + 264, + 483 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/41ebdd2d30f388edfd89eb5765d19b1696df76dcf09f9c25c465b7abb7791c2c.jpg", + "image_caption": [ + "(e) Relighting outdoor Churches", + "Figure 10. StyLitGAN extends to finding relighting directions for StyleGANs trained on other datasets." + ], + "image_footnote": [], + "bbox": [ + 86, + 497, + 267, + 633 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/345544a3e5e24f28b193a84fae75a982292eb0e373c4b94c3494e93f648e7955.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 281, + 194, + 460, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fdeb4c976808fe9c8c6d5bfed230d28b807fe438aab779b4a93459b9a4805f28.jpg", + "image_caption": [ + "(b) Kitchen Relighting" + ], + "image_footnote": [], + "bbox": [ + 281, + 345, + 459, + 482 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/86a2438cb584c617b683c965ca94228ad9a05d7ece2cf54a21e1e3ca5fad0ccf.jpg", + "image_caption": [ + "(d) Dining Room Relighting", + "(f) Portrait Relighting" + ], + "image_footnote": [], + "bbox": [ + 281, + 497, + 460, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Downstream Applications", + "text_level": 1, + "bbox": [ + 76, + 691, + 318, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lighting Variance in Surface Normal Prediction: The Multilum dataset [29] provides images of 1000 various indoor scenes, each under 25 lighting conditions, physically relit and captured. A SOTA normal predictor (Omnidata [12, 21]) applied to the test set produces surface normal predictions that vary significantly with changes of light in a fixed scene (Figure 11, purple bar). Finetuning Omnidata with the Multilum training dataset significantly reduces this variance (Figure 11, orange bar); but multiple lightings of a fixed scene are very hard to find. Finetuning with StyLitGAN relights produces comparable improvements; using seven distinct relights (Figure 11, pale orange bar) is slightly", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b18315a4c356fb926ef337ffe428642de1853a27d7aba075bb8b6ab8715c80e0.jpg", + "image_caption": [ + "Figure 11. Normal variance to lighting reduces when fine-tuning on our relighting dataset. Purple boxplot shows normal variance under relighting for real test scenes from Multilum using the Omnidata normal predictor; orange shows the result of finetuning using Multilum training data; light orange and yellow show the result of finetuning using StyLitGAN images (7 and 25 per scene respectively). The measure is angular error in radians from the mean prediction of a scene for each relit image in the Multilum test set (30 scenes, 25 lightings each)." + ], + "image_footnote": [], + "bbox": [ + 550, + 78, + 836, + 272 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/648cf2ffc644a4acbef8476ef546b65980865d6af2bf979b6e9b5ecb93fd8ac4.jpg", + "image_caption": [ + "Figure 12. A surface normal predictor is finetuned to increase prediction consistency for relit images of the same scene. General performance shown on Taskonomy test images parallels that of the original model. Mean across buildings is given in the last column." + ], + "image_footnote": [], + "bbox": [ + 542, + 410, + 846, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "worse than using 25 distinct relights (Figure 11, yellow bar). The resulting improvement is not at the cost of base accuracy. Figure 12 compares the accuracy of various finetuned models on the Taskonomy test set [50] (recall this involves thousands of frames each in 10 blocks; we show results by block). Note that finetuned methods mostly show slight accuracy improvements over the base model, but losses in some blocks result in means that match.", + "bbox": [ + 496, + 654, + 893, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our follow-up work, we also show that StyleGAN \"knows\" intrinsic images and can be easily extracted [4].", + "bbox": [ + 496, + 773, + 893, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 818, + 651, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We thank Aniruddha Kembhavi, Derek Hoiem, Min Jin Chong and Shenlong Wang. This material is based upon work supported by the National Science Foundation under Grant No. 2106825 and by a gift from Boeing Corporation.", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4238", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2", + "[2] Peter N Belhumeur and David J Kriegman. What is the set of images of an object under all possible illumination conditions? International Journal of Computer Vision, 1998. 4", + "[3] Anand Bhattachad and D.A. Forsyth. Cut-and-paste object insertion by enabling deep image prior for reshading. In 2022 International Conference on 3D Vision (3DV). IEEE, 2022. 2, 3, 5", + "[4] Anand Bhattachad, Daniel McKee, Derek Hoiem, and David Forsyth. Stylegan knows normal, depth, albedo, and more. Advances in Neural Information Processing Systems, 36, 2024. 8", + "[5] Anand Bhattachad, Viraj Shah, Derek Hoiem, and David A Forsyth. Make it so: Steering stylegan for any image inversion and editing. arXiv preprint arXiv:2304.14403, 2023. 6, 1", + "[6] Prashanth Chandran, Sebastian Winberg, Gaspard Zoss, Jérémy Riviere, Markus Gross, Paulo Gotardo, and Derek Bradley. Rendering with style: combining traditional and neural approaches for high-quality face rendering. ACM Transactions on Graphics (ToG), 40(6):1-14, 2021. 2", + "[7] Min Jin Chong and David Forsyth. Jojogan: One shot face stylization. arXiv preprint arXiv:2112.11641, 2021. 3", + "[8] Min Jin Chong, Hsin-Ying Lee, and David Forsyth. Stylegan of all trades: Image manipulation with only pretrained stylegan. arXiv preprint arXiv:2111.01619, 2021. 2", + "[9] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019.", + "[10] Aditya Deshpande, Jiajun Lu, Mao-Chuang Yeh, Min Jin Chong, and David Forsyth. Learning diverse image colorization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6837-6845, 2017. 2", + "[11] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 341-346, 2001. 2", + "[12] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 8", + "[13] Dave Epstein, Taesung Park, Richard Zhang, Eli Shechtman, and Alexei A Efros. Blobgan: Spatially disentangled scene representations. arXiv preprint arXiv:2205.02837, 2022. 2, 3, 7", + "[14] D.A. Forsyth and Jason J Rock. Intrinsic image decomposition using paradigms. TPAMI, 2022 in press. 2, 3, 5", + "[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and" + ], + "bbox": [ + 78, + 114, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pattern recognition, pages 2414-2423, 2016. 2", + "[16] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. arXiv preprint arXiv:1406.2661, 2014. 2", + "[17] Aaron Hertzmann, Charles E Jacobs, Nuria Oliver, Brian Curless, and David H Salesin. Image analogies. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 327-340, 2001. 2", + "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint arXiv:1706.08500, 2017. 7", + "[19] Le Hui, Jia Yuan, Mingmei Cheng, Jin Xie, Xiaoya Zhang, and Jian Yang. Superpoint network for point cloud oversegmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5510-5519, 2021. 4", + "[20] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, 2016. 3", + "[21] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 2, 8", + "[22] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 1, 2", + "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019.", + "[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 1, 2", + "[25] Edwin H Land. The retina theory of color vision. Scientific american, 1977. 3", + "[26] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Miloš Hasan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. arXiv preprint arXiv:2205.09343, 2022. 6, 1", + "[27] Zicheng Liao, Hugues Hoppe, David Forsyth, and Yizhou Yu. A subdivision-based representation for vector image editing. IEEE transactions on visualization and computer graphics, 2012. 2", + "[28] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. Advances in Neural Information Processing Systems, 34, 2021. 2", + "[29] Lukas Murmann, Michael Gharbi, Miika Aittala, and Fredo Durand. A multi-illumination dataset of indoor object appearance. In 2019 IEEE International Conference on Computer Vision (ICCV), Oct 2019. 2, 8", + "[30] Thomas Nestmeyer, Jean-François Lalonde, Iain Matthews, Epic Games, Andreas Lehrmann, and AI Borealis. Learning physics-guided face relighting under directional light. 2020. 2" + ], + "bbox": [ + 503, + 93, + 893, + 888 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4239", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Xingang Pan, Xudong Xu, Chen Change Loy, Christian Theobalt, and Bo Dai. A shading-guided generative implicit model for shape-accurate 3d-aware image synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2", + "[32] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 2", + "[33] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[34] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 7", + "[35] Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images. IEEE Computer graphics and applications, 21(5):34-41, 2001. 2", + "[36] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 2", + "[37] Soumyadip Sengupta, Brian Curless, Ira Kemelmacher-Shlizerman, and Steven M Seitz. A light stage on every desk. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2", + "[38] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 2020. 1, 2", + "[39] Alon Shoshan, Nadav Bhonker, Igor Kviatkovsky, and Gerard Medioni. Gan-control: Explicitly controllable gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14083-14093, 2021. 2, 5, 7, 8", + "[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. *ICLR*, 2015. 3", + "[41] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics, 2019. 2", + "[42] Feitong Tan, Sean Fanello, Abhimitra Meka, Sergio Orts-Escolano, Danhang Tang, Rohit Pandey, Jonathan Taylor, Ping Tan, and Yinda Zhang. Volux-gan: A generative model for 3d face synthesis with hdri relighting. arXiv preprint arXiv:2201.04873, 2022. 2", + "[43] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan for 3d control over portrait images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6142-6151, 2020. 2", + "[44] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2", + "[45] Andrey Voynov and Artem Babenko. Unsupervised discovery" + ], + "bbox": [ + 78, + 90, + 470, + 891 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of interpretable directions in the gan latent space. In International conference on machine learning, pages 9786-9796. PMLR, 2020. 1, 2", + "[46] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 1, 2", + "[47] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. arXiv preprint arXiv: 2101.05278, 2021. 2", + "[48] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. International Journal of Computer Vision, 2020. 2, 5", + "[49] Ning Yu, Guilin Liu, Aysegul Dundar, Andrew Tao, Bryan Catanzaro, Larry S Davis, and Mario Fritz. Dual contrastive loss and attention for gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6731-6742, 2021. 3", + "[50] Amir R. Zamir, Alexander Sax, William B. Shen, Leonidas J. Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2018. 8", + "[51] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 3", + "[52] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE International Conference on Computer Vision, 2019. 2", + "[53] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 1, 2", + "[54] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, 2017. 2" + ], + "bbox": [ + 501, + 92, + 893, + 633 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4240", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/StyLitGAN_ Image-Based Relighting via Latent Control/37a6c7df-f82e-4da6-aa4d-10d6b8e2b178_model.json b/2024/StyLitGAN_ Image-Based Relighting via Latent Control/37a6c7df-f82e-4da6-aa4d-10d6b8e2b178_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f8ea2e2508be6904a46ef56dd71d0cb6de1f92f8 --- /dev/null +++ b/2024/StyLitGAN_ Image-Based Relighting via Latent Control/37a6c7df-f82e-4da6-aa4d-10d6b8e2b178_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.204, + 0.131, + 0.768, + 0.154 + ], + "angle": 0, + "content": "StyLitGAN: Image-based Relighting via Latent Control" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.182, + 0.696, + 0.199 + ], + "angle": 0, + "content": "Anand Bhattad James Soole D.A. Forsyth" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.2, + 0.649, + 0.218 + ], + "angle": 0, + "content": "University of Illinois Urbana-Champaign" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.218, + 0.726, + 0.234 + ], + "angle": 0, + "content": "https://anandbhattad.github.io/stylitgan/" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.256, + 0.228, + 0.267 + ], + "angle": 0, + "content": "Generated Image" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.256, + 0.842, + 0.269 + ], + "angle": 0, + "content": "Relit-1 \\((\\mathbf{w}^{+} + d_{1})\\) Relit-2 \\((\\mathbf{w}^{+} + d_{2})\\) Relit-3 \\((\\mathbf{w}^{+} + d_{3})\\) Relit-4 \\((\\mathbf{w}^{+} + d_{4})\\) Relit-5 \\((\\mathbf{w}^{+} + d_{5})\\)" + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.268, + 0.848, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.552, + 0.895, + 0.651 + ], + "angle": 0, + "content": "Figure 1. StyLitGAN identifies directional vectors \\((d_i)\\) within StyleGAN's style space \\((\\mathcal{W}^{+})\\) which, when added to the \\(\\mathbf{w}^{+}\\)style code, effectively modify the lighting of generated images while preserving their geometry and albedo. This process eliminates the need for per-image search or model fine-tuning. The first column displays images generated from StyleGAN2; subsequent columns illustrate the same scene, each relit using a specific direction. These relighting directions \\((d_i)\\) are derived through a forward selection method, ensuring diversity and avoiding cherry-picking. The directional effects are consistent across different scenes: for instance, \\(d_1\\) activates an orange-tinged bedside lamp, \\(d_2\\) a less intense white-tinged lamp, \\(d_3\\) introduces strong directional light from the window, and so on, demonstrating diverse relighting capabilities of StyLitGAN." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.662, + 0.314, + 0.679 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.689, + 0.475, + 0.903 + ], + "angle": 0, + "content": "We describe a novel method, StyLitGAN, for relighting and resurfacing images in the absence of labeled data. StyLitGAN generates images with realistic lighting effects, including cast shadows, soft shadows, inter-reflections, and glossy effects, without the need for paired or CGI data. StyLitGAN uses an intrinsic image method to decompose an image, followed by a search of the latent space of a pretrained StyleGAN to identify a set of directions. By prompting the model to fix one component (e.g., albedo) and vary another (e.g., shading), we generate relighted images by adding the identified directions to the latent style codes. Quantitative metrics of change in albedo and lighting diversity allow us to choose effective directions using a forward selection process. Qualitative evaluation confirms the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.664, + 0.631, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.825 + ], + "angle": 0, + "content": "Scene appearance shifts dramatically with varying lighting conditions - a sunlit room takes on a different character as daylight fades, and interior spaces transform with the flick of a switch. Similarly, surface changes, like a wall's paint color, change not only the wall's appearance but also the overall image due to light reflection. Despite the impressive realism achieved by current generative models like StyleGAN [22-24], they fall short in dynamically controlling scene lighting, a key aspect of realistic image generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this work, we present StyLitGAN, a novel approach that extends the editing capabilities of StyleGAN [38, 45, 46, 53]. StyLitGAN uniquely manipulates style codes to selectively change lighting while preserving other image attributes like albedo and geometry. This selective editing" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4231" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.137 + ], + "angle": 0, + "content": "addresses a critical gap in current generative methods, which typically lack the precision to control individual scene components independently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.47, + 0.272 + ], + "angle": 0, + "content": "Our method, StyLitGAN, first uses StyleGAN to produce a set of images and then decomposes these generated images into albedo, diffuse shading, and glossy effects using an off-the-shelf, self-supervised network [14]. We then search for style code edits by prompting StyleGAN to produce images that (a) are diverse, but (b) have the same albedo (and so geometry and material) as the original generated images. Our search selects the most effective relighting directions in a data-driven manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.471, + 0.439 + ], + "angle": 0, + "content": "Our approach generates images with realistic lighting effects, including cast shadows, soft shadows, inter-reflections, and glossy effects. Importantly, we observe that style code edits produce consistent effects across images. For instance, as seen in Fig. 1, adding the first and second directions tends to switch on bedside lamps (columns Relit-1 and Relit-2), while adding the fourth direction increases the light intensity from outside the window (column Relit-4). Since StyLitGAN can generate any image that a vanilla StyleGAN can, it also generate images that are out of distribution, one would expect FID scores to increase over StyleGAN; this happens." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.514 + ], + "angle": 0, + "content": "The recent method GAN-control [39] controls lighting on face images using an attribute procedure. We show, in contrast to StyLitGAN, GAN-control fails on indoor scenes, likely because the attribute vocabulary is too easily subverted by the complex lighting effects in indoor scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.471, + 0.636 + ], + "angle": 0, + "content": "We demonstrate applications of StyLitGAN to standard vision problems. Using the Multi-Ilum dataset of [29], we show that predictions from a SOTA surface normal predictor [21] vary significantly when lighting is changed. Finetuning this normal predictor using StyLitGAN images suppresses this effect. The improvement is comparable with that obtained by finetuning with true multi-illuminant images (which are very difficult to obtain in quantity)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.649, + 0.219, + 0.665 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Image Manipulation: A significant literature deals with manipulating and editing images [3, 10, 11, 15, 17, 27, 33, 35, 44, 54]. Editing procedures for generative image models [16] are important, because they demand compact image representations with useful, disentangled interpretations. StyleGAN [22-24] is currently de facto state-of-the-art for editing generated images, likely because its mapping of initial noise vectors to style codes which control entire feature layers produces latent spaces that are heavily disentangled and so easy to manipulate. Recent editing methods include [8, 36, 38, 45, 46, 53], with a survey in [47]. The architecture can be adapted to incorporate spatial priors for authoring novel and edited images [13, 28, 43]. In contrast to this literature, we show how to fix one physically meaningful image factor while changing another. Doing so is difficult" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.303 + ], + "angle": 0, + "content": "because the latent spaces are not perfectly disentangled, and we must produce a diverse set of changes in only one factor. Relighting using StyleGAN: Relighting faces using StyleGAN can be achieved with Stylerig [43], but this method requires a 3D morphable face model. In contrast, StyLitGAN does not require a 3D model and can be extended to complex indoor scenes, which is not possible with Stylerig. Yang et al. [48] uses semantic label attributes to train a binary classifier to find latent space directions that represent indoor and natural lighting, but this method cannot produce diverse relighting effects. We also find Yang et al's relighting to change color or albedo. In contrast, StyLitGAN generates diverse realistic relighting effects without changing albedo and without requiring any labeled attributes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.894, + 0.529 + ], + "angle": 0, + "content": "StyleFlow [1] and GAN-control [39] require a parametric model to express lighting, such as spherical harmonics. These methods are limited to relighting faces and do not result in realistic relighting of rooms. Our experiments using GAN-control for rooms result in large geometry and albedo change. In contrast, StyLitGAN can produce relighted images without changing geometry or albedo. We also note that rooms are more challenging to relight than faces due to significant long-scale inter-reflection effects, diverse shadow patterns, stylized luminaires, stylized surface albedos, and surface brightnesses that are not a function of surface normal alone. These factors make it difficult to apply GAN control directly to rooms. Also, none of these methods can resurface or recolor rooms, though StyLitGAN can also edit color or materials while preserving the scene's lighting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.894, + 0.681 + ], + "angle": 0, + "content": "Other Face Relighting methods use carefully collected supervisory data from light-stages or parametric spherical harmonics [30, 32, 37, 41, 52]. ShadeGAN [31], Rendering with Style [6], and Volux-GAN [42] use a volumetric rendering approach to learn the 3D structure of the face and the illumination encoding. Volux-GAN [42] also requires image decomposition from [32] that is trained using carefully curated light-stage data. In comparison, we neither require any explicit 3D modeling of the scene nor labeled and curated data for training the image decomposition model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.609, + 0.712 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.894, + 0.885 + ], + "angle": 0, + "content": "We follow convention and manipulate StyleGAN [24] by adjusting the \\(\\mathbf{w}^{+}\\) latent variables. We do not modify StyleGAN weights, but instead, seek a set of lighting directions \\(\\mathbf{d}_i\\) (same shape as \\(\\mathbf{w}^{+}\\)) which are independent of \\(\\mathbf{w}^{+}\\) and have desired effects on the generated image. We obtain these directions by constructing losses that capture the desired outcomes, and then search for directions that minimize these losses. We find all directions only once and use 2000 randomly generated images for this search. Once found, these lighting directions apply to all other generated images. Our search procedure only sees each image once." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fig. 2 summarizes our procedure and we call our model" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4232" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.088, + 0.845, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.26, + 0.892, + 0.345 + ], + "angle": 0, + "content": "Figure 2. How StyLitGAN works: We generate an image from random Gaussian noise using a pretrained StyleGAN. We also generate novel relighted versions (16 in our case) of the same image using randomly initialized latent directions (d) that are added to \\(\\mathbf{w}^{+}\\)latent style codes. We train a classifier (F) that takes in all the pairs of relighted and original images and predicts the relighting direction applied to them. We apply a distinction loss and jointly update the latent directions and the classifier. Next, we generate the decomposition of these images from a pretrained decomposition model (D). We then apply losses that force StyLitGAN to find latent directions such that the albedo does not change (consistency loss), but the image does (diversity loss)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.473, + 0.576 + ], + "angle": 0, + "content": "StyLitGAN. Our method consists of two stages. The first stage involves decomposing images using a pretrained model. The second stage jointly searches for directions and trains a classifier \\( \\mathrm{F} \\). The classifier \\( \\mathrm{F} \\) predicts the latent direction applied to image pairs. It's a classification task with a fixed number of jointly learned latent directions. While classifier \\( \\mathrm{F} \\) doesn't directly know whether latent directions relate to lighting, our use of image decomposition losses ensures these directions are lighting-related. The consistency loss maintains albedo, and the diversity loss ensures diverse shading changes, both properties expected when changing lighting conditions. Thus, while jointly updating the classifier \\( \\mathrm{F} \\) and latent directions, these losses ensure the discovered directions pertain to relighting. We now elaborate on our search for directions and losses in detail." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.472, + 0.656 + ], + "angle": 0, + "content": "Base StyleGAN Models: We use baseline pretrained models from [49] that use a dual-contrastive loss to train StyleGAN for bedrooms, faces, and churches. We also use baseline pretrained StyleGAN2 models from [13] for conference rooms, kitchens, dining rooms, and living rooms." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.472, + 0.869 + ], + "angle": 0, + "content": "Decomposition: We decompose images into albedo, shading, and gloss maps (gloss only when available) as \\( A \\times S + G \\), where \\( A \\) models albedo effects and \\( S \\) and \\( G \\) model shading and gloss effects respectively. We use the method of Forsyth and Rock [14] and its variant from Bhattad and Forsyth [3], which is easily adapted because it is self-supervised and uses only samples from statistical models derived from Land's Retinex theory [25]. By changing the statistical spatial model parameters, we can construct many decompositions using their approach. We evaluate many such decomposition models under several hyperparameter settings and create a large pool of relighting directions. We finalize our directions using a forward selection process that provides minimal albedo and geometry shift with a large relighting diversity (Section 4)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Relighting a scene should produce a new, realistic image where the shading has changed but the albedo has" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.895, + 0.578 + ], + "angle": 0, + "content": "not. Write \\( I(\\mathbf{w}^{+}) \\) for the image produced by StyleGAN given style codes \\( \\mathbf{w}^{+} \\), and \\( A(I) \\), \\( S(I) \\), and \\( G(I) \\) for the albedo, shading, and gloss respectively recovered from image \\( I \\). We search for multiple directions \\( \\mathbf{d}_i \\) such that: (a) \\( A(I(\\mathbf{w}^{+} + \\mathbf{d}_i)) \\) is very close to \\( A(I(\\mathbf{w}^{+})) \\) – so the image is a relighted version of \\( I(\\mathbf{w}^{+}) \\), a property we call persistent consistency; (b) the images produced by the different directions are linearly independent – relighting diversity; (c) which direction was used can be determined from the image, so that different directions have visibly distinct effects – distinctive relighting; and (d) the new shading field (map) is not strongly correlated to the albedo – independent relighting. Not every shading field can be paired with a given albedo, otherwise there would be nothing to do. We assume that edited \\( \\mathbf{w}^{+} \\) will result in realistic images [7]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.894, + 0.654 + ], + "angle": 0, + "content": "Recoloring: Alternatively, we may wish to edit scenes where the colors or materials of objects have changed, but the lighting hasn't. Because shading conveys a great deal of information about shape, we can find these edits using modified losses by seeking consistency in the shading field." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.895, + 0.779 + ], + "angle": 0, + "content": "Persistent Consistency: The albedo decomposition of both the relighted scene: \\( A_{R} = A(I(\\mathbf{w}^{+} + \\mathbf{d}_{i})) \\) and the original: \\( A_{O} = A(I(\\mathbf{w}^{+})) \\) must be the same; where R refers to relighted images and O refers to StyleGAN generated images. We use a Huber loss and a perceptual feature loss [20, 51] from a VGG feature extractor (Φ) [40] at various feature layers (j) to preserve persistent effects (geometry, appearance, and texture) in the scene." + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.788, + 0.891, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n s t}} \\left(A _ {O}, A _ {R}\\right) = \\left\\{ \\begin{array}{c} \\frac {1}{2} \\left[ A _ {O} - A _ {R} \\right] ^ {2} \\quad \\text {f o r} | A _ {O} - A _ {R} | \\leq \\delta , \\\\ \\delta \\left(| A _ {O} - A _ {R} | - \\delta / 2\\right) \\quad \\text {o t h e r w i s e .} \\end{array} \\right. (1) \\\\ \\mathcal {L} _ {p e r} \\left(A _ {O}, A _ {R}\\right) = \\left| \\left| \\Phi_ {j} \\left(A _ {O}\\right) - \\Phi_ {j} \\left(A _ {R}\\right) \\right| \\right| _ {2}. (2) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Relighting Diversity: We want the set of relighted images produced by the directions to be diverse on a long scale so that regions that were in shadow in one image might be" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4233" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.168 + ], + "angle": 0, + "content": "bright in another. For each \\( S(\\mathbf{w}^{+} + \\mathbf{d}_{i}) \\), we stack the two shading and gloss: \\( S \\) and \\( G \\), and compute a smoothed and downsampled vector \\( \\mathbf{t}_{i} \\) from these maps. We then compute \\( \\mathcal{L}_{div}(S,G) \\) (diversity loss) which compels these \\( \\mathbf{t}_{i} \\) to be linearly independent and encourages diversity in relighting." + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.177, + 0.469, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i v}} (S, G) = - \\log \\det N \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.201, + 0.371, + 0.218 + ], + "angle": 0, + "content": "where \\(i^{th}\\) & \\(j^{th}\\) component of \\(\\mathbf{N}\\) is \\(t_i^\\top t_j\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.222, + 0.47, + 0.329 + ], + "angle": 0, + "content": "Distinctive Relighting: A network might try to cheat by making minimal changes to the image. Directions \\(\\mathbf{d}_i\\) should have the property that \\(\\mathbf{d}_i\\) is easy to impute from \\(I(\\mathbf{w}^{+} + \\mathbf{d}_{i})\\). Joint with the search for directions, we train a classifier to categorize the applied direction. This classifier accepts \\(I(\\mathbf{w}^{+})\\) and \\(I(\\mathbf{w}^{+} + \\mathbf{d}_{i})\\) and must predict \\(i\\). Its cross-entropy supplies our loss:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.338, + 0.469, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {l, F} \\mathcal {L} _ {\\text {d i s t}} (I (\\mathbf {w} ^ {+}), I (\\mathbf {w} ^ {+} + d _ {i})) \\\\ = - \\sum_ {i = 1} ^ {M} y _ {i} \\log F \\left(I \\left(\\mathbf {w} ^ {+}\\right), I \\left(\\mathbf {w} ^ {+} + d _ {i}\\right)\\right) \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.405, + 0.47, + 0.466 + ], + "angle": 0, + "content": "Saturation Penalty: Our diversity loss might cheat and obtain high diversity by generating blocks of over-saturated or under-saturated pixels. So, we apply a saturation penalty over several pixels within a certain threshold." + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.477, + 0.469, + 0.564 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {s a t} = \\lambda_ {\\text {o v e r s a t}} \\left[ \\frac {1}{H \\times W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\max (0, I _ {i, j} - s) ^ {2} \\right] \\tag {5} \\\\ + \\lambda_ {\\text {u n d e r s a t}} \\left[ \\frac {1}{H \\times W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\max (0, s - I _ {i, j}) ^ {2} \\right] \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.47, + 0.68 + ], + "angle": 0, + "content": "where \\(\\lambda_{oversat}\\) and \\(\\lambda_{undersat}\\) are the penalty weights for over-saturation and under-saturation respectively, \\(H\\) and \\(W\\) are the height and width of the images, \\(I_{i,j}\\) is the pixel intensity at pixel location \\((i,j)\\), and \\(s\\) is the saturation threshold (i.e., the maximum allowed pixel intensity). The penalty is computed as the mean squared difference between the pixel intensity and the saturation threshold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.471, + 0.773 + ], + "angle": 0, + "content": "Recoloring requires swapping albedo and shading components in all losses, except we do not use decorrelation loss while recoloring. Obtaining good results requires quite a careful choice of loss weights (\\(\\lambda\\) coefficients). We experiment with several \\(\\lambda\\) coefficients for both these edits (Section 4 and Supplementary)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.363, + 0.801 + ], + "angle": 0, + "content": "4. Model and Directions Selection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We prompt StyleGAN to find style code directions that: (a) do not change the albedo, and (b) strongly change the image. We use many image decomposition models to obtain directions across multiple hyperparameter settings. We have no particular reason to believe that a single model will give only good directions or all good directions. We then find" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.198 + ], + "angle": 0, + "content": "a subset of admissible models. We must choose admissible models using a plot of albedo change versus diversity because there is no way to weigh these effects against one another. However, relatively few methods are admissible – see Figure 3. We then pool all directions from all of the admissible models and use forward selection to find a small set of polished directions in this pool." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.205, + 0.894, + 0.31 + ], + "angle": 0, + "content": "Scoring Albedo Change: We use SuperPoints [19] to find 100 interest points in the original StyleGAN-generated image. Around each interest point, we form a \\(8 \\times 8\\) patch. We then compare these patches with patches in the same locations for multiple different relightings of that image. If the albedo in the image does not change, then each patch will have the same albedo but different lighting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.894, + 0.54 + ], + "angle": 0, + "content": "Given two color image patches \\(\\mathbf{p}\\) and \\(\\mathbf{q}\\), viewed under different lights, we must measure the difference between their albedos \\(d_{a}(\\mathbf{p},\\mathbf{q})\\). Write \\(p_{ij}\\) for the RGB vector at the \\(i\\), \\(j\\)'th location (\\(1 \\leq i \\leq M\\), \\(1 \\leq j \\leq N\\)) and write \\(p_{ij,k}\\) for the \\(k\\)'th RGB component at that location. The intensity of the light may change without the albedo changing, so this problem is homogeneous (i.e. for \\(\\lambda, \\mu > 0\\), \\(d_{a}(\\mathbf{p},\\mathbf{q}) = d_{a}(\\lambda \\mathbf{p},\\mu \\mathbf{q}))\\). Assume that the illumination intensity changes, but not color. The patches are small, so the illumination field on a patch can be modeled as a linear function, so there are albedos \\(\\mathbf{a}\\), \\(\\mathbf{b}\\) such that \\(p_{ij} = (p_{x}i + p_{y}j + p_{c})a_{ij}\\) and \\(q_{ij} = (q_{x}i + q_{y}j + q_{c})b_{ij}\\). If the two patches have similar albedo, there will be \\(p_{x}\\) etc. such that \\(p_{ij}' = (q_{x}i + q_{y}j + q_{c})p_{ij}\\) is the same as \\(q_{ij}' = (p_{x}i + p_{y}j + p_{c})q_{ij}\\). We measure the cosine distance" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.552, + 0.892, + 0.598 + ], + "angle": 0, + "content": "\\[\nd _ {a} (\\mathbf {p}, \\mathbf {q}) = 1 - \\max _ {p _ {x}, \\dots q _ {c}} \\frac {\\sum_ {i j k} p _ {i j k} ^ {\\prime} q _ {i j k} ^ {\\prime}}{\\sqrt {\\sum_ {i j k} \\left(p _ {i j k} ^ {\\prime}\\right) ^ {2}} \\sqrt {\\sum_ {i j k} \\left(p _ {i j k} ^ {\\prime}\\right) ^ {2}}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.609, + 0.892, + 0.641 + ], + "angle": 0, + "content": "The relevant maximum can be calculated by analogy with canonical correlation analysis (Supplementary)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.892, + 0.845 + ], + "angle": 0, + "content": "Scoring Lighting Diversity: Illumination cone theory [2] yields that any non-negative linear combination of \\( k \\) shadings is a physically plausible shading. To determine if an image is new, we relax the non-negativity constraint and so must ensure that it cannot be expressed as a linear combination of existing images. In turn, we seek a measure of the linear independence of a set of images. This measure should: be large when there is a strong linear dependency; and not grow too fast when the images are scaled. Write \\( \\mathbf{x}_i \\) for the \\( i \\)'th image, and \\( \\mathcal{X} \\) for the matrix whose \\( i \\), \\( j \\)'th component is \\( \\mathbf{x}_i\\mathbf{x}_j \\). Then \\( -\\log \\det \\mathcal{X} \\) is very large when the \\( \\mathbf{x}_i \\) is close to linearly dependent, but does not scale too fast when the images are scaled." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Decomposition Models Investigated: We searched 25 instances in total obtained with different hyperparameter settings from three families of decomposition. The first family" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4234" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.092, + 0.422, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.326, + 0.473, + 0.562 + ], + "angle": 0, + "content": "Figure 3. Each model from StyLitGAN produces 16 directions. Models differ by choice of hyperparameters and intrinsic image decomposition. We evaluate models by albedo change and by diversity, averaged across a small fixed validation set of test scenes. As the figure shows, there is typically a payoff, but some models are not admissible. Figure 4 shows examples from some of the models considered here. We exclude inadmissible models, then pool all directions from all other models, and apply a forward selection procedure (section 4). This yields 16 strong relighting directions (the gold star). In our comparison with Yang et al. [48], we focus on the changes in albedo since their method identifies only a single relighting direction. This limits the scope for evaluating relighting diversity. Yang et al. [48] aggressively change scene albedo, while our StyLitGAN ensures only lighting changes. Additionally, we compare with GAN-control [39], which, while attempting relighting, often changes the scene layout, leading to large albedo change and increased diversity score due to layout variations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.472, + 0.724 + ], + "angle": 0, + "content": "is the SOTA unsupervised model of [14], which decomposes images into albedo and shading using example images drawn from statistical models. The second is a variant of that family that decomposes into albedo, shading, and gloss decomposition [3]. The third is an albedo, shading, and gloss decomposition that models fine edges in the albedo rather than the shading field. These models were chosen to represent a range of possible decompositions, but others could yield better results. The key point is that we can choose a model from a collection by a rational process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Selecting Directions: Our approach for selecting directions involves creating a scatter plot of 25 instances with various hyperparameters and image decomposition models. We find 16 directions for each instance in our final experiments, and the search for 16 directions takes about 14 minutes on an A40 GPU. We experimented with different numbers of directions of order \\(2^{n}\\) for \\(n = 2, 3, 4, 5, 6, 7\\) and found that 16 directions \\((n = 4)\\) strike a better balance between relighting diversity and albedo change. However, finding multiple directions is challenging because the search space is complex and high-dimensional, and we lack ground truth to supervise" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.089, + 0.88, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.264, + 0.892, + 0.362 + ], + "angle": 0, + "content": "Figure 4. The bottom row shows scene relightings obtained using our final, forward selected, set of directions (star of Figure 3). For comparison, we also show scene relightings from different models obtained from StyLitGAN shown in Figure 3 (model numbers correspond to numbers on that figure). Note how most models are capable of producing some good directions, but not all directions from a given model may be good." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.892, + 0.403 + ], + "angle": 0, + "content": "the search. Therefore, we apply a two-step process to find effective directions and filter out any bad directions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.893, + 0.524 + ], + "angle": 0, + "content": "We first identify and discard inadmissible models that are located behind the Pareto frontier. We then select the top 10 admissible models based on their average albedo change when applied over a large set of fixed validation images. Our goal is to select the best relighting directions from these admissible models. To achieve this, a forward selection process is employed, which involves selecting a subset of directions from the set of admissible directions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.894, + 0.621 + ], + "angle": 0, + "content": "Forward Selection Process: To select the best directions, we begin with all directions from the admissible models, resulting in 160 directions from 10 models. These directions are then filtered to remove \"bad\" directions that produce relighting similar to the original image or shading that does not vary across pixels, resulting in 108 directions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.894, + 0.727 + ], + "angle": 0, + "content": "Next, we use a greedy process to select the best 16 directions from the remaining 108 directions. We evaluate each direction one at a time and add it to the pool if it provides a large diversity score while incurring a small penalty for large albedo change. This process continues until the desired number of directions are selected. The forward selection process is fast and efficient, taking less than a minute." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.727, + 0.894, + 0.787 + ], + "angle": 0, + "content": "The resulting scores from the forward select 16 directions are marked with a star in golden color in Fig. 3. The directions obtained with this process are significantly better than individual models alone. A qualitative ablation is in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.633, + 0.817 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Qualitative Evaluation: StyLitGAN produces realistic images that are out of distribution but known to exist for straightforward physical reasons. Because they're out of distribution, current quantitative evaluation tools do not apply. We evaluate realism qualitatively. Further, there is no direct" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4235" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.088, + 0.179, + 0.1 + ], + "angle": 0, + "content": "Generated Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.089, + 0.273, + 0.1 + ], + "angle": 0, + "content": "Relit-1 \\((+d_{1})\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.089, + 0.376, + 0.1 + ], + "angle": 0, + "content": "Relit-2 \\((+d_{2})\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.089, + 0.476, + 0.1 + ], + "angle": 0, + "content": "Relit-3 \\((+d_{3})\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.089, + 0.578, + 0.1 + ], + "angle": 0, + "content": "Relit-4 \\((+d_{4})\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.089, + 0.679, + 0.1 + ], + "angle": 0, + "content": "Relit-5 \\((+d_{5})\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.089, + 0.781, + 0.1 + ], + "angle": 0, + "content": "Relit-6 \\((+d_{6})\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.089, + 0.882, + 0.1 + ], + "angle": 0, + "content": "Relit-7 \\((+d_{7})\\)" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.101, + 0.894, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.493, + 0.893, + 0.577 + ], + "angle": 0, + "content": "Figure 5. First column images generated by the original StyleGAN. Other columns show images obtained from \\(\\mathbf{w}^{+} + \\mathbf{d}_{i}\\), our relighting directions added to the style codes \\((\\mathbf{w}^{+})\\) of the images in the first column. These directions have been chosen to fix albedo, but change shading. Note: each row shows the same scenes but with different illumination. Lighting varies aggressively, and the individual latent direction \\(\\mathbf{d}_{i}\\) has persistent semantics - each column corresponds to a type of illumination. For example, the second \\((d_{1})\\) and third \\((d_{2})\\) column switches on bedside lamps. It is worth noting the presence of soft shadows, cast shadows, inter-reflections, and glossy changes across relights. Another important observation to note is that all relightings are with respect to world coordinates, not camera coordinates." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.582, + 0.871, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.747, + 0.893, + 0.817 + ], + "angle": 0, + "content": "Figure 6. Resurfacing Generated Images. Instead of relighting images, we can generate resurfaced images by swapping our consistency and diversity loss. We apply diversity loss to change the albedo and consistency loss to maintain the shading and global illumination. The first column shows images generated by the original StyleGAN, and the other columns show images obtained from \\(\\mathbf{w}^{+} + \\mathbf{d}_{i}\\) for our resurfacing or recoloring directions. Each column shows the same scene as in the first column, but with varying surface colors and materials, while the individual latent direction \\(\\mathbf{d}_i\\) retains its semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.83, + 0.473, + 0.892 + ], + "angle": 0, + "content": "comparable method. However, we show relighting comparisons to a recent SOTA method that is physically motivated and trained with CGI data [26] by using a SOTA inversion method [5] (Fig. 14 in Supplementary). For relighting, our" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.83, + 0.895, + 0.892 + ], + "angle": 0, + "content": "method should generate images that: are clearly relightings of a scene; fix geometry and albedo but visibly change shading; and display complicated illumination effects, including soft shadows, cast shadows and gloss. For resurfacing, our" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4236" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.872, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.894, + 0.314 + ], + "angle": 0, + "content": "Figure 7. Controllability: The first and last columns of the figure show relighted images generated using our Relit-1 and Relit-5 directions. The bottom section of the figure features a user-controllable slider that enables adjusting the weight of the relighting effects produced by these two directions. Moving the slider from left to right results in a seamless interpolation between the two lighting directions and provides precise control over the relighting of the generated images." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.32, + 0.877, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.491, + 0.893, + 0.547 + ], + "angle": 0, + "content": "Figure 8. Scaling Directions: The figure depicts the persistent and smooth effects of applying the direction at different scalar coefficients. We use our Relit-2 direction. A slider at the bottom allows the user to adjust the weight of the relighting direction, producing a seamless interpolation when increasing or decreasing the intensity of the chosen direction. The relighting effects range from a well-lit room with the bedside lamp off to weak external lighting with a bedside lamp on." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.473, + 0.727 + ], + "angle": 0, + "content": "method should generate images that: are clearly images of the original layout, but with different materials or colors or changes in furniture; and display illumination effects that are consistent with these color changes. As Figure 6 shows, our method meets these goals. Figure 7 and Figure 8 show interpolation sequences for a relighting between two directions and scaling only one direction. Note that the lighting changes smoothly, as one would expect. Figure 15 in Supplementary shows that our relighting and recoloring directions are largely disentangled. An ablation for individual losses is also given in Supplementary Figure 13." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.472, + 0.818 + ], + "angle": 0, + "content": "Quantitative Evaluation: Figure 3 shows how we can evaluate albedo and lighting shifts. In Table 1 we show we can generate image datasets with increased FID [18, 34] (clean-FID) from the base comparison set. This is strong evidence our method can produce a set of images that is a strict superset of those that the vanilla StyleGAN can produce." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.818, + 0.471, + 0.864 + ], + "angle": 0, + "content": "Generality: We have applied our method to StyleGAN trained on Conference Room, Kitchen, Living Room, Dining Room, Church, and Face datasets (results in Figure 10)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Comparison to GAN-control: GAN-control (GC) [39] represents lighting with a spherical harmonic predictor pre" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.619 + ], + "angle": 0, + "content": "Table 1. FID measures distribution shift and not realism. Our generated images are realistic and are out-of-distribution because of large illumination and color changes in the images. This results in large FID scores. \\( KDL \\) in the table is for kitchen, dining and living room which are jointly trained [13]." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.619, + 0.89, + 0.687 + ], + "angle": 0, + "content": "
TypeBedroomKDLConferenceChurchFaces
StyleGAN (SG)5.015.869.353.805.02
SG + Relighting (RL)14.236.8710.4812.1237.87
SG + Resurfacing (RS)17.039.4110.6318.6034.06
SG + RL + RS21.3911.6812.7121.0837.40
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.901 + ], + "angle": 0, + "content": "trained on a parameterized 3D face reconstruction model [9]. This model does not apply to indoor lighting (among other problems, it predicts all points with the same normal have the same shading). We trained a GC model on the LSUN Bedroom dataset with 2 subspaces \\( z^k \\) - illumination corresponding to spherical harmonic coefficients, and other to represent all other structural information. The model was trained for 800 epochs. GC produces images whose structure varies wildly with any lighting changes, resulting in large albedo changes (GC in Figures 3, 9). The difficulty appears to be that the attribute predictor is easily subverted; if the lighting representation cannot produce an image that is (say) dark on the left side, the albedo is adjusted instead." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4237" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.452, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.157, + 0.47, + 0.185 + ], + "angle": 0, + "content": "Figure 9. GAN-control (GC) [39] cannot relight complex scenes like bedrooms; it completely changes the scene's layout." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.195, + 0.268, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.334, + 0.266, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.486, + 0.252, + 0.497 + ], + "angle": 0, + "content": "(c) Living Room Relighting" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.498, + 0.268, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.637, + 0.264, + 0.648 + ], + "angle": 0, + "content": "(e) Relighting outdoor Churches" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.195, + 0.461, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.334, + 0.431, + 0.346 + ], + "angle": 0, + "content": "(b) Kitchen Relighting" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.347, + 0.46, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.485, + 0.446, + 0.497 + ], + "angle": 0, + "content": "(d) Dining Room Relighting" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.498, + 0.461, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.637, + 0.428, + 0.649 + ], + "angle": 0, + "content": "(f) Portrait Relighting" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.65, + 0.47, + 0.678 + ], + "angle": 0, + "content": "Figure 10. StyLitGAN extends to finding relighting directions for StyleGANs trained on other datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.693, + 0.319, + 0.71 + ], + "angle": 0, + "content": "6. Downstream Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Lighting Variance in Surface Normal Prediction: The Multilum dataset [29] provides images of 1000 various indoor scenes, each under 25 lighting conditions, physically relit and captured. A SOTA normal predictor (Omnidata [12, 21]) applied to the test set produces surface normal predictions that vary significantly with changes of light in a fixed scene (Figure 11, purple bar). Finetuning Omnidata with the Multilum training dataset significantly reduces this variance (Figure 11, orange bar); but multiple lightings of a fixed scene are very hard to find. Finetuning with StyLitGAN relights produces comparable improvements; using seven distinct relights (Figure 11, pale orange bar) is slightly" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.079, + 0.837, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.277, + 0.896, + 0.403 + ], + "angle": 0, + "content": "Figure 11. Normal variance to lighting reduces when fine-tuning on our relighting dataset. Purple boxplot shows normal variance under relighting for real test scenes from Multilum using the Omnidata normal predictor; orange shows the result of finetuning using Multilum training data; light orange and yellow show the result of finetuning using StyLitGAN images (7 and 25 per scene respectively). The measure is angular error in radians from the mean prediction of a scene for each relit image in the Multilum test set (30 scenes, 25 lightings each)." + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.411, + 0.847, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.588, + 0.894, + 0.645 + ], + "angle": 0, + "content": "Figure 12. A surface normal predictor is finetuned to increase prediction consistency for relit images of the same scene. General performance shown on Taskonomy test images parallels that of the original model. Mean across buildings is given in the last column." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.895, + 0.774 + ], + "angle": 0, + "content": "worse than using 25 distinct relights (Figure 11, yellow bar). The resulting improvement is not at the cost of base accuracy. Figure 12 compares the accuracy of various finetuned models on the Taskonomy test set [50] (recall this involves thousands of frames each in 10 blocks; we show results by block). Note that finetuned methods mostly show slight accuracy improvements over the base model, but losses in some blocks result in means that match." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.775, + 0.894, + 0.807 + ], + "angle": 0, + "content": "In our follow-up work, we also show that StyleGAN \"knows\" intrinsic images and can be easily extracted [4]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.819, + 0.652, + 0.836 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "We thank Aniruddha Kembhavi, Derek Hoiem, Min Jin Chong and Shenlong Wang. This material is based upon work supported by the National Science Foundation under Grant No. 2106825 and by a gift from Boeing Corporation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4238" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.183 + ], + "angle": 0, + "content": "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.183, + 0.472, + 0.224 + ], + "angle": 0, + "content": "[2] Peter N Belhumeur and David J Kriegman. What is the set of images of an object under all possible illumination conditions? International Journal of Computer Vision, 1998. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.224, + 0.472, + 0.277 + ], + "angle": 0, + "content": "[3] Anand Bhattachad and D.A. Forsyth. Cut-and-paste object insertion by enabling deep image prior for reshading. In 2022 International Conference on 3D Vision (3DV). IEEE, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.278, + 0.472, + 0.33 + ], + "angle": 0, + "content": "[4] Anand Bhattachad, Daniel McKee, Derek Hoiem, and David Forsyth. Stylegan knows normal, depth, albedo, and more. Advances in Neural Information Processing Systems, 36, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.472, + 0.386 + ], + "angle": 0, + "content": "[5] Anand Bhattachad, Viraj Shah, Derek Hoiem, and David A Forsyth. Make it so: Steering stylegan for any image inversion and editing. arXiv preprint arXiv:2304.14403, 2023. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.386, + 0.472, + 0.455 + ], + "angle": 0, + "content": "[6] Prashanth Chandran, Sebastian Winberg, Gaspard Zoss, Jérémy Riviere, Markus Gross, Paulo Gotardo, and Derek Bradley. Rendering with style: combining traditional and neural approaches for high-quality face rendering. ACM Transactions on Graphics (ToG), 40(6):1-14, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.455, + 0.472, + 0.481 + ], + "angle": 0, + "content": "[7] Min Jin Chong and David Forsyth. Jojogan: One shot face stylization. arXiv preprint arXiv:2112.11641, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.472, + 0.522 + ], + "angle": 0, + "content": "[8] Min Jin Chong, Hsin-Ying Lee, and David Forsyth. Stylegan of all trades: Image manipulation with only pretrained stylegan. arXiv preprint arXiv:2111.01619, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.472, + 0.587 + ], + "angle": 0, + "content": "[9] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.588, + 0.472, + 0.655 + ], + "angle": 0, + "content": "[10] Aditya Deshpande, Jiajun Lu, Mao-Chuang Yeh, Min Jin Chong, and David Forsyth. Learning diverse image colorization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6837-6845, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.657, + 0.472, + 0.711 + ], + "angle": 0, + "content": "[11] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 341-346, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.711, + 0.472, + 0.779 + ], + "angle": 0, + "content": "[12] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.779, + 0.472, + 0.831 + ], + "angle": 0, + "content": "[13] Dave Epstein, Taesung Park, Richard Zhang, Eli Shechtman, and Alexei A Efros. Blobgan: Spatially disentangled scene representations. arXiv preprint arXiv:2205.02837, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.832, + 0.472, + 0.86 + ], + "angle": 0, + "content": "[14] D.A. Forsyth and Jason J Rock. Intrinsic image decomposition using paradigms. TPAMI, 2022 in press. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.86, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.815, + 0.106 + ], + "angle": 0, + "content": "pattern recognition, pages 2414-2423, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.106, + 0.893, + 0.16 + ], + "angle": 0, + "content": "[16] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. arXiv preprint arXiv:1406.2661, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.16, + 0.894, + 0.214 + ], + "angle": 0, + "content": "[17] Aaron Hertzmann, Charles E Jacobs, Nuria Oliver, Brian Curless, and David H Salesin. Image analogies. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 327-340, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.214, + 0.894, + 0.268 + ], + "angle": 0, + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint arXiv:1706.08500, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.268, + 0.894, + 0.322 + ], + "angle": 0, + "content": "[19] Le Hui, Jia Yuan, Mingmei Cheng, Jin Xie, Xiaoya Zhang, and Jian Yang. Superpoint network for point cloud oversegmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5510-5519, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.322, + 0.893, + 0.363 + ], + "angle": 0, + "content": "[20] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.894, + 0.429 + ], + "angle": 0, + "content": "[21] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.43, + 0.893, + 0.485 + ], + "angle": 0, + "content": "[22] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.893, + 0.539 + ], + "angle": 0, + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.893, + 0.58 + ], + "angle": 0, + "content": "[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.58, + 0.893, + 0.605 + ], + "angle": 0, + "content": "[25] Edwin H Land. The retina theory of color vision. Scientific american, 1977. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.605, + 0.893, + 0.673 + ], + "angle": 0, + "content": "[26] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Miloš Hasan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. arXiv preprint arXiv:2205.09343, 2022. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.673, + 0.893, + 0.727 + ], + "angle": 0, + "content": "[27] Zicheng Liao, Hugues Hoppe, David Forsyth, and Yizhou Yu. A subdivision-based representation for vector image editing. IEEE transactions on visualization and computer graphics, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.727, + 0.893, + 0.782 + ], + "angle": 0, + "content": "[28] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. Advances in Neural Information Processing Systems, 34, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.782, + 0.893, + 0.836 + ], + "angle": 0, + "content": "[29] Lukas Murmann, Michael Gharbi, Miika Aittala, and Fredo Durand. A multi-illumination dataset of indoor object appearance. In 2019 IEEE International Conference on Computer Vision (ICCV), Oct 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.836, + 0.893, + 0.889 + ], + "angle": 0, + "content": "[30] Thomas Nestmeyer, Jean-François Lalonde, Iain Matthews, Epic Games, Andreas Lehrmann, and AI Borealis. Learning physics-guided face relighting under directional light. 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.894, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4239" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.16 + ], + "angle": 0, + "content": "[31] Xingang Pan, Xudong Xu, Chen Change Loy, Christian Theobalt, and Bo Dai. A shading-guided generative implicit model for shape-accurate 3d-aware image synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.16, + 0.472, + 0.229 + ], + "angle": 0, + "content": "[32] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.229, + 0.472, + 0.284 + ], + "angle": 0, + "content": "[33] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.284, + 0.472, + 0.322 + ], + "angle": 0, + "content": "[34] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.472, + 0.365 + ], + "angle": 0, + "content": "[35] Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images. IEEE Computer graphics and applications, 21(5):34-41, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.365, + 0.472, + 0.432 + ], + "angle": 0, + "content": "[36] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.432, + 0.472, + 0.486 + ], + "angle": 0, + "content": "[37] Soumyadip Sengupta, Brian Curless, Ira Kemelmacher-Shlizerman, and Steven M Seitz. A light stage on every desk. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.486, + 0.472, + 0.54 + ], + "angle": 0, + "content": "[38] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.54, + 0.472, + 0.595 + ], + "angle": 0, + "content": "[39] Alon Shoshan, Nadav Bhonker, Igor Kviatkovsky, and Gerard Medioni. Gan-control: Explicitly controllable gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14083-14093, 2021. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.472, + 0.633 + ], + "angle": 0, + "content": "[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. *ICLR*, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.472, + 0.689 + ], + "angle": 0, + "content": "[41] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.472, + 0.756 + ], + "angle": 0, + "content": "[42] Feitong Tan, Sean Fanello, Abhimitra Meka, Sergio Orts-Escolano, Danhang Tang, Rohit Pandey, Jonathan Taylor, Ping Tan, and Yinda Zhang. Volux-gan: A generative model for 3d face synthesis with hdri relighting. arXiv preprint arXiv:2201.04873, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.756, + 0.472, + 0.839 + ], + "angle": 0, + "content": "[43] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan for 3d control over portrait images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6142-6151, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.839, + 0.472, + 0.879 + ], + "angle": 0, + "content": "[44] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.879, + 0.472, + 0.892 + ], + "angle": 0, + "content": "[45] Andrey Voynov and Artem Babenko. Unsupervised discovery" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.894, + 0.133 + ], + "angle": 0, + "content": "of interpretable directions in the gan latent space. In International conference on machine learning, pages 9786-9796. PMLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.133, + 0.894, + 0.201 + ], + "angle": 0, + "content": "[46] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.201, + 0.894, + 0.242 + ], + "angle": 0, + "content": "[47] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. arXiv preprint arXiv: 2101.05278, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.242, + 0.894, + 0.295 + ], + "angle": 0, + "content": "[48] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. International Journal of Computer Vision, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.295, + 0.894, + 0.363 + ], + "angle": 0, + "content": "[49] Ning Yu, Guilin Liu, Aysegul Dundar, Andrew Tao, Bryan Catanzaro, Larry S Davis, and Mario Fritz. Dual contrastive loss and attention for gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6731-6742, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.894, + 0.431 + ], + "angle": 0, + "content": "[50] Amir R. Zamir, Alexander Sax, William B. Shen, Leonidas J. Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.431, + 0.894, + 0.472 + ], + "angle": 0, + "content": "[51] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.472, + 0.894, + 0.525 + ], + "angle": 0, + "content": "[52] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.894, + 0.58 + ], + "angle": 0, + "content": "[53] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.58, + 0.894, + 0.635 + ], + "angle": 0, + "content": "[54] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4240" + } + ] +] \ No newline at end of file diff --git a/2024/StyLitGAN_ Image-Based Relighting via Latent Control/37a6c7df-f82e-4da6-aa4d-10d6b8e2b178_origin.pdf b/2024/StyLitGAN_ Image-Based Relighting via Latent Control/37a6c7df-f82e-4da6-aa4d-10d6b8e2b178_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..61c5cdd2b7ce61e04c825af2cfa8de00d0a38840 --- /dev/null +++ b/2024/StyLitGAN_ Image-Based Relighting via Latent Control/37a6c7df-f82e-4da6-aa4d-10d6b8e2b178_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b612a2417e244bf3685c80b4e5e49b8134a122562ef469501182a252939d6fc4 +size 10342680 diff --git a/2024/StyLitGAN_ Image-Based Relighting via Latent Control/full.md b/2024/StyLitGAN_ Image-Based Relighting via Latent Control/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5015c8ae4bf5d757acc10805d02e0da511b34b47 --- /dev/null +++ b/2024/StyLitGAN_ Image-Based Relighting via Latent Control/full.md @@ -0,0 +1,280 @@ +# StyLitGAN: Image-based Relighting via Latent Control + +Anand Bhattad James Soole D.A. Forsyth + +University of Illinois Urbana-Champaign + +https://anandbhattad.github.io/stylitgan/ + +Generated Image + +Relit-1 $(\mathbf{w}^{+} + d_{1})$ Relit-2 $(\mathbf{w}^{+} + d_{2})$ Relit-3 $(\mathbf{w}^{+} + d_{3})$ Relit-4 $(\mathbf{w}^{+} + d_{4})$ Relit-5 $(\mathbf{w}^{+} + d_{5})$ + +![](images/ca50ceef1b4c4acb1b024de96b1a74ca80e39eb45597d829b2b3de7949c1e4ba.jpg) +Figure 1. StyLitGAN identifies directional vectors $(d_i)$ within StyleGAN's style space $(\mathcal{W}^{+})$ which, when added to the $\mathbf{w}^{+}$ style code, effectively modify the lighting of generated images while preserving their geometry and albedo. This process eliminates the need for per-image search or model fine-tuning. The first column displays images generated from StyleGAN2; subsequent columns illustrate the same scene, each relit using a specific direction. These relighting directions $(d_i)$ are derived through a forward selection method, ensuring diversity and avoiding cherry-picking. The directional effects are consistent across different scenes: for instance, $d_1$ activates an orange-tinged bedside lamp, $d_2$ a less intense white-tinged lamp, $d_3$ introduces strong directional light from the window, and so on, demonstrating diverse relighting capabilities of StyLitGAN. + +# Abstract + +We describe a novel method, StyLitGAN, for relighting and resurfacing images in the absence of labeled data. StyLitGAN generates images with realistic lighting effects, including cast shadows, soft shadows, inter-reflections, and glossy effects, without the need for paired or CGI data. StyLitGAN uses an intrinsic image method to decompose an image, followed by a search of the latent space of a pretrained StyleGAN to identify a set of directions. By prompting the model to fix one component (e.g., albedo) and vary another (e.g., shading), we generate relighted images by adding the identified directions to the latent style codes. Quantitative metrics of change in albedo and lighting diversity allow us to choose effective directions using a forward selection process. Qualitative evaluation confirms the effectiveness of our method. + +# 1. Introduction + +Scene appearance shifts dramatically with varying lighting conditions - a sunlit room takes on a different character as daylight fades, and interior spaces transform with the flick of a switch. Similarly, surface changes, like a wall's paint color, change not only the wall's appearance but also the overall image due to light reflection. Despite the impressive realism achieved by current generative models like StyleGAN [22-24], they fall short in dynamically controlling scene lighting, a key aspect of realistic image generation. + +In this work, we present StyLitGAN, a novel approach that extends the editing capabilities of StyleGAN [38, 45, 46, 53]. StyLitGAN uniquely manipulates style codes to selectively change lighting while preserving other image attributes like albedo and geometry. This selective editing + +addresses a critical gap in current generative methods, which typically lack the precision to control individual scene components independently. + +Our method, StyLitGAN, first uses StyleGAN to produce a set of images and then decomposes these generated images into albedo, diffuse shading, and glossy effects using an off-the-shelf, self-supervised network [14]. We then search for style code edits by prompting StyleGAN to produce images that (a) are diverse, but (b) have the same albedo (and so geometry and material) as the original generated images. Our search selects the most effective relighting directions in a data-driven manner. + +Our approach generates images with realistic lighting effects, including cast shadows, soft shadows, inter-reflections, and glossy effects. Importantly, we observe that style code edits produce consistent effects across images. For instance, as seen in Fig. 1, adding the first and second directions tends to switch on bedside lamps (columns Relit-1 and Relit-2), while adding the fourth direction increases the light intensity from outside the window (column Relit-4). Since StyLitGAN can generate any image that a vanilla StyleGAN can, it also generate images that are out of distribution, one would expect FID scores to increase over StyleGAN; this happens. + +The recent method GAN-control [39] controls lighting on face images using an attribute procedure. We show, in contrast to StyLitGAN, GAN-control fails on indoor scenes, likely because the attribute vocabulary is too easily subverted by the complex lighting effects in indoor scenes. + +We demonstrate applications of StyLitGAN to standard vision problems. Using the Multi-Ilum dataset of [29], we show that predictions from a SOTA surface normal predictor [21] vary significantly when lighting is changed. Finetuning this normal predictor using StyLitGAN images suppresses this effect. The improvement is comparable with that obtained by finetuning with true multi-illuminant images (which are very difficult to obtain in quantity). + +# 2. Related Work + +Image Manipulation: A significant literature deals with manipulating and editing images [3, 10, 11, 15, 17, 27, 33, 35, 44, 54]. Editing procedures for generative image models [16] are important, because they demand compact image representations with useful, disentangled interpretations. StyleGAN [22-24] is currently de facto state-of-the-art for editing generated images, likely because its mapping of initial noise vectors to style codes which control entire feature layers produces latent spaces that are heavily disentangled and so easy to manipulate. Recent editing methods include [8, 36, 38, 45, 46, 53], with a survey in [47]. The architecture can be adapted to incorporate spatial priors for authoring novel and edited images [13, 28, 43]. In contrast to this literature, we show how to fix one physically meaningful image factor while changing another. Doing so is difficult + +because the latent spaces are not perfectly disentangled, and we must produce a diverse set of changes in only one factor. Relighting using StyleGAN: Relighting faces using StyleGAN can be achieved with Stylerig [43], but this method requires a 3D morphable face model. In contrast, StyLitGAN does not require a 3D model and can be extended to complex indoor scenes, which is not possible with Stylerig. Yang et al. [48] uses semantic label attributes to train a binary classifier to find latent space directions that represent indoor and natural lighting, but this method cannot produce diverse relighting effects. We also find Yang et al's relighting to change color or albedo. In contrast, StyLitGAN generates diverse realistic relighting effects without changing albedo and without requiring any labeled attributes. + +StyleFlow [1] and GAN-control [39] require a parametric model to express lighting, such as spherical harmonics. These methods are limited to relighting faces and do not result in realistic relighting of rooms. Our experiments using GAN-control for rooms result in large geometry and albedo change. In contrast, StyLitGAN can produce relighted images without changing geometry or albedo. We also note that rooms are more challenging to relight than faces due to significant long-scale inter-reflection effects, diverse shadow patterns, stylized luminaires, stylized surface albedos, and surface brightnesses that are not a function of surface normal alone. These factors make it difficult to apply GAN control directly to rooms. Also, none of these methods can resurface or recolor rooms, though StyLitGAN can also edit color or materials while preserving the scene's lighting. + +Other Face Relighting methods use carefully collected supervisory data from light-stages or parametric spherical harmonics [30, 32, 37, 41, 52]. ShadeGAN [31], Rendering with Style [6], and Volux-GAN [42] use a volumetric rendering approach to learn the 3D structure of the face and the illumination encoding. Volux-GAN [42] also requires image decomposition from [32] that is trained using carefully curated light-stage data. In comparison, we neither require any explicit 3D modeling of the scene nor labeled and curated data for training the image decomposition model. + +# 3. Approach + +We follow convention and manipulate StyleGAN [24] by adjusting the $\mathbf{w}^{+}$ latent variables. We do not modify StyleGAN weights, but instead, seek a set of lighting directions $\mathbf{d}_i$ (same shape as $\mathbf{w}^{+}$ ) which are independent of $\mathbf{w}^{+}$ and have desired effects on the generated image. We obtain these directions by constructing losses that capture the desired outcomes, and then search for directions that minimize these losses. We find all directions only once and use 2000 randomly generated images for this search. Once found, these lighting directions apply to all other generated images. Our search procedure only sees each image once. + +Fig. 2 summarizes our procedure and we call our model + +![](images/ef00eae8814e93b30e2a78df87c02ad12b5ad1ac95088f9a84c9e6b3b31bbaa6.jpg) +Figure 2. How StyLitGAN works: We generate an image from random Gaussian noise using a pretrained StyleGAN. We also generate novel relighted versions (16 in our case) of the same image using randomly initialized latent directions (d) that are added to $\mathbf{w}^{+}$ latent style codes. We train a classifier (F) that takes in all the pairs of relighted and original images and predicts the relighting direction applied to them. We apply a distinction loss and jointly update the latent directions and the classifier. Next, we generate the decomposition of these images from a pretrained decomposition model (D). We then apply losses that force StyLitGAN to find latent directions such that the albedo does not change (consistency loss), but the image does (diversity loss). + +StyLitGAN. Our method consists of two stages. The first stage involves decomposing images using a pretrained model. The second stage jointly searches for directions and trains a classifier $\mathrm{F}$ . The classifier $\mathrm{F}$ predicts the latent direction applied to image pairs. It's a classification task with a fixed number of jointly learned latent directions. While classifier $\mathrm{F}$ doesn't directly know whether latent directions relate to lighting, our use of image decomposition losses ensures these directions are lighting-related. The consistency loss maintains albedo, and the diversity loss ensures diverse shading changes, both properties expected when changing lighting conditions. Thus, while jointly updating the classifier $\mathrm{F}$ and latent directions, these losses ensure the discovered directions pertain to relighting. We now elaborate on our search for directions and losses in detail. + +Base StyleGAN Models: We use baseline pretrained models from [49] that use a dual-contrastive loss to train StyleGAN for bedrooms, faces, and churches. We also use baseline pretrained StyleGAN2 models from [13] for conference rooms, kitchens, dining rooms, and living rooms. + +Decomposition: We decompose images into albedo, shading, and gloss maps (gloss only when available) as $A \times S + G$ , where $A$ models albedo effects and $S$ and $G$ model shading and gloss effects respectively. We use the method of Forsyth and Rock [14] and its variant from Bhattad and Forsyth [3], which is easily adapted because it is self-supervised and uses only samples from statistical models derived from Land's Retinex theory [25]. By changing the statistical spatial model parameters, we can construct many decompositions using their approach. We evaluate many such decomposition models under several hyperparameter settings and create a large pool of relighting directions. We finalize our directions using a forward selection process that provides minimal albedo and geometry shift with a large relighting diversity (Section 4). + +Relighting a scene should produce a new, realistic image where the shading has changed but the albedo has + +not. Write $I(\mathbf{w}^{+})$ for the image produced by StyleGAN given style codes $\mathbf{w}^{+}$ , and $A(I)$ , $S(I)$ , and $G(I)$ for the albedo, shading, and gloss respectively recovered from image $I$ . We search for multiple directions $\mathbf{d}_i$ such that: (a) $A(I(\mathbf{w}^{+} + \mathbf{d}_i))$ is very close to $A(I(\mathbf{w}^{+}))$ – so the image is a relighted version of $I(\mathbf{w}^{+})$ , a property we call persistent consistency; (b) the images produced by the different directions are linearly independent – relighting diversity; (c) which direction was used can be determined from the image, so that different directions have visibly distinct effects – distinctive relighting; and (d) the new shading field (map) is not strongly correlated to the albedo – independent relighting. Not every shading field can be paired with a given albedo, otherwise there would be nothing to do. We assume that edited $\mathbf{w}^{+}$ will result in realistic images [7]. + +Recoloring: Alternatively, we may wish to edit scenes where the colors or materials of objects have changed, but the lighting hasn't. Because shading conveys a great deal of information about shape, we can find these edits using modified losses by seeking consistency in the shading field. + +Persistent Consistency: The albedo decomposition of both the relighted scene: $A_{R} = A(I(\mathbf{w}^{+} + \mathbf{d}_{i}))$ and the original: $A_{O} = A(I(\mathbf{w}^{+}))$ must be the same; where R refers to relighted images and O refers to StyleGAN generated images. We use a Huber loss and a perceptual feature loss [20, 51] from a VGG feature extractor (Φ) [40] at various feature layers (j) to preserve persistent effects (geometry, appearance, and texture) in the scene. + +$$ +\begin{array}{l} \mathcal {L} _ {\text {c o n s t}} \left(A _ {O}, A _ {R}\right) = \left\{ \begin{array}{c} \frac {1}{2} \left[ A _ {O} - A _ {R} \right] ^ {2} \quad \text {f o r} | A _ {O} - A _ {R} | \leq \delta , \\ \delta \left(| A _ {O} - A _ {R} | - \delta / 2\right) \quad \text {o t h e r w i s e .} \end{array} \right. (1) \\ \mathcal {L} _ {p e r} \left(A _ {O}, A _ {R}\right) = \left| \left| \Phi_ {j} \left(A _ {O}\right) - \Phi_ {j} \left(A _ {R}\right) \right| \right| _ {2}. (2) \\ \end{array} +$$ + +Relighting Diversity: We want the set of relighted images produced by the directions to be diverse on a long scale so that regions that were in shadow in one image might be + +bright in another. For each $S(\mathbf{w}^{+} + \mathbf{d}_{i})$ , we stack the two shading and gloss: $S$ and $G$ , and compute a smoothed and downsampled vector $\mathbf{t}_{i}$ from these maps. We then compute $\mathcal{L}_{div}(S,G)$ (diversity loss) which compels these $\mathbf{t}_{i}$ to be linearly independent and encourages diversity in relighting. + +$$ +\mathcal {L} _ {\text {d i v}} (S, G) = - \log \det N \tag {3} +$$ + +where $i^{th}$ & $j^{th}$ component of $\mathbf{N}$ is $t_i^\top t_j$ + +Distinctive Relighting: A network might try to cheat by making minimal changes to the image. Directions $\mathbf{d}_i$ should have the property that $\mathbf{d}_i$ is easy to impute from $I(\mathbf{w}^{+} + \mathbf{d}_{i})$ . Joint with the search for directions, we train a classifier to categorize the applied direction. This classifier accepts $I(\mathbf{w}^{+})$ and $I(\mathbf{w}^{+} + \mathbf{d}_{i})$ and must predict $i$ . Its cross-entropy supplies our loss: + +$$ +\begin{array}{l} \min _ {l, F} \mathcal {L} _ {\text {d i s t}} (I (\mathbf {w} ^ {+}), I (\mathbf {w} ^ {+} + d _ {i})) \\ = - \sum_ {i = 1} ^ {M} y _ {i} \log F \left(I \left(\mathbf {w} ^ {+}\right), I \left(\mathbf {w} ^ {+} + d _ {i}\right)\right) \tag {4} \\ \end{array} +$$ + +Saturation Penalty: Our diversity loss might cheat and obtain high diversity by generating blocks of over-saturated or under-saturated pixels. So, we apply a saturation penalty over several pixels within a certain threshold. + +$$ +\begin{array}{l} \mathcal {L} _ {s a t} = \lambda_ {\text {o v e r s a t}} \left[ \frac {1}{H \times W} \sum_ {i = 1} ^ {H} \sum_ {j = 1} ^ {W} \max (0, I _ {i, j} - s) ^ {2} \right] \tag {5} \\ + \lambda_ {\text {u n d e r s a t}} \left[ \frac {1}{H \times W} \sum_ {i = 1} ^ {H} \sum_ {j = 1} ^ {W} \max (0, s - I _ {i, j}) ^ {2} \right] \\ \end{array} +$$ + +where $\lambda_{oversat}$ and $\lambda_{undersat}$ are the penalty weights for over-saturation and under-saturation respectively, $H$ and $W$ are the height and width of the images, $I_{i,j}$ is the pixel intensity at pixel location $(i,j)$ , and $s$ is the saturation threshold (i.e., the maximum allowed pixel intensity). The penalty is computed as the mean squared difference between the pixel intensity and the saturation threshold. + +Recoloring requires swapping albedo and shading components in all losses, except we do not use decorrelation loss while recoloring. Obtaining good results requires quite a careful choice of loss weights ( $\lambda$ coefficients). We experiment with several $\lambda$ coefficients for both these edits (Section 4 and Supplementary). + +# 4. Model and Directions Selection + +We prompt StyleGAN to find style code directions that: (a) do not change the albedo, and (b) strongly change the image. We use many image decomposition models to obtain directions across multiple hyperparameter settings. We have no particular reason to believe that a single model will give only good directions or all good directions. We then find + +a subset of admissible models. We must choose admissible models using a plot of albedo change versus diversity because there is no way to weigh these effects against one another. However, relatively few methods are admissible – see Figure 3. We then pool all directions from all of the admissible models and use forward selection to find a small set of polished directions in this pool. + +Scoring Albedo Change: We use SuperPoints [19] to find 100 interest points in the original StyleGAN-generated image. Around each interest point, we form a $8 \times 8$ patch. We then compare these patches with patches in the same locations for multiple different relightings of that image. If the albedo in the image does not change, then each patch will have the same albedo but different lighting. + +Given two color image patches $\mathbf{p}$ and $\mathbf{q}$ , viewed under different lights, we must measure the difference between their albedos $d_{a}(\mathbf{p},\mathbf{q})$ . Write $p_{ij}$ for the RGB vector at the $i$ , $j$ 'th location ( $1 \leq i \leq M$ , $1 \leq j \leq N$ ) and write $p_{ij,k}$ for the $k$ 'th RGB component at that location. The intensity of the light may change without the albedo changing, so this problem is homogeneous (i.e. for $\lambda, \mu > 0$ , $d_{a}(\mathbf{p},\mathbf{q}) = d_{a}(\lambda \mathbf{p},\mu \mathbf{q}))$ . Assume that the illumination intensity changes, but not color. The patches are small, so the illumination field on a patch can be modeled as a linear function, so there are albedos $\mathbf{a}$ , $\mathbf{b}$ such that $p_{ij} = (p_{x}i + p_{y}j + p_{c})a_{ij}$ and $q_{ij} = (q_{x}i + q_{y}j + q_{c})b_{ij}$ . If the two patches have similar albedo, there will be $p_{x}$ etc. such that $p_{ij}' = (q_{x}i + q_{y}j + q_{c})p_{ij}$ is the same as $q_{ij}' = (p_{x}i + p_{y}j + p_{c})q_{ij}$ . We measure the cosine distance + +$$ +d _ {a} (\mathbf {p}, \mathbf {q}) = 1 - \max _ {p _ {x}, \dots q _ {c}} \frac {\sum_ {i j k} p _ {i j k} ^ {\prime} q _ {i j k} ^ {\prime}}{\sqrt {\sum_ {i j k} \left(p _ {i j k} ^ {\prime}\right) ^ {2}} \sqrt {\sum_ {i j k} \left(p _ {i j k} ^ {\prime}\right) ^ {2}}} \tag {6} +$$ + +The relevant maximum can be calculated by analogy with canonical correlation analysis (Supplementary). + +Scoring Lighting Diversity: Illumination cone theory [2] yields that any non-negative linear combination of $k$ shadings is a physically plausible shading. To determine if an image is new, we relax the non-negativity constraint and so must ensure that it cannot be expressed as a linear combination of existing images. In turn, we seek a measure of the linear independence of a set of images. This measure should: be large when there is a strong linear dependency; and not grow too fast when the images are scaled. Write $\mathbf{x}_i$ for the $i$ 'th image, and $\mathcal{X}$ for the matrix whose $i$ , $j$ 'th component is $\mathbf{x}_i\mathbf{x}_j$ . Then $-\log \det \mathcal{X}$ is very large when the $\mathbf{x}_i$ is close to linearly dependent, but does not scale too fast when the images are scaled. + +Decomposition Models Investigated: We searched 25 instances in total obtained with different hyperparameter settings from three families of decomposition. The first family + +![](images/b5252d7045a48fc04cdc2dbabcad7e250e0ce1d2b8033904dcbac07430850133.jpg) +Figure 3. Each model from StyLitGAN produces 16 directions. Models differ by choice of hyperparameters and intrinsic image decomposition. We evaluate models by albedo change and by diversity, averaged across a small fixed validation set of test scenes. As the figure shows, there is typically a payoff, but some models are not admissible. Figure 4 shows examples from some of the models considered here. We exclude inadmissible models, then pool all directions from all other models, and apply a forward selection procedure (section 4). This yields 16 strong relighting directions (the gold star). In our comparison with Yang et al. [48], we focus on the changes in albedo since their method identifies only a single relighting direction. This limits the scope for evaluating relighting diversity. Yang et al. [48] aggressively change scene albedo, while our StyLitGAN ensures only lighting changes. Additionally, we compare with GAN-control [39], which, while attempting relighting, often changes the scene layout, leading to large albedo change and increased diversity score due to layout variations. + +is the SOTA unsupervised model of [14], which decomposes images into albedo and shading using example images drawn from statistical models. The second is a variant of that family that decomposes into albedo, shading, and gloss decomposition [3]. The third is an albedo, shading, and gloss decomposition that models fine edges in the albedo rather than the shading field. These models were chosen to represent a range of possible decompositions, but others could yield better results. The key point is that we can choose a model from a collection by a rational process. + +Selecting Directions: Our approach for selecting directions involves creating a scatter plot of 25 instances with various hyperparameters and image decomposition models. We find 16 directions for each instance in our final experiments, and the search for 16 directions takes about 14 minutes on an A40 GPU. We experimented with different numbers of directions of order $2^{n}$ for $n = 2, 3, 4, 5, 6, 7$ and found that 16 directions $(n = 4)$ strike a better balance between relighting diversity and albedo change. However, finding multiple directions is challenging because the search space is complex and high-dimensional, and we lack ground truth to supervise + +![](images/7f5e3de54231b4937aaace146c50da9cea0c22c9600791c8d69cdae838a7e777.jpg) +Figure 4. The bottom row shows scene relightings obtained using our final, forward selected, set of directions (star of Figure 3). For comparison, we also show scene relightings from different models obtained from StyLitGAN shown in Figure 3 (model numbers correspond to numbers on that figure). Note how most models are capable of producing some good directions, but not all directions from a given model may be good. + +the search. Therefore, we apply a two-step process to find effective directions and filter out any bad directions. + +We first identify and discard inadmissible models that are located behind the Pareto frontier. We then select the top 10 admissible models based on their average albedo change when applied over a large set of fixed validation images. Our goal is to select the best relighting directions from these admissible models. To achieve this, a forward selection process is employed, which involves selecting a subset of directions from the set of admissible directions. + +Forward Selection Process: To select the best directions, we begin with all directions from the admissible models, resulting in 160 directions from 10 models. These directions are then filtered to remove "bad" directions that produce relighting similar to the original image or shading that does not vary across pixels, resulting in 108 directions. + +Next, we use a greedy process to select the best 16 directions from the remaining 108 directions. We evaluate each direction one at a time and add it to the pool if it provides a large diversity score while incurring a small penalty for large albedo change. This process continues until the desired number of directions are selected. The forward selection process is fast and efficient, taking less than a minute. + +The resulting scores from the forward select 16 directions are marked with a star in golden color in Fig. 3. The directions obtained with this process are significantly better than individual models alone. A qualitative ablation is in Fig. 4. + +# 5. Experiments + +Qualitative Evaluation: StyLitGAN produces realistic images that are out of distribution but known to exist for straightforward physical reasons. Because they're out of distribution, current quantitative evaluation tools do not apply. We evaluate realism qualitatively. Further, there is no direct + +![](images/7166d8d77f767b023bc10ba80437020a9bbcd67714d16b616b5c6ea92a4fd88f.jpg) +Generated Image +Relit-1 $(+d_{1})$ +Relit-2 $(+d_{2})$ +Relit-3 $(+d_{3})$ +Relit-4 $(+d_{4})$ +Relit-5 $(+d_{5})$ +Relit-6 $(+d_{6})$ +Relit-7 $(+d_{7})$ +Figure 5. First column images generated by the original StyleGAN. Other columns show images obtained from $\mathbf{w}^{+} + \mathbf{d}_{i}$ , our relighting directions added to the style codes $(\mathbf{w}^{+})$ of the images in the first column. These directions have been chosen to fix albedo, but change shading. Note: each row shows the same scenes but with different illumination. Lighting varies aggressively, and the individual latent direction $\mathbf{d}_{i}$ has persistent semantics - each column corresponds to a type of illumination. For example, the second $(d_{1})$ and third $(d_{2})$ column switches on bedside lamps. It is worth noting the presence of soft shadows, cast shadows, inter-reflections, and glossy changes across relights. Another important observation to note is that all relightings are with respect to world coordinates, not camera coordinates. + +![](images/d1bc21dc64fd1c6702e3e66077b3aaa869db1ebc164fc3a4c97c19c1a94e7b06.jpg) +Figure 6. Resurfacing Generated Images. Instead of relighting images, we can generate resurfaced images by swapping our consistency and diversity loss. We apply diversity loss to change the albedo and consistency loss to maintain the shading and global illumination. The first column shows images generated by the original StyleGAN, and the other columns show images obtained from $\mathbf{w}^{+} + \mathbf{d}_{i}$ for our resurfacing or recoloring directions. Each column shows the same scene as in the first column, but with varying surface colors and materials, while the individual latent direction $\mathbf{d}_i$ retains its semantics. + +comparable method. However, we show relighting comparisons to a recent SOTA method that is physically motivated and trained with CGI data [26] by using a SOTA inversion method [5] (Fig. 14 in Supplementary). For relighting, our + +method should generate images that: are clearly relightings of a scene; fix geometry and albedo but visibly change shading; and display complicated illumination effects, including soft shadows, cast shadows and gloss. For resurfacing, our + +![](images/2fd4b9a12e44e0c1910ee4e27fdfba0f92dd8940b5ac0a140ab0b22371286c64.jpg) +Figure 7. Controllability: The first and last columns of the figure show relighted images generated using our Relit-1 and Relit-5 directions. The bottom section of the figure features a user-controllable slider that enables adjusting the weight of the relighting effects produced by these two directions. Moving the slider from left to right results in a seamless interpolation between the two lighting directions and provides precise control over the relighting of the generated images. + +![](images/ea7eef6570ef1ec9bd1a3fc1856d73f98a7de1090b9b33827b3fb51fa5f55fae.jpg) +Figure 8. Scaling Directions: The figure depicts the persistent and smooth effects of applying the direction at different scalar coefficients. We use our Relit-2 direction. A slider at the bottom allows the user to adjust the weight of the relighting direction, producing a seamless interpolation when increasing or decreasing the intensity of the chosen direction. The relighting effects range from a well-lit room with the bedside lamp off to weak external lighting with a bedside lamp on. + +method should generate images that: are clearly images of the original layout, but with different materials or colors or changes in furniture; and display illumination effects that are consistent with these color changes. As Figure 6 shows, our method meets these goals. Figure 7 and Figure 8 show interpolation sequences for a relighting between two directions and scaling only one direction. Note that the lighting changes smoothly, as one would expect. Figure 15 in Supplementary shows that our relighting and recoloring directions are largely disentangled. An ablation for individual losses is also given in Supplementary Figure 13. + +Quantitative Evaluation: Figure 3 shows how we can evaluate albedo and lighting shifts. In Table 1 we show we can generate image datasets with increased FID [18, 34] (clean-FID) from the base comparison set. This is strong evidence our method can produce a set of images that is a strict superset of those that the vanilla StyleGAN can produce. + +Generality: We have applied our method to StyleGAN trained on Conference Room, Kitchen, Living Room, Dining Room, Church, and Face datasets (results in Figure 10). + +Comparison to GAN-control: GAN-control (GC) [39] represents lighting with a spherical harmonic predictor pre + +Table 1. FID measures distribution shift and not realism. Our generated images are realistic and are out-of-distribution because of large illumination and color changes in the images. This results in large FID scores. $KDL$ in the table is for kitchen, dining and living room which are jointly trained [13]. + +
TypeBedroomKDLConferenceChurchFaces
StyleGAN (SG)5.015.869.353.805.02
SG + Relighting (RL)14.236.8710.4812.1237.87
SG + Resurfacing (RS)17.039.4110.6318.6034.06
SG + RL + RS21.3911.6812.7121.0837.40
+ +trained on a parameterized 3D face reconstruction model [9]. This model does not apply to indoor lighting (among other problems, it predicts all points with the same normal have the same shading). We trained a GC model on the LSUN Bedroom dataset with 2 subspaces $z^k$ - illumination corresponding to spherical harmonic coefficients, and other to represent all other structural information. The model was trained for 800 epochs. GC produces images whose structure varies wildly with any lighting changes, resulting in large albedo changes (GC in Figures 3, 9). The difficulty appears to be that the attribute predictor is easily subverted; if the lighting representation cannot produce an image that is (say) dark on the left side, the albedo is adjusted instead. + +![](images/616e562904dbfc4791c0582ec40a42c8d2ed3c14537c46285ea6b0cee9bcadd9.jpg) +Figure 9. GAN-control (GC) [39] cannot relight complex scenes like bedrooms; it completely changes the scene's layout. + +![](images/136a1a252d87a1153f2c7be876b8d88a7fcd64fcff843b5203840378e6914d27.jpg) + +![](images/4ee88bae8989c9c8a895b5d15a0e25021316baa5246f7afedb1db4702bcb771b.jpg) +(c) Living Room Relighting + +![](images/41ebdd2d30f388edfd89eb5765d19b1696df76dcf09f9c25c465b7abb7791c2c.jpg) +(e) Relighting outdoor Churches +Figure 10. StyLitGAN extends to finding relighting directions for StyleGANs trained on other datasets. + +![](images/345544a3e5e24f28b193a84fae75a982292eb0e373c4b94c3494e93f648e7955.jpg) + +![](images/fdeb4c976808fe9c8c6d5bfed230d28b807fe438aab779b4a93459b9a4805f28.jpg) +(b) Kitchen Relighting + +![](images/86a2438cb584c617b683c965ca94228ad9a05d7ece2cf54a21e1e3ca5fad0ccf.jpg) +(d) Dining Room Relighting +(f) Portrait Relighting + +# 6. Downstream Applications + +Lighting Variance in Surface Normal Prediction: The Multilum dataset [29] provides images of 1000 various indoor scenes, each under 25 lighting conditions, physically relit and captured. A SOTA normal predictor (Omnidata [12, 21]) applied to the test set produces surface normal predictions that vary significantly with changes of light in a fixed scene (Figure 11, purple bar). Finetuning Omnidata with the Multilum training dataset significantly reduces this variance (Figure 11, orange bar); but multiple lightings of a fixed scene are very hard to find. Finetuning with StyLitGAN relights produces comparable improvements; using seven distinct relights (Figure 11, pale orange bar) is slightly + +![](images/b18315a4c356fb926ef337ffe428642de1853a27d7aba075bb8b6ab8715c80e0.jpg) +Figure 11. Normal variance to lighting reduces when fine-tuning on our relighting dataset. Purple boxplot shows normal variance under relighting for real test scenes from Multilum using the Omnidata normal predictor; orange shows the result of finetuning using Multilum training data; light orange and yellow show the result of finetuning using StyLitGAN images (7 and 25 per scene respectively). The measure is angular error in radians from the mean prediction of a scene for each relit image in the Multilum test set (30 scenes, 25 lightings each). + +![](images/648cf2ffc644a4acbef8476ef546b65980865d6af2bf979b6e9b5ecb93fd8ac4.jpg) +Figure 12. A surface normal predictor is finetuned to increase prediction consistency for relit images of the same scene. General performance shown on Taskonomy test images parallels that of the original model. Mean across buildings is given in the last column. + +worse than using 25 distinct relights (Figure 11, yellow bar). The resulting improvement is not at the cost of base accuracy. Figure 12 compares the accuracy of various finetuned models on the Taskonomy test set [50] (recall this involves thousands of frames each in 10 blocks; we show results by block). Note that finetuned methods mostly show slight accuracy improvements over the base model, but losses in some blocks result in means that match. + +In our follow-up work, we also show that StyleGAN "knows" intrinsic images and can be easily extracted [4]. + +# Acknowledgment + +We thank Aniruddha Kembhavi, Derek Hoiem, Min Jin Chong and Shenlong Wang. This material is based upon work supported by the National Science Foundation under Grant No. 2106825 and by a gift from Boeing Corporation. + +# References + +[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2 +[2] Peter N Belhumeur and David J Kriegman. What is the set of images of an object under all possible illumination conditions? International Journal of Computer Vision, 1998. 4 +[3] Anand Bhattachad and D.A. Forsyth. Cut-and-paste object insertion by enabling deep image prior for reshading. In 2022 International Conference on 3D Vision (3DV). IEEE, 2022. 2, 3, 5 +[4] Anand Bhattachad, Daniel McKee, Derek Hoiem, and David Forsyth. Stylegan knows normal, depth, albedo, and more. Advances in Neural Information Processing Systems, 36, 2024. 8 +[5] Anand Bhattachad, Viraj Shah, Derek Hoiem, and David A Forsyth. Make it so: Steering stylegan for any image inversion and editing. arXiv preprint arXiv:2304.14403, 2023. 6, 1 +[6] Prashanth Chandran, Sebastian Winberg, Gaspard Zoss, Jérémy Riviere, Markus Gross, Paulo Gotardo, and Derek Bradley. Rendering with style: combining traditional and neural approaches for high-quality face rendering. ACM Transactions on Graphics (ToG), 40(6):1-14, 2021. 2 +[7] Min Jin Chong and David Forsyth. Jojogan: One shot face stylization. arXiv preprint arXiv:2112.11641, 2021. 3 +[8] Min Jin Chong, Hsin-Ying Lee, and David Forsyth. Stylegan of all trades: Image manipulation with only pretrained stylegan. arXiv preprint arXiv:2111.01619, 2021. 2 +[9] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. +[10] Aditya Deshpande, Jiajun Lu, Mao-Chuang Yeh, Min Jin Chong, and David Forsyth. Learning diverse image colorization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6837-6845, 2017. 2 +[11] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 341-346, 2001. 2 +[12] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 8 +[13] Dave Epstein, Taesung Park, Richard Zhang, Eli Shechtman, and Alexei A Efros. Blobgan: Spatially disentangled scene representations. arXiv preprint arXiv:2205.02837, 2022. 2, 3, 7 +[14] D.A. Forsyth and Jason J Rock. Intrinsic image decomposition using paradigms. TPAMI, 2022 in press. 2, 3, 5 +[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and + +pattern recognition, pages 2414-2423, 2016. 2 +[16] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. arXiv preprint arXiv:1406.2661, 2014. 2 +[17] Aaron Hertzmann, Charles E Jacobs, Nuria Oliver, Brian Curless, and David H Salesin. Image analogies. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 327-340, 2001. 2 +[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint arXiv:1706.08500, 2017. 7 +[19] Le Hui, Jia Yuan, Mingmei Cheng, Jin Xie, Xiaoya Zhang, and Jian Yang. Superpoint network for point cloud oversegmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5510-5519, 2021. 4 +[20] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, 2016. 3 +[21] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 2, 8 +[22] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 1, 2 +[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. +[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 1, 2 +[25] Edwin H Land. The retina theory of color vision. Scientific american, 1977. 3 +[26] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Miloš Hasan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. arXiv preprint arXiv:2205.09343, 2022. 6, 1 +[27] Zicheng Liao, Hugues Hoppe, David Forsyth, and Yizhou Yu. A subdivision-based representation for vector image editing. IEEE transactions on visualization and computer graphics, 2012. 2 +[28] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. Advances in Neural Information Processing Systems, 34, 2021. 2 +[29] Lukas Murmann, Michael Gharbi, Miika Aittala, and Fredo Durand. A multi-illumination dataset of indoor object appearance. In 2019 IEEE International Conference on Computer Vision (ICCV), Oct 2019. 2, 8 +[30] Thomas Nestmeyer, Jean-François Lalonde, Iain Matthews, Epic Games, Andreas Lehrmann, and AI Borealis. Learning physics-guided face relighting under directional light. 2020. 2 + +[31] Xingang Pan, Xudong Xu, Chen Change Loy, Christian Theobalt, and Bo Dai. A shading-guided generative implicit model for shape-accurate 3d-aware image synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2 +[32] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 2 +[33] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2 +[34] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 7 +[35] Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images. IEEE Computer graphics and applications, 21(5):34-41, 2001. 2 +[36] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 2 +[37] Soumyadip Sengupta, Brian Curless, Ira Kemelmacher-Shlizerman, and Steven M Seitz. A light stage on every desk. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2 +[38] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 2020. 1, 2 +[39] Alon Shoshan, Nadav Bhonker, Igor Kviatkovsky, and Gerard Medioni. Gan-control: Explicitly controllable gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14083-14093, 2021. 2, 5, 7, 8 +[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. *ICLR*, 2015. 3 +[41] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics, 2019. 2 +[42] Feitong Tan, Sean Fanello, Abhimitra Meka, Sergio Orts-Escolano, Danhang Tang, Rohit Pandey, Jonathan Taylor, Ping Tan, and Yinda Zhang. Volux-gan: A generative model for 3d face synthesis with hdri relighting. arXiv preprint arXiv:2201.04873, 2022. 2 +[43] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan for 3d control over portrait images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6142-6151, 2020. 2 +[44] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2 +[45] Andrey Voynov and Artem Babenko. Unsupervised discovery + +of interpretable directions in the gan latent space. In International conference on machine learning, pages 9786-9796. PMLR, 2020. 1, 2 +[46] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 1, 2 +[47] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. arXiv preprint arXiv: 2101.05278, 2021. 2 +[48] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. International Journal of Computer Vision, 2020. 2, 5 +[49] Ning Yu, Guilin Liu, Aysegul Dundar, Andrew Tao, Bryan Catanzaro, Larry S Davis, and Mario Fritz. Dual contrastive loss and attention for gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6731-6742, 2021. 3 +[50] Amir R. Zamir, Alexander Sax, William B. Shen, Leonidas J. Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2018. 8 +[51] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 3 +[52] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE International Conference on Computer Vision, 2019. 2 +[53] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 1, 2 +[54] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, 2017. 2 \ No newline at end of file diff --git a/2024/StyLitGAN_ Image-Based Relighting via Latent Control/images.zip b/2024/StyLitGAN_ Image-Based Relighting via Latent Control/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6db151dc40ed1be70d037d60458b64798dfbad79 --- /dev/null +++ b/2024/StyLitGAN_ Image-Based Relighting via Latent Control/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76108c8f7346cfad3376e7288a5febac304c95a82d7f3ef8a7b80b74a8962f3f +size 997540 diff --git a/2024/StyLitGAN_ Image-Based Relighting via Latent Control/layout.json b/2024/StyLitGAN_ Image-Based Relighting via Latent Control/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9ba9913ca322d66917656dc433f96dd24daee0c5 --- /dev/null +++ b/2024/StyLitGAN_ Image-Based Relighting via Latent Control/layout.json @@ -0,0 +1,8662 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 470, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 470, + 121 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 470, + 121 + ], + "type": "text", + "content": "StyLitGAN: Image-based Relighting via Latent Control" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 168, + 144, + 425, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 144, + 425, + 157 + ], + "spans": [ + { + "bbox": [ + 168, + 144, + 425, + 157 + ], + "type": "text", + "content": "Anand Bhattad James Soole D.A. Forsyth" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 197, + 158, + 397, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 158, + 397, + 172 + ], + "spans": [ + { + "bbox": [ + 197, + 158, + 397, + 172 + ], + "type": "text", + "content": "University of Illinois Urbana-Champaign" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 148, + 172, + 444, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 172, + 444, + 185 + ], + "spans": [ + { + "bbox": [ + 148, + 172, + 444, + 185 + ], + "type": "text", + "content": "https://anandbhattad.github.io/stylitgan/" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 202, + 139, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 202, + 139, + 211 + ], + "spans": [ + { + "bbox": [ + 82, + 202, + 139, + 211 + ], + "type": "text", + "content": "Generated Image" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 150, + 202, + 515, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 202, + 515, + 213 + ], + "spans": [ + { + "bbox": [ + 150, + 202, + 515, + 213 + ], + "type": "text", + "content": "Relit-1 " + }, + { + "bbox": [ + 150, + 202, + 515, + 213 + ], + "type": "inline_equation", + "content": "(\\mathbf{w}^{+} + d_{1})" + }, + { + "bbox": [ + 150, + 202, + 515, + 213 + ], + "type": "text", + "content": " Relit-2 " + }, + { + "bbox": [ + 150, + 202, + 515, + 213 + ], + "type": "inline_equation", + "content": "(\\mathbf{w}^{+} + d_{2})" + }, + { + "bbox": [ + 150, + 202, + 515, + 213 + ], + "type": "text", + "content": " Relit-3 " + }, + { + "bbox": [ + 150, + 202, + 515, + 213 + ], + "type": "inline_equation", + "content": "(\\mathbf{w}^{+} + d_{3})" + }, + { + "bbox": [ + 150, + 202, + 515, + 213 + ], + "type": "text", + "content": " Relit-4 " + }, + { + "bbox": [ + 150, + 202, + 515, + 213 + ], + "type": "inline_equation", + "content": "(\\mathbf{w}^{+} + d_{4})" + }, + { + "bbox": [ + 150, + 202, + 515, + 213 + ], + "type": "text", + "content": " Relit-5 " + }, + { + "bbox": [ + 150, + 202, + 515, + 213 + ], + "type": "inline_equation", + "content": "(\\mathbf{w}^{+} + d_{5})" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 73, + 212, + 518, + 435 + ], + "blocks": [ + { + "bbox": [ + 73, + 212, + 518, + 435 + ], + "lines": [ + { + "bbox": [ + 73, + 212, + 518, + 435 + ], + "spans": [ + { + "bbox": [ + 73, + 212, + 518, + 435 + ], + "type": "image", + "image_path": "ca50ceef1b4c4acb1b024de96b1a74ca80e39eb45597d829b2b3de7949c1e4ba.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 437, + 547, + 515 + ], + "lines": [ + { + "bbox": [ + 45, + 437, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 45, + 437, + 547, + 515 + ], + "type": "text", + "content": "Figure 1. StyLitGAN identifies directional vectors " + }, + { + "bbox": [ + 45, + 437, + 547, + 515 + ], + "type": "inline_equation", + "content": "(d_i)" + }, + { + "bbox": [ + 45, + 437, + 547, + 515 + ], + "type": "text", + "content": " within StyleGAN's style space " + }, + { + "bbox": [ + 45, + 437, + 547, + 515 + ], + "type": "inline_equation", + "content": "(\\mathcal{W}^{+})" + }, + { + "bbox": [ + 45, + 437, + 547, + 515 + ], + "type": "text", + "content": " which, when added to the " + }, + { + "bbox": [ + 45, + 437, + 547, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{+}" + }, + { + "bbox": [ + 45, + 437, + 547, + 515 + ], + "type": "text", + "content": "style code, effectively modify the lighting of generated images while preserving their geometry and albedo. This process eliminates the need for per-image search or model fine-tuning. The first column displays images generated from StyleGAN2; subsequent columns illustrate the same scene, each relit using a specific direction. These relighting directions " + }, + { + "bbox": [ + 45, + 437, + 547, + 515 + ], + "type": "inline_equation", + "content": "(d_i)" + }, + { + "bbox": [ + 45, + 437, + 547, + 515 + ], + "type": "text", + "content": " are derived through a forward selection method, ensuring diversity and avoiding cherry-picking. The directional effects are consistent across different scenes: for instance, " + }, + { + "bbox": [ + 45, + 437, + 547, + 515 + ], + "type": "inline_equation", + "content": "d_1" + }, + { + "bbox": [ + 45, + 437, + 547, + 515 + ], + "type": "text", + "content": " activates an orange-tinged bedside lamp, " + }, + { + "bbox": [ + 45, + 437, + 547, + 515 + ], + "type": "inline_equation", + "content": "d_2" + }, + { + "bbox": [ + 45, + 437, + 547, + 515 + ], + "type": "text", + "content": " a less intense white-tinged lamp, " + }, + { + "bbox": [ + 45, + 437, + 547, + 515 + ], + "type": "inline_equation", + "content": "d_3" + }, + { + "bbox": [ + 45, + 437, + 547, + 515 + ], + "type": "text", + "content": " introduces strong directional light from the window, and so on, demonstrating diverse relighting capabilities of StyLitGAN." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 524, + 192, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 524, + 192, + 537 + ], + "spans": [ + { + "bbox": [ + 143, + 524, + 192, + 537 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 545, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 545, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 545, + 290, + 715 + ], + "type": "text", + "content": "We describe a novel method, StyLitGAN, for relighting and resurfacing images in the absence of labeled data. StyLitGAN generates images with realistic lighting effects, including cast shadows, soft shadows, inter-reflections, and glossy effects, without the need for paired or CGI data. StyLitGAN uses an intrinsic image method to decompose an image, followed by a search of the latent space of a pretrained StyleGAN to identify a set of directions. By prompting the model to fix one component (e.g., albedo) and vary another (e.g., shading), we generate relighted images by adding the identified directions to the latent style codes. Quantitative metrics of change in albedo and lighting diversity allow us to choose effective directions using a forward selection process. Qualitative evaluation confirms the effectiveness of our method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 525, + 386, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 386, + 538 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 386, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 545, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 653 + ], + "type": "text", + "content": "Scene appearance shifts dramatically with varying lighting conditions - a sunlit room takes on a different character as daylight fades, and interior spaces transform with the flick of a switch. Similarly, surface changes, like a wall's paint color, change not only the wall's appearance but also the overall image due to light reflection. Despite the impressive realism achieved by current generative models like StyleGAN [22-24], they fall short in dynamically controlling scene lighting, a key aspect of realistic image generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "In this work, we present StyLitGAN, a novel approach that extends the editing capabilities of StyleGAN [38, 45, 46, 53]. StyLitGAN uniquely manipulates style codes to selectively change lighting while preserving other image attributes like albedo and geometry. This selective editing" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4231" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "type": "text", + "content": "addresses a critical gap in current generative methods, which typically lack the precision to control individual scene components independently." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 215 + ], + "type": "text", + "content": "Our method, StyLitGAN, first uses StyleGAN to produce a set of images and then decomposes these generated images into albedo, diffuse shading, and glossy effects using an off-the-shelf, self-supervised network [14]. We then search for style code edits by prompting StyleGAN to produce images that (a) are diverse, but (b) have the same albedo (and so geometry and material) as the original generated images. Our search selects the most effective relighting directions in a data-driven manner." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 216, + 288, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 288, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 288, + 347 + ], + "type": "text", + "content": "Our approach generates images with realistic lighting effects, including cast shadows, soft shadows, inter-reflections, and glossy effects. Importantly, we observe that style code edits produce consistent effects across images. For instance, as seen in Fig. 1, adding the first and second directions tends to switch on bedside lamps (columns Relit-1 and Relit-2), while adding the fourth direction increases the light intensity from outside the window (column Relit-4). Since StyLitGAN can generate any image that a vanilla StyleGAN can, it also generate images that are out of distribution, one would expect FID scores to increase over StyleGAN; this happens." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 347, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 407 + ], + "type": "text", + "content": "The recent method GAN-control [39] controls lighting on face images using an attribute procedure. We show, in contrast to StyLitGAN, GAN-control fails on indoor scenes, likely because the attribute vocabulary is too easily subverted by the complex lighting effects in indoor scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 407, + 288, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 288, + 503 + ], + "type": "text", + "content": "We demonstrate applications of StyLitGAN to standard vision problems. Using the Multi-Ilum dataset of [29], we show that predictions from a SOTA surface normal predictor [21] vary significantly when lighting is changed. Finetuning this normal predictor using StyLitGAN images suppresses this effect. The improvement is comparable with that obtained by finetuning with true multi-illuminant images (which are very difficult to obtain in quantity)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 514, + 134, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 134, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 134, + 526 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": "Image Manipulation: A significant literature deals with manipulating and editing images [3, 10, 11, 15, 17, 27, 33, 35, 44, 54]. Editing procedures for generative image models [16] are important, because they demand compact image representations with useful, disentangled interpretations. StyleGAN [22-24] is currently de facto state-of-the-art for editing generated images, likely because its mapping of initial noise vectors to style codes which control entire feature layers produces latent spaces that are heavily disentangled and so easy to manipulate. Recent editing methods include [8, 36, 38, 45, 46, 53], with a survey in [47]. The architecture can be adapted to incorporate spatial priors for authoring novel and edited images [13, 28, 43]. In contrast to this literature, we show how to fix one physically meaningful image factor while changing another. Doing so is difficult" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "text", + "content": "because the latent spaces are not perfectly disentangled, and we must produce a diverse set of changes in only one factor. Relighting using StyleGAN: Relighting faces using StyleGAN can be achieved with Stylerig [43], but this method requires a 3D morphable face model. In contrast, StyLitGAN does not require a 3D model and can be extended to complex indoor scenes, which is not possible with Stylerig. Yang et al. [48] uses semantic label attributes to train a binary classifier to find latent space directions that represent indoor and natural lighting, but this method cannot produce diverse relighting effects. We also find Yang et al's relighting to change color or albedo. In contrast, StyLitGAN generates diverse realistic relighting effects without changing albedo and without requiring any labeled attributes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 239, + 547, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 547, + 418 + ], + "type": "text", + "content": "StyleFlow [1] and GAN-control [39] require a parametric model to express lighting, such as spherical harmonics. These methods are limited to relighting faces and do not result in realistic relighting of rooms. Our experiments using GAN-control for rooms result in large geometry and albedo change. In contrast, StyLitGAN can produce relighted images without changing geometry or albedo. We also note that rooms are more challenging to relight than faces due to significant long-scale inter-reflection effects, diverse shadow patterns, stylized luminaires, stylized surface albedos, and surface brightnesses that are not a function of surface normal alone. These factors make it difficult to apply GAN control directly to rooms. Also, none of these methods can resurface or recolor rooms, though StyLitGAN can also edit color or materials while preserving the scene's lighting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 419, + 547, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 547, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 547, + 539 + ], + "type": "text", + "content": "Other Face Relighting methods use carefully collected supervisory data from light-stages or parametric spherical harmonics [30, 32, 37, 41, 52]. ShadeGAN [31], Rendering with Style [6], and Volux-GAN [42] use a volumetric rendering approach to learn the 3D structure of the face and the illumination encoding. Volux-GAN [42] also requires image decomposition from [32] that is trained using carefully curated light-stage data. In comparison, we neither require any explicit 3D modeling of the scene nor labeled and curated data for training the image decomposition model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 550, + 372, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 372, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 372, + 563 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 700 + ], + "type": "text", + "content": "We follow convention and manipulate StyleGAN [24] by adjusting the " + }, + { + "bbox": [ + 304, + 570, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{+}" + }, + { + "bbox": [ + 304, + 570, + 547, + 700 + ], + "type": "text", + "content": " latent variables. We do not modify StyleGAN weights, but instead, seek a set of lighting directions " + }, + { + "bbox": [ + 304, + 570, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i" + }, + { + "bbox": [ + 304, + 570, + 547, + 700 + ], + "type": "text", + "content": " (same shape as " + }, + { + "bbox": [ + 304, + 570, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{+}" + }, + { + "bbox": [ + 304, + 570, + 547, + 700 + ], + "type": "text", + "content": ") which are independent of " + }, + { + "bbox": [ + 304, + 570, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{+}" + }, + { + "bbox": [ + 304, + 570, + 547, + 700 + ], + "type": "text", + "content": " and have desired effects on the generated image. We obtain these directions by constructing losses that capture the desired outcomes, and then search for directions that minimize these losses. We find all directions only once and use 2000 randomly generated images for this search. Once found, these lighting directions apply to all other generated images. Our search procedure only sees each image once." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Fig. 2 summarizes our procedure and we call our model" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4232" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 69, + 517, + 200 + ], + "blocks": [ + { + "bbox": [ + 72, + 69, + 517, + 200 + ], + "lines": [ + { + "bbox": [ + 72, + 69, + 517, + 200 + ], + "spans": [ + { + "bbox": [ + 72, + 69, + 517, + 200 + ], + "type": "image", + "image_path": "ef00eae8814e93b30e2a78df87c02ad12b5ad1ac95088f9a84c9e6b3b31bbaa6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 545, + 273 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 545, + 273 + ], + "type": "text", + "content": "Figure 2. How StyLitGAN works: We generate an image from random Gaussian noise using a pretrained StyleGAN. We also generate novel relighted versions (16 in our case) of the same image using randomly initialized latent directions (d) that are added to " + }, + { + "bbox": [ + 46, + 205, + 545, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{+}" + }, + { + "bbox": [ + 46, + 205, + 545, + 273 + ], + "type": "text", + "content": "latent style codes. We train a classifier (F) that takes in all the pairs of relighted and original images and predicts the relighting direction applied to them. We apply a distinction loss and jointly update the latent directions and the classifier. Next, we generate the decomposition of these images from a pretrained decomposition model (D). We then apply losses that force StyLitGAN to find latent directions such that the albedo does not change (consistency loss), but the image does (diversity loss)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 456 + ], + "type": "text", + "content": "StyLitGAN. Our method consists of two stages. The first stage involves decomposing images using a pretrained model. The second stage jointly searches for directions and trains a classifier " + }, + { + "bbox": [ + 46, + 277, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\mathrm{F}" + }, + { + "bbox": [ + 46, + 277, + 289, + 456 + ], + "type": "text", + "content": ". The classifier " + }, + { + "bbox": [ + 46, + 277, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\mathrm{F}" + }, + { + "bbox": [ + 46, + 277, + 289, + 456 + ], + "type": "text", + "content": " predicts the latent direction applied to image pairs. It's a classification task with a fixed number of jointly learned latent directions. While classifier " + }, + { + "bbox": [ + 46, + 277, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\mathrm{F}" + }, + { + "bbox": [ + 46, + 277, + 289, + 456 + ], + "type": "text", + "content": " doesn't directly know whether latent directions relate to lighting, our use of image decomposition losses ensures these directions are lighting-related. The consistency loss maintains albedo, and the diversity loss ensures diverse shading changes, both properties expected when changing lighting conditions. Thus, while jointly updating the classifier " + }, + { + "bbox": [ + 46, + 277, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\mathrm{F}" + }, + { + "bbox": [ + 46, + 277, + 289, + 456 + ], + "type": "text", + "content": " and latent directions, these losses ensure the discovered directions pertain to relighting. We now elaborate on our search for directions and losses in detail." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 459, + 288, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 288, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 288, + 519 + ], + "type": "text", + "content": "Base StyleGAN Models: We use baseline pretrained models from [49] that use a dual-contrastive loss to train StyleGAN for bedrooms, faces, and churches. We also use baseline pretrained StyleGAN2 models from [13] for conference rooms, kitchens, dining rooms, and living rooms." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 520, + 288, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 288, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 288, + 688 + ], + "type": "text", + "content": "Decomposition: We decompose images into albedo, shading, and gloss maps (gloss only when available) as " + }, + { + "bbox": [ + 46, + 520, + 288, + 688 + ], + "type": "inline_equation", + "content": "A \\times S + G" + }, + { + "bbox": [ + 46, + 520, + 288, + 688 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 520, + 288, + 688 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 520, + 288, + 688 + ], + "type": "text", + "content": " models albedo effects and " + }, + { + "bbox": [ + 46, + 520, + 288, + 688 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 520, + 288, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 520, + 288, + 688 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 520, + 288, + 688 + ], + "type": "text", + "content": " model shading and gloss effects respectively. We use the method of Forsyth and Rock [14] and its variant from Bhattad and Forsyth [3], which is easily adapted because it is self-supervised and uses only samples from statistical models derived from Land's Retinex theory [25]. By changing the statistical spatial model parameters, we can construct many decompositions using their approach. We evaluate many such decomposition models under several hyperparameter settings and create a large pool of relighting directions. We finalize our directions using a forward selection process that provides minimal albedo and geometry shift with a large relighting diversity (Section 4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Relighting a scene should produce a new, realistic image where the shading has changed but the albedo has" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "text", + "content": "not. Write " + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "inline_equation", + "content": "I(\\mathbf{w}^{+})" + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "text", + "content": " for the image produced by StyleGAN given style codes " + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{+}" + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "inline_equation", + "content": "A(I)" + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "inline_equation", + "content": "S(I)" + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "inline_equation", + "content": "G(I)" + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "text", + "content": " for the albedo, shading, and gloss respectively recovered from image " + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "text", + "content": ". We search for multiple directions " + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i" + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "text", + "content": " such that: (a) " + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "inline_equation", + "content": "A(I(\\mathbf{w}^{+} + \\mathbf{d}_i))" + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "text", + "content": " is very close to " + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "inline_equation", + "content": "A(I(\\mathbf{w}^{+}))" + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "text", + "content": " – so the image is a relighted version of " + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "inline_equation", + "content": "I(\\mathbf{w}^{+})" + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "text", + "content": ", a property we call persistent consistency; (b) the images produced by the different directions are linearly independent – relighting diversity; (c) which direction was used can be determined from the image, so that different directions have visibly distinct effects – distinctive relighting; and (d) the new shading field (map) is not strongly correlated to the albedo – independent relighting. Not every shading field can be paired with a given albedo, otherwise there would be nothing to do. We assume that edited " + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{+}" + }, + { + "bbox": [ + 304, + 277, + 547, + 457 + ], + "type": "text", + "content": " will result in realistic images [7]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 457, + 547, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 547, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 547, + 517 + ], + "type": "text", + "content": "Recoloring: Alternatively, we may wish to edit scenes where the colors or materials of objects have changed, but the lighting hasn't. Because shading conveys a great deal of information about shape, we can find these edits using modified losses by seeking consistency in the shading field." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 521, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 616 + ], + "type": "text", + "content": "Persistent Consistency: The albedo decomposition of both the relighted scene: " + }, + { + "bbox": [ + 304, + 521, + 547, + 616 + ], + "type": "inline_equation", + "content": "A_{R} = A(I(\\mathbf{w}^{+} + \\mathbf{d}_{i}))" + }, + { + "bbox": [ + 304, + 521, + 547, + 616 + ], + "type": "text", + "content": " and the original: " + }, + { + "bbox": [ + 304, + 521, + 547, + 616 + ], + "type": "inline_equation", + "content": "A_{O} = A(I(\\mathbf{w}^{+}))" + }, + { + "bbox": [ + 304, + 521, + 547, + 616 + ], + "type": "text", + "content": " must be the same; where R refers to relighted images and O refers to StyleGAN generated images. We use a Huber loss and a perceptual feature loss [20, 51] from a VGG feature extractor (Φ) [40] at various feature layers (j) to preserve persistent effects (geometry, appearance, and texture) in the scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 624, + 545, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 624, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 624, + 545, + 673 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {c o n s t}} \\left(A _ {O}, A _ {R}\\right) = \\left\\{ \\begin{array}{c} \\frac {1}{2} \\left[ A _ {O} - A _ {R} \\right] ^ {2} \\quad \\text {f o r} | A _ {O} - A _ {R} | \\leq \\delta , \\\\ \\delta \\left(| A _ {O} - A _ {R} | - \\delta / 2\\right) \\quad \\text {o t h e r w i s e .} \\end{array} \\right. (1) \\\\ \\mathcal {L} _ {p e r} \\left(A _ {O}, A _ {R}\\right) = \\left| \\left| \\Phi_ {j} \\left(A _ {O}\\right) - \\Phi_ {j} \\left(A _ {R}\\right) \\right| \\right| _ {2}. (2) \\\\ \\end{array}", + "image_path": "999cd4df1388a0eaadb471152f335d98ab466ec2585c25da334599c9f4a0d9ca.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Relighting Diversity: We want the set of relighted images produced by the directions to be diverse on a long scale so that regions that were in shadow in one image might be" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4233" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": "bright in another. For each " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "S(\\mathbf{w}^{+} + \\mathbf{d}_{i})" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ", we stack the two shading and gloss: " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ", and compute a smoothed and downsampled vector " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{i}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " from these maps. We then compute " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{div}(S,G)" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " (diversity loss) which compels these " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{i}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " to be linearly independent and encourages diversity in relighting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 140, + 287, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 140, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 116, + 140, + 287, + 152 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i v}} (S, G) = - \\log \\det N \\tag {3}", + "image_path": "88156a9efbd14b4fdc55013b6a84d6592a07fe06be1120b1bcc505f67cb00428.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 159, + 227, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 159, + 227, + 172 + ], + "spans": [ + { + "bbox": [ + 59, + 159, + 227, + 172 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 59, + 159, + 227, + 172 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 59, + 159, + 227, + 172 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 59, + 159, + 227, + 172 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 59, + 159, + 227, + 172 + ], + "type": "text", + "content": " component of " + }, + { + "bbox": [ + 59, + 159, + 227, + 172 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 59, + 159, + 227, + 172 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 59, + 159, + 227, + 172 + ], + "type": "inline_equation", + "content": "t_i^\\top t_j" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 175, + 287, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 287, + 260 + ], + "type": "text", + "content": "Distinctive Relighting: A network might try to cheat by making minimal changes to the image. Directions " + }, + { + "bbox": [ + 47, + 175, + 287, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i" + }, + { + "bbox": [ + 47, + 175, + 287, + 260 + ], + "type": "text", + "content": " should have the property that " + }, + { + "bbox": [ + 47, + 175, + 287, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i" + }, + { + "bbox": [ + 47, + 175, + 287, + 260 + ], + "type": "text", + "content": " is easy to impute from " + }, + { + "bbox": [ + 47, + 175, + 287, + 260 + ], + "type": "inline_equation", + "content": "I(\\mathbf{w}^{+} + \\mathbf{d}_{i})" + }, + { + "bbox": [ + 47, + 175, + 287, + 260 + ], + "type": "text", + "content": ". Joint with the search for directions, we train a classifier to categorize the applied direction. This classifier accepts " + }, + { + "bbox": [ + 47, + 175, + 287, + 260 + ], + "type": "inline_equation", + "content": "I(\\mathbf{w}^{+})" + }, + { + "bbox": [ + 47, + 175, + 287, + 260 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 175, + 287, + 260 + ], + "type": "inline_equation", + "content": "I(\\mathbf{w}^{+} + \\mathbf{d}_{i})" + }, + { + "bbox": [ + 47, + 175, + 287, + 260 + ], + "type": "text", + "content": " and must predict " + }, + { + "bbox": [ + 47, + 175, + 287, + 260 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 175, + 287, + 260 + ], + "type": "text", + "content": ". Its cross-entropy supplies our loss:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 267, + 287, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 267, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 93, + 267, + 287, + 316 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {l, F} \\mathcal {L} _ {\\text {d i s t}} (I (\\mathbf {w} ^ {+}), I (\\mathbf {w} ^ {+} + d _ {i})) \\\\ = - \\sum_ {i = 1} ^ {M} y _ {i} \\log F \\left(I \\left(\\mathbf {w} ^ {+}\\right), I \\left(\\mathbf {w} ^ {+} + d _ {i}\\right)\\right) \\tag {4} \\\\ \\end{array}", + "image_path": "ded74f30f6fecc719feb54369d3816d09301851a16c53f4b8ce448b8c8793d9a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 320, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 320, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 320, + 287, + 369 + ], + "type": "text", + "content": "Saturation Penalty: Our diversity loss might cheat and obtain high diversity by generating blocks of over-saturated or under-saturated pixels. So, we apply a saturation penalty over several pixels within a certain threshold." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 377, + 287, + 446 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 377, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 287, + 446 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {s a t} = \\lambda_ {\\text {o v e r s a t}} \\left[ \\frac {1}{H \\times W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\max (0, I _ {i, j} - s) ^ {2} \\right] \\tag {5} \\\\ + \\lambda_ {\\text {u n d e r s a t}} \\left[ \\frac {1}{H \\times W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\max (0, s - I _ {i, j}) ^ {2} \\right] \\\\ \\end{array}", + "image_path": "acd211325bb086e838f5558e5283a6fe429f01b1f8a6f06847dd2ca5c314fc22.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 456, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 538 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 456, + 287, + 538 + ], + "type": "inline_equation", + "content": "\\lambda_{oversat}" + }, + { + "bbox": [ + 46, + 456, + 287, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 456, + 287, + 538 + ], + "type": "inline_equation", + "content": "\\lambda_{undersat}" + }, + { + "bbox": [ + 46, + 456, + 287, + 538 + ], + "type": "text", + "content": " are the penalty weights for over-saturation and under-saturation respectively, " + }, + { + "bbox": [ + 46, + 456, + 287, + 538 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 456, + 287, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 456, + 287, + 538 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 456, + 287, + 538 + ], + "type": "text", + "content": " are the height and width of the images, " + }, + { + "bbox": [ + 46, + 456, + 287, + 538 + ], + "type": "inline_equation", + "content": "I_{i,j}" + }, + { + "bbox": [ + 46, + 456, + 287, + 538 + ], + "type": "text", + "content": " is the pixel intensity at pixel location " + }, + { + "bbox": [ + 46, + 456, + 287, + 538 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 46, + 456, + 287, + 538 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 456, + 287, + 538 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 456, + 287, + 538 + ], + "type": "text", + "content": " is the saturation threshold (i.e., the maximum allowed pixel intensity). The penalty is computed as the mean squared difference between the pixel intensity and the saturation threshold." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 539, + 288, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 288, + 612 + ], + "type": "text", + "content": "Recoloring requires swapping albedo and shading components in all losses, except we do not use decorrelation loss while recoloring. Obtaining good results requires quite a careful choice of loss weights (" + }, + { + "bbox": [ + 46, + 539, + 288, + 612 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 539, + 288, + 612 + ], + "type": "text", + "content": " coefficients). We experiment with several " + }, + { + "bbox": [ + 46, + 539, + 288, + 612 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 539, + 288, + 612 + ], + "type": "text", + "content": " coefficients for both these edits (Section 4 and Supplementary)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 621, + 222, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 222, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 222, + 634 + ], + "type": "text", + "content": "4. Model and Directions Selection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "content": "We prompt StyleGAN to find style code directions that: (a) do not change the albedo, and (b) strongly change the image. We use many image decomposition models to obtain directions across multiple hyperparameter settings. We have no particular reason to believe that a single model will give only good directions or all good directions. We then find" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "a subset of admissible models. We must choose admissible models using a plot of albedo change versus diversity because there is no way to weigh these effects against one another. However, relatively few methods are admissible – see Figure 3. We then pool all directions from all of the admissible models and use forward selection to find a small set of polished directions in this pool." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 162, + 547, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 162, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 304, + 162, + 547, + 245 + ], + "type": "text", + "content": "Scoring Albedo Change: We use SuperPoints [19] to find 100 interest points in the original StyleGAN-generated image. Around each interest point, we form a " + }, + { + "bbox": [ + 304, + 162, + 547, + 245 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 304, + 162, + 547, + 245 + ], + "type": "text", + "content": " patch. We then compare these patches with patches in the same locations for multiple different relightings of that image. If the albedo in the image does not change, then each patch will have the same albedo but different lighting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": "Given two color image patches " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": ", viewed under different lights, we must measure the difference between their albedos " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "d_{a}(\\mathbf{p},\\mathbf{q})" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": ". Write " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "p_{ij}" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": " for the RGB vector at the " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": "'th location (" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "1 \\leq i \\leq M" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "1 \\leq j \\leq N" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": ") and write " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "p_{ij,k}" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": "'th RGB component at that location. The intensity of the light may change without the albedo changing, so this problem is homogeneous (i.e. for " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "\\lambda, \\mu > 0" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "d_{a}(\\mathbf{p},\\mathbf{q}) = d_{a}(\\lambda \\mathbf{p},\\mu \\mathbf{q}))" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": ". Assume that the illumination intensity changes, but not color. The patches are small, so the illumination field on a patch can be modeled as a linear function, so there are albedos " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "p_{ij} = (p_{x}i + p_{y}j + p_{c})a_{ij}" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "q_{ij} = (q_{x}i + q_{y}j + q_{c})b_{ij}" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": ". If the two patches have similar albedo, there will be " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "p_{x}" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": " etc. such that " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "p_{ij}' = (q_{x}i + q_{y}j + q_{c})p_{ij}" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": " is the same as " + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "inline_equation", + "content": "q_{ij}' = (p_{x}i + p_{y}j + p_{c})q_{ij}" + }, + { + "bbox": [ + 304, + 247, + 547, + 427 + ], + "type": "text", + "content": ". We measure the cosine distance" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 310, + 437, + 545, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 437, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 310, + 437, + 545, + 473 + ], + "type": "interline_equation", + "content": "d _ {a} (\\mathbf {p}, \\mathbf {q}) = 1 - \\max _ {p _ {x}, \\dots q _ {c}} \\frac {\\sum_ {i j k} p _ {i j k} ^ {\\prime} q _ {i j k} ^ {\\prime}}{\\sqrt {\\sum_ {i j k} \\left(p _ {i j k} ^ {\\prime}\\right) ^ {2}} \\sqrt {\\sum_ {i j k} \\left(p _ {i j k} ^ {\\prime}\\right) ^ {2}}} \\tag {6}", + "image_path": "594d8ad05cd8c8943a33f4fdd34b46c2be9bd29b089d5e0890e24c1ca30fa0a8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 482, + 545, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 482, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 305, + 482, + 545, + 507 + ], + "type": "text", + "content": "The relevant maximum can be calculated by analogy with canonical correlation analysis (Supplementary)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "text", + "content": "Scoring Lighting Diversity: Illumination cone theory [2] yields that any non-negative linear combination of " + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "text", + "content": " shadings is a physically plausible shading. To determine if an image is new, we relax the non-negativity constraint and so must ensure that it cannot be expressed as a linear combination of existing images. In turn, we seek a measure of the linear independence of a set of images. This measure should: be large when there is a strong linear dependency; and not grow too fast when the images are scaled. Write " + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "text", + "content": "'th image, and " + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "text", + "content": " for the matrix whose " + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "text", + "content": "'th component is " + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i\\mathbf{x}_j" + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "inline_equation", + "content": "-\\log \\det \\mathcal{X}" + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "text", + "content": " is very large when the " + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 514, + 545, + 669 + ], + "type": "text", + "content": " is close to linearly dependent, but does not scale too fast when the images are scaled." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Decomposition Models Investigated: We searched 25 instances in total obtained with different hyperparameter settings from three families of decomposition. The first family" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4234" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 72, + 258, + 255 + ], + "blocks": [ + { + "bbox": [ + 74, + 72, + 258, + 255 + ], + "lines": [ + { + "bbox": [ + 74, + 72, + 258, + 255 + ], + "spans": [ + { + "bbox": [ + 74, + 72, + 258, + 255 + ], + "type": "image", + "image_path": "b5252d7045a48fc04cdc2dbabcad7e250e0ce1d2b8033904dcbac07430850133.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 258, + 289, + 445 + ], + "lines": [ + { + "bbox": [ + 46, + 258, + 289, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 289, + 445 + ], + "type": "text", + "content": "Figure 3. Each model from StyLitGAN produces 16 directions. Models differ by choice of hyperparameters and intrinsic image decomposition. We evaluate models by albedo change and by diversity, averaged across a small fixed validation set of test scenes. As the figure shows, there is typically a payoff, but some models are not admissible. Figure 4 shows examples from some of the models considered here. We exclude inadmissible models, then pool all directions from all other models, and apply a forward selection procedure (section 4). This yields 16 strong relighting directions (the gold star). In our comparison with Yang et al. [48], we focus on the changes in albedo since their method identifies only a single relighting direction. This limits the scope for evaluating relighting diversity. Yang et al. [48] aggressively change scene albedo, while our StyLitGAN ensures only lighting changes. Additionally, we compare with GAN-control [39], which, while attempting relighting, often changes the scene layout, leading to large albedo change and increased diversity score due to layout variations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 453, + 288, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 288, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 288, + 573 + ], + "type": "text", + "content": "is the SOTA unsupervised model of [14], which decomposes images into albedo and shading using example images drawn from statistical models. The second is a variant of that family that decomposes into albedo, shading, and gloss decomposition [3]. The third is an albedo, shading, and gloss decomposition that models fine edges in the albedo rather than the shading field. These models were chosen to represent a range of possible decompositions, but others could yield better results. The key point is that we can choose a model from a collection by a rational process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Selecting Directions: Our approach for selecting directions involves creating a scatter plot of 25 instances with various hyperparameters and image decomposition models. We find 16 directions for each instance in our final experiments, and the search for 16 directions takes about 14 minutes on an A40 GPU. We experimented with different numbers of directions of order " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "2^{n}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "n = 2, 3, 4, 5, 6, 7" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " and found that 16 directions " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "(n = 4)" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " strike a better balance between relighting diversity and albedo change. However, finding multiple directions is challenging because the search space is complex and high-dimensional, and we lack ground truth to supervise" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 312, + 70, + 538, + 207 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 538, + 207 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 538, + 207 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 538, + 207 + ], + "type": "image", + "image_path": "7f5e3de54231b4937aaace146c50da9cea0c22c9600791c8d69cdae838a7e777.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 209, + 545, + 286 + ], + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 286 + ], + "type": "text", + "content": "Figure 4. The bottom row shows scene relightings obtained using our final, forward selected, set of directions (star of Figure 3). For comparison, we also show scene relightings from different models obtained from StyLitGAN shown in Figure 3 (model numbers correspond to numbers on that figure). Note how most models are capable of producing some good directions, but not all directions from a given model may be good." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 295, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 545, + 319 + ], + "type": "text", + "content": "the search. Therefore, we apply a two-step process to find effective directions and filter out any bad directions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 546, + 415 + ], + "type": "text", + "content": "We first identify and discard inadmissible models that are located behind the Pareto frontier. We then select the top 10 admissible models based on their average albedo change when applied over a large set of fixed validation images. Our goal is to select the best relighting directions from these admissible models. To achieve this, a forward selection process is employed, which involves selecting a subset of directions from the set of admissible directions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 419, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 547, + 491 + ], + "type": "text", + "content": "Forward Selection Process: To select the best directions, we begin with all directions from the admissible models, resulting in 160 directions from 10 models. These directions are then filtered to remove \"bad\" directions that produce relighting similar to the original image or shading that does not vary across pixels, resulting in 108 directions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 491, + 547, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 547, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 547, + 575 + ], + "type": "text", + "content": "Next, we use a greedy process to select the best 16 directions from the remaining 108 directions. We evaluate each direction one at a time and add it to the pool if it provides a large diversity score while incurring a small penalty for large albedo change. This process continues until the desired number of directions are selected. The forward selection process is fast and efficient, taking less than a minute." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 575, + 547, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 575, + 547, + 623 + ], + "spans": [ + { + "bbox": [ + 304, + 575, + 547, + 623 + ], + "type": "text", + "content": "The resulting scores from the forward select 16 directions are marked with a star in golden color in Fig. 3. The directions obtained with this process are significantly better than individual models alone. A qualitative ablation is in Fig. 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 634, + 387, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 387, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 387, + 647 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Qualitative Evaluation: StyLitGAN produces realistic images that are out of distribution but known to exist for straightforward physical reasons. Because they're out of distribution, current quantitative evaluation tools do not apply. We evaluate realism qualitatively. Further, there is no direct" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4235" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 79, + 547, + 388 + ], + "blocks": [ + { + "bbox": [ + 51, + 69, + 109, + 79 + ], + "lines": [ + { + "bbox": [ + 51, + 69, + 109, + 79 + ], + "spans": [ + { + "bbox": [ + 51, + 69, + 109, + 79 + ], + "type": "text", + "content": "Generated Image" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 117, + 70, + 167, + 79 + ], + "lines": [ + { + "bbox": [ + 117, + 70, + 167, + 79 + ], + "spans": [ + { + "bbox": [ + 117, + 70, + 167, + 79 + ], + "type": "text", + "content": "Relit-1 " + }, + { + "bbox": [ + 117, + 70, + 167, + 79 + ], + "type": "inline_equation", + "content": "(+d_{1})" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 179, + 70, + 230, + 79 + ], + "lines": [ + { + "bbox": [ + 179, + 70, + 230, + 79 + ], + "spans": [ + { + "bbox": [ + 179, + 70, + 230, + 79 + ], + "type": "text", + "content": "Relit-2 " + }, + { + "bbox": [ + 179, + 70, + 230, + 79 + ], + "type": "inline_equation", + "content": "(+d_{2})" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 241, + 70, + 291, + 79 + ], + "lines": [ + { + "bbox": [ + 241, + 70, + 291, + 79 + ], + "spans": [ + { + "bbox": [ + 241, + 70, + 291, + 79 + ], + "type": "text", + "content": "Relit-3 " + }, + { + "bbox": [ + 241, + 70, + 291, + 79 + ], + "type": "inline_equation", + "content": "(+d_{3})" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 70, + 353, + 79 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 353, + 79 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 353, + 79 + ], + "type": "text", + "content": "Relit-4 " + }, + { + "bbox": [ + 304, + 70, + 353, + 79 + ], + "type": "inline_equation", + "content": "(+d_{4})" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 70, + 415, + 79 + ], + "lines": [ + { + "bbox": [ + 365, + 70, + 415, + 79 + ], + "spans": [ + { + "bbox": [ + 365, + 70, + 415, + 79 + ], + "type": "text", + "content": "Relit-5 " + }, + { + "bbox": [ + 365, + 70, + 415, + 79 + ], + "type": "inline_equation", + "content": "(+d_{5})" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 70, + 477, + 79 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 477, + 79 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 477, + 79 + ], + "type": "text", + "content": "Relit-6 " + }, + { + "bbox": [ + 427, + 70, + 477, + 79 + ], + "type": "inline_equation", + "content": "(+d_{6})" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 490, + 70, + 539, + 79 + ], + "lines": [ + { + "bbox": [ + 490, + 70, + 539, + 79 + ], + "spans": [ + { + "bbox": [ + 490, + 70, + 539, + 79 + ], + "type": "text", + "content": "Relit-7 " + }, + { + "bbox": [ + 490, + 70, + 539, + 79 + ], + "type": "inline_equation", + "content": "(+d_{7})" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 79, + 547, + 388 + ], + "lines": [ + { + "bbox": [ + 49, + 79, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 49, + 79, + 547, + 388 + ], + "type": "image", + "image_path": "7166d8d77f767b023bc10ba80437020a9bbcd67714d16b616b5c6ea92a4fd88f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 390, + 546, + 456 + ], + "lines": [ + { + "bbox": [ + 46, + 390, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 546, + 456 + ], + "type": "text", + "content": "Figure 5. First column images generated by the original StyleGAN. Other columns show images obtained from " + }, + { + "bbox": [ + 46, + 390, + 546, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{+} + \\mathbf{d}_{i}" + }, + { + "bbox": [ + 46, + 390, + 546, + 456 + ], + "type": "text", + "content": ", our relighting directions added to the style codes " + }, + { + "bbox": [ + 46, + 390, + 546, + 456 + ], + "type": "inline_equation", + "content": "(\\mathbf{w}^{+})" + }, + { + "bbox": [ + 46, + 390, + 546, + 456 + ], + "type": "text", + "content": " of the images in the first column. These directions have been chosen to fix albedo, but change shading. Note: each row shows the same scenes but with different illumination. Lighting varies aggressively, and the individual latent direction " + }, + { + "bbox": [ + 46, + 390, + 546, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{i}" + }, + { + "bbox": [ + 46, + 390, + 546, + 456 + ], + "type": "text", + "content": " has persistent semantics - each column corresponds to a type of illumination. For example, the second " + }, + { + "bbox": [ + 46, + 390, + 546, + 456 + ], + "type": "inline_equation", + "content": "(d_{1})" + }, + { + "bbox": [ + 46, + 390, + 546, + 456 + ], + "type": "text", + "content": " and third " + }, + { + "bbox": [ + 46, + 390, + 546, + 456 + ], + "type": "inline_equation", + "content": "(d_{2})" + }, + { + "bbox": [ + 46, + 390, + 546, + 456 + ], + "type": "text", + "content": " column switches on bedside lamps. It is worth noting the presence of soft shadows, cast shadows, inter-reflections, and glossy changes across relights. Another important observation to note is that all relightings are with respect to world coordinates, not camera coordinates." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 58, + 460, + 533, + 589 + ], + "blocks": [ + { + "bbox": [ + 58, + 460, + 533, + 589 + ], + "lines": [ + { + "bbox": [ + 58, + 460, + 533, + 589 + ], + "spans": [ + { + "bbox": [ + 58, + 460, + 533, + 589 + ], + "type": "image", + "image_path": "d1bc21dc64fd1c6702e3e66077b3aaa869db1ebc164fc3a4c97c19c1a94e7b06.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 591, + 546, + 647 + ], + "lines": [ + { + "bbox": [ + 46, + 591, + 546, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 546, + 647 + ], + "type": "text", + "content": "Figure 6. Resurfacing Generated Images. Instead of relighting images, we can generate resurfaced images by swapping our consistency and diversity loss. We apply diversity loss to change the albedo and consistency loss to maintain the shading and global illumination. The first column shows images generated by the original StyleGAN, and the other columns show images obtained from " + }, + { + "bbox": [ + 46, + 591, + 546, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{+} + \\mathbf{d}_{i}" + }, + { + "bbox": [ + 46, + 591, + 546, + 647 + ], + "type": "text", + "content": " for our resurfacing or recoloring directions. Each column shows the same scene as in the first column, but with varying surface colors and materials, while the individual latent direction " + }, + { + "bbox": [ + 46, + 591, + 546, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i" + }, + { + "bbox": [ + 46, + 591, + 546, + 647 + ], + "type": "text", + "content": " retains its semantics." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 657, + 289, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 657, + 289, + 706 + ], + "spans": [ + { + "bbox": [ + 46, + 657, + 289, + 706 + ], + "type": "text", + "content": "comparable method. However, we show relighting comparisons to a recent SOTA method that is physically motivated and trained with CGI data [26] by using a SOTA inversion method [5] (Fig. 14 in Supplementary). For relighting, our" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 657, + 547, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 657, + 547, + 706 + ], + "spans": [ + { + "bbox": [ + 305, + 657, + 547, + 706 + ], + "type": "text", + "content": "method should generate images that: are clearly relightings of a scene; fix geometry and albedo but visibly change shading; and display complicated illumination effects, including soft shadows, cast shadows and gloss. For resurfacing, our" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4236" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 533, + 202 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 533, + 202 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 533, + 202 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 533, + 202 + ], + "type": "image", + "image_path": "2fd4b9a12e44e0c1910ee4e27fdfba0f92dd8940b5ac0a140ab0b22371286c64.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 204, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 547, + 248 + ], + "type": "text", + "content": "Figure 7. Controllability: The first and last columns of the figure show relighted images generated using our Relit-1 and Relit-5 directions. The bottom section of the figure features a user-controllable slider that enables adjusting the weight of the relighting effects produced by these two directions. Moving the slider from left to right results in a seamless interpolation between the two lighting directions and provides precise control over the relighting of the generated images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 253, + 536, + 385 + ], + "blocks": [ + { + "bbox": [ + 58, + 253, + 536, + 385 + ], + "lines": [ + { + "bbox": [ + 58, + 253, + 536, + 385 + ], + "spans": [ + { + "bbox": [ + 58, + 253, + 536, + 385 + ], + "type": "image", + "image_path": "ea7eef6570ef1ec9bd1a3fc1856d73f98a7de1090b9b33827b3fb51fa5f55fae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 388, + 546, + 433 + ], + "lines": [ + { + "bbox": [ + 46, + 388, + 546, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 546, + 433 + ], + "type": "text", + "content": "Figure 8. Scaling Directions: The figure depicts the persistent and smooth effects of applying the direction at different scalar coefficients. We use our Relit-2 direction. A slider at the bottom allows the user to adjust the weight of the relighting direction, producing a seamless interpolation when increasing or decreasing the intensity of the chosen direction. The relighting effects range from a well-lit room with the bedside lamp off to weak external lighting with a bedside lamp on." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 444, + 289, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 289, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 289, + 575 + ], + "type": "text", + "content": "method should generate images that: are clearly images of the original layout, but with different materials or colors or changes in furniture; and display illumination effects that are consistent with these color changes. As Figure 6 shows, our method meets these goals. Figure 7 and Figure 8 show interpolation sequences for a relighting between two directions and scaling only one direction. Note that the lighting changes smoothly, as one would expect. Figure 15 in Supplementary shows that our relighting and recoloring directions are largely disentangled. An ablation for individual losses is also given in Supplementary Figure 13." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 576, + 288, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 576, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 288, + 647 + ], + "type": "text", + "content": "Quantitative Evaluation: Figure 3 shows how we can evaluate albedo and lighting shifts. In Table 1 we show we can generate image datasets with increased FID [18, 34] (clean-FID) from the base comparison set. This is strong evidence our method can produce a set of images that is a strict superset of those that the vanilla StyleGAN can produce." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 647, + 288, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 288, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 288, + 684 + ], + "type": "text", + "content": "Generality: We have applied our method to StyleGAN trained on Conference Room, Kitchen, Living Room, Dining Room, Church, and Face datasets (results in Figure 10)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 289, + 714 + ], + "type": "text", + "content": "Comparison to GAN-control: GAN-control (GC) [39] represents lighting with a spherical harmonic predictor pre" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 490, + 544, + 544 + ], + "blocks": [ + { + "bbox": [ + 304, + 435, + 545, + 490 + ], + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 490 + ], + "type": "text", + "content": "Table 1. FID measures distribution shift and not realism. Our generated images are realistic and are out-of-distribution because of large illumination and color changes in the images. This results in large FID scores. " + }, + { + "bbox": [ + 304, + 435, + 545, + 490 + ], + "type": "inline_equation", + "content": "KDL" + }, + { + "bbox": [ + 304, + 435, + 545, + 490 + ], + "type": "text", + "content": " in the table is for kitchen, dining and living room which are jointly trained [13]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 490, + 544, + 544 + ], + "lines": [ + { + "bbox": [ + 307, + 490, + 544, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 544, + 544 + ], + "type": "table", + "html": "
TypeBedroomKDLConferenceChurchFaces
StyleGAN (SG)5.015.869.353.805.02
SG + Relighting (RL)14.236.8710.4812.1237.87
SG + Resurfacing (RS)17.039.4110.6318.6034.06
SG + RL + RS21.3911.6812.7121.0837.40
", + "image_path": "53cbc0b29954b52114849d5ddae759d90d3ec9e264e16a1f6efdee545ce62eb0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": "trained on a parameterized 3D face reconstruction model [9]. This model does not apply to indoor lighting (among other problems, it predicts all points with the same normal have the same shading). We trained a GC model on the LSUN Bedroom dataset with 2 subspaces " + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "inline_equation", + "content": "z^k" + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": " - illumination corresponding to spherical harmonic coefficients, and other to represent all other structural information. The model was trained for 800 epochs. GC produces images whose structure varies wildly with any lighting changes, resulting in large albedo changes (GC in Figures 3, 9). The difficulty appears to be that the attribute predictor is easily subverted; if the lighting representation cannot produce an image that is (say) dark on the left side, the albedo is adjusted instead." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4237" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 276, + 121 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 276, + 121 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 276, + 121 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 276, + 121 + ], + "type": "image", + "image_path": "616e562904dbfc4791c0582ec40a42c8d2ed3c14537c46285ea6b0cee9bcadd9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 124, + 287, + 146 + ], + "lines": [ + { + "bbox": [ + 47, + 124, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 47, + 124, + 287, + 146 + ], + "type": "text", + "content": "Figure 9. GAN-control (GC) [39] cannot relight complex scenes like bedrooms; it completely changes the scene's layout." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 154, + 164, + 262 + ], + "blocks": [ + { + "bbox": [ + 53, + 154, + 164, + 262 + ], + "lines": [ + { + "bbox": [ + 53, + 154, + 164, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 154, + 164, + 262 + ], + "type": "image", + "image_path": "136a1a252d87a1153f2c7be876b8d88a7fcd64fcff843b5203840378e6914d27.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 54, + 264, + 162, + 383 + ], + "blocks": [ + { + "bbox": [ + 54, + 264, + 162, + 383 + ], + "lines": [ + { + "bbox": [ + 54, + 264, + 162, + 383 + ], + "spans": [ + { + "bbox": [ + 54, + 264, + 162, + 383 + ], + "type": "image", + "image_path": "4ee88bae8989c9c8a895b5d15a0e25021316baa5246f7afedb1db4702bcb771b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 384, + 154, + 393 + ], + "lines": [ + { + "bbox": [ + 63, + 384, + 154, + 393 + ], + "spans": [ + { + "bbox": [ + 63, + 384, + 154, + 393 + ], + "type": "text", + "content": "(c) Living Room Relighting" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 53, + 394, + 164, + 502 + ], + "blocks": [ + { + "bbox": [ + 53, + 394, + 164, + 502 + ], + "lines": [ + { + "bbox": [ + 53, + 394, + 164, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 394, + 164, + 502 + ], + "type": "image", + "image_path": "41ebdd2d30f388edfd89eb5765d19b1696df76dcf09f9c25c465b7abb7791c2c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 504, + 161, + 513 + ], + "lines": [ + { + "bbox": [ + 56, + 504, + 161, + 513 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 161, + 513 + ], + "type": "text", + "content": "(e) Relighting outdoor Churches" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 514, + 287, + 536 + ], + "lines": [ + { + "bbox": [ + 47, + 514, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 287, + 536 + ], + "type": "text", + "content": "Figure 10. StyLitGAN extends to finding relighting directions for StyleGANs trained on other datasets." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 172, + 154, + 282, + 262 + ], + "blocks": [ + { + "bbox": [ + 172, + 154, + 282, + 262 + ], + "lines": [ + { + "bbox": [ + 172, + 154, + 282, + 262 + ], + "spans": [ + { + "bbox": [ + 172, + 154, + 282, + 262 + ], + "type": "image", + "image_path": "345544a3e5e24f28b193a84fae75a982292eb0e373c4b94c3494e93f648e7955.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 172, + 274, + 281, + 382 + ], + "blocks": [ + { + "bbox": [ + 189, + 264, + 263, + 274 + ], + "lines": [ + { + "bbox": [ + 189, + 264, + 263, + 274 + ], + "spans": [ + { + "bbox": [ + 189, + 264, + 263, + 274 + ], + "type": "text", + "content": "(b) Kitchen Relighting" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 172, + 274, + 281, + 382 + ], + "lines": [ + { + "bbox": [ + 172, + 274, + 281, + 382 + ], + "spans": [ + { + "bbox": [ + 172, + 274, + 281, + 382 + ], + "type": "image", + "image_path": "fdeb4c976808fe9c8c6d5bfed230d28b807fe438aab779b4a93459b9a4805f28.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 172, + 394, + 282, + 502 + ], + "blocks": [ + { + "bbox": [ + 181, + 384, + 272, + 393 + ], + "lines": [ + { + "bbox": [ + 181, + 384, + 272, + 393 + ], + "spans": [ + { + "bbox": [ + 181, + 384, + 272, + 393 + ], + "type": "text", + "content": "(d) Dining Room Relighting" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 172, + 394, + 282, + 502 + ], + "lines": [ + { + "bbox": [ + 172, + 394, + 282, + 502 + ], + "spans": [ + { + "bbox": [ + 172, + 394, + 282, + 502 + ], + "type": "image", + "image_path": "86a2438cb584c617b683c965ca94228ad9a05d7ece2cf54a21e1e3ca5fad0ccf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 504, + 261, + 514 + ], + "lines": [ + { + "bbox": [ + 190, + 504, + 261, + 514 + ], + "spans": [ + { + "bbox": [ + 190, + 504, + 261, + 514 + ], + "type": "text", + "content": "(f) Portrait Relighting" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 548, + 195, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 195, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 195, + 562 + ], + "type": "text", + "content": "6. Downstream Applications" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "Lighting Variance in Surface Normal Prediction: The Multilum dataset [29] provides images of 1000 various indoor scenes, each under 25 lighting conditions, physically relit and captured. A SOTA normal predictor (Omnidata [12, 21]) applied to the test set produces surface normal predictions that vary significantly with changes of light in a fixed scene (Figure 11, purple bar). Finetuning Omnidata with the Multilum training dataset significantly reduces this variance (Figure 11, orange bar); but multiple lightings of a fixed scene are very hard to find. Finetuning with StyLitGAN relights produces comparable improvements; using seven distinct relights (Figure 11, pale orange bar) is slightly" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 337, + 62, + 512, + 216 + ], + "blocks": [ + { + "bbox": [ + 337, + 62, + 512, + 216 + ], + "lines": [ + { + "bbox": [ + 337, + 62, + 512, + 216 + ], + "spans": [ + { + "bbox": [ + 337, + 62, + 512, + 216 + ], + "type": "image", + "image_path": "b18315a4c356fb926ef337ffe428642de1853a27d7aba075bb8b6ab8715c80e0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 219, + 548, + 319 + ], + "lines": [ + { + "bbox": [ + 305, + 219, + 548, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 219, + 548, + 319 + ], + "type": "text", + "content": "Figure 11. Normal variance to lighting reduces when fine-tuning on our relighting dataset. Purple boxplot shows normal variance under relighting for real test scenes from Multilum using the Omnidata normal predictor; orange shows the result of finetuning using Multilum training data; light orange and yellow show the result of finetuning using StyLitGAN images (7 and 25 per scene respectively). The measure is angular error in radians from the mean prediction of a scene for each relit image in the Multilum test set (30 scenes, 25 lightings each)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 332, + 325, + 518, + 464 + ], + "blocks": [ + { + "bbox": [ + 332, + 325, + 518, + 464 + ], + "lines": [ + { + "bbox": [ + 332, + 325, + 518, + 464 + ], + "spans": [ + { + "bbox": [ + 332, + 325, + 518, + 464 + ], + "type": "image", + "image_path": "648cf2ffc644a4acbef8476ef546b65980865d6af2bf979b6e9b5ecb93fd8ac4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 465, + 547, + 510 + ], + "lines": [ + { + "bbox": [ + 305, + 465, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 465, + 547, + 510 + ], + "type": "text", + "content": "Figure 12. A surface normal predictor is finetuned to increase prediction consistency for relit images of the same scene. General performance shown on Taskonomy test images parallels that of the original model. Mean across buildings is given in the last column." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 518, + 547, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 547, + 613 + ], + "type": "text", + "content": "worse than using 25 distinct relights (Figure 11, yellow bar). The resulting improvement is not at the cost of base accuracy. Figure 12 compares the accuracy of various finetuned models on the Taskonomy test set [50] (recall this involves thousands of frames each in 10 blocks; we show results by block). Note that finetuned methods mostly show slight accuracy improvements over the base model, but losses in some blocks result in means that match." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 613, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 547, + 639 + ], + "type": "text", + "content": "In our follow-up work, we also show that StyleGAN \"knows\" intrinsic images and can be easily extracted [4]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 648, + 399, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 648, + 399, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 648, + 399, + 662 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "We thank Aniruddha Kembhavi, Derek Hoiem, Min Jin Chong and Shenlong Wang. This material is based upon work supported by the National Science Foundation under Grant No. 2106825 and by a gift from Boeing Corporation." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4238" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 144 + ], + "type": "text", + "content": "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 144, + 288, + 177 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 144, + 288, + 177 + ], + "spans": [ + { + "bbox": [ + 53, + 144, + 288, + 177 + ], + "type": "text", + "content": "[2] Peter N Belhumeur and David J Kriegman. What is the set of images of an object under all possible illumination conditions? International Journal of Computer Vision, 1998. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 177, + 288, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 177, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 53, + 177, + 288, + 219 + ], + "type": "text", + "content": "[3] Anand Bhattachad and D.A. Forsyth. Cut-and-paste object insertion by enabling deep image prior for reshading. In 2022 International Conference on 3D Vision (3DV). IEEE, 2022. 2, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 220, + 288, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 288, + 261 + ], + "type": "text", + "content": "[4] Anand Bhattachad, Daniel McKee, Derek Hoiem, and David Forsyth. Stylegan knows normal, depth, albedo, and more. Advances in Neural Information Processing Systems, 36, 2024. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 288, + 305 + ], + "type": "text", + "content": "[5] Anand Bhattachad, Viraj Shah, Derek Hoiem, and David A Forsyth. Make it so: Steering stylegan for any image inversion and editing. arXiv preprint arXiv:2304.14403, 2023. 6, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 305, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 305, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 305, + 288, + 360 + ], + "type": "text", + "content": "[6] Prashanth Chandran, Sebastian Winberg, Gaspard Zoss, Jérémy Riviere, Markus Gross, Paulo Gotardo, and Derek Bradley. Rendering with style: combining traditional and neural approaches for high-quality face rendering. ACM Transactions on Graphics (ToG), 40(6):1-14, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 360, + 288, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 360, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 360, + 288, + 380 + ], + "type": "text", + "content": "[7] Min Jin Chong and David Forsyth. Jojogan: One shot face stylization. arXiv preprint arXiv:2112.11641, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "type": "text", + "content": "[8] Min Jin Chong, Hsin-Ying Lee, and David Forsyth. Stylegan of all trades: Image manipulation with only pretrained stylegan. arXiv preprint arXiv:2111.01619, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 413, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 464 + ], + "type": "text", + "content": "[9] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 465, + 288, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 288, + 518 + ], + "type": "text", + "content": "[10] Aditya Deshpande, Jiajun Lu, Mao-Chuang Yeh, Min Jin Chong, and David Forsyth. Learning diverse image colorization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6837-6845, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 520, + 288, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 288, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 288, + 563 + ], + "type": "text", + "content": "[11] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 341-346, 2001. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 563, + 288, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 563, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 48, + 563, + 288, + 616 + ], + "type": "text", + "content": "[12] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 616, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 616, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 616, + 288, + 658 + ], + "type": "text", + "content": "[13] Dave Epstein, Taesung Park, Richard Zhang, Eli Shechtman, and Alexei A Efros. Blobgan: Spatially disentangled scene representations. arXiv preprint arXiv:2205.02837, 2022. 2, 3, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 288, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 681 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 681 + ], + "type": "text", + "content": "[14] D.A. Forsyth and Jason J Rock. Intrinsic image decomposition using paradigms. TPAMI, 2022 in press. 2, 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 547, + 704 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 326, + 74, + 498, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 498, + 83 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 498, + 83 + ], + "type": "text", + "content": "pattern recognition, pages 2414-2423, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 83, + 546, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 83, + 546, + 126 + ], + "spans": [ + { + "bbox": [ + 308, + 83, + 546, + 126 + ], + "type": "text", + "content": "[16] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. arXiv preprint arXiv:1406.2661, 2014. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 126, + 547, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 126, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 308, + 126, + 547, + 169 + ], + "type": "text", + "content": "[17] Aaron Hertzmann, Charles E Jacobs, Nuria Oliver, Brian Curless, and David H Salesin. Image analogies. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 327-340, 2001. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 169, + 547, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 169, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 308, + 169, + 547, + 212 + ], + "type": "text", + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint arXiv:1706.08500, 2017. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 212, + 547, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 212, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 308, + 212, + 547, + 255 + ], + "type": "text", + "content": "[19] Le Hui, Jia Yuan, Mingmei Cheng, Jin Xie, Xiaoya Zhang, and Jian Yang. Superpoint network for point cloud oversegmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5510-5519, 2021. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 255, + 546, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 255, + 546, + 287 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 546, + 287 + ], + "type": "text", + "content": "[20] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, 2016. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 287, + 547, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 547, + 339 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 547, + 339 + ], + "type": "text", + "content": "[21] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 2, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 340, + 546, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 340, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 340, + 546, + 384 + ], + "type": "text", + "content": "[22] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 383, + 546, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 546, + 426 + ], + "type": "text", + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 426, + 546, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 546, + 459 + ], + "type": "text", + "content": "[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 459, + 546, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 459, + 546, + 479 + ], + "spans": [ + { + "bbox": [ + 308, + 459, + 546, + 479 + ], + "type": "text", + "content": "[25] Edwin H Land. The retina theory of color vision. Scientific american, 1977. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 479, + 546, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 546, + 533 + ], + "type": "text", + "content": "[26] Zhengqin Li, Jia Shi, Sai Bi, Rui Zhu, Kalyan Sunkavalli, Miloš Hasan, Zexiang Xu, Ravi Ramamoorthi, and Manmohan Chandraker. Physically-based editing of indoor scene lighting from a single image. arXiv preprint arXiv:2205.09343, 2022. 6, 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 533, + 546, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 546, + 575 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 546, + 575 + ], + "type": "text", + "content": "[27] Zicheng Liao, Hugues Hoppe, David Forsyth, and Yizhou Yu. A subdivision-based representation for vector image editing. IEEE transactions on visualization and computer graphics, 2012. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 575, + 546, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 575, + 546, + 619 + ], + "spans": [ + { + "bbox": [ + 308, + 575, + 546, + 619 + ], + "type": "text", + "content": "[28] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. Advances in Neural Information Processing Systems, 34, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 619, + 546, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 619, + 546, + 662 + ], + "spans": [ + { + "bbox": [ + 308, + 619, + 546, + 662 + ], + "type": "text", + "content": "[29] Lukas Murmann, Michael Gharbi, Miika Aittala, and Fredo Durand. A multi-illumination dataset of indoor object appearance. In 2019 IEEE International Conference on Computer Vision (ICCV), Oct 2019. 2, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 662, + 546, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 662, + 546, + 704 + ], + "spans": [ + { + "bbox": [ + 308, + 662, + 546, + 704 + ], + "type": "text", + "content": "[30] Thomas Nestmeyer, Jean-François Lalonde, Iain Matthews, Epic Games, Andreas Lehrmann, and AI Borealis. Learning physics-guided face relighting under directional light. 2020. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4239" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 706 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 126 + ], + "type": "text", + "content": "[31] Xingang Pan, Xudong Xu, Chen Change Loy, Christian Theobalt, and Bo Dai. A shading-guided generative implicit model for shape-accurate 3d-aware image synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 126, + 288, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 126, + 288, + 181 + ], + "spans": [ + { + "bbox": [ + 48, + 126, + 288, + 181 + ], + "type": "text", + "content": "[32] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 181, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 181, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 48, + 181, + 288, + 224 + ], + "type": "text", + "content": "[33] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 224, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 224, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 224, + 288, + 255 + ], + "type": "text", + "content": "[34] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 255, + 288, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 288, + 289 + ], + "type": "text", + "content": "[35] Erik Reinhard, Michael Adhikhmin, Bruce Gooch, and Peter Shirley. Color transfer between images. IEEE Computer graphics and applications, 21(5):34-41, 2001. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 289, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 289, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 288, + 342 + ], + "type": "text", + "content": "[36] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 342, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 384 + ], + "type": "text", + "content": "[37] Soumyadip Sengupta, Brian Curless, Ira Kemelmacher-Shlizerman, and Steven M Seitz. A light stage on every desk. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 384, + 288, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 384, + 288, + 427 + ], + "spans": [ + { + "bbox": [ + 48, + 384, + 288, + 427 + ], + "type": "text", + "content": "[38] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 427, + 288, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 427, + 288, + 471 + ], + "spans": [ + { + "bbox": [ + 48, + 427, + 288, + 471 + ], + "type": "text", + "content": "[39] Alon Shoshan, Nadav Bhonker, Igor Kviatkovsky, and Gerard Medioni. Gan-control: Explicitly controllable gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14083-14093, 2021. 2, 5, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 288, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 501 + ], + "type": "text", + "content": "[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. *ICLR*, 2015. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 502, + 288, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 288, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 288, + 545 + ], + "type": "text", + "content": "[41] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 288, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 288, + 598 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 288, + 598 + ], + "type": "text", + "content": "[42] Feitong Tan, Sean Fanello, Abhimitra Meka, Sergio Orts-Escolano, Danhang Tang, Rohit Pandey, Jonathan Taylor, Ping Tan, and Yinda Zhang. Volux-gan: A generative model for 3d face synthesis with hdri relighting. arXiv preprint arXiv:2201.04873, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 598, + 288, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 598, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 48, + 598, + 288, + 664 + ], + "type": "text", + "content": "[43] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan for 3d control over portrait images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6142-6151, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 664, + 288, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 664, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 48, + 664, + 288, + 696 + ], + "type": "text", + "content": "[44] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 696, + 288, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 696, + 288, + 706 + ], + "spans": [ + { + "bbox": [ + 48, + 696, + 288, + 706 + ], + "type": "text", + "content": "[45] Andrey Voynov and Artem Babenko. Unsupervised discovery" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 502 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "type": "text", + "content": "of interpretable directions in the gan latent space. In International conference on machine learning, pages 9786-9796. PMLR, 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 105, + 547, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 105, + 547, + 159 + ], + "spans": [ + { + "bbox": [ + 307, + 105, + 547, + 159 + ], + "type": "text", + "content": "[46] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 159, + 547, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 159, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 159, + 547, + 191 + ], + "type": "text", + "content": "[47] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. arXiv preprint arXiv: 2101.05278, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 191, + 547, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 191, + 547, + 233 + ], + "spans": [ + { + "bbox": [ + 307, + 191, + 547, + 233 + ], + "type": "text", + "content": "[48] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. International Journal of Computer Vision, 2020. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 233, + 547, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 233, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 307, + 233, + 547, + 287 + ], + "type": "text", + "content": "[49] Ning Yu, Guilin Liu, Aysegul Dundar, Andrew Tao, Bryan Catanzaro, Larry S Davis, and Mario Fritz. Dual contrastive loss and attention for gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6731-6742, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 287, + 547, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 547, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 547, + 341 + ], + "type": "text", + "content": "[50] Amir R. Zamir, Alexander Sax, William B. Shen, Leonidas J. Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2018. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 341, + 547, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 341, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 341, + 547, + 373 + ], + "type": "text", + "content": "[51] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 373, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 373, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 373, + 547, + 415 + ], + "type": "text", + "content": "[52] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 415, + 547, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 547, + 459 + ], + "type": "text", + "content": "[53] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 459, + 547, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 459, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 459, + 547, + 502 + ], + "type": "text", + "content": "[54] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, 2017. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4240" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Style Aligned Image Generation via Shared Attention/0cae3168-e63f-4f71-908a-5bdeae8d8200_content_list.json b/2024/Style Aligned Image Generation via Shared Attention/0cae3168-e63f-4f71-908a-5bdeae8d8200_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..16f14106961b63f69b43d964ad4db76658042959 --- /dev/null +++ b/2024/Style Aligned Image Generation via Shared Attention/0cae3168-e63f-4f71-908a-5bdeae8d8200_content_list.json @@ -0,0 +1,2463 @@ +[ + { + "type": "text", + "text": "Style Aligned Image Generation via Shared Attention", + "text_level": 1, + "bbox": [ + 214, + 130, + 756, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Amir Hertz\\*, Andrey Voynov\\*, Shlomi Fruchter\\*, and Daniel Cohen-Or\\*1,2 \n1 Google Research \n2 Tel Aviv University", + "bbox": [ + 173, + 179, + 795, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale Text-to-Image (T2I) models have rapidly gained prominence across creative fields, generating visually compelling outputs from textual prompts. However, controlling these models to ensure consistent style remains challenging, with existing methods necessitating fine-tuning and manual intervention to disentangle content and style. In this paper, we introduce StyleAligned, a novel technique designed to establish style alignment among a series of generated images. By employing minimal 'attention sharing' during the diffusion process, our method maintains style consistency across images within T2I models. This approach allows for the creation of style-consistent images using a reference style through a straightforward inversion operation. Our method's evaluation across diverse styles and text prompts demonstrates high-quality synthesis and fidelity, underscoring its efficacy in achieving consistent style across various inputs.", + "bbox": [ + 75, + 300, + 473, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 584, + 209, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale Text-to-Image (T2I) generative models [37, 39, 45] have emerged as an essential tool across creative disciplines, such as art, graphic design, animation, architecture, gaming and more. These models show tremendous capabilities of translating an input text into an appealing visual result that is aligned with the input description.", + "bbox": [ + 75, + 609, + 468, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "An envisioned application of T2I models revolves around the rendition of various concepts in a way that shares a consistent style and character, as though all were created by the same artist and method (see Fig. 1). While proficient in aligning with the textual description of the style, state-of-the-art T2I models often create images that diverge significantly in their interpretations of the same stylistic descriptor, as depicted in Fig. 2.", + "bbox": [ + 75, + 700, + 468, + 820 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent methods mitigate this by fine-tuning the T2I model over a set of images that share the same style [13,49]. This optimization is computationally expensive and usually", + "bbox": [ + 75, + 821, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/befe72dd08ea13e65a43000c1dc25189dcdb8c0c0bf181c36f354ee9aa00c776.jpg", + "image_caption": [ + "Figure 1. Style aligned image set generation. By fusing the features of the toy train image (left) during the diffusion process, we can generate an image set of different content that shares the style." + ], + "image_footnote": [], + "bbox": [ + 500, + 266, + 893, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "requires human input in order to find a plausible subset of images and texts that enables the disentanglement of content and style.", + "bbox": [ + 496, + 696, + 893, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce StyleAligned, a method that enables consistent style interpretation across a set of generated images (Fig. 1). Our method requires no optimization and can be applied to any attention-based text-to-image diffusion model. We show that adding minimal attention sharing operations along the diffusion process, from each generated image to the first one in a batch, leads to a style-consistent set. Moreover, using diffusion inversion, our method can be applied to generate style-consistent images given a reference style image, with no optimization or fine-tuning.", + "bbox": [ + 496, + 742, + 893, + 893 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 99, + 875, + 210, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Equal Advising.", + "bbox": [ + 99, + 887, + 197, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4775", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cca67eff1325ad439ed8bffe4784e7bbf26f6fce3d87078922c54c3404fd3b16.jpg", + "image_caption": [ + "Figure 2. Standard text-to-image vs. StyleAligned set generation. Given style description of \"minimal origami\", standard text-to-image generation (top) results with an unaligned image set while our method (bottom) can generate variety of style aligned content." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 468, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present our results over diverse styles and text prompts, demonstrating high-quality synthesis and fidelity to the prompts and reference style. We show diverse examples of generated images that share their style with a reference image that can possibly be a given input image. Importantly, our technique stands as a zero-shot solution, distinct from other personalization techniques, as it operates without any form of optimization or fine-tuning. For our code and more examples, please visit the project page style-aligned-gen.github.io", + "bbox": [ + 75, + 491, + 468, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 656, + 218, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-image generation. Text conditioned image generative models [7, 31, 38] show unprecedented capabilities of generating high quality images from text descriptions. In particular, T2I diffusion models [35, 38, 46] are pushing the state of the art and they are quickly adopted for different generative visual tasks like inpainting [2, 44], imaged-to-image translation [55, 59], local image editing [9, 24], subject-driven image generation [42, 51] and more.", + "bbox": [ + 75, + 681, + 468, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Attention control in diffusion models. Hertz et al. [17] have shown how cross and self-attention maps within the diffusion process determine the layout and content of the generated images. Moreover, they showed how the attention maps can be used for controlled image generation. Other studies have leveraged modifications in attention lay", + "bbox": [ + 75, + 801, + 468, + 892 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ers to enhance the fidelity or diversity of generated images [8, 34], or apply attention control for image editing [5, 12, 30, 32, 33, 53]. However, in contrast to prior approaches that primarily enable structure-preserving image editing, our method excels at generating images with diverse structures and content while maintaining a consistent style interpretation.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Style transfer. Transferring a style from a reference image to a target content image is well studied subject in computer graphics. Classic works [10, 11, 18, 26] rely on optimization of handcrafted features and texture resampling algorithms from an input texture image, combined with structure constrains of a content image. With the progress of deep learning research, another line of works utilizes deep neural priors for style transfer optimization using deep features of pre-trained networks [15, 52], or injecting attention features from a style image to a target one [1]. More related to our approach, Huang et al. [22] introduced a real time style transfer network based on Adaptive Instance Normalization layers (AdaIN) that are used to normalize deep features of a target image using deep features statistics of a reference style image. Follow-up works, employ the AdaIN layer for additional unsupervised learning tasks, like style-based image generation [25] and Image2Image translation [23, 29].", + "bbox": [ + 496, + 196, + 892, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "T2I personalization To generalize T2I over new visual concepts, several works developed different optimization techniques over a small collection of input images that share the same concept [13, 16, 42, 56]. In instances where the collection shares a consistent style, the acquired concept becomes the style itself, affecting subsequent generations. Most close to our work is StyleDrop [49], a style personalization method that relies on fine-tuning of light weight adapter layers [20] at the end of each attention block in a non-autoregressive generative text-to-image transformer [7]. StyleDrop can generate a set of images in the same style of by training the adapter layers over a collection of images that share the same style. However, it struggles to generate a consistent image set of different content when training on a single image.", + "bbox": [ + 496, + 468, + 892, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method can generate a consistent image set without optimization phase and without relying on several images for training. To skip the training phase, recent works developed dedicated personalization encoders [14, 27, 47, 58, 59] that can directly inject new priors from a single input image to the T2I model. However, these methods encounter challenges to disentangle style from content as they focus on generating the same subject as in the input image.", + "bbox": [ + 496, + 695, + 890, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method overview", + "text_level": 1, + "bbox": [ + 500, + 829, + 668, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the following section we start with an overview of the T2I diffusion process, and in particular the self-attention mechanism Sec. 3.1. We continue by present", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4776", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6bfec6c2924b31a525aed02432c3bbe38680437dfe55778583f7be18ac076611.jpg", + "image_caption": [ + "Figure 3. Style Aligned Diffusion. Generation of images with a style aligned to the reference image on the left. In each diffusion denoising step all the images, except the reference, perform a shared self-attention with the reference image." + ], + "image_footnote": [], + "bbox": [ + 76, + 84, + 872, + 349 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/40d62b020e5800a0fb2d06cae575f19a45f0d7e49c04264b0542055e71d15eec.jpg", + "image_caption": [ + "Figure 4. Shared attention layer. The target images attend to the reference image by applying AdaIN over their queries and keys using the reference queries and keys respectively. Then, we apply shared attention where the target features are updated by both the target values $V_{t}$ and the reference values $V_{r}$ ." + ], + "image_footnote": [], + "bbox": [ + 142, + 410, + 411, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing our attention-sharing operation within the self-attention layers that enable style aligned image set generation.", + "bbox": [ + 75, + 825, + 470, + 858 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 412, + 640, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models [19, 48] are generative latent variable models that aim to model a distribution $p_{\\theta}(x_0)$ that approximates the data distribution $q(x_0)$ and are easy to sample from. Diffusion models are trained to reverse the diffusion \"forward process\":", + "bbox": [ + 498, + 435, + 892, + 511 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {\\alpha_ {t}} x _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\epsilon \\sim N (0, I),\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 521, + 831, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t \\in [0, \\infty)$ and the values of $\\alpha_{t}$ are determined by a scheduler such that $\\alpha_{0} = 1$ and $\\lim_{t \\to \\infty} \\alpha_{t} = 0$ . During inference, we sample an image by gradually denoising an input noise image $x_{T} \\sim \\mathcal{N}(0, I)$ via the reverse process:", + "bbox": [ + 498, + 549, + 893, + 611 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = \\mu_ {t - 1} + \\sigma_ {t} z, z \\sim N (0, I),\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 622, + 815, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the value of $\\sigma_{t}$ is determined by the sampler and $\\mu_{t - 1}$ is given by", + "bbox": [ + 498, + 648, + 890, + 680 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {t - 1} = \\frac {\\sqrt {\\alpha_ {t - 1}} x _ {t}}{\\sqrt {\\alpha_ {t}}} + \\left(\\sqrt {1 - \\alpha_ {t - 1}} - \\frac {\\sqrt {1 - \\alpha_ {t}}}{\\sqrt {\\alpha_ {t}}}\\right) \\epsilon_ {\\theta} (x _ {t}, t),\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 689, + 890, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}(x_t,t)$ is the output of a diffusion model parameterized by $\\theta$", + "bbox": [ + 498, + 734, + 890, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, this process can be generalized for learning a marginal distribution using an additional input condition. That leads text-to-image diffusion models (T2I), where the output of the model $\\epsilon_{\\theta}(x_t,t,y)$ is conditioned on a text prompt $y$ .", + "bbox": [ + 498, + 765, + 890, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Attention in T2I diffusion models. State-of-the-art T2I diffusion models [4,35,46] employ a U-Net architecture [40] that consists of convolution layers and transformer attention blocks [54]. In these attention mechanisms, deep", + "bbox": [ + 498, + 840, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4777", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ba3e52212d6c2d820ad98b6bb7c6961b73723f21fa0dc4ed42a668f250f4f670.jpg", + "image_caption": [ + "Figure 5. Ablation study - qualitative comparison. Each pair of rows shows two sets of images generated by the same set of prompts \"...in minimal flat design illustration\" using different configurations of our method, and each row in a pair uses a different seed. Sharing the self-attention between all images in the set (bottom) results with some diversity loss (style collapse across many seeds) and content leakage within each set (colors from one image leak to another). Disabling the queries keys AdALN operation results with less consistent image sets compared to our full method (top) which keeps on both diversity between different sets and consistency within each set." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 472, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "image features $\\phi \\in \\mathbb{R}^{m\\times d_h}$ attend to each other via selfattention layers and to contextual text embedding via crossattention layers.", + "bbox": [ + 75, + 734, + 468, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our work operates at the self-attention layers where deep features are being updated by attending to each other. First, the features are projected into queries $Q \\in m \\times d_k$ , keys $K \\in m \\times d_k$ and values $V \\in m \\times d_h$ through learned linear layers. Then, the attention is computed by the scaled", + "bbox": [ + 76, + 781, + 470, + 857 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dot-product attention:", + "bbox": [ + 500, + 90, + 647, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}} V\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 116, + 844, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_{k}$ is the dimension of $Q$ and $K$ . Intuitively, each image feature is updated by a weighted sum of $V$ , where the weight depends on the correlation between the projected query $q$ and the keys $K$ . In practice, each self-attention layer consists of several attention heads, and then the residual is computed by concatenating and projecting the attention heads output back to the image feature space $d_{h}$ :", + "bbox": [ + 496, + 160, + 893, + 267 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\phi} = \\phi + \\text {M u l t i - H e a d - A t t e n t i o n} (\\phi).\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 277, + 813, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Style Aligned Image Set Generation", + "text_level": 1, + "bbox": [ + 500, + 306, + 808, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal of our method is to generate a set of images $\\mathcal{I}_1\\ldots \\mathcal{I}_n$ that are aligned with an input set of text prompts $y_{1}\\dots y_{n}$ and share a consistent style interpretation with each other. For example, see the garnered image set of toy objects in Fig. 3 that are style-aligned with each other and to the input text on top. A naive way to generate a style aligned image set of different content is to use the same style description in the text prompts. As can be seen at the bottom of Fig. 2, generating different images using a shared style description of \"in minimal origami style\" results in an unaligned set, since each image is unaware of the exact appearance of other images in the set during the generation process.", + "bbox": [ + 496, + 329, + 890, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key insight underlying our approach is the utilization of the self-attention mechanism to allow communication among various generated images. This is achieved by sharing attention layers across the generated images.", + "bbox": [ + 496, + 526, + 890, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, let $Q_{i}$ , $K_{i}$ , and $V_{i}$ be the queries, keys, and values, projected from deep features $\\phi_{i}$ of $\\mathcal{I}_{i}$ in the set, then, the attention update for $\\phi_{i}$ is given by:", + "bbox": [ + 496, + 585, + 890, + 632 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} \\left(Q _ {i}, K _ {1 \\dots n}, V _ {1 \\dots n}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 643, + 890, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} K _ {1 \\dots n} = \\left[ \\begin{array}{c} K _ {1} \\\\ K _ {2} \\\\ \\vdots \\\\ K _ {n} \\end{array} \\right] \\text {a n d} V _ {1 \\dots n} = \\left[ \\begin{array}{c} V _ {1} \\\\ V _ {2} \\\\ \\vdots \\\\ V _ {n} \\end{array} \\right]. \\text {H o w e v e r},\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 670, + 890, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "have noticed that by enabling full attention sharing, we may harm the quality of the generated set. As can be seen in Fig. 5 (bottom rows), full attention sharing results in content leakage among the images. For example, the unicorns got green paint from the garnered dino in the set. Moreover, full attention sharing results with less diverse sets of the same set of prompts, see the two sets in Fig. 5 in bottom rows compared to the sets above.", + "bbox": [ + 496, + 746, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To restrict the content leakage and allow diverse sets, we share the attention to only one image in the generated set", + "bbox": [ + 500, + 867, + 890, + 897 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4778", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/925b1dcfbe4f36bc77e342a38b71d524bffa46834c75e5708976954f11d2d13e.jpg", + "image_caption": [ + "Figure 6. Quantitative Comparison. We compare the results of the different methods (blue marks) and our ablation experiments (orange marks) in terms of text alignment (CLIP score) and set consistency (DINO embedding similarity)." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 454, + 251 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/247a44494f02213f2b2b7c91ffa2bee923897c85a965aa4f01f500d88bcf6752.jpg", + "table_caption": [ + "Table 1. User evaluation for style aligned image set generation. In each question, the user was asked to select between two image sets, Which is better in terms of style consistency and match to the text descriptions (see Sec. 4). We report the percentage of judgments in favor of StyleAligned over 800 answers (2400 in total)." + ], + "table_footnote": [], + "table_body": "
StyleDrop (unofficial MUSE)StyleDrop (SDXL)DreamBooth-LoRA (SDXL)
85.2 %67.1 %61.3%
", + "bbox": [ + 96, + 407, + 444, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(the first in the batch). That is, target image features $\\phi_t$ are attending to themselves and to the features of only one reference image in the set using Eq. 1. As can be seen in Fig. 5 (middle), sharing the attention to only one image in the set results in diverse sets that share a similar style. However, in that case, we have noticed that the style of different images is not well aligned. We suspect that this is due to low attention flow from the reference to the target image.", + "bbox": [ + 75, + 469, + 468, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 4, to enable balanced attention reference, we normalize the queries $Q_{t}$ and keys $K_{t}$ of the target image using the queries $Q_{r}$ and keys $K_{r}$ of the reference image using the adaptive normalization operation (AdaIN) [22]:", + "bbox": [ + 75, + 590, + 468, + 666 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Q} _ {t} = \\operatorname {A d a I N} \\left(Q _ {t}, Q _ {r}\\right) \\quad \\hat {K} _ {t} = \\operatorname {A d a I N} \\left(K _ {t}, K _ {r}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 676, + 431, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the AdaIn operation is given by:", + "bbox": [ + 76, + 705, + 336, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A d a I N} (x, y) = \\sigma (y) \\left(\\frac {x - \\mu (x)}{\\sigma (x)}\\right) + \\mu_ {y},\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 732, + 413, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $\\mu (x),\\sigma (x)\\in \\mathbb{R}^{d_k}$ are the mean and the standard deviation of queries and keys across different pixels. Finally, our shared attention is given by", + "bbox": [ + 75, + 776, + 468, + 823 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} \\left(\\hat {Q} _ {t}, K _ {r t} ^ {T}, V _ {r t}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 832, + 357, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $K_{rt} = \\left[ \\begin{array}{c}K_{r}\\\\ \\hat{K}_{t} \\end{array} \\right]$ and $V_{rt} = \\left[ \\begin{array}{c}V_{r}\\\\ V_{t} \\end{array} \\right]$", + "bbox": [ + 76, + 861, + 326, + 896 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Evaluations and Experiments", + "text_level": 1, + "bbox": [ + 500, + 89, + 769, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We have implemented our method over Stable Diffusion XL (SDXL) [35] by applying our attention sharing overall 70 self-attention layers of the model. The generation of a four images set takes 29 seconds on a single A100 GPU. Notice that since the generation of the reference image is not influenced by other images in the batch, we can generate larger sets by fixing the prompt and seed of the reference image across the set generation.", + "bbox": [ + 496, + 114, + 890, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For example, see the sets in Fig. 2 and 3.", + "bbox": [ + 517, + 234, + 792, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation set. With the support of ChatGPT, we have generated 100 text prompts describing different image styles over four random objects. For example, “{A guitar, A hot air balloon, A sailboat, A mountain} in papercut art style.” For each style and set of objects, we use our method to generate a set of images. The full list of prompts is provided in the appendix.", + "bbox": [ + 496, + 251, + 890, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. To verify that each generated image contains its specified object, we measure the CLIP similarity [36] between the image and the text description of the object. In addition, we evaluate the style consistency of each generated set, by measuring the pairwise average cosine similarity between DINO VIT-B/8 [6] embeddings of the generated images in each set. Following [41, 56], we used DINO embeddings instead of CLIP image embeddings for measuring image similarity, since CLIP was trained with class labels and therefore it might give a high score for different images in the set that have similar content but with a different style. On the other hand, DINO better distinguishes between different styles due to its self-supervised training.", + "bbox": [ + 496, + 356, + 890, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 560, + 653, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The quantitative results are summarized in Fig. 6, where the right-top place on the chart means better text similarity and style consistency, respectively. As a reference, we report the score obtained by generating the set of images using SDXL (T2I Reference) using the same seeds without any intervention. As can be seen, our method achieves a much higher style consistency score at the expense of text similarity. See qualitative comparison in Fig. 2", + "bbox": [ + 496, + 583, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, we compared our method to additional two variants of the shared attention as described in Sec. 3.2. The first variant uses full attention sharing (Full Attn. Share) where the keys and values are shared between each pair of images in the set. In the second variant (W.A. AdaIN) we omit the AdaIN operation over queries and keys. As expected, this Full Attn. Share variant, results with higher style consistency and lower text alignment. As can be seen in Fig. 5, Full Attn. Share harms the overall quality of the image sets and diversity across sets. Moreover, our method without the use of AdaIN results in much lower style consistency. Qualitative results can be seen in Fig. 5 and in the supplementary..", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4779", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/90b4c6812a23bc64e0005d692fb5d98823a2a1d002f9f220cbd0b2c2e09b02ba.jpg", + "image_caption": [ + "Reference image", + "A guitar" + ], + "image_footnote": [], + "bbox": [ + 86, + 102, + 200, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9182dc023ae4f02985381ae8250dd67f27a32d6b2304ba17eb545d23e84d78b2.jpg", + "image_caption": [ + "StyleAligned", + "A hot air balloon" + ], + "image_footnote": [], + "bbox": [ + 205, + 103, + 320, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5aca3abfb4d09a6d3e5db3cba5ca2de4f2cc89a6531e18621a8d0d39d20800c7.jpg", + "image_caption": [ + "A mountain" + ], + "image_footnote": [], + "bbox": [ + 320, + 103, + 372, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c7948ce73d00ded74b3d54ff0bbc607a0b02b28b6b1d46b61271a0f2a65985d0.jpg", + "image_caption": [ + "StyleDrop (SDXL)", + "A hot air balloon" + ], + "image_footnote": [], + "bbox": [ + 375, + 102, + 488, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a5994e42c5975622b839fe5edb1ae76d3ccbb6ab0107ba916a23dc3bad6ba32b.jpg", + "image_caption": [ + "A mountain" + ], + "image_footnote": [], + "bbox": [ + 500, + 102, + 542, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5a0e47a4805c3521b0a200742e4d7a6efc924c7cb5a56dac468920f576d2922a.jpg", + "image_caption": [ + "StyleDrop (MUSE unofficial)", + "A hot air balloon" + ], + "image_footnote": [], + "bbox": [ + 547, + 103, + 658, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/46b6a33a75f26a0b999c095607dae2c6762844b5d675645eb58f683a47db58dc.jpg", + "image_caption": [ + "A mountain" + ], + "image_footnote": [], + "bbox": [ + 658, + 103, + 712, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cf7b7ed8e76a13b03bed044b822d2266a9b6f5900b9b0d0dd4546a02b814c5f7.jpg", + "image_caption": [ + "Dreambooth-LoRA (SDXL)", + "A hot air balloon" + ], + "image_footnote": [], + "bbox": [ + 715, + 103, + 828, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5df865f298975953ec22baa0012e6361e197e23bf7faab5934792aaa96698c2c.jpg", + "image_caption": [ + "A mountain" + ], + "image_footnote": [], + "bbox": [ + 831, + 103, + 883, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/20e2aaaf0c1e6ab47123702fef3e6d7744bc6664e05e999232247491622d8124.jpg", + "image_caption": [ + "A woman walking her dogs" + ], + "image_footnote": [], + "bbox": [ + 86, + 198, + 200, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2900abc58cb353417dbcacb0858899b10ac88bc6db44cd9e37deb49575279ad4.jpg", + "image_caption": [ + "A friendly robot", + "Cherryblossom" + ], + "image_footnote": [], + "bbox": [ + 205, + 198, + 320, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/91f969d04a9a14bea047a577c124bb468e71870f92e46a74591a2accfd49efbf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 320, + 198, + 372, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/934d4e99253578a4646ad35ef4a76dc408e1ce411b4fae9ab69fc49920db8ff8.jpg", + "image_caption": [ + "A friendly robot" + ], + "image_footnote": [], + "bbox": [ + 375, + 198, + 488, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cddaf90b91fa69e42670e5a911d7462cb4857ff150b046a76dfa7ec3e37516d5.jpg", + "image_caption": [ + "Cherryblossom" + ], + "image_footnote": [], + "bbox": [ + 490, + 198, + 542, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2ecc2accac53d7b2fc69a9b04e6409aabeb20377e722c3e746a29ad3763ec90b.jpg", + "image_caption": [ + "A robot" + ], + "image_footnote": [], + "bbox": [ + 549, + 198, + 658, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eb1f356d8d9dbb4832264617ad7181e028801061eb592887032ae1c4e14a004b.jpg", + "image_caption": [ + "Cherryblos" + ], + "image_footnote": [], + "bbox": [ + 658, + 198, + 712, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4a444f5f41709352b61c2f29db3900fe49fb1982d985bb2a52e3d55197d629c5.jpg", + "image_caption": [ + "A friendly robot", + "Cherryblossom" + ], + "image_footnote": [], + "bbox": [ + 715, + 198, + 828, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/610c9ba3314ca1226805491b5a9ce683756ab16a5b5f16aeebefaf025038cf13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 198, + 883, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8840c5026e15cd714d61ba2d56d55ab5f9290a381cb36c9d92013488ec9d9cd7.jpg", + "image_caption": [ + "A duck" + ], + "image_footnote": [], + "bbox": [ + 86, + 294, + 200, + 378 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/27d3adda1e9c1e8abb94cd5d3919471f7d4377e42f229c39b65589ea91c646c5.jpg", + "image_caption": [ + "A ship" + ], + "image_footnote": [], + "bbox": [ + 205, + 294, + 320, + 378 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5a84f2c0a5fbafa354a5977511fe2a587e3984e96500383873d23ee76b6ca112.jpg", + "image_caption": [ + "A rocket" + ], + "image_footnote": [], + "bbox": [ + 320, + 294, + 372, + 378 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7f110362dbb4193ed2c122926456ca32cf0ebff2b338a2f79b2ba1a19ea9abe1.jpg", + "image_caption": [ + "A ship" + ], + "image_footnote": [], + "bbox": [ + 375, + 294, + 488, + 378 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e719c9a00052dcead3a2edfa21f5306cf25728b1e523e32e19639193911e3f21.jpg", + "image_caption": [ + "A rocket" + ], + "image_footnote": [], + "bbox": [ + 490, + 294, + 542, + 378 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/520144230c3f6cda53ff73278bfc318dc00a4a7b4f25acc8f6aad269a22713e0.jpg", + "image_caption": [ + "A ship" + ], + "image_footnote": [], + "bbox": [ + 549, + 294, + 658, + 378 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f30b62b042d2e1a7bc31ea8397e082c8f27c08c36af31b9f8e439353ddb0b106.jpg", + "image_caption": [ + "A car" + ], + "image_footnote": [], + "bbox": [ + 658, + 294, + 712, + 378 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c120d4a94419eaaf18f09eb87f57ff0e51bc783cbfc3b7df039a20de156ed2fa.jpg", + "image_caption": [ + "A ship" + ], + "image_footnote": [], + "bbox": [ + 715, + 294, + 828, + 378 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9d94bbdfcfdd7eeb1b3052b31dc2b9ffcd4e5c7236d8b171d9ae33c032af0cff.jpg", + "image_caption": [ + "A rocket" + ], + "image_footnote": [], + "bbox": [ + 828, + 294, + 883, + 378 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d7ddbb290dfd3012f5aa02c56956acb097eef5b8f80243a3deaff12c12b286c4.jpg", + "image_caption": [ + "A cute bear" + ], + "image_footnote": [], + "bbox": [ + 86, + 387, + 200, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eb8f5046cf9749c732d4fd406a4560c5b3b793e5ae5ad161bdccdc6d44f34b1b.jpg", + "image_caption": [ + "A moose" + ], + "image_footnote": [], + "bbox": [ + 205, + 388, + 320, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/101e1170f7064b5edc9f04d96143713764a19191953362df139e3913fa60e039.jpg", + "image_caption": [ + "A baby pen" + ], + "image_footnote": [], + "bbox": [ + 320, + 388, + 372, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0f3d2fa0c62c9e0e5edf09ed204fd59daf986792ce49f36c091cb0568c688e05.jpg", + "image_caption": [ + "A moose", + "A baby penguin" + ], + "image_footnote": [], + "bbox": [ + 375, + 390, + 488, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1e88d7841435824dd9ec845f5d2d1d4131c53799ec14308bbce9203f29322916.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 390, + 542, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/066e32b3d3e53446c8e21b056da640992280644c433de7ec203d5b58ab4d1c94.jpg", + "image_caption": [ + "A moose", + "A baby penguin" + ], + "image_footnote": [], + "bbox": [ + 549, + 390, + 658, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d493c3f5b38c916496d273246a57c83b152ee12941106b89e07e94b18f98934f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 390, + 712, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/98bd1600692848c1d1e32c87ffbbaad7e8831f1325c1d591c3d4836fded90ab6.jpg", + "image_caption": [ + "A moose", + "A baby penguin" + ], + "image_footnote": [], + "bbox": [ + 715, + 390, + 828, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/07dab4a6b25a0ac834d7e5208ec8f5ee690714b8ca83182f7753e0c5f5c91b88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 390, + 883, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0e75922bafab63fd0dc43b2a236d62036d8df5d8480aa1a4c0b98594e1613c20.jpg", + "image_caption": [ + "A camera" + ], + "image_footnote": [], + "bbox": [ + 86, + 484, + 200, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fde80ce0ea8f158811b6028048d98b0700eaf3363d4247ad20510ff14bdce6ce.jpg", + "image_caption": [ + "A fireman" + ], + "image_footnote": [], + "bbox": [ + 205, + 484, + 320, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3613566046215aec7374cc7ae12d282ad8e812100a330aa8355d1ff5806af360.jpg", + "image_caption": [ + "A tall hill" + ], + "image_footnote": [], + "bbox": [ + 320, + 484, + 372, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/687587c656d7c57e108c490ea6e0f9a58ca98ec07e812b9a7d6dd6c239aad61e.jpg", + "image_caption": [ + "A fireman" + ], + "image_footnote": [], + "bbox": [ + 375, + 484, + 488, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5f858526fbb5f99a5735067d56df198d3cd99d2fc790db7840d84b2fe9da5d37.jpg", + "image_caption": [ + "A tall hill" + ], + "image_footnote": [], + "bbox": [ + 490, + 484, + 542, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a5ecc311b5dbeb6b2baedaef213a0d6b54334591474cb728be06ab329d4edeef.jpg", + "image_caption": [ + "A fireman" + ], + "image_footnote": [], + "bbox": [ + 549, + 484, + 658, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5f0710c2366220dd26304908de510520344952d40686136117fcf3bf7c907660.jpg", + "image_caption": [ + "A tall hill" + ], + "image_footnote": [], + "bbox": [ + 658, + 484, + 712, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/de633264efa8f257c089060d149edbe67c2102974cb50c9648219b00e7729e36.jpg", + "image_caption": [ + "A fireman" + ], + "image_footnote": [], + "bbox": [ + 715, + 484, + 828, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/848ac7555642c0cea82a2611ee7dab8c5f40775618c28cab35f1d537b3a2204a.jpg", + "image_caption": [ + "A tall hill" + ], + "image_footnote": [], + "bbox": [ + 828, + 484, + 883, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/218c906898b3a5ae3af8a0c59cbfe3f2e6a876f559a6ed9dbed5f74cff59aca5.jpg", + "image_caption": [ + "A saxophone" + ], + "image_footnote": [], + "bbox": [ + 86, + 578, + 200, + 665 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5a326baa1c71afd4fa92e7b945445d5df1a917fb68392dc96bde8df9060c0670.jpg", + "image_caption": [ + "A sunflower" + ], + "image_footnote": [], + "bbox": [ + 205, + 578, + 320, + 665 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2e24b65414432bef1e741090cb6d689be11819594a2cdbe6f42fe4398c0cb96f.jpg", + "image_caption": [ + "A compass" + ], + "image_footnote": [], + "bbox": [ + 320, + 578, + 372, + 665 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3bd7a2838367c664a9e337c39d2c1f17bc4b4d674a887d56572386b522ef23a5.jpg", + "image_caption": [ + "A sunflower" + ], + "image_footnote": [], + "bbox": [ + 375, + 578, + 488, + 665 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a909ea9a7f355a26af4d2503b0d77e6884a4592309406b7b8a2b5cbf9376bc47.jpg", + "image_caption": [ + "A compass" + ], + "image_footnote": [], + "bbox": [ + 490, + 578, + 542, + 665 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/245f0efac9754680077bce0f2084c5a09d0700a318b35ddbc7df9e82dd856f22.jpg", + "image_caption": [ + "A sunflower" + ], + "image_footnote": [], + "bbox": [ + 549, + 578, + 658, + 665 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1b207c7f42fabcce392a07367ce06138d3678cf13f55abd0b1abdda1ce539d27.jpg", + "image_caption": [ + "A compass" + ], + "image_footnote": [], + "bbox": [ + 658, + 578, + 712, + 665 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0e392ae6870353c15e4c33cb5e500f566f4a5367ee5622f56e67fdfb217bce2f.jpg", + "image_caption": [ + "A sunflower" + ], + "image_footnote": [], + "bbox": [ + 715, + 578, + 828, + 665 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2a20746fe4b8fc5b42a361f53ef12c4712aebf9df993f31740b2c8624cbe67e5.jpg", + "image_caption": [ + "A compass" + ], + "image_footnote": [], + "bbox": [ + 828, + 578, + 883, + 665 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b266b1b05456e22c05ece494dafbad1df540c9dd01234ce23f0846dcfdab7187.jpg", + "image_caption": [ + "A bottle of wine" + ], + "image_footnote": [], + "bbox": [ + 86, + 675, + 200, + 760 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0cc1655eb494e44e0a04cd133ad64293c8f42e2d1a50b974f0592bdb3ef10b6a.jpg", + "image_caption": [ + "Scones" + ], + "image_footnote": [], + "bbox": [ + 205, + 675, + 320, + 760 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c0febaa1d6f4af1489962cdf2c022e226c99f50d6390739b462f03a7e5894d83.jpg", + "image_caption": [ + "A muffin" + ], + "image_footnote": [], + "bbox": [ + 320, + 675, + 372, + 760 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a67f1819a5ac9832c842ffec3c5c2bea10ef041df432dc1197e1a531b7a6f866.jpg", + "image_caption": [ + "Scones" + ], + "image_footnote": [], + "bbox": [ + 375, + 675, + 488, + 760 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/273b7d99f4027b80fb94d2a712e1526dccae55b952daf95ebaf79781277e4638.jpg", + "image_caption": [ + "A muffin" + ], + "image_footnote": [], + "bbox": [ + 490, + 675, + 542, + 760 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9729ac558578c2cc718793e23350acb7c4e090b725969a2dc8c12431c4a194f0.jpg", + "image_caption": [ + "A fireman" + ], + "image_footnote": [], + "bbox": [ + 549, + 675, + 658, + 760 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8fd9cff8bb4ccfef3049b2a2e4eb07b940bcfdadfe156730a753ba68f0cf99eb.jpg", + "image_caption": [ + "A muffin" + ], + "image_footnote": [], + "bbox": [ + 658, + 675, + 712, + 760 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3bc108ed9e3f17a44c0e1c114055d5844ede0e28a61f0b1c4879b10da89fb618.jpg", + "image_caption": [ + "Scones" + ], + "image_footnote": [], + "bbox": [ + 715, + 675, + 828, + 760 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e20646a086cffac4ae116315a50305af6784ccc6c59bd2082f4bdeee1ee46165.jpg", + "image_caption": [ + "A muffin" + ], + "image_footnote": [], + "bbox": [ + 828, + 675, + 883, + 760 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2fa74eae159321dd188194042693392198fcb61858fa854a7b9382491534aef9.jpg", + "image_caption": [ + "A wise owl" + ], + "image_footnote": [], + "bbox": [ + 86, + 770, + 200, + 854 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4f563e863ce68571d9fa9ceae9bb07ae33f3fb60e146a0dab8488c410f327a6e.jpg", + "image_caption": [ + "Full moon", + "Figure 7. Qualitative comparison to personalization based methods." + ], + "image_footnote": [], + "bbox": [ + 205, + 770, + 320, + 854 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/39fa8d35e7e3ba2427e87d43ffa4924cc93aae4c1d0d493c156b7c5d1acb4cdd.jpg", + "image_caption": [ + "A book" + ], + "image_footnote": [], + "bbox": [ + 320, + 770, + 372, + 854 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2bb7b8d91f4af638659972a0ca014ec10d525a15fc9566d0869f376937cbc8a6.jpg", + "image_caption": [ + "Full moon" + ], + "image_footnote": [], + "bbox": [ + 375, + 770, + 488, + 854 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bc6a9fbaf50c348093b27ffe01d00455c1a6a322d45a787eed6f99fd26ddc061.jpg", + "image_caption": [ + "A book" + ], + "image_footnote": [], + "bbox": [ + 490, + 770, + 542, + 854 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d5733e81801d09ddb91dd13aa468c0647906a7565c02d87e74e2a016c17152db.jpg", + "image_caption": [ + "Full moon" + ], + "image_footnote": [], + "bbox": [ + 549, + 770, + 658, + 854 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3a71945b845f2970193a4628958143b266d4488bfa7f69cf7b4ced283b29b9c6.jpg", + "image_caption": [ + "A book" + ], + "image_footnote": [], + "bbox": [ + 658, + 770, + 712, + 854 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ec4bb150faf2d7acfdb5c2bcb832dccc3d4f65619f949822f0f87ff929759a87.jpg", + "image_caption": [ + "Full moon" + ], + "image_footnote": [], + "bbox": [ + 715, + 770, + 828, + 854 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/991f0ea54f6cbd853d334702a4074807d20a62f8c656ffa058f21edf2e85dc5d.jpg", + "image_caption": [ + "A book" + ], + "image_footnote": [], + "bbox": [ + 828, + 770, + 883, + 854 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4780", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ada735eb1902ca79d5ab11a26e4fe25d02bb9659ddec19de02a6b75738006b8c.jpg", + "image_caption": [ + "Figure 8. Varying level of attention sharing. By reducing the number of shared attention layers, i.e., allowing only self-attention in part of the layers, we can get more varied results (bottom rows) at the expense of style alignment (top row)." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 468, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparisons", + "text_level": 1, + "bbox": [ + 76, + 491, + 217, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For baselines, we compare our method to T2I personalization methods. We trained StyleDrop [49] and DreamBooth [41] over the first image in each set of our evaluation data, and use the trained personalized weights to generate the additional three images in each set. We use a public unofficial implementation of StyleDrop * (SDRP-unofficial) over non-regressive T2I model. Due to the large quality gap between the unofficial MUSE model † to the official one [7], we follow StyleDrop and implement an adapter model over SDXL (SDRP-SDXL) *, where we train a low rank linear layer after each Feed-Forward layer at the model's attention blocks. For training DreamBooth, we adapt the LoRA [21, 43] variant (DB-LoRA) over SDXL using the public huggingface-diffusers implementation ‡. We follow the hyperparameters tuning reported in [49] and train both SDRP-SDXL and DB-LoRA for 400 steps to prevent overfitting to the style training image.", + "bbox": [ + 75, + 513, + 468, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As can be seen in the qualitative comparison, Fig. 7, the image sets generated by our method, are more consistent across style attributes like color palette, drawing style, and composition. Moreover, the personalization-based methods", + "bbox": [ + 75, + 771, + 470, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "may leak the content of the training reference image (on the left) when generating the new images. For example, see the repeated woman and dogs in the results of DB-LoRA and SDRP-SDXL at the second row or the repeated owl at the bottom row. Similarly, because of the content leakage, these methods obtained lower text similarity scores and higher set consistency scores compared to our method.", + "bbox": [ + 500, + 90, + 890, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also apply two encoder-based personalization methods ELITE [57], IP-Adapter [59], and BLIP-Diffusion [27] over our evaluation set. These methods receive as input the first image in each set and use its embeddings to generate images with other content. Unlike the optimization-based techniques, these methods operate in a much faster feedforward diffusion loop, like our method. However, as can be seen in Fig. 6, their performance for style aligned image generation is poor compared to the other baselines. We argue that current encoder-based personalization techniques struggle to disentangle the content and the style of the input image. We supply qualitative results in the supplementary.", + "bbox": [ + 496, + 196, + 892, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User study. In addition to the automatic evaluation, we conducted a user study over the results of our method, StyleDrop (unofficial MUSE), StyleDrop (SDXL), and DreamBooth-LoRA (SDXL). In each question, we randomly sample one of the evaluation examples and show", + "bbox": [ + 498, + 378, + 890, + 453 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/68ce1e879762b8e47f9f2caa41262a4af139d672c99f321df933fc63e9d4f341.jpg", + "image_caption": [ + "Figure 9. Style aligned image generation to an input image. Given an input reference image (left column) and text description, we first apply DDIM inversion over the image to get the inverted diffusion trajectory $x_{T}$ , $x_{T-1} \\ldots x_{0}$ . Then, starting from $x_{T}$ and a new set of prompts, we apply our method to generate new content (right columns) with an aligned style to the input." + ], + "image_footnote": [], + "bbox": [ + 501, + 476, + 890, + 791 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "\\*github.com/aim-uofa/StyleDrop-PyTorch", + "bbox": [ + 99, + 840, + 401, + 853 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "github.com/baaivision/MUSE-Pytorch", + "bbox": [ + 102, + 854, + 377, + 864 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "\\*github.com/amirhertz/smlx_adapter", + "bbox": [ + 102, + 864, + 367, + 876 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ github.com/huggingface/diffusers", + "bbox": [ + 102, + 877, + 359, + 888 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4781", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the user the 4 image set that resulted from our and another method (in a random order). The user had to choose which set is better in terms of style consistency, and text alignment. A print screen of the user study format is provided in the supplementary. Overall, we collected 2400 answers from 100 users using the Amazon Mechanical Turk service. The results are summarized in Tab. 1 where for each method, we report the percentage of judgments in our favor. As can be seen, most participants favored our method by a large margin. More information about our user study can be found in the supplementary.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Additional Results", + "text_level": 1, + "bbox": [ + 76, + 266, + 256, + 279 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Style alignment control. We provide means of control over the style alignment to the reference image by applying the shared attention over only part of the self-attention layers. As can be seen in Fig. 8, reducing the number of shared attention layers results with a more diverse image set, which still shares common attributes with the reference image.", + "bbox": [ + 75, + 289, + 468, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "StyleAligned from an input image. To generate style-aligned images to an input image, we apply DDIM inversion [50] using a provided text caption. Then, we apply our method to generate new images in the style of the input using the inverted diffusion trajectory $x_{T}, x_{T-1}, \\ldots, x_{0}$ for the reference image. Examples are shown in Fig. 9, where we use BLIP captioning [28] to get a caption for each input image. For example, we used the prompt \"A render of a house with a yellow roof\" for the DDIM inversion of the top example and replaced the word house with other objects to generate the style-aligned images of a car, a cat, and a cactus. Notice that this method does not require any optimization. However, DDIM inversion may fail [30] or results with an erroneous trajectory [24]. More results and analysis, are provided in the supplementary.", + "bbox": [ + 75, + 378, + 468, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "StyleAligned with other methods. Since our method doesn't require training or optimization, it can be easily combined on top of other diffusion based methods to generate style-consistent image sets. Fig. 10 shows several such examples where we combine our method with ControlNet [60], DreamBooth [42] and MultiDiffusion [3]. More examples and details about the integration of StyleAligned with other methods can be found in supplementary.", + "bbox": [ + 75, + 606, + 468, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 739, + 204, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented StyleAligned, which addresses the challenge of achieving style-aligned image generation within the realm of large-scale Text-to-Image models. By introducing minimal attention sharing operations with AdaIN modulation during the diffusion process, our method successfully establishes style-consistency and visual coherence across generated images. The demonstrated efficacy of StyleAligned in producing high-quality, style-consistent images across diverse styles and textual prompts under", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48eab7e15d523b557b2e445a5c7f8acebe22b8794caccc6a4756084a368b7c75.jpg", + "image_caption": [ + "Figure 10. StyleAligned with other methods. On top, StyleAligned is combined with ControlNet to generate style-aligned images conditioned on depth maps. In the middle, our method combined with MultiDiffusion to generate panorama images that share multiple styles. On the bottom, style consistent and personalized content created by combining our method with pretrained personalized DreamBooth-LoRA models." + ], + "image_footnote": [], + "bbox": [ + 500, + 85, + 893, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "scores its potential in creative domains and practical applications. Our results affirm StyleAligned capability to faithfully adhere to provided descriptions and reference styles while maintaining impressive synthesis quality.", + "bbox": [ + 496, + 650, + 890, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the future we would like to explore the scalability and adaptability of StyleAligned to have more control over the shape and appearance similarity among the generated images. Additionally, due to the limitation of current diffusion inversion methods, a promising direction is to leverage StyleAligned to assemble a style-aligned dataset which then can be used to train style condition text-to-image models.", + "bbox": [ + 496, + 710, + 890, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 830, + 679, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We thank Or Patashnik, Matan Cohen, Yael Pritch, and Yael Vinker for their valuable inputs that helped improve this work.", + "bbox": [ + 496, + 853, + 890, + 895 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4782", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer, 2023. 2", + "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Trans. Graph., 42(4), jul 2023. 2", + "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 8", + "[4] James Betker, Gabriel Goh, Li Jing, TimBrooks, Jianfeng Wang, Linjie Li, LongOuyang, JuntangZhuang, JoyceLee, YufeiGuo, WesamManassra, PrafullaDhariwal, CaseyChu, YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 3", + "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. MasaCtrl: tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22560–22570, October 2023. 2", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5", + "[7] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, José Lezama, Lu Jiang, Ming Yang, Kevin P. Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. In International Conference on Machine Learning, 2023. 2, 7", + "[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42:1 - 10, 2023. 2", + "[9] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In The Eleventh International Conference on Learning Representations, 2022. 2", + "[10] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 571-576. 2023. 2", + "[11] Alexei A Efros and Thomas K Leung. Texture synthesis by non-parametric sampling. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 1033-1038. IEEE, 1999. 2", + "[12] Dave Epstein, Allan Jabri, Ben Poole, Alexei A Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 2", + "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, 2022. 1, 2", + "[14] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 2", + "[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2", + "[16] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris N. Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. ArXiv, abs/2303.11305, 2023. 2", + "[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2", + "[18] Aaron Hertzmann, Charles E Jacobs, Nuria Oliver, Brian Curless, and David H Salesin. Image analogies. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 557-570. 2023. 2", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proc. NeurIPS, 2020. 3", + "[20] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2", + "[21] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 7", + "[22] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 5", + "[23] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European conference on computer vision (ECCV), pages 172-189, 2018. 2", + "[24] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. arXiv preprint arXiv:2304.06140, 2023. 2, 8", + "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2", + "[26] Hochang Lee, Sanghyun Seo, Seungtaek Ryoo, and Kyunghyun Yoon. Directional texture transfer. In Proceedings of the 8th International Symposium on Non-Photorealistic Animation and Rendering, pages 43–48, 2010. 2", + "[27] Dongxu Li, Junnan Li, and Steven C. H. Hoi. BLIP-Diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. ArXiv, abs/2305.14720, 2023. 2, 7", + "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 8" + ], + "bbox": [ + 503, + 92, + 893, + 885 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4783", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsupervised image-to-image translation. In IEEE International Conference on Computer Vision (ICCV), 2019. 2", + "[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2, 8", + "[31] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, 2021. 2", + "[32] Dong Huk Park*, Grace Luo*, Clayton Toste, Samaneh Azadi, Xihui Liu, Maka Karalashvili, Anna Rohrbach, and Trevor Darrell. Shape-guided diffusion with inside-outside attention. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 2", + "[33] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2", + "[34] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. ArXiv, abs/2303.11306, 2023. 2", + "[35] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 2, 3, 5", + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 5", + "[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1", + "[38] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 2", + "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1", + "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference," + ], + "bbox": [ + 78, + 90, + 470, + 893 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3", + "[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 5, 7", + "[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 2, 8", + "[43] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora, 2022.7", + "[44] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. ACM SIGGRAPH 2022 Conference Proceedings, 2021. 2", + "[45] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1", + "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2, 3", + "[47] Jing Shi, Wei Xiong, Zhe L. Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. ArXiv, abs/2304.03411, 2023. 2", + "[48] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3", + "[49] Kihyuk Sohn, Nataniel Ruiz, Kimin Lee, Daniel Castro Chin, Irina Blok, Huiwen Chang, Jarred Barber, Lu Jiang, Glenn Entis, Yuanzhen Li, et al. Styledrop: Text-to-image generation in any style. arXiv preprint arXiv:2306.00983, 2023. 1, 2, 7", + "[50] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 8", + "[51] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. SIGGRAPH 2023 Conference Proceedings, 2023. 2", + "[52] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10748-10757, 2022. 2" + ], + "bbox": [ + 501, + 92, + 893, + 893 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4784", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2", + "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 3", + "[55] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. arXiv preprint arXiv:2211.13752, 2022. 2", + "[56] Andrey Voynov, Q. Chu, Daniel Cohen-Or, and Kfir Aberman. P+: Extended textual conditioning in text-to-image generation. ArXiv, abs/2303.09522, 2023. 2, 5", + "[57] Yuxiang Wei. Official implementation of ELITE. https://github.com/csyxwei/ELITE, 2023. 7", + "[58] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: Encoding visual concepts into textual embeddings for customized text-to-image generation. ArXiv, abs/2302.13848, 2023. 2", + "[59] Hu Ye, Jun Zhang, Siyi Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. ArXiv, abs/2308.06721, 2023. 2, 7", + "[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 8" + ], + "bbox": [ + 78, + 90, + 468, + 527 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4785", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Style Aligned Image Generation via Shared Attention/0cae3168-e63f-4f71-908a-5bdeae8d8200_model.json b/2024/Style Aligned Image Generation via Shared Attention/0cae3168-e63f-4f71-908a-5bdeae8d8200_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6ffc31cbfc7b2e3c3bab8d4519af066ec0ec231f --- /dev/null +++ b/2024/Style Aligned Image Generation via Shared Attention/0cae3168-e63f-4f71-908a-5bdeae8d8200_model.json @@ -0,0 +1,3731 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.131, + 0.757, + 0.154 + ], + "angle": 0, + "content": "Style Aligned Image Generation via Shared Attention" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.18, + 0.796, + 0.235 + ], + "angle": 0, + "content": "Amir Hertz\\*, Andrey Voynov\\*, Shlomi Fruchter\\*, and Daniel Cohen-Or\\*1,2 \n1 Google Research \n2 Tel Aviv University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.558 + ], + "angle": 0, + "content": "Large-scale Text-to-Image (T2I) models have rapidly gained prominence across creative fields, generating visually compelling outputs from textual prompts. However, controlling these models to ensure consistent style remains challenging, with existing methods necessitating fine-tuning and manual intervention to disentangle content and style. In this paper, we introduce StyleAligned, a novel technique designed to establish style alignment among a series of generated images. By employing minimal 'attention sharing' during the diffusion process, our method maintains style consistency across images within T2I models. This approach allows for the creation of style-consistent images using a reference style through a straightforward inversion operation. Our method's evaluation across diverse styles and text prompts demonstrates high-quality synthesis and fidelity, underscoring its efficacy in achieving consistent style across various inputs." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.585, + 0.21, + 0.601 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.47, + 0.701 + ], + "angle": 0, + "content": "Large-scale Text-to-Image (T2I) generative models [37, 39, 45] have emerged as an essential tool across creative disciplines, such as art, graphic design, animation, architecture, gaming and more. These models show tremendous capabilities of translating an input text into an appealing visual result that is aligned with the input description." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.47, + 0.821 + ], + "angle": 0, + "content": "An envisioned application of T2I models revolves around the rendition of various concepts in a way that shares a consistent style and character, as though all were created by the same artist and method (see Fig. 1). While proficient in aligning with the textual description of the style, state-of-the-art T2I models often create images that diverge significantly in their interpretations of the same stylistic descriptor, as depicted in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.822, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Recent methods mitigate this by fine-tuning the T2I model over a set of images that share the same style [13,49]. This optimization is computationally expensive and usually" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.267, + 0.895, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.629, + 0.895, + 0.672 + ], + "angle": 0, + "content": "Figure 1. Style aligned image set generation. By fusing the features of the toy train image (left) during the diffusion process, we can generate an image set of different content that shares the style." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.895, + 0.743 + ], + "angle": 0, + "content": "requires human input in order to find a plausible subset of images and texts that enables the disentanglement of content and style." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.743, + 0.895, + 0.895 + ], + "angle": 0, + "content": "We introduce StyleAligned, a method that enables consistent style interpretation across a set of generated images (Fig. 1). Our method requires no optimization and can be applied to any attention-based text-to-image diffusion model. We show that adding minimal attention sharing operations along the diffusion process, from each generated image to the first one in a batch, leads to a style-consistent set. Moreover, using diffusion inversion, our method can be applied to generate style-consistent images given a reference style image, with no optimization or fine-tuning." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.875, + 0.212, + 0.887 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.888, + 0.198, + 0.901 + ], + "angle": 0, + "content": "Equal Advising." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.875, + 0.212, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4775" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.47, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.396, + 0.471, + 0.466 + ], + "angle": 0, + "content": "Figure 2. Standard text-to-image vs. StyleAligned set generation. Given style description of \"minimal origami\", standard text-to-image generation (top) results with an unaligned image set while our method (bottom) can generate variety of style aligned content." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.645 + ], + "angle": 0, + "content": "We present our results over diverse styles and text prompts, demonstrating high-quality synthesis and fidelity to the prompts and reference style. We show diverse examples of generated images that share their style with a reference image that can possibly be a given input image. Importantly, our technique stands as a zero-shot solution, distinct from other personalization techniques, as it operates without any form of optimization or fine-tuning. For our code and more examples, please visit the project page style-aligned-gen.github.io" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.657, + 0.22, + 0.673 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.469, + 0.802 + ], + "angle": 0, + "content": "Text-to-image generation. Text conditioned image generative models [7, 31, 38] show unprecedented capabilities of generating high quality images from text descriptions. In particular, T2I diffusion models [35, 38, 46] are pushing the state of the art and they are quickly adopted for different generative visual tasks like inpainting [2, 44], imaged-to-image translation [55, 59], local image editing [9, 24], subject-driven image generation [42, 51] and more." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.803, + 0.469, + 0.893 + ], + "angle": 0, + "content": "Attention control in diffusion models. Hertz et al. [17] have shown how cross and self-attention maps within the diffusion process determine the layout and content of the generated images. Moreover, they showed how the attention maps can be used for controlled image generation. Other studies have leveraged modifications in attention lay" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.197 + ], + "angle": 0, + "content": "ers to enhance the fidelity or diversity of generated images [8, 34], or apply attention control for image editing [5, 12, 30, 32, 33, 53]. However, in contrast to prior approaches that primarily enable structure-preserving image editing, our method excels at generating images with diverse structures and content while maintaining a consistent style interpretation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.197, + 0.893, + 0.469 + ], + "angle": 0, + "content": "Style transfer. Transferring a style from a reference image to a target content image is well studied subject in computer graphics. Classic works [10, 11, 18, 26] rely on optimization of handcrafted features and texture resampling algorithms from an input texture image, combined with structure constrains of a content image. With the progress of deep learning research, another line of works utilizes deep neural priors for style transfer optimization using deep features of pre-trained networks [15, 52], or injecting attention features from a style image to a target one [1]. More related to our approach, Huang et al. [22] introduced a real time style transfer network based on Adaptive Instance Normalization layers (AdaIN) that are used to normalize deep features of a target image using deep features statistics of a reference style image. Follow-up works, employ the AdaIN layer for additional unsupervised learning tasks, like style-based image generation [25] and Image2Image translation [23, 29]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.893, + 0.695 + ], + "angle": 0, + "content": "T2I personalization To generalize T2I over new visual concepts, several works developed different optimization techniques over a small collection of input images that share the same concept [13, 16, 42, 56]. In instances where the collection shares a consistent style, the acquired concept becomes the style itself, affecting subsequent generations. Most close to our work is StyleDrop [49], a style personalization method that relies on fine-tuning of light weight adapter layers [20] at the end of each attention block in a non-autoregressive generative text-to-image transformer [7]. StyleDrop can generate a set of images in the same style of by training the adapter layers over a collection of images that share the same style. However, it struggles to generate a consistent image set of different content when training on a single image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.696, + 0.892, + 0.816 + ], + "angle": 0, + "content": "Our method can generate a consistent image set without optimization phase and without relying on several images for training. To skip the training phase, recent works developed dedicated personalization encoders [14, 27, 47, 58, 59] that can directly inject new priors from a single input image to the T2I model. However, these methods encounter challenges to disentangle style from content as they focus on generating the same subject as in the input image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.669, + 0.845 + ], + "angle": 0, + "content": "3. Method overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.855, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the following section we start with an overview of the T2I diffusion process, and in particular the self-attention mechanism Sec. 3.1. We continue by present" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4776" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.085, + 0.873, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.359, + 0.895, + 0.389 + ], + "angle": 0, + "content": "Figure 3. Style Aligned Diffusion. Generation of images with a style aligned to the reference image on the left. In each diffusion denoising step all the images, except the reference, perform a shared self-attention with the reference image." + }, + { + "type": "image", + "bbox": [ + 0.143, + 0.411, + 0.413, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.73, + 0.473, + 0.801 + ], + "angle": 0, + "content": "Figure 4. Shared attention layer. The target images attend to the reference image by applying AdaIN over their queries and keys using the reference queries and keys respectively. Then, we apply shared attention where the target features are updated by both the target values \\( V_{t} \\) and the reference values \\( V_{r} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.827, + 0.471, + 0.859 + ], + "angle": 0, + "content": "ing our attention-sharing operation within the self-attention layers that enable style aligned image set generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.413, + 0.642, + 0.427 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.436, + 0.893, + 0.512 + ], + "angle": 0, + "content": "Diffusion models [19, 48] are generative latent variable models that aim to model a distribution \\( p_{\\theta}(x_0) \\) that approximates the data distribution \\( q(x_0) \\) and are easy to sample from. Diffusion models are trained to reverse the diffusion \"forward process\":" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.522, + 0.833, + 0.54 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {\\alpha_ {t}} x _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\epsilon \\sim N (0, I),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.55, + 0.894, + 0.612 + ], + "angle": 0, + "content": "where \\(t \\in [0, \\infty)\\) and the values of \\(\\alpha_{t}\\) are determined by a scheduler such that \\(\\alpha_{0} = 1\\) and \\(\\lim_{t \\to \\infty} \\alpha_{t} = 0\\). During inference, we sample an image by gradually denoising an input noise image \\(x_{T} \\sim \\mathcal{N}(0, I)\\) via the reverse process:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.623, + 0.816, + 0.64 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = \\mu_ {t - 1} + \\sigma_ {t} z, z \\sim N (0, I),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.65, + 0.891, + 0.681 + ], + "angle": 0, + "content": "where the value of \\(\\sigma_{t}\\) is determined by the sampler and \\(\\mu_{t - 1}\\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.69, + 0.891, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mu_ {t - 1} = \\frac {\\sqrt {\\alpha_ {t - 1}} x _ {t}}{\\sqrt {\\alpha_ {t}}} + \\left(\\sqrt {1 - \\alpha_ {t - 1}} - \\frac {\\sqrt {1 - \\alpha_ {t}}}{\\sqrt {\\alpha_ {t}}}\\right) \\epsilon_ {\\theta} (x _ {t}, t),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.735, + 0.891, + 0.765 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}(x_t,t)\\) is the output of a diffusion model parameterized by \\(\\theta\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.766, + 0.892, + 0.841 + ], + "angle": 0, + "content": "Moreover, this process can be generalized for learning a marginal distribution using an additional input condition. That leads text-to-image diffusion models (T2I), where the output of the model \\(\\epsilon_{\\theta}(x_t,t,y)\\) is conditioned on a text prompt \\(y\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.841, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Self-Attention in T2I diffusion models. State-of-the-art T2I diffusion models [4,35,46] employ a U-Net architecture [40] that consists of convolution layers and transformer attention blocks [54]. In these attention mechanisms, deep" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4777" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.473, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.557, + 0.472, + 0.71 + ], + "angle": 0, + "content": "Figure 5. Ablation study - qualitative comparison. Each pair of rows shows two sets of images generated by the same set of prompts \"...in minimal flat design illustration\" using different configurations of our method, and each row in a pair uses a different seed. Sharing the self-attention between all images in the set (bottom) results with some diversity loss (style collapse across many seeds) and content leakage within each set (colors from one image leak to another). Disabling the queries keys AdALN operation results with less consistent image sets compared to our full method (top) which keeps on both diversity between different sets and consistency within each set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.736, + 0.47, + 0.782 + ], + "angle": 0, + "content": "image features \\(\\phi \\in \\mathbb{R}^{m\\times d_h}\\) attend to each other via selfattention layers and to contextual text embedding via crossattention layers." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.782, + 0.471, + 0.858 + ], + "angle": 0, + "content": "Our work operates at the self-attention layers where deep features are being updated by attending to each other. First, the features are projected into queries \\( Q \\in m \\times d_k \\), keys \\( K \\in m \\times d_k \\) and values \\( V \\in m \\times d_h \\) through learned linear layers. Then, the attention is computed by the scaled" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.648, + 0.106 + ], + "angle": 0, + "content": "dot-product attention:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.117, + 0.845, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}} V\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.161, + 0.894, + 0.268 + ], + "angle": 0, + "content": "where \\( d_{k} \\) is the dimension of \\( Q \\) and \\( K \\). Intuitively, each image feature is updated by a weighted sum of \\( V \\), where the weight depends on the correlation between the projected query \\( q \\) and the keys \\( K \\). In practice, each self-attention layer consists of several attention heads, and then the residual is computed by concatenating and projecting the attention heads output back to the image feature space \\( d_{h} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.278, + 0.815, + 0.296 + ], + "angle": 0, + "content": "\\[\n\\hat {\\phi} = \\phi + \\text {M u l t i - H e a d - A t t e n t i o n} (\\phi).\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.307, + 0.81, + 0.323 + ], + "angle": 0, + "content": "3.2. Style Aligned Image Set Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.33, + 0.892, + 0.526 + ], + "angle": 0, + "content": "The goal of our method is to generate a set of images \\(\\mathcal{I}_1\\ldots \\mathcal{I}_n\\) that are aligned with an input set of text prompts \\(y_{1}\\dots y_{n}\\) and share a consistent style interpretation with each other. For example, see the garnered image set of toy objects in Fig. 3 that are style-aligned with each other and to the input text on top. A naive way to generate a style aligned image set of different content is to use the same style description in the text prompts. As can be seen at the bottom of Fig. 2, generating different images using a shared style description of \"in minimal origami style\" results in an unaligned set, since each image is unaware of the exact appearance of other images in the set during the generation process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.892, + 0.587 + ], + "angle": 0, + "content": "The key insight underlying our approach is the utilization of the self-attention mechanism to allow communication among various generated images. This is achieved by sharing attention layers across the generated images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.892, + 0.633 + ], + "angle": 0, + "content": "Formally, let \\( Q_{i} \\), \\( K_{i} \\), and \\( V_{i} \\) be the queries, keys, and values, projected from deep features \\( \\phi_{i} \\) of \\( \\mathcal{I}_{i} \\) in the set, then, the attention update for \\( \\phi_{i} \\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.645, + 0.892, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} \\left(Q _ {i}, K _ {1 \\dots n}, V _ {1 \\dots n}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.671, + 0.892, + 0.744 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} K _ {1 \\dots n} = \\left[ \\begin{array}{c} K _ {1} \\\\ K _ {2} \\\\ \\vdots \\\\ K _ {n} \\end{array} \\right] \\text {a n d} V _ {1 \\dots n} = \\left[ \\begin{array}{c} V _ {1} \\\\ V _ {2} \\\\ \\vdots \\\\ V _ {n} \\end{array} \\right]. \\text {H o w e v e r},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.892, + 0.867 + ], + "angle": 0, + "content": "have noticed that by enabling full attention sharing, we may harm the quality of the generated set. As can be seen in Fig. 5 (bottom rows), full attention sharing results in content leakage among the images. For example, the unicorns got green paint from the garnered dino in the set. Moreover, full attention sharing results with less diverse sets of the same set of prompts, see the two sets in Fig. 5 in bottom rows compared to the sets above." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.868, + 0.892, + 0.898 + ], + "angle": 0, + "content": "To restrict the content leakage and allow diverse sets, we share the attention to only one image in the generated set" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4778" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.455, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.264, + 0.473, + 0.32 + ], + "angle": 0, + "content": "Figure 6. Quantitative Comparison. We compare the results of the different methods (blue marks) and our ablation experiments (orange marks) in terms of text alignment (CLIP score) and set consistency (DINO embedding similarity)." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.328, + 0.47, + 0.398 + ], + "angle": 0, + "content": "Table 1. User evaluation for style aligned image set generation. In each question, the user was asked to select between two image sets, Which is better in terms of style consistency and match to the text descriptions (see Sec. 4). We report the percentage of judgments in favor of StyleAligned over 800 answers (2400 in total)." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.409, + 0.446, + 0.456 + ], + "angle": 0, + "content": "
StyleDrop (unofficial MUSE)StyleDrop (SDXL)DreamBooth-LoRA (SDXL)
85.2 %67.1 %61.3%
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.47, + 0.59 + ], + "angle": 0, + "content": "(the first in the batch). That is, target image features \\(\\phi_t\\) are attending to themselves and to the features of only one reference image in the set using Eq. 1. As can be seen in Fig. 5 (middle), sharing the attention to only one image in the set results in diverse sets that share a similar style. However, in that case, we have noticed that the style of different images is not well aligned. We suspect that this is due to low attention flow from the reference to the target image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.47, + 0.667 + ], + "angle": 0, + "content": "As illustrated in Fig. 4, to enable balanced attention reference, we normalize the queries \\( Q_{t} \\) and keys \\( K_{t} \\) of the target image using the queries \\( Q_{r} \\) and keys \\( K_{r} \\) of the reference image using the adaptive normalization operation (AdaIN) [22]:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.677, + 0.433, + 0.696 + ], + "angle": 0, + "content": "\\[\n\\hat {Q} _ {t} = \\operatorname {A d a I N} \\left(Q _ {t}, Q _ {r}\\right) \\quad \\hat {K} _ {t} = \\operatorname {A d a I N} \\left(K _ {t}, K _ {r}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.707, + 0.338, + 0.722 + ], + "angle": 0, + "content": "where the AdaIn operation is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.733, + 0.414, + 0.766 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A d a I N} (x, y) = \\sigma (y) \\left(\\frac {x - \\mu (x)}{\\sigma (x)}\\right) + \\mu_ {y},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.777, + 0.47, + 0.824 + ], + "angle": 0, + "content": "and \\(\\mu (x),\\sigma (x)\\in \\mathbb{R}^{d_k}\\) are the mean and the standard deviation of queries and keys across different pixels. Finally, our shared attention is given by" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.833, + 0.359, + 0.852 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} \\left(\\hat {Q} _ {t}, K _ {r t} ^ {T}, V _ {r t}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.862, + 0.327, + 0.897 + ], + "angle": 0, + "content": "where \\(K_{rt} = \\left[ \\begin{array}{c}K_{r}\\\\ \\hat{K}_{t} \\end{array} \\right]\\) and \\(V_{rt} = \\left[ \\begin{array}{c}V_{r}\\\\ V_{t} \\end{array} \\right]\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.771, + 0.108 + ], + "angle": 0, + "content": "4. Evaluations and Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.235 + ], + "angle": 0, + "content": "We have implemented our method over Stable Diffusion XL (SDXL) [35] by applying our attention sharing overall 70 self-attention layers of the model. The generation of a four images set takes 29 seconds on a single A100 GPU. Notice that since the generation of the reference image is not influenced by other images in the batch, we can generate larger sets by fixing the prompt and seed of the reference image across the set generation." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.236, + 0.794, + 0.251 + ], + "angle": 0, + "content": "For example, see the sets in Fig. 2 and 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.252, + 0.892, + 0.356 + ], + "angle": 0, + "content": "Evaluation set. With the support of ChatGPT, we have generated 100 text prompts describing different image styles over four random objects. For example, “{A guitar, A hot air balloon, A sailboat, A mountain} in papercut art style.” For each style and set of objects, we use our method to generate a set of images. The full list of prompts is provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Metrics. To verify that each generated image contains its specified object, we measure the CLIP similarity [36] between the image and the text description of the object. In addition, we evaluate the style consistency of each generated set, by measuring the pairwise average cosine similarity between DINO VIT-B/8 [6] embeddings of the generated images in each set. Following [41, 56], we used DINO embeddings instead of CLIP image embeddings for measuring image similarity, since CLIP was trained with class labels and therefore it might give a high score for different images in the set that have similar content but with a different style. On the other hand, DINO better distinguishes between different styles due to its self-supervised training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.654, + 0.577 + ], + "angle": 0, + "content": "4.1. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.704 + ], + "angle": 0, + "content": "The quantitative results are summarized in Fig. 6, where the right-top place on the chart means better text similarity and style consistency, respectively. As a reference, we report the score obtained by generating the set of images using SDXL (T2I Reference) using the same seeds without any intervention. As can be seen, our method achieves a much higher style consistency score at the expense of text similarity. See qualitative comparison in Fig. 2" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In addition, we compared our method to additional two variants of the shared attention as described in Sec. 3.2. The first variant uses full attention sharing (Full Attn. Share) where the keys and values are shared between each pair of images in the set. In the second variant (W.A. AdaIN) we omit the AdaIN operation over queries and keys. As expected, this Full Attn. Share variant, results with higher style consistency and lower text alignment. As can be seen in Fig. 5, Full Attn. Share harms the overall quality of the image sets and diversity across sets. Moreover, our method without the use of AdaIN results in much lower style consistency. Qualitative results can be seen in Fig. 5 and in the supplementary.." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4779" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.091, + 0.192, + 0.102 + ], + "angle": 0, + "content": "Reference image" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.103, + 0.201, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.19, + 0.158, + 0.197 + ], + "angle": 0, + "content": "A guitar" + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.091, + 0.326, + 0.102 + ], + "angle": 0, + "content": "StyleAligned" + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.104, + 0.321, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.19, + 0.292, + 0.196 + ], + "angle": 0, + "content": "A hot air balloon" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.104, + 0.374, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.19, + 0.368, + 0.196 + ], + "angle": 0, + "content": "A mountain" + }, + { + "type": "image_caption", + "bbox": [ + 0.41, + 0.091, + 0.509, + 0.102 + ], + "angle": 0, + "content": "StyleDrop (SDXL)" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.103, + 0.49, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.19, + 0.462, + 0.196 + ], + "angle": 0, + "content": "A hot air balloon" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.103, + 0.544, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.19, + 0.537, + 0.196 + ], + "angle": 0, + "content": "A mountain" + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.091, + 0.706, + 0.102 + ], + "angle": 0, + "content": "StyleDrop (MUSE unofficial)" + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.104, + 0.659, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.19, + 0.632, + 0.196 + ], + "angle": 0, + "content": "A hot air balloon" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.104, + 0.714, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.19, + 0.708, + 0.196 + ], + "angle": 0, + "content": "A mountain" + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.091, + 0.879, + 0.102 + ], + "angle": 0, + "content": "Dreambooth-LoRA (SDXL)" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.104, + 0.83, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.19, + 0.8, + 0.196 + ], + "angle": 0, + "content": "A hot air balloon" + }, + { + "type": "image", + "bbox": [ + 0.832, + 0.104, + 0.884, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.836, + 0.19, + 0.878, + 0.196 + ], + "angle": 0, + "content": "A mountain" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.199, + 0.201, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.285, + 0.191, + 0.292 + ], + "angle": 0, + "content": "A woman walking her dogs" + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.199, + 0.321, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.285, + 0.291, + 0.292 + ], + "angle": 0, + "content": "A friendly robot" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.199, + 0.374, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.285, + 0.372, + 0.292 + ], + "angle": 0, + "content": "Cherryblossom" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.199, + 0.49, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.285, + 0.461, + 0.292 + ], + "angle": 0, + "content": "A friendly robot" + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.199, + 0.544, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.285, + 0.543, + 0.292 + ], + "angle": 0, + "content": "Cherryblossom" + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.199, + 0.659, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.285, + 0.617, + 0.292 + ], + "angle": 0, + "content": "A robot" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.199, + 0.714, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.285, + 0.707, + 0.292 + ], + "angle": 0, + "content": "Cherryblos" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.199, + 0.829, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.285, + 0.8, + 0.292 + ], + "angle": 0, + "content": "A friendly robot" + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.199, + 0.884, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.285, + 0.882, + 0.292 + ], + "angle": 0, + "content": "Cherryblossom" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.295, + 0.201, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.38, + 0.157, + 0.386 + ], + "angle": 0, + "content": "A duck" + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.295, + 0.321, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.38, + 0.276, + 0.387 + ], + "angle": 0, + "content": "A ship" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.295, + 0.374, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.38, + 0.362, + 0.387 + ], + "angle": 0, + "content": "A rocket" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.295, + 0.49, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.38, + 0.446, + 0.387 + ], + "angle": 0, + "content": "A ship" + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.295, + 0.544, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.502, + 0.38, + 0.533, + 0.387 + ], + "angle": 0, + "content": "A rocket" + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.295, + 0.659, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.38, + 0.615, + 0.387 + ], + "angle": 0, + "content": "A ship" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.295, + 0.714, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.38, + 0.699, + 0.387 + ], + "angle": 0, + "content": "A car" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.295, + 0.829, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.763, + 0.38, + 0.785, + 0.387 + ], + "angle": 0, + "content": "A ship" + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.295, + 0.884, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.841, + 0.38, + 0.872, + 0.387 + ], + "angle": 0, + "content": "A rocket" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.388, + 0.201, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.477, + 0.164, + 0.483 + ], + "angle": 0, + "content": "A cute bear" + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.389, + 0.321, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.477, + 0.28, + 0.483 + ], + "angle": 0, + "content": "A moose" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.389, + 0.374, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.477, + 0.369, + 0.483 + ], + "angle": 0, + "content": "A baby pen" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.391, + 0.49, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.477, + 0.45, + 0.483 + ], + "angle": 0, + "content": "A moose" + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.391, + 0.544, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.477, + 0.542, + 0.483 + ], + "angle": 0, + "content": "A baby penguin" + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.391, + 0.659, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.477, + 0.619, + 0.483 + ], + "angle": 0, + "content": "A moose" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.391, + 0.714, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.477, + 0.714, + 0.483 + ], + "angle": 0, + "content": "A baby penguin" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.391, + 0.829, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.477, + 0.789, + 0.483 + ], + "angle": 0, + "content": "A moose" + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.391, + 0.884, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.477, + 0.883, + 0.483 + ], + "angle": 0, + "content": "A baby penguin" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.485, + 0.201, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.572, + 0.161, + 0.578 + ], + "angle": 0, + "content": "A camera" + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.485, + 0.321, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.572, + 0.28, + 0.578 + ], + "angle": 0, + "content": "A fireman" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.485, + 0.374, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.572, + 0.362, + 0.578 + ], + "angle": 0, + "content": "A tall hill" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.485, + 0.49, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.572, + 0.451, + 0.578 + ], + "angle": 0, + "content": "A fireman" + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.485, + 0.544, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.572, + 0.532, + 0.578 + ], + "angle": 0, + "content": "A tall hill" + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.485, + 0.659, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.572, + 0.621, + 0.578 + ], + "angle": 0, + "content": "A fireman" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.485, + 0.714, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.672, + 0.572, + 0.701, + 0.578 + ], + "angle": 0, + "content": "A tall hill" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.485, + 0.829, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.572, + 0.791, + 0.578 + ], + "angle": 0, + "content": "A fireman" + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.485, + 0.884, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.841, + 0.572, + 0.872, + 0.578 + ], + "angle": 0, + "content": "A tall hill" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.579, + 0.201, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.667, + 0.167, + 0.673 + ], + "angle": 0, + "content": "A saxophone" + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.579, + 0.321, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.667, + 0.284, + 0.673 + ], + "angle": 0, + "content": "A sunflower" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.579, + 0.374, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.667, + 0.368, + 0.673 + ], + "angle": 0, + "content": "A compass" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.579, + 0.49, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.667, + 0.454, + 0.673 + ], + "angle": 0, + "content": "A sunflower" + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.579, + 0.544, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.667, + 0.536, + 0.673 + ], + "angle": 0, + "content": "A compass" + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.579, + 0.659, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.667, + 0.624, + 0.673 + ], + "angle": 0, + "content": "A sunflower" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.579, + 0.714, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.667, + 0.706, + 0.673 + ], + "angle": 0, + "content": "A compass" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.579, + 0.829, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.667, + 0.793, + 0.673 + ], + "angle": 0, + "content": "A sunflower" + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.579, + 0.884, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.667, + 0.875, + 0.673 + ], + "angle": 0, + "content": "A compass" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.676, + 0.201, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.762, + 0.171, + 0.768 + ], + "angle": 0, + "content": "A bottle of wine" + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.676, + 0.321, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.762, + 0.277, + 0.768 + ], + "angle": 0, + "content": "Scones" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.676, + 0.374, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.762, + 0.362, + 0.768 + ], + "angle": 0, + "content": "A muffin" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.676, + 0.49, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.762, + 0.447, + 0.768 + ], + "angle": 0, + "content": "Scones" + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.676, + 0.544, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.762, + 0.533, + 0.768 + ], + "angle": 0, + "content": "A muffin" + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.676, + 0.659, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.762, + 0.621, + 0.768 + ], + "angle": 0, + "content": "A fireman" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.676, + 0.714, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.762, + 0.701, + 0.768 + ], + "angle": 0, + "content": "A muffin" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.676, + 0.829, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.762, + 0.786, + 0.768 + ], + "angle": 0, + "content": "Scones" + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.676, + 0.884, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.841, + 0.762, + 0.872, + 0.768 + ], + "angle": 0, + "content": "A muffin" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.771, + 0.201, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.857, + 0.163, + 0.863 + ], + "angle": 0, + "content": "A wise owl" + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.771, + 0.321, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.857, + 0.28, + 0.863 + ], + "angle": 0, + "content": "Full moon" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.771, + 0.374, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.857, + 0.36, + 0.863 + ], + "angle": 0, + "content": "A book" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.771, + 0.49, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.857, + 0.45, + 0.863 + ], + "angle": 0, + "content": "Full moon" + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.771, + 0.544, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.857, + 0.531, + 0.863 + ], + "angle": 0, + "content": "A book" + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.771, + 0.659, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.857, + 0.621, + 0.863 + ], + "angle": 0, + "content": "Full moon" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.771, + 0.714, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.857, + 0.698, + 0.863 + ], + "angle": 0, + "content": "A book" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.771, + 0.829, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.857, + 0.791, + 0.863 + ], + "angle": 0, + "content": "Full moon" + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.771, + 0.884, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.843, + 0.857, + 0.869, + 0.863 + ], + "angle": 0, + "content": "A book" + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.877, + 0.697, + 0.891 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison to personalization based methods." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4780" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.47, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.424, + 0.471, + 0.48 + ], + "angle": 0, + "content": "Figure 8. Varying level of attention sharing. By reducing the number of shared attention layers, i.e., allowing only self-attention in part of the layers, we can get more varied results (bottom rows) at the expense of style alignment (top row)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.492, + 0.218, + 0.508 + ], + "angle": 0, + "content": "4.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.47, + 0.771 + ], + "angle": 0, + "content": "For baselines, we compare our method to T2I personalization methods. We trained StyleDrop [49] and DreamBooth [41] over the first image in each set of our evaluation data, and use the trained personalized weights to generate the additional three images in each set. We use a public unofficial implementation of StyleDrop * (SDRP-unofficial) over non-regressive T2I model. Due to the large quality gap between the unofficial MUSE model † to the official one [7], we follow StyleDrop and implement an adapter model over SDXL (SDRP-SDXL) *, where we train a low rank linear layer after each Feed-Forward layer at the model's attention blocks. For training DreamBooth, we adapt the LoRA [21, 43] variant (DB-LoRA) over SDXL using the public huggingface-diffusers implementation ‡. We follow the hyperparameters tuning reported in [49] and train both SDRP-SDXL and DB-LoRA for 400 steps to prevent overfitting to the style training image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.772, + 0.471, + 0.833 + ], + "angle": 0, + "content": "As can be seen in the qualitative comparison, Fig. 7, the image sets generated by our method, are more consistent across style attributes like color palette, drawing style, and composition. Moreover, the personalization-based methods" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "may leak the content of the training reference image (on the left) when generating the new images. For example, see the repeated woman and dogs in the results of DB-LoRA and SDRP-SDXL at the second row or the repeated owl at the bottom row. Similarly, because of the content leakage, these methods obtained lower text similarity scores and higher set consistency scores compared to our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.893, + 0.379 + ], + "angle": 0, + "content": "We also apply two encoder-based personalization methods ELITE [57], IP-Adapter [59], and BLIP-Diffusion [27] over our evaluation set. These methods receive as input the first image in each set and use its embeddings to generate images with other content. Unlike the optimization-based techniques, these methods operate in a much faster feedforward diffusion loop, like our method. However, as can be seen in Fig. 6, their performance for style aligned image generation is poor compared to the other baselines. We argue that current encoder-based personalization techniques struggle to disentangle the content and the style of the input image. We supply qualitative results in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.379, + 0.892, + 0.454 + ], + "angle": 0, + "content": "User study. In addition to the automatic evaluation, we conducted a user study over the results of our method, StyleDrop (unofficial MUSE), StyleDrop (SDXL), and DreamBooth-LoRA (SDXL). In each question, we randomly sample one of the evaluation examples and show" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.477, + 0.892, + 0.792 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.8, + 0.894, + 0.884 + ], + "angle": 0, + "content": "Figure 9. Style aligned image generation to an input image. Given an input reference image (left column) and text description, we first apply DDIM inversion over the image to get the inverted diffusion trajectory \\( x_{T} \\), \\( x_{T-1} \\ldots x_{0} \\). Then, starting from \\( x_{T} \\) and a new set of prompts, we apply our method to generate new content (right columns) with an aligned style to the input." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.841, + 0.402, + 0.854 + ], + "angle": 0, + "content": "\\*github.com/aim-uofa/StyleDrop-PyTorch" + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.855, + 0.378, + 0.865 + ], + "angle": 0, + "content": "github.com/baaivision/MUSE-Pytorch" + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.866, + 0.369, + 0.877 + ], + "angle": 0, + "content": "\\*github.com/amirhertz/smlx_adapter" + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.878, + 0.361, + 0.89 + ], + "angle": 0, + "content": "\\(\\ddagger\\)github.com/huggingface/diffusers" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.841, + 0.402, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4781" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "the user the 4 image set that resulted from our and another method (in a random order). The user had to choose which set is better in terms of style consistency, and text alignment. A print screen of the user study format is provided in the supplementary. Overall, we collected 2400 answers from 100 users using the Amazon Mechanical Turk service. The results are summarized in Tab. 1 where for each method, we report the percentage of judgments in our favor. As can be seen, most participants favored our method by a large margin. More information about our user study can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.267, + 0.258, + 0.28 + ], + "angle": 0, + "content": "4.3. Additional Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.47, + 0.38 + ], + "angle": 0, + "content": "Style alignment control. We provide means of control over the style alignment to the reference image by applying the shared attention over only part of the self-attention layers. As can be seen in Fig. 8, reducing the number of shared attention layers results with a more diverse image set, which still shares common attributes with the reference image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.47, + 0.606 + ], + "angle": 0, + "content": "StyleAligned from an input image. To generate style-aligned images to an input image, we apply DDIM inversion [50] using a provided text caption. Then, we apply our method to generate new images in the style of the input using the inverted diffusion trajectory \\( x_{T}, x_{T-1}, \\ldots, x_{0} \\) for the reference image. Examples are shown in Fig. 9, where we use BLIP captioning [28] to get a caption for each input image. For example, we used the prompt \"A render of a house with a yellow roof\" for the DDIM inversion of the top example and replaced the word house with other objects to generate the style-aligned images of a car, a cat, and a cactus. Notice that this method does not require any optimization. However, DDIM inversion may fail [30] or results with an erroneous trajectory [24]. More results and analysis, are provided in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.47, + 0.729 + ], + "angle": 0, + "content": "StyleAligned with other methods. Since our method doesn't require training or optimization, it can be easily combined on top of other diffusion based methods to generate style-consistent image sets. Fig. 10 shows several such examples where we combine our method with ControlNet [60], DreamBooth [42] and MultiDiffusion [3]. More examples and details about the integration of StyleAligned with other methods can be found in supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.205, + 0.755 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We have presented StyleAligned, which addresses the challenge of achieving style-aligned image generation within the realm of large-scale Text-to-Image models. By introducing minimal attention sharing operations with AdaIN modulation during the diffusion process, our method successfully establishes style-consistency and visual coherence across generated images. The demonstrated efficacy of StyleAligned in producing high-quality, style-consistent images across diverse styles and textual prompts under" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.087, + 0.895, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.541, + 0.895, + 0.639 + ], + "angle": 0, + "content": "Figure 10. StyleAligned with other methods. On top, StyleAligned is combined with ControlNet to generate style-aligned images conditioned on depth maps. In the middle, our method combined with MultiDiffusion to generate panorama images that share multiple styles. On the bottom, style consistent and personalized content created by combining our method with pretrained personalized DreamBooth-LoRA models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.892, + 0.711 + ], + "angle": 0, + "content": "scores its potential in creative domains and practical applications. Our results affirm StyleAligned capability to faithfully adhere to provided descriptions and reference styles while maintaining impressive synthesis quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.711, + 0.892, + 0.817 + ], + "angle": 0, + "content": "In the future we would like to explore the scalability and adaptability of StyleAligned to have more control over the shape and appearance similarity among the generated images. Additionally, due to the limitation of current diffusion inversion methods, a promising direction is to leverage StyleAligned to assemble a style-aligned dataset which then can be used to train style condition text-to-image models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.681, + 0.848 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.854, + 0.892, + 0.896 + ], + "angle": 0, + "content": "We thank Or Patashnik, Matan Cohen, Yael Pritch, and Yael Vinker for their valuable inputs that helped improve this work." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4782" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.471, + 0.186 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Trans. Graph., 42(4), jul 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.471, + 0.227 + ], + "angle": 0, + "content": "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.298 + ], + "angle": 0, + "content": "[4] James Betker, Gabriel Goh, Li Jing, TimBrooks, Jianfeng Wang, Linjie Li, LongOuyang, JuntangZhuang, JoyceLee, YufeiGuo, WesamManassra, PrafullaDhariwal, CaseyChu, YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.381 + ], + "angle": 0, + "content": "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. MasaCtrl: tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22560–22570, October 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.452 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.471, + 0.536 + ], + "angle": 0, + "content": "[7] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, José Lezama, Lu Jiang, Ming Yang, Kevin P. Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. In International Conference on Machine Learning, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42:1 - 10, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.592, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[9] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In The Eleventh International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.47, + 0.69 + ], + "angle": 0, + "content": "[10] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 571-576. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[11] Alexei A Efros and Thomas K Leung. Texture synthesis by non-parametric sampling. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 1033-1038. IEEE, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[12] Dave Epstein, Allan Jabri, Ben Poole, Alexei A Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[14] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[16] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris N. Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. ArXiv, abs/2303.11305, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[18] Aaron Hertzmann, Charles E Jacobs, Nuria Oliver, Brian Curless, and David H Salesin. Image analogies. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 557-570. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.348, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proc. NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[20] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[21] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[22] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.56, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[23] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European conference on computer vision (ECCV), pages 172-189, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[24] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. arXiv preprint arXiv:2304.06140, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.716, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[26] Hochang Lee, Sanghyun Seo, Seungtaek Ryoo, and Kyunghyun Yoon. Directional texture transfer. In Proceedings of the 8th International Symposium on Non-Photorealistic Animation and Rendering, pages 43–48, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.786, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[27] Dongxu Li, Junnan Li, and Steven C. H. Hoi. BLIP-Diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. ArXiv, abs/2305.14720, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.843, + 0.892, + 0.886 + ], + "angle": 0, + "content": "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.886 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4783" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsupervised image-to-image translation. In IEEE International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.471, + 0.289 + ], + "angle": 0, + "content": "[31] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.29, + 0.471, + 0.359 + ], + "angle": 0, + "content": "[32] Dong Huk Park*, Grace Luo*, Clayton Toste, Samaneh Azadi, Xihui Liu, Maka Karalashvili, Anna Rohrbach, and Trevor Darrell. Shape-guided diffusion with inside-outside attention. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.36, + 0.47, + 0.416 + ], + "angle": 0, + "content": "[33] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[34] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. ArXiv, abs/2303.11306, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.473, + 0.47, + 0.543 + ], + "angle": 0, + "content": "[35] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.545, + 0.47, + 0.627 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.47, + 0.696 + ], + "angle": 0, + "content": "[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.699, + 0.47, + 0.767 + ], + "angle": 0, + "content": "[38] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.77, + 0.471, + 0.838 + ], + "angle": 0, + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.84, + 0.47, + 0.895 + ], + "angle": 0, + "content": "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[43] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora, 2022.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.319, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[44] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. ACM SIGGRAPH 2022 Conference Proceedings, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.39, + 0.892, + 0.471 + ], + "angle": 0, + "content": "[45] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.474, + 0.892, + 0.557 + ], + "angle": 0, + "content": "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.558, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[47] Jing Shi, Wei Xiong, Zhe L. Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. ArXiv, abs/2304.03411, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.601, + 0.892, + 0.668 + ], + "angle": 0, + "content": "[48] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.671, + 0.892, + 0.739 + ], + "angle": 0, + "content": "[49] Kihyuk Sohn, Nataniel Ruiz, Kimin Lee, Daniel Castro Chin, Irina Blok, Huiwen Chang, Jarred Barber, Lu Jiang, Glenn Entis, Yuanzhen Li, et al. Styledrop: Text-to-image generation in any style. arXiv preprint arXiv:2306.00983, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.742, + 0.892, + 0.783 + ], + "angle": 0, + "content": "[50] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.785, + 0.892, + 0.826 + ], + "angle": 0, + "content": "[51] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. SIGGRAPH 2023 Conference Proceedings, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.828, + 0.892, + 0.895 + ], + "angle": 0, + "content": "[52] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10748-10757, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4784" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[53] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.259 + ], + "angle": 0, + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.262, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[55] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. arXiv preprint arXiv:2211.13752, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.47, + 0.346 + ], + "angle": 0, + "content": "[56] Andrey Voynov, Q. Chu, Daniel Cohen-Or, and Kfir Aberman. P+: Extended textual conditioning in text-to-image generation. ArXiv, abs/2303.09522, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.468, + 0.374 + ], + "angle": 0, + "content": "[57] Yuxiang Wei. Official implementation of ELITE. https://github.com/csyxwei/ELITE, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.47, + 0.431 + ], + "angle": 0, + "content": "[58] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: Encoding visual concepts into textual embeddings for customized text-to-image generation. ArXiv, abs/2302.13848, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.47, + 0.486 + ], + "angle": 0, + "content": "[59] Hu Ye, Jun Zhang, Siyi Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. ArXiv, abs/2308.06721, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.47, + 0.529 + ], + "angle": 0, + "content": "[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "4785" + } + ] +] \ No newline at end of file diff --git a/2024/Style Aligned Image Generation via Shared Attention/0cae3168-e63f-4f71-908a-5bdeae8d8200_origin.pdf b/2024/Style Aligned Image Generation via Shared Attention/0cae3168-e63f-4f71-908a-5bdeae8d8200_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7789c510222db3a31948ae77a30e80d3748f2686 --- /dev/null +++ b/2024/Style Aligned Image Generation via Shared Attention/0cae3168-e63f-4f71-908a-5bdeae8d8200_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fac94e9345423fc93f0ae9927fb2ed3c56b1c790779364354863699ed0e86eed +size 9245834 diff --git a/2024/Style Aligned Image Generation via Shared Attention/full.md b/2024/Style Aligned Image Generation via Shared Attention/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a5ec4dc119711a506f35d601f5eef45ccc41ecaa --- /dev/null +++ b/2024/Style Aligned Image Generation via Shared Attention/full.md @@ -0,0 +1,499 @@ +# Style Aligned Image Generation via Shared Attention + +Amir Hertz\*, Andrey Voynov\*, Shlomi Fruchter\*, and Daniel Cohen-Or\*1,2 +1 Google Research +2 Tel Aviv University + +# Abstract + +Large-scale Text-to-Image (T2I) models have rapidly gained prominence across creative fields, generating visually compelling outputs from textual prompts. However, controlling these models to ensure consistent style remains challenging, with existing methods necessitating fine-tuning and manual intervention to disentangle content and style. In this paper, we introduce StyleAligned, a novel technique designed to establish style alignment among a series of generated images. By employing minimal 'attention sharing' during the diffusion process, our method maintains style consistency across images within T2I models. This approach allows for the creation of style-consistent images using a reference style through a straightforward inversion operation. Our method's evaluation across diverse styles and text prompts demonstrates high-quality synthesis and fidelity, underscoring its efficacy in achieving consistent style across various inputs. + +# 1. Introduction + +Large-scale Text-to-Image (T2I) generative models [37, 39, 45] have emerged as an essential tool across creative disciplines, such as art, graphic design, animation, architecture, gaming and more. These models show tremendous capabilities of translating an input text into an appealing visual result that is aligned with the input description. + +An envisioned application of T2I models revolves around the rendition of various concepts in a way that shares a consistent style and character, as though all were created by the same artist and method (see Fig. 1). While proficient in aligning with the textual description of the style, state-of-the-art T2I models often create images that diverge significantly in their interpretations of the same stylistic descriptor, as depicted in Fig. 2. + +Recent methods mitigate this by fine-tuning the T2I model over a set of images that share the same style [13,49]. This optimization is computationally expensive and usually + +![](images/befe72dd08ea13e65a43000c1dc25189dcdb8c0c0bf181c36f354ee9aa00c776.jpg) +Figure 1. Style aligned image set generation. By fusing the features of the toy train image (left) during the diffusion process, we can generate an image set of different content that shares the style. + +requires human input in order to find a plausible subset of images and texts that enables the disentanglement of content and style. + +We introduce StyleAligned, a method that enables consistent style interpretation across a set of generated images (Fig. 1). Our method requires no optimization and can be applied to any attention-based text-to-image diffusion model. We show that adding minimal attention sharing operations along the diffusion process, from each generated image to the first one in a batch, leads to a style-consistent set. Moreover, using diffusion inversion, our method can be applied to generate style-consistent images given a reference style image, with no optimization or fine-tuning. + +![](images/cca67eff1325ad439ed8bffe4784e7bbf26f6fce3d87078922c54c3404fd3b16.jpg) +Figure 2. Standard text-to-image vs. StyleAligned set generation. Given style description of "minimal origami", standard text-to-image generation (top) results with an unaligned image set while our method (bottom) can generate variety of style aligned content. + +We present our results over diverse styles and text prompts, demonstrating high-quality synthesis and fidelity to the prompts and reference style. We show diverse examples of generated images that share their style with a reference image that can possibly be a given input image. Importantly, our technique stands as a zero-shot solution, distinct from other personalization techniques, as it operates without any form of optimization or fine-tuning. For our code and more examples, please visit the project page style-aligned-gen.github.io + +# 2. Related Work + +Text-to-image generation. Text conditioned image generative models [7, 31, 38] show unprecedented capabilities of generating high quality images from text descriptions. In particular, T2I diffusion models [35, 38, 46] are pushing the state of the art and they are quickly adopted for different generative visual tasks like inpainting [2, 44], imaged-to-image translation [55, 59], local image editing [9, 24], subject-driven image generation [42, 51] and more. + +Attention control in diffusion models. Hertz et al. [17] have shown how cross and self-attention maps within the diffusion process determine the layout and content of the generated images. Moreover, they showed how the attention maps can be used for controlled image generation. Other studies have leveraged modifications in attention lay + +ers to enhance the fidelity or diversity of generated images [8, 34], or apply attention control for image editing [5, 12, 30, 32, 33, 53]. However, in contrast to prior approaches that primarily enable structure-preserving image editing, our method excels at generating images with diverse structures and content while maintaining a consistent style interpretation. + +Style transfer. Transferring a style from a reference image to a target content image is well studied subject in computer graphics. Classic works [10, 11, 18, 26] rely on optimization of handcrafted features and texture resampling algorithms from an input texture image, combined with structure constrains of a content image. With the progress of deep learning research, another line of works utilizes deep neural priors for style transfer optimization using deep features of pre-trained networks [15, 52], or injecting attention features from a style image to a target one [1]. More related to our approach, Huang et al. [22] introduced a real time style transfer network based on Adaptive Instance Normalization layers (AdaIN) that are used to normalize deep features of a target image using deep features statistics of a reference style image. Follow-up works, employ the AdaIN layer for additional unsupervised learning tasks, like style-based image generation [25] and Image2Image translation [23, 29]. + +T2I personalization To generalize T2I over new visual concepts, several works developed different optimization techniques over a small collection of input images that share the same concept [13, 16, 42, 56]. In instances where the collection shares a consistent style, the acquired concept becomes the style itself, affecting subsequent generations. Most close to our work is StyleDrop [49], a style personalization method that relies on fine-tuning of light weight adapter layers [20] at the end of each attention block in a non-autoregressive generative text-to-image transformer [7]. StyleDrop can generate a set of images in the same style of by training the adapter layers over a collection of images that share the same style. However, it struggles to generate a consistent image set of different content when training on a single image. + +Our method can generate a consistent image set without optimization phase and without relying on several images for training. To skip the training phase, recent works developed dedicated personalization encoders [14, 27, 47, 58, 59] that can directly inject new priors from a single input image to the T2I model. However, these methods encounter challenges to disentangle style from content as they focus on generating the same subject as in the input image. + +# 3. Method overview + +In the following section we start with an overview of the T2I diffusion process, and in particular the self-attention mechanism Sec. 3.1. We continue by present + +![](images/6bfec6c2924b31a525aed02432c3bbe38680437dfe55778583f7be18ac076611.jpg) +Figure 3. Style Aligned Diffusion. Generation of images with a style aligned to the reference image on the left. In each diffusion denoising step all the images, except the reference, perform a shared self-attention with the reference image. + +![](images/40d62b020e5800a0fb2d06cae575f19a45f0d7e49c04264b0542055e71d15eec.jpg) +Figure 4. Shared attention layer. The target images attend to the reference image by applying AdaIN over their queries and keys using the reference queries and keys respectively. Then, we apply shared attention where the target features are updated by both the target values $V_{t}$ and the reference values $V_{r}$ . + +ing our attention-sharing operation within the self-attention layers that enable style aligned image set generation. + +# 3.1. Preliminaries + +Diffusion models [19, 48] are generative latent variable models that aim to model a distribution $p_{\theta}(x_0)$ that approximates the data distribution $q(x_0)$ and are easy to sample from. Diffusion models are trained to reverse the diffusion "forward process": + +$$ +x _ {t} = \sqrt {\alpha_ {t}} x _ {0} + \sqrt {1 - \alpha_ {t}} \epsilon , \epsilon \sim N (0, I), +$$ + +where $t \in [0, \infty)$ and the values of $\alpha_{t}$ are determined by a scheduler such that $\alpha_{0} = 1$ and $\lim_{t \to \infty} \alpha_{t} = 0$ . During inference, we sample an image by gradually denoising an input noise image $x_{T} \sim \mathcal{N}(0, I)$ via the reverse process: + +$$ +x _ {t - 1} = \mu_ {t - 1} + \sigma_ {t} z, z \sim N (0, I), +$$ + +where the value of $\sigma_{t}$ is determined by the sampler and $\mu_{t - 1}$ is given by + +$$ +\mu_ {t - 1} = \frac {\sqrt {\alpha_ {t - 1}} x _ {t}}{\sqrt {\alpha_ {t}}} + \left(\sqrt {1 - \alpha_ {t - 1}} - \frac {\sqrt {1 - \alpha_ {t}}}{\sqrt {\alpha_ {t}}}\right) \epsilon_ {\theta} (x _ {t}, t), +$$ + +where $\epsilon_{\theta}(x_t,t)$ is the output of a diffusion model parameterized by $\theta$ + +Moreover, this process can be generalized for learning a marginal distribution using an additional input condition. That leads text-to-image diffusion models (T2I), where the output of the model $\epsilon_{\theta}(x_t,t,y)$ is conditioned on a text prompt $y$ . + +Self-Attention in T2I diffusion models. State-of-the-art T2I diffusion models [4,35,46] employ a U-Net architecture [40] that consists of convolution layers and transformer attention blocks [54]. In these attention mechanisms, deep + +![](images/ba3e52212d6c2d820ad98b6bb7c6961b73723f21fa0dc4ed42a668f250f4f670.jpg) +Figure 5. Ablation study - qualitative comparison. Each pair of rows shows two sets of images generated by the same set of prompts "...in minimal flat design illustration" using different configurations of our method, and each row in a pair uses a different seed. Sharing the self-attention between all images in the set (bottom) results with some diversity loss (style collapse across many seeds) and content leakage within each set (colors from one image leak to another). Disabling the queries keys AdALN operation results with less consistent image sets compared to our full method (top) which keeps on both diversity between different sets and consistency within each set. + +image features $\phi \in \mathbb{R}^{m\times d_h}$ attend to each other via selfattention layers and to contextual text embedding via crossattention layers. + +Our work operates at the self-attention layers where deep features are being updated by attending to each other. First, the features are projected into queries $Q \in m \times d_k$ , keys $K \in m \times d_k$ and values $V \in m \times d_h$ through learned linear layers. Then, the attention is computed by the scaled + +dot-product attention: + +$$ +\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}} V\right), +$$ + +where $d_{k}$ is the dimension of $Q$ and $K$ . Intuitively, each image feature is updated by a weighted sum of $V$ , where the weight depends on the correlation between the projected query $q$ and the keys $K$ . In practice, each self-attention layer consists of several attention heads, and then the residual is computed by concatenating and projecting the attention heads output back to the image feature space $d_{h}$ : + +$$ +\hat {\phi} = \phi + \text {M u l t i - H e a d - A t t e n t i o n} (\phi). +$$ + +# 3.2. Style Aligned Image Set Generation + +The goal of our method is to generate a set of images $\mathcal{I}_1\ldots \mathcal{I}_n$ that are aligned with an input set of text prompts $y_{1}\dots y_{n}$ and share a consistent style interpretation with each other. For example, see the garnered image set of toy objects in Fig. 3 that are style-aligned with each other and to the input text on top. A naive way to generate a style aligned image set of different content is to use the same style description in the text prompts. As can be seen at the bottom of Fig. 2, generating different images using a shared style description of "in minimal origami style" results in an unaligned set, since each image is unaware of the exact appearance of other images in the set during the generation process. + +The key insight underlying our approach is the utilization of the self-attention mechanism to allow communication among various generated images. This is achieved by sharing attention layers across the generated images. + +Formally, let $Q_{i}$ , $K_{i}$ , and $V_{i}$ be the queries, keys, and values, projected from deep features $\phi_{i}$ of $\mathcal{I}_{i}$ in the set, then, the attention update for $\phi_{i}$ is given by: + +$$ +\operatorname {A t t e n t i o n} \left(Q _ {i}, K _ {1 \dots n}, V _ {1 \dots n}\right), \tag {1} +$$ + +$$ +\text {w h e r e} K _ {1 \dots n} = \left[ \begin{array}{c} K _ {1} \\ K _ {2} \\ \vdots \\ K _ {n} \end{array} \right] \text {a n d} V _ {1 \dots n} = \left[ \begin{array}{c} V _ {1} \\ V _ {2} \\ \vdots \\ V _ {n} \end{array} \right]. \text {H o w e v e r}, +$$ + +have noticed that by enabling full attention sharing, we may harm the quality of the generated set. As can be seen in Fig. 5 (bottom rows), full attention sharing results in content leakage among the images. For example, the unicorns got green paint from the garnered dino in the set. Moreover, full attention sharing results with less diverse sets of the same set of prompts, see the two sets in Fig. 5 in bottom rows compared to the sets above. + +To restrict the content leakage and allow diverse sets, we share the attention to only one image in the generated set + +![](images/925b1dcfbe4f36bc77e342a38b71d524bffa46834c75e5708976954f11d2d13e.jpg) +Figure 6. Quantitative Comparison. We compare the results of the different methods (blue marks) and our ablation experiments (orange marks) in terms of text alignment (CLIP score) and set consistency (DINO embedding similarity). + +Table 1. User evaluation for style aligned image set generation. In each question, the user was asked to select between two image sets, Which is better in terms of style consistency and match to the text descriptions (see Sec. 4). We report the percentage of judgments in favor of StyleAligned over 800 answers (2400 in total). + +
StyleDrop (unofficial MUSE)StyleDrop (SDXL)DreamBooth-LoRA (SDXL)
85.2 %67.1 %61.3%
+ +(the first in the batch). That is, target image features $\phi_t$ are attending to themselves and to the features of only one reference image in the set using Eq. 1. As can be seen in Fig. 5 (middle), sharing the attention to only one image in the set results in diverse sets that share a similar style. However, in that case, we have noticed that the style of different images is not well aligned. We suspect that this is due to low attention flow from the reference to the target image. + +As illustrated in Fig. 4, to enable balanced attention reference, we normalize the queries $Q_{t}$ and keys $K_{t}$ of the target image using the queries $Q_{r}$ and keys $K_{r}$ of the reference image using the adaptive normalization operation (AdaIN) [22]: + +$$ +\hat {Q} _ {t} = \operatorname {A d a I N} \left(Q _ {t}, Q _ {r}\right) \quad \hat {K} _ {t} = \operatorname {A d a I N} \left(K _ {t}, K _ {r}\right), +$$ + +where the AdaIn operation is given by: + +$$ +\operatorname {A d a I N} (x, y) = \sigma (y) \left(\frac {x - \mu (x)}{\sigma (x)}\right) + \mu_ {y}, +$$ + +and $\mu (x),\sigma (x)\in \mathbb{R}^{d_k}$ are the mean and the standard deviation of queries and keys across different pixels. Finally, our shared attention is given by + +$$ +\operatorname {A t t e n t i o n} \left(\hat {Q} _ {t}, K _ {r t} ^ {T}, V _ {r t}\right), +$$ + +where $K_{rt} = \left[ \begin{array}{c}K_{r}\\ \hat{K}_{t} \end{array} \right]$ and $V_{rt} = \left[ \begin{array}{c}V_{r}\\ V_{t} \end{array} \right]$ + +# 4. Evaluations and Experiments + +We have implemented our method over Stable Diffusion XL (SDXL) [35] by applying our attention sharing overall 70 self-attention layers of the model. The generation of a four images set takes 29 seconds on a single A100 GPU. Notice that since the generation of the reference image is not influenced by other images in the batch, we can generate larger sets by fixing the prompt and seed of the reference image across the set generation. + +For example, see the sets in Fig. 2 and 3. + +Evaluation set. With the support of ChatGPT, we have generated 100 text prompts describing different image styles over four random objects. For example, “{A guitar, A hot air balloon, A sailboat, A mountain} in papercut art style.” For each style and set of objects, we use our method to generate a set of images. The full list of prompts is provided in the appendix. + +Metrics. To verify that each generated image contains its specified object, we measure the CLIP similarity [36] between the image and the text description of the object. In addition, we evaluate the style consistency of each generated set, by measuring the pairwise average cosine similarity between DINO VIT-B/8 [6] embeddings of the generated images in each set. Following [41, 56], we used DINO embeddings instead of CLIP image embeddings for measuring image similarity, since CLIP was trained with class labels and therefore it might give a high score for different images in the set that have similar content but with a different style. On the other hand, DINO better distinguishes between different styles due to its self-supervised training. + +# 4.1. Ablation Study + +The quantitative results are summarized in Fig. 6, where the right-top place on the chart means better text similarity and style consistency, respectively. As a reference, we report the score obtained by generating the set of images using SDXL (T2I Reference) using the same seeds without any intervention. As can be seen, our method achieves a much higher style consistency score at the expense of text similarity. See qualitative comparison in Fig. 2 + +In addition, we compared our method to additional two variants of the shared attention as described in Sec. 3.2. The first variant uses full attention sharing (Full Attn. Share) where the keys and values are shared between each pair of images in the set. In the second variant (W.A. AdaIN) we omit the AdaIN operation over queries and keys. As expected, this Full Attn. Share variant, results with higher style consistency and lower text alignment. As can be seen in Fig. 5, Full Attn. Share harms the overall quality of the image sets and diversity across sets. Moreover, our method without the use of AdaIN results in much lower style consistency. Qualitative results can be seen in Fig. 5 and in the supplementary.. + +![](images/90b4c6812a23bc64e0005d692fb5d98823a2a1d002f9f220cbd0b2c2e09b02ba.jpg) +Reference image +A guitar + +![](images/9182dc023ae4f02985381ae8250dd67f27a32d6b2304ba17eb545d23e84d78b2.jpg) +StyleAligned +A hot air balloon + +![](images/5aca3abfb4d09a6d3e5db3cba5ca2de4f2cc89a6531e18621a8d0d39d20800c7.jpg) +A mountain + +![](images/c7948ce73d00ded74b3d54ff0bbc607a0b02b28b6b1d46b61271a0f2a65985d0.jpg) +StyleDrop (SDXL) +A hot air balloon + +![](images/a5994e42c5975622b839fe5edb1ae76d3ccbb6ab0107ba916a23dc3bad6ba32b.jpg) +A mountain + +![](images/5a0e47a4805c3521b0a200742e4d7a6efc924c7cb5a56dac468920f576d2922a.jpg) +StyleDrop (MUSE unofficial) +A hot air balloon + +![](images/46b6a33a75f26a0b999c095607dae2c6762844b5d675645eb58f683a47db58dc.jpg) +A mountain + +![](images/cf7b7ed8e76a13b03bed044b822d2266a9b6f5900b9b0d0dd4546a02b814c5f7.jpg) +Dreambooth-LoRA (SDXL) +A hot air balloon + +![](images/5df865f298975953ec22baa0012e6361e197e23bf7faab5934792aaa96698c2c.jpg) +A mountain + +![](images/20e2aaaf0c1e6ab47123702fef3e6d7744bc6664e05e999232247491622d8124.jpg) +A woman walking her dogs + +![](images/2900abc58cb353417dbcacb0858899b10ac88bc6db44cd9e37deb49575279ad4.jpg) +A friendly robot +Cherryblossom + +![](images/91f969d04a9a14bea047a577c124bb468e71870f92e46a74591a2accfd49efbf.jpg) + +![](images/934d4e99253578a4646ad35ef4a76dc408e1ce411b4fae9ab69fc49920db8ff8.jpg) +A friendly robot + +![](images/cddaf90b91fa69e42670e5a911d7462cb4857ff150b046a76dfa7ec3e37516d5.jpg) +Cherryblossom + +![](images/2ecc2accac53d7b2fc69a9b04e6409aabeb20377e722c3e746a29ad3763ec90b.jpg) +A robot + +![](images/eb1f356d8d9dbb4832264617ad7181e028801061eb592887032ae1c4e14a004b.jpg) +Cherryblos + +![](images/4a444f5f41709352b61c2f29db3900fe49fb1982d985bb2a52e3d55197d629c5.jpg) +A friendly robot +Cherryblossom + +![](images/610c9ba3314ca1226805491b5a9ce683756ab16a5b5f16aeebefaf025038cf13.jpg) + +![](images/8840c5026e15cd714d61ba2d56d55ab5f9290a381cb36c9d92013488ec9d9cd7.jpg) +A duck + +![](images/27d3adda1e9c1e8abb94cd5d3919471f7d4377e42f229c39b65589ea91c646c5.jpg) +A ship + +![](images/5a84f2c0a5fbafa354a5977511fe2a587e3984e96500383873d23ee76b6ca112.jpg) +A rocket + +![](images/7f110362dbb4193ed2c122926456ca32cf0ebff2b338a2f79b2ba1a19ea9abe1.jpg) +A ship + +![](images/e719c9a00052dcead3a2edfa21f5306cf25728b1e523e32e19639193911e3f21.jpg) +A rocket + +![](images/520144230c3f6cda53ff73278bfc318dc00a4a7b4f25acc8f6aad269a22713e0.jpg) +A ship + +![](images/f30b62b042d2e1a7bc31ea8397e082c8f27c08c36af31b9f8e439353ddb0b106.jpg) +A car + +![](images/c120d4a94419eaaf18f09eb87f57ff0e51bc783cbfc3b7df039a20de156ed2fa.jpg) +A ship + +![](images/9d94bbdfcfdd7eeb1b3052b31dc2b9ffcd4e5c7236d8b171d9ae33c032af0cff.jpg) +A rocket + +![](images/d7ddbb290dfd3012f5aa02c56956acb097eef5b8f80243a3deaff12c12b286c4.jpg) +A cute bear + +![](images/eb8f5046cf9749c732d4fd406a4560c5b3b793e5ae5ad161bdccdc6d44f34b1b.jpg) +A moose + +![](images/101e1170f7064b5edc9f04d96143713764a19191953362df139e3913fa60e039.jpg) +A baby pen + +![](images/0f3d2fa0c62c9e0e5edf09ed204fd59daf986792ce49f36c091cb0568c688e05.jpg) +A moose +A baby penguin + +![](images/1e88d7841435824dd9ec845f5d2d1d4131c53799ec14308bbce9203f29322916.jpg) + +![](images/066e32b3d3e53446c8e21b056da640992280644c433de7ec203d5b58ab4d1c94.jpg) +A moose +A baby penguin + +![](images/d493c3f5b38c916496d273246a57c83b152ee12941106b89e07e94b18f98934f.jpg) + +![](images/98bd1600692848c1d1e32c87ffbbaad7e8831f1325c1d591c3d4836fded90ab6.jpg) +A moose +A baby penguin + +![](images/07dab4a6b25a0ac834d7e5208ec8f5ee690714b8ca83182f7753e0c5f5c91b88.jpg) + +![](images/0e75922bafab63fd0dc43b2a236d62036d8df5d8480aa1a4c0b98594e1613c20.jpg) +A camera + +![](images/fde80ce0ea8f158811b6028048d98b0700eaf3363d4247ad20510ff14bdce6ce.jpg) +A fireman + +![](images/3613566046215aec7374cc7ae12d282ad8e812100a330aa8355d1ff5806af360.jpg) +A tall hill + +![](images/687587c656d7c57e108c490ea6e0f9a58ca98ec07e812b9a7d6dd6c239aad61e.jpg) +A fireman + +![](images/5f858526fbb5f99a5735067d56df198d3cd99d2fc790db7840d84b2fe9da5d37.jpg) +A tall hill + +![](images/a5ecc311b5dbeb6b2baedaef213a0d6b54334591474cb728be06ab329d4edeef.jpg) +A fireman + +![](images/5f0710c2366220dd26304908de510520344952d40686136117fcf3bf7c907660.jpg) +A tall hill + +![](images/de633264efa8f257c089060d149edbe67c2102974cb50c9648219b00e7729e36.jpg) +A fireman + +![](images/848ac7555642c0cea82a2611ee7dab8c5f40775618c28cab35f1d537b3a2204a.jpg) +A tall hill + +![](images/218c906898b3a5ae3af8a0c59cbfe3f2e6a876f559a6ed9dbed5f74cff59aca5.jpg) +A saxophone + +![](images/5a326baa1c71afd4fa92e7b945445d5df1a917fb68392dc96bde8df9060c0670.jpg) +A sunflower + +![](images/2e24b65414432bef1e741090cb6d689be11819594a2cdbe6f42fe4398c0cb96f.jpg) +A compass + +![](images/3bd7a2838367c664a9e337c39d2c1f17bc4b4d674a887d56572386b522ef23a5.jpg) +A sunflower + +![](images/a909ea9a7f355a26af4d2503b0d77e6884a4592309406b7b8a2b5cbf9376bc47.jpg) +A compass + +![](images/245f0efac9754680077bce0f2084c5a09d0700a318b35ddbc7df9e82dd856f22.jpg) +A sunflower + +![](images/1b207c7f42fabcce392a07367ce06138d3678cf13f55abd0b1abdda1ce539d27.jpg) +A compass + +![](images/0e392ae6870353c15e4c33cb5e500f566f4a5367ee5622f56e67fdfb217bce2f.jpg) +A sunflower + +![](images/2a20746fe4b8fc5b42a361f53ef12c4712aebf9df993f31740b2c8624cbe67e5.jpg) +A compass + +![](images/b266b1b05456e22c05ece494dafbad1df540c9dd01234ce23f0846dcfdab7187.jpg) +A bottle of wine + +![](images/0cc1655eb494e44e0a04cd133ad64293c8f42e2d1a50b974f0592bdb3ef10b6a.jpg) +Scones + +![](images/c0febaa1d6f4af1489962cdf2c022e226c99f50d6390739b462f03a7e5894d83.jpg) +A muffin + +![](images/a67f1819a5ac9832c842ffec3c5c2bea10ef041df432dc1197e1a531b7a6f866.jpg) +Scones + +![](images/273b7d99f4027b80fb94d2a712e1526dccae55b952daf95ebaf79781277e4638.jpg) +A muffin + +![](images/9729ac558578c2cc718793e23350acb7c4e090b725969a2dc8c12431c4a194f0.jpg) +A fireman + +![](images/8fd9cff8bb4ccfef3049b2a2e4eb07b940bcfdadfe156730a753ba68f0cf99eb.jpg) +A muffin + +![](images/3bc108ed9e3f17a44c0e1c114055d5844ede0e28a61f0b1c4879b10da89fb618.jpg) +Scones + +![](images/e20646a086cffac4ae116315a50305af6784ccc6c59bd2082f4bdeee1ee46165.jpg) +A muffin + +![](images/2fa74eae159321dd188194042693392198fcb61858fa854a7b9382491534aef9.jpg) +A wise owl + +![](images/4f563e863ce68571d9fa9ceae9bb07ae33f3fb60e146a0dab8488c410f327a6e.jpg) +Full moon +Figure 7. Qualitative comparison to personalization based methods. + +![](images/39fa8d35e7e3ba2427e87d43ffa4924cc93aae4c1d0d493c156b7c5d1acb4cdd.jpg) +A book + +![](images/2bb7b8d91f4af638659972a0ca014ec10d525a15fc9566d0869f376937cbc8a6.jpg) +Full moon + +![](images/bc6a9fbaf50c348093b27ffe01d00455c1a6a322d45a787eed6f99fd26ddc061.jpg) +A book + +![](images/d5733e81801d09ddb91dd13aa468c0647906a7565c02d87e74e2a016c17152db.jpg) +Full moon + +![](images/3a71945b845f2970193a4628958143b266d4488bfa7f69cf7b4ced283b29b9c6.jpg) +A book + +![](images/ec4bb150faf2d7acfdb5c2bcb832dccc3d4f65619f949822f0f87ff929759a87.jpg) +Full moon + +![](images/991f0ea54f6cbd853d334702a4074807d20a62f8c656ffa058f21edf2e85dc5d.jpg) +A book + +![](images/ada735eb1902ca79d5ab11a26e4fe25d02bb9659ddec19de02a6b75738006b8c.jpg) +Figure 8. Varying level of attention sharing. By reducing the number of shared attention layers, i.e., allowing only self-attention in part of the layers, we can get more varied results (bottom rows) at the expense of style alignment (top row). + +# 4.2. Comparisons + +For baselines, we compare our method to T2I personalization methods. We trained StyleDrop [49] and DreamBooth [41] over the first image in each set of our evaluation data, and use the trained personalized weights to generate the additional three images in each set. We use a public unofficial implementation of StyleDrop * (SDRP-unofficial) over non-regressive T2I model. Due to the large quality gap between the unofficial MUSE model † to the official one [7], we follow StyleDrop and implement an adapter model over SDXL (SDRP-SDXL) *, where we train a low rank linear layer after each Feed-Forward layer at the model's attention blocks. For training DreamBooth, we adapt the LoRA [21, 43] variant (DB-LoRA) over SDXL using the public huggingface-diffusers implementation ‡. We follow the hyperparameters tuning reported in [49] and train both SDRP-SDXL and DB-LoRA for 400 steps to prevent overfitting to the style training image. + +As can be seen in the qualitative comparison, Fig. 7, the image sets generated by our method, are more consistent across style attributes like color palette, drawing style, and composition. Moreover, the personalization-based methods + +may leak the content of the training reference image (on the left) when generating the new images. For example, see the repeated woman and dogs in the results of DB-LoRA and SDRP-SDXL at the second row or the repeated owl at the bottom row. Similarly, because of the content leakage, these methods obtained lower text similarity scores and higher set consistency scores compared to our method. + +We also apply two encoder-based personalization methods ELITE [57], IP-Adapter [59], and BLIP-Diffusion [27] over our evaluation set. These methods receive as input the first image in each set and use its embeddings to generate images with other content. Unlike the optimization-based techniques, these methods operate in a much faster feedforward diffusion loop, like our method. However, as can be seen in Fig. 6, their performance for style aligned image generation is poor compared to the other baselines. We argue that current encoder-based personalization techniques struggle to disentangle the content and the style of the input image. We supply qualitative results in the supplementary. + +User study. In addition to the automatic evaluation, we conducted a user study over the results of our method, StyleDrop (unofficial MUSE), StyleDrop (SDXL), and DreamBooth-LoRA (SDXL). In each question, we randomly sample one of the evaluation examples and show + +![](images/68ce1e879762b8e47f9f2caa41262a4af139d672c99f321df933fc63e9d4f341.jpg) +Figure 9. Style aligned image generation to an input image. Given an input reference image (left column) and text description, we first apply DDIM inversion over the image to get the inverted diffusion trajectory $x_{T}$ , $x_{T-1} \ldots x_{0}$ . Then, starting from $x_{T}$ and a new set of prompts, we apply our method to generate new content (right columns) with an aligned style to the input. + +the user the 4 image set that resulted from our and another method (in a random order). The user had to choose which set is better in terms of style consistency, and text alignment. A print screen of the user study format is provided in the supplementary. Overall, we collected 2400 answers from 100 users using the Amazon Mechanical Turk service. The results are summarized in Tab. 1 where for each method, we report the percentage of judgments in our favor. As can be seen, most participants favored our method by a large margin. More information about our user study can be found in the supplementary. + +# 4.3. Additional Results + +Style alignment control. We provide means of control over the style alignment to the reference image by applying the shared attention over only part of the self-attention layers. As can be seen in Fig. 8, reducing the number of shared attention layers results with a more diverse image set, which still shares common attributes with the reference image. + +StyleAligned from an input image. To generate style-aligned images to an input image, we apply DDIM inversion [50] using a provided text caption. Then, we apply our method to generate new images in the style of the input using the inverted diffusion trajectory $x_{T}, x_{T-1}, \ldots, x_{0}$ for the reference image. Examples are shown in Fig. 9, where we use BLIP captioning [28] to get a caption for each input image. For example, we used the prompt "A render of a house with a yellow roof" for the DDIM inversion of the top example and replaced the word house with other objects to generate the style-aligned images of a car, a cat, and a cactus. Notice that this method does not require any optimization. However, DDIM inversion may fail [30] or results with an erroneous trajectory [24]. More results and analysis, are provided in the supplementary. + +StyleAligned with other methods. Since our method doesn't require training or optimization, it can be easily combined on top of other diffusion based methods to generate style-consistent image sets. Fig. 10 shows several such examples where we combine our method with ControlNet [60], DreamBooth [42] and MultiDiffusion [3]. More examples and details about the integration of StyleAligned with other methods can be found in supplementary. + +# 5. Conclusions + +We have presented StyleAligned, which addresses the challenge of achieving style-aligned image generation within the realm of large-scale Text-to-Image models. By introducing minimal attention sharing operations with AdaIN modulation during the diffusion process, our method successfully establishes style-consistency and visual coherence across generated images. The demonstrated efficacy of StyleAligned in producing high-quality, style-consistent images across diverse styles and textual prompts under + +![](images/48eab7e15d523b557b2e445a5c7f8acebe22b8794caccc6a4756084a368b7c75.jpg) +Figure 10. StyleAligned with other methods. On top, StyleAligned is combined with ControlNet to generate style-aligned images conditioned on depth maps. In the middle, our method combined with MultiDiffusion to generate panorama images that share multiple styles. On the bottom, style consistent and personalized content created by combining our method with pretrained personalized DreamBooth-LoRA models. + +scores its potential in creative domains and practical applications. Our results affirm StyleAligned capability to faithfully adhere to provided descriptions and reference styles while maintaining impressive synthesis quality. + +In the future we would like to explore the scalability and adaptability of StyleAligned to have more control over the shape and appearance similarity among the generated images. Additionally, due to the limitation of current diffusion inversion methods, a promising direction is to leverage StyleAligned to assemble a style-aligned dataset which then can be used to train style condition text-to-image models. + +# 6. Acknowledgement + +We thank Or Patashnik, Matan Cohen, Yael Pritch, and Yael Vinker for their valuable inputs that helped improve this work. + +# References + +[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer, 2023. 2 +[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Trans. Graph., 42(4), jul 2023. 2 +[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 8 +[4] James Betker, Gabriel Goh, Li Jing, TimBrooks, Jianfeng Wang, Linjie Li, LongOuyang, JuntangZhuang, JoyceLee, YufeiGuo, WesamManassra, PrafullaDhariwal, CaseyChu, YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 3 +[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. MasaCtrl: tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22560–22570, October 2023. 2 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5 +[7] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, José Lezama, Lu Jiang, Ming Yang, Kevin P. Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. In International Conference on Machine Learning, 2023. 2, 7 +[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42:1 - 10, 2023. 2 +[9] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In The Eleventh International Conference on Learning Representations, 2022. 2 +[10] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 571-576. 2023. 2 +[11] Alexei A Efros and Thomas K Leung. Texture synthesis by non-parametric sampling. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 1033-1038. IEEE, 1999. 2 +[12] Dave Epstein, Allan Jabri, Ben Poole, Alexei A Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 2 +[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, 2022. 1, 2 +[14] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain + +tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 2 +[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2 +[16] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris N. Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. ArXiv, abs/2303.11305, 2023. 2 +[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2 +[18] Aaron Hertzmann, Charles E Jacobs, Nuria Oliver, Brian Curless, and David H Salesin. Image analogies. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 557-570. 2023. 2 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proc. NeurIPS, 2020. 3 +[20] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2 +[21] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 7 +[22] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 5 +[23] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European conference on computer vision (ECCV), pages 172-189, 2018. 2 +[24] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. arXiv preprint arXiv:2304.06140, 2023. 2, 8 +[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2 +[26] Hochang Lee, Sanghyun Seo, Seungtaek Ryoo, and Kyunghyun Yoon. Directional texture transfer. In Proceedings of the 8th International Symposium on Non-Photorealistic Animation and Rendering, pages 43–48, 2010. 2 +[27] Dongxu Li, Junnan Li, and Steven C. H. Hoi. BLIP-Diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. ArXiv, abs/2305.14720, 2023. 2, 7 +[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 8 + +[29] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsupervised image-to-image translation. In IEEE International Conference on Computer Vision (ICCV), 2019. 2 +[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2, 8 +[31] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, 2021. 2 +[32] Dong Huk Park*, Grace Luo*, Clayton Toste, Samaneh Azadi, Xihui Liu, Maka Karalashvili, Anna Rohrbach, and Trevor Darrell. Shape-guided diffusion with inside-outside attention. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 2 +[33] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2 +[34] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. ArXiv, abs/2303.11306, 2023. 2 +[35] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 2, 3, 5 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 5 +[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1 +[38] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 2 +[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1 +[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, + +Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3 +[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 5, 7 +[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 2, 8 +[43] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora, 2022.7 +[44] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. ACM SIGGRAPH 2022 Conference Proceedings, 2021. 2 +[45] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1 +[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2, 3 +[47] Jing Shi, Wei Xiong, Zhe L. Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. ArXiv, abs/2304.03411, 2023. 2 +[48] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3 +[49] Kihyuk Sohn, Nataniel Ruiz, Kimin Lee, Daniel Castro Chin, Irina Blok, Huiwen Chang, Jarred Barber, Lu Jiang, Glenn Entis, Yuanzhen Li, et al. Styledrop: Text-to-image generation in any style. arXiv preprint arXiv:2306.00983, 2023. 1, 2, 7 +[50] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 8 +[51] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. SIGGRAPH 2023 Conference Proceedings, 2023. 2 +[52] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10748-10757, 2022. 2 + +[53] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2 +[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 3 +[55] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. arXiv preprint arXiv:2211.13752, 2022. 2 +[56] Andrey Voynov, Q. Chu, Daniel Cohen-Or, and Kfir Aberman. P+: Extended textual conditioning in text-to-image generation. ArXiv, abs/2303.09522, 2023. 2, 5 +[57] Yuxiang Wei. Official implementation of ELITE. https://github.com/csyxwei/ELITE, 2023. 7 +[58] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: Encoding visual concepts into textual embeddings for customized text-to-image generation. ArXiv, abs/2302.13848, 2023. 2 +[59] Hu Ye, Jun Zhang, Siyi Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. ArXiv, abs/2308.06721, 2023. 2, 7 +[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 8 \ No newline at end of file diff --git a/2024/Style Aligned Image Generation via Shared Attention/images.zip b/2024/Style Aligned Image Generation via Shared Attention/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9db38157b0923c47a74c0bbe147a4770767ff561 --- /dev/null +++ b/2024/Style Aligned Image Generation via Shared Attention/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fc5af7e8bacb904de7cfdcd2f59cacb00c2ad3e365d4925b98b8ec5d448a95f +size 1404150 diff --git a/2024/Style Aligned Image Generation via Shared Attention/layout.json b/2024/Style Aligned Image Generation via Shared Attention/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..93d47bfdcbfc16360424cbaf061c75f97ec7a541 --- /dev/null +++ b/2024/Style Aligned Image Generation via Shared Attention/layout.json @@ -0,0 +1,12992 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 131, + 103, + 463, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 103, + 463, + 121 + ], + "spans": [ + { + "bbox": [ + 131, + 103, + 463, + 121 + ], + "type": "text", + "content": "Style Aligned Image Generation via Shared Attention" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 142, + 487, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 142, + 487, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 487, + 186 + ], + "type": "text", + "content": "Amir Hertz\\*, Andrey Voynov\\*, Shlomi Fruchter\\*, and Daniel Cohen-Or\\*1,2 \n1 Google Research \n2 Tel Aviv University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 238, + 290, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 441 + ], + "type": "text", + "content": "Large-scale Text-to-Image (T2I) models have rapidly gained prominence across creative fields, generating visually compelling outputs from textual prompts. However, controlling these models to ensure consistent style remains challenging, with existing methods necessitating fine-tuning and manual intervention to disentangle content and style. In this paper, we introduce StyleAligned, a novel technique designed to establish style alignment among a series of generated images. By employing minimal 'attention sharing' during the diffusion process, our method maintains style consistency across images within T2I models. This approach allows for the creation of style-consistent images using a reference style through a straightforward inversion operation. Our method's evaluation across diverse styles and text prompts demonstrates high-quality synthesis and fidelity, underscoring its efficacy in achieving consistent style across various inputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 463, + 128, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 463, + 128, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 463, + 128, + 475 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 483, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 555 + ], + "type": "text", + "content": "Large-scale Text-to-Image (T2I) generative models [37, 39, 45] have emerged as an essential tool across creative disciplines, such as art, graphic design, animation, architecture, gaming and more. These models show tremendous capabilities of translating an input text into an appealing visual result that is aligned with the input description." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 555, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 650 + ], + "type": "text", + "content": "An envisioned application of T2I models revolves around the rendition of various concepts in a way that shares a consistent style and character, as though all were created by the same artist and method (see Fig. 1). While proficient in aligning with the textual description of the style, state-of-the-art T2I models often create images that diverge significantly in their interpretations of the same stylistic descriptor, as depicted in Fig. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 651, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 651, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 651, + 287, + 687 + ], + "type": "text", + "content": "Recent methods mitigate this by fine-tuning the T2I model over a set of images that share the same style [13,49]. This optimization is computationally expensive and usually" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 211, + 547, + 488 + ], + "blocks": [ + { + "bbox": [ + 306, + 211, + 547, + 488 + ], + "lines": [ + { + "bbox": [ + 306, + 211, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 306, + 211, + 547, + 488 + ], + "type": "image", + "image_path": "befe72dd08ea13e65a43000c1dc25189dcdb8c0c0bf181c36f354ee9aa00c776.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 498, + 547, + 532 + ], + "lines": [ + { + "bbox": [ + 305, + 498, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 547, + 532 + ], + "type": "text", + "content": "Figure 1. Style aligned image set generation. By fusing the features of the toy train image (left) during the diffusion process, we can generate an image set of different content that shares the style." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 552, + 547, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 547, + 588 + ], + "type": "text", + "content": "requires human input in order to find a plausible subset of images and texts that enables the disentanglement of content and style." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 588, + 547, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 547, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 547, + 708 + ], + "type": "text", + "content": "We introduce StyleAligned, a method that enables consistent style interpretation across a set of generated images (Fig. 1). Our method requires no optimization and can be applied to any attention-based text-to-image diffusion model. We show that adding minimal attention sharing operations along the diffusion process, from each generated image to the first one in a batch, leads to a style-consistent set. Moreover, using diffusion inversion, our method can be applied to generate style-consistent images given a reference style image, with no optimization or fine-tuning." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 129, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 129, + 702 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 129, + 702 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 61, + 703, + 121, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 121, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 121, + 713 + ], + "type": "text", + "content": "Equal Advising." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4775" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 287, + 305 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 287, + 305 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 287, + 305 + ], + "type": "image", + "image_path": "cca67eff1325ad439ed8bffe4784e7bbf26f6fce3d87078922c54c3404fd3b16.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 313, + 288, + 369 + ], + "lines": [ + { + "bbox": [ + 46, + 313, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 288, + 369 + ], + "type": "text", + "content": "Figure 2. Standard text-to-image vs. StyleAligned set generation. Given style description of \"minimal origami\", standard text-to-image generation (top) results with an unaligned image set while our method (bottom) can generate variety of style aligned content." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 389, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 510 + ], + "type": "text", + "content": "We present our results over diverse styles and text prompts, demonstrating high-quality synthesis and fidelity to the prompts and reference style. We show diverse examples of generated images that share their style with a reference image that can possibly be a given input image. Importantly, our technique stands as a zero-shot solution, distinct from other personalization techniques, as it operates without any form of optimization or fine-tuning. For our code and more examples, please visit the project page style-aligned-gen.github.io" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 520, + 134, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 134, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 134, + 533 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 540, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 635 + ], + "type": "text", + "content": "Text-to-image generation. Text conditioned image generative models [7, 31, 38] show unprecedented capabilities of generating high quality images from text descriptions. In particular, T2I diffusion models [35, 38, 46] are pushing the state of the art and they are quickly adopted for different generative visual tasks like inpainting [2, 44], imaged-to-image translation [55, 59], local image editing [9, 24], subject-driven image generation [42, 51] and more." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 635, + 287, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 287, + 707 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 287, + 707 + ], + "type": "text", + "content": "Attention control in diffusion models. Hertz et al. [17] have shown how cross and self-attention maps within the diffusion process determine the layout and content of the generated images. Moreover, they showed how the attention maps can be used for controlled image generation. Other studies have leveraged modifications in attention lay" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "ers to enhance the fidelity or diversity of generated images [8, 34], or apply attention control for image editing [5, 12, 30, 32, 33, 53]. However, in contrast to prior approaches that primarily enable structure-preserving image editing, our method excels at generating images with diverse structures and content while maintaining a consistent style interpretation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 156, + 546, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 546, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 546, + 371 + ], + "type": "text", + "content": "Style transfer. Transferring a style from a reference image to a target content image is well studied subject in computer graphics. Classic works [10, 11, 18, 26] rely on optimization of handcrafted features and texture resampling algorithms from an input texture image, combined with structure constrains of a content image. With the progress of deep learning research, another line of works utilizes deep neural priors for style transfer optimization using deep features of pre-trained networks [15, 52], or injecting attention features from a style image to a target one [1]. More related to our approach, Huang et al. [22] introduced a real time style transfer network based on Adaptive Instance Normalization layers (AdaIN) that are used to normalize deep features of a target image using deep features statistics of a reference style image. Follow-up works, employ the AdaIN layer for additional unsupervised learning tasks, like style-based image generation [25] and Image2Image translation [23, 29]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 371, + 546, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 546, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 546, + 550 + ], + "type": "text", + "content": "T2I personalization To generalize T2I over new visual concepts, several works developed different optimization techniques over a small collection of input images that share the same concept [13, 16, 42, 56]. In instances where the collection shares a consistent style, the acquired concept becomes the style itself, affecting subsequent generations. Most close to our work is StyleDrop [49], a style personalization method that relies on fine-tuning of light weight adapter layers [20] at the end of each attention block in a non-autoregressive generative text-to-image transformer [7]. StyleDrop can generate a set of images in the same style of by training the adapter layers over a collection of images that share the same style. However, it struggles to generate a consistent image set of different content when training on a single image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 551, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 551, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 551, + 545, + 646 + ], + "type": "text", + "content": "Our method can generate a consistent image set without optimization phase and without relying on several images for training. To skip the training phase, recent works developed dedicated personalization encoders [14, 27, 47, 58, 59] that can directly inject new priors from a single input image to the T2I model. However, these methods encounter challenges to disentangle style from content as they focus on generating the same subject as in the input image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 657, + 409, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 409, + 669 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 409, + 669 + ], + "type": "text", + "content": "3. Method overview" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "In the following section we start with an overview of the T2I diffusion process, and in particular the self-attention mechanism Sec. 3.1. We continue by present" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4776" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 67, + 534, + 277 + ], + "blocks": [ + { + "bbox": [ + 47, + 67, + 534, + 277 + ], + "lines": [ + { + "bbox": [ + 47, + 67, + 534, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 67, + 534, + 277 + ], + "type": "image", + "image_path": "6bfec6c2924b31a525aed02432c3bbe38680437dfe55778583f7be18ac076611.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 284, + 547, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 284, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 547, + 308 + ], + "type": "text", + "content": "Figure 3. Style Aligned Diffusion. Generation of images with a style aligned to the reference image on the left. In each diffusion denoising step all the images, except the reference, perform a shared self-attention with the reference image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 87, + 325, + 252, + 571 + ], + "blocks": [ + { + "bbox": [ + 87, + 325, + 252, + 571 + ], + "lines": [ + { + "bbox": [ + 87, + 325, + 252, + 571 + ], + "spans": [ + { + "bbox": [ + 87, + 325, + 252, + 571 + ], + "type": "image", + "image_path": "40d62b020e5800a0fb2d06cae575f19a45f0d7e49c04264b0542055e71d15eec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 578, + 289, + 634 + ], + "lines": [ + { + "bbox": [ + 46, + 578, + 289, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 289, + 634 + ], + "type": "text", + "content": "Figure 4. Shared attention layer. The target images attend to the reference image by applying AdaIN over their queries and keys using the reference queries and keys respectively. Then, we apply shared attention where the target features are updated by both the target values " + }, + { + "bbox": [ + 46, + 578, + 289, + 634 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 46, + 578, + 289, + 634 + ], + "type": "text", + "content": " and the reference values " + }, + { + "bbox": [ + 46, + 578, + 289, + 634 + ], + "type": "inline_equation", + "content": "V_{r}" + }, + { + "bbox": [ + 46, + 578, + 289, + 634 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 654, + 288, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 680 + ], + "type": "text", + "content": "ing our attention-sharing operation within the self-attention layers that enable style aligned image set generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 327, + 392, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 327, + 392, + 338 + ], + "spans": [ + { + "bbox": [ + 306, + 327, + 392, + 338 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 345, + 546, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 345, + 546, + 405 + ], + "spans": [ + { + "bbox": [ + 305, + 345, + 546, + 405 + ], + "type": "text", + "content": "Diffusion models [19, 48] are generative latent variable models that aim to model a distribution " + }, + { + "bbox": [ + 305, + 345, + 546, + 405 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_0)" + }, + { + "bbox": [ + 305, + 345, + 546, + 405 + ], + "type": "text", + "content": " that approximates the data distribution " + }, + { + "bbox": [ + 305, + 345, + 546, + 405 + ], + "type": "inline_equation", + "content": "q(x_0)" + }, + { + "bbox": [ + 305, + 345, + 546, + 405 + ], + "type": "text", + "content": " and are easy to sample from. Diffusion models are trained to reverse the diffusion \"forward process\":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 341, + 413, + 509, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 413, + 509, + 427 + ], + "spans": [ + { + "bbox": [ + 341, + 413, + 509, + 427 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {\\alpha_ {t}} x _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\epsilon \\sim N (0, I),", + "image_path": "6bac531cd6adf14b1e4adcb9319b225f0b3bdbf21aa2290594e69f71bf7749cd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 435, + 547, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 435, + 547, + 484 + ], + "spans": [ + { + "bbox": [ + 305, + 435, + 547, + 484 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 435, + 547, + 484 + ], + "type": "inline_equation", + "content": "t \\in [0, \\infty)" + }, + { + "bbox": [ + 305, + 435, + 547, + 484 + ], + "type": "text", + "content": " and the values of " + }, + { + "bbox": [ + 305, + 435, + 547, + 484 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 305, + 435, + 547, + 484 + ], + "type": "text", + "content": " are determined by a scheduler such that " + }, + { + "bbox": [ + 305, + 435, + 547, + 484 + ], + "type": "inline_equation", + "content": "\\alpha_{0} = 1" + }, + { + "bbox": [ + 305, + 435, + 547, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 435, + 547, + 484 + ], + "type": "inline_equation", + "content": "\\lim_{t \\to \\infty} \\alpha_{t} = 0" + }, + { + "bbox": [ + 305, + 435, + 547, + 484 + ], + "type": "text", + "content": ". During inference, we sample an image by gradually denoising an input noise image " + }, + { + "bbox": [ + 305, + 435, + 547, + 484 + ], + "type": "inline_equation", + "content": "x_{T} \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 305, + 435, + 547, + 484 + ], + "type": "text", + "content": " via the reverse process:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 353, + 493, + 499, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 493, + 499, + 506 + ], + "spans": [ + { + "bbox": [ + 353, + 493, + 499, + 506 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = \\mu_ {t - 1} + \\sigma_ {t} z, z \\sim N (0, I),", + "image_path": "e0ef31116c3ae65437be7c81e35162ec507f32256e0a195fbf8657a279a28630.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 514, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 514, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 305, + 514, + 545, + 539 + ], + "type": "text", + "content": "where the value of " + }, + { + "bbox": [ + 305, + 514, + 545, + 539 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 305, + 514, + 545, + 539 + ], + "type": "text", + "content": " is determined by the sampler and " + }, + { + "bbox": [ + 305, + 514, + 545, + 539 + ], + "type": "inline_equation", + "content": "\\mu_{t - 1}" + }, + { + "bbox": [ + 305, + 514, + 545, + 539 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 546, + 545, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 546, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 546, + 545, + 574 + ], + "type": "interline_equation", + "content": "\\mu_ {t - 1} = \\frac {\\sqrt {\\alpha_ {t - 1}} x _ {t}}{\\sqrt {\\alpha_ {t}}} + \\left(\\sqrt {1 - \\alpha_ {t - 1}} - \\frac {\\sqrt {1 - \\alpha_ {t}}}{\\sqrt {\\alpha_ {t}}}\\right) \\epsilon_ {\\theta} (x _ {t}, t),", + "image_path": "b582e3ab7e8620a34911c0f07cd04738dd113c5bf450f96d5ec3701fd7be41fa.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 582, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 545, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 582, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 305, + 582, + 545, + 605 + ], + "type": "text", + "content": " is the output of a diffusion model parameterized by " + }, + { + "bbox": [ + 305, + 582, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\theta" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 606, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 606, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 305, + 606, + 545, + 666 + ], + "type": "text", + "content": "Moreover, this process can be generalized for learning a marginal distribution using an additional input condition. That leads text-to-image diffusion models (T2I), where the output of the model " + }, + { + "bbox": [ + 305, + 606, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,t,y)" + }, + { + "bbox": [ + 305, + 606, + 545, + 666 + ], + "type": "text", + "content": " is conditioned on a text prompt " + }, + { + "bbox": [ + 305, + 606, + 545, + 666 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 606, + 545, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 666, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 666, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 666, + 547, + 714 + ], + "type": "text", + "content": "Self-Attention in T2I diffusion models. State-of-the-art T2I diffusion models [4,35,46] employ a U-Net architecture [40] that consists of convolution layers and transformer attention blocks [54]. In these attention mechanisms, deep" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4777" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 289, + 432 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 289, + 432 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 289, + 432 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 289, + 432 + ], + "type": "image", + "image_path": "ba3e52212d6c2d820ad98b6bb7c6961b73723f21fa0dc4ed42a668f250f4f670.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 441, + 288, + 562 + ], + "lines": [ + { + "bbox": [ + 46, + 441, + 288, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 288, + 562 + ], + "type": "text", + "content": "Figure 5. Ablation study - qualitative comparison. Each pair of rows shows two sets of images generated by the same set of prompts \"...in minimal flat design illustration\" using different configurations of our method, and each row in a pair uses a different seed. Sharing the self-attention between all images in the set (bottom) results with some diversity loss (style collapse across many seeds) and content leakage within each set (colors from one image leak to another). Disabling the queries keys AdALN operation results with less consistent image sets compared to our full method (top) which keeps on both diversity between different sets and consistency within each set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "type": "text", + "content": "image features " + }, + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "type": "inline_equation", + "content": "\\phi \\in \\mathbb{R}^{m\\times d_h}" + }, + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "type": "text", + "content": " attend to each other via selfattention layers and to contextual text embedding via crossattention layers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 619, + 288, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 619, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 47, + 619, + 288, + 679 + ], + "type": "text", + "content": "Our work operates at the self-attention layers where deep features are being updated by attending to each other. First, the features are projected into queries " + }, + { + "bbox": [ + 47, + 619, + 288, + 679 + ], + "type": "inline_equation", + "content": "Q \\in m \\times d_k" + }, + { + "bbox": [ + 47, + 619, + 288, + 679 + ], + "type": "text", + "content": ", keys " + }, + { + "bbox": [ + 47, + 619, + 288, + 679 + ], + "type": "inline_equation", + "content": "K \\in m \\times d_k" + }, + { + "bbox": [ + 47, + 619, + 288, + 679 + ], + "type": "text", + "content": " and values " + }, + { + "bbox": [ + 47, + 619, + 288, + 679 + ], + "type": "inline_equation", + "content": "V \\in m \\times d_h" + }, + { + "bbox": [ + 47, + 619, + 288, + 679 + ], + "type": "text", + "content": " through learned linear layers. Then, the attention is computed by the scaled" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 72, + 396, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 396, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 396, + 83 + ], + "type": "text", + "content": "dot-product attention:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 334, + 92, + 517, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 92, + 517, + 121 + ], + "spans": [ + { + "bbox": [ + 334, + 92, + 517, + 121 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}} V\\right),", + "image_path": "1c7b75381121fdcbc02b5a54bb7e6f007b612d5e803664ebd6b774d09309a155.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "text", + "content": " is the dimension of " + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "text", + "content": ". Intuitively, each image feature is updated by a weighted sum of " + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "text", + "content": ", where the weight depends on the correlation between the projected query " + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "text", + "content": " and the keys " + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "text", + "content": ". In practice, each self-attention layer consists of several attention heads, and then the residual is computed by concatenating and projecting the attention heads output back to the image feature space " + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "inline_equation", + "content": "d_{h}" + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 353, + 220, + 498, + 234 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 220, + 498, + 234 + ], + "spans": [ + { + "bbox": [ + 353, + 220, + 498, + 234 + ], + "type": "interline_equation", + "content": "\\hat {\\phi} = \\phi + \\text {M u l t i - H e a d - A t t e n t i o n} (\\phi).", + "image_path": "d4314d92b450ebe0f541a99ce4af3fc16d47f1ad3fd5e26941ef8ef72f152d48.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 243, + 495, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 495, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 495, + 255 + ], + "type": "text", + "content": "3.2. Style Aligned Image Set Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 261, + 545, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 545, + 416 + ], + "type": "text", + "content": "The goal of our method is to generate a set of images " + }, + { + "bbox": [ + 304, + 261, + 545, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_1\\ldots \\mathcal{I}_n" + }, + { + "bbox": [ + 304, + 261, + 545, + 416 + ], + "type": "text", + "content": " that are aligned with an input set of text prompts " + }, + { + "bbox": [ + 304, + 261, + 545, + 416 + ], + "type": "inline_equation", + "content": "y_{1}\\dots y_{n}" + }, + { + "bbox": [ + 304, + 261, + 545, + 416 + ], + "type": "text", + "content": " and share a consistent style interpretation with each other. For example, see the garnered image set of toy objects in Fig. 3 that are style-aligned with each other and to the input text on top. A naive way to generate a style aligned image set of different content is to use the same style description in the text prompts. As can be seen at the bottom of Fig. 2, generating different images using a shared style description of \"in minimal origami style\" results in an unaligned set, since each image is unaware of the exact appearance of other images in the set during the generation process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 417, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 545, + 464 + ], + "type": "text", + "content": "The key insight underlying our approach is the utilization of the self-attention mechanism to allow communication among various generated images. This is achieved by sharing attention layers across the generated images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 464, + 545, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 501 + ], + "type": "text", + "content": "Formally, let " + }, + { + "bbox": [ + 304, + 464, + 545, + 501 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 304, + 464, + 545, + 501 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 464, + 545, + 501 + ], + "type": "inline_equation", + "content": "K_{i}" + }, + { + "bbox": [ + 304, + 464, + 545, + 501 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 464, + 545, + 501 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 304, + 464, + 545, + 501 + ], + "type": "text", + "content": " be the queries, keys, and values, projected from deep features " + }, + { + "bbox": [ + 304, + 464, + 545, + 501 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 304, + 464, + 545, + 501 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 464, + 545, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{i}" + }, + { + "bbox": [ + 304, + 464, + 545, + 501 + ], + "type": "text", + "content": " in the set, then, the attention update for " + }, + { + "bbox": [ + 304, + 464, + 545, + 501 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 304, + 464, + 545, + 501 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 366, + 510, + 545, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 510, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 366, + 510, + 545, + 525 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} \\left(Q _ {i}, K _ {1 \\dots n}, V _ {1 \\dots n}\\right), \\tag {1}", + "image_path": "9caaf50f91e972c82e2d494899ca902ebfc3ec8f485df20128941cc59d99eac7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 531, + 545, + 589 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 531, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 306, + 531, + 545, + 589 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} K _ {1 \\dots n} = \\left[ \\begin{array}{c} K _ {1} \\\\ K _ {2} \\\\ \\vdots \\\\ K _ {n} \\end{array} \\right] \\text {a n d} V _ {1 \\dots n} = \\left[ \\begin{array}{c} V _ {1} \\\\ V _ {2} \\\\ \\vdots \\\\ V _ {n} \\end{array} \\right]. \\text {H o w e v e r},", + "image_path": "24a4ee67a9a24beb8aa84ea100f1d365a45239ecc4f00f0156ca1da9cdd75f92.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 591, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 545, + 686 + ], + "type": "text", + "content": "have noticed that by enabling full attention sharing, we may harm the quality of the generated set. As can be seen in Fig. 5 (bottom rows), full attention sharing results in content leakage among the images. For example, the unicorns got green paint from the garnered dino in the set. Moreover, full attention sharing results with less diverse sets of the same set of prompts, see the two sets in Fig. 5 in bottom rows compared to the sets above." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 687, + 545, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 687, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 306, + 687, + 545, + 711 + ], + "type": "text", + "content": "To restrict the content leakage and allow diverse sets, we share the attention to only one image in the generated set" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4778" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 278, + 199 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 278, + 199 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 278, + 199 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 278, + 199 + ], + "type": "image", + "image_path": "925b1dcfbe4f36bc77e342a38b71d524bffa46834c75e5708976954f11d2d13e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 289, + 253 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 289, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 289, + 253 + ], + "type": "text", + "content": "Figure 6. Quantitative Comparison. We compare the results of the different methods (blue marks) and our ablation experiments (orange marks) in terms of text alignment (CLIP score) and set consistency (DINO embedding similarity)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 323, + 272, + 361 + ], + "blocks": [ + { + "bbox": [ + 46, + 259, + 287, + 315 + ], + "lines": [ + { + "bbox": [ + 46, + 259, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 287, + 315 + ], + "type": "text", + "content": "Table 1. User evaluation for style aligned image set generation. In each question, the user was asked to select between two image sets, Which is better in terms of style consistency and match to the text descriptions (see Sec. 4). We report the percentage of judgments in favor of StyleAligned over 800 answers (2400 in total)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 323, + 272, + 361 + ], + "lines": [ + { + "bbox": [ + 59, + 323, + 272, + 361 + ], + "spans": [ + { + "bbox": [ + 59, + 323, + 272, + 361 + ], + "type": "table", + "html": "
StyleDrop (unofficial MUSE)StyleDrop (SDXL)DreamBooth-LoRA (SDXL)
85.2 %67.1 %61.3%
", + "image_path": "247a44494f02213f2b2b7c91ffa2bee923897c85a965aa4f01f500d88bcf6752.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 372, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 467 + ], + "type": "text", + "content": "(the first in the batch). That is, target image features " + }, + { + "bbox": [ + 46, + 372, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\phi_t" + }, + { + "bbox": [ + 46, + 372, + 287, + 467 + ], + "type": "text", + "content": " are attending to themselves and to the features of only one reference image in the set using Eq. 1. As can be seen in Fig. 5 (middle), sharing the attention to only one image in the set results in diverse sets that share a similar style. However, in that case, we have noticed that the style of different images is not well aligned. We suspect that this is due to low attention flow from the reference to the target image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "text", + "content": "As illustrated in Fig. 4, to enable balanced attention reference, we normalize the queries " + }, + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "inline_equation", + "content": "Q_{t}" + }, + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "text", + "content": " and keys " + }, + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "inline_equation", + "content": "K_{t}" + }, + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "text", + "content": " of the target image using the queries " + }, + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "inline_equation", + "content": "Q_{r}" + }, + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "text", + "content": " and keys " + }, + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "inline_equation", + "content": "K_{r}" + }, + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "text", + "content": " of the reference image using the adaptive normalization operation (AdaIN) [22]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 536, + 264, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 536, + 264, + 551 + ], + "spans": [ + { + "bbox": [ + 69, + 536, + 264, + 551 + ], + "type": "interline_equation", + "content": "\\hat {Q} _ {t} = \\operatorname {A d a I N} \\left(Q _ {t}, Q _ {r}\\right) \\quad \\hat {K} _ {t} = \\operatorname {A d a I N} \\left(K _ {t}, K _ {r}\\right),", + "image_path": "e6db49057c09ea4718172f50951a20ec9859b237aa988216ae15f3dbc668c41f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 559, + 206, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 206, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 206, + 571 + ], + "type": "text", + "content": "where the AdaIn operation is given by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 82, + 580, + 253, + 606 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 580, + 253, + 606 + ], + "spans": [ + { + "bbox": [ + 82, + 580, + 253, + 606 + ], + "type": "interline_equation", + "content": "\\operatorname {A d a I N} (x, y) = \\sigma (y) \\left(\\frac {x - \\mu (x)}{\\sigma (x)}\\right) + \\mu_ {y},", + "image_path": "0c6433a89491ffb91c7cdabdef5c64ca76ee07c92428a7e8cbbbfc01554547b0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 615, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 615, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 287, + 652 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 46, + 615, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mu (x),\\sigma (x)\\in \\mathbb{R}^{d_k}" + }, + { + "bbox": [ + 46, + 615, + 287, + 652 + ], + "type": "text", + "content": " are the mean and the standard deviation of queries and keys across different pixels. Finally, our shared attention is given by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 659, + 219, + 674 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 659, + 219, + 674 + ], + "spans": [ + { + "bbox": [ + 115, + 659, + 219, + 674 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} \\left(\\hat {Q} _ {t}, K _ {r t} ^ {T}, V _ {r t}\\right),", + "image_path": "23af8d1dda00a1834e4113c6b291258790593d653033051466b97dd08404b6aa.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 682, + 200, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 682, + 200, + 710 + ], + "spans": [ + { + "bbox": [ + 47, + 682, + 200, + 710 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 682, + 200, + 710 + ], + "type": "inline_equation", + "content": "K_{rt} = \\left[ \\begin{array}{c}K_{r}\\\\ \\hat{K}_{t} \\end{array} \\right]" + }, + { + "bbox": [ + 47, + 682, + 200, + 710 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 682, + 200, + 710 + ], + "type": "inline_equation", + "content": "V_{rt} = \\left[ \\begin{array}{c}V_{r}\\\\ V_{t} \\end{array} \\right]" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 71, + 471, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 471, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 471, + 85 + ], + "type": "text", + "content": "4. Evaluations and Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 91, + 545, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 186 + ], + "type": "text", + "content": "We have implemented our method over Stable Diffusion XL (SDXL) [35] by applying our attention sharing overall 70 self-attention layers of the model. The generation of a four images set takes 29 seconds on a single A100 GPU. Notice that since the generation of the reference image is not influenced by other images in the batch, we can generate larger sets by fixing the prompt and seed of the reference image across the set generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 186, + 485, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 485, + 198 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 485, + 198 + ], + "type": "text", + "content": "For example, see the sets in Fig. 2 and 3." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 199, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 199, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 199, + 545, + 281 + ], + "type": "text", + "content": "Evaluation set. With the support of ChatGPT, we have generated 100 text prompts describing different image styles over four random objects. For example, “{A guitar, A hot air balloon, A sailboat, A mountain} in papercut art style.” For each style and set of objects, we use our method to generate a set of images. The full list of prompts is provided in the appendix." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 282, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 545, + 437 + ], + "type": "text", + "content": "Metrics. To verify that each generated image contains its specified object, we measure the CLIP similarity [36] between the image and the text description of the object. In addition, we evaluate the style consistency of each generated set, by measuring the pairwise average cosine similarity between DINO VIT-B/8 [6] embeddings of the generated images in each set. Following [41, 56], we used DINO embeddings instead of CLIP image embeddings for measuring image similarity, since CLIP was trained with class labels and therefore it might give a high score for different images in the set that have similar content but with a different style. On the other hand, DINO better distinguishes between different styles due to its self-supervised training." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 444, + 400, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 400, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 400, + 456 + ], + "type": "text", + "content": "4.1. Ablation Study" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "text", + "content": "The quantitative results are summarized in Fig. 6, where the right-top place on the chart means better text similarity and style consistency, respectively. As a reference, we report the score obtained by generating the set of images using SDXL (T2I Reference) using the same seeds without any intervention. As can be seen, our method achieves a much higher style consistency score at the expense of text similarity. See qualitative comparison in Fig. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "In addition, we compared our method to additional two variants of the shared attention as described in Sec. 3.2. The first variant uses full attention sharing (Full Attn. Share) where the keys and values are shared between each pair of images in the set. In the second variant (W.A. AdaIN) we omit the AdaIN operation over queries and keys. As expected, this Full Attn. Share variant, results with higher style consistency and lower text alignment. As can be seen in Fig. 5, Full Attn. Share harms the overall quality of the image sets and diversity across sets. Moreover, our method without the use of AdaIN results in much lower style consistency. Qualitative results can be seen in Fig. 5 and in the supplementary.." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4779" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 81, + 123, + 150 + ], + "blocks": [ + { + "bbox": [ + 57, + 72, + 117, + 80 + ], + "lines": [ + { + "bbox": [ + 57, + 72, + 117, + 80 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 117, + 80 + ], + "type": "text", + "content": "Reference image" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 81, + 123, + 150 + ], + "lines": [ + { + "bbox": [ + 53, + 81, + 123, + 150 + ], + "spans": [ + { + "bbox": [ + 53, + 81, + 123, + 150 + ], + "type": "image", + "image_path": "90b4c6812a23bc64e0005d692fb5d98823a2a1d002f9f220cbd0b2c2e09b02ba.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 150, + 96, + 156 + ], + "lines": [ + { + "bbox": [ + 78, + 150, + 96, + 156 + ], + "spans": [ + { + "bbox": [ + 78, + 150, + 96, + 156 + ], + "type": "text", + "content": "A guitar" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 126, + 82, + 196, + 150 + ], + "blocks": [ + { + "bbox": [ + 154, + 72, + 199, + 80 + ], + "lines": [ + { + "bbox": [ + 154, + 72, + 199, + 80 + ], + "spans": [ + { + "bbox": [ + 154, + 72, + 199, + 80 + ], + "type": "text", + "content": "StyleAligned" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 126, + 82, + 196, + 150 + ], + "lines": [ + { + "bbox": [ + 126, + 82, + 196, + 150 + ], + "spans": [ + { + "bbox": [ + 126, + 82, + 196, + 150 + ], + "type": "image", + "image_path": "9182dc023ae4f02985381ae8250dd67f27a32d6b2304ba17eb545d23e84d78b2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 150, + 178, + 155 + ], + "lines": [ + { + "bbox": [ + 144, + 150, + 178, + 155 + ], + "spans": [ + { + "bbox": [ + 144, + 150, + 178, + 155 + ], + "type": "text", + "content": "A hot air balloon" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 196, + 82, + 228, + 150 + ], + "blocks": [ + { + "bbox": [ + 196, + 82, + 228, + 150 + ], + "lines": [ + { + "bbox": [ + 196, + 82, + 228, + 150 + ], + "spans": [ + { + "bbox": [ + 196, + 82, + 228, + 150 + ], + "type": "image", + "image_path": "5aca3abfb4d09a6d3e5db3cba5ca2de4f2cc89a6531e18621a8d0d39d20800c7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 150, + 225, + 155 + ], + "lines": [ + { + "bbox": [ + 200, + 150, + 225, + 155 + ], + "spans": [ + { + "bbox": [ + 200, + 150, + 225, + 155 + ], + "type": "text", + "content": "A mountain" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 230, + 81, + 299, + 150 + ], + "blocks": [ + { + "bbox": [ + 250, + 72, + 311, + 80 + ], + "lines": [ + { + "bbox": [ + 250, + 72, + 311, + 80 + ], + "spans": [ + { + "bbox": [ + 250, + 72, + 311, + 80 + ], + "type": "text", + "content": "StyleDrop (SDXL)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 230, + 81, + 299, + 150 + ], + "lines": [ + { + "bbox": [ + 230, + 81, + 299, + 150 + ], + "spans": [ + { + "bbox": [ + 230, + 81, + 299, + 150 + ], + "type": "image", + "image_path": "c7948ce73d00ded74b3d54ff0bbc607a0b02b28b6b1d46b61271a0f2a65985d0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 150, + 282, + 155 + ], + "lines": [ + { + "bbox": [ + 249, + 150, + 282, + 155 + ], + "spans": [ + { + "bbox": [ + 249, + 150, + 282, + 155 + ], + "type": "text", + "content": "A hot air balloon" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 81, + 332, + 150 + ], + "blocks": [ + { + "bbox": [ + 306, + 81, + 332, + 150 + ], + "lines": [ + { + "bbox": [ + 306, + 81, + 332, + 150 + ], + "spans": [ + { + "bbox": [ + 306, + 81, + 332, + 150 + ], + "type": "image", + "image_path": "a5994e42c5975622b839fe5edb1ae76d3ccbb6ab0107ba916a23dc3bad6ba32b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 150, + 328, + 155 + ], + "lines": [ + { + "bbox": [ + 304, + 150, + 328, + 155 + ], + "spans": [ + { + "bbox": [ + 304, + 150, + 328, + 155 + ], + "type": "text", + "content": "A mountain" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 335, + 82, + 403, + 150 + ], + "blocks": [ + { + "bbox": [ + 336, + 72, + 432, + 80 + ], + "lines": [ + { + "bbox": [ + 336, + 72, + 432, + 80 + ], + "spans": [ + { + "bbox": [ + 336, + 72, + 432, + 80 + ], + "type": "text", + "content": "StyleDrop (MUSE unofficial)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 335, + 82, + 403, + 150 + ], + "lines": [ + { + "bbox": [ + 335, + 82, + 403, + 150 + ], + "spans": [ + { + "bbox": [ + 335, + 82, + 403, + 150 + ], + "type": "image", + "image_path": "5a0e47a4805c3521b0a200742e4d7a6efc924c7cb5a56dac468920f576d2922a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 150, + 386, + 155 + ], + "lines": [ + { + "bbox": [ + 352, + 150, + 386, + 155 + ], + "spans": [ + { + "bbox": [ + 352, + 150, + 386, + 155 + ], + "type": "text", + "content": "A hot air balloon" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 403, + 82, + 436, + 150 + ], + "blocks": [ + { + "bbox": [ + 403, + 82, + 436, + 150 + ], + "lines": [ + { + "bbox": [ + 403, + 82, + 436, + 150 + ], + "spans": [ + { + "bbox": [ + 403, + 82, + 436, + 150 + ], + "type": "image", + "image_path": "46b6a33a75f26a0b999c095607dae2c6762844b5d675645eb58f683a47db58dc.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 150, + 433, + 155 + ], + "lines": [ + { + "bbox": [ + 408, + 150, + 433, + 155 + ], + "spans": [ + { + "bbox": [ + 408, + 150, + 433, + 155 + ], + "type": "text", + "content": "A mountain" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 438, + 82, + 507, + 150 + ], + "blocks": [ + { + "bbox": [ + 444, + 72, + 537, + 80 + ], + "lines": [ + { + "bbox": [ + 444, + 72, + 537, + 80 + ], + "spans": [ + { + "bbox": [ + 444, + 72, + 537, + 80 + ], + "type": "text", + "content": "Dreambooth-LoRA (SDXL)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 438, + 82, + 507, + 150 + ], + "lines": [ + { + "bbox": [ + 438, + 82, + 507, + 150 + ], + "spans": [ + { + "bbox": [ + 438, + 82, + 507, + 150 + ], + "type": "image", + "image_path": "cf7b7ed8e76a13b03bed044b822d2266a9b6f5900b9b0d0dd4546a02b814c5f7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 150, + 489, + 155 + ], + "lines": [ + { + "bbox": [ + 456, + 150, + 489, + 155 + ], + "spans": [ + { + "bbox": [ + 456, + 150, + 489, + 155 + ], + "type": "text", + "content": "A hot air balloon" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 509, + 82, + 541, + 150 + ], + "blocks": [ + { + "bbox": [ + 509, + 82, + 541, + 150 + ], + "lines": [ + { + "bbox": [ + 509, + 82, + 541, + 150 + ], + "spans": [ + { + "bbox": [ + 509, + 82, + 541, + 150 + ], + "type": "image", + "image_path": "5df865f298975953ec22baa0012e6361e197e23bf7faab5934792aaa96698c2c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 511, + 150, + 537, + 155 + ], + "lines": [ + { + "bbox": [ + 511, + 150, + 537, + 155 + ], + "spans": [ + { + "bbox": [ + 511, + 150, + 537, + 155 + ], + "type": "text", + "content": "A mountain" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 53, + 157, + 123, + 224 + ], + "blocks": [ + { + "bbox": [ + 53, + 157, + 123, + 224 + ], + "lines": [ + { + "bbox": [ + 53, + 157, + 123, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 157, + 123, + 224 + ], + "type": "image", + "image_path": "20e2aaaf0c1e6ab47123702fef3e6d7744bc6664e05e999232247491622d8124.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 225, + 116, + 231 + ], + "lines": [ + { + "bbox": [ + 59, + 225, + 116, + 231 + ], + "spans": [ + { + "bbox": [ + 59, + 225, + 116, + 231 + ], + "type": "text", + "content": "A woman walking her dogs" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 126, + 157, + 196, + 224 + ], + "blocks": [ + { + "bbox": [ + 126, + 157, + 196, + 224 + ], + "lines": [ + { + "bbox": [ + 126, + 157, + 196, + 224 + ], + "spans": [ + { + "bbox": [ + 126, + 157, + 196, + 224 + ], + "type": "image", + "image_path": "2900abc58cb353417dbcacb0858899b10ac88bc6db44cd9e37deb49575279ad4.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 225, + 178, + 231 + ], + "lines": [ + { + "bbox": [ + 145, + 225, + 178, + 231 + ], + "spans": [ + { + "bbox": [ + 145, + 225, + 178, + 231 + ], + "type": "text", + "content": "A friendly robot" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 196, + 225, + 227, + 231 + ], + "lines": [ + { + "bbox": [ + 196, + 225, + 227, + 231 + ], + "spans": [ + { + "bbox": [ + 196, + 225, + 227, + 231 + ], + "type": "text", + "content": "Cherryblossom" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 196, + 157, + 228, + 224 + ], + "blocks": [ + { + "bbox": [ + 196, + 157, + 228, + 224 + ], + "lines": [ + { + "bbox": [ + 196, + 157, + 228, + 224 + ], + "spans": [ + { + "bbox": [ + 196, + 157, + 228, + 224 + ], + "type": "image", + "image_path": "91f969d04a9a14bea047a577c124bb468e71870f92e46a74591a2accfd49efbf.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 230, + 157, + 299, + 224 + ], + "blocks": [ + { + "bbox": [ + 230, + 157, + 299, + 224 + ], + "lines": [ + { + "bbox": [ + 230, + 157, + 299, + 224 + ], + "spans": [ + { + "bbox": [ + 230, + 157, + 299, + 224 + ], + "type": "image", + "image_path": "934d4e99253578a4646ad35ef4a76dc408e1ce411b4fae9ab69fc49920db8ff8.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 225, + 282, + 231 + ], + "lines": [ + { + "bbox": [ + 249, + 225, + 282, + 231 + ], + "spans": [ + { + "bbox": [ + 249, + 225, + 282, + 231 + ], + "type": "text", + "content": "A friendly robot" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 300, + 157, + 332, + 224 + ], + "blocks": [ + { + "bbox": [ + 300, + 157, + 332, + 224 + ], + "lines": [ + { + "bbox": [ + 300, + 157, + 332, + 224 + ], + "spans": [ + { + "bbox": [ + 300, + 157, + 332, + 224 + ], + "type": "image", + "image_path": "cddaf90b91fa69e42670e5a911d7462cb4857ff150b046a76dfa7ec3e37516d5.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 225, + 332, + 231 + ], + "lines": [ + { + "bbox": [ + 301, + 225, + 332, + 231 + ], + "spans": [ + { + "bbox": [ + 301, + 225, + 332, + 231 + ], + "type": "text", + "content": "Cherryblossom" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 336, + 157, + 403, + 224 + ], + "blocks": [ + { + "bbox": [ + 336, + 157, + 403, + 224 + ], + "lines": [ + { + "bbox": [ + 336, + 157, + 403, + 224 + ], + "spans": [ + { + "bbox": [ + 336, + 157, + 403, + 224 + ], + "type": "image", + "image_path": "2ecc2accac53d7b2fc69a9b04e6409aabeb20377e722c3e746a29ad3763ec90b.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 225, + 377, + 231 + ], + "lines": [ + { + "bbox": [ + 361, + 225, + 377, + 231 + ], + "spans": [ + { + "bbox": [ + 361, + 225, + 377, + 231 + ], + "type": "text", + "content": "A robot" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 403, + 157, + 436, + 224 + ], + "blocks": [ + { + "bbox": [ + 403, + 157, + 436, + 224 + ], + "lines": [ + { + "bbox": [ + 403, + 157, + 436, + 224 + ], + "spans": [ + { + "bbox": [ + 403, + 157, + 436, + 224 + ], + "type": "image", + "image_path": "eb1f356d8d9dbb4832264617ad7181e028801061eb592887032ae1c4e14a004b.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 407, + 225, + 432, + 231 + ], + "lines": [ + { + "bbox": [ + 407, + 225, + 432, + 231 + ], + "spans": [ + { + "bbox": [ + 407, + 225, + 432, + 231 + ], + "type": "text", + "content": "Cherryblos" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 438, + 157, + 507, + 224 + ], + "blocks": [ + { + "bbox": [ + 438, + 157, + 507, + 224 + ], + "lines": [ + { + "bbox": [ + 438, + 157, + 507, + 224 + ], + "spans": [ + { + "bbox": [ + 438, + 157, + 507, + 224 + ], + "type": "image", + "image_path": "4a444f5f41709352b61c2f29db3900fe49fb1982d985bb2a52e3d55197d629c5.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 225, + 489, + 231 + ], + "lines": [ + { + "bbox": [ + 457, + 225, + 489, + 231 + ], + "spans": [ + { + "bbox": [ + 457, + 225, + 489, + 231 + ], + "type": "text", + "content": "A friendly robot" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 507, + 225, + 539, + 231 + ], + "lines": [ + { + "bbox": [ + 507, + 225, + 539, + 231 + ], + "spans": [ + { + "bbox": [ + 507, + 225, + 539, + 231 + ], + "type": "text", + "content": "Cherryblossom" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 507, + 157, + 541, + 224 + ], + "blocks": [ + { + "bbox": [ + 507, + 157, + 541, + 224 + ], + "lines": [ + { + "bbox": [ + 507, + 157, + 541, + 224 + ], + "spans": [ + { + "bbox": [ + 507, + 157, + 541, + 224 + ], + "type": "image", + "image_path": "610c9ba3314ca1226805491b5a9ce683756ab16a5b5f16aeebefaf025038cf13.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 53, + 233, + 123, + 300 + ], + "blocks": [ + { + "bbox": [ + 53, + 233, + 123, + 300 + ], + "lines": [ + { + "bbox": [ + 53, + 233, + 123, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 233, + 123, + 300 + ], + "type": "image", + "image_path": "8840c5026e15cd714d61ba2d56d55ab5f9290a381cb36c9d92013488ec9d9cd7.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 300, + 96, + 305 + ], + "lines": [ + { + "bbox": [ + 80, + 300, + 96, + 305 + ], + "spans": [ + { + "bbox": [ + 80, + 300, + 96, + 305 + ], + "type": "text", + "content": "A duck" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 126, + 233, + 196, + 300 + ], + "blocks": [ + { + "bbox": [ + 126, + 233, + 196, + 300 + ], + "lines": [ + { + "bbox": [ + 126, + 233, + 196, + 300 + ], + "spans": [ + { + "bbox": [ + 126, + 233, + 196, + 300 + ], + "type": "image", + "image_path": "27d3adda1e9c1e8abb94cd5d3919471f7d4377e42f229c39b65589ea91c646c5.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 300, + 168, + 306 + ], + "lines": [ + { + "bbox": [ + 155, + 300, + 168, + 306 + ], + "spans": [ + { + "bbox": [ + 155, + 300, + 168, + 306 + ], + "type": "text", + "content": "A ship" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 196, + 233, + 228, + 300 + ], + "blocks": [ + { + "bbox": [ + 196, + 233, + 228, + 300 + ], + "lines": [ + { + "bbox": [ + 196, + 233, + 228, + 300 + ], + "spans": [ + { + "bbox": [ + 196, + 233, + 228, + 300 + ], + "type": "image", + "image_path": "5a84f2c0a5fbafa354a5977511fe2a587e3984e96500383873d23ee76b6ca112.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 300, + 221, + 306 + ], + "lines": [ + { + "bbox": [ + 203, + 300, + 221, + 306 + ], + "spans": [ + { + "bbox": [ + 203, + 300, + 221, + 306 + ], + "type": "text", + "content": "A rocket" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 230, + 233, + 299, + 300 + ], + "blocks": [ + { + "bbox": [ + 230, + 233, + 299, + 300 + ], + "lines": [ + { + "bbox": [ + 230, + 233, + 299, + 300 + ], + "spans": [ + { + "bbox": [ + 230, + 233, + 299, + 300 + ], + "type": "image", + "image_path": "7f110362dbb4193ed2c122926456ca32cf0ebff2b338a2f79b2ba1a19ea9abe1.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 300, + 272, + 306 + ], + "lines": [ + { + "bbox": [ + 258, + 300, + 272, + 306 + ], + "spans": [ + { + "bbox": [ + 258, + 300, + 272, + 306 + ], + "type": "text", + "content": "A ship" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 300, + 233, + 332, + 300 + ], + "blocks": [ + { + "bbox": [ + 300, + 233, + 332, + 300 + ], + "lines": [ + { + "bbox": [ + 300, + 233, + 332, + 300 + ], + "spans": [ + { + "bbox": [ + 300, + 233, + 332, + 300 + ], + "type": "image", + "image_path": "e719c9a00052dcead3a2edfa21f5306cf25728b1e523e32e19639193911e3f21.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 300, + 326, + 306 + ], + "lines": [ + { + "bbox": [ + 307, + 300, + 326, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 300, + 326, + 306 + ], + "type": "text", + "content": "A rocket" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 336, + 233, + 403, + 300 + ], + "blocks": [ + { + "bbox": [ + 336, + 233, + 403, + 300 + ], + "lines": [ + { + "bbox": [ + 336, + 233, + 403, + 300 + ], + "spans": [ + { + "bbox": [ + 336, + 233, + 403, + 300 + ], + "type": "image", + "image_path": "520144230c3f6cda53ff73278bfc318dc00a4a7b4f25acc8f6aad269a22713e0.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 300, + 376, + 306 + ], + "lines": [ + { + "bbox": [ + 362, + 300, + 376, + 306 + ], + "spans": [ + { + "bbox": [ + 362, + 300, + 376, + 306 + ], + "type": "text", + "content": "A ship" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 403, + 233, + 436, + 300 + ], + "blocks": [ + { + "bbox": [ + 403, + 233, + 436, + 300 + ], + "lines": [ + { + "bbox": [ + 403, + 233, + 436, + 300 + ], + "spans": [ + { + "bbox": [ + 403, + 233, + 436, + 300 + ], + "type": "image", + "image_path": "f30b62b042d2e1a7bc31ea8397e082c8f27c08c36af31b9f8e439353ddb0b106.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 300, + 427, + 306 + ], + "lines": [ + { + "bbox": [ + 414, + 300, + 427, + 306 + ], + "spans": [ + { + "bbox": [ + 414, + 300, + 427, + 306 + ], + "type": "text", + "content": "A car" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 438, + 233, + 507, + 300 + ], + "blocks": [ + { + "bbox": [ + 438, + 233, + 507, + 300 + ], + "lines": [ + { + "bbox": [ + 438, + 233, + 507, + 300 + ], + "spans": [ + { + "bbox": [ + 438, + 233, + 507, + 300 + ], + "type": "image", + "image_path": "c120d4a94419eaaf18f09eb87f57ff0e51bc783cbfc3b7df039a20de156ed2fa.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 300, + 480, + 306 + ], + "lines": [ + { + "bbox": [ + 466, + 300, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 466, + 300, + 480, + 306 + ], + "type": "text", + "content": "A ship" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 507, + 233, + 541, + 300 + ], + "blocks": [ + { + "bbox": [ + 507, + 233, + 541, + 300 + ], + "lines": [ + { + "bbox": [ + 507, + 233, + 541, + 300 + ], + "spans": [ + { + "bbox": [ + 507, + 233, + 541, + 300 + ], + "type": "image", + "image_path": "9d94bbdfcfdd7eeb1b3052b31dc2b9ffcd4e5c7236d8b171d9ae33c032af0cff.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 514, + 300, + 533, + 306 + ], + "lines": [ + { + "bbox": [ + 514, + 300, + 533, + 306 + ], + "spans": [ + { + "bbox": [ + 514, + 300, + 533, + 306 + ], + "type": "text", + "content": "A rocket" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 53, + 307, + 123, + 376 + ], + "blocks": [ + { + "bbox": [ + 53, + 307, + 123, + 376 + ], + "lines": [ + { + "bbox": [ + 53, + 307, + 123, + 376 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 123, + 376 + ], + "type": "image", + "image_path": "d7ddbb290dfd3012f5aa02c56956acb097eef5b8f80243a3deaff12c12b286c4.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 377, + 100, + 382 + ], + "lines": [ + { + "bbox": [ + 76, + 377, + 100, + 382 + ], + "spans": [ + { + "bbox": [ + 76, + 377, + 100, + 382 + ], + "type": "text", + "content": "A cute bear" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 126, + 308, + 196, + 376 + ], + "blocks": [ + { + "bbox": [ + 126, + 308, + 196, + 376 + ], + "lines": [ + { + "bbox": [ + 126, + 308, + 196, + 376 + ], + "spans": [ + { + "bbox": [ + 126, + 308, + 196, + 376 + ], + "type": "image", + "image_path": "eb8f5046cf9749c732d4fd406a4560c5b3b793e5ae5ad161bdccdc6d44f34b1b.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 377, + 171, + 382 + ], + "lines": [ + { + "bbox": [ + 151, + 377, + 171, + 382 + ], + "spans": [ + { + "bbox": [ + 151, + 377, + 171, + 382 + ], + "type": "text", + "content": "A moose" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_caption" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 196, + 308, + 228, + 376 + ], + "blocks": [ + { + "bbox": [ + 196, + 308, + 228, + 376 + ], + "lines": [ + { + "bbox": [ + 196, + 308, + 228, + 376 + ], + "spans": [ + { + "bbox": [ + 196, + 308, + 228, + 376 + ], + "type": "image", + "image_path": "101e1170f7064b5edc9f04d96143713764a19191953362df139e3913fa60e039.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 377, + 225, + 382 + ], + "lines": [ + { + "bbox": [ + 199, + 377, + 225, + 382 + ], + "spans": [ + { + "bbox": [ + 199, + 377, + 225, + 382 + ], + "type": "text", + "content": "A baby pen" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 230, + 309, + 299, + 376 + ], + "blocks": [ + { + "bbox": [ + 230, + 309, + 299, + 376 + ], + "lines": [ + { + "bbox": [ + 230, + 309, + 299, + 376 + ], + "spans": [ + { + "bbox": [ + 230, + 309, + 299, + 376 + ], + "type": "image", + "image_path": "0f3d2fa0c62c9e0e5edf09ed204fd59daf986792ce49f36c091cb0568c688e05.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 377, + 275, + 382 + ], + "lines": [ + { + "bbox": [ + 255, + 377, + 275, + 382 + ], + "spans": [ + { + "bbox": [ + 255, + 377, + 275, + 382 + ], + "type": "text", + "content": "A moose" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 377, + 331, + 382 + ], + "lines": [ + { + "bbox": [ + 299, + 377, + 331, + 382 + ], + "spans": [ + { + "bbox": [ + 299, + 377, + 331, + 382 + ], + "type": "text", + "content": "A baby penguin" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_caption" + } + ], + "index": 65 + }, + { + "type": "image", + "bbox": [ + 300, + 309, + 332, + 376 + ], + "blocks": [ + { + "bbox": [ + 300, + 309, + 332, + 376 + ], + "lines": [ + { + "bbox": [ + 300, + 309, + 332, + 376 + ], + "spans": [ + { + "bbox": [ + 300, + 309, + 332, + 376 + ], + "type": "image", + "image_path": "1e88d7841435824dd9ec845f5d2d1d4131c53799ec14308bbce9203f29322916.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + } + ], + "index": 67 + }, + { + "type": "image", + "bbox": [ + 336, + 309, + 403, + 376 + ], + "blocks": [ + { + "bbox": [ + 336, + 309, + 403, + 376 + ], + "lines": [ + { + "bbox": [ + 336, + 309, + 403, + 376 + ], + "spans": [ + { + "bbox": [ + 336, + 309, + 403, + 376 + ], + "type": "image", + "image_path": "066e32b3d3e53446c8e21b056da640992280644c433de7ec203d5b58ab4d1c94.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 377, + 378, + 382 + ], + "lines": [ + { + "bbox": [ + 359, + 377, + 378, + 382 + ], + "spans": [ + { + "bbox": [ + 359, + 377, + 378, + 382 + ], + "type": "text", + "content": "A moose" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 403, + 377, + 436, + 382 + ], + "lines": [ + { + "bbox": [ + 403, + 377, + 436, + 382 + ], + "spans": [ + { + "bbox": [ + 403, + 377, + 436, + 382 + ], + "type": "text", + "content": "A baby penguin" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_caption" + } + ], + "index": 69 + }, + { + "type": "image", + "bbox": [ + 403, + 309, + 436, + 376 + ], + "blocks": [ + { + "bbox": [ + 403, + 309, + 436, + 376 + ], + "lines": [ + { + "bbox": [ + 403, + 309, + 436, + 376 + ], + "spans": [ + { + "bbox": [ + 403, + 309, + 436, + 376 + ], + "type": "image", + "image_path": "d493c3f5b38c916496d273246a57c83b152ee12941106b89e07e94b18f98934f.jpg" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_body" + } + ], + "index": 71 + }, + { + "type": "image", + "bbox": [ + 438, + 309, + 507, + 376 + ], + "blocks": [ + { + "bbox": [ + 438, + 309, + 507, + 376 + ], + "lines": [ + { + "bbox": [ + 438, + 309, + 507, + 376 + ], + "spans": [ + { + "bbox": [ + 438, + 309, + 507, + 376 + ], + "type": "image", + "image_path": "98bd1600692848c1d1e32c87ffbbaad7e8831f1325c1d591c3d4836fded90ab6.jpg" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 377, + 482, + 382 + ], + "lines": [ + { + "bbox": [ + 463, + 377, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 463, + 377, + 482, + 382 + ], + "type": "text", + "content": "A moose" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 507, + 377, + 540, + 382 + ], + "lines": [ + { + "bbox": [ + 507, + 377, + 540, + 382 + ], + "spans": [ + { + "bbox": [ + 507, + 377, + 540, + 382 + ], + "type": "text", + "content": "A baby penguin" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_caption" + } + ], + "index": 73 + }, + { + "type": "image", + "bbox": [ + 507, + 309, + 541, + 376 + ], + "blocks": [ + { + "bbox": [ + 507, + 309, + 541, + 376 + ], + "lines": [ + { + "bbox": [ + 507, + 309, + 541, + 376 + ], + "spans": [ + { + "bbox": [ + 507, + 309, + 541, + 376 + ], + "type": "image", + "image_path": "07dab4a6b25a0ac834d7e5208ec8f5ee690714b8ca83182f7753e0c5f5c91b88.jpg" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_body" + } + ], + "index": 75 + }, + { + "type": "image", + "bbox": [ + 53, + 384, + 123, + 452 + ], + "blocks": [ + { + "bbox": [ + 53, + 384, + 123, + 452 + ], + "lines": [ + { + "bbox": [ + 53, + 384, + 123, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 123, + 452 + ], + "type": "image", + "image_path": "0e75922bafab63fd0dc43b2a236d62036d8df5d8480aa1a4c0b98594e1613c20.jpg" + } + ] + } + ], + "index": 77, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 453, + 98, + 457 + ], + "lines": [ + { + "bbox": [ + 77, + 453, + 98, + 457 + ], + "spans": [ + { + "bbox": [ + 77, + 453, + 98, + 457 + ], + "type": "text", + "content": "A camera" + } + ] + } + ], + "index": 78, + "angle": 0, + "type": "image_caption" + } + ], + "index": 77 + }, + { + "type": "image", + "bbox": [ + 126, + 384, + 196, + 452 + ], + "blocks": [ + { + "bbox": [ + 126, + 384, + 196, + 452 + ], + "lines": [ + { + "bbox": [ + 126, + 384, + 196, + 452 + ], + "spans": [ + { + "bbox": [ + 126, + 384, + 196, + 452 + ], + "type": "image", + "image_path": "fde80ce0ea8f158811b6028048d98b0700eaf3363d4247ad20510ff14bdce6ce.jpg" + } + ] + } + ], + "index": 79, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 453, + 171, + 457 + ], + "lines": [ + { + "bbox": [ + 151, + 453, + 171, + 457 + ], + "spans": [ + { + "bbox": [ + 151, + 453, + 171, + 457 + ], + "type": "text", + "content": "A fireman" + } + ] + } + ], + "index": 80, + "angle": 0, + "type": "image_caption" + } + ], + "index": 79 + }, + { + "type": "image", + "bbox": [ + 196, + 384, + 228, + 452 + ], + "blocks": [ + { + "bbox": [ + 196, + 384, + 228, + 452 + ], + "lines": [ + { + "bbox": [ + 196, + 384, + 228, + 452 + ], + "spans": [ + { + "bbox": [ + 196, + 384, + 228, + 452 + ], + "type": "image", + "image_path": "3613566046215aec7374cc7ae12d282ad8e812100a330aa8355d1ff5806af360.jpg" + } + ] + } + ], + "index": 81, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 453, + 221, + 457 + ], + "lines": [ + { + "bbox": [ + 203, + 453, + 221, + 457 + ], + "spans": [ + { + "bbox": [ + 203, + 453, + 221, + 457 + ], + "type": "text", + "content": "A tall hill" + } + ] + } + ], + "index": 82, + "angle": 0, + "type": "image_caption" + } + ], + "index": 81 + }, + { + "type": "image", + "bbox": [ + 230, + 384, + 299, + 452 + ], + "blocks": [ + { + "bbox": [ + 230, + 384, + 299, + 452 + ], + "lines": [ + { + "bbox": [ + 230, + 384, + 299, + 452 + ], + "spans": [ + { + "bbox": [ + 230, + 384, + 299, + 452 + ], + "type": "image", + "image_path": "687587c656d7c57e108c490ea6e0f9a58ca98ec07e812b9a7d6dd6c239aad61e.jpg" + } + ] + } + ], + "index": 83, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 453, + 276, + 457 + ], + "lines": [ + { + "bbox": [ + 255, + 453, + 276, + 457 + ], + "spans": [ + { + "bbox": [ + 255, + 453, + 276, + 457 + ], + "type": "text", + "content": "A fireman" + } + ] + } + ], + "index": 84, + "angle": 0, + "type": "image_caption" + } + ], + "index": 83 + }, + { + "type": "image", + "bbox": [ + 300, + 384, + 332, + 452 + ], + "blocks": [ + { + "bbox": [ + 300, + 384, + 332, + 452 + ], + "lines": [ + { + "bbox": [ + 300, + 384, + 332, + 452 + ], + "spans": [ + { + "bbox": [ + 300, + 384, + 332, + 452 + ], + "type": "image", + "image_path": "5f858526fbb5f99a5735067d56df198d3cd99d2fc790db7840d84b2fe9da5d37.jpg" + } + ] + } + ], + "index": 85, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 453, + 325, + 457 + ], + "lines": [ + { + "bbox": [ + 307, + 453, + 325, + 457 + ], + "spans": [ + { + "bbox": [ + 307, + 453, + 325, + 457 + ], + "type": "text", + "content": "A tall hill" + } + ] + } + ], + "index": 86, + "angle": 0, + "type": "image_caption" + } + ], + "index": 85 + }, + { + "type": "image", + "bbox": [ + 336, + 384, + 403, + 452 + ], + "blocks": [ + { + "bbox": [ + 336, + 384, + 403, + 452 + ], + "lines": [ + { + "bbox": [ + 336, + 384, + 403, + 452 + ], + "spans": [ + { + "bbox": [ + 336, + 384, + 403, + 452 + ], + "type": "image", + "image_path": "a5ecc311b5dbeb6b2baedaef213a0d6b54334591474cb728be06ab329d4edeef.jpg" + } + ] + } + ], + "index": 87, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 453, + 380, + 457 + ], + "lines": [ + { + "bbox": [ + 358, + 453, + 380, + 457 + ], + "spans": [ + { + "bbox": [ + 358, + 453, + 380, + 457 + ], + "type": "text", + "content": "A fireman" + } + ] + } + ], + "index": 88, + "angle": 0, + "type": "image_caption" + } + ], + "index": 87 + }, + { + "type": "image", + "bbox": [ + 403, + 384, + 436, + 452 + ], + "blocks": [ + { + "bbox": [ + 403, + 384, + 436, + 452 + ], + "lines": [ + { + "bbox": [ + 403, + 384, + 436, + 452 + ], + "spans": [ + { + "bbox": [ + 403, + 384, + 436, + 452 + ], + "type": "image", + "image_path": "5f0710c2366220dd26304908de510520344952d40686136117fcf3bf7c907660.jpg" + } + ] + } + ], + "index": 89, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 453, + 429, + 457 + ], + "lines": [ + { + "bbox": [ + 411, + 453, + 429, + 457 + ], + "spans": [ + { + "bbox": [ + 411, + 453, + 429, + 457 + ], + "type": "text", + "content": "A tall hill" + } + ] + } + ], + "index": 90, + "angle": 0, + "type": "image_caption" + } + ], + "index": 89 + }, + { + "type": "image", + "bbox": [ + 438, + 384, + 507, + 452 + ], + "blocks": [ + { + "bbox": [ + 438, + 384, + 507, + 452 + ], + "lines": [ + { + "bbox": [ + 438, + 384, + 507, + 452 + ], + "spans": [ + { + "bbox": [ + 438, + 384, + 507, + 452 + ], + "type": "image", + "image_path": "de633264efa8f257c089060d149edbe67c2102974cb50c9648219b00e7729e36.jpg" + } + ] + } + ], + "index": 91, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 453, + 484, + 457 + ], + "lines": [ + { + "bbox": [ + 462, + 453, + 484, + 457 + ], + "spans": [ + { + "bbox": [ + 462, + 453, + 484, + 457 + ], + "type": "text", + "content": "A fireman" + } + ] + } + ], + "index": 92, + "angle": 0, + "type": "image_caption" + } + ], + "index": 91 + }, + { + "type": "image", + "bbox": [ + 507, + 384, + 541, + 452 + ], + "blocks": [ + { + "bbox": [ + 507, + 384, + 541, + 452 + ], + "lines": [ + { + "bbox": [ + 507, + 384, + 541, + 452 + ], + "spans": [ + { + "bbox": [ + 507, + 384, + 541, + 452 + ], + "type": "image", + "image_path": "848ac7555642c0cea82a2611ee7dab8c5f40775618c28cab35f1d537b3a2204a.jpg" + } + ] + } + ], + "index": 93, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 514, + 453, + 533, + 457 + ], + "lines": [ + { + "bbox": [ + 514, + 453, + 533, + 457 + ], + "spans": [ + { + "bbox": [ + 514, + 453, + 533, + 457 + ], + "type": "text", + "content": "A tall hill" + } + ] + } + ], + "index": 94, + "angle": 0, + "type": "image_caption" + } + ], + "index": 93 + }, + { + "type": "image", + "bbox": [ + 53, + 458, + 123, + 527 + ], + "blocks": [ + { + "bbox": [ + 53, + 458, + 123, + 527 + ], + "lines": [ + { + "bbox": [ + 53, + 458, + 123, + 527 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 123, + 527 + ], + "type": "image", + "image_path": "218c906898b3a5ae3af8a0c59cbfe3f2e6a876f559a6ed9dbed5f74cff59aca5.jpg" + } + ] + } + ], + "index": 95, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 528, + 102, + 533 + ], + "lines": [ + { + "bbox": [ + 74, + 528, + 102, + 533 + ], + "spans": [ + { + "bbox": [ + 74, + 528, + 102, + 533 + ], + "type": "text", + "content": "A saxophone" + } + ] + } + ], + "index": 96, + "angle": 0, + "type": "image_caption" + } + ], + "index": 95 + }, + { + "type": "image", + "bbox": [ + 126, + 458, + 196, + 527 + ], + "blocks": [ + { + "bbox": [ + 126, + 458, + 196, + 527 + ], + "lines": [ + { + "bbox": [ + 126, + 458, + 196, + 527 + ], + "spans": [ + { + "bbox": [ + 126, + 458, + 196, + 527 + ], + "type": "image", + "image_path": "5a326baa1c71afd4fa92e7b945445d5df1a917fb68392dc96bde8df9060c0670.jpg" + } + ] + } + ], + "index": 97, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 528, + 173, + 533 + ], + "lines": [ + { + "bbox": [ + 148, + 528, + 173, + 533 + ], + "spans": [ + { + "bbox": [ + 148, + 528, + 173, + 533 + ], + "type": "text", + "content": "A sunflower" + } + ] + } + ], + "index": 98, + "angle": 0, + "type": "image_caption" + } + ], + "index": 97 + }, + { + "type": "image", + "bbox": [ + 196, + 458, + 228, + 527 + ], + "blocks": [ + { + "bbox": [ + 196, + 458, + 228, + 527 + ], + "lines": [ + { + "bbox": [ + 196, + 458, + 228, + 527 + ], + "spans": [ + { + "bbox": [ + 196, + 458, + 228, + 527 + ], + "type": "image", + "image_path": "2e24b65414432bef1e741090cb6d689be11819594a2cdbe6f42fe4398c0cb96f.jpg" + } + ] + } + ], + "index": 99, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 528, + 225, + 533 + ], + "lines": [ + { + "bbox": [ + 200, + 528, + 225, + 533 + ], + "spans": [ + { + "bbox": [ + 200, + 528, + 225, + 533 + ], + "type": "text", + "content": "A compass" + } + ] + } + ], + "index": 100, + "angle": 0, + "type": "image_caption" + } + ], + "index": 99 + }, + { + "type": "image", + "bbox": [ + 230, + 458, + 299, + 527 + ], + "blocks": [ + { + "bbox": [ + 230, + 458, + 299, + 527 + ], + "lines": [ + { + "bbox": [ + 230, + 458, + 299, + 527 + ], + "spans": [ + { + "bbox": [ + 230, + 458, + 299, + 527 + ], + "type": "image", + "image_path": "3bd7a2838367c664a9e337c39d2c1f17bc4b4d674a887d56572386b522ef23a5.jpg" + } + ] + } + ], + "index": 101, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 528, + 277, + 533 + ], + "lines": [ + { + "bbox": [ + 253, + 528, + 277, + 533 + ], + "spans": [ + { + "bbox": [ + 253, + 528, + 277, + 533 + ], + "type": "text", + "content": "A sunflower" + } + ] + } + ], + "index": 102, + "angle": 0, + "type": "image_caption" + } + ], + "index": 101 + }, + { + "type": "image", + "bbox": [ + 300, + 458, + 332, + 527 + ], + "blocks": [ + { + "bbox": [ + 300, + 458, + 332, + 527 + ], + "lines": [ + { + "bbox": [ + 300, + 458, + 332, + 527 + ], + "spans": [ + { + "bbox": [ + 300, + 458, + 332, + 527 + ], + "type": "image", + "image_path": "a909ea9a7f355a26af4d2503b0d77e6884a4592309406b7b8a2b5cbf9376bc47.jpg" + } + ] + } + ], + "index": 103, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 528, + 328, + 533 + ], + "lines": [ + { + "bbox": [ + 305, + 528, + 328, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 528, + 328, + 533 + ], + "type": "text", + "content": "A compass" + } + ] + } + ], + "index": 104, + "angle": 0, + "type": "image_caption" + } + ], + "index": 103 + }, + { + "type": "image", + "bbox": [ + 336, + 458, + 403, + 527 + ], + "blocks": [ + { + "bbox": [ + 336, + 458, + 403, + 527 + ], + "lines": [ + { + "bbox": [ + 336, + 458, + 403, + 527 + ], + "spans": [ + { + "bbox": [ + 336, + 458, + 403, + 527 + ], + "type": "image", + "image_path": "245f0efac9754680077bce0f2084c5a09d0700a318b35ddbc7df9e82dd856f22.jpg" + } + ] + } + ], + "index": 105, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 528, + 381, + 533 + ], + "lines": [ + { + "bbox": [ + 356, + 528, + 381, + 533 + ], + "spans": [ + { + "bbox": [ + 356, + 528, + 381, + 533 + ], + "type": "text", + "content": "A sunflower" + } + ] + } + ], + "index": 106, + "angle": 0, + "type": "image_caption" + } + ], + "index": 105 + }, + { + "type": "image", + "bbox": [ + 403, + 458, + 436, + 527 + ], + "blocks": [ + { + "bbox": [ + 403, + 458, + 436, + 527 + ], + "lines": [ + { + "bbox": [ + 403, + 458, + 436, + 527 + ], + "spans": [ + { + "bbox": [ + 403, + 458, + 436, + 527 + ], + "type": "image", + "image_path": "1b207c7f42fabcce392a07367ce06138d3678cf13f55abd0b1abdda1ce539d27.jpg" + } + ] + } + ], + "index": 107, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 528, + 432, + 533 + ], + "lines": [ + { + "bbox": [ + 408, + 528, + 432, + 533 + ], + "spans": [ + { + "bbox": [ + 408, + 528, + 432, + 533 + ], + "type": "text", + "content": "A compass" + } + ] + } + ], + "index": 108, + "angle": 0, + "type": "image_caption" + } + ], + "index": 107 + }, + { + "type": "image", + "bbox": [ + 438, + 458, + 507, + 527 + ], + "blocks": [ + { + "bbox": [ + 438, + 458, + 507, + 527 + ], + "lines": [ + { + "bbox": [ + 438, + 458, + 507, + 527 + ], + "spans": [ + { + "bbox": [ + 438, + 458, + 507, + 527 + ], + "type": "image", + "image_path": "0e392ae6870353c15e4c33cb5e500f566f4a5367ee5622f56e67fdfb217bce2f.jpg" + } + ] + } + ], + "index": 109, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 528, + 485, + 533 + ], + "lines": [ + { + "bbox": [ + 460, + 528, + 485, + 533 + ], + "spans": [ + { + "bbox": [ + 460, + 528, + 485, + 533 + ], + "type": "text", + "content": "A sunflower" + } + ] + } + ], + "index": 110, + "angle": 0, + "type": "image_caption" + } + ], + "index": 109 + }, + { + "type": "image", + "bbox": [ + 507, + 458, + 541, + 527 + ], + "blocks": [ + { + "bbox": [ + 507, + 458, + 541, + 527 + ], + "lines": [ + { + "bbox": [ + 507, + 458, + 541, + 527 + ], + "spans": [ + { + "bbox": [ + 507, + 458, + 541, + 527 + ], + "type": "image", + "image_path": "2a20746fe4b8fc5b42a361f53ef12c4712aebf9df993f31740b2c8624cbe67e5.jpg" + } + ] + } + ], + "index": 111, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 528, + 535, + 533 + ], + "lines": [ + { + "bbox": [ + 512, + 528, + 535, + 533 + ], + "spans": [ + { + "bbox": [ + 512, + 528, + 535, + 533 + ], + "type": "text", + "content": "A compass" + } + ] + } + ], + "index": 112, + "angle": 0, + "type": "image_caption" + } + ], + "index": 111 + }, + { + "type": "image", + "bbox": [ + 53, + 535, + 123, + 602 + ], + "blocks": [ + { + "bbox": [ + 53, + 535, + 123, + 602 + ], + "lines": [ + { + "bbox": [ + 53, + 535, + 123, + 602 + ], + "spans": [ + { + "bbox": [ + 53, + 535, + 123, + 602 + ], + "type": "image", + "image_path": "b266b1b05456e22c05ece494dafbad1df540c9dd01234ce23f0846dcfdab7187.jpg" + } + ] + } + ], + "index": 113, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 603, + 104, + 608 + ], + "lines": [ + { + "bbox": [ + 71, + 603, + 104, + 608 + ], + "spans": [ + { + "bbox": [ + 71, + 603, + 104, + 608 + ], + "type": "text", + "content": "A bottle of wine" + } + ] + } + ], + "index": 114, + "angle": 0, + "type": "image_caption" + } + ], + "index": 113 + }, + { + "type": "image", + "bbox": [ + 126, + 535, + 196, + 602 + ], + "blocks": [ + { + "bbox": [ + 126, + 535, + 196, + 602 + ], + "lines": [ + { + "bbox": [ + 126, + 535, + 196, + 602 + ], + "spans": [ + { + "bbox": [ + 126, + 535, + 196, + 602 + ], + "type": "image", + "image_path": "0cc1655eb494e44e0a04cd133ad64293c8f42e2d1a50b974f0592bdb3ef10b6a.jpg" + } + ] + } + ], + "index": 115, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 603, + 169, + 608 + ], + "lines": [ + { + "bbox": [ + 153, + 603, + 169, + 608 + ], + "spans": [ + { + "bbox": [ + 153, + 603, + 169, + 608 + ], + "type": "text", + "content": "Scones" + } + ] + } + ], + "index": 116, + "angle": 0, + "type": "image_caption" + } + ], + "index": 115 + }, + { + "type": "image", + "bbox": [ + 196, + 535, + 228, + 602 + ], + "blocks": [ + { + "bbox": [ + 196, + 535, + 228, + 602 + ], + "lines": [ + { + "bbox": [ + 196, + 535, + 228, + 602 + ], + "spans": [ + { + "bbox": [ + 196, + 535, + 228, + 602 + ], + "type": "image", + "image_path": "c0febaa1d6f4af1489962cdf2c022e226c99f50d6390739b462f03a7e5894d83.jpg" + } + ] + } + ], + "index": 117, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 603, + 221, + 608 + ], + "lines": [ + { + "bbox": [ + 203, + 603, + 221, + 608 + ], + "spans": [ + { + "bbox": [ + 203, + 603, + 221, + 608 + ], + "type": "text", + "content": "A muffin" + } + ] + } + ], + "index": 118, + "angle": 0, + "type": "image_caption" + } + ], + "index": 117 + }, + { + "type": "image", + "bbox": [ + 230, + 535, + 299, + 602 + ], + "blocks": [ + { + "bbox": [ + 230, + 535, + 299, + 602 + ], + "lines": [ + { + "bbox": [ + 230, + 535, + 299, + 602 + ], + "spans": [ + { + "bbox": [ + 230, + 535, + 299, + 602 + ], + "type": "image", + "image_path": "a67f1819a5ac9832c842ffec3c5c2bea10ef041df432dc1197e1a531b7a6f866.jpg" + } + ] + } + ], + "index": 119, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 603, + 273, + 608 + ], + "lines": [ + { + "bbox": [ + 257, + 603, + 273, + 608 + ], + "spans": [ + { + "bbox": [ + 257, + 603, + 273, + 608 + ], + "type": "text", + "content": "Scones" + } + ] + } + ], + "index": 120, + "angle": 0, + "type": "image_caption" + } + ], + "index": 119 + }, + { + "type": "image", + "bbox": [ + 300, + 535, + 332, + 602 + ], + "blocks": [ + { + "bbox": [ + 300, + 535, + 332, + 602 + ], + "lines": [ + { + "bbox": [ + 300, + 535, + 332, + 602 + ], + "spans": [ + { + "bbox": [ + 300, + 535, + 332, + 602 + ], + "type": "image", + "image_path": "273b7d99f4027b80fb94d2a712e1526dccae55b952daf95ebaf79781277e4638.jpg" + } + ] + } + ], + "index": 121, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 603, + 326, + 608 + ], + "lines": [ + { + "bbox": [ + 307, + 603, + 326, + 608 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 326, + 608 + ], + "type": "text", + "content": "A muffin" + } + ] + } + ], + "index": 122, + "angle": 0, + "type": "image_caption" + } + ], + "index": 121 + }, + { + "type": "image", + "bbox": [ + 336, + 535, + 403, + 602 + ], + "blocks": [ + { + "bbox": [ + 336, + 535, + 403, + 602 + ], + "lines": [ + { + "bbox": [ + 336, + 535, + 403, + 602 + ], + "spans": [ + { + "bbox": [ + 336, + 535, + 403, + 602 + ], + "type": "image", + "image_path": "9729ac558578c2cc718793e23350acb7c4e090b725969a2dc8c12431c4a194f0.jpg" + } + ] + } + ], + "index": 123, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 603, + 380, + 608 + ], + "lines": [ + { + "bbox": [ + 359, + 603, + 380, + 608 + ], + "spans": [ + { + "bbox": [ + 359, + 603, + 380, + 608 + ], + "type": "text", + "content": "A fireman" + } + ] + } + ], + "index": 124, + "angle": 0, + "type": "image_caption" + } + ], + "index": 123 + }, + { + "type": "image", + "bbox": [ + 403, + 535, + 436, + 602 + ], + "blocks": [ + { + "bbox": [ + 403, + 535, + 436, + 602 + ], + "lines": [ + { + "bbox": [ + 403, + 535, + 436, + 602 + ], + "spans": [ + { + "bbox": [ + 403, + 535, + 436, + 602 + ], + "type": "image", + "image_path": "8fd9cff8bb4ccfef3049b2a2e4eb07b940bcfdadfe156730a753ba68f0cf99eb.jpg" + } + ] + } + ], + "index": 125, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 603, + 429, + 608 + ], + "lines": [ + { + "bbox": [ + 410, + 603, + 429, + 608 + ], + "spans": [ + { + "bbox": [ + 410, + 603, + 429, + 608 + ], + "type": "text", + "content": "A muffin" + } + ] + } + ], + "index": 126, + "angle": 0, + "type": "image_caption" + } + ], + "index": 125 + }, + { + "type": "image", + "bbox": [ + 438, + 535, + 507, + 602 + ], + "blocks": [ + { + "bbox": [ + 438, + 535, + 507, + 602 + ], + "lines": [ + { + "bbox": [ + 438, + 535, + 507, + 602 + ], + "spans": [ + { + "bbox": [ + 438, + 535, + 507, + 602 + ], + "type": "image", + "image_path": "3bc108ed9e3f17a44c0e1c114055d5844ede0e28a61f0b1c4879b10da89fb618.jpg" + } + ] + } + ], + "index": 127, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 603, + 481, + 608 + ], + "lines": [ + { + "bbox": [ + 465, + 603, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 465, + 603, + 481, + 608 + ], + "type": "text", + "content": "Scones" + } + ] + } + ], + "index": 128, + "angle": 0, + "type": "image_caption" + } + ], + "index": 127 + }, + { + "type": "image", + "bbox": [ + 507, + 535, + 541, + 602 + ], + "blocks": [ + { + "bbox": [ + 507, + 535, + 541, + 602 + ], + "lines": [ + { + "bbox": [ + 507, + 535, + 541, + 602 + ], + "spans": [ + { + "bbox": [ + 507, + 535, + 541, + 602 + ], + "type": "image", + "image_path": "e20646a086cffac4ae116315a50305af6784ccc6c59bd2082f4bdeee1ee46165.jpg" + } + ] + } + ], + "index": 129, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 514, + 603, + 533, + 608 + ], + "lines": [ + { + "bbox": [ + 514, + 603, + 533, + 608 + ], + "spans": [ + { + "bbox": [ + 514, + 603, + 533, + 608 + ], + "type": "text", + "content": "A muffin" + } + ] + } + ], + "index": 130, + "angle": 0, + "type": "image_caption" + } + ], + "index": 129 + }, + { + "type": "image", + "bbox": [ + 53, + 610, + 123, + 677 + ], + "blocks": [ + { + "bbox": [ + 53, + 610, + 123, + 677 + ], + "lines": [ + { + "bbox": [ + 53, + 610, + 123, + 677 + ], + "spans": [ + { + "bbox": [ + 53, + 610, + 123, + 677 + ], + "type": "image", + "image_path": "2fa74eae159321dd188194042693392198fcb61858fa854a7b9382491534aef9.jpg" + } + ] + } + ], + "index": 131, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 678, + 99, + 683 + ], + "lines": [ + { + "bbox": [ + 76, + 678, + 99, + 683 + ], + "spans": [ + { + "bbox": [ + 76, + 678, + 99, + 683 + ], + "type": "text", + "content": "A wise owl" + } + ] + } + ], + "index": 132, + "angle": 0, + "type": "image_caption" + } + ], + "index": 131 + }, + { + "type": "image", + "bbox": [ + 126, + 610, + 196, + 677 + ], + "blocks": [ + { + "bbox": [ + 126, + 610, + 196, + 677 + ], + "lines": [ + { + "bbox": [ + 126, + 610, + 196, + 677 + ], + "spans": [ + { + "bbox": [ + 126, + 610, + 196, + 677 + ], + "type": "image", + "image_path": "4f563e863ce68571d9fa9ceae9bb07ae33f3fb60e146a0dab8488c410f327a6e.jpg" + } + ] + } + ], + "index": 133, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 678, + 171, + 683 + ], + "lines": [ + { + "bbox": [ + 151, + 678, + 171, + 683 + ], + "spans": [ + { + "bbox": [ + 151, + 678, + 171, + 683 + ], + "type": "text", + "content": "Full moon" + } + ] + } + ], + "index": 134, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 166, + 694, + 426, + 705 + ], + "lines": [ + { + "bbox": [ + 166, + 694, + 426, + 705 + ], + "spans": [ + { + "bbox": [ + 166, + 694, + 426, + 705 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison to personalization based methods." + } + ] + } + ], + "index": 149, + "angle": 0, + "type": "image_caption" + } + ], + "index": 133 + }, + { + "type": "image", + "bbox": [ + 196, + 610, + 228, + 677 + ], + "blocks": [ + { + "bbox": [ + 196, + 610, + 228, + 677 + ], + "lines": [ + { + "bbox": [ + 196, + 610, + 228, + 677 + ], + "spans": [ + { + "bbox": [ + 196, + 610, + 228, + 677 + ], + "type": "image", + "image_path": "39fa8d35e7e3ba2427e87d43ffa4924cc93aae4c1d0d493c156b7c5d1acb4cdd.jpg" + } + ] + } + ], + "index": 135, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 678, + 220, + 683 + ], + "lines": [ + { + "bbox": [ + 204, + 678, + 220, + 683 + ], + "spans": [ + { + "bbox": [ + 204, + 678, + 220, + 683 + ], + "type": "text", + "content": "A book" + } + ] + } + ], + "index": 136, + "angle": 0, + "type": "image_caption" + } + ], + "index": 135 + }, + { + "type": "image", + "bbox": [ + 230, + 610, + 299, + 677 + ], + "blocks": [ + { + "bbox": [ + 230, + 610, + 299, + 677 + ], + "lines": [ + { + "bbox": [ + 230, + 610, + 299, + 677 + ], + "spans": [ + { + "bbox": [ + 230, + 610, + 299, + 677 + ], + "type": "image", + "image_path": "2bb7b8d91f4af638659972a0ca014ec10d525a15fc9566d0869f376937cbc8a6.jpg" + } + ] + } + ], + "index": 137, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 678, + 275, + 683 + ], + "lines": [ + { + "bbox": [ + 255, + 678, + 275, + 683 + ], + "spans": [ + { + "bbox": [ + 255, + 678, + 275, + 683 + ], + "type": "text", + "content": "Full moon" + } + ] + } + ], + "index": 138, + "angle": 0, + "type": "image_caption" + } + ], + "index": 137 + }, + { + "type": "image", + "bbox": [ + 300, + 610, + 332, + 677 + ], + "blocks": [ + { + "bbox": [ + 300, + 610, + 332, + 677 + ], + "lines": [ + { + "bbox": [ + 300, + 610, + 332, + 677 + ], + "spans": [ + { + "bbox": [ + 300, + 610, + 332, + 677 + ], + "type": "image", + "image_path": "bc6a9fbaf50c348093b27ffe01d00455c1a6a322d45a787eed6f99fd26ddc061.jpg" + } + ] + } + ], + "index": 139, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 678, + 324, + 683 + ], + "lines": [ + { + "bbox": [ + 308, + 678, + 324, + 683 + ], + "spans": [ + { + "bbox": [ + 308, + 678, + 324, + 683 + ], + "type": "text", + "content": "A book" + } + ] + } + ], + "index": 140, + "angle": 0, + "type": "image_caption" + } + ], + "index": 139 + }, + { + "type": "image", + "bbox": [ + 336, + 610, + 403, + 677 + ], + "blocks": [ + { + "bbox": [ + 336, + 610, + 403, + 677 + ], + "lines": [ + { + "bbox": [ + 336, + 610, + 403, + 677 + ], + "spans": [ + { + "bbox": [ + 336, + 610, + 403, + 677 + ], + "type": "image", + "image_path": "d5733e81801d09ddb91dd13aa468c0647906a7565c02d87e74e2a016c17152db.jpg" + } + ] + } + ], + "index": 141, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 678, + 380, + 683 + ], + "lines": [ + { + "bbox": [ + 359, + 678, + 380, + 683 + ], + "spans": [ + { + "bbox": [ + 359, + 678, + 380, + 683 + ], + "type": "text", + "content": "Full moon" + } + ] + } + ], + "index": 142, + "angle": 0, + "type": "image_caption" + } + ], + "index": 141 + }, + { + "type": "image", + "bbox": [ + 403, + 610, + 436, + 677 + ], + "blocks": [ + { + "bbox": [ + 403, + 610, + 436, + 677 + ], + "lines": [ + { + "bbox": [ + 403, + 610, + 436, + 677 + ], + "spans": [ + { + "bbox": [ + 403, + 610, + 436, + 677 + ], + "type": "image", + "image_path": "3a71945b845f2970193a4628958143b266d4488bfa7f69cf7b4ced283b29b9c6.jpg" + } + ] + } + ], + "index": 143, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 678, + 427, + 683 + ], + "lines": [ + { + "bbox": [ + 412, + 678, + 427, + 683 + ], + "spans": [ + { + "bbox": [ + 412, + 678, + 427, + 683 + ], + "type": "text", + "content": "A book" + } + ] + } + ], + "index": 144, + "angle": 0, + "type": "image_caption" + } + ], + "index": 143 + }, + { + "type": "image", + "bbox": [ + 438, + 610, + 507, + 677 + ], + "blocks": [ + { + "bbox": [ + 438, + 610, + 507, + 677 + ], + "lines": [ + { + "bbox": [ + 438, + 610, + 507, + 677 + ], + "spans": [ + { + "bbox": [ + 438, + 610, + 507, + 677 + ], + "type": "image", + "image_path": "ec4bb150faf2d7acfdb5c2bcb832dccc3d4f65619f949822f0f87ff929759a87.jpg" + } + ] + } + ], + "index": 145, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 678, + 484, + 683 + ], + "lines": [ + { + "bbox": [ + 463, + 678, + 484, + 683 + ], + "spans": [ + { + "bbox": [ + 463, + 678, + 484, + 683 + ], + "type": "text", + "content": "Full moon" + } + ] + } + ], + "index": 146, + "angle": 0, + "type": "image_caption" + } + ], + "index": 145 + }, + { + "type": "image", + "bbox": [ + 507, + 610, + 541, + 677 + ], + "blocks": [ + { + "bbox": [ + 507, + 610, + 541, + 677 + ], + "lines": [ + { + "bbox": [ + 507, + 610, + 541, + 677 + ], + "spans": [ + { + "bbox": [ + 507, + 610, + 541, + 677 + ], + "type": "image", + "image_path": "991f0ea54f6cbd853d334702a4074807d20a62f8c656ffa058f21edf2e85dc5d.jpg" + } + ] + } + ], + "index": 147, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 515, + 678, + 531, + 683 + ], + "lines": [ + { + "bbox": [ + 515, + 678, + 531, + 683 + ], + "spans": [ + { + "bbox": [ + 515, + 678, + 531, + 683 + ], + "type": "text", + "content": "A book" + } + ] + } + ], + "index": 148, + "angle": 0, + "type": "image_caption" + } + ], + "index": 147 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4780" + } + ] + } + ], + "index": 150 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 287, + 327 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 287, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 287, + 327 + ], + "type": "image", + "image_path": "ada735eb1902ca79d5ab11a26e4fe25d02bb9659ddec19de02a6b75738006b8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 335, + 288, + 380 + ], + "lines": [ + { + "bbox": [ + 46, + 335, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 288, + 380 + ], + "type": "text", + "content": "Figure 8. Varying level of attention sharing. By reducing the number of shared attention layers, i.e., allowing only self-attention in part of the layers, we can get more varied results (bottom rows) at the expense of style alignment (top row)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 389, + 133, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 389, + 133, + 402 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 133, + 402 + ], + "type": "text", + "content": "4.2. Comparisons" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 407, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 610 + ], + "type": "text", + "content": "For baselines, we compare our method to T2I personalization methods. We trained StyleDrop [49] and DreamBooth [41] over the first image in each set of our evaluation data, and use the trained personalized weights to generate the additional three images in each set. We use a public unofficial implementation of StyleDrop * (SDRP-unofficial) over non-regressive T2I model. Due to the large quality gap between the unofficial MUSE model † to the official one [7], we follow StyleDrop and implement an adapter model over SDXL (SDRP-SDXL) *, where we train a low rank linear layer after each Feed-Forward layer at the model's attention blocks. For training DreamBooth, we adapt the LoRA [21, 43] variant (DB-LoRA) over SDXL using the public huggingface-diffusers implementation ‡. We follow the hyperparameters tuning reported in [49] and train both SDRP-SDXL and DB-LoRA for 400 steps to prevent overfitting to the style training image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 611, + 288, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 611, + 288, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 611, + 288, + 659 + ], + "type": "text", + "content": "As can be seen in the qualitative comparison, Fig. 7, the image sets generated by our method, are more consistent across style attributes like color palette, drawing style, and composition. Moreover, the personalization-based methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "content": "may leak the content of the training reference image (on the left) when generating the new images. For example, see the repeated woman and dogs in the results of DB-LoRA and SDRP-SDXL at the second row or the repeated owl at the bottom row. Similarly, because of the content leakage, these methods obtained lower text similarity scores and higher set consistency scores compared to our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 156, + 546, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 546, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 546, + 300 + ], + "type": "text", + "content": "We also apply two encoder-based personalization methods ELITE [57], IP-Adapter [59], and BLIP-Diffusion [27] over our evaluation set. These methods receive as input the first image in each set and use its embeddings to generate images with other content. Unlike the optimization-based techniques, these methods operate in a much faster feedforward diffusion loop, like our method. However, as can be seen in Fig. 6, their performance for style aligned image generation is poor compared to the other baselines. We argue that current encoder-based personalization techniques struggle to disentangle the content and the style of the input image. We supply qualitative results in the supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 300, + 545, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 300, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 545, + 359 + ], + "type": "text", + "content": "User study. In addition to the automatic evaluation, we conducted a user study over the results of our method, StyleDrop (unofficial MUSE), StyleDrop (SDXL), and DreamBooth-LoRA (SDXL). In each question, we randomly sample one of the evaluation examples and show" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 377, + 545, + 627 + ], + "blocks": [ + { + "bbox": [ + 307, + 377, + 545, + 627 + ], + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 627 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 627 + ], + "type": "image", + "image_path": "68ce1e879762b8e47f9f2caa41262a4af139d672c99f321df933fc63e9d4f341.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 633, + 547, + 700 + ], + "lines": [ + { + "bbox": [ + 305, + 633, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 305, + 633, + 547, + 700 + ], + "type": "text", + "content": "Figure 9. Style aligned image generation to an input image. Given an input reference image (left column) and text description, we first apply DDIM inversion over the image to get the inverted diffusion trajectory " + }, + { + "bbox": [ + 305, + 633, + 547, + 700 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 305, + 633, + 547, + 700 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 633, + 547, + 700 + ], + "type": "inline_equation", + "content": "x_{T-1} \\ldots x_{0}" + }, + { + "bbox": [ + 305, + 633, + 547, + 700 + ], + "type": "text", + "content": ". Then, starting from " + }, + { + "bbox": [ + 305, + 633, + 547, + 700 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 305, + 633, + 547, + 700 + ], + "type": "text", + "content": " and a new set of prompts, we apply our method to generate new content (right columns) with an aligned style to the input." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 61, + 666, + 246, + 676 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 666, + 246, + 676 + ], + "spans": [ + { + "bbox": [ + 61, + 666, + 246, + 676 + ], + "type": "text", + "content": "\\*github.com/aim-uofa/StyleDrop-PyTorch" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 63, + 677, + 231, + 685 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 677, + 231, + 685 + ], + "spans": [ + { + "bbox": [ + 63, + 677, + 231, + 685 + ], + "type": "text", + "content": "github.com/baaivision/MUSE-Pytorch" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 63, + 685, + 225, + 694 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 685, + 225, + 694 + ], + "spans": [ + { + "bbox": [ + 63, + 685, + 225, + 694 + ], + "type": "text", + "content": "\\*github.com/amirhertz/smlx_adapter" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 63, + 695, + 220, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 695, + 220, + 704 + ], + "spans": [ + { + "bbox": [ + 63, + 695, + 220, + 704 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 63, + 695, + 220, + 704 + ], + "type": "text", + "content": "github.com/huggingface/diffusers" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4781" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "the user the 4 image set that resulted from our and another method (in a random order). The user had to choose which set is better in terms of style consistency, and text alignment. A print screen of the user study format is provided in the supplementary. Overall, we collected 2400 answers from 100 users using the Amazon Mechanical Turk service. The results are summarized in Tab. 1 where for each method, we report the percentage of judgments in our favor. As can be seen, most participants favored our method by a large margin. More information about our user study can be found in the supplementary." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 211, + 157, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 211, + 157, + 221 + ], + "spans": [ + { + "bbox": [ + 47, + 211, + 157, + 221 + ], + "type": "text", + "content": "4.3. Additional Results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 229, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 287, + 300 + ], + "type": "text", + "content": "Style alignment control. We provide means of control over the style alignment to the reference image by applying the shared attention over only part of the self-attention layers. As can be seen in Fig. 8, reducing the number of shared attention layers results with a more diverse image set, which still shares common attributes with the reference image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 300, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 479 + ], + "type": "text", + "content": "StyleAligned from an input image. To generate style-aligned images to an input image, we apply DDIM inversion [50] using a provided text caption. Then, we apply our method to generate new images in the style of the input using the inverted diffusion trajectory " + }, + { + "bbox": [ + 46, + 300, + 287, + 479 + ], + "type": "inline_equation", + "content": "x_{T}, x_{T-1}, \\ldots, x_{0}" + }, + { + "bbox": [ + 46, + 300, + 287, + 479 + ], + "type": "text", + "content": " for the reference image. Examples are shown in Fig. 9, where we use BLIP captioning [28] to get a caption for each input image. For example, we used the prompt \"A render of a house with a yellow roof\" for the DDIM inversion of the top example and replaced the word house with other objects to generate the style-aligned images of a car, a cat, and a cactus. Notice that this method does not require any optimization. However, DDIM inversion may fail [30] or results with an erroneous trajectory [24]. More results and analysis, are provided in the supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 480, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 287, + 577 + ], + "type": "text", + "content": "StyleAligned with other methods. Since our method doesn't require training or optimization, it can be easily combined on top of other diffusion based methods to generate style-consistent image sets. Fig. 10 shows several such examples where we combine our method with ControlNet [60], DreamBooth [42] and MultiDiffusion [3]. More examples and details about the integration of StyleAligned with other methods can be found in supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 586, + 125, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 125, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 125, + 597 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "We have presented StyleAligned, which addresses the challenge of achieving style-aligned image generation within the realm of large-scale Text-to-Image models. By introducing minimal attention sharing operations with AdaIN modulation during the diffusion process, our method successfully establishes style-consistency and visual coherence across generated images. The demonstrated efficacy of StyleAligned in producing high-quality, style-consistent images across diverse styles and textual prompts under" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 68, + 547, + 419 + ], + "blocks": [ + { + "bbox": [ + 306, + 68, + 547, + 419 + ], + "lines": [ + { + "bbox": [ + 306, + 68, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 68, + 547, + 419 + ], + "type": "image", + "image_path": "48eab7e15d523b557b2e445a5c7f8acebe22b8794caccc6a4756084a368b7c75.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 428, + 547, + 506 + ], + "lines": [ + { + "bbox": [ + 304, + 428, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 547, + 506 + ], + "type": "text", + "content": "Figure 10. StyleAligned with other methods. On top, StyleAligned is combined with ControlNet to generate style-aligned images conditioned on depth maps. In the middle, our method combined with MultiDiffusion to generate panorama images that share multiple styles. On the bottom, style consistent and personalized content created by combining our method with pretrained personalized DreamBooth-LoRA models." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 515, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 545, + 563 + ], + "type": "text", + "content": "scores its potential in creative domains and practical applications. Our results affirm StyleAligned capability to faithfully adhere to provided descriptions and reference styles while maintaining impressive synthesis quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "text", + "content": "In the future we would like to explore the scalability and adaptability of StyleAligned to have more control over the shape and appearance similarity among the generated images. Additionally, due to the limitation of current diffusion inversion methods, a promising direction is to leverage StyleAligned to assemble a style-aligned dataset which then can be used to train style condition text-to-image models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 658, + 416, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 416, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 416, + 671 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 676, + 545, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 676, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 304, + 676, + 545, + 709 + ], + "type": "text", + "content": "We thank Or Patashnik, Matan Cohen, Yael Pritch, and Yael Vinker for their valuable inputs that helped improve this work." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4782" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 147 + ], + "type": "text", + "content": "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM Trans. Graph., 42(4), jul 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 147, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 179 + ], + "type": "text", + "content": "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 180, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 236 + ], + "type": "text", + "content": "[4] James Betker, Gabriel Goh, Li Jing, TimBrooks, Jianfeng Wang, Linjie Li, LongOuyang, JuntangZhuang, JoyceLee, YufeiGuo, WesamManassra, PrafullaDhariwal, CaseyChu, YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 288, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 301 + ], + "type": "text", + "content": "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. MasaCtrl: tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22560–22570, October 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 357, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 424 + ], + "type": "text", + "content": "[7] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, José Lezama, Lu Jiang, Ming Yang, Kevin P. Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. In International Conference on Machine Learning, 2023. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "text", + "content": "[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42:1 - 10, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 468, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 288, + 513 + ], + "type": "text", + "content": "[9] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In The Eleventh International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "type": "text", + "content": "[10] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 571-576. 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "type": "text", + "content": "[11] Alexei A Efros and Thomas K Leung. Texture synthesis by non-parametric sampling. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 1033-1038. IEEE, 1999. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "text", + "content": "[12] Dave Epstein, Allan Jabri, Ben Poole, Alexei A Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. arXiv preprint arXiv:2306.00986, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 691 + ], + "type": "text", + "content": "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, 2022. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[14] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 701 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 140 + ], + "type": "text", + "content": "[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "text", + "content": "[16] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris N. Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. ArXiv, abs/2303.11305, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 185, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 229 + ], + "type": "text", + "content": "[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 230, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 274 + ], + "type": "text", + "content": "[18] Aaron Hertzmann, Charles E Jacobs, Nuria Oliver, Brian Curless, and David H Salesin. Image analogies. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 557-570. 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 275, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 275, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 275, + 545, + 297 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proc. NeurIPS, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 298, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 354 + ], + "type": "text", + "content": "[20] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "type": "text", + "content": "[21] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "text", + "content": "[22] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 443, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 443, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 443, + 545, + 487 + ], + "type": "text", + "content": "[23] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European conference on computer vision (ECCV), pages 172-189, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 488, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 545, + 521 + ], + "type": "text", + "content": "[24] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. arXiv preprint arXiv:2304.06140, 2023. 2, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 522, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 545, + 567 + ], + "type": "text", + "content": "[25] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 567, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 621 + ], + "type": "text", + "content": "[26] Hochang Lee, Sanghyun Seo, Seungtaek Ryoo, and Kyunghyun Yoon. Directional texture transfer. In Proceedings of the 8th International Symposium on Non-Photorealistic Animation and Rendering, pages 43–48, 2010. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 622, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 622, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 308, + 622, + 545, + 666 + ], + "type": "text", + "content": "[27] Dongxu Li, Junnan Li, and Steven C. H. Hoi. BLIP-Diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. ArXiv, abs/2305.14720, 2023. 2, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 667, + 545, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 667, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 308, + 667, + 545, + 701 + ], + "type": "text", + "content": "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 8" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4783" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 708 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsupervised image-to-image translation. In IEEE International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "text", + "content": "[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 288, + 228 + ], + "type": "text", + "content": "[31] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 229, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 229, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 229, + 288, + 284 + ], + "type": "text", + "content": "[32] Dong Huk Park*, Grace Luo*, Clayton Toste, Samaneh Azadi, Xihui Liu, Maka Karalashvili, Anna Rohrbach, and Trevor Darrell. Shape-guided diffusion with inside-outside attention. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 287, + 329 + ], + "type": "text", + "content": "[33] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 373 + ], + "type": "text", + "content": "[34] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. ArXiv, abs/2303.11306, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 374, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 374, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 374, + 287, + 430 + ], + "type": "text", + "content": "[35] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 431, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 431, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 48, + 431, + 287, + 496 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 498, + 287, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 551 + ], + "type": "text", + "content": "[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 553, + 287, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 553, + 287, + 607 + ], + "spans": [ + { + "bbox": [ + 48, + 553, + 287, + 607 + ], + "type": "text", + "content": "[38] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 609, + 288, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 609, + 288, + 663 + ], + "spans": [ + { + "bbox": [ + 48, + 609, + 288, + 663 + ], + "type": "text", + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 665, + 287, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 665, + 287, + 708 + ], + "spans": [ + { + "bbox": [ + 48, + 665, + 287, + 708 + ], + "type": "text", + "content": "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 708 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 161 + ], + "type": "text", + "content": "[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 5, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "text", + "content": "[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 2, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "text", + "content": "[43] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora, 2022.7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 252, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 252, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 308, + 252, + 545, + 306 + ], + "type": "text", + "content": "[44] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. ACM SIGGRAPH 2022 Conference Proceedings, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 308, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 373 + ], + "type": "text", + "content": "[45] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 375, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 375, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 375, + 545, + 441 + ], + "type": "text", + "content": "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 441, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 441, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 308, + 441, + 545, + 474 + ], + "type": "text", + "content": "[47] Jing Shi, Wei Xiong, Zhe L. Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. ArXiv, abs/2304.03411, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 475, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 545, + 529 + ], + "type": "text", + "content": "[48] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 531, + 545, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 531, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 308, + 531, + 545, + 585 + ], + "type": "text", + "content": "[49] Kihyuk Sohn, Nataniel Ruiz, Kimin Lee, Daniel Castro Chin, Irina Blok, Huiwen Chang, Jarred Barber, Lu Jiang, Glenn Entis, Yuanzhen Li, et al. Styledrop: Text-to-image generation in any style. arXiv preprint arXiv:2306.00983, 2023. 1, 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 587, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 587, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 308, + 587, + 545, + 620 + ], + "type": "text", + "content": "[50] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 621, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 621, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 308, + 621, + 545, + 654 + ], + "type": "text", + "content": "[51] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. SIGGRAPH 2023 Conference Proceedings, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 655, + 545, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 655, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 308, + 655, + 545, + 708 + ], + "type": "text", + "content": "[52] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10748-10757, 2022. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4784" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 418 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[53] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 205 + ], + "type": "text", + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 207, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 207, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 207, + 287, + 239 + ], + "type": "text", + "content": "[55] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. arXiv preprint arXiv:2211.13752, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 287, + 274 + ], + "type": "text", + "content": "[56] Andrey Voynov, Q. Chu, Daniel Cohen-Or, and Kfir Aberman. P+: Extended textual conditioning in text-to-image generation. ArXiv, abs/2303.09522, 2023. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 286, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 286, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 286, + 296 + ], + "type": "text", + "content": "[57] Yuxiang Wei. Official implementation of ELITE. https://github.com/csyxwei/ELITE, 2023. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "text", + "content": "[58] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: Encoding visual concepts into textual embeddings for customized text-to-image generation. ArXiv, abs/2302.13848, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 342, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 384 + ], + "type": "text", + "content": "[59] Hu Ye, Jun Zhang, Siyi Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. ArXiv, abs/2308.06721, 2023. 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 387, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 287, + 418 + ], + "type": "text", + "content": "[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 8" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4785" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/1e340b0f-2934-49e8-9933-238d4fb891ff_content_list.json b/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/1e340b0f-2934-49e8-9933-238d4fb891ff_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..89b78aeb701f4a84aacf7e900322538c860473db --- /dev/null +++ b/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/1e340b0f-2934-49e8-9933-238d4fb891ff_content_list.json @@ -0,0 +1,1743 @@ +[ + { + "type": "text", + "text": "Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning", + "text_level": 1, + "bbox": [ + 107, + 130, + 861, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Woo-Jin Ahn1 Geun-Yeong Yang1 Hyun-Duck Choi2* Myo-Taeg Lim1* \n1Korea University 2Chonnam National University", + "bbox": [ + 183, + 202, + 790, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}\\{wjahn, hggofficial, mlim\\} @korea.ac.kr$ $^{2}$ ducky.choi@jnu.ac.kr", + "bbox": [ + 218, + 239, + 750, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning models for semantic segmentation often experience performance degradation when deployed to unseen target domains unidentified during the training phase. This is mainly due to variations in image texture (i.e. style) from different data sources. To tackle this challenge, existing domain generalized semantic segmentation (DGSS) methods attempt to remove style variations from the feature. However, these approaches struggle with the entanglement of style and content, which may lead to the unintentional removal of crucial content information, causing performance degradation. This study addresses this limitation by proposing BlindNet, a novel DGSS approach that blinds the style without external modules or datasets. The main idea behind our proposed approach is to alleviate the effect of style in the encoder whilst facilitating robust segmentation in the decoder. To achieve this, BlindNet comprises two key components: covariance alignment and semantic consistency contrastive learning. Specifically, the covariance alignment trains the encoder to uniformly recognize various styles and preserve the content information of the feature, rather than removing the style-sensitive factor. Meanwhile, semantic consistency contrastive learning enables the decoder to construct discriminative class embedding space and disentangles features that are vulnerable to misclassification. Through extensive experiments, our approach outperforms existing DGSS methods, exhibiting robustness and superior performance for semantic segmentation on unseen target domains. The code is available at https://github.com/root0yang/BlindNet.", + "bbox": [ + 75, + 323, + 472, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 805, + 209, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation, a technique that classifies each pixel in an image into predefined categories, has garnered significant attention due to its potential applications in vari-", + "bbox": [ + 75, + 830, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f1ff12b29489673484c490594c0b0065fa431db390eaa18e1fb3b77b2196fbd2.jpg", + "image_caption": [ + "Figure 1. Comparison of semantic segmentation results between the baseline (DeepLabV3+ with ResNet50 backbone) and our BlindNet. Both models are trained on the source domain (GTAV [43]) and tested on the target domain (Cityscapes [8])." + ], + "image_footnote": [], + "bbox": [ + 500, + 287, + 893, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ous fields. Particularly, it plays a crucial role in autonomous driving [1, 18] and robotic systems [34, 35]. Besides, with the advent of large datasets, deep neural networks have emerged as a trending approach for semantic segmentation tasks, achieving impressive results [3, 5, 44, 54]. However, there remains a major bottleneck detailing the meticulous and labor-intensive process of dataset labeling. More particularly, this process not only consumes time but also poses economic challenges [8, 46]. To address this challenge, synthetic datasets have emerged as a compelling alternative. Specifically, these datasets, generated using three-dimensional (3D) rendering techniques, offer vast amounts of easily accessible data, eliminating the need for manual labeling [43, 45]. However, a challenge arises when models trained on synthetic datasets are deployed in real-world scenarios. More precisely, a domain shift problem arises due to style factor discrepancies (e.g. texture, illumination, and image quality) between synthetic and real-world data, which affects the performance of the model, as shown in Fig. 1.", + "bbox": [ + 496, + 506, + 893, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the domain shift problem, domain adaptation semantic segmentation (DASS) has been introduced [15, 17, 22, 25, 51, 62]. Specifically, DASS aims to bridge the gap between source and target domains by aligning their data distributions. However, a significant limitation of DASS is its dependency on the target domain during train-", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Authors", + "bbox": [ + 94, + 886, + 228, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3616", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing. For DASS to function effectively, target domain samples must be available during the training phase.", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, domain generalized semantic segmentation (DGSS) has been introduced as an alternative approach to tackle the domain shift problem. Unlike DASS, DGSS is trained only with the source domain, aiming to extract domain-invariant features. To achieve this, two main techniques have been employed: domain randomization (DR) and feature normalization (FN).", + "bbox": [ + 75, + 123, + 467, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DR augments the training set by introducing variability, either by altering the image style [41, 60] or by modifying the feature representation [24, 53]. By exposing the model to a wider variety of styles via DR, the network is less likely to overfit to the specific styles present in the training data. Consequently, the robustness of the model is improved, making it more adept at generalizing to new, unseen domains. Nevertheless, a crucial limitation of DR is its significant dependence on auxiliary domains.", + "bbox": [ + 75, + 229, + 467, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "FN methods, converse to DR, regularize the features to prevent the model from overfitting to the distinct styles or characteristics of the training data. This is achieved by removing domain-specific style information using feature statistics, such as instance normalization [38] or whitening transformation [7, 19, 39, 42]. While these approaches effectively remove style-related information, they simultaneously pose the challenge of removing semantic content because content and style information are entangled. Consequently, the model fails to capture the essential patterns or features required for accurate segmentation prediction.", + "bbox": [ + 75, + 367, + 467, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this problem, we propose BlindNet, a model that blinds the style within the encoder and improves the robustness of the decoder, without requiring auxiliary datasets or external modules. Specifically, our proposed BlindNet consists of two components: covariance alignment for the encoder and semantic consistency contrastive learning for the decoder. Precisely, the covariance alignment facilitates the generation of style-invariant features with the proposed covariance matching loss function (CML) and the cross-covariance loss function (CCL). Specifically, CML mitigates the effects of style variations, while CCL focuses on preserving content information, effectively addressing the prevalent content information loss observed in the FN method. To further improve the generalization ability, we develop semantic consistency contrastive learning, which consists of class-wise contrastive learning (CWCL) and semantic disentanglement contrastive learning (SDCL). Particularly, the CWCL constructs a discriminative class embedding space, while SDCL disentangles features of similar classes that often lead to prediction errors. Extensive experiments across various datasets demonstrate that the proposed BlindNet outperforms existing DGSS methods.", + "bbox": [ + 75, + 535, + 467, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 96, + 868, + 398, + 882 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a covariance alignment within the encoder,", + "bbox": [ + 76, + 885, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "comprising CML and CCL. Specifically, the CML aims to mitigate the effects of style variations, while CCL ensures the preservation of content information, together facilitating the generation of style-agnostic features.", + "bbox": [ + 509, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose semantic consistency contrastive learning within the decoder that comprises CWCL and SDCL, utilizing segmentation masks. Specifically, CWCL generates discriminative embeddings, whereas SDCL disentangles features of similar classes, enhancing the robustness of the model.", + "- Through extensive experiments, we demonstrate the superiority of our approach in DGSS, without the need to alter the network architecture or rely on external datasets." + ], + "bbox": [ + 500, + 152, + 890, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 301, + 648, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain adaptation and generalization for semantic segmentation. Domain adaptation (DA) aims at minimizing the distribution discrepancy between different domains, enabling a model to generalize from a source to a target domain. For DASS, adversarial training and cross-domain self-training strategies are commonly used. Particularly, adversarial-based methods [15, 29] employ generative adversarial networks [11] to close the feature distribution gap between source and target domains. Meanwhile, cross-domain self-training methods [16, 37, 59, 62] generate pseudo-labels for target domain data using pre-trained models, and employ them as training data, thereby expanding the training data and reducing distribution differences between the source and target domain.", + "bbox": [ + 498, + 325, + 890, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain generalization (DG) methods [26, 28, 30, 50, 61] aim to improve the generalization ability of the model without accessing the target domain during training. Since the difference in style of the image is the main cause of the disparity, most existing domain generalized semantic segmentation methods utilize the style information of the image for domain-invariant learning. Feature statistics (e.g. mean, variance, covariance, gram matrix, etc.) which are commonly used in style transfer [10, 21, 27, 57] are employed to capture the style information. Interestingly, existing DGSS methods can be separated into two parts: i) domain randomization to expand the distribution of style or ii) feature normalization to remove style.", + "bbox": [ + 496, + 537, + 890, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DR involves randomizing either the image or its features through stylization to learn domain-invariant features from various styles. For example, Peng et al. [42] extended the source domain data by stylizing images in the style of unreal paintings. Similarly, Yue et al. [60] and Huang et al. [22] attempted to enhance generalization by synthesizing images with diverse styles in the image space. In another study, Lee et al. [24] adopted ImageNet data [9] as wild data and performed randomization via synthesis in the feature space. Meanwhile, Wu et al. [53] diversified the trainable feature space by mixing the statistics of the feature and its color-", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3617", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "jittered feature with Ada-IN [21].", + "bbox": [ + 75, + 90, + 299, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FN methods aim to remove domain-specific styles from features, extracting only domain-invariant content. For instance, Pan et al. [39] first attempted the DGSS method, combining batch normalization (BN) [23] and instance normalization (IN) [48] in the network layer. While BN preserves the content information within discriminative features, IN focuses on removing domain-specific style information from features. In a study, Choi et al. [7] addressed the limitations of previous whitening transformation [6, 31] that can eliminate the content information from the feature. Specifically, they proposed an instance-selective whitening approach designed to remove covariance components that are sensitive to domain shifts. Peng et al. [41] developed semantic-aware normalization that performs on class-wise and semantic-aware whitening that aligns channels based on the prediction through group whitening transformation [6]. In a study, Xu et al. [55] introduced the prior guided attention module and guided feature whitening to re-calibrate the feature and remove domain-specific style effects, respectively. Unlike the FN methods that directly remove the style component, our work explores a covariance alignment method that mitigates the effect of the style's effect while preserving the content information. Our method achieves the DGSS without any additional modules or auxiliary datasets.", + "bbox": [ + 75, + 109, + 472, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive Learning. Contrastive learning aims to learn representations by maximizing the similarity between positive pairs of samples while minimizing the similarity between negative pairs. In recent years, it has attracted significant attention for its effectiveness in learning discriminative representations across various tasks [2, 4, 12, 14]. Oord et al. [36] were the first to introduce the infoNCE loss, a type of contrastive loss function designed for self-contrastive learning. In a work, Park et al. [40] introduced patch-level contrastive learning for image translation, using co-located patches as positive pairs and spatially distant patches as negatives to maintain image context. For the semantic segmentation task, Wang et al. [52] introduced class-wise contrastive learning to aid the model in learning the embedding space of each class. Specifically, they sampled the classes existing in the images and applied contrastive learning based on the class label. For DGSS, Lee et al. [24] adopted contrastive learning to learn the ImageNet information in their model. Specifically, they set the ImageNet data as wild and applied the contrastive learning method by setting the wild-stylized feature and its closest wild content as positive samples. In another study, Yang et al. [56] developed multi-level contrastive learning, which designed instance prototypes and class prototypes for contrastive learning. Specifically, they sample each class's pixel features and apply contrastive learning with a transition-probability ability matrix. Unlike recent DGSS works that embed the origi", + "bbox": [ + 75, + 493, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "nal image, we define contrastive learning for the augmented image. Specifically, the proposed method builds a robust embedding space by preserving the semantic consistency of the feature representation across various domains.", + "bbox": [ + 496, + 90, + 893, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 498, + 164, + 591, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of the proposed method is to train a segmentation model $\\varphi$ on a given source domain $S$ and generalize well to the unseen target domain $T$ . Precisely, the source domain $S = \\{(x,y)\\}$ contains a paired image $x \\in \\mathbb{R}^{H \\times W \\times 3}$ and segmentation label $y \\in \\mathbb{R}^{H \\times W \\times C}$ , where $H$ , $W$ , and $C$ denote the height of the image, the width of the image, and the class number of the segmentation map, respectively. The model $\\varphi$ takes an original $x$ and its augmented counterpart $x_{a}$ , which have the same content but different styles, and uses the feature information to enhance its generalization ability.", + "bbox": [ + 496, + 190, + 893, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, our method leverages the feature information through two main approaches: covariance alignment and semantic consistency contrastive learning. Specifically, the covariance matching ensures that features, having different styles but the same content, contain similar information. Additionally, the semantic consistency contrastive learning embeds the generalized features into discriminative representation based on the segmentation label.", + "bbox": [ + 496, + 356, + 893, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Covariance Alignment", + "text_level": 1, + "bbox": [ + 498, + 486, + 710, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The domain shift in semantic segmentation results from changes in the visual characteristics of the image, known as style. This style information is typically detected in the shallow layers of networks [38], which are the encoders of the segmentation model. Based on this understanding, our method targets the encoder to handle style variations by employing the proposed covariance matching loss and cross-covariance loss function.", + "bbox": [ + 496, + 508, + 893, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Covariance Matching Loss", + "text_level": 1, + "bbox": [ + 498, + 648, + 692, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train the network to uniformly recognize various styles without removing content information, we introduce the CML. Specifically, the loss aims to minimize the difference between covariance matrices derived from different styles of image features. Given an image pair $(x,x_{a})$ , the features from $i^{th}$ block of the encoder are represented as $F^i\\in \\mathbb{R}^{(H^i\\times W^i)\\times C^i}$ and $F_{a}^{i}\\in \\mathbb{R}^{(H^{i}\\times W^{i})\\times C^{i}}$ , respectively. Further, following the methodologies of [7, 41], we compute the covariance matrices using instance normalized features [48], which ensures consistent scaling across features. The feature maps are normalized and flattened into $\\bar{F}$ and $\\bar{F}_a\\in \\mathbb{R}^{(HW)\\times C}$ , which are given by:", + "bbox": [ + 496, + 672, + 893, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {F} = \\frac {(F - \\mu (F))}{\\sigma (F)}, \\bar {F} _ {a} = \\frac {(F _ {a} - \\mu (F _ {a}))}{\\sigma (F _ {a})}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 864, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3618", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5fcbbb724d3720e68e61b5e9b5a91dabb76d3372b3e54fd809de935bffa75425.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed BlindNet. The network processes a pair of images - the original image $x$ and its augmented counterpart $x_{a}$ . It employs covariance alignment to treat encoder features and utilizes semantic consistency contrastive learning for the processing of decoder features." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 890, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu (\\cdot)\\in \\mathbb{R}^{C}$ and $\\sigma (\\cdot)\\in \\mathbb{R}^C$ denote the mean and standard deviation of the features. Utilizing the normalized features, we evaluate the covariance matrices for the original and augmented image features as:", + "bbox": [ + 75, + 481, + 468, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma_ {x, x} ^ {i} = \\bar {F} ^ {i} ^ {T} \\cdot \\bar {F} ^ {i}, \\quad \\Sigma_ {x _ {a}, x _ {a}} ^ {i} = \\bar {F} _ {a} ^ {i} ^ {T} \\cdot \\bar {F} _ {a} ^ {i}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 553, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The CML is then formulated to align these covariance matrices, ensuring that the network maintains consistency in the presence of style variations. The CML is defined as:", + "bbox": [ + 75, + 585, + 468, + 631 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C M} = \\sum_ {i = 1} ^ {n _ {e}} \\| \\Sigma_ {x, x} ^ {i} - \\Sigma_ {x _ {a}, x _ {a}} ^ {i} \\| _ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 643, + 468, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $n_e$ denotes the number of blocks of the encoder.", + "bbox": [ + 76, + 691, + 437, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-covariance Loss", + "text_level": 1, + "bbox": [ + 76, + 726, + 236, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the CML effectively aligns the internal distributions of features within the same image, it does not fully account for the direct correlations across paired images. To complement this, we introduce CCL, which aims to encode the consistent content information of an image pair $(x,x_{a})$ by utilizing the cross-covariance of the image pair. Given the normalized feature pair $(\\bar{F}^i,\\bar{F}_a^i)$ , the cross-covariance of the feature pair can be expressed as:", + "bbox": [ + 75, + 748, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma_ {x, x _ {a}} ^ {i} = \\bar {F} ^ {i} ^ {T} \\cdot \\bar {F} _ {a} ^ {i}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 880, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The cross-covariance is expected to exhibit an identity matrix, as the feature pair should contain identical information. Nonetheless, the proposed CCL converges only the diagonal component of the covariance matrix to one. This is to prevent the drawbacks of the existing FN methods [7, 42] from removing content information. The CCL function is thus defined as:", + "bbox": [ + 498, + 482, + 890, + 587 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C C} = \\sum_ {i = 1} ^ {n _ {e}} \\| \\operatorname {d i a g} \\left(\\Sigma_ {x, x _ {a}} ^ {i}\\right) - \\mathbb {1} \\| _ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 592, + 890, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{diag}(\\Sigma_c) \\in \\mathbb{R}^C$ denotes the column vector comprising diagonal elements of $\\Sigma_{x,x_a}^i$ and $\\mathbb{1} \\in \\mathbb{R}^C$ denotes the one vector.", + "bbox": [ + 498, + 638, + 890, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Semantic Consistence Contrastive Learning", + "text_level": 1, + "bbox": [ + 500, + 691, + 870, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the encoder focuses on generating style-blinded features, the decoder aims to improve the robustness of the segmentation prediction against domain shifts. For the decoder, we employ a contrastive learning approach, which has demonstrated effectiveness in extracting discriminative features [4]. Specifically, we utilize the InfoNCE loss [36], which is formulated as:", + "bbox": [ + 498, + 714, + 890, + 819 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I N} (a, p, n) = - \\log \\left(\\frac {e ^ {(a \\cdot p / \\tau)}}{e ^ {(a \\cdot p / \\tau)} + \\sum_ {n} ^ {N ^ {-}} e ^ {(a \\cdot n / \\tau)}}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 823, + 890, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $a, p, n$ , and $N^{-}$ denote anchor, positive sample, negative sample, and negative sample set, respectively.", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3619", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f9fb00733b6b0939d8e914bad891b9b9d71654c5baf4d18638084e017a3d791f.jpg", + "image_caption": [ + "Figure 3. Illustration of semantic consistency contrastive learning: The mask $M$ represents the error mask derived from the augmented segmentation map. CWCL conducts contrastive learning by sampling per segmentation class and SDCL conducts contrastive learning based on the $M$ . Both methods share a projection head $\\pi$ for the semantic representation." + ], + "image_footnote": [], + "bbox": [ + 98, + 85, + 872, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve consistent feature representation in DGSS across various styles, we introduce semantic consistency contrastive learning. Specifically, the anchor is derived from the augmented image $x_{a}$ , while the positive sample is extracted from the corresponding pixel of the original image $x$ . Our method consists of two main components based on the negative sample as shown in Fig. 3: class-wise contrastive learning and semantic disentanglement contrastive learning.", + "bbox": [ + 75, + 321, + 470, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Class-wise Contrastive Learning", + "text_level": 1, + "bbox": [ + 76, + 491, + 308, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our CWCL aims to build a discriminative embedding space for each segmentation class using different classes of the original image as negatives. Given an image pair $(x, x_{a})$ , the features from the $j^{th}$ block of the decoder at pixel position $(m, n)$ are denoted as $F_{(m, n)}^{j}$ and $F_{a, (m, n)}^{j} \\in \\mathbb{R}^{(1 \\times 1) \\times C^{j}}$ , where $C^{j}$ indicates the channel length of the feature. As mentioned above, we take $F_{a, (m, n)}^{j}$ as the anchor and $F_{(m, n)}^{j}$ as the positive sample, since they represent the same content at the corresponding spatial location. To obtain the negative samples from $F^{j}$ , we leverage the resized segmentation class label $y^{j} \\in \\mathbb{R}^{(H^{j} \\times W^{j}) \\times C}$ to collect the different class samples. The samples are passed through the projection head, denoted as $\\pi$ , resulting in the projected features $\\tilde{F}$ and $\\tilde{F}_{a}$ . We define our CWCL as:", + "bbox": [ + 75, + 518, + 470, + 743 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C W C L} = \\sum_ {j} ^ {n _ {d}} \\mathcal {L} _ {I N} \\left(\\tilde {F} _ {a, (m, n)} ^ {j}, \\tilde {F} _ {(m, n)} ^ {j}, \\tilde {F} _ {(p, q)} ^ {j}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 758, + 468, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(p,q)\\in \\{(p,q)\\in P|y_{(p,q)}^j\\neq y_{(m,n)}^j\\}$", + "bbox": [ + 122, + 801, + 429, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The set $P$ represents all pixel positions in feature $F^j$ , with dimensions $H^j \\times W^j$ corresponding to the height and width. Additionally, $n_d$ denotes the total number of blocks in the decoder.", + "bbox": [ + 75, + 839, + 470, + 898 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/364e0bcd00ec452f09a91d3c336a3abe80fbdab1f326b463d00f6dc6623bce24.jpg", + "image_caption": [ + "Figure 4. t-SNE [49] visualization comparing scenarios with and without $\\mathcal{L}_{SCCL}$ . In (b), the application of SCCL results in a clear separation between the sidewalk (pink), the road (purple), and the building (gray)." + ], + "image_footnote": [], + "bbox": [ + 540, + 318, + 694, + 431 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1185b7eb568978d73c2f83c1c0f9c882983f91b10edd00fd456149b4d75b6763.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 329, + 852, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic Disentanglement Contrastive Learning", + "text_level": 1, + "bbox": [ + 500, + 522, + 843, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Domain shifts can lead to the entanglement of similar classes, causing the model to misclassify, as illustrated in Fig. 4. To mitigate this issue, we introduce the SDCL, specifically designed to disentangle the feature $x_{a}$ that has been misclassified, making it closer to the correct class and far from the misclassified class to achieve effective disentanglement. To further ensure consistent feature space and capture the semantic meaning, we share the projection head $\\pi$ used in the CWCL loss. Given the predicted segmentation map of the augmented image, represented as $\\hat{y}_a = \\varphi (x_a)$ , we resize it to $\\hat{y}_{a,(m,n)}^j\\in \\mathbb{R}^{(H^j\\times W^j)\\times C}$ . Similarly, $y_{(m,n)}^j$ represents the ground truth segmentation map. Using these segmentation maps, we set the anchor at positions where $\\hat{y}_{a,(m,n)}^j\\neq y_{(m,n)}^j$ . Negative samples are selected from the augmented image features corresponding to the anchor's misclassified class. The samples go through the projection head $\\pi$ . Our SDCL loss is defined as follows:", + "bbox": [ + 496, + 546, + 890, + 803 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S D C L} = \\sum_ {j} ^ {n _ {d}} \\mathcal {L} _ {I N} \\left(\\tilde {F} _ {a, (m, n)} ^ {j}, \\tilde {F} _ {(m, n)} ^ {j}, \\tilde {F} _ {a, (r, s)} ^ {j}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 811, + 890, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(r,s)\\in \\{(r,s)\\in P|y_{(r,s)}^j = \\hat{y}_{(m,n)}^j\\}$", + "bbox": [ + 544, + 857, + 844, + 877 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, combining the cross-entropy segmentation loss", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3620", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/644a98889183e0ef6b889a9e2199e4b2bb590eeae1aa7f4a56ec1ca5acd5517c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneMethodsExternalTrained on GTAV (G)Trained on Cityscapes (C)
DatasetModuleCBMSBMSG
ResNet50 [13]Baseline [3]--28.9525.1428.1826.2344.9651.6823.2942.55
IBN-Net [38]--33.8532.3037.7527.9048.5657.0426.1445.06
RobustNet [7]--37.3135.2040.3328.3050.7358.6426.2045.00
SiamDoGe [53]--42.9637.5440.6428.3451.5359.0026.6745.08
DIRL [55]-✓41.0439.1541.60-51.80-26.5046.52
WildNet [24]✓-44.6238.4246.0931.3450.9458.7927.9547.01
SANSAW [42]-✓39.7537.3441.8630.7952.9559.8128.3247.28
SPC [20]-✓44.1040.4645.51-----
DPCL [56]-✓44.7440.5946.3330.8150.9758.5925.8546.00
Ours--45.7241.3247.0831.3951.8460.1828.5147.97
ShuffleNetV2 [32]Baseline [3]--25.5622.1728.6023.3336.8443.1321.5636.95
IBN-Net [38]--27.1031.8234.8925.5641.8946.3522.9940.91
RobustNet [7]--30.9832.0635.3124.3141.9446.9722.8240.17
SiamDoGe [53]--34.4034.2335.8721.9542.6147.4823.1340.93
DIRL [55]-✓31.8832.5736.12-42.55-23.7441.23
DPCL [56]-✓36.6634.3539.9222.6643.9048.9522.4741.07
Ours--38.5634.5140.1125.6444.2249.6923.5441.10
", + "bbox": [ + 78, + 87, + 895, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison of mIoU (%) between DGSS methods. External dataset denotes the necessity of an auxiliary dataset during training and External module denotes the requirement of an additional module during inference. G, C, B, M, and S denote GTAV, Cityscapes, BDD100K, Mapillary, and SYNTHIA, respectively. The best and second-best results are bolded and underlined, respectively.", + "bbox": [ + 75, + 375, + 892, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{CE}$ with other loss components, the total is defined as:", + "bbox": [ + 76, + 444, + 447, + 459 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {C E} + \\omega_ {1} \\mathcal {L} _ {C M} + \\omega_ {2} \\mathcal {L} _ {C C} \\tag {9} \\\\ + \\omega_ {3} \\mathcal {L} _ {C W C L} + \\omega_ {4} \\mathcal {L} _ {S D C L} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 468, + 468, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\omega_{1},\\omega_{2},\\omega_{3}$ , and $\\omega_{4}$ denote the weighting factor of each loss functions.", + "bbox": [ + 75, + 512, + 468, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 555, + 200, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we describe the implementation details, the experimental setup for comparison with existing DGSS methods, and the ablation study conducted to further validate the effectiveness of our approach.", + "bbox": [ + 75, + 580, + 468, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 648, + 294, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt DeepLabV3+ [3] for the segmentation architecture and use ResNet-50 [13], ShuffleNetV2 [32], and MobileNetV2 [47] as the backbone network of the segmentation network. The model is trained for 40K iterations with a batch size of 8 using the SGD optimizer, which has a momentum of 0.9 and a weight decay of 5e-4. We employ a polynomial learning rate schedule with an initial rate of 1e-2 and a power of 0.9. For the simulation of domain shift, we augment the image $x_{a}$ using strong color jittering transformation similar to [7]. The weighting parameters of (9), $\\omega_{1}$ , $\\omega_{2}$ , $\\omega_{3}$ and $\\omega_{4}$ , are set as 0.2, 0.2, 0.3, and 0.3 respectively.", + "bbox": [ + 75, + 672, + 468, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Datasets", + "text_level": 1, + "bbox": [ + 76, + 847, + 179, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use two synthetic datasets (GTA [43] and SYNTHIA [45]), and three real-world datasets (Cityscapes [8],", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BDD-100K [58], and Mapillary [33]) for the experiment. All segmentation labels are evaluated based on 19 object categories.", + "bbox": [ + 496, + 444, + 890, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GTAV (G) [43] is a large-scale dataset generated from the Grand Theft Auto V (GTAV) game engine. It comprises 24,966 images, split into 12,403 for training, 6,382 for validation, and 6,181 for testing with a resolution of $1914 \\times 1052$ .", + "bbox": [ + 496, + 489, + 892, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SYNTHIA (S) [45] is a virtual, photo-realistic urban scene dataset comprising 9,400 images with a resolution of $960 \\times 720$ . Among these, 2,820 images are designated for evaluation.", + "bbox": [ + 496, + 565, + 893, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cityscapes (C) [8] is a large-scale urban scene dataset captured from 50 cities, primarily in Germany. Particularly, it contains 5,000 high-resolution images with a resolution of $2048 \\times 1024$ . The dataset is divided into 2,975 images for training, 500 for validation, and 1,525 for testing.", + "bbox": [ + 496, + 626, + 893, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BDD-100K (B) [58] is another real-world urban scene dataset that contains more diverse 10000 urban driving scene images with a resolution of $1280 \\times 720$ . Specifically, the validation split (1,000 images) is used for evaluation.", + "bbox": [ + 496, + 700, + 890, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mapillary (M) [33] contains 25,000 images with a minimum resolution of $1920 \\times 1080$ , collected from various locations worldwide. Specifically, the validation split of 2,000 images is used for evaluation.", + "bbox": [ + 496, + 761, + 890, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with DGSS methods", + "text_level": 1, + "bbox": [ + 498, + 832, + 790, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our methods with other state-of-the-art DGSS methods: Baseline (DeepLabV3+ [3] trained on the source domain), IBN-Net [38], RobustNet [7], SiamDoGe [53],", + "bbox": [ + 496, + 854, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3621", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a8c1edacd555b9aeb544e8b1e27336634ab38752437a712ef806a02175809d01.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsExternal ModuleTrained on GTAV (G)
CBMMean
Baseline [3]25.9425.7326.4526.04
IBN-Net [38]30.1427.6627.0728.29
RobustNet [7]30.8630.0530.6730.52
SiamDoGe [53]34.1534.5032.3433.67
DIRL [55]✓34.6732.7834.3133.92
DPCL [56]✓37.5735.4540.3037.77
Ours37.6636.1040.4038.05
", + "bbox": [ + 78, + 88, + 467, + 214 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DIRL [55], WildNet [24], SANSAW [42], SPC [20], and DPCL [56]. To evaluate the generalization ability of the model on arbitrary unseen domains, we conduct the experiment on two scenarios: i) trained on GTAV, tested on Cityscapes, BDD-100K, and Mapillary, and ii) trained on Cityscapes, tested on BDD-100K, Mapillary, and SYNTHIA. The quantitative results are computed with mean intersection over union (mIoU). Additionally, we compared the method trained on the backbone of ResNet-50 [13], ShuffleNetV2 [32], and MoblieNet [47], pre-trained on ImageNet [9].", + "bbox": [ + 75, + 280, + 467, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative and Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 468, + 336, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 summarizes the quantitative results. Our method outperforms all other methods when trained on GTAV, using ResNet-50 as the backbone. When compared with FN methods that remove domain-specific styles, we demonstrate that our approach minimizes the loss of content information. We also show that our method effectively shows generalization ability when trained on Cityscapes. We further evaluate our methods with different backbones, showing the wide applicability of our method. When trained with ShuffleNetV2, our method achieves the first or second-best performance among unseen target domains. Table 2 shows the results of our method trained on GTAV with MobileNetV2, demonstrating the superiority of our method.", + "bbox": [ + 75, + 492, + 467, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For qualitative evaluation, we compare the visual result between DGSS methods and ours. As depicted in Fig. 5, our method demonstrates superior results compared to other approaches, particularly in its overall prediction accuracy. Notably, our proposed techniques enable distinct prediction of features such as on road and sidewalk, yielding clearer segmentation boundaries. Please refer to the supplementary material for more qualitative results.", + "bbox": [ + 75, + 688, + 467, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computational cost analysis", + "text_level": 1, + "bbox": [ + 76, + 830, + 277, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To confirm that our approach does not incur additional computational overhead, we provide the number of parameters, GFLOPS, and average inference time of each method. As", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cde9f4931872aa504e6d8ec0edfca433b12f3ceb3ea1a2ee335bce871e1abc73.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of mIoU $(\\%)$ using MobileNetV2 [47] backbone trained on the GTAV dataset." + ], + "table_footnote": [], + "table_body": "
MethodsExternal ModuleParams (M)GFLOPSTime (ms)
Baseline [13]45.08277.7710.01
SANSAW [42]✓25.63421.8668.96
SPC [20]✓45.22286.0912.24
DIRL [55]✓45.41278.1111.69
DPCL [56]✓56.461188.64823.78
Ours45.08277.7810.03
", + "bbox": [ + 501, + 88, + 890, + 196 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/31a215e998aaa44615ecc9f72d54f3cf3b11c9da13b663716cf47cf64d5fa051.jpg", + "table_caption": [ + "Table 3. Computational cost comparison conducted using DeepLabV3+ with a ResNet-50 backbone on an NVIDIA Tesla V100 GPU with an image resolution of $2048 \\times 1024$ . Inference time is averaged over 400 trials." + ], + "table_footnote": [], + "table_body": "
LCMCCLCWCLSDCLCBM
28.9525.1428.18
✓38.0836.6540.62
✓✓40.4237.8143.91
✓42.0338.2744.02
✓✓43.1638.5945.38
✓✓✓43.1738.2344.84
✓✓✓✓45.7241.3247.08
", + "bbox": [ + 501, + 277, + 890, + 392 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study on proposed losses. The experiments were conducted using DeepLabV3+ with ResNet-50 backbone, trained on the GTAV dataset. The losses are detailed in $\\mathcal{L}_{CM}$ : (3), $\\mathcal{L}_{CC}$ : (5), $\\mathcal{L}_{CWCL}$ : (7), $\\mathcal{L}_{SDCL}$ : (8)", + "bbox": [ + 496, + 402, + 890, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "detailed in Table 3, our method operates comparably to baseline models by learning features intrinsically without adopting a separate module.", + "bbox": [ + 496, + 487, + 890, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 544, + 663, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we conducted a series of ablation studies to demonstrate the individual contribution and effectiveness of each component within our method. Specifically, we investigate the impact of the following components: $\\mathcal{L}_{CM}$ , $\\mathcal{L}_{CC}$ , $\\mathcal{L}_{CWCL}$ , $\\mathcal{L}_{SDCL}$ . For the study, we use a scenario where the DeepLabV3+ model with backbone ResNet-50 model is trained on GTA and tested on Cityscapes, BDD100K, and Mapillary.", + "bbox": [ + 496, + 566, + 890, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 presents the impact of various proposed losses on domain generalization performance. Specifically, the baseline model, trained solely with cross-entropy loss, exhibits suboptimal performance on target domains because of overfitting to the source domain. Conversely, the integration of any proposed loss mechanisms leads to a marked enhancement in performance. More specifically, the incorporation of the covariance alignment $(\\mathcal{L}_{CM},\\mathcal{L}_{CC})$ shows its efficacy in preserving essential content information by correlating features of paired images. The differential impact of the semantic consistency constrastive learning $(\\mathcal{L}_{CWCL},\\mathcal{L}_{SDCL})$ is also evident, as it significantly aids in disentangling features of similar classes, thereby constructing a more robust embedding space.", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3622", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5743df130013ec2edea812c12d2a7e375fd2bbd88be13afab8d1a8d9272aea9c.jpg", + "image_caption": [ + "Target Data" + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 192, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/337f71e74c8508b424ee737397e719b5030e129644c3e795c1a4ea71ba7d1e6a.jpg", + "image_caption": [ + "Baseline" + ], + "image_footnote": [], + "bbox": [ + 196, + 88, + 310, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/49c2a2eee91c74f79070a020372836d10140ed59f2a9e5dc8645dd7751c88c18.jpg", + "image_caption": [ + "RobustNet" + ], + "image_footnote": [], + "bbox": [ + 312, + 88, + 424, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95e5f7abd34e2a68c19572c4e0fa1674385464d3501bd35dfd4aa3a6bd004b39.jpg", + "image_caption": [ + "WildNet" + ], + "image_footnote": [], + "bbox": [ + 428, + 88, + 542, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ba681e1508b8da213ef145d65f63839dc55f7d2eace8e17b8c6c0de9e132f18f.jpg", + "image_caption": [ + "SPC" + ], + "image_footnote": [], + "bbox": [ + 544, + 88, + 656, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56a57369c29ab1ee852bc71832d6b55c8c7c94a5be7f39e7bf518f526eb5fc04.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 658, + 88, + 772, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e75373901ae793412627ba1d0ec111765d7b497f114899c12c47858de9371ae6.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 776, + 88, + 890, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/97b7de89b1fade3976ea20663122c5a175bbbba4817f004ecaee8fedc9661f42.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison between DGSS methods trained on GTAV (G) and tested on unseen target domains of Cityscapes (C) using DeeplabV3+ with ResNet50 backbone.", + "Figure 6. t-SNE [49] visualization comparing the covariance with and without $\\mathcal{L}_{CML}$ ." + ], + "image_footnote": [ + "Baseline - Original $x$", + "Baseline - Augmented $x_{a}$", + "Ours - Original $x$", + "Ours-Augmented $x_{a}$" + ], + "bbox": [ + 117, + 232, + 310, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Covariance Matching Loss. Fig. 6 presents t-SNE plots of covariances for original and augmented images, before and after the application of CML. The baseline network perceives original and augmented images differently from a style perspective. However, after applying CML, the distribution becomes more intermixed, indicating that our proposed CML effectively ensures similar recognition of different style images.", + "bbox": [ + 75, + 414, + 468, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Calculation of CCL. Table 5a demonstrates that our proposed cross-covariance method, which converges the diagonal components to 1, yields superior performance. As mentioned before, removing non-diagonal components, which contain content information actually degrades performance.", + "bbox": [ + 75, + 536, + 468, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sampling number in CWCL. Table 5b and Table 5c show the impact of varying the number of classes sampled per image and the number of samples per class in CWCL, respectively. As shown in Table 5b, the performance improves with an increase in the diversity of classes sampled in CWCL. This suggests that contrasting a broader array of classes enhances the model's discriminative capability. Furthermore, Table 5c demonstrates that a balanced number of negative samples per class leads to optimal performance.", + "bbox": [ + 75, + 612, + 470, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Projection Head for SDCL. The influence of different project head configurations on the SDCL is investigated. We experimented with three distinct approaches: individual projection head, copying the weights of CWCL's (stop gradient), and shared projection head of CWCL. As demonstrated in Table 5d, sharing the projection head yielded the most superior results. The results indicate that SDCL not only relies on the semantic information from CWCL for effective disentanglement of similar classes but also enhances the embedding space learned by CWCL.", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8690ec98ec23f09059bc8e81f3d41c1140c4b223d47e8a5b7f326677aa559de0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Cross-covariance loss
MethodCBM
Whitening38.6836.9142.12
Diagonal40.4237.8143.91
", + "bbox": [ + 500, + 232, + 709, + 289 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3d78f9782c24c7a5e8b367c88b83157faf680aa15989be82e67c3294f27d90c3.jpg", + "table_caption": [ + "(a)" + ], + "table_footnote": [], + "table_body": "
# of classes
#CBM
1045.5738.8846.37
1545.7241.3247.08
", + "bbox": [ + 712, + 232, + 880, + 289 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/02ede73468dbdebefc060f9a495d9c8eb9276b5f007fc4eb3a4ddbc706d8d9e6.jpg", + "table_caption": [ + "(b)" + ], + "table_footnote": [], + "table_body": "
# of negative samples
#CBM
1044.7638.2146.46
5045.7241.3247.08
10044.4439.1446.29
", + "bbox": [ + 506, + 309, + 669, + 373 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/865e936593eaf7e7c54b4fd622556dd46df755addabcc6363305b953c7327d9d.jpg", + "table_caption": [ + "(c)", + "Projection Head of SDCL" + ], + "table_footnote": [], + "table_body": "
MLPCBM
Individual44.9138.4546.31
Shared (SG)44.0338.1546.64
Shared45.7241.3247.08
", + "bbox": [ + 679, + 321, + 883, + 373 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(d)", + "bbox": [ + 774, + 375, + 789, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation studies. (a) Calculation of $\\mathcal{L}_{CC}$ . (b) Number of classes for $\\mathcal{L}_{CWCL}$ . (c) Number of negative samples for $\\mathcal{L}_{CWCL}$ . (d) Projection head of $\\mathcal{L}_{SDCL}$ . \"SG\" indicates stop gradient.", + "bbox": [ + 498, + 395, + 892, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 462, + 617, + 478 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel BlindNet with covariance alignment and semantic consistency contrastive learning. By introducing covariance alignment, our method effectively addresses style variations, ensuring the extraction of features that are consistent across different styles. Furthermore, with the proposed semantic consistency contrastive learning, we not only facilitate the extraction of discriminative features but also enhance the generalization capabilities of the model in semantic segmentation predictions. Comprehensive experimental results validate the effectiveness of our approach, demonstrating its ability to generalize across multiple unseen target domains without requiring auxiliary domains or additional modules. Our future work will be improving and stabilizing the covariance alignment method.", + "bbox": [ + 496, + 488, + 892, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported in part by the Basic Science Research Program through National Research Foundation of Korea (NRF) (Grants No. NRF2022R1F1A1073543), the MSIT(Ministry of Science and ICT), Korea, under the ICAN(ICT Challenge and Advanced Network of HRD) support program(RS-2022-00156385) supervised by the IITP(Institute for Information & Communications Technology Planning & Evaluation), and Innovative Human Resource Development for Local Intellectualization program through the Institute of Information & Communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT)(IITP-2024-00156287).", + "bbox": [ + 496, + 703, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3623", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Florent Bartoccioni, Éloi Zablocki, Andrei Bursuc, Patrick Pérez, Matthieu Cord, and Karteek Alahari. Lara: Latents and rays for multi-camera bird's-eye-view semantic segmentation. In Conference on Robot Learning, pages 1663-1672. PMLR, 2023. 1", + "[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 3", + "[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1, 6, 7", + "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3, 4", + "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1", + "[6] Wonwooong Cho, Sungha Choi, David Keetae Park, Inkyu Shin, and Jaegul Choo. Image-to-image translation via group-wise deep whitening-and-coloring transformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10639-10647, 2019. 3", + "[7] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11580-11590, 2021. 2, 3, 4, 6, 7", + "[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 6", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 7", + "[10] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2", + "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2", + "[12] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7", + "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3", + "[15] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 1, 2", + "[16] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9924-9935, 2022. 2", + "[17] Lukas Hoyer, Dengxin Dai, Qin Wang, Yuhua Chen, and Luc Van Gool. Improving semi-supervised and domain-adaptive semantic segmentation with self-supervised depth estimation. International Journal of Computer Vision, pages 1-27, 2023. 1", + "[18] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1", + "[19] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6891-6902, 2021. 2", + "[20] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3061-3071, 2023. 6, 7", + "[21] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 3", + "[22] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European conference on computer vision (ECCV), pages 172-189, 2018. 1, 2", + "[23] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 3", + "[24] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3624", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 9936-9946, 2022. 2, 3, 6, 7", + "[25] Guangrui Li, Guoliang Kang, Wu Liu, Yunchao Wei, and Yi Yang. Content-consistent matching for domain adaptive semantic segmentation. In European conference on computer vision, pages 440-456. Springer, 2020. 1", + "[26] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 224-233, 2021. 2", + "[27] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 2", + "[28] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European conference on computer vision (ECCV), pages 624-639, 2018. 2", + "[29] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6936-6945, 2019. 2", + "[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2", + "[31] Ping Luo. Learning deep architectures via generalized whitened neural networks. In International Conference on Machine Learning, pages 2238-2246. PMLR, 2017. 3", + "[32] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European conference on computer vision (ECCV), pages 116-131, 2018. 6, 7", + "[33] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990-4999, 2017. 6", + "[34] David Nilsson, Aleksis Pirinen, Erik Gartner, and Cristian Sminchisescu. Embodied visual active learning for semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2373–2383, 2021. 1", + "[35] Yuya Onozuka, Ryosuke Matsumi, and Motoki Shino. Autonomous mobile robot navigation independent of road boundary using driving recommendation map. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4501-4508. IEEE, 2021. 1", + "[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3, 4", + "[37] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3764-3773, 2020. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464-479, 2018. 2, 3, 6, 7", + "[39] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1863-1871, 2019. 2, 3", + "[40] Taesung Park, Alexei A Efros, Richard Zhang, and JunYan Zhu. Contrastive learning for unpaired image-to-image translation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 319-345. Springer, 2020. 3", + "[41] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE Transactions on Image Processing, 30:6594-6608, 2021. 2, 3", + "[42] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2594-2605, 2022. 2, 4, 6, 7", + "[43] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 102-118. Springer, 2016. 1, 6", + "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 1", + "[45] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 1, 6", + "[46] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10765-10775, 2021. 1", + "[47] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6, 7", + "[48] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 3", + "[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 5, 8" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3625", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. Advances in neural information processing systems, 31, 2018. 2", + "[51] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2517-2526, 2019. 1", + "[52] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7303-7313, 2021. 3", + "[53] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. Siamdoge: Domain generalizable semantic segmentation using siamese network. In European Conference on Computer Vision, pages 603-620. Springer, 2022. 2, 6, 7", + "[54] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 1", + "[55] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2884-2892, 2022. 3, 6, 7", + "[56] Liwei Yang, Xiang Gu, and Jian Sun. Generalized semantic segmentation by self-supervised source domain projection and multi-level contrastive learning. arXiv preprint arXiv:2303.01906, 2023. 3, 6, 7", + "[57] Jaejun Yoo, Youngjung Uh, Sanghyuk Chun, Byeongkyu Kang, and Jung-Woo Ha. Photorealistic style transfer via wavelet transforms. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9036-9045, 2019. 2", + "[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 6", + "[59] Fei Yu, Mo Zhang, Hexin Dong, Sheng Hu, Bin Dong, and Li Zhang. Dast: Unsupervised domain adaptation in semantic segmentation based on discriminator attention and self-training. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10754–10762, 2021. 2", + "[60] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2100-2110, 2019. 2", + "[61] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Do" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "main generalization with mixstyle. In International Conference on Learning Representations, 2020. 2", + "[62] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018. 1, 2" + ], + "bbox": [ + 503, + 92, + 890, + 189 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3626", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/1e340b0f-2934-49e8-9933-238d4fb891ff_model.json b/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/1e340b0f-2934-49e8-9933-238d4fb891ff_model.json new file mode 100644 index 0000000000000000000000000000000000000000..029ff13c8b32f3876e7bf173a613edf4c7c24e3a --- /dev/null +++ b/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/1e340b0f-2934-49e8-9933-238d4fb891ff_model.json @@ -0,0 +1,2565 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.131, + 0.862, + 0.177 + ], + "angle": 0, + "content": "Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.203, + 0.792, + 0.24 + ], + "angle": 0, + "content": "Woo-Jin Ahn1 Geun-Yeong Yang1 Hyun-Duck Choi2* Myo-Taeg Lim1* \n1Korea University 2Chonnam National University" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.241, + 0.751, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\{wjahn, hggofficial, mlim\\} @korea.ac.kr\\) \\(^{2}\\)ducky.choi@jnu.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.473, + 0.764 + ], + "angle": 0, + "content": "Deep learning models for semantic segmentation often experience performance degradation when deployed to unseen target domains unidentified during the training phase. This is mainly due to variations in image texture (i.e. style) from different data sources. To tackle this challenge, existing domain generalized semantic segmentation (DGSS) methods attempt to remove style variations from the feature. However, these approaches struggle with the entanglement of style and content, which may lead to the unintentional removal of crucial content information, causing performance degradation. This study addresses this limitation by proposing BlindNet, a novel DGSS approach that blinds the style without external modules or datasets. The main idea behind our proposed approach is to alleviate the effect of style in the encoder whilst facilitating robust segmentation in the decoder. To achieve this, BlindNet comprises two key components: covariance alignment and semantic consistency contrastive learning. Specifically, the covariance alignment trains the encoder to uniformly recognize various styles and preserve the content information of the feature, rather than removing the style-sensitive factor. Meanwhile, semantic consistency contrastive learning enables the decoder to construct discriminative class embedding space and disentangles features that are vulnerable to misclassification. Through extensive experiments, our approach outperforms existing DGSS methods, exhibiting robustness and superior performance for semantic segmentation on unseen target domains. The code is available at https://github.com/root0yang/BlindNet." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.806, + 0.21, + 0.822 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.832, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Semantic segmentation, a technique that classifies each pixel in an image into predefined categories, has garnered significant attention due to its potential applications in vari-" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.289, + 0.895, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.422, + 0.895, + 0.48 + ], + "angle": 0, + "content": "Figure 1. Comparison of semantic segmentation results between the baseline (DeepLabV3+ with ResNet50 backbone) and our BlindNet. Both models are trained on the source domain (GTAV [43]) and tested on the target domain (Cityscapes [8])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.895, + 0.809 + ], + "angle": 0, + "content": "ous fields. Particularly, it plays a crucial role in autonomous driving [1, 18] and robotic systems [34, 35]. Besides, with the advent of large datasets, deep neural networks have emerged as a trending approach for semantic segmentation tasks, achieving impressive results [3, 5, 44, 54]. However, there remains a major bottleneck detailing the meticulous and labor-intensive process of dataset labeling. More particularly, this process not only consumes time but also poses economic challenges [8, 46]. To address this challenge, synthetic datasets have emerged as a compelling alternative. Specifically, these datasets, generated using three-dimensional (3D) rendering techniques, offer vast amounts of easily accessible data, eliminating the need for manual labeling [43, 45]. However, a challenge arises when models trained on synthetic datasets are deployed in real-world scenarios. More precisely, a domain shift problem arises due to style factor discrepancies (e.g. texture, illumination, and image quality) between synthetic and real-world data, which affects the performance of the model, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.902 + ], + "angle": 0, + "content": "To address the domain shift problem, domain adaptation semantic segmentation (DASS) has been introduced [15, 17, 22, 25, 51, 62]. Specifically, DASS aims to bridge the gap between source and target domains by aligning their data distributions. However, a significant limitation of DASS is its dependency on the target domain during train-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.23, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Authors" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3616" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "ing. For DASS to function effectively, target domain samples must be available during the training phase." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.124, + 0.468, + 0.228 + ], + "angle": 0, + "content": "Meanwhile, domain generalized semantic segmentation (DGSS) has been introduced as an alternative approach to tackle the domain shift problem. Unlike DASS, DGSS is trained only with the source domain, aiming to extract domain-invariant features. To achieve this, two main techniques have been employed: domain randomization (DR) and feature normalization (FN)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.468, + 0.366 + ], + "angle": 0, + "content": "DR augments the training set by introducing variability, either by altering the image style [41, 60] or by modifying the feature representation [24, 53]. By exposing the model to a wider variety of styles via DR, the network is less likely to overfit to the specific styles present in the training data. Consequently, the robustness of the model is improved, making it more adept at generalizing to new, unseen domains. Nevertheless, a crucial limitation of DR is its significant dependence on auxiliary domains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.468, + 0.534 + ], + "angle": 0, + "content": "FN methods, converse to DR, regularize the features to prevent the model from overfitting to the distinct styles or characteristics of the training data. This is achieved by removing domain-specific style information using feature statistics, such as instance normalization [38] or whitening transformation [7, 19, 39, 42]. While these approaches effectively remove style-related information, they simultaneously pose the challenge of removing semantic content because content and style information are entangled. Consequently, the model fails to capture the essential patterns or features required for accurate segmentation prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.468, + 0.867 + ], + "angle": 0, + "content": "To address this problem, we propose BlindNet, a model that blinds the style within the encoder and improves the robustness of the decoder, without requiring auxiliary datasets or external modules. Specifically, our proposed BlindNet consists of two components: covariance alignment for the encoder and semantic consistency contrastive learning for the decoder. Precisely, the covariance alignment facilitates the generation of style-invariant features with the proposed covariance matching loss function (CML) and the cross-covariance loss function (CCL). Specifically, CML mitigates the effects of style variations, while CCL focuses on preserving content information, effectively addressing the prevalent content information loss observed in the FN method. To further improve the generalization ability, we develop semantic consistency contrastive learning, which consists of class-wise contrastive learning (CWCL) and semantic disentanglement contrastive learning (SDCL). Particularly, the CWCL constructs a discriminative class embedding space, while SDCL disentangles features of similar classes that often lead to prediction errors. Extensive experiments across various datasets demonstrate that the proposed BlindNet outperforms existing DGSS methods." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.869, + 0.399, + 0.883 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- We propose a covariance alignment within the encoder," + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "comprising CML and CCL. Specifically, the CML aims to mitigate the effects of style variations, while CCL ensures the preservation of content information, together facilitating the generation of style-agnostic features." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.153, + 0.892, + 0.241 + ], + "angle": 0, + "content": "- We propose semantic consistency contrastive learning within the decoder that comprises CWCL and SDCL, utilizing segmentation masks. Specifically, CWCL generates discriminative embeddings, whereas SDCL disentangles features of similar classes, enhancing the robustness of the model." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.243, + 0.892, + 0.287 + ], + "angle": 0, + "content": "- Through extensive experiments, we demonstrate the superiority of our approach in DGSS, without the need to alter the network architecture or rely on external datasets." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.153, + 0.892, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.302, + 0.65, + 0.317 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.327, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Domain adaptation and generalization for semantic segmentation. Domain adaptation (DA) aims at minimizing the distribution discrepancy between different domains, enabling a model to generalize from a source to a target domain. For DASS, adversarial training and cross-domain self-training strategies are commonly used. Particularly, adversarial-based methods [15, 29] employ generative adversarial networks [11] to close the feature distribution gap between source and target domains. Meanwhile, cross-domain self-training methods [16, 37, 59, 62] generate pseudo-labels for target domain data using pre-trained models, and employ them as training data, thereby expanding the training data and reducing distribution differences between the source and target domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Domain generalization (DG) methods [26, 28, 30, 50, 61] aim to improve the generalization ability of the model without accessing the target domain during training. Since the difference in style of the image is the main cause of the disparity, most existing domain generalized semantic segmentation methods utilize the style information of the image for domain-invariant learning. Feature statistics (e.g. mean, variance, covariance, gram matrix, etc.) which are commonly used in style transfer [10, 21, 27, 57] are employed to capture the style information. Interestingly, existing DGSS methods can be separated into two parts: i) domain randomization to expand the distribution of style or ii) feature normalization to remove style." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "DR involves randomizing either the image or its features through stylization to learn domain-invariant features from various styles. For example, Peng et al. [42] extended the source domain data by stylizing images in the style of unreal paintings. Similarly, Yue et al. [60] and Huang et al. [22] attempted to enhance generalization by synthesizing images with diverse styles in the image space. In another study, Lee et al. [24] adopted ImageNet data [9] as wild data and performed randomization via synthesis in the feature space. Meanwhile, Wu et al. [53] diversified the trainable feature space by mixing the statistics of the feature and its color-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3617" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.3, + 0.106 + ], + "angle": 0, + "content": "jittered feature with Ada-IN [21]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.111, + 0.473, + 0.489 + ], + "angle": 0, + "content": "FN methods aim to remove domain-specific styles from features, extracting only domain-invariant content. For instance, Pan et al. [39] first attempted the DGSS method, combining batch normalization (BN) [23] and instance normalization (IN) [48] in the network layer. While BN preserves the content information within discriminative features, IN focuses on removing domain-specific style information from features. In a study, Choi et al. [7] addressed the limitations of previous whitening transformation [6, 31] that can eliminate the content information from the feature. Specifically, they proposed an instance-selective whitening approach designed to remove covariance components that are sensitive to domain shifts. Peng et al. [41] developed semantic-aware normalization that performs on class-wise and semantic-aware whitening that aligns channels based on the prediction through group whitening transformation [6]. In a study, Xu et al. [55] introduced the prior guided attention module and guided feature whitening to re-calibrate the feature and remove domain-specific style effects, respectively. Unlike the FN methods that directly remove the style component, our work explores a covariance alignment method that mitigates the effect of the style's effect while preserving the content information. Our method achieves the DGSS without any additional modules or auxiliary datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Contrastive Learning. Contrastive learning aims to learn representations by maximizing the similarity between positive pairs of samples while minimizing the similarity between negative pairs. In recent years, it has attracted significant attention for its effectiveness in learning discriminative representations across various tasks [2, 4, 12, 14]. Oord et al. [36] were the first to introduce the infoNCE loss, a type of contrastive loss function designed for self-contrastive learning. In a work, Park et al. [40] introduced patch-level contrastive learning for image translation, using co-located patches as positive pairs and spatially distant patches as negatives to maintain image context. For the semantic segmentation task, Wang et al. [52] introduced class-wise contrastive learning to aid the model in learning the embedding space of each class. Specifically, they sampled the classes existing in the images and applied contrastive learning based on the class label. For DGSS, Lee et al. [24] adopted contrastive learning to learn the ImageNet information in their model. Specifically, they set the ImageNet data as wild and applied the contrastive learning method by setting the wild-stylized feature and its closest wild content as positive samples. In another study, Yang et al. [56] developed multi-level contrastive learning, which designed instance prototypes and class prototypes for contrastive learning. Specifically, they sample each class's pixel features and apply contrastive learning with a transition-probability ability matrix. Unlike recent DGSS works that embed the origi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.152 + ], + "angle": 0, + "content": "nal image, we define contrastive learning for the augmented image. Specifically, the proposed method builds a robust embedding space by preserving the semantic consistency of the feature representation across various domains." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.165, + 0.593, + 0.181 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.191, + 0.895, + 0.356 + ], + "angle": 0, + "content": "The goal of the proposed method is to train a segmentation model \\(\\varphi\\) on a given source domain \\(S\\) and generalize well to the unseen target domain \\(T\\). Precisely, the source domain \\(S = \\{(x,y)\\}\\) contains a paired image \\(x \\in \\mathbb{R}^{H \\times W \\times 3}\\) and segmentation label \\(y \\in \\mathbb{R}^{H \\times W \\times C}\\), where \\(H\\), \\(W\\), and \\(C\\) denote the height of the image, the width of the image, and the class number of the segmentation map, respectively. The model \\(\\varphi\\) takes an original \\(x\\) and its augmented counterpart \\(x_{a}\\), which have the same content but different styles, and uses the feature information to enhance its generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.895, + 0.479 + ], + "angle": 0, + "content": "As shown in Fig. 2, our method leverages the feature information through two main approaches: covariance alignment and semantic consistency contrastive learning. Specifically, the covariance matching ensures that features, having different styles but the same content, contain similar information. Additionally, the semantic consistency contrastive learning embeds the generalized features into discriminative representation based on the segmentation label." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.487, + 0.712, + 0.503 + ], + "angle": 0, + "content": "3.1. Covariance Alignment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.894, + 0.631 + ], + "angle": 0, + "content": "The domain shift in semantic segmentation results from changes in the visual characteristics of the image, known as style. This style information is typically detected in the shallow layers of networks [38], which are the encoders of the segmentation model. Based on this understanding, our method targets the encoder to handle style variations by employing the proposed covariance matching loss and cross-covariance loss function." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.649, + 0.693, + 0.665 + ], + "angle": 0, + "content": "Covariance Matching Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.895, + 0.856 + ], + "angle": 0, + "content": "To train the network to uniformly recognize various styles without removing content information, we introduce the CML. Specifically, the loss aims to minimize the difference between covariance matrices derived from different styles of image features. Given an image pair \\((x,x_{a})\\), the features from \\(i^{th}\\) block of the encoder are represented as \\(F^i\\in \\mathbb{R}^{(H^i\\times W^i)\\times C^i}\\) and \\(F_{a}^{i}\\in \\mathbb{R}^{(H^{i}\\times W^{i})\\times C^{i}}\\), respectively. Further, following the methodologies of [7, 41], we compute the covariance matrices using instance normalized features [48], which ensures consistent scaling across features. The feature maps are normalized and flattened into \\(\\bar{F}\\) and \\(\\bar{F}_a\\in \\mathbb{R}^{(HW)\\times C}\\), which are given by:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.865, + 0.892, + 0.9 + ], + "angle": 0, + "content": "\\[\n\\bar {F} = \\frac {(F - \\mu (F))}{\\sigma (F)}, \\bar {F} _ {a} = \\frac {(F _ {a} - \\mu (F _ {a}))}{\\sigma (F _ {a})}, \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3618" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.891, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.414, + 0.893, + 0.456 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed BlindNet. The network processes a pair of images - the original image \\( x \\) and its augmented counterpart \\( x_{a} \\). It employs covariance alignment to treat encoder features and utilizes semantic consistency contrastive learning for the processing of decoder features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.47, + 0.545 + ], + "angle": 0, + "content": "where \\(\\mu (\\cdot)\\in \\mathbb{R}^{C}\\) and \\(\\sigma (\\cdot)\\in \\mathbb{R}^C\\) denote the mean and standard deviation of the features. Utilizing the normalized features, we evaluate the covariance matrices for the original and augmented image features as:" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.554, + 0.47, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\Sigma_ {x, x} ^ {i} = \\bar {F} ^ {i} ^ {T} \\cdot \\bar {F} ^ {i}, \\quad \\Sigma_ {x _ {a}, x _ {a}} ^ {i} = \\bar {F} _ {a} ^ {i} ^ {T} \\cdot \\bar {F} _ {a} ^ {i}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.47, + 0.632 + ], + "angle": 0, + "content": "The CML is then formulated to align these covariance matrices, ensuring that the network maintains consistency in the presence of style variations. The CML is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.644, + 0.469, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C M} = \\sum_ {i = 1} ^ {n _ {e}} \\| \\Sigma_ {x, x} ^ {i} - \\Sigma_ {x _ {a}, x _ {a}} ^ {i} \\| _ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.693, + 0.438, + 0.707 + ], + "angle": 0, + "content": "where \\( n_e \\) denotes the number of blocks of the encoder." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.238, + 0.74 + ], + "angle": 0, + "content": "Cross-covariance Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.47, + 0.87 + ], + "angle": 0, + "content": "While the CML effectively aligns the internal distributions of features within the same image, it does not fully account for the direct correlations across paired images. To complement this, we introduce CCL, which aims to encode the consistent content information of an image pair \\((x,x_{a})\\) by utilizing the cross-covariance of the image pair. Given the normalized feature pair \\((\\bar{F}^i,\\bar{F}_a^i)\\), the cross-covariance of the feature pair can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.881, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\Sigma_ {x, x _ {a}} ^ {i} = \\bar {F} ^ {i} ^ {T} \\cdot \\bar {F} _ {a} ^ {i}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.483, + 0.892, + 0.588 + ], + "angle": 0, + "content": "The cross-covariance is expected to exhibit an identity matrix, as the feature pair should contain identical information. Nonetheless, the proposed CCL converges only the diagonal component of the covariance matrix to one. This is to prevent the drawbacks of the existing FN methods [7, 42] from removing content information. The CCL function is thus defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.593, + 0.891, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C C} = \\sum_ {i = 1} ^ {n _ {e}} \\| \\operatorname {d i a g} \\left(\\Sigma_ {x, x _ {a}} ^ {i}\\right) - \\mathbb {1} \\| _ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.639, + 0.891, + 0.682 + ], + "angle": 0, + "content": "where \\(\\mathrm{diag}(\\Sigma_c) \\in \\mathbb{R}^C\\) denotes the column vector comprising diagonal elements of \\(\\Sigma_{x,x_a}^i\\) and \\(\\mathbb{1} \\in \\mathbb{R}^C\\) denotes the one vector." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.692, + 0.871, + 0.708 + ], + "angle": 0, + "content": "3.2. Semantic Consistence Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.715, + 0.892, + 0.82 + ], + "angle": 0, + "content": "While the encoder focuses on generating style-blinded features, the decoder aims to improve the robustness of the segmentation prediction against domain shifts. For the decoder, we employ a contrastive learning approach, which has demonstrated effectiveness in extracting discriminative features [4]. Specifically, we utilize the InfoNCE loss [36], which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.824, + 0.891, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I N} (a, p, n) = - \\log \\left(\\frac {e ^ {(a \\cdot p / \\tau)}}{e ^ {(a \\cdot p / \\tau)} + \\sum_ {n} ^ {N ^ {-}} e ^ {(a \\cdot n / \\tau)}}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "where \\(a, p, n\\), and \\(N^{-}\\) denote anchor, positive sample, negative sample, and negative sample set, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3619" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.087, + 0.873, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.893, + 0.298 + ], + "angle": 0, + "content": "Figure 3. Illustration of semantic consistency contrastive learning: The mask \\(M\\) represents the error mask derived from the augmented segmentation map. CWCL conducts contrastive learning by sampling per segmentation class and SDCL conducts contrastive learning based on the \\(M\\). Both methods share a projection head \\(\\pi\\) for the semantic representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.471, + 0.46 + ], + "angle": 0, + "content": "To achieve consistent feature representation in DGSS across various styles, we introduce semantic consistency contrastive learning. Specifically, the anchor is derived from the augmented image \\( x_{a} \\), while the positive sample is extracted from the corresponding pixel of the original image \\( x \\). Our method consists of two main components based on the negative sample as shown in Fig. 3: class-wise contrastive learning and semantic disentanglement contrastive learning." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.492, + 0.31, + 0.508 + ], + "angle": 0, + "content": "Class-wise Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.471, + 0.744 + ], + "angle": 0, + "content": "Our CWCL aims to build a discriminative embedding space for each segmentation class using different classes of the original image as negatives. Given an image pair \\((x, x_{a})\\), the features from the \\(j^{th}\\) block of the decoder at pixel position \\((m, n)\\) are denoted as \\(F_{(m, n)}^{j}\\) and \\(F_{a, (m, n)}^{j} \\in \\mathbb{R}^{(1 \\times 1) \\times C^{j}}\\), where \\(C^{j}\\) indicates the channel length of the feature. As mentioned above, we take \\(F_{a, (m, n)}^{j}\\) as the anchor and \\(F_{(m, n)}^{j}\\) as the positive sample, since they represent the same content at the corresponding spatial location. To obtain the negative samples from \\(F^{j}\\), we leverage the resized segmentation class label \\(y^{j} \\in \\mathbb{R}^{(H^{j} \\times W^{j}) \\times C}\\) to collect the different class samples. The samples are passed through the projection head, denoted as \\(\\pi\\), resulting in the projected features \\(\\tilde{F}\\) and \\(\\tilde{F}_{a}\\). We define our CWCL as:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.759, + 0.47, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C W C L} = \\sum_ {j} ^ {n _ {d}} \\mathcal {L} _ {I N} \\left(\\tilde {F} _ {a, (m, n)} ^ {j}, \\tilde {F} _ {(m, n)} ^ {j}, \\tilde {F} _ {(p, q)} ^ {j}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.803, + 0.43, + 0.825 + ], + "angle": 0, + "content": "where \\((p,q)\\in \\{(p,q)\\in P|y_{(p,q)}^j\\neq y_{(m,n)}^j\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.9 + ], + "angle": 0, + "content": "The set \\( P \\) represents all pixel positions in feature \\( F^j \\), with dimensions \\( H^j \\times W^j \\) corresponding to the height and width. Additionally, \\( n_d \\) denotes the total number of blocks in the decoder." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.319, + 0.695, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.33, + 0.853, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.441, + 0.893, + 0.497 + ], + "angle": 0, + "content": "Figure 4. t-SNE [49] visualization comparing scenarios with and without \\(\\mathcal{L}_{SCCL}\\). In (b), the application of SCCL results in a clear separation between the sidewalk (pink), the road (purple), and the building (gray)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.523, + 0.844, + 0.539 + ], + "angle": 0, + "content": "Semantic Disentanglement Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.804 + ], + "angle": 0, + "content": "Domain shifts can lead to the entanglement of similar classes, causing the model to misclassify, as illustrated in Fig. 4. To mitigate this issue, we introduce the SDCL, specifically designed to disentangle the feature \\( x_{a} \\) that has been misclassified, making it closer to the correct class and far from the misclassified class to achieve effective disentanglement. To further ensure consistent feature space and capture the semantic meaning, we share the projection head \\( \\pi \\) used in the CWCL loss. Given the predicted segmentation map of the augmented image, represented as \\( \\hat{y}_a = \\varphi (x_a) \\), we resize it to \\( \\hat{y}_{a,(m,n)}^j\\in \\mathbb{R}^{(H^j\\times W^j)\\times C} \\). Similarly, \\( y_{(m,n)}^j \\) represents the ground truth segmentation map. Using these segmentation maps, we set the anchor at positions where \\( \\hat{y}_{a,(m,n)}^j\\neq y_{(m,n)}^j \\). Negative samples are selected from the augmented image features corresponding to the anchor's misclassified class. The samples go through the projection head \\( \\pi \\). Our SDCL loss is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.813, + 0.892, + 0.854 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S D C L} = \\sum_ {j} ^ {n _ {d}} \\mathcal {L} _ {I N} \\left(\\tilde {F} _ {a, (m, n)} ^ {j}, \\tilde {F} _ {(m, n)} ^ {j}, \\tilde {F} _ {a, (r, s)} ^ {j}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.545, + 0.858, + 0.846, + 0.878 + ], + "angle": 0, + "content": "where \\((r,s)\\in \\{(r,s)\\in P|y_{(r,s)}^j = \\hat{y}_{(m,n)}^j\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Finally, combining the cross-entropy segmentation loss" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3620" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.088, + 0.896, + 0.367 + ], + "angle": 0, + "content": "
BackboneMethodsExternalTrained on GTAV (G)Trained on Cityscapes (C)
DatasetModuleCBMSBMSG
ResNet50 [13]Baseline [3]--28.9525.1428.1826.2344.9651.6823.2942.55
IBN-Net [38]--33.8532.3037.7527.9048.5657.0426.1445.06
RobustNet [7]--37.3135.2040.3328.3050.7358.6426.2045.00
SiamDoGe [53]--42.9637.5440.6428.3451.5359.0026.6745.08
DIRL [55]-✓41.0439.1541.60-51.80-26.5046.52
WildNet [24]✓-44.6238.4246.0931.3450.9458.7927.9547.01
SANSAW [42]-✓39.7537.3441.8630.7952.9559.8128.3247.28
SPC [20]-✓44.1040.4645.51-----
DPCL [56]-✓44.7440.5946.3330.8150.9758.5925.8546.00
Ours--45.7241.3247.0831.3951.8460.1828.5147.97
ShuffleNetV2 [32]Baseline [3]--25.5622.1728.6023.3336.8443.1321.5636.95
IBN-Net [38]--27.1031.8234.8925.5641.8946.3522.9940.91
RobustNet [7]--30.9832.0635.3124.3141.9446.9722.8240.17
SiamDoGe [53]--34.4034.2335.8721.9542.6147.4823.1340.93
DIRL [55]-✓31.8832.5736.12-42.55-23.7441.23
DPCL [56]-✓36.6634.3539.9222.6643.9048.9522.4741.07
Ours--38.5634.5140.1125.6444.2249.6923.5441.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.376, + 0.893, + 0.419 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of mIoU (%) between DGSS methods. External dataset denotes the necessity of an auxiliary dataset during training and External module denotes the requirement of an additional module during inference. G, C, B, M, and S denote GTAV, Cityscapes, BDD100K, Mapillary, and SYNTHIA, respectively. The best and second-best results are bolded and underlined, respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.445, + 0.449, + 0.46 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{CE}\\) with other loss components, the total is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.469, + 0.469, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {C E} + \\omega_ {1} \\mathcal {L} _ {C M} + \\omega_ {2} \\mathcal {L} _ {C C} \\tag {9} \\\\ + \\omega_ {3} \\mathcal {L} _ {C W C L} + \\omega_ {4} \\mathcal {L} _ {S D C L} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.47, + 0.543 + ], + "angle": 0, + "content": "where \\(\\omega_{1},\\omega_{2},\\omega_{3}\\) , and \\(\\omega_{4}\\) denote the weighting factor of each loss functions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.556, + 0.202, + 0.573 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.469, + 0.643 + ], + "angle": 0, + "content": "In this section, we describe the implementation details, the experimental setup for comparison with existing DGSS methods, and the ablation study conducted to further validate the effectiveness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.295, + 0.667 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.84 + ], + "angle": 0, + "content": "We adopt DeepLabV3+ [3] for the segmentation architecture and use ResNet-50 [13], ShuffleNetV2 [32], and MobileNetV2 [47] as the backbone network of the segmentation network. The model is trained for 40K iterations with a batch size of 8 using the SGD optimizer, which has a momentum of 0.9 and a weight decay of 5e-4. We employ a polynomial learning rate schedule with an initial rate of 1e-2 and a power of 0.9. For the simulation of domain shift, we augment the image \\( x_{a} \\) using strong color jittering transformation similar to [7]. The weighting parameters of (9), \\( \\omega_{1} \\), \\( \\omega_{2} \\), \\( \\omega_{3} \\) and \\( \\omega_{4} \\), are set as 0.2, 0.2, 0.3, and 0.3 respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.18, + 0.862 + ], + "angle": 0, + "content": "4.2. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We use two synthetic datasets (GTA [43] and SYNTHIA [45]), and three real-world datasets (Cityscapes [8]," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.49 + ], + "angle": 0, + "content": "BDD-100K [58], and Mapillary [33]) for the experiment. All segmentation labels are evaluated based on 19 object categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.893, + 0.564 + ], + "angle": 0, + "content": "GTAV (G) [43] is a large-scale dataset generated from the Grand Theft Auto V (GTAV) game engine. It comprises 24,966 images, split into 12,403 for training, 6,382 for validation, and 6,181 for testing with a resolution of \\(1914 \\times 1052\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.894, + 0.625 + ], + "angle": 0, + "content": "SYNTHIA (S) [45] is a virtual, photo-realistic urban scene dataset comprising 9,400 images with a resolution of \\(960 \\times 720\\). Among these, 2,820 images are designated for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.894, + 0.701 + ], + "angle": 0, + "content": "Cityscapes (C) [8] is a large-scale urban scene dataset captured from 50 cities, primarily in Germany. Particularly, it contains 5,000 high-resolution images with a resolution of \\(2048 \\times 1024\\). The dataset is divided into 2,975 images for training, 500 for validation, and 1,525 for testing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.762 + ], + "angle": 0, + "content": "BDD-100K (B) [58] is another real-world urban scene dataset that contains more diverse 10000 urban driving scene images with a resolution of \\(1280 \\times 720\\). Specifically, the validation split (1,000 images) is used for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Mapillary (M) [33] contains 25,000 images with a minimum resolution of \\(1920 \\times 1080\\), collected from various locations worldwide. Specifically, the validation split of 2,000 images is used for evaluation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.833, + 0.792, + 0.849 + ], + "angle": 0, + "content": "4.3. Comparison with DGSS methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.902 + ], + "angle": 0, + "content": "We compare our methods with other state-of-the-art DGSS methods: Baseline (DeepLabV3+ [3] trained on the source domain), IBN-Net [38], RobustNet [7], SiamDoGe [53]," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3621" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.468, + 0.215 + ], + "angle": 0, + "content": "
MethodsExternal ModuleTrained on GTAV (G)
CBMMean
Baseline [3]25.9425.7326.4526.04
IBN-Net [38]30.1427.6627.0728.29
RobustNet [7]30.8630.0530.6730.52
SiamDoGe [53]34.1534.5032.3433.67
DIRL [55]✓34.6732.7834.3133.92
DPCL [56]✓37.5735.4540.3037.77
Ours37.6636.1040.4038.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.226, + 0.468, + 0.253 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of mIoU \\((\\%)\\) using MobileNetV2 [47] backbone trained on the GTAV dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.281, + 0.468, + 0.448 + ], + "angle": 0, + "content": "DIRL [55], WildNet [24], SANSAW [42], SPC [20], and DPCL [56]. To evaluate the generalization ability of the model on arbitrary unseen domains, we conduct the experiment on two scenarios: i) trained on GTAV, tested on Cityscapes, BDD-100K, and Mapillary, and ii) trained on Cityscapes, tested on BDD-100K, Mapillary, and SYNTHIA. The quantitative results are computed with mean intersection over union (mIoU). Additionally, we compared the method trained on the backbone of ResNet-50 [13], ShuffleNetV2 [32], and MoblieNet [47], pre-trained on ImageNet [9]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.469, + 0.338, + 0.484 + ], + "angle": 0, + "content": "Quantitative and Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.468, + 0.688 + ], + "angle": 0, + "content": "Table 1 summarizes the quantitative results. Our method outperforms all other methods when trained on GTAV, using ResNet-50 as the backbone. When compared with FN methods that remove domain-specific styles, we demonstrate that our approach minimizes the loss of content information. We also show that our method effectively shows generalization ability when trained on Cityscapes. We further evaluate our methods with different backbones, showing the wide applicability of our method. When trained with ShuffleNetV2, our method achieves the first or second-best performance among unseen target domains. Table 2 shows the results of our method trained on GTAV with MobileNetV2, demonstrating the superiority of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.468, + 0.81 + ], + "angle": 0, + "content": "For qualitative evaluation, we compare the visual result between DGSS methods and ours. As depicted in Fig. 5, our method demonstrates superior results compared to other approaches, particularly in its overall prediction accuracy. Notably, our proposed techniques enable distinct prediction of features such as on road and sidewalk, yielding clearer segmentation boundaries. Please refer to the supplementary material for more qualitative results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.279, + 0.846 + ], + "angle": 0, + "content": "Computational cost analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To confirm that our approach does not incur additional computational overhead, we provide the number of parameters, GFLOPS, and average inference time of each method. As" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.089, + 0.892, + 0.197 + ], + "angle": 0, + "content": "
MethodsExternal ModuleParams (M)GFLOPSTime (ms)
Baseline [13]45.08277.7710.01
SANSAW [42]✓25.63421.8668.96
SPC [20]✓45.22286.0912.24
DIRL [55]✓45.41278.1111.69
DPCL [56]✓56.461188.64823.78
Ours45.08277.7810.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.208, + 0.892, + 0.263 + ], + "angle": 0, + "content": "Table 3. Computational cost comparison conducted using DeepLabV3+ with a ResNet-50 backbone on an NVIDIA Tesla V100 GPU with an image resolution of \\(2048 \\times 1024\\). Inference time is averaged over 400 trials." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.393 + ], + "angle": 0, + "content": "
LCMCCLCWCLSDCLCBM
28.9525.1428.18
✓38.0836.6540.62
✓✓40.4237.8143.91
✓42.0338.2744.02
✓✓43.1638.5945.38
✓✓✓43.1738.2344.84
✓✓✓✓45.7241.3247.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.404, + 0.892, + 0.46 + ], + "angle": 0, + "content": "Table 4. Ablation study on proposed losses. The experiments were conducted using DeepLabV3+ with ResNet-50 backbone, trained on the GTAV dataset. The losses are detailed in \\(\\mathcal{L}_{CM}\\): (3), \\(\\mathcal{L}_{CC}\\): (5), \\(\\mathcal{L}_{CWCL}\\): (7), \\(\\mathcal{L}_{SDCL}\\): (8)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.891, + 0.533 + ], + "angle": 0, + "content": "detailed in Table 3, our method operates comparably to baseline models by learning features intrinsically without adopting a separate module." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.665, + 0.559 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.689 + ], + "angle": 0, + "content": "In this subsection, we conducted a series of ablation studies to demonstrate the individual contribution and effectiveness of each component within our method. Specifically, we investigate the impact of the following components: \\(\\mathcal{L}_{CM}\\), \\(\\mathcal{L}_{CC}\\), \\(\\mathcal{L}_{CWCL}\\), \\(\\mathcal{L}_{SDCL}\\). For the study, we use a scenario where the DeepLabV3+ model with backbone ResNet-50 model is trained on GTA and tested on Cityscapes, BDD100K, and Mapillary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Table 4 presents the impact of various proposed losses on domain generalization performance. Specifically, the baseline model, trained solely with cross-entropy loss, exhibits suboptimal performance on target domains because of overfitting to the source domain. Conversely, the integration of any proposed loss mechanisms leads to a marked enhancement in performance. More specifically, the incorporation of the covariance alignment \\((\\mathcal{L}_{CM},\\mathcal{L}_{CC})\\) shows its efficacy in preserving essential content information by correlating features of paired images. The differential impact of the semantic consistency constrastive learning \\((\\mathcal{L}_{CWCL},\\mathcal{L}_{SDCL})\\) is also evident, as it significantly aids in disentangling features of similar classes, thereby constructing a more robust embedding space." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3622" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.194, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.157, + 0.169, + 0.169 + ], + "angle": 0, + "content": "Target Data" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.089, + 0.311, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.157, + 0.277, + 0.169 + ], + "angle": 0, + "content": "Baseline" + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.089, + 0.426, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.157, + 0.399, + 0.169 + ], + "angle": 0, + "content": "RobustNet" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.089, + 0.543, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.157, + 0.509, + 0.169 + ], + "angle": 0, + "content": "WildNet" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.089, + 0.658, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.157, + 0.614, + 0.169 + ], + "angle": 0, + "content": "SPC" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.089, + 0.774, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.703, + 0.157, + 0.731, + 0.168 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.089, + 0.891, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.796, + 0.157, + 0.871, + 0.169 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.181, + 0.893, + 0.209 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison between DGSS methods trained on GTAV (G) and tested on unseen target domains of Cityscapes (C) using DeeplabV3+ with ResNet50 backbone." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.233, + 0.311, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.315, + 0.253, + 0.414, + 0.264 + ], + "angle": 0, + "content": "Baseline - Original \\(x\\)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.315, + 0.272, + 0.43, + 0.284 + ], + "angle": 0, + "content": "Baseline - Augmented \\(x_{a}\\)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.315, + 0.299, + 0.398, + 0.31 + ], + "angle": 0, + "content": "Ours - Original \\(x\\)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.315, + 0.318, + 0.415, + 0.328 + ], + "angle": 0, + "content": "Ours-Augmented \\(x_{a}\\)" + }, + { + "type": "list", + "bbox": [ + 0.315, + 0.253, + 0.43, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.359, + 0.47, + 0.387 + ], + "angle": 0, + "content": "Figure 6. t-SNE [49] visualization comparing the covariance with and without \\(\\mathcal{L}_{CML}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.469, + 0.536 + ], + "angle": 0, + "content": "Covariance Matching Loss. Fig. 6 presents t-SNE plots of covariances for original and augmented images, before and after the application of CML. The baseline network perceives original and augmented images differently from a style perspective. However, after applying CML, the distribution becomes more intermixed, indicating that our proposed CML effectively ensures similar recognition of different style images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.612 + ], + "angle": 0, + "content": "Calculation of CCL. Table 5a demonstrates that our proposed cross-covariance method, which converges the diagonal components to 1, yields superior performance. As mentioned before, removing non-diagonal components, which contain content information actually degrades performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.471, + 0.749 + ], + "angle": 0, + "content": "Sampling number in CWCL. Table 5b and Table 5c show the impact of varying the number of classes sampled per image and the number of samples per class in CWCL, respectively. As shown in Table 5b, the performance improves with an increase in the diversity of classes sampled in CWCL. This suggests that contrasting a broader array of classes enhances the model's discriminative capability. Furthermore, Table 5c demonstrates that a balanced number of negative samples per class leads to optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Projection Head for SDCL. The influence of different project head configurations on the SDCL is investigated. We experimented with three distinct approaches: individual projection head, copying the weights of CWCL's (stop gradient), and shared projection head of CWCL. As demonstrated in Table 5d, sharing the projection head yielded the most superior results. The results indicate that SDCL not only relies on the semantic information from CWCL for effective disentanglement of similar classes but also enhances the embedding space learned by CWCL." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.233, + 0.71, + 0.29 + ], + "angle": 0, + "content": "
Cross-covariance loss
MethodCBM
Whitening38.6836.9142.12
Diagonal40.4237.8143.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.596, + 0.292, + 0.611, + 0.302 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.713, + 0.233, + 0.882, + 0.29 + ], + "angle": 0, + "content": "
# of classes
#CBM
1045.5738.8846.37
1545.7241.3247.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.79, + 0.291, + 0.805, + 0.301 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.31, + 0.671, + 0.374 + ], + "angle": 0, + "content": "
# of negative samples
#CBM
1044.7638.2146.46
5045.7241.3247.08
10044.4439.1446.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.583, + 0.376, + 0.596, + 0.384 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "table_caption", + "bbox": [ + 0.718, + 0.31, + 0.847, + 0.321 + ], + "angle": 0, + "content": "Projection Head of SDCL" + }, + { + "type": "table", + "bbox": [ + 0.681, + 0.322, + 0.885, + 0.374 + ], + "angle": 0, + "content": "
MLPCBM
Individual44.9138.4546.31
Shared (SG)44.0338.1546.64
Shared45.7241.3247.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.776, + 0.375, + 0.79, + 0.384 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.396, + 0.893, + 0.438 + ], + "angle": 0, + "content": "Table 5. Ablation studies. (a) Calculation of \\(\\mathcal{L}_{CC}\\). (b) Number of classes for \\(\\mathcal{L}_{CWCL}\\). (c) Number of negative samples for \\(\\mathcal{L}_{CWCL}\\). (d) Projection head of \\(\\mathcal{L}_{SDCL}\\). \"SG\" indicates stop gradient." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.463, + 0.619, + 0.479 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.893, + 0.701 + ], + "angle": 0, + "content": "In this paper, we propose a novel BlindNet with covariance alignment and semantic consistency contrastive learning. By introducing covariance alignment, our method effectively addresses style variations, ensuring the extraction of features that are consistent across different styles. Furthermore, with the proposed semantic consistency contrastive learning, we not only facilitate the extraction of discriminative features but also enhance the generalization capabilities of the model in semantic segmentation predictions. Comprehensive experimental results validate the effectiveness of our approach, demonstrating its ability to generalize across multiple unseen target domains without requiring auxiliary domains or additional modules. Our future work will be improving and stabilizing the covariance alignment method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported in part by the Basic Science Research Program through National Research Foundation of Korea (NRF) (Grants No. NRF2022R1F1A1073543), the MSIT(Ministry of Science and ICT), Korea, under the ICAN(ICT Challenge and Advanced Network of HRD) support program(RS-2022-00156385) supervised by the IITP(Institute for Information & Communications Technology Planning & Evaluation), and Innovative Human Resource Development for Local Intellectualization program through the Institute of Information & Communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT)(IITP-2024-00156287)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3623" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Florent Bartoccioni, Éloi Zablocki, Andrei Bursuc, Patrick Pérez, Matthieu Cord, and Karteek Alahari. Lara: Latents and rays for multi-camera bird's-eye-view semantic segmentation. In Conference on Robot Learning, pages 1663-1672. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.254 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.31 + ], + "angle": 0, + "content": "[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.368 + ], + "angle": 0, + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.471, + 0.438 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.471, + 0.508 + ], + "angle": 0, + "content": "[6] Wonwooong Cho, Sungha Choi, David Keetae Park, Inkyu Shin, and Jaegul Choo. Image-to-image translation via group-wise deep whitening-and-coloring transformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10639-10647, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[7] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11580-11590, 2021. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.677, + 0.471, + 0.732 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[10] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[12] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[15] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[16] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9924-9935, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[17] Lukas Hoyer, Dengxin Dai, Qin Wang, Yuhua Chen, and Luc Van Gool. Improving semi-supervised and domain-adaptive semantic segmentation with self-supervised depth estimation. International Journal of Computer Vision, pages 1-27, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[18] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[19] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6891-6902, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[20] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3061-3071, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[21] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[22] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European conference on computer vision (ECCV), pages 172-189, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[23] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[24] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3624" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 9936-9946, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[25] Guangrui Li, Guoliang Kang, Wu Liu, Yunchao Wei, and Yi Yang. Content-consistent matching for domain adaptive semantic segmentation. In European conference on computer vision, pages 440-456. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.468, + 0.26 + ], + "angle": 0, + "content": "[26] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 224-233, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.468, + 0.318 + ], + "angle": 0, + "content": "[27] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.468, + 0.389 + ], + "angle": 0, + "content": "[28] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European conference on computer vision (ECCV), pages 624-639, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.468, + 0.446 + ], + "angle": 0, + "content": "[29] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6936-6945, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.468, + 0.489 + ], + "angle": 0, + "content": "[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.468, + 0.533 + ], + "angle": 0, + "content": "[31] Ping Luo. Learning deep architectures via generalized whitened neural networks. In International Conference on Machine Learning, pages 2238-2246. PMLR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.535, + 0.468, + 0.589 + ], + "angle": 0, + "content": "[32] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European conference on computer vision (ECCV), pages 116-131, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.468, + 0.659 + ], + "angle": 0, + "content": "[33] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990-4999, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.468, + 0.716 + ], + "angle": 0, + "content": "[34] David Nilsson, Aleksis Pirinen, Erik Gartner, and Cristian Sminchisescu. Embodied visual active learning for semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2373–2383, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.468, + 0.787 + ], + "angle": 0, + "content": "[35] Yuya Onozuka, Ryosuke Matsumi, and Motoki Shino. Autonomous mobile robot navigation independent of road boundary using driving recommendation map. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4501-4508. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.468, + 0.83 + ], + "angle": 0, + "content": "[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[37] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3764-3773, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[38] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464-479, 2018. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[39] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1863-1871, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[40] Taesung Park, Alexei A Efros, Richard Zhang, and JunYan Zhu. Contrastive learning for unpaired image-to-image translation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 319-345. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[41] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE Transactions on Image Processing, 30:6594-6608, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[42] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2594-2605, 2022. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.501 + ], + "angle": 0, + "content": "[43] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 102-118. Springer, 2016. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[45] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.674, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[46] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10765-10775, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[47] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[48] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 5, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3625" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[50] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.164, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[51] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2517-2526, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.469, + 0.304 + ], + "angle": 0, + "content": "[52] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7303-7313, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.469, + 0.362 + ], + "angle": 0, + "content": "[53] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. Siamdoge: Domain generalizable semantic segmentation using siamese network. In European Conference on Computer Vision, pages 603-620. Springer, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.363, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[54] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.433, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[55] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2884-2892, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.519, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[56] Liwei Yang, Xiang Gu, and Jian Sun. Generalized semantic segmentation by self-supervised source domain projection and multi-level contrastive learning. arXiv preprint arXiv:2303.01906, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.576, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[57] Jaejun Yoo, Youngjung Uh, Sanghyuk Chun, Byeongkyu Kang, and Jung-Woo Ha. Photorealistic style transfer via wavelet transforms. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9036-9045, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.647, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.731, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[59] Fei Yu, Mo Zhang, Hexin Dong, Sheng Hu, Bin Dong, and Li Zhang. Dast: Unsupervised domain adaptation in semantic segmentation based on discriminator attention and self-training. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10754–10762, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.802, + 0.469, + 0.886 + ], + "angle": 0, + "content": "[60] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2100-2110, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.888, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[61] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Do" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "main generalization with mixstyle. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[62] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3626" + } + ] +] \ No newline at end of file diff --git a/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/1e340b0f-2934-49e8-9933-238d4fb891ff_origin.pdf b/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/1e340b0f-2934-49e8-9933-238d4fb891ff_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e157ea49f1d22c5ecbc3b2e4b224658069370068 --- /dev/null +++ b/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/1e340b0f-2934-49e8-9933-238d4fb891ff_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:210ab1c261d3c8d8d2f0c4875b0e27e76ea5891f0ba80c40572122e79e3059f8 +size 3461227 diff --git a/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/full.md b/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..64372bb750b3396719e46d844b78fe78fe228836 --- /dev/null +++ b/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/full.md @@ -0,0 +1,358 @@ +# Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning + +Woo-Jin Ahn1 Geun-Yeong Yang1 Hyun-Duck Choi2* Myo-Taeg Lim1* +1Korea University 2Chonnam National University + +$^{1}\{wjahn, hggofficial, mlim\} @korea.ac.kr$ $^{2}$ ducky.choi@jnu.ac.kr + +# Abstract + +Deep learning models for semantic segmentation often experience performance degradation when deployed to unseen target domains unidentified during the training phase. This is mainly due to variations in image texture (i.e. style) from different data sources. To tackle this challenge, existing domain generalized semantic segmentation (DGSS) methods attempt to remove style variations from the feature. However, these approaches struggle with the entanglement of style and content, which may lead to the unintentional removal of crucial content information, causing performance degradation. This study addresses this limitation by proposing BlindNet, a novel DGSS approach that blinds the style without external modules or datasets. The main idea behind our proposed approach is to alleviate the effect of style in the encoder whilst facilitating robust segmentation in the decoder. To achieve this, BlindNet comprises two key components: covariance alignment and semantic consistency contrastive learning. Specifically, the covariance alignment trains the encoder to uniformly recognize various styles and preserve the content information of the feature, rather than removing the style-sensitive factor. Meanwhile, semantic consistency contrastive learning enables the decoder to construct discriminative class embedding space and disentangles features that are vulnerable to misclassification. Through extensive experiments, our approach outperforms existing DGSS methods, exhibiting robustness and superior performance for semantic segmentation on unseen target domains. The code is available at https://github.com/root0yang/BlindNet. + +# 1. Introduction + +Semantic segmentation, a technique that classifies each pixel in an image into predefined categories, has garnered significant attention due to its potential applications in vari- + +![](images/f1ff12b29489673484c490594c0b0065fa431db390eaa18e1fb3b77b2196fbd2.jpg) +Figure 1. Comparison of semantic segmentation results between the baseline (DeepLabV3+ with ResNet50 backbone) and our BlindNet. Both models are trained on the source domain (GTAV [43]) and tested on the target domain (Cityscapes [8]). + +ous fields. Particularly, it plays a crucial role in autonomous driving [1, 18] and robotic systems [34, 35]. Besides, with the advent of large datasets, deep neural networks have emerged as a trending approach for semantic segmentation tasks, achieving impressive results [3, 5, 44, 54]. However, there remains a major bottleneck detailing the meticulous and labor-intensive process of dataset labeling. More particularly, this process not only consumes time but also poses economic challenges [8, 46]. To address this challenge, synthetic datasets have emerged as a compelling alternative. Specifically, these datasets, generated using three-dimensional (3D) rendering techniques, offer vast amounts of easily accessible data, eliminating the need for manual labeling [43, 45]. However, a challenge arises when models trained on synthetic datasets are deployed in real-world scenarios. More precisely, a domain shift problem arises due to style factor discrepancies (e.g. texture, illumination, and image quality) between synthetic and real-world data, which affects the performance of the model, as shown in Fig. 1. + +To address the domain shift problem, domain adaptation semantic segmentation (DASS) has been introduced [15, 17, 22, 25, 51, 62]. Specifically, DASS aims to bridge the gap between source and target domains by aligning their data distributions. However, a significant limitation of DASS is its dependency on the target domain during train- + +ing. For DASS to function effectively, target domain samples must be available during the training phase. + +Meanwhile, domain generalized semantic segmentation (DGSS) has been introduced as an alternative approach to tackle the domain shift problem. Unlike DASS, DGSS is trained only with the source domain, aiming to extract domain-invariant features. To achieve this, two main techniques have been employed: domain randomization (DR) and feature normalization (FN). + +DR augments the training set by introducing variability, either by altering the image style [41, 60] or by modifying the feature representation [24, 53]. By exposing the model to a wider variety of styles via DR, the network is less likely to overfit to the specific styles present in the training data. Consequently, the robustness of the model is improved, making it more adept at generalizing to new, unseen domains. Nevertheless, a crucial limitation of DR is its significant dependence on auxiliary domains. + +FN methods, converse to DR, regularize the features to prevent the model from overfitting to the distinct styles or characteristics of the training data. This is achieved by removing domain-specific style information using feature statistics, such as instance normalization [38] or whitening transformation [7, 19, 39, 42]. While these approaches effectively remove style-related information, they simultaneously pose the challenge of removing semantic content because content and style information are entangled. Consequently, the model fails to capture the essential patterns or features required for accurate segmentation prediction. + +To address this problem, we propose BlindNet, a model that blinds the style within the encoder and improves the robustness of the decoder, without requiring auxiliary datasets or external modules. Specifically, our proposed BlindNet consists of two components: covariance alignment for the encoder and semantic consistency contrastive learning for the decoder. Precisely, the covariance alignment facilitates the generation of style-invariant features with the proposed covariance matching loss function (CML) and the cross-covariance loss function (CCL). Specifically, CML mitigates the effects of style variations, while CCL focuses on preserving content information, effectively addressing the prevalent content information loss observed in the FN method. To further improve the generalization ability, we develop semantic consistency contrastive learning, which consists of class-wise contrastive learning (CWCL) and semantic disentanglement contrastive learning (SDCL). Particularly, the CWCL constructs a discriminative class embedding space, while SDCL disentangles features of similar classes that often lead to prediction errors. Extensive experiments across various datasets demonstrate that the proposed BlindNet outperforms existing DGSS methods. + +Our contributions are summarized as follows: + +- We propose a covariance alignment within the encoder, + +comprising CML and CCL. Specifically, the CML aims to mitigate the effects of style variations, while CCL ensures the preservation of content information, together facilitating the generation of style-agnostic features. + +- We propose semantic consistency contrastive learning within the decoder that comprises CWCL and SDCL, utilizing segmentation masks. Specifically, CWCL generates discriminative embeddings, whereas SDCL disentangles features of similar classes, enhancing the robustness of the model. +- Through extensive experiments, we demonstrate the superiority of our approach in DGSS, without the need to alter the network architecture or rely on external datasets. + +# 2. Related Works + +Domain adaptation and generalization for semantic segmentation. Domain adaptation (DA) aims at minimizing the distribution discrepancy between different domains, enabling a model to generalize from a source to a target domain. For DASS, adversarial training and cross-domain self-training strategies are commonly used. Particularly, adversarial-based methods [15, 29] employ generative adversarial networks [11] to close the feature distribution gap between source and target domains. Meanwhile, cross-domain self-training methods [16, 37, 59, 62] generate pseudo-labels for target domain data using pre-trained models, and employ them as training data, thereby expanding the training data and reducing distribution differences between the source and target domain. + +Domain generalization (DG) methods [26, 28, 30, 50, 61] aim to improve the generalization ability of the model without accessing the target domain during training. Since the difference in style of the image is the main cause of the disparity, most existing domain generalized semantic segmentation methods utilize the style information of the image for domain-invariant learning. Feature statistics (e.g. mean, variance, covariance, gram matrix, etc.) which are commonly used in style transfer [10, 21, 27, 57] are employed to capture the style information. Interestingly, existing DGSS methods can be separated into two parts: i) domain randomization to expand the distribution of style or ii) feature normalization to remove style. + +DR involves randomizing either the image or its features through stylization to learn domain-invariant features from various styles. For example, Peng et al. [42] extended the source domain data by stylizing images in the style of unreal paintings. Similarly, Yue et al. [60] and Huang et al. [22] attempted to enhance generalization by synthesizing images with diverse styles in the image space. In another study, Lee et al. [24] adopted ImageNet data [9] as wild data and performed randomization via synthesis in the feature space. Meanwhile, Wu et al. [53] diversified the trainable feature space by mixing the statistics of the feature and its color- + +jittered feature with Ada-IN [21]. + +FN methods aim to remove domain-specific styles from features, extracting only domain-invariant content. For instance, Pan et al. [39] first attempted the DGSS method, combining batch normalization (BN) [23] and instance normalization (IN) [48] in the network layer. While BN preserves the content information within discriminative features, IN focuses on removing domain-specific style information from features. In a study, Choi et al. [7] addressed the limitations of previous whitening transformation [6, 31] that can eliminate the content information from the feature. Specifically, they proposed an instance-selective whitening approach designed to remove covariance components that are sensitive to domain shifts. Peng et al. [41] developed semantic-aware normalization that performs on class-wise and semantic-aware whitening that aligns channels based on the prediction through group whitening transformation [6]. In a study, Xu et al. [55] introduced the prior guided attention module and guided feature whitening to re-calibrate the feature and remove domain-specific style effects, respectively. Unlike the FN methods that directly remove the style component, our work explores a covariance alignment method that mitigates the effect of the style's effect while preserving the content information. Our method achieves the DGSS without any additional modules or auxiliary datasets. + +Contrastive Learning. Contrastive learning aims to learn representations by maximizing the similarity between positive pairs of samples while minimizing the similarity between negative pairs. In recent years, it has attracted significant attention for its effectiveness in learning discriminative representations across various tasks [2, 4, 12, 14]. Oord et al. [36] were the first to introduce the infoNCE loss, a type of contrastive loss function designed for self-contrastive learning. In a work, Park et al. [40] introduced patch-level contrastive learning for image translation, using co-located patches as positive pairs and spatially distant patches as negatives to maintain image context. For the semantic segmentation task, Wang et al. [52] introduced class-wise contrastive learning to aid the model in learning the embedding space of each class. Specifically, they sampled the classes existing in the images and applied contrastive learning based on the class label. For DGSS, Lee et al. [24] adopted contrastive learning to learn the ImageNet information in their model. Specifically, they set the ImageNet data as wild and applied the contrastive learning method by setting the wild-stylized feature and its closest wild content as positive samples. In another study, Yang et al. [56] developed multi-level contrastive learning, which designed instance prototypes and class prototypes for contrastive learning. Specifically, they sample each class's pixel features and apply contrastive learning with a transition-probability ability matrix. Unlike recent DGSS works that embed the origi + +nal image, we define contrastive learning for the augmented image. Specifically, the proposed method builds a robust embedding space by preserving the semantic consistency of the feature representation across various domains. + +# 3. Method + +The goal of the proposed method is to train a segmentation model $\varphi$ on a given source domain $S$ and generalize well to the unseen target domain $T$ . Precisely, the source domain $S = \{(x,y)\}$ contains a paired image $x \in \mathbb{R}^{H \times W \times 3}$ and segmentation label $y \in \mathbb{R}^{H \times W \times C}$ , where $H$ , $W$ , and $C$ denote the height of the image, the width of the image, and the class number of the segmentation map, respectively. The model $\varphi$ takes an original $x$ and its augmented counterpart $x_{a}$ , which have the same content but different styles, and uses the feature information to enhance its generalization ability. + +As shown in Fig. 2, our method leverages the feature information through two main approaches: covariance alignment and semantic consistency contrastive learning. Specifically, the covariance matching ensures that features, having different styles but the same content, contain similar information. Additionally, the semantic consistency contrastive learning embeds the generalized features into discriminative representation based on the segmentation label. + +# 3.1. Covariance Alignment + +The domain shift in semantic segmentation results from changes in the visual characteristics of the image, known as style. This style information is typically detected in the shallow layers of networks [38], which are the encoders of the segmentation model. Based on this understanding, our method targets the encoder to handle style variations by employing the proposed covariance matching loss and cross-covariance loss function. + +# Covariance Matching Loss + +To train the network to uniformly recognize various styles without removing content information, we introduce the CML. Specifically, the loss aims to minimize the difference between covariance matrices derived from different styles of image features. Given an image pair $(x,x_{a})$ , the features from $i^{th}$ block of the encoder are represented as $F^i\in \mathbb{R}^{(H^i\times W^i)\times C^i}$ and $F_{a}^{i}\in \mathbb{R}^{(H^{i}\times W^{i})\times C^{i}}$ , respectively. Further, following the methodologies of [7, 41], we compute the covariance matrices using instance normalized features [48], which ensures consistent scaling across features. The feature maps are normalized and flattened into $\bar{F}$ and $\bar{F}_a\in \mathbb{R}^{(HW)\times C}$ , which are given by: + +$$ +\bar {F} = \frac {(F - \mu (F))}{\sigma (F)}, \bar {F} _ {a} = \frac {(F _ {a} - \mu (F _ {a}))}{\sigma (F _ {a})}, \tag {1} +$$ + +![](images/5fcbbb724d3720e68e61b5e9b5a91dabb76d3372b3e54fd809de935bffa75425.jpg) +Figure 2. Overview of the proposed BlindNet. The network processes a pair of images - the original image $x$ and its augmented counterpart $x_{a}$ . It employs covariance alignment to treat encoder features and utilizes semantic consistency contrastive learning for the processing of decoder features. + +where $\mu (\cdot)\in \mathbb{R}^{C}$ and $\sigma (\cdot)\in \mathbb{R}^C$ denote the mean and standard deviation of the features. Utilizing the normalized features, we evaluate the covariance matrices for the original and augmented image features as: + +$$ +\Sigma_ {x, x} ^ {i} = \bar {F} ^ {i} ^ {T} \cdot \bar {F} ^ {i}, \quad \Sigma_ {x _ {a}, x _ {a}} ^ {i} = \bar {F} _ {a} ^ {i} ^ {T} \cdot \bar {F} _ {a} ^ {i}. \tag {2} +$$ + +The CML is then formulated to align these covariance matrices, ensuring that the network maintains consistency in the presence of style variations. The CML is defined as: + +$$ +\mathcal {L} _ {C M} = \sum_ {i = 1} ^ {n _ {e}} \| \Sigma_ {x, x} ^ {i} - \Sigma_ {x _ {a}, x _ {a}} ^ {i} \| _ {2}, \tag {3} +$$ + +where $n_e$ denotes the number of blocks of the encoder. + +# Cross-covariance Loss + +While the CML effectively aligns the internal distributions of features within the same image, it does not fully account for the direct correlations across paired images. To complement this, we introduce CCL, which aims to encode the consistent content information of an image pair $(x,x_{a})$ by utilizing the cross-covariance of the image pair. Given the normalized feature pair $(\bar{F}^i,\bar{F}_a^i)$ , the cross-covariance of the feature pair can be expressed as: + +$$ +\Sigma_ {x, x _ {a}} ^ {i} = \bar {F} ^ {i} ^ {T} \cdot \bar {F} _ {a} ^ {i}. \tag {4} +$$ + +The cross-covariance is expected to exhibit an identity matrix, as the feature pair should contain identical information. Nonetheless, the proposed CCL converges only the diagonal component of the covariance matrix to one. This is to prevent the drawbacks of the existing FN methods [7, 42] from removing content information. The CCL function is thus defined as: + +$$ +\mathcal {L} _ {C C} = \sum_ {i = 1} ^ {n _ {e}} \| \operatorname {d i a g} \left(\Sigma_ {x, x _ {a}} ^ {i}\right) - \mathbb {1} \| _ {2}, \tag {5} +$$ + +where $\mathrm{diag}(\Sigma_c) \in \mathbb{R}^C$ denotes the column vector comprising diagonal elements of $\Sigma_{x,x_a}^i$ and $\mathbb{1} \in \mathbb{R}^C$ denotes the one vector. + +# 3.2. Semantic Consistence Contrastive Learning + +While the encoder focuses on generating style-blinded features, the decoder aims to improve the robustness of the segmentation prediction against domain shifts. For the decoder, we employ a contrastive learning approach, which has demonstrated effectiveness in extracting discriminative features [4]. Specifically, we utilize the InfoNCE loss [36], which is formulated as: + +$$ +\mathcal {L} _ {I N} (a, p, n) = - \log \left(\frac {e ^ {(a \cdot p / \tau)}}{e ^ {(a \cdot p / \tau)} + \sum_ {n} ^ {N ^ {-}} e ^ {(a \cdot n / \tau)}}\right) \tag {6} +$$ + +where $a, p, n$ , and $N^{-}$ denote anchor, positive sample, negative sample, and negative sample set, respectively. + +![](images/f9fb00733b6b0939d8e914bad891b9b9d71654c5baf4d18638084e017a3d791f.jpg) +Figure 3. Illustration of semantic consistency contrastive learning: The mask $M$ represents the error mask derived from the augmented segmentation map. CWCL conducts contrastive learning by sampling per segmentation class and SDCL conducts contrastive learning based on the $M$ . Both methods share a projection head $\pi$ for the semantic representation. + +To achieve consistent feature representation in DGSS across various styles, we introduce semantic consistency contrastive learning. Specifically, the anchor is derived from the augmented image $x_{a}$ , while the positive sample is extracted from the corresponding pixel of the original image $x$ . Our method consists of two main components based on the negative sample as shown in Fig. 3: class-wise contrastive learning and semantic disentanglement contrastive learning. + +# Class-wise Contrastive Learning + +Our CWCL aims to build a discriminative embedding space for each segmentation class using different classes of the original image as negatives. Given an image pair $(x, x_{a})$ , the features from the $j^{th}$ block of the decoder at pixel position $(m, n)$ are denoted as $F_{(m, n)}^{j}$ and $F_{a, (m, n)}^{j} \in \mathbb{R}^{(1 \times 1) \times C^{j}}$ , where $C^{j}$ indicates the channel length of the feature. As mentioned above, we take $F_{a, (m, n)}^{j}$ as the anchor and $F_{(m, n)}^{j}$ as the positive sample, since they represent the same content at the corresponding spatial location. To obtain the negative samples from $F^{j}$ , we leverage the resized segmentation class label $y^{j} \in \mathbb{R}^{(H^{j} \times W^{j}) \times C}$ to collect the different class samples. The samples are passed through the projection head, denoted as $\pi$ , resulting in the projected features $\tilde{F}$ and $\tilde{F}_{a}$ . We define our CWCL as: + +$$ +\mathcal {L} _ {C W C L} = \sum_ {j} ^ {n _ {d}} \mathcal {L} _ {I N} \left(\tilde {F} _ {a, (m, n)} ^ {j}, \tilde {F} _ {(m, n)} ^ {j}, \tilde {F} _ {(p, q)} ^ {j}\right) \tag {7} +$$ + +where $(p,q)\in \{(p,q)\in P|y_{(p,q)}^j\neq y_{(m,n)}^j\}$ + +The set $P$ represents all pixel positions in feature $F^j$ , with dimensions $H^j \times W^j$ corresponding to the height and width. Additionally, $n_d$ denotes the total number of blocks in the decoder. + +![](images/364e0bcd00ec452f09a91d3c336a3abe80fbdab1f326b463d00f6dc6623bce24.jpg) +Figure 4. t-SNE [49] visualization comparing scenarios with and without $\mathcal{L}_{SCCL}$ . In (b), the application of SCCL results in a clear separation between the sidewalk (pink), the road (purple), and the building (gray). + +![](images/1185b7eb568978d73c2f83c1c0f9c882983f91b10edd00fd456149b4d75b6763.jpg) + +# Semantic Disentanglement Contrastive Learning + +Domain shifts can lead to the entanglement of similar classes, causing the model to misclassify, as illustrated in Fig. 4. To mitigate this issue, we introduce the SDCL, specifically designed to disentangle the feature $x_{a}$ that has been misclassified, making it closer to the correct class and far from the misclassified class to achieve effective disentanglement. To further ensure consistent feature space and capture the semantic meaning, we share the projection head $\pi$ used in the CWCL loss. Given the predicted segmentation map of the augmented image, represented as $\hat{y}_a = \varphi (x_a)$ , we resize it to $\hat{y}_{a,(m,n)}^j\in \mathbb{R}^{(H^j\times W^j)\times C}$ . Similarly, $y_{(m,n)}^j$ represents the ground truth segmentation map. Using these segmentation maps, we set the anchor at positions where $\hat{y}_{a,(m,n)}^j\neq y_{(m,n)}^j$ . Negative samples are selected from the augmented image features corresponding to the anchor's misclassified class. The samples go through the projection head $\pi$ . Our SDCL loss is defined as follows: + +$$ +\mathcal {L} _ {S D C L} = \sum_ {j} ^ {n _ {d}} \mathcal {L} _ {I N} \left(\tilde {F} _ {a, (m, n)} ^ {j}, \tilde {F} _ {(m, n)} ^ {j}, \tilde {F} _ {a, (r, s)} ^ {j}\right) \tag {8} +$$ + +where $(r,s)\in \{(r,s)\in P|y_{(r,s)}^j = \hat{y}_{(m,n)}^j\}$ + +Finally, combining the cross-entropy segmentation loss + +
BackboneMethodsExternalTrained on GTAV (G)Trained on Cityscapes (C)
DatasetModuleCBMSBMSG
ResNet50 [13]Baseline [3]--28.9525.1428.1826.2344.9651.6823.2942.55
IBN-Net [38]--33.8532.3037.7527.9048.5657.0426.1445.06
RobustNet [7]--37.3135.2040.3328.3050.7358.6426.2045.00
SiamDoGe [53]--42.9637.5440.6428.3451.5359.0026.6745.08
DIRL [55]-✓41.0439.1541.60-51.80-26.5046.52
WildNet [24]✓-44.6238.4246.0931.3450.9458.7927.9547.01
SANSAW [42]-✓39.7537.3441.8630.7952.9559.8128.3247.28
SPC [20]-✓44.1040.4645.51-----
DPCL [56]-✓44.7440.5946.3330.8150.9758.5925.8546.00
Ours--45.7241.3247.0831.3951.8460.1828.5147.97
ShuffleNetV2 [32]Baseline [3]--25.5622.1728.6023.3336.8443.1321.5636.95
IBN-Net [38]--27.1031.8234.8925.5641.8946.3522.9940.91
RobustNet [7]--30.9832.0635.3124.3141.9446.9722.8240.17
SiamDoGe [53]--34.4034.2335.8721.9542.6147.4823.1340.93
DIRL [55]-✓31.8832.5736.12-42.55-23.7441.23
DPCL [56]-✓36.6634.3539.9222.6643.9048.9522.4741.07
Ours--38.5634.5140.1125.6444.2249.6923.5441.10
+ +Table 1. Quantitative comparison of mIoU (%) between DGSS methods. External dataset denotes the necessity of an auxiliary dataset during training and External module denotes the requirement of an additional module during inference. G, C, B, M, and S denote GTAV, Cityscapes, BDD100K, Mapillary, and SYNTHIA, respectively. The best and second-best results are bolded and underlined, respectively. + +$\mathcal{L}_{CE}$ with other loss components, the total is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {C E} + \omega_ {1} \mathcal {L} _ {C M} + \omega_ {2} \mathcal {L} _ {C C} \tag {9} \\ + \omega_ {3} \mathcal {L} _ {C W C L} + \omega_ {4} \mathcal {L} _ {S D C L} \\ \end{array} +$$ + +where $\omega_{1},\omega_{2},\omega_{3}$ , and $\omega_{4}$ denote the weighting factor of each loss functions. + +# 4. Experiment + +In this section, we describe the implementation details, the experimental setup for comparison with existing DGSS methods, and the ablation study conducted to further validate the effectiveness of our approach. + +# 4.1. Implementation Details + +We adopt DeepLabV3+ [3] for the segmentation architecture and use ResNet-50 [13], ShuffleNetV2 [32], and MobileNetV2 [47] as the backbone network of the segmentation network. The model is trained for 40K iterations with a batch size of 8 using the SGD optimizer, which has a momentum of 0.9 and a weight decay of 5e-4. We employ a polynomial learning rate schedule with an initial rate of 1e-2 and a power of 0.9. For the simulation of domain shift, we augment the image $x_{a}$ using strong color jittering transformation similar to [7]. The weighting parameters of (9), $\omega_{1}$ , $\omega_{2}$ , $\omega_{3}$ and $\omega_{4}$ , are set as 0.2, 0.2, 0.3, and 0.3 respectively. + +# 4.2. Datasets + +We use two synthetic datasets (GTA [43] and SYNTHIA [45]), and three real-world datasets (Cityscapes [8], + +BDD-100K [58], and Mapillary [33]) for the experiment. All segmentation labels are evaluated based on 19 object categories. + +GTAV (G) [43] is a large-scale dataset generated from the Grand Theft Auto V (GTAV) game engine. It comprises 24,966 images, split into 12,403 for training, 6,382 for validation, and 6,181 for testing with a resolution of $1914 \times 1052$ . + +SYNTHIA (S) [45] is a virtual, photo-realistic urban scene dataset comprising 9,400 images with a resolution of $960 \times 720$ . Among these, 2,820 images are designated for evaluation. + +Cityscapes (C) [8] is a large-scale urban scene dataset captured from 50 cities, primarily in Germany. Particularly, it contains 5,000 high-resolution images with a resolution of $2048 \times 1024$ . The dataset is divided into 2,975 images for training, 500 for validation, and 1,525 for testing. + +BDD-100K (B) [58] is another real-world urban scene dataset that contains more diverse 10000 urban driving scene images with a resolution of $1280 \times 720$ . Specifically, the validation split (1,000 images) is used for evaluation. + +Mapillary (M) [33] contains 25,000 images with a minimum resolution of $1920 \times 1080$ , collected from various locations worldwide. Specifically, the validation split of 2,000 images is used for evaluation. + +# 4.3. Comparison with DGSS methods + +We compare our methods with other state-of-the-art DGSS methods: Baseline (DeepLabV3+ [3] trained on the source domain), IBN-Net [38], RobustNet [7], SiamDoGe [53], + +
MethodsExternal ModuleTrained on GTAV (G)
CBMMean
Baseline [3]25.9425.7326.4526.04
IBN-Net [38]30.1427.6627.0728.29
RobustNet [7]30.8630.0530.6730.52
SiamDoGe [53]34.1534.5032.3433.67
DIRL [55]✓34.6732.7834.3133.92
DPCL [56]✓37.5735.4540.3037.77
Ours37.6636.1040.4038.05
+ +DIRL [55], WildNet [24], SANSAW [42], SPC [20], and DPCL [56]. To evaluate the generalization ability of the model on arbitrary unseen domains, we conduct the experiment on two scenarios: i) trained on GTAV, tested on Cityscapes, BDD-100K, and Mapillary, and ii) trained on Cityscapes, tested on BDD-100K, Mapillary, and SYNTHIA. The quantitative results are computed with mean intersection over union (mIoU). Additionally, we compared the method trained on the backbone of ResNet-50 [13], ShuffleNetV2 [32], and MoblieNet [47], pre-trained on ImageNet [9]. + +# Quantitative and Qualitative Results + +Table 1 summarizes the quantitative results. Our method outperforms all other methods when trained on GTAV, using ResNet-50 as the backbone. When compared with FN methods that remove domain-specific styles, we demonstrate that our approach minimizes the loss of content information. We also show that our method effectively shows generalization ability when trained on Cityscapes. We further evaluate our methods with different backbones, showing the wide applicability of our method. When trained with ShuffleNetV2, our method achieves the first or second-best performance among unseen target domains. Table 2 shows the results of our method trained on GTAV with MobileNetV2, demonstrating the superiority of our method. + +For qualitative evaluation, we compare the visual result between DGSS methods and ours. As depicted in Fig. 5, our method demonstrates superior results compared to other approaches, particularly in its overall prediction accuracy. Notably, our proposed techniques enable distinct prediction of features such as on road and sidewalk, yielding clearer segmentation boundaries. Please refer to the supplementary material for more qualitative results. + +# Computational cost analysis + +To confirm that our approach does not incur additional computational overhead, we provide the number of parameters, GFLOPS, and average inference time of each method. As + +Table 2. Quantitative comparison of mIoU $(\%)$ using MobileNetV2 [47] backbone trained on the GTAV dataset. + +
MethodsExternal ModuleParams (M)GFLOPSTime (ms)
Baseline [13]45.08277.7710.01
SANSAW [42]✓25.63421.8668.96
SPC [20]✓45.22286.0912.24
DIRL [55]✓45.41278.1111.69
DPCL [56]✓56.461188.64823.78
Ours45.08277.7810.03
+ +Table 3. Computational cost comparison conducted using DeepLabV3+ with a ResNet-50 backbone on an NVIDIA Tesla V100 GPU with an image resolution of $2048 \times 1024$ . Inference time is averaged over 400 trials. + +
LCMCCLCWCLSDCLCBM
28.9525.1428.18
✓38.0836.6540.62
✓✓40.4237.8143.91
✓42.0338.2744.02
✓✓43.1638.5945.38
✓✓✓43.1738.2344.84
✓✓✓✓45.7241.3247.08
+ +Table 4. Ablation study on proposed losses. The experiments were conducted using DeepLabV3+ with ResNet-50 backbone, trained on the GTAV dataset. The losses are detailed in $\mathcal{L}_{CM}$ : (3), $\mathcal{L}_{CC}$ : (5), $\mathcal{L}_{CWCL}$ : (7), $\mathcal{L}_{SDCL}$ : (8) + +detailed in Table 3, our method operates comparably to baseline models by learning features intrinsically without adopting a separate module. + +# 4.4. Ablation Studies + +In this subsection, we conducted a series of ablation studies to demonstrate the individual contribution and effectiveness of each component within our method. Specifically, we investigate the impact of the following components: $\mathcal{L}_{CM}$ , $\mathcal{L}_{CC}$ , $\mathcal{L}_{CWCL}$ , $\mathcal{L}_{SDCL}$ . For the study, we use a scenario where the DeepLabV3+ model with backbone ResNet-50 model is trained on GTA and tested on Cityscapes, BDD100K, and Mapillary. + +Table 4 presents the impact of various proposed losses on domain generalization performance. Specifically, the baseline model, trained solely with cross-entropy loss, exhibits suboptimal performance on target domains because of overfitting to the source domain. Conversely, the integration of any proposed loss mechanisms leads to a marked enhancement in performance. More specifically, the incorporation of the covariance alignment $(\mathcal{L}_{CM},\mathcal{L}_{CC})$ shows its efficacy in preserving essential content information by correlating features of paired images. The differential impact of the semantic consistency constrastive learning $(\mathcal{L}_{CWCL},\mathcal{L}_{SDCL})$ is also evident, as it significantly aids in disentangling features of similar classes, thereby constructing a more robust embedding space. + +![](images/5743df130013ec2edea812c12d2a7e375fd2bbd88be13afab8d1a8d9272aea9c.jpg) +Target Data + +![](images/337f71e74c8508b424ee737397e719b5030e129644c3e795c1a4ea71ba7d1e6a.jpg) +Baseline + +![](images/49c2a2eee91c74f79070a020372836d10140ed59f2a9e5dc8645dd7751c88c18.jpg) +RobustNet + +![](images/95e5f7abd34e2a68c19572c4e0fa1674385464d3501bd35dfd4aa3a6bd004b39.jpg) +WildNet + +![](images/ba681e1508b8da213ef145d65f63839dc55f7d2eace8e17b8c6c0de9e132f18f.jpg) +SPC + +![](images/56a57369c29ab1ee852bc71832d6b55c8c7c94a5be7f39e7bf518f526eb5fc04.jpg) +Ours + +![](images/e75373901ae793412627ba1d0ec111765d7b497f114899c12c47858de9371ae6.jpg) +Ground Truth + +Figure 5. Qualitative comparison between DGSS methods trained on GTAV (G) and tested on unseen target domains of Cityscapes (C) using DeeplabV3+ with ResNet50 backbone. +Figure 6. t-SNE [49] visualization comparing the covariance with and without $\mathcal{L}_{CML}$ . +![](images/97b7de89b1fade3976ea20663122c5a175bbbba4817f004ecaee8fedc9661f42.jpg) +Baseline - Original $x$ +Baseline - Augmented $x_{a}$ +Ours - Original $x$ +Ours-Augmented $x_{a}$ + +Covariance Matching Loss. Fig. 6 presents t-SNE plots of covariances for original and augmented images, before and after the application of CML. The baseline network perceives original and augmented images differently from a style perspective. However, after applying CML, the distribution becomes more intermixed, indicating that our proposed CML effectively ensures similar recognition of different style images. + +Calculation of CCL. Table 5a demonstrates that our proposed cross-covariance method, which converges the diagonal components to 1, yields superior performance. As mentioned before, removing non-diagonal components, which contain content information actually degrades performance. + +Sampling number in CWCL. Table 5b and Table 5c show the impact of varying the number of classes sampled per image and the number of samples per class in CWCL, respectively. As shown in Table 5b, the performance improves with an increase in the diversity of classes sampled in CWCL. This suggests that contrasting a broader array of classes enhances the model's discriminative capability. Furthermore, Table 5c demonstrates that a balanced number of negative samples per class leads to optimal performance. + +Projection Head for SDCL. The influence of different project head configurations on the SDCL is investigated. We experimented with three distinct approaches: individual projection head, copying the weights of CWCL's (stop gradient), and shared projection head of CWCL. As demonstrated in Table 5d, sharing the projection head yielded the most superior results. The results indicate that SDCL not only relies on the semantic information from CWCL for effective disentanglement of similar classes but also enhances the embedding space learned by CWCL. + +
Cross-covariance loss
MethodCBM
Whitening38.6836.9142.12
Diagonal40.4237.8143.91
+ +(a) + +
# of classes
#CBM
1045.5738.8846.37
1545.7241.3247.08
+ +(b) + +
# of negative samples
#CBM
1044.7638.2146.46
5045.7241.3247.08
10044.4439.1446.29
+ +(c) +Projection Head of SDCL + +
MLPCBM
Individual44.9138.4546.31
Shared (SG)44.0338.1546.64
Shared45.7241.3247.08
+ +(d) + +Table 5. Ablation studies. (a) Calculation of $\mathcal{L}_{CC}$ . (b) Number of classes for $\mathcal{L}_{CWCL}$ . (c) Number of negative samples for $\mathcal{L}_{CWCL}$ . (d) Projection head of $\mathcal{L}_{SDCL}$ . "SG" indicates stop gradient. + +# 5. Conclusion + +In this paper, we propose a novel BlindNet with covariance alignment and semantic consistency contrastive learning. By introducing covariance alignment, our method effectively addresses style variations, ensuring the extraction of features that are consistent across different styles. Furthermore, with the proposed semantic consistency contrastive learning, we not only facilitate the extraction of discriminative features but also enhance the generalization capabilities of the model in semantic segmentation predictions. Comprehensive experimental results validate the effectiveness of our approach, demonstrating its ability to generalize across multiple unseen target domains without requiring auxiliary domains or additional modules. Our future work will be improving and stabilizing the covariance alignment method. + +Acknowledgement. This work was supported in part by the Basic Science Research Program through National Research Foundation of Korea (NRF) (Grants No. NRF2022R1F1A1073543), the MSIT(Ministry of Science and ICT), Korea, under the ICAN(ICT Challenge and Advanced Network of HRD) support program(RS-2022-00156385) supervised by the IITP(Institute for Information & Communications Technology Planning & Evaluation), and Innovative Human Resource Development for Local Intellectualization program through the Institute of Information & Communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT)(IITP-2024-00156287). + +# References + +[1] Florent Bartoccioni, Éloi Zablocki, Andrei Bursuc, Patrick Pérez, Matthieu Cord, and Karteek Alahari. Lara: Latents and rays for multi-camera bird's-eye-view semantic segmentation. In Conference on Robot Learning, pages 1663-1672. PMLR, 2023. 1 +[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 3 +[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1, 6, 7 +[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3, 4 +[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1 +[6] Wonwooong Cho, Sungha Choi, David Keetae Park, Inkyu Shin, and Jaegul Choo. Image-to-image translation via group-wise deep whitening-and-coloring transformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10639-10647, 2019. 3 +[7] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11580-11590, 2021. 2, 3, 4, 6, 7 +[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 6 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 7 +[10] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2 +[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[12] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach + +to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7 +[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3 +[15] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 1, 2 +[16] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9924-9935, 2022. 2 +[17] Lukas Hoyer, Dengxin Dai, Qin Wang, Yuhua Chen, and Luc Van Gool. Improving semi-supervised and domain-adaptive semantic segmentation with self-supervised depth estimation. International Journal of Computer Vision, pages 1-27, 2023. 1 +[18] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1 +[19] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6891-6902, 2021. 2 +[20] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3061-3071, 2023. 6, 7 +[21] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 3 +[22] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European conference on computer vision (ECCV), pages 172-189, 2018. 1, 2 +[23] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 3 +[24] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 9936-9946, 2022. 2, 3, 6, 7 +[25] Guangrui Li, Guoliang Kang, Wu Liu, Yunchao Wei, and Yi Yang. Content-consistent matching for domain adaptive semantic segmentation. In European conference on computer vision, pages 440-456. Springer, 2020. 1 +[26] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 224-233, 2021. 2 +[27] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 2 +[28] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European conference on computer vision (ECCV), pages 624-639, 2018. 2 +[29] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6936-6945, 2019. 2 +[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2 +[31] Ping Luo. Learning deep architectures via generalized whitened neural networks. In International Conference on Machine Learning, pages 2238-2246. PMLR, 2017. 3 +[32] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European conference on computer vision (ECCV), pages 116-131, 2018. 6, 7 +[33] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990-4999, 2017. 6 +[34] David Nilsson, Aleksis Pirinen, Erik Gartner, and Cristian Sminchisescu. Embodied visual active learning for semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2373–2383, 2021. 1 +[35] Yuya Onozuka, Ryosuke Matsumi, and Motoki Shino. Autonomous mobile robot navigation independent of road boundary using driving recommendation map. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4501-4508. IEEE, 2021. 1 +[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3, 4 +[37] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3764-3773, 2020. 2 + +[38] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464-479, 2018. 2, 3, 6, 7 +[39] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1863-1871, 2019. 2, 3 +[40] Taesung Park, Alexei A Efros, Richard Zhang, and JunYan Zhu. Contrastive learning for unpaired image-to-image translation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 319-345. Springer, 2020. 3 +[41] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE Transactions on Image Processing, 30:6594-6608, 2021. 2, 3 +[42] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2594-2605, 2022. 2, 4, 6, 7 +[43] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 102-118. Springer, 2016. 1, 6 +[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 1 +[45] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 1, 6 +[46] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10765-10775, 2021. 1 +[47] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6, 7 +[48] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 3 +[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 5, 8 + +[50] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. Advances in neural information processing systems, 31, 2018. 2 +[51] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2517-2526, 2019. 1 +[52] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7303-7313, 2021. 3 +[53] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. Siamdoge: Domain generalizable semantic segmentation using siamese network. In European Conference on Computer Vision, pages 603-620. Springer, 2022. 2, 6, 7 +[54] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 1 +[55] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2884-2892, 2022. 3, 6, 7 +[56] Liwei Yang, Xiang Gu, and Jian Sun. Generalized semantic segmentation by self-supervised source domain projection and multi-level contrastive learning. arXiv preprint arXiv:2303.01906, 2023. 3, 6, 7 +[57] Jaejun Yoo, Youngjung Uh, Sanghyuk Chun, Byeongkyu Kang, and Jung-Woo Ha. Photorealistic style transfer via wavelet transforms. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9036-9045, 2019. 2 +[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 6 +[59] Fei Yu, Mo Zhang, Hexin Dong, Sheng Hu, Bin Dong, and Li Zhang. Dast: Unsupervised domain adaptation in semantic segmentation based on discriminator attention and self-training. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10754–10762, 2021. 2 +[60] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2100-2110, 2019. 2 +[61] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Do + +main generalization with mixstyle. In International Conference on Learning Representations, 2020. 2 +[62] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018. 1, 2 \ No newline at end of file diff --git a/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/images.zip b/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4762967e5eb8ffe60a3b507c91960d801884f382 --- /dev/null +++ b/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0993954cb406f0074349f9ac5f0d37a6051081431de193a4bab76cf489098ed8 +size 578235 diff --git a/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/layout.json b/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e72539aaec4e858e2293379d04f2b586e3850989 --- /dev/null +++ b/2024/Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning/layout.json @@ -0,0 +1,9723 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 527, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 527, + 140 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 527, + 140 + ], + "type": "text", + "content": "Style Blind Domain Generalized Semantic Segmentation via Covariance Alignment and Semantic Consistence Contrastive Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 160, + 484, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 160, + 484, + 190 + ], + "spans": [ + { + "bbox": [ + 112, + 160, + 484, + 190 + ], + "type": "text", + "content": "Woo-Jin Ahn1 Geun-Yeong Yang1 Hyun-Duck Choi2* Myo-Taeg Lim1* \n1Korea University 2Chonnam National University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 190, + 459, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 190, + 459, + 204 + ], + "spans": [ + { + "bbox": [ + 134, + 190, + 459, + 204 + ], + "type": "inline_equation", + "content": "^{1}\\{wjahn, hggofficial, mlim\\} @korea.ac.kr" + }, + { + "bbox": [ + 134, + 190, + 459, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 190, + 459, + 204 + ], + "type": "text", + "content": "ducky.choi@jnu.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 256, + 289, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 289, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 289, + 605 + ], + "type": "text", + "content": "Deep learning models for semantic segmentation often experience performance degradation when deployed to unseen target domains unidentified during the training phase. This is mainly due to variations in image texture (i.e. style) from different data sources. To tackle this challenge, existing domain generalized semantic segmentation (DGSS) methods attempt to remove style variations from the feature. However, these approaches struggle with the entanglement of style and content, which may lead to the unintentional removal of crucial content information, causing performance degradation. This study addresses this limitation by proposing BlindNet, a novel DGSS approach that blinds the style without external modules or datasets. The main idea behind our proposed approach is to alleviate the effect of style in the encoder whilst facilitating robust segmentation in the decoder. To achieve this, BlindNet comprises two key components: covariance alignment and semantic consistency contrastive learning. Specifically, the covariance alignment trains the encoder to uniformly recognize various styles and preserve the content information of the feature, rather than removing the style-sensitive factor. Meanwhile, semantic consistency contrastive learning enables the decoder to construct discriminative class embedding space and disentangles features that are vulnerable to misclassification. Through extensive experiments, our approach outperforms existing DGSS methods, exhibiting robustness and superior performance for semantic segmentation on unseen target domains. The code is available at https://github.com/root0yang/BlindNet." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 638, + 128, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 128, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 128, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 658, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 658, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 658, + 287, + 696 + ], + "type": "text", + "content": "Semantic segmentation, a technique that classifies each pixel in an image into predefined categories, has garnered significant attention due to its potential applications in vari-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 228, + 547, + 327 + ], + "blocks": [ + { + "bbox": [ + 306, + 228, + 547, + 327 + ], + "lines": [ + { + "bbox": [ + 306, + 228, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 547, + 327 + ], + "type": "image", + "image_path": "f1ff12b29489673484c490594c0b0065fa431db390eaa18e1fb3b77b2196fbd2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 334, + 547, + 380 + ], + "lines": [ + { + "bbox": [ + 305, + 334, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 547, + 380 + ], + "type": "text", + "content": "Figure 1. Comparison of semantic segmentation results between the baseline (DeepLabV3+ with ResNet50 backbone) and our BlindNet. Both models are trained on the source domain (GTAV [43]) and tested on the target domain (Cityscapes [8])." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 401, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 547, + 640 + ], + "type": "text", + "content": "ous fields. Particularly, it plays a crucial role in autonomous driving [1, 18] and robotic systems [34, 35]. Besides, with the advent of large datasets, deep neural networks have emerged as a trending approach for semantic segmentation tasks, achieving impressive results [3, 5, 44, 54]. However, there remains a major bottleneck detailing the meticulous and labor-intensive process of dataset labeling. More particularly, this process not only consumes time but also poses economic challenges [8, 46]. To address this challenge, synthetic datasets have emerged as a compelling alternative. Specifically, these datasets, generated using three-dimensional (3D) rendering techniques, offer vast amounts of easily accessible data, eliminating the need for manual labeling [43, 45]. However, a challenge arises when models trained on synthetic datasets are deployed in real-world scenarios. More precisely, a domain shift problem arises due to style factor discrepancies (e.g. texture, illumination, and image quality) between synthetic and real-world data, which affects the performance of the model, as shown in Fig. 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "To address the domain shift problem, domain adaptation semantic segmentation (DASS) has been introduced [15, 17, 22, 25, 51, 62]. Specifically, DASS aims to bridge the gap between source and target domains by aligning their data distributions. However, a significant limitation of DASS is its dependency on the target domain during train-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 140, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 140, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 140, + 712 + ], + "type": "text", + "content": "*Corresponding Authors" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3616" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "ing. For DASS to function effectively, target domain samples must be available during the training phase." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 98, + 286, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 98, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 98, + 286, + 180 + ], + "type": "text", + "content": "Meanwhile, domain generalized semantic segmentation (DGSS) has been introduced as an alternative approach to tackle the domain shift problem. Unlike DASS, DGSS is trained only with the source domain, aiming to extract domain-invariant features. To achieve this, two main techniques have been employed: domain randomization (DR) and feature normalization (FN)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 182, + 286, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 286, + 289 + ], + "type": "text", + "content": "DR augments the training set by introducing variability, either by altering the image style [41, 60] or by modifying the feature representation [24, 53]. By exposing the model to a wider variety of styles via DR, the network is less likely to overfit to the specific styles present in the training data. Consequently, the robustness of the model is improved, making it more adept at generalizing to new, unseen domains. Nevertheless, a crucial limitation of DR is its significant dependence on auxiliary domains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 291, + 286, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 286, + 422 + ], + "type": "text", + "content": "FN methods, converse to DR, regularize the features to prevent the model from overfitting to the distinct styles or characteristics of the training data. This is achieved by removing domain-specific style information using feature statistics, such as instance normalization [38] or whitening transformation [7, 19, 39, 42]. While these approaches effectively remove style-related information, they simultaneously pose the challenge of removing semantic content because content and style information are entangled. Consequently, the model fails to capture the essential patterns or features required for accurate segmentation prediction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 424, + 286, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 286, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 286, + 686 + ], + "type": "text", + "content": "To address this problem, we propose BlindNet, a model that blinds the style within the encoder and improves the robustness of the decoder, without requiring auxiliary datasets or external modules. Specifically, our proposed BlindNet consists of two components: covariance alignment for the encoder and semantic consistency contrastive learning for the decoder. Precisely, the covariance alignment facilitates the generation of style-invariant features with the proposed covariance matching loss function (CML) and the cross-covariance loss function (CCL). Specifically, CML mitigates the effects of style variations, while CCL focuses on preserving content information, effectively addressing the prevalent content information loss observed in the FN method. To further improve the generalization ability, we develop semantic consistency contrastive learning, which consists of class-wise contrastive learning (CWCL) and semantic disentanglement contrastive learning (SDCL). Particularly, the CWCL constructs a discriminative class embedding space, while SDCL disentangles features of similar classes that often lead to prediction errors. Extensive experiments across various datasets demonstrate that the proposed BlindNet outperforms existing DGSS methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 688, + 244, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 688, + 244, + 699 + ], + "spans": [ + { + "bbox": [ + 59, + 688, + 244, + 699 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "content": "- We propose a covariance alignment within the encoder," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 545, + 120 + ], + "type": "text", + "content": "comprising CML and CCL. Specifically, the CML aims to mitigate the effects of style variations, while CCL ensures the preservation of content information, together facilitating the generation of style-agnostic features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 121, + 545, + 227 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 121, + 545, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 121, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 121, + 545, + 190 + ], + "type": "text", + "content": "- We propose semantic consistency contrastive learning within the decoder that comprises CWCL and SDCL, utilizing segmentation masks. Specifically, CWCL generates discriminative embeddings, whereas SDCL disentangles features of similar classes, enhancing the robustness of the model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 192, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 192, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 192, + 545, + 227 + ], + "type": "text", + "content": "- Through extensive experiments, we demonstrate the superiority of our approach in DGSS, without the need to alter the network architecture or rely on external datasets." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 239, + 397, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 397, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 397, + 251 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 258, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 545, + 426 + ], + "type": "text", + "content": "Domain adaptation and generalization for semantic segmentation. Domain adaptation (DA) aims at minimizing the distribution discrepancy between different domains, enabling a model to generalize from a source to a target domain. For DASS, adversarial training and cross-domain self-training strategies are commonly used. Particularly, adversarial-based methods [15, 29] employ generative adversarial networks [11] to close the feature distribution gap between source and target domains. Meanwhile, cross-domain self-training methods [16, 37, 59, 62] generate pseudo-labels for target domain data using pre-trained models, and employ them as training data, thereby expanding the training data and reducing distribution differences between the source and target domain." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": "Domain generalization (DG) methods [26, 28, 30, 50, 61] aim to improve the generalization ability of the model without accessing the target domain during training. Since the difference in style of the image is the main cause of the disparity, most existing domain generalized semantic segmentation methods utilize the style information of the image for domain-invariant learning. Feature statistics (e.g. mean, variance, covariance, gram matrix, etc.) which are commonly used in style transfer [10, 21, 27, 57] are employed to capture the style information. Interestingly, existing DGSS methods can be separated into two parts: i) domain randomization to expand the distribution of style or ii) feature normalization to remove style." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "DR involves randomizing either the image or its features through stylization to learn domain-invariant features from various styles. For example, Peng et al. [42] extended the source domain data by stylizing images in the style of unreal paintings. Similarly, Yue et al. [60] and Huang et al. [22] attempted to enhance generalization by synthesizing images with diverse styles in the image space. In another study, Lee et al. [24] adopted ImageNet data [9] as wild data and performed randomization via synthesis in the feature space. Meanwhile, Wu et al. [53] diversified the trainable feature space by mixing the statistics of the feature and its color-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3617" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 183, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 183, + 83 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 183, + 83 + ], + "type": "text", + "content": "jittered feature with Ada-IN [21]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 87, + 289, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 87, + 289, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 87, + 289, + 387 + ], + "type": "text", + "content": "FN methods aim to remove domain-specific styles from features, extracting only domain-invariant content. For instance, Pan et al. [39] first attempted the DGSS method, combining batch normalization (BN) [23] and instance normalization (IN) [48] in the network layer. While BN preserves the content information within discriminative features, IN focuses on removing domain-specific style information from features. In a study, Choi et al. [7] addressed the limitations of previous whitening transformation [6, 31] that can eliminate the content information from the feature. Specifically, they proposed an instance-selective whitening approach designed to remove covariance components that are sensitive to domain shifts. Peng et al. [41] developed semantic-aware normalization that performs on class-wise and semantic-aware whitening that aligns channels based on the prediction through group whitening transformation [6]. In a study, Xu et al. [55] introduced the prior guided attention module and guided feature whitening to re-calibrate the feature and remove domain-specific style effects, respectively. Unlike the FN methods that directly remove the style component, our work explores a covariance alignment method that mitigates the effect of the style's effect while preserving the content information. Our method achieves the DGSS without any additional modules or auxiliary datasets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "text", + "content": "Contrastive Learning. Contrastive learning aims to learn representations by maximizing the similarity between positive pairs of samples while minimizing the similarity between negative pairs. In recent years, it has attracted significant attention for its effectiveness in learning discriminative representations across various tasks [2, 4, 12, 14]. Oord et al. [36] were the first to introduce the infoNCE loss, a type of contrastive loss function designed for self-contrastive learning. In a work, Park et al. [40] introduced patch-level contrastive learning for image translation, using co-located patches as positive pairs and spatially distant patches as negatives to maintain image context. For the semantic segmentation task, Wang et al. [52] introduced class-wise contrastive learning to aid the model in learning the embedding space of each class. Specifically, they sampled the classes existing in the images and applied contrastive learning based on the class label. For DGSS, Lee et al. [24] adopted contrastive learning to learn the ImageNet information in their model. Specifically, they set the ImageNet data as wild and applied the contrastive learning method by setting the wild-stylized feature and its closest wild content as positive samples. In another study, Yang et al. [56] developed multi-level contrastive learning, which designed instance prototypes and class prototypes for contrastive learning. Specifically, they sample each class's pixel features and apply contrastive learning with a transition-probability ability matrix. Unlike recent DGSS works that embed the origi" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": "nal image, we define contrastive learning for the augmented image. Specifically, the proposed method builds a robust embedding space by preserving the semantic consistency of the feature representation across various domains." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 130, + 362, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 130, + 362, + 143 + ], + "spans": [ + { + "bbox": [ + 305, + 130, + 362, + 143 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "text", + "content": "The goal of the proposed method is to train a segmentation model " + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "text", + "content": " on a given source domain " + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "text", + "content": " and generalize well to the unseen target domain " + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "text", + "content": ". Precisely, the source domain " + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "inline_equation", + "content": "S = \\{(x,y)\\}" + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "text", + "content": " contains a paired image " + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "text", + "content": " and segmentation label " + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "inline_equation", + "content": "y \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "text", + "content": " denote the height of the image, the width of the image, and the class number of the segmentation map, respectively. The model " + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "text", + "content": " takes an original " + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "text", + "content": " and its augmented counterpart " + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 304, + 151, + 547, + 281 + ], + "type": "text", + "content": ", which have the same content but different styles, and uses the feature information to enhance its generalization ability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 282, + 547, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 547, + 379 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 547, + 379 + ], + "type": "text", + "content": "As shown in Fig. 2, our method leverages the feature information through two main approaches: covariance alignment and semantic consistency contrastive learning. Specifically, the covariance matching ensures that features, having different styles but the same content, contain similar information. Additionally, the semantic consistency contrastive learning embeds the generalized features into discriminative representation based on the segmentation label." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 385, + 435, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 385, + 435, + 398 + ], + "spans": [ + { + "bbox": [ + 305, + 385, + 435, + 398 + ], + "type": "text", + "content": "3.1. Covariance Alignment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 403, + 547, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 547, + 499 + ], + "type": "text", + "content": "The domain shift in semantic segmentation results from changes in the visual characteristics of the image, known as style. This style information is typically detected in the shallow layers of networks [38], which are the encoders of the segmentation model. Based on this understanding, our method targets the encoder to handle style variations by employing the proposed covariance matching loss and cross-covariance loss function." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 514, + 424, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 514, + 424, + 526 + ], + "spans": [ + { + "bbox": [ + 305, + 514, + 424, + 526 + ], + "type": "text", + "content": "Covariance Matching Loss" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "text", + "content": "To train the network to uniformly recognize various styles without removing content information, we introduce the CML. Specifically, the loss aims to minimize the difference between covariance matrices derived from different styles of image features. Given an image pair " + }, + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "inline_equation", + "content": "(x,x_{a})" + }, + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "text", + "content": ", the features from " + }, + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "text", + "content": " block of the encoder are represented as " + }, + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "inline_equation", + "content": "F^i\\in \\mathbb{R}^{(H^i\\times W^i)\\times C^i}" + }, + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "inline_equation", + "content": "F_{a}^{i}\\in \\mathbb{R}^{(H^{i}\\times W^{i})\\times C^{i}}" + }, + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "text", + "content": ", respectively. Further, following the methodologies of [7, 41], we compute the covariance matrices using instance normalized features [48], which ensures consistent scaling across features. The feature maps are normalized and flattened into " + }, + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\bar{F}" + }, + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\bar{F}_a\\in \\mathbb{R}^{(HW)\\times C}" + }, + { + "bbox": [ + 304, + 533, + 547, + 677 + ], + "type": "text", + "content": ", which are given by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 342, + 685, + 545, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 685, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 342, + 685, + 545, + 712 + ], + "type": "interline_equation", + "content": "\\bar {F} = \\frac {(F - \\mu (F))}{\\sigma (F)}, \\bar {F} _ {a} = \\frac {(F _ {a} - \\mu (F _ {a}))}{\\sigma (F _ {a})}, \\tag {1}", + "image_path": "2f353ece5e5c0554e195dfed7507dedd6c429565cf5ab2cfcd5206a8b77ea2d7.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3618" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 545, + 320 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 545, + 320 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 545, + 320 + ], + "type": "image", + "image_path": "5fcbbb724d3720e68e61b5e9b5a91dabb76d3372b3e54fd809de935bffa75425.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 327, + 546, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 327, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 546, + 361 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed BlindNet. The network processes a pair of images - the original image " + }, + { + "bbox": [ + 46, + 327, + 546, + 361 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 327, + 546, + 361 + ], + "type": "text", + "content": " and its augmented counterpart " + }, + { + "bbox": [ + 46, + 327, + 546, + 361 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 46, + 327, + 546, + 361 + ], + "type": "text", + "content": ". It employs covariance alignment to treat encoder features and utilizes semantic consistency contrastive learning for the processing of decoder features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 381, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 381, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mu (\\cdot)\\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 46, + 381, + 287, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 381, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 46, + 381, + 287, + 431 + ], + "type": "text", + "content": " denote the mean and standard deviation of the features. Utilizing the normalized features, we evaluate the covariance matrices for the original and augmented image features as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 438, + 287, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 438, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 75, + 438, + 287, + 456 + ], + "type": "interline_equation", + "content": "\\Sigma_ {x, x} ^ {i} = \\bar {F} ^ {i} ^ {T} \\cdot \\bar {F} ^ {i}, \\quad \\Sigma_ {x _ {a}, x _ {a}} ^ {i} = \\bar {F} _ {a} ^ {i} ^ {T} \\cdot \\bar {F} _ {a} ^ {i}. \\tag {2}", + "image_path": "80800837feca206a60efb11057ad539abdff1899eb805467641e09ae1c4d6fe9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 464, + 287, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 500 + ], + "type": "text", + "content": "The CML is then formulated to align these covariance matrices, ensuring that the network maintains consistency in the presence of style variations. The CML is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 101, + 510, + 287, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 510, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 101, + 510, + 287, + 540 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C M} = \\sum_ {i = 1} ^ {n _ {e}} \\| \\Sigma_ {x, x} ^ {i} - \\Sigma_ {x _ {a}, x _ {a}} ^ {i} \\| _ {2}, \\tag {3}", + "image_path": "4623802dc1bf53f71b8f6dcab09a92b4542b0c952c8c829c45f3c375b07364ba.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 548, + 268, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 268, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 268, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 548, + 268, + 559 + ], + "type": "inline_equation", + "content": "n_e" + }, + { + "bbox": [ + 47, + 548, + 268, + 559 + ], + "type": "text", + "content": " denotes the number of blocks of the encoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 575, + 145, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 145, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 145, + 586 + ], + "type": "text", + "content": "Cross-covariance Loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "text", + "content": "While the CML effectively aligns the internal distributions of features within the same image, it does not fully account for the direct correlations across paired images. To complement this, we introduce CCL, which aims to encode the consistent content information of an image pair " + }, + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "inline_equation", + "content": "(x,x_{a})" + }, + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "text", + "content": " by utilizing the cross-covariance of the image pair. Given the normalized feature pair " + }, + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "inline_equation", + "content": "(\\bar{F}^i,\\bar{F}_a^i)" + }, + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "text", + "content": ", the cross-covariance of the feature pair can be expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 697, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 697, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 127, + 697, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\Sigma_ {x, x _ {a}} ^ {i} = \\bar {F} ^ {i} ^ {T} \\cdot \\bar {F} _ {a} ^ {i}. \\tag {4}", + "image_path": "ee5d64b188f826f16a90b60969e58f06425daed783148e4bf8d85de928df858d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 382, + 545, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 382, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 305, + 382, + 545, + 465 + ], + "type": "text", + "content": "The cross-covariance is expected to exhibit an identity matrix, as the feature pair should contain identical information. Nonetheless, the proposed CCL converges only the diagonal component of the covariance matrix to one. This is to prevent the drawbacks of the existing FN methods [7, 42] from removing content information. The CCL function is thus defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 358, + 469, + 545, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 469, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 358, + 469, + 545, + 500 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C C} = \\sum_ {i = 1} ^ {n _ {e}} \\| \\operatorname {d i a g} \\left(\\Sigma_ {x, x _ {a}} ^ {i}\\right) - \\mathbb {1} \\| _ {2}, \\tag {5}", + "image_path": "c282b8883b21f691d2d00a55b50b649a7041d17c79723471fdc4479688f88326.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 506, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 506, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 506, + 545, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 506, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathrm{diag}(\\Sigma_c) \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 305, + 506, + 545, + 540 + ], + "type": "text", + "content": " denotes the column vector comprising diagonal elements of " + }, + { + "bbox": [ + 305, + 506, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\Sigma_{x,x_a}^i" + }, + { + "bbox": [ + 305, + 506, + 545, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 506, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbb{1} \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 305, + 506, + 545, + 540 + ], + "type": "text", + "content": " denotes the one vector." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 548, + 533, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 533, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 533, + 560 + ], + "type": "text", + "content": "3.2. Semantic Consistence Contrastive Learning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 566, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 566, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 305, + 566, + 545, + 649 + ], + "type": "text", + "content": "While the encoder focuses on generating style-blinded features, the decoder aims to improve the robustness of the segmentation prediction against domain shifts. For the decoder, we employ a contrastive learning approach, which has demonstrated effectiveness in extracting discriminative features [4]. Specifically, we utilize the InfoNCE loss [36], which is formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 652, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 652, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 314, + 652, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I N} (a, p, n) = - \\log \\left(\\frac {e ^ {(a \\cdot p / \\tau)}}{e ^ {(a \\cdot p / \\tau)} + \\sum_ {n} ^ {N ^ {-}} e ^ {(a \\cdot n / \\tau)}}\\right) \\tag {6}", + "image_path": "725b5d89c90a0264a3a8a74c70269d848ea2ebc79737ddbcaf18346d253c3ebb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "a, p, n" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "N^{-}" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " denote anchor, positive sample, negative sample, and negative sample set, respectively." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3619" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 68, + 534, + 194 + ], + "blocks": [ + { + "bbox": [ + 60, + 68, + 534, + 194 + ], + "lines": [ + { + "bbox": [ + 60, + 68, + 534, + 194 + ], + "spans": [ + { + "bbox": [ + 60, + 68, + 534, + 194 + ], + "type": "image", + "image_path": "f9fb00733b6b0939d8e914bad891b9b9d71654c5baf4d18638084e017a3d791f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 546, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 546, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 546, + 236 + ], + "type": "text", + "content": "Figure 3. Illustration of semantic consistency contrastive learning: The mask " + }, + { + "bbox": [ + 46, + 201, + 546, + 236 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 201, + 546, + 236 + ], + "type": "text", + "content": " represents the error mask derived from the augmented segmentation map. CWCL conducts contrastive learning by sampling per segmentation class and SDCL conducts contrastive learning based on the " + }, + { + "bbox": [ + 46, + 201, + 546, + 236 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 201, + 546, + 236 + ], + "type": "text", + "content": ". Both methods share a projection head " + }, + { + "bbox": [ + 46, + 201, + 546, + 236 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 201, + 546, + 236 + ], + "type": "text", + "content": " for the semantic representation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 255, + 288, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 288, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 288, + 364 + ], + "type": "text", + "content": "To achieve consistent feature representation in DGSS across various styles, we introduce semantic consistency contrastive learning. Specifically, the anchor is derived from the augmented image " + }, + { + "bbox": [ + 46, + 255, + 288, + 364 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 46, + 255, + 288, + 364 + ], + "type": "text", + "content": ", while the positive sample is extracted from the corresponding pixel of the original image " + }, + { + "bbox": [ + 46, + 255, + 288, + 364 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 255, + 288, + 364 + ], + "type": "text", + "content": ". Our method consists of two main components based on the negative sample as shown in Fig. 3: class-wise contrastive learning and semantic disentanglement contrastive learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 389, + 189, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 389, + 189, + 402 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 189, + 402 + ], + "type": "text", + "content": "Class-wise Contrastive Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "text", + "content": "Our CWCL aims to build a discriminative embedding space for each segmentation class using different classes of the original image as negatives. Given an image pair " + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "inline_equation", + "content": "(x, x_{a})" + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "text", + "content": ", the features from the " + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "text", + "content": " block of the decoder at pixel position " + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "inline_equation", + "content": "(m, n)" + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "text", + "content": " are denoted as " + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "inline_equation", + "content": "F_{(m, n)}^{j}" + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "inline_equation", + "content": "F_{a, (m, n)}^{j} \\in \\mathbb{R}^{(1 \\times 1) \\times C^{j}}" + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "inline_equation", + "content": "C^{j}" + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "text", + "content": " indicates the channel length of the feature. As mentioned above, we take " + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "inline_equation", + "content": "F_{a, (m, n)}^{j}" + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "text", + "content": " as the anchor and " + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "inline_equation", + "content": "F_{(m, n)}^{j}" + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "text", + "content": " as the positive sample, since they represent the same content at the corresponding spatial location. To obtain the negative samples from " + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "inline_equation", + "content": "F^{j}" + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "text", + "content": ", we leverage the resized segmentation class label " + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "inline_equation", + "content": "y^{j} \\in \\mathbb{R}^{(H^{j} \\times W^{j}) \\times C}" + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "text", + "content": " to collect the different class samples. The samples are passed through the projection head, denoted as " + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "text", + "content": ", resulting in the projected features " + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\tilde{F}" + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{a}" + }, + { + "bbox": [ + 46, + 411, + 288, + 589 + ], + "type": "text", + "content": ". We define our CWCL as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 601, + 287, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 601, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 70, + 601, + 287, + 633 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C W C L} = \\sum_ {j} ^ {n _ {d}} \\mathcal {L} _ {I N} \\left(\\tilde {F} _ {a, (m, n)} ^ {j}, \\tilde {F} _ {(m, n)} ^ {j}, \\tilde {F} _ {(p, q)} ^ {j}\\right) \\tag {7}", + "image_path": "a03e67a09f13f81c5e478540de593e4bde07984248bd77bd514e2a4a30e14745.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 75, + 635, + 263, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 635, + 263, + 653 + ], + "spans": [ + { + "bbox": [ + 75, + 635, + 263, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 75, + 635, + 263, + 653 + ], + "type": "inline_equation", + "content": "(p,q)\\in \\{(p,q)\\in P|y_{(p,q)}^j\\neq y_{(m,n)}^j\\}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "content": "The set " + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "content": " represents all pixel positions in feature " + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "inline_equation", + "content": "F^j" + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "content": ", with dimensions " + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "inline_equation", + "content": "H^j \\times W^j" + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "content": " corresponding to the height and width. Additionally, " + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "inline_equation", + "content": "n_d" + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "content": " denotes the total number of blocks in the decoder." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 331, + 252, + 425, + 342 + ], + "blocks": [ + { + "bbox": [ + 331, + 252, + 425, + 342 + ], + "lines": [ + { + "bbox": [ + 331, + 252, + 425, + 342 + ], + "spans": [ + { + "bbox": [ + 331, + 252, + 425, + 342 + ], + "type": "image", + "image_path": "364e0bcd00ec452f09a91d3c336a3abe80fbdab1f326b463d00f6dc6623bce24.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 349, + 546, + 393 + ], + "lines": [ + { + "bbox": [ + 305, + 349, + 546, + 393 + ], + "spans": [ + { + "bbox": [ + 305, + 349, + 546, + 393 + ], + "type": "text", + "content": "Figure 4. t-SNE [49] visualization comparing scenarios with and without " + }, + { + "bbox": [ + 305, + 349, + 546, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SCCL}" + }, + { + "bbox": [ + 305, + 349, + 546, + 393 + ], + "type": "text", + "content": ". In (b), the application of SCCL results in a clear separation between the sidewalk (pink), the road (purple), and the building (gray)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 427, + 261, + 522, + 342 + ], + "blocks": [ + { + "bbox": [ + 427, + 261, + 522, + 342 + ], + "lines": [ + { + "bbox": [ + 427, + 261, + 522, + 342 + ], + "spans": [ + { + "bbox": [ + 427, + 261, + 522, + 342 + ], + "type": "image", + "image_path": "1185b7eb568978d73c2f83c1c0f9c882983f91b10edd00fd456149b4d75b6763.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 414, + 516, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 414, + 516, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 414, + 516, + 426 + ], + "type": "text", + "content": "Semantic Disentanglement Contrastive Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "text", + "content": "Domain shifts can lead to the entanglement of similar classes, causing the model to misclassify, as illustrated in Fig. 4. To mitigate this issue, we introduce the SDCL, specifically designed to disentangle the feature " + }, + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "text", + "content": " that has been misclassified, making it closer to the correct class and far from the misclassified class to achieve effective disentanglement. To further ensure consistent feature space and capture the semantic meaning, we share the projection head " + }, + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "text", + "content": " used in the CWCL loss. Given the predicted segmentation map of the augmented image, represented as " + }, + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "inline_equation", + "content": "\\hat{y}_a = \\varphi (x_a)" + }, + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "text", + "content": ", we resize it to " + }, + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "inline_equation", + "content": "\\hat{y}_{a,(m,n)}^j\\in \\mathbb{R}^{(H^j\\times W^j)\\times C}" + }, + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "text", + "content": ". Similarly, " + }, + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "inline_equation", + "content": "y_{(m,n)}^j" + }, + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "text", + "content": " represents the ground truth segmentation map. Using these segmentation maps, we set the anchor at positions where " + }, + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "inline_equation", + "content": "\\hat{y}_{a,(m,n)}^j\\neq y_{(m,n)}^j" + }, + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "text", + "content": ". Negative samples are selected from the augmented image features corresponding to the anchor's misclassified class. The samples go through the projection head " + }, + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "text", + "content": ". Our SDCL loss is defined as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 643, + 545, + 676 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 643, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 321, + 643, + 545, + 676 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S D C L} = \\sum_ {j} ^ {n _ {d}} \\mathcal {L} _ {I N} \\left(\\tilde {F} _ {a, (m, n)} ^ {j}, \\tilde {F} _ {(m, n)} ^ {j}, \\tilde {F} _ {a, (r, s)} ^ {j}\\right) \\tag {8}", + "image_path": "a05976ade526f53a5b38d528532f477caf80d7b84c4aa7d99f963fcd033f24e2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 333, + 679, + 517, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 679, + 517, + 695 + ], + "spans": [ + { + "bbox": [ + 333, + 679, + 517, + 695 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 333, + 679, + 517, + 695 + ], + "type": "inline_equation", + "content": "(r,s)\\in \\{(r,s)\\in P|y_{(r,s)}^j = \\hat{y}_{(m,n)}^j\\}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "Finally, combining the cross-entropy segmentation loss" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3620" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 69, + 548, + 290 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 548, + 290 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 548, + 290 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 548, + 290 + ], + "type": "table", + "html": "
BackboneMethodsExternalTrained on GTAV (G)Trained on Cityscapes (C)
DatasetModuleCBMSBMSG
ResNet50 [13]Baseline [3]--28.9525.1428.1826.2344.9651.6823.2942.55
IBN-Net [38]--33.8532.3037.7527.9048.5657.0426.1445.06
RobustNet [7]--37.3135.2040.3328.3050.7358.6426.2045.00
SiamDoGe [53]--42.9637.5440.6428.3451.5359.0026.6745.08
DIRL [55]-✓41.0439.1541.60-51.80-26.5046.52
WildNet [24]✓-44.6238.4246.0931.3450.9458.7927.9547.01
SANSAW [42]-✓39.7537.3441.8630.7952.9559.8128.3247.28
SPC [20]-✓44.1040.4645.51-----
DPCL [56]-✓44.7440.5946.3330.8150.9758.5925.8546.00
Ours--45.7241.3247.0831.3951.8460.1828.5147.97
ShuffleNetV2 [32]Baseline [3]--25.5622.1728.6023.3336.8443.1321.5636.95
IBN-Net [38]--27.1031.8234.8925.5641.8946.3522.9940.91
RobustNet [7]--30.9832.0635.3124.3141.9446.9722.8240.17
SiamDoGe [53]--34.4034.2335.8721.9542.6147.4823.1340.93
DIRL [55]-✓31.8832.5736.12-42.55-23.7441.23
DPCL [56]-✓36.6634.3539.9222.6643.9048.9522.4741.07
Ours--38.5634.5140.1125.6444.2249.6923.5441.10
", + "image_path": "644a98889183e0ef6b889a9e2199e4b2bb590eeae1aa7f4a56ec1ca5acd5517c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 297, + 546, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 297, + 546, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 546, + 331 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of mIoU (%) between DGSS methods. External dataset denotes the necessity of an auxiliary dataset during training and External module denotes the requirement of an additional module during inference. G, C, B, M, and S denote GTAV, Cityscapes, BDD100K, Mapillary, and SYNTHIA, respectively. The best and second-best results are bolded and underlined, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 352, + 274, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 352, + 274, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 352, + 274, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 47, + 352, + 274, + 364 + ], + "type": "text", + "content": " with other loss components, the total is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 371, + 287, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 371, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 83, + 371, + 287, + 398 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {C E} + \\omega_ {1} \\mathcal {L} _ {C M} + \\omega_ {2} \\mathcal {L} _ {C C} \\tag {9} \\\\ + \\omega_ {3} \\mathcal {L} _ {C W C L} + \\omega_ {4} \\mathcal {L} _ {S D C L} \\\\ \\end{array}", + "image_path": "d92344071967dd7e67b709af207e608ae928b88bb5bbe62ab4c3b4ef5ba84a24.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\omega_{1},\\omega_{2},\\omega_{3}" + }, + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\omega_{4}" + }, + { + "bbox": [ + 46, + 406, + 287, + 430 + ], + "type": "text", + "content": " denote the weighting factor of each loss functions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 440, + 123, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 440, + 123, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 440, + 123, + 453 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 460, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 509 + ], + "type": "text", + "content": "In this section, we describe the implementation details, the experimental setup for comparison with existing DGSS methods, and the ablation study conducted to further validate the effectiveness of our approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 514, + 180, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 180, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 180, + 528 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "text", + "content": "We adopt DeepLabV3+ [3] for the segmentation architecture and use ResNet-50 [13], ShuffleNetV2 [32], and MobileNetV2 [47] as the backbone network of the segmentation network. The model is trained for 40K iterations with a batch size of 8 using the SGD optimizer, which has a momentum of 0.9 and a weight decay of 5e-4. We employ a polynomial learning rate schedule with an initial rate of 1e-2 and a power of 0.9. For the simulation of domain shift, we augment the image " + }, + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "text", + "content": " using strong color jittering transformation similar to [7]. The weighting parameters of (9), " + }, + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\omega_{1}" + }, + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\omega_{2}" + }, + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\omega_{3}" + }, + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\omega_{4}" + }, + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "text", + "content": ", are set as 0.2, 0.2, 0.3, and 0.3 respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 671, + 110, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 110, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 110, + 682 + ], + "type": "text", + "content": "4.2. Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "We use two synthetic datasets (GTA [43] and SYNTHIA [45]), and three real-world datasets (Cityscapes [8]," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 352, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 388 + ], + "type": "text", + "content": "BDD-100K [58], and Mapillary [33]) for the experiment. All segmentation labels are evaluated based on 19 object categories." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 388, + 546, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 546, + 446 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 546, + 446 + ], + "type": "text", + "content": "GTAV (G) [43] is a large-scale dataset generated from the Grand Theft Auto V (GTAV) game engine. It comprises 24,966 images, split into 12,403 for training, 6,382 for validation, and 6,181 for testing with a resolution of " + }, + { + "bbox": [ + 304, + 388, + 546, + 446 + ], + "type": "inline_equation", + "content": "1914 \\times 1052" + }, + { + "bbox": [ + 304, + 388, + 546, + 446 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 448, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 547, + 495 + ], + "type": "text", + "content": "SYNTHIA (S) [45] is a virtual, photo-realistic urban scene dataset comprising 9,400 images with a resolution of " + }, + { + "bbox": [ + 304, + 448, + 547, + 495 + ], + "type": "inline_equation", + "content": "960 \\times 720" + }, + { + "bbox": [ + 304, + 448, + 547, + 495 + ], + "type": "text", + "content": ". Among these, 2,820 images are designated for evaluation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 496, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 547, + 555 + ], + "type": "text", + "content": "Cityscapes (C) [8] is a large-scale urban scene dataset captured from 50 cities, primarily in Germany. Particularly, it contains 5,000 high-resolution images with a resolution of " + }, + { + "bbox": [ + 304, + 496, + 547, + 555 + ], + "type": "inline_equation", + "content": "2048 \\times 1024" + }, + { + "bbox": [ + 304, + 496, + 547, + 555 + ], + "type": "text", + "content": ". The dataset is divided into 2,975 images for training, 500 for validation, and 1,525 for testing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "text", + "content": "BDD-100K (B) [58] is another real-world urban scene dataset that contains more diverse 10000 urban driving scene images with a resolution of " + }, + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "text", + "content": ". Specifically, the validation split (1,000 images) is used for evaluation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 603, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 603, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 545, + 651 + ], + "type": "text", + "content": "Mapillary (M) [33] contains 25,000 images with a minimum resolution of " + }, + { + "bbox": [ + 304, + 603, + 545, + 651 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 304, + 603, + 545, + 651 + ], + "type": "text", + "content": ", collected from various locations worldwide. Specifically, the validation split of 2,000 images is used for evaluation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 659, + 484, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 659, + 484, + 672 + ], + "spans": [ + { + "bbox": [ + 305, + 659, + 484, + 672 + ], + "type": "text", + "content": "4.3. Comparison with DGSS methods" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "text", + "content": "We compare our methods with other state-of-the-art DGSS methods: Baseline (DeepLabV3+ [3] trained on the source domain), IBN-Net [38], RobustNet [7], SiamDoGe [53]," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3621" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 286, + 170 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 286, + 170 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 286, + 170 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 286, + 170 + ], + "type": "table", + "html": "
MethodsExternal ModuleTrained on GTAV (G)
CBMMean
Baseline [3]25.9425.7326.4526.04
IBN-Net [38]30.1427.6627.0728.29
RobustNet [7]30.8630.0530.6730.52
SiamDoGe [53]34.1534.5032.3433.67
DIRL [55]✓34.6732.7834.3133.92
DPCL [56]✓37.5735.4540.3037.77
Ours37.6636.1040.4038.05
", + "image_path": "a8c1edacd555b9aeb544e8b1e27336634ab38752437a712ef806a02175809d01.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 222, + 286, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 222, + 286, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 286, + 354 + ], + "type": "text", + "content": "DIRL [55], WildNet [24], SANSAW [42], SPC [20], and DPCL [56]. To evaluate the generalization ability of the model on arbitrary unseen domains, we conduct the experiment on two scenarios: i) trained on GTAV, tested on Cityscapes, BDD-100K, and Mapillary, and ii) trained on Cityscapes, tested on BDD-100K, Mapillary, and SYNTHIA. The quantitative results are computed with mean intersection over union (mIoU). Additionally, we compared the method trained on the backbone of ResNet-50 [13], ShuffleNetV2 [32], and MoblieNet [47], pre-trained on ImageNet [9]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 371, + 206, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 206, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 206, + 383 + ], + "type": "text", + "content": "Quantitative and Qualitative Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 390, + 286, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 286, + 544 + ], + "type": "text", + "content": "Table 1 summarizes the quantitative results. Our method outperforms all other methods when trained on GTAV, using ResNet-50 as the backbone. When compared with FN methods that remove domain-specific styles, we demonstrate that our approach minimizes the loss of content information. We also show that our method effectively shows generalization ability when trained on Cityscapes. We further evaluate our methods with different backbones, showing the wide applicability of our method. When trained with ShuffleNetV2, our method achieves the first or second-best performance among unseen target domains. Table 2 shows the results of our method trained on GTAV with MobileNetV2, demonstrating the superiority of our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 286, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 286, + 641 + ], + "type": "text", + "content": "For qualitative evaluation, we compare the visual result between DGSS methods and ours. As depicted in Fig. 5, our method demonstrates superior results compared to other approaches, particularly in its overall prediction accuracy. Notably, our proposed techniques enable distinct prediction of features such as on road and sidewalk, yielding clearer segmentation boundaries. Please refer to the supplementary material for more qualitative results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 658, + 170, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 170, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 170, + 670 + ], + "type": "text", + "content": "Computational cost analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "To confirm that our approach does not incur additional computational overhead, we provide the number of parameters, GFLOPS, and average inference time of each method. As" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 545, + 156 + ], + "blocks": [ + { + "bbox": [ + 47, + 178, + 286, + 200 + ], + "lines": [ + { + "bbox": [ + 47, + 178, + 286, + 200 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 286, + 200 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of mIoU " + }, + { + "bbox": [ + 47, + 178, + 286, + 200 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 47, + 178, + 286, + 200 + ], + "type": "text", + "content": " using MobileNetV2 [47] backbone trained on the GTAV dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 70, + 545, + 156 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 156 + ], + "type": "table", + "html": "
MethodsExternal ModuleParams (M)GFLOPSTime (ms)
Baseline [13]45.08277.7710.01
SANSAW [42]✓25.63421.8668.96
SPC [20]✓45.22286.0912.24
DIRL [55]✓45.41278.1111.69
DPCL [56]✓56.461188.64823.78
Ours45.08277.7810.03
", + "image_path": "cde9f4931872aa504e6d8ec0edfca433b12f3ceb3ea1a2ee335bce871e1abc73.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 220, + 545, + 311 + ], + "blocks": [ + { + "bbox": [ + 304, + 164, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 304, + 164, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 304, + 164, + 545, + 208 + ], + "type": "text", + "content": "Table 3. Computational cost comparison conducted using DeepLabV3+ with a ResNet-50 backbone on an NVIDIA Tesla V100 GPU with an image resolution of " + }, + { + "bbox": [ + 304, + 164, + 545, + 208 + ], + "type": "inline_equation", + "content": "2048 \\times 1024" + }, + { + "bbox": [ + 304, + 164, + 545, + 208 + ], + "type": "text", + "content": ". Inference time is averaged over 400 trials." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 220, + 545, + 311 + ], + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 311 + ], + "type": "table", + "html": "
LCMCCLCWCLSDCLCBM
28.9525.1428.18
✓38.0836.6540.62
✓✓40.4237.8143.91
✓42.0338.2744.02
✓✓43.1638.5945.38
✓✓✓43.1738.2344.84
✓✓✓✓45.7241.3247.08
", + "image_path": "31a215e998aaa44615ecc9f72d54f3cf3b11c9da13b663716cf47cf64d5fa051.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 319, + 545, + 364 + ], + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 364 + ], + "type": "text", + "content": "Table 4. Ablation study on proposed losses. The experiments were conducted using DeepLabV3+ with ResNet-50 backbone, trained on the GTAV dataset. The losses are detailed in " + }, + { + "bbox": [ + 304, + 319, + 545, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CM}" + }, + { + "bbox": [ + 304, + 319, + 545, + 364 + ], + "type": "text", + "content": ": (3), " + }, + { + "bbox": [ + 304, + 319, + 545, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CC}" + }, + { + "bbox": [ + 304, + 319, + 545, + 364 + ], + "type": "text", + "content": ": (5), " + }, + { + "bbox": [ + 304, + 319, + 545, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CWCL}" + }, + { + "bbox": [ + 304, + 319, + 545, + 364 + ], + "type": "text", + "content": ": (7), " + }, + { + "bbox": [ + 304, + 319, + 545, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SDCL}" + }, + { + "bbox": [ + 304, + 319, + 545, + 364 + ], + "type": "text", + "content": ": (8)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 386, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 545, + 422 + ], + "type": "text", + "content": "detailed in Table 3, our method operates comparably to baseline models by learning features intrinsically without adopting a separate module." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 431, + 406, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 406, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 406, + 442 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 449, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 545 + ], + "type": "text", + "content": "In this subsection, we conducted a series of ablation studies to demonstrate the individual contribution and effectiveness of each component within our method. Specifically, we investigate the impact of the following components: " + }, + { + "bbox": [ + 304, + 449, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CM}" + }, + { + "bbox": [ + 304, + 449, + 545, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 449, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CC}" + }, + { + "bbox": [ + 304, + 449, + 545, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 449, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CWCL}" + }, + { + "bbox": [ + 304, + 449, + 545, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 449, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SDCL}" + }, + { + "bbox": [ + 304, + 449, + 545, + 545 + ], + "type": "text", + "content": ". For the study, we use a scenario where the DeepLabV3+ model with backbone ResNet-50 model is trained on GTA and tested on Cityscapes, BDD100K, and Mapillary." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "Table 4 presents the impact of various proposed losses on domain generalization performance. Specifically, the baseline model, trained solely with cross-entropy loss, exhibits suboptimal performance on target domains because of overfitting to the source domain. Conversely, the integration of any proposed loss mechanisms leads to a marked enhancement in performance. More specifically, the incorporation of the covariance alignment " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{CM},\\mathcal{L}_{CC})" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " shows its efficacy in preserving essential content information by correlating features of paired images. The differential impact of the semantic consistency constrastive learning " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{CWCL},\\mathcal{L}_{SDCL})" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " is also evident, as it significantly aids in disentangling features of similar classes, thereby constructing a more robust embedding space." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3622" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 118, + 123 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 118, + 123 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 118, + 123 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 118, + 123 + ], + "type": "image", + "image_path": "5743df130013ec2edea812c12d2a7e375fd2bbd88be13afab8d1a8d9272aea9c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 124, + 103, + 133 + ], + "lines": [ + { + "bbox": [ + 64, + 124, + 103, + 133 + ], + "spans": [ + { + "bbox": [ + 64, + 124, + 103, + 133 + ], + "type": "text", + "content": "Target Data" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 120, + 70, + 190, + 123 + ], + "blocks": [ + { + "bbox": [ + 120, + 70, + 190, + 123 + ], + "lines": [ + { + "bbox": [ + 120, + 70, + 190, + 123 + ], + "spans": [ + { + "bbox": [ + 120, + 70, + 190, + 123 + ], + "type": "image", + "image_path": "337f71e74c8508b424ee737397e719b5030e129644c3e795c1a4ea71ba7d1e6a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 124, + 169, + 133 + ], + "lines": [ + { + "bbox": [ + 140, + 124, + 169, + 133 + ], + "spans": [ + { + "bbox": [ + 140, + 124, + 169, + 133 + ], + "type": "text", + "content": "Baseline" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 191, + 70, + 260, + 123 + ], + "blocks": [ + { + "bbox": [ + 191, + 70, + 260, + 123 + ], + "lines": [ + { + "bbox": [ + 191, + 70, + 260, + 123 + ], + "spans": [ + { + "bbox": [ + 191, + 70, + 260, + 123 + ], + "type": "image", + "image_path": "49c2a2eee91c74f79070a020372836d10140ed59f2a9e5dc8645dd7751c88c18.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 124, + 244, + 133 + ], + "lines": [ + { + "bbox": [ + 207, + 124, + 244, + 133 + ], + "spans": [ + { + "bbox": [ + 207, + 124, + 244, + 133 + ], + "type": "text", + "content": "RobustNet" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 262, + 70, + 332, + 123 + ], + "blocks": [ + { + "bbox": [ + 262, + 70, + 332, + 123 + ], + "lines": [ + { + "bbox": [ + 262, + 70, + 332, + 123 + ], + "spans": [ + { + "bbox": [ + 262, + 70, + 332, + 123 + ], + "type": "image", + "image_path": "95e5f7abd34e2a68c19572c4e0fa1674385464d3501bd35dfd4aa3a6bd004b39.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 124, + 311, + 133 + ], + "lines": [ + { + "bbox": [ + 282, + 124, + 311, + 133 + ], + "spans": [ + { + "bbox": [ + 282, + 124, + 311, + 133 + ], + "type": "text", + "content": "WildNet" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 333, + 70, + 402, + 123 + ], + "blocks": [ + { + "bbox": [ + 333, + 70, + 402, + 123 + ], + "lines": [ + { + "bbox": [ + 333, + 70, + 402, + 123 + ], + "spans": [ + { + "bbox": [ + 333, + 70, + 402, + 123 + ], + "type": "image", + "image_path": "ba681e1508b8da213ef145d65f63839dc55f7d2eace8e17b8c6c0de9e132f18f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 124, + 375, + 133 + ], + "lines": [ + { + "bbox": [ + 359, + 124, + 375, + 133 + ], + "spans": [ + { + "bbox": [ + 359, + 124, + 375, + 133 + ], + "type": "text", + "content": "SPC" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 403, + 70, + 473, + 123 + ], + "blocks": [ + { + "bbox": [ + 403, + 70, + 473, + 123 + ], + "lines": [ + { + "bbox": [ + 403, + 70, + 473, + 123 + ], + "spans": [ + { + "bbox": [ + 403, + 70, + 473, + 123 + ], + "type": "image", + "image_path": "56a57369c29ab1ee852bc71832d6b55c8c7c94a5be7f39e7bf518f526eb5fc04.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 124, + 447, + 133 + ], + "lines": [ + { + "bbox": [ + 430, + 124, + 447, + 133 + ], + "spans": [ + { + "bbox": [ + 430, + 124, + 447, + 133 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 475, + 70, + 545, + 123 + ], + "blocks": [ + { + "bbox": [ + 475, + 70, + 545, + 123 + ], + "lines": [ + { + "bbox": [ + 475, + 70, + 545, + 123 + ], + "spans": [ + { + "bbox": [ + 475, + 70, + 545, + 123 + ], + "type": "image", + "image_path": "e75373901ae793412627ba1d0ec111765d7b497f114899c12c47858de9371ae6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 124, + 533, + 133 + ], + "lines": [ + { + "bbox": [ + 487, + 124, + 533, + 133 + ], + "spans": [ + { + "bbox": [ + 487, + 124, + 533, + 133 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 72, + 184, + 190, + 275 + ], + "blocks": [ + { + "bbox": [ + 46, + 143, + 546, + 165 + ], + "lines": [ + { + "bbox": [ + 46, + 143, + 546, + 165 + ], + "spans": [ + { + "bbox": [ + 46, + 143, + 546, + 165 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison between DGSS methods trained on GTAV (G) and tested on unseen target domains of Cityscapes (C) using DeeplabV3+ with ResNet50 backbone." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 72, + 184, + 190, + 275 + ], + "lines": [ + { + "bbox": [ + 72, + 184, + 190, + 275 + ], + "spans": [ + { + "bbox": [ + 72, + 184, + 190, + 275 + ], + "type": "image", + "image_path": "97b7de89b1fade3976ea20663122c5a175bbbba4817f004ecaee8fedc9661f42.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 200, + 253, + 209 + ], + "lines": [ + { + "bbox": [ + 192, + 200, + 253, + 209 + ], + "spans": [ + { + "bbox": [ + 192, + 200, + 253, + 209 + ], + "type": "text", + "content": "Baseline - Original " + }, + { + "bbox": [ + 192, + 200, + 253, + 209 + ], + "type": "inline_equation", + "content": "x" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 192, + 215, + 263, + 224 + ], + "lines": [ + { + "bbox": [ + 192, + 215, + 263, + 224 + ], + "spans": [ + { + "bbox": [ + 192, + 215, + 263, + 224 + ], + "type": "text", + "content": "Baseline - Augmented " + }, + { + "bbox": [ + 192, + 215, + 263, + 224 + ], + "type": "inline_equation", + "content": "x_{a}" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 192, + 236, + 243, + 245 + ], + "lines": [ + { + "bbox": [ + 192, + 236, + 243, + 245 + ], + "spans": [ + { + "bbox": [ + 192, + 236, + 243, + 245 + ], + "type": "text", + "content": "Ours - Original " + }, + { + "bbox": [ + 192, + 236, + 243, + 245 + ], + "type": "inline_equation", + "content": "x" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 192, + 251, + 253, + 259 + ], + "lines": [ + { + "bbox": [ + 192, + 251, + 253, + 259 + ], + "spans": [ + { + "bbox": [ + 192, + 251, + 253, + 259 + ], + "type": "text", + "content": "Ours-Augmented " + }, + { + "bbox": [ + 192, + 251, + 253, + 259 + ], + "type": "inline_equation", + "content": "x_{a}" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 284, + 287, + 306 + ], + "lines": [ + { + "bbox": [ + 46, + 284, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 287, + 306 + ], + "type": "text", + "content": "Figure 6. t-SNE [49] visualization comparing the covariance with and without " + }, + { + "bbox": [ + 46, + 284, + 287, + 306 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CML}" + }, + { + "bbox": [ + 46, + 284, + 287, + 306 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 328, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 424 + ], + "type": "text", + "content": "Covariance Matching Loss. Fig. 6 presents t-SNE plots of covariances for original and augmented images, before and after the application of CML. The baseline network perceives original and augmented images differently from a style perspective. However, after applying CML, the distribution becomes more intermixed, indicating that our proposed CML effectively ensures similar recognition of different style images." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "text", + "content": "Calculation of CCL. Table 5a demonstrates that our proposed cross-covariance method, which converges the diagonal components to 1, yields superior performance. As mentioned before, removing non-diagonal components, which contain content information actually degrades performance." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 485, + 288, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 288, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 288, + 593 + ], + "type": "text", + "content": "Sampling number in CWCL. Table 5b and Table 5c show the impact of varying the number of classes sampled per image and the number of samples per class in CWCL, respectively. As shown in Table 5b, the performance improves with an increase in the diversity of classes sampled in CWCL. This suggests that contrasting a broader array of classes enhances the model's discriminative capability. Furthermore, Table 5c demonstrates that a balanced number of negative samples per class leads to optimal performance." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Projection Head for SDCL. The influence of different project head configurations on the SDCL is investigated. We experimented with three distinct approaches: individual projection head, copying the weights of CWCL's (stop gradient), and shared projection head of CWCL. As demonstrated in Table 5d, sharing the projection head yielded the most superior results. The results indicate that SDCL not only relies on the semantic information from CWCL for effective disentanglement of similar classes but also enhances the embedding space learned by CWCL." + } + ] + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 306, + 184, + 434, + 229 + ], + "blocks": [ + { + "bbox": [ + 306, + 184, + 434, + 229 + ], + "lines": [ + { + "bbox": [ + 306, + 184, + 434, + 229 + ], + "spans": [ + { + "bbox": [ + 306, + 184, + 434, + 229 + ], + "type": "table", + "html": "
Cross-covariance loss
MethodCBM
Whitening38.6836.9142.12
Diagonal40.4237.8143.91
", + "image_path": "8690ec98ec23f09059bc8e81f3d41c1140c4b223d47e8a5b7f326677aa559de0.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "type": "table", + "bbox": [ + 436, + 184, + 539, + 229 + ], + "blocks": [ + { + "bbox": [ + 364, + 231, + 373, + 239 + ], + "lines": [ + { + "bbox": [ + 364, + 231, + 373, + 239 + ], + "spans": [ + { + "bbox": [ + 364, + 231, + 373, + 239 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 436, + 184, + 539, + 229 + ], + "lines": [ + { + "bbox": [ + 436, + 184, + 539, + 229 + ], + "spans": [ + { + "bbox": [ + 436, + 184, + 539, + 229 + ], + "type": "table", + "html": "
# of classes
#CBM
1045.5738.8846.37
1545.7241.3247.08
", + "image_path": "3d78f9782c24c7a5e8b367c88b83157faf680aa15989be82e67c3294f27d90c3.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "type": "table", + "bbox": [ + 310, + 245, + 410, + 296 + ], + "blocks": [ + { + "bbox": [ + 483, + 230, + 492, + 238 + ], + "lines": [ + { + "bbox": [ + 483, + 230, + 492, + 238 + ], + "spans": [ + { + "bbox": [ + 483, + 230, + 492, + 238 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 245, + 410, + 296 + ], + "lines": [ + { + "bbox": [ + 310, + 245, + 410, + 296 + ], + "spans": [ + { + "bbox": [ + 310, + 245, + 410, + 296 + ], + "type": "table", + "html": "
# of negative samples
#CBM
1044.7638.2146.46
5045.7241.3247.08
10044.4439.1446.29
", + "image_path": "02ede73468dbdebefc060f9a495d9c8eb9276b5f007fc4eb3a4ddbc706d8d9e6.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "table_body" + } + ], + "index": 30 + }, + { + "type": "table", + "bbox": [ + 416, + 255, + 541, + 296 + ], + "blocks": [ + { + "bbox": [ + 356, + 297, + 364, + 304 + ], + "lines": [ + { + "bbox": [ + 356, + 297, + 364, + 304 + ], + "spans": [ + { + "bbox": [ + 356, + 297, + 364, + 304 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 439, + 245, + 518, + 254 + ], + "lines": [ + { + "bbox": [ + 439, + 245, + 518, + 254 + ], + "spans": [ + { + "bbox": [ + 439, + 245, + 518, + 254 + ], + "type": "text", + "content": "Projection Head of SDCL" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 416, + 255, + 541, + 296 + ], + "lines": [ + { + "bbox": [ + 416, + 255, + 541, + 296 + ], + "spans": [ + { + "bbox": [ + 416, + 255, + 541, + 296 + ], + "type": "table", + "html": "
MLPCBM
Individual44.9138.4546.31
Shared (SG)44.0338.1546.64
Shared45.7241.3247.08
", + "image_path": "865e936593eaf7e7c54b4fd622556dd46df755addabcc6363305b953c7327d9d.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "table_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 474, + 297, + 483, + 304 + ], + "lines": [ + { + "bbox": [ + 474, + 297, + 483, + 304 + ], + "spans": [ + { + "bbox": [ + 474, + 297, + 483, + 304 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 313, + 546, + 346 + ], + "lines": [ + { + "bbox": [ + 305, + 313, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 313, + 546, + 346 + ], + "type": "text", + "content": "Table 5. Ablation studies. (a) Calculation of " + }, + { + "bbox": [ + 305, + 313, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CC}" + }, + { + "bbox": [ + 305, + 313, + 546, + 346 + ], + "type": "text", + "content": ". (b) Number of classes for " + }, + { + "bbox": [ + 305, + 313, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CWCL}" + }, + { + "bbox": [ + 305, + 313, + 546, + 346 + ], + "type": "text", + "content": ". (c) Number of negative samples for " + }, + { + "bbox": [ + 305, + 313, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CWCL}" + }, + { + "bbox": [ + 305, + 313, + 546, + 346 + ], + "type": "text", + "content": ". (d) Projection head of " + }, + { + "bbox": [ + 305, + 313, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SDCL}" + }, + { + "bbox": [ + 305, + 313, + 546, + 346 + ], + "type": "text", + "content": ". \"SG\" indicates stop gradient." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 366, + 378, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 366, + 378, + 379 + ], + "spans": [ + { + "bbox": [ + 306, + 366, + 378, + 379 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 387, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 546, + 555 + ], + "type": "text", + "content": "In this paper, we propose a novel BlindNet with covariance alignment and semantic consistency contrastive learning. By introducing covariance alignment, our method effectively addresses style variations, ensuring the extraction of features that are consistent across different styles. Furthermore, with the proposed semantic consistency contrastive learning, we not only facilitate the extraction of discriminative features but also enhance the generalization capabilities of the model in semantic segmentation predictions. Comprehensive experimental results validate the effectiveness of our approach, demonstrating its ability to generalize across multiple unseen target domains without requiring auxiliary domains or additional modules. Our future work will be improving and stabilizing the covariance alignment method." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 557, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 712 + ], + "type": "text", + "content": "Acknowledgement. This work was supported in part by the Basic Science Research Program through National Research Foundation of Korea (NRF) (Grants No. NRF2022R1F1A1073543), the MSIT(Ministry of Science and ICT), Korea, under the ICAN(ICT Challenge and Advanced Network of HRD) support program(RS-2022-00156385) supervised by the IITP(Institute for Information & Communications Technology Planning & Evaluation), and Innovative Human Resource Development for Local Intellectualization program through the Institute of Information & Communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT)(IITP-2024-00156287)." + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3623" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Florent Bartoccioni, Éloi Zablocki, Andrei Bursuc, Patrick Pérez, Matthieu Cord, and Karteek Alahari. Lara: Latents and rays for multi-camera bird's-eye-view semantic segmentation. In Conference on Robot Learning, pages 1663-1672. PMLR, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 201 + ], + "type": "text", + "content": "[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 245 + ], + "type": "text", + "content": "[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "type": "text", + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 292, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 288, + 346 + ], + "type": "text", + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 347, + 288, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 288, + 402 + ], + "type": "text", + "content": "[6] Wonwooong Cho, Sungha Choi, David Keetae Park, Inkyu Shin, and Jaegul Choo. Image-to-image translation via group-wise deep whitening-and-coloring transformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10639-10647, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 468 + ], + "type": "text", + "content": "[7] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11580-11590, 2021. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 469, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 288, + 534 + ], + "type": "text", + "content": "[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 536, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 536, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 53, + 536, + 288, + 579 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "text", + "content": "[10] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "text", + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "type": "text", + "content": "[12] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "text", + "content": "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 198, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 252 + ], + "type": "text", + "content": "[15] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "text", + "content": "[16] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9924-9935, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 310, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 363 + ], + "type": "text", + "content": "[17] Lukas Hoyer, Dengxin Dai, Qin Wang, Yuhua Chen, and Luc Van Gool. Improving semi-supervised and domain-adaptive semantic segmentation with self-supervised depth estimation. International Journal of Computer Vision, pages 1-27, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 365, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 421 + ], + "type": "text", + "content": "[18] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "type": "text", + "content": "[19] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6891-6902, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 478, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 544 + ], + "type": "text", + "content": "[20] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3061-3071, 2023. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "type": "text", + "content": "[21] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "text", + "content": "[22] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European conference on computer vision (ECCV), pages 172-189, 2018. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "text", + "content": "[23] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 681, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 714 + ], + "type": "text", + "content": "[24] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3624" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 9936-9946, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[25] Guangrui Li, Guoliang Kang, Wu Liu, Yunchao Wei, and Yi Yang. Content-consistent matching for domain adaptive semantic segmentation. In European conference on computer vision, pages 440-456. Springer, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 286, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 286, + 205 + ], + "type": "text", + "content": "[26] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 224-233, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 286, + 251 + ], + "type": "text", + "content": "[27] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 286, + 308 + ], + "type": "text", + "content": "[28] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European conference on computer vision (ECCV), pages 624-639, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 286, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 286, + 353 + ], + "type": "text", + "content": "[29] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6936-6945, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 286, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 286, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 286, + 387 + ], + "type": "text", + "content": "[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 286, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 286, + 422 + ], + "type": "text", + "content": "[31] Ping Luo. Learning deep architectures via generalized whitened neural networks. In International Conference on Machine Learning, pages 2238-2246. PMLR, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 423, + 286, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 423, + 286, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 423, + 286, + 466 + ], + "type": "text", + "content": "[32] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European conference on computer vision (ECCV), pages 116-131, 2018. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 286, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 286, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 286, + 521 + ], + "type": "text", + "content": "[33] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990-4999, 2017. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 286, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 286, + 567 + ], + "type": "text", + "content": "[34] David Nilsson, Aleksis Pirinen, Erik Gartner, and Cristian Sminchisescu. Embodied visual active learning for semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2373–2383, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 286, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 286, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 286, + 623 + ], + "type": "text", + "content": "[35] Yuya Onozuka, Ryosuke Matsumi, and Motoki Shino. Autonomous mobile robot navigation independent of road boundary using driving recommendation map. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4501-4508. IEEE, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 286, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 286, + 657 + ], + "type": "text", + "content": "[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "type": "text", + "content": "[37] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3764-3773, 2020. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[38] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464-479, 2018. 2, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 119, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 171 + ], + "type": "text", + "content": "[39] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1863-1871, 2019. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "text", + "content": "[40] Taesung Park, Alexei A Efros, Richard Zhang, and JunYan Zhu. Contrastive learning for unpaired image-to-image translation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 319-345. Springer, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "text", + "content": "[41] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE Transactions on Image Processing, 30:6594-6608, 2021. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "type": "text", + "content": "[42] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2594-2605, 2022. 2, 4, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 332, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 396 + ], + "type": "text", + "content": "[43] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 102-118. Springer, 2016. 1, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 399, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 465 + ], + "type": "text", + "content": "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 467, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 531 + ], + "type": "text", + "content": "[45] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 1, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 533, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 545, + 588 + ], + "type": "text", + "content": "[46] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10765-10775, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "type": "text", + "content": "[47] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 678 + ], + "type": "text", + "content": "[48] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 5, 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3625" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[50] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 129, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 287, + 185 + ], + "type": "text", + "content": "[51] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2517-2526, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 186, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 240 + ], + "type": "text", + "content": "[52] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7303-7313, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 241, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 287, + 286 + ], + "type": "text", + "content": "[53] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. Siamdoge: Domain generalizable semantic segmentation using siamese network. In European Conference on Computer Vision, pages 603-620. Springer, 2022. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 287, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 287, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 49, + 287, + 287, + 341 + ], + "type": "text", + "content": "[54] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 342, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 342, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 49, + 342, + 287, + 408 + ], + "type": "text", + "content": "[55] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2884-2892, 2022. 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 411, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 411, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 411, + 287, + 453 + ], + "type": "text", + "content": "[56] Liwei Yang, Xiang Gu, and Jian Sun. Generalized semantic segmentation by self-supervised source domain projection and multi-level contrastive learning. arXiv preprint arXiv:2303.01906, 2023. 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 456, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 456, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 49, + 456, + 287, + 510 + ], + "type": "text", + "content": "[57] Jaejun Yoo, Youngjung Uh, Sanghyuk Chun, Byeongkyu Kang, and Jung-Woo Ha. Photorealistic style transfer via wavelet transforms. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9036-9045, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 512, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 512, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 49, + 512, + 287, + 577 + ], + "type": "text", + "content": "[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 578, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 578, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 49, + 578, + 287, + 633 + ], + "type": "text", + "content": "[59] Fei Yu, Mo Zhang, Hexin Dong, Sheng Hu, Bin Dong, and Li Zhang. Dast: Unsupervised domain adaptation in semantic segmentation based on discriminator attention and self-training. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10754–10762, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 635, + 287, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 635, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 49, + 635, + 287, + 701 + ], + "type": "text", + "content": "[60] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2100-2110, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 703, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 703, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 703, + 287, + 713 + ], + "type": "text", + "content": "[61] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Do" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 150 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "main generalization with mixstyle. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[62] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018. 1, 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3626" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/43978830-73bd-4db5-8b8c-57888ce93c6e_content_list.json b/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/43978830-73bd-4db5-8b8c-57888ce93c6e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e50921680821a2f8a4fc2a32793c26c7a39a1502 --- /dev/null +++ b/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/43978830-73bd-4db5-8b8c-57888ce93c6e_content_list.json @@ -0,0 +1,1665 @@ +[ + { + "type": "text", + "text": "Style Injection in Diffusion: A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer", + "text_level": 1, + "bbox": [ + 176, + 130, + 794, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiwoo Chung*, Sangeek Hyun*, Jae-Pil Heo† Sungkyunkwan University", + "bbox": [ + 305, + 202, + 661, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wldn0202, hsi1032, jaepilheo}@g.skku.edu", + "bbox": [ + 303, + 241, + 665, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the impressive generative capabilities of diffusion models, existing diffusion model-based style transfer methods require inference-stage optimization (e.g. fine-tuning or textual inversion of style) which is time-consuming, or fails to leverage the generative ability of large-scale diffusion models. To address these issues, we introduce a novel artistic style transfer method based on a pre-trained large-scale diffusion model without any optimization. Specifically, we manipulate the features of self-attention layers as the way the cross-attention mechanism works; in the generation process, substituting the key and value of content with those of style image. This approach provides several desirable characteristics for style transfer including 1) preservation of content by transferring similar styles into similar image patches and 2) transfer of style based on similarity of local texture (e.g. edge) between content and style images. Furthermore, we introduce query preservation and attention temperature scaling to mitigate the issue of disruption of original content, and initial latent Adaptive Instance Normalization (AdaIN) to deal with the disharmonious color (failure to transfer the colors of style). Our experimental results demonstrate that our proposed method surpasses state-of-the-art methods in both conventional and diffusion-based style transfer baselines. Codes are available at github.com/jiwoogit/StyleID.", + "bbox": [ + 75, + 323, + 473, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 734, + 209, + 751 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in Diffusion Models (DMs) have led to breakthroughs in various generative applications such as text-to-image synthesis [30, 33, 35] and image or video editing [2, 4, 6, 14, 19, 41, 48]. One of these efforts is also applied to the task of style transfer [10, 18, 45, 47, 53]; given style and content images, modifying the style of the content image to possess the given style.", + "bbox": [ + 75, + 761, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/64e6849128fd581115042f13f5ea0b408d3614d988eeffcf75b7b26502729014.jpg", + "image_caption": [ + "(a) General usage of self-attention in DM" + ], + "image_footnote": [], + "bbox": [ + 509, + 295, + 888, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ede24d878fbeafce1263d42f2ba33274a2148939c7c478f43bbeff71c501e9c.jpg", + "image_caption": [ + "(b) Our adaptation of self-attention for style transfer", + "Figure 1. Manipulation of self-attention features for style transfer. (a) General self-attention (SA) deploys the query, key, and value features from a single image in both the training and inference phases. (b) At inference phase, we suggest that manipulating features of self-attention of pre-trained large-scale DM is an effective way to transfer the styles; injection of key and value of styles into SA of contents is a proper way for transferring styles. As a result, style-injected content $z_{t - 1}^{c}$ would maintain contents while modifying its style to resemble the target style." + ], + "image_footnote": [], + "bbox": [ + 506, + 377, + 888, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "General approaches for diffusion model-based style transfer leverage the generative capability of pre-trained DM. Some of these works focus on explicit disentangling style and content for interpretable and controllable style transfer [45], or inversion of the style image into the textual latent space of a large-scale text-to-image DM [53]. However, these methods additionally require gradient-based optimization for fine-tuning and textual inversion [34] for each style image, which is time-consuming. Without this issue, DiffStyle [18] introduces training-free style transfer, but they are known to be hardly applicable to Latent Diffusion Model [33] which is widely adopted for training large-scale text-to-image DM such as Stable Diffusion [33], hindering the users from taking advantage of the prominent generative ability of large-scale models.", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 101, + 875, + 212, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 102, + 887, + 228, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8795", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a5ef20a21d0b05940f620e5cd1ca924f81d6dc810fb5c5f55b2724dc2eda0ae7.jpg", + "image_caption": [ + "(b) query-key similarity between patches of similar texture", + "Figure 2. Desirable attributes of self-attention (SA) for style transfer. (a) Visualization of query by PCA shows that query features well-reflect similarities among patches. That is, style transfer employing SA can preserve the original content, as content patches with similarities tend to receive similar attention scores from a corresponding style image patch. (b) We visualize a similarity map between the blue box (edge) query of the content image, and key of the style image. Thanks to the features representation of large-scale DM encompassing texture and semantics, a query exhibits higher similarity to keys that share a similar style, such as edges." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 468, + 228 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we focus on extending the training-free style transfer to its application on large-scale pre-trained DM. We start from the observation of recent advances in image-to-image translation based on large-scale DM; they uncover the image editing capability of attention layers. Notably, Plug-and-play [41] shows that the residual block and the attention map of self-attention (SA) determine the spatial layout of generated images. Also, Prompt-to-Prompt [14] locally edits the image by replacing key and value of cross-attention (CA) obtained from text prompt, while keeping their original attention maps. That is, all these works suggest that 1) attention maps determine the spatial layout and 2) key and value of CA adjust the content to fill.", + "bbox": [ + 75, + 400, + 468, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the aforementioned methods, we newly argue that manipulating the SA layer is an effective way to transfer the styles (Fig. 1). Specifically, similar to CA, we substitute the key and value of SA and observe that the generated images are still visually plausible and naturally incorporate the elements of the substituted image into the original image. This observation motivates us to propose a style transfer technique based on self-attention, which combines the styles (textures) of a specific image with the content (semantics and spatial layout) of different images. Furthermore, we highlight that SA layer has desirable characteristics for style transfer. First, as shown in Fig. 2 (a), in SA-based style transfer, the content image patches (query) that share semantic similarities engage with a similar style (key), thereby maintaining the relationship among these content image patches after the transfer. Next, thanks to the powerful feature representation of large-scale DM [49], each patch of the query reveals higher similarity to keys which has similar texture and semantics. For instance, in Fig. 2 (b),", + "bbox": [ + 75, + 613, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we can observe that the query feature of content within the blue box exhibits a high similarity to the key features of style with similar edge texture. This encourages the model to transfer style based on the similarity of local texture (e.g. edge) between content and style.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a result, our method aims to transfer the textures of the style image to the content images by manipulating self-attention features of pre-trained large-scale DM without any optimization. To this end, we first propose an attention-based style injection method. The basic idea of it is substituting the content's key and value of SA with those of the style image, especially layers in the latter part of decoder which are relevant to the local textures. As mentioned in above paragraph, exchanged styles are well aligned with the content and texture of original image, exploiting the similarity-based attention mechanism. With the proposed style injection, we observe that the local texture patterns are successfully transferred, but there still are remaining problems such as disruption of original content and disharmonious colors. To handle these problems, we additionally propose the following techniques; query preservation, attention temperature scaling, and initial latent AdaIN. Query preservation makes the reverse diffusion process to retain the spatial structure of original content by preserving the query of the content image in the SA. Attention temperature scaling also aims to keep the structure of content by dealing with the blurred self-attention map introduced from the substitution of key. Lastly, initial latent AdaIN corrects inharmonious color problem, referring that the color distribution of style images is not properly transferred, by modulating the statistics of initial noise in the diffusion model.", + "bbox": [ + 496, + 167, + 890, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 517, + 561, + 859, + 575 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a style transfer method exploiting the large-scale pre-trained DM by simple manipulation of the features in self-attention; substituting key and value of content with those of styles without any requirements of optimization or supervision (e.g. text).", + "- We further improve the naive approach for style transfer to properly adapt the styles by proposing three components; query preservation, attention temperature scaling, and initial latent AdaIN.", + "- Extensive experiments on the style transfer dataset validate the proposed method significantly outperforms previous methods and achieves state-of-the-art performance." + ], + "bbox": [ + 496, + 579, + 890, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 773, + 640, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion Model-based Neural Style Transfer", + "text_level": 1, + "bbox": [ + 500, + 800, + 879, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural style transfer [11, 24-26, 29, 31, 43, 44, 52] is an example-guided image generation task that transfers the style of one image onto another while retaining the content of the original. In the realm of diffusion models, neural style transfer has evolved by leveraging the generative capability", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8796", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of pre-trained diffusion models. For instance, InST [53] introduced a textual inversion-based approach, aiming to map a given style into corresponding textual embeddings. StyleDiffusion [45] aimed to disentangle style and content by introducing CLIP-based style disentanglement loss for fine-tuning DM for style transfer. Also, several approaches utilize the text input as a style condition or for determining the content to synthesize [10, 47].", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conversely, DiffStyle [18] proposed a training-free style transfer method that leverages $h$ -space [23] and adjusts skip connections for effectively conveying style and content information, respectively. However, when DiffStyle is applied to Stable Diffusion [33, 42], their behavior is quite different from typical style-transfer methods; not only textures but also semantics such as spatial layout are also changed.", + "bbox": [ + 75, + 214, + 468, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these limitations, we propose a novel algorithm that harmoniously merges style and content features within the self-attention layers of Stable Diffusion without any optimization process.", + "bbox": [ + 75, + 323, + 468, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Attention-based Image Editing in DM", + "text_level": 1, + "bbox": [ + 76, + 402, + 403, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the remarkable advances achieved by pre-trained text-to-image DMs [32, 42], there have been numerous image editing works [2, 6, 19, 37] utilizing these DMs. Notably, Prompt-to-Prompt [14] proposed text-based local image editing by manipulating the cross-attention map. Specifically, they observe that cross-attention largely contributes to modeling the relation between the spatial layout of the image to each word in the prompt. Hence, they substitute the original words and cross-attention map with desirable ones, obtaining edited images matched with text conditions. Subsequently, Plug-and-play [41] introduces text-guided image-to-image translation method. They found that the spatial features (i.e. feature from residual block) and self-attention map determine the spatial layout of the synthesized image. Thus, while generating a new image with the given text condition, they guide the diffusion model with features and attention map from the original image for preserving the original spatial layout. Recently, MasaCtrl [3] proposes mutual self-attention control for consistent image editing using text prompts. In detail, they retain the source image's key and value of the self-attention layers, while conditioning the model with desired text prompts.", + "bbox": [ + 75, + 429, + 468, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Along with these works, we recognize the potential of attention maps in representing spatial information. However, different from the aforementioned methods concentrating on exploiting textual condition, we focus on conditioning by style and content images composed of two images from distinct styles. By combining the features in self-attention layers of both style and content images with precise adjustment of statistics in intermediate representations, we transfer the texture of the content image to the given style.", + "bbox": [ + 75, + 765, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 500, + 89, + 627, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent Diffusion Model (LDM) [33] is a type of diffusion model trained in the low dimensional latent space to focus on semantic bits of data and reduce computation costs. Given an image $x \\in \\mathbb{R}^{H \\times W \\times 3}$ , the encoder $\\mathcal{E}$ encodes $x$ into the latent representation $z \\in \\mathbb{R}^{h \\times w \\times c}$ and decoder reconstructs the image from the latent.", + "bbox": [ + 496, + 116, + 890, + 207 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the pretrained encoder, they encode the entire images in the dataset and train a diffusion model on latent space $z$ , by predicting noise $\\epsilon$ from the noised version of latent $z_{t}$ at time step $t$ . The corresponding training objective is", + "bbox": [ + 496, + 209, + 890, + 282 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {L D M}} = \\mathbb {E} _ {z, \\epsilon , t} [ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, y) \\| _ {2} ^ {2} ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 289, + 890, + 308 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon \\in \\mathcal{N}(0,1)$ is a noise, $t$ is the number of time steps which uniformly sampled from $\\{1,\\dots,T\\}$ , $y$ is a condition, and $\\epsilon_{\\theta}$ is a neural network which predicts the noise added to the $z$ .", + "bbox": [ + 496, + 319, + 890, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our work, we utilize Stable Diffusion (SD) [33] which is the only publicized large-scale pre-trained DM. In the case of SD, $y$ is a text, and $\\epsilon_{\\theta}$ is a U-Net architecture in which a block for each resolution comprises a residual block, self-attention block (SA), and cross-attention block (CA), sequentially. Among these modules, we focus on the SA block to transfer the styles, as discussed in Sec. 1. Given a feature $\\phi$ after the residual block, the self-attention block performs as follows:", + "bbox": [ + 496, + 381, + 890, + 517 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nQ = W _ {Q} (\\phi), K = W _ {K} (\\phi), V = W _ {V} (\\phi),\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 532, + 841, + 550 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {\\text {o u t}} = \\operatorname {A t t n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) \\cdot V, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 551, + 890, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d$ denotes the dimension of the projected query, and $W_{(\\cdot)}$ is a projection layer. Note that, we don't use any text conditions, so the variable $y$ is always an empty text prompt $(\\ldots)$ .", + "bbox": [ + 500, + 599, + 890, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 676, + 589, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we aim to solve artistic style transfer by leveraging the generative capability of a pre-trained large-scale text-to-image diffusion model. Briefly, artistic style transfer is the task of modifying the style of a given content image $I^{c}$ to that of style image $I^{s}$ . Then, the stylized image $I^{cs}$ would maintain the semantic content of $I^{c}$ while its style (such as texture) is transferred from $I^{s}$ . For simplicity, we skip the explanations about the encoding and decoding process of the autoencoder in the LDM. Instead, we focus on elaborating the proposed method in the aspect of the diffusion process. Thus, in the following sections, we regard the content, style, and stylized images same as their encoded counterparts $z_{0}^{c}, z_{0}^{s}$ , and $z_{0}^{cs}$ .", + "bbox": [ + 496, + 704, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8797", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d557a0ed5ba8b09332a9a9cb85c7f7e3dad6ad5ddc3172013a934b5322b0b4d1.jpg", + "image_caption": [ + "Figure 3. Overall framework. (Left) Illustration for the proposed style transfer method. We first invert content image $z_0^c$ and style image $z_0^s$ into the latent noise space as $z_T^c$ and $z_T^s$ , respectively. Then, we initialize the initial noise of stylized image $z_T^{cs}$ from initial latent AdaIN (Sec. 4.3) which combines the content and style noise, $z_T^c$ and $z_T^s$ . While performing the reverse diffusion process with $z_T^{cs}$ , we inject the information of content and style by attention-based style injection (Sec. 4.1) and attention temperature scaling (Sec. 4.2). (Right) Detailed explanation of style injection and initial noise AdaIN. Style injection is basically the manipulation of self-attention (SA) layer during the reverse diffusion process. Specifically, at time step $t$ , we substitute the key $(K_t^{cs})$ and value $(V_t^{cs})$ in SA of stylized image with those of style features, $K_t^s$ and $V_t^s$ , from identical timestep $t$ . At the same time, we preserve the content information by blending the query of content $Q_t^c$ and query of stylized image $Q_t^{cs}$ . Finally, we scale the magnitude of the attention map to deal with the magnitude decrease that the substitution of feature leads to. Initial latent AdaIN produces the initial noise $z_T^{cs}$ by combining style noise $z_T^s$ and content noise $z_T^s$ . Specifically, we modify the channel statistics of $z_T^c$ to resemble the statistics of $z_T^s$ and regard it as $z_T^{cs}$ . We observe this operation enables us to keep the spatial layout of content image while well-reflecting the color tones of a given style image." + ], + "image_footnote": [], + "bbox": [ + 83, + 92, + 580, + 361 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/25341082c1bb4fa13bb877698f1b3db935c09e0b80ac2ef541c5291fe7f10777.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 92, + 883, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Attention-based Style Injection", + "text_level": 1, + "bbox": [ + 76, + 532, + 352, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We start from the observation in previous image-to-image translation methods, especially Prompt-to-Prompt [14]. The key idea of their method is changing the text condition for cross-attention (CA) while keeping the attention map. Since the attention map affects the spatial layout of output, substituted text conditions determine what to draw in the generated image, and these conditions are actually key and value in CA. Inspired by them, we manipulate the features in self-attention layer as like cross-attention, regarding the features from style image $I^{s}$ as the condition. Specifically, in the generation process, we substitute the key and value of content image with those of style for transferring the texture of style image into the content image.", + "bbox": [ + 75, + 559, + 468, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this end, we first obtain the latent for content and style images with DDIM inversion [39], and then collect the SA features of style image over the DDIM inversion process. Specifically, for pre-defined timesteps $t = \\{0,\\dots,T\\}$ , style and content images $z_0^c$ and $z_0^s$ are inverted from image $(t = 0)$ to gaussian noise $(t = T)$ . During DDIM inversion, we also collect query features of content $(Q_t^c)$ and key and value features of style $(K_t^s,V_t^s)$ at every time steps.", + "bbox": [ + 75, + 760, + 468, + 881 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After that, we initialize stylized latent noise $z_T^{cs}$ by copy-", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing content latent noise $z_{T}^{c}$ . Then, we transfer the target style to the stylized latent by injecting the key $K_{t}^{s}$ and value $V_{t}^{s}$ collected from the style into SA layer, instead of the original key $K_{t}^{cs}$ and value $V_{t}^{cs}$ , when performing the entire reverse process of stylized latent $z_{t}^{cs}$ . However, only applying this substitution can lead to content disruption, since the content of stylized latent would be progressively changed as attended value changes. Hence, we propose query preservation to maintain original content. Simply, we blend query of stylized latent $Q_{t}^{cs}$ and that of content $Q_{t}^{c}$ for the entire reverse process. These style injection and query preservation processes at time step $t$ are expressed as follows:", + "bbox": [ + 496, + 532, + 890, + 714 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {Q} _ {t} ^ {c s} = \\gamma \\times Q _ {t} ^ {c} + (1 - \\gamma) \\times Q _ {t} ^ {c s}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 728, + 890, + 746 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {\\text {o u t}} ^ {\\mathrm {c s}} = \\operatorname {A t t n} \\left(\\tilde {Q} _ {t} ^ {c s}, K _ {t} ^ {s}, V _ {t} ^ {s}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 765, + 890, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ is degree of blending in range of [0, 1]. In addition, we apply these operations on the latter layers of decoder (7- $12^{\\mathrm{th}}$ decoder layers in SD) relevant to local textures. We also highlight that the proposed method can adjust the degree of style transfer by changing query preservation ratio $\\gamma$ . Specifically, higher $\\gamma$ maintains more content, while lower $\\gamma$ strengthens effects of style transfer.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8798", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3dfeb95f4b4a72fd4f26c9fa31b3133f0dc69f21163d9f1bc6c7575f44d06a02.jpg", + "image_caption": [ + "(a) Std. of attention map", + "(b) Ratio of std. of attention map" + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 294, + 189 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ac32451fdf99d903083809667e55399f7e6415ded3047660913ea5c999dfe958.jpg", + "image_caption": [ + "Figure 4. Visualization of the standard deviation of attention map before softmax. (a) Attention-based style injection reduces the standard deviation of self-attention map. Original denotes SA maps from the generation process without style injection. We use both style and content images for generation. (b) We compute the ratio between attention maps w/ and w/o style injection. For the std of original image, we use averaged std. of content and style." + ], + "image_footnote": [], + "bbox": [ + 294, + 90, + 464, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Attention Temperature Scaling", + "text_level": 1, + "bbox": [ + 76, + 320, + 352, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attention map is computed by scaled dot-product between query and key features. During training, query and key features in the SA layer originate from an identical image. However, if we substitute the key features with those of style images, the magnitude of similarity would be overall lowered as style and content are highly likely to be irrelevant. Thus, the computed attention map can be blurred or smoothed, and it would further make output images unsharp, which is detrimental to capturing both content and style information.", + "bbox": [ + 75, + 343, + 468, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To quantify this issue, we measure the standard deviation of attention map, while ablating the attention-based style injection. In detail, we calculate the attention map before applying softmax, which is scaled-dot product between query and key. As shown in Fig. 4 (a), we validate that this style injection tends to lower the standard deviation of the attention map over the entire timesteps. That is, attention maps after softmax with style injection would be overly smooth.", + "bbox": [ + 75, + 494, + 468, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To rectify the attention map sharper, we introduce an attention temperature scaling parameter. In detail, we multiply the attention map before softmax by a constant temperature scaling parameter $\\tau$ larger than 1. Thus, the attention map after softmax would be sharper than its original values. The modified attention process is represented as follows:", + "bbox": [ + 75, + 614, + 468, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} _ {\\tau} \\left(\\tilde {Q _ {t} ^ {c s}}, K _ {t} ^ {s}, V _ {t} ^ {s}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {\\tau \\tilde {Q _ {t} ^ {c s}} \\left(K _ {t} ^ {s}\\right) ^ {T}}{\\sqrt {d}}\\right) \\cdot V _ {t} ^ {s}, \\tau > 1. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 715, + 467, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use $\\tau = 1.5$ as a default setting, which is the average ratio over entire timesteps. As reported in Fig. 4 (b), we confirm that it effectively calibrates the standard deviation of attention map similar to its original values.", + "bbox": [ + 75, + 763, + 468, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Initial Latent AdaIN", + "text_level": 1, + "bbox": [ + 76, + 832, + 272, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In artistic style transfer, the color tone generally takes up a significant portion of the style information. In this context, we observe that the style transfer only with attention-based", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8cfdad2fb268773913895461a5819b09eacd13147ba70e421af6bead6f86e5ba.jpg", + "image_caption": [ + "(a) Generated results only w/ style injection $(\\mathrm{K}^s,V^s)$" + ], + "image_footnote": [], + "bbox": [ + 519, + 89, + 746, + 180 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dde322ca136330611fdd146218e09cf6e5cd3e362c095d788c8b9605543f7f15.jpg", + "image_caption": [ + "njection $(\\mathrm{K}^s,V^s)$" + ], + "image_footnote": [], + "bbox": [ + 746, + 90, + 867, + 180 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7c4e3c4344e80896acccab14266518ab7f61d2f69ac574ece9e3eb52bc3f048a.jpg", + "image_caption": [ + "(b) Generated results only w/ style injection $(\\mathrm{Q}^s,\\mathrm{K}^s,V^s)$", + "Figure 5. Generated results only w/ style injection. (a) We observe that generated images only with attention-based style injection do not harmonize with the given style in the aspect of color tone. (b) To identify the effects of every feature in SA on color tones, we additionally include query in the style injection process. However, color tones still resemble those of content, concluding features in self-attention have less effect on the color tones." + ], + "image_footnote": [], + "bbox": [ + 519, + 199, + 746, + 291 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8c8d4174f83169458e3f8dc0035282fc426522048821c800fcd49fd4a8947b3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 199, + 867, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "style injection often fails in terms of capturing the color tone of the given style. As shown in Fig. 5 (a), textures and local patterns are successfully transferred to the content image while the color tone of the content image still remains. Furthermore, even with injecting the query, key, and value of styles, the resulting images still preserve the color tone of the content, as shown in Fig. 5 (b).", + "bbox": [ + 496, + 422, + 893, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As substituting the self-attention features has less effects to color tone, we analyze the other vital part of DM; initial latent noise. One of the recent discoveries in DM is that the DM struggles to synthesize purely white or black images [13]. Instead, they tend to generate images of median color as the initial noise is sampled from zero mean and unit variance. Thus, we hypothesize the statistics of initial noise largely affect the colors and brightness of generated images.", + "bbox": [ + 496, + 527, + 892, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on this assumption, we attempt to use the initial latent of style $z_{T}^{s}$ for the style transfer process. However, if we simply start to generate the image from style latent $z_{T}^{s}$ , the structural information of synthesized results also follows the style image and loses the structure of the content. To harness valuable information in both initial latents, we consider that the tone information is intricately connected with the channel statistics of the initial latent, following the principle underlying Style Loss [11] and AdaIN [17]. Thus, we employ AdaIN to modulate the initial latent for effective tone information transfer, represented as:", + "bbox": [ + 496, + 648, + 893, + 814 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {T} ^ {c s} = \\sigma \\left(z _ {T} ^ {s}\\right) \\left(\\frac {z _ {T} ^ {c} - \\mu \\left(z _ {T} ^ {c}\\right)}{\\sigma \\left(z _ {T} ^ {c}\\right)}\\right) + \\mu \\left(z _ {T} ^ {s}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 825, + 890, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu (\\cdot),\\sigma (\\cdot)$ denote channel-wise mean and standard deviation, respectively. Based on this, the initial latent $z_{T}^{cs}$", + "bbox": [ + 498, + 869, + 892, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8799", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/338e8deaf2ebc4d404dadea79f5e31bf92496db89fb348e14e2423eaae02ebae.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricOursAesPA-NetCASTStyTR2EFDMMASTAdaAttnArtFlowAdaConvAdaINDiffuseITInSTDiffStyle
ArtFID ↓28.80131.42034.68530.72034.60531.28230.35034.63031.85630.93340.72140.63341.464
FID ↓18.13119.76020.39518.89020.06218.19918.65821.25219.02218.24223.06521.57120.903
LPIPS ↓0.50550.51350.62120.54450.64300.62930.54390.55620.59100.60760.69210.80020.8931
CFSD ↓0.22810.24640.29180.30110.33460.30430.28620.29200.36000.31550.34280.67590.2819
", + "bbox": [ + 78, + 89, + 888, + 171 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison with conventional $\\left( {{3}^{\\text{rd }} - {11}^{\\text{th }}\\text{columns}}\\right)$ and diffusion model baselines $\\left( {{12}^{\\text{th }} - {14}^{\\text{th }}\\text{columns}}\\right)$", + "bbox": [ + 124, + 174, + 841, + 188 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "preserves content information from $z_T^c$ while aligning the channel-wise mean and standard deviation with $z_T^s$ .", + "bbox": [ + 75, + 200, + 468, + 233 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 244, + 209, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 270, + 282, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct all experiments in Stable Diffusion 1.4 pretrained on LAION dataset [36] and adopt DDIM sampling [39] with a total 50 timesteps ( $t = \\{1, \\dots, 50\\}$ ). For default settings for hyperparameters, we use $\\gamma = 0.75$ and $\\tau = 1.5$ , if they are not mentioned separately.", + "bbox": [ + 75, + 292, + 470, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Evaluation Protocol", + "text_level": 1, + "bbox": [ + 76, + 378, + 266, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Conventional style transfer methods typically utilize Style Loss [11] as both training objective and evaluation metric, so their results tend to overfit the Style Loss. Thus, for a fair comparison, we employ a recently proposed metric, ArtFID [46] which evaluates overall style transfer performances with consideration of both content and style preservation and also is known as strongly coinciding with human judgment. Specifically, ArtFID is computed as $(\\text{ArtFID} = (1 + \\text{LPIPS}) \\cdot (1 + \\text{FID}))$ . LPIPS [50] measures content fidelity between the stylized image and the corresponding content image, and FID [15] assesses the style fidelity between the stylized image and the corresponding style image.", + "bbox": [ + 75, + 402, + 468, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. Our evaluations employ content images from MS-COCO [27] dataset and style images from WikiArt [40] dataset. All input images are center-cropped to $512 \\times 512$ resolution. Also, for quantitative comparison, we randomly selected 20 content and 40 style images from each dataset, yielding 800 stylized images as StyTR² [9] has done.", + "bbox": [ + 75, + 590, + 468, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Content Feature Structural Distance (CFSD). In the style transfer evaluation, the assessment of content fidelity often relies on the LPIPS distance. However, since LPIPS utilizes the feature space of AlexNet [20] pre-trained for classification task on ImageNet [7], which is known as texture-biased [12]. Thus, the style information of the images can affect the LPIPS score. To mitigate this style influence, we additionally introduce Content Feature Structural Distance (CFSD) which is a distance measure that only considers the spatial correlation between image patches.", + "bbox": [ + 75, + 688, + 468, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In detail, we first define the correlation map between image patch features as follows. For a given image $I$ , we obtain feature maps $F \\in \\mathbb{R}^{hw \\times c}$ , which is the output feature of conv3 in VGG19 [38]. Then, we calculate the patch sim", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The similarity map $M = F \\times F^T$ , $M \\in \\mathbb{R}^{hw \\times hw}$ , which is a similarity map between every pair of features in $F$ . After that, for computing the distance between two patch similarity maps, we model the similarity between a single patch and the others as a probability distribution by applying softmax operation. Finally, the correlation map is represented as $S = [\\text{softmax}(M_i)]_{i=1}^{hw}$ , $S \\in \\mathbb{R}^{hw \\times hw}$ , where $M_i \\in \\mathbb{R}^{1 \\times hw}$ is a similarity map between $i^{\\text{th}}$ patch and the other patches.", + "bbox": [ + 496, + 200, + 890, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, CFSD is defined as KL-divergence between two correlation maps. In our case, we compute CFSD between the correlation map of the content $(S^c)$ and stylized images $(S^{cs})$ as follows:", + "bbox": [ + 496, + 321, + 890, + 382 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C F S D} = \\frac {1}{h w} \\sum_ {i = 1} ^ {h w} D _ {\\mathrm {K L}} \\left(S _ {i} ^ {c} | | S _ {i} ^ {c s}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 393, + 890, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3.Quantitative Comparison", + "text_level": 1, + "bbox": [ + 498, + 454, + 732, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our proposed method through comparison with twelve state-of-the-art methods, including nine conventional style transfer methods (AesPA-Net [16], CAST [52], StyTR $^2$ [9], EFDM [51], MAST [8], AdaAttN [28], ArtFlow [1], AdaConv [5], AdaIN [17]) and three diffusion-based style transfer methods (DiffuseIT [22], InST [53], DiffStyle [18]), which have a style image as input. We employ the publicly available implementations of all baselines, using their recommended configurations.", + "bbox": [ + 496, + 478, + 890, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with Conventional Style Transfer. As shown in Tab. 1, our method largely surpasses the conventional style transfer methods in terms of ArtFID, which is known as coinciding the human preference. In addition, the proposed method records the lowest FID, which denotes that stylized images highly resemble the target styles. For content fidelity metrics, ours shows superior scores in both CFSD and LPIPS. We point out that ours achieves much lower CFSD compared to other methods, which is the metric to only consider the spatial correlation.", + "bbox": [ + 496, + 621, + 890, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, we also emphasize that the proposed method can arbitrarily adjust the degree of style transfer by changing the $\\gamma$ , and the proposed method significantly surpasses all the other methods in terms of FID (style), when we match the value of LPIPS (content) (Fig. 10).", + "bbox": [ + 496, + 772, + 890, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with Diffusion-based Style Transfer. Our method demonstrates the best performance in terms of LPIPS, FID, and their combination (ArtFID) with a large", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8800", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Content Style Ours AesPA-Net CAST StyTR2 EFDM MAST AdaAttN AdaIN DiffuseIT InST DiffStyle", + "bbox": [ + 83, + 87, + 885, + 101 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/860ff3210846250b16d879d171e60f15c4647f18d2abd94bffd6a5026d2fa2ad.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison with conventional $(4^{\\mathrm{th}} - 10^{\\mathrm{th}}$ columns) and diffusion model baselines $(11^{\\mathrm{th}} - 13^{\\mathrm{th}}$ columns)" + ], + "image_footnote": [], + "bbox": [ + 78, + 101, + 890, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2838e8ad98d425397cce2a5d39e7ef935a30c31d0c61a70bce8bd7758a67ed21.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison with best ArtFID (AdaAttN) and most recently proposed baselines (AesPA-Net) with additional zooming details" + ], + "image_footnote": [], + "bbox": [ + 94, + 318, + 450, + 449 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a075caf6dea678e94b4921075cb4d81fd5903e126d48e55406b153bbf8e7277b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricDiffuseITInSTDiffStyleOurs
Time (sec)792.8816.9355.912.4
", + "bbox": [ + 133, + 498, + 410, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison of inference time of diffusion-based methods for style transferring a given style and content pair", + "bbox": [ + 75, + 542, + 467, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "margin, as shown in Tab. 1. One significant factor for diffusion models is their running times, since they require several steps to synthesize a single image and it requires inevitable time cost. Hence, we measure the inference time for a pair of content and style images on a single Titan RTX GPU, as shown in Tab. 2. Our method requires a total of 12.4 seconds, with 8.2 seconds for DDIM inversions and 4.2 seconds for sampling costs. As reported, we validate the proposed method significantly faster than other methods, even exploiting large-scale DM. This faster speed comes from the fact that the proposed methods can use the much smaller steps of DDIM inversion, because we additionally utilize the features collected during inversion steps, largely reducing the necessity for perfect inversion of content and style.", + "bbox": [ + 75, + 579, + 467, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 76, + 801, + 297, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Conventional Style Transfer. As shown in Fig. 6, we observe that our method tends to highly preserve the structural information of the content image, while also transferring the style well. For instance, as shown in the third row, ours retains the structure of the", + "bbox": [ + 75, + 825, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "bridge, but the baselines struggle to preserve structure or transfer the style. We also provide the qualitative comparison with zooming details in Fig. 7 and Supplementary.", + "bbox": [ + 498, + 320, + 890, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Diffusion-based Style Transfer. We also compare our method with recent diffusion-based style transfer baselines [18, 23, 53]. As shown in Fig. 6, we observe the proposed technique transfers the style to the content well. On the other hand, baselines often lose the structure of content or fail to transfer the style, when an arbitrary content style pair is given. For instance, DiffuseIT and DiffStyle suffer from generating shape and visually plausible images or drop the original content. Differently, InST synthesizes the realistic images, while struggling to transfer style (1 $^{\\text{st}}$ row) or change content of image (2 $^{\\text{nd}}$ , 3 $^{\\text{rd}}$ rows).", + "bbox": [ + 498, + 372, + 890, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 551, + 651, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the effectiveness of the proposed components, we conduct ablation studies in both quantitative and qualitative ways. As shown in Fig. 8 and Tab. 3, style injection is significant for guiding the style and content of given images (Config. B). Besides, initial latent AdaIN has a large portion of transferring the color tone of style (Config. D). Attention temperature scaling is in charge of enhancement of quality in synthesized results such as sharpening details and resolving blurriness. For instance, this scaling jointly reduces the FID and LPIPS (Config. $\\mathrm{A}^*$ vs. C in Tab. 3). For more detailed analysis, we provide quantitative metrics with the style-content trade-off, while changing the attention scaling parameter $\\tau$ in Fig. 10 (b). As reported, attention scaling effectively reduces both FID and LPIPS, proving its effects on the preservation of content and capability of style transfer ( $\\tau = 1.0$ vs. $\\tau = 1.5$ ).", + "bbox": [ + 496, + 575, + 890, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.6. Additional Analysis", + "text_level": 1, + "bbox": [ + 500, + 830, + 687, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Content-Style Trade-Off. Our proposed method offers flexible control of the trade-off relation between content and style fidelity by adjusting the parameter $\\gamma$ , as discussed in", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8801", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5b1b936a45b01d2bdd11131aae08a19b15765285713166401ef2ee2900e905a9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ConfigurationArtFIDFIDLPIPS
AOurs (γ = 0.75, default)28.8018.130.505
A*Ours (γ = 0.6)27.9717.210.535
B- Style Inject. (Sec. 4.1)43.7227.130.554
C- Attn. Scaling (Sec. 4.2)29.0217.810.542
D- Init. AdaIN (Sec. 4.3)29.2620.050.390
", + "bbox": [ + 78, + 88, + 475, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study on proposed components", + "bbox": [ + 125, + 200, + 419, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b89ab7b4bcfdd501d4f0c4ece99c1f74adbac5ab3e9aafff8a18fd52f34f13c6.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison with ablation studies" + ], + "image_footnote": [], + "bbox": [ + 84, + 222, + 459, + 363 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4cf5a63c61f5638d6fffff5dc9acbbc0b67f136c825e93b8776c146c9759bf5c.jpg", + "image_caption": [ + "Figure 9. Visualization of effects of query preservation ratio $\\gamma$" + ], + "image_footnote": [], + "bbox": [ + 76, + 386, + 460, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sec. 4.1. In detail, we compute the FID and LPIPS while varying $\\gamma$ within the range [0.3, 1] with a step size of 0.1. As shown in Fig. 10 (a), our method surpasses baseline methods across all ranges of content and style fidelity. This result implies the proposed method significantly outperforms the other methods, when we match style or content metric to the compared model by adjusting the $\\gamma$ of ours. Note that, dotted lines refer to our model reported in Tab. 1.", + "bbox": [ + 75, + 585, + 468, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also visualize the effects of the style-content trade-off by synthesizing images by adjusting $\\gamma$ . As shown in Fig. 9, the lower $\\gamma$ highly reflects the style while losing the content of the given image, and vice versa. This characteristic of the proposed method suggests that the users can adjust the degree of style by following their personal preferences.", + "bbox": [ + 75, + 709, + 468, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Study on the value of $\\tau$ . We observe that the gradual increase of $\\tau$ enhances the performance of style transfer, although its effects on enhancement become smaller as $\\tau$ goes larger, as shown in Fig. 10 (b). This result implies that the attention temperature scaling effectively works with a simple modification of the magnitude of the attention map.", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0574ed7e223feb7c0b3309243ecdcabf39ec3c7cba7af2ebe674151a020c830e.jpg", + "image_caption": [ + "Figure 10. Style-content trade-offs" + ], + "image_footnote": [], + "bbox": [ + 491, + 92, + 736, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d13886104303c34f63fe8e4d7a7cc55500466e7e10422acf2cfaad9f948dd77d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 92, + 883, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f76cbcee65f8d474e662ff237a8efaf946791fe47816741a406344a17535b5e.jpg", + "image_caption": [ + "Figure 11. Comparison with text-guided style transfer methods" + ], + "image_footnote": [], + "bbox": [ + 516, + 268, + 879, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with text-guided style transfer. We additionally compare the proposed method with the style transfer methods [21, 41] which are conditioned on the textual inputs. As text-guided methods tend to modify the style largely, we use $\\gamma = 0.3$ for this experiment. Since the text condition hardly contains all the information in the style image such as texture and color tones, the transferred results less resemble the target style, as shown in Fig. 11. Differently, we validate that the proposed method successfully transfers the style with high fidelity.", + "bbox": [ + 496, + 385, + 890, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 545, + 617, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our work addresses the challenges associated with diffusion model-based style transfer methods, which often require time-consuming optimization steps or struggle to leverage the generative potential of large-scale diffusion models. To this end, we propose the method of adapting the pre-trained large-scale diffusion model on style transfer in a training-free way. Our method focuses on manipulating the features of self-attention layers, akin to the cross-attention mechanism, by substituting the key and value during the content generation process with those of the style. Furthermore, we propose the query preservation and attention temperature scaling to mitigate the issue of disruption of content, and initial latent AdaIN to handle the disharmonious color (failure to transfer the colors of style). Experimental results show the superiority of our proposed method over state-of-the-art techniques in previous baselines.", + "bbox": [ + 496, + 571, + 890, + 813 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 821, + 658, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by MSIT/IITP (No. 2022-0-00680, 2019-0-00421, 2020-0-01821, 2021-0-02068), and MSIT&KNPA/KIPoT (Police Lab 2.0, No. 210121M06).", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8802", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jie An, Siyu Huang, Yibing Song, Dejing Dou, Wei Liu, and Jiebo Luo. Artflow: Unbiased image style transfer via reversible neural flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 862-871, 2021. 6", + "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 1, 3", + "[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22560-22570, 2023. 3", + "[4] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23040-23050, 2023. 1", + "[5] Prashanth Chandran, Gaspard Zoss, Paulo Gotardo, Markus Gross, and Derek Bradley. Adaptive convolutions for structure-aware style transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7972-7981, 2021. 6", + "[6] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In ICLR 2023 (Eleventh International Conference on Learning Representations), 2023. 1, 3", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6", + "[8] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM international conference on multimedia, pages 2719-2727, 2020. 6", + "[9] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11326-11336, 2022. 6", + "[10] Martin Nicolas Everaert, Marco Bocchio, Sami Arpa, Sabine Susstrunk, and Radhakrishna Achanta. Diffusion in style. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2251-2261, 2023. 1, 3", + "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2, 5, 6", + "[12] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 6" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Nocholas Guttenberg. Diffusion with offset noise. https://www.crosslabs.org/blog/diffusion-with-offset-noise, 2023.5", + "[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 3, 4", + "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[16] Kibeom Hong, Seogkyu Jeon, Junsoo Lee, Namhyuk Ahn, Kunhee Kim, Pilhyeon Lee, Daesik Kim, Youngjung Uh, and Hyeran Byun. Aespa-net: Aesthetic pattern-aware style transfer networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22758-22767, 2023. 6", + "[17] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 5, 6", + "[18] Jaeseok Jeong, Mingi Kwon, and Youngjung Uh. Training-free style transfer emerges from h-space in diffusion models. arXiv preprint arXiv:2303.15403, 2023. 1, 3, 6, 7", + "[19] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Conference on Computer Vision and Pattern Recognition 2023, 2023. 1, 3", + "[20] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 6", + "[21] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 8", + "[22] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. arXiv preprint arXiv:2209.15264, 2022. 6", + "[23] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 3, 7", + "[24] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 2", + "[25] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017.", + "[26] Yijun Li, Ming-Yu Liu, Xueting Li, Ming-Hsuan Yang, and Jan Kautz. A closed-form solution to photorealistic image" + ], + "bbox": [ + 501, + 92, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8803", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "stylization. In Proceedings of the European conference on computer vision (ECCV), pages 453-468, 2018. 2", + "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[28] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 6", + "[29] Ming Lu, Hao Zhao, Anbang Yao, Yurong Chen, Feng Xu, and Li Zhang. A closed-form solution to universal style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5952-5961, 2019. 2", + "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1", + "[31] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880–5888, 2019. 2", + "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. 3", + "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3", + "[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1", + "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1", + "[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 6", + "[37] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6", + "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 4, 6", + "[40] Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394-409, 2019. 6", + "[41] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2, 3, 8", + "[42] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/ diffusers, 2022.3", + "[43] Huan Wang, Yijun Li, Yuehai Wang, Haoji Hu, and Ming-Hsuan Yang. Collaborative distillation for ultra-resolution universal style transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1860-1869, 2020. 2", + "[44] Zhizhong Wang, Lei Zhao, Haibo Chen, Lihong Qiu, Qihang Mo, Sihuan Lin, Wei Xing, and Dongming Lu. Diversified arbitrary style transfer via deep feature perturbation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7789-7798, 2020. 2", + "[45] Zhizhong Wang, Lei Zhao, and Wei Xing. *Style diffusions: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 1, 3", + "[46] Matthias Wright and Björn Ommer. Artfid: Quantitative evaluation of neural style transfer. In DAGM German Conference on Pattern Recognition, pages 560-576. Springer, 2022. 6", + "[47] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. arXiv preprint arXiv:2303.08622, 2023. 1, 3", + "[48] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 1", + "[49] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. arXiv preprint arXiv:2305.15347, 2023. 2", + "[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[51] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In Proceedings of" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8804", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8035-8045, 2022. 6", + "[52] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022. 2, 6", + "[53] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10146-10156, 2023. 1, 3, 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 271 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8805", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/43978830-73bd-4db5-8b8c-57888ce93c6e_model.json b/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/43978830-73bd-4db5-8b8c-57888ce93c6e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b9e3bb054d28afd13931a1c8019569cdd9f97756 --- /dev/null +++ b/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/43978830-73bd-4db5-8b8c-57888ce93c6e_model.json @@ -0,0 +1,2367 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.177, + 0.131, + 0.795, + 0.177 + ], + "angle": 0, + "content": "Style Injection in Diffusion: A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.203, + 0.663, + 0.24 + ], + "angle": 0, + "content": "Jiwoo Chung*, Sangeek Hyun*, Jae-Pil Heo† Sungkyunkwan University" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.242, + 0.666, + 0.256 + ], + "angle": 0, + "content": "{wldn0202, hsi1032, jaepilheo}@g.skku.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.702 + ], + "angle": 0, + "content": "Despite the impressive generative capabilities of diffusion models, existing diffusion model-based style transfer methods require inference-stage optimization (e.g. fine-tuning or textual inversion of style) which is time-consuming, or fails to leverage the generative ability of large-scale diffusion models. To address these issues, we introduce a novel artistic style transfer method based on a pre-trained large-scale diffusion model without any optimization. Specifically, we manipulate the features of self-attention layers as the way the cross-attention mechanism works; in the generation process, substituting the key and value of content with those of style image. This approach provides several desirable characteristics for style transfer including 1) preservation of content by transferring similar styles into similar image patches and 2) transfer of style based on similarity of local texture (e.g. edge) between content and style images. Furthermore, we introduce query preservation and attention temperature scaling to mitigate the issue of disruption of original content, and initial latent Adaptive Instance Normalization (AdaIN) to deal with the disharmonious color (failure to transfer the colors of style). Our experimental results demonstrate that our proposed method surpasses state-of-the-art methods in both conventional and diffusion-based style transfer baselines. Codes are available at github.com/jiwoogit/StyleID." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.736, + 0.21, + 0.752 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.762, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Recent advances in Diffusion Models (DMs) have led to breakthroughs in various generative applications such as text-to-image synthesis [30, 33, 35] and image or video editing [2, 4, 6, 14, 19, 41, 48]. One of these efforts is also applied to the task of style transfer [10, 18, 45, 47, 53]; given style and content images, modifying the style of the content image to possess the given style." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.296, + 0.889, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.363, + 0.835, + 0.375 + ], + "angle": 0, + "content": "(a) General usage of self-attention in DM" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.378, + 0.889, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.513, + 0.869, + 0.527 + ], + "angle": 0, + "content": "(b) Our adaptation of self-attention for style transfer" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.538, + 0.895, + 0.664 + ], + "angle": 0, + "content": "Figure 1. Manipulation of self-attention features for style transfer. (a) General self-attention (SA) deploys the query, key, and value features from a single image in both the training and inference phases. (b) At inference phase, we suggest that manipulating features of self-attention of pre-trained large-scale DM is an effective way to transfer the styles; injection of key and value of styles into SA of contents is a proper way for transferring styles. As a result, style-injected content \\( z_{t - 1}^{c} \\) would maintain contents while modifying its style to resemble the target style." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "General approaches for diffusion model-based style transfer leverage the generative capability of pre-trained DM. Some of these works focus on explicit disentangling style and content for interpretable and controllable style transfer [45], or inversion of the style image into the textual latent space of a large-scale text-to-image DM [53]. However, these methods additionally require gradient-based optimization for fine-tuning and textual inversion [34] for each style image, which is time-consuming. Without this issue, DiffStyle [18] introduces training-free style transfer, but they are known to be hardly applicable to Latent Diffusion Model [33] which is widely adopted for training large-scale text-to-image DM such as Stable Diffusion [33], hindering the users from taking advantage of the prominent generative ability of large-scale models." + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.875, + 0.214, + 0.888 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.888, + 0.23, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.875, + 0.23, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "8795" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.47, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.231, + 0.414, + 0.242 + ], + "angle": 0, + "content": "(b) query-key similarity between patches of similar texture" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.249, + 0.47, + 0.388 + ], + "angle": 0, + "content": "Figure 2. Desirable attributes of self-attention (SA) for style transfer. (a) Visualization of query by PCA shows that query features well-reflect similarities among patches. That is, style transfer employing SA can preserve the original content, as content patches with similarities tend to receive similar attention scores from a corresponding style image patch. (b) We visualize a similarity map between the blue box (edge) query of the content image, and key of the style image. Thanks to the features representation of large-scale DM encompassing texture and semantics, a query exhibits higher similarity to keys that share a similar style, such as edges." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.469, + 0.61 + ], + "angle": 0, + "content": "In this paper, we focus on extending the training-free style transfer to its application on large-scale pre-trained DM. We start from the observation of recent advances in image-to-image translation based on large-scale DM; they uncover the image editing capability of attention layers. Notably, Plug-and-play [41] shows that the residual block and the attention map of self-attention (SA) determine the spatial layout of generated images. Also, Prompt-to-Prompt [14] locally edits the image by replacing key and value of cross-attention (CA) obtained from text prompt, while keeping their original attention maps. That is, all these works suggest that 1) attention maps determine the spatial layout and 2) key and value of CA adjust the content to fill." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Inspired by the aforementioned methods, we newly argue that manipulating the SA layer is an effective way to transfer the styles (Fig. 1). Specifically, similar to CA, we substitute the key and value of SA and observe that the generated images are still visually plausible and naturally incorporate the elements of the substituted image into the original image. This observation motivates us to propose a style transfer technique based on self-attention, which combines the styles (textures) of a specific image with the content (semantics and spatial layout) of different images. Furthermore, we highlight that SA layer has desirable characteristics for style transfer. First, as shown in Fig. 2 (a), in SA-based style transfer, the content image patches (query) that share semantic similarities engage with a similar style (key), thereby maintaining the relationship among these content image patches after the transfer. Next, thanks to the powerful feature representation of large-scale DM [49], each patch of the query reveals higher similarity to keys which has similar texture and semantics. For instance, in Fig. 2 (b)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "we can observe that the query feature of content within the blue box exhibits a high similarity to the key features of style with similar edge texture. This encourages the model to transfer style based on the similarity of local texture (e.g. edge) between content and style." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.169, + 0.892, + 0.56 + ], + "angle": 0, + "content": "As a result, our method aims to transfer the textures of the style image to the content images by manipulating self-attention features of pre-trained large-scale DM without any optimization. To this end, we first propose an attention-based style injection method. The basic idea of it is substituting the content's key and value of SA with those of the style image, especially layers in the latter part of decoder which are relevant to the local textures. As mentioned in above paragraph, exchanged styles are well aligned with the content and texture of original image, exploiting the similarity-based attention mechanism. With the proposed style injection, we observe that the local texture patterns are successfully transferred, but there still are remaining problems such as disruption of original content and disharmonious colors. To handle these problems, we additionally propose the following techniques; query preservation, attention temperature scaling, and initial latent AdaIN. Query preservation makes the reverse diffusion process to retain the spatial structure of original content by preserving the query of the content image in the SA. Attention temperature scaling also aims to keep the structure of content by dealing with the blurred self-attention map introduced from the substitution of key. Lastly, initial latent AdaIN corrects inharmonious color problem, referring that the color distribution of style images is not properly transferred, by modulating the statistics of initial noise in the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.563, + 0.86, + 0.577 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.891, + 0.654 + ], + "angle": 0, + "content": "- We propose a style transfer method exploiting the large-scale pre-trained DM by simple manipulation of the features in self-attention; substituting key and value of content with those of styles without any requirements of optimization or supervision (e.g. text)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.891, + 0.713 + ], + "angle": 0, + "content": "- We further improve the naive approach for style transfer to properly adapt the styles by proposing three components; query preservation, attention temperature scaling, and initial latent AdaIN." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.891, + 0.759 + ], + "angle": 0, + "content": "- Extensive experiments on the style transfer dataset validate the proposed method significantly outperforms previous methods and achieves state-of-the-art performance." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.58, + 0.891, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.774, + 0.641, + 0.79 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.88, + 0.816 + ], + "angle": 0, + "content": "2.1. Diffusion Model-based Neural Style Transfer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Neural style transfer [11, 24-26, 29, 31, 43, 44, 52] is an example-guided image generation task that transfers the style of one image onto another while retaining the content of the original. In the realm of diffusion models, neural style transfer has evolved by leveraging the generative capability" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8796" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.212 + ], + "angle": 0, + "content": "of pre-trained diffusion models. For instance, InST [53] introduced a textual inversion-based approach, aiming to map a given style into corresponding textual embeddings. StyleDiffusion [45] aimed to disentangle style and content by introducing CLIP-based style disentanglement loss for fine-tuning DM for style transfer. Also, several approaches utilize the text input as a style condition or for determining the content to synthesize [10, 47]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.469, + 0.322 + ], + "angle": 0, + "content": "Conversely, DiffStyle [18] proposed a training-free style transfer method that leverages \\( h \\)-space [23] and adjusts skip connections for effectively conveying style and content information, respectively. However, when DiffStyle is applied to Stable Diffusion [33, 42], their behavior is quite different from typical style-transfer methods; not only textures but also semantics such as spatial layout are also changed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.469, + 0.386 + ], + "angle": 0, + "content": "To address these limitations, we propose a novel algorithm that harmoniously merges style and content features within the self-attention layers of Stable Diffusion without any optimization process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.404, + 0.405, + 0.421 + ], + "angle": 0, + "content": "2.2. Attention-based Image Editing in DM" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Following the remarkable advances achieved by pre-trained text-to-image DMs [32, 42], there have been numerous image editing works [2, 6, 19, 37] utilizing these DMs. Notably, Prompt-to-Prompt [14] proposed text-based local image editing by manipulating the cross-attention map. Specifically, they observe that cross-attention largely contributes to modeling the relation between the spatial layout of the image to each word in the prompt. Hence, they substitute the original words and cross-attention map with desirable ones, obtaining edited images matched with text conditions. Subsequently, Plug-and-play [41] introduces text-guided image-to-image translation method. They found that the spatial features (i.e. feature from residual block) and self-attention map determine the spatial layout of the synthesized image. Thus, while generating a new image with the given text condition, they guide the diffusion model with features and attention map from the original image for preserving the original spatial layout. Recently, MasaCtrl [3] proposes mutual self-attention control for consistent image editing using text prompts. In detail, they retain the source image's key and value of the self-attention layers, while conditioning the model with desired text prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Along with these works, we recognize the potential of attention maps in representing spatial information. However, different from the aforementioned methods concentrating on exploiting textual condition, we focus on conditioning by style and content images composed of two images from distinct styles. By combining the features in self-attention layers of both style and content images with precise adjustment of statistics in intermediate representations, we transfer the texture of the content image to the given style." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.628, + 0.108 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.117, + 0.892, + 0.208 + ], + "angle": 0, + "content": "Latent Diffusion Model (LDM) [33] is a type of diffusion model trained in the low dimensional latent space to focus on semantic bits of data and reduce computation costs. Given an image \\( x \\in \\mathbb{R}^{H \\times W \\times 3} \\), the encoder \\( \\mathcal{E} \\) encodes \\( x \\) into the latent representation \\( z \\in \\mathbb{R}^{h \\times w \\times c} \\) and decoder reconstructs the image from the latent." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.21, + 0.892, + 0.284 + ], + "angle": 0, + "content": "With the pretrained encoder, they encode the entire images in the dataset and train a diffusion model on latent space \\( z \\), by predicting noise \\( \\epsilon \\) from the noised version of latent \\( z_{t} \\) at time step \\( t \\). The corresponding training objective is" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.29, + 0.892, + 0.309 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {L D M}} = \\mathbb {E} _ {z, \\epsilon , t} [ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, y) \\| _ {2} ^ {2} ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.892, + 0.38 + ], + "angle": 0, + "content": "where \\(\\epsilon \\in \\mathcal{N}(0,1)\\) is a noise, \\(t\\) is the number of time steps which uniformly sampled from \\(\\{1,\\dots,T\\}\\), \\(y\\) is a condition, and \\(\\epsilon_{\\theta}\\) is a neural network which predicts the noise added to the \\(z\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.892, + 0.518 + ], + "angle": 0, + "content": "In our work, we utilize Stable Diffusion (SD) [33] which is the only publicized large-scale pre-trained DM. In the case of SD, \\( y \\) is a text, and \\( \\epsilon_{\\theta} \\) is a U-Net architecture in which a block for each resolution comprises a residual block, self-attention block (SA), and cross-attention block (CA), sequentially. Among these modules, we focus on the SA block to transfer the styles, as discussed in Sec. 1. Given a feature \\( \\phi \\) after the residual block, the self-attention block performs as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.533, + 0.843, + 0.551 + ], + "angle": 0, + "content": "\\[\nQ = W _ {Q} (\\phi), K = W _ {K} (\\phi), V = W _ {V} (\\phi),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.553, + 0.891, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\phi_ {\\text {o u t}} = \\operatorname {A t t n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) \\cdot V, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.601, + 0.892, + 0.662 + ], + "angle": 0, + "content": "where \\(d\\) denotes the dimension of the projected query, and \\(W_{(\\cdot)}\\) is a projection layer. Note that, we don't use any text conditions, so the variable \\(y\\) is always an empty text prompt \\((\\ldots)\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.678, + 0.59, + 0.693 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In this paper, we aim to solve artistic style transfer by leveraging the generative capability of a pre-trained large-scale text-to-image diffusion model. Briefly, artistic style transfer is the task of modifying the style of a given content image \\( I^{c} \\) to that of style image \\( I^{s} \\). Then, the stylized image \\( I^{cs} \\) would maintain the semantic content of \\( I^{c} \\) while its style (such as texture) is transferred from \\( I^{s} \\). For simplicity, we skip the explanations about the encoding and decoding process of the autoencoder in the LDM. Instead, we focus on elaborating the proposed method in the aspect of the diffusion process. Thus, in the following sections, we regard the content, style, and stylized images same as their encoded counterparts \\( z_{0}^{c}, z_{0}^{s} \\), and \\( z_{0}^{cs} \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8797" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.093, + 0.581, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.093, + 0.885, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.368, + 0.892, + 0.522 + ], + "angle": 0, + "content": "Figure 3. Overall framework. (Left) Illustration for the proposed style transfer method. We first invert content image \\( z_0^c \\) and style image \\( z_0^s \\) into the latent noise space as \\( z_T^c \\) and \\( z_T^s \\), respectively. Then, we initialize the initial noise of stylized image \\( z_T^{cs} \\) from initial latent AdaIN (Sec. 4.3) which combines the content and style noise, \\( z_T^c \\) and \\( z_T^s \\). While performing the reverse diffusion process with \\( z_T^{cs} \\), we inject the information of content and style by attention-based style injection (Sec. 4.1) and attention temperature scaling (Sec. 4.2). (Right) Detailed explanation of style injection and initial noise AdaIN. Style injection is basically the manipulation of self-attention (SA) layer during the reverse diffusion process. Specifically, at time step \\( t \\), we substitute the key \\( (K_t^{cs}) \\) and value \\( (V_t^{cs}) \\) in SA of stylized image with those of style features, \\( K_t^s \\) and \\( V_t^s \\), from identical timestep \\( t \\). At the same time, we preserve the content information by blending the query of content \\( Q_t^c \\) and query of stylized image \\( Q_t^{cs} \\). Finally, we scale the magnitude of the attention map to deal with the magnitude decrease that the substitution of feature leads to. Initial latent AdaIN produces the initial noise \\( z_T^{cs} \\) by combining style noise \\( z_T^s \\) and content noise \\( z_T^s \\). Specifically, we modify the channel statistics of \\( z_T^c \\) to resemble the statistics of \\( z_T^s \\) and regard it as \\( z_T^{cs} \\). We observe this operation enables us to keep the spatial layout of content image while well-reflecting the color tones of a given style image." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.533, + 0.353, + 0.549 + ], + "angle": 0, + "content": "4.1. Attention-based Style Injection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.756 + ], + "angle": 0, + "content": "We start from the observation in previous image-to-image translation methods, especially Prompt-to-Prompt [14]. The key idea of their method is changing the text condition for cross-attention (CA) while keeping the attention map. Since the attention map affects the spatial layout of output, substituted text conditions determine what to draw in the generated image, and these conditions are actually key and value in CA. Inspired by them, we manipulate the features in self-attention layer as like cross-attention, regarding the features from style image \\( I^{s} \\) as the condition. Specifically, in the generation process, we substitute the key and value of content image with those of style for transferring the texture of style image into the content image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.761, + 0.469, + 0.882 + ], + "angle": 0, + "content": "To this end, we first obtain the latent for content and style images with DDIM inversion [39], and then collect the SA features of style image over the DDIM inversion process. Specifically, for pre-defined timesteps \\(t = \\{0,\\dots,T\\}\\), style and content images \\(z_0^c\\) and \\(z_0^s\\) are inverted from image \\((t = 0)\\) to gaussian noise \\((t = T)\\). During DDIM inversion, we also collect query features of content \\((Q_t^c)\\) and key and value features of style \\((K_t^s,V_t^s)\\) at every time steps." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "After that, we initialize stylized latent noise \\( z_T^{cs} \\) by copy-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.715 + ], + "angle": 0, + "content": "ing content latent noise \\( z_{T}^{c} \\). Then, we transfer the target style to the stylized latent by injecting the key \\( K_{t}^{s} \\) and value \\( V_{t}^{s} \\) collected from the style into SA layer, instead of the original key \\( K_{t}^{cs} \\) and value \\( V_{t}^{cs} \\), when performing the entire reverse process of stylized latent \\( z_{t}^{cs} \\). However, only applying this substitution can lead to content disruption, since the content of stylized latent would be progressively changed as attended value changes. Hence, we propose query preservation to maintain original content. Simply, we blend query of stylized latent \\( Q_{t}^{cs} \\) and that of content \\( Q_{t}^{c} \\) for the entire reverse process. These style injection and query preservation processes at time step \\( t \\) are expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.729, + 0.891, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\tilde {Q} _ {t} ^ {c s} = \\gamma \\times Q _ {t} ^ {c} + (1 - \\gamma) \\times Q _ {t} ^ {c s}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.766, + 0.891, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\phi_ {\\text {o u t}} ^ {\\mathrm {c s}} = \\operatorname {A t t n} \\left(\\tilde {Q} _ {t} ^ {c s}, K _ {t} ^ {s}, V _ {t} ^ {s}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is degree of blending in range of [0, 1]. In addition, we apply these operations on the latter layers of decoder (7- \\(12^{\\mathrm{th}}\\) decoder layers in SD) relevant to local textures. We also highlight that the proposed method can adjust the degree of style transfer by changing query preservation ratio \\(\\gamma\\). Specifically, higher \\(\\gamma\\) maintains more content, while lower \\(\\gamma\\) strengthens effects of style transfer." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8798" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.295, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.194, + 0.255, + 0.206 + ], + "angle": 0, + "content": "(a) Std. of attention map" + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.091, + 0.465, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.194, + 0.468, + 0.206 + ], + "angle": 0, + "content": "(b) Ratio of std. of attention map" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.212, + 0.471, + 0.31 + ], + "angle": 0, + "content": "Figure 4. Visualization of the standard deviation of attention map before softmax. (a) Attention-based style injection reduces the standard deviation of self-attention map. Original denotes SA maps from the generation process without style injection. We use both style and content images for generation. (b) We compute the ratio between attention maps w/ and w/o style injection. For the std of original image, we use averaged std. of content and style." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.321, + 0.353, + 0.337 + ], + "angle": 0, + "content": "4.2. Attention Temperature Scaling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.344, + 0.47, + 0.493 + ], + "angle": 0, + "content": "Attention map is computed by scaled dot-product between query and key features. During training, query and key features in the SA layer originate from an identical image. However, if we substitute the key features with those of style images, the magnitude of similarity would be overall lowered as style and content are highly likely to be irrelevant. Thus, the computed attention map can be blurred or smoothed, and it would further make output images unsharp, which is detrimental to capturing both content and style information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.495, + 0.47, + 0.616 + ], + "angle": 0, + "content": "To quantify this issue, we measure the standard deviation of attention map, while ablating the attention-based style injection. In detail, we calculate the attention map before applying softmax, which is scaled-dot product between query and key. As shown in Fig. 4 (a), we validate that this style injection tends to lower the standard deviation of the attention map over the entire timesteps. That is, attention maps after softmax with style injection would be overly smooth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.616, + 0.47, + 0.706 + ], + "angle": 0, + "content": "To rectify the attention map sharper, we introduce an attention temperature scaling parameter. In detail, we multiply the attention map before softmax by a constant temperature scaling parameter \\(\\tau\\) larger than 1. Thus, the attention map after softmax would be sharper than its original values. The modified attention process is represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.716, + 0.468, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} _ {\\tau} \\left(\\tilde {Q _ {t} ^ {c s}}, K _ {t} ^ {s}, V _ {t} ^ {s}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {\\tau \\tilde {Q _ {t} ^ {c s}} \\left(K _ {t} ^ {s}\\right) ^ {T}}{\\sqrt {d}}\\right) \\cdot V _ {t} ^ {s}, \\tau > 1. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.47, + 0.824 + ], + "angle": 0, + "content": "We use \\(\\tau = 1.5\\) as a default setting, which is the average ratio over entire timesteps. As reported in Fig. 4 (b), we confirm that it effectively calibrates the standard deviation of attention map similar to its original values." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.273, + 0.847 + ], + "angle": 0, + "content": "4.3. Initial Latent AdaIN" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In artistic style transfer, the color tone generally takes up a significant portion of the style information. In this context, we observe that the style transfer only with attention-based" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.09, + 0.747, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.182, + 0.842, + 0.194 + ], + "angle": 0, + "content": "(a) Generated results only w/ style injection \\((\\mathrm{K}^s,V^s)\\)" + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.091, + 0.869, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.182, + 0.866, + 0.194 + ], + "angle": 0, + "content": "njection \\((\\mathrm{K}^s,V^s)\\)" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.2, + 0.747, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.54, + 0.294, + 0.853, + 0.308 + ], + "angle": 0, + "content": "(b) Generated results only w/ style injection \\((\\mathrm{Q}^s,\\mathrm{K}^s,V^s)\\)" + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.2, + 0.869, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.313, + 0.894, + 0.41 + ], + "angle": 0, + "content": "Figure 5. Generated results only w/ style injection. (a) We observe that generated images only with attention-based style injection do not harmonize with the given style in the aspect of color tone. (b) To identify the effects of every feature in SA on color tones, we additionally include query in the style injection process. However, color tones still resemble those of content, concluding features in self-attention have less effect on the color tones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.423, + 0.894, + 0.528 + ], + "angle": 0, + "content": "style injection often fails in terms of capturing the color tone of the given style. As shown in Fig. 5 (a), textures and local patterns are successfully transferred to the content image while the color tone of the content image still remains. Furthermore, even with injecting the query, key, and value of styles, the resulting images still preserve the color tone of the content, as shown in Fig. 5 (b)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.893, + 0.649 + ], + "angle": 0, + "content": "As substituting the self-attention features has less effects to color tone, we analyze the other vital part of DM; initial latent noise. One of the recent discoveries in DM is that the DM struggles to synthesize purely white or black images [13]. Instead, they tend to generate images of median color as the initial noise is sampled from zero mean and unit variance. Thus, we hypothesize the statistics of initial noise largely affect the colors and brightness of generated images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.894, + 0.815 + ], + "angle": 0, + "content": "Based on this assumption, we attempt to use the initial latent of style \\( z_{T}^{s} \\) for the style transfer process. However, if we simply start to generate the image from style latent \\( z_{T}^{s} \\), the structural information of synthesized results also follows the style image and loses the structure of the content. To harness valuable information in both initial latents, we consider that the tone information is intricately connected with the channel statistics of the initial latent, following the principle underlying Style Loss [11] and AdaIN [17]. Thus, we employ AdaIN to modulate the initial latent for effective tone information transfer, represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.826, + 0.892, + 0.86 + ], + "angle": 0, + "content": "\\[\nz _ {T} ^ {c s} = \\sigma \\left(z _ {T} ^ {s}\\right) \\left(\\frac {z _ {T} ^ {c} - \\mu \\left(z _ {T} ^ {c}\\right)}{\\sigma \\left(z _ {T} ^ {c}\\right)}\\right) + \\mu \\left(z _ {T} ^ {s}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.893, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mu (\\cdot),\\sigma (\\cdot)\\) denote channel-wise mean and standard deviation, respectively. Based on this, the initial latent \\(z_{T}^{cs}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8799" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.09, + 0.89, + 0.172 + ], + "angle": 0, + "content": "
MetricOursAesPA-NetCASTStyTR2EFDMMASTAdaAttnArtFlowAdaConvAdaINDiffuseITInSTDiffStyle
ArtFID ↓28.80131.42034.68530.72034.60531.28230.35034.63031.85630.93340.72140.63341.464
FID ↓18.13119.76020.39518.89020.06218.19918.65821.25219.02218.24223.06521.57120.903
LPIPS ↓0.50550.51350.62120.54450.64300.62930.54390.55620.59100.60760.69210.80020.8931
CFSD ↓0.22810.24640.29180.30110.33460.30430.28620.29200.36000.31550.34280.67590.2819
" + }, + { + "type": "table_caption", + "bbox": [ + 0.125, + 0.175, + 0.843, + 0.189 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with conventional \\( \\left( {{3}^{\\text{rd }} - {11}^{\\text{th }}\\text{columns}}\\right) \\) and diffusion model baselines \\( \\left( {{12}^{\\text{th }} - {14}^{\\text{th }}\\text{columns}}\\right) \\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.202, + 0.47, + 0.234 + ], + "angle": 0, + "content": "preserves content information from \\( z_T^c \\) while aligning the channel-wise mean and standard deviation with \\( z_T^s \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.246, + 0.21, + 0.263 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.271, + 0.284, + 0.287 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.294, + 0.471, + 0.37 + ], + "angle": 0, + "content": "We conduct all experiments in Stable Diffusion 1.4 pretrained on LAION dataset [36] and adopt DDIM sampling [39] with a total 50 timesteps (\\(t = \\{1, \\dots, 50\\}\\)). For default settings for hyperparameters, we use \\(\\gamma = 0.75\\) and \\(\\tau = 1.5\\), if they are not mentioned separately." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.38, + 0.267, + 0.394 + ], + "angle": 0, + "content": "5.2. Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.587 + ], + "angle": 0, + "content": "Conventional style transfer methods typically utilize Style Loss [11] as both training objective and evaluation metric, so their results tend to overfit the Style Loss. Thus, for a fair comparison, we employ a recently proposed metric, ArtFID [46] which evaluates overall style transfer performances with consideration of both content and style preservation and also is known as strongly coinciding with human judgment. Specifically, ArtFID is computed as \\((\\text{ArtFID} = (1 + \\text{LPIPS}) \\cdot (1 + \\text{FID}))\\). LPIPS [50] measures content fidelity between the stylized image and the corresponding content image, and FID [15] assesses the style fidelity between the stylized image and the corresponding style image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.47, + 0.683 + ], + "angle": 0, + "content": "Dataset. Our evaluations employ content images from MS-COCO [27] dataset and style images from WikiArt [40] dataset. All input images are center-cropped to \\(512 \\times 512\\) resolution. Also, for quantitative comparison, we randomly selected 20 content and 40 style images from each dataset, yielding 800 stylized images as StyTR² [9] has done." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Content Feature Structural Distance (CFSD). In the style transfer evaluation, the assessment of content fidelity often relies on the LPIPS distance. However, since LPIPS utilizes the feature space of AlexNet [20] pre-trained for classification task on ImageNet [7], which is known as texture-biased [12]. Thus, the style information of the images can affect the LPIPS score. To mitigate this style influence, we additionally introduce Content Feature Structural Distance (CFSD) which is a distance measure that only considers the spatial correlation between image patches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In detail, we first define the correlation map between image patch features as follows. For a given image \\(I\\), we obtain feature maps \\(F \\in \\mathbb{R}^{hw \\times c}\\), which is the output feature of conv3 in VGG19 [38]. Then, we calculate the patch sim" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.892, + 0.322 + ], + "angle": 0, + "content": "The similarity map \\( M = F \\times F^T \\), \\( M \\in \\mathbb{R}^{hw \\times hw} \\), which is a similarity map between every pair of features in \\( F \\). After that, for computing the distance between two patch similarity maps, we model the similarity between a single patch and the others as a probability distribution by applying softmax operation. Finally, the correlation map is represented as \\( S = [\\text{softmax}(M_i)]_{i=1}^{hw} \\), \\( S \\in \\mathbb{R}^{hw \\times hw} \\), where \\( M_i \\in \\mathbb{R}^{1 \\times hw} \\) is a similarity map between \\( i^{\\text{th}} \\) patch and the other patches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.892, + 0.383 + ], + "angle": 0, + "content": "Then, CFSD is defined as KL-divergence between two correlation maps. In our case, we compute CFSD between the correlation map of the content \\((S^c)\\) and stylized images \\((S^{cs})\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.394, + 0.892, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C F S D} = \\frac {1}{h w} \\sum_ {i = 1} ^ {h w} D _ {\\mathrm {K L}} \\left(S _ {i} ^ {c} | | S _ {i} ^ {c s}\\right), \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.455, + 0.733, + 0.472 + ], + "angle": 0, + "content": "5.3.Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.892, + 0.616 + ], + "angle": 0, + "content": "We evaluate our proposed method through comparison with twelve state-of-the-art methods, including nine conventional style transfer methods (AesPA-Net [16], CAST [52], StyTR\\(^2\\) [9], EFDM [51], MAST [8], AdaAttN [28], ArtFlow [1], AdaConv [5], AdaIN [17]) and three diffusion-based style transfer methods (DiffuseIT [22], InST [53], DiffStyle [18]), which have a style image as input. We employ the publicly available implementations of all baselines, using their recommended configurations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.892, + 0.772 + ], + "angle": 0, + "content": "Comparison with Conventional Style Transfer. As shown in Tab. 1, our method largely surpasses the conventional style transfer methods in terms of ArtFID, which is known as coinciding the human preference. In addition, the proposed method records the lowest FID, which denotes that stylized images highly resemble the target styles. For content fidelity metrics, ours shows superior scores in both CFSD and LPIPS. We point out that ours achieves much lower CFSD compared to other methods, which is the metric to only consider the spatial correlation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.773, + 0.892, + 0.85 + ], + "angle": 0, + "content": "In addition, we also emphasize that the proposed method can arbitrarily adjust the degree of style transfer by changing the \\(\\gamma\\), and the proposed method significantly surpasses all the other methods in terms of FID (style), when we match the value of LPIPS (content) (Fig. 10)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Comparison with Diffusion-based Style Transfer. Our method demonstrates the best performance in terms of LPIPS, FID, and their combination (ArtFID) with a large" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8800" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.084, + 0.088, + 0.887, + 0.102 + ], + "angle": 0, + "content": "Content Style Ours AesPA-Net CAST StyTR2 EFDM MAST AdaAttN AdaIN DiffuseIT InST DiffStyle" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.102, + 0.891, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.297, + 0.842, + 0.311 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison with conventional \\((4^{\\mathrm{th}} - 10^{\\mathrm{th}}\\) columns) and diffusion model baselines \\((11^{\\mathrm{th}} - 13^{\\mathrm{th}}\\) columns)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.319, + 0.452, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.451, + 0.468, + 0.491 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison with best ArtFID (AdaAttN) and most recently proposed baselines (AesPA-Net) with additional zooming details" + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.499, + 0.411, + 0.539 + ], + "angle": 0, + "content": "
MetricDiffuseITInSTDiffStyleOurs
Time (sec)792.8816.9355.912.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.543, + 0.468, + 0.571 + ], + "angle": 0, + "content": "Table 2. Comparison of inference time of diffusion-based methods for style transferring a given style and content pair" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.468, + 0.791 + ], + "angle": 0, + "content": "margin, as shown in Tab. 1. One significant factor for diffusion models is their running times, since they require several steps to synthesize a single image and it requires inevitable time cost. Hence, we measure the inference time for a pair of content and style images on a single Titan RTX GPU, as shown in Tab. 2. Our method requires a total of 12.4 seconds, with 8.2 seconds for DDIM inversions and 4.2 seconds for sampling costs. As reported, we validate the proposed method significantly faster than other methods, even exploiting large-scale DM. This faster speed comes from the fact that the proposed methods can use the much smaller steps of DDIM inversion, because we additionally utilize the features collected during inversion steps, largely reducing the necessity for perfect inversion of content and style." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.299, + 0.818 + ], + "angle": 0, + "content": "5.4. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.468, + 0.899 + ], + "angle": 0, + "content": "Comparison with Conventional Style Transfer. As shown in Fig. 6, we observe that our method tends to highly preserve the structural information of the content image, while also transferring the style well. For instance, as shown in the third row, ours retains the structure of the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.321, + 0.892, + 0.366 + ], + "angle": 0, + "content": "bridge, but the baselines struggle to preserve structure or transfer the style. We also provide the qualitative comparison with zooming details in Fig. 7 and Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.373, + 0.892, + 0.54 + ], + "angle": 0, + "content": "Comparison with Diffusion-based Style Transfer. We also compare our method with recent diffusion-based style transfer baselines [18, 23, 53]. As shown in Fig. 6, we observe the proposed technique transfers the style to the content well. On the other hand, baselines often lose the structure of content or fail to transfer the style, when an arbitrary content style pair is given. For instance, DiffuseIT and DiffStyle suffer from generating shape and visually plausible images or drop the original content. Differently, InST synthesizes the realistic images, while struggling to transfer style (1\\(^{\\text{st}}\\) row) or change content of image (2\\(^{\\text{nd}}\\), 3\\(^{\\text{rd}}\\) rows)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.552, + 0.653, + 0.569 + ], + "angle": 0, + "content": "5.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.818 + ], + "angle": 0, + "content": "To validate the effectiveness of the proposed components, we conduct ablation studies in both quantitative and qualitative ways. As shown in Fig. 8 and Tab. 3, style injection is significant for guiding the style and content of given images (Config. B). Besides, initial latent AdaIN has a large portion of transferring the color tone of style (Config. D). Attention temperature scaling is in charge of enhancement of quality in synthesized results such as sharpening details and resolving blurriness. For instance, this scaling jointly reduces the FID and LPIPS (Config. \\(\\mathrm{A}^*\\) vs. C in Tab. 3). For more detailed analysis, we provide quantitative metrics with the style-content trade-off, while changing the attention scaling parameter \\(\\tau\\) in Fig. 10 (b). As reported, attention scaling effectively reduces both FID and LPIPS, proving its effects on the preservation of content and capability of style transfer (\\(\\tau = 1.0\\) vs. \\(\\tau = 1.5\\))." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.688, + 0.847 + ], + "angle": 0, + "content": "5.6. Additional Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Content-Style Trade-Off. Our proposed method offers flexible control of the trade-off relation between content and style fidelity by adjusting the parameter \\(\\gamma\\), as discussed in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8801" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.477, + 0.196 + ], + "angle": 0, + "content": "
ConfigurationArtFIDFIDLPIPS
AOurs (γ = 0.75, default)28.8018.130.505
A*Ours (γ = 0.6)27.9717.210.535
B- Style Inject. (Sec. 4.1)43.7227.130.554
C- Attn. Scaling (Sec. 4.2)29.0217.810.542
D- Init. AdaIN (Sec. 4.3)29.2620.050.390
" + }, + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.201, + 0.42, + 0.215 + ], + "angle": 0, + "content": "Table 3. Ablation study on proposed components" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.223, + 0.46, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.369, + 0.436, + 0.383 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison with ablation studies" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.387, + 0.462, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.554, + 0.458, + 0.57 + ], + "angle": 0, + "content": "Figure 9. Visualization of effects of query preservation ratio \\(\\gamma\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.47, + 0.708 + ], + "angle": 0, + "content": "Sec. 4.1. In detail, we compute the FID and LPIPS while varying \\(\\gamma\\) within the range [0.3, 1] with a step size of 0.1. As shown in Fig. 10 (a), our method surpasses baseline methods across all ranges of content and style fidelity. This result implies the proposed method significantly outperforms the other methods, when we match style or content metric to the compared model by adjusting the \\(\\gamma\\) of ours. Note that, dotted lines refer to our model reported in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.71, + 0.47, + 0.803 + ], + "angle": 0, + "content": "We also visualize the effects of the style-content trade-off by synthesizing images by adjusting \\(\\gamma\\). As shown in Fig. 9, the lower \\(\\gamma\\) highly reflects the style while losing the content of the given image, and vice versa. This characteristic of the proposed method suggests that the users can adjust the degree of style by following their personal preferences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Study on the value of \\(\\tau\\). We observe that the gradual increase of \\(\\tau\\) enhances the performance of style transfer, although its effects on enhancement become smaller as \\(\\tau\\) goes larger, as shown in Fig. 10 (b). This result implies that the attention temperature scaling effectively works with a simple modification of the magnitude of the attention map." + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.093, + 0.738, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.093, + 0.885, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.249, + 0.8, + 0.263 + ], + "angle": 0, + "content": "Figure 10. Style-content trade-offs" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.269, + 0.88, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.362, + 0.884, + 0.375 + ], + "angle": 0, + "content": "Figure 11. Comparison with text-guided style transfer methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Comparison with text-guided style transfer. We additionally compare the proposed method with the style transfer methods [21, 41] which are conditioned on the textual inputs. As text-guided methods tend to modify the style largely, we use \\(\\gamma = 0.3\\) for this experiment. Since the text condition hardly contains all the information in the style image such as texture and color tones, the transferred results less resemble the target style, as shown in Fig. 11. Differently, we validate that the proposed method successfully transfers the style with high fidelity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.546, + 0.619, + 0.561 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.814 + ], + "angle": 0, + "content": "Our work addresses the challenges associated with diffusion model-based style transfer methods, which often require time-consuming optimization steps or struggle to leverage the generative potential of large-scale diffusion models. To this end, we propose the method of adapting the pre-trained large-scale diffusion model on style transfer in a training-free way. Our method focuses on manipulating the features of self-attention layers, akin to the cross-attention mechanism, by substituting the key and value during the content generation process with those of the style. Furthermore, we propose the query preservation and attention temperature scaling to mitigate the issue of disruption of content, and initial latent AdaIN to handle the disharmonious color (failure to transfer the colors of style). Experimental results show the superiority of our proposed method over state-of-the-art techniques in previous baselines." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.823, + 0.66, + 0.839 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported in part by MSIT/IITP (No. 2022-0-00680, 2019-0-00421, 2020-0-01821, 2021-0-02068), and MSIT&KNPA/KIPoT (Police Lab 2.0, No. 210121M06)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8802" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Jie An, Siyu Huang, Yibing Song, Dejing Dou, Wei Liu, and Jiebo Luo. Artflow: Unbiased image style transfer via reversible neural flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 862-871, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.472, + 0.324 + ], + "angle": 0, + "content": "[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22560-22570, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.472, + 0.382 + ], + "angle": 0, + "content": "[4] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23040-23050, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.472, + 0.452 + ], + "angle": 0, + "content": "[5] Prashanth Chandran, Gaspard Zoss, Paulo Gotardo, Markus Gross, and Derek Bradley. Adaptive convolutions for structure-aware style transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7972-7981, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.472, + 0.52 + ], + "angle": 0, + "content": "[6] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In ICLR 2023 (Eleventh International Conference on Learning Representations), 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.472, + 0.579 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.472, + 0.647 + ], + "angle": 0, + "content": "[8] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM international conference on multimedia, pages 2719-2727, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.472, + 0.718 + ], + "angle": 0, + "content": "[9] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11326-11336, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.775 + ], + "angle": 0, + "content": "[10] Martin Nicolas Everaert, Marco Bocchio, Sami Arpa, Sabine Susstrunk, and Radhakrishna Achanta. Diffusion in style. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2251-2261, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[13] Nocholas Guttenberg. Diffusion with offset noise. https://www.crosslabs.org/blog/diffusion-with-offset-noise, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[16] Kibeom Hong, Seogkyu Jeon, Junsoo Lee, Namhyuk Ahn, Kunhee Kim, Pilhyeon Lee, Daesik Kim, Youngjung Uh, and Hyeran Byun. Aespa-net: Aesthetic pattern-aware style transfer networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22758-22767, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[17] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[18] Jaeseok Jeong, Mingi Kwon, and Youngjung Uh. Training-free style transfer emerges from h-space in diffusion models. arXiv preprint arXiv:2303.15403, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[19] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Conference on Computer Vision and Pattern Recognition 2023, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[20] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.894, + 0.645 + ], + "angle": 0, + "content": "[21] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[22] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. arXiv preprint arXiv:2209.15264, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[23] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[24] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[25] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[26] Yijun Li, Ming-Yu Liu, Xueting Li, Ming-Hsuan Yang, and Jan Kautz. A closed-form solution to photorealistic image" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8803" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.121 + ], + "angle": 0, + "content": "stylization. In Proceedings of the European conference on computer vision (ECCV), pages 453-468, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.469, + 0.207 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.469, + 0.279 + ], + "angle": 0, + "content": "[28] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.282, + 0.469, + 0.338 + ], + "angle": 0, + "content": "[29] Ming Lu, Hao Zhao, Anbang Yao, Yurong Chen, Feng Xu, and Li Zhang. A closed-form solution to universal style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5952-5961, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.34, + 0.469, + 0.409 + ], + "angle": 0, + "content": "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.412, + 0.47, + 0.468 + ], + "angle": 0, + "content": "[31] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880–5888, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.469, + 0.513 + ], + "angle": 0, + "content": "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.515, + 0.47, + 0.585 + ], + "angle": 0, + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.587, + 0.469, + 0.669 + ], + "angle": 0, + "content": "[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[38] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[40] Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394-409, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[41] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[42] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/ diffusers, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[43] Huan Wang, Yijun Li, Yuehai Wang, Haoji Hu, and Ming-Hsuan Yang. Collaborative distillation for ultra-resolution universal style transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1860-1869, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[44] Zhizhong Wang, Lei Zhao, Haibo Chen, Lihong Qiu, Qihang Mo, Sihuan Lin, Wei Xing, and Dongming Lu. Diversified arbitrary style transfer via deep feature perturbation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7789-7798, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[45] Zhizhong Wang, Lei Zhao, and Wei Xing. *Style diffusions: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[46] Matthias Wright and Björn Ommer. Artfid: Quantitative evaluation of neural style transfer. In DAGM German Conference on Pattern Recognition, pages 560-576. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[47] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. arXiv preprint arXiv:2303.08622, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[48] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[49] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. arXiv preprint arXiv:2305.15347, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[51] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8804" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8035-8045, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[52] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[53] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10146-10156, 2023. 1, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "8805" + } + ] +] \ No newline at end of file diff --git a/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/43978830-73bd-4db5-8b8c-57888ce93c6e_origin.pdf b/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/43978830-73bd-4db5-8b8c-57888ce93c6e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a5ef2c9c4bb995490412fcf9ec4207b5ae04703c --- /dev/null +++ b/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/43978830-73bd-4db5-8b8c-57888ce93c6e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d41ef29d6d3fd5ebfa0e6f37a8297ed18bec92b91b1c7a7846ed5433fb5681da +size 7797603 diff --git a/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/full.md b/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2ae223825788da33530b73984d19e22bd651b140 --- /dev/null +++ b/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/full.md @@ -0,0 +1,326 @@ +# Style Injection in Diffusion: A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer + +Jiwoo Chung*, Sangeek Hyun*, Jae-Pil Heo† Sungkyunkwan University + +{wldn0202, hsi1032, jaepilheo}@g.skku.edu + +# Abstract + +Despite the impressive generative capabilities of diffusion models, existing diffusion model-based style transfer methods require inference-stage optimization (e.g. fine-tuning or textual inversion of style) which is time-consuming, or fails to leverage the generative ability of large-scale diffusion models. To address these issues, we introduce a novel artistic style transfer method based on a pre-trained large-scale diffusion model without any optimization. Specifically, we manipulate the features of self-attention layers as the way the cross-attention mechanism works; in the generation process, substituting the key and value of content with those of style image. This approach provides several desirable characteristics for style transfer including 1) preservation of content by transferring similar styles into similar image patches and 2) transfer of style based on similarity of local texture (e.g. edge) between content and style images. Furthermore, we introduce query preservation and attention temperature scaling to mitigate the issue of disruption of original content, and initial latent Adaptive Instance Normalization (AdaIN) to deal with the disharmonious color (failure to transfer the colors of style). Our experimental results demonstrate that our proposed method surpasses state-of-the-art methods in both conventional and diffusion-based style transfer baselines. Codes are available at github.com/jiwoogit/StyleID. + +# 1. Introduction + +Recent advances in Diffusion Models (DMs) have led to breakthroughs in various generative applications such as text-to-image synthesis [30, 33, 35] and image or video editing [2, 4, 6, 14, 19, 41, 48]. One of these efforts is also applied to the task of style transfer [10, 18, 45, 47, 53]; given style and content images, modifying the style of the content image to possess the given style. + +![](images/64e6849128fd581115042f13f5ea0b408d3614d988eeffcf75b7b26502729014.jpg) +(a) General usage of self-attention in DM + +![](images/8ede24d878fbeafce1263d42f2ba33274a2148939c7c478f43bbeff71c501e9c.jpg) +(b) Our adaptation of self-attention for style transfer +Figure 1. Manipulation of self-attention features for style transfer. (a) General self-attention (SA) deploys the query, key, and value features from a single image in both the training and inference phases. (b) At inference phase, we suggest that manipulating features of self-attention of pre-trained large-scale DM is an effective way to transfer the styles; injection of key and value of styles into SA of contents is a proper way for transferring styles. As a result, style-injected content $z_{t - 1}^{c}$ would maintain contents while modifying its style to resemble the target style. + +General approaches for diffusion model-based style transfer leverage the generative capability of pre-trained DM. Some of these works focus on explicit disentangling style and content for interpretable and controllable style transfer [45], or inversion of the style image into the textual latent space of a large-scale text-to-image DM [53]. However, these methods additionally require gradient-based optimization for fine-tuning and textual inversion [34] for each style image, which is time-consuming. Without this issue, DiffStyle [18] introduces training-free style transfer, but they are known to be hardly applicable to Latent Diffusion Model [33] which is widely adopted for training large-scale text-to-image DM such as Stable Diffusion [33], hindering the users from taking advantage of the prominent generative ability of large-scale models. + +![](images/a5ef20a21d0b05940f620e5cd1ca924f81d6dc810fb5c5f55b2724dc2eda0ae7.jpg) +(b) query-key similarity between patches of similar texture +Figure 2. Desirable attributes of self-attention (SA) for style transfer. (a) Visualization of query by PCA shows that query features well-reflect similarities among patches. That is, style transfer employing SA can preserve the original content, as content patches with similarities tend to receive similar attention scores from a corresponding style image patch. (b) We visualize a similarity map between the blue box (edge) query of the content image, and key of the style image. Thanks to the features representation of large-scale DM encompassing texture and semantics, a query exhibits higher similarity to keys that share a similar style, such as edges. + +In this paper, we focus on extending the training-free style transfer to its application on large-scale pre-trained DM. We start from the observation of recent advances in image-to-image translation based on large-scale DM; they uncover the image editing capability of attention layers. Notably, Plug-and-play [41] shows that the residual block and the attention map of self-attention (SA) determine the spatial layout of generated images. Also, Prompt-to-Prompt [14] locally edits the image by replacing key and value of cross-attention (CA) obtained from text prompt, while keeping their original attention maps. That is, all these works suggest that 1) attention maps determine the spatial layout and 2) key and value of CA adjust the content to fill. + +Inspired by the aforementioned methods, we newly argue that manipulating the SA layer is an effective way to transfer the styles (Fig. 1). Specifically, similar to CA, we substitute the key and value of SA and observe that the generated images are still visually plausible and naturally incorporate the elements of the substituted image into the original image. This observation motivates us to propose a style transfer technique based on self-attention, which combines the styles (textures) of a specific image with the content (semantics and spatial layout) of different images. Furthermore, we highlight that SA layer has desirable characteristics for style transfer. First, as shown in Fig. 2 (a), in SA-based style transfer, the content image patches (query) that share semantic similarities engage with a similar style (key), thereby maintaining the relationship among these content image patches after the transfer. Next, thanks to the powerful feature representation of large-scale DM [49], each patch of the query reveals higher similarity to keys which has similar texture and semantics. For instance, in Fig. 2 (b), + +we can observe that the query feature of content within the blue box exhibits a high similarity to the key features of style with similar edge texture. This encourages the model to transfer style based on the similarity of local texture (e.g. edge) between content and style. + +As a result, our method aims to transfer the textures of the style image to the content images by manipulating self-attention features of pre-trained large-scale DM without any optimization. To this end, we first propose an attention-based style injection method. The basic idea of it is substituting the content's key and value of SA with those of the style image, especially layers in the latter part of decoder which are relevant to the local textures. As mentioned in above paragraph, exchanged styles are well aligned with the content and texture of original image, exploiting the similarity-based attention mechanism. With the proposed style injection, we observe that the local texture patterns are successfully transferred, but there still are remaining problems such as disruption of original content and disharmonious colors. To handle these problems, we additionally propose the following techniques; query preservation, attention temperature scaling, and initial latent AdaIN. Query preservation makes the reverse diffusion process to retain the spatial structure of original content by preserving the query of the content image in the SA. Attention temperature scaling also aims to keep the structure of content by dealing with the blurred self-attention map introduced from the substitution of key. Lastly, initial latent AdaIN corrects inharmonious color problem, referring that the color distribution of style images is not properly transferred, by modulating the statistics of initial noise in the diffusion model. + +Our main contributions are summarized as follows: + +- We propose a style transfer method exploiting the large-scale pre-trained DM by simple manipulation of the features in self-attention; substituting key and value of content with those of styles without any requirements of optimization or supervision (e.g. text). +- We further improve the naive approach for style transfer to properly adapt the styles by proposing three components; query preservation, attention temperature scaling, and initial latent AdaIN. +- Extensive experiments on the style transfer dataset validate the proposed method significantly outperforms previous methods and achieves state-of-the-art performance. + +# 2. Related Work + +# 2.1. Diffusion Model-based Neural Style Transfer + +Neural style transfer [11, 24-26, 29, 31, 43, 44, 52] is an example-guided image generation task that transfers the style of one image onto another while retaining the content of the original. In the realm of diffusion models, neural style transfer has evolved by leveraging the generative capability + +of pre-trained diffusion models. For instance, InST [53] introduced a textual inversion-based approach, aiming to map a given style into corresponding textual embeddings. StyleDiffusion [45] aimed to disentangle style and content by introducing CLIP-based style disentanglement loss for fine-tuning DM for style transfer. Also, several approaches utilize the text input as a style condition or for determining the content to synthesize [10, 47]. + +Conversely, DiffStyle [18] proposed a training-free style transfer method that leverages $h$ -space [23] and adjusts skip connections for effectively conveying style and content information, respectively. However, when DiffStyle is applied to Stable Diffusion [33, 42], their behavior is quite different from typical style-transfer methods; not only textures but also semantics such as spatial layout are also changed. + +To address these limitations, we propose a novel algorithm that harmoniously merges style and content features within the self-attention layers of Stable Diffusion without any optimization process. + +# 2.2. Attention-based Image Editing in DM + +Following the remarkable advances achieved by pre-trained text-to-image DMs [32, 42], there have been numerous image editing works [2, 6, 19, 37] utilizing these DMs. Notably, Prompt-to-Prompt [14] proposed text-based local image editing by manipulating the cross-attention map. Specifically, they observe that cross-attention largely contributes to modeling the relation between the spatial layout of the image to each word in the prompt. Hence, they substitute the original words and cross-attention map with desirable ones, obtaining edited images matched with text conditions. Subsequently, Plug-and-play [41] introduces text-guided image-to-image translation method. They found that the spatial features (i.e. feature from residual block) and self-attention map determine the spatial layout of the synthesized image. Thus, while generating a new image with the given text condition, they guide the diffusion model with features and attention map from the original image for preserving the original spatial layout. Recently, MasaCtrl [3] proposes mutual self-attention control for consistent image editing using text prompts. In detail, they retain the source image's key and value of the self-attention layers, while conditioning the model with desired text prompts. + +Along with these works, we recognize the potential of attention maps in representing spatial information. However, different from the aforementioned methods concentrating on exploiting textual condition, we focus on conditioning by style and content images composed of two images from distinct styles. By combining the features in self-attention layers of both style and content images with precise adjustment of statistics in intermediate representations, we transfer the texture of the content image to the given style. + +# 3. Background + +Latent Diffusion Model (LDM) [33] is a type of diffusion model trained in the low dimensional latent space to focus on semantic bits of data and reduce computation costs. Given an image $x \in \mathbb{R}^{H \times W \times 3}$ , the encoder $\mathcal{E}$ encodes $x$ into the latent representation $z \in \mathbb{R}^{h \times w \times c}$ and decoder reconstructs the image from the latent. + +With the pretrained encoder, they encode the entire images in the dataset and train a diffusion model on latent space $z$ , by predicting noise $\epsilon$ from the noised version of latent $z_{t}$ at time step $t$ . The corresponding training objective is + +$$ +L _ {\mathrm {L D M}} = \mathbb {E} _ {z, \epsilon , t} [ \| \epsilon - \epsilon_ {\theta} (z _ {t}, t, y) \| _ {2} ^ {2} ], \tag {1} +$$ + +where $\epsilon \in \mathcal{N}(0,1)$ is a noise, $t$ is the number of time steps which uniformly sampled from $\{1,\dots,T\}$ , $y$ is a condition, and $\epsilon_{\theta}$ is a neural network which predicts the noise added to the $z$ . + +In our work, we utilize Stable Diffusion (SD) [33] which is the only publicized large-scale pre-trained DM. In the case of SD, $y$ is a text, and $\epsilon_{\theta}$ is a U-Net architecture in which a block for each resolution comprises a residual block, self-attention block (SA), and cross-attention block (CA), sequentially. Among these modules, we focus on the SA block to transfer the styles, as discussed in Sec. 1. Given a feature $\phi$ after the residual block, the self-attention block performs as follows: + +$$ +Q = W _ {Q} (\phi), K = W _ {K} (\phi), V = W _ {V} (\phi), +$$ + +$$ +\phi_ {\text {o u t}} = \operatorname {A t t n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}}\right) \cdot V, \tag {2} +$$ + +where $d$ denotes the dimension of the projected query, and $W_{(\cdot)}$ is a projection layer. Note that, we don't use any text conditions, so the variable $y$ is always an empty text prompt $(\ldots)$ . + +# 4. Method + +In this paper, we aim to solve artistic style transfer by leveraging the generative capability of a pre-trained large-scale text-to-image diffusion model. Briefly, artistic style transfer is the task of modifying the style of a given content image $I^{c}$ to that of style image $I^{s}$ . Then, the stylized image $I^{cs}$ would maintain the semantic content of $I^{c}$ while its style (such as texture) is transferred from $I^{s}$ . For simplicity, we skip the explanations about the encoding and decoding process of the autoencoder in the LDM. Instead, we focus on elaborating the proposed method in the aspect of the diffusion process. Thus, in the following sections, we regard the content, style, and stylized images same as their encoded counterparts $z_{0}^{c}, z_{0}^{s}$ , and $z_{0}^{cs}$ . + +![](images/d557a0ed5ba8b09332a9a9cb85c7f7e3dad6ad5ddc3172013a934b5322b0b4d1.jpg) +Figure 3. Overall framework. (Left) Illustration for the proposed style transfer method. We first invert content image $z_0^c$ and style image $z_0^s$ into the latent noise space as $z_T^c$ and $z_T^s$ , respectively. Then, we initialize the initial noise of stylized image $z_T^{cs}$ from initial latent AdaIN (Sec. 4.3) which combines the content and style noise, $z_T^c$ and $z_T^s$ . While performing the reverse diffusion process with $z_T^{cs}$ , we inject the information of content and style by attention-based style injection (Sec. 4.1) and attention temperature scaling (Sec. 4.2). (Right) Detailed explanation of style injection and initial noise AdaIN. Style injection is basically the manipulation of self-attention (SA) layer during the reverse diffusion process. Specifically, at time step $t$ , we substitute the key $(K_t^{cs})$ and value $(V_t^{cs})$ in SA of stylized image with those of style features, $K_t^s$ and $V_t^s$ , from identical timestep $t$ . At the same time, we preserve the content information by blending the query of content $Q_t^c$ and query of stylized image $Q_t^{cs}$ . Finally, we scale the magnitude of the attention map to deal with the magnitude decrease that the substitution of feature leads to. Initial latent AdaIN produces the initial noise $z_T^{cs}$ by combining style noise $z_T^s$ and content noise $z_T^s$ . Specifically, we modify the channel statistics of $z_T^c$ to resemble the statistics of $z_T^s$ and regard it as $z_T^{cs}$ . We observe this operation enables us to keep the spatial layout of content image while well-reflecting the color tones of a given style image. + +![](images/25341082c1bb4fa13bb877698f1b3db935c09e0b80ac2ef541c5291fe7f10777.jpg) + +# 4.1. Attention-based Style Injection + +We start from the observation in previous image-to-image translation methods, especially Prompt-to-Prompt [14]. The key idea of their method is changing the text condition for cross-attention (CA) while keeping the attention map. Since the attention map affects the spatial layout of output, substituted text conditions determine what to draw in the generated image, and these conditions are actually key and value in CA. Inspired by them, we manipulate the features in self-attention layer as like cross-attention, regarding the features from style image $I^{s}$ as the condition. Specifically, in the generation process, we substitute the key and value of content image with those of style for transferring the texture of style image into the content image. + +To this end, we first obtain the latent for content and style images with DDIM inversion [39], and then collect the SA features of style image over the DDIM inversion process. Specifically, for pre-defined timesteps $t = \{0,\dots,T\}$ , style and content images $z_0^c$ and $z_0^s$ are inverted from image $(t = 0)$ to gaussian noise $(t = T)$ . During DDIM inversion, we also collect query features of content $(Q_t^c)$ and key and value features of style $(K_t^s,V_t^s)$ at every time steps. + +After that, we initialize stylized latent noise $z_T^{cs}$ by copy- + +ing content latent noise $z_{T}^{c}$ . Then, we transfer the target style to the stylized latent by injecting the key $K_{t}^{s}$ and value $V_{t}^{s}$ collected from the style into SA layer, instead of the original key $K_{t}^{cs}$ and value $V_{t}^{cs}$ , when performing the entire reverse process of stylized latent $z_{t}^{cs}$ . However, only applying this substitution can lead to content disruption, since the content of stylized latent would be progressively changed as attended value changes. Hence, we propose query preservation to maintain original content. Simply, we blend query of stylized latent $Q_{t}^{cs}$ and that of content $Q_{t}^{c}$ for the entire reverse process. These style injection and query preservation processes at time step $t$ are expressed as follows: + +$$ +\tilde {Q} _ {t} ^ {c s} = \gamma \times Q _ {t} ^ {c} + (1 - \gamma) \times Q _ {t} ^ {c s}, \tag {3} +$$ + +$$ +\phi_ {\text {o u t}} ^ {\mathrm {c s}} = \operatorname {A t t n} \left(\tilde {Q} _ {t} ^ {c s}, K _ {t} ^ {s}, V _ {t} ^ {s}\right), \tag {4} +$$ + +where $\gamma$ is degree of blending in range of [0, 1]. In addition, we apply these operations on the latter layers of decoder (7- $12^{\mathrm{th}}$ decoder layers in SD) relevant to local textures. We also highlight that the proposed method can adjust the degree of style transfer by changing query preservation ratio $\gamma$ . Specifically, higher $\gamma$ maintains more content, while lower $\gamma$ strengthens effects of style transfer. + +![](images/3dfeb95f4b4a72fd4f26c9fa31b3133f0dc69f21163d9f1bc6c7575f44d06a02.jpg) +(a) Std. of attention map +(b) Ratio of std. of attention map + +![](images/ac32451fdf99d903083809667e55399f7e6415ded3047660913ea5c999dfe958.jpg) +Figure 4. Visualization of the standard deviation of attention map before softmax. (a) Attention-based style injection reduces the standard deviation of self-attention map. Original denotes SA maps from the generation process without style injection. We use both style and content images for generation. (b) We compute the ratio between attention maps w/ and w/o style injection. For the std of original image, we use averaged std. of content and style. + +# 4.2. Attention Temperature Scaling + +Attention map is computed by scaled dot-product between query and key features. During training, query and key features in the SA layer originate from an identical image. However, if we substitute the key features with those of style images, the magnitude of similarity would be overall lowered as style and content are highly likely to be irrelevant. Thus, the computed attention map can be blurred or smoothed, and it would further make output images unsharp, which is detrimental to capturing both content and style information. + +To quantify this issue, we measure the standard deviation of attention map, while ablating the attention-based style injection. In detail, we calculate the attention map before applying softmax, which is scaled-dot product between query and key. As shown in Fig. 4 (a), we validate that this style injection tends to lower the standard deviation of the attention map over the entire timesteps. That is, attention maps after softmax with style injection would be overly smooth. + +To rectify the attention map sharper, we introduce an attention temperature scaling parameter. In detail, we multiply the attention map before softmax by a constant temperature scaling parameter $\tau$ larger than 1. Thus, the attention map after softmax would be sharper than its original values. The modified attention process is represented as follows: + +$$ +\operatorname {A t t n} _ {\tau} \left(\tilde {Q _ {t} ^ {c s}}, K _ {t} ^ {s}, V _ {t} ^ {s}\right) = \operatorname {s o f t m a x} \left(\frac {\tau \tilde {Q _ {t} ^ {c s}} \left(K _ {t} ^ {s}\right) ^ {T}}{\sqrt {d}}\right) \cdot V _ {t} ^ {s}, \tau > 1. \tag {5} +$$ + +We use $\tau = 1.5$ as a default setting, which is the average ratio over entire timesteps. As reported in Fig. 4 (b), we confirm that it effectively calibrates the standard deviation of attention map similar to its original values. + +# 4.3. Initial Latent AdaIN + +In artistic style transfer, the color tone generally takes up a significant portion of the style information. In this context, we observe that the style transfer only with attention-based + +![](images/8cfdad2fb268773913895461a5819b09eacd13147ba70e421af6bead6f86e5ba.jpg) +(a) Generated results only w/ style injection $(\mathrm{K}^s,V^s)$ + +![](images/dde322ca136330611fdd146218e09cf6e5cd3e362c095d788c8b9605543f7f15.jpg) +njection $(\mathrm{K}^s,V^s)$ + +![](images/7c4e3c4344e80896acccab14266518ab7f61d2f69ac574ece9e3eb52bc3f048a.jpg) +(b) Generated results only w/ style injection $(\mathrm{Q}^s,\mathrm{K}^s,V^s)$ +Figure 5. Generated results only w/ style injection. (a) We observe that generated images only with attention-based style injection do not harmonize with the given style in the aspect of color tone. (b) To identify the effects of every feature in SA on color tones, we additionally include query in the style injection process. However, color tones still resemble those of content, concluding features in self-attention have less effect on the color tones. + +![](images/8c8d4174f83169458e3f8dc0035282fc426522048821c800fcd49fd4a8947b3f.jpg) + +style injection often fails in terms of capturing the color tone of the given style. As shown in Fig. 5 (a), textures and local patterns are successfully transferred to the content image while the color tone of the content image still remains. Furthermore, even with injecting the query, key, and value of styles, the resulting images still preserve the color tone of the content, as shown in Fig. 5 (b). + +As substituting the self-attention features has less effects to color tone, we analyze the other vital part of DM; initial latent noise. One of the recent discoveries in DM is that the DM struggles to synthesize purely white or black images [13]. Instead, they tend to generate images of median color as the initial noise is sampled from zero mean and unit variance. Thus, we hypothesize the statistics of initial noise largely affect the colors and brightness of generated images. + +Based on this assumption, we attempt to use the initial latent of style $z_{T}^{s}$ for the style transfer process. However, if we simply start to generate the image from style latent $z_{T}^{s}$ , the structural information of synthesized results also follows the style image and loses the structure of the content. To harness valuable information in both initial latents, we consider that the tone information is intricately connected with the channel statistics of the initial latent, following the principle underlying Style Loss [11] and AdaIN [17]. Thus, we employ AdaIN to modulate the initial latent for effective tone information transfer, represented as: + +$$ +z _ {T} ^ {c s} = \sigma \left(z _ {T} ^ {s}\right) \left(\frac {z _ {T} ^ {c} - \mu \left(z _ {T} ^ {c}\right)}{\sigma \left(z _ {T} ^ {c}\right)}\right) + \mu \left(z _ {T} ^ {s}\right), \tag {6} +$$ + +where $\mu (\cdot),\sigma (\cdot)$ denote channel-wise mean and standard deviation, respectively. Based on this, the initial latent $z_{T}^{cs}$ + +
MetricOursAesPA-NetCASTStyTR2EFDMMASTAdaAttnArtFlowAdaConvAdaINDiffuseITInSTDiffStyle
ArtFID ↓28.80131.42034.68530.72034.60531.28230.35034.63031.85630.93340.72140.63341.464
FID ↓18.13119.76020.39518.89020.06218.19918.65821.25219.02218.24223.06521.57120.903
LPIPS ↓0.50550.51350.62120.54450.64300.62930.54390.55620.59100.60760.69210.80020.8931
CFSD ↓0.22810.24640.29180.30110.33460.30430.28620.29200.36000.31550.34280.67590.2819
+ +Table 1. Quantitative comparison with conventional $\left( {{3}^{\text{rd }} - {11}^{\text{th }}\text{columns}}\right)$ and diffusion model baselines $\left( {{12}^{\text{th }} - {14}^{\text{th }}\text{columns}}\right)$ + +preserves content information from $z_T^c$ while aligning the channel-wise mean and standard deviation with $z_T^s$ . + +# 5. Experiments + +# 5.1. Experimental Settings + +We conduct all experiments in Stable Diffusion 1.4 pretrained on LAION dataset [36] and adopt DDIM sampling [39] with a total 50 timesteps ( $t = \{1, \dots, 50\}$ ). For default settings for hyperparameters, we use $\gamma = 0.75$ and $\tau = 1.5$ , if they are not mentioned separately. + +# 5.2. Evaluation Protocol + +Conventional style transfer methods typically utilize Style Loss [11] as both training objective and evaluation metric, so their results tend to overfit the Style Loss. Thus, for a fair comparison, we employ a recently proposed metric, ArtFID [46] which evaluates overall style transfer performances with consideration of both content and style preservation and also is known as strongly coinciding with human judgment. Specifically, ArtFID is computed as $(\text{ArtFID} = (1 + \text{LPIPS}) \cdot (1 + \text{FID}))$ . LPIPS [50] measures content fidelity between the stylized image and the corresponding content image, and FID [15] assesses the style fidelity between the stylized image and the corresponding style image. + +Dataset. Our evaluations employ content images from MS-COCO [27] dataset and style images from WikiArt [40] dataset. All input images are center-cropped to $512 \times 512$ resolution. Also, for quantitative comparison, we randomly selected 20 content and 40 style images from each dataset, yielding 800 stylized images as StyTR² [9] has done. + +Content Feature Structural Distance (CFSD). In the style transfer evaluation, the assessment of content fidelity often relies on the LPIPS distance. However, since LPIPS utilizes the feature space of AlexNet [20] pre-trained for classification task on ImageNet [7], which is known as texture-biased [12]. Thus, the style information of the images can affect the LPIPS score. To mitigate this style influence, we additionally introduce Content Feature Structural Distance (CFSD) which is a distance measure that only considers the spatial correlation between image patches. + +In detail, we first define the correlation map between image patch features as follows. For a given image $I$ , we obtain feature maps $F \in \mathbb{R}^{hw \times c}$ , which is the output feature of conv3 in VGG19 [38]. Then, we calculate the patch sim + +The similarity map $M = F \times F^T$ , $M \in \mathbb{R}^{hw \times hw}$ , which is a similarity map between every pair of features in $F$ . After that, for computing the distance between two patch similarity maps, we model the similarity between a single patch and the others as a probability distribution by applying softmax operation. Finally, the correlation map is represented as $S = [\text{softmax}(M_i)]_{i=1}^{hw}$ , $S \in \mathbb{R}^{hw \times hw}$ , where $M_i \in \mathbb{R}^{1 \times hw}$ is a similarity map between $i^{\text{th}}$ patch and the other patches. + +Then, CFSD is defined as KL-divergence between two correlation maps. In our case, we compute CFSD between the correlation map of the content $(S^c)$ and stylized images $(S^{cs})$ as follows: + +$$ +\mathrm {C F S D} = \frac {1}{h w} \sum_ {i = 1} ^ {h w} D _ {\mathrm {K L}} \left(S _ {i} ^ {c} | | S _ {i} ^ {c s}\right), \tag {7} +$$ + +# 5.3.Quantitative Comparison + +We evaluate our proposed method through comparison with twelve state-of-the-art methods, including nine conventional style transfer methods (AesPA-Net [16], CAST [52], StyTR $^2$ [9], EFDM [51], MAST [8], AdaAttN [28], ArtFlow [1], AdaConv [5], AdaIN [17]) and three diffusion-based style transfer methods (DiffuseIT [22], InST [53], DiffStyle [18]), which have a style image as input. We employ the publicly available implementations of all baselines, using their recommended configurations. + +Comparison with Conventional Style Transfer. As shown in Tab. 1, our method largely surpasses the conventional style transfer methods in terms of ArtFID, which is known as coinciding the human preference. In addition, the proposed method records the lowest FID, which denotes that stylized images highly resemble the target styles. For content fidelity metrics, ours shows superior scores in both CFSD and LPIPS. We point out that ours achieves much lower CFSD compared to other methods, which is the metric to only consider the spatial correlation. + +In addition, we also emphasize that the proposed method can arbitrarily adjust the degree of style transfer by changing the $\gamma$ , and the proposed method significantly surpasses all the other methods in terms of FID (style), when we match the value of LPIPS (content) (Fig. 10). + +Comparison with Diffusion-based Style Transfer. Our method demonstrates the best performance in terms of LPIPS, FID, and their combination (ArtFID) with a large + +Content Style Ours AesPA-Net CAST StyTR2 EFDM MAST AdaAttN AdaIN DiffuseIT InST DiffStyle + +![](images/860ff3210846250b16d879d171e60f15c4647f18d2abd94bffd6a5026d2fa2ad.jpg) +Figure 6. Qualitative comparison with conventional $(4^{\mathrm{th}} - 10^{\mathrm{th}}$ columns) and diffusion model baselines $(11^{\mathrm{th}} - 13^{\mathrm{th}}$ columns) + +![](images/2838e8ad98d425397cce2a5d39e7ef935a30c31d0c61a70bce8bd7758a67ed21.jpg) +Figure 7. Qualitative comparison with best ArtFID (AdaAttN) and most recently proposed baselines (AesPA-Net) with additional zooming details + +
MetricDiffuseITInSTDiffStyleOurs
Time (sec)792.8816.9355.912.4
+ +Table 2. Comparison of inference time of diffusion-based methods for style transferring a given style and content pair + +margin, as shown in Tab. 1. One significant factor for diffusion models is their running times, since they require several steps to synthesize a single image and it requires inevitable time cost. Hence, we measure the inference time for a pair of content and style images on a single Titan RTX GPU, as shown in Tab. 2. Our method requires a total of 12.4 seconds, with 8.2 seconds for DDIM inversions and 4.2 seconds for sampling costs. As reported, we validate the proposed method significantly faster than other methods, even exploiting large-scale DM. This faster speed comes from the fact that the proposed methods can use the much smaller steps of DDIM inversion, because we additionally utilize the features collected during inversion steps, largely reducing the necessity for perfect inversion of content and style. + +# 5.4. Qualitative Comparison + +Comparison with Conventional Style Transfer. As shown in Fig. 6, we observe that our method tends to highly preserve the structural information of the content image, while also transferring the style well. For instance, as shown in the third row, ours retains the structure of the + +bridge, but the baselines struggle to preserve structure or transfer the style. We also provide the qualitative comparison with zooming details in Fig. 7 and Supplementary. + +Comparison with Diffusion-based Style Transfer. We also compare our method with recent diffusion-based style transfer baselines [18, 23, 53]. As shown in Fig. 6, we observe the proposed technique transfers the style to the content well. On the other hand, baselines often lose the structure of content or fail to transfer the style, when an arbitrary content style pair is given. For instance, DiffuseIT and DiffStyle suffer from generating shape and visually plausible images or drop the original content. Differently, InST synthesizes the realistic images, while struggling to transfer style (1 $^{\text{st}}$ row) or change content of image (2 $^{\text{nd}}$ , 3 $^{\text{rd}}$ rows). + +# 5.5. Ablation Study + +To validate the effectiveness of the proposed components, we conduct ablation studies in both quantitative and qualitative ways. As shown in Fig. 8 and Tab. 3, style injection is significant for guiding the style and content of given images (Config. B). Besides, initial latent AdaIN has a large portion of transferring the color tone of style (Config. D). Attention temperature scaling is in charge of enhancement of quality in synthesized results such as sharpening details and resolving blurriness. For instance, this scaling jointly reduces the FID and LPIPS (Config. $\mathrm{A}^*$ vs. C in Tab. 3). For more detailed analysis, we provide quantitative metrics with the style-content trade-off, while changing the attention scaling parameter $\tau$ in Fig. 10 (b). As reported, attention scaling effectively reduces both FID and LPIPS, proving its effects on the preservation of content and capability of style transfer ( $\tau = 1.0$ vs. $\tau = 1.5$ ). + +# 5.6. Additional Analysis + +Content-Style Trade-Off. Our proposed method offers flexible control of the trade-off relation between content and style fidelity by adjusting the parameter $\gamma$ , as discussed in + +
ConfigurationArtFIDFIDLPIPS
AOurs (γ = 0.75, default)28.8018.130.505
A*Ours (γ = 0.6)27.9717.210.535
B- Style Inject. (Sec. 4.1)43.7227.130.554
C- Attn. Scaling (Sec. 4.2)29.0217.810.542
D- Init. AdaIN (Sec. 4.3)29.2620.050.390
+ +Table 3. Ablation study on proposed components + +![](images/b89ab7b4bcfdd501d4f0c4ece99c1f74adbac5ab3e9aafff8a18fd52f34f13c6.jpg) +Figure 8. Qualitative comparison with ablation studies + +![](images/4cf5a63c61f5638d6fffff5dc9acbbc0b67f136c825e93b8776c146c9759bf5c.jpg) +Figure 9. Visualization of effects of query preservation ratio $\gamma$ + +Sec. 4.1. In detail, we compute the FID and LPIPS while varying $\gamma$ within the range [0.3, 1] with a step size of 0.1. As shown in Fig. 10 (a), our method surpasses baseline methods across all ranges of content and style fidelity. This result implies the proposed method significantly outperforms the other methods, when we match style or content metric to the compared model by adjusting the $\gamma$ of ours. Note that, dotted lines refer to our model reported in Tab. 1. + +We also visualize the effects of the style-content trade-off by synthesizing images by adjusting $\gamma$ . As shown in Fig. 9, the lower $\gamma$ highly reflects the style while losing the content of the given image, and vice versa. This characteristic of the proposed method suggests that the users can adjust the degree of style by following their personal preferences. + +Study on the value of $\tau$ . We observe that the gradual increase of $\tau$ enhances the performance of style transfer, although its effects on enhancement become smaller as $\tau$ goes larger, as shown in Fig. 10 (b). This result implies that the attention temperature scaling effectively works with a simple modification of the magnitude of the attention map. + +![](images/0574ed7e223feb7c0b3309243ecdcabf39ec3c7cba7af2ebe674151a020c830e.jpg) +Figure 10. Style-content trade-offs + +![](images/d13886104303c34f63fe8e4d7a7cc55500466e7e10422acf2cfaad9f948dd77d.jpg) + +![](images/6f76cbcee65f8d474e662ff237a8efaf946791fe47816741a406344a17535b5e.jpg) +Figure 11. Comparison with text-guided style transfer methods + +Comparison with text-guided style transfer. We additionally compare the proposed method with the style transfer methods [21, 41] which are conditioned on the textual inputs. As text-guided methods tend to modify the style largely, we use $\gamma = 0.3$ for this experiment. Since the text condition hardly contains all the information in the style image such as texture and color tones, the transferred results less resemble the target style, as shown in Fig. 11. Differently, we validate that the proposed method successfully transfers the style with high fidelity. + +# 6. Conclusion + +Our work addresses the challenges associated with diffusion model-based style transfer methods, which often require time-consuming optimization steps or struggle to leverage the generative potential of large-scale diffusion models. To this end, we propose the method of adapting the pre-trained large-scale diffusion model on style transfer in a training-free way. Our method focuses on manipulating the features of self-attention layers, akin to the cross-attention mechanism, by substituting the key and value during the content generation process with those of the style. Furthermore, we propose the query preservation and attention temperature scaling to mitigate the issue of disruption of content, and initial latent AdaIN to handle the disharmonious color (failure to transfer the colors of style). Experimental results show the superiority of our proposed method over state-of-the-art techniques in previous baselines. + +# Acknowledgments + +This work was supported in part by MSIT/IITP (No. 2022-0-00680, 2019-0-00421, 2020-0-01821, 2021-0-02068), and MSIT&KNPA/KIPoT (Police Lab 2.0, No. 210121M06). + +# References + +[1] Jie An, Siyu Huang, Yibing Song, Dejing Dou, Wei Liu, and Jiebo Luo. Artflow: Unbiased image style transfer via reversible neural flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 862-871, 2021. 6 +[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 1, 3 +[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22560-22570, 2023. 3 +[4] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23040-23050, 2023. 1 +[5] Prashanth Chandran, Gaspard Zoss, Paulo Gotardo, Markus Gross, and Derek Bradley. Adaptive convolutions for structure-aware style transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7972-7981, 2021. 6 +[6] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In ICLR 2023 (Eleventh International Conference on Learning Representations), 2023. 1, 3 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6 +[8] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM international conference on multimedia, pages 2719-2727, 2020. 6 +[9] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11326-11336, 2022. 6 +[10] Martin Nicolas Everaert, Marco Bocchio, Sami Arpa, Sabine Susstrunk, and Radhakrishna Achanta. Diffusion in style. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2251-2261, 2023. 1, 3 +[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2, 5, 6 +[12] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 6 + +[13] Nocholas Guttenberg. Diffusion with offset noise. https://www.crosslabs.org/blog/diffusion-with-offset-noise, 2023.5 +[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 3, 4 +[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[16] Kibeom Hong, Seogkyu Jeon, Junsoo Lee, Namhyuk Ahn, Kunhee Kim, Pilhyeon Lee, Daesik Kim, Youngjung Uh, and Hyeran Byun. Aespa-net: Aesthetic pattern-aware style transfer networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22758-22767, 2023. 6 +[17] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 5, 6 +[18] Jaeseok Jeong, Mingi Kwon, and Youngjung Uh. Training-free style transfer emerges from h-space in diffusion models. arXiv preprint arXiv:2303.15403, 2023. 1, 3, 6, 7 +[19] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Conference on Computer Vision and Pattern Recognition 2023, 2023. 1, 3 +[20] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 6 +[21] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 8 +[22] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. arXiv preprint arXiv:2209.15264, 2022. 6 +[23] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 3, 7 +[24] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 2 +[25] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. +[26] Yijun Li, Ming-Yu Liu, Xueting Li, Ming-Hsuan Yang, and Jan Kautz. A closed-form solution to photorealistic image + +stylization. In Proceedings of the European conference on computer vision (ECCV), pages 453-468, 2018. 2 +[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[28] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 6 +[29] Ming Lu, Hao Zhao, Anbang Yao, Yurong Chen, Feng Xu, and Li Zhang. A closed-form solution to universal style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5952-5961, 2019. 2 +[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1 +[31] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880–5888, 2019. 2 +[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. 3 +[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3 +[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1 +[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1 +[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 6 +[37] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 3 + +[38] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6 +[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 4, 6 +[40] Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394-409, 2019. 6 +[41] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2, 3, 8 +[42] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/ diffusers, 2022.3 +[43] Huan Wang, Yijun Li, Yuehai Wang, Haoji Hu, and Ming-Hsuan Yang. Collaborative distillation for ultra-resolution universal style transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1860-1869, 2020. 2 +[44] Zhizhong Wang, Lei Zhao, Haibo Chen, Lihong Qiu, Qihang Mo, Sihuan Lin, Wei Xing, and Dongming Lu. Diversified arbitrary style transfer via deep feature perturbation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7789-7798, 2020. 2 +[45] Zhizhong Wang, Lei Zhao, and Wei Xing. *Style diffusions: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 1, 3 +[46] Matthias Wright and Björn Ommer. Artfid: Quantitative evaluation of neural style transfer. In DAGM German Conference on Pattern Recognition, pages 560-576. Springer, 2022. 6 +[47] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. arXiv preprint arXiv:2303.08622, 2023. 1, 3 +[48] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 1 +[49] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. arXiv preprint arXiv:2305.15347, 2023. 2 +[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[51] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8035-8045, 2022. 6 +[52] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022. 2, 6 +[53] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10146-10156, 2023. 1, 3, 6, 7 \ No newline at end of file diff --git a/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/images.zip b/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..45de208649ad8276c1aff8dca62a605b8377628e --- /dev/null +++ b/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adb39747c2d5a9d5ab7721b345788144334ce4e33206e10f4ddea5fad134f18d +size 843834 diff --git a/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/layout.json b/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f31a52486642b9a19fdfb8f60a55a1fcd899bb64 --- /dev/null +++ b/2024/Style Injection in Diffusion_ A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer/layout.json @@ -0,0 +1,9712 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 108, + 103, + 486, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 103, + 486, + 140 + ], + "spans": [ + { + "bbox": [ + 108, + 103, + 486, + 140 + ], + "type": "text", + "content": "Style Injection in Diffusion: A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 187, + 160, + 405, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 160, + 405, + 190 + ], + "spans": [ + { + "bbox": [ + 187, + 160, + 405, + 190 + ], + "type": "text", + "content": "Jiwoo Chung*, Sangeek Hyun*, Jae-Pil Heo† Sungkyunkwan University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 186, + 191, + 407, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 191, + 407, + 202 + ], + "spans": [ + { + "bbox": [ + 186, + 191, + 407, + 202 + ], + "type": "text", + "content": "{wldn0202, hsi1032, jaepilheo}@g.skku.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 256, + 290, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 555 + ], + "type": "text", + "content": "Despite the impressive generative capabilities of diffusion models, existing diffusion model-based style transfer methods require inference-stage optimization (e.g. fine-tuning or textual inversion of style) which is time-consuming, or fails to leverage the generative ability of large-scale diffusion models. To address these issues, we introduce a novel artistic style transfer method based on a pre-trained large-scale diffusion model without any optimization. Specifically, we manipulate the features of self-attention layers as the way the cross-attention mechanism works; in the generation process, substituting the key and value of content with those of style image. This approach provides several desirable characteristics for style transfer including 1) preservation of content by transferring similar styles into similar image patches and 2) transfer of style based on similarity of local texture (e.g. edge) between content and style images. Furthermore, we introduce query preservation and attention temperature scaling to mitigate the issue of disruption of original content, and initial latent Adaptive Instance Normalization (AdaIN) to deal with the disharmonious color (failure to transfer the colors of style). Our experimental results demonstrate that our proposed method surpasses state-of-the-art methods in both conventional and diffusion-based style transfer baselines. Codes are available at github.com/jiwoogit/StyleID." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 582, + 128, + 595 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 128, + 595 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 128, + 595 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 603, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 603, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 603, + 287, + 687 + ], + "type": "text", + "content": "Recent advances in Diffusion Models (DMs) have led to breakthroughs in various generative applications such as text-to-image synthesis [30, 33, 35] and image or video editing [2, 4, 6, 14, 19, 41, 48]. One of these efforts is also applied to the task of style transfer [10, 18, 45, 47, 53]; given style and content images, modifying the style of the content image to possess the given style." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 234, + 544, + 285 + ], + "blocks": [ + { + "bbox": [ + 312, + 234, + 544, + 285 + ], + "lines": [ + { + "bbox": [ + 312, + 234, + 544, + 285 + ], + "spans": [ + { + "bbox": [ + 312, + 234, + 544, + 285 + ], + "type": "image", + "image_path": "64e6849128fd581115042f13f5ea0b408d3614d988eeffcf75b7b26502729014.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 287, + 511, + 297 + ], + "lines": [ + { + "bbox": [ + 350, + 287, + 511, + 297 + ], + "spans": [ + { + "bbox": [ + 350, + 287, + 511, + 297 + ], + "type": "text", + "content": "(a) General usage of self-attention in DM" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 299, + 544, + 403 + ], + "blocks": [ + { + "bbox": [ + 310, + 299, + 544, + 403 + ], + "lines": [ + { + "bbox": [ + 310, + 299, + 544, + 403 + ], + "spans": [ + { + "bbox": [ + 310, + 299, + 544, + 403 + ], + "type": "image", + "image_path": "8ede24d878fbeafce1263d42f2ba33274a2148939c7c478f43bbeff71c501e9c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 406, + 531, + 417 + ], + "lines": [ + { + "bbox": [ + 329, + 406, + 531, + 417 + ], + "spans": [ + { + "bbox": [ + 329, + 406, + 531, + 417 + ], + "type": "text", + "content": "(b) Our adaptation of self-attention for style transfer" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 426, + 547, + 525 + ], + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 525 + ], + "type": "text", + "content": "Figure 1. Manipulation of self-attention features for style transfer. (a) General self-attention (SA) deploys the query, key, and value features from a single image in both the training and inference phases. (b) At inference phase, we suggest that manipulating features of self-attention of pre-trained large-scale DM is an effective way to transfer the styles; injection of key and value of styles into SA of contents is a proper way for transferring styles. As a result, style-injected content " + }, + { + "bbox": [ + 304, + 426, + 547, + 525 + ], + "type": "inline_equation", + "content": "z_{t - 1}^{c}" + }, + { + "bbox": [ + 304, + 426, + 547, + 525 + ], + "type": "text", + "content": " would maintain contents while modifying its style to resemble the target style." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "General approaches for diffusion model-based style transfer leverage the generative capability of pre-trained DM. Some of these works focus on explicit disentangling style and content for interpretable and controllable style transfer [45], or inversion of the style image into the textual latent space of a large-scale text-to-image DM [53]. However, these methods additionally require gradient-based optimization for fine-tuning and textual inversion [34] for each style image, which is time-consuming. Without this issue, DiffStyle [18] introduces training-free style transfer, but they are known to be hardly applicable to Latent Diffusion Model [33] which is widely adopted for training large-scale text-to-image DM such as Stable Diffusion [33], hindering the users from taking advantage of the prominent generative ability of large-scale models." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 693, + 130, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 693, + 130, + 703 + ], + "spans": [ + { + "bbox": [ + 62, + 693, + 130, + 703 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 63, + 703, + 140, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 703, + 140, + 713 + ], + "spans": [ + { + "bbox": [ + 63, + 703, + 140, + 713 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "8795" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 287, + 181 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 287, + 181 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 287, + 181 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 287, + 181 + ], + "type": "image", + "image_path": "a5ef20a21d0b05940f620e5cd1ca924f81d6dc810fb5c5f55b2724dc2eda0ae7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 182, + 253, + 191 + ], + "lines": [ + { + "bbox": [ + 80, + 182, + 253, + 191 + ], + "spans": [ + { + "bbox": [ + 80, + 182, + 253, + 191 + ], + "type": "text", + "content": "(b) query-key similarity between patches of similar texture" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 197, + 287, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 197, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 287, + 307 + ], + "type": "text", + "content": "Figure 2. Desirable attributes of self-attention (SA) for style transfer. (a) Visualization of query by PCA shows that query features well-reflect similarities among patches. That is, style transfer employing SA can preserve the original content, as content patches with similarities tend to receive similar attention scores from a corresponding style image patch. (b) We visualize a similarity map between the blue box (edge) query of the content image, and key of the style image. Thanks to the features representation of large-scale DM encompassing texture and semantics, a query exhibits higher similarity to keys that share a similar style, such as edges." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 317, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 287, + 483 + ], + "type": "text", + "content": "In this paper, we focus on extending the training-free style transfer to its application on large-scale pre-trained DM. We start from the observation of recent advances in image-to-image translation based on large-scale DM; they uncover the image editing capability of attention layers. Notably, Plug-and-play [41] shows that the residual block and the attention map of self-attention (SA) determine the spatial layout of generated images. Also, Prompt-to-Prompt [14] locally edits the image by replacing key and value of cross-attention (CA) obtained from text prompt, while keeping their original attention maps. That is, all these works suggest that 1) attention maps determine the spatial layout and 2) key and value of CA adjust the content to fill." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "type": "text", + "content": "Inspired by the aforementioned methods, we newly argue that manipulating the SA layer is an effective way to transfer the styles (Fig. 1). Specifically, similar to CA, we substitute the key and value of SA and observe that the generated images are still visually plausible and naturally incorporate the elements of the substituted image into the original image. This observation motivates us to propose a style transfer technique based on self-attention, which combines the styles (textures) of a specific image with the content (semantics and spatial layout) of different images. Furthermore, we highlight that SA layer has desirable characteristics for style transfer. First, as shown in Fig. 2 (a), in SA-based style transfer, the content image patches (query) that share semantic similarities engage with a similar style (key), thereby maintaining the relationship among these content image patches after the transfer. Next, thanks to the powerful feature representation of large-scale DM [49], each patch of the query reveals higher similarity to keys which has similar texture and semantics. For instance, in Fig. 2 (b)," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "we can observe that the query feature of content within the blue box exhibits a high similarity to the key features of style with similar edge texture. This encourages the model to transfer style based on the similarity of local texture (e.g. edge) between content and style." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 133, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 443 + ], + "type": "text", + "content": "As a result, our method aims to transfer the textures of the style image to the content images by manipulating self-attention features of pre-trained large-scale DM without any optimization. To this end, we first propose an attention-based style injection method. The basic idea of it is substituting the content's key and value of SA with those of the style image, especially layers in the latter part of decoder which are relevant to the local textures. As mentioned in above paragraph, exchanged styles are well aligned with the content and texture of original image, exploiting the similarity-based attention mechanism. With the proposed style injection, we observe that the local texture patterns are successfully transferred, but there still are remaining problems such as disruption of original content and disharmonious colors. To handle these problems, we additionally propose the following techniques; query preservation, attention temperature scaling, and initial latent AdaIN. Query preservation makes the reverse diffusion process to retain the spatial structure of original content by preserving the query of the content image in the SA. Attention temperature scaling also aims to keep the structure of content by dealing with the blurred self-attention map introduced from the substitution of key. Lastly, initial latent AdaIN corrects inharmonious color problem, referring that the color distribution of style images is not properly transferred, by modulating the statistics of initial noise in the diffusion model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 445, + 526, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 526, + 456 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 526, + 456 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 459, + 545, + 601 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 304, + 459, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 517 + ], + "type": "text", + "content": "- We propose a style transfer method exploiting the large-scale pre-trained DM by simple manipulation of the features in self-attention; substituting key and value of content with those of styles without any requirements of optimization or supervision (e.g. text)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 518, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 564 + ], + "type": "text", + "content": "- We further improve the naive approach for style transfer to properly adapt the styles by proposing three components; query preservation, attention temperature scaling, and initial latent AdaIN." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 566, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 545, + 601 + ], + "type": "text", + "content": "- Extensive experiments on the style transfer dataset validate the proposed method significantly outperforms previous methods and achieves state-of-the-art performance." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 613, + 392, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 613, + 392, + 625 + ], + "spans": [ + { + "bbox": [ + 306, + 613, + 392, + 625 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 634, + 538, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 538, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 538, + 646 + ], + "type": "text", + "content": "2.1. Diffusion Model-based Neural Style Transfer" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Neural style transfer [11, 24-26, 29, 31, 43, 44, 52] is an example-guided image generation task that transfers the style of one image onto another while retaining the content of the original. In the realm of diffusion models, neural style transfer has evolved by leveraging the generative capability" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "8796" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "of pre-trained diffusion models. For instance, InST [53] introduced a textual inversion-based approach, aiming to map a given style into corresponding textual embeddings. StyleDiffusion [45] aimed to disentangle style and content by introducing CLIP-based style disentanglement loss for fine-tuning DM for style transfer. Also, several approaches utilize the text input as a style condition or for determining the content to synthesize [10, 47]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 287, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 287, + 255 + ], + "type": "text", + "content": "Conversely, DiffStyle [18] proposed a training-free style transfer method that leverages " + }, + { + "bbox": [ + 46, + 170, + 287, + 255 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 170, + 287, + 255 + ], + "type": "text", + "content": "-space [23] and adjusts skip connections for effectively conveying style and content information, respectively. However, when DiffStyle is applied to Stable Diffusion [33, 42], their behavior is quite different from typical style-transfer methods; not only textures but also semantics such as spatial layout are also changed." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 256, + 287, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 287, + 305 + ], + "type": "text", + "content": "To address these limitations, we propose a novel algorithm that harmoniously merges style and content features within the self-attention layers of Stable Diffusion without any optimization process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 319, + 247, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 247, + 333 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 247, + 333 + ], + "type": "text", + "content": "2.2. Attention-based Image Editing in DM" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 340, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 604 + ], + "type": "text", + "content": "Following the remarkable advances achieved by pre-trained text-to-image DMs [32, 42], there have been numerous image editing works [2, 6, 19, 37] utilizing these DMs. Notably, Prompt-to-Prompt [14] proposed text-based local image editing by manipulating the cross-attention map. Specifically, they observe that cross-attention largely contributes to modeling the relation between the spatial layout of the image to each word in the prompt. Hence, they substitute the original words and cross-attention map with desirable ones, obtaining edited images matched with text conditions. Subsequently, Plug-and-play [41] introduces text-guided image-to-image translation method. They found that the spatial features (i.e. feature from residual block) and self-attention map determine the spatial layout of the synthesized image. Thus, while generating a new image with the given text condition, they guide the diffusion model with features and attention map from the original image for preserving the original spatial layout. Recently, MasaCtrl [3] proposes mutual self-attention control for consistent image editing using text prompts. In detail, they retain the source image's key and value of the self-attention layers, while conditioning the model with desired text prompts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 606, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 287, + 715 + ], + "type": "text", + "content": "Along with these works, we recognize the potential of attention maps in representing spatial information. However, different from the aforementioned methods concentrating on exploiting textual condition, we focus on conditioning by style and content images composed of two images from distinct styles. By combining the features in self-attention layers of both style and content images with precise adjustment of statistics in intermediate representations, we transfer the texture of the content image to the given style." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 71, + 384, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 384, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 384, + 85 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 92, + 545, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 92, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 304, + 92, + 545, + 164 + ], + "type": "text", + "content": "Latent Diffusion Model (LDM) [33] is a type of diffusion model trained in the low dimensional latent space to focus on semantic bits of data and reduce computation costs. Given an image " + }, + { + "bbox": [ + 304, + 92, + 545, + 164 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 92, + 545, + 164 + ], + "type": "text", + "content": ", the encoder " + }, + { + "bbox": [ + 304, + 92, + 545, + 164 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 92, + 545, + 164 + ], + "type": "text", + "content": " encodes " + }, + { + "bbox": [ + 304, + 92, + 545, + 164 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 92, + 545, + 164 + ], + "type": "text", + "content": " into the latent representation " + }, + { + "bbox": [ + 304, + 92, + 545, + 164 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 304, + 92, + 545, + 164 + ], + "type": "text", + "content": " and decoder reconstructs the image from the latent." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 166, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 166, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 166, + 545, + 224 + ], + "type": "text", + "content": "With the pretrained encoder, they encode the entire images in the dataset and train a diffusion model on latent space " + }, + { + "bbox": [ + 304, + 166, + 545, + 224 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 166, + 545, + 224 + ], + "type": "text", + "content": ", by predicting noise " + }, + { + "bbox": [ + 304, + 166, + 545, + 224 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 166, + 545, + 224 + ], + "type": "text", + "content": " from the noised version of latent " + }, + { + "bbox": [ + 304, + 166, + 545, + 224 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 166, + 545, + 224 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 304, + 166, + 545, + 224 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 166, + 545, + 224 + ], + "type": "text", + "content": ". The corresponding training objective is" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 354, + 229, + 545, + 244 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 229, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 354, + 229, + 545, + 244 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {L D M}} = \\mathbb {E} _ {z, \\epsilon , t} [ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, y) \\| _ {2} ^ {2} ], \\tag {1}", + "image_path": "48434552d9df9f6e40631fe97f2e53a2b0a2f1e00cd9f7adc09ad90e343a4570.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\epsilon \\in \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "text", + "content": " is a noise, " + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "text", + "content": " is the number of time steps which uniformly sampled from " + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\{1,\\dots,T\\}" + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "text", + "content": " is a condition, and " + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "text", + "content": " is a neural network which predicts the noise added to the " + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "text", + "content": "In our work, we utilize Stable Diffusion (SD) [33] which is the only publicized large-scale pre-trained DM. In the case of SD, " + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "text", + "content": " is a text, and " + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "text", + "content": " is a U-Net architecture in which a block for each resolution comprises a residual block, self-attention block (SA), and cross-attention block (CA), sequentially. Among these modules, we focus on the SA block to transfer the styles, as discussed in Sec. 1. Given a feature " + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "text", + "content": " after the residual block, the self-attention block performs as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 422, + 515, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 422, + 515, + 436 + ], + "spans": [ + { + "bbox": [ + 347, + 422, + 515, + 436 + ], + "type": "interline_equation", + "content": "Q = W _ {Q} (\\phi), K = W _ {K} (\\phi), V = W _ {V} (\\phi),", + "image_path": "5e18ec0fc208f09ce82e0ce9b5597fb250b15f9bc5629da1d99b704ee9ae4073.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 324, + 437, + 545, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 437, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 324, + 437, + 545, + 464 + ], + "type": "interline_equation", + "content": "\\phi_ {\\text {o u t}} = \\operatorname {A t t n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) \\cdot V, \\tag {2}", + "image_path": "c44f22579f58de6fa673aa4140dc397f5c19036d6e8208f5f809079e0f24f7fe.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 475, + 545, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 475, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 545, + 524 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 475, + 545, + 524 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 306, + 475, + 545, + 524 + ], + "type": "text", + "content": " denotes the dimension of the projected query, and " + }, + { + "bbox": [ + 306, + 475, + 545, + 524 + ], + "type": "inline_equation", + "content": "W_{(\\cdot)}" + }, + { + "bbox": [ + 306, + 475, + 545, + 524 + ], + "type": "text", + "content": " is a projection layer. Note that, we don't use any text conditions, so the variable " + }, + { + "bbox": [ + 306, + 475, + 545, + 524 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 306, + 475, + 545, + 524 + ], + "type": "text", + "content": " is always an empty text prompt " + }, + { + "bbox": [ + 306, + 475, + 545, + 524 + ], + "type": "inline_equation", + "content": "(\\ldots)" + }, + { + "bbox": [ + 306, + 475, + 545, + 524 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 536, + 361, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 361, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 361, + 548 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 558, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 715 + ], + "type": "text", + "content": "In this paper, we aim to solve artistic style transfer by leveraging the generative capability of a pre-trained large-scale text-to-image diffusion model. Briefly, artistic style transfer is the task of modifying the style of a given content image " + }, + { + "bbox": [ + 304, + 558, + 545, + 715 + ], + "type": "inline_equation", + "content": "I^{c}" + }, + { + "bbox": [ + 304, + 558, + 545, + 715 + ], + "type": "text", + "content": " to that of style image " + }, + { + "bbox": [ + 304, + 558, + 545, + 715 + ], + "type": "inline_equation", + "content": "I^{s}" + }, + { + "bbox": [ + 304, + 558, + 545, + 715 + ], + "type": "text", + "content": ". Then, the stylized image " + }, + { + "bbox": [ + 304, + 558, + 545, + 715 + ], + "type": "inline_equation", + "content": "I^{cs}" + }, + { + "bbox": [ + 304, + 558, + 545, + 715 + ], + "type": "text", + "content": " would maintain the semantic content of " + }, + { + "bbox": [ + 304, + 558, + 545, + 715 + ], + "type": "inline_equation", + "content": "I^{c}" + }, + { + "bbox": [ + 304, + 558, + 545, + 715 + ], + "type": "text", + "content": " while its style (such as texture) is transferred from " + }, + { + "bbox": [ + 304, + 558, + 545, + 715 + ], + "type": "inline_equation", + "content": "I^{s}" + }, + { + "bbox": [ + 304, + 558, + 545, + 715 + ], + "type": "text", + "content": ". For simplicity, we skip the explanations about the encoding and decoding process of the autoencoder in the LDM. Instead, we focus on elaborating the proposed method in the aspect of the diffusion process. Thus, in the following sections, we regard the content, style, and stylized images same as their encoded counterparts " + }, + { + "bbox": [ + 304, + 558, + 545, + 715 + ], + "type": "inline_equation", + "content": "z_{0}^{c}, z_{0}^{s}" + }, + { + "bbox": [ + 304, + 558, + 545, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 558, + 545, + 715 + ], + "type": "inline_equation", + "content": "z_{0}^{cs}" + }, + { + "bbox": [ + 304, + 558, + 545, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8797" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 73, + 355, + 286 + ], + "blocks": [ + { + "bbox": [ + 51, + 73, + 355, + 286 + ], + "lines": [ + { + "bbox": [ + 51, + 73, + 355, + 286 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 355, + 286 + ], + "type": "image", + "image_path": "d557a0ed5ba8b09332a9a9cb85c7f7e3dad6ad5ddc3172013a934b5322b0b4d1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "lines": [ + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": "Figure 3. Overall framework. (Left) Illustration for the proposed style transfer method. We first invert content image " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "z_0^c" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": " and style image " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "z_0^s" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": " into the latent noise space as " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "z_T^c" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "z_T^s" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": ", respectively. Then, we initialize the initial noise of stylized image " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "z_T^{cs}" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": " from initial latent AdaIN (Sec. 4.3) which combines the content and style noise, " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "z_T^c" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "z_T^s" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": ". While performing the reverse diffusion process with " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "z_T^{cs}" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": ", we inject the information of content and style by attention-based style injection (Sec. 4.1) and attention temperature scaling (Sec. 4.2). (Right) Detailed explanation of style injection and initial noise AdaIN. Style injection is basically the manipulation of self-attention (SA) layer during the reverse diffusion process. Specifically, at time step " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": ", we substitute the key " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "(K_t^{cs})" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "(V_t^{cs})" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": " in SA of stylized image with those of style features, " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "K_t^s" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "V_t^s" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": ", from identical timestep " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": ". At the same time, we preserve the content information by blending the query of content " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "Q_t^c" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": " and query of stylized image " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "Q_t^{cs}" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": ". Finally, we scale the magnitude of the attention map to deal with the magnitude decrease that the substitution of feature leads to. Initial latent AdaIN produces the initial noise " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "z_T^{cs}" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": " by combining style noise " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "z_T^s" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": " and content noise " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "z_T^s" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": ". Specifically, we modify the channel statistics of " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "z_T^c" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": " to resemble the statistics of " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "z_T^s" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": " and regard it as " + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "inline_equation", + "content": "z_T^{cs}" + }, + { + "bbox": [ + 46, + 291, + 545, + 413 + ], + "type": "text", + "content": ". We observe this operation enables us to keep the spatial layout of content image while well-reflecting the color tones of a given style image." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 357, + 73, + 541, + 285 + ], + "blocks": [ + { + "bbox": [ + 357, + 73, + 541, + 285 + ], + "lines": [ + { + "bbox": [ + 357, + 73, + 541, + 285 + ], + "spans": [ + { + "bbox": [ + 357, + 73, + 541, + 285 + ], + "type": "image", + "image_path": "25341082c1bb4fa13bb877698f1b3db935c09e0b80ac2ef541c5291fe7f10777.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 422, + 216, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 216, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 216, + 434 + ], + "type": "text", + "content": "4.1. Attention-based Style Injection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "text", + "content": "We start from the observation in previous image-to-image translation methods, especially Prompt-to-Prompt [14]. The key idea of their method is changing the text condition for cross-attention (CA) while keeping the attention map. Since the attention map affects the spatial layout of output, substituted text conditions determine what to draw in the generated image, and these conditions are actually key and value in CA. Inspired by them, we manipulate the features in self-attention layer as like cross-attention, regarding the features from style image " + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "inline_equation", + "content": "I^{s}" + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "text", + "content": " as the condition. Specifically, in the generation process, we substitute the key and value of content image with those of style for transferring the texture of style image into the content image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 602, + 287, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 602, + 287, + 698 + ], + "spans": [ + { + "bbox": [ + 46, + 602, + 287, + 698 + ], + "type": "text", + "content": "To this end, we first obtain the latent for content and style images with DDIM inversion [39], and then collect the SA features of style image over the DDIM inversion process. Specifically, for pre-defined timesteps " + }, + { + "bbox": [ + 46, + 602, + 287, + 698 + ], + "type": "inline_equation", + "content": "t = \\{0,\\dots,T\\}" + }, + { + "bbox": [ + 46, + 602, + 287, + 698 + ], + "type": "text", + "content": ", style and content images " + }, + { + "bbox": [ + 46, + 602, + 287, + 698 + ], + "type": "inline_equation", + "content": "z_0^c" + }, + { + "bbox": [ + 46, + 602, + 287, + 698 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 602, + 287, + 698 + ], + "type": "inline_equation", + "content": "z_0^s" + }, + { + "bbox": [ + 46, + 602, + 287, + 698 + ], + "type": "text", + "content": " are inverted from image " + }, + { + "bbox": [ + 46, + 602, + 287, + 698 + ], + "type": "inline_equation", + "content": "(t = 0)" + }, + { + "bbox": [ + 46, + 602, + 287, + 698 + ], + "type": "text", + "content": " to gaussian noise " + }, + { + "bbox": [ + 46, + 602, + 287, + 698 + ], + "type": "inline_equation", + "content": "(t = T)" + }, + { + "bbox": [ + 46, + 602, + 287, + 698 + ], + "type": "text", + "content": ". During DDIM inversion, we also collect query features of content " + }, + { + "bbox": [ + 46, + 602, + 287, + 698 + ], + "type": "inline_equation", + "content": "(Q_t^c)" + }, + { + "bbox": [ + 46, + 602, + 287, + 698 + ], + "type": "text", + "content": " and key and value features of style " + }, + { + "bbox": [ + 46, + 602, + 287, + 698 + ], + "type": "inline_equation", + "content": "(K_t^s,V_t^s)" + }, + { + "bbox": [ + 46, + 602, + 287, + 698 + ], + "type": "text", + "content": " at every time steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "After that, we initialize stylized latent noise " + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "z_T^{cs}" + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": " by copy-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "text", + "content": "ing content latent noise " + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "inline_equation", + "content": "z_{T}^{c}" + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "text", + "content": ". Then, we transfer the target style to the stylized latent by injecting the key " + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "inline_equation", + "content": "K_{t}^{s}" + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "inline_equation", + "content": "V_{t}^{s}" + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "text", + "content": " collected from the style into SA layer, instead of the original key " + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "inline_equation", + "content": "K_{t}^{cs}" + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "inline_equation", + "content": "V_{t}^{cs}" + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "text", + "content": ", when performing the entire reverse process of stylized latent " + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "inline_equation", + "content": "z_{t}^{cs}" + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "text", + "content": ". However, only applying this substitution can lead to content disruption, since the content of stylized latent would be progressively changed as attended value changes. Hence, we propose query preservation to maintain original content. Simply, we blend query of stylized latent " + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "inline_equation", + "content": "Q_{t}^{cs}" + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "text", + "content": " and that of content " + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "inline_equation", + "content": "Q_{t}^{c}" + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "text", + "content": " for the entire reverse process. These style injection and query preservation processes at time step " + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "text", + "content": " are expressed as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 358, + 577, + 545, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 577, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 358, + 577, + 545, + 591 + ], + "type": "interline_equation", + "content": "\\tilde {Q} _ {t} ^ {c s} = \\gamma \\times Q _ {t} ^ {c} + (1 - \\gamma) \\times Q _ {t} ^ {c s}, \\tag {3}", + "image_path": "034761c6a84638026ecceee178d49ec69d5de432c501db36029259b17ee227ec.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 370, + 606, + 545, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 606, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 370, + 606, + 545, + 620 + ], + "type": "interline_equation", + "content": "\\phi_ {\\text {o u t}} ^ {\\mathrm {c s}} = \\operatorname {A t t n} \\left(\\tilde {Q} _ {t} ^ {c s}, K _ {t} ^ {s}, V _ {t} ^ {s}\\right), \\tag {4}", + "image_path": "c31eac7cba1ef274f208c4fd152324e07bcc42e8529630557900a041e707b670.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " is degree of blending in range of [0, 1]. In addition, we apply these operations on the latter layers of decoder (7- " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "12^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " decoder layers in SD) relevant to local textures. We also highlight that the proposed method can adjust the degree of style transfer by changing query preservation ratio " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ". Specifically, higher " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " maintains more content, while lower " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " strengthens effects of style transfer." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8798" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 180, + 150 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 180, + 150 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 180, + 150 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 180, + 150 + ], + "type": "image", + "image_path": "3dfeb95f4b4a72fd4f26c9fa31b3133f0dc69f21163d9f1bc6c7575f44d06a02.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 153, + 156, + 163 + ], + "lines": [ + { + "bbox": [ + 75, + 153, + 156, + 163 + ], + "spans": [ + { + "bbox": [ + 75, + 153, + 156, + 163 + ], + "type": "text", + "content": "(a) Std. of attention map" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 179, + 153, + 286, + 163 + ], + "lines": [ + { + "bbox": [ + 179, + 153, + 286, + 163 + ], + "spans": [ + { + "bbox": [ + 179, + 153, + 286, + 163 + ], + "type": "text", + "content": "(b) Ratio of std. of attention map" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 180, + 72, + 284, + 150 + ], + "blocks": [ + { + "bbox": [ + 180, + 72, + 284, + 150 + ], + "lines": [ + { + "bbox": [ + 180, + 72, + 284, + 150 + ], + "spans": [ + { + "bbox": [ + 180, + 72, + 284, + 150 + ], + "type": "image", + "image_path": "ac32451fdf99d903083809667e55399f7e6415ded3047660913ea5c999dfe958.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 167, + 288, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 288, + 245 + ], + "type": "text", + "content": "Figure 4. Visualization of the standard deviation of attention map before softmax. (a) Attention-based style injection reduces the standard deviation of self-attention map. Original denotes SA maps from the generation process without style injection. We use both style and content images for generation. (b) We compute the ratio between attention maps w/ and w/o style injection. For the std of original image, we use averaged std. of content and style." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 254, + 216, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 254, + 216, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 254, + 216, + 266 + ], + "type": "text", + "content": "4.2. Attention Temperature Scaling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 272, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 272, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 287, + 390 + ], + "type": "text", + "content": "Attention map is computed by scaled dot-product between query and key features. During training, query and key features in the SA layer originate from an identical image. However, if we substitute the key features with those of style images, the magnitude of similarity would be overall lowered as style and content are highly likely to be irrelevant. Thus, the computed attention map can be blurred or smoothed, and it would further make output images unsharp, which is detrimental to capturing both content and style information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 392, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 487 + ], + "type": "text", + "content": "To quantify this issue, we measure the standard deviation of attention map, while ablating the attention-based style injection. In detail, we calculate the attention map before applying softmax, which is scaled-dot product between query and key. As shown in Fig. 4 (a), we validate that this style injection tends to lower the standard deviation of the attention map over the entire timesteps. That is, attention maps after softmax with style injection would be overly smooth." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 487, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 559 + ], + "type": "text", + "content": "To rectify the attention map sharper, we introduce an attention temperature scaling parameter. In detail, we multiply the attention map before softmax by a constant temperature scaling parameter " + }, + { + "bbox": [ + 46, + 487, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 487, + 287, + 559 + ], + "type": "text", + "content": " larger than 1. Thus, the attention map after softmax would be sharper than its original values. The modified attention process is represented as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 567, + 286, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 286, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 286, + 604 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} _ {\\tau} \\left(\\tilde {Q _ {t} ^ {c s}}, K _ {t} ^ {s}, V _ {t} ^ {s}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {\\tau \\tilde {Q _ {t} ^ {c s}} \\left(K _ {t} ^ {s}\\right) ^ {T}}{\\sqrt {d}}\\right) \\cdot V _ {t} ^ {s}, \\tau > 1. \\tag {5}", + "image_path": "e29bcc9c0bd42242b1143f91824732ba30d89b8641b618d3ec0889d2bffad2e5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 652 + ], + "type": "text", + "content": "We use " + }, + { + "bbox": [ + 46, + 605, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\tau = 1.5" + }, + { + "bbox": [ + 46, + 605, + 287, + 652 + ], + "type": "text", + "content": " as a default setting, which is the average ratio over entire timesteps. As reported in Fig. 4 (b), we confirm that it effectively calibrates the standard deviation of attention map similar to its original values." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 659, + 167, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 167, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 167, + 670 + ], + "type": "text", + "content": "4.3. Initial Latent AdaIN" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "In artistic style transfer, the color tone generally takes up a significant portion of the style information. In this context, we observe that the style transfer only with attention-based" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 71, + 457, + 143 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 457, + 143 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 457, + 143 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 457, + 143 + ], + "type": "image", + "image_path": "8cfdad2fb268773913895461a5819b09eacd13147ba70e421af6bead6f86e5ba.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 144, + 515, + 153 + ], + "lines": [ + { + "bbox": [ + 337, + 144, + 515, + 153 + ], + "spans": [ + { + "bbox": [ + 337, + 144, + 515, + 153 + ], + "type": "text", + "content": "(a) Generated results only w/ style injection " + }, + { + "bbox": [ + 337, + 144, + 515, + 153 + ], + "type": "inline_equation", + "content": "(\\mathrm{K}^s,V^s)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 457, + 72, + 531, + 143 + ], + "blocks": [ + { + "bbox": [ + 457, + 72, + 531, + 143 + ], + "lines": [ + { + "bbox": [ + 457, + 72, + 531, + 143 + ], + "spans": [ + { + "bbox": [ + 457, + 72, + 531, + 143 + ], + "type": "image", + "image_path": "dde322ca136330611fdd146218e09cf6e5cd3e362c095d788c8b9605543f7f15.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 144, + 529, + 153 + ], + "lines": [ + { + "bbox": [ + 459, + 144, + 529, + 153 + ], + "spans": [ + { + "bbox": [ + 459, + 144, + 529, + 153 + ], + "type": "text", + "content": "njection " + }, + { + "bbox": [ + 459, + 144, + 529, + 153 + ], + "type": "inline_equation", + "content": "(\\mathrm{K}^s,V^s)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 318, + 158, + 457, + 231 + ], + "blocks": [ + { + "bbox": [ + 318, + 158, + 457, + 231 + ], + "lines": [ + { + "bbox": [ + 318, + 158, + 457, + 231 + ], + "spans": [ + { + "bbox": [ + 318, + 158, + 457, + 231 + ], + "type": "image", + "image_path": "7c4e3c4344e80896acccab14266518ab7f61d2f69ac574ece9e3eb52bc3f048a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 330, + 232, + 522, + 243 + ], + "lines": [ + { + "bbox": [ + 330, + 232, + 522, + 243 + ], + "spans": [ + { + "bbox": [ + 330, + 232, + 522, + 243 + ], + "type": "text", + "content": "(b) Generated results only w/ style injection " + }, + { + "bbox": [ + 330, + 232, + 522, + 243 + ], + "type": "inline_equation", + "content": "(\\mathrm{Q}^s,\\mathrm{K}^s,V^s)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 247, + 547, + 324 + ], + "lines": [ + { + "bbox": [ + 305, + 247, + 547, + 324 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 547, + 324 + ], + "type": "text", + "content": "Figure 5. Generated results only w/ style injection. (a) We observe that generated images only with attention-based style injection do not harmonize with the given style in the aspect of color tone. (b) To identify the effects of every feature in SA on color tones, we additionally include query in the style injection process. However, color tones still resemble those of content, concluding features in self-attention have less effect on the color tones." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 457, + 158, + 531, + 230 + ], + "blocks": [ + { + "bbox": [ + 457, + 158, + 531, + 230 + ], + "lines": [ + { + "bbox": [ + 457, + 158, + 531, + 230 + ], + "spans": [ + { + "bbox": [ + 457, + 158, + 531, + 230 + ], + "type": "image", + "image_path": "8c8d4174f83169458e3f8dc0035282fc426522048821c800fcd49fd4a8947b3f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 335, + 547, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 547, + 418 + ], + "type": "text", + "content": "style injection often fails in terms of capturing the color tone of the given style. As shown in Fig. 5 (a), textures and local patterns are successfully transferred to the content image while the color tone of the content image still remains. Furthermore, even with injecting the query, key, and value of styles, the resulting images still preserve the color tone of the content, as shown in Fig. 5 (b)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 418, + 546, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 546, + 514 + ], + "type": "text", + "content": "As substituting the self-attention features has less effects to color tone, we analyze the other vital part of DM; initial latent noise. One of the recent discoveries in DM is that the DM struggles to synthesize purely white or black images [13]. Instead, they tend to generate images of median color as the initial noise is sampled from zero mean and unit variance. Thus, we hypothesize the statistics of initial noise largely affect the colors and brightness of generated images." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 514, + 547, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 645 + ], + "type": "text", + "content": "Based on this assumption, we attempt to use the initial latent of style " + }, + { + "bbox": [ + 304, + 514, + 547, + 645 + ], + "type": "inline_equation", + "content": "z_{T}^{s}" + }, + { + "bbox": [ + 304, + 514, + 547, + 645 + ], + "type": "text", + "content": " for the style transfer process. However, if we simply start to generate the image from style latent " + }, + { + "bbox": [ + 304, + 514, + 547, + 645 + ], + "type": "inline_equation", + "content": "z_{T}^{s}" + }, + { + "bbox": [ + 304, + 514, + 547, + 645 + ], + "type": "text", + "content": ", the structural information of synthesized results also follows the style image and loses the structure of the content. To harness valuable information in both initial latents, we consider that the tone information is intricately connected with the channel statistics of the initial latent, following the principle underlying Style Loss [11] and AdaIN [17]. Thus, we employ AdaIN to modulate the initial latent for effective tone information transfer, represented as:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 345, + 654, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 654, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 345, + 654, + 545, + 681 + ], + "type": "interline_equation", + "content": "z _ {T} ^ {c s} = \\sigma \\left(z _ {T} ^ {s}\\right) \\left(\\frac {z _ {T} ^ {c} - \\mu \\left(z _ {T} ^ {c}\\right)}{\\sigma \\left(z _ {T} ^ {c}\\right)}\\right) + \\mu \\left(z _ {T} ^ {s}\\right), \\tag {6}", + "image_path": "851a0051d0f443e8d52f0caea733bf1443c7fd1f2ebbbe5d3f2bfab965997a8e.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mu (\\cdot),\\sigma (\\cdot)" + }, + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "type": "text", + "content": " denote channel-wise mean and standard deviation, respectively. Based on this, the initial latent " + }, + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "type": "inline_equation", + "content": "z_{T}^{cs}" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8799" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 544, + 136 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 544, + 136 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 544, + 136 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 544, + 136 + ], + "type": "table", + "html": "
MetricOursAesPA-NetCASTStyTR2EFDMMASTAdaAttnArtFlowAdaConvAdaINDiffuseITInSTDiffStyle
ArtFID ↓28.80131.42034.68530.72034.60531.28230.35034.63031.85630.93340.72140.63341.464
FID ↓18.13119.76020.39518.89020.06218.19918.65821.25219.02218.24223.06521.57120.903
LPIPS ↓0.50550.51350.62120.54450.64300.62930.54390.55620.59100.60760.69210.80020.8931
CFSD ↓0.22810.24640.29180.30110.33460.30430.28620.29200.36000.31550.34280.67590.2819
", + "image_path": "338e8deaf2ebc4d404dadea79f5e31bf92496db89fb348e14e2423eaae02ebae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 76, + 138, + 515, + 149 + ], + "lines": [ + { + "bbox": [ + 76, + 138, + 515, + 149 + ], + "spans": [ + { + "bbox": [ + 76, + 138, + 515, + 149 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with conventional " + }, + { + "bbox": [ + 76, + 138, + 515, + 149 + ], + "type": "inline_equation", + "content": "\\left( {{3}^{\\text{rd }} - {11}^{\\text{th }}\\text{columns}}\\right)" + }, + { + "bbox": [ + 76, + 138, + 515, + 149 + ], + "type": "text", + "content": " and diffusion model baselines " + }, + { + "bbox": [ + 76, + 138, + 515, + 149 + ], + "type": "inline_equation", + "content": "\\left( {{12}^{\\text{th }} - {14}^{\\text{th }}\\text{columns}}\\right)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 159, + 287, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 159, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 287, + 185 + ], + "type": "text", + "content": "preserves content information from " + }, + { + "bbox": [ + 46, + 159, + 287, + 185 + ], + "type": "inline_equation", + "content": "z_T^c" + }, + { + "bbox": [ + 46, + 159, + 287, + 185 + ], + "type": "text", + "content": " while aligning the channel-wise mean and standard deviation with " + }, + { + "bbox": [ + 46, + 159, + 287, + 185 + ], + "type": "inline_equation", + "content": "z_T^s" + }, + { + "bbox": [ + 46, + 159, + 287, + 185 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 194, + 128, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 194, + 128, + 208 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 128, + 208 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 214, + 173, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 214, + 173, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 214, + 173, + 227 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 232, + 288, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 288, + 293 + ], + "type": "text", + "content": "We conduct all experiments in Stable Diffusion 1.4 pretrained on LAION dataset [36] and adopt DDIM sampling [39] with a total 50 timesteps (" + }, + { + "bbox": [ + 46, + 232, + 288, + 293 + ], + "type": "inline_equation", + "content": "t = \\{1, \\dots, 50\\}" + }, + { + "bbox": [ + 46, + 232, + 288, + 293 + ], + "type": "text", + "content": "). For default settings for hyperparameters, we use " + }, + { + "bbox": [ + 46, + 232, + 288, + 293 + ], + "type": "inline_equation", + "content": "\\gamma = 0.75" + }, + { + "bbox": [ + 46, + 232, + 288, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 232, + 288, + 293 + ], + "type": "inline_equation", + "content": "\\tau = 1.5" + }, + { + "bbox": [ + 46, + 232, + 288, + 293 + ], + "type": "text", + "content": ", if they are not mentioned separately." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 300, + 163, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 163, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 163, + 312 + ], + "type": "text", + "content": "5.2. Evaluation Protocol" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 319, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 464 + ], + "type": "text", + "content": "Conventional style transfer methods typically utilize Style Loss [11] as both training objective and evaluation metric, so their results tend to overfit the Style Loss. Thus, for a fair comparison, we employ a recently proposed metric, ArtFID [46] which evaluates overall style transfer performances with consideration of both content and style preservation and also is known as strongly coinciding with human judgment. Specifically, ArtFID is computed as " + }, + { + "bbox": [ + 46, + 319, + 287, + 464 + ], + "type": "inline_equation", + "content": "(\\text{ArtFID} = (1 + \\text{LPIPS}) \\cdot (1 + \\text{FID}))" + }, + { + "bbox": [ + 46, + 319, + 287, + 464 + ], + "type": "text", + "content": ". LPIPS [50] measures content fidelity between the stylized image and the corresponding content image, and FID [15] assesses the style fidelity between the stylized image and the corresponding style image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 468, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 540 + ], + "type": "text", + "content": "Dataset. Our evaluations employ content images from MS-COCO [27] dataset and style images from WikiArt [40] dataset. All input images are center-cropped to " + }, + { + "bbox": [ + 46, + 468, + 287, + 540 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 468, + 287, + 540 + ], + "type": "text", + "content": " resolution. Also, for quantitative comparison, we randomly selected 20 content and 40 style images from each dataset, yielding 800 stylized images as StyTR² [9] has done." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": "Content Feature Structural Distance (CFSD). In the style transfer evaluation, the assessment of content fidelity often relies on the LPIPS distance. However, since LPIPS utilizes the feature space of AlexNet [20] pre-trained for classification task on ImageNet [7], which is known as texture-biased [12]. Thus, the style information of the images can affect the LPIPS score. To mitigate this style influence, we additionally introduce Content Feature Structural Distance (CFSD) which is a distance measure that only considers the spatial correlation between image patches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "In detail, we first define the correlation map between image patch features as follows. For a given image " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": ", we obtain feature maps " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{hw \\times c}" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": ", which is the output feature of conv3 in VGG19 [38]. Then, we calculate the patch sim" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "text", + "content": "The similarity map " + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "inline_equation", + "content": "M = F \\times F^T" + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{hw \\times hw}" + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "text", + "content": ", which is a similarity map between every pair of features in " + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "text", + "content": ". After that, for computing the distance between two patch similarity maps, we model the similarity between a single patch and the others as a probability distribution by applying softmax operation. Finally, the correlation map is represented as " + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "inline_equation", + "content": "S = [\\text{softmax}(M_i)]_{i=1}^{hw}" + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{hw \\times hw}" + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "inline_equation", + "content": "M_i \\in \\mathbb{R}^{1 \\times hw}" + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "text", + "content": " is a similarity map between " + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "text", + "content": " patch and the other patches." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 255, + 545, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 303 + ], + "type": "text", + "content": "Then, CFSD is defined as KL-divergence between two correlation maps. In our case, we compute CFSD between the correlation map of the content " + }, + { + "bbox": [ + 304, + 255, + 545, + 303 + ], + "type": "inline_equation", + "content": "(S^c)" + }, + { + "bbox": [ + 304, + 255, + 545, + 303 + ], + "type": "text", + "content": " and stylized images " + }, + { + "bbox": [ + 304, + 255, + 545, + 303 + ], + "type": "inline_equation", + "content": "(S^{cs})" + }, + { + "bbox": [ + 304, + 255, + 545, + 303 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 358, + 312, + 545, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 312, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 358, + 312, + 545, + 345 + ], + "type": "interline_equation", + "content": "\\mathrm {C F S D} = \\frac {1}{h w} \\sum_ {i = 1} ^ {h w} D _ {\\mathrm {K L}} \\left(S _ {i} ^ {c} | | S _ {i} ^ {c s}\\right), \\tag {7}", + "image_path": "2b6d2ef42cb4cab9c9e93c022e8ae10d04415e6c5d9069f0bc05381fa3b5976f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 360, + 448, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 360, + 448, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 360, + 448, + 373 + ], + "type": "text", + "content": "5.3.Quantitative Comparison" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 379, + 545, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 545, + 487 + ], + "type": "text", + "content": "We evaluate our proposed method through comparison with twelve state-of-the-art methods, including nine conventional style transfer methods (AesPA-Net [16], CAST [52], StyTR" + }, + { + "bbox": [ + 304, + 379, + 545, + 487 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 379, + 545, + 487 + ], + "type": "text", + "content": " [9], EFDM [51], MAST [8], AdaAttN [28], ArtFlow [1], AdaConv [5], AdaIN [17]) and three diffusion-based style transfer methods (DiffuseIT [22], InST [53], DiffStyle [18]), which have a style image as input. We employ the publicly available implementations of all baselines, using their recommended configurations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 492, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 545, + 611 + ], + "type": "text", + "content": "Comparison with Conventional Style Transfer. As shown in Tab. 1, our method largely surpasses the conventional style transfer methods in terms of ArtFID, which is known as coinciding the human preference. In addition, the proposed method records the lowest FID, which denotes that stylized images highly resemble the target styles. For content fidelity metrics, ours shows superior scores in both CFSD and LPIPS. We point out that ours achieves much lower CFSD compared to other methods, which is the metric to only consider the spatial correlation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 612, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 612, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 545, + 673 + ], + "type": "text", + "content": "In addition, we also emphasize that the proposed method can arbitrarily adjust the degree of style transfer by changing the " + }, + { + "bbox": [ + 304, + 612, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 612, + 545, + 673 + ], + "type": "text", + "content": ", and the proposed method significantly surpasses all the other methods in terms of FID (style), when we match the value of LPIPS (content) (Fig. 10)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Comparison with Diffusion-based Style Transfer. Our method demonstrates the best performance in terms of LPIPS, FID, and their combination (ArtFID) with a large" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8800" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 51, + 69, + 542, + 80 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 69, + 542, + 80 + ], + "spans": [ + { + "bbox": [ + 51, + 69, + 542, + 80 + ], + "type": "text", + "content": "Content Style Ours AesPA-Net CAST StyTR2 EFDM MAST AdaAttN AdaIN DiffuseIT InST DiffStyle" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 80, + 545, + 233 + ], + "blocks": [ + { + "bbox": [ + 48, + 80, + 545, + 233 + ], + "lines": [ + { + "bbox": [ + 48, + 80, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 48, + 80, + 545, + 233 + ], + "type": "image", + "image_path": "860ff3210846250b16d879d171e60f15c4647f18d2abd94bffd6a5026d2fa2ad.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 235, + 515, + 246 + ], + "lines": [ + { + "bbox": [ + 77, + 235, + 515, + 246 + ], + "spans": [ + { + "bbox": [ + 77, + 235, + 515, + 246 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison with conventional " + }, + { + "bbox": [ + 77, + 235, + 515, + 246 + ], + "type": "inline_equation", + "content": "(4^{\\mathrm{th}} - 10^{\\mathrm{th}}" + }, + { + "bbox": [ + 77, + 235, + 515, + 246 + ], + "type": "text", + "content": " columns) and diffusion model baselines " + }, + { + "bbox": [ + 77, + 235, + 515, + 246 + ], + "type": "inline_equation", + "content": "(11^{\\mathrm{th}} - 13^{\\mathrm{th}}" + }, + { + "bbox": [ + 77, + 235, + 515, + 246 + ], + "type": "text", + "content": " columns)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 252, + 276, + 356 + ], + "blocks": [ + { + "bbox": [ + 58, + 252, + 276, + 356 + ], + "lines": [ + { + "bbox": [ + 58, + 252, + 276, + 356 + ], + "spans": [ + { + "bbox": [ + 58, + 252, + 276, + 356 + ], + "type": "image", + "image_path": "2838e8ad98d425397cce2a5d39e7ef935a30c31d0c61a70bce8bd7758a67ed21.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 357, + 286, + 388 + ], + "lines": [ + { + "bbox": [ + 46, + 357, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 286, + 388 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison with best ArtFID (AdaAttN) and most recently proposed baselines (AesPA-Net) with additional zooming details" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 82, + 395, + 251, + 426 + ], + "blocks": [ + { + "bbox": [ + 82, + 395, + 251, + 426 + ], + "lines": [ + { + "bbox": [ + 82, + 395, + 251, + 426 + ], + "spans": [ + { + "bbox": [ + 82, + 395, + 251, + 426 + ], + "type": "table", + "html": "
MetricDiffuseITInSTDiffStyleOurs
Time (sec)792.8816.9355.912.4
", + "image_path": "a075caf6dea678e94b4921075cb4d81fd5903e126d48e55406b153bbf8e7277b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 430, + 286, + 452 + ], + "lines": [ + { + "bbox": [ + 46, + 430, + 286, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 286, + 452 + ], + "type": "text", + "content": "Table 2. Comparison of inference time of diffusion-based methods for style transferring a given style and content pair" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 459, + 286, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 286, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 286, + 626 + ], + "type": "text", + "content": "margin, as shown in Tab. 1. One significant factor for diffusion models is their running times, since they require several steps to synthesize a single image and it requires inevitable time cost. Hence, we measure the inference time for a pair of content and style images on a single Titan RTX GPU, as shown in Tab. 2. Our method requires a total of 12.4 seconds, with 8.2 seconds for DDIM inversions and 4.2 seconds for sampling costs. As reported, we validate the proposed method significantly faster than other methods, even exploiting large-scale DM. This faster speed comes from the fact that the proposed methods can use the much smaller steps of DDIM inversion, because we additionally utilize the features collected during inversion steps, largely reducing the necessity for perfect inversion of content and style." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 635, + 182, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 182, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 182, + 647 + ], + "type": "text", + "content": "5.4. Qualitative Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 654, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 286, + 712 + ], + "type": "text", + "content": "Comparison with Conventional Style Transfer. As shown in Fig. 6, we observe that our method tends to highly preserve the structural information of the content image, while also transferring the style well. For instance, as shown in the third row, ours retains the structure of the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 254, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 254, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 254, + 545, + 289 + ], + "type": "text", + "content": "bridge, but the baselines struggle to preserve structure or transfer the style. We also provide the qualitative comparison with zooming details in Fig. 7 and Supplementary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 295, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 295, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 305, + 295, + 545, + 427 + ], + "type": "text", + "content": "Comparison with Diffusion-based Style Transfer. We also compare our method with recent diffusion-based style transfer baselines [18, 23, 53]. As shown in Fig. 6, we observe the proposed technique transfers the style to the content well. On the other hand, baselines often lose the structure of content or fail to transfer the style, when an arbitrary content style pair is given. For instance, DiffuseIT and DiffStyle suffer from generating shape and visually plausible images or drop the original content. Differently, InST synthesizes the realistic images, while struggling to transfer style (1" + }, + { + "bbox": [ + 305, + 295, + 545, + 427 + ], + "type": "inline_equation", + "content": "^{\\text{st}}" + }, + { + "bbox": [ + 305, + 295, + 545, + 427 + ], + "type": "text", + "content": " row) or change content of image (2" + }, + { + "bbox": [ + 305, + 295, + 545, + 427 + ], + "type": "inline_equation", + "content": "^{\\text{nd}}" + }, + { + "bbox": [ + 305, + 295, + 545, + 427 + ], + "type": "text", + "content": ", 3" + }, + { + "bbox": [ + 305, + 295, + 545, + 427 + ], + "type": "inline_equation", + "content": "^{\\text{rd}}" + }, + { + "bbox": [ + 305, + 295, + 545, + 427 + ], + "type": "text", + "content": " rows)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 437, + 399, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 399, + 450 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 399, + 450 + ], + "type": "text", + "content": "5.5. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 456, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 647 + ], + "type": "text", + "content": "To validate the effectiveness of the proposed components, we conduct ablation studies in both quantitative and qualitative ways. As shown in Fig. 8 and Tab. 3, style injection is significant for guiding the style and content of given images (Config. B). Besides, initial latent AdaIN has a large portion of transferring the color tone of style (Config. D). Attention temperature scaling is in charge of enhancement of quality in synthesized results such as sharpening details and resolving blurriness. For instance, this scaling jointly reduces the FID and LPIPS (Config. " + }, + { + "bbox": [ + 304, + 456, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^*" + }, + { + "bbox": [ + 304, + 456, + 545, + 647 + ], + "type": "text", + "content": " vs. C in Tab. 3). For more detailed analysis, we provide quantitative metrics with the style-content trade-off, while changing the attention scaling parameter " + }, + { + "bbox": [ + 304, + 456, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 456, + 545, + 647 + ], + "type": "text", + "content": " in Fig. 10 (b). As reported, attention scaling effectively reduces both FID and LPIPS, proving its effects on the preservation of content and capability of style transfer (" + }, + { + "bbox": [ + 304, + 456, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\tau = 1.0" + }, + { + "bbox": [ + 304, + 456, + 545, + 647 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 304, + 456, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\tau = 1.5" + }, + { + "bbox": [ + 304, + 456, + 545, + 647 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 658, + 421, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 421, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 421, + 670 + ], + "type": "text", + "content": "5.6. Additional Analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Content-Style Trade-Off. Our proposed method offers flexible control of the trade-off relation between content and style fidelity by adjusting the parameter " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": ", as discussed in" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8801" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 291, + 155 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 291, + 155 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 291, + 155 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 291, + 155 + ], + "type": "table", + "html": "
ConfigurationArtFIDFIDLPIPS
AOurs (γ = 0.75, default)28.8018.130.505
A*Ours (γ = 0.6)27.9717.210.535
B- Style Inject. (Sec. 4.1)43.7227.130.554
C- Attn. Scaling (Sec. 4.2)29.0217.810.542
D- Init. AdaIN (Sec. 4.3)29.2620.050.390
", + "image_path": "5b1b936a45b01d2bdd11131aae08a19b15765285713166401ef2ee2900e905a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 159, + 257, + 170 + ], + "lines": [ + { + "bbox": [ + 77, + 159, + 257, + 170 + ], + "spans": [ + { + "bbox": [ + 77, + 159, + 257, + 170 + ], + "type": "text", + "content": "Table 3. Ablation study on proposed components" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 52, + 176, + 281, + 288 + ], + "blocks": [ + { + "bbox": [ + 52, + 176, + 281, + 288 + ], + "lines": [ + { + "bbox": [ + 52, + 176, + 281, + 288 + ], + "spans": [ + { + "bbox": [ + 52, + 176, + 281, + 288 + ], + "type": "image", + "image_path": "b89ab7b4bcfdd501d4f0c4ece99c1f74adbac5ab3e9aafff8a18fd52f34f13c6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 292, + 266, + 303 + ], + "lines": [ + { + "bbox": [ + 67, + 292, + 266, + 303 + ], + "spans": [ + { + "bbox": [ + 67, + 292, + 266, + 303 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison with ablation studies" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 47, + 306, + 282, + 435 + ], + "blocks": [ + { + "bbox": [ + 47, + 306, + 282, + 435 + ], + "lines": [ + { + "bbox": [ + 47, + 306, + 282, + 435 + ], + "spans": [ + { + "bbox": [ + 47, + 306, + 282, + 435 + ], + "type": "image", + "image_path": "4cf5a63c61f5638d6fffff5dc9acbbc0b67f136c825e93b8776c146c9759bf5c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 438, + 280, + 451 + ], + "lines": [ + { + "bbox": [ + 53, + 438, + 280, + 451 + ], + "spans": [ + { + "bbox": [ + 53, + 438, + 280, + 451 + ], + "type": "text", + "content": "Figure 9. Visualization of effects of query preservation ratio " + }, + { + "bbox": [ + 53, + 438, + 280, + 451 + ], + "type": "inline_equation", + "content": "\\gamma" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 464, + 287, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 560 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 560 + ], + "type": "text", + "content": "Sec. 4.1. In detail, we compute the FID and LPIPS while varying " + }, + { + "bbox": [ + 46, + 464, + 287, + 560 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 464, + 287, + 560 + ], + "type": "text", + "content": " within the range [0.3, 1] with a step size of 0.1. As shown in Fig. 10 (a), our method surpasses baseline methods across all ranges of content and style fidelity. This result implies the proposed method significantly outperforms the other methods, when we match style or content metric to the compared model by adjusting the " + }, + { + "bbox": [ + 46, + 464, + 287, + 560 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 464, + 287, + 560 + ], + "type": "text", + "content": " of ours. Note that, dotted lines refer to our model reported in Tab. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 562, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 562, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 562, + 287, + 635 + ], + "type": "text", + "content": "We also visualize the effects of the style-content trade-off by synthesizing images by adjusting " + }, + { + "bbox": [ + 46, + 562, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 562, + 287, + 635 + ], + "type": "text", + "content": ". As shown in Fig. 9, the lower " + }, + { + "bbox": [ + 46, + 562, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 562, + 287, + 635 + ], + "type": "text", + "content": " highly reflects the style while losing the content of the given image, and vice versa. This characteristic of the proposed method suggests that the users can adjust the degree of style by following their personal preferences." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Study on the value of " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": ". We observe that the gradual increase of " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " enhances the performance of style transfer, although its effects on enhancement become smaller as " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " goes larger, as shown in Fig. 10 (b). This result implies that the attention temperature scaling effectively works with a simple modification of the magnitude of the attention map." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 301, + 73, + 451, + 194 + ], + "blocks": [ + { + "bbox": [ + 301, + 73, + 451, + 194 + ], + "lines": [ + { + "bbox": [ + 301, + 73, + 451, + 194 + ], + "spans": [ + { + "bbox": [ + 301, + 73, + 451, + 194 + ], + "type": "image", + "image_path": "0574ed7e223feb7c0b3309243ecdcabf39ec3c7cba7af2ebe674151a020c830e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 197, + 489, + 208 + ], + "lines": [ + { + "bbox": [ + 361, + 197, + 489, + 208 + ], + "spans": [ + { + "bbox": [ + 361, + 197, + 489, + 208 + ], + "type": "text", + "content": "Figure 10. Style-content trade-offs" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 452, + 73, + 541, + 194 + ], + "blocks": [ + { + "bbox": [ + 452, + 73, + 541, + 194 + ], + "lines": [ + { + "bbox": [ + 452, + 73, + 541, + 194 + ], + "spans": [ + { + "bbox": [ + 452, + 73, + 541, + 194 + ], + "type": "image", + "image_path": "d13886104303c34f63fe8e4d7a7cc55500466e7e10422acf2cfaad9f948dd77d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 213, + 538, + 285 + ], + "blocks": [ + { + "bbox": [ + 316, + 213, + 538, + 285 + ], + "lines": [ + { + "bbox": [ + 316, + 213, + 538, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 538, + 285 + ], + "type": "image", + "image_path": "6f76cbcee65f8d474e662ff237a8efaf946791fe47816741a406344a17535b5e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 286, + 541, + 297 + ], + "lines": [ + { + "bbox": [ + 311, + 286, + 541, + 297 + ], + "spans": [ + { + "bbox": [ + 311, + 286, + 541, + 297 + ], + "type": "text", + "content": "Figure 11. Comparison with text-guided style transfer methods" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 305, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 545, + 426 + ], + "type": "text", + "content": "Comparison with text-guided style transfer. We additionally compare the proposed method with the style transfer methods [21, 41] which are conditioned on the textual inputs. As text-guided methods tend to modify the style largely, we use " + }, + { + "bbox": [ + 304, + 305, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\gamma = 0.3" + }, + { + "bbox": [ + 304, + 305, + 545, + 426 + ], + "type": "text", + "content": " for this experiment. Since the text condition hardly contains all the information in the style image such as texture and color tones, the transferred results less resemble the target style, as shown in Fig. 11. Differently, we validate that the proposed method successfully transfers the style with high fidelity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 432, + 378, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 432, + 378, + 444 + ], + "spans": [ + { + "bbox": [ + 306, + 432, + 378, + 444 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "text", + "content": "Our work addresses the challenges associated with diffusion model-based style transfer methods, which often require time-consuming optimization steps or struggle to leverage the generative potential of large-scale diffusion models. To this end, we propose the method of adapting the pre-trained large-scale diffusion model on style transfer in a training-free way. Our method focuses on manipulating the features of self-attention layers, akin to the cross-attention mechanism, by substituting the key and value during the content generation process with those of the style. Furthermore, we propose the query preservation and attention temperature scaling to mitigate the issue of disruption of content, and initial latent AdaIN to handle the disharmonious color (failure to transfer the colors of style). Experimental results show the superiority of our proposed method over state-of-the-art techniques in previous baselines." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 651, + 403, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 403, + 664 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 403, + 664 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "This work was supported in part by MSIT/IITP (No. 2022-0-00680, 2019-0-00421, 2020-0-01821, 2021-0-02068), and MSIT&KNPA/KIPoT (Police Lab 2.0, No. 210121M06)." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8802" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Jie An, Siyu Huang, Yibing Song, Dejing Dou, Wei Liu, and Jiebo Luo. Artflow: Unbiased image style transfer via reversible neural flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 862-871, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "text", + "content": "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 256 + ], + "type": "text", + "content": "[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22560-22570, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 288, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 302 + ], + "type": "text", + "content": "[4] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23040-23050, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "type": "text", + "content": "[5] Prashanth Chandran, Gaspard Zoss, Paulo Gotardo, Markus Gross, and Derek Bradley. Adaptive convolutions for structure-aware style transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7972-7981, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 288, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 411 + ], + "type": "text", + "content": "[6] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In ICLR 2023 (Eleventh International Conference on Learning Representations), 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 413, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 458 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 288, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 288, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 288, + 512 + ], + "type": "text", + "content": "[8] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM international conference on multimedia, pages 2719-2727, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 288, + 568 + ], + "type": "text", + "content": "[9] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11326-11336, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "text", + "content": "[10] Martin Nicolas Everaert, Marco Bocchio, Sami Arpa, Sabine Susstrunk, and Radhakrishna Achanta. Diffusion in style. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2251-2261, 2023. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "text", + "content": "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[13] Nocholas Guttenberg. Diffusion with offset noise. https://www.crosslabs.org/blog/diffusion-with-offset-noise, 2023.5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 162 + ], + "type": "text", + "content": "[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "text", + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 283 + ], + "type": "text", + "content": "[16] Kibeom Hong, Seogkyu Jeon, Junsoo Lee, Namhyuk Ahn, Kunhee Kim, Pilhyeon Lee, Daesik Kim, Youngjung Uh, and Hyeran Byun. Aespa-net: Aesthetic pattern-aware style transfer networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22758-22767, 2023. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 285, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 330 + ], + "type": "text", + "content": "[17] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "type": "text", + "content": "[18] Jaeseok Jeong, Mingi Kwon, and Youngjung Uh. Training-free style transfer emerges from h-space in diffusion models. arXiv preprint arXiv:2303.15403, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 365, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 419 + ], + "type": "text", + "content": "[19] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Conference on Computer Vision and Pattern Recognition 2023, 2023. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "type": "text", + "content": "[20] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 466, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 547, + 510 + ], + "type": "text", + "content": "[21] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 511, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 544 + ], + "type": "text", + "content": "[22] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. arXiv preprint arXiv:2209.15264, 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "type": "text", + "content": "[23] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In The Eleventh International Conference on Learning Representations, 2023. 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 590, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 644 + ], + "type": "text", + "content": "[24] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[25] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "text", + "content": "[26] Yijun Li, Ming-Yu Liu, Xueting Li, Ming-Hsuan Yang, and Jan Kautz. A closed-form solution to photorealistic image" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8803" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "stylization. In Proceedings of the European conference on computer vision (ECCV), pages 453-468, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 287, + 163 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 287, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 287, + 220 + ], + "type": "text", + "content": "[28] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 223, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 223, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 48, + 223, + 287, + 267 + ], + "type": "text", + "content": "[29] Ming Lu, Hao Zhao, Anbang Yao, Yurong Chen, Feng Xu, and Li Zhang. A closed-form solution to universal style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5952-5961, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 269, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 269, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 48, + 269, + 287, + 323 + ], + "type": "text", + "content": "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 326, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 326, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 326, + 287, + 370 + ], + "type": "text", + "content": "[31] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880–5888, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 372, + 287, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 287, + 406 + ], + "type": "text", + "content": "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 407, + 287, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 407, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 48, + 407, + 287, + 463 + ], + "type": "text", + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 464, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 464, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 464, + 287, + 529 + ], + "type": "text", + "content": "[34] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 533, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 599 + ], + "type": "text", + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 601, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 667 + ], + "type": "text", + "content": "[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[37] Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435, 2023. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[38] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 4, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[40] Wei Ren Tan, Chee Seng Chan, Hernan Aguirre, and Kiyoshi Tanaka. Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing, 28(1):394-409, 2019. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 186, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 241 + ], + "type": "text", + "content": "[41] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2, 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "type": "text", + "content": "[42] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/ diffusers, 2022.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 298, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 354 + ], + "type": "text", + "content": "[43] Huan Wang, Yijun Li, Yuehai Wang, Haoji Hu, and Ming-Hsuan Yang. Collaborative distillation for ultra-resolution universal style transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1860-1869, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 354, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 409 + ], + "type": "text", + "content": "[44] Zhizhong Wang, Lei Zhao, Haibo Chen, Lihong Qiu, Qihang Mo, Sihuan Lin, Wei Xing, and Dongming Lu. Diversified arbitrary style transfer via deep feature perturbation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7789-7798, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 410, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 454 + ], + "type": "text", + "content": "[45] Zhizhong Wang, Lei Zhao, and Wei Xing. *Style diffusions: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "text", + "content": "[46] Matthias Wright and Björn Ommer. Artfid: Quantitative evaluation of neural style transfer. In DAGM German Conference on Pattern Recognition, pages 560-576. Springer, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 500, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 533 + ], + "type": "text", + "content": "[47] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. arXiv preprint arXiv:2303.08622, 2023. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "type": "text", + "content": "[48] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. arXiv preprint arXiv:2306.07954, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 568, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 622 + ], + "type": "text", + "content": "[49] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. arXiv preprint arXiv:2305.15347, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 624, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 679 + ], + "type": "text", + "content": "[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 681, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 714 + ], + "type": "text", + "content": "[51] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In Proceedings of" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8804" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 215 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8035-8045, 2022. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "text", + "content": "[52] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 215 + ], + "type": "text", + "content": "[53] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10146-10156, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8805" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/51ff4f8c-4976-4d05-85df-cf04bbc75017_content_list.json b/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/51ff4f8c-4976-4d05-85df-cf04bbc75017_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4934c25034db015f85faed15cc89d485f1e61de0 --- /dev/null +++ b/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/51ff4f8c-4976-4d05-85df-cf04bbc75017_content_list.json @@ -0,0 +1,1594 @@ +[ + { + "type": "text", + "text": "StyleCineGAN: Landscape Cinemagraph Generation using a Pre-trained StyleGAN", + "text_level": 1, + "bbox": [ + 215, + 130, + 754, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jongwoo Choi", + "bbox": [ + 158, + 203, + 279, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kwanggyoon Seo", + "bbox": [ + 318, + 203, + 462, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Amirsaman Ashtari", + "bbox": [ + 501, + 203, + 660, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junyong Noh", + "bbox": [ + 699, + 203, + 807, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Media Lab, KAIST", + "bbox": [ + 379, + 222, + 589, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jeolpyeoni0,skg1023,a.s.ashtari,junyongnoh}@kaist.ac.kr", + "bbox": [ + 235, + 241, + 728, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1. Given a landscape image, StyleCineGAN generates a seamless cinematograph at $1024 \\times 1024$ resolution. This figure contains video clips, thus consider viewing it using Adobe Reader. The same results are also included in the supplementary video.", + "bbox": [ + 75, + 465, + 893, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 508, + 313, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a method that can generate cinematographs automatically from a still landscape image using a pretrained StyleGAN. Inspired by the success of recent unconditional video generation, we leverage a powerful pretrained image generator to synthesize high-quality cinematographs. Unlike previous approaches that mainly utilize the latent space of a pre-trained StyleGAN, our approach utilizes its deep feature space for both GAN inversion and cinematograph generation. Specifically, we propose multi-scale deep feature warping (MSDFW), which warps the intermediate features of a pre-trained StyleGAN at different resolutions. By using MSDFW, the generated cinematographs are of high resolution and exhibit plausible looping animation. We demonstrate the superiority of our method through user studies and quantitative comparisons with state-of-the-art cinematograph generation methods and a video generation method that uses a pre-trained StyleGAN.", + "bbox": [ + 75, + 527, + 472, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 801, + 209, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cinemagraph is a unique form of media that combines a still image and video. While most of the scene remains still, subtle and repeated movements in only a small region effectively highlight a particular object or an important event that occurred at the time of capture. Because this mixture", + "bbox": [ + 75, + 824, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of still and moving elements in one scene creates an interesting eye-catching effect, cinematographs have recently attracted increasing attention in social media.", + "bbox": [ + 496, + 508, + 890, + 555 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the growing popularity of cinematograph, its creation heavily relies on a manual process. To create a motion in part of the input image, one typically utilizes an image manipulation tool to stretch, scale, rotate, or shift the part of the image in a physically plausible and aesthetically pleasing way. This process is often time-consuming and requires a high level of skill in the use of image editing tools. Therefore, creating a cinematograph has generally been considered to be a personal project of professionals so far.", + "bbox": [ + 496, + 561, + 892, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To allow ordinary users to create a cinematograph with a single image, automatic methods have been proposed. One line of research uses a reference video as guidance [14, 25, 33]. These methods are capable of producing realistic motions in the target scene similar to a reference video. Recent methods obviate the need for a reference video, by training deep generative models [4, 6, 9, 15, 17-19, 23]. These methods decompose the task into two processes of learning motion and spatial information individually from separate datasets. The decomposition effectively reduces the complexity of simultaneously learning both temporal and spatial information, and improves the perceptual quality of the generated videos. However, these models", + "bbox": [ + 496, + 703, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7872", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "must be trained from scratch, which sometimes requires several days if not weeks on a modern GPU. Moreover, these models are not specifically designed to generate high-resolution of $1024 \\times 1024$ cinemagrams, because of the significant requirements in memory and processing power.", + "bbox": [ + 75, + 90, + 467, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose the first approach to high-quality one-shot landscape cinematograph generation based on a pre-trained StyleGAN [12]. By using the learned image prior of StyleGAN, our method removes the need for training a large model from scratch and systematically improves the resolution of the generated cinematographs to $1024 \\times 1024$ . Moreover, our method enables a user to easily edit the style and appearance of the generated cinematograph by leveraging the properties of StyleGAN for image stylization and editing.", + "bbox": [ + 75, + 166, + 467, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method is inspired by recent unconditional video generation methods [7, 34] which allow to navigate the latent space of a pre-trained image generator, to synthesize a high-quality temporally coherent video. Unlike these methods that utilize the latent codes of StyleGAN, we opt to use the deep features that are generated by convolution operations in each layer of StyleGAN. We use these deep features for two reasons. First, we observed that highly detailed landscape images cannot be reconstructed accurately from the latent codes using GAN inversion methods because these latent codes are low-dimensional [30]. Second, a plausible motion that preserves the content cannot be created by only navigating the latent space, because this space is highly semantic-condensed and lacks explicit spatial prior [38].", + "bbox": [ + 75, + 316, + 467, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve these problems, for the first time in the cinematograph generation domain, we utilize the deep features of a pre-trained StyleGAN. These deep features preserve spatial information and encode both high-level semantic and low-level style appearances across high and low-level convolutional layers. To produce cinematographs from these deep features, we propose a multi-scale deep feature warping (MSDFW) to apply the motions generated using a motion generator to the deep feature space of StyleGAN.", + "bbox": [ + 75, + 542, + 467, + 676 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness and advantages of our method by comparing it to state-of-the-art methods in cinematagraph generation [6, 9, 17, 19] and to an unconditional video generation method that uses a pre-trained StyleGAN [34], using various metrics. We also performed a user study to verify the effectiveness of our method for creating visually pleasing cinematagraph effects. Both qualitative and quantitative results confirm that our method substantially outperforms all existing baselines.", + "bbox": [ + 75, + 678, + 467, + 811 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 824, + 218, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Cinemagraph Generation", + "text_level": 1, + "bbox": [ + 76, + 849, + 310, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One of the early studies for cinemagraph generation used a procedural approach to decompose each object into layers", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and apply time-varying stochastic motion [5]. This method can handle a diverse range of images such as photos and paintings by relying on manual interaction from the user. Another approach is to use a reference video to animate the given image [20-22, 25]. These methods rely on a statistical motion analysis of videos to transfer their periodic motion property to the desired image. Another method distills a dynamic NeRF to render looping 3D video textures from the representative motion found in the reference video [14]. Without any reference video as guidance, Halperin et al. [8] proposed a framework to animate arbitrary objects with periodic patterns in a simple motion direction.", + "bbox": [ + 496, + 90, + 890, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent approaches use the capacity of deep learning to automatically create cinemagrams from a single image. One study trained an image-based renderer to generate water animation, utilizing the water simulation results [32]. Another research first predicts a sequence of normal maps, then generates the corresponding RGB images that show a looping animation of a garment as if it is blown in the wind [4]. Another line of research trains a generator to produce a motion field [6, 9, 15, 18, 19]. Our work is similar to these deep learning-based approaches in that a motion generator is utilized to synthesize the cinemagrams. The difference is that our method leverages the deep features of a pre-trained image generator and thus can generate plausible high-quality results without fine-tuning or training a separate synthesis network. At the same time, our method systematically improves the resolution of the generated cinemagrams to $1024 \\times 1024$ .", + "bbox": [ + 496, + 272, + 890, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Unconditional Video Generation", + "text_level": 1, + "bbox": [ + 500, + 532, + 784, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The task of video generation is known to be difficult because the generation process has to take into account both spatial and temporal information. To ease the problem, previous approaches [27, 36] decompose video generation into content and motion generation. These methods first predict the latent trajectory for motion, then generate a video from the set of predicted latent codes using the image generator. Instead of training the generation model from scratch, MoCoGAN-HD [34] and StyleVideoGAN [7] leverage a pre-trained image generator model, StyleGAN [12]. These methods use the property of well-constructed latent space to morph one frame toward the next frames [10]. Utilizing the image generation capability of a pre-trained generator, MoCoGAN-HD and StyleVideoGAN only need to train a motion trajectory generator, which greatly reduces the training time. These methods can generate high-quality videos in multiple domains such as faces, cars, and even sky. However, the content details are not well preserved, which has hindered the methods' application for cinematograph generation. Our method builds upon a similar concept of using a pre-trained StyleGAN. We leverage the deep features instead of the latent codes of StyleGAN to generate looping and content-preserving cinematographs.", + "bbox": [ + 496, + 551, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7873", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Video Synthesis using pre-trained StyleGAN", + "text_level": 1, + "bbox": [ + 76, + 90, + 454, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A pre-trained generative model, StyleGAN [11-13], has been actively employed for downstream image and video editing applications. For GAN-based applications, one of the essential steps is to project the desired images or videos to the latent space of the pre-trained GAN model, which is known as GAN inversion. The latent codes are obtained using either an optimization technique [1, 2], a trained encoder [26, 35], or a hybrid of both approaches [16]. For video editing applications, previous methods [3, 29, 37, 42] have utilized a pre-trained image encoder to project all of the frames to the latent space. While most of the methods operate in $\\mathcal{W}+$ space [2], we opt to use both $w^{+}$ and deep features [24, 43, 44] of the pre-trained StyleGAN to accurately reconstruct the original image and synthesize motion.", + "bbox": [ + 75, + 114, + 472, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 76, + 343, + 176, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of our method is shown in Figure 2. Given a landscape image $I$ , our method generates a seamlessly looping cinemagraph $V = \\{\\hat{I}_0, \\dots, \\hat{I}_N\\}$ using a pre-trained StyleGAN. At the core of our method, we use a pre-trained generator without additional fine-tuning. The overall process is described below.", + "bbox": [ + 75, + 369, + 468, + 460 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(A) First, we project the landscape image into both the latent space and the feature space of StyleGAN. To this end, we train an encoder that outputs both latent codes $w^{+}$ and intermediate features $D^{10}$ of StyleGAN (Sec. 3.1).", + "(B) In addition, we predict a mask $S$ to divide the image into static and dynamic regions (Sec. 3.2).", + "(C) Next, we use a motion generator that accepts the landscape image $I$ to synthesize a motion field $M$ , which defines the position of each pixel in the future frames (Sec. 3.3).", + "(D) Lastly, we generate the final cinematograph frames using the pre-trained StyleGAN with deep feature warping (DFW) operation added in between the original layers. (Sec. 3.4)." + ], + "bbox": [ + 60, + 465, + 468, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following sections, we will describe the details of each process.", + "bbox": [ + 75, + 681, + 468, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. GAN Inversion", + "text_level": 1, + "bbox": [ + 76, + 724, + 230, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first step for generating a cinematograph is to project the input image to the latent space of a pre-trained StyleGAN. This process is necessary because StyleGAN is an unconditional generator that cannot take a conditioning input. Many previous methods have used $\\mathcal{W} +$ [2] latent space, which is an extended space of native latent space $\\mathcal{W}$ . However, we observed that $w^{+} \\in \\mathcal{W}+$ is not expressive enough to reconstruct the original high-frequency details of the landscape images. Therefore, we chose to use the generated deep features $D^{10}$ of StyleGAN as well as $w^{+}$ . The use of $D^{10}$", + "bbox": [ + 75, + 750, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "enables us to recover details in the original input as shown in Figure 8.", + "bbox": [ + 496, + 90, + 890, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To project an image $I$ to both the latent space and feature space and obtain $w^{+}$ and $D^{10}$ , we train an encoder $E$ similarly to Yao et al. [43], in which the encoder takes $I$ as input and predicts $(w^{+}, D^{10})$ . Additional training details can be found in the supplementary material.", + "bbox": [ + 496, + 121, + 893, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Mask Prediction", + "text_level": 1, + "bbox": [ + 498, + 208, + 663, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We improve the quality of GAN inversion described in Sec. 3.1 by training an additional classifier that predicts a mask to separate the static and dynamic regions. Using the mask, the structure of the static regions can be preserved. To this end, we train a multi-layer-perceptron (MLP) classifier that accepts the deep features of StyleGAN as its input and outputs the mask that specifies each dynamic region in the image, as performed in DatasetGAN [46].", + "bbox": [ + 496, + 229, + 890, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train the classifier, we manually annotate 32 segmentation masks $S$ of the selected images. We then follow the same procedure as DatasetGAN but using the deep features $D^{i}$ where $i \\in \\{10,11,\\dots,18\\}$ . We resize all these deep features followed by concatenating them in the channel dimension to construct the input feature $D^{*}$ . In the end, paired data $(D^{*},S)$ is constructed. An ensemble of 10 MLP classifiers is trained with a cross-entropy loss. To further improve the performance of the previous work, at inference, we also refine the predicted mask. For additional details, please refer to the supplementary material.", + "bbox": [ + 496, + 353, + 892, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Motion Generation", + "text_level": 1, + "bbox": [ + 498, + 529, + 684, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To generate motion from a static landscape image $I$ , we use an Eulerian motion field $M$ which represents the immediate velocity of each pixel. To predict $M$ given $I$ , an imaged-to-image translation network [39] is trained as a motion generator. We train the network with paired data $(I, M)$ as performed in the previous method [9]. Adding controls in motion generation is also possible, either by using text prompts [19] or drawing arrows [18].", + "bbox": [ + 496, + 551, + 890, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While the trained network works well to a certain degree, we observed that the predicted motion field $M$ does not align with the boundaries between the static and dynamic regions and often contains motion in static regions. Thus, we further refine $M$ by multiplying it with the segmentation mask $S$ , which effectively removes most errors. The refined $M$ is then used to simulate the motion of a pixel at time $t$ using the Euler integration as follows:", + "bbox": [ + 496, + 674, + 890, + 794 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x _ {t + 1} = x _ {t} + M \\left(x _ {t}\\right), \\quad M \\left(x _ {t}\\right) = F _ {t \\rightarrow t + 1} \\left(x _ {t}\\right), \\\\ F _ {0 \\rightarrow t} (x _ {0}) = F _ {0 \\rightarrow t - 1} (x _ {0}) + M \\left(x _ {0} + F _ {0 \\rightarrow t - 1} (x _ {0})\\right), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 806, + 890, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x_0$ is the source pixel position at time 0, $F_{0\\rightarrow t}$ is the displacement field that defines the displacement of the pixel position $x_0$ to the pixel position $x_t$ at time $t$ .", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7874", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4a67e97121df12534c46aef569f41f801346f128f3bc0015d37c5f820bb34afe.jpg", + "image_caption": [ + "Motion Generation" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 524, + 311 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/db2dd263e1cfeda263e26d7f00f69f00cf198b488f3b2ce313b548f3c3422577.jpg", + "image_caption": [ + "Cinemagraph Generation", + "Figure 2. Overview of StyleCineGAN. Given an input landscape image $I$ , our goal is to generate a cienemagraph using a fixed pre-trained StyleGAN $G$ . We project the image into both latent codes $w^{+}$ and deep features $D^{10}$ of $G$ . Using the deep features $D^{*}$ , a mask predictor predicts a segmentation mask $S$ . To animate the input image, we use a motion generator to predict the motion field $M$ from $I$ . $M$ is refined using $S$ . Through Euler integration, $M$ produces the future and past displacement fields $F_{0\\rightarrow t}$ and $F_{N\\rightarrow t}$ . To synthesize cienemagraph frames, we add a DFW layer in between the layers of $G$ . DFW refers to Eqns. 2 and 3. This modification enables the intermediate features of $G$ to be warped according to $F_{0\\rightarrow t}$ and $F_{N\\rightarrow t}$ using a joint splatting method at different resolutions, specifically for the StyleGAN layers indexed with $i\\in [10,12,14,16,18]$ . The warped deep features are used to synthesize frames $\\hat{I}_t$ resulting in the final cienemagraph video." + ], + "image_footnote": [], + "bbox": [ + 531, + 89, + 890, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Cinemagraph Generation", + "text_level": 1, + "bbox": [ + 76, + 452, + 310, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After acquiring the displacement field $F_{0\\rightarrow t}$ (Sec. 3.3), latent codes $w^{+}$ , and deep features $D^{10}$ of StyleGAN (Sec. 3.1), we feed these data into the pre-trained StyleGAN with our DFW layer added. We apply forward warping (or splatting) to the original content to generate a cinematograph. Here, the application of forward warping directly to an RGB image often results in images with tearing artifacts. To reduce the artifacts, we warp the deep features $D_0^i$ of StyleGAN in different resolutions or scales:", + "bbox": [ + 75, + 476, + 468, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {t} ^ {i} = \\operatorname {W a r p} \\left(D _ {0} ^ {i}, F _ {0 \\rightarrow t}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 617, + 468, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\text{Warp}$ is the forward warping function and $D_{t}^{i}$ is the warped deep feature at time $t$ and scale $i$ . We observed that warping a single deep feature (e.g., only $D_{0}^{10}$ ) results in blurry textures in the generated cinemagrams. Therefore, we opt to warp the multi-scale deep features; we call this operation a MSDFW. Specifically, we apply warping to the deep features $D_{0}^{i}$ where $i \\in [10, 12, 14, 16, 18]$ . The deep feature $D_{0}^{10}$ is acquired using GAN inversion, and the consequent deep features $D_{0}^{12}, D_{0}^{14}, \\ldots$ , and $D_{0}^{18}$ are subsequently generated using both $D_{0}^{10}$ and $w^{+}$ , as shown in the rightmost column of Figure 2.", + "bbox": [ + 75, + 643, + 468, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Looping video $V$ with frame length $N + 1$ is synthesized using the joint splating technique [9] using the future and past displacement fields $F_{0\\rightarrow t}$ and $F_{N\\rightarrow t}$ , generated through Euler integration. The displacement fields are computed with $t$ times of integration on $M$ for $F_{0\\rightarrow t}$ , and $N - t$ times of integration on $-M$ for $F_{N\\rightarrow t}$ . The multi-scale fea", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ture $D_0^i$ is warped in both directions and is composited to form $D_t^i$ . Specifically, $D_t^i$ is computed as a weighted sum of $Warp(D_0^i,F_{0\\to t})$ and $Warp(D_0^i,F_{N\\to t})$ as follows:", + "bbox": [ + 498, + 452, + 890, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} D _ {t} ^ {i} \\left(x ^ {\\prime}\\right) = \\frac {\\sum_ {x \\in \\chi} \\alpha_ {t} \\cdot W a r p \\left(D _ {0} ^ {i} (x) , F _ {0 \\rightarrow t}\\right)}{\\sum_ {x \\in \\chi} \\alpha_ {t}} \\\\ + \\frac {\\sum_ {x \\in \\chi} \\left(1 - \\alpha_ {t}\\right) \\cdot W a r p \\left(D _ {0} ^ {i} (x) , F _ {N \\rightarrow t}\\right)}{\\sum_ {x \\in \\chi} \\left(1 - \\alpha_ {t}\\right)}, \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 513, + 890, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x \\in \\chi$ is a set of pixels being mapped to the same destination pixel $x'$ , and $\\alpha_{t}$ is the looping weight defined as $(1 - \\frac{t}{N})$ .", + "bbox": [ + 498, + 602, + 890, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the above DFW module, we can generate a video frame given the predicted deep features $D^{10}$ , latent code $w^{+}$ , and mask $S$ :", + "bbox": [ + 498, + 648, + 890, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {I} _ {t} = S \\odot \\left(G _ {\\text {w a r p}} \\left(D _ {0} ^ {1 0}, w ^ {+}, F _ {0 \\rightarrow t}, F _ {N \\rightarrow t}\\right)\\right) \\\\ + (\\mathbf {1} - S) \\odot I, \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 700, + 890, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ is an element-wise multiplication and $G_{\\text{warp}}$ is a fixed pre-trained StyleGAN that incorporates our DFW module.", + "bbox": [ + 498, + 750, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Style Interpolation In addition to motion generation, the change of appearance style is an additional feature often observed in a landscape cinematograph. This can be achieved by interpolating the latent code with the target latent code as follows:", + "bbox": [ + 496, + 797, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {s} ^ {+} = w ^ {+} \\cdot (1 - \\beta) + w _ {t} ^ {+} \\cdot \\beta , \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 883, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7875", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{s}^{+}$ is the interpolated latent code, $w_{t}^{+}$ is the latent code of the target style, and $\\beta$ is the interpolation weight. This is possible because the later layers of the StyleGAN only modify the color of the synthesized image while the original structure is preserved.", + "bbox": [ + 75, + 236, + 468, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With modified Equation 4, we prevent the visual mismatch between the static and dynamic regions by also reflecting the changes to the static regions as follows:", + "bbox": [ + 76, + 313, + 468, + 359 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {I} _ {t} = S \\odot \\left(G _ {\\text {w a r p}} \\left(D _ {0} ^ {1 0}, w _ {s} ^ {+}, F _ {0 \\rightarrow t}, F _ {N \\rightarrow t}\\right)\\right) \\tag {6} \\\\ + (\\mathbf {1} - S) \\odot (I + \\Delta I _ {s}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 371, + 468, + 409 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta I _ {s} = G \\left(D _ {0} ^ {1 0}, w _ {s} ^ {+}\\right) - G \\left(D _ {0} ^ {1 0}, w ^ {+}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 411, + 468, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta I_{s}$ is the color difference between the two images generated from latent codes $w_{s}^{+}$ and $w^{+}$ .", + "bbox": [ + 76, + 443, + 468, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 489, + 209, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we compare our method with state-of-the-art landscape cinematograph generation methods [6, 9, 17, 19] (Sec. 4.1) and with an unconditional video generation method [34] (Sec. 4.2). In addition, we show the importance of the components used in our method through ablation studies (Sec. 4.3). To observe more various results, we recommend readers see the supplementary video. Example results are presented in Figures 1 and 3.", + "bbox": [ + 75, + 516, + 468, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Comparisons with Cinemagraph Generation Methods", + "text_level": 1, + "bbox": [ + 76, + 648, + 468, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compared our method with state-of-the-art cinemagraph generation methods: Animating Landscape (AL) [6], Deep Landscape (DL) [17], Eulerian Motion Field (EMF) [9], and Text2Cinemagraph (T2C) [19]. AL is a learning-based method that uses backward warping to generate cinemagraph frames. DL trains a style-based generator to synthesize images and dynamic motion in a scene. EMF and T2C train an encoder-decoder architecture paired with a motion generator to produce looping cinemagrams. We used official implementations of AL, DL, and T2C, and faithfully reproduced EMF based on the provided training details and hyper-parameters. In the following, we will examine the qualitative difference, evaluate the results using two metrics, and report the results of human perceptual studies.", + "bbox": [ + 75, + 689, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0969637ceb9ca37a33559ba2e6c9207ff7047175487bea4ca60a64b79dfa815f.jpg", + "image_caption": [ + "Figure 3. Generated cinematagraph results. This figure contains video clips, thus consider viewing it using Adobe Reader. The first two are cinematographs without appearance change, and the last two are cinematographs with appearance change." + ], + "image_footnote": [], + "bbox": [ + 521, + 90, + 869, + 169 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/39a0213e700ef9bc56ba9bc73a6eeddde37538bac4a9f16f4257e2c9eadcc04f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 171, + 867, + 251 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2b2dff4bbec86926ff16f3a9189aec4cf3d1734c21855e584c6339d7e0200f38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 253, + 867, + 333 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fd20f8b3e221aecb6ce65ca529f33fe1dbf2ad802f07ce17154bbfe3efe03578.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison with state-of-the-art cinematograph generation methods. Please refer to the supplementary video for more examples." + ], + "image_footnote": [], + "bbox": [ + 522, + 335, + 867, + 415 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/491143a31de31d72610b9141bf036fd9cfeed33cf9e9ae62ad2c04a6ea78a207.jpg", + "table_caption": [ + "Table 1. Quantitative comparison with the state-of-the-art cinematograph generation methods. We compared our method with AL, EMF, DL, and T2C. The best scores are bolded." + ], + "table_footnote": [], + "table_body": "
MethodStatic ConsistencyMotion Quality
LPIPS↓MS-SSIM↑RMSE↓LPIPS↓MS-SSIM↑RMSE↓FID↓
AL0.04770.95248.89560.06170.681925.108953.6893
EMF0.01030.97236.04400.05330.710222.893245.8035
DL0.00710.99312.10440.05040.706221.801141.6374
T2C0.00630.97734.87850.01590.722421.278440.1186
Ours0.00620.99621.94300.01310.723720.894839.2113
", + "bbox": [ + 509, + 523, + 880, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative Comparisons Figure 4 shows visual results from AL, EMF, DL, T2C, and our method. AL exhibits stretching artifacts (row 1) due to recurrent backward warping, whereas our method uses forward warping that prevents the appearance of such distortions. EMF struggles with precise reconstruction (row 2), because its encoder-decoder architecture is trained from scratch on a low-quality video dataset. In contrast, our method leverages a StyleGAN model pre-trained on high-quality images, resulting in cinemagrams with high perceptual quality. DL often loses fine details in textures (row 3) as it performs GAN inversion using the latent codes of StyleGAN, and fine-tunes the generator. Our method, utilizing deep features for GAN inversion, retains original texture details. T2C encounters difficulties in preserving high-frequency details, and accurate motion segmentation between static and dynamic regions (row 4). In contrast, our method utilizes a pre-trained StyleGAN for cinematograph generation, and uses its features for mask prediction, preserving both fine and structural details of the source image.", + "bbox": [ + 496, + 606, + 890, + 909 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7876", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Comparisons The evaluation was performed considering two aspects: (1) static consistency and (2) motion quality. Static consistency measures the consistency over time of non-movable regions such as buildings and mountains. Motion quality refers to both the image quality of the animated regions and the animation plausibility. For a fair comparison, we normalized the speed of motion in the generated results for all methods to match the average speed in the test dataset. AL, EMF, DL, and T2C were provided with the first frame $I_0$ of the test video to generate $512 \\times 512$ videos. Our method generates $1024 \\times 1024$ results, thus we downsampled it to the size of $512 \\times 512$ .", + "bbox": [ + 75, + 90, + 468, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a quantitative evaluation, we used 224 test videos from the Sky Time-Lapse dataset [41]. To measure the static consistency, we computed LPIPS [45], Root-Mean-Squared Error (RMSE), and MS-SSIM [40] between generated frame $\\hat{I}_n$ and the input image $I_0$ with the sky masked out. To measure the motion quality, we used the same evaluation metrics along with Frechet inception distance (FID) [28] between $\\hat{I}_n$ and the ground-truth frame $I_n$ with the static parts masked out.", + "bbox": [ + 75, + 272, + 467, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 reveals the quantitative evaluation results. For both static consistency and motion quality, our method outperforms the other approaches. Our cinematographs represent static consistency better because use of the mask improves the accuracy in the detection of static regions. In addition, the quality of texture details are improved in the generated frames due to GAN inversion and the MSDFW based on the deep features of StyleGAN. The use of our motion generator and DFW leads to improved motion quality for each scene, compared with that of the results of previous approaches.", + "bbox": [ + 75, + 407, + 467, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "User Study We conducted a user study with 17 participants to subjectively evaluate our method in comparison with previous cinematograph generation methods. We did not target any specific demographic groups, as the evaluation of cinematographs should not be dependent on particular demographic distributions. We conducted two evaluations: score evaluation and side-by-side comparison. Both evaluations assessed static consistency and motion quality. In the score evaluation, participants rated the video presented with a reference image on a 1-to-5 scale, with \"very bad\" being 1 and \"very good\" being 5. In the side-by-side comparison, participants selected the preferred cinematograph from two videos generated using different methods, presented with a reference image. For both evaluations, we used ten samples randomly chosen from the LHQ [31] dataset. To eliminate bias, distinct samples were used for each evaluation, and the positions of the cinematographs were randomized in all questions.", + "bbox": [ + 75, + 582, + 467, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tables 2 and 3 summarize the statistics of the user studies. For both score evaluation and human preference, our method outperforms the previous approaches by a substan", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f8057d5bc126095eab08959849b67219fc35cf69206b4c6e6da21bc3815500da.jpg", + "table_caption": [ + "Table 2. Human perceptual study results for score evaluation. The best scores are bolded." + ], + "table_footnote": [], + "table_body": "
MethodStatic ConsistencyMotion Quality
AL1.59 ± 0.201.76 ± 0.45
EMF2.35 ± 0.892.24 ± 0.99
DL3.35 ± 0.833.25 ± 0.67
T2C3.75 ± 0.692.96 ± 0.66
Ours4.37 ± 0.183.86 ± 0.53
", + "bbox": [ + 560, + 119, + 831, + 202 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/86e87569558fbeb09657d161c9158875cc390d45ad0ca0494cf9e3109fc09b26.jpg", + "table_caption": [ + "Table 3. Human perceptual study results for side-by-side comparison. The percentage of ours chosen against each competing method is reported." + ], + "table_footnote": [], + "table_body": "
MethodHuman Preference (Ours %)
Static ConsistencyMotion Quality
vs AL99.35%92.81%
vs EMF96.08%86.93%
vs DL94.77%84.97%
vs T2C77.78%73.86%
", + "bbox": [ + 557, + 261, + 828, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tial margin. We also performed statistical analysis (Kruskal-Wallis H-test) on the resulting evaluation scores. The results showed that our method achieved significantly higher scores in all pairs of comparisons with every previous method $(p < 0.001$ in post-hoc analysis). Both user study results reveal the advantages of our method for generating cinematographs with high image quality and plausible animation.", + "bbox": [ + 496, + 354, + 890, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with Video Generation Methods", + "text_level": 1, + "bbox": [ + 498, + 472, + 885, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compared our approach with MoCoGAN-HD [34], a video generation method that predicts the trajectory within the latent space of a pre-trained StyleGAN. For content generation, we used the same pre-trained StyleGAN as that used by MoCoGAN-HD, which was trained on the Sky Time-lapse [41] dataset to generate $128 \\times 128$ images. The comparison was made in an unconditional manner using 300 randomly sampled latent codes, at a resolution of $128 \\times 128$ . We compared the first 60 frames of the generated videos with those of MoCoGAN-HD, as our method produces looping videos.", + "bbox": [ + 496, + 496, + 890, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparisons Figure 5 presents a qualitative comparison. As shown in the first and third rows, the videos generated using our method exhibit static consistencies, with the clouds moving while the ground remains static. In contrast, as shown in the second and fourth rows, the frames generated by MoCoGAN-HD deviate significantly from the original image, making the method unsuitable for cinematograph generation.", + "bbox": [ + 496, + 667, + 890, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparisons We compared the content preservation ability of both methods by measuring LPIPS [45], RMSE, and MS-SSIM [40] between the first frame $\\hat{I}_0$ and the subsequent frames $\\hat{I}_n$ . Because the main content to be preserved in kinemagrams is the static regions, we defined content preservation as static consistency. We masked out the sky for all image frames according to the", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7877", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6019b077d7cbd2e52e32a8f09f5c75e1aed65c3732c400e006c828ad8c966ef1.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison with the state-of-the-art video generation method, MoCoGAN-HD [34]. For more examples from this comparison, please refer to the supplementary video." + ], + "image_footnote": [], + "bbox": [ + 106, + 87, + 436, + 347 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a794bf00fcc35105d17098df0a2401efecfbcc2e2b8133156a66ab6a93226d33.jpg", + "image_caption": [ + "Figure 6. Quantitative comparison of content preservation with state-of-the-art video generation method MoCoGAN-HD [34]." + ], + "image_footnote": [], + "bbox": [ + 96, + 401, + 447, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "segmentation mask and compared only the static parts. We used 219 videos in which the scene could be horizontally divided into static and animated regions. LPIPS, RMSE, and MS-SSIM were all computed for each pair of $\\hat{I}_0$ and $\\hat{I}_n$ , and were averaged over the number of samples. Figure 6 shows that MoCoGAN-HD exhibits significant divergence in terms of LPIPS, RMSE, and MS-SSIM over time. This indicates that MoCoGAN-HD is unable to generate motion that preserves the content of the original image. In contrast, our method exhibits a small diverging trend, confirming its superiority in preserving the content of the original image over time, by utilizing deep features.", + "bbox": [ + 75, + 532, + 470, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 728, + 230, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our design choices for the proposed method, we conducted a series of ablation studies. Specifically, we focused on the effectiveness of our warping and GAN inversion methods. To evaluate warping, we compared the image frames generated with and without the use of 1) forward warping, 2) DFW, 3) MSDFW, and 4) segmentation mask. The first frames of 224 test videos from the Sky Time-Lapse [41] dataset were given as input to generate $1024 \\times 1024$ videos, and we used the first 60 frames for the evaluation. For the evaluation of GAN inversion, we", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb5b3cdaa79bcc779d32b3d97b9387f720c1d68cd27af3559c9ae022d5cdd0ab.jpg", + "image_caption": [ + "Figure 7. Results of qualitative comparison in ablation study. Please see the supplementary video for the animated results." + ], + "image_footnote": [], + "bbox": [ + 524, + 88, + 692, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ef88055d4ef5ca0a0e9822c91ff7031dd6c9f05db397b92441a12a59e9fd8f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 89, + 864, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7f42b19cf715c06c7d4e413353f86ecd73d80170520fac7dc3477bb3184a3f1.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison results on GAN inversion." + ], + "image_footnote": [], + "bbox": [ + 527, + 362, + 859, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "compared images reconstructed with and without the use of the deep features. A total of 256 images, 128 from the Sky Time-Lapse and 128 from the Eulerian [9] dataset were provided to generate $1024 \\times 1024$ reconstructed images.", + "bbox": [ + 496, + 500, + 890, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Forward Warping We compared the image frames generated with and without forward warping (FW). For the case without FW, we warped the deep features of StyleGAN using backward warping. The first row in Figure 7 shows the qualitative results of this comparison. As revealed in the figure, without FW, the results usually contain unrealistically stretched textures. For quantitative comparisons, we computed LPIPS, MS-SSIM, and RMSE between $I_{n}$ and $\\hat{I}_n$ . Table 4 shows that excluding forward warping resulted in significant degradation in terms of MS-SSIM. This indicates that the stretched textures resulted in huge structural distortions in the generated images.", + "bbox": [ + 496, + 571, + 892, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Deep Feature Warping We then compared frames generated with and without DFW. For the case without DFW, we directly warped RGB images using predicted motion fields for comparison. The second row in Figure 7 presents the qualitative comparison. As shown in the figure, the frame generated without DFW contained tearing artifacts in the animated regions. Table 4 shows an increase in the LPIPS score, which indicates that excluding DFW degraded the perceptual quality by introducing tearing artifacts.", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7878", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0f1926ecc6f88fc688d361c20efbe8fa56288085c64664b12f1383dad8e827db.jpg", + "table_caption": [ + "Table 4. Results of quantitative evaluation in ablation study. The best scores are bolded." + ], + "table_footnote": [], + "table_body": "
MethodLPIPS↓MS-SSIM↑RMSE↓
Ours - Full0.05110.716521.8188
Ours - w/o FW0.05240.685322.9816
Ours - w/o DFW0.05370.690822.4993
Ours - w/o MSDFW0.06290.694622.3227
Ours - w/o Mask0.05640.698022.6213
", + "bbox": [ + 117, + 121, + 429, + 204 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1bb14de5cb3824060f668bd40cc7402e33027587a02326286e5d4f0fb2c41a93.jpg", + "table_caption": [ + "Table 5. Quantitative evaluation results on GAN inversion. The best scores are bolded." + ], + "table_footnote": [], + "table_body": "
w+D10FID↓LPIPS↓RMSE↓MS-SSIM↑
✓-40.82760.015318.59010.7891
✓✓11.82310.00195.64100.9907
", + "bbox": [ + 114, + 243, + 426, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-scale Deep Feature Warping Image frames generated with and without MSDFW were also compared. For the case without MSDFW, only a single deep feature $D^{10}$ was warped and propagated through the next blocks of StyleGAN. The results of the qualitative comparison are shown in the third row of Figure 7. As revealed in the figure, excluding MSDFW resulted in blurry textures in the dynamic region. The quantitative comparison reported in Table 4 shows a significant increase in the LPIPS score. This indicates that excluding MSDFW degraded the perceptual quality, especially the texture details.", + "bbox": [ + 75, + 295, + 467, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Segmentation Mask We compared the image frames generated with and without a segmentation mask. For the case without a mask, we used the initially predicted motion field for warping. The fourth row in Figure 7 shows the qualitative result of this comparison. As shown in the figure, excluding the mask resulted in erroneous movement in the static regions. The quantitative comparison reported in Table 4 reveals a significant degradation of MS-SSIM, which indicates the structural distortion caused by the erroneous motion.", + "bbox": [ + 75, + 465, + 467, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Deep Feature Inversion To evaluate the effectiveness of GAN inversion, we compared the images reconstructed with and without the use of the deep features $D^{10}$ . For the case without deep feature inversion, only the latent codes $w^{+}$ of a pre-trained StyleGAN were used to reconstruct the input landscape image. Figure 8 shows a qualitative result of this comparison. As revealed in the figure, using only $w^{+}$ failed to accurately reconstruct the details of the original images. For quantitative comparisons, FID, LPIPS, RMSE, and MS-SSIM were computed between $I$ and $\\hat{I}$ . Table 5 reveals significant improvements in the perceptual quality of the results when reconstructed using the deep features $D^{10}$ .", + "bbox": [ + 75, + 621, + 467, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 76, + 818, + 349, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our mask predictor and motion generator generally performed well for landscape images. However, automatic prediction cannot be accurate for all images because most landscape images contain inherent ambiguity in their motion", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0a3178b499e93c16913a8a8349e6d82741ef8a510389777e8ab9dc7d67285d5f.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 535, + 90, + 692, + 151 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/99266baccdda3e0ba47175c59d0df00a990c5ee099f3af6ed316f18498209f98.jpg", + "image_caption": [ + "(b)", + "Figure 9. Limitations of StyleCineGAN: (a) automatic prediction of motion cannot be accurate for all images, and (b) the motion of a very thin structured object is hard to be isolated." + ], + "image_footnote": [], + "bbox": [ + 718, + 90, + 795, + 150 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/66837910d83aaf923e84f8e453c57f1e4b218fed97a2f2fc7b454e387d80de0e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 90, + 872, + 150 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "directions except for some obvious cases (e.g., waterfalls). The failure case is illustrated in Figure 9 (a). User-defined motion hints can be used to resolve such ambiguities and provide further control capability during the generation process [18, 19]. In addition, it is hard for our method to isolate the motion of a very thin structured object placed within the animated region as shown in Figure 9 (b). This is because our method performs warping of features at multiple resolutions, in which low-resolution features cannot spatially represent the thin structures.", + "bbox": [ + 496, + 205, + 890, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we mainly focused on animating the landscape images, particularly skies and fluids, while putting other types of animations outside the scope. In future developments, we would like to expand the capabilities to include other forms of motion. Investigating the rotating hands of a clock, the playing arm of a guitarist moving up and down, and a flag or the wings of a bird fluttering for cinematograph generation would be a very interesting direction to pursue.", + "bbox": [ + 496, + 356, + 890, + 478 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 489, + 617, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed the first approach that leverages a pre-trained StyleGAN for high-quality one-shot landscape cinematograph generation. In contrast to previous studies, our method does not require training a large image generator from scratch, and also systematically improves the resolution of the generated cinematographs to $1024 \\times 1024$ . At the core of our method, we utilized the deep features of a pretrained StyleGAN, because those features can help preserve spatial information and encode both high-level semantic and low-level style appearances. Using our MSDFW approach, we applied the predicted motion to the deep feature space of StyleGAN. Both qualitative and quantitative results confirm that our method substantially outperforms existing baselines.", + "bbox": [ + 496, + 512, + 890, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This research was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023 (Project Name: Development of Universal Fashion Creation Platform Technology for Avatar Personality Expression, Project Number: RS-2023-00228331, Contribution Rate: $90\\%$ ) and Korea Institute for Advancement of Technology(KIAT) grant funded by the Korea Government(MOTIE) (Project Name: The Competency Development Program for Industry Specialist, Project Number: P0012746, Contribution Rate: $10\\%$ ).", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7879", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 3", + "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? CoRR, abs/1911.11544, 2019.3", + "[3] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3. arXiv preprint arXiv:2201.13433, 2022. 3", + "[4] Hugo Bertiche, Niloy J. Mitra, Kuldeep Kulkarni, ChunHao Paul Huang, Tuanfeng Y. Wang, Meysam Madadi, Sergio Escalera, and Duygu Ceylan. Blowing in the wind: Cyclenet for human cinematographs from still images, 2023. 1, 2", + "[5] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. In ACM SIGGRAPH 2005 Papers, pages 853-860. 2005. 2", + "[6] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. ACM Transactions on Graphics (Proc. of SIGGRAPH ASIA 2019), 38(6):175:1-175:19, 2019. 1, 2, 5", + "[7] Gereon Fox, Ayush Tewari, Mohamed Elgharib, and Christian Theobalt. Stylevideogan: A temporal generative model using a pretrained stylegan, 2021. 2", + "[8] Tavi Halperin, Hanit Hakim, Orestis Vantzos, Gershon Hochman, Netai Benaim, Lior Sassy, Michael Kupchik, Ofir Bibi, and Ohad Fried. Endless loops: detecting and animating periodic patterns in still images. ACM Transactions on Graphics (TOG), 40(4):1-12, 2021. 2", + "[9] Aleksander Holynski, Brian L. Curless, Steven M. Seitz, and Richard Szeliski. Animating pictures with eulerian motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5810-5819, 2021. 1, 2, 3, 4, 5, 7", + "[10] Ali Jahanian, Lucy Chai, and Phillip Isola. \"On the \"steerability\" of generative adversarial networks. In International Conference on Learning Representations, 2020. 2", + "[11] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. CoRR, abs/1710.10196, 2017. 3", + "[12] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2", + "[13] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 3", + "[14] Alexander Kristoffersen. Loopnerf: Exploring temporal compression for 3d video textures. Master's thesis, EECS Department, University of California, Berkeley, 2023. 1, 2", + "[15] Xingyi Li, Zhiguo Cao, Huiqiang Sun, Jianming Zhang, Ke Xian, and Guosheng Lin. 3d cinematography from a single" + ], + "bbox": [ + 78, + 114, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4595-4605, 2023. 1, 2", + "[16] Ji Lin, Richard Zhang, Frieder Ganz, Song Han, and JunYan Zhu. Anycost gans for interactive image synthesis and editing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[17] Elizaveta Logacheva, Roman Suvorov, Oleg Khomenko, Anton Mashikhin, and Victor Lempitsky. Deeplandscape: Adversarial modeling of landscape videos. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 5", + "[18] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 8", + "[19] Aniruddha Mahapatra, Aliaksandr Siarohin, Hsin-Ying Lee, Sergey Tulyakov, and Jun-Yan Zhu. Text-guided synthesis of eulerian cinemagrams. 2023. 1, 2, 3, 5, 8", + "[20] Makoto Okabe, Ken Anjyo, Takeo Igarashi, and Hans-Peter Seidel. Animating pictures of fluid using video examples. In Computer Graphics Forum, pages 677-686. Wiley Online Library, 2009. 2", + "[21] Makoto Okabe, Ken Anjyor, and Rikio Onai. Creating fluid animation from a single image using video database. In Computer Graphics Forum, pages 1973-1982. Wiley Online Library, 2011.", + "[22] Makoto Okabe, Yoshinori Dobashi, and Ken Anjyo. Animating pictures of water scenes using video retrieval. The Visual Computer, 34(3):347-358, 2018. 2", + "[23] Ehsan Pajouheshgar, Yitao Xu, Tong Zhang, and Sabine Süssstrunk. Dynca: Real-time dynamic texture synthesis using neural cellular automata. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1", + "[24] Gaurav Parmar, Yijun Li, Jingwan Lu, Richard Zhang, Jun-Yan Zhu, and Krishna Kumar Singh. Spatially-adaptive multilayer selection for gan inversion and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3", + "[25] Ekta Prashnani, Maneli Noorkami, Daniel Vaquero, and Pradeep Sen. A phase-based approach for animating images using video examples. In Computer Graphics Forum, pages 303-311. Wiley Online Library, 2017. 1, 2", + "[26] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[27] Masaki Saito, Shunta Saito, Masanori Koyama, and Sosuke Kobayashi. Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal gan, 2020. 2", + "[28] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.6" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7880", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Kwanggyoon Seo, Seoung Wug Oh, Jingwan Lu, Joon-Young Lee, Seonghyeon Kim, and Junyong Noh. Styleportraitvideo: Editing portrait videos with expression optimization. 41(7), 2022. 3", + "[30] Claude E. Shannon. Coding Theorems for a Discrete Source With a Fidelity Criterion Institute of Radio Engineers, International Convention Record, vol. 7, 1959., pages 325-350. 1993. 2", + "[31] Ivan Skorokhodov, Grigorii Sotnikov, and Mohamed Elhoseiny. Aligning latent and image spaces to connect the unconnectable. arXiv preprint arXiv:2104.06954, 2021. 6", + "[32] Ryusuke Sugimoto, Mingming He, Jing Liao, and Pedro V. Sander. Water simulation and rendering from a still photograph. In SIGGRAPH Asia 2022 Conference Papers, New York, NY, USA, 2022. Association for Computing Machinery. 2", + "[33] Matthew Tesfaldet, Marcus A. Brubaker, and Konstantinos G. Derpanis. Two-stream convolutional networks for dynamic texture synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1", + "[34] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. In International Conference on Learning Representations, 2021. 2, 5, 6, 7", + "[35] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4): 1-14, 2021. 3", + "[36] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1526-1535, 2018. 2", + "[37] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. Stitch it in time: Gan-based facial editing of real videos. In SIGGRAPH Asia 2022 Conference Papers, New York, NY, USA, 2022. Association for Computing Machinery. 3", + "[38] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[39] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 3", + "[40] Zhou Wang, Eero P. Simoncelli, and Alan Conrad Bovik. Multiscale structural similarity for image quality assessment. The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, 2:1398-1402 Vol.2, 2003. 6", + "[41] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6, 7" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally consistent semantic video editing. pages 357-374. Springer, 2022. 3", + "[43] Xu Yao, Alasdair Newson, Yann Gousseau, and Pierre Hellier. A style-based gan encoder for high fidelity reconstruction of images and videos. European conference on computer vision, 2022. 3", + "[44] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 85-101. Springer, 2022. 3", + "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6", + "[46] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10145-10155, 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 415 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7881", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/51ff4f8c-4976-4d05-85df-cf04bbc75017_model.json b/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/51ff4f8c-4976-4d05-85df-cf04bbc75017_model.json new file mode 100644 index 0000000000000000000000000000000000000000..658d8baff23b92a3abe1533a1a914c8d0a17ab40 --- /dev/null +++ b/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/51ff4f8c-4976-4d05-85df-cf04bbc75017_model.json @@ -0,0 +1,2200 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.131, + 0.756, + 0.177 + ], + "angle": 0, + "content": "StyleCineGAN: Landscape Cinemagraph Generation using a Pre-trained StyleGAN" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.204, + 0.28, + 0.222 + ], + "angle": 0, + "content": "Jongwoo Choi" + }, + { + "type": "text", + "bbox": [ + 0.319, + 0.204, + 0.463, + 0.222 + ], + "angle": 0, + "content": "Kwanggyoon Seo" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.204, + 0.661, + 0.222 + ], + "angle": 0, + "content": "Amirsaman Ashtari" + }, + { + "type": "text", + "bbox": [ + 0.7, + 0.204, + 0.808, + 0.222 + ], + "angle": 0, + "content": "Junyong Noh" + }, + { + "type": "text", + "bbox": [ + 0.38, + 0.223, + 0.591, + 0.239 + ], + "angle": 0, + "content": "Visual Media Lab, KAIST" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.242, + 0.73, + 0.258 + ], + "angle": 0, + "content": "{jeolpyeoni0,skg1023,a.s.ashtari,junyongnoh}@kaist.ac.kr" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.467, + 0.895, + 0.497 + ], + "angle": 0, + "content": "Figure 1. Given a landscape image, StyleCineGAN generates a seamless cinematograph at \\(1024 \\times 1024\\) resolution. This figure contains video clips, thus consider viewing it using Adobe Reader. The same results are also included in the supplementary video." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.509, + 0.314, + 0.525 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.473, + 0.786 + ], + "angle": 0, + "content": "We propose a method that can generate cinematographs automatically from a still landscape image using a pretrained StyleGAN. Inspired by the success of recent unconditional video generation, we leverage a powerful pretrained image generator to synthesize high-quality cinematographs. Unlike previous approaches that mainly utilize the latent space of a pre-trained StyleGAN, our approach utilizes its deep feature space for both GAN inversion and cinematograph generation. Specifically, we propose multi-scale deep feature warping (MSDFW), which warps the intermediate features of a pre-trained StyleGAN at different resolutions. By using MSDFW, the generated cinematographs are of high resolution and exhibit plausible looping animation. We demonstrate the superiority of our method through user studies and quantitative comparisons with state-of-the-art cinematograph generation methods and a video generation method that uses a pre-trained StyleGAN." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.802, + 0.21, + 0.818 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Cinemagraph is a unique form of media that combines a still image and video. While most of the scene remains still, subtle and repeated movements in only a small region effectively highlight a particular object or an important event that occurred at the time of capture. Because this mixture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.892, + 0.556 + ], + "angle": 0, + "content": "of still and moving elements in one scene creates an interesting eye-catching effect, cinematographs have recently attracted increasing attention in social media." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.893, + 0.699 + ], + "angle": 0, + "content": "Despite the growing popularity of cinematograph, its creation heavily relies on a manual process. To create a motion in part of the input image, one typically utilizes an image manipulation tool to stretch, scale, rotate, or shift the part of the image in a physically plausible and aesthetically pleasing way. This process is often time-consuming and requires a high level of skill in the use of image editing tools. Therefore, creating a cinematograph has generally been considered to be a personal project of professionals so far." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.902 + ], + "angle": 0, + "content": "To allow ordinary users to create a cinematograph with a single image, automatic methods have been proposed. One line of research uses a reference video as guidance [14, 25, 33]. These methods are capable of producing realistic motions in the target scene similar to a reference video. Recent methods obviate the need for a reference video, by training deep generative models [4, 6, 9, 15, 17-19, 23]. These methods decompose the task into two processes of learning motion and spatial information individually from separate datasets. The decomposition effectively reduces the complexity of simultaneously learning both temporal and spatial information, and improves the perceptual quality of the generated videos. However, these models" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7872" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.167 + ], + "angle": 0, + "content": "must be trained from scratch, which sometimes requires several days if not weeks on a modern GPU. Moreover, these models are not specifically designed to generate high-resolution of \\(1024 \\times 1024\\) cinemagrams, because of the significant requirements in memory and processing power." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.167, + 0.468, + 0.317 + ], + "angle": 0, + "content": "In this paper, we propose the first approach to high-quality one-shot landscape cinematograph generation based on a pre-trained StyleGAN [12]. By using the learned image prior of StyleGAN, our method removes the need for training a large model from scratch and systematically improves the resolution of the generated cinematographs to \\(1024 \\times 1024\\). Moreover, our method enables a user to easily edit the style and appearance of the generated cinematograph by leveraging the properties of StyleGAN for image stylization and editing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.317, + 0.468, + 0.543 + ], + "angle": 0, + "content": "Our method is inspired by recent unconditional video generation methods [7, 34] which allow to navigate the latent space of a pre-trained image generator, to synthesize a high-quality temporally coherent video. Unlike these methods that utilize the latent codes of StyleGAN, we opt to use the deep features that are generated by convolution operations in each layer of StyleGAN. We use these deep features for two reasons. First, we observed that highly detailed landscape images cannot be reconstructed accurately from the latent codes using GAN inversion methods because these latent codes are low-dimensional [30]. Second, a plausible motion that preserves the content cannot be created by only navigating the latent space, because this space is highly semantic-condensed and lacks explicit spatial prior [38]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.468, + 0.678 + ], + "angle": 0, + "content": "To solve these problems, for the first time in the cinematograph generation domain, we utilize the deep features of a pre-trained StyleGAN. These deep features preserve spatial information and encode both high-level semantic and low-level style appearances across high and low-level convolutional layers. To produce cinematographs from these deep features, we propose a multi-scale deep feature warping (MSDFW) to apply the motions generated using a motion generator to the deep feature space of StyleGAN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.468, + 0.813 + ], + "angle": 0, + "content": "We demonstrate the effectiveness and advantages of our method by comparing it to state-of-the-art methods in cinematagraph generation [6, 9, 17, 19] and to an unconditional video generation method that uses a pre-trained StyleGAN [34], using various metrics. We also performed a user study to verify the effectiveness of our method for creating visually pleasing cinematagraph effects. Both qualitative and quantitative results confirm that our method substantially outperforms all existing baselines." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.825, + 0.219, + 0.84 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.85, + 0.311, + 0.866 + ], + "angle": 0, + "content": "2.1. Cinemagraph Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "One of the early studies for cinemagraph generation used a procedural approach to decompose each object into layers" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.272 + ], + "angle": 0, + "content": "and apply time-varying stochastic motion [5]. This method can handle a diverse range of images such as photos and paintings by relying on manual interaction from the user. Another approach is to use a reference video to animate the given image [20-22, 25]. These methods rely on a statistical motion analysis of videos to transfer their periodic motion property to the desired image. Another method distills a dynamic NeRF to render looping 3D video textures from the representative motion found in the reference video [14]. Without any reference video as guidance, Halperin et al. [8] proposed a framework to animate arbitrary objects with periodic patterns in a simple motion direction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.892, + 0.529 + ], + "angle": 0, + "content": "Recent approaches use the capacity of deep learning to automatically create cinemagrams from a single image. One study trained an image-based renderer to generate water animation, utilizing the water simulation results [32]. Another research first predicts a sequence of normal maps, then generates the corresponding RGB images that show a looping animation of a garment as if it is blown in the wind [4]. Another line of research trains a generator to produce a motion field [6, 9, 15, 18, 19]. Our work is similar to these deep learning-based approaches in that a motion generator is utilized to synthesize the cinemagrams. The difference is that our method leverages the deep features of a pre-trained image generator and thus can generate plausible high-quality results without fine-tuning or training a separate synthesis network. At the same time, our method systematically improves the resolution of the generated cinemagrams to \\(1024 \\times 1024\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.534, + 0.785, + 0.549 + ], + "angle": 0, + "content": "2.2. Unconditional Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The task of video generation is known to be difficult because the generation process has to take into account both spatial and temporal information. To ease the problem, previous approaches [27, 36] decompose video generation into content and motion generation. These methods first predict the latent trajectory for motion, then generate a video from the set of predicted latent codes using the image generator. Instead of training the generation model from scratch, MoCoGAN-HD [34] and StyleVideoGAN [7] leverage a pre-trained image generator model, StyleGAN [12]. These methods use the property of well-constructed latent space to morph one frame toward the next frames [10]. Utilizing the image generation capability of a pre-trained generator, MoCoGAN-HD and StyleVideoGAN only need to train a motion trajectory generator, which greatly reduces the training time. These methods can generate high-quality videos in multiple domains such as faces, cars, and even sky. However, the content details are not well preserved, which has hindered the methods' application for cinematograph generation. Our method builds upon a similar concept of using a pre-trained StyleGAN. We leverage the deep features instead of the latent codes of StyleGAN to generate looping and content-preserving cinematographs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7873" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.455, + 0.108 + ], + "angle": 0, + "content": "2.3. Video Synthesis using pre-trained StyleGAN" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.473, + 0.328 + ], + "angle": 0, + "content": "A pre-trained generative model, StyleGAN [11-13], has been actively employed for downstream image and video editing applications. For GAN-based applications, one of the essential steps is to project the desired images or videos to the latent space of the pre-trained GAN model, which is known as GAN inversion. The latent codes are obtained using either an optimization technique [1, 2], a trained encoder [26, 35], or a hybrid of both approaches [16]. For video editing applications, previous methods [3, 29, 37, 42] have utilized a pre-trained image encoder to project all of the frames to the latent space. While most of the methods operate in \\(\\mathcal{W}+\\) space [2], we opt to use both \\(w^{+}\\) and deep features [24, 43, 44] of the pre-trained StyleGAN to accurately reconstruct the original image and synthesize motion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.344, + 0.178, + 0.36 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.47, + 0.461 + ], + "angle": 0, + "content": "An overview of our method is shown in Figure 2. Given a landscape image \\(I\\), our method generates a seamlessly looping cinemagraph \\(V = \\{\\hat{I}_0, \\dots, \\hat{I}_N\\}\\) using a pre-trained StyleGAN. At the core of our method, we use a pre-trained generator without additional fine-tuning. The overall process is described below." + }, + { + "type": "text", + "bbox": [ + 0.061, + 0.466, + 0.47, + 0.526 + ], + "angle": 0, + "content": "(A) First, we project the landscape image into both the latent space and the feature space of StyleGAN. To this end, we train an encoder that outputs both latent codes \\( w^{+} \\) and intermediate features \\( D^{10} \\) of StyleGAN (Sec. 3.1)." + }, + { + "type": "text", + "bbox": [ + 0.061, + 0.527, + 0.469, + 0.556 + ], + "angle": 0, + "content": "(B) In addition, we predict a mask \\( S \\) to divide the image into static and dynamic regions (Sec. 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.062, + 0.557, + 0.469, + 0.616 + ], + "angle": 0, + "content": "(C) Next, we use a motion generator that accepts the landscape image \\(I\\) to synthesize a motion field \\(M\\), which defines the position of each pixel in the future frames (Sec. 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.061, + 0.618, + 0.469, + 0.676 + ], + "angle": 0, + "content": "(D) Lastly, we generate the final cinematograph frames using the pre-trained StyleGAN with deep feature warping (DFW) operation added in between the original layers. (Sec. 3.4)." + }, + { + "type": "list", + "bbox": [ + 0.061, + 0.466, + 0.47, + 0.676 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.713 + ], + "angle": 0, + "content": "In the following sections, we will describe the details of each process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.231, + 0.74 + ], + "angle": 0, + "content": "3.1. GAN Inversion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.902 + ], + "angle": 0, + "content": "The first step for generating a cinematograph is to project the input image to the latent space of a pre-trained StyleGAN. This process is necessary because StyleGAN is an unconditional generator that cannot take a conditioning input. Many previous methods have used \\(\\mathcal{W} +\\) [2] latent space, which is an extended space of native latent space \\(\\mathcal{W}\\). However, we observed that \\(w^{+} \\in \\mathcal{W}+\\) is not expressive enough to reconstruct the original high-frequency details of the landscape images. Therefore, we chose to use the generated deep features \\(D^{10}\\) of StyleGAN as well as \\(w^{+}\\). The use of \\(D^{10}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "enables us to recover details in the original input as shown in Figure 8." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.894, + 0.199 + ], + "angle": 0, + "content": "To project an image \\(I\\) to both the latent space and feature space and obtain \\(w^{+}\\) and \\(D^{10}\\), we train an encoder \\(E\\) similarly to Yao et al. [43], in which the encoder takes \\(I\\) as input and predicts \\((w^{+}, D^{10})\\). Additional training details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.209, + 0.665, + 0.223 + ], + "angle": 0, + "content": "3.2. Mask Prediction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.231, + 0.892, + 0.352 + ], + "angle": 0, + "content": "We improve the quality of GAN inversion described in Sec. 3.1 by training an additional classifier that predicts a mask to separate the static and dynamic regions. Using the mask, the structure of the static regions can be preserved. To this end, we train a multi-layer-perceptron (MLP) classifier that accepts the deep features of StyleGAN as its input and outputs the mask that specifies each dynamic region in the image, as performed in DatasetGAN [46]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.893, + 0.52 + ], + "angle": 0, + "content": "To train the classifier, we manually annotate 32 segmentation masks \\( S \\) of the selected images. We then follow the same procedure as DatasetGAN but using the deep features \\( D^{i} \\) where \\( i \\in \\{10,11,\\dots,18\\} \\). We resize all these deep features followed by concatenating them in the channel dimension to construct the input feature \\( D^{*} \\). In the end, paired data \\( (D^{*},S) \\) is constructed. An ensemble of 10 MLP classifiers is trained with a cross-entropy loss. To further improve the performance of the previous work, at inference, we also refine the predicted mask. For additional details, please refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.53, + 0.685, + 0.544 + ], + "angle": 0, + "content": "3.3. Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.674 + ], + "angle": 0, + "content": "To generate motion from a static landscape image \\( I \\), we use an Eulerian motion field \\( M \\) which represents the immediate velocity of each pixel. To predict \\( M \\) given \\( I \\), an imaged-to-image translation network [39] is trained as a motion generator. We train the network with paired data \\( (I, M) \\) as performed in the previous method [9]. Adding controls in motion generation is also possible, either by using text prompts [19] or drawing arrows [18]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.795 + ], + "angle": 0, + "content": "While the trained network works well to a certain degree, we observed that the predicted motion field \\(M\\) does not align with the boundaries between the static and dynamic regions and often contains motion in static regions. Thus, we further refine \\(M\\) by multiplying it with the segmentation mask \\(S\\), which effectively removes most errors. The refined \\(M\\) is then used to simulate the motion of a pixel at time \\(t\\) using the Euler integration as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.808, + 0.892, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x _ {t + 1} = x _ {t} + M \\left(x _ {t}\\right), \\quad M \\left(x _ {t}\\right) = F _ {t \\rightarrow t + 1} \\left(x _ {t}\\right), \\\\ F _ {0 \\rightarrow t} (x _ {0}) = F _ {0 \\rightarrow t - 1} (x _ {0}) + M \\left(x _ {0} + F _ {0 \\rightarrow t - 1} (x _ {0})\\right), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\( x_0 \\) is the source pixel position at time 0, \\( F_{0\\rightarrow t} \\) is the displacement field that defines the displacement of the pixel position \\( x_0 \\) to the pixel position \\( x_t \\) at time \\( t \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7874" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.526, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.316, + 0.387, + 0.326 + ], + "angle": 0, + "content": "Motion Generation" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.09, + 0.891, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.316, + 0.666, + 0.327 + ], + "angle": 0, + "content": "Cinemagraph Generation" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.33, + 0.892, + 0.44 + ], + "angle": 0, + "content": "Figure 2. Overview of StyleCineGAN. Given an input landscape image \\( I \\), our goal is to generate a cienemagraph using a fixed pre-trained StyleGAN \\( G \\). We project the image into both latent codes \\( w^{+} \\) and deep features \\( D^{10} \\) of \\( G \\). Using the deep features \\( D^{*} \\), a mask predictor predicts a segmentation mask \\( S \\). To animate the input image, we use a motion generator to predict the motion field \\( M \\) from \\( I \\). \\( M \\) is refined using \\( S \\). Through Euler integration, \\( M \\) produces the future and past displacement fields \\( F_{0\\rightarrow t} \\) and \\( F_{N\\rightarrow t} \\). To synthesize cienemagraph frames, we add a DFW layer in between the layers of \\( G \\). DFW refers to Eqns. 2 and 3. This modification enables the intermediate features of \\( G \\) to be warped according to \\( F_{0\\rightarrow t} \\) and \\( F_{N\\rightarrow t} \\) using a joint splatting method at different resolutions, specifically for the StyleGAN layers indexed with \\( i\\in [10,12,14,16,18] \\). The warped deep features are used to synthesize frames \\( \\hat{I}_t \\) resulting in the final cienemagraph video." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.453, + 0.312, + 0.469 + ], + "angle": 0, + "content": "3.4. Cinemagraph Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.47, + 0.611 + ], + "angle": 0, + "content": "After acquiring the displacement field \\( F_{0\\rightarrow t} \\) (Sec. 3.3), latent codes \\( w^{+} \\), and deep features \\( D^{10} \\) of StyleGAN (Sec. 3.1), we feed these data into the pre-trained StyleGAN with our DFW layer added. We apply forward warping (or splatting) to the original content to generate a cinematograph. Here, the application of forward warping directly to an RGB image often results in images with tearing artifacts. To reduce the artifacts, we warp the deep features \\( D_0^i \\) of StyleGAN in different resolutions or scales:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.618, + 0.469, + 0.637 + ], + "angle": 0, + "content": "\\[\nD _ {t} ^ {i} = \\operatorname {W a r p} \\left(D _ {0} ^ {i}, F _ {0 \\rightarrow t}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.469, + 0.81 + ], + "angle": 0, + "content": "where \\( \\text{Warp} \\) is the forward warping function and \\( D_{t}^{i} \\) is the warped deep feature at time \\( t \\) and scale \\( i \\). We observed that warping a single deep feature (e.g., only \\( D_{0}^{10} \\)) results in blurry textures in the generated cinemagrams. Therefore, we opt to warp the multi-scale deep features; we call this operation a MSDFW. Specifically, we apply warping to the deep features \\( D_{0}^{i} \\) where \\( i \\in [10, 12, 14, 16, 18] \\). The deep feature \\( D_{0}^{10} \\) is acquired using GAN inversion, and the consequent deep features \\( D_{0}^{12}, D_{0}^{14}, \\ldots \\), and \\( D_{0}^{18} \\) are subsequently generated using both \\( D_{0}^{10} \\) and \\( w^{+} \\), as shown in the rightmost column of Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Looping video \\(V\\) with frame length \\(N + 1\\) is synthesized using the joint splating technique [9] using the future and past displacement fields \\(F_{0\\rightarrow t}\\) and \\(F_{N\\rightarrow t}\\), generated through Euler integration. The displacement fields are computed with \\(t\\) times of integration on \\(M\\) for \\(F_{0\\rightarrow t}\\), and \\(N - t\\) times of integration on \\(-M\\) for \\(F_{N\\rightarrow t}\\). The multi-scale fea" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.453, + 0.892, + 0.501 + ], + "angle": 0, + "content": "ture \\(D_0^i\\) is warped in both directions and is composited to form \\(D_t^i\\). Specifically, \\(D_t^i\\) is computed as a weighted sum of \\(Warp(D_0^i,F_{0\\to t})\\) and \\(Warp(D_0^i,F_{N\\to t})\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.514, + 0.891, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} D _ {t} ^ {i} \\left(x ^ {\\prime}\\right) = \\frac {\\sum_ {x \\in \\chi} \\alpha_ {t} \\cdot W a r p \\left(D _ {0} ^ {i} (x) , F _ {0 \\rightarrow t}\\right)}{\\sum_ {x \\in \\chi} \\alpha_ {t}} \\\\ + \\frac {\\sum_ {x \\in \\chi} \\left(1 - \\alpha_ {t}\\right) \\cdot W a r p \\left(D _ {0} ^ {i} (x) , F _ {N \\rightarrow t}\\right)}{\\sum_ {x \\in \\chi} \\left(1 - \\alpha_ {t}\\right)}, \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.603, + 0.891, + 0.649 + ], + "angle": 0, + "content": "where \\( x \\in \\chi \\) is a set of pixels being mapped to the same destination pixel \\( x' \\), and \\( \\alpha_{t} \\) is the looping weight defined as \\( (1 - \\frac{t}{N}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.649, + 0.891, + 0.692 + ], + "angle": 0, + "content": "With the above DFW module, we can generate a video frame given the predicted deep features \\(D^{10}\\), latent code \\(w^{+}\\), and mask \\(S\\):" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.701, + 0.891, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {I} _ {t} = S \\odot \\left(G _ {\\text {w a r p}} \\left(D _ {0} ^ {1 0}, w ^ {+}, F _ {0 \\rightarrow t}, F _ {N \\rightarrow t}\\right)\\right) \\\\ + (\\mathbf {1} - S) \\odot I, \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.75, + 0.891, + 0.793 + ], + "angle": 0, + "content": "where \\(\\odot\\) is an element-wise multiplication and \\(G_{\\text{warp}}\\) is a fixed pre-trained StyleGAN that incorporates our DFW module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.798, + 0.891, + 0.871 + ], + "angle": 0, + "content": "Style Interpolation In addition to motion generation, the change of appearance style is an additional feature often observed in a landscape cinematograph. This can be achieved by interpolating the latent code with the target latent code as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.885, + 0.891, + 0.902 + ], + "angle": 0, + "content": "\\[\nw _ {s} ^ {+} = w ^ {+} \\cdot (1 - \\beta) + w _ {t} ^ {+} \\cdot \\beta , \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "7875" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.17, + 0.473, + 0.228 + ], + "angle": 0, + "content": "Figure 3. Generated cinematagraph results. This figure contains video clips, thus consider viewing it using Adobe Reader. The first two are cinematographs without appearance change, and the last two are cinematographs with appearance change." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.237, + 0.47, + 0.313 + ], + "angle": 0, + "content": "where \\( w_{s}^{+} \\) is the interpolated latent code, \\( w_{t}^{+} \\) is the latent code of the target style, and \\( \\beta \\) is the interpolation weight. This is possible because the later layers of the StyleGAN only modify the color of the synthesized image while the original structure is preserved." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.314, + 0.47, + 0.36 + ], + "angle": 0, + "content": "With modified Equation 4, we prevent the visual mismatch between the static and dynamic regions by also reflecting the changes to the static regions as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.372, + 0.47, + 0.41 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {I} _ {t} = S \\odot \\left(G _ {\\text {w a r p}} \\left(D _ {0} ^ {1 0}, w _ {s} ^ {+}, F _ {0 \\rightarrow t}, F _ {N \\rightarrow t}\\right)\\right) \\tag {6} \\\\ + (\\mathbf {1} - S) \\odot (I + \\Delta I _ {s}), \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.412, + 0.469, + 0.431 + ], + "angle": 0, + "content": "\\[\n\\Delta I _ {s} = G \\left(D _ {0} ^ {1 0}, w _ {s} ^ {+}\\right) - G \\left(D _ {0} ^ {1 0}, w ^ {+}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.444, + 0.47, + 0.475 + ], + "angle": 0, + "content": "where \\(\\Delta I_{s}\\) is the color difference between the two images generated from latent codes \\(w_{s}^{+}\\) and \\(w^{+}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.491, + 0.21, + 0.508 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.47, + 0.639 + ], + "angle": 0, + "content": "In this section, we compare our method with state-of-the-art landscape cinematograph generation methods [6, 9, 17, 19] (Sec. 4.1) and with an unconditional video generation method [34] (Sec. 4.2). In addition, we show the importance of the components used in our method through ablation studies (Sec. 4.3). To observe more various results, we recommend readers see the supplementary video. Example results are presented in Figures 1 and 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.47, + 0.68 + ], + "angle": 0, + "content": "4.1. Comparisons with Cinemagraph Generation Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We compared our method with state-of-the-art cinemagraph generation methods: Animating Landscape (AL) [6], Deep Landscape (DL) [17], Eulerian Motion Field (EMF) [9], and Text2Cinemagraph (T2C) [19]. AL is a learning-based method that uses backward warping to generate cinemagraph frames. DL trains a style-based generator to synthesize images and dynamic motion in a scene. EMF and T2C train an encoder-decoder architecture paired with a motion generator to produce looping cinemagrams. We used official implementations of AL, DL, and T2C, and faithfully reproduced EMF based on the provided training details and hyper-parameters. In the following, we will examine the qualitative difference, evaluate the results using two metrics, and report the results of human perceptual studies." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.092, + 0.87, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.172, + 0.869, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.254, + 0.868, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.337, + 0.868, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.42, + 0.892, + 0.462 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison with state-of-the-art cinematograph generation methods. Please refer to the supplementary video for more examples." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.472, + 0.892, + 0.514 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with the state-of-the-art cinematograph generation methods. We compared our method with AL, EMF, DL, and T2C. The best scores are bolded." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.525, + 0.882, + 0.59 + ], + "angle": 0, + "content": "
MethodStatic ConsistencyMotion Quality
LPIPS↓MS-SSIM↑RMSE↓LPIPS↓MS-SSIM↑RMSE↓FID↓
AL0.04770.95248.89560.06170.681925.108953.6893
EMF0.01030.97236.04400.05330.710222.893245.8035
DL0.00710.99312.10440.05040.706221.801141.6374
T2C0.00630.97734.87850.01590.722421.278440.1186
Ours0.00620.99621.94300.01310.723720.894839.2113
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.892, + 0.91 + ], + "angle": 0, + "content": "Qualitative Comparisons Figure 4 shows visual results from AL, EMF, DL, T2C, and our method. AL exhibits stretching artifacts (row 1) due to recurrent backward warping, whereas our method uses forward warping that prevents the appearance of such distortions. EMF struggles with precise reconstruction (row 2), because its encoder-decoder architecture is trained from scratch on a low-quality video dataset. In contrast, our method leverages a StyleGAN model pre-trained on high-quality images, resulting in cinemagrams with high perceptual quality. DL often loses fine details in textures (row 3) as it performs GAN inversion using the latent codes of StyleGAN, and fine-tunes the generator. Our method, utilizing deep features for GAN inversion, retains original texture details. T2C encounters difficulties in preserving high-frequency details, and accurate motion segmentation between static and dynamic regions (row 4). In contrast, our method utilizes a pre-trained StyleGAN for cinematograph generation, and uses its features for mask prediction, preserving both fine and structural details of the source image." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7876" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.272 + ], + "angle": 0, + "content": "Quantitative Comparisons The evaluation was performed considering two aspects: (1) static consistency and (2) motion quality. Static consistency measures the consistency over time of non-movable regions such as buildings and mountains. Motion quality refers to both the image quality of the animated regions and the animation plausibility. For a fair comparison, we normalized the speed of motion in the generated results for all methods to match the average speed in the test dataset. AL, EMF, DL, and T2C were provided with the first frame \\( I_0 \\) of the test video to generate \\( 512 \\times 512 \\) videos. Our method generates \\( 1024 \\times 1024 \\) results, thus we downsampled it to the size of \\( 512 \\times 512 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.468, + 0.408 + ], + "angle": 0, + "content": "For a quantitative evaluation, we used 224 test videos from the Sky Time-Lapse dataset [41]. To measure the static consistency, we computed LPIPS [45], Root-Mean-Squared Error (RMSE), and MS-SSIM [40] between generated frame \\(\\hat{I}_n\\) and the input image \\(I_0\\) with the sky masked out. To measure the motion quality, we used the same evaluation metrics along with Frechet inception distance (FID) [28] between \\(\\hat{I}_n\\) and the ground-truth frame \\(I_n\\) with the static parts masked out." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.468, + 0.575 + ], + "angle": 0, + "content": "Table 1 reveals the quantitative evaluation results. For both static consistency and motion quality, our method outperforms the other approaches. Our cinematographs represent static consistency better because use of the mask improves the accuracy in the detection of static regions. In addition, the quality of texture details are improved in the generated frames due to GAN inversion and the MSDFW based on the deep features of StyleGAN. The use of our motion generator and DFW leads to improved motion quality for each scene, compared with that of the results of previous approaches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.468, + 0.853 + ], + "angle": 0, + "content": "User Study We conducted a user study with 17 participants to subjectively evaluate our method in comparison with previous cinematograph generation methods. We did not target any specific demographic groups, as the evaluation of cinematographs should not be dependent on particular demographic distributions. We conducted two evaluations: score evaluation and side-by-side comparison. Both evaluations assessed static consistency and motion quality. In the score evaluation, participants rated the video presented with a reference image on a 1-to-5 scale, with \"very bad\" being 1 and \"very good\" being 5. In the side-by-side comparison, participants selected the preferred cinematograph from two videos generated using different methods, presented with a reference image. For both evaluations, we used ten samples randomly chosen from the LHQ [31] dataset. To eliminate bias, distinct samples were used for each evaluation, and the positions of the cinematographs were randomized in all questions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Tables 2 and 3 summarize the statistics of the user studies. For both score evaluation and human preference, our method outperforms the previous approaches by a substan" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.116 + ], + "angle": 0, + "content": "Table 2. Human perceptual study results for score evaluation. The best scores are bolded." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.121, + 0.832, + 0.203 + ], + "angle": 0, + "content": "
MethodStatic ConsistencyMotion Quality
AL1.59 ± 0.201.76 ± 0.45
EMF2.35 ± 0.892.24 ± 0.99
DL3.35 ± 0.833.25 ± 0.67
T2C3.75 ± 0.692.96 ± 0.66
Ours4.37 ± 0.183.86 ± 0.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.217, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Table 3. Human perceptual study results for side-by-side comparison. The percentage of ours chosen against each competing method is reported." + }, + { + "type": "table", + "bbox": [ + 0.558, + 0.262, + 0.83, + 0.344 + ], + "angle": 0, + "content": "
MethodHuman Preference (Ours %)
Static ConsistencyMotion Quality
vs AL99.35%92.81%
vs EMF96.08%86.93%
vs DL94.77%84.97%
vs T2C77.78%73.86%
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.461 + ], + "angle": 0, + "content": "tial margin. We also performed statistical analysis (Kruskal-Wallis H-test) on the resulting evaluation scores. The results showed that our method achieved significantly higher scores in all pairs of comparisons with every previous method \\((p < 0.001\\) in post-hoc analysis). Both user study results reveal the advantages of our method for generating cinematographs with high image quality and plausible animation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.473, + 0.887, + 0.489 + ], + "angle": 0, + "content": "4.2. Comparisons with Video Generation Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.892, + 0.663 + ], + "angle": 0, + "content": "We compared our approach with MoCoGAN-HD [34], a video generation method that predicts the trajectory within the latent space of a pre-trained StyleGAN. For content generation, we used the same pre-trained StyleGAN as that used by MoCoGAN-HD, which was trained on the Sky Time-lapse [41] dataset to generate \\(128 \\times 128\\) images. The comparison was made in an unconditional manner using 300 randomly sampled latent codes, at a resolution of \\(128 \\times 128\\). We compared the first 60 frames of the generated videos with those of MoCoGAN-HD, as our method produces looping videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.892, + 0.789 + ], + "angle": 0, + "content": "Qualitative Comparisons Figure 5 presents a qualitative comparison. As shown in the first and third rows, the videos generated using our method exhibit static consistencies, with the clouds moving while the ground remains static. In contrast, as shown in the second and fourth rows, the frames generated by MoCoGAN-HD deviate significantly from the original image, making the method unsuitable for cinematograph generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Quantitative Comparisons We compared the content preservation ability of both methods by measuring LPIPS [45], RMSE, and MS-SSIM [40] between the first frame \\(\\hat{I}_0\\) and the subsequent frames \\(\\hat{I}_n\\). Because the main content to be preserved in kinemagrams is the static regions, we defined content preservation as static consistency. We masked out the sky for all image frames according to the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7877" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.088, + 0.437, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.352, + 0.47, + 0.395 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison with the state-of-the-art video generation method, MoCoGAN-HD [34]. For more examples from this comparison, please refer to the supplementary video." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.402, + 0.449, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.495, + 0.47, + 0.523 + ], + "angle": 0, + "content": "Figure 6. Quantitative comparison of content preservation with state-of-the-art video generation method MoCoGAN-HD [34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.471, + 0.716 + ], + "angle": 0, + "content": "segmentation mask and compared only the static parts. We used 219 videos in which the scene could be horizontally divided into static and animated regions. LPIPS, RMSE, and MS-SSIM were all computed for each pair of \\(\\hat{I}_0\\) and \\(\\hat{I}_n\\), and were averaged over the number of samples. Figure 6 shows that MoCoGAN-HD exhibits significant divergence in terms of LPIPS, RMSE, and MS-SSIM over time. This indicates that MoCoGAN-HD is unable to generate motion that preserves the content of the original image. In contrast, our method exhibits a small diverging trend, confirming its superiority in preserving the content of the original image over time, by utilizing deep features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.729, + 0.231, + 0.745 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our design choices for the proposed method, we conducted a series of ablation studies. Specifically, we focused on the effectiveness of our warping and GAN inversion methods. To evaluate warping, we compared the image frames generated with and without the use of 1) forward warping, 2) DFW, 3) MSDFW, and 4) segmentation mask. The first frames of 224 test videos from the Sky Time-Lapse [41] dataset were given as input to generate \\(1024 \\times 1024\\) videos, and we used the first 60 frames for the evaluation. For the evaluation of GAN inversion, we" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.089, + 0.693, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.09, + 0.865, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.33, + 0.892, + 0.358 + ], + "angle": 0, + "content": "Figure 7. Results of qualitative comparison in ablation study. Please see the supplementary video for the animated results." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.363, + 0.86, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.481, + 0.874, + 0.495 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison results on GAN inversion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.563 + ], + "angle": 0, + "content": "compared images reconstructed with and without the use of the deep features. A total of 256 images, 128 from the Sky Time-Lapse and 128 from the Eulerian [9] dataset were provided to generate \\(1024 \\times 1024\\) reconstructed images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.573, + 0.893, + 0.756 + ], + "angle": 0, + "content": "Forward Warping We compared the image frames generated with and without forward warping (FW). For the case without FW, we warped the deep features of StyleGAN using backward warping. The first row in Figure 7 shows the qualitative results of this comparison. As revealed in the figure, without FW, the results usually contain unrealistically stretched textures. For quantitative comparisons, we computed LPIPS, MS-SSIM, and RMSE between \\( I_{n} \\) and \\( \\hat{I}_n \\). Table 4 shows that excluding forward warping resulted in significant degradation in terms of MS-SSIM. This indicates that the stretched textures resulted in huge structural distortions in the generated images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Deep Feature Warping We then compared frames generated with and without DFW. For the case without DFW, we directly warped RGB images using predicted motion fields for comparison. The second row in Figure 7 presents the qualitative comparison. As shown in the figure, the frame generated without DFW contained tearing artifacts in the animated regions. Table 4 shows an increase in the LPIPS score, which indicates that excluding DFW degraded the perceptual quality by introducing tearing artifacts." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7878" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.117 + ], + "angle": 0, + "content": "Table 4. Results of quantitative evaluation in ablation study. The best scores are bolded." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.122, + 0.43, + 0.205 + ], + "angle": 0, + "content": "
MethodLPIPS↓MS-SSIM↑RMSE↓
Ours - Full0.05110.716521.8188
Ours - w/o FW0.05240.685322.9816
Ours - w/o DFW0.05370.690822.4993
Ours - w/o MSDFW0.06290.694622.3227
Ours - w/o Mask0.05640.698022.6213
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.213, + 0.468, + 0.24 + ], + "angle": 0, + "content": "Table 5. Quantitative evaluation results on GAN inversion. The best scores are bolded." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.244, + 0.428, + 0.284 + ], + "angle": 0, + "content": "
w+D10FID↓LPIPS↓RMSE↓MS-SSIM↑
✓-40.82760.015318.59010.7891
✓✓11.82310.00195.64100.9907
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.468, + 0.462 + ], + "angle": 0, + "content": "Multi-scale Deep Feature Warping Image frames generated with and without MSDFW were also compared. For the case without MSDFW, only a single deep feature \\(D^{10}\\) was warped and propagated through the next blocks of StyleGAN. The results of the qualitative comparison are shown in the third row of Figure 7. As revealed in the figure, excluding MSDFW resulted in blurry textures in the dynamic region. The quantitative comparison reported in Table 4 shows a significant increase in the LPIPS score. This indicates that excluding MSDFW degraded the perceptual quality, especially the texture details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.468, + 0.616 + ], + "angle": 0, + "content": "Segmentation Mask We compared the image frames generated with and without a segmentation mask. For the case without a mask, we used the initially predicted motion field for warping. The fourth row in Figure 7 shows the qualitative result of this comparison. As shown in the figure, excluding the mask resulted in erroneous movement in the static regions. The quantitative comparison reported in Table 4 reveals a significant degradation of MS-SSIM, which indicates the structural distortion caused by the erroneous motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.468, + 0.803 + ], + "angle": 0, + "content": "Deep Feature Inversion To evaluate the effectiveness of GAN inversion, we compared the images reconstructed with and without the use of the deep features \\(D^{10}\\). For the case without deep feature inversion, only the latent codes \\(w^{+}\\) of a pre-trained StyleGAN were used to reconstruct the input landscape image. Figure 8 shows a qualitative result of this comparison. As revealed in the figure, using only \\(w^{+}\\) failed to accurately reconstruct the details of the original images. For quantitative comparisons, FID, LPIPS, RMSE, and MS-SSIM were computed between \\(I\\) and \\(\\hat{I}\\). Table 5 reveals significant improvements in the perceptual quality of the results when reconstructed using the deep features \\(D^{10}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.819, + 0.35, + 0.835 + ], + "angle": 0, + "content": "5. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Our mask predictor and motion generator generally performed well for landscape images. However, automatic prediction cannot be accurate for all images because most landscape images contain inherent ambiguity in their motion" + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.117, + 0.532, + 0.127 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.092, + 0.693, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.117, + 0.718, + 0.127 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.092, + 0.796, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.798, + 0.092, + 0.874, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.156, + 0.892, + 0.198 + ], + "angle": 0, + "content": "Figure 9. Limitations of StyleCineGAN: (a) automatic prediction of motion cannot be accurate for all images, and (b) the motion of a very thin structured object is hard to be isolated." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.206, + 0.892, + 0.356 + ], + "angle": 0, + "content": "directions except for some obvious cases (e.g., waterfalls). The failure case is illustrated in Figure 9 (a). User-defined motion hints can be used to resolve such ambiguities and provide further control capability during the generation process [18, 19]. In addition, it is hard for our method to isolate the motion of a very thin structured object placed within the animated region as shown in Figure 9 (b). This is because our method performs warping of features at multiple resolutions, in which low-resolution features cannot spatially represent the thin structures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.892, + 0.479 + ], + "angle": 0, + "content": "In this work, we mainly focused on animating the landscape images, particularly skies and fluids, while putting other types of animations outside the scope. In future developments, we would like to expand the capabilities to include other forms of motion. Investigating the rotating hands of a clock, the playing arm of a guitarist moving up and down, and a flag or the wings of a bird fluttering for cinematograph generation would be a very interesting direction to pursue." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.491, + 0.619, + 0.506 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.892, + 0.724 + ], + "angle": 0, + "content": "We proposed the first approach that leverages a pre-trained StyleGAN for high-quality one-shot landscape cinematograph generation. In contrast to previous studies, our method does not require training a large image generator from scratch, and also systematically improves the resolution of the generated cinematographs to \\(1024 \\times 1024\\). At the core of our method, we utilized the deep features of a pretrained StyleGAN, because those features can help preserve spatial information and encode both high-level semantic and low-level style appearances. Using our MSDFW approach, we applied the predicted motion to the deep feature space of StyleGAN. Both qualitative and quantitative results confirm that our method substantially outperforms existing baselines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements This research was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023 (Project Name: Development of Universal Fashion Creation Platform Technology for Avatar Personality Expression, Project Number: RS-2023-00228331, Contribution Rate: \\(90\\%\\) ) and Korea Institute for Advancement of Technology(KIAT) grant funded by the Korea Government(MOTIE) (Project Name: The Competency Development Program for Industry Specialist, Project Number: P0012746, Contribution Rate: \\(10\\%\\))." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7879" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.469, + 0.212 + ], + "angle": 0, + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? CoRR, abs/1911.11544, 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.469, + 0.269 + ], + "angle": 0, + "content": "[3] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3. arXiv preprint arXiv:2201.13433, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.469, + 0.337 + ], + "angle": 0, + "content": "[4] Hugo Bertiche, Niloy J. Mitra, Kuldeep Kulkarni, ChunHao Paul Huang, Tuanfeng Y. Wang, Meysam Madadi, Sergio Escalera, and Duygu Ceylan. Blowing in the wind: Cyclenet for human cinematographs from still images, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.469, + 0.396 + ], + "angle": 0, + "content": "[5] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. In ACM SIGGRAPH 2005 Papers, pages 853-860. 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.469, + 0.464 + ], + "angle": 0, + "content": "[6] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. ACM Transactions on Graphics (Proc. of SIGGRAPH ASIA 2019), 38(6):175:1-175:19, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.469, + 0.507 + ], + "angle": 0, + "content": "[7] Gereon Fox, Ayush Tewari, Mohamed Elgharib, and Christian Theobalt. Stylevideogan: A temporal generative model using a pretrained stylegan, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.469, + 0.577 + ], + "angle": 0, + "content": "[8] Tavi Halperin, Hanit Hakim, Orestis Vantzos, Gershon Hochman, Netai Benaim, Lior Sassy, Michael Kupchik, Ofir Bibi, and Ohad Fried. Endless loops: detecting and animating periodic patterns in still images. ACM Transactions on Graphics (TOG), 40(4):1-12, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.578, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[9] Aleksander Holynski, Brian L. Curless, Steven M. Seitz, and Richard Szeliski. Animating pictures with eulerian motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5810-5819, 2021. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[10] Ali Jahanian, Lucy Chai, and Phillip Isola. \"On the \"steerability\" of generative adversarial networks. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[11] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. CoRR, abs/1710.10196, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[12] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[13] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[14] Alexander Kristoffersen. Loopnerf: Exploring temporal compression for 3d video textures. Master's thesis, EECS Department, University of California, Berkeley, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[15] Xingyi Li, Zhiguo Cao, Huiqiang Sun, Jianming Zhang, Ke Xian, and Guosheng Lin. 3d cinematography from a single" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4595-4605, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.193 + ], + "angle": 0, + "content": "[16] Ji Lin, Richard Zhang, Frieder Ganz, Song Han, and JunYan Zhu. Anycost gans for interactive image synthesis and editing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.195, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[17] Elizaveta Logacheva, Roman Suvorov, Oleg Khomenko, Anton Mashikhin, and Victor Lempitsky. Deeplandscape: Adversarial modeling of landscape videos. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.266, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[18] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.324, + 0.892, + 0.366 + ], + "angle": 0, + "content": "[19] Aniruddha Mahapatra, Aliaksandr Siarohin, Hsin-Ying Lee, Sergey Tulyakov, and Jun-Yan Zhu. Text-guided synthesis of eulerian cinemagrams. 2023. 1, 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.368, + 0.892, + 0.424 + ], + "angle": 0, + "content": "[20] Makoto Okabe, Ken Anjyo, Takeo Igarashi, and Hans-Peter Seidel. Animating pictures of fluid using video examples. In Computer Graphics Forum, pages 677-686. Wiley Online Library, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.426, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[21] Makoto Okabe, Ken Anjyor, and Rikio Onai. Creating fluid animation from a single image using video database. In Computer Graphics Forum, pages 1973-1982. Wiley Online Library, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.526 + ], + "angle": 0, + "content": "[22] Makoto Okabe, Yoshinori Dobashi, and Ken Anjyo. Animating pictures of water scenes using video retrieval. The Visual Computer, 34(3):347-358, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.528, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[23] Ehsan Pajouheshgar, Yitao Xu, Tong Zhang, and Sabine Süssstrunk. Dynca: Real-time dynamic texture synthesis using neural cellular automata. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.6, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[24] Gaurav Parmar, Yijun Li, Jingwan Lu, Richard Zhang, Jun-Yan Zhu, and Krishna Kumar Singh. Spatially-adaptive multilayer selection for gan inversion and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.671, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[25] Ekta Prashnani, Maneli Noorkami, Daniel Vaquero, and Pradeep Sen. A phase-based approach for animating images using video examples. In Computer Graphics Forum, pages 303-311. Wiley Online Library, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.729, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[26] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.801, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[27] Masaki Saito, Shunta Saito, Masanori Koyama, and Sosuke Kobayashi. Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal gan, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7880" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Kwanggyoon Seo, Seoung Wug Oh, Jingwan Lu, Joon-Young Lee, Seonghyeon Kim, and Junyong Noh. Styleportraitvideo: Editing portrait videos with expression optimization. 41(7), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.472, + 0.207 + ], + "angle": 0, + "content": "[30] Claude E. Shannon. Coding Theorems for a Discrete Source With a Fidelity Criterion Institute of Radio Engineers, International Convention Record, vol. 7, 1959., pages 325-350. 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.472, + 0.251 + ], + "angle": 0, + "content": "[31] Ivan Skorokhodov, Grigorii Sotnikov, and Mohamed Elhoseiny. Aligning latent and image spaces to connect the unconnectable. arXiv preprint arXiv:2104.06954, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.253, + 0.472, + 0.323 + ], + "angle": 0, + "content": "[32] Ryusuke Sugimoto, Mingming He, Jing Liao, and Pedro V. Sander. Water simulation and rendering from a still photograph. In SIGGRAPH Asia 2022 Conference Papers, New York, NY, USA, 2022. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.325, + 0.472, + 0.381 + ], + "angle": 0, + "content": "[33] Matthew Tesfaldet, Marcus A. Brubaker, and Konstantinos G. Derpanis. Two-stream convolutional networks for dynamic texture synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.383, + 0.472, + 0.453 + ], + "angle": 0, + "content": "[34] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. In International Conference on Learning Representations, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.472, + 0.51 + ], + "angle": 0, + "content": "[35] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4): 1-14, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.513, + 0.472, + 0.569 + ], + "angle": 0, + "content": "[36] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1526-1535, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.572, + 0.472, + 0.642 + ], + "angle": 0, + "content": "[37] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. Stitch it in time: Gan-based facial editing of real videos. In SIGGRAPH Asia 2022 Conference Papers, New York, NY, USA, 2022. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.472, + 0.699 + ], + "angle": 0, + "content": "[38] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.472, + 0.771 + ], + "angle": 0, + "content": "[39] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.472, + 0.829 + ], + "angle": 0, + "content": "[40] Zhou Wang, Eero P. Simoncelli, and Alan Conrad Bovik. Multiscale structural similarity for image quality assessment. The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, 2:1398-1402 Vol.2, 2003. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[41] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[42] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally consistent semantic video editing. pages 357-374. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[43] Xu Yao, Alasdair Newson, Yann Gousseau, and Pierre Hellier. A style-based gan encoder for high fidelity reconstruction of images and videos. European conference on computer vision, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[44] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 85-101. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[46] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10145-10155, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "7881" + } + ] +] \ No newline at end of file diff --git a/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/51ff4f8c-4976-4d05-85df-cf04bbc75017_origin.pdf b/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/51ff4f8c-4976-4d05-85df-cf04bbc75017_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..98e6a46ccd4918a71ccf7492e31c0ffd6462a460 --- /dev/null +++ b/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/51ff4f8c-4976-4d05-85df-cf04bbc75017_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2cb5e8c2101a58160e1e26c2b8a997e4486434d4d60529675d9d1e3546906096 +size 9571833 diff --git a/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/full.md b/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c2cd7b3f9c9f7a4d645626cc52010458707cbca6 --- /dev/null +++ b/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/full.md @@ -0,0 +1,317 @@ +# StyleCineGAN: Landscape Cinemagraph Generation using a Pre-trained StyleGAN + +Jongwoo Choi + +Kwanggyoon Seo + +Amirsaman Ashtari + +Junyong Noh + +Visual Media Lab, KAIST + +{jeolpyeoni0,skg1023,a.s.ashtari,junyongnoh}@kaist.ac.kr + +Figure 1. Given a landscape image, StyleCineGAN generates a seamless cinematograph at $1024 \times 1024$ resolution. This figure contains video clips, thus consider viewing it using Adobe Reader. The same results are also included in the supplementary video. + +# Abstract + +We propose a method that can generate cinematographs automatically from a still landscape image using a pretrained StyleGAN. Inspired by the success of recent unconditional video generation, we leverage a powerful pretrained image generator to synthesize high-quality cinematographs. Unlike previous approaches that mainly utilize the latent space of a pre-trained StyleGAN, our approach utilizes its deep feature space for both GAN inversion and cinematograph generation. Specifically, we propose multi-scale deep feature warping (MSDFW), which warps the intermediate features of a pre-trained StyleGAN at different resolutions. By using MSDFW, the generated cinematographs are of high resolution and exhibit plausible looping animation. We demonstrate the superiority of our method through user studies and quantitative comparisons with state-of-the-art cinematograph generation methods and a video generation method that uses a pre-trained StyleGAN. + +# 1. Introduction + +Cinemagraph is a unique form of media that combines a still image and video. While most of the scene remains still, subtle and repeated movements in only a small region effectively highlight a particular object or an important event that occurred at the time of capture. Because this mixture + +of still and moving elements in one scene creates an interesting eye-catching effect, cinematographs have recently attracted increasing attention in social media. + +Despite the growing popularity of cinematograph, its creation heavily relies on a manual process. To create a motion in part of the input image, one typically utilizes an image manipulation tool to stretch, scale, rotate, or shift the part of the image in a physically plausible and aesthetically pleasing way. This process is often time-consuming and requires a high level of skill in the use of image editing tools. Therefore, creating a cinematograph has generally been considered to be a personal project of professionals so far. + +To allow ordinary users to create a cinematograph with a single image, automatic methods have been proposed. One line of research uses a reference video as guidance [14, 25, 33]. These methods are capable of producing realistic motions in the target scene similar to a reference video. Recent methods obviate the need for a reference video, by training deep generative models [4, 6, 9, 15, 17-19, 23]. These methods decompose the task into two processes of learning motion and spatial information individually from separate datasets. The decomposition effectively reduces the complexity of simultaneously learning both temporal and spatial information, and improves the perceptual quality of the generated videos. However, these models + +must be trained from scratch, which sometimes requires several days if not weeks on a modern GPU. Moreover, these models are not specifically designed to generate high-resolution of $1024 \times 1024$ cinemagrams, because of the significant requirements in memory and processing power. + +In this paper, we propose the first approach to high-quality one-shot landscape cinematograph generation based on a pre-trained StyleGAN [12]. By using the learned image prior of StyleGAN, our method removes the need for training a large model from scratch and systematically improves the resolution of the generated cinematographs to $1024 \times 1024$ . Moreover, our method enables a user to easily edit the style and appearance of the generated cinematograph by leveraging the properties of StyleGAN for image stylization and editing. + +Our method is inspired by recent unconditional video generation methods [7, 34] which allow to navigate the latent space of a pre-trained image generator, to synthesize a high-quality temporally coherent video. Unlike these methods that utilize the latent codes of StyleGAN, we opt to use the deep features that are generated by convolution operations in each layer of StyleGAN. We use these deep features for two reasons. First, we observed that highly detailed landscape images cannot be reconstructed accurately from the latent codes using GAN inversion methods because these latent codes are low-dimensional [30]. Second, a plausible motion that preserves the content cannot be created by only navigating the latent space, because this space is highly semantic-condensed and lacks explicit spatial prior [38]. + +To solve these problems, for the first time in the cinematograph generation domain, we utilize the deep features of a pre-trained StyleGAN. These deep features preserve spatial information and encode both high-level semantic and low-level style appearances across high and low-level convolutional layers. To produce cinematographs from these deep features, we propose a multi-scale deep feature warping (MSDFW) to apply the motions generated using a motion generator to the deep feature space of StyleGAN. + +We demonstrate the effectiveness and advantages of our method by comparing it to state-of-the-art methods in cinematagraph generation [6, 9, 17, 19] and to an unconditional video generation method that uses a pre-trained StyleGAN [34], using various metrics. We also performed a user study to verify the effectiveness of our method for creating visually pleasing cinematagraph effects. Both qualitative and quantitative results confirm that our method substantially outperforms all existing baselines. + +# 2. Related Work + +# 2.1. Cinemagraph Generation + +One of the early studies for cinemagraph generation used a procedural approach to decompose each object into layers + +and apply time-varying stochastic motion [5]. This method can handle a diverse range of images such as photos and paintings by relying on manual interaction from the user. Another approach is to use a reference video to animate the given image [20-22, 25]. These methods rely on a statistical motion analysis of videos to transfer their periodic motion property to the desired image. Another method distills a dynamic NeRF to render looping 3D video textures from the representative motion found in the reference video [14]. Without any reference video as guidance, Halperin et al. [8] proposed a framework to animate arbitrary objects with periodic patterns in a simple motion direction. + +Recent approaches use the capacity of deep learning to automatically create cinemagrams from a single image. One study trained an image-based renderer to generate water animation, utilizing the water simulation results [32]. Another research first predicts a sequence of normal maps, then generates the corresponding RGB images that show a looping animation of a garment as if it is blown in the wind [4]. Another line of research trains a generator to produce a motion field [6, 9, 15, 18, 19]. Our work is similar to these deep learning-based approaches in that a motion generator is utilized to synthesize the cinemagrams. The difference is that our method leverages the deep features of a pre-trained image generator and thus can generate plausible high-quality results without fine-tuning or training a separate synthesis network. At the same time, our method systematically improves the resolution of the generated cinemagrams to $1024 \times 1024$ . + +# 2.2. Unconditional Video Generation + +The task of video generation is known to be difficult because the generation process has to take into account both spatial and temporal information. To ease the problem, previous approaches [27, 36] decompose video generation into content and motion generation. These methods first predict the latent trajectory for motion, then generate a video from the set of predicted latent codes using the image generator. Instead of training the generation model from scratch, MoCoGAN-HD [34] and StyleVideoGAN [7] leverage a pre-trained image generator model, StyleGAN [12]. These methods use the property of well-constructed latent space to morph one frame toward the next frames [10]. Utilizing the image generation capability of a pre-trained generator, MoCoGAN-HD and StyleVideoGAN only need to train a motion trajectory generator, which greatly reduces the training time. These methods can generate high-quality videos in multiple domains such as faces, cars, and even sky. However, the content details are not well preserved, which has hindered the methods' application for cinematograph generation. Our method builds upon a similar concept of using a pre-trained StyleGAN. We leverage the deep features instead of the latent codes of StyleGAN to generate looping and content-preserving cinematographs. + +# 2.3. Video Synthesis using pre-trained StyleGAN + +A pre-trained generative model, StyleGAN [11-13], has been actively employed for downstream image and video editing applications. For GAN-based applications, one of the essential steps is to project the desired images or videos to the latent space of the pre-trained GAN model, which is known as GAN inversion. The latent codes are obtained using either an optimization technique [1, 2], a trained encoder [26, 35], or a hybrid of both approaches [16]. For video editing applications, previous methods [3, 29, 37, 42] have utilized a pre-trained image encoder to project all of the frames to the latent space. While most of the methods operate in $\mathcal{W}+$ space [2], we opt to use both $w^{+}$ and deep features [24, 43, 44] of the pre-trained StyleGAN to accurately reconstruct the original image and synthesize motion. + +# 3. Methods + +An overview of our method is shown in Figure 2. Given a landscape image $I$ , our method generates a seamlessly looping cinemagraph $V = \{\hat{I}_0, \dots, \hat{I}_N\}$ using a pre-trained StyleGAN. At the core of our method, we use a pre-trained generator without additional fine-tuning. The overall process is described below. + +(A) First, we project the landscape image into both the latent space and the feature space of StyleGAN. To this end, we train an encoder that outputs both latent codes $w^{+}$ and intermediate features $D^{10}$ of StyleGAN (Sec. 3.1). +(B) In addition, we predict a mask $S$ to divide the image into static and dynamic regions (Sec. 3.2). +(C) Next, we use a motion generator that accepts the landscape image $I$ to synthesize a motion field $M$ , which defines the position of each pixel in the future frames (Sec. 3.3). +(D) Lastly, we generate the final cinematograph frames using the pre-trained StyleGAN with deep feature warping (DFW) operation added in between the original layers. (Sec. 3.4). + +In the following sections, we will describe the details of each process. + +# 3.1. GAN Inversion + +The first step for generating a cinematograph is to project the input image to the latent space of a pre-trained StyleGAN. This process is necessary because StyleGAN is an unconditional generator that cannot take a conditioning input. Many previous methods have used $\mathcal{W} +$ [2] latent space, which is an extended space of native latent space $\mathcal{W}$ . However, we observed that $w^{+} \in \mathcal{W}+$ is not expressive enough to reconstruct the original high-frequency details of the landscape images. Therefore, we chose to use the generated deep features $D^{10}$ of StyleGAN as well as $w^{+}$ . The use of $D^{10}$ + +enables us to recover details in the original input as shown in Figure 8. + +To project an image $I$ to both the latent space and feature space and obtain $w^{+}$ and $D^{10}$ , we train an encoder $E$ similarly to Yao et al. [43], in which the encoder takes $I$ as input and predicts $(w^{+}, D^{10})$ . Additional training details can be found in the supplementary material. + +# 3.2. Mask Prediction + +We improve the quality of GAN inversion described in Sec. 3.1 by training an additional classifier that predicts a mask to separate the static and dynamic regions. Using the mask, the structure of the static regions can be preserved. To this end, we train a multi-layer-perceptron (MLP) classifier that accepts the deep features of StyleGAN as its input and outputs the mask that specifies each dynamic region in the image, as performed in DatasetGAN [46]. + +To train the classifier, we manually annotate 32 segmentation masks $S$ of the selected images. We then follow the same procedure as DatasetGAN but using the deep features $D^{i}$ where $i \in \{10,11,\dots,18\}$ . We resize all these deep features followed by concatenating them in the channel dimension to construct the input feature $D^{*}$ . In the end, paired data $(D^{*},S)$ is constructed. An ensemble of 10 MLP classifiers is trained with a cross-entropy loss. To further improve the performance of the previous work, at inference, we also refine the predicted mask. For additional details, please refer to the supplementary material. + +# 3.3. Motion Generation + +To generate motion from a static landscape image $I$ , we use an Eulerian motion field $M$ which represents the immediate velocity of each pixel. To predict $M$ given $I$ , an imaged-to-image translation network [39] is trained as a motion generator. We train the network with paired data $(I, M)$ as performed in the previous method [9]. Adding controls in motion generation is also possible, either by using text prompts [19] or drawing arrows [18]. + +While the trained network works well to a certain degree, we observed that the predicted motion field $M$ does not align with the boundaries between the static and dynamic regions and often contains motion in static regions. Thus, we further refine $M$ by multiplying it with the segmentation mask $S$ , which effectively removes most errors. The refined $M$ is then used to simulate the motion of a pixel at time $t$ using the Euler integration as follows: + +$$ +\begin{array}{l} x _ {t + 1} = x _ {t} + M \left(x _ {t}\right), \quad M \left(x _ {t}\right) = F _ {t \rightarrow t + 1} \left(x _ {t}\right), \\ F _ {0 \rightarrow t} (x _ {0}) = F _ {0 \rightarrow t - 1} (x _ {0}) + M \left(x _ {0} + F _ {0 \rightarrow t - 1} (x _ {0})\right), \tag {1} \\ \end{array} +$$ + +where $x_0$ is the source pixel position at time 0, $F_{0\rightarrow t}$ is the displacement field that defines the displacement of the pixel position $x_0$ to the pixel position $x_t$ at time $t$ . + +![](images/4a67e97121df12534c46aef569f41f801346f128f3bc0015d37c5f820bb34afe.jpg) +Motion Generation + +![](images/db2dd263e1cfeda263e26d7f00f69f00cf198b488f3b2ce313b548f3c3422577.jpg) +Cinemagraph Generation +Figure 2. Overview of StyleCineGAN. Given an input landscape image $I$ , our goal is to generate a cienemagraph using a fixed pre-trained StyleGAN $G$ . We project the image into both latent codes $w^{+}$ and deep features $D^{10}$ of $G$ . Using the deep features $D^{*}$ , a mask predictor predicts a segmentation mask $S$ . To animate the input image, we use a motion generator to predict the motion field $M$ from $I$ . $M$ is refined using $S$ . Through Euler integration, $M$ produces the future and past displacement fields $F_{0\rightarrow t}$ and $F_{N\rightarrow t}$ . To synthesize cienemagraph frames, we add a DFW layer in between the layers of $G$ . DFW refers to Eqns. 2 and 3. This modification enables the intermediate features of $G$ to be warped according to $F_{0\rightarrow t}$ and $F_{N\rightarrow t}$ using a joint splatting method at different resolutions, specifically for the StyleGAN layers indexed with $i\in [10,12,14,16,18]$ . The warped deep features are used to synthesize frames $\hat{I}_t$ resulting in the final cienemagraph video. + +# 3.4. Cinemagraph Generation + +After acquiring the displacement field $F_{0\rightarrow t}$ (Sec. 3.3), latent codes $w^{+}$ , and deep features $D^{10}$ of StyleGAN (Sec. 3.1), we feed these data into the pre-trained StyleGAN with our DFW layer added. We apply forward warping (or splatting) to the original content to generate a cinematograph. Here, the application of forward warping directly to an RGB image often results in images with tearing artifacts. To reduce the artifacts, we warp the deep features $D_0^i$ of StyleGAN in different resolutions or scales: + +$$ +D _ {t} ^ {i} = \operatorname {W a r p} \left(D _ {0} ^ {i}, F _ {0 \rightarrow t}\right), \tag {2} +$$ + +where $\text{Warp}$ is the forward warping function and $D_{t}^{i}$ is the warped deep feature at time $t$ and scale $i$ . We observed that warping a single deep feature (e.g., only $D_{0}^{10}$ ) results in blurry textures in the generated cinemagrams. Therefore, we opt to warp the multi-scale deep features; we call this operation a MSDFW. Specifically, we apply warping to the deep features $D_{0}^{i}$ where $i \in [10, 12, 14, 16, 18]$ . The deep feature $D_{0}^{10}$ is acquired using GAN inversion, and the consequent deep features $D_{0}^{12}, D_{0}^{14}, \ldots$ , and $D_{0}^{18}$ are subsequently generated using both $D_{0}^{10}$ and $w^{+}$ , as shown in the rightmost column of Figure 2. + +Looping video $V$ with frame length $N + 1$ is synthesized using the joint splating technique [9] using the future and past displacement fields $F_{0\rightarrow t}$ and $F_{N\rightarrow t}$ , generated through Euler integration. The displacement fields are computed with $t$ times of integration on $M$ for $F_{0\rightarrow t}$ , and $N - t$ times of integration on $-M$ for $F_{N\rightarrow t}$ . The multi-scale fea + +ture $D_0^i$ is warped in both directions and is composited to form $D_t^i$ . Specifically, $D_t^i$ is computed as a weighted sum of $Warp(D_0^i,F_{0\to t})$ and $Warp(D_0^i,F_{N\to t})$ as follows: + +$$ +\begin{array}{l} D _ {t} ^ {i} \left(x ^ {\prime}\right) = \frac {\sum_ {x \in \chi} \alpha_ {t} \cdot W a r p \left(D _ {0} ^ {i} (x) , F _ {0 \rightarrow t}\right)}{\sum_ {x \in \chi} \alpha_ {t}} \\ + \frac {\sum_ {x \in \chi} \left(1 - \alpha_ {t}\right) \cdot W a r p \left(D _ {0} ^ {i} (x) , F _ {N \rightarrow t}\right)}{\sum_ {x \in \chi} \left(1 - \alpha_ {t}\right)}, \tag {3} \\ \end{array} +$$ + +where $x \in \chi$ is a set of pixels being mapped to the same destination pixel $x'$ , and $\alpha_{t}$ is the looping weight defined as $(1 - \frac{t}{N})$ . + +With the above DFW module, we can generate a video frame given the predicted deep features $D^{10}$ , latent code $w^{+}$ , and mask $S$ : + +$$ +\begin{array}{l} \hat {I} _ {t} = S \odot \left(G _ {\text {w a r p}} \left(D _ {0} ^ {1 0}, w ^ {+}, F _ {0 \rightarrow t}, F _ {N \rightarrow t}\right)\right) \\ + (\mathbf {1} - S) \odot I, \tag {4} \\ \end{array} +$$ + +where $\odot$ is an element-wise multiplication and $G_{\text{warp}}$ is a fixed pre-trained StyleGAN that incorporates our DFW module. + +Style Interpolation In addition to motion generation, the change of appearance style is an additional feature often observed in a landscape cinematograph. This can be achieved by interpolating the latent code with the target latent code as follows: + +$$ +w _ {s} ^ {+} = w ^ {+} \cdot (1 - \beta) + w _ {t} ^ {+} \cdot \beta , \tag {5} +$$ + +where $w_{s}^{+}$ is the interpolated latent code, $w_{t}^{+}$ is the latent code of the target style, and $\beta$ is the interpolation weight. This is possible because the later layers of the StyleGAN only modify the color of the synthesized image while the original structure is preserved. + +With modified Equation 4, we prevent the visual mismatch between the static and dynamic regions by also reflecting the changes to the static regions as follows: + +$$ +\begin{array}{l} \hat {I} _ {t} = S \odot \left(G _ {\text {w a r p}} \left(D _ {0} ^ {1 0}, w _ {s} ^ {+}, F _ {0 \rightarrow t}, F _ {N \rightarrow t}\right)\right) \tag {6} \\ + (\mathbf {1} - S) \odot (I + \Delta I _ {s}), \\ \end{array} +$$ + +$$ +\Delta I _ {s} = G \left(D _ {0} ^ {1 0}, w _ {s} ^ {+}\right) - G \left(D _ {0} ^ {1 0}, w ^ {+}\right), \tag {7} +$$ + +where $\Delta I_{s}$ is the color difference between the two images generated from latent codes $w_{s}^{+}$ and $w^{+}$ . + +# 4. Experiments + +In this section, we compare our method with state-of-the-art landscape cinematograph generation methods [6, 9, 17, 19] (Sec. 4.1) and with an unconditional video generation method [34] (Sec. 4.2). In addition, we show the importance of the components used in our method through ablation studies (Sec. 4.3). To observe more various results, we recommend readers see the supplementary video. Example results are presented in Figures 1 and 3. + +# 4.1. Comparisons with Cinemagraph Generation Methods + +We compared our method with state-of-the-art cinemagraph generation methods: Animating Landscape (AL) [6], Deep Landscape (DL) [17], Eulerian Motion Field (EMF) [9], and Text2Cinemagraph (T2C) [19]. AL is a learning-based method that uses backward warping to generate cinemagraph frames. DL trains a style-based generator to synthesize images and dynamic motion in a scene. EMF and T2C train an encoder-decoder architecture paired with a motion generator to produce looping cinemagrams. We used official implementations of AL, DL, and T2C, and faithfully reproduced EMF based on the provided training details and hyper-parameters. In the following, we will examine the qualitative difference, evaluate the results using two metrics, and report the results of human perceptual studies. + +![](images/0969637ceb9ca37a33559ba2e6c9207ff7047175487bea4ca60a64b79dfa815f.jpg) +Figure 3. Generated cinematagraph results. This figure contains video clips, thus consider viewing it using Adobe Reader. The first two are cinematographs without appearance change, and the last two are cinematographs with appearance change. + +![](images/39a0213e700ef9bc56ba9bc73a6eeddde37538bac4a9f16f4257e2c9eadcc04f.jpg) + +![](images/2b2dff4bbec86926ff16f3a9189aec4cf3d1734c21855e584c6339d7e0200f38.jpg) + +![](images/fd20f8b3e221aecb6ce65ca529f33fe1dbf2ad802f07ce17154bbfe3efe03578.jpg) +Figure 4. Qualitative comparison with state-of-the-art cinematograph generation methods. Please refer to the supplementary video for more examples. + +Table 1. Quantitative comparison with the state-of-the-art cinematograph generation methods. We compared our method with AL, EMF, DL, and T2C. The best scores are bolded. + +
MethodStatic ConsistencyMotion Quality
LPIPS↓MS-SSIM↑RMSE↓LPIPS↓MS-SSIM↑RMSE↓FID↓
AL0.04770.95248.89560.06170.681925.108953.6893
EMF0.01030.97236.04400.05330.710222.893245.8035
DL0.00710.99312.10440.05040.706221.801141.6374
T2C0.00630.97734.87850.01590.722421.278440.1186
Ours0.00620.99621.94300.01310.723720.894839.2113
+ +Qualitative Comparisons Figure 4 shows visual results from AL, EMF, DL, T2C, and our method. AL exhibits stretching artifacts (row 1) due to recurrent backward warping, whereas our method uses forward warping that prevents the appearance of such distortions. EMF struggles with precise reconstruction (row 2), because its encoder-decoder architecture is trained from scratch on a low-quality video dataset. In contrast, our method leverages a StyleGAN model pre-trained on high-quality images, resulting in cinemagrams with high perceptual quality. DL often loses fine details in textures (row 3) as it performs GAN inversion using the latent codes of StyleGAN, and fine-tunes the generator. Our method, utilizing deep features for GAN inversion, retains original texture details. T2C encounters difficulties in preserving high-frequency details, and accurate motion segmentation between static and dynamic regions (row 4). In contrast, our method utilizes a pre-trained StyleGAN for cinematograph generation, and uses its features for mask prediction, preserving both fine and structural details of the source image. + +Quantitative Comparisons The evaluation was performed considering two aspects: (1) static consistency and (2) motion quality. Static consistency measures the consistency over time of non-movable regions such as buildings and mountains. Motion quality refers to both the image quality of the animated regions and the animation plausibility. For a fair comparison, we normalized the speed of motion in the generated results for all methods to match the average speed in the test dataset. AL, EMF, DL, and T2C were provided with the first frame $I_0$ of the test video to generate $512 \times 512$ videos. Our method generates $1024 \times 1024$ results, thus we downsampled it to the size of $512 \times 512$ . + +For a quantitative evaluation, we used 224 test videos from the Sky Time-Lapse dataset [41]. To measure the static consistency, we computed LPIPS [45], Root-Mean-Squared Error (RMSE), and MS-SSIM [40] between generated frame $\hat{I}_n$ and the input image $I_0$ with the sky masked out. To measure the motion quality, we used the same evaluation metrics along with Frechet inception distance (FID) [28] between $\hat{I}_n$ and the ground-truth frame $I_n$ with the static parts masked out. + +Table 1 reveals the quantitative evaluation results. For both static consistency and motion quality, our method outperforms the other approaches. Our cinematographs represent static consistency better because use of the mask improves the accuracy in the detection of static regions. In addition, the quality of texture details are improved in the generated frames due to GAN inversion and the MSDFW based on the deep features of StyleGAN. The use of our motion generator and DFW leads to improved motion quality for each scene, compared with that of the results of previous approaches. + +User Study We conducted a user study with 17 participants to subjectively evaluate our method in comparison with previous cinematograph generation methods. We did not target any specific demographic groups, as the evaluation of cinematographs should not be dependent on particular demographic distributions. We conducted two evaluations: score evaluation and side-by-side comparison. Both evaluations assessed static consistency and motion quality. In the score evaluation, participants rated the video presented with a reference image on a 1-to-5 scale, with "very bad" being 1 and "very good" being 5. In the side-by-side comparison, participants selected the preferred cinematograph from two videos generated using different methods, presented with a reference image. For both evaluations, we used ten samples randomly chosen from the LHQ [31] dataset. To eliminate bias, distinct samples were used for each evaluation, and the positions of the cinematographs were randomized in all questions. + +Tables 2 and 3 summarize the statistics of the user studies. For both score evaluation and human preference, our method outperforms the previous approaches by a substan + +Table 2. Human perceptual study results for score evaluation. The best scores are bolded. + +
MethodStatic ConsistencyMotion Quality
AL1.59 ± 0.201.76 ± 0.45
EMF2.35 ± 0.892.24 ± 0.99
DL3.35 ± 0.833.25 ± 0.67
T2C3.75 ± 0.692.96 ± 0.66
Ours4.37 ± 0.183.86 ± 0.53
+ +Table 3. Human perceptual study results for side-by-side comparison. The percentage of ours chosen against each competing method is reported. + +
MethodHuman Preference (Ours %)
Static ConsistencyMotion Quality
vs AL99.35%92.81%
vs EMF96.08%86.93%
vs DL94.77%84.97%
vs T2C77.78%73.86%
+ +tial margin. We also performed statistical analysis (Kruskal-Wallis H-test) on the resulting evaluation scores. The results showed that our method achieved significantly higher scores in all pairs of comparisons with every previous method $(p < 0.001$ in post-hoc analysis). Both user study results reveal the advantages of our method for generating cinematographs with high image quality and plausible animation. + +# 4.2. Comparisons with Video Generation Methods + +We compared our approach with MoCoGAN-HD [34], a video generation method that predicts the trajectory within the latent space of a pre-trained StyleGAN. For content generation, we used the same pre-trained StyleGAN as that used by MoCoGAN-HD, which was trained on the Sky Time-lapse [41] dataset to generate $128 \times 128$ images. The comparison was made in an unconditional manner using 300 randomly sampled latent codes, at a resolution of $128 \times 128$ . We compared the first 60 frames of the generated videos with those of MoCoGAN-HD, as our method produces looping videos. + +Qualitative Comparisons Figure 5 presents a qualitative comparison. As shown in the first and third rows, the videos generated using our method exhibit static consistencies, with the clouds moving while the ground remains static. In contrast, as shown in the second and fourth rows, the frames generated by MoCoGAN-HD deviate significantly from the original image, making the method unsuitable for cinematograph generation. + +Quantitative Comparisons We compared the content preservation ability of both methods by measuring LPIPS [45], RMSE, and MS-SSIM [40] between the first frame $\hat{I}_0$ and the subsequent frames $\hat{I}_n$ . Because the main content to be preserved in kinemagrams is the static regions, we defined content preservation as static consistency. We masked out the sky for all image frames according to the + +![](images/6019b077d7cbd2e52e32a8f09f5c75e1aed65c3732c400e006c828ad8c966ef1.jpg) +Figure 5. Qualitative comparison with the state-of-the-art video generation method, MoCoGAN-HD [34]. For more examples from this comparison, please refer to the supplementary video. + +![](images/a794bf00fcc35105d17098df0a2401efecfbcc2e2b8133156a66ab6a93226d33.jpg) +Figure 6. Quantitative comparison of content preservation with state-of-the-art video generation method MoCoGAN-HD [34]. + +segmentation mask and compared only the static parts. We used 219 videos in which the scene could be horizontally divided into static and animated regions. LPIPS, RMSE, and MS-SSIM were all computed for each pair of $\hat{I}_0$ and $\hat{I}_n$ , and were averaged over the number of samples. Figure 6 shows that MoCoGAN-HD exhibits significant divergence in terms of LPIPS, RMSE, and MS-SSIM over time. This indicates that MoCoGAN-HD is unable to generate motion that preserves the content of the original image. In contrast, our method exhibits a small diverging trend, confirming its superiority in preserving the content of the original image over time, by utilizing deep features. + +# 4.3. Ablation Study + +To demonstrate the effectiveness of our design choices for the proposed method, we conducted a series of ablation studies. Specifically, we focused on the effectiveness of our warping and GAN inversion methods. To evaluate warping, we compared the image frames generated with and without the use of 1) forward warping, 2) DFW, 3) MSDFW, and 4) segmentation mask. The first frames of 224 test videos from the Sky Time-Lapse [41] dataset were given as input to generate $1024 \times 1024$ videos, and we used the first 60 frames for the evaluation. For the evaluation of GAN inversion, we + +![](images/eb5b3cdaa79bcc779d32b3d97b9387f720c1d68cd27af3559c9ae022d5cdd0ab.jpg) +Figure 7. Results of qualitative comparison in ablation study. Please see the supplementary video for the animated results. + +![](images/1ef88055d4ef5ca0a0e9822c91ff7031dd6c9f05db397b92441a12a59e9fd8f4.jpg) + +![](images/a7f42b19cf715c06c7d4e413353f86ecd73d80170520fac7dc3477bb3184a3f1.jpg) +Figure 8. Qualitative comparison results on GAN inversion. + +compared images reconstructed with and without the use of the deep features. A total of 256 images, 128 from the Sky Time-Lapse and 128 from the Eulerian [9] dataset were provided to generate $1024 \times 1024$ reconstructed images. + +Forward Warping We compared the image frames generated with and without forward warping (FW). For the case without FW, we warped the deep features of StyleGAN using backward warping. The first row in Figure 7 shows the qualitative results of this comparison. As revealed in the figure, without FW, the results usually contain unrealistically stretched textures. For quantitative comparisons, we computed LPIPS, MS-SSIM, and RMSE between $I_{n}$ and $\hat{I}_n$ . Table 4 shows that excluding forward warping resulted in significant degradation in terms of MS-SSIM. This indicates that the stretched textures resulted in huge structural distortions in the generated images. + +Deep Feature Warping We then compared frames generated with and without DFW. For the case without DFW, we directly warped RGB images using predicted motion fields for comparison. The second row in Figure 7 presents the qualitative comparison. As shown in the figure, the frame generated without DFW contained tearing artifacts in the animated regions. Table 4 shows an increase in the LPIPS score, which indicates that excluding DFW degraded the perceptual quality by introducing tearing artifacts. + +Table 4. Results of quantitative evaluation in ablation study. The best scores are bolded. + +
MethodLPIPS↓MS-SSIM↑RMSE↓
Ours - Full0.05110.716521.8188
Ours - w/o FW0.05240.685322.9816
Ours - w/o DFW0.05370.690822.4993
Ours - w/o MSDFW0.06290.694622.3227
Ours - w/o Mask0.05640.698022.6213
+ +Table 5. Quantitative evaluation results on GAN inversion. The best scores are bolded. + +
w+D10FID↓LPIPS↓RMSE↓MS-SSIM↑
✓-40.82760.015318.59010.7891
✓✓11.82310.00195.64100.9907
+ +Multi-scale Deep Feature Warping Image frames generated with and without MSDFW were also compared. For the case without MSDFW, only a single deep feature $D^{10}$ was warped and propagated through the next blocks of StyleGAN. The results of the qualitative comparison are shown in the third row of Figure 7. As revealed in the figure, excluding MSDFW resulted in blurry textures in the dynamic region. The quantitative comparison reported in Table 4 shows a significant increase in the LPIPS score. This indicates that excluding MSDFW degraded the perceptual quality, especially the texture details. + +Segmentation Mask We compared the image frames generated with and without a segmentation mask. For the case without a mask, we used the initially predicted motion field for warping. The fourth row in Figure 7 shows the qualitative result of this comparison. As shown in the figure, excluding the mask resulted in erroneous movement in the static regions. The quantitative comparison reported in Table 4 reveals a significant degradation of MS-SSIM, which indicates the structural distortion caused by the erroneous motion. + +Deep Feature Inversion To evaluate the effectiveness of GAN inversion, we compared the images reconstructed with and without the use of the deep features $D^{10}$ . For the case without deep feature inversion, only the latent codes $w^{+}$ of a pre-trained StyleGAN were used to reconstruct the input landscape image. Figure 8 shows a qualitative result of this comparison. As revealed in the figure, using only $w^{+}$ failed to accurately reconstruct the details of the original images. For quantitative comparisons, FID, LPIPS, RMSE, and MS-SSIM were computed between $I$ and $\hat{I}$ . Table 5 reveals significant improvements in the perceptual quality of the results when reconstructed using the deep features $D^{10}$ . + +# 5. Limitations and Future Work + +Our mask predictor and motion generator generally performed well for landscape images. However, automatic prediction cannot be accurate for all images because most landscape images contain inherent ambiguity in their motion + +![](images/0a3178b499e93c16913a8a8349e6d82741ef8a510389777e8ab9dc7d67285d5f.jpg) +(a) + +![](images/99266baccdda3e0ba47175c59d0df00a990c5ee099f3af6ed316f18498209f98.jpg) +(b) +Figure 9. Limitations of StyleCineGAN: (a) automatic prediction of motion cannot be accurate for all images, and (b) the motion of a very thin structured object is hard to be isolated. + +![](images/66837910d83aaf923e84f8e453c57f1e4b218fed97a2f2fc7b454e387d80de0e.jpg) + +directions except for some obvious cases (e.g., waterfalls). The failure case is illustrated in Figure 9 (a). User-defined motion hints can be used to resolve such ambiguities and provide further control capability during the generation process [18, 19]. In addition, it is hard for our method to isolate the motion of a very thin structured object placed within the animated region as shown in Figure 9 (b). This is because our method performs warping of features at multiple resolutions, in which low-resolution features cannot spatially represent the thin structures. + +In this work, we mainly focused on animating the landscape images, particularly skies and fluids, while putting other types of animations outside the scope. In future developments, we would like to expand the capabilities to include other forms of motion. Investigating the rotating hands of a clock, the playing arm of a guitarist moving up and down, and a flag or the wings of a bird fluttering for cinematograph generation would be a very interesting direction to pursue. + +# 6. Conclusion + +We proposed the first approach that leverages a pre-trained StyleGAN for high-quality one-shot landscape cinematograph generation. In contrast to previous studies, our method does not require training a large image generator from scratch, and also systematically improves the resolution of the generated cinematographs to $1024 \times 1024$ . At the core of our method, we utilized the deep features of a pretrained StyleGAN, because those features can help preserve spatial information and encode both high-level semantic and low-level style appearances. Using our MSDFW approach, we applied the predicted motion to the deep feature space of StyleGAN. Both qualitative and quantitative results confirm that our method substantially outperforms existing baselines. + +Acknowledgements This research was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023 (Project Name: Development of Universal Fashion Creation Platform Technology for Avatar Personality Expression, Project Number: RS-2023-00228331, Contribution Rate: $90\%$ ) and Korea Institute for Advancement of Technology(KIAT) grant funded by the Korea Government(MOTIE) (Project Name: The Competency Development Program for Industry Specialist, Project Number: P0012746, Contribution Rate: $10\%$ ). + +# References + +[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 3 +[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? CoRR, abs/1911.11544, 2019.3 +[3] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3. arXiv preprint arXiv:2201.13433, 2022. 3 +[4] Hugo Bertiche, Niloy J. Mitra, Kuldeep Kulkarni, ChunHao Paul Huang, Tuanfeng Y. Wang, Meysam Madadi, Sergio Escalera, and Duygu Ceylan. Blowing in the wind: Cyclenet for human cinematographs from still images, 2023. 1, 2 +[5] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. In ACM SIGGRAPH 2005 Papers, pages 853-860. 2005. 2 +[6] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. ACM Transactions on Graphics (Proc. of SIGGRAPH ASIA 2019), 38(6):175:1-175:19, 2019. 1, 2, 5 +[7] Gereon Fox, Ayush Tewari, Mohamed Elgharib, and Christian Theobalt. Stylevideogan: A temporal generative model using a pretrained stylegan, 2021. 2 +[8] Tavi Halperin, Hanit Hakim, Orestis Vantzos, Gershon Hochman, Netai Benaim, Lior Sassy, Michael Kupchik, Ofir Bibi, and Ohad Fried. Endless loops: detecting and animating periodic patterns in still images. ACM Transactions on Graphics (TOG), 40(4):1-12, 2021. 2 +[9] Aleksander Holynski, Brian L. Curless, Steven M. Seitz, and Richard Szeliski. Animating pictures with eulerian motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5810-5819, 2021. 1, 2, 3, 4, 5, 7 +[10] Ali Jahanian, Lucy Chai, and Phillip Isola. "On the "steerability" of generative adversarial networks. In International Conference on Learning Representations, 2020. 2 +[11] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. CoRR, abs/1710.10196, 2017. 3 +[12] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2 +[13] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 3 +[14] Alexander Kristoffersen. Loopnerf: Exploring temporal compression for 3d video textures. Master's thesis, EECS Department, University of California, Berkeley, 2023. 1, 2 +[15] Xingyi Li, Zhiguo Cao, Huiqiang Sun, Jianming Zhang, Ke Xian, and Guosheng Lin. 3d cinematography from a single + +image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4595-4605, 2023. 1, 2 +[16] Ji Lin, Richard Zhang, Frieder Ganz, Song Han, and JunYan Zhu. Anycost gans for interactive image synthesis and editing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[17] Elizaveta Logacheva, Roman Suvorov, Oleg Khomenko, Anton Mashikhin, and Victor Lempitsky. Deeplandscape: Adversarial modeling of landscape videos. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 5 +[18] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 8 +[19] Aniruddha Mahapatra, Aliaksandr Siarohin, Hsin-Ying Lee, Sergey Tulyakov, and Jun-Yan Zhu. Text-guided synthesis of eulerian cinemagrams. 2023. 1, 2, 3, 5, 8 +[20] Makoto Okabe, Ken Anjyo, Takeo Igarashi, and Hans-Peter Seidel. Animating pictures of fluid using video examples. In Computer Graphics Forum, pages 677-686. Wiley Online Library, 2009. 2 +[21] Makoto Okabe, Ken Anjyor, and Rikio Onai. Creating fluid animation from a single image using video database. In Computer Graphics Forum, pages 1973-1982. Wiley Online Library, 2011. +[22] Makoto Okabe, Yoshinori Dobashi, and Ken Anjyo. Animating pictures of water scenes using video retrieval. The Visual Computer, 34(3):347-358, 2018. 2 +[23] Ehsan Pajouheshgar, Yitao Xu, Tong Zhang, and Sabine Süssstrunk. Dynca: Real-time dynamic texture synthesis using neural cellular automata. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1 +[24] Gaurav Parmar, Yijun Li, Jingwan Lu, Richard Zhang, Jun-Yan Zhu, and Krishna Kumar Singh. Spatially-adaptive multilayer selection for gan inversion and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3 +[25] Ekta Prashnani, Maneli Noorkami, Daniel Vaquero, and Pradeep Sen. A phase-based approach for animating images using video examples. In Computer Graphics Forum, pages 303-311. Wiley Online Library, 2017. 1, 2 +[26] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[27] Masaki Saito, Shunta Saito, Masanori Koyama, and Sosuke Kobayashi. Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal gan, 2020. 2 +[28] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.6 + +[29] Kwanggyoon Seo, Seoung Wug Oh, Jingwan Lu, Joon-Young Lee, Seonghyeon Kim, and Junyong Noh. Styleportraitvideo: Editing portrait videos with expression optimization. 41(7), 2022. 3 +[30] Claude E. Shannon. Coding Theorems for a Discrete Source With a Fidelity Criterion Institute of Radio Engineers, International Convention Record, vol. 7, 1959., pages 325-350. 1993. 2 +[31] Ivan Skorokhodov, Grigorii Sotnikov, and Mohamed Elhoseiny. Aligning latent and image spaces to connect the unconnectable. arXiv preprint arXiv:2104.06954, 2021. 6 +[32] Ryusuke Sugimoto, Mingming He, Jing Liao, and Pedro V. Sander. Water simulation and rendering from a still photograph. In SIGGRAPH Asia 2022 Conference Papers, New York, NY, USA, 2022. Association for Computing Machinery. 2 +[33] Matthew Tesfaldet, Marcus A. Brubaker, and Konstantinos G. Derpanis. Two-stream convolutional networks for dynamic texture synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1 +[34] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. In International Conference on Learning Representations, 2021. 2, 5, 6, 7 +[35] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4): 1-14, 2021. 3 +[36] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1526-1535, 2018. 2 +[37] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. Stitch it in time: Gan-based facial editing of real videos. In SIGGRAPH Asia 2022 Conference Papers, New York, NY, USA, 2022. Association for Computing Machinery. 3 +[38] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[39] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 3 +[40] Zhou Wang, Eero P. Simoncelli, and Alan Conrad Bovik. Multiscale structural similarity for image quality assessment. The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, 2:1398-1402 Vol.2, 2003. 6 +[41] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6, 7 + +[42] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally consistent semantic video editing. pages 357-374. Springer, 2022. 3 +[43] Xu Yao, Alasdair Newson, Yann Gousseau, and Pierre Hellier. A style-based gan encoder for high fidelity reconstruction of images and videos. European conference on computer vision, 2022. 3 +[44] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 85-101. Springer, 2022. 3 +[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[46] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10145-10155, 2021. 3 \ No newline at end of file diff --git a/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/images.zip b/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..88894b832aaafaca28f9fc5edc071d9e7cd36cbc --- /dev/null +++ b/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ccb6bc7c92ecd16747151c7ff7851f45159be93ade091329d03ffd6efec9af4f +size 430388 diff --git a/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/layout.json b/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3d24c3208aacab9148a1568bb870863b8834763a --- /dev/null +++ b/2024/StyleCineGAN_ Landscape Cinemagraph Generation using a Pre-trained StyleGAN/layout.json @@ -0,0 +1,9585 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 132, + 103, + 462, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 103, + 462, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 103, + 462, + 140 + ], + "type": "text", + "content": "StyleCineGAN: Landscape Cinemagraph Generation using a Pre-trained StyleGAN" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 161, + 171, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 161, + 171, + 175 + ], + "spans": [ + { + "bbox": [ + 97, + 161, + 171, + 175 + ], + "type": "text", + "content": "Jongwoo Choi" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 195, + 161, + 283, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 161, + 283, + 175 + ], + "spans": [ + { + "bbox": [ + 195, + 161, + 283, + 175 + ], + "type": "text", + "content": "Kwanggyoon Seo" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 307, + 161, + 404, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 161, + 404, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 161, + 404, + 175 + ], + "type": "text", + "content": "Amirsaman Ashtari" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 428, + 161, + 494, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 428, + 161, + 494, + 175 + ], + "spans": [ + { + "bbox": [ + 428, + 161, + 494, + 175 + ], + "type": "text", + "content": "Junyong Noh" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 232, + 176, + 361, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 176, + 361, + 189 + ], + "spans": [ + { + "bbox": [ + 232, + 176, + 361, + 189 + ], + "type": "text", + "content": "Visual Media Lab, KAIST" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 144, + 191, + 446, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 191, + 446, + 204 + ], + "spans": [ + { + "bbox": [ + 144, + 191, + 446, + 204 + ], + "type": "text", + "content": "{jeolpyeoni0,skg1023,a.s.ashtari,junyongnoh}@kaist.ac.kr" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 369, + 547, + 393 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 547, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 547, + 393 + ], + "type": "text", + "content": "Figure 1. Given a landscape image, StyleCineGAN generates a seamless cinematograph at " + }, + { + "bbox": [ + 46, + 369, + 547, + 393 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 46, + 369, + 547, + 393 + ], + "type": "text", + "content": " resolution. This figure contains video clips, thus consider viewing it using Adobe Reader. The same results are also included in the supplementary video." + } + ] + } + ], + "index": 9, + "type": "text" + }, + { + "bbox": [ + 143, + 403, + 192, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 403, + 192, + 415 + ], + "spans": [ + { + "bbox": [ + 143, + 403, + 192, + 415 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 418, + 289, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 289, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 289, + 622 + ], + "type": "text", + "content": "We propose a method that can generate cinematographs automatically from a still landscape image using a pretrained StyleGAN. Inspired by the success of recent unconditional video generation, we leverage a powerful pretrained image generator to synthesize high-quality cinematographs. Unlike previous approaches that mainly utilize the latent space of a pre-trained StyleGAN, our approach utilizes its deep feature space for both GAN inversion and cinematograph generation. Specifically, we propose multi-scale deep feature warping (MSDFW), which warps the intermediate features of a pre-trained StyleGAN at different resolutions. By using MSDFW, the generated cinematographs are of high resolution and exhibit plausible looping animation. We demonstrate the superiority of our method through user studies and quantitative comparisons with state-of-the-art cinematograph generation methods and a video generation method that uses a pre-trained StyleGAN." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 635, + 128, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 128, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 128, + 647 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": "Cinemagraph is a unique form of media that combines a still image and video. While most of the scene remains still, subtle and repeated movements in only a small region effectively highlight a particular object or an important event that occurred at the time of capture. Because this mixture" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 403, + 545, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 440 + ], + "type": "text", + "content": "of still and moving elements in one scene creates an interesting eye-catching effect, cinematographs have recently attracted increasing attention in social media." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 445, + 546, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 546, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 546, + 553 + ], + "type": "text", + "content": "Despite the growing popularity of cinematograph, its creation heavily relies on a manual process. To create a motion in part of the input image, one typically utilizes an image manipulation tool to stretch, scale, rotate, or shift the part of the image in a physically plausible and aesthetically pleasing way. This process is often time-consuming and requires a high level of skill in the use of image editing tools. Therefore, creating a cinematograph has generally been considered to be a personal project of professionals so far." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 557, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 714 + ], + "type": "text", + "content": "To allow ordinary users to create a cinematograph with a single image, automatic methods have been proposed. One line of research uses a reference video as guidance [14, 25, 33]. These methods are capable of producing realistic motions in the target scene similar to a reference video. Recent methods obviate the need for a reference video, by training deep generative models [4, 6, 9, 15, 17-19, 23]. These methods decompose the task into two processes of learning motion and spatial information individually from separate datasets. The decomposition effectively reduces the complexity of simultaneously learning both temporal and spatial information, and improves the perceptual quality of the generated videos. However, these models" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7872" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "content": "must be trained from scratch, which sometimes requires several days if not weeks on a modern GPU. Moreover, these models are not specifically designed to generate high-resolution of " + }, + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "content": " cinemagrams, because of the significant requirements in memory and processing power." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 132, + 286, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 132, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 132, + 286, + 251 + ], + "type": "text", + "content": "In this paper, we propose the first approach to high-quality one-shot landscape cinematograph generation based on a pre-trained StyleGAN [12]. By using the learned image prior of StyleGAN, our method removes the need for training a large model from scratch and systematically improves the resolution of the generated cinematographs to " + }, + { + "bbox": [ + 46, + 132, + 286, + 251 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 46, + 132, + 286, + 251 + ], + "type": "text", + "content": ". Moreover, our method enables a user to easily edit the style and appearance of the generated cinematograph by leveraging the properties of StyleGAN for image stylization and editing." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 251, + 286, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 286, + 430 + ], + "type": "text", + "content": "Our method is inspired by recent unconditional video generation methods [7, 34] which allow to navigate the latent space of a pre-trained image generator, to synthesize a high-quality temporally coherent video. Unlike these methods that utilize the latent codes of StyleGAN, we opt to use the deep features that are generated by convolution operations in each layer of StyleGAN. We use these deep features for two reasons. First, we observed that highly detailed landscape images cannot be reconstructed accurately from the latent codes using GAN inversion methods because these latent codes are low-dimensional [30]. Second, a plausible motion that preserves the content cannot be created by only navigating the latent space, because this space is highly semantic-condensed and lacks explicit spatial prior [38]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 430, + 286, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 286, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 286, + 536 + ], + "type": "text", + "content": "To solve these problems, for the first time in the cinematograph generation domain, we utilize the deep features of a pre-trained StyleGAN. These deep features preserve spatial information and encode both high-level semantic and low-level style appearances across high and low-level convolutional layers. To produce cinematographs from these deep features, we propose a multi-scale deep feature warping (MSDFW) to apply the motions generated using a motion generator to the deep feature space of StyleGAN." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 537, + 286, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 286, + 643 + ], + "type": "text", + "content": "We demonstrate the effectiveness and advantages of our method by comparing it to state-of-the-art methods in cinematagraph generation [6, 9, 17, 19] and to an unconditional video generation method that uses a pre-trained StyleGAN [34], using various metrics. We also performed a user study to verify the effectiveness of our method for creating visually pleasing cinematagraph effects. Both qualitative and quantitative results confirm that our method substantially outperforms all existing baselines." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 653, + 134, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 134, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 134, + 665 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 673, + 190, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 190, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 190, + 685 + ], + "type": "text", + "content": "2.1. Cinemagraph Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "One of the early studies for cinemagraph generation used a procedural approach to decompose each object into layers" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": "and apply time-varying stochastic motion [5]. This method can handle a diverse range of images such as photos and paintings by relying on manual interaction from the user. Another approach is to use a reference video to animate the given image [20-22, 25]. These methods rely on a statistical motion analysis of videos to transfer their periodic motion property to the desired image. Another method distills a dynamic NeRF to render looping 3D video textures from the representative motion found in the reference video [14]. Without any reference video as guidance, Halperin et al. [8] proposed a framework to animate arbitrary objects with periodic patterns in a simple motion direction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 216, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 545, + 418 + ], + "type": "text", + "content": "Recent approaches use the capacity of deep learning to automatically create cinemagrams from a single image. One study trained an image-based renderer to generate water animation, utilizing the water simulation results [32]. Another research first predicts a sequence of normal maps, then generates the corresponding RGB images that show a looping animation of a garment as if it is blown in the wind [4]. Another line of research trains a generator to produce a motion field [6, 9, 15, 18, 19]. Our work is similar to these deep learning-based approaches in that a motion generator is utilized to synthesize the cinemagrams. The difference is that our method leverages the deep features of a pre-trained image generator and thus can generate plausible high-quality results without fine-tuning or training a separate synthesis network. At the same time, our method systematically improves the resolution of the generated cinemagrams to " + }, + { + "bbox": [ + 304, + 216, + 545, + 418 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 216, + 545, + 418 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 422, + 480, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 480, + 434 + ], + "type": "text", + "content": "2.2. Unconditional Video Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 713 + ], + "type": "text", + "content": "The task of video generation is known to be difficult because the generation process has to take into account both spatial and temporal information. To ease the problem, previous approaches [27, 36] decompose video generation into content and motion generation. These methods first predict the latent trajectory for motion, then generate a video from the set of predicted latent codes using the image generator. Instead of training the generation model from scratch, MoCoGAN-HD [34] and StyleVideoGAN [7] leverage a pre-trained image generator model, StyleGAN [12]. These methods use the property of well-constructed latent space to morph one frame toward the next frames [10]. Utilizing the image generation capability of a pre-trained generator, MoCoGAN-HD and StyleVideoGAN only need to train a motion trajectory generator, which greatly reduces the training time. These methods can generate high-quality videos in multiple domains such as faces, cars, and even sky. However, the content details are not well preserved, which has hindered the methods' application for cinematograph generation. Our method builds upon a similar concept of using a pre-trained StyleGAN. We leverage the deep features instead of the latent codes of StyleGAN to generate looping and content-preserving cinematographs." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7873" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 278, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 278, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 278, + 85 + ], + "type": "text", + "content": "2.3. Video Synthesis using pre-trained StyleGAN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 259 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 259 + ], + "type": "text", + "content": "A pre-trained generative model, StyleGAN [11-13], has been actively employed for downstream image and video editing applications. For GAN-based applications, one of the essential steps is to project the desired images or videos to the latent space of the pre-trained GAN model, which is known as GAN inversion. The latent codes are obtained using either an optimization technique [1, 2], a trained encoder [26, 35], or a hybrid of both approaches [16]. For video editing applications, previous methods [3, 29, 37, 42] have utilized a pre-trained image encoder to project all of the frames to the latent space. While most of the methods operate in " + }, + { + "bbox": [ + 46, + 91, + 289, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 46, + 91, + 289, + 259 + ], + "type": "text", + "content": " space [2], we opt to use both " + }, + { + "bbox": [ + 46, + 91, + 289, + 259 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 46, + 91, + 289, + 259 + ], + "type": "text", + "content": " and deep features [24, 43, 44] of the pre-trained StyleGAN to accurately reconstruct the original image and synthesize motion." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 272, + 108, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 272, + 108, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 108, + 285 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "text", + "content": "An overview of our method is shown in Figure 2. Given a landscape image " + }, + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "text", + "content": ", our method generates a seamlessly looping cinemagraph " + }, + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "inline_equation", + "content": "V = \\{\\hat{I}_0, \\dots, \\hat{I}_N\\}" + }, + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "text", + "content": " using a pre-trained StyleGAN. At the core of our method, we use a pre-trained generator without additional fine-tuning. The overall process is described below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 37, + 369, + 287, + 535 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 37, + 369, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 369, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 37, + 369, + 287, + 416 + ], + "type": "text", + "content": "(A) First, we project the landscape image into both the latent space and the feature space of StyleGAN. To this end, we train an encoder that outputs both latent codes " + }, + { + "bbox": [ + 37, + 369, + 287, + 416 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 37, + 369, + 287, + 416 + ], + "type": "text", + "content": " and intermediate features " + }, + { + "bbox": [ + 37, + 369, + 287, + 416 + ], + "type": "inline_equation", + "content": "D^{10}" + }, + { + "bbox": [ + 37, + 369, + 287, + 416 + ], + "type": "text", + "content": " of StyleGAN (Sec. 3.1)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 37, + 417, + 287, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 417, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 37, + 417, + 287, + 440 + ], + "type": "text", + "content": "(B) In addition, we predict a mask " + }, + { + "bbox": [ + 37, + 417, + 287, + 440 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 37, + 417, + 287, + 440 + ], + "type": "text", + "content": " to divide the image into static and dynamic regions (Sec. 3.2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 37, + 441, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 441, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 37, + 441, + 287, + 487 + ], + "type": "text", + "content": "(C) Next, we use a motion generator that accepts the landscape image " + }, + { + "bbox": [ + 37, + 441, + 287, + 487 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 37, + 441, + 287, + 487 + ], + "type": "text", + "content": " to synthesize a motion field " + }, + { + "bbox": [ + 37, + 441, + 287, + 487 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 37, + 441, + 287, + 487 + ], + "type": "text", + "content": ", which defines the position of each pixel in the future frames (Sec. 3.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 37, + 489, + 287, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 489, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 37, + 489, + 287, + 535 + ], + "type": "text", + "content": "(D) Lastly, we generate the final cinematograph frames using the pre-trained StyleGAN with deep feature warping (DFW) operation added in between the original layers. (Sec. 3.4)." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 540, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 564 + ], + "type": "text", + "content": "In the following sections, we will describe the details of each process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 574, + 141, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 141, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 141, + 586 + ], + "type": "text", + "content": "3.1. GAN Inversion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": "The first step for generating a cinematograph is to project the input image to the latent space of a pre-trained StyleGAN. This process is necessary because StyleGAN is an unconditional generator that cannot take a conditioning input. Many previous methods have used " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W} +" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": " [2] latent space, which is an extended space of native latent space " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": ". However, we observed that " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "w^{+} \\in \\mathcal{W}+" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": " is not expressive enough to reconstruct the original high-frequency details of the landscape images. Therefore, we chose to use the generated deep features " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "D^{10}" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": " of StyleGAN as well as " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": ". The use of " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "D^{10}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": "enables us to recover details in the original input as shown in Figure 8." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 96, + 547, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 547, + 157 + ], + "type": "text", + "content": "To project an image " + }, + { + "bbox": [ + 304, + 96, + 547, + 157 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 96, + 547, + 157 + ], + "type": "text", + "content": " to both the latent space and feature space and obtain " + }, + { + "bbox": [ + 304, + 96, + 547, + 157 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 304, + 96, + 547, + 157 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 96, + 547, + 157 + ], + "type": "inline_equation", + "content": "D^{10}" + }, + { + "bbox": [ + 304, + 96, + 547, + 157 + ], + "type": "text", + "content": ", we train an encoder " + }, + { + "bbox": [ + 304, + 96, + 547, + 157 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 96, + 547, + 157 + ], + "type": "text", + "content": " similarly to Yao et al. [43], in which the encoder takes " + }, + { + "bbox": [ + 304, + 96, + 547, + 157 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 96, + 547, + 157 + ], + "type": "text", + "content": " as input and predicts " + }, + { + "bbox": [ + 304, + 96, + 547, + 157 + ], + "type": "inline_equation", + "content": "(w^{+}, D^{10})" + }, + { + "bbox": [ + 304, + 96, + 547, + 157 + ], + "type": "text", + "content": ". Additional training details can be found in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 165, + 406, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 165, + 406, + 176 + ], + "spans": [ + { + "bbox": [ + 305, + 165, + 406, + 176 + ], + "type": "text", + "content": "3.2. Mask Prediction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 182, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 278 + ], + "type": "text", + "content": "We improve the quality of GAN inversion described in Sec. 3.1 by training an additional classifier that predicts a mask to separate the static and dynamic regions. Using the mask, the structure of the static regions can be preserved. To this end, we train a multi-layer-perceptron (MLP) classifier that accepts the deep features of StyleGAN as its input and outputs the mask that specifies each dynamic region in the image, as performed in DatasetGAN [46]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 280, + 546, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 546, + 411 + ], + "type": "text", + "content": "To train the classifier, we manually annotate 32 segmentation masks " + }, + { + "bbox": [ + 304, + 280, + 546, + 411 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 280, + 546, + 411 + ], + "type": "text", + "content": " of the selected images. We then follow the same procedure as DatasetGAN but using the deep features " + }, + { + "bbox": [ + 304, + 280, + 546, + 411 + ], + "type": "inline_equation", + "content": "D^{i}" + }, + { + "bbox": [ + 304, + 280, + 546, + 411 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 280, + 546, + 411 + ], + "type": "inline_equation", + "content": "i \\in \\{10,11,\\dots,18\\}" + }, + { + "bbox": [ + 304, + 280, + 546, + 411 + ], + "type": "text", + "content": ". We resize all these deep features followed by concatenating them in the channel dimension to construct the input feature " + }, + { + "bbox": [ + 304, + 280, + 546, + 411 + ], + "type": "inline_equation", + "content": "D^{*}" + }, + { + "bbox": [ + 304, + 280, + 546, + 411 + ], + "type": "text", + "content": ". In the end, paired data " + }, + { + "bbox": [ + 304, + 280, + 546, + 411 + ], + "type": "inline_equation", + "content": "(D^{*},S)" + }, + { + "bbox": [ + 304, + 280, + 546, + 411 + ], + "type": "text", + "content": " is constructed. An ensemble of 10 MLP classifiers is trained with a cross-entropy loss. To further improve the performance of the previous work, at inference, we also refine the predicted mask. For additional details, please refer to the supplementary material." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 419, + 419, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 419, + 419, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 419, + 419, + 430 + ], + "type": "text", + "content": "3.3. Motion Generation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "content": "To generate motion from a static landscape image " + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "content": ", we use an Eulerian motion field " + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "content": " which represents the immediate velocity of each pixel. To predict " + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "content": ", an imaged-to-image translation network [39] is trained as a motion generator. We train the network with paired data " + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "(I, M)" + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "content": " as performed in the previous method [9]. Adding controls in motion generation is also possible, either by using text prompts [19] or drawing arrows [18]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "content": "While the trained network works well to a certain degree, we observed that the predicted motion field " + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "content": " does not align with the boundaries between the static and dynamic regions and often contains motion in static regions. Thus, we further refine " + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "content": " by multiplying it with the segmentation mask " + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "content": ", which effectively removes most errors. The refined " + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "content": " is then used to simulate the motion of a pixel at time " + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 534, + 545, + 629 + ], + "type": "text", + "content": " using the Euler integration as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 639, + 545, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 545, + 668 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x _ {t + 1} = x _ {t} + M \\left(x _ {t}\\right), \\quad M \\left(x _ {t}\\right) = F _ {t \\rightarrow t + 1} \\left(x _ {t}\\right), \\\\ F _ {0 \\rightarrow t} (x _ {0}) = F _ {0 \\rightarrow t - 1} (x _ {0}) + M \\left(x _ {0} + F _ {0 \\rightarrow t - 1} (x _ {0})\\right), \\tag {1} \\\\ \\end{array}", + "image_path": "a44a27bca4e95760b4cb10bd101f70ea5f9f2c36f290d48cebd07e5c5da0583a.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " is the source pixel position at time 0, " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "F_{0\\rightarrow t}" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " is the displacement field that defines the displacement of the pixel position " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " to the pixel position " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7874" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 321, + 247 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 321, + 247 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 321, + 247 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 321, + 247 + ], + "type": "image", + "image_path": "4a67e97121df12534c46aef569f41f801346f128f3bc0015d37c5f820bb34afe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 250, + 236, + 258 + ], + "lines": [ + { + "bbox": [ + 182, + 250, + 236, + 258 + ], + "spans": [ + { + "bbox": [ + 182, + 250, + 236, + 258 + ], + "type": "text", + "content": "Motion Generation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 325, + 71, + 545, + 245 + ], + "blocks": [ + { + "bbox": [ + 325, + 71, + 545, + 245 + ], + "lines": [ + { + "bbox": [ + 325, + 71, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 325, + 71, + 545, + 245 + ], + "type": "image", + "image_path": "db2dd263e1cfeda263e26d7f00f69f00cf198b488f3b2ce313b548f3c3422577.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 250, + 407, + 258 + ], + "lines": [ + { + "bbox": [ + 334, + 250, + 407, + 258 + ], + "spans": [ + { + "bbox": [ + 334, + 250, + 407, + 258 + ], + "type": "text", + "content": "Cinemagraph Generation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": "Figure 2. Overview of StyleCineGAN. Given an input landscape image " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": ", our goal is to generate a cienemagraph using a fixed pre-trained StyleGAN " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": ". We project the image into both latent codes " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": " and deep features " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "D^{10}" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": ". Using the deep features " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "D^{*}" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": ", a mask predictor predicts a segmentation mask " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": ". To animate the input image, we use a motion generator to predict the motion field " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": " is refined using " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": ". Through Euler integration, " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": " produces the future and past displacement fields " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "F_{0\\rightarrow t}" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "F_{N\\rightarrow t}" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": ". To synthesize cienemagraph frames, we add a DFW layer in between the layers of " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": ". DFW refers to Eqns. 2 and 3. This modification enables the intermediate features of " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": " to be warped according to " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "F_{0\\rightarrow t}" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "F_{N\\rightarrow t}" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": " using a joint splatting method at different resolutions, specifically for the StyleGAN layers indexed with " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "i\\in [10,12,14,16,18]" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": ". The warped deep features are used to synthesize frames " + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "inline_equation", + "content": "\\hat{I}_t" + }, + { + "bbox": [ + 46, + 261, + 545, + 348 + ], + "type": "text", + "content": " resulting in the final cienemagraph video." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 358, + 190, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 190, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 190, + 371 + ], + "type": "text", + "content": "3.4. Cinemagraph Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 377, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 483 + ], + "type": "text", + "content": "After acquiring the displacement field " + }, + { + "bbox": [ + 46, + 377, + 287, + 483 + ], + "type": "inline_equation", + "content": "F_{0\\rightarrow t}" + }, + { + "bbox": [ + 46, + 377, + 287, + 483 + ], + "type": "text", + "content": " (Sec. 3.3), latent codes " + }, + { + "bbox": [ + 46, + 377, + 287, + 483 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 46, + 377, + 287, + 483 + ], + "type": "text", + "content": ", and deep features " + }, + { + "bbox": [ + 46, + 377, + 287, + 483 + ], + "type": "inline_equation", + "content": "D^{10}" + }, + { + "bbox": [ + 46, + 377, + 287, + 483 + ], + "type": "text", + "content": " of StyleGAN (Sec. 3.1), we feed these data into the pre-trained StyleGAN with our DFW layer added. We apply forward warping (or splatting) to the original content to generate a cinematograph. Here, the application of forward warping directly to an RGB image often results in images with tearing artifacts. To reduce the artifacts, we warp the deep features " + }, + { + "bbox": [ + 46, + 377, + 287, + 483 + ], + "type": "inline_equation", + "content": "D_0^i" + }, + { + "bbox": [ + 46, + 377, + 287, + 483 + ], + "type": "text", + "content": " of StyleGAN in different resolutions or scales:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 489, + 287, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 489, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 115, + 489, + 287, + 504 + ], + "type": "interline_equation", + "content": "D _ {t} ^ {i} = \\operatorname {W a r p} \\left(D _ {0} ^ {i}, F _ {0 \\rightarrow t}\\right), \\tag {2}", + "image_path": "181c5eb7e488de27e08a921a30392e21460ffad7034e3edb7c4565d9b6d69c0f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\text{Warp}" + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": " is the forward warping function and " + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "inline_equation", + "content": "D_{t}^{i}" + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": " is the warped deep feature at time " + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": ". We observed that warping a single deep feature (e.g., only " + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "inline_equation", + "content": "D_{0}^{10}" + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": ") results in blurry textures in the generated cinemagrams. Therefore, we opt to warp the multi-scale deep features; we call this operation a MSDFW. Specifically, we apply warping to the deep features " + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "inline_equation", + "content": "D_{0}^{i}" + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "inline_equation", + "content": "i \\in [10, 12, 14, 16, 18]" + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": ". The deep feature " + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "inline_equation", + "content": "D_{0}^{10}" + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": " is acquired using GAN inversion, and the consequent deep features " + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "inline_equation", + "content": "D_{0}^{12}, D_{0}^{14}, \\ldots" + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "inline_equation", + "content": "D_{0}^{18}" + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": " are subsequently generated using both " + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "inline_equation", + "content": "D_{0}^{10}" + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": ", as shown in the rightmost column of Figure 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "Looping video " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " with frame length " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "N + 1" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " is synthesized using the joint splating technique [9] using the future and past displacement fields " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "F_{0\\rightarrow t}" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "F_{N\\rightarrow t}" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": ", generated through Euler integration. The displacement fields are computed with " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " times of integration on " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "F_{0\\rightarrow t}" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "N - t" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " times of integration on " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "-M" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "F_{N\\rightarrow t}" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": ". The multi-scale fea" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 358, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 358, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 358, + 545, + 396 + ], + "type": "text", + "content": "ture " + }, + { + "bbox": [ + 305, + 358, + 545, + 396 + ], + "type": "inline_equation", + "content": "D_0^i" + }, + { + "bbox": [ + 305, + 358, + 545, + 396 + ], + "type": "text", + "content": " is warped in both directions and is composited to form " + }, + { + "bbox": [ + 305, + 358, + 545, + 396 + ], + "type": "inline_equation", + "content": "D_t^i" + }, + { + "bbox": [ + 305, + 358, + 545, + 396 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 305, + 358, + 545, + 396 + ], + "type": "inline_equation", + "content": "D_t^i" + }, + { + "bbox": [ + 305, + 358, + 545, + 396 + ], + "type": "text", + "content": " is computed as a weighted sum of " + }, + { + "bbox": [ + 305, + 358, + 545, + 396 + ], + "type": "inline_equation", + "content": "Warp(D_0^i,F_{0\\to t})" + }, + { + "bbox": [ + 305, + 358, + 545, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 358, + 545, + 396 + ], + "type": "inline_equation", + "content": "Warp(D_0^i,F_{N\\to t})" + }, + { + "bbox": [ + 305, + 358, + 545, + 396 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 407, + 545, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 407, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 315, + 407, + 545, + 468 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} D _ {t} ^ {i} \\left(x ^ {\\prime}\\right) = \\frac {\\sum_ {x \\in \\chi} \\alpha_ {t} \\cdot W a r p \\left(D _ {0} ^ {i} (x) , F _ {0 \\rightarrow t}\\right)}{\\sum_ {x \\in \\chi} \\alpha_ {t}} \\\\ + \\frac {\\sum_ {x \\in \\chi} \\left(1 - \\alpha_ {t}\\right) \\cdot W a r p \\left(D _ {0} ^ {i} (x) , F _ {N \\rightarrow t}\\right)}{\\sum_ {x \\in \\chi} \\left(1 - \\alpha_ {t}\\right)}, \\tag {3} \\\\ \\end{array}", + "image_path": "dcdae4980e12ab8309edf853b5f945e1a732b818a2bc50b8cb77c2c26029f7ec.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 477, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 477, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 477, + 545, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 477, + 545, + 514 + ], + "type": "inline_equation", + "content": "x \\in \\chi" + }, + { + "bbox": [ + 305, + 477, + 545, + 514 + ], + "type": "text", + "content": " is a set of pixels being mapped to the same destination pixel " + }, + { + "bbox": [ + 305, + 477, + 545, + 514 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 305, + 477, + 545, + 514 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 477, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 305, + 477, + 545, + 514 + ], + "type": "text", + "content": " is the looping weight defined as " + }, + { + "bbox": [ + 305, + 477, + 545, + 514 + ], + "type": "inline_equation", + "content": "(1 - \\frac{t}{N})" + }, + { + "bbox": [ + 305, + 477, + 545, + 514 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 514, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 514, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 305, + 514, + 545, + 548 + ], + "type": "text", + "content": "With the above DFW module, we can generate a video frame given the predicted deep features " + }, + { + "bbox": [ + 305, + 514, + 545, + 548 + ], + "type": "inline_equation", + "content": "D^{10}" + }, + { + "bbox": [ + 305, + 514, + 545, + 548 + ], + "type": "text", + "content": ", latent code " + }, + { + "bbox": [ + 305, + 514, + 545, + 548 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 305, + 514, + 545, + 548 + ], + "type": "text", + "content": ", and mask " + }, + { + "bbox": [ + 305, + 514, + 545, + 548 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 514, + 545, + 548 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 555, + 545, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 555, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 315, + 555, + 545, + 583 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {I} _ {t} = S \\odot \\left(G _ {\\text {w a r p}} \\left(D _ {0} ^ {1 0}, w ^ {+}, F _ {0 \\rightarrow t}, F _ {N \\rightarrow t}\\right)\\right) \\\\ + (\\mathbf {1} - S) \\odot I, \\tag {4} \\\\ \\end{array}", + "image_path": "1a4d9d93d89cdb7c4cfbbd6656e14d7f5371b358d364f152371ac4eb8b77bf4f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 594, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 545, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 594, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 305, + 594, + 545, + 628 + ], + "type": "text", + "content": " is an element-wise multiplication and " + }, + { + "bbox": [ + 305, + 594, + 545, + 628 + ], + "type": "inline_equation", + "content": "G_{\\text{warp}}" + }, + { + "bbox": [ + 305, + 594, + 545, + 628 + ], + "type": "text", + "content": " is a fixed pre-trained StyleGAN that incorporates our DFW module." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 632, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 545, + 689 + ], + "type": "text", + "content": "Style Interpolation In addition to motion generation, the change of appearance style is an additional feature often observed in a landscape cinematograph. This can be achieved by interpolating the latent code with the target latent code as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 363, + 700, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 700, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 363, + 700, + 545, + 714 + ], + "type": "interline_equation", + "content": "w _ {s} ^ {+} = w ^ {+} \\cdot (1 - \\beta) + w _ {t} ^ {+} \\cdot \\beta , \\tag {5}", + "image_path": "1ddb0d0d3bf3ef720a62a83b66d096600b1020a2722e501ff2584d2940dc79c8.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7875" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 187, + 287, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 187, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 287, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 187, + 287, + 247 + ], + "type": "inline_equation", + "content": "w_{s}^{+}" + }, + { + "bbox": [ + 46, + 187, + 287, + 247 + ], + "type": "text", + "content": " is the interpolated latent code, " + }, + { + "bbox": [ + 46, + 187, + 287, + 247 + ], + "type": "inline_equation", + "content": "w_{t}^{+}" + }, + { + "bbox": [ + 46, + 187, + 287, + 247 + ], + "type": "text", + "content": " is the latent code of the target style, and " + }, + { + "bbox": [ + 46, + 187, + 287, + 247 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 187, + 287, + 247 + ], + "type": "text", + "content": " is the interpolation weight. This is possible because the later layers of the StyleGAN only modify the color of the synthesized image while the original structure is preserved." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 248, + 287, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 287, + 285 + ], + "type": "text", + "content": "With modified Equation 4, we prevent the visual mismatch between the static and dynamic regions by also reflecting the changes to the static regions as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 294, + 287, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 294, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 79, + 294, + 287, + 324 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {I} _ {t} = S \\odot \\left(G _ {\\text {w a r p}} \\left(D _ {0} ^ {1 0}, w _ {s} ^ {+}, F _ {0 \\rightarrow t}, F _ {N \\rightarrow t}\\right)\\right) \\tag {6} \\\\ + (\\mathbf {1} - S) \\odot (I + \\Delta I _ {s}), \\\\ \\end{array}", + "image_path": "8c843e406f2adcde2638d042c231a2b0d62a30a01ef70b12d5d16220be800d37.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 326, + 287, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 326, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 100, + 326, + 287, + 341 + ], + "type": "interline_equation", + "content": "\\Delta I _ {s} = G \\left(D _ {0} ^ {1 0}, w _ {s} ^ {+}\\right) - G \\left(D _ {0} ^ {1 0}, w ^ {+}\\right), \\tag {7}", + "image_path": "dfe1a45fd88f8b95fb4b818664b28d1639f2a1fda2977fef81ee2f098ec88982.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 351, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 351, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 351, + 287, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 351, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\Delta I_{s}" + }, + { + "bbox": [ + 47, + 351, + 287, + 376 + ], + "type": "text", + "content": " is the color difference between the two images generated from latent codes " + }, + { + "bbox": [ + 47, + 351, + 287, + 376 + ], + "type": "inline_equation", + "content": "w_{s}^{+}" + }, + { + "bbox": [ + 47, + 351, + 287, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 351, + 287, + 376 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 47, + 351, + 287, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 388, + 128, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 128, + 402 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 128, + 402 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 409, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 287, + 506 + ], + "type": "text", + "content": "In this section, we compare our method with state-of-the-art landscape cinematograph generation methods [6, 9, 17, 19] (Sec. 4.1) and with an unconditional video generation method [34] (Sec. 4.2). In addition, we show the importance of the components used in our method through ablation studies (Sec. 4.3). To observe more various results, we recommend readers see the supplementary video. Example results are presented in Figures 1 and 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 514, + 287, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 287, + 538 + ], + "type": "text", + "content": "4.1. Comparisons with Cinemagraph Generation Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 546, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 288, + 714 + ], + "type": "text", + "content": "We compared our method with state-of-the-art cinemagraph generation methods: Animating Landscape (AL) [6], Deep Landscape (DL) [17], Eulerian Motion Field (EMF) [9], and Text2Cinemagraph (T2C) [19]. AL is a learning-based method that uses backward warping to generate cinemagraph frames. DL trains a style-based generator to synthesize images and dynamic motion in a scene. EMF and T2C train an encoder-decoder architecture paired with a motion generator to produce looping cinemagrams. We used official implementations of AL, DL, and T2C, and faithfully reproduced EMF based on the provided training details and hyper-parameters. In the following, we will examine the qualitative difference, evaluate the results using two metrics, and report the results of human perceptual studies." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 72, + 532, + 134 + ], + "blocks": [ + { + "bbox": [ + 47, + 134, + 289, + 180 + ], + "lines": [ + { + "bbox": [ + 47, + 134, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 134, + 289, + 180 + ], + "type": "text", + "content": "Figure 3. Generated cinematagraph results. This figure contains video clips, thus consider viewing it using Adobe Reader. The first two are cinematographs without appearance change, and the last two are cinematographs with appearance change." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 72, + 532, + 134 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 532, + 134 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 532, + 134 + ], + "type": "image", + "image_path": "0969637ceb9ca37a33559ba2e6c9207ff7047175487bea4ca60a64b79dfa815f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 136, + 531, + 199 + ], + "blocks": [ + { + "bbox": [ + 320, + 136, + 531, + 199 + ], + "lines": [ + { + "bbox": [ + 320, + 136, + 531, + 199 + ], + "spans": [ + { + "bbox": [ + 320, + 136, + 531, + 199 + ], + "type": "image", + "image_path": "39a0213e700ef9bc56ba9bc73a6eeddde37538bac4a9f16f4257e2c9eadcc04f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 320, + 201, + 531, + 264 + ], + "blocks": [ + { + "bbox": [ + 320, + 201, + 531, + 264 + ], + "lines": [ + { + "bbox": [ + 320, + 201, + 531, + 264 + ], + "spans": [ + { + "bbox": [ + 320, + 201, + 531, + 264 + ], + "type": "image", + "image_path": "2b2dff4bbec86926ff16f3a9189aec4cf3d1734c21855e584c6339d7e0200f38.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 320, + 266, + 531, + 329 + ], + "blocks": [ + { + "bbox": [ + 320, + 266, + 531, + 329 + ], + "lines": [ + { + "bbox": [ + 320, + 266, + 531, + 329 + ], + "spans": [ + { + "bbox": [ + 320, + 266, + 531, + 329 + ], + "type": "image", + "image_path": "fd20f8b3e221aecb6ce65ca529f33fe1dbf2ad802f07ce17154bbfe3efe03578.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 332, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 305, + 332, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 332, + 545, + 365 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison with state-of-the-art cinematograph generation methods. Please refer to the supplementary video for more examples." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 312, + 415, + 539, + 467 + ], + "blocks": [ + { + "bbox": [ + 306, + 373, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 306, + 373, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 545, + 407 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with the state-of-the-art cinematograph generation methods. We compared our method with AL, EMF, DL, and T2C. The best scores are bolded." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 415, + 539, + 467 + ], + "lines": [ + { + "bbox": [ + 312, + 415, + 539, + 467 + ], + "spans": [ + { + "bbox": [ + 312, + 415, + 539, + 467 + ], + "type": "table", + "html": "
MethodStatic ConsistencyMotion Quality
LPIPS↓MS-SSIM↑RMSE↓LPIPS↓MS-SSIM↑RMSE↓FID↓
AL0.04770.95248.89560.06170.681925.108953.6893
EMF0.01030.97236.04400.05330.710222.893245.8035
DL0.00710.99312.10440.05040.706221.801141.6374
T2C0.00630.97734.87850.01590.722421.278440.1186
Ours0.00620.99621.94300.01310.723720.894839.2113
", + "image_path": "491143a31de31d72610b9141bf036fd9cfeed33cf9e9ae62ad2c04a6ea78a207.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 480, + 545, + 720 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 545, + 720 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 545, + 720 + ], + "type": "text", + "content": "Qualitative Comparisons Figure 4 shows visual results from AL, EMF, DL, T2C, and our method. AL exhibits stretching artifacts (row 1) due to recurrent backward warping, whereas our method uses forward warping that prevents the appearance of such distortions. EMF struggles with precise reconstruction (row 2), because its encoder-decoder architecture is trained from scratch on a low-quality video dataset. In contrast, our method leverages a StyleGAN model pre-trained on high-quality images, resulting in cinemagrams with high perceptual quality. DL often loses fine details in textures (row 3) as it performs GAN inversion using the latent codes of StyleGAN, and fine-tunes the generator. Our method, utilizing deep features for GAN inversion, retains original texture details. T2C encounters difficulties in preserving high-frequency details, and accurate motion segmentation between static and dynamic regions (row 4). In contrast, our method utilizes a pre-trained StyleGAN for cinematograph generation, and uses its features for mask prediction, preserving both fine and structural details of the source image." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7876" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 215 + ], + "type": "text", + "content": "Quantitative Comparisons The evaluation was performed considering two aspects: (1) static consistency and (2) motion quality. Static consistency measures the consistency over time of non-movable regions such as buildings and mountains. Motion quality refers to both the image quality of the animated regions and the animation plausibility. For a fair comparison, we normalized the speed of motion in the generated results for all methods to match the average speed in the test dataset. AL, EMF, DL, and T2C were provided with the first frame " + }, + { + "bbox": [ + 46, + 72, + 287, + 215 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 46, + 72, + 287, + 215 + ], + "type": "text", + "content": " of the test video to generate " + }, + { + "bbox": [ + 46, + 72, + 287, + 215 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 72, + 287, + 215 + ], + "type": "text", + "content": " videos. Our method generates " + }, + { + "bbox": [ + 46, + 72, + 287, + 215 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 46, + 72, + 287, + 215 + ], + "type": "text", + "content": " results, thus we downsampled it to the size of " + }, + { + "bbox": [ + 46, + 72, + 287, + 215 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 72, + 287, + 215 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 286, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 286, + 323 + ], + "type": "text", + "content": "For a quantitative evaluation, we used 224 test videos from the Sky Time-Lapse dataset [41]. To measure the static consistency, we computed LPIPS [45], Root-Mean-Squared Error (RMSE), and MS-SSIM [40] between generated frame " + }, + { + "bbox": [ + 46, + 216, + 286, + 323 + ], + "type": "inline_equation", + "content": "\\hat{I}_n" + }, + { + "bbox": [ + 46, + 216, + 286, + 323 + ], + "type": "text", + "content": " and the input image " + }, + { + "bbox": [ + 46, + 216, + 286, + 323 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 46, + 216, + 286, + 323 + ], + "type": "text", + "content": " with the sky masked out. To measure the motion quality, we used the same evaluation metrics along with Frechet inception distance (FID) [28] between " + }, + { + "bbox": [ + 46, + 216, + 286, + 323 + ], + "type": "inline_equation", + "content": "\\hat{I}_n" + }, + { + "bbox": [ + 46, + 216, + 286, + 323 + ], + "type": "text", + "content": " and the ground-truth frame " + }, + { + "bbox": [ + 46, + 216, + 286, + 323 + ], + "type": "inline_equation", + "content": "I_n" + }, + { + "bbox": [ + 46, + 216, + 286, + 323 + ], + "type": "text", + "content": " with the static parts masked out." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 323, + 286, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 286, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 286, + 455 + ], + "type": "text", + "content": "Table 1 reveals the quantitative evaluation results. For both static consistency and motion quality, our method outperforms the other approaches. Our cinematographs represent static consistency better because use of the mask improves the accuracy in the detection of static regions. In addition, the quality of texture details are improved in the generated frames due to GAN inversion and the MSDFW based on the deep features of StyleGAN. The use of our motion generator and DFW leads to improved motion quality for each scene, compared with that of the results of previous approaches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 461, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 286, + 675 + ], + "type": "text", + "content": "User Study We conducted a user study with 17 participants to subjectively evaluate our method in comparison with previous cinematograph generation methods. We did not target any specific demographic groups, as the evaluation of cinematographs should not be dependent on particular demographic distributions. We conducted two evaluations: score evaluation and side-by-side comparison. Both evaluations assessed static consistency and motion quality. In the score evaluation, participants rated the video presented with a reference image on a 1-to-5 scale, with \"very bad\" being 1 and \"very good\" being 5. In the side-by-side comparison, participants selected the preferred cinematograph from two videos generated using different methods, presented with a reference image. For both evaluations, we used ten samples randomly chosen from the LHQ [31] dataset. To eliminate bias, distinct samples were used for each evaluation, and the positions of the cinematographs were randomized in all questions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "Tables 2 and 3 summarize the statistics of the user studies. For both score evaluation and human preference, our method outperforms the previous approaches by a substan" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 343, + 95, + 509, + 160 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 91 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 91 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 91 + ], + "type": "text", + "content": "Table 2. Human perceptual study results for score evaluation. The best scores are bolded." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 95, + 509, + 160 + ], + "lines": [ + { + "bbox": [ + 343, + 95, + 509, + 160 + ], + "spans": [ + { + "bbox": [ + 343, + 95, + 509, + 160 + ], + "type": "table", + "html": "
MethodStatic ConsistencyMotion Quality
AL1.59 ± 0.201.76 ± 0.45
EMF2.35 ± 0.892.24 ± 0.99
DL3.35 ± 0.833.25 ± 0.67
T2C3.75 ± 0.692.96 ± 0.66
Ours4.37 ± 0.183.86 ± 0.53
", + "image_path": "f8057d5bc126095eab08959849b67219fc35cf69206b4c6e6da21bc3815500da.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 341, + 207, + 507, + 272 + ], + "blocks": [ + { + "bbox": [ + 305, + 171, + 545, + 204 + ], + "lines": [ + { + "bbox": [ + 305, + 171, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 171, + 545, + 204 + ], + "type": "text", + "content": "Table 3. Human perceptual study results for side-by-side comparison. The percentage of ours chosen against each competing method is reported." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 341, + 207, + 507, + 272 + ], + "lines": [ + { + "bbox": [ + 341, + 207, + 507, + 272 + ], + "spans": [ + { + "bbox": [ + 341, + 207, + 507, + 272 + ], + "type": "table", + "html": "
MethodHuman Preference (Ours %)
Static ConsistencyMotion Quality
vs AL99.35%92.81%
vs EMF96.08%86.93%
vs DL94.77%84.97%
vs T2C77.78%73.86%
", + "image_path": "86e87569558fbeb09657d161c9158875cc390d45ad0ca0494cf9e3109fc09b26.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "text", + "content": "tial margin. We also performed statistical analysis (Kruskal-Wallis H-test) on the resulting evaluation scores. The results showed that our method achieved significantly higher scores in all pairs of comparisons with every previous method " + }, + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "inline_equation", + "content": "(p < 0.001" + }, + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "text", + "content": " in post-hoc analysis). Both user study results reveal the advantages of our method for generating cinematographs with high image quality and plausible animation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 374, + 542, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 374, + 542, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 374, + 542, + 387 + ], + "type": "text", + "content": "4.2. Comparisons with Video Generation Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 393, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 545, + 525 + ], + "type": "text", + "content": "We compared our approach with MoCoGAN-HD [34], a video generation method that predicts the trajectory within the latent space of a pre-trained StyleGAN. For content generation, we used the same pre-trained StyleGAN as that used by MoCoGAN-HD, which was trained on the Sky Time-lapse [41] dataset to generate " + }, + { + "bbox": [ + 304, + 393, + 545, + 525 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 304, + 393, + 545, + 525 + ], + "type": "text", + "content": " images. The comparison was made in an unconditional manner using 300 randomly sampled latent codes, at a resolution of " + }, + { + "bbox": [ + 304, + 393, + 545, + 525 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 304, + 393, + 545, + 525 + ], + "type": "text", + "content": ". We compared the first 60 frames of the generated videos with those of MoCoGAN-HD, as our method produces looping videos." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 529, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 624 + ], + "type": "text", + "content": "Qualitative Comparisons Figure 5 presents a qualitative comparison. As shown in the first and third rows, the videos generated using our method exhibit static consistencies, with the clouds moving while the ground remains static. In contrast, as shown in the second and fourth rows, the frames generated by MoCoGAN-HD deviate significantly from the original image, making the method unsuitable for cinematograph generation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Quantitative Comparisons We compared the content preservation ability of both methods by measuring LPIPS [45], RMSE, and MS-SSIM [40] between the first frame " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{I}_0" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " and the subsequent frames " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{I}_n" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ". Because the main content to be preserved in kinemagrams is the static regions, we defined content preservation as static consistency. We masked out the sky for all image frames according to the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7877" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 69, + 267, + 275 + ], + "blocks": [ + { + "bbox": [ + 65, + 69, + 267, + 275 + ], + "lines": [ + { + "bbox": [ + 65, + 69, + 267, + 275 + ], + "spans": [ + { + "bbox": [ + 65, + 69, + 267, + 275 + ], + "type": "image", + "image_path": "6019b077d7cbd2e52e32a8f09f5c75e1aed65c3732c400e006c828ad8c966ef1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 278, + 287, + 312 + ], + "lines": [ + { + "bbox": [ + 47, + 278, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 287, + 312 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison with the state-of-the-art video generation method, MoCoGAN-HD [34]. For more examples from this comparison, please refer to the supplementary video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 318, + 274, + 388 + ], + "blocks": [ + { + "bbox": [ + 59, + 318, + 274, + 388 + ], + "lines": [ + { + "bbox": [ + 59, + 318, + 274, + 388 + ], + "spans": [ + { + "bbox": [ + 59, + 318, + 274, + 388 + ], + "type": "image", + "image_path": "a794bf00fcc35105d17098df0a2401efecfbcc2e2b8133156a66ab6a93226d33.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 392, + 287, + 414 + ], + "lines": [ + { + "bbox": [ + 47, + 392, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 287, + 414 + ], + "type": "text", + "content": "Figure 6. Quantitative comparison of content preservation with state-of-the-art video generation method MoCoGAN-HD [34]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 422, + 288, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 288, + 567 + ], + "type": "text", + "content": "segmentation mask and compared only the static parts. We used 219 videos in which the scene could be horizontally divided into static and animated regions. LPIPS, RMSE, and MS-SSIM were all computed for each pair of " + }, + { + "bbox": [ + 46, + 422, + 288, + 567 + ], + "type": "inline_equation", + "content": "\\hat{I}_0" + }, + { + "bbox": [ + 46, + 422, + 288, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 422, + 288, + 567 + ], + "type": "inline_equation", + "content": "\\hat{I}_n" + }, + { + "bbox": [ + 46, + 422, + 288, + 567 + ], + "type": "text", + "content": ", and were averaged over the number of samples. Figure 6 shows that MoCoGAN-HD exhibits significant divergence in terms of LPIPS, RMSE, and MS-SSIM over time. This indicates that MoCoGAN-HD is unable to generate motion that preserves the content of the original image. In contrast, our method exhibits a small diverging trend, confirming its superiority in preserving the content of the original image over time, by utilizing deep features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 577, + 141, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 577, + 141, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 577, + 141, + 590 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our design choices for the proposed method, we conducted a series of ablation studies. Specifically, we focused on the effectiveness of our warping and GAN inversion methods. To evaluate warping, we compared the image frames generated with and without the use of 1) forward warping, 2) DFW, 3) MSDFW, and 4) segmentation mask. The first frames of 224 test videos from the Sky Time-Lapse [41] dataset were given as input to generate " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " videos, and we used the first 60 frames for the evaluation. For the evaluation of GAN inversion, we" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 321, + 70, + 424, + 258 + ], + "blocks": [ + { + "bbox": [ + 321, + 70, + 424, + 258 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 424, + 258 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 424, + 258 + ], + "type": "image", + "image_path": "eb5b3cdaa79bcc779d32b3d97b9387f720c1d68cd27af3559c9ae022d5cdd0ab.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 261, + 545, + 283 + ], + "lines": [ + { + "bbox": [ + 306, + 261, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 306, + 261, + 545, + 283 + ], + "type": "text", + "content": "Figure 7. Results of qualitative comparison in ablation study. Please see the supplementary video for the animated results." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 426, + 71, + 529, + 258 + ], + "blocks": [ + { + "bbox": [ + 426, + 71, + 529, + 258 + ], + "lines": [ + { + "bbox": [ + 426, + 71, + 529, + 258 + ], + "spans": [ + { + "bbox": [ + 426, + 71, + 529, + 258 + ], + "type": "image", + "image_path": "1ef88055d4ef5ca0a0e9822c91ff7031dd6c9f05db397b92441a12a59e9fd8f4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 323, + 287, + 526, + 376 + ], + "blocks": [ + { + "bbox": [ + 323, + 287, + 526, + 376 + ], + "lines": [ + { + "bbox": [ + 323, + 287, + 526, + 376 + ], + "spans": [ + { + "bbox": [ + 323, + 287, + 526, + 376 + ], + "type": "image", + "image_path": "a7f42b19cf715c06c7d4e413353f86ecd73d80170520fac7dc3477bb3184a3f1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 380, + 534, + 392 + ], + "lines": [ + { + "bbox": [ + 316, + 380, + 534, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 380, + 534, + 392 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison results on GAN inversion." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 396, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 445 + ], + "type": "text", + "content": "compared images reconstructed with and without the use of the deep features. A total of 256 images, 128 from the Sky Time-Lapse and 128 from the Eulerian [9] dataset were provided to generate " + }, + { + "bbox": [ + 304, + 396, + 545, + 445 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 396, + 545, + 445 + ], + "type": "text", + "content": " reconstructed images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 453, + 546, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 546, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 546, + 598 + ], + "type": "text", + "content": "Forward Warping We compared the image frames generated with and without forward warping (FW). For the case without FW, we warped the deep features of StyleGAN using backward warping. The first row in Figure 7 shows the qualitative results of this comparison. As revealed in the figure, without FW, the results usually contain unrealistically stretched textures. For quantitative comparisons, we computed LPIPS, MS-SSIM, and RMSE between " + }, + { + "bbox": [ + 304, + 453, + 546, + 598 + ], + "type": "inline_equation", + "content": "I_{n}" + }, + { + "bbox": [ + 304, + 453, + 546, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 453, + 546, + 598 + ], + "type": "inline_equation", + "content": "\\hat{I}_n" + }, + { + "bbox": [ + 304, + 453, + 546, + 598 + ], + "type": "text", + "content": ". Table 4 shows that excluding forward warping resulted in significant degradation in terms of MS-SSIM. This indicates that the stretched textures resulted in huge structural distortions in the generated images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "Deep Feature Warping We then compared frames generated with and without DFW. For the case without DFW, we directly warped RGB images using predicted motion fields for comparison. The second row in Figure 7 presents the qualitative comparison. As shown in the figure, the frame generated without DFW contained tearing artifacts in the animated regions. Table 4 shows an increase in the LPIPS score, which indicates that excluding DFW degraded the perceptual quality by introducing tearing artifacts." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7878" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 96, + 263, + 162 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "type": "text", + "content": "Table 4. Results of quantitative evaluation in ablation study. The best scores are bolded." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 96, + 263, + 162 + ], + "lines": [ + { + "bbox": [ + 72, + 96, + 263, + 162 + ], + "spans": [ + { + "bbox": [ + 72, + 96, + 263, + 162 + ], + "type": "table", + "html": "
MethodLPIPS↓MS-SSIM↑RMSE↓
Ours - Full0.05110.716521.8188
Ours - w/o FW0.05240.685322.9816
Ours - w/o DFW0.05370.690822.4993
Ours - w/o MSDFW0.06290.694622.3227
Ours - w/o Mask0.05640.698022.6213
", + "image_path": "0f1926ecc6f88fc688d361c20efbe8fa56288085c64664b12f1383dad8e827db.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 70, + 193, + 261, + 224 + ], + "blocks": [ + { + "bbox": [ + 47, + 168, + 286, + 190 + ], + "lines": [ + { + "bbox": [ + 47, + 168, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 286, + 190 + ], + "type": "text", + "content": "Table 5. Quantitative evaluation results on GAN inversion. The best scores are bolded." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 193, + 261, + 224 + ], + "lines": [ + { + "bbox": [ + 70, + 193, + 261, + 224 + ], + "spans": [ + { + "bbox": [ + 70, + 193, + 261, + 224 + ], + "type": "table", + "html": "
w+D10FID↓LPIPS↓RMSE↓MS-SSIM↑
✓-40.82760.015318.59010.7891
✓✓11.82310.00195.64100.9907
", + "image_path": "1bb14de5cb3824060f668bd40cc7402e33027587a02326286e5d4f0fb2c41a93.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 234, + 286, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 286, + 365 + ], + "type": "text", + "content": "Multi-scale Deep Feature Warping Image frames generated with and without MSDFW were also compared. For the case without MSDFW, only a single deep feature " + }, + { + "bbox": [ + 46, + 234, + 286, + 365 + ], + "type": "inline_equation", + "content": "D^{10}" + }, + { + "bbox": [ + 46, + 234, + 286, + 365 + ], + "type": "text", + "content": " was warped and propagated through the next blocks of StyleGAN. The results of the qualitative comparison are shown in the third row of Figure 7. As revealed in the figure, excluding MSDFW resulted in blurry textures in the dynamic region. The quantitative comparison reported in Table 4 shows a significant increase in the LPIPS score. This indicates that excluding MSDFW degraded the perceptual quality, especially the texture details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 369, + 286, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 286, + 487 + ], + "type": "text", + "content": "Segmentation Mask We compared the image frames generated with and without a segmentation mask. For the case without a mask, we used the initially predicted motion field for warping. The fourth row in Figure 7 shows the qualitative result of this comparison. As shown in the figure, excluding the mask resulted in erroneous movement in the static regions. The quantitative comparison reported in Table 4 reveals a significant degradation of MS-SSIM, which indicates the structural distortion caused by the erroneous motion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 492, + 286, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 286, + 635 + ], + "type": "text", + "content": "Deep Feature Inversion To evaluate the effectiveness of GAN inversion, we compared the images reconstructed with and without the use of the deep features " + }, + { + "bbox": [ + 46, + 492, + 286, + 635 + ], + "type": "inline_equation", + "content": "D^{10}" + }, + { + "bbox": [ + 46, + 492, + 286, + 635 + ], + "type": "text", + "content": ". For the case without deep feature inversion, only the latent codes " + }, + { + "bbox": [ + 46, + 492, + 286, + 635 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 46, + 492, + 286, + 635 + ], + "type": "text", + "content": " of a pre-trained StyleGAN were used to reconstruct the input landscape image. Figure 8 shows a qualitative result of this comparison. As revealed in the figure, using only " + }, + { + "bbox": [ + 46, + 492, + 286, + 635 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 46, + 492, + 286, + 635 + ], + "type": "text", + "content": " failed to accurately reconstruct the details of the original images. For quantitative comparisons, FID, LPIPS, RMSE, and MS-SSIM were computed between " + }, + { + "bbox": [ + 46, + 492, + 286, + 635 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 492, + 286, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 492, + 286, + 635 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 46, + 492, + 286, + 635 + ], + "type": "text", + "content": ". Table 5 reveals significant improvements in the perceptual quality of the results when reconstructed using the deep features " + }, + { + "bbox": [ + 46, + 492, + 286, + 635 + ], + "type": "inline_equation", + "content": "D^{10}" + }, + { + "bbox": [ + 46, + 492, + 286, + 635 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 648, + 214, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 648, + 214, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 648, + 214, + 661 + ], + "type": "text", + "content": "5. Limitations and Future Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Our mask predictor and motion generator generally performed well for landscape images. However, automatic prediction cannot be accurate for all images because most landscape images contain inherent ambiguity in their motion" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 72, + 424, + 120 + ], + "blocks": [ + { + "bbox": [ + 316, + 92, + 325, + 100 + ], + "lines": [ + { + "bbox": [ + 316, + 92, + 325, + 100 + ], + "spans": [ + { + "bbox": [ + 316, + 92, + 325, + 100 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 72, + 424, + 120 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 424, + 120 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 424, + 120 + ], + "type": "image", + "image_path": "0a3178b499e93c16913a8a8349e6d82741ef8a510389777e8ab9dc7d67285d5f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 440, + 72, + 487, + 119 + ], + "blocks": [ + { + "bbox": [ + 429, + 92, + 439, + 100 + ], + "lines": [ + { + "bbox": [ + 429, + 92, + 439, + 100 + ], + "spans": [ + { + "bbox": [ + 429, + 92, + 439, + 100 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 72, + 487, + 119 + ], + "lines": [ + { + "bbox": [ + 440, + 72, + 487, + 119 + ], + "spans": [ + { + "bbox": [ + 440, + 72, + 487, + 119 + ], + "type": "image", + "image_path": "99266baccdda3e0ba47175c59d0df00a990c5ee099f3af6ed316f18498209f98.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 123, + 545, + 156 + ], + "lines": [ + { + "bbox": [ + 305, + 123, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 305, + 123, + 545, + 156 + ], + "type": "text", + "content": "Figure 9. Limitations of StyleCineGAN: (a) automatic prediction of motion cannot be accurate for all images, and (b) the motion of a very thin structured object is hard to be isolated." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 488, + 72, + 534, + 119 + ], + "blocks": [ + { + "bbox": [ + 488, + 72, + 534, + 119 + ], + "lines": [ + { + "bbox": [ + 488, + 72, + 534, + 119 + ], + "spans": [ + { + "bbox": [ + 488, + 72, + 534, + 119 + ], + "type": "image", + "image_path": "66837910d83aaf923e84f8e453c57f1e4b218fed97a2f2fc7b454e387d80de0e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 163, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 545, + 281 + ], + "type": "text", + "content": "directions except for some obvious cases (e.g., waterfalls). The failure case is illustrated in Figure 9 (a). User-defined motion hints can be used to resolve such ambiguities and provide further control capability during the generation process [18, 19]. In addition, it is hard for our method to isolate the motion of a very thin structured object placed within the animated region as shown in Figure 9 (b). This is because our method performs warping of features at multiple resolutions, in which low-resolution features cannot spatially represent the thin structures." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 282, + 545, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 545, + 379 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 545, + 379 + ], + "type": "text", + "content": "In this work, we mainly focused on animating the landscape images, particularly skies and fluids, while putting other types of animations outside the scope. In future developments, we would like to expand the capabilities to include other forms of motion. Investigating the rotating hands of a clock, the playing arm of a guitarist moving up and down, and a flag or the wings of a bird fluttering for cinematograph generation would be a very interesting direction to pursue." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 388, + 378, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 388, + 378, + 400 + ], + "spans": [ + { + "bbox": [ + 306, + 388, + 378, + 400 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 406, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 545, + 573 + ], + "type": "text", + "content": "We proposed the first approach that leverages a pre-trained StyleGAN for high-quality one-shot landscape cinematograph generation. In contrast to previous studies, our method does not require training a large image generator from scratch, and also systematically improves the resolution of the generated cinematographs to " + }, + { + "bbox": [ + 304, + 406, + 545, + 573 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 406, + 545, + 573 + ], + "type": "text", + "content": ". At the core of our method, we utilized the deep features of a pretrained StyleGAN, because those features can help preserve spatial information and encode both high-level semantic and low-level style appearances. Using our MSDFW approach, we applied the predicted motion to the deep feature space of StyleGAN. Both qualitative and quantitative results confirm that our method substantially outperforms existing baselines." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements This research was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023 (Project Name: Development of Universal Fashion Creation Platform Technology for Avatar Personality Expression, Project Number: RS-2023-00228331, Contribution Rate: " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " ) and Korea Institute for Advancement of Technology(KIAT) grant funded by the Korea Government(MOTIE) (Project Name: The Competency Development Program for Industry Specialist, Project Number: P0012746, Contribution Rate: " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7879" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 167 + ], + "type": "text", + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? CoRR, abs/1911.11544, 2019.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 287, + 213 + ], + "type": "text", + "content": "[3] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3. arXiv preprint arXiv:2201.13433, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 287, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 266 + ], + "type": "text", + "content": "[4] Hugo Bertiche, Niloy J. Mitra, Kuldeep Kulkarni, ChunHao Paul Huang, Tuanfeng Y. Wang, Meysam Madadi, Sergio Escalera, and Duygu Ceylan. Blowing in the wind: Cyclenet for human cinematographs from still images, 2023. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 287, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 287, + 313 + ], + "type": "text", + "content": "[5] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. In ACM SIGGRAPH 2005 Papers, pages 853-860. 2005. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 367 + ], + "type": "text", + "content": "[6] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. ACM Transactions on Graphics (Proc. of SIGGRAPH ASIA 2019), 38(6):175:1-175:19, 2019. 1, 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 401 + ], + "type": "text", + "content": "[7] Gereon Fox, Ayush Tewari, Mohamed Elgharib, and Christian Theobalt. Stylevideogan: A temporal generative model using a pretrained stylegan, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 402, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 287, + 456 + ], + "type": "text", + "content": "[8] Tavi Halperin, Hanit Hakim, Orestis Vantzos, Gershon Hochman, Netai Benaim, Lior Sassy, Michael Kupchik, Ofir Bibi, and Ohad Fried. Endless loops: detecting and animating periodic patterns in still images. ACM Transactions on Graphics (TOG), 40(4):1-12, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 457, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 457, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 53, + 457, + 287, + 511 + ], + "type": "text", + "content": "[9] Aleksander Holynski, Brian L. Curless, Steven M. Seitz, and Richard Szeliski. Animating pictures with eulerian motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5810-5819, 2021. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "type": "text", + "content": "[10] Ali Jahanian, Lucy Chai, and Phillip Isola. \"On the \"steerability\" of generative adversarial networks. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 546, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 579 + ], + "type": "text", + "content": "[11] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. CoRR, abs/1710.10196, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "text", + "content": "[12] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "text", + "content": "[13] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "text", + "content": "[14] Alexander Kristoffersen. Loopnerf: Exploring temporal compression for 3d video textures. Master's thesis, EECS Department, University of California, Berkeley, 2023. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[15] Xingyi Li, Zhiguo Cao, Huiqiang Sun, Jianming Zhang, Ke Xian, and Guosheng Lin. 3d cinematography from a single" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4595-4605, 2023. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 108, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 152 + ], + "type": "text", + "content": "[16] Ji Lin, Richard Zhang, Frieder Ganz, Song Han, and JunYan Zhu. Anycost gans for interactive image synthesis and editing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 154, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 154, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 308, + 154, + 545, + 208 + ], + "type": "text", + "content": "[17] Elizaveta Logacheva, Roman Suvorov, Oleg Khomenko, Anton Mashikhin, and Victor Lempitsky. Deeplandscape: Adversarial modeling of landscape videos. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 210, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 210, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 308, + 210, + 545, + 255 + ], + "type": "text", + "content": "[18] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 256, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 256, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 308, + 256, + 545, + 289 + ], + "type": "text", + "content": "[19] Aniruddha Mahapatra, Aliaksandr Siarohin, Hsin-Ying Lee, Sergey Tulyakov, and Jun-Yan Zhu. Text-guided synthesis of eulerian cinemagrams. 2023. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 291, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 291, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 308, + 291, + 545, + 335 + ], + "type": "text", + "content": "[20] Makoto Okabe, Ken Anjyo, Takeo Igarashi, and Hans-Peter Seidel. Animating pictures of fluid using video examples. In Computer Graphics Forum, pages 677-686. Wiley Online Library, 2009. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 337, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 337, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 337, + 545, + 381 + ], + "type": "text", + "content": "[21] Makoto Okabe, Ken Anjyor, and Rikio Onai. Creating fluid animation from a single image using video database. In Computer Graphics Forum, pages 1973-1982. Wiley Online Library, 2011." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 383, + 545, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 416 + ], + "type": "text", + "content": "[22] Makoto Okabe, Yoshinori Dobashi, and Ken Anjyo. Animating pictures of water scenes using video retrieval. The Visual Computer, 34(3):347-358, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 418, + 545, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 418, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 308, + 418, + 545, + 473 + ], + "type": "text", + "content": "[23] Ehsan Pajouheshgar, Yitao Xu, Tong Zhang, and Sabine Süssstrunk. Dynca: Real-time dynamic texture synthesis using neural cellular automata. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 475, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 545, + 529 + ], + "type": "text", + "content": "[24] Gaurav Parmar, Yijun Li, Jingwan Lu, Richard Zhang, Jun-Yan Zhu, and Krishna Kumar Singh. Spatially-adaptive multilayer selection for gan inversion and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 531, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 531, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 308, + 531, + 545, + 576 + ], + "type": "text", + "content": "[25] Ekta Prashnani, Maneli Noorkami, Daniel Vaquero, and Pradeep Sen. A phase-based approach for animating images using video examples. In Computer Graphics Forum, pages 303-311. Wiley Online Library, 2017. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 577, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 577, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 577, + 545, + 632 + ], + "type": "text", + "content": "[26] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 634, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 634, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 634, + 545, + 678 + ], + "type": "text", + "content": "[27] Masaki Saito, Shunta Saito, Masanori Koyama, and Sosuke Kobayashi. Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal gan, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "text", + "content": "[28] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7880" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Kwanggyoon Seo, Seoung Wug Oh, Jingwan Lu, Joon-Young Lee, Seonghyeon Kim, and Junyong Noh. Styleportraitvideo: Editing portrait videos with expression optimization. 41(7), 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 288, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 288, + 163 + ], + "type": "text", + "content": "[30] Claude E. Shannon. Coding Theorems for a Discrete Source With a Fidelity Criterion Institute of Radio Engineers, International Convention Record, vol. 7, 1959., pages 325-350. 1993. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 288, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 288, + 198 + ], + "type": "text", + "content": "[31] Ivan Skorokhodov, Grigorii Sotnikov, and Mohamed Elhoseiny. Aligning latent and image spaces to connect the unconnectable. arXiv preprint arXiv:2104.06954, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 200, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 200, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 200, + 288, + 255 + ], + "type": "text", + "content": "[32] Ryusuke Sugimoto, Mingming He, Jing Liao, and Pedro V. Sander. Water simulation and rendering from a still photograph. In SIGGRAPH Asia 2022 Conference Papers, New York, NY, USA, 2022. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 257, + 288, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 257, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 48, + 257, + 288, + 301 + ], + "type": "text", + "content": "[33] Matthew Tesfaldet, Marcus A. Brubaker, and Konstantinos G. Derpanis. Two-stream convolutional networks for dynamic texture synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 303, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 303, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 303, + 288, + 358 + ], + "type": "text", + "content": "[34] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. In International Conference on Learning Representations, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "type": "text", + "content": "[35] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4): 1-14, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 406, + 288, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 406, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 288, + 450 + ], + "type": "text", + "content": "[36] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1526-1535, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 453, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 288, + 508 + ], + "type": "text", + "content": "[37] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. Stitch it in time: Gan-based facial editing of real videos. In SIGGRAPH Asia 2022 Conference Papers, New York, NY, USA, 2022. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 510, + 288, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 288, + 553 + ], + "type": "text", + "content": "[38] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 288, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 610 + ], + "type": "text", + "content": "[39] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 612, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 288, + 656 + ], + "type": "text", + "content": "[40] Zhou Wang, Eero P. Simoncelli, and Alan Conrad Bovik. Multiscale structural similarity for image quality assessment. The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, 2:1398-1402 Vol.2, 2003. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[41] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 329 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[42] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally consistent semantic video editing. pages 357-374. Springer, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[43] Xu Yao, Alasdair Newson, Yann Gousseau, and Pierre Hellier. A style-based gan encoder for high fidelity reconstruction of images and videos. European conference on computer vision, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 228 + ], + "type": "text", + "content": "[44] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 85-101. Springer, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "text", + "content": "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 263, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 545, + 329 + ], + "type": "text", + "content": "[46] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10145-10155, 2021. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "7881" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/8c37e125-7381-4a71-8f2b-015c3d8a066a_content_list.json b/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/8c37e125-7381-4a71-8f2b-015c3d8a066a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b1ad9ae05b7fd2f469458459d0a4213daf21dadc --- /dev/null +++ b/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/8c37e125-7381-4a71-8f2b-015c3d8a066a_content_list.json @@ -0,0 +1,1812 @@ +[ + { + "type": "text", + "text": "SuGaR: Surface-Aligned Gaussian Splitting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering", + "text_level": 1, + "bbox": [ + 135, + 130, + 833, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Antoine Guédon Vincent Lepetit \nLIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, France \nhttps://anttwo.github.io/sugar/", + "bbox": [ + 243, + 203, + 725, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6673930ff670e4525dde86ea891322dc4edf5ae835eac2957ca13ae2f137f242.jpg", + "image_caption": [ + "Figure 1. We introduce a method that extracts accurate and editable meshes from 3D Gaussian Splitting representations within minutes on a single GPU. The meshes can be edited, animated, composited, etc. with very realistic Gaussian Splitting rendering, offering new possibilities for Computer Graphics. Note for example that we changed the posture of the robot between the captured scene on the bottom left and the composited scene on the right. The supplementary material provides more examples, including a video illustrating our results." + ], + "image_footnote": [], + "bbox": [ + 78, + 267, + 202, + 429 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1e71b34782521a295dba22014fb33764264018e29489fec3a41da095d26a740b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 268, + 504, + 429 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/10648b5890419237a4fb6653f09ebb64111dc67816d3d828c0fbbb2c4b6de2cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 268, + 890, + 429 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 510, + 313, + 525 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a method to allow precise and extremely fast mesh extraction from 3D Gaussian Splatting [15]. Gaussian Splatting has recently become very popular as it yields realistic rendering while being significantly faster to train than NeRFs. It is however challenging to extract a mesh from the millions of tiny 3D Gaussians as these Gaussians tend to be unorganized after optimization and no method has been proposed so far. Our first key contribution is a regularization term that encourages the Gaussians to align well with the surface of the scene. We then introduce a method that exploits this alignment to extract a mesh from the Gaussians using Poisson reconstruction, which is fast, scalable, and preserves details, in contrast to the Marching Cubes algorithm usually applied to extract meshes from Neural SDFs. Finally, we introduce an optional refinement strategy that binds Gaussians to the surface of the mesh, and jointly optimizes these Gaussians and the mesh through Gaussian splatting rendering. This enables easy editing, sculpting, animating, and relighting of the Gaussians by manipulating the mesh instead of the Gaussians themselves. Retrieving such an editable mesh for realistic rendering is done within minutes with our method, compared to hours with the state-of-the-art method on SDFs, while providing a better rendering quality.", + "bbox": [ + 73, + 531, + 472, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 510, + 630, + 525 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "After NeRFs [22], 3D Gaussian Splatting [15] has recently become very popular for capturing a 3D scene and rendering it from novel points of view. 3D Gaussian Splatting optimizes the positions, orientations, appearances (represented as spherical harmonics), and alpha blending of many tiny 3D Gaussians on the basis of a set of training images of the scene to capture the scene geometry and appearance. Because rendering the Gaussians is much faster than rendering a neural field, 3D Gaussian Splatting is much faster than NeRFs and can capture a scene in a few minutes.", + "bbox": [ + 496, + 542, + 890, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While the Gaussians allow very realistic renderings of the scene, it is still however challenging to extract the surface of the scene from them: As shown in Figure 3, after optimization by 3D Gaussian Splatting, the Gaussians do not take an ordered structure in general and do not correspond well to the actual surface of the scene. In addition to the surface itself, it is also often desirable to represent the scene as a mesh, which remains the representation of choice in many pipelines: A mesh-based representation allows for powerful tools for editing, sculpting, animating, and relighting the scene. Because the Gaussians after Gaussian Splatting are unstructured, it is very challenging to extract a mesh from them. Note that this is also challenging with NeRFs albeit for different reasons.", + "bbox": [ + 496, + 695, + 892, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5354", + "bbox": [ + 482, + 943, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/965e4b706f98280301b6f99dcf08dc1d3cc1bda348d4aebf93eb5600c76ab7b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 89, + 272, + 170 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/63849b2eb6568bab393e1eb9db3e9606a0f34ca11c35142170ea5710115c3897.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 89, + 473, + 170 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7f0927d0b0e6405cd930616ebc9dd362849d6074393daefa14f02ea6d4638a45.jpg", + "image_caption": [ + "Figure 2. Our algorithm can extract a highly detailed mesh from any 3D Gaussian Splatting scene [15] within minutes on a single GPU (top: Renderings of our meshes without texture, bottom: Renderings of the meshes with bound Gaussians)." + ], + "image_footnote": [], + "bbox": [ + 86, + 172, + 272, + 253 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/414eed21c26dc56e63563eb1defede617594decabb74654cbaab4632ce6bb7ca.jpg", + "image_caption": [ + "Figure 3. Extracting a mesh from Gaussians. Without regularization, the Gaussians have no special arrangement after optimization, which makes extracting a mesh very difficult. Without our regularization term, Marching Cubes fail to extract an acceptable mesh. With our regularization term, Marching Cubes recover an extremely noisy mesh even with a very fine 3D grid. Our scalable extraction method obtains a mesh even without our regularization term. Still, the mesh is noisy. By contrast, our full method succeeds in reconstructing an accurate mesh very efficiently." + ], + "image_footnote": [], + "bbox": [ + 287, + 172, + 473, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we first propose a regularization term that encourages the Gaussians to be well distributed over the scene surface so that the Gaussians capture much better the scene geometry, as shown in Figure 3. Our approach is to derive a volume density from the Gaussians under the assumption that the Gaussians are flat and well distributed over the scene surface. By minimizing the difference between this density and the actual one computed from the Gaussians during optimization, we encourage the 3D Gaussians to represent well the surface geometry.", + "bbox": [ + 75, + 351, + 467, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thanks to this regularization term, it becomes easier to extract a mesh from the Gaussians. In fact, since we introduce a density function to evaluate our regularization term, a natural approach would be to extract level sets of this density function. However, Gaussian Splatting performs densification in order to capture details of the scene with high fidelity, which results in a drastic increase in the number of Gaussians. Real scenes typically end up with one or several millions of 3D Gaussians with different scales and rotations, the majority of them being extremely small in order to reproduce texture and details in the scene. This results in a density function that is close to zero almost everywhere, and the Marching Cubes algorithm [21] fails to extract proper level sets of such a sparse density function even with a fine voxel grid, as also shown in Figure 3.", + "bbox": [ + 75, + 500, + 467, + 719 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instead, we introduce a method that very efficiently samples points on the visible part of a level set of the density function, allowing us to run the Poisson reconstruction algorithm [14] on these points to obtain a triangle mesh. This approach is scalable, by contrast with the Marching Cubes algorithm for example, and reconstructs a surface mesh within minutes on a single GPU, compared to other state of the art methods relying on Neural SDFs for extracting meshes from radiance fields, that require at least 24 hours on one GPU [20, 36, 38, 39] and rely on multiple GPUs to speed up the process [26].", + "bbox": [ + 75, + 722, + 467, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As illustrated in Figures 2 and 4, our method produces", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fab143a17035a4621c16f71a04dc960390996ce00a5ee09eef27a7f69788960d.jpg", + "image_caption": [ + "without our regularization term" + ], + "image_footnote": [], + "bbox": [ + 514, + 101, + 632, + 165 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/62c6b2f08a971899191a5b562162f4a4a0d84c2fac41173d6dd4fdd4d21db07f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 101, + 756, + 165 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ad83aca945c2ef13508a04bdafb0e6aa0808eddaf7faabbba997f79174034c89.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 758, + 101, + 875, + 165 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4015513f1d9b4cc0061301924259fc7103dea1c2a531de2b078dd3f682cdbdbe.jpg", + "image_caption": [ + "with our regularization term", + "zoom on Gaussians on a planar surface" + ], + "image_footnote": [], + "bbox": [ + 514, + 181, + 632, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f9718805da1261a9b074b2a8f184814cf14f0acfe7762c7ef3b4057ab334615d.jpg", + "image_caption": [ + "mesh with Marching Cubes" + ], + "image_footnote": [], + "bbox": [ + 637, + 181, + 756, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/38688d6186e8082e57213c850b4fedc5c366bbb99ad6bf40978505ff29e8c3eb.jpg", + "image_caption": [ + "mesh with our extraction method" + ], + "image_footnote": [], + "bbox": [ + 758, + 181, + 874, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "high quality meshes. The challenge is in efficiently identifying points lying on the level set. To do this, we rely on the Gaussians depth maps seen from the training viewpoints. These depth maps can be obtained by extending the Gaussian Splatting rasterizer, and we show how to accurately sample points on the level set starting from these depth maps.", + "bbox": [ + 496, + 436, + 890, + 539 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, after extracting this mesh, we propose an optional refinement strategy that jointly optimizes the mesh and a set of 3D Gaussians through Gaussian splatting rendering only. This optimization enables high-quality rendering of the mesh using Gaussian splatting rendering rather than traditional textured mesh rendering. This results in higher performance in terms of rendering quality than other radiance field models relying on an underlying mesh at inference [6, 26, 39]. As shown in Figure 1, this makes possible the use of traditional mesh-editing tools for editing a Gaussian Splatting representation of a scene, offering endless possibilities for Computer Graphics.", + "bbox": [ + 496, + 541, + 890, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are:", + "bbox": [ + 517, + 719, + 764, + 732 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- a regularization term that makes the Gaussians capture accurately the geometry of the scene;", + "- an efficient algorithm that extracts an accurate mesh from the Gaussians within minutes;", + "- a method to bind the Gaussians to the mesh, resulting in a more accurate mesh, higher rendering quality than state of the art methods using a mesh for Novel View Synthesis [6, 26, 39], and allowing editing the scene in many different ways." + ], + "bbox": [ + 500, + 736, + 890, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We call our approach SuGaR. In the remainder of the paper, we discuss related work, give a brief overview of vanilla", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5355", + "bbox": [ + 482, + 943, + 514, + 954 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Gaussian Splatting, describe SuGaR, and compare it to the state of the art.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 137, + 218, + 152 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-based rendering (IBR) methods rely on a set of two-dimensional images of a scene to generate a representation of the scene and render novel views. The very first novel-view synthesis approaches were based on light fields [19], and developed the concept of volume rendering for novel views. Their work emphasized the importance of efficiently traversing volumetric data to produce realistic images.", + "bbox": [ + 75, + 164, + 467, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Various scene representations have been proposed since, such as triangle meshes, point clouds, voxel grids, multiplane images, or neural implicit functions.", + "bbox": [ + 76, + 270, + 467, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditional mesh-based IBR methods. Structure-from-motion (SfM) [32] and subsequent multi-view stereo (MVS) [10] allow for 3D reconstruction of surfaces, leading to the development of several view synthesis algorithms relying on triangle meshes as the primary 3D representation of scenes. Such algorithms consider textured triangles or warp and blend captured images on the mesh surface to generate novel views [4, 12, 37]. [29, 30] consider deep learning-based mesh representations for better view synthesis, bridging the gap between traditional graphics and modern machine learning techniques. While these mesh-based methods take advantage of existing graphics hardware and software for efficient rendering, they struggle with the capture of accurate geometry and appearance in complex regions.", + "bbox": [ + 75, + 337, + 467, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Volumetric IBR methods. Volumetric methods use voxel grids, multiplane images, or neural networks to represent scenes as continuous volumetric functions of density and color. Recently, Neural Radiance Fields (NeRF) [22] introduced a novel scene representation based on a continuous volumetric function parameterized by a multilayer perceptron (MLP). NeRF produces photorealistic renderings with fine details and view-dependent effects, achieved through volumetric ray tracing. However, the original NeRF is computationally expensive and memory intensive.", + "bbox": [ + 75, + 585, + 467, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these challenges, several works have improved NeRF's performance and scalability. These methods leverage discretized or sparse volumetric representations like voxel grids and hash tables as ways to store learnable features acting as positional encodings for 3D points [5, 13, 23, 34, 41], hierarchical sampling strategies [2, 11, 28, 40], or low-rank approximations [5]. However, they still rely on volumetric ray marching, which is incompatible with standard graphics hardware and software designed for rendering polygonal surfaces. Recent works have proposed modifying the NeRF's representation", + "bbox": [ + 75, + 736, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of geometry and emitted radiance to allow for better reconstruction of specular materials [35] or relighting the scene through an explicit decomposition into material and lighting properties [3, 18, 33, 43].", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hybrid IBR methods. Some methods build on differentiable rendering to combine the advantages of mesh-based and volumetric methods, and allow for surface reconstruction as well as better editability. They use a hybrid volume-surface representation, which enables high-quality meshes suitable for downstream graphics applications while efficiently modeling view-dependent appearance. In particular, some works optimize neural signed distance functions (SDF) by training neural radiance fields in which the density is derived as a differentiable transformation of the SDF [7, 8, 20, 24, 36, 38]. A triangle mesh can finally be reconstructed from the SDF by applying the Marching Cubes algorithm [21]. However, most of these methods do not target real-time rendering.", + "bbox": [ + 496, + 172, + 890, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Alternatively, other approaches \"bake\" the rendering capacity of an optimized NeRF or neural SDF into a much efficient structure relying on an underlying triangle mesh [6] that could benefit from the traditional triangle rasterization pipeline. In particular, the recent BakedSDF [39] reconstructs high quality meshes by optimizing a full neural SDF model, baking it into a high-resolution triangle mesh that combines mesh rendering for interpolating features and deep learning to translate these features into images, and finally optimizes a view-dependent appearance model.", + "bbox": [ + 496, + 383, + 890, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, even though it achieves real-time rendering and produces impressive meshes of the surface of the scene, this model demands training a full neural SDF with an architecture identical to Mip-NeRF360 [1], which necessitates 48 hours of training.", + "bbox": [ + 496, + 532, + 890, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarly, the recent method NeRFMeshing [26] proposes to also bake any NeRF model into a mesh structure, achieving real-time rendering. However, the meshing performed in this method lowers the quality of the rendering and results in a PSNR much lower than our method. Additionally, this method still requires training a full NeRF model beforehand, and needs approximately an hour of training on 8 V100 NVIDIA GPUs to allow for mesh training and extraction.", + "bbox": [ + 496, + 608, + 890, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method is much faster at retrieving a 3D mesh from 3D Gaussian Splatting, which is itself much faster than NeRFs. As our experiments show, our rendering done by bounding Gaussians to the mesh results in higher quality than previous solutions based on meshes.", + "bbox": [ + 496, + 743, + 890, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point-based IBR methods. Alternatively, point-based representations for radiance field excel at modeling thin geometry and leverage fast point rasterization pipelines to render images using $\\alpha$ -blending rather than ray-marching [17,", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5356", + "bbox": [ + 482, + 943, + 514, + 954 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "31]. In particular, the very recent 3D Gaussian Splitting model [15] allows for optimizing and rendering scenes with speed and quality never seen before.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 76, + 148, + 290, + 166 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the sake of completeness, we briefly describe the original 3D Gaussian Splitting method here. The scene is represented as a (large) set of Gaussians, where each Gaussian $g$ is represented by its mean $\\mu_{g}$ and its covariance $\\Sigma_{g}$ is parameterized by a scaling vector $s_g\\in \\mathbb{R}^3$ and a quaternion $q_{g}\\in \\mathbb{R}^{4}$ encoding the rotation of the Gaussian. In addition, each Gaussian is associated with its opacity $\\alpha_{g}\\in [0,1]$ and a set of spherical harmonics coordinates describing the colors emitted by the Gaussian for all directions.", + "bbox": [ + 75, + 172, + 468, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An image of a set of Gaussians can be rendered from a given viewpoint thanks to a rasterizer. This rasterizer splats the 3D Gaussians into 2D Gaussians parallel to the image plane for rendering, which results in an extremely fast rendering process. This is the key component that makes 3D Gaussian Splating much faster than NeRFs, as it is much faster than the ray-marching compositing required in the optimization of NeRFs.", + "bbox": [ + 75, + 308, + 468, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a set of images, the set of Gaussians is initialized from the point cloud produced by SfM [32]. The Gaussians' parameters (means, quaternions, scaling vectors, but also opacities and spherical harmonics parameters) are optimized to make the renderings of the Gaussians match the input images. During optimization, more Gaussians are added to better fit the scene's geometry. As a consequence, Gaussian Splatting generally produces scenes with millions of Gaussians that can be extremely small.", + "bbox": [ + 75, + 428, + 468, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 574, + 166, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present our SuGaR in this section:", + "bbox": [ + 76, + 599, + 331, + 614 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- First, we detail our loss term that enforces the alignment of the 3D Gaussians with the surface of the scene during the optimization of Gaussian Splatting.", + "- We then detail our method that exploits this alignment for extracting a highly detailed mesh from the Gaussians within minutes on a single GPU.", + "- Finally, we describe our optional refinement strategy that jointly optimizes the mesh and 3D Gaussians located on the surface of the mesh using Gaussian Splatting rendering. This strategy results in a new set of Gaussians bound to an editable mesh." + ], + "bbox": [ + 76, + 614, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Aligning the Gaussians with the Surface", + "text_level": 1, + "bbox": [ + 75, + 787, + 419, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As discussed in the introduction, to facilitate the creation of a mesh from the Gaussians, we introduce a regularization term into the Gaussian Splitting optimization that encourages the Gaussians to be aligned with the surface of the scene and well distributed over this surface. Our approach is to derive an SDF from the Gaussians under the assumption", + "bbox": [ + 75, + 811, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that the Gaussians have the desired properties. By minimizing the difference between this SDF and the actual SDF computed for the Gaussians, we encourage the Gaussians to have these properties.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a given Gaussian Splitting scene, we start by considering the corresponding density function $d:\\mathbb{R}^3\\to \\mathbb{R}_+$ computed as the sum of the Gaussian values weighted by their alpha-blending coefficients at any space location $p$ ..", + "bbox": [ + 496, + 162, + 890, + 222 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd (p) = \\sum_ {g} \\alpha_ {g} \\exp \\left(- \\frac {1}{2} \\left(p - \\mu_ {g}\\right) ^ {T} \\Sigma_ {g} ^ {- 1} \\left(p - \\mu_ {g}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 232, + 890, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $\\mu_g, \\Sigma_g$ , and $\\alpha_g$ are the centers, covariances, and alpha-blending coefficients of the Gaussians, respectively. Let us consider what this density function becomes if the Gaussians are well distributed and aligned with the surface.", + "bbox": [ + 496, + 281, + 890, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, in such scenario, the Gaussians would have limited overlap with their neighbors. As illustrated in Figure 3 (top-left), this is not the case in general. Then, for any point $p \\in \\mathbb{R}^3$ close to the surface of the scene, the Gaussian $g^{*}$ closest to the point $p$ is likely to contribute much more than others to the density value $d(p)$ . We could then approximate the Gaussian density at $p$ by:", + "bbox": [ + 496, + 351, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {g ^ {*}} \\exp \\left(- \\frac {1}{2} (p - \\mu_ {g ^ {*}}) ^ {T} \\Sigma_ {g ^ {*}} ^ {- 1} (p - \\mu_ {g ^ {*}})\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 467, + 890, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the \"closest Gaussian\" $g^{*}$ is taken as the Gaussian with the largest contribution at point $p$ :", + "bbox": [ + 496, + 510, + 890, + 541 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng ^ {*} = \\arg \\min _ {g} \\left\\{\\left(p - \\mu_ {g}\\right) ^ {T} \\Sigma_ {g} ^ {- 1} \\left(p - \\mu_ {g}\\right) \\right\\}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 550, + 890, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eq. (2) thus considers that the contribution of the closest Gaussian $g^{*}$ to the density at $p$ is much higher than the contribution of the other Gaussians. This will help us encourage the Gaussians to be well spread.", + "bbox": [ + 496, + 585, + 890, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also would like the 3D Gaussians to be flat, as they would then be aligned more closely with the surface of the mesh. Consequently, every Gaussian $g$ would have one of its three scaling factors close to 0 and:", + "bbox": [ + 496, + 656, + 890, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(p - \\mu_ {g}\\right) ^ {T} \\Sigma_ {g} ^ {- 1} \\left(p - \\mu_ {g}\\right) \\approx \\frac {1}{s _ {g} ^ {2}} \\langle p - \\mu_ {g}, n _ {g} \\rangle^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 726, + 890, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s_g$ the smallest scaling factor of the Gaussian and $n_g$ the direction of the corresponding axis. Moreover, because we want Gaussians to describe the true surface of the scene, we need to avoid semi-transparent Gaussians. Therefore, we want Gaussians to be either opaque or fully transparent, in which case we can drop them for rendering. Consequently, we want to have $\\alpha_g = 1$ for any Gaussian $g$ .", + "bbox": [ + 496, + 770, + 890, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In such scenario, the density of the Gaussians could fi-", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5357", + "bbox": [ + 482, + 943, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a56292c9d827b58edba578199500579430f2a6afc2342e78c183f2cabb75b7e6.jpg", + "image_caption": [ + "(a) Mesh & Gaussians" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 210, + 263 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0fcc44c7ccb311ac823d6d50fb0fc3578e6568e81d27799a90f6d5a308775a37.jpg", + "image_caption": [ + "(b) Mesh (No Texture)" + ], + "image_footnote": [], + "bbox": [ + 212, + 88, + 344, + 263 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d0edac57d5ee016f3498c03b58541f242fc507dbc35d2b68fac779908d8555be.jpg", + "image_caption": [ + "(c) Mesh normals" + ], + "image_footnote": [], + "bbox": [ + 346, + 88, + 478, + 263 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c917145fa695713ee2573a050eaa62a4f8dc031b2f14f7df00ab8096b2a22b0e.jpg", + "image_caption": [ + "(a) Mesh & Gaussians", + "Figure 4. Examples of (a) renderings and (b) reconstructed meshes with SuGaR. The (c) normal maps help visualize the geometry." + ], + "image_footnote": [], + "bbox": [ + 491, + 88, + 622, + 263 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3cb31b2b075cc43fa4311d547582d177406bfe27cb552cf805412325a9c0c952.jpg", + "image_caption": [ + "(b) Mesh (No Texture)" + ], + "image_footnote": [], + "bbox": [ + 625, + 88, + 756, + 263 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4a714642ab921d433fa94db5a38240ef13762090f89437b638b27ee572ab763f.jpg", + "image_caption": [ + "(c) Mesh normals" + ], + "image_footnote": [], + "bbox": [ + 759, + 88, + 890, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "nally be approximated by density $\\bar{d} (p)$ with:", + "bbox": [ + 76, + 330, + 372, + 348 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {d} (p) = \\exp \\left(- \\frac {1}{2 s _ {g ^ {*}} ^ {2}} \\langle p - \\mu_ {g ^ {*}}, n _ {g ^ {*}} \\rangle^ {2}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 357, + 468, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A first strategy to enforce our regularization is to add term $|d(p) - \\bar{d}(p)|$ to the optimization loss. While this approach works well to align Gaussians with the surface, we noticed that computing a slightly different loss relying on an SDF rather than on density further increases the alignment of Gaussians with the surface of the scene. For a given flat Gaussian, i.e., $s_g = 0$ , considering level sets is meaningless since all level sets would degenerate toward the plane passing through the center of the Gaussian $\\mu_g$ with normal $n_g$ . The distance between point $p$ and the true surface of the scene would be approximately $|\\langle p - \\mu_{g'}, n_{g'} \\rangle|$ , the distance from $p$ to this plane. Consequently, the zero-crossings of the Signed Distance Function", + "bbox": [ + 75, + 412, + 468, + 606 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {f} (p) = \\pm s _ {g *} \\sqrt {- 2 \\log (\\bar {d} (p))} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 618, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "corresponds to the surface of the scene. More generally, we define", + "bbox": [ + 75, + 652, + 468, + 680 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf (p) = \\pm s _ {g *} \\sqrt {- 2 \\log (d (p))} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 681, + 468, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as the \"ideal\" distance function associated with the density function $d$ . This distance function corresponds to the true surface of the scene in an ideal scenario where $d = \\bar{d}$ . We therefore take our regularization term $\\mathcal{R}$ as", + "bbox": [ + 75, + 705, + 468, + 765 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} = \\frac {1}{| \\mathcal {P} |} \\sum_ {p \\in \\mathcal {P}} \\left| | \\hat {f} (p) | - | f (p) | \\right|, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 775, + 468, + 813 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by sampling 3D points $p$ and summing the differences at these points between the ideal SDF $f(p)$ and an estimate $\\hat{f}(p)$ of the SDF of the surface created by the current Gaussians. $\\mathcal{P}$ refers to the set of sampled points. During optimization, we simply want the zero-level sets of both the", + "bbox": [ + 75, + 823, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0fda687334d3dc4b93586b3ae1e96bd344c0ae51d878b5b833c5a193d5922480.jpg", + "image_caption": [ + "Figure 5. Efficiently estimating $\\hat{f}(p)$ of the SDF of the surface generated from Gaussians. We render depth maps of the Gaussians, sample points $p$ in the viewpoint according to the distribution of the Gaussians. Value $\\hat{f}(p)$ is taken as the 3D distance between $p$ and the intersection between the line of sight for $p$ and the depth map." + ], + "image_footnote": [], + "bbox": [ + 588, + 349, + 799, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SDF estimator $\\hat{f}$ and the ideal SDF $f$ to align. Since a zero-level set is entirely determined by an unsigned distance function, we actually do not need to compute the sign of the SDFs and use absolute values.", + "bbox": [ + 496, + 603, + 890, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Computing efficiently $\\hat{f}(p)$ is a priori challenging. To do so, we propose to use the depth maps of the Gaussians from the viewpoints used for training—these depth maps can be rendered efficiently by extending the splatting rasterizer. Then, as shown in Figure 5, for a point $p$ visible from a training viewpoint, $\\hat{f}(p)$ is the difference between the depth of $p$ and the depth in the corresponding depth map at the projection of $p$ . Moreover, we sample points $p$ following the distribution of the Gaussians:", + "bbox": [ + 496, + 667, + 890, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\sim \\prod_ {g} \\mathcal {N} \\left(.; \\mu_ {g}, \\Sigma_ {g}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 811, + 890, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\mathcal{N}(:,\\mu_g,\\Sigma_g)$ the Gaussian distribution of mean $\\mu_{g}$ and covariance $\\Sigma_{g}$ as these points are likely to correspond to a high gradient for $\\mathcal{R}$ .", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5358", + "bbox": [ + 482, + 943, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7793310e8a1663a1bc425bb1121acfd80c404fd4cb0d52af2380bd5663ebf64a.jpg", + "image_caption": [ + "Figure 6. Sampling points on a level set for Poisson reconstruction. Left: We sample points on the depth maps of the Gaussians and refine the point locations to move the points on the level set. Right: Comparison between the extracted mesh without (left) and with (right) our refinement step. Since splatted depth maps are not exact, using directly the depth points for reconstruction usually results in a large amount of noise and missing details." + ], + "image_footnote": [], + "bbox": [ + 81, + 93, + 251, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d262f28448b4d3af275cf2b66155ba011da0fa7967794a9bd0b45cb786fba190.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 92, + 370, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2f706608e69c018da35d9debfa62abf0fe7d326101fcafab0e862679a53f0601.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 92, + 467, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also add a regularization term to encourage the normals of SDF $f$ and the normals of SDF $\\bar{f}$ to be similar:", + "bbox": [ + 76, + 359, + 468, + 390 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {\\text {N o r m}} = \\frac {1}{| \\mathcal {P} |} \\sum_ {p \\in \\mathcal {P}} \\left\\| \\frac {\\nabla f (p)}{\\| \\nabla f (p) \\| _ {2}} - n _ {g ^ {*}} \\right\\| _ {2} ^ {2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 400, + 468, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Efficient Mesh Extraction", + "text_level": 1, + "bbox": [ + 76, + 454, + 310, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To create a mesh from the Gaussians obtained after optimization using our regularization terms in Eq. (8) and Eq. (10), we sample 3D points on a level set of the density computed from the Gaussians. The level set depends on a level parameter $\\lambda$ . Then, we obtain a mesh by simply running a Poisson reconstruction [14] on these points. Note that we can also easily assign the points with the normals of the SDF, which improves the mesh quality.", + "bbox": [ + 75, + 479, + 468, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The challenge is in efficiently identifying points lying on the level set. For this, as shown in Figure 6, we again rely on the depth maps of the Gaussians as seen from the training viewpoints. We first randomly sample pixels from each depth map. For each pixel $m$ , we sample its line of sight to find a 3D point on the level set. Formally, we sample $n$ points $p + t_i v$ , where $p$ is the 3D point in the depth map that reprojects on pixel $m$ , $v$ is the direction of the line of sight, and $t_i \\in [-3\\sigma_g(v), 3\\sigma_g(v)]$ where $\\sigma_g(v)$ is the standard deviation of the 3D Gaussian $g$ in the direction of the camera. The interval $[-3\\sigma_g(v), 3\\sigma_g(v)]$ is the confidence interval for the 99.7 confidence level of the 1D Gaussian function of $t$ along the ray.", + "bbox": [ + 75, + 599, + 468, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, we compute the density values $d_{i} = d(p + t_{i}v)$ from Eq. (1) of these sampled points. If there exist $i,j$ such that $d_{i} < \\lambda < d_{j}$ , then there is a level set point located in this range. If so, we use linear interpolation to compute the coefficient $t^{*}$ such that $p + t^{*}v$ is the level set point closest to the camera, verifying $d(p + t^{*}v) = \\lambda$ . We also compute the normals of the surface at points $\\hat{p}$ , which we naturally", + "bbox": [ + 75, + 796, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e1246092c67462f639fac33a618dd8fa91f5355a55569554ece5a95bc25dd3d.jpg", + "image_caption": [ + "Figure 7. Joint refinement of mesh and Gaussians. Left: We bind Gaussians to the triangles of the mesh. Depending on the number of triangles in the scene, we bind a different number of Gaussians per triangle, with predefined barycentric coordinates. Right: Mesh before and after joint refinement." + ], + "image_footnote": [], + "bbox": [ + 540, + 88, + 874, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "define as the normalized analytical gradient of the density $\\frac{\\nabla d(\\hat{p})}{\\|\\nabla d(\\hat{p})\\|_2}$ .", + "bbox": [ + 500, + 361, + 888, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we apply Poisson reconstruction to reconstruct a surface mesh from the level set points and their normals.", + "bbox": [ + 500, + 401, + 890, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Binding New 3D Gaussians to the Mesh", + "text_level": 1, + "bbox": [ + 500, + 460, + 836, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Once we have extracted a first mesh, we can refine this mesh by binding new Gaussians to the mesh triangles and optimize the Gaussians and the mesh jointly using the Gaussian Splatting rasterizer. This enables the edition of the Gaussian splatting scene with popular mesh editing tools while keeping high-quality rendering thanks to the Gaussians.", + "bbox": [ + 498, + 491, + 890, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the initial mesh, we instantiate new 3D Gaussians on the mesh. More exactly, we associate a set of $n$ thin 3D Gaussians to each triangle of the mesh, sampled on the surface of the triangle, as illustrated in Figure 7. To do so, we slightly modify the structure of the original 3D Gaussian Splitting model.", + "bbox": [ + 498, + 587, + 890, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We explicitly compute the means of the Gaussians from the mesh vertices using predefined barycentric coordinates in the corresponding triangles during optimization. Also, the Gaussians have only 2 learnable scaling factors instead of 3 and only 1 learnable 2D rotation encoded with a complex number rather than a quaternion, to keep the Gaussians flat and aligned with the mesh triangles. More details about this parameterisation are given in the supplementary material. Like the original model, we also optimize an opacity value and a set of spherical harmonics for every Gaussian to encode the color emitted in all directions.", + "bbox": [ + 496, + 684, + 890, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 7 shows an example of a mesh before and after refinement. Figure 1 and the supplementary material give examples of what can be done by editing the mesh.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5359", + "bbox": [ + 482, + 943, + 514, + 954 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 90, + 209, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 114, + 290, + 130 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All our models are optimized on a single GPU Nvidia Tesla V100 SXM2 32 Go.", + "bbox": [ + 76, + 137, + 468, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regularization. For all scenes, we start by optimizing a Gaussian Splatting with no regularization for 7,000 iterations to let the 3D Gaussians position themselves without any additional constraint. Then, we perform 2,000 iterations with an additional entropy loss on the opacities $\\alpha_{g}$ of the Gaussians, as a way to enforce them to become binary.", + "bbox": [ + 76, + 186, + 467, + 276 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we remove Gaussians with opacity values under 0.5 and perform 6,000 iterations with the regularization term introduced in Subsection 4.1, which makes a total of 15,000 iterations. To compute the density values of points from a Gaussian $g$ , we sum only the Gaussian functions from the 16 nearest Gaussians of $g$ and update the list of nearest neighbors every 500 iterations. Optimization takes between 15 and 45 minutes depending on the scene.", + "bbox": [ + 75, + 277, + 467, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mesh extraction. For all experiments except the ablation presented in Table 2, we extract the $\\lambda$ -level set of the density function for $\\lambda = 0.3$ . We perform Poisson reconstruction with depth 10 and apply mesh simplification using quadric error metrics [9] to decrease the resolution of the meshes. Mesh extraction generally takes between 5 and 10 minutes depending on the scene.", + "bbox": [ + 75, + 414, + 467, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Joint refinement. We jointly refine the mesh and the bound 3D Gaussians for either 2,000, 7,000 or 15,000 iterations. Depending on the number of iterations, the duration of refinement goes from a few minutes to an hour.", + "bbox": [ + 76, + 537, + 467, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Real-Time Rendering of Real Scenes", + "text_level": 1, + "bbox": [ + 76, + 607, + 393, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For evaluating our model, we follow the approach from the original 3D Gaussian Splatting paper [15] and compare the performance of several variations of our method SuGaR after refinement on real 3D scenes from 3 different datasets: Mip-NeRF360 [1], DeepBlending [12] and Tanks&Temples [16]. We call R-SuGaR-NK a refined SuGaR model optimized for $N$ iterations during refinement.", + "bbox": [ + 76, + 631, + 467, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following [15], we select the same sets of 2 scenes from Tanks&Temples (Truck and Train) and 2 scenes from DeepBlending (Playroom and Dr. Johnson). However, due to licensing issues and the unavailability of the scenes Flowers and Treehill, we perform the evaluation of all methods only on 7 scenes from Mip-NeRF360 instead of the full set of 9 scenes.", + "bbox": [ + 76, + 736, + 467, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compute the standard metrics PSNR, SSIM and LPIPS [44] to evaluate the quality of SuGaR's rendering using our extracted meshes and their bound surface Gaussians. Note that [6, 26, 39] also do not use plain textured", + "bbox": [ + 76, + 840, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mesh rendering. We compare to several baselines, some of them focusing only on Novel View Synthesis [2, 15, 23, 41] and others relying on a reconstructed mesh [6, 26, 39], just like SuGaR. Results on the Mip-NeRF360 dataset are given in Table 1. Results on Tanks&Temple and DeepBlending are similar and can be found in the supplementary material.", + "bbox": [ + 496, + 90, + 890, + 180 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Even though SuGaR focuses on aligning 3D Gaussians for reconstructing a high quality mesh during the first stage of its optimization, it significantly outperforms the state of the art methods for Novel View Synthesis using a mesh and reaches better performance than several famous models that focus only on rendering, such as I-NGP [23] and Plenoxels [41]. This performance is remarkable as SuGaR is able to extract a mesh significantly faster than other methods.", + "bbox": [ + 496, + 181, + 890, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, SuGaR even reaches performance similar to state-of-the-art models for rendering quality [2, 15] on some of the scenes used for evaluation. Two main reasons explain this performance. First, the mesh extracted after the first stage of optimization serves as an excellent initialization for positioning Gaussians when starting the refinement phase. Then, the Gaussians constrained to remain on the surface during refinement greatly increase the rendering quality as they play the role of an efficient texturing tool and help reconstructing very fine details missing in the extracted mesh. Additional qualitative results are available in Figure 4.", + "bbox": [ + 496, + 301, + 890, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Mesh Extraction", + "text_level": 1, + "bbox": [ + 500, + 479, + 663, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the ability of our mesh extraction method for reconstructing high-quality meshes that are well-suited for view synthesis, we compare different mesh extraction algorithms. In particular, we optimize several variations of SuGaR by following the exact same pipeline as our standard model, except for the mesh extraction process: We either extract the mesh using a very fine marching cubes algorithm [21], by applying Poisson reconstruction [14] using the centers of the 3D Gaussians as the surface point cloud, or by applying our mesh extraction method on different level sets. Quantitative results are available in Table 2 and show the clear superiority of our approach for meshing 3D Gaussians. Figure 3 also illustrates how the marching cubes algorithm fails in this context.", + "bbox": [ + 496, + 503, + 890, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Mesh Rendering Ablation", + "text_level": 1, + "bbox": [ + 500, + 726, + 735, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3 provides additional results to quantify how various parameters impact rendering performance. In particular, we evaluate how the resolution of the mesh extraction, i.e., the number of triangles, modifies the rendering quality. For fair comparison, we increase the number of surface-aligned Gaussians per triangle when we decrease the number of triangles. Results show that increasing the number of vertices increases the quality of rendering with surface Gaussians, but meshes with less triangles are already able to reach state of the art results.", + "bbox": [ + 496, + 750, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5360", + "bbox": [ + 482, + 943, + 514, + 954 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/671fe6d63757a6e2852b35c099121a5965e771f4c7633e1e4ba7462c5e273efb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Indoor scenesOutdoor scenesAverage on all scenes
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
No mesh (except SuGaR)
Plenoxels [42]24.830.7660.42622.020.5420.46523.620.6700.443
INGP-Base [23]28.650.8400.28123.470.5710.41626.430.7250.339
INGP-Big [23]29.140.8630.24223.570.6020.37526.750.7510.299
Mip-NeRF360 [2]31.580.9140.18225.790.7460.24729.090.8420.210
3DGS [15]30.410.9200.18926.400.8050.17328.690.8700.182
R-SuGaR-15K (Ours)29.430.9100.21624.400.6990.30127.270.8200.253
With mesh
Mobile-NeRF [6]---21.950.4700.470---
NeRFMeshing [26]23.83--22.23--23.15--
BakedSDF [39]27.060.8360.258------
R-SuGaR-2K (Ours)26.290.8720.26222.970.6480.36024.870.7760.304
R-SuGaR-7K (Ours)28.730.9040.22624.160.6910.31326.770.8130.263
R-SuGaR-15K (Ours)29.430.9100.21624.400.6990.30127.270.8200.253
", + "bbox": [ + 107, + 88, + 857, + 351 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/64196f0b1efc41c3b79491cb9c860df5e66b20cb9461243b7e45ba98b9b2501a.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation of rendering quality on the Mip-NeRF360 dataset [2]. SuGaR is best among the methods that recover a mesh, and still performs well compared to NeRF methods and vanilla 3D Gaussian Splatting." + ], + "table_footnote": [], + "table_body": "
Extraction methodPSNR ↑SSIM ↑LPIPS ↓
Marching Cubes [21]23.910.7030.392
Poisson (centers) [14]23.760.7560.340
Ours (Surface level 0.1)24.620.7650.313
Ours (Surface level 0.3)24.870.7760.304
Ours (Surface level 0.5)24.910.7770.304
", + "bbox": [ + 91, + 414, + 450, + 513 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1d39f4556690dd0181417cf11f5b127f9c425cf3c7ba4eda4b8cdf0933ee8866.jpg", + "table_caption": [ + "Table 2. Ablation for different mesh extraction methods on the Mip-NeRF360 dataset [2] after applying our regularization term. For 'Poisson (centers)', we apply Poisson reconstruction [14] using as surface points the centers of the 3D Gaussians. For fair comparison, we calibrate the methods to enforce all extracted meshes to have approximately 1,000,000 vertices." + ], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPS ↓
1M vertices (3DGS)24.510.7680.295
1M vertices (UV)21.240.6090.478
200K vertices (3DGS)24.240.7570.300
200K vertices (UV)21.440.6560.419
", + "bbox": [ + 96, + 625, + 444, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Comparison between surface-aligned 3D Gaussians and an optimized traditional UV texture on the Mip-NeRF360 dataset [2]. For fair comparison, we only use the diffuse spherical harmonics component when rendering images with SuGaR. Using 3D Gaussians bound to the mesh greatly improves rendering quality, even though it contains less parameters than the UV texture.", + "bbox": [ + 75, + 727, + 468, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Then, we illustrate the benefits of using Gaussians aligned on the surface as a texturing tool for rendering meshes. To this end, we also optimize traditional UV textures on our meshes using differentiable mesh rendering", + "bbox": [ + 75, + 840, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with traditional triangle rasterization. Even though rendering with surface-aligned Gaussians provides better performance, rendering our meshes with traditional UV textures still produces satisfying results, which further illustrates the quality of our extracted meshes. Qualitative comparisons are provided in the supplementary material.", + "bbox": [ + 498, + 416, + 890, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 531, + 619, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed a very fast algorithm to obtain an accurate 3D triangle mesh for a scene via Gaussian Splatting. Moreover, by combining meshing and Gaussian Splatting, we make possible intuitive manipulation of the captured scenes and realistic rendering, offering new possibilities for creators. SuGaR does not come without limitations: Gaussians do tend to \"cheat\" on the geometry and depth by creating cavities to reproduce specular effects, instead of relying on spherical harmonics. SuGaR's regularization mitigates this issue, but Gaussians can still distort the surface in regions with intense specularity. In addition, SuGaR's assumption that the scene can be represented as a surface complicates the rendering of volumetric effects or fuzzy materials.", + "bbox": [ + 496, + 559, + 890, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements.", + "text_level": 1, + "bbox": [ + 500, + 781, + 669, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This project was funded in part by the European Union (ERC Advanced Grant explorer Funding ID #101097259). This work was granted access to the HPC resources of IDRIS under the allocation 2023-AD011013387R1 made by GENCI. We thank George Drettakis and Elliot Vincent for inspiring discussions and valuable feedback.", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5361", + "bbox": [ + 482, + 943, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T. Barron. Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. In International Conference on Computer Vision, 2021. 3, 7", + "[2] Jonathan T. Barron. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In Conference on Computer Vision and Pattern Recognition, 2022. 3, 7, 8, 2, 4", + "[3] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. NeRD: Neural Reflectance Decomposition from Image Collections. In International Conference on Computer Vision, 2021. 3", + "[4] Chris Buehler, Michael Bosse, Leonard Mcmillan, Steven Gortler, and Michael Cohen. Unstructured Lumigraph Rendering. In ACM SIGGRAPH, 2001. 3", + "[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial Radiance Fields. In European Conference on Computer Vision, 2022. 3", + "[6] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. MobileNeRF: Exploiting the Polygon Rasterization Pipeline for Efficient Neural Field Rendering on Mobile Architectures. In Conference on Computer Vision and Pattern Recognition, 2023. 2, 3, 7, 8", + "[7] Chong Bao and Bangbang Yang, Zeng Junyi, Bao Hujun, Zhang Yinda, Cui Zhaopeng, and Zhang Guofeng. NeuMesh: Learning Disentangled Neural Mesh-Based Implicit Field for Geometry and Texture Editing. In European Conference on Computer Vision, 2022. 3", + "[8] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving Neural Implicit Surfaces Geometry with Patch Warping. In Conference on Computer Vision and Pattern Recognition, 2022. 3", + "[9] Michael Garland and Paul S. Heckbert. Surface Simplification Using Quadric Error Metrics. In ACM SIGGRAPH, 1997. 7", + "[10] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven Seitz. Multi-View Stereo for Community Photo Collections. In International Conference on Computer Vision, 2007. 3", + "[11] Peter Hedman and Pratul P. Srinivasan. Baking Neural Radiance Fields for Real-Time View Synthesis. In International Conference on Computer Vision, 2021. 3", + "[12] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep Blending for Free-Viewpoint Image-Based Rendering. In ACM SIGGRAPH, 2018. 3, 7, 2, 4", + "[13] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Mitra. ReLU Fields: The Little Non-Linearity That Could. In ACM SIGGRAPH, 2022. 3", + "[14] Michael M. Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson Surface Reconstruction. In Eurographics, 2006. 2, 6, 7, 8", + "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. In ACM SIGGRAPH, 2023. 1, 2, 4, 7, 8" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking Large-Scale Scene Reconstruction. In ACM SIGGRAPH, 2017. 7, 2, 4", + "[17] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-Based Neural Rendering with Per-View Optimization. In Computer Graphics Forum, 2021. 3", + "[18] Zhengfei Kuang, Kyle Olszewski, Menglei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. NeROIC: Neural Rendering of Objects from Online Image Collections. In ACM SIGGRAPH, 2022. 3", + "[19] Marc Levoy and Pat Hanrahan. Light Field Rendering. In ACM SIGGRAPH, 1996. 3", + "[20] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-Fidelity Neural Surface Reconstruction. In Conference on Computer Vision and Pattern Recognition, 2023. 2, 3", + "[21] William E. Lorensen and Harvey E. Cline. Marching Cubes: A High Resolution 3D Surface Construction Algorithm. In ACM SIGGRAPH, 1987. 2, 3, 7, 8", + "[22] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, 2020. 1, 3", + "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. In ACM SIGGRAPH, 2022. 3, 7, 8, 2", + "[24] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying Neural Implicit Surfaces and Radiance Fields for Multi-View Reconstruction. In International Conference on Computer Vision, 2021. 3", + "[25] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary Devito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An Imperative Style, High-Performance Deep Learning Library. In Advances in Neural Information Processing Systems. Curran Associates Inc., 2019. 2", + "[26] Marie-Julie Rakotosaona, Fabian Manhardt, Diego Martin Arroyo, Michael Niemeyer, Abhijit Kundu, and Federico Tombari. NeRFMeshing: Distilling Neural Radiance Fields into Geometrically-Accurate 3D Meshes. In DV, 2023. 2, 3, 7, 8", + "[27] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3D Deep Learning with PyTorch3D. In arXiv Preprint, 2020. 2", + "[28] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. KiloNeRF: Speeding Up Neural Radiance Fields with Thousands of Tiny MLPs. In International Conference on Computer Vision, 2021. 3", + "[29] Gernot Riegler and Vladlen Koltun. Free View Synthesis. In European Conference on Computer Vision, 2020. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5362", + "bbox": [ + 482, + 943, + 514, + 954 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Gernot Riegler and Vladlen Koltun. Stable View Synthesis. In Conference on Computer Vision and Pattern Recognition, 2021. 3", + "[31] Darius Rückert, Linus Franke, and Marc Stamminger. ADOP: Approximate Differentiable One-Pixel Point Rendering. In ACM SIGGRAPH, 2022. 4", + "[32] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo Tourism: Exploring Photo Collections in 3D. In ACM SIGGRAPH, 2006. 3, 4", + "[33] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural Reflectance and Visibility Fields for Relighting and View Synthesis. In Conference on Computer Vision and Pattern Recognition, 2021. 3", + "[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct Voxel Grid Optimization: Super-Fast Convergence for Radiance Fields Reconstruction. In Conference on Computer Vision and Pattern Recognition, 2022. 3", + "[35] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields. In Conference on Computer Vision and Pattern Recognition, 2022. 3", + "[36] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-View Reconstruction. In Advances in Neural Information Processing Systems, 2021. 2, 3", + "[37] Daniel N. Wood, Daniel I. Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H. Salesin, and Werner Stuetzle. Surface Light Fields for 3D Photography. In ACM SIGGRAPH, 2000. 3", + "[38] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume Rendering of Neural Implicit Surfaces. In Advances in Neural Information Processing Systems, 2021. 2, 3", + "[39] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, and Jonathan T. Barron. BakedSDF: Meshing Neural SDFs for Real-Time View Synthesis. In ACM SIGGRAPH, 2023. 2, 3, 7, 8", + "[40] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees For Real-Time Rendering of Neural Radiance Fields. In International Conference on Computer Vision, 2021. 3", + "[41] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. In Conference on Computer Vision and Pattern Recognition, 2022. 3, 7", + "[42] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. In Conference on Computer Vision and Pattern Recognition, 2022. 8, 2", + "[43] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse Rendering with Spherical Gaussians for Physics-Based Material Editing and Relighting. In Conference on Computer Vision and Pattern Recognition, 2021. 3", + "[44] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Deep Features as a Perceptual Metric. In Conference on Computer Vision and Pattern Recognition, 2018. 7, 2, 4", + "bbox": [ + 535, + 92, + 890, + 119 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5363", + "bbox": [ + 482, + 943, + 514, + 954 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/8c37e125-7381-4a71-8f2b-015c3d8a066a_model.json b/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/8c37e125-7381-4a71-8f2b-015c3d8a066a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0b1e60195e7cfb3bc51be370c8a2b1ac3a73bebf --- /dev/null +++ b/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/8c37e125-7381-4a71-8f2b-015c3d8a066a_model.json @@ -0,0 +1,2420 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.131, + 0.834, + 0.177 + ], + "angle": 0, + "content": "SuGaR: Surface-Aligned Gaussian Splitting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.204, + 0.726, + 0.257 + ], + "angle": 0, + "content": "Antoine Guédon Vincent Lepetit \nLIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, France \nhttps://anttwo.github.io/sugar/" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.268, + 0.203, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.269, + 0.505, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.269, + 0.892, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.437, + 0.894, + 0.493 + ], + "angle": 0, + "content": "Figure 1. We introduce a method that extracts accurate and editable meshes from 3D Gaussian Splitting representations within minutes on a single GPU. The meshes can be edited, animated, composited, etc. with very realistic Gaussian Splitting rendering, offering new possibilities for Computer Graphics. Note for example that we changed the posture of the robot between the captured scene on the bottom left and the composited scene on the right. The supplementary material provides more examples, including a video illustrating our results." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.511, + 0.314, + 0.526 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.532, + 0.473, + 0.884 + ], + "angle": 0, + "content": "We propose a method to allow precise and extremely fast mesh extraction from 3D Gaussian Splatting [15]. Gaussian Splatting has recently become very popular as it yields realistic rendering while being significantly faster to train than NeRFs. It is however challenging to extract a mesh from the millions of tiny 3D Gaussians as these Gaussians tend to be unorganized after optimization and no method has been proposed so far. Our first key contribution is a regularization term that encourages the Gaussians to align well with the surface of the scene. We then introduce a method that exploits this alignment to extract a mesh from the Gaussians using Poisson reconstruction, which is fast, scalable, and preserves details, in contrast to the Marching Cubes algorithm usually applied to extract meshes from Neural SDFs. Finally, we introduce an optional refinement strategy that binds Gaussians to the surface of the mesh, and jointly optimizes these Gaussians and the mesh through Gaussian splatting rendering. This enables easy editing, sculpting, animating, and relighting of the Gaussians by manipulating the mesh instead of the Gaussians themselves. Retrieving such an editable mesh for realistic rendering is done within minutes with our method, compared to hours with the state-of-the-art method on SDFs, while providing a better rendering quality." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.511, + 0.631, + 0.526 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.892, + 0.691 + ], + "angle": 0, + "content": "After NeRFs [22], 3D Gaussian Splatting [15] has recently become very popular for capturing a 3D scene and rendering it from novel points of view. 3D Gaussian Splatting optimizes the positions, orientations, appearances (represented as spherical harmonics), and alpha blending of many tiny 3D Gaussians on the basis of a set of training images of the scene to capture the scene geometry and appearance. Because rendering the Gaussians is much faster than rendering a neural field, 3D Gaussian Splatting is much faster than NeRFs and can capture a scene in a few minutes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.696, + 0.893, + 0.9 + ], + "angle": 0, + "content": "While the Gaussians allow very realistic renderings of the scene, it is still however challenging to extract the surface of the scene from them: As shown in Figure 3, after optimization by 3D Gaussian Splatting, the Gaussians do not take an ordered structure in general and do not correspond well to the actual surface of the scene. In addition to the surface itself, it is also often desirable to represent the scene as a mesh, which remains the representation of choice in many pipelines: A mesh-based representation allows for powerful tools for editing, sculpting, animating, and relighting the scene. Because the Gaussians after Gaussian Splatting are unstructured, it is very challenging to extract a mesh from them. Note that this is also challenging with NeRFs albeit for different reasons." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5354" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.273, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.09, + 0.474, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.173, + 0.274, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.173, + 0.474, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.265, + 0.47, + 0.321 + ], + "angle": 0, + "content": "Figure 2. Our algorithm can extract a highly detailed mesh from any 3D Gaussian Splatting scene [15] within minutes on a single GPU (top: Renderings of our meshes without texture, bottom: Renderings of the meshes with bound Gaussians)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.468, + 0.499 + ], + "angle": 0, + "content": "In this paper, we first propose a regularization term that encourages the Gaussians to be well distributed over the scene surface so that the Gaussians capture much better the scene geometry, as shown in Figure 3. Our approach is to derive a volume density from the Gaussians under the assumption that the Gaussians are flat and well distributed over the scene surface. By minimizing the difference between this density and the actual one computed from the Gaussians during optimization, we encourage the 3D Gaussians to represent well the surface geometry." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.468, + 0.72 + ], + "angle": 0, + "content": "Thanks to this regularization term, it becomes easier to extract a mesh from the Gaussians. In fact, since we introduce a density function to evaluate our regularization term, a natural approach would be to extract level sets of this density function. However, Gaussian Splatting performs densification in order to capture details of the scene with high fidelity, which results in a drastic increase in the number of Gaussians. Real scenes typically end up with one or several millions of 3D Gaussians with different scales and rotations, the majority of them being extremely small in order to reproduce texture and details in the scene. This results in a density function that is close to zero almost everywhere, and the Marching Cubes algorithm [21] fails to extract proper level sets of such a sparse density function even with a fine voxel grid, as also shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.468, + 0.884 + ], + "angle": 0, + "content": "Instead, we introduce a method that very efficiently samples points on the visible part of a level set of the density function, allowing us to run the Poisson reconstruction algorithm [14] on these points to obtain a triangle mesh. This approach is scalable, by contrast with the Marching Cubes algorithm for example, and reconstructs a surface mesh within minutes on a single GPU, compared to other state of the art methods relying on Neural SDFs for extracting meshes from radiance fields, that require at least 24 hours on one GPU [20, 36, 38, 39] and rely on multiple GPUs to speed up the process [26]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Figures 2 and 4, our method produces" + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.09, + 0.778, + 0.101 + ], + "angle": 0, + "content": "without our regularization term" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.102, + 0.633, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.102, + 0.757, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.102, + 0.876, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.17, + 0.77, + 0.181 + ], + "angle": 0, + "content": "with our regularization term" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.182, + 0.633, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.249, + 0.626, + 0.268 + ], + "angle": 0, + "content": "zoom on Gaussians on a planar surface" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.182, + 0.757, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.249, + 0.742, + 0.268 + ], + "angle": 0, + "content": "mesh with Marching Cubes" + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.182, + 0.875, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.249, + 0.865, + 0.268 + ], + "angle": 0, + "content": "mesh with our extraction method" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.281, + 0.892, + 0.405 + ], + "angle": 0, + "content": "Figure 3. Extracting a mesh from Gaussians. Without regularization, the Gaussians have no special arrangement after optimization, which makes extracting a mesh very difficult. Without our regularization term, Marching Cubes fail to extract an acceptable mesh. With our regularization term, Marching Cubes recover an extremely noisy mesh even with a very fine 3D grid. Our scalable extraction method obtains a mesh even without our regularization term. Still, the mesh is noisy. By contrast, our full method succeeds in reconstructing an accurate mesh very efficiently." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.891, + 0.54 + ], + "angle": 0, + "content": "high quality meshes. The challenge is in efficiently identifying points lying on the level set. To do this, we rely on the Gaussians depth maps seen from the training viewpoints. These depth maps can be obtained by extending the Gaussian Splatting rasterizer, and we show how to accurately sample points on the level set starting from these depth maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Finally, after extracting this mesh, we propose an optional refinement strategy that jointly optimizes the mesh and a set of 3D Gaussians through Gaussian splatting rendering only. This optimization enables high-quality rendering of the mesh using Gaussian splatting rendering rather than traditional textured mesh rendering. This results in higher performance in terms of rendering quality than other radiance field models relying on an underlying mesh at inference [6, 26, 39]. As shown in Figure 1, this makes possible the use of traditional mesh-editing tools for editing a Gaussian Splatting representation of a scene, offering endless possibilities for Computer Graphics." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.72, + 0.766, + 0.733 + ], + "angle": 0, + "content": "To summarize, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.737, + 0.891, + 0.765 + ], + "angle": 0, + "content": "- a regularization term that makes the Gaussians capture accurately the geometry of the scene;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.766, + 0.89, + 0.794 + ], + "angle": 0, + "content": "- an efficient algorithm that extracts an accurate mesh from the Gaussians within minutes;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.796, + 0.891, + 0.868 + ], + "angle": 0, + "content": "- a method to bind the Gaussians to the mesh, resulting in a more accurate mesh, higher rendering quality than state of the art methods using a mesh for Novel View Synthesis [6, 26, 39], and allowing editing the scene in many different ways." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.737, + 0.891, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "We call our approach SuGaR. In the remainder of the paper, we discuss related work, give a brief overview of vanilla" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.515, + 0.955 + ], + "angle": 0, + "content": "5355" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "3D Gaussian Splatting, describe SuGaR, and compare it to the state of the art." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.138, + 0.22, + 0.154 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.165, + 0.468, + 0.269 + ], + "angle": 0, + "content": "Image-based rendering (IBR) methods rely on a set of two-dimensional images of a scene to generate a representation of the scene and render novel views. The very first novel-view synthesis approaches were based on light fields [19], and developed the concept of volume rendering for novel views. Their work emphasized the importance of efficiently traversing volumetric data to produce realistic images." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.271, + 0.468, + 0.315 + ], + "angle": 0, + "content": "Various scene representations have been proposed since, such as triangle meshes, point clouds, voxel grids, multiplane images, or neural implicit functions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.468, + 0.563 + ], + "angle": 0, + "content": "Traditional mesh-based IBR methods. Structure-from-motion (SfM) [32] and subsequent multi-view stereo (MVS) [10] allow for 3D reconstruction of surfaces, leading to the development of several view synthesis algorithms relying on triangle meshes as the primary 3D representation of scenes. Such algorithms consider textured triangles or warp and blend captured images on the mesh surface to generate novel views [4, 12, 37]. [29, 30] consider deep learning-based mesh representations for better view synthesis, bridging the gap between traditional graphics and modern machine learning techniques. While these mesh-based methods take advantage of existing graphics hardware and software for efficient rendering, they struggle with the capture of accurate geometry and appearance in complex regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.586, + 0.468, + 0.734 + ], + "angle": 0, + "content": "Volumetric IBR methods. Volumetric methods use voxel grids, multiplane images, or neural networks to represent scenes as continuous volumetric functions of density and color. Recently, Neural Radiance Fields (NeRF) [22] introduced a novel scene representation based on a continuous volumetric function parameterized by a multilayer perceptron (MLP). NeRF produces photorealistic renderings with fine details and view-dependent effects, achieved through volumetric ray tracing. However, the original NeRF is computationally expensive and memory intensive." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.737, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To address these challenges, several works have improved NeRF's performance and scalability. These methods leverage discretized or sparse volumetric representations like voxel grids and hash tables as ways to store learnable features acting as positional encodings for 3D points [5, 13, 23, 34, 41], hierarchical sampling strategies [2, 11, 28, 40], or low-rank approximations [5]. However, they still rely on volumetric ray marching, which is incompatible with standard graphics hardware and software designed for rendering polygonal surfaces. Recent works have proposed modifying the NeRF's representation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "of geometry and emitted radiance to allow for better reconstruction of specular materials [35] or relighting the scene through an explicit decomposition into material and lighting properties [3, 18, 33, 43]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.173, + 0.892, + 0.382 + ], + "angle": 0, + "content": "Hybrid IBR methods. Some methods build on differentiable rendering to combine the advantages of mesh-based and volumetric methods, and allow for surface reconstruction as well as better editability. They use a hybrid volume-surface representation, which enables high-quality meshes suitable for downstream graphics applications while efficiently modeling view-dependent appearance. In particular, some works optimize neural signed distance functions (SDF) by training neural radiance fields in which the density is derived as a differentiable transformation of the SDF [7, 8, 20, 24, 36, 38]. A triangle mesh can finally be reconstructed from the SDF by applying the Marching Cubes algorithm [21]. However, most of these methods do not target real-time rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Alternatively, other approaches \"bake\" the rendering capacity of an optimized NeRF or neural SDF into a much efficient structure relying on an underlying triangle mesh [6] that could benefit from the traditional triangle rasterization pipeline. In particular, the recent BakedSDF [39] reconstructs high quality meshes by optimizing a full neural SDF model, baking it into a high-resolution triangle mesh that combines mesh rendering for interpolating features and deep learning to translate these features into images, and finally optimizes a view-dependent appearance model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.608 + ], + "angle": 0, + "content": "However, even though it achieves real-time rendering and produces impressive meshes of the surface of the scene, this model demands training a full neural SDF with an architecture identical to Mip-NeRF360 [1], which necessitates 48 hours of training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.892, + 0.743 + ], + "angle": 0, + "content": "Similarly, the recent method NeRFMeshing [26] proposes to also bake any NeRF model into a mesh structure, achieving real-time rendering. However, the meshing performed in this method lowers the quality of the rendering and results in a PSNR much lower than our method. Additionally, this method still requires training a full NeRF model beforehand, and needs approximately an hour of training on 8 V100 NVIDIA GPUs to allow for mesh training and extraction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.744, + 0.892, + 0.819 + ], + "angle": 0, + "content": "Our method is much faster at retrieving a 3D mesh from 3D Gaussian Splatting, which is itself much faster than NeRFs. As our experiments show, our rendering done by bounding Gaussians to the mesh results in higher quality than previous solutions based on meshes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Point-based IBR methods. Alternatively, point-based representations for radiance field excel at modeling thin geometry and leverage fast point rasterization pipelines to render images using \\(\\alpha\\)-blending rather than ray-marching [17," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.516, + 0.955 + ], + "angle": 0, + "content": "5356" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "31]. In particular, the very recent 3D Gaussian Splitting model [15] allows for optimizing and rendering scenes with speed and quality never seen before." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.149, + 0.291, + 0.167 + ], + "angle": 0, + "content": "3. 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.174, + 0.469, + 0.308 + ], + "angle": 0, + "content": "For the sake of completeness, we briefly describe the original 3D Gaussian Splitting method here. The scene is represented as a (large) set of Gaussians, where each Gaussian \\(g\\) is represented by its mean \\(\\mu_{g}\\) and its covariance \\(\\Sigma_{g}\\) is parameterized by a scaling vector \\(s_g\\in \\mathbb{R}^3\\) and a quaternion \\(q_{g}\\in \\mathbb{R}^{4}\\) encoding the rotation of the Gaussian. In addition, each Gaussian is associated with its opacity \\(\\alpha_{g}\\in [0,1]\\) and a set of spherical harmonics coordinates describing the colors emitted by the Gaussian for all directions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.469, + 0.427 + ], + "angle": 0, + "content": "An image of a set of Gaussians can be rendered from a given viewpoint thanks to a rasterizer. This rasterizer splats the 3D Gaussians into 2D Gaussians parallel to the image plane for rendering, which results in an extremely fast rendering process. This is the key component that makes 3D Gaussian Splating much faster than NeRFs, as it is much faster than the ray-marching compositing required in the optimization of NeRFs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.563 + ], + "angle": 0, + "content": "Given a set of images, the set of Gaussians is initialized from the point cloud produced by SfM [32]. The Gaussians' parameters (means, quaternions, scaling vectors, but also opacities and spherical harmonics parameters) are optimized to make the renderings of the Gaussians match the input images. During optimization, more Gaussians are added to better fit the scene's geometry. As a consequence, Gaussian Splatting generally produces scenes with millions of Gaussians that can be extremely small." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.575, + 0.168, + 0.59 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.601, + 0.332, + 0.615 + ], + "angle": 0, + "content": "We present our SuGaR in this section:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.616, + 0.469, + 0.66 + ], + "angle": 0, + "content": "- First, we detail our loss term that enforces the alignment of the 3D Gaussians with the surface of the scene during the optimization of Gaussian Splatting." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.661, + 0.469, + 0.704 + ], + "angle": 0, + "content": "- We then detail our method that exploits this alignment for extracting a highly detailed mesh from the Gaussians within minutes on a single GPU." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.706, + 0.469, + 0.779 + ], + "angle": 0, + "content": "- Finally, we describe our optional refinement strategy that jointly optimizes the mesh and 3D Gaussians located on the surface of the mesh using Gaussian Splatting rendering. This strategy results in a new set of Gaussians bound to an editable mesh." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.616, + 0.469, + 0.779 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.788, + 0.421, + 0.804 + ], + "angle": 0, + "content": "4.1. Aligning the Gaussians with the Surface" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.812, + 0.469, + 0.901 + ], + "angle": 0, + "content": "As discussed in the introduction, to facilitate the creation of a mesh from the Gaussians, we introduce a regularization term into the Gaussian Splitting optimization that encourages the Gaussians to be aligned with the surface of the scene and well distributed over this surface. Our approach is to derive an SDF from the Gaussians under the assumption" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "that the Gaussians have the desired properties. By minimizing the difference between this SDF and the actual SDF computed for the Gaussians, we encourage the Gaussians to have these properties." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.163, + 0.892, + 0.223 + ], + "angle": 0, + "content": "For a given Gaussian Splitting scene, we start by considering the corresponding density function \\(d:\\mathbb{R}^3\\to \\mathbb{R}_+\\) computed as the sum of the Gaussian values weighted by their alpha-blending coefficients at any space location \\(p\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.233, + 0.892, + 0.27 + ], + "angle": 0, + "content": "\\[\nd (p) = \\sum_ {g} \\alpha_ {g} \\exp \\left(- \\frac {1}{2} \\left(p - \\mu_ {g}\\right) ^ {T} \\Sigma_ {g} ^ {- 1} \\left(p - \\mu_ {g}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.892, + 0.341 + ], + "angle": 0, + "content": "where the \\(\\mu_g, \\Sigma_g\\), and \\(\\alpha_g\\) are the centers, covariances, and alpha-blending coefficients of the Gaussians, respectively. Let us consider what this density function becomes if the Gaussians are well distributed and aligned with the surface." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.892, + 0.457 + ], + "angle": 0, + "content": "First, in such scenario, the Gaussians would have limited overlap with their neighbors. As illustrated in Figure 3 (top-left), this is not the case in general. Then, for any point \\( p \\in \\mathbb{R}^3 \\) close to the surface of the scene, the Gaussian \\( g^{*} \\) closest to the point \\( p \\) is likely to contribute much more than others to the density value \\( d(p) \\). We could then approximate the Gaussian density at \\( p \\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.468, + 0.892, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {g ^ {*}} \\exp \\left(- \\frac {1}{2} (p - \\mu_ {g ^ {*}}) ^ {T} \\Sigma_ {g ^ {*}} ^ {- 1} (p - \\mu_ {g ^ {*}})\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.892, + 0.542 + ], + "angle": 0, + "content": "where the \"closest Gaussian\" \\( g^{*} \\) is taken as the Gaussian with the largest contribution at point \\( p \\):" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.551, + 0.892, + 0.576 + ], + "angle": 0, + "content": "\\[\ng ^ {*} = \\arg \\min _ {g} \\left\\{\\left(p - \\mu_ {g}\\right) ^ {T} \\Sigma_ {g} ^ {- 1} \\left(p - \\mu_ {g}\\right) \\right\\}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.892, + 0.646 + ], + "angle": 0, + "content": "Eq. (2) thus considers that the contribution of the closest Gaussian \\( g^{*} \\) to the density at \\( p \\) is much higher than the contribution of the other Gaussians. This will help us encourage the Gaussians to be well spread." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.717 + ], + "angle": 0, + "content": "We also would like the 3D Gaussians to be flat, as they would then be aligned more closely with the surface of the mesh. Consequently, every Gaussian \\( g \\) would have one of its three scaling factors close to 0 and:" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.727, + 0.892, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\left(p - \\mu_ {g}\\right) ^ {T} \\Sigma_ {g} ^ {- 1} \\left(p - \\mu_ {g}\\right) \\approx \\frac {1}{s _ {g} ^ {2}} \\langle p - \\mu_ {g}, n _ {g} \\rangle^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.771, + 0.892, + 0.876 + ], + "angle": 0, + "content": "where \\( s_g \\) the smallest scaling factor of the Gaussian and \\( n_g \\) the direction of the corresponding axis. Moreover, because we want Gaussians to describe the true surface of the scene, we need to avoid semi-transparent Gaussians. Therefore, we want Gaussians to be either opaque or fully transparent, in which case we can drop them for rendering. Consequently, we want to have \\( \\alpha_g = 1 \\) for any Gaussian \\( g \\)." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In such scenario, the density of the Gaussians could fi-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5357" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.212, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.267, + 0.205, + 0.279 + ], + "angle": 0, + "content": "(a) Mesh & Gaussians" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.089, + 0.345, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.268, + 0.34, + 0.279 + ], + "angle": 0, + "content": "(b) Mesh (No Texture)" + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.089, + 0.479, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.268, + 0.46, + 0.279 + ], + "angle": 0, + "content": "(c) Mesh normals" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.089, + 0.624, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.268, + 0.617, + 0.279 + ], + "angle": 0, + "content": "(a) Mesh & Gaussians" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.089, + 0.757, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.267, + 0.752, + 0.279 + ], + "angle": 0, + "content": "(b) Mesh (No Texture)" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.089, + 0.892, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.268, + 0.872, + 0.279 + ], + "angle": 0, + "content": "(c) Mesh normals" + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.292, + 0.882, + 0.307 + ], + "angle": 0, + "content": "Figure 4. Examples of (a) renderings and (b) reconstructed meshes with SuGaR. The (c) normal maps help visualize the geometry." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.332, + 0.373, + 0.349 + ], + "angle": 0, + "content": "nally be approximated by density \\(\\bar{d} (p)\\) with:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.358, + 0.469, + 0.394 + ], + "angle": 0, + "content": "\\[\n\\bar {d} (p) = \\exp \\left(- \\frac {1}{2 s _ {g ^ {*}} ^ {2}} \\langle p - \\mu_ {g ^ {*}}, n _ {g ^ {*}} \\rangle^ {2}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.47, + 0.607 + ], + "angle": 0, + "content": "A first strategy to enforce our regularization is to add term \\( |d(p) - \\bar{d}(p)| \\) to the optimization loss. While this approach works well to align Gaussians with the surface, we noticed that computing a slightly different loss relying on an SDF rather than on density further increases the alignment of Gaussians with the surface of the scene. For a given flat Gaussian, i.e., \\( s_g = 0 \\), considering level sets is meaningless since all level sets would degenerate toward the plane passing through the center of the Gaussian \\( \\mu_g \\) with normal \\( n_g \\). The distance between point \\( p \\) and the true surface of the scene would be approximately \\( |\\langle p - \\mu_{g'}, n_{g'} \\rangle| \\), the distance from \\( p \\) to this plane. Consequently, the zero-crossings of the Signed Distance Function" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.619, + 0.469, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\bar {f} (p) = \\pm s _ {g *} \\sqrt {- 2 \\log (\\bar {d} (p))} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.469, + 0.681 + ], + "angle": 0, + "content": "corresponds to the surface of the scene. More generally, we define" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.682, + 0.469, + 0.701 + ], + "angle": 0, + "content": "\\[\nf (p) = \\pm s _ {g *} \\sqrt {- 2 \\log (d (p))} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.706, + 0.469, + 0.766 + ], + "angle": 0, + "content": "as the \"ideal\" distance function associated with the density function \\( d \\). This distance function corresponds to the true surface of the scene in an ideal scenario where \\( d = \\bar{d} \\). We therefore take our regularization term \\( \\mathcal{R} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.776, + 0.469, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} = \\frac {1}{| \\mathcal {P} |} \\sum_ {p \\in \\mathcal {P}} \\left| | \\hat {f} (p) | - | f (p) | \\right|, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.824, + 0.47, + 0.901 + ], + "angle": 0, + "content": "by sampling 3D points \\( p \\) and summing the differences at these points between the ideal SDF \\( f(p) \\) and an estimate \\( \\hat{f}(p) \\) of the SDF of the surface created by the current Gaussians. \\( \\mathcal{P} \\) refers to the set of sampled points. During optimization, we simply want the zero-level sets of both the" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.35, + 0.8, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.573 + ], + "angle": 0, + "content": "Figure 5. Efficiently estimating \\(\\hat{f}(p)\\) of the SDF of the surface generated from Gaussians. We render depth maps of the Gaussians, sample points \\(p\\) in the viewpoint according to the distribution of the Gaussians. Value \\(\\hat{f}(p)\\) is taken as the 3D distance between \\(p\\) and the intersection between the line of sight for \\(p\\) and the depth map." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.892, + 0.664 + ], + "angle": 0, + "content": "SDF estimator \\(\\hat{f}\\) and the ideal SDF \\(f\\) to align. Since a zero-level set is entirely determined by an unsigned distance function, we actually do not need to compute the sign of the SDFs and use absolute values." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.668, + 0.892, + 0.803 + ], + "angle": 0, + "content": "Computing efficiently \\(\\hat{f}(p)\\) is a priori challenging. To do so, we propose to use the depth maps of the Gaussians from the viewpoints used for training—these depth maps can be rendered efficiently by extending the splatting rasterizer. Then, as shown in Figure 5, for a point \\(p\\) visible from a training viewpoint, \\(\\hat{f}(p)\\) is the difference between the depth of \\(p\\) and the depth in the corresponding depth map at the projection of \\(p\\). Moreover, we sample points \\(p\\) following the distribution of the Gaussians:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.812, + 0.891, + 0.845 + ], + "angle": 0, + "content": "\\[\np \\sim \\prod_ {g} \\mathcal {N} \\left(.; \\mu_ {g}, \\Sigma_ {g}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.855, + 0.892, + 0.901 + ], + "angle": 0, + "content": "with \\(\\mathcal{N}(:,\\mu_g,\\Sigma_g)\\) the Gaussian distribution of mean \\(\\mu_{g}\\) and covariance \\(\\Sigma_{g}\\) as these points are likely to correspond to a high gradient for \\(\\mathcal{R}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5358" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.094, + 0.252, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.093, + 0.371, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.093, + 0.468, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.234, + 0.47, + 0.331 + ], + "angle": 0, + "content": "Figure 6. Sampling points on a level set for Poisson reconstruction. Left: We sample points on the depth maps of the Gaussians and refine the point locations to move the points on the level set. Right: Comparison between the extracted mesh without (left) and with (right) our refinement step. Since splatted depth maps are not exact, using directly the depth points for reconstruction usually results in a large amount of noise and missing details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.361, + 0.469, + 0.391 + ], + "angle": 0, + "content": "We also add a regularization term to encourage the normals of SDF \\( f \\) and the normals of SDF \\( \\bar{f} \\) to be similar:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.401, + 0.469, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {\\text {N o r m}} = \\frac {1}{| \\mathcal {P} |} \\sum_ {p \\in \\mathcal {P}} \\left\\| \\frac {\\nabla f (p)}{\\| \\nabla f (p) \\| _ {2}} - n _ {g ^ {*}} \\right\\| _ {2} ^ {2}. \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.455, + 0.312, + 0.469 + ], + "angle": 0, + "content": "4.2. Efficient Mesh Extraction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.48, + 0.47, + 0.599 + ], + "angle": 0, + "content": "To create a mesh from the Gaussians obtained after optimization using our regularization terms in Eq. (8) and Eq. (10), we sample 3D points on a level set of the density computed from the Gaussians. The level set depends on a level parameter \\(\\lambda\\). Then, we obtain a mesh by simply running a Poisson reconstruction [14] on these points. Note that we can also easily assign the points with the normals of the SDF, which improves the mesh quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.601, + 0.469, + 0.795 + ], + "angle": 0, + "content": "The challenge is in efficiently identifying points lying on the level set. For this, as shown in Figure 6, we again rely on the depth maps of the Gaussians as seen from the training viewpoints. We first randomly sample pixels from each depth map. For each pixel \\( m \\), we sample its line of sight to find a 3D point on the level set. Formally, we sample \\( n \\) points \\( p + t_i v \\), where \\( p \\) is the 3D point in the depth map that reprojects on pixel \\( m \\), \\( v \\) is the direction of the line of sight, and \\( t_i \\in [-3\\sigma_g(v), 3\\sigma_g(v)] \\) where \\( \\sigma_g(v) \\) is the standard deviation of the 3D Gaussian \\( g \\) in the direction of the camera. The interval \\( [-3\\sigma_g(v), 3\\sigma_g(v)] \\) is the confidence interval for the 99.7 confidence level of the 1D Gaussian function of \\( t \\) along the ray." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.797, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Then, we compute the density values \\( d_{i} = d(p + t_{i}v) \\) from Eq. (1) of these sampled points. If there exist \\( i,j \\) such that \\( d_{i} < \\lambda < d_{j} \\), then there is a level set point located in this range. If so, we use linear interpolation to compute the coefficient \\( t^{*} \\) such that \\( p + t^{*}v \\) is the level set point closest to the camera, verifying \\( d(p + t^{*}v) = \\lambda \\). We also compute the normals of the surface at points \\( \\hat{p} \\), which we naturally" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.089, + 0.875, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.251, + 0.892, + 0.321 + ], + "angle": 0, + "content": "Figure 7. Joint refinement of mesh and Gaussians. Left: We bind Gaussians to the triangles of the mesh. Depending on the number of triangles in the scene, we bind a different number of Gaussians per triangle, with predefined barycentric coordinates. Right: Mesh before and after joint refinement." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.362, + 0.89, + 0.398 + ], + "angle": 0, + "content": "define as the normalized analytical gradient of the density \\(\\frac{\\nabla d(\\hat{p})}{\\|\\nabla d(\\hat{p})\\|_2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.402, + 0.891, + 0.431 + ], + "angle": 0, + "content": "Finally, we apply Poisson reconstruction to reconstruct a surface mesh from the level set points and their normals." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.462, + 0.838, + 0.478 + ], + "angle": 0, + "content": "4.3. Binding New 3D Gaussians to the Mesh" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.492, + 0.891, + 0.581 + ], + "angle": 0, + "content": "Once we have extracted a first mesh, we can refine this mesh by binding new Gaussians to the mesh triangles and optimize the Gaussians and the mesh jointly using the Gaussian Splatting rasterizer. This enables the edition of the Gaussian splatting scene with popular mesh editing tools while keeping high-quality rendering thanks to the Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.588, + 0.892, + 0.678 + ], + "angle": 0, + "content": "Given the initial mesh, we instantiate new 3D Gaussians on the mesh. More exactly, we associate a set of \\( n \\) thin 3D Gaussians to each triangle of the mesh, sampled on the surface of the triangle, as illustrated in Figure 7. To do so, we slightly modify the structure of the original 3D Gaussian Splitting model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.849 + ], + "angle": 0, + "content": "We explicitly compute the means of the Gaussians from the mesh vertices using predefined barycentric coordinates in the corresponding triangles during optimization. Also, the Gaussians have only 2 learnable scaling factors instead of 3 and only 1 learnable 2D rotation encoded with a complex number rather than a quaternion, to keep the Gaussians flat and aligned with the mesh triangles. More details about this parameterisation are given in the supplementary material. Like the original model, we also optimize an opacity value and a set of spherical harmonics for every Gaussian to encode the color emitted in all directions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Figure 7 shows an example of a mesh before and after refinement. Figure 1 and the supplementary material give examples of what can be done by editing the mesh." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.516, + 0.955 + ], + "angle": 0, + "content": "5359" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.21, + 0.108 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.115, + 0.292, + 0.131 + ], + "angle": 0, + "content": "5.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.138, + 0.47, + 0.168 + ], + "angle": 0, + "content": "All our models are optimized on a single GPU Nvidia Tesla V100 SXM2 32 Go." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.187, + 0.468, + 0.277 + ], + "angle": 0, + "content": "Regularization. For all scenes, we start by optimizing a Gaussian Splatting with no regularization for 7,000 iterations to let the 3D Gaussians position themselves without any additional constraint. Then, we perform 2,000 iterations with an additional entropy loss on the opacities \\(\\alpha_{g}\\) of the Gaussians, as a way to enforce them to become binary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.278, + 0.468, + 0.397 + ], + "angle": 0, + "content": "Finally, we remove Gaussians with opacity values under 0.5 and perform 6,000 iterations with the regularization term introduced in Subsection 4.1, which makes a total of 15,000 iterations. To compute the density values of points from a Gaussian \\( g \\), we sum only the Gaussian functions from the 16 nearest Gaussians of \\( g \\) and update the list of nearest neighbors every 500 iterations. Optimization takes between 15 and 45 minutes depending on the scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.468, + 0.52 + ], + "angle": 0, + "content": "Mesh extraction. For all experiments except the ablation presented in Table 2, we extract the \\(\\lambda\\)-level set of the density function for \\(\\lambda = 0.3\\). We perform Poisson reconstruction with depth 10 and apply mesh simplification using quadric error metrics [9] to decrease the resolution of the meshes. Mesh extraction generally takes between 5 and 10 minutes depending on the scene." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.539, + 0.468, + 0.599 + ], + "angle": 0, + "content": "Joint refinement. We jointly refine the mesh and the bound 3D Gaussians for either 2,000, 7,000 or 15,000 iterations. Depending on the number of iterations, the duration of refinement goes from a few minutes to an hour." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.608, + 0.394, + 0.625 + ], + "angle": 0, + "content": "5.2. Real-Time Rendering of Real Scenes" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.632, + 0.468, + 0.736 + ], + "angle": 0, + "content": "For evaluating our model, we follow the approach from the original 3D Gaussian Splatting paper [15] and compare the performance of several variations of our method SuGaR after refinement on real 3D scenes from 3 different datasets: Mip-NeRF360 [1], DeepBlending [12] and Tanks&Temples [16]. We call R-SuGaR-NK a refined SuGaR model optimized for \\(N\\) iterations during refinement." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.737, + 0.468, + 0.839 + ], + "angle": 0, + "content": "Following [15], we select the same sets of 2 scenes from Tanks&Temples (Truck and Train) and 2 scenes from DeepBlending (Playroom and Dr. Johnson). However, due to licensing issues and the unavailability of the scenes Flowers and Treehill, we perform the evaluation of all methods only on 7 scenes from Mip-NeRF360 instead of the full set of 9 scenes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.841, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We compute the standard metrics PSNR, SSIM and LPIPS [44] to evaluate the quality of SuGaR's rendering using our extracted meshes and their bound surface Gaussians. Note that [6, 26, 39] also do not use plain textured" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.181 + ], + "angle": 0, + "content": "mesh rendering. We compare to several baselines, some of them focusing only on Novel View Synthesis [2, 15, 23, 41] and others relying on a reconstructed mesh [6, 26, 39], just like SuGaR. Results on the Mip-NeRF360 dataset are given in Table 1. Results on Tanks&Temple and DeepBlending are similar and can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.892, + 0.302 + ], + "angle": 0, + "content": "Even though SuGaR focuses on aligning 3D Gaussians for reconstructing a high quality mesh during the first stage of its optimization, it significantly outperforms the state of the art methods for Novel View Synthesis using a mesh and reaches better performance than several famous models that focus only on rendering, such as I-NGP [23] and Plenoxels [41]. This performance is remarkable as SuGaR is able to extract a mesh significantly faster than other methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.892, + 0.468 + ], + "angle": 0, + "content": "Moreover, SuGaR even reaches performance similar to state-of-the-art models for rendering quality [2, 15] on some of the scenes used for evaluation. Two main reasons explain this performance. First, the mesh extracted after the first stage of optimization serves as an excellent initialization for positioning Gaussians when starting the refinement phase. Then, the Gaussians constrained to remain on the surface during refinement greatly increase the rendering quality as they play the role of an efficient texturing tool and help reconstructing very fine details missing in the extracted mesh. Additional qualitative results are available in Figure 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.481, + 0.665, + 0.495 + ], + "angle": 0, + "content": "5.3. Mesh Extraction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.714 + ], + "angle": 0, + "content": "To demonstrate the ability of our mesh extraction method for reconstructing high-quality meshes that are well-suited for view synthesis, we compare different mesh extraction algorithms. In particular, we optimize several variations of SuGaR by following the exact same pipeline as our standard model, except for the mesh extraction process: We either extract the mesh using a very fine marching cubes algorithm [21], by applying Poisson reconstruction [14] using the centers of the 3D Gaussians as the surface point cloud, or by applying our mesh extraction method on different level sets. Quantitative results are available in Table 2 and show the clear superiority of our approach for meshing 3D Gaussians. Figure 3 also illustrates how the marching cubes algorithm fails in this context." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.736, + 0.743 + ], + "angle": 0, + "content": "5.4. Mesh Rendering Ablation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.751, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Table 3 provides additional results to quantify how various parameters impact rendering performance. In particular, we evaluate how the resolution of the mesh extraction, i.e., the number of triangles, modifies the rendering quality. For fair comparison, we increase the number of surface-aligned Gaussians per triangle when we decrease the number of triangles. Results show that increasing the number of vertices increases the quality of rendering with surface Gaussians, but meshes with less triangles are already able to reach state of the art results." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.516, + 0.955 + ], + "angle": 0, + "content": "5360" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.108, + 0.089, + 0.859, + 0.352 + ], + "angle": 0, + "content": "
Indoor scenesOutdoor scenesAverage on all scenes
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
No mesh (except SuGaR)
Plenoxels [42]24.830.7660.42622.020.5420.46523.620.6700.443
INGP-Base [23]28.650.8400.28123.470.5710.41626.430.7250.339
INGP-Big [23]29.140.8630.24223.570.6020.37526.750.7510.299
Mip-NeRF360 [2]31.580.9140.18225.790.7460.24729.090.8420.210
3DGS [15]30.410.9200.18926.400.8050.17328.690.8700.182
R-SuGaR-15K (Ours)29.430.9100.21624.400.6990.30127.270.8200.253
With mesh
Mobile-NeRF [6]---21.950.4700.470---
NeRFMeshing [26]23.83--22.23--23.15--
BakedSDF [39]27.060.8360.258------
R-SuGaR-2K (Ours)26.290.8720.26222.970.6480.36024.870.7760.304
R-SuGaR-7K (Ours)28.730.9040.22624.160.6910.31326.770.8130.263
R-SuGaR-15K (Ours)29.430.9100.21624.400.6990.30127.270.8200.253
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.363, + 0.893, + 0.391 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation of rendering quality on the Mip-NeRF360 dataset [2]. SuGaR is best among the methods that recover a mesh, and still performs well compared to NeRF methods and vanilla 3D Gaussian Splatting." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.415, + 0.451, + 0.514 + ], + "angle": 0, + "content": "
Extraction methodPSNR ↑SSIM ↑LPIPS ↓
Marching Cubes [21]23.910.7030.392
Poisson (centers) [14]23.760.7560.340
Ours (Surface level 0.1)24.620.7650.313
Ours (Surface level 0.3)24.870.7760.304
Ours (Surface level 0.5)24.910.7770.304
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.525, + 0.47, + 0.609 + ], + "angle": 0, + "content": "Table 2. Ablation for different mesh extraction methods on the Mip-NeRF360 dataset [2] after applying our regularization term. For 'Poisson (centers)', we apply Poisson reconstruction [14] using as surface points the centers of the 3D Gaussians. For fair comparison, we calibrate the methods to enforce all extracted meshes to have approximately 1,000,000 vertices." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.625, + 0.446, + 0.718 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPS ↓
1M vertices (3DGS)24.510.7680.295
1M vertices (UV)21.240.6090.478
200K vertices (3DGS)24.240.7570.300
200K vertices (UV)21.440.6560.419
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.728, + 0.47, + 0.812 + ], + "angle": 0, + "content": "Table 3. Comparison between surface-aligned 3D Gaussians and an optimized traditional UV texture on the Mip-NeRF360 dataset [2]. For fair comparison, we only use the diffuse spherical harmonics component when rendering images with SuGaR. Using 3D Gaussians bound to the mesh greatly improves rendering quality, even though it contains less parameters than the UV texture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Then, we illustrate the benefits of using Gaussians aligned on the surface as a texturing tool for rendering meshes. To this end, we also optimize traditional UV textures on our meshes using differentiable mesh rendering" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.417, + 0.892, + 0.509 + ], + "angle": 0, + "content": "with traditional triangle rasterization. Even though rendering with surface-aligned Gaussians provides better performance, rendering our meshes with traditional UV textures still produces satisfying results, which further illustrates the quality of our extracted meshes. Qualitative comparisons are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.532, + 0.62, + 0.547 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.56, + 0.892, + 0.758 + ], + "angle": 0, + "content": "We proposed a very fast algorithm to obtain an accurate 3D triangle mesh for a scene via Gaussian Splatting. Moreover, by combining meshing and Gaussian Splatting, we make possible intuitive manipulation of the captured scenes and realistic rendering, offering new possibilities for creators. SuGaR does not come without limitations: Gaussians do tend to \"cheat\" on the geometry and depth by creating cavities to reproduce specular effects, instead of relying on spherical harmonics. SuGaR's regularization mitigates this issue, but Gaussians can still distort the surface in regions with intense specularity. In addition, SuGaR's assumption that the scene can be represented as a surface complicates the rendering of volumetric effects or fuzzy materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.782, + 0.671, + 0.799 + ], + "angle": 0, + "content": "Acknowledgements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.901 + ], + "angle": 0, + "content": "This project was funded in part by the European Union (ERC Advanced Grant explorer Funding ID #101097259). This work was granted access to the HPC resources of IDRIS under the allocation 2023-AD011013387R1 made by GENCI. We thank George Drettakis and Elliot Vincent for inspiring discussions and valuable feedback." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.514, + 0.956 + ], + "angle": 0, + "content": "5361" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Jonathan T. Barron. Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. In International Conference on Computer Vision, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.469, + 0.2 + ], + "angle": 0, + "content": "[2] Jonathan T. Barron. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In Conference on Computer Vision and Pattern Recognition, 2022. 3, 7, 8, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.204, + 0.469, + 0.259 + ], + "angle": 0, + "content": "[3] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. NeRD: Neural Reflectance Decomposition from Image Collections. In International Conference on Computer Vision, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.262, + 0.468, + 0.303 + ], + "angle": 0, + "content": "[4] Chris Buehler, Michael Bosse, Leonard Mcmillan, Steven Gortler, and Michael Cohen. Unstructured Lumigraph Rendering. In ACM SIGGRAPH, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.307, + 0.468, + 0.347 + ], + "angle": 0, + "content": "[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial Radiance Fields. In European Conference on Computer Vision, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.351, + 0.468, + 0.419 + ], + "angle": 0, + "content": "[6] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. MobileNeRF: Exploiting the Polygon Rasterization Pipeline for Efficient Neural Field Rendering on Mobile Architectures. In Conference on Computer Vision and Pattern Recognition, 2023. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.423, + 0.468, + 0.491 + ], + "angle": 0, + "content": "[7] Chong Bao and Bangbang Yang, Zeng Junyi, Bao Hujun, Zhang Yinda, Cui Zhaopeng, and Zhang Guofeng. NeuMesh: Learning Disentangled Neural Mesh-Based Implicit Field for Geometry and Texture Editing. In European Conference on Computer Vision, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.468, + 0.549 + ], + "angle": 0, + "content": "[8] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving Neural Implicit Surfaces Geometry with Patch Warping. In Conference on Computer Vision and Pattern Recognition, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.553, + 0.468, + 0.593 + ], + "angle": 0, + "content": "[9] Michael Garland and Paul S. Heckbert. Surface Simplification Using Quadric Error Metrics. In ACM SIGGRAPH, 1997. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.597, + 0.469, + 0.651 + ], + "angle": 0, + "content": "[10] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven Seitz. Multi-View Stereo for Community Photo Collections. In International Conference on Computer Vision, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.655, + 0.469, + 0.696 + ], + "angle": 0, + "content": "[11] Peter Hedman and Pratul P. Srinivasan. Baking Neural Radiance Fields for Real-Time View Synthesis. In International Conference on Computer Vision, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.699, + 0.469, + 0.753 + ], + "angle": 0, + "content": "[12] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep Blending for Free-Viewpoint Image-Based Rendering. In ACM SIGGRAPH, 2018. 3, 7, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.757, + 0.468, + 0.797 + ], + "angle": 0, + "content": "[13] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Mitra. ReLU Fields: The Little Non-Linearity That Could. In ACM SIGGRAPH, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.468, + 0.841 + ], + "angle": 0, + "content": "[14] Michael M. Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson Surface Reconstruction. In Eurographics, 2006. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. In ACM SIGGRAPH, 2023. 1, 2, 4, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[16] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking Large-Scale Scene Reconstruction. In ACM SIGGRAPH, 2017. 7, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[17] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-Based Neural Rendering with Per-View Optimization. In Computer Graphics Forum, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[18] Zhengfei Kuang, Kyle Olszewski, Menglei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. NeROIC: Neural Rendering of Objects from Online Image Collections. In ACM SIGGRAPH, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[19] Marc Levoy and Pat Hanrahan. Light Field Rendering. In ACM SIGGRAPH, 1996. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.265, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[20] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-Fidelity Neural Surface Reconstruction. In Conference on Computer Vision and Pattern Recognition, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.336, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[21] William E. Lorensen and Harvey E. Cline. Marching Cubes: A High Resolution 3D Surface Construction Algorithm. In ACM SIGGRAPH, 1987. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.379, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[22] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. In ACM SIGGRAPH, 2022. 3, 7, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.507, + 0.892, + 0.561 + ], + "angle": 0, + "content": "[24] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying Neural Implicit Surfaces and Radiance Fields for Multi-View Reconstruction. In International Conference on Computer Vision, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.564, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[25] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary Devito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An Imperative Style, High-Performance Deep Learning Library. In Advances in Neural Information Processing Systems. Curran Associates Inc., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[26] Marie-Julie Rakotosaona, Fabian Manhardt, Diego Martin Arroyo, Michael Niemeyer, Abhijit Kundu, and Federico Tombari. NeRFMeshing: Distilling Neural Radiance Fields into Geometrically-Accurate 3D Meshes. In DV, 2023. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[27] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3D Deep Learning with PyTorch3D. In arXiv Preprint, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[28] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. KiloNeRF: Speeding Up Neural Radiance Fields with Thousands of Tiny MLPs. In International Conference on Computer Vision, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Gernot Riegler and Vladlen Koltun. Free View Synthesis. In European Conference on Computer Vision, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.516, + 0.955 + ], + "angle": 0, + "content": "5362" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "[30] Gernot Riegler and Vladlen Koltun. Stable View Synthesis. In Conference on Computer Vision and Pattern Recognition, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[31] Darius Rückert, Linus Franke, and Marc Stamminger. ADOP: Approximate Differentiable One-Pixel Point Rendering. In ACM SIGGRAPH, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.217 + ], + "angle": 0, + "content": "[32] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo Tourism: Exploring Photo Collections in 3D. In ACM SIGGRAPH, 2006. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[33] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural Reflectance and Visibility Fields for Relighting and View Synthesis. In Conference on Computer Vision and Pattern Recognition, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct Voxel Grid Optimization: Super-Fast Convergence for Radiance Fields Reconstruction. In Conference on Computer Vision and Pattern Recognition, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.412 + ], + "angle": 0, + "content": "[35] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields. In Conference on Computer Vision and Pattern Recognition, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[36] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-View Reconstruction. In Advances in Neural Information Processing Systems, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[37] Daniel N. Wood, Daniel I. Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H. Salesin, and Werner Stuetzle. Surface Light Fields for 3D Photography. In ACM SIGGRAPH, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[38] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume Rendering of Neural Implicit Surfaces. In Advances in Neural Information Processing Systems, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[39] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, and Jonathan T. Barron. BakedSDF: Meshing Neural SDFs for Real-Time View Synthesis. In ACM SIGGRAPH, 2023. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[40] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees For Real-Time Rendering of Neural Radiance Fields. In International Conference on Computer Vision, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[41] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. In Conference on Computer Vision and Pattern Recognition, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[42] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. In Conference on Computer Vision and Pattern Recognition, 2022. 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[43] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse Rendering with Spherical Gaussians for Physics-Based Material Editing and Relighting. In Conference on Computer Vision and Pattern Recognition, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[44] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Deep Features as a Perceptual Metric. In Conference on Computer Vision and Pattern Recognition, 2018. 7, 2, 4" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.515, + 0.955 + ], + "angle": 0, + "content": "5363" + } + ] +] \ No newline at end of file diff --git a/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/8c37e125-7381-4a71-8f2b-015c3d8a066a_origin.pdf b/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/8c37e125-7381-4a71-8f2b-015c3d8a066a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e647437d30089154e2c43d102f35053d9d3824ad --- /dev/null +++ b/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/8c37e125-7381-4a71-8f2b-015c3d8a066a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95cb88bfe88ab982bb3b46b3764cc4aff0d828ccde2be668f05551ea56ca40f3 +size 6804928 diff --git a/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/full.md b/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..63895c49b062f118c22e3ade919351e313f413a5 --- /dev/null +++ b/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/full.md @@ -0,0 +1,360 @@ +# SuGaR: Surface-Aligned Gaussian Splitting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering + +Antoine Guédon Vincent Lepetit +LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, France +https://anttwo.github.io/sugar/ + +![](images/6673930ff670e4525dde86ea891322dc4edf5ae835eac2957ca13ae2f137f242.jpg) +Figure 1. We introduce a method that extracts accurate and editable meshes from 3D Gaussian Splitting representations within minutes on a single GPU. The meshes can be edited, animated, composited, etc. with very realistic Gaussian Splitting rendering, offering new possibilities for Computer Graphics. Note for example that we changed the posture of the robot between the captured scene on the bottom left and the composited scene on the right. The supplementary material provides more examples, including a video illustrating our results. + +![](images/1e71b34782521a295dba22014fb33764264018e29489fec3a41da095d26a740b.jpg) + +![](images/10648b5890419237a4fb6653f09ebb64111dc67816d3d828c0fbbb2c4b6de2cf.jpg) + +# Abstract + +We propose a method to allow precise and extremely fast mesh extraction from 3D Gaussian Splatting [15]. Gaussian Splatting has recently become very popular as it yields realistic rendering while being significantly faster to train than NeRFs. It is however challenging to extract a mesh from the millions of tiny 3D Gaussians as these Gaussians tend to be unorganized after optimization and no method has been proposed so far. Our first key contribution is a regularization term that encourages the Gaussians to align well with the surface of the scene. We then introduce a method that exploits this alignment to extract a mesh from the Gaussians using Poisson reconstruction, which is fast, scalable, and preserves details, in contrast to the Marching Cubes algorithm usually applied to extract meshes from Neural SDFs. Finally, we introduce an optional refinement strategy that binds Gaussians to the surface of the mesh, and jointly optimizes these Gaussians and the mesh through Gaussian splatting rendering. This enables easy editing, sculpting, animating, and relighting of the Gaussians by manipulating the mesh instead of the Gaussians themselves. Retrieving such an editable mesh for realistic rendering is done within minutes with our method, compared to hours with the state-of-the-art method on SDFs, while providing a better rendering quality. + +# 1. Introduction + +After NeRFs [22], 3D Gaussian Splatting [15] has recently become very popular for capturing a 3D scene and rendering it from novel points of view. 3D Gaussian Splatting optimizes the positions, orientations, appearances (represented as spherical harmonics), and alpha blending of many tiny 3D Gaussians on the basis of a set of training images of the scene to capture the scene geometry and appearance. Because rendering the Gaussians is much faster than rendering a neural field, 3D Gaussian Splatting is much faster than NeRFs and can capture a scene in a few minutes. + +While the Gaussians allow very realistic renderings of the scene, it is still however challenging to extract the surface of the scene from them: As shown in Figure 3, after optimization by 3D Gaussian Splatting, the Gaussians do not take an ordered structure in general and do not correspond well to the actual surface of the scene. In addition to the surface itself, it is also often desirable to represent the scene as a mesh, which remains the representation of choice in many pipelines: A mesh-based representation allows for powerful tools for editing, sculpting, animating, and relighting the scene. Because the Gaussians after Gaussian Splatting are unstructured, it is very challenging to extract a mesh from them. Note that this is also challenging with NeRFs albeit for different reasons. + +![](images/965e4b706f98280301b6f99dcf08dc1d3cc1bda348d4aebf93eb5600c76ab7b7.jpg) + +![](images/63849b2eb6568bab393e1eb9db3e9606a0f34ca11c35142170ea5710115c3897.jpg) + +![](images/7f0927d0b0e6405cd930616ebc9dd362849d6074393daefa14f02ea6d4638a45.jpg) +Figure 2. Our algorithm can extract a highly detailed mesh from any 3D Gaussian Splatting scene [15] within minutes on a single GPU (top: Renderings of our meshes without texture, bottom: Renderings of the meshes with bound Gaussians). + +![](images/414eed21c26dc56e63563eb1defede617594decabb74654cbaab4632ce6bb7ca.jpg) +Figure 3. Extracting a mesh from Gaussians. Without regularization, the Gaussians have no special arrangement after optimization, which makes extracting a mesh very difficult. Without our regularization term, Marching Cubes fail to extract an acceptable mesh. With our regularization term, Marching Cubes recover an extremely noisy mesh even with a very fine 3D grid. Our scalable extraction method obtains a mesh even without our regularization term. Still, the mesh is noisy. By contrast, our full method succeeds in reconstructing an accurate mesh very efficiently. + +In this paper, we first propose a regularization term that encourages the Gaussians to be well distributed over the scene surface so that the Gaussians capture much better the scene geometry, as shown in Figure 3. Our approach is to derive a volume density from the Gaussians under the assumption that the Gaussians are flat and well distributed over the scene surface. By minimizing the difference between this density and the actual one computed from the Gaussians during optimization, we encourage the 3D Gaussians to represent well the surface geometry. + +Thanks to this regularization term, it becomes easier to extract a mesh from the Gaussians. In fact, since we introduce a density function to evaluate our regularization term, a natural approach would be to extract level sets of this density function. However, Gaussian Splatting performs densification in order to capture details of the scene with high fidelity, which results in a drastic increase in the number of Gaussians. Real scenes typically end up with one or several millions of 3D Gaussians with different scales and rotations, the majority of them being extremely small in order to reproduce texture and details in the scene. This results in a density function that is close to zero almost everywhere, and the Marching Cubes algorithm [21] fails to extract proper level sets of such a sparse density function even with a fine voxel grid, as also shown in Figure 3. + +Instead, we introduce a method that very efficiently samples points on the visible part of a level set of the density function, allowing us to run the Poisson reconstruction algorithm [14] on these points to obtain a triangle mesh. This approach is scalable, by contrast with the Marching Cubes algorithm for example, and reconstructs a surface mesh within minutes on a single GPU, compared to other state of the art methods relying on Neural SDFs for extracting meshes from radiance fields, that require at least 24 hours on one GPU [20, 36, 38, 39] and rely on multiple GPUs to speed up the process [26]. + +As illustrated in Figures 2 and 4, our method produces + +![](images/fab143a17035a4621c16f71a04dc960390996ce00a5ee09eef27a7f69788960d.jpg) +without our regularization term + +![](images/62c6b2f08a971899191a5b562162f4a4a0d84c2fac41173d6dd4fdd4d21db07f.jpg) + +![](images/ad83aca945c2ef13508a04bdafb0e6aa0808eddaf7faabbba997f79174034c89.jpg) + +![](images/4015513f1d9b4cc0061301924259fc7103dea1c2a531de2b078dd3f682cdbdbe.jpg) +with our regularization term +zoom on Gaussians on a planar surface + +![](images/f9718805da1261a9b074b2a8f184814cf14f0acfe7762c7ef3b4057ab334615d.jpg) +mesh with Marching Cubes + +![](images/38688d6186e8082e57213c850b4fedc5c366bbb99ad6bf40978505ff29e8c3eb.jpg) +mesh with our extraction method + +high quality meshes. The challenge is in efficiently identifying points lying on the level set. To do this, we rely on the Gaussians depth maps seen from the training viewpoints. These depth maps can be obtained by extending the Gaussian Splatting rasterizer, and we show how to accurately sample points on the level set starting from these depth maps. + +Finally, after extracting this mesh, we propose an optional refinement strategy that jointly optimizes the mesh and a set of 3D Gaussians through Gaussian splatting rendering only. This optimization enables high-quality rendering of the mesh using Gaussian splatting rendering rather than traditional textured mesh rendering. This results in higher performance in terms of rendering quality than other radiance field models relying on an underlying mesh at inference [6, 26, 39]. As shown in Figure 1, this makes possible the use of traditional mesh-editing tools for editing a Gaussian Splatting representation of a scene, offering endless possibilities for Computer Graphics. + +To summarize, our contributions are: + +- a regularization term that makes the Gaussians capture accurately the geometry of the scene; +- an efficient algorithm that extracts an accurate mesh from the Gaussians within minutes; +- a method to bind the Gaussians to the mesh, resulting in a more accurate mesh, higher rendering quality than state of the art methods using a mesh for Novel View Synthesis [6, 26, 39], and allowing editing the scene in many different ways. + +We call our approach SuGaR. In the remainder of the paper, we discuss related work, give a brief overview of vanilla + +3D Gaussian Splatting, describe SuGaR, and compare it to the state of the art. + +# 2. Related Work + +Image-based rendering (IBR) methods rely on a set of two-dimensional images of a scene to generate a representation of the scene and render novel views. The very first novel-view synthesis approaches were based on light fields [19], and developed the concept of volume rendering for novel views. Their work emphasized the importance of efficiently traversing volumetric data to produce realistic images. + +Various scene representations have been proposed since, such as triangle meshes, point clouds, voxel grids, multiplane images, or neural implicit functions. + +Traditional mesh-based IBR methods. Structure-from-motion (SfM) [32] and subsequent multi-view stereo (MVS) [10] allow for 3D reconstruction of surfaces, leading to the development of several view synthesis algorithms relying on triangle meshes as the primary 3D representation of scenes. Such algorithms consider textured triangles or warp and blend captured images on the mesh surface to generate novel views [4, 12, 37]. [29, 30] consider deep learning-based mesh representations for better view synthesis, bridging the gap between traditional graphics and modern machine learning techniques. While these mesh-based methods take advantage of existing graphics hardware and software for efficient rendering, they struggle with the capture of accurate geometry and appearance in complex regions. + +Volumetric IBR methods. Volumetric methods use voxel grids, multiplane images, or neural networks to represent scenes as continuous volumetric functions of density and color. Recently, Neural Radiance Fields (NeRF) [22] introduced a novel scene representation based on a continuous volumetric function parameterized by a multilayer perceptron (MLP). NeRF produces photorealistic renderings with fine details and view-dependent effects, achieved through volumetric ray tracing. However, the original NeRF is computationally expensive and memory intensive. + +To address these challenges, several works have improved NeRF's performance and scalability. These methods leverage discretized or sparse volumetric representations like voxel grids and hash tables as ways to store learnable features acting as positional encodings for 3D points [5, 13, 23, 34, 41], hierarchical sampling strategies [2, 11, 28, 40], or low-rank approximations [5]. However, they still rely on volumetric ray marching, which is incompatible with standard graphics hardware and software designed for rendering polygonal surfaces. Recent works have proposed modifying the NeRF's representation + +of geometry and emitted radiance to allow for better reconstruction of specular materials [35] or relighting the scene through an explicit decomposition into material and lighting properties [3, 18, 33, 43]. + +Hybrid IBR methods. Some methods build on differentiable rendering to combine the advantages of mesh-based and volumetric methods, and allow for surface reconstruction as well as better editability. They use a hybrid volume-surface representation, which enables high-quality meshes suitable for downstream graphics applications while efficiently modeling view-dependent appearance. In particular, some works optimize neural signed distance functions (SDF) by training neural radiance fields in which the density is derived as a differentiable transformation of the SDF [7, 8, 20, 24, 36, 38]. A triangle mesh can finally be reconstructed from the SDF by applying the Marching Cubes algorithm [21]. However, most of these methods do not target real-time rendering. + +Alternatively, other approaches "bake" the rendering capacity of an optimized NeRF or neural SDF into a much efficient structure relying on an underlying triangle mesh [6] that could benefit from the traditional triangle rasterization pipeline. In particular, the recent BakedSDF [39] reconstructs high quality meshes by optimizing a full neural SDF model, baking it into a high-resolution triangle mesh that combines mesh rendering for interpolating features and deep learning to translate these features into images, and finally optimizes a view-dependent appearance model. + +However, even though it achieves real-time rendering and produces impressive meshes of the surface of the scene, this model demands training a full neural SDF with an architecture identical to Mip-NeRF360 [1], which necessitates 48 hours of training. + +Similarly, the recent method NeRFMeshing [26] proposes to also bake any NeRF model into a mesh structure, achieving real-time rendering. However, the meshing performed in this method lowers the quality of the rendering and results in a PSNR much lower than our method. Additionally, this method still requires training a full NeRF model beforehand, and needs approximately an hour of training on 8 V100 NVIDIA GPUs to allow for mesh training and extraction. + +Our method is much faster at retrieving a 3D mesh from 3D Gaussian Splatting, which is itself much faster than NeRFs. As our experiments show, our rendering done by bounding Gaussians to the mesh results in higher quality than previous solutions based on meshes. + +Point-based IBR methods. Alternatively, point-based representations for radiance field excel at modeling thin geometry and leverage fast point rasterization pipelines to render images using $\alpha$ -blending rather than ray-marching [17, + +31]. In particular, the very recent 3D Gaussian Splitting model [15] allows for optimizing and rendering scenes with speed and quality never seen before. + +# 3. 3D Gaussian Splitting + +For the sake of completeness, we briefly describe the original 3D Gaussian Splitting method here. The scene is represented as a (large) set of Gaussians, where each Gaussian $g$ is represented by its mean $\mu_{g}$ and its covariance $\Sigma_{g}$ is parameterized by a scaling vector $s_g\in \mathbb{R}^3$ and a quaternion $q_{g}\in \mathbb{R}^{4}$ encoding the rotation of the Gaussian. In addition, each Gaussian is associated with its opacity $\alpha_{g}\in [0,1]$ and a set of spherical harmonics coordinates describing the colors emitted by the Gaussian for all directions. + +An image of a set of Gaussians can be rendered from a given viewpoint thanks to a rasterizer. This rasterizer splats the 3D Gaussians into 2D Gaussians parallel to the image plane for rendering, which results in an extremely fast rendering process. This is the key component that makes 3D Gaussian Splating much faster than NeRFs, as it is much faster than the ray-marching compositing required in the optimization of NeRFs. + +Given a set of images, the set of Gaussians is initialized from the point cloud produced by SfM [32]. The Gaussians' parameters (means, quaternions, scaling vectors, but also opacities and spherical harmonics parameters) are optimized to make the renderings of the Gaussians match the input images. During optimization, more Gaussians are added to better fit the scene's geometry. As a consequence, Gaussian Splatting generally produces scenes with millions of Gaussians that can be extremely small. + +# 4. Method + +We present our SuGaR in this section: + +- First, we detail our loss term that enforces the alignment of the 3D Gaussians with the surface of the scene during the optimization of Gaussian Splatting. +- We then detail our method that exploits this alignment for extracting a highly detailed mesh from the Gaussians within minutes on a single GPU. +- Finally, we describe our optional refinement strategy that jointly optimizes the mesh and 3D Gaussians located on the surface of the mesh using Gaussian Splatting rendering. This strategy results in a new set of Gaussians bound to an editable mesh. + +# 4.1. Aligning the Gaussians with the Surface + +As discussed in the introduction, to facilitate the creation of a mesh from the Gaussians, we introduce a regularization term into the Gaussian Splitting optimization that encourages the Gaussians to be aligned with the surface of the scene and well distributed over this surface. Our approach is to derive an SDF from the Gaussians under the assumption + +that the Gaussians have the desired properties. By minimizing the difference between this SDF and the actual SDF computed for the Gaussians, we encourage the Gaussians to have these properties. + +For a given Gaussian Splitting scene, we start by considering the corresponding density function $d:\mathbb{R}^3\to \mathbb{R}_+$ computed as the sum of the Gaussian values weighted by their alpha-blending coefficients at any space location $p$ .. + +$$ +d (p) = \sum_ {g} \alpha_ {g} \exp \left(- \frac {1}{2} \left(p - \mu_ {g}\right) ^ {T} \Sigma_ {g} ^ {- 1} \left(p - \mu_ {g}\right)\right), \tag {1} +$$ + +where the $\mu_g, \Sigma_g$ , and $\alpha_g$ are the centers, covariances, and alpha-blending coefficients of the Gaussians, respectively. Let us consider what this density function becomes if the Gaussians are well distributed and aligned with the surface. + +First, in such scenario, the Gaussians would have limited overlap with their neighbors. As illustrated in Figure 3 (top-left), this is not the case in general. Then, for any point $p \in \mathbb{R}^3$ close to the surface of the scene, the Gaussian $g^{*}$ closest to the point $p$ is likely to contribute much more than others to the density value $d(p)$ . We could then approximate the Gaussian density at $p$ by: + +$$ +\alpha_ {g ^ {*}} \exp \left(- \frac {1}{2} (p - \mu_ {g ^ {*}}) ^ {T} \Sigma_ {g ^ {*}} ^ {- 1} (p - \mu_ {g ^ {*}})\right), \tag {2} +$$ + +where the "closest Gaussian" $g^{*}$ is taken as the Gaussian with the largest contribution at point $p$ : + +$$ +g ^ {*} = \arg \min _ {g} \left\{\left(p - \mu_ {g}\right) ^ {T} \Sigma_ {g} ^ {- 1} \left(p - \mu_ {g}\right) \right\}. \tag {3} +$$ + +Eq. (2) thus considers that the contribution of the closest Gaussian $g^{*}$ to the density at $p$ is much higher than the contribution of the other Gaussians. This will help us encourage the Gaussians to be well spread. + +We also would like the 3D Gaussians to be flat, as they would then be aligned more closely with the surface of the mesh. Consequently, every Gaussian $g$ would have one of its three scaling factors close to 0 and: + +$$ +\left(p - \mu_ {g}\right) ^ {T} \Sigma_ {g} ^ {- 1} \left(p - \mu_ {g}\right) \approx \frac {1}{s _ {g} ^ {2}} \langle p - \mu_ {g}, n _ {g} \rangle^ {2}, \tag {4} +$$ + +where $s_g$ the smallest scaling factor of the Gaussian and $n_g$ the direction of the corresponding axis. Moreover, because we want Gaussians to describe the true surface of the scene, we need to avoid semi-transparent Gaussians. Therefore, we want Gaussians to be either opaque or fully transparent, in which case we can drop them for rendering. Consequently, we want to have $\alpha_g = 1$ for any Gaussian $g$ . + +In such scenario, the density of the Gaussians could fi- + +![](images/a56292c9d827b58edba578199500579430f2a6afc2342e78c183f2cabb75b7e6.jpg) +(a) Mesh & Gaussians + +![](images/0fcc44c7ccb311ac823d6d50fb0fc3578e6568e81d27799a90f6d5a308775a37.jpg) +(b) Mesh (No Texture) + +![](images/d0edac57d5ee016f3498c03b58541f242fc507dbc35d2b68fac779908d8555be.jpg) +(c) Mesh normals + +![](images/c917145fa695713ee2573a050eaa62a4f8dc031b2f14f7df00ab8096b2a22b0e.jpg) +(a) Mesh & Gaussians +Figure 4. Examples of (a) renderings and (b) reconstructed meshes with SuGaR. The (c) normal maps help visualize the geometry. + +![](images/3cb31b2b075cc43fa4311d547582d177406bfe27cb552cf805412325a9c0c952.jpg) +(b) Mesh (No Texture) + +![](images/4a714642ab921d433fa94db5a38240ef13762090f89437b638b27ee572ab763f.jpg) +(c) Mesh normals + +nally be approximated by density $\bar{d} (p)$ with: + +$$ +\bar {d} (p) = \exp \left(- \frac {1}{2 s _ {g ^ {*}} ^ {2}} \langle p - \mu_ {g ^ {*}}, n _ {g ^ {*}} \rangle^ {2}\right). \tag {5} +$$ + +A first strategy to enforce our regularization is to add term $|d(p) - \bar{d}(p)|$ to the optimization loss. While this approach works well to align Gaussians with the surface, we noticed that computing a slightly different loss relying on an SDF rather than on density further increases the alignment of Gaussians with the surface of the scene. For a given flat Gaussian, i.e., $s_g = 0$ , considering level sets is meaningless since all level sets would degenerate toward the plane passing through the center of the Gaussian $\mu_g$ with normal $n_g$ . The distance between point $p$ and the true surface of the scene would be approximately $|\langle p - \mu_{g'}, n_{g'} \rangle|$ , the distance from $p$ to this plane. Consequently, the zero-crossings of the Signed Distance Function + +$$ +\bar {f} (p) = \pm s _ {g *} \sqrt {- 2 \log (\bar {d} (p))} \tag {6} +$$ + +corresponds to the surface of the scene. More generally, we define + +$$ +f (p) = \pm s _ {g *} \sqrt {- 2 \log (d (p))} \tag {7} +$$ + +as the "ideal" distance function associated with the density function $d$ . This distance function corresponds to the true surface of the scene in an ideal scenario where $d = \bar{d}$ . We therefore take our regularization term $\mathcal{R}$ as + +$$ +\mathcal {R} = \frac {1}{| \mathcal {P} |} \sum_ {p \in \mathcal {P}} \left| | \hat {f} (p) | - | f (p) | \right|, \tag {8} +$$ + +by sampling 3D points $p$ and summing the differences at these points between the ideal SDF $f(p)$ and an estimate $\hat{f}(p)$ of the SDF of the surface created by the current Gaussians. $\mathcal{P}$ refers to the set of sampled points. During optimization, we simply want the zero-level sets of both the + +![](images/0fda687334d3dc4b93586b3ae1e96bd344c0ae51d878b5b833c5a193d5922480.jpg) +Figure 5. Efficiently estimating $\hat{f}(p)$ of the SDF of the surface generated from Gaussians. We render depth maps of the Gaussians, sample points $p$ in the viewpoint according to the distribution of the Gaussians. Value $\hat{f}(p)$ is taken as the 3D distance between $p$ and the intersection between the line of sight for $p$ and the depth map. + +SDF estimator $\hat{f}$ and the ideal SDF $f$ to align. Since a zero-level set is entirely determined by an unsigned distance function, we actually do not need to compute the sign of the SDFs and use absolute values. + +Computing efficiently $\hat{f}(p)$ is a priori challenging. To do so, we propose to use the depth maps of the Gaussians from the viewpoints used for training—these depth maps can be rendered efficiently by extending the splatting rasterizer. Then, as shown in Figure 5, for a point $p$ visible from a training viewpoint, $\hat{f}(p)$ is the difference between the depth of $p$ and the depth in the corresponding depth map at the projection of $p$ . Moreover, we sample points $p$ following the distribution of the Gaussians: + +$$ +p \sim \prod_ {g} \mathcal {N} \left(.; \mu_ {g}, \Sigma_ {g}\right), \tag {9} +$$ + +with $\mathcal{N}(:,\mu_g,\Sigma_g)$ the Gaussian distribution of mean $\mu_{g}$ and covariance $\Sigma_{g}$ as these points are likely to correspond to a high gradient for $\mathcal{R}$ . + +![](images/7793310e8a1663a1bc425bb1121acfd80c404fd4cb0d52af2380bd5663ebf64a.jpg) +Figure 6. Sampling points on a level set for Poisson reconstruction. Left: We sample points on the depth maps of the Gaussians and refine the point locations to move the points on the level set. Right: Comparison between the extracted mesh without (left) and with (right) our refinement step. Since splatted depth maps are not exact, using directly the depth points for reconstruction usually results in a large amount of noise and missing details. + +![](images/d262f28448b4d3af275cf2b66155ba011da0fa7967794a9bd0b45cb786fba190.jpg) + +![](images/2f706608e69c018da35d9debfa62abf0fe7d326101fcafab0e862679a53f0601.jpg) + +We also add a regularization term to encourage the normals of SDF $f$ and the normals of SDF $\bar{f}$ to be similar: + +$$ +\mathcal {R} _ {\text {N o r m}} = \frac {1}{| \mathcal {P} |} \sum_ {p \in \mathcal {P}} \left\| \frac {\nabla f (p)}{\| \nabla f (p) \| _ {2}} - n _ {g ^ {*}} \right\| _ {2} ^ {2}. \tag {10} +$$ + +# 4.2. Efficient Mesh Extraction + +To create a mesh from the Gaussians obtained after optimization using our regularization terms in Eq. (8) and Eq. (10), we sample 3D points on a level set of the density computed from the Gaussians. The level set depends on a level parameter $\lambda$ . Then, we obtain a mesh by simply running a Poisson reconstruction [14] on these points. Note that we can also easily assign the points with the normals of the SDF, which improves the mesh quality. + +The challenge is in efficiently identifying points lying on the level set. For this, as shown in Figure 6, we again rely on the depth maps of the Gaussians as seen from the training viewpoints. We first randomly sample pixels from each depth map. For each pixel $m$ , we sample its line of sight to find a 3D point on the level set. Formally, we sample $n$ points $p + t_i v$ , where $p$ is the 3D point in the depth map that reprojects on pixel $m$ , $v$ is the direction of the line of sight, and $t_i \in [-3\sigma_g(v), 3\sigma_g(v)]$ where $\sigma_g(v)$ is the standard deviation of the 3D Gaussian $g$ in the direction of the camera. The interval $[-3\sigma_g(v), 3\sigma_g(v)]$ is the confidence interval for the 99.7 confidence level of the 1D Gaussian function of $t$ along the ray. + +Then, we compute the density values $d_{i} = d(p + t_{i}v)$ from Eq. (1) of these sampled points. If there exist $i,j$ such that $d_{i} < \lambda < d_{j}$ , then there is a level set point located in this range. If so, we use linear interpolation to compute the coefficient $t^{*}$ such that $p + t^{*}v$ is the level set point closest to the camera, verifying $d(p + t^{*}v) = \lambda$ . We also compute the normals of the surface at points $\hat{p}$ , which we naturally + +![](images/5e1246092c67462f639fac33a618dd8fa91f5355a55569554ece5a95bc25dd3d.jpg) +Figure 7. Joint refinement of mesh and Gaussians. Left: We bind Gaussians to the triangles of the mesh. Depending on the number of triangles in the scene, we bind a different number of Gaussians per triangle, with predefined barycentric coordinates. Right: Mesh before and after joint refinement. + +define as the normalized analytical gradient of the density $\frac{\nabla d(\hat{p})}{\|\nabla d(\hat{p})\|_2}$ . + +Finally, we apply Poisson reconstruction to reconstruct a surface mesh from the level set points and their normals. + +# 4.3. Binding New 3D Gaussians to the Mesh + +Once we have extracted a first mesh, we can refine this mesh by binding new Gaussians to the mesh triangles and optimize the Gaussians and the mesh jointly using the Gaussian Splatting rasterizer. This enables the edition of the Gaussian splatting scene with popular mesh editing tools while keeping high-quality rendering thanks to the Gaussians. + +Given the initial mesh, we instantiate new 3D Gaussians on the mesh. More exactly, we associate a set of $n$ thin 3D Gaussians to each triangle of the mesh, sampled on the surface of the triangle, as illustrated in Figure 7. To do so, we slightly modify the structure of the original 3D Gaussian Splitting model. + +We explicitly compute the means of the Gaussians from the mesh vertices using predefined barycentric coordinates in the corresponding triangles during optimization. Also, the Gaussians have only 2 learnable scaling factors instead of 3 and only 1 learnable 2D rotation encoded with a complex number rather than a quaternion, to keep the Gaussians flat and aligned with the mesh triangles. More details about this parameterisation are given in the supplementary material. Like the original model, we also optimize an opacity value and a set of spherical harmonics for every Gaussian to encode the color emitted in all directions. + +Figure 7 shows an example of a mesh before and after refinement. Figure 1 and the supplementary material give examples of what can be done by editing the mesh. + +# 5. Experiments + +# 5.1. Implementation details + +All our models are optimized on a single GPU Nvidia Tesla V100 SXM2 32 Go. + +Regularization. For all scenes, we start by optimizing a Gaussian Splatting with no regularization for 7,000 iterations to let the 3D Gaussians position themselves without any additional constraint. Then, we perform 2,000 iterations with an additional entropy loss on the opacities $\alpha_{g}$ of the Gaussians, as a way to enforce them to become binary. + +Finally, we remove Gaussians with opacity values under 0.5 and perform 6,000 iterations with the regularization term introduced in Subsection 4.1, which makes a total of 15,000 iterations. To compute the density values of points from a Gaussian $g$ , we sum only the Gaussian functions from the 16 nearest Gaussians of $g$ and update the list of nearest neighbors every 500 iterations. Optimization takes between 15 and 45 minutes depending on the scene. + +Mesh extraction. For all experiments except the ablation presented in Table 2, we extract the $\lambda$ -level set of the density function for $\lambda = 0.3$ . We perform Poisson reconstruction with depth 10 and apply mesh simplification using quadric error metrics [9] to decrease the resolution of the meshes. Mesh extraction generally takes between 5 and 10 minutes depending on the scene. + +Joint refinement. We jointly refine the mesh and the bound 3D Gaussians for either 2,000, 7,000 or 15,000 iterations. Depending on the number of iterations, the duration of refinement goes from a few minutes to an hour. + +# 5.2. Real-Time Rendering of Real Scenes + +For evaluating our model, we follow the approach from the original 3D Gaussian Splatting paper [15] and compare the performance of several variations of our method SuGaR after refinement on real 3D scenes from 3 different datasets: Mip-NeRF360 [1], DeepBlending [12] and Tanks&Temples [16]. We call R-SuGaR-NK a refined SuGaR model optimized for $N$ iterations during refinement. + +Following [15], we select the same sets of 2 scenes from Tanks&Temples (Truck and Train) and 2 scenes from DeepBlending (Playroom and Dr. Johnson). However, due to licensing issues and the unavailability of the scenes Flowers and Treehill, we perform the evaluation of all methods only on 7 scenes from Mip-NeRF360 instead of the full set of 9 scenes. + +We compute the standard metrics PSNR, SSIM and LPIPS [44] to evaluate the quality of SuGaR's rendering using our extracted meshes and their bound surface Gaussians. Note that [6, 26, 39] also do not use plain textured + +mesh rendering. We compare to several baselines, some of them focusing only on Novel View Synthesis [2, 15, 23, 41] and others relying on a reconstructed mesh [6, 26, 39], just like SuGaR. Results on the Mip-NeRF360 dataset are given in Table 1. Results on Tanks&Temple and DeepBlending are similar and can be found in the supplementary material. + +Even though SuGaR focuses on aligning 3D Gaussians for reconstructing a high quality mesh during the first stage of its optimization, it significantly outperforms the state of the art methods for Novel View Synthesis using a mesh and reaches better performance than several famous models that focus only on rendering, such as I-NGP [23] and Plenoxels [41]. This performance is remarkable as SuGaR is able to extract a mesh significantly faster than other methods. + +Moreover, SuGaR even reaches performance similar to state-of-the-art models for rendering quality [2, 15] on some of the scenes used for evaluation. Two main reasons explain this performance. First, the mesh extracted after the first stage of optimization serves as an excellent initialization for positioning Gaussians when starting the refinement phase. Then, the Gaussians constrained to remain on the surface during refinement greatly increase the rendering quality as they play the role of an efficient texturing tool and help reconstructing very fine details missing in the extracted mesh. Additional qualitative results are available in Figure 4. + +# 5.3. Mesh Extraction + +To demonstrate the ability of our mesh extraction method for reconstructing high-quality meshes that are well-suited for view synthesis, we compare different mesh extraction algorithms. In particular, we optimize several variations of SuGaR by following the exact same pipeline as our standard model, except for the mesh extraction process: We either extract the mesh using a very fine marching cubes algorithm [21], by applying Poisson reconstruction [14] using the centers of the 3D Gaussians as the surface point cloud, or by applying our mesh extraction method on different level sets. Quantitative results are available in Table 2 and show the clear superiority of our approach for meshing 3D Gaussians. Figure 3 also illustrates how the marching cubes algorithm fails in this context. + +# 5.4. Mesh Rendering Ablation + +Table 3 provides additional results to quantify how various parameters impact rendering performance. In particular, we evaluate how the resolution of the mesh extraction, i.e., the number of triangles, modifies the rendering quality. For fair comparison, we increase the number of surface-aligned Gaussians per triangle when we decrease the number of triangles. Results show that increasing the number of vertices increases the quality of rendering with surface Gaussians, but meshes with less triangles are already able to reach state of the art results. + +
Indoor scenesOutdoor scenesAverage on all scenes
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
No mesh (except SuGaR)
Plenoxels [42]24.830.7660.42622.020.5420.46523.620.6700.443
INGP-Base [23]28.650.8400.28123.470.5710.41626.430.7250.339
INGP-Big [23]29.140.8630.24223.570.6020.37526.750.7510.299
Mip-NeRF360 [2]31.580.9140.18225.790.7460.24729.090.8420.210
3DGS [15]30.410.9200.18926.400.8050.17328.690.8700.182
R-SuGaR-15K (Ours)29.430.9100.21624.400.6990.30127.270.8200.253
With mesh
Mobile-NeRF [6]---21.950.4700.470---
NeRFMeshing [26]23.83--22.23--23.15--
BakedSDF [39]27.060.8360.258------
R-SuGaR-2K (Ours)26.290.8720.26222.970.6480.36024.870.7760.304
R-SuGaR-7K (Ours)28.730.9040.22624.160.6910.31326.770.8130.263
R-SuGaR-15K (Ours)29.430.9100.21624.400.6990.30127.270.8200.253
+ +Table 1. Quantitative evaluation of rendering quality on the Mip-NeRF360 dataset [2]. SuGaR is best among the methods that recover a mesh, and still performs well compared to NeRF methods and vanilla 3D Gaussian Splatting. + +
Extraction methodPSNR ↑SSIM ↑LPIPS ↓
Marching Cubes [21]23.910.7030.392
Poisson (centers) [14]23.760.7560.340
Ours (Surface level 0.1)24.620.7650.313
Ours (Surface level 0.3)24.870.7760.304
Ours (Surface level 0.5)24.910.7770.304
+ +Table 2. Ablation for different mesh extraction methods on the Mip-NeRF360 dataset [2] after applying our regularization term. For 'Poisson (centers)', we apply Poisson reconstruction [14] using as surface points the centers of the 3D Gaussians. For fair comparison, we calibrate the methods to enforce all extracted meshes to have approximately 1,000,000 vertices. + +
PSNR ↑SSIM ↑LPIPS ↓
1M vertices (3DGS)24.510.7680.295
1M vertices (UV)21.240.6090.478
200K vertices (3DGS)24.240.7570.300
200K vertices (UV)21.440.6560.419
+ +Table 3. Comparison between surface-aligned 3D Gaussians and an optimized traditional UV texture on the Mip-NeRF360 dataset [2]. For fair comparison, we only use the diffuse spherical harmonics component when rendering images with SuGaR. Using 3D Gaussians bound to the mesh greatly improves rendering quality, even though it contains less parameters than the UV texture. + +Then, we illustrate the benefits of using Gaussians aligned on the surface as a texturing tool for rendering meshes. To this end, we also optimize traditional UV textures on our meshes using differentiable mesh rendering + +with traditional triangle rasterization. Even though rendering with surface-aligned Gaussians provides better performance, rendering our meshes with traditional UV textures still produces satisfying results, which further illustrates the quality of our extracted meshes. Qualitative comparisons are provided in the supplementary material. + +# 6. Conclusion + +We proposed a very fast algorithm to obtain an accurate 3D triangle mesh for a scene via Gaussian Splatting. Moreover, by combining meshing and Gaussian Splatting, we make possible intuitive manipulation of the captured scenes and realistic rendering, offering new possibilities for creators. SuGaR does not come without limitations: Gaussians do tend to "cheat" on the geometry and depth by creating cavities to reproduce specular effects, instead of relying on spherical harmonics. SuGaR's regularization mitigates this issue, but Gaussians can still distort the surface in regions with intense specularity. In addition, SuGaR's assumption that the scene can be represented as a surface complicates the rendering of volumetric effects or fuzzy materials. + +# Acknowledgements. + +This project was funded in part by the European Union (ERC Advanced Grant explorer Funding ID #101097259). This work was granted access to the HPC resources of IDRIS under the allocation 2023-AD011013387R1 made by GENCI. We thank George Drettakis and Elliot Vincent for inspiring discussions and valuable feedback. + +# References + +[1] Jonathan T. Barron. Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. In International Conference on Computer Vision, 2021. 3, 7 +[2] Jonathan T. Barron. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In Conference on Computer Vision and Pattern Recognition, 2022. 3, 7, 8, 2, 4 +[3] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. NeRD: Neural Reflectance Decomposition from Image Collections. In International Conference on Computer Vision, 2021. 3 +[4] Chris Buehler, Michael Bosse, Leonard Mcmillan, Steven Gortler, and Michael Cohen. Unstructured Lumigraph Rendering. In ACM SIGGRAPH, 2001. 3 +[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial Radiance Fields. In European Conference on Computer Vision, 2022. 3 +[6] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. MobileNeRF: Exploiting the Polygon Rasterization Pipeline for Efficient Neural Field Rendering on Mobile Architectures. In Conference on Computer Vision and Pattern Recognition, 2023. 2, 3, 7, 8 +[7] Chong Bao and Bangbang Yang, Zeng Junyi, Bao Hujun, Zhang Yinda, Cui Zhaopeng, and Zhang Guofeng. NeuMesh: Learning Disentangled Neural Mesh-Based Implicit Field for Geometry and Texture Editing. In European Conference on Computer Vision, 2022. 3 +[8] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving Neural Implicit Surfaces Geometry with Patch Warping. In Conference on Computer Vision and Pattern Recognition, 2022. 3 +[9] Michael Garland and Paul S. Heckbert. Surface Simplification Using Quadric Error Metrics. In ACM SIGGRAPH, 1997. 7 +[10] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven Seitz. Multi-View Stereo for Community Photo Collections. In International Conference on Computer Vision, 2007. 3 +[11] Peter Hedman and Pratul P. Srinivasan. Baking Neural Radiance Fields for Real-Time View Synthesis. In International Conference on Computer Vision, 2021. 3 +[12] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep Blending for Free-Viewpoint Image-Based Rendering. In ACM SIGGRAPH, 2018. 3, 7, 2, 4 +[13] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Mitra. ReLU Fields: The Little Non-Linearity That Could. In ACM SIGGRAPH, 2022. 3 +[14] Michael M. Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson Surface Reconstruction. In Eurographics, 2006. 2, 6, 7, 8 +[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. In ACM SIGGRAPH, 2023. 1, 2, 4, 7, 8 + +[16] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking Large-Scale Scene Reconstruction. In ACM SIGGRAPH, 2017. 7, 2, 4 +[17] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-Based Neural Rendering with Per-View Optimization. In Computer Graphics Forum, 2021. 3 +[18] Zhengfei Kuang, Kyle Olszewski, Menglei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. NeROIC: Neural Rendering of Objects from Online Image Collections. In ACM SIGGRAPH, 2022. 3 +[19] Marc Levoy and Pat Hanrahan. Light Field Rendering. In ACM SIGGRAPH, 1996. 3 +[20] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-Fidelity Neural Surface Reconstruction. In Conference on Computer Vision and Pattern Recognition, 2023. 2, 3 +[21] William E. Lorensen and Harvey E. Cline. Marching Cubes: A High Resolution 3D Surface Construction Algorithm. In ACM SIGGRAPH, 1987. 2, 3, 7, 8 +[22] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, 2020. 1, 3 +[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. In ACM SIGGRAPH, 2022. 3, 7, 8, 2 +[24] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying Neural Implicit Surfaces and Radiance Fields for Multi-View Reconstruction. In International Conference on Computer Vision, 2021. 3 +[25] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary Devito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An Imperative Style, High-Performance Deep Learning Library. In Advances in Neural Information Processing Systems. Curran Associates Inc., 2019. 2 +[26] Marie-Julie Rakotosaona, Fabian Manhardt, Diego Martin Arroyo, Michael Niemeyer, Abhijit Kundu, and Federico Tombari. NeRFMeshing: Distilling Neural Radiance Fields into Geometrically-Accurate 3D Meshes. In DV, 2023. 2, 3, 7, 8 +[27] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3D Deep Learning with PyTorch3D. In arXiv Preprint, 2020. 2 +[28] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. KiloNeRF: Speeding Up Neural Radiance Fields with Thousands of Tiny MLPs. In International Conference on Computer Vision, 2021. 3 +[29] Gernot Riegler and Vladlen Koltun. Free View Synthesis. In European Conference on Computer Vision, 2020. 3 + +[30] Gernot Riegler and Vladlen Koltun. Stable View Synthesis. In Conference on Computer Vision and Pattern Recognition, 2021. 3 +[31] Darius Rückert, Linus Franke, and Marc Stamminger. ADOP: Approximate Differentiable One-Pixel Point Rendering. In ACM SIGGRAPH, 2022. 4 +[32] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo Tourism: Exploring Photo Collections in 3D. In ACM SIGGRAPH, 2006. 3, 4 +[33] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural Reflectance and Visibility Fields for Relighting and View Synthesis. In Conference on Computer Vision and Pattern Recognition, 2021. 3 +[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct Voxel Grid Optimization: Super-Fast Convergence for Radiance Fields Reconstruction. In Conference on Computer Vision and Pattern Recognition, 2022. 3 +[35] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields. In Conference on Computer Vision and Pattern Recognition, 2022. 3 +[36] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-View Reconstruction. In Advances in Neural Information Processing Systems, 2021. 2, 3 +[37] Daniel N. Wood, Daniel I. Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H. Salesin, and Werner Stuetzle. Surface Light Fields for 3D Photography. In ACM SIGGRAPH, 2000. 3 +[38] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume Rendering of Neural Implicit Surfaces. In Advances in Neural Information Processing Systems, 2021. 2, 3 +[39] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, and Jonathan T. Barron. BakedSDF: Meshing Neural SDFs for Real-Time View Synthesis. In ACM SIGGRAPH, 2023. 2, 3, 7, 8 +[40] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees For Real-Time Rendering of Neural Radiance Fields. In International Conference on Computer Vision, 2021. 3 +[41] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. In Conference on Computer Vision and Pattern Recognition, 2022. 3, 7 +[42] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. In Conference on Computer Vision and Pattern Recognition, 2022. 8, 2 +[43] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse Rendering with Spherical Gaussians for Physics-Based Material Editing and Relighting. In Conference on Computer Vision and Pattern Recognition, 2021. 3 +[44] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of + +Deep Features as a Perceptual Metric. In Conference on Computer Vision and Pattern Recognition, 2018. 7, 2, 4 \ No newline at end of file diff --git a/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/images.zip b/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a34dee54934c79b008a72fe693fed3fe4e527613 --- /dev/null +++ b/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a598b088d88250ed975ce16a174df3f2d3c36b25502354a51e5947f94661f440 +size 563285 diff --git a/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/layout.json b/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fc60635880b79188a3de816d292baebd6908ae6f --- /dev/null +++ b/2024/SuGaR_ Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering/layout.json @@ -0,0 +1,9272 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 510, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 510, + 140 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 510, + 140 + ], + "type": "text", + "content": "SuGaR: Surface-Aligned Gaussian Splitting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 161, + 444, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 161, + 444, + 203 + ], + "spans": [ + { + "bbox": [ + 149, + 161, + 444, + 203 + ], + "type": "text", + "content": "Antoine Guédon Vincent Lepetit \nLIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, France \nhttps://anttwo.github.io/sugar/" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 212, + 124, + 340 + ], + "blocks": [ + { + "bbox": [ + 48, + 212, + 124, + 340 + ], + "lines": [ + { + "bbox": [ + 48, + 212, + 124, + 340 + ], + "spans": [ + { + "bbox": [ + 48, + 212, + 124, + 340 + ], + "type": "image", + "image_path": "6673930ff670e4525dde86ea891322dc4edf5ae835eac2957ca13ae2f137f242.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 346, + 547, + 390 + ], + "lines": [ + { + "bbox": [ + 45, + 346, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 45, + 346, + 547, + 390 + ], + "type": "text", + "content": "Figure 1. We introduce a method that extracts accurate and editable meshes from 3D Gaussian Splitting representations within minutes on a single GPU. The meshes can be edited, animated, composited, etc. with very realistic Gaussian Splitting rendering, offering new possibilities for Computer Graphics. Note for example that we changed the posture of the robot between the captured scene on the bottom left and the composited scene on the right. The supplementary material provides more examples, including a video illustrating our results." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 213, + 309, + 340 + ], + "blocks": [ + { + "bbox": [ + 133, + 213, + 309, + 340 + ], + "lines": [ + { + "bbox": [ + 133, + 213, + 309, + 340 + ], + "spans": [ + { + "bbox": [ + 133, + 213, + 309, + 340 + ], + "type": "image", + "image_path": "1e71b34782521a295dba22014fb33764264018e29489fec3a41da095d26a740b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 213, + 545, + 340 + ], + "blocks": [ + { + "bbox": [ + 318, + 213, + 545, + 340 + ], + "lines": [ + { + "bbox": [ + 318, + 213, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 318, + 213, + 545, + 340 + ], + "type": "image", + "image_path": "10648b5890419237a4fb6653f09ebb64111dc67816d3d828c0fbbb2c4b6de2cf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 404, + 192, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 404, + 192, + 416 + ], + "spans": [ + { + "bbox": [ + 143, + 404, + 192, + 416 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 421, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 421, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 45, + 421, + 289, + 700 + ], + "type": "text", + "content": "We propose a method to allow precise and extremely fast mesh extraction from 3D Gaussian Splatting [15]. Gaussian Splatting has recently become very popular as it yields realistic rendering while being significantly faster to train than NeRFs. It is however challenging to extract a mesh from the millions of tiny 3D Gaussians as these Gaussians tend to be unorganized after optimization and no method has been proposed so far. Our first key contribution is a regularization term that encourages the Gaussians to align well with the surface of the scene. We then introduce a method that exploits this alignment to extract a mesh from the Gaussians using Poisson reconstruction, which is fast, scalable, and preserves details, in contrast to the Marching Cubes algorithm usually applied to extract meshes from Neural SDFs. Finally, we introduce an optional refinement strategy that binds Gaussians to the surface of the mesh, and jointly optimizes these Gaussians and the mesh through Gaussian splatting rendering. This enables easy editing, sculpting, animating, and relighting of the Gaussians by manipulating the mesh instead of the Gaussians themselves. Retrieving such an editable mesh for realistic rendering is done within minutes with our method, compared to hours with the state-of-the-art method on SDFs, while providing a better rendering quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 404, + 386, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 386, + 416 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 386, + 416 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 430, + 545, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 547 + ], + "type": "text", + "content": "After NeRFs [22], 3D Gaussian Splatting [15] has recently become very popular for capturing a 3D scene and rendering it from novel points of view. 3D Gaussian Splatting optimizes the positions, orientations, appearances (represented as spherical harmonics), and alpha blending of many tiny 3D Gaussians on the basis of a set of training images of the scene to capture the scene geometry and appearance. Because rendering the Gaussians is much faster than rendering a neural field, 3D Gaussian Splatting is much faster than NeRFs and can capture a scene in a few minutes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 551, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 551, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 551, + 546, + 712 + ], + "type": "text", + "content": "While the Gaussians allow very realistic renderings of the scene, it is still however challenging to extract the surface of the scene from them: As shown in Figure 3, after optimization by 3D Gaussian Splatting, the Gaussians do not take an ordered structure in general and do not correspond well to the actual surface of the scene. In addition to the surface itself, it is also often desirable to represent the scene as a mesh, which remains the representation of choice in many pipelines: A mesh-based representation allows for powerful tools for editing, sculpting, animating, and relighting the scene. Because the Gaussians after Gaussian Splatting are unstructured, it is very challenging to extract a mesh from them. Note that this is also challenging with NeRFs albeit for different reasons." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "text", + "content": "5354" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 167, + 135 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 167, + 135 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 167, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 167, + 135 + ], + "type": "image", + "image_path": "965e4b706f98280301b6f99dcf08dc1d3cc1bda348d4aebf93eb5600c76ab7b7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 71, + 290, + 135 + ], + "blocks": [ + { + "bbox": [ + 176, + 71, + 290, + 135 + ], + "lines": [ + { + "bbox": [ + 176, + 71, + 290, + 135 + ], + "spans": [ + { + "bbox": [ + 176, + 71, + 290, + 135 + ], + "type": "image", + "image_path": "63849b2eb6568bab393e1eb9db3e9606a0f34ca11c35142170ea5710115c3897.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 137, + 167, + 201 + ], + "blocks": [ + { + "bbox": [ + 53, + 137, + 167, + 201 + ], + "lines": [ + { + "bbox": [ + 53, + 137, + 167, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 167, + 201 + ], + "type": "image", + "image_path": "7f0927d0b0e6405cd930616ebc9dd362849d6074393daefa14f02ea6d4638a45.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 287, + 254 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 254 + ], + "type": "text", + "content": "Figure 2. Our algorithm can extract a highly detailed mesh from any 3D Gaussian Splatting scene [15] within minutes on a single GPU (top: Renderings of our meshes without texture, bottom: Renderings of the meshes with bound Gaussians)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 137, + 290, + 201 + ], + "blocks": [ + { + "bbox": [ + 176, + 137, + 290, + 201 + ], + "lines": [ + { + "bbox": [ + 176, + 137, + 290, + 201 + ], + "spans": [ + { + "bbox": [ + 176, + 137, + 290, + 201 + ], + "type": "image", + "image_path": "414eed21c26dc56e63563eb1defede617594decabb74654cbaab4632ce6bb7ca.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 222, + 545, + 320 + ], + "lines": [ + { + "bbox": [ + 304, + 222, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 545, + 320 + ], + "type": "text", + "content": "Figure 3. Extracting a mesh from Gaussians. Without regularization, the Gaussians have no special arrangement after optimization, which makes extracting a mesh very difficult. Without our regularization term, Marching Cubes fail to extract an acceptable mesh. With our regularization term, Marching Cubes recover an extremely noisy mesh even with a very fine 3D grid. Our scalable extraction method obtains a mesh even without our regularization term. Still, the mesh is noisy. By contrast, our full method succeeds in reconstructing an accurate mesh very efficiently." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 278, + 286, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 286, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 286, + 395 + ], + "type": "text", + "content": "In this paper, we first propose a regularization term that encourages the Gaussians to be well distributed over the scene surface so that the Gaussians capture much better the scene geometry, as shown in Figure 3. Our approach is to derive a volume density from the Gaussians under the assumption that the Gaussians are flat and well distributed over the scene surface. By minimizing the difference between this density and the actual one computed from the Gaussians during optimization, we encourage the 3D Gaussians to represent well the surface geometry." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 396, + 286, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 286, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 286, + 570 + ], + "type": "text", + "content": "Thanks to this regularization term, it becomes easier to extract a mesh from the Gaussians. In fact, since we introduce a density function to evaluate our regularization term, a natural approach would be to extract level sets of this density function. However, Gaussian Splatting performs densification in order to capture details of the scene with high fidelity, which results in a drastic increase in the number of Gaussians. Real scenes typically end up with one or several millions of 3D Gaussians with different scales and rotations, the majority of them being extremely small in order to reproduce texture and details in the scene. This results in a density function that is close to zero almost everywhere, and the Marching Cubes algorithm [21] fails to extract proper level sets of such a sparse density function even with a fine voxel grid, as also shown in Figure 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 572, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 572, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 286, + 700 + ], + "type": "text", + "content": "Instead, we introduce a method that very efficiently samples points on the visible part of a level set of the density function, allowing us to run the Poisson reconstruction algorithm [14] on these points to obtain a triangle mesh. This approach is scalable, by contrast with the Marching Cubes algorithm for example, and reconstructs a surface mesh within minutes on a single GPU, compared to other state of the art methods relying on Neural SDFs for extracting meshes from radiance fields, that require at least 24 hours on one GPU [20, 36, 38, 39] and rely on multiple GPUs to speed up the process [26]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "As illustrated in Figures 2 and 4, our method produces" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 80, + 387, + 131 + ], + "blocks": [ + { + "bbox": [ + 374, + 71, + 476, + 79 + ], + "lines": [ + { + "bbox": [ + 374, + 71, + 476, + 79 + ], + "spans": [ + { + "bbox": [ + 374, + 71, + 476, + 79 + ], + "type": "text", + "content": "without our regularization term" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 80, + 387, + 131 + ], + "lines": [ + { + "bbox": [ + 315, + 80, + 387, + 131 + ], + "spans": [ + { + "bbox": [ + 315, + 80, + 387, + 131 + ], + "type": "image", + "image_path": "fab143a17035a4621c16f71a04dc960390996ce00a5ee09eef27a7f69788960d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 390, + 80, + 463, + 131 + ], + "blocks": [ + { + "bbox": [ + 390, + 80, + 463, + 131 + ], + "lines": [ + { + "bbox": [ + 390, + 80, + 463, + 131 + ], + "spans": [ + { + "bbox": [ + 390, + 80, + 463, + 131 + ], + "type": "image", + "image_path": "62c6b2f08a971899191a5b562162f4a4a0d84c2fac41173d6dd4fdd4d21db07f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 464, + 80, + 536, + 131 + ], + "blocks": [ + { + "bbox": [ + 464, + 80, + 536, + 131 + ], + "lines": [ + { + "bbox": [ + 464, + 80, + 536, + 131 + ], + "spans": [ + { + "bbox": [ + 464, + 80, + 536, + 131 + ], + "type": "image", + "image_path": "ad83aca945c2ef13508a04bdafb0e6aa0808eddaf7faabbba997f79174034c89.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 315, + 144, + 387, + 194 + ], + "blocks": [ + { + "bbox": [ + 379, + 134, + 471, + 143 + ], + "lines": [ + { + "bbox": [ + 379, + 134, + 471, + 143 + ], + "spans": [ + { + "bbox": [ + 379, + 134, + 471, + 143 + ], + "type": "text", + "content": "with our regularization term" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 144, + 387, + 194 + ], + "lines": [ + { + "bbox": [ + 315, + 144, + 387, + 194 + ], + "spans": [ + { + "bbox": [ + 315, + 144, + 387, + 194 + ], + "type": "image", + "image_path": "4015513f1d9b4cc0061301924259fc7103dea1c2a531de2b078dd3f682cdbdbe.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 197, + 383, + 212 + ], + "lines": [ + { + "bbox": [ + 319, + 197, + 383, + 212 + ], + "spans": [ + { + "bbox": [ + 319, + 197, + 383, + 212 + ], + "type": "text", + "content": "zoom on Gaussians on a planar surface" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 390, + 144, + 463, + 194 + ], + "blocks": [ + { + "bbox": [ + 390, + 144, + 463, + 194 + ], + "lines": [ + { + "bbox": [ + 390, + 144, + 463, + 194 + ], + "spans": [ + { + "bbox": [ + 390, + 144, + 463, + 194 + ], + "type": "image", + "image_path": "f9718805da1261a9b074b2a8f184814cf14f0acfe7762c7ef3b4057ab334615d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 197, + 454, + 212 + ], + "lines": [ + { + "bbox": [ + 399, + 197, + 454, + 212 + ], + "spans": [ + { + "bbox": [ + 399, + 197, + 454, + 212 + ], + "type": "text", + "content": "mesh with Marching Cubes" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 464, + 144, + 535, + 194 + ], + "blocks": [ + { + "bbox": [ + 464, + 144, + 535, + 194 + ], + "lines": [ + { + "bbox": [ + 464, + 144, + 535, + 194 + ], + "spans": [ + { + "bbox": [ + 464, + 144, + 535, + 194 + ], + "type": "image", + "image_path": "38688d6186e8082e57213c850b4fedc5c366bbb99ad6bf40978505ff29e8c3eb.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 197, + 529, + 212 + ], + "lines": [ + { + "bbox": [ + 470, + 197, + 529, + 212 + ], + "spans": [ + { + "bbox": [ + 470, + 197, + 529, + 212 + ], + "type": "text", + "content": "mesh with our extraction method" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 346, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 427 + ], + "type": "text", + "content": "high quality meshes. The challenge is in efficiently identifying points lying on the level set. To do this, we rely on the Gaussians depth maps seen from the training viewpoints. These depth maps can be obtained by extending the Gaussian Splatting rasterizer, and we show how to accurately sample points on the level set starting from these depth maps." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 429, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 568 + ], + "type": "text", + "content": "Finally, after extracting this mesh, we propose an optional refinement strategy that jointly optimizes the mesh and a set of 3D Gaussians through Gaussian splatting rendering only. This optimization enables high-quality rendering of the mesh using Gaussian splatting rendering rather than traditional textured mesh rendering. This results in higher performance in terms of rendering quality than other radiance field models relying on an underlying mesh at inference [6, 26, 39]. As shown in Figure 1, this makes possible the use of traditional mesh-editing tools for editing a Gaussian Splatting representation of a scene, offering endless possibilities for Computer Graphics." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 570, + 468, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 468, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 468, + 580 + ], + "type": "text", + "content": "To summarize, our contributions are:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 583, + 545, + 687 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 306, + 583, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 583, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 306, + 583, + 545, + 605 + ], + "type": "text", + "content": "- a regularization term that makes the Gaussians capture accurately the geometry of the scene;" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 606, + 544, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 606, + 544, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 606, + 544, + 628 + ], + "type": "text", + "content": "- an efficient algorithm that extracts an accurate mesh from the Gaussians within minutes;" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 630, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 630, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 306, + 630, + 545, + 687 + ], + "type": "text", + "content": "- a method to bind the Gaussians to the mesh, resulting in a more accurate mesh, higher rendering quality than state of the art methods using a mesh for Novel View Synthesis [6, 26, 39], and allowing editing the scene in many different ways." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We call our approach SuGaR. In the remainder of the paper, we discuss related work, give a brief overview of vanilla" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "text", + "content": "5355" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "3D Gaussian Splatting, describe SuGaR, and compare it to the state of the art." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 109, + 134, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 109, + 134, + 121 + ], + "spans": [ + { + "bbox": [ + 47, + 109, + 134, + 121 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 130, + 286, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 130, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 130, + 286, + 213 + ], + "type": "text", + "content": "Image-based rendering (IBR) methods rely on a set of two-dimensional images of a scene to generate a representation of the scene and render novel views. The very first novel-view synthesis approaches were based on light fields [19], and developed the concept of volume rendering for novel views. Their work emphasized the importance of efficiently traversing volumetric data to produce realistic images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 214, + 286, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 214, + 286, + 249 + ], + "spans": [ + { + "bbox": [ + 47, + 214, + 286, + 249 + ], + "type": "text", + "content": "Various scene representations have been proposed since, such as triangle meshes, point clouds, voxel grids, multiplane images, or neural implicit functions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 267, + 286, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 286, + 445 + ], + "type": "text", + "content": "Traditional mesh-based IBR methods. Structure-from-motion (SfM) [32] and subsequent multi-view stereo (MVS) [10] allow for 3D reconstruction of surfaces, leading to the development of several view synthesis algorithms relying on triangle meshes as the primary 3D representation of scenes. Such algorithms consider textured triangles or warp and blend captured images on the mesh surface to generate novel views [4, 12, 37]. [29, 30] consider deep learning-based mesh representations for better view synthesis, bridging the gap between traditional graphics and modern machine learning techniques. While these mesh-based methods take advantage of existing graphics hardware and software for efficient rendering, they struggle with the capture of accurate geometry and appearance in complex regions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 464, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 286, + 581 + ], + "type": "text", + "content": "Volumetric IBR methods. Volumetric methods use voxel grids, multiplane images, or neural networks to represent scenes as continuous volumetric functions of density and color. Recently, Neural Radiance Fields (NeRF) [22] introduced a novel scene representation based on a continuous volumetric function parameterized by a multilayer perceptron (MLP). NeRF produces photorealistic renderings with fine details and view-dependent effects, achieved through volumetric ray tracing. However, the original NeRF is computationally expensive and memory intensive." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 583, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 583, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 583, + 286, + 713 + ], + "type": "text", + "content": "To address these challenges, several works have improved NeRF's performance and scalability. These methods leverage discretized or sparse volumetric representations like voxel grids and hash tables as ways to store learnable features acting as positional encodings for 3D points [5, 13, 23, 34, 41], hierarchical sampling strategies [2, 11, 28, 40], or low-rank approximations [5]. However, they still rely on volumetric ray marching, which is incompatible with standard graphics hardware and software designed for rendering polygonal surfaces. Recent works have proposed modifying the NeRF's representation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "of geometry and emitted radiance to allow for better reconstruction of specular materials [35] or relighting the scene through an explicit decomposition into material and lighting properties [3, 18, 33, 43]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 137, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 137, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 137, + 545, + 302 + ], + "type": "text", + "content": "Hybrid IBR methods. Some methods build on differentiable rendering to combine the advantages of mesh-based and volumetric methods, and allow for surface reconstruction as well as better editability. They use a hybrid volume-surface representation, which enables high-quality meshes suitable for downstream graphics applications while efficiently modeling view-dependent appearance. In particular, some works optimize neural signed distance functions (SDF) by training neural radiance fields in which the density is derived as a differentiable transformation of the SDF [7, 8, 20, 24, 36, 38]. A triangle mesh can finally be reconstructed from the SDF by applying the Marching Cubes algorithm [21]. However, most of these methods do not target real-time rendering." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 304, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 422 + ], + "type": "text", + "content": "Alternatively, other approaches \"bake\" the rendering capacity of an optimized NeRF or neural SDF into a much efficient structure relying on an underlying triangle mesh [6] that could benefit from the traditional triangle rasterization pipeline. In particular, the recent BakedSDF [39] reconstructs high quality meshes by optimizing a full neural SDF model, baking it into a high-resolution triangle mesh that combines mesh rendering for interpolating features and deep learning to translate these features into images, and finally optimizes a view-dependent appearance model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 422, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 481 + ], + "type": "text", + "content": "However, even though it achieves real-time rendering and produces impressive meshes of the surface of the scene, this model demands training a full neural SDF with an architecture identical to Mip-NeRF360 [1], which necessitates 48 hours of training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 482, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 545, + 588 + ], + "type": "text", + "content": "Similarly, the recent method NeRFMeshing [26] proposes to also bake any NeRF model into a mesh structure, achieving real-time rendering. However, the meshing performed in this method lowers the quality of the rendering and results in a PSNR much lower than our method. Additionally, this method still requires training a full NeRF model beforehand, and needs approximately an hour of training on 8 V100 NVIDIA GPUs to allow for mesh training and extraction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 589, + 545, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 589, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 545, + 648 + ], + "type": "text", + "content": "Our method is much faster at retrieving a 3D mesh from 3D Gaussian Splatting, which is itself much faster than NeRFs. As our experiments show, our rendering done by bounding Gaussians to the mesh results in higher quality than previous solutions based on meshes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "Point-based IBR methods. Alternatively, point-based representations for radiance field excel at modeling thin geometry and leverage fast point rasterization pipelines to render images using " + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "-blending rather than ray-marching [17," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "text", + "content": "5356" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "31]. In particular, the very recent 3D Gaussian Splitting model [15] allows for optimizing and rendering scenes with speed and quality never seen before." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 118, + 178, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 118, + 178, + 132 + ], + "spans": [ + { + "bbox": [ + 47, + 118, + 178, + 132 + ], + "type": "text", + "content": "3. 3D Gaussian Splitting" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 137, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 137, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 137, + 287, + 243 + ], + "type": "text", + "content": "For the sake of completeness, we briefly describe the original 3D Gaussian Splitting method here. The scene is represented as a (large) set of Gaussians, where each Gaussian " + }, + { + "bbox": [ + 46, + 137, + 287, + 243 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 137, + 287, + 243 + ], + "type": "text", + "content": " is represented by its mean " + }, + { + "bbox": [ + 46, + 137, + 287, + 243 + ], + "type": "inline_equation", + "content": "\\mu_{g}" + }, + { + "bbox": [ + 46, + 137, + 287, + 243 + ], + "type": "text", + "content": " and its covariance " + }, + { + "bbox": [ + 46, + 137, + 287, + 243 + ], + "type": "inline_equation", + "content": "\\Sigma_{g}" + }, + { + "bbox": [ + 46, + 137, + 287, + 243 + ], + "type": "text", + "content": " is parameterized by a scaling vector " + }, + { + "bbox": [ + 46, + 137, + 287, + 243 + ], + "type": "inline_equation", + "content": "s_g\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 137, + 287, + 243 + ], + "type": "text", + "content": " and a quaternion " + }, + { + "bbox": [ + 46, + 137, + 287, + 243 + ], + "type": "inline_equation", + "content": "q_{g}\\in \\mathbb{R}^{4}" + }, + { + "bbox": [ + 46, + 137, + 287, + 243 + ], + "type": "text", + "content": " encoding the rotation of the Gaussian. In addition, each Gaussian is associated with its opacity " + }, + { + "bbox": [ + 46, + 137, + 287, + 243 + ], + "type": "inline_equation", + "content": "\\alpha_{g}\\in [0,1]" + }, + { + "bbox": [ + 46, + 137, + 287, + 243 + ], + "type": "text", + "content": " and a set of spherical harmonics coordinates describing the colors emitted by the Gaussian for all directions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 244, + 287, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 244, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 287, + 338 + ], + "type": "text", + "content": "An image of a set of Gaussians can be rendered from a given viewpoint thanks to a rasterizer. This rasterizer splats the 3D Gaussians into 2D Gaussians parallel to the image plane for rendering, which results in an extremely fast rendering process. This is the key component that makes 3D Gaussian Splating much faster than NeRFs, as it is much faster than the ray-marching compositing required in the optimization of NeRFs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 339, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 445 + ], + "type": "text", + "content": "Given a set of images, the set of Gaussians is initialized from the point cloud produced by SfM [32]. The Gaussians' parameters (means, quaternions, scaling vectors, but also opacities and spherical harmonics parameters) are optimized to make the renderings of the Gaussians match the input images. During optimization, more Gaussians are added to better fit the scene's geometry. As a consequence, Gaussian Splatting generally produces scenes with millions of Gaussians that can be extremely small." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 455, + 102, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 455, + 102, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 102, + 467 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 475, + 203, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 203, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 203, + 487 + ], + "type": "text", + "content": "We present our SuGaR in this section:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 487, + 287, + 616 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 47, + 487, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 287, + 522 + ], + "type": "text", + "content": "- First, we detail our loss term that enforces the alignment of the 3D Gaussians with the surface of the scene during the optimization of Gaussian Splatting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 523, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 287, + 557 + ], + "type": "text", + "content": "- We then detail our method that exploits this alignment for extracting a highly detailed mesh from the Gaussians within minutes on a single GPU." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 559, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 287, + 616 + ], + "type": "text", + "content": "- Finally, we describe our optional refinement strategy that jointly optimizes the mesh and 3D Gaussians located on the surface of the mesh using Gaussian Splatting rendering. This strategy results in a new set of Gaussians bound to an editable mesh." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 624, + 257, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 257, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 257, + 636 + ], + "type": "text", + "content": "4.1. Aligning the Gaussians with the Surface" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 643, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 287, + 713 + ], + "type": "text", + "content": "As discussed in the introduction, to facilitate the creation of a mesh from the Gaussians, we introduce a regularization term into the Gaussian Splitting optimization that encourages the Gaussians to be aligned with the surface of the scene and well distributed over this surface. Our approach is to derive an SDF from the Gaussians under the assumption" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "that the Gaussians have the desired properties. By minimizing the difference between this SDF and the actual SDF computed for the Gaussians, we encourage the Gaussians to have these properties." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 129, + 545, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 129, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 304, + 129, + 545, + 176 + ], + "type": "text", + "content": "For a given Gaussian Splitting scene, we start by considering the corresponding density function " + }, + { + "bbox": [ + 304, + 129, + 545, + 176 + ], + "type": "inline_equation", + "content": "d:\\mathbb{R}^3\\to \\mathbb{R}_+" + }, + { + "bbox": [ + 304, + 129, + 545, + 176 + ], + "type": "text", + "content": " computed as the sum of the Gaussian values weighted by their alpha-blending coefficients at any space location " + }, + { + "bbox": [ + 304, + 129, + 545, + 176 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 129, + 545, + 176 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 184, + 545, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 545, + 213 + ], + "type": "interline_equation", + "content": "d (p) = \\sum_ {g} \\alpha_ {g} \\exp \\left(- \\frac {1}{2} \\left(p - \\mu_ {g}\\right) ^ {T} \\Sigma_ {g} ^ {- 1} \\left(p - \\mu_ {g}\\right)\\right), \\tag {1}", + "image_path": "f99d8c2aec860b88368202c4c709b690bce74dca7b78315fd435325707f65aef.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 223, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 545, + 270 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 304, + 223, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mu_g, \\Sigma_g" + }, + { + "bbox": [ + 304, + 223, + 545, + 270 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 223, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\alpha_g" + }, + { + "bbox": [ + 304, + 223, + 545, + 270 + ], + "type": "text", + "content": " are the centers, covariances, and alpha-blending coefficients of the Gaussians, respectively. Let us consider what this density function becomes if the Gaussians are well distributed and aligned with the surface." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 278, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 545, + 361 + ], + "type": "text", + "content": "First, in such scenario, the Gaussians would have limited overlap with their neighbors. As illustrated in Figure 3 (top-left), this is not the case in general. Then, for any point " + }, + { + "bbox": [ + 304, + 278, + 545, + 361 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 278, + 545, + 361 + ], + "type": "text", + "content": " close to the surface of the scene, the Gaussian " + }, + { + "bbox": [ + 304, + 278, + 545, + 361 + ], + "type": "inline_equation", + "content": "g^{*}" + }, + { + "bbox": [ + 304, + 278, + 545, + 361 + ], + "type": "text", + "content": " closest to the point " + }, + { + "bbox": [ + 304, + 278, + 545, + 361 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 278, + 545, + 361 + ], + "type": "text", + "content": " is likely to contribute much more than others to the density value " + }, + { + "bbox": [ + 304, + 278, + 545, + 361 + ], + "type": "inline_equation", + "content": "d(p)" + }, + { + "bbox": [ + 304, + 278, + 545, + 361 + ], + "type": "text", + "content": ". We could then approximate the Gaussian density at " + }, + { + "bbox": [ + 304, + 278, + 545, + 361 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 278, + 545, + 361 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 336, + 370, + 545, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 370, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 336, + 370, + 545, + 397 + ], + "type": "interline_equation", + "content": "\\alpha_ {g ^ {*}} \\exp \\left(- \\frac {1}{2} (p - \\mu_ {g ^ {*}}) ^ {T} \\Sigma_ {g ^ {*}} ^ {- 1} (p - \\mu_ {g ^ {*}})\\right), \\tag {2}", + "image_path": "4c2f2c860acbe02ac9e8f270433b717e28e83b533749b237fb271f77468265fa.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 404, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 545, + 429 + ], + "type": "text", + "content": "where the \"closest Gaussian\" " + }, + { + "bbox": [ + 304, + 404, + 545, + 429 + ], + "type": "inline_equation", + "content": "g^{*}" + }, + { + "bbox": [ + 304, + 404, + 545, + 429 + ], + "type": "text", + "content": " is taken as the Gaussian with the largest contribution at point " + }, + { + "bbox": [ + 304, + 404, + 545, + 429 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 404, + 545, + 429 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 339, + 436, + 545, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 436, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 339, + 436, + 545, + 456 + ], + "type": "interline_equation", + "content": "g ^ {*} = \\arg \\min _ {g} \\left\\{\\left(p - \\mu_ {g}\\right) ^ {T} \\Sigma_ {g} ^ {- 1} \\left(p - \\mu_ {g}\\right) \\right\\}. \\tag {3}", + "image_path": "c2bc45ca4c881f7f4f92d1c8e84cd36660c8a5dd3270294fe908276059c63d77.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 464, + 545, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 511 + ], + "type": "text", + "content": "Eq. (2) thus considers that the contribution of the closest Gaussian " + }, + { + "bbox": [ + 304, + 464, + 545, + 511 + ], + "type": "inline_equation", + "content": "g^{*}" + }, + { + "bbox": [ + 304, + 464, + 545, + 511 + ], + "type": "text", + "content": " to the density at " + }, + { + "bbox": [ + 304, + 464, + 545, + 511 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 464, + 545, + 511 + ], + "type": "text", + "content": " is much higher than the contribution of the other Gaussians. This will help us encourage the Gaussians to be well spread." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 520, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 567 + ], + "type": "text", + "content": "We also would like the 3D Gaussians to be flat, as they would then be aligned more closely with the surface of the mesh. Consequently, every Gaussian " + }, + { + "bbox": [ + 304, + 520, + 545, + 567 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 520, + 545, + 567 + ], + "type": "text", + "content": " would have one of its three scaling factors close to 0 and:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 334, + 575, + 545, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 575, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 334, + 575, + 545, + 602 + ], + "type": "interline_equation", + "content": "\\left(p - \\mu_ {g}\\right) ^ {T} \\Sigma_ {g} ^ {- 1} \\left(p - \\mu_ {g}\\right) \\approx \\frac {1}{s _ {g} ^ {2}} \\langle p - \\mu_ {g}, n _ {g} \\rangle^ {2}, \\tag {4}", + "image_path": "9869e56a9bf45714d9c94f26c55a9f42c2bded8aee390d2612c1c7765e57ba74.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 610, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 610, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 610, + 545, + 693 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 610, + 545, + 693 + ], + "type": "inline_equation", + "content": "s_g" + }, + { + "bbox": [ + 304, + 610, + 545, + 693 + ], + "type": "text", + "content": " the smallest scaling factor of the Gaussian and " + }, + { + "bbox": [ + 304, + 610, + 545, + 693 + ], + "type": "inline_equation", + "content": "n_g" + }, + { + "bbox": [ + 304, + 610, + 545, + 693 + ], + "type": "text", + "content": " the direction of the corresponding axis. Moreover, because we want Gaussians to describe the true surface of the scene, we need to avoid semi-transparent Gaussians. Therefore, we want Gaussians to be either opaque or fully transparent, in which case we can drop them for rendering. Consequently, we want to have " + }, + { + "bbox": [ + 304, + 610, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\alpha_g = 1" + }, + { + "bbox": [ + 304, + 610, + 545, + 693 + ], + "type": "text", + "content": " for any Gaussian " + }, + { + "bbox": [ + 304, + 610, + 545, + 693 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 610, + 545, + 693 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "In such scenario, the density of the Gaussians could fi-" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "text", + "content": "5357" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 129, + 209 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 129, + 209 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 129, + 209 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 129, + 209 + ], + "type": "image", + "image_path": "a56292c9d827b58edba578199500579430f2a6afc2342e78c183f2cabb75b7e6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 211, + 125, + 220 + ], + "lines": [ + { + "bbox": [ + 52, + 211, + 125, + 220 + ], + "spans": [ + { + "bbox": [ + 52, + 211, + 125, + 220 + ], + "type": "text", + "content": "(a) Mesh & Gaussians" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 130, + 70, + 211, + 209 + ], + "blocks": [ + { + "bbox": [ + 130, + 70, + 211, + 209 + ], + "lines": [ + { + "bbox": [ + 130, + 70, + 211, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 70, + 211, + 209 + ], + "type": "image", + "image_path": "0fcc44c7ccb311ac823d6d50fb0fc3578e6568e81d27799a90f6d5a308775a37.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 212, + 208, + 220 + ], + "lines": [ + { + "bbox": [ + 134, + 212, + 208, + 220 + ], + "spans": [ + { + "bbox": [ + 134, + 212, + 208, + 220 + ], + "type": "text", + "content": "(b) Mesh (No Texture)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 212, + 70, + 293, + 209 + ], + "blocks": [ + { + "bbox": [ + 212, + 70, + 293, + 209 + ], + "lines": [ + { + "bbox": [ + 212, + 70, + 293, + 209 + ], + "spans": [ + { + "bbox": [ + 212, + 70, + 293, + 209 + ], + "type": "image", + "image_path": "d0edac57d5ee016f3498c03b58541f242fc507dbc35d2b68fac779908d8555be.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 212, + 281, + 220 + ], + "lines": [ + { + "bbox": [ + 223, + 212, + 281, + 220 + ], + "spans": [ + { + "bbox": [ + 223, + 212, + 281, + 220 + ], + "type": "text", + "content": "(c) Mesh normals" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 301, + 70, + 381, + 209 + ], + "blocks": [ + { + "bbox": [ + 301, + 70, + 381, + 209 + ], + "lines": [ + { + "bbox": [ + 301, + 70, + 381, + 209 + ], + "spans": [ + { + "bbox": [ + 301, + 70, + 381, + 209 + ], + "type": "image", + "image_path": "c917145fa695713ee2573a050eaa62a4f8dc031b2f14f7df00ab8096b2a22b0e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 212, + 377, + 220 + ], + "lines": [ + { + "bbox": [ + 306, + 212, + 377, + 220 + ], + "spans": [ + { + "bbox": [ + 306, + 212, + 377, + 220 + ], + "type": "text", + "content": "(a) Mesh & Gaussians" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 231, + 539, + 243 + ], + "lines": [ + { + "bbox": [ + 51, + 231, + 539, + 243 + ], + "spans": [ + { + "bbox": [ + 51, + 231, + 539, + 243 + ], + "type": "text", + "content": "Figure 4. Examples of (a) renderings and (b) reconstructed meshes with SuGaR. The (c) normal maps help visualize the geometry." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 383, + 70, + 463, + 209 + ], + "blocks": [ + { + "bbox": [ + 383, + 70, + 463, + 209 + ], + "lines": [ + { + "bbox": [ + 383, + 70, + 463, + 209 + ], + "spans": [ + { + "bbox": [ + 383, + 70, + 463, + 209 + ], + "type": "image", + "image_path": "3cb31b2b075cc43fa4311d547582d177406bfe27cb552cf805412325a9c0c952.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 211, + 460, + 220 + ], + "lines": [ + { + "bbox": [ + 387, + 211, + 460, + 220 + ], + "spans": [ + { + "bbox": [ + 387, + 211, + 460, + 220 + ], + "type": "text", + "content": "(b) Mesh (No Texture)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 465, + 70, + 545, + 209 + ], + "blocks": [ + { + "bbox": [ + 465, + 70, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 465, + 70, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 465, + 70, + 545, + 209 + ], + "type": "image", + "image_path": "4a714642ab921d433fa94db5a38240ef13762090f89437b638b27ee572ab763f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 212, + 533, + 220 + ], + "lines": [ + { + "bbox": [ + 476, + 212, + 533, + 220 + ], + "spans": [ + { + "bbox": [ + 476, + 212, + 533, + 220 + ], + "type": "text", + "content": "(c) Mesh normals" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 262, + 228, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 228, + 276 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 228, + 276 + ], + "type": "text", + "content": "nally be approximated by density " + }, + { + "bbox": [ + 47, + 262, + 228, + 276 + ], + "type": "inline_equation", + "content": "\\bar{d} (p)" + }, + { + "bbox": [ + 47, + 262, + 228, + 276 + ], + "type": "text", + "content": " with:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 84, + 283, + 287, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 283, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 84, + 283, + 287, + 312 + ], + "type": "interline_equation", + "content": "\\bar {d} (p) = \\exp \\left(- \\frac {1}{2 s _ {g ^ {*}} ^ {2}} \\langle p - \\mu_ {g ^ {*}}, n _ {g ^ {*}} \\rangle^ {2}\\right). \\tag {5}", + "image_path": "0ddb52b34aa8e84119ddecac78c01d788d8dd3a6032731cfabc011a0c1458a65.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 327, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 480 + ], + "type": "text", + "content": "A first strategy to enforce our regularization is to add term " + }, + { + "bbox": [ + 46, + 327, + 287, + 480 + ], + "type": "inline_equation", + "content": "|d(p) - \\bar{d}(p)|" + }, + { + "bbox": [ + 46, + 327, + 287, + 480 + ], + "type": "text", + "content": " to the optimization loss. While this approach works well to align Gaussians with the surface, we noticed that computing a slightly different loss relying on an SDF rather than on density further increases the alignment of Gaussians with the surface of the scene. For a given flat Gaussian, i.e., " + }, + { + "bbox": [ + 46, + 327, + 287, + 480 + ], + "type": "inline_equation", + "content": "s_g = 0" + }, + { + "bbox": [ + 46, + 327, + 287, + 480 + ], + "type": "text", + "content": ", considering level sets is meaningless since all level sets would degenerate toward the plane passing through the center of the Gaussian " + }, + { + "bbox": [ + 46, + 327, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mu_g" + }, + { + "bbox": [ + 46, + 327, + 287, + 480 + ], + "type": "text", + "content": " with normal " + }, + { + "bbox": [ + 46, + 327, + 287, + 480 + ], + "type": "inline_equation", + "content": "n_g" + }, + { + "bbox": [ + 46, + 327, + 287, + 480 + ], + "type": "text", + "content": ". The distance between point " + }, + { + "bbox": [ + 46, + 327, + 287, + 480 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 327, + 287, + 480 + ], + "type": "text", + "content": " and the true surface of the scene would be approximately " + }, + { + "bbox": [ + 46, + 327, + 287, + 480 + ], + "type": "inline_equation", + "content": "|\\langle p - \\mu_{g'}, n_{g'} \\rangle|" + }, + { + "bbox": [ + 46, + 327, + 287, + 480 + ], + "type": "text", + "content": ", the distance from " + }, + { + "bbox": [ + 46, + 327, + 287, + 480 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 327, + 287, + 480 + ], + "type": "text", + "content": " to this plane. Consequently, the zero-crossings of the Signed Distance Function" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 490, + 287, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 490, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 287, + 510 + ], + "type": "interline_equation", + "content": "\\bar {f} (p) = \\pm s _ {g *} \\sqrt {- 2 \\log (\\bar {d} (p))} \\tag {6}", + "image_path": "8220a80fadc899291f1b61070cb17817ce369f9370b6f7bd74119accced8eca2.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 517, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 539 + ], + "type": "text", + "content": "corresponds to the surface of the scene. More generally, we define" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 540, + 287, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 540, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 107, + 540, + 287, + 555 + ], + "type": "interline_equation", + "content": "f (p) = \\pm s _ {g *} \\sqrt {- 2 \\log (d (p))} \\tag {7}", + "image_path": "4ab6123b84a4b90ab1d8181282fb0d1624ec0e95780338338650ba2112ed8a46.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 559, + 287, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 287, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 287, + 606 + ], + "type": "text", + "content": "as the \"ideal\" distance function associated with the density function " + }, + { + "bbox": [ + 46, + 559, + 287, + 606 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 559, + 287, + 606 + ], + "type": "text", + "content": ". This distance function corresponds to the true surface of the scene in an ideal scenario where " + }, + { + "bbox": [ + 46, + 559, + 287, + 606 + ], + "type": "inline_equation", + "content": "d = \\bar{d}" + }, + { + "bbox": [ + 46, + 559, + 287, + 606 + ], + "type": "text", + "content": ". We therefore take our regularization term " + }, + { + "bbox": [ + 46, + 559, + 287, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 559, + 287, + 606 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 99, + 614, + 287, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 614, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 99, + 614, + 287, + 644 + ], + "type": "interline_equation", + "content": "\\mathcal {R} = \\frac {1}{| \\mathcal {P} |} \\sum_ {p \\in \\mathcal {P}} \\left| | \\hat {f} (p) | - | f (p) | \\right|, \\tag {8}", + "image_path": "10854395c07130cdea8f398b75469c2c5af1a3808dec6dbff992c67dab9a24a6.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "text", + "content": "by sampling 3D points " + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "text", + "content": " and summing the differences at these points between the ideal SDF " + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "inline_equation", + "content": "f(p)" + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "text", + "content": " and an estimate " + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{f}(p)" + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "text", + "content": " of the SDF of the surface created by the current Gaussians. " + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "text", + "content": " refers to the set of sampled points. During optimization, we simply want the zero-level sets of both the" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 360, + 277, + 489, + 377 + ], + "blocks": [ + { + "bbox": [ + 360, + 277, + 489, + 377 + ], + "lines": [ + { + "bbox": [ + 360, + 277, + 489, + 377 + ], + "spans": [ + { + "bbox": [ + 360, + 277, + 489, + 377 + ], + "type": "image", + "image_path": "0fda687334d3dc4b93586b3ae1e96bd344c0ae51d878b5b833c5a193d5922480.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 387, + 545, + 453 + ], + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 453 + ], + "type": "text", + "content": "Figure 5. Efficiently estimating " + }, + { + "bbox": [ + 304, + 387, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\hat{f}(p)" + }, + { + "bbox": [ + 304, + 387, + 545, + 453 + ], + "type": "text", + "content": " of the SDF of the surface generated from Gaussians. We render depth maps of the Gaussians, sample points " + }, + { + "bbox": [ + 304, + 387, + 545, + 453 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 387, + 545, + 453 + ], + "type": "text", + "content": " in the viewpoint according to the distribution of the Gaussians. Value " + }, + { + "bbox": [ + 304, + 387, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\hat{f}(p)" + }, + { + "bbox": [ + 304, + 387, + 545, + 453 + ], + "type": "text", + "content": " is taken as the 3D distance between " + }, + { + "bbox": [ + 304, + 387, + 545, + 453 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 387, + 545, + 453 + ], + "type": "text", + "content": " and the intersection between the line of sight for " + }, + { + "bbox": [ + 304, + 387, + 545, + 453 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 387, + 545, + 453 + ], + "type": "text", + "content": " and the depth map." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 478, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 478, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 545, + 525 + ], + "type": "text", + "content": "SDF estimator " + }, + { + "bbox": [ + 304, + 478, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\hat{f}" + }, + { + "bbox": [ + 304, + 478, + 545, + 525 + ], + "type": "text", + "content": " and the ideal SDF " + }, + { + "bbox": [ + 304, + 478, + 545, + 525 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 478, + 545, + 525 + ], + "type": "text", + "content": " to align. Since a zero-level set is entirely determined by an unsigned distance function, we actually do not need to compute the sign of the SDFs and use absolute values." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 529, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 635 + ], + "type": "text", + "content": "Computing efficiently " + }, + { + "bbox": [ + 304, + 529, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\hat{f}(p)" + }, + { + "bbox": [ + 304, + 529, + 545, + 635 + ], + "type": "text", + "content": " is a priori challenging. To do so, we propose to use the depth maps of the Gaussians from the viewpoints used for training—these depth maps can be rendered efficiently by extending the splatting rasterizer. Then, as shown in Figure 5, for a point " + }, + { + "bbox": [ + 304, + 529, + 545, + 635 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 529, + 545, + 635 + ], + "type": "text", + "content": " visible from a training viewpoint, " + }, + { + "bbox": [ + 304, + 529, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\hat{f}(p)" + }, + { + "bbox": [ + 304, + 529, + 545, + 635 + ], + "type": "text", + "content": " is the difference between the depth of " + }, + { + "bbox": [ + 304, + 529, + 545, + 635 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 529, + 545, + 635 + ], + "type": "text", + "content": " and the depth in the corresponding depth map at the projection of " + }, + { + "bbox": [ + 304, + 529, + 545, + 635 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 529, + 545, + 635 + ], + "type": "text", + "content": ". Moreover, we sample points " + }, + { + "bbox": [ + 304, + 529, + 545, + 635 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 529, + 545, + 635 + ], + "type": "text", + "content": " following the distribution of the Gaussians:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 378, + 643, + 545, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 643, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 378, + 643, + 545, + 669 + ], + "type": "interline_equation", + "content": "p \\sim \\prod_ {g} \\mathcal {N} \\left(.; \\mu_ {g}, \\Sigma_ {g}\\right), \\tag {9}", + "image_path": "05340f15dd4881f1b10cfef30dc2a4672a204fbbe760d12365fca68aa493b5b6.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(:,\\mu_g,\\Sigma_g)" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " the Gaussian distribution of mean " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mu_{g}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and covariance " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Sigma_{g}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " as these points are likely to correspond to a high gradient for " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "text", + "content": "5358" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 74, + 154, + 178 + ], + "blocks": [ + { + "bbox": [ + 50, + 74, + 154, + 178 + ], + "lines": [ + { + "bbox": [ + 50, + 74, + 154, + 178 + ], + "spans": [ + { + "bbox": [ + 50, + 74, + 154, + 178 + ], + "type": "image", + "image_path": "7793310e8a1663a1bc425bb1121acfd80c404fd4cb0d52af2380bd5663ebf64a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 185, + 287, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 185, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 287, + 262 + ], + "type": "text", + "content": "Figure 6. Sampling points on a level set for Poisson reconstruction. Left: We sample points on the depth maps of the Gaussians and refine the point locations to move the points on the level set. Right: Comparison between the extracted mesh without (left) and with (right) our refinement step. Since splatted depth maps are not exact, using directly the depth points for reconstruction usually results in a large amount of noise and missing details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 73, + 227, + 176 + ], + "blocks": [ + { + "bbox": [ + 167, + 73, + 227, + 176 + ], + "lines": [ + { + "bbox": [ + 167, + 73, + 227, + 176 + ], + "spans": [ + { + "bbox": [ + 167, + 73, + 227, + 176 + ], + "type": "image", + "image_path": "d262f28448b4d3af275cf2b66155ba011da0fa7967794a9bd0b45cb786fba190.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 227, + 73, + 286, + 176 + ], + "blocks": [ + { + "bbox": [ + 227, + 73, + 286, + 176 + ], + "lines": [ + { + "bbox": [ + 227, + 73, + 286, + 176 + ], + "spans": [ + { + "bbox": [ + 227, + 73, + 286, + 176 + ], + "type": "image", + "image_path": "2f706608e69c018da35d9debfa62abf0fe7d326101fcafab0e862679a53f0601.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 285, + 287, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 287, + 309 + ], + "type": "text", + "content": "We also add a regularization term to encourage the normals of SDF " + }, + { + "bbox": [ + 47, + 285, + 287, + 309 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 285, + 287, + 309 + ], + "type": "text", + "content": " and the normals of SDF " + }, + { + "bbox": [ + 47, + 285, + 287, + 309 + ], + "type": "inline_equation", + "content": "\\bar{f}" + }, + { + "bbox": [ + 47, + 285, + 287, + 309 + ], + "type": "text", + "content": " to be similar:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 317, + 287, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 317, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 82, + 317, + 287, + 350 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {\\text {N o r m}} = \\frac {1}{| \\mathcal {P} |} \\sum_ {p \\in \\mathcal {P}} \\left\\| \\frac {\\nabla f (p)}{\\| \\nabla f (p) \\| _ {2}} - n _ {g ^ {*}} \\right\\| _ {2} ^ {2}. \\tag {10}", + "image_path": "1abab89c79f87f6008b58c124418a641c4f89b231b4c5739bb56853d72ddf510.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 360, + 190, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 190, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 190, + 371 + ], + "type": "text", + "content": "4.2. Efficient Mesh Extraction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 380, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 474 + ], + "type": "text", + "content": "To create a mesh from the Gaussians obtained after optimization using our regularization terms in Eq. (8) and Eq. (10), we sample 3D points on a level set of the density computed from the Gaussians. The level set depends on a level parameter " + }, + { + "bbox": [ + 46, + 380, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 380, + 287, + 474 + ], + "type": "text", + "content": ". Then, we obtain a mesh by simply running a Poisson reconstruction [14] on these points. Note that we can also easily assign the points with the normals of the SDF, which improves the mesh quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "text", + "content": "The challenge is in efficiently identifying points lying on the level set. For this, as shown in Figure 6, we again rely on the depth maps of the Gaussians as seen from the training viewpoints. We first randomly sample pixels from each depth map. For each pixel " + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "text", + "content": ", we sample its line of sight to find a 3D point on the level set. Formally, we sample " + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "text", + "content": " points " + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "inline_equation", + "content": "p + t_i v" + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "text", + "content": " is the 3D point in the depth map that reprojects on pixel " + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "text", + "content": " is the direction of the line of sight, and " + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "inline_equation", + "content": "t_i \\in [-3\\sigma_g(v), 3\\sigma_g(v)]" + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\sigma_g(v)" + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "text", + "content": " is the standard deviation of the 3D Gaussian " + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "text", + "content": " in the direction of the camera. The interval " + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "inline_equation", + "content": "[-3\\sigma_g(v), 3\\sigma_g(v)]" + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "text", + "content": " is the confidence interval for the 99.7 confidence level of the 1D Gaussian function of " + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 475, + 287, + 629 + ], + "type": "text", + "content": " along the ray." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 631, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 631, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 631, + 287, + 713 + ], + "type": "text", + "content": "Then, we compute the density values " + }, + { + "bbox": [ + 46, + 631, + 287, + 713 + ], + "type": "inline_equation", + "content": "d_{i} = d(p + t_{i}v)" + }, + { + "bbox": [ + 46, + 631, + 287, + 713 + ], + "type": "text", + "content": " from Eq. (1) of these sampled points. If there exist " + }, + { + "bbox": [ + 46, + 631, + 287, + 713 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 46, + 631, + 287, + 713 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 631, + 287, + 713 + ], + "type": "inline_equation", + "content": "d_{i} < \\lambda < d_{j}" + }, + { + "bbox": [ + 46, + 631, + 287, + 713 + ], + "type": "text", + "content": ", then there is a level set point located in this range. If so, we use linear interpolation to compute the coefficient " + }, + { + "bbox": [ + 46, + 631, + 287, + 713 + ], + "type": "inline_equation", + "content": "t^{*}" + }, + { + "bbox": [ + 46, + 631, + 287, + 713 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 631, + 287, + 713 + ], + "type": "inline_equation", + "content": "p + t^{*}v" + }, + { + "bbox": [ + 46, + 631, + 287, + 713 + ], + "type": "text", + "content": " is the level set point closest to the camera, verifying " + }, + { + "bbox": [ + 46, + 631, + 287, + 713 + ], + "type": "inline_equation", + "content": "d(p + t^{*}v) = \\lambda" + }, + { + "bbox": [ + 46, + 631, + 287, + 713 + ], + "type": "text", + "content": ". We also compute the normals of the surface at points " + }, + { + "bbox": [ + 46, + 631, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{p}" + }, + { + "bbox": [ + 46, + 631, + 287, + 713 + ], + "type": "text", + "content": ", which we naturally" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 331, + 70, + 535, + 191 + ], + "blocks": [ + { + "bbox": [ + 331, + 70, + 535, + 191 + ], + "lines": [ + { + "bbox": [ + 331, + 70, + 535, + 191 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 535, + 191 + ], + "type": "image", + "image_path": "5e1246092c67462f639fac33a618dd8fa91f5355a55569554ece5a95bc25dd3d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 198, + 545, + 254 + ], + "lines": [ + { + "bbox": [ + 305, + 198, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 545, + 254 + ], + "type": "text", + "content": "Figure 7. Joint refinement of mesh and Gaussians. Left: We bind Gaussians to the triangles of the mesh. Depending on the number of triangles in the scene, we bind a different number of Gaussians per triangle, with predefined barycentric coordinates. Right: Mesh before and after joint refinement." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 286, + 544, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 286, + 544, + 315 + ], + "spans": [ + { + "bbox": [ + 306, + 286, + 544, + 315 + ], + "type": "text", + "content": "define as the normalized analytical gradient of the density " + }, + { + "bbox": [ + 306, + 286, + 544, + 315 + ], + "type": "inline_equation", + "content": "\\frac{\\nabla d(\\hat{p})}{\\|\\nabla d(\\hat{p})\\|_2}" + }, + { + "bbox": [ + 306, + 286, + 544, + 315 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 318, + 545, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 318, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 306, + 318, + 545, + 341 + ], + "type": "text", + "content": "Finally, we apply Poisson reconstruction to reconstruct a surface mesh from the level set points and their normals." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 365, + 512, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 512, + 378 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 512, + 378 + ], + "type": "text", + "content": "4.3. Binding New 3D Gaussians to the Mesh" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 389, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 389, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 389, + 545, + 460 + ], + "type": "text", + "content": "Once we have extracted a first mesh, we can refine this mesh by binding new Gaussians to the mesh triangles and optimize the Gaussians and the mesh jointly using the Gaussian Splatting rasterizer. This enables the edition of the Gaussian splatting scene with popular mesh editing tools while keeping high-quality rendering thanks to the Gaussians." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 465, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 465, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 465, + 545, + 536 + ], + "type": "text", + "content": "Given the initial mesh, we instantiate new 3D Gaussians on the mesh. More exactly, we associate a set of " + }, + { + "bbox": [ + 305, + 465, + 545, + 536 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 465, + 545, + 536 + ], + "type": "text", + "content": " thin 3D Gaussians to each triangle of the mesh, sampled on the surface of the triangle, as illustrated in Figure 7. To do so, we slightly modify the structure of the original 3D Gaussian Splitting model." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 542, + 545, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 672 + ], + "type": "text", + "content": "We explicitly compute the means of the Gaussians from the mesh vertices using predefined barycentric coordinates in the corresponding triangles during optimization. Also, the Gaussians have only 2 learnable scaling factors instead of 3 and only 1 learnable 2D rotation encoded with a complex number rather than a quaternion, to keep the Gaussians flat and aligned with the mesh triangles. More details about this parameterisation are given in the supplementary material. Like the original model, we also optimize an opacity value and a set of spherical harmonics for every Gaussian to encode the color emitted in all directions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Figure 7 shows an example of a mesh before and after refinement. Figure 1 and the supplementary material give examples of what can be done by editing the mesh." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "text", + "content": "5359" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 128, + 85 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 178, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 178, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 178, + 103 + ], + "type": "text", + "content": "5.1. Implementation details" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 109, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 109, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 109, + 287, + 133 + ], + "type": "text", + "content": "All our models are optimized on a single GPU Nvidia Tesla V100 SXM2 32 Go." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 148, + 286, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 148, + 286, + 219 + ], + "spans": [ + { + "bbox": [ + 47, + 148, + 286, + 219 + ], + "type": "text", + "content": "Regularization. For all scenes, we start by optimizing a Gaussian Splatting with no regularization for 7,000 iterations to let the 3D Gaussians position themselves without any additional constraint. Then, we perform 2,000 iterations with an additional entropy loss on the opacities " + }, + { + "bbox": [ + 47, + 148, + 286, + 219 + ], + "type": "inline_equation", + "content": "\\alpha_{g}" + }, + { + "bbox": [ + 47, + 148, + 286, + 219 + ], + "type": "text", + "content": " of the Gaussians, as a way to enforce them to become binary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 220, + 286, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 286, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 286, + 314 + ], + "type": "text", + "content": "Finally, we remove Gaussians with opacity values under 0.5 and perform 6,000 iterations with the regularization term introduced in Subsection 4.1, which makes a total of 15,000 iterations. To compute the density values of points from a Gaussian " + }, + { + "bbox": [ + 46, + 220, + 286, + 314 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 220, + 286, + 314 + ], + "type": "text", + "content": ", we sum only the Gaussian functions from the 16 nearest Gaussians of " + }, + { + "bbox": [ + 46, + 220, + 286, + 314 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 220, + 286, + 314 + ], + "type": "text", + "content": " and update the list of nearest neighbors every 500 iterations. Optimization takes between 15 and 45 minutes depending on the scene." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "type": "text", + "content": "Mesh extraction. For all experiments except the ablation presented in Table 2, we extract the " + }, + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "type": "text", + "content": "-level set of the density function for " + }, + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "type": "inline_equation", + "content": "\\lambda = 0.3" + }, + { + "bbox": [ + 46, + 328, + 286, + 411 + ], + "type": "text", + "content": ". We perform Poisson reconstruction with depth 10 and apply mesh simplification using quadric error metrics [9] to decrease the resolution of the meshes. Mesh extraction generally takes between 5 and 10 minutes depending on the scene." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 426, + 286, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 286, + 474 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 286, + 474 + ], + "type": "text", + "content": "Joint refinement. We jointly refine the mesh and the bound 3D Gaussians for either 2,000, 7,000 or 15,000 iterations. Depending on the number of iterations, the duration of refinement goes from a few minutes to an hour." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 481, + 241, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 481, + 241, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 481, + 241, + 495 + ], + "type": "text", + "content": "5.2. Real-Time Rendering of Real Scenes" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 500, + 286, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 500, + 286, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 500, + 286, + 582 + ], + "type": "text", + "content": "For evaluating our model, we follow the approach from the original 3D Gaussian Splatting paper [15] and compare the performance of several variations of our method SuGaR after refinement on real 3D scenes from 3 different datasets: Mip-NeRF360 [1], DeepBlending [12] and Tanks&Temples [16]. We call R-SuGaR-NK a refined SuGaR model optimized for " + }, + { + "bbox": [ + 47, + 500, + 286, + 582 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 500, + 286, + 582 + ], + "type": "text", + "content": " iterations during refinement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 583, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 583, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 583, + 286, + 664 + ], + "type": "text", + "content": "Following [15], we select the same sets of 2 scenes from Tanks&Temples (Truck and Train) and 2 scenes from DeepBlending (Playroom and Dr. Johnson). However, due to licensing issues and the unavailability of the scenes Flowers and Treehill, we perform the evaluation of all methods only on 7 scenes from Mip-NeRF360 instead of the full set of 9 scenes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 666, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 286, + 713 + ], + "type": "text", + "content": "We compute the standard metrics PSNR, SSIM and LPIPS [44] to evaluate the quality of SuGaR's rendering using our extracted meshes and their bound surface Gaussians. Note that [6, 26, 39] also do not use plain textured" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": "mesh rendering. We compare to several baselines, some of them focusing only on Novel View Synthesis [2, 15, 23, 41] and others relying on a reconstructed mesh [6, 26, 39], just like SuGaR. Results on the Mip-NeRF360 dataset are given in Table 1. Results on Tanks&Temple and DeepBlending are similar and can be found in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 144, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 239 + ], + "type": "text", + "content": "Even though SuGaR focuses on aligning 3D Gaussians for reconstructing a high quality mesh during the first stage of its optimization, it significantly outperforms the state of the art methods for Novel View Synthesis using a mesh and reaches better performance than several famous models that focus only on rendering, such as I-NGP [23] and Plenoxels [41]. This performance is remarkable as SuGaR is able to extract a mesh significantly faster than other methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 239, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 545, + 370 + ], + "type": "text", + "content": "Moreover, SuGaR even reaches performance similar to state-of-the-art models for rendering quality [2, 15] on some of the scenes used for evaluation. Two main reasons explain this performance. First, the mesh extracted after the first stage of optimization serves as an excellent initialization for positioning Gaussians when starting the refinement phase. Then, the Gaussians constrained to remain on the surface during refinement greatly increase the rendering quality as they play the role of an efficient texturing tool and help reconstructing very fine details missing in the extracted mesh. Additional qualitative results are available in Figure 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 380, + 406, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 406, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 406, + 392 + ], + "type": "text", + "content": "5.3. Mesh Extraction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 399, + 545, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 565 + ], + "type": "text", + "content": "To demonstrate the ability of our mesh extraction method for reconstructing high-quality meshes that are well-suited for view synthesis, we compare different mesh extraction algorithms. In particular, we optimize several variations of SuGaR by following the exact same pipeline as our standard model, except for the mesh extraction process: We either extract the mesh using a very fine marching cubes algorithm [21], by applying Poisson reconstruction [14] using the centers of the 3D Gaussians as the surface point cloud, or by applying our mesh extraction method on different level sets. Quantitative results are available in Table 2 and show the clear superiority of our approach for meshing 3D Gaussians. Figure 3 also illustrates how the marching cubes algorithm fails in this context." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 575, + 450, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 450, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 450, + 588 + ], + "type": "text", + "content": "5.4. Mesh Rendering Ablation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "text", + "content": "Table 3 provides additional results to quantify how various parameters impact rendering performance. In particular, we evaluate how the resolution of the mesh extraction, i.e., the number of triangles, modifies the rendering quality. For fair comparison, we increase the number of surface-aligned Gaussians per triangle when we decrease the number of triangles. Results show that increasing the number of vertices increases the quality of rendering with surface Gaussians, but meshes with less triangles are already able to reach state of the art results." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "text", + "content": "5360" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 525, + 278 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 525, + 278 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 525, + 278 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 525, + 278 + ], + "type": "table", + "html": "
Indoor scenesOutdoor scenesAverage on all scenes
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
No mesh (except SuGaR)
Plenoxels [42]24.830.7660.42622.020.5420.46523.620.6700.443
INGP-Base [23]28.650.8400.28123.470.5710.41626.430.7250.339
INGP-Big [23]29.140.8630.24223.570.6020.37526.750.7510.299
Mip-NeRF360 [2]31.580.9140.18225.790.7460.24729.090.8420.210
3DGS [15]30.410.9200.18926.400.8050.17328.690.8700.182
R-SuGaR-15K (Ours)29.430.9100.21624.400.6990.30127.270.8200.253
With mesh
Mobile-NeRF [6]---21.950.4700.470---
NeRFMeshing [26]23.83--22.23--23.15--
BakedSDF [39]27.060.8360.258------
R-SuGaR-2K (Ours)26.290.8720.26222.970.6480.36024.870.7760.304
R-SuGaR-7K (Ours)28.730.9040.22624.160.6910.31326.770.8130.263
R-SuGaR-15K (Ours)29.430.9100.21624.400.6990.30127.270.8200.253
", + "image_path": "671fe6d63757a6e2852b35c099121a5965e771f4c7633e1e4ba7462c5e273efb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 328, + 276, + 407 + ], + "blocks": [ + { + "bbox": [ + 46, + 287, + 546, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 287, + 546, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 546, + 309 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation of rendering quality on the Mip-NeRF360 dataset [2]. SuGaR is best among the methods that recover a mesh, and still performs well compared to NeRF methods and vanilla 3D Gaussian Splatting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 328, + 276, + 407 + ], + "lines": [ + { + "bbox": [ + 56, + 328, + 276, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 328, + 276, + 407 + ], + "type": "table", + "html": "
Extraction methodPSNR ↑SSIM ↑LPIPS ↓
Marching Cubes [21]23.910.7030.392
Poisson (centers) [14]23.760.7560.340
Ours (Surface level 0.1)24.620.7650.313
Ours (Surface level 0.3)24.870.7760.304
Ours (Surface level 0.5)24.910.7770.304
", + "image_path": "64196f0b1efc41c3b79491cb9c860df5e66b20cb9461243b7e45ba98b9b2501a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 59, + 495, + 272, + 568 + ], + "blocks": [ + { + "bbox": [ + 46, + 415, + 287, + 482 + ], + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 482 + ], + "type": "text", + "content": "Table 2. Ablation for different mesh extraction methods on the Mip-NeRF360 dataset [2] after applying our regularization term. For 'Poisson (centers)', we apply Poisson reconstruction [14] using as surface points the centers of the 3D Gaussians. For fair comparison, we calibrate the methods to enforce all extracted meshes to have approximately 1,000,000 vertices." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 495, + 272, + 568 + ], + "lines": [ + { + "bbox": [ + 59, + 495, + 272, + 568 + ], + "spans": [ + { + "bbox": [ + 59, + 495, + 272, + 568 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPS ↓
1M vertices (3DGS)24.510.7680.295
1M vertices (UV)21.240.6090.478
200K vertices (3DGS)24.240.7570.300
200K vertices (UV)21.440.6560.419
", + "image_path": "1d39f4556690dd0181417cf11f5b127f9c425cf3c7ba4eda4b8cdf0933ee8866.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 576, + 287, + 643 + ], + "lines": [ + { + "bbox": [ + 46, + 576, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 287, + 643 + ], + "type": "text", + "content": "Table 3. Comparison between surface-aligned 3D Gaussians and an optimized traditional UV texture on the Mip-NeRF360 dataset [2]. For fair comparison, we only use the diffuse spherical harmonics component when rendering images with SuGaR. Using 3D Gaussians bound to the mesh greatly improves rendering quality, even though it contains less parameters than the UV texture." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "content": "Then, we illustrate the benefits of using Gaussians aligned on the surface as a texturing tool for rendering meshes. To this end, we also optimize traditional UV textures on our meshes using differentiable mesh rendering" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 330, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 330, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 305, + 330, + 545, + 403 + ], + "type": "text", + "content": "with traditional triangle rasterization. Even though rendering with surface-aligned Gaussians provides better performance, rendering our meshes with traditional UV textures still produces satisfying results, which further illustrates the quality of our extracted meshes. Qualitative comparisons are provided in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 421, + 379, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 421, + 379, + 433 + ], + "spans": [ + { + "bbox": [ + 306, + 421, + 379, + 433 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 443, + 545, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 545, + 600 + ], + "type": "text", + "content": "We proposed a very fast algorithm to obtain an accurate 3D triangle mesh for a scene via Gaussian Splatting. Moreover, by combining meshing and Gaussian Splatting, we make possible intuitive manipulation of the captured scenes and realistic rendering, offering new possibilities for creators. SuGaR does not come without limitations: Gaussians do tend to \"cheat\" on the geometry and depth by creating cavities to reproduce specular effects, instead of relying on spherical harmonics. SuGaR's regularization mitigates this issue, but Gaussians can still distort the surface in regions with intense specularity. In addition, SuGaR's assumption that the scene can be represented as a surface complicates the rendering of volumetric effects or fuzzy materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 619, + 410, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 619, + 410, + 632 + ], + "spans": [ + { + "bbox": [ + 306, + 619, + 410, + 632 + ], + "type": "text", + "content": "Acknowledgements." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "This project was funded in part by the European Union (ERC Advanced Grant explorer Funding ID #101097259). This work was granted access to the HPC resources of IDRIS under the allocation 2023-AD011013387R1 made by GENCI. We thank George Drettakis and Elliot Vincent for inspiring discussions and valuable feedback." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 747, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 314, + 757 + ], + "type": "text", + "content": "5361" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Jonathan T. Barron. Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. In International Conference on Computer Vision, 2021. 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 287, + 158 + ], + "type": "text", + "content": "[2] Jonathan T. Barron. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In Conference on Computer Vision and Pattern Recognition, 2022. 3, 7, 8, 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 161, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 161, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 161, + 287, + 205 + ], + "type": "text", + "content": "[3] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. NeRD: Neural Reflectance Decomposition from Image Collections. In International Conference on Computer Vision, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 207, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 207, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 207, + 286, + 239 + ], + "type": "text", + "content": "[4] Chris Buehler, Michael Bosse, Leonard Mcmillan, Steven Gortler, and Michael Cohen. Unstructured Lumigraph Rendering. In ACM SIGGRAPH, 2001. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 243, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 243, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 286, + 274 + ], + "type": "text", + "content": "[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial Radiance Fields. In European Conference on Computer Vision, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 277, + 286, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 277, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 53, + 277, + 286, + 331 + ], + "type": "text", + "content": "[6] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. MobileNeRF: Exploiting the Polygon Rasterization Pipeline for Efficient Neural Field Rendering on Mobile Architectures. In Conference on Computer Vision and Pattern Recognition, 2023. 2, 3, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 286, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 286, + 388 + ], + "type": "text", + "content": "[7] Chong Bao and Bangbang Yang, Zeng Junyi, Bao Hujun, Zhang Yinda, Cui Zhaopeng, and Zhang Guofeng. NeuMesh: Learning Disentangled Neural Mesh-Based Implicit Field for Geometry and Texture Editing. In European Conference on Computer Vision, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 392, + 286, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 286, + 434 + ], + "type": "text", + "content": "[8] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving Neural Implicit Surfaces Geometry with Patch Warping. In Conference on Computer Vision and Pattern Recognition, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 437, + 286, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 437, + 286, + 469 + ], + "spans": [ + { + "bbox": [ + 53, + 437, + 286, + 469 + ], + "type": "text", + "content": "[9] Michael Garland and Paul S. Heckbert. Surface Simplification Using Quadric Error Metrics. In ACM SIGGRAPH, 1997. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 472, + 287, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 472, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 287, + 515 + ], + "type": "text", + "content": "[10] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven Seitz. Multi-View Stereo for Community Photo Collections. In International Conference on Computer Vision, 2007. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 518, + 287, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 518, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 48, + 518, + 287, + 551 + ], + "type": "text", + "content": "[11] Peter Hedman and Pratul P. Srinivasan. Baking Neural Radiance Fields for Real-Time View Synthesis. In International Conference on Computer Vision, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 553, + 287, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 553, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 48, + 553, + 287, + 596 + ], + "type": "text", + "content": "[12] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep Blending for Free-Viewpoint Image-Based Rendering. In ACM SIGGRAPH, 2018. 3, 7, 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 599, + 286, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 599, + 286, + 631 + ], + "spans": [ + { + "bbox": [ + 48, + 599, + 286, + 631 + ], + "type": "text", + "content": "[13] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Mitra. ReLU Fields: The Little Non-Linearity That Could. In ACM SIGGRAPH, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 286, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 286, + 666 + ], + "type": "text", + "content": "[14] Michael M. Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson Surface Reconstruction. In Eurographics, 2006. 2, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. In ACM SIGGRAPH, 2023. 1, 2, 4, 7, 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking Large-Scale Scene Reconstruction. In ACM SIGGRAPH, 2017. 7, 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[17] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-Based Neural Rendering with Per-View Optimization. In Computer Graphics Forum, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[18] Zhengfei Kuang, Kyle Olszewski, Menglei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. NeROIC: Neural Rendering of Objects from Online Image Collections. In ACM SIGGRAPH, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "type": "text", + "content": "[19] Marc Levoy and Pat Hanrahan. Light Field Rendering. In ACM SIGGRAPH, 1996. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 262 + ], + "type": "text", + "content": "[20] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-Fidelity Neural Surface Reconstruction. In Conference on Computer Vision and Pattern Recognition, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 266, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 266, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 545, + 297 + ], + "type": "text", + "content": "[21] William E. Lorensen and Harvey E. Cline. Marching Cubes: A High Resolution 3D Surface Construction Algorithm. In ACM SIGGRAPH, 1987. 2, 3, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 300, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 300, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 300, + 545, + 353 + ], + "type": "text", + "content": "[22] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, 2020. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "type": "text", + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. In ACM SIGGRAPH, 2022. 3, 7, 8, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 401, + 545, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 401, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 307, + 401, + 545, + 444 + ], + "type": "text", + "content": "[24] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying Neural Implicit Surfaces and Radiance Fields for Multi-View Reconstruction. In International Conference on Computer Vision, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 446, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 446, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 446, + 545, + 543 + ], + "type": "text", + "content": "[25] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary Devito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An Imperative Style, High-Performance Deep Learning Library. In Advances in Neural Information Processing Systems. Curran Associates Inc., 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 545, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 599 + ], + "type": "text", + "content": "[26] Marie-Julie Rakotosaona, Fabian Manhardt, Diego Martin Arroyo, Michael Niemeyer, Abhijit Kundu, and Federico Tombari. NeRFMeshing: Distilling Neural Radiance Fields into Geometrically-Accurate 3D Meshes. In DV, 2023. 2, 3, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "text", + "content": "[27] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3D Deep Learning with PyTorch3D. In arXiv Preprint, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[28] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. KiloNeRF: Speeding Up Neural Radiance Fields with Thousands of Tiny MLPs. In International Conference on Computer Vision, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[29] Gernot Riegler and Vladlen Koltun. Free View Synthesis. In European Conference on Computer Vision, 2020. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "text", + "content": "5362" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[30] Gernot Riegler and Vladlen Koltun. Stable View Synthesis. In Conference on Computer Vision and Pattern Recognition, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "text", + "content": "[31] Darius Rückert, Linus Franke, and Marc Stamminger. ADOP: Approximate Differentiable One-Pixel Point Rendering. In ACM SIGGRAPH, 2022. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 171 + ], + "type": "text", + "content": "[32] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo Tourism: Exploring Photo Collections in 3D. In ACM SIGGRAPH, 2006. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "text", + "content": "[33] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural Reflectance and Visibility Fields for Relighting and View Synthesis. In Conference on Computer Vision and Pattern Recognition, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "text", + "content": "[34] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct Voxel Grid Optimization: Super-Fast Convergence for Radiance Fields Reconstruction. In Conference on Computer Vision and Pattern Recognition, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 326 + ], + "type": "text", + "content": "[35] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields. In Conference on Computer Vision and Pattern Recognition, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 381 + ], + "type": "text", + "content": "[36] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-View Reconstruction. In Advances in Neural Information Processing Systems, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 382, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 425 + ], + "type": "text", + "content": "[37] Daniel N. Wood, Daniel I. Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H. Salesin, and Werner Stuetzle. Surface Light Fields for 3D Photography. In ACM SIGGRAPH, 2000. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "text", + "content": "[38] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume Rendering of Neural Implicit Surfaces. In Advances in Neural Information Processing Systems, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 459, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 502 + ], + "type": "text", + "content": "[39] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, and Jonathan T. Barron. BakedSDF: Meshing Neural SDFs for Real-Time View Synthesis. In ACM SIGGRAPH, 2023. 2, 3, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "type": "text", + "content": "[40] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees For Real-Time Rendering of Neural Radiance Fields. In International Conference on Computer Vision, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "text", + "content": "[41] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. In Conference on Computer Vision and Pattern Recognition, 2022. 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "text", + "content": "[42] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. In Conference on Computer Vision and Pattern Recognition, 2022. 8, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[43] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse Rendering with Spherical Gaussians for Physics-Based Material Editing and Relighting. In Conference on Computer Vision and Pattern Recognition, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "text", + "content": "[44] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 328, + 73, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 328, + 73, + 545, + 95 + ], + "type": "text", + "content": "Deep Features as a Perceptual Metric. In Conference on Computer Vision and Pattern Recognition, 2018. 7, 2, 4" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "text", + "content": "5363" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/80ad799c-9412-49c0-a322-3a5e2a9d1a8f_content_list.json b/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/80ad799c-9412-49c0-a322-3a5e2a9d1a8f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d7ae1b4eea54c9f32f74c80ed05caac665eb2db8 --- /dev/null +++ b/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/80ad799c-9412-49c0-a322-3a5e2a9d1a8f_content_list.json @@ -0,0 +1,1294 @@ +[ + { + "type": "text", + "text": "SubT-MRS Dataset: Pushing SLAM Towards All-weather Environments https://superodometry.com/datasets", + "text_level": 1, + "bbox": [ + 117, + 128, + 851, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shibo Zhao\\*, Yuanjun Gao, Tianhao Wu, Damanpreet Singh, Rushan Jiang, Haoxiang Sun, Mansi Sarawata, Yuheng Qiu, Warren Whittaker, Ian Higgins, Yi Du, Shaoshu Su, Can Xu, John Keller, Jay Karhade, Lucas Nogueira, Sourojit Saha, Ji Zhang, Wenshan Wang, Chen Wang, Sebastian Scherer", + "bbox": [ + 84, + 203, + 895, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Carnegie Mellon University $^{2}$ University at Buffalo", + "bbox": [ + 289, + 261, + 687, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 314, + 313, + 329 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Simultaneous localization and mapping (SLAM) is a fundamental task for numerous applications such as autonomous navigation and exploration. Despite many SLAM datasets have been released, current SLAM solutions still struggle to have sustained and resilient performance. One major issue is the absence of high-quality datasets including diverse all-weather conditions and a reliable metric for assessing robustness. This limitation significantly restricts the scalability and generalizability of SLAM technologies, impacting their development, validation, and deployment. To address this problem, we present SubT-MRS, an extremely challenging real-world dataset designed to push SLAM towards all-weather environments to pursue the most robust SLAM performance. It contains multi-degraded environments including over 30 diverse scenes such as structureless corridors, varying lighting conditions, and perceptual obscurants like smoke and dust; multimodal sensors such as LiDAR, fisheye camera, IMU, and thermal camera; and multiple locomotions like aerial, legged, and wheeled robots. We developed accuracy and robustness evaluation tracks for SLAM and introduced novel robustness metrics. Comprehensive studies are performed, revealing new observations, challenges, and opportunities for future research.", + "bbox": [ + 75, + 345, + 473, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 718, + 209, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Simultaneous Localization and Mapping (SLAM) is essential in robotics since it provides foundational perception and spatial awareness, and enables machines to understand and interact with the physical world in real-time. Therefore, it has a wide range of applications such as autonomous driving and space exploration. Despite significant advancements in both geometric [5, 34, 41] and data-driven SLAM methods [43, 48], existing solutions remain fragile in challenging conditions. One main reason is that existing SLAM", + "bbox": [ + 75, + 739, + 468, + 877 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "algorithms are often developed and evaluated with datasets from controlled environments [3, 12, 18, 21, 33, 35, 39, 54]. For example, KITTI [18, 21] dataset is mostly collected in sunny weather and EuROC-MAV dataset [3] is collected in small well-lit rooms. These datasets, unfortunately, fail to capture the challenges in real-world scenarios, hindering the development of robust SLAM solutions.", + "bbox": [ + 500, + 315, + 893, + 421 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To bridge this gap and push SLAM towards all-weather environments, we present an extremely challenging dataset, SubT-MRS, including scenarios featuring various sensor degradation, aggressive locomotions, and extreme-weather conditions. The SubT-MRS dataset comprises 3 years of data from the DARPA Subterranean (SubT) Challenge [1] (2019-2021) and extends with an additional 2 years of diverse environments (2022-2023), containing mixed indoor and outdoor settings, including long corridors, off-road scenario, tunnels, caves, deserts, forests, and bushlands. Cumulatively, this forms a 5-year dataset encompassing over 500 hours and 100 kilometers of terrain subjected to multimodal sensors including LiDAR, fisheye cameras, depth cameras, thermal cameras, and IMU; heterogeneous platforms including RC cars, legged robots, aerial robots, and wheeled robots; and extreme obscurant conditions such as dense fog, dust, smoke, and heavy snow.", + "bbox": [ + 496, + 426, + 893, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Additionally, we find there is no well-established metric to evaluate the robustness of a SLAM system. Existing evaluation metrics such as absolute trajectory error (ATE) [3] are not representative of the actual performance in robotic applications. We argue that to ensure the safety control of a robot, SLAM system evaluation should not only focus on poses but also its velocity. Taking robot localization as an example, while a momentary pose error spike may only slightly undermine the overall ATE, however, it can lead to a catastrophic crash in an aerial robotic platform. To effectively gauge the performance of a SLAM algorithm, we introduce a new robustness metric to evaluate the reliability of the SLAM system, particularly examining smoothness and accuracy of velocity estimation.", + "bbox": [ + 496, + 688, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: shiboz@andrew.cmu.edu", + "bbox": [ + 94, + 886, + 359, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22647", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b3d861d86989a475f5b89f18a2bceb82e3f55d46da24ddf4bfe5e6348b1bff74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 89, + 893, + 292 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3e2278a11c657ec21d4a1b1d9bd38063d647760e8f90618979c79422d7d0a8b0.jpg", + "image_caption": [ + "SubT-Tunnel", + "SubT-Urban", + "SubT-Cavern", + "Multi Floors" + ], + "image_footnote": [], + "bbox": [ + 80, + 292, + 893, + 333 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4ff2e7e8ea5727e51f0d28853e2951db92306ff0c8c78c86400d2dbee3f76ece.jpg", + "image_caption": [ + "Dark Long Corridor", + "Dense Fog", + "Laurel Cavern", + "Textureless Corridor", + "Darkness Stairs with Jerk Motion", + "Dense Dust and Smoke", + "Offroad with Aggressive Motion", + "Snow with Slipping Motion", + "Figure 1. Dense reconstruction from the SubT-MRS dataset, achieved through collaboration with diverse robots equipped with multimodal sensors. Colors represent different challenging environments (tunnels, caves, urban, confined spaces) captured by various robot types (aerial, legged, wheeled). The bottom section showcases a gallery with diverse visual, LiDAR, and mixed degradations." + ], + "image_footnote": [], + "bbox": [ + 80, + 345, + 893, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lastly, we perform extensive experiments using proposed degradation datasets to benchmark visual and LiDAR SLAM algorithms. These experiments identify the limitations of the existing SLAM systems and evaluate their robustness across the degradation. Our contributions include:", + "bbox": [ + 75, + 479, + 468, + 556 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- All-weather Environments To push SLAM toward all-weather environments, we introduce SubT-MRS, an extremely challenging dataset. Spanning five years, it comprises over 500 hours and 100 kilometers of accurately measured trajectories. To the best of our knowledge, SubT-MRS is the first real-world dataset that specifically addresses failure scenarios of SLAM by incorporating a variety of degraded conditions, multiple robotic platforms, and diverse sets of multimodal sensors.", + "- Robustness Metric To evaluate the actual performance running on robots, we propose a robustness metric, which, to the best of our knowledge, is the first metric evaluating the reliability, safety, and resilience of SLAM.", + "- SLAM Challenge We provide a comprehensive benchmark1 based on the SubT-MRS dataset by conducting ICCV'23 SLAM Challenge. Our evaluation reveals the limitations of state-of-the-art visual and LiDAR SLAM solutions. Furthermore, we conduct extensive experiments to verify robustness metric against various sensor degradation and length of trajectories." + ], + "bbox": [ + 76, + 556, + 468, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 478, + 640, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal Sensors Multimodal sensor datasets are crucial for the development of robust SLAM systems, as single sensor modalities are not comprehensive for all scenarios. Existing datasets typically focus on a limited range of modalities, such as monocular or stereo cameras [3, 49], LiDAR [19], event cameras [29], and depth cameras [42]. The KITTI dataset [19], although covering most modalities, is geared towards on-road scenarios and lacks thermal cameras, limiting its applicability to simple, controlled environments. Conversely, the ViViD++ dataset [23] incorporates thermal images in outdoor settings but is deficient in hardware synchronization, posing challenges for SLAM system development. The Weichen dataset [10] offers thermal images in indoor environments with accurate ground truth poses but is confined to motion capture room settings. In contrast, our SubT-MRS dataset delivers a comprehensive suite of time-synchronized multimodal sensors, including LiDAR, monocular cameras, thermal cameras, depth cameras, and IMU, along with centimeter-level ground truth, catering to a diverse range of research needs.", + "bbox": [ + 496, + 503, + 890, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Heterogeneous Platforms Developing a robust SLAM system necessitates versatility in handling different motion patterns and various robotic platforms. However, most existing datasets are tailored for single-robot scenarios. While there are a few multi-robot datasets, they typically involve homogeneous platforms [9, 15, 24, 44], posing challenges", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1SubT-MRS dataset is completely real-world data. Simulated data in the benchmark is to ensure evaluation completeness", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22648", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/01f3a26eaab9dae558cba6309166aa16b55eafc0d3c44dd7989cb8309dc889dd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMulti-SpectrumMulti-DegradationMulti-Robot
CameraIMULiDAR/DepthThermalIlluminationSnow SmokeStructurelessSubTAggressive MotionVehicleDroneLeggedHandheld
EuRoC MAV [3]✓✓XX✓XXXXX✓XX
PennCOsYVIO [33]✓✓XXXXXXXXXX✓
TUM VIO [39]✓✓XX✓XXXXXXX✓
UZH-FPV [12]✓✓XXXXXX✓X✓XX
College Dataset [52]✓✓✓XXXXXXXXX✓
RobotCar [28]✓✓✓X✓✓XXX✓XXX
UMA VI [54]✓✓XX✓X✓XXXXX✓
UMich [6]✓✓✓X✓XXXX✓XXX
KITTI [19]✓✓✓XXX✓X✓✓XXX
Virtual KITTI [16]✓✓✓X✓X✓X✓✓XXX
DARPA SubT [37]✓✓✓X✓X✓✓X✓XXX
Hilti SLAM [21]✓✓✓X✓X✓XXXXX✓
M3ED [7]✓✓✓X✓XXXX✓✓✓X
TartanAir [47]✓✓✓X✓✓✓X✓X✓XX
SubT-MRS (Ours)✓✓✓✓✓✓✓✓✓✓✓✓✓
", + "bbox": [ + 80, + 90, + 890, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparison of SLAM datasets on multi-sensors, multi robots, and multi degradation.", + "bbox": [ + 204, + 287, + 759, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in evaluating SLAM performance across varied robotic platforms. The AirMuseum dataset [13] includes drones and ground robots but lacks data from legged robots. As a comparison, SubT-MRS dataset features diverse robotic platforms including legged robots, aerial robots, and wheeled robots, operating in varied environments and sensor setups.", + "bbox": [ + 75, + 328, + 467, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extreme Environmental Conditions Developing and testing SLAM systems in extreme environmental conditions is crucial for mitigating potential real-world failures. However, most existing SLAM datasets have been primarily limited to single, controlled environments. The TUM-VI [39] and UMA-VI [54] datasets, being indoor-outdoor visual-inertial datasets, pose challenges due to varying illumination and low-texture environments. The EuRoC MAV [3] and UZH-FPV drone racing [12] datasets, popular in the SLAM community, offer data on aggressive drone movements but usually in consistently lit conditions. The KITTI dataset [19], a staple in autonomous driving research for its outdoor LiDAR-Visual-Inertial data. The Hilti SLAM dataset [21] includes both indoor and outdoor LiDAR-visual-inertial datasets with dynamic lighting and confined spaces. However, both of them focus more on accuracy than on a variety of environmental degradations. TartanAir [47] covers most of the challenging environments but it is a simulation dataset that will pose the sim-to-real gap.", + "bbox": [ + 75, + 425, + 467, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, the SubT-MRS dataset provides 30 varied scenes, encompassing a wide array of environments. These include SubT tunnels, urban areas, caverns, multifloor structures, dark corridors, and foggy conditions. It also features textureless surfaces, dusty and smoky environments, off-road areas with aggressive motion, and snowy terrains prone to slippage. We include challenging lighting conditions such as extreme darkness and overexposure, along with geometrically challenging scenarios like featureless corridors, staircases, and self-similar cave layouts. Additionally, the dataset incorporates perceptual challenges posed by smoke, fog, dust, and various weather conditions.", + "bbox": [ + 75, + 719, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SubT-MRS Dataset", + "text_level": 1, + "bbox": [ + 498, + 325, + 686, + 342 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We next present the SubT-MRS dataset from three aspects including its collecting environments and settings, ground truth collection, and the new robustness metric. A detailed comparison with other datasets is listed in Table 1.", + "bbox": [ + 496, + 348, + 890, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. All-weather Environments", + "text_level": 1, + "bbox": [ + 498, + 407, + 738, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As mentioned in Sec. 2, the SubT-MRS dataset distinguishes itself through its multimodal sensor setups, heterogeneous robotic platforms, and extreme environmental conditions. This section emphasizes this comprehensive nature pushing SLAM towards all-weather environments.", + "bbox": [ + 496, + 425, + 890, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Multimodal Sensors", + "text_level": 1, + "bbox": [ + 498, + 503, + 702, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal sensors provide a wealth of information for SLAM systems operating in all-weather environments. This diversity improves scene understanding and strengthens the system's perception ability. Therefore, We incorporate diverse sensors to ensure system robustness.", + "bbox": [ + 496, + 521, + 890, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sensor Pack We embedded 4 Leopard Imaging RGB fisheye cameras, 1 Velodyne puck, 1 Epson M-G365 IMU, 1 FLIR Boson thermal camera, and NVIDIA Jetson AGX Xavier as a sensor pack shown in Figure 2.", + "bbox": [ + 496, + 597, + 890, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Time Synchronization To ensure the overall consistency of the fused data, we meticulously synchronize the sensors' time using the 'pulse per second (PPS)' technique. As illustrated in Figure 2, the IMU, LiDAR, and thermal camera are directly synchronized with the CPU clock, while the four RGB camera are synchronized using an FPGA board. Consequently, we can effectively manage the time synchronization gap between each pair of sensors not to exceed 3ms.", + "bbox": [ + 496, + 657, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fisheye and Thermal Camera Calibration Camera calibration plays an essential role in the efficiency of a SLAM system. For calibrating the fisheye cameras, we employed the open-source toolkit Kalibr [36], focusing on the intrinsic and extrinsic parameters. Specifically, we use the radial-tangential distortion model to rectify the omnidirectional fisheye camera model. Calibrating thermal cameras, however, poses unique challenges, especially in gathering high-", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22649", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4a5dcfa2e9f9e3cf0bef64f230e655cdef81aa0eee5146db6cfaca25eed1ea7d.jpg", + "image_caption": [ + "Figure 2. An overview of the sensor pack used in SubT-MRS dataset. It is equipped with a Xavier processing unit with hardware time synchronization for multimodal sensors including LiDAR, fisheye cameras, thermal cameras, depth cameras (option), and an IMU." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 851, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "quality thermal data. To tackle this, we set up a $7 \\times 9$ chessboard, heated by direct sunlight, to generate high-contrast thermal data, ensuring accuracy in the calibration process.", + "bbox": [ + 75, + 321, + 468, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "IMU Calibration and Extrinsic Calibration For Lidar-IMU extrinsic calibration, we utilize the CAD model to obtain the calibration parameters. In the case of Camera-IMU extrinsic calibration, we employ the Kalibr toolbox [36] to estimate the extrinsic matrix. To estimate the sensors' bias and the random walk noise of the gyroscope and accelerometer, we collected static data from the IMU and calibrated it using an Allan variance-based IMU tool [14, 17].", + "bbox": [ + 75, + 366, + 468, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2. Multi-degraded Environments", + "text_level": 1, + "bbox": [ + 76, + 489, + 361, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SubT-MRS includes multiple challenging environments including visually degraded environments, geometrically degraded environments, and their combination.", + "bbox": [ + 75, + 507, + 468, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual Degradation Poor-quality visual features can significantly hinder the performance of feature extraction processes and disrupt the accuracy of feature matching. Such issues arise in low-light conditions with inconsistent brightness or image noise introduced by air obscurants. SubT-MRS encompasses these challenges and provides a wide range of visual degradation. This includes environments with limited lighting, such as hospital interiors and caves (Figure 3 A-F), as well as smoky or dusty conditions that cause visual obstruction (Figure 3 M-N), and snowy areas with reduced visibility (Figure 3 H, K, and L).", + "bbox": [ + 75, + 551, + 468, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geometric Degradation Lack of geometrical features poses significant challenges to LiDAR odometry. The root of this issue often lies in the limited sensing capabilities of LiDAR sensors and the constraints due to their mechanical installation. SubT-MRS captures a variety of environments that exemplify such challenges. It includes long, featureless corridors (Figure 3 E) and staircases (Figure 3 C and G). These scenarios illustrate various forms of geometric degradation, contributing to the improvement of LiDAR odometry systems in challenging conditions.", + "bbox": [ + 75, + 719, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mixed Degradation A mixed visual and geometric degradation can further hinder the performance of SLAM sys", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tems. Examples in SubT-MRS include long, dimly lit corridors (Figure 3 A, E, and F), poorly illuminated staircases (Figure 3 C), and environments affected by snowy weather (Figure 3 H, K, and L). In these settings, both LiDAR odometry and visual odometry are prone to failure due to the compounded effects of mixed degradation. The inclusion of such scenarios in SubT-MRS is crucial for evaluating and improving the resilience of multimodal SLAM algorithms.", + "bbox": [ + 496, + 321, + 890, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.3. Heterogeneous Robot Platforms", + "text_level": 1, + "bbox": [ + 498, + 446, + 795, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Most of current datasets focus on single-robot systems, limiting multi-robot SLAM development as shown in Table 1. To address this, we employed diverse robot platforms, including aerial, legged, and wheeled robots, navigating through various environments from urban campuses to medical facilities and natural terrains like caves. This diversity offers a range of scenarios to enhance SLAM algorithms for effective operation in challenging conditions.", + "bbox": [ + 496, + 464, + 890, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Extrinsic Calibration for Multiple Robots To ensure the multi-robot system shares a common coordinate system, we perform the extrinsic calibration process [38]. It has 2 steps: First, each robot runs Super Odometry [53] to generate its local map using LiDAR data and share it with other robots through a wireless network; Second, the remaining robots identify overlapping regions between their local maps and estimate the extrinsic parameters using GICP [40].", + "bbox": [ + 496, + 585, + 890, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Ground Truth", + "text_level": 1, + "bbox": [ + 500, + 714, + 645, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ground Truth Map As depicted in Figure 4, we utilized the high-precision FARO Focus 3D S120 3D scanner for creating ground-truth maps. This cutting-edge 3D scanner can measure distances up to $120\\mathrm{m}$ , with a maximum measurement rate of $976\\mathrm{K}$ points per second. The accuracy of the ground truth map is noteworthy, maintaining a range error within $\\pm 2\\mathrm{mm}$ . We have developed ground truth models for diverse environments, including subterranean areas like urban areas $(350\\mathrm{m} \\times 350\\mathrm{m})$ , caves $(150\\mathrm{m} \\times 200\\mathrm{m})$ and tunnels $(100 \\times 200\\mathrm{m})$ , as well as indoor-outdoor spaces $(200\\mathrm{m} \\times 200\\mathrm{m})$ within the CMU campus. To reduce drift, a", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22650", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e99cad20d18ec8f0048688cceaa90f07e28c3d2bab5ffea665868a82026408ba.jpg", + "image_caption": [ + "Figure 3. The SubT-MRS datasets were collected across diverse seasons, capturing environments with perceptual challenges such as poor illumination, darkness, and water puddles, where visual sensors falter. They also include geometrically complex areas like long featureless corridors and steep multi-floor structures, challenging LiDAR systems with potential drift. Moreover, these datasets cover conditions with airborne obscurants like dust, fog, snow, and smoke in tough environments, including caves, deserts, long tunnels, and off-road areas." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 890, + 410 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3c0d36f7d01aeff96cc49b14894396b1d84ab5655b90321852f2810c4bddc139.jpg", + "image_caption": [ + "Figure 4. The SubT-MRS dataset facilitates the generation of high-precision ground truth maps and trajectories. Figure (a) shows the ground truth trajectory in multi-floor settings. Figure (b) displays ground truth maps for indoor and outdoor areas, encompassing long corridors, multi-floor structures, and open spaces. Figure (c) features photo-realistic scans based on our ground truth maps in cave environments." + ], + "image_footnote": [], + "bbox": [ + 117, + 491, + 851, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "loop closure algorithm was applied to correct the poses. Notably, $96\\%$ of the scans achieved a position uncertainty of less than $2\\mathrm{mm}$ . This ground-truth map is crucial for SLAM development where precision is of utmost importance.", + "bbox": [ + 75, + 773, + 470, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ground Truth Trajectory Ground truth trajectories for all sequences are generated based on the ground truth map. In this process, we establish point-to-point, point-to-plane, and point-to-line correspondences[40] between the Ground", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Truth map and the current LiDAR scan. Our estimated ground truth trajectory is a comprehensive fusion of various constraints derived from the ground truth map, visual odometry, LiDAR odometry, and IMU measurements. This approach ensures the robustness of our solution across diverse environmental conditions. To verify accuracy, we reconstruct LiDAR maps from ground truth trajectory and compare it with our ground truth maps.", + "bbox": [ + 496, + 773, + 893, + 897 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22651", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fdec7445e1994ac9e22a65403478b80a8d9b0efe3e082ff2af469ac640c8c455.jpg", + "table_caption": [ + "Table 2. SLAM Challenge Results (Blue shadings indicate rankings of ATE and Robustness Metric; L: LiDAR I: IMU C: Camera)" + ], + "table_footnote": [], + "table_body": "
#TeamMethodOdometry TypeDeviceRealTime (s)CPU/GPU (%)RAM (GB)ATE↓\\( {R}_{v} \\uparrow /{R}_{w} \\uparrow \\)Sensors L I C
1Liu et alFAST-LIO2 [51], HBA [27]FilterIntel i7-9700K51.31098.667 / 04.0520.5880.517/0.770✓✓
2Yibin et alLIO-EKF [45]FilterIntel i7-107000.00652.167 / 00.0724.3130.441/0.574✓✓
3Weitong et alFAST-LIO[2], Pose Graph[11]FilterIntel Xeon(R)E3-1240v50.12522.63 / 04.3050.6630.473/0.747✓✓
4Kim et alFAST-LIO2[50], Point-LIO[20], Quatro[25]FilterIntel i5-125000.268101.108 / 055.643.8250.479/0.615✓✓
5Zhong et alDLO[8], Scan-Context++[22]SW OptAMD Ryzen 9 5900x0.02713.289 / 01.1741.2090.276/0.486✓✓
1Peng et alDVI-SLAM [32]LearningIntel i9-12900183.233- / 14911 (4)0.5470.473/0.788✓✓
2Jiang et alLET-NET[26], VINS-Mono[34]HybridIntel i5-94000.06440.35 / 04.3371.0930.078/0.322✓✓
3Thien et alVR-SLAM[31]SW OptIntel i9-129000.142176.44 / 09.1113.0370.083/0.372✓✓
4Li et alORB-SLAM3[4]SW Opt.Intel i7-107000.01965.028 / 00.3868.9750.163/0.474✓✓
", + "bbox": [ + 80, + 114, + 903, + 260 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2ed6115a3035f29a5745994d8f0ae28f3eb62609250d5bd3571af50b6022515a.jpg", + "table_caption": [ + "Table 3. Accuracy Performance on Geometric Degradation. Red numbers represent ATE ranking. * denotes incomplete submissions." + ], + "table_footnote": [], + "table_body": "
TeamGeometric Degradation (Real World)SimulationMix DegradationAverage
UrbanTunnelCaveNuclear_1Nuclear_2Laurel_CavernsFactoryOceanSewerageLong CorridorMulti FloorBlock Lidar
Liu et al10.3070.0950.6290.1220.2350.2600.8890.7570.9781.4540.4010.9340.588
Weitong et al20.260.0960.6170.1200.2220.4020.9980.7701.5861.2540.5771.0560.663
Kim et al30.3310.0920.7870.1230.2700.27910.62822.4257.1472.1000.6501.0683.825
Yibin et al41.0600.2200.7500.4700.6209.1404.9200.28024.4602.9905.5001.3404.312
Zhong et al51.2050.695-1.1751.722.080.8890.7781.13---1.209*
Average0.6330.2400.6960.4020.61342.4323.6655.0027.0601.9501.7821.099
", + "bbox": [ + 81, + 299, + 883, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Robustness Metric", + "text_level": 1, + "bbox": [ + 76, + 436, + 256, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As discussed in Sec. 1, existing evaluation metrics such as absolute trajectory error (ATE) [3] have limitations in evaluating the SLAM's robustness in real-world applications. ATE primarily focuses on the accuracy of trajectory and does not consider the completeness of trajectory (recall). Also, it can not effectively capture velocity changes that have a direct impact on the robot's safety. For example, a pose spike error might not be immediately noticeable in ATE evaluation, it may result in a crash in an aerial robot control. To address these gaps, we introduce a new robustness metric based on estimated velocity. That is because velocity estimation from SLAM is crucial for the robot's control system which directly impacts stability, affecting the robot's safety. The new robustness metric is the area under the curve (AUC) of the F-1 score:", + "bbox": [ + 75, + 454, + 472, + 680 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {1} (e) = \\frac {2 P (e < T) R (e < T)}{P (e < T) + R (e < T)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 691, + 468, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the precision $P$ quantifies the precision of the estimated velocity as a percentage: how closely the estimated velocity points lie to the ground truth point and the recall rate $R$ quantifies the estimated velocity's completeness: to what extent all the ground-truth points are covered. A high F-1 score can only be achieved by the velocity estimation that is both accurate and complete throughout the entire run. Specifically, an estimated error $e$ is regarded as an inlier, if it is smaller than a threshold $T$ . To scale the threshold $T$ within the range of [0, 1], we apply exponential mapping $\\exp(-10T)$ when calculating robustness metric.", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Position Robustness $R_{p}$ & Rotation Robustness $R_{r}$ The AUC of F-1 score can be defined using both linear velocity and angular velocity, which can reflect the robustness of position and rotation estimation, respectively. Specifically, we define $R_{p} = \\mathrm{AUC}(F_{1}(\\mathrm{v}_{\\mathrm{e}}))$ and $R_{r} = \\mathrm{AUC}(F_{1}(\\omega_{\\mathrm{e}}))$ , where $v_{e}$ and $\\omega_{e}$ are the estimated error of linear velocity and rotation velocity, respectively. Note that not all SLAM solutions can output the velocity at the desired frequency. To address this, we use B-splines [30, 46] to obtain smooth trajectories, compute derivatives for smooth trajectories, and derive estimated linear velocity and angular velocity. For more details, please refer to the supplementary.", + "bbox": [ + 496, + 436, + 890, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Open SLAM Challenge Results", + "text_level": 1, + "bbox": [ + 498, + 628, + 785, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we will present our works from two perspectives: accuracy evaluation and robustness evaluation.", + "bbox": [ + 496, + 648, + 890, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Accuracy Evaluation", + "text_level": 1, + "bbox": [ + 498, + 686, + 697, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results of the ICCV SLAM challenge underscore the necessity for advancements in system robustness. From 29 submissions, we identified 5 winners in the LiDAR category and 4 in the visual category. However, in the sensor fusion track, which addresses both visual and geometric degradation, no submissions met the criteria for success. This result reveals the existing SLAM systems still have lots of space to improve. Sequence characteristics will be detailed in the supplementary material. Table 2 shows the results for both the LiDAR and visual tracks. Unfortunately, there are no current solutions that can balance high accuracy and real-time performance in challenging environments. Since several environments lack geometric features or visual features,", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22652", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/811b48b0e0fffb482362b1dad99129a7b483227a851a403173d18320f4a4f3dc.jpg", + "table_caption": [ + "Table 4. Accuracy Performance on Visual Degradation. Red numbers represent ATE ranking. * denotes incomplete submissions." + ], + "table_footnote": [], + "table_body": "
TeamVisual Degradation (Real World)SimulationAverage
Lowlight 1Lowlight 2Over ExposureFlash LightSmoke RoomOutdoor NightEnd of WorldMoonWestern Desert
Peng et al11.0631.6370.5030.440.1530.8270.0380.1950.0700.547
Thien et al21.0812.0541.7331.05410.5327.6920.7531.2281.2093.037
Jiang et al31.0191.1261.9112.3413.75711.8212.1540.6044.0103.193
Li et al45.7687.8341.7571.2955.37010.766-30.07-8.98*
Average2.2323.1631.4761.2824.9537.7760.9828.0241.763
", + "bbox": [ + 119, + 114, + 846, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b8e8b8ab874deecdf5e7d6a75971d02e8e9b765ff81b297d02a906c91599ca0.jpg", + "image_caption": [ + "Figure 5. From left to right, it shows robustness metric $R_{p}$ and $R_{r}$ for LiDAR and visual sequences respectively. Note: This is a summary of results for all sequences, with weights based on the trajectory length. The area under the curve (AUC) represents the robustness $(R_{p}, R_{r})$ . The x-axis shows velocity thresholds for classifying estimated velocities as inliers and the y-axis is F-1 score." + ], + "image_footnote": [], + "bbox": [ + 138, + 224, + 316, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55cfc6031a312bd26283cd3f30215cbe3a40100a153c8fd52f3f5b533cf7c593.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 226, + 482, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ddcbb645616be2d30870b196461b2b337a9d4108afc101e9f4d046bbd9ca7d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 226, + 658, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddea1daf4391967666e91c1708573d8df339b08b3c6cbb9141b9eeb3029f5647.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 226, + 831, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the algorithms required more processing times to achieve reasonably high accuracy. Liu et al. at the University of Hong Kong, who achieved the highest accuracy, recorded an Absolute Trajectory Error (ATE) of $0.588\\mathrm{m}$ , but each iteration took 51.3 seconds. In the visual track, Peng et al. from the Samsung Research Center attained a leading ATE of $0.547\\mathrm{m}$ , but each iteration required 183 seconds.", + "bbox": [ + 75, + 429, + 470, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LiDAR Track Discussion To enhance robustness in all-weather environments, a pivotal question arises: what are the limits of current LiDAR SLAM algorithms? Table 3 presents a summary of the ATE/RPE errors observed in real-world geometrically degraded, simulated, and mixed degradation scenarios across the top five teams. In real-world geometrically degraded environments, we observed the first limitation: existing LiDAR solutions struggle in confined spaces such as caves. Almost all algorithms perform worse in Cave (Figure 3S) and Laureal_Cavern environments (Figure 3N) with average ATE 0.696 and 2.432 meters respectively. One reason is that cave environments is the most confined space which may lack sufficient geometric features, compared with other environments like tunnel (Figure 3J) and Urban (Figure 3B) and Nuclear scenes (Figure 3U).", + "bbox": [ + 75, + 536, + 468, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In simulated environments, we encountered the second limitation: existing methods are tailored for fixed motion patterns like vehicles and struggle with unpredictable motion patterns. Since our simulation sequence is derived from TartanAir [47], featuring aggressive and random motion patterns, it might not follow the usual velocity distribution domain and has not been thoroughly tested. As shown in Table 3 (simulation), it resulted in significantly higher ATE errors $(5.239\\mathrm{m})$ compared to real-world's $(0.836\\mathrm{m})$ .", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In mixed degradation environments, the third limitation we found is that existing methods cannot actively select the most informative measurements to adapt to new environments. For example, in the Long Corridor sequence (Figure 3 A) with low lighting and a featureless environment, the scenario is intricately designed for collaborative usage of LiDAR and Visual sensors. Algorithms are expected to utilize visual information in geometrically degraded environments while disregarding it in darkness. Similarly, the Block LiDAR sequences (Figure 3 V) aim to simulate sensor drop scenarios frequently encountered in robotic applications. These sequences involve alternating periods of LiDAR and Visual data loss, challenging algorithms to promptly detect sensor failures and switch to the other modality for SLAM. However, the average ATE error $(1.61\\mathrm{m})$ in Table 3 (mixed degradation) is much higher than in geometric degradation scenarios $(0.836\\mathrm{m})$ , suggesting the above limitations. Details are in the supplementary.", + "bbox": [ + 496, + 429, + 890, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual Track Discussion To improve the robustness, a pivotal question arises: what are the limits of current visual SLAM? To what extent does image quality impact a visual SLAM system? Table 4 shows a summary of ATE errors from real-world and simulation scenarios from the awarded 4 teams. In real-world scenarios, the first limitation is the lack of anti-noise capability in current methods, especially in low-quality image settings. We assessed visual odometry accuracy in visually degraded environments, including low lighting, sunlight overexposure, flashing lights, smoke-filled rooms, and nighttime outdoor settings (Figure 3). Trajectories in conditions like the Smoke Room and Outdoor Night exhibit significantly higher average ATE er", + "bbox": [ + 496, + 703, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22653", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/673a090df4bcba9895606a892977e11654621b21cdf8b4bf2da0799f720b5c65.jpg", + "table_caption": [ + "Table 5. Robustness Performance on Geometric Degradation. Red numbers are robustness ranking. Larger values indicate better robustness." + ], + "table_footnote": [], + "table_body": "
TeamGeometric Degradation (Real World)SimulationMix DegradationAverage
UrbanTunnelCaveNuclear_1Nuclear_2Laurel_CavernsFactoryOceanSewerageLong CorridorMulti FloorBlock Lidar
Liu et al1Rp↑0.8110.8650.7360.7470.5040.8160.1570.1350.1350.3960.5290.3710.516
Weitong et al20.7730.8380.6900.8710.7390.1520.1530.1490.0810.4070.4560.3450.471
Kim et al40.7470.8650.7370.8700.6890.7770.0140.0010.0100.4100.2850.3450.479
Yibin et al30.6500.7830.5570.7210.4900.4810.2190.5600.0190.1750.2600.3700.440
Zhong et al50.5670.6830.2040.6800.4260.4470.1030.0840.1220000.276
Liu et al1Rr↑0.8880.8930.8160.8570.7780.8610.6920.7190.6520.7530.6890.6430.770
Weitong et al20.8860.8920.8110.8930.8400.5980.6880.7200.5140.7660.7060.6440.746
Kim et al40.6800.8930.8160.8930.8030.8080.1980.1250.3000.8020.4220.6420.615
Yibin et al30.6240.7070.4620.7330.5400.4670.8620.9620.4960.4770.3420.2130.573
Zhong et al50.7310.7450.4370.7570.6030.5230.6880.7040.6430000.485
AverageRp0.7100.8070.5850.7780.5700.5350.1300.1860.0730.2780.3060.286
Average Rr0.7620.8260.6680.8270.7130.6510.6260.6460.5210.5600.4320.428
", + "bbox": [ + 112, + 111, + 854, + 277 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0554e414eda0c184fb5086e9c578d4cb0574e8ba3786ffd0e627be43dfee6a5e.jpg", + "table_caption": [ + "Table 6. Robustness Performance on Visual Degradation. Red numbers are robustness ranking. Larger values indicate better robustness." + ], + "table_footnote": [], + "table_body": "
TeamVisual Degradation (Real World)SimulationAverage
Lowlight 1Lowlight 2Over ExposureFlash LightSmoke RoomOutdoor NightEnd of WorldMoonWestern Desert
Peng et al10.3570.2270.2640.2030.5360.2700.6990.8930.8060.472
Thien et al30.0450.0700.2400.1560.1310.07500.03100.083
Jiang et al4Rp↑0.0460.0390.1940.0880.2420.0950000.078
Li et al20.3420.1870.2570.2080.3220.14200.00600.162
Peng et al10.6410.5810.7440.6500.8780.6700.9750.9750.9740.787
Thien et al30.4130.4450.6100.2690.4740.4870.1770.31500.354
Jiang et al4Rr↑0.4530.4520.6190.2520.5420.5770.00200.1570.339
Li et al20.6420.5740.6570.6510.7730.6600.00.30500.473
AverageRp0.1980.1310.2390.1640.3080.1460.1750.2320.202
Average Rr0.5370.5130.6580.4560.6670.5980.2880.3990.283
", + "bbox": [ + 106, + 311, + 854, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rors (4.95m, 7.76m respectively). Smoke and night scenes challenge feature extraction and tracking and introduce sensor noise, emphasizing the need for robust anti-noise algorithms. The second limitation is that existing methods struggle to overcome aggressive motion. Even in simulations with relatively good image quality, most methods still show significant average ATE errors (8.024m) on the Moon sequence (Figure 3 Y). See more details in supplementary.", + "bbox": [ + 75, + 484, + 468, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Robustness Evaluation", + "text_level": 1, + "bbox": [ + 76, + 611, + 289, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The F-1 robustness curve provides a detailed evaluation under different error tolerances, while $R_{p}$ and $R_{r}$ provide an overall measure of robustness that is not dependent on a specific decision threshold. This suggests the flexibility of our new robustness metric. Figure 5 clearly shows the robustness performance of all teams regarding position and rotation, highlighting the differences in robustness between the methods. The values of $R_{p}$ and $R_{r}$ , which are the AUC under F-1 curves, are displayed in the brackets of the legends. It reveals that Liu et al. and Peng et al. are the most robust solutions for LiDAR and visual tracks, respectively.", + "bbox": [ + 75, + 628, + 468, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Is the robustness metric robust? We did extensive experiments on our robustness metric against diverse sensor degradation settings. Table 5 and 6 present a summary of robustness performance in real-world, simulated, and mixed degradation scenarios for both LiDAR and visual sequences. In Table 5, we observe that the average values of $R_{p}$ and $R_{r}$ for the mixed degradation environment (0.290,", + "bbox": [ + 75, + 795, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "0.473) are significantly lower than geometric degradation environments (0.664, 0.754). This observation suggests that the majority of SLAM algorithms exhibit reduced robustness in mixed degradation environments as compared to geometrically degraded ones. This aligns with our expectations and verifies the effectiveness of the robustness metric.", + "bbox": [ + 496, + 484, + 890, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our robustness ranking, indicated by small red numbers, differs from those on ATE ranking shown in Table 3 and 4. This is because our F-1 score-based metrics $R_{p}$ and $R_{r}$ jointly consider precision and recall rate to provide a balanced evaluation of SLAM performance, considering both metrics across the full spectrum of thresholds. In contrast, the ATE focuses solely on precision, neglecting the trajectory's completeness (recall). We also evaluate our metric on various synthetic trajectories in the supplementary.", + "bbox": [ + 496, + 580, + 890, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 739, + 617, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce SubT-MRS, a comprehensive SLAM dataset with various sensor data, locomotion patterns, and over 30 degradation types in simulation and real-world settings to push SLAM towards all-weather environments. Additionally, we introduce a new robustness metric to evaluate the reliability of SLAM systems, enhancing robot safety control. 29 teams have tested SubT-MRS in the organized SLAM challenge and we expect that it will serve as a critical benchmark for future SLAM development.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22654", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 80, + 90, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] https://www.darpa.mil/program/darpa-subterranean-challenge. 1", + "[2] Chunge Bai, Tao Xiao, Yajie Chen, Haoqian Wang, Fang Zhang, and Xiang Gao. Faster-lio: Lightweight tightly coupled lidar-inertial odometry using parallel sparse incremental voxels. IEEE Robotics and Automation Letters, 7(2):4861-4868, 2022. 6", + "[3] Michael Burri, Janosch Nikolic, Pascal Gohl, Thomas Schneider, Joern Rehder, Sammy Omari, Markus W Achtelik, and Roland Siegwart. The EuRoC micro aerial vehicle datasets. The International Journal of Robotics Research, 2016. 1, 2, 3, 6", + "[4] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 6", + "[5] Carlos Campos, Richard Elvira, Juan J. Gómez Rodríguez, José M. M. Montiel, and Juan D. Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6): 1874-1890, 2021. 1", + "[6] Nicholas Carlevaris-Bianco, Arash K Ushani, and Ryan M Eustice. University of michigan north campus long-term vision and lidar dataset. The International Journal of Robotics Research, 35(9):1023-1035, 2016. 3", + "[7] Kenneth Chaney, Fernando Cladera, Ziyun Wang, Anthony Bisulco, M Ani Hsieh, Christopher Korpela, Vijay Kumar, Camillo J Taylor, and Kostas Daniilidis. M3ed: Multi-robot, multi-sensor, multi-environment event dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015–4022, 2023. 3", + "[8] Kenny Chen, Brett T. Lopez, Ali-akbar Agha-mohammadi, and Ankur Mehta. Direct lidar odometry: Fast localization with dense point clouds. IEEE Robotics and Automation Letters, 7(2):2000–2007, 2022. 6", + "[9] Long Chen, Yuchen Li, Chao Huang, Bai Li, Yang Xing, Daxin Tian, Li Li, Zhongxu Hu, Xiaoxiang Na, Zixuan Li, et al. Milestones in autonomous driving and intelligent vehicles: Survey of surveys. IEEE Transactions on Intelligent Vehicles, 8(2):1046-1056, 2022. 2", + "[10] Weichen Dai, Yu Zhang, Shenzhou Chen, Donglei Sun, and Da Kong. A multi-spectral dataset for evaluating motion estimation systems. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 5560-5566. IEEE, 2021. 2", + "[11] Frank Dellaert, Michael Kaess, et al. Factor graphs for robot perception. Foundations and Trends® in Robotics, 6(1-2): 1-139, 2017. 6", + "[12] Jeffrey Delmerico, Titus Cieslewski, Henri Rebecq, Matthias Faessler, and Davide Scaramuzza. Are we ready for autonomous drone racing? the UZH-FPV drone racing dataset. In IEEE Int. Conf. Robot. Autom. (ICRA), 2019. 1, 3", + "[13] Rodolphe Dubois, Alexandre Eudes, and Vincent Frémont. Airmuseum: a heterogeneous multi-robot dataset for stereovisual and inertial simultaneous localization and mapping. In" + ], + "bbox": [ + 78, + 109, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2020 IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI), pages 166-172, 2020. 3", + "[14] Naser El-Sheimi, Haiying Hou, and Xiaoji Niu. Analysis and modeling of inertial sensors using allan variance. IEEE Transactions on Instrumentation and Measurement, 57(1): 140-149, 2008. 4", + "[15] Dapeng Feng, Yuhua Qi, Shipeng Zhong, Zhiqiang Chen, Yudu Jiao, Qiming Chen, Tao Jiang, and Hongbo Chen. S3e: A large-scale multimodal dataset for collaborative slam. arXiv preprint arXiv:2210.13723, 2022. 2", + "[16] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4340-4349, 2016. 3", + "[17] Wenliang Gao. A ros package tool to analyze the imu performance, 2018. 4", + "[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI Vision Benchmark Suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 1", + "[19] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 3", + "[20] Dongjiao He, Wei Xu, Nan Chen, Fanze Kong, Chongjian Yuan, and Fu Zhang. Point-lio: Robust high-bandwidth light detection and ranging inertial odometry. Advanced Intelligent Systems, page 2200459, 2023. 6", + "[21] Michael Helmberger, Kristian Morin, Beda Berner, Nitish Kumar, Giovanni Cioffi, and Davide Scaramuzza. The hilti slam challenge dataset. IEEE Robotics and Automation Letters, 7(3):7518-7525, 2022. 1, 3", + "[22] Giseop Kim and Ayoung Kim. Scan context: Egocentric spatial descriptor for place recognition within 3d point cloud map. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4802-4809. IEEE, 2018. 6", + "[23] Alex Junho Lee, Younggun Cho, Young-sik Shin, Ayoung Kim, and Hyun Myung. Vivid++: Vision for visibility dataset. IEEE Robotics and Automation Letters, 7(3):6282-6289, 2022. 2", + "[24] Keith YK Leung, Yoni Halpern, Timothy D Barfoot, and Hugh HT Liu. The utias multi-robot cooperative localization and mapping dataset. The International Journal of Robotics Research, 30(8):969–974, 2011. 2", + "[25] Hyungtae Lim, Suyong Yeon, Suyong Ryu, Yonghan Lee, Youngji Kim, Jaeseong Yun, Euigon Jung, Donghwan Lee, and Hyun Myung. A single correspondence is enough: Robust global registration to avoid degeneracy in urban environments. page Accepted. To appear, 2022. 6", + "[26] Yicheng Lin, Shuo Wang, Yunlong Jiang, and Bin Han. Breaking of brightness consistency in optical flow with a lightweight cnn network, 2023. 6", + "[27] Zheng Liu, Xiyuan Liu, and Fu Zhang. Efficient and consistent bundle adjustment on lidar point clouds. IEEE Transactions on Robotics, 2023. 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22655", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Will Maddern, Geoffrey Pascoe, Chris Linegar, and Paul Newman. 1 year, $1000\\mathrm{km}$ : The oxford robotcar dataset. The International Journal of Robotics Research, 36(1):3-15, 2017. 3", + "[29] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. The International Journal of Robotics Research, 36(2):142-149, 2017. 2", + "[30] Elias Mueggler, Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. Continuous-time visual-inertial odometry for event cameras. IEEE Transactions on Robotics, 34(6): 1425-1440, 2018. 6", + "[31] Thien Hoang Nguyen, Shenghai Yuan, and Lihua Xie. Vrslam: A visual-range simultaneous localization and mapping system using monocular camera and ultra-wideband sensors, 2023. 6", + "[32] Xiongfeng Peng, Zhihua Liu, Weiming Li, Ping Tan, SoonYong Cho, and Qiang Wang. Dvi-slam: A dual visual inertial slam network, 2023. 6", + "[33] Bernd Pfrommer, Nitin Sanket, Kostas Daniilidis, and Jonas Cleveland. Penncosyvio: A challenging visual inertial odometry benchmark. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 3847-3854. IEEE, 2017. 1, 3", + "[34] Tong Qin, Peiliang Li, and Shaojie Shen. Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics, 34(4):1004-1020, 2018. 1, 6", + "[35] Milad Ramezani, Yiduo Wang, Marco Camurri, David Wisth, Matias Mattamala, and Maurice Fallon. The newer college dataset: Handheld lidar, inertial and vision with ground truth. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4353-4360. IEEE, 2020. 1", + "[36] Joern Rehder, Janosch Nikolic, Thomas Schneider, Timo Hinzmann, and Roland Siegwart. Extending kalibr: Calibrating the extrinsic of multiple imus and of individual axes. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 4304-4311. IEEE, 2016. 3, 4", + "[37] John G Rogers, Arthur Schang, Carlos Nieto-Granda, John Ware, John Carter, Jonathan Fink, and Ethan Stump. The darpa subt urban circuit mapping dataset and evaluation metric. In Experimental Robotics: The 17th International Symposium, pages 391–401. Springer, 2021. 3", + "[38] Sebastian Scherer, Vasu Agrawal, Graeme Best, Chao Cao, Katarina Cujic, R Darnley, R DeBortoli, E Dexheimer, B Drozd, R Garg, et al. Resilient and modular subterranean exploration with a team of roving and flying robots. Field Robotics, 2022. 4", + "[39] David Schubert, Thore Goll, Nikolaus Demmel, Vladyslav Usenko, Jörg Stuckler, and Daniel Cremers. The tum vi benchmark for evaluating visual-inertial odometry. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1680-1687. IEEE, 2018. 1, 3", + "[40] Aleksandr Segal, Dirk Haehnel, and Sebastian Thrun. Generalized-icp. In Robotics: science and systems, page 435. Seattle, WA, 2009. 4, 5" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Tixiao Shan, Brendan Englot, Drew Meyers, Wei Wang, Carlo Ratti, and Rus Daniela. Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5135-5142. IEEE, 2020. 1", + "[42] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A benchmark for the evaluation of rgb-d slam systems. In Proc. of the International Conference on Intelligent Robot Systems (IROS), 2012. 2", + "[43] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 1", + "[44] Yulun Tian, Yun Chang, Long Quang, Arthur Schang, Carlos Nieto-Granda, Jonathan P How, and Luca Carlone. Resilient and distributed multi-robot visual slam: Datasets, experiments, and lessons learned. arXiv preprint arXiv:2304.04362, 2023. 2", + "[45] Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Louis Wiesmann, Jens Behley, and Cyril Stachniss. Kiss-icp: In defense of point-to-point icp-simple, accurate, and robust registration if done the right way. IEEE Robotics and Automation Letters, 8(2):1029–1036, 2023. 6", + "[46] Chen Wang, Dasong Gao, Kuan Xu, Junyi Geng, Yaoyu Hu, Yuheng Qiu, Bowen Li, Fan Yang, Brady Moon, Abhinav Pandey, Aryan, Jiahe Xu, Tianhao Wu, Haonan He, Daning Huang, Zhongqiang Ren, Shibo Zhao, Taimeng Fu, Pranay Reddy, Xiao Lin, Wenshan Wang, Jingnan Shi, Rajat Talak, Kun Cao, Yi Du, Han Wang, Huai Yu, Shanzhao Wang, Siyu Chen, Ananth Kashyap, Rohan Bandaru, Karthik Dantu, Jiajun Wu, Lihua Xie, Luca Carlone, Marco Hutter, and Sebastian Scherer. PyPose: A library for robot learning with physics-based optimization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6", + "[47] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual SLAM. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 3, 7", + "[48] Wenshan Wang, Yaoyu Hu, and Sebastian Scherer. Tartanvo: A generalizable learning-based vo. In Conference on Robot Learning, pages 1761-1772. PMLR, 2021. 1", + "[49] P. Wenzel, R. Wang, N. Yang, Q. Cheng, Q. Khan, L. von Stumberg, N. Zeller, and D. Cremers. 4Seasons: A cross-season dataset for multi-weather SLAM in autonomous driving. In Proceedings of the German Conference on Pattern Recognition (GCPR), 2020. 2", + "[50] Wei Xu, Yixi Cai, Dongjiao He, Jiarong Lin, and Fu Zhang. FAST-LIO2: fast direct lidar-inertial odometry. IEEE Transactions on Robotics, abs/2107.06829, 2022. 6", + "[51] Wei Xu, Yixi Cai, Dongjiao He, Jiarong Lin, and Fu Zhang. Fast-lio2: Fast direct lidar-inertial odometry. IEEE Transactions on Robotics, 38(4):2053–2073, 2022. 6", + "[52] Lintong Zhang, Marco Camurri, and Maurice Fallon. Multicamera lidar inertial extension to the newer college dataset. arXiv preprint arXiv:2112.08854, 2021. 3", + "[53] Shibo Zhao, Hengrui Zhang, Peng Wang, Lucas Nogueira, and Sebastian Scherer. Super odometry: Imu-centric" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22656", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "lidar-visual-inertial estimator for challenging environments. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8729-8736. IEEE, 2021. 4", + "[54] David Zuniga-Noël, Alberto Jaenal, Ruben Gomez-Ojeda, and Javier Gonzalez-Jimenez. The uma-vi dataset: Visual-inertial odometry in low-textured and dynamic illumination environments. The International Journal of Robotics Research, 39(9):1052-1060, 2020. 1, 3" + ], + "bbox": [ + 78, + 90, + 468, + 218 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22657", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/80ad799c-9412-49c0-a322-3a5e2a9d1a8f_model.json b/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/80ad799c-9412-49c0-a322-3a5e2a9d1a8f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dd8269a714753ac3c6b6d1b1c5f180aa13442fa4 --- /dev/null +++ b/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/80ad799c-9412-49c0-a322-3a5e2a9d1a8f_model.json @@ -0,0 +1,2048 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.13, + 0.852, + 0.176 + ], + "angle": 0, + "content": "SubT-MRS Dataset: Pushing SLAM Towards All-weather Environments https://superodometry.com/datasets" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.204, + 0.897, + 0.258 + ], + "angle": 0, + "content": "Shibo Zhao\\*, Yuanjun Gao, Tianhao Wu, Damanpreet Singh, Rushan Jiang, Haoxiang Sun, Mansi Sarawata, Yuheng Qiu, Warren Whittaker, Ian Higgins, Yi Du, Shaoshu Su, Can Xu, John Keller, Jay Karhade, Lucas Nogueira, Sourojit Saha, Ji Zhang, Wenshan Wang, Chen Wang, Sebastian Scherer" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.262, + 0.688, + 0.28 + ], + "angle": 0, + "content": "\\(^{1}\\)Carnegie Mellon University \\(^{2}\\)University at Buffalo" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.315, + 0.314, + 0.33 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.347, + 0.474, + 0.695 + ], + "angle": 0, + "content": "Simultaneous localization and mapping (SLAM) is a fundamental task for numerous applications such as autonomous navigation and exploration. Despite many SLAM datasets have been released, current SLAM solutions still struggle to have sustained and resilient performance. One major issue is the absence of high-quality datasets including diverse all-weather conditions and a reliable metric for assessing robustness. This limitation significantly restricts the scalability and generalizability of SLAM technologies, impacting their development, validation, and deployment. To address this problem, we present SubT-MRS, an extremely challenging real-world dataset designed to push SLAM towards all-weather environments to pursue the most robust SLAM performance. It contains multi-degraded environments including over 30 diverse scenes such as structureless corridors, varying lighting conditions, and perceptual obscurants like smoke and dust; multimodal sensors such as LiDAR, fisheye camera, IMU, and thermal camera; and multiple locomotions like aerial, legged, and wheeled robots. We developed accuracy and robustness evaluation tracks for SLAM and introduced novel robustness metrics. Comprehensive studies are performed, revealing new observations, challenges, and opportunities for future research." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.719, + 0.21, + 0.735 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.74, + 0.47, + 0.878 + ], + "angle": 0, + "content": "Simultaneous Localization and Mapping (SLAM) is essential in robotics since it provides foundational perception and spatial awareness, and enables machines to understand and interact with the physical world in real-time. Therefore, it has a wide range of applications such as autonomous driving and space exploration. Despite significant advancements in both geometric [5, 34, 41] and data-driven SLAM methods [43, 48], existing solutions remain fragile in challenging conditions. One main reason is that existing SLAM" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.316, + 0.895, + 0.422 + ], + "angle": 0, + "content": "algorithms are often developed and evaluated with datasets from controlled environments [3, 12, 18, 21, 33, 35, 39, 54]. For example, KITTI [18, 21] dataset is mostly collected in sunny weather and EuROC-MAV dataset [3] is collected in small well-lit rooms. These datasets, unfortunately, fail to capture the challenges in real-world scenarios, hindering the development of robust SLAM solutions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.895, + 0.685 + ], + "angle": 0, + "content": "To bridge this gap and push SLAM towards all-weather environments, we present an extremely challenging dataset, SubT-MRS, including scenarios featuring various sensor degradation, aggressive locomotions, and extreme-weather conditions. The SubT-MRS dataset comprises 3 years of data from the DARPA Subterranean (SubT) Challenge [1] (2019-2021) and extends with an additional 2 years of diverse environments (2022-2023), containing mixed indoor and outdoor settings, including long corridors, off-road scenario, tunnels, caves, deserts, forests, and bushlands. Cumulatively, this forms a 5-year dataset encompassing over 500 hours and 100 kilometers of terrain subjected to multimodal sensors including LiDAR, fisheye cameras, depth cameras, thermal cameras, and IMU; heterogeneous platforms including RC cars, legged robots, aerial robots, and wheeled robots; and extreme obscurant conditions such as dense fog, dust, smoke, and heavy snow." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Additionally, we find there is no well-established metric to evaluate the robustness of a SLAM system. Existing evaluation metrics such as absolute trajectory error (ATE) [3] are not representative of the actual performance in robotic applications. We argue that to ensure the safety control of a robot, SLAM system evaluation should not only focus on poses but also its velocity. Taking robot localization as an example, while a momentary pose error spike may only slightly undermine the overall ATE, however, it can lead to a catastrophic crash in an aerial robotic platform. To effectively gauge the performance of a SLAM algorithm, we introduce a new robustness metric to evaluate the reliability of the SLAM system, particularly examining smoothness and accuracy of velocity estimation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.361, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: shiboz@andrew.cmu.edu" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22647" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.894, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.293, + 0.894, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.336, + 0.175, + 0.344 + ], + "angle": 0, + "content": "SubT-Tunnel" + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.336, + 0.301, + 0.344 + ], + "angle": 0, + "content": "SubT-Urban" + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.336, + 0.439, + 0.344 + ], + "angle": 0, + "content": "SubT-Cavern" + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.336, + 0.574, + 0.344 + ], + "angle": 0, + "content": "Multi Floors" + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.336, + 0.73, + 0.345 + ], + "angle": 0, + "content": "Dark Long Corridor" + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.336, + 0.843, + 0.345 + ], + "angle": 0, + "content": "Dense Fog" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.347, + 0.894, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.392, + 0.174, + 0.4 + ], + "angle": 0, + "content": "Laurel Cavern" + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.392, + 0.314, + 0.401 + ], + "angle": 0, + "content": "Textureless Corridor" + }, + { + "type": "image_caption", + "bbox": [ + 0.355, + 0.392, + 0.48, + 0.4 + ], + "angle": 0, + "content": "Darkness Stairs with Jerk Motion" + }, + { + "type": "image_caption", + "bbox": [ + 0.484, + 0.392, + 0.596, + 0.401 + ], + "angle": 0, + "content": "Dense Dust and Smoke" + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.392, + 0.751, + 0.401 + ], + "angle": 0, + "content": "Offroad with Aggressive Motion" + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.392, + 0.875, + 0.401 + ], + "angle": 0, + "content": "Snow with Slipping Motion" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.412, + 0.893, + 0.456 + ], + "angle": 0, + "content": "Figure 1. Dense reconstruction from the SubT-MRS dataset, achieved through collaboration with diverse robots equipped with multimodal sensors. Colors represent different challenging environments (tunnels, caves, urban, confined spaces) captured by various robot types (aerial, legged, wheeled). The bottom section showcases a gallery with diverse visual, LiDAR, and mixed degradations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.469, + 0.557 + ], + "angle": 0, + "content": "Lastly, we perform extensive experiments using proposed degradation datasets to benchmark visual and LiDAR SLAM algorithms. These experiments identify the limitations of the existing SLAM systems and evaluate their robustness across the degradation. Our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.558, + 0.469, + 0.693 + ], + "angle": 0, + "content": "- All-weather Environments To push SLAM toward all-weather environments, we introduce SubT-MRS, an extremely challenging dataset. Spanning five years, it comprises over 500 hours and 100 kilometers of accurately measured trajectories. To the best of our knowledge, SubT-MRS is the first real-world dataset that specifically addresses failure scenarios of SLAM by incorporating a variety of degraded conditions, multiple robotic platforms, and diverse sets of multimodal sensors." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.694, + 0.469, + 0.754 + ], + "angle": 0, + "content": "- Robustness Metric To evaluate the actual performance running on robots, we propose a robustness metric, which, to the best of our knowledge, is the first metric evaluating the reliability, safety, and resilience of SLAM." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.755, + 0.469, + 0.86 + ], + "angle": 0, + "content": "- SLAM Challenge We provide a comprehensive benchmark1 based on the SubT-MRS dataset by conducting ICCV'23 SLAM Challenge. Our evaluation reveals the limitations of state-of-the-art visual and LiDAR SLAM solutions. Furthermore, we conduct extensive experiments to verify robustness metric against various sensor degradation and length of trajectories." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.558, + 0.469, + 0.86 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.479, + 0.642, + 0.495 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.892, + 0.806 + ], + "angle": 0, + "content": "Multimodal Sensors Multimodal sensor datasets are crucial for the development of robust SLAM systems, as single sensor modalities are not comprehensive for all scenarios. Existing datasets typically focus on a limited range of modalities, such as monocular or stereo cameras [3, 49], LiDAR [19], event cameras [29], and depth cameras [42]. The KITTI dataset [19], although covering most modalities, is geared towards on-road scenarios and lacks thermal cameras, limiting its applicability to simple, controlled environments. Conversely, the ViViD++ dataset [23] incorporates thermal images in outdoor settings but is deficient in hardware synchronization, posing challenges for SLAM system development. The Weichen dataset [10] offers thermal images in indoor environments with accurate ground truth poses but is confined to motion capture room settings. In contrast, our SubT-MRS dataset delivers a comprehensive suite of time-synchronized multimodal sensors, including LiDAR, monocular cameras, thermal cameras, depth cameras, and IMU, along with centimeter-level ground truth, catering to a diverse range of research needs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Heterogeneous Platforms Developing a robust SLAM system necessitates versatility in handling different motion patterns and various robotic platforms. However, most existing datasets are tailored for single-robot scenarios. While there are a few multi-robot datasets, they typically involve homogeneous platforms [9, 15, 24, 44], posing challenges" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": "1SubT-MRS dataset is completely real-world data. Simulated data in the benchmark is to ensure evaluation completeness" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22648" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.091, + 0.891, + 0.282 + ], + "angle": 0, + "content": "
DatasetMulti-SpectrumMulti-DegradationMulti-Robot
CameraIMULiDAR/DepthThermalIlluminationSnow SmokeStructurelessSubTAggressive MotionVehicleDroneLeggedHandheld
EuRoC MAV [3]✓✓XX✓XXXXX✓XX
PennCOsYVIO [33]✓✓XXXXXXXXXX✓
TUM VIO [39]✓✓XX✓XXXXXXX✓
UZH-FPV [12]✓✓XXXXXX✓X✓XX
College Dataset [52]✓✓✓XXXXXXXXX✓
RobotCar [28]✓✓✓X✓✓XXX✓XXX
UMA VI [54]✓✓XX✓X✓XXXXX✓
UMich [6]✓✓✓X✓XXXX✓XXX
KITTI [19]✓✓✓XXX✓X✓✓XXX
Virtual KITTI [16]✓✓✓X✓X✓X✓✓XXX
DARPA SubT [37]✓✓✓X✓X✓✓X✓XXX
Hilti SLAM [21]✓✓✓X✓X✓XXXXX✓
M3ED [7]✓✓✓X✓XXXX✓✓✓X
TartanAir [47]✓✓✓X✓✓✓X✓X✓XX
SubT-MRS (Ours)✓✓✓✓✓✓✓✓✓✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.205, + 0.288, + 0.761, + 0.302 + ], + "angle": 0, + "content": "Table 1. Comparison of SLAM datasets on multi-sensors, multi robots, and multi degradation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.329, + 0.468, + 0.42 + ], + "angle": 0, + "content": "in evaluating SLAM performance across varied robotic platforms. The AirMuseum dataset [13] includes drones and ground robots but lacks data from legged robots. As a comparison, SubT-MRS dataset features diverse robotic platforms including legged robots, aerial robots, and wheeled robots, operating in varied environments and sensor setups." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.468, + 0.713 + ], + "angle": 0, + "content": "Extreme Environmental Conditions Developing and testing SLAM systems in extreme environmental conditions is crucial for mitigating potential real-world failures. However, most existing SLAM datasets have been primarily limited to single, controlled environments. The TUM-VI [39] and UMA-VI [54] datasets, being indoor-outdoor visual-inertial datasets, pose challenges due to varying illumination and low-texture environments. The EuRoC MAV [3] and UZH-FPV drone racing [12] datasets, popular in the SLAM community, offer data on aggressive drone movements but usually in consistently lit conditions. The KITTI dataset [19], a staple in autonomous driving research for its outdoor LiDAR-Visual-Inertial data. The Hilti SLAM dataset [21] includes both indoor and outdoor LiDAR-visual-inertial datasets with dynamic lighting and confined spaces. However, both of them focus more on accuracy than on a variety of environmental degradations. TartanAir [47] covers most of the challenging environments but it is a simulation dataset that will pose the sim-to-real gap." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In contrast, the SubT-MRS dataset provides 30 varied scenes, encompassing a wide array of environments. These include SubT tunnels, urban areas, caverns, multifloor structures, dark corridors, and foggy conditions. It also features textureless surfaces, dusty and smoky environments, off-road areas with aggressive motion, and snowy terrains prone to slippage. We include challenging lighting conditions such as extreme darkness and overexposure, along with geometrically challenging scenarios like featureless corridors, staircases, and self-similar cave layouts. Additionally, the dataset incorporates perceptual challenges posed by smoke, fog, dust, and various weather conditions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.327, + 0.687, + 0.343 + ], + "angle": 0, + "content": "3. SubT-MRS Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.892, + 0.407 + ], + "angle": 0, + "content": "We next present the SubT-MRS dataset from three aspects including its collecting environments and settings, ground truth collection, and the new robustness metric. A detailed comparison with other datasets is listed in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.408, + 0.74, + 0.423 + ], + "angle": 0, + "content": "3.1. All-weather Environments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.892, + 0.501 + ], + "angle": 0, + "content": "As mentioned in Sec. 2, the SubT-MRS dataset distinguishes itself through its multimodal sensor setups, heterogeneous robotic platforms, and extreme environmental conditions. This section emphasizes this comprehensive nature pushing SLAM towards all-weather environments." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.505, + 0.704, + 0.519 + ], + "angle": 0, + "content": "3.1.1. Multimodal Sensors" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.597 + ], + "angle": 0, + "content": "Multimodal sensors provide a wealth of information for SLAM systems operating in all-weather environments. This diversity improves scene understanding and strengthens the system's perception ability. Therefore, We incorporate diverse sensors to ensure system robustness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Sensor Pack We embedded 4 Leopard Imaging RGB fisheye cameras, 1 Velodyne puck, 1 Epson M-G365 IMU, 1 FLIR Boson thermal camera, and NVIDIA Jetson AGX Xavier as a sensor pack shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Time Synchronization To ensure the overall consistency of the fused data, we meticulously synchronize the sensors' time using the 'pulse per second (PPS)' technique. As illustrated in Figure 2, the IMU, LiDAR, and thermal camera are directly synchronized with the CPU clock, while the four RGB camera are synchronized using an FPGA board. Consequently, we can effectively manage the time synchronization gap between each pair of sensors not to exceed 3ms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fisheye and Thermal Camera Calibration Camera calibration plays an essential role in the efficiency of a SLAM system. For calibrating the fisheye cameras, we employed the open-source toolkit Kalibr [36], focusing on the intrinsic and extrinsic parameters. Specifically, we use the radial-tangential distortion model to rectify the omnidirectional fisheye camera model. Calibrating thermal cameras, however, poses unique challenges, especially in gathering high-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22649" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.089, + 0.852, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.267, + 0.895, + 0.297 + ], + "angle": 0, + "content": "Figure 2. An overview of the sensor pack used in SubT-MRS dataset. It is equipped with a Xavier processing unit with hardware time synchronization for multimodal sensors including LiDAR, fisheye cameras, thermal cameras, depth cameras (option), and an IMU." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.47, + 0.367 + ], + "angle": 0, + "content": "quality thermal data. To tackle this, we set up a \\(7 \\times 9\\) chessboard, heated by direct sunlight, to generate high-contrast thermal data, ensuring accuracy in the calibration process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.47, + 0.488 + ], + "angle": 0, + "content": "IMU Calibration and Extrinsic Calibration For Lidar-IMU extrinsic calibration, we utilize the CAD model to obtain the calibration parameters. In the case of Camera-IMU extrinsic calibration, we employ the Kalibr toolbox [36] to estimate the extrinsic matrix. To estimate the sensors' bias and the random walk noise of the gyroscope and accelerometer, we collected static data from the IMU and calibrated it using an Allan variance-based IMU tool [14, 17]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.491, + 0.362, + 0.507 + ], + "angle": 0, + "content": "3.1.2. Multi-degraded Environments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.553 + ], + "angle": 0, + "content": "SubT-MRS includes multiple challenging environments including visually degraded environments, geometrically degraded environments, and their combination." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.718 + ], + "angle": 0, + "content": "Visual Degradation Poor-quality visual features can significantly hinder the performance of feature extraction processes and disrupt the accuracy of feature matching. Such issues arise in low-light conditions with inconsistent brightness or image noise introduced by air obscurants. SubT-MRS encompasses these challenges and provides a wide range of visual degradation. This includes environments with limited lighting, such as hospital interiors and caves (Figure 3 A-F), as well as smoky or dusty conditions that cause visual obstruction (Figure 3 M-N), and snowy areas with reduced visibility (Figure 3 H, K, and L)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Geometric Degradation Lack of geometrical features poses significant challenges to LiDAR odometry. The root of this issue often lies in the limited sensing capabilities of LiDAR sensors and the constraints due to their mechanical installation. SubT-MRS captures a variety of environments that exemplify such challenges. It includes long, featureless corridors (Figure 3 E) and staircases (Figure 3 C and G). These scenarios illustrate various forms of geometric degradation, contributing to the improvement of LiDAR odometry systems in challenging conditions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Mixed Degradation A mixed visual and geometric degradation can further hinder the performance of SLAM sys" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.892, + 0.444 + ], + "angle": 0, + "content": "tems. Examples in SubT-MRS include long, dimly lit corridors (Figure 3 A, E, and F), poorly illuminated staircases (Figure 3 C), and environments affected by snowy weather (Figure 3 H, K, and L). In these settings, both LiDAR odometry and visual odometry are prone to failure due to the compounded effects of mixed degradation. The inclusion of such scenarios in SubT-MRS is crucial for evaluating and improving the resilience of multimodal SLAM algorithms." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.447, + 0.796, + 0.463 + ], + "angle": 0, + "content": "3.1.3. Heterogeneous Robot Platforms" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.892, + 0.585 + ], + "angle": 0, + "content": "Most of current datasets focus on single-robot systems, limiting multi-robot SLAM development as shown in Table 1. To address this, we employed diverse robot platforms, including aerial, legged, and wheeled robots, navigating through various environments from urban campuses to medical facilities and natural terrains like caves. This diversity offers a range of scenarios to enhance SLAM algorithms for effective operation in challenging conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.892, + 0.709 + ], + "angle": 0, + "content": "Extrinsic Calibration for Multiple Robots To ensure the multi-robot system shares a common coordinate system, we perform the extrinsic calibration process [38]. It has 2 steps: First, each robot runs Super Odometry [53] to generate its local map using LiDAR data and share it with other robots through a wireless network; Second, the remaining robots identify overlapping regions between their local maps and estimate the extrinsic parameters using GICP [40]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.715, + 0.646, + 0.73 + ], + "angle": 0, + "content": "3.2. Ground Truth" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Ground Truth Map As depicted in Figure 4, we utilized the high-precision FARO Focus 3D S120 3D scanner for creating ground-truth maps. This cutting-edge 3D scanner can measure distances up to \\(120\\mathrm{m}\\), with a maximum measurement rate of \\(976\\mathrm{K}\\) points per second. The accuracy of the ground truth map is noteworthy, maintaining a range error within \\(\\pm 2\\mathrm{mm}\\). We have developed ground truth models for diverse environments, including subterranean areas like urban areas \\((350\\mathrm{m} \\times 350\\mathrm{m})\\), caves \\((150\\mathrm{m} \\times 200\\mathrm{m})\\) and tunnels \\((100 \\times 200\\mathrm{m})\\), as well as indoor-outdoor spaces \\((200\\mathrm{m} \\times 200\\mathrm{m})\\) within the CMU campus. To reduce drift, a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22650" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.891, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.421, + 0.896, + 0.478 + ], + "angle": 0, + "content": "Figure 3. The SubT-MRS datasets were collected across diverse seasons, capturing environments with perceptual challenges such as poor illumination, darkness, and water puddles, where visual sensors falter. They also include geometrically complex areas like long featureless corridors and steep multi-floor structures, challenging LiDAR systems with potential drift. Moreover, these datasets cover conditions with airborne obscurants like dust, fog, snow, and smoke in tough environments, including caves, deserts, long tunnels, and off-road areas." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.492, + 0.852, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.707, + 0.895, + 0.751 + ], + "angle": 0, + "content": "Figure 4. The SubT-MRS dataset facilitates the generation of high-precision ground truth maps and trajectories. Figure (a) shows the ground truth trajectory in multi-floor settings. Figure (b) displays ground truth maps for indoor and outdoor areas, encompassing long corridors, multi-floor structures, and open spaces. Figure (c) features photo-realistic scans based on our ground truth maps in cave environments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.471, + 0.837 + ], + "angle": 0, + "content": "loop closure algorithm was applied to correct the poses. Notably, \\(96\\%\\) of the scans achieved a position uncertainty of less than \\(2\\mathrm{mm}\\). This ground-truth map is crucial for SLAM development where precision is of utmost importance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Ground Truth Trajectory Ground truth trajectories for all sequences are generated based on the ground truth map. In this process, we establish point-to-point, point-to-plane, and point-to-line correspondences[40] between the Ground" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.775, + 0.895, + 0.898 + ], + "angle": 0, + "content": "Truth map and the current LiDAR scan. Our estimated ground truth trajectory is a comprehensive fusion of various constraints derived from the ground truth map, visual odometry, LiDAR odometry, and IMU measurements. This approach ensures the robustness of our solution across diverse environmental conditions. To verify accuracy, we reconstruct LiDAR maps from ground truth trajectory and compare it with our ground truth maps." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22651" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.09, + 0.872, + 0.104 + ], + "angle": 0, + "content": "Table 2. SLAM Challenge Results (Blue shadings indicate rankings of ATE and Robustness Metric; L: LiDAR I: IMU C: Camera)" + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.115, + 0.905, + 0.261 + ], + "angle": 0, + "content": "
#TeamMethodOdometry TypeDeviceRealTime (s)CPU/GPU (%)RAM (GB)ATE↓\\( {R}_{v} \\uparrow /{R}_{w} \\uparrow \\)Sensors L I C
1Liu et alFAST-LIO2 [51], HBA [27]FilterIntel i7-9700K51.31098.667 / 04.0520.5880.517/0.770✓✓
2Yibin et alLIO-EKF [45]FilterIntel i7-107000.00652.167 / 00.0724.3130.441/0.574✓✓
3Weitong et alFAST-LIO[2], Pose Graph[11]FilterIntel Xeon(R)E3-1240v50.12522.63 / 04.3050.6630.473/0.747✓✓
4Kim et alFAST-LIO2[50], Point-LIO[20], Quatro[25]FilterIntel i5-125000.268101.108 / 055.643.8250.479/0.615✓✓
5Zhong et alDLO[8], Scan-Context++[22]SW OptAMD Ryzen 9 5900x0.02713.289 / 01.1741.2090.276/0.486✓✓
1Peng et alDVI-SLAM [32]LearningIntel i9-12900183.233- / 14911 (4)0.5470.473/0.788✓✓
2Jiang et alLET-NET[26], VINS-Mono[34]HybridIntel i5-94000.06440.35 / 04.3371.0930.078/0.322✓✓
3Thien et alVR-SLAM[31]SW OptIntel i9-129000.142176.44 / 09.1113.0370.083/0.372✓✓
4Li et alORB-SLAM3[4]SW Opt.Intel i7-107000.01965.028 / 00.3868.9750.163/0.474✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.274, + 0.877, + 0.289 + ], + "angle": 0, + "content": "Table 3. Accuracy Performance on Geometric Degradation. Red numbers represent ATE ranking. * denotes incomplete submissions." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.3, + 0.885, + 0.413 + ], + "angle": 0, + "content": "
TeamGeometric Degradation (Real World)SimulationMix DegradationAverage
UrbanTunnelCaveNuclear_1Nuclear_2Laurel_CavernsFactoryOceanSewerageLong CorridorMulti FloorBlock Lidar
Liu et al10.3070.0950.6290.1220.2350.2600.8890.7570.9781.4540.4010.9340.588
Weitong et al20.260.0960.6170.1200.2220.4020.9980.7701.5861.2540.5771.0560.663
Kim et al30.3310.0920.7870.1230.2700.27910.62822.4257.1472.1000.6501.0683.825
Yibin et al41.0600.2200.7500.4700.6209.1404.9200.28024.4602.9905.5001.3404.312
Zhong et al51.2050.695-1.1751.722.080.8890.7781.13---1.209*
Average0.6330.2400.6960.4020.61342.4323.6655.0027.0601.9501.7821.099
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.437, + 0.258, + 0.451 + ], + "angle": 0, + "content": "3.3. Robustness Metric" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.473, + 0.681 + ], + "angle": 0, + "content": "As discussed in Sec. 1, existing evaluation metrics such as absolute trajectory error (ATE) [3] have limitations in evaluating the SLAM's robustness in real-world applications. ATE primarily focuses on the accuracy of trajectory and does not consider the completeness of trajectory (recall). Also, it can not effectively capture velocity changes that have a direct impact on the robot's safety. For example, a pose spike error might not be immediately noticeable in ATE evaluation, it may result in a crash in an aerial robot control. To address these gaps, we introduce a new robustness metric based on estimated velocity. That is because velocity estimation from SLAM is crucial for the robot's control system which directly impacts stability, affecting the robot's safety. The new robustness metric is the area under the curve (AUC) of the F-1 score:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.692, + 0.47, + 0.725 + ], + "angle": 0, + "content": "\\[\nF _ {1} (e) = \\frac {2 P (e < T) R (e < T)}{P (e < T) + R (e < T)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "where the precision \\( P \\) quantifies the precision of the estimated velocity as a percentage: how closely the estimated velocity points lie to the ground truth point and the recall rate \\( R \\) quantifies the estimated velocity's completeness: to what extent all the ground-truth points are covered. A high F-1 score can only be achieved by the velocity estimation that is both accurate and complete throughout the entire run. Specifically, an estimated error \\( e \\) is regarded as an inlier, if it is smaller than a threshold \\( T \\). To scale the threshold \\( T \\) within the range of [0, 1], we apply exponential mapping \\( \\exp(-10T) \\) when calculating robustness metric." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.619 + ], + "angle": 0, + "content": "Position Robustness \\( R_{p} \\) & Rotation Robustness \\( R_{r} \\) The AUC of F-1 score can be defined using both linear velocity and angular velocity, which can reflect the robustness of position and rotation estimation, respectively. Specifically, we define \\( R_{p} = \\mathrm{AUC}(F_{1}(\\mathrm{v}_{\\mathrm{e}})) \\) and \\( R_{r} = \\mathrm{AUC}(F_{1}(\\omega_{\\mathrm{e}})) \\), where \\( v_{e} \\) and \\( \\omega_{e} \\) are the estimated error of linear velocity and rotation velocity, respectively. Note that not all SLAM solutions can output the velocity at the desired frequency. To address this, we use B-splines [30, 46] to obtain smooth trajectories, compute derivatives for smooth trajectories, and derive estimated linear velocity and angular velocity. For more details, please refer to the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.629, + 0.787, + 0.646 + ], + "angle": 0, + "content": "4. Open SLAM Challenge Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.892, + 0.68 + ], + "angle": 0, + "content": "In this section, we will present our works from two perspectives: accuracy evaluation and robustness evaluation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.687, + 0.698, + 0.702 + ], + "angle": 0, + "content": "4.1. Accuracy Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The results of the ICCV SLAM challenge underscore the necessity for advancements in system robustness. From 29 submissions, we identified 5 winners in the LiDAR category and 4 in the visual category. However, in the sensor fusion track, which addresses both visual and geometric degradation, no submissions met the criteria for success. This result reveals the existing SLAM systems still have lots of space to improve. Sequence characteristics will be detailed in the supplementary material. Table 2 shows the results for both the LiDAR and visual tracks. Unfortunately, there are no current solutions that can balance high accuracy and real-time performance in challenging environments. Since several environments lack geometric features or visual features," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22652" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.104, + 0.09, + 0.864, + 0.104 + ], + "angle": 0, + "content": "Table 4. Accuracy Performance on Visual Degradation. Red numbers represent ATE ranking. * denotes incomplete submissions." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.115, + 0.847, + 0.214 + ], + "angle": 0, + "content": "
TeamVisual Degradation (Real World)SimulationAverage
Lowlight 1Lowlight 2Over ExposureFlash LightSmoke RoomOutdoor NightEnd of WorldMoonWestern Desert
Peng et al11.0631.6370.5030.440.1530.8270.0380.1950.0700.547
Thien et al21.0812.0541.7331.05410.5327.6920.7531.2281.2093.037
Jiang et al31.0191.1261.9112.3413.75711.8212.1540.6044.0103.193
Li et al45.7687.8341.7571.2955.37010.766-30.07-8.98*
Average2.2323.1631.4761.2824.9537.7760.9828.0241.763
" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.226, + 0.318, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.227, + 0.483, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.227, + 0.659, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.227, + 0.832, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.372, + 0.895, + 0.415 + ], + "angle": 0, + "content": "Figure 5. From left to right, it shows robustness metric \\( R_{p} \\) and \\( R_{r} \\) for LiDAR and visual sequences respectively. Note: This is a summary of results for all sequences, with weights based on the trajectory length. The area under the curve (AUC) represents the robustness \\( (R_{p}, R_{r}) \\). The x-axis shows velocity thresholds for classifying estimated velocities as inliers and the y-axis is F-1 score." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.471, + 0.536 + ], + "angle": 0, + "content": "the algorithms required more processing times to achieve reasonably high accuracy. Liu et al. at the University of Hong Kong, who achieved the highest accuracy, recorded an Absolute Trajectory Error (ATE) of \\(0.588\\mathrm{m}\\), but each iteration took 51.3 seconds. In the visual track, Peng et al. from the Samsung Research Center attained a leading ATE of \\(0.547\\mathrm{m}\\), but each iteration required 183 seconds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.765 + ], + "angle": 0, + "content": "LiDAR Track Discussion To enhance robustness in all-weather environments, a pivotal question arises: what are the limits of current LiDAR SLAM algorithms? Table 3 presents a summary of the ATE/RPE errors observed in real-world geometrically degraded, simulated, and mixed degradation scenarios across the top five teams. In real-world geometrically degraded environments, we observed the first limitation: existing LiDAR solutions struggle in confined spaces such as caves. Almost all algorithms perform worse in Cave (Figure 3S) and Laureal_Cavern environments (Figure 3N) with average ATE 0.696 and 2.432 meters respectively. One reason is that cave environments is the most confined space which may lack sufficient geometric features, compared with other environments like tunnel (Figure 3J) and Urban (Figure 3B) and Nuclear scenes (Figure 3U)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In simulated environments, we encountered the second limitation: existing methods are tailored for fixed motion patterns like vehicles and struggle with unpredictable motion patterns. Since our simulation sequence is derived from TartanAir [47], featuring aggressive and random motion patterns, it might not follow the usual velocity distribution domain and has not been thoroughly tested. As shown in Table 3 (simulation), it resulted in significantly higher ATE errors \\((5.239\\mathrm{m})\\) compared to real-world's \\((0.836\\mathrm{m})\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.702 + ], + "angle": 0, + "content": "In mixed degradation environments, the third limitation we found is that existing methods cannot actively select the most informative measurements to adapt to new environments. For example, in the Long Corridor sequence (Figure 3 A) with low lighting and a featureless environment, the scenario is intricately designed for collaborative usage of LiDAR and Visual sensors. Algorithms are expected to utilize visual information in geometrically degraded environments while disregarding it in darkness. Similarly, the Block LiDAR sequences (Figure 3 V) aim to simulate sensor drop scenarios frequently encountered in robotic applications. These sequences involve alternating periods of LiDAR and Visual data loss, challenging algorithms to promptly detect sensor failures and switch to the other modality for SLAM. However, the average ATE error \\((1.61\\mathrm{m})\\) in Table 3 (mixed degradation) is much higher than in geometric degradation scenarios \\((0.836\\mathrm{m})\\), suggesting the above limitations. Details are in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Visual Track Discussion To improve the robustness, a pivotal question arises: what are the limits of current visual SLAM? To what extent does image quality impact a visual SLAM system? Table 4 shows a summary of ATE errors from real-world and simulation scenarios from the awarded 4 teams. In real-world scenarios, the first limitation is the lack of anti-noise capability in current methods, especially in low-quality image settings. We assessed visual odometry accuracy in visually degraded environments, including low lighting, sunlight overexposure, flashing lights, smoke-filled rooms, and nighttime outdoor settings (Figure 3). Trajectories in conditions like the Smoke Room and Outdoor Night exhibit significantly higher average ATE er" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22653" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.09, + 0.892, + 0.104 + ], + "angle": 0, + "content": "Table 5. Robustness Performance on Geometric Degradation. Red numbers are robustness ranking. Larger values indicate better robustness." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.112, + 0.855, + 0.279 + ], + "angle": 0, + "content": "
TeamGeometric Degradation (Real World)SimulationMix DegradationAverage
UrbanTunnelCaveNuclear_1Nuclear_2Laurel_CavernsFactoryOceanSewerageLong CorridorMulti FloorBlock Lidar
Liu et al1Rp↑0.8110.8650.7360.7470.5040.8160.1570.1350.1350.3960.5290.3710.516
Weitong et al20.7730.8380.6900.8710.7390.1520.1530.1490.0810.4070.4560.3450.471
Kim et al40.7470.8650.7370.8700.6890.7770.0140.0010.0100.4100.2850.3450.479
Yibin et al30.6500.7830.5570.7210.4900.4810.2190.5600.0190.1750.2600.3700.440
Zhong et al50.5670.6830.2040.6800.4260.4470.1030.0840.1220000.276
Liu et al1Rr↑0.8880.8930.8160.8570.7780.8610.6920.7190.6520.7530.6890.6430.770
Weitong et al20.8860.8920.8110.8930.8400.5980.6880.7200.5140.7660.7060.6440.746
Kim et al40.6800.8930.8160.8930.8030.8080.1980.1250.3000.8020.4220.6420.615
Yibin et al30.6240.7070.4620.7330.5400.4670.8620.9620.4960.4770.3420.2130.573
Zhong et al50.7310.7450.4370.7570.6030.5230.6880.7040.6430000.485
AverageRp0.7100.8070.5850.7780.5700.5350.1300.1860.0730.2780.3060.286
Average Rr0.7620.8260.6680.8270.7130.6510.6260.6460.5210.5600.4320.428
" + }, + { + "type": "table_caption", + "bbox": [ + 0.084, + 0.292, + 0.885, + 0.306 + ], + "angle": 0, + "content": "Table 6. Robustness Performance on Visual Degradation. Red numbers are robustness ranking. Larger values indicate better robustness." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.313, + 0.855, + 0.472 + ], + "angle": 0, + "content": "
TeamVisual Degradation (Real World)SimulationAverage
Lowlight 1Lowlight 2Over ExposureFlash LightSmoke RoomOutdoor NightEnd of WorldMoonWestern Desert
Peng et al10.3570.2270.2640.2030.5360.2700.6990.8930.8060.472
Thien et al30.0450.0700.2400.1560.1310.07500.03100.083
Jiang et al4Rp↑0.0460.0390.1940.0880.2420.0950000.078
Li et al20.3420.1870.2570.2080.3220.14200.00600.162
Peng et al10.6410.5810.7440.6500.8780.6700.9750.9750.9740.787
Thien et al30.4130.4450.6100.2690.4740.4870.1770.31500.354
Jiang et al4Rr↑0.4530.4520.6190.2520.5420.5770.00200.1570.339
Li et al20.6420.5740.6570.6510.7730.6600.00.30500.473
AverageRp0.1980.1310.2390.1640.3080.1460.1750.2320.202
Average Rr0.5370.5130.6580.4560.6670.5980.2880.3990.283
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.47, + 0.608 + ], + "angle": 0, + "content": "rors (4.95m, 7.76m respectively). Smoke and night scenes challenge feature extraction and tracking and introduce sensor noise, emphasizing the need for robust anti-noise algorithms. The second limitation is that existing methods struggle to overcome aggressive motion. Even in simulations with relatively good image quality, most methods still show significant average ATE errors (8.024m) on the Moon sequence (Figure 3 Y). See more details in supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.612, + 0.29, + 0.626 + ], + "angle": 0, + "content": "4.2. Robustness Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.795 + ], + "angle": 0, + "content": "The F-1 robustness curve provides a detailed evaluation under different error tolerances, while \\( R_{p} \\) and \\( R_{r} \\) provide an overall measure of robustness that is not dependent on a specific decision threshold. This suggests the flexibility of our new robustness metric. Figure 5 clearly shows the robustness performance of all teams regarding position and rotation, highlighting the differences in robustness between the methods. The values of \\( R_{p} \\) and \\( R_{r} \\), which are the AUC under F-1 curves, are displayed in the brackets of the legends. It reveals that Liu et al. and Peng et al. are the most robust solutions for LiDAR and visual tracks, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Is the robustness metric robust? We did extensive experiments on our robustness metric against diverse sensor degradation settings. Table 5 and 6 present a summary of robustness performance in real-world, simulated, and mixed degradation scenarios for both LiDAR and visual sequences. In Table 5, we observe that the average values of \\( R_{p} \\) and \\( R_{r} \\) for the mixed degradation environment (0.290," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.892, + 0.577 + ], + "angle": 0, + "content": "0.473) are significantly lower than geometric degradation environments (0.664, 0.754). This observation suggests that the majority of SLAM algorithms exhibit reduced robustness in mixed degradation environments as compared to geometrically degraded ones. This aligns with our expectations and verifies the effectiveness of the robustness metric." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Our robustness ranking, indicated by small red numbers, differs from those on ATE ranking shown in Table 3 and 4. This is because our F-1 score-based metrics \\( R_{p} \\) and \\( R_{r} \\) jointly consider precision and recall rate to provide a balanced evaluation of SLAM performance, considering both metrics across the full spectrum of thresholds. In contrast, the ATE focuses solely on precision, neglecting the trajectory's completeness (recall). We also evaluate our metric on various synthetic trajectories in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.619, + 0.755 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduce SubT-MRS, a comprehensive SLAM dataset with various sensor data, locomotion patterns, and over 30 degradation types in simulation and real-world settings to push SLAM towards all-weather environments. Additionally, we introduce a new robustness metric to evaluate the reliability of SLAM systems, enhancing robot safety control. 29 teams have tested SubT-MRS in the organized SLAM challenge and we expect that it will serve as a critical benchmark for future SLAM development." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22654" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.091, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.11, + 0.422, + 0.138 + ], + "angle": 0, + "content": "[1] https://www.darpa.mil/program/darpa-subterranean-challenge. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.14, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[2] Chunge Bai, Tao Xiao, Yajie Chen, Haoqian Wang, Fang Zhang, and Xiang Gao. Faster-lio: Lightweight tightly coupled lidar-inertial odometry using parallel sparse incremental voxels. IEEE Robotics and Automation Letters, 7(2):4861-4868, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.21, + 0.471, + 0.278 + ], + "angle": 0, + "content": "[3] Michael Burri, Janosch Nikolic, Pascal Gohl, Thomas Schneider, Joern Rehder, Sammy Omari, Markus W Achtelik, and Roland Siegwart. The EuRoC micro aerial vehicle datasets. The International Journal of Robotics Research, 2016. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.281, + 0.471, + 0.348 + ], + "angle": 0, + "content": "[4] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.351, + 0.469, + 0.419 + ], + "angle": 0, + "content": "[5] Carlos Campos, Richard Elvira, Juan J. Gómez Rodríguez, José M. M. Montiel, and Juan D. Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6): 1874-1890, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.421, + 0.469, + 0.476 + ], + "angle": 0, + "content": "[6] Nicholas Carlevaris-Bianco, Arash K Ushani, and Ryan M Eustice. University of michigan north campus long-term vision and lidar dataset. The International Journal of Robotics Research, 35(9):1023-1035, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.478, + 0.469, + 0.561 + ], + "angle": 0, + "content": "[7] Kenneth Chaney, Fernando Cladera, Ziyun Wang, Anthony Bisulco, M Ani Hsieh, Christopher Korpela, Vijay Kumar, Camillo J Taylor, and Kostas Daniilidis. M3ed: Multi-robot, multi-sensor, multi-environment event dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015–4022, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.563, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[8] Kenny Chen, Brett T. Lopez, Ali-akbar Agha-mohammadi, and Ankur Mehta. Direct lidar odometry: Fast localization with dense point clouds. IEEE Robotics and Automation Letters, 7(2):2000–2007, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.619, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[9] Long Chen, Yuchen Li, Chao Huang, Bai Li, Yang Xing, Daxin Tian, Li Li, Zhongxu Hu, Xiaoxiang Na, Zixuan Li, et al. Milestones in autonomous driving and intelligent vehicles: Survey of surveys. IEEE Transactions on Intelligent Vehicles, 8(2):1046-1056, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[10] Weichen Dai, Yu Zhang, Shenzhou Chen, Donglei Sun, and Da Kong. A multi-spectral dataset for evaluating motion estimation systems. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 5560-5566. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[11] Frank Dellaert, Michael Kaess, et al. Factor graphs for robot perception. Foundations and Trends® in Robotics, 6(1-2): 1-139, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[12] Jeffrey Delmerico, Titus Cieslewski, Henri Rebecq, Matthias Faessler, and Davide Scaramuzza. Are we ready for autonomous drone racing? the UZH-FPV drone racing dataset. In IEEE Int. Conf. Robot. Autom. (ICRA), 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Rodolphe Dubois, Alexandre Eudes, and Vincent Frémont. Airmuseum: a heterogeneous multi-robot dataset for stereovisual and inertial simultaneous localization and mapping. In" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.11, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "2020 IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI), pages 166-172, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[14] Naser El-Sheimi, Haiying Hou, and Xiaoji Niu. Analysis and modeling of inertial sensors using allan variance. IEEE Transactions on Instrumentation and Measurement, 57(1): 140-149, 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[15] Dapeng Feng, Yuhua Qi, Shipeng Zhong, Zhiqiang Chen, Yudu Jiao, Qiming Chen, Tao Jiang, and Hongbo Chen. S3e: A large-scale multimodal dataset for collaborative slam. arXiv preprint arXiv:2210.13723, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[16] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4340-4349, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[17] Wenliang Gao. A ros package tool to analyze the imu performance, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI Vision Benchmark Suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[19] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[20] Dongjiao He, Wei Xu, Nan Chen, Fanze Kong, Chongjian Yuan, and Fu Zhang. Point-lio: Robust high-bandwidth light detection and ranging inertial odometry. Advanced Intelligent Systems, page 2200459, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.506, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[21] Michael Helmberger, Kristian Morin, Beda Berner, Nitish Kumar, Giovanni Cioffi, and Davide Scaramuzza. The hilti slam challenge dataset. IEEE Robotics and Automation Letters, 7(3):7518-7525, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.563, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[22] Giseop Kim and Ayoung Kim. Scan context: Egocentric spatial descriptor for place recognition within 3d point cloud map. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4802-4809. IEEE, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[23] Alex Junho Lee, Younggun Cho, Young-sik Shin, Ayoung Kim, and Hyun Myung. Vivid++: Vision for visibility dataset. IEEE Robotics and Automation Letters, 7(3):6282-6289, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[24] Keith YK Leung, Yoni Halpern, Timothy D Barfoot, and Hugh HT Liu. The utias multi-robot cooperative localization and mapping dataset. The International Journal of Robotics Research, 30(8):969–974, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[25] Hyungtae Lim, Suyong Yeon, Suyong Ryu, Yonghan Lee, Youngji Kim, Jaeseong Yun, Euigon Jung, Donghwan Lee, and Hyun Myung. A single correspondence is enough: Robust global registration to avoid degeneracy in urban environments. page Accepted. To appear, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[26] Yicheng Lin, Shuo Wang, Yunlong Jiang, and Bin Han. Breaking of brightness consistency in optical flow with a lightweight cnn network, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Zheng Liu, Xiyuan Liu, and Fu Zhang. Efficient and consistent bundle adjustment on lidar point clouds. IEEE Transactions on Robotics, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22655" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[28] Will Maddern, Geoffrey Pascoe, Chris Linegar, and Paul Newman. 1 year, \\(1000\\mathrm{km}\\): The oxford robotcar dataset. The International Journal of Robotics Research, 36(1):3-15, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[29] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. The International Journal of Robotics Research, 36(2):142-149, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[30] Elias Mueggler, Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. Continuous-time visual-inertial odometry for event cameras. IEEE Transactions on Robotics, 34(6): 1425-1440, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.327 + ], + "angle": 0, + "content": "[31] Thien Hoang Nguyen, Shenghai Yuan, and Lihua Xie. Vrslam: A visual-range simultaneous localization and mapping system using monocular camera and ultra-wideband sensors, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[32] Xiongfeng Peng, Zhihua Liu, Weiming Li, Ping Tan, SoonYong Cho, and Qiang Wang. Dvi-slam: A dual visual inertial slam network, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.439 + ], + "angle": 0, + "content": "[33] Bernd Pfrommer, Nitin Sanket, Kostas Daniilidis, and Jonas Cleveland. Penncosyvio: A challenging visual inertial odometry benchmark. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 3847-3854. IEEE, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[34] Tong Qin, Peiliang Li, and Shaojie Shen. Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics, 34(4):1004-1020, 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[35] Milad Ramezani, Yiduo Wang, Marco Camurri, David Wisth, Matias Mattamala, and Maurice Fallon. The newer college dataset: Handheld lidar, inertial and vision with ground truth. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4353-4360. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[36] Joern Rehder, Janosch Nikolic, Thomas Schneider, Timo Hinzmann, and Roland Siegwart. Extending kalibr: Calibrating the extrinsic of multiple imus and of individual axes. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 4304-4311. IEEE, 2016. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.719 + ], + "angle": 0, + "content": "[37] John G Rogers, Arthur Schang, Carlos Nieto-Granda, John Ware, John Carter, Jonathan Fink, and Ethan Stump. The darpa subt urban circuit mapping dataset and evaluation metric. In Experimental Robotics: The 17th International Symposium, pages 391–401. Springer, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[38] Sebastian Scherer, Vasu Agrawal, Graeme Best, Chao Cao, Katarina Cujic, R Darnley, R DeBortoli, E Dexheimer, B Drozd, R Garg, et al. Resilient and modular subterranean exploration with a team of roving and flying robots. Field Robotics, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[39] David Schubert, Thore Goll, Nikolaus Demmel, Vladyslav Usenko, Jörg Stuckler, and Daniel Cremers. The tum vi benchmark for evaluating visual-inertial odometry. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1680-1687. IEEE, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[40] Aleksandr Segal, Dirk Haehnel, and Sebastian Thrun. Generalized-icp. In Robotics: science and systems, page 435. Seattle, WA, 2009. 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[41] Tixiao Shan, Brendan Englot, Drew Meyers, Wei Wang, Carlo Ratti, and Rus Daniela. Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5135-5142. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[42] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A benchmark for the evaluation of rgb-d slam systems. In Proc. of the International Conference on Intelligent Robot Systems (IROS), 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[43] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[44] Yulun Tian, Yun Chang, Long Quang, Arthur Schang, Carlos Nieto-Granda, Jonathan P How, and Luca Carlone. Resilient and distributed multi-robot visual slam: Datasets, experiments, and lessons learned. arXiv preprint arXiv:2304.04362, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[45] Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Louis Wiesmann, Jens Behley, and Cyril Stachniss. Kiss-icp: In defense of point-to-point icp-simple, accurate, and robust registration if done the right way. IEEE Robotics and Automation Letters, 8(2):1029–1036, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[46] Chen Wang, Dasong Gao, Kuan Xu, Junyi Geng, Yaoyu Hu, Yuheng Qiu, Bowen Li, Fan Yang, Brady Moon, Abhinav Pandey, Aryan, Jiahe Xu, Tianhao Wu, Haonan He, Daning Huang, Zhongqiang Ren, Shibo Zhao, Taimeng Fu, Pranay Reddy, Xiao Lin, Wenshan Wang, Jingnan Shi, Rajat Talak, Kun Cao, Yi Du, Han Wang, Huai Yu, Shanzhao Wang, Siyu Chen, Ananth Kashyap, Rohan Bandaru, Karthik Dantu, Jiajun Wu, Lihua Xie, Luca Carlone, Marco Hutter, and Sebastian Scherer. PyPose: A library for robot learning with physics-based optimization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[47] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual SLAM. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[48] Wenshan Wang, Yaoyu Hu, and Sebastian Scherer. Tartanvo: A generalizable learning-based vo. In Conference on Robot Learning, pages 1761-1772. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[49] P. Wenzel, R. Wang, N. Yang, Q. Cheng, Q. Khan, L. von Stumberg, N. Zeller, and D. Cremers. 4Seasons: A cross-season dataset for multi-weather SLAM in autonomous driving. In Proceedings of the German Conference on Pattern Recognition (GCPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[50] Wei Xu, Yixi Cai, Dongjiao He, Jiarong Lin, and Fu Zhang. FAST-LIO2: fast direct lidar-inertial odometry. IEEE Transactions on Robotics, abs/2107.06829, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[51] Wei Xu, Yixi Cai, Dongjiao He, Jiarong Lin, and Fu Zhang. Fast-lio2: Fast direct lidar-inertial odometry. IEEE Transactions on Robotics, 38(4):2053–2073, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[52] Lintong Zhang, Marco Camurri, and Maurice Fallon. Multicamera lidar inertial extension to the newer college dataset. arXiv preprint arXiv:2112.08854, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[53] Shibo Zhao, Hengrui Zhang, Peng Wang, Lucas Nogueira, and Sebastian Scherer. Super odometry: Imu-centric" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22656" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "lidar-visual-inertial estimator for challenging environments. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8729-8736. IEEE, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[54] David Zuniga-Noël, Alberto Jaenal, Ruben Gomez-Ojeda, and Javier Gonzalez-Jimenez. The uma-vi dataset: Visual-inertial odometry in low-textured and dynamic illumination environments. The International Journal of Robotics Research, 39(9):1052-1060, 2020. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22657" + } + ] +] \ No newline at end of file diff --git a/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/80ad799c-9412-49c0-a322-3a5e2a9d1a8f_origin.pdf b/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/80ad799c-9412-49c0-a322-3a5e2a9d1a8f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ed28476d479c6d78067732a3b2331ef1bf8d2fac --- /dev/null +++ b/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/80ad799c-9412-49c0-a322-3a5e2a9d1a8f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2e2ca39b1212cdf2f8344c703f529f689000dd620f0e1a806abab261cd57466 +size 2825928 diff --git a/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/full.md b/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/full.md new file mode 100644 index 0000000000000000000000000000000000000000..243c8c2a06e0e5e814055f8107e7af4f69638bbc --- /dev/null +++ b/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/full.md @@ -0,0 +1,257 @@ +# SubT-MRS Dataset: Pushing SLAM Towards All-weather Environments https://superodometry.com/datasets + +Shibo Zhao\*, Yuanjun Gao, Tianhao Wu, Damanpreet Singh, Rushan Jiang, Haoxiang Sun, Mansi Sarawata, Yuheng Qiu, Warren Whittaker, Ian Higgins, Yi Du, Shaoshu Su, Can Xu, John Keller, Jay Karhade, Lucas Nogueira, Sourojit Saha, Ji Zhang, Wenshan Wang, Chen Wang, Sebastian Scherer + +$^{1}$ Carnegie Mellon University $^{2}$ University at Buffalo + +# Abstract + +Simultaneous localization and mapping (SLAM) is a fundamental task for numerous applications such as autonomous navigation and exploration. Despite many SLAM datasets have been released, current SLAM solutions still struggle to have sustained and resilient performance. One major issue is the absence of high-quality datasets including diverse all-weather conditions and a reliable metric for assessing robustness. This limitation significantly restricts the scalability and generalizability of SLAM technologies, impacting their development, validation, and deployment. To address this problem, we present SubT-MRS, an extremely challenging real-world dataset designed to push SLAM towards all-weather environments to pursue the most robust SLAM performance. It contains multi-degraded environments including over 30 diverse scenes such as structureless corridors, varying lighting conditions, and perceptual obscurants like smoke and dust; multimodal sensors such as LiDAR, fisheye camera, IMU, and thermal camera; and multiple locomotions like aerial, legged, and wheeled robots. We developed accuracy and robustness evaluation tracks for SLAM and introduced novel robustness metrics. Comprehensive studies are performed, revealing new observations, challenges, and opportunities for future research. + +# 1. Introduction + +Simultaneous Localization and Mapping (SLAM) is essential in robotics since it provides foundational perception and spatial awareness, and enables machines to understand and interact with the physical world in real-time. Therefore, it has a wide range of applications such as autonomous driving and space exploration. Despite significant advancements in both geometric [5, 34, 41] and data-driven SLAM methods [43, 48], existing solutions remain fragile in challenging conditions. One main reason is that existing SLAM + +algorithms are often developed and evaluated with datasets from controlled environments [3, 12, 18, 21, 33, 35, 39, 54]. For example, KITTI [18, 21] dataset is mostly collected in sunny weather and EuROC-MAV dataset [3] is collected in small well-lit rooms. These datasets, unfortunately, fail to capture the challenges in real-world scenarios, hindering the development of robust SLAM solutions. + +To bridge this gap and push SLAM towards all-weather environments, we present an extremely challenging dataset, SubT-MRS, including scenarios featuring various sensor degradation, aggressive locomotions, and extreme-weather conditions. The SubT-MRS dataset comprises 3 years of data from the DARPA Subterranean (SubT) Challenge [1] (2019-2021) and extends with an additional 2 years of diverse environments (2022-2023), containing mixed indoor and outdoor settings, including long corridors, off-road scenario, tunnels, caves, deserts, forests, and bushlands. Cumulatively, this forms a 5-year dataset encompassing over 500 hours and 100 kilometers of terrain subjected to multimodal sensors including LiDAR, fisheye cameras, depth cameras, thermal cameras, and IMU; heterogeneous platforms including RC cars, legged robots, aerial robots, and wheeled robots; and extreme obscurant conditions such as dense fog, dust, smoke, and heavy snow. + +Additionally, we find there is no well-established metric to evaluate the robustness of a SLAM system. Existing evaluation metrics such as absolute trajectory error (ATE) [3] are not representative of the actual performance in robotic applications. We argue that to ensure the safety control of a robot, SLAM system evaluation should not only focus on poses but also its velocity. Taking robot localization as an example, while a momentary pose error spike may only slightly undermine the overall ATE, however, it can lead to a catastrophic crash in an aerial robotic platform. To effectively gauge the performance of a SLAM algorithm, we introduce a new robustness metric to evaluate the reliability of the SLAM system, particularly examining smoothness and accuracy of velocity estimation. + +![](images/b3d861d86989a475f5b89f18a2bceb82e3f55d46da24ddf4bfe5e6348b1bff74.jpg) + +![](images/3e2278a11c657ec21d4a1b1d9bd38063d647760e8f90618979c79422d7d0a8b0.jpg) +SubT-Tunnel +SubT-Urban +SubT-Cavern +Multi Floors + +![](images/4ff2e7e8ea5727e51f0d28853e2951db92306ff0c8c78c86400d2dbee3f76ece.jpg) +Dark Long Corridor +Dense Fog +Laurel Cavern +Textureless Corridor +Darkness Stairs with Jerk Motion +Dense Dust and Smoke +Offroad with Aggressive Motion +Snow with Slipping Motion +Figure 1. Dense reconstruction from the SubT-MRS dataset, achieved through collaboration with diverse robots equipped with multimodal sensors. Colors represent different challenging environments (tunnels, caves, urban, confined spaces) captured by various robot types (aerial, legged, wheeled). The bottom section showcases a gallery with diverse visual, LiDAR, and mixed degradations. + +Lastly, we perform extensive experiments using proposed degradation datasets to benchmark visual and LiDAR SLAM algorithms. These experiments identify the limitations of the existing SLAM systems and evaluate their robustness across the degradation. Our contributions include: + +- All-weather Environments To push SLAM toward all-weather environments, we introduce SubT-MRS, an extremely challenging dataset. Spanning five years, it comprises over 500 hours and 100 kilometers of accurately measured trajectories. To the best of our knowledge, SubT-MRS is the first real-world dataset that specifically addresses failure scenarios of SLAM by incorporating a variety of degraded conditions, multiple robotic platforms, and diverse sets of multimodal sensors. +- Robustness Metric To evaluate the actual performance running on robots, we propose a robustness metric, which, to the best of our knowledge, is the first metric evaluating the reliability, safety, and resilience of SLAM. +- SLAM Challenge We provide a comprehensive benchmark1 based on the SubT-MRS dataset by conducting ICCV'23 SLAM Challenge. Our evaluation reveals the limitations of state-of-the-art visual and LiDAR SLAM solutions. Furthermore, we conduct extensive experiments to verify robustness metric against various sensor degradation and length of trajectories. + +# 2. Related Work + +Multimodal Sensors Multimodal sensor datasets are crucial for the development of robust SLAM systems, as single sensor modalities are not comprehensive for all scenarios. Existing datasets typically focus on a limited range of modalities, such as monocular or stereo cameras [3, 49], LiDAR [19], event cameras [29], and depth cameras [42]. The KITTI dataset [19], although covering most modalities, is geared towards on-road scenarios and lacks thermal cameras, limiting its applicability to simple, controlled environments. Conversely, the ViViD++ dataset [23] incorporates thermal images in outdoor settings but is deficient in hardware synchronization, posing challenges for SLAM system development. The Weichen dataset [10] offers thermal images in indoor environments with accurate ground truth poses but is confined to motion capture room settings. In contrast, our SubT-MRS dataset delivers a comprehensive suite of time-synchronized multimodal sensors, including LiDAR, monocular cameras, thermal cameras, depth cameras, and IMU, along with centimeter-level ground truth, catering to a diverse range of research needs. + +Heterogeneous Platforms Developing a robust SLAM system necessitates versatility in handling different motion patterns and various robotic platforms. However, most existing datasets are tailored for single-robot scenarios. While there are a few multi-robot datasets, they typically involve homogeneous platforms [9, 15, 24, 44], posing challenges + +
DatasetMulti-SpectrumMulti-DegradationMulti-Robot
CameraIMULiDAR/DepthThermalIlluminationSnow SmokeStructurelessSubTAggressive MotionVehicleDroneLeggedHandheld
EuRoC MAV [3]✓✓XX✓XXXXX✓XX
PennCOsYVIO [33]✓✓XXXXXXXXXX✓
TUM VIO [39]✓✓XX✓XXXXXXX✓
UZH-FPV [12]✓✓XXXXXX✓X✓XX
College Dataset [52]✓✓✓XXXXXXXXX✓
RobotCar [28]✓✓✓X✓✓XXX✓XXX
UMA VI [54]✓✓XX✓X✓XXXXX✓
UMich [6]✓✓✓X✓XXXX✓XXX
KITTI [19]✓✓✓XXX✓X✓✓XXX
Virtual KITTI [16]✓✓✓X✓X✓X✓✓XXX
DARPA SubT [37]✓✓✓X✓X✓✓X✓XXX
Hilti SLAM [21]✓✓✓X✓X✓XXXXX✓
M3ED [7]✓✓✓X✓XXXX✓✓✓X
TartanAir [47]✓✓✓X✓✓✓X✓X✓XX
SubT-MRS (Ours)✓✓✓✓✓✓✓✓✓✓✓✓✓
+ +Table 1. Comparison of SLAM datasets on multi-sensors, multi robots, and multi degradation. + +in evaluating SLAM performance across varied robotic platforms. The AirMuseum dataset [13] includes drones and ground robots but lacks data from legged robots. As a comparison, SubT-MRS dataset features diverse robotic platforms including legged robots, aerial robots, and wheeled robots, operating in varied environments and sensor setups. + +Extreme Environmental Conditions Developing and testing SLAM systems in extreme environmental conditions is crucial for mitigating potential real-world failures. However, most existing SLAM datasets have been primarily limited to single, controlled environments. The TUM-VI [39] and UMA-VI [54] datasets, being indoor-outdoor visual-inertial datasets, pose challenges due to varying illumination and low-texture environments. The EuRoC MAV [3] and UZH-FPV drone racing [12] datasets, popular in the SLAM community, offer data on aggressive drone movements but usually in consistently lit conditions. The KITTI dataset [19], a staple in autonomous driving research for its outdoor LiDAR-Visual-Inertial data. The Hilti SLAM dataset [21] includes both indoor and outdoor LiDAR-visual-inertial datasets with dynamic lighting and confined spaces. However, both of them focus more on accuracy than on a variety of environmental degradations. TartanAir [47] covers most of the challenging environments but it is a simulation dataset that will pose the sim-to-real gap. + +In contrast, the SubT-MRS dataset provides 30 varied scenes, encompassing a wide array of environments. These include SubT tunnels, urban areas, caverns, multifloor structures, dark corridors, and foggy conditions. It also features textureless surfaces, dusty and smoky environments, off-road areas with aggressive motion, and snowy terrains prone to slippage. We include challenging lighting conditions such as extreme darkness and overexposure, along with geometrically challenging scenarios like featureless corridors, staircases, and self-similar cave layouts. Additionally, the dataset incorporates perceptual challenges posed by smoke, fog, dust, and various weather conditions. + +# 3. SubT-MRS Dataset + +We next present the SubT-MRS dataset from three aspects including its collecting environments and settings, ground truth collection, and the new robustness metric. A detailed comparison with other datasets is listed in Table 1. + +# 3.1. All-weather Environments + +As mentioned in Sec. 2, the SubT-MRS dataset distinguishes itself through its multimodal sensor setups, heterogeneous robotic platforms, and extreme environmental conditions. This section emphasizes this comprehensive nature pushing SLAM towards all-weather environments. + +# 3.1.1. Multimodal Sensors + +Multimodal sensors provide a wealth of information for SLAM systems operating in all-weather environments. This diversity improves scene understanding and strengthens the system's perception ability. Therefore, We incorporate diverse sensors to ensure system robustness. + +Sensor Pack We embedded 4 Leopard Imaging RGB fisheye cameras, 1 Velodyne puck, 1 Epson M-G365 IMU, 1 FLIR Boson thermal camera, and NVIDIA Jetson AGX Xavier as a sensor pack shown in Figure 2. + +Time Synchronization To ensure the overall consistency of the fused data, we meticulously synchronize the sensors' time using the 'pulse per second (PPS)' technique. As illustrated in Figure 2, the IMU, LiDAR, and thermal camera are directly synchronized with the CPU clock, while the four RGB camera are synchronized using an FPGA board. Consequently, we can effectively manage the time synchronization gap between each pair of sensors not to exceed 3ms. + +Fisheye and Thermal Camera Calibration Camera calibration plays an essential role in the efficiency of a SLAM system. For calibrating the fisheye cameras, we employed the open-source toolkit Kalibr [36], focusing on the intrinsic and extrinsic parameters. Specifically, we use the radial-tangential distortion model to rectify the omnidirectional fisheye camera model. Calibrating thermal cameras, however, poses unique challenges, especially in gathering high- + +![](images/4a5dcfa2e9f9e3cf0bef64f230e655cdef81aa0eee5146db6cfaca25eed1ea7d.jpg) +Figure 2. An overview of the sensor pack used in SubT-MRS dataset. It is equipped with a Xavier processing unit with hardware time synchronization for multimodal sensors including LiDAR, fisheye cameras, thermal cameras, depth cameras (option), and an IMU. + +quality thermal data. To tackle this, we set up a $7 \times 9$ chessboard, heated by direct sunlight, to generate high-contrast thermal data, ensuring accuracy in the calibration process. + +IMU Calibration and Extrinsic Calibration For Lidar-IMU extrinsic calibration, we utilize the CAD model to obtain the calibration parameters. In the case of Camera-IMU extrinsic calibration, we employ the Kalibr toolbox [36] to estimate the extrinsic matrix. To estimate the sensors' bias and the random walk noise of the gyroscope and accelerometer, we collected static data from the IMU and calibrated it using an Allan variance-based IMU tool [14, 17]. + +# 3.1.2. Multi-degraded Environments + +SubT-MRS includes multiple challenging environments including visually degraded environments, geometrically degraded environments, and their combination. + +Visual Degradation Poor-quality visual features can significantly hinder the performance of feature extraction processes and disrupt the accuracy of feature matching. Such issues arise in low-light conditions with inconsistent brightness or image noise introduced by air obscurants. SubT-MRS encompasses these challenges and provides a wide range of visual degradation. This includes environments with limited lighting, such as hospital interiors and caves (Figure 3 A-F), as well as smoky or dusty conditions that cause visual obstruction (Figure 3 M-N), and snowy areas with reduced visibility (Figure 3 H, K, and L). + +Geometric Degradation Lack of geometrical features poses significant challenges to LiDAR odometry. The root of this issue often lies in the limited sensing capabilities of LiDAR sensors and the constraints due to their mechanical installation. SubT-MRS captures a variety of environments that exemplify such challenges. It includes long, featureless corridors (Figure 3 E) and staircases (Figure 3 C and G). These scenarios illustrate various forms of geometric degradation, contributing to the improvement of LiDAR odometry systems in challenging conditions. + +Mixed Degradation A mixed visual and geometric degradation can further hinder the performance of SLAM sys + +tems. Examples in SubT-MRS include long, dimly lit corridors (Figure 3 A, E, and F), poorly illuminated staircases (Figure 3 C), and environments affected by snowy weather (Figure 3 H, K, and L). In these settings, both LiDAR odometry and visual odometry are prone to failure due to the compounded effects of mixed degradation. The inclusion of such scenarios in SubT-MRS is crucial for evaluating and improving the resilience of multimodal SLAM algorithms. + +# 3.1.3. Heterogeneous Robot Platforms + +Most of current datasets focus on single-robot systems, limiting multi-robot SLAM development as shown in Table 1. To address this, we employed diverse robot platforms, including aerial, legged, and wheeled robots, navigating through various environments from urban campuses to medical facilities and natural terrains like caves. This diversity offers a range of scenarios to enhance SLAM algorithms for effective operation in challenging conditions. + +Extrinsic Calibration for Multiple Robots To ensure the multi-robot system shares a common coordinate system, we perform the extrinsic calibration process [38]. It has 2 steps: First, each robot runs Super Odometry [53] to generate its local map using LiDAR data and share it with other robots through a wireless network; Second, the remaining robots identify overlapping regions between their local maps and estimate the extrinsic parameters using GICP [40]. + +# 3.2. Ground Truth + +Ground Truth Map As depicted in Figure 4, we utilized the high-precision FARO Focus 3D S120 3D scanner for creating ground-truth maps. This cutting-edge 3D scanner can measure distances up to $120\mathrm{m}$ , with a maximum measurement rate of $976\mathrm{K}$ points per second. The accuracy of the ground truth map is noteworthy, maintaining a range error within $\pm 2\mathrm{mm}$ . We have developed ground truth models for diverse environments, including subterranean areas like urban areas $(350\mathrm{m} \times 350\mathrm{m})$ , caves $(150\mathrm{m} \times 200\mathrm{m})$ and tunnels $(100 \times 200\mathrm{m})$ , as well as indoor-outdoor spaces $(200\mathrm{m} \times 200\mathrm{m})$ within the CMU campus. To reduce drift, a + +![](images/e99cad20d18ec8f0048688cceaa90f07e28c3d2bab5ffea665868a82026408ba.jpg) +Figure 3. The SubT-MRS datasets were collected across diverse seasons, capturing environments with perceptual challenges such as poor illumination, darkness, and water puddles, where visual sensors falter. They also include geometrically complex areas like long featureless corridors and steep multi-floor structures, challenging LiDAR systems with potential drift. Moreover, these datasets cover conditions with airborne obscurants like dust, fog, snow, and smoke in tough environments, including caves, deserts, long tunnels, and off-road areas. + +![](images/3c0d36f7d01aeff96cc49b14894396b1d84ab5655b90321852f2810c4bddc139.jpg) +Figure 4. The SubT-MRS dataset facilitates the generation of high-precision ground truth maps and trajectories. Figure (a) shows the ground truth trajectory in multi-floor settings. Figure (b) displays ground truth maps for indoor and outdoor areas, encompassing long corridors, multi-floor structures, and open spaces. Figure (c) features photo-realistic scans based on our ground truth maps in cave environments. + +loop closure algorithm was applied to correct the poses. Notably, $96\%$ of the scans achieved a position uncertainty of less than $2\mathrm{mm}$ . This ground-truth map is crucial for SLAM development where precision is of utmost importance. + +Ground Truth Trajectory Ground truth trajectories for all sequences are generated based on the ground truth map. In this process, we establish point-to-point, point-to-plane, and point-to-line correspondences[40] between the Ground + +Truth map and the current LiDAR scan. Our estimated ground truth trajectory is a comprehensive fusion of various constraints derived from the ground truth map, visual odometry, LiDAR odometry, and IMU measurements. This approach ensures the robustness of our solution across diverse environmental conditions. To verify accuracy, we reconstruct LiDAR maps from ground truth trajectory and compare it with our ground truth maps. + +Table 2. SLAM Challenge Results (Blue shadings indicate rankings of ATE and Robustness Metric; L: LiDAR I: IMU C: Camera) + +
#TeamMethodOdometry TypeDeviceRealTime (s)CPU/GPU (%)RAM (GB)ATE↓\( {R}_{v} \uparrow /{R}_{w} \uparrow \)Sensors L I C
1Liu et alFAST-LIO2 [51], HBA [27]FilterIntel i7-9700K51.31098.667 / 04.0520.5880.517/0.770✓✓
2Yibin et alLIO-EKF [45]FilterIntel i7-107000.00652.167 / 00.0724.3130.441/0.574✓✓
3Weitong et alFAST-LIO[2], Pose Graph[11]FilterIntel Xeon(R)E3-1240v50.12522.63 / 04.3050.6630.473/0.747✓✓
4Kim et alFAST-LIO2[50], Point-LIO[20], Quatro[25]FilterIntel i5-125000.268101.108 / 055.643.8250.479/0.615✓✓
5Zhong et alDLO[8], Scan-Context++[22]SW OptAMD Ryzen 9 5900x0.02713.289 / 01.1741.2090.276/0.486✓✓
1Peng et alDVI-SLAM [32]LearningIntel i9-12900183.233- / 14911 (4)0.5470.473/0.788✓✓
2Jiang et alLET-NET[26], VINS-Mono[34]HybridIntel i5-94000.06440.35 / 04.3371.0930.078/0.322✓✓
3Thien et alVR-SLAM[31]SW OptIntel i9-129000.142176.44 / 09.1113.0370.083/0.372✓✓
4Li et alORB-SLAM3[4]SW Opt.Intel i7-107000.01965.028 / 00.3868.9750.163/0.474✓✓
+ +Table 3. Accuracy Performance on Geometric Degradation. Red numbers represent ATE ranking. * denotes incomplete submissions. + +
TeamGeometric Degradation (Real World)SimulationMix DegradationAverage
UrbanTunnelCaveNuclear_1Nuclear_2Laurel_CavernsFactoryOceanSewerageLong CorridorMulti FloorBlock Lidar
Liu et al10.3070.0950.6290.1220.2350.2600.8890.7570.9781.4540.4010.9340.588
Weitong et al20.260.0960.6170.1200.2220.4020.9980.7701.5861.2540.5771.0560.663
Kim et al30.3310.0920.7870.1230.2700.27910.62822.4257.1472.1000.6501.0683.825
Yibin et al41.0600.2200.7500.4700.6209.1404.9200.28024.4602.9905.5001.3404.312
Zhong et al51.2050.695-1.1751.722.080.8890.7781.13---1.209*
Average0.6330.2400.6960.4020.61342.4323.6655.0027.0601.9501.7821.099
+ +# 3.3. Robustness Metric + +As discussed in Sec. 1, existing evaluation metrics such as absolute trajectory error (ATE) [3] have limitations in evaluating the SLAM's robustness in real-world applications. ATE primarily focuses on the accuracy of trajectory and does not consider the completeness of trajectory (recall). Also, it can not effectively capture velocity changes that have a direct impact on the robot's safety. For example, a pose spike error might not be immediately noticeable in ATE evaluation, it may result in a crash in an aerial robot control. To address these gaps, we introduce a new robustness metric based on estimated velocity. That is because velocity estimation from SLAM is crucial for the robot's control system which directly impacts stability, affecting the robot's safety. The new robustness metric is the area under the curve (AUC) of the F-1 score: + +$$ +F _ {1} (e) = \frac {2 P (e < T) R (e < T)}{P (e < T) + R (e < T)}, \tag {1} +$$ + +where the precision $P$ quantifies the precision of the estimated velocity as a percentage: how closely the estimated velocity points lie to the ground truth point and the recall rate $R$ quantifies the estimated velocity's completeness: to what extent all the ground-truth points are covered. A high F-1 score can only be achieved by the velocity estimation that is both accurate and complete throughout the entire run. Specifically, an estimated error $e$ is regarded as an inlier, if it is smaller than a threshold $T$ . To scale the threshold $T$ within the range of [0, 1], we apply exponential mapping $\exp(-10T)$ when calculating robustness metric. + +Position Robustness $R_{p}$ & Rotation Robustness $R_{r}$ The AUC of F-1 score can be defined using both linear velocity and angular velocity, which can reflect the robustness of position and rotation estimation, respectively. Specifically, we define $R_{p} = \mathrm{AUC}(F_{1}(\mathrm{v}_{\mathrm{e}}))$ and $R_{r} = \mathrm{AUC}(F_{1}(\omega_{\mathrm{e}}))$ , where $v_{e}$ and $\omega_{e}$ are the estimated error of linear velocity and rotation velocity, respectively. Note that not all SLAM solutions can output the velocity at the desired frequency. To address this, we use B-splines [30, 46] to obtain smooth trajectories, compute derivatives for smooth trajectories, and derive estimated linear velocity and angular velocity. For more details, please refer to the supplementary. + +# 4. Open SLAM Challenge Results + +In this section, we will present our works from two perspectives: accuracy evaluation and robustness evaluation. + +# 4.1. Accuracy Evaluation + +The results of the ICCV SLAM challenge underscore the necessity for advancements in system robustness. From 29 submissions, we identified 5 winners in the LiDAR category and 4 in the visual category. However, in the sensor fusion track, which addresses both visual and geometric degradation, no submissions met the criteria for success. This result reveals the existing SLAM systems still have lots of space to improve. Sequence characteristics will be detailed in the supplementary material. Table 2 shows the results for both the LiDAR and visual tracks. Unfortunately, there are no current solutions that can balance high accuracy and real-time performance in challenging environments. Since several environments lack geometric features or visual features, + +Table 4. Accuracy Performance on Visual Degradation. Red numbers represent ATE ranking. * denotes incomplete submissions. + +
TeamVisual Degradation (Real World)SimulationAverage
Lowlight 1Lowlight 2Over ExposureFlash LightSmoke RoomOutdoor NightEnd of WorldMoonWestern Desert
Peng et al11.0631.6370.5030.440.1530.8270.0380.1950.0700.547
Thien et al21.0812.0541.7331.05410.5327.6920.7531.2281.2093.037
Jiang et al31.0191.1261.9112.3413.75711.8212.1540.6044.0103.193
Li et al45.7687.8341.7571.2955.37010.766-30.07-8.98*
Average2.2323.1631.4761.2824.9537.7760.9828.0241.763
+ +![](images/7b8e8b8ab874deecdf5e7d6a75971d02e8e9b765ff81b297d02a906c91599ca0.jpg) +Figure 5. From left to right, it shows robustness metric $R_{p}$ and $R_{r}$ for LiDAR and visual sequences respectively. Note: This is a summary of results for all sequences, with weights based on the trajectory length. The area under the curve (AUC) represents the robustness $(R_{p}, R_{r})$ . The x-axis shows velocity thresholds for classifying estimated velocities as inliers and the y-axis is F-1 score. + +![](images/55cfc6031a312bd26283cd3f30215cbe3a40100a153c8fd52f3f5b533cf7c593.jpg) + +![](images/1ddcbb645616be2d30870b196461b2b337a9d4108afc101e9f4d046bbd9ca7d8.jpg) + +![](images/ddea1daf4391967666e91c1708573d8df339b08b3c6cbb9141b9eeb3029f5647.jpg) + +the algorithms required more processing times to achieve reasonably high accuracy. Liu et al. at the University of Hong Kong, who achieved the highest accuracy, recorded an Absolute Trajectory Error (ATE) of $0.588\mathrm{m}$ , but each iteration took 51.3 seconds. In the visual track, Peng et al. from the Samsung Research Center attained a leading ATE of $0.547\mathrm{m}$ , but each iteration required 183 seconds. + +LiDAR Track Discussion To enhance robustness in all-weather environments, a pivotal question arises: what are the limits of current LiDAR SLAM algorithms? Table 3 presents a summary of the ATE/RPE errors observed in real-world geometrically degraded, simulated, and mixed degradation scenarios across the top five teams. In real-world geometrically degraded environments, we observed the first limitation: existing LiDAR solutions struggle in confined spaces such as caves. Almost all algorithms perform worse in Cave (Figure 3S) and Laureal_Cavern environments (Figure 3N) with average ATE 0.696 and 2.432 meters respectively. One reason is that cave environments is the most confined space which may lack sufficient geometric features, compared with other environments like tunnel (Figure 3J) and Urban (Figure 3B) and Nuclear scenes (Figure 3U). + +In simulated environments, we encountered the second limitation: existing methods are tailored for fixed motion patterns like vehicles and struggle with unpredictable motion patterns. Since our simulation sequence is derived from TartanAir [47], featuring aggressive and random motion patterns, it might not follow the usual velocity distribution domain and has not been thoroughly tested. As shown in Table 3 (simulation), it resulted in significantly higher ATE errors $(5.239\mathrm{m})$ compared to real-world's $(0.836\mathrm{m})$ . + +In mixed degradation environments, the third limitation we found is that existing methods cannot actively select the most informative measurements to adapt to new environments. For example, in the Long Corridor sequence (Figure 3 A) with low lighting and a featureless environment, the scenario is intricately designed for collaborative usage of LiDAR and Visual sensors. Algorithms are expected to utilize visual information in geometrically degraded environments while disregarding it in darkness. Similarly, the Block LiDAR sequences (Figure 3 V) aim to simulate sensor drop scenarios frequently encountered in robotic applications. These sequences involve alternating periods of LiDAR and Visual data loss, challenging algorithms to promptly detect sensor failures and switch to the other modality for SLAM. However, the average ATE error $(1.61\mathrm{m})$ in Table 3 (mixed degradation) is much higher than in geometric degradation scenarios $(0.836\mathrm{m})$ , suggesting the above limitations. Details are in the supplementary. + +Visual Track Discussion To improve the robustness, a pivotal question arises: what are the limits of current visual SLAM? To what extent does image quality impact a visual SLAM system? Table 4 shows a summary of ATE errors from real-world and simulation scenarios from the awarded 4 teams. In real-world scenarios, the first limitation is the lack of anti-noise capability in current methods, especially in low-quality image settings. We assessed visual odometry accuracy in visually degraded environments, including low lighting, sunlight overexposure, flashing lights, smoke-filled rooms, and nighttime outdoor settings (Figure 3). Trajectories in conditions like the Smoke Room and Outdoor Night exhibit significantly higher average ATE er + +Table 5. Robustness Performance on Geometric Degradation. Red numbers are robustness ranking. Larger values indicate better robustness. + +
TeamGeometric Degradation (Real World)SimulationMix DegradationAverage
UrbanTunnelCaveNuclear_1Nuclear_2Laurel_CavernsFactoryOceanSewerageLong CorridorMulti FloorBlock Lidar
Liu et al1Rp↑0.8110.8650.7360.7470.5040.8160.1570.1350.1350.3960.5290.3710.516
Weitong et al20.7730.8380.6900.8710.7390.1520.1530.1490.0810.4070.4560.3450.471
Kim et al40.7470.8650.7370.8700.6890.7770.0140.0010.0100.4100.2850.3450.479
Yibin et al30.6500.7830.5570.7210.4900.4810.2190.5600.0190.1750.2600.3700.440
Zhong et al50.5670.6830.2040.6800.4260.4470.1030.0840.1220000.276
Liu et al1Rr↑0.8880.8930.8160.8570.7780.8610.6920.7190.6520.7530.6890.6430.770
Weitong et al20.8860.8920.8110.8930.8400.5980.6880.7200.5140.7660.7060.6440.746
Kim et al40.6800.8930.8160.8930.8030.8080.1980.1250.3000.8020.4220.6420.615
Yibin et al30.6240.7070.4620.7330.5400.4670.8620.9620.4960.4770.3420.2130.573
Zhong et al50.7310.7450.4370.7570.6030.5230.6880.7040.6430000.485
AverageRp0.7100.8070.5850.7780.5700.5350.1300.1860.0730.2780.3060.286
Average Rr0.7620.8260.6680.8270.7130.6510.6260.6460.5210.5600.4320.428
+ +Table 6. Robustness Performance on Visual Degradation. Red numbers are robustness ranking. Larger values indicate better robustness. + +
TeamVisual Degradation (Real World)SimulationAverage
Lowlight 1Lowlight 2Over ExposureFlash LightSmoke RoomOutdoor NightEnd of WorldMoonWestern Desert
Peng et al10.3570.2270.2640.2030.5360.2700.6990.8930.8060.472
Thien et al30.0450.0700.2400.1560.1310.07500.03100.083
Jiang et al4Rp↑0.0460.0390.1940.0880.2420.0950000.078
Li et al20.3420.1870.2570.2080.3220.14200.00600.162
Peng et al10.6410.5810.7440.6500.8780.6700.9750.9750.9740.787
Thien et al30.4130.4450.6100.2690.4740.4870.1770.31500.354
Jiang et al4Rr↑0.4530.4520.6190.2520.5420.5770.00200.1570.339
Li et al20.6420.5740.6570.6510.7730.6600.00.30500.473
AverageRp0.1980.1310.2390.1640.3080.1460.1750.2320.202
Average Rr0.5370.5130.6580.4560.6670.5980.2880.3990.283
+ +rors (4.95m, 7.76m respectively). Smoke and night scenes challenge feature extraction and tracking and introduce sensor noise, emphasizing the need for robust anti-noise algorithms. The second limitation is that existing methods struggle to overcome aggressive motion. Even in simulations with relatively good image quality, most methods still show significant average ATE errors (8.024m) on the Moon sequence (Figure 3 Y). See more details in supplementary. + +# 4.2. Robustness Evaluation + +The F-1 robustness curve provides a detailed evaluation under different error tolerances, while $R_{p}$ and $R_{r}$ provide an overall measure of robustness that is not dependent on a specific decision threshold. This suggests the flexibility of our new robustness metric. Figure 5 clearly shows the robustness performance of all teams regarding position and rotation, highlighting the differences in robustness between the methods. The values of $R_{p}$ and $R_{r}$ , which are the AUC under F-1 curves, are displayed in the brackets of the legends. It reveals that Liu et al. and Peng et al. are the most robust solutions for LiDAR and visual tracks, respectively. + +Is the robustness metric robust? We did extensive experiments on our robustness metric against diverse sensor degradation settings. Table 5 and 6 present a summary of robustness performance in real-world, simulated, and mixed degradation scenarios for both LiDAR and visual sequences. In Table 5, we observe that the average values of $R_{p}$ and $R_{r}$ for the mixed degradation environment (0.290, + +0.473) are significantly lower than geometric degradation environments (0.664, 0.754). This observation suggests that the majority of SLAM algorithms exhibit reduced robustness in mixed degradation environments as compared to geometrically degraded ones. This aligns with our expectations and verifies the effectiveness of the robustness metric. + +Our robustness ranking, indicated by small red numbers, differs from those on ATE ranking shown in Table 3 and 4. This is because our F-1 score-based metrics $R_{p}$ and $R_{r}$ jointly consider precision and recall rate to provide a balanced evaluation of SLAM performance, considering both metrics across the full spectrum of thresholds. In contrast, the ATE focuses solely on precision, neglecting the trajectory's completeness (recall). We also evaluate our metric on various synthetic trajectories in the supplementary. + +# 5. Conclusion + +We introduce SubT-MRS, a comprehensive SLAM dataset with various sensor data, locomotion patterns, and over 30 degradation types in simulation and real-world settings to push SLAM towards all-weather environments. Additionally, we introduce a new robustness metric to evaluate the reliability of SLAM systems, enhancing robot safety control. 29 teams have tested SubT-MRS in the organized SLAM challenge and we expect that it will serve as a critical benchmark for future SLAM development. + +# References + +[1] https://www.darpa.mil/program/darpa-subterranean-challenge. 1 +[2] Chunge Bai, Tao Xiao, Yajie Chen, Haoqian Wang, Fang Zhang, and Xiang Gao. Faster-lio: Lightweight tightly coupled lidar-inertial odometry using parallel sparse incremental voxels. IEEE Robotics and Automation Letters, 7(2):4861-4868, 2022. 6 +[3] Michael Burri, Janosch Nikolic, Pascal Gohl, Thomas Schneider, Joern Rehder, Sammy Omari, Markus W Achtelik, and Roland Siegwart. The EuRoC micro aerial vehicle datasets. The International Journal of Robotics Research, 2016. 1, 2, 3, 6 +[4] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 6 +[5] Carlos Campos, Richard Elvira, Juan J. Gómez Rodríguez, José M. M. Montiel, and Juan D. Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6): 1874-1890, 2021. 1 +[6] Nicholas Carlevaris-Bianco, Arash K Ushani, and Ryan M Eustice. University of michigan north campus long-term vision and lidar dataset. The International Journal of Robotics Research, 35(9):1023-1035, 2016. 3 +[7] Kenneth Chaney, Fernando Cladera, Ziyun Wang, Anthony Bisulco, M Ani Hsieh, Christopher Korpela, Vijay Kumar, Camillo J Taylor, and Kostas Daniilidis. M3ed: Multi-robot, multi-sensor, multi-environment event dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015–4022, 2023. 3 +[8] Kenny Chen, Brett T. Lopez, Ali-akbar Agha-mohammadi, and Ankur Mehta. Direct lidar odometry: Fast localization with dense point clouds. IEEE Robotics and Automation Letters, 7(2):2000–2007, 2022. 6 +[9] Long Chen, Yuchen Li, Chao Huang, Bai Li, Yang Xing, Daxin Tian, Li Li, Zhongxu Hu, Xiaoxiang Na, Zixuan Li, et al. Milestones in autonomous driving and intelligent vehicles: Survey of surveys. IEEE Transactions on Intelligent Vehicles, 8(2):1046-1056, 2022. 2 +[10] Weichen Dai, Yu Zhang, Shenzhou Chen, Donglei Sun, and Da Kong. A multi-spectral dataset for evaluating motion estimation systems. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 5560-5566. IEEE, 2021. 2 +[11] Frank Dellaert, Michael Kaess, et al. Factor graphs for robot perception. Foundations and Trends® in Robotics, 6(1-2): 1-139, 2017. 6 +[12] Jeffrey Delmerico, Titus Cieslewski, Henri Rebecq, Matthias Faessler, and Davide Scaramuzza. Are we ready for autonomous drone racing? the UZH-FPV drone racing dataset. In IEEE Int. Conf. Robot. Autom. (ICRA), 2019. 1, 3 +[13] Rodolphe Dubois, Alexandre Eudes, and Vincent Frémont. Airmuseum: a heterogeneous multi-robot dataset for stereovisual and inertial simultaneous localization and mapping. In + +2020 IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI), pages 166-172, 2020. 3 +[14] Naser El-Sheimi, Haiying Hou, and Xiaoji Niu. Analysis and modeling of inertial sensors using allan variance. IEEE Transactions on Instrumentation and Measurement, 57(1): 140-149, 2008. 4 +[15] Dapeng Feng, Yuhua Qi, Shipeng Zhong, Zhiqiang Chen, Yudu Jiao, Qiming Chen, Tao Jiang, and Hongbo Chen. S3e: A large-scale multimodal dataset for collaborative slam. arXiv preprint arXiv:2210.13723, 2022. 2 +[16] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4340-4349, 2016. 3 +[17] Wenliang Gao. A ros package tool to analyze the imu performance, 2018. 4 +[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI Vision Benchmark Suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 1 +[19] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 3 +[20] Dongjiao He, Wei Xu, Nan Chen, Fanze Kong, Chongjian Yuan, and Fu Zhang. Point-lio: Robust high-bandwidth light detection and ranging inertial odometry. Advanced Intelligent Systems, page 2200459, 2023. 6 +[21] Michael Helmberger, Kristian Morin, Beda Berner, Nitish Kumar, Giovanni Cioffi, and Davide Scaramuzza. The hilti slam challenge dataset. IEEE Robotics and Automation Letters, 7(3):7518-7525, 2022. 1, 3 +[22] Giseop Kim and Ayoung Kim. Scan context: Egocentric spatial descriptor for place recognition within 3d point cloud map. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4802-4809. IEEE, 2018. 6 +[23] Alex Junho Lee, Younggun Cho, Young-sik Shin, Ayoung Kim, and Hyun Myung. Vivid++: Vision for visibility dataset. IEEE Robotics and Automation Letters, 7(3):6282-6289, 2022. 2 +[24] Keith YK Leung, Yoni Halpern, Timothy D Barfoot, and Hugh HT Liu. The utias multi-robot cooperative localization and mapping dataset. The International Journal of Robotics Research, 30(8):969–974, 2011. 2 +[25] Hyungtae Lim, Suyong Yeon, Suyong Ryu, Yonghan Lee, Youngji Kim, Jaeseong Yun, Euigon Jung, Donghwan Lee, and Hyun Myung. A single correspondence is enough: Robust global registration to avoid degeneracy in urban environments. page Accepted. To appear, 2022. 6 +[26] Yicheng Lin, Shuo Wang, Yunlong Jiang, and Bin Han. Breaking of brightness consistency in optical flow with a lightweight cnn network, 2023. 6 +[27] Zheng Liu, Xiyuan Liu, and Fu Zhang. Efficient and consistent bundle adjustment on lidar point clouds. IEEE Transactions on Robotics, 2023. 6 + +[28] Will Maddern, Geoffrey Pascoe, Chris Linegar, and Paul Newman. 1 year, $1000\mathrm{km}$ : The oxford robotcar dataset. The International Journal of Robotics Research, 36(1):3-15, 2017. 3 +[29] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. The International Journal of Robotics Research, 36(2):142-149, 2017. 2 +[30] Elias Mueggler, Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. Continuous-time visual-inertial odometry for event cameras. IEEE Transactions on Robotics, 34(6): 1425-1440, 2018. 6 +[31] Thien Hoang Nguyen, Shenghai Yuan, and Lihua Xie. Vrslam: A visual-range simultaneous localization and mapping system using monocular camera and ultra-wideband sensors, 2023. 6 +[32] Xiongfeng Peng, Zhihua Liu, Weiming Li, Ping Tan, SoonYong Cho, and Qiang Wang. Dvi-slam: A dual visual inertial slam network, 2023. 6 +[33] Bernd Pfrommer, Nitin Sanket, Kostas Daniilidis, and Jonas Cleveland. Penncosyvio: A challenging visual inertial odometry benchmark. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 3847-3854. IEEE, 2017. 1, 3 +[34] Tong Qin, Peiliang Li, and Shaojie Shen. Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics, 34(4):1004-1020, 2018. 1, 6 +[35] Milad Ramezani, Yiduo Wang, Marco Camurri, David Wisth, Matias Mattamala, and Maurice Fallon. The newer college dataset: Handheld lidar, inertial and vision with ground truth. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4353-4360. IEEE, 2020. 1 +[36] Joern Rehder, Janosch Nikolic, Thomas Schneider, Timo Hinzmann, and Roland Siegwart. Extending kalibr: Calibrating the extrinsic of multiple imus and of individual axes. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 4304-4311. IEEE, 2016. 3, 4 +[37] John G Rogers, Arthur Schang, Carlos Nieto-Granda, John Ware, John Carter, Jonathan Fink, and Ethan Stump. The darpa subt urban circuit mapping dataset and evaluation metric. In Experimental Robotics: The 17th International Symposium, pages 391–401. Springer, 2021. 3 +[38] Sebastian Scherer, Vasu Agrawal, Graeme Best, Chao Cao, Katarina Cujic, R Darnley, R DeBortoli, E Dexheimer, B Drozd, R Garg, et al. Resilient and modular subterranean exploration with a team of roving and flying robots. Field Robotics, 2022. 4 +[39] David Schubert, Thore Goll, Nikolaus Demmel, Vladyslav Usenko, Jörg Stuckler, and Daniel Cremers. The tum vi benchmark for evaluating visual-inertial odometry. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1680-1687. IEEE, 2018. 1, 3 +[40] Aleksandr Segal, Dirk Haehnel, and Sebastian Thrun. Generalized-icp. In Robotics: science and systems, page 435. Seattle, WA, 2009. 4, 5 + +[41] Tixiao Shan, Brendan Englot, Drew Meyers, Wei Wang, Carlo Ratti, and Rus Daniela. Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5135-5142. IEEE, 2020. 1 +[42] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A benchmark for the evaluation of rgb-d slam systems. In Proc. of the International Conference on Intelligent Robot Systems (IROS), 2012. 2 +[43] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 1 +[44] Yulun Tian, Yun Chang, Long Quang, Arthur Schang, Carlos Nieto-Granda, Jonathan P How, and Luca Carlone. Resilient and distributed multi-robot visual slam: Datasets, experiments, and lessons learned. arXiv preprint arXiv:2304.04362, 2023. 2 +[45] Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Louis Wiesmann, Jens Behley, and Cyril Stachniss. Kiss-icp: In defense of point-to-point icp-simple, accurate, and robust registration if done the right way. IEEE Robotics and Automation Letters, 8(2):1029–1036, 2023. 6 +[46] Chen Wang, Dasong Gao, Kuan Xu, Junyi Geng, Yaoyu Hu, Yuheng Qiu, Bowen Li, Fan Yang, Brady Moon, Abhinav Pandey, Aryan, Jiahe Xu, Tianhao Wu, Haonan He, Daning Huang, Zhongqiang Ren, Shibo Zhao, Taimeng Fu, Pranay Reddy, Xiao Lin, Wenshan Wang, Jingnan Shi, Rajat Talak, Kun Cao, Yi Du, Han Wang, Huai Yu, Shanzhao Wang, Siyu Chen, Ananth Kashyap, Rohan Bandaru, Karthik Dantu, Jiajun Wu, Lihua Xie, Luca Carlone, Marco Hutter, and Sebastian Scherer. PyPose: A library for robot learning with physics-based optimization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6 +[47] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual SLAM. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 3, 7 +[48] Wenshan Wang, Yaoyu Hu, and Sebastian Scherer. Tartanvo: A generalizable learning-based vo. In Conference on Robot Learning, pages 1761-1772. PMLR, 2021. 1 +[49] P. Wenzel, R. Wang, N. Yang, Q. Cheng, Q. Khan, L. von Stumberg, N. Zeller, and D. Cremers. 4Seasons: A cross-season dataset for multi-weather SLAM in autonomous driving. In Proceedings of the German Conference on Pattern Recognition (GCPR), 2020. 2 +[50] Wei Xu, Yixi Cai, Dongjiao He, Jiarong Lin, and Fu Zhang. FAST-LIO2: fast direct lidar-inertial odometry. IEEE Transactions on Robotics, abs/2107.06829, 2022. 6 +[51] Wei Xu, Yixi Cai, Dongjiao He, Jiarong Lin, and Fu Zhang. Fast-lio2: Fast direct lidar-inertial odometry. IEEE Transactions on Robotics, 38(4):2053–2073, 2022. 6 +[52] Lintong Zhang, Marco Camurri, and Maurice Fallon. Multicamera lidar inertial extension to the newer college dataset. arXiv preprint arXiv:2112.08854, 2021. 3 +[53] Shibo Zhao, Hengrui Zhang, Peng Wang, Lucas Nogueira, and Sebastian Scherer. Super odometry: Imu-centric + +lidar-visual-inertial estimator for challenging environments. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8729-8736. IEEE, 2021. 4 +[54] David Zuniga-Noël, Alberto Jaenal, Ruben Gomez-Ojeda, and Javier Gonzalez-Jimenez. The uma-vi dataset: Visual-inertial odometry in low-textured and dynamic illumination environments. The International Journal of Robotics Research, 39(9):1052-1060, 2020. 1, 3 \ No newline at end of file diff --git a/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/images.zip b/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ad021b5b9ef2740b1d30deb3523e800c4c57fa34 --- /dev/null +++ b/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dca19bf33ee81173ed05a5c0bc919e488852a9738c80476b1513dca7daa91725 +size 961835 diff --git a/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/layout.json b/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0bbdf4422b2ce85b153d574af02f780b4646c7cb --- /dev/null +++ b/2024/SubT-MRS Dataset_ Pushing SLAM Towards All-weather Environments/layout.json @@ -0,0 +1,7114 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 521, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 521, + 139 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 521, + 139 + ], + "type": "text", + "content": "SubT-MRS Dataset: Pushing SLAM Towards All-weather Environments https://superodometry.com/datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 161, + 548, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 161, + 548, + 204 + ], + "spans": [ + { + "bbox": [ + 52, + 161, + 548, + 204 + ], + "type": "text", + "content": "Shibo Zhao\\*, Yuanjun Gao, Tianhao Wu, Damanpreet Singh, Rushan Jiang, Haoxiang Sun, Mansi Sarawata, Yuheng Qiu, Warren Whittaker, Ian Higgins, Yi Du, Shaoshu Su, Can Xu, John Keller, Jay Karhade, Lucas Nogueira, Sourojit Saha, Ji Zhang, Wenshan Wang, Chen Wang, Sebastian Scherer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 177, + 207, + 421, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 207, + 421, + 221 + ], + "spans": [ + { + "bbox": [ + 177, + 207, + 421, + 221 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 177, + 207, + 421, + 221 + ], + "type": "text", + "content": "Carnegie Mellon University " + }, + { + "bbox": [ + 177, + 207, + 421, + 221 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 177, + 207, + 421, + 221 + ], + "type": "text", + "content": "University at Buffalo" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 249, + 192, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 249, + 192, + 261 + ], + "spans": [ + { + "bbox": [ + 143, + 249, + 192, + 261 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 274, + 290, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 290, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 290, + 550 + ], + "type": "text", + "content": "Simultaneous localization and mapping (SLAM) is a fundamental task for numerous applications such as autonomous navigation and exploration. Despite many SLAM datasets have been released, current SLAM solutions still struggle to have sustained and resilient performance. One major issue is the absence of high-quality datasets including diverse all-weather conditions and a reliable metric for assessing robustness. This limitation significantly restricts the scalability and generalizability of SLAM technologies, impacting their development, validation, and deployment. To address this problem, we present SubT-MRS, an extremely challenging real-world dataset designed to push SLAM towards all-weather environments to pursue the most robust SLAM performance. It contains multi-degraded environments including over 30 diverse scenes such as structureless corridors, varying lighting conditions, and perceptual obscurants like smoke and dust; multimodal sensors such as LiDAR, fisheye camera, IMU, and thermal camera; and multiple locomotions like aerial, legged, and wheeled robots. We developed accuracy and robustness evaluation tracks for SLAM and introduced novel robustness metrics. Comprehensive studies are performed, revealing new observations, challenges, and opportunities for future research." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 569, + 128, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 569, + 128, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 569, + 128, + 582 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 586, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 695 + ], + "type": "text", + "content": "Simultaneous Localization and Mapping (SLAM) is essential in robotics since it provides foundational perception and spatial awareness, and enables machines to understand and interact with the physical world in real-time. Therefore, it has a wide range of applications such as autonomous driving and space exploration. Despite significant advancements in both geometric [5, 34, 41] and data-driven SLAM methods [43, 48], existing solutions remain fragile in challenging conditions. One main reason is that existing SLAM" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 250, + 547, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 250, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 306, + 250, + 547, + 334 + ], + "type": "text", + "content": "algorithms are often developed and evaluated with datasets from controlled environments [3, 12, 18, 21, 33, 35, 39, 54]. For example, KITTI [18, 21] dataset is mostly collected in sunny weather and EuROC-MAV dataset [3] is collected in small well-lit rooms. These datasets, unfortunately, fail to capture the challenges in real-world scenarios, hindering the development of robust SLAM solutions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 338, + 547, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 547, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 547, + 542 + ], + "type": "text", + "content": "To bridge this gap and push SLAM towards all-weather environments, we present an extremely challenging dataset, SubT-MRS, including scenarios featuring various sensor degradation, aggressive locomotions, and extreme-weather conditions. The SubT-MRS dataset comprises 3 years of data from the DARPA Subterranean (SubT) Challenge [1] (2019-2021) and extends with an additional 2 years of diverse environments (2022-2023), containing mixed indoor and outdoor settings, including long corridors, off-road scenario, tunnels, caves, deserts, forests, and bushlands. Cumulatively, this forms a 5-year dataset encompassing over 500 hours and 100 kilometers of terrain subjected to multimodal sensors including LiDAR, fisheye cameras, depth cameras, thermal cameras, and IMU; heterogeneous platforms including RC cars, legged robots, aerial robots, and wheeled robots; and extreme obscurant conditions such as dense fog, dust, smoke, and heavy snow." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "content": "Additionally, we find there is no well-established metric to evaluate the robustness of a SLAM system. Existing evaluation metrics such as absolute trajectory error (ATE) [3] are not representative of the actual performance in robotic applications. We argue that to ensure the safety control of a robot, SLAM system evaluation should not only focus on poses but also its velocity. Taking robot localization as an example, while a momentary pose error spike may only slightly undermine the overall ATE, however, it can lead to a catastrophic crash in an aerial robotic platform. To effectively gauge the performance of a SLAM algorithm, we introduce a new robustness metric to evaluate the reliability of the SLAM system, particularly examining smoothness and accuracy of velocity estimation." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 220, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 220, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 220, + 713 + ], + "type": "text", + "content": "*Corresponding author: shiboz@andrew.cmu.edu" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22647" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 547, + 232 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 547, + 232 + ], + "type": "image", + "image_path": "b3d861d86989a475f5b89f18a2bceb82e3f55d46da24ddf4bfe5e6348b1bff74.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 232, + 547, + 264 + ], + "blocks": [ + { + "bbox": [ + 49, + 232, + 547, + 264 + ], + "lines": [ + { + "bbox": [ + 49, + 232, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 49, + 232, + 547, + 264 + ], + "type": "image", + "image_path": "3e2278a11c657ec21d4a1b1d9bd38063d647760e8f90618979c79422d7d0a8b0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 266, + 107, + 272 + ], + "lines": [ + { + "bbox": [ + 75, + 266, + 107, + 272 + ], + "spans": [ + { + "bbox": [ + 75, + 266, + 107, + 272 + ], + "type": "text", + "content": "SubT-Tunnel" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 266, + 184, + 272 + ], + "lines": [ + { + "bbox": [ + 154, + 266, + 184, + 272 + ], + "spans": [ + { + "bbox": [ + 154, + 266, + 184, + 272 + ], + "type": "text", + "content": "SubT-Urban" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 237, + 266, + 268, + 272 + ], + "lines": [ + { + "bbox": [ + 237, + 266, + 268, + 272 + ], + "spans": [ + { + "bbox": [ + 237, + 266, + 268, + 272 + ], + "type": "text", + "content": "SubT-Cavern" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 266, + 351, + 272 + ], + "lines": [ + { + "bbox": [ + 321, + 266, + 351, + 272 + ], + "spans": [ + { + "bbox": [ + 321, + 266, + 351, + 272 + ], + "type": "text", + "content": "Multi Floors" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 274, + 547, + 308 + ], + "blocks": [ + { + "bbox": [ + 399, + 266, + 446, + 273 + ], + "lines": [ + { + "bbox": [ + 399, + 266, + 446, + 273 + ], + "spans": [ + { + "bbox": [ + 399, + 266, + 446, + 273 + ], + "type": "text", + "content": "Dark Long Corridor" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 490, + 266, + 515, + 273 + ], + "lines": [ + { + "bbox": [ + 490, + 266, + 515, + 273 + ], + "spans": [ + { + "bbox": [ + 490, + 266, + 515, + 273 + ], + "type": "text", + "content": "Dense Fog" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 274, + 547, + 308 + ], + "lines": [ + { + "bbox": [ + 49, + 274, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 274, + 547, + 308 + ], + "type": "image", + "image_path": "4ff2e7e8ea5727e51f0d28853e2951db92306ff0c8c78c86400d2dbee3f76ece.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 310, + 106, + 316 + ], + "lines": [ + { + "bbox": [ + 72, + 310, + 106, + 316 + ], + "spans": [ + { + "bbox": [ + 72, + 310, + 106, + 316 + ], + "type": "text", + "content": "Laurel Cavern" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 143, + 310, + 192, + 317 + ], + "lines": [ + { + "bbox": [ + 143, + 310, + 192, + 317 + ], + "spans": [ + { + "bbox": [ + 143, + 310, + 192, + 317 + ], + "type": "text", + "content": "Textureless Corridor" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 217, + 310, + 293, + 316 + ], + "lines": [ + { + "bbox": [ + 217, + 310, + 293, + 316 + ], + "spans": [ + { + "bbox": [ + 217, + 310, + 293, + 316 + ], + "type": "text", + "content": "Darkness Stairs with Jerk Motion" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 310, + 364, + 317 + ], + "lines": [ + { + "bbox": [ + 296, + 310, + 364, + 317 + ], + "spans": [ + { + "bbox": [ + 296, + 310, + 364, + 317 + ], + "type": "text", + "content": "Dense Dust and Smoke" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 382, + 310, + 459, + 317 + ], + "lines": [ + { + "bbox": [ + 382, + 310, + 459, + 317 + ], + "spans": [ + { + "bbox": [ + 382, + 310, + 459, + 317 + ], + "type": "text", + "content": "Offroad with Aggressive Motion" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 471, + 310, + 535, + 317 + ], + "lines": [ + { + "bbox": [ + 471, + 310, + 535, + 317 + ], + "spans": [ + { + "bbox": [ + 471, + 310, + 535, + 317 + ], + "type": "text", + "content": "Snow with Slipping Motion" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 326, + 546, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 326, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 546, + 361 + ], + "type": "text", + "content": "Figure 1. Dense reconstruction from the SubT-MRS dataset, achieved through collaboration with diverse robots equipped with multimodal sensors. Colors represent different challenging environments (tunnels, caves, urban, confined spaces) captured by various robot types (aerial, legged, wheeled). The bottom section showcases a gallery with diverse visual, LiDAR, and mixed degradations." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "text", + "content": "Lastly, we perform extensive experiments using proposed degradation datasets to benchmark visual and LiDAR SLAM algorithms. These experiments identify the limitations of the existing SLAM systems and evaluate their robustness across the degradation. Our contributions include:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 441, + 287, + 681 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 47, + 441, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 287, + 548 + ], + "type": "text", + "content": "- All-weather Environments To push SLAM toward all-weather environments, we introduce SubT-MRS, an extremely challenging dataset. Spanning five years, it comprises over 500 hours and 100 kilometers of accurately measured trajectories. To the best of our knowledge, SubT-MRS is the first real-world dataset that specifically addresses failure scenarios of SLAM by incorporating a variety of degraded conditions, multiple robotic platforms, and diverse sets of multimodal sensors." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 549, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 287, + 597 + ], + "type": "text", + "content": "- Robustness Metric To evaluate the actual performance running on robots, we propose a robustness metric, which, to the best of our knowledge, is the first metric evaluating the reliability, safety, and resilience of SLAM." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 597, + 287, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 287, + 681 + ], + "type": "text", + "content": "- SLAM Challenge We provide a comprehensive benchmark1 based on the SubT-MRS dataset by conducting ICCV'23 SLAM Challenge. Our evaluation reveals the limitations of state-of-the-art visual and LiDAR SLAM solutions. Furthermore, we conduct extensive experiments to verify robustness metric against various sensor degradation and length of trajectories." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 379, + 392, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 379, + 392, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 379, + 392, + 392 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 399, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 638 + ], + "type": "text", + "content": "Multimodal Sensors Multimodal sensor datasets are crucial for the development of robust SLAM systems, as single sensor modalities are not comprehensive for all scenarios. Existing datasets typically focus on a limited range of modalities, such as monocular or stereo cameras [3, 49], LiDAR [19], event cameras [29], and depth cameras [42]. The KITTI dataset [19], although covering most modalities, is geared towards on-road scenarios and lacks thermal cameras, limiting its applicability to simple, controlled environments. Conversely, the ViViD++ dataset [23] incorporates thermal images in outdoor settings but is deficient in hardware synchronization, posing challenges for SLAM system development. The Weichen dataset [10] offers thermal images in indoor environments with accurate ground truth poses but is confined to motion capture room settings. In contrast, our SubT-MRS dataset delivers a comprehensive suite of time-synchronized multimodal sensors, including LiDAR, monocular cameras, thermal cameras, depth cameras, and IMU, along with centimeter-level ground truth, catering to a diverse range of research needs." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "Heterogeneous Platforms Developing a robust SLAM system necessitates versatility in handling different motion patterns and various robotic platforms. However, most existing datasets are tailored for single-robot scenarios. While there are a few multi-robot datasets, they typically involve homogeneous platforms [9, 15, 24, 44], posing challenges" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": "1SubT-MRS dataset is completely real-world data. Simulated data in the benchmark is to ensure evaluation completeness" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22648" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 72, + 545, + 223 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 545, + 223 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 545, + 223 + ], + "type": "table", + "html": "
DatasetMulti-SpectrumMulti-DegradationMulti-Robot
CameraIMULiDAR/DepthThermalIlluminationSnow SmokeStructurelessSubTAggressive MotionVehicleDroneLeggedHandheld
EuRoC MAV [3]✓✓XX✓XXXXX✓XX
PennCOsYVIO [33]✓✓XXXXXXXXXX✓
TUM VIO [39]✓✓XX✓XXXXXXX✓
UZH-FPV [12]✓✓XXXXXX✓X✓XX
College Dataset [52]✓✓✓XXXXXXXXX✓
RobotCar [28]✓✓✓X✓✓XXX✓XXX
UMA VI [54]✓✓XX✓X✓XXXXX✓
UMich [6]✓✓✓X✓XXXX✓XXX
KITTI [19]✓✓✓XXX✓X✓✓XXX
Virtual KITTI [16]✓✓✓X✓X✓X✓✓XXX
DARPA SubT [37]✓✓✓X✓X✓✓X✓XXX
Hilti SLAM [21]✓✓✓X✓X✓XXXXX✓
M3ED [7]✓✓✓X✓XXXX✓✓✓X
TartanAir [47]✓✓✓X✓✓✓X✓X✓XX
SubT-MRS (Ours)✓✓✓✓✓✓✓✓✓✓✓✓✓
", + "image_path": "01f3a26eaab9dae558cba6309166aa16b55eafc0d3c44dd7989cb8309dc889dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 228, + 465, + 239 + ], + "lines": [ + { + "bbox": [ + 125, + 228, + 465, + 239 + ], + "spans": [ + { + "bbox": [ + 125, + 228, + 465, + 239 + ], + "type": "text", + "content": "Table 1. Comparison of SLAM datasets on multi-sensors, multi robots, and multi degradation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 260, + 286, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 260, + 286, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 286, + 332 + ], + "type": "text", + "content": "in evaluating SLAM performance across varied robotic platforms. The AirMuseum dataset [13] includes drones and ground robots but lacks data from legged robots. As a comparison, SubT-MRS dataset features diverse robotic platforms including legged robots, aerial robots, and wheeled robots, operating in varied environments and sensor setups." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 337, + 286, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 286, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 286, + 564 + ], + "type": "text", + "content": "Extreme Environmental Conditions Developing and testing SLAM systems in extreme environmental conditions is crucial for mitigating potential real-world failures. However, most existing SLAM datasets have been primarily limited to single, controlled environments. The TUM-VI [39] and UMA-VI [54] datasets, being indoor-outdoor visual-inertial datasets, pose challenges due to varying illumination and low-texture environments. The EuRoC MAV [3] and UZH-FPV drone racing [12] datasets, popular in the SLAM community, offer data on aggressive drone movements but usually in consistently lit conditions. The KITTI dataset [19], a staple in autonomous driving research for its outdoor LiDAR-Visual-Inertial data. The Hilti SLAM dataset [21] includes both indoor and outdoor LiDAR-visual-inertial datasets with dynamic lighting and confined spaces. However, both of them focus more on accuracy than on a variety of environmental degradations. TartanAir [47] covers most of the challenging environments but it is a simulation dataset that will pose the sim-to-real gap." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": "In contrast, the SubT-MRS dataset provides 30 varied scenes, encompassing a wide array of environments. These include SubT tunnels, urban areas, caverns, multifloor structures, dark corridors, and foggy conditions. It also features textureless surfaces, dusty and smoky environments, off-road areas with aggressive motion, and snowy terrains prone to slippage. We include challenging lighting conditions such as extreme darkness and overexposure, along with geometrically challenging scenarios like featureless corridors, staircases, and self-similar cave layouts. Additionally, the dataset incorporates perceptual challenges posed by smoke, fog, dust, and various weather conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 258, + 420, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 420, + 271 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 420, + 271 + ], + "type": "text", + "content": "3. SubT-MRS Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 276, + 545, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 545, + 322 + ], + "type": "text", + "content": "We next present the SubT-MRS dataset from three aspects including its collecting environments and settings, ground truth collection, and the new robustness metric. A detailed comparison with other datasets is listed in Table 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 323, + 452, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 323, + 452, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 452, + 335 + ], + "type": "text", + "content": "3.1. All-weather Environments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 337, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 545, + 396 + ], + "type": "text", + "content": "As mentioned in Sec. 2, the SubT-MRS dataset distinguishes itself through its multimodal sensor setups, heterogeneous robotic platforms, and extreme environmental conditions. This section emphasizes this comprehensive nature pushing SLAM towards all-weather environments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 399, + 430, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 430, + 411 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 430, + 411 + ], + "type": "text", + "content": "3.1.1. Multimodal Sensors" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 413, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 472 + ], + "type": "text", + "content": "Multimodal sensors provide a wealth of information for SLAM systems operating in all-weather environments. This diversity improves scene understanding and strengthens the system's perception ability. Therefore, We incorporate diverse sensors to ensure system robustness." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 473, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 521 + ], + "type": "text", + "content": "Sensor Pack We embedded 4 Leopard Imaging RGB fisheye cameras, 1 Velodyne puck, 1 Epson M-G365 IMU, 1 FLIR Boson thermal camera, and NVIDIA Jetson AGX Xavier as a sensor pack shown in Figure 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 617 + ], + "type": "text", + "content": "Time Synchronization To ensure the overall consistency of the fused data, we meticulously synchronize the sensors' time using the 'pulse per second (PPS)' technique. As illustrated in Figure 2, the IMU, LiDAR, and thermal camera are directly synchronized with the CPU clock, while the four RGB camera are synchronized using an FPGA board. Consequently, we can effectively manage the time synchronization gap between each pair of sensors not to exceed 3ms." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Fisheye and Thermal Camera Calibration Camera calibration plays an essential role in the efficiency of a SLAM system. For calibrating the fisheye cameras, we employed the open-source toolkit Kalibr [36], focusing on the intrinsic and extrinsic parameters. Specifically, we use the radial-tangential distortion model to rectify the omnidirectional fisheye camera model. Calibrating thermal cameras, however, poses unique challenges, especially in gathering high-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22649" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 521, + 203 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 521, + 203 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 521, + 203 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 521, + 203 + ], + "type": "image", + "image_path": "4a5dcfa2e9f9e3cf0bef64f230e655cdef81aa0eee5146db6cfaca25eed1ea7d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 211, + 547, + 235 + ], + "lines": [ + { + "bbox": [ + 46, + 211, + 547, + 235 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 547, + 235 + ], + "type": "text", + "content": "Figure 2. An overview of the sensor pack used in SubT-MRS dataset. It is equipped with a Xavier processing unit with hardware time synchronization for multimodal sensors including LiDAR, fisheye cameras, thermal cameras, depth cameras (option), and an IMU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 255, + 287, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 290 + ], + "type": "text", + "content": "quality thermal data. To tackle this, we set up a " + }, + { + "bbox": [ + 46, + 255, + 287, + 290 + ], + "type": "inline_equation", + "content": "7 \\times 9" + }, + { + "bbox": [ + 46, + 255, + 287, + 290 + ], + "type": "text", + "content": " chessboard, heated by direct sunlight, to generate high-contrast thermal data, ensuring accuracy in the calibration process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 290, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 287, + 386 + ], + "type": "text", + "content": "IMU Calibration and Extrinsic Calibration For Lidar-IMU extrinsic calibration, we utilize the CAD model to obtain the calibration parameters. In the case of Camera-IMU extrinsic calibration, we employ the Kalibr toolbox [36] to estimate the extrinsic matrix. To estimate the sensors' bias and the random walk noise of the gyroscope and accelerometer, we collected static data from the IMU and calibrated it using an Allan variance-based IMU tool [14, 17]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 388, + 221, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 221, + 401 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 221, + 401 + ], + "type": "text", + "content": "3.1.2. Multi-degraded Environments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 402, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 437 + ], + "type": "text", + "content": "SubT-MRS includes multiple challenging environments including visually degraded environments, geometrically degraded environments, and their combination." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 437, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 568 + ], + "type": "text", + "content": "Visual Degradation Poor-quality visual features can significantly hinder the performance of feature extraction processes and disrupt the accuracy of feature matching. Such issues arise in low-light conditions with inconsistent brightness or image noise introduced by air obscurants. SubT-MRS encompasses these challenges and provides a wide range of visual degradation. This includes environments with limited lighting, such as hospital interiors and caves (Figure 3 A-F), as well as smoky or dusty conditions that cause visual obstruction (Figure 3 M-N), and snowy areas with reduced visibility (Figure 3 H, K, and L)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": "Geometric Degradation Lack of geometrical features poses significant challenges to LiDAR odometry. The root of this issue often lies in the limited sensing capabilities of LiDAR sensors and the constraints due to their mechanical installation. SubT-MRS captures a variety of environments that exemplify such challenges. It includes long, featureless corridors (Figure 3 E) and staircases (Figure 3 C and G). These scenarios illustrate various forms of geometric degradation, contributing to the improvement of LiDAR odometry systems in challenging conditions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Mixed Degradation A mixed visual and geometric degradation can further hinder the performance of SLAM sys" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "text", + "content": "tems. Examples in SubT-MRS include long, dimly lit corridors (Figure 3 A, E, and F), poorly illuminated staircases (Figure 3 C), and environments affected by snowy weather (Figure 3 H, K, and L). In these settings, both LiDAR odometry and visual odometry are prone to failure due to the compounded effects of mixed degradation. The inclusion of such scenarios in SubT-MRS is crucial for evaluating and improving the resilience of multimodal SLAM algorithms." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 354, + 487, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 487, + 366 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 487, + 366 + ], + "type": "text", + "content": "3.1.3. Heterogeneous Robot Platforms" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 368, + 545, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 545, + 463 + ], + "type": "text", + "content": "Most of current datasets focus on single-robot systems, limiting multi-robot SLAM development as shown in Table 1. To address this, we employed diverse robot platforms, including aerial, legged, and wheeled robots, navigating through various environments from urban campuses to medical facilities and natural terrains like caves. This diversity offers a range of scenarios to enhance SLAM algorithms for effective operation in challenging conditions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 464, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 561 + ], + "type": "text", + "content": "Extrinsic Calibration for Multiple Robots To ensure the multi-robot system shares a common coordinate system, we perform the extrinsic calibration process [38]. It has 2 steps: First, each robot runs Super Odometry [53] to generate its local map using LiDAR data and share it with other robots through a wireless network; Second, the remaining robots identify overlapping regions between their local maps and estimate the extrinsic parameters using GICP [40]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 566, + 395, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 566, + 395, + 578 + ], + "spans": [ + { + "bbox": [ + 306, + 566, + 395, + 578 + ], + "type": "text", + "content": "3.2. Ground Truth" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Ground Truth Map As depicted in Figure 4, we utilized the high-precision FARO Focus 3D S120 3D scanner for creating ground-truth maps. This cutting-edge 3D scanner can measure distances up to " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "120\\mathrm{m}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", with a maximum measurement rate of " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "976\\mathrm{K}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " points per second. The accuracy of the ground truth map is noteworthy, maintaining a range error within " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pm 2\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ". We have developed ground truth models for diverse environments, including subterranean areas like urban areas " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "(350\\mathrm{m} \\times 350\\mathrm{m})" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", caves " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "(150\\mathrm{m} \\times 200\\mathrm{m})" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and tunnels " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "(100 \\times 200\\mathrm{m})" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", as well as indoor-outdoor spaces " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "(200\\mathrm{m} \\times 200\\mathrm{m})" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " within the CMU campus. To reduce drift, a" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22650" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 545, + 325 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 325 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 325 + ], + "type": "image", + "image_path": "e99cad20d18ec8f0048688cceaa90f07e28c3d2bab5ffea665868a82026408ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 333, + 548, + 378 + ], + "lines": [ + { + "bbox": [ + 46, + 333, + 548, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 548, + 378 + ], + "type": "text", + "content": "Figure 3. The SubT-MRS datasets were collected across diverse seasons, capturing environments with perceptual challenges such as poor illumination, darkness, and water puddles, where visual sensors falter. They also include geometrically complex areas like long featureless corridors and steep multi-floor structures, challenging LiDAR systems with potential drift. Moreover, these datasets cover conditions with airborne obscurants like dust, fog, snow, and smoke in tough environments, including caves, deserts, long tunnels, and off-road areas." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 72, + 389, + 521, + 550 + ], + "blocks": [ + { + "bbox": [ + 72, + 389, + 521, + 550 + ], + "lines": [ + { + "bbox": [ + 72, + 389, + 521, + 550 + ], + "spans": [ + { + "bbox": [ + 72, + 389, + 521, + 550 + ], + "type": "image", + "image_path": "3c0d36f7d01aeff96cc49b14894396b1d84ab5655b90321852f2810c4bddc139.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 559, + 547, + 594 + ], + "lines": [ + { + "bbox": [ + 46, + 559, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 547, + 594 + ], + "type": "text", + "content": "Figure 4. The SubT-MRS dataset facilitates the generation of high-precision ground truth maps and trajectories. Figure (a) shows the ground truth trajectory in multi-floor settings. Figure (b) displays ground truth maps for indoor and outdoor areas, encompassing long corridors, multi-floor structures, and open spaces. Figure (c) features photo-realistic scans based on our ground truth maps in cave environments." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 613, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 288, + 662 + ], + "type": "text", + "content": "loop closure algorithm was applied to correct the poses. Notably, " + }, + { + "bbox": [ + 46, + 613, + 288, + 662 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 46, + 613, + 288, + 662 + ], + "type": "text", + "content": " of the scans achieved a position uncertainty of less than " + }, + { + "bbox": [ + 46, + 613, + 288, + 662 + ], + "type": "inline_equation", + "content": "2\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 613, + 288, + 662 + ], + "type": "text", + "content": ". This ground-truth map is crucial for SLAM development where precision is of utmost importance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Ground Truth Trajectory Ground truth trajectories for all sequences are generated based on the ground truth map. In this process, we establish point-to-point, point-to-plane, and point-to-line correspondences[40] between the Ground" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 613, + 547, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 547, + 711 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 547, + 711 + ], + "type": "text", + "content": "Truth map and the current LiDAR scan. Our estimated ground truth trajectory is a comprehensive fusion of various constraints derived from the ground truth map, visual odometry, LiDAR odometry, and IMU measurements. This approach ensures the robustness of our solution across diverse environmental conditions. To verify accuracy, we reconstruct LiDAR maps from ground truth trajectory and compare it with our ground truth maps." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22651" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 91, + 553, + 206 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 533, + 82 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 533, + 82 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 533, + 82 + ], + "type": "text", + "content": "Table 2. SLAM Challenge Results (Blue shadings indicate rankings of ATE and Robustness Metric; L: LiDAR I: IMU C: Camera)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 91, + 553, + 206 + ], + "lines": [ + { + "bbox": [ + 49, + 91, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 91, + 553, + 206 + ], + "type": "table", + "html": "
#TeamMethodOdometry TypeDeviceRealTime (s)CPU/GPU (%)RAM (GB)ATE↓\\( {R}_{v} \\uparrow /{R}_{w} \\uparrow \\)Sensors L I C
1Liu et alFAST-LIO2 [51], HBA [27]FilterIntel i7-9700K51.31098.667 / 04.0520.5880.517/0.770✓✓
2Yibin et alLIO-EKF [45]FilterIntel i7-107000.00652.167 / 00.0724.3130.441/0.574✓✓
3Weitong et alFAST-LIO[2], Pose Graph[11]FilterIntel Xeon(R)E3-1240v50.12522.63 / 04.3050.6630.473/0.747✓✓
4Kim et alFAST-LIO2[50], Point-LIO[20], Quatro[25]FilterIntel i5-125000.268101.108 / 055.643.8250.479/0.615✓✓
5Zhong et alDLO[8], Scan-Context++[22]SW OptAMD Ryzen 9 5900x0.02713.289 / 01.1741.2090.276/0.486✓✓
1Peng et alDVI-SLAM [32]LearningIntel i9-12900183.233- / 14911 (4)0.5470.473/0.788✓✓
2Jiang et alLET-NET[26], VINS-Mono[34]HybridIntel i5-94000.06440.35 / 04.3371.0930.078/0.322✓✓
3Thien et alVR-SLAM[31]SW OptIntel i9-129000.142176.44 / 09.1113.0370.083/0.372✓✓
4Li et alORB-SLAM3[4]SW Opt.Intel i7-107000.01965.028 / 00.3868.9750.163/0.474✓✓
", + "image_path": "fdec7445e1994ac9e22a65403478b80a8d9b0efe3e082ff2af469ac640c8c455.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 237, + 541, + 327 + ], + "blocks": [ + { + "bbox": [ + 55, + 217, + 536, + 228 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 536, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 536, + 228 + ], + "type": "text", + "content": "Table 3. Accuracy Performance on Geometric Degradation. Red numbers represent ATE ranking. * denotes incomplete submissions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 237, + 541, + 327 + ], + "lines": [ + { + "bbox": [ + 50, + 237, + 541, + 327 + ], + "spans": [ + { + "bbox": [ + 50, + 237, + 541, + 327 + ], + "type": "table", + "html": "
TeamGeometric Degradation (Real World)SimulationMix DegradationAverage
UrbanTunnelCaveNuclear_1Nuclear_2Laurel_CavernsFactoryOceanSewerageLong CorridorMulti FloorBlock Lidar
Liu et al10.3070.0950.6290.1220.2350.2600.8890.7570.9781.4540.4010.9340.588
Weitong et al20.260.0960.6170.1200.2220.4020.9980.7701.5861.2540.5771.0560.663
Kim et al30.3310.0920.7870.1230.2700.27910.62822.4257.1472.1000.6501.0683.825
Yibin et al41.0600.2200.7500.4700.6209.1404.9200.28024.4602.9905.5001.3404.312
Zhong et al51.2050.695-1.1751.722.080.8890.7781.13---1.209*
Average0.6330.2400.6960.4020.61342.4323.6655.0027.0601.9501.7821.099
", + "image_path": "2ed6115a3035f29a5745994d8f0ae28f3eb62609250d5bd3571af50b6022515a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 346, + 157, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 157, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 157, + 357 + ], + "type": "text", + "content": "3.3. Robustness Metric" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "text", + "content": "As discussed in Sec. 1, existing evaluation metrics such as absolute trajectory error (ATE) [3] have limitations in evaluating the SLAM's robustness in real-world applications. ATE primarily focuses on the accuracy of trajectory and does not consider the completeness of trajectory (recall). Also, it can not effectively capture velocity changes that have a direct impact on the robot's safety. For example, a pose spike error might not be immediately noticeable in ATE evaluation, it may result in a crash in an aerial robot control. To address these gaps, we introduce a new robustness metric based on estimated velocity. That is because velocity estimation from SLAM is crucial for the robot's control system which directly impacts stability, affecting the robot's safety. The new robustness metric is the area under the curve (AUC) of the F-1 score:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 548, + 287, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 548, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 97, + 548, + 287, + 574 + ], + "type": "interline_equation", + "content": "F _ {1} (e) = \\frac {2 P (e < T) R (e < T)}{P (e < T) + R (e < T)}, \\tag {1}", + "image_path": "8b94c39cb19e6df03174a29af8c2b47f40bcb44dc20cd69fe2b86df2c44912c0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "where the precision " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " quantifies the precision of the estimated velocity as a percentage: how closely the estimated velocity points lie to the ground truth point and the recall rate " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " quantifies the estimated velocity's completeness: to what extent all the ground-truth points are covered. A high F-1 score can only be achieved by the velocity estimation that is both accurate and complete throughout the entire run. Specifically, an estimated error " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " is regarded as an inlier, if it is smaller than a threshold " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ". To scale the threshold " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " within the range of [0, 1], we apply exponential mapping " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\exp(-10T)" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " when calculating robustness metric." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 346, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 490 + ], + "type": "text", + "content": "Position Robustness " + }, + { + "bbox": [ + 304, + 346, + 545, + 490 + ], + "type": "inline_equation", + "content": "R_{p}" + }, + { + "bbox": [ + 304, + 346, + 545, + 490 + ], + "type": "text", + "content": " & Rotation Robustness " + }, + { + "bbox": [ + 304, + 346, + 545, + 490 + ], + "type": "inline_equation", + "content": "R_{r}" + }, + { + "bbox": [ + 304, + 346, + 545, + 490 + ], + "type": "text", + "content": " The AUC of F-1 score can be defined using both linear velocity and angular velocity, which can reflect the robustness of position and rotation estimation, respectively. Specifically, we define " + }, + { + "bbox": [ + 304, + 346, + 545, + 490 + ], + "type": "inline_equation", + "content": "R_{p} = \\mathrm{AUC}(F_{1}(\\mathrm{v}_{\\mathrm{e}}))" + }, + { + "bbox": [ + 304, + 346, + 545, + 490 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 346, + 545, + 490 + ], + "type": "inline_equation", + "content": "R_{r} = \\mathrm{AUC}(F_{1}(\\omega_{\\mathrm{e}}))" + }, + { + "bbox": [ + 304, + 346, + 545, + 490 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 346, + 545, + 490 + ], + "type": "inline_equation", + "content": "v_{e}" + }, + { + "bbox": [ + 304, + 346, + 545, + 490 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 346, + 545, + 490 + ], + "type": "inline_equation", + "content": "\\omega_{e}" + }, + { + "bbox": [ + 304, + 346, + 545, + 490 + ], + "type": "text", + "content": " are the estimated error of linear velocity and rotation velocity, respectively. Note that not all SLAM solutions can output the velocity at the desired frequency. To address this, we use B-splines [30, 46] to obtain smooth trajectories, compute derivatives for smooth trajectories, and derive estimated linear velocity and angular velocity. For more details, please refer to the supplementary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 498, + 481, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 481, + 511 + ], + "type": "text", + "content": "4. Open SLAM Challenge Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 514, + 545, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 538 + ], + "type": "text", + "content": "In this section, we will present our works from two perspectives: accuracy evaluation and robustness evaluation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 544, + 427, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 544, + 427, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 544, + 427, + 555 + ], + "type": "text", + "content": "4.1. Accuracy Evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "The results of the ICCV SLAM challenge underscore the necessity for advancements in system robustness. From 29 submissions, we identified 5 winners in the LiDAR category and 4 in the visual category. However, in the sensor fusion track, which addresses both visual and geometric degradation, no submissions met the criteria for success. This result reveals the existing SLAM systems still have lots of space to improve. Sequence characteristics will be detailed in the supplementary material. Table 2 shows the results for both the LiDAR and visual tracks. Unfortunately, there are no current solutions that can balance high accuracy and real-time performance in challenging environments. Since several environments lack geometric features or visual features," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22652" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 91, + 518, + 169 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 528, + 82 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 528, + 82 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 528, + 82 + ], + "type": "text", + "content": "Table 4. Accuracy Performance on Visual Degradation. Red numbers represent ATE ranking. * denotes incomplete submissions." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 91, + 518, + 169 + ], + "lines": [ + { + "bbox": [ + 73, + 91, + 518, + 169 + ], + "spans": [ + { + "bbox": [ + 73, + 91, + 518, + 169 + ], + "type": "table", + "html": "
TeamVisual Degradation (Real World)SimulationAverage
Lowlight 1Lowlight 2Over ExposureFlash LightSmoke RoomOutdoor NightEnd of WorldMoonWestern Desert
Peng et al11.0631.6370.5030.440.1530.8270.0380.1950.0700.547
Thien et al21.0812.0541.7331.05410.5327.6920.7531.2281.2093.037
Jiang et al31.0191.1261.9112.3413.75711.8212.1540.6044.0103.193
Li et al45.7687.8341.7571.2955.37010.766-30.07-8.98*
Average2.2323.1631.4761.2824.9537.7760.9828.0241.763
", + "image_path": "811b48b0e0fffb482362b1dad99129a7b483227a851a403173d18320f4a4f3dc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 85, + 178, + 194, + 285 + ], + "blocks": [ + { + "bbox": [ + 85, + 178, + 194, + 285 + ], + "lines": [ + { + "bbox": [ + 85, + 178, + 194, + 285 + ], + "spans": [ + { + "bbox": [ + 85, + 178, + 194, + 285 + ], + "type": "image", + "image_path": "7b8e8b8ab874deecdf5e7d6a75971d02e8e9b765ff81b297d02a906c91599ca0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 294, + 547, + 328 + ], + "lines": [ + { + "bbox": [ + 46, + 294, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 547, + 328 + ], + "type": "text", + "content": "Figure 5. From left to right, it shows robustness metric " + }, + { + "bbox": [ + 46, + 294, + 547, + 328 + ], + "type": "inline_equation", + "content": "R_{p}" + }, + { + "bbox": [ + 46, + 294, + 547, + 328 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 294, + 547, + 328 + ], + "type": "inline_equation", + "content": "R_{r}" + }, + { + "bbox": [ + 46, + 294, + 547, + 328 + ], + "type": "text", + "content": " for LiDAR and visual sequences respectively. Note: This is a summary of results for all sequences, with weights based on the trajectory length. The area under the curve (AUC) represents the robustness " + }, + { + "bbox": [ + 46, + 294, + 547, + 328 + ], + "type": "inline_equation", + "content": "(R_{p}, R_{r})" + }, + { + "bbox": [ + 46, + 294, + 547, + 328 + ], + "type": "text", + "content": ". The x-axis shows velocity thresholds for classifying estimated velocities as inliers and the y-axis is F-1 score." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 195, + 179, + 295, + 285 + ], + "blocks": [ + { + "bbox": [ + 195, + 179, + 295, + 285 + ], + "lines": [ + { + "bbox": [ + 195, + 179, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 195, + 179, + 295, + 285 + ], + "type": "image", + "image_path": "55cfc6031a312bd26283cd3f30215cbe3a40100a153c8fd52f3f5b533cf7c593.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 296, + 179, + 403, + 285 + ], + "blocks": [ + { + "bbox": [ + 296, + 179, + 403, + 285 + ], + "lines": [ + { + "bbox": [ + 296, + 179, + 403, + 285 + ], + "spans": [ + { + "bbox": [ + 296, + 179, + 403, + 285 + ], + "type": "image", + "image_path": "1ddcbb645616be2d30870b196461b2b337a9d4108afc101e9f4d046bbd9ca7d8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 405, + 179, + 509, + 285 + ], + "blocks": [ + { + "bbox": [ + 405, + 179, + 509, + 285 + ], + "lines": [ + { + "bbox": [ + 405, + 179, + 509, + 285 + ], + "spans": [ + { + "bbox": [ + 405, + 179, + 509, + 285 + ], + "type": "image", + "image_path": "ddea1daf4391967666e91c1708573d8df339b08b3c6cbb9141b9eeb3029f5647.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 340, + 288, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 288, + 424 + ], + "type": "text", + "content": "the algorithms required more processing times to achieve reasonably high accuracy. Liu et al. at the University of Hong Kong, who achieved the highest accuracy, recorded an Absolute Trajectory Error (ATE) of " + }, + { + "bbox": [ + 46, + 340, + 288, + 424 + ], + "type": "inline_equation", + "content": "0.588\\mathrm{m}" + }, + { + "bbox": [ + 46, + 340, + 288, + 424 + ], + "type": "text", + "content": ", but each iteration took 51.3 seconds. In the visual track, Peng et al. from the Samsung Research Center attained a leading ATE of " + }, + { + "bbox": [ + 46, + 340, + 288, + 424 + ], + "type": "inline_equation", + "content": "0.547\\mathrm{m}" + }, + { + "bbox": [ + 46, + 340, + 288, + 424 + ], + "type": "text", + "content": ", but each iteration required 183 seconds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 425, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 605 + ], + "type": "text", + "content": "LiDAR Track Discussion To enhance robustness in all-weather environments, a pivotal question arises: what are the limits of current LiDAR SLAM algorithms? Table 3 presents a summary of the ATE/RPE errors observed in real-world geometrically degraded, simulated, and mixed degradation scenarios across the top five teams. In real-world geometrically degraded environments, we observed the first limitation: existing LiDAR solutions struggle in confined spaces such as caves. Almost all algorithms perform worse in Cave (Figure 3S) and Laureal_Cavern environments (Figure 3N) with average ATE 0.696 and 2.432 meters respectively. One reason is that cave environments is the most confined space which may lack sufficient geometric features, compared with other environments like tunnel (Figure 3J) and Urban (Figure 3B) and Nuclear scenes (Figure 3U)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "In simulated environments, we encountered the second limitation: existing methods are tailored for fixed motion patterns like vehicles and struggle with unpredictable motion patterns. Since our simulation sequence is derived from TartanAir [47], featuring aggressive and random motion patterns, it might not follow the usual velocity distribution domain and has not been thoroughly tested. As shown in Table 3 (simulation), it resulted in significantly higher ATE errors " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "(5.239\\mathrm{m})" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " compared to real-world's " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "(0.836\\mathrm{m})" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 340, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 555 + ], + "type": "text", + "content": "In mixed degradation environments, the third limitation we found is that existing methods cannot actively select the most informative measurements to adapt to new environments. For example, in the Long Corridor sequence (Figure 3 A) with low lighting and a featureless environment, the scenario is intricately designed for collaborative usage of LiDAR and Visual sensors. Algorithms are expected to utilize visual information in geometrically degraded environments while disregarding it in darkness. Similarly, the Block LiDAR sequences (Figure 3 V) aim to simulate sensor drop scenarios frequently encountered in robotic applications. These sequences involve alternating periods of LiDAR and Visual data loss, challenging algorithms to promptly detect sensor failures and switch to the other modality for SLAM. However, the average ATE error " + }, + { + "bbox": [ + 304, + 340, + 545, + 555 + ], + "type": "inline_equation", + "content": "(1.61\\mathrm{m})" + }, + { + "bbox": [ + 304, + 340, + 545, + 555 + ], + "type": "text", + "content": " in Table 3 (mixed degradation) is much higher than in geometric degradation scenarios " + }, + { + "bbox": [ + 304, + 340, + 545, + 555 + ], + "type": "inline_equation", + "content": "(0.836\\mathrm{m})" + }, + { + "bbox": [ + 304, + 340, + 545, + 555 + ], + "type": "text", + "content": ", suggesting the above limitations. Details are in the supplementary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 557, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 714 + ], + "type": "text", + "content": "Visual Track Discussion To improve the robustness, a pivotal question arises: what are the limits of current visual SLAM? To what extent does image quality impact a visual SLAM system? Table 4 shows a summary of ATE errors from real-world and simulation scenarios from the awarded 4 teams. In real-world scenarios, the first limitation is the lack of anti-noise capability in current methods, especially in low-quality image settings. We assessed visual odometry accuracy in visually degraded environments, including low lighting, sunlight overexposure, flashing lights, smoke-filled rooms, and nighttime outdoor settings (Figure 3). Trajectories in conditions like the Smoke Room and Outdoor Night exhibit significantly higher average ATE er" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22653" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 88, + 523, + 220 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 545, + 82 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 545, + 82 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 545, + 82 + ], + "type": "text", + "content": "Table 5. Robustness Performance on Geometric Degradation. Red numbers are robustness ranking. Larger values indicate better robustness." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 88, + 523, + 220 + ], + "lines": [ + { + "bbox": [ + 69, + 88, + 523, + 220 + ], + "spans": [ + { + "bbox": [ + 69, + 88, + 523, + 220 + ], + "type": "table", + "html": "
TeamGeometric Degradation (Real World)SimulationMix DegradationAverage
UrbanTunnelCaveNuclear_1Nuclear_2Laurel_CavernsFactoryOceanSewerageLong CorridorMulti FloorBlock Lidar
Liu et al1Rp↑0.8110.8650.7360.7470.5040.8160.1570.1350.1350.3960.5290.3710.516
Weitong et al20.7730.8380.6900.8710.7390.1520.1530.1490.0810.4070.4560.3450.471
Kim et al40.7470.8650.7370.8700.6890.7770.0140.0010.0100.4100.2850.3450.479
Yibin et al30.6500.7830.5570.7210.4900.4810.2190.5600.0190.1750.2600.3700.440
Zhong et al50.5670.6830.2040.6800.4260.4470.1030.0840.1220000.276
Liu et al1Rr↑0.8880.8930.8160.8570.7780.8610.6920.7190.6520.7530.6890.6430.770
Weitong et al20.8860.8920.8110.8930.8400.5980.6880.7200.5140.7660.7060.6440.746
Kim et al40.6800.8930.8160.8930.8030.8080.1980.1250.3000.8020.4220.6420.615
Yibin et al30.6240.7070.4620.7330.5400.4670.8620.9620.4960.4770.3420.2130.573
Zhong et al50.7310.7450.4370.7570.6030.5230.6880.7040.6430000.485
AverageRp0.7100.8070.5850.7780.5700.5350.1300.1860.0730.2780.3060.286
Average Rr0.7620.8260.6680.8270.7130.6510.6260.6460.5210.5600.4320.428
", + "image_path": "673a090df4bcba9895606a892977e11654621b21cdf8b4bf2da0799f720b5c65.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 65, + 247, + 523, + 373 + ], + "blocks": [ + { + "bbox": [ + 51, + 231, + 541, + 242 + ], + "lines": [ + { + "bbox": [ + 51, + 231, + 541, + 242 + ], + "spans": [ + { + "bbox": [ + 51, + 231, + 541, + 242 + ], + "type": "text", + "content": "Table 6. Robustness Performance on Visual Degradation. Red numbers are robustness ranking. Larger values indicate better robustness." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 247, + 523, + 373 + ], + "lines": [ + { + "bbox": [ + 65, + 247, + 523, + 373 + ], + "spans": [ + { + "bbox": [ + 65, + 247, + 523, + 373 + ], + "type": "table", + "html": "
TeamVisual Degradation (Real World)SimulationAverage
Lowlight 1Lowlight 2Over ExposureFlash LightSmoke RoomOutdoor NightEnd of WorldMoonWestern Desert
Peng et al10.3570.2270.2640.2030.5360.2700.6990.8930.8060.472
Thien et al30.0450.0700.2400.1560.1310.07500.03100.083
Jiang et al4Rp↑0.0460.0390.1940.0880.2420.0950000.078
Li et al20.3420.1870.2570.2080.3220.14200.00600.162
Peng et al10.6410.5810.7440.6500.8780.6700.9750.9750.9740.787
Thien et al30.4130.4450.6100.2690.4740.4870.1770.31500.354
Jiang et al4Rr↑0.4530.4520.6190.2520.5420.5770.00200.1570.339
Li et al20.6420.5740.6570.6510.7730.6600.00.30500.473
AverageRp0.1980.1310.2390.1640.3080.1460.1750.2320.202
Average Rr0.5370.5130.6580.4560.6670.5980.2880.3990.283
", + "image_path": "0554e414eda0c184fb5086e9c578d4cb0574e8ba3786ffd0e627be43dfee6a5e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 384, + 287, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 481 + ], + "type": "text", + "content": "rors (4.95m, 7.76m respectively). Smoke and night scenes challenge feature extraction and tracking and introduce sensor noise, emphasizing the need for robust anti-noise algorithms. The second limitation is that existing methods struggle to overcome aggressive motion. Even in simulations with relatively good image quality, most methods still show significant average ATE errors (8.024m) on the Moon sequence (Figure 3 Y). See more details in supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 484, + 177, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 177, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 177, + 495 + ], + "type": "text", + "content": "4.2. Robustness Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "type": "text", + "content": "The F-1 robustness curve provides a detailed evaluation under different error tolerances, while " + }, + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "type": "inline_equation", + "content": "R_{p}" + }, + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "type": "inline_equation", + "content": "R_{r}" + }, + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "type": "text", + "content": " provide an overall measure of robustness that is not dependent on a specific decision threshold. This suggests the flexibility of our new robustness metric. Figure 5 clearly shows the robustness performance of all teams regarding position and rotation, highlighting the differences in robustness between the methods. The values of " + }, + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "type": "inline_equation", + "content": "R_{p}" + }, + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "type": "inline_equation", + "content": "R_{r}" + }, + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "type": "text", + "content": ", which are the AUC under F-1 curves, are displayed in the brackets of the legends. It reveals that Liu et al. and Peng et al. are the most robust solutions for LiDAR and visual tracks, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "Is the robustness metric robust? We did extensive experiments on our robustness metric against diverse sensor degradation settings. Table 5 and 6 present a summary of robustness performance in real-world, simulated, and mixed degradation scenarios for both LiDAR and visual sequences. In Table 5, we observe that the average values of " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "R_{p}" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "R_{r}" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": " for the mixed degradation environment (0.290," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 384, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 456 + ], + "type": "text", + "content": "0.473) are significantly lower than geometric degradation environments (0.664, 0.754). This observation suggests that the majority of SLAM algorithms exhibit reduced robustness in mixed degradation environments as compared to geometrically degraded ones. This aligns with our expectations and verifies the effectiveness of the robustness metric." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 460, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 568 + ], + "type": "text", + "content": "Our robustness ranking, indicated by small red numbers, differs from those on ATE ranking shown in Table 3 and 4. This is because our F-1 score-based metrics " + }, + { + "bbox": [ + 304, + 460, + 545, + 568 + ], + "type": "inline_equation", + "content": "R_{p}" + }, + { + "bbox": [ + 304, + 460, + 545, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 460, + 545, + 568 + ], + "type": "inline_equation", + "content": "R_{r}" + }, + { + "bbox": [ + 304, + 460, + 545, + 568 + ], + "type": "text", + "content": " jointly consider precision and recall rate to provide a balanced evaluation of SLAM performance, considering both metrics across the full spectrum of thresholds. In contrast, the ATE focuses solely on precision, neglecting the trajectory's completeness (recall). We also evaluate our metric on various synthetic trajectories in the supplementary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "We introduce SubT-MRS, a comprehensive SLAM dataset with various sensor data, locomotion patterns, and over 30 degradation types in simulation and real-world settings to push SLAM towards all-weather environments. Additionally, we introduce a new robustness metric to evaluate the reliability of SLAM systems, enhancing robot safety control. 29 teams have tested SubT-MRS in the organized SLAM challenge and we expect that it will serve as a critical benchmark for future SLAM development." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22654" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 87, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 87, + 258, + 109 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 87, + 258, + 109 + ], + "spans": [ + { + "bbox": [ + 53, + 87, + 258, + 109 + ], + "type": "text", + "content": "[1] https://www.darpa.mil/program/darpa-subterranean-challenge. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 110, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 110, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 53, + 110, + 287, + 163 + ], + "type": "text", + "content": "[2] Chunge Bai, Tao Xiao, Yajie Chen, Haoqian Wang, Fang Zhang, and Xiang Gao. Faster-lio: Lightweight tightly coupled lidar-inertial odometry using parallel sparse incremental voxels. IEEE Robotics and Automation Letters, 7(2):4861-4868, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 166, + 288, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 166, + 288, + 220 + ], + "spans": [ + { + "bbox": [ + 53, + 166, + 288, + 220 + ], + "type": "text", + "content": "[3] Michael Burri, Janosch Nikolic, Pascal Gohl, Thomas Schneider, Joern Rehder, Sammy Omari, Markus W Achtelik, and Roland Siegwart. The EuRoC micro aerial vehicle datasets. The International Journal of Robotics Research, 2016. 1, 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 222, + 288, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 222, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 53, + 222, + 288, + 275 + ], + "type": "text", + "content": "[4] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 277, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 277, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 53, + 277, + 287, + 331 + ], + "type": "text", + "content": "[5] Carlos Campos, Richard Elvira, Juan J. Gómez Rodríguez, José M. M. Montiel, and Juan D. Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6): 1874-1890, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 333, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 333, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 53, + 333, + 287, + 376 + ], + "type": "text", + "content": "[6] Nicholas Carlevaris-Bianco, Arash K Ushani, and Ryan M Eustice. University of michigan north campus long-term vision and lidar dataset. The International Journal of Robotics Research, 35(9):1023-1035, 2016. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 378, + 287, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 378, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 53, + 378, + 287, + 444 + ], + "type": "text", + "content": "[7] Kenneth Chaney, Fernando Cladera, Ziyun Wang, Anthony Bisulco, M Ani Hsieh, Christopher Korpela, Vijay Kumar, Camillo J Taylor, and Kostas Daniilidis. M3ed: Multi-robot, multi-sensor, multi-environment event dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015–4022, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 445, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 445, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 53, + 445, + 287, + 488 + ], + "type": "text", + "content": "[8] Kenny Chen, Brett T. Lopez, Ali-akbar Agha-mohammadi, and Ankur Mehta. Direct lidar odometry: Fast localization with dense point clouds. IEEE Robotics and Automation Letters, 7(2):2000–2007, 2022. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 490, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 490, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 490, + 287, + 544 + ], + "type": "text", + "content": "[9] Long Chen, Yuchen Li, Chao Huang, Bai Li, Yang Xing, Daxin Tian, Li Li, Zhongxu Hu, Xiaoxiang Na, Zixuan Li, et al. Milestones in autonomous driving and intelligent vehicles: Survey of surveys. IEEE Transactions on Intelligent Vehicles, 8(2):1046-1056, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 599 + ], + "type": "text", + "content": "[10] Weichen Dai, Yu Zhang, Shenzhou Chen, Donglei Sun, and Da Kong. A multi-spectral dataset for evaluating motion estimation systems. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 5560-5566. IEEE, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "type": "text", + "content": "[11] Frank Dellaert, Michael Kaess, et al. Factor graphs for robot perception. Foundations and Trends® in Robotics, 6(1-2): 1-139, 2017. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[12] Jeffrey Delmerico, Titus Cieslewski, Henri Rebecq, Matthias Faessler, and Davide Scaramuzza. Are we ready for autonomous drone racing? the UZH-FPV drone racing dataset. In IEEE Int. Conf. Robot. Autom. (ICRA), 2019. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[13] Rodolphe Dubois, Alexandre Eudes, and Vincent Frémont. Airmuseum: a heterogeneous multi-robot dataset for stereovisual and inertial simultaneous localization and mapping. In" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "text", + "content": "2020 IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI), pages 166-172, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 149 + ], + "type": "text", + "content": "[14] Naser El-Sheimi, Haiying Hou, and Xiaoji Niu. Analysis and modeling of inertial sensors using allan variance. IEEE Transactions on Instrumentation and Measurement, 57(1): 140-149, 2008. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[15] Dapeng Feng, Yuhua Qi, Shipeng Zhong, Zhiqiang Chen, Yudu Jiao, Qiming Chen, Tao Jiang, and Hongbo Chen. S3e: A large-scale multimodal dataset for collaborative slam. arXiv preprint arXiv:2210.13723, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 198, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 240 + ], + "type": "text", + "content": "[16] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4340-4349, 2016. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 243, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 262 + ], + "type": "text", + "content": "[17] Wenliang Gao. A ros package tool to analyze the imu performance, 2018. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 265, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 545, + 308 + ], + "type": "text", + "content": "[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI Vision Benchmark Suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 310, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 352 + ], + "type": "text", + "content": "[19] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 355, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 398 + ], + "type": "text", + "content": "[20] Dongjiao He, Wei Xu, Nan Chen, Fanze Kong, Chongjian Yuan, and Fu Zhang. Point-lio: Robust high-bandwidth light detection and ranging inertial odometry. Advanced Intelligent Systems, page 2200459, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 400, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 400, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 400, + 545, + 442 + ], + "type": "text", + "content": "[21] Michael Helmberger, Kristian Morin, Beda Berner, Nitish Kumar, Giovanni Cioffi, and Davide Scaramuzza. The hilti slam challenge dataset. IEEE Robotics and Automation Letters, 7(3):7518-7525, 2022. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "type": "text", + "content": "[22] Giseop Kim and Ayoung Kim. Scan context: Egocentric spatial descriptor for place recognition within 3d point cloud map. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4802-4809. IEEE, 2018. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "text", + "content": "[23] Alex Junho Lee, Younggun Cho, Young-sik Shin, Ayoung Kim, and Hyun Myung. Vivid++: Vision for visibility dataset. IEEE Robotics and Automation Letters, 7(3):6282-6289, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 546, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 545, + 588 + ], + "type": "text", + "content": "[24] Keith YK Leung, Yoni Halpern, Timothy D Barfoot, and Hugh HT Liu. The utias multi-robot cooperative localization and mapping dataset. The International Journal of Robotics Research, 30(8):969–974, 2011. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "type": "text", + "content": "[25] Hyungtae Lim, Suyong Yeon, Suyong Ryu, Yonghan Lee, Youngji Kim, Jaeseong Yun, Euigon Jung, Donghwan Lee, and Hyun Myung. A single correspondence is enough: Robust global registration to avoid degeneracy in urban environments. page Accepted. To appear, 2022. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 678 + ], + "type": "text", + "content": "[26] Yicheng Lin, Shuo Wang, Yunlong Jiang, and Bin Han. Breaking of brightness consistency in optical flow with a lightweight cnn network, 2023. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[27] Zheng Liu, Xiyuan Liu, and Fu Zhang. Efficient and consistent bundle adjustment on lidar point clouds. IEEE Transactions on Robotics, 2023. 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22655" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[28] Will Maddern, Geoffrey Pascoe, Chris Linegar, and Paul Newman. 1 year, " + }, + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "inline_equation", + "content": "1000\\mathrm{km}" + }, + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": ": The oxford robotcar dataset. The International Journal of Robotics Research, 36(1):3-15, 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 171 + ], + "type": "text", + "content": "[29] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. The International Journal of Robotics Research, 36(2):142-149, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 215 + ], + "type": "text", + "content": "[30] Elias Mueggler, Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. Continuous-time visual-inertial odometry for event cameras. IEEE Transactions on Robotics, 34(6): 1425-1440, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 258 + ], + "type": "text", + "content": "[31] Thien Hoang Nguyen, Shenghai Yuan, and Lihua Xie. Vrslam: A visual-range simultaneous localization and mapping system using monocular camera and ultra-wideband sensors, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "type": "text", + "content": "[32] Xiongfeng Peng, Zhihua Liu, Weiming Li, Ping Tan, SoonYong Cho, and Qiang Wang. Dvi-slam: A dual visual inertial slam network, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 287, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 347 + ], + "type": "text", + "content": "[33] Bernd Pfrommer, Nitin Sanket, Kostas Daniilidis, and Jonas Cleveland. Penncosyvio: A challenging visual inertial odometry benchmark. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 3847-3854. IEEE, 2017. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "text", + "content": "[34] Tong Qin, Peiliang Li, and Shaojie Shen. Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics, 34(4):1004-1020, 2018. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 458 + ], + "type": "text", + "content": "[35] Milad Ramezani, Yiduo Wang, Marco Camurri, David Wisth, Matias Mattamala, and Maurice Fallon. The newer college dataset: Handheld lidar, inertial and vision with ground truth. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4353-4360. IEEE, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "text", + "content": "[36] Joern Rehder, Janosch Nikolic, Thomas Schneider, Timo Hinzmann, and Roland Siegwart. Extending kalibr: Calibrating the extrinsic of multiple imus and of individual axes. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 4304-4311. IEEE, 2016. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 515, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 569 + ], + "type": "text", + "content": "[37] John G Rogers, Arthur Schang, Carlos Nieto-Granda, John Ware, John Carter, Jonathan Fink, and Ethan Stump. The darpa subt urban circuit mapping dataset and evaluation metric. In Experimental Robotics: The 17th International Symposium, pages 391–401. Springer, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 571, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 623 + ], + "type": "text", + "content": "[38] Sebastian Scherer, Vasu Agrawal, Graeme Best, Chao Cao, Katarina Cujic, R Darnley, R DeBortoli, E Dexheimer, B Drozd, R Garg, et al. Resilient and modular subterranean exploration with a team of roving and flying robots. Field Robotics, 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "type": "text", + "content": "[39] David Schubert, Thore Goll, Nikolaus Demmel, Vladyslav Usenko, Jörg Stuckler, and Daniel Cremers. The tum vi benchmark for evaluating visual-inertial odometry. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1680-1687. IEEE, 2018. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[40] Aleksandr Segal, Dirk Haehnel, and Sebastian Thrun. Generalized-icp. In Robotics: science and systems, page 435. Seattle, WA, 2009. 4, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[41] Tixiao Shan, Brendan Englot, Drew Meyers, Wei Wang, Carlo Ratti, and Rus Daniela. Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5135-5142. IEEE, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "type": "text", + "content": "[42] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A benchmark for the evaluation of rgb-d slam systems. In Proc. of the International Conference on Intelligent Robot Systems (IROS), 2012. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 173, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 215 + ], + "type": "text", + "content": "[43] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 217, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 270 + ], + "type": "text", + "content": "[44] Yulun Tian, Yun Chang, Long Quang, Arthur Schang, Carlos Nieto-Granda, Jonathan P How, and Luca Carlone. Resilient and distributed multi-robot visual slam: Datasets, experiments, and lessons learned. arXiv preprint arXiv:2304.04362, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 272, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 326 + ], + "type": "text", + "content": "[45] Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Louis Wiesmann, Jens Behley, and Cyril Stachniss. Kiss-icp: In defense of point-to-point icp-simple, accurate, and robust registration if done the right way. IEEE Robotics and Automation Letters, 8(2):1029–1036, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 327, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 447 + ], + "type": "text", + "content": "[46] Chen Wang, Dasong Gao, Kuan Xu, Junyi Geng, Yaoyu Hu, Yuheng Qiu, Bowen Li, Fan Yang, Brady Moon, Abhinav Pandey, Aryan, Jiahe Xu, Tianhao Wu, Haonan He, Daning Huang, Zhongqiang Ren, Shibo Zhao, Taimeng Fu, Pranay Reddy, Xiao Lin, Wenshan Wang, Jingnan Shi, Rajat Talak, Kun Cao, Yi Du, Han Wang, Huai Yu, Shanzhao Wang, Siyu Chen, Ananth Kashyap, Rohan Bandaru, Karthik Dantu, Jiajun Wu, Lihua Xie, Luca Carlone, Marco Hutter, and Sebastian Scherer. PyPose: A library for robot learning with physics-based optimization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "type": "text", + "content": "[47] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual SLAM. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "type": "text", + "content": "[48] Wenshan Wang, Yaoyu Hu, and Sebastian Scherer. Tartanvo: A generalizable learning-based vo. In Conference on Robot Learning, pages 1761-1772. PMLR, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 537, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 590 + ], + "type": "text", + "content": "[49] P. Wenzel, R. Wang, N. Yang, Q. Cheng, Q. Khan, L. von Stumberg, N. Zeller, and D. Cremers. 4Seasons: A cross-season dataset for multi-weather SLAM in autonomous driving. In Proceedings of the German Conference on Pattern Recognition (GCPR), 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 592, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 545, + 623 + ], + "type": "text", + "content": "[50] Wei Xu, Yixi Cai, Dongjiao He, Jiarong Lin, and Fu Zhang. FAST-LIO2: fast direct lidar-inertial odometry. IEEE Transactions on Robotics, abs/2107.06829, 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 625, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 657 + ], + "type": "text", + "content": "[51] Wei Xu, Yixi Cai, Dongjiao He, Jiarong Lin, and Fu Zhang. Fast-lio2: Fast direct lidar-inertial odometry. IEEE Transactions on Robotics, 38(4):2053–2073, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "type": "text", + "content": "[52] Lintong Zhang, Marco Camurri, and Maurice Fallon. Multicamera lidar inertial extension to the newer college dataset. arXiv preprint arXiv:2112.08854, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 692, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 712 + ], + "type": "text", + "content": "[53] Shibo Zhao, Hengrui Zhang, Peng Wang, Lucas Nogueira, and Sebastian Scherer. Super odometry: Imu-centric" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22656" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 173 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 115 + ], + "type": "text", + "content": "lidar-visual-inertial estimator for challenging environments. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8729-8736. IEEE, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "type": "text", + "content": "[54] David Zuniga-Noël, Alberto Jaenal, Ruben Gomez-Ojeda, and Javier Gonzalez-Jimenez. The uma-vi dataset: Visual-inertial odometry in low-textured and dynamic illumination environments. The International Journal of Robotics Research, 39(9):1052-1060, 2020. 1, 3" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22657" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/5f051557-0067-4cd2-89f8-ff27dddafa6f_content_list.json b/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/5f051557-0067-4cd2-89f8-ff27dddafa6f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bd02f6a75db38caf23f01d7d0a9e9f7107885c4e --- /dev/null +++ b/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/5f051557-0067-4cd2-89f8-ff27dddafa6f_content_list.json @@ -0,0 +1,1430 @@ +[ + { + "type": "text", + "text": "Summarize the Past to Predict the Future: Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation", + "text_level": 1, + "bbox": [ + 96, + 130, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Razvan-George Pasca $^{1*}$ Alexey Gavyushin $^{1*}$ Muhammad Hamza $^{2}$ Yen-Ling Kuo $^{3}$ Kaichun Mo $^{4}$ Luc Van Gool $^{1,5,6}$ Otmar Hilliges $^{1}$ Xi Wang $^{1}$ ETH Zurich Univ. of Zurich Univ. of Virginia NVIDIA KU Leuven INSAIT, Sofia", + "bbox": [ + 145, + 202, + 831, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f23b4421a71e33743c550be111aa1ad4300ae983910743e70882576c09612d4.jpg", + "image_caption": [ + "Figure 1. TransFusion: Multimodal fusion transformer for short-term object interaction anticipation. Given a video sequence of past observations, the task aims to predict a set of objects visible in the current frame that will be interacted with in the future, i.e. in the activation frame that is $\\Delta$ frames away from the current prediction frame. Additionally, the task requires estimating the bounding box, the associated action described by a verb-noun pair, and the time to contact for each predicted object. We propose TransFusion, a multimodal fusion architecture that uses language summaries of past actions and salient objects to effectively predict future object interactions." + ], + "image_footnote": [], + "bbox": [ + 147, + 267, + 831, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 604, + 313, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We study object interaction anticipation in egocentric videos. This task requires an understanding of the spatiotemporal context formed by past actions on objects, coined action context. We propose TransFusion, a multimodal transformer-based architecture for short-term object interaction anticipation. Our method exploits the representational power of language by summarizing the action context textually, after leveraging pre-trained vision-language foundation models to extract the action context from past video frames. The summarized action context and the last observed video frame are processed by the multimodal fusion module to forecast the next object interaction. Experiments on the Ego4D next active object interaction dataset show the effectiveness of our multimodal fusion model and highlight the benefits of using the power of foundation models and language-based context summaries in a task where", + "bbox": [ + 75, + 635, + 472, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vision may appear to suffice. Our novel approach outperforms all state-of-the-art methods on both versions of the Ego4D dataset. A project video and code are available at https://eth-ait.github.io/transfusion-proj/.", + "bbox": [ + 496, + 604, + 893, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 691, + 633, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The ability to predict plausible future human-object interactions is important for many assistive technologies. The task of short-term object interaction anticipation [18, 22] is defined as predicting which object a user will interact with next, what action will be performed and when in the future, given an egocentric video input. Providing an effective solution would help artificial agents, such as care robots, assist humans in their daily activities, as well as enable safety applications such as child and elderly monitoring.", + "bbox": [ + 496, + 718, + 893, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This task is challenging because it requires reasoning about user intent [6, 35, 43], which is not directly observable. For example, if the video shows the user holding", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Authors contributed equally.", + "bbox": [ + 94, + 886, + 254, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18286", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a vegetable, the most likely next action is to slice it, but only with a knife and cutting board present. Without those, there is significantly more uncertainty. Therefore, understanding what has happened in the past and what relevant objects are present is critical for predicting the next object interaction. We collectively refer to these concepts as action context. Existing interaction forecasting methods do not explicitly model action context and rely on neural networks to extract that information from fixed-sized video chunks [5, 18, 22, 32].", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose to use language summarization of the past as an explicit representation of action context. This filters out ambiguities caused by visual clutter. It also puts focus on objects and actions directly relevant to the ongoing and likely future activities. For the object interaction anticipation, we next introduce a multimodal fusion approach. Specifically, we propose a transformer-based multimodal fusion model, TransFusion. As shown in Figure 1, our model takes language summaries of the action context and the last observed single frame of a video as input and predicts future object interactions. We employ an existing image captioning system [53] to generate a concise description of past actions given by pairs of verbs and nouns. We use CLIP [38], a vision-language model trained on pairs of images and captions, to detect salient objects in the prediction frame of interest, and deem these objects task-relevant. The verb-noun pairs together with the list of salient objects form the summary of the past action context. Leveraging the vast knowledge contained in pre-trained vision-language foundation models yields more generic representations of the past action context and enables us to better generalize towards various scenarios, in particular rare ones not well-represented in the dataset.", + "bbox": [ + 75, + 243, + 472, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments show that our language-based summaries can effectively represent the past action context. We evaluate on both versions of the challenging egocentric dataset Ego4D [22]. Our approach improves interaction prediction accuracy and outperforms state-of-the-art models pre-trained on large video datasets [7, 40]. Such improvements show that large vision-language models can be useful even for a task that seems purely visual.", + "bbox": [ + 75, + 592, + 470, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are:", + "bbox": [ + 76, + 714, + 307, + 729 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A strategy for generating concise language-based summaries of action context over longer timespans from frame-wise language-based summaries, for instance obtained using vision-language foundation models.", + "- TransFusion, a multimodal fusion model for short-term object interaction anticipation combining action context descriptions and visual features.", + "- Experiments on both versions of the Ego4D dataset showing improvements over state-of-the-art methods on the short-term object interaction anticipation task, in particular for long-tailed classes." + ], + "bbox": [ + 94, + 732, + 467, + 896 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 89, + 637, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object interaction anticipation. Several works have been proposed for human-object interaction prediction. Furnari et al. [18] first introduced next-active-object (NAO) prediction, which predicts the active/not active label for each object using its motion trajectory features, but doesn't consider action anticipation. Closely related to NAO, Bertasius et al. [5] introduced action-objects, which considers objects that capture the human actor's visual attention or tactile interaction. Bertasius et al. use the data from a stereo camera system in a two-stream network, integrating RGB and depth information. They show that certain aspects of human actions can be predicted by exploiting the spatial configuration of the objects and the actor's head. Their approach is not easily scalable due to the need for stereo data.", + "bbox": [ + 496, + 118, + 890, + 329 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to different tasks, Liu et al. [32] propose a model that improves NAO prediction by incorporating future hand trajectory modeling. They obtain impressive results using 3D CNNs, confirming that hand motion cues and scene dynamics have important explanatory power in predicting the short-term future. Nonetheless, their approach also encounters scaling limitations because additional ground-truth labels for hand trajectories are needed.", + "bbox": [ + 496, + 333, + 890, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Ego4D dataset [22] proposed object interaction anticipation. Apart from locating the NAO, the task in Ego4D also requires the prediction of associated nouns and verbs, as well as the time to contact (TTC) for the future interacted objects. The official Ego4D baseline uses the current frame and past video frames as input and employs a two-stage approach to first learn noun and box prediction, and afterwards learn verb and TTC prediction. InternVideo [7] similarly leverages a DINO [62] model to detect objects in the first stage, and a VideoMAE [49]-pretrained ViT [14] spacetime attention model to predict verbs and TTC conditioned on the object predictions in the second stage. StillFast [40] presents an end-to-end method for NAO prediction, fusing multi-scale high-resolution image and low-resolution video features and combining local RoI features with global scene features to boost the interaction anticipation performance. GANOv2 [46, 47] and its predecessor [48] extract features from objects detected in past observed frames and fuse them with global frame features, forwarding them to a StillFast predictor and a motion-object-separating encoder-decoder architecture, respectively.", + "bbox": [ + 496, + 457, + 890, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Action anticipation. Besides localizing future object interactions, predicting the next action steps is an equally important task that intelligent systems need to perform to interact with humans. A longer prediction timespan requires a better grasp of the structure behind the succession of actions. A pioneering advancement in this field is the introduction of the EPIC-Kitchens 100 (EK100) [9] dataset, which facilitates research in action anticipation within egocentric", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18287", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "videos. Significant research efforts have been dedicated to exploring the action anticipation task utilizing EK100. Anticipative Video Transformer (AVT) [20] is one of the best-performing architectures on the EK100 [9] action anticipation benchmark. AVT uses a pre-trained Vision Transformer [14] as the image feature backbone and attends to the previously observed video frames to anticipate future actions. MeMViT [54] improves AVT by exploiting a longer temporal context, highlighting the importance of modeling a long sequence of previous actions. Other model architectures such as recurrent neural networks [17, 55], multimodal temporal CNNs [25] and transformers [19, 51] are also used to model the temporal contexts. Note that these methods are designed for action classification using video data and adapting them for object detection is not trivial. It is worth noting that the action anticipation task from the EK100 benchmark is a subtask of our short-term object interaction anticipation task. Our task not only predicts the action label of verb-noun pairs, but we also predict the location of the object and the time to contact, which is more challenging than the action anticipation task itself. TransFusion focuses on detecting the next active objects, as all evaluation metrics are conditioned on the box IoU. Therefore, comparing our work with the state-of-the-art methods for the action anticipation task on EK100 is out of the scope of this work. The Ego4D dataset itself further introduces a long-term action anticipation benchmark, which [61] tackles using pre-trained visual grounding models.", + "bbox": [ + 76, + 90, + 472, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-language models. Our approach combines the visual and language modalities to improve object interaction anticipation. Most recent multimodal vision-language architectures are based on Transformer [50] fusion schemes applied on features extracted by various encoders [4, 12, 26]. We follow their spirit by employing self-attention-based modality fusion. Our work differs in one significant aspect: the language features represent the past action context. This serves as an additional signal to disambiguate the actor's intent given similar visual scenes. The model must learn to predict from the multimodal fusion of language context summaries and visual cues in a considerably smaller data regime.", + "bbox": [ + 76, + 521, + 468, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More broadly, there are numerous transformer-based works aimed at learning cross-modal representations between vision and language [24, 28, 36, 38, 44, 45, 58, 59, 64] using terabyte-scale datasets crawled from multiple internet sources. When combined with large language models such as GPT-2 [37], BERT [13], Sentence-BERT [41], and the T5 variants [8, 39], these vision-language models can be applied to several downstream tasks such as VQA [3] and image captioning [56, 57]. Similar to other vision-language models that are trained with language input, our approach leverages state-of-the-art image captioning models [53] to provide a consistent description of past actions.", + "bbox": [ + 76, + 719, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Language-guided prediction", + "text_level": 1, + "bbox": [ + 498, + 89, + 759, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use language-based context summaries to represent the spatial-temporal context formed by past actions and object relationships. The context summary offers an explicit yet compact representation, which allows us to focus on task-relevant content and filter out unimportant information from the likely cluttered environment. To this end, we propose TransFusion, a transformer-based model that leverages a joint-attention fusion mechanism to effectively learn to predict object interactions in the future from the past action context. In the following, we describe our framework, which takes video clips as input and anticipates the next object interaction. We begin by defining the object interaction anticipation task (Sec. 3.1). We then show how to generate a summary for a prediction frame from an egocentric video to represent the action context provided by the past frames (Sec. 3.2). Finally, we present our multimodal fusion model's architecture (Sec. 3.3).", + "bbox": [ + 496, + 114, + 893, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Task definition", + "text_level": 1, + "bbox": [ + 498, + 385, + 650, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a video $\\mathcal{V} = \\{f_1,\\dots ,f_T\\}$ of length $T$ as input, our goal is to predict the next object interaction $\\mathcal{O}$ in the last observed video frame $f_{T}$ , possibly using older frames to extract context information. We call $f_{T}$ the prediction frame.", + "bbox": [ + 496, + 409, + 890, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Figure 1, the task is to anticipate the object interaction in $f_{T}$ . $\\mathcal{O}_T = \\{o_1^T,\\dots ,o_N^T\\}$ denotes the set of future object interactions, with $N > 1$ whenever multiple objects are interacted with simultaneously. In other words, multiple $o_i^T$ can be detected in the prediction frame $f_{T}$ . For example, $N = 2$ when one hand is holding two mugs in a kitchen scene, or the left hand is touching a plant while the right hand is moving a pot in a garden scene. Each object interaction $o_i^T = (b_i^T,n_i^T,v_i^T,t_i^T)$ consists of the object bounding box $b_{i}^{T}$ , the semantic object label (noun) $n_i^T$ , the action label (verb) $v_{i}^{T}$ , and the time to contact (TTC) $t_i^T$ . The task is to predict the next active object location $b_{i}^{T}$ in the prediction frame $f_{T}$ , the associated action described by the verb-noun pair $(n_i^T,v_i^T)$ , and the time $t_i^T$ to the point of contact when the interaction is going to start. The task is set up such that the actual object interaction takes place after a buffer time $\\Delta$ on frame $f_{T + \\Delta}$ and the predicted time to contact $t_i^T = \\Delta$ . Ego4D uses $\\Delta \\geq 0.033s$ . This requires anticipating future actions and increases the task difficulty.", + "bbox": [ + 496, + 472, + 893, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Summarizing the past", + "text_level": 1, + "bbox": [ + 500, + 770, + 705, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Anticipating object interactions in the future requires an understanding of what happened in the past. We first infer the per-frame action context from each egocentric video. This process consists of extracting action descriptions and held objects. These information sources are generated and aggregated independently across frames and finally forwarded as the action context to our anticipation model.", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18288", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dd1f4c7a5e8616dfd8819f244e47559a7fb781f4fd0cbae7c8aa31790cffbb6f.jpg", + "image_caption": [ + "Figure 2. Overview of the TransFusion model. TransFusion takes the prediction frame and the action context summary as input and predicts the bounding box of the next-active object, the noun-verb pair describing the associated action, and the time to contact (TTC). Feature maps of different scales are extracted from the visual encoder and then fused via a multimodal fusion module with the encoded language features. Their output is then processed by a regular feature pyramid network (FPN), denoted as multi-scale region proposal networks, before being fed into the Faster R-CNN detector." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 893, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Extracting frame-wise action descriptions. Inspired by the label format used in egocentric video datasets [10, 22], we aim to describe the past actions by sequences of verb-noun pairs (e.g. \"wash tomato\", \"cut tomato\"), so-called action descriptions. For each video frame $f$ , we extract the action currently performed by the agent, represented by a single pair $(v, n)$ of a verb $v$ and a noun $n$ .", + "bbox": [ + 75, + 343, + 472, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use an off-the-shelf image captioning model [53] to generate multiple full-sentence captions for each frame of interest, so as to find repeatedly occurring outputs and reduce noise: we perform part-of-speech tagging [2] on these captions and collect candidate pairs consisting of verbs followed by nouns within some cutoff distance $d \\in \\mathbb{N}_0^+$ . Setting $d > 0$ allows us to account for additional words appearing between a verb-noun pair (e.g. \"eat apple\" from \"A person eating a red apple\") and thereby obtain more valid candidate pairs from the captions. However, larger cutoff distances may introduce more spurious detections (e.g. \"eat gathering\", from \"A person eating while at a gathering\"). We use $d = 4$ in our experiments. The most frequent candidate pair across the frame's captions is selected as the verb-noun action description of this frame. If no valid verb-noun pair is found, the frame will have no action description.", + "bbox": [ + 75, + 459, + 472, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Extracting frame-wise salient objects. Another characteristic shared by many daily action sequences is that the next active object is likely to have already appeared in the actor's view before the upcoming interaction. The current environment in the recording may also provide model conditioning. Therefore, we select a set of salient objects in each frame as (1) the potential candidates for future active objects and (2) a proxy of the surroundings.", + "bbox": [ + 75, + 703, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Encouraged by its previous successful use in action-related tasks [30, 52], we use CLIP [38], a pre-trained open-vocabulary zero-shot classifier, to compute the cosine similarities between the CLIP image embedding of a frame and the embeddings of the object categories from the Ego4D do", + "bbox": [ + 75, + 825, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "main. We choose to keep the highest-scoring $k$ objects as the set of salient objects for a given frame. A small $k$ may fail to provide sufficient information, while a large $k$ makes capturing the true salient objects more likely, yet introduces more noise due to false detections. We set $k = 3$ .", + "bbox": [ + 496, + 343, + 893, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Aggregating summaries over multiple frames. After the frame-wise summarization step, we have, for each frame $t$ , at most one verb-noun pair describing the frame-wise action $a_{t} = (n,v) \\in \\mathcal{A}_{t}^{F}$ . We then apply a cross-frame aggregation scheme to all $\\mathcal{A}_{t}^{F}$ so as to identify contiguous segments of frames exhibiting the same activity, while accounting for individual noisy frames possibly not matching the framework summaries of their neighbors. The aggregation processes the frames in temporal order. We consider frames belonging to the same atomic action as an action segment, which is initiated by observing a number of reoccurrences of some verb-noun pair and terminated by a lack of its occurrence within a number of contiguous frames. After the aggregation, we obtain $\\mathcal{A}$ , the sequence of all action segments in the video. For each prediction frame, we construct its action context from a number of action segments preceding the frame, and the objects from $\\mathcal{N}_s$ in that frame.", + "bbox": [ + 496, + 431, + 895, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By aggregating across frames, we thus capture dominant actions in the past, reduce noise in the generated action context and represent long timespans of similar activity by short textual descriptions. The complete action context is provided to our model by concatenating the textual representations of the aggregation results of any desired subset of the two context representations $\\{\\mathcal{A},\\mathcal{N}_s\\}$ , where we drop the index $t$ for convenience. For more details, see the Supp. Mat.", + "bbox": [ + 496, + 688, + 893, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. TransFusion: Multimodal fusion", + "text_level": 1, + "bbox": [ + 500, + 816, + 790, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 2 visualizes the architecture of the proposed multimodal TransFusion model. TransFusion employs two different base encoders, for language and image input respectively. The input prediction frame $f_{T}$ is processed by a regu", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18289", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ec1da6b14eb441a83d93e834c2f4a065392b12cc24c9c1f485faa788592b3b40.jpg", + "image_caption": [ + "Figure 3. Multimodal fusion module. We first project the CNN feature map and the language tokens to a common dimensionality before adding the specific embeddings. We concatenate the visual and language tokens and feed them to $M$ self-attention layers. At the output, the fused visual tokens are projected back into the initial feature map shape." + ], + "image_footnote": [], + "bbox": [ + 117, + 87, + 433, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "lar CNN-based visual encoder [23], producing a set of multiscale feature maps $\\mathcal{F}_{\\mathcal{V}}^{s}(f_{T})$ , with $s$ indicating the scale. The frame's action context is encoded by a language encoder [41] and yields $\\mathcal{F}_{\\mathcal{L}}(\\mathcal{C}_L)$ . The feature maps belonging to different scale levels from the visual encoder are fused with the encoded language features via a joint-attention scheme. The result is then processed by a regular feature pyramid network (FPN) [29] before being forwarded to the Faster R-CNN [42] detector.", + "bbox": [ + 75, + 455, + 468, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multimodal fusion. For simplicity, we describe the multimodal fusion module for a single input scale $s$ , which we hence omit from notation. See Figure 3 for a detailed view of a single Transformer encoder layer together with the input projection stages. For the visual modality, we first split the image into a sequence of tokens. Specifically, the visual features $\\mathcal{F}_{\\mathcal{V}}(f_T) \\in \\mathbb{R}^{C \\times H \\times W}$ are reshaped to $\\mathbb{R}^{N \\times (P^2 \\cdot C)}$ where $P$ is the token size and $N = \\frac{H \\times W}{P^2}$ the number of tokens. We add unique type embeddings as well as a positional encoding to each modality. Both visual and language tokens are first mapped to $\\mathbb{R}^D$ via a linear projection before being fed to the fusion module in a concatenated form. The fusion block consists of a regular Transformer encoder layer replicated $M$ times. Finally, the visual tokens are regrouped into their initial shape to compose the feature map for the feature pyramid network (FPN).", + "bbox": [ + 75, + 594, + 472, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multiscale fusion. The multimodal fusion block outputs multiple fused feature maps of varying scales: $\\mathcal{F}_{fi} \\in \\mathbb{R}^{C_i \\times H_i \\times W_i}$ , where $i$ is the scale index. They are fed into the FPN, which has different parameters for each feature", + "bbox": [ + 76, + 839, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "map, allowing the network to learn specialized features for each downsampling ratio. The output of the FPN is processed by a Faster R-CNN detector.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prediction head networks. The Faster R-CNN detector outputs the bounding box $b$ through a regression layer, an associated object $n$ and verb $v$ , and time to contact $t$ .", + "bbox": [ + 496, + 138, + 890, + 184 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training objective. We train TransFusion using the standard Faster R-CNN [42] objective $\\mathcal{L}_{box}$ for bounding box prediction, which consists of localization and objectness terms; further, cross-entropy losses $\\mathcal{L}_{noun}$ and $\\mathcal{L}_{verb}$ for noun and verb prediction respectively, and an L1 loss $\\mathcal{L}_{ttc}$ for time to contact prediction. The overall objective is", + "bbox": [ + 496, + 186, + 890, + 277 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {b o x}} + \\mathcal {L} _ {\\text {n o u n}} + \\mathcal {L} _ {\\text {v e r b}} + \\mathcal {L} _ {\\text {t t c}}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 289, + 890, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As defined in Faster R-CNN, $\\mathcal{L}_{box}$ is", + "bbox": [ + 519, + 316, + 764, + 330 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {b o x}} = \\frac {1}{N _ {\\mathrm {c l s}}} \\sum_ {i} \\mathcal {L} _ {\\mathrm {c l s}} \\left(p _ {i}, p _ {i} ^ {*}\\right) + \\frac {\\lambda}{N _ {\\mathrm {r e g}}} \\sum_ {i} p _ {i} ^ {*} \\cdot L _ {\\mathrm {r e g}} \\left(b _ {i} - b _ {i} ^ {*}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 340, + 890, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_i$ is the predicted probability of a box being classified as a foreground object (true positive), $p_i^*$ is the ground-truth label; $b_i$ and $b_i^*$ are the predicted and ground-truth box coordinates. $N_{\\mathrm{reg}}$ is the number of bounding boxes used in training, $N_{\\mathrm{cls}}$ the detection network's image batch size. We leave $\\lambda = 11$ as in the reference implementation.", + "bbox": [ + 496, + 390, + 890, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 494, + 632, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We use both versions (v1 and v2) of the dataset provided for the short-term object interaction anticipation task in Ego4D [22]. It contains various activities ranging from gardening work to cooking or car parts changing. There are in total 64,798 (v2: 165,451) annotated video clips of $\\sim 5$ minutes each with frame rate $fps = 30$ . The dataset has a highly imbalanced long tail distribution, including 87 nouns and 74 verbs (v2: 128 and 81 resp.). For instance, the \"take\" class encompasses almost $40\\%$ of all the verb labels in version 1, while \"mold\" appears only in one video. The train set consists of about 28k (v2: 98k) samples, while the rest are roughly equally split between validation and test.", + "bbox": [ + 496, + 518, + 890, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation. As in the official Ego4D benchmark [22], we use the Top-5 mean Average Precision (mAP) to evaluate the performance of individual predictions. It considers two constraints: The IoU constraint requires that the predicted boxes are counted as hits only if there is an intersection over union IoU $\\geq 0.5$ with the ground-truth box. The TTC constraint requires that the predicted time-to-contact $t_i$ is close enough to the ground-truth $\\hat{t}_i$ , i.e. $|t_i - \\hat{t}_i| < T_\\delta, T_\\delta = 0.25$ . The following metrics are considered:", + "bbox": [ + 496, + 718, + 890, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Noun refers to the Noun-Box Top-5 mAP. In addition to the IoU constraint, a correct match with the ground-truth noun is required.", + "bbox": [ + 517, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18290", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3d5187d91dda355a3a175239bb85fe7d7ec5cf23bc0fd5522819e758861d34ff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Ver.ModelN↑N-V ↑N-T ↑A↑
v1FRCNN+Rnd. [22]20.452.223.860.44
FRCNN+SF [22]20.456.786.172.45
InternVideo [7]24.609.187.643.40
StillFast [40]19.519.956.453.49
TF (ours)24.9010.067.573.54
v2FRCNN+SF [22]26.159.458.693.61
StillFast [40]25.0613.299.145.12
GANov2 [46]25.6713.609.025.16
TF (ours)30.4313.4510.385.18
", + "bbox": [ + 78, + 88, + 470, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison with the state-of-the-art. Our method TransFusion (TF) outperforms all state-of-the-art approaches on 3 out of 4 metrics, on both versions of the Ego4D test set. We report the top-5 mean average precision for the Noun (N), NounVerb (N-V), Noun-TTC (N-T), and Noun-Verb-TTC (A) metrics.", + "bbox": [ + 75, + 239, + 468, + 309 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Noun-Verb refers to the Noun-Verb-Box Top-5 mAP. Correct noun and verb match with the IoU constraint.", + "- Noun-TTC refers to the Noun-Box Top-5 mAP. Correct noun matches with the TTC constraint.", + "- Noun-Verb-TTC (\"Overall\") refers to the overall Top-5 mAP, intersecting correct noun-verb matches with IoU and TTC constraints." + ], + "bbox": [ + 94, + 327, + 467, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that since detecting the next-active objects is the main focus of the task, all evaluation metrics are conditioned on the IoU constraint.", + "bbox": [ + 75, + 439, + 467, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We make use of the pre-trained Torchvision [34] Faster R-CNN. ResNet-50 [23] is used as the visual encoder and Sentence-BERT (SBERT) [41] is used as the language encoder. We freeze ResNet-50 and apply the multimodal fusion scheme on top. We use the RAdam [31] optimizer with a weight decay of $2e^{-4}$ , and an effective batch size of 32. Besides the default Faster R-CNN background class in nouns, we also add it in the verb class, which improves the Noun-Verb mAP by almost 2 points. We reduce the region proposal network's sampling batch size and the detection network's image batch size and use data augmentation techniques to reduce overfitting. The number of verb-noun pairs used in the input is defined as the context length $L_{c}$ . On average, an action segment described by one verb-noun phrase corresponds to one second with 30 frames, as computed from the GT annotations. Unless specified otherwise, we use $L_{c} = 3$ . We perform postprocessing involving non-maximum suppression on the predicted bounding boxes, as elaborated in the Supp. Mat.", + "bbox": [ + 75, + 486, + 468, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison with state-of-the-art", + "text_level": 1, + "bbox": [ + 76, + 785, + 366, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first compare our method against the baselines proposed in the Ego4D work [22]. The FRCNN + SF method employs a two-stage approach: a ResNet-based Faster R-CNN detector first produces bounding box regressions and noun predictions from the prediction frame, without using any video features. In the second stage, a SlowFast [16] 3D", + "bbox": [ + 75, + 809, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/074104493356d0743a5070d5e10fd5fb5890bc65383919de0f728cf803c48a64.jpg", + "image_caption": [ + "Figure 4. Prediction of Noun-Verb in relation to time-to-contact. Histogram of the time-to-contact labels in Ego4Dv1 is shown in blue bars. Performance measured in Noun-Verb is plotted as a function of time-to-contact. We compare our method to Ego4D's official FRCNN + SF [22] baseline." + ], + "image_footnote": [], + "bbox": [ + 509, + 89, + 885, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CNN video processing module performs region-of-interest (RoI) pooling [21] on the corresponding SlowFast 3D CNN video features. The pooled video features are then used by RoI heads to predict the verb and TTC values. The FR-CNN + Rnd. baseline, with results published on v1 only, drops the second stage of FRCNN + SF and uses randomly drawn verb and TTC predictions, following their respective distributions in the training set.", + "bbox": [ + 496, + 332, + 890, + 453 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Another baseline on v1 is given by InternVideo [7], a method achieving strong results on the short-term object interaction anticipation task with a two-stage strategy similar to FRCNN+SF: bounding boxes and noun labels are first obtained using a DINO-based object detector [62], after which a VideoMAE-pretrained [49] ViT [14] model leverages space-time attention to predict the corresponding verb classes and TTC values.", + "bbox": [ + 496, + 455, + 890, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further compare TransFusion to the recent StillFast [40] approach on both dataset versions. StillFast uses a Faster-RCNN-based spatial stream operating on a high-resolution version of the input image, and an X3D-based [15] temporal stream operating on a low-resolution version of the input video. The two streams each produce a multiscale feature map. Both feature maps are fused into a combined feature pyramid, from which local RoI features and global scene features are extracted, merged by a fusion network, and finally used by linear layers to predict each boxes' noun, verb and TTC.", + "bbox": [ + 496, + 578, + 890, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GANOV2 [46] enhances StillFast by adding a multi-head guided attention mechanism to its temporal stream, achieving favorable performance on version 2 of the dataset. The mechanism attends to object detections produced from the network's temporal branch, together with patches from the 3-D temporal feature stack, in a cross-fusion manner, and uses the fused tokens as the input to StillFast's feature pyramid and subsequent architecture.", + "bbox": [ + 496, + 747, + 890, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 demonstrates that TransFusion outperforms all state-of-the-art methods in terms of the N, N-V and A met-", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18291", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/edbb1320c1404778505b6b5f81388d0d766cb2538b2dca253dc38b7479457ed9.jpg", + "image_caption": [ + "Figure 5. Classification performance on top/tail categories. We show the relative and absolute gains of TransFusion over Ego4D's FRCNN + SF baseline [22] on the validation set, for Noun-Only and Verb-Only mAP (without any box IoU constraint). Relative improvements are written on top of the red bars." + ], + "image_footnote": [], + "bbox": [ + 112, + 80, + 439, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rics on version 1 of the dataset, and approaches the state-of-the art performance on the N-T metric while requiring a fraction of the competitor's resources during training [7] (see the Supp. Mat. for details). On the dataset's second version, TransFusion performs best in terms of N, N-T and A, while approaching the runner-up method on N-V. These favorable results validate the advantages of modeling action context explicitly for the interaction anticipation task.", + "bbox": [ + 75, + 287, + 468, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We henceforth analyze our performance gains over the official Ego4D FRCNN + SF baseline on the first version of the Ego4D dataset. Figure 4 shows the long tail distribution of the time-to-contact labels in the validation set, split into buckets. We see that TransFusion consistently outperforms the baseline on the N-V metric for all TTC ranges. The results in Figure 5 indicate that our method not only outperforms the Ego4D method on top noun and verb classes, but also achieves bigger gains in the tail classes, demonstrating the generalization ability of using our language-based context summaries. Top and tail buckets represent roughly $50\\%$ of the samples in the validation set.", + "bbox": [ + 75, + 410, + 468, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 6 shows that the Ego4D approach manages to detect the most salient objects in the environment, but fails to reliably associate the bounding boxes to the correct action phrase. In contrast, TransFusion is able to correctly associate the drawer bounding box to \"open drawer\" and to predict the correct action associated with the phone. More examples are provided in the Supp. Mat.", + "bbox": [ + 75, + 592, + 468, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation studies", + "text_level": 1, + "bbox": [ + 76, + 709, + 236, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation of action context representation. In Table 2, we ablate different variants of context representation and compare against the representation that uses the groundtruth (GT) action labels provided by Ego4D in the action context, to assess the quality of our generated context summaries. The observed TTC Average Precision difference measures around $2\\%$ and was thus omitted from the comparison. We report on the Noun and Noun-Verb metrics as we expect major differences in semantic-related tasks.", + "bbox": [ + 75, + 733, + 468, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experiments show that we achieve results approaching GT, demonstrating the effectiveness of our context repre", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2760f6325fb033de5848b12ab2a0db695cc01daa0951789615e11bb639c27c08.jpg", + "image_caption": [ + "GT: open drawer 0.1s", + "(a) Ego4D: take container 0.78s, take container 0.46s, take pot 0.42s, take lid 0.46s" + ], + "image_footnote": [], + "bbox": [ + 506, + 102, + 687, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f347836900e00440eff16f08c3b1a20da37d3039bebd62ce7fbfd2bcac128d1c.jpg", + "image_caption": [ + "GT: take phone 1.0s", + "(b) Ego4D: take table 0.78s, take container 0.85s, take computer 0.86s, take phone 1.11s" + ], + "image_footnote": [], + "bbox": [ + 700, + 102, + 880, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f275f6477bf296d642ce09639b28fbbdb8f8c32b7a37ecd75be9cb706813e18f.jpg", + "image_caption": [ + "(c) TF (ours): open drawer 0.56s, open drawer 0.36s, open drawer 0.44s, take lid 0.6s" + ], + "image_footnote": [], + "bbox": [ + 506, + 250, + 684, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e1888e52b63b20e82773356a1746d7e8c773d7b1012884f11f87efe7f330a24.jpg", + "image_caption": [ + "(d) TF (ours): take phone 0.87s, open computer 0.63s, operate tablet 0.73s, take phone 0.98s", + "Figure 6. Qualitative results of the proposed approach. Results of Ego4D's FRCNN + SF baseline [22] are shown in the top row and TransFusion in the bottom row. Green denotes the ground truth bounding box, with the labels shown on top of each column." + ], + "image_footnote": [], + "bbox": [ + 699, + 250, + 879, + 353 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a61616aac5c4d615da69ed141d1e3f176671e6c2f7d79f6a060e06ca8f2cfc20.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Rep.∅ANhNsA+NhA+NsGT
N↑17.7120.0919.3319.7519.0320.1921.71
N-V↑6.147.166.867.057.127.557.86
", + "bbox": [ + 504, + 464, + 883, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison of different action context representations. We evaluate different variations of context representations on the predicted noun (N) and noun-verb (N-V), and compare them against ground-truth annotations (GT) provided by v1 of the Ego4D validation set. GT only contains action descriptions of verb-noun pairs. Action context represented by $\\mathcal{A} + \\mathcal{N}_s$ achieves the best N-V performance.", + "bbox": [ + 498, + 513, + 890, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sentations. Using a vision-only version of our model $(\\emptyset)$ yields markedly worse results than when adding any type of language conditioning. Action descriptions $\\mathcal{A}$ consistently boost the verb classification performance. Adding salient objects $\\mathcal{N}_s$ to $\\mathcal{A}$ further improves task performance. These together offer a reasonable picture of using language summaries to represent the action context. In addition, we notice a domain gap in the choice of words between the generated labels and the ground-truth ones. In some cases, even if the description is satisfactory for a human evaluator, the information may be too vague to be directly useful to the model. Phrasing differences between the generated captions and the training corpus of the language encoders may further help to improve the performance.", + "bbox": [ + 496, + 613, + 890, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As an additional type of action context, we hypothesize that the sequence of objects held by the actor may be indicative of the overall task and can help the model better infer the immediate next steps. Motivated by this, we use an existing hand-object interaction predictor [11] jointly with an", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18292", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/940a5f7209592e370d384e48631731a1fa40247041117243d0139f6e40d41cee.jpg", + "image_caption": [ + "Figure 7. Comparison to video features. TransFusion outperforms TransFusion-Video over different context lengths measured by Noun-Verb mAP. TransFusion using GT-based context summaries as language input represents an upper performance bound." + ], + "image_footnote": [], + "bbox": [ + 122, + 84, + 411, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "object detector [65] to look for held objects and their semantic labels in each frame. Usually, such objects occur only in a subset of frames (e.g. due to limited visibility of the actor's hands), which results in a lower performance of the resulting $\\mathcal{N}_h$ and $\\mathcal{A} + \\mathcal{N}_h$ representations. We hence use $\\mathcal{A} + \\mathcal{N}_s$ to represent the action context, considering the good performance achieved. Ablations on the hyperparameters $d$ , the cutoff distance, and $k$ , the number of salient objects per frame, show that the performance is less sensitive to their settings. More details can be found in the Supp. Mat.", + "bbox": [ + 75, + 287, + 468, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison to video features. We evaluate whether language summaries can better represent action context than video features for the object-anticipation task. To do so, we implement TransFusion-Video, which takes video features extracted from the Ego4D SlowFast [16] model as input and uses an identical fusion module as in TransFusion.", + "bbox": [ + 75, + 441, + 468, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 7 shows that TransFusion consistently outperforms TransFusion-Video. $L_{c} \\times 30$ frames are taken in TransFusion-Video. For video features, the domain gap between pre-training and target datasets seems to be too large to allow effective generalization to a diverse dataset such as Ego4D. We suspect that the feature representations are more entangled due to the need to represent temporal, visual, and semantic aspects simultaneously. With language, temporal information is directly encoded in the syntax, i.e. the succession of words. Given a video clip of one second, language can summarize it in two words whereas corresponding video features use more than 12 times the space (e.g. smaller SBERT [41] features of $2 \\times 384$ compared to SlowFast [16] features of $4 \\times 2304$ ). Therefore, our language-based context representation has an advantage when it comes to longer video sequences. Overall, TransFusion obtains a $21\\%$ boost in N-V performance. See the Supp. Mat. for details.", + "bbox": [ + 75, + 535, + 468, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Length of action context. A significant parameter in obtaining good performance is the length of action context, i.e., how many past actions to include in the input. An adequate context length is important to allow the model to pick up more specific activity patterns. Some activities that have a more structured nature or certain repeating patterns (e.g.", + "bbox": [ + 75, + 810, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a85093ff60befe94cb1db30be749e073ca104ce1fc6217a43d1fa2f1130037a9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Language EncoderToken sizeN↑N-V↑
SBERT/BERT [13, 41]38420.197.55
SBERT/RoBERTa [33, 41]76819.787.78
GPT-2 [37]76818.987.08
Flan-T5 [8]102418.266.61
", + "bbox": [ + 509, + 88, + 883, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Language encoder ablations. We compare SBERT using BERT and RoBERTa backbones, with GPT-2 and Flan-T5.", + "bbox": [ + 498, + 172, + 890, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cooking or repairing a bicycle) are easier to model with an increased context length compared to more random activities such as playing basketball. Figure 7 shows that the overall performance plateaus at $L_{c} = 2$ and higher.", + "bbox": [ + 496, + 210, + 890, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Language encoder. A language encoder is used to process the context summaries in TransFusion. To understand the impact of different language encoders, we experiment with multiple LLMs including SBERT with a RoBERTa backbone [33, 41], GPT-2 [37], and FLAN-T5 [8], using version 1 of the validation set. Table 3 summarizes the performance of different language encoders, showing that the SBERT models perform best. We consider TTC less relevant here. We hypothesize that the difference comes from the training objectives of language models. SBERT is trained to map semantically similar sentences closer to each other in the embedding space. This is particularly useful when the model needs to understand descriptions with semantically close words, e.g. \"cut carrots\" and \"slice carrots\". GPT-2 and FLAN-T5 are optimized using generative objectives, which makes them overly sensitive to the input word choice.", + "bbox": [ + 496, + 273, + 892, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fusion module. We provide ablations regarding the fusion module in the Supp. Mat.", + "bbox": [ + 498, + 517, + 890, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and discussion", + "text_level": 1, + "bbox": [ + 500, + 568, + 743, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose TransFusion, a multimodal fusion model that anticipates object interactions by considering the past action context represented by language summaries. Our successful experiments on a challenging egocentric video dataset demonstrate how language summaries and the vast knowledge contained in vision-language foundation models can improve object interaction anticipation, highlighting the representational power of language. It appears promising to investigate fusing even more modalities, such as motion cues from optical flow. We emphasize that the proposed summarization approach is not limited to this task. Language provides a universal interface to complement the visual input with information encoded in foundation models or task-specific pipelines. Future work can leverage language summarization for other video reasoning tasks [60, 63], or extend it temporally to describe possible future activities [1, 27] for long-term interaction anticipation.", + "bbox": [ + 496, + 595, + 890, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by an ETH Zurich Postdoctoral Fellowship. We thank Adrian Spurr for the helpful comments and Velko Vechev for the voiceover.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18293", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 90, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chaitanya Ahuja and Louis-Philippe Morency. Language2Pose: Natural Language Grounded Pose Forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719–728, 2019. 8", + "[2] Alan Akbik, Duncan Blythe, and Roland Vollgraf. Contextual String Embeddings for Sequence Labeling. In Proceedings of the 27th international conference on computational linguistics, pages 1638-1649, 2018. 4", + "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. VQA: Visual Question Answering. In Proceedings of the IEEE International Conference on Computer Vision, pages 2425-2433, 2015. 3", + "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. CoRR, abs/2104.00650, 2021. 3", + "[5] Gedas Bertasius, Hyun Soo Park, Stella X. Yu, and Jianbo Shi. First-Person Action-Object Detection with EgoNet. In Proceedings of Robotics: Science and Systems, 2017. 2", + "[6] Jerome S Bruner. Intention in the Structure of Action and Interaction. Advances in Infancy Research, 1981. 1", + "[7] Guo Chen et al. InternVideo-Ego4D: A Pack of Champion Solutions to Ego4D Challenges. arXiv preprint arXiv:2211.09529, 2022.2,6,7", + "[8] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling Instruction-Finetuned Language Models, 2022. 3, 8", + "[9] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling Egocentric Vision. CoRR, abs/2006.13256, 2020. 2, 3", + "[10] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling Egocentric Vision: Collection, Pipeline and Challenges for EPIC-Kitchens-100. International Journal of Computer Vision, 130(1):33-55, 2022. 4", + "[11] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. EPIC-KITCHENS VISOR Benchmark: VVideo Segmentations and Object Relations. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2022. 7", + "[12] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. TransVG: End-to-End Visual Grounding with Transformers. CoRR, abs/2104.08541, 2021.3", + "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Transformers for Language Understanding. CoRR, abs/1810.04805, 2018.3,8", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. CoRR, abs/2010.11929, 2020. 2, 3, 6", + "[15] Christoph Feichtenhofer. X3D: Expanding Architectures for Efficient Video Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 203-213, 2020. 6", + "[16] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast Networks for Video Recognition. CoRR, abs/1812.03982, 2018. 6, 8", + "[17] Antonino Furnari and Giovanni Maria Farinella. Rolling-Unrolling LSTMs for Action Anticipation from First-Person Video. CoRR, abs/2005.02190, 2020. 3", + "[18] Antonino Furnari, Sebastiano Battiato, Kristen Grauman, and Giovanni Maria Farinella. Next-Active-Object Prediction from Egocentric Videos. CoRR, abs/1904.05250, 2019. 1, 2", + "[19] H. Girase, N. Agarwal, C. Choi, and K. Mangalam. Latency Matters: Real-Time Action Forecasting Transformer. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18759-18769, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 3", + "[20] Rohit Girdhar and Kristen Grauman. Anticipative Video Transformer. In ICCV, 2021. 3", + "[21] Ross B. Girshick. Fast R-CNN. CoRR, abs/1504.08083, 2015.6", + "[22] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Wesley Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4D: Around the World in 3,000 Hours of Egocentric Video. CoRR, abs/2110.07058, 2021. 1, 2, 4, 5, 6, 7" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18294", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. CoRR, abs/1512.03385, 2015. 5, 6", + "[24] Zhicheng Huang, Zhaoyang Zeng, Bei Liu, Dongmei Fu, and Jianlong Fu. Pixel-BERT: Aligning Image Pixels with Text by Deep Multi-Modal Transformers. CoRR, abs/2004.00849, 2020.3", + "[25] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. EPIC-Fusion: Audio-Visual Temporal Binding for Egocentric Action Recognition. CoRR, abs/1908.08498, 2019. 3", + "[26] Weicheng Kuo, Fred Bertsch, Wei Li, AJ Piergiovanni, Mohammad Saffar, and Anelia Angelova. FindIt: Generalized Localization with Natural Language Queries. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI, pages 502-520. Springer, 2022. 3", + "[27] Yen-Ling Kuo, Xin Huang, Andrei Barbu, Stephen G McGill, Boris Katz, John J Leonard, and Guy Rosman. Trajectory Prediction with Linguistic Representations. In 2022 International Conference on Robotics and Automation (ICRA), pages 2868-2875. IEEE, 2022. 8", + "[28] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. Less is More: ClipBERT for Video-and-Language Learning via Sparse Sampling. CoRR, abs/2102.06183, 2021. 3", + "[29] Tsung-Yi Lin, Piotr Dólar, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature Pyramid Networks for Object Detection. CoRR, abs/1612.03144, 2016. 5", + "[30] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen Clip Models Are Efficient Video Learners. In European Conference on Computer Vision, pages 388-404. Springer, 2022. 4", + "[31] Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the Variance of the Adaptive Learning Rate and Beyond. CoRR, abs/1908.03265, 2019. 6", + "[32] Miao Liu, Siyu Tang, Yin Li, and James M. Rehg. Forecasting Human Object Interaction: Joint Prediction of Motor Attention and Egocentric Activity. CoRR, abs/1911.10967, 2019. 2", + "[33] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A Robustly Optimized BERT Pretraining Approach. CoRR, abs/1907.11692, 2019.8", + "[34] TorchVision maintainers and contributors. TorchVision: PyTorch's Computer Vision Library. https://github.com/pytorch/vision, 2016.6", + "[35] David W Orr. The Nature of Design: Ecology, Culture, and Human Intention. Oxford University Press, 2002. 1", + "[36] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou, Rama Chellappa, and Pengchuan Zhang. EGOVLPV2: Egocentric Video-Language Pre-training With Fusion in the Backbone, 2023. 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language Models are Unsupervised Multitask Learners. OpenAI Blog, 2019. 3, 8", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. CoRR, abs/2103.00020, 2021. 2, 3, 4", + "[39] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. CoRR, abs/1910.10683, 2019. 3", + "[40] Francesco Ragusa et al. StillFast: An End-to-End Approach for Short-Term Object Interaction Anticipation. arXiv preprint arXiv:2304.03959, 2023. 2, 6", + "[41] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. CoRR, abs/1908.10084, 2019. 3, 5, 6, 8", + "[42] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. CoRR, abs/1506.01497, 2015. 5", + "[43] Paschal Sheeran. Intention-Behavior Relations: A Conceptual and Empirical Review. European Review of Social Psychology, 12(1):1-36, 2002. 1", + "[44] Chen Sun, Fabien Baradel, Kevin Murphy, and Cordelia Schmid. Contrastive Bidirectional Transformer for Temporal Representation Learning. CoRR, abs/1906.05743, 2019. 3", + "[45] Hao Tan and Mohit Bansal. LXMERT: Learning Cross-Modality Encoder Representations from Transformers. CoRR, abs/1908.07490, 2019. 3", + "[46] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Enhancing Next Active Object-Based Egocentric Action Anticipation With Guided Attention. In 2023 IEEE International Conference on Image Processing (ICIP), pages 1450-1454. IEEE, 2023. 2, 6", + "[47] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Guided Attention for Next Active Object @ EGO4D STA Challenge. arXiv preprint arXiv:2305.16066, 2023. 2", + "[48] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Leveraging Next-Active Objects for Context-Aware Anticipation in Egocentric Videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 8657–8666, 2024. 2", + "[49] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked Autoencoders Are Data-Efficient Learners for Self-Supervised Video Pre-Training. Advances in Neural Information Processing Systems, 35:10078–10093, 2022. 2, 6", + "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. CoRR, abs/1706.03762, 2017. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18295", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-Anticipation Transformer for Online Action Understanding, 2023. 3", + "[52] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Action-CLIP: A New Paradigm for Video Action Recognition. arXiv preprint arXiv:2109.08472, 2021. 4", + "[53] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework. CoRR, abs/2202.03052, 2022. 2, 3, 4", + "[54] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. MeMViT: Memory-Augmented Multiscale Vision Transformer for Efficient Long-Term Video Recognition. CoRR, abs/2201.08383, 2022. 3", + "[55] Yu Wu, Linchao Zhu, Xiaohan Wang, Yi Yang, and Fei Wu. Learning to Anticipate Egocentric Actions by Imagination. CoRR, abs/2101.04924, 2021. 3", + "[56] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, Attend and Tell: Neural Image Caption Generation With Visual Attention. In International Conference on Machine Learning, pages 2048-2057. PMLR, 2015. 3", + "[57] Ting Yao, Yingwei Pan, Yehao Li, Zhaofan Qiu, and Tao Mei. Boosting Image Captioning With Attributes. In Proceedings of the IEEE International Conference on Computer Vision, pages 4894-4902, 2017. 3", + "[58] Moon Ye-Bin, Jisoo Kim, Hongyeob Kim, Kilho Son, and Tae-Hyun Oh. TextMania: Enriching Visual Feature by Text-driven Manifold Augmentation, 2023. 3", + "[59] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. MERLOT: Multimodal Neural Script Knowledge Models. CoRR, abs/2106.02636, 2021. 3", + "[60] Kuo-Hao Zeng, Tseng-Hung Chen, Ching-Yao Chuang, Yuan-Hong Liao, Juan Carlos Niebles, and Min Sun. Leveraging Video Descriptions to Learn Video Question Answering. In Thirty-First AAAI Conference on Artificial Intelligence, 2017. 8", + "[61] Ce Zhang, Changcheng Fu, Shijie Wang, Nakul Agarwal, Kwonjoon Lee, Chiho Choi, and Chen Sun. Object-Centric Video Representation for Long-term Action Anticipation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6751-6761, 2024. 3", + "[62] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection. In The Eleventh International Conference on Learning Representations, 2022. 2, 6", + "[63] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. Video Summarization With Long Short-Term Memory. In European Conference on Computer Vision, pages 766-782. Springer, 2016. 8", + "[64] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning Video Representations From Large Language Models, 2022. 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[65] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple Multi-Dataset Detection. CoRR, abs/2102.13086, 2021. 8", + "bbox": [ + 501, + 90, + 890, + 132 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18296", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/5f051557-0067-4cd2-89f8-ff27dddafa6f_model.json b/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/5f051557-0067-4cd2-89f8-ff27dddafa6f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c2877905d29b212f00e82f4d47564685a060f5d9 --- /dev/null +++ b/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/5f051557-0067-4cd2-89f8-ff27dddafa6f_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.877, + 0.178 + ], + "angle": 0, + "content": "Summarize the Past to Predict the Future: Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.203, + 0.832, + 0.257 + ], + "angle": 0, + "content": "Razvan-George Pasca\\(^{1*}\\) Alexey Gavyushin\\(^{1*}\\) Muhammad Hamza\\(^{2}\\) Yen-Ling Kuo\\(^{3}\\) Kaichun Mo\\(^{4}\\) Luc Van Gool\\(^{1,5,6}\\) Otmar Hilliges\\(^{1}\\) Xi Wang\\(^{1}\\) ETH Zurich Univ. of Zurich Univ. of Virginia NVIDIA KU Leuven INSAIT, Sofia" + }, + { + "type": "image", + "bbox": [ + 0.148, + 0.268, + 0.833, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.511, + 0.896, + 0.582 + ], + "angle": 0, + "content": "Figure 1. TransFusion: Multimodal fusion transformer for short-term object interaction anticipation. Given a video sequence of past observations, the task aims to predict a set of objects visible in the current frame that will be interacted with in the future, i.e. in the activation frame that is \\(\\Delta\\) frames away from the current prediction frame. Additionally, the task requires estimating the bounding box, the associated action described by a verb-noun pair, and the time to contact for each predicted object. We propose TransFusion, a multimodal fusion architecture that uses language summaries of past actions and salient objects to effectively predict future object interactions." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.605, + 0.314, + 0.621 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.473, + 0.879 + ], + "angle": 0, + "content": "We study object interaction anticipation in egocentric videos. This task requires an understanding of the spatiotemporal context formed by past actions on objects, coined action context. We propose TransFusion, a multimodal transformer-based architecture for short-term object interaction anticipation. Our method exploits the representational power of language by summarizing the action context textually, after leveraging pre-trained vision-language foundation models to extract the action context from past video frames. The summarized action context and the last observed video frame are processed by the multimodal fusion module to forecast the next object interaction. Experiments on the Ego4D next active object interaction dataset show the effectiveness of our multimodal fusion model and highlight the benefits of using the power of foundation models and language-based context summaries in a task where" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.895, + 0.668 + ], + "angle": 0, + "content": "vision may appear to suffice. Our novel approach outperforms all state-of-the-art methods on both versions of the Ego4D dataset. A project video and code are available at https://eth-ait.github.io/transfusion-proj/." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.693, + 0.634, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.895, + 0.856 + ], + "angle": 0, + "content": "The ability to predict plausible future human-object interactions is important for many assistive technologies. The task of short-term object interaction anticipation [18, 22] is defined as predicting which object a user will interact with next, what action will be performed and when in the future, given an egocentric video input. Providing an effective solution would help artificial agents, such as care robots, assist humans in their daily activities, as well as enable safety applications such as child and elderly monitoring." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "This task is challenging because it requires reasoning about user intent [6, 35, 43], which is not directly observable. For example, if the video shows the user holding" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.255, + 0.902 + ], + "angle": 0, + "content": "*Authors contributed equally." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18286" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "a vegetable, the most likely next action is to slice it, but only with a knife and cutting board present. Without those, there is significantly more uncertainty. Therefore, understanding what has happened in the past and what relevant objects are present is critical for predicting the next object interaction. We collectively refer to these concepts as action context. Existing interaction forecasting methods do not explicitly model action context and rely on neural networks to extract that information from fixed-sized video chunks [5, 18, 22, 32]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.473, + 0.59 + ], + "angle": 0, + "content": "In this paper, we propose to use language summarization of the past as an explicit representation of action context. This filters out ambiguities caused by visual clutter. It also puts focus on objects and actions directly relevant to the ongoing and likely future activities. For the object interaction anticipation, we next introduce a multimodal fusion approach. Specifically, we propose a transformer-based multimodal fusion model, TransFusion. As shown in Figure 1, our model takes language summaries of the action context and the last observed single frame of a video as input and predicts future object interactions. We employ an existing image captioning system [53] to generate a concise description of past actions given by pairs of verbs and nouns. We use CLIP [38], a vision-language model trained on pairs of images and captions, to detect salient objects in the prediction frame of interest, and deem these objects task-relevant. The verb-noun pairs together with the list of salient objects form the summary of the past action context. Leveraging the vast knowledge contained in pre-trained vision-language foundation models yields more generic representations of the past action context and enables us to better generalize towards various scenarios, in particular rare ones not well-represented in the dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.472, + 0.713 + ], + "angle": 0, + "content": "Our experiments show that our language-based summaries can effectively represent the past action context. We evaluate on both versions of the challenging egocentric dataset Ego4D [22]. Our approach improves interaction prediction accuracy and outperforms state-of-the-art models pre-trained on large video datasets [7, 40]. Such improvements show that large vision-language models can be useful even for a task that seems purely visual." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.715, + 0.308, + 0.73 + ], + "angle": 0, + "content": "In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.733, + 0.468, + 0.791 + ], + "angle": 0, + "content": "- A strategy for generating concise language-based summaries of action context over longer timespans from frame-wise language-based summaries, for instance obtained using vision-language foundation models." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.794, + 0.468, + 0.835 + ], + "angle": 0, + "content": "- TransFusion, a multimodal fusion model for short-term object interaction anticipation combining action context descriptions and visual features." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.838, + 0.468, + 0.897 + ], + "angle": 0, + "content": "- Experiments on both versions of the Ego4D dataset showing improvements over state-of-the-art methods on the short-term object interaction anticipation task, in particular for long-tailed classes." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.733, + 0.468, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.638, + 0.106 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.119, + 0.892, + 0.33 + ], + "angle": 0, + "content": "Object interaction anticipation. Several works have been proposed for human-object interaction prediction. Furnari et al. [18] first introduced next-active-object (NAO) prediction, which predicts the active/not active label for each object using its motion trajectory features, but doesn't consider action anticipation. Closely related to NAO, Bertasius et al. [5] introduced action-objects, which considers objects that capture the human actor's visual attention or tactile interaction. Bertasius et al. use the data from a stereo camera system in a two-stream network, integrating RGB and depth information. They show that certain aspects of human actions can be predicted by exploiting the spatial configuration of the objects and the actor's head. Their approach is not easily scalable due to the need for stereo data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.456 + ], + "angle": 0, + "content": "In addition to different tasks, Liu et al. [32] propose a model that improves NAO prediction by incorporating future hand trajectory modeling. They obtain impressive results using 3D CNNs, confirming that hand motion cues and scene dynamics have important explanatory power in predicting the short-term future. Nonetheless, their approach also encounters scaling limitations because additional ground-truth labels for hand trajectories are needed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.892, + 0.775 + ], + "angle": 0, + "content": "The Ego4D dataset [22] proposed object interaction anticipation. Apart from locating the NAO, the task in Ego4D also requires the prediction of associated nouns and verbs, as well as the time to contact (TTC) for the future interacted objects. The official Ego4D baseline uses the current frame and past video frames as input and employs a two-stage approach to first learn noun and box prediction, and afterwards learn verb and TTC prediction. InternVideo [7] similarly leverages a DINO [62] model to detect objects in the first stage, and a VideoMAE [49]-pretrained ViT [14] spacetime attention model to predict verbs and TTC conditioned on the object predictions in the second stage. StillFast [40] presents an end-to-end method for NAO prediction, fusing multi-scale high-resolution image and low-resolution video features and combining local RoI features with global scene features to boost the interaction anticipation performance. GANOv2 [46, 47] and its predecessor [48] extract features from objects detected in past observed frames and fuse them with global frame features, forwarding them to a StillFast predictor and a motion-object-separating encoder-decoder architecture, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Action anticipation. Besides localizing future object interactions, predicting the next action steps is an equally important task that intelligent systems need to perform to interact with humans. A longer prediction timespan requires a better grasp of the structure behind the succession of actions. A pioneering advancement in this field is the introduction of the EPIC-Kitchens 100 (EK100) [9] dataset, which facilitates research in action anticipation within egocentric" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18287" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.514 + ], + "angle": 0, + "content": "videos. Significant research efforts have been dedicated to exploring the action anticipation task utilizing EK100. Anticipative Video Transformer (AVT) [20] is one of the best-performing architectures on the EK100 [9] action anticipation benchmark. AVT uses a pre-trained Vision Transformer [14] as the image feature backbone and attends to the previously observed video frames to anticipate future actions. MeMViT [54] improves AVT by exploiting a longer temporal context, highlighting the importance of modeling a long sequence of previous actions. Other model architectures such as recurrent neural networks [17, 55], multimodal temporal CNNs [25] and transformers [19, 51] are also used to model the temporal contexts. Note that these methods are designed for action classification using video data and adapting them for object detection is not trivial. It is worth noting that the action anticipation task from the EK100 benchmark is a subtask of our short-term object interaction anticipation task. Our task not only predicts the action label of verb-noun pairs, but we also predict the location of the object and the time to contact, which is more challenging than the action anticipation task itself. TransFusion focuses on detecting the next active objects, as all evaluation metrics are conditioned on the box IoU. Therefore, comparing our work with the state-of-the-art methods for the action anticipation task on EK100 is out of the scope of this work. The Ego4D dataset itself further introduces a long-term action anticipation benchmark, which [61] tackles using pre-trained visual grounding models." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.522, + 0.47, + 0.719 + ], + "angle": 0, + "content": "Vision-language models. Our approach combines the visual and language modalities to improve object interaction anticipation. Most recent multimodal vision-language architectures are based on Transformer [50] fusion schemes applied on features extracted by various encoders [4, 12, 26]. We follow their spirit by employing self-attention-based modality fusion. Our work differs in one significant aspect: the language features represent the past action context. This serves as an additional signal to disambiguate the actor's intent given similar visual scenes. The model must learn to predict from the multimodal fusion of language context summaries and visual cues in a considerably smaller data regime." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "More broadly, there are numerous transformer-based works aimed at learning cross-modal representations between vision and language [24, 28, 36, 38, 44, 45, 58, 59, 64] using terabyte-scale datasets crawled from multiple internet sources. When combined with large language models such as GPT-2 [37], BERT [13], Sentence-BERT [41], and the T5 variants [8, 39], these vision-language models can be applied to several downstream tasks such as VQA [3] and image captioning [56, 57]. Similar to other vision-language models that are trained with language input, our approach leverages state-of-the-art image captioning models [53] to provide a consistent description of past actions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.09, + 0.761, + 0.108 + ], + "angle": 0, + "content": "3. Language-guided prediction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.895, + 0.375 + ], + "angle": 0, + "content": "We use language-based context summaries to represent the spatial-temporal context formed by past actions and object relationships. The context summary offers an explicit yet compact representation, which allows us to focus on task-relevant content and filter out unimportant information from the likely cluttered environment. To this end, we propose TransFusion, a transformer-based model that leverages a joint-attention fusion mechanism to effectively learn to predict object interactions in the future from the past action context. In the following, we describe our framework, which takes video clips as input and anticipates the next object interaction. We begin by defining the object interaction anticipation task (Sec. 3.1). We then show how to generate a summary for a prediction frame from an egocentric video to represent the action context provided by the past frames (Sec. 3.2). Finally, we present our multimodal fusion model's architecture (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.386, + 0.651, + 0.401 + ], + "angle": 0, + "content": "3.1. Task definition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Given a video \\(\\mathcal{V} = \\{f_1,\\dots ,f_T\\}\\) of length \\(T\\) as input, our goal is to predict the next object interaction \\(\\mathcal{O}\\) in the last observed video frame \\(f_{T}\\), possibly using older frames to extract context information. We call \\(f_{T}\\) the prediction frame." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.895, + 0.76 + ], + "angle": 0, + "content": "As illustrated in Figure 1, the task is to anticipate the object interaction in \\( f_{T} \\). \\( \\mathcal{O}_T = \\{o_1^T,\\dots ,o_N^T\\} \\) denotes the set of future object interactions, with \\( N > 1 \\) whenever multiple objects are interacted with simultaneously. In other words, multiple \\( o_i^T \\) can be detected in the prediction frame \\( f_{T} \\). For example, \\( N = 2 \\) when one hand is holding two mugs in a kitchen scene, or the left hand is touching a plant while the right hand is moving a pot in a garden scene. Each object interaction \\( o_i^T = (b_i^T,n_i^T,v_i^T,t_i^T) \\) consists of the object bounding box \\( b_{i}^{T} \\), the semantic object label (noun) \\( n_i^T \\), the action label (verb) \\( v_{i}^{T} \\), and the time to contact (TTC) \\( t_i^T \\). The task is to predict the next active object location \\( b_{i}^{T} \\) in the prediction frame \\( f_{T} \\), the associated action described by the verb-noun pair \\( (n_i^T,v_i^T) \\), and the time \\( t_i^T \\) to the point of contact when the interaction is going to start. The task is set up such that the actual object interaction takes place after a buffer time \\( \\Delta \\) on frame \\( f_{T + \\Delta} \\) and the predicted time to contact \\( t_i^T = \\Delta \\). Ego4D uses \\( \\Delta \\geq 0.033s \\). This requires anticipating future actions and increases the task difficulty." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.707, + 0.788 + ], + "angle": 0, + "content": "3.2. Summarizing the past" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Anticipating object interactions in the future requires an understanding of what happened in the past. We first infer the per-frame action context from each egocentric video. This process consists of extracting action descriptions and held objects. These information sources are generated and aggregated independently across frames and finally forwarded as the action context to our anticipation model." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18288" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.895, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.265, + 0.895, + 0.337 + ], + "angle": 0, + "content": "Figure 2. Overview of the TransFusion model. TransFusion takes the prediction frame and the action context summary as input and predicts the bounding box of the next-active object, the noun-verb pair describing the associated action, and the time to contact (TTC). Feature maps of different scales are extracted from the visual encoder and then fused via a multimodal fusion module with the encoded language features. Their output is then processed by a regular feature pyramid network (FPN), denoted as multi-scale region proposal networks, before being fed into the Faster R-CNN detector." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.344, + 0.473, + 0.451 + ], + "angle": 0, + "content": "Extracting frame-wise action descriptions. Inspired by the label format used in egocentric video datasets [10, 22], we aim to describe the past actions by sequences of verb-noun pairs (e.g. \"wash tomato\", \"cut tomato\"), so-called action descriptions. For each video frame \\( f \\), we extract the action currently performed by the agent, represented by a single pair \\( (v, n) \\) of a verb \\( v \\) and a noun \\( n \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.473, + 0.702 + ], + "angle": 0, + "content": "We use an off-the-shelf image captioning model [53] to generate multiple full-sentence captions for each frame of interest, so as to find repeatedly occurring outputs and reduce noise: we perform part-of-speech tagging [2] on these captions and collect candidate pairs consisting of verbs followed by nouns within some cutoff distance \\( d \\in \\mathbb{N}_0^+ \\). Setting \\( d > 0 \\) allows us to account for additional words appearing between a verb-noun pair (e.g. \"eat apple\" from \"A person eating a red apple\") and thereby obtain more valid candidate pairs from the captions. However, larger cutoff distances may introduce more spurious detections (e.g. \"eat gathering\", from \"A person eating while at a gathering\"). We use \\( d = 4 \\) in our experiments. The most frequent candidate pair across the frame's captions is selected as the verb-noun action description of this frame. If no valid verb-noun pair is found, the frame will have no action description." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Extracting frame-wise salient objects. Another characteristic shared by many daily action sequences is that the next active object is likely to have already appeared in the actor's view before the upcoming interaction. The current environment in the recording may also provide model conditioning. Therefore, we select a set of salient objects in each frame as (1) the potential candidates for future active objects and (2) a proxy of the surroundings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Encouraged by its previous successful use in action-related tasks [30, 52], we use CLIP [38], a pre-trained open-vocabulary zero-shot classifier, to compute the cosine similarities between the CLIP image embedding of a frame and the embeddings of the object categories from the Ego4D do" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.344, + 0.895, + 0.42 + ], + "angle": 0, + "content": "main. We choose to keep the highest-scoring \\( k \\) objects as the set of salient objects for a given frame. A small \\( k \\) may fail to provide sufficient information, while a large \\( k \\) makes capturing the true salient objects more likely, yet introduces more noise due to false detections. We set \\( k = 3 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.896, + 0.689 + ], + "angle": 0, + "content": "Aggregating summaries over multiple frames. After the frame-wise summarization step, we have, for each frame \\( t \\), at most one verb-noun pair describing the frame-wise action \\( a_{t} = (n,v) \\in \\mathcal{A}_{t}^{F} \\). We then apply a cross-frame aggregation scheme to all \\( \\mathcal{A}_{t}^{F} \\) so as to identify contiguous segments of frames exhibiting the same activity, while accounting for individual noisy frames possibly not matching the framework summaries of their neighbors. The aggregation processes the frames in temporal order. We consider frames belonging to the same atomic action as an action segment, which is initiated by observing a number of reoccurrences of some verb-noun pair and terminated by a lack of its occurrence within a number of contiguous frames. After the aggregation, we obtain \\( \\mathcal{A} \\), the sequence of all action segments in the video. For each prediction frame, we construct its action context from a number of action segments preceding the frame, and the objects from \\( \\mathcal{N}_s \\) in that frame." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.81 + ], + "angle": 0, + "content": "By aggregating across frames, we thus capture dominant actions in the past, reduce noise in the generated action context and represent long timespans of similar activity by short textual descriptions. The complete action context is provided to our model by concatenating the textual representations of the aggregation results of any desired subset of the two context representations \\(\\{\\mathcal{A},\\mathcal{N}_s\\}\\), where we drop the index \\(t\\) for convenience. For more details, see the Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.792, + 0.833 + ], + "angle": 0, + "content": "3.3. TransFusion: Multimodal fusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Figure 2 visualizes the architecture of the proposed multimodal TransFusion model. TransFusion employs two different base encoders, for language and image input respectively. The input prediction frame \\( f_{T} \\) is processed by a regu" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18289" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.088, + 0.434, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.359, + 0.473, + 0.444 + ], + "angle": 0, + "content": "Figure 3. Multimodal fusion module. We first project the CNN feature map and the language tokens to a common dimensionality before adding the specific embeddings. We concatenate the visual and language tokens and feed them to \\( M \\) self-attention layers. At the output, the fused visual tokens are projected back into the initial feature map shape." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.47, + 0.592 + ], + "angle": 0, + "content": "lar CNN-based visual encoder [23], producing a set of multiscale feature maps \\(\\mathcal{F}_{\\mathcal{V}}^{s}(f_{T})\\), with \\(s\\) indicating the scale. The frame's action context is encoded by a language encoder [41] and yields \\(\\mathcal{F}_{\\mathcal{L}}(\\mathcal{C}_L)\\). The feature maps belonging to different scale levels from the visual encoder are fused with the encoded language features via a joint-attention scheme. The result is then processed by a regular feature pyramid network (FPN) [29] before being forwarded to the Faster R-CNN [42] detector." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.473, + 0.839 + ], + "angle": 0, + "content": "Multimodal fusion. For simplicity, we describe the multimodal fusion module for a single input scale \\(s\\), which we hence omit from notation. See Figure 3 for a detailed view of a single Transformer encoder layer together with the input projection stages. For the visual modality, we first split the image into a sequence of tokens. Specifically, the visual features \\(\\mathcal{F}_{\\mathcal{V}}(f_T) \\in \\mathbb{R}^{C \\times H \\times W}\\) are reshaped to \\(\\mathbb{R}^{N \\times (P^2 \\cdot C)}\\) where \\(P\\) is the token size and \\(N = \\frac{H \\times W}{P^2}\\) the number of tokens. We add unique type embeddings as well as a positional encoding to each modality. Both visual and language tokens are first mapped to \\(\\mathbb{R}^D\\) via a linear projection before being fed to the fusion module in a concatenated form. The fusion block consists of a regular Transformer encoder layer replicated \\(M\\) times. Finally, the visual tokens are regrouped into their initial shape to compose the feature map for the feature pyramid network (FPN)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Multiscale fusion. The multimodal fusion block outputs multiple fused feature maps of varying scales: \\(\\mathcal{F}_{fi} \\in \\mathbb{R}^{C_i \\times H_i \\times W_i}\\), where \\(i\\) is the scale index. They are fed into the FPN, which has different parameters for each feature" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "map, allowing the network to learn specialized features for each downsampling ratio. The output of the FPN is processed by a Faster R-CNN detector." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.892, + 0.185 + ], + "angle": 0, + "content": "Prediction head networks. The Faster R-CNN detector outputs the bounding box \\( b \\) through a regression layer, an associated object \\( n \\) and verb \\( v \\), and time to contact \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.187, + 0.892, + 0.278 + ], + "angle": 0, + "content": "Training objective. We train TransFusion using the standard Faster R-CNN [42] objective \\(\\mathcal{L}_{box}\\) for bounding box prediction, which consists of localization and objectness terms; further, cross-entropy losses \\(\\mathcal{L}_{noun}\\) and \\(\\mathcal{L}_{verb}\\) for noun and verb prediction respectively, and an L1 loss \\(\\mathcal{L}_{ttc}\\) for time to contact prediction. The overall objective is" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.29, + 0.892, + 0.307 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {b o x}} + \\mathcal {L} _ {\\text {n o u n}} + \\mathcal {L} _ {\\text {v e r b}} + \\mathcal {L} _ {\\text {t t c}}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.317, + 0.766, + 0.332 + ], + "angle": 0, + "content": "As defined in Faster R-CNN, \\(\\mathcal{L}_{box}\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.342, + 0.892, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {b o x}} = \\frac {1}{N _ {\\mathrm {c l s}}} \\sum_ {i} \\mathcal {L} _ {\\mathrm {c l s}} \\left(p _ {i}, p _ {i} ^ {*}\\right) + \\frac {\\lambda}{N _ {\\mathrm {r e g}}} \\sum_ {i} p _ {i} ^ {*} \\cdot L _ {\\mathrm {r e g}} \\left(b _ {i} - b _ {i} ^ {*}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.391, + 0.892, + 0.482 + ], + "angle": 0, + "content": "where \\( p_i \\) is the predicted probability of a box being classified as a foreground object (true positive), \\( p_i^* \\) is the ground-truth label; \\( b_i \\) and \\( b_i^* \\) are the predicted and ground-truth box coordinates. \\( N_{\\mathrm{reg}} \\) is the number of bounding boxes used in training, \\( N_{\\mathrm{cls}} \\) the detection network's image batch size. We leave \\( \\lambda = 11 \\) as in the reference implementation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.495, + 0.633, + 0.512 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Dataset. We use both versions (v1 and v2) of the dataset provided for the short-term object interaction anticipation task in Ego4D [22]. It contains various activities ranging from gardening work to cooking or car parts changing. There are in total 64,798 (v2: 165,451) annotated video clips of \\(\\sim 5\\) minutes each with frame rate \\(fps = 30\\). The dataset has a highly imbalanced long tail distribution, including 87 nouns and 74 verbs (v2: 128 and 81 resp.). For instance, the \"take\" class encompasses almost \\(40\\%\\) of all the verb labels in version 1, while \"mold\" appears only in one video. The train set consists of about 28k (v2: 98k) samples, while the rest are roughly equally split between validation and test." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Evaluation. As in the official Ego4D benchmark [22], we use the Top-5 mean Average Precision (mAP) to evaluate the performance of individual predictions. It considers two constraints: The IoU constraint requires that the predicted boxes are counted as hits only if there is an intersection over union IoU \\(\\geq 0.5\\) with the ground-truth box. The TTC constraint requires that the predicted time-to-contact \\(t_i\\) is close enough to the ground-truth \\(\\hat{t}_i\\), i.e. \\(|t_i - \\hat{t}_i| < T_\\delta, T_\\delta = 0.25\\). The following metrics are considered:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- Noun refers to the Noun-Box Top-5 mAP. In addition to the IoU constraint, a correct match with the ground-truth noun is required." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18290" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.472, + 0.233 + ], + "angle": 0, + "content": "
Ver.ModelN↑N-V ↑N-T ↑A↑
v1FRCNN+Rnd. [22]20.452.223.860.44
FRCNN+SF [22]20.456.786.172.45
InternVideo [7]24.609.187.643.40
StillFast [40]19.519.956.453.49
TF (ours)24.9010.067.573.54
v2FRCNN+SF [22]26.159.458.693.61
StillFast [40]25.0613.299.145.12
GANov2 [46]25.6713.609.025.16
TF (ours)30.4313.4510.385.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.24, + 0.47, + 0.31 + ], + "angle": 0, + "content": "Table 1. Comparison with the state-of-the-art. Our method TransFusion (TF) outperforms all state-of-the-art approaches on 3 out of 4 metrics, on both versions of the Ego4D test set. We report the top-5 mean average precision for the Noun (N), NounVerb (N-V), Noun-TTC (N-T), and Noun-Verb-TTC (A) metrics." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.328, + 0.468, + 0.357 + ], + "angle": 0, + "content": "- Noun-Verb refers to the Noun-Verb-Box Top-5 mAP. Correct noun and verb match with the IoU constraint." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.36, + 0.468, + 0.389 + ], + "angle": 0, + "content": "- Noun-TTC refers to the Noun-Box Top-5 mAP. Correct noun matches with the TTC constraint." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.393, + 0.468, + 0.436 + ], + "angle": 0, + "content": "- Noun-Verb-TTC (\"Overall\") refers to the overall Top-5 mAP, intersecting correct noun-verb matches with IoU and TTC constraints." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.328, + 0.468, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.468, + 0.484 + ], + "angle": 0, + "content": "Note that since detecting the next-active objects is the main focus of the task, all evaluation metrics are conditioned on the IoU constraint." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.775 + ], + "angle": 0, + "content": "Implementation details. We make use of the pre-trained Torchvision [34] Faster R-CNN. ResNet-50 [23] is used as the visual encoder and Sentence-BERT (SBERT) [41] is used as the language encoder. We freeze ResNet-50 and apply the multimodal fusion scheme on top. We use the RAdam [31] optimizer with a weight decay of \\( 2e^{-4} \\), and an effective batch size of 32. Besides the default Faster R-CNN background class in nouns, we also add it in the verb class, which improves the Noun-Verb mAP by almost 2 points. We reduce the region proposal network's sampling batch size and the detection network's image batch size and use data augmentation techniques to reduce overfitting. The number of verb-noun pairs used in the input is defined as the context length \\( L_{c} \\). On average, an action segment described by one verb-noun phrase corresponds to one second with 30 frames, as computed from the GT annotations. Unless specified otherwise, we use \\( L_{c} = 3 \\). We perform postprocessing involving non-maximum suppression on the predicted bounding boxes, as elaborated in the Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.367, + 0.803 + ], + "angle": 0, + "content": "4.1. Comparison with state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We first compare our method against the baselines proposed in the Ego4D work [22]. The FRCNN + SF method employs a two-stage approach: a ResNet-based Faster R-CNN detector first produces bounding box regressions and noun predictions from the prediction frame, without using any video features. In the second stage, a SlowFast [16] 3D" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.09, + 0.887, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.257, + 0.892, + 0.327 + ], + "angle": 0, + "content": "Figure 4. Prediction of Noun-Verb in relation to time-to-contact. Histogram of the time-to-contact labels in Ego4Dv1 is shown in blue bars. Performance measured in Noun-Verb is plotted as a function of time-to-contact. We compare our method to Ego4D's official FRCNN + SF [22] baseline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.892, + 0.454 + ], + "angle": 0, + "content": "CNN video processing module performs region-of-interest (RoI) pooling [21] on the corresponding SlowFast 3D CNN video features. The pooled video features are then used by RoI heads to predict the verb and TTC values. The FR-CNN + Rnd. baseline, with results published on v1 only, drops the second stage of FRCNN + SF and uses randomly drawn verb and TTC predictions, following their respective distributions in the training set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.892, + 0.576 + ], + "angle": 0, + "content": "Another baseline on v1 is given by InternVideo [7], a method achieving strong results on the short-term object interaction anticipation task with a two-stage strategy similar to FRCNN+SF: bounding boxes and noun labels are first obtained using a DINO-based object detector [62], after which a VideoMAE-pretrained [49] ViT [14] model leverages space-time attention to predict the corresponding verb classes and TTC values." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.744 + ], + "angle": 0, + "content": "We further compare TransFusion to the recent StillFast [40] approach on both dataset versions. StillFast uses a Faster-RCNN-based spatial stream operating on a high-resolution version of the input image, and an X3D-based [15] temporal stream operating on a low-resolution version of the input video. The two streams each produce a multiscale feature map. Both feature maps are fused into a combined feature pyramid, from which local RoI features and global scene features are extracted, merged by a fusion network, and finally used by linear layers to predict each boxes' noun, verb and TTC." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.892, + 0.868 + ], + "angle": 0, + "content": "GANOV2 [46] enhances StillFast by adding a multi-head guided attention mechanism to its temporal stream, achieving favorable performance on version 2 of the dataset. The mechanism attends to object detections produced from the network's temporal branch, together with patches from the 3-D temporal feature stack, in a cross-fusion manner, and uses the fused tokens as the input to StillFast's feature pyramid and subsequent architecture." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Table 1 demonstrates that TransFusion outperforms all state-of-the-art methods in terms of the N, N-V and A met-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18291" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.082, + 0.441, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.216, + 0.47, + 0.286 + ], + "angle": 0, + "content": "Figure 5. Classification performance on top/tail categories. We show the relative and absolute gains of TransFusion over Ego4D's FRCNN + SF baseline [22] on the validation set, for Noun-Only and Verb-Only mAP (without any box IoU constraint). Relative improvements are written on top of the red bars." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.469, + 0.41 + ], + "angle": 0, + "content": "rics on version 1 of the dataset, and approaches the state-of-the art performance on the N-T metric while requiring a fraction of the competitor's resources during training [7] (see the Supp. Mat. for details). On the dataset's second version, TransFusion performs best in terms of N, N-T and A, while approaching the runner-up method on N-V. These favorable results validate the advantages of modeling action context explicitly for the interaction anticipation task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.469, + 0.59 + ], + "angle": 0, + "content": "We henceforth analyze our performance gains over the official Ego4D FRCNN + SF baseline on the first version of the Ego4D dataset. Figure 4 shows the long tail distribution of the time-to-contact labels in the validation set, split into buckets. We see that TransFusion consistently outperforms the baseline on the N-V metric for all TTC ranges. The results in Figure 5 indicate that our method not only outperforms the Ego4D method on top noun and verb classes, but also achieves bigger gains in the tail classes, demonstrating the generalization ability of using our language-based context summaries. Top and tail buckets represent roughly \\(50\\%\\) of the samples in the validation set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.469, + 0.699 + ], + "angle": 0, + "content": "Figure 6 shows that the Ego4D approach manages to detect the most salient objects in the environment, but fails to reliably associate the bounding boxes to the correct action phrase. In contrast, TransFusion is able to correctly associate the drawer bounding box to \"open drawer\" and to predict the correct action associated with the phone. More examples are provided in the Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.238, + 0.725 + ], + "angle": 0, + "content": "4.2. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.469, + 0.869 + ], + "angle": 0, + "content": "Evaluation of action context representation. In Table 2, we ablate different variants of context representation and compare against the representation that uses the groundtruth (GT) action labels provided by Ego4D in the action context, to assess the quality of our generated context summaries. The observed TTC Average Precision difference measures around \\(2\\%\\) and was thus omitted from the comparison. We report on the Noun and Noun-Verb metrics as we expect major differences in semantic-related tasks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Experiments show that we achieve results approaching GT, demonstrating the effectiveness of our context repre" + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.088, + 0.669, + 0.102 + ], + "angle": 0, + "content": "GT: open drawer 0.1s" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.103, + 0.689, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.208, + 0.693, + 0.247 + ], + "angle": 0, + "content": "(a) Ego4D: take container 0.78s, take container 0.46s, take pot 0.42s, take lid 0.46s" + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.089, + 0.854, + 0.102 + ], + "angle": 0, + "content": "GT: take phone 1.0s" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.103, + 0.882, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.208, + 0.889, + 0.247 + ], + "angle": 0, + "content": "(b) Ego4D: take table 0.78s, take container 0.85s, take computer 0.86s, take phone 1.11s" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.25, + 0.686, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.354, + 0.695, + 0.393 + ], + "angle": 0, + "content": "(c) TF (ours): open drawer 0.56s, open drawer 0.36s, open drawer 0.44s, take lid 0.6s" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.25, + 0.88, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.354, + 0.888, + 0.394 + ], + "angle": 0, + "content": "(d) TF (ours): take phone 0.87s, open computer 0.63s, operate tablet 0.73s, take phone 0.98s" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.399, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of the proposed approach. Results of Ego4D's FRCNN + SF baseline [22] are shown in the top row and TransFusion in the bottom row. Green denotes the ground truth bounding box, with the labels shown on top of each column." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.465, + 0.885, + 0.511 + ], + "angle": 0, + "content": "
Rep.∅ANhNsA+NhA+NsGT
N↑17.7120.0919.3319.7519.0320.1921.71
N-V↑6.147.166.867.057.127.557.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.515, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Table 2. Comparison of different action context representations. We evaluate different variations of context representations on the predicted noun (N) and noun-verb (N-V), and compare them against ground-truth annotations (GT) provided by v1 of the Ego4D validation set. GT only contains action descriptions of verb-noun pairs. Action context represented by \\(\\mathcal{A} + \\mathcal{N}_s\\) achieves the best N-V performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.825 + ], + "angle": 0, + "content": "sentations. Using a vision-only version of our model \\((\\emptyset)\\) yields markedly worse results than when adding any type of language conditioning. Action descriptions \\(\\mathcal{A}\\) consistently boost the verb classification performance. Adding salient objects \\(\\mathcal{N}_s\\) to \\(\\mathcal{A}\\) further improves task performance. These together offer a reasonable picture of using language summaries to represent the action context. In addition, we notice a domain gap in the choice of words between the generated labels and the ground-truth ones. In some cases, even if the description is satisfactory for a human evaluator, the information may be too vague to be directly useful to the model. Phrasing differences between the generated captions and the training corpus of the language encoders may further help to improve the performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As an additional type of action context, we hypothesize that the sequence of objects held by the actor may be indicative of the overall task and can help the model better infer the immediate next steps. Motivated by this, we use an existing hand-object interaction predictor [11] jointly with an" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18292" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.124, + 0.085, + 0.413, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.225, + 0.47, + 0.281 + ], + "angle": 0, + "content": "Figure 7. Comparison to video features. TransFusion outperforms TransFusion-Video over different context lengths measured by Noun-Verb mAP. TransFusion using GT-based context summaries as language input represents an upper performance bound." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.47, + 0.44 + ], + "angle": 0, + "content": "object detector [65] to look for held objects and their semantic labels in each frame. Usually, such objects occur only in a subset of frames (e.g. due to limited visibility of the actor's hands), which results in a lower performance of the resulting \\(\\mathcal{N}_h\\) and \\(\\mathcal{A} + \\mathcal{N}_h\\) representations. We hence use \\(\\mathcal{A} + \\mathcal{N}_s\\) to represent the action context, considering the good performance achieved. Ablations on the hyperparameters \\(d\\), the cutoff distance, and \\(k\\), the number of salient objects per frame, show that the performance is less sensitive to their settings. More details can be found in the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.469, + 0.532 + ], + "angle": 0, + "content": "Comparison to video features. We evaluate whether language summaries can better represent action context than video features for the object-anticipation task. To do so, we implement TransFusion-Video, which takes video features extracted from the Ego4D SlowFast [16] model as input and uses an identical fusion module as in TransFusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.469, + 0.807 + ], + "angle": 0, + "content": "Figure 7 shows that TransFusion consistently outperforms TransFusion-Video. \\( L_{c} \\times 30 \\) frames are taken in TransFusion-Video. For video features, the domain gap between pre-training and target datasets seems to be too large to allow effective generalization to a diverse dataset such as Ego4D. We suspect that the feature representations are more entangled due to the need to represent temporal, visual, and semantic aspects simultaneously. With language, temporal information is directly encoded in the syntax, i.e. the succession of words. Given a video clip of one second, language can summarize it in two words whereas corresponding video features use more than 12 times the space (e.g. smaller SBERT [41] features of \\( 2 \\times 384 \\) compared to SlowFast [16] features of \\( 4 \\times 2304 \\)). Therefore, our language-based context representation has an advantage when it comes to longer video sequences. Overall, TransFusion obtains a \\( 21\\% \\) boost in N-V performance. See the Supp. Mat. for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Length of action context. A significant parameter in obtaining good performance is the length of action context, i.e., how many past actions to include in the input. An adequate context length is important to allow the model to pick up more specific activity patterns. Some activities that have a more structured nature or certain repeating patterns (e.g." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.089, + 0.885, + 0.162 + ], + "angle": 0, + "content": "
Language EncoderToken sizeN↑N-V↑
SBERT/BERT [13, 41]38420.197.55
SBERT/RoBERTa [33, 41]76819.787.78
GPT-2 [37]76818.987.08
Flan-T5 [8]102418.266.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.173, + 0.892, + 0.201 + ], + "angle": 0, + "content": "Table 3. Language encoder ablations. We compare SBERT using BERT and RoBERTa backbones, with GPT-2 and Flan-T5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.211, + 0.892, + 0.271 + ], + "angle": 0, + "content": "cooking or repairing a bicycle) are easier to model with an increased context length compared to more random activities such as playing basketball. Figure 7 shows that the overall performance plateaus at \\( L_{c} = 2 \\) and higher." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.893, + 0.516 + ], + "angle": 0, + "content": "Language encoder. A language encoder is used to process the context summaries in TransFusion. To understand the impact of different language encoders, we experiment with multiple LLMs including SBERT with a RoBERTa backbone [33, 41], GPT-2 [37], and FLAN-T5 [8], using version 1 of the validation set. Table 3 summarizes the performance of different language encoders, showing that the SBERT models perform best. We consider TTC less relevant here. We hypothesize that the difference comes from the training objectives of language models. SBERT is trained to map semantically similar sentences closer to each other in the embedding space. This is particularly useful when the model needs to understand descriptions with semantically close words, e.g. \"cut carrots\" and \"slice carrots\". GPT-2 and FLAN-T5 are optimized using generative objectives, which makes them overly sensitive to the input word choice." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.518, + 0.892, + 0.548 + ], + "angle": 0, + "content": "Fusion module. We provide ablations regarding the fusion module in the Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.569, + 0.744, + 0.585 + ], + "angle": 0, + "content": "5. Conclusion and discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.853 + ], + "angle": 0, + "content": "We propose TransFusion, a multimodal fusion model that anticipates object interactions by considering the past action context represented by language summaries. Our successful experiments on a challenging egocentric video dataset demonstrate how language summaries and the vast knowledge contained in vision-language foundation models can improve object interaction anticipation, highlighting the representational power of language. It appears promising to investigate fusing even more modalities, such as motion cues from optical flow. We emphasize that the proposed summarization approach is not limited to this task. Language provides a universal interface to complement the visual input with information encoded in foundation models or task-specific pipelines. Future work can leverage language summarization for other video reasoning tasks [60, 63], or extend it temporally to describe possible future activities [1, 27] for long-term interaction anticipation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by an ETH Zurich Postdoctoral Fellowship. We thank Adrian Spurr for the helpful comments and Velko Vechev for the voiceover." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18293" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.091, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Chaitanya Ahuja and Louis-Philippe Morency. Language2Pose: Natural Language Grounded Pose Forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719–728, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.471, + 0.227 + ], + "angle": 0, + "content": "[2] Alan Akbik, Duncan Blythe, and Roland Vollgraf. Contextual String Embeddings for Sequence Labeling. In Proceedings of the 27th international conference on computational linguistics, pages 1638-1649, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.296 + ], + "angle": 0, + "content": "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. VQA: Visual Question Answering. In Proceedings of the IEEE International Conference on Computer Vision, pages 2425-2433, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.339 + ], + "angle": 0, + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. CoRR, abs/2104.00650, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.471, + 0.382 + ], + "angle": 0, + "content": "[5] Gedas Bertasius, Hyun Soo Park, Stella X. Yu, and Jianbo Shi. First-Person Action-Object Detection with EgoNet. In Proceedings of Robotics: Science and Systems, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.471, + 0.41 + ], + "angle": 0, + "content": "[6] Jerome S Bruner. Intention in the Structure of Action and Interaction. Advances in Infancy Research, 1981. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.412, + 0.471, + 0.452 + ], + "angle": 0, + "content": "[7] Guo Chen et al. InternVideo-Ego4D: A Pack of Champion Solutions to Ego4D Challenges. arXiv preprint arXiv:2211.09529, 2022.2,6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.454, + 0.471, + 0.578 + ], + "angle": 0, + "content": "[8] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling Instruction-Finetuned Language Models, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.58, + 0.471, + 0.647 + ], + "angle": 0, + "content": "[9] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling Egocentric Vision. CoRR, abs/2006.13256, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.471, + 0.732 + ], + "angle": 0, + "content": "[10] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling Egocentric Vision: Collection, Pipeline and Challenges for EPIC-Kitchens-100. International Journal of Computer Vision, 130(1):33-55, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.471, + 0.815 + ], + "angle": 0, + "content": "[11] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. EPIC-KITCHENS VISOR Benchmark: VVideo Segmentations and Object Relations. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[12] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. TransVG: End-to-End Visual Grounding with Transformers. CoRR, abs/2104.08541, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Transformers for Language Understanding. CoRR, abs/1810.04805, 2018.3,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. CoRR, abs/2010.11929, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[15] Christoph Feichtenhofer. X3D: Expanding Architectures for Efficient Video Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 203-213, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.265, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[16] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast Networks for Video Recognition. CoRR, abs/1812.03982, 2018. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.309, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[17] Antonino Furnari and Giovanni Maria Farinella. Rolling-Unrolling LSTMs for Action Anticipation from First-Person Video. CoRR, abs/2005.02190, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.352, + 0.892, + 0.407 + ], + "angle": 0, + "content": "[18] Antonino Furnari, Sebastiano Battiato, Kristen Grauman, and Giovanni Maria Farinella. Next-Active-Object Prediction from Egocentric Videos. CoRR, abs/1904.05250, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.41, + 0.892, + 0.479 + ], + "angle": 0, + "content": "[19] H. Girase, N. Agarwal, C. Choi, and K. Mangalam. Latency Matters: Real-Time Action Forecasting Transformer. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18759-18769, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.481, + 0.892, + 0.508 + ], + "angle": 0, + "content": "[20] Rohit Girdhar and Kristen Grauman. Anticipative Video Transformer. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[21] Ross B. Girshick. Fast R-CNN. CoRR, abs/1504.08083, 2015.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.541, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[22] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Wesley Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4D: Around the World in 3,000 Hours of Egocentric Video. CoRR, abs/2110.07058, 2021. 1, 2, 4, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18294" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. CoRR, abs/1512.03385, 2015. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.134, + 0.469, + 0.188 + ], + "angle": 0, + "content": "[24] Zhicheng Huang, Zhaoyang Zeng, Bei Liu, Dongmei Fu, and Jianlong Fu. Pixel-BERT: Aligning Image Pixels with Text by Deep Multi-Modal Transformers. CoRR, abs/2004.00849, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.19, + 0.469, + 0.244 + ], + "angle": 0, + "content": "[25] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. EPIC-Fusion: Audio-Visual Temporal Binding for Egocentric Action Recognition. CoRR, abs/1908.08498, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.246, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[26] Weicheng Kuo, Fred Bertsch, Wei Li, AJ Piergiovanni, Mohammad Saffar, and Anelia Angelova. FindIt: Generalized Localization with Natural Language Queries. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI, pages 502-520. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.33, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[27] Yen-Ling Kuo, Xin Huang, Andrei Barbu, Stephen G McGill, Boris Katz, John J Leonard, and Guy Rosman. Trajectory Prediction with Linguistic Representations. In 2022 International Conference on Robotics and Automation (ICRA), pages 2868-2875. IEEE, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.4, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[28] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. Less is More: ClipBERT for Video-and-Language Learning via Sparse Sampling. CoRR, abs/2102.06183, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.455, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Piotr Dólar, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature Pyramid Networks for Object Detection. CoRR, abs/1612.03144, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.511, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[30] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen Clip Models Are Efficient Video Learners. In European Conference on Computer Vision, pages 388-404. Springer, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.581, + 0.469, + 0.634 + ], + "angle": 0, + "content": "[31] Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the Variance of the Adaptive Learning Rate and Beyond. CoRR, abs/1908.03265, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.636, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[32] Miao Liu, Siyu Tang, Yin Li, and James M. Rehg. Forecasting Human Object Interaction: Joint Prediction of Motor Attention and Egocentric Activity. CoRR, abs/1911.10967, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.692, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[33] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A Robustly Optimized BERT Pretraining Approach. CoRR, abs/1907.11692, 2019.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.762, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[34] TorchVision maintainers and contributors. TorchVision: PyTorch's Computer Vision Library. https://github.com/pytorch/vision, 2016.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.804, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[35] David W Orr. The Nature of Design: Ecology, Culture, and Human Intention. Oxford University Press, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.832, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[36] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou, Rama Chellappa, and Pengchuan Zhang. EGOVLPV2: Egocentric Video-Language Pre-training With Fusion in the Backbone, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[37] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language Models are Unsupervised Multitask Learners. OpenAI Blog, 2019. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. CoRR, abs/2103.00020, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[39] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. CoRR, abs/1910.10683, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.294, + 0.892, + 0.335 + ], + "angle": 0, + "content": "[40] Francesco Ragusa et al. StillFast: An End-to-End Approach for Short-Term Object Interaction Anticipation. arXiv preprint arXiv:2304.03959, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.338, + 0.892, + 0.379 + ], + "angle": 0, + "content": "[41] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. CoRR, abs/1908.10084, 2019. 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.382, + 0.892, + 0.424 + ], + "angle": 0, + "content": "[42] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. CoRR, abs/1506.01497, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.426, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[43] Paschal Sheeran. Intention-Behavior Relations: A Conceptual and Empirical Review. European Review of Social Psychology, 12(1):1-36, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[44] Chen Sun, Fabien Baradel, Kevin Murphy, and Cordelia Schmid. Contrastive Bidirectional Transformer for Temporal Representation Learning. CoRR, abs/1906.05743, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[45] Hao Tan and Mohit Bansal. LXMERT: Learning Cross-Modality Encoder Representations from Transformers. CoRR, abs/1908.07490, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.558, + 0.892, + 0.627 + ], + "angle": 0, + "content": "[46] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Enhancing Next Active Object-Based Egocentric Action Anticipation With Guided Attention. In 2023 IEEE International Conference on Image Processing (ICIP), pages 1450-1454. IEEE, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[47] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Guided Attention for Next Active Object @ EGO4D STA Challenge. arXiv preprint arXiv:2305.16066, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.769 + ], + "angle": 0, + "content": "[48] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Leveraging Next-Active Objects for Context-Aware Anticipation in Egocentric Videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 8657–8666, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.841 + ], + "angle": 0, + "content": "[49] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked Autoencoders Are Data-Efficient Learners for Self-Supervised Video Pre-Training. Advances in Neural Information Processing Systems, 35:10078–10093, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. CoRR, abs/1706.03762, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18295" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[51] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-Anticipation Transformer for Online Action Understanding, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[52] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Action-CLIP: A New Paradigm for Video Action Recognition. arXiv preprint arXiv:2109.08472, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.468, + 0.248 + ], + "angle": 0, + "content": "[53] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework. CoRR, abs/2202.03052, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.468, + 0.318 + ], + "angle": 0, + "content": "[54] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. MeMViT: Memory-Augmented Multiscale Vision Transformer for Efficient Long-Term Video Recognition. CoRR, abs/2201.08383, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.321, + 0.468, + 0.361 + ], + "angle": 0, + "content": "[55] Yu Wu, Linchao Zhu, Xiaohan Wang, Yi Yang, and Fei Wu. Learning to Anticipate Egocentric Actions by Imagination. CoRR, abs/2101.04924, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.363, + 0.468, + 0.432 + ], + "angle": 0, + "content": "[56] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, Attend and Tell: Neural Image Caption Generation With Visual Attention. In International Conference on Machine Learning, pages 2048-2057. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.434, + 0.468, + 0.489 + ], + "angle": 0, + "content": "[57] Ting Yao, Yingwei Pan, Yehao Li, Zhaofan Qiu, and Tao Mei. Boosting Image Captioning With Attributes. In Proceedings of the IEEE International Conference on Computer Vision, pages 4894-4902, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.491, + 0.468, + 0.532 + ], + "angle": 0, + "content": "[58] Moon Ye-Bin, Jisoo Kim, Hongyeob Kim, Kilho Son, and Tae-Hyun Oh. TextMania: Enriching Visual Feature by Text-driven Manifold Augmentation, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.534, + 0.468, + 0.588 + ], + "angle": 0, + "content": "[59] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. MERLOT: Multimodal Neural Script Knowledge Models. CoRR, abs/2106.02636, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.591, + 0.468, + 0.659 + ], + "angle": 0, + "content": "[60] Kuo-Hao Zeng, Tseng-Hung Chen, Ching-Yao Chuang, Yuan-Hong Liao, Juan Carlos Niebles, and Min Sun. Leveraging Video Descriptions to Learn Video Question Answering. In Thirty-First AAAI Conference on Artificial Intelligence, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.661, + 0.468, + 0.731 + ], + "angle": 0, + "content": "[61] Ce Zhang, Changcheng Fu, Shijie Wang, Nakul Agarwal, Kwonjoon Lee, Chiho Choi, and Chen Sun. Object-Centric Video Representation for Long-term Action Anticipation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6751-6761, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.732, + 0.468, + 0.8 + ], + "angle": 0, + "content": "[62] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection. In The Eleventh International Conference on Learning Representations, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.803, + 0.468, + 0.858 + ], + "angle": 0, + "content": "[63] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. Video Summarization With Long Short-Term Memory. In European Conference on Computer Vision, pages 766-782. Springer, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[64] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning Video Representations From Large Language Models, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[65] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple Multi-Dataset Detection. CoRR, abs/2102.13086, 2021. 8" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18296" + } + ] +] \ No newline at end of file diff --git a/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/5f051557-0067-4cd2-89f8-ff27dddafa6f_origin.pdf b/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/5f051557-0067-4cd2-89f8-ff27dddafa6f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..eb4cc0dd3b1e57c927c108585a7e31034293aefd --- /dev/null +++ b/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/5f051557-0067-4cd2-89f8-ff27dddafa6f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ea12b98dd3d195cf6543f1cc8c990f35091d14d96904292904f82b86b50b569 +size 5265644 diff --git a/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/full.md b/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..232292b49c5c04942fa1396a36ebbd43abf34e9d --- /dev/null +++ b/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/full.md @@ -0,0 +1,284 @@ +# Summarize the Past to Predict the Future: Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation + +Razvan-George Pasca $^{1*}$ Alexey Gavyushin $^{1*}$ Muhammad Hamza $^{2}$ Yen-Ling Kuo $^{3}$ Kaichun Mo $^{4}$ Luc Van Gool $^{1,5,6}$ Otmar Hilliges $^{1}$ Xi Wang $^{1}$ ETH Zurich Univ. of Zurich Univ. of Virginia NVIDIA KU Leuven INSAIT, Sofia + +![](images/0f23b4421a71e33743c550be111aa1ad4300ae983910743e70882576c09612d4.jpg) +Figure 1. TransFusion: Multimodal fusion transformer for short-term object interaction anticipation. Given a video sequence of past observations, the task aims to predict a set of objects visible in the current frame that will be interacted with in the future, i.e. in the activation frame that is $\Delta$ frames away from the current prediction frame. Additionally, the task requires estimating the bounding box, the associated action described by a verb-noun pair, and the time to contact for each predicted object. We propose TransFusion, a multimodal fusion architecture that uses language summaries of past actions and salient objects to effectively predict future object interactions. + +# Abstract + +We study object interaction anticipation in egocentric videos. This task requires an understanding of the spatiotemporal context formed by past actions on objects, coined action context. We propose TransFusion, a multimodal transformer-based architecture for short-term object interaction anticipation. Our method exploits the representational power of language by summarizing the action context textually, after leveraging pre-trained vision-language foundation models to extract the action context from past video frames. The summarized action context and the last observed video frame are processed by the multimodal fusion module to forecast the next object interaction. Experiments on the Ego4D next active object interaction dataset show the effectiveness of our multimodal fusion model and highlight the benefits of using the power of foundation models and language-based context summaries in a task where + +vision may appear to suffice. Our novel approach outperforms all state-of-the-art methods on both versions of the Ego4D dataset. A project video and code are available at https://eth-ait.github.io/transfusion-proj/. + +# 1. Introduction + +The ability to predict plausible future human-object interactions is important for many assistive technologies. The task of short-term object interaction anticipation [18, 22] is defined as predicting which object a user will interact with next, what action will be performed and when in the future, given an egocentric video input. Providing an effective solution would help artificial agents, such as care robots, assist humans in their daily activities, as well as enable safety applications such as child and elderly monitoring. + +This task is challenging because it requires reasoning about user intent [6, 35, 43], which is not directly observable. For example, if the video shows the user holding + +a vegetable, the most likely next action is to slice it, but only with a knife and cutting board present. Without those, there is significantly more uncertainty. Therefore, understanding what has happened in the past and what relevant objects are present is critical for predicting the next object interaction. We collectively refer to these concepts as action context. Existing interaction forecasting methods do not explicitly model action context and rely on neural networks to extract that information from fixed-sized video chunks [5, 18, 22, 32]. + +In this paper, we propose to use language summarization of the past as an explicit representation of action context. This filters out ambiguities caused by visual clutter. It also puts focus on objects and actions directly relevant to the ongoing and likely future activities. For the object interaction anticipation, we next introduce a multimodal fusion approach. Specifically, we propose a transformer-based multimodal fusion model, TransFusion. As shown in Figure 1, our model takes language summaries of the action context and the last observed single frame of a video as input and predicts future object interactions. We employ an existing image captioning system [53] to generate a concise description of past actions given by pairs of verbs and nouns. We use CLIP [38], a vision-language model trained on pairs of images and captions, to detect salient objects in the prediction frame of interest, and deem these objects task-relevant. The verb-noun pairs together with the list of salient objects form the summary of the past action context. Leveraging the vast knowledge contained in pre-trained vision-language foundation models yields more generic representations of the past action context and enables us to better generalize towards various scenarios, in particular rare ones not well-represented in the dataset. + +Our experiments show that our language-based summaries can effectively represent the past action context. We evaluate on both versions of the challenging egocentric dataset Ego4D [22]. Our approach improves interaction prediction accuracy and outperforms state-of-the-art models pre-trained on large video datasets [7, 40]. Such improvements show that large vision-language models can be useful even for a task that seems purely visual. + +In summary, our contributions are: + +- A strategy for generating concise language-based summaries of action context over longer timespans from frame-wise language-based summaries, for instance obtained using vision-language foundation models. +- TransFusion, a multimodal fusion model for short-term object interaction anticipation combining action context descriptions and visual features. +- Experiments on both versions of the Ego4D dataset showing improvements over state-of-the-art methods on the short-term object interaction anticipation task, in particular for long-tailed classes. + +# 2. Related work + +Object interaction anticipation. Several works have been proposed for human-object interaction prediction. Furnari et al. [18] first introduced next-active-object (NAO) prediction, which predicts the active/not active label for each object using its motion trajectory features, but doesn't consider action anticipation. Closely related to NAO, Bertasius et al. [5] introduced action-objects, which considers objects that capture the human actor's visual attention or tactile interaction. Bertasius et al. use the data from a stereo camera system in a two-stream network, integrating RGB and depth information. They show that certain aspects of human actions can be predicted by exploiting the spatial configuration of the objects and the actor's head. Their approach is not easily scalable due to the need for stereo data. + +In addition to different tasks, Liu et al. [32] propose a model that improves NAO prediction by incorporating future hand trajectory modeling. They obtain impressive results using 3D CNNs, confirming that hand motion cues and scene dynamics have important explanatory power in predicting the short-term future. Nonetheless, their approach also encounters scaling limitations because additional ground-truth labels for hand trajectories are needed. + +The Ego4D dataset [22] proposed object interaction anticipation. Apart from locating the NAO, the task in Ego4D also requires the prediction of associated nouns and verbs, as well as the time to contact (TTC) for the future interacted objects. The official Ego4D baseline uses the current frame and past video frames as input and employs a two-stage approach to first learn noun and box prediction, and afterwards learn verb and TTC prediction. InternVideo [7] similarly leverages a DINO [62] model to detect objects in the first stage, and a VideoMAE [49]-pretrained ViT [14] spacetime attention model to predict verbs and TTC conditioned on the object predictions in the second stage. StillFast [40] presents an end-to-end method for NAO prediction, fusing multi-scale high-resolution image and low-resolution video features and combining local RoI features with global scene features to boost the interaction anticipation performance. GANOv2 [46, 47] and its predecessor [48] extract features from objects detected in past observed frames and fuse them with global frame features, forwarding them to a StillFast predictor and a motion-object-separating encoder-decoder architecture, respectively. + +Action anticipation. Besides localizing future object interactions, predicting the next action steps is an equally important task that intelligent systems need to perform to interact with humans. A longer prediction timespan requires a better grasp of the structure behind the succession of actions. A pioneering advancement in this field is the introduction of the EPIC-Kitchens 100 (EK100) [9] dataset, which facilitates research in action anticipation within egocentric + +videos. Significant research efforts have been dedicated to exploring the action anticipation task utilizing EK100. Anticipative Video Transformer (AVT) [20] is one of the best-performing architectures on the EK100 [9] action anticipation benchmark. AVT uses a pre-trained Vision Transformer [14] as the image feature backbone and attends to the previously observed video frames to anticipate future actions. MeMViT [54] improves AVT by exploiting a longer temporal context, highlighting the importance of modeling a long sequence of previous actions. Other model architectures such as recurrent neural networks [17, 55], multimodal temporal CNNs [25] and transformers [19, 51] are also used to model the temporal contexts. Note that these methods are designed for action classification using video data and adapting them for object detection is not trivial. It is worth noting that the action anticipation task from the EK100 benchmark is a subtask of our short-term object interaction anticipation task. Our task not only predicts the action label of verb-noun pairs, but we also predict the location of the object and the time to contact, which is more challenging than the action anticipation task itself. TransFusion focuses on detecting the next active objects, as all evaluation metrics are conditioned on the box IoU. Therefore, comparing our work with the state-of-the-art methods for the action anticipation task on EK100 is out of the scope of this work. The Ego4D dataset itself further introduces a long-term action anticipation benchmark, which [61] tackles using pre-trained visual grounding models. + +Vision-language models. Our approach combines the visual and language modalities to improve object interaction anticipation. Most recent multimodal vision-language architectures are based on Transformer [50] fusion schemes applied on features extracted by various encoders [4, 12, 26]. We follow their spirit by employing self-attention-based modality fusion. Our work differs in one significant aspect: the language features represent the past action context. This serves as an additional signal to disambiguate the actor's intent given similar visual scenes. The model must learn to predict from the multimodal fusion of language context summaries and visual cues in a considerably smaller data regime. + +More broadly, there are numerous transformer-based works aimed at learning cross-modal representations between vision and language [24, 28, 36, 38, 44, 45, 58, 59, 64] using terabyte-scale datasets crawled from multiple internet sources. When combined with large language models such as GPT-2 [37], BERT [13], Sentence-BERT [41], and the T5 variants [8, 39], these vision-language models can be applied to several downstream tasks such as VQA [3] and image captioning [56, 57]. Similar to other vision-language models that are trained with language input, our approach leverages state-of-the-art image captioning models [53] to provide a consistent description of past actions. + +# 3. Language-guided prediction + +We use language-based context summaries to represent the spatial-temporal context formed by past actions and object relationships. The context summary offers an explicit yet compact representation, which allows us to focus on task-relevant content and filter out unimportant information from the likely cluttered environment. To this end, we propose TransFusion, a transformer-based model that leverages a joint-attention fusion mechanism to effectively learn to predict object interactions in the future from the past action context. In the following, we describe our framework, which takes video clips as input and anticipates the next object interaction. We begin by defining the object interaction anticipation task (Sec. 3.1). We then show how to generate a summary for a prediction frame from an egocentric video to represent the action context provided by the past frames (Sec. 3.2). Finally, we present our multimodal fusion model's architecture (Sec. 3.3). + +# 3.1. Task definition + +Given a video $\mathcal{V} = \{f_1,\dots ,f_T\}$ of length $T$ as input, our goal is to predict the next object interaction $\mathcal{O}$ in the last observed video frame $f_{T}$ , possibly using older frames to extract context information. We call $f_{T}$ the prediction frame. + +As illustrated in Figure 1, the task is to anticipate the object interaction in $f_{T}$ . $\mathcal{O}_T = \{o_1^T,\dots ,o_N^T\}$ denotes the set of future object interactions, with $N > 1$ whenever multiple objects are interacted with simultaneously. In other words, multiple $o_i^T$ can be detected in the prediction frame $f_{T}$ . For example, $N = 2$ when one hand is holding two mugs in a kitchen scene, or the left hand is touching a plant while the right hand is moving a pot in a garden scene. Each object interaction $o_i^T = (b_i^T,n_i^T,v_i^T,t_i^T)$ consists of the object bounding box $b_{i}^{T}$ , the semantic object label (noun) $n_i^T$ , the action label (verb) $v_{i}^{T}$ , and the time to contact (TTC) $t_i^T$ . The task is to predict the next active object location $b_{i}^{T}$ in the prediction frame $f_{T}$ , the associated action described by the verb-noun pair $(n_i^T,v_i^T)$ , and the time $t_i^T$ to the point of contact when the interaction is going to start. The task is set up such that the actual object interaction takes place after a buffer time $\Delta$ on frame $f_{T + \Delta}$ and the predicted time to contact $t_i^T = \Delta$ . Ego4D uses $\Delta \geq 0.033s$ . This requires anticipating future actions and increases the task difficulty. + +# 3.2. Summarizing the past + +Anticipating object interactions in the future requires an understanding of what happened in the past. We first infer the per-frame action context from each egocentric video. This process consists of extracting action descriptions and held objects. These information sources are generated and aggregated independently across frames and finally forwarded as the action context to our anticipation model. + +![](images/dd1f4c7a5e8616dfd8819f244e47559a7fb781f4fd0cbae7c8aa31790cffbb6f.jpg) +Figure 2. Overview of the TransFusion model. TransFusion takes the prediction frame and the action context summary as input and predicts the bounding box of the next-active object, the noun-verb pair describing the associated action, and the time to contact (TTC). Feature maps of different scales are extracted from the visual encoder and then fused via a multimodal fusion module with the encoded language features. Their output is then processed by a regular feature pyramid network (FPN), denoted as multi-scale region proposal networks, before being fed into the Faster R-CNN detector. + +Extracting frame-wise action descriptions. Inspired by the label format used in egocentric video datasets [10, 22], we aim to describe the past actions by sequences of verb-noun pairs (e.g. "wash tomato", "cut tomato"), so-called action descriptions. For each video frame $f$ , we extract the action currently performed by the agent, represented by a single pair $(v, n)$ of a verb $v$ and a noun $n$ . + +We use an off-the-shelf image captioning model [53] to generate multiple full-sentence captions for each frame of interest, so as to find repeatedly occurring outputs and reduce noise: we perform part-of-speech tagging [2] on these captions and collect candidate pairs consisting of verbs followed by nouns within some cutoff distance $d \in \mathbb{N}_0^+$ . Setting $d > 0$ allows us to account for additional words appearing between a verb-noun pair (e.g. "eat apple" from "A person eating a red apple") and thereby obtain more valid candidate pairs from the captions. However, larger cutoff distances may introduce more spurious detections (e.g. "eat gathering", from "A person eating while at a gathering"). We use $d = 4$ in our experiments. The most frequent candidate pair across the frame's captions is selected as the verb-noun action description of this frame. If no valid verb-noun pair is found, the frame will have no action description. + +Extracting frame-wise salient objects. Another characteristic shared by many daily action sequences is that the next active object is likely to have already appeared in the actor's view before the upcoming interaction. The current environment in the recording may also provide model conditioning. Therefore, we select a set of salient objects in each frame as (1) the potential candidates for future active objects and (2) a proxy of the surroundings. + +Encouraged by its previous successful use in action-related tasks [30, 52], we use CLIP [38], a pre-trained open-vocabulary zero-shot classifier, to compute the cosine similarities between the CLIP image embedding of a frame and the embeddings of the object categories from the Ego4D do + +main. We choose to keep the highest-scoring $k$ objects as the set of salient objects for a given frame. A small $k$ may fail to provide sufficient information, while a large $k$ makes capturing the true salient objects more likely, yet introduces more noise due to false detections. We set $k = 3$ . + +Aggregating summaries over multiple frames. After the frame-wise summarization step, we have, for each frame $t$ , at most one verb-noun pair describing the frame-wise action $a_{t} = (n,v) \in \mathcal{A}_{t}^{F}$ . We then apply a cross-frame aggregation scheme to all $\mathcal{A}_{t}^{F}$ so as to identify contiguous segments of frames exhibiting the same activity, while accounting for individual noisy frames possibly not matching the framework summaries of their neighbors. The aggregation processes the frames in temporal order. We consider frames belonging to the same atomic action as an action segment, which is initiated by observing a number of reoccurrences of some verb-noun pair and terminated by a lack of its occurrence within a number of contiguous frames. After the aggregation, we obtain $\mathcal{A}$ , the sequence of all action segments in the video. For each prediction frame, we construct its action context from a number of action segments preceding the frame, and the objects from $\mathcal{N}_s$ in that frame. + +By aggregating across frames, we thus capture dominant actions in the past, reduce noise in the generated action context and represent long timespans of similar activity by short textual descriptions. The complete action context is provided to our model by concatenating the textual representations of the aggregation results of any desired subset of the two context representations $\{\mathcal{A},\mathcal{N}_s\}$ , where we drop the index $t$ for convenience. For more details, see the Supp. Mat. + +# 3.3. TransFusion: Multimodal fusion + +Figure 2 visualizes the architecture of the proposed multimodal TransFusion model. TransFusion employs two different base encoders, for language and image input respectively. The input prediction frame $f_{T}$ is processed by a regu + +![](images/ec1da6b14eb441a83d93e834c2f4a065392b12cc24c9c1f485faa788592b3b40.jpg) +Figure 3. Multimodal fusion module. We first project the CNN feature map and the language tokens to a common dimensionality before adding the specific embeddings. We concatenate the visual and language tokens and feed them to $M$ self-attention layers. At the output, the fused visual tokens are projected back into the initial feature map shape. + +lar CNN-based visual encoder [23], producing a set of multiscale feature maps $\mathcal{F}_{\mathcal{V}}^{s}(f_{T})$ , with $s$ indicating the scale. The frame's action context is encoded by a language encoder [41] and yields $\mathcal{F}_{\mathcal{L}}(\mathcal{C}_L)$ . The feature maps belonging to different scale levels from the visual encoder are fused with the encoded language features via a joint-attention scheme. The result is then processed by a regular feature pyramid network (FPN) [29] before being forwarded to the Faster R-CNN [42] detector. + +Multimodal fusion. For simplicity, we describe the multimodal fusion module for a single input scale $s$ , which we hence omit from notation. See Figure 3 for a detailed view of a single Transformer encoder layer together with the input projection stages. For the visual modality, we first split the image into a sequence of tokens. Specifically, the visual features $\mathcal{F}_{\mathcal{V}}(f_T) \in \mathbb{R}^{C \times H \times W}$ are reshaped to $\mathbb{R}^{N \times (P^2 \cdot C)}$ where $P$ is the token size and $N = \frac{H \times W}{P^2}$ the number of tokens. We add unique type embeddings as well as a positional encoding to each modality. Both visual and language tokens are first mapped to $\mathbb{R}^D$ via a linear projection before being fed to the fusion module in a concatenated form. The fusion block consists of a regular Transformer encoder layer replicated $M$ times. Finally, the visual tokens are regrouped into their initial shape to compose the feature map for the feature pyramid network (FPN). + +Multiscale fusion. The multimodal fusion block outputs multiple fused feature maps of varying scales: $\mathcal{F}_{fi} \in \mathbb{R}^{C_i \times H_i \times W_i}$ , where $i$ is the scale index. They are fed into the FPN, which has different parameters for each feature + +map, allowing the network to learn specialized features for each downsampling ratio. The output of the FPN is processed by a Faster R-CNN detector. + +Prediction head networks. The Faster R-CNN detector outputs the bounding box $b$ through a regression layer, an associated object $n$ and verb $v$ , and time to contact $t$ . + +Training objective. We train TransFusion using the standard Faster R-CNN [42] objective $\mathcal{L}_{box}$ for bounding box prediction, which consists of localization and objectness terms; further, cross-entropy losses $\mathcal{L}_{noun}$ and $\mathcal{L}_{verb}$ for noun and verb prediction respectively, and an L1 loss $\mathcal{L}_{ttc}$ for time to contact prediction. The overall objective is + +$$ +\mathcal {L} = \mathcal {L} _ {\text {b o x}} + \mathcal {L} _ {\text {n o u n}} + \mathcal {L} _ {\text {v e r b}} + \mathcal {L} _ {\text {t t c}}. \tag {1} +$$ + +As defined in Faster R-CNN, $\mathcal{L}_{box}$ is + +$$ +\mathcal {L} _ {\text {b o x}} = \frac {1}{N _ {\mathrm {c l s}}} \sum_ {i} \mathcal {L} _ {\mathrm {c l s}} \left(p _ {i}, p _ {i} ^ {*}\right) + \frac {\lambda}{N _ {\mathrm {r e g}}} \sum_ {i} p _ {i} ^ {*} \cdot L _ {\mathrm {r e g}} \left(b _ {i} - b _ {i} ^ {*}\right), \tag {2} +$$ + +where $p_i$ is the predicted probability of a box being classified as a foreground object (true positive), $p_i^*$ is the ground-truth label; $b_i$ and $b_i^*$ are the predicted and ground-truth box coordinates. $N_{\mathrm{reg}}$ is the number of bounding boxes used in training, $N_{\mathrm{cls}}$ the detection network's image batch size. We leave $\lambda = 11$ as in the reference implementation. + +# 4. Experiments + +Dataset. We use both versions (v1 and v2) of the dataset provided for the short-term object interaction anticipation task in Ego4D [22]. It contains various activities ranging from gardening work to cooking or car parts changing. There are in total 64,798 (v2: 165,451) annotated video clips of $\sim 5$ minutes each with frame rate $fps = 30$ . The dataset has a highly imbalanced long tail distribution, including 87 nouns and 74 verbs (v2: 128 and 81 resp.). For instance, the "take" class encompasses almost $40\%$ of all the verb labels in version 1, while "mold" appears only in one video. The train set consists of about 28k (v2: 98k) samples, while the rest are roughly equally split between validation and test. + +Evaluation. As in the official Ego4D benchmark [22], we use the Top-5 mean Average Precision (mAP) to evaluate the performance of individual predictions. It considers two constraints: The IoU constraint requires that the predicted boxes are counted as hits only if there is an intersection over union IoU $\geq 0.5$ with the ground-truth box. The TTC constraint requires that the predicted time-to-contact $t_i$ is close enough to the ground-truth $\hat{t}_i$ , i.e. $|t_i - \hat{t}_i| < T_\delta, T_\delta = 0.25$ . The following metrics are considered: + +- Noun refers to the Noun-Box Top-5 mAP. In addition to the IoU constraint, a correct match with the ground-truth noun is required. + +
Ver.ModelN↑N-V ↑N-T ↑A↑
v1FRCNN+Rnd. [22]20.452.223.860.44
FRCNN+SF [22]20.456.786.172.45
InternVideo [7]24.609.187.643.40
StillFast [40]19.519.956.453.49
TF (ours)24.9010.067.573.54
v2FRCNN+SF [22]26.159.458.693.61
StillFast [40]25.0613.299.145.12
GANov2 [46]25.6713.609.025.16
TF (ours)30.4313.4510.385.18
+ +Table 1. Comparison with the state-of-the-art. Our method TransFusion (TF) outperforms all state-of-the-art approaches on 3 out of 4 metrics, on both versions of the Ego4D test set. We report the top-5 mean average precision for the Noun (N), NounVerb (N-V), Noun-TTC (N-T), and Noun-Verb-TTC (A) metrics. + +- Noun-Verb refers to the Noun-Verb-Box Top-5 mAP. Correct noun and verb match with the IoU constraint. +- Noun-TTC refers to the Noun-Box Top-5 mAP. Correct noun matches with the TTC constraint. +- Noun-Verb-TTC ("Overall") refers to the overall Top-5 mAP, intersecting correct noun-verb matches with IoU and TTC constraints. + +Note that since detecting the next-active objects is the main focus of the task, all evaluation metrics are conditioned on the IoU constraint. + +Implementation details. We make use of the pre-trained Torchvision [34] Faster R-CNN. ResNet-50 [23] is used as the visual encoder and Sentence-BERT (SBERT) [41] is used as the language encoder. We freeze ResNet-50 and apply the multimodal fusion scheme on top. We use the RAdam [31] optimizer with a weight decay of $2e^{-4}$ , and an effective batch size of 32. Besides the default Faster R-CNN background class in nouns, we also add it in the verb class, which improves the Noun-Verb mAP by almost 2 points. We reduce the region proposal network's sampling batch size and the detection network's image batch size and use data augmentation techniques to reduce overfitting. The number of verb-noun pairs used in the input is defined as the context length $L_{c}$ . On average, an action segment described by one verb-noun phrase corresponds to one second with 30 frames, as computed from the GT annotations. Unless specified otherwise, we use $L_{c} = 3$ . We perform postprocessing involving non-maximum suppression on the predicted bounding boxes, as elaborated in the Supp. Mat. + +# 4.1. Comparison with state-of-the-art + +We first compare our method against the baselines proposed in the Ego4D work [22]. The FRCNN + SF method employs a two-stage approach: a ResNet-based Faster R-CNN detector first produces bounding box regressions and noun predictions from the prediction frame, without using any video features. In the second stage, a SlowFast [16] 3D + +![](images/074104493356d0743a5070d5e10fd5fb5890bc65383919de0f728cf803c48a64.jpg) +Figure 4. Prediction of Noun-Verb in relation to time-to-contact. Histogram of the time-to-contact labels in Ego4Dv1 is shown in blue bars. Performance measured in Noun-Verb is plotted as a function of time-to-contact. We compare our method to Ego4D's official FRCNN + SF [22] baseline. + +CNN video processing module performs region-of-interest (RoI) pooling [21] on the corresponding SlowFast 3D CNN video features. The pooled video features are then used by RoI heads to predict the verb and TTC values. The FR-CNN + Rnd. baseline, with results published on v1 only, drops the second stage of FRCNN + SF and uses randomly drawn verb and TTC predictions, following their respective distributions in the training set. + +Another baseline on v1 is given by InternVideo [7], a method achieving strong results on the short-term object interaction anticipation task with a two-stage strategy similar to FRCNN+SF: bounding boxes and noun labels are first obtained using a DINO-based object detector [62], after which a VideoMAE-pretrained [49] ViT [14] model leverages space-time attention to predict the corresponding verb classes and TTC values. + +We further compare TransFusion to the recent StillFast [40] approach on both dataset versions. StillFast uses a Faster-RCNN-based spatial stream operating on a high-resolution version of the input image, and an X3D-based [15] temporal stream operating on a low-resolution version of the input video. The two streams each produce a multiscale feature map. Both feature maps are fused into a combined feature pyramid, from which local RoI features and global scene features are extracted, merged by a fusion network, and finally used by linear layers to predict each boxes' noun, verb and TTC. + +GANOV2 [46] enhances StillFast by adding a multi-head guided attention mechanism to its temporal stream, achieving favorable performance on version 2 of the dataset. The mechanism attends to object detections produced from the network's temporal branch, together with patches from the 3-D temporal feature stack, in a cross-fusion manner, and uses the fused tokens as the input to StillFast's feature pyramid and subsequent architecture. + +Table 1 demonstrates that TransFusion outperforms all state-of-the-art methods in terms of the N, N-V and A met- + +![](images/edbb1320c1404778505b6b5f81388d0d766cb2538b2dca253dc38b7479457ed9.jpg) +Figure 5. Classification performance on top/tail categories. We show the relative and absolute gains of TransFusion over Ego4D's FRCNN + SF baseline [22] on the validation set, for Noun-Only and Verb-Only mAP (without any box IoU constraint). Relative improvements are written on top of the red bars. + +rics on version 1 of the dataset, and approaches the state-of-the art performance on the N-T metric while requiring a fraction of the competitor's resources during training [7] (see the Supp. Mat. for details). On the dataset's second version, TransFusion performs best in terms of N, N-T and A, while approaching the runner-up method on N-V. These favorable results validate the advantages of modeling action context explicitly for the interaction anticipation task. + +We henceforth analyze our performance gains over the official Ego4D FRCNN + SF baseline on the first version of the Ego4D dataset. Figure 4 shows the long tail distribution of the time-to-contact labels in the validation set, split into buckets. We see that TransFusion consistently outperforms the baseline on the N-V metric for all TTC ranges. The results in Figure 5 indicate that our method not only outperforms the Ego4D method on top noun and verb classes, but also achieves bigger gains in the tail classes, demonstrating the generalization ability of using our language-based context summaries. Top and tail buckets represent roughly $50\%$ of the samples in the validation set. + +Figure 6 shows that the Ego4D approach manages to detect the most salient objects in the environment, but fails to reliably associate the bounding boxes to the correct action phrase. In contrast, TransFusion is able to correctly associate the drawer bounding box to "open drawer" and to predict the correct action associated with the phone. More examples are provided in the Supp. Mat. + +# 4.2. Ablation studies + +Evaluation of action context representation. In Table 2, we ablate different variants of context representation and compare against the representation that uses the groundtruth (GT) action labels provided by Ego4D in the action context, to assess the quality of our generated context summaries. The observed TTC Average Precision difference measures around $2\%$ and was thus omitted from the comparison. We report on the Noun and Noun-Verb metrics as we expect major differences in semantic-related tasks. + +Experiments show that we achieve results approaching GT, demonstrating the effectiveness of our context repre + +![](images/2760f6325fb033de5848b12ab2a0db695cc01daa0951789615e11bb639c27c08.jpg) +GT: open drawer 0.1s +(a) Ego4D: take container 0.78s, take container 0.46s, take pot 0.42s, take lid 0.46s + +![](images/f347836900e00440eff16f08c3b1a20da37d3039bebd62ce7fbfd2bcac128d1c.jpg) +GT: take phone 1.0s +(b) Ego4D: take table 0.78s, take container 0.85s, take computer 0.86s, take phone 1.11s + +![](images/f275f6477bf296d642ce09639b28fbbdb8f8c32b7a37ecd75be9cb706813e18f.jpg) +(c) TF (ours): open drawer 0.56s, open drawer 0.36s, open drawer 0.44s, take lid 0.6s + +![](images/5e1888e52b63b20e82773356a1746d7e8c773d7b1012884f11f87efe7f330a24.jpg) +(d) TF (ours): take phone 0.87s, open computer 0.63s, operate tablet 0.73s, take phone 0.98s +Figure 6. Qualitative results of the proposed approach. Results of Ego4D's FRCNN + SF baseline [22] are shown in the top row and TransFusion in the bottom row. Green denotes the ground truth bounding box, with the labels shown on top of each column. + +
Rep.∅ANhNsA+NhA+NsGT
N↑17.7120.0919.3319.7519.0320.1921.71
N-V↑6.147.166.867.057.127.557.86
+ +Table 2. Comparison of different action context representations. We evaluate different variations of context representations on the predicted noun (N) and noun-verb (N-V), and compare them against ground-truth annotations (GT) provided by v1 of the Ego4D validation set. GT only contains action descriptions of verb-noun pairs. Action context represented by $\mathcal{A} + \mathcal{N}_s$ achieves the best N-V performance. + +sentations. Using a vision-only version of our model $(\emptyset)$ yields markedly worse results than when adding any type of language conditioning. Action descriptions $\mathcal{A}$ consistently boost the verb classification performance. Adding salient objects $\mathcal{N}_s$ to $\mathcal{A}$ further improves task performance. These together offer a reasonable picture of using language summaries to represent the action context. In addition, we notice a domain gap in the choice of words between the generated labels and the ground-truth ones. In some cases, even if the description is satisfactory for a human evaluator, the information may be too vague to be directly useful to the model. Phrasing differences between the generated captions and the training corpus of the language encoders may further help to improve the performance. + +As an additional type of action context, we hypothesize that the sequence of objects held by the actor may be indicative of the overall task and can help the model better infer the immediate next steps. Motivated by this, we use an existing hand-object interaction predictor [11] jointly with an + +![](images/940a5f7209592e370d384e48631731a1fa40247041117243d0139f6e40d41cee.jpg) +Figure 7. Comparison to video features. TransFusion outperforms TransFusion-Video over different context lengths measured by Noun-Verb mAP. TransFusion using GT-based context summaries as language input represents an upper performance bound. + +object detector [65] to look for held objects and their semantic labels in each frame. Usually, such objects occur only in a subset of frames (e.g. due to limited visibility of the actor's hands), which results in a lower performance of the resulting $\mathcal{N}_h$ and $\mathcal{A} + \mathcal{N}_h$ representations. We hence use $\mathcal{A} + \mathcal{N}_s$ to represent the action context, considering the good performance achieved. Ablations on the hyperparameters $d$ , the cutoff distance, and $k$ , the number of salient objects per frame, show that the performance is less sensitive to their settings. More details can be found in the Supp. Mat. + +Comparison to video features. We evaluate whether language summaries can better represent action context than video features for the object-anticipation task. To do so, we implement TransFusion-Video, which takes video features extracted from the Ego4D SlowFast [16] model as input and uses an identical fusion module as in TransFusion. + +Figure 7 shows that TransFusion consistently outperforms TransFusion-Video. $L_{c} \times 30$ frames are taken in TransFusion-Video. For video features, the domain gap between pre-training and target datasets seems to be too large to allow effective generalization to a diverse dataset such as Ego4D. We suspect that the feature representations are more entangled due to the need to represent temporal, visual, and semantic aspects simultaneously. With language, temporal information is directly encoded in the syntax, i.e. the succession of words. Given a video clip of one second, language can summarize it in two words whereas corresponding video features use more than 12 times the space (e.g. smaller SBERT [41] features of $2 \times 384$ compared to SlowFast [16] features of $4 \times 2304$ ). Therefore, our language-based context representation has an advantage when it comes to longer video sequences. Overall, TransFusion obtains a $21\%$ boost in N-V performance. See the Supp. Mat. for details. + +Length of action context. A significant parameter in obtaining good performance is the length of action context, i.e., how many past actions to include in the input. An adequate context length is important to allow the model to pick up more specific activity patterns. Some activities that have a more structured nature or certain repeating patterns (e.g. + +
Language EncoderToken sizeN↑N-V↑
SBERT/BERT [13, 41]38420.197.55
SBERT/RoBERTa [33, 41]76819.787.78
GPT-2 [37]76818.987.08
Flan-T5 [8]102418.266.61
+ +Table 3. Language encoder ablations. We compare SBERT using BERT and RoBERTa backbones, with GPT-2 and Flan-T5. + +cooking or repairing a bicycle) are easier to model with an increased context length compared to more random activities such as playing basketball. Figure 7 shows that the overall performance plateaus at $L_{c} = 2$ and higher. + +Language encoder. A language encoder is used to process the context summaries in TransFusion. To understand the impact of different language encoders, we experiment with multiple LLMs including SBERT with a RoBERTa backbone [33, 41], GPT-2 [37], and FLAN-T5 [8], using version 1 of the validation set. Table 3 summarizes the performance of different language encoders, showing that the SBERT models perform best. We consider TTC less relevant here. We hypothesize that the difference comes from the training objectives of language models. SBERT is trained to map semantically similar sentences closer to each other in the embedding space. This is particularly useful when the model needs to understand descriptions with semantically close words, e.g. "cut carrots" and "slice carrots". GPT-2 and FLAN-T5 are optimized using generative objectives, which makes them overly sensitive to the input word choice. + +Fusion module. We provide ablations regarding the fusion module in the Supp. Mat. + +# 5. Conclusion and discussion + +We propose TransFusion, a multimodal fusion model that anticipates object interactions by considering the past action context represented by language summaries. Our successful experiments on a challenging egocentric video dataset demonstrate how language summaries and the vast knowledge contained in vision-language foundation models can improve object interaction anticipation, highlighting the representational power of language. It appears promising to investigate fusing even more modalities, such as motion cues from optical flow. We emphasize that the proposed summarization approach is not limited to this task. Language provides a universal interface to complement the visual input with information encoded in foundation models or task-specific pipelines. Future work can leverage language summarization for other video reasoning tasks [60, 63], or extend it temporally to describe possible future activities [1, 27] for long-term interaction anticipation. + +Acknowledgements. This work was supported by an ETH Zurich Postdoctoral Fellowship. We thank Adrian Spurr for the helpful comments and Velko Vechev for the voiceover. + +# References + +[1] Chaitanya Ahuja and Louis-Philippe Morency. Language2Pose: Natural Language Grounded Pose Forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719–728, 2019. 8 +[2] Alan Akbik, Duncan Blythe, and Roland Vollgraf. Contextual String Embeddings for Sequence Labeling. In Proceedings of the 27th international conference on computational linguistics, pages 1638-1649, 2018. 4 +[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. VQA: Visual Question Answering. In Proceedings of the IEEE International Conference on Computer Vision, pages 2425-2433, 2015. 3 +[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. CoRR, abs/2104.00650, 2021. 3 +[5] Gedas Bertasius, Hyun Soo Park, Stella X. Yu, and Jianbo Shi. First-Person Action-Object Detection with EgoNet. In Proceedings of Robotics: Science and Systems, 2017. 2 +[6] Jerome S Bruner. Intention in the Structure of Action and Interaction. Advances in Infancy Research, 1981. 1 +[7] Guo Chen et al. InternVideo-Ego4D: A Pack of Champion Solutions to Ego4D Challenges. arXiv preprint arXiv:2211.09529, 2022.2,6,7 +[8] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling Instruction-Finetuned Language Models, 2022. 3, 8 +[9] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling Egocentric Vision. CoRR, abs/2006.13256, 2020. 2, 3 +[10] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling Egocentric Vision: Collection, Pipeline and Challenges for EPIC-Kitchens-100. International Journal of Computer Vision, 130(1):33-55, 2022. 4 +[11] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. EPIC-KITCHENS VISOR Benchmark: VVideo Segmentations and Object Relations. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2022. 7 +[12] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. TransVG: End-to-End Visual Grounding with Transformers. CoRR, abs/2104.08541, 2021.3 +[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional + +Transformers for Language Understanding. CoRR, abs/1810.04805, 2018.3,8 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. CoRR, abs/2010.11929, 2020. 2, 3, 6 +[15] Christoph Feichtenhofer. X3D: Expanding Architectures for Efficient Video Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 203-213, 2020. 6 +[16] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast Networks for Video Recognition. CoRR, abs/1812.03982, 2018. 6, 8 +[17] Antonino Furnari and Giovanni Maria Farinella. Rolling-Unrolling LSTMs for Action Anticipation from First-Person Video. CoRR, abs/2005.02190, 2020. 3 +[18] Antonino Furnari, Sebastiano Battiato, Kristen Grauman, and Giovanni Maria Farinella. Next-Active-Object Prediction from Egocentric Videos. CoRR, abs/1904.05250, 2019. 1, 2 +[19] H. Girase, N. Agarwal, C. Choi, and K. Mangalam. Latency Matters: Real-Time Action Forecasting Transformer. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18759-18769, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 3 +[20] Rohit Girdhar and Kristen Grauman. Anticipative Video Transformer. In ICCV, 2021. 3 +[21] Ross B. Girshick. Fast R-CNN. CoRR, abs/1504.08083, 2015.6 +[22] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Wesley Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4D: Around the World in 3,000 Hours of Egocentric Video. CoRR, abs/2110.07058, 2021. 1, 2, 4, 5, 6, 7 + +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. CoRR, abs/1512.03385, 2015. 5, 6 +[24] Zhicheng Huang, Zhaoyang Zeng, Bei Liu, Dongmei Fu, and Jianlong Fu. Pixel-BERT: Aligning Image Pixels with Text by Deep Multi-Modal Transformers. CoRR, abs/2004.00849, 2020.3 +[25] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. EPIC-Fusion: Audio-Visual Temporal Binding for Egocentric Action Recognition. CoRR, abs/1908.08498, 2019. 3 +[26] Weicheng Kuo, Fred Bertsch, Wei Li, AJ Piergiovanni, Mohammad Saffar, and Anelia Angelova. FindIt: Generalized Localization with Natural Language Queries. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI, pages 502-520. Springer, 2022. 3 +[27] Yen-Ling Kuo, Xin Huang, Andrei Barbu, Stephen G McGill, Boris Katz, John J Leonard, and Guy Rosman. Trajectory Prediction with Linguistic Representations. In 2022 International Conference on Robotics and Automation (ICRA), pages 2868-2875. IEEE, 2022. 8 +[28] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. Less is More: ClipBERT for Video-and-Language Learning via Sparse Sampling. CoRR, abs/2102.06183, 2021. 3 +[29] Tsung-Yi Lin, Piotr Dólar, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature Pyramid Networks for Object Detection. CoRR, abs/1612.03144, 2016. 5 +[30] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen Clip Models Are Efficient Video Learners. In European Conference on Computer Vision, pages 388-404. Springer, 2022. 4 +[31] Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the Variance of the Adaptive Learning Rate and Beyond. CoRR, abs/1908.03265, 2019. 6 +[32] Miao Liu, Siyu Tang, Yin Li, and James M. Rehg. Forecasting Human Object Interaction: Joint Prediction of Motor Attention and Egocentric Activity. CoRR, abs/1911.10967, 2019. 2 +[33] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A Robustly Optimized BERT Pretraining Approach. CoRR, abs/1907.11692, 2019.8 +[34] TorchVision maintainers and contributors. TorchVision: PyTorch's Computer Vision Library. https://github.com/pytorch/vision, 2016.6 +[35] David W Orr. The Nature of Design: Ecology, Culture, and Human Intention. Oxford University Press, 2002. 1 +[36] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou, Rama Chellappa, and Pengchuan Zhang. EGOVLPV2: Egocentric Video-Language Pre-training With Fusion in the Backbone, 2023. 3 + +[37] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language Models are Unsupervised Multitask Learners. OpenAI Blog, 2019. 3, 8 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. CoRR, abs/2103.00020, 2021. 2, 3, 4 +[39] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. CoRR, abs/1910.10683, 2019. 3 +[40] Francesco Ragusa et al. StillFast: An End-to-End Approach for Short-Term Object Interaction Anticipation. arXiv preprint arXiv:2304.03959, 2023. 2, 6 +[41] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. CoRR, abs/1908.10084, 2019. 3, 5, 6, 8 +[42] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. CoRR, abs/1506.01497, 2015. 5 +[43] Paschal Sheeran. Intention-Behavior Relations: A Conceptual and Empirical Review. European Review of Social Psychology, 12(1):1-36, 2002. 1 +[44] Chen Sun, Fabien Baradel, Kevin Murphy, and Cordelia Schmid. Contrastive Bidirectional Transformer for Temporal Representation Learning. CoRR, abs/1906.05743, 2019. 3 +[45] Hao Tan and Mohit Bansal. LXMERT: Learning Cross-Modality Encoder Representations from Transformers. CoRR, abs/1908.07490, 2019. 3 +[46] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Enhancing Next Active Object-Based Egocentric Action Anticipation With Guided Attention. In 2023 IEEE International Conference on Image Processing (ICIP), pages 1450-1454. IEEE, 2023. 2, 6 +[47] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Guided Attention for Next Active Object @ EGO4D STA Challenge. arXiv preprint arXiv:2305.16066, 2023. 2 +[48] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Leveraging Next-Active Objects for Context-Aware Anticipation in Egocentric Videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 8657–8666, 2024. 2 +[49] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked Autoencoders Are Data-Efficient Learners for Self-Supervised Video Pre-Training. Advances in Neural Information Processing Systems, 35:10078–10093, 2022. 2, 6 +[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. CoRR, abs/1706.03762, 2017. 3 + +[51] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-Anticipation Transformer for Online Action Understanding, 2023. 3 +[52] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Action-CLIP: A New Paradigm for Video Action Recognition. arXiv preprint arXiv:2109.08472, 2021. 4 +[53] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework. CoRR, abs/2202.03052, 2022. 2, 3, 4 +[54] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. MeMViT: Memory-Augmented Multiscale Vision Transformer for Efficient Long-Term Video Recognition. CoRR, abs/2201.08383, 2022. 3 +[55] Yu Wu, Linchao Zhu, Xiaohan Wang, Yi Yang, and Fei Wu. Learning to Anticipate Egocentric Actions by Imagination. CoRR, abs/2101.04924, 2021. 3 +[56] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, Attend and Tell: Neural Image Caption Generation With Visual Attention. In International Conference on Machine Learning, pages 2048-2057. PMLR, 2015. 3 +[57] Ting Yao, Yingwei Pan, Yehao Li, Zhaofan Qiu, and Tao Mei. Boosting Image Captioning With Attributes. In Proceedings of the IEEE International Conference on Computer Vision, pages 4894-4902, 2017. 3 +[58] Moon Ye-Bin, Jisoo Kim, Hongyeob Kim, Kilho Son, and Tae-Hyun Oh. TextMania: Enriching Visual Feature by Text-driven Manifold Augmentation, 2023. 3 +[59] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. MERLOT: Multimodal Neural Script Knowledge Models. CoRR, abs/2106.02636, 2021. 3 +[60] Kuo-Hao Zeng, Tseng-Hung Chen, Ching-Yao Chuang, Yuan-Hong Liao, Juan Carlos Niebles, and Min Sun. Leveraging Video Descriptions to Learn Video Question Answering. In Thirty-First AAAI Conference on Artificial Intelligence, 2017. 8 +[61] Ce Zhang, Changcheng Fu, Shijie Wang, Nakul Agarwal, Kwonjoon Lee, Chiho Choi, and Chen Sun. Object-Centric Video Representation for Long-term Action Anticipation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6751-6761, 2024. 3 +[62] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection. In The Eleventh International Conference on Learning Representations, 2022. 2, 6 +[63] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. Video Summarization With Long Short-Term Memory. In European Conference on Computer Vision, pages 766-782. Springer, 2016. 8 +[64] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning Video Representations From Large Language Models, 2022. 3 + +[65] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple Multi-Dataset Detection. CoRR, abs/2102.13086, 2021. 8 \ No newline at end of file diff --git a/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/images.zip b/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e13053c0aaeab0295c85ea684032c976962e152f --- /dev/null +++ b/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9e199a90a70feea0412f24b19383cc114fd920f9bc7895a365e299cbbd53c70 +size 362875 diff --git a/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/layout.json b/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..583dc171015b34c74d68d92c8f8a99f9a256c90c --- /dev/null +++ b/2024/Summarize the Past to Predict the Future_ Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation/layout.json @@ -0,0 +1,8915 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 536, + 140 + ], + "type": "text", + "content": "Summarize the Past to Predict the Future: Natural Language Descriptions of Context Boost Multimodal Object Interaction Anticipation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "spans": [ + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "text", + "content": "Razvan-George Pasca" + }, + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "text", + "content": " Alexey Gavyushin" + }, + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "text", + "content": " Muhammad Hamza" + }, + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "text", + "content": " Yen-Ling Kuo" + }, + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "text", + "content": " Kaichun Mo" + }, + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "text", + "content": " Luc Van Gool" + }, + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "inline_equation", + "content": "^{1,5,6}" + }, + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "text", + "content": " Otmar Hilliges" + }, + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "text", + "content": " Xi Wang" + }, + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 160, + 509, + 203 + ], + "type": "text", + "content": " ETH Zurich Univ. of Zurich Univ. of Virginia NVIDIA KU Leuven INSAIT, Sofia" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 90, + 212, + 509, + 396 + ], + "blocks": [ + { + "bbox": [ + 90, + 212, + 509, + 396 + ], + "lines": [ + { + "bbox": [ + 90, + 212, + 509, + 396 + ], + "spans": [ + { + "bbox": [ + 90, + 212, + 509, + 396 + ], + "type": "image", + "image_path": "0f23b4421a71e33743c550be111aa1ad4300ae983910743e70882576c09612d4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 404, + 548, + 460 + ], + "lines": [ + { + "bbox": [ + 45, + 404, + 548, + 460 + ], + "spans": [ + { + "bbox": [ + 45, + 404, + 548, + 460 + ], + "type": "text", + "content": "Figure 1. TransFusion: Multimodal fusion transformer for short-term object interaction anticipation. Given a video sequence of past observations, the task aims to predict a set of objects visible in the current frame that will be interacted with in the future, i.e. in the activation frame that is " + }, + { + "bbox": [ + 45, + 404, + 548, + 460 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 45, + 404, + 548, + 460 + ], + "type": "text", + "content": " frames away from the current prediction frame. Additionally, the task requires estimating the bounding box, the associated action described by a verb-noun pair, and the time to contact for each predicted object. We propose TransFusion, a multimodal fusion architecture that uses language summaries of past actions and salient objects to effectively predict future object interactions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 479, + 192, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 479, + 192, + 491 + ], + "spans": [ + { + "bbox": [ + 143, + 479, + 192, + 491 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 503, + 289, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 289, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 289, + 696 + ], + "type": "text", + "content": "We study object interaction anticipation in egocentric videos. This task requires an understanding of the spatiotemporal context formed by past actions on objects, coined action context. We propose TransFusion, a multimodal transformer-based architecture for short-term object interaction anticipation. Our method exploits the representational power of language by summarizing the action context textually, after leveraging pre-trained vision-language foundation models to extract the action context from past video frames. The summarized action context and the last observed video frame are processed by the multimodal fusion module to forecast the next object interaction. Experiments on the Ego4D next active object interaction dataset show the effectiveness of our multimodal fusion model and highlight the benefits of using the power of foundation models and language-based context summaries in a task where" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 479, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 547, + 529 + ], + "type": "text", + "content": "vision may appear to suffice. Our novel approach outperforms all state-of-the-art methods on both versions of the Ego4D dataset. A project video and code are available at https://eth-ait.github.io/transfusion-proj/." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 548, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 388, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 569, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 547, + 677 + ], + "type": "text", + "content": "The ability to predict plausible future human-object interactions is important for many assistive technologies. The task of short-term object interaction anticipation [18, 22] is defined as predicting which object a user will interact with next, what action will be performed and when in the future, given an egocentric video input. Providing an effective solution would help artificial agents, such as care robots, assist humans in their daily activities, as well as enable safety applications such as child and elderly monitoring." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "This task is challenging because it requires reasoning about user intent [6, 35, 43], which is not directly observable. For example, if the video shows the user holding" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 156, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 156, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 156, + 714 + ], + "type": "text", + "content": "*Authors contributed equally." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18286" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "a vegetable, the most likely next action is to slice it, but only with a knife and cutting board present. Without those, there is significantly more uncertainty. Therefore, understanding what has happened in the past and what relevant objects are present is critical for predicting the next object interaction. We collectively refer to these concepts as action context. Existing interaction forecasting methods do not explicitly model action context and rely on neural networks to extract that information from fixed-sized video chunks [5, 18, 22, 32]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 289, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 289, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 289, + 467 + ], + "type": "text", + "content": "In this paper, we propose to use language summarization of the past as an explicit representation of action context. This filters out ambiguities caused by visual clutter. It also puts focus on objects and actions directly relevant to the ongoing and likely future activities. For the object interaction anticipation, we next introduce a multimodal fusion approach. Specifically, we propose a transformer-based multimodal fusion model, TransFusion. As shown in Figure 1, our model takes language summaries of the action context and the last observed single frame of a video as input and predicts future object interactions. We employ an existing image captioning system [53] to generate a concise description of past actions given by pairs of verbs and nouns. We use CLIP [38], a vision-language model trained on pairs of images and captions, to detect salient objects in the prediction frame of interest, and deem these objects task-relevant. The verb-noun pairs together with the list of salient objects form the summary of the past action context. Leveraging the vast knowledge contained in pre-trained vision-language foundation models yields more generic representations of the past action context and enables us to better generalize towards various scenarios, in particular rare ones not well-represented in the dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 469, + 288, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 288, + 564 + ], + "type": "text", + "content": "Our experiments show that our language-based summaries can effectively represent the past action context. We evaluate on both versions of the challenging egocentric dataset Ego4D [22]. Our approach improves interaction prediction accuracy and outperforms state-of-the-art models pre-trained on large video datasets [7, 40]. Such improvements show that large vision-language models can be useful even for a task that seems purely visual." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 566, + 188, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 566, + 188, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 188, + 578 + ], + "type": "text", + "content": "In summary, our contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 580, + 286, + 710 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 580, + 286, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 580, + 286, + 626 + ], + "spans": [ + { + "bbox": [ + 58, + 580, + 286, + 626 + ], + "type": "text", + "content": "- A strategy for generating concise language-based summaries of action context over longer timespans from frame-wise language-based summaries, for instance obtained using vision-language foundation models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 628, + 286, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 628, + 286, + 661 + ], + "spans": [ + { + "bbox": [ + 58, + 628, + 286, + 661 + ], + "type": "text", + "content": "- TransFusion, a multimodal fusion model for short-term object interaction anticipation combining action context descriptions and visual features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 663, + 286, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 663, + 286, + 710 + ], + "spans": [ + { + "bbox": [ + 58, + 663, + 286, + 710 + ], + "type": "text", + "content": "- Experiments on both versions of the Ego4D dataset showing improvements over state-of-the-art methods on the short-term object interaction anticipation task, in particular for long-tailed classes." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 390, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 390, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 390, + 83 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 94, + 545, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 94, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 304, + 94, + 545, + 261 + ], + "type": "text", + "content": "Object interaction anticipation. Several works have been proposed for human-object interaction prediction. Furnari et al. [18] first introduced next-active-object (NAO) prediction, which predicts the active/not active label for each object using its motion trajectory features, but doesn't consider action anticipation. Closely related to NAO, Bertasius et al. [5] introduced action-objects, which considers objects that capture the human actor's visual attention or tactile interaction. Bertasius et al. use the data from a stereo camera system in a two-stream network, integrating RGB and depth information. They show that certain aspects of human actions can be predicted by exploiting the spatial configuration of the objects and the actor's head. Their approach is not easily scalable due to the need for stereo data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 264, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 361 + ], + "type": "text", + "content": "In addition to different tasks, Liu et al. [32] propose a model that improves NAO prediction by incorporating future hand trajectory modeling. They obtain impressive results using 3D CNNs, confirming that hand motion cues and scene dynamics have important explanatory power in predicting the short-term future. Nonetheless, their approach also encounters scaling limitations because additional ground-truth labels for hand trajectories are needed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 362, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 545, + 613 + ], + "type": "text", + "content": "The Ego4D dataset [22] proposed object interaction anticipation. Apart from locating the NAO, the task in Ego4D also requires the prediction of associated nouns and verbs, as well as the time to contact (TTC) for the future interacted objects. The official Ego4D baseline uses the current frame and past video frames as input and employs a two-stage approach to first learn noun and box prediction, and afterwards learn verb and TTC prediction. InternVideo [7] similarly leverages a DINO [62] model to detect objects in the first stage, and a VideoMAE [49]-pretrained ViT [14] spacetime attention model to predict verbs and TTC conditioned on the object predictions in the second stage. StillFast [40] presents an end-to-end method for NAO prediction, fusing multi-scale high-resolution image and low-resolution video features and combining local RoI features with global scene features to boost the interaction anticipation performance. GANOv2 [46, 47] and its predecessor [48] extract features from objects detected in past observed frames and fuse them with global frame features, forwarding them to a StillFast predictor and a motion-object-separating encoder-decoder architecture, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Action anticipation. Besides localizing future object interactions, predicting the next action steps is an equally important task that intelligent systems need to perform to interact with humans. A longer prediction timespan requires a better grasp of the structure behind the succession of actions. A pioneering advancement in this field is the introduction of the EPIC-Kitchens 100 (EK100) [9] dataset, which facilitates research in action anticipation within egocentric" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18287" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": "videos. Significant research efforts have been dedicated to exploring the action anticipation task utilizing EK100. Anticipative Video Transformer (AVT) [20] is one of the best-performing architectures on the EK100 [9] action anticipation benchmark. AVT uses a pre-trained Vision Transformer [14] as the image feature backbone and attends to the previously observed video frames to anticipate future actions. MeMViT [54] improves AVT by exploiting a longer temporal context, highlighting the importance of modeling a long sequence of previous actions. Other model architectures such as recurrent neural networks [17, 55], multimodal temporal CNNs [25] and transformers [19, 51] are also used to model the temporal contexts. Note that these methods are designed for action classification using video data and adapting them for object detection is not trivial. It is worth noting that the action anticipation task from the EK100 benchmark is a subtask of our short-term object interaction anticipation task. Our task not only predicts the action label of verb-noun pairs, but we also predict the location of the object and the time to contact, which is more challenging than the action anticipation task itself. TransFusion focuses on detecting the next active objects, as all evaluation metrics are conditioned on the box IoU. Therefore, comparing our work with the state-of-the-art methods for the action anticipation task on EK100 is out of the scope of this work. The Ego4D dataset itself further introduces a long-term action anticipation benchmark, which [61] tackles using pre-trained visual grounding models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 413, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 413, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 413, + 287, + 569 + ], + "type": "text", + "content": "Vision-language models. Our approach combines the visual and language modalities to improve object interaction anticipation. Most recent multimodal vision-language architectures are based on Transformer [50] fusion schemes applied on features extracted by various encoders [4, 12, 26]. We follow their spirit by employing self-attention-based modality fusion. Our work differs in one significant aspect: the language features represent the past action context. This serves as an additional signal to disambiguate the actor's intent given similar visual scenes. The model must learn to predict from the multimodal fusion of language context summaries and visual cues in a considerably smaller data regime." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 289, + 714 + ], + "type": "text", + "content": "More broadly, there are numerous transformer-based works aimed at learning cross-modal representations between vision and language [24, 28, 36, 38, 44, 45, 58, 59, 64] using terabyte-scale datasets crawled from multiple internet sources. When combined with large language models such as GPT-2 [37], BERT [13], Sentence-BERT [41], and the T5 variants [8, 39], these vision-language models can be applied to several downstream tasks such as VQA [3] and image captioning [56, 57]. Similar to other vision-language models that are trained with language input, our approach leverages state-of-the-art image captioning models [53] to provide a consistent description of past actions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 71, + 465, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 465, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 465, + 85 + ], + "type": "text", + "content": "3. Language-guided prediction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 91, + 547, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 297 + ], + "type": "text", + "content": "We use language-based context summaries to represent the spatial-temporal context formed by past actions and object relationships. The context summary offers an explicit yet compact representation, which allows us to focus on task-relevant content and filter out unimportant information from the likely cluttered environment. To this end, we propose TransFusion, a transformer-based model that leverages a joint-attention fusion mechanism to effectively learn to predict object interactions in the future from the past action context. In the following, we describe our framework, which takes video clips as input and anticipates the next object interaction. We begin by defining the object interaction anticipation task (Sec. 3.1). We then show how to generate a summary for a prediction frame from an egocentric video to represent the action context provided by the past frames (Sec. 3.2). Finally, we present our multimodal fusion model's architecture (Sec. 3.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 305, + 398, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 305, + 398, + 317 + ], + "spans": [ + { + "bbox": [ + 305, + 305, + 398, + 317 + ], + "type": "text", + "content": "3.1. Task definition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 324, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 545, + 373 + ], + "type": "text", + "content": "Given a video " + }, + { + "bbox": [ + 304, + 324, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{f_1,\\dots ,f_T\\}" + }, + { + "bbox": [ + 304, + 324, + 545, + 373 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 304, + 324, + 545, + 373 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 324, + 545, + 373 + ], + "type": "text", + "content": " as input, our goal is to predict the next object interaction " + }, + { + "bbox": [ + 304, + 324, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 304, + 324, + 545, + 373 + ], + "type": "text", + "content": " in the last observed video frame " + }, + { + "bbox": [ + 304, + 324, + 545, + 373 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 304, + 324, + 545, + 373 + ], + "type": "text", + "content": ", possibly using older frames to extract context information. We call " + }, + { + "bbox": [ + 304, + 324, + 545, + 373 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 304, + 324, + 545, + 373 + ], + "type": "text", + "content": " the prediction frame." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": "As illustrated in Figure 1, the task is to anticipate the object interaction in " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_T = \\{o_1^T,\\dots ,o_N^T\\}" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": " denotes the set of future object interactions, with " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "N > 1" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": " whenever multiple objects are interacted with simultaneously. In other words, multiple " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "o_i^T" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": " can be detected in the prediction frame " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": ". For example, " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "N = 2" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": " when one hand is holding two mugs in a kitchen scene, or the left hand is touching a plant while the right hand is moving a pot in a garden scene. Each object interaction " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "o_i^T = (b_i^T,n_i^T,v_i^T,t_i^T)" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": " consists of the object bounding box " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "b_{i}^{T}" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": ", the semantic object label (noun) " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "n_i^T" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": ", the action label (verb) " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "v_{i}^{T}" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": ", and the time to contact (TTC) " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "t_i^T" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": ". The task is to predict the next active object location " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "b_{i}^{T}" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": " in the prediction frame " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": ", the associated action described by the verb-noun pair " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "(n_i^T,v_i^T)" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": ", and the time " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "t_i^T" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": " to the point of contact when the interaction is going to start. The task is set up such that the actual object interaction takes place after a buffer time " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": " on frame " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "f_{T + \\Delta}" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": " and the predicted time to contact " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "t_i^T = \\Delta" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": ". Ego4D uses " + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "inline_equation", + "content": "\\Delta \\geq 0.033s" + }, + { + "bbox": [ + 304, + 374, + 547, + 601 + ], + "type": "text", + "content": ". This requires anticipating future actions and increases the task difficulty." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 610, + 432, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 432, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 432, + 624 + ], + "type": "text", + "content": "3.2. Summarizing the past" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Anticipating object interactions in the future requires an understanding of what happened in the past. We first infer the per-frame action context from each egocentric video. This process consists of extracting action descriptions and held objects. These information sources are generated and aggregated independently across frames and finally forwarded as the action context to our anticipation model." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18288" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 547, + 202 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 547, + 202 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 547, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 547, + 202 + ], + "type": "image", + "image_path": "dd1f4c7a5e8616dfd8819f244e47559a7fb781f4fd0cbae7c8aa31790cffbb6f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 547, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 547, + 266 + ], + "type": "text", + "content": "Figure 2. Overview of the TransFusion model. TransFusion takes the prediction frame and the action context summary as input and predicts the bounding box of the next-active object, the noun-verb pair describing the associated action, and the time to contact (TTC). Feature maps of different scales are extracted from the visual encoder and then fused via a multimodal fusion module with the encoded language features. Their output is then processed by a regular feature pyramid network (FPN), denoted as multi-scale region proposal networks, before being fed into the Faster R-CNN detector." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 272, + 289, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 272, + 289, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 289, + 357 + ], + "type": "text", + "content": "Extracting frame-wise action descriptions. Inspired by the label format used in egocentric video datasets [10, 22], we aim to describe the past actions by sequences of verb-noun pairs (e.g. \"wash tomato\", \"cut tomato\"), so-called action descriptions. For each video frame " + }, + { + "bbox": [ + 46, + 272, + 289, + 357 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 272, + 289, + 357 + ], + "type": "text", + "content": ", we extract the action currently performed by the agent, represented by a single pair " + }, + { + "bbox": [ + 46, + 272, + 289, + 357 + ], + "type": "inline_equation", + "content": "(v, n)" + }, + { + "bbox": [ + 46, + 272, + 289, + 357 + ], + "type": "text", + "content": " of a verb " + }, + { + "bbox": [ + 46, + 272, + 289, + 357 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 272, + 289, + 357 + ], + "type": "text", + "content": " and a noun " + }, + { + "bbox": [ + 46, + 272, + 289, + 357 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 272, + 289, + 357 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 364, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 289, + 555 + ], + "type": "text", + "content": "We use an off-the-shelf image captioning model [53] to generate multiple full-sentence captions for each frame of interest, so as to find repeatedly occurring outputs and reduce noise: we perform part-of-speech tagging [2] on these captions and collect candidate pairs consisting of verbs followed by nouns within some cutoff distance " + }, + { + "bbox": [ + 46, + 364, + 289, + 555 + ], + "type": "inline_equation", + "content": "d \\in \\mathbb{N}_0^+" + }, + { + "bbox": [ + 46, + 364, + 289, + 555 + ], + "type": "text", + "content": ". Setting " + }, + { + "bbox": [ + 46, + 364, + 289, + 555 + ], + "type": "inline_equation", + "content": "d > 0" + }, + { + "bbox": [ + 46, + 364, + 289, + 555 + ], + "type": "text", + "content": " allows us to account for additional words appearing between a verb-noun pair (e.g. \"eat apple\" from \"A person eating a red apple\") and thereby obtain more valid candidate pairs from the captions. However, larger cutoff distances may introduce more spurious detections (e.g. \"eat gathering\", from \"A person eating while at a gathering\"). We use " + }, + { + "bbox": [ + 46, + 364, + 289, + 555 + ], + "type": "inline_equation", + "content": "d = 4" + }, + { + "bbox": [ + 46, + 364, + 289, + 555 + ], + "type": "text", + "content": " in our experiments. The most frequent candidate pair across the frame's captions is selected as the verb-noun action description of this frame. If no valid verb-noun pair is found, the frame will have no action description." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "content": "Extracting frame-wise salient objects. Another characteristic shared by many daily action sequences is that the next active object is likely to have already appeared in the actor's view before the upcoming interaction. The current environment in the recording may also provide model conditioning. Therefore, we select a set of salient objects in each frame as (1) the potential candidates for future active objects and (2) a proxy of the surroundings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": "Encouraged by its previous successful use in action-related tasks [30, 52], we use CLIP [38], a pre-trained open-vocabulary zero-shot classifier, to compute the cosine similarities between the CLIP image embedding of a frame and the embeddings of the object categories from the Ego4D do" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 272, + 547, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 547, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 547, + 332 + ], + "type": "text", + "content": "main. We choose to keep the highest-scoring " + }, + { + "bbox": [ + 304, + 272, + 547, + 332 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 272, + 547, + 332 + ], + "type": "text", + "content": " objects as the set of salient objects for a given frame. A small " + }, + { + "bbox": [ + 304, + 272, + 547, + 332 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 272, + 547, + 332 + ], + "type": "text", + "content": " may fail to provide sufficient information, while a large " + }, + { + "bbox": [ + 304, + 272, + 547, + 332 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 272, + 547, + 332 + ], + "type": "text", + "content": " makes capturing the true salient objects more likely, yet introduces more noise due to false detections. We set " + }, + { + "bbox": [ + 304, + 272, + 547, + 332 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 304, + 272, + 547, + 332 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 342, + 548, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 548, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 548, + 545 + ], + "type": "text", + "content": "Aggregating summaries over multiple frames. After the frame-wise summarization step, we have, for each frame " + }, + { + "bbox": [ + 304, + 342, + 548, + 545 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 342, + 548, + 545 + ], + "type": "text", + "content": ", at most one verb-noun pair describing the frame-wise action " + }, + { + "bbox": [ + 304, + 342, + 548, + 545 + ], + "type": "inline_equation", + "content": "a_{t} = (n,v) \\in \\mathcal{A}_{t}^{F}" + }, + { + "bbox": [ + 304, + 342, + 548, + 545 + ], + "type": "text", + "content": ". We then apply a cross-frame aggregation scheme to all " + }, + { + "bbox": [ + 304, + 342, + 548, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{t}^{F}" + }, + { + "bbox": [ + 304, + 342, + 548, + 545 + ], + "type": "text", + "content": " so as to identify contiguous segments of frames exhibiting the same activity, while accounting for individual noisy frames possibly not matching the framework summaries of their neighbors. The aggregation processes the frames in temporal order. We consider frames belonging to the same atomic action as an action segment, which is initiated by observing a number of reoccurrences of some verb-noun pair and terminated by a lack of its occurrence within a number of contiguous frames. After the aggregation, we obtain " + }, + { + "bbox": [ + 304, + 342, + 548, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 342, + 548, + 545 + ], + "type": "text", + "content": ", the sequence of all action segments in the video. For each prediction frame, we construct its action context from a number of action segments preceding the frame, and the objects from " + }, + { + "bbox": [ + 304, + 342, + 548, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_s" + }, + { + "bbox": [ + 304, + 342, + 548, + 545 + ], + "type": "text", + "content": " in that frame." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 545, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 641 + ], + "type": "text", + "content": "By aggregating across frames, we thus capture dominant actions in the past, reduce noise in the generated action context and represent long timespans of similar activity by short textual descriptions. The complete action context is provided to our model by concatenating the textual representations of the aggregation results of any desired subset of the two context representations " + }, + { + "bbox": [ + 304, + 545, + 547, + 641 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{A},\\mathcal{N}_s\\}" + }, + { + "bbox": [ + 304, + 545, + 547, + 641 + ], + "type": "text", + "content": ", where we drop the index " + }, + { + "bbox": [ + 304, + 545, + 547, + 641 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 545, + 547, + 641 + ], + "type": "text", + "content": " for convenience. For more details, see the Supp. Mat." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 647, + 484, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 484, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 484, + 659 + ], + "type": "text", + "content": "3.3. TransFusion: Multimodal fusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Figure 2 visualizes the architecture of the proposed multimodal TransFusion model. TransFusion employs two different base encoders, for language and image input respectively. The input prediction frame " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " is processed by a regu" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18289" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 69, + 265, + 276 + ], + "blocks": [ + { + "bbox": [ + 72, + 69, + 265, + 276 + ], + "lines": [ + { + "bbox": [ + 72, + 69, + 265, + 276 + ], + "spans": [ + { + "bbox": [ + 72, + 69, + 265, + 276 + ], + "type": "image", + "image_path": "ec1da6b14eb441a83d93e834c2f4a065392b12cc24c9c1f485faa788592b3b40.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 284, + 289, + 351 + ], + "lines": [ + { + "bbox": [ + 46, + 284, + 289, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 289, + 351 + ], + "type": "text", + "content": "Figure 3. Multimodal fusion module. We first project the CNN feature map and the language tokens to a common dimensionality before adding the specific embeddings. We concatenate the visual and language tokens and feed them to " + }, + { + "bbox": [ + 46, + 284, + 289, + 351 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 284, + 289, + 351 + ], + "type": "text", + "content": " self-attention layers. At the output, the fused visual tokens are projected back into the initial feature map shape." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 361, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 468 + ], + "type": "text", + "content": "lar CNN-based visual encoder [23], producing a set of multiscale feature maps " + }, + { + "bbox": [ + 46, + 361, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{V}}^{s}(f_{T})" + }, + { + "bbox": [ + 46, + 361, + 287, + 468 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 361, + 287, + 468 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 361, + 287, + 468 + ], + "type": "text", + "content": " indicating the scale. The frame's action context is encoded by a language encoder [41] and yields " + }, + { + "bbox": [ + 46, + 361, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{L}}(\\mathcal{C}_L)" + }, + { + "bbox": [ + 46, + 361, + 287, + 468 + ], + "type": "text", + "content": ". The feature maps belonging to different scale levels from the visual encoder are fused with the encoded language features via a joint-attention scheme. The result is then processed by a regular feature pyramid network (FPN) [29] before being forwarded to the Faster R-CNN [42] detector." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 471, + 289, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 289, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 289, + 664 + ], + "type": "text", + "content": "Multimodal fusion. For simplicity, we describe the multimodal fusion module for a single input scale " + }, + { + "bbox": [ + 46, + 471, + 289, + 664 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 471, + 289, + 664 + ], + "type": "text", + "content": ", which we hence omit from notation. See Figure 3 for a detailed view of a single Transformer encoder layer together with the input projection stages. For the visual modality, we first split the image into a sequence of tokens. Specifically, the visual features " + }, + { + "bbox": [ + 46, + 471, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{V}}(f_T) \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 46, + 471, + 289, + 664 + ], + "type": "text", + "content": " are reshaped to " + }, + { + "bbox": [ + 46, + 471, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{N \\times (P^2 \\cdot C)}" + }, + { + "bbox": [ + 46, + 471, + 289, + 664 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 471, + 289, + 664 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 471, + 289, + 664 + ], + "type": "text", + "content": " is the token size and " + }, + { + "bbox": [ + 46, + 471, + 289, + 664 + ], + "type": "inline_equation", + "content": "N = \\frac{H \\times W}{P^2}" + }, + { + "bbox": [ + 46, + 471, + 289, + 664 + ], + "type": "text", + "content": " the number of tokens. We add unique type embeddings as well as a positional encoding to each modality. Both visual and language tokens are first mapped to " + }, + { + "bbox": [ + 46, + 471, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 471, + 289, + 664 + ], + "type": "text", + "content": " via a linear projection before being fed to the fusion module in a concatenated form. The fusion block consists of a regular Transformer encoder layer replicated " + }, + { + "bbox": [ + 46, + 471, + 289, + 664 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 471, + 289, + 664 + ], + "type": "text", + "content": " times. Finally, the visual tokens are regrouped into their initial shape to compose the feature map for the feature pyramid network (FPN)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 665, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 289, + 715 + ], + "type": "text", + "content": "Multiscale fusion. The multimodal fusion block outputs multiple fused feature maps of varying scales: " + }, + { + "bbox": [ + 47, + 665, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{fi} \\in \\mathbb{R}^{C_i \\times H_i \\times W_i}" + }, + { + "bbox": [ + 47, + 665, + 289, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 665, + 289, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 665, + 289, + 715 + ], + "type": "text", + "content": " is the scale index. They are fed into the FPN, which has different parameters for each feature" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "map, allowing the network to learn specialized features for each downsampling ratio. The output of the FPN is processed by a Faster R-CNN detector." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 110, + 545, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 146 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 146 + ], + "type": "text", + "content": "Prediction head networks. The Faster R-CNN detector outputs the bounding box " + }, + { + "bbox": [ + 304, + 110, + 545, + 146 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 110, + 545, + 146 + ], + "type": "text", + "content": " through a regression layer, an associated object " + }, + { + "bbox": [ + 304, + 110, + 545, + 146 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 110, + 545, + 146 + ], + "type": "text", + "content": " and verb " + }, + { + "bbox": [ + 304, + 110, + 545, + 146 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 110, + 545, + 146 + ], + "type": "text", + "content": ", and time to contact " + }, + { + "bbox": [ + 304, + 110, + 545, + 146 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 110, + 545, + 146 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 148, + 545, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 545, + 220 + ], + "type": "text", + "content": "Training objective. We train TransFusion using the standard Faster R-CNN [42] objective " + }, + { + "bbox": [ + 304, + 148, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{box}" + }, + { + "bbox": [ + 304, + 148, + 545, + 220 + ], + "type": "text", + "content": " for bounding box prediction, which consists of localization and objectness terms; further, cross-entropy losses " + }, + { + "bbox": [ + 304, + 148, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{noun}" + }, + { + "bbox": [ + 304, + 148, + 545, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 148, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{verb}" + }, + { + "bbox": [ + 304, + 148, + 545, + 220 + ], + "type": "text", + "content": " for noun and verb prediction respectively, and an L1 loss " + }, + { + "bbox": [ + 304, + 148, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ttc}" + }, + { + "bbox": [ + 304, + 148, + 545, + 220 + ], + "type": "text", + "content": " for time to contact prediction. The overall objective is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 351, + 229, + 545, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 229, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 351, + 229, + 545, + 243 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {b o x}} + \\mathcal {L} _ {\\text {n o u n}} + \\mathcal {L} _ {\\text {v e r b}} + \\mathcal {L} _ {\\text {t t c}}. \\tag {1}", + "image_path": "5ac13dc9a693afa29328517a082c1a91db01391bfe3d1ca7df7fb16597b2abcf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 251, + 468, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 251, + 468, + 262 + ], + "spans": [ + { + "bbox": [ + 318, + 251, + 468, + 262 + ], + "type": "text", + "content": "As defined in Faster R-CNN, " + }, + { + "bbox": [ + 318, + 251, + 468, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{box}" + }, + { + "bbox": [ + 318, + 251, + 468, + 262 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 270, + 545, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 545, + 308 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {b o x}} = \\frac {1}{N _ {\\mathrm {c l s}}} \\sum_ {i} \\mathcal {L} _ {\\mathrm {c l s}} \\left(p _ {i}, p _ {i} ^ {*}\\right) + \\frac {\\lambda}{N _ {\\mathrm {r e g}}} \\sum_ {i} p _ {i} ^ {*} \\cdot L _ {\\mathrm {r e g}} \\left(b _ {i} - b _ {i} ^ {*}\\right), \\tag {2}", + "image_path": "20dff26c9d72c109f7c1d491c140c1747bc3d4866ded11f0b7191a388d04f015.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 309, + 545, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 309, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 309, + 545, + 381 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 309, + 545, + 381 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 309, + 545, + 381 + ], + "type": "text", + "content": " is the predicted probability of a box being classified as a foreground object (true positive), " + }, + { + "bbox": [ + 304, + 309, + 545, + 381 + ], + "type": "inline_equation", + "content": "p_i^*" + }, + { + "bbox": [ + 304, + 309, + 545, + 381 + ], + "type": "text", + "content": " is the ground-truth label; " + }, + { + "bbox": [ + 304, + 309, + 545, + 381 + ], + "type": "inline_equation", + "content": "b_i" + }, + { + "bbox": [ + 304, + 309, + 545, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 309, + 545, + 381 + ], + "type": "inline_equation", + "content": "b_i^*" + }, + { + "bbox": [ + 304, + 309, + 545, + 381 + ], + "type": "text", + "content": " are the predicted and ground-truth box coordinates. " + }, + { + "bbox": [ + 304, + 309, + 545, + 381 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{reg}}" + }, + { + "bbox": [ + 304, + 309, + 545, + 381 + ], + "type": "text", + "content": " is the number of bounding boxes used in training, " + }, + { + "bbox": [ + 304, + 309, + 545, + 381 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{cls}}" + }, + { + "bbox": [ + 304, + 309, + 545, + 381 + ], + "type": "text", + "content": " the detection network's image batch size. We leave " + }, + { + "bbox": [ + 304, + 309, + 545, + 381 + ], + "type": "inline_equation", + "content": "\\lambda = 11" + }, + { + "bbox": [ + 304, + 309, + 545, + 381 + ], + "type": "text", + "content": " as in the reference implementation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 392, + 387, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 387, + 405 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 387, + 405 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 411, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 567 + ], + "type": "text", + "content": "Dataset. We use both versions (v1 and v2) of the dataset provided for the short-term object interaction anticipation task in Ego4D [22]. It contains various activities ranging from gardening work to cooking or car parts changing. There are in total 64,798 (v2: 165,451) annotated video clips of " + }, + { + "bbox": [ + 304, + 411, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\sim 5" + }, + { + "bbox": [ + 304, + 411, + 545, + 567 + ], + "type": "text", + "content": " minutes each with frame rate " + }, + { + "bbox": [ + 304, + 411, + 545, + 567 + ], + "type": "inline_equation", + "content": "fps = 30" + }, + { + "bbox": [ + 304, + 411, + 545, + 567 + ], + "type": "text", + "content": ". The dataset has a highly imbalanced long tail distribution, including 87 nouns and 74 verbs (v2: 128 and 81 resp.). For instance, the \"take\" class encompasses almost " + }, + { + "bbox": [ + 304, + 411, + 545, + 567 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 411, + 545, + 567 + ], + "type": "text", + "content": " of all the verb labels in version 1, while \"mold\" appears only in one video. The train set consists of about 28k (v2: 98k) samples, while the rest are roughly equally split between validation and test." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 569, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 676 + ], + "type": "text", + "content": "Evaluation. As in the official Ego4D benchmark [22], we use the Top-5 mean Average Precision (mAP) to evaluate the performance of individual predictions. It considers two constraints: The IoU constraint requires that the predicted boxes are counted as hits only if there is an intersection over union IoU " + }, + { + "bbox": [ + 304, + 569, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\geq 0.5" + }, + { + "bbox": [ + 304, + 569, + 545, + 676 + ], + "type": "text", + "content": " with the ground-truth box. The TTC constraint requires that the predicted time-to-contact " + }, + { + "bbox": [ + 304, + 569, + 545, + 676 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 304, + 569, + 545, + 676 + ], + "type": "text", + "content": " is close enough to the ground-truth " + }, + { + "bbox": [ + 304, + 569, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\hat{t}_i" + }, + { + "bbox": [ + 304, + 569, + 545, + 676 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 304, + 569, + 545, + 676 + ], + "type": "inline_equation", + "content": "|t_i - \\hat{t}_i| < T_\\delta, T_\\delta = 0.25" + }, + { + "bbox": [ + 304, + 569, + 545, + 676 + ], + "type": "text", + "content": ". The following metrics are considered:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 677, + 545, + 713 + ], + "type": "text", + "content": "- Noun refers to the Noun-Box Top-5 mAP. In addition to the IoU constraint, a correct match with the ground-truth noun is required." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18290" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 184 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 184 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 184 + ], + "type": "table", + "html": "
Ver.ModelN↑N-V ↑N-T ↑A↑
v1FRCNN+Rnd. [22]20.452.223.860.44
FRCNN+SF [22]20.456.786.172.45
InternVideo [7]24.609.187.643.40
StillFast [40]19.519.956.453.49
TF (ours)24.9010.067.573.54
v2FRCNN+SF [22]26.159.458.693.61
StillFast [40]25.0613.299.145.12
GANov2 [46]25.6713.609.025.16
TF (ours)30.4313.4510.385.18
", + "image_path": "3d5187d91dda355a3a175239bb85fe7d7ec5cf23bc0fd5522819e758861d34ff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 190, + 287, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 287, + 245 + ], + "type": "text", + "content": "Table 1. Comparison with the state-of-the-art. Our method TransFusion (TF) outperforms all state-of-the-art approaches on 3 out of 4 metrics, on both versions of the Ego4D test set. We report the top-5 mean average precision for the Noun (N), NounVerb (N-V), Noun-TTC (N-T), and Noun-Verb-TTC (A) metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 58, + 259, + 286, + 345 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 58, + 259, + 286, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 259, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 58, + 259, + 286, + 282 + ], + "type": "text", + "content": "- Noun-Verb refers to the Noun-Verb-Box Top-5 mAP. Correct noun and verb match with the IoU constraint." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 285, + 286, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 285, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 58, + 285, + 286, + 308 + ], + "type": "text", + "content": "- Noun-TTC refers to the Noun-Box Top-5 mAP. Correct noun matches with the TTC constraint." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 311, + 286, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 311, + 286, + 345 + ], + "spans": [ + { + "bbox": [ + 58, + 311, + 286, + 345 + ], + "type": "text", + "content": "- Noun-Verb-TTC (\"Overall\") refers to the overall Top-5 mAP, intersecting correct noun-verb matches with IoU and TTC constraints." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 348, + 286, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 286, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 286, + 383 + ], + "type": "text", + "content": "Note that since detecting the next-active objects is the main focus of the task, all evaluation metrics are conditioned on the IoU constraint." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 385, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 613 + ], + "type": "text", + "content": "Implementation details. We make use of the pre-trained Torchvision [34] Faster R-CNN. ResNet-50 [23] is used as the visual encoder and Sentence-BERT (SBERT) [41] is used as the language encoder. We freeze ResNet-50 and apply the multimodal fusion scheme on top. We use the RAdam [31] optimizer with a weight decay of " + }, + { + "bbox": [ + 46, + 385, + 287, + 613 + ], + "type": "inline_equation", + "content": "2e^{-4}" + }, + { + "bbox": [ + 46, + 385, + 287, + 613 + ], + "type": "text", + "content": ", and an effective batch size of 32. Besides the default Faster R-CNN background class in nouns, we also add it in the verb class, which improves the Noun-Verb mAP by almost 2 points. We reduce the region proposal network's sampling batch size and the detection network's image batch size and use data augmentation techniques to reduce overfitting. The number of verb-noun pairs used in the input is defined as the context length " + }, + { + "bbox": [ + 46, + 385, + 287, + 613 + ], + "type": "inline_equation", + "content": "L_{c}" + }, + { + "bbox": [ + 46, + 385, + 287, + 613 + ], + "type": "text", + "content": ". On average, an action segment described by one verb-noun phrase corresponds to one second with 30 frames, as computed from the GT annotations. Unless specified otherwise, we use " + }, + { + "bbox": [ + 46, + 385, + 287, + 613 + ], + "type": "inline_equation", + "content": "L_{c} = 3" + }, + { + "bbox": [ + 46, + 385, + 287, + 613 + ], + "type": "text", + "content": ". We perform postprocessing involving non-maximum suppression on the predicted bounding boxes, as elaborated in the Supp. Mat." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 622, + 224, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 224, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 224, + 635 + ], + "type": "text", + "content": "4.1. Comparison with state-of-the-art" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": "We first compare our method against the baselines proposed in the Ego4D work [22]. The FRCNN + SF method employs a two-stage approach: a ResNet-based Faster R-CNN detector first produces bounding box regressions and noun predictions from the prediction frame, without using any video features. In the second stage, a SlowFast [16] 3D" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 71, + 542, + 199 + ], + "blocks": [ + { + "bbox": [ + 312, + 71, + 542, + 199 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 542, + 199 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 542, + 199 + ], + "type": "image", + "image_path": "074104493356d0743a5070d5e10fd5fb5890bc65383919de0f728cf803c48a64.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 203, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 304, + 203, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 203, + 545, + 258 + ], + "type": "text", + "content": "Figure 4. Prediction of Noun-Verb in relation to time-to-contact. Histogram of the time-to-contact labels in Ego4Dv1 is shown in blue bars. Performance measured in Noun-Verb is plotted as a function of time-to-contact. We compare our method to Ego4D's official FRCNN + SF [22] baseline." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 263, + 545, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 545, + 359 + ], + "type": "text", + "content": "CNN video processing module performs region-of-interest (RoI) pooling [21] on the corresponding SlowFast 3D CNN video features. The pooled video features are then used by RoI heads to predict the verb and TTC values. The FR-CNN + Rnd. baseline, with results published on v1 only, drops the second stage of FRCNN + SF and uses randomly drawn verb and TTC predictions, following their respective distributions in the training set." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 361, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 456 + ], + "type": "text", + "content": "Another baseline on v1 is given by InternVideo [7], a method achieving strong results on the short-term object interaction anticipation task with a two-stage strategy similar to FRCNN+SF: bounding boxes and noun labels are first obtained using a DINO-based object detector [62], after which a VideoMAE-pretrained [49] ViT [14] model leverages space-time attention to predict the corresponding verb classes and TTC values." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 458, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 589 + ], + "type": "text", + "content": "We further compare TransFusion to the recent StillFast [40] approach on both dataset versions. StillFast uses a Faster-RCNN-based spatial stream operating on a high-resolution version of the input image, and an X3D-based [15] temporal stream operating on a low-resolution version of the input video. The two streams each produce a multiscale feature map. Both feature maps are fused into a combined feature pyramid, from which local RoI features and global scene features are extracted, merged by a fusion network, and finally used by linear layers to predict each boxes' noun, verb and TTC." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 592, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 687 + ], + "type": "text", + "content": "GANOV2 [46] enhances StillFast by adding a multi-head guided attention mechanism to its temporal stream, achieving favorable performance on version 2 of the dataset. The mechanism attends to object detections produced from the network's temporal branch, together with patches from the 3-D temporal feature stack, in a cross-fusion manner, and uses the fused tokens as the input to StillFast's feature pyramid and subsequent architecture." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "Table 1 demonstrates that TransFusion outperforms all state-of-the-art methods in terms of the N, N-V and A met-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18291" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 64, + 269, + 168 + ], + "blocks": [ + { + "bbox": [ + 69, + 64, + 269, + 168 + ], + "lines": [ + { + "bbox": [ + 69, + 64, + 269, + 168 + ], + "spans": [ + { + "bbox": [ + 69, + 64, + 269, + 168 + ], + "type": "image", + "image_path": "edbb1320c1404778505b6b5f81388d0d766cb2538b2dca253dc38b7479457ed9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 171, + 287, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 226 + ], + "type": "text", + "content": "Figure 5. Classification performance on top/tail categories. We show the relative and absolute gains of TransFusion over Ego4D's FRCNN + SF baseline [22] on the validation set, for Noun-Only and Verb-Only mAP (without any box IoU constraint). Relative improvements are written on top of the red bars." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 324 + ], + "type": "text", + "content": "rics on version 1 of the dataset, and approaches the state-of-the art performance on the N-T metric while requiring a fraction of the competitor's resources during training [7] (see the Supp. Mat. for details). On the dataset's second version, TransFusion performs best in terms of N, N-T and A, while approaching the runner-up method on N-V. These favorable results validate the advantages of modeling action context explicitly for the interaction anticipation task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 325, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 287, + 467 + ], + "type": "text", + "content": "We henceforth analyze our performance gains over the official Ego4D FRCNN + SF baseline on the first version of the Ego4D dataset. Figure 4 shows the long tail distribution of the time-to-contact labels in the validation set, split into buckets. We see that TransFusion consistently outperforms the baseline on the N-V metric for all TTC ranges. The results in Figure 5 indicate that our method not only outperforms the Ego4D method on top noun and verb classes, but also achieves bigger gains in the tail classes, demonstrating the generalization ability of using our language-based context summaries. Top and tail buckets represent roughly " + }, + { + "bbox": [ + 46, + 325, + 287, + 467 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 325, + 287, + 467 + ], + "type": "text", + "content": " of the samples in the validation set." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 469, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 553 + ], + "type": "text", + "content": "Figure 6 shows that the Ego4D approach manages to detect the most salient objects in the environment, but fails to reliably associate the bounding boxes to the correct action phrase. In contrast, TransFusion is able to correctly associate the drawer bounding box to \"open drawer\" and to predict the correct action associated with the phone. More examples are provided in the Supp. Mat." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 562, + 145, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 145, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 145, + 574 + ], + "type": "text", + "content": "4.2. Ablation studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 581, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 688 + ], + "type": "text", + "content": "Evaluation of action context representation. In Table 2, we ablate different variants of context representation and compare against the representation that uses the groundtruth (GT) action labels provided by Ego4D in the action context, to assess the quality of our generated context summaries. The observed TTC Average Precision difference measures around " + }, + { + "bbox": [ + 46, + 581, + 287, + 688 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 581, + 287, + 688 + ], + "type": "text", + "content": " and was thus omitted from the comparison. We report on the Noun and Noun-Verb metrics as we expect major differences in semantic-related tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Experiments show that we achieve results approaching GT, demonstrating the effectiveness of our context repre" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 81, + 421, + 164 + ], + "blocks": [ + { + "bbox": [ + 321, + 69, + 409, + 80 + ], + "lines": [ + { + "bbox": [ + 321, + 69, + 409, + 80 + ], + "spans": [ + { + "bbox": [ + 321, + 69, + 409, + 80 + ], + "type": "text", + "content": "GT: open drawer 0.1s" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 81, + 421, + 164 + ], + "lines": [ + { + "bbox": [ + 310, + 81, + 421, + 164 + ], + "spans": [ + { + "bbox": [ + 310, + 81, + 421, + 164 + ], + "type": "image", + "image_path": "2760f6325fb033de5848b12ab2a0db695cc01daa0951789615e11bb639c27c08.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 164, + 424, + 195 + ], + "lines": [ + { + "bbox": [ + 308, + 164, + 424, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 424, + 195 + ], + "type": "text", + "content": "(a) Ego4D: take container 0.78s, take container 0.46s, take pot 0.42s, take lid 0.46s" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 429, + 81, + 539, + 163 + ], + "blocks": [ + { + "bbox": [ + 441, + 70, + 522, + 80 + ], + "lines": [ + { + "bbox": [ + 441, + 70, + 522, + 80 + ], + "spans": [ + { + "bbox": [ + 441, + 70, + 522, + 80 + ], + "type": "text", + "content": "GT: take phone 1.0s" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 429, + 81, + 539, + 163 + ], + "lines": [ + { + "bbox": [ + 429, + 81, + 539, + 163 + ], + "spans": [ + { + "bbox": [ + 429, + 81, + 539, + 163 + ], + "type": "image", + "image_path": "f347836900e00440eff16f08c3b1a20da37d3039bebd62ce7fbfd2bcac128d1c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 164, + 544, + 195 + ], + "lines": [ + { + "bbox": [ + 426, + 164, + 544, + 195 + ], + "spans": [ + { + "bbox": [ + 426, + 164, + 544, + 195 + ], + "type": "text", + "content": "(b) Ego4D: take table 0.78s, take container 0.85s, take computer 0.86s, take phone 1.11s" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 198, + 419, + 279 + ], + "blocks": [ + { + "bbox": [ + 310, + 198, + 419, + 279 + ], + "lines": [ + { + "bbox": [ + 310, + 198, + 419, + 279 + ], + "spans": [ + { + "bbox": [ + 310, + 198, + 419, + 279 + ], + "type": "image", + "image_path": "f275f6477bf296d642ce09639b28fbbdb8f8c32b7a37ecd75be9cb706813e18f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 280, + 425, + 311 + ], + "lines": [ + { + "bbox": [ + 306, + 280, + 425, + 311 + ], + "spans": [ + { + "bbox": [ + 306, + 280, + 425, + 311 + ], + "type": "text", + "content": "(c) TF (ours): open drawer 0.56s, open drawer 0.36s, open drawer 0.44s, take lid 0.6s" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 428, + 198, + 538, + 280 + ], + "blocks": [ + { + "bbox": [ + 428, + 198, + 538, + 280 + ], + "lines": [ + { + "bbox": [ + 428, + 198, + 538, + 280 + ], + "spans": [ + { + "bbox": [ + 428, + 198, + 538, + 280 + ], + "type": "image", + "image_path": "5e1888e52b63b20e82773356a1746d7e8c773d7b1012884f11f87efe7f330a24.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 280, + 543, + 312 + ], + "lines": [ + { + "bbox": [ + 426, + 280, + 543, + 312 + ], + "spans": [ + { + "bbox": [ + 426, + 280, + 543, + 312 + ], + "type": "text", + "content": "(d) TF (ours): take phone 0.87s, open computer 0.63s, operate tablet 0.73s, take phone 0.98s" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 316, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 305, + 316, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 545, + 361 + ], + "type": "text", + "content": "Figure 6. Qualitative results of the proposed approach. Results of Ego4D's FRCNN + SF baseline [22] are shown in the top row and TransFusion in the bottom row. Green denotes the ground truth bounding box, with the labels shown on top of each column." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 309, + 368, + 541, + 404 + ], + "blocks": [ + { + "bbox": [ + 309, + 368, + 541, + 404 + ], + "lines": [ + { + "bbox": [ + 309, + 368, + 541, + 404 + ], + "spans": [ + { + "bbox": [ + 309, + 368, + 541, + 404 + ], + "type": "table", + "html": "
Rep.∅ANhNsA+NhA+NsGT
N↑17.7120.0919.3319.7519.0320.1921.71
N-V↑6.147.166.867.057.127.557.86
", + "image_path": "a61616aac5c4d615da69ed141d1e3f176671e6c2f7d79f6a060e06ca8f2cfc20.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 407, + 545, + 484 + ], + "lines": [ + { + "bbox": [ + 305, + 407, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 545, + 484 + ], + "type": "text", + "content": "Table 2. Comparison of different action context representations. We evaluate different variations of context representations on the predicted noun (N) and noun-verb (N-V), and compare them against ground-truth annotations (GT) provided by v1 of the Ego4D validation set. GT only contains action descriptions of verb-noun pairs. Action context represented by " + }, + { + "bbox": [ + 305, + 407, + 545, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{A} + \\mathcal{N}_s" + }, + { + "bbox": [ + 305, + 407, + 545, + 484 + ], + "type": "text", + "content": " achieves the best N-V performance." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "text", + "content": "sentations. Using a vision-only version of our model " + }, + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "inline_equation", + "content": "(\\emptyset)" + }, + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "text", + "content": " yields markedly worse results than when adding any type of language conditioning. Action descriptions " + }, + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "text", + "content": " consistently boost the verb classification performance. Adding salient objects " + }, + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_s" + }, + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "text", + "content": " further improves task performance. These together offer a reasonable picture of using language summaries to represent the action context. In addition, we notice a domain gap in the choice of words between the generated labels and the ground-truth ones. In some cases, even if the description is satisfactory for a human evaluator, the information may be too vague to be directly useful to the model. Phrasing differences between the generated captions and the training corpus of the language encoders may further help to improve the performance." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "As an additional type of action context, we hypothesize that the sequence of objects held by the actor may be indicative of the overall task and can help the model better infer the immediate next steps. Motivated by this, we use an existing hand-object interaction predictor [11] jointly with an" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18292" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 67, + 252, + 174 + ], + "blocks": [ + { + "bbox": [ + 75, + 67, + 252, + 174 + ], + "lines": [ + { + "bbox": [ + 75, + 67, + 252, + 174 + ], + "spans": [ + { + "bbox": [ + 75, + 67, + 252, + 174 + ], + "type": "image", + "image_path": "940a5f7209592e370d384e48631731a1fa40247041117243d0139f6e40d41cee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 178, + 287, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 287, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 287, + 222 + ], + "type": "text", + "content": "Figure 7. Comparison to video features. TransFusion outperforms TransFusion-Video over different context lengths measured by Noun-Verb mAP. TransFusion using GT-based context summaries as language input represents an upper performance bound." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 348 + ], + "type": "text", + "content": "object detector [65] to look for held objects and their semantic labels in each frame. Usually, such objects occur only in a subset of frames (e.g. due to limited visibility of the actor's hands), which results in a lower performance of the resulting " + }, + { + "bbox": [ + 46, + 228, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_h" + }, + { + "bbox": [ + 46, + 228, + 287, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 228, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{A} + \\mathcal{N}_h" + }, + { + "bbox": [ + 46, + 228, + 287, + 348 + ], + "type": "text", + "content": " representations. We hence use " + }, + { + "bbox": [ + 46, + 228, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{A} + \\mathcal{N}_s" + }, + { + "bbox": [ + 46, + 228, + 287, + 348 + ], + "type": "text", + "content": " to represent the action context, considering the good performance achieved. Ablations on the hyperparameters " + }, + { + "bbox": [ + 46, + 228, + 287, + 348 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 228, + 287, + 348 + ], + "type": "text", + "content": ", the cutoff distance, and " + }, + { + "bbox": [ + 46, + 228, + 287, + 348 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 228, + 287, + 348 + ], + "type": "text", + "content": ", the number of salient objects per frame, show that the performance is less sensitive to their settings. More details can be found in the Supp. Mat." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 287, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 421 + ], + "type": "text", + "content": "Comparison to video features. We evaluate whether language summaries can better represent action context than video features for the object-anticipation task. To do so, we implement TransFusion-Video, which takes video features extracted from the Ego4D SlowFast [16] model as input and uses an identical fusion module as in TransFusion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 424, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 639 + ], + "type": "text", + "content": "Figure 7 shows that TransFusion consistently outperforms TransFusion-Video. " + }, + { + "bbox": [ + 46, + 424, + 287, + 639 + ], + "type": "inline_equation", + "content": "L_{c} \\times 30" + }, + { + "bbox": [ + 46, + 424, + 287, + 639 + ], + "type": "text", + "content": " frames are taken in TransFusion-Video. For video features, the domain gap between pre-training and target datasets seems to be too large to allow effective generalization to a diverse dataset such as Ego4D. We suspect that the feature representations are more entangled due to the need to represent temporal, visual, and semantic aspects simultaneously. With language, temporal information is directly encoded in the syntax, i.e. the succession of words. Given a video clip of one second, language can summarize it in two words whereas corresponding video features use more than 12 times the space (e.g. smaller SBERT [41] features of " + }, + { + "bbox": [ + 46, + 424, + 287, + 639 + ], + "type": "inline_equation", + "content": "2 \\times 384" + }, + { + "bbox": [ + 46, + 424, + 287, + 639 + ], + "type": "text", + "content": " compared to SlowFast [16] features of " + }, + { + "bbox": [ + 46, + 424, + 287, + 639 + ], + "type": "inline_equation", + "content": "4 \\times 2304" + }, + { + "bbox": [ + 46, + 424, + 287, + 639 + ], + "type": "text", + "content": "). Therefore, our language-based context representation has an advantage when it comes to longer video sequences. Overall, TransFusion obtains a " + }, + { + "bbox": [ + 46, + 424, + 287, + 639 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 46, + 424, + 287, + 639 + ], + "type": "text", + "content": " boost in N-V performance. See the Supp. Mat. for details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": "Length of action context. A significant parameter in obtaining good performance is the length of action context, i.e., how many past actions to include in the input. An adequate context length is important to allow the model to pick up more specific activity patterns. Some activities that have a more structured nature or certain repeating patterns (e.g." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 312, + 70, + 541, + 128 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 541, + 128 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 541, + 128 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 541, + 128 + ], + "type": "table", + "html": "
Language EncoderToken sizeN↑N-V↑
SBERT/BERT [13, 41]38420.197.55
SBERT/RoBERTa [33, 41]76819.787.78
GPT-2 [37]76818.987.08
Flan-T5 [8]102418.266.61
", + "image_path": "a85093ff60befe94cb1db30be749e073ca104ce1fc6217a43d1fa2f1130037a9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 137, + 545, + 159 + ], + "lines": [ + { + "bbox": [ + 305, + 137, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 305, + 137, + 545, + 159 + ], + "type": "text", + "content": "Table 3. Language encoder ablations. We compare SBERT using BERT and RoBERTa backbones, with GPT-2 and Flan-T5." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 167, + 545, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 545, + 214 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 545, + 214 + ], + "type": "text", + "content": "cooking or repairing a bicycle) are easier to model with an increased context length compared to more random activities such as playing basketball. Figure 7 shows that the overall performance plateaus at " + }, + { + "bbox": [ + 304, + 167, + 545, + 214 + ], + "type": "inline_equation", + "content": "L_{c} = 2" + }, + { + "bbox": [ + 304, + 167, + 545, + 214 + ], + "type": "text", + "content": " and higher." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 217, + 546, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 546, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 546, + 408 + ], + "type": "text", + "content": "Language encoder. A language encoder is used to process the context summaries in TransFusion. To understand the impact of different language encoders, we experiment with multiple LLMs including SBERT with a RoBERTa backbone [33, 41], GPT-2 [37], and FLAN-T5 [8], using version 1 of the validation set. Table 3 summarizes the performance of different language encoders, showing that the SBERT models perform best. We consider TTC less relevant here. We hypothesize that the difference comes from the training objectives of language models. SBERT is trained to map semantically similar sentences closer to each other in the embedding space. This is particularly useful when the model needs to understand descriptions with semantically close words, e.g. \"cut carrots\" and \"slice carrots\". GPT-2 and FLAN-T5 are optimized using generative objectives, which makes them overly sensitive to the input word choice." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 410, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 410, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 410, + 545, + 434 + ], + "type": "text", + "content": "Fusion module. We provide ablations regarding the fusion module in the Supp. Mat." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 450, + 455, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 450, + 455, + 463 + ], + "spans": [ + { + "bbox": [ + 306, + 450, + 455, + 463 + ], + "type": "text", + "content": "5. Conclusion and discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 472, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 675 + ], + "type": "text", + "content": "We propose TransFusion, a multimodal fusion model that anticipates object interactions by considering the past action context represented by language summaries. Our successful experiments on a challenging egocentric video dataset demonstrate how language summaries and the vast knowledge contained in vision-language foundation models can improve object interaction anticipation, highlighting the representational power of language. It appears promising to investigate fusing even more modalities, such as motion cues from optical flow. We emphasize that the proposed summarization approach is not limited to this task. Language provides a universal interface to complement the visual input with information encoded in foundation models or task-specific pipelines. Future work can leverage language summarization for other video reasoning tasks [60, 63], or extend it temporally to describe possible future activities [1, 27] for long-term interaction anticipation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by an ETH Zurich Postdoctoral Fellowship. We thank Adrian Spurr for the helpful comments and Velko Vechev for the voiceover." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18293" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Chaitanya Ahuja and Louis-Philippe Morency. Language2Pose: Natural Language Grounded Pose Forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719–728, 2019. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 179 + ], + "type": "text", + "content": "[2] Alan Akbik, Duncan Blythe, and Roland Vollgraf. Contextual String Embeddings for Sequence Labeling. In Proceedings of the 27th international conference on computational linguistics, pages 1638-1649, 2018. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 234 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 234 + ], + "type": "text", + "content": "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. VQA: Visual Question Answering. In Proceedings of the IEEE International Conference on Computer Vision, pages 2425-2433, 2015. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 268 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 268 + ], + "type": "text", + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. CoRR, abs/2104.00650, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 288, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 288, + 302 + ], + "type": "text", + "content": "[5] Gedas Bertasius, Hyun Soo Park, Stella X. Yu, and Jianbo Shi. First-Person Action-Object Detection with EgoNet. In Proceedings of Robotics: Science and Systems, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 303, + 288, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 288, + 324 + ], + "type": "text", + "content": "[6] Jerome S Bruner. Intention in the Structure of Action and Interaction. Advances in Infancy Research, 1981. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 326, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 326, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 326, + 288, + 357 + ], + "type": "text", + "content": "[7] Guo Chen et al. InternVideo-Ego4D: A Pack of Champion Solutions to Ego4D Challenges. arXiv preprint arXiv:2211.09529, 2022.2,6,7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 359, + 288, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 359, + 288, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 359, + 288, + 457 + ], + "type": "text", + "content": "[8] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling Instruction-Finetuned Language Models, 2022. 3, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 459, + 288, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 459, + 288, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 459, + 288, + 512 + ], + "type": "text", + "content": "[9] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling Egocentric Vision. CoRR, abs/2006.13256, 2020. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 579 + ], + "type": "text", + "content": "[10] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling Egocentric Vision: Collection, Pipeline and Challenges for EPIC-Kitchens-100. International Journal of Computer Vision, 130(1):33-55, 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 288, + 645 + ], + "type": "text", + "content": "[11] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. EPIC-KITCHENS VISOR Benchmark: VVideo Segmentations and Object Relations. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2022. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "text", + "content": "[12] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. TransVG: End-to-End Visual Grounding with Transformers. CoRR, abs/2104.08541, 2021.3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "Transformers for Language Understanding. CoRR, abs/1810.04805, 2018.3,8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 162 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. CoRR, abs/2010.11929, 2020. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 164, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 208 + ], + "type": "text", + "content": "[15] Christoph Feichtenhofer. X3D: Expanding Architectures for Efficient Video Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 203-213, 2020. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 209, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 242 + ], + "type": "text", + "content": "[16] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast Networks for Video Recognition. CoRR, abs/1812.03982, 2018. 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 244, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 244, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 244, + 545, + 277 + ], + "type": "text", + "content": "[17] Antonino Furnari and Giovanni Maria Farinella. Rolling-Unrolling LSTMs for Action Anticipation from First-Person Video. CoRR, abs/2005.02190, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 278, + 545, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 278, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 307, + 278, + 545, + 322 + ], + "type": "text", + "content": "[18] Antonino Furnari, Sebastiano Battiato, Kristen Grauman, and Giovanni Maria Farinella. Next-Active-Object Prediction from Egocentric Videos. CoRR, abs/1904.05250, 2019. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 324, + 545, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 324, + 545, + 379 + ], + "spans": [ + { + "bbox": [ + 307, + 324, + 545, + 379 + ], + "type": "text", + "content": "[19] H. Girase, N. Agarwal, C. Choi, and K. Mangalam. Latency Matters: Real-Time Action Forecasting Transformer. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18759-18769, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 380, + 545, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 380, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 307, + 380, + 545, + 402 + ], + "type": "text", + "content": "[20] Rohit Girdhar and Kristen Grauman. Anticipative Video Transformer. In ICCV, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 404, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 425 + ], + "type": "text", + "content": "[21] Ross B. Girshick. Fast R-CNN. CoRR, abs/1504.08083, 2015.6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 428, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 428, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 428, + 547, + 713 + ], + "type": "text", + "content": "[22] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Wesley Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4D: Around the World in 3,000 Hours of Egocentric Video. CoRR, abs/2110.07058, 2021. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18294" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. CoRR, abs/1512.03385, 2015. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 106, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 106, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 49, + 106, + 287, + 148 + ], + "type": "text", + "content": "[24] Zhicheng Huang, Zhaoyang Zeng, Bei Liu, Dongmei Fu, and Jianlong Fu. Pixel-BERT: Aligning Image Pixels with Text by Deep Multi-Modal Transformers. CoRR, abs/2004.00849, 2020.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 150, + 287, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 150, + 287, + 193 + ], + "spans": [ + { + "bbox": [ + 49, + 150, + 287, + 193 + ], + "type": "text", + "content": "[25] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. EPIC-Fusion: Audio-Visual Temporal Binding for Egocentric Action Recognition. CoRR, abs/1908.08498, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 194, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 194, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 49, + 194, + 287, + 259 + ], + "type": "text", + "content": "[26] Weicheng Kuo, Fred Bertsch, Wei Li, AJ Piergiovanni, Mohammad Saffar, and Anelia Angelova. FindIt: Generalized Localization with Natural Language Queries. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI, pages 502-520. Springer, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 261, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 261, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 49, + 261, + 287, + 315 + ], + "type": "text", + "content": "[27] Yen-Ling Kuo, Xin Huang, Andrei Barbu, Stephen G McGill, Boris Katz, John J Leonard, and Guy Rosman. Trajectory Prediction with Linguistic Representations. In 2022 International Conference on Robotics and Automation (ICRA), pages 2868-2875. IEEE, 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 316, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 316, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 49, + 316, + 287, + 358 + ], + "type": "text", + "content": "[28] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. Less is More: ClipBERT for Video-and-Language Learning via Sparse Sampling. CoRR, abs/2102.06183, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 360, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 360, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 49, + 360, + 287, + 403 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Piotr Dólar, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature Pyramid Networks for Object Detection. CoRR, abs/1612.03144, 2016. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 404, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 404, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 49, + 404, + 287, + 458 + ], + "type": "text", + "content": "[30] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen Clip Models Are Efficient Video Learners. In European Conference on Computer Vision, pages 388-404. Springer, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 460, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 460, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 49, + 460, + 287, + 502 + ], + "type": "text", + "content": "[31] Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the Variance of the Adaptive Learning Rate and Beyond. CoRR, abs/1908.03265, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 503, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 503, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 49, + 503, + 287, + 545 + ], + "type": "text", + "content": "[32] Miao Liu, Siyu Tang, Yin Li, and James M. Rehg. Forecasting Human Object Interaction: Joint Prediction of Motor Attention and Egocentric Activity. CoRR, abs/1911.10967, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 548, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 548, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 49, + 548, + 287, + 601 + ], + "type": "text", + "content": "[33] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A Robustly Optimized BERT Pretraining Approach. CoRR, abs/1907.11692, 2019.8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 603, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 603, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 49, + 603, + 287, + 635 + ], + "type": "text", + "content": "[34] TorchVision maintainers and contributors. TorchVision: PyTorch's Computer Vision Library. https://github.com/pytorch/vision, 2016.6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 636, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 636, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 49, + 636, + 287, + 657 + ], + "type": "text", + "content": "[35] David W Orr. The Nature of Design: Ecology, Culture, and Human Intention. Oxford University Press, 2002. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "type": "text", + "content": "[36] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou, Rama Chellappa, and Pengchuan Zhang. EGOVLPV2: Egocentric Video-Language Pre-training With Fusion in the Backbone, 2023. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[37] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language Models are Unsupervised Multitask Learners. OpenAI Blog, 2019. 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 173 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. CoRR, abs/2103.00020, 2021. 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "text", + "content": "[39] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. CoRR, abs/1910.10683, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 232, + 545, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 265 + ], + "type": "text", + "content": "[40] Francesco Ragusa et al. StillFast: An End-to-End Approach for Short-Term Object Interaction Anticipation. arXiv preprint arXiv:2304.03959, 2023. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 267, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 267, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 267, + 545, + 300 + ], + "type": "text", + "content": "[41] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. CoRR, abs/1908.10084, 2019. 3, 5, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 302, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 302, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 307, + 302, + 545, + 335 + ], + "type": "text", + "content": "[42] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. CoRR, abs/1506.01497, 2015. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 337, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 337, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 545, + 369 + ], + "type": "text", + "content": "[43] Paschal Sheeran. Intention-Behavior Relations: A Conceptual and Empirical Review. European Review of Social Psychology, 12(1):1-36, 2002. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 372, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 404 + ], + "type": "text", + "content": "[44] Chen Sun, Fabien Baradel, Kevin Murphy, and Cordelia Schmid. Contrastive Bidirectional Transformer for Temporal Representation Learning. CoRR, abs/1906.05743, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 406, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 545, + 438 + ], + "type": "text", + "content": "[45] Hao Tan and Mohit Bansal. LXMERT: Learning Cross-Modality Encoder Representations from Transformers. CoRR, abs/1908.07490, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 441, + 545, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 545, + 496 + ], + "type": "text", + "content": "[46] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Enhancing Next Active Object-Based Egocentric Action Anticipation With Guided Attention. In 2023 IEEE International Conference on Image Processing (ICIP), pages 1450-1454. IEEE, 2023. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 498, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 542 + ], + "type": "text", + "content": "[47] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Guided Attention for Next Active Object @ EGO4D STA Challenge. arXiv preprint arXiv:2305.16066, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 544, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 609 + ], + "type": "text", + "content": "[48] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Leveraging Next-Active Objects for Context-Aware Anticipation in Egocentric Videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 8657–8666, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 612, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 666 + ], + "type": "text", + "content": "[49] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked Autoencoders Are Data-Efficient Learners for Self-Supervised Video Pre-Training. Advances in Neural Information Processing Systems, 35:10078–10093, 2022. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. CoRR, abs/1706.03762, 2017. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18295" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[51] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-Anticipation Transformer for Online Action Understanding, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "text", + "content": "[52] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Action-CLIP: A New Paradigm for Video Action Recognition. arXiv preprint arXiv:2109.08472, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 286, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 286, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 286, + 196 + ], + "type": "text", + "content": "[53] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework. CoRR, abs/2202.03052, 2022. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 286, + 251 + ], + "type": "text", + "content": "[54] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. MeMViT: Memory-Augmented Multiscale Vision Transformer for Efficient Long-Term Video Recognition. CoRR, abs/2201.08383, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 254, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 254, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 254, + 286, + 285 + ], + "type": "text", + "content": "[55] Yu Wu, Linchao Zhu, Xiaohan Wang, Yi Yang, and Fei Wu. Learning to Anticipate Egocentric Actions by Imagination. CoRR, abs/2101.04924, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 287, + 286, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 287, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 287, + 286, + 342 + ], + "type": "text", + "content": "[56] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, Attend and Tell: Neural Image Caption Generation With Visual Attention. In International Conference on Machine Learning, pages 2048-2057. PMLR, 2015. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 343, + 286, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 343, + 286, + 387 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 286, + 387 + ], + "type": "text", + "content": "[57] Ting Yao, Yingwei Pan, Yehao Li, Zhaofan Qiu, and Tao Mei. Boosting Image Captioning With Attributes. In Proceedings of the IEEE International Conference on Computer Vision, pages 4894-4902, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 388, + 286, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 286, + 421 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 286, + 421 + ], + "type": "text", + "content": "[58] Moon Ye-Bin, Jisoo Kim, Hongyeob Kim, Kilho Son, and Tae-Hyun Oh. TextMania: Enriching Visual Feature by Text-driven Manifold Augmentation, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 422, + 286, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 422, + 286, + 465 + ], + "spans": [ + { + "bbox": [ + 49, + 422, + 286, + 465 + ], + "type": "text", + "content": "[59] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. MERLOT: Multimodal Neural Script Knowledge Models. CoRR, abs/2106.02636, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 468, + 286, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 468, + 286, + 521 + ], + "spans": [ + { + "bbox": [ + 49, + 468, + 286, + 521 + ], + "type": "text", + "content": "[60] Kuo-Hao Zeng, Tseng-Hung Chen, Ching-Yao Chuang, Yuan-Hong Liao, Juan Carlos Niebles, and Min Sun. Leveraging Video Descriptions to Learn Video Question Answering. In Thirty-First AAAI Conference on Artificial Intelligence, 2017. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 523, + 286, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 523, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 49, + 523, + 286, + 578 + ], + "type": "text", + "content": "[61] Ce Zhang, Changcheng Fu, Shijie Wang, Nakul Agarwal, Kwonjoon Lee, Chiho Choi, and Chen Sun. Object-Centric Video Representation for Long-term Action Anticipation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6751-6761, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 579, + 286, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 579, + 286, + 633 + ], + "spans": [ + { + "bbox": [ + 49, + 579, + 286, + 633 + ], + "type": "text", + "content": "[62] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection. In The Eleventh International Conference on Learning Representations, 2022. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 635, + 286, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 635, + 286, + 679 + ], + "spans": [ + { + "bbox": [ + 49, + 635, + 286, + 679 + ], + "type": "text", + "content": "[63] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. Video Summarization With Long Short-Term Memory. In European Conference on Computer Vision, pages 766-782. Springer, 2016. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "type": "text", + "content": "[64] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning Video Representations From Large Language Models, 2022. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 105 + ], + "type": "text", + "content": "[65] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple Multi-Dataset Detection. CoRR, abs/2102.13086, 2021. 8" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18296" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/ac654108-cae7-444a-a7e8-a0faf89a682d_content_list.json b/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/ac654108-cae7-444a-a7e8-a0faf89a682d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7d2ef0ca6f68f8cd25473c92eb254cd168d53168 --- /dev/null +++ b/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/ac654108-cae7-444a-a7e8-a0faf89a682d_content_list.json @@ -0,0 +1,1483 @@ +[ + { + "type": "text", + "text": "Super-Resolution Reconstruction from Bayer-Pattern Spike Streams", + "text_level": 1, + "bbox": [ + 138, + 130, + 828, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yanchen Dong $^{1}$ , Ruiqin Xiong $^{1*}$ , Jian Zhang $^{2}$ , Zhaofei Yu $^{3}$ , Xiaopeng Fan $^{4}$ , Shuyuan Zhu $^{5}$ , Tiejun Huang $^{1,3}$ $^{1}$ School of Computer Science, Peking University \n $^{2}$ School of Electronic and Computer Engineering, Peking University \n $^{3}$ Institute for Artificial Intelligence, Peking University \n $^{4}$ School of Computer Science and Technology, Harbin Institute of Technology \n $^{5}$ University of Electronic Science and Technology of China", + "bbox": [ + 86, + 178, + 929, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yanchendong@stu.pku.edu.cn, {rqxiong, zhangjian.sz, yuzf12, tjhuang}@pku.edu.cn, fxp@hit.edu.cn, eezsy@uestc.edu.cn", + "bbox": [ + 155, + 289, + 859, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 356, + 313, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spike camera is a neuromorphic vision sensor that can capture highly dynamic scenes by generating a continuous stream of binary spikes to represent the arrival of photons at very high temporal resolution. Equipped with Bayer color filter array (CFA), color spike camera (CSC) has been invented to capture color information. Although spike camera has already demonstrated great potential for high-speed imaging, its spatial resolution is limited compared with conventional digital cameras. This paper proposes a Color Spike Camera Super-Resolution (CSCSR) network to super-resolve higher-resolution color images from spike camera streams with Bayer CFA. To be specific, we first propose a representation for Bayer-pattern spike streams, exploring local temporal information with global perception to represent the binary data. Then we exploit the CFA layout and sub-pixel level motion to collect temporal pixels for the spatial super-resolution of each color channel. In particular, a residual-based module for feature refinement is developed to reduce the impact of motion estimation errors. Considering color correlation, we jointly utilize the multi-stage temporal-pixel features of color channels to reconstruct the high-resolution color image. Experimental results demonstrate that the proposed scheme can reconstruct satisfactory color images with both high temporal and spatial resolution from low-resolution Bayer-pattern spike streams. The source codes are available at https://github.com/csycdong/CSCSR.", + "bbox": [ + 76, + 388, + 473, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 823, + 209, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the quick development of high-speed vision applications such as autonomous driving and unmanned aerial ve", + "bbox": [ + 76, + 848, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/faeac15f9e3171a483a0abc4ff01fa9c5151a9eda65618a1ee817961e5b6c468.jpg", + "image_caption": [ + "Figure 1. Pipeline of super-resolution reconstruction from Bayer-pattern spike streams. The super-resolution scale is denoted as $s$ ." + ], + "image_footnote": [], + "bbox": [ + 503, + 354, + 890, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hicles, the demand for cameras that can capture high-speed motion and respond quickly is increasing. Most conventional digital cameras use a certain exposure time window to accumulate photoelectric signals and create a snapshot image, which requires the scene to be still during the exposure interval. Due to the exposure time, conventional cameras suffer from motion blur when capturing high-speed scenes. To be specific, a given point of moving objects can move during exposure time, bringing undesired motion blur to the image. Therefore, it's hard for conventional digital cameras to meet the requirements of high-speed imaging.", + "bbox": [ + 496, + 672, + 892, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mimicking the structure of human vision, neuromorphic cameras with ultra-high temporal resolution (e.g., $20,000\\mathrm{Hz}$ ) show great potential in capturing high-speed scenes. A well-known type of neuromorphic camera called", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24871", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "event camera [19, 21, 24] is sensitive to motion in dynamic scenes, which can record relative light intensity changes. To record more texture details, another type of neuromorphic camera called spike camera [8, 9, 49] captures absolute light intensity via an \"integrate-and-fire\" mechanism. As the first-generation spike camera can only reproduce gray-level signals, color spike camera, which employs the Bayer-pattern color filter array (CFA), has been invented to record dynamic scenes with color information and meanwhile enjoy the benefits of high-speed imaging.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although capable of capturing color information like conventional cameras, the spatial resolution of color spike cameras is limited compared to recent digital cameras, which is a trade-off for high temporal resolution and low latency. Therefore, we aim to reconstruct high-resolution (HR) color images from low-resolution (LR) Bayer-pattern spike streams, achieving both high temporal and spatial resolution with a color spike camera. An intuitive way is to combine color spike camera reconstruction methods [10] with image super-resolution (SR) algorithms [7, 18, 20]. However, information among temporal neighboring frames that is essential for SR will be degraded in the first reconstruction stage of this scheme. Besides, high-speed motion and quantization noise within the Bayer-pattern spike streams also bring challenges to the SR task. As a result, how to improve the spatial resolution as well as the imaging quality for color spike cameras is worth studying.", + "bbox": [ + 75, + 244, + 472, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a Color Spike Camera SuperResolution (CSCSR) network to super-resolve dynamic scenes from Bayer-pattern spike streams. To represent the binary data with color and motion information, we first develop a Bayer-pattern Spike Stream Representation (BSSR) to explore local temporal information with global perception. As the pixels of a single frame are limited for spatial super-resolution, we propose a Motion-guided SuperResolution (MSR) module to search temporal pixels of each color channel according to the Bayer-pattern color layout and sub-pixel level motion jointly estimated from color channels. In most cases, there are errors in the estimated motion, causing an undesired impact on the temporal pixel search. Therefore, a Feature Refinement (FR) module is designed based on the residual between a temporal average reference and the temporal pixels, resulting in refined temporal-pixel features. Finally, we further extract features and jointly utilize multi-stage features of each color channel to restore the final HR image, achieved by our Color Correlation-based Reconstruction (CCR) module.", + "bbox": [ + 75, + 505, + 472, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To train and evaluate our models, we design a simulator based on the mechanism of color spike cameras, which can generate data pairs for SR reconstruction. Besides, we also capture some real-world Bayer-pattern spike streams for further evaluation. Experimental results on both synthetic and real-world captured data show that our method", + "bbox": [ + 75, + 810, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "can reconstruct HR color images from LR Bayer-pattern spike streams, with good texture details. The main contribution of our work can be summarized as follows:", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop a color spike camera super-resolution network to reconstruct high-resolution color images from low-resolution Bayer-pattern spike streams, which is the first attempt to the best of our knowledge.", + "- To reconstruct details beyond the sensor resolution, we explore the sub-pixel level motion of each color channel and propose a motion-guided super-resolution module to collect temporal pixels according to the color layout.", + "- Experiments demonstrate that our method outperforms all the existing methods, achieving better quantitative and qualitative results in color spike camera super-resolution." + ], + "bbox": [ + 500, + 138, + 893, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 319, + 640, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 344, + 653, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event Camera Reconstruction. Event camera [19, 21, 24] is a kind of neuromorphic camera that monitors relative light intensity changes, making it promising in capturing motion. With the high temporal resolution, reconstructing intensity images with an event camera is an active topic. Extended Kalman Filter [16] is the early attempt at event camera reconstruction, which is based on photometric constancy. Bardow et al. [1] estimate intensity as well as optical flow via the primal-dual algorithm. Inspired by the success of deep learning, E2VID [25] and FireNet [26] were presented with promoting reconstruction performance.", + "bbox": [ + 496, + 368, + 890, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Spike Camera Reconstruction. Spike camera [8, 9, 33, 41, 49] is another type of neuromorphic camera. Different from event cameras, spike camera produces a continuous spike stream to record the dynamic scenes by accumulating photons and firing spike signals. To reconstruct images from spike streams, some exploration [11, 39, 40, 44] has been performed. Zhu et al. [48] proposed to utilize the number of spike signals within a temporal window (TFP) and neighboring spike intervals (TFI). For better performance, some work [37, 43, 47] tried to design reconstruction networks for spike cameras. In addition, there are also attempts at hybrid input [5, 32, 50] for reconstruction. Focusing on the newly invented color spike camera, 3DRI [10] was proposed for scene reconstruction with color information.", + "bbox": [ + 496, + 534, + 893, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Super-Resolution", + "text_level": 1, + "bbox": [ + 500, + 756, + 669, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image and Video Super-Resolution. In the last ten years, many researchers have been devoted to learning to superresolve images. Dong et al. [7] proposed SRNet as the first SR network, bringing great performance gains. In recent years, Liang et al. [18] have achieved significant improvement based on Transformer [28]. Besides the progress in image SR, there are also some video SR work [2, 3, 14, 15] proposed to restore HR frames from LR video frames.", + "bbox": [ + 496, + 779, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24872", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event Camera Super-Resolution. Event camera SR includes upsampling from LR events to HR events [12, 17] and HR image reconstruction from LR events [6, 29, 31]. Li et al. [17] designed a two-stage scheme to solve the spatial SR problem of spatiotemporal events. Then a deep neural framework EventZoom [12] was proposed for event SR. To restore SR images from events directly, Choi et al. [6] proposed a network E2SRI and trained models using synthetic data. In contrast, another deep network EventSR [29] was implemented without ground truth HR images.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spike Camera Super-Resolution. As a trade-off of temporal resolution, the spatial resolution of spike cameras is limited. To restore images with both high temporal and spatial resolution, there is some research about spike camera super-resolution. Zhao et al. [42] exploited relative motion and derived the relationship between light intensity and each spike for super-resolution. Xiang et al. [34] proposed an end-to-end network VidarSR to reconstruct HR images from LR spike streams. Besides, another network SpikeSR-Net [45] is designed based on the observation model of spike camera, achieving state-of-the-art performance.", + "bbox": [ + 75, + 242, + 473, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 419, + 204, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Color Spike Camera", + "text_level": 1, + "bbox": [ + 76, + 444, + 272, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Color spike camera (CSC) is a recently implemented neuromorphic sensor with $H \\times W$ pixels that work independently. To capture color information of dynamic scenes, color filter array is employed on the sensor. As a result, each pixel of the sensor corresponds to one of the three colors, red, green and blue, according to the Bayer-pattern color layout.", + "bbox": [ + 75, + 467, + 470, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 1, CSC is a neuromorphic vision sensor that mimics the structure of human vision. Conventional digital cameras usually use a certain time window for exposure to accumulate photoelectric signals and compact them into a snapshot. In contrast, CSCs accumulate photons and fire spikes continuously to record dynamic scenes, resulting in a Bayer-pattern spike stream. As shown in Fig. 2, every individual pixel denoted as $(x,y)$ within the sensor accumulates photons continuously and converts the instantaneous intensity into an electric signal. When the accumulated electric signals reach a predetermined threshold $\\theta$ , the pixel triggers a flag indicating firing a spike. Then the pixel will be reset and continue to accumulate photons after the flag is checked, starting a new accumulating and firing process. The accumulated electric signal of a certain pixel $(x,y)$ at arbitrary time point $t$ can be formulated as", + "bbox": [ + 75, + 559, + 472, + 801 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} (t, x, y) = \\int_ {0} ^ {t} \\eta \\cdot \\mathbf {I} _ {C} (\\tau , x, y) d \\tau \\mod \\theta , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 806, + 468, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\eta$ denotes the photoelectric conversion rate of the sensor, $\\mathbf{I}_C(\\tau ,x,y)$ denotes the instantaneous intensity of a certain color at time $\\tau$ , and $C\\in \\{R,G,B\\}$ denotes the color determined by the Bayer-pattern color layout.", + "bbox": [ + 75, + 840, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ebe6083a21a8ed0969bbc6a0d959e9ae3035f09fb660a023250dbbfa93702840.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 88, + 697, + 183 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2cdc79f3cf68e88aca59222bec342bf8e8587a4e0872c378444e4c6486230657.jpg", + "image_caption": [ + "Figure 2. Illustration of the spike camera model. (a) The instantaneous intensity of a certain pixel. (b) The accumulated electric signal of the pixel. (c) The number of spikes that have been fired at the pixel. (d) The spike stream by the pixel, controlled by a clock." + ], + "image_footnote": [], + "bbox": [ + 500, + 183, + 700, + 280 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7ef3f81ac81dbaa81432ac5ffd3e8801f95ea1eec331d9fdb85a6ba073551f15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 88, + 890, + 183 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1cfbbc745a72c592aa62958dcd8bcbdd45e173845d65131dcff7bfdff0d28ee7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 184, + 883, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ideally, the pixels on the sensor would be reset immediately after firing a spike. The number of spikes that have been fired before arbitrary time point $t$ can be written as", + "bbox": [ + 496, + 364, + 892, + 410 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi (t, x, y) = \\left\\lfloor \\frac {1}{\\theta} \\int_ {0} ^ {t} \\mathbf {I} _ {C} (\\tau , x, y) d \\tau \\right\\rfloor . \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 416, + 890, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then the read-out spike signal at the $n$ -th $(n \\in \\mathbb{N}^{+})$ frame can be formulated as follows:", + "bbox": [ + 498, + 453, + 890, + 482 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {n} (x, y) = \\left\\{ \\begin{array}{l l} 1, & \\Phi (t _ {n}, x, y) - \\Phi (t _ {n - 1}, x, y) > 0 \\\\ 0, & \\Phi (t _ {n}, x, y) - \\Phi (t _ {n - 1}, x, y) = 0 \\end{array} , \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 487, + 890, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t_n$ denotes the time point according to the $n$ -th spike frame. However, the flags of firing spikes are checked and reset at discrete time points in real hardware implementation, bringing undesired quantization errors. To be specific, CSCs can only check them at discrete time points with a fixed sensor period $T$ , which is controlled by a clock. If the flag of the pixel $(x,y)$ is set up between time points $t = n \\cdot T$ and $t' = (n - 1) \\cdot T$ , we can read out the spike signal $\\mathbf{S}_n(x,y)$ as 1. Otherwise, we have $\\mathbf{S}_n(x,y) = 0$ . After $N$ times of the checking and resetting process, CSC outputs a binary Bayer-pattern spike stream $\\{\\mathbf{S}_i\\}_{i=1}^N$ with a spatial-temporal shape of $N \\times H \\times W$ .", + "bbox": [ + 496, + 529, + 893, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Bayer-Pattern Spike Stream to Raw Images", + "text_level": 1, + "bbox": [ + 498, + 719, + 870, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To reconstruct color images from the Bayer-pattern spike stream, an intuitive way is to convert it to raw images for demosaicing. To infer the raw images, we can employ some spike camera reconstruction methods that don't fuse values of spatial neighboring pixels of multiple colors, such as the spike interval-based method (TFI [48]). Specifically, the spike interval of two temporally neighboring spikes that covers the $n$ -th spike frame can be formulated as", + "bbox": [ + 496, + 742, + 893, + 863 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Psi_ {n} (x, y) = \\min \\left\\{k \\mid \\mathbf {S} _ {i} (x, y) = 1, k \\geq n \\right\\} \\tag {4} \\\\ - \\max \\{k | \\mathbf {S} _ {i} (x, y) = 1, k < n \\}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 869, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24873", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9d2066606d3f5164ee5e339268ef64638f64e04f2a88bf4a5cfd9f1bebfb20a9.jpg", + "image_caption": [ + "Figure 3. The overall architecture of CSCSR network. First, the Bayer-pattern spike stream representation (BSSR) is employed to represent the spike stream clip $\\{\\mathbf{S}_i\\}_{i=1}^N$ . Then the features are passed to the motion-guided super-resolution (MSR) module based on the temporal neighboring pixels of each color. Finally, the HR color image is restored by our color correlation-based reconstruction (CCR) module." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 890, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then the $n$ -th Bayer-pattern raw image can be estimated according to the intervals, which can be written as", + "bbox": [ + 76, + 335, + 470, + 364 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} _ {n} (x, y) = \\frac {\\delta}{\\Psi_ {n} (x , y)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 371, + 468, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\delta$ denotes the maximum dynamic range. As to a spike number-based method (TFP [48]), the raw image $\\mathbf{R}_n(x,y)$ can also be inferred based on the spike number within a temporal window with an odd window length $w$ :", + "bbox": [ + 75, + 406, + 468, + 465 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} _ {n} (x, y) = \\frac {\\delta}{w} \\sum_ {i = n + 1} ^ {n + w} \\mathbf {S} _ {i - \\frac {w + 1}{2}} (x, y). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 472, + 468, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 518, + 166, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 76, + 544, + 272, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To reconstruct HR color images from LR Bayer-pattern spike stream, we develop a CSCSR network whose overall architecture is shown in Fig. 3. The input of the network is a clip of the spike stream with a shape of $N \\times H \\times W$ , denoted as $\\{\\mathbf{S}_i\\}_{i=1}^N$ . To represent color and motion information within the binary input, we first pass it to our proposed Bayer-pattern spike stream representation (BSSR) module. Considering the motion consistency of color channels, we split out four sequences according to the color layout and jointly estimate sub-pixel level motion from them. Guided by the motion information, temporal neighboring pixels of each color channel are searched in the motion-guided super-resolution (MSR) module, resulting in HR temporal-pixel features. Finally, the temporal-pixel features of the three color channels are fused to reconstruct the final HR image, exploiting the correlation of color channels.", + "bbox": [ + 75, + 566, + 470, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Bayer-Pattern Spike Stream Representation", + "text_level": 1, + "bbox": [ + 76, + 816, + 449, + 833 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The spike signals in the Bayer-pattern spike stream mean the arrival of a certain photon amount for a certain color channel, which may suffer from quantization errors as introduced in Sec. 3.1. Considering the motion and color", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "information within the binary data, how to represent it appropriately is worth studying. As shown in Fig. 4, we propose a Bayer-pattern spike stream representation module to represent the information contained in the data.", + "bbox": [ + 496, + 335, + 890, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To keep the color information, we first perform elementwise multiplication between each frame of the spike sequence and the mask of each color, splitting the data into three sequences. For the sequence of each color channel, we develop an encoder to extract the temporal features and retain the relative motion. To be specific, we employ a sliding temporal window to extract the features of each time point. As the impact of quantization errors can be reduced by more temporal information, we extract the features from the whole clip serving as an additional global perception (GP) to the features of each time point. With the temporal window sliding, features of each color are obtained.", + "bbox": [ + 496, + 396, + 892, + 577 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d26537fb7a69453c9d7ba8d807d62dcfa7f308d35a986bf954547a1ff56cb4d4.jpg", + "image_caption": [ + "Figure 4. Illustration of the Bayer-pattern spike stream representation module, with a temporal sliding window in encoders. In the figure, \"ConvBlock\" denotes a $3 \\times 3$ convolution layer followed by a ReLU activation function. \"ResBlock\" refers to [13]." + ], + "image_footnote": [], + "bbox": [ + 501, + 587, + 890, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Motion-Guided Super-Resolution", + "text_level": 1, + "bbox": [ + 498, + 756, + 792, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Joint Motion Estimation. Though the spatial pixels are limited, some pixels from the temporal domain are available for SR, which makes it essential to analyze the motion within the Bayer-pattern spike clip. Therefore, we employ the spike camera optical flow method Spike2Flow [46]. To apply the method to the Bayer-pattern spike stream, we split it into four single-channel spike sequences and estimate motion from each channel as shown in Fig. 3, resulting in op", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24874", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/435f4ea8aea956ffc3632b0f61cdc700f235c00a85beea121fb16237fb6ec76a.jpg", + "image_caption": [ + "Figure 5. Illustration of the motion-guided super-resolution (MSR) module, taking the blue channel as an example. (a) Overview of the MSR module. (b) Motion-guided temporal pixel search, with three steps. (c) Structure of the feature refinement (FR) module." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 566, + 238 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/308f45367f4748e58496a9bcc7a9b471fd8f72f08647cc6a9fddcad456fd6023.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 88, + 890, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tical flows $\\{\\mathbf{w}_{j,k}^{C}\\}, C \\in \\{R, G_{1}, G_{2}, B\\}$ , from the middle time point $j$ to the last time point $k$ . By reversing the spike sequences, we then estimate optical flows $\\{\\mathbf{w}_{j,i}^{C}\\}$ from the middle time point $j$ to the first time point $i$ . Though coming from different intensities, the color channels of the dynamic scene share the same motion. For more robust motion estimation, we jointly consider the color channels and obtain the final sub-pixel level optical flows as", + "bbox": [ + 75, + 289, + 472, + 411 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {w} _ {j, i}, \\mathbf {w} _ {j, k}\\right) = \\left(\\frac {1}{4} \\sum_ {C} \\mathbf {w} _ {j, i} ^ {C}, \\frac {1}{4} \\sum_ {C} \\mathbf {w} _ {j, k} ^ {C}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 420, + 468, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal Pixel Search. To provide features for the final reconstruction of the middle time point, we propose to fill the gap of spatial pixels from LR to HR based on the known temporal pixels at non-middle time points, considering the motion and color layout within the spike stream clip.", + "bbox": [ + 75, + 465, + 468, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The positions of known spatial pixels of a certain color can be inferred according to the Bayer-pattern color mask. Denoting the SR scale as $s$ , the target spatial resolution is $sH \\times sW$ . In the target, one originally known spatial position corresponds to $s^2$ neighboring positions, resulting in an upsampled color mask as shown in Fig. 5 (b). To get the extra known pixels after upsampling, we apply bilinear interpolation to the features, resulting in the super-resolved features $\\mathbf{F}$ . Utilizing the estimated optical flows $(\\mathbf{w}_{j,i}, \\mathbf{w}_{j,k})$ , we can infer the new spatial position of every pixel at the middle time point $j$ when moving to a certain time point $t$ . To be specific, the optical flow $\\mathbf{w}_{j,i}$ can be formulated as the pixel offset along $x$ direction and $y$ direction, which can be written as $[\\mathbf{U}_j^i, \\mathbf{V}_j^i]$ . Considering the spatial upsampling, the offsets should be multiplied by the scale $s$ . Since the time span from time point $i$ or $k$ to $j$ is extremely small $(0.025\\mathrm{ms} \\times N)$ , we assume the motion is linear. As a result, after moving to a certain time point $t$ , the new position of a certain original position $(x,y)$ at the middle time point $j$ can be written as follows:", + "bbox": [ + 75, + 542, + 472, + 843 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} \\left(x + s \\cdot \\frac {t - j}{i - j} \\mathbf {U} _ {j} ^ {i} (x, y), y + s \\cdot \\frac {t - j}{i - j} \\mathbf {V} _ {j} ^ {i} (x, y)\\right), & t < j, \\\\ \\left(x + s \\cdot \\frac {t - j}{k - j} \\mathbf {U} _ {j} ^ {k} (x, y), y + s \\cdot \\frac {t - j}{k - j} \\mathbf {V} _ {j} ^ {k} (x, y)\\right), & t > j. \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 848, + 468, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "According to Eqn. (8), we can map all the pixels at the middle time point to their new positions at any other time point. However, the new positions are usually sub-pixel. Thus we are going to reconstruct the temporal feature value at the sub-pixel position by the values of known integer-pixels according to the upsampled color mask. In particular, the value of the new sub-pixel position $(x_{m},y_{m})$ at time point $t$ can be calculated based on the 4 nearest known pixels $\\{(x_i,y_i)\\}_{i = 1}^4$ , which can be formulated as follows:", + "bbox": [ + 496, + 290, + 890, + 428 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} (t, x, y) = \\sum_ {i = 1} ^ {4} \\frac {\\phi_ {i}}{\\sum_ {i = 1} ^ {4} \\phi_ {i}} \\mathbf {F} (t, x _ {i}, y _ {i}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 440, + 890, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{T}$ denotes the temporal-pixel features, $(x,y)$ denotes the original position at the middle time point, and $\\phi_{i}$ denotes the weight of each known value, which can be written as", + "bbox": [ + 498, + 489, + 890, + 535 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {i} = \\frac {1}{\\sqrt {\\left(x _ {m} - x _ {i}\\right) ^ {2} + \\left(y _ {m} - y _ {i}\\right) ^ {2}}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 546, + 890, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the temporal pixel search process, we can obtain the temporal-pixel features of each color channel for the following HR reconstruction of the middle time point.", + "bbox": [ + 496, + 590, + 890, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Refinement. In most cases, there are errors in the estimated motion, causing an undesired impact on the temporal pixel search. Therefore, we propose a feature refinement as shown in Fig. 5 (c). By averaging the input temporal-pixel features $\\mathbf{T}$ temporally, a rough temporal average result of the channel as a reference can be obtained. Then we compute residuals between the reference and $\\mathbf{T}$ to represent the pixel fluctuation of the frames. At last, we extract deeper features from the residuals and the based $\\mathbf{T}$ , and add them together for refined temporal-pixel features $\\hat{\\mathbf{T}}$ .", + "bbox": [ + 496, + 637, + 890, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Color Correlation-Based Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 800, + 846, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the temporal-pixel features, we are going to reconstruct the final HR image. Despite having different intensities, there is a correlation between color channels. Thus we can restore the image by integrating the features, achieving mutual information complementation of color channels.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24875", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/13975d8102b76e23a5899aa043b8995970dc4634aa4dee4f39bda9477b092218.jpg", + "image_caption": [ + "Figure 6. Illustration of three structures of the reconstruction module. (a) Our proposed color correlation-based reconstruction (CCR) module. (b) A reconstruction module considering no color correlation. (c) An intuitive design considering color correlation." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 467, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 6 (a), we design a reconstruction module utilizing multi-stage features of color channels. In particular, we first apply residual blocks to extract deeper features from the features of each color channel. To fuse the multi-color features of each stage for better aggregation, we develop a color fusion (CF) module to encode the features of color channels with the output of the previous CF stage. After several stages of CF, we employ residual blocks in the end to generate the final image from the fused features.", + "bbox": [ + 75, + 329, + 468, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 477, + 209, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Color Spike Camera Simulator", + "text_level": 1, + "bbox": [ + 76, + 501, + 352, + 517 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eff25f3a2bcd92b935d9d2498c874409453d61b66e9d8fb041ab4f0c6ad94dbc.jpg", + "image_caption": [ + "Figure 7. The pipeline of the color spike camera simulator, with the green channel as an example of signal accumulation." + ], + "image_footnote": [], + "bbox": [ + 81, + 535, + 467, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To train models, a large amount of LR spike stream-HR color image pairs are required. However, it's hard to collect corresponding high-quality HR images, especially in scenes with high-speed motion. Inspired by the success of gray-scale spike camera simulators [43, 45], we also developed a simulator for color spike cameras. The pipeline of our simulator is shown in Fig. 7. To generate spike streams, we regard the input video as the dynamic scene to be recorded. As the frame rate of the input video is usually limited, the temporal information is not enough for the simulator of CSC with ultra-high temporal resolution to generate Bayer-pattern spike streams. To address this issue, we", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "use a frame interpolation method [27] to generate latent intensity frames between the original video frames. To generate spike signals, we follow the mechanism introduced in Sec. 3.1 to accumulate light intensity from the latent frames and compare the accumulated value with the threshold periodically for firing spikes. As a result, a Bayer-pattern spike stream is generated from the input video. In particular, the pixel count used to accumulate intensity signals is determined by SR scales. For example, we accumulate signals of 4 pixels to generate the signal of one target pixel in the $\\times 2$ case. Considering the CFA, the 4 pixels come from one of the channels according to the Bayer-pattern color layout.", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 280, + 705, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. In our implementation, we randomly crop the Bayer-pattern spike frames into patches with spatial resolution $96 \\times 96$ for training, with the RGGB Bayer pattern. We set the number of spike frames $N$ to 41 and the batch size to 4. The length $w$ and stride $s$ of the sliding temporal window are set to 11 and 3. We employ $\\ell_{1}$ loss function and use Adam optimizer with an initial learning rate of $10^{-4}$ , which will decay to 0.8 times after every 7500 iterations. Besides, we employ the PyTorch framework and train the models via an NVIDIA RTX3090 GPU.", + "bbox": [ + 496, + 303, + 890, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training and Evaluation Datasets. Using our developed CSC simulator, we generated the training dataset based on 240 scenes for training from REDS (120fps) [23], resulting in $5 \\times 240 = 1200$ clips for each super-resolution scale ( $\\times 2$ , $\\times 3$ and $\\times 4$ ). Then we generate $5 \\times 30 = 150$ clips for evaluation based on 30 videos of the REDS evaluation dataset. To demonstrate the generalization to non-REDS-based data, we also generated evaluation datasets from two video SR datasets, Vid4 [22] and Viceo [35]. Besides, we also captured some real-world Bayer-pattern spike streams to verify the performance further. The details of the synthetic evaluation datasets can be found in Table 1.", + "bbox": [ + 496, + 454, + 890, + 633 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/70f12afecdadfbde52c6ca7df6b9955eed1e5af1630e2575880bbb299e0b8364.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SourceScenesSamplesResolutionScales
REDS [23]30150720×12802, 3, 4
Vid4 [22]44-2, 3, 4
Vimeo [35]78247824448×2564
", + "bbox": [ + 506, + 645, + 885, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Details of the three synthetic evaluation datasets.", + "bbox": [ + 522, + 713, + 864, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison Methods. As there are no SR methods for CSCs, we first choose the state-of-the-art gray-scale spike camera SR networks SpikeSR-Net [45] and VidarSR [34] for comparison. To adapt the gray-scale methods to the task, we adopt two strategies: using Bayer-pattern spike streams for end-to-end training and handling each color channel separately. We denote the networks with the second strategy as SpikeSR-Net* and VidarSR*. Then we combine a CSC reconstruction method 3DRI [10] and an image/video SR method SwinIR [18]/BasicVSR [3], resulting in another two comparison methods 3DRI+SwinIR and 3DRI+BasicVSR.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24876", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c3d8ad461bb41a5d7b3f8dd15a40223401727da929a26e501642fd0eeb89aa2d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScaleMethodREDS-based DatasetVid4-based DatasetVimeo-based Dataset
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
×2TFI [48] +TSCNN [4]29.24dB0.81730.312625.06dB0.77860.3196---
TFP[48] +TSCNN27.43dB0.75650.367124.60dB0.76660.3142---
TFI+Real-RawVSR [36]31.03dB0.86270.247826.79dB0.84220.2508---
TFP+Real-RawVSR30.81dB0.85350.260127.10dB0.85640.2329---
3DRI [10] +SwinIR [18]31.13dB0.88090.219126.57dB0.86680.2100---
3DRI+BasicVSR [3]31.60dB0.89120.205227.03dB0.87480.2021---
VidarSR [34]30.81dB0.89910.188426.44dB0.89110.1791---
VidarSR*30.27dB0.84300.263125.28dB0.81060.2538---
SpikeSR-Net [45]32.38dB0.89130.209928.34dB0.89570.1833---
SpikeSR-Net*29.66dB0.82490.286124.62dB0.78540.2721---
CSCSR (ours)33.39dB0.91210.176428.77dB0.91100.1612---
×3TFI+TSCNN27.22dB0.75050.384222.91dB0.66090.4058---
TFP+TSCNN26.55dB0.71760.413723.22dB0.68000.3808---
TFI+Real-RawVSR28.68dB0.79420.329224.00dB0.72250.3474---
TFP+Real-RawVSR28.66dB0.78830.334424.39dB0.74070.3326---
3DRI+SwinIR28.38dB0.81360.297723.97dB0.75340.3160---
3DRI+BasicVSR28.65dB0.82080.290524.08dB0.75680.3135---
VidarSR29.36dB0.84140.267924.66dB0.80090.2816---
VidarSR*27.53dB0.76700.355422.78dB0.67420.3680---
SpikeSR-Net29.79dB0.82700.297025.33dB0.79150.2989---
SpikeSR-Net*26.73dB0.73670.390822.12dB0.63230.4067---
CSCSR (ours)29.92dB0.85130.265325.68dB0.82400.2761---
×4TFI+TSCNN25.98dB0.70100.431021.69dB0.57220.465924.77dB0.73040.3554
TFP+TSCNN26.55dB0.71760.413722.53dB0.62400.417725.56dB0.75870.3557
TFI+Real-RawVSR27.27dB0.73910.385122.53dB0.62400.417725.12dB0.74920.3244
TFP+Real-RawVSR27.32dB0.73470.388522.81dB0.64080.363425.15dB0.74930.3324
3DRI+SwinIR26.94dB0.75440.360222.47dB0.65330.390125.21dB0.74230.3218
3DRI+BasicVSR26.94dB0.75700.360522.52dB0.65070.397325.05dB0.74010.3204
VidarSR27.56dB0.76920.350522.69dB0.67120.386825.19dB0.75650.3098
VidarSR*26.51dB0.71350.413521.58dB0.58800.439524.64dB0.73520.3423
SpikeSR-Net28.28dB0.77550.351723.59dB0.69830.372725.27dB0.76110.3119
SpikeSR-Net*25.93dB0.69130.440321.01dB0.54510.471824.46dB0.72400.3593
CSCSR (ours)28.77dB0.79630.325323.79dB0.72230.355525.34dB0.76600.2967
", + "bbox": [ + 83, + 89, + 888, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison. Red and blue indicate the best and the second-best performance, respectively. For each scene of Video [35], there is a sequence of LR frames but only one corresponding HR frame $(\\times 4)$ . Thus only the dataset of $\\times 4$ scale is generated for Video.", + "bbox": [ + 76, + 584, + 892, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It's also intuitive to combine methods of estimating raw images from Bayer-pattern spike streams (Sec. 3.2) and joint demosaicing and super-resolution (JDSR) methods. Thus we employ TFI or TFP [48] to infer raw images, followed by an image JDSR method TSCNN [4] or a video JDSR method Real-RawVSR [36] to reconstruct color images. In addition, we employ PSNR, SSIM [30] and LPIPS [38] (lower means better) as quantitative metrics for comparison.", + "bbox": [ + 75, + 628, + 470, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Comparative Results", + "text_level": 1, + "bbox": [ + 76, + 756, + 274, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. As shown in Table 2, we conduct comparative experiments on the three SR scales. It can be found that our method achieves the best quantitative results in most experiments. The state-of-the-art spike camera SR methods VidarSR and SpikeSR-Net achieve competitive performance. However, VidarSR* and SpikeSR-Net* that handle each color channel separately don't perform as well as VidarSR and SpikeSR-Net, due to the lack of considera", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion of color correlation. With the temporal information, the multi-frame method 3DRI+BasicVSR performs better than the single-frame method 3DRI+SwinIR in most cases. This is also true for the multi-frame JDSR method RawVSR and the single-frame JDSR method TSCNN. Besides, with the increase of the SR scale, TFP-based methods show growing competitiveness to TFI-based methods.", + "bbox": [ + 496, + 628, + 890, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual Results. Fig. 8 presents the visual comparison $(\\times 4)$ on the synthetic data. The visual quality of results by our proposed method is better than the other methods, with less motion blur and more accurate color. There are artifacts in the results produced by VidarSR* and SpikeSR-Net*, which demonstrates the importance of the consideration of color correlation. In addition, the multi-frame methods show better visual quality. To demonstrate model generalization, we conduct experiments on real-world captured Bayer-pattern spike streams as shown in Fig. 9. In contrast, our method can restore better textures and details.", + "bbox": [ + 496, + 734, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24877", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/162dd7e3ffc318ca8a901b5537b8712e68d331223f48887cc9bbb89a24eabf8e.jpg", + "image_caption": [ + "Figure 8. Visual comparison $(\\times 4)$ on the synthetic Bayer-pattern spike streams. The first sample comes from the REDS-based dataset, while the second one comes from the Vid4-based dataset. Please enlarge the figure for better comparison." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 893, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8fc7be26a481a16f935452ccaa5173c4daf18051e6f369b1dce80f9aa338060d.jpg", + "image_caption": [ + "Figure 9. Visual comparison $(\\times 4)$ on the real-world Bayer-pattern spike streams. The first spike stream sample is captured by a fast-moving CSC. The second one records a fast-rotating water bottle. Please enlarge the figure for better comparison." + ], + "image_footnote": [], + "bbox": [ + 78, + 282, + 890, + 448 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bde87938c9df89e0bb21ea90c2d542a6b8b99037d1414685da478fb6455ba7f9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CaseSetting Description×2×3×4
(A)Removing BSSR module33.0429.5728.57
(B)Removing GP from BSSR33.1729.7428.68
(C)Interval-based representation32.0528.1928.06
(D)Independent motion estimation33.2229.5528.59
(E)Assembled motion estimation30.2427.0826.66
(F)Removing MSR module32.6629.4228.39
(G)Removing FR from MSR33.0729.7028.56
(H)Reconstruction module (b)30.1927.3626.22
(I)Reconstruction module (c)33.2429.6928.66
(J)Our final network33.3929.9228.77
", + "bbox": [ + 84, + 493, + 467, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study on the REDS-based evaluation dataset. Greener blocks represent higher PSNR(dB) performance.", + "bbox": [ + 76, + 654, + 468, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 696, + 230, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To verify the effectiveness of our methodology, we perform ablation studies on our modules and motion estimation strategy. Case (J) is our final CSCSR network. First, we remove the BSSR module in (A) to demonstrate the proposed representation. Then we remove the additional GP branch from BSSR in (B). We also replace BSSR with the interval-based representation [46, 48] in (C). To demonstrate our joint motion estimation strategy, we employ independent motion estimation for each color channel in (D) and an optical flow assembly strategy in (E). To be specific, the optical flows of color channels are assembled according to the Bayer pattern. After that, we implement studies on the MSR", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "module. In (F), MSR is removed from the network. In (G), the FR module is removed from MSE to verify its effectiveness. Finally, (H) and (I) are the studies on the proposed reconstruction module, where the module is replaced by the structures in (b) and (c) of Fig. 6, with the same number of residual blocks. The study results are shown in Table 3.", + "bbox": [ + 496, + 493, + 893, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 498, + 595, + 627, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a deep network to restore HR color images from LR Bayer-pattern spike streams. To represent the binary data, we develop a representation that utilizes local temporal information with global perception. Then we propose to collect temporal pixels for spatial super-resolution according to the color layout and sub-pixel level motion. To reduce the impact of motion estimation errors, we design a feature refinement module based on residuals. Finally, multi-stage temporal-pixel features of each color channel are jointly considered, resulting in the final HR color image. Experimental results on both synthetic and real-world captured data demonstrate our performance.", + "bbox": [ + 496, + 621, + 893, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 815, + 658, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grants 62072009, 22127807, and in part by the National Key R&D Program of China under Grant 2021YFF0900501.", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24878", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Patrick Bardow, Andrew J Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 884-892, 2016. 2", + "[2] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Real-time video super-resolution with spatio-temporal networks and motion compensation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4778-4787, 2017. 2", + "[3] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. BasicVSR: The search for essential components in video super-resolution and beyond. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4947-4956, 2021. 2, 6, 7", + "[4] Kan Chang, Hengxin Li, Yufei Tan, Pak Lun Kevin Ding, and Baoxin Li. A two-stage convolutional neural network for joint demosaicking and super-resolution. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4238-4254, 2021. 7", + "[5] Yakun Chang, Chu Zhou, Yuchen Hong, Liwen Hu, Chao Xu, Tiejun Huang, and Boxin Shi. 1000 FPS HDR video with a spike-rgb hybrid camera. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 22180-22190, 2023. 2", + "[6] Jonghyun Choi, Kuk-Jin Yoon, et al. Learning to super resolve intensity images from events. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2768-2776, 2020. 3", + "[7] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Proceedings of European Conference on Computer Vision (ECCV), 2014. 2", + "[8] Siwei Dong, Tiejun Huang, and Yonghong Tian. Spike camera and its coding methods. In Data Compression Conference (DCC), page 437, 2017. 2", + "[9] Siwei Dong, Lin Zhu, Daoyuan Xu, Yonghong Tian, and Tiejun Huang. An efficient coding method for spike camera using inter-spike intervals. In Data Compression Conference (DCC), page 568, 2019. 2", + "[10] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. 3D residual interpolation for spike camera demosaicing. In 2022 IEEE International Conference on Image Processing (ICIP), pages 1461-1465. IEEE, 2022. 2, 6, 7", + "[11] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. High-speed scene reconstruction from low-light spike streams. In 2022 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2022. 2", + "[12] Peiqi Duan, Zihao W Wang, Xinyu Zhou, Yi Ma, and Boxin Shi. EventZoom: Learning to denoise and super resolve neuromorphic events. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12824-12833, 2021. 3", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 4", + "[14] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3224-3232, 2018. 2", + "[15] Armin Kappeler, Seunghwan Yoo, Qiqin Dai, and Aggelos K Katsaggelos. Video super-resolution with convolutional neural networks. IEEE Transactions on Computational Imaging, 2(2):109-122, 2016. 2", + "[16] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew J Davison. Simultaneous mosaicing and tracking with an event camera. J. Solid State Circ, 43:566-576, 2008. 2", + "[17] Hongmin Li, Guoqi Li, and Luping Shi. Super-resolution of spatiotemporal event-stream image. Neurocomputing, 335: 206-214, 2019. 3", + "[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 2, 6, 7", + "[19] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \\times 128$ 120 db $15\\mu s$ latency asynchronous temporal contrast vision sensor. IEEE journal of solid-state circuits, 43 (2):566-576, 2008. 2", + "[20] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In IEEE conference on computer vision and pattern recognition workshops (CVPRW), pages 136-144, 2017. 2", + "[21] Martin Litzenberger, Christoph Posch, D Bauer, Ahmed Nabil Belbachir, P Schon, B Kohn, and H Garn. Embedded vision system for real-time object tracking using an asynchronous transient vision sensor. In 2006 IEEE 12th Digital Signal Processing Workshop & 4th IEEE Signal Processing Education Workshop, pages 173-178. IEEE, 2006. 2", + "[22] Ce Liu and Deqing Sun. On bayesian adaptive video super resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(2):346-360, 2013. 6", + "[23] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. NTIRE 2019 challenge on video deblurring and superresolution: Dataset and study. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 0–0, 2019. 6", + "[24] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2010. 2", + "[25] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3857-3866, 2019. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24879", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast image reconstruction with an event camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 156-163, 2020. 2", + "[27] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. XVFI: extreme video frame interpolation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14489-14498, 2021. 6", + "[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[29] Lin Wang, Tae-Kyun Kim, and Kuk-Jin Yoon. EventSR: From asynchronous events to image reconstruction, restoration, and super-resolution via end-to-end adversarial learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8315-8325, 2020. 3", + "[30] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 7", + "[31] Zihao W. Wang, Peiqi Duan, Oliver Cossairt, Aggelos Katsaggelos, Tiejun Huang, and Boxin Shi. Joint filtering of intensity images and neuromorphic events for high-resolution noise-robust imaging. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1609–1619, 2020. 3", + "[32] Lujie Xia, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. SVFI: Spiking-based video frame interpolation for high-speed motion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2910-2918, 2023. 2", + "[33] Lujie Xia, Ziluo Ding, Rui Zhao, Jiyuan Zhang, Lei Ma, Zhaofei Yu, Tiejun Huang, and Ruiqin Xiong. Unsupervised optical flow estimation with dynamic timing representation for spike camera. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[34] Xijie Xiang, Lin Zhu, Jianing Li, Yixuan Wang, Tiejun Huang, and Yonghong Tian. Learning super-resolution reconstruction for high temporal resolution spike stream. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 3, 6, 7", + "[35] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. Int. J. Comput. Vis, 127:1106-1125, 2019. 6, 7", + "[36] Huanjing Yue, Zhiming Zhang, and Jingyu Yang. Realrawvsr: Real-world raw video super-resolution with a benchmark dataset. In European Conference on Computer Vision, pages 608-624. Springer, 2022. 7", + "[37] Jiyuan Zhang, Shanshan Jia, Zhaofei Yu, and Tiejun Huang. Learning temporal-ordered representation for spike streams based on discrete wavelet transforms. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 137–147, 2023. 2", + "[38] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 7", + "[39] Yiyang Zhang, Ruiqin Xiong, and Tiejun Huang. Spike signal reconstruction based on inter-spike similarity. In 2022 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2022. 2", + "[40] Jing Zhao, Ruiqin Xiong, and Tiejun Huang. High-speed motion scene reconstruction for spike camera via motion aligned filtering. In IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-5, 2020. 2", + "[41] Jing Zhao, Ruiqin Xiong, Rui Zhao, Jin Wang, Siwei Ma, and Tiejun Huang. Motion estimation for spike camera data sequence via spike interval analysis. In 2020 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 371-374. IEEE, 2020. 2", + "[42] Jing Zhao, Jiyu Xie, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, and Tiejun Huang. Super resolve dynamic scene from continuous spike streams. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2533-2542, 2021. 3", + "[43] Jing Zhao, Ruiqin Xiong, Hangfan Liu, Jian Zhang, and Tiejun Huang. Spk2ImgNet: Learning to reconstruct dynamic scene from continuous spike stream. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 11996-12005, 2021. 2, 6", + "[44] Jing Zhao, Ruiqin Xiong, Jiyu Xie, Boxin Shi, Zhaofei Yu, Wen Gao, and Tiejun Huang. Reconstructing clear image for high-speed motion scene with a retina-inspired spike camera. IEEE Transactions on Computational Imaging, 8:12-27, 2021. 2", + "[45] Jing Zhao, Ruiqin Xiong, Jian Zhang, Rui Zhao, Hangfan Liu, and Tiejun Huang. Learning to super-resolve dynamic scenes for neuromorphic spike camera. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3579-3587, 2023. 3, 6, 7", + "[46] Rui Zhao, Ruiqin Xiong, Jing Zhao, Zhaofei Yu, Xiaopeng Fan, and Tiejun Huang. Learning optical flow from continuous spike streams. Advances in Neural Information Processing Systems, 35:7905-7920, 2022. 4, 8", + "[47] Rui Zhao, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, Shuyuan Zhu, Lei Ma, and Tiejun Huang. Spike camera image reconstruction using deep spiking neural networks. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2", + "[48] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. A retina-inspired sampling method for visual texture reconstruction. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1432-1437. IEEE, 2019. 2, 3, 4, 7, 8", + "[49] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. Hybrid coding of spatiotemporal spike data for a bio-inspired camera. IEEE Transactions on Circuits and Systems for Video Technology, 31(7):2837-2851, 2021. 2", + "[50] Lin Zhu, Jianing Li, Xiao Wang, Tiejun Huang, and Yonghong Tian. NeuSpike-Net: High speed video reconstruction via bio-inspired neuromorphic cameras. In IEEE International Conference on Computer Vision (ICCV), pages 2400–2409, 2021. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24880", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/ac654108-cae7-444a-a7e8-a0faf89a682d_model.json b/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/ac654108-cae7-444a-a7e8-a0faf89a682d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dbb1a0b881a29916ad284533c956661067304e93 --- /dev/null +++ b/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/ac654108-cae7-444a-a7e8-a0faf89a682d_model.json @@ -0,0 +1,2057 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.14, + 0.131, + 0.83, + 0.154 + ], + "angle": 0, + "content": "Super-Resolution Reconstruction from Bayer-Pattern Spike Streams" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.179, + 0.93, + 0.288 + ], + "angle": 0, + "content": "Yanchen Dong\\(^{1}\\), Ruiqin Xiong\\(^{1*}\\), Jian Zhang\\(^{2}\\), Zhaofei Yu\\(^{3}\\), Xiaopeng Fan\\(^{4}\\), Shuyuan Zhu\\(^{5}\\), Tiejun Huang\\(^{1,3}\\) \n\\(^{1}\\)School of Computer Science, Peking University \n\\(^{2}\\)School of Electronic and Computer Engineering, Peking University \n\\(^{3}\\)Institute for Artificial Intelligence, Peking University \n\\(^{4}\\)School of Computer Science and Technology, Harbin Institute of Technology \n\\(^{5}\\)University of Electronic Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.29, + 0.86, + 0.324 + ], + "angle": 0, + "content": "yanchendong@stu.pku.edu.cn, {rqxiong, zhangjian.sz, yuzf12, tjhuang}@pku.edu.cn, fxp@hit.edu.cn, eezsy@uestc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.357, + 0.314, + 0.373 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.389, + 0.474, + 0.798 + ], + "angle": 0, + "content": "Spike camera is a neuromorphic vision sensor that can capture highly dynamic scenes by generating a continuous stream of binary spikes to represent the arrival of photons at very high temporal resolution. Equipped with Bayer color filter array (CFA), color spike camera (CSC) has been invented to capture color information. Although spike camera has already demonstrated great potential for high-speed imaging, its spatial resolution is limited compared with conventional digital cameras. This paper proposes a Color Spike Camera Super-Resolution (CSCSR) network to super-resolve higher-resolution color images from spike camera streams with Bayer CFA. To be specific, we first propose a representation for Bayer-pattern spike streams, exploring local temporal information with global perception to represent the binary data. Then we exploit the CFA layout and sub-pixel level motion to collect temporal pixels for the spatial super-resolution of each color channel. In particular, a residual-based module for feature refinement is developed to reduce the impact of motion estimation errors. Considering color correlation, we jointly utilize the multi-stage temporal-pixel features of color channels to reconstruct the high-resolution color image. Experimental results demonstrate that the proposed scheme can reconstruct satisfactory color images with both high temporal and spatial resolution from low-resolution Bayer-pattern spike streams. The source codes are available at https://github.com/csycdong/CSCSR." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.824, + 0.21, + 0.839 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.849, + 0.47, + 0.88 + ], + "angle": 0, + "content": "With the quick development of high-speed vision applications such as autonomous driving and unmanned aerial ve" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.355, + 0.891, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.626, + 0.892, + 0.654 + ], + "angle": 0, + "content": "Figure 1. Pipeline of super-resolution reconstruction from Bayer-pattern spike streams. The super-resolution scale is denoted as \\(s\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.893, + 0.84 + ], + "angle": 0, + "content": "hicles, the demand for cameras that can capture high-speed motion and respond quickly is increasing. Most conventional digital cameras use a certain exposure time window to accumulate photoelectric signals and create a snapshot image, which requires the scene to be still during the exposure interval. Due to the exposure time, conventional cameras suffer from motion blur when capturing high-speed scenes. To be specific, a given point of moving objects can move during exposure time, bringing undesired motion blur to the image. Therefore, it's hard for conventional digital cameras to meet the requirements of high-speed imaging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Mimicking the structure of human vision, neuromorphic cameras with ultra-high temporal resolution (e.g., \\(20,000\\mathrm{Hz}\\)) show great potential in capturing high-speed scenes. A well-known type of neuromorphic camera called" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24871" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "event camera [19, 21, 24] is sensitive to motion in dynamic scenes, which can record relative light intensity changes. To record more texture details, another type of neuromorphic camera called spike camera [8, 9, 49] captures absolute light intensity via an \"integrate-and-fire\" mechanism. As the first-generation spike camera can only reproduce gray-level signals, color spike camera, which employs the Bayer-pattern color filter array (CFA), has been invented to record dynamic scenes with color information and meanwhile enjoy the benefits of high-speed imaging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.473, + 0.502 + ], + "angle": 0, + "content": "Although capable of capturing color information like conventional cameras, the spatial resolution of color spike cameras is limited compared to recent digital cameras, which is a trade-off for high temporal resolution and low latency. Therefore, we aim to reconstruct high-resolution (HR) color images from low-resolution (LR) Bayer-pattern spike streams, achieving both high temporal and spatial resolution with a color spike camera. An intuitive way is to combine color spike camera reconstruction methods [10] with image super-resolution (SR) algorithms [7, 18, 20]. However, information among temporal neighboring frames that is essential for SR will be degraded in the first reconstruction stage of this scheme. Besides, high-speed motion and quantization noise within the Bayer-pattern spike streams also bring challenges to the SR task. As a result, how to improve the spatial resolution as well as the imaging quality for color spike cameras is worth studying." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.473, + 0.807 + ], + "angle": 0, + "content": "In this paper, we propose a Color Spike Camera SuperResolution (CSCSR) network to super-resolve dynamic scenes from Bayer-pattern spike streams. To represent the binary data with color and motion information, we first develop a Bayer-pattern Spike Stream Representation (BSSR) to explore local temporal information with global perception. As the pixels of a single frame are limited for spatial super-resolution, we propose a Motion-guided SuperResolution (MSR) module to search temporal pixels of each color channel according to the Bayer-pattern color layout and sub-pixel level motion jointly estimated from color channels. In most cases, there are errors in the estimated motion, causing an undesired impact on the temporal pixel search. Therefore, a Feature Refinement (FR) module is designed based on the residual between a temporal average reference and the temporal pixels, resulting in refined temporal-pixel features. Finally, we further extract features and jointly utilize multi-stage features of each color channel to restore the final HR image, achieved by our Color Correlation-based Reconstruction (CCR) module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To train and evaluate our models, we design a simulator based on the mechanism of color spike cameras, which can generate data pairs for SR reconstruction. Besides, we also capture some real-world Bayer-pattern spike streams for further evaluation. Experimental results on both synthetic and real-world captured data show that our method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "can reconstruct HR color images from LR Bayer-pattern spike streams, with good texture details. The main contribution of our work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.14, + 0.894, + 0.2 + ], + "angle": 0, + "content": "- We develop a color spike camera super-resolution network to reconstruct high-resolution color images from low-resolution Bayer-pattern spike streams, which is the first attempt to the best of our knowledge." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.201, + 0.894, + 0.261 + ], + "angle": 0, + "content": "- To reconstruct details beyond the sensor resolution, we explore the sub-pixel level motion of each color channel and propose a motion-guided super-resolution module to collect temporal pixels according to the color layout." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.261, + 0.895, + 0.307 + ], + "angle": 0, + "content": "- Experiments demonstrate that our method outperforms all the existing methods, achieving better quantitative and qualitative results in color spike camera super-resolution." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.14, + 0.895, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.32, + 0.642, + 0.336 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.345, + 0.655, + 0.36 + ], + "angle": 0, + "content": "2.1. Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.892, + 0.534 + ], + "angle": 0, + "content": "Event Camera Reconstruction. Event camera [19, 21, 24] is a kind of neuromorphic camera that monitors relative light intensity changes, making it promising in capturing motion. With the high temporal resolution, reconstructing intensity images with an event camera is an active topic. Extended Kalman Filter [16] is the early attempt at event camera reconstruction, which is based on photometric constancy. Bardow et al. [1] estimate intensity as well as optical flow via the primal-dual algorithm. Inspired by the success of deep learning, E2VID [25] and FireNet [26] were presented with promoting reconstruction performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.894, + 0.747 + ], + "angle": 0, + "content": "Spike Camera Reconstruction. Spike camera [8, 9, 33, 41, 49] is another type of neuromorphic camera. Different from event cameras, spike camera produces a continuous spike stream to record the dynamic scenes by accumulating photons and firing spike signals. To reconstruct images from spike streams, some exploration [11, 39, 40, 44] has been performed. Zhu et al. [48] proposed to utilize the number of spike signals within a temporal window (TFP) and neighboring spike intervals (TFI). For better performance, some work [37, 43, 47] tried to design reconstruction networks for spike cameras. In addition, there are also attempts at hybrid input [5, 32, 50] for reconstruction. Focusing on the newly invented color spike camera, 3DRI [10] was proposed for scene reconstruction with color information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.671, + 0.773 + ], + "angle": 0, + "content": "2.2. Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Image and Video Super-Resolution. In the last ten years, many researchers have been devoted to learning to superresolve images. Dong et al. [7] proposed SRNet as the first SR network, bringing great performance gains. In recent years, Liang et al. [18] have achieved significant improvement based on Transformer [28]. Besides the progress in image SR, there are also some video SR work [2, 3, 14, 15] proposed to restore HR frames from LR video frames." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24872" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "Event Camera Super-Resolution. Event camera SR includes upsampling from LR events to HR events [12, 17] and HR image reconstruction from LR events [6, 29, 31]. Li et al. [17] designed a two-stage scheme to solve the spatial SR problem of spatiotemporal events. Then a deep neural framework EventZoom [12] was proposed for event SR. To restore SR images from events directly, Choi et al. [6] proposed a network E2SRI and trained models using synthetic data. In contrast, another deep network EventSR [29] was implemented without ground truth HR images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.474, + 0.41 + ], + "angle": 0, + "content": "Spike Camera Super-Resolution. As a trade-off of temporal resolution, the spatial resolution of spike cameras is limited. To restore images with both high temporal and spatial resolution, there is some research about spike camera super-resolution. Zhao et al. [42] exploited relative motion and derived the relationship between light intensity and each spike for super-resolution. Xiang et al. [34] proposed an end-to-end network VidarSR to reconstruct HR images from LR spike streams. Besides, another network SpikeSR-Net [45] is designed based on the observation model of spike camera, achieving state-of-the-art performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.42, + 0.205, + 0.437 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.445, + 0.273, + 0.461 + ], + "angle": 0, + "content": "3.1. Color Spike Camera" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.468, + 0.471, + 0.559 + ], + "angle": 0, + "content": "Color spike camera (CSC) is a recently implemented neuromorphic sensor with \\( H \\times W \\) pixels that work independently. To capture color information of dynamic scenes, color filter array is employed on the sensor. As a result, each pixel of the sensor corresponds to one of the three colors, red, green and blue, according to the Bayer-pattern color layout." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.473, + 0.802 + ], + "angle": 0, + "content": "As shown in Fig. 1, CSC is a neuromorphic vision sensor that mimics the structure of human vision. Conventional digital cameras usually use a certain time window for exposure to accumulate photoelectric signals and compact them into a snapshot. In contrast, CSCs accumulate photons and fire spikes continuously to record dynamic scenes, resulting in a Bayer-pattern spike stream. As shown in Fig. 2, every individual pixel denoted as \\((x,y)\\) within the sensor accumulates photons continuously and converts the instantaneous intensity into an electric signal. When the accumulated electric signals reach a predetermined threshold \\(\\theta\\), the pixel triggers a flag indicating firing a spike. Then the pixel will be reset and continue to accumulate photons after the flag is checked, starting a new accumulating and firing process. The accumulated electric signal of a certain pixel \\((x,y)\\) at arbitrary time point \\(t\\) can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.807, + 0.47, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} (t, x, y) = \\int_ {0} ^ {t} \\eta \\cdot \\mathbf {I} _ {C} (\\tau , x, y) d \\tau \\mod \\theta , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\(\\eta\\) denotes the photoelectric conversion rate of the sensor, \\(\\mathbf{I}_C(\\tau ,x,y)\\) denotes the instantaneous intensity of a certain color at time \\(\\tau\\), and \\(C\\in \\{R,G,B\\}\\) denotes the color determined by the Bayer-pattern color layout." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.698, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.184, + 0.702, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.089, + 0.892, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.185, + 0.885, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.292, + 0.894, + 0.35 + ], + "angle": 0, + "content": "Figure 2. Illustration of the spike camera model. (a) The instantaneous intensity of a certain pixel. (b) The accumulated electric signal of the pixel. (c) The number of spikes that have been fired at the pixel. (d) The spike stream by the pixel, controlled by a clock." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.893, + 0.411 + ], + "angle": 0, + "content": "Ideally, the pixels on the sensor would be reset immediately after firing a spike. The number of spikes that have been fired before arbitrary time point \\( t \\) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.417, + 0.892, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\Phi (t, x, y) = \\left\\lfloor \\frac {1}{\\theta} \\int_ {0} ^ {t} \\mathbf {I} _ {C} (\\tau , x, y) d \\tau \\right\\rfloor . \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.454, + 0.892, + 0.483 + ], + "angle": 0, + "content": "Then the read-out spike signal at the \\(n\\)-th \\((n \\in \\mathbb{N}^{+})\\) frame can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.488, + 0.892, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {n} (x, y) = \\left\\{ \\begin{array}{l l} 1, & \\Phi (t _ {n}, x, y) - \\Phi (t _ {n - 1}, x, y) > 0 \\\\ 0, & \\Phi (t _ {n}, x, y) - \\Phi (t _ {n - 1}, x, y) = 0 \\end{array} , \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.895, + 0.712 + ], + "angle": 0, + "content": "where \\( t_n \\) denotes the time point according to the \\( n \\)-th spike frame. However, the flags of firing spikes are checked and reset at discrete time points in real hardware implementation, bringing undesired quantization errors. To be specific, CSCs can only check them at discrete time points with a fixed sensor period \\( T \\), which is controlled by a clock. If the flag of the pixel \\( (x,y) \\) is set up between time points \\( t = n \\cdot T \\) and \\( t' = (n - 1) \\cdot T \\), we can read out the spike signal \\( \\mathbf{S}_n(x,y) \\) as 1. Otherwise, we have \\( \\mathbf{S}_n(x,y) = 0 \\). After \\( N \\) times of the checking and resetting process, CSC outputs a binary Bayer-pattern spike stream \\( \\{\\mathbf{S}_i\\}_{i=1}^N \\) with a spatial-temporal shape of \\( N \\times H \\times W \\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.72, + 0.871, + 0.736 + ], + "angle": 0, + "content": "3.2. Bayer-Pattern Spike Stream to Raw Images" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.743, + 0.894, + 0.864 + ], + "angle": 0, + "content": "To reconstruct color images from the Bayer-pattern spike stream, an intuitive way is to convert it to raw images for demosaicing. To infer the raw images, we can employ some spike camera reconstruction methods that don't fuse values of spatial neighboring pixels of multiple colors, such as the spike interval-based method (TFI [48]). Specifically, the spike interval of two temporally neighboring spikes that covers the \\( n \\)-th spike frame can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.871, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Psi_ {n} (x, y) = \\min \\left\\{k \\mid \\mathbf {S} _ {i} (x, y) = 1, k \\geq n \\right\\} \\tag {4} \\\\ - \\max \\{k | \\mathbf {S} _ {i} (x, y) = 1, k < n \\}. \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24873" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.891, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.275, + 0.893, + 0.318 + ], + "angle": 0, + "content": "Figure 3. The overall architecture of CSCSR network. First, the Bayer-pattern spike stream representation (BSSR) is employed to represent the spike stream clip \\(\\{\\mathbf{S}_i\\}_{i=1}^N\\). Then the features are passed to the motion-guided super-resolution (MSR) module based on the temporal neighboring pixels of each color. Finally, the HR color image is restored by our color correlation-based reconstruction (CCR) module." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.337, + 0.471, + 0.366 + ], + "angle": 0, + "content": "Then the \\(n\\)-th Bayer-pattern raw image can be estimated according to the intervals, which can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.372, + 0.47, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} _ {n} (x, y) = \\frac {\\delta}{\\Psi_ {n} (x , y)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.47, + 0.467 + ], + "angle": 0, + "content": "where \\(\\delta\\) denotes the maximum dynamic range. As to a spike number-based method (TFP [48]), the raw image \\(\\mathbf{R}_n(x,y)\\) can also be inferred based on the spike number within a temporal window with an odd window length \\(w\\):" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.473, + 0.47, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} _ {n} (x, y) = \\frac {\\delta}{w} \\sum_ {i = n + 1} ^ {n + w} \\mathbf {S} _ {i - \\frac {w + 1}{2}} (x, y). \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.52, + 0.168, + 0.535 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.274, + 0.559 + ], + "angle": 0, + "content": "4.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.471, + 0.81 + ], + "angle": 0, + "content": "To reconstruct HR color images from LR Bayer-pattern spike stream, we develop a CSCSR network whose overall architecture is shown in Fig. 3. The input of the network is a clip of the spike stream with a shape of \\( N \\times H \\times W \\), denoted as \\( \\{\\mathbf{S}_i\\}_{i=1}^N \\). To represent color and motion information within the binary input, we first pass it to our proposed Bayer-pattern spike stream representation (BSSR) module. Considering the motion consistency of color channels, we split out four sequences according to the color layout and jointly estimate sub-pixel level motion from them. Guided by the motion information, temporal neighboring pixels of each color channel are searched in the motion-guided super-resolution (MSR) module, resulting in HR temporal-pixel features. Finally, the temporal-pixel features of the three color channels are fused to reconstruct the final HR image, exploiting the correlation of color channels." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.45, + 0.834 + ], + "angle": 0, + "content": "4.2. Bayer-Pattern Spike Stream Representation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The spike signals in the Bayer-pattern spike stream mean the arrival of a certain photon amount for a certain color channel, which may suffer from quantization errors as introduced in Sec. 3.1. Considering the motion and color" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.892, + 0.396 + ], + "angle": 0, + "content": "information within the binary data, how to represent it appropriately is worth studying. As shown in Fig. 4, we propose a Bayer-pattern spike stream representation module to represent the information contained in the data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.893, + 0.578 + ], + "angle": 0, + "content": "To keep the color information, we first perform elementwise multiplication between each frame of the spike sequence and the mask of each color, splitting the data into three sequences. For the sequence of each color channel, we develop an encoder to extract the temporal features and retain the relative motion. To be specific, we employ a sliding temporal window to extract the features of each time point. As the impact of quantization errors can be reduced by more temporal information, we extract the features from the whole clip serving as an additional global perception (GP) to the features of each time point. With the temporal window sliding, features of each color are obtained." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.588, + 0.891, + 0.676 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.741 + ], + "angle": 0, + "content": "Figure 4. Illustration of the Bayer-pattern spike stream representation module, with a temporal sliding window in encoders. In the figure, \"ConvBlock\" denotes a \\(3 \\times 3\\) convolution layer followed by a ReLU activation function. \"ResBlock\" refers to [13]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.757, + 0.794, + 0.773 + ], + "angle": 0, + "content": "4.3. Motion-Guided Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Joint Motion Estimation. Though the spatial pixels are limited, some pixels from the temporal domain are available for SR, which makes it essential to analyze the motion within the Bayer-pattern spike clip. Therefore, we employ the spike camera optical flow method Spike2Flow [46]. To apply the method to the Bayer-pattern spike stream, we split it into four single-channel spike sequences and estimate motion from each channel as shown in Fig. 3, resulting in op" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24874" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.568, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.089, + 0.892, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.244, + 0.894, + 0.272 + ], + "angle": 0, + "content": "Figure 5. Illustration of the motion-guided super-resolution (MSR) module, taking the blue channel as an example. (a) Overview of the MSR module. (b) Motion-guided temporal pixel search, with three steps. (c) Structure of the feature refinement (FR) module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.473, + 0.412 + ], + "angle": 0, + "content": "tical flows \\(\\{\\mathbf{w}_{j,k}^{C}\\}, C \\in \\{R, G_{1}, G_{2}, B\\}\\), from the middle time point \\(j\\) to the last time point \\(k\\). By reversing the spike sequences, we then estimate optical flows \\(\\{\\mathbf{w}_{j,i}^{C}\\}\\) from the middle time point \\(j\\) to the first time point \\(i\\). Though coming from different intensities, the color channels of the dynamic scene share the same motion. For more robust motion estimation, we jointly consider the color channels and obtain the final sub-pixel level optical flows as" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.421, + 0.47, + 0.459 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {w} _ {j, i}, \\mathbf {w} _ {j, k}\\right) = \\left(\\frac {1}{4} \\sum_ {C} \\mathbf {w} _ {j, i} ^ {C}, \\frac {1}{4} \\sum_ {C} \\mathbf {w} _ {j, k} ^ {C}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.47, + 0.543 + ], + "angle": 0, + "content": "Temporal Pixel Search. To provide features for the final reconstruction of the middle time point, we propose to fill the gap of spatial pixels from LR to HR based on the known temporal pixels at non-middle time points, considering the motion and color layout within the spike stream clip." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.473, + 0.844 + ], + "angle": 0, + "content": "The positions of known spatial pixels of a certain color can be inferred according to the Bayer-pattern color mask. Denoting the SR scale as \\( s \\), the target spatial resolution is \\( sH \\times sW \\). In the target, one originally known spatial position corresponds to \\( s^2 \\) neighboring positions, resulting in an upsampled color mask as shown in Fig. 5 (b). To get the extra known pixels after upsampling, we apply bilinear interpolation to the features, resulting in the super-resolved features \\( \\mathbf{F} \\). Utilizing the estimated optical flows \\( (\\mathbf{w}_{j,i}, \\mathbf{w}_{j,k}) \\), we can infer the new spatial position of every pixel at the middle time point \\( j \\) when moving to a certain time point \\( t \\). To be specific, the optical flow \\( \\mathbf{w}_{j,i} \\) can be formulated as the pixel offset along \\( x \\) direction and \\( y \\) direction, which can be written as \\( [\\mathbf{U}_j^i, \\mathbf{V}_j^i] \\). Considering the spatial upsampling, the offsets should be multiplied by the scale \\( s \\). Since the time span from time point \\( i \\) or \\( k \\) to \\( j \\) is extremely small \\( (0.025\\mathrm{ms} \\times N) \\), we assume the motion is linear. As a result, after moving to a certain time point \\( t \\), the new position of a certain original position \\( (x,y) \\) at the middle time point \\( j \\) can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.849, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} \\left(x + s \\cdot \\frac {t - j}{i - j} \\mathbf {U} _ {j} ^ {i} (x, y), y + s \\cdot \\frac {t - j}{i - j} \\mathbf {V} _ {j} ^ {i} (x, y)\\right), & t < j, \\\\ \\left(x + s \\cdot \\frac {t - j}{k - j} \\mathbf {U} _ {j} ^ {k} (x, y), y + s \\cdot \\frac {t - j}{k - j} \\mathbf {V} _ {j} ^ {k} (x, y)\\right), & t > j. \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.892, + 0.429 + ], + "angle": 0, + "content": "According to Eqn. (8), we can map all the pixels at the middle time point to their new positions at any other time point. However, the new positions are usually sub-pixel. Thus we are going to reconstruct the temporal feature value at the sub-pixel position by the values of known integer-pixels according to the upsampled color mask. In particular, the value of the new sub-pixel position \\((x_{m},y_{m})\\) at time point \\(t\\) can be calculated based on the 4 nearest known pixels \\(\\{(x_i,y_i)\\}_{i = 1}^4\\), which can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.441, + 0.892, + 0.478 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} (t, x, y) = \\sum_ {i = 1} ^ {4} \\frac {\\phi_ {i}}{\\sum_ {i = 1} ^ {4} \\phi_ {i}} \\mathbf {F} (t, x _ {i}, y _ {i}), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.49, + 0.892, + 0.536 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}\\) denotes the temporal-pixel features, \\((x,y)\\) denotes the original position at the middle time point, and \\(\\phi_{i}\\) denotes the weight of each known value, which can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.547, + 0.892, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\phi_ {i} = \\frac {1}{\\sqrt {\\left(x _ {m} - x _ {i}\\right) ^ {2} + \\left(y _ {m} - y _ {i}\\right) ^ {2}}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.892, + 0.636 + ], + "angle": 0, + "content": "After the temporal pixel search process, we can obtain the temporal-pixel features of each color channel for the following HR reconstruction of the middle time point." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.892, + 0.789 + ], + "angle": 0, + "content": "Feature Refinement. In most cases, there are errors in the estimated motion, causing an undesired impact on the temporal pixel search. Therefore, we propose a feature refinement as shown in Fig. 5 (c). By averaging the input temporal-pixel features \\(\\mathbf{T}\\) temporally, a rough temporal average result of the channel as a reference can be obtained. Then we compute residuals between the reference and \\(\\mathbf{T}\\) to represent the pixel fluctuation of the frames. At last, we extract deeper features from the residuals and the based \\(\\mathbf{T}\\), and add them together for refined temporal-pixel features \\(\\hat{\\mathbf{T}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.848, + 0.816 + ], + "angle": 0, + "content": "4.4. Color Correlation-Based Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "With the temporal-pixel features, we are going to reconstruct the final HR image. Despite having different intensities, there is a correlation between color channels. Thus we can restore the image by integrating the features, achieving mutual information complementation of color channels." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24875" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.266, + 0.47, + 0.322 + ], + "angle": 0, + "content": "Figure 6. Illustration of three structures of the reconstruction module. (a) Our proposed color correlation-based reconstruction (CCR) module. (b) A reconstruction module considering no color correlation. (c) An intuitive design considering color correlation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.33, + 0.469, + 0.466 + ], + "angle": 0, + "content": "As shown in Fig. 6 (a), we design a reconstruction module utilizing multi-stage features of color channels. In particular, we first apply residual blocks to extract deeper features from the features of each color channel. To fuse the multi-color features of each stage for better aggregation, we develop a color fusion (CF) module to encode the features of color channels with the output of the previous CF stage. After several stages of CF, we employ residual blocks in the end to generate the final image from the fused features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.478, + 0.21, + 0.495 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.502, + 0.353, + 0.518 + ], + "angle": 0, + "content": "5.1. Color Spike Camera Simulator" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.536, + 0.468, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.684, + 0.468, + 0.711 + ], + "angle": 0, + "content": "Figure 7. The pipeline of the color spike camera simulator, with the green channel as an example of signal accumulation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To train models, a large amount of LR spike stream-HR color image pairs are required. However, it's hard to collect corresponding high-quality HR images, especially in scenes with high-speed motion. Inspired by the success of gray-scale spike camera simulators [43, 45], we also developed a simulator for color spike cameras. The pipeline of our simulator is shown in Fig. 7. To generate spike streams, we regard the input video as the dynamic scene to be recorded. As the frame rate of the input video is usually limited, the temporal information is not enough for the simulator of CSC with ultra-high temporal resolution to generate Bayer-pattern spike streams. To address this issue, we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "use a frame interpolation method [27] to generate latent intensity frames between the original video frames. To generate spike signals, we follow the mechanism introduced in Sec. 3.1 to accumulate light intensity from the latent frames and compare the accumulated value with the threshold periodically for firing spikes. As a result, a Bayer-pattern spike stream is generated from the input video. In particular, the pixel count used to accumulate intensity signals is determined by SR scales. For example, we accumulate signals of 4 pixels to generate the signal of one target pixel in the \\(\\times 2\\) case. Considering the CFA, the 4 pixels come from one of the channels according to the Bayer-pattern color layout." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.281, + 0.707, + 0.297 + ], + "angle": 0, + "content": "5.2. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Implementation Details. In our implementation, we randomly crop the Bayer-pattern spike frames into patches with spatial resolution \\(96 \\times 96\\) for training, with the RGGB Bayer pattern. We set the number of spike frames \\(N\\) to 41 and the batch size to 4. The length \\(w\\) and stride \\(s\\) of the sliding temporal window are set to 11 and 3. We employ \\(\\ell_{1}\\) loss function and use Adam optimizer with an initial learning rate of \\(10^{-4}\\), which will decay to 0.8 times after every 7500 iterations. Besides, we employ the PyTorch framework and train the models via an NVIDIA RTX3090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.635 + ], + "angle": 0, + "content": "Training and Evaluation Datasets. Using our developed CSC simulator, we generated the training dataset based on 240 scenes for training from REDS (120fps) [23], resulting in \\(5 \\times 240 = 1200\\) clips for each super-resolution scale (\\(\\times 2\\), \\(\\times 3\\) and \\(\\times 4\\)). Then we generate \\(5 \\times 30 = 150\\) clips for evaluation based on 30 videos of the REDS evaluation dataset. To demonstrate the generalization to non-REDS-based data, we also generated evaluation datasets from two video SR datasets, Vid4 [22] and Viceo [35]. Besides, we also captured some real-world Bayer-pattern spike streams to verify the performance further. The details of the synthetic evaluation datasets can be found in Table 1." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.646, + 0.887, + 0.708 + ], + "angle": 0, + "content": "
SourceScenesSamplesResolutionScales
REDS [23]30150720×12802, 3, 4
Vid4 [22]44-2, 3, 4
Vimeo [35]78247824448×2564
" + }, + { + "type": "table_caption", + "bbox": [ + 0.524, + 0.714, + 0.866, + 0.728 + ], + "angle": 0, + "content": "Table 1. Details of the three synthetic evaluation datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparison Methods. As there are no SR methods for CSCs, we first choose the state-of-the-art gray-scale spike camera SR networks SpikeSR-Net [45] and VidarSR [34] for comparison. To adapt the gray-scale methods to the task, we adopt two strategies: using Bayer-pattern spike streams for end-to-end training and handling each color channel separately. We denote the networks with the second strategy as SpikeSR-Net* and VidarSR*. Then we combine a CSC reconstruction method 3DRI [10] and an image/video SR method SwinIR [18]/BasicVSR [3], resulting in another two comparison methods 3DRI+SwinIR and 3DRI+BasicVSR." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24876" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.09, + 0.889, + 0.581 + ], + "angle": 0, + "content": "
ScaleMethodREDS-based DatasetVid4-based DatasetVimeo-based Dataset
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
×2TFI [48] +TSCNN [4]29.24dB0.81730.312625.06dB0.77860.3196---
TFP[48] +TSCNN27.43dB0.75650.367124.60dB0.76660.3142---
TFI+Real-RawVSR [36]31.03dB0.86270.247826.79dB0.84220.2508---
TFP+Real-RawVSR30.81dB0.85350.260127.10dB0.85640.2329---
3DRI [10] +SwinIR [18]31.13dB0.88090.219126.57dB0.86680.2100---
3DRI+BasicVSR [3]31.60dB0.89120.205227.03dB0.87480.2021---
VidarSR [34]30.81dB0.89910.188426.44dB0.89110.1791---
VidarSR*30.27dB0.84300.263125.28dB0.81060.2538---
SpikeSR-Net [45]32.38dB0.89130.209928.34dB0.89570.1833---
SpikeSR-Net*29.66dB0.82490.286124.62dB0.78540.2721---
CSCSR (ours)33.39dB0.91210.176428.77dB0.91100.1612---
×3TFI+TSCNN27.22dB0.75050.384222.91dB0.66090.4058---
TFP+TSCNN26.55dB0.71760.413723.22dB0.68000.3808---
TFI+Real-RawVSR28.68dB0.79420.329224.00dB0.72250.3474---
TFP+Real-RawVSR28.66dB0.78830.334424.39dB0.74070.3326---
3DRI+SwinIR28.38dB0.81360.297723.97dB0.75340.3160---
3DRI+BasicVSR28.65dB0.82080.290524.08dB0.75680.3135---
VidarSR29.36dB0.84140.267924.66dB0.80090.2816---
VidarSR*27.53dB0.76700.355422.78dB0.67420.3680---
SpikeSR-Net29.79dB0.82700.297025.33dB0.79150.2989---
SpikeSR-Net*26.73dB0.73670.390822.12dB0.63230.4067---
CSCSR (ours)29.92dB0.85130.265325.68dB0.82400.2761---
×4TFI+TSCNN25.98dB0.70100.431021.69dB0.57220.465924.77dB0.73040.3554
TFP+TSCNN26.55dB0.71760.413722.53dB0.62400.417725.56dB0.75870.3557
TFI+Real-RawVSR27.27dB0.73910.385122.53dB0.62400.417725.12dB0.74920.3244
TFP+Real-RawVSR27.32dB0.73470.388522.81dB0.64080.363425.15dB0.74930.3324
3DRI+SwinIR26.94dB0.75440.360222.47dB0.65330.390125.21dB0.74230.3218
3DRI+BasicVSR26.94dB0.75700.360522.52dB0.65070.397325.05dB0.74010.3204
VidarSR27.56dB0.76920.350522.69dB0.67120.386825.19dB0.75650.3098
VidarSR*26.51dB0.71350.413521.58dB0.58800.439524.64dB0.73520.3423
SpikeSR-Net28.28dB0.77550.351723.59dB0.69830.372725.27dB0.76110.3119
SpikeSR-Net*25.93dB0.69130.440321.01dB0.54510.471824.46dB0.72400.3593
CSCSR (ours)28.77dB0.79630.325323.79dB0.72230.355525.34dB0.76600.2967
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.585, + 0.893, + 0.614 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison. Red and blue indicate the best and the second-best performance, respectively. For each scene of Video [35], there is a sequence of LR frames but only one corresponding HR frame \\((\\times 4)\\). Thus only the dataset of \\(\\times 4\\) scale is generated for Video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.471, + 0.751 + ], + "angle": 0, + "content": "It's also intuitive to combine methods of estimating raw images from Bayer-pattern spike streams (Sec. 3.2) and joint demosaicing and super-resolution (JDSR) methods. Thus we employ TFI or TFP [48] to infer raw images, followed by an image JDSR method TSCNN [4] or a video JDSR method Real-RawVSR [36] to reconstruct color images. In addition, we employ PSNR, SSIM [30] and LPIPS [38] (lower means better) as quantitative metrics for comparison." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.276, + 0.773 + ], + "angle": 0, + "content": "5.3. Comparative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. As shown in Table 2, we conduct comparative experiments on the three SR scales. It can be found that our method achieves the best quantitative results in most experiments. The state-of-the-art spike camera SR methods VidarSR and SpikeSR-Net achieve competitive performance. However, VidarSR* and SpikeSR-Net* that handle each color channel separately don't perform as well as VidarSR and SpikeSR-Net, due to the lack of considera" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.735 + ], + "angle": 0, + "content": "tion of color correlation. With the temporal information, the multi-frame method 3DRI+BasicVSR performs better than the single-frame method 3DRI+SwinIR in most cases. This is also true for the multi-frame JDSR method RawVSR and the single-frame JDSR method TSCNN. Besides, with the increase of the SR scale, TFP-based methods show growing competitiveness to TFI-based methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Visual Results. Fig. 8 presents the visual comparison \\((\\times 4)\\) on the synthetic data. The visual quality of results by our proposed method is better than the other methods, with less motion blur and more accurate color. There are artifacts in the results produced by VidarSR* and SpikeSR-Net*, which demonstrates the importance of the consideration of color correlation. In addition, the multi-frame methods show better visual quality. To demonstrate model generalization, we conduct experiments on real-world captured Bayer-pattern spike streams as shown in Fig. 9. In contrast, our method can restore better textures and details." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24877" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.895, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.249, + 0.895, + 0.278 + ], + "angle": 0, + "content": "Figure 8. Visual comparison \\((\\times 4)\\) on the synthetic Bayer-pattern spike streams. The first sample comes from the REDS-based dataset, while the second one comes from the Vid4-based dataset. Please enlarge the figure for better comparison." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.283, + 0.891, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.451, + 0.893, + 0.479 + ], + "angle": 0, + "content": "Figure 9. Visual comparison \\((\\times 4)\\) on the real-world Bayer-pattern spike streams. The first spike stream sample is captured by a fast-moving CSC. The second one records a fast-rotating water bottle. Please enlarge the figure for better comparison." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.494, + 0.468, + 0.652 + ], + "angle": 0, + "content": "
CaseSetting Description×2×3×4
(A)Removing BSSR module33.0429.5728.57
(B)Removing GP from BSSR33.1729.7428.68
(C)Interval-based representation32.0528.1928.06
(D)Independent motion estimation33.2229.5528.59
(E)Assembled motion estimation30.2427.0826.66
(F)Removing MSR module32.6629.4228.39
(G)Removing FR from MSR33.0729.7028.56
(H)Reconstruction module (b)30.1927.3626.22
(I)Reconstruction module (c)33.2429.6928.66
(J)Our final network33.3929.9228.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.655, + 0.47, + 0.684 + ], + "angle": 0, + "content": "Table 3. Ablation study on the REDS-based evaluation dataset. Greener blocks represent higher PSNR(dB) performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.232, + 0.714 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "To verify the effectiveness of our methodology, we perform ablation studies on our modules and motion estimation strategy. Case (J) is our final CSCSR network. First, we remove the BSSR module in (A) to demonstrate the proposed representation. Then we remove the additional GP branch from BSSR in (B). We also replace BSSR with the interval-based representation [46, 48] in (C). To demonstrate our joint motion estimation strategy, we employ independent motion estimation for each color channel in (D) and an optical flow assembly strategy in (E). To be specific, the optical flows of color channels are assembled according to the Bayer pattern. After that, we implement studies on the MSR" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.895, + 0.587 + ], + "angle": 0, + "content": "module. In (F), MSR is removed from the network. In (G), the FR module is removed from MSE to verify its effectiveness. Finally, (H) and (I) are the studies on the proposed reconstruction module, where the module is replaced by the structures in (b) and (c) of Fig. 6, with the same number of residual blocks. The study results are shown in Table 3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.597, + 0.628, + 0.613 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.895, + 0.804 + ], + "angle": 0, + "content": "In this paper, we present a deep network to restore HR color images from LR Bayer-pattern spike streams. To represent the binary data, we develop a representation that utilizes local temporal information with global perception. Then we propose to collect temporal pixels for spatial super-resolution according to the color layout and sub-pixel level motion. To reduce the impact of motion estimation errors, we design a feature refinement module based on residuals. Finally, multi-stage temporal-pixel features of each color channel are jointly considered, resulting in the final HR color image. Experimental results on both synthetic and real-world captured data demonstrate our performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.66, + 0.833 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grants 62072009, 22127807, and in part by the National Key R&D Program of China under Grant 2021YFF0900501." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24878" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Patrick Bardow, Andrew J Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 884-892, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.253 + ], + "angle": 0, + "content": "[2] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Real-time video super-resolution with spatio-temporal networks and motion compensation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4778-4787, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.325 + ], + "angle": 0, + "content": "[3] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. BasicVSR: The search for essential components in video super-resolution and beyond. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4947-4956, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.393 + ], + "angle": 0, + "content": "[4] Kan Chang, Hengxin Li, Yufei Tan, Pak Lun Kevin Ding, and Baoxin Li. A two-stage convolutional neural network for joint demosaicking and super-resolution. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4238-4254, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.395, + 0.47, + 0.464 + ], + "angle": 0, + "content": "[5] Yakun Chang, Chu Zhou, Yuchen Hong, Liwen Hu, Chao Xu, Tiejun Huang, and Boxin Shi. 1000 FPS HDR video with a spike-rgb hybrid camera. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 22180-22190, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.47, + 0.52 + ], + "angle": 0, + "content": "[6] Jonghyun Choi, Kuk-Jin Yoon, et al. Learning to super resolve intensity images from events. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2768-2776, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.47, + 0.577 + ], + "angle": 0, + "content": "[7] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Proceedings of European Conference on Computer Vision (ECCV), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.47, + 0.619 + ], + "angle": 0, + "content": "[8] Siwei Dong, Tiejun Huang, and Yonghong Tian. Spike camera and its coding methods. In Data Compression Conference (DCC), page 437, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[9] Siwei Dong, Lin Zhu, Daoyuan Xu, Yonghong Tian, and Tiejun Huang. An efficient coding method for spike camera using inter-spike intervals. In Data Compression Conference (DCC), page 568, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[10] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. 3D residual interpolation for spike camera demosaicing. In 2022 IEEE International Conference on Image Processing (ICIP), pages 1461-1465. IEEE, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[11] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. High-speed scene reconstruction from low-light spike streams. In 2022 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[12] Peiqi Duan, Zihao W Wang, Xinyu Zhou, Yi Ma, and Boxin Shi. EventZoom: Learning to denoise and super resolve neuromorphic events. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12824-12833, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[14] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3224-3232, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[15] Armin Kappeler, Seunghwan Yoo, Qiqin Dai, and Aggelos K Katsaggelos. Video super-resolution with convolutional neural networks. IEEE Transactions on Computational Imaging, 2(2):109-122, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[16] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew J Davison. Simultaneous mosaicing and tracking with an event camera. J. Solid State Circ, 43:566-576, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[17] Hongmin Li, Guoqi Li, and Luping Shi. Super-resolution of spatiotemporal event-stream image. Neurocomputing, 335: 206-214, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[19] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A \\(128 \\times 128\\) 120 db \\(15\\mu s\\) latency asynchronous temporal contrast vision sensor. IEEE journal of solid-state circuits, 43 (2):566-576, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[20] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In IEEE conference on computer vision and pattern recognition workshops (CVPRW), pages 136-144, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[21] Martin Litzenberger, Christoph Posch, D Bauer, Ahmed Nabil Belbachir, P Schon, B Kohn, and H Garn. Embedded vision system for real-time object tracking using an asynchronous transient vision sensor. In 2006 IEEE 12th Digital Signal Processing Workshop & 4th IEEE Signal Processing Education Workshop, pages 173-178. IEEE, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[22] Ce Liu and Deqing Sun. On bayesian adaptive video super resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(2):346-360, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[23] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. NTIRE 2019 challenge on video deblurring and superresolution: Dataset and study. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 0–0, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[24] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[25] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3857-3866, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24879" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[26] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast image reconstruction with an event camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 156-163, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[27] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. XVFI: extreme video frame interpolation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14489-14498, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.472, + 0.276 + ], + "angle": 0, + "content": "[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.472, + 0.347 + ], + "angle": 0, + "content": "[29] Lin Wang, Tae-Kyun Kim, and Kuk-Jin Yoon. EventSR: From asynchronous events to image reconstruction, restoration, and super-resolution via end-to-end adversarial learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8315-8325, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.472, + 0.403 + ], + "angle": 0, + "content": "[30] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.472, + 0.487 + ], + "angle": 0, + "content": "[31] Zihao W. Wang, Peiqi Duan, Oliver Cossairt, Aggelos Katsaggelos, Tiejun Huang, and Boxin Shi. Joint filtering of intensity images and neuromorphic events for high-resolution noise-robust imaging. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1609–1619, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.472, + 0.545 + ], + "angle": 0, + "content": "[32] Lujie Xia, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. SVFI: Spiking-based video frame interpolation for high-speed motion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2910-2918, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.472, + 0.615 + ], + "angle": 0, + "content": "[33] Lujie Xia, Ziluo Ding, Rui Zhao, Jiyuan Zhang, Lei Ma, Zhaofei Yu, Tiejun Huang, and Ruiqin Xiong. Unsupervised optical flow estimation with dynamic timing representation for spike camera. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.472, + 0.686 + ], + "angle": 0, + "content": "[34] Xijie Xiang, Lin Zhu, Jianing Li, Yixuan Wang, Tiejun Huang, and Yonghong Tian. Learning super-resolution reconstruction for high temporal resolution spike stream. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.472, + 0.73 + ], + "angle": 0, + "content": "[35] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. Int. J. Comput. Vis, 127:1106-1125, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.472, + 0.787 + ], + "angle": 0, + "content": "[36] Huanjing Yue, Zhiming Zhang, and Jingyu Yang. Realrawvsr: Real-world raw video super-resolution with a benchmark dataset. In European Conference on Computer Vision, pages 608-624. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.472, + 0.856 + ], + "angle": 0, + "content": "[37] Jiyuan Zhang, Shanshan Jia, Zhaofei Yu, and Tiejun Huang. Learning temporal-ordered representation for spike streams based on discrete wavelet transforms. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 137–147, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[38] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[39] Yiyang Zhang, Ruiqin Xiong, and Tiejun Huang. Spike signal reconstruction based on inter-spike similarity. In 2022 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[40] Jing Zhao, Ruiqin Xiong, and Tiejun Huang. High-speed motion scene reconstruction for spike camera via motion aligned filtering. In IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-5, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[41] Jing Zhao, Ruiqin Xiong, Rui Zhao, Jin Wang, Siwei Ma, and Tiejun Huang. Motion estimation for spike camera data sequence via spike interval analysis. In 2020 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 371-374. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[42] Jing Zhao, Jiyu Xie, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, and Tiejun Huang. Super resolve dynamic scene from continuous spike streams. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2533-2542, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[43] Jing Zhao, Ruiqin Xiong, Hangfan Liu, Jian Zhang, and Tiejun Huang. Spk2ImgNet: Learning to reconstruct dynamic scene from continuous spike stream. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 11996-12005, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[44] Jing Zhao, Ruiqin Xiong, Jiyu Xie, Boxin Shi, Zhaofei Yu, Wen Gao, and Tiejun Huang. Reconstructing clear image for high-speed motion scene with a retina-inspired spike camera. IEEE Transactions on Computational Imaging, 8:12-27, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[45] Jing Zhao, Ruiqin Xiong, Jian Zhang, Rui Zhao, Hangfan Liu, and Tiejun Huang. Learning to super-resolve dynamic scenes for neuromorphic spike camera. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3579-3587, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[46] Rui Zhao, Ruiqin Xiong, Jing Zhao, Zhaofei Yu, Xiaopeng Fan, and Tiejun Huang. Learning optical flow from continuous spike streams. Advances in Neural Information Processing Systems, 35:7905-7920, 2022. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.892, + 0.703 + ], + "angle": 0, + "content": "[47] Rui Zhao, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, Shuyuan Zhu, Lei Ma, and Tiejun Huang. Spike camera image reconstruction using deep spiking neural networks. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[48] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. A retina-inspired sampling method for visual texture reconstruction. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1432-1437. IEEE, 2019. 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[49] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. Hybrid coding of spatiotemporal spike data for a bio-inspired camera. IEEE Transactions on Circuits and Systems for Video Technology, 31(7):2837-2851, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.833, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[50] Lin Zhu, Jianing Li, Xiao Wang, Tiejun Huang, and Yonghong Tian. NeuSpike-Net: High speed video reconstruction via bio-inspired neuromorphic cameras. In IEEE International Conference on Computer Vision (ICCV), pages 2400–2409, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24880" + } + ] +] \ No newline at end of file diff --git a/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/ac654108-cae7-444a-a7e8-a0faf89a682d_origin.pdf b/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/ac654108-cae7-444a-a7e8-a0faf89a682d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6e46c82139103f68ae16fdc40471c4daea123847 --- /dev/null +++ b/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/ac654108-cae7-444a-a7e8-a0faf89a682d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:254383369ff9ed8ad494e4c4c54c2ec0ccf1eae21df73c15aea8ee318c479906 +size 2489611 diff --git a/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/full.md b/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0dd9c8d153e17151861d605ec816ddf34814cde8 --- /dev/null +++ b/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/full.md @@ -0,0 +1,298 @@ +# Super-Resolution Reconstruction from Bayer-Pattern Spike Streams + +Yanchen Dong $^{1}$ , Ruiqin Xiong $^{1*}$ , Jian Zhang $^{2}$ , Zhaofei Yu $^{3}$ , Xiaopeng Fan $^{4}$ , Shuyuan Zhu $^{5}$ , Tiejun Huang $^{1,3}$ $^{1}$ School of Computer Science, Peking University + $^{2}$ School of Electronic and Computer Engineering, Peking University + $^{3}$ Institute for Artificial Intelligence, Peking University + $^{4}$ School of Computer Science and Technology, Harbin Institute of Technology + $^{5}$ University of Electronic Science and Technology of China + +yanchendong@stu.pku.edu.cn, {rqxiong, zhangjian.sz, yuzf12, tjhuang}@pku.edu.cn, fxp@hit.edu.cn, eezsy@uestc.edu.cn + +# Abstract + +Spike camera is a neuromorphic vision sensor that can capture highly dynamic scenes by generating a continuous stream of binary spikes to represent the arrival of photons at very high temporal resolution. Equipped with Bayer color filter array (CFA), color spike camera (CSC) has been invented to capture color information. Although spike camera has already demonstrated great potential for high-speed imaging, its spatial resolution is limited compared with conventional digital cameras. This paper proposes a Color Spike Camera Super-Resolution (CSCSR) network to super-resolve higher-resolution color images from spike camera streams with Bayer CFA. To be specific, we first propose a representation for Bayer-pattern spike streams, exploring local temporal information with global perception to represent the binary data. Then we exploit the CFA layout and sub-pixel level motion to collect temporal pixels for the spatial super-resolution of each color channel. In particular, a residual-based module for feature refinement is developed to reduce the impact of motion estimation errors. Considering color correlation, we jointly utilize the multi-stage temporal-pixel features of color channels to reconstruct the high-resolution color image. Experimental results demonstrate that the proposed scheme can reconstruct satisfactory color images with both high temporal and spatial resolution from low-resolution Bayer-pattern spike streams. The source codes are available at https://github.com/csycdong/CSCSR. + +# 1. Introduction + +With the quick development of high-speed vision applications such as autonomous driving and unmanned aerial ve + +![](images/faeac15f9e3171a483a0abc4ff01fa9c5151a9eda65618a1ee817961e5b6c468.jpg) +Figure 1. Pipeline of super-resolution reconstruction from Bayer-pattern spike streams. The super-resolution scale is denoted as $s$ . + +hicles, the demand for cameras that can capture high-speed motion and respond quickly is increasing. Most conventional digital cameras use a certain exposure time window to accumulate photoelectric signals and create a snapshot image, which requires the scene to be still during the exposure interval. Due to the exposure time, conventional cameras suffer from motion blur when capturing high-speed scenes. To be specific, a given point of moving objects can move during exposure time, bringing undesired motion blur to the image. Therefore, it's hard for conventional digital cameras to meet the requirements of high-speed imaging. + +Mimicking the structure of human vision, neuromorphic cameras with ultra-high temporal resolution (e.g., $20,000\mathrm{Hz}$ ) show great potential in capturing high-speed scenes. A well-known type of neuromorphic camera called + +event camera [19, 21, 24] is sensitive to motion in dynamic scenes, which can record relative light intensity changes. To record more texture details, another type of neuromorphic camera called spike camera [8, 9, 49] captures absolute light intensity via an "integrate-and-fire" mechanism. As the first-generation spike camera can only reproduce gray-level signals, color spike camera, which employs the Bayer-pattern color filter array (CFA), has been invented to record dynamic scenes with color information and meanwhile enjoy the benefits of high-speed imaging. + +Although capable of capturing color information like conventional cameras, the spatial resolution of color spike cameras is limited compared to recent digital cameras, which is a trade-off for high temporal resolution and low latency. Therefore, we aim to reconstruct high-resolution (HR) color images from low-resolution (LR) Bayer-pattern spike streams, achieving both high temporal and spatial resolution with a color spike camera. An intuitive way is to combine color spike camera reconstruction methods [10] with image super-resolution (SR) algorithms [7, 18, 20]. However, information among temporal neighboring frames that is essential for SR will be degraded in the first reconstruction stage of this scheme. Besides, high-speed motion and quantization noise within the Bayer-pattern spike streams also bring challenges to the SR task. As a result, how to improve the spatial resolution as well as the imaging quality for color spike cameras is worth studying. + +In this paper, we propose a Color Spike Camera SuperResolution (CSCSR) network to super-resolve dynamic scenes from Bayer-pattern spike streams. To represent the binary data with color and motion information, we first develop a Bayer-pattern Spike Stream Representation (BSSR) to explore local temporal information with global perception. As the pixels of a single frame are limited for spatial super-resolution, we propose a Motion-guided SuperResolution (MSR) module to search temporal pixels of each color channel according to the Bayer-pattern color layout and sub-pixel level motion jointly estimated from color channels. In most cases, there are errors in the estimated motion, causing an undesired impact on the temporal pixel search. Therefore, a Feature Refinement (FR) module is designed based on the residual between a temporal average reference and the temporal pixels, resulting in refined temporal-pixel features. Finally, we further extract features and jointly utilize multi-stage features of each color channel to restore the final HR image, achieved by our Color Correlation-based Reconstruction (CCR) module. + +To train and evaluate our models, we design a simulator based on the mechanism of color spike cameras, which can generate data pairs for SR reconstruction. Besides, we also capture some real-world Bayer-pattern spike streams for further evaluation. Experimental results on both synthetic and real-world captured data show that our method + +can reconstruct HR color images from LR Bayer-pattern spike streams, with good texture details. The main contribution of our work can be summarized as follows: + +- We develop a color spike camera super-resolution network to reconstruct high-resolution color images from low-resolution Bayer-pattern spike streams, which is the first attempt to the best of our knowledge. +- To reconstruct details beyond the sensor resolution, we explore the sub-pixel level motion of each color channel and propose a motion-guided super-resolution module to collect temporal pixels according to the color layout. +- Experiments demonstrate that our method outperforms all the existing methods, achieving better quantitative and qualitative results in color spike camera super-resolution. + +# 2. Related Work + +# 2.1. Reconstruction + +Event Camera Reconstruction. Event camera [19, 21, 24] is a kind of neuromorphic camera that monitors relative light intensity changes, making it promising in capturing motion. With the high temporal resolution, reconstructing intensity images with an event camera is an active topic. Extended Kalman Filter [16] is the early attempt at event camera reconstruction, which is based on photometric constancy. Bardow et al. [1] estimate intensity as well as optical flow via the primal-dual algorithm. Inspired by the success of deep learning, E2VID [25] and FireNet [26] were presented with promoting reconstruction performance. + +Spike Camera Reconstruction. Spike camera [8, 9, 33, 41, 49] is another type of neuromorphic camera. Different from event cameras, spike camera produces a continuous spike stream to record the dynamic scenes by accumulating photons and firing spike signals. To reconstruct images from spike streams, some exploration [11, 39, 40, 44] has been performed. Zhu et al. [48] proposed to utilize the number of spike signals within a temporal window (TFP) and neighboring spike intervals (TFI). For better performance, some work [37, 43, 47] tried to design reconstruction networks for spike cameras. In addition, there are also attempts at hybrid input [5, 32, 50] for reconstruction. Focusing on the newly invented color spike camera, 3DRI [10] was proposed for scene reconstruction with color information. + +# 2.2. Super-Resolution + +Image and Video Super-Resolution. In the last ten years, many researchers have been devoted to learning to superresolve images. Dong et al. [7] proposed SRNet as the first SR network, bringing great performance gains. In recent years, Liang et al. [18] have achieved significant improvement based on Transformer [28]. Besides the progress in image SR, there are also some video SR work [2, 3, 14, 15] proposed to restore HR frames from LR video frames. + +Event Camera Super-Resolution. Event camera SR includes upsampling from LR events to HR events [12, 17] and HR image reconstruction from LR events [6, 29, 31]. Li et al. [17] designed a two-stage scheme to solve the spatial SR problem of spatiotemporal events. Then a deep neural framework EventZoom [12] was proposed for event SR. To restore SR images from events directly, Choi et al. [6] proposed a network E2SRI and trained models using synthetic data. In contrast, another deep network EventSR [29] was implemented without ground truth HR images. + +Spike Camera Super-Resolution. As a trade-off of temporal resolution, the spatial resolution of spike cameras is limited. To restore images with both high temporal and spatial resolution, there is some research about spike camera super-resolution. Zhao et al. [42] exploited relative motion and derived the relationship between light intensity and each spike for super-resolution. Xiang et al. [34] proposed an end-to-end network VidarSR to reconstruct HR images from LR spike streams. Besides, another network SpikeSR-Net [45] is designed based on the observation model of spike camera, achieving state-of-the-art performance. + +# 3. Preliminary + +# 3.1. Color Spike Camera + +Color spike camera (CSC) is a recently implemented neuromorphic sensor with $H \times W$ pixels that work independently. To capture color information of dynamic scenes, color filter array is employed on the sensor. As a result, each pixel of the sensor corresponds to one of the three colors, red, green and blue, according to the Bayer-pattern color layout. + +As shown in Fig. 1, CSC is a neuromorphic vision sensor that mimics the structure of human vision. Conventional digital cameras usually use a certain time window for exposure to accumulate photoelectric signals and compact them into a snapshot. In contrast, CSCs accumulate photons and fire spikes continuously to record dynamic scenes, resulting in a Bayer-pattern spike stream. As shown in Fig. 2, every individual pixel denoted as $(x,y)$ within the sensor accumulates photons continuously and converts the instantaneous intensity into an electric signal. When the accumulated electric signals reach a predetermined threshold $\theta$ , the pixel triggers a flag indicating firing a spike. Then the pixel will be reset and continue to accumulate photons after the flag is checked, starting a new accumulating and firing process. The accumulated electric signal of a certain pixel $(x,y)$ at arbitrary time point $t$ can be formulated as + +$$ +\mathbf {A} (t, x, y) = \int_ {0} ^ {t} \eta \cdot \mathbf {I} _ {C} (\tau , x, y) d \tau \mod \theta , \tag {1} +$$ + +where $\eta$ denotes the photoelectric conversion rate of the sensor, $\mathbf{I}_C(\tau ,x,y)$ denotes the instantaneous intensity of a certain color at time $\tau$ , and $C\in \{R,G,B\}$ denotes the color determined by the Bayer-pattern color layout. + +![](images/ebe6083a21a8ed0969bbc6a0d959e9ae3035f09fb660a023250dbbfa93702840.jpg) + +![](images/2cdc79f3cf68e88aca59222bec342bf8e8587a4e0872c378444e4c6486230657.jpg) +Figure 2. Illustration of the spike camera model. (a) The instantaneous intensity of a certain pixel. (b) The accumulated electric signal of the pixel. (c) The number of spikes that have been fired at the pixel. (d) The spike stream by the pixel, controlled by a clock. + +![](images/7ef3f81ac81dbaa81432ac5ffd3e8801f95ea1eec331d9fdb85a6ba073551f15.jpg) + +![](images/1cfbbc745a72c592aa62958dcd8bcbdd45e173845d65131dcff7bfdff0d28ee7.jpg) + +Ideally, the pixels on the sensor would be reset immediately after firing a spike. The number of spikes that have been fired before arbitrary time point $t$ can be written as + +$$ +\Phi (t, x, y) = \left\lfloor \frac {1}{\theta} \int_ {0} ^ {t} \mathbf {I} _ {C} (\tau , x, y) d \tau \right\rfloor . \tag {2} +$$ + +Then the read-out spike signal at the $n$ -th $(n \in \mathbb{N}^{+})$ frame can be formulated as follows: + +$$ +\mathbf {S} _ {n} (x, y) = \left\{ \begin{array}{l l} 1, & \Phi (t _ {n}, x, y) - \Phi (t _ {n - 1}, x, y) > 0 \\ 0, & \Phi (t _ {n}, x, y) - \Phi (t _ {n - 1}, x, y) = 0 \end{array} , \right. \tag {3} +$$ + +where $t_n$ denotes the time point according to the $n$ -th spike frame. However, the flags of firing spikes are checked and reset at discrete time points in real hardware implementation, bringing undesired quantization errors. To be specific, CSCs can only check them at discrete time points with a fixed sensor period $T$ , which is controlled by a clock. If the flag of the pixel $(x,y)$ is set up between time points $t = n \cdot T$ and $t' = (n - 1) \cdot T$ , we can read out the spike signal $\mathbf{S}_n(x,y)$ as 1. Otherwise, we have $\mathbf{S}_n(x,y) = 0$ . After $N$ times of the checking and resetting process, CSC outputs a binary Bayer-pattern spike stream $\{\mathbf{S}_i\}_{i=1}^N$ with a spatial-temporal shape of $N \times H \times W$ . + +# 3.2. Bayer-Pattern Spike Stream to Raw Images + +To reconstruct color images from the Bayer-pattern spike stream, an intuitive way is to convert it to raw images for demosaicing. To infer the raw images, we can employ some spike camera reconstruction methods that don't fuse values of spatial neighboring pixels of multiple colors, such as the spike interval-based method (TFI [48]). Specifically, the spike interval of two temporally neighboring spikes that covers the $n$ -th spike frame can be formulated as + +$$ +\begin{array}{l} \Psi_ {n} (x, y) = \min \left\{k \mid \mathbf {S} _ {i} (x, y) = 1, k \geq n \right\} \tag {4} \\ - \max \{k | \mathbf {S} _ {i} (x, y) = 1, k < n \}. \\ \end{array} +$$ + +![](images/9d2066606d3f5164ee5e339268ef64638f64e04f2a88bf4a5cfd9f1bebfb20a9.jpg) +Figure 3. The overall architecture of CSCSR network. First, the Bayer-pattern spike stream representation (BSSR) is employed to represent the spike stream clip $\{\mathbf{S}_i\}_{i=1}^N$ . Then the features are passed to the motion-guided super-resolution (MSR) module based on the temporal neighboring pixels of each color. Finally, the HR color image is restored by our color correlation-based reconstruction (CCR) module. + +Then the $n$ -th Bayer-pattern raw image can be estimated according to the intervals, which can be written as + +$$ +\mathbf {R} _ {n} (x, y) = \frac {\delta}{\Psi_ {n} (x , y)}, \tag {5} +$$ + +where $\delta$ denotes the maximum dynamic range. As to a spike number-based method (TFP [48]), the raw image $\mathbf{R}_n(x,y)$ can also be inferred based on the spike number within a temporal window with an odd window length $w$ : + +$$ +\mathbf {R} _ {n} (x, y) = \frac {\delta}{w} \sum_ {i = n + 1} ^ {n + w} \mathbf {S} _ {i - \frac {w + 1}{2}} (x, y). \tag {6} +$$ + +# 4. Method + +# 4.1. Overall Architecture + +To reconstruct HR color images from LR Bayer-pattern spike stream, we develop a CSCSR network whose overall architecture is shown in Fig. 3. The input of the network is a clip of the spike stream with a shape of $N \times H \times W$ , denoted as $\{\mathbf{S}_i\}_{i=1}^N$ . To represent color and motion information within the binary input, we first pass it to our proposed Bayer-pattern spike stream representation (BSSR) module. Considering the motion consistency of color channels, we split out four sequences according to the color layout and jointly estimate sub-pixel level motion from them. Guided by the motion information, temporal neighboring pixels of each color channel are searched in the motion-guided super-resolution (MSR) module, resulting in HR temporal-pixel features. Finally, the temporal-pixel features of the three color channels are fused to reconstruct the final HR image, exploiting the correlation of color channels. + +# 4.2. Bayer-Pattern Spike Stream Representation + +The spike signals in the Bayer-pattern spike stream mean the arrival of a certain photon amount for a certain color channel, which may suffer from quantization errors as introduced in Sec. 3.1. Considering the motion and color + +information within the binary data, how to represent it appropriately is worth studying. As shown in Fig. 4, we propose a Bayer-pattern spike stream representation module to represent the information contained in the data. + +To keep the color information, we first perform elementwise multiplication between each frame of the spike sequence and the mask of each color, splitting the data into three sequences. For the sequence of each color channel, we develop an encoder to extract the temporal features and retain the relative motion. To be specific, we employ a sliding temporal window to extract the features of each time point. As the impact of quantization errors can be reduced by more temporal information, we extract the features from the whole clip serving as an additional global perception (GP) to the features of each time point. With the temporal window sliding, features of each color are obtained. + +![](images/d26537fb7a69453c9d7ba8d807d62dcfa7f308d35a986bf954547a1ff56cb4d4.jpg) +Figure 4. Illustration of the Bayer-pattern spike stream representation module, with a temporal sliding window in encoders. In the figure, "ConvBlock" denotes a $3 \times 3$ convolution layer followed by a ReLU activation function. "ResBlock" refers to [13]. + +# 4.3. Motion-Guided Super-Resolution + +Joint Motion Estimation. Though the spatial pixels are limited, some pixels from the temporal domain are available for SR, which makes it essential to analyze the motion within the Bayer-pattern spike clip. Therefore, we employ the spike camera optical flow method Spike2Flow [46]. To apply the method to the Bayer-pattern spike stream, we split it into four single-channel spike sequences and estimate motion from each channel as shown in Fig. 3, resulting in op + +![](images/435f4ea8aea956ffc3632b0f61cdc700f235c00a85beea121fb16237fb6ec76a.jpg) +Figure 5. Illustration of the motion-guided super-resolution (MSR) module, taking the blue channel as an example. (a) Overview of the MSR module. (b) Motion-guided temporal pixel search, with three steps. (c) Structure of the feature refinement (FR) module. + +![](images/308f45367f4748e58496a9bcc7a9b471fd8f72f08647cc6a9fddcad456fd6023.jpg) + +tical flows $\{\mathbf{w}_{j,k}^{C}\}, C \in \{R, G_{1}, G_{2}, B\}$ , from the middle time point $j$ to the last time point $k$ . By reversing the spike sequences, we then estimate optical flows $\{\mathbf{w}_{j,i}^{C}\}$ from the middle time point $j$ to the first time point $i$ . Though coming from different intensities, the color channels of the dynamic scene share the same motion. For more robust motion estimation, we jointly consider the color channels and obtain the final sub-pixel level optical flows as + +$$ +\left(\mathbf {w} _ {j, i}, \mathbf {w} _ {j, k}\right) = \left(\frac {1}{4} \sum_ {C} \mathbf {w} _ {j, i} ^ {C}, \frac {1}{4} \sum_ {C} \mathbf {w} _ {j, k} ^ {C}\right). \tag {7} +$$ + +Temporal Pixel Search. To provide features for the final reconstruction of the middle time point, we propose to fill the gap of spatial pixels from LR to HR based on the known temporal pixels at non-middle time points, considering the motion and color layout within the spike stream clip. + +The positions of known spatial pixels of a certain color can be inferred according to the Bayer-pattern color mask. Denoting the SR scale as $s$ , the target spatial resolution is $sH \times sW$ . In the target, one originally known spatial position corresponds to $s^2$ neighboring positions, resulting in an upsampled color mask as shown in Fig. 5 (b). To get the extra known pixels after upsampling, we apply bilinear interpolation to the features, resulting in the super-resolved features $\mathbf{F}$ . Utilizing the estimated optical flows $(\mathbf{w}_{j,i}, \mathbf{w}_{j,k})$ , we can infer the new spatial position of every pixel at the middle time point $j$ when moving to a certain time point $t$ . To be specific, the optical flow $\mathbf{w}_{j,i}$ can be formulated as the pixel offset along $x$ direction and $y$ direction, which can be written as $[\mathbf{U}_j^i, \mathbf{V}_j^i]$ . Considering the spatial upsampling, the offsets should be multiplied by the scale $s$ . Since the time span from time point $i$ or $k$ to $j$ is extremely small $(0.025\mathrm{ms} \times N)$ , we assume the motion is linear. As a result, after moving to a certain time point $t$ , the new position of a certain original position $(x,y)$ at the middle time point $j$ can be written as follows: + +$$ +\left\{ \begin{array}{l l} \left(x + s \cdot \frac {t - j}{i - j} \mathbf {U} _ {j} ^ {i} (x, y), y + s \cdot \frac {t - j}{i - j} \mathbf {V} _ {j} ^ {i} (x, y)\right), & t < j, \\ \left(x + s \cdot \frac {t - j}{k - j} \mathbf {U} _ {j} ^ {k} (x, y), y + s \cdot \frac {t - j}{k - j} \mathbf {V} _ {j} ^ {k} (x, y)\right), & t > j. \end{array} \right. \tag {8} +$$ + +According to Eqn. (8), we can map all the pixels at the middle time point to their new positions at any other time point. However, the new positions are usually sub-pixel. Thus we are going to reconstruct the temporal feature value at the sub-pixel position by the values of known integer-pixels according to the upsampled color mask. In particular, the value of the new sub-pixel position $(x_{m},y_{m})$ at time point $t$ can be calculated based on the 4 nearest known pixels $\{(x_i,y_i)\}_{i = 1}^4$ , which can be formulated as follows: + +$$ +\mathbf {T} (t, x, y) = \sum_ {i = 1} ^ {4} \frac {\phi_ {i}}{\sum_ {i = 1} ^ {4} \phi_ {i}} \mathbf {F} (t, x _ {i}, y _ {i}), \tag {9} +$$ + +where $\mathbf{T}$ denotes the temporal-pixel features, $(x,y)$ denotes the original position at the middle time point, and $\phi_{i}$ denotes the weight of each known value, which can be written as + +$$ +\phi_ {i} = \frac {1}{\sqrt {\left(x _ {m} - x _ {i}\right) ^ {2} + \left(y _ {m} - y _ {i}\right) ^ {2}}}. \tag {10} +$$ + +After the temporal pixel search process, we can obtain the temporal-pixel features of each color channel for the following HR reconstruction of the middle time point. + +Feature Refinement. In most cases, there are errors in the estimated motion, causing an undesired impact on the temporal pixel search. Therefore, we propose a feature refinement as shown in Fig. 5 (c). By averaging the input temporal-pixel features $\mathbf{T}$ temporally, a rough temporal average result of the channel as a reference can be obtained. Then we compute residuals between the reference and $\mathbf{T}$ to represent the pixel fluctuation of the frames. At last, we extract deeper features from the residuals and the based $\mathbf{T}$ , and add them together for refined temporal-pixel features $\hat{\mathbf{T}}$ . + +# 4.4. Color Correlation-Based Reconstruction + +With the temporal-pixel features, we are going to reconstruct the final HR image. Despite having different intensities, there is a correlation between color channels. Thus we can restore the image by integrating the features, achieving mutual information complementation of color channels. + +![](images/13975d8102b76e23a5899aa043b8995970dc4634aa4dee4f39bda9477b092218.jpg) +Figure 6. Illustration of three structures of the reconstruction module. (a) Our proposed color correlation-based reconstruction (CCR) module. (b) A reconstruction module considering no color correlation. (c) An intuitive design considering color correlation. + +As shown in Fig. 6 (a), we design a reconstruction module utilizing multi-stage features of color channels. In particular, we first apply residual blocks to extract deeper features from the features of each color channel. To fuse the multi-color features of each stage for better aggregation, we develop a color fusion (CF) module to encode the features of color channels with the output of the previous CF stage. After several stages of CF, we employ residual blocks in the end to generate the final image from the fused features. + +# 5. Experiments + +# 5.1. Color Spike Camera Simulator + +![](images/eff25f3a2bcd92b935d9d2498c874409453d61b66e9d8fb041ab4f0c6ad94dbc.jpg) +Figure 7. The pipeline of the color spike camera simulator, with the green channel as an example of signal accumulation. + +To train models, a large amount of LR spike stream-HR color image pairs are required. However, it's hard to collect corresponding high-quality HR images, especially in scenes with high-speed motion. Inspired by the success of gray-scale spike camera simulators [43, 45], we also developed a simulator for color spike cameras. The pipeline of our simulator is shown in Fig. 7. To generate spike streams, we regard the input video as the dynamic scene to be recorded. As the frame rate of the input video is usually limited, the temporal information is not enough for the simulator of CSC with ultra-high temporal resolution to generate Bayer-pattern spike streams. To address this issue, we + +use a frame interpolation method [27] to generate latent intensity frames between the original video frames. To generate spike signals, we follow the mechanism introduced in Sec. 3.1 to accumulate light intensity from the latent frames and compare the accumulated value with the threshold periodically for firing spikes. As a result, a Bayer-pattern spike stream is generated from the input video. In particular, the pixel count used to accumulate intensity signals is determined by SR scales. For example, we accumulate signals of 4 pixels to generate the signal of one target pixel in the $\times 2$ case. Considering the CFA, the 4 pixels come from one of the channels according to the Bayer-pattern color layout. + +# 5.2. Experimental Settings + +Implementation Details. In our implementation, we randomly crop the Bayer-pattern spike frames into patches with spatial resolution $96 \times 96$ for training, with the RGGB Bayer pattern. We set the number of spike frames $N$ to 41 and the batch size to 4. The length $w$ and stride $s$ of the sliding temporal window are set to 11 and 3. We employ $\ell_{1}$ loss function and use Adam optimizer with an initial learning rate of $10^{-4}$ , which will decay to 0.8 times after every 7500 iterations. Besides, we employ the PyTorch framework and train the models via an NVIDIA RTX3090 GPU. + +Training and Evaluation Datasets. Using our developed CSC simulator, we generated the training dataset based on 240 scenes for training from REDS (120fps) [23], resulting in $5 \times 240 = 1200$ clips for each super-resolution scale ( $\times 2$ , $\times 3$ and $\times 4$ ). Then we generate $5 \times 30 = 150$ clips for evaluation based on 30 videos of the REDS evaluation dataset. To demonstrate the generalization to non-REDS-based data, we also generated evaluation datasets from two video SR datasets, Vid4 [22] and Viceo [35]. Besides, we also captured some real-world Bayer-pattern spike streams to verify the performance further. The details of the synthetic evaluation datasets can be found in Table 1. + +
SourceScenesSamplesResolutionScales
REDS [23]30150720×12802, 3, 4
Vid4 [22]44-2, 3, 4
Vimeo [35]78247824448×2564
+ +Table 1. Details of the three synthetic evaluation datasets. + +Comparison Methods. As there are no SR methods for CSCs, we first choose the state-of-the-art gray-scale spike camera SR networks SpikeSR-Net [45] and VidarSR [34] for comparison. To adapt the gray-scale methods to the task, we adopt two strategies: using Bayer-pattern spike streams for end-to-end training and handling each color channel separately. We denote the networks with the second strategy as SpikeSR-Net* and VidarSR*. Then we combine a CSC reconstruction method 3DRI [10] and an image/video SR method SwinIR [18]/BasicVSR [3], resulting in another two comparison methods 3DRI+SwinIR and 3DRI+BasicVSR. + +
ScaleMethodREDS-based DatasetVid4-based DatasetVimeo-based Dataset
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
×2TFI [48] +TSCNN [4]29.24dB0.81730.312625.06dB0.77860.3196---
TFP[48] +TSCNN27.43dB0.75650.367124.60dB0.76660.3142---
TFI+Real-RawVSR [36]31.03dB0.86270.247826.79dB0.84220.2508---
TFP+Real-RawVSR30.81dB0.85350.260127.10dB0.85640.2329---
3DRI [10] +SwinIR [18]31.13dB0.88090.219126.57dB0.86680.2100---
3DRI+BasicVSR [3]31.60dB0.89120.205227.03dB0.87480.2021---
VidarSR [34]30.81dB0.89910.188426.44dB0.89110.1791---
VidarSR*30.27dB0.84300.263125.28dB0.81060.2538---
SpikeSR-Net [45]32.38dB0.89130.209928.34dB0.89570.1833---
SpikeSR-Net*29.66dB0.82490.286124.62dB0.78540.2721---
CSCSR (ours)33.39dB0.91210.176428.77dB0.91100.1612---
×3TFI+TSCNN27.22dB0.75050.384222.91dB0.66090.4058---
TFP+TSCNN26.55dB0.71760.413723.22dB0.68000.3808---
TFI+Real-RawVSR28.68dB0.79420.329224.00dB0.72250.3474---
TFP+Real-RawVSR28.66dB0.78830.334424.39dB0.74070.3326---
3DRI+SwinIR28.38dB0.81360.297723.97dB0.75340.3160---
3DRI+BasicVSR28.65dB0.82080.290524.08dB0.75680.3135---
VidarSR29.36dB0.84140.267924.66dB0.80090.2816---
VidarSR*27.53dB0.76700.355422.78dB0.67420.3680---
SpikeSR-Net29.79dB0.82700.297025.33dB0.79150.2989---
SpikeSR-Net*26.73dB0.73670.390822.12dB0.63230.4067---
CSCSR (ours)29.92dB0.85130.265325.68dB0.82400.2761---
×4TFI+TSCNN25.98dB0.70100.431021.69dB0.57220.465924.77dB0.73040.3554
TFP+TSCNN26.55dB0.71760.413722.53dB0.62400.417725.56dB0.75870.3557
TFI+Real-RawVSR27.27dB0.73910.385122.53dB0.62400.417725.12dB0.74920.3244
TFP+Real-RawVSR27.32dB0.73470.388522.81dB0.64080.363425.15dB0.74930.3324
3DRI+SwinIR26.94dB0.75440.360222.47dB0.65330.390125.21dB0.74230.3218
3DRI+BasicVSR26.94dB0.75700.360522.52dB0.65070.397325.05dB0.74010.3204
VidarSR27.56dB0.76920.350522.69dB0.67120.386825.19dB0.75650.3098
VidarSR*26.51dB0.71350.413521.58dB0.58800.439524.64dB0.73520.3423
SpikeSR-Net28.28dB0.77550.351723.59dB0.69830.372725.27dB0.76110.3119
SpikeSR-Net*25.93dB0.69130.440321.01dB0.54510.471824.46dB0.72400.3593
CSCSR (ours)28.77dB0.79630.325323.79dB0.72230.355525.34dB0.76600.2967
+ +Table 2. Quantitative comparison. Red and blue indicate the best and the second-best performance, respectively. For each scene of Video [35], there is a sequence of LR frames but only one corresponding HR frame $(\times 4)$ . Thus only the dataset of $\times 4$ scale is generated for Video. + +It's also intuitive to combine methods of estimating raw images from Bayer-pattern spike streams (Sec. 3.2) and joint demosaicing and super-resolution (JDSR) methods. Thus we employ TFI or TFP [48] to infer raw images, followed by an image JDSR method TSCNN [4] or a video JDSR method Real-RawVSR [36] to reconstruct color images. In addition, we employ PSNR, SSIM [30] and LPIPS [38] (lower means better) as quantitative metrics for comparison. + +# 5.3. Comparative Results + +Quantitative Results. As shown in Table 2, we conduct comparative experiments on the three SR scales. It can be found that our method achieves the best quantitative results in most experiments. The state-of-the-art spike camera SR methods VidarSR and SpikeSR-Net achieve competitive performance. However, VidarSR* and SpikeSR-Net* that handle each color channel separately don't perform as well as VidarSR and SpikeSR-Net, due to the lack of considera + +tion of color correlation. With the temporal information, the multi-frame method 3DRI+BasicVSR performs better than the single-frame method 3DRI+SwinIR in most cases. This is also true for the multi-frame JDSR method RawVSR and the single-frame JDSR method TSCNN. Besides, with the increase of the SR scale, TFP-based methods show growing competitiveness to TFI-based methods. + +Visual Results. Fig. 8 presents the visual comparison $(\times 4)$ on the synthetic data. The visual quality of results by our proposed method is better than the other methods, with less motion blur and more accurate color. There are artifacts in the results produced by VidarSR* and SpikeSR-Net*, which demonstrates the importance of the consideration of color correlation. In addition, the multi-frame methods show better visual quality. To demonstrate model generalization, we conduct experiments on real-world captured Bayer-pattern spike streams as shown in Fig. 9. In contrast, our method can restore better textures and details. + +![](images/162dd7e3ffc318ca8a901b5537b8712e68d331223f48887cc9bbb89a24eabf8e.jpg) +Figure 8. Visual comparison $(\times 4)$ on the synthetic Bayer-pattern spike streams. The first sample comes from the REDS-based dataset, while the second one comes from the Vid4-based dataset. Please enlarge the figure for better comparison. + +![](images/8fc7be26a481a16f935452ccaa5173c4daf18051e6f369b1dce80f9aa338060d.jpg) +Figure 9. Visual comparison $(\times 4)$ on the real-world Bayer-pattern spike streams. The first spike stream sample is captured by a fast-moving CSC. The second one records a fast-rotating water bottle. Please enlarge the figure for better comparison. + +
CaseSetting Description×2×3×4
(A)Removing BSSR module33.0429.5728.57
(B)Removing GP from BSSR33.1729.7428.68
(C)Interval-based representation32.0528.1928.06
(D)Independent motion estimation33.2229.5528.59
(E)Assembled motion estimation30.2427.0826.66
(F)Removing MSR module32.6629.4228.39
(G)Removing FR from MSR33.0729.7028.56
(H)Reconstruction module (b)30.1927.3626.22
(I)Reconstruction module (c)33.2429.6928.66
(J)Our final network33.3929.9228.77
+ +Table 3. Ablation study on the REDS-based evaluation dataset. Greener blocks represent higher PSNR(dB) performance. + +# 5.4. Ablation Study + +To verify the effectiveness of our methodology, we perform ablation studies on our modules and motion estimation strategy. Case (J) is our final CSCSR network. First, we remove the BSSR module in (A) to demonstrate the proposed representation. Then we remove the additional GP branch from BSSR in (B). We also replace BSSR with the interval-based representation [46, 48] in (C). To demonstrate our joint motion estimation strategy, we employ independent motion estimation for each color channel in (D) and an optical flow assembly strategy in (E). To be specific, the optical flows of color channels are assembled according to the Bayer pattern. After that, we implement studies on the MSR + +module. In (F), MSR is removed from the network. In (G), the FR module is removed from MSE to verify its effectiveness. Finally, (H) and (I) are the studies on the proposed reconstruction module, where the module is replaced by the structures in (b) and (c) of Fig. 6, with the same number of residual blocks. The study results are shown in Table 3. + +# 6. Conclusions + +In this paper, we present a deep network to restore HR color images from LR Bayer-pattern spike streams. To represent the binary data, we develop a representation that utilizes local temporal information with global perception. Then we propose to collect temporal pixels for spatial super-resolution according to the color layout and sub-pixel level motion. To reduce the impact of motion estimation errors, we design a feature refinement module based on residuals. Finally, multi-stage temporal-pixel features of each color channel are jointly considered, resulting in the final HR color image. Experimental results on both synthetic and real-world captured data demonstrate our performance. + +# Acknowledgments + +This work was supported in part by the National Natural Science Foundation of China under Grants 62072009, 22127807, and in part by the National Key R&D Program of China under Grant 2021YFF0900501. + +# References + +[1] Patrick Bardow, Andrew J Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 884-892, 2016. 2 +[2] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Real-time video super-resolution with spatio-temporal networks and motion compensation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4778-4787, 2017. 2 +[3] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. BasicVSR: The search for essential components in video super-resolution and beyond. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4947-4956, 2021. 2, 6, 7 +[4] Kan Chang, Hengxin Li, Yufei Tan, Pak Lun Kevin Ding, and Baoxin Li. A two-stage convolutional neural network for joint demosaicking and super-resolution. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4238-4254, 2021. 7 +[5] Yakun Chang, Chu Zhou, Yuchen Hong, Liwen Hu, Chao Xu, Tiejun Huang, and Boxin Shi. 1000 FPS HDR video with a spike-rgb hybrid camera. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 22180-22190, 2023. 2 +[6] Jonghyun Choi, Kuk-Jin Yoon, et al. Learning to super resolve intensity images from events. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2768-2776, 2020. 3 +[7] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Proceedings of European Conference on Computer Vision (ECCV), 2014. 2 +[8] Siwei Dong, Tiejun Huang, and Yonghong Tian. Spike camera and its coding methods. In Data Compression Conference (DCC), page 437, 2017. 2 +[9] Siwei Dong, Lin Zhu, Daoyuan Xu, Yonghong Tian, and Tiejun Huang. An efficient coding method for spike camera using inter-spike intervals. In Data Compression Conference (DCC), page 568, 2019. 2 +[10] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. 3D residual interpolation for spike camera demosaicing. In 2022 IEEE International Conference on Image Processing (ICIP), pages 1461-1465. IEEE, 2022. 2, 6, 7 +[11] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. High-speed scene reconstruction from low-light spike streams. In 2022 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2022. 2 +[12] Peiqi Duan, Zihao W Wang, Xinyu Zhou, Yi Ma, and Boxin Shi. EventZoom: Learning to denoise and super resolve neuromorphic events. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12824-12833, 2021. 3 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE + +Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 4 +[14] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3224-3232, 2018. 2 +[15] Armin Kappeler, Seunghwan Yoo, Qiqin Dai, and Aggelos K Katsaggelos. Video super-resolution with convolutional neural networks. IEEE Transactions on Computational Imaging, 2(2):109-122, 2016. 2 +[16] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew J Davison. Simultaneous mosaicing and tracking with an event camera. J. Solid State Circ, 43:566-576, 2008. 2 +[17] Hongmin Li, Guoqi Li, and Luping Shi. Super-resolution of spatiotemporal event-stream image. Neurocomputing, 335: 206-214, 2019. 3 +[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 2, 6, 7 +[19] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \times 128$ 120 db $15\mu s$ latency asynchronous temporal contrast vision sensor. IEEE journal of solid-state circuits, 43 (2):566-576, 2008. 2 +[20] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In IEEE conference on computer vision and pattern recognition workshops (CVPRW), pages 136-144, 2017. 2 +[21] Martin Litzenberger, Christoph Posch, D Bauer, Ahmed Nabil Belbachir, P Schon, B Kohn, and H Garn. Embedded vision system for real-time object tracking using an asynchronous transient vision sensor. In 2006 IEEE 12th Digital Signal Processing Workshop & 4th IEEE Signal Processing Education Workshop, pages 173-178. IEEE, 2006. 2 +[22] Ce Liu and Deqing Sun. On bayesian adaptive video super resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(2):346-360, 2013. 6 +[23] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. NTIRE 2019 challenge on video deblurring and superresolution: Dataset and study. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 0–0, 2019. 6 +[24] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2010. 2 +[25] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3857-3866, 2019. 2 + +[26] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast image reconstruction with an event camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 156-163, 2020. 2 +[27] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. XVFI: extreme video frame interpolation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14489-14498, 2021. 6 +[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[29] Lin Wang, Tae-Kyun Kim, and Kuk-Jin Yoon. EventSR: From asynchronous events to image reconstruction, restoration, and super-resolution via end-to-end adversarial learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8315-8325, 2020. 3 +[30] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 7 +[31] Zihao W. Wang, Peiqi Duan, Oliver Cossairt, Aggelos Katsaggelos, Tiejun Huang, and Boxin Shi. Joint filtering of intensity images and neuromorphic events for high-resolution noise-robust imaging. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1609–1619, 2020. 3 +[32] Lujie Xia, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. SVFI: Spiking-based video frame interpolation for high-speed motion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2910-2918, 2023. 2 +[33] Lujie Xia, Ziluo Ding, Rui Zhao, Jiyuan Zhang, Lei Ma, Zhaofei Yu, Tiejun Huang, and Ruiqin Xiong. Unsupervised optical flow estimation with dynamic timing representation for spike camera. Advances in Neural Information Processing Systems, 36, 2024. 2 +[34] Xijie Xiang, Lin Zhu, Jianing Li, Yixuan Wang, Tiejun Huang, and Yonghong Tian. Learning super-resolution reconstruction for high temporal resolution spike stream. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 3, 6, 7 +[35] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. Int. J. Comput. Vis, 127:1106-1125, 2019. 6, 7 +[36] Huanjing Yue, Zhiming Zhang, and Jingyu Yang. Realrawvsr: Real-world raw video super-resolution with a benchmark dataset. In European Conference on Computer Vision, pages 608-624. Springer, 2022. 7 +[37] Jiyuan Zhang, Shanshan Jia, Zhaofei Yu, and Tiejun Huang. Learning temporal-ordered representation for spike streams based on discrete wavelet transforms. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 137–147, 2023. 2 +[38] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference + +on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 7 +[39] Yiyang Zhang, Ruiqin Xiong, and Tiejun Huang. Spike signal reconstruction based on inter-spike similarity. In 2022 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2022. 2 +[40] Jing Zhao, Ruiqin Xiong, and Tiejun Huang. High-speed motion scene reconstruction for spike camera via motion aligned filtering. In IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-5, 2020. 2 +[41] Jing Zhao, Ruiqin Xiong, Rui Zhao, Jin Wang, Siwei Ma, and Tiejun Huang. Motion estimation for spike camera data sequence via spike interval analysis. In 2020 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 371-374. IEEE, 2020. 2 +[42] Jing Zhao, Jiyu Xie, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, and Tiejun Huang. Super resolve dynamic scene from continuous spike streams. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2533-2542, 2021. 3 +[43] Jing Zhao, Ruiqin Xiong, Hangfan Liu, Jian Zhang, and Tiejun Huang. Spk2ImgNet: Learning to reconstruct dynamic scene from continuous spike stream. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 11996-12005, 2021. 2, 6 +[44] Jing Zhao, Ruiqin Xiong, Jiyu Xie, Boxin Shi, Zhaofei Yu, Wen Gao, and Tiejun Huang. Reconstructing clear image for high-speed motion scene with a retina-inspired spike camera. IEEE Transactions on Computational Imaging, 8:12-27, 2021. 2 +[45] Jing Zhao, Ruiqin Xiong, Jian Zhang, Rui Zhao, Hangfan Liu, and Tiejun Huang. Learning to super-resolve dynamic scenes for neuromorphic spike camera. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3579-3587, 2023. 3, 6, 7 +[46] Rui Zhao, Ruiqin Xiong, Jing Zhao, Zhaofei Yu, Xiaopeng Fan, and Tiejun Huang. Learning optical flow from continuous spike streams. Advances in Neural Information Processing Systems, 35:7905-7920, 2022. 4, 8 +[47] Rui Zhao, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, Shuyuan Zhu, Lei Ma, and Tiejun Huang. Spike camera image reconstruction using deep spiking neural networks. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2 +[48] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. A retina-inspired sampling method for visual texture reconstruction. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1432-1437. IEEE, 2019. 2, 3, 4, 7, 8 +[49] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. Hybrid coding of spatiotemporal spike data for a bio-inspired camera. IEEE Transactions on Circuits and Systems for Video Technology, 31(7):2837-2851, 2021. 2 +[50] Lin Zhu, Jianing Li, Xiao Wang, Tiejun Huang, and Yonghong Tian. NeuSpike-Net: High speed video reconstruction via bio-inspired neuromorphic cameras. In IEEE International Conference on Computer Vision (ICCV), pages 2400–2409, 2021. 2 \ No newline at end of file diff --git a/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/images.zip b/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..56205ad47fb8fc66442b967ec4ff756ba6d2abb5 --- /dev/null +++ b/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20fbe610859fbfc88ddfdef438be16f17040bf999356e0498a30a07cb8aa013c +size 1012814 diff --git a/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/layout.json b/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..390a5ea0b9b1ec06e425912018b1d4694473bbf0 --- /dev/null +++ b/2024/Super-Resolution Reconstruction from Bayer-Pattern Spike Streams/layout.json @@ -0,0 +1,8321 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 507, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 507, + 121 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 507, + 121 + ], + "type": "text", + "content": "Super-Resolution Reconstruction from Bayer-Pattern Spike Streams" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "text", + "content": "Yanchen Dong" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "text", + "content": ", Ruiqin Xiong" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "text", + "content": ", Jian Zhang" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "text", + "content": ", Zhaofei Yu" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "text", + "content": ", Xiaopeng Fan" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "text", + "content": ", Shuyuan Zhu" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "text", + "content": ", Tiejun Huang" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "text", + "content": "School of Computer Science, Peking University \n" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "text", + "content": "School of Electronic and Computer Engineering, Peking University \n" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "text", + "content": "Institute for Artificial Intelligence, Peking University \n" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "text", + "content": "School of Computer Science and Technology, Harbin Institute of Technology \n" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 53, + 141, + 569, + 228 + ], + "type": "text", + "content": "University of Electronic Science and Technology of China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 95, + 229, + 526, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 229, + 526, + 256 + ], + "spans": [ + { + "bbox": [ + 95, + 229, + 526, + 256 + ], + "type": "text", + "content": "yanchendong@stu.pku.edu.cn, {rqxiong, zhangjian.sz, yuzf12, tjhuang}@pku.edu.cn, fxp@hit.edu.cn, eezsy@uestc.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "spans": [ + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 308, + 290, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 290, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 290, + 632 + ], + "type": "text", + "content": "Spike camera is a neuromorphic vision sensor that can capture highly dynamic scenes by generating a continuous stream of binary spikes to represent the arrival of photons at very high temporal resolution. Equipped with Bayer color filter array (CFA), color spike camera (CSC) has been invented to capture color information. Although spike camera has already demonstrated great potential for high-speed imaging, its spatial resolution is limited compared with conventional digital cameras. This paper proposes a Color Spike Camera Super-Resolution (CSCSR) network to super-resolve higher-resolution color images from spike camera streams with Bayer CFA. To be specific, we first propose a representation for Bayer-pattern spike streams, exploring local temporal information with global perception to represent the binary data. Then we exploit the CFA layout and sub-pixel level motion to collect temporal pixels for the spatial super-resolution of each color channel. In particular, a residual-based module for feature refinement is developed to reduce the impact of motion estimation errors. Considering color correlation, we jointly utilize the multi-stage temporal-pixel features of color channels to reconstruct the high-resolution color image. Experimental results demonstrate that the proposed scheme can reconstruct satisfactory color images with both high temporal and spatial resolution from low-resolution Bayer-pattern spike streams. The source codes are available at https://github.com/csycdong/CSCSR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 652, + 128, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 652, + 128, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 652, + 128, + 664 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "type": "text", + "content": "With the quick development of high-speed vision applications such as autonomous driving and unmanned aerial ve" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 281, + 545, + 488 + ], + "blocks": [ + { + "bbox": [ + 308, + 281, + 545, + 488 + ], + "lines": [ + { + "bbox": [ + 308, + 281, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 281, + 545, + 488 + ], + "type": "image", + "image_path": "faeac15f9e3171a483a0abc4ff01fa9c5151a9eda65618a1ee817961e5b6c468.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 495, + 545, + 517 + ], + "lines": [ + { + "bbox": [ + 305, + 495, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 305, + 495, + 545, + 517 + ], + "type": "text", + "content": "Figure 1. Pipeline of super-resolution reconstruction from Bayer-pattern spike streams. The super-resolution scale is denoted as " + }, + { + "bbox": [ + 305, + 495, + 545, + 517 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 495, + 545, + 517 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 533, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 665 + ], + "type": "text", + "content": "hicles, the demand for cameras that can capture high-speed motion and respond quickly is increasing. Most conventional digital cameras use a certain exposure time window to accumulate photoelectric signals and create a snapshot image, which requires the scene to be still during the exposure interval. Due to the exposure time, conventional cameras suffer from motion blur when capturing high-speed scenes. To be specific, a given point of moving objects can move during exposure time, bringing undesired motion blur to the image. Therefore, it's hard for conventional digital cameras to meet the requirements of high-speed imaging." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Mimicking the structure of human vision, neuromorphic cameras with ultra-high temporal resolution (e.g., " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "20,000\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": ") show great potential in capturing high-speed scenes. A well-known type of neuromorphic camera called" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24871" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "event camera [19, 21, 24] is sensitive to motion in dynamic scenes, which can record relative light intensity changes. To record more texture details, another type of neuromorphic camera called spike camera [8, 9, 49] captures absolute light intensity via an \"integrate-and-fire\" mechanism. As the first-generation spike camera can only reproduce gray-level signals, color spike camera, which employs the Bayer-pattern color filter array (CFA), has been invented to record dynamic scenes with color information and meanwhile enjoy the benefits of high-speed imaging." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 289, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 397 + ], + "type": "text", + "content": "Although capable of capturing color information like conventional cameras, the spatial resolution of color spike cameras is limited compared to recent digital cameras, which is a trade-off for high temporal resolution and low latency. Therefore, we aim to reconstruct high-resolution (HR) color images from low-resolution (LR) Bayer-pattern spike streams, achieving both high temporal and spatial resolution with a color spike camera. An intuitive way is to combine color spike camera reconstruction methods [10] with image super-resolution (SR) algorithms [7, 18, 20]. However, information among temporal neighboring frames that is essential for SR will be degraded in the first reconstruction stage of this scheme. Besides, high-speed motion and quantization noise within the Bayer-pattern spike streams also bring challenges to the SR task. As a result, how to improve the spatial resolution as well as the imaging quality for color spike cameras is worth studying." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 400, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 289, + 639 + ], + "type": "text", + "content": "In this paper, we propose a Color Spike Camera SuperResolution (CSCSR) network to super-resolve dynamic scenes from Bayer-pattern spike streams. To represent the binary data with color and motion information, we first develop a Bayer-pattern Spike Stream Representation (BSSR) to explore local temporal information with global perception. As the pixels of a single frame are limited for spatial super-resolution, we propose a Motion-guided SuperResolution (MSR) module to search temporal pixels of each color channel according to the Bayer-pattern color layout and sub-pixel level motion jointly estimated from color channels. In most cases, there are errors in the estimated motion, causing an undesired impact on the temporal pixel search. Therefore, a Feature Refinement (FR) module is designed based on the residual between a temporal average reference and the temporal pixels, resulting in refined temporal-pixel features. Finally, we further extract features and jointly utilize multi-stage features of each color channel to restore the final HR image, achieved by our Color Correlation-based Reconstruction (CCR) module." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "content": "To train and evaluate our models, we design a simulator based on the mechanism of color spike cameras, which can generate data pairs for SR reconstruction. Besides, we also capture some real-world Bayer-pattern spike streams for further evaluation. Experimental results on both synthetic and real-world captured data show that our method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "can reconstruct HR color images from LR Bayer-pattern spike streams, with good texture details. The main contribution of our work can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 110, + 547, + 243 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 110, + 547, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 110, + 547, + 158 + ], + "spans": [ + { + "bbox": [ + 306, + 110, + 547, + 158 + ], + "type": "text", + "content": "- We develop a color spike camera super-resolution network to reconstruct high-resolution color images from low-resolution Bayer-pattern spike streams, which is the first attempt to the best of our knowledge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 159, + 547, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 159, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 306, + 159, + 547, + 206 + ], + "type": "text", + "content": "- To reconstruct details beyond the sensor resolution, we explore the sub-pixel level motion of each color channel and propose a motion-guided super-resolution module to collect temporal pixels according to the color layout." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 206, + 547, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 206, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 306, + 206, + 547, + 243 + ], + "type": "text", + "content": "- Experiments demonstrate that our method outperforms all the existing methods, achieving better quantitative and qualitative results in color spike camera super-resolution." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 253, + 392, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 253, + 392, + 266 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 392, + 266 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 273, + 400, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 273, + 400, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 273, + 400, + 285 + ], + "type": "text", + "content": "2.1. Reconstruction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 292, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 545, + 422 + ], + "type": "text", + "content": "Event Camera Reconstruction. Event camera [19, 21, 24] is a kind of neuromorphic camera that monitors relative light intensity changes, making it promising in capturing motion. With the high temporal resolution, reconstructing intensity images with an event camera is an active topic. Extended Kalman Filter [16] is the early attempt at event camera reconstruction, which is based on photometric constancy. Bardow et al. [1] estimate intensity as well as optical flow via the primal-dual algorithm. Inspired by the success of deep learning, E2VID [25] and FireNet [26] were presented with promoting reconstruction performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 423, + 547, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 547, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 547, + 591 + ], + "type": "text", + "content": "Spike Camera Reconstruction. Spike camera [8, 9, 33, 41, 49] is another type of neuromorphic camera. Different from event cameras, spike camera produces a continuous spike stream to record the dynamic scenes by accumulating photons and firing spike signals. To reconstruct images from spike streams, some exploration [11, 39, 40, 44] has been performed. Zhu et al. [48] proposed to utilize the number of spike signals within a temporal window (TFP) and neighboring spike intervals (TFI). For better performance, some work [37, 43, 47] tried to design reconstruction networks for spike cameras. In addition, there are also attempts at hybrid input [5, 32, 50] for reconstruction. Focusing on the newly invented color spike camera, 3DRI [10] was proposed for scene reconstruction with color information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 599, + 410, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 410, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 410, + 612 + ], + "type": "text", + "content": "2.2. Super-Resolution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": "Image and Video Super-Resolution. In the last ten years, many researchers have been devoted to learning to superresolve images. Dong et al. [7] proposed SRNet as the first SR network, bringing great performance gains. In recent years, Liang et al. [18] have achieved significant improvement based on Transformer [28]. Besides the progress in image SR, there are also some video SR work [2, 3, 14, 15] proposed to restore HR frames from LR video frames." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24872" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "Event Camera Super-Resolution. Event camera SR includes upsampling from LR events to HR events [12, 17] and HR image reconstruction from LR events [6, 29, 31]. Li et al. [17] designed a two-stage scheme to solve the spatial SR problem of spatiotemporal events. Then a deep neural framework EventZoom [12] was proposed for event SR. To restore SR images from events directly, Choi et al. [6] proposed a network E2SRI and trained models using synthetic data. In contrast, another deep network EventSR [29] was implemented without ground truth HR images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 290, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 290, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 290, + 324 + ], + "type": "text", + "content": "Spike Camera Super-Resolution. As a trade-off of temporal resolution, the spatial resolution of spike cameras is limited. To restore images with both high temporal and spatial resolution, there is some research about spike camera super-resolution. Zhao et al. [42] exploited relative motion and derived the relationship between light intensity and each spike for super-resolution. Xiang et al. [34] proposed an end-to-end network VidarSR to reconstruct HR images from LR spike streams. Besides, another network SpikeSR-Net [45] is designed based on the observation model of spike camera, achieving state-of-the-art performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 332, + 125, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 332, + 125, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 332, + 125, + 346 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 352, + 167, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 352, + 167, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 352, + 167, + 365 + ], + "type": "text", + "content": "3.1. Color Spike Camera" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 370, + 288, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 370, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 46, + 370, + 288, + 442 + ], + "type": "text", + "content": "Color spike camera (CSC) is a recently implemented neuromorphic sensor with " + }, + { + "bbox": [ + 46, + 370, + 288, + 442 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 370, + 288, + 442 + ], + "type": "text", + "content": " pixels that work independently. To capture color information of dynamic scenes, color filter array is employed on the sensor. As a result, each pixel of the sensor corresponds to one of the three colors, red, green and blue, according to the Bayer-pattern color layout." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 443, + 289, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 289, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 289, + 635 + ], + "type": "text", + "content": "As shown in Fig. 1, CSC is a neuromorphic vision sensor that mimics the structure of human vision. Conventional digital cameras usually use a certain time window for exposure to accumulate photoelectric signals and compact them into a snapshot. In contrast, CSCs accumulate photons and fire spikes continuously to record dynamic scenes, resulting in a Bayer-pattern spike stream. As shown in Fig. 2, every individual pixel denoted as " + }, + { + "bbox": [ + 46, + 443, + 289, + 635 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 443, + 289, + 635 + ], + "type": "text", + "content": " within the sensor accumulates photons continuously and converts the instantaneous intensity into an electric signal. When the accumulated electric signals reach a predetermined threshold " + }, + { + "bbox": [ + 46, + 443, + 289, + 635 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 443, + 289, + 635 + ], + "type": "text", + "content": ", the pixel triggers a flag indicating firing a spike. Then the pixel will be reset and continue to accumulate photons after the flag is checked, starting a new accumulating and firing process. The accumulated electric signal of a certain pixel " + }, + { + "bbox": [ + 46, + 443, + 289, + 635 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 443, + 289, + 635 + ], + "type": "text", + "content": " at arbitrary time point " + }, + { + "bbox": [ + 46, + 443, + 289, + 635 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 443, + 289, + 635 + ], + "type": "text", + "content": " can be formulated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 639, + 287, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 639, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 84, + 639, + 287, + 663 + ], + "type": "interline_equation", + "content": "\\mathbf {A} (t, x, y) = \\int_ {0} ^ {t} \\eta \\cdot \\mathbf {I} _ {C} (\\tau , x, y) d \\tau \\mod \\theta , \\tag {1}", + "image_path": "658e849f08e359721d5db8438ebc9efd07f6b5a40007b3c9cecc769d3d5d2865.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": " denotes the photoelectric conversion rate of the sensor, " + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_C(\\tau ,x,y)" + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": " denotes the instantaneous intensity of a certain color at time " + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "inline_equation", + "content": "C\\in \\{R,G,B\\}" + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": " denotes the color determined by the Bayer-pattern color layout." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 427, + 145 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 427, + 145 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 427, + 145 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 427, + 145 + ], + "type": "image", + "image_path": "ebe6083a21a8ed0969bbc6a0d959e9ae3035f09fb660a023250dbbfa93702840.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 145, + 429, + 222 + ], + "blocks": [ + { + "bbox": [ + 306, + 145, + 429, + 222 + ], + "lines": [ + { + "bbox": [ + 306, + 145, + 429, + 222 + ], + "spans": [ + { + "bbox": [ + 306, + 145, + 429, + 222 + ], + "type": "image", + "image_path": "2cdc79f3cf68e88aca59222bec342bf8e8587a4e0872c378444e4c6486230657.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 231, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 304, + 231, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 547, + 277 + ], + "type": "text", + "content": "Figure 2. Illustration of the spike camera model. (a) The instantaneous intensity of a certain pixel. (b) The accumulated electric signal of the pixel. (c) The number of spikes that have been fired at the pixel. (d) The spike stream by the pixel, controlled by a clock." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 425, + 70, + 545, + 145 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 545, + 145 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 545, + 145 + ], + "type": "image", + "image_path": "7ef3f81ac81dbaa81432ac5ffd3e8801f95ea1eec331d9fdb85a6ba073551f15.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 423, + 146, + 541, + 222 + ], + "blocks": [ + { + "bbox": [ + 423, + 146, + 541, + 222 + ], + "lines": [ + { + "bbox": [ + 423, + 146, + 541, + 222 + ], + "spans": [ + { + "bbox": [ + 423, + 146, + 541, + 222 + ], + "type": "image", + "image_path": "1cfbbc745a72c592aa62958dcd8bcbdd45e173845d65131dcff7bfdff0d28ee7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 289, + 546, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 546, + 325 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 546, + 325 + ], + "type": "text", + "content": "Ideally, the pixels on the sensor would be reset immediately after firing a spike. The number of spikes that have been fired before arbitrary time point " + }, + { + "bbox": [ + 304, + 289, + 546, + 325 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 289, + 546, + 325 + ], + "type": "text", + "content": " can be written as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 358, + 330, + 545, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 330, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 358, + 330, + 545, + 354 + ], + "type": "interline_equation", + "content": "\\Phi (t, x, y) = \\left\\lfloor \\frac {1}{\\theta} \\int_ {0} ^ {t} \\mathbf {I} _ {C} (\\tau , x, y) d \\tau \\right\\rfloor . \\tag {2}", + "image_path": "43be394ecfdf557460464297768e1cb0c1785fe4ef3c91d20a36c1e79ee24c08.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 359, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 359, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 305, + 359, + 545, + 382 + ], + "type": "text", + "content": "Then the read-out spike signal at the " + }, + { + "bbox": [ + 305, + 359, + 545, + 382 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 359, + 545, + 382 + ], + "type": "text", + "content": "-th " + }, + { + "bbox": [ + 305, + 359, + 545, + 382 + ], + "type": "inline_equation", + "content": "(n \\in \\mathbb{N}^{+})" + }, + { + "bbox": [ + 305, + 359, + 545, + 382 + ], + "type": "text", + "content": " frame can be formulated as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 327, + 386, + 545, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 386, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 327, + 386, + 545, + 415 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {n} (x, y) = \\left\\{ \\begin{array}{l l} 1, & \\Phi (t _ {n}, x, y) - \\Phi (t _ {n - 1}, x, y) > 0 \\\\ 0, & \\Phi (t _ {n}, x, y) - \\Phi (t _ {n - 1}, x, y) = 0 \\end{array} , \\right. \\tag {3}", + "image_path": "b4ba02a3645f2314e3222a51ff5e0085b45827028e0eaa091b3fee9242dc3c66.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "text", + "content": " denotes the time point according to the " + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "text", + "content": "-th spike frame. However, the flags of firing spikes are checked and reset at discrete time points in real hardware implementation, bringing undesired quantization errors. To be specific, CSCs can only check them at discrete time points with a fixed sensor period " + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "text", + "content": ", which is controlled by a clock. If the flag of the pixel " + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "text", + "content": " is set up between time points " + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "inline_equation", + "content": "t = n \\cdot T" + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "inline_equation", + "content": "t' = (n - 1) \\cdot T" + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "text", + "content": ", we can read out the spike signal " + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_n(x,y)" + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "text", + "content": " as 1. Otherwise, we have " + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_n(x,y) = 0" + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "text", + "content": ". After " + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "text", + "content": " times of the checking and resetting process, CSC outputs a binary Bayer-pattern spike stream " + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "text", + "content": " with a spatial-temporal shape of " + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "inline_equation", + "content": "N \\times H \\times W" + }, + { + "bbox": [ + 304, + 419, + 547, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 570, + 533, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 570, + 533, + 582 + ], + "spans": [ + { + "bbox": [ + 305, + 570, + 533, + 582 + ], + "type": "text", + "content": "3.2. Bayer-Pattern Spike Stream to Raw Images" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 588, + 547, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 547, + 684 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 547, + 684 + ], + "type": "text", + "content": "To reconstruct color images from the Bayer-pattern spike stream, an intuitive way is to convert it to raw images for demosaicing. To infer the raw images, we can employ some spike camera reconstruction methods that don't fuse values of spatial neighboring pixels of multiple colors, such as the spike interval-based method (TFI [48]). Specifically, the spike interval of two temporally neighboring spikes that covers the " + }, + { + "bbox": [ + 304, + 588, + 547, + 684 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 588, + 547, + 684 + ], + "type": "text", + "content": "-th spike frame can be formulated as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 345, + 689, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 345, + 689, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Psi_ {n} (x, y) = \\min \\left\\{k \\mid \\mathbf {S} _ {i} (x, y) = 1, k \\geq n \\right\\} \\tag {4} \\\\ - \\max \\{k | \\mathbf {S} _ {i} (x, y) = 1, k < n \\}. \\\\ \\end{array}", + "image_path": "8383e7a6196761581ca54f3d32f99c6321f89b828e5c2d4bd38b93cab1898389.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24873" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 545, + 213 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 545, + 213 + ], + "type": "image", + "image_path": "9d2066606d3f5164ee5e339268ef64638f64e04f2a88bf4a5cfd9f1bebfb20a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 546, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 546, + 251 + ], + "type": "text", + "content": "Figure 3. The overall architecture of CSCSR network. First, the Bayer-pattern spike stream representation (BSSR) is employed to represent the spike stream clip " + }, + { + "bbox": [ + 46, + 217, + 546, + 251 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 217, + 546, + 251 + ], + "type": "text", + "content": ". Then the features are passed to the motion-guided super-resolution (MSR) module based on the temporal neighboring pixels of each color. Finally, the HR color image is restored by our color correlation-based reconstruction (CCR) module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 266, + 288, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 288, + 289 + ], + "type": "text", + "content": "Then the " + }, + { + "bbox": [ + 47, + 266, + 288, + 289 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 266, + 288, + 289 + ], + "type": "text", + "content": "-th Bayer-pattern raw image can be estimated according to the intervals, which can be written as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 294, + 287, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 294, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 123, + 294, + 287, + 317 + ], + "type": "interline_equation", + "content": "\\mathbf {R} _ {n} (x, y) = \\frac {\\delta}{\\Psi_ {n} (x , y)}, \\tag {5}", + "image_path": "5fda5a2df6c27e5ab1024f0e2ee88e15cf33ce190e7d14bffb97e14fbf850ca0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 322, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 287, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 322, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 322, + 287, + 369 + ], + "type": "text", + "content": " denotes the maximum dynamic range. As to a spike number-based method (TFP [48]), the raw image " + }, + { + "bbox": [ + 46, + 322, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_n(x,y)" + }, + { + "bbox": [ + 46, + 322, + 287, + 369 + ], + "type": "text", + "content": " can also be inferred based on the spike number within a temporal window with an odd window length " + }, + { + "bbox": [ + 46, + 322, + 287, + 369 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 322, + 287, + 369 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 374, + 287, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 374, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 97, + 374, + 287, + 403 + ], + "type": "interline_equation", + "content": "\\mathbf {R} _ {n} (x, y) = \\frac {\\delta}{w} \\sum_ {i = n + 1} ^ {n + w} \\mathbf {S} _ {i - \\frac {w + 1}{2}} (x, y). \\tag {6}", + "image_path": "4d5b1f381da093b47b03d2c63ba5f4ac36995718172694c76dc09d25cc362913.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 411, + 102, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 102, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 102, + 423 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 431, + 167, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 167, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 167, + 442 + ], + "type": "text", + "content": "4.1. Overall Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 449, + 288, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 641 + ], + "type": "text", + "content": "To reconstruct HR color images from LR Bayer-pattern spike stream, we develop a CSCSR network whose overall architecture is shown in Fig. 3. The input of the network is a clip of the spike stream with a shape of " + }, + { + "bbox": [ + 46, + 449, + 288, + 641 + ], + "type": "inline_equation", + "content": "N \\times H \\times W" + }, + { + "bbox": [ + 46, + 449, + 288, + 641 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 46, + 449, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 449, + 288, + 641 + ], + "type": "text", + "content": ". To represent color and motion information within the binary input, we first pass it to our proposed Bayer-pattern spike stream representation (BSSR) module. Considering the motion consistency of color channels, we split out four sequences according to the color layout and jointly estimate sub-pixel level motion from them. Guided by the motion information, temporal neighboring pixels of each color channel are searched in the motion-guided super-resolution (MSR) module, resulting in HR temporal-pixel features. Finally, the temporal-pixel features of the three color channels are fused to reconstruct the final HR image, exploiting the correlation of color channels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 647, + 275, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 275, + 660 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 275, + 660 + ], + "type": "text", + "content": "4.2. Bayer-Pattern Spike Stream Representation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "The spike signals in the Bayer-pattern spike stream mean the arrival of a certain photon amount for a certain color channel, which may suffer from quantization errors as introduced in Sec. 3.1. Considering the motion and color" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 266, + 545, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 313 + ], + "type": "text", + "content": "information within the binary data, how to represent it appropriately is worth studying. As shown in Fig. 4, we propose a Bayer-pattern spike stream representation module to represent the information contained in the data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 314, + 546, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 546, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 546, + 457 + ], + "type": "text", + "content": "To keep the color information, we first perform elementwise multiplication between each frame of the spike sequence and the mask of each color, splitting the data into three sequences. For the sequence of each color channel, we develop an encoder to extract the temporal features and retain the relative motion. To be specific, we employ a sliding temporal window to extract the features of each time point. As the impact of quantization errors can be reduced by more temporal information, we extract the features from the whole clip serving as an additional global perception (GP) to the features of each time point. With the temporal window sliding, features of each color are obtained." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 465, + 545, + 535 + ], + "blocks": [ + { + "bbox": [ + 307, + 465, + 545, + 535 + ], + "lines": [ + { + "bbox": [ + 307, + 465, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 307, + 465, + 545, + 535 + ], + "type": "image", + "image_path": "d26537fb7a69453c9d7ba8d807d62dcfa7f308d35a986bf954547a1ff56cb4d4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 542, + 545, + 586 + ], + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 586 + ], + "type": "text", + "content": "Figure 4. Illustration of the Bayer-pattern spike stream representation module, with a temporal sliding window in encoders. In the figure, \"ConvBlock\" denotes a " + }, + { + "bbox": [ + 304, + 542, + 545, + 586 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 542, + 545, + 586 + ], + "type": "text", + "content": " convolution layer followed by a ReLU activation function. \"ResBlock\" refers to [13]." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 599, + 485, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 599, + 485, + 612 + ], + "spans": [ + { + "bbox": [ + 305, + 599, + 485, + 612 + ], + "type": "text", + "content": "4.3. Motion-Guided Super-Resolution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Joint Motion Estimation. Though the spatial pixels are limited, some pixels from the temporal domain are available for SR, which makes it essential to analyze the motion within the Bayer-pattern spike clip. Therefore, we employ the spike camera optical flow method Spike2Flow [46]. To apply the method to the Bayer-pattern spike stream, we split it into four single-channel spike sequences and estimate motion from each channel as shown in Fig. 3, resulting in op" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24874" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 347, + 189 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 347, + 189 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 347, + 189 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 347, + 189 + ], + "type": "image", + "image_path": "435f4ea8aea956ffc3632b0f61cdc700f235c00a85beea121fb16237fb6ec76a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 193, + 547, + 215 + ], + "lines": [ + { + "bbox": [ + 46, + 193, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 547, + 215 + ], + "type": "text", + "content": "Figure 5. Illustration of the motion-guided super-resolution (MSR) module, taking the blue channel as an example. (a) Overview of the MSR module. (b) Motion-guided temporal pixel search, with three steps. (c) Structure of the feature refinement (FR) module." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 349, + 70, + 545, + 188 + ], + "blocks": [ + { + "bbox": [ + 349, + 70, + 545, + 188 + ], + "lines": [ + { + "bbox": [ + 349, + 70, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 349, + 70, + 545, + 188 + ], + "type": "image", + "image_path": "308f45367f4748e58496a9bcc7a9b471fd8f72f08647cc6a9fddcad456fd6023.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 229, + 289, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 289, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 289, + 326 + ], + "type": "text", + "content": "tical flows " + }, + { + "bbox": [ + 46, + 229, + 289, + 326 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}_{j,k}^{C}\\}, C \\in \\{R, G_{1}, G_{2}, B\\}" + }, + { + "bbox": [ + 46, + 229, + 289, + 326 + ], + "type": "text", + "content": ", from the middle time point " + }, + { + "bbox": [ + 46, + 229, + 289, + 326 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 229, + 289, + 326 + ], + "type": "text", + "content": " to the last time point " + }, + { + "bbox": [ + 46, + 229, + 289, + 326 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 229, + 289, + 326 + ], + "type": "text", + "content": ". By reversing the spike sequences, we then estimate optical flows " + }, + { + "bbox": [ + 46, + 229, + 289, + 326 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}_{j,i}^{C}\\}" + }, + { + "bbox": [ + 46, + 229, + 289, + 326 + ], + "type": "text", + "content": " from the middle time point " + }, + { + "bbox": [ + 46, + 229, + 289, + 326 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 229, + 289, + 326 + ], + "type": "text", + "content": " to the first time point " + }, + { + "bbox": [ + 46, + 229, + 289, + 326 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 229, + 289, + 326 + ], + "type": "text", + "content": ". Though coming from different intensities, the color channels of the dynamic scene share the same motion. For more robust motion estimation, we jointly consider the color channels and obtain the final sub-pixel level optical flows as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 333, + 287, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 333, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 85, + 333, + 287, + 363 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {w} _ {j, i}, \\mathbf {w} _ {j, k}\\right) = \\left(\\frac {1}{4} \\sum_ {C} \\mathbf {w} _ {j, i} ^ {C}, \\frac {1}{4} \\sum_ {C} \\mathbf {w} _ {j, k} ^ {C}\\right). \\tag {7}", + "image_path": "2329380942288bc74cf241077bc5a40512cef32ba1d8fe81a82ac6c05346cb38.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 369, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 430 + ], + "type": "text", + "content": "Temporal Pixel Search. To provide features for the final reconstruction of the middle time point, we propose to fill the gap of spatial pixels from LR to HR based on the known temporal pixels at non-middle time points, considering the motion and color layout within the spike stream clip." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": "The positions of known spatial pixels of a certain color can be inferred according to the Bayer-pattern color mask. Denoting the SR scale as " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": ", the target spatial resolution is " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "sH \\times sW" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": ". In the target, one originally known spatial position corresponds to " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "s^2" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": " neighboring positions, resulting in an upsampled color mask as shown in Fig. 5 (b). To get the extra known pixels after upsampling, we apply bilinear interpolation to the features, resulting in the super-resolved features " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": ". Utilizing the estimated optical flows " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "(\\mathbf{w}_{j,i}, \\mathbf{w}_{j,k})" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": ", we can infer the new spatial position of every pixel at the middle time point " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": " when moving to a certain time point " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": ". To be specific, the optical flow " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{j,i}" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": " can be formulated as the pixel offset along " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": " direction and " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": " direction, which can be written as " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "[\\mathbf{U}_j^i, \\mathbf{V}_j^i]" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": ". Considering the spatial upsampling, the offsets should be multiplied by the scale " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": ". Since the time span from time point " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": " is extremely small " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "(0.025\\mathrm{ms} \\times N)" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": ", we assume the motion is linear. As a result, after moving to a certain time point " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": ", the new position of a certain original position " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": " at the middle time point " + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 430, + 289, + 668 + ], + "type": "text", + "content": " can be written as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 672, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 672, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 53, + 672, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} \\left(x + s \\cdot \\frac {t - j}{i - j} \\mathbf {U} _ {j} ^ {i} (x, y), y + s \\cdot \\frac {t - j}{i - j} \\mathbf {V} _ {j} ^ {i} (x, y)\\right), & t < j, \\\\ \\left(x + s \\cdot \\frac {t - j}{k - j} \\mathbf {U} _ {j} ^ {k} (x, y), y + s \\cdot \\frac {t - j}{k - j} \\mathbf {V} _ {j} ^ {k} (x, y)\\right), & t > j. \\end{array} \\right. \\tag {8}", + "image_path": "baa5ad9968d9a8b9dfc11475bdb89a7fb3eeb27296a07594d91a50717ab5e47d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 230, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 545, + 339 + ], + "type": "text", + "content": "According to Eqn. (8), we can map all the pixels at the middle time point to their new positions at any other time point. However, the new positions are usually sub-pixel. Thus we are going to reconstruct the temporal feature value at the sub-pixel position by the values of known integer-pixels according to the upsampled color mask. In particular, the value of the new sub-pixel position " + }, + { + "bbox": [ + 304, + 230, + 545, + 339 + ], + "type": "inline_equation", + "content": "(x_{m},y_{m})" + }, + { + "bbox": [ + 304, + 230, + 545, + 339 + ], + "type": "text", + "content": " at time point " + }, + { + "bbox": [ + 304, + 230, + 545, + 339 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 230, + 545, + 339 + ], + "type": "text", + "content": " can be calculated based on the 4 nearest known pixels " + }, + { + "bbox": [ + 304, + 230, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\{(x_i,y_i)\\}_{i = 1}^4" + }, + { + "bbox": [ + 304, + 230, + 545, + 339 + ], + "type": "text", + "content": ", which can be formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 353, + 349, + 545, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 349, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 353, + 349, + 545, + 378 + ], + "type": "interline_equation", + "content": "\\mathbf {T} (t, x, y) = \\sum_ {i = 1} ^ {4} \\frac {\\phi_ {i}}{\\sum_ {i = 1} ^ {4} \\phi_ {i}} \\mathbf {F} (t, x _ {i}, y _ {i}), \\tag {9}", + "image_path": "63a805bdfe0fdc0c3924f359670c783256be74acc840b7726ba123dbc9437798.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "type": "text", + "content": " denotes the temporal-pixel features, " + }, + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "type": "text", + "content": " denotes the original position at the middle time point, and " + }, + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 305, + 388, + 545, + 424 + ], + "type": "text", + "content": " denotes the weight of each known value, which can be written as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 358, + 433, + 545, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 433, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 358, + 433, + 545, + 459 + ], + "type": "interline_equation", + "content": "\\phi_ {i} = \\frac {1}{\\sqrt {\\left(x _ {m} - x _ {i}\\right) ^ {2} + \\left(y _ {m} - y _ {i}\\right) ^ {2}}}. \\tag {10}", + "image_path": "d103aee0a308729ec6ace3c614c3f38da48275d222c027037ac6732c3ea3e6b4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 468, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 503 + ], + "type": "text", + "content": "After the temporal pixel search process, we can obtain the temporal-pixel features of each color channel for the following HR reconstruction of the middle time point." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 505, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 545, + 624 + ], + "type": "text", + "content": "Feature Refinement. In most cases, there are errors in the estimated motion, causing an undesired impact on the temporal pixel search. Therefore, we propose a feature refinement as shown in Fig. 5 (c). By averaging the input temporal-pixel features " + }, + { + "bbox": [ + 304, + 505, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 505, + 545, + 624 + ], + "type": "text", + "content": " temporally, a rough temporal average result of the channel as a reference can be obtained. Then we compute residuals between the reference and " + }, + { + "bbox": [ + 304, + 505, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 505, + 545, + 624 + ], + "type": "text", + "content": " to represent the pixel fluctuation of the frames. At last, we extract deeper features from the residuals and the based " + }, + { + "bbox": [ + 304, + 505, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 505, + 545, + 624 + ], + "type": "text", + "content": ", and add them together for refined temporal-pixel features " + }, + { + "bbox": [ + 304, + 505, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}" + }, + { + "bbox": [ + 304, + 505, + 545, + 624 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 634, + 518, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 518, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 518, + 646 + ], + "type": "text", + "content": "4.4. Color Correlation-Based Reconstruction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "With the temporal-pixel features, we are going to reconstruct the final HR image. Despite having different intensities, there is a correlation between color channels. Thus we can restore the image by integrating the features, achieving mutual information complementation of color channels." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24875" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 286, + 208 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 208 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 208 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 208 + ], + "type": "image", + "image_path": "13975d8102b76e23a5899aa043b8995970dc4634aa4dee4f39bda9477b092218.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 210, + 287, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 210, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 287, + 255 + ], + "type": "text", + "content": "Figure 6. Illustration of three structures of the reconstruction module. (a) Our proposed color correlation-based reconstruction (CCR) module. (b) A reconstruction module considering no color correlation. (c) An intuitive design considering color correlation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 261, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 261, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 287, + 369 + ], + "type": "text", + "content": "As shown in Fig. 6 (a), we design a reconstruction module utilizing multi-stage features of color channels. In particular, we first apply residual blocks to extract deeper features from the features of each color channel. To fuse the multi-color features of each stage for better aggregation, we develop a color fusion (CF) module to encode the features of color channels with the output of the previous CF stage. After several stages of CF, we employ residual blocks in the end to generate the final image from the fused features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 378, + 128, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 378, + 128, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 378, + 128, + 392 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 397, + 216, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 397, + 216, + 410 + ], + "spans": [ + { + "bbox": [ + 47, + 397, + 216, + 410 + ], + "type": "text", + "content": "5.1. Color Spike Camera Simulator" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 424, + 286, + 538 + ], + "blocks": [ + { + "bbox": [ + 50, + 424, + 286, + 538 + ], + "lines": [ + { + "bbox": [ + 50, + 424, + 286, + 538 + ], + "spans": [ + { + "bbox": [ + 50, + 424, + 286, + 538 + ], + "type": "image", + "image_path": "eff25f3a2bcd92b935d9d2498c874409453d61b66e9d8fb041ab4f0c6ad94dbc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 541, + 286, + 563 + ], + "lines": [ + { + "bbox": [ + 46, + 541, + 286, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 286, + 563 + ], + "type": "text", + "content": "Figure 7. The pipeline of the color spike camera simulator, with the green channel as an example of signal accumulation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "To train models, a large amount of LR spike stream-HR color image pairs are required. However, it's hard to collect corresponding high-quality HR images, especially in scenes with high-speed motion. Inspired by the success of gray-scale spike camera simulators [43, 45], we also developed a simulator for color spike cameras. The pipeline of our simulator is shown in Fig. 7. To generate spike streams, we regard the input video as the dynamic scene to be recorded. As the frame rate of the input video is usually limited, the temporal information is not enough for the simulator of CSC with ultra-high temporal resolution to generate Bayer-pattern spike streams. To address this issue, we" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "use a frame interpolation method [27] to generate latent intensity frames between the original video frames. To generate spike signals, we follow the mechanism introduced in Sec. 3.1 to accumulate light intensity from the latent frames and compare the accumulated value with the threshold periodically for firing spikes. As a result, a Bayer-pattern spike stream is generated from the input video. In particular, the pixel count used to accumulate intensity signals is determined by SR scales. For example, we accumulate signals of 4 pixels to generate the signal of one target pixel in the " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " case. Considering the CFA, the 4 pixels come from one of the channels according to the Bayer-pattern color layout." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 222, + 432, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 222, + 432, + 235 + ], + "spans": [ + { + "bbox": [ + 306, + 222, + 432, + 235 + ], + "type": "text", + "content": "5.2. Experimental Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 240, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 358 + ], + "type": "text", + "content": "Implementation Details. In our implementation, we randomly crop the Bayer-pattern spike frames into patches with spatial resolution " + }, + { + "bbox": [ + 304, + 240, + 545, + 358 + ], + "type": "inline_equation", + "content": "96 \\times 96" + }, + { + "bbox": [ + 304, + 240, + 545, + 358 + ], + "type": "text", + "content": " for training, with the RGGB Bayer pattern. We set the number of spike frames " + }, + { + "bbox": [ + 304, + 240, + 545, + 358 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 240, + 545, + 358 + ], + "type": "text", + "content": " to 41 and the batch size to 4. The length " + }, + { + "bbox": [ + 304, + 240, + 545, + 358 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 240, + 545, + 358 + ], + "type": "text", + "content": " and stride " + }, + { + "bbox": [ + 304, + 240, + 545, + 358 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 240, + 545, + 358 + ], + "type": "text", + "content": " of the sliding temporal window are set to 11 and 3. We employ " + }, + { + "bbox": [ + 304, + 240, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\ell_{1}" + }, + { + "bbox": [ + 304, + 240, + 545, + 358 + ], + "type": "text", + "content": " loss function and use Adam optimizer with an initial learning rate of " + }, + { + "bbox": [ + 304, + 240, + 545, + 358 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 240, + 545, + 358 + ], + "type": "text", + "content": ", which will decay to 0.8 times after every 7500 iterations. Besides, we employ the PyTorch framework and train the models via an NVIDIA RTX3090 GPU." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 360, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 502 + ], + "type": "text", + "content": "Training and Evaluation Datasets. Using our developed CSC simulator, we generated the training dataset based on 240 scenes for training from REDS (120fps) [23], resulting in " + }, + { + "bbox": [ + 304, + 360, + 545, + 502 + ], + "type": "inline_equation", + "content": "5 \\times 240 = 1200" + }, + { + "bbox": [ + 304, + 360, + 545, + 502 + ], + "type": "text", + "content": " clips for each super-resolution scale (" + }, + { + "bbox": [ + 304, + 360, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 304, + 360, + 545, + 502 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 360, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 304, + 360, + 545, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 360, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 304, + 360, + 545, + 502 + ], + "type": "text", + "content": "). Then we generate " + }, + { + "bbox": [ + 304, + 360, + 545, + 502 + ], + "type": "inline_equation", + "content": "5 \\times 30 = 150" + }, + { + "bbox": [ + 304, + 360, + 545, + 502 + ], + "type": "text", + "content": " clips for evaluation based on 30 videos of the REDS evaluation dataset. To demonstrate the generalization to non-REDS-based data, we also generated evaluation datasets from two video SR datasets, Vid4 [22] and Viceo [35]. Besides, we also captured some real-world Bayer-pattern spike streams to verify the performance further. The details of the synthetic evaluation datasets can be found in Table 1." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 310, + 511, + 542, + 560 + ], + "blocks": [ + { + "bbox": [ + 310, + 511, + 542, + 560 + ], + "lines": [ + { + "bbox": [ + 310, + 511, + 542, + 560 + ], + "spans": [ + { + "bbox": [ + 310, + 511, + 542, + 560 + ], + "type": "table", + "html": "
SourceScenesSamplesResolutionScales
REDS [23]30150720×12802, 3, 4
Vid4 [22]44-2, 3, 4
Vimeo [35]78247824448×2564
", + "image_path": "70f12afecdadfbde52c6ca7df6b9955eed1e5af1630e2575880bbb299e0b8364.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 565, + 529, + 576 + ], + "lines": [ + { + "bbox": [ + 320, + 565, + 529, + 576 + ], + "spans": [ + { + "bbox": [ + 320, + 565, + 529, + 576 + ], + "type": "text", + "content": "Table 1. Details of the three synthetic evaluation datasets." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Comparison Methods. As there are no SR methods for CSCs, we first choose the state-of-the-art gray-scale spike camera SR networks SpikeSR-Net [45] and VidarSR [34] for comparison. To adapt the gray-scale methods to the task, we adopt two strategies: using Bayer-pattern spike streams for end-to-end training and handling each color channel separately. We denote the networks with the second strategy as SpikeSR-Net* and VidarSR*. Then we combine a CSC reconstruction method 3DRI [10] and an image/video SR method SwinIR [18]/BasicVSR [3], resulting in another two comparison methods 3DRI+SwinIR and 3DRI+BasicVSR." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24876" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 71, + 544, + 460 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 544, + 460 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 544, + 460 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 544, + 460 + ], + "type": "table", + "html": "
ScaleMethodREDS-based DatasetVid4-based DatasetVimeo-based Dataset
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
×2TFI [48] +TSCNN [4]29.24dB0.81730.312625.06dB0.77860.3196---
TFP[48] +TSCNN27.43dB0.75650.367124.60dB0.76660.3142---
TFI+Real-RawVSR [36]31.03dB0.86270.247826.79dB0.84220.2508---
TFP+Real-RawVSR30.81dB0.85350.260127.10dB0.85640.2329---
3DRI [10] +SwinIR [18]31.13dB0.88090.219126.57dB0.86680.2100---
3DRI+BasicVSR [3]31.60dB0.89120.205227.03dB0.87480.2021---
VidarSR [34]30.81dB0.89910.188426.44dB0.89110.1791---
VidarSR*30.27dB0.84300.263125.28dB0.81060.2538---
SpikeSR-Net [45]32.38dB0.89130.209928.34dB0.89570.1833---
SpikeSR-Net*29.66dB0.82490.286124.62dB0.78540.2721---
CSCSR (ours)33.39dB0.91210.176428.77dB0.91100.1612---
×3TFI+TSCNN27.22dB0.75050.384222.91dB0.66090.4058---
TFP+TSCNN26.55dB0.71760.413723.22dB0.68000.3808---
TFI+Real-RawVSR28.68dB0.79420.329224.00dB0.72250.3474---
TFP+Real-RawVSR28.66dB0.78830.334424.39dB0.74070.3326---
3DRI+SwinIR28.38dB0.81360.297723.97dB0.75340.3160---
3DRI+BasicVSR28.65dB0.82080.290524.08dB0.75680.3135---
VidarSR29.36dB0.84140.267924.66dB0.80090.2816---
VidarSR*27.53dB0.76700.355422.78dB0.67420.3680---
SpikeSR-Net29.79dB0.82700.297025.33dB0.79150.2989---
SpikeSR-Net*26.73dB0.73670.390822.12dB0.63230.4067---
CSCSR (ours)29.92dB0.85130.265325.68dB0.82400.2761---
×4TFI+TSCNN25.98dB0.70100.431021.69dB0.57220.465924.77dB0.73040.3554
TFP+TSCNN26.55dB0.71760.413722.53dB0.62400.417725.56dB0.75870.3557
TFI+Real-RawVSR27.27dB0.73910.385122.53dB0.62400.417725.12dB0.74920.3244
TFP+Real-RawVSR27.32dB0.73470.388522.81dB0.64080.363425.15dB0.74930.3324
3DRI+SwinIR26.94dB0.75440.360222.47dB0.65330.390125.21dB0.74230.3218
3DRI+BasicVSR26.94dB0.75700.360522.52dB0.65070.397325.05dB0.74010.3204
VidarSR27.56dB0.76920.350522.69dB0.67120.386825.19dB0.75650.3098
VidarSR*26.51dB0.71350.413521.58dB0.58800.439524.64dB0.73520.3423
SpikeSR-Net28.28dB0.77550.351723.59dB0.69830.372725.27dB0.76110.3119
SpikeSR-Net*25.93dB0.69130.440321.01dB0.54510.471824.46dB0.72400.3593
CSCSR (ours)28.77dB0.79630.325323.79dB0.72230.355525.34dB0.76600.2967
", + "image_path": "c3d8ad461bb41a5d7b3f8dd15a40223401727da929a26e501642fd0eeb89aa2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 463, + 546, + 486 + ], + "lines": [ + { + "bbox": [ + 47, + 463, + 546, + 486 + ], + "spans": [ + { + "bbox": [ + 47, + 463, + 546, + 486 + ], + "type": "text", + "content": "Table 2. Quantitative comparison. Red and blue indicate the best and the second-best performance, respectively. For each scene of Video [35], there is a sequence of LR frames but only one corresponding HR frame " + }, + { + "bbox": [ + 47, + 463, + 546, + 486 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 47, + 463, + 546, + 486 + ], + "type": "text", + "content": ". Thus only the dataset of " + }, + { + "bbox": [ + 47, + 463, + 546, + 486 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 47, + 463, + 546, + 486 + ], + "type": "text", + "content": " scale is generated for Video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 498, + 288, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 594 + ], + "type": "text", + "content": "It's also intuitive to combine methods of estimating raw images from Bayer-pattern spike streams (Sec. 3.2) and joint demosaicing and super-resolution (JDSR) methods. Thus we employ TFI or TFP [48] to infer raw images, followed by an image JDSR method TSCNN [4] or a video JDSR method Real-RawVSR [36] to reconstruct color images. In addition, we employ PSNR, SSIM [30] and LPIPS [38] (lower means better) as quantitative metrics for comparison." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 599, + 168, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 168, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 168, + 612 + ], + "type": "text", + "content": "5.3. Comparative Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "Quantitative Results. As shown in Table 2, we conduct comparative experiments on the three SR scales. It can be found that our method achieves the best quantitative results in most experiments. The state-of-the-art spike camera SR methods VidarSR and SpikeSR-Net achieve competitive performance. However, VidarSR* and SpikeSR-Net* that handle each color channel separately don't perform as well as VidarSR and SpikeSR-Net, due to the lack of considera" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 498, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 582 + ], + "type": "text", + "content": "tion of color correlation. With the temporal information, the multi-frame method 3DRI+BasicVSR performs better than the single-frame method 3DRI+SwinIR in most cases. This is also true for the multi-frame JDSR method RawVSR and the single-frame JDSR method TSCNN. Besides, with the increase of the SR scale, TFP-based methods show growing competitiveness to TFI-based methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": "Visual Results. Fig. 8 presents the visual comparison " + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": " on the synthetic data. The visual quality of results by our proposed method is better than the other methods, with less motion blur and more accurate color. There are artifacts in the results produced by VidarSR* and SpikeSR-Net*, which demonstrates the importance of the consideration of color correlation. In addition, the multi-frame methods show better visual quality. To demonstrate model generalization, we conduct experiments on real-world captured Bayer-pattern spike streams as shown in Fig. 9. In contrast, our method can restore better textures and details." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24877" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 547, + 195 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 547, + 195 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 547, + 195 + ], + "type": "image", + "image_path": "162dd7e3ffc318ca8a901b5537b8712e68d331223f48887cc9bbb89a24eabf8e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 197, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 197, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 547, + 220 + ], + "type": "text", + "content": "Figure 8. Visual comparison " + }, + { + "bbox": [ + 46, + 197, + 547, + 220 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 46, + 197, + 547, + 220 + ], + "type": "text", + "content": " on the synthetic Bayer-pattern spike streams. The first sample comes from the REDS-based dataset, while the second one comes from the Vid4-based dataset. Please enlarge the figure for better comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 224, + 545, + 355 + ], + "blocks": [ + { + "bbox": [ + 48, + 224, + 545, + 355 + ], + "lines": [ + { + "bbox": [ + 48, + 224, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 48, + 224, + 545, + 355 + ], + "type": "image", + "image_path": "8fc7be26a481a16f935452ccaa5173c4daf18051e6f369b1dce80f9aa338060d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 357, + 546, + 379 + ], + "lines": [ + { + "bbox": [ + 46, + 357, + 546, + 379 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 546, + 379 + ], + "type": "text", + "content": "Figure 9. Visual comparison " + }, + { + "bbox": [ + 46, + 357, + 546, + 379 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 46, + 357, + 546, + 379 + ], + "type": "text", + "content": " on the real-world Bayer-pattern spike streams. The first spike stream sample is captured by a fast-moving CSC. The second one records a fast-rotating water bottle. Please enlarge the figure for better comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 52, + 391, + 286, + 516 + ], + "blocks": [ + { + "bbox": [ + 52, + 391, + 286, + 516 + ], + "lines": [ + { + "bbox": [ + 52, + 391, + 286, + 516 + ], + "spans": [ + { + "bbox": [ + 52, + 391, + 286, + 516 + ], + "type": "table", + "html": "
CaseSetting Description×2×3×4
(A)Removing BSSR module33.0429.5728.57
(B)Removing GP from BSSR33.1729.7428.68
(C)Interval-based representation32.0528.1928.06
(D)Independent motion estimation33.2229.5528.59
(E)Assembled motion estimation30.2427.0826.66
(F)Removing MSR module32.6629.4228.39
(G)Removing FR from MSR33.0729.7028.56
(H)Reconstruction module (b)30.1927.3626.22
(I)Reconstruction module (c)33.2429.6928.66
(J)Our final network33.3929.9228.77
", + "image_path": "bde87938c9df89e0bb21ea90c2d542a6b8b99037d1414685da478fb6455ba7f9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 518, + 287, + 541 + ], + "lines": [ + { + "bbox": [ + 47, + 518, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 287, + 541 + ], + "type": "text", + "content": "Table 3. Ablation study on the REDS-based evaluation dataset. Greener blocks represent higher PSNR(dB) performance." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 552, + 141, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 141, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 141, + 565 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "To verify the effectiveness of our methodology, we perform ablation studies on our modules and motion estimation strategy. Case (J) is our final CSCSR network. First, we remove the BSSR module in (A) to demonstrate the proposed representation. Then we remove the additional GP branch from BSSR in (B). We also replace BSSR with the interval-based representation [46, 48] in (C). To demonstrate our joint motion estimation strategy, we employ independent motion estimation for each color channel in (D) and an optical flow assembly strategy in (E). To be specific, the optical flows of color channels are assembled according to the Bayer pattern. After that, we implement studies on the MSR" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 391, + 547, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 547, + 464 + ], + "type": "text", + "content": "module. In (F), MSR is removed from the network. In (G), the FR module is removed from MSE to verify its effectiveness. Finally, (H) and (I) are the studies on the proposed reconstruction module, where the module is replaced by the structures in (b) and (c) of Fig. 6, with the same number of residual blocks. The study results are shown in Table 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 472, + 384, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 472, + 384, + 485 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 384, + 485 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 492, + 547, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 547, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 547, + 636 + ], + "type": "text", + "content": "In this paper, we present a deep network to restore HR color images from LR Bayer-pattern spike streams. To represent the binary data, we develop a representation that utilizes local temporal information with global perception. Then we propose to collect temporal pixels for spatial super-resolution according to the color layout and sub-pixel level motion. To reduce the impact of motion estimation errors, we design a feature refinement module based on residuals. Finally, multi-stage temporal-pixel features of each color channel are jointly considered, resulting in the final HR color image. Experimental results on both synthetic and real-world captured data demonstrate our performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 646, + 403, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 403, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 403, + 659 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grants 62072009, 22127807, and in part by the National Key R&D Program of China under Grant 2021YFF0900501." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24878" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Patrick Bardow, Andrew J Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 884-892, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 200 + ], + "type": "text", + "content": "[2] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Real-time video super-resolution with spatio-temporal networks and motion compensation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4778-4787, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 257 + ], + "type": "text", + "content": "[3] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. BasicVSR: The search for essential components in video super-resolution and beyond. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4947-4956, 2021. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 287, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 311 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 311 + ], + "type": "text", + "content": "[4] Kan Chang, Hengxin Li, Yufei Tan, Pak Lun Kevin Ding, and Baoxin Li. A two-stage convolutional neural network for joint demosaicking and super-resolution. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4238-4254, 2021. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 312, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 312, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 312, + 287, + 367 + ], + "type": "text", + "content": "[5] Yakun Chang, Chu Zhou, Yuchen Hong, Liwen Hu, Chao Xu, Tiejun Huang, and Boxin Shi. 1000 FPS HDR video with a spike-rgb hybrid camera. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 22180-22190, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 287, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 411 + ], + "type": "text", + "content": "[6] Jonghyun Choi, Kuk-Jin Yoon, et al. Learning to super resolve intensity images from events. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2768-2776, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 413, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 287, + 456 + ], + "type": "text", + "content": "[7] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Proceedings of European Conference on Computer Vision (ECCV), 2014. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 287, + 490 + ], + "type": "text", + "content": "[8] Siwei Dong, Tiejun Huang, and Yonghong Tian. Spike camera and its coding methods. In Data Compression Conference (DCC), page 437, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 287, + 535 + ], + "type": "text", + "content": "[9] Siwei Dong, Lin Zhu, Daoyuan Xu, Yonghong Tian, and Tiejun Huang. An efficient coding method for spike camera using inter-spike intervals. In Data Compression Conference (DCC), page 568, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "type": "text", + "content": "[10] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. 3D residual interpolation for spike camera demosaicing. In 2022 IEEE International Conference on Image Processing (ICIP), pages 1461-1465. IEEE, 2022. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 634 + ], + "type": "text", + "content": "[11] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. High-speed scene reconstruction from low-light spike streams. In 2022 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[12] Peiqi Duan, Zihao W Wang, Xinyu Zhou, Yi Ma, and Boxin Shi. EventZoom: Learning to denoise and super resolve neuromorphic events. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12824-12833, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "text", + "content": "[14] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3224-3232, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "text", + "content": "[15] Armin Kappeler, Seunghwan Yoo, Qiqin Dai, and Aggelos K Katsaggelos. Video super-resolution with convolutional neural networks. IEEE Transactions on Computational Imaging, 2(2):109-122, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "text", + "content": "[16] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew J Davison. Simultaneous mosaicing and tracking with an event camera. J. Solid State Circ, 43:566-576, 2008. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 239, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 270 + ], + "type": "text", + "content": "[17] Hongmin Li, Guoqi Li, and Luping Shi. Super-resolution of spatiotemporal event-stream image. Neurocomputing, 335: 206-214, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 272, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 326 + ], + "type": "text", + "content": "[18] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "text", + "content": "[19] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A " + }, + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "text", + "content": " 120 db " + }, + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "inline_equation", + "content": "15\\mu s" + }, + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "text", + "content": " latency asynchronous temporal contrast vision sensor. IEEE journal of solid-state circuits, 43 (2):566-576, 2008. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "text", + "content": "[20] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In IEEE conference on computer vision and pattern recognition workshops (CVPRW), pages 136-144, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 426, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 502 + ], + "type": "text", + "content": "[21] Martin Litzenberger, Christoph Posch, D Bauer, Ahmed Nabil Belbachir, P Schon, B Kohn, and H Garn. Embedded vision system for real-time object tracking using an asynchronous transient vision sensor. In 2006 IEEE 12th Digital Signal Processing Workshop & 4th IEEE Signal Processing Education Workshop, pages 173-178. IEEE, 2006. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "text", + "content": "[22] Ce Liu and Deqing Sun. On bayesian adaptive video super resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(2):346-360, 2013. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 537, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 602 + ], + "type": "text", + "content": "[23] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. NTIRE 2019 challenge on video deblurring and superresolution: Dataset and study. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 0–0, 2019. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "type": "text", + "content": "[24] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2010. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[25] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3857-3866, 2019. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24879" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[26] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast image reconstruction with an event camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 156-163, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "text", + "content": "[27] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. XVFI: extreme video frame interpolation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14489-14498, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 218 + ], + "type": "text", + "content": "[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 274 + ], + "type": "text", + "content": "[29] Lin Wang, Tae-Kyun Kim, and Kuk-Jin Yoon. EventSR: From asynchronous events to image reconstruction, restoration, and super-resolution via end-to-end adversarial learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8315-8325, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 319 + ], + "type": "text", + "content": "[30] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 288, + 385 + ], + "type": "text", + "content": "[31] Zihao W. Wang, Peiqi Duan, Oliver Cossairt, Aggelos Katsaggelos, Tiejun Huang, and Boxin Shi. Joint filtering of intensity images and neuromorphic events for high-resolution noise-robust imaging. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1609–1619, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "type": "text", + "content": "[32] Lujie Xia, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. SVFI: Spiking-based video frame interpolation for high-speed motion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2910-2918, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 434, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 487 + ], + "type": "text", + "content": "[33] Lujie Xia, Ziluo Ding, Rui Zhao, Jiyuan Zhang, Lei Ma, Zhaofei Yu, Tiejun Huang, and Ruiqin Xiong. Unsupervised optical flow estimation with dynamic timing representation for spike camera. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 489, + 288, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 288, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 288, + 543 + ], + "type": "text", + "content": "[34] Xijie Xiang, Lin Zhu, Jianing Li, Yixuan Wang, Tiejun Huang, and Yonghong Tian. Learning super-resolution reconstruction for high temporal resolution spike stream. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 544, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 578 + ], + "type": "text", + "content": "[35] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. Int. J. Comput. Vis, 127:1106-1125, 2019. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 288, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 623 + ], + "type": "text", + "content": "[36] Huanjing Yue, Zhiming Zhang, and Jingyu Yang. Realrawvsr: Real-world raw video super-resolution with a benchmark dataset. In European Conference on Computer Vision, pages 608-624. Springer, 2022. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 288, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 677 + ], + "type": "text", + "content": "[37] Jiyuan Zhang, Shanshan Jia, Zhaofei Yu, and Tiejun Huang. Learning temporal-ordered representation for spike streams based on discrete wavelet transforms. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 137–147, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[38] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[39] Yiyang Zhang, Ruiqin Xiong, and Tiejun Huang. Spike signal reconstruction based on inter-spike similarity. In 2022 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "text", + "content": "[40] Jing Zhao, Ruiqin Xiong, and Tiejun Huang. High-speed motion scene reconstruction for spike camera via motion aligned filtering. In IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-5, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 185, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 238 + ], + "type": "text", + "content": "[41] Jing Zhao, Ruiqin Xiong, Rui Zhao, Jin Wang, Siwei Ma, and Tiejun Huang. Motion estimation for spike camera data sequence via spike interval analysis. In 2020 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 371-374. IEEE, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "text", + "content": "[42] Jing Zhao, Jiyu Xie, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, and Tiejun Huang. Super resolve dynamic scene from continuous spike streams. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2533-2542, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "type": "text", + "content": "[43] Jing Zhao, Ruiqin Xiong, Hangfan Liu, Jian Zhang, and Tiejun Huang. Spk2ImgNet: Learning to reconstruct dynamic scene from continuous spike stream. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 11996-12005, 2021. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[44] Jing Zhao, Ruiqin Xiong, Jiyu Xie, Boxin Shi, Zhaofei Yu, Wen Gao, and Tiejun Huang. Reconstructing clear image for high-speed motion scene with a retina-inspired spike camera. IEEE Transactions on Computational Imaging, 8:12-27, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 404, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 458 + ], + "type": "text", + "content": "[45] Jing Zhao, Ruiqin Xiong, Jian Zhang, Rui Zhao, Hangfan Liu, and Tiejun Huang. Learning to super-resolve dynamic scenes for neuromorphic spike camera. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3579-3587, 2023. 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "text", + "content": "[46] Rui Zhao, Ruiqin Xiong, Jing Zhao, Zhaofei Yu, Xiaopeng Fan, and Tiejun Huang. Learning optical flow from continuous spike streams. Advances in Neural Information Processing Systems, 35:7905-7920, 2022. 4, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 505, + 545, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 556 + ], + "type": "text", + "content": "[47] Rui Zhao, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, Shuyuan Zhu, Lei Ma, and Tiejun Huang. Spike camera image reconstruction using deep spiking neural networks. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "text", + "content": "[48] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. A retina-inspired sampling method for visual texture reconstruction. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1432-1437. IEEE, 2019. 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "text", + "content": "[49] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. Hybrid coding of spatiotemporal spike data for a bio-inspired camera. IEEE Transactions on Circuits and Systems for Video Technology, 31(7):2837-2851, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "text", + "content": "[50] Lin Zhu, Jianing Li, Xiao Wang, Tiejun Huang, and Yonghong Tian. NeuSpike-Net: High speed video reconstruction via bio-inspired neuromorphic cameras. In IEEE International Conference on Computer Vision (ICCV), pages 2400–2409, 2021. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24880" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/a4b33f00-d1f9-45e2-a62c-bb25956b58ec_content_list.json b/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/a4b33f00-d1f9-45e2-a62c-bb25956b58ec_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..14dbdf895896173ba2b7baefc112f1c5724a9508 --- /dev/null +++ b/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/a4b33f00-d1f9-45e2-a62c-bb25956b58ec_content_list.json @@ -0,0 +1,1627 @@ +[ + { + "type": "text", + "text": "SuperNormal: Neural Surface Reconstruction via Multi-View Normal Integration", + "text_level": 1, + "bbox": [ + 251, + 130, + 718, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xu Cao Takafumi Taketomi", + "bbox": [ + 362, + 203, + 604, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CyberAgent", + "bbox": [ + 434, + 222, + 535, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xu_cao, taketomi_takafumi}@cyberagent.co.jp", + "bbox": [ + 294, + 241, + 674, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present SuperNormal, a fast, high-fidelity approach to multi-view 3D reconstruction using surface normal maps. With a few minutes, SuperNormal produces detailed surfaces on par with 3D scanners. We harness volume rendering to optimize a neural signed distance function (SDF) powered by multi-resolution hash encoding. To accelerate training, we propose directional finite difference and patch-based ray marching to approximate the SDF gradients numerically. While not compromising reconstruction quality, this strategy is nearly twice as efficient as analytical gradients and about three times faster than axis-aligned finite difference. Experiments on the benchmark dataset demonstrate the superiority of SuperNormal in efficiency and accuracy compared to existing multi-view photometric stereo methods. On our captured objects, SuperNormal produces more fine-grained geometry than recent neural 3D reconstruction methods. Our code is available at https://github.com/CyberAgentAILab/SuperNormal.", + "bbox": [ + 75, + 324, + 473, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 630, + 209, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recovering high-quality 3D geometry of real-world objects from their multi-view images is a long-standing challenge in computer vision. Recently, neural implicit surface-based methods have shown remarkable reconstruction results. Compared to traditional multi-view stereo (MVS) methods [25], neural methods are more robust to view-dependent observations and textureless surfaces [29]. Furthermore, the reconstruction procedure has become highly efficient [30] by introducing multi-resolution hash coding [20]. However, even though this spatial encoding allows fine-grained geometry to be represented, shape reconstruction from multi-view images tends to smooth out high-frequency surface details, as shown in Fig. 1 bottom right.", + "bbox": [ + 75, + 657, + 468, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-view photometric stereo (MVPS), on the other hand, aims to recover pixel-level high-frequency surface detail by introducing additional lighting conditions during", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd4a7330b81a55a9ab69dd7d0dd6a582b2dafb4ed9a6f47cfb78157305ba94d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 290, + 890, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c479ca7ec93608d18f5b3a6564724cec7e349ef64abcbf2f3db03bd6dae3ed92.jpg", + "image_caption": [ + "Figure 1. (Top) From multi-view surface normal maps, SuperNormal recovers fine-grained surface details comparable to 3D scanners while being orders of magnitude faster than the existing MVPS method PS-NeRF [31]. (Bottom) Using normal maps produces more faithful high-frequency details than the MVS-based method NeuS2 [30], although both use multi-resolution hash encoding [20]." + ], + "image_footnote": [], + "bbox": [ + 506, + 424, + 890, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the image acquisition [9]. In a typical workflow, a surface normal map is first recovered for each view, recording per-pixel surface orientation information. The normal maps estimated at different viewpoints are then fused into a 3D model, also known as multi-view normal integration [5]. However, existing normal fusion methods struggle to reflect the details of the normal maps in the shape of the recovered 3D model. Moreover, the reconstruction process takes hours even when only a few low-resolution normal maps are used [4, 31]. Due to the lack of a fast and accurate multiview normal fusion method, current multi-view photometric stereo results remain unsatisfactory.", + "bbox": [ + 496, + 672, + 890, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents SuperNormal to unite the best of both worlds. Normal maps with pixel-level surface details are utilized to exploit the expressive power and efficiency", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20581", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of multi-resolution hash encoding-powered neural implicit surface. Using normal maps alone avoids per-scene reflectance modeling and optimization, thus improving training efficiency compared to using color images. In a few minutes, our method can produce fine-grained geometry on par with commercial 3D scanners.", + "bbox": [ + 75, + 90, + 468, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we use volume rendering to optimize a neural signed distance function (SDF) such that its rendered normal maps are consistent with input normal maps. To further improve the training efficiency, we propose a directional finite difference and patch-based ray marching strategy to approximate the SDF gradients numerically. This strategy kills two birds with one stone: SDF values evaluated at sampled points are all used both for SDF-based volume rendering and SDF gradient approximation. This way, we avoid second-order derivative computation during training. Unlike traditional axis-aligned finite difference, directional finite difference avoids redundant SDF evaluations (Fig. 2). Without compromising reconstruction quality, our method is nearly twice as fast as using analytical gradients and about three times faster than finite difference.", + "bbox": [ + 75, + 181, + 470, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to existing neural MVPS methods, our method is two orders of magnitude faster and recovers better surface details (Fig. 1 above). We also capture objects with complex geometry and show that our method recovers more faithful surface details than the recent neural 3D reconstruction method [30] (Fig. 1 bottom). Taking advantage of recent advances in photometric stereo, we can estimate high-quality normal maps from images captured under common lighting conditions with inexpensive equipment [10]. Combined with SuperNormal, high-fidelity 3D objects can be reconstructed using inexpensive equipment in widely accessible environments.", + "bbox": [ + 75, + 407, + 468, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, this paper's key contributions are:", + "bbox": [ + 96, + 589, + 406, + 604 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- An effective neural 3D reconstruction approach using multi-view normal maps, enhancing the reconstruction quality to near 3D scanner levels;", + "- An efficient method for computing numerical gradients in SDF-based neural rendering, speeding up the reconstruction process; and", + "- A comprehensive evaluation using benchmark objects with MVPS approaches and an assessment using our captured objects with MVS approaches." + ], + "bbox": [ + 76, + 604, + 468, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 762, + 214, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multi-view photometric stereo (MVPS)", + "text_level": 1, + "bbox": [ + 76, + 786, + 411, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MVPS aims at high-fidelity 3D reconstruction using multi-view observations under varying light conditions [9] (usually one-light-at-time (OLAT) images [16, 26]).", + "bbox": [ + 75, + 809, + 468, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing methods differ in what geometry-related information is inferred in each view and how information from different views is fused into a complete 3D model. Chang et", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cfac92fe142e4ba7f830a110d9ce6d48012fe1d807e03238eef7b829e7827220.jpg", + "image_caption": [ + "Figure 2. (Left) In SDF-based volume rendering, approximating SDF gradients using axis-aligned finite difference [18, 36] requires additional SDF samples at neighboring positions of volume rendering points, increasing computational complexity and training time. (Right) Our strategy of patch-based ray marching and directional finite difference effectively uses all SDF samples for both SDF gradient approximation and volume rendering." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 890, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "al. [5] proposes a level set method [22] to fuse multi-view normal maps into a volumetric representation of the shape. R-MVPS [23] uses the multi-view normal maps to refine a coarse mesh estimated by multi-view stereo. B-MVPS [16] initialize the shape as a sparse point cloud estimated by structure from motion (SfM), then propagates the known point spatially using the normal maps.", + "bbox": [ + 496, + 380, + 890, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent MVPS methods have undergone a paradigm shift since the advent of NeRF [19] by using a neural implicit scene representation, usually parameterized as a multi-layer perception (MLP). PS-NeRF [31] recovers a neural field based on UNISURF [21] by regularizing the gradients of the neural field using the normal maps. UA-MVPS [12] fuses the multi-view depth and normal maps into a neural SDF using neural inverse rendering while considering per-pixel confidence. MVAS [4] shows utilizing the azimuth components of multi-view normal maps is sufficient to recover the neural implicit surface. While most methods only use images in the same view for normal map estimation, MVPSNet [35] utilizes neighboring views to infer per-view depth and normal maps jointly and applies screened Poisson surface reconstruction (sPSR) [13] to obtain the shape.", + "bbox": [ + 496, + 487, + 892, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the reconstructed surfaces of these methods tend to be smoothed and do not reflect the high-frequency details of the input normal maps. Our approach introduces multi-resolution hash encoding and directional finite difference to improve the implicit neural field's expressive power and training efficiency.", + "bbox": [ + 496, + 714, + 890, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Neural 3D reconstruction", + "text_level": 1, + "bbox": [ + 500, + 816, + 732, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The seminal work NeRF [19] and follow-up works [2, 3, 27] have demonstrated remarkable novel view synthesis quality by volume rendering a neural implicit field. For better geometry reconstruction quality, IDR [32] imposes eikonal", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20582", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "regularization [8] such that the neural field approximates a valid SDF. Later, the concurrent works VolSDF [33] and NeuS [29] develop SDF-based volume rendering, which bridges SDF values to volume rendering weights and improves the geometry fidelity. However, these methods require several hours to recover one object due to optimizing a deep coordinate-based MLP.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-resolution hash encoding has been proposed to improve expressive power and accelerate the learning of neural implicit field [20]. Several efforts have been made to incorporate this expressive spatial encoding into neural SDF learning. Since the official CUDA implementation does not support the double propagation deduced by the eikonal regularization, PermutoSDF [24] and NeuS2 [30] re-implement the CUDA programming. InstantNSR [36] and Neuralangelo [18] use finite difference instead of automatic differentiation to avoid double propagation.", + "bbox": [ + 75, + 198, + 470, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose directional finite difference, a numerical way to compute SDF gradients that avoids double propagation and is even faster than automatic differentiation when combined with a tailored patch-based sampling strategy.", + "bbox": [ + 76, + 351, + 468, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Neural 3D reconstruction with normals", + "text_level": 1, + "bbox": [ + 76, + 426, + 413, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Normal maps have been used as auxiliary information to facilitate neural scene reconstruction in MonoSDF [34] and NeuRIS [28]. These works target scene-level reconstruction, and normal maps help infer the geometry of textureless smooth regions (e.g., walls) where using color images is likely to break down.", + "bbox": [ + 75, + 450, + 468, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead, our method is particularly useful for single objects with complex surface details. We use photometric stereo [10] to estimate normal maps from images under varying light conditions. This produces high-quality normal maps that can be used alone for shape reconstruction.", + "bbox": [ + 75, + 542, + 468, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 637, + 184, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to recover the surface given a set of normal maps, object masks, and camera intrinsic and extrinsic parameters. Since photometric stereo methods estimate the normal maps in camera space, we first apply the camera-to-world rotation to normal maps to obtain world-space normal maps. Figure 3 depicts our pipeline. We represent the geometry as a parametric SDF and optimize SDF parameters such that its rendered gradient and opacity values are consistent with input normal vectors and mask values, respectively.", + "bbox": [ + 75, + 664, + 468, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Pipeline", + "text_level": 1, + "bbox": [ + 76, + 815, + 174, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SDF-based volume rendering (Forward pass) A signed distance function (SDF) implicitly captures the geometry by assigning each spatial point with the signed distance to its closest surface point. The surface $\\mathcal{M}$ can then be repre", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sented as the zero-level-set of the SDF:", + "bbox": [ + 500, + 90, + 759, + 104 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} = \\{\\mathbf {x} \\mid f (\\mathbf {x}) = 0 \\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 122, + 890, + 138 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We parameterize the SDF using multi-resolution hash encoding followed by a shallow MLP. In multi-resolution hash encoding, the 3D space is discretized into a set of multi-resolution 3D grids. Each grid cell is associated with a unique hash value mapping the 3D coordinates to a learnable feature vector. Specifically, $h(\\mathbf{x};\\phi)$ is the feature vector obtained by concatenating feature vectors of different levels $h(\\mathbf{x}) = [\\mathbf{x},h_1(\\mathbf{x}),\\dots,h_L(\\mathbf{x})]$ . Then, the SDF can be parameterized as", + "bbox": [ + 496, + 154, + 890, + 289 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf (\\mathbf {x}) = f (h (\\mathbf {x}; \\phi); \\theta), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 306, + 890, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\phi$ and $\\theta$ are learnable hash-encoding feature vectors and MLP parameters. Multi-resolution hash encoding has proven to speed up training and improve the expressive power of the neural SDF [20, 30, 36].", + "bbox": [ + 496, + 338, + 890, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During training, we evaluate the SDF values at points sampled on the rays cast from the camera centers toward the scene. NeuS [29] proposes an unbiased and occlusion-aware way to transfer the SDF samples to volume rendering opacities. Given $N$ ordered 3D points $\\{\\mathbf{x}_i\\}_{i=0}^N$ on a ray and their SDF values $\\{f(\\mathbf{x}_i)\\}_{i=0}^N$ , the opacity at each point is designed as", + "bbox": [ + 496, + 400, + 890, + 506 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = \\max \\left(\\frac {\\phi_ {s} (f (\\mathbf {x} _ {i})) - \\phi_ {s} (f (\\mathbf {x} _ {i + 1}))}{\\phi_ {s} (f (\\mathbf {x} _ {i}))}, 0\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 520, + 890, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\phi_s(x) = \\frac{1}{1 + \\exp(-sx)}$ is the sigmoid function with a trainable sharpness $s$ .", + "bbox": [ + 496, + 569, + 890, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With this SDF-based opacity, we can render the normal and opacity for the pixel $\\mathbf{p}$ as:", + "bbox": [ + 496, + 604, + 890, + 635 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {n}} (\\mathbf {p}) = \\sum_ {i = 0} ^ {N} T _ {i} \\alpha_ {i} \\nabla f (\\mathbf {x} _ {i}), \\quad \\hat {o} (\\mathbf {p}) = \\sum_ {i = 0} ^ {N} T _ {i} \\alpha_ {i} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 648, + 890, + 699 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w i t h} \\quad T _ {i} = \\prod_ {j = 0} ^ {i - 1} (1 - \\alpha_ {j}).\n$$\n", + "text_format": "latex", + "bbox": [ + 674, + 694, + 844, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From Eqs. (3) and (4), the neural SDF must be evaluated at least once at sampled points for volume rendering. In Secs. 3.2 and 3.3, we show that these SDF samples can be reused for SDF gradient approximation, thus reducing the computational amount and speeding up the training.", + "bbox": [ + 496, + 750, + 890, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural SDF Training (Backward pass) To train the neural SDF, we supervise the rendered normals and opacities using input ones $\\mathbf{n}(\\mathbf{p})$ and $\\hat{o} (\\mathbf{p})$ . The loss function consists", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "footer", + "text": "20583", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8a20c984d4c8cc382b0dd432ebb04c4d3a096b69675c56f6399dfe04c36dc528.jpg", + "image_caption": [ + "Figure 3. Approach overview. (a) Given input views, we randomly sample patches of pixels, cast rays from those pixels, and march a plane along the patch of rays (Sec. 3.3). We treat the ray-plane intersections as sampled points and evaluate their SDF values via the neural SDF. (b) SDF samples neighbor to a point are used to compute the SDF gradients using directional finite difference (Sec. 3.2). (c) SDF samples on the same rays are used for SDF-based volume rendering (Sec. 3.1)." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 890, + 276 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of three terms: The normal, the mask, and the eikonal term:", + "bbox": [ + 76, + 359, + 468, + 373 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\overbrace {\\sum_ {\\mathbf {p}} \\left(\\hat {\\mathbf {n}} (\\mathbf {p}) - \\mathbf {n} (\\mathbf {p})\\right) ^ {2}} ^ {\\text {n o r m a l}} + \\lambda_ {1} \\overbrace {\\sum_ {\\mathbf {p}} \\operatorname {B C E} \\left(\\hat {o} (\\mathbf {p}) , o (\\mathbf {p})\\right)} ^ {\\text {m a s k}} \\tag {5} \\\\ + \\lambda_ {2} \\underbrace {\\sum_ {\\mathbf {x}} \\left(\\| \\nabla f (\\mathbf {x}) \\| _ {2} - 1\\right) ^ {2}} _ {e i k o n a l}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 397, + 468, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{p}$ is the sampled pixels in a batch, $\\mathbf{x}$ is the sampled points on rays, and BCE is the binary cross entropy function. The normal term encourages the rendered SDF gradients to be consistent with the input normal vectors so that accurate geometry details are recovered. The mask term aligns the shape's silhouette with input masks and can be considered the boundary condition. The eikonal term enforces the SDF gradient norm to be close to 1 almost everywhere so that the neural SDF is approximately valid [22].", + "bbox": [ + 75, + 523, + 468, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Directional Finite Difference", + "text_level": 1, + "bbox": [ + 76, + 689, + 333, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As seen, both the normal and eikonal terms backpropagate the loss gradient via the SDF gradient to neural SDF parameters (i.e., double backpropagation $\\frac{\\partial\\mathcal{L}}{\\partial\\nabla f}\\frac{\\partial\\nabla f}{\\partial\\theta}$ ). The most common way is to use Pytorch's automatic differentiation (AD) due to implementation simplicity. However, the multi-resolution hash encoding package does not officially support double propagation. Common workarounds include CUDA programming [24, 30] or use finite difference (FD) [18, 36] to bypass double propagation. However, CUDA programming can be inflexible in that it shrinks the design space (e.g., NeuS2 [30] requires ReLU activation), and FD will slow down the training. Consider the ordinary", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "axis-aligned FD:", + "bbox": [ + 500, + 359, + 612, + 373 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla f (\\mathbf {x}) \\approx \\left[ \\begin{array}{l} \\frac {f (\\mathbf {x} + \\epsilon_ {x}) - f (\\mathbf {x} - \\epsilon_ {x})}{2 \\epsilon} \\\\ \\frac {f (\\mathbf {x} + \\epsilon_ {y}) - f (\\mathbf {x} - \\epsilon_ {y})}{2 \\epsilon} \\\\ \\frac {f (\\mathbf {x} + \\epsilon_ {z}) - f (\\mathbf {x} - \\epsilon_ {z})}{2 \\epsilon} \\end{array} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 380, + 890, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon_{* \\in \\{x, y, z\\}}$ is an offset vector along $x, y$ , or $z$ axis with a tiny step size $\\epsilon$ . Computing the gradient at one point requires evaluating the SDF at its nearby points. These additional SDF samples are also included in the computation graph in the forward pass. Consequently, using FD significantly slows the forward rendering and backward optimization even if the efficient spatial encoding is used [18].", + "bbox": [ + 496, + 443, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To avoid redundant SDF evaluations and double backpropagation, we propose directional finite difference (DFD) to reuse the volume rendering SDF samples for SDF gradients approximation. Consider the SDF value at a 3D point $\\mathbf{x}$ parameterized by the ray origin $\\mathbf{o}$ and direction $\\mathbf{v}$ :", + "bbox": [ + 496, + 549, + 890, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf (\\mathbf {x}) = f (\\mathbf {o} + t \\mathbf {v}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 632, + 890, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t$ is the distance to the ray origin. From Eq. (7), we have", + "bbox": [ + 496, + 656, + 890, + 683 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {d f}{d t} = \\nabla f (\\mathbf {x}) ^ {\\top} \\mathbf {v} := \\nabla_ {\\mathbf {v}} f. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 681, + 890, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Equation (8) is exactly the directional derivative of the SDF, which projects the SDF gradient onto the ray direction. Suppose we parameterize the same point along three different directions as", + "bbox": [ + 496, + 714, + 890, + 772 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} = \\mathbf {o} _ {i} + t _ {i} \\mathbf {v} _ {i}, \\quad i \\in \\{1, 2, 3 \\}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 782, + 890, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Applying Eq. (8) then yields three linear equations:", + "bbox": [ + 500, + 806, + 839, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underbrace {\\left[ \\begin{array}{l} - \\mathbf {v} _ {1} ^ {\\top} - \\\\ - \\mathbf {v} _ {2} ^ {\\top} - \\\\ - \\mathbf {v} _ {3} ^ {\\top} - \\end{array} \\right]} _ {\\mathbf {V}} \\nabla f = \\underbrace {\\left[ \\begin{array}{l} \\frac {d f}{d t _ {1}} \\\\ \\frac {d f}{d t _ {2}} \\\\ \\frac {d f}{d t _ {3}} \\end{array} \\right]} _ {\\nabla_ {\\mathbf {V}} f}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 828, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20584", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From Eq. (10), we can compute the SDF gradient by", + "bbox": [ + 76, + 90, + 426, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla f = \\mathbf {V} ^ {- 1} \\nabla_ {\\mathbf {V}} f. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 114, + 468, + 132 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Numerically, we can approximate the directional derivative by", + "bbox": [ + 76, + 143, + 468, + 174 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\frac {d f}{d t _ {i}} \\approx \\frac {f (\\mathbf {o} _ {i} + (t _ {i} + \\Delta t) \\mathbf {v} _ {i}) - f (\\mathbf {o} _ {i} + (t _ {i} - \\Delta t) \\mathbf {v} _ {i})}{2 \\Delta t} \\\\ = \\frac {f \\left(\\mathbf {x} _ {i} + \\Delta t \\mathbf {v} _ {i}\\right) - f \\left(\\mathbf {x} _ {i} - \\Delta t \\mathbf {v} _ {i}\\right)}{2 \\Delta t}, \\quad i \\in \\{1, 2, 3 \\} \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 181, + 468, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Equations (11) and (12) imply that, to numerically approximate the SDF gradient, we can first compute finite differences along three directions $\\mathbf{V}$ , then linearly transform the concatenated finite difference by $\\mathbf{V}^{-1}$ . We call this SDF gradient approximation directional finite difference. Equation (11) has a unique solution if the three directions are non-coplanar.", + "bbox": [ + 75, + 270, + 468, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The ordinary finite difference can be viewed as the special case of Eq. (11) by choosing the three coordinate axes as the ray directions. In this case, $\\mathbf{V}$ becomes the identity matrix. Both DFD and FD require nearby SDF samples to approximate the SDF gradient. However, by tailoring the sampling strategy, we can reuse the SDF samples in volume rendering for DFD without additional SDF samples, as described in the following.", + "bbox": [ + 75, + 377, + 468, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Patch-based Ray Marching", + "text_level": 1, + "bbox": [ + 76, + 506, + 323, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DFD requires computing the finite difference in three directions. In volume rendering, we naturally have SDF samples along the viewing ray direction that can be used for finite difference. For the other two directions, we propose to sample patches of pixels instead of single pixels and march a plane on each patch of rays to locate the sampled points on parallel planes.", + "bbox": [ + 75, + 529, + 468, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we perform ordinary ray marching for the center pixel/ray of the patch and treat the sampled points as the intersections between a marching plane and the center ray. The points to be sampled on the remaining rays within the patch can then be found as the intersection of the planes and the rays:", + "bbox": [ + 75, + 635, + 468, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nt _ {j} = \\frac {t _ {i} \\mathbf {v} _ {i} ^ {\\top} \\mathbf {m}}{\\mathbf {v} _ {j} ^ {\\top} \\mathbf {m}}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 724, + 468, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{m}$ is the unit normal direction of the marching plane, $t_i$ is the ray marching distance on the center ray with direction $\\mathbf{v}_i$ , and $t_j$ is the distance from the intersection point to the ray origin (i.e., camera center) along the direction $\\mathbf{v}_j$ ; for derivation see Sec. B.2 in SupMat.", + "bbox": [ + 75, + 763, + 468, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We choose the marching plane to parallel the image plane where the patch is located (Fig. 3). With this sampling strategy, the three directions for DFD computation are the viewing ray direction and the world-space directions of $x$", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $y$ -axis of the camera coordinates. The marching plane normal $\\mathbf{m}$ becomes the world-space direction of the $z$ -axis of the camera coordinates and is the same for all pixels/rays from the same view. This way, the SDF samples on the same ray can be used for volume rendering as Eq. (4), and the SDF samples neighbor to a point can be used for SDF gradient computation as Eq. (11). We use the central difference (i.e., Eq. (12)) for points with neighbors in both directions and forward/backward difference for boundary points.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This sampling strategy is beneficial in three aspects. First, it allows us to reuse the SDF samples from neighboring rays to compute DFD. Second, it stabilizes training by ensuring the existence of $\\mathbf{V}^{-1}$ , since the viewing ray direction will never co-planer to the image plane. Third, it reduces the computation overhead during training because the three ray directions are identical for samples from the same pixel/ray. The inverse of ray directions $\\mathbf{V}^{-1}$ can be pre-computed for each pixel before training. In our experiments, the computation takes about 0.6 sec for two million pixels. During training, only the $3\\times 3$ matrix-vector multiplication Eq. (11) is required, which can almost be ignored compared to AD or FD.", + "bbox": [ + 496, + 227, + 890, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 439, + 633, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method quantitatively on the MVPS benchmark in Sec. 4.1, investigate different components of our method in Sec. 4.2 and evaluate our method qualitatively on our captured real-world data in Sec. 4.3.", + "bbox": [ + 496, + 465, + 890, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details We apply SDM-UniPS [10] to images in each view to obtain normal maps. In each batch, we randomly sample 2048 patches of $3 \\times 3$ pixels from all input normal maps. Like INGP [20], we skip the empty and occluded space during ray marching using NerfAcc [17]. The scene is bounded within a unit sphere, and we use a coarse-to-fine ray marching step decreasing log-linearly from $1e^{-2}$ to $1e^{-3}$ . We optimize the shape of DiLiGenT-MV [16] benchmark objects for 5000 batches (roughly equal to 20 epochs), and 30000 batches for our captured data. The weights of loss terms are set as 1. We use the Adam optimizer with an initial learning rate $5e^{-3}$ . More details can be found in Sec. B in SupMat.", + "bbox": [ + 496, + 532, + 890, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Quantitative evaluation", + "text_level": 1, + "bbox": [ + 498, + 739, + 718, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset We evaluate our method using the only MVPS benchmark dataset DiLiGenT-MV [16]. It consists of five objects captured from 20 views, and 96 OLAT images are captured in each view, yielding a total of 1920 images per object. Each image is $512 \\times 612$ , and the total number of foreground object pixels is about 2.2 million. The camera is fixed during capture, and the object is placed on a turntable in a darkroom. The \"ground truth\" meshes are created using a 3D scanner.", + "bbox": [ + 496, + 763, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20585", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines We benchmark SuperNormal against several MVPS methods grouped into three categories. (1) Those that refine an initial point cloud or mesh and require manual effort, including R-MVPS [23] and B-MVPS [16]. (2) Those that use Poisson surface reconstruction (PSR) [13] to fuse multi-view depth and normal maps, including MVPSNet [35]. (3) Those that optimize an implicit neural surface representation, including UA-MVPS [12], PS-NeRF [31], MVAS [4], and our method.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics We use L2 Chamfer distance (CD) and F-score with threshold $\\tau_{F} = 0.5\\mathrm{mm}$ to evaluate geometry accuracy [12, 15]. For CD and F-score, we only consider visible points from the input views by casting rays for all pixels and finding the first ray-mesh intersections (More details in Sec. B.4 in SupMat.)", + "bbox": [ + 75, + 234, + 468, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results As reported in Table 1, our method achieves the best geometry accuracy among all compared MVPS methods (See Sec. A in SupMat. for normal accuracy). We attribute this superior performance to two aspects: 1) the powerful expressive capability of multi-resolution hash encoding [20] and 2) the fine-grained normal estimation method SDM-UniPS [10], as discussed in Sec. 4.2. Figure 4 visualizes the recovered shapes of different methods. Our method recovers fine-grained surface details on par with the \"GT\" meshes produced by a 3D scanner.", + "bbox": [ + 75, + 332, + 468, + 482 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 also reports the average runtime measured on our machine with one RTX 4090Ti graphics card. The runtime of R-MVPS [23] and B-MVPS [16] is difficult to measure since both methods require intensive manual effort. The code of UA-MVPS [12] is not publicly available, and several hours per object are required, as reported in MVPSNet [35]. Both PS-NeRF [31] and MVAS [4] require hours per object since they optimize a dense MLP. Our method is the fastest among neural implicit representation-based methods, with less than one minute per object.", + "bbox": [ + 75, + 484, + 468, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On the other hand, MVPSNet [35] is highly efficient because its pipeline mainly consists of feedforward inferencing a trained neural network and an sPSR [13] step, the latter of which has been optimized over one decade. Although our method is slower than MVPSNet [35], we achieve better reconstruction quality within a reasonable time.", + "bbox": [ + 75, + 636, + 468, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 739, + 225, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of normal maps Since multi-view normal integration is not bound to specific normal estimation methods, we investigate the effect of different normal estimation methods on normal integration methods. We test three sets of normal maps, estimated by SDPS [6] or SDM-UniPS [10], and the one rendered from the scanned mesh (i.e., \"GT\" normal maps provided by DiLiGenT-MV [16]). Table 2 reports the L2-Chamfer distance and F-score averaged on five DiLiGenT-MV [16] objects. Our method consistently", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "achieves better results than PS-NeRF [31] and MVAS [4] when using the same set of normal maps. This verifies that our method can better fuse the information from 2.5D normal maps to 3D shapes. Using the SoTA method SDM-UniPS [10] for normal estimation, our method realizes the best reconstruction quality.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of spatial encoding Table 2 also compares the geometry accuracy between multi-resolution hash encoding (hash enc.) and positional encoding (pos. enc.). Our results verify again that multi-resolution hash encoding surpasses positional encoding in expressiveness, enhancing geometry accuracy. Therefore, the keys to superior MVPS performance are high-quality normal estimation, an expressive surface representation, and an effective normal fusion approach.", + "bbox": [ + 496, + 186, + 890, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of SDF gradient computation Table 3 compares the geometry accuracy and runtime using finite difference (FD), Pytorch's automatic differentiation (AD), and directional finite difference (DFD) to compute the SDF gradient. All other parameter settings are kept the same. We do not observe a significant difference regarding reconstruction quality. However, DFD performs consistently faster than FD and AD by accelerating both the forward rendering and backward optimization.", + "bbox": [ + 496, + 314, + 890, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generalizability of DFD In principle, DFD and patch-based sampling strategy apply to any SDF-based volume rendering like NeuS [29] or Neuralangelo [18]. To verify this, we incorporate DFD and patch-based sampling into Neuralangelo $[18]^{1}$ . We follow the coarse-to-fine sampling strategy of Neuralangelo [18] but use DFD instead of FD for SDF gradient approximation. We do not use mask supervision for Neuralangelo [18]. As shown in Figure 5, training time is reduced by $25\\%$ on a scene from the DTU benchmark [11], even $2.25\\times$ more pixels are looped in training. The result suggests that DFD could be a general strategy to accelerate NeuS-like multi-view reconstruction [29].", + "bbox": [ + 496, + 455, + 890, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative results on our captured data", + "text_level": 1, + "bbox": [ + 498, + 647, + 843, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further demonstrate the performance of SuperNormal in fine-grained surface detail recovery, we also captured objects with more complex details, as shown in Fig. 6. Thanks to the generalizability of SDM-UniPS [10], we can capture the objects under casual light conditions (e.g., within an apartment) without ensuring a darkroom setup. We manually move a video light during the capture without any photometric or geometric light calibration. In total, we capture 18 viewpoints $\\times 13$ light conditions for each object (1 under ambient light for SfM). We use Metashape [1] to calibrate camera parameters, use SAM [14] to create foreground masks, and apply SDM-UniPS [10] to every 12 image to infer the per-view normal map.", + "bbox": [ + 496, + 670, + 890, + 867 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1Our modified code of Neuralangelo [18] is available at https:// github.com/xucao-42/Neuralangelo_DFD.git.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20586", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d3bde6629ff85c8b95b728fc3091d52c14e12a022a38410f5f160cc5bb767f5a.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation on DiLiGenT-MV [16] benchmark. Red and orange cells indicate the best and the second best results respectively. On average, our approach achieves the best reconstruction quality at the second fastest speed." + ], + "table_footnote": [], + "table_body": "
MethodsL2 Chamfer distance [mm] (↓)F-score (τF=0.5 mm) (↑)Runtime (↓) Average
BearBuddhaCowPot2ReadingAverageBearBuddhaCowPot2ReadingAverage
R-MVPS [23]1.0700.3970.4401.5040.5610.7940.2620.6980.7600.1980.5190.487NA
B-MVPS [16]0.2120.2540.0910.2010.2590.2030.9580.9020.9860.9460.9140.941NA
MVPSSNet [35]0.3170.2790.2550.3100.2480.2820.8130.8660.8890.8380.9180.86537 secs
UA-MVPS [12]0.4140.4520.3260.4140.3820.3980.7070.6690.7980.7310.7620.733several hrs
PS-NeRF [31]0.2600.3140.2870.2540.3520.2930.8980.8060.8560.9190.7850.8538 hrs
MVAS [4]0.2430.3570.2160.1970.5220.3070.9090.7540.9070.9620.5460.81670 mins
Ours0.1840.2180.1930.1500.2230.1940.9830.9500.9770.9940.9160.96248 secs
", + "bbox": [ + 78, + 122, + 890, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0cdb1c61c4408a135140062303efda8dce5b4344baf5c534ec20c2b9ffc6b89f.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison of recovered shapes of Buddha and Pot2 and their error maps. Best viewed on screen." + ], + "image_footnote": [], + "bbox": [ + 78, + 243, + 890, + 696 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/df881f485d86c96488007ec3b5d1c403c1ec442f31c5f6e4059d0c2c5f3c456e.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation using normal maps estimated by different methods as inputs. The metrics are averaged over five DiLiGenT-MV [16] objects. Darker colors indicate better results." + ], + "table_footnote": [], + "table_body": "
L2 Chamfer distance [mm] (↓)F-score (τF=0.5 mm) (↑)
SDPS [6]SDM-UniPS [10]GTSDPS [6]SDM-UniPS [10]GT
MVAS [4]0.3070.3320.2920.8160.8120.859
PS-NeRF [31]0.2930.2960.2240.8530.8440.938
Ours (pos, enc.)0.3010.2280.1950.8490.9420.977
Ours (hash enc.)0.2810.1940.1140.8600.9620.998
", + "bbox": [ + 78, + 797, + 468, + 861 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5927e38fee03c5e9dc7d9779983a82a63e5ed685e2215e57463cb6c85e23ce4.jpg", + "image_caption": [ + "Figure 5. DFD and patch-based sampling accelerate Neuralangelo's training without quality degradation. We train both alternatives for $20\\mathrm{k}$ batches. For FD, 512 random rays are sampled per batch; for DFD, 128 random patches of $3\\times 3$ rays per batch." + ], + "image_footnote": [], + "bbox": [ + 500, + 743, + 890, + 801 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20587", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7e2703c87f857879dc1849a776eace2d859c06bd31315baa67a6f01154f52bdd.jpg", + "table_caption": [ + "Table 3. Ablation study using different strategies for SDF gradients computation in volume rendering. We report the metrics averaged over five DiLiGenT-MV [16] objects. FD: Finite difference, AD: Automatic differentiation, DFD: Directional finite difference." + ], + "table_footnote": [], + "table_body": "
AccuracyRuntime (Sec)
CDF-ScoreForwardBackwardTotal
FD0.1920.96347.770.3124.6
AD0.1950.96129.442.478.1
DFD0.1940.96119.622.548.3
", + "bbox": [ + 117, + 156, + 426, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/38318f8bec8399f933f859ec2e8bdd664f5731d6dcae7705950fc61f9002425c.jpg", + "image_caption": [ + "Figure 6. Our real-world data capture setup. The target object is placed on a turntable before a dark sheet under general indoor illumination. There is no need to prepare a darkroom. We use an iPhone to take images and move a video light to illuminate the object from the camera side. Objects are $6\\mathrm{cm}$ to $15\\mathrm{cm}$ high." + ], + "image_footnote": [], + "bbox": [ + 84, + 242, + 272, + 352 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3cea14033fa30364ecc607094b1aeab094a1e9a4dcd60b16e48dfb07d9c7987d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 241, + 460, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figures 1 and 7 display the results from NeuS2 [30], SuperNormal, and a structured-light based commercial 3D scanner EinScan SE. Although NeuS2 [30] and our method use the same multi-resolution hash encoding for fine-grained scene representation, the surface details recovered from multi-view color images are smoothed out. Our method, benefitting from the normal map inputs, recovers more fine-grained surface details.", + "bbox": [ + 75, + 448, + 468, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We find that our scanner performs poorly in highly concave regions (Fig. 7 above). This is because the scanner uses a wide baseline stereo camera (about $15\\mathrm{cm}$ apart), and small concave regions create self-occlusions, thus reducing the reconstruction quality. In principle, normal estimation does not suffer from self-occlusion because it is based on single-view observations. Fusing multi-view normal maps can then produce better results in concave regions.", + "bbox": [ + 75, + 570, + 468, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 705, + 191, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have demonstrated SuperNormal, a fast, high-fidelity 3D reconstruction approach using multi-view normal maps estimated by photometric stereo. SuperNormal outperforms other neural MVPS methods in speed and reconstruction quality on the benchmark dataset. Compared to multi-view neural reconstruction methods, using normal maps can exploit the expressive power of multi-resolution hash encoding, and produce scanner-level surface details. Further, we have shown the effectiveness of directional finite differences in expediting training without sacrificing quality.", + "bbox": [ + 75, + 732, + 468, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SuperNormal provides a promising direction to make", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c13d95a870b71c1cdb722c7c9ecd34ba454525d9334c13f2f53c0d2f5e68843a.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison using our objects. Our approach recovers better surface detail than multi-view stereo approaches and achieves competitive results with commercial 3D scanners." + ], + "image_footnote": [], + "bbox": [ + 501, + 87, + 890, + 452 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0cfc59baf60775ffccca5f973d02bf866d9e9f00c5a44ab97ee07affe14405ce.jpg", + "image_caption": [ + "Ours", + "Figure 8. Limitations. (Left) Fault artifacts are highlighted in red circles. (Right) Our method is sensitive to the inaccuracy in normal maps estimated by monocular methods (e.g., OmniData [7])." + ], + "image_footnote": [], + "bbox": [ + 504, + 512, + 702, + 628 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e13b646eb49d1ada047b5971a0c1d4b99296474726c8ff9fe7100f76ef325ffd.jpg", + "image_caption": [ + "OmniData normal" + ], + "image_footnote": [], + "bbox": [ + 714, + 513, + 813, + 627 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f03c38d05d82dad187b9984e45f66d220d7e02ef2cf676b8605da37d2b39283c.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 813, + 513, + 885, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "high-quality 3D scanning more accessible. Unlike costly commercial 3D scanners, our approach is viable with everyday devices like smartphones and basic equipment such as a tripod, video light, and a turntable.", + "bbox": [ + 496, + 705, + 890, + 766 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations Our method may produce fault-like artifacts, as shown in Fig. 8. The artifacts exist no matter our SDF gradient computation approach (i.e., AD, FD, or DFD) and require further investigation. Moreover, the performance depends on mask supervision and high-quality normal maps (e.g., those obtained from photometric stereo). Our method does not work well with the monocular normal estimation methods like OmniData [7] (Fig. 8).", + "bbox": [ + 496, + 777, + 892, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20588", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Agisoft metashape. https://www.agisoft.com.6", + "[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 2", + "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-NeRF: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 2", + "[4] Xu Cao, Hiroaki Santo, Fumio Okura, and Yasuyuki Matsushita. Multi-view azimuth stereo via tangent space consistency. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 825-834, 2023. 1, 2, 6, 7, 11, 12", + "[5] Ju Yong Chang, Kyoung Mu Lee, and Sang Uk Lee. Multiview normal field integration using level set methods. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 1-8. IEEE, 2007. 1, 2", + "[6] Guanying Chen, Kai Han, Boxin Shi, Yasuyuki Matsushita, and Kwan-Yee K. Wong. SDPS-Net: Self-calibrating deep photometric stereo networks. In CVPR, 2019. 6, 7", + "[7] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In ICCV, pages 10786-10796, 2021. 8", + "[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of Machine Learning and Systems 2020, pages 3569-3579. 2020. 3, 11", + "[9] Carlos Hernandez, George Vogiatzis, and Roberto Cipolla. Multiview photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 30(3):548-554, 2008. 1, 2", + "[10] Satoshi Ikehata. Scalable, detailed and mask-free universal photometric stereo. In Proc. of Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5, 6, 7", + "[11] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, pages 406-413. IEEE, 2014. 6", + "[12] Berk Kaya, Suryansh Kumar, Carlos Oliveira, Vittorio Ferrari, and Luc Van Gool. Uncertainty-aware deep multi-view photometric stereo. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 12601-12611, 2022. 2, 6, 7", + "[13] Michael Kazhdan and Hugues Hoppe. Screened Poisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 2, 6", + "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proc. of International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 6", + "[15] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 6", + "[16] Min Li, Zhenglong Zhou, Zhe Wu, Boxin Shi, Changyu Diao, and Ping Tan. Multi-view photometric stereo: A ro" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "bust solution and benchmark dataset for spatially varying isotropic materials. IEEE Transactions on Image Processing, 29:4159-4173, 2020. 2, 5, 6, 7, 8, 11, 12", + "[17] Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. NerfAcc: Efficient sampling accelerates nerfs. In Proc. of International Conference on Computer Vision (ICCV), pages 18537-18546, 2023. 5, 13", + "[18] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 8456-8465, 2023. 2, 3, 4, 6", + "[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. of European Conference on Computer Vision (ECCV), 2020. 2", + "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 3, 5, 6", + "[21] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proc. of International Conference on Computer Vision (ICCV), 2021. 2", + "[22] Stanley Osher, Ronald Fedkiw, and K Piechor. Level set methods and dynamic implicit surfaces. Appl. Mech. Rev., 57(3):B15-B15, 2004. 2, 4", + "[23] Jaesik Park, Sudipta N Sinha, Yasuyuki Matsushita, Yu-Wing Tai, and In So Kweon. Robust multiview photometric stereo using planar mesh parameterization. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 39(8):1591-1604, 2016. 2, 6, 7, 11, 12", + "[24] Radu Alexandru Rosu and Sven Behnke. Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices. In CVPR, 2023. 3, 4", + "[25] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. of European Conference on Computer Vision (ECCV), 2016. 1", + "[26] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my NeRF: A dataset for novel view synthesis and relighting of real world objects. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 20762-20772, 2023. 2", + "[27] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 2", + "[28] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. In ECCV, 2022. 3", + "[29] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view recon" + ], + "bbox": [ + 501, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20589", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "struction. Advances in Neural Information Processing Systems (NeurIPS), 2021. 1, 3, 6", + "[30] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. NeuS2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proc. of International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 4, 8, 14, 15", + "[31] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K. Wong. PS-NeRF: Neural inverse rendering for multi-view photometric stereo. In Proc. of European Conference on Computer Vision (ECCV), 2022. 1, 2, 6, 7, 11, 12, 14", + "[32] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 2", + "[33] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems (NeurIPS), 34:4805-4815, 2021. 3", + "[34] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. NIPS, 2022. 3", + "[35] Dongxu Zhao, Daniel Lichy, Pierre-Nicolas Perrin, Jan-Michael Frahm, and Soumyadip Sengupta. MVPSNet: Fast generalizable multi-view photometric stereo. In Proc. of International Conference on Computer Vision (ICCV), pages 12525-12536, 2023. 2, 6, 7", + "[36] Fuqiang Zhao, Yuheng Jiang, Kaixin Yao, Jiakai Zhang, Liao Wang, Haizhao Dai, Yuhui Zhong, Yingliang Zhang, Minye Wu, Lan Xu, and Jingyi Yu. Human performance modeling and rendering via neural animated mesh. ACM Trans. Graph., 41(6), 2, 3, 4" + ], + "bbox": [ + 78, + 90, + 468, + 584 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20590", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/a4b33f00-d1f9-45e2-a62c-bb25956b58ec_model.json b/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/a4b33f00-d1f9-45e2-a62c-bb25956b58ec_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e16349b08d530b6262c382db596faca368d132cb --- /dev/null +++ b/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/a4b33f00-d1f9-45e2-a62c-bb25956b58ec_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.252, + 0.131, + 0.72, + 0.175 + ], + "angle": 0, + "content": "SuperNormal: Neural Surface Reconstruction via Multi-View Normal Integration" + }, + { + "type": "text", + "bbox": [ + 0.363, + 0.204, + 0.605, + 0.22 + ], + "angle": 0, + "content": "Xu Cao Takafumi Taketomi" + }, + { + "type": "text", + "bbox": [ + 0.435, + 0.223, + 0.536, + 0.24 + ], + "angle": 0, + "content": "CyberAgent" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.242, + 0.675, + 0.257 + ], + "angle": 0, + "content": "{xu_cao, taketomi_takafumi}@cyberagent.co.jp" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.474, + 0.598 + ], + "angle": 0, + "content": "We present SuperNormal, a fast, high-fidelity approach to multi-view 3D reconstruction using surface normal maps. With a few minutes, SuperNormal produces detailed surfaces on par with 3D scanners. We harness volume rendering to optimize a neural signed distance function (SDF) powered by multi-resolution hash encoding. To accelerate training, we propose directional finite difference and patch-based ray marching to approximate the SDF gradients numerically. While not compromising reconstruction quality, this strategy is nearly twice as efficient as analytical gradients and about three times faster than axis-aligned finite difference. Experiments on the benchmark dataset demonstrate the superiority of SuperNormal in efficiency and accuracy compared to existing multi-view photometric stereo methods. On our captured objects, SuperNormal produces more fine-grained geometry than recent neural 3D reconstruction methods. Our code is available at https://github.com/CyberAgentAILab/SuperNormal." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.631, + 0.21, + 0.646 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Recovering high-quality 3D geometry of real-world objects from their multi-view images is a long-standing challenge in computer vision. Recently, neural implicit surface-based methods have shown remarkable reconstruction results. Compared to traditional multi-view stereo (MVS) methods [25], neural methods are more robust to view-dependent observations and textureless surfaces [29]. Furthermore, the reconstruction procedure has become highly efficient [30] by introducing multi-resolution hash coding [20]. However, even though this spatial encoding allows fine-grained geometry to be represented, shape reconstruction from multi-view images tends to smooth out high-frequency surface details, as shown in Fig. 1 bottom right." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Multi-view photometric stereo (MVPS), on the other hand, aims to recover pixel-level high-frequency surface detail by introducing additional lighting conditions during" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.291, + 0.892, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.425, + 0.891, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.55, + 0.895, + 0.648 + ], + "angle": 0, + "content": "Figure 1. (Top) From multi-view surface normal maps, SuperNormal recovers fine-grained surface details comparable to 3D scanners while being orders of magnitude faster than the existing MVPS method PS-NeRF [31]. (Bottom) Using normal maps produces more faithful high-frequency details than the MVS-based method NeuS2 [30], although both use multi-resolution hash encoding [20]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.854 + ], + "angle": 0, + "content": "the image acquisition [9]. In a typical workflow, a surface normal map is first recovered for each view, recording per-pixel surface orientation information. The normal maps estimated at different viewpoints are then fused into a 3D model, also known as multi-view normal integration [5]. However, existing normal fusion methods struggle to reflect the details of the normal maps in the shape of the recovered 3D model. Moreover, the reconstruction process takes hours even when only a few low-resolution normal maps are used [4, 31]. Due to the lack of a fast and accurate multiview normal fusion method, current multi-view photometric stereo results remain unsatisfactory." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "This paper presents SuperNormal to unite the best of both worlds. Normal maps with pixel-level surface details are utilized to exploit the expressive power and efficiency" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20581" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.181 + ], + "angle": 0, + "content": "of multi-resolution hash encoding-powered neural implicit surface. Using normal maps alone avoids per-scene reflectance modeling and optimization, thus improving training efficiency compared to using color images. In a few minutes, our method can produce fine-grained geometry on par with commercial 3D scanners." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.471, + 0.408 + ], + "angle": 0, + "content": "Specifically, we use volume rendering to optimize a neural signed distance function (SDF) such that its rendered normal maps are consistent with input normal maps. To further improve the training efficiency, we propose a directional finite difference and patch-based ray marching strategy to approximate the SDF gradients numerically. This strategy kills two birds with one stone: SDF values evaluated at sampled points are all used both for SDF-based volume rendering and SDF gradient approximation. This way, we avoid second-order derivative computation during training. Unlike traditional axis-aligned finite difference, directional finite difference avoids redundant SDF evaluations (Fig. 2). Without compromising reconstruction quality, our method is nearly twice as fast as using analytical gradients and about three times faster than finite difference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.47, + 0.589 + ], + "angle": 0, + "content": "Compared to existing neural MVPS methods, our method is two orders of magnitude faster and recovers better surface details (Fig. 1 above). We also capture objects with complex geometry and show that our method recovers more faithful surface details than the recent neural 3D reconstruction method [30] (Fig. 1 bottom). Taking advantage of recent advances in photometric stereo, we can estimate high-quality normal maps from images captured under common lighting conditions with inexpensive equipment [10]. Combined with SuperNormal, high-fidelity 3D objects can be reconstructed using inexpensive equipment in widely accessible environments." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.59, + 0.408, + 0.605 + ], + "angle": 0, + "content": "In summary, this paper's key contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.606, + 0.469, + 0.65 + ], + "angle": 0, + "content": "- An effective neural 3D reconstruction approach using multi-view normal maps, enhancing the reconstruction quality to near 3D scanner levels;" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.656, + 0.469, + 0.7 + ], + "angle": 0, + "content": "- An efficient method for computing numerical gradients in SDF-based neural rendering, speeding up the reconstruction process; and" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.706, + 0.469, + 0.751 + ], + "angle": 0, + "content": "- A comprehensive evaluation using benchmark objects with MVPS approaches and an assessment using our captured objects with MVS approaches." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.606, + 0.469, + 0.751 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.763, + 0.215, + 0.779 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.412, + 0.803 + ], + "angle": 0, + "content": "2.1. Multi-view photometric stereo (MVPS)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.856 + ], + "angle": 0, + "content": "MVPS aims at high-fidelity 3D reconstruction using multi-view observations under varying light conditions [9] (usually one-light-at-time (OLAT) images [16, 26])." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Existing methods differ in what geometry-related information is inferred in each view and how information from different views is fused into a complete 3D model. Chang et" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.891, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.257, + 0.892, + 0.354 + ], + "angle": 0, + "content": "Figure 2. (Left) In SDF-based volume rendering, approximating SDF gradients using axis-aligned finite difference [18, 36] requires additional SDF samples at neighboring positions of volume rendering points, increasing computational complexity and training time. (Right) Our strategy of patch-based ray marching and directional finite difference effectively uses all SDF samples for both SDF gradient approximation and volume rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.488 + ], + "angle": 0, + "content": "al. [5] proposes a level set method [22] to fuse multi-view normal maps into a volumetric representation of the shape. R-MVPS [23] uses the multi-view normal maps to refine a coarse mesh estimated by multi-view stereo. B-MVPS [16] initialize the shape as a sparse point cloud estimated by structure from motion (SfM), then propagates the known point spatially using the normal maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.893, + 0.715 + ], + "angle": 0, + "content": "Recent MVPS methods have undergone a paradigm shift since the advent of NeRF [19] by using a neural implicit scene representation, usually parameterized as a multi-layer perception (MLP). PS-NeRF [31] recovers a neural field based on UNISURF [21] by regularizing the gradients of the neural field using the normal maps. UA-MVPS [12] fuses the multi-view depth and normal maps into a neural SDF using neural inverse rendering while considering per-pixel confidence. MVAS [4] shows utilizing the azimuth components of multi-view normal maps is sufficient to recover the neural implicit surface. While most methods only use images in the same view for normal map estimation, MVPSNet [35] utilizes neighboring views to infer per-view depth and normal maps jointly and applies screened Poisson surface reconstruction (sPSR) [13] to obtain the shape." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.892, + 0.806 + ], + "angle": 0, + "content": "However, the reconstructed surfaces of these methods tend to be smoothed and do not reflect the high-frequency details of the input normal maps. Our approach introduces multi-resolution hash encoding and directional finite difference to improve the implicit neural field's expressive power and training efficiency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.733, + 0.832 + ], + "angle": 0, + "content": "2.2. Neural 3D reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The seminal work NeRF [19] and follow-up works [2, 3, 27] have demonstrated remarkable novel view synthesis quality by volume rendering a neural implicit field. For better geometry reconstruction quality, IDR [32] imposes eikonal" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20582" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "regularization [8] such that the neural field approximates a valid SDF. Later, the concurrent works VolSDF [33] and NeuS [29] develop SDF-based volume rendering, which bridges SDF values to volume rendering weights and improves the geometry fidelity. However, these methods require several hours to recover one object due to optimizing a deep coordinate-based MLP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.471, + 0.351 + ], + "angle": 0, + "content": "Multi-resolution hash encoding has been proposed to improve expressive power and accelerate the learning of neural implicit field [20]. Several efforts have been made to incorporate this expressive spatial encoding into neural SDF learning. Since the official CUDA implementation does not support the double propagation deduced by the eikonal regularization, PermutoSDF [24] and NeuS2 [30] re-implement the CUDA programming. InstantNSR [36] and Neuralangelo [18] use finite difference instead of automatic differentiation to avoid double propagation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.352, + 0.47, + 0.413 + ], + "angle": 0, + "content": "We propose directional finite difference, a numerical way to compute SDF gradients that avoids double propagation and is even faster than automatic differentiation when combined with a tailored patch-based sampling strategy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.427, + 0.415, + 0.442 + ], + "angle": 0, + "content": "2.3. Neural 3D reconstruction with normals" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.469, + 0.542 + ], + "angle": 0, + "content": "Normal maps have been used as auxiliary information to facilitate neural scene reconstruction in MonoSDF [34] and NeuRIS [28]. These works target scene-level reconstruction, and normal maps help infer the geometry of textureless smooth regions (e.g., walls) where using color images is likely to break down." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.469, + 0.62 + ], + "angle": 0, + "content": "Instead, our method is particularly useful for single objects with complex surface details. We use photometric stereo [10] to estimate normal maps from images under varying light conditions. This produces high-quality normal maps that can be used alone for shape reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.638, + 0.185, + 0.656 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.665, + 0.469, + 0.802 + ], + "angle": 0, + "content": "We aim to recover the surface given a set of normal maps, object masks, and camera intrinsic and extrinsic parameters. Since photometric stereo methods estimate the normal maps in camera space, we first apply the camera-to-world rotation to normal maps to obtain world-space normal maps. Figure 3 depicts our pipeline. We represent the geometry as a parametric SDF and optimize SDF parameters such that its rendered gradient and opacity values are consistent with input normal vectors and mask values, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.176, + 0.832 + ], + "angle": 0, + "content": "3.1. Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "SDF-based volume rendering (Forward pass) A signed distance function (SDF) implicitly captures the geometry by assigning each spatial point with the signed distance to its closest surface point. The surface \\(\\mathcal{M}\\) can then be repre" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.761, + 0.106 + ], + "angle": 0, + "content": "sented as the zero-level-set of the SDF:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.123, + 0.892, + 0.139 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} = \\{\\mathbf {x} \\mid f (\\mathbf {x}) = 0 \\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.155, + 0.892, + 0.29 + ], + "angle": 0, + "content": "We parameterize the SDF using multi-resolution hash encoding followed by a shallow MLP. In multi-resolution hash encoding, the 3D space is discretized into a set of multi-resolution 3D grids. Each grid cell is associated with a unique hash value mapping the 3D coordinates to a learnable feature vector. Specifically, \\( h(\\mathbf{x};\\phi) \\) is the feature vector obtained by concatenating feature vectors of different levels \\( h(\\mathbf{x}) = [\\mathbf{x},h_1(\\mathbf{x}),\\dots,h_L(\\mathbf{x})] \\). Then, the SDF can be parameterized as" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.307, + 0.892, + 0.323 + ], + "angle": 0, + "content": "\\[\nf (\\mathbf {x}) = f (h (\\mathbf {x}; \\phi); \\theta), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.891, + 0.399 + ], + "angle": 0, + "content": "where \\(\\phi\\) and \\(\\theta\\) are learnable hash-encoding feature vectors and MLP parameters. Multi-resolution hash encoding has proven to speed up training and improve the expressive power of the neural SDF [20, 30, 36]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.507 + ], + "angle": 0, + "content": "During training, we evaluate the SDF values at points sampled on the rays cast from the camera centers toward the scene. NeuS [29] proposes an unbiased and occlusion-aware way to transfer the SDF samples to volume rendering opacities. Given \\(N\\) ordered 3D points \\(\\{\\mathbf{x}_i\\}_{i=0}^N\\) on a ray and their SDF values \\(\\{f(\\mathbf{x}_i)\\}_{i=0}^N\\), the opacity at each point is designed as" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.521, + 0.891, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = \\max \\left(\\frac {\\phi_ {s} (f (\\mathbf {x} _ {i})) - \\phi_ {s} (f (\\mathbf {x} _ {i + 1}))}{\\phi_ {s} (f (\\mathbf {x} _ {i}))}, 0\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.891, + 0.602 + ], + "angle": 0, + "content": "where \\(\\phi_s(x) = \\frac{1}{1 + \\exp(-sx)}\\) is the sigmoid function with a trainable sharpness \\(s\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.891, + 0.636 + ], + "angle": 0, + "content": "With this SDF-based opacity, we can render the normal and opacity for the pixel \\(\\mathbf{p}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.65, + 0.892, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {n}} (\\mathbf {p}) = \\sum_ {i = 0} ^ {N} T _ {i} \\alpha_ {i} \\nabla f (\\mathbf {x} _ {i}), \\quad \\hat {o} (\\mathbf {p}) = \\sum_ {i = 0} ^ {N} T _ {i} \\alpha_ {i} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.676, + 0.695, + 0.846, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\text {w i t h} \\quad T _ {i} = \\prod_ {j = 0} ^ {i - 1} (1 - \\alpha_ {j}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.751, + 0.892, + 0.827 + ], + "angle": 0, + "content": "From Eqs. (3) and (4), the neural SDF must be evaluated at least once at sampled points for volume rendering. In Secs. 3.2 and 3.3, we show that these SDF samples can be reused for SDF gradient approximation, thus reducing the computational amount and speeding up the training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Neural SDF Training (Backward pass) To train the neural SDF, we supervise the rendered normals and opacities using input ones \\(\\mathbf{n}(\\mathbf{p})\\) and \\(\\hat{o} (\\mathbf{p})\\). The loss function consists" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20583" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.892, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.278, + 0.894, + 0.334 + ], + "angle": 0, + "content": "Figure 3. Approach overview. (a) Given input views, we randomly sample patches of pixels, cast rays from those pixels, and march a plane along the patch of rays (Sec. 3.3). We treat the ray-plane intersections as sampled points and evaluate their SDF values via the neural SDF. (b) SDF samples neighbor to a point are used to compute the SDF gradients using directional finite difference (Sec. 3.2). (c) SDF samples on the same rays are used for SDF-based volume rendering (Sec. 3.1)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.36, + 0.47, + 0.374 + ], + "angle": 0, + "content": "of three terms: The normal, the mask, and the eikonal term:" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.398, + 0.469, + 0.501 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\overbrace {\\sum_ {\\mathbf {p}} \\left(\\hat {\\mathbf {n}} (\\mathbf {p}) - \\mathbf {n} (\\mathbf {p})\\right) ^ {2}} ^ {\\text {n o r m a l}} + \\lambda_ {1} \\overbrace {\\sum_ {\\mathbf {p}} \\operatorname {B C E} \\left(\\hat {o} (\\mathbf {p}) , o (\\mathbf {p})\\right)} ^ {\\text {m a s k}} \\tag {5} \\\\ + \\lambda_ {2} \\underbrace {\\sum_ {\\mathbf {x}} \\left(\\| \\nabla f (\\mathbf {x}) \\| _ {2} - 1\\right) ^ {2}} _ {e i k o n a l}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.525, + 0.47, + 0.66 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}\\) is the sampled pixels in a batch, \\(\\mathbf{x}\\) is the sampled points on rays, and BCE is the binary cross entropy function. The normal term encourages the rendered SDF gradients to be consistent with the input normal vectors so that accurate geometry details are recovered. The mask term aligns the shape's silhouette with input masks and can be considered the boundary condition. The eikonal term enforces the SDF gradient norm to be close to 1 almost everywhere so that the neural SDF is approximately valid [22]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.69, + 0.334, + 0.705 + ], + "angle": 0, + "content": "3.2. Directional Finite Difference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "As seen, both the normal and eikonal terms backpropagate the loss gradient via the SDF gradient to neural SDF parameters (i.e., double backpropagation \\(\\frac{\\partial\\mathcal{L}}{\\partial\\nabla f}\\frac{\\partial\\nabla f}{\\partial\\theta}\\)). The most common way is to use Pytorch's automatic differentiation (AD) due to implementation simplicity. However, the multi-resolution hash encoding package does not officially support double propagation. Common workarounds include CUDA programming [24, 30] or use finite difference (FD) [18, 36] to bypass double propagation. However, CUDA programming can be inflexible in that it shrinks the design space (e.g., NeuS2 [30] requires ReLU activation), and FD will slow down the training. Consider the ordinary" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.36, + 0.614, + 0.374 + ], + "angle": 0, + "content": "axis-aligned FD:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.381, + 0.892, + 0.438 + ], + "angle": 0, + "content": "\\[\n\\nabla f (\\mathbf {x}) \\approx \\left[ \\begin{array}{l} \\frac {f (\\mathbf {x} + \\epsilon_ {x}) - f (\\mathbf {x} - \\epsilon_ {x})}{2 \\epsilon} \\\\ \\frac {f (\\mathbf {x} + \\epsilon_ {y}) - f (\\mathbf {x} - \\epsilon_ {y})}{2 \\epsilon} \\\\ \\frac {f (\\mathbf {x} + \\epsilon_ {z}) - f (\\mathbf {x} - \\epsilon_ {z})}{2 \\epsilon} \\end{array} \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.548 + ], + "angle": 0, + "content": "where \\(\\epsilon_{* \\in \\{x, y, z\\}}\\) is an offset vector along \\(x, y\\), or \\(z\\) axis with a tiny step size \\(\\epsilon\\). Computing the gradient at one point requires evaluating the SDF at its nearby points. These additional SDF samples are also included in the computation graph in the forward pass. Consequently, using FD significantly slows the forward rendering and backward optimization even if the efficient spatial encoding is used [18]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.625 + ], + "angle": 0, + "content": "To avoid redundant SDF evaluations and double backpropagation, we propose directional finite difference (DFD) to reuse the volume rendering SDF samples for SDF gradients approximation. Consider the SDF value at a 3D point \\(\\mathbf{x}\\) parameterized by the ray origin \\(\\mathbf{o}\\) and direction \\(\\mathbf{v}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.633, + 0.892, + 0.65 + ], + "angle": 0, + "content": "\\[\nf (\\mathbf {x}) = f (\\mathbf {o} + t \\mathbf {v}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.891, + 0.684 + ], + "angle": 0, + "content": "where \\(t\\) is the distance to the ray origin. From Eq. (7), we have" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.683, + 0.891, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\frac {d f}{d t} = \\nabla f (\\mathbf {x}) ^ {\\top} \\mathbf {v} := \\nabla_ {\\mathbf {v}} f. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.891, + 0.773 + ], + "angle": 0, + "content": "Equation (8) is exactly the directional derivative of the SDF, which projects the SDF gradient onto the ray direction. Suppose we parameterize the same point along three different directions as" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.784, + 0.891, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} = \\mathbf {o} _ {i} + t _ {i} \\mathbf {v} _ {i}, \\quad i \\in \\{1, 2, 3 \\}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.808, + 0.84, + 0.823 + ], + "angle": 0, + "content": "Applying Eq. (8) then yields three linear equations:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.829, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\underbrace {\\left[ \\begin{array}{l} - \\mathbf {v} _ {1} ^ {\\top} - \\\\ - \\mathbf {v} _ {2} ^ {\\top} - \\\\ - \\mathbf {v} _ {3} ^ {\\top} - \\end{array} \\right]} _ {\\mathbf {V}} \\nabla f = \\underbrace {\\left[ \\begin{array}{l} \\frac {d f}{d t _ {1}} \\\\ \\frac {d f}{d t _ {2}} \\\\ \\frac {d f}{d t _ {3}} \\end{array} \\right]} _ {\\nabla_ {\\mathbf {V}} f}. \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20584" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.427, + 0.107 + ], + "angle": 0, + "content": "From Eq. (10), we can compute the SDF gradient by" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.116, + 0.47, + 0.133 + ], + "angle": 0, + "content": "\\[\n\\nabla f = \\mathbf {V} ^ {- 1} \\nabla_ {\\mathbf {V}} f. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.144, + 0.47, + 0.175 + ], + "angle": 0, + "content": "Numerically, we can approximate the directional derivative by" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.182, + 0.469, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\frac {d f}{d t _ {i}} \\approx \\frac {f (\\mathbf {o} _ {i} + (t _ {i} + \\Delta t) \\mathbf {v} _ {i}) - f (\\mathbf {o} _ {i} + (t _ {i} - \\Delta t) \\mathbf {v} _ {i})}{2 \\Delta t} \\\\ = \\frac {f \\left(\\mathbf {x} _ {i} + \\Delta t \\mathbf {v} _ {i}\\right) - f \\left(\\mathbf {x} _ {i} - \\Delta t \\mathbf {v} _ {i}\\right)}{2 \\Delta t}, \\quad i \\in \\{1, 2, 3 \\} \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.271, + 0.469, + 0.377 + ], + "angle": 0, + "content": "Equations (11) and (12) imply that, to numerically approximate the SDF gradient, we can first compute finite differences along three directions \\(\\mathbf{V}\\), then linearly transform the concatenated finite difference by \\(\\mathbf{V}^{-1}\\). We call this SDF gradient approximation directional finite difference. Equation (11) has a unique solution if the three directions are non-coplanar." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.378, + 0.469, + 0.499 + ], + "angle": 0, + "content": "The ordinary finite difference can be viewed as the special case of Eq. (11) by choosing the three coordinate axes as the ray directions. In this case, \\(\\mathbf{V}\\) becomes the identity matrix. Both DFD and FD require nearby SDF samples to approximate the SDF gradient. However, by tailoring the sampling strategy, we can reuse the SDF samples in volume rendering for DFD without additional SDF samples, as described in the following." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.507, + 0.325, + 0.523 + ], + "angle": 0, + "content": "3.3. Patch-based Ray Marching" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.469, + 0.635 + ], + "angle": 0, + "content": "DFD requires computing the finite difference in three directions. In volume rendering, we naturally have SDF samples along the viewing ray direction that can be used for finite difference. For the other two directions, we propose to sample patches of pixels instead of single pixels and march a plane on each patch of rays to locate the sampled points on parallel planes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.469, + 0.725 + ], + "angle": 0, + "content": "Specifically, we perform ordinary ray marching for the center pixel/ray of the patch and treat the sampled points as the intersections between a marching plane and the center ray. The points to be sampled on the remaining rays within the patch can then be found as the intersection of the planes and the rays:" + }, + { + "type": "equation", + "bbox": [ + 0.224, + 0.725, + 0.469, + 0.761 + ], + "angle": 0, + "content": "\\[\nt _ {j} = \\frac {t _ {i} \\mathbf {v} _ {i} ^ {\\top} \\mathbf {m}}{\\mathbf {v} _ {j} ^ {\\top} \\mathbf {m}}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.84 + ], + "angle": 0, + "content": "where \\(\\mathbf{m}\\) is the unit normal direction of the marching plane, \\(t_i\\) is the ray marching distance on the center ray with direction \\(\\mathbf{v}_i\\), and \\(t_j\\) is the distance from the intersection point to the ray origin (i.e., camera center) along the direction \\(\\mathbf{v}_j\\); for derivation see Sec. B.2 in SupMat." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We choose the marching plane to parallel the image plane where the patch is located (Fig. 3). With this sampling strategy, the three directions for DFD computation are the viewing ray direction and the world-space directions of \\( x \\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "and \\(y\\)-axis of the camera coordinates. The marching plane normal \\(\\mathbf{m}\\) becomes the world-space direction of the \\(z\\)-axis of the camera coordinates and is the same for all pixels/rays from the same view. This way, the SDF samples on the same ray can be used for volume rendering as Eq. (4), and the SDF samples neighbor to a point can be used for SDF gradient computation as Eq. (11). We use the central difference (i.e., Eq. (12)) for points with neighbors in both directions and forward/backward difference for boundary points." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.424 + ], + "angle": 0, + "content": "This sampling strategy is beneficial in three aspects. First, it allows us to reuse the SDF samples from neighboring rays to compute DFD. Second, it stabilizes training by ensuring the existence of \\(\\mathbf{V}^{-1}\\), since the viewing ray direction will never co-planer to the image plane. Third, it reduces the computation overhead during training because the three ray directions are identical for samples from the same pixel/ray. The inverse of ray directions \\(\\mathbf{V}^{-1}\\) can be pre-computed for each pixel before training. In our experiments, the computation takes about 0.6 sec for two million pixels. During training, only the \\(3\\times 3\\) matrix-vector multiplication Eq. (11) is required, which can almost be ignored compared to AD or FD." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.44, + 0.634, + 0.457 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.892, + 0.527 + ], + "angle": 0, + "content": "We evaluate our method quantitatively on the MVPS benchmark in Sec. 4.1, investigate different components of our method in Sec. 4.2 and evaluate our method qualitatively on our captured real-world data in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.73 + ], + "angle": 0, + "content": "Implementation details We apply SDM-UniPS [10] to images in each view to obtain normal maps. In each batch, we randomly sample 2048 patches of \\(3 \\times 3\\) pixels from all input normal maps. Like INGP [20], we skip the empty and occluded space during ray marching using NerfAcc [17]. The scene is bounded within a unit sphere, and we use a coarse-to-fine ray marching step decreasing log-linearly from \\(1e^{-2}\\) to \\(1e^{-3}\\). We optimize the shape of DiLiGenT-MV [16] benchmark objects for 5000 batches (roughly equal to 20 epochs), and 30000 batches for our captured data. The weights of loss terms are set as 1. We use the Adam optimizer with an initial learning rate \\(5e^{-3}\\). More details can be found in Sec. B in SupMat." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.741, + 0.719, + 0.757 + ], + "angle": 0, + "content": "4.1. Quantitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Dataset We evaluate our method using the only MVPS benchmark dataset DiLiGenT-MV [16]. It consists of five objects captured from 20 views, and 96 OLAT images are captured in each view, yielding a total of 1920 images per object. Each image is \\(512 \\times 612\\), and the total number of foreground object pixels is about 2.2 million. The camera is fixed during capture, and the object is placed on a turntable in a darkroom. The \"ground truth\" meshes are created using a 3D scanner." + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20585" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "Baselines We benchmark SuperNormal against several MVPS methods grouped into three categories. (1) Those that refine an initial point cloud or mesh and require manual effort, including R-MVPS [23] and B-MVPS [16]. (2) Those that use Poisson surface reconstruction (PSR) [13] to fuse multi-view depth and normal maps, including MVPSNet [35]. (3) Those that optimize an implicit neural surface representation, including UA-MVPS [12], PS-NeRF [31], MVAS [4], and our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.235, + 0.47, + 0.326 + ], + "angle": 0, + "content": "Evaluation metrics We use L2 Chamfer distance (CD) and F-score with threshold \\(\\tau_{F} = 0.5\\mathrm{mm}\\) to evaluate geometry accuracy [12, 15]. For CD and F-score, we only consider visible points from the input views by casting rays for all pixels and finding the first ray-mesh intersections (More details in Sec. B.4 in SupMat.)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.47, + 0.483 + ], + "angle": 0, + "content": "Results As reported in Table 1, our method achieves the best geometry accuracy among all compared MVPS methods (See Sec. A in SupMat. for normal accuracy). We attribute this superior performance to two aspects: 1) the powerful expressive capability of multi-resolution hash encoding [20] and 2) the fine-grained normal estimation method SDM-UniPS [10], as discussed in Sec. 4.2. Figure 4 visualizes the recovered shapes of different methods. Our method recovers fine-grained surface details on par with the \"GT\" meshes produced by a 3D scanner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.47, + 0.636 + ], + "angle": 0, + "content": "Table 1 also reports the average runtime measured on our machine with one RTX 4090Ti graphics card. The runtime of R-MVPS [23] and B-MVPS [16] is difficult to measure since both methods require intensive manual effort. The code of UA-MVPS [12] is not publicly available, and several hours per object are required, as reported in MVPSNet [35]. Both PS-NeRF [31] and MVAS [4] require hours per object since they optimize a dense MLP. Our method is the fastest among neural implicit representation-based methods, with less than one minute per object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.47, + 0.728 + ], + "angle": 0, + "content": "On the other hand, MVPSNet [35] is highly efficient because its pipeline mainly consists of feedforward inferencing a trained neural network and an sPSR [13] step, the latter of which has been optimized over one decade. Although our method is slower than MVPSNet [35], we achieve better reconstruction quality within a reasonable time." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.227, + 0.756 + ], + "angle": 0, + "content": "4.2. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Effect of normal maps Since multi-view normal integration is not bound to specific normal estimation methods, we investigate the effect of different normal estimation methods on normal integration methods. We test three sets of normal maps, estimated by SDPS [6] or SDM-UniPS [10], and the one rendered from the scanned mesh (i.e., \"GT\" normal maps provided by DiLiGenT-MV [16]). Table 2 reports the L2-Chamfer distance and F-score averaged on five DiLiGenT-MV [16] objects. Our method consistently" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "achieves better results than PS-NeRF [31] and MVAS [4] when using the same set of normal maps. This verifies that our method can better fuse the information from 2.5D normal maps to 3D shapes. Using the SoTA method SDM-UniPS [10] for normal estimation, our method realizes the best reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.188, + 0.892, + 0.31 + ], + "angle": 0, + "content": "Effect of spatial encoding Table 2 also compares the geometry accuracy between multi-resolution hash encoding (hash enc.) and positional encoding (pos. enc.). Our results verify again that multi-resolution hash encoding surpasses positional encoding in expressiveness, enhancing geometry accuracy. Therefore, the keys to superior MVPS performance are high-quality normal estimation, an expressive surface representation, and an effective normal fusion approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.315, + 0.892, + 0.452 + ], + "angle": 0, + "content": "Effect of SDF gradient computation Table 3 compares the geometry accuracy and runtime using finite difference (FD), Pytorch's automatic differentiation (AD), and directional finite difference (DFD) to compute the SDF gradient. All other parameter settings are kept the same. We do not observe a significant difference regarding reconstruction quality. However, DFD performs consistently faster than FD and AD by accelerating both the forward rendering and backward optimization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.639 + ], + "angle": 0, + "content": "Generalizability of DFD In principle, DFD and patch-based sampling strategy apply to any SDF-based volume rendering like NeuS [29] or Neuralangelo [18]. To verify this, we incorporate DFD and patch-based sampling into Neuralangelo \\([18]^{1}\\). We follow the coarse-to-fine sampling strategy of Neuralangelo [18] but use DFD instead of FD for SDF gradient approximation. We do not use mask supervision for Neuralangelo [18]. As shown in Figure 5, training time is reduced by \\(25\\%\\) on a scene from the DTU benchmark [11], even \\(2.25\\times\\) more pixels are looped in training. The result suggests that DFD could be a general strategy to accelerate NeuS-like multi-view reconstruction [29]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.648, + 0.844, + 0.664 + ], + "angle": 0, + "content": "4.3. Qualitative results on our captured data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.868 + ], + "angle": 0, + "content": "To further demonstrate the performance of SuperNormal in fine-grained surface detail recovery, we also captured objects with more complex details, as shown in Fig. 6. Thanks to the generalizability of SDM-UniPS [10], we can capture the objects under casual light conditions (e.g., within an apartment) without ensuring a darkroom setup. We manually move a video light during the capture without any photometric or geometric light calibration. In total, we capture 18 viewpoints \\(\\times 13\\) light conditions for each object (1 under ambient light for SfM). We use Metashape [1] to calibrate camera parameters, use SAM [14] to create foreground masks, and apply SDM-UniPS [10] to every 12 image to infer the per-view normal map." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1Our modified code of Neuralangelo [18] is available at https:// github.com/xucao-42/Neuralangelo_DFD.git." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20586" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.894, + 0.12 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation on DiLiGenT-MV [16] benchmark. Red and orange cells indicate the best and the second best results respectively. On average, our approach achieves the best reconstruction quality at the second fastest speed." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.123, + 0.891, + 0.233 + ], + "angle": 0, + "content": "
MethodsL2 Chamfer distance [mm] (↓)F-score (τF=0.5 mm) (↑)Runtime (↓) Average
BearBuddhaCowPot2ReadingAverageBearBuddhaCowPot2ReadingAverage
R-MVPS [23]1.0700.3970.4401.5040.5610.7940.2620.6980.7600.1980.5190.487NA
B-MVPS [16]0.2120.2540.0910.2010.2590.2030.9580.9020.9860.9460.9140.941NA
MVPSSNet [35]0.3170.2790.2550.3100.2480.2820.8130.8660.8890.8380.9180.86537 secs
UA-MVPS [12]0.4140.4520.3260.4140.3820.3980.7070.6690.7980.7310.7620.733several hrs
PS-NeRF [31]0.2600.3140.2870.2540.3520.2930.8980.8060.8560.9190.7850.8538 hrs
MVAS [4]0.2430.3570.2160.1970.5220.3070.9090.7540.9070.9620.5460.81670 mins
Ours0.1840.2180.1930.1500.2230.1940.9830.9500.9770.9940.9160.96248 secs
" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.244, + 0.892, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.707, + 0.838, + 0.723 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of recovered shapes of Buddha and Pot2 and their error maps. Best viewed on screen." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.746, + 0.47, + 0.788 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation using normal maps estimated by different methods as inputs. The metrics are averaged over five DiLiGenT-MV [16] objects. Darker colors indicate better results." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.799, + 0.469, + 0.862 + ], + "angle": 0, + "content": "
L2 Chamfer distance [mm] (↓)F-score (τF=0.5 mm) (↑)
SDPS [6]SDM-UniPS [10]GTSDPS [6]SDM-UniPS [10]GT
MVAS [4]0.3070.3320.2920.8160.8120.859
PS-NeRF [31]0.2930.2960.2240.8530.8440.938
Ours (pos, enc.)0.3010.2280.1950.8490.9420.977
Ours (hash enc.)0.2810.1940.1140.8600.9620.998
" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.744, + 0.892, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.807, + 0.894, + 0.863 + ], + "angle": 0, + "content": "Figure 5. DFD and patch-based sampling accelerate Neuralangelo's training without quality degradation. We train both alternatives for \\(20\\mathrm{k}\\) batches. For FD, 512 random rays are sampled per batch; for DFD, 128 random patches of \\(3\\times 3\\) rays per batch." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20587" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.145 + ], + "angle": 0, + "content": "Table 3. Ablation study using different strategies for SDF gradients computation in volume rendering. We report the metrics averaged over five DiLiGenT-MV [16] objects. FD: Finite difference, AD: Automatic differentiation, DFD: Directional finite difference." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.157, + 0.428, + 0.228 + ], + "angle": 0, + "content": "
AccuracyRuntime (Sec)
CDF-ScoreForwardBackwardTotal
FD0.1920.96347.770.3124.6
AD0.1950.96129.442.478.1
DFD0.1940.96119.622.548.3
" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.243, + 0.273, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.242, + 0.462, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.363, + 0.47, + 0.434 + ], + "angle": 0, + "content": "Figure 6. Our real-world data capture setup. The target object is placed on a turntable before a dark sheet under general indoor illumination. There is no need to prepare a darkroom. We use an iPhone to take images and move a video light to illuminate the object from the camera side. Objects are \\(6\\mathrm{cm}\\) to \\(15\\mathrm{cm}\\) high." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.47, + 0.569 + ], + "angle": 0, + "content": "Figures 1 and 7 display the results from NeuS2 [30], SuperNormal, and a structured-light based commercial 3D scanner EinScan SE. Although NeuS2 [30] and our method use the same multi-resolution hash encoding for fine-grained scene representation, the surface details recovered from multi-view color images are smoothed out. Our method, benefitting from the normal map inputs, recovers more fine-grained surface details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.571, + 0.469, + 0.692 + ], + "angle": 0, + "content": "We find that our scanner performs poorly in highly concave regions (Fig. 7 above). This is because the scanner uses a wide baseline stereo camera (about \\(15\\mathrm{cm}\\) apart), and small concave regions create self-occlusions, thus reducing the reconstruction quality. In principle, normal estimation does not suffer from self-occlusion because it is based on single-view observations. Fusing multi-view normal maps can then produce better results in concave regions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.707, + 0.192, + 0.723 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.469, + 0.885 + ], + "angle": 0, + "content": "We have demonstrated SuperNormal, a fast, high-fidelity 3D reconstruction approach using multi-view normal maps estimated by photometric stereo. SuperNormal outperforms other neural MVPS methods in speed and reconstruction quality on the benchmark dataset. Compared to multi-view neural reconstruction methods, using normal maps can exploit the expressive power of multi-resolution hash encoding, and produce scanner-level surface details. Further, we have shown the effectiveness of directional finite differences in expediting training without sacrificing quality." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "SuperNormal provides a promising direction to make" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.088, + 0.892, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.458, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison using our objects. Our approach recovers better surface detail than multi-view stereo approaches and achieves competitive results with commercial 3D scanners." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.513, + 0.703, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.63, + 0.564, + 0.638 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.514, + 0.814, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.629, + 0.803, + 0.639 + ], + "angle": 0, + "content": "OmniData normal" + }, + { + "type": "image", + "bbox": [ + 0.814, + 0.515, + 0.887, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.63, + 0.864, + 0.638 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.65, + 0.892, + 0.692 + ], + "angle": 0, + "content": "Figure 8. Limitations. (Left) Fault artifacts are highlighted in red circles. (Right) Our method is sensitive to the inaccuracy in normal maps estimated by monocular methods (e.g., OmniData [7])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.706, + 0.892, + 0.767 + ], + "angle": 0, + "content": "high-quality 3D scanning more accessible. Unlike costly commercial 3D scanners, our approach is viable with everyday devices like smartphones and basic equipment such as a tripod, video light, and a turntable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Limitations Our method may produce fault-like artifacts, as shown in Fig. 8. The artifacts exist no matter our SDF gradient computation approach (i.e., AD, FD, or DFD) and require further investigation. Moreover, the performance depends on mask supervision and high-quality normal maps (e.g., those obtained from photometric stereo). Our method does not work well with the monocular normal estimation methods like OmniData [7] (Fig. 8)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20588" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.452, + 0.129 + ], + "angle": 0, + "content": "[1] Agisoft metashape. https://www.agisoft.com.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.13, + 0.469, + 0.184 + ], + "angle": 0, + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.469, + 0.226 + ], + "angle": 0, + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-NeRF: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.229, + 0.469, + 0.283 + ], + "angle": 0, + "content": "[4] Xu Cao, Hiroaki Santo, Fumio Okura, and Yasuyuki Matsushita. Multi-view azimuth stereo via tangent space consistency. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 825-834, 2023. 1, 2, 6, 7, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.469, + 0.34 + ], + "angle": 0, + "content": "[5] Ju Yong Chang, Kyoung Mu Lee, and Sang Uk Lee. Multiview normal field integration using level set methods. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 1-8. IEEE, 2007. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.342, + 0.469, + 0.382 + ], + "angle": 0, + "content": "[6] Guanying Chen, Kai Han, Boxin Shi, Yasuyuki Matsushita, and Kwan-Yee K. Wong. SDPS-Net: Self-calibrating deep photometric stereo networks. In CVPR, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.469, + 0.436 + ], + "angle": 0, + "content": "[7] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In ICCV, pages 10786-10796, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.469, + 0.493 + ], + "angle": 0, + "content": "[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of Machine Learning and Systems 2020, pages 3569-3579. 2020. 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.496, + 0.469, + 0.549 + ], + "angle": 0, + "content": "[9] Carlos Hernandez, George Vogiatzis, and Roberto Cipolla. Multiview photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 30(3):548-554, 2008. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.551, + 0.469, + 0.592 + ], + "angle": 0, + "content": "[10] Satoshi Ikehata. Scalable, detailed and mask-free universal photometric stereo. In Proc. of Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.594, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[11] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, pages 406-413. IEEE, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[12] Berk Kaya, Suryansh Kumar, Carlos Oliveira, Vittorio Ferrari, and Luc Van Gool. Uncertainty-aware deep multi-view photometric stereo. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 12601-12611, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[13] Michael Kazhdan and Hugues Hoppe. Screened Poisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proc. of International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[15] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] Min Li, Zhenglong Zhou, Zhe Wu, Boxin Shi, Changyu Diao, and Ping Tan. Multi-view photometric stereo: A ro" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.134 + ], + "angle": 0, + "content": "bust solution and benchmark dataset for spatially varying isotropic materials. IEEE Transactions on Image Processing, 29:4159-4173, 2020. 2, 5, 6, 7, 8, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[17] Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. NerfAcc: Efficient sampling accelerates nerfs. In Proc. of International Conference on Computer Vision (ICCV), pages 18537-18546, 2023. 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[18] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 8456-8465, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. of European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[21] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proc. of International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[22] Stanley Osher, Ronald Fedkiw, and K Piechor. Level set methods and dynamic implicit surfaces. Appl. Mech. Rev., 57(3):B15-B15, 2004. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[23] Jaesik Park, Sudipta N Sinha, Yasuyuki Matsushita, Yu-Wing Tai, and In So Kweon. Robust multiview photometric stereo using planar mesh parameterization. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 39(8):1591-1604, 2016. 2, 6, 7, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[24] Radu Alexandru Rosu and Sven Behnke. Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices. In CVPR, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[25] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. of European Conference on Computer Vision (ECCV), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[26] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my NeRF: A dataset for novel view synthesis and relighting of real world objects. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 20762-20772, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[27] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[28] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view recon" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20589" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "struction. Advances in Neural Information Processing Systems (NeurIPS), 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[30] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. NeuS2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proc. of International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 4, 8, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.468, + 0.26 + ], + "angle": 0, + "content": "[31] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K. Wong. PS-NeRF: Neural inverse rendering for multi-view photometric stereo. In Proc. of European Conference on Computer Vision (ECCV), 2022. 1, 2, 6, 7, 11, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.468, + 0.331 + ], + "angle": 0, + "content": "[32] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.468, + 0.387 + ], + "angle": 0, + "content": "[33] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems (NeurIPS), 34:4805-4815, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.468, + 0.444 + ], + "angle": 0, + "content": "[34] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. NIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.468, + 0.514 + ], + "angle": 0, + "content": "[35] Dongxu Zhao, Daniel Lichy, Pierre-Nicolas Perrin, Jan-Michael Frahm, and Soumyadip Sengupta. MVPSNet: Fast generalizable multi-view photometric stereo. In Proc. of International Conference on Computer Vision (ICCV), pages 12525-12536, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.468, + 0.585 + ], + "angle": 0, + "content": "[36] Fuqiang Zhao, Yuheng Jiang, Kaixin Yao, Jiakai Zhang, Liao Wang, Haizhao Dai, Yuhui Zhong, Yingliang Zhang, Minye Wu, Lan Xu, and Jingyi Yu. Human performance modeling and rendering via neural animated mesh. ACM Trans. Graph., 41(6), 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20590" + } + ] +] \ No newline at end of file diff --git a/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/a4b33f00-d1f9-45e2-a62c-bb25956b58ec_origin.pdf b/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/a4b33f00-d1f9-45e2-a62c-bb25956b58ec_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b2cb839cbca54131806ddcd49f85a1ccb0de95bf --- /dev/null +++ b/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/a4b33f00-d1f9-45e2-a62c-bb25956b58ec_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:666a3bedc1356ee514cbd0a312a8d902ba3125d1709ab8b99341a046f62866ed +size 9931946 diff --git a/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/full.md b/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e681021f3ac9019631ad7b2421bd377bedca1907 --- /dev/null +++ b/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/full.md @@ -0,0 +1,325 @@ +# SuperNormal: Neural Surface Reconstruction via Multi-View Normal Integration + +Xu Cao Takafumi Taketomi + +CyberAgent + +{xu_cao, taketomi_takafumi}@cyberagent.co.jp + +# Abstract + +We present SuperNormal, a fast, high-fidelity approach to multi-view 3D reconstruction using surface normal maps. With a few minutes, SuperNormal produces detailed surfaces on par with 3D scanners. We harness volume rendering to optimize a neural signed distance function (SDF) powered by multi-resolution hash encoding. To accelerate training, we propose directional finite difference and patch-based ray marching to approximate the SDF gradients numerically. While not compromising reconstruction quality, this strategy is nearly twice as efficient as analytical gradients and about three times faster than axis-aligned finite difference. Experiments on the benchmark dataset demonstrate the superiority of SuperNormal in efficiency and accuracy compared to existing multi-view photometric stereo methods. On our captured objects, SuperNormal produces more fine-grained geometry than recent neural 3D reconstruction methods. Our code is available at https://github.com/CyberAgentAILab/SuperNormal. + +# 1. Introduction + +Recovering high-quality 3D geometry of real-world objects from their multi-view images is a long-standing challenge in computer vision. Recently, neural implicit surface-based methods have shown remarkable reconstruction results. Compared to traditional multi-view stereo (MVS) methods [25], neural methods are more robust to view-dependent observations and textureless surfaces [29]. Furthermore, the reconstruction procedure has become highly efficient [30] by introducing multi-resolution hash coding [20]. However, even though this spatial encoding allows fine-grained geometry to be represented, shape reconstruction from multi-view images tends to smooth out high-frequency surface details, as shown in Fig. 1 bottom right. + +Multi-view photometric stereo (MVPS), on the other hand, aims to recover pixel-level high-frequency surface detail by introducing additional lighting conditions during + +![](images/fd4a7330b81a55a9ab69dd7d0dd6a582b2dafb4ed9a6f47cfb78157305ba94d2.jpg) + +![](images/c479ca7ec93608d18f5b3a6564724cec7e349ef64abcbf2f3db03bd6dae3ed92.jpg) +Figure 1. (Top) From multi-view surface normal maps, SuperNormal recovers fine-grained surface details comparable to 3D scanners while being orders of magnitude faster than the existing MVPS method PS-NeRF [31]. (Bottom) Using normal maps produces more faithful high-frequency details than the MVS-based method NeuS2 [30], although both use multi-resolution hash encoding [20]. + +the image acquisition [9]. In a typical workflow, a surface normal map is first recovered for each view, recording per-pixel surface orientation information. The normal maps estimated at different viewpoints are then fused into a 3D model, also known as multi-view normal integration [5]. However, existing normal fusion methods struggle to reflect the details of the normal maps in the shape of the recovered 3D model. Moreover, the reconstruction process takes hours even when only a few low-resolution normal maps are used [4, 31]. Due to the lack of a fast and accurate multiview normal fusion method, current multi-view photometric stereo results remain unsatisfactory. + +This paper presents SuperNormal to unite the best of both worlds. Normal maps with pixel-level surface details are utilized to exploit the expressive power and efficiency + +of multi-resolution hash encoding-powered neural implicit surface. Using normal maps alone avoids per-scene reflectance modeling and optimization, thus improving training efficiency compared to using color images. In a few minutes, our method can produce fine-grained geometry on par with commercial 3D scanners. + +Specifically, we use volume rendering to optimize a neural signed distance function (SDF) such that its rendered normal maps are consistent with input normal maps. To further improve the training efficiency, we propose a directional finite difference and patch-based ray marching strategy to approximate the SDF gradients numerically. This strategy kills two birds with one stone: SDF values evaluated at sampled points are all used both for SDF-based volume rendering and SDF gradient approximation. This way, we avoid second-order derivative computation during training. Unlike traditional axis-aligned finite difference, directional finite difference avoids redundant SDF evaluations (Fig. 2). Without compromising reconstruction quality, our method is nearly twice as fast as using analytical gradients and about three times faster than finite difference. + +Compared to existing neural MVPS methods, our method is two orders of magnitude faster and recovers better surface details (Fig. 1 above). We also capture objects with complex geometry and show that our method recovers more faithful surface details than the recent neural 3D reconstruction method [30] (Fig. 1 bottom). Taking advantage of recent advances in photometric stereo, we can estimate high-quality normal maps from images captured under common lighting conditions with inexpensive equipment [10]. Combined with SuperNormal, high-fidelity 3D objects can be reconstructed using inexpensive equipment in widely accessible environments. + +In summary, this paper's key contributions are: + +- An effective neural 3D reconstruction approach using multi-view normal maps, enhancing the reconstruction quality to near 3D scanner levels; +- An efficient method for computing numerical gradients in SDF-based neural rendering, speeding up the reconstruction process; and +- A comprehensive evaluation using benchmark objects with MVPS approaches and an assessment using our captured objects with MVS approaches. + +# 2. Related work + +# 2.1. Multi-view photometric stereo (MVPS) + +MVPS aims at high-fidelity 3D reconstruction using multi-view observations under varying light conditions [9] (usually one-light-at-time (OLAT) images [16, 26]). + +Existing methods differ in what geometry-related information is inferred in each view and how information from different views is fused into a complete 3D model. Chang et + +![](images/cfac92fe142e4ba7f830a110d9ce6d48012fe1d807e03238eef7b829e7827220.jpg) +Figure 2. (Left) In SDF-based volume rendering, approximating SDF gradients using axis-aligned finite difference [18, 36] requires additional SDF samples at neighboring positions of volume rendering points, increasing computational complexity and training time. (Right) Our strategy of patch-based ray marching and directional finite difference effectively uses all SDF samples for both SDF gradient approximation and volume rendering. + +al. [5] proposes a level set method [22] to fuse multi-view normal maps into a volumetric representation of the shape. R-MVPS [23] uses the multi-view normal maps to refine a coarse mesh estimated by multi-view stereo. B-MVPS [16] initialize the shape as a sparse point cloud estimated by structure from motion (SfM), then propagates the known point spatially using the normal maps. + +Recent MVPS methods have undergone a paradigm shift since the advent of NeRF [19] by using a neural implicit scene representation, usually parameterized as a multi-layer perception (MLP). PS-NeRF [31] recovers a neural field based on UNISURF [21] by regularizing the gradients of the neural field using the normal maps. UA-MVPS [12] fuses the multi-view depth and normal maps into a neural SDF using neural inverse rendering while considering per-pixel confidence. MVAS [4] shows utilizing the azimuth components of multi-view normal maps is sufficient to recover the neural implicit surface. While most methods only use images in the same view for normal map estimation, MVPSNet [35] utilizes neighboring views to infer per-view depth and normal maps jointly and applies screened Poisson surface reconstruction (sPSR) [13] to obtain the shape. + +However, the reconstructed surfaces of these methods tend to be smoothed and do not reflect the high-frequency details of the input normal maps. Our approach introduces multi-resolution hash encoding and directional finite difference to improve the implicit neural field's expressive power and training efficiency. + +# 2.2. Neural 3D reconstruction + +The seminal work NeRF [19] and follow-up works [2, 3, 27] have demonstrated remarkable novel view synthesis quality by volume rendering a neural implicit field. For better geometry reconstruction quality, IDR [32] imposes eikonal + +regularization [8] such that the neural field approximates a valid SDF. Later, the concurrent works VolSDF [33] and NeuS [29] develop SDF-based volume rendering, which bridges SDF values to volume rendering weights and improves the geometry fidelity. However, these methods require several hours to recover one object due to optimizing a deep coordinate-based MLP. + +Multi-resolution hash encoding has been proposed to improve expressive power and accelerate the learning of neural implicit field [20]. Several efforts have been made to incorporate this expressive spatial encoding into neural SDF learning. Since the official CUDA implementation does not support the double propagation deduced by the eikonal regularization, PermutoSDF [24] and NeuS2 [30] re-implement the CUDA programming. InstantNSR [36] and Neuralangelo [18] use finite difference instead of automatic differentiation to avoid double propagation. + +We propose directional finite difference, a numerical way to compute SDF gradients that avoids double propagation and is even faster than automatic differentiation when combined with a tailored patch-based sampling strategy. + +# 2.3. Neural 3D reconstruction with normals + +Normal maps have been used as auxiliary information to facilitate neural scene reconstruction in MonoSDF [34] and NeuRIS [28]. These works target scene-level reconstruction, and normal maps help infer the geometry of textureless smooth regions (e.g., walls) where using color images is likely to break down. + +Instead, our method is particularly useful for single objects with complex surface details. We use photometric stereo [10] to estimate normal maps from images under varying light conditions. This produces high-quality normal maps that can be used alone for shape reconstruction. + +# 3. Approach + +We aim to recover the surface given a set of normal maps, object masks, and camera intrinsic and extrinsic parameters. Since photometric stereo methods estimate the normal maps in camera space, we first apply the camera-to-world rotation to normal maps to obtain world-space normal maps. Figure 3 depicts our pipeline. We represent the geometry as a parametric SDF and optimize SDF parameters such that its rendered gradient and opacity values are consistent with input normal vectors and mask values, respectively. + +# 3.1. Pipeline + +SDF-based volume rendering (Forward pass) A signed distance function (SDF) implicitly captures the geometry by assigning each spatial point with the signed distance to its closest surface point. The surface $\mathcal{M}$ can then be repre + +sented as the zero-level-set of the SDF: + +$$ +\mathcal {M} = \{\mathbf {x} \mid f (\mathbf {x}) = 0 \}. \tag {1} +$$ + +We parameterize the SDF using multi-resolution hash encoding followed by a shallow MLP. In multi-resolution hash encoding, the 3D space is discretized into a set of multi-resolution 3D grids. Each grid cell is associated with a unique hash value mapping the 3D coordinates to a learnable feature vector. Specifically, $h(\mathbf{x};\phi)$ is the feature vector obtained by concatenating feature vectors of different levels $h(\mathbf{x}) = [\mathbf{x},h_1(\mathbf{x}),\dots,h_L(\mathbf{x})]$ . Then, the SDF can be parameterized as + +$$ +f (\mathbf {x}) = f (h (\mathbf {x}; \phi); \theta), \tag {2} +$$ + +where $\phi$ and $\theta$ are learnable hash-encoding feature vectors and MLP parameters. Multi-resolution hash encoding has proven to speed up training and improve the expressive power of the neural SDF [20, 30, 36]. + +During training, we evaluate the SDF values at points sampled on the rays cast from the camera centers toward the scene. NeuS [29] proposes an unbiased and occlusion-aware way to transfer the SDF samples to volume rendering opacities. Given $N$ ordered 3D points $\{\mathbf{x}_i\}_{i=0}^N$ on a ray and their SDF values $\{f(\mathbf{x}_i)\}_{i=0}^N$ , the opacity at each point is designed as + +$$ +\alpha_ {i} = \max \left(\frac {\phi_ {s} (f (\mathbf {x} _ {i})) - \phi_ {s} (f (\mathbf {x} _ {i + 1}))}{\phi_ {s} (f (\mathbf {x} _ {i}))}, 0\right), \tag {3} +$$ + +where $\phi_s(x) = \frac{1}{1 + \exp(-sx)}$ is the sigmoid function with a trainable sharpness $s$ . + +With this SDF-based opacity, we can render the normal and opacity for the pixel $\mathbf{p}$ as: + +$$ +\hat {\mathbf {n}} (\mathbf {p}) = \sum_ {i = 0} ^ {N} T _ {i} \alpha_ {i} \nabla f (\mathbf {x} _ {i}), \quad \hat {o} (\mathbf {p}) = \sum_ {i = 0} ^ {N} T _ {i} \alpha_ {i} \tag {4} +$$ + +$$ +\text {w i t h} \quad T _ {i} = \prod_ {j = 0} ^ {i - 1} (1 - \alpha_ {j}). +$$ + +From Eqs. (3) and (4), the neural SDF must be evaluated at least once at sampled points for volume rendering. In Secs. 3.2 and 3.3, we show that these SDF samples can be reused for SDF gradient approximation, thus reducing the computational amount and speeding up the training. + +Neural SDF Training (Backward pass) To train the neural SDF, we supervise the rendered normals and opacities using input ones $\mathbf{n}(\mathbf{p})$ and $\hat{o} (\mathbf{p})$ . The loss function consists + +![](images/8a20c984d4c8cc382b0dd432ebb04c4d3a096b69675c56f6399dfe04c36dc528.jpg) +Figure 3. Approach overview. (a) Given input views, we randomly sample patches of pixels, cast rays from those pixels, and march a plane along the patch of rays (Sec. 3.3). We treat the ray-plane intersections as sampled points and evaluate their SDF values via the neural SDF. (b) SDF samples neighbor to a point are used to compute the SDF gradients using directional finite difference (Sec. 3.2). (c) SDF samples on the same rays are used for SDF-based volume rendering (Sec. 3.1). + +of three terms: The normal, the mask, and the eikonal term: + +$$ +\begin{array}{l} \mathcal {L} = \overbrace {\sum_ {\mathbf {p}} \left(\hat {\mathbf {n}} (\mathbf {p}) - \mathbf {n} (\mathbf {p})\right) ^ {2}} ^ {\text {n o r m a l}} + \lambda_ {1} \overbrace {\sum_ {\mathbf {p}} \operatorname {B C E} \left(\hat {o} (\mathbf {p}) , o (\mathbf {p})\right)} ^ {\text {m a s k}} \tag {5} \\ + \lambda_ {2} \underbrace {\sum_ {\mathbf {x}} \left(\| \nabla f (\mathbf {x}) \| _ {2} - 1\right) ^ {2}} _ {e i k o n a l}, \\ \end{array} +$$ + +where $\mathbf{p}$ is the sampled pixels in a batch, $\mathbf{x}$ is the sampled points on rays, and BCE is the binary cross entropy function. The normal term encourages the rendered SDF gradients to be consistent with the input normal vectors so that accurate geometry details are recovered. The mask term aligns the shape's silhouette with input masks and can be considered the boundary condition. The eikonal term enforces the SDF gradient norm to be close to 1 almost everywhere so that the neural SDF is approximately valid [22]. + +# 3.2. Directional Finite Difference + +As seen, both the normal and eikonal terms backpropagate the loss gradient via the SDF gradient to neural SDF parameters (i.e., double backpropagation $\frac{\partial\mathcal{L}}{\partial\nabla f}\frac{\partial\nabla f}{\partial\theta}$ ). The most common way is to use Pytorch's automatic differentiation (AD) due to implementation simplicity. However, the multi-resolution hash encoding package does not officially support double propagation. Common workarounds include CUDA programming [24, 30] or use finite difference (FD) [18, 36] to bypass double propagation. However, CUDA programming can be inflexible in that it shrinks the design space (e.g., NeuS2 [30] requires ReLU activation), and FD will slow down the training. Consider the ordinary + +axis-aligned FD: + +$$ +\nabla f (\mathbf {x}) \approx \left[ \begin{array}{l} \frac {f (\mathbf {x} + \epsilon_ {x}) - f (\mathbf {x} - \epsilon_ {x})}{2 \epsilon} \\ \frac {f (\mathbf {x} + \epsilon_ {y}) - f (\mathbf {x} - \epsilon_ {y})}{2 \epsilon} \\ \frac {f (\mathbf {x} + \epsilon_ {z}) - f (\mathbf {x} - \epsilon_ {z})}{2 \epsilon} \end{array} \right], \tag {6} +$$ + +where $\epsilon_{* \in \{x, y, z\}}$ is an offset vector along $x, y$ , or $z$ axis with a tiny step size $\epsilon$ . Computing the gradient at one point requires evaluating the SDF at its nearby points. These additional SDF samples are also included in the computation graph in the forward pass. Consequently, using FD significantly slows the forward rendering and backward optimization even if the efficient spatial encoding is used [18]. + +To avoid redundant SDF evaluations and double backpropagation, we propose directional finite difference (DFD) to reuse the volume rendering SDF samples for SDF gradients approximation. Consider the SDF value at a 3D point $\mathbf{x}$ parameterized by the ray origin $\mathbf{o}$ and direction $\mathbf{v}$ : + +$$ +f (\mathbf {x}) = f (\mathbf {o} + t \mathbf {v}), \tag {7} +$$ + +where $t$ is the distance to the ray origin. From Eq. (7), we have + +$$ +\frac {d f}{d t} = \nabla f (\mathbf {x}) ^ {\top} \mathbf {v} := \nabla_ {\mathbf {v}} f. \tag {8} +$$ + +Equation (8) is exactly the directional derivative of the SDF, which projects the SDF gradient onto the ray direction. Suppose we parameterize the same point along three different directions as + +$$ +\mathbf {x} = \mathbf {o} _ {i} + t _ {i} \mathbf {v} _ {i}, \quad i \in \{1, 2, 3 \}. \tag {9} +$$ + +Applying Eq. (8) then yields three linear equations: + +$$ +\underbrace {\left[ \begin{array}{l} - \mathbf {v} _ {1} ^ {\top} - \\ - \mathbf {v} _ {2} ^ {\top} - \\ - \mathbf {v} _ {3} ^ {\top} - \end{array} \right]} _ {\mathbf {V}} \nabla f = \underbrace {\left[ \begin{array}{l} \frac {d f}{d t _ {1}} \\ \frac {d f}{d t _ {2}} \\ \frac {d f}{d t _ {3}} \end{array} \right]} _ {\nabla_ {\mathbf {V}} f}. \tag {10} +$$ + +From Eq. (10), we can compute the SDF gradient by + +$$ +\nabla f = \mathbf {V} ^ {- 1} \nabla_ {\mathbf {V}} f. \tag {11} +$$ + +Numerically, we can approximate the directional derivative by + +$$ +\begin{array}{l} \frac {d f}{d t _ {i}} \approx \frac {f (\mathbf {o} _ {i} + (t _ {i} + \Delta t) \mathbf {v} _ {i}) - f (\mathbf {o} _ {i} + (t _ {i} - \Delta t) \mathbf {v} _ {i})}{2 \Delta t} \\ = \frac {f \left(\mathbf {x} _ {i} + \Delta t \mathbf {v} _ {i}\right) - f \left(\mathbf {x} _ {i} - \Delta t \mathbf {v} _ {i}\right)}{2 \Delta t}, \quad i \in \{1, 2, 3 \} \tag {12} \\ \end{array} +$$ + +Equations (11) and (12) imply that, to numerically approximate the SDF gradient, we can first compute finite differences along three directions $\mathbf{V}$ , then linearly transform the concatenated finite difference by $\mathbf{V}^{-1}$ . We call this SDF gradient approximation directional finite difference. Equation (11) has a unique solution if the three directions are non-coplanar. + +The ordinary finite difference can be viewed as the special case of Eq. (11) by choosing the three coordinate axes as the ray directions. In this case, $\mathbf{V}$ becomes the identity matrix. Both DFD and FD require nearby SDF samples to approximate the SDF gradient. However, by tailoring the sampling strategy, we can reuse the SDF samples in volume rendering for DFD without additional SDF samples, as described in the following. + +# 3.3. Patch-based Ray Marching + +DFD requires computing the finite difference in three directions. In volume rendering, we naturally have SDF samples along the viewing ray direction that can be used for finite difference. For the other two directions, we propose to sample patches of pixels instead of single pixels and march a plane on each patch of rays to locate the sampled points on parallel planes. + +Specifically, we perform ordinary ray marching for the center pixel/ray of the patch and treat the sampled points as the intersections between a marching plane and the center ray. The points to be sampled on the remaining rays within the patch can then be found as the intersection of the planes and the rays: + +$$ +t _ {j} = \frac {t _ {i} \mathbf {v} _ {i} ^ {\top} \mathbf {m}}{\mathbf {v} _ {j} ^ {\top} \mathbf {m}}, \tag {13} +$$ + +where $\mathbf{m}$ is the unit normal direction of the marching plane, $t_i$ is the ray marching distance on the center ray with direction $\mathbf{v}_i$ , and $t_j$ is the distance from the intersection point to the ray origin (i.e., camera center) along the direction $\mathbf{v}_j$ ; for derivation see Sec. B.2 in SupMat. + +We choose the marching plane to parallel the image plane where the patch is located (Fig. 3). With this sampling strategy, the three directions for DFD computation are the viewing ray direction and the world-space directions of $x$ + +and $y$ -axis of the camera coordinates. The marching plane normal $\mathbf{m}$ becomes the world-space direction of the $z$ -axis of the camera coordinates and is the same for all pixels/rays from the same view. This way, the SDF samples on the same ray can be used for volume rendering as Eq. (4), and the SDF samples neighbor to a point can be used for SDF gradient computation as Eq. (11). We use the central difference (i.e., Eq. (12)) for points with neighbors in both directions and forward/backward difference for boundary points. + +This sampling strategy is beneficial in three aspects. First, it allows us to reuse the SDF samples from neighboring rays to compute DFD. Second, it stabilizes training by ensuring the existence of $\mathbf{V}^{-1}$ , since the viewing ray direction will never co-planer to the image plane. Third, it reduces the computation overhead during training because the three ray directions are identical for samples from the same pixel/ray. The inverse of ray directions $\mathbf{V}^{-1}$ can be pre-computed for each pixel before training. In our experiments, the computation takes about 0.6 sec for two million pixels. During training, only the $3\times 3$ matrix-vector multiplication Eq. (11) is required, which can almost be ignored compared to AD or FD. + +# 4. Experiments + +We evaluate our method quantitatively on the MVPS benchmark in Sec. 4.1, investigate different components of our method in Sec. 4.2 and evaluate our method qualitatively on our captured real-world data in Sec. 4.3. + +Implementation details We apply SDM-UniPS [10] to images in each view to obtain normal maps. In each batch, we randomly sample 2048 patches of $3 \times 3$ pixels from all input normal maps. Like INGP [20], we skip the empty and occluded space during ray marching using NerfAcc [17]. The scene is bounded within a unit sphere, and we use a coarse-to-fine ray marching step decreasing log-linearly from $1e^{-2}$ to $1e^{-3}$ . We optimize the shape of DiLiGenT-MV [16] benchmark objects for 5000 batches (roughly equal to 20 epochs), and 30000 batches for our captured data. The weights of loss terms are set as 1. We use the Adam optimizer with an initial learning rate $5e^{-3}$ . More details can be found in Sec. B in SupMat. + +# 4.1. Quantitative evaluation + +Dataset We evaluate our method using the only MVPS benchmark dataset DiLiGenT-MV [16]. It consists of five objects captured from 20 views, and 96 OLAT images are captured in each view, yielding a total of 1920 images per object. Each image is $512 \times 612$ , and the total number of foreground object pixels is about 2.2 million. The camera is fixed during capture, and the object is placed on a turntable in a darkroom. The "ground truth" meshes are created using a 3D scanner. + +Baselines We benchmark SuperNormal against several MVPS methods grouped into three categories. (1) Those that refine an initial point cloud or mesh and require manual effort, including R-MVPS [23] and B-MVPS [16]. (2) Those that use Poisson surface reconstruction (PSR) [13] to fuse multi-view depth and normal maps, including MVPSNet [35]. (3) Those that optimize an implicit neural surface representation, including UA-MVPS [12], PS-NeRF [31], MVAS [4], and our method. + +Evaluation metrics We use L2 Chamfer distance (CD) and F-score with threshold $\tau_{F} = 0.5\mathrm{mm}$ to evaluate geometry accuracy [12, 15]. For CD and F-score, we only consider visible points from the input views by casting rays for all pixels and finding the first ray-mesh intersections (More details in Sec. B.4 in SupMat.) + +Results As reported in Table 1, our method achieves the best geometry accuracy among all compared MVPS methods (See Sec. A in SupMat. for normal accuracy). We attribute this superior performance to two aspects: 1) the powerful expressive capability of multi-resolution hash encoding [20] and 2) the fine-grained normal estimation method SDM-UniPS [10], as discussed in Sec. 4.2. Figure 4 visualizes the recovered shapes of different methods. Our method recovers fine-grained surface details on par with the "GT" meshes produced by a 3D scanner. + +Table 1 also reports the average runtime measured on our machine with one RTX 4090Ti graphics card. The runtime of R-MVPS [23] and B-MVPS [16] is difficult to measure since both methods require intensive manual effort. The code of UA-MVPS [12] is not publicly available, and several hours per object are required, as reported in MVPSNet [35]. Both PS-NeRF [31] and MVAS [4] require hours per object since they optimize a dense MLP. Our method is the fastest among neural implicit representation-based methods, with less than one minute per object. + +On the other hand, MVPSNet [35] is highly efficient because its pipeline mainly consists of feedforward inferencing a trained neural network and an sPSR [13] step, the latter of which has been optimized over one decade. Although our method is slower than MVPSNet [35], we achieve better reconstruction quality within a reasonable time. + +# 4.2. Ablation study + +Effect of normal maps Since multi-view normal integration is not bound to specific normal estimation methods, we investigate the effect of different normal estimation methods on normal integration methods. We test three sets of normal maps, estimated by SDPS [6] or SDM-UniPS [10], and the one rendered from the scanned mesh (i.e., "GT" normal maps provided by DiLiGenT-MV [16]). Table 2 reports the L2-Chamfer distance and F-score averaged on five DiLiGenT-MV [16] objects. Our method consistently + +achieves better results than PS-NeRF [31] and MVAS [4] when using the same set of normal maps. This verifies that our method can better fuse the information from 2.5D normal maps to 3D shapes. Using the SoTA method SDM-UniPS [10] for normal estimation, our method realizes the best reconstruction quality. + +Effect of spatial encoding Table 2 also compares the geometry accuracy between multi-resolution hash encoding (hash enc.) and positional encoding (pos. enc.). Our results verify again that multi-resolution hash encoding surpasses positional encoding in expressiveness, enhancing geometry accuracy. Therefore, the keys to superior MVPS performance are high-quality normal estimation, an expressive surface representation, and an effective normal fusion approach. + +Effect of SDF gradient computation Table 3 compares the geometry accuracy and runtime using finite difference (FD), Pytorch's automatic differentiation (AD), and directional finite difference (DFD) to compute the SDF gradient. All other parameter settings are kept the same. We do not observe a significant difference regarding reconstruction quality. However, DFD performs consistently faster than FD and AD by accelerating both the forward rendering and backward optimization. + +Generalizability of DFD In principle, DFD and patch-based sampling strategy apply to any SDF-based volume rendering like NeuS [29] or Neuralangelo [18]. To verify this, we incorporate DFD and patch-based sampling into Neuralangelo $[18]^{1}$ . We follow the coarse-to-fine sampling strategy of Neuralangelo [18] but use DFD instead of FD for SDF gradient approximation. We do not use mask supervision for Neuralangelo [18]. As shown in Figure 5, training time is reduced by $25\%$ on a scene from the DTU benchmark [11], even $2.25\times$ more pixels are looped in training. The result suggests that DFD could be a general strategy to accelerate NeuS-like multi-view reconstruction [29]. + +# 4.3. Qualitative results on our captured data + +To further demonstrate the performance of SuperNormal in fine-grained surface detail recovery, we also captured objects with more complex details, as shown in Fig. 6. Thanks to the generalizability of SDM-UniPS [10], we can capture the objects under casual light conditions (e.g., within an apartment) without ensuring a darkroom setup. We manually move a video light during the capture without any photometric or geometric light calibration. In total, we capture 18 viewpoints $\times 13$ light conditions for each object (1 under ambient light for SfM). We use Metashape [1] to calibrate camera parameters, use SAM [14] to create foreground masks, and apply SDM-UniPS [10] to every 12 image to infer the per-view normal map. + +Table 1. Quantitative evaluation on DiLiGenT-MV [16] benchmark. Red and orange cells indicate the best and the second best results respectively. On average, our approach achieves the best reconstruction quality at the second fastest speed. + +
MethodsL2 Chamfer distance [mm] (↓)F-score (τF=0.5 mm) (↑)Runtime (↓) Average
BearBuddhaCowPot2ReadingAverageBearBuddhaCowPot2ReadingAverage
R-MVPS [23]1.0700.3970.4401.5040.5610.7940.2620.6980.7600.1980.5190.487NA
B-MVPS [16]0.2120.2540.0910.2010.2590.2030.9580.9020.9860.9460.9140.941NA
MVPSSNet [35]0.3170.2790.2550.3100.2480.2820.8130.8660.8890.8380.9180.86537 secs
UA-MVPS [12]0.4140.4520.3260.4140.3820.3980.7070.6690.7980.7310.7620.733several hrs
PS-NeRF [31]0.2600.3140.2870.2540.3520.2930.8980.8060.8560.9190.7850.8538 hrs
MVAS [4]0.2430.3570.2160.1970.5220.3070.9090.7540.9070.9620.5460.81670 mins
Ours0.1840.2180.1930.1500.2230.1940.9830.9500.9770.9940.9160.96248 secs
+ +![](images/0cdb1c61c4408a135140062303efda8dce5b4344baf5c534ec20c2b9ffc6b89f.jpg) +Figure 4. Qualitative comparison of recovered shapes of Buddha and Pot2 and their error maps. Best viewed on screen. + +Table 2. Quantitative evaluation using normal maps estimated by different methods as inputs. The metrics are averaged over five DiLiGenT-MV [16] objects. Darker colors indicate better results. + +
L2 Chamfer distance [mm] (↓)F-score (τF=0.5 mm) (↑)
SDPS [6]SDM-UniPS [10]GTSDPS [6]SDM-UniPS [10]GT
MVAS [4]0.3070.3320.2920.8160.8120.859
PS-NeRF [31]0.2930.2960.2240.8530.8440.938
Ours (pos, enc.)0.3010.2280.1950.8490.9420.977
Ours (hash enc.)0.2810.1940.1140.8600.9620.998
+ +![](images/b5927e38fee03c5e9dc7d9779983a82a63e5ed685e2215e57463cb6c85e23ce4.jpg) +Figure 5. DFD and patch-based sampling accelerate Neuralangelo's training without quality degradation. We train both alternatives for $20\mathrm{k}$ batches. For FD, 512 random rays are sampled per batch; for DFD, 128 random patches of $3\times 3$ rays per batch. + +Table 3. Ablation study using different strategies for SDF gradients computation in volume rendering. We report the metrics averaged over five DiLiGenT-MV [16] objects. FD: Finite difference, AD: Automatic differentiation, DFD: Directional finite difference. + +
AccuracyRuntime (Sec)
CDF-ScoreForwardBackwardTotal
FD0.1920.96347.770.3124.6
AD0.1950.96129.442.478.1
DFD0.1940.96119.622.548.3
+ +![](images/38318f8bec8399f933f859ec2e8bdd664f5731d6dcae7705950fc61f9002425c.jpg) +Figure 6. Our real-world data capture setup. The target object is placed on a turntable before a dark sheet under general indoor illumination. There is no need to prepare a darkroom. We use an iPhone to take images and move a video light to illuminate the object from the camera side. Objects are $6\mathrm{cm}$ to $15\mathrm{cm}$ high. + +![](images/3cea14033fa30364ecc607094b1aeab094a1e9a4dcd60b16e48dfb07d9c7987d.jpg) + +Figures 1 and 7 display the results from NeuS2 [30], SuperNormal, and a structured-light based commercial 3D scanner EinScan SE. Although NeuS2 [30] and our method use the same multi-resolution hash encoding for fine-grained scene representation, the surface details recovered from multi-view color images are smoothed out. Our method, benefitting from the normal map inputs, recovers more fine-grained surface details. + +We find that our scanner performs poorly in highly concave regions (Fig. 7 above). This is because the scanner uses a wide baseline stereo camera (about $15\mathrm{cm}$ apart), and small concave regions create self-occlusions, thus reducing the reconstruction quality. In principle, normal estimation does not suffer from self-occlusion because it is based on single-view observations. Fusing multi-view normal maps can then produce better results in concave regions. + +# 5. Discussion + +We have demonstrated SuperNormal, a fast, high-fidelity 3D reconstruction approach using multi-view normal maps estimated by photometric stereo. SuperNormal outperforms other neural MVPS methods in speed and reconstruction quality on the benchmark dataset. Compared to multi-view neural reconstruction methods, using normal maps can exploit the expressive power of multi-resolution hash encoding, and produce scanner-level surface details. Further, we have shown the effectiveness of directional finite differences in expediting training without sacrificing quality. + +SuperNormal provides a promising direction to make + +![](images/c13d95a870b71c1cdb722c7c9ecd34ba454525d9334c13f2f53c0d2f5e68843a.jpg) +Figure 7. Qualitative comparison using our objects. Our approach recovers better surface detail than multi-view stereo approaches and achieves competitive results with commercial 3D scanners. + +![](images/0cfc59baf60775ffccca5f973d02bf866d9e9f00c5a44ab97ee07affe14405ce.jpg) +Ours +Figure 8. Limitations. (Left) Fault artifacts are highlighted in red circles. (Right) Our method is sensitive to the inaccuracy in normal maps estimated by monocular methods (e.g., OmniData [7]). + +![](images/e13b646eb49d1ada047b5971a0c1d4b99296474726c8ff9fe7100f76ef325ffd.jpg) +OmniData normal + +![](images/f03c38d05d82dad187b9984e45f66d220d7e02ef2cf676b8605da37d2b39283c.jpg) +Ours + +high-quality 3D scanning more accessible. Unlike costly commercial 3D scanners, our approach is viable with everyday devices like smartphones and basic equipment such as a tripod, video light, and a turntable. + +Limitations Our method may produce fault-like artifacts, as shown in Fig. 8. The artifacts exist no matter our SDF gradient computation approach (i.e., AD, FD, or DFD) and require further investigation. Moreover, the performance depends on mask supervision and high-quality normal maps (e.g., those obtained from photometric stereo). Our method does not work well with the monocular normal estimation methods like OmniData [7] (Fig. 8). + +# References + +[1] Agisoft metashape. https://www.agisoft.com.6 +[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 2 +[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-NeRF: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 2 +[4] Xu Cao, Hiroaki Santo, Fumio Okura, and Yasuyuki Matsushita. Multi-view azimuth stereo via tangent space consistency. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 825-834, 2023. 1, 2, 6, 7, 11, 12 +[5] Ju Yong Chang, Kyoung Mu Lee, and Sang Uk Lee. Multiview normal field integration using level set methods. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 1-8. IEEE, 2007. 1, 2 +[6] Guanying Chen, Kai Han, Boxin Shi, Yasuyuki Matsushita, and Kwan-Yee K. Wong. SDPS-Net: Self-calibrating deep photometric stereo networks. In CVPR, 2019. 6, 7 +[7] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In ICCV, pages 10786-10796, 2021. 8 +[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of Machine Learning and Systems 2020, pages 3569-3579. 2020. 3, 11 +[9] Carlos Hernandez, George Vogiatzis, and Roberto Cipolla. Multiview photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 30(3):548-554, 2008. 1, 2 +[10] Satoshi Ikehata. Scalable, detailed and mask-free universal photometric stereo. In Proc. of Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5, 6, 7 +[11] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, pages 406-413. IEEE, 2014. 6 +[12] Berk Kaya, Suryansh Kumar, Carlos Oliveira, Vittorio Ferrari, and Luc Van Gool. Uncertainty-aware deep multi-view photometric stereo. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 12601-12611, 2022. 2, 6, 7 +[13] Michael Kazhdan and Hugues Hoppe. Screened Poisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 2, 6 +[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proc. of International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 6 +[15] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 6 +[16] Min Li, Zhenglong Zhou, Zhe Wu, Boxin Shi, Changyu Diao, and Ping Tan. Multi-view photometric stereo: A ro + +bust solution and benchmark dataset for spatially varying isotropic materials. IEEE Transactions on Image Processing, 29:4159-4173, 2020. 2, 5, 6, 7, 8, 11, 12 +[17] Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. NerfAcc: Efficient sampling accelerates nerfs. In Proc. of International Conference on Computer Vision (ICCV), pages 18537-18546, 2023. 5, 13 +[18] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 8456-8465, 2023. 2, 3, 4, 6 +[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. of European Conference on Computer Vision (ECCV), 2020. 2 +[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 3, 5, 6 +[21] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proc. of International Conference on Computer Vision (ICCV), 2021. 2 +[22] Stanley Osher, Ronald Fedkiw, and K Piechor. Level set methods and dynamic implicit surfaces. Appl. Mech. Rev., 57(3):B15-B15, 2004. 2, 4 +[23] Jaesik Park, Sudipta N Sinha, Yasuyuki Matsushita, Yu-Wing Tai, and In So Kweon. Robust multiview photometric stereo using planar mesh parameterization. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 39(8):1591-1604, 2016. 2, 6, 7, 11, 12 +[24] Radu Alexandru Rosu and Sven Behnke. Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices. In CVPR, 2023. 3, 4 +[25] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. of European Conference on Computer Vision (ECCV), 2016. 1 +[26] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my NeRF: A dataset for novel view synthesis and relighting of real world objects. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 20762-20772, 2023. 2 +[27] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 2 +[28] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. In ECCV, 2022. 3 +[29] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view recon + +struction. Advances in Neural Information Processing Systems (NeurIPS), 2021. 1, 3, 6 +[30] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. NeuS2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proc. of International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 4, 8, 14, 15 +[31] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K. Wong. PS-NeRF: Neural inverse rendering for multi-view photometric stereo. In Proc. of European Conference on Computer Vision (ECCV), 2022. 1, 2, 6, 7, 11, 12, 14 +[32] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 2 +[33] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems (NeurIPS), 34:4805-4815, 2021. 3 +[34] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. NIPS, 2022. 3 +[35] Dongxu Zhao, Daniel Lichy, Pierre-Nicolas Perrin, Jan-Michael Frahm, and Soumyadip Sengupta. MVPSNet: Fast generalizable multi-view photometric stereo. In Proc. of International Conference on Computer Vision (ICCV), pages 12525-12536, 2023. 2, 6, 7 +[36] Fuqiang Zhao, Yuheng Jiang, Kaixin Yao, Jiakai Zhang, Liao Wang, Haizhao Dai, Yuhui Zhong, Yingliang Zhang, Minye Wu, Lan Xu, and Jingyi Yu. Human performance modeling and rendering via neural animated mesh. ACM Trans. Graph., 41(6), 2, 3, 4 \ No newline at end of file diff --git a/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/images.zip b/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cbb6decdbfc04c8672954a7d8335eacadd767bc1 --- /dev/null +++ b/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:771ab3498b2eac319ddec3e90071b0b892ad3bf4c558d1209626a93600dbf613 +size 752691 diff --git a/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/layout.json b/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..db3b882a300ac10b8db43bd231456144b86db663 --- /dev/null +++ b/2024/SuperNormal_ Neural Surface Reconstruction via Multi-View Normal Integration/layout.json @@ -0,0 +1,7577 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 154, + 103, + 440, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 103, + 440, + 138 + ], + "spans": [ + { + "bbox": [ + 154, + 103, + 440, + 138 + ], + "type": "text", + "content": "SuperNormal: Neural Surface Reconstruction via Multi-View Normal Integration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 222, + 161, + 370, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 161, + 370, + 174 + ], + "spans": [ + { + "bbox": [ + 222, + 161, + 370, + 174 + ], + "type": "text", + "content": "Xu Cao Takafumi Taketomi" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 266, + 176, + 328, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 176, + 328, + 190 + ], + "spans": [ + { + "bbox": [ + 266, + 176, + 328, + 190 + ], + "type": "text", + "content": "CyberAgent" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 180, + 191, + 413, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 191, + 413, + 203 + ], + "spans": [ + { + "bbox": [ + 180, + 191, + 413, + 203 + ], + "type": "text", + "content": "{xu_cao, taketomi_takafumi}@cyberagent.co.jp" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 257, + 290, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 290, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 290, + 473 + ], + "type": "text", + "content": "We present SuperNormal, a fast, high-fidelity approach to multi-view 3D reconstruction using surface normal maps. With a few minutes, SuperNormal produces detailed surfaces on par with 3D scanners. We harness volume rendering to optimize a neural signed distance function (SDF) powered by multi-resolution hash encoding. To accelerate training, we propose directional finite difference and patch-based ray marching to approximate the SDF gradients numerically. While not compromising reconstruction quality, this strategy is nearly twice as efficient as analytical gradients and about three times faster than axis-aligned finite difference. Experiments on the benchmark dataset demonstrate the superiority of SuperNormal in efficiency and accuracy compared to existing multi-view photometric stereo methods. On our captured objects, SuperNormal produces more fine-grained geometry than recent neural 3D reconstruction methods. Our code is available at https://github.com/CyberAgentAILab/SuperNormal." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 521, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 676 + ], + "type": "text", + "content": "Recovering high-quality 3D geometry of real-world objects from their multi-view images is a long-standing challenge in computer vision. Recently, neural implicit surface-based methods have shown remarkable reconstruction results. Compared to traditional multi-view stereo (MVS) methods [25], neural methods are more robust to view-dependent observations and textureless surfaces [29]. Furthermore, the reconstruction procedure has become highly efficient [30] by introducing multi-resolution hash coding [20]. However, even though this spatial encoding allows fine-grained geometry to be represented, shape reconstruction from multi-view images tends to smooth out high-frequency surface details, as shown in Fig. 1 bottom right." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "Multi-view photometric stereo (MVPS), on the other hand, aims to recover pixel-level high-frequency surface detail by introducing additional lighting conditions during" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 230, + 545, + 331 + ], + "blocks": [ + { + "bbox": [ + 306, + 230, + 545, + 331 + ], + "lines": [ + { + "bbox": [ + 306, + 230, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 306, + 230, + 545, + 331 + ], + "type": "image", + "image_path": "fd4a7330b81a55a9ab69dd7d0dd6a582b2dafb4ed9a6f47cfb78157305ba94d2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 336, + 545, + 426 + ], + "blocks": [ + { + "bbox": [ + 310, + 336, + 545, + 426 + ], + "lines": [ + { + "bbox": [ + 310, + 336, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 310, + 336, + 545, + 426 + ], + "type": "image", + "image_path": "c479ca7ec93608d18f5b3a6564724cec7e349ef64abcbf2f3db03bd6dae3ed92.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 435, + 547, + 513 + ], + "lines": [ + { + "bbox": [ + 304, + 435, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 547, + 513 + ], + "type": "text", + "content": "Figure 1. (Top) From multi-view surface normal maps, SuperNormal recovers fine-grained surface details comparable to 3D scanners while being orders of magnitude faster than the existing MVPS method PS-NeRF [31]. (Bottom) Using normal maps produces more faithful high-frequency details than the MVS-based method NeuS2 [30], although both use multi-resolution hash encoding [20]." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "type": "text", + "content": "the image acquisition [9]. In a typical workflow, a surface normal map is first recovered for each view, recording per-pixel surface orientation information. The normal maps estimated at different viewpoints are then fused into a 3D model, also known as multi-view normal integration [5]. However, existing normal fusion methods struggle to reflect the details of the normal maps in the shape of the recovered 3D model. Moreover, the reconstruction process takes hours even when only a few low-resolution normal maps are used [4, 31]. Due to the lack of a fast and accurate multiview normal fusion method, current multi-view photometric stereo results remain unsatisfactory." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "This paper presents SuperNormal to unite the best of both worlds. Normal maps with pixel-level surface details are utilized to exploit the expressive power and efficiency" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20581" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "type": "text", + "content": "of multi-resolution hash encoding-powered neural implicit surface. Using normal maps alone avoids per-scene reflectance modeling and optimization, thus improving training efficiency compared to using color images. In a few minutes, our method can produce fine-grained geometry on par with commercial 3D scanners." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 288, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 323 + ], + "type": "text", + "content": "Specifically, we use volume rendering to optimize a neural signed distance function (SDF) such that its rendered normal maps are consistent with input normal maps. To further improve the training efficiency, we propose a directional finite difference and patch-based ray marching strategy to approximate the SDF gradients numerically. This strategy kills two birds with one stone: SDF values evaluated at sampled points are all used both for SDF-based volume rendering and SDF gradient approximation. This way, we avoid second-order derivative computation during training. Unlike traditional axis-aligned finite difference, directional finite difference avoids redundant SDF evaluations (Fig. 2). Without compromising reconstruction quality, our method is nearly twice as fast as using analytical gradients and about three times faster than finite difference." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 323, + 287, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 287, + 466 + ], + "type": "text", + "content": "Compared to existing neural MVPS methods, our method is two orders of magnitude faster and recovers better surface details (Fig. 1 above). We also capture objects with complex geometry and show that our method recovers more faithful surface details than the recent neural 3D reconstruction method [30] (Fig. 1 bottom). Taking advantage of recent advances in photometric stereo, we can estimate high-quality normal maps from images captured under common lighting conditions with inexpensive equipment [10]. Combined with SuperNormal, high-fidelity 3D objects can be reconstructed using inexpensive equipment in widely accessible environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 467, + 249, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 467, + 249, + 479 + ], + "spans": [ + { + "bbox": [ + 59, + 467, + 249, + 479 + ], + "type": "text", + "content": "In summary, this paper's key contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 479, + 287, + 594 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 479, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 287, + 514 + ], + "type": "text", + "content": "- An effective neural 3D reconstruction approach using multi-view normal maps, enhancing the reconstruction quality to near 3D scanner levels;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 519, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 519, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 47, + 519, + 287, + 554 + ], + "type": "text", + "content": "- An efficient method for computing numerical gradients in SDF-based neural rendering, speeding up the reconstruction process; and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 559, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 287, + 594 + ], + "type": "text", + "content": "- A comprehensive evaluation using benchmark objects with MVPS approaches and an assessment using our captured objects with MVS approaches." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 604, + 131, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 131, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 131, + 616 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 623, + 252, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 252, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 252, + 635 + ], + "type": "text", + "content": "2.1. Multi-view photometric stereo (MVPS)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "content": "MVPS aims at high-fidelity 3D reconstruction using multi-view observations under varying light conditions [9] (usually one-light-at-time (OLAT) images [16, 26])." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Existing methods differ in what geometry-related information is inferred in each view and how information from different views is fused into a complete 3D model. Chang et" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 545, + 194 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 545, + 194 + ], + "type": "image", + "image_path": "cfac92fe142e4ba7f830a110d9ce6d48012fe1d807e03238eef7b829e7827220.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 203, + 545, + 280 + ], + "lines": [ + { + "bbox": [ + 304, + 203, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 203, + 545, + 280 + ], + "type": "text", + "content": "Figure 2. (Left) In SDF-based volume rendering, approximating SDF gradients using axis-aligned finite difference [18, 36] requires additional SDF samples at neighboring positions of volume rendering points, increasing computational complexity and training time. (Right) Our strategy of patch-based ray marching and directional finite difference effectively uses all SDF samples for both SDF gradient approximation and volume rendering." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 301, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 386 + ], + "type": "text", + "content": "al. [5] proposes a level set method [22] to fuse multi-view normal maps into a volumetric representation of the shape. R-MVPS [23] uses the multi-view normal maps to refine a coarse mesh estimated by multi-view stereo. B-MVPS [16] initialize the shape as a sparse point cloud estimated by structure from motion (SfM), then propagates the known point spatially using the normal maps." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 386, + 546, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 546, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 546, + 566 + ], + "type": "text", + "content": "Recent MVPS methods have undergone a paradigm shift since the advent of NeRF [19] by using a neural implicit scene representation, usually parameterized as a multi-layer perception (MLP). PS-NeRF [31] recovers a neural field based on UNISURF [21] by regularizing the gradients of the neural field using the normal maps. UA-MVPS [12] fuses the multi-view depth and normal maps into a neural SDF using neural inverse rendering while considering per-pixel confidence. MVAS [4] shows utilizing the azimuth components of multi-view normal maps is sufficient to recover the neural implicit surface. While most methods only use images in the same view for normal map estimation, MVPSNet [35] utilizes neighboring views to infer per-view depth and normal maps jointly and applies screened Poisson surface reconstruction (sPSR) [13] to obtain the shape." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 566, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 545, + 638 + ], + "type": "text", + "content": "However, the reconstructed surfaces of these methods tend to be smoothed and do not reflect the high-frequency details of the input normal maps. Our approach introduces multi-resolution hash encoding and directional finite difference to improve the implicit neural field's expressive power and training efficiency." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 647, + 448, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 448, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 448, + 658 + ], + "type": "text", + "content": "2.2. Neural 3D reconstruction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "The seminal work NeRF [19] and follow-up works [2, 3, 27] have demonstrated remarkable novel view synthesis quality by volume rendering a neural implicit field. For better geometry reconstruction quality, IDR [32] imposes eikonal" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20582" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "regularization [8] such that the neural field approximates a valid SDF. Later, the concurrent works VolSDF [33] and NeuS [29] develop SDF-based volume rendering, which bridges SDF values to volume rendering weights and improves the geometry fidelity. However, these methods require several hours to recover one object due to optimizing a deep coordinate-based MLP." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 288, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 288, + 277 + ], + "type": "text", + "content": "Multi-resolution hash encoding has been proposed to improve expressive power and accelerate the learning of neural implicit field [20]. Several efforts have been made to incorporate this expressive spatial encoding into neural SDF learning. Since the official CUDA implementation does not support the double propagation deduced by the eikonal regularization, PermutoSDF [24] and NeuS2 [30] re-implement the CUDA programming. InstantNSR [36] and Neuralangelo [18] use finite difference instead of automatic differentiation to avoid double propagation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 278, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 278, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 287, + 327 + ], + "type": "text", + "content": "We propose directional finite difference, a numerical way to compute SDF gradients that avoids double propagation and is even faster than automatic differentiation when combined with a tailored patch-based sampling strategy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 338, + 253, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 253, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 253, + 350 + ], + "type": "text", + "content": "2.3. Neural 3D reconstruction with normals" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 357, + 287, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 429 + ], + "type": "text", + "content": "Normal maps have been used as auxiliary information to facilitate neural scene reconstruction in MonoSDF [34] and NeuRIS [28]. These works target scene-level reconstruction, and normal maps help infer the geometry of textureless smooth regions (e.g., walls) where using color images is likely to break down." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 430, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 491 + ], + "type": "text", + "content": "Instead, our method is particularly useful for single objects with complex surface details. We use photometric stereo [10] to estimate normal maps from images under varying light conditions. This produces high-quality normal maps that can be used alone for shape reconstruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 505, + 113, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 505, + 113, + 519 + ], + "spans": [ + { + "bbox": [ + 47, + 505, + 113, + 519 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 526, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 526, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 287, + 635 + ], + "type": "text", + "content": "We aim to recover the surface given a set of normal maps, object masks, and camera intrinsic and extrinsic parameters. Since photometric stereo methods estimate the normal maps in camera space, we first apply the camera-to-world rotation to normal maps to obtain world-space normal maps. Figure 3 depicts our pipeline. We represent the geometry as a parametric SDF and optimize SDF parameters such that its rendered gradient and opacity values are consistent with input normal vectors and mask values, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 646, + 107, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 107, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 107, + 658 + ], + "type": "text", + "content": "3.1. Pipeline" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "SDF-based volume rendering (Forward pass) A signed distance function (SDF) implicitly captures the geometry by assigning each spatial point with the signed distance to its closest surface point. The surface " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " can then be repre" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 465, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 465, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 465, + 83 + ], + "type": "text", + "content": "sented as the zero-level-set of the SDF:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 378, + 97, + 545, + 110 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 97, + 545, + 110 + ], + "spans": [ + { + "bbox": [ + 378, + 97, + 545, + 110 + ], + "type": "interline_equation", + "content": "\\mathcal {M} = \\{\\mathbf {x} \\mid f (\\mathbf {x}) = 0 \\}. \\tag {1}", + "image_path": "f019273f23425f0db67dcf046427f96be2b9f67fa4c584fed844ac55a69c09be.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "text", + "content": "We parameterize the SDF using multi-resolution hash encoding followed by a shallow MLP. In multi-resolution hash encoding, the 3D space is discretized into a set of multi-resolution 3D grids. Each grid cell is associated with a unique hash value mapping the 3D coordinates to a learnable feature vector. Specifically, " + }, + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "inline_equation", + "content": "h(\\mathbf{x};\\phi)" + }, + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "text", + "content": " is the feature vector obtained by concatenating feature vectors of different levels " + }, + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "inline_equation", + "content": "h(\\mathbf{x}) = [\\mathbf{x},h_1(\\mathbf{x}),\\dots,h_L(\\mathbf{x})]" + }, + { + "bbox": [ + 304, + 122, + 545, + 229 + ], + "type": "text", + "content": ". Then, the SDF can be parameterized as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 379, + 243, + 545, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 243, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 379, + 243, + 545, + 255 + ], + "type": "interline_equation", + "content": "f (\\mathbf {x}) = f (h (\\mathbf {x}; \\phi); \\theta), \\tag {2}", + "image_path": "d38861bedf7db355e4c28a6f2695d8c6143d492efbf8bc61f028bdc7dca564c3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 268, + 545, + 316 + ], + "type": "text", + "content": " are learnable hash-encoding feature vectors and MLP parameters. Multi-resolution hash encoding has proven to speed up training and improve the expressive power of the neural SDF [20, 30, 36]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 317, + 545, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 401 + ], + "type": "text", + "content": "During training, we evaluate the SDF values at points sampled on the rays cast from the camera centers toward the scene. NeuS [29] proposes an unbiased and occlusion-aware way to transfer the SDF samples to volume rendering opacities. Given " + }, + { + "bbox": [ + 304, + 317, + 545, + 401 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 317, + 545, + 401 + ], + "type": "text", + "content": " ordered 3D points " + }, + { + "bbox": [ + 304, + 317, + 545, + 401 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i\\}_{i=0}^N" + }, + { + "bbox": [ + 304, + 317, + 545, + 401 + ], + "type": "text", + "content": " on a ray and their SDF values " + }, + { + "bbox": [ + 304, + 317, + 545, + 401 + ], + "type": "inline_equation", + "content": "\\{f(\\mathbf{x}_i)\\}_{i=0}^N" + }, + { + "bbox": [ + 304, + 317, + 545, + 401 + ], + "type": "text", + "content": ", the opacity at each point is designed as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 334, + 412, + 545, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 412, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 334, + 412, + 545, + 440 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = \\max \\left(\\frac {\\phi_ {s} (f (\\mathbf {x} _ {i})) - \\phi_ {s} (f (\\mathbf {x} _ {i + 1}))}{\\phi_ {s} (f (\\mathbf {x} _ {i}))}, 0\\right), \\tag {3}", + "image_path": "96130bd8662d6bb673ba81c3672fad5af1503195b3539beeab299430ddd8d208.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 451, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 545, + 476 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 451, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\phi_s(x) = \\frac{1}{1 + \\exp(-sx)}" + }, + { + "bbox": [ + 304, + 451, + 545, + 476 + ], + "type": "text", + "content": " is the sigmoid function with a trainable sharpness " + }, + { + "bbox": [ + 304, + 451, + 545, + 476 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 451, + 545, + 476 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 479, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 503 + ], + "type": "text", + "content": "With this SDF-based opacity, we can render the normal and opacity for the pixel " + }, + { + "bbox": [ + 304, + 479, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 479, + 545, + 503 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 332, + 514, + 545, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 514, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 332, + 514, + 545, + 554 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {n}} (\\mathbf {p}) = \\sum_ {i = 0} ^ {N} T _ {i} \\alpha_ {i} \\nabla f (\\mathbf {x} _ {i}), \\quad \\hat {o} (\\mathbf {p}) = \\sum_ {i = 0} ^ {N} T _ {i} \\alpha_ {i} \\tag {4}", + "image_path": "66a84be1967cf82e7a3acaa266a1d5bc797a47b019dc570e51082f2a0ccb88f6.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 413, + 550, + 517, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 550, + 517, + 582 + ], + "spans": [ + { + "bbox": [ + 413, + 550, + 517, + 582 + ], + "type": "interline_equation", + "content": "\\text {w i t h} \\quad T _ {i} = \\prod_ {j = 0} ^ {i - 1} (1 - \\alpha_ {j}).", + "image_path": "0b799c8cf490c43ef18d3f8d8026312763299dc881bc08273631a1247e90c63d.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 594, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 654 + ], + "type": "text", + "content": "From Eqs. (3) and (4), the neural SDF must be evaluated at least once at sampled points for volume rendering. In Secs. 3.2 and 3.3, we show that these SDF samples can be reused for SDF gradient approximation, thus reducing the computational amount and speeding up the training." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "Neural SDF Training (Backward pass) To train the neural SDF, we supervise the rendered normals and opacities using input ones " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{n}(\\mathbf{p})" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{o} (\\mathbf{p})" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": ". The loss function consists" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20583" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 545, + 219 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 545, + 219 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 545, + 219 + ], + "type": "image", + "image_path": "8a20c984d4c8cc382b0dd432ebb04c4d3a096b69675c56f6399dfe04c36dc528.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 547, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 547, + 264 + ], + "type": "text", + "content": "Figure 3. Approach overview. (a) Given input views, we randomly sample patches of pixels, cast rays from those pixels, and march a plane along the patch of rays (Sec. 3.3). We treat the ray-plane intersections as sampled points and evaluate their SDF values via the neural SDF. (b) SDF samples neighbor to a point are used to compute the SDF gradients using directional finite difference (Sec. 3.2). (c) SDF samples on the same rays are used for SDF-based volume rendering (Sec. 3.1)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 285, + 287, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 287, + 296 + ], + "type": "text", + "content": "of three terms: The normal, the mask, and the eikonal term:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 315, + 287, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 315, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 315, + 287, + 396 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\overbrace {\\sum_ {\\mathbf {p}} \\left(\\hat {\\mathbf {n}} (\\mathbf {p}) - \\mathbf {n} (\\mathbf {p})\\right) ^ {2}} ^ {\\text {n o r m a l}} + \\lambda_ {1} \\overbrace {\\sum_ {\\mathbf {p}} \\operatorname {B C E} \\left(\\hat {o} (\\mathbf {p}) , o (\\mathbf {p})\\right)} ^ {\\text {m a s k}} \\tag {5} \\\\ + \\lambda_ {2} \\underbrace {\\sum_ {\\mathbf {x}} \\left(\\| \\nabla f (\\mathbf {x}) \\| _ {2} - 1\\right) ^ {2}} _ {e i k o n a l}, \\\\ \\end{array}", + "image_path": "7af39221a028447910c31906386ecaeb085c40361e5a9b1aef4477067dbf8b90.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 415, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 522 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 415, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 415, + 287, + 522 + ], + "type": "text", + "content": " is the sampled pixels in a batch, " + }, + { + "bbox": [ + 46, + 415, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 415, + 287, + 522 + ], + "type": "text", + "content": " is the sampled points on rays, and BCE is the binary cross entropy function. The normal term encourages the rendered SDF gradients to be consistent with the input normal vectors so that accurate geometry details are recovered. The mask term aligns the shape's silhouette with input masks and can be considered the boundary condition. The eikonal term enforces the SDF gradient norm to be close to 1 almost everywhere so that the neural SDF is approximately valid [22]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 546, + 204, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 204, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 204, + 558 + ], + "type": "text", + "content": "3.2. Directional Finite Difference" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "As seen, both the normal and eikonal terms backpropagate the loss gradient via the SDF gradient to neural SDF parameters (i.e., double backpropagation " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathcal{L}}{\\partial\\nabla f}\\frac{\\partial\\nabla f}{\\partial\\theta}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "). The most common way is to use Pytorch's automatic differentiation (AD) due to implementation simplicity. However, the multi-resolution hash encoding package does not officially support double propagation. Common workarounds include CUDA programming [24, 30] or use finite difference (FD) [18, 36] to bypass double propagation. However, CUDA programming can be inflexible in that it shrinks the design space (e.g., NeuS2 [30] requires ReLU activation), and FD will slow down the training. Consider the ordinary" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 285, + 375, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 285, + 375, + 296 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 375, + 296 + ], + "type": "text", + "content": "axis-aligned FD:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 359, + 301, + 545, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 301, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 359, + 301, + 545, + 346 + ], + "type": "interline_equation", + "content": "\\nabla f (\\mathbf {x}) \\approx \\left[ \\begin{array}{l} \\frac {f (\\mathbf {x} + \\epsilon_ {x}) - f (\\mathbf {x} - \\epsilon_ {x})}{2 \\epsilon} \\\\ \\frac {f (\\mathbf {x} + \\epsilon_ {y}) - f (\\mathbf {x} - \\epsilon_ {y})}{2 \\epsilon} \\\\ \\frac {f (\\mathbf {x} + \\epsilon_ {z}) - f (\\mathbf {x} - \\epsilon_ {z})}{2 \\epsilon} \\end{array} \\right], \\tag {6}", + "image_path": "470cf0078858564ac500f0e544da6801329e1bb95f8dd7d8e0933ce8a095b387.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\epsilon_{* \\in \\{x, y, z\\}}" + }, + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "type": "text", + "content": " is an offset vector along " + }, + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "type": "inline_equation", + "content": "x, y" + }, + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "type": "text", + "content": ", or " + }, + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "type": "text", + "content": " axis with a tiny step size " + }, + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 351, + 545, + 434 + ], + "type": "text", + "content": ". Computing the gradient at one point requires evaluating the SDF at its nearby points. These additional SDF samples are also included in the computation graph in the forward pass. Consequently, using FD significantly slows the forward rendering and backward optimization even if the efficient spatial encoding is used [18]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": "To avoid redundant SDF evaluations and double backpropagation, we propose directional finite difference (DFD) to reuse the volume rendering SDF samples for SDF gradients approximation. Consider the SDF value at a 3D point " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": " parameterized by the ray origin " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": " and direction " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 385, + 501, + 545, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 501, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 385, + 501, + 545, + 514 + ], + "type": "interline_equation", + "content": "f (\\mathbf {x}) = f (\\mathbf {o} + t \\mathbf {v}), \\tag {7}", + "image_path": "761918bbf7017454da22db7d818de3f1162dcee1b6b5d7d7e1f335dfa21aee0d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 520, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 541 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 520, + 545, + 541 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 520, + 545, + 541 + ], + "type": "text", + "content": " is the distance to the ray origin. From Eq. (7), we have" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 372, + 540, + 545, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 540, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 372, + 540, + 545, + 563 + ], + "type": "interline_equation", + "content": "\\frac {d f}{d t} = \\nabla f (\\mathbf {x}) ^ {\\top} \\mathbf {v} := \\nabla_ {\\mathbf {v}} f. \\tag {8}", + "image_path": "c83657241624a20c51022057f32874d44a0f24fa85097f489a1b653fe2c2a667.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 566, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 545, + 612 + ], + "type": "text", + "content": "Equation (8) is exactly the directional derivative of the SDF, which projects the SDF gradient onto the ray direction. Suppose we parameterize the same point along three different directions as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 620, + 545, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 620, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 361, + 620, + 545, + 633 + ], + "type": "interline_equation", + "content": "\\mathbf {x} = \\mathbf {o} _ {i} + t _ {i} \\mathbf {v} _ {i}, \\quad i \\in \\{1, 2, 3 \\}. \\tag {9}", + "image_path": "4c1affa380fdd696bc752d877f1f9da4e2f7040e420caae243a57c29498f5c51.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 639, + 514, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 514, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 514, + 651 + ], + "type": "text", + "content": "Applying Eq. (8) then yields three linear equations:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 372, + 656, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 656, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 372, + 656, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\underbrace {\\left[ \\begin{array}{l} - \\mathbf {v} _ {1} ^ {\\top} - \\\\ - \\mathbf {v} _ {2} ^ {\\top} - \\\\ - \\mathbf {v} _ {3} ^ {\\top} - \\end{array} \\right]} _ {\\mathbf {V}} \\nabla f = \\underbrace {\\left[ \\begin{array}{l} \\frac {d f}{d t _ {1}} \\\\ \\frac {d f}{d t _ {2}} \\\\ \\frac {d f}{d t _ {3}} \\end{array} \\right]} _ {\\nabla_ {\\mathbf {V}} f}. \\tag {10}", + "image_path": "74be9064b182c9b966d0d2be163010ccbadba6a5be5b0bca7417ccc4e1726e13.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20584" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 261, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 261, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 261, + 84 + ], + "type": "text", + "content": "From Eq. (10), we can compute the SDF gradient by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 129, + 91, + 287, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 91, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 129, + 91, + 287, + 105 + ], + "type": "interline_equation", + "content": "\\nabla f = \\mathbf {V} ^ {- 1} \\nabla_ {\\mathbf {V}} f. \\tag {11}", + "image_path": "41d1e147616dec694445e8c4897715e27c58ede1b7f3d96107c8cef54a8b1a84.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 114, + 287, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 114, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 47, + 114, + 287, + 138 + ], + "type": "text", + "content": "Numerically, we can approximate the directional derivative by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 144, + 287, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 144, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 287, + 206 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\frac {d f}{d t _ {i}} \\approx \\frac {f (\\mathbf {o} _ {i} + (t _ {i} + \\Delta t) \\mathbf {v} _ {i}) - f (\\mathbf {o} _ {i} + (t _ {i} - \\Delta t) \\mathbf {v} _ {i})}{2 \\Delta t} \\\\ = \\frac {f \\left(\\mathbf {x} _ {i} + \\Delta t \\mathbf {v} _ {i}\\right) - f \\left(\\mathbf {x} _ {i} - \\Delta t \\mathbf {v} _ {i}\\right)}{2 \\Delta t}, \\quad i \\in \\{1, 2, 3 \\} \\tag {12} \\\\ \\end{array}", + "image_path": "4fdf86e503e37a253b1483ab1eef02ad651ae32a78e9d8f1b41a998354060807.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 214, + 287, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 214, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 287, + 298 + ], + "type": "text", + "content": "Equations (11) and (12) imply that, to numerically approximate the SDF gradient, we can first compute finite differences along three directions " + }, + { + "bbox": [ + 46, + 214, + 287, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 214, + 287, + 298 + ], + "type": "text", + "content": ", then linearly transform the concatenated finite difference by " + }, + { + "bbox": [ + 46, + 214, + 287, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{-1}" + }, + { + "bbox": [ + 46, + 214, + 287, + 298 + ], + "type": "text", + "content": ". We call this SDF gradient approximation directional finite difference. Equation (11) has a unique solution if the three directions are non-coplanar." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 299, + 287, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 299, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 287, + 395 + ], + "type": "text", + "content": "The ordinary finite difference can be viewed as the special case of Eq. (11) by choosing the three coordinate axes as the ray directions. In this case, " + }, + { + "bbox": [ + 46, + 299, + 287, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 299, + 287, + 395 + ], + "type": "text", + "content": " becomes the identity matrix. Both DFD and FD require nearby SDF samples to approximate the SDF gradient. However, by tailoring the sampling strategy, we can reuse the SDF samples in volume rendering for DFD without additional SDF samples, as described in the following." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 401, + 198, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 401, + 198, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 401, + 198, + 414 + ], + "type": "text", + "content": "3.3. Patch-based Ray Marching" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 419, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 502 + ], + "type": "text", + "content": "DFD requires computing the finite difference in three directions. In volume rendering, we naturally have SDF samples along the viewing ray direction that can be used for finite difference. For the other two directions, we propose to sample patches of pixels instead of single pixels and march a plane on each patch of rays to locate the sampled points on parallel planes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 503, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 287, + 574 + ], + "type": "text", + "content": "Specifically, we perform ordinary ray marching for the center pixel/ray of the patch and treat the sampled points as the intersections between a marching plane and the center ray. The points to be sampled on the remaining rays within the patch can then be found as the intersection of the planes and the rays:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 574, + 287, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 574, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 137, + 574, + 287, + 602 + ], + "type": "interline_equation", + "content": "t _ {j} = \\frac {t _ {i} \\mathbf {v} _ {i} ^ {\\top} \\mathbf {m}}{\\mathbf {v} _ {j} ^ {\\top} \\mathbf {m}}, \\tag {13}", + "image_path": "c7446eb128fab13391efb8d8d062b8ed4c808d33a3a50f2d929a5427a5172811.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "text", + "content": " is the unit normal direction of the marching plane, " + }, + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "text", + "content": " is the ray marching distance on the center ray with direction " + }, + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "text", + "content": " is the distance from the intersection point to the ray origin (i.e., camera center) along the direction " + }, + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_j" + }, + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "text", + "content": "; for derivation see Sec. B.2 in SupMat." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "We choose the marching plane to parallel the image plane where the patch is located (Fig. 3). With this sampling strategy, the three directions for DFD computation are the viewing ray direction and the world-space directions of " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "x" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "-axis of the camera coordinates. The marching plane normal " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " becomes the world-space direction of the " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "-axis of the camera coordinates and is the same for all pixels/rays from the same view. This way, the SDF samples on the same ray can be used for volume rendering as Eq. (4), and the SDF samples neighbor to a point can be used for SDF gradient computation as Eq. (11). We use the central difference (i.e., Eq. (12)) for points with neighbors in both directions and forward/backward difference for boundary points." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 180, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 335 + ], + "type": "text", + "content": "This sampling strategy is beneficial in three aspects. First, it allows us to reuse the SDF samples from neighboring rays to compute DFD. Second, it stabilizes training by ensuring the existence of " + }, + { + "bbox": [ + 304, + 180, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{-1}" + }, + { + "bbox": [ + 304, + 180, + 545, + 335 + ], + "type": "text", + "content": ", since the viewing ray direction will never co-planer to the image plane. Third, it reduces the computation overhead during training because the three ray directions are identical for samples from the same pixel/ray. The inverse of ray directions " + }, + { + "bbox": [ + 304, + 180, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{-1}" + }, + { + "bbox": [ + 304, + 180, + 545, + 335 + ], + "type": "text", + "content": " can be pre-computed for each pixel before training. In our experiments, the computation takes about 0.6 sec for two million pixels. During training, only the " + }, + { + "bbox": [ + 304, + 180, + 545, + 335 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 304, + 180, + 545, + 335 + ], + "type": "text", + "content": " matrix-vector multiplication Eq. (11) is required, which can almost be ignored compared to AD or FD." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 348, + 388, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 348, + 388, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 388, + 361 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 369, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 417 + ], + "type": "text", + "content": "We evaluate our method quantitatively on the MVPS benchmark in Sec. 4.1, investigate different components of our method in Sec. 4.2 and evaluate our method qualitatively on our captured real-world data in Sec. 4.3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 422, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 578 + ], + "type": "text", + "content": "Implementation details We apply SDM-UniPS [10] to images in each view to obtain normal maps. In each batch, we randomly sample 2048 patches of " + }, + { + "bbox": [ + 304, + 422, + 545, + 578 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 422, + 545, + 578 + ], + "type": "text", + "content": " pixels from all input normal maps. Like INGP [20], we skip the empty and occluded space during ray marching using NerfAcc [17]. The scene is bounded within a unit sphere, and we use a coarse-to-fine ray marching step decreasing log-linearly from " + }, + { + "bbox": [ + 304, + 422, + 545, + 578 + ], + "type": "inline_equation", + "content": "1e^{-2}" + }, + { + "bbox": [ + 304, + 422, + 545, + 578 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 422, + 545, + 578 + ], + "type": "inline_equation", + "content": "1e^{-3}" + }, + { + "bbox": [ + 304, + 422, + 545, + 578 + ], + "type": "text", + "content": ". We optimize the shape of DiLiGenT-MV [16] benchmark objects for 5000 batches (roughly equal to 20 epochs), and 30000 batches for our captured data. The weights of loss terms are set as 1. We use the Adam optimizer with an initial learning rate " + }, + { + "bbox": [ + 304, + 422, + 545, + 578 + ], + "type": "inline_equation", + "content": "5e^{-3}" + }, + { + "bbox": [ + 304, + 422, + 545, + 578 + ], + "type": "text", + "content": ". More details can be found in Sec. B in SupMat." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 586, + 440, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 440, + 599 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 440, + 599 + ], + "type": "text", + "content": "4.1. Quantitative evaluation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "content": "Dataset We evaluate our method using the only MVPS benchmark dataset DiLiGenT-MV [16]. It consists of five objects captured from 20 views, and 96 OLAT images are captured in each view, yielding a total of 1920 images per object. Each image is " + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "inline_equation", + "content": "512 \\times 612" + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "content": ", and the total number of foreground object pixels is about 2.2 million. The camera is fixed during capture, and the object is placed on a turntable in a darkroom. The \"ground truth\" meshes are created using a 3D scanner." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20585" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "Baselines We benchmark SuperNormal against several MVPS methods grouped into three categories. (1) Those that refine an initial point cloud or mesh and require manual effort, including R-MVPS [23] and B-MVPS [16]. (2) Those that use Poisson surface reconstruction (PSR) [13] to fuse multi-view depth and normal maps, including MVPSNet [35]. (3) Those that optimize an implicit neural surface representation, including UA-MVPS [12], PS-NeRF [31], MVAS [4], and our method." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "text", + "content": "Evaluation metrics We use L2 Chamfer distance (CD) and F-score with threshold " + }, + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\tau_{F} = 0.5\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "text", + "content": " to evaluate geometry accuracy [12, 15]. For CD and F-score, we only consider visible points from the input views by casting rays for all pixels and finding the first ray-mesh intersections (More details in Sec. B.4 in SupMat.)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 287, + 382 + ], + "type": "text", + "content": "Results As reported in Table 1, our method achieves the best geometry accuracy among all compared MVPS methods (See Sec. A in SupMat. for normal accuracy). We attribute this superior performance to two aspects: 1) the powerful expressive capability of multi-resolution hash encoding [20] and 2) the fine-grained normal estimation method SDM-UniPS [10], as discussed in Sec. 4.2. Figure 4 visualizes the recovered shapes of different methods. Our method recovers fine-grained surface details on par with the \"GT\" meshes produced by a 3D scanner." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 384, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 503 + ], + "type": "text", + "content": "Table 1 also reports the average runtime measured on our machine with one RTX 4090Ti graphics card. The runtime of R-MVPS [23] and B-MVPS [16] is difficult to measure since both methods require intensive manual effort. The code of UA-MVPS [12] is not publicly available, and several hours per object are required, as reported in MVPSNet [35]. Both PS-NeRF [31] and MVAS [4] require hours per object since they optimize a dense MLP. Our method is the fastest among neural implicit representation-based methods, with less than one minute per object." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 504, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 287, + 576 + ], + "type": "text", + "content": "On the other hand, MVPSNet [35] is highly efficient because its pipeline mainly consists of feedforward inferencing a trained neural network and an sPSR [13] step, the latter of which has been optimized over one decade. Although our method is slower than MVPSNet [35], we achieve better reconstruction quality within a reasonable time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 586, + 138, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 138, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 138, + 598 + ], + "type": "text", + "content": "4.2. Ablation study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "Effect of normal maps Since multi-view normal integration is not bound to specific normal estimation methods, we investigate the effect of different normal estimation methods on normal integration methods. We test three sets of normal maps, estimated by SDPS [6] or SDM-UniPS [10], and the one rendered from the scanned mesh (i.e., \"GT\" normal maps provided by DiLiGenT-MV [16]). Table 2 reports the L2-Chamfer distance and F-score averaged on five DiLiGenT-MV [16] objects. Our method consistently" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "achieves better results than PS-NeRF [31] and MVAS [4] when using the same set of normal maps. This verifies that our method can better fuse the information from 2.5D normal maps to 3D shapes. Using the SoTA method SDM-UniPS [10] for normal estimation, our method realizes the best reconstruction quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 148, + 545, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 545, + 245 + ], + "type": "text", + "content": "Effect of spatial encoding Table 2 also compares the geometry accuracy between multi-resolution hash encoding (hash enc.) and positional encoding (pos. enc.). Our results verify again that multi-resolution hash encoding surpasses positional encoding in expressiveness, enhancing geometry accuracy. Therefore, the keys to superior MVPS performance are high-quality normal estimation, an expressive surface representation, and an effective normal fusion approach." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 249, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 249, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 249, + 545, + 357 + ], + "type": "text", + "content": "Effect of SDF gradient computation Table 3 compares the geometry accuracy and runtime using finite difference (FD), Pytorch's automatic differentiation (AD), and directional finite difference (DFD) to compute the SDF gradient. All other parameter settings are kept the same. We do not observe a significant difference regarding reconstruction quality. However, DFD performs consistently faster than FD and AD by accelerating both the forward rendering and backward optimization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "text", + "content": "Generalizability of DFD In principle, DFD and patch-based sampling strategy apply to any SDF-based volume rendering like NeuS [29] or Neuralangelo [18]. To verify this, we incorporate DFD and patch-based sampling into Neuralangelo " + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "inline_equation", + "content": "[18]^{1}" + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "text", + "content": ". We follow the coarse-to-fine sampling strategy of Neuralangelo [18] but use DFD instead of FD for SDF gradient approximation. We do not use mask supervision for Neuralangelo [18]. As shown in Figure 5, training time is reduced by " + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "text", + "content": " on a scene from the DTU benchmark [11], even " + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "inline_equation", + "content": "2.25\\times" + }, + { + "bbox": [ + 304, + 361, + 545, + 506 + ], + "type": "text", + "content": " more pixels are looped in training. The result suggests that DFD could be a general strategy to accelerate NeuS-like multi-view reconstruction [29]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 513, + 516, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 513, + 516, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 513, + 516, + 525 + ], + "type": "text", + "content": "4.3. Qualitative results on our captured data" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "text", + "content": "To further demonstrate the performance of SuperNormal in fine-grained surface detail recovery, we also captured objects with more complex details, as shown in Fig. 6. Thanks to the generalizability of SDM-UniPS [10], we can capture the objects under casual light conditions (e.g., within an apartment) without ensuring a darkroom setup. We manually move a video light during the capture without any photometric or geometric light calibration. In total, we capture 18 viewpoints " + }, + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\times 13" + }, + { + "bbox": [ + 304, + 531, + 545, + 687 + ], + "type": "text", + "content": " light conditions for each object (1 under ambient light for SfM). We use Metashape [1] to calibrate camera parameters, use SAM [14] to create foreground masks, and apply SDM-UniPS [10] to every 12 image to infer the per-view normal map." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "1Our modified code of Neuralangelo [18] is available at https:// github.com/xucao-42/Neuralangelo_DFD.git." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20586" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 97, + 545, + 184 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 547, + 95 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 547, + 95 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation on DiLiGenT-MV [16] benchmark. Red and orange cells indicate the best and the second best results respectively. On average, our approach achieves the best reconstruction quality at the second fastest speed." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 97, + 545, + 184 + ], + "lines": [ + { + "bbox": [ + 48, + 97, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 545, + 184 + ], + "type": "table", + "html": "
MethodsL2 Chamfer distance [mm] (↓)F-score (τF=0.5 mm) (↑)Runtime (↓) Average
BearBuddhaCowPot2ReadingAverageBearBuddhaCowPot2ReadingAverage
R-MVPS [23]1.0700.3970.4401.5040.5610.7940.2620.6980.7600.1980.5190.487NA
B-MVPS [16]0.2120.2540.0910.2010.2590.2030.9580.9020.9860.9460.9140.941NA
MVPSSNet [35]0.3170.2790.2550.3100.2480.2820.8130.8660.8890.8380.9180.86537 secs
UA-MVPS [12]0.4140.4520.3260.4140.3820.3980.7070.6690.7980.7310.7620.733several hrs
PS-NeRF [31]0.2600.3140.2870.2540.3520.2930.8980.8060.8560.9190.7850.8538 hrs
MVAS [4]0.2430.3570.2160.1970.5220.3070.9090.7540.9070.9620.5460.81670 mins
Ours0.1840.2180.1930.1500.2230.1940.9830.9500.9770.9940.9160.96248 secs
", + "image_path": "d3bde6629ff85c8b95b728fc3091d52c14e12a022a38410f5f160cc5bb767f5a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 193, + 545, + 552 + ], + "blocks": [ + { + "bbox": [ + 48, + 193, + 545, + 552 + ], + "lines": [ + { + "bbox": [ + 48, + 193, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 193, + 545, + 552 + ], + "type": "image", + "image_path": "0cdb1c61c4408a135140062303efda8dce5b4344baf5c534ec20c2b9ffc6b89f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 559, + 512, + 572 + ], + "lines": [ + { + "bbox": [ + 80, + 559, + 512, + 572 + ], + "spans": [ + { + "bbox": [ + 80, + 559, + 512, + 572 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of recovered shapes of Buddha and Pot2 and their error maps. Best viewed on screen." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 632, + 287, + 682 + ], + "blocks": [ + { + "bbox": [ + 47, + 590, + 287, + 624 + ], + "lines": [ + { + "bbox": [ + 47, + 590, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 287, + 624 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation using normal maps estimated by different methods as inputs. The metrics are averaged over five DiLiGenT-MV [16] objects. Darker colors indicate better results." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 632, + 287, + 682 + ], + "lines": [ + { + "bbox": [ + 48, + 632, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 48, + 632, + 287, + 682 + ], + "type": "table", + "html": "
L2 Chamfer distance [mm] (↓)F-score (τF=0.5 mm) (↑)
SDPS [6]SDM-UniPS [10]GTSDPS [6]SDM-UniPS [10]GT
MVAS [4]0.3070.3320.2920.8160.8120.859
PS-NeRF [31]0.2930.2960.2240.8530.8440.938
Ours (pos, enc.)0.3010.2280.1950.8490.9420.977
Ours (hash enc.)0.2810.1940.1140.8600.9620.998
", + "image_path": "df881f485d86c96488007ec3b5d1c403c1ec442f31c5f6e4059d0c2c5f3c456e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 589, + 545, + 635 + ], + "blocks": [ + { + "bbox": [ + 306, + 589, + 545, + 635 + ], + "lines": [ + { + "bbox": [ + 306, + 589, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 589, + 545, + 635 + ], + "type": "image", + "image_path": "b5927e38fee03c5e9dc7d9779983a82a63e5ed685e2215e57463cb6c85e23ce4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 639, + 547, + 683 + ], + "lines": [ + { + "bbox": [ + 304, + 639, + 547, + 683 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 547, + 683 + ], + "type": "text", + "content": "Figure 5. DFD and patch-based sampling accelerate Neuralangelo's training without quality degradation. We train both alternatives for " + }, + { + "bbox": [ + 304, + 639, + 547, + 683 + ], + "type": "inline_equation", + "content": "20\\mathrm{k}" + }, + { + "bbox": [ + 304, + 639, + 547, + 683 + ], + "type": "text", + "content": " batches. For FD, 512 random rays are sampled per batch; for DFD, 128 random patches of " + }, + { + "bbox": [ + 304, + 639, + 547, + 683 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 304, + 639, + 547, + 683 + ], + "type": "text", + "content": " rays per batch." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20587" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 124, + 261, + 180 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 114 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 114 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 114 + ], + "type": "text", + "content": "Table 3. Ablation study using different strategies for SDF gradients computation in volume rendering. We report the metrics averaged over five DiLiGenT-MV [16] objects. FD: Finite difference, AD: Automatic differentiation, DFD: Directional finite difference." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 124, + 261, + 180 + ], + "lines": [ + { + "bbox": [ + 72, + 124, + 261, + 180 + ], + "spans": [ + { + "bbox": [ + 72, + 124, + 261, + 180 + ], + "type": "table", + "html": "
AccuracyRuntime (Sec)
CDF-ScoreForwardBackwardTotal
FD0.1920.96347.770.3124.6
AD0.1950.96129.442.478.1
DFD0.1940.96119.622.548.3
", + "image_path": "7e2703c87f857879dc1849a776eace2d859c06bd31315baa67a6f01154f52bdd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 52, + 192, + 167, + 279 + ], + "blocks": [ + { + "bbox": [ + 52, + 192, + 167, + 279 + ], + "lines": [ + { + "bbox": [ + 52, + 192, + 167, + 279 + ], + "spans": [ + { + "bbox": [ + 52, + 192, + 167, + 279 + ], + "type": "image", + "image_path": "38318f8bec8399f933f859ec2e8bdd664f5731d6dcae7705950fc61f9002425c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 287, + 287, + 343 + ], + "lines": [ + { + "bbox": [ + 46, + 287, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 287, + 343 + ], + "type": "text", + "content": "Figure 6. Our real-world data capture setup. The target object is placed on a turntable before a dark sheet under general indoor illumination. There is no need to prepare a darkroom. We use an iPhone to take images and move a video light to illuminate the object from the camera side. Objects are " + }, + { + "bbox": [ + 46, + 287, + 287, + 343 + ], + "type": "inline_equation", + "content": "6\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 287, + 287, + 343 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 287, + 287, + 343 + ], + "type": "inline_equation", + "content": "15\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 287, + 287, + 343 + ], + "type": "text", + "content": " high." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 191, + 282, + 279 + ], + "blocks": [ + { + "bbox": [ + 167, + 191, + 282, + 279 + ], + "lines": [ + { + "bbox": [ + 167, + 191, + 282, + 279 + ], + "spans": [ + { + "bbox": [ + 167, + 191, + 282, + 279 + ], + "type": "image", + "image_path": "3cea14033fa30364ecc607094b1aeab094a1e9a4dcd60b16e48dfb07d9c7987d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 355, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 287, + 450 + ], + "type": "text", + "content": "Figures 1 and 7 display the results from NeuS2 [30], SuperNormal, and a structured-light based commercial 3D scanner EinScan SE. Although NeuS2 [30] and our method use the same multi-resolution hash encoding for fine-grained scene representation, the surface details recovered from multi-view color images are smoothed out. Our method, benefitting from the normal map inputs, recovers more fine-grained surface details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 452, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 452, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 452, + 287, + 548 + ], + "type": "text", + "content": "We find that our scanner performs poorly in highly concave regions (Fig. 7 above). This is because the scanner uses a wide baseline stereo camera (about " + }, + { + "bbox": [ + 46, + 452, + 287, + 548 + ], + "type": "inline_equation", + "content": "15\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 452, + 287, + 548 + ], + "type": "text", + "content": " apart), and small concave regions create self-occlusions, thus reducing the reconstruction quality. In principle, normal estimation does not suffer from self-occlusion because it is based on single-view observations. Fusing multi-view normal maps can then produce better results in concave regions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 559, + 117, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 117, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 117, + 572 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "type": "text", + "content": "We have demonstrated SuperNormal, a fast, high-fidelity 3D reconstruction approach using multi-view normal maps estimated by photometric stereo. SuperNormal outperforms other neural MVPS methods in speed and reconstruction quality on the benchmark dataset. Compared to multi-view neural reconstruction methods, using normal maps can exploit the expressive power of multi-resolution hash encoding, and produce scanner-level surface details. Further, we have shown the effectiveness of directional finite differences in expediting training without sacrificing quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "SuperNormal provides a promising direction to make" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 545, + 358 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 545, + 358 + ], + "type": "image", + "image_path": "c13d95a870b71c1cdb722c7c9ecd34ba454525d9334c13f2f53c0d2f5e68843a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 362, + 545, + 396 + ], + "lines": [ + { + "bbox": [ + 305, + 362, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 362, + 545, + 396 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison using our objects. Our approach recovers better surface detail than multi-view stereo approaches and achieves competitive results with commercial 3D scanners." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 406, + 430, + 498 + ], + "blocks": [ + { + "bbox": [ + 309, + 406, + 430, + 498 + ], + "lines": [ + { + "bbox": [ + 309, + 406, + 430, + 498 + ], + "spans": [ + { + "bbox": [ + 309, + 406, + 430, + 498 + ], + "type": "image", + "image_path": "0cfc59baf60775ffccca5f973d02bf866d9e9f00c5a44ab97ee07affe14405ce.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 498, + 345, + 505 + ], + "lines": [ + { + "bbox": [ + 329, + 498, + 345, + 505 + ], + "spans": [ + { + "bbox": [ + 329, + 498, + 345, + 505 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 514, + 545, + 548 + ], + "lines": [ + { + "bbox": [ + 305, + 514, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 305, + 514, + 545, + 548 + ], + "type": "text", + "content": "Figure 8. Limitations. (Left) Fault artifacts are highlighted in red circles. (Right) Our method is sensitive to the inaccuracy in normal maps estimated by monocular methods (e.g., OmniData [7])." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 437, + 407, + 498, + 497 + ], + "blocks": [ + { + "bbox": [ + 437, + 407, + 498, + 497 + ], + "lines": [ + { + "bbox": [ + 437, + 407, + 498, + 497 + ], + "spans": [ + { + "bbox": [ + 437, + 407, + 498, + 497 + ], + "type": "image", + "image_path": "e13b646eb49d1ada047b5971a0c1d4b99296474726c8ff9fe7100f76ef325ffd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 498, + 491, + 506 + ], + "lines": [ + { + "bbox": [ + 438, + 498, + 491, + 506 + ], + "spans": [ + { + "bbox": [ + 438, + 498, + 491, + 506 + ], + "type": "text", + "content": "OmniData normal" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 498, + 407, + 542, + 497 + ], + "blocks": [ + { + "bbox": [ + 498, + 407, + 542, + 497 + ], + "lines": [ + { + "bbox": [ + 498, + 407, + 542, + 497 + ], + "spans": [ + { + "bbox": [ + 498, + 407, + 542, + 497 + ], + "type": "image", + "image_path": "f03c38d05d82dad187b9984e45f66d220d7e02ef2cf676b8605da37d2b39283c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 498, + 528, + 505 + ], + "lines": [ + { + "bbox": [ + 512, + 498, + 528, + 505 + ], + "spans": [ + { + "bbox": [ + 512, + 498, + 528, + 505 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 559, + 545, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 545, + 607 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 545, + 607 + ], + "type": "text", + "content": "high-quality 3D scanning more accessible. Unlike costly commercial 3D scanners, our approach is viable with everyday devices like smartphones and basic equipment such as a tripod, video light, and a turntable." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 616, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 546, + 712 + ], + "type": "text", + "content": "Limitations Our method may produce fault-like artifacts, as shown in Fig. 8. The artifacts exist no matter our SDF gradient computation approach (i.e., AD, FD, or DFD) and require further investigation. Moreover, the performance depends on mask supervision and high-quality normal maps (e.g., those obtained from photometric stereo). Our method does not work well with the monocular normal estimation methods like OmniData [7] (Fig. 8)." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20588" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 276, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 276, + 102 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 276, + 102 + ], + "type": "text", + "content": "[1] Agisoft metashape. https://www.agisoft.com.6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 102, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 102, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 102, + 287, + 145 + ], + "type": "text", + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 287, + 178 + ], + "type": "text", + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-NeRF: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 181, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 181, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 181, + 287, + 224 + ], + "type": "text", + "content": "[4] Xu Cao, Hiroaki Santo, Fumio Okura, and Yasuyuki Matsushita. Multi-view azimuth stereo via tangent space consistency. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 825-834, 2023. 1, 2, 6, 7, 11, 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 225, + 287, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 287, + 269 + ], + "type": "text", + "content": "[5] Ju Yong Chang, Kyoung Mu Lee, and Sang Uk Lee. Multiview normal field integration using level set methods. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 1-8. IEEE, 2007. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 270, + 287, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 287, + 302 + ], + "type": "text", + "content": "[6] Guanying Chen, Kai Han, Boxin Shi, Yasuyuki Matsushita, and Kwan-Yee K. Wong. SDPS-Net: Self-calibrating deep photometric stereo networks. In CVPR, 2019. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 303, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 287, + 345 + ], + "type": "text", + "content": "[7] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In ICCV, pages 10786-10796, 2021. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 347, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 287, + 390 + ], + "type": "text", + "content": "[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of Machine Learning and Systems 2020, pages 3569-3579. 2020. 3, 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 392, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 287, + 434 + ], + "type": "text", + "content": "[9] Carlos Hernandez, George Vogiatzis, and Roberto Cipolla. Multiview photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 30(3):548-554, 2008. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 436, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 436, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 436, + 287, + 468 + ], + "type": "text", + "content": "[10] Satoshi Ikehata. Scalable, detailed and mask-free universal photometric stereo. In Proc. of Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 470, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 470, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 470, + 287, + 502 + ], + "type": "text", + "content": "[11] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, pages 406-413. IEEE, 2014. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 503, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 546 + ], + "type": "text", + "content": "[12] Berk Kaya, Suryansh Kumar, Carlos Oliveira, Vittorio Ferrari, and Luc Van Gool. Uncertainty-aware deep multi-view photometric stereo. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 12601-12611, 2022. 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 548, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 579 + ], + "type": "text", + "content": "[13] Michael Kazhdan and Hugues Hoppe. Screened Poisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 581, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 645 + ], + "type": "text", + "content": "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proc. of International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[15] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[16] Min Li, Zhenglong Zhou, Zhe Wu, Boxin Shi, Changyu Diao, and Ping Tan. Multi-view photometric stereo: A ro" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 106 + ], + "type": "text", + "content": "bust solution and benchmark dataset for spatially varying isotropic materials. IEEE Transactions on Image Processing, 29:4159-4173, 2020. 2, 5, 6, 7, 8, 11, 12" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 150 + ], + "type": "text", + "content": "[17] Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. NerfAcc: Efficient sampling accelerates nerfs. In Proc. of International Conference on Computer Vision (ICCV), pages 18537-18546, 2023. 5, 13" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[18] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 8456-8465, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "text", + "content": "[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. of European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "text", + "content": "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 309, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 353 + ], + "type": "text", + "content": "[21] Michael Oechsle, Songyou Peng, and Andreas Geiger. UNISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proc. of International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 387 + ], + "type": "text", + "content": "[22] Stanley Osher, Ronald Fedkiw, and K Piechor. Level set methods and dynamic implicit surfaces. Appl. Mech. Rev., 57(3):B15-B15, 2004. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "type": "text", + "content": "[23] Jaesik Park, Sudipta N Sinha, Yasuyuki Matsushita, Yu-Wing Tai, and In So Kweon. Robust multiview photometric stereo using planar mesh parameterization. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 39(8):1591-1604, 2016. 2, 6, 7, 11, 12" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "type": "text", + "content": "[24] Radu Alexandru Rosu and Sven Behnke. Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices. In CVPR, 2023. 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "text", + "content": "[25] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. of European Conference on Computer Vision (ECCV), 2016. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 523, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 588 + ], + "type": "text", + "content": "[26] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my NeRF: A dataset for novel view synthesis and relighting of real world objects. In Proc. of Computer Vision and Pattern Recognition (CVPR), pages 20762-20772, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "type": "text", + "content": "[27] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "text", + "content": "[28] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. In ECCV, 2022. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[29] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view recon" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20589" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 463 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "struction. Advances in Neural Information Processing Systems (NeurIPS), 2021. 1, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "text", + "content": "[30] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. NeuS2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proc. of International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 4, 8, 14, 15" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 286, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 286, + 205 + ], + "type": "text", + "content": "[31] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K. Wong. PS-NeRF: Neural inverse rendering for multi-view photometric stereo. In Proc. of European Conference on Computer Vision (ECCV), 2022. 1, 2, 6, 7, 11, 12, 14" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 286, + 262 + ], + "type": "text", + "content": "[32] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 286, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 286, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 286, + 306 + ], + "type": "text", + "content": "[33] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems (NeurIPS), 34:4805-4815, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 286, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 286, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 286, + 351 + ], + "type": "text", + "content": "[34] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. NIPS, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 353, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 286, + 407 + ], + "type": "text", + "content": "[35] Dongxu Zhao, Daniel Lichy, Pierre-Nicolas Perrin, Jan-Michael Frahm, and Soumyadip Sengupta. MVPSNet: Fast generalizable multi-view photometric stereo. In Proc. of International Conference on Computer Vision (ICCV), pages 12525-12536, 2023. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 409, + 286, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 286, + 463 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 286, + 463 + ], + "type": "text", + "content": "[36] Fuqiang Zhao, Yuheng Jiang, Kaixin Yao, Jiakai Zhang, Liao Wang, Haizhao Dai, Yuhui Zhong, Yingliang Zhang, Minye Wu, Lan Xu, and Jingyi Yu. Human performance modeling and rendering via neural animated mesh. ACM Trans. Graph., 41(6), 2, 3, 4" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20590" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/556fb01c-1445-4c08-b116-0426ba52ea17_content_list.json b/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/556fb01c-1445-4c08-b116-0426ba52ea17_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3933d026ff1535221caacfe9821f6a57f2e5eef8 --- /dev/null +++ b/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/556fb01c-1445-4c08-b116-0426ba52ea17_content_list.json @@ -0,0 +1,1746 @@ +[ + { + "type": "text", + "text": "SuperPrimitive: Scene Reconstruction at a Primitive Level", + "text_level": 1, + "bbox": [ + 189, + 130, + 779, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kirill Mazur, Gwangbin Bae, Andrew J. Davison Dyson Robotics Lab, Imperial College London", + "bbox": [ + 269, + 180, + 697, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{k.mazur21, g.bae, a.davison}@imperial.ac.uk", + "bbox": [ + 287, + 219, + 676, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Joint camera pose and dense geometry estimation from a set of images or a monocular video remains a challenging problem due to its computational complexity and inherent visual ambiguities. Most dense incremental reconstruction systems operate directly on image pixels and solve for their 3D positions using multi-view geometry cues. Such pixel-level approaches suffer from ambiguities or violations of multi-view consistency (e.g. caused by textureless or specular surfaces).", + "bbox": [ + 75, + 300, + 470, + 435 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We address this issue with a new image representation which we call a SuperPrimitive. SuperPrimitives are obtained by splitting images into semantically correlated local regions and enhancing them with estimated surface normal directions, both of which are predicted by state-of-the-art single image neural networks. This provides a local geometry estimate per SuperPrimitive, while their relative positions are adjusted based on multi-view observations.", + "bbox": [ + 75, + 436, + 468, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We demonstrate the versatility of our new representation by addressing three 3D reconstruction tasks: depth completion, few-view structure from motion, and monocular dense visual odometry. Project page: https://makezur.github.io/SuperPrimitive/", + "bbox": [ + 75, + 556, + 468, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 648, + 209, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Enriching monocular incremental reconstruction with prior world knowledge is essential for resolving visual ambiguities. This issue is particularly prevalent in scenarios with scarce data observations available: a notable example would be monocular visual SLAM, where images are being streamed from a camera into the system in real-time.", + "bbox": [ + 75, + 674, + 468, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When a monocular vision system encounters a new scene region, it must estimate the region's geometry based on a very limited number of observations. Without this, continuous camera motion tracking would not be possible. Once the scene region is thoroughly observed, the initial geometry estimate should be refined to better explain the multiview information.", + "bbox": [ + 75, + 763, + 468, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This naturally leads to a question: what sort of priors are effective in both providing reliable initial geometry es", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/374776498f3c7f0c4217067d977995de1c873b9ec63f7fe2fce7212e571eb247.jpg", + "image_caption": [ + "Figure 1. Multi-View Geometry with SuperPrimitives. SuperPrimitives are extracted from an input frame by dividing it into image segments equipped with estimated surface normal directions (bottom-left). Each SuperPrimitive induces a dense reconstruction within the corresponding image segment up to a priori unknown scale. Different possible reconstructions are shown in light blue. The scales are then jointly optimised together with a relative camera pose to fit multi-view photometric constraints (visualised in green and red). The resulting dense reconstruction of the reference frame is shown in the top." + ], + "image_footnote": [], + "bbox": [ + 509, + 273, + 870, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "timates and supporting multi-view consistency? Geometric priors generally fall into one of two categories: local and global. Local priors, such as smoothness assumption [30] or surface normal regularisation [52], impose additional constraints within a small neighbourhood. Global priors, on the other hand, aim to impose constraints on a larger scale, such as depth prediction [2, 11].", + "bbox": [ + 496, + 744, + 890, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our key observation is that some of the geometrical correlations are more reliable than the others, and therefore could be safely \"locked in\" together within local regions", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4979", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "based on a single-view prediction. Points belonging to the same rigid body are strongly correlated, making it unnecessary to determine their depth independently. In contrast, distinct and unrelated objects can be placed arbitrarily in the scene. As the number of objects increases, learning a reliable global prior on their relative positions becomes an increasingly complex problem.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we show that purely local but strong priors are enough to achieve excellent performance across a variety of geometric vision tasks. For that purpose, we introduce a novel representation, SuperPrimitives. A Super-Primitive represents a local image segment, coupled with a dense shape estimate which is determined up to a scale factor. The scale factor can be further adjusted based on information observed from other views or additional measurements.", + "bbox": [ + 75, + 196, + 468, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We show that SuperPrimitives can be efficiently constructed using a front-end which consists of two single-image neural networks, extracting image segmentation and surface normal prediction. Effectively, our neural front-end predicts whether adjacent pixels belong to the same geometrical entity through image segmentation and estimates a surface normal at this point, thereby providing an infinitesimal geometry estimate.", + "bbox": [ + 75, + 334, + 467, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We delegate global, scene-level alignment to our streamlined multi-view, iterative, optimisation-based back-end. The resulting front-end / back-end tandem combines the flexibility of multi-view based optimisation methods with the observation efficiency common in prior-driven systems.", + "bbox": [ + 75, + 455, + 467, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our new representation showcases its versatility in three key applications, where single-view ambiguity is resolved via additional measurements or viewpoints:", + "bbox": [ + 75, + 531, + 467, + 575 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Firstly, it adeptly handles zero-shot depth completion tasks in real-world scenarios, matching the performance of state-of-the-art methods tailored for depth completion;", + "- Secondly, it facilitates joint pose and depth estimation using a limited set of unstructured images, surpassing its nearest competitor even in the absence of global priors;", + "- Thirdly, our method outperforms previous monocular visual odometry systems on the challenging TUM dataset, and exhibits robustness across various domains." + ], + "bbox": [ + 76, + 577, + 467, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 723, + 217, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Monocular Reconstruction. Both offline monocular reconstruction systems, such as COLMAP [43] and [16], or online systems, such as MonoSLAM [9] and DSO [22] track, filter, and reconstruct only well-constrained points with high and reliable photometric information. This often involves fine-grained and sophisticated point management to reliably resolve visual ambiguities and filter unreliable visual observations. DTAM [30] demonstrated feasibility of incremental dense reconstruction in the monocular scenario. DTAM employed a hand-crafted local smoothness", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "prior to handle regions with poor texture. Subsequently, other local priors [32, 54] have been extensively explored to regularise multi-view geometry estimation problems.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Global Priors. We are interested in exploring the space of possible geometric priors from single-view networks. Depth prediction [12, 37, 38] is the most obvious choice but leads to a rigid per-image reconstruction which is difficult to feed into multi-view optimisation. In the recent years, deep-learning based approaches sought to replace explicit multi-view geometry estimating with learning-based methods [3, 42, 48]. These methods, however, assume known poses and are, therefore, not suitable for joint pose and geometry estimation. Notably, CodeSLAM [2] introduced depth prediction conditioned on latent codes, which are then optimised to achieve cross-view consistency. However, CodeSLAM still struggles with out-of-domain data, as depth prediction networks are known to struggle with generalisation [51].", + "bbox": [ + 496, + 138, + 890, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Higher-Level Mapping. Introducing parametric primitives, such as lines [36, 50], planes [6, 17, 19, 25, 26, 45] or even high-order algebraic shapes [7, 23, 31] to better constrain multi-view geometry problems have been thoroughly explored over the last few decades. They all however use assumptions which may not often hold for all 3D scenes, especially for dense reconstruction.", + "bbox": [ + 496, + 367, + 890, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides parametric algebraic primitives object-level mapping has been explored in the last decade. SLAM++ [41] represented a map with a set of CAD models retrieved from an existing database and tracked camera position against these models. This method, however, could not be applied in a variety of settings due to the limited size of the CAD databases. Even objects of the same class can vary in their geometric appearance from instance to instance. This was later approached by [15, 47], who also represent their map via a set of objects, but learn a latent space of possible geometric variations per object class. Fusion++ [29] extended this approach into a more versatile per object depth fusion using an RGB-D sensor and mask proposals from a pre-trained neural network. These methods rely on an RGB-D inputs and assume a pre-defined set of object classes, therefore lacking generality.", + "bbox": [ + 496, + 473, + 890, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 724, + 589, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Firstly, we introduce the concept of a SuperPrimitive and explain how an image is processed into a set of its SuperPrimitives, referring to this part of our method as the front-end. In the second part of this section, we describe how multi-view geometry problems can be reformulated at the level of SuperPrimitives instead of pixels. We refer to this stage of SuperPrimitive alignment as the back-end.", + "bbox": [ + 496, + 750, + 890, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The core of our method proposes splitting a given image into a set of (possibly overlapping) minimal segments, image regions which are likely to have strongly correlated", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4980", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2394536898c6d7b5ebe1603afd03c5680681346861012fb5086549ac85b35474.jpg", + "image_caption": [ + "Figure 2. SuperPrimitives Extraction. (left) Our front-end processor extracts SuperPrimitives from an image by dividing it into a set of image regions with surface normal directions estimated for each image pixel within the segment. (right) Highlighted SuperPrimitives extracted from the image are visualised by showing their estimated normal and colour maps side by side. Note some of them are scaled either up or down for better viewing. While some of the SuperPrimitives are akin to object-level segmentation, the others tend to represent more low-level image segments." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 261, + 311 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/29acfc2df01382d0829f980c75c3a252ad07ed79291b184ab759c9f6ccd96a22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 89, + 883, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "geometry. We repurpose the recent state-of-the-art Segment Anything (SAM) [21] model into predicting these minimal segments.", + "bbox": [ + 75, + 393, + 468, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our key idea is to estimate local geometry within each segment from a single view, while leaving the relative positioning of the segments to be estimated via multi-view photometric consistency optimisation. We refer to these geometrically enhanced minimal image segments as Super-Primitives, since they are inspired by both superpixels [39] and geometric 3D primitives.", + "bbox": [ + 75, + 440, + 468, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For per-segment local geometry estimation we employ an off-the-shelf surface normal prediction network [1], to estimate infinitesimal geometry for each image pixel. The surface normals within an image segment can be used to estimate its depth via simple integration up to a scale factor. We set these scale factors — depth scales — to be optimisable parameters, which are either optimised via multi-view cues or explicit depth measurements for the depth completion experiments.", + "bbox": [ + 75, + 546, + 468, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thus, our method combines the strong priors provided by state-of-the-art neural networks in the front-end with the flexibility and consistency offered by multi-view optimisation.", + "bbox": [ + 75, + 681, + 468, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Conventions", + "text_level": 1, + "bbox": [ + 76, + 746, + 209, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unless stated otherwise, we use lowercase letters for scalar values, bold letters for vectors, and uppercase letters for matrices. We consider images $\\mathbf{I} \\in \\mathbb{R}^{3 \\times h \\times w}$ of height $h$ and width $w$ captured by a camera with a known calibration matrix $\\mathbf{K} = \\left( \\begin{array}{ccc} f_u & 0 & c_u \\\\ 0 & f_v & c_v \\\\ 0 & 0 & 1 \\end{array} \\right) \\in \\mathbb{R}^{3 \\times 3}$ . Image pixels are parameterised as $\\mathbf{u} = (u, v) \\in [0, h - 1] \\times [0, w - 1]$ . Given a per pixel depth function $z(\\mathbf{u})$ : $\\mathbf{I} \\to \\mathbb{R}^+$ , we define image's unprojection onto the 3D space to be $\\pi^{-1}(\\mathbf{u}) = z(\\mathbf{u})\\mathbf{K}^{-1}\\dot{\\mathbf{u}}$ ,", + "bbox": [ + 75, + 768, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where dot is a homogenisation operator $\\dot{\\pmb{p}} = (\\pmb{p},1)$ . Conversely, for a set of 3D points $\\pmb{x} = (x,y,z) \\in \\mathbb{R}^3$ we define the projection $\\pi(\\pmb{x}) = \\gamma\\left(\\frac{1}{z} K\\pmb{x}\\right)$ , where the $\\gamma$ function drops the last coordinate $\\gamma((x,y,z)^T) = (x,y)^T$ .", + "bbox": [ + 496, + 393, + 890, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We represent camera poses as matrices $\\mathrm{T}_{WC} \\in \\mathrm{SE}(3)$ in the camera-to-world coordinates unless stated otherwise. During pose optimisation we store linearised pose increments $\\xi \\in \\mathfrak{se}(3)$ as Lie algebra elements. These increments are used then to update current camera pose estimates, see [13, 14] for more detail.", + "bbox": [ + 496, + 455, + 890, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. SuperPrimitives", + "text_level": 1, + "bbox": [ + 500, + 556, + 661, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We split the input image I into a set of possibly overlapping connected regions (segments), each of which is equipped with its local 3D geometry. More formally, a SuperPrimitive $P = (\\Omega, \\mathfrak{D})$ is a connected image region $\\Omega \\subseteq I$ which is also equipped with an unscaled depth map $\\mathfrak{D} : \\Omega \\to [0, \\infty)$ for each pixel within the region $\\Omega$ . Here, unscaled means that each SuperPrimitive's depth $\\mathfrak{D}$ differs from its ground truth depth $D$ by an a priori unknown scalar. In other words, there exists a single scalar $s$ such that:", + "bbox": [ + 496, + 580, + 890, + 715 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD = s \\cdot \\mathfrak {D} \\quad \\text {f o r a l l p i x e l s} \\mathrm {p} \\in \\Omega \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 719, + 890, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We say that an image $\\mathrm{I}$ is primified into super-primitives $\\mathcal{P}(\\mathrm{I}) = \\{P_i = (\\Omega_i, \\mathfrak{D}_i)\\}$ if it has $n$ , possibly overlapping primitives that lie within the image $\\mathrm{I}$ , i.e. $\\bigcup \\Omega_i \\subseteq \\mathrm{I}$ .", + "bbox": [ + 496, + 739, + 890, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For brevity we use the words primitive and SuperPrimitive interchangeably throughout the rest of the paper.", + "bbox": [ + 496, + 786, + 890, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Representing Geometry with SuperPrimitives. A set of primitives $\\mathcal{P}$ itself is not enough to extract dense image geometry due to scale/depth ambiguity. Therefore, we introduce the concept of optimisable depth scales which anchor the correct depth scale for each primitive.", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4981", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a scalar depth scale $s$ for a super-primitive $P$ , one can infer its depth as $D(p) = s \\cdot \\mathfrak{D}(p)$ for every point $p \\in \\Omega$ . In practice, we employ a log-depth representation to represent depth throughout the whole system. This means that we store optimisable log-depth scales $\\log s$ and unscaled log-depth $\\log \\mathfrak{D}$ , and so depth inference reduces to a simple shift operation $\\log D = \\log s + \\log \\mathfrak{D}$ .", + "bbox": [ + 75, + 90, + 467, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to other dense methods, we estimate the dense geometry $\\mathcal{G}$ of the image in the form of a point cloud rather than a depth image. This choice is driven by the fact that multiple primitives' supporting segments $\\Omega_{i}$ and $\\Omega_{j}$ might overlap, which would lead to different depth estimates within their intersection $\\Omega_{i} \\cap \\Omega_{j}$ .", + "bbox": [ + 75, + 196, + 467, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a set of depth scaled primitives $\\mathbb{P} = \\{(s_i, P_i)\\}$ , the geometry $\\mathcal{G}$ of the image $I$ can be estimated as follows:", + "bbox": [ + 76, + 289, + 467, + 319 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} := \\bigcup_ {i} \\pi^ {- 1} \\left(s _ {i} \\cdot \\mathfrak {D} _ {i}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 334, + 467, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain estimated depth $\\hat{D}$ , we average depth values along the corresponding camera rays.", + "bbox": [ + 76, + 369, + 467, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Converting an Image into SuperPrimitives. Next we explain how SuperPrimitives are obtained in practice. An input image $I$ is first split into a set of minimal image segments $\\Omega_{i}$ produced by an image segmentation model. We use the Segment Anything [21] model in this work, since it captures low-level semantic correlations highly accurately. Given the extracted segments, we independently estimate the local geometry $\\mathfrak{D}_i$ within each segment by integrating surface normals predicted by another state-of-the-art neural network [1] within this region.", + "bbox": [ + 75, + 409, + 467, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that this approach is premised on the assumption of geometry being continuous within each predicted image segment $\\Omega_{i}$ , hence its unscaled depth values could be obtained by integrating its surface normals. Although this is not guaranteed a priori, we observed that a more fine-grained mask selection yields a compelling correlation with geometrical continuity. We discuss this in more detail in the following section.", + "bbox": [ + 75, + 561, + 467, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image Segment Retrieval. An ideal neural network which estimates image segments $\\{\\Omega_i\\}$ should predict regions of geometric continuity, where depth could be obtained via surface normal integration. However, to the best of our knowledge, such a network does not exist. Our approach aims to approximate this behaviour by utilising the Segment Anything model, coupled with a specialised mask selection process. Specifically, for each SAM query, we strive to select the smallest predicted mask surrounding that point. While such over-segmentation could potentially increase the dimensionality of the multi-view optimisation problem, under-segmentation might lead to incorrect geometry estimate within the primitive. Since we only adjust the scale of a primitive after the normal integration is com", + "bbox": [ + 75, + 689, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pleted, any incorrect geometry estimates within a primitive cannot be compensated at this stage.", + "bbox": [ + 498, + 90, + 888, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The segmentation model employed in this work outputs three binary segmentation masks given a query point $q \\in \\mathbf{I}$ . For each query $q$ we first filter predicted masks using the post-processing introduced in [21], such as stability filtering and Non-Maximum-Suppression (NMS). Even though our method allows redundant regions, we employed NMS to remove similar segments and save compute in the image alignment stage. With filtering done, the mask with the smallest area is selected. If the filtered set is empty, we discard the query point.", + "bbox": [ + 496, + 121, + 890, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We query the SAM backbone feature extractor once per image. Then we first sample 300 query points randomly across the image, followed by the filtering discussed above. Then the image coverage mask is calculated and an additional 100 mask query points are actively sampled in the uncovered regions.", + "bbox": [ + 496, + 272, + 890, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Normal Integration. To estimate each segment's local geometry, we first pass the image through a surface normal estimation network. This network effectively predicts the derivative of desired depth values. Then, we \"integrate\" these surface normal vectors within each image segment to obtain its unscaled depth map.", + "bbox": [ + 496, + 364, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each image pixel $\\boldsymbol{u} = (u,v)^T \\in \\mathrm{I}$ we estimate its surface normal vector $\\boldsymbol{n} = (n_x, n_y, n_z)^T$ with a pretrained state-of-the-art CNN [1]. As was demonstrated in [5], the log-depth $\\tilde{z} = \\log(z(\\boldsymbol{u}))$ satisfies the following PDEs within a segment $\\Omega$ :", + "bbox": [ + 496, + 455, + 890, + 530 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {n} _ {z} \\partial_ {u} \\tilde {z} + n _ {x} = 0 \\quad \\text {a n d} \\quad \\tilde {n} _ {z} \\partial_ {v} \\tilde {z} + n _ {y} = 0 \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 534, + 890, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{n}_z = n_x(u - c_u) + n_y(v - c_v) + n_zf$ . The depth values within the segment $\\Omega$ can therefore be obtained via minimising the following functional:", + "bbox": [ + 496, + 553, + 888, + 598 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {z} \\iint_ {\\Omega} \\left(\\tilde {n} _ {z} \\partial_ {u} \\tilde {z} + n _ {x}\\right) ^ {2} + \\left(\\tilde {n} _ {z} \\partial_ {u} \\tilde {z} + n _ {y}\\right) ^ {2} d u d v \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 599, + 890, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that this leads to a family of solutions $\\tilde{z} + \\text{Const}$ which differ by a shift constant. This is due to the fact that only partial derivatives of $\\tilde{z}$ are used in the functional. That means that the actual depth will be estimated up to scale, since $z = \\exp(\\tilde{z})$ by definition.", + "bbox": [ + 496, + 642, + 890, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This ambiguity prompted us to introduce the notion of depth scales, to allow each segment to be adjusted towards its true depth. We implemented batched normal integration, efficiently solving all optimisation problems as a single sparse linear system using conjugate gradient method [18]. This implementation enables the integration of approximately 100 to 300 segments within a total time frame of $\\sim 100\\mathrm{ms}$ .", + "bbox": [ + 496, + 718, + 890, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Primitive-based Image Alignment", + "text_level": 1, + "bbox": [ + 500, + 847, + 794, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now we explain how our new representation can be used as a building block for dense multi-view geometry and pose", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4982", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "estimation, combining an optimisation-based mindset with learned single view priors. To make optimisation computationally tractable, we design our image alignment to fully abstract away any knowledge about the neural networks involved in the image primification stage. Hence, even though there has been early evidence that SAM is independently capable of establishing (coarse) mask correspondences across neighbouring frames, we rely on photometric information only during the image alignment stage. We believe that more semantic primitive association would be \"heavier\" and coarser since it would not provide per pixel correspondences, and would therefore not be a good fit to be used in a multi-view optimisation loop. However, we speculate that such a method could be employed to further enhance our system, e.g. to improve occlusion handling, particularly as these models become more computationally affordable.", + "bbox": [ + 75, + 90, + 472, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1 Two-view SfM on SuperPrimitives", + "text_level": 1, + "bbox": [ + 76, + 347, + 372, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the sake of clarity, we formulate our method in the simplest case of two frames observing the scene, although it could be trivially extended into a setting with a higher number of views. At the core of our Structure-from-Motion (SfM) approach lies a per-primitive photometric alignment, which contrasts with widely accepted per-pixel photometric alignment [14, 30] techniques. Informally speaking, our proposition is to treat each image primitive as a \"rigid\" piece, which is only allowed to be scaled up or down. This degree of freedom is due to the scale ambiguity present for each primitive. Thus, instead of estimating a depth value per pixel, we only estimate a depth-scale per primitive, which is illustrated in Fig. 1. This greatly reduces the dimensionality of the optimisation problem, especially in the case of an unknown relative pose. Note that our method does not require the target image to be primified nor does it not require any pre-established correspondences (e.g. primitive to primitive).", + "bbox": [ + 75, + 371, + 472, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We assume a primified reference image $\\mathrm{I}_{\\mathrm{ref}}$ with its set of primitives $\\mathcal{P}(\\mathrm{I}) = \\{(\\Omega_i, \\mathfrak{D}_i)\\}$ . Given an unposed target image $\\mathrm{I}_{\\mathrm{t}}$ , our goal is to jointly estimate its relative pose $\\mathrm{T}_{tr}$ with respect to the reference image $\\mathrm{I}_{\\mathrm{ref}}$ as well as the dense geometry $\\mathcal{G}_{\\mathrm{ref}} = \\bigcup_{i} \\pi^{-1}(s_i \\cdot \\mathfrak{D}_i)$ of the reference frame.", + "bbox": [ + 75, + 647, + 470, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While unscaled depths $\\mathfrak{D}_i$ are given by a per-image preprocessor after the primification of the reference image, the set of depth scales $s_i$ are yet to be estimated. In our case, we jointly estimate both depth scales $s_i$ and a relative image pose $\\mathrm{T}_{tr}$ by solving a photometric consistency optimisation problem. First, we warp each depth-scaled primitive $(s_i, P_i)$ from the reference frame $\\mathrm{I}_{\\mathrm{ref}}$ into the target frame $\\mathrm{I}_{\\mathrm{t}}$ :", + "bbox": [ + 75, + 726, + 470, + 847 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {P} _ {i} [ \\boldsymbol {u} ] = \\pi \\left(\\mathrm {T} _ {t r} \\pi^ {- 1} \\left(\\boldsymbol {u}, s _ {i} \\cdot \\mathfrak {D} _ {i}\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 849, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then a per-segment photometric residual for the primitive $P_{i} = (\\Omega_{i},\\mathfrak{D}_{i})$ is defined by averaging all photometric re", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "projection $\\ell^1$ errors across every pixel $\\pmb{u} \\in \\Omega_i$ :", + "bbox": [ + 498, + 90, + 808, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr \\left(P _ {i}, s _ {i}, I _ {\\mathrm {t}}, T _ {t r}\\right) = \\frac {1}{\\left| \\Omega_ {i} \\right|} \\sum_ {\\boldsymbol {u} \\in \\Omega_ {i}} \\left\\| I _ {\\text {r e f}} (\\boldsymbol {u}) - I _ {\\mathrm {t}} \\left(\\hat {P} _ {i} [ \\boldsymbol {u} ]\\right) \\right\\| _ {1} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 117, + 892, + 156 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The resulting photometric cost, aggregated across all depth-scaled primitives $(s_i, P_i) \\in \\mathcal{P}(\\mathrm{I})$ , is:", + "bbox": [ + 498, + 166, + 890, + 198 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\text {p h o t o}} = \\frac {1}{| \\mathcal {P} (\\mathrm {I}) |} \\sum_ {P _ {i} = \\left(\\Omega_ {i}, \\mathfrak {D}\\right)} r \\left(P _ {i}, s _ {i}, \\mathrm {I} _ {\\mathrm {t}}, \\mathrm {T} _ {t r}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 208, + 892, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that we abstain from explicit occlusion handling in our photometric alignment, as our per-segment alignment is by design robust to pixel-level occlusion. We, however, expect our system could further be refined with explicit primitive correspondence checks.", + "bbox": [ + 496, + 258, + 893, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, to obtain the relative pose and depth scales, we minimise the photometric cost using the Adam [20] optimiser:", + "bbox": [ + 498, + 335, + 892, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{s _ {i} ^ {\\text {o p t}}, \\mathrm {T} _ {t r} ^ {\\text {o p t}} \\right\\} = \\operatorname {a r g m i n} _ {s _ {i}, \\mathrm {T} _ {t r}} E _ {\\text {p h o t o}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 382, + 890, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach trivially extends to multiple reference and target views via photometric cost summation, and we can jointly solve all depth scales and poses using the resulting cost function.", + "bbox": [ + 496, + 402, + 893, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Monocular Visual Odometry", + "text_level": 1, + "bbox": [ + 498, + 476, + 754, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate that our representation is suitable for simultaneous geometry and pose estimation, we design a novel monocular visual odometry system which operates directly on primitives. Informally, we incrementally build a local 3D map out of primitives and then track new incoming frames against this map, also in a per-primitive manner.", + "bbox": [ + 496, + 501, + 890, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visual Odometry (VO) consists of three components: initialisation, tracking, and mapping. Below, we cover each of these components. We interleave tracking and mapping in a single thread.", + "bbox": [ + 496, + 592, + 893, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt a keyframe-based approach [22], which means that the 3D map is represented by a set of posed keyframes $\\{\\mathrm{I}_{\\mathrm{kf}}^i, \\mathrm{T}_{\\mathrm{kf}}^i, \\mathcal{G}^i\\}$ . Note that each keyframe also has its estimated dense geometry $\\mathcal{G}^i$ , which results in a three-dimensional local map of the scene. We map in a sliding window fashion, with a window size of 5 keyframes. When the window is full, the earliest keyframe is popped from the window.", + "bbox": [ + 498, + 652, + 893, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initialisation. Initialisation of a VO system typically involves estimating the geometry of the first keyframe and its pose with respect to the subsequent few frames. Initialisation of a monocular incremental SfM or SLAM system is often hard, because one has to solve a chicken-and-egg problem: poses are required to reliably estimate geometry and vice versa. Therefore many VO / SLAM systems employ a wide range of heuristics in order to initialise their", + "bbox": [ + 496, + 779, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4983", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "system. However, using our new representation no special treatment is required in this case. In order to bootstrap our system, we simply create the two first keyframes and then employ the Structure-from-Motion method proposed in Sec. 3.3.1. We keep the pose of the first keyframe fixed to fix gauge freedom.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tracking. Given a map, the goal of tracking is to estimate the pose $\\mathrm{T}_{\\text{track}} \\in \\mathrm{SE}(3)$ of a new incoming frame $\\mathrm{I}_{\\text{track}}$ . We adapt Lucas-Kanade [28] tracking onto our per-primitive formulation. At each step we track a new incoming frame against the latest keyframe in the odometry sliding window.", + "bbox": [ + 76, + 186, + 467, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We solve the same photometric cost formulated in Eq. (8), but for the pose only. The latest keyframe serves as the reference frame $\\mathrm{I}_{\\mathrm{ref}}$ whereas the new incoming image is set to be the target frame $\\mathrm{I}_{\\mathrm{t}}$ . The depth scales of the latest keyframe have already been estimated at the mapping stage.", + "bbox": [ + 75, + 277, + 467, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since target images are not assumed to be primified, our tracking is essentially equivalent to the classical Lucas-Kanade method and hence could be implemented efficiently on modern hardware. Our proof-of-concept implementation does tracking at 2-3 FPS, but we expect high performance gains from employing standard machinery, such as Gauss-Newton optimisation [24].", + "bbox": [ + 75, + 354, + 467, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mapping. The mapping stage ensures geometric consistency within the keyframe sliding window of size $n$ . Concretely, our mapping refines depth scales $\\{s_j\\}$ and a pose $\\mathrm{T}_{\\mathrm{kf}}^i$ for each keyframe $\\mathrm{I}_{\\mathrm{kf}}^i$ where $0 \\leq i \\leq n$ .", + "bbox": [ + 76, + 465, + 467, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each keyframe, we define a set of target frames against which the photometric cost is being optimised with a connectivity function $\\mathcal{M}(t) = \\{\\mathrm{I}_s,\\mathrm{T}_s\\}$ . In our implementation, $\\mathcal{M}(t)$ includes temporally neighbouring keyframes $\\mathrm{I}_{\\mathrm{kf}}^{t - 1}$ and $\\mathrm{I}_{\\mathrm{kf}}^{t + 1}$ (if they exist) used as supporting frames, as well as 4 additional supplementary views, for which only the pose is being estimated. We employ these views to better constrain the geometry and observed that increasing the number of supplementary views does not induce high computational cost.", + "bbox": [ + 75, + 526, + 467, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Thus, the mapping stage is done by solving a joint photometric cost for all keyframes:", + "bbox": [ + 76, + 676, + 467, + 704 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\text {m a p p i n g}} = \\sum_ {t = 0} ^ {n} \\sum_ {\\mathrm {I} _ {s} \\in \\mathcal {M} (t)} \\sum_ {P _ {i} \\in \\mathcal {P} \\left(\\mathrm {I} _ {k f} ^ {t}\\right)} r \\left(P _ {i}, s _ {i}, \\mathrm {I} _ {s}, \\mathrm {T} _ {s} ^ {- 1} \\mathrm {T} _ {k f}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 708, + 467, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 761, + 207, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Sparse Depth Completion", + "text_level": 1, + "bbox": [ + 76, + 786, + 310, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our SuperPrimitives representation could be seamlessly applied to depth completion with no pre-training required, thereby solving it in a zero-shot manner. For each primitive $P_{i} \\in \\mathcal{P}(\\mathrm{I})$ we adjust its depth scale to minimise depth discrepancy $\\| s_i \\cdot \\mathfrak{D}_i - \\hat{D} \\|$ with given ground truth sparse depth $\\hat{D}$ across all valid depth points within the segment $\\Omega_{i}$ .", + "bbox": [ + 75, + 809, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "If a segment lacks valid depth measurements, it is discarded. For image regions not covered by any valid primitives, we generate a dense depth prediction by simply bilinearly interpolating depth values.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Many existing depth completion studies evaluate their methods in artificial scenarios, such as selecting input depth points randomly from a known ground truth. In contrast, our approach is tested using the real-world Void benchmark [57]. The benchmark provides video sequences captured with a RealSense D435i camera together with sparse metric depth measurements acquired from an external visual-inertial SfM system. This setup exhibits noise and biases that might be present in the sparse depth inputs. These measurements are already provided by the dataset itself and are therefore shared across all the methods being compared.", + "bbox": [ + 496, + 152, + 890, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our results with a recent state-of-the-art model [55], which significantly outperforms previous depth completion methods, particularly in zero-shot generalisation contexts. Other methods evaluated in [55] are also reported in the table. Our depth completion method also operates in a zero-shot manner both task-wise and dataset-wise. That means our method was not trained for the depth completion task and neither of our surface normal prediction network nor segmentation model were trained on the Void dataset.", + "bbox": [ + 496, + 333, + 890, + 482 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We focus on the most challenging \"150 points\" density setting, characterised by minimal sparse depth points per image. The depth is estimated in full $480 \\times 640$ image resolution, following [55].", + "bbox": [ + 496, + 484, + 890, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1 our method quantitatively performs on par with a recent state-of-the-art method, which is enabled by a monocular depth predictor pre-trained on a vast mixture of datasets [37, 38] and fine-tined on the Void train set. In the zero-shot setting, we outperform VI-Depth [55], which uses a DPT-Hybrid pre-trained backbone, on three out of four metrics. Note that no training was done the for depth completion task for our method. Compared to the ground truth depth maps obtained via a noisy sensor, our predictions show (Fig. 3) sharper object boundaries and are better at preserving straight lines and perpendicular structures.", + "bbox": [ + 496, + 545, + 890, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Few-View Structure-from-Motion", + "text_level": 1, + "bbox": [ + 500, + 720, + 795, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Contrasted with depth completion, where global geometry structure is roughly given, we also test our method on few view Structure-from-Motion. Given a set of unposed images $\\{\\mathrm{I}_{\\mathrm{ref}},\\mathrm{I}_s^0,\\dots ,\\mathrm{I}_s^{(n - 1)}\\}$ captured within a small timeframe, the goal is to estimate the depth of the reference image $\\mathrm{I_{ref}}$ . In that setup, the set of unposed supplementary views $\\{\\mathrm{I}_s^0,\\ldots ,\\mathrm{I}_s^{(n - 1)}\\}$ will be leveraged for multi-view geometry estimation.", + "bbox": [ + 496, + 744, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We choose the test set of the ScanNet dataset for this evaluation. In each test sequence we select every 200-th", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4984", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d14519d12162bbcf13ae7584fcf4c7af00e2af919923dd34972278d57ccf74ef.jpg", + "image_caption": [ + "Figure 3. Depth Completion on Void. We visualise the coloured unprojections of ground truth depth maps provided by a sensor (top row) and the geometry estimated by our method (bottom row). Sparse depth input points are visualised as red dots (electronic zoom-in recommended). Qualitatively, we achieve sharper geometry estimates than from a commodity depth sensor." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 893, + 343 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/05d7efb282ef91aca1dfb07b7ae8e204759884f99e37f5b3cad2e33a20b7fc4d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMAERMSEiMAEiRMSE
VOICED [57]174.04253.1487.39126.3
NLSPN [34]--143238.10
ScaffNet [58]150.65255.0880.79133.33
KBNet [56]131.54263.5466.84128.33
MonDi [27]104.96225.6048.4496.79
VI-Depth (DPT-Hybrid + NYUv2)--55.9085.20
Ours109.0204.1547.3283.40
VI-Depth* (MiDaS) [55]113.27193.3853.8684.82
VI-Depth* (DPT-Hybrid) [55]97.03167.8246.6274.67
", + "bbox": [ + 86, + 393, + 460, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Depth Completion on Void. We report four most widely used metrics for depth completion on the Void benchmark. The methods which did not use the Void dataset for training are in the top section. Methods trained on Void are marked with an asterisk*. Our method demonstrates superior performance on three out of the four metrics within the zero-shot group. It is second-best on two out of the four metrics overall.", + "bbox": [ + 75, + 521, + 468, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "frame to be the reference frame $\\mathrm{I}_{\\mathrm{ref}}$ . Supporting frames are gathered from the neighbouring frames. Then we discard the frame sets with not enough motion to remove mostly static video clips, where SfM could not be performed. This resulted in $\\sim 500$ reference frames. Note that in these experiments we use a surface normal neural network pre-trained only on the synthetic HyperSim dataset [40].", + "bbox": [ + 75, + 628, + 468, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since the multi-view depth estimation problem has scale ambiguity, we employ median-scaling to align estimated depth to the metric scale for evaluation. We report iMAE and iRMSE in Fig. 6 for our method with varying number of supplementary views. We compare our depth estimation quality against the method closest to ours, DeepV2D [49], which can also estimate depth together with supplementary frames poses. We demonstrate that our geometry quickly saturates after observing as little as 2 supporting views. Unlike DeepV2D we do not use any external tracking or initial relative pose estimation, yet still our method demonstrates", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cd37b939f9f2698aba14071cafe4e0bb7da26534354b69e3d371f8e2f075ad25.jpg", + "image_caption": [ + "Figure 4. 3-View SfM on ScanNet. We provide the visualisations of unprojected reference frame depth maps predicted by our method for few-view SfM using one reference and 2 supplementary views. Note that we used surface normal prediction network which was only pretrained on HyperSim [40] for this experiment." + ], + "image_footnote": [], + "bbox": [ + 506, + 395, + 883, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "consistent improvement over DeepV2D. Additionally, our method does not have any global prior on relative object positions, unlike DeepV2D. Our approach therefore could be used as an VO / SLAM initialisation mechanism, estimating joint relative poses and geometry.", + "bbox": [ + 496, + 729, + 893, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Monocular Visual Odometry", + "text_level": 1, + "bbox": [ + 498, + 815, + 756, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Monocular Visual Odometry requires both accurate pose and geometry estimation to successfully track camera motion across long trajectories. Minor pose estimation inaccuracies can accumulate over time resulting in what is called", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4985", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d5cc49e732b8b85c274699b0a714af00cc62af10092bcaeed63bff256c49a9c4.jpg", + "image_caption": [ + "Figure 5. TUM Reconstruction Results. Examples of reconstructions produced by our monocular VO system on the TUM dataset. Each image shows a coloured point cloud of the geometry estimated on an odometry keyframe." + ], + "image_footnote": [], + "bbox": [ + 83, + 92, + 361, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b5b1216a0518c73a217ad202abf211c3fcdf98c0a102b61b31bd8d336b4f4ad8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 92, + 635, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9ad305d9bf382664072aaaf3b4b98eafa66b75052d7d24b38b5f8f53d7586f6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 92, + 885, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bfce09d04e29f96cfc14063379b589fc2ffed88683671bca5b31e22c65e2f0fe.jpg", + "image_caption": [ + "Figure 6. Few-View SfM Depth Estimation Quality. We evaluate the quality of our depth estimation method from an unstructured set of images on the ScanNet dataset. On the $x$ -axis the number of supporting views is shown, while on the $y$ -axis the corresponding depth reconstruction quality metric is reported. We show that the quality of our depth estimation quickly saturates and consistently outperforms its closest competitor, DeepV2D." + ], + "image_footnote": [], + "bbox": [ + 83, + 309, + 272, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ba15942e098593b50d20088664b42e89400cb9c1ad254bbe2496efa5955f7088.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 309, + 467, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "drift. With geometry being incorrectly estimated, accurate pose tracking becomes impossible, and vice versa. Our SuperPrimitive representation allows estimating both pose and geometry, enabling us to build a simple monocular VO system which performs better even in hard conditions.", + "bbox": [ + 75, + 566, + 468, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate our monocular odometry on the TUM RGB-D [46] dataset. The dataset was captured with a handheld camera in indoor scenes. It is renowned for being incredibly challenging (especially for dense reconstruction systems), due to motion blur, rolling shutter artefacts, and abundance of pure rotational motion. We show that thanks to the strong priors encapsulated in our SuperPrimitives, our simple monocular odometry system can handle the TUM dataset without any special treatment (e.g. our method does not involve any special motion blur handling).", + "bbox": [ + 75, + 643, + 468, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare against other VO systems that do not have global bundle adjustment, following the protocol of [10] and evaluate on 8 sequences from the Friedburg 1 split. We use only RGB images as the input to our system and down-sample them to $120 \\times 160$ for efficiency purposes. Since our method is purely monocular, the estimated trajectory lacks global scale and we first use $\\mathrm{Sim}(3)$ alignment to the", + "bbox": [ + 75, + 795, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/82c20203c750d764b3732275a1859d51ed8ed9f30ae2793ee1135b812a489a6c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Seq.TartanVO [53]DeepV2D [49]DeepFactors [8]DepthCov [10]Ours
3600.1780.2430.1590.1280.173
desk0.1250.1660.1700.0560.085
desk20.1220.3790.2530.0480.108
plant0.2970.2030.3050.2610.153
room0.3330.2460.3640.2570.363
rpy0.0490.1050.0430.0520.055
teddy0.3390.3160.6010.4750.253
xyz0.0620.0640.0350.0560.036
mean0.1880.2150.2410.1670.153
", + "bbox": [ + 504, + 309, + 895, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Trajectory Estimation Error on TUM. Average Trajectory Error (ATE) is compared against other monocular odometry systems on the TUM Friedburg 1 split. The best and second best results are highlighted in bold and underscored correspondingly. Our method outperforms others in terms of the ATE averaged across all trajectories.", + "bbox": [ + 498, + 465, + 892, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ground truth scale, following standard practice [4]. Tab. 2 shows that, despite the simplicity of our odometry system, it outperforms all other methods in terms of Average Trajectory Error (ATE) [46], averaged across all trajectories. Additionally, our VO is either the best or second best on five out of eight sequences. Besides quantitative evaluation, we also demonstrate reconstruction results in Fig. 5.", + "bbox": [ + 498, + 556, + 892, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 675, + 617, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a new representation, SuperPrimitive, which demonstrates how recent advances in building strong single image priors could be incorporated into pose and dense geometry estimation problems. We show that incorporating these priors alleviates the need for sophisticated handcrafted heuristics and paves the way into monocular reconstruction with relative ease.", + "bbox": [ + 498, + 700, + 890, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. Research presented in this paper was supported by Dyson Technology Ltd. The authors would like to thank Hidenobu Matsuki, Eric Dexheimer, Riku Murai, and other members of the Dyson Robotics Lab for countless fruitful discussions.", + "bbox": [ + 498, + 816, + 892, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4986", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Gwangbin Bae, Ignas Budvytis, and Roberto Cipolla. Estimating and exploiting the aleatoric uncertainty in surface normal estimation. In International Conference on Computer Vision (ICCV), 2021. 3, 4", + "[2] M. Bloesch, J. Czarnowski, R. Clark, S. Leutenegger, and A. J. Davison. CodeSLAM — learning a compact, optimisable representation for dense visual SLAM. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1, 2, 12", + "[3] Aljaž Božić, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. Transformerfusion: Monocular rgb scene reconstruction using transformers. Proc. Neural Information Processing Systems (NeurIPS), 2021. 2", + "[4] Carlos Campos, Richard Elvira, Juan J. Gomez, Jose M. M. Montiel, and Juan D. Tardos. ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Transactions on Robotics, 2021. 8", + "[5] Xu Cao, Hiroaki Santo, Boxin Shi, Fumio Okura, and Yasuyuki Matsushita. Bilateral normal integration. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 4", + "[6] Alejo Concha and Javier Civera. Using superpixels in monocular SLAM. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2014. 2", + "[7] G. Cross and A. Zisserman. Quadric Reconstruction from Dual-Space Geometry. In Proceedings of the International Conference on Computer Vision (ICCV), 1998. 2", + "[8] J. Czarnowski, T. Laidlow, R. Clark, and A. J. Davison. Deepfactors: Real-time probabilistic dense monocular SLAM. In IEEE Robotics and Automation Letters, 2020. 8, 12", + "[9] A. J. Davison, N. D. Molton, I. Reid, and O. Stasse. MonoSLAM: Real-Time Single Camera SLAM. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2007. 2", + "[10] Eric Dexheimer and Andrew J. Davison. Learning a depth covariance function. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 8", + "[11] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In Neural Information Processing Systems (NeurIPS), 2014. 1", + "[12] David Eigen, Christian Puhrsch, and Rob Fergus. Depth Map Prediction from a Single Image using a Multi-Scale Deep Network. In Neural Information Processing Systems (NeurIPS), 2014. 2", + "[13] Jakob Engel, Thomas Schoeps, and Daniel Cremers. LSD-SLAM: Large-scale direct monocular SLAM. In Proceedings of the European Conference on Computer Vision (ECCV), 2014. 3", + "[14] J. Engel, V. Koltun, and D. Cremers. Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2017. 3, 5", + "[15] Jiahui Fu, Yilun Du, Kurran Singh, Joshua B Tenenbaum, and John J Leonard. Neuse: Neural se (3)-equivariant em" + ], + "bbox": [ + 76, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "bedding for consistent spatial understanding with objects. In Proceedings of Robotics: Science and Systems (RSS), 2023. 2", + "[16] Y. Furukawa, B. Curless, S. M. Seitz, and R. Szeliski. Towards Internet-scale multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2010. 2", + "[17] D. Gallup, J.-M. Frahm, and M. Pollefeys. Piecewise planar and non-planar stereo for urban scene reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2010. 2", + "[18] Magnus R. Hestenes and Eduard Stiefel. Methods of conjugate gradients for solving linear systems. Journal of research of the National Bureau of Standards, 1952. 4", + "[19] M Kaess. Simultaneous localization and mapping with infinite planes. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2015. 2", + "[20] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015. 5", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3, 4", + "[22] G. Klein and D. W. Murray. Parallel Tracking and Mapping for Small AR Workspaces. In Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR), 2007. 2, 5", + "[23] T. Laidlow and A. J. Davison. Simultaneous localisation and mapping with quadric surfaces. In International Conference on 3D Vision (3DV), 2022. 2", + "[24] Stefan Leutenegger, Simon Lynen, Michael Bosse, Roland Siegwart, and Paul Furgale. Keyframe-based visual-inertial odometry using nonlinear optimization. International Journal of Robotics Research (IJRR), 2014. 6", + "[25] Chen Liu, Jimei Yang, Duygu Ceylan, Ersin Yumer, and Yasutaka Furukawa. Planenet: Piece-wise planar reconstruction from a single rgb image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[26] Chen Liu, Kihwan Kim, Jinwei Gu, Yasutaka Furukawa, and Jan Kautz. Planercnn: 3d plane detection and reconstruction from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[27] Tianlin Liu, Parth T. Agrawal, Allison Chen, Byung-Woo Hong, and A. Wong. Monitored distillation for positive congruent depth completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 7", + "[28] B. D. Lucas and T. Kanade. An Iterative Image Registration Technique with an Application to Stereo Vision. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), 1981. 6", + "[29] J. McCormac, R. Clark, M. Bloesch, A. J. Davison, and S. Leutenegger. Fusion++: volumetric object-level slam. In Proceedings of the International Conference on 3D Vision (3DV), 2018. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4987", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] R. A. Newcombe, S. Lovegrove, and A. J. Davison. DTAM: Dense Tracking and Mapping in Real-Time. In Proceedings of the International Conference on Computer Vision (ICCV), 2011. 1, 2, 5", + "[31] Lachlan Nicholson, Michael Milford, and Niko Sunderhauf. Quadricslam: Dual quadrics from object detections as landmarks in object-oriented slam. IEEE Robotics and Automation Letters, 2019. 2", + "[32] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[33] Ryosuke Okuta, Yuya Unno, Daisuke Nishino, Shohei Hido, and Crissman Loomis. Copy: A numpy-compatible library for nvidiagpu calculations. In Proceedings of Workshop on Machine Learning Systems (LearningSys) in The Thirty-first Annual Conference on Neural Information Processing Systems (NeurIPS), 2017. 13", + "[34] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 7", + "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. In Neural Information Processing Systems (NeurIPS), 2019. 13", + "[36] Albert Pumarola, Alexander Vakhitov, Antonio Agudo, Alberto Sanfeliu, and Francese Moreno-Noguer. Pl-slam: Real-time monocular visual slam with points and lines. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2017. 2", + "[37] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2, 6, 12", + "[38] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2022. 2, 6", + "[39] Xiaofeng Ren and Jitendra Malik. Learning a classification model for segmentation. In Proceedings of the International Conference on Computer Vision (ICCV), 2003. 3", + "[40] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 7", + "[41] R. F. Salas-Moreno, R. A. Newcombe, H. Strasdat, P. H. J. Kelly, and A. J. Davison. SLAM++: Simultaneous Localisation and Mapping at the Level of Objects. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2013. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Mohamed Sayed, John Gibson, Jamie Watson, Victor Prisacariu, Michael Firman, and Clément Godard. Simplerecon: 3d reconstruction without 3d convolutions. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2", + "[43] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2", + "[44] Jonathan Shade, Steven Gortler, Li-wei He, and Richard Szeliski. Layered depth images. In Proceedings of the 25th annual conference on Computer graphics and interactive techniques, 1998. 12", + "[45] Jingjia Shi, Shuaifeng Zhi, and Kai Xu. Planerectr: Unified query learning for 3d plane recovery from a single view. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2", + "[46] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A Benchmark for the Evaluation of RGB-D SLAM Systems. In Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS), 2012. 8, 13", + "[47] E. Sucar, K. Wada, and A. J. Davison. NodeSLAM: Neural object descriptors for multi-view shape reconstruction. In Proceedings of the International Conference on 3D Vision (3DV), 2020. 2", + "[48] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. NeuralRecon: Real-time coherent 3D reconstruction from monocular video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[49] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. In Proceedings of the International Conference on Learning Representations (ICLR), 2020. 7, 8", + "[50] Alexander Vakhitov, Jan Funke, and Francesc Moreno-Noguer. Accurate and linear time pose estimation from points and lines. In Proceedings of the European Conference on Computer Vision (ECCV), 2016. 2", + "[51] Tom van Dijk and Guido C.H.E. de Croon. How do neural networks see depth in single images? In Proceedings of the International Conference on Computer Vision (ICCV), 2019.", + "[52] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1", + "[53] Wenshan Wang, Yaoyu Hu, and Sebastian Scherer. TartanVO: A Generalizable Learning-based VO. In Conference on Robot Learning (CoRL), 2020. 8", + "[54] C. S. Weerasekera, Y. Latif, R. Garg, and I. Reid. Dense monocular reconstruction using surface normals. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2017. 2", + "[55] Wofk, Diana and Ranftl, René and Müller, Matthias and Koltun, Vladlen. Monocular Visual-Inertial Depth Estimation. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2023. 6, 7, 12" + ], + "bbox": [ + 501, + 92, + 892, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4988", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Alex Wong and Stefano Soatto. Unsupervised depth completion with calibrated backprojection layers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7", + "[57] Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters, 2020. 6, 7", + "[58] Alex Wong, Safa Cicek, and Stefano Soatto. Learning topology from synthetic data for unsupervised depth completion. IEEE Robotics and Automation Letters, 2021. 7", + "[59] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In Proceedings of SIGGRAPH, 2018. 12" + ], + "bbox": [ + 78, + 90, + 468, + 301 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4989", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/556fb01c-1445-4c08-b116-0426ba52ea17_model.json b/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/556fb01c-1445-4c08-b116-0426ba52ea17_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dcbc8888c1b3fad2d905953b7d599c3d997b3c89 --- /dev/null +++ b/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/556fb01c-1445-4c08-b116-0426ba52ea17_model.json @@ -0,0 +1,2389 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.19, + 0.131, + 0.78, + 0.152 + ], + "angle": 0, + "content": "SuperPrimitive: Scene Reconstruction at a Primitive Level" + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.181, + 0.699, + 0.217 + ], + "angle": 0, + "content": "Kirill Mazur, Gwangbin Bae, Andrew J. Davison Dyson Robotics Lab, Imperial College London" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.22, + 0.678, + 0.234 + ], + "angle": 0, + "content": "{k.mazur21, g.bae, a.davison}@imperial.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.471, + 0.436 + ], + "angle": 0, + "content": "Joint camera pose and dense geometry estimation from a set of images or a monocular video remains a challenging problem due to its computational complexity and inherent visual ambiguities. Most dense incremental reconstruction systems operate directly on image pixels and solve for their 3D positions using multi-view geometry cues. Such pixel-level approaches suffer from ambiguities or violations of multi-view consistency (e.g. caused by textureless or specular surfaces)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.437, + 0.47, + 0.557 + ], + "angle": 0, + "content": "We address this issue with a new image representation which we call a SuperPrimitive. SuperPrimitives are obtained by splitting images into semantically correlated local regions and enhancing them with estimated surface normal directions, both of which are predicted by state-of-the-art single image neural networks. This provides a local geometry estimate per SuperPrimitive, while their relative positions are adjusted based on multi-view observations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.47, + 0.633 + ], + "angle": 0, + "content": "We demonstrate the versatility of our new representation by addressing three 3D reconstruction tasks: depth completion, few-view structure from motion, and monocular dense visual odometry. Project page: https://makezur.github.io/SuperPrimitive/" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.649, + 0.21, + 0.665 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.764 + ], + "angle": 0, + "content": "Enriching monocular incremental reconstruction with prior world knowledge is essential for resolving visual ambiguities. This issue is particularly prevalent in scenarios with scarce data observations available: a notable example would be monocular visual SLAM, where images are being streamed from a camera into the system in real-time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.869 + ], + "angle": 0, + "content": "When a monocular vision system encounters a new scene region, it must estimate the region's geometry based on a very limited number of observations. Without this, continuous camera motion tracking would not be possible. Once the scene region is thoroughly observed, the initial geometry estimate should be refined to better explain the multiview information." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "This naturally leads to a question: what sort of priors are effective in both providing reliable initial geometry es" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.274, + 0.871, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.593, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Figure 1. Multi-View Geometry with SuperPrimitives. SuperPrimitives are extracted from an input frame by dividing it into image segments equipped with estimated surface normal directions (bottom-left). Each SuperPrimitive induces a dense reconstruction within the corresponding image segment up to a priori unknown scale. Different possible reconstructions are shown in light blue. The scales are then jointly optimised together with a relative camera pose to fit multi-view photometric constraints (visualised in green and red). The resulting dense reconstruction of the reference frame is shown in the top." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.892, + 0.852 + ], + "angle": 0, + "content": "timates and supporting multi-view consistency? Geometric priors generally fall into one of two categories: local and global. Local priors, such as smoothness assumption [30] or surface normal regularisation [52], impose additional constraints within a small neighbourhood. Global priors, on the other hand, aim to impose constraints on a larger scale, such as depth prediction [2, 11]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our key observation is that some of the geometrical correlations are more reliable than the others, and therefore could be safely \"locked in\" together within local regions" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4979" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "based on a single-view prediction. Points belonging to the same rigid body are strongly correlated, making it unnecessary to determine their depth independently. In contrast, distinct and unrelated objects can be placed arbitrarily in the scene. As the number of objects increases, learning a reliable global prior on their relative positions becomes an increasingly complex problem." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.47, + 0.333 + ], + "angle": 0, + "content": "In this work, we show that purely local but strong priors are enough to achieve excellent performance across a variety of geometric vision tasks. For that purpose, we introduce a novel representation, SuperPrimitives. A Super-Primitive represents a local image segment, coupled with a dense shape estimate which is determined up to a scale factor. The scale factor can be further adjusted based on information observed from other views or additional measurements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.468, + 0.456 + ], + "angle": 0, + "content": "We show that SuperPrimitives can be efficiently constructed using a front-end which consists of two single-image neural networks, extracting image segmentation and surface normal prediction. Effectively, our neural front-end predicts whether adjacent pixels belong to the same geometrical entity through image segmentation and estimates a surface normal at this point, thereby providing an infinitesimal geometry estimate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.468, + 0.532 + ], + "angle": 0, + "content": "We delegate global, scene-level alignment to our streamlined multi-view, iterative, optimisation-based back-end. The resulting front-end / back-end tandem combines the flexibility of multi-view based optimisation methods with the observation efficiency common in prior-driven systems." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.468, + 0.577 + ], + "angle": 0, + "content": "Our new representation showcases its versatility in three key applications, where single-view ambiguity is resolved via additional measurements or viewpoints:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.578, + 0.468, + 0.623 + ], + "angle": 0, + "content": "- Firstly, it adeptly handles zero-shot depth completion tasks in real-world scenarios, matching the performance of state-of-the-art methods tailored for depth completion;" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.624, + 0.468, + 0.668 + ], + "angle": 0, + "content": "- Secondly, it facilitates joint pose and depth estimation using a limited set of unstructured images, surpassing its nearest competitor even in the absence of global priors;" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.669, + 0.468, + 0.713 + ], + "angle": 0, + "content": "- Thirdly, our method outperforms previous monocular visual odometry systems on the challenging TUM dataset, and exhibits robustness across various domains." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.578, + 0.468, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.724, + 0.218, + 0.739 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Monocular Reconstruction. Both offline monocular reconstruction systems, such as COLMAP [43] and [16], or online systems, such as MonoSLAM [9] and DSO [22] track, filter, and reconstruct only well-constrained points with high and reliable photometric information. This often involves fine-grained and sophisticated point management to reliably resolve visual ambiguities and filter unreliable visual observations. DTAM [30] demonstrated feasibility of incremental dense reconstruction in the monocular scenario. DTAM employed a hand-crafted local smoothness" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "prior to handle regions with poor texture. Subsequently, other local priors [32, 54] have been extensively explored to regularise multi-view geometry estimation problems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.892, + 0.366 + ], + "angle": 0, + "content": "Global Priors. We are interested in exploring the space of possible geometric priors from single-view networks. Depth prediction [12, 37, 38] is the most obvious choice but leads to a rigid per-image reconstruction which is difficult to feed into multi-view optimisation. In the recent years, deep-learning based approaches sought to replace explicit multi-view geometry estimating with learning-based methods [3, 42, 48]. These methods, however, assume known poses and are, therefore, not suitable for joint pose and geometry estimation. Notably, CodeSLAM [2] introduced depth prediction conditioned on latent codes, which are then optimised to achieve cross-view consistency. However, CodeSLAM still struggles with out-of-domain data, as depth prediction networks are known to struggle with generalisation [51]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.892, + 0.473 + ], + "angle": 0, + "content": "Higher-Level Mapping. Introducing parametric primitives, such as lines [36, 50], planes [6, 17, 19, 25, 26, 45] or even high-order algebraic shapes [7, 23, 31] to better constrain multi-view geometry problems have been thoroughly explored over the last few decades. They all however use assumptions which may not often hold for all 3D scenes, especially for dense reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Besides parametric algebraic primitives object-level mapping has been explored in the last decade. SLAM++ [41] represented a map with a set of CAD models retrieved from an existing database and tracked camera position against these models. This method, however, could not be applied in a variety of settings due to the limited size of the CAD databases. Even objects of the same class can vary in their geometric appearance from instance to instance. This was later approached by [15, 47], who also represent their map via a set of objects, but learn a latent space of possible geometric variations per object class. Fusion++ [29] extended this approach into a more versatile per object depth fusion using an RGB-D sensor and mask proposals from a pre-trained neural network. These methods rely on an RGB-D inputs and assume a pre-defined set of object classes, therefore lacking generality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.59, + 0.74 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Firstly, we introduce the concept of a SuperPrimitive and explain how an image is processed into a set of its SuperPrimitives, referring to this part of our method as the front-end. In the second part of this section, we describe how multi-view geometry problems can be reformulated at the level of SuperPrimitives instead of pixels. We refer to this stage of SuperPrimitive alignment as the back-end." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The core of our method proposes splitting a given image into a set of (possibly overlapping) minimal segments, image regions which are likely to have strongly correlated" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4980" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.262, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.09, + 0.885, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.316, + 0.895, + 0.386 + ], + "angle": 0, + "content": "Figure 2. SuperPrimitives Extraction. (left) Our front-end processor extracts SuperPrimitives from an image by dividing it into a set of image regions with surface normal directions estimated for each image pixel within the segment. (right) Highlighted SuperPrimitives extracted from the image are visualised by showing their estimated normal and colour maps side by side. Note some of them are scaled either up or down for better viewing. While some of the SuperPrimitives are akin to object-level segmentation, the others tend to represent more low-level image segments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.47, + 0.44 + ], + "angle": 0, + "content": "geometry. We repurpose the recent state-of-the-art Segment Anything (SAM) [21] model into predicting these minimal segments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.47, + 0.546 + ], + "angle": 0, + "content": "Our key idea is to estimate local geometry within each segment from a single view, while leaving the relative positioning of the segments to be estimated via multi-view photometric consistency optimisation. We refer to these geometrically enhanced minimal image segments as Super-Primitives, since they are inspired by both superpixels [39] and geometric 3D primitives." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.682 + ], + "angle": 0, + "content": "For per-segment local geometry estimation we employ an off-the-shelf surface normal prediction network [1], to estimate infinitesimal geometry for each image pixel. The surface normals within an image segment can be used to estimate its depth via simple integration up to a scale factor. We set these scale factors — depth scales — to be optimisable parameters, which are either optimised via multi-view cues or explicit depth measurements for the depth completion experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.47, + 0.743 + ], + "angle": 0, + "content": "Thus, our method combines the strong priors provided by state-of-the-art neural networks in the front-end with the flexibility and consistency offered by multi-view optimisation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.747, + 0.21, + 0.761 + ], + "angle": 0, + "content": "3.1. Conventions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Unless stated otherwise, we use lowercase letters for scalar values, bold letters for vectors, and uppercase letters for matrices. We consider images \\(\\mathbf{I} \\in \\mathbb{R}^{3 \\times h \\times w}\\) of height \\(h\\) and width \\(w\\) captured by a camera with a known calibration matrix \\(\\mathbf{K} = \\left( \\begin{array}{ccc} f_u & 0 & c_u \\\\ 0 & f_v & c_v \\\\ 0 & 0 & 1 \\end{array} \\right) \\in \\mathbb{R}^{3 \\times 3}\\). Image pixels are parameterised as \\(\\mathbf{u} = (u, v) \\in [0, h - 1] \\times [0, w - 1]\\). Given a per pixel depth function \\(z(\\mathbf{u})\\): \\(\\mathbf{I} \\to \\mathbb{R}^+\\), we define image's unprojection onto the 3D space to be \\(\\pi^{-1}(\\mathbf{u}) = z(\\mathbf{u})\\mathbf{K}^{-1}\\dot{\\mathbf{u}}\\)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.892, + 0.456 + ], + "angle": 0, + "content": "where dot is a homogenisation operator \\(\\dot{\\pmb{p}} = (\\pmb{p},1)\\). Conversely, for a set of 3D points \\(\\pmb{x} = (x,y,z) \\in \\mathbb{R}^3\\) we define the projection \\(\\pi(\\pmb{x}) = \\gamma\\left(\\frac{1}{z} K\\pmb{x}\\right)\\), where the \\(\\gamma\\) function drops the last coordinate \\(\\gamma((x,y,z)^T) = (x,y)^T\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.547 + ], + "angle": 0, + "content": "We represent camera poses as matrices \\(\\mathrm{T}_{WC} \\in \\mathrm{SE}(3)\\) in the camera-to-world coordinates unless stated otherwise. During pose optimisation we store linearised pose increments \\(\\xi \\in \\mathfrak{se}(3)\\) as Lie algebra elements. These increments are used then to update current camera pose estimates, see [13, 14] for more detail." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.557, + 0.662, + 0.573 + ], + "angle": 0, + "content": "3.2. SuperPrimitives" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.717 + ], + "angle": 0, + "content": "We split the input image I into a set of possibly overlapping connected regions (segments), each of which is equipped with its local 3D geometry. More formally, a SuperPrimitive \\( P = (\\Omega, \\mathfrak{D}) \\) is a connected image region \\( \\Omega \\subseteq I \\) which is also equipped with an unscaled depth map \\( \\mathfrak{D} : \\Omega \\to [0, \\infty) \\) for each pixel within the region \\( \\Omega \\). Here, unscaled means that each SuperPrimitive's depth \\( \\mathfrak{D} \\) differs from its ground truth depth \\( D \\) by an a priori unknown scalar. In other words, there exists a single scalar \\( s \\) such that:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.72, + 0.892, + 0.737 + ], + "angle": 0, + "content": "\\[\nD = s \\cdot \\mathfrak {D} \\quad \\text {f o r a l l p i x e l s} \\mathrm {p} \\in \\Omega \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.741, + 0.892, + 0.786 + ], + "angle": 0, + "content": "We say that an image \\( \\mathrm{I} \\) is primified into super-primitives \\( \\mathcal{P}(\\mathrm{I}) = \\{P_i = (\\Omega_i, \\mathfrak{D}_i)\\} \\) if it has \\( n \\), possibly overlapping primitives that lie within the image \\( \\mathrm{I} \\), i.e. \\( \\bigcup \\Omega_i \\subseteq \\mathrm{I} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.787, + 0.892, + 0.818 + ], + "angle": 0, + "content": "For brevity we use the words primitive and SuperPrimitive interchangeably throughout the rest of the paper." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Representing Geometry with SuperPrimitives. A set of primitives \\(\\mathcal{P}\\) itself is not enough to extract dense image geometry due to scale/depth ambiguity. Therefore, we introduce the concept of optimisable depth scales which anchor the correct depth scale for each primitive." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4981" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.198 + ], + "angle": 0, + "content": "Given a scalar depth scale \\(s\\) for a super-primitive \\(P\\), one can infer its depth as \\(D(p) = s \\cdot \\mathfrak{D}(p)\\) for every point \\(p \\in \\Omega\\). In practice, we employ a log-depth representation to represent depth throughout the whole system. This means that we store optimisable log-depth scales \\(\\log s\\) and unscaled log-depth \\(\\log \\mathfrak{D}\\), and so depth inference reduces to a simple shift operation \\(\\log D = \\log s + \\log \\mathfrak{D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.468, + 0.289 + ], + "angle": 0, + "content": "In contrast to other dense methods, we estimate the dense geometry \\(\\mathcal{G}\\) of the image in the form of a point cloud rather than a depth image. This choice is driven by the fact that multiple primitives' supporting segments \\(\\Omega_{i}\\) and \\(\\Omega_{j}\\) might overlap, which would lead to different depth estimates within their intersection \\(\\Omega_{i} \\cap \\Omega_{j}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.29, + 0.468, + 0.32 + ], + "angle": 0, + "content": "Given a set of depth scaled primitives \\(\\mathbb{P} = \\{(s_i, P_i)\\}\\), the geometry \\(\\mathcal{G}\\) of the image \\(I\\) can be estimated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.335, + 0.468, + 0.366 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} := \\bigcup_ {i} \\pi^ {- 1} \\left(s _ {i} \\cdot \\mathfrak {D} _ {i}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.371, + 0.468, + 0.402 + ], + "angle": 0, + "content": "To obtain estimated depth \\(\\hat{D}\\), we average depth values along the corresponding camera rays." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.468, + 0.56 + ], + "angle": 0, + "content": "Converting an Image into SuperPrimitives. Next we explain how SuperPrimitives are obtained in practice. An input image \\(I\\) is first split into a set of minimal image segments \\(\\Omega_{i}\\) produced by an image segmentation model. We use the Segment Anything [21] model in this work, since it captures low-level semantic correlations highly accurately. Given the extracted segments, we independently estimate the local geometry \\(\\mathfrak{D}_i\\) within each segment by integrating surface normals predicted by another state-of-the-art neural network [1] within this region." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.468, + 0.682 + ], + "angle": 0, + "content": "Note that this approach is premised on the assumption of geometry being continuous within each predicted image segment \\(\\Omega_{i}\\), hence its unscaled depth values could be obtained by integrating its surface normals. Although this is not guaranteed a priori, we observed that a more fine-grained mask selection yields a compelling correlation with geometrical continuity. We discuss this in more detail in the following section." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Image Segment Retrieval. An ideal neural network which estimates image segments \\(\\{\\Omega_i\\}\\) should predict regions of geometric continuity, where depth could be obtained via surface normal integration. However, to the best of our knowledge, such a network does not exist. Our approach aims to approximate this behaviour by utilising the Segment Anything model, coupled with a specialised mask selection process. Specifically, for each SAM query, we strive to select the smallest predicted mask surrounding that point. While such over-segmentation could potentially increase the dimensionality of the multi-view optimisation problem, under-segmentation might lead to incorrect geometry estimate within the primitive. Since we only adjust the scale of a primitive after the normal integration is com" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "pleted, any incorrect geometry estimates within a primitive cannot be compensated at this stage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.891, + 0.272 + ], + "angle": 0, + "content": "The segmentation model employed in this work outputs three binary segmentation masks given a query point \\( q \\in \\mathbf{I} \\). For each query \\( q \\) we first filter predicted masks using the post-processing introduced in [21], such as stability filtering and Non-Maximum-Suppression (NMS). Even though our method allows redundant regions, we employed NMS to remove similar segments and save compute in the image alignment stage. With filtering done, the mask with the smallest area is selected. If the filtered set is empty, we discard the query point." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.891, + 0.363 + ], + "angle": 0, + "content": "We query the SAM backbone feature extractor once per image. Then we first sample 300 query points randomly across the image, followed by the filtering discussed above. Then the image coverage mask is calculated and an additional 100 mask query points are actively sampled in the uncovered regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.891, + 0.456 + ], + "angle": 0, + "content": "Normal Integration. To estimate each segment's local geometry, we first pass the image through a surface normal estimation network. This network effectively predicts the derivative of desired depth values. Then, we \"integrate\" these surface normal vectors within each image segment to obtain its unscaled depth map." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.891, + 0.531 + ], + "angle": 0, + "content": "For each image pixel \\(\\boldsymbol{u} = (u,v)^T \\in \\mathrm{I}\\) we estimate its surface normal vector \\(\\boldsymbol{n} = (n_x, n_y, n_z)^T\\) with a pretrained state-of-the-art CNN [1]. As was demonstrated in [5], the log-depth \\(\\tilde{z} = \\log(z(\\boldsymbol{u}))\\) satisfies the following PDEs within a segment \\(\\Omega\\):" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.535, + 0.891, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\tilde {n} _ {z} \\partial_ {u} \\tilde {z} + n _ {x} = 0 \\quad \\text {a n d} \\quad \\tilde {n} _ {z} \\partial_ {v} \\tilde {z} + n _ {y} = 0 \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.89, + 0.599 + ], + "angle": 0, + "content": "where \\(\\tilde{n}_z = n_x(u - c_u) + n_y(v - c_v) + n_zf\\). The depth values within the segment \\(\\Omega\\) can therefore be obtained via minimising the following functional:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.601, + 0.891, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\min _ {z} \\iint_ {\\Omega} \\left(\\tilde {n} _ {z} \\partial_ {u} \\tilde {z} + n _ {x}\\right) ^ {2} + \\left(\\tilde {n} _ {z} \\partial_ {u} \\tilde {z} + n _ {y}\\right) ^ {2} d u d v \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.891, + 0.718 + ], + "angle": 0, + "content": "Note that this leads to a family of solutions \\(\\tilde{z} + \\text{Const}\\) which differ by a shift constant. This is due to the fact that only partial derivatives of \\(\\tilde{z}\\) are used in the functional. That means that the actual depth will be estimated up to scale, since \\(z = \\exp(\\tilde{z})\\) by definition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.838 + ], + "angle": 0, + "content": "This ambiguity prompted us to introduce the notion of depth scales, to allow each segment to be adjusted towards its true depth. We implemented batched normal integration, efficiently solving all optimisation problems as a single sparse linear system using conjugate gradient method [18]. This implementation enables the integration of approximately 100 to 300 segments within a total time frame of \\(\\sim 100\\mathrm{ms}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.795, + 0.864 + ], + "angle": 0, + "content": "3.3. Primitive-based Image Alignment" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Now we explain how our new representation can be used as a building block for dense multi-view geometry and pose" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4982" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "estimation, combining an optimisation-based mindset with learned single view priors. To make optimisation computationally tractable, we design our image alignment to fully abstract away any knowledge about the neural networks involved in the image primification stage. Hence, even though there has been early evidence that SAM is independently capable of establishing (coarse) mask correspondences across neighbouring frames, we rely on photometric information only during the image alignment stage. We believe that more semantic primitive association would be \"heavier\" and coarser since it would not provide per pixel correspondences, and would therefore not be a good fit to be used in a multi-view optimisation loop. However, we speculate that such a method could be employed to further enhance our system, e.g. to improve occlusion handling, particularly as these models become more computationally affordable." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.348, + 0.373, + 0.364 + ], + "angle": 0, + "content": "3.3.1 Two-view SfM on SuperPrimitives" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.473, + 0.644 + ], + "angle": 0, + "content": "For the sake of clarity, we formulate our method in the simplest case of two frames observing the scene, although it could be trivially extended into a setting with a higher number of views. At the core of our Structure-from-Motion (SfM) approach lies a per-primitive photometric alignment, which contrasts with widely accepted per-pixel photometric alignment [14, 30] techniques. Informally speaking, our proposition is to treat each image primitive as a \"rigid\" piece, which is only allowed to be scaled up or down. This degree of freedom is due to the scale ambiguity present for each primitive. Thus, instead of estimating a depth value per pixel, we only estimate a depth-scale per primitive, which is illustrated in Fig. 1. This greatly reduces the dimensionality of the optimisation problem, especially in the case of an unknown relative pose. Note that our method does not require the target image to be primified nor does it not require any pre-established correspondences (e.g. primitive to primitive)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.648, + 0.471, + 0.725 + ], + "angle": 0, + "content": "We assume a primified reference image \\( \\mathrm{I}_{\\mathrm{ref}} \\) with its set of primitives \\( \\mathcal{P}(\\mathrm{I}) = \\{(\\Omega_i, \\mathfrak{D}_i)\\} \\). Given an unposed target image \\( \\mathrm{I}_{\\mathrm{t}} \\), our goal is to jointly estimate its relative pose \\( \\mathrm{T}_{tr} \\) with respect to the reference image \\( \\mathrm{I}_{\\mathrm{ref}} \\) as well as the dense geometry \\( \\mathcal{G}_{\\mathrm{ref}} = \\bigcup_{i} \\pi^{-1}(s_i \\cdot \\mathfrak{D}_i) \\) of the reference frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.472, + 0.848 + ], + "angle": 0, + "content": "While unscaled depths \\(\\mathfrak{D}_i\\) are given by a per-image preprocessor after the primification of the reference image, the set of depth scales \\(s_i\\) are yet to be estimated. In our case, we jointly estimate both depth scales \\(s_i\\) and a relative image pose \\(\\mathrm{T}_{tr}\\) by solving a photometric consistency optimisation problem. First, we warp each depth-scaled primitive \\((s_i, P_i)\\) from the reference frame \\(\\mathrm{I}_{\\mathrm{ref}}\\) into the target frame \\(\\mathrm{I}_{\\mathrm{t}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.85, + 0.47, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\hat {P} _ {i} [ \\boldsymbol {u} ] = \\pi \\left(\\mathrm {T} _ {t r} \\pi^ {- 1} \\left(\\boldsymbol {u}, s _ {i} \\cdot \\mathfrak {D} _ {i}\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Then a per-segment photometric residual for the primitive \\(P_{i} = (\\Omega_{i},\\mathfrak{D}_{i})\\) is defined by averaging all photometric re" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.81, + 0.108 + ], + "angle": 0, + "content": "projection \\(\\ell^1\\) errors across every pixel \\(\\pmb{u} \\in \\Omega_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.118, + 0.893, + 0.157 + ], + "angle": 0, + "content": "\\[\nr \\left(P _ {i}, s _ {i}, I _ {\\mathrm {t}}, T _ {t r}\\right) = \\frac {1}{\\left| \\Omega_ {i} \\right|} \\sum_ {\\boldsymbol {u} \\in \\Omega_ {i}} \\left\\| I _ {\\text {r e f}} (\\boldsymbol {u}) - I _ {\\mathrm {t}} \\left(\\hat {P} _ {i} [ \\boldsymbol {u} ]\\right) \\right\\| _ {1} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.167, + 0.892, + 0.199 + ], + "angle": 0, + "content": "The resulting photometric cost, aggregated across all depth-scaled primitives \\((s_i, P_i) \\in \\mathcal{P}(\\mathrm{I})\\), is:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.209, + 0.893, + 0.248 + ], + "angle": 0, + "content": "\\[\nE _ {\\text {p h o t o}} = \\frac {1}{| \\mathcal {P} (\\mathrm {I}) |} \\sum_ {P _ {i} = \\left(\\Omega_ {i}, \\mathfrak {D}\\right)} r \\left(P _ {i}, s _ {i}, \\mathrm {I} _ {\\mathrm {t}}, \\mathrm {T} _ {t r}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.26, + 0.894, + 0.335 + ], + "angle": 0, + "content": "Note that we abstain from explicit occlusion handling in our photometric alignment, as our per-segment alignment is by design robust to pixel-level occlusion. We, however, expect our system could further be refined with explicit primitive correspondence checks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.336, + 0.893, + 0.379 + ], + "angle": 0, + "content": "Finally, to obtain the relative pose and depth scales, we minimise the photometric cost using the Adam [20] optimiser:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.383, + 0.892, + 0.403 + ], + "angle": 0, + "content": "\\[\n\\left\\{s _ {i} ^ {\\text {o p t}}, \\mathrm {T} _ {t r} ^ {\\text {o p t}} \\right\\} = \\operatorname {a r g m i n} _ {s _ {i}, \\mathrm {T} _ {t r}} E _ {\\text {p h o t o}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.894, + 0.464 + ], + "angle": 0, + "content": "Our approach trivially extends to multiple reference and target views via photometric cost summation, and we can jointly solve all depth scales and poses using the resulting cost function." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.477, + 0.756, + 0.493 + ], + "angle": 0, + "content": "3.3.2 Monocular Visual Odometry" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.592 + ], + "angle": 0, + "content": "To demonstrate that our representation is suitable for simultaneous geometry and pose estimation, we design a novel monocular visual odometry system which operates directly on primitives. Informally, we incrementally build a local 3D map out of primitives and then track new incoming frames against this map, also in a per-primitive manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.894, + 0.652 + ], + "angle": 0, + "content": "Visual Odometry (VO) consists of three components: initialisation, tracking, and mapping. Below, we cover each of these components. We interleave tracking and mapping in a single thread." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.654, + 0.894, + 0.774 + ], + "angle": 0, + "content": "We adopt a keyframe-based approach [22], which means that the 3D map is represented by a set of posed keyframes \\(\\{\\mathrm{I}_{\\mathrm{kf}}^i, \\mathrm{T}_{\\mathrm{kf}}^i, \\mathcal{G}^i\\}\\). Note that each keyframe also has its estimated dense geometry \\(\\mathcal{G}^i\\), which results in a three-dimensional local map of the scene. We map in a sliding window fashion, with a window size of 5 keyframes. When the window is full, the earliest keyframe is popped from the window." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Initialisation. Initialisation of a VO system typically involves estimating the geometry of the first keyframe and its pose with respect to the subsequent few frames. Initialisation of a monocular incremental SfM or SLAM system is often hard, because one has to solve a chicken-and-egg problem: poses are required to reliably estimate geometry and vice versa. Therefore many VO / SLAM systems employ a wide range of heuristics in order to initialise their" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4983" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "system. However, using our new representation no special treatment is required in this case. In order to bootstrap our system, we simply create the two first keyframes and then employ the Structure-from-Motion method proposed in Sec. 3.3.1. We keep the pose of the first keyframe fixed to fix gauge freedom." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.187, + 0.468, + 0.277 + ], + "angle": 0, + "content": "Tracking. Given a map, the goal of tracking is to estimate the pose \\( \\mathrm{T}_{\\text{track}} \\in \\mathrm{SE}(3) \\) of a new incoming frame \\( \\mathrm{I}_{\\text{track}} \\). We adapt Lucas-Kanade [28] tracking onto our per-primitive formulation. At each step we track a new incoming frame against the latest keyframe in the odometry sliding window." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.279, + 0.468, + 0.355 + ], + "angle": 0, + "content": "We solve the same photometric cost formulated in Eq. (8), but for the pose only. The latest keyframe serves as the reference frame \\( \\mathrm{I}_{\\mathrm{ref}} \\) whereas the new incoming image is set to be the target frame \\( \\mathrm{I}_{\\mathrm{t}} \\). The depth scales of the latest keyframe have already been estimated at the mapping stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.468, + 0.46 + ], + "angle": 0, + "content": "Since target images are not assumed to be primified, our tracking is essentially equivalent to the classical Lucas-Kanade method and hence could be implemented efficiently on modern hardware. Our proof-of-concept implementation does tracking at 2-3 FPS, but we expect high performance gains from employing standard machinery, such as Gauss-Newton optimisation [24]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.466, + 0.468, + 0.526 + ], + "angle": 0, + "content": "Mapping. The mapping stage ensures geometric consistency within the keyframe sliding window of size \\( n \\). Concretely, our mapping refines depth scales \\( \\{s_j\\} \\) and a pose \\( \\mathrm{T}_{\\mathrm{kf}}^i \\) for each keyframe \\( \\mathrm{I}_{\\mathrm{kf}}^i \\) where \\( 0 \\leq i \\leq n \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.527, + 0.468, + 0.675 + ], + "angle": 0, + "content": "For each keyframe, we define a set of target frames against which the photometric cost is being optimised with a connectivity function \\(\\mathcal{M}(t) = \\{\\mathrm{I}_s,\\mathrm{T}_s\\}\\). In our implementation, \\(\\mathcal{M}(t)\\) includes temporally neighbouring keyframes \\(\\mathrm{I}_{\\mathrm{kf}}^{t - 1}\\) and \\(\\mathrm{I}_{\\mathrm{kf}}^{t + 1}\\) (if they exist) used as supporting frames, as well as 4 additional supplementary views, for which only the pose is being estimated. We employ these views to better constrain the geometry and observed that increasing the number of supplementary views does not induce high computational cost." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.677, + 0.468, + 0.705 + ], + "angle": 0, + "content": "Thus, the mapping stage is done by solving a joint photometric cost for all keyframes:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.709, + 0.468, + 0.748 + ], + "angle": 0, + "content": "\\[\nE _ {\\text {m a p p i n g}} = \\sum_ {t = 0} ^ {n} \\sum_ {\\mathrm {I} _ {s} \\in \\mathcal {M} (t)} \\sum_ {P _ {i} \\in \\mathcal {P} \\left(\\mathrm {I} _ {k f} ^ {t}\\right)} r \\left(P _ {i}, s _ {i}, \\mathrm {I} _ {s}, \\mathrm {T} _ {s} ^ {- 1} \\mathrm {T} _ {k f}\\right) \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.762, + 0.209, + 0.779 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.311, + 0.803 + ], + "angle": 0, + "content": "4.1. Sparse Depth Completion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Our SuperPrimitives representation could be seamlessly applied to depth completion with no pre-training required, thereby solving it in a zero-shot manner. For each primitive \\( P_{i} \\in \\mathcal{P}(\\mathrm{I}) \\) we adjust its depth scale to minimise depth discrepancy \\( \\| s_i \\cdot \\mathfrak{D}_i - \\hat{D} \\| \\) with given ground truth sparse depth \\( \\hat{D} \\) across all valid depth points within the segment \\( \\Omega_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.891, + 0.152 + ], + "angle": 0, + "content": "If a segment lacks valid depth measurements, it is discarded. For image regions not covered by any valid primitives, we generate a dense depth prediction by simply bilinearly interpolating depth values." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.892, + 0.333 + ], + "angle": 0, + "content": "Many existing depth completion studies evaluate their methods in artificial scenarios, such as selecting input depth points randomly from a known ground truth. In contrast, our approach is tested using the real-world Void benchmark [57]. The benchmark provides video sequences captured with a RealSense D435i camera together with sparse metric depth measurements acquired from an external visual-inertial SfM system. This setup exhibits noise and biases that might be present in the sparse depth inputs. These measurements are already provided by the dataset itself and are therefore shared across all the methods being compared." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.483 + ], + "angle": 0, + "content": "We compare our results with a recent state-of-the-art model [55], which significantly outperforms previous depth completion methods, particularly in zero-shot generalisation contexts. Other methods evaluated in [55] are also reported in the table. Our depth completion method also operates in a zero-shot manner both task-wise and dataset-wise. That means our method was not trained for the depth completion task and neither of our surface normal prediction network nor segmentation model were trained on the Void dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.891, + 0.545 + ], + "angle": 0, + "content": "We focus on the most challenging \"150 points\" density setting, characterised by minimal sparse depth points per image. The depth is estimated in full \\(480 \\times 640\\) image resolution, following [55]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.712 + ], + "angle": 0, + "content": "In Tab. 1 our method quantitatively performs on par with a recent state-of-the-art method, which is enabled by a monocular depth predictor pre-trained on a vast mixture of datasets [37, 38] and fine-tined on the Void train set. In the zero-shot setting, we outperform VI-Depth [55], which uses a DPT-Hybrid pre-trained backbone, on three out of four metrics. Note that no training was done the for depth completion task for our method. Compared to the ground truth depth maps obtained via a noisy sensor, our predictions show (Fig. 3) sharper object boundaries and are better at preserving straight lines and perpendicular structures." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.722, + 0.796, + 0.737 + ], + "angle": 0, + "content": "4.2. Few-View Structure-from-Motion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Contrasted with depth completion, where global geometry structure is roughly given, we also test our method on few view Structure-from-Motion. Given a set of unposed images \\(\\{\\mathrm{I}_{\\mathrm{ref}},\\mathrm{I}_s^0,\\dots ,\\mathrm{I}_s^{(n - 1)}\\}\\) captured within a small timeframe, the goal is to estimate the depth of the reference image \\(\\mathrm{I_{ref}}\\). In that setup, the set of unposed supplementary views \\(\\{\\mathrm{I}_s^0,\\ldots ,\\mathrm{I}_s^{(n - 1)}\\}\\) will be leveraged for multi-view geometry estimation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We choose the test set of the ScanNet dataset for this evaluation. In each test sequence we select every 200-th" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4984" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.895, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.346, + 0.895, + 0.391 + ], + "angle": 0, + "content": "Figure 3. Depth Completion on Void. We visualise the coloured unprojections of ground truth depth maps provided by a sensor (top row) and the geometry estimated by our method (bottom row). Sparse depth input points are visualised as red dots (electronic zoom-in recommended). Qualitatively, we achieve sharper geometry estimates than from a commodity depth sensor." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.395, + 0.462, + 0.52 + ], + "angle": 0, + "content": "
MethodMAERMSEiMAEiRMSE
VOICED [57]174.04253.1487.39126.3
NLSPN [34]--143238.10
ScaffNet [58]150.65255.0880.79133.33
KBNet [56]131.54263.5466.84128.33
MonDi [27]104.96225.6048.4496.79
VI-Depth (DPT-Hybrid + NYUv2)--55.9085.20
Ours109.0204.1547.3283.40
VI-Depth* (MiDaS) [55]113.27193.3853.8684.82
VI-Depth* (DPT-Hybrid) [55]97.03167.8246.6274.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.62 + ], + "angle": 0, + "content": "Table 1. Depth Completion on Void. We report four most widely used metrics for depth completion on the Void benchmark. The methods which did not use the Void dataset for training are in the top section. Methods trained on Void are marked with an asterisk*. Our method demonstrates superior performance on three out of the four metrics within the zero-shot group. It is second-best on two out of the four metrics overall." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.734 + ], + "angle": 0, + "content": "frame to be the reference frame \\( \\mathrm{I}_{\\mathrm{ref}} \\). Supporting frames are gathered from the neighbouring frames. Then we discard the frame sets with not enough motion to remove mostly static video clips, where SfM could not be performed. This resulted in \\( \\sim 500 \\) reference frames. Note that in these experiments we use a surface normal neural network pre-trained only on the synthetic HyperSim dataset [40]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Since the multi-view depth estimation problem has scale ambiguity, we employ median-scaling to align estimated depth to the metric scale for evaluation. We report iMAE and iRMSE in Fig. 6 for our method with varying number of supplementary views. We compare our depth estimation quality against the method closest to ours, DeepV2D [49], which can also estimate depth together with supplementary frames poses. We demonstrate that our geometry quickly saturates after observing as little as 2 supporting views. Unlike DeepV2D we do not use any external tracking or initial relative pose estimation, yet still our method demonstrates" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.396, + 0.884, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.649, + 0.894, + 0.72 + ], + "angle": 0, + "content": "Figure 4. 3-View SfM on ScanNet. We provide the visualisations of unprojected reference frame depth maps predicted by our method for few-view SfM using one reference and 2 supplementary views. Note that we used surface normal prediction network which was only pretrained on HyperSim [40] for this experiment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.73, + 0.894, + 0.807 + ], + "angle": 0, + "content": "consistent improvement over DeepV2D. Additionally, our method does not have any global prior on relative object positions, unlike DeepV2D. Our approach therefore could be used as an VO / SLAM initialisation mechanism, estimating joint relative poses and geometry." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.816, + 0.757, + 0.833 + ], + "angle": 0, + "content": "4.3. Monocular Visual Odometry" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Monocular Visual Odometry requires both accurate pose and geometry estimation to successfully track camera motion across long trajectories. Minor pose estimation inaccuracies can accumulate over time resulting in what is called" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4985" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.093, + 0.362, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.093, + 0.637, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.093, + 0.887, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.275, + 0.893, + 0.304 + ], + "angle": 0, + "content": "Figure 5. TUM Reconstruction Results. Examples of reconstructions produced by our monocular VO system on the TUM dataset. Each image shows a coloured point cloud of the geometry estimated on an odometry keyframe." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.31, + 0.273, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.31, + 0.468, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.56 + ], + "angle": 0, + "content": "Figure 6. Few-View SfM Depth Estimation Quality. We evaluate the quality of our depth estimation method from an unstructured set of images on the ScanNet dataset. On the \\( x \\)-axis the number of supporting views is shown, while on the \\( y \\)-axis the corresponding depth reconstruction quality metric is reported. We show that the quality of our depth estimation quickly saturates and consistently outperforms its closest competitor, DeepV2D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.469, + 0.644 + ], + "angle": 0, + "content": "drift. With geometry being incorrectly estimated, accurate pose tracking becomes impossible, and vice versa. Our SuperPrimitive representation allows estimating both pose and geometry, enabling us to build a simple monocular VO system which performs better even in hard conditions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.469, + 0.795 + ], + "angle": 0, + "content": "We evaluate our monocular odometry on the TUM RGB-D [46] dataset. The dataset was captured with a handheld camera in indoor scenes. It is renowned for being incredibly challenging (especially for dense reconstruction systems), due to motion blur, rolling shutter artefacts, and abundance of pure rotational motion. We show that thanks to the strong priors encapsulated in our SuperPrimitives, our simple monocular odometry system can handle the TUM dataset without any special treatment (e.g. our method does not involve any special motion blur handling)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.903 + ], + "angle": 0, + "content": "We compare against other VO systems that do not have global bundle adjustment, following the protocol of [10] and evaluate on 8 sequences from the Friedburg 1 split. We use only RGB images as the input to our system and down-sample them to \\(120 \\times 160\\) for efficiency purposes. Since our method is purely monocular, the estimated trajectory lacks global scale and we first use \\(\\mathrm{Sim}(3)\\) alignment to the" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.31, + 0.896, + 0.463 + ], + "angle": 0, + "content": "
Seq.TartanVO [53]DeepV2D [49]DeepFactors [8]DepthCov [10]Ours
3600.1780.2430.1590.1280.173
desk0.1250.1660.1700.0560.085
desk20.1220.3790.2530.0480.108
plant0.2970.2030.3050.2610.153
room0.3330.2460.3640.2570.363
rpy0.0490.1050.0430.0520.055
teddy0.3390.3160.6010.4750.253
xyz0.0620.0640.0350.0560.036
mean0.1880.2150.2410.1670.153
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.466, + 0.893, + 0.55 + ], + "angle": 0, + "content": "Table 2. Trajectory Estimation Error on TUM. Average Trajectory Error (ATE) is compared against other monocular odometry systems on the TUM Friedburg 1 split. The best and second best results are highlighted in bold and underscored correspondingly. Our method outperforms others in terms of the ATE averaged across all trajectories." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.557, + 0.893, + 0.664 + ], + "angle": 0, + "content": "ground truth scale, following standard practice [4]. Tab. 2 shows that, despite the simplicity of our odometry system, it outperforms all other methods in terms of Average Trajectory Error (ATE) [46], averaged across all trajectories. Additionally, our VO is either the best or second best on five out of eight sequences. Besides quantitative evaluation, we also demonstrate reconstruction results in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.676, + 0.619, + 0.692 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.702, + 0.892, + 0.807 + ], + "angle": 0, + "content": "We presented a new representation, SuperPrimitive, which demonstrates how recent advances in building strong single image priors could be incorporated into pose and dense geometry estimation problems. We show that incorporating these priors alleviates the need for sophisticated handcrafted heuristics and paves the way into monocular reconstruction with relative ease." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.817, + 0.893, + 0.888 + ], + "angle": 0, + "content": "Acknowledgments. Research presented in this paper was supported by Dyson Technology Ltd. The authors would like to thank Hidenobu Matsuki, Eric Dexheimer, Riku Murai, and other members of the Dyson Robotics Lab for countless fruitful discussions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4986" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Gwangbin Bae, Ignas Budvytis, and Roberto Cipolla. Estimating and exploiting the aleatoric uncertainty in surface normal estimation. In International Conference on Computer Vision (ICCV), 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] M. Bloesch, J. Czarnowski, R. Clark, S. Leutenegger, and A. J. Davison. CodeSLAM — learning a compact, optimisable representation for dense visual SLAM. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1, 2, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Aljaž Božić, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. Transformerfusion: Monocular rgb scene reconstruction using transformers. Proc. Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.352 + ], + "angle": 0, + "content": "[4] Carlos Campos, Richard Elvira, Juan J. Gomez, Jose M. M. Montiel, and Juan D. Tardos. ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Transactions on Robotics, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.47, + 0.408 + ], + "angle": 0, + "content": "[5] Xu Cao, Hiroaki Santo, Boxin Shi, Fumio Okura, and Yasuyuki Matsushita. Bilateral normal integration. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.47, + 0.451 + ], + "angle": 0, + "content": "[6] Alejo Concha and Javier Civera. Using superpixels in monocular SLAM. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.47, + 0.493 + ], + "angle": 0, + "content": "[7] G. Cross and A. Zisserman. Quadric Reconstruction from Dual-Space Geometry. In Proceedings of the International Conference on Computer Vision (ICCV), 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.47, + 0.548 + ], + "angle": 0, + "content": "[8] J. Czarnowski, T. Laidlow, R. Clark, and A. J. Davison. Deepfactors: Real-time probabilistic dense monocular SLAM. In IEEE Robotics and Automation Letters, 2020. 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.47, + 0.606 + ], + "angle": 0, + "content": "[9] A. J. Davison, N. D. Molton, I. Reid, and O. Stasse. MonoSLAM: Real-Time Single Camera SLAM. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.607, + 0.47, + 0.66 + ], + "angle": 0, + "content": "[10] Eric Dexheimer and Andrew J. Davison. Learning a depth covariance function. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.662, + 0.47, + 0.717 + ], + "angle": 0, + "content": "[11] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In Neural Information Processing Systems (NeurIPS), 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.718, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[12] David Eigen, Christian Puhrsch, and Rob Fergus. Depth Map Prediction from a Single Image using a Multi-Scale Deep Network. In Neural Information Processing Systems (NeurIPS), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.775, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[13] Jakob Engel, Thomas Schoeps, and Daniel Cremers. LSD-SLAM: Large-scale direct monocular SLAM. In Proceedings of the European Conference on Computer Vision (ECCV), 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.831, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[14] J. Engel, V. Koltun, and D. Cremers. Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[15] Jiahui Fu, Yilun Du, Kurran Singh, Joshua B Tenenbaum, and John J Leonard. Neuse: Neural se (3)-equivariant em" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.115, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "bedding for consistent spatial understanding with objects. In Proceedings of Robotics: Science and Systems (RSS), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[16] Y. Furukawa, B. Curless, S. M. Seitz, and R. Szeliski. Towards Internet-scale multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[17] D. Gallup, J.-M. Frahm, and M. Pollefeys. Piecewise planar and non-planar stereo for urban scene reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.246, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[18] Magnus R. Hestenes and Eduard Stiefel. Methods of conjugate gradients for solving linear systems. Journal of research of the National Bureau of Standards, 1952. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[19] M Kaess. Simultaneous localization and mapping with infinite planes. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[20] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[22] G. Klein and D. W. Murray. Parallel Tracking and Mapping for Small AR Workspaces. In Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR), 2007. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[23] T. Laidlow and A. J. Davison. Simultaneous localisation and mapping with quadric surfaces. In International Conference on 3D Vision (3DV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[24] Stefan Leutenegger, Simon Lynen, Michael Bosse, Roland Siegwart, and Paul Furgale. Keyframe-based visual-inertial odometry using nonlinear optimization. International Journal of Robotics Research (IJRR), 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[25] Chen Liu, Jimei Yang, Duygu Ceylan, Ersin Yumer, and Yasutaka Furukawa. Planenet: Piece-wise planar reconstruction from a single rgb image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[26] Chen Liu, Kihwan Kim, Jinwei Gu, Yasutaka Furukawa, and Jan Kautz. Planercnn: 3d plane detection and reconstruction from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[27] Tianlin Liu, Parth T. Agrawal, Allison Chen, Byung-Woo Hong, and A. Wong. Monitored distillation for positive congruent depth completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[28] B. D. Lucas and T. Kanade. An Iterative Image Registration Technique with an Application to Stereo Vision. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), 1981. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] J. McCormac, R. Clark, M. Bloesch, A. J. Davison, and S. Leutenegger. Fusion++: volumetric object-level slam. In Proceedings of the International Conference on 3D Vision (3DV), 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4987" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[30] R. A. Newcombe, S. Lovegrove, and A. J. Davison. DTAM: Dense Tracking and Mapping in Real-Time. In Proceedings of the International Conference on Computer Vision (ICCV), 2011. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[31] Lachlan Nicholson, Michael Milford, and Niko Sunderhauf. Quadricslam: Dual quadrics from object detections as landmarks in object-oriented slam. IEEE Robotics and Automation Letters, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[32] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.471, + 0.375 + ], + "angle": 0, + "content": "[33] Ryosuke Okuta, Yuya Unno, Daisuke Nishino, Shohei Hido, and Crissman Loomis. Copy: A numpy-compatible library for nvidiagpu calculations. In Proceedings of Workshop on Machine Learning Systems (LearningSys) in The Thirty-first Annual Conference on Neural Information Processing Systems (NeurIPS), 2017. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.471, + 0.432 + ], + "angle": 0, + "content": "[34] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.471, + 0.503 + ], + "angle": 0, + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. In Neural Information Processing Systems (NeurIPS), 2019. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.471, + 0.574 + ], + "angle": 0, + "content": "[36] Albert Pumarola, Alexander Vakhitov, Antonio Agudo, Alberto Sanfeliu, and Francese Moreno-Noguer. Pl-slam: Real-time monocular visual slam with points and lines. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.471, + 0.63 + ], + "angle": 0, + "content": "[37] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[38] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.471, + 0.745 + ], + "angle": 0, + "content": "[39] Xiaofeng Ren and Jitendra Malik. Learning a classification model for segmentation. In Proceedings of the International Conference on Computer Vision (ICCV), 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.471, + 0.83 + ], + "angle": 0, + "content": "[40] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[41] R. F. Salas-Moreno, R. A. Newcombe, H. Strasdat, P. H. J. Kelly, and A. J. Davison. SLAM++: Simultaneous Localisation and Mapping at the Level of Objects. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2013. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[42] Mohamed Sayed, John Gibson, Jamie Watson, Victor Prisacariu, Michael Firman, and Clément Godard. Simplerecon: 3d reconstruction without 3d convolutions. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.217 + ], + "angle": 0, + "content": "[43] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.893, + 0.273 + ], + "angle": 0, + "content": "[44] Jonathan Shade, Steven Gortler, Li-wei He, and Richard Szeliski. Layered depth images. In Proceedings of the 25th annual conference on Computer graphics and interactive techniques, 1998. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.893, + 0.328 + ], + "angle": 0, + "content": "[45] Jingjia Shi, Shuaifeng Zhi, and Kai Xu. Planerectr: Unified query learning for 3d plane recovery from a single view. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.893, + 0.385 + ], + "angle": 0, + "content": "[46] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A Benchmark for the Evaluation of RGB-D SLAM Systems. In Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS), 2012. 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.893, + 0.441 + ], + "angle": 0, + "content": "[47] E. Sucar, K. Wada, and A. J. Davison. NodeSLAM: Neural object descriptors for multi-view shape reconstruction. In Proceedings of the International Conference on 3D Vision (3DV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.893, + 0.509 + ], + "angle": 0, + "content": "[48] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. NeuralRecon: Real-time coherent 3D reconstruction from monocular video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.893, + 0.566 + ], + "angle": 0, + "content": "[49] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. In Proceedings of the International Conference on Learning Representations (ICLR), 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.893, + 0.621 + ], + "angle": 0, + "content": "[50] Alexander Vakhitov, Jan Funke, and Francesc Moreno-Noguer. Accurate and linear time pose estimation from points and lines. In Proceedings of the European Conference on Computer Vision (ECCV), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.893, + 0.676 + ], + "angle": 0, + "content": "[51] Tom van Dijk and Guido C.H.E. de Croon. How do neural networks see depth in single images? In Proceedings of the International Conference on Computer Vision (ICCV), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.678, + 0.893, + 0.747 + ], + "angle": 0, + "content": "[52] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.893, + 0.789 + ], + "angle": 0, + "content": "[53] Wenshan Wang, Yaoyu Hu, and Sebastian Scherer. TartanVO: A Generalizable Learning-based VO. In Conference on Robot Learning (CoRL), 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.893, + 0.844 + ], + "angle": 0, + "content": "[54] C. S. Weerasekera, Y. Latif, R. Garg, and I. Reid. Dense monocular reconstruction using surface normals. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[55] Wofk, Diana and Ranftl, René and Müller, Matthias and Koltun, Vladlen. Monocular Visual-Inertial Depth Estimation. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2023. 6, 7, 12" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4988" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[56] Alex Wong and Stefano Soatto. Unsupervised depth completion with calibrated backprojection layers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[57] Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.47, + 0.247 + ], + "angle": 0, + "content": "[58] Alex Wong, Safa Cicek, and Stefano Soatto. Learning topology from synthetic data for unsupervised depth completion. IEEE Robotics and Automation Letters, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[59] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In Proceedings of SIGGRAPH, 2018. 12" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "4989" + } + ] +] \ No newline at end of file diff --git a/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/556fb01c-1445-4c08-b116-0426ba52ea17_origin.pdf b/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/556fb01c-1445-4c08-b116-0426ba52ea17_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b966b01766e28dcf9f0fd5ff69a8368f0fd2cf09 --- /dev/null +++ b/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/556fb01c-1445-4c08-b116-0426ba52ea17_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e61c3e97193167d42929d10d1aa926c384aae7bf2d6863f06da7eba38fb9dfd +size 4400146 diff --git a/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/full.md b/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7cb9747e1c9b0d1d7f7db4faf0aea4bfd1c4c601 --- /dev/null +++ b/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/full.md @@ -0,0 +1,346 @@ +# SuperPrimitive: Scene Reconstruction at a Primitive Level + +Kirill Mazur, Gwangbin Bae, Andrew J. Davison Dyson Robotics Lab, Imperial College London + +{k.mazur21, g.bae, a.davison}@imperial.ac.uk + +# Abstract + +Joint camera pose and dense geometry estimation from a set of images or a monocular video remains a challenging problem due to its computational complexity and inherent visual ambiguities. Most dense incremental reconstruction systems operate directly on image pixels and solve for their 3D positions using multi-view geometry cues. Such pixel-level approaches suffer from ambiguities or violations of multi-view consistency (e.g. caused by textureless or specular surfaces). + +We address this issue with a new image representation which we call a SuperPrimitive. SuperPrimitives are obtained by splitting images into semantically correlated local regions and enhancing them with estimated surface normal directions, both of which are predicted by state-of-the-art single image neural networks. This provides a local geometry estimate per SuperPrimitive, while their relative positions are adjusted based on multi-view observations. + +We demonstrate the versatility of our new representation by addressing three 3D reconstruction tasks: depth completion, few-view structure from motion, and monocular dense visual odometry. Project page: https://makezur.github.io/SuperPrimitive/ + +# 1. Introduction + +Enriching monocular incremental reconstruction with prior world knowledge is essential for resolving visual ambiguities. This issue is particularly prevalent in scenarios with scarce data observations available: a notable example would be monocular visual SLAM, where images are being streamed from a camera into the system in real-time. + +When a monocular vision system encounters a new scene region, it must estimate the region's geometry based on a very limited number of observations. Without this, continuous camera motion tracking would not be possible. Once the scene region is thoroughly observed, the initial geometry estimate should be refined to better explain the multiview information. + +This naturally leads to a question: what sort of priors are effective in both providing reliable initial geometry es + +![](images/374776498f3c7f0c4217067d977995de1c873b9ec63f7fe2fce7212e571eb247.jpg) +Figure 1. Multi-View Geometry with SuperPrimitives. SuperPrimitives are extracted from an input frame by dividing it into image segments equipped with estimated surface normal directions (bottom-left). Each SuperPrimitive induces a dense reconstruction within the corresponding image segment up to a priori unknown scale. Different possible reconstructions are shown in light blue. The scales are then jointly optimised together with a relative camera pose to fit multi-view photometric constraints (visualised in green and red). The resulting dense reconstruction of the reference frame is shown in the top. + +timates and supporting multi-view consistency? Geometric priors generally fall into one of two categories: local and global. Local priors, such as smoothness assumption [30] or surface normal regularisation [52], impose additional constraints within a small neighbourhood. Global priors, on the other hand, aim to impose constraints on a larger scale, such as depth prediction [2, 11]. + +Our key observation is that some of the geometrical correlations are more reliable than the others, and therefore could be safely "locked in" together within local regions + +based on a single-view prediction. Points belonging to the same rigid body are strongly correlated, making it unnecessary to determine their depth independently. In contrast, distinct and unrelated objects can be placed arbitrarily in the scene. As the number of objects increases, learning a reliable global prior on their relative positions becomes an increasingly complex problem. + +In this work, we show that purely local but strong priors are enough to achieve excellent performance across a variety of geometric vision tasks. For that purpose, we introduce a novel representation, SuperPrimitives. A Super-Primitive represents a local image segment, coupled with a dense shape estimate which is determined up to a scale factor. The scale factor can be further adjusted based on information observed from other views or additional measurements. + +We show that SuperPrimitives can be efficiently constructed using a front-end which consists of two single-image neural networks, extracting image segmentation and surface normal prediction. Effectively, our neural front-end predicts whether adjacent pixels belong to the same geometrical entity through image segmentation and estimates a surface normal at this point, thereby providing an infinitesimal geometry estimate. + +We delegate global, scene-level alignment to our streamlined multi-view, iterative, optimisation-based back-end. The resulting front-end / back-end tandem combines the flexibility of multi-view based optimisation methods with the observation efficiency common in prior-driven systems. + +Our new representation showcases its versatility in three key applications, where single-view ambiguity is resolved via additional measurements or viewpoints: + +- Firstly, it adeptly handles zero-shot depth completion tasks in real-world scenarios, matching the performance of state-of-the-art methods tailored for depth completion; +- Secondly, it facilitates joint pose and depth estimation using a limited set of unstructured images, surpassing its nearest competitor even in the absence of global priors; +- Thirdly, our method outperforms previous monocular visual odometry systems on the challenging TUM dataset, and exhibits robustness across various domains. + +# 2. Related Work + +Monocular Reconstruction. Both offline monocular reconstruction systems, such as COLMAP [43] and [16], or online systems, such as MonoSLAM [9] and DSO [22] track, filter, and reconstruct only well-constrained points with high and reliable photometric information. This often involves fine-grained and sophisticated point management to reliably resolve visual ambiguities and filter unreliable visual observations. DTAM [30] demonstrated feasibility of incremental dense reconstruction in the monocular scenario. DTAM employed a hand-crafted local smoothness + +prior to handle regions with poor texture. Subsequently, other local priors [32, 54] have been extensively explored to regularise multi-view geometry estimation problems. + +Global Priors. We are interested in exploring the space of possible geometric priors from single-view networks. Depth prediction [12, 37, 38] is the most obvious choice but leads to a rigid per-image reconstruction which is difficult to feed into multi-view optimisation. In the recent years, deep-learning based approaches sought to replace explicit multi-view geometry estimating with learning-based methods [3, 42, 48]. These methods, however, assume known poses and are, therefore, not suitable for joint pose and geometry estimation. Notably, CodeSLAM [2] introduced depth prediction conditioned on latent codes, which are then optimised to achieve cross-view consistency. However, CodeSLAM still struggles with out-of-domain data, as depth prediction networks are known to struggle with generalisation [51]. + +Higher-Level Mapping. Introducing parametric primitives, such as lines [36, 50], planes [6, 17, 19, 25, 26, 45] or even high-order algebraic shapes [7, 23, 31] to better constrain multi-view geometry problems have been thoroughly explored over the last few decades. They all however use assumptions which may not often hold for all 3D scenes, especially for dense reconstruction. + +Besides parametric algebraic primitives object-level mapping has been explored in the last decade. SLAM++ [41] represented a map with a set of CAD models retrieved from an existing database and tracked camera position against these models. This method, however, could not be applied in a variety of settings due to the limited size of the CAD databases. Even objects of the same class can vary in their geometric appearance from instance to instance. This was later approached by [15, 47], who also represent their map via a set of objects, but learn a latent space of possible geometric variations per object class. Fusion++ [29] extended this approach into a more versatile per object depth fusion using an RGB-D sensor and mask proposals from a pre-trained neural network. These methods rely on an RGB-D inputs and assume a pre-defined set of object classes, therefore lacking generality. + +# 3. Method + +Firstly, we introduce the concept of a SuperPrimitive and explain how an image is processed into a set of its SuperPrimitives, referring to this part of our method as the front-end. In the second part of this section, we describe how multi-view geometry problems can be reformulated at the level of SuperPrimitives instead of pixels. We refer to this stage of SuperPrimitive alignment as the back-end. + +The core of our method proposes splitting a given image into a set of (possibly overlapping) minimal segments, image regions which are likely to have strongly correlated + +![](images/2394536898c6d7b5ebe1603afd03c5680681346861012fb5086549ac85b35474.jpg) +Figure 2. SuperPrimitives Extraction. (left) Our front-end processor extracts SuperPrimitives from an image by dividing it into a set of image regions with surface normal directions estimated for each image pixel within the segment. (right) Highlighted SuperPrimitives extracted from the image are visualised by showing their estimated normal and colour maps side by side. Note some of them are scaled either up or down for better viewing. While some of the SuperPrimitives are akin to object-level segmentation, the others tend to represent more low-level image segments. + +![](images/29acfc2df01382d0829f980c75c3a252ad07ed79291b184ab759c9f6ccd96a22.jpg) + +geometry. We repurpose the recent state-of-the-art Segment Anything (SAM) [21] model into predicting these minimal segments. + +Our key idea is to estimate local geometry within each segment from a single view, while leaving the relative positioning of the segments to be estimated via multi-view photometric consistency optimisation. We refer to these geometrically enhanced minimal image segments as Super-Primitives, since they are inspired by both superpixels [39] and geometric 3D primitives. + +For per-segment local geometry estimation we employ an off-the-shelf surface normal prediction network [1], to estimate infinitesimal geometry for each image pixel. The surface normals within an image segment can be used to estimate its depth via simple integration up to a scale factor. We set these scale factors — depth scales — to be optimisable parameters, which are either optimised via multi-view cues or explicit depth measurements for the depth completion experiments. + +Thus, our method combines the strong priors provided by state-of-the-art neural networks in the front-end with the flexibility and consistency offered by multi-view optimisation. + +# 3.1. Conventions + +Unless stated otherwise, we use lowercase letters for scalar values, bold letters for vectors, and uppercase letters for matrices. We consider images $\mathbf{I} \in \mathbb{R}^{3 \times h \times w}$ of height $h$ and width $w$ captured by a camera with a known calibration matrix $\mathbf{K} = \left( \begin{array}{ccc} f_u & 0 & c_u \\ 0 & f_v & c_v \\ 0 & 0 & 1 \end{array} \right) \in \mathbb{R}^{3 \times 3}$ . Image pixels are parameterised as $\mathbf{u} = (u, v) \in [0, h - 1] \times [0, w - 1]$ . Given a per pixel depth function $z(\mathbf{u})$ : $\mathbf{I} \to \mathbb{R}^+$ , we define image's unprojection onto the 3D space to be $\pi^{-1}(\mathbf{u}) = z(\mathbf{u})\mathbf{K}^{-1}\dot{\mathbf{u}}$ , + +where dot is a homogenisation operator $\dot{\pmb{p}} = (\pmb{p},1)$ . Conversely, for a set of 3D points $\pmb{x} = (x,y,z) \in \mathbb{R}^3$ we define the projection $\pi(\pmb{x}) = \gamma\left(\frac{1}{z} K\pmb{x}\right)$ , where the $\gamma$ function drops the last coordinate $\gamma((x,y,z)^T) = (x,y)^T$ . + +We represent camera poses as matrices $\mathrm{T}_{WC} \in \mathrm{SE}(3)$ in the camera-to-world coordinates unless stated otherwise. During pose optimisation we store linearised pose increments $\xi \in \mathfrak{se}(3)$ as Lie algebra elements. These increments are used then to update current camera pose estimates, see [13, 14] for more detail. + +# 3.2. SuperPrimitives + +We split the input image I into a set of possibly overlapping connected regions (segments), each of which is equipped with its local 3D geometry. More formally, a SuperPrimitive $P = (\Omega, \mathfrak{D})$ is a connected image region $\Omega \subseteq I$ which is also equipped with an unscaled depth map $\mathfrak{D} : \Omega \to [0, \infty)$ for each pixel within the region $\Omega$ . Here, unscaled means that each SuperPrimitive's depth $\mathfrak{D}$ differs from its ground truth depth $D$ by an a priori unknown scalar. In other words, there exists a single scalar $s$ such that: + +$$ +D = s \cdot \mathfrak {D} \quad \text {f o r a l l p i x e l s} \mathrm {p} \in \Omega \tag {1} +$$ + +We say that an image $\mathrm{I}$ is primified into super-primitives $\mathcal{P}(\mathrm{I}) = \{P_i = (\Omega_i, \mathfrak{D}_i)\}$ if it has $n$ , possibly overlapping primitives that lie within the image $\mathrm{I}$ , i.e. $\bigcup \Omega_i \subseteq \mathrm{I}$ . + +For brevity we use the words primitive and SuperPrimitive interchangeably throughout the rest of the paper. + +Representing Geometry with SuperPrimitives. A set of primitives $\mathcal{P}$ itself is not enough to extract dense image geometry due to scale/depth ambiguity. Therefore, we introduce the concept of optimisable depth scales which anchor the correct depth scale for each primitive. + +Given a scalar depth scale $s$ for a super-primitive $P$ , one can infer its depth as $D(p) = s \cdot \mathfrak{D}(p)$ for every point $p \in \Omega$ . In practice, we employ a log-depth representation to represent depth throughout the whole system. This means that we store optimisable log-depth scales $\log s$ and unscaled log-depth $\log \mathfrak{D}$ , and so depth inference reduces to a simple shift operation $\log D = \log s + \log \mathfrak{D}$ . + +In contrast to other dense methods, we estimate the dense geometry $\mathcal{G}$ of the image in the form of a point cloud rather than a depth image. This choice is driven by the fact that multiple primitives' supporting segments $\Omega_{i}$ and $\Omega_{j}$ might overlap, which would lead to different depth estimates within their intersection $\Omega_{i} \cap \Omega_{j}$ . + +Given a set of depth scaled primitives $\mathbb{P} = \{(s_i, P_i)\}$ , the geometry $\mathcal{G}$ of the image $I$ can be estimated as follows: + +$$ +\mathcal {G} := \bigcup_ {i} \pi^ {- 1} \left(s _ {i} \cdot \mathfrak {D} _ {i}\right) \tag {2} +$$ + +To obtain estimated depth $\hat{D}$ , we average depth values along the corresponding camera rays. + +Converting an Image into SuperPrimitives. Next we explain how SuperPrimitives are obtained in practice. An input image $I$ is first split into a set of minimal image segments $\Omega_{i}$ produced by an image segmentation model. We use the Segment Anything [21] model in this work, since it captures low-level semantic correlations highly accurately. Given the extracted segments, we independently estimate the local geometry $\mathfrak{D}_i$ within each segment by integrating surface normals predicted by another state-of-the-art neural network [1] within this region. + +Note that this approach is premised on the assumption of geometry being continuous within each predicted image segment $\Omega_{i}$ , hence its unscaled depth values could be obtained by integrating its surface normals. Although this is not guaranteed a priori, we observed that a more fine-grained mask selection yields a compelling correlation with geometrical continuity. We discuss this in more detail in the following section. + +Image Segment Retrieval. An ideal neural network which estimates image segments $\{\Omega_i\}$ should predict regions of geometric continuity, where depth could be obtained via surface normal integration. However, to the best of our knowledge, such a network does not exist. Our approach aims to approximate this behaviour by utilising the Segment Anything model, coupled with a specialised mask selection process. Specifically, for each SAM query, we strive to select the smallest predicted mask surrounding that point. While such over-segmentation could potentially increase the dimensionality of the multi-view optimisation problem, under-segmentation might lead to incorrect geometry estimate within the primitive. Since we only adjust the scale of a primitive after the normal integration is com + +pleted, any incorrect geometry estimates within a primitive cannot be compensated at this stage. + +The segmentation model employed in this work outputs three binary segmentation masks given a query point $q \in \mathbf{I}$ . For each query $q$ we first filter predicted masks using the post-processing introduced in [21], such as stability filtering and Non-Maximum-Suppression (NMS). Even though our method allows redundant regions, we employed NMS to remove similar segments and save compute in the image alignment stage. With filtering done, the mask with the smallest area is selected. If the filtered set is empty, we discard the query point. + +We query the SAM backbone feature extractor once per image. Then we first sample 300 query points randomly across the image, followed by the filtering discussed above. Then the image coverage mask is calculated and an additional 100 mask query points are actively sampled in the uncovered regions. + +Normal Integration. To estimate each segment's local geometry, we first pass the image through a surface normal estimation network. This network effectively predicts the derivative of desired depth values. Then, we "integrate" these surface normal vectors within each image segment to obtain its unscaled depth map. + +For each image pixel $\boldsymbol{u} = (u,v)^T \in \mathrm{I}$ we estimate its surface normal vector $\boldsymbol{n} = (n_x, n_y, n_z)^T$ with a pretrained state-of-the-art CNN [1]. As was demonstrated in [5], the log-depth $\tilde{z} = \log(z(\boldsymbol{u}))$ satisfies the following PDEs within a segment $\Omega$ : + +$$ +\tilde {n} _ {z} \partial_ {u} \tilde {z} + n _ {x} = 0 \quad \text {a n d} \quad \tilde {n} _ {z} \partial_ {v} \tilde {z} + n _ {y} = 0 \tag {3} +$$ + +where $\tilde{n}_z = n_x(u - c_u) + n_y(v - c_v) + n_zf$ . The depth values within the segment $\Omega$ can therefore be obtained via minimising the following functional: + +$$ +\min _ {z} \iint_ {\Omega} \left(\tilde {n} _ {z} \partial_ {u} \tilde {z} + n _ {x}\right) ^ {2} + \left(\tilde {n} _ {z} \partial_ {u} \tilde {z} + n _ {y}\right) ^ {2} d u d v \tag {4} +$$ + +Note that this leads to a family of solutions $\tilde{z} + \text{Const}$ which differ by a shift constant. This is due to the fact that only partial derivatives of $\tilde{z}$ are used in the functional. That means that the actual depth will be estimated up to scale, since $z = \exp(\tilde{z})$ by definition. + +This ambiguity prompted us to introduce the notion of depth scales, to allow each segment to be adjusted towards its true depth. We implemented batched normal integration, efficiently solving all optimisation problems as a single sparse linear system using conjugate gradient method [18]. This implementation enables the integration of approximately 100 to 300 segments within a total time frame of $\sim 100\mathrm{ms}$ . + +# 3.3. Primitive-based Image Alignment + +Now we explain how our new representation can be used as a building block for dense multi-view geometry and pose + +estimation, combining an optimisation-based mindset with learned single view priors. To make optimisation computationally tractable, we design our image alignment to fully abstract away any knowledge about the neural networks involved in the image primification stage. Hence, even though there has been early evidence that SAM is independently capable of establishing (coarse) mask correspondences across neighbouring frames, we rely on photometric information only during the image alignment stage. We believe that more semantic primitive association would be "heavier" and coarser since it would not provide per pixel correspondences, and would therefore not be a good fit to be used in a multi-view optimisation loop. However, we speculate that such a method could be employed to further enhance our system, e.g. to improve occlusion handling, particularly as these models become more computationally affordable. + +# 3.3.1 Two-view SfM on SuperPrimitives + +For the sake of clarity, we formulate our method in the simplest case of two frames observing the scene, although it could be trivially extended into a setting with a higher number of views. At the core of our Structure-from-Motion (SfM) approach lies a per-primitive photometric alignment, which contrasts with widely accepted per-pixel photometric alignment [14, 30] techniques. Informally speaking, our proposition is to treat each image primitive as a "rigid" piece, which is only allowed to be scaled up or down. This degree of freedom is due to the scale ambiguity present for each primitive. Thus, instead of estimating a depth value per pixel, we only estimate a depth-scale per primitive, which is illustrated in Fig. 1. This greatly reduces the dimensionality of the optimisation problem, especially in the case of an unknown relative pose. Note that our method does not require the target image to be primified nor does it not require any pre-established correspondences (e.g. primitive to primitive). + +We assume a primified reference image $\mathrm{I}_{\mathrm{ref}}$ with its set of primitives $\mathcal{P}(\mathrm{I}) = \{(\Omega_i, \mathfrak{D}_i)\}$ . Given an unposed target image $\mathrm{I}_{\mathrm{t}}$ , our goal is to jointly estimate its relative pose $\mathrm{T}_{tr}$ with respect to the reference image $\mathrm{I}_{\mathrm{ref}}$ as well as the dense geometry $\mathcal{G}_{\mathrm{ref}} = \bigcup_{i} \pi^{-1}(s_i \cdot \mathfrak{D}_i)$ of the reference frame. + +While unscaled depths $\mathfrak{D}_i$ are given by a per-image preprocessor after the primification of the reference image, the set of depth scales $s_i$ are yet to be estimated. In our case, we jointly estimate both depth scales $s_i$ and a relative image pose $\mathrm{T}_{tr}$ by solving a photometric consistency optimisation problem. First, we warp each depth-scaled primitive $(s_i, P_i)$ from the reference frame $\mathrm{I}_{\mathrm{ref}}$ into the target frame $\mathrm{I}_{\mathrm{t}}$ : + +$$ +\hat {P} _ {i} [ \boldsymbol {u} ] = \pi \left(\mathrm {T} _ {t r} \pi^ {- 1} \left(\boldsymbol {u}, s _ {i} \cdot \mathfrak {D} _ {i}\right)\right) \tag {5} +$$ + +Then a per-segment photometric residual for the primitive $P_{i} = (\Omega_{i},\mathfrak{D}_{i})$ is defined by averaging all photometric re + +projection $\ell^1$ errors across every pixel $\pmb{u} \in \Omega_i$ : + +$$ +r \left(P _ {i}, s _ {i}, I _ {\mathrm {t}}, T _ {t r}\right) = \frac {1}{\left| \Omega_ {i} \right|} \sum_ {\boldsymbol {u} \in \Omega_ {i}} \left\| I _ {\text {r e f}} (\boldsymbol {u}) - I _ {\mathrm {t}} \left(\hat {P} _ {i} [ \boldsymbol {u} ]\right) \right\| _ {1} \tag {6} +$$ + +The resulting photometric cost, aggregated across all depth-scaled primitives $(s_i, P_i) \in \mathcal{P}(\mathrm{I})$ , is: + +$$ +E _ {\text {p h o t o}} = \frac {1}{| \mathcal {P} (\mathrm {I}) |} \sum_ {P _ {i} = \left(\Omega_ {i}, \mathfrak {D}\right)} r \left(P _ {i}, s _ {i}, \mathrm {I} _ {\mathrm {t}}, \mathrm {T} _ {t r}\right) \tag {7} +$$ + +Note that we abstain from explicit occlusion handling in our photometric alignment, as our per-segment alignment is by design robust to pixel-level occlusion. We, however, expect our system could further be refined with explicit primitive correspondence checks. + +Finally, to obtain the relative pose and depth scales, we minimise the photometric cost using the Adam [20] optimiser: + +$$ +\left\{s _ {i} ^ {\text {o p t}}, \mathrm {T} _ {t r} ^ {\text {o p t}} \right\} = \operatorname {a r g m i n} _ {s _ {i}, \mathrm {T} _ {t r}} E _ {\text {p h o t o}} \tag {8} +$$ + +Our approach trivially extends to multiple reference and target views via photometric cost summation, and we can jointly solve all depth scales and poses using the resulting cost function. + +# 3.3.2 Monocular Visual Odometry + +To demonstrate that our representation is suitable for simultaneous geometry and pose estimation, we design a novel monocular visual odometry system which operates directly on primitives. Informally, we incrementally build a local 3D map out of primitives and then track new incoming frames against this map, also in a per-primitive manner. + +Visual Odometry (VO) consists of three components: initialisation, tracking, and mapping. Below, we cover each of these components. We interleave tracking and mapping in a single thread. + +We adopt a keyframe-based approach [22], which means that the 3D map is represented by a set of posed keyframes $\{\mathrm{I}_{\mathrm{kf}}^i, \mathrm{T}_{\mathrm{kf}}^i, \mathcal{G}^i\}$ . Note that each keyframe also has its estimated dense geometry $\mathcal{G}^i$ , which results in a three-dimensional local map of the scene. We map in a sliding window fashion, with a window size of 5 keyframes. When the window is full, the earliest keyframe is popped from the window. + +Initialisation. Initialisation of a VO system typically involves estimating the geometry of the first keyframe and its pose with respect to the subsequent few frames. Initialisation of a monocular incremental SfM or SLAM system is often hard, because one has to solve a chicken-and-egg problem: poses are required to reliably estimate geometry and vice versa. Therefore many VO / SLAM systems employ a wide range of heuristics in order to initialise their + +system. However, using our new representation no special treatment is required in this case. In order to bootstrap our system, we simply create the two first keyframes and then employ the Structure-from-Motion method proposed in Sec. 3.3.1. We keep the pose of the first keyframe fixed to fix gauge freedom. + +Tracking. Given a map, the goal of tracking is to estimate the pose $\mathrm{T}_{\text{track}} \in \mathrm{SE}(3)$ of a new incoming frame $\mathrm{I}_{\text{track}}$ . We adapt Lucas-Kanade [28] tracking onto our per-primitive formulation. At each step we track a new incoming frame against the latest keyframe in the odometry sliding window. + +We solve the same photometric cost formulated in Eq. (8), but for the pose only. The latest keyframe serves as the reference frame $\mathrm{I}_{\mathrm{ref}}$ whereas the new incoming image is set to be the target frame $\mathrm{I}_{\mathrm{t}}$ . The depth scales of the latest keyframe have already been estimated at the mapping stage. + +Since target images are not assumed to be primified, our tracking is essentially equivalent to the classical Lucas-Kanade method and hence could be implemented efficiently on modern hardware. Our proof-of-concept implementation does tracking at 2-3 FPS, but we expect high performance gains from employing standard machinery, such as Gauss-Newton optimisation [24]. + +Mapping. The mapping stage ensures geometric consistency within the keyframe sliding window of size $n$ . Concretely, our mapping refines depth scales $\{s_j\}$ and a pose $\mathrm{T}_{\mathrm{kf}}^i$ for each keyframe $\mathrm{I}_{\mathrm{kf}}^i$ where $0 \leq i \leq n$ . + +For each keyframe, we define a set of target frames against which the photometric cost is being optimised with a connectivity function $\mathcal{M}(t) = \{\mathrm{I}_s,\mathrm{T}_s\}$ . In our implementation, $\mathcal{M}(t)$ includes temporally neighbouring keyframes $\mathrm{I}_{\mathrm{kf}}^{t - 1}$ and $\mathrm{I}_{\mathrm{kf}}^{t + 1}$ (if they exist) used as supporting frames, as well as 4 additional supplementary views, for which only the pose is being estimated. We employ these views to better constrain the geometry and observed that increasing the number of supplementary views does not induce high computational cost. + +Thus, the mapping stage is done by solving a joint photometric cost for all keyframes: + +$$ +E _ {\text {m a p p i n g}} = \sum_ {t = 0} ^ {n} \sum_ {\mathrm {I} _ {s} \in \mathcal {M} (t)} \sum_ {P _ {i} \in \mathcal {P} \left(\mathrm {I} _ {k f} ^ {t}\right)} r \left(P _ {i}, s _ {i}, \mathrm {I} _ {s}, \mathrm {T} _ {s} ^ {- 1} \mathrm {T} _ {k f}\right) \tag {9} +$$ + +# 4. Experiments + +# 4.1. Sparse Depth Completion + +Our SuperPrimitives representation could be seamlessly applied to depth completion with no pre-training required, thereby solving it in a zero-shot manner. For each primitive $P_{i} \in \mathcal{P}(\mathrm{I})$ we adjust its depth scale to minimise depth discrepancy $\| s_i \cdot \mathfrak{D}_i - \hat{D} \|$ with given ground truth sparse depth $\hat{D}$ across all valid depth points within the segment $\Omega_{i}$ . + +If a segment lacks valid depth measurements, it is discarded. For image regions not covered by any valid primitives, we generate a dense depth prediction by simply bilinearly interpolating depth values. + +Many existing depth completion studies evaluate their methods in artificial scenarios, such as selecting input depth points randomly from a known ground truth. In contrast, our approach is tested using the real-world Void benchmark [57]. The benchmark provides video sequences captured with a RealSense D435i camera together with sparse metric depth measurements acquired from an external visual-inertial SfM system. This setup exhibits noise and biases that might be present in the sparse depth inputs. These measurements are already provided by the dataset itself and are therefore shared across all the methods being compared. + +We compare our results with a recent state-of-the-art model [55], which significantly outperforms previous depth completion methods, particularly in zero-shot generalisation contexts. Other methods evaluated in [55] are also reported in the table. Our depth completion method also operates in a zero-shot manner both task-wise and dataset-wise. That means our method was not trained for the depth completion task and neither of our surface normal prediction network nor segmentation model were trained on the Void dataset. + +We focus on the most challenging "150 points" density setting, characterised by minimal sparse depth points per image. The depth is estimated in full $480 \times 640$ image resolution, following [55]. + +In Tab. 1 our method quantitatively performs on par with a recent state-of-the-art method, which is enabled by a monocular depth predictor pre-trained on a vast mixture of datasets [37, 38] and fine-tined on the Void train set. In the zero-shot setting, we outperform VI-Depth [55], which uses a DPT-Hybrid pre-trained backbone, on three out of four metrics. Note that no training was done the for depth completion task for our method. Compared to the ground truth depth maps obtained via a noisy sensor, our predictions show (Fig. 3) sharper object boundaries and are better at preserving straight lines and perpendicular structures. + +# 4.2. Few-View Structure-from-Motion + +Contrasted with depth completion, where global geometry structure is roughly given, we also test our method on few view Structure-from-Motion. Given a set of unposed images $\{\mathrm{I}_{\mathrm{ref}},\mathrm{I}_s^0,\dots ,\mathrm{I}_s^{(n - 1)}\}$ captured within a small timeframe, the goal is to estimate the depth of the reference image $\mathrm{I_{ref}}$ . In that setup, the set of unposed supplementary views $\{\mathrm{I}_s^0,\ldots ,\mathrm{I}_s^{(n - 1)}\}$ will be leveraged for multi-view geometry estimation. + +We choose the test set of the ScanNet dataset for this evaluation. In each test sequence we select every 200-th + +![](images/d14519d12162bbcf13ae7584fcf4c7af00e2af919923dd34972278d57ccf74ef.jpg) +Figure 3. Depth Completion on Void. We visualise the coloured unprojections of ground truth depth maps provided by a sensor (top row) and the geometry estimated by our method (bottom row). Sparse depth input points are visualised as red dots (electronic zoom-in recommended). Qualitatively, we achieve sharper geometry estimates than from a commodity depth sensor. + +
MethodMAERMSEiMAEiRMSE
VOICED [57]174.04253.1487.39126.3
NLSPN [34]--143238.10
ScaffNet [58]150.65255.0880.79133.33
KBNet [56]131.54263.5466.84128.33
MonDi [27]104.96225.6048.4496.79
VI-Depth (DPT-Hybrid + NYUv2)--55.9085.20
Ours109.0204.1547.3283.40
VI-Depth* (MiDaS) [55]113.27193.3853.8684.82
VI-Depth* (DPT-Hybrid) [55]97.03167.8246.6274.67
+ +Table 1. Depth Completion on Void. We report four most widely used metrics for depth completion on the Void benchmark. The methods which did not use the Void dataset for training are in the top section. Methods trained on Void are marked with an asterisk*. Our method demonstrates superior performance on three out of the four metrics within the zero-shot group. It is second-best on two out of the four metrics overall. + +frame to be the reference frame $\mathrm{I}_{\mathrm{ref}}$ . Supporting frames are gathered from the neighbouring frames. Then we discard the frame sets with not enough motion to remove mostly static video clips, where SfM could not be performed. This resulted in $\sim 500$ reference frames. Note that in these experiments we use a surface normal neural network pre-trained only on the synthetic HyperSim dataset [40]. + +Since the multi-view depth estimation problem has scale ambiguity, we employ median-scaling to align estimated depth to the metric scale for evaluation. We report iMAE and iRMSE in Fig. 6 for our method with varying number of supplementary views. We compare our depth estimation quality against the method closest to ours, DeepV2D [49], which can also estimate depth together with supplementary frames poses. We demonstrate that our geometry quickly saturates after observing as little as 2 supporting views. Unlike DeepV2D we do not use any external tracking or initial relative pose estimation, yet still our method demonstrates + +![](images/cd37b939f9f2698aba14071cafe4e0bb7da26534354b69e3d371f8e2f075ad25.jpg) +Figure 4. 3-View SfM on ScanNet. We provide the visualisations of unprojected reference frame depth maps predicted by our method for few-view SfM using one reference and 2 supplementary views. Note that we used surface normal prediction network which was only pretrained on HyperSim [40] for this experiment. + +consistent improvement over DeepV2D. Additionally, our method does not have any global prior on relative object positions, unlike DeepV2D. Our approach therefore could be used as an VO / SLAM initialisation mechanism, estimating joint relative poses and geometry. + +# 4.3. Monocular Visual Odometry + +Monocular Visual Odometry requires both accurate pose and geometry estimation to successfully track camera motion across long trajectories. Minor pose estimation inaccuracies can accumulate over time resulting in what is called + +![](images/d5cc49e732b8b85c274699b0a714af00cc62af10092bcaeed63bff256c49a9c4.jpg) +Figure 5. TUM Reconstruction Results. Examples of reconstructions produced by our monocular VO system on the TUM dataset. Each image shows a coloured point cloud of the geometry estimated on an odometry keyframe. + +![](images/b5b1216a0518c73a217ad202abf211c3fcdf98c0a102b61b31bd8d336b4f4ad8.jpg) + +![](images/9ad305d9bf382664072aaaf3b4b98eafa66b75052d7d24b38b5f8f53d7586f6c.jpg) + +![](images/bfce09d04e29f96cfc14063379b589fc2ffed88683671bca5b31e22c65e2f0fe.jpg) +Figure 6. Few-View SfM Depth Estimation Quality. We evaluate the quality of our depth estimation method from an unstructured set of images on the ScanNet dataset. On the $x$ -axis the number of supporting views is shown, while on the $y$ -axis the corresponding depth reconstruction quality metric is reported. We show that the quality of our depth estimation quickly saturates and consistently outperforms its closest competitor, DeepV2D. + +![](images/ba15942e098593b50d20088664b42e89400cb9c1ad254bbe2496efa5955f7088.jpg) + +drift. With geometry being incorrectly estimated, accurate pose tracking becomes impossible, and vice versa. Our SuperPrimitive representation allows estimating both pose and geometry, enabling us to build a simple monocular VO system which performs better even in hard conditions. + +We evaluate our monocular odometry on the TUM RGB-D [46] dataset. The dataset was captured with a handheld camera in indoor scenes. It is renowned for being incredibly challenging (especially for dense reconstruction systems), due to motion blur, rolling shutter artefacts, and abundance of pure rotational motion. We show that thanks to the strong priors encapsulated in our SuperPrimitives, our simple monocular odometry system can handle the TUM dataset without any special treatment (e.g. our method does not involve any special motion blur handling). + +We compare against other VO systems that do not have global bundle adjustment, following the protocol of [10] and evaluate on 8 sequences from the Friedburg 1 split. We use only RGB images as the input to our system and down-sample them to $120 \times 160$ for efficiency purposes. Since our method is purely monocular, the estimated trajectory lacks global scale and we first use $\mathrm{Sim}(3)$ alignment to the + +
Seq.TartanVO [53]DeepV2D [49]DeepFactors [8]DepthCov [10]Ours
3600.1780.2430.1590.1280.173
desk0.1250.1660.1700.0560.085
desk20.1220.3790.2530.0480.108
plant0.2970.2030.3050.2610.153
room0.3330.2460.3640.2570.363
rpy0.0490.1050.0430.0520.055
teddy0.3390.3160.6010.4750.253
xyz0.0620.0640.0350.0560.036
mean0.1880.2150.2410.1670.153
+ +Table 2. Trajectory Estimation Error on TUM. Average Trajectory Error (ATE) is compared against other monocular odometry systems on the TUM Friedburg 1 split. The best and second best results are highlighted in bold and underscored correspondingly. Our method outperforms others in terms of the ATE averaged across all trajectories. + +ground truth scale, following standard practice [4]. Tab. 2 shows that, despite the simplicity of our odometry system, it outperforms all other methods in terms of Average Trajectory Error (ATE) [46], averaged across all trajectories. Additionally, our VO is either the best or second best on five out of eight sequences. Besides quantitative evaluation, we also demonstrate reconstruction results in Fig. 5. + +# 5. Conclusion + +We presented a new representation, SuperPrimitive, which demonstrates how recent advances in building strong single image priors could be incorporated into pose and dense geometry estimation problems. We show that incorporating these priors alleviates the need for sophisticated handcrafted heuristics and paves the way into monocular reconstruction with relative ease. + +Acknowledgments. Research presented in this paper was supported by Dyson Technology Ltd. The authors would like to thank Hidenobu Matsuki, Eric Dexheimer, Riku Murai, and other members of the Dyson Robotics Lab for countless fruitful discussions. + +# References + +[1] Gwangbin Bae, Ignas Budvytis, and Roberto Cipolla. Estimating and exploiting the aleatoric uncertainty in surface normal estimation. In International Conference on Computer Vision (ICCV), 2021. 3, 4 +[2] M. Bloesch, J. Czarnowski, R. Clark, S. Leutenegger, and A. J. Davison. CodeSLAM — learning a compact, optimisable representation for dense visual SLAM. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1, 2, 12 +[3] Aljaž Božić, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. Transformerfusion: Monocular rgb scene reconstruction using transformers. Proc. Neural Information Processing Systems (NeurIPS), 2021. 2 +[4] Carlos Campos, Richard Elvira, Juan J. Gomez, Jose M. M. Montiel, and Juan D. Tardos. ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Transactions on Robotics, 2021. 8 +[5] Xu Cao, Hiroaki Santo, Boxin Shi, Fumio Okura, and Yasuyuki Matsushita. Bilateral normal integration. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 4 +[6] Alejo Concha and Javier Civera. Using superpixels in monocular SLAM. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2014. 2 +[7] G. Cross and A. Zisserman. Quadric Reconstruction from Dual-Space Geometry. In Proceedings of the International Conference on Computer Vision (ICCV), 1998. 2 +[8] J. Czarnowski, T. Laidlow, R. Clark, and A. J. Davison. Deepfactors: Real-time probabilistic dense monocular SLAM. In IEEE Robotics and Automation Letters, 2020. 8, 12 +[9] A. J. Davison, N. D. Molton, I. Reid, and O. Stasse. MonoSLAM: Real-Time Single Camera SLAM. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2007. 2 +[10] Eric Dexheimer and Andrew J. Davison. Learning a depth covariance function. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 8 +[11] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In Neural Information Processing Systems (NeurIPS), 2014. 1 +[12] David Eigen, Christian Puhrsch, and Rob Fergus. Depth Map Prediction from a Single Image using a Multi-Scale Deep Network. In Neural Information Processing Systems (NeurIPS), 2014. 2 +[13] Jakob Engel, Thomas Schoeps, and Daniel Cremers. LSD-SLAM: Large-scale direct monocular SLAM. In Proceedings of the European Conference on Computer Vision (ECCV), 2014. 3 +[14] J. Engel, V. Koltun, and D. Cremers. Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2017. 3, 5 +[15] Jiahui Fu, Yilun Du, Kurran Singh, Joshua B Tenenbaum, and John J Leonard. Neuse: Neural se (3)-equivariant em + +bedding for consistent spatial understanding with objects. In Proceedings of Robotics: Science and Systems (RSS), 2023. 2 +[16] Y. Furukawa, B. Curless, S. M. Seitz, and R. Szeliski. Towards Internet-scale multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2010. 2 +[17] D. Gallup, J.-M. Frahm, and M. Pollefeys. Piecewise planar and non-planar stereo for urban scene reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2010. 2 +[18] Magnus R. Hestenes and Eduard Stiefel. Methods of conjugate gradients for solving linear systems. Journal of research of the National Bureau of Standards, 1952. 4 +[19] M Kaess. Simultaneous localization and mapping with infinite planes. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2015. 2 +[20] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015. 5 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3, 4 +[22] G. Klein and D. W. Murray. Parallel Tracking and Mapping for Small AR Workspaces. In Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR), 2007. 2, 5 +[23] T. Laidlow and A. J. Davison. Simultaneous localisation and mapping with quadric surfaces. In International Conference on 3D Vision (3DV), 2022. 2 +[24] Stefan Leutenegger, Simon Lynen, Michael Bosse, Roland Siegwart, and Paul Furgale. Keyframe-based visual-inertial odometry using nonlinear optimization. International Journal of Robotics Research (IJRR), 2014. 6 +[25] Chen Liu, Jimei Yang, Duygu Ceylan, Ersin Yumer, and Yasutaka Furukawa. Planenet: Piece-wise planar reconstruction from a single rgb image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[26] Chen Liu, Kihwan Kim, Jinwei Gu, Yasutaka Furukawa, and Jan Kautz. Planercnn: 3d plane detection and reconstruction from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[27] Tianlin Liu, Parth T. Agrawal, Allison Chen, Byung-Woo Hong, and A. Wong. Monitored distillation for positive congruent depth completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 7 +[28] B. D. Lucas and T. Kanade. An Iterative Image Registration Technique with an Application to Stereo Vision. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), 1981. 6 +[29] J. McCormac, R. Clark, M. Bloesch, A. J. Davison, and S. Leutenegger. Fusion++: volumetric object-level slam. In Proceedings of the International Conference on 3D Vision (3DV), 2018. 2 + +[30] R. A. Newcombe, S. Lovegrove, and A. J. Davison. DTAM: Dense Tracking and Mapping in Real-Time. In Proceedings of the International Conference on Computer Vision (ICCV), 2011. 1, 2, 5 +[31] Lachlan Nicholson, Michael Milford, and Niko Sunderhauf. Quadricslam: Dual quadrics from object detections as landmarks in object-oriented slam. IEEE Robotics and Automation Letters, 2019. 2 +[32] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[33] Ryosuke Okuta, Yuya Unno, Daisuke Nishino, Shohei Hido, and Crissman Loomis. Copy: A numpy-compatible library for nvidiagpu calculations. In Proceedings of Workshop on Machine Learning Systems (LearningSys) in The Thirty-first Annual Conference on Neural Information Processing Systems (NeurIPS), 2017. 13 +[34] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 7 +[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. In Neural Information Processing Systems (NeurIPS), 2019. 13 +[36] Albert Pumarola, Alexander Vakhitov, Antonio Agudo, Alberto Sanfeliu, and Francese Moreno-Noguer. Pl-slam: Real-time monocular visual slam with points and lines. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2017. 2 +[37] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2, 6, 12 +[38] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2022. 2, 6 +[39] Xiaofeng Ren and Jitendra Malik. Learning a classification model for segmentation. In Proceedings of the International Conference on Computer Vision (ICCV), 2003. 3 +[40] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 7 +[41] R. F. Salas-Moreno, R. A. Newcombe, H. Strasdat, P. H. J. Kelly, and A. J. Davison. SLAM++: Simultaneous Localisation and Mapping at the Level of Objects. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2013. 2 + +[42] Mohamed Sayed, John Gibson, Jamie Watson, Victor Prisacariu, Michael Firman, and Clément Godard. Simplerecon: 3d reconstruction without 3d convolutions. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2 +[43] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2 +[44] Jonathan Shade, Steven Gortler, Li-wei He, and Richard Szeliski. Layered depth images. In Proceedings of the 25th annual conference on Computer graphics and interactive techniques, 1998. 12 +[45] Jingjia Shi, Shuaifeng Zhi, and Kai Xu. Planerectr: Unified query learning for 3d plane recovery from a single view. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2 +[46] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A Benchmark for the Evaluation of RGB-D SLAM Systems. In Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS), 2012. 8, 13 +[47] E. Sucar, K. Wada, and A. J. Davison. NodeSLAM: Neural object descriptors for multi-view shape reconstruction. In Proceedings of the International Conference on 3D Vision (3DV), 2020. 2 +[48] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. NeuralRecon: Real-time coherent 3D reconstruction from monocular video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[49] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. In Proceedings of the International Conference on Learning Representations (ICLR), 2020. 7, 8 +[50] Alexander Vakhitov, Jan Funke, and Francesc Moreno-Noguer. Accurate and linear time pose estimation from points and lines. In Proceedings of the European Conference on Computer Vision (ECCV), 2016. 2 +[51] Tom van Dijk and Guido C.H.E. de Croon. How do neural networks see depth in single images? In Proceedings of the International Conference on Computer Vision (ICCV), 2019. +[52] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1 +[53] Wenshan Wang, Yaoyu Hu, and Sebastian Scherer. TartanVO: A Generalizable Learning-based VO. In Conference on Robot Learning (CoRL), 2020. 8 +[54] C. S. Weerasekera, Y. Latif, R. Garg, and I. Reid. Dense monocular reconstruction using surface normals. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2017. 2 +[55] Wofk, Diana and Ranftl, René and Müller, Matthias and Koltun, Vladlen. Monocular Visual-Inertial Depth Estimation. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2023. 6, 7, 12 + +[56] Alex Wong and Stefano Soatto. Unsupervised depth completion with calibrated backprojection layers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7 +[57] Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters, 2020. 6, 7 +[58] Alex Wong, Safa Cicek, and Stefano Soatto. Learning topology from synthetic data for unsupervised depth completion. IEEE Robotics and Automation Letters, 2021. 7 +[59] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In Proceedings of SIGGRAPH, 2018. 12 \ No newline at end of file diff --git a/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/images.zip b/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a8485d0b83c35432e18dd9ec3baf8b51978f0c84 --- /dev/null +++ b/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56b3e9d3db08c6829c4fbaa14c72e409c1716fa7b16857107d524e6c937ab9bc +size 473104 diff --git a/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/layout.json b/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9ace3bb790a968074ababbea6d3a1ea014d9888d --- /dev/null +++ b/2024/SuperPrimitive_ Scene Reconstruction at a Primitive Level/layout.json @@ -0,0 +1,9383 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 103, + 477, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 103, + 477, + 120 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 477, + 120 + ], + "type": "text", + "content": "SuperPrimitive: Scene Reconstruction at a Primitive Level" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 165, + 143, + 427, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 143, + 427, + 171 + ], + "spans": [ + { + "bbox": [ + 165, + 143, + 427, + 171 + ], + "type": "text", + "content": "Kirill Mazur, Gwangbin Bae, Andrew J. Davison Dyson Robotics Lab, Imperial College London" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 176, + 174, + 414, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 174, + 414, + 185 + ], + "spans": [ + { + "bbox": [ + 176, + 174, + 414, + 185 + ], + "type": "text", + "content": "{k.mazur21, g.bae, a.davison}@imperial.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 238, + 288, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 288, + 345 + ], + "type": "text", + "content": "Joint camera pose and dense geometry estimation from a set of images or a monocular video remains a challenging problem due to its computational complexity and inherent visual ambiguities. Most dense incremental reconstruction systems operate directly on image pixels and solve for their 3D positions using multi-view geometry cues. Such pixel-level approaches suffer from ambiguities or violations of multi-view consistency (e.g. caused by textureless or specular surfaces)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 346, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 441 + ], + "type": "text", + "content": "We address this issue with a new image representation which we call a SuperPrimitive. SuperPrimitives are obtained by splitting images into semantically correlated local regions and enhancing them with estimated surface normal directions, both of which are predicted by state-of-the-art single image neural networks. This provides a local geometry estimate per SuperPrimitive, while their relative positions are adjusted based on multi-view observations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 441, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 501 + ], + "type": "text", + "content": "We demonstrate the versatility of our new representation by addressing three 3D reconstruction tasks: depth completion, few-view structure from motion, and monocular dense visual odometry. Project page: https://makezur.github.io/SuperPrimitive/" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 514, + 128, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 128, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 128, + 526 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 534, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 605 + ], + "type": "text", + "content": "Enriching monocular incremental reconstruction with prior world knowledge is essential for resolving visual ambiguities. This issue is particularly prevalent in scenarios with scarce data observations available: a notable example would be monocular visual SLAM, where images are being streamed from a camera into the system in real-time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "content": "When a monocular vision system encounters a new scene region, it must estimate the region's geometry based on a very limited number of observations. Without this, continuous camera motion tracking would not be possible. Once the scene region is thoroughly observed, the initial geometry estimate should be refined to better explain the multiview information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "This naturally leads to a question: what sort of priors are effective in both providing reliable initial geometry es" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 312, + 217, + 533, + 466 + ], + "blocks": [ + { + "bbox": [ + 312, + 217, + 533, + 466 + ], + "lines": [ + { + "bbox": [ + 312, + 217, + 533, + 466 + ], + "spans": [ + { + "bbox": [ + 312, + 217, + 533, + 466 + ], + "type": "image", + "image_path": "374776498f3c7f0c4217067d977995de1c873b9ec63f7fe2fce7212e571eb247.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 469, + 545, + 579 + ], + "lines": [ + { + "bbox": [ + 304, + 469, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 545, + 579 + ], + "type": "text", + "content": "Figure 1. Multi-View Geometry with SuperPrimitives. SuperPrimitives are extracted from an input frame by dividing it into image segments equipped with estimated surface normal directions (bottom-left). Each SuperPrimitive induces a dense reconstruction within the corresponding image segment up to a priori unknown scale. Different possible reconstructions are shown in light blue. The scales are then jointly optimised together with a relative camera pose to fit multi-view photometric constraints (visualised in green and red). The resulting dense reconstruction of the reference frame is shown in the top." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 590, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 545, + 674 + ], + "type": "text", + "content": "timates and supporting multi-view consistency? Geometric priors generally fall into one of two categories: local and global. Local priors, such as smoothness assumption [30] or surface normal regularisation [52], impose additional constraints within a small neighbourhood. Global priors, on the other hand, aim to impose constraints on a larger scale, such as depth prediction [2, 11]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Our key observation is that some of the geometrical correlations are more reliable than the others, and therefore could be safely \"locked in\" together within local regions" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4979" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "based on a single-view prediction. Points belonging to the same rigid body are strongly correlated, making it unnecessary to determine their depth independently. In contrast, distinct and unrelated objects can be placed arbitrarily in the scene. As the number of objects increases, learning a reliable global prior on their relative positions becomes an increasingly complex problem." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 287, + 263 + ], + "type": "text", + "content": "In this work, we show that purely local but strong priors are enough to achieve excellent performance across a variety of geometric vision tasks. For that purpose, we introduce a novel representation, SuperPrimitives. A Super-Primitive represents a local image segment, coupled with a dense shape estimate which is determined up to a scale factor. The scale factor can be further adjusted based on information observed from other views or additional measurements." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "text", + "content": "We show that SuperPrimitives can be efficiently constructed using a front-end which consists of two single-image neural networks, extracting image segmentation and surface normal prediction. Effectively, our neural front-end predicts whether adjacent pixels belong to the same geometrical entity through image segmentation and estimates a surface normal at this point, thereby providing an infinitesimal geometry estimate." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 286, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 286, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 286, + 421 + ], + "type": "text", + "content": "We delegate global, scene-level alignment to our streamlined multi-view, iterative, optimisation-based back-end. The resulting front-end / back-end tandem combines the flexibility of multi-view based optimisation methods with the observation efficiency common in prior-driven systems." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 421, + 286, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 286, + 456 + ], + "type": "text", + "content": "Our new representation showcases its versatility in three key applications, where single-view ambiguity is resolved via additional measurements or viewpoints:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 457, + 286, + 564 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 457, + 286, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 457, + 286, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 457, + 286, + 493 + ], + "type": "text", + "content": "- Firstly, it adeptly handles zero-shot depth completion tasks in real-world scenarios, matching the performance of state-of-the-art methods tailored for depth completion;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 494, + 286, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 494, + 286, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 494, + 286, + 529 + ], + "type": "text", + "content": "- Secondly, it facilitates joint pose and depth estimation using a limited set of unstructured images, surpassing its nearest competitor even in the absence of global priors;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 529, + 286, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 286, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 286, + 564 + ], + "type": "text", + "content": "- Thirdly, our method outperforms previous monocular visual odometry systems on the challenging TUM dataset, and exhibits robustness across various domains." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 573, + 133, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 133, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 133, + 585 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "Monocular Reconstruction. Both offline monocular reconstruction systems, such as COLMAP [43] and [16], or online systems, such as MonoSLAM [9] and DSO [22] track, filter, and reconstruct only well-constrained points with high and reliable photometric information. This often involves fine-grained and sophisticated point management to reliably resolve visual ambiguities and filter unreliable visual observations. DTAM [30] demonstrated feasibility of incremental dense reconstruction in the monocular scenario. DTAM employed a hand-crafted local smoothness" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "prior to handle regions with poor texture. Subsequently, other local priors [32, 54] have been extensively explored to regularise multi-view geometry estimation problems." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 110, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 289 + ], + "type": "text", + "content": "Global Priors. We are interested in exploring the space of possible geometric priors from single-view networks. Depth prediction [12, 37, 38] is the most obvious choice but leads to a rigid per-image reconstruction which is difficult to feed into multi-view optimisation. In the recent years, deep-learning based approaches sought to replace explicit multi-view geometry estimating with learning-based methods [3, 42, 48]. These methods, however, assume known poses and are, therefore, not suitable for joint pose and geometry estimation. Notably, CodeSLAM [2] introduced depth prediction conditioned on latent codes, which are then optimised to achieve cross-view consistency. However, CodeSLAM still struggles with out-of-domain data, as depth prediction networks are known to struggle with generalisation [51]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 291, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 545, + 374 + ], + "type": "text", + "content": "Higher-Level Mapping. Introducing parametric primitives, such as lines [36, 50], planes [6, 17, 19, 25, 26, 45] or even high-order algebraic shapes [7, 23, 31] to better constrain multi-view geometry problems have been thoroughly explored over the last few decades. They all however use assumptions which may not often hold for all 3D scenes, especially for dense reconstruction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 375, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 545, + 567 + ], + "type": "text", + "content": "Besides parametric algebraic primitives object-level mapping has been explored in the last decade. SLAM++ [41] represented a map with a set of CAD models retrieved from an existing database and tracked camera position against these models. This method, however, could not be applied in a variety of settings due to the limited size of the CAD databases. Even objects of the same class can vary in their geometric appearance from instance to instance. This was later approached by [15, 47], who also represent their map via a set of objects, but learn a latent space of possible geometric variations per object class. Fusion++ [29] extended this approach into a more versatile per object depth fusion using an RGB-D sensor and mask proposals from a pre-trained neural network. These methods rely on an RGB-D inputs and assume a pre-defined set of object classes, therefore lacking generality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 574, + 361, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 361, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 361, + 586 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": "Firstly, we introduce the concept of a SuperPrimitive and explain how an image is processed into a set of its SuperPrimitives, referring to this part of our method as the front-end. In the second part of this section, we describe how multi-view geometry problems can be reformulated at the level of SuperPrimitives instead of pixels. We refer to this stage of SuperPrimitive alignment as the back-end." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "The core of our method proposes splitting a given image into a set of (possibly overlapping) minimal segments, image regions which are likely to have strongly correlated" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4980" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 160, + 247 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 160, + 247 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 160, + 247 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 160, + 247 + ], + "type": "image", + "image_path": "2394536898c6d7b5ebe1603afd03c5680681346861012fb5086549ac85b35474.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 250, + 547, + 305 + ], + "lines": [ + { + "bbox": [ + 46, + 250, + 547, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 547, + 305 + ], + "type": "text", + "content": "Figure 2. SuperPrimitives Extraction. (left) Our front-end processor extracts SuperPrimitives from an image by dividing it into a set of image regions with surface normal directions estimated for each image pixel within the segment. (right) Highlighted SuperPrimitives extracted from the image are visualised by showing their estimated normal and colour maps side by side. Note some of them are scaled either up or down for better viewing. While some of the SuperPrimitives are akin to object-level segmentation, the others tend to represent more low-level image segments." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 161, + 71, + 541, + 246 + ], + "blocks": [ + { + "bbox": [ + 161, + 71, + 541, + 246 + ], + "lines": [ + { + "bbox": [ + 161, + 71, + 541, + 246 + ], + "spans": [ + { + "bbox": [ + 161, + 71, + 541, + 246 + ], + "type": "image", + "image_path": "29acfc2df01382d0829f980c75c3a252ad07ed79291b184ab759c9f6ccd96a22.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 312, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 348 + ], + "type": "text", + "content": "geometry. We repurpose the recent state-of-the-art Segment Anything (SAM) [21] model into predicting these minimal segments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 349, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 287, + 432 + ], + "type": "text", + "content": "Our key idea is to estimate local geometry within each segment from a single view, while leaving the relative positioning of the segments to be estimated via multi-view photometric consistency optimisation. We refer to these geometrically enhanced minimal image segments as Super-Primitives, since they are inspired by both superpixels [39] and geometric 3D primitives." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 433, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 540 + ], + "type": "text", + "content": "For per-segment local geometry estimation we employ an off-the-shelf surface normal prediction network [1], to estimate infinitesimal geometry for each image pixel. The surface normals within an image segment can be used to estimate its depth via simple integration up to a scale factor. We set these scale factors — depth scales — to be optimisable parameters, which are either optimised via multi-view cues or explicit depth measurements for the depth completion experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 540, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 588 + ], + "type": "text", + "content": "Thus, our method combines the strong priors provided by state-of-the-art neural networks in the front-end with the flexibility and consistency offered by multi-view optimisation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 591, + 128, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 128, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 128, + 602 + ], + "type": "text", + "content": "3.1. Conventions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "text", + "content": "Unless stated otherwise, we use lowercase letters for scalar values, bold letters for vectors, and uppercase letters for matrices. We consider images " + }, + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{3 \\times h \\times w}" + }, + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "text", + "content": " of height " + }, + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "text", + "content": " captured by a camera with a known calibration matrix " + }, + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{K} = \\left( \\begin{array}{ccc} f_u & 0 & c_u \\\\ 0 & f_v & c_v \\\\ 0 & 0 & 1 \\end{array} \\right) \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "text", + "content": ". Image pixels are parameterised as " + }, + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = (u, v) \\in [0, h - 1] \\times [0, w - 1]" + }, + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "text", + "content": ". Given a per pixel depth function " + }, + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "inline_equation", + "content": "z(\\mathbf{u})" + }, + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\to \\mathbb{R}^+" + }, + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "text", + "content": ", we define image's unprojection onto the 3D space to be " + }, + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\pi^{-1}(\\mathbf{u}) = z(\\mathbf{u})\\mathbf{K}^{-1}\\dot{\\mathbf{u}}" + }, + { + "bbox": [ + 46, + 609, + 288, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "text", + "content": "where dot is a homogenisation operator " + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\dot{\\pmb{p}} = (\\pmb{p},1)" + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "text", + "content": ". Conversely, for a set of 3D points " + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{x} = (x,y,z) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "text", + "content": " we define the projection " + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\pi(\\pmb{x}) = \\gamma\\left(\\frac{1}{z} K\\pmb{x}\\right)" + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "text", + "content": ", where the " + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "text", + "content": " function drops the last coordinate " + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\gamma((x,y,z)^T) = (x,y)^T" + }, + { + "bbox": [ + 304, + 312, + 545, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 361, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 433 + ], + "type": "text", + "content": "We represent camera poses as matrices " + }, + { + "bbox": [ + 304, + 361, + 545, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_{WC} \\in \\mathrm{SE}(3)" + }, + { + "bbox": [ + 304, + 361, + 545, + 433 + ], + "type": "text", + "content": " in the camera-to-world coordinates unless stated otherwise. During pose optimisation we store linearised pose increments " + }, + { + "bbox": [ + 304, + 361, + 545, + 433 + ], + "type": "inline_equation", + "content": "\\xi \\in \\mathfrak{se}(3)" + }, + { + "bbox": [ + 304, + 361, + 545, + 433 + ], + "type": "text", + "content": " as Lie algebra elements. These increments are used then to update current camera pose estimates, see [13, 14] for more detail." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 441, + 405, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 405, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 405, + 453 + ], + "type": "text", + "content": "3.2. SuperPrimitives" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "content": "We split the input image I into a set of possibly overlapping connected regions (segments), each of which is equipped with its local 3D geometry. More formally, a SuperPrimitive " + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "inline_equation", + "content": "P = (\\Omega, \\mathfrak{D})" + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "content": " is a connected image region " + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\Omega \\subseteq I" + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "content": " which is also equipped with an unscaled depth map " + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathfrak{D} : \\Omega \\to [0, \\infty)" + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "content": " for each pixel within the region " + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "content": ". Here, unscaled means that each SuperPrimitive's depth " + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathfrak{D}" + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "content": " differs from its ground truth depth " + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "content": " by an a priori unknown scalar. In other words, there exists a single scalar " + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "content": " such that:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 344, + 570, + 545, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 570, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 344, + 570, + 545, + 583 + ], + "type": "interline_equation", + "content": "D = s \\cdot \\mathfrak {D} \\quad \\text {f o r a l l p i x e l s} \\mathrm {p} \\in \\Omega \\tag {1}", + "image_path": "2d537606b84ad6a7936b25409270bcc5b14225544f0ce4e94163bd9361c33607.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 586, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 545, + 622 + ], + "type": "text", + "content": "We say that an image " + }, + { + "bbox": [ + 304, + 586, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathrm{I}" + }, + { + "bbox": [ + 304, + 586, + 545, + 622 + ], + "type": "text", + "content": " is primified into super-primitives " + }, + { + "bbox": [ + 304, + 586, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\mathrm{I}) = \\{P_i = (\\Omega_i, \\mathfrak{D}_i)\\}" + }, + { + "bbox": [ + 304, + 586, + 545, + 622 + ], + "type": "text", + "content": " if it has " + }, + { + "bbox": [ + 304, + 586, + 545, + 622 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 586, + 545, + 622 + ], + "type": "text", + "content": ", possibly overlapping primitives that lie within the image " + }, + { + "bbox": [ + 304, + 586, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathrm{I}" + }, + { + "bbox": [ + 304, + 586, + 545, + 622 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 304, + 586, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\bigcup \\Omega_i \\subseteq \\mathrm{I}" + }, + { + "bbox": [ + 304, + 586, + 545, + 622 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 623, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 623, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 623, + 545, + 647 + ], + "type": "text", + "content": "For brevity we use the words primitive and SuperPrimitive interchangeably throughout the rest of the paper." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Representing Geometry with SuperPrimitives. A set of primitives " + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": " itself is not enough to extract dense image geometry due to scale/depth ambiguity. Therefore, we introduce the concept of optimisable depth scales which anchor the correct depth scale for each primitive." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4981" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": "Given a scalar depth scale " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": " for a super-primitive " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": ", one can infer its depth as " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "D(p) = s \\cdot \\mathfrak{D}(p)" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": " for every point " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "p \\in \\Omega" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": ". In practice, we employ a log-depth representation to represent depth throughout the whole system. This means that we store optimisable log-depth scales " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "\\log s" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": " and unscaled log-depth " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "\\log \\mathfrak{D}" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": ", and so depth inference reduces to a simple shift operation " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "\\log D = \\log s + \\log \\mathfrak{D}" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 286, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 286, + 228 + ], + "type": "text", + "content": "In contrast to other dense methods, we estimate the dense geometry " + }, + { + "bbox": [ + 46, + 156, + 286, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 156, + 286, + 228 + ], + "type": "text", + "content": " of the image in the form of a point cloud rather than a depth image. This choice is driven by the fact that multiple primitives' supporting segments " + }, + { + "bbox": [ + 46, + 156, + 286, + 228 + ], + "type": "inline_equation", + "content": "\\Omega_{i}" + }, + { + "bbox": [ + 46, + 156, + 286, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 156, + 286, + 228 + ], + "type": "inline_equation", + "content": "\\Omega_{j}" + }, + { + "bbox": [ + 46, + 156, + 286, + 228 + ], + "type": "text", + "content": " might overlap, which would lead to different depth estimates within their intersection " + }, + { + "bbox": [ + 46, + 156, + 286, + 228 + ], + "type": "inline_equation", + "content": "\\Omega_{i} \\cap \\Omega_{j}" + }, + { + "bbox": [ + 46, + 156, + 286, + 228 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 229, + 286, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 229, + 286, + 253 + ], + "spans": [ + { + "bbox": [ + 47, + 229, + 286, + 253 + ], + "type": "text", + "content": "Given a set of depth scaled primitives " + }, + { + "bbox": [ + 47, + 229, + 286, + 253 + ], + "type": "inline_equation", + "content": "\\mathbb{P} = \\{(s_i, P_i)\\}" + }, + { + "bbox": [ + 47, + 229, + 286, + 253 + ], + "type": "text", + "content": ", the geometry " + }, + { + "bbox": [ + 47, + 229, + 286, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 47, + 229, + 286, + 253 + ], + "type": "text", + "content": " of the image " + }, + { + "bbox": [ + 47, + 229, + 286, + 253 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 229, + 286, + 253 + ], + "type": "text", + "content": " can be estimated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 265, + 286, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 265, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 121, + 265, + 286, + 289 + ], + "type": "interline_equation", + "content": "\\mathcal {G} := \\bigcup_ {i} \\pi^ {- 1} \\left(s _ {i} \\cdot \\mathfrak {D} _ {i}\\right) \\tag {2}", + "image_path": "a4bb5aa5f02866560b3fbc0d19b444100dc380ff1f34ecc9620953efc5d486c3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 293, + 286, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 286, + 318 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 286, + 318 + ], + "type": "text", + "content": "To obtain estimated depth " + }, + { + "bbox": [ + 47, + 293, + 286, + 318 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 47, + 293, + 286, + 318 + ], + "type": "text", + "content": ", we average depth values along the corresponding camera rays." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 324, + 286, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 286, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 286, + 443 + ], + "type": "text", + "content": "Converting an Image into SuperPrimitives. Next we explain how SuperPrimitives are obtained in practice. An input image " + }, + { + "bbox": [ + 46, + 324, + 286, + 443 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 324, + 286, + 443 + ], + "type": "text", + "content": " is first split into a set of minimal image segments " + }, + { + "bbox": [ + 46, + 324, + 286, + 443 + ], + "type": "inline_equation", + "content": "\\Omega_{i}" + }, + { + "bbox": [ + 46, + 324, + 286, + 443 + ], + "type": "text", + "content": " produced by an image segmentation model. We use the Segment Anything [21] model in this work, since it captures low-level semantic correlations highly accurately. Given the extracted segments, we independently estimate the local geometry " + }, + { + "bbox": [ + 46, + 324, + 286, + 443 + ], + "type": "inline_equation", + "content": "\\mathfrak{D}_i" + }, + { + "bbox": [ + 46, + 324, + 286, + 443 + ], + "type": "text", + "content": " within each segment by integrating surface normals predicted by another state-of-the-art neural network [1] within this region." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 445, + 286, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 286, + 540 + ], + "type": "text", + "content": "Note that this approach is premised on the assumption of geometry being continuous within each predicted image segment " + }, + { + "bbox": [ + 46, + 445, + 286, + 540 + ], + "type": "inline_equation", + "content": "\\Omega_{i}" + }, + { + "bbox": [ + 46, + 445, + 286, + 540 + ], + "type": "text", + "content": ", hence its unscaled depth values could be obtained by integrating its surface normals. Although this is not guaranteed a priori, we observed that a more fine-grained mask selection yields a compelling correlation with geometrical continuity. We discuss this in more detail in the following section." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 546, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 286, + 713 + ], + "type": "text", + "content": "Image Segment Retrieval. An ideal neural network which estimates image segments " + }, + { + "bbox": [ + 46, + 546, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\{\\Omega_i\\}" + }, + { + "bbox": [ + 46, + 546, + 286, + 713 + ], + "type": "text", + "content": " should predict regions of geometric continuity, where depth could be obtained via surface normal integration. However, to the best of our knowledge, such a network does not exist. Our approach aims to approximate this behaviour by utilising the Segment Anything model, coupled with a specialised mask selection process. Specifically, for each SAM query, we strive to select the smallest predicted mask surrounding that point. While such over-segmentation could potentially increase the dimensionality of the multi-view optimisation problem, under-segmentation might lead to incorrect geometry estimate within the primitive. Since we only adjust the scale of a primitive after the normal integration is com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": "pleted, any incorrect geometry estimates within a primitive cannot be compensated at this stage." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 96, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 545, + 215 + ], + "type": "text", + "content": "The segmentation model employed in this work outputs three binary segmentation masks given a query point " + }, + { + "bbox": [ + 304, + 96, + 545, + 215 + ], + "type": "inline_equation", + "content": "q \\in \\mathbf{I}" + }, + { + "bbox": [ + 304, + 96, + 545, + 215 + ], + "type": "text", + "content": ". For each query " + }, + { + "bbox": [ + 304, + 96, + 545, + 215 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 96, + 545, + 215 + ], + "type": "text", + "content": " we first filter predicted masks using the post-processing introduced in [21], such as stability filtering and Non-Maximum-Suppression (NMS). Even though our method allows redundant regions, we employed NMS to remove similar segments and save compute in the image alignment stage. With filtering done, the mask with the smallest area is selected. If the filtered set is empty, we discard the query point." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 216, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 545, + 287 + ], + "type": "text", + "content": "We query the SAM backbone feature extractor once per image. Then we first sample 300 query points randomly across the image, followed by the filtering discussed above. Then the image coverage mask is calculated and an additional 100 mask query points are actively sampled in the uncovered regions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 289, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 361 + ], + "type": "text", + "content": "Normal Integration. To estimate each segment's local geometry, we first pass the image through a surface normal estimation network. This network effectively predicts the derivative of desired depth values. Then, we \"integrate\" these surface normal vectors within each image segment to obtain its unscaled depth map." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 361, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 420 + ], + "type": "text", + "content": "For each image pixel " + }, + { + "bbox": [ + 304, + 361, + 545, + 420 + ], + "type": "inline_equation", + "content": "\\boldsymbol{u} = (u,v)^T \\in \\mathrm{I}" + }, + { + "bbox": [ + 304, + 361, + 545, + 420 + ], + "type": "text", + "content": " we estimate its surface normal vector " + }, + { + "bbox": [ + 304, + 361, + 545, + 420 + ], + "type": "inline_equation", + "content": "\\boldsymbol{n} = (n_x, n_y, n_z)^T" + }, + { + "bbox": [ + 304, + 361, + 545, + 420 + ], + "type": "text", + "content": " with a pretrained state-of-the-art CNN [1]. As was demonstrated in [5], the log-depth " + }, + { + "bbox": [ + 304, + 361, + 545, + 420 + ], + "type": "inline_equation", + "content": "\\tilde{z} = \\log(z(\\boldsymbol{u}))" + }, + { + "bbox": [ + 304, + 361, + 545, + 420 + ], + "type": "text", + "content": " satisfies the following PDEs within a segment " + }, + { + "bbox": [ + 304, + 361, + 545, + 420 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 304, + 361, + 545, + 420 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 324, + 423, + 545, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 423, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 324, + 423, + 545, + 436 + ], + "type": "interline_equation", + "content": "\\tilde {n} _ {z} \\partial_ {u} \\tilde {z} + n _ {x} = 0 \\quad \\text {a n d} \\quad \\tilde {n} _ {z} \\partial_ {v} \\tilde {z} + n _ {y} = 0 \\tag {3}", + "image_path": "3936c935ca160a4a314b0f21c2b9681ee2c11d9c48618dd4a7aa2b9e2e8c2756.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 438, + 544, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 544, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 544, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 438, + 544, + 474 + ], + "type": "inline_equation", + "content": "\\tilde{n}_z = n_x(u - c_u) + n_y(v - c_v) + n_zf" + }, + { + "bbox": [ + 304, + 438, + 544, + 474 + ], + "type": "text", + "content": ". The depth values within the segment " + }, + { + "bbox": [ + 304, + 438, + 544, + 474 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 304, + 438, + 544, + 474 + ], + "type": "text", + "content": " can therefore be obtained via minimising the following functional:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 475, + 545, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 475, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 320, + 475, + 545, + 506 + ], + "type": "interline_equation", + "content": "\\min _ {z} \\iint_ {\\Omega} \\left(\\tilde {n} _ {z} \\partial_ {u} \\tilde {z} + n _ {x}\\right) ^ {2} + \\left(\\tilde {n} _ {z} \\partial_ {u} \\tilde {z} + n _ {y}\\right) ^ {2} d u d v \\tag {4}", + "image_path": "95cd6af576d801cb4d657c59d7283c69dee24b9f5037ebebe0f2554c7f801061.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 509, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 568 + ], + "type": "text", + "content": "Note that this leads to a family of solutions " + }, + { + "bbox": [ + 304, + 509, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\tilde{z} + \\text{Const}" + }, + { + "bbox": [ + 304, + 509, + 545, + 568 + ], + "type": "text", + "content": " which differ by a shift constant. This is due to the fact that only partial derivatives of " + }, + { + "bbox": [ + 304, + 509, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 304, + 509, + 545, + 568 + ], + "type": "text", + "content": " are used in the functional. That means that the actual depth will be estimated up to scale, since " + }, + { + "bbox": [ + 304, + 509, + 545, + 568 + ], + "type": "inline_equation", + "content": "z = \\exp(\\tilde{z})" + }, + { + "bbox": [ + 304, + 509, + 545, + 568 + ], + "type": "text", + "content": " by definition." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "text", + "content": "This ambiguity prompted us to introduce the notion of depth scales, to allow each segment to be adjusted towards its true depth. We implemented batched normal integration, efficiently solving all optimisation problems as a single sparse linear system using conjugate gradient method [18]. This implementation enables the integration of approximately 100 to 300 segments within a total time frame of " + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\sim 100\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 671, + 486, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 486, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 486, + 684 + ], + "type": "text", + "content": "3.3. Primitive-based Image Alignment" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Now we explain how our new representation can be used as a building block for dense multi-view geometry and pose" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4982" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": "estimation, combining an optimisation-based mindset with learned single view priors. To make optimisation computationally tractable, we design our image alignment to fully abstract away any knowledge about the neural networks involved in the image primification stage. Hence, even though there has been early evidence that SAM is independently capable of establishing (coarse) mask correspondences across neighbouring frames, we rely on photometric information only during the image alignment stage. We believe that more semantic primitive association would be \"heavier\" and coarser since it would not provide per pixel correspondences, and would therefore not be a good fit to be used in a multi-view optimisation loop. However, we speculate that such a method could be employed to further enhance our system, e.g. to improve occlusion handling, particularly as these models become more computationally affordable." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 275, + 228, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 228, + 288 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 228, + 288 + ], + "type": "text", + "content": "3.3.1 Two-view SfM on SuperPrimitives" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 294, + 289, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 289, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 289, + 510 + ], + "type": "text", + "content": "For the sake of clarity, we formulate our method in the simplest case of two frames observing the scene, although it could be trivially extended into a setting with a higher number of views. At the core of our Structure-from-Motion (SfM) approach lies a per-primitive photometric alignment, which contrasts with widely accepted per-pixel photometric alignment [14, 30] techniques. Informally speaking, our proposition is to treat each image primitive as a \"rigid\" piece, which is only allowed to be scaled up or down. This degree of freedom is due to the scale ambiguity present for each primitive. Thus, instead of estimating a depth value per pixel, we only estimate a depth-scale per primitive, which is illustrated in Fig. 1. This greatly reduces the dimensionality of the optimisation problem, especially in the case of an unknown relative pose. Note that our method does not require the target image to be primified nor does it not require any pre-established correspondences (e.g. primitive to primitive)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 513, + 288, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 513, + 288, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 513, + 288, + 574 + ], + "type": "text", + "content": "We assume a primified reference image " + }, + { + "bbox": [ + 46, + 513, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 513, + 288, + 574 + ], + "type": "text", + "content": " with its set of primitives " + }, + { + "bbox": [ + 46, + 513, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\mathrm{I}) = \\{(\\Omega_i, \\mathfrak{D}_i)\\}" + }, + { + "bbox": [ + 46, + 513, + 288, + 574 + ], + "type": "text", + "content": ". Given an unposed target image " + }, + { + "bbox": [ + 46, + 513, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{t}}" + }, + { + "bbox": [ + 46, + 513, + 288, + 574 + ], + "type": "text", + "content": ", our goal is to jointly estimate its relative pose " + }, + { + "bbox": [ + 46, + 513, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_{tr}" + }, + { + "bbox": [ + 46, + 513, + 288, + 574 + ], + "type": "text", + "content": " with respect to the reference image " + }, + { + "bbox": [ + 46, + 513, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 513, + 288, + 574 + ], + "type": "text", + "content": " as well as the dense geometry " + }, + { + "bbox": [ + 46, + 513, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathrm{ref}} = \\bigcup_{i} \\pi^{-1}(s_i \\cdot \\mathfrak{D}_i)" + }, + { + "bbox": [ + 46, + 513, + 288, + 574 + ], + "type": "text", + "content": " of the reference frame." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 575, + 288, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 288, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 288, + 671 + ], + "type": "text", + "content": "While unscaled depths " + }, + { + "bbox": [ + 46, + 575, + 288, + 671 + ], + "type": "inline_equation", + "content": "\\mathfrak{D}_i" + }, + { + "bbox": [ + 46, + 575, + 288, + 671 + ], + "type": "text", + "content": " are given by a per-image preprocessor after the primification of the reference image, the set of depth scales " + }, + { + "bbox": [ + 46, + 575, + 288, + 671 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 575, + 288, + 671 + ], + "type": "text", + "content": " are yet to be estimated. In our case, we jointly estimate both depth scales " + }, + { + "bbox": [ + 46, + 575, + 288, + 671 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 575, + 288, + 671 + ], + "type": "text", + "content": " and a relative image pose " + }, + { + "bbox": [ + 46, + 575, + 288, + 671 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_{tr}" + }, + { + "bbox": [ + 46, + 575, + 288, + 671 + ], + "type": "text", + "content": " by solving a photometric consistency optimisation problem. First, we warp each depth-scaled primitive " + }, + { + "bbox": [ + 46, + 575, + 288, + 671 + ], + "type": "inline_equation", + "content": "(s_i, P_i)" + }, + { + "bbox": [ + 46, + 575, + 288, + 671 + ], + "type": "text", + "content": " from the reference frame " + }, + { + "bbox": [ + 46, + 575, + 288, + 671 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 575, + 288, + 671 + ], + "type": "text", + "content": " into the target frame " + }, + { + "bbox": [ + 46, + 575, + 288, + 671 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{t}}" + }, + { + "bbox": [ + 46, + 575, + 288, + 671 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 673, + 287, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 673, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 100, + 673, + 287, + 689 + ], + "type": "interline_equation", + "content": "\\hat {P} _ {i} [ \\boldsymbol {u} ] = \\pi \\left(\\mathrm {T} _ {t r} \\pi^ {- 1} \\left(\\boldsymbol {u}, s _ {i} \\cdot \\mathfrak {D} _ {i}\\right)\\right) \\tag {5}", + "image_path": "603bf5eb72166eca8cfd6e88b332c85d88305e09fe550cbb0f44411d9aba60b1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Then a per-segment photometric residual for the primitive " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "P_{i} = (\\Omega_{i},\\mathfrak{D}_{i})" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": " is defined by averaging all photometric re" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 495, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 495, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 495, + 85 + ], + "type": "text", + "content": "projection " + }, + { + "bbox": [ + 305, + 72, + 495, + 85 + ], + "type": "inline_equation", + "content": "\\ell^1" + }, + { + "bbox": [ + 305, + 72, + 495, + 85 + ], + "type": "text", + "content": " errors across every pixel " + }, + { + "bbox": [ + 305, + 72, + 495, + 85 + ], + "type": "inline_equation", + "content": "\\pmb{u} \\in \\Omega_i" + }, + { + "bbox": [ + 305, + 72, + 495, + 85 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 93, + 546, + 124 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 93, + 546, + 124 + ], + "spans": [ + { + "bbox": [ + 313, + 93, + 546, + 124 + ], + "type": "interline_equation", + "content": "r \\left(P _ {i}, s _ {i}, I _ {\\mathrm {t}}, T _ {t r}\\right) = \\frac {1}{\\left| \\Omega_ {i} \\right|} \\sum_ {\\boldsymbol {u} \\in \\Omega_ {i}} \\left\\| I _ {\\text {r e f}} (\\boldsymbol {u}) - I _ {\\mathrm {t}} \\left(\\hat {P} _ {i} [ \\boldsymbol {u} ]\\right) \\right\\| _ {1} \\tag {6}", + "image_path": "a1abf266ad14d4d22d8334487c41f3dcddabdfdcecb5e1d9dc72810fc3e998b1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 132, + 545, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 132, + 545, + 157 + ], + "spans": [ + { + "bbox": [ + 305, + 132, + 545, + 157 + ], + "type": "text", + "content": "The resulting photometric cost, aggregated across all depth-scaled primitives " + }, + { + "bbox": [ + 305, + 132, + 545, + 157 + ], + "type": "inline_equation", + "content": "(s_i, P_i) \\in \\mathcal{P}(\\mathrm{I})" + }, + { + "bbox": [ + 305, + 132, + 545, + 157 + ], + "type": "text", + "content": ", is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 339, + 165, + 546, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 165, + 546, + 196 + ], + "spans": [ + { + "bbox": [ + 339, + 165, + 546, + 196 + ], + "type": "interline_equation", + "content": "E _ {\\text {p h o t o}} = \\frac {1}{| \\mathcal {P} (\\mathrm {I}) |} \\sum_ {P _ {i} = \\left(\\Omega_ {i}, \\mathfrak {D}\\right)} r \\left(P _ {i}, s _ {i}, \\mathrm {I} _ {\\mathrm {t}}, \\mathrm {T} _ {t r}\\right) \\tag {7}", + "image_path": "81769f7f921e172fc3ff33caabe1083db1caf9c872a815f35db2267f53b4c415.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 205, + 547, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 547, + 265 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 547, + 265 + ], + "type": "text", + "content": "Note that we abstain from explicit occlusion handling in our photometric alignment, as our per-segment alignment is by design robust to pixel-level occlusion. We, however, expect our system could further be refined with explicit primitive correspondence checks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 266, + 546, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 546, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 546, + 300 + ], + "type": "text", + "content": "Finally, to obtain the relative pose and depth scales, we minimise the photometric cost using the Adam [20] optimiser:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 354, + 303, + 545, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 303, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 354, + 303, + 545, + 319 + ], + "type": "interline_equation", + "content": "\\left\\{s _ {i} ^ {\\text {o p t}}, \\mathrm {T} _ {t r} ^ {\\text {o p t}} \\right\\} = \\operatorname {a r g m i n} _ {s _ {i}, \\mathrm {T} _ {t r}} E _ {\\text {p h o t o}} \\tag {8}", + "image_path": "9e7f7b5b566c7baaa9f7ab14544e345c9c4b6bcf37c63f973b2d582bd1a27fdf.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 319, + 547, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 547, + 367 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 547, + 367 + ], + "type": "text", + "content": "Our approach trivially extends to multiple reference and target views via photometric cost summation, and we can jointly solve all depth scales and poses using the resulting cost function." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 377, + 462, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 377, + 462, + 390 + ], + "spans": [ + { + "bbox": [ + 305, + 377, + 462, + 390 + ], + "type": "text", + "content": "3.3.2 Monocular Visual Odometry" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 397, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 468 + ], + "type": "text", + "content": "To demonstrate that our representation is suitable for simultaneous geometry and pose estimation, we design a novel monocular visual odometry system which operates directly on primitives. Informally, we incrementally build a local 3D map out of primitives and then track new incoming frames against this map, also in a per-primitive manner." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 469, + 547, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 547, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 547, + 516 + ], + "type": "text", + "content": "Visual Odometry (VO) consists of three components: initialisation, tracking, and mapping. Below, we cover each of these components. We interleave tracking and mapping in a single thread." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 517, + 547, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 517, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 305, + 517, + 547, + 613 + ], + "type": "text", + "content": "We adopt a keyframe-based approach [22], which means that the 3D map is represented by a set of posed keyframes " + }, + { + "bbox": [ + 305, + 517, + 547, + 613 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{I}_{\\mathrm{kf}}^i, \\mathrm{T}_{\\mathrm{kf}}^i, \\mathcal{G}^i\\}" + }, + { + "bbox": [ + 305, + 517, + 547, + 613 + ], + "type": "text", + "content": ". Note that each keyframe also has its estimated dense geometry " + }, + { + "bbox": [ + 305, + 517, + 547, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^i" + }, + { + "bbox": [ + 305, + 517, + 547, + 613 + ], + "type": "text", + "content": ", which results in a three-dimensional local map of the scene. We map in a sliding window fashion, with a window size of 5 keyframes. When the window is full, the earliest keyframe is popped from the window." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": "Initialisation. Initialisation of a VO system typically involves estimating the geometry of the first keyframe and its pose with respect to the subsequent few frames. Initialisation of a monocular incremental SfM or SLAM system is often hard, because one has to solve a chicken-and-egg problem: poses are required to reliably estimate geometry and vice versa. Therefore many VO / SLAM systems employ a wide range of heuristics in order to initialise their" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4983" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "system. However, using our new representation no special treatment is required in this case. In order to bootstrap our system, we simply create the two first keyframes and then employ the Structure-from-Motion method proposed in Sec. 3.3.1. We keep the pose of the first keyframe fixed to fix gauge freedom." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 148, + 286, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 148, + 286, + 219 + ], + "spans": [ + { + "bbox": [ + 47, + 148, + 286, + 219 + ], + "type": "text", + "content": "Tracking. Given a map, the goal of tracking is to estimate the pose " + }, + { + "bbox": [ + 47, + 148, + 286, + 219 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_{\\text{track}} \\in \\mathrm{SE}(3)" + }, + { + "bbox": [ + 47, + 148, + 286, + 219 + ], + "type": "text", + "content": " of a new incoming frame " + }, + { + "bbox": [ + 47, + 148, + 286, + 219 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\text{track}}" + }, + { + "bbox": [ + 47, + 148, + 286, + 219 + ], + "type": "text", + "content": ". We adapt Lucas-Kanade [28] tracking onto our per-primitive formulation. At each step we track a new incoming frame against the latest keyframe in the odometry sliding window." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 220, + 286, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 286, + 281 + ], + "type": "text", + "content": "We solve the same photometric cost formulated in Eq. (8), but for the pose only. The latest keyframe serves as the reference frame " + }, + { + "bbox": [ + 46, + 220, + 286, + 281 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 220, + 286, + 281 + ], + "type": "text", + "content": " whereas the new incoming image is set to be the target frame " + }, + { + "bbox": [ + 46, + 220, + 286, + 281 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{t}}" + }, + { + "bbox": [ + 46, + 220, + 286, + 281 + ], + "type": "text", + "content": ". The depth scales of the latest keyframe have already been estimated at the mapping stage." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 286, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 286, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 286, + 364 + ], + "type": "text", + "content": "Since target images are not assumed to be primified, our tracking is essentially equivalent to the classical Lucas-Kanade method and hence could be implemented efficiently on modern hardware. Our proof-of-concept implementation does tracking at 2-3 FPS, but we expect high performance gains from employing standard machinery, such as Gauss-Newton optimisation [24]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 369, + 286, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 286, + 416 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 286, + 416 + ], + "type": "text", + "content": "Mapping. The mapping stage ensures geometric consistency within the keyframe sliding window of size " + }, + { + "bbox": [ + 47, + 369, + 286, + 416 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 369, + 286, + 416 + ], + "type": "text", + "content": ". Concretely, our mapping refines depth scales " + }, + { + "bbox": [ + 47, + 369, + 286, + 416 + ], + "type": "inline_equation", + "content": "\\{s_j\\}" + }, + { + "bbox": [ + 47, + 369, + 286, + 416 + ], + "type": "text", + "content": " and a pose " + }, + { + "bbox": [ + 47, + 369, + 286, + 416 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_{\\mathrm{kf}}^i" + }, + { + "bbox": [ + 47, + 369, + 286, + 416 + ], + "type": "text", + "content": " for each keyframe " + }, + { + "bbox": [ + 47, + 369, + 286, + 416 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{kf}}^i" + }, + { + "bbox": [ + 47, + 369, + 286, + 416 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 47, + 369, + 286, + 416 + ], + "type": "inline_equation", + "content": "0 \\leq i \\leq n" + }, + { + "bbox": [ + 47, + 369, + 286, + 416 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 417, + 286, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 417, + 286, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 417, + 286, + 534 + ], + "type": "text", + "content": "For each keyframe, we define a set of target frames against which the photometric cost is being optimised with a connectivity function " + }, + { + "bbox": [ + 46, + 417, + 286, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(t) = \\{\\mathrm{I}_s,\\mathrm{T}_s\\}" + }, + { + "bbox": [ + 46, + 417, + 286, + 534 + ], + "type": "text", + "content": ". In our implementation, " + }, + { + "bbox": [ + 46, + 417, + 286, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(t)" + }, + { + "bbox": [ + 46, + 417, + 286, + 534 + ], + "type": "text", + "content": " includes temporally neighbouring keyframes " + }, + { + "bbox": [ + 46, + 417, + 286, + 534 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{kf}}^{t - 1}" + }, + { + "bbox": [ + 46, + 417, + 286, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 417, + 286, + 534 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{kf}}^{t + 1}" + }, + { + "bbox": [ + 46, + 417, + 286, + 534 + ], + "type": "text", + "content": " (if they exist) used as supporting frames, as well as 4 additional supplementary views, for which only the pose is being estimated. We employ these views to better constrain the geometry and observed that increasing the number of supplementary views does not induce high computational cost." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 536, + 286, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 286, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 286, + 558 + ], + "type": "text", + "content": "Thus, the mapping stage is done by solving a joint photometric cost for all keyframes:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 561, + 286, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 561, + 286, + 592 + ], + "spans": [ + { + "bbox": [ + 59, + 561, + 286, + 592 + ], + "type": "interline_equation", + "content": "E _ {\\text {m a p p i n g}} = \\sum_ {t = 0} ^ {n} \\sum_ {\\mathrm {I} _ {s} \\in \\mathcal {M} (t)} \\sum_ {P _ {i} \\in \\mathcal {P} \\left(\\mathrm {I} _ {k f} ^ {t}\\right)} r \\left(P _ {i}, s _ {i}, \\mathrm {I} _ {s}, \\mathrm {T} _ {s} ^ {- 1} \\mathrm {T} _ {k f}\\right) \\tag {9}", + "image_path": "4cfa5409edc89a2652e1a0890205b7e0025f2dc01c85ee9f2d34bdf4f74abae1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 603, + 127, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 603, + 127, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 603, + 127, + 616 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 623, + 190, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 190, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 190, + 635 + ], + "type": "text", + "content": "4.1. Sparse Depth Completion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": "Our SuperPrimitives representation could be seamlessly applied to depth completion with no pre-training required, thereby solving it in a zero-shot manner. For each primitive " + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "inline_equation", + "content": "P_{i} \\in \\mathcal{P}(\\mathrm{I})" + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": " we adjust its depth scale to minimise depth discrepancy " + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\| s_i \\cdot \\mathfrak{D}_i - \\hat{D} \\|" + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": " with given ground truth sparse depth " + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": " across all valid depth points within the segment " + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\Omega_{i}" + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "If a segment lacks valid depth measurements, it is discarded. For image regions not covered by any valid primitives, we generate a dense depth prediction by simply bilinearly interpolating depth values." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 121, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 263 + ], + "type": "text", + "content": "Many existing depth completion studies evaluate their methods in artificial scenarios, such as selecting input depth points randomly from a known ground truth. In contrast, our approach is tested using the real-world Void benchmark [57]. The benchmark provides video sequences captured with a RealSense D435i camera together with sparse metric depth measurements acquired from an external visual-inertial SfM system. This setup exhibits noise and biases that might be present in the sparse depth inputs. These measurements are already provided by the dataset itself and are therefore shared across all the methods being compared." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 264, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 382 + ], + "type": "text", + "content": "We compare our results with a recent state-of-the-art model [55], which significantly outperforms previous depth completion methods, particularly in zero-shot generalisation contexts. Other methods evaluated in [55] are also reported in the table. Our depth completion method also operates in a zero-shot manner both task-wise and dataset-wise. That means our method was not trained for the depth completion task and neither of our surface normal prediction network nor segmentation model were trained on the Void dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 384, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 431 + ], + "type": "text", + "content": "We focus on the most challenging \"150 points\" density setting, characterised by minimal sparse depth points per image. The depth is estimated in full " + }, + { + "bbox": [ + 304, + 384, + 545, + 431 + ], + "type": "inline_equation", + "content": "480 \\times 640" + }, + { + "bbox": [ + 304, + 384, + 545, + 431 + ], + "type": "text", + "content": " image resolution, following [55]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 432, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 563 + ], + "type": "text", + "content": "In Tab. 1 our method quantitatively performs on par with a recent state-of-the-art method, which is enabled by a monocular depth predictor pre-trained on a vast mixture of datasets [37, 38] and fine-tined on the Void train set. In the zero-shot setting, we outperform VI-Depth [55], which uses a DPT-Hybrid pre-trained backbone, on three out of four metrics. Note that no training was done the for depth completion task for our method. Compared to the ground truth depth maps obtained via a noisy sensor, our predictions show (Fig. 3) sharper object boundaries and are better at preserving straight lines and perpendicular structures." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 571, + 487, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 571, + 487, + 583 + ], + "spans": [ + { + "bbox": [ + 306, + 571, + 487, + 583 + ], + "type": "text", + "content": "4.2. Few-View Structure-from-Motion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 590, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 545, + 689 + ], + "type": "text", + "content": "Contrasted with depth completion, where global geometry structure is roughly given, we also test our method on few view Structure-from-Motion. Given a set of unposed images " + }, + { + "bbox": [ + 304, + 590, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{I}_{\\mathrm{ref}},\\mathrm{I}_s^0,\\dots ,\\mathrm{I}_s^{(n - 1)}\\}" + }, + { + "bbox": [ + 304, + 590, + 545, + 689 + ], + "type": "text", + "content": " captured within a small timeframe, the goal is to estimate the depth of the reference image " + }, + { + "bbox": [ + 304, + 590, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{I_{ref}}" + }, + { + "bbox": [ + 304, + 590, + 545, + 689 + ], + "type": "text", + "content": ". In that setup, the set of unposed supplementary views " + }, + { + "bbox": [ + 304, + 590, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{I}_s^0,\\ldots ,\\mathrm{I}_s^{(n - 1)}\\}" + }, + { + "bbox": [ + 304, + 590, + 545, + 689 + ], + "type": "text", + "content": " will be leveraged for multi-view geometry estimation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We choose the test set of the ScanNet dataset for this evaluation. In each test sequence we select every 200-th" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4984" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 547, + 272 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 547, + 272 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 547, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 547, + 272 + ], + "type": "image", + "image_path": "d14519d12162bbcf13ae7584fcf4c7af00e2af919923dd34972278d57ccf74ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 274, + 547, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 547, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 547, + 309 + ], + "type": "text", + "content": "Figure 3. Depth Completion on Void. We visualise the coloured unprojections of ground truth depth maps provided by a sensor (top row) and the geometry estimated by our method (bottom row). Sparse depth input points are visualised as red dots (electronic zoom-in recommended). Qualitatively, we achieve sharper geometry estimates than from a commodity depth sensor." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 312, + 282, + 411 + ], + "blocks": [ + { + "bbox": [ + 53, + 312, + 282, + 411 + ], + "lines": [ + { + "bbox": [ + 53, + 312, + 282, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 312, + 282, + 411 + ], + "type": "table", + "html": "
MethodMAERMSEiMAEiRMSE
VOICED [57]174.04253.1487.39126.3
NLSPN [34]--143238.10
ScaffNet [58]150.65255.0880.79133.33
KBNet [56]131.54263.5466.84128.33
MonDi [27]104.96225.6048.4496.79
VI-Depth (DPT-Hybrid + NYUv2)--55.9085.20
Ours109.0204.1547.3283.40
VI-Depth* (MiDaS) [55]113.27193.3853.8684.82
VI-Depth* (DPT-Hybrid) [55]97.03167.8246.6274.67
", + "image_path": "05d7efb282ef91aca1dfb07b7ae8e204759884f99e37f5b3cad2e33a20b7fc4d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 413, + 287, + 491 + ], + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 491 + ], + "type": "text", + "content": "Table 1. Depth Completion on Void. We report four most widely used metrics for depth completion on the Void benchmark. The methods which did not use the Void dataset for training are in the top section. Methods trained on Void are marked with an asterisk*. Our method demonstrates superior performance on three out of the four metrics within the zero-shot group. It is second-best on two out of the four metrics overall." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 498, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 581 + ], + "type": "text", + "content": "frame to be the reference frame " + }, + { + "bbox": [ + 46, + 498, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 581 + ], + "type": "text", + "content": ". Supporting frames are gathered from the neighbouring frames. Then we discard the frame sets with not enough motion to remove mostly static video clips, where SfM could not be performed. This resulted in " + }, + { + "bbox": [ + 46, + 498, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\sim 500" + }, + { + "bbox": [ + 46, + 498, + 287, + 581 + ], + "type": "text", + "content": " reference frames. Note that in these experiments we use a surface normal neural network pre-trained only on the synthetic HyperSim dataset [40]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "Since the multi-view depth estimation problem has scale ambiguity, we employ median-scaling to align estimated depth to the metric scale for evaluation. We report iMAE and iRMSE in Fig. 6 for our method with varying number of supplementary views. We compare our depth estimation quality against the method closest to ours, DeepV2D [49], which can also estimate depth together with supplementary frames poses. We demonstrate that our geometry quickly saturates after observing as little as 2 supporting views. Unlike DeepV2D we do not use any external tracking or initial relative pose estimation, yet still our method demonstrates" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 313, + 541, + 509 + ], + "blocks": [ + { + "bbox": [ + 310, + 313, + 541, + 509 + ], + "lines": [ + { + "bbox": [ + 310, + 313, + 541, + 509 + ], + "spans": [ + { + "bbox": [ + 310, + 313, + 541, + 509 + ], + "type": "image", + "image_path": "cd37b939f9f2698aba14071cafe4e0bb7da26534354b69e3d371f8e2f075ad25.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 514, + 547, + 570 + ], + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 570 + ], + "type": "text", + "content": "Figure 4. 3-View SfM on ScanNet. We provide the visualisations of unprojected reference frame depth maps predicted by our method for few-view SfM using one reference and 2 supplementary views. Note that we used surface normal prediction network which was only pretrained on HyperSim [40] for this experiment." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 578, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 547, + 639 + ], + "type": "text", + "content": "consistent improvement over DeepV2D. Additionally, our method does not have any global prior on relative object positions, unlike DeepV2D. Our approach therefore could be used as an VO / SLAM initialisation mechanism, estimating joint relative poses and geometry." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 646, + 463, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 646, + 463, + 659 + ], + "spans": [ + { + "bbox": [ + 305, + 646, + 463, + 659 + ], + "type": "text", + "content": "4.3. Monocular Visual Odometry" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Monocular Visual Odometry requires both accurate pose and geometry estimation to successfully track camera motion across long trajectories. Minor pose estimation inaccuracies can accumulate over time resulting in what is called" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4985" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 73, + 221, + 210 + ], + "blocks": [ + { + "bbox": [ + 51, + 73, + 221, + 210 + ], + "lines": [ + { + "bbox": [ + 51, + 73, + 221, + 210 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 221, + 210 + ], + "type": "image", + "image_path": "d5cc49e732b8b85c274699b0a714af00cc62af10092bcaeed63bff256c49a9c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 546, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 546, + 240 + ], + "type": "text", + "content": "Figure 5. TUM Reconstruction Results. Examples of reconstructions produced by our monocular VO system on the TUM dataset. Each image shows a coloured point cloud of the geometry estimated on an odometry keyframe." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 73, + 389, + 209 + ], + "blocks": [ + { + "bbox": [ + 225, + 73, + 389, + 209 + ], + "lines": [ + { + "bbox": [ + 225, + 73, + 389, + 209 + ], + "spans": [ + { + "bbox": [ + 225, + 73, + 389, + 209 + ], + "type": "image", + "image_path": "b5b1216a0518c73a217ad202abf211c3fcdf98c0a102b61b31bd8d336b4f4ad8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 391, + 73, + 542, + 210 + ], + "blocks": [ + { + "bbox": [ + 391, + 73, + 542, + 210 + ], + "lines": [ + { + "bbox": [ + 391, + 73, + 542, + 210 + ], + "spans": [ + { + "bbox": [ + 391, + 73, + 542, + 210 + ], + "type": "image", + "image_path": "9ad305d9bf382664072aaaf3b4b98eafa66b75052d7d24b38b5f8f53d7586f6c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 51, + 245, + 167, + 361 + ], + "blocks": [ + { + "bbox": [ + 51, + 245, + 167, + 361 + ], + "lines": [ + { + "bbox": [ + 51, + 245, + 167, + 361 + ], + "spans": [ + { + "bbox": [ + 51, + 245, + 167, + 361 + ], + "type": "image", + "image_path": "bfce09d04e29f96cfc14063379b589fc2ffed88683671bca5b31e22c65e2f0fe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 365, + 287, + 443 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 443 + ], + "type": "text", + "content": "Figure 6. Few-View SfM Depth Estimation Quality. We evaluate the quality of our depth estimation method from an unstructured set of images on the ScanNet dataset. On the " + }, + { + "bbox": [ + 46, + 365, + 287, + 443 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 365, + 287, + 443 + ], + "type": "text", + "content": "-axis the number of supporting views is shown, while on the " + }, + { + "bbox": [ + 46, + 365, + 287, + 443 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 365, + 287, + 443 + ], + "type": "text", + "content": "-axis the corresponding depth reconstruction quality metric is reported. We show that the quality of our depth estimation quickly saturates and consistently outperforms its closest competitor, DeepV2D." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 169, + 245, + 286, + 361 + ], + "blocks": [ + { + "bbox": [ + 169, + 245, + 286, + 361 + ], + "lines": [ + { + "bbox": [ + 169, + 245, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 169, + 245, + 286, + 361 + ], + "type": "image", + "image_path": "ba15942e098593b50d20088664b42e89400cb9c1ad254bbe2496efa5955f7088.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 449, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 510 + ], + "type": "text", + "content": "drift. With geometry being incorrectly estimated, accurate pose tracking becomes impossible, and vice versa. Our SuperPrimitive representation allows estimating both pose and geometry, enabling us to build a simple monocular VO system which performs better even in hard conditions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 629 + ], + "type": "text", + "content": "We evaluate our monocular odometry on the TUM RGB-D [46] dataset. The dataset was captured with a handheld camera in indoor scenes. It is renowned for being incredibly challenging (especially for dense reconstruction systems), due to motion blur, rolling shutter artefacts, and abundance of pure rotational motion. We show that thanks to the strong priors encapsulated in our SuperPrimitives, our simple monocular odometry system can handle the TUM dataset without any special treatment (e.g. our method does not involve any special motion blur handling)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": "We compare against other VO systems that do not have global bundle adjustment, following the protocol of [10] and evaluate on 8 sequences from the Friedburg 1 split. We use only RGB images as the input to our system and down-sample them to " + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "inline_equation", + "content": "120 \\times 160" + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": " for efficiency purposes. Since our method is purely monocular, the estimated trajectory lacks global scale and we first use " + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{Sim}(3)" + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": " alignment to the" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 245, + 548, + 366 + ], + "blocks": [ + { + "bbox": [ + 309, + 245, + 548, + 366 + ], + "lines": [ + { + "bbox": [ + 309, + 245, + 548, + 366 + ], + "spans": [ + { + "bbox": [ + 309, + 245, + 548, + 366 + ], + "type": "table", + "html": "
Seq.TartanVO [53]DeepV2D [49]DeepFactors [8]DepthCov [10]Ours
3600.1780.2430.1590.1280.173
desk0.1250.1660.1700.0560.085
desk20.1220.3790.2530.0480.108
plant0.2970.2030.3050.2610.153
room0.3330.2460.3640.2570.363
rpy0.0490.1050.0430.0520.055
teddy0.3390.3160.6010.4750.253
xyz0.0620.0640.0350.0560.036
mean0.1880.2150.2410.1670.153
", + "image_path": "82c20203c750d764b3732275a1859d51ed8ed9f30ae2793ee1135b812a489a6c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 369, + 546, + 435 + ], + "lines": [ + { + "bbox": [ + 305, + 369, + 546, + 435 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 546, + 435 + ], + "type": "text", + "content": "Table 2. Trajectory Estimation Error on TUM. Average Trajectory Error (ATE) is compared against other monocular odometry systems on the TUM Friedburg 1 split. The best and second best results are highlighted in bold and underscored correspondingly. Our method outperforms others in terms of the ATE averaged across all trajectories." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 441, + 546, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 546, + 525 + ], + "type": "text", + "content": "ground truth scale, following standard practice [4]. Tab. 2 shows that, despite the simplicity of our odometry system, it outperforms all other methods in terms of Average Trajectory Error (ATE) [46], averaged across all trajectories. Additionally, our VO is either the best or second best on five out of eight sequences. Besides quantitative evaluation, we also demonstrate reconstruction results in Fig. 5." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 535, + 378, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 535, + 378, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 535, + 378, + 548 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 555, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 555, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 305, + 555, + 545, + 639 + ], + "type": "text", + "content": "We presented a new representation, SuperPrimitive, which demonstrates how recent advances in building strong single image priors could be incorporated into pose and dense geometry estimation problems. We show that incorporating these priors alleviates the need for sophisticated handcrafted heuristics and paves the way into monocular reconstruction with relative ease." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 647, + 546, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 647, + 546, + 703 + ], + "spans": [ + { + "bbox": [ + 305, + 647, + 546, + 703 + ], + "type": "text", + "content": "Acknowledgments. Research presented in this paper was supported by Dyson Technology Ltd. The authors would like to thank Hidenobu Matsuki, Eric Dexheimer, Riku Murai, and other members of the Dyson Robotics Lab for countless fruitful discussions." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4986" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Gwangbin Bae, Ignas Budvytis, and Roberto Cipolla. Estimating and exploiting the aleatoric uncertainty in surface normal estimation. In International Conference on Computer Vision (ICCV), 2021. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "text", + "content": "[2] M. Bloesch, J. Czarnowski, R. Clark, S. Leutenegger, and A. J. Davison. CodeSLAM — learning a compact, optimisable representation for dense visual SLAM. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 1, 2, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "text", + "content": "[3] Aljaž Božić, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. Transformerfusion: Monocular rgb scene reconstruction using transformers. Proc. Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 278 + ], + "type": "text", + "content": "[4] Carlos Campos, Richard Elvira, Juan J. Gomez, Jose M. M. Montiel, and Juan D. Tardos. ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Transactions on Robotics, 2021. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 323 + ], + "type": "text", + "content": "[5] Xu Cao, Hiroaki Santo, Boxin Shi, Fumio Okura, and Yasuyuki Matsushita. Bilateral normal integration. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 357 + ], + "type": "text", + "content": "[6] Alejo Concha and Javier Civera. Using superpixels in monocular SLAM. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2014. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 357, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 390 + ], + "type": "text", + "content": "[7] G. Cross and A. Zisserman. Quadric Reconstruction from Dual-Space Geometry. In Proceedings of the International Conference on Computer Vision (ICCV), 1998. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "text", + "content": "[8] J. Czarnowski, T. Laidlow, R. Clark, and A. J. Davison. Deepfactors: Real-time probabilistic dense monocular SLAM. In IEEE Robotics and Automation Letters, 2020. 8, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 435, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 287, + 479 + ], + "type": "text", + "content": "[9] A. J. Davison, N. D. Molton, I. Reid, and O. Stasse. MonoSLAM: Real-Time Single Camera SLAM. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2007. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 480, + 287, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 480, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 287, + 522 + ], + "type": "text", + "content": "[10] Eric Dexheimer and Andrew J. Davison. Learning a depth covariance function. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 524, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 524, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 524, + 287, + 567 + ], + "type": "text", + "content": "[11] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In Neural Information Processing Systems (NeurIPS), 2014. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 568, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 287, + 612 + ], + "type": "text", + "content": "[12] David Eigen, Christian Puhrsch, and Rob Fergus. Depth Map Prediction from a Single Image using a Multi-Scale Deep Network. In Neural Information Processing Systems (NeurIPS), 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 613, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 287, + 657 + ], + "type": "text", + "content": "[13] Jakob Engel, Thomas Schoeps, and Daniel Cremers. LSD-SLAM: Large-scale direct monocular SLAM. In Proceedings of the European Conference on Computer Vision (ECCV), 2014. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 287, + 690 + ], + "type": "text", + "content": "[14] J. Engel, V. Koltun, and D. Cremers. Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2017. 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 691, + 287, + 715 + ], + "type": "text", + "content": "[15] Jiahui Fu, Yilun Du, Kurran Singh, Joshua B Tenenbaum, and John J Leonard. Neuse: Neural se (3)-equivariant em" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "text", + "content": "bedding for consistent spatial understanding with objects. In Proceedings of Robotics: Science and Systems (RSS), 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "type": "text", + "content": "[16] Y. Furukawa, B. Curless, S. M. Seitz, and R. Szeliski. Towards Internet-scale multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2010. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "type": "text", + "content": "[17] D. Gallup, J.-M. Frahm, and M. Pollefeys. Piecewise planar and non-planar stereo for urban scene reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2010. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 194, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 194, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 194, + 545, + 227 + ], + "type": "text", + "content": "[18] Magnus R. Hestenes and Eduard Stiefel. Methods of conjugate gradients for solving linear systems. Journal of research of the National Bureau of Standards, 1952. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "text", + "content": "[19] M Kaess. Simultaneous localization and mapping with infinite planes. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2015. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 261, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 294 + ], + "type": "text", + "content": "[20] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[22] G. Klein and D. W. Murray. Parallel Tracking and Mapping for Small AR Workspaces. In Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR), 2007. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 393, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 426 + ], + "type": "text", + "content": "[23] T. Laidlow and A. J. Davison. Simultaneous localisation and mapping with quadric surfaces. In International Conference on 3D Vision (3DV), 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "type": "text", + "content": "[24] Stefan Leutenegger, Simon Lynen, Michael Bosse, Roland Siegwart, and Paul Furgale. Keyframe-based visual-inertial odometry using nonlinear optimization. International Journal of Robotics Research (IJRR), 2014. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "text", + "content": "[25] Chen Liu, Jimei Yang, Duygu Ceylan, Ersin Yumer, and Yasutaka Furukawa. Planenet: Piece-wise planar reconstruction from a single rgb image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 525, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 580 + ], + "type": "text", + "content": "[26] Chen Liu, Kihwan Kim, Jinwei Gu, Yasutaka Furukawa, and Jan Kautz. Planercnn: 3d plane detection and reconstruction from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 581, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 624 + ], + "type": "text", + "content": "[27] Tianlin Liu, Parth T. Agrawal, Allison Chen, Byung-Woo Hong, and A. Wong. Monitored distillation for positive congruent depth completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[28] B. D. Lucas and T. Kanade. An Iterative Image Registration Technique with an Application to Stereo Vision. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), 1981. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[29] J. McCormac, R. Clark, M. Bloesch, A. J. Davison, and S. Leutenegger. Fusion++: volumetric object-level slam. In Proceedings of the International Conference on 3D Vision (3DV), 2018. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4987" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[30] R. A. Newcombe, S. Lovegrove, and A. J. Davison. DTAM: Dense Tracking and Mapping in Real-Time. In Proceedings of the International Conference on Computer Vision (ICCV), 2011. 1, 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[31] Lachlan Nicholson, Michael Milford, and Niko Sunderhauf. Quadricslam: Dual quadrics from object detections as landmarks in object-oriented slam. IEEE Robotics and Automation Letters, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 229 + ], + "type": "text", + "content": "[32] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 297 + ], + "type": "text", + "content": "[33] Ryosuke Okuta, Yuya Unno, Daisuke Nishino, Shohei Hido, and Crissman Loomis. Copy: A numpy-compatible library for nvidiagpu calculations. In Proceedings of Workshop on Machine Learning Systems (LearningSys) in The Thirty-first Annual Conference on Neural Information Processing Systems (NeurIPS), 2017. 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "type": "text", + "content": "[34] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 343, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 288, + 398 + ], + "type": "text", + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. In Neural Information Processing Systems (NeurIPS), 2019. 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "type": "text", + "content": "[36] Albert Pumarola, Alexander Vakhitov, Antonio Agudo, Alberto Sanfeliu, and Francese Moreno-Noguer. Pl-slam: Real-time monocular visual slam with points and lines. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 456, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 498 + ], + "type": "text", + "content": "[37] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 2, 6, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 500, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 288, + 555 + ], + "type": "text", + "content": "[38] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2022. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 557, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 590 + ], + "type": "text", + "content": "[39] Xiaofeng Ren and Jitendra Malik. Learning a classification model for segmentation. In Proceedings of the International Conference on Computer Vision (ICCV), 2003. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 591, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 657 + ], + "type": "text", + "content": "[40] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[41] R. F. Salas-Moreno, R. A. Newcombe, H. Strasdat, P. H. J. Kelly, and A. J. Davison. SLAM++: Simultaneous Localisation and Mapping at the Level of Objects. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2013. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "text", + "content": "[42] Mohamed Sayed, John Gibson, Jamie Watson, Victor Prisacariu, Michael Firman, and Clément Godard. Simplerecon: 3d reconstruction without 3d convolutions. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 546, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 171 + ], + "type": "text", + "content": "[43] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 173, + 546, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 546, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 546, + 216 + ], + "type": "text", + "content": "[44] Jonathan Shade, Steven Gortler, Li-wei He, and Richard Szeliski. Layered depth images. In Proceedings of the 25th annual conference on Computer graphics and interactive techniques, 1998. 12" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 217, + 546, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 546, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 546, + 259 + ], + "type": "text", + "content": "[45] Jingjia Shi, Shuaifeng Zhi, and Kai Xu. Planerectr: Unified query learning for 3d plane recovery from a single view. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 261, + 546, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 546, + 304 + ], + "type": "text", + "content": "[46] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A Benchmark for the Evaluation of RGB-D SLAM Systems. In Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS), 2012. 8, 13" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 305, + 546, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 546, + 349 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 546, + 349 + ], + "type": "text", + "content": "[47] E. Sucar, K. Wada, and A. J. Davison. NodeSLAM: Neural object descriptors for multi-view shape reconstruction. In Proceedings of the International Conference on 3D Vision (3DV), 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 350, + 546, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 546, + 403 + ], + "type": "text", + "content": "[48] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. NeuralRecon: Real-time coherent 3D reconstruction from monocular video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 404, + 546, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 546, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 546, + 448 + ], + "type": "text", + "content": "[49] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. In Proceedings of the International Conference on Learning Representations (ICLR), 2020. 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 449, + 546, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 546, + 491 + ], + "type": "text", + "content": "[50] Alexander Vakhitov, Jan Funke, and Francesc Moreno-Noguer. Accurate and linear time pose estimation from points and lines. In Proceedings of the European Conference on Computer Vision (ECCV), 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 492, + 546, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 546, + 535 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 546, + 535 + ], + "type": "text", + "content": "[51] Tom van Dijk and Guido C.H.E. de Croon. How do neural networks see depth in single images? In Proceedings of the International Conference on Computer Vision (ICCV), 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 536, + 546, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 536, + 546, + 591 + ], + "spans": [ + { + "bbox": [ + 307, + 536, + 546, + 591 + ], + "type": "text", + "content": "[52] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 592, + 546, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 546, + 624 + ], + "type": "text", + "content": "[53] Wenshan Wang, Yaoyu Hu, and Sebastian Scherer. TartanVO: A Generalizable Learning-based VO. In Conference on Robot Learning (CoRL), 2020. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 625, + 546, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 546, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 546, + 668 + ], + "type": "text", + "content": "[54] C. S. Weerasekera, Y. Latif, R. Garg, and I. Reid. Dense monocular reconstruction using surface normals. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 669, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 546, + 713 + ], + "type": "text", + "content": "[55] Wofk, Diana and Ranftl, René and Müller, Matthias and Koltun, Vladlen. Monocular Visual-Inertial Depth Estimation. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2023. 6, 7, 12" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4988" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 239 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[56] Alex Wong and Stefano Soatto. Unsupervised depth completion with calibrated backprojection layers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "type": "text", + "content": "[57] Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters, 2020. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 195 + ], + "type": "text", + "content": "[58] Alex Wong, Safa Cicek, and Stefano Soatto. Learning topology from synthetic data for unsupervised depth completion. IEEE Robotics and Automation Letters, 2021. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "text", + "content": "[59] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In Proceedings of SIGGRAPH, 2018. 12" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4989" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/3fa1e5c3-5890-4208-976c-19acdb43c796_content_list.json b/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/3fa1e5c3-5890-4208-976c-19acdb43c796_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d7ec1d38a5c753305875ceba22c4fb4ef8d1e733 --- /dev/null +++ b/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/3fa1e5c3-5890-4208-976c-19acdb43c796_content_list.json @@ -0,0 +1,1478 @@ +[ + { + "type": "text", + "text": "SuperSVG: Superpixel-based Scalable Vector Graphics Synthesis", + "text_level": 1, + "bbox": [ + 153, + 130, + 815, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Teng Hu $^{1}$ , Ran Yi $^{1*}$ , Baihong Qian $^{1}$ , Jiangning Zhang $^{2}$ , Paul L. Rosin $^{3}$ , Yu-Kun Lai $^{3}$ $^{1}$ Shanghai Jiao Tong University, $^{2}$ Youtu Lab, Tencent, $^{3}$ Cardiff University \n{hu-teng, ranyi, cherry_qbh} $@$ sjtu.edu.cn, vtzhang@tencent.com, {RosinPL, LaiY4} $@$ cardiff.ac.uk", + "bbox": [ + 86, + 179, + 908, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SVG (Scalable Vector Graphics) is a widely used graphics format that possesses excellent scalability and editability. Image vectorization, which aims to convert raster images to SVGs, is an important yet challenging problem in computer vision and graphics. Existing image vectorization methods either suffer from low reconstruction accuracy for complex images or require long computation time. To address this issue, we propose SuperSVG, a superpixel-based vectorization model that achieves fast and high-precision image vectorization. Specifically, we decompose the input image into superpixels to help the model focus on areas with similar colors and textures. Then, we propose a two-stage self-training framework, where a coarse-stage model is employed to reconstruct the main structure and a refinement-stage model is used for enriching the details. Moreover, we propose a novel dynamic path warping loss to help the refinement-stage model to inherit knowledge from the coarse-stage model. Extensive qualitative and quantitative experiments demonstrate the superior performance of our method in terms of reconstruction accuracy and inference time compared to state-of-the-art approaches. The code is available in https://github.com/sjtuplayer/SuperSVG.", + "bbox": [ + 76, + 301, + 473, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 680, + 209, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scalable Vector Graphics, commonly known as SVG, is a widely used vector image format that has a wide range of applications and advantages within the domains of web design, graphic design, mobile applications, data visualization, and various other contexts. Compared with raster images that represent content by pixels, SVG describes images by parameterized vectors and benefits from its scalability and editability where it can be resized to any resolution without losing quality and can be easily manipulated by its layer-wise topological information.", + "bbox": [ + 75, + 705, + 468, + 857 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given the superior capabilities of Scalable Vector Graph", + "bbox": [ + 96, + 858, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ba3dbbac75d07c23024b79669a896d16b08d24cf6ddc49ef9d48e3f017c5305.jpg", + "image_caption": [ + "Figure 1. Overview of our SuperSVG: our model first decomposes the image to be vectorized into superpixels, each containing pixels sharing similar colors and contents. The coarse-stage model predicts the path parameters to reconstruct the main structure, and then the coarse paths guided refinement model enriches the details by learning the knowledge from the coarse-stage model. Compared to the previous methods, our SuperSVG achieves both a high vectorization accuracy and fast computation speed." + ], + "image_footnote": [], + "bbox": [ + 514, + 267, + 874, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ics (SVG) in image representation and editing, there is much research on the topic of image vectorization, which aims to convert rasterized images into SVG. The existing methods can be categorized into three classes: 1) Traditional algorithm-based methods [15, 18, 24, 28, 33, 37], where conventional algorithms are employed to fit images, but they usually suffer from a lower vectorization quality. 2) Deep-learning-based methods [9, 10, 12, 19, 21, 22], which parameterize raster images using deep neural networks for reconstruction. They are efficient and can handle the vectorization of simple graphics or characters (e.g., icons and emojis), but struggle to vectorize complex images. 3) Optimization-based methods [17, 20, 26, 38], which optimize SVG parameters to fit the target image, yielding relatively superior reconstruction quality. However, these methods entail a substantial amount of time and computational resources, making them impractical for timely processing of large-scale data. In summary, previous image vectorization methods either suffer from low reconstruction quality for complex images, or demand extensive computation time, imposing significant constraints on their practical utility.", + "bbox": [ + 496, + 568, + 893, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To achieve good vectorization quality with high ef", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24892", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ficiency, we propose SuperSVG, a deep-learning-based method that translates images into scalable vector graphics (SVG) in a coarse-to-fine manner. Since neural networks have difficulties in directly vectorizing complex images, we decompose the input image into different parts in the form of superpixels wherein the pixels share similar colors and textures and then vectorize each part. Then, we propose a Two-Stage Self-Teaching Training framework to vectorize the superpixels, where the coarse-stage model is trained to reconstruct the main structure of the image and the refinement-stage model is trained to enrich the image details. We make use of the predicted paths from the coarse-stage model to guide the refinement-stage model in image vectorization. Furthermore, we propose a novel Dynamic Path Warping loss which helps the refinement-stage to inherit the knowledge of the coarse-stage model. With the help of the superpixel-based image decomposition and the two-stage self-teaching framework, our SuperSVG can keep the image structure well and reconstruct more details at high speed. Extensive quantitative and qualitative experiments validate the effectiveness of our model.", + "bbox": [ + 76, + 90, + 472, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of our work are four-fold:", + "bbox": [ + 96, + 409, + 426, + 422 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SuperSVG, a novel superpixel-based vectorization model that translates the rasterized images into scalable vector graphics (SVG) based on superpixels and vectorizes the superpixels in a coarse-to-fine manner.", + "- We design a Two-Stage Self-Teaching Training framework, where we employ a coarse-stage model to reconstruct the main structures and a refinement-stage model to enrich the image details based on the coarse-stage output.", + "- We propose a Coarse Paths Guided Training strategy to guide the refinement-stage model to inherit the knowledge from the coarse-stage model, which greatly improves the performance of the refinement-stage model and avoids converging to suboptimal local minimum.", + "- We propose Dynamic Path Warping (DPW) loss, which measures the distance between the predicted paths from the refinement-stage model and the pseudo ground truth approximated with coarse paths. By minimizing the DPW loss, the refinement-stage model can distill the knowledge from the coarse-stage model." + ], + "bbox": [ + 76, + 428, + 468, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 744, + 218, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Vectorization", + "text_level": 1, + "bbox": [ + 76, + 770, + 267, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image vectorization aims to transform a rasterized image into scalable vector graphics (SVG) composed of parameterized vectors. Different from raster images that may become blurry when zooming in, SVG can be rendered at any resolution without losing quality and is convenient to edit, widely used in web design, graphic design, etc. The existing vectorization methods can be classified into 3 categories:", + "bbox": [ + 76, + 794, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional Algorithm-based Image Vectorization Methods can be classified into mesh-based and curve-based ones. The mesh-based methods [15, 18, 28, 37] segment an input image into non-overlapping patches, and infer the color and the boundary location for each region. The curve-based methods [1, 5, 24, 32, 36] employ Bézier curves with different colors defined on either side to create the vector image. Potrace [24] is a representative method of this type that projects the smooth outlines into Bézier paths, and merge the adjacent paths together. However, the vectorization quality of these methods still needs improvements.", + "bbox": [ + 496, + 90, + 890, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep-learning-based Image Vectorization Methods use neural networks to project a raster image into SVG. Im2Vec [21] employs a variational auto-encoder (VAE) [14] to embed the input image and then maps it into path parameters by a Long Short-Term Memory (LSTM) module [23]. Raster2Vec [19] is focused on vectorization of rasterized floor plans using a ResNet [11]. Gao et al. [9] rely on a pre-trained VGG network [25] and a hierarchical Recurrent Neural Network (RNN) to output parametric curves of different sizes. But these methods only focus on simple images and cannot vectorize complex images well. In comparison, our SuperSVG is the first deep-learning-based method that can vectorize images with complex details, thanks to our superpixel decomposition and coarse-path guided refinement that substantially reduce the learning difficulties.", + "bbox": [ + 496, + 257, + 892, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optimization-based Image Vectorization Methods. DiffVG [17] proposes a differentiable renderer that renders the SVG parameters into images. Based on this, DiffVG minimizes the distance between the rasterized and vector images by optimizing the SVG parameters using gradient descent. LIVE [20] and SAMVG [38] further introduce a layer-wise optimization framework, which achieves better vectorization quality over the previous methods. However, due to the low optimization efficiency, they suffer from a long optimization time. In contrast, our SuperSVG achieves both a good vectorization quality and high efficiency.", + "bbox": [ + 496, + 484, + 890, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Superpixel Decomposition", + "text_level": 1, + "bbox": [ + 500, + 657, + 736, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Superpixel decomposition is usually used for data preprocessing in vision tasks. Existing superpixel decomposition methods can be categorized into methods based on traditional algorithm or deep learning. For the traditional algorithm based methods, diverse strategies have been employed, e.g., energy-driven sampling [30], geometric flows [16] and clustering [3]. Some recent works [13, 29, 34] employ neural networks to enhance the performance in superpixel decomposition, which shows great potential in this task.", + "bbox": [ + 496, + 681, + 890, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 844, + 589, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image vectorization aims to translate a rasterized image $I$ into a Scalable Vector Graphic (SVG). An SVG is com", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24893", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f228249a05e0d8180b1b9ff0e963639311473963883245fbda2d937d28655add.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 85, + 851, + 228 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/38f6b6096415ea91de2579883a6ff84fc00d4e87ae10f9f385c01b8edda6b5e2.jpg", + "image_caption": [ + "Figure 2. Main framework of our SuperSVG: we decompose the target image into superpixels and vectorize each superpixel separately. We employ an attention-based coarse-stage model to predict SVG paths that reconstruct the main structure of the superpixel. Then, a refinement-stage model guided by the coarse paths is designed to predict more SVG paths to refine details based on the coarse image. Finally, by combining all the predicted SVGs for each superpixel, we obtain an output SVG with good structure and fine details." + ], + "image_footnote": [], + "bbox": [ + 117, + 229, + 844, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "posed of many vector primitives, which can be SVG paths, ellipses, circles, or rectangles, etc. Following previous works [17, 20, 21], we employ the SVG paths as the shape primitive, where each SVG path defines a region constructed by multiple cubic Bezier curves connected end-to-end with certain color. With the parameters of these SVG paths, the rasterized image can be rendered in any resolution. To obtain the path parameters, some previous methods [17, 20] optimize the path parameters directly to minimize the distance between the input image and the rendered image, which achieves good reconstruction quality but requires long optimization time. To speed up the vectorization process, some deep-learning-based methods [21] employ a deep-learning model to predict the SVG path parameters, but struggle to vectorize complex images.", + "bbox": [ + 75, + 438, + 468, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve good vectorization quality with high efficiency, we propose SuperSVG, a deep-learning-based image vectorization method that translates images into SVG path parameters automatically. To improve the model ability to vectorize complex images, we segment the input image into different parts, within which the pixels share similar colors and textures, and then vectorize each part separately, where superpixels are used for image segmentation as they tend to maintain compactness, uniformity and regularity, particularly suitable for our task. For each superpixel $x \\in \\mathcal{X}$ where $\\mathcal{X}$ is the set of all superpixels, our model converts it into a sequence of path parameters, where each path is composed of several cubic Bezier curves and has a fill color, with a total of $N_{p}$ parameters. With the predicted path parameters for each superpixel, we employ the differ", + "bbox": [ + 75, + 674, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "entiable renderer $R(\\cdot)$ from DiffVG [17] to get the rendered image $\\hat{I}$ in pixel space, which is expected to be close to the input image $I$ .", + "bbox": [ + 498, + 438, + 890, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a two-stage self-teaching framework, composed of a coarse-stage model $E_{c}$ to reconstruct the main structure and a refinement model $E_{r}$ to enrich the details, where the predicted paths from coarse-stage model are used to guide the refinement model in vectorization. $E_{c}$ takes the superpixel $x$ as input and outputs $n$ paths $S = \\{s_1,s_2,\\dots s_n\\}$ to reconstruct main structure; while $E_{r}$ takes both the rendered image $R(S)$ and target superpixel $x$ as inputs, and outputs $m$ paths $S^{\\prime} = \\{s_{1}^{\\prime},s_{2}^{\\prime}\\dots s_{m}^{\\prime}\\}$ to refine details. Combining all the predicted $S$ and $S^{\\prime}$ for each superpixel produces the final SVG result.", + "bbox": [ + 498, + 484, + 892, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Superpixel-based Coarse Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 664, + 846, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Superpixel decomposition. Considering the optimization-based methods suffer from a long optimization time, our SuperSVG builds upon neural networks to efficiently predict SVG paths. However, as neural networks have difficulties in directly vectorizing complex images [8] we therefore simplify the task to vectorizing a certain part of the image containing homogeneous colors and textures. Since superpixel algorithms provide a good tool to decompose images based on local pixel color and also ensure alignment of the regions with the image boundaries, we segment the input image into superpixels, and our model reconstructs each superpixel with scalable vectors. Superpixels also tend to be more regular, making them easier for vectorization. Specifically, we utilize SLIC [3] to decompose the input image", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24894", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8807daf335b8d788c8ebb22f45ddb52a6652ad9434a59c59f88c3849cd7ba8ee.jpg", + "image_caption": [ + "Figure 3. Illustration of our boundary loss $\\mathcal{L}_{\\text{Bound}}$ , which computes the area of the SVG paths that are outside the superpixel mask, and guides the paths to be inside the superpixel." + ], + "image_footnote": [], + "bbox": [ + 107, + 87, + 436, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "into superpixels. We set the compactness parameter as 30 for SLIC to make superpixels more regular.", + "bbox": [ + 75, + 265, + 468, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Coarse-stage model. For a superpixel $x$ with mask mask (indicating those pixels within $x$ with 1 and 0 otherwise), we first design the coarse-stage model $E_{c}$ to vectorize the main structure by predicting the SVG path sequence $E_{c}(x) = S = \\{s_{1}, s_{2}, \\dots, s_{n}\\}$ . Inspired by AttnPainter [27], $E_{c}$ is composed of a Vision Transformer (ViT) encoder [7] and a cross-attention module followed by a self-attention module, which is shown in Fig. 2.", + "bbox": [ + 75, + 295, + 468, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the ViT encoder first encodes the input superpixel $x$ into image feature $T_{f}$ . To control the number of output paths $(n)$ and the parameter number in each path $(N_{p})$ , we employ a cross-attention module to calculate the correlation between the image feature $T_{f}$ and $n N_{p}$ -dimensional learnable path queries $T_{l}$ , and output an intermediate feature $T_{f}^{\\prime}$ with the shape of path parameters $(n \\times N_{p})$ , which is formulated as:", + "bbox": [ + 75, + 416, + 468, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {f} ^ {\\prime} = \\operatorname {S o f t m a x} \\left(\\frac {\\left(W _ {Q} T _ {l}\\right) \\left(W _ {K} T _ {f}\\right) ^ {T}}{\\sqrt {d}}\\right) \\left(W _ {V} T _ {f}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 547, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{Q}, W_{K}$ and $W_{V}$ are learnable query, key and value matrices.", + "bbox": [ + 75, + 590, + 468, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, a self-attention module is employed to process the image feature $T_f'$ and project it into the parameter space with $n$ paths, where each path contains $N_p$ parameters.", + "bbox": [ + 75, + 621, + 468, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training objectives. 1) Normalized Reconstruction Loss: We employ the differentiable renderer $R(\\cdot)$ in DiffVG [17] to render a raster image $\\hat{x} = R(S)$ from the predicted path parameters $S$ . Then, we train the coarse-stage model $E_{c}$ by minimizing the normalized $\\mathcal{L}_2$ distance between the rendering $\\hat{x} = R(S)$ and the target image:", + "bbox": [ + 75, + 667, + 468, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {2} = \\| \\hat {x} - x \\| ^ {2} \\cdot \\frac {\\sum_ {p} \\operatorname {m a s k} (p)}{w h}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 768, + 468, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w$ and $h$ are the width and height of the superpixel image $x$ and the superpixel mask mask, and $\\text{mask}(p)$ indicates the mask value (1 or 0) for pixel $p$ .", + "bbox": [ + 75, + 809, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2) Boundary Loss: To avoid the SVG path from crossing the superpixel boundary, we propose boundary loss to guide the paths to be inside the superpixel. We set the color of all", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/28ca982cacb49dad098847c5be0b79419cfec7462967d7ca8c273652d8f001cb.jpg", + "image_caption": [ + "Figure 4. Problem of training the refinement model with $\\mathcal{L}_2$ loss alone: optimizing a newly-added path on the canvas by $\\mathcal{L}_2$ gradually pulls it to disappear (as a suboptimal local minimum). With our proposed coarse paths guided training and DPW loss, the added path is successfully optimized to resemble the target." + ], + "image_footnote": [], + "bbox": [ + 545, + 85, + 846, + 178 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "predicted SVG paths to 1 (white) to get a new path sequence $S_{binary}$ . Then, we compute the boundary loss by:", + "bbox": [ + 498, + 258, + 890, + 291 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {B o u n d}} = \\underset {p \\sim \\operatorname {m a s k}} {\\mathbb {E}} (R \\left(S _ {\\text {b i n a r y}}\\right) \\cdot (1 - \\operatorname {m a s k})) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 300, + 890, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since $(1 - \\text{mask})$ is 1 outside the superpixel mask and 0 inside the mask, the loss term calculates the area of the paths that are outside the superpixel. When some SVG paths cross the superpixel boundary, the path area outside is penalized; while when all SVG paths are inside the superpixel, $\\mathcal{L}_{\\text{Bound}}$ reaches 0 (Fig. 3).", + "bbox": [ + 496, + 333, + 890, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3) Path Efficiency Loss: To enable our model to reconstruct the maximum amount of information with the fewest paths, we propose the path efficiency loss $\\mathcal{L}_{PE}$ . Specifically, for each path $i$ , an additional opacity parameter $\\beta_{i}$ is predicted. We treat the path as visible if $\\beta_{i} \\geq 0.5$ , and the loss $L_{PE}$ penalizes the case with more visible paths, i.e., to encourage reconstructing the image with as few paths as possible, calculated as:", + "bbox": [ + 496, + 424, + 890, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {P E} = \\sum_ {i = 1} ^ {n} \\operatorname {S i g n} \\left(\\beta_ {i} - 0. 5\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 553, + 890, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{Sign}(\\cdot)$ is the sign function. Since $\\operatorname{Sign}(\\beta_i)$ is not differentiable, we approximate $\\frac{\\partial \\operatorname{Sign}(\\beta_i)}{\\partial \\beta_i}$ by $\\operatorname{Sig}(\\beta_i)(1 - \\operatorname{Sig}(\\beta_i))$ , where $\\operatorname{Sig}(\\cdot)$ is the sigmoid function.", + "bbox": [ + 498, + 579, + 890, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final training objective is formulated as:", + "bbox": [ + 517, + 630, + 812, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {c} ^ {*} = \\underset {E _ {c}} {\\arg \\min } \\mathcal {L} _ {2} + \\lambda_ {\\text {B o u n d}} \\mathcal {L} _ {\\text {B o u n d}} + \\lambda_ {P E} \\mathcal {L} _ {P E}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 652, + 890, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Coarse Paths Guided Refinement Stage", + "text_level": 1, + "bbox": [ + 500, + 696, + 836, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the coarse reconstruction stage, our coarse-stage model $E_{c}$ can output an SVG that captures and reconstructs the main structure of the input superpixel $x$ . The rendered image from the SVG (denoted as $c_{1}$ ) resembles $x$ in general, but lacks some image details, especially when the superpixel is complex. To enrich image details, we employ a refinement model $E_{r}$ to predict more SVG paths to add more details based on the current canvas $c_{1}$ .", + "bbox": [ + 496, + 719, + 890, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model framework. Different from the coarse-stage model $E_{c}$ that only takes the target superpixel $x$ as input, the refinement model $E_{r}$ takes both the current canvas $c_{1}$ (rendered from the coarse stage output) and the target superpixel", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24895", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$x$ as inputs, and predict new paths $S^{\\prime} = \\{s_1^{\\prime}, s_2^{\\prime} \\cdots s_m^{\\prime}\\}$ to be overlaid onto the canvas to refine details. Specifically, 3 convolution layers followed by ReLU activations are employed to fuse the two input images into a feature map. After getting the fused feature map, $E_{r}$ shares the same structure as the coarse-stage model $E_{c}$ , which encodes the fused feature map by a ViT Encoder and maps the encoded features into path parameters by a cross-attention and a self-attention layer. To accelerate the training process, we inherit the weights of the ViT encoder in $E_{c}$ as an initialization.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Local optimal solution with $\\mathcal{L}_2$ loss. The goal of the refinement model is to reconstruct more details of the input superpixel based on the current canvas. A simple $\\mathcal{L}_2$ loss defined as follows is used as the reconstruction loss:", + "bbox": [ + 75, + 242, + 470, + 301 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {2} = \\| x - R ([ E _ {c} (x), E _ {r} (x, c _ {1}) ]) \\| ^ {2} \\cdot \\frac {\\sum_ {p} m a s k (p)}{w h}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 305, + 468, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the predicted path sequences by coarse-stage model $E_{c}$ and refinement model $E_{r}$ are concatenated together to get the final SVG result, and the rendered image of which is expected to resemble the input superpixel $x$ .", + "bbox": [ + 76, + 340, + 468, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the refinement model $E_r$ trained with Eq.(6) alone tends to predict paths that are extremely small in area, or even invisible. In Fig. 4, we use an example to illustrate this phenomenon more clearly: we newly add a path onto the canvas and optimize the path parameters with $\\mathcal{L}_2$ loss; it can be seen that the new path gradually shrinks and finally disappears in the canvas. A possible reason is that the coarse stage result $c_1$ is already close to $x$ , and a local optimal solution for $E_r$ is to overlap nothing onto $c_1$ , which is better than adding a sub-optimal path and can prevent the $\\mathcal{L}_2$ distance from increasing.", + "bbox": [ + 75, + 402, + 468, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Coarse paths guided training framework. To avoid the refinement model from falling into poor local optimum, we propose a coarse paths guided framework, which inherits the knowledge from the coarse-stage model to help train the refinement model with an additional constraint on the SVG path parameters. As illustrated in Fig. 2, for an input superpixel $x$ , we first use the coarse-stage model to predict a coarse level path sequence $S = \\{s_1, s_2 \\dots s_n\\}$ . Then, we randomly choose a value $k \\in (1, n - m)$ and split the predicted path sequence into two subsequences: $S_1 = \\{s_1, s_2, \\dots, s_k\\}$ and $S_2 = \\{s_{k + 1}, s_{k + 2}, \\dots, s_n\\}$ . The subsequence $S_1$ is then rendered into $c_1 = R(S_1)$ and used as the input canvas for the refinement model $E_r$ , while the remaining subsequence $S_2$ can be regarded as a pseudo ground truth for the output path sequence of $E_r$ . Specifically, when training $E_r$ , in addition to the previous constraints that operate in the pixel space, we add a new constraint on the path parameter space, which minimizes the distance between path sequences $S'$ and $S_2$ .", + "bbox": [ + 75, + 568, + 470, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In detail, during training, the refinement model $E_{r}$ takes both the canvas $c_{1} = R(S_{1})$ and the target image $x$ as inputs, and outputs the path sequence $S^{\\prime} = \\{s_1^{\\prime},s_2^{\\prime},\\dots ,s_m^{\\prime}\\}$ .", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Forward pass to efficiently compute $\\mathbf{dpw}_{\\gamma}(\\mathbf{S},\\mathbf{S}^{\\prime})$", + "bbox": [ + 500, + 90, + 890, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: $S, S'$ , smoothing $\\gamma \\geq 0$ , distance function $d$", + "bbox": [ + 500, + 111, + 813, + 125 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} 1: p _ {0, j} = 0; p _ {i, 0} = q _ {i, 0} = q _ {0, j} = \\infty , i \\in [ [ n ] ], j \\in [ [ m ] ] \\\\ 2: \\text {f o r} j = 1, \\dots , m \\text {d o} \\\\ 3: \\quad \\text {f o r} i = 1, \\dots , n \\mathbf {d o} \\\\ 4: \\quad p _ {i, j} = d _ {i, j} + \\min ^ {\\gamma} \\left(q _ {i, j - 1}, p _ {i, j - 1}\\right) \\\\ 5: \\quad q _ {i, j} = \\min ^ {\\gamma} \\left(q _ {i - 1, j}, p _ {i - 1, j}\\right) \\\\ 6: \\quad e n d f o r \\\\ 7: \\text {e n d f o r} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 126, + 839, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: $(\\min^{\\gamma}(p_{n,m},q_{n,m}),P,Q)$", + "bbox": [ + 501, + 222, + 720, + 237 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8ec5c60eed98ffe5f022368c197173c389f6a1d4b15c5d5a37d5fd21d67eea2e.jpg", + "image_caption": [ + "(a) Difference in algorithms" + ], + "image_footnote": [], + "bbox": [ + 519, + 252, + 867, + 406 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9991f250cb023d47a691ae2f8399607781639e282d071ee56d6d8363b063bfcf.jpg", + "image_caption": [ + "Figure 5. Difference between DTW and our DPW. a) Both the DTW and DPW loss calculate the sum of distances of elements colored yellow. The difference is that one generated path can only match one target path in DPW to avoid averaging several target paths. b) The comparison between the training processes." + ], + "image_footnote": [], + "bbox": [ + 535, + 417, + 844, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We minimize both the distance in pixel space and path parameter space, with the total loss formulated as:", + "bbox": [ + 496, + 595, + 890, + 627 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {r} ^ {*} = \\underset {E _ {r}} {\\arg \\min } \\mathcal {L} _ {2} + \\lambda_ {D P W} \\mathcal {L} _ {D P W} + \\lambda_ {B o u n d} \\mathcal {L} _ {B o u n d}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 635, + 890, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{DPW}$ (detailed in Sec. 3.3) is the distance between the path sequences $S_{2}$ and $S^{\\prime}$ in the path parameter space.", + "bbox": [ + 496, + 674, + 890, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Dynamic Path Warping", + "text_level": 1, + "bbox": [ + 500, + 713, + 718, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To calculate the distance between the target path sequence $S = \\{s_1, s_2, \\dots, s_n\\}$ and the predicted path sequence $S' = \\{s_1', s_2', \\dots, s_m'\\}$ , Dynamic Time Warping (DTW) [4] is a commonly used metric, which finds an optimal matching match between the two ordered sequences (the yellow points in Fig. 5(a)) to minimize the accumulated distance $\\sum_{i}^{n} \\| s_i - s_{match(i)}' \\|^2$ where match(i + 1) ≥ match(i).", + "bbox": [ + 496, + 736, + 890, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that the monotonicity of the matching function cannot be ignored since the path sequences are well-ordered where latter paths are overlaid on former paths.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24896", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our coarse paths guided framework, we expect the generated path sequence to be a subsequence of the target path sequence. However, in DTW, one generated path $s_j'$ can correspond to several target paths $s_{i1},\\ldots ,s_{il}$ . Therefore, when trained with DTW, one generated path tends to become the average of several target paths. As shown in Fig. 5(b), when optimizing 3 paths to match the target emoji (with 4 paths), one path becomes the average of the two eyebrow paths, which is not our desired case.", + "bbox": [ + 75, + 90, + 472, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address this issue, we propose Dynamic Path Warping (DPW), where each generated path should match one and only one target path, and some target paths can be skipped (to learn a subsequence), as shown in Fig. 5(a), each horizontal line only passes through one matching point (yellow). To compute the DPW, we define $p_{i,j}$ as the minimum accumulated distance when $s_i$ matches $s_j'$ , and $q_{i,j}$ as the minimum accumulated distance when $s_j'$ has been matched to one path before $s_i$ (not including $s_i$ ). We employ dynamic programming to compute the final DPW loss $p_{n,m}$ as shown in Alg. 1. For each $p_{i,j}$ , the distance $d_{i,j}$ between $s_i$ and $s_j'$ is added to the smaller one of $q_{i,j-1}$ and $p_{i,j-1}$ . And for each $q_{i,j}$ , its value takes the smaller one between $q_{i-1,j}$ and $p_{i-1,j}$ (more explanations are provided in the supplementary material). Moreover, to make Alg. 1 differentiable, we follow SoftDTW [4] to substitute the $\\min(\\cdot)$ operation:", + "bbox": [ + 75, + 226, + 472, + 470 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min^ {\\gamma} (a _ {0}, a _ {1},..., a _ {n}) = \\left\\{ \\begin{array}{l l} \\min _ {i \\leq n} a _ {i} & \\gamma = 0, \\\\ - \\gamma \\log \\sum_ {i = 1} ^ {n} e ^ {- a _ {i} / \\gamma} & \\gamma > 0. \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 478, + 457, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 535, + 209, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Setting", + "text_level": 1, + "bbox": [ + 76, + 560, + 264, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We use SVG paths composed by cubic Bézier curves as the vector primitive, where each SVG path is closed, composed of 4 cubic Bézier curves connected end-to-end and has a fill color. Each SVG path has 28 parameters (24 for shape, 3 for color, and 1 for visibility). The coarse-stage model is trained to predict 128 paths for each superpixel first. Then, the refinement model is trained to predict 8 paths at one time. We train both the coarse-stage model and refinement model on ImageNet dataset [6]. We set batch size as 64 and learning rate as $2.5 \\times 10^{-4}$ . We train the coarse-stage model for 200K iterations with 5K warm up iterations, and train the refinement model for 200K iterations with $\\lambda_{DPW}$ decreasing from $1 \\times 10^{-3}$ to 0 in 10K iterations uniformly. In the following experiments, we implement our model with two versions: 1) SuperSVG-B, that decomposes the image into superpixels and vectorizes them by the coarse-stage model and refinement model, and 2) SuperSVG-F, which finetunes the SVG parameters from SuperSVG-B with $\\mathcal{L}_2$ loss, which takes around 10 seconds for optimization. All experiments are carried out on an NVIDIA GeForce RTX 4090 24GB GPU.", + "bbox": [ + 75, + 583, + 472, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Details. For quantitative evaluation and comparison, we test our model on 1,000 images randomly selected from ImageNet test set, and convert each image into SVGs with 500, 2,000 and 4,000 paths respectively. With the predicted SVG paths, we evaluate the reconstruction accuracy of the output SVG with the following 4 metrics: 1) MSE Distance and 2) PSNR to measure the pixel distance between the input image and the rendering from SVG; 3) LPIPS [35] to evaluate the perceptual distance, and 4) SSIM [31] to measure the structural distance. We further compare with Im2Vec [21] on EMOJIS dataset [2].", + "bbox": [ + 496, + 90, + 893, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Image to SVG Comparison", + "text_level": 1, + "bbox": [ + 498, + 267, + 743, + 284 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "State-of-the-art methods. The state-of-the-art methods can be classified into 3 categories: 1) Algorithm-based methods: Potrace [24] employs edge tracing to vectorize binary images. To process color images, we follow Color Trace1 to first quantize color images into different layers and then convert each layer to SVG using Potrace. Adobe Illustrator [1] is a widely-used commercial software which converts an image into SVG through image tracing. 2) Deep-learning-based methods: Im2Vec [21] encodes the target image into latent and predicts the vector paths with LSTM (#suppl.); and 3) Optimization-based methods: DiffVG [17] optimizes path parameters with random initialization and LIVE [20] employs layer-wise optimization to ensure a good vectorization structure. We use the official codes of these methods and default settings for comparison.", + "bbox": [ + 496, + 291, + 893, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparison on ImageNet. We compare with the state-of-the-art vectorization methods in reconstruction accuracy on ImageNet. Specifically, we conduct the comparison experiments under path numbers 500, 2,000 and $4,000^2$ . The qualitative results are shown in Fig. 6. It can be seen that Potrace cannot reconstruct the image well. LIVE loses a lot of details in relatively smooth areas due to its emphasis on regions with substantial color variations. DiffVG and Adobe work better when the path number increases, but they reconstruct fewer details than our SuperSVG. In comparison, our SuperSVG-B reconstructs most of the details with a short inference time. And by optimizing the SVG parameters from SuperSVG-B with only 10 seconds, our SuperSVG-F achieves the best reconstruction accuracy under different SVG path numbers.", + "bbox": [ + 496, + 518, + 893, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison on ImageNet. We further conduct quantitative comparison on 1,000 images randomly sampled from ImageNet [6] dataset (50 images for LIVE due to the extremely long optimization time: each input image takes about 6 GPU hours to optimize under 500 SVG paths). The quantitative results are presented in Tab. 1. Our SuperSVG achieves the best image vectorization results.", + "bbox": [ + 496, + 746, + 893, + 851 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2Since Adobe and Potrace outputs have different number of parameters in each path, we keep their output parameter number the same as ours.", + "bbox": [ + 501, + 875, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24897", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/18fc9e563be8472eb8c120edfa49fd705936706691eaf5b89ec0c76d06f66dc6.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison with the state-of-the-art methods in image vectorization with different number of SVG paths." + ], + "image_footnote": [], + "bbox": [ + 178, + 89, + 795, + 506 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a3300ac45e8309cc4cedb3b18f263f2b7461b596729b1098344db555ff5669b5.jpg", + "table_caption": [ + "Table 1. Quantitative comparison between the state-of-the-arts and Ours. Bold and underline for the best and the second best results." + ], + "table_footnote": [], + "table_body": "
#PathsMethodTime (s) ↓MSE ↓PSNR ↑LPIPS ↓SSIM ↑
500LIVE [20]20,0000.003924.100.44670.7983
DiffVG [17]19.290.006921.420.53190.6671
Adobe [1]0.870.006721.820.55950.6939
Potrace [24]0.980.020817.850.51150.6920
SuperSVG-B (Ours)0.310.004424.800.44520.7687
SuperSVG-F (Ours)10.000.002627.460.35580.8383
2,000LIVE [20]120,0000.002526.980.39940.8431
DiffVG [17]73.650.003625.880.46830.7710
Adobe [1]2.150.003326.230.39610.7229
Potrace [24]3.100.016019.650.43550.6997
SuperSVG-B (Ours)0.710.002427.250.36480.8446
SuperSVG-F (Ours)10.000.001729.120.29310.8828
4,000LIVE [20]300,0000.002426.800.39810.8446
DiffVG [17]140.340.002527.290.38580.8492
Adobe [1]3.120.003525.930.34080.7297
Potrace [24]5.110.011320.680.43800.7374
SuperSVG-B (Ours)1.040.001928.450.31870.8757
SuperSVG-F (Ours)10.000.001429.960.24960.9028
", + "bbox": [ + 81, + 555, + 468, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 845, + 230, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study on the Superpixel-based Framework. We first validate the effectiveness of the superpixel-based", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vectorization framework. We train a model that directly predicts the SVG paths for an input image, without superpixel segmentation. Then, we test the model in two ways: 1) predict the SVG paths for the whole input image and 2) uniformly divide the input image into $4 \\times 4$ blocks and vectorize each block separately. We compare our model with these two models with the same number of paths (1,000). The results are shown in Fig. 7 and Tab. 2(a). The model without superpixel segmentation loses many image details. By introducing the block division, the model enriches the details, but the regions near block boundaries are discontinuous (shown in red box), producing unnatural results. In comparison, our superpixel-based SuperSVG-B reconstructs most details and keeps the image structures well.", + "bbox": [ + 496, + 536, + 890, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study on the Coarse-stage Model. We then evaluate the effectiveness of the boundary loss $\\mathcal{L}_{\\text{Bound}}$ and the path efficiency loss $\\mathcal{L}_{PE}$ in the coarse-stage model. We train 2 ablated models: 1) the model without $\\mathcal{L}_{\\text{Bound}}$ ; and 2) the model without $\\mathcal{L}_{PE}$ , and compare them with our coarse-stage model under 500 SVG paths. In this comparison, we only compare vectorization using the coarse-stage model, without using the refinement model. The results are shown in Fig. 8 and Tab. 2(b). The model without $L_{\\text{Bound}}$ predicts some paths that cross superpixel boundaries, re", + "bbox": [ + 496, + 750, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24898", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb0bd8a4ac218addf3223fbdf0c775c6426e2b4643261beb199062825959c733.jpg", + "image_caption": [ + "Figure 7. Ablation study on the superpixel-based image vectorization framework. The ablated models either cannot reconstruct most details or suffer from boundary inconsistency problem." + ], + "image_footnote": [], + "bbox": [ + 116, + 88, + 426, + 226 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1dc0cd3060dd4f4aa0b2993cc730c141b811fcbba9e4de89350b92de5f4b7449.jpg", + "image_caption": [ + "Figure 8. Ablation study on the coarse-stage model. The ablated models either predict paths crossing superpixel boundaries or reconstruct less details than ours." + ], + "image_footnote": [], + "bbox": [ + 116, + 282, + 426, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sulting in worse reconstruction. The model without $L_{PE}$ has a poorer performance, which is validated by the metric results. In comparison, our model outperforms the two ablated models, validating the effectiveness of the losses $\\mathcal{L}_{\\text{Bound}}$ and $\\mathcal{L}_{PE}$ in the coarse-stage model.", + "bbox": [ + 75, + 489, + 468, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study on the Refinement-stage Model. Finally, we validate the effectiveness of the refinement stage and the DPW loss. We train 3 ablated models: 1) the model without the refinement stage; 2) the model without the DPW loss $\\mathcal{L}_{DPW}$ (i.e., without coarse paths guidance, with pixel-wise loss only); and 3) the model replacing DPW loss $\\mathcal{L}_{DPW}$ with $\\mathcal{L}_2$ loss in path parameter space. The results are shown in Fig. 9 and Tab. 2(c). The model without refinement-stage cannot reconstruct as many details as ours. The ablated model without $\\mathcal{L}_{DPW}$ predicts paths with a very small area or even invisible, as explained in Sec.3.2, thus the results look similar to the coarse-stage results. For the ablated model replacing DPW loss $\\mathcal{L}_{DPW}$ with $\\mathcal{L}_2$ in path parameter space, which enforces one-to-one alignment between two paths' control points, since the constraint is too strict, the loss cannot function well in experiments, and the results look alike the results without $\\mathcal{L}_{DPW}$ . In comparison, our model outperforms the ablated models, validating the effectiveness of the refinement and DPW loss.", + "bbox": [ + 75, + 568, + 468, + 854 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/402d531a0b8f73adc39f9178bae87fd012bdf2623cb514abf1919e4ec0b2eba7.jpg", + "image_caption": [ + "Figure 9. Ablation study on the refinement-stage model. The ablated models either lose details or converge to the poor local optimum described in Sec. 3.2." + ], + "image_footnote": [], + "bbox": [ + 540, + 88, + 849, + 226 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9adec7c9fc1f008ae0f5a5ead7ab5df338b9d1236686c411c1544152177c510d.jpg", + "table_caption": [ + "Table 2. Quantitative results of ablation studies." + ], + "table_footnote": [], + "table_body": "
MethodMSE ↓PSNR ↑LPIPS ↓SSIM ↑
w/o superpixel0.008321.720.50570.6900
4 × 4 blocks0.003829.500.42990.7834
Ours0.003226.040.40750.8111
(a) Ablation on superpixel-based framework.
MethodMSE ↓PSNR ↑LPIPS ↓SSIM ↑
w/o LBound0.053413.380.50410.6210
w/o LPE0.006322.980.48300.7219
Ours (Coarse)0.004524.490.44520.7673
", + "bbox": [ + 522, + 284, + 870, + 415 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/05ece90da5f0b3255aa2ac0226430fb847cf447b6bec773c58c6d59b71c99259.jpg", + "table_caption": [ + "(b) Ablation on coarse-stage model (results are obtained by the coarse-stage model alone, without refinement model)." + ], + "table_footnote": [], + "table_body": "
MethodMSE ↓PSNR ↑LPIPS ↓SSIM ↑
w/o refine0.004125.020.43750.7770
w/o LDPW0.004524.490.44520.7673
L2 replacing LDPW0.004524.490.44520.7673
Ours0.003226.040.40750.8111
", + "bbox": [ + 522, + 445, + 867, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(c) Ablation on refinement-stage model.", + "bbox": [ + 576, + 515, + 813, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 534, + 617, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose SuperSVG, a novel superpixel-based vectorization model that decomposes a raster image into superpixels and then vectorizes each separately, achieving fast and accurate image vectorization. We propose a two-stage self-teaching framework, where a coarse-stage model reconstructs main structure and a refinement model enriches details, with a novel dynamic path warping loss that guides the refinement model by inheriting knowledge from coarse paths. Extensive experiments demonstrate that SuperSVG achieves the state-of-the-art performance on vectorization.", + "bbox": [ + 496, + 559, + 890, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 724, + 658, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by National Natural Science Foundation of China (62302297, 72192821, 62272447), Shanghai Sailing Program (22YF1420300), Young Elite Scientists Sponsorship Program by CAST (2022QNRC001), Beijing Natural Science Foundation (L222117), the Fundamental Research Funds for the Central Universities (YG2023QNB17), Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), Shanghai Science and Technology Commission (21511101200).", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "3Since the model without refinement does not contain refinement paths, we increase the number of coarse paths to keep the path number consistent.", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24899", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adobe Illustrator. https://www.adobe.com/products/illustrator.html.2,6,7", + "[2] Noto Emoji. https://github.com/googlefonts/ noto-emoji.6", + "[3] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and Sabine Susstrunk. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 2274-2282, 2012. 2, 3", + "[4] Marco Cuturi and Mathieu Blondel. Soft-DTW: a differentiable loss function for time-series. In International conference on machine learning, pages 894-903. PMLR, 2017. 5, 6", + "[5] Wen Dai, Tao Luo, and Jianbing Shen. Automatic image vectorization using superpixels and random walkers. In 2013 6th International Congress on Image and Signal Processing (CISP), 2013. 2", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 6", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth $16 \\times 16$ words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4", + "[8] Maria Dziuba, Ivan Jarsky, Valeria Efimova, and Andrey Filchenkov. Image vectorization: a review. arXiv preprint arXiv:2306.06441, 2023. 3", + "[9] Jun Gao, Chengcheng Tang, Vignesh Ganapathi-Subramanian, Jiahui Huang, Hao Su, and LeonidasJ. Guibas. DeepSpline: Data-driven reconstruction of parametric curves and surfaces. 2019. 1, 2", + "[10] David Ha and Douglas Eck. A neural representation of sketch drawings. Learning, Learning, 2017. 1", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2", + "[12] Teng Hu, Ran Yi, Haokun Zhu, Liang Liu, Jinlong Peng, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. Stroke-based neural painting and stylization with dynamically predicted painting region. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7470-7480, 2023. 1", + "[13] Varun Jampani, Deqing Sun, Ming-Yu Liu, Ming-Hsuan Yang, and Jan Kautz. Superpixel sampling networks. arXiv: Computer Vision and Pattern Recognition, 2018. 2", + "[14] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2", + "[15] Yu-Kun Lai, Shi-Min Hu, and Ralph R. Martin. Automatic and topology-preserving gradient mesh generation for image vectorization. ACM Transactions on Graphics, 28(3):1-8, 2009. 1, 2" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] A. Levinshtein, A. Stere, K.N. Kutulakos, D.J. Fleet, S.J. Dickinson, and K. Siddiqi. TurboPixels: Fast superpixels using geometric flows. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 2290-2297, 2009. 2", + "[17] Tzu-Mao Li, Michal Lukáč, Michael Gharbi, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for editing and learning. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 1, 2, 3, 4, 6, 7", + "[18] Zicheng Liao, H. Hoppe, D. Forsyth, and Yizhou Yu. A subdivision-based representation for vector image editing. IEEE Transactions on Visualization and Computer Graphics, 18(11):1858-1867, 2012. 1, 2", + "[19] Chen Liu, Jiajun Wu, Pushmeet Kohli, and Yasutaka Furukawa. Raster-to-vector: Revisiting floorplan transformation. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 1, 2", + "[20] Xu Ma, Yuqian Zhou, Xingqian Xu, Bin Sun, Valerii Filev, Nikita Orlov, Yun Fu, and Humphrey Shi. Towards layerwise image vectorization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16314-16323, 2022. 1, 2, 3, 6, 7", + "[21] Pradyumna Reddy, Michael Gharbi, Michal Lukac, and Niloy J Mitra. Im2Vec: Synthesizing vector graphics without vector supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7342-7351, 2021. 1, 2, 3, 6", + "[22] Leo Sampaio Ferraz Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based representation for sketched structure. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1", + "[23] Mike Schuster and Kuldip K Paliwal. Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing, 45(11):2673-2681, 1997. 2", + "[24] Peter Selinger. Potrace: a polygon-based tracing algorithm, 2003. 1, 2, 6, 7", + "[25] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2", + "[26] Bohao Tang, Teng Hu, Yuzhen Du, Ran Yu, and Lizhuang Ma. Curved-stroke-based neural painting and stylization through thin plate spline interpolation. Scientia Sinica Informationis, 54(2):301-315, 2024. 1", + "[27] Yizhe Tang, Yue Wang, Teng Hu, Xin Tan, and Ran Yi. AttnPainter: A scalable stroke predictor for any scene painting. arXiv preprint, 2024. 4", + "[28] Sebastian Thrun and James Diebel. Bayesian image vectorization: the probabilistic inversion of vector image rasterization. 2008. 1, 2", + "[29] Wei-Chih Tu, Ming-Yu Liu, Varun Jampani, Deqing Sun, Shao-Yi Chien, Ming-Hsuan Yang, and Jan Kautz. Learning superpixels with segmentation-aware affinity loss. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 568-576, 2018. 2", + "[30] Michael Van den Bergh, Xavier Boix, Gemma Roig, Benjamin De Capitani, and Luc Van Gool. SEEDS: Superpixels" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24900", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "extracted via energy-driven sampling. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision (ECCV), pages 13-26. Springer, 2012. 2", + "[31] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6", + "[32] Guofu Xie, Xin Sun, Xin Tong, and Derek Nowrouzezahrai. Hierarchical diffusion curves for accurate automatic image vectorization. ACM Transactions on Graphics, 33(6):1-11, 2014. 2", + "[33] Ming Yang, Hongyang Chao, Chi Zhang, Jun Guo, Lu Yuan, and Jian Sun. Effective clipart image vectorization through direct optimization of bezigons. IEEE Transactions on Visualization and Computer Graphics, page 1063-1075, 2016. 1", + "[34] Zipeng Ye, Ran Yi, Minjing Yu, Yong-Jin Liu, and Ying He. Fast computation of content-sensitive superpixels and supervoxels using q-distances. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3770-3779, 2019. 2", + "[35] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 6", + "[36] Shuang Zhao, Frédo Durand, and Changxi Zheng. Inverse diffusion curves using shape optimization. IEEE Transactions on Visualization and Computer Graphics, 2016. 2", + "[37] Hailing Zhou, Jianmin Zheng, and Lei Wei. Representing images using curvilinear feature driven subdivision surfaces. IEEE Transactions on Image Processing, 23(8):3268-3280, 2014. 1, 2", + "[38] Haokun Zhu, Juang Ian Chong, Teng Hu, Ran Yi, Yu-Kun Lai, and Paul L Rosin. SAMVG: A multi-stage image vectorization model with the segment-anything model. arXiv preprint arXiv:2311.05276, 2023. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 614 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24901", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/3fa1e5c3-5890-4208-976c-19acdb43c796_model.json b/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/3fa1e5c3-5890-4208-976c-19acdb43c796_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ce94d173d8e85befbb73970c9d8658ea716f21c0 --- /dev/null +++ b/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/3fa1e5c3-5890-4208-976c-19acdb43c796_model.json @@ -0,0 +1,1991 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.155, + 0.131, + 0.816, + 0.154 + ], + "angle": 0, + "content": "SuperSVG: Superpixel-based Scalable Vector Graphics Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.18, + 0.91, + 0.236 + ], + "angle": 0, + "content": "Teng Hu\\(^{1}\\), Ran Yi\\(^{1*}\\), Baihong Qian\\(^{1}\\), Jiangning Zhang\\(^{2}\\), Paul L. Rosin\\(^{3}\\), Yu-Kun Lai\\(^{3}\\) \n\\(^{1}\\)Shanghai Jiao Tong University, \\(^{2}\\)Youtu Lab, Tencent, \\(^{3}\\)Cardiff University \n{hu-teng, ranyi, cherry_qbh} \\(@\\) sjtu.edu.cn, vtzhang@tencent.com, {RosinPL, LaiY4} \\(@\\) cardiff.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.302, + 0.474, + 0.65 + ], + "angle": 0, + "content": "SVG (Scalable Vector Graphics) is a widely used graphics format that possesses excellent scalability and editability. Image vectorization, which aims to convert raster images to SVGs, is an important yet challenging problem in computer vision and graphics. Existing image vectorization methods either suffer from low reconstruction accuracy for complex images or require long computation time. To address this issue, we propose SuperSVG, a superpixel-based vectorization model that achieves fast and high-precision image vectorization. Specifically, we decompose the input image into superpixels to help the model focus on areas with similar colors and textures. Then, we propose a two-stage self-training framework, where a coarse-stage model is employed to reconstruct the main structure and a refinement-stage model is used for enriching the details. Moreover, we propose a novel dynamic path warping loss to help the refinement-stage model to inherit knowledge from the coarse-stage model. Extensive qualitative and quantitative experiments demonstrate the superior performance of our method in terms of reconstruction accuracy and inference time compared to state-of-the-art approaches. The code is available in https://github.com/sjtuplayer/SuperSVG." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.681, + 0.21, + 0.697 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.47, + 0.858 + ], + "angle": 0, + "content": "Scalable Vector Graphics, commonly known as SVG, is a widely used vector image format that has a wide range of applications and advantages within the domains of web design, graphic design, mobile applications, data visualization, and various other contexts. Compared with raster images that represent content by pixels, SVG describes images by parameterized vectors and benefits from its scalability and editability where it can be resized to any resolution without losing quality and can be easily manipulated by its layer-wise topological information." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.859, + 0.469, + 0.875 + ], + "angle": 0, + "content": "Given the superior capabilities of Scalable Vector Graph" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.268, + 0.875, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.444, + 0.895, + 0.556 + ], + "angle": 0, + "content": "Figure 1. Overview of our SuperSVG: our model first decomposes the image to be vectorized into superpixels, each containing pixels sharing similar colors and contents. The coarse-stage model predicts the path parameters to reconstruct the main structure, and then the coarse paths guided refinement model enriches the details by learning the knowledge from the coarse-stage model. Compared to the previous methods, our SuperSVG achieves both a high vectorization accuracy and fast computation speed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.886 + ], + "angle": 0, + "content": "ics (SVG) in image representation and editing, there is much research on the topic of image vectorization, which aims to convert rasterized images into SVG. The existing methods can be categorized into three classes: 1) Traditional algorithm-based methods [15, 18, 24, 28, 33, 37], where conventional algorithms are employed to fit images, but they usually suffer from a lower vectorization quality. 2) Deep-learning-based methods [9, 10, 12, 19, 21, 22], which parameterize raster images using deep neural networks for reconstruction. They are efficient and can handle the vectorization of simple graphics or characters (e.g., icons and emojis), but struggle to vectorize complex images. 3) Optimization-based methods [17, 20, 26, 38], which optimize SVG parameters to fit the target image, yielding relatively superior reconstruction quality. However, these methods entail a substantial amount of time and computational resources, making them impractical for timely processing of large-scale data. In summary, previous image vectorization methods either suffer from low reconstruction quality for complex images, or demand extensive computation time, imposing significant constraints on their practical utility." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To achieve good vectorization quality with high ef" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24892" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.408 + ], + "angle": 0, + "content": "ficiency, we propose SuperSVG, a deep-learning-based method that translates images into scalable vector graphics (SVG) in a coarse-to-fine manner. Since neural networks have difficulties in directly vectorizing complex images, we decompose the input image into different parts in the form of superpixels wherein the pixels share similar colors and textures and then vectorize each part. Then, we propose a Two-Stage Self-Teaching Training framework to vectorize the superpixels, where the coarse-stage model is trained to reconstruct the main structure of the image and the refinement-stage model is trained to enrich the image details. We make use of the predicted paths from the coarse-stage model to guide the refinement-stage model in image vectorization. Furthermore, we propose a novel Dynamic Path Warping loss which helps the refinement-stage to inherit the knowledge of the coarse-stage model. With the help of the superpixel-based image decomposition and the two-stage self-teaching framework, our SuperSVG can keep the image structure well and reconstruct more details at high speed. Extensive quantitative and qualitative experiments validate the effectiveness of our model." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.41, + 0.427, + 0.424 + ], + "angle": 0, + "content": "The main contributions of our work are four-fold:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.429, + 0.47, + 0.488 + ], + "angle": 0, + "content": "- We propose SuperSVG, a novel superpixel-based vectorization model that translates the rasterized images into scalable vector graphics (SVG) based on superpixels and vectorizes the superpixels in a coarse-to-fine manner." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.489, + 0.47, + 0.564 + ], + "angle": 0, + "content": "- We design a Two-Stage Self-Teaching Training framework, where we employ a coarse-stage model to reconstruct the main structures and a refinement-stage model to enrich the image details based on the coarse-stage output." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.565, + 0.47, + 0.639 + ], + "angle": 0, + "content": "- We propose a Coarse Paths Guided Training strategy to guide the refinement-stage model to inherit the knowledge from the coarse-stage model, which greatly improves the performance of the refinement-stage model and avoids converging to suboptimal local minimum." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.64, + 0.47, + 0.73 + ], + "angle": 0, + "content": "- We propose Dynamic Path Warping (DPW) loss, which measures the distance between the predicted paths from the refinement-stage model and the pseudo ground truth approximated with coarse paths. By minimizing the DPW loss, the refinement-stage model can distill the knowledge from the coarse-stage model." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.429, + 0.47, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.746, + 0.22, + 0.761 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.771, + 0.268, + 0.787 + ], + "angle": 0, + "content": "2.1. Image Vectorization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Image vectorization aims to transform a rasterized image into scalable vector graphics (SVG) composed of parameterized vectors. Different from raster images that may become blurry when zooming in, SVG can be rendered at any resolution without losing quality and is convenient to edit, widely used in web design, graphic design, etc. The existing vectorization methods can be classified into 3 categories:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.257 + ], + "angle": 0, + "content": "Traditional Algorithm-based Image Vectorization Methods can be classified into mesh-based and curve-based ones. The mesh-based methods [15, 18, 28, 37] segment an input image into non-overlapping patches, and infer the color and the boundary location for each region. The curve-based methods [1, 5, 24, 32, 36] employ Bézier curves with different colors defined on either side to create the vector image. Potrace [24] is a representative method of this type that projects the smooth outlines into Bézier paths, and merge the adjacent paths together. However, the vectorization quality of these methods still needs improvements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.893, + 0.483 + ], + "angle": 0, + "content": "Deep-learning-based Image Vectorization Methods use neural networks to project a raster image into SVG. Im2Vec [21] employs a variational auto-encoder (VAE) [14] to embed the input image and then maps it into path parameters by a Long Short-Term Memory (LSTM) module [23]. Raster2Vec [19] is focused on vectorization of rasterized floor plans using a ResNet [11]. Gao et al. [9] rely on a pre-trained VGG network [25] and a hierarchical Recurrent Neural Network (RNN) to output parametric curves of different sizes. But these methods only focus on simple images and cannot vectorize complex images well. In comparison, our SuperSVG is the first deep-learning-based method that can vectorize images with complex details, thanks to our superpixel decomposition and coarse-path guided refinement that substantially reduce the learning difficulties." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.892, + 0.651 + ], + "angle": 0, + "content": "Optimization-based Image Vectorization Methods. DiffVG [17] proposes a differentiable renderer that renders the SVG parameters into images. Based on this, DiffVG minimizes the distance between the rasterized and vector images by optimizing the SVG parameters using gradient descent. LIVE [20] and SAMVG [38] further introduce a layer-wise optimization framework, which achieves better vectorization quality over the previous methods. However, due to the low optimization efficiency, they suffer from a long optimization time. In contrast, our SuperSVG achieves both a good vectorization quality and high efficiency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.659, + 0.738, + 0.675 + ], + "angle": 0, + "content": "2.2. Superpixel Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.892, + 0.832 + ], + "angle": 0, + "content": "Superpixel decomposition is usually used for data preprocessing in vision tasks. Existing superpixel decomposition methods can be categorized into methods based on traditional algorithm or deep learning. For the traditional algorithm based methods, diverse strategies have been employed, e.g., energy-driven sampling [30], geometric flows [16] and clustering [3]. Some recent works [13, 29, 34] employ neural networks to enhance the performance in superpixel decomposition, which shows great potential in this task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.591, + 0.86 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Image vectorization aims to translate a rasterized image \\(I\\) into a Scalable Vector Graphic (SVG). An SVG is com" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24893" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.087, + 0.852, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.23, + 0.846, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.37, + 0.892, + 0.427 + ], + "angle": 0, + "content": "Figure 2. Main framework of our SuperSVG: we decompose the target image into superpixels and vectorize each superpixel separately. We employ an attention-based coarse-stage model to predict SVG paths that reconstruct the main structure of the superpixel. Then, a refinement-stage model guided by the coarse paths is designed to predict more SVG paths to refine details based on the coarse image. Finally, by combining all the predicted SVGs for each superpixel, we obtain an output SVG with good structure and fine details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.665 + ], + "angle": 0, + "content": "posed of many vector primitives, which can be SVG paths, ellipses, circles, or rectangles, etc. Following previous works [17, 20, 21], we employ the SVG paths as the shape primitive, where each SVG path defines a region constructed by multiple cubic Bezier curves connected end-to-end with certain color. With the parameters of these SVG paths, the rasterized image can be rendered in any resolution. To obtain the path parameters, some previous methods [17, 20] optimize the path parameters directly to minimize the distance between the input image and the rendered image, which achieves good reconstruction quality but requires long optimization time. To speed up the vectorization process, some deep-learning-based methods [21] employ a deep-learning model to predict the SVG path parameters, but struggle to vectorize complex images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To achieve good vectorization quality with high efficiency, we propose SuperSVG, a deep-learning-based image vectorization method that translates images into SVG path parameters automatically. To improve the model ability to vectorize complex images, we segment the input image into different parts, within which the pixels share similar colors and textures, and then vectorize each part separately, where superpixels are used for image segmentation as they tend to maintain compactness, uniformity and regularity, particularly suitable for our task. For each superpixel \\( x \\in \\mathcal{X} \\) where \\( \\mathcal{X} \\) is the set of all superpixels, our model converts it into a sequence of path parameters, where each path is composed of several cubic Bezier curves and has a fill color, with a total of \\( N_{p} \\) parameters. With the predicted path parameters for each superpixel, we employ the differ" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.439, + 0.892, + 0.484 + ], + "angle": 0, + "content": "entiable renderer \\( R(\\cdot) \\) from DiffVG [17] to get the rendered image \\( \\hat{I} \\) in pixel space, which is expected to be close to the input image \\( I \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.486, + 0.893, + 0.651 + ], + "angle": 0, + "content": "We propose a two-stage self-teaching framework, composed of a coarse-stage model \\( E_{c} \\) to reconstruct the main structure and a refinement model \\( E_{r} \\) to enrich the details, where the predicted paths from coarse-stage model are used to guide the refinement model in vectorization. \\( E_{c} \\) takes the superpixel \\( x \\) as input and outputs \\( n \\) paths \\( S = \\{s_1,s_2,\\dots s_n\\} \\) to reconstruct main structure; while \\( E_{r} \\) takes both the rendered image \\( R(S) \\) and target superpixel \\( x \\) as inputs, and outputs \\( m \\) paths \\( S^{\\prime} = \\{s_{1}^{\\prime},s_{2}^{\\prime}\\dots s_{m}^{\\prime}\\} \\) to refine details. Combining all the predicted \\( S \\) and \\( S^{\\prime} \\) for each superpixel produces the final SVG result." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.848, + 0.681 + ], + "angle": 0, + "content": "3.1. Superpixel-based Coarse Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Superpixel decomposition. Considering the optimization-based methods suffer from a long optimization time, our SuperSVG builds upon neural networks to efficiently predict SVG paths. However, as neural networks have difficulties in directly vectorizing complex images [8] we therefore simplify the task to vectorizing a certain part of the image containing homogeneous colors and textures. Since superpixel algorithms provide a good tool to decompose images based on local pixel color and also ensure alignment of the regions with the image boundaries, we segment the input image into superpixels, and our model reconstructs each superpixel with scalable vectors. Superpixels also tend to be more regular, making them easier for vectorization. Specifically, we utilize SLIC [3] to decompose the input image" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24894" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.088, + 0.437, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.212, + 0.47, + 0.253 + ], + "angle": 0, + "content": "Figure 3. Illustration of our boundary loss \\(\\mathcal{L}_{\\text{Bound}}\\), which computes the area of the SVG paths that are outside the superpixel mask, and guides the paths to be inside the superpixel." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.47, + 0.296 + ], + "angle": 0, + "content": "into superpixels. We set the compactness parameter as 30 for SLIC to make superpixels more regular." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.469, + 0.417 + ], + "angle": 0, + "content": "Coarse-stage model. For a superpixel \\( x \\) with mask mask (indicating those pixels within \\( x \\) with 1 and 0 otherwise), we first design the coarse-stage model \\( E_{c} \\) to vectorize the main structure by predicting the SVG path sequence \\( E_{c}(x) = S = \\{s_{1}, s_{2}, \\dots, s_{n}\\} \\). Inspired by AttnPainter [27], \\( E_{c} \\) is composed of a Vision Transformer (ViT) encoder [7] and a cross-attention module followed by a self-attention module, which is shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.469, + 0.538 + ], + "angle": 0, + "content": "Specifically, the ViT encoder first encodes the input superpixel \\( x \\) into image feature \\( T_{f} \\). To control the number of output paths \\( (n) \\) and the parameter number in each path \\( (N_{p}) \\), we employ a cross-attention module to calculate the correlation between the image feature \\( T_{f} \\) and \\( n N_{p} \\)-dimensional learnable path queries \\( T_{l} \\), and output an intermediate feature \\( T_{f}^{\\prime} \\) with the shape of path parameters \\( (n \\times N_{p}) \\), which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.549, + 0.47, + 0.582 + ], + "angle": 0, + "content": "\\[\nT _ {f} ^ {\\prime} = \\operatorname {S o f t m a x} \\left(\\frac {\\left(W _ {Q} T _ {l}\\right) \\left(W _ {K} T _ {f}\\right) ^ {T}}{\\sqrt {d}}\\right) \\left(W _ {V} T _ {f}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.469, + 0.621 + ], + "angle": 0, + "content": "where \\( W_{Q}, W_{K} \\) and \\( W_{V} \\) are learnable query, key and value matrices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.469, + 0.668 + ], + "angle": 0, + "content": "Then, a self-attention module is employed to process the image feature \\( T_f' \\) and project it into the parameter space with \\( n \\) paths, where each path contains \\( N_p \\) parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.469, + 0.759 + ], + "angle": 0, + "content": "Training objectives. 1) Normalized Reconstruction Loss: We employ the differentiable renderer \\( R(\\cdot) \\) in DiffVG [17] to render a raster image \\( \\hat{x} = R(S) \\) from the predicted path parameters \\( S \\). Then, we train the coarse-stage model \\( E_{c} \\) by minimizing the normalized \\( \\mathcal{L}_2 \\) distance between the rendering \\( \\hat{x} = R(S) \\) and the target image:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.769, + 0.469, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {2} = \\| \\hat {x} - x \\| ^ {2} \\cdot \\frac {\\sum_ {p} \\operatorname {m a s k} (p)}{w h}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.856 + ], + "angle": 0, + "content": "where \\( w \\) and \\( h \\) are the width and height of the superpixel image \\( x \\) and the superpixel mask mask, and \\( \\text{mask}(p) \\) indicates the mask value (1 or 0) for pixel \\( p \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "2) Boundary Loss: To avoid the SVG path from crossing the superpixel boundary, we propose boundary loss to guide the paths to be inside the superpixel. We set the color of all" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.087, + 0.847, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.18, + 0.892, + 0.25 + ], + "angle": 0, + "content": "Figure 4. Problem of training the refinement model with \\(\\mathcal{L}_2\\) loss alone: optimizing a newly-added path on the canvas by \\(\\mathcal{L}_2\\) gradually pulls it to disappear (as a suboptimal local minimum). With our proposed coarse paths guided training and DPW loss, the added path is successfully optimized to resemble the target." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.26, + 0.892, + 0.292 + ], + "angle": 0, + "content": "predicted SVG paths to 1 (white) to get a new path sequence \\( S_{binary} \\). Then, we compute the boundary loss by:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.301, + 0.892, + 0.325 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {B o u n d}} = \\underset {p \\sim \\operatorname {m a s k}} {\\mathbb {E}} (R \\left(S _ {\\text {b i n a r y}}\\right) \\cdot (1 - \\operatorname {m a s k})) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.424 + ], + "angle": 0, + "content": "Since \\((1 - \\text{mask})\\) is 1 outside the superpixel mask and 0 inside the mask, the loss term calculates the area of the paths that are outside the superpixel. When some SVG paths cross the superpixel boundary, the path area outside is penalized; while when all SVG paths are inside the superpixel, \\(\\mathcal{L}_{\\text{Bound}}\\) reaches 0 (Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.545 + ], + "angle": 0, + "content": "3) Path Efficiency Loss: To enable our model to reconstruct the maximum amount of information with the fewest paths, we propose the path efficiency loss \\(\\mathcal{L}_{PE}\\). Specifically, for each path \\(i\\), an additional opacity parameter \\(\\beta_{i}\\) is predicted. We treat the path as visible if \\(\\beta_{i} \\geq 0.5\\), and the loss \\(L_{PE}\\) penalizes the case with more visible paths, i.e., to encourage reconstructing the image with as few paths as possible, calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.554, + 0.892, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {P E} = \\sum_ {i = 1} ^ {n} \\operatorname {S i g n} \\left(\\beta_ {i} - 0. 5\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.58, + 0.892, + 0.63 + ], + "angle": 0, + "content": "where \\( \\operatorname{Sign}(\\cdot) \\) is the sign function. Since \\( \\operatorname{Sign}(\\beta_i) \\) is not differentiable, we approximate \\( \\frac{\\partial \\operatorname{Sign}(\\beta_i)}{\\partial \\beta_i} \\) by \\( \\operatorname{Sig}(\\beta_i)(1 - \\operatorname{Sig}(\\beta_i)) \\), where \\( \\operatorname{Sig}(\\cdot) \\) is the sigmoid function." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.631, + 0.813, + 0.645 + ], + "angle": 0, + "content": "The final training objective is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.654, + 0.892, + 0.679 + ], + "angle": 0, + "content": "\\[\nE _ {c} ^ {*} = \\underset {E _ {c}} {\\arg \\min } \\mathcal {L} _ {2} + \\lambda_ {\\text {B o u n d}} \\mathcal {L} _ {\\text {B o u n d}} + \\lambda_ {P E} \\mathcal {L} _ {P E}. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.838, + 0.713 + ], + "angle": 0, + "content": "3.2. Coarse Paths Guided Refinement Stage" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.84 + ], + "angle": 0, + "content": "In the coarse reconstruction stage, our coarse-stage model \\( E_{c} \\) can output an SVG that captures and reconstructs the main structure of the input superpixel \\( x \\). The rendered image from the SVG (denoted as \\( c_{1} \\)) resembles \\( x \\) in general, but lacks some image details, especially when the superpixel is complex. To enrich image details, we employ a refinement model \\( E_{r} \\) to predict more SVG paths to add more details based on the current canvas \\( c_{1} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Model framework. Different from the coarse-stage model \\( E_{c} \\) that only takes the target superpixel \\( x \\) as input, the refinement model \\( E_{r} \\) takes both the current canvas \\( c_{1} \\) (rendered from the coarse stage output) and the target superpixel" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24895" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.242 + ], + "angle": 0, + "content": "\\(x\\) as inputs, and predict new paths \\(S^{\\prime} = \\{s_1^{\\prime}, s_2^{\\prime} \\cdots s_m^{\\prime}\\}\\) to be overlaid onto the canvas to refine details. Specifically, 3 convolution layers followed by ReLU activations are employed to fuse the two input images into a feature map. After getting the fused feature map, \\(E_{r}\\) shares the same structure as the coarse-stage model \\(E_{c}\\), which encodes the fused feature map by a ViT Encoder and maps the encoded features into path parameters by a cross-attention and a self-attention layer. To accelerate the training process, we inherit the weights of the ViT encoder in \\(E_{c}\\) as an initialization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.472, + 0.303 + ], + "angle": 0, + "content": "Local optimal solution with \\(\\mathcal{L}_2\\) loss. The goal of the refinement model is to reconstruct more details of the input superpixel based on the current canvas. A simple \\(\\mathcal{L}_2\\) loss defined as follows is used as the reconstruction loss:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.306, + 0.47, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {2} = \\| x - R ([ E _ {c} (x), E _ {r} (x, c _ {1}) ]) \\| ^ {2} \\cdot \\frac {\\sum_ {p} m a s k (p)}{w h}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.342, + 0.47, + 0.403 + ], + "angle": 0, + "content": "where the predicted path sequences by coarse-stage model \\( E_{c} \\) and refinement model \\( E_{r} \\) are concatenated together to get the final SVG result, and the rendered image of which is expected to resemble the input superpixel \\( x \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.569 + ], + "angle": 0, + "content": "However, the refinement model \\( E_r \\) trained with Eq.(6) alone tends to predict paths that are extremely small in area, or even invisible. In Fig. 4, we use an example to illustrate this phenomenon more clearly: we newly add a path onto the canvas and optimize the path parameters with \\( \\mathcal{L}_2 \\) loss; it can be seen that the new path gradually shrinks and finally disappears in the canvas. A possible reason is that the coarse stage result \\( c_1 \\) is already close to \\( x \\), and a local optimal solution for \\( E_r \\) is to overlap nothing onto \\( c_1 \\), which is better than adding a sub-optimal path and can prevent the \\( \\mathcal{L}_2 \\) distance from increasing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.471, + 0.856 + ], + "angle": 0, + "content": "Coarse paths guided training framework. To avoid the refinement model from falling into poor local optimum, we propose a coarse paths guided framework, which inherits the knowledge from the coarse-stage model to help train the refinement model with an additional constraint on the SVG path parameters. As illustrated in Fig. 2, for an input superpixel \\( x \\), we first use the coarse-stage model to predict a coarse level path sequence \\( S = \\{s_1, s_2 \\dots s_n\\} \\). Then, we randomly choose a value \\( k \\in (1, n - m) \\) and split the predicted path sequence into two subsequences: \\( S_1 = \\{s_1, s_2, \\dots, s_k\\} \\) and \\( S_2 = \\{s_{k + 1}, s_{k + 2}, \\dots, s_n\\} \\). The subsequence \\( S_1 \\) is then rendered into \\( c_1 = R(S_1) \\) and used as the input canvas for the refinement model \\( E_r \\), while the remaining subsequence \\( S_2 \\) can be regarded as a pseudo ground truth for the output path sequence of \\( E_r \\). Specifically, when training \\( E_r \\), in addition to the previous constraints that operate in the pixel space, we add a new constraint on the path parameter space, which minimizes the distance between path sequences \\( S' \\) and \\( S_2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In detail, during training, the refinement model \\(E_{r}\\) takes both the canvas \\(c_{1} = R(S_{1})\\) and the target image \\(x\\) as inputs, and outputs the path sequence \\(S^{\\prime} = \\{s_1^{\\prime},s_2^{\\prime},\\dots ,s_m^{\\prime}\\}\\)." + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.092, + 0.892, + 0.108 + ], + "angle": 0, + "content": "Algorithm 1 Forward pass to efficiently compute \\(\\mathbf{dpw}_{\\gamma}(\\mathbf{S},\\mathbf{S}^{\\prime})\\)" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.112, + 0.815, + 0.126 + ], + "angle": 0, + "content": "Input: \\(S, S'\\), smoothing \\(\\gamma \\geq 0\\), distance function \\(d\\)" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.127, + 0.841, + 0.221 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} 1: p _ {0, j} = 0; p _ {i, 0} = q _ {i, 0} = q _ {0, j} = \\infty , i \\in [ [ n ] ], j \\in [ [ m ] ] \\\\ 2: \\text {f o r} j = 1, \\dots , m \\text {d o} \\\\ 3: \\quad \\text {f o r} i = 1, \\dots , n \\mathbf {d o} \\\\ 4: \\quad p _ {i, j} = d _ {i, j} + \\min ^ {\\gamma} \\left(q _ {i, j - 1}, p _ {i, j - 1}\\right) \\\\ 5: \\quad q _ {i, j} = \\min ^ {\\gamma} \\left(q _ {i - 1, j}, p _ {i - 1, j}\\right) \\\\ 6: \\quad e n d f o r \\\\ 7: \\text {e n d f o r} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.223, + 0.721, + 0.238 + ], + "angle": 0, + "content": "Output: \\((\\min^{\\gamma}(p_{n,m},q_{n,m}),P,Q)\\)" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.253, + 0.868, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.408, + 0.765, + 0.417 + ], + "angle": 0, + "content": "(a) Difference in algorithms" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.419, + 0.846, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.591 + ], + "angle": 0, + "content": "Figure 5. Difference between DTW and our DPW. a) Both the DTW and DPW loss calculate the sum of distances of elements colored yellow. The difference is that one generated path can only match one target path in DPW to avoid averaging several target paths. b) The comparison between the training processes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.891, + 0.628 + ], + "angle": 0, + "content": "We minimize both the distance in pixel space and path parameter space, with the total loss formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.636, + 0.892, + 0.674 + ], + "angle": 0, + "content": "\\[\nE _ {r} ^ {*} = \\underset {E _ {r}} {\\arg \\min } \\mathcal {L} _ {2} + \\lambda_ {D P W} \\mathcal {L} _ {D P W} + \\lambda_ {B o u n d} \\mathcal {L} _ {B o u n d}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.891, + 0.706 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{DPW}\\) (detailed in Sec. 3.3) is the distance between the path sequences \\(S_{2}\\) and \\(S^{\\prime}\\) in the path parameter space." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.714, + 0.719, + 0.731 + ], + "angle": 0, + "content": "3.3. Dynamic Path Warping" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.737, + 0.892, + 0.853 + ], + "angle": 0, + "content": "To calculate the distance between the target path sequence \\( S = \\{s_1, s_2, \\dots, s_n\\} \\) and the predicted path sequence \\( S' = \\{s_1', s_2', \\dots, s_m'\\} \\), Dynamic Time Warping (DTW) [4] is a commonly used metric, which finds an optimal matching match between the two ordered sequences (the yellow points in Fig. 5(a)) to minimize the accumulated distance \\( \\sum_{i}^{n} \\| s_i - s_{match(i)}' \\|^2 \\) where match(i + 1) ≥ match(i)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Note that the monotonicity of the matching function cannot be ignored since the path sequences are well-ordered where latter paths are overlaid on former paths." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24896" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.226 + ], + "angle": 0, + "content": "In our coarse paths guided framework, we expect the generated path sequence to be a subsequence of the target path sequence. However, in DTW, one generated path \\( s_j' \\) can correspond to several target paths \\( s_{i1},\\ldots ,s_{il} \\). Therefore, when trained with DTW, one generated path tends to become the average of several target paths. As shown in Fig. 5(b), when optimizing 3 paths to match the target emoji (with 4 paths), one path becomes the average of the two eyebrow paths, which is not our desired case." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.227, + 0.473, + 0.471 + ], + "angle": 0, + "content": "To address this issue, we propose Dynamic Path Warping (DPW), where each generated path should match one and only one target path, and some target paths can be skipped (to learn a subsequence), as shown in Fig. 5(a), each horizontal line only passes through one matching point (yellow). To compute the DPW, we define \\( p_{i,j} \\) as the minimum accumulated distance when \\( s_i \\) matches \\( s_j' \\), and \\( q_{i,j} \\) as the minimum accumulated distance when \\( s_j' \\) has been matched to one path before \\( s_i \\) (not including \\( s_i \\)). We employ dynamic programming to compute the final DPW loss \\( p_{n,m} \\) as shown in Alg. 1. For each \\( p_{i,j} \\), the distance \\( d_{i,j} \\) between \\( s_i \\) and \\( s_j' \\) is added to the smaller one of \\( q_{i,j-1} \\) and \\( p_{i,j-1} \\). And for each \\( q_{i,j} \\), its value takes the smaller one between \\( q_{i-1,j} \\) and \\( p_{i-1,j} \\) (more explanations are provided in the supplementary material). Moreover, to make Alg. 1 differentiable, we follow SoftDTW [4] to substitute the \\( \\min(\\cdot) \\) operation:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.479, + 0.459, + 0.519 + ], + "angle": 0, + "content": "\\[\n\\min^ {\\gamma} (a _ {0}, a _ {1},..., a _ {n}) = \\left\\{ \\begin{array}{l l} \\min _ {i \\leq n} a _ {i} & \\gamma = 0, \\\\ - \\gamma \\log \\sum_ {i = 1} ^ {n} e ^ {- a _ {i} / \\gamma} & \\gamma > 0. \\end{array} \\right.\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.536, + 0.21, + 0.552 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.561, + 0.265, + 0.577 + ], + "angle": 0, + "content": "4.1. Experiment Setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We use SVG paths composed by cubic Bézier curves as the vector primitive, where each SVG path is closed, composed of 4 cubic Bézier curves connected end-to-end and has a fill color. Each SVG path has 28 parameters (24 for shape, 3 for color, and 1 for visibility). The coarse-stage model is trained to predict 128 paths for each superpixel first. Then, the refinement model is trained to predict 8 paths at one time. We train both the coarse-stage model and refinement model on ImageNet dataset [6]. We set batch size as 64 and learning rate as \\(2.5 \\times 10^{-4}\\). We train the coarse-stage model for 200K iterations with 5K warm up iterations, and train the refinement model for 200K iterations with \\(\\lambda_{DPW}\\) decreasing from \\(1 \\times 10^{-3}\\) to 0 in 10K iterations uniformly. In the following experiments, we implement our model with two versions: 1) SuperSVG-B, that decomposes the image into superpixels and vectorizes them by the coarse-stage model and refinement model, and 2) SuperSVG-F, which finetunes the SVG parameters from SuperSVG-B with \\(\\mathcal{L}_2\\) loss, which takes around 10 seconds for optimization. All experiments are carried out on an NVIDIA GeForce RTX 4090 24GB GPU." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.259 + ], + "angle": 0, + "content": "Evaluation Details. For quantitative evaluation and comparison, we test our model on 1,000 images randomly selected from ImageNet test set, and convert each image into SVGs with 500, 2,000 and 4,000 paths respectively. With the predicted SVG paths, we evaluate the reconstruction accuracy of the output SVG with the following 4 metrics: 1) MSE Distance and 2) PSNR to measure the pixel distance between the input image and the rendering from SVG; 3) LPIPS [35] to evaluate the perceptual distance, and 4) SSIM [31] to measure the structural distance. We further compare with Im2Vec [21] on EMOJIS dataset [2]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.268, + 0.744, + 0.285 + ], + "angle": 0, + "content": "4.2. Image to SVG Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.895, + 0.519 + ], + "angle": 0, + "content": "State-of-the-art methods. The state-of-the-art methods can be classified into 3 categories: 1) Algorithm-based methods: Potrace [24] employs edge tracing to vectorize binary images. To process color images, we follow Color Trace1 to first quantize color images into different layers and then convert each layer to SVG using Potrace. Adobe Illustrator [1] is a widely-used commercial software which converts an image into SVG through image tracing. 2) Deep-learning-based methods: Im2Vec [21] encodes the target image into latent and predicts the vector paths with LSTM (#suppl.); and 3) Optimization-based methods: DiffVG [17] optimizes path parameters with random initialization and LIVE [20] employs layer-wise optimization to ensure a good vectorization structure. We use the official codes of these methods and default settings for comparison." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.895, + 0.746 + ], + "angle": 0, + "content": "Qualitative Comparison on ImageNet. We compare with the state-of-the-art vectorization methods in reconstruction accuracy on ImageNet. Specifically, we conduct the comparison experiments under path numbers 500, 2,000 and \\(4,000^2\\). The qualitative results are shown in Fig. 6. It can be seen that Potrace cannot reconstruct the image well. LIVE loses a lot of details in relatively smooth areas due to its emphasis on regions with substantial color variations. DiffVG and Adobe work better when the path number increases, but they reconstruct fewer details than our SuperSVG. In comparison, our SuperSVG-B reconstructs most of the details with a short inference time. And by optimizing the SVG parameters from SuperSVG-B with only 10 seconds, our SuperSVG-F achieves the best reconstruction accuracy under different SVG path numbers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.895, + 0.852 + ], + "angle": 0, + "content": "Quantitative Comparison on ImageNet. We further conduct quantitative comparison on 1,000 images randomly sampled from ImageNet [6] dataset (50 images for LIVE due to the extremely long optimization time: each input image takes about 6 GPU hours to optimize under 500 SVG paths). The quantitative results are presented in Tab. 1. Our SuperSVG achieves the best image vectorization results." + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.863, + 0.85, + 0.876 + ], + "angle": 0, + "content": "2Since Adobe and Potrace outputs have different number of parameters in each path, we keep their output parameter number the same as ours." + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.863, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24897" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.179, + 0.09, + 0.797, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.51, + 0.858, + 0.524 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison with the state-of-the-art methods in image vectorization with different number of SVG paths." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.528, + 0.47, + 0.554 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison between the state-of-the-arts and Ours. Bold and underline for the best and the second best results." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.556, + 0.47, + 0.835 + ], + "angle": 0, + "content": "
#PathsMethodTime (s) ↓MSE ↓PSNR ↑LPIPS ↓SSIM ↑
500LIVE [20]20,0000.003924.100.44670.7983
DiffVG [17]19.290.006921.420.53190.6671
Adobe [1]0.870.006721.820.55950.6939
Potrace [24]0.980.020817.850.51150.6920
SuperSVG-B (Ours)0.310.004424.800.44520.7687
SuperSVG-F (Ours)10.000.002627.460.35580.8383
2,000LIVE [20]120,0000.002526.980.39940.8431
DiffVG [17]73.650.003625.880.46830.7710
Adobe [1]2.150.003326.230.39610.7229
Potrace [24]3.100.016019.650.43550.6997
SuperSVG-B (Ours)0.710.002427.250.36480.8446
SuperSVG-F (Ours)10.000.001729.120.29310.8828
4,000LIVE [20]300,0000.002426.800.39810.8446
DiffVG [17]140.340.002527.290.38580.8492
Adobe [1]3.120.003525.930.34080.7297
Potrace [24]5.110.011320.680.43800.7374
SuperSVG-B (Ours)1.040.001928.450.31870.8757
SuperSVG-F (Ours)10.000.001429.960.24960.9028
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.231, + 0.864 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Ablation Study on the Superpixel-based Framework. We first validate the effectiveness of the superpixel-based" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.748 + ], + "angle": 0, + "content": "vectorization framework. We train a model that directly predicts the SVG paths for an input image, without superpixel segmentation. Then, we test the model in two ways: 1) predict the SVG paths for the whole input image and 2) uniformly divide the input image into \\(4 \\times 4\\) blocks and vectorize each block separately. We compare our model with these two models with the same number of paths (1,000). The results are shown in Fig. 7 and Tab. 2(a). The model without superpixel segmentation loses many image details. By introducing the block division, the model enriches the details, but the regions near block boundaries are discontinuous (shown in red box), producing unnatural results. In comparison, our superpixel-based SuperSVG-B reconstructs most details and keeps the image structures well." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Ablation Study on the Coarse-stage Model. We then evaluate the effectiveness of the boundary loss \\(\\mathcal{L}_{\\text{Bound}}\\) and the path efficiency loss \\(\\mathcal{L}_{PE}\\) in the coarse-stage model. We train 2 ablated models: 1) the model without \\(\\mathcal{L}_{\\text{Bound}}\\); and 2) the model without \\(\\mathcal{L}_{PE}\\), and compare them with our coarse-stage model under 500 SVG paths. In this comparison, we only compare vectorization using the coarse-stage model, without using the refinement model. The results are shown in Fig. 8 and Tab. 2(b). The model without \\(L_{\\text{Bound}}\\) predicts some paths that cross superpixel boundaries, re" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24898" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.089, + 0.428, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.232, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Figure 7. Ablation study on the superpixel-based image vectorization framework. The ablated models either cannot reconstruct most details or suffer from boundary inconsistency problem." + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.283, + 0.428, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.433, + 0.47, + 0.473 + ], + "angle": 0, + "content": "Figure 8. Ablation study on the coarse-stage model. The ablated models either predict paths crossing superpixel boundaries or reconstruct less details than ours." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.469, + 0.566 + ], + "angle": 0, + "content": "sulting in worse reconstruction. The model without \\(L_{PE}\\) has a poorer performance, which is validated by the metric results. In comparison, our model outperforms the two ablated models, validating the effectiveness of the losses \\(\\mathcal{L}_{\\text{Bound}}\\) and \\(\\mathcal{L}_{PE}\\) in the coarse-stage model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.855 + ], + "angle": 0, + "content": "Ablation Study on the Refinement-stage Model. Finally, we validate the effectiveness of the refinement stage and the DPW loss. We train 3 ablated models: 1) the model without the refinement stage; 2) the model without the DPW loss \\(\\mathcal{L}_{DPW}\\) (i.e., without coarse paths guidance, with pixel-wise loss only); and 3) the model replacing DPW loss \\(\\mathcal{L}_{DPW}\\) with \\(\\mathcal{L}_2\\) loss in path parameter space. The results are shown in Fig. 9 and Tab. 2(c). The model without refinement-stage cannot reconstruct as many details as ours. The ablated model without \\(\\mathcal{L}_{DPW}\\) predicts paths with a very small area or even invisible, as explained in Sec.3.2, thus the results look similar to the coarse-stage results. For the ablated model replacing DPW loss \\(\\mathcal{L}_{DPW}\\) with \\(\\mathcal{L}_2\\) in path parameter space, which enforces one-to-one alignment between two paths' control points, since the constraint is too strict, the loss cannot function well in experiments, and the results look alike the results without \\(\\mathcal{L}_{DPW}\\). In comparison, our model outperforms the ablated models, validating the effectiveness of the refinement and DPW loss." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.089, + 0.851, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.228, + 0.892, + 0.268 + ], + "angle": 0, + "content": "Figure 9. Ablation study on the refinement-stage model. The ablated models either lose details or converge to the poor local optimum described in Sec. 3.2." + }, + { + "type": "table_caption", + "bbox": [ + 0.553, + 0.269, + 0.838, + 0.282 + ], + "angle": 0, + "content": "Table 2. Quantitative results of ablation studies." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.285, + 0.872, + 0.416 + ], + "angle": 0, + "content": "
MethodMSE ↓PSNR ↑LPIPS ↓SSIM ↑
w/o superpixel0.008321.720.50570.6900
4 × 4 blocks0.003829.500.42990.7834
Ours0.003226.040.40750.8111
(a) Ablation on superpixel-based framework.
MethodMSE ↓PSNR ↑LPIPS ↓SSIM ↑
w/o LBound0.053413.380.50410.6210
w/o LPE0.006322.980.48300.7219
Ours (Coarse)0.004524.490.44520.7673
" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.417, + 0.872, + 0.444 + ], + "angle": 0, + "content": "(b) Ablation on coarse-stage model (results are obtained by the coarse-stage model alone, without refinement model)." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.446, + 0.869, + 0.514 + ], + "angle": 0, + "content": "
MethodMSE ↓PSNR ↑LPIPS ↓SSIM ↑
w/o refine0.004125.020.43750.7770
w/o LDPW0.004524.490.44520.7673
L2 replacing LDPW0.004524.490.44520.7673
Ours0.003226.040.40750.8111
" + }, + { + "type": "table_caption", + "bbox": [ + 0.577, + 0.516, + 0.815, + 0.528 + ], + "angle": 0, + "content": "(c) Ablation on refinement-stage model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.535, + 0.619, + 0.55 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.56, + 0.892, + 0.711 + ], + "angle": 0, + "content": "We propose SuperSVG, a novel superpixel-based vectorization model that decomposes a raster image into superpixels and then vectorizes each separately, achieving fast and accurate image vectorization. We propose a two-stage self-teaching framework, where a coarse-stage model reconstructs main structure and a refinement model enriches details, with a novel dynamic path warping loss that guides the refinement model by inheriting knowledge from coarse paths. Extensive experiments demonstrate that SuperSVG achieves the state-of-the-art performance on vectorization." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.659, + 0.741 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported by National Natural Science Foundation of China (62302297, 72192821, 62272447), Shanghai Sailing Program (22YF1420300), Young Elite Scientists Sponsorship Program by CAST (2022QNRC001), Beijing Natural Science Foundation (L222117), the Fundamental Research Funds for the Central Universities (YG2023QNB17), Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), Shanghai Science and Technology Commission (21511101200)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": "3Since the model without refinement does not contain refinement paths, we increase the number of coarse paths to keep the path number consistent." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24899" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Adobe Illustrator. https://www.adobe.com/products/illustrator.html.2,6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.468, + 0.171 + ], + "angle": 0, + "content": "[2] Noto Emoji. https://github.com/googlefonts/ noto-emoji.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.468, + 0.24 + ], + "angle": 0, + "content": "[3] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and Sabine Susstrunk. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 2274-2282, 2012. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.468, + 0.296 + ], + "angle": 0, + "content": "[4] Marco Cuturi and Mathieu Blondel. Soft-DTW: a differentiable loss function for time-series. In International conference on machine learning, pages 894-903. PMLR, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.468, + 0.354 + ], + "angle": 0, + "content": "[5] Wen Dai, Tao Luo, and Jianbing Shen. Automatic image vectorization using superpixels and random walkers. In 2013 6th International Congress on Image and Signal Processing (CISP), 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.468, + 0.422 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.468, + 0.506 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth \\(16 \\times 16\\) words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.468, + 0.548 + ], + "angle": 0, + "content": "[8] Maria Dziuba, Ivan Jarsky, Valeria Efimova, and Andrey Filchenkov. Image vectorization: a review. arXiv preprint arXiv:2306.06441, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.468, + 0.606 + ], + "angle": 0, + "content": "[9] Jun Gao, Chengcheng Tang, Vignesh Ganapathi-Subramanian, Jiahui Huang, Hao Su, and LeonidasJ. Guibas. DeepSpline: Data-driven reconstruction of parametric curves and surfaces. 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.607, + 0.468, + 0.634 + ], + "angle": 0, + "content": "[10] David Ha and Douglas Eck. A neural representation of sketch drawings. Learning, Learning, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.635, + 0.468, + 0.69 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.691, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[12] Teng Hu, Ran Yi, Haokun Zhu, Liang Liu, Jinlong Peng, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. Stroke-based neural painting and stylization with dynamically predicted painting region. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7470-7480, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.775, + 0.468, + 0.816 + ], + "angle": 0, + "content": "[13] Varun Jampani, Deqing Sun, Ming-Yu Liu, Ming-Hsuan Yang, and Jan Kautz. Superpixel sampling networks. arXiv: Computer Vision and Pattern Recognition, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.817, + 0.468, + 0.845 + ], + "angle": 0, + "content": "[14] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[15] Yu-Kun Lai, Shi-Min Hu, and Ralph R. Martin. Automatic and topology-preserving gradient mesh generation for image vectorization. ACM Transactions on Graphics, 28(3):1-8, 2009. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[16] A. Levinshtein, A. Stere, K.N. Kutulakos, D.J. Fleet, S.J. Dickinson, and K. Siddiqi. TurboPixels: Fast superpixels using geometric flows. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 2290-2297, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[17] Tzu-Mao Li, Michal Lukáč, Michael Gharbi, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for editing and learning. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.209, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[18] Zicheng Liao, H. Hoppe, D. Forsyth, and Yizhou Yu. A subdivision-based representation for vector image editing. IEEE Transactions on Visualization and Computer Graphics, 18(11):1858-1867, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.266, + 0.892, + 0.321 + ], + "angle": 0, + "content": "[19] Chen Liu, Jiajun Wu, Pushmeet Kohli, and Yasutaka Furukawa. Raster-to-vector: Revisiting floorplan transformation. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.324, + 0.892, + 0.393 + ], + "angle": 0, + "content": "[20] Xu Ma, Yuqian Zhou, Xingqian Xu, Bin Sun, Valerii Filev, Nikita Orlov, Yun Fu, and Humphrey Shi. Towards layerwise image vectorization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16314-16323, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.395, + 0.892, + 0.464 + ], + "angle": 0, + "content": "[21] Pradyumna Reddy, Michael Gharbi, Michal Lukac, and Niloy J Mitra. Im2Vec: Synthesizing vector graphics without vector supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7342-7351, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.467, + 0.892, + 0.534 + ], + "angle": 0, + "content": "[22] Leo Sampaio Ferraz Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based representation for sketched structure. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.538, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[23] Mike Schuster and Kuldip K Paliwal. Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing, 45(11):2673-2681, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.582, + 0.892, + 0.609 + ], + "angle": 0, + "content": "[24] Peter Selinger. Potrace: a polygon-based tracing algorithm, 2003. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.612, + 0.892, + 0.653 + ], + "angle": 0, + "content": "[25] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.656, + 0.892, + 0.711 + ], + "angle": 0, + "content": "[26] Bohao Tang, Teng Hu, Yuzhen Du, Ran Yu, and Lizhuang Ma. Curved-stroke-based neural painting and stylization through thin plate spline interpolation. Scientia Sinica Informationis, 54(2):301-315, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.714, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[27] Yizhe Tang, Yue Wang, Teng Hu, Xin Tan, and Ran Yi. AttnPainter: A scalable stroke predictor for any scene painting. arXiv preprint, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.758, + 0.892, + 0.798 + ], + "angle": 0, + "content": "[28] Sebastian Thrun and James Diebel. Bayesian image vectorization: the probabilistic inversion of vector image rasterization. 2008. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.801, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[29] Wei-Chih Tu, Ming-Yu Liu, Varun Jampani, Deqing Sun, Shao-Yi Chien, Ming-Hsuan Yang, and Jan Kautz. Learning superpixels with segmentation-aware affinity loss. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 568-576, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[30] Michael Van den Bergh, Xavier Boix, Gemma Roig, Benjamin De Capitani, and Luc Van Gool. SEEDS: Superpixels" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24900" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.135 + ], + "angle": 0, + "content": "extracted via energy-driven sampling. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision (ECCV), pages 13-26. Springer, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[31] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.246 + ], + "angle": 0, + "content": "[32] Guofu Xie, Xin Sun, Xin Tong, and Derek Nowrouzezahrai. Hierarchical diffusion curves for accurate automatic image vectorization. ACM Transactions on Graphics, 33(6):1-11, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.317 + ], + "angle": 0, + "content": "[33] Ming Yang, Hongyang Chao, Chi Zhang, Jun Guo, Lu Yuan, and Jian Sun. Effective clipart image vectorization through direct optimization of bezigons. IEEE Transactions on Visualization and Computer Graphics, page 1063-1075, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.319, + 0.469, + 0.387 + ], + "angle": 0, + "content": "[34] Zipeng Ye, Ran Yi, Minjing Yu, Yong-Jin Liu, and Ying He. Fast computation of content-sensitive superpixels and supervoxels using q-distances. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3770-3779, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.47, + 0.459 + ], + "angle": 0, + "content": "[35] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[36] Shuang Zhao, Frédo Durand, and Changxi Zheng. Inverse diffusion curves using shape optimization. IEEE Transactions on Visualization and Computer Graphics, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.557 + ], + "angle": 0, + "content": "[37] Hailing Zhou, Jianmin Zheng, and Lei Wei. Representing images using curvilinear feature driven subdivision surfaces. IEEE Transactions on Image Processing, 23(8):3268-3280, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.56, + 0.469, + 0.615 + ], + "angle": 0, + "content": "[38] Haokun Zhu, Juang Ian Chong, Teng Hu, Ran Yi, Yu-Kun Lai, and Paul L Rosin. SAMVG: A multi-stage image vectorization model with the segment-anything model. arXiv preprint arXiv:2311.05276, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24901" + } + ] +] \ No newline at end of file diff --git a/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/3fa1e5c3-5890-4208-976c-19acdb43c796_origin.pdf b/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/3fa1e5c3-5890-4208-976c-19acdb43c796_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dd2564a8087645b5b6e4021d70ff5ff5d0fb4cda --- /dev/null +++ b/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/3fa1e5c3-5890-4208-976c-19acdb43c796_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a9f8dcf93f1c0ff2b148606938f2da00faa9bea83fb72745240f5d213fe5d8a +size 5854053 diff --git a/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/full.md b/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..534da74ac4e47065e2a27968181a88db8d3ef93d --- /dev/null +++ b/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/full.md @@ -0,0 +1,285 @@ +# SuperSVG: Superpixel-based Scalable Vector Graphics Synthesis + +Teng Hu $^{1}$ , Ran Yi $^{1*}$ , Baihong Qian $^{1}$ , Jiangning Zhang $^{2}$ , Paul L. Rosin $^{3}$ , Yu-Kun Lai $^{3}$ $^{1}$ Shanghai Jiao Tong University, $^{2}$ Youtu Lab, Tencent, $^{3}$ Cardiff University +{hu-teng, ranyi, cherry_qbh} $@$ sjtu.edu.cn, vtzhang@tencent.com, {RosinPL, LaiY4} $@$ cardiff.ac.uk + +# Abstract + +SVG (Scalable Vector Graphics) is a widely used graphics format that possesses excellent scalability and editability. Image vectorization, which aims to convert raster images to SVGs, is an important yet challenging problem in computer vision and graphics. Existing image vectorization methods either suffer from low reconstruction accuracy for complex images or require long computation time. To address this issue, we propose SuperSVG, a superpixel-based vectorization model that achieves fast and high-precision image vectorization. Specifically, we decompose the input image into superpixels to help the model focus on areas with similar colors and textures. Then, we propose a two-stage self-training framework, where a coarse-stage model is employed to reconstruct the main structure and a refinement-stage model is used for enriching the details. Moreover, we propose a novel dynamic path warping loss to help the refinement-stage model to inherit knowledge from the coarse-stage model. Extensive qualitative and quantitative experiments demonstrate the superior performance of our method in terms of reconstruction accuracy and inference time compared to state-of-the-art approaches. The code is available in https://github.com/sjtuplayer/SuperSVG. + +# 1. Introduction + +Scalable Vector Graphics, commonly known as SVG, is a widely used vector image format that has a wide range of applications and advantages within the domains of web design, graphic design, mobile applications, data visualization, and various other contexts. Compared with raster images that represent content by pixels, SVG describes images by parameterized vectors and benefits from its scalability and editability where it can be resized to any resolution without losing quality and can be easily manipulated by its layer-wise topological information. + +Given the superior capabilities of Scalable Vector Graph + +![](images/8ba3dbbac75d07c23024b79669a896d16b08d24cf6ddc49ef9d48e3f017c5305.jpg) +Figure 1. Overview of our SuperSVG: our model first decomposes the image to be vectorized into superpixels, each containing pixels sharing similar colors and contents. The coarse-stage model predicts the path parameters to reconstruct the main structure, and then the coarse paths guided refinement model enriches the details by learning the knowledge from the coarse-stage model. Compared to the previous methods, our SuperSVG achieves both a high vectorization accuracy and fast computation speed. + +ics (SVG) in image representation and editing, there is much research on the topic of image vectorization, which aims to convert rasterized images into SVG. The existing methods can be categorized into three classes: 1) Traditional algorithm-based methods [15, 18, 24, 28, 33, 37], where conventional algorithms are employed to fit images, but they usually suffer from a lower vectorization quality. 2) Deep-learning-based methods [9, 10, 12, 19, 21, 22], which parameterize raster images using deep neural networks for reconstruction. They are efficient and can handle the vectorization of simple graphics or characters (e.g., icons and emojis), but struggle to vectorize complex images. 3) Optimization-based methods [17, 20, 26, 38], which optimize SVG parameters to fit the target image, yielding relatively superior reconstruction quality. However, these methods entail a substantial amount of time and computational resources, making them impractical for timely processing of large-scale data. In summary, previous image vectorization methods either suffer from low reconstruction quality for complex images, or demand extensive computation time, imposing significant constraints on their practical utility. + +To achieve good vectorization quality with high ef + +ficiency, we propose SuperSVG, a deep-learning-based method that translates images into scalable vector graphics (SVG) in a coarse-to-fine manner. Since neural networks have difficulties in directly vectorizing complex images, we decompose the input image into different parts in the form of superpixels wherein the pixels share similar colors and textures and then vectorize each part. Then, we propose a Two-Stage Self-Teaching Training framework to vectorize the superpixels, where the coarse-stage model is trained to reconstruct the main structure of the image and the refinement-stage model is trained to enrich the image details. We make use of the predicted paths from the coarse-stage model to guide the refinement-stage model in image vectorization. Furthermore, we propose a novel Dynamic Path Warping loss which helps the refinement-stage to inherit the knowledge of the coarse-stage model. With the help of the superpixel-based image decomposition and the two-stage self-teaching framework, our SuperSVG can keep the image structure well and reconstruct more details at high speed. Extensive quantitative and qualitative experiments validate the effectiveness of our model. + +The main contributions of our work are four-fold: + +- We propose SuperSVG, a novel superpixel-based vectorization model that translates the rasterized images into scalable vector graphics (SVG) based on superpixels and vectorizes the superpixels in a coarse-to-fine manner. +- We design a Two-Stage Self-Teaching Training framework, where we employ a coarse-stage model to reconstruct the main structures and a refinement-stage model to enrich the image details based on the coarse-stage output. +- We propose a Coarse Paths Guided Training strategy to guide the refinement-stage model to inherit the knowledge from the coarse-stage model, which greatly improves the performance of the refinement-stage model and avoids converging to suboptimal local minimum. +- We propose Dynamic Path Warping (DPW) loss, which measures the distance between the predicted paths from the refinement-stage model and the pseudo ground truth approximated with coarse paths. By minimizing the DPW loss, the refinement-stage model can distill the knowledge from the coarse-stage model. + +# 2. Related Work + +# 2.1. Image Vectorization + +Image vectorization aims to transform a rasterized image into scalable vector graphics (SVG) composed of parameterized vectors. Different from raster images that may become blurry when zooming in, SVG can be rendered at any resolution without losing quality and is convenient to edit, widely used in web design, graphic design, etc. The existing vectorization methods can be classified into 3 categories: + +Traditional Algorithm-based Image Vectorization Methods can be classified into mesh-based and curve-based ones. The mesh-based methods [15, 18, 28, 37] segment an input image into non-overlapping patches, and infer the color and the boundary location for each region. The curve-based methods [1, 5, 24, 32, 36] employ Bézier curves with different colors defined on either side to create the vector image. Potrace [24] is a representative method of this type that projects the smooth outlines into Bézier paths, and merge the adjacent paths together. However, the vectorization quality of these methods still needs improvements. + +Deep-learning-based Image Vectorization Methods use neural networks to project a raster image into SVG. Im2Vec [21] employs a variational auto-encoder (VAE) [14] to embed the input image and then maps it into path parameters by a Long Short-Term Memory (LSTM) module [23]. Raster2Vec [19] is focused on vectorization of rasterized floor plans using a ResNet [11]. Gao et al. [9] rely on a pre-trained VGG network [25] and a hierarchical Recurrent Neural Network (RNN) to output parametric curves of different sizes. But these methods only focus on simple images and cannot vectorize complex images well. In comparison, our SuperSVG is the first deep-learning-based method that can vectorize images with complex details, thanks to our superpixel decomposition and coarse-path guided refinement that substantially reduce the learning difficulties. + +Optimization-based Image Vectorization Methods. DiffVG [17] proposes a differentiable renderer that renders the SVG parameters into images. Based on this, DiffVG minimizes the distance between the rasterized and vector images by optimizing the SVG parameters using gradient descent. LIVE [20] and SAMVG [38] further introduce a layer-wise optimization framework, which achieves better vectorization quality over the previous methods. However, due to the low optimization efficiency, they suffer from a long optimization time. In contrast, our SuperSVG achieves both a good vectorization quality and high efficiency. + +# 2.2. Superpixel Decomposition + +Superpixel decomposition is usually used for data preprocessing in vision tasks. Existing superpixel decomposition methods can be categorized into methods based on traditional algorithm or deep learning. For the traditional algorithm based methods, diverse strategies have been employed, e.g., energy-driven sampling [30], geometric flows [16] and clustering [3]. Some recent works [13, 29, 34] employ neural networks to enhance the performance in superpixel decomposition, which shows great potential in this task. + +# 3. Method + +Image vectorization aims to translate a rasterized image $I$ into a Scalable Vector Graphic (SVG). An SVG is com + +![](images/f228249a05e0d8180b1b9ff0e963639311473963883245fbda2d937d28655add.jpg) + +![](images/38f6b6096415ea91de2579883a6ff84fc00d4e87ae10f9f385c01b8edda6b5e2.jpg) +Figure 2. Main framework of our SuperSVG: we decompose the target image into superpixels and vectorize each superpixel separately. We employ an attention-based coarse-stage model to predict SVG paths that reconstruct the main structure of the superpixel. Then, a refinement-stage model guided by the coarse paths is designed to predict more SVG paths to refine details based on the coarse image. Finally, by combining all the predicted SVGs for each superpixel, we obtain an output SVG with good structure and fine details. + +posed of many vector primitives, which can be SVG paths, ellipses, circles, or rectangles, etc. Following previous works [17, 20, 21], we employ the SVG paths as the shape primitive, where each SVG path defines a region constructed by multiple cubic Bezier curves connected end-to-end with certain color. With the parameters of these SVG paths, the rasterized image can be rendered in any resolution. To obtain the path parameters, some previous methods [17, 20] optimize the path parameters directly to minimize the distance between the input image and the rendered image, which achieves good reconstruction quality but requires long optimization time. To speed up the vectorization process, some deep-learning-based methods [21] employ a deep-learning model to predict the SVG path parameters, but struggle to vectorize complex images. + +To achieve good vectorization quality with high efficiency, we propose SuperSVG, a deep-learning-based image vectorization method that translates images into SVG path parameters automatically. To improve the model ability to vectorize complex images, we segment the input image into different parts, within which the pixels share similar colors and textures, and then vectorize each part separately, where superpixels are used for image segmentation as they tend to maintain compactness, uniformity and regularity, particularly suitable for our task. For each superpixel $x \in \mathcal{X}$ where $\mathcal{X}$ is the set of all superpixels, our model converts it into a sequence of path parameters, where each path is composed of several cubic Bezier curves and has a fill color, with a total of $N_{p}$ parameters. With the predicted path parameters for each superpixel, we employ the differ + +entiable renderer $R(\cdot)$ from DiffVG [17] to get the rendered image $\hat{I}$ in pixel space, which is expected to be close to the input image $I$ . + +We propose a two-stage self-teaching framework, composed of a coarse-stage model $E_{c}$ to reconstruct the main structure and a refinement model $E_{r}$ to enrich the details, where the predicted paths from coarse-stage model are used to guide the refinement model in vectorization. $E_{c}$ takes the superpixel $x$ as input and outputs $n$ paths $S = \{s_1,s_2,\dots s_n\}$ to reconstruct main structure; while $E_{r}$ takes both the rendered image $R(S)$ and target superpixel $x$ as inputs, and outputs $m$ paths $S^{\prime} = \{s_{1}^{\prime},s_{2}^{\prime}\dots s_{m}^{\prime}\}$ to refine details. Combining all the predicted $S$ and $S^{\prime}$ for each superpixel produces the final SVG result. + +# 3.1. Superpixel-based Coarse Reconstruction + +Superpixel decomposition. Considering the optimization-based methods suffer from a long optimization time, our SuperSVG builds upon neural networks to efficiently predict SVG paths. However, as neural networks have difficulties in directly vectorizing complex images [8] we therefore simplify the task to vectorizing a certain part of the image containing homogeneous colors and textures. Since superpixel algorithms provide a good tool to decompose images based on local pixel color and also ensure alignment of the regions with the image boundaries, we segment the input image into superpixels, and our model reconstructs each superpixel with scalable vectors. Superpixels also tend to be more regular, making them easier for vectorization. Specifically, we utilize SLIC [3] to decompose the input image + +![](images/8807daf335b8d788c8ebb22f45ddb52a6652ad9434a59c59f88c3849cd7ba8ee.jpg) +Figure 3. Illustration of our boundary loss $\mathcal{L}_{\text{Bound}}$ , which computes the area of the SVG paths that are outside the superpixel mask, and guides the paths to be inside the superpixel. + +into superpixels. We set the compactness parameter as 30 for SLIC to make superpixels more regular. + +Coarse-stage model. For a superpixel $x$ with mask mask (indicating those pixels within $x$ with 1 and 0 otherwise), we first design the coarse-stage model $E_{c}$ to vectorize the main structure by predicting the SVG path sequence $E_{c}(x) = S = \{s_{1}, s_{2}, \dots, s_{n}\}$ . Inspired by AttnPainter [27], $E_{c}$ is composed of a Vision Transformer (ViT) encoder [7] and a cross-attention module followed by a self-attention module, which is shown in Fig. 2. + +Specifically, the ViT encoder first encodes the input superpixel $x$ into image feature $T_{f}$ . To control the number of output paths $(n)$ and the parameter number in each path $(N_{p})$ , we employ a cross-attention module to calculate the correlation between the image feature $T_{f}$ and $n N_{p}$ -dimensional learnable path queries $T_{l}$ , and output an intermediate feature $T_{f}^{\prime}$ with the shape of path parameters $(n \times N_{p})$ , which is formulated as: + +$$ +T _ {f} ^ {\prime} = \operatorname {S o f t m a x} \left(\frac {\left(W _ {Q} T _ {l}\right) \left(W _ {K} T _ {f}\right) ^ {T}}{\sqrt {d}}\right) \left(W _ {V} T _ {f}\right), \tag {1} +$$ + +where $W_{Q}, W_{K}$ and $W_{V}$ are learnable query, key and value matrices. + +Then, a self-attention module is employed to process the image feature $T_f'$ and project it into the parameter space with $n$ paths, where each path contains $N_p$ parameters. + +Training objectives. 1) Normalized Reconstruction Loss: We employ the differentiable renderer $R(\cdot)$ in DiffVG [17] to render a raster image $\hat{x} = R(S)$ from the predicted path parameters $S$ . Then, we train the coarse-stage model $E_{c}$ by minimizing the normalized $\mathcal{L}_2$ distance between the rendering $\hat{x} = R(S)$ and the target image: + +$$ +\mathcal {L} _ {2} = \| \hat {x} - x \| ^ {2} \cdot \frac {\sum_ {p} \operatorname {m a s k} (p)}{w h}, \tag {2} +$$ + +where $w$ and $h$ are the width and height of the superpixel image $x$ and the superpixel mask mask, and $\text{mask}(p)$ indicates the mask value (1 or 0) for pixel $p$ . + +2) Boundary Loss: To avoid the SVG path from crossing the superpixel boundary, we propose boundary loss to guide the paths to be inside the superpixel. We set the color of all + +![](images/28ca982cacb49dad098847c5be0b79419cfec7462967d7ca8c273652d8f001cb.jpg) +Figure 4. Problem of training the refinement model with $\mathcal{L}_2$ loss alone: optimizing a newly-added path on the canvas by $\mathcal{L}_2$ gradually pulls it to disappear (as a suboptimal local minimum). With our proposed coarse paths guided training and DPW loss, the added path is successfully optimized to resemble the target. + +predicted SVG paths to 1 (white) to get a new path sequence $S_{binary}$ . Then, we compute the boundary loss by: + +$$ +\mathcal {L} _ {\text {B o u n d}} = \underset {p \sim \operatorname {m a s k}} {\mathbb {E}} (R \left(S _ {\text {b i n a r y}}\right) \cdot (1 - \operatorname {m a s k})) \tag {3} +$$ + +Since $(1 - \text{mask})$ is 1 outside the superpixel mask and 0 inside the mask, the loss term calculates the area of the paths that are outside the superpixel. When some SVG paths cross the superpixel boundary, the path area outside is penalized; while when all SVG paths are inside the superpixel, $\mathcal{L}_{\text{Bound}}$ reaches 0 (Fig. 3). + +3) Path Efficiency Loss: To enable our model to reconstruct the maximum amount of information with the fewest paths, we propose the path efficiency loss $\mathcal{L}_{PE}$ . Specifically, for each path $i$ , an additional opacity parameter $\beta_{i}$ is predicted. We treat the path as visible if $\beta_{i} \geq 0.5$ , and the loss $L_{PE}$ penalizes the case with more visible paths, i.e., to encourage reconstructing the image with as few paths as possible, calculated as: + +$$ +\mathcal {L} _ {P E} = \sum_ {i = 1} ^ {n} \operatorname {S i g n} \left(\beta_ {i} - 0. 5\right), \tag {4} +$$ + +where $\operatorname{Sign}(\cdot)$ is the sign function. Since $\operatorname{Sign}(\beta_i)$ is not differentiable, we approximate $\frac{\partial \operatorname{Sign}(\beta_i)}{\partial \beta_i}$ by $\operatorname{Sig}(\beta_i)(1 - \operatorname{Sig}(\beta_i))$ , where $\operatorname{Sig}(\cdot)$ is the sigmoid function. + +The final training objective is formulated as: + +$$ +E _ {c} ^ {*} = \underset {E _ {c}} {\arg \min } \mathcal {L} _ {2} + \lambda_ {\text {B o u n d}} \mathcal {L} _ {\text {B o u n d}} + \lambda_ {P E} \mathcal {L} _ {P E}. \tag {5} +$$ + +# 3.2. Coarse Paths Guided Refinement Stage + +In the coarse reconstruction stage, our coarse-stage model $E_{c}$ can output an SVG that captures and reconstructs the main structure of the input superpixel $x$ . The rendered image from the SVG (denoted as $c_{1}$ ) resembles $x$ in general, but lacks some image details, especially when the superpixel is complex. To enrich image details, we employ a refinement model $E_{r}$ to predict more SVG paths to add more details based on the current canvas $c_{1}$ . + +Model framework. Different from the coarse-stage model $E_{c}$ that only takes the target superpixel $x$ as input, the refinement model $E_{r}$ takes both the current canvas $c_{1}$ (rendered from the coarse stage output) and the target superpixel + +$x$ as inputs, and predict new paths $S^{\prime} = \{s_1^{\prime}, s_2^{\prime} \cdots s_m^{\prime}\}$ to be overlaid onto the canvas to refine details. Specifically, 3 convolution layers followed by ReLU activations are employed to fuse the two input images into a feature map. After getting the fused feature map, $E_{r}$ shares the same structure as the coarse-stage model $E_{c}$ , which encodes the fused feature map by a ViT Encoder and maps the encoded features into path parameters by a cross-attention and a self-attention layer. To accelerate the training process, we inherit the weights of the ViT encoder in $E_{c}$ as an initialization. + +Local optimal solution with $\mathcal{L}_2$ loss. The goal of the refinement model is to reconstruct more details of the input superpixel based on the current canvas. A simple $\mathcal{L}_2$ loss defined as follows is used as the reconstruction loss: + +$$ +\mathcal {L} _ {2} = \| x - R ([ E _ {c} (x), E _ {r} (x, c _ {1}) ]) \| ^ {2} \cdot \frac {\sum_ {p} m a s k (p)}{w h}, \tag {6} +$$ + +where the predicted path sequences by coarse-stage model $E_{c}$ and refinement model $E_{r}$ are concatenated together to get the final SVG result, and the rendered image of which is expected to resemble the input superpixel $x$ . + +However, the refinement model $E_r$ trained with Eq.(6) alone tends to predict paths that are extremely small in area, or even invisible. In Fig. 4, we use an example to illustrate this phenomenon more clearly: we newly add a path onto the canvas and optimize the path parameters with $\mathcal{L}_2$ loss; it can be seen that the new path gradually shrinks and finally disappears in the canvas. A possible reason is that the coarse stage result $c_1$ is already close to $x$ , and a local optimal solution for $E_r$ is to overlap nothing onto $c_1$ , which is better than adding a sub-optimal path and can prevent the $\mathcal{L}_2$ distance from increasing. + +Coarse paths guided training framework. To avoid the refinement model from falling into poor local optimum, we propose a coarse paths guided framework, which inherits the knowledge from the coarse-stage model to help train the refinement model with an additional constraint on the SVG path parameters. As illustrated in Fig. 2, for an input superpixel $x$ , we first use the coarse-stage model to predict a coarse level path sequence $S = \{s_1, s_2 \dots s_n\}$ . Then, we randomly choose a value $k \in (1, n - m)$ and split the predicted path sequence into two subsequences: $S_1 = \{s_1, s_2, \dots, s_k\}$ and $S_2 = \{s_{k + 1}, s_{k + 2}, \dots, s_n\}$ . The subsequence $S_1$ is then rendered into $c_1 = R(S_1)$ and used as the input canvas for the refinement model $E_r$ , while the remaining subsequence $S_2$ can be regarded as a pseudo ground truth for the output path sequence of $E_r$ . Specifically, when training $E_r$ , in addition to the previous constraints that operate in the pixel space, we add a new constraint on the path parameter space, which minimizes the distance between path sequences $S'$ and $S_2$ . + +In detail, during training, the refinement model $E_{r}$ takes both the canvas $c_{1} = R(S_{1})$ and the target image $x$ as inputs, and outputs the path sequence $S^{\prime} = \{s_1^{\prime},s_2^{\prime},\dots ,s_m^{\prime}\}$ . + +Algorithm 1 Forward pass to efficiently compute $\mathbf{dpw}_{\gamma}(\mathbf{S},\mathbf{S}^{\prime})$ + +Input: $S, S'$ , smoothing $\gamma \geq 0$ , distance function $d$ + +$$ +\begin{array}{l} 1: p _ {0, j} = 0; p _ {i, 0} = q _ {i, 0} = q _ {0, j} = \infty , i \in [ [ n ] ], j \in [ [ m ] ] \\ 2: \text {f o r} j = 1, \dots , m \text {d o} \\ 3: \quad \text {f o r} i = 1, \dots , n \mathbf {d o} \\ 4: \quad p _ {i, j} = d _ {i, j} + \min ^ {\gamma} \left(q _ {i, j - 1}, p _ {i, j - 1}\right) \\ 5: \quad q _ {i, j} = \min ^ {\gamma} \left(q _ {i - 1, j}, p _ {i - 1, j}\right) \\ 6: \quad e n d f o r \\ 7: \text {e n d f o r} \\ \end{array} +$$ + +Output: $(\min^{\gamma}(p_{n,m},q_{n,m}),P,Q)$ + +![](images/8ec5c60eed98ffe5f022368c197173c389f6a1d4b15c5d5a37d5fd21d67eea2e.jpg) +(a) Difference in algorithms + +![](images/9991f250cb023d47a691ae2f8399607781639e282d071ee56d6d8363b063bfcf.jpg) +Figure 5. Difference between DTW and our DPW. a) Both the DTW and DPW loss calculate the sum of distances of elements colored yellow. The difference is that one generated path can only match one target path in DPW to avoid averaging several target paths. b) The comparison between the training processes. + +We minimize both the distance in pixel space and path parameter space, with the total loss formulated as: + +$$ +E _ {r} ^ {*} = \underset {E _ {r}} {\arg \min } \mathcal {L} _ {2} + \lambda_ {D P W} \mathcal {L} _ {D P W} + \lambda_ {B o u n d} \mathcal {L} _ {B o u n d}, \tag {7} +$$ + +where $\mathcal{L}_{DPW}$ (detailed in Sec. 3.3) is the distance between the path sequences $S_{2}$ and $S^{\prime}$ in the path parameter space. + +# 3.3. Dynamic Path Warping + +To calculate the distance between the target path sequence $S = \{s_1, s_2, \dots, s_n\}$ and the predicted path sequence $S' = \{s_1', s_2', \dots, s_m'\}$ , Dynamic Time Warping (DTW) [4] is a commonly used metric, which finds an optimal matching match between the two ordered sequences (the yellow points in Fig. 5(a)) to minimize the accumulated distance $\sum_{i}^{n} \| s_i - s_{match(i)}' \|^2$ where match(i + 1) ≥ match(i). + +Note that the monotonicity of the matching function cannot be ignored since the path sequences are well-ordered where latter paths are overlaid on former paths. + +In our coarse paths guided framework, we expect the generated path sequence to be a subsequence of the target path sequence. However, in DTW, one generated path $s_j'$ can correspond to several target paths $s_{i1},\ldots ,s_{il}$ . Therefore, when trained with DTW, one generated path tends to become the average of several target paths. As shown in Fig. 5(b), when optimizing 3 paths to match the target emoji (with 4 paths), one path becomes the average of the two eyebrow paths, which is not our desired case. + +To address this issue, we propose Dynamic Path Warping (DPW), where each generated path should match one and only one target path, and some target paths can be skipped (to learn a subsequence), as shown in Fig. 5(a), each horizontal line only passes through one matching point (yellow). To compute the DPW, we define $p_{i,j}$ as the minimum accumulated distance when $s_i$ matches $s_j'$ , and $q_{i,j}$ as the minimum accumulated distance when $s_j'$ has been matched to one path before $s_i$ (not including $s_i$ ). We employ dynamic programming to compute the final DPW loss $p_{n,m}$ as shown in Alg. 1. For each $p_{i,j}$ , the distance $d_{i,j}$ between $s_i$ and $s_j'$ is added to the smaller one of $q_{i,j-1}$ and $p_{i,j-1}$ . And for each $q_{i,j}$ , its value takes the smaller one between $q_{i-1,j}$ and $p_{i-1,j}$ (more explanations are provided in the supplementary material). Moreover, to make Alg. 1 differentiable, we follow SoftDTW [4] to substitute the $\min(\cdot)$ operation: + +$$ +\min^ {\gamma} (a _ {0}, a _ {1},..., a _ {n}) = \left\{ \begin{array}{l l} \min _ {i \leq n} a _ {i} & \gamma = 0, \\ - \gamma \log \sum_ {i = 1} ^ {n} e ^ {- a _ {i} / \gamma} & \gamma > 0. \end{array} \right. +$$ + +# 4. Experiments + +# 4.1. Experiment Setting + +Implementation Details. We use SVG paths composed by cubic Bézier curves as the vector primitive, where each SVG path is closed, composed of 4 cubic Bézier curves connected end-to-end and has a fill color. Each SVG path has 28 parameters (24 for shape, 3 for color, and 1 for visibility). The coarse-stage model is trained to predict 128 paths for each superpixel first. Then, the refinement model is trained to predict 8 paths at one time. We train both the coarse-stage model and refinement model on ImageNet dataset [6]. We set batch size as 64 and learning rate as $2.5 \times 10^{-4}$ . We train the coarse-stage model for 200K iterations with 5K warm up iterations, and train the refinement model for 200K iterations with $\lambda_{DPW}$ decreasing from $1 \times 10^{-3}$ to 0 in 10K iterations uniformly. In the following experiments, we implement our model with two versions: 1) SuperSVG-B, that decomposes the image into superpixels and vectorizes them by the coarse-stage model and refinement model, and 2) SuperSVG-F, which finetunes the SVG parameters from SuperSVG-B with $\mathcal{L}_2$ loss, which takes around 10 seconds for optimization. All experiments are carried out on an NVIDIA GeForce RTX 4090 24GB GPU. + +Evaluation Details. For quantitative evaluation and comparison, we test our model on 1,000 images randomly selected from ImageNet test set, and convert each image into SVGs with 500, 2,000 and 4,000 paths respectively. With the predicted SVG paths, we evaluate the reconstruction accuracy of the output SVG with the following 4 metrics: 1) MSE Distance and 2) PSNR to measure the pixel distance between the input image and the rendering from SVG; 3) LPIPS [35] to evaluate the perceptual distance, and 4) SSIM [31] to measure the structural distance. We further compare with Im2Vec [21] on EMOJIS dataset [2]. + +# 4.2. Image to SVG Comparison + +State-of-the-art methods. The state-of-the-art methods can be classified into 3 categories: 1) Algorithm-based methods: Potrace [24] employs edge tracing to vectorize binary images. To process color images, we follow Color Trace1 to first quantize color images into different layers and then convert each layer to SVG using Potrace. Adobe Illustrator [1] is a widely-used commercial software which converts an image into SVG through image tracing. 2) Deep-learning-based methods: Im2Vec [21] encodes the target image into latent and predicts the vector paths with LSTM (#suppl.); and 3) Optimization-based methods: DiffVG [17] optimizes path parameters with random initialization and LIVE [20] employs layer-wise optimization to ensure a good vectorization structure. We use the official codes of these methods and default settings for comparison. + +Qualitative Comparison on ImageNet. We compare with the state-of-the-art vectorization methods in reconstruction accuracy on ImageNet. Specifically, we conduct the comparison experiments under path numbers 500, 2,000 and $4,000^2$ . The qualitative results are shown in Fig. 6. It can be seen that Potrace cannot reconstruct the image well. LIVE loses a lot of details in relatively smooth areas due to its emphasis on regions with substantial color variations. DiffVG and Adobe work better when the path number increases, but they reconstruct fewer details than our SuperSVG. In comparison, our SuperSVG-B reconstructs most of the details with a short inference time. And by optimizing the SVG parameters from SuperSVG-B with only 10 seconds, our SuperSVG-F achieves the best reconstruction accuracy under different SVG path numbers. + +Quantitative Comparison on ImageNet. We further conduct quantitative comparison on 1,000 images randomly sampled from ImageNet [6] dataset (50 images for LIVE due to the extremely long optimization time: each input image takes about 6 GPU hours to optimize under 500 SVG paths). The quantitative results are presented in Tab. 1. Our SuperSVG achieves the best image vectorization results. + +![](images/18fc9e563be8472eb8c120edfa49fd705936706691eaf5b89ec0c76d06f66dc6.jpg) +Figure 6. Qualitative comparison with the state-of-the-art methods in image vectorization with different number of SVG paths. + +Table 1. Quantitative comparison between the state-of-the-arts and Ours. Bold and underline for the best and the second best results. + +
#PathsMethodTime (s) ↓MSE ↓PSNR ↑LPIPS ↓SSIM ↑
500LIVE [20]20,0000.003924.100.44670.7983
DiffVG [17]19.290.006921.420.53190.6671
Adobe [1]0.870.006721.820.55950.6939
Potrace [24]0.980.020817.850.51150.6920
SuperSVG-B (Ours)0.310.004424.800.44520.7687
SuperSVG-F (Ours)10.000.002627.460.35580.8383
2,000LIVE [20]120,0000.002526.980.39940.8431
DiffVG [17]73.650.003625.880.46830.7710
Adobe [1]2.150.003326.230.39610.7229
Potrace [24]3.100.016019.650.43550.6997
SuperSVG-B (Ours)0.710.002427.250.36480.8446
SuperSVG-F (Ours)10.000.001729.120.29310.8828
4,000LIVE [20]300,0000.002426.800.39810.8446
DiffVG [17]140.340.002527.290.38580.8492
Adobe [1]3.120.003525.930.34080.7297
Potrace [24]5.110.011320.680.43800.7374
SuperSVG-B (Ours)1.040.001928.450.31870.8757
SuperSVG-F (Ours)10.000.001429.960.24960.9028
+ +# 4.3. Ablation Study + +Ablation Study on the Superpixel-based Framework. We first validate the effectiveness of the superpixel-based + +vectorization framework. We train a model that directly predicts the SVG paths for an input image, without superpixel segmentation. Then, we test the model in two ways: 1) predict the SVG paths for the whole input image and 2) uniformly divide the input image into $4 \times 4$ blocks and vectorize each block separately. We compare our model with these two models with the same number of paths (1,000). The results are shown in Fig. 7 and Tab. 2(a). The model without superpixel segmentation loses many image details. By introducing the block division, the model enriches the details, but the regions near block boundaries are discontinuous (shown in red box), producing unnatural results. In comparison, our superpixel-based SuperSVG-B reconstructs most details and keeps the image structures well. + +Ablation Study on the Coarse-stage Model. We then evaluate the effectiveness of the boundary loss $\mathcal{L}_{\text{Bound}}$ and the path efficiency loss $\mathcal{L}_{PE}$ in the coarse-stage model. We train 2 ablated models: 1) the model without $\mathcal{L}_{\text{Bound}}$ ; and 2) the model without $\mathcal{L}_{PE}$ , and compare them with our coarse-stage model under 500 SVG paths. In this comparison, we only compare vectorization using the coarse-stage model, without using the refinement model. The results are shown in Fig. 8 and Tab. 2(b). The model without $L_{\text{Bound}}$ predicts some paths that cross superpixel boundaries, re + +![](images/eb0bd8a4ac218addf3223fbdf0c775c6426e2b4643261beb199062825959c733.jpg) +Figure 7. Ablation study on the superpixel-based image vectorization framework. The ablated models either cannot reconstruct most details or suffer from boundary inconsistency problem. + +![](images/1dc0cd3060dd4f4aa0b2993cc730c141b811fcbba9e4de89350b92de5f4b7449.jpg) +Figure 8. Ablation study on the coarse-stage model. The ablated models either predict paths crossing superpixel boundaries or reconstruct less details than ours. + +sulting in worse reconstruction. The model without $L_{PE}$ has a poorer performance, which is validated by the metric results. In comparison, our model outperforms the two ablated models, validating the effectiveness of the losses $\mathcal{L}_{\text{Bound}}$ and $\mathcal{L}_{PE}$ in the coarse-stage model. + +Ablation Study on the Refinement-stage Model. Finally, we validate the effectiveness of the refinement stage and the DPW loss. We train 3 ablated models: 1) the model without the refinement stage; 2) the model without the DPW loss $\mathcal{L}_{DPW}$ (i.e., without coarse paths guidance, with pixel-wise loss only); and 3) the model replacing DPW loss $\mathcal{L}_{DPW}$ with $\mathcal{L}_2$ loss in path parameter space. The results are shown in Fig. 9 and Tab. 2(c). The model without refinement-stage cannot reconstruct as many details as ours. The ablated model without $\mathcal{L}_{DPW}$ predicts paths with a very small area or even invisible, as explained in Sec.3.2, thus the results look similar to the coarse-stage results. For the ablated model replacing DPW loss $\mathcal{L}_{DPW}$ with $\mathcal{L}_2$ in path parameter space, which enforces one-to-one alignment between two paths' control points, since the constraint is too strict, the loss cannot function well in experiments, and the results look alike the results without $\mathcal{L}_{DPW}$ . In comparison, our model outperforms the ablated models, validating the effectiveness of the refinement and DPW loss. + +![](images/402d531a0b8f73adc39f9178bae87fd012bdf2623cb514abf1919e4ec0b2eba7.jpg) +Figure 9. Ablation study on the refinement-stage model. The ablated models either lose details or converge to the poor local optimum described in Sec. 3.2. + +Table 2. Quantitative results of ablation studies. + +
MethodMSE ↓PSNR ↑LPIPS ↓SSIM ↑
w/o superpixel0.008321.720.50570.6900
4 × 4 blocks0.003829.500.42990.7834
Ours0.003226.040.40750.8111
(a) Ablation on superpixel-based framework.
MethodMSE ↓PSNR ↑LPIPS ↓SSIM ↑
w/o LBound0.053413.380.50410.6210
w/o LPE0.006322.980.48300.7219
Ours (Coarse)0.004524.490.44520.7673
+ +(b) Ablation on coarse-stage model (results are obtained by the coarse-stage model alone, without refinement model). + +
MethodMSE ↓PSNR ↑LPIPS ↓SSIM ↑
w/o refine0.004125.020.43750.7770
w/o LDPW0.004524.490.44520.7673
L2 replacing LDPW0.004524.490.44520.7673
Ours0.003226.040.40750.8111
+ +(c) Ablation on refinement-stage model. + +# 5. Conclusion + +We propose SuperSVG, a novel superpixel-based vectorization model that decomposes a raster image into superpixels and then vectorizes each separately, achieving fast and accurate image vectorization. We propose a two-stage self-teaching framework, where a coarse-stage model reconstructs main structure and a refinement model enriches details, with a novel dynamic path warping loss that guides the refinement model by inheriting knowledge from coarse paths. Extensive experiments demonstrate that SuperSVG achieves the state-of-the-art performance on vectorization. + +# Acknowledgments + +This work was supported by National Natural Science Foundation of China (62302297, 72192821, 62272447), Shanghai Sailing Program (22YF1420300), Young Elite Scientists Sponsorship Program by CAST (2022QNRC001), Beijing Natural Science Foundation (L222117), the Fundamental Research Funds for the Central Universities (YG2023QNB17), Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), Shanghai Science and Technology Commission (21511101200). + +# References + +[1] Adobe Illustrator. https://www.adobe.com/products/illustrator.html.2,6,7 +[2] Noto Emoji. https://github.com/googlefonts/ noto-emoji.6 +[3] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and Sabine Susstrunk. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 2274-2282, 2012. 2, 3 +[4] Marco Cuturi and Mathieu Blondel. Soft-DTW: a differentiable loss function for time-series. In International conference on machine learning, pages 894-903. PMLR, 2017. 5, 6 +[5] Wen Dai, Tao Luo, and Jianbing Shen. Automatic image vectorization using superpixels and random walkers. In 2013 6th International Congress on Image and Signal Processing (CISP), 2013. 2 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 6 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth $16 \times 16$ words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4 +[8] Maria Dziuba, Ivan Jarsky, Valeria Efimova, and Andrey Filchenkov. Image vectorization: a review. arXiv preprint arXiv:2306.06441, 2023. 3 +[9] Jun Gao, Chengcheng Tang, Vignesh Ganapathi-Subramanian, Jiahui Huang, Hao Su, and LeonidasJ. Guibas. DeepSpline: Data-driven reconstruction of parametric curves and surfaces. 2019. 1, 2 +[10] David Ha and Douglas Eck. A neural representation of sketch drawings. Learning, Learning, 2017. 1 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2 +[12] Teng Hu, Ran Yi, Haokun Zhu, Liang Liu, Jinlong Peng, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. Stroke-based neural painting and stylization with dynamically predicted painting region. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7470-7480, 2023. 1 +[13] Varun Jampani, Deqing Sun, Ming-Yu Liu, Ming-Hsuan Yang, and Jan Kautz. Superpixel sampling networks. arXiv: Computer Vision and Pattern Recognition, 2018. 2 +[14] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2 +[15] Yu-Kun Lai, Shi-Min Hu, and Ralph R. Martin. Automatic and topology-preserving gradient mesh generation for image vectorization. ACM Transactions on Graphics, 28(3):1-8, 2009. 1, 2 + +[16] A. Levinshtein, A. Stere, K.N. Kutulakos, D.J. Fleet, S.J. Dickinson, and K. Siddiqi. TurboPixels: Fast superpixels using geometric flows. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 2290-2297, 2009. 2 +[17] Tzu-Mao Li, Michal Lukáč, Michael Gharbi, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for editing and learning. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 1, 2, 3, 4, 6, 7 +[18] Zicheng Liao, H. Hoppe, D. Forsyth, and Yizhou Yu. A subdivision-based representation for vector image editing. IEEE Transactions on Visualization and Computer Graphics, 18(11):1858-1867, 2012. 1, 2 +[19] Chen Liu, Jiajun Wu, Pushmeet Kohli, and Yasutaka Furukawa. Raster-to-vector: Revisiting floorplan transformation. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 1, 2 +[20] Xu Ma, Yuqian Zhou, Xingqian Xu, Bin Sun, Valerii Filev, Nikita Orlov, Yun Fu, and Humphrey Shi. Towards layerwise image vectorization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16314-16323, 2022. 1, 2, 3, 6, 7 +[21] Pradyumna Reddy, Michael Gharbi, Michal Lukac, and Niloy J Mitra. Im2Vec: Synthesizing vector graphics without vector supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7342-7351, 2021. 1, 2, 3, 6 +[22] Leo Sampaio Ferraz Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based representation for sketched structure. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1 +[23] Mike Schuster and Kuldip K Paliwal. Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing, 45(11):2673-2681, 1997. 2 +[24] Peter Selinger. Potrace: a polygon-based tracing algorithm, 2003. 1, 2, 6, 7 +[25] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2 +[26] Bohao Tang, Teng Hu, Yuzhen Du, Ran Yu, and Lizhuang Ma. Curved-stroke-based neural painting and stylization through thin plate spline interpolation. Scientia Sinica Informationis, 54(2):301-315, 2024. 1 +[27] Yizhe Tang, Yue Wang, Teng Hu, Xin Tan, and Ran Yi. AttnPainter: A scalable stroke predictor for any scene painting. arXiv preprint, 2024. 4 +[28] Sebastian Thrun and James Diebel. Bayesian image vectorization: the probabilistic inversion of vector image rasterization. 2008. 1, 2 +[29] Wei-Chih Tu, Ming-Yu Liu, Varun Jampani, Deqing Sun, Shao-Yi Chien, Ming-Hsuan Yang, and Jan Kautz. Learning superpixels with segmentation-aware affinity loss. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 568-576, 2018. 2 +[30] Michael Van den Bergh, Xavier Boix, Gemma Roig, Benjamin De Capitani, and Luc Van Gool. SEEDS: Superpixels + +extracted via energy-driven sampling. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision (ECCV), pages 13-26. Springer, 2012. 2 +[31] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6 +[32] Guofu Xie, Xin Sun, Xin Tong, and Derek Nowrouzezahrai. Hierarchical diffusion curves for accurate automatic image vectorization. ACM Transactions on Graphics, 33(6):1-11, 2014. 2 +[33] Ming Yang, Hongyang Chao, Chi Zhang, Jun Guo, Lu Yuan, and Jian Sun. Effective clipart image vectorization through direct optimization of bezigons. IEEE Transactions on Visualization and Computer Graphics, page 1063-1075, 2016. 1 +[34] Zipeng Ye, Ran Yi, Minjing Yu, Yong-Jin Liu, and Ying He. Fast computation of content-sensitive superpixels and supervoxels using q-distances. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3770-3779, 2019. 2 +[35] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 6 +[36] Shuang Zhao, Frédo Durand, and Changxi Zheng. Inverse diffusion curves using shape optimization. IEEE Transactions on Visualization and Computer Graphics, 2016. 2 +[37] Hailing Zhou, Jianmin Zheng, and Lei Wei. Representing images using curvilinear feature driven subdivision surfaces. IEEE Transactions on Image Processing, 23(8):3268-3280, 2014. 1, 2 +[38] Haokun Zhu, Juang Ian Chong, Teng Hu, Ran Yi, Yu-Kun Lai, and Paul L Rosin. SAMVG: A multi-stage image vectorization model with the segment-anything model. arXiv preprint arXiv:2311.05276, 2023. 1, 2 \ No newline at end of file diff --git a/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/images.zip b/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a799d04c51b2a68906a0b90a0092dd1bfe1fe73a --- /dev/null +++ b/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:641eb2b9299b688d99b5efa6bde40e875345b2914846fb18506a02cc66192cec +size 761713 diff --git a/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/layout.json b/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5f49c6753ac72237af47a6eb4918e2aed28b2a9a --- /dev/null +++ b/2024/SuperSVG_ Superpixel-based Scalable Vector Graphics Synthesis/layout.json @@ -0,0 +1,9524 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 103, + 499, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 103, + 499, + 121 + ], + "spans": [ + { + "bbox": [ + 94, + 103, + 499, + 121 + ], + "type": "text", + "content": "SuperSVG: Superpixel-based Scalable Vector Graphics Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "spans": [ + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "text", + "content": "Teng Hu" + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "text", + "content": ", Ran Yi" + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "text", + "content": ", Baihong Qian" + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "text", + "content": ", Jiangning Zhang" + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "text", + "content": ", Paul L. Rosin" + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "text", + "content": ", Yu-Kun Lai" + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "text", + "content": "Shanghai Jiao Tong University, " + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "text", + "content": "Youtu Lab, Tencent, " + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "text", + "content": "Cardiff University \n{hu-teng, ranyi, cherry_qbh} " + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "text", + "content": " sjtu.edu.cn, vtzhang@tencent.com, {RosinPL, LaiY4} " + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 53, + 142, + 556, + 186 + ], + "type": "text", + "content": " cardiff.ac.uk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 239, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 290, + 514 + ], + "type": "text", + "content": "SVG (Scalable Vector Graphics) is a widely used graphics format that possesses excellent scalability and editability. Image vectorization, which aims to convert raster images to SVGs, is an important yet challenging problem in computer vision and graphics. Existing image vectorization methods either suffer from low reconstruction accuracy for complex images or require long computation time. To address this issue, we propose SuperSVG, a superpixel-based vectorization model that achieves fast and high-precision image vectorization. Specifically, we decompose the input image into superpixels to help the model focus on areas with similar colors and textures. Then, we propose a two-stage self-training framework, where a coarse-stage model is employed to reconstruct the main structure and a refinement-stage model is used for enriching the details. Moreover, we propose a novel dynamic path warping loss to help the refinement-stage model to inherit knowledge from the coarse-stage model. Extensive qualitative and quantitative experiments demonstrate the superior performance of our method in terms of reconstruction accuracy and inference time compared to state-of-the-art approaches. The code is available in https://github.com/sjtuplayer/SuperSVG." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 559, + 287, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 287, + 679 + ], + "type": "text", + "content": "Scalable Vector Graphics, commonly known as SVG, is a widely used vector image format that has a wide range of applications and advantages within the domains of web design, graphic design, mobile applications, data visualization, and various other contexts. Compared with raster images that represent content by pixels, SVG describes images by parameterized vectors and benefits from its scalability and editability where it can be resized to any resolution without losing quality and can be easily manipulated by its layer-wise topological information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 680, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 680, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 59, + 680, + 287, + 693 + ], + "type": "text", + "content": "Given the superior capabilities of Scalable Vector Graph" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 212, + 535, + 350 + ], + "blocks": [ + { + "bbox": [ + 315, + 212, + 535, + 350 + ], + "lines": [ + { + "bbox": [ + 315, + 212, + 535, + 350 + ], + "spans": [ + { + "bbox": [ + 315, + 212, + 535, + 350 + ], + "type": "image", + "image_path": "8ba3dbbac75d07c23024b79669a896d16b08d24cf6ddc49ef9d48e3f017c5305.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 351, + 547, + 440 + ], + "lines": [ + { + "bbox": [ + 304, + 351, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 547, + 440 + ], + "type": "text", + "content": "Figure 1. Overview of our SuperSVG: our model first decomposes the image to be vectorized into superpixels, each containing pixels sharing similar colors and contents. The coarse-stage model predicts the path parameters to reconstruct the main structure, and then the coarse paths guided refinement model enriches the details by learning the knowledge from the coarse-stage model. Compared to the previous methods, our SuperSVG achieves both a high vectorization accuracy and fast computation speed." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 450, + 547, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 701 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 701 + ], + "type": "text", + "content": "ics (SVG) in image representation and editing, there is much research on the topic of image vectorization, which aims to convert rasterized images into SVG. The existing methods can be categorized into three classes: 1) Traditional algorithm-based methods [15, 18, 24, 28, 33, 37], where conventional algorithms are employed to fit images, but they usually suffer from a lower vectorization quality. 2) Deep-learning-based methods [9, 10, 12, 19, 21, 22], which parameterize raster images using deep neural networks for reconstruction. They are efficient and can handle the vectorization of simple graphics or characters (e.g., icons and emojis), but struggle to vectorize complex images. 3) Optimization-based methods [17, 20, 26, 38], which optimize SVG parameters to fit the target image, yielding relatively superior reconstruction quality. However, these methods entail a substantial amount of time and computational resources, making them impractical for timely processing of large-scale data. In summary, previous image vectorization methods either suffer from low reconstruction quality for complex images, or demand extensive computation time, imposing significant constraints on their practical utility." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "To achieve good vectorization quality with high ef" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24892" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 323 + ], + "type": "text", + "content": "ficiency, we propose SuperSVG, a deep-learning-based method that translates images into scalable vector graphics (SVG) in a coarse-to-fine manner. Since neural networks have difficulties in directly vectorizing complex images, we decompose the input image into different parts in the form of superpixels wherein the pixels share similar colors and textures and then vectorize each part. Then, we propose a Two-Stage Self-Teaching Training framework to vectorize the superpixels, where the coarse-stage model is trained to reconstruct the main structure of the image and the refinement-stage model is trained to enrich the image details. We make use of the predicted paths from the coarse-stage model to guide the refinement-stage model in image vectorization. Furthermore, we propose a novel Dynamic Path Warping loss which helps the refinement-stage to inherit the knowledge of the coarse-stage model. With the help of the superpixel-based image decomposition and the two-stage self-teaching framework, our SuperSVG can keep the image structure well and reconstruct more details at high speed. Extensive quantitative and qualitative experiments validate the effectiveness of our model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 324, + 261, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 324, + 261, + 335 + ], + "spans": [ + { + "bbox": [ + 59, + 324, + 261, + 335 + ], + "type": "text", + "content": "The main contributions of our work are four-fold:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 339, + 287, + 578 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 339, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 339, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 339, + 287, + 386 + ], + "type": "text", + "content": "- We propose SuperSVG, a novel superpixel-based vectorization model that translates the rasterized images into scalable vector graphics (SVG) based on superpixels and vectorizes the superpixels in a coarse-to-fine manner." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 387, + 287, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 287, + 446 + ], + "type": "text", + "content": "- We design a Two-Stage Self-Teaching Training framework, where we employ a coarse-stage model to reconstruct the main structures and a refinement-stage model to enrich the image details based on the coarse-stage output." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 447, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 287, + 506 + ], + "type": "text", + "content": "- We propose a Coarse Paths Guided Training strategy to guide the refinement-stage model to inherit the knowledge from the coarse-stage model, which greatly improves the performance of the refinement-stage model and avoids converging to suboptimal local minimum." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 287, + 578 + ], + "type": "text", + "content": "- We propose Dynamic Path Warping (DPW) loss, which measures the distance between the predicted paths from the refinement-stage model and the pseudo ground truth approximated with coarse paths. By minimizing the DPW loss, the refinement-stage model can distill the knowledge from the coarse-stage model." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 590, + 134, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 134, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 134, + 602 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 610, + 164, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 164, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 164, + 623 + ], + "type": "text", + "content": "2.1. Image Vectorization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "text", + "content": "Image vectorization aims to transform a rasterized image into scalable vector graphics (SVG) composed of parameterized vectors. Different from raster images that may become blurry when zooming in, SVG can be rendered at any resolution without losing quality and is convenient to edit, widely used in web design, graphic design, etc. The existing vectorization methods can be classified into 3 categories:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "Traditional Algorithm-based Image Vectorization Methods can be classified into mesh-based and curve-based ones. The mesh-based methods [15, 18, 28, 37] segment an input image into non-overlapping patches, and infer the color and the boundary location for each region. The curve-based methods [1, 5, 24, 32, 36] employ Bézier curves with different colors defined on either side to create the vector image. Potrace [24] is a representative method of this type that projects the smooth outlines into Bézier paths, and merge the adjacent paths together. However, the vectorization quality of these methods still needs improvements." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 204, + 546, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 546, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 546, + 382 + ], + "type": "text", + "content": "Deep-learning-based Image Vectorization Methods use neural networks to project a raster image into SVG. Im2Vec [21] employs a variational auto-encoder (VAE) [14] to embed the input image and then maps it into path parameters by a Long Short-Term Memory (LSTM) module [23]. Raster2Vec [19] is focused on vectorization of rasterized floor plans using a ResNet [11]. Gao et al. [9] rely on a pre-trained VGG network [25] and a hierarchical Recurrent Neural Network (RNN) to output parametric curves of different sizes. But these methods only focus on simple images and cannot vectorize complex images well. In comparison, our SuperSVG is the first deep-learning-based method that can vectorize images with complex details, thanks to our superpixel decomposition and coarse-path guided refinement that substantially reduce the learning difficulties." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 384, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 515 + ], + "type": "text", + "content": "Optimization-based Image Vectorization Methods. DiffVG [17] proposes a differentiable renderer that renders the SVG parameters into images. Based on this, DiffVG minimizes the distance between the rasterized and vector images by optimizing the SVG parameters using gradient descent. LIVE [20] and SAMVG [38] further introduce a layer-wise optimization framework, which achieves better vectorization quality over the previous methods. However, due to the low optimization efficiency, they suffer from a long optimization time. In contrast, our SuperSVG achieves both a good vectorization quality and high efficiency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 521, + 451, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 521, + 451, + 534 + ], + "spans": [ + { + "bbox": [ + 306, + 521, + 451, + 534 + ], + "type": "text", + "content": "2.2. Superpixel Decomposition" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 540, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 658 + ], + "type": "text", + "content": "Superpixel decomposition is usually used for data preprocessing in vision tasks. Existing superpixel decomposition methods can be categorized into methods based on traditional algorithm or deep learning. For the traditional algorithm based methods, diverse strategies have been employed, e.g., energy-driven sampling [30], geometric flows [16] and clustering [3]. Some recent works [13, 29, 34] employ neural networks to enhance the performance in superpixel decomposition, which shows great potential in this task." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 669, + 361, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 361, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 361, + 681 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Image vectorization aims to translate a rasterized image " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " into a Scalable Vector Graphic (SVG). An SVG is com" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24893" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 68, + 521, + 181 + ], + "blocks": [ + { + "bbox": [ + 72, + 68, + 521, + 181 + ], + "lines": [ + { + "bbox": [ + 72, + 68, + 521, + 181 + ], + "spans": [ + { + "bbox": [ + 72, + 68, + 521, + 181 + ], + "type": "image", + "image_path": "f228249a05e0d8180b1b9ff0e963639311473963883245fbda2d937d28655add.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 72, + 182, + 517, + 291 + ], + "blocks": [ + { + "bbox": [ + 72, + 182, + 517, + 291 + ], + "lines": [ + { + "bbox": [ + 72, + 182, + 517, + 291 + ], + "spans": [ + { + "bbox": [ + 72, + 182, + 517, + 291 + ], + "type": "image", + "image_path": "38f6b6096415ea91de2579883a6ff84fc00d4e87ae10f9f385c01b8edda6b5e2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 293, + 545, + 338 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 545, + 338 + ], + "type": "text", + "content": "Figure 2. Main framework of our SuperSVG: we decompose the target image into superpixels and vectorize each superpixel separately. We employ an attention-based coarse-stage model to predict SVG paths that reconstruct the main structure of the superpixel. Then, a refinement-stage model guided by the coarse paths is designed to predict more SVG paths to refine details based on the coarse image. Finally, by combining all the predicted SVGs for each superpixel, we obtain an output SVG with good structure and fine details." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 347, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 526 + ], + "type": "text", + "content": "posed of many vector primitives, which can be SVG paths, ellipses, circles, or rectangles, etc. Following previous works [17, 20, 21], we employ the SVG paths as the shape primitive, where each SVG path defines a region constructed by multiple cubic Bezier curves connected end-to-end with certain color. With the parameters of these SVG paths, the rasterized image can be rendered in any resolution. To obtain the path parameters, some previous methods [17, 20] optimize the path parameters directly to minimize the distance between the input image and the rendered image, which achieves good reconstruction quality but requires long optimization time. To speed up the vectorization process, some deep-learning-based methods [21] employ a deep-learning model to predict the SVG path parameters, but struggle to vectorize complex images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "content": "To achieve good vectorization quality with high efficiency, we propose SuperSVG, a deep-learning-based image vectorization method that translates images into SVG path parameters automatically. To improve the model ability to vectorize complex images, we segment the input image into different parts, within which the pixels share similar colors and textures, and then vectorize each part separately, where superpixels are used for image segmentation as they tend to maintain compactness, uniformity and regularity, particularly suitable for our task. For each superpixel " + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{X}" + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "content": " is the set of all superpixels, our model converts it into a sequence of path parameters, where each path is composed of several cubic Bezier curves and has a fill color, with a total of " + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "content": " parameters. With the predicted path parameters for each superpixel, we employ the differ" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 347, + 545, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 347, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 545, + 383 + ], + "type": "text", + "content": "entiable renderer " + }, + { + "bbox": [ + 305, + 347, + 545, + 383 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 305, + 347, + 545, + 383 + ], + "type": "text", + "content": " from DiffVG [17] to get the rendered image " + }, + { + "bbox": [ + 305, + 347, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 305, + 347, + 545, + 383 + ], + "type": "text", + "content": " in pixel space, which is expected to be close to the input image " + }, + { + "bbox": [ + 305, + 347, + 545, + 383 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 305, + 347, + 545, + 383 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "text", + "content": "We propose a two-stage self-teaching framework, composed of a coarse-stage model " + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "text", + "content": " to reconstruct the main structure and a refinement model " + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "inline_equation", + "content": "E_{r}" + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "text", + "content": " to enrich the details, where the predicted paths from coarse-stage model are used to guide the refinement model in vectorization. " + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "text", + "content": " takes the superpixel " + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "text", + "content": " as input and outputs " + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "text", + "content": " paths " + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "inline_equation", + "content": "S = \\{s_1,s_2,\\dots s_n\\}" + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "text", + "content": " to reconstruct main structure; while " + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "inline_equation", + "content": "E_{r}" + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "text", + "content": " takes both the rendered image " + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "inline_equation", + "content": "R(S)" + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "text", + "content": " and target superpixel " + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "text", + "content": " as inputs, and outputs " + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "text", + "content": " paths " + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "inline_equation", + "content": "S^{\\prime} = \\{s_{1}^{\\prime},s_{2}^{\\prime}\\dots s_{m}^{\\prime}\\}" + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "text", + "content": " to refine details. Combining all the predicted " + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "inline_equation", + "content": "S^{\\prime}" + }, + { + "bbox": [ + 305, + 384, + 546, + 515 + ], + "type": "text", + "content": " for each superpixel produces the final SVG result." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 526, + 518, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 518, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 518, + 539 + ], + "type": "text", + "content": "3.1. Superpixel-based Coarse Reconstruction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Superpixel decomposition. Considering the optimization-based methods suffer from a long optimization time, our SuperSVG builds upon neural networks to efficiently predict SVG paths. However, as neural networks have difficulties in directly vectorizing complex images [8] we therefore simplify the task to vectorizing a certain part of the image containing homogeneous colors and textures. Since superpixel algorithms provide a good tool to decompose images based on local pixel color and also ensure alignment of the regions with the image boundaries, we segment the input image into superpixels, and our model reconstructs each superpixel with scalable vectors. Superpixels also tend to be more regular, making them easier for vectorization. Specifically, we utilize SLIC [3] to decompose the input image" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24894" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 69, + 267, + 166 + ], + "blocks": [ + { + "bbox": [ + 66, + 69, + 267, + 166 + ], + "lines": [ + { + "bbox": [ + 66, + 69, + 267, + 166 + ], + "spans": [ + { + "bbox": [ + 66, + 69, + 267, + 166 + ], + "type": "image", + "image_path": "8807daf335b8d788c8ebb22f45ddb52a6652ad9434a59c59f88c3849cd7ba8ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 167, + 287, + 200 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 287, + 200 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 287, + 200 + ], + "type": "text", + "content": "Figure 3. Illustration of our boundary loss " + }, + { + "bbox": [ + 46, + 167, + 287, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{Bound}}" + }, + { + "bbox": [ + 46, + 167, + 287, + 200 + ], + "type": "text", + "content": ", which computes the area of the SVG paths that are outside the superpixel mask, and guides the paths to be inside the superpixel." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 210, + 287, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 210, + 287, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 287, + 234 + ], + "type": "text", + "content": "into superpixels. We set the compactness parameter as 30 for SLIC to make superpixels more regular." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 234, + 287, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 287, + 330 + ], + "type": "text", + "content": "Coarse-stage model. For a superpixel " + }, + { + "bbox": [ + 46, + 234, + 287, + 330 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 234, + 287, + 330 + ], + "type": "text", + "content": " with mask mask (indicating those pixels within " + }, + { + "bbox": [ + 46, + 234, + 287, + 330 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 234, + 287, + 330 + ], + "type": "text", + "content": " with 1 and 0 otherwise), we first design the coarse-stage model " + }, + { + "bbox": [ + 46, + 234, + 287, + 330 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 46, + 234, + 287, + 330 + ], + "type": "text", + "content": " to vectorize the main structure by predicting the SVG path sequence " + }, + { + "bbox": [ + 46, + 234, + 287, + 330 + ], + "type": "inline_equation", + "content": "E_{c}(x) = S = \\{s_{1}, s_{2}, \\dots, s_{n}\\}" + }, + { + "bbox": [ + 46, + 234, + 287, + 330 + ], + "type": "text", + "content": ". Inspired by AttnPainter [27], " + }, + { + "bbox": [ + 46, + 234, + 287, + 330 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 46, + 234, + 287, + 330 + ], + "type": "text", + "content": " is composed of a Vision Transformer (ViT) encoder [7] and a cross-attention module followed by a self-attention module, which is shown in Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "content": "Specifically, the ViT encoder first encodes the input superpixel " + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "content": " into image feature " + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "inline_equation", + "content": "T_{f}" + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "content": ". To control the number of output paths " + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "inline_equation", + "content": "(n)" + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "content": " and the parameter number in each path " + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "inline_equation", + "content": "(N_{p})" + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "content": ", we employ a cross-attention module to calculate the correlation between the image feature " + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "inline_equation", + "content": "T_{f}" + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "inline_equation", + "content": "n N_{p}" + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "content": "-dimensional learnable path queries " + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "inline_equation", + "content": "T_{l}" + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "content": ", and output an intermediate feature " + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "inline_equation", + "content": "T_{f}^{\\prime}" + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "content": " with the shape of path parameters " + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "inline_equation", + "content": "(n \\times N_{p})" + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "content": ", which is formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 434, + 287, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 434, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 70, + 434, + 287, + 460 + ], + "type": "interline_equation", + "content": "T _ {f} ^ {\\prime} = \\operatorname {S o f t m a x} \\left(\\frac {\\left(W _ {Q} T _ {l}\\right) \\left(W _ {K} T _ {f}\\right) ^ {T}}{\\sqrt {d}}\\right) \\left(W _ {V} T _ {f}\\right), \\tag {1}", + "image_path": "4915665c87d8469ab6d3d3142eb3348dabb74d39fa6d468ce216464dcc771a6a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 468, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 468, + 287, + 491 + ], + "type": "inline_equation", + "content": "W_{Q}, W_{K}" + }, + { + "bbox": [ + 46, + 468, + 287, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 468, + 287, + 491 + ], + "type": "inline_equation", + "content": "W_{V}" + }, + { + "bbox": [ + 46, + 468, + 287, + 491 + ], + "type": "text", + "content": " are learnable query, key and value matrices." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 492, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 287, + 529 + ], + "type": "text", + "content": "Then, a self-attention module is employed to process the image feature " + }, + { + "bbox": [ + 46, + 492, + 287, + 529 + ], + "type": "inline_equation", + "content": "T_f'" + }, + { + "bbox": [ + 46, + 492, + 287, + 529 + ], + "type": "text", + "content": " and project it into the parameter space with " + }, + { + "bbox": [ + 46, + 492, + 287, + 529 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 492, + 287, + 529 + ], + "type": "text", + "content": " paths, where each path contains " + }, + { + "bbox": [ + 46, + 492, + 287, + 529 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 46, + 492, + 287, + 529 + ], + "type": "text", + "content": " parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "content": "Training objectives. 1) Normalized Reconstruction Loss: We employ the differentiable renderer " + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "content": " in DiffVG [17] to render a raster image " + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\hat{x} = R(S)" + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "content": " from the predicted path parameters " + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "content": ". Then, we train the coarse-stage model " + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "content": " by minimizing the normalized " + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "content": " distance between the rendering " + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\hat{x} = R(S)" + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "content": " and the target image:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 103, + 609, + 287, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 609, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 103, + 609, + 287, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {2} = \\| \\hat {x} - x \\| ^ {2} \\cdot \\frac {\\sum_ {p} \\operatorname {m a s k} (p)}{w h}, \\tag {2}", + "image_path": "152af55bd3e27df1a71db925089943da0c546d94f6bf0cd0a414fb51f3a8e88c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "content": " are the width and height of the superpixel image " + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "content": " and the superpixel mask mask, and " + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\text{mask}(p)" + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "content": " indicates the mask value (1 or 0) for pixel " + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "2) Boundary Loss: To avoid the SVG path from crossing the superpixel boundary, we propose boundary loss to guide the paths to be inside the superpixel. We set the color of all" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 334, + 68, + 518, + 141 + ], + "blocks": [ + { + "bbox": [ + 334, + 68, + 518, + 141 + ], + "lines": [ + { + "bbox": [ + 334, + 68, + 518, + 141 + ], + "spans": [ + { + "bbox": [ + 334, + 68, + 518, + 141 + ], + "type": "image", + "image_path": "28ca982cacb49dad098847c5be0b79419cfec7462967d7ca8c273652d8f001cb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 142, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 305, + 142, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 305, + 142, + 545, + 198 + ], + "type": "text", + "content": "Figure 4. Problem of training the refinement model with " + }, + { + "bbox": [ + 305, + 142, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 305, + 142, + 545, + 198 + ], + "type": "text", + "content": " loss alone: optimizing a newly-added path on the canvas by " + }, + { + "bbox": [ + 305, + 142, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 305, + 142, + 545, + 198 + ], + "type": "text", + "content": " gradually pulls it to disappear (as a suboptimal local minimum). With our proposed coarse paths guided training and DPW loss, the added path is successfully optimized to resemble the target." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 205, + 545, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 205, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 545, + 231 + ], + "type": "text", + "content": "predicted SVG paths to 1 (white) to get a new path sequence " + }, + { + "bbox": [ + 305, + 205, + 545, + 231 + ], + "type": "inline_equation", + "content": "S_{binary}" + }, + { + "bbox": [ + 305, + 205, + 545, + 231 + ], + "type": "text", + "content": ". Then, we compute the boundary loss by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 330, + 238, + 545, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 238, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 330, + 238, + 545, + 257 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {B o u n d}} = \\underset {p \\sim \\operatorname {m a s k}} {\\mathbb {E}} (R \\left(S _ {\\text {b i n a r y}}\\right) \\cdot (1 - \\operatorname {m a s k})) \\tag {3}", + "image_path": "7eae66172593cd7192db6a4bb3d2c6c3e354d04e71079876ef0ec56938e10d6e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 264, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 335 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 304, + 264, + 545, + 335 + ], + "type": "inline_equation", + "content": "(1 - \\text{mask})" + }, + { + "bbox": [ + 304, + 264, + 545, + 335 + ], + "type": "text", + "content": " is 1 outside the superpixel mask and 0 inside the mask, the loss term calculates the area of the paths that are outside the superpixel. When some SVG paths cross the superpixel boundary, the path area outside is penalized; while when all SVG paths are inside the superpixel, " + }, + { + "bbox": [ + 304, + 264, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{Bound}}" + }, + { + "bbox": [ + 304, + 264, + 545, + 335 + ], + "type": "text", + "content": " reaches 0 (Fig. 3)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 336, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 431 + ], + "type": "text", + "content": "3) Path Efficiency Loss: To enable our model to reconstruct the maximum amount of information with the fewest paths, we propose the path efficiency loss " + }, + { + "bbox": [ + 304, + 336, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PE}" + }, + { + "bbox": [ + 304, + 336, + 545, + 431 + ], + "type": "text", + "content": ". Specifically, for each path " + }, + { + "bbox": [ + 304, + 336, + 545, + 431 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 336, + 545, + 431 + ], + "type": "text", + "content": ", an additional opacity parameter " + }, + { + "bbox": [ + 304, + 336, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 304, + 336, + 545, + 431 + ], + "type": "text", + "content": " is predicted. We treat the path as visible if " + }, + { + "bbox": [ + 304, + 336, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\beta_{i} \\geq 0.5" + }, + { + "bbox": [ + 304, + 336, + 545, + 431 + ], + "type": "text", + "content": ", and the loss " + }, + { + "bbox": [ + 304, + 336, + 545, + 431 + ], + "type": "inline_equation", + "content": "L_{PE}" + }, + { + "bbox": [ + 304, + 336, + 545, + 431 + ], + "type": "text", + "content": " penalizes the case with more visible paths, i.e., to encourage reconstructing the image with as few paths as possible, calculated as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 361, + 438, + 545, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 438, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 361, + 438, + 545, + 453 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {P E} = \\sum_ {i = 1} ^ {n} \\operatorname {S i g n} \\left(\\beta_ {i} - 0. 5\\right), \\tag {4}", + "image_path": "79be3ae1464ec8dc7c8975a16c59ea3349759bc63d10ce3ae08b9c6bfccaec89.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 459, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 459, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 459, + 545, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 459, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\operatorname{Sign}(\\cdot)" + }, + { + "bbox": [ + 305, + 459, + 545, + 498 + ], + "type": "text", + "content": " is the sign function. Since " + }, + { + "bbox": [ + 305, + 459, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\operatorname{Sign}(\\beta_i)" + }, + { + "bbox": [ + 305, + 459, + 545, + 498 + ], + "type": "text", + "content": " is not differentiable, we approximate " + }, + { + "bbox": [ + 305, + 459, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\frac{\\partial \\operatorname{Sign}(\\beta_i)}{\\partial \\beta_i}" + }, + { + "bbox": [ + 305, + 459, + 545, + 498 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 305, + 459, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\operatorname{Sig}(\\beta_i)(1 - \\operatorname{Sig}(\\beta_i))" + }, + { + "bbox": [ + 305, + 459, + 545, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 459, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\operatorname{Sig}(\\cdot)" + }, + { + "bbox": [ + 305, + 459, + 545, + 498 + ], + "type": "text", + "content": " is the sigmoid function." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 499, + 497, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 499, + 497, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 499, + 497, + 510 + ], + "type": "text", + "content": "The final training objective is formulated as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 517, + 545, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 517, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 318, + 517, + 545, + 537 + ], + "type": "interline_equation", + "content": "E _ {c} ^ {*} = \\underset {E _ {c}} {\\arg \\min } \\mathcal {L} _ {2} + \\lambda_ {\\text {B o u n d}} \\mathcal {L} _ {\\text {B o u n d}} + \\lambda_ {P E} \\mathcal {L} _ {P E}. \\tag {5}", + "image_path": "b25d89d3ec1c9b95d062e906899637623fca2546f38f6a5c1523c2234f1f4756.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 552, + 512, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 512, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 512, + 564 + ], + "type": "text", + "content": "3.2. Coarse Paths Guided Refinement Stage" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "content": "In the coarse reconstruction stage, our coarse-stage model " + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "content": " can output an SVG that captures and reconstructs the main structure of the input superpixel " + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "content": ". The rendered image from the SVG (denoted as " + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "inline_equation", + "content": "c_{1}" + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "content": ") resembles " + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "content": " in general, but lacks some image details, especially when the superpixel is complex. To enrich image details, we employ a refinement model " + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "inline_equation", + "content": "E_{r}" + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "content": " to predict more SVG paths to add more details based on the current canvas " + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "inline_equation", + "content": "c_{1}" + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Model framework. Different from the coarse-stage model " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " that only takes the target superpixel " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " as input, the refinement model " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "E_{r}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " takes both the current canvas " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "c_{1}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " (rendered from the coarse stage output) and the target superpixel" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24895" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " as inputs, and predict new paths " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "S^{\\prime} = \\{s_1^{\\prime}, s_2^{\\prime} \\cdots s_m^{\\prime}\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " to be overlaid onto the canvas to refine details. Specifically, 3 convolution layers followed by ReLU activations are employed to fuse the two input images into a feature map. After getting the fused feature map, " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "E_{r}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " shares the same structure as the coarse-stage model " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ", which encodes the fused feature map by a ViT Encoder and maps the encoded features into path parameters by a cross-attention and a self-attention layer. To accelerate the training process, we inherit the weights of the ViT encoder in " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " as an initialization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 288, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 288, + 239 + ], + "type": "text", + "content": "Local optimal solution with " + }, + { + "bbox": [ + 46, + 192, + 288, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 46, + 192, + 288, + 239 + ], + "type": "text", + "content": " loss. The goal of the refinement model is to reconstruct more details of the input superpixel based on the current canvas. A simple " + }, + { + "bbox": [ + 46, + 192, + 288, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 46, + 192, + 288, + 239 + ], + "type": "text", + "content": " loss defined as follows is used as the reconstruction loss:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 242, + 287, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 242, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 53, + 242, + 287, + 268 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {2} = \\| x - R ([ E _ {c} (x), E _ {r} (x, c _ {1}) ]) \\| ^ {2} \\cdot \\frac {\\sum_ {p} m a s k (p)}{w h}, \\tag {6}", + "image_path": "954ea27a407a3eb53ab745b9ed298550df1417a2eabc0f9cfdb74c696fe18d04.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 270, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 287, + 319 + ], + "type": "text", + "content": "where the predicted path sequences by coarse-stage model " + }, + { + "bbox": [ + 47, + 270, + 287, + 319 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 47, + 270, + 287, + 319 + ], + "type": "text", + "content": " and refinement model " + }, + { + "bbox": [ + 47, + 270, + 287, + 319 + ], + "type": "inline_equation", + "content": "E_{r}" + }, + { + "bbox": [ + 47, + 270, + 287, + 319 + ], + "type": "text", + "content": " are concatenated together to get the final SVG result, and the rendered image of which is expected to resemble the input superpixel " + }, + { + "bbox": [ + 47, + 270, + 287, + 319 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 270, + 287, + 319 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "text", + "content": "However, the refinement model " + }, + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "inline_equation", + "content": "E_r" + }, + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "text", + "content": " trained with Eq.(6) alone tends to predict paths that are extremely small in area, or even invisible. In Fig. 4, we use an example to illustrate this phenomenon more clearly: we newly add a path onto the canvas and optimize the path parameters with " + }, + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "text", + "content": " loss; it can be seen that the new path gradually shrinks and finally disappears in the canvas. A possible reason is that the coarse stage result " + }, + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "inline_equation", + "content": "c_1" + }, + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "text", + "content": " is already close to " + }, + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "text", + "content": ", and a local optimal solution for " + }, + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "inline_equation", + "content": "E_r" + }, + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "text", + "content": " is to overlap nothing onto " + }, + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "inline_equation", + "content": "c_1" + }, + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "text", + "content": ", which is better than adding a sub-optimal path and can prevent the " + }, + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "text", + "content": " distance from increasing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "text", + "content": "Coarse paths guided training framework. To avoid the refinement model from falling into poor local optimum, we propose a coarse paths guided framework, which inherits the knowledge from the coarse-stage model to help train the refinement model with an additional constraint on the SVG path parameters. As illustrated in Fig. 2, for an input superpixel " + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "text", + "content": ", we first use the coarse-stage model to predict a coarse level path sequence " + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "inline_equation", + "content": "S = \\{s_1, s_2 \\dots s_n\\}" + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "text", + "content": ". Then, we randomly choose a value " + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "inline_equation", + "content": "k \\in (1, n - m)" + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "text", + "content": " and split the predicted path sequence into two subsequences: " + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "inline_equation", + "content": "S_1 = \\{s_1, s_2, \\dots, s_k\\}" + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "inline_equation", + "content": "S_2 = \\{s_{k + 1}, s_{k + 2}, \\dots, s_n\\}" + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "text", + "content": ". The subsequence " + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "inline_equation", + "content": "S_1" + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "text", + "content": " is then rendered into " + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "inline_equation", + "content": "c_1 = R(S_1)" + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "text", + "content": " and used as the input canvas for the refinement model " + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "inline_equation", + "content": "E_r" + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "text", + "content": ", while the remaining subsequence " + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "inline_equation", + "content": "S_2" + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "text", + "content": " can be regarded as a pseudo ground truth for the output path sequence of " + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "inline_equation", + "content": "E_r" + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "text", + "content": ". Specifically, when training " + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "inline_equation", + "content": "E_r" + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "text", + "content": ", in addition to the previous constraints that operate in the pixel space, we add a new constraint on the path parameter space, which minimizes the distance between path sequences " + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "inline_equation", + "content": "S'" + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "inline_equation", + "content": "S_2" + }, + { + "bbox": [ + 46, + 450, + 288, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "In detail, during training, the refinement model " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "E_{r}" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " takes both the canvas " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "c_{1} = R(S_{1})" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " and the target image " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " as inputs, and outputs the path sequence " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "S^{\\prime} = \\{s_1^{\\prime},s_2^{\\prime},\\dots ,s_m^{\\prime}\\}" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 545, + 85 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 85 + ], + "type": "text", + "content": "Algorithm 1 Forward pass to efficiently compute " + }, + { + "bbox": [ + 306, + 72, + 545, + 85 + ], + "type": "inline_equation", + "content": "\\mathbf{dpw}_{\\gamma}(\\mathbf{S},\\mathbf{S}^{\\prime})" + } + ] + } + ], + "index": 7, + "type": "text" + }, + { + "bbox": [ + 306, + 88, + 498, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 88, + 498, + 99 + ], + "spans": [ + { + "bbox": [ + 306, + 88, + 498, + 99 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 306, + 88, + 498, + 99 + ], + "type": "inline_equation", + "content": "S, S'" + }, + { + "bbox": [ + 306, + 88, + 498, + 99 + ], + "type": "text", + "content": ", smoothing " + }, + { + "bbox": [ + 306, + 88, + 498, + 99 + ], + "type": "inline_equation", + "content": "\\gamma \\geq 0" + }, + { + "bbox": [ + 306, + 88, + 498, + 99 + ], + "type": "text", + "content": ", distance function " + }, + { + "bbox": [ + 306, + 88, + 498, + 99 + ], + "type": "inline_equation", + "content": "d" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 100, + 514, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 100, + 514, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 100, + 514, + 175 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} 1: p _ {0, j} = 0; p _ {i, 0} = q _ {i, 0} = q _ {0, j} = \\infty , i \\in [ [ n ] ], j \\in [ [ m ] ] \\\\ 2: \\text {f o r} j = 1, \\dots , m \\text {d o} \\\\ 3: \\quad \\text {f o r} i = 1, \\dots , n \\mathbf {d o} \\\\ 4: \\quad p _ {i, j} = d _ {i, j} + \\min ^ {\\gamma} \\left(q _ {i, j - 1}, p _ {i, j - 1}\\right) \\\\ 5: \\quad q _ {i, j} = \\min ^ {\\gamma} \\left(q _ {i - 1, j}, p _ {i - 1, j}\\right) \\\\ 6: \\quad e n d f o r \\\\ 7: \\text {e n d f o r} \\\\ \\end{array}", + "image_path": "1aa09455f88610ba797c72cfbfa908459e9ab540e33fa45306a457dd4d5d12a6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 176, + 441, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 176, + 441, + 188 + ], + "spans": [ + { + "bbox": [ + 307, + 176, + 441, + 188 + ], + "type": "text", + "content": "Output: " + }, + { + "bbox": [ + 307, + 176, + 441, + 188 + ], + "type": "inline_equation", + "content": "(\\min^{\\gamma}(p_{n,m},q_{n,m}),P,Q)" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 200, + 531, + 322 + ], + "blocks": [ + { + "bbox": [ + 318, + 200, + 531, + 322 + ], + "lines": [ + { + "bbox": [ + 318, + 200, + 531, + 322 + ], + "spans": [ + { + "bbox": [ + 318, + 200, + 531, + 322 + ], + "type": "image", + "image_path": "8ec5c60eed98ffe5f022368c197173c389f6a1d4b15c5d5a37d5fd21d67eea2e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 323, + 468, + 330 + ], + "lines": [ + { + "bbox": [ + 391, + 323, + 468, + 330 + ], + "spans": [ + { + "bbox": [ + 391, + 323, + 468, + 330 + ], + "type": "text", + "content": "(a) Difference in algorithms" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 328, + 331, + 517, + 407 + ], + "blocks": [ + { + "bbox": [ + 328, + 331, + 517, + 407 + ], + "lines": [ + { + "bbox": [ + 328, + 331, + 517, + 407 + ], + "spans": [ + { + "bbox": [ + 328, + 331, + 517, + 407 + ], + "type": "image", + "image_path": "9991f250cb023d47a691ae2f8399607781639e282d071ee56d6d8363b063bfcf.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 412, + 545, + 468 + ], + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 468 + ], + "type": "text", + "content": "Figure 5. Difference between DTW and our DPW. a) Both the DTW and DPW loss calculate the sum of distances of elements colored yellow. The difference is that one generated path can only match one target path in DPW to avoid averaging several target paths. b) The comparison between the training processes." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 472, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 497 + ], + "type": "text", + "content": "We minimize both the distance in pixel space and path parameter space, with the total loss formulated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 503, + 545, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 503, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 315, + 503, + 545, + 533 + ], + "type": "interline_equation", + "content": "E _ {r} ^ {*} = \\underset {E _ {r}} {\\arg \\min } \\mathcal {L} _ {2} + \\lambda_ {D P W} \\mathcal {L} _ {D P W} + \\lambda_ {B o u n d} \\mathcal {L} _ {B o u n d}, \\tag {7}", + "image_path": "08d8e59563c821acbb66e1ce73ce3a58744739c1c9f944b410f263d95a0b3063.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 534, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 534, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DPW}" + }, + { + "bbox": [ + 304, + 534, + 545, + 559 + ], + "type": "text", + "content": " (detailed in Sec. 3.3) is the distance between the path sequences " + }, + { + "bbox": [ + 304, + 534, + 545, + 559 + ], + "type": "inline_equation", + "content": "S_{2}" + }, + { + "bbox": [ + 304, + 534, + 545, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 545, + 559 + ], + "type": "inline_equation", + "content": "S^{\\prime}" + }, + { + "bbox": [ + 304, + 534, + 545, + 559 + ], + "type": "text", + "content": " in the path parameter space." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 565, + 440, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 565, + 440, + 578 + ], + "spans": [ + { + "bbox": [ + 306, + 565, + 440, + 578 + ], + "type": "text", + "content": "3.3. Dynamic Path Warping" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 583, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 583, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 583, + 545, + 675 + ], + "type": "text", + "content": "To calculate the distance between the target path sequence " + }, + { + "bbox": [ + 304, + 583, + 545, + 675 + ], + "type": "inline_equation", + "content": "S = \\{s_1, s_2, \\dots, s_n\\}" + }, + { + "bbox": [ + 304, + 583, + 545, + 675 + ], + "type": "text", + "content": " and the predicted path sequence " + }, + { + "bbox": [ + 304, + 583, + 545, + 675 + ], + "type": "inline_equation", + "content": "S' = \\{s_1', s_2', \\dots, s_m'\\}" + }, + { + "bbox": [ + 304, + 583, + 545, + 675 + ], + "type": "text", + "content": ", Dynamic Time Warping (DTW) [4] is a commonly used metric, which finds an optimal matching match between the two ordered sequences (the yellow points in Fig. 5(a)) to minimize the accumulated distance " + }, + { + "bbox": [ + 304, + 583, + 545, + 675 + ], + "type": "inline_equation", + "content": "\\sum_{i}^{n} \\| s_i - s_{match(i)}' \\|^2" + }, + { + "bbox": [ + 304, + 583, + 545, + 675 + ], + "type": "text", + "content": " where match(i + 1) ≥ match(i)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Note that the monotonicity of the matching function cannot be ignored since the path sequences are well-ordered where latter paths are overlaid on former paths." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24896" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": "In our coarse paths guided framework, we expect the generated path sequence to be a subsequence of the target path sequence. However, in DTW, one generated path " + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "s_j'" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": " can correspond to several target paths " + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "s_{i1},\\ldots ,s_{il}" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": ". Therefore, when trained with DTW, one generated path tends to become the average of several target paths. As shown in Fig. 5(b), when optimizing 3 paths to match the target emoji (with 4 paths), one path becomes the average of the two eyebrow paths, which is not our desired case." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": "To address this issue, we propose Dynamic Path Warping (DPW), where each generated path should match one and only one target path, and some target paths can be skipped (to learn a subsequence), as shown in Fig. 5(a), each horizontal line only passes through one matching point (yellow). To compute the DPW, we define " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "p_{i,j}" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": " as the minimum accumulated distance when " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": " matches " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "s_j'" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "q_{i,j}" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": " as the minimum accumulated distance when " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "s_j'" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": " has been matched to one path before " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": " (not including " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": "). We employ dynamic programming to compute the final DPW loss " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "p_{n,m}" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": " as shown in Alg. 1. For each " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "p_{i,j}" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": ", the distance " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "d_{i,j}" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "s_j'" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": " is added to the smaller one of " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "q_{i,j-1}" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "p_{i,j-1}" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": ". And for each " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "q_{i,j}" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": ", its value takes the smaller one between " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "q_{i-1,j}" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "p_{i-1,j}" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": " (more explanations are provided in the supplementary material). Moreover, to make Alg. 1 differentiable, we follow SoftDTW [4] to substitute the " + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "inline_equation", + "content": "\\min(\\cdot)" + }, + { + "bbox": [ + 46, + 179, + 289, + 373 + ], + "type": "text", + "content": " operation:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 379, + 280, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 379, + 280, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 379, + 280, + 411 + ], + "type": "interline_equation", + "content": "\\min^ {\\gamma} (a _ {0}, a _ {1},..., a _ {n}) = \\left\\{ \\begin{array}{l l} \\min _ {i \\leq n} a _ {i} & \\gamma = 0, \\\\ - \\gamma \\log \\sum_ {i = 1} ^ {n} e ^ {- a _ {i} / \\gamma} & \\gamma > 0. \\end{array} \\right.", + "image_path": "f33dc85d358bd5f1ed4ee3a651b3ddd2b72fedf4e0faa24ee7d30c5fb365dbe3.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 424, + 128, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 424, + 128, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 424, + 128, + 437 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 444, + 162, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 162, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 162, + 456 + ], + "type": "text", + "content": "4.1. Experiment Setting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 462, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 289, + 713 + ], + "type": "text", + "content": "Implementation Details. We use SVG paths composed by cubic Bézier curves as the vector primitive, where each SVG path is closed, composed of 4 cubic Bézier curves connected end-to-end and has a fill color. Each SVG path has 28 parameters (24 for shape, 3 for color, and 1 for visibility). The coarse-stage model is trained to predict 128 paths for each superpixel first. Then, the refinement model is trained to predict 8 paths at one time. We train both the coarse-stage model and refinement model on ImageNet dataset [6]. We set batch size as 64 and learning rate as " + }, + { + "bbox": [ + 46, + 462, + 289, + 713 + ], + "type": "inline_equation", + "content": "2.5 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 462, + 289, + 713 + ], + "type": "text", + "content": ". We train the coarse-stage model for 200K iterations with 5K warm up iterations, and train the refinement model for 200K iterations with " + }, + { + "bbox": [ + 46, + 462, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{DPW}" + }, + { + "bbox": [ + 46, + 462, + 289, + 713 + ], + "type": "text", + "content": " decreasing from " + }, + { + "bbox": [ + 46, + 462, + 289, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 46, + 462, + 289, + 713 + ], + "type": "text", + "content": " to 0 in 10K iterations uniformly. In the following experiments, we implement our model with two versions: 1) SuperSVG-B, that decomposes the image into superpixels and vectorizes them by the coarse-stage model and refinement model, and 2) SuperSVG-F, which finetunes the SVG parameters from SuperSVG-B with " + }, + { + "bbox": [ + 46, + 462, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 46, + 462, + 289, + 713 + ], + "type": "text", + "content": " loss, which takes around 10 seconds for optimization. All experiments are carried out on an NVIDIA GeForce RTX 4090 24GB GPU." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": "Evaluation Details. For quantitative evaluation and comparison, we test our model on 1,000 images randomly selected from ImageNet test set, and convert each image into SVGs with 500, 2,000 and 4,000 paths respectively. With the predicted SVG paths, we evaluate the reconstruction accuracy of the output SVG with the following 4 metrics: 1) MSE Distance and 2) PSNR to measure the pixel distance between the input image and the rendering from SVG; 3) LPIPS [35] to evaluate the perceptual distance, and 4) SSIM [31] to measure the structural distance. We further compare with Im2Vec [21] on EMOJIS dataset [2]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 212, + 455, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 212, + 455, + 225 + ], + "spans": [ + { + "bbox": [ + 305, + 212, + 455, + 225 + ], + "type": "text", + "content": "4.2. Image to SVG Comparison" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 231, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 547, + 411 + ], + "type": "text", + "content": "State-of-the-art methods. The state-of-the-art methods can be classified into 3 categories: 1) Algorithm-based methods: Potrace [24] employs edge tracing to vectorize binary images. To process color images, we follow Color Trace1 to first quantize color images into different layers and then convert each layer to SVG using Potrace. Adobe Illustrator [1] is a widely-used commercial software which converts an image into SVG through image tracing. 2) Deep-learning-based methods: Im2Vec [21] encodes the target image into latent and predicts the vector paths with LSTM (#suppl.); and 3) Optimization-based methods: DiffVG [17] optimizes path parameters with random initialization and LIVE [20] employs layer-wise optimization to ensure a good vectorization structure. We use the official codes of these methods and default settings for comparison." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 411, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 590 + ], + "type": "text", + "content": "Qualitative Comparison on ImageNet. We compare with the state-of-the-art vectorization methods in reconstruction accuracy on ImageNet. Specifically, we conduct the comparison experiments under path numbers 500, 2,000 and " + }, + { + "bbox": [ + 304, + 411, + 547, + 590 + ], + "type": "inline_equation", + "content": "4,000^2" + }, + { + "bbox": [ + 304, + 411, + 547, + 590 + ], + "type": "text", + "content": ". The qualitative results are shown in Fig. 6. It can be seen that Potrace cannot reconstruct the image well. LIVE loses a lot of details in relatively smooth areas due to its emphasis on regions with substantial color variations. DiffVG and Adobe work better when the path number increases, but they reconstruct fewer details than our SuperSVG. In comparison, our SuperSVG-B reconstructs most of the details with a short inference time. And by optimizing the SVG parameters from SuperSVG-B with only 10 seconds, our SuperSVG-F achieves the best reconstruction accuracy under different SVG path numbers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 591, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 547, + 674 + ], + "type": "text", + "content": "Quantitative Comparison on ImageNet. We further conduct quantitative comparison on 1,000 images randomly sampled from ImageNet [6] dataset (50 images for LIVE due to the extremely long optimization time: each input image takes about 6 GPU hours to optimize under 500 SVG paths). The quantitative results are presented in Tab. 1. Our SuperSVG achieves the best image vectorization results." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 683, + 520, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 683, + 520, + 693 + ], + "spans": [ + { + "bbox": [ + 317, + 683, + 520, + 693 + ], + "type": "text", + "content": "2Since Adobe and Potrace outputs have different number of parameters in each path, we keep their output parameter number the same as ours." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24897" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 71, + 487, + 401 + ], + "blocks": [ + { + "bbox": [ + 109, + 71, + 487, + 401 + ], + "lines": [ + { + "bbox": [ + 109, + 71, + 487, + 401 + ], + "spans": [ + { + "bbox": [ + 109, + 71, + 487, + 401 + ], + "type": "image", + "image_path": "18fc9e563be8472eb8c120edfa49fd705936706691eaf5b89ec0c76d06f66dc6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 403, + 525, + 415 + ], + "lines": [ + { + "bbox": [ + 67, + 403, + 525, + 415 + ], + "spans": [ + { + "bbox": [ + 67, + 403, + 525, + 415 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison with the state-of-the-art methods in image vectorization with different number of SVG paths." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 440, + 287, + 661 + ], + "blocks": [ + { + "bbox": [ + 47, + 418, + 287, + 438 + ], + "lines": [ + { + "bbox": [ + 47, + 418, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 287, + 438 + ], + "type": "text", + "content": "Table 1. Quantitative comparison between the state-of-the-arts and Ours. Bold and underline for the best and the second best results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 440, + 287, + 661 + ], + "lines": [ + { + "bbox": [ + 50, + 440, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 50, + 440, + 287, + 661 + ], + "type": "table", + "html": "
#PathsMethodTime (s) ↓MSE ↓PSNR ↑LPIPS ↓SSIM ↑
500LIVE [20]20,0000.003924.100.44670.7983
DiffVG [17]19.290.006921.420.53190.6671
Adobe [1]0.870.006721.820.55950.6939
Potrace [24]0.980.020817.850.51150.6920
SuperSVG-B (Ours)0.310.004424.800.44520.7687
SuperSVG-F (Ours)10.000.002627.460.35580.8383
2,000LIVE [20]120,0000.002526.980.39940.8431
DiffVG [17]73.650.003625.880.46830.7710
Adobe [1]2.150.003326.230.39610.7229
Potrace [24]3.100.016019.650.43550.6997
SuperSVG-B (Ours)0.710.002427.250.36480.8446
SuperSVG-F (Ours)10.000.001729.120.29310.8828
4,000LIVE [20]300,0000.002426.800.39810.8446
DiffVG [17]140.340.002527.290.38580.8492
Adobe [1]3.120.003525.930.34080.7297
Potrace [24]5.110.011320.680.43800.7374
SuperSVG-B (Ours)1.040.001928.450.31870.8757
SuperSVG-F (Ours)10.000.001429.960.24960.9028
", + "image_path": "a3300ac45e8309cc4cedb3b18f263f2b7461b596729b1098344db555ff5669b5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 670, + 141, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 141, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 141, + 684 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Ablation Study on the Superpixel-based Framework. We first validate the effectiveness of the superpixel-based" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 425, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 592 + ], + "type": "text", + "content": "vectorization framework. We train a model that directly predicts the SVG paths for an input image, without superpixel segmentation. Then, we test the model in two ways: 1) predict the SVG paths for the whole input image and 2) uniformly divide the input image into " + }, + { + "bbox": [ + 304, + 425, + 545, + 592 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 304, + 425, + 545, + 592 + ], + "type": "text", + "content": " blocks and vectorize each block separately. We compare our model with these two models with the same number of paths (1,000). The results are shown in Fig. 7 and Tab. 2(a). The model without superpixel segmentation loses many image details. By introducing the block division, the model enriches the details, but the regions near block boundaries are discontinuous (shown in red box), producing unnatural results. In comparison, our superpixel-based SuperSVG-B reconstructs most details and keeps the image structures well." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "text", + "content": "Ablation Study on the Coarse-stage Model. We then evaluate the effectiveness of the boundary loss " + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{Bound}}" + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "text", + "content": " and the path efficiency loss " + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PE}" + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "text", + "content": " in the coarse-stage model. We train 2 ablated models: 1) the model without " + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{Bound}}" + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "text", + "content": "; and 2) the model without " + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PE}" + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "text", + "content": ", and compare them with our coarse-stage model under 500 SVG paths. In this comparison, we only compare vectorization using the coarse-stage model, without using the refinement model. The results are shown in Fig. 8 and Tab. 2(b). The model without " + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "inline_equation", + "content": "L_{\\text{Bound}}" + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "text", + "content": " predicts some paths that cross superpixel boundaries, re" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24898" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 70, + 261, + 179 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 261, + 179 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 261, + 179 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 261, + 179 + ], + "type": "image", + "image_path": "eb0bd8a4ac218addf3223fbdf0c775c6426e2b4643261beb199062825959c733.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 183, + 287, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 183, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 287, + 216 + ], + "type": "text", + "content": "Figure 7. Ablation study on the superpixel-based image vectorization framework. The ablated models either cannot reconstruct most details or suffer from boundary inconsistency problem." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 71, + 224, + 261, + 341 + ], + "blocks": [ + { + "bbox": [ + 71, + 224, + 261, + 341 + ], + "lines": [ + { + "bbox": [ + 71, + 224, + 261, + 341 + ], + "spans": [ + { + "bbox": [ + 71, + 224, + 261, + 341 + ], + "type": "image", + "image_path": "1dc0cd3060dd4f4aa0b2993cc730c141b811fcbba9e4de89350b92de5f4b7449.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 342, + 287, + 374 + ], + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 374 + ], + "type": "text", + "content": "Figure 8. Ablation study on the coarse-stage model. The ablated models either predict paths crossing superpixel boundaries or reconstruct less details than ours." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 388, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 448 + ], + "type": "text", + "content": "sulting in worse reconstruction. The model without " + }, + { + "bbox": [ + 46, + 388, + 287, + 448 + ], + "type": "inline_equation", + "content": "L_{PE}" + }, + { + "bbox": [ + 46, + 388, + 287, + 448 + ], + "type": "text", + "content": " has a poorer performance, which is validated by the metric results. In comparison, our model outperforms the two ablated models, validating the effectiveness of the losses " + }, + { + "bbox": [ + 46, + 388, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{Bound}}" + }, + { + "bbox": [ + 46, + 388, + 287, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 388, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PE}" + }, + { + "bbox": [ + 46, + 388, + 287, + 448 + ], + "type": "text", + "content": " in the coarse-stage model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 677 + ], + "type": "text", + "content": "Ablation Study on the Refinement-stage Model. Finally, we validate the effectiveness of the refinement stage and the DPW loss. We train 3 ablated models: 1) the model without the refinement stage; 2) the model without the DPW loss " + }, + { + "bbox": [ + 46, + 450, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DPW}" + }, + { + "bbox": [ + 46, + 450, + 287, + 677 + ], + "type": "text", + "content": " (i.e., without coarse paths guidance, with pixel-wise loss only); and 3) the model replacing DPW loss " + }, + { + "bbox": [ + 46, + 450, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DPW}" + }, + { + "bbox": [ + 46, + 450, + 287, + 677 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 450, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 46, + 450, + 287, + 677 + ], + "type": "text", + "content": " loss in path parameter space. The results are shown in Fig. 9 and Tab. 2(c). The model without refinement-stage cannot reconstruct as many details as ours. The ablated model without " + }, + { + "bbox": [ + 46, + 450, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DPW}" + }, + { + "bbox": [ + 46, + 450, + 287, + 677 + ], + "type": "text", + "content": " predicts paths with a very small area or even invisible, as explained in Sec.3.2, thus the results look similar to the coarse-stage results. For the ablated model replacing DPW loss " + }, + { + "bbox": [ + 46, + 450, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DPW}" + }, + { + "bbox": [ + 46, + 450, + 287, + 677 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 450, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 46, + 450, + 287, + 677 + ], + "type": "text", + "content": " in path parameter space, which enforces one-to-one alignment between two paths' control points, since the constraint is too strict, the loss cannot function well in experiments, and the results look alike the results without " + }, + { + "bbox": [ + 46, + 450, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DPW}" + }, + { + "bbox": [ + 46, + 450, + 287, + 677 + ], + "type": "text", + "content": ". In comparison, our model outperforms the ablated models, validating the effectiveness of the refinement and DPW loss." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 331, + 70, + 520, + 179 + ], + "blocks": [ + { + "bbox": [ + 331, + 70, + 520, + 179 + ], + "lines": [ + { + "bbox": [ + 331, + 70, + 520, + 179 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 520, + 179 + ], + "type": "image", + "image_path": "402d531a0b8f73adc39f9178bae87fd012bdf2623cb514abf1919e4ec0b2eba7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 180, + 545, + 212 + ], + "lines": [ + { + "bbox": [ + 305, + 180, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 305, + 180, + 545, + 212 + ], + "type": "text", + "content": "Figure 9. Ablation study on the refinement-stage model. The ablated models either lose details or converge to the poor local optimum described in Sec. 3.2." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 320, + 225, + 533, + 329 + ], + "blocks": [ + { + "bbox": [ + 338, + 213, + 512, + 223 + ], + "lines": [ + { + "bbox": [ + 338, + 213, + 512, + 223 + ], + "spans": [ + { + "bbox": [ + 338, + 213, + 512, + 223 + ], + "type": "text", + "content": "Table 2. Quantitative results of ablation studies." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 225, + 533, + 329 + ], + "lines": [ + { + "bbox": [ + 320, + 225, + 533, + 329 + ], + "spans": [ + { + "bbox": [ + 320, + 225, + 533, + 329 + ], + "type": "table", + "html": "
MethodMSE ↓PSNR ↑LPIPS ↓SSIM ↑
w/o superpixel0.008321.720.50570.6900
4 × 4 blocks0.003829.500.42990.7834
Ours0.003226.040.40750.8111
(a) Ablation on superpixel-based framework.
MethodMSE ↓PSNR ↑LPIPS ↓SSIM ↑
w/o LBound0.053413.380.50410.6210
w/o LPE0.006322.980.48300.7219
Ours (Coarse)0.004524.490.44520.7673
", + "image_path": "9adec7c9fc1f008ae0f5a5ead7ab5df338b9d1236686c411c1544152177c510d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 320, + 353, + 531, + 407 + ], + "blocks": [ + { + "bbox": [ + 318, + 330, + 533, + 351 + ], + "lines": [ + { + "bbox": [ + 318, + 330, + 533, + 351 + ], + "spans": [ + { + "bbox": [ + 318, + 330, + 533, + 351 + ], + "type": "text", + "content": "(b) Ablation on coarse-stage model (results are obtained by the coarse-stage model alone, without refinement model)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 353, + 531, + 407 + ], + "lines": [ + { + "bbox": [ + 320, + 353, + 531, + 407 + ], + "spans": [ + { + "bbox": [ + 320, + 353, + 531, + 407 + ], + "type": "table", + "html": "
MethodMSE ↓PSNR ↑LPIPS ↓SSIM ↑
w/o refine0.004125.020.43750.7770
w/o LDPW0.004524.490.44520.7673
L2 replacing LDPW0.004524.490.44520.7673
Ours0.003226.040.40750.8111
", + "image_path": "05ece90da5f0b3255aa2ac0226430fb847cf447b6bec773c58c6d59b71c99259.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 353, + 408, + 498, + 418 + ], + "lines": [ + { + "bbox": [ + 353, + 408, + 498, + 418 + ], + "spans": [ + { + "bbox": [ + 353, + 408, + 498, + 418 + ], + "type": "text", + "content": "(c) Ablation on refinement-stage model." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 423, + 378, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 423, + 378, + 435 + ], + "spans": [ + { + "bbox": [ + 306, + 423, + 378, + 435 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 443, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 545, + 563 + ], + "type": "text", + "content": "We propose SuperSVG, a novel superpixel-based vectorization model that decomposes a raster image into superpixels and then vectorizes each separately, achieving fast and accurate image vectorization. We propose a two-stage self-teaching framework, where a coarse-stage model reconstructs main structure and a refinement model enriches details, with a novel dynamic path warping loss that guides the refinement model by inheriting knowledge from coarse paths. Extensive experiments demonstrate that SuperSVG achieves the state-of-the-art performance on vectorization." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 574, + 403, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 403, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 403, + 586 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "This work was supported by National Natural Science Foundation of China (62302297, 72192821, 62272447), Shanghai Sailing Program (22YF1420300), Young Elite Scientists Sponsorship Program by CAST (2022QNRC001), Beijing Natural Science Foundation (L222117), the Fundamental Research Funds for the Central Universities (YG2023QNB17), Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), Shanghai Science and Technology Commission (21511101200)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": "3Since the model without refinement does not contain refinement paths, we increase the number of coarse paths to keep the path number consistent." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24899" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Adobe Illustrator. https://www.adobe.com/products/illustrator.html.2,6,7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "type": "text", + "content": "[2] Noto Emoji. https://github.com/googlefonts/ noto-emoji.6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 137, + 286, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 286, + 190 + ], + "type": "text", + "content": "[3] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and Sabine Susstrunk. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 2274-2282, 2012. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 192, + 286, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 286, + 234 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 286, + 234 + ], + "type": "text", + "content": "[4] Marco Cuturi and Mathieu Blondel. Soft-DTW: a differentiable loss function for time-series. In International conference on machine learning, pages 894-903. PMLR, 2017. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 286, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 286, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 286, + 280 + ], + "type": "text", + "content": "[5] Wen Dai, Tao Luo, and Jianbing Shen. Automatic image vectorization using superpixels and random walkers. In 2013 6th International Congress on Image and Signal Processing (CISP), 2013. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 281, + 286, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 286, + 334 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 286, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 286, + 400 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 286, + 400 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth " + }, + { + "bbox": [ + 53, + 335, + 286, + 400 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 53, + 335, + 286, + 400 + ], + "type": "text", + "content": " words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 402, + 286, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 286, + 434 + ], + "type": "text", + "content": "[8] Maria Dziuba, Ivan Jarsky, Valeria Efimova, and Andrey Filchenkov. Image vectorization: a review. arXiv preprint arXiv:2306.06441, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 436, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 286, + 479 + ], + "type": "text", + "content": "[9] Jun Gao, Chengcheng Tang, Vignesh Ganapathi-Subramanian, Jiahui Huang, Hao Su, and LeonidasJ. Guibas. DeepSpline: Data-driven reconstruction of parametric curves and surfaces. 2019. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 480, + 286, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 286, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 286, + 502 + ], + "type": "text", + "content": "[10] David Ha and Douglas Eck. A neural representation of sketch drawings. Learning, Learning, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 502, + 286, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 286, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 286, + 546 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 547, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 286, + 612 + ], + "type": "text", + "content": "[12] Teng Hu, Ran Yi, Haokun Zhu, Liang Liu, Jinlong Peng, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. Stroke-based neural painting and stylization with dynamically predicted painting region. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7470-7480, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 286, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 646 + ], + "type": "text", + "content": "[13] Varun Jampani, Deqing Sun, Ming-Yu Liu, Ming-Hsuan Yang, and Jan Kautz. Superpixel sampling networks. arXiv: Computer Vision and Pattern Recognition, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 286, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 669 + ], + "type": "text", + "content": "[14] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[15] Yu-Kun Lai, Shi-Min Hu, and Ralph R. Martin. Automatic and topology-preserving gradient mesh generation for image vectorization. ACM Transactions on Graphics, 28(3):1-8, 2009. 1, 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[16] A. Levinshtein, A. Stere, K.N. Kutulakos, D.J. Fleet, S.J. Dickinson, and K. Siddiqi. TurboPixels: Fast superpixels using geometric flows. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 2290-2297, 2009. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "type": "text", + "content": "[17] Tzu-Mao Li, Michal Lukáč, Michael Gharbi, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for editing and learning. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 165, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 165, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 165, + 545, + 208 + ], + "type": "text", + "content": "[18] Zicheng Liao, H. Hoppe, D. Forsyth, and Yizhou Yu. A subdivision-based representation for vector image editing. IEEE Transactions on Visualization and Computer Graphics, 18(11):1858-1867, 2012. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 210, + 545, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 210, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 308, + 210, + 545, + 254 + ], + "type": "text", + "content": "[19] Chen Liu, Jiajun Wu, Pushmeet Kohli, and Yasutaka Furukawa. Raster-to-vector: Revisiting floorplan transformation. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 256, + 545, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 256, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 308, + 256, + 545, + 311 + ], + "type": "text", + "content": "[20] Xu Ma, Yuqian Zhou, Xingqian Xu, Bin Sun, Valerii Filev, Nikita Orlov, Yun Fu, and Humphrey Shi. Towards layerwise image vectorization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16314-16323, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 312, + 545, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 312, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 308, + 312, + 545, + 367 + ], + "type": "text", + "content": "[21] Pradyumna Reddy, Michael Gharbi, Michal Lukac, and Niloy J Mitra. Im2Vec: Synthesizing vector graphics without vector supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7342-7351, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 369, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 369, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 308, + 369, + 545, + 422 + ], + "type": "text", + "content": "[22] Leo Sampaio Ferraz Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based representation for sketched structure. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 426, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 458 + ], + "type": "text", + "content": "[23] Mike Schuster and Kuldip K Paliwal. Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing, 45(11):2673-2681, 1997. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 460, + 545, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 482 + ], + "type": "text", + "content": "[24] Peter Selinger. Potrace: a polygon-based tracing algorithm, 2003. 1, 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 484, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 484, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 308, + 484, + 545, + 517 + ], + "type": "text", + "content": "[25] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 519, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 519, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 308, + 519, + 545, + 563 + ], + "type": "text", + "content": "[26] Bohao Tang, Teng Hu, Yuzhen Du, Ran Yu, and Lizhuang Ma. Curved-stroke-based neural painting and stylization through thin plate spline interpolation. Scientia Sinica Informationis, 54(2):301-315, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 565, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 565, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 308, + 565, + 545, + 597 + ], + "type": "text", + "content": "[27] Yizhe Tang, Yue Wang, Teng Hu, Xin Tan, and Ran Yi. AttnPainter: A scalable stroke predictor for any scene painting. arXiv preprint, 2024. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 600, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 600, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 600, + 545, + 632 + ], + "type": "text", + "content": "[28] Sebastian Thrun and James Diebel. Bayesian image vectorization: the probabilistic inversion of vector image rasterization. 2008. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 634, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 634, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 634, + 545, + 689 + ], + "type": "text", + "content": "[29] Wei-Chih Tu, Ming-Yu Liu, Varun Jampani, Deqing Sun, Shao-Yi Chien, Ming-Hsuan Yang, and Jan Kautz. Learning superpixels with segmentation-aware affinity loss. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 568-576, 2018. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "text", + "content": "[30] Michael Van den Bergh, Xavier Boix, Gemma Roig, Benjamin De Capitani, and Luc Van Gool. SEEDS: Superpixels" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24900" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 487 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "content": "extracted via energy-driven sampling. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision (ECCV), pages 13-26. Springer, 2012. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[31] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 194 + ], + "type": "text", + "content": "[32] Guofu Xie, Xin Sun, Xin Tong, and Derek Nowrouzezahrai. Hierarchical diffusion curves for accurate automatic image vectorization. ACM Transactions on Graphics, 33(6):1-11, 2014. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "type": "text", + "content": "[33] Ming Yang, Hongyang Chao, Chi Zhang, Jun Guo, Lu Yuan, and Jian Sun. Effective clipart image vectorization through direct optimization of bezigons. IEEE Transactions on Visualization and Computer Graphics, page 1063-1075, 2016. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 252, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 252, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 252, + 287, + 306 + ], + "type": "text", + "content": "[34] Zipeng Ye, Ran Yi, Minjing Yu, Yong-Jin Liu, and Ying He. Fast computation of content-sensitive superpixels and supervoxels using q-distances. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3770-3779, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 363 + ], + "type": "text", + "content": "[35] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 397 + ], + "type": "text", + "content": "[36] Shuang Zhao, Frédo Durand, and Changxi Zheng. Inverse diffusion curves using shape optimization. IEEE Transactions on Visualization and Computer Graphics, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "type": "text", + "content": "[37] Hailing Zhou, Jianmin Zheng, and Lei Wei. Representing images using curvilinear feature driven subdivision surfaces. IEEE Transactions on Image Processing, 23(8):3268-3280, 2014. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 443, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 443, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 443, + 287, + 487 + ], + "type": "text", + "content": "[38] Haokun Zhu, Juang Ian Chong, Teng Hu, Ran Yi, Yu-Kun Lai, and Paul L Rosin. SAMVG: A multi-stage image vectorization model with the segment-anything model. arXiv preprint arXiv:2311.05276, 2023. 1, 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24901" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Supervised Anomaly Detection for Complex Industrial Images/ef358e69-3d45-4c6c-aa99-f5a5ceb11e08_content_list.json b/2024/Supervised Anomaly Detection for Complex Industrial Images/ef358e69-3d45-4c6c-aa99-f5a5ceb11e08_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3c7589bf7e5b007009cb20510d52acd2775d532f --- /dev/null +++ b/2024/Supervised Anomaly Detection for Complex Industrial Images/ef358e69-3d45-4c6c-aa99-f5a5ceb11e08_content_list.json @@ -0,0 +1,1496 @@ +[ + { + "type": "text", + "text": "Supervised Anomaly Detection for Complex Industrial Images", + "text_level": 1, + "bbox": [ + 169, + 130, + 799, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aimira Baitieva1, David Hurych2, Victor Besnier2, Olivier Bernard1", + "bbox": [ + 215, + 179, + 750, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ valeo, $^{2}$ valeo.ai", + "bbox": [ + 416, + 214, + 550, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{aimira.baitieva, david.hurych, victor.besnier, olivier.bernard}@valeo.com", + "bbox": [ + 155, + 236, + 808, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Automating visual inspection in industrial production lines is essential for increasing product quality across various industries. Anomaly detection (AD) methods serve as robust tools for this purpose. However, existing public datasets primarily consist of images without anomalies, limiting the practical application of AD methods in production settings. To address this challenge, we present (1) the Valeo Anomaly Dataset (VAD), a novel real-world industrial dataset comprising 5000 images, including 2000 instances of challenging real defects across more than 20 subclasses. Acknowledging that traditional AD methods struggle with this dataset, we introduce (2) Segmentation-based Anomaly Detector (SegAD). First, SegAD leverages anomaly maps as well as segmentation maps to compute local statistics. Next, SegAD uses these statistics and an optional supervised classifier score as input features for a Boosted Random Forest (BRF) classifier, yielding the final anomaly score. Our SegAD achieves state-of-the-art performance on both VAD (+2.1% AUROC) and the VisA dataset (+0.4% AUROC). The code and the models are publicly available1.", + "bbox": [ + 76, + 318, + 473, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 666, + 209, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Within the manufacturing process, industrial visual inspection plays a crucial role in identifying defects in produced components. This operation holds significant importance in minimizing costs by identifying and removing faulty parts early in the production stages and, more importantly, in preventing the dispatch of defective components to customers. Traditionally, this task has relied on human operators; however, the likelihood of overlooking certain defects can be as high as $25\\%$ for specific defect types [9].", + "bbox": [ + 75, + 691, + 468, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When the inspected product comprises numerous components, its examination may create a bottleneck in the production process, causing delays across the entire line. While", + "bbox": [ + 76, + 829, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "conventional computer vision methods applied to this task demonstrate superior speed and lower error rates compared to human operators [16], their inflexibility and lack of satisfactory accuracy [15] limit their effectiveness.", + "bbox": [ + 500, + 287, + 890, + 348 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Industrial deep learning anomaly detection has been an active research field in recent years. Most of anomaly detection methods use only good images for training and try to detect deviations from the training data [1, 7, 23, 29]. Development of the new methods is restrained by publicly available datasets. Recent industrial anomaly detection datasets typically contain approximately one hundred (or even fewer) abnormal images, showcasing defects in the testing set only [2, 32]. This poses a challenge for supervised anomaly detection methods aiming to utilize both normal and defective parts during training. Supervised models often undergo training with just ten abnormal images from the testing set, resulting in overfitting and reduced sensitivity to previously unseen defects [28].", + "bbox": [ + 496, + 351, + 892, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In response to the limitations of current datasets, we introduce and publicly release VAD (Valeo Anomaly Dataset), which contains 1000 bad and 2000 good parts in the training set and 1000 bad and 1000 good parts in the test set, see Fig. 1(a). Among the defective parts used for testing, 165 contain specific defect types not present in the training data; these parts are explicitly labeled in the released dataset. All images in VAD are captured from an actual production line, showcasing a diverse range of defects, from highly obvious to extremely subtle. This dataset bridges the gap between the academic community and the industry, offering researchers the chance to advance the performance of methods in tackling more intricate real-world challenges.", + "bbox": [ + 496, + 565, + 893, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current approaches to supervised anomaly detection either yield unsatisfactory results; see Tab. 3, or demand pixel-level labels for defective parts, as seen in [28, 31]. Consequently, we introduce a novel method named SegAD (Segmentation-based Anomaly Detector), which is described in Fig. 1(b). The employed approach eliminates the need for pixel-level labels, requiring only a flag for each image. Anomaly map scores from each segment are used", + "bbox": [ + 496, + 779, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/abc-125/segad", + "bbox": [ + 94, + 886, + 357, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17754", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/993389f2b9e558e13cbe5c56ca4f5b671efb80b9d3662e8a3e743b2ade7ee9ba.jpg", + "image_caption": [ + "(a)", + "(b)", + "Figure 1. Overview of our contributions. (a) VAD, a real-world industrial dataset designed for supervised anomaly detection with complex defects. (b) SegAD, our method that leverages anomaly maps extracted from segmented outputs of one or more anomaly detectors. Higher-level statistical features are computed on these maps, such as skewness, kurtosis, or mean, to generate the local anomaly features. Additionally, our SegAD provides the flexibility to use the output of a supervised classifier score with the local anomaly features, creating input for a final Boosted Random Forest (BRF) classifier that yields the final score." + ], + "image_footnote": [], + "bbox": [ + 106, + 94, + 872, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to delineate the parameters of the distribution, emphasizing relatively low scores in important areas while disregarding higher scores in noisy regions. In this context, the precise position and appearance of the anomaly diminishes in importance as long as it resides within a designated segment. Our main contributions can be summarized as follows:", + "bbox": [ + 75, + 411, + 467, + 500 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a supervised anomaly detection dataset with complex objects and a large variety of defects. As an addition, we establish a one-class anomaly detection benchmark, which is more challenging than most of the current ones, a supervised benchmark with a high number of bad images for training (1000 images) and a supervised benchmark with a low number of bad images for training (100 images). We share the dataset freely with the research community.", + "- We create a novel supervised anomaly detection method called SegAD. This approach performs noticeably better than recent anomaly detectors alone while retaining the ability to detect unknown defects. SegAD reaches SOTA results on VAD as well as on the established VisA dataset [32]." + ], + "bbox": [ + 76, + 503, + 467, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 746, + 218, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first discuss existing datasets and explain how VAD is different. Next, we give a quick overview of existing anomaly detection methods. We show in Table 1 a global comparison between our VAD and existing datasets.", + "bbox": [ + 75, + 773, + 467, + 833 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 845, + 178, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In 2019, the MVTec AD dataset [2] was introduced. This dataset significantly advanced anomaly detection methods", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "by providing images with authentic defects and realistic objects, in contrast to earlier datasets predominantly comprising textures and simple defects [4, 26, 27]. MVtec LOCO AD [3] highlighted the challenge of logical defects, such as missing or misplaced parts, necessitating anomaly detection methods capable of capturing the global context of the image. The more recent VisA dataset [32] expands on this by incorporating diverse objects with complex structures, multiple instances, and variations in scale.", + "bbox": [ + 496, + 410, + 890, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing datasets often contain simulated defects [2, 3, 32], creating a domain gap between research and practical applications. This complicates the deployment of anomaly detection methods in real-world settings. To promote supervised anomaly detection, we propose the Valeo Anomaly Dataset (VAD), a real-world industrial dataset with various defects, including logical ones.", + "bbox": [ + 496, + 547, + 890, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another notable issue stems from the saturation of solved datasets, where other methods are influenced more by dataset-specific design choices than general applicability. Compared to this, our dataset introduces a broader range of challenging and diverse anomalies and high intra-class variability of good images.", + "bbox": [ + 496, + 654, + 890, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Methods", + "text_level": 1, + "bbox": [ + 500, + 755, + 602, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To define the terms we are working with, \"one-class\" methods are commonly called unsupervised in the context of anomaly detection [10, 15, 29]. We use \"one-class\" as a more accurate name since these methods use labeled images for training. It is important to distinguish such methods from ones that work with unlabelled data, as in [17]. Models that use bad and good images for training are usually called \"supervised\", which is the name we also adopt.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17755", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/a775f8af6adf8393a567976ff47e98ded84b505b7520616782728c004a1beada.jpg", + "table_caption": [ + "Table 1. Comparison of different anomaly detection datasets. Tr. good and Tr. bad show the average number of images per class for training. tex = texture, obj = object, Log. def. = Logical defects, Pix. labels = pixel-level labels and wl = weakly labelled (ellipse around the defect)." + ], + "table_footnote": [], + "table_body": "
DatasetYearClassesImagesTypesDefectsTr. goodTr. badLog. def.Pix. labels
DAGM [27]20071011500texsynthetic50075nowl
KSSD [26]20191399texreal-world11517noyes
MVTec AD [2]2019155354tex, objsimulated2420noyes
KSSD2 [4]202113335texreal-world2085246noyes
BTAD [19]202132830tex, objreal-world6000noyes
LOCO AD [3]202253644objsimulated3540yesyes
VisA [32]20221210821objsimulated7200noyes
VAD (ours)202315000objreal-world20001000yesno
", + "bbox": [ + 76, + 141, + 893, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One-class Anomaly Detection. PatchCore [23] relies on a pretrained feature extractor model to extract features from the training set into a memory bank and reduce the size of a memory bank using coreset subsampling. Features extracted from the new input are compared to their nearest neighbors in the memory bank. FastFlow [29] uses a similar feature extractor model and maps extracted features to the Gaussian distribution using normalizing flow [22]. RD4AD [7] utilizes a teacher-student architecture, which combines features from several layers of feature extractor to eliminate redundant ones. EfficientAD [1] introduces many innovations, including using an autoencoder to detect logical defects as an addition to a classic teacher-student architecture and their own pretrained feature extractor, which imitates the behavior of a bigger model with a noticeably lesser inference time.", + "bbox": [ + 75, + 311, + 472, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Supervised Anomaly Detection. In theory, leveraging defective parts for training is advantageous for refining class boundaries [24]. However, practical challenges arise, such as the small size of anomalies and the impossibility of collecting all potential defects [31]. Several recent anomaly detection methods use both good and bad images [21, 28, 31], but some of them overfit to seen defects, and others require pixel masks for defects to calculate loss or to generate new defects, which can be problematic in a real-world case. Real-world datasets (as VAD) might contain logical defects that cannot be generated by existing defects generation strategies, which usually cut and paste existing defects with various modifications [28, 31], which is not going to improve results for a wrong wire shape as or other similar types of defects.", + "bbox": [ + 75, + 561, + 472, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Very few supervised anomaly detection benchmarks are available; the most popular is Supervised Anomaly Detection on MVTec AD. DevNet [21] is one of the first supervised anomaly detection models that tried to solve industrial MVTec AD dataset, compared to earlier methods which were mostly addressing Out-of-Distribution problem [11, 24] using non-industrial datasets such as MNIST [14]", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "or CIFAR-10 [13]. DRA [8] uses several heads to learn both seen and pseudo anomalies, as well as normal examples, which should reduce overfitting on seen anomalies. These methods perform well on MVtec AD but might fail on more complex problems, as shown in Tab. 3 and Tab. 5. Compared to modern supervised anomaly detection methods, SegAD (ours) performs better than current SOTA methods, even on complex problems without pixel-level labels, extensive augmentation, and long training.", + "bbox": [ + 496, + 311, + 893, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Valeo Anomaly Dataset (VAD)", + "text_level": 1, + "bbox": [ + 498, + 465, + 779, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VAD consists of one class with predefined training and testing sets. The training set contains 1000 bad and 2000 good images, and the testing set contains 1000 bad, 165 of them are unseen bad, and 1000 good images. Unseen bad images in the test dataset refer to several rare defect types that are not present in the training data, several examples of which are shown in Fig. 4. Having such images is important to avoid turning this anomaly detection problem into a supervised classification problem. Some images might have a thin black border at the bottom due to how they were filmed, an example can be seen in Fig. 2. Defects might occur in the whole area of the image. Image-level annotation is provided, but there is no pixel-level annotation due to the complexity of defects and the fact there is no exact position for missing or misplaced components. All images are $512 \\times 512$ pixels in PNG format. Examples of all types of defects can be found in Appendix C.", + "bbox": [ + 496, + 491, + 893, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Good parts consist of two wires connected to two pins on one side and two soldering dots on the other, placed on a piezo. Piezo means piezoelectric element, a big round area under the other components. Wires should have the right shape, not too straight, too long, or bent too much. Wires should connect the soldering dots as close to their centers as possible. Some amount of deviation is allowed, but if the wire is too close to the side of a solder dot, it is a defect. Soldering should be well-rounded, not too small, and placed in the correct location, not overlapping with the", + "bbox": [ + 496, + 750, + 895, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17756", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d3105d75fc2c8b7cb52aa9267bc13e366d596b1b232487a26966f68dfde09f67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 88, + 259, + 222 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/37a362d3d40e4339cdb5a9c8d810d436ffd2deb4ff3d73e53db6244eca149be3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 88, + 434, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b551e15ba0358def0273f3b67e614fb5d83867b6726c043a4ab65fc498a63a11.jpg", + "image_caption": [ + "Figure 2. Good parts. Small scratches on the piezo are allowed, as well as wire being closer to the side of the soldering. (VAD)" + ], + "image_footnote": [], + "bbox": [ + 86, + 224, + 259, + 357 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f451ca90a852a8c1812aab828cdfd6d7cd2e49edb845d9198f7bbbcd8b58dcf1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 224, + 434, + 357 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/595e8758846c627d9045bb394391f9ac01e448ede99dcb7f376de17dccee8e28.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 441, + 269, + 507 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/aaf296721c7a7c1d47b3de0bc562f81bc4297747d948a8cd2f52023ccb4d6da5.jpg", + "image_caption": [ + "Logical defects", + "Figure 3. Variety of defects: logical, from the top left: wire on the side, wire out of solder dot, missing wire, bad wire shape (bent too much). Structural: soldering paste pollution, crack on a piezo, burned area under the solder dot, broken piezo. (VAD)" + ], + "image_footnote": [], + "bbox": [ + 94, + 508, + 267, + 571 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/04fab07105a0f1a51830e5bd15f8fed0185abd6e05f19e935939e177f395a8ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 441, + 452, + 507 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e27f961208d89d5fa335c0ded39c876204df4177a84e7abfe5ba808693a4298e.jpg", + "image_caption": [ + "Structural defects" + ], + "image_footnote": [], + "bbox": [ + 292, + 508, + 450, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "piezo contours. The wire should be properly connected to the pin without any pieces of it visible on the side of the pin. The piezo can have different textures. Any kind of pollution is considered to be an anomaly, which makes the part defective.", + "bbox": [ + 75, + 700, + 468, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Logical defects result from components being misplaced or misshapen rather than being damaged, e.g. bad wire shape, bad soldering shape, wrong wire position, wrong soldering position, and missing parts. Structural defects include various cracks and broken parts of the piezo, as well as a variety of burns and pollution. For both types of defects, see Fig. 3. The importance of logical defects is often overlooked in the existing datasets (excluding MVTec", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4bdb7c339de780f876a09b1a68620c6a1c8aadedf62f2bd1e3ed6978ede888c2.jpg", + "image_caption": [ + "Figure 4. Unseen defects for left to right: burned solder dot, burned wire, extra unnecessary wire. (VAD)" + ], + "image_footnote": [], + "bbox": [ + 517, + 94, + 642, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/53c3be7cf04fb199446ce0b90d8591a07cbc5fe7ebb1f183b67607b7ad3c31cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 95, + 746, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1b815771df7936ec659412c3c916e7270a94a4c06f5193737f11f26d75326185.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 95, + 875, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LOCO AD) because they are hard to imitate if the dataset is filmed in the lab, but also hard to detect. VAD provides the research community the opportunity to work with real logical constraints of a complex object.", + "bbox": [ + 498, + 251, + 890, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.Benchmarks", + "text_level": 1, + "bbox": [ + 500, + 320, + 633, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We establish three benchmarks for the VAD dataset. The first is one-class anomaly detection. As shown in Tab 2, even SOTA methods struggle to perform well on our dataset compared to other popular datasets, which leaves an opening for further research. For this benchmark, no bad images are used for training. The second benchmark is high-shot supervised anomaly detection, which uses all 1000 bad images from the training set. The last benchmark is low-shot supervised anomaly detection. Only 100 bad images from the training set are used. These images are randomly selected using 5 seeds, with the average result over all seeds being expected. The same 2000 good parts are used for the training in all benchmarks.", + "bbox": [ + 498, + 343, + 890, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 551, + 589, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SegAD works with image data at its input, and after inference, it is expected to output a score that describes a probability of anomaly appearance. Let $\\mathcal{I} = \\left[0,\\dots ,255\\right]^{W,H,3}$ be a space of all images of width $W$ , height $H$ and with 3 color channels. The $i$ -th image from a dataset is marked as $I^{(i)}\\in \\mathcal{I}$ . We denote $\\mathbf{x}$ to be a 2D pixel coordinate vector. SegAD inference consists of three consecutive stages in a pipeline. First is anomaly map(s) calculation with optional supervised classifier score calculation. Next is the calculation of simple statistics from anomaly maps over individual segments. Lastly, the obtained statistics, and optionally also the classifier score, are used as input features for a BRF classifier [18] that delivers the final result.", + "bbox": [ + 498, + 577, + 890, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let us denote $K$ as the number of used pixel-wise anomaly detectors and mark them as functions $f_{k}(I):\\mathcal{I}\\to \\mathbb{R}^{W,H},k\\in \\{1,\\ldots ,K\\}$ . A supervised classifier is a function $g(I):\\mathbb{R}^{W,H}\\to \\mathbb{R}$ . Next, we introduce a set of $L$ mutually exclusive segments $S = \\{s_1,\\dots,s_L\\}$ , where $s_l\\in [0,1]^{W,H}$ .", + "bbox": [ + 498, + 776, + 890, + 871 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the calculation of anomaly maps for all $f_{k}(I)$ , the masks $s_l$ are used for segment-wise", + "bbox": [ + 500, + 871, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17757", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "extraction of results, followed by the calculation of several statistics. The final feature vector $\\mathbf{f} = [g(I),[\\mathbf{q}^L ]_{k = 1}^K,[\\mathbf{z}^L ]_{k = 1}^K,[\\mathbf{c}^L ]_{k = 1}^K,[\\mathbf{m}^L ]_{k = 1}^K]$ extracted from one image, is the concatenation of 4 vectors containing different statistics. $\\mathbf{q}_k^L$ stands for a vector of $L$ numbers where each stands for a $99.5\\%$ quantile from anomaly map $f_{k}(I)$ over pixels where $s_l(\\mathbf{x}) = 1$ Similarly, for every anomaly map $f_{k}(I)$ we calculate a vector for skew $\\mathbf{z}_k^L$ , kurtosis $\\mathbf{c}_k^L$ and mean $\\mathbf{m}_k^L$ . The length of the feature vector $\\mathbf{f}$ equals to $K*L*4 + 1$", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For notation simplicity, we keep the supervised classifier score $g(I)$ always present in $\\mathbf{f}$ , but it may be omitted, which is clearly defined in every experiment setup. Feature vectors $\\mathbf{f}^{(i)}$ are used as inputs for training the final classifier [18], which, at inference time, delivers the final result.", + "bbox": [ + 76, + 242, + 470, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Training", + "text_level": 1, + "bbox": [ + 76, + 328, + 181, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training SegAD requires two separate subsets from the training dataset. The first subset is used to train base models $f_{k}$ , the second subset is used to train the BRF classifier. To create these subsets from one training dataset, we split available good images into two parts. For VAD, we report average results over 5 different splits to show more reliable results.", + "bbox": [ + 75, + 352, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each model $(f_{k}, g, \\mathrm{BRF})$ is trained independently. Training can be separated into two stages. In the first stage, anomaly detector(s) $f_{k}$ are trained using the same data. In the same stage, a segmentation map $S$ must be defined. As will be elaborated in more detail in Sec. 5, the segmentation map can be static or produced by a segmentation model, depending on the dataset. A supervised classifier $g$ is trained separately on $80\\%$ of the full training set (1600 good and 800 bad parts for VAD) to ensure that the training sets for this model and SegAD are not fully overlapping. The next stage is to train a final BRF classifier in SegAD, which requires both good and bad images, bad images can be replaced with artificial defects in some cases, as shown in Tab. 2. BRF is trained with the output of the models from the previous stage.", + "bbox": [ + 75, + 458, + 468, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An important thing to note is that SegAD is very fast to train as long as you already have other models available. Extraction of features and training the final classifier takes noticeably less time than training anomaly detectors. The same is true for the inference; SegAD adds a negligible time to processing the image while improving results significantly.", + "bbox": [ + 75, + 685, + 468, + 791 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Setups for SegAD", + "text_level": 1, + "bbox": [ + 76, + 801, + 248, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SegAD can use any anomaly detector(s) $f_{k}$ as long as it produces an anomaly map showing pixel-level anomaly scores, although, in this paper, we use only one-class anomaly detectors because they require less data to train. Any supervised classifier $g$ that returns a classification score can be", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "used. In experiments, we use several setups, which will be described here in detail. SegAD is denoted as Ours. Supervised classifier $g$ is not used unless stated otherwise.", + "bbox": [ + 496, + 90, + 890, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Anomaly Detector + Ours means $K = 1$ , $f_{1} =$ \"Anomaly Detector name\". This setup allows us to show improvement for one anomaly detection model in particular and demonstrate that this improvement can be achieved for different types of anomaly detectors.", + "bbox": [ + 496, + 136, + 890, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All $\\mathbf{AD} + \\mathbf{Ours}$ denotes that all anomaly detectors from the list $AllAD$ of length $N$ are used to produce anomaly maps. Such a list can be found in the caption of the table. In that case $K = N$ , $f_{1} = AllAD_{1},\\ldots ,f_{N} = AllAD_{N}$ . Anomaly detectors might use an ensemble of feature extractors, similarly to [23], which makes them slower but can improve the result. In our case, anomaly detectors themselves can be successfully assembled in a supervised way.", + "bbox": [ + 496, + 212, + 890, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All AD + Supervised Classifier + Ours denotes a setup in which $g =$ \"Supervised Classifier name\", and the rest is the same as in the previous setup. A supervised classifier allows us to detect seen defects, which can be invisible to a one-class classifier, similarly to [25]. Such a setup shows the best result for academic purposes, but the speed of inference can be unsatisfying for real-world applications. We expect that, in that case, the number and selection of anomaly detectors can be optimized to fit the problem as well as the time constraints.", + "bbox": [ + 496, + 333, + 890, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Anomaly Detector + Supervised Classifier + Ours mean a setup in which $g =$ \"Supervised Classifier name\", $K = 1$ , $f_{1} =$ \"Anomaly Detector name\". It shows an example of an optimization mentioned in the previous setup. Only one anomaly detector gives a higher inference speed compared to the ensemble of anomaly detectors while giving a satisfying result.", + "bbox": [ + 496, + 484, + 890, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 603, + 632, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. We compare results using the following metrics: (1) Area Under the Receiver Operating Characteristic Curve (AUROC) [5], (2) False Positive Rate (FPR) at $95\\%$ True Positive Rate (TPR) denoted as FPR@95TPR. Classification AUROC, which describes the performance on the image level for multiple thresholds, denoted by Cl. AUROC. FPR@95TPR shows the percentage of misclassified good parts (false positive) at $95\\%$ of bad parts classified correctly (true positive). We report mean results $\\pm$ standard deviation.", + "bbox": [ + 496, + 628, + 890, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation and evaluation details. SegAD uses the output of one-class detectors PatchCore [23], FastFlow [29], Reverse Distillation [7] (referred to as RD4AD), EfficientAD [1] either alone or several anomaly detectors at once. SegAD results are compared to the results of these models alone, as well as supervised anomaly detectors DevNet [21] and DRA [8]. We also compare our results to the supervised classifier Wide ResNet50 [30] (denoted as", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17758", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "WRN). The final classifier BRF was trained with the same parameters for all datasets using XGBoost library [6]. More details on the implementation can be found in Appendixes A and B.", + "bbox": [ + 75, + 90, + 468, + 150 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Images are resized to $\\mathrm{W} = \\mathrm{H} = 256$ pixels to make the comparison less biased. The DRA model originally uses a resolution $448 \\times 448$ , which shows better results on VAD than $256 \\times 256$ , 92.8 CI. AUROC, which is still lower than the top results. The higher resolution also improves results for some of the other models, so we do not use it. $256 \\times 256$ is sufficient for defects to remain visible, but it also allows to have a reasonable speed of inference, especially for anomaly detection models.", + "bbox": [ + 75, + 152, + 467, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For VAD, results averaged for 5 seeds and training runs are reported. These seeds are used for training and splitting the training set for SegAD into two sets, as described in Subsection 4.1. 2000 good parts from the train set were split in half: 1000 for training the base model(s) and 1000 for training SegAD. 0, 1000, or 100 bad parts (depending on the benchmark) were also used to train SegAD. Details on the VisA benchmark can be found in Subsection 5.4. No augmentations were used for training, and existing augmentations were removed from used methods. Augmentations tend to be suitable only for specific tasks (as rotation in DRA or DevNet, removing it improved results by these models on the VAD drastically), and using augmentations makes it hard to compare models themselves.", + "bbox": [ + 75, + 287, + 467, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We include three benchmarks for the new VAD and one additional for the VisA dataset [32]. We compare results for VAD for one-class, high-shot, and low-shot supervised benchmarks, as described in Subsection 3.1. The benchmark for the VisA dataset is explained in 5.4. The segmen", + "bbox": [ + 75, + 500, + 467, + 575 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/abc22acaf126607a77657c5ce1b4aea02cd5a88f9c5bedd859b46e19176465f4.jpg", + "image_caption": [ + "Figure 5. Static segmentation map for objects in VAD dataset, overlayed over the image. The image is separated into $L = 7$ segments: background, outer half-circle, piezo border, solder dots area, wires area, pins area, and piezo in the middle." + ], + "image_footnote": [], + "bbox": [ + 156, + 589, + 390, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tation maps for VAD is the same static image for every input image from the dataset. The segmentation map can be seen in Fig. 5. Due to the fact that the object in the VAD is al", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ways centered with the constant size and component placement, this is the simplest choice, which may be sub-optimal, but serves the purpose and does not require any manual or model-inferred annotation of the data. It was created to separate the anomaly maps into several meaningful parts, such as piezo border, wire areas, soldering area, etc., which can contain different defects and different levels of falsely high scores on the anomaly map. Another source of segmentation maps can be a specially trained segmentation model or a zero-shot segmentation model.", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. VAD, one-class benchmark", + "text_level": 1, + "bbox": [ + 498, + 257, + 741, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This benchmark allows only good images for training, but SegAD requires bad images as well, so we try to replace them with artificially generated defects. Because SegAD works with anomaly maps, generating realistic defects is unnecessary. For this reason, we have applied several simple augmentations to the 1000 good images available for training SegAD. These augmentations include Gaussian blur or a randomly placed, randomly sized thin rectangle of a random shade of gray. They were applied to randomly selected segments or random parts of the segments; segments are defined with the segmentation map. Code for generating defects can be found in the SegAD GitHub repository.", + "bbox": [ + 496, + 282, + 890, + 464 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/728cdf57302ac6c3914e9c492b2074778ec529907b7b9a1e1274d480d6215740.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCl. AUROC ↑FPR@95TPR ↓
One-class Anomaly Detection
FastFlow79.1±1.776.2±1.9
RD4AD84.5±0.265.8±2
PatchCore88.3±0.360.4±2.8
EfficientAD89.1±0.844.5±2.2
SegAD (Ours)
FastFlow + Ours83.3±1.371.5±5
RD4AD + Ours87.5±1.261.0±7.9
PatchCore + Ours89.7±0.858.8±3.7
EfficientAD + Ours85.6±1.6100±0
All AD + Ours90.4±0.552.4±4.5
", + "bbox": [ + 500, + 479, + 890, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. One-class benchmark (VAD). The best result is marked in bold. All AD means PatchCore, FastFlow, and RD4AD.", + "bbox": [ + 500, + 700, + 890, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results in Table 2 show that even such a naive approach improves results compared to anomaly detection methods alone. Even greater improvement can be achieved by using several anomaly detectors to produce anomaly maps for SegAD. This strategy fails for EfficientAD + Ours, anomaly maps produced by EfficientAD for the generated defects are too different from anomaly maps for real defects, creating a large discrepancy between training and test distributions for SegAD and causing worse results compared to anomaly detector alone.", + "bbox": [ + 496, + 750, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17759", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. VAD, high-shot supervised benchmark", + "text_level": 1, + "bbox": [ + 76, + 90, + 410, + 107 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c27cf7d857f1f5bfb13ffa528c27e01b04abbd1aeffe27984edcb56e323b8bbd.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. High-shot supervised benchmark (VAD). The best result is marked in bold. All AD means PatchCore, FastFlow, RD4AD, and EfficientAD. Improvement calculated compared to base method results in Tab. 2." + ], + "table_body": "
MethodCl. AUROC ↑FPR@95TPR ↓
Supervised Anomaly Detection
DevNet86.9±1.266.6±2.7
DRA87.4±0.660.9±4.4
Supervised Classifier
WRN95.0±0.632.8±4.3
SegAD (Ours)
FastFlow + Ours86.9±0.556.9±3.9
RD4AD + Ours90.2±0.346.9±2.6
PatchCore + Ours91.7±0.348.0±2.9
EfficientAD + Ours91.4±0.236.8±1
AllAD+WRN+Ours96.5±0.316.4±1.9
", + "bbox": [ + 78, + 123, + 468, + 344 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The high-shot benchmark allows training supervised classifier $g$ , which shows competitive results as can be seen in Tab. 3 in the row \"WRN\", but it won't be able to detect unseen defects. Histograms in Fig. 6 visualize this problem. On the left is the one-class anomaly detector, which can detect unseen defects (in red) but shows a bad separation between good and bad parts. In the center, the supervised classifier $g$ shows a better separation of good and bad parts, but unseen defects distribution is shifted to the left because their scores are relatively low and such defects cannot be detected. On the right, SegAD shows a good separation between classes, but also unseen defects have a similar distribution to the seen defects. We still compare the supervised classifier Wide ResNet50 (WRN) with other methods, but the ability to detect unseen defects is crucial in real-world applications.", + "bbox": [ + 75, + 426, + 472, + 670 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b6e41dc57d33938baf62ad26303d3a531562e8669653edc7e43b6250cb64a215.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 91, + 684, + 202, + 758 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87b12149746b30f361d2d85e89cd9b02146e549bd0f99c86b0923721ffda84d7.jpg", + "image_caption": [ + "(b)", + "Figure 6. Distributions of scores in the VAD test set. Green color shows good parts, orange bad parts, and red bad parts with unseen defects. (a) shows one-class anomaly detector PatchCore, (b) supervised classifier Wide ResNet, (c) SegAD (PatchCore + WRN + Ours). X axis is score values, Y axis is frequency." + ], + "image_footnote": [], + "bbox": [ + 217, + 684, + 330, + 758 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6f3e7fed76afdc8428586670d383e9aabc288c58284f41305df1341e99ce7d08.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 344, + 684, + 455, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Combining SegAD, supervised classifier WRN, and all one-class anomaly detection models showed the best results", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "presented in Table 3. A more practical setting can be EfficientAD + WRN + Ours, which can achieve Cl. AUROC 96 and FPR@95TPR 19.4 with a much higher inference speed.", + "bbox": [ + 496, + 90, + 892, + 137 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. VAD, low-shot supervised benchmark", + "text_level": 1, + "bbox": [ + 498, + 146, + 825, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This benchmark uses just 100 bad images for training, which can be a relatively low number for a dataset with more than 20 types of defects. However, such a setting is closer to real-world applications, where a limited amount of defective images can be available. In Table 4, WRN shows much lower results due to a smaller training dataset, as well as DevNet and DRA. SegAD also performs worse, than with a high-shot benchmark, but it still can visibly improve the performance of other models. We do not use WRN in SegAD for this benchmark, because it showed low results by itself.", + "bbox": [ + 496, + 169, + 893, + 335 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/925ff91b5a4532f4588b82edab18a74c6fb9b056ba564acbde825f843f5f4c4f.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Low-shot supervised benchmark (VAD). The best result is marked in bold. All AD means PatchCore, FastFlow, RD4AD, and EfficientAD." + ], + "table_body": "
MethodCl. AUROC ↑FPR@95TPR ↓
Supervised Anomaly Detection
DevNet73.1±2.388.3±2.3
DRA78.9±1.874.6±1.9
Supervised Classifier
WRN78.7±1.877.9±4.4
SegAD (ours)
FastFlow + Ours84.3±1.464.5±2.5
RD4AD + Ours88.6±0.251.1±0.9
PatchCore + Ours90.8±0.552.7±4.3
EfficientAD + Ours90.6±0.740.1±2.4
All AD + Ours92.7±0.536.8±3.1
", + "bbox": [ + 500, + 348, + 890, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Results on VisA dataset", + "text_level": 1, + "bbox": [ + 500, + 650, + 718, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VisA contains fewer good images per class for training than VAD (500-1000 in VisA, compared to 2000 in VAD). The training set consists of good images only. For this reason, we use a different strategy to split the training dataset to make sure we have enough images to train the base anomaly detection model with a competitive result. For RD4AD, the training set was divided into $90\\%$ of images to train the RD4AD itself and $10\\%$ to train SegAD. EfficientAD uses a validation set ( $10\\%$ of good images from the train set) to normalize anomaly maps. In our case, the same images were used to train SegAD. The segmentation maps for this dataset were created with the Segment Anything Model (SAM) [12]. For most classes, it is a binary mask for the object and background, and for pcb1-4, static components were mapped on top of it; see Fig. 7.", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17760", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/faed8c248e9be65f9aa2750548993ae5d11f7d6f05544f5a94410540b5917c72.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 89, + 93, + 205, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce9cb5d1b85db512db2232c5280e543f58cecde19401f5aa932d5dc510f8ca9f.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 215, + 93, + 330, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a0ea3a90541364f5f291409e93ce9176edf77ffd775758bdf9f5658ed918d6c9.jpg", + "image_caption": [ + "(c)", + "Figure 7. VisA. Examples of segmentation maps created using SAM for (a) cashew, (b) macaroni1, (c) pb2." + ], + "image_footnote": [], + "bbox": [ + 339, + 93, + 455, + 181 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/38102a446d1d623fce49734ea46f6e219a65b13d25dabe8bdb8ab7d0b1514201.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCl. AUROC↑FPR@95TPR↓
One-class Anomaly Detection
RD4AD94.225.5
EfficientAD98.08.1
Supervised Anomaly Detection
DRA88.942.4
DevNet89.344.0
SegAD (ours)
RD4AD + Ours96.115.5
EfficientAD + Ours98.36.2
All AD + Ours98.47.5
", + "bbox": [ + 76, + 257, + 468, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With VisA, we create a new, challenging, supervised benchmark on a public dataset with 12 different classes. In a similar way to the supervised benchmark on MVTec AD [20], we move 10 randomly selected bad images from the test set to the train set, reporting results over 10 different seeds. In Table 5, this benchmark shows how SegAD performs on a different data from VAD and creates a new possibility to compete on a difficult supervised anomaly detection problem. Detailed results per class can be found in Appendix D.", + "bbox": [ + 75, + 546, + 468, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 709, + 225, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform an ablation study on the VAD high-shot supervised benchmark. We use the average result of the Patch-Core, FastFlow, RD4AD, and EfficientAD as the baseline in the Table 6, denoted as \"An.Det.\" These anomaly detectors use the maximum value from the anomaly map to calculate the score. On the contrary, we compute our proposed features $\\mathbf{f}$ , which are described in Section 4, and use BRF to yield a final score. This shows an improvement of 1.3 CI. AUROC in the row \"One Seg.\" In addition, we evaluate the segmentation map's impact by computing the anomaly map's maximum value per segment. Doing this gives an im", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "provement in 2.1 Cl. AUROC, which can be seen in the row \"Max.\" This shows how both features and segmentation maps are important for SegAD. We also show the strength of BRF by comparing it against other algorithms. We compare BRF with BT and RF using a segmentation map and features. The results show that the BRF is better than these methods by 1.7 and 0.3 Cl. AUROC, respectively.", + "bbox": [ + 496, + 90, + 890, + 198 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/adffd947c55868173cf1c01ba9758807a466da7397157812495609065d0bea5d.jpg", + "table_caption": [ + "Table 5. Results on VisA dataset, supervised benchmark. The best result is marked in bold. All AD means RD4AD and EfficientAD." + ], + "table_footnote": [], + "table_body": "
DescriptionFeat. fSeg. SBRFCl. AUROC ↑
An. Det.87.3
One Seg.✓✓88.6
Max.✓✓89.4
BT✓✓88.4
RF✓✓89.8
SegAD✓✓✓90.1
", + "bbox": [ + 501, + 207, + 890, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study on the methodical differences between our method (SegAD) and anomaly detection methods. An.Det. = Anomaly Detectors. One Seg. = One Segment. Max. means using maximum value. BT = Boosted Tree. RF = Random Forest. BRF = Boosted Random Forest. SegAD denotes the average result from PatchCore + Ours, FastFlow + Ours, RD4AD + Ours, EfficientAD + Ours.", + "bbox": [ + 496, + 354, + 893, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations: SegAD (ours) requires to have segmentation maps. Static maps can be used for objects like the product in VAD. Obtaining segmentation maps for unaligned objects can be more difficult, yet SAM [12] or a specially trained segmentation model may be a solution. The performance improvement of SegAD depends on the complexity and structure of the objects. For less complex objects in VisA, improvement is lower than for VAD, as can be seen in Tab. 3 and Tab. 5. Another possible limitation can be the difference in the anomaly maps between training and test data. Due to the nature of anomaly detection, this problem might be impossible to mitigate, because our task is to detect such differences.", + "bbox": [ + 496, + 484, + 890, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 694, + 617, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This study introduces VAD, a brand new supervised anomaly detection dataset derived from real production, offering challenging benchmarks for the research community to address real-world defect detection. Additionally, we propose SegAD, an innovative supervised anomaly detection method that achieves state-of-the-art performance on VAD as well as on a supervised benchmark on standard VisA dataset. Experimental results reveal the limitations of existing supervised anomaly detection methods in handling complex problems, while the integration of SegAD with top-performing one-class anomaly detection models further enhances their results.", + "bbox": [ + 496, + 719, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17761", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kilian Batzner, Lars Heckler, and Rebecca König. Efficientad: Accurate visual anomaly detection at millisecond-level latencies, 2023. 1, 3, 5", + "[2] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad — a comprehensive real-world dataset for unsupervised anomaly detection. In CVPR, pages 9584–9592, 2019. 1, 2, 3", + "[3] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. IJCV, 130(4):947-969, 2022. 2, 3", + "[4] Jakob Božić, Domen Tabernik, and Danijel Skočaj. Mixed supervision for surface-defect detection: from weakly to fully supervised learning. Computers in Industry, 2021. 2, 3", + "[5] Andrew P. Bradley. The use of the area under the roc curve in the evaluation of machine learning algorithms. PR, 30(7): 1145-1159, 1997. 5", + "[6] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In SIGKDD, pages 785-794, 2016. 6", + "[7] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In CVPR, pages 9737-9746, 2022. 1, 3, 5", + "[8] Chouro Ding, Guansong Pang, and Chunhua Shen. Catching both gray and black swans: Open-set supervised anomaly detection. In CVPR, 2022. 3, 5", + "[9] Colin Drury and Murray Sinclair. Human and machine performance in an inspection task. Human Factors: The Journal of the Human Factors and Ergonomics Society, 25:391-399, 1983. 1", + "[10] Lars Heckler, Rebecca König, and Paul Bergmann. Exploring the importance of pretrained feature extractors for unsupervised anomaly detection and localization. In CVPRW, pages 2917-2926, 2023. 2", + "[11] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In ICLR, 2018. 3", + "[12] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, pages 4015-4026, 2023. 7, 8", + "[13] Alex Krizhevsky et al. Learning multiple layers of features from tiny images. 2009. 3", + "[14] Yann LeCun, Corinna Cortes, Chris Burges, et al. Mnist handwritten digit database, 2010. 3", + "[15] Jiaqi Liu, Guoyang Xie, Jingbao Wang, Shangnian Li, Chengjie Wang, Feng Zheng, and Yaochu Jin. Deep industrial image anomaly detection: A survey. arXiv e-prints, pages arXiv-2301, 2023. 1, 2", + "[16] Muriel Mazzetto, Marcelo Teixeira, Érick Oliveira Rodrigues, and Dalcimar Casanova. Deep learning models for visual inspection on automotive assembling line. arXiv preprint arXiv:2007.01857, 2020. 1", + "[17] Declan McIntosh and Alexandra Branzan Albu. Interrealization channels: Unsupervised anomaly detection in images beyond one-class classification, 2023. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Yohei Mishina, Masamitsu Tsuchiya, and Hironobu Fujiyoshi. Boosted random forest. In VISAPP, pages 594-598, 2014. 4, 5", + "[19] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. VT-ADL: A vision transformer network for image anomaly detection and localization. In ISIE, 2021. 3", + "[20] Guansong Pang, Chunhua Shen, and Anton van den Hengel. Deep anomaly detection with deviation networks. In SIGKDD, pages 353-362, 2019. 8", + "[21] Guansong Pang, Choubo Ding, Chunhua Shen, and Anton van den Hengel. Explainable deep few-shot anomaly detection with deviation networks. arXiv preprint arXiv:2108.00462, 2021. 3, 5", + "[22] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In ICML, pages 1530-1538. PMLR, 2015. 3", + "[23] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Scholkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In CVPR, pages 14318-14328, 2022. 1, 3, 5", + "[24] Lukas Ruff, Robert A Vandermeulen, Nico Gornitz, Alexander Binder, Emmanuel Müller, Klaus-Robert Müller, and Marius Kloft. Deep semi-supervised anomaly detection. In ICLR, 2019. 3", + "[25] Maximilian Schwab, Charles Madeline-Dérou, Steffen Klarmann, Nils Thielen, Sven Meier, Jörg Franke, Sandan Chintanippu, and Wilhelm Stork. Multi-model machine learning based industrial vision framework for assembly part quality control. In *ETFA*, pages 1-4, 2022. 5", + "[26] Domen Tabernik, Samo Šela, Jure Skvarč, and Danijel Skočaj. Segmentation-based deep-learning approach for surface-defect detection. In Journal of Intelligent Manufacturing, 31(3):759-776, 2020. 2, 3", + "[27] Matthias Wieler and Tobias Hahn. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, 2007. 2, 3", + "[28] Xincheng Yao, Ruoqi Li, Jing Zhang, Jun Sun, and Chongyang Zhang. Explicit boundary guided semi-push-pull contrastive learning for supervised anomaly detection. In CVPR, pages 24490-24499, 2023. 1, 3", + "[29] Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows, 2021. 1, 2, 3, 5", + "[30] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC. British Machine Vision Association, 2016. 5", + "[31] Hui Zhang, Zuxuan Wu, Zheng Wang, Zhineng Chen, and Yu-Gang Jiang. Prototypical residual networks for anomaly detection and localization. In CVPR, pages 16281-16291, 2023. 1, 3", + "[32] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In ECCV, pages 392-408. Springer, 2022. 1, 2, 3, 6" + ], + "bbox": [ + 501, + 92, + 890, + 883 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17762", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2024/Supervised Anomaly Detection for Complex Industrial Images/ef358e69-3d45-4c6c-aa99-f5a5ceb11e08_model.json b/2024/Supervised Anomaly Detection for Complex Industrial Images/ef358e69-3d45-4c6c-aa99-f5a5ceb11e08_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d94b5517db38787fcdb0b80ad4c2f3c29d11b2bc --- /dev/null +++ b/2024/Supervised Anomaly Detection for Complex Industrial Images/ef358e69-3d45-4c6c-aa99-f5a5ceb11e08_model.json @@ -0,0 +1,1967 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.171, + 0.131, + 0.8, + 0.154 + ], + "angle": 0, + "content": "Supervised Anomaly Detection for Complex Industrial Images" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.18, + 0.751, + 0.2 + ], + "angle": 0, + "content": "Aimira Baitieva1, David Hurych2, Victor Besnier2, Olivier Bernard1" + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.215, + 0.551, + 0.233 + ], + "angle": 0, + "content": "\\(^{1}\\)valeo, \\(^{2}\\)valeo.ai" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.237, + 0.81, + 0.253 + ], + "angle": 0, + "content": "{aimira.baitieva, david.hurych, victor.besnier, olivier.bernard}@valeo.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.319, + 0.474, + 0.635 + ], + "angle": 0, + "content": "Automating visual inspection in industrial production lines is essential for increasing product quality across various industries. Anomaly detection (AD) methods serve as robust tools for this purpose. However, existing public datasets primarily consist of images without anomalies, limiting the practical application of AD methods in production settings. To address this challenge, we present (1) the Valeo Anomaly Dataset (VAD), a novel real-world industrial dataset comprising 5000 images, including 2000 instances of challenging real defects across more than 20 subclasses. Acknowledging that traditional AD methods struggle with this dataset, we introduce (2) Segmentation-based Anomaly Detector (SegAD). First, SegAD leverages anomaly maps as well as segmentation maps to compute local statistics. Next, SegAD uses these statistics and an optional supervised classifier score as input features for a Boosted Random Forest (BRF) classifier, yielding the final anomaly score. Our SegAD achieves state-of-the-art performance on both VAD (+2.1% AUROC) and the VisA dataset (+0.4% AUROC). The code and the models are publicly available1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.667, + 0.21, + 0.683 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.693, + 0.47, + 0.829 + ], + "angle": 0, + "content": "Within the manufacturing process, industrial visual inspection plays a crucial role in identifying defects in produced components. This operation holds significant importance in minimizing costs by identifying and removing faulty parts early in the production stages and, more importantly, in preventing the dispatch of defective components to customers. Traditionally, this task has relied on human operators; however, the likelihood of overlooking certain defects can be as high as \\(25\\%\\) for specific defect types [9]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.83, + 0.47, + 0.875 + ], + "angle": 0, + "content": "When the inspected product comprises numerous components, its examination may create a bottleneck in the production process, causing delays across the entire line. While" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.288, + 0.892, + 0.349 + ], + "angle": 0, + "content": "conventional computer vision methods applied to this task demonstrate superior speed and lower error rates compared to human operators [16], their inflexibility and lack of satisfactory accuracy [15] limit their effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.893, + 0.563 + ], + "angle": 0, + "content": "Industrial deep learning anomaly detection has been an active research field in recent years. Most of anomaly detection methods use only good images for training and try to detect deviations from the training data [1, 7, 23, 29]. Development of the new methods is restrained by publicly available datasets. Recent industrial anomaly detection datasets typically contain approximately one hundred (or even fewer) abnormal images, showcasing defects in the testing set only [2, 32]. This poses a challenge for supervised anomaly detection methods aiming to utilize both normal and defective parts during training. Supervised models often undergo training with just ten abnormal images from the testing set, resulting in overfitting and reduced sensitivity to previously unseen defects [28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.895, + 0.779 + ], + "angle": 0, + "content": "In response to the limitations of current datasets, we introduce and publicly release VAD (Valeo Anomaly Dataset), which contains 1000 bad and 2000 good parts in the training set and 1000 bad and 1000 good parts in the test set, see Fig. 1(a). Among the defective parts used for testing, 165 contain specific defect types not present in the training data; these parts are explicitly labeled in the released dataset. All images in VAD are captured from an actual production line, showcasing a diverse range of defects, from highly obvious to extremely subtle. This dataset bridges the gap between the academic community and the industry, offering researchers the chance to advance the performance of methods in tackling more intricate real-world challenges." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Current approaches to supervised anomaly detection either yield unsatisfactory results; see Tab. 3, or demand pixel-level labels for defective parts, as seen in [28, 31]. Consequently, we introduce a novel method named SegAD (Segmentation-based Anomaly Detector), which is described in Fig. 1(b). The employed approach eliminates the need for pixel-level labels, requiring only a flag for each image. Anomaly map scores from each segment are used" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.358, + 0.901 + ], + "angle": 0, + "content": "1https://github.com/abc-125/segad" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17754" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.095, + 0.874, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.286, + 0.235, + 0.298 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.287, + 0.594, + 0.3 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.315, + 0.892, + 0.384 + ], + "angle": 0, + "content": "Figure 1. Overview of our contributions. (a) VAD, a real-world industrial dataset designed for supervised anomaly detection with complex defects. (b) SegAD, our method that leverages anomaly maps extracted from segmented outputs of one or more anomaly detectors. Higher-level statistical features are computed on these maps, such as skewness, kurtosis, or mean, to generate the local anomaly features. Additionally, our SegAD provides the flexibility to use the output of a supervised classifier score with the local anomaly features, creating input for a final Boosted Random Forest (BRF) classifier that yields the final score." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.468, + 0.501 + ], + "angle": 0, + "content": "to delineate the parameters of the distribution, emphasizing relatively low scores in important areas while disregarding higher scores in noisy regions. In this context, the precise position and appearance of the anomaly diminishes in importance as long as it resides within a designated segment. Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.504, + 0.468, + 0.639 + ], + "angle": 0, + "content": "- We propose a supervised anomaly detection dataset with complex objects and a large variety of defects. As an addition, we establish a one-class anomaly detection benchmark, which is more challenging than most of the current ones, a supervised benchmark with a high number of bad images for training (1000 images) and a supervised benchmark with a low number of bad images for training (100 images). We share the dataset freely with the research community." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.64, + 0.468, + 0.73 + ], + "angle": 0, + "content": "- We create a novel supervised anomaly detection method called SegAD. This approach performs noticeably better than recent anomaly detectors alone while retaining the ability to detect unknown defects. SegAD reaches SOTA results on VAD as well as on the established VisA dataset [32]." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.504, + 0.468, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.747, + 0.219, + 0.763 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.774, + 0.468, + 0.834 + ], + "angle": 0, + "content": "In this section, we first discuss existing datasets and explain how VAD is different. Next, we give a quick overview of existing anomaly detection methods. We show in Table 1 a global comparison between our VAD and existing datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.847, + 0.179, + 0.86 + ], + "angle": 0, + "content": "2.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In 2019, the MVTec AD dataset [2] was introduced. This dataset significantly advanced anomaly detection methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.546 + ], + "angle": 0, + "content": "by providing images with authentic defects and realistic objects, in contrast to earlier datasets predominantly comprising textures and simple defects [4, 26, 27]. MVtec LOCO AD [3] highlighted the challenge of logical defects, such as missing or misplaced parts, necessitating anomaly detection methods capable of capturing the global context of the image. The more recent VisA dataset [32] expands on this by incorporating diverse objects with complex structures, multiple instances, and variations in scale." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Existing datasets often contain simulated defects [2, 3, 32], creating a domain gap between research and practical applications. This complicates the deployment of anomaly detection methods in real-world settings. To promote supervised anomaly detection, we propose the Valeo Anomaly Dataset (VAD), a real-world industrial dataset with various defects, including logical ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.744 + ], + "angle": 0, + "content": "Another notable issue stems from the saturation of solved datasets, where other methods are influenced more by dataset-specific design choices than general applicability. Compared to this, our dataset introduces a broader range of challenging and diverse anomalies and high intra-class variability of good images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.604, + 0.77 + ], + "angle": 0, + "content": "2.2. Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To define the terms we are working with, \"one-class\" methods are commonly called unsupervised in the context of anomaly detection [10, 15, 29]. We use \"one-class\" as a more accurate name since these methods use labeled images for training. It is important to distinguish such methods from ones that work with unlabelled data, as in [17]. Models that use bad and good images for training are usually called \"supervised\", which is the name we also adopt." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17755" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.131 + ], + "angle": 0, + "content": "Table 1. Comparison of different anomaly detection datasets. Tr. good and Tr. bad show the average number of images per class for training. tex = texture, obj = object, Log. def. = Logical defects, Pix. labels = pixel-level labels and wl = weakly labelled (ellipse around the defect)." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.142, + 0.894, + 0.29 + ], + "angle": 0, + "content": "
DatasetYearClassesImagesTypesDefectsTr. goodTr. badLog. def.Pix. labels
DAGM [27]20071011500texsynthetic50075nowl
KSSD [26]20191399texreal-world11517noyes
MVTec AD [2]2019155354tex, objsimulated2420noyes
KSSD2 [4]202113335texreal-world2085246noyes
BTAD [19]202132830tex, objreal-world6000noyes
LOCO AD [3]202253644objsimulated3540yesyes
VisA [32]20221210821objsimulated7200noyes
VAD (ours)202315000objreal-world20001000yesno
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.473, + 0.554 + ], + "angle": 0, + "content": "One-class Anomaly Detection. PatchCore [23] relies on a pretrained feature extractor model to extract features from the training set into a memory bank and reduce the size of a memory bank using coreset subsampling. Features extracted from the new input are compared to their nearest neighbors in the memory bank. FastFlow [29] uses a similar feature extractor model and maps extracted features to the Gaussian distribution using normalizing flow [22]. RD4AD [7] utilizes a teacher-student architecture, which combines features from several layers of feature extractor to eliminate redundant ones. EfficientAD [1] introduces many innovations, including using an autoencoder to detect logical defects as an addition to a classic teacher-student architecture and their own pretrained feature extractor, which imitates the behavior of a bigger model with a noticeably lesser inference time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.473, + 0.789 + ], + "angle": 0, + "content": "Supervised Anomaly Detection. In theory, leveraging defective parts for training is advantageous for refining class boundaries [24]. However, practical challenges arise, such as the small size of anomalies and the impossibility of collecting all potential defects [31]. Several recent anomaly detection methods use both good and bad images [21, 28, 31], but some of them overfit to seen defects, and others require pixel masks for defects to calculate loss or to generate new defects, which can be problematic in a real-world case. Real-world datasets (as VAD) might contain logical defects that cannot be generated by existing defects generation strategies, which usually cut and paste existing defects with various modifications [28, 31], which is not going to improve results for a wrong wire shape as or other similar types of defects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.904 + ], + "angle": 0, + "content": "Very few supervised anomaly detection benchmarks are available; the most popular is Supervised Anomaly Detection on MVTec AD. DevNet [21] is one of the first supervised anomaly detection models that tried to solve industrial MVTec AD dataset, compared to earlier methods which were mostly addressing Out-of-Distribution problem [11, 24] using non-industrial datasets such as MNIST [14]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.895, + 0.451 + ], + "angle": 0, + "content": "or CIFAR-10 [13]. DRA [8] uses several heads to learn both seen and pseudo anomalies, as well as normal examples, which should reduce overfitting on seen anomalies. These methods perform well on MVtec AD but might fail on more complex problems, as shown in Tab. 3 and Tab. 5. Compared to modern supervised anomaly detection methods, SegAD (ours) performs better than current SOTA methods, even on complex problems without pixel-level labels, extensive augmentation, and long training." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.466, + 0.78, + 0.483 + ], + "angle": 0, + "content": "3. Valeo Anomaly Dataset (VAD)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.895, + 0.749 + ], + "angle": 0, + "content": "VAD consists of one class with predefined training and testing sets. The training set contains 1000 bad and 2000 good images, and the testing set contains 1000 bad, 165 of them are unseen bad, and 1000 good images. Unseen bad images in the test dataset refer to several rare defect types that are not present in the training data, several examples of which are shown in Fig. 4. Having such images is important to avoid turning this anomaly detection problem into a supervised classification problem. Some images might have a thin black border at the bottom due to how they were filmed, an example can be seen in Fig. 2. Defects might occur in the whole area of the image. Image-level annotation is provided, but there is no pixel-level annotation due to the complexity of defects and the fact there is no exact position for missing or misplaced components. All images are \\(512 \\times 512\\) pixels in PNG format. Examples of all types of defects can be found in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Good parts consist of two wires connected to two pins on one side and two soldering dots on the other, placed on a piezo. Piezo means piezoelectric element, a big round area under the other components. Wires should have the right shape, not too straight, too long, or bent too much. Wires should connect the soldering dots as close to their centers as possible. Some amount of deviation is allowed, but if the wire is too close to the side of a solder dot, it is a defect. Soldering should be well-rounded, not too small, and placed in the correct location, not overlapping with the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "17756" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.089, + 0.26, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.089, + 0.436, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.225, + 0.26, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.225, + 0.436, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.388, + 0.47, + 0.417 + ], + "angle": 0, + "content": "Figure 2. Good parts. Small scratches on the piezo are allowed, as well as wire being closer to the side of the soldering. (VAD)" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.443, + 0.27, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.509, + 0.269, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.587, + 0.218, + 0.599 + ], + "angle": 0, + "content": "Logical defects" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.443, + 0.453, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.509, + 0.452, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.587, + 0.425, + 0.598 + ], + "angle": 0, + "content": "Structural defects" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.671 + ], + "angle": 0, + "content": "Figure 3. Variety of defects: logical, from the top left: wire on the side, wire out of solder dot, missing wire, bad wire shape (bent too much). Structural: soldering paste pollution, crack on a piezo, burned area under the solder dot, broken piezo. (VAD)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.47, + 0.776 + ], + "angle": 0, + "content": "piezo contours. The wire should be properly connected to the pin without any pieces of it visible on the side of the pin. The piezo can have different textures. Any kind of pollution is considered to be an anomaly, which makes the part defective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Logical defects result from components being misplaced or misshapen rather than being damaged, e.g. bad wire shape, bad soldering shape, wrong wire position, wrong soldering position, and missing parts. Structural defects include various cracks and broken parts of the piezo, as well as a variety of burns and pollution. For both types of defects, see Fig. 3. The importance of logical defects is often overlooked in the existing datasets (excluding MVTec" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.095, + 0.643, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.096, + 0.748, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.096, + 0.876, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.199, + 0.892, + 0.227 + ], + "angle": 0, + "content": "Figure 4. Unseen defects for left to right: burned solder dot, burned wire, extra unnecessary wire. (VAD)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.252, + 0.892, + 0.313 + ], + "angle": 0, + "content": "LOCO AD) because they are hard to imitate if the dataset is filmed in the lab, but also hard to detect. VAD provides the research community the opportunity to work with real logical constraints of a complex object." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.321, + 0.634, + 0.336 + ], + "angle": 0, + "content": "3.1.Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.344, + 0.892, + 0.541 + ], + "angle": 0, + "content": "We establish three benchmarks for the VAD dataset. The first is one-class anomaly detection. As shown in Tab 2, even SOTA methods struggle to perform well on our dataset compared to other popular datasets, which leaves an opening for further research. For this benchmark, no bad images are used for training. The second benchmark is high-shot supervised anomaly detection, which uses all 1000 bad images from the training set. The last benchmark is low-shot supervised anomaly detection. Only 100 bad images from the training set are used. These images are randomly selected using 5 seeds, with the average result over all seeds being expected. The same 2000 good parts are used for the training in all benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.553, + 0.591, + 0.568 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.578, + 0.892, + 0.776 + ], + "angle": 0, + "content": "SegAD works with image data at its input, and after inference, it is expected to output a score that describes a probability of anomaly appearance. Let \\(\\mathcal{I} = \\left[0,\\dots ,255\\right]^{W,H,3}\\) be a space of all images of width \\(W\\), height \\(H\\) and with 3 color channels. The \\(i\\)-th image from a dataset is marked as \\(I^{(i)}\\in \\mathcal{I}\\). We denote \\(\\mathbf{x}\\) to be a 2D pixel coordinate vector. SegAD inference consists of three consecutive stages in a pipeline. First is anomaly map(s) calculation with optional supervised classifier score calculation. Next is the calculation of simple statistics from anomaly maps over individual segments. Lastly, the obtained statistics, and optionally also the classifier score, are used as input features for a BRF classifier [18] that delivers the final result." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.777, + 0.892, + 0.872 + ], + "angle": 0, + "content": "Let us denote \\(K\\) as the number of used pixel-wise anomaly detectors and mark them as functions \\(f_{k}(I):\\mathcal{I}\\to \\mathbb{R}^{W,H},k\\in \\{1,\\ldots ,K\\}\\). A supervised classifier is a function \\(g(I):\\mathbb{R}^{W,H}\\to \\mathbb{R}\\). Next, we introduce a set of \\(L\\) mutually exclusive segments \\(S = \\{s_1,\\dots,s_L\\}\\), where \\(s_l\\in [0,1]^{W,H}\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "After the calculation of anomaly maps for all \\( f_{k}(I) \\), the masks \\( s_l \\) are used for segment-wise" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17757" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "extraction of results, followed by the calculation of several statistics. The final feature vector \\(\\mathbf{f} = [g(I),[\\mathbf{q}^L ]_{k = 1}^K,[\\mathbf{z}^L ]_{k = 1}^K,[\\mathbf{c}^L ]_{k = 1}^K,[\\mathbf{m}^L ]_{k = 1}^K]\\) extracted from one image, is the concatenation of 4 vectors containing different statistics. \\(\\mathbf{q}_k^L\\) stands for a vector of \\(L\\) numbers where each stands for a \\(99.5\\%\\) quantile from anomaly map \\(f_{k}(I)\\) over pixels where \\(s_l(\\mathbf{x}) = 1\\) Similarly, for every anomaly map \\(f_{k}(I)\\) we calculate a vector for skew \\(\\mathbf{z}_k^L\\) , kurtosis \\(\\mathbf{c}_k^L\\) and mean \\(\\mathbf{m}_k^L\\) . The length of the feature vector \\(\\mathbf{f}\\) equals to \\(K*L*4 + 1\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.243, + 0.472, + 0.318 + ], + "angle": 0, + "content": "For notation simplicity, we keep the supervised classifier score \\( g(I) \\) always present in \\( \\mathbf{f} \\), but it may be omitted, which is clearly defined in every experiment setup. Feature vectors \\( \\mathbf{f}^{(i)} \\) are used as inputs for training the final classifier [18], which, at inference time, delivers the final result." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.329, + 0.182, + 0.345 + ], + "angle": 0, + "content": "4.1. Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.457 + ], + "angle": 0, + "content": "Training SegAD requires two separate subsets from the training dataset. The first subset is used to train base models \\( f_{k} \\), the second subset is used to train the BRF classifier. To create these subsets from one training dataset, we split available good images into two parts. For VAD, we report average results over 5 different splits to show more reliable results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.685 + ], + "angle": 0, + "content": "Each model \\((f_{k}, g, \\mathrm{BRF})\\) is trained independently. Training can be separated into two stages. In the first stage, anomaly detector(s) \\(f_{k}\\) are trained using the same data. In the same stage, a segmentation map \\(S\\) must be defined. As will be elaborated in more detail in Sec. 5, the segmentation map can be static or produced by a segmentation model, depending on the dataset. A supervised classifier \\(g\\) is trained separately on \\(80\\%\\) of the full training set (1600 good and 800 bad parts for VAD) to ensure that the training sets for this model and SegAD are not fully overlapping. The next stage is to train a final BRF classifier in SegAD, which requires both good and bad images, bad images can be replaced with artificial defects in some cases, as shown in Tab. 2. BRF is trained with the output of the models from the previous stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.469, + 0.792 + ], + "angle": 0, + "content": "An important thing to note is that SegAD is very fast to train as long as you already have other models available. Extraction of features and training the final classifier takes noticeably less time than training anomaly detectors. The same is true for the inference; SegAD adds a negligible time to processing the image while improving results significantly." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.249, + 0.818 + ], + "angle": 0, + "content": "4.2. Setups for SegAD" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "SegAD can use any anomaly detector(s) \\( f_{k} \\) as long as it produces an anomaly map showing pixel-level anomaly scores, although, in this paper, we use only one-class anomaly detectors because they require less data to train. Any supervised classifier \\( g \\) that returns a classification score can be" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "used. In experiments, we use several setups, which will be described here in detail. SegAD is denoted as Ours. Supervised classifier \\( g \\) is not used unless stated otherwise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.212 + ], + "angle": 0, + "content": "Anomaly Detector + Ours means \\( K = 1 \\), \\( f_{1} = \\) \"Anomaly Detector name\". This setup allows us to show improvement for one anomaly detection model in particular and demonstrate that this improvement can be achieved for different types of anomaly detectors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.333 + ], + "angle": 0, + "content": "All \\(\\mathbf{AD} + \\mathbf{Ours}\\) denotes that all anomaly detectors from the list \\(AllAD\\) of length \\(N\\) are used to produce anomaly maps. Such a list can be found in the caption of the table. In that case \\(K = N\\), \\(f_{1} = AllAD_{1},\\ldots ,f_{N} = AllAD_{N}\\). Anomaly detectors might use an ensemble of feature extractors, similarly to [23], which makes them slower but can improve the result. In our case, anomaly detectors themselves can be successfully assembled in a supervised way." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.483 + ], + "angle": 0, + "content": "All AD + Supervised Classifier + Ours denotes a setup in which \\( g = \\) \"Supervised Classifier name\", and the rest is the same as in the previous setup. A supervised classifier allows us to detect seen defects, which can be invisible to a one-class classifier, similarly to [25]. Such a setup shows the best result for academic purposes, but the speed of inference can be unsatisfying for real-world applications. We expect that, in that case, the number and selection of anomaly detectors can be optimized to fit the problem as well as the time constraints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.892, + 0.59 + ], + "angle": 0, + "content": "Anomaly Detector + Supervised Classifier + Ours mean a setup in which \\( g = \\) \"Supervised Classifier name\", \\( K = 1 \\), \\( f_{1} = \\) \"Anomaly Detector name\". It shows an example of an optimization mentioned in the previous setup. Only one anomaly detector gives a higher inference speed compared to the ensemble of anomaly detectors while giving a satisfying result." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.633, + 0.621 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Evaluation metrics. We compare results using the following metrics: (1) Area Under the Receiver Operating Characteristic Curve (AUROC) [5], (2) False Positive Rate (FPR) at \\(95\\%\\) True Positive Rate (TPR) denoted as FPR@95TPR. Classification AUROC, which describes the performance on the image level for multiple thresholds, denoted by Cl. AUROC. FPR@95TPR shows the percentage of misclassified good parts (false positive) at \\(95\\%\\) of bad parts classified correctly (true positive). We report mean results \\(\\pm\\) standard deviation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation and evaluation details. SegAD uses the output of one-class detectors PatchCore [23], FastFlow [29], Reverse Distillation [7] (referred to as RD4AD), EfficientAD [1] either alone or several anomaly detectors at once. SegAD results are compared to the results of these models alone, as well as supervised anomaly detectors DevNet [21] and DRA [8]. We also compare our results to the supervised classifier Wide ResNet50 [30] (denoted as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17758" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "WRN). The final classifier BRF was trained with the same parameters for all datasets using XGBoost library [6]. More details on the implementation can be found in Appendixes A and B." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.468, + 0.287 + ], + "angle": 0, + "content": "Images are resized to \\( \\mathrm{W} = \\mathrm{H} = 256 \\) pixels to make the comparison less biased. The DRA model originally uses a resolution \\( 448 \\times 448 \\), which shows better results on VAD than \\( 256 \\times 256 \\), 92.8 CI. AUROC, which is still lower than the top results. The higher resolution also improves results for some of the other models, so we do not use it. \\( 256 \\times 256 \\) is sufficient for defects to remain visible, but it also allows to have a reasonable speed of inference, especially for anomaly detection models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.468, + 0.5 + ], + "angle": 0, + "content": "For VAD, results averaged for 5 seeds and training runs are reported. These seeds are used for training and splitting the training set for SegAD into two sets, as described in Subsection 4.1. 2000 good parts from the train set were split in half: 1000 for training the base model(s) and 1000 for training SegAD. 0, 1000, or 100 bad parts (depending on the benchmark) were also used to train SegAD. Details on the VisA benchmark can be found in Subsection 5.4. No augmentations were used for training, and existing augmentations were removed from used methods. Augmentations tend to be suitable only for specific tasks (as rotation in DRA or DevNet, removing it improved results by these models on the VAD drastically), and using augmentations makes it hard to compare models themselves." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.468, + 0.577 + ], + "angle": 0, + "content": "We include three benchmarks for the new VAD and one additional for the VisA dataset [32]. We compare results for VAD for one-class, high-shot, and low-shot supervised benchmarks, as described in Subsection 3.1. The benchmark for the VisA dataset is explained in 5.4. The segmen" + }, + { + "type": "image", + "bbox": [ + 0.157, + 0.59, + 0.391, + 0.772 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.782, + 0.468, + 0.839 + ], + "angle": 0, + "content": "Figure 5. Static segmentation map for objects in VAD dataset, overlayed over the image. The image is separated into \\( L = 7 \\) segments: background, outer half-circle, piezo border, solder dots area, wires area, pins area, and piezo in the middle." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "tation maps for VAD is the same static image for every input image from the dataset. The segmentation map can be seen in Fig. 5. Due to the fact that the object in the VAD is al" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "ways centered with the constant size and component placement, this is the simplest choice, which may be sub-optimal, but serves the purpose and does not require any manual or model-inferred annotation of the data. It was created to separate the anomaly maps into several meaningful parts, such as piezo border, wire areas, soldering area, etc., which can contain different defects and different levels of falsely high scores on the anomaly map. Another source of segmentation maps can be a specially trained segmentation model or a zero-shot segmentation model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.258, + 0.743, + 0.273 + ], + "angle": 0, + "content": "5.1. VAD, one-class benchmark" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.283, + 0.892, + 0.465 + ], + "angle": 0, + "content": "This benchmark allows only good images for training, but SegAD requires bad images as well, so we try to replace them with artificially generated defects. Because SegAD works with anomaly maps, generating realistic defects is unnecessary. For this reason, we have applied several simple augmentations to the 1000 good images available for training SegAD. These augmentations include Gaussian blur or a randomly placed, randomly sized thin rectangle of a random shade of gray. They were applied to randomly selected segments or random parts of the segments; segments are defined with the segmentation map. Code for generating defects can be found in the SegAD GitHub repository." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.481, + 0.892, + 0.691 + ], + "angle": 0, + "content": "
MethodCl. AUROC ↑FPR@95TPR ↓
One-class Anomaly Detection
FastFlow79.1±1.776.2±1.9
RD4AD84.5±0.265.8±2
PatchCore88.3±0.360.4±2.8
EfficientAD89.1±0.844.5±2.2
SegAD (Ours)
FastFlow + Ours83.3±1.371.5±5
RD4AD + Ours87.5±1.261.0±7.9
PatchCore + Ours89.7±0.858.8±3.7
EfficientAD + Ours85.6±1.6100±0
All AD + Ours90.4±0.552.4±4.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.701, + 0.892, + 0.728 + ], + "angle": 0, + "content": "Table 2. One-class benchmark (VAD). The best result is marked in bold. All AD means PatchCore, FastFlow, and RD4AD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Results in Table 2 show that even such a naive approach improves results compared to anomaly detection methods alone. Even greater improvement can be achieved by using several anomaly detectors to produce anomaly maps for SegAD. This strategy fails for EfficientAD + Ours, anomaly maps produced by EfficientAD for the generated defects are too different from anomaly maps for real defects, creating a large discrepancy between training and test distributions for SegAD and causing worse results compared to anomaly detector alone." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17759" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.411, + 0.108 + ], + "angle": 0, + "content": "5.2. VAD, high-shot supervised benchmark" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.124, + 0.47, + 0.345 + ], + "angle": 0, + "content": "
MethodCl. AUROC ↑FPR@95TPR ↓
Supervised Anomaly Detection
DevNet86.9±1.266.6±2.7
DRA87.4±0.660.9±4.4
Supervised Classifier
WRN95.0±0.632.8±4.3
SegAD (Ours)
FastFlow + Ours86.9±0.556.9±3.9
RD4AD + Ours90.2±0.346.9±2.6
PatchCore + Ours91.7±0.348.0±2.9
EfficientAD + Ours91.4±0.236.8±1
AllAD+WRN+Ours96.5±0.316.4±1.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.414 + ], + "angle": 0, + "content": "Table 3. High-shot supervised benchmark (VAD). The best result is marked in bold. All AD means PatchCore, FastFlow, RD4AD, and EfficientAD. Improvement calculated compared to base method results in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.473, + 0.671 + ], + "angle": 0, + "content": "The high-shot benchmark allows training supervised classifier \\( g \\), which shows competitive results as can be seen in Tab. 3 in the row \"WRN\", but it won't be able to detect unseen defects. Histograms in Fig. 6 visualize this problem. On the left is the one-class anomaly detector, which can detect unseen defects (in red) but shows a bad separation between good and bad parts. In the center, the supervised classifier \\( g \\) shows a better separation of good and bad parts, but unseen defects distribution is shifted to the left because their scores are relatively low and such defects cannot be detected. On the right, SegAD shows a good separation between classes, but also unseen defects have a similar distribution to the seen defects. We still compare the supervised classifier Wide ResNet50 (WRN) with other methods, but the ability to detect unseen defects is crucial in real-world applications." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.685, + 0.203, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.765, + 0.153, + 0.774 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.685, + 0.331, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.765, + 0.278, + 0.774 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.685, + 0.457, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.765, + 0.411, + 0.774 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.786, + 0.47, + 0.858 + ], + "angle": 0, + "content": "Figure 6. Distributions of scores in the VAD test set. Green color shows good parts, orange bad parts, and red bad parts with unseen defects. (a) shows one-class anomaly detector PatchCore, (b) supervised classifier Wide ResNet, (c) SegAD (PatchCore + WRN + Ours). X axis is score values, Y axis is frequency." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Combining SegAD, supervised classifier WRN, and all one-class anomaly detection models showed the best results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.138 + ], + "angle": 0, + "content": "presented in Table 3. A more practical setting can be EfficientAD + WRN + Ours, which can achieve Cl. AUROC 96 and FPR@95TPR 19.4 with a much higher inference speed." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.147, + 0.826, + 0.164 + ], + "angle": 0, + "content": "5.3. VAD, low-shot supervised benchmark" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.17, + 0.895, + 0.337 + ], + "angle": 0, + "content": "This benchmark uses just 100 bad images for training, which can be a relatively low number for a dataset with more than 20 types of defects. However, such a setting is closer to real-world applications, where a limited amount of defective images can be available. In Table 4, WRN shows much lower results due to a smaller training dataset, as well as DevNet and DRA. SegAD also performs worse, than with a high-shot benchmark, but it still can visibly improve the performance of other models. We do not use WRN in SegAD for this benchmark, because it showed low results by itself." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.349, + 0.892, + 0.571 + ], + "angle": 0, + "content": "
MethodCl. AUROC ↑FPR@95TPR ↓
Supervised Anomaly Detection
DevNet73.1±2.388.3±2.3
DRA78.9±1.874.6±1.9
Supervised Classifier
WRN78.7±1.877.9±4.4
SegAD (ours)
FastFlow + Ours84.3±1.464.5±2.5
RD4AD + Ours88.6±0.251.1±0.9
PatchCore + Ours90.8±0.552.7±4.3
EfficientAD + Ours90.6±0.740.1±2.4
All AD + Ours92.7±0.536.8±3.1
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.498, + 0.582, + 0.893, + 0.624 + ], + "angle": 0, + "content": "Table 4. Low-shot supervised benchmark (VAD). The best result is marked in bold. All AD means PatchCore, FastFlow, RD4AD, and EfficientAD." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.719, + 0.667 + ], + "angle": 0, + "content": "5.4. Results on VisA dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "VisA contains fewer good images per class for training than VAD (500-1000 in VisA, compared to 2000 in VAD). The training set consists of good images only. For this reason, we use a different strategy to split the training dataset to make sure we have enough images to train the base anomaly detection model with a competitive result. For RD4AD, the training set was divided into \\(90\\%\\) of images to train the RD4AD itself and \\(10\\%\\) to train SegAD. EfficientAD uses a validation set (\\(10\\%\\) of good images from the train set) to normalize anomaly maps. In our case, the same images were used to train SegAD. The segmentation maps for this dataset were created with the Segment Anything Model (SAM) [12]. For most classes, it is a binary mask for the object and background, and for pcb1-4, static components were mapped on top of it; see Fig. 7." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17760" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.094, + 0.207, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.188, + 0.155, + 0.199 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.094, + 0.331, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.188, + 0.281, + 0.2 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.094, + 0.456, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.188, + 0.406, + 0.199 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.214, + 0.47, + 0.243 + ], + "angle": 0, + "content": "Figure 7. VisA. Examples of segmentation maps created using SAM for (a) cashew, (b) macaroni1, (c) pb2." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.258, + 0.47, + 0.464 + ], + "angle": 0, + "content": "
MethodCl. AUROC↑FPR@95TPR↓
One-class Anomaly Detection
RD4AD94.225.5
EfficientAD98.08.1
Supervised Anomaly Detection
DRA88.942.4
DevNet89.344.0
SegAD (ours)
RD4AD + Ours96.115.5
EfficientAD + Ours98.36.2
All AD + Ours98.47.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.476, + 0.469, + 0.503 + ], + "angle": 0, + "content": "Table 5. Results on VisA dataset, supervised benchmark. The best result is marked in bold. All AD means RD4AD and EfficientAD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.698 + ], + "angle": 0, + "content": "With VisA, we create a new, challenging, supervised benchmark on a public dataset with 12 different classes. In a similar way to the supervised benchmark on MVTec AD [20], we move 10 randomly selected bad images from the test set to the train set, reporting results over 10 different seeds. In Table 5, this benchmark shows how SegAD performs on a different data from VAD and creates a new possibility to compete on a difficult supervised anomaly detection problem. Detailed results per class can be found in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.227, + 0.727 + ], + "angle": 0, + "content": "5.5. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We perform an ablation study on the VAD high-shot supervised benchmark. We use the average result of the Patch-Core, FastFlow, RD4AD, and EfficientAD as the baseline in the Table 6, denoted as \"An.Det.\" These anomaly detectors use the maximum value from the anomaly map to calculate the score. On the contrary, we compute our proposed features \\( \\mathbf{f} \\), which are described in Section 4, and use BRF to yield a final score. This shows an improvement of 1.3 CI. AUROC in the row \"One Seg.\" In addition, we evaluate the segmentation map's impact by computing the anomaly map's maximum value per segment. Doing this gives an im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.199 + ], + "angle": 0, + "content": "provement in 2.1 Cl. AUROC, which can be seen in the row \"Max.\" This shows how both features and segmentation maps are important for SegAD. We also show the strength of BRF by comparing it against other algorithms. We compare BRF with BT and RF using a segmentation map and features. The results show that the BRF is better than these methods by 1.7 and 0.3 Cl. AUROC, respectively." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.208, + 0.892, + 0.346 + ], + "angle": 0, + "content": "
DescriptionFeat. fSeg. SBRFCl. AUROC ↑
An. Det.87.3
One Seg.✓✓88.6
Max.✓✓89.4
BT✓✓88.4
RF✓✓89.8
SegAD✓✓✓90.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.356, + 0.894, + 0.453 + ], + "angle": 0, + "content": "Table 6. Ablation study on the methodical differences between our method (SegAD) and anomaly detection methods. An.Det. = Anomaly Detectors. One Seg. = One Segment. Max. means using maximum value. BT = Boosted Tree. RF = Random Forest. BRF = Boosted Random Forest. SegAD denotes the average result from PatchCore + Ours, FastFlow + Ours, RD4AD + Ours, EfficientAD + Ours." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.892, + 0.681 + ], + "angle": 0, + "content": "Limitations: SegAD (ours) requires to have segmentation maps. Static maps can be used for objects like the product in VAD. Obtaining segmentation maps for unaligned objects can be more difficult, yet SAM [12] or a specially trained segmentation model may be a solution. The performance improvement of SegAD depends on the complexity and structure of the objects. For less complex objects in VisA, improvement is lower than for VAD, as can be seen in Tab. 3 and Tab. 5. Another possible limitation can be the difference in the anomaly maps between training and test data. Due to the nature of anomaly detection, this problem might be impossible to mitigate, because our task is to detect such differences." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.619, + 0.71 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This study introduces VAD, a brand new supervised anomaly detection dataset derived from real production, offering challenging benchmarks for the research community to address real-world defect detection. Additionally, we propose SegAD, an innovative supervised anomaly detection method that achieves state-of-the-art performance on VAD as well as on a supervised benchmark on standard VisA dataset. Experimental results reveal the limitations of existing supervised anomaly detection methods in handling complex problems, while the integration of SegAD with top-performing one-class anomaly detection models further enhances their results." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17761" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Kilian Batzner, Lars Heckler, and Rebecca König. Efficientad: Accurate visual anomaly detection at millisecond-level latencies, 2023. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.212 + ], + "angle": 0, + "content": "[2] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad — a comprehensive real-world dataset for unsupervised anomaly detection. In CVPR, pages 9584–9592, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.268 + ], + "angle": 0, + "content": "[3] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. IJCV, 130(4):947-969, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[4] Jakob Božić, Domen Tabernik, and Danijel Skočaj. Mixed supervision for surface-defect detection: from weakly to fully supervised learning. Computers in Industry, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.352 + ], + "angle": 0, + "content": "[5] Andrew P. Bradley. The use of the area under the roc curve in the evaluation of machine learning algorithms. PR, 30(7): 1145-1159, 1997. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.471, + 0.382 + ], + "angle": 0, + "content": "[6] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In SIGKDD, pages 785-794, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.471, + 0.422 + ], + "angle": 0, + "content": "[7] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In CVPR, pages 9737-9746, 2022. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.465 + ], + "angle": 0, + "content": "[8] Chouro Ding, Guansong Pang, and Chunhua Shen. Catching both gray and black swans: Open-set supervised anomaly detection. In CVPR, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.471, + 0.52 + ], + "angle": 0, + "content": "[9] Colin Drury and Murray Sinclair. Human and machine performance in an inspection task. Human Factors: The Journal of the Human Factors and Ergonomics Society, 25:391-399, 1983. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.522, + 0.47, + 0.578 + ], + "angle": 0, + "content": "[10] Lars Heckler, Rebecca König, and Paul Bergmann. Exploring the importance of pretrained feature extractors for unsupervised anomaly detection and localization. In CVPRW, pages 2917-2926, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.579, + 0.469, + 0.618 + ], + "angle": 0, + "content": "[11] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In ICLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.47, + 0.689 + ], + "angle": 0, + "content": "[12] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, pages 4015-4026, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[13] Alex Krizhevsky et al. Learning multiple layers of features from tiny images. 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[14] Yann LeCun, Corinna Cortes, Chris Burges, et al. Mnist handwritten digit database, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[15] Jiaqi Liu, Guoyang Xie, Jingbao Wang, Shangnian Li, Chengjie Wang, Feng Zheng, and Yaochu Jin. Deep industrial image anomaly detection: A survey. arXiv e-prints, pages arXiv-2301, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[16] Muriel Mazzetto, Marcelo Teixeira, Érick Oliveira Rodrigues, and Dalcimar Casanova. Deep learning models for visual inspection on automotive assembling line. arXiv preprint arXiv:2007.01857, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[17] Declan McIntosh and Alexandra Branzan Albu. Interrealization channels: Unsupervised anomaly detection in images beyond one-class classification, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[18] Yohei Mishina, Masamitsu Tsuchiya, and Hironobu Fujiyoshi. Boosted random forest. In VISAPP, pages 594-598, 2014. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[19] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. VT-ADL: A vision transformer network for image anomaly detection and localization. In ISIE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[20] Guansong Pang, Chunhua Shen, and Anton van den Hengel. Deep anomaly detection with deviation networks. In SIGKDD, pages 353-362, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[21] Guansong Pang, Choubo Ding, Chunhua Shen, and Anton van den Hengel. Explainable deep few-shot anomaly detection with deviation networks. arXiv preprint arXiv:2108.00462, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[22] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In ICML, pages 1530-1538. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[23] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Scholkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In CVPR, pages 14318-14328, 2022. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[24] Lukas Ruff, Robert A Vandermeulen, Nico Gornitz, Alexander Binder, Emmanuel Müller, Klaus-Robert Müller, and Marius Kloft. Deep semi-supervised anomaly detection. In ICLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[25] Maximilian Schwab, Charles Madeline-Dérou, Steffen Klarmann, Nils Thielen, Sven Meier, Jörg Franke, Sandan Chintanippu, and Wilhelm Stork. Multi-model machine learning based industrial vision framework for assembly part quality control. In *ETFA*, pages 1-4, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[26] Domen Tabernik, Samo Šela, Jure Skvarč, and Danijel Skočaj. Segmentation-based deep-learning approach for surface-defect detection. In Journal of Intelligent Manufacturing, 31(3):759-776, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.614 + ], + "angle": 0, + "content": "[27] Matthias Wieler and Tobias Hahn. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, 2007. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[28] Xincheng Yao, Ruoqi Li, Jing Zhang, Jun Sun, and Chongyang Zhang. Explicit boundary guided semi-push-pull contrastive learning for supervised anomaly detection. In CVPR, pages 24490-24499, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[29] Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows, 2021. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[30] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC. British Machine Vision Association, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[31] Hui Zhang, Zuxuan Wu, Zheng Wang, Zhineng Chen, and Yu-Gang Jiang. Prototypical residual networks for anomaly detection and localization. In CVPR, pages 16281-16291, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.83, + 0.892, + 0.885 + ], + "angle": 0, + "content": "[32] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In ECCV, pages 392-408. Springer, 2022. 1, 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.885 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17762" + } + ] +] \ No newline at end of file diff --git a/2024/Supervised Anomaly Detection for Complex Industrial Images/ef358e69-3d45-4c6c-aa99-f5a5ceb11e08_origin.pdf b/2024/Supervised Anomaly Detection for Complex Industrial Images/ef358e69-3d45-4c6c-aa99-f5a5ceb11e08_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fd47a73f2521722dc2e259fdf928271ca5788b5b --- /dev/null +++ b/2024/Supervised Anomaly Detection for Complex Industrial Images/ef358e69-3d45-4c6c-aa99-f5a5ceb11e08_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e24086b13ce40a57abc55f35a57f7087b5600a2b5d2d97db9a5a0e1b6491d47 +size 7590767 diff --git a/2024/Supervised Anomaly Detection for Complex Industrial Images/full.md b/2024/Supervised Anomaly Detection for Complex Industrial Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b83aac8c45d84e6613f04ffe6126b8665784db83 --- /dev/null +++ b/2024/Supervised Anomaly Detection for Complex Industrial Images/full.md @@ -0,0 +1,275 @@ +# Supervised Anomaly Detection for Complex Industrial Images + +Aimira Baitieva1, David Hurych2, Victor Besnier2, Olivier Bernard1 + +$^{1}$ valeo, $^{2}$ valeo.ai + +{aimira.baitieva, david.hurych, victor.besnier, olivier.bernard}@valeo.com + +# Abstract + +Automating visual inspection in industrial production lines is essential for increasing product quality across various industries. Anomaly detection (AD) methods serve as robust tools for this purpose. However, existing public datasets primarily consist of images without anomalies, limiting the practical application of AD methods in production settings. To address this challenge, we present (1) the Valeo Anomaly Dataset (VAD), a novel real-world industrial dataset comprising 5000 images, including 2000 instances of challenging real defects across more than 20 subclasses. Acknowledging that traditional AD methods struggle with this dataset, we introduce (2) Segmentation-based Anomaly Detector (SegAD). First, SegAD leverages anomaly maps as well as segmentation maps to compute local statistics. Next, SegAD uses these statistics and an optional supervised classifier score as input features for a Boosted Random Forest (BRF) classifier, yielding the final anomaly score. Our SegAD achieves state-of-the-art performance on both VAD (+2.1% AUROC) and the VisA dataset (+0.4% AUROC). The code and the models are publicly available1. + +# 1. Introduction + +Within the manufacturing process, industrial visual inspection plays a crucial role in identifying defects in produced components. This operation holds significant importance in minimizing costs by identifying and removing faulty parts early in the production stages and, more importantly, in preventing the dispatch of defective components to customers. Traditionally, this task has relied on human operators; however, the likelihood of overlooking certain defects can be as high as $25\%$ for specific defect types [9]. + +When the inspected product comprises numerous components, its examination may create a bottleneck in the production process, causing delays across the entire line. While + +conventional computer vision methods applied to this task demonstrate superior speed and lower error rates compared to human operators [16], their inflexibility and lack of satisfactory accuracy [15] limit their effectiveness. + +Industrial deep learning anomaly detection has been an active research field in recent years. Most of anomaly detection methods use only good images for training and try to detect deviations from the training data [1, 7, 23, 29]. Development of the new methods is restrained by publicly available datasets. Recent industrial anomaly detection datasets typically contain approximately one hundred (or even fewer) abnormal images, showcasing defects in the testing set only [2, 32]. This poses a challenge for supervised anomaly detection methods aiming to utilize both normal and defective parts during training. Supervised models often undergo training with just ten abnormal images from the testing set, resulting in overfitting and reduced sensitivity to previously unseen defects [28]. + +In response to the limitations of current datasets, we introduce and publicly release VAD (Valeo Anomaly Dataset), which contains 1000 bad and 2000 good parts in the training set and 1000 bad and 1000 good parts in the test set, see Fig. 1(a). Among the defective parts used for testing, 165 contain specific defect types not present in the training data; these parts are explicitly labeled in the released dataset. All images in VAD are captured from an actual production line, showcasing a diverse range of defects, from highly obvious to extremely subtle. This dataset bridges the gap between the academic community and the industry, offering researchers the chance to advance the performance of methods in tackling more intricate real-world challenges. + +Current approaches to supervised anomaly detection either yield unsatisfactory results; see Tab. 3, or demand pixel-level labels for defective parts, as seen in [28, 31]. Consequently, we introduce a novel method named SegAD (Segmentation-based Anomaly Detector), which is described in Fig. 1(b). The employed approach eliminates the need for pixel-level labels, requiring only a flag for each image. Anomaly map scores from each segment are used + +![](images/993389f2b9e558e13cbe5c56ca4f5b671efb80b9d3662e8a3e743b2ade7ee9ba.jpg) +(a) +(b) +Figure 1. Overview of our contributions. (a) VAD, a real-world industrial dataset designed for supervised anomaly detection with complex defects. (b) SegAD, our method that leverages anomaly maps extracted from segmented outputs of one or more anomaly detectors. Higher-level statistical features are computed on these maps, such as skewness, kurtosis, or mean, to generate the local anomaly features. Additionally, our SegAD provides the flexibility to use the output of a supervised classifier score with the local anomaly features, creating input for a final Boosted Random Forest (BRF) classifier that yields the final score. + +to delineate the parameters of the distribution, emphasizing relatively low scores in important areas while disregarding higher scores in noisy regions. In this context, the precise position and appearance of the anomaly diminishes in importance as long as it resides within a designated segment. Our main contributions can be summarized as follows: + +- We propose a supervised anomaly detection dataset with complex objects and a large variety of defects. As an addition, we establish a one-class anomaly detection benchmark, which is more challenging than most of the current ones, a supervised benchmark with a high number of bad images for training (1000 images) and a supervised benchmark with a low number of bad images for training (100 images). We share the dataset freely with the research community. +- We create a novel supervised anomaly detection method called SegAD. This approach performs noticeably better than recent anomaly detectors alone while retaining the ability to detect unknown defects. SegAD reaches SOTA results on VAD as well as on the established VisA dataset [32]. + +# 2. Related Work + +In this section, we first discuss existing datasets and explain how VAD is different. Next, we give a quick overview of existing anomaly detection methods. We show in Table 1 a global comparison between our VAD and existing datasets. + +# 2.1. Datasets + +In 2019, the MVTec AD dataset [2] was introduced. This dataset significantly advanced anomaly detection methods + +by providing images with authentic defects and realistic objects, in contrast to earlier datasets predominantly comprising textures and simple defects [4, 26, 27]. MVtec LOCO AD [3] highlighted the challenge of logical defects, such as missing or misplaced parts, necessitating anomaly detection methods capable of capturing the global context of the image. The more recent VisA dataset [32] expands on this by incorporating diverse objects with complex structures, multiple instances, and variations in scale. + +Existing datasets often contain simulated defects [2, 3, 32], creating a domain gap between research and practical applications. This complicates the deployment of anomaly detection methods in real-world settings. To promote supervised anomaly detection, we propose the Valeo Anomaly Dataset (VAD), a real-world industrial dataset with various defects, including logical ones. + +Another notable issue stems from the saturation of solved datasets, where other methods are influenced more by dataset-specific design choices than general applicability. Compared to this, our dataset introduces a broader range of challenging and diverse anomalies and high intra-class variability of good images. + +# 2.2. Methods + +To define the terms we are working with, "one-class" methods are commonly called unsupervised in the context of anomaly detection [10, 15, 29]. We use "one-class" as a more accurate name since these methods use labeled images for training. It is important to distinguish such methods from ones that work with unlabelled data, as in [17]. Models that use bad and good images for training are usually called "supervised", which is the name we also adopt. + +Table 1. Comparison of different anomaly detection datasets. Tr. good and Tr. bad show the average number of images per class for training. tex = texture, obj = object, Log. def. = Logical defects, Pix. labels = pixel-level labels and wl = weakly labelled (ellipse around the defect). + +
DatasetYearClassesImagesTypesDefectsTr. goodTr. badLog. def.Pix. labels
DAGM [27]20071011500texsynthetic50075nowl
KSSD [26]20191399texreal-world11517noyes
MVTec AD [2]2019155354tex, objsimulated2420noyes
KSSD2 [4]202113335texreal-world2085246noyes
BTAD [19]202132830tex, objreal-world6000noyes
LOCO AD [3]202253644objsimulated3540yesyes
VisA [32]20221210821objsimulated7200noyes
VAD (ours)202315000objreal-world20001000yesno
+ +One-class Anomaly Detection. PatchCore [23] relies on a pretrained feature extractor model to extract features from the training set into a memory bank and reduce the size of a memory bank using coreset subsampling. Features extracted from the new input are compared to their nearest neighbors in the memory bank. FastFlow [29] uses a similar feature extractor model and maps extracted features to the Gaussian distribution using normalizing flow [22]. RD4AD [7] utilizes a teacher-student architecture, which combines features from several layers of feature extractor to eliminate redundant ones. EfficientAD [1] introduces many innovations, including using an autoencoder to detect logical defects as an addition to a classic teacher-student architecture and their own pretrained feature extractor, which imitates the behavior of a bigger model with a noticeably lesser inference time. + +Supervised Anomaly Detection. In theory, leveraging defective parts for training is advantageous for refining class boundaries [24]. However, practical challenges arise, such as the small size of anomalies and the impossibility of collecting all potential defects [31]. Several recent anomaly detection methods use both good and bad images [21, 28, 31], but some of them overfit to seen defects, and others require pixel masks for defects to calculate loss or to generate new defects, which can be problematic in a real-world case. Real-world datasets (as VAD) might contain logical defects that cannot be generated by existing defects generation strategies, which usually cut and paste existing defects with various modifications [28, 31], which is not going to improve results for a wrong wire shape as or other similar types of defects. + +Very few supervised anomaly detection benchmarks are available; the most popular is Supervised Anomaly Detection on MVTec AD. DevNet [21] is one of the first supervised anomaly detection models that tried to solve industrial MVTec AD dataset, compared to earlier methods which were mostly addressing Out-of-Distribution problem [11, 24] using non-industrial datasets such as MNIST [14] + +or CIFAR-10 [13]. DRA [8] uses several heads to learn both seen and pseudo anomalies, as well as normal examples, which should reduce overfitting on seen anomalies. These methods perform well on MVtec AD but might fail on more complex problems, as shown in Tab. 3 and Tab. 5. Compared to modern supervised anomaly detection methods, SegAD (ours) performs better than current SOTA methods, even on complex problems without pixel-level labels, extensive augmentation, and long training. + +# 3. Valeo Anomaly Dataset (VAD) + +VAD consists of one class with predefined training and testing sets. The training set contains 1000 bad and 2000 good images, and the testing set contains 1000 bad, 165 of them are unseen bad, and 1000 good images. Unseen bad images in the test dataset refer to several rare defect types that are not present in the training data, several examples of which are shown in Fig. 4. Having such images is important to avoid turning this anomaly detection problem into a supervised classification problem. Some images might have a thin black border at the bottom due to how they were filmed, an example can be seen in Fig. 2. Defects might occur in the whole area of the image. Image-level annotation is provided, but there is no pixel-level annotation due to the complexity of defects and the fact there is no exact position for missing or misplaced components. All images are $512 \times 512$ pixels in PNG format. Examples of all types of defects can be found in Appendix C. + +Good parts consist of two wires connected to two pins on one side and two soldering dots on the other, placed on a piezo. Piezo means piezoelectric element, a big round area under the other components. Wires should have the right shape, not too straight, too long, or bent too much. Wires should connect the soldering dots as close to their centers as possible. Some amount of deviation is allowed, but if the wire is too close to the side of a solder dot, it is a defect. Soldering should be well-rounded, not too small, and placed in the correct location, not overlapping with the + +![](images/d3105d75fc2c8b7cb52aa9267bc13e366d596b1b232487a26966f68dfde09f67.jpg) + +![](images/37a362d3d40e4339cdb5a9c8d810d436ffd2deb4ff3d73e53db6244eca149be3.jpg) + +![](images/b551e15ba0358def0273f3b67e614fb5d83867b6726c043a4ab65fc498a63a11.jpg) +Figure 2. Good parts. Small scratches on the piezo are allowed, as well as wire being closer to the side of the soldering. (VAD) + +![](images/f451ca90a852a8c1812aab828cdfd6d7cd2e49edb845d9198f7bbbcd8b58dcf1.jpg) + +![](images/595e8758846c627d9045bb394391f9ac01e448ede99dcb7f376de17dccee8e28.jpg) + +![](images/aaf296721c7a7c1d47b3de0bc562f81bc4297747d948a8cd2f52023ccb4d6da5.jpg) +Logical defects +Figure 3. Variety of defects: logical, from the top left: wire on the side, wire out of solder dot, missing wire, bad wire shape (bent too much). Structural: soldering paste pollution, crack on a piezo, burned area under the solder dot, broken piezo. (VAD) + +![](images/04fab07105a0f1a51830e5bd15f8fed0185abd6e05f19e935939e177f395a8ec.jpg) + +![](images/e27f961208d89d5fa335c0ded39c876204df4177a84e7abfe5ba808693a4298e.jpg) +Structural defects + +piezo contours. The wire should be properly connected to the pin without any pieces of it visible on the side of the pin. The piezo can have different textures. Any kind of pollution is considered to be an anomaly, which makes the part defective. + +Logical defects result from components being misplaced or misshapen rather than being damaged, e.g. bad wire shape, bad soldering shape, wrong wire position, wrong soldering position, and missing parts. Structural defects include various cracks and broken parts of the piezo, as well as a variety of burns and pollution. For both types of defects, see Fig. 3. The importance of logical defects is often overlooked in the existing datasets (excluding MVTec + +![](images/4bdb7c339de780f876a09b1a68620c6a1c8aadedf62f2bd1e3ed6978ede888c2.jpg) +Figure 4. Unseen defects for left to right: burned solder dot, burned wire, extra unnecessary wire. (VAD) + +![](images/53c3be7cf04fb199446ce0b90d8591a07cbc5fe7ebb1f183b67607b7ad3c31cd.jpg) + +![](images/1b815771df7936ec659412c3c916e7270a94a4c06f5193737f11f26d75326185.jpg) + +LOCO AD) because they are hard to imitate if the dataset is filmed in the lab, but also hard to detect. VAD provides the research community the opportunity to work with real logical constraints of a complex object. + +# 3.1.Benchmarks + +We establish three benchmarks for the VAD dataset. The first is one-class anomaly detection. As shown in Tab 2, even SOTA methods struggle to perform well on our dataset compared to other popular datasets, which leaves an opening for further research. For this benchmark, no bad images are used for training. The second benchmark is high-shot supervised anomaly detection, which uses all 1000 bad images from the training set. The last benchmark is low-shot supervised anomaly detection. Only 100 bad images from the training set are used. These images are randomly selected using 5 seeds, with the average result over all seeds being expected. The same 2000 good parts are used for the training in all benchmarks. + +# 4. Method + +SegAD works with image data at its input, and after inference, it is expected to output a score that describes a probability of anomaly appearance. Let $\mathcal{I} = \left[0,\dots ,255\right]^{W,H,3}$ be a space of all images of width $W$ , height $H$ and with 3 color channels. The $i$ -th image from a dataset is marked as $I^{(i)}\in \mathcal{I}$ . We denote $\mathbf{x}$ to be a 2D pixel coordinate vector. SegAD inference consists of three consecutive stages in a pipeline. First is anomaly map(s) calculation with optional supervised classifier score calculation. Next is the calculation of simple statistics from anomaly maps over individual segments. Lastly, the obtained statistics, and optionally also the classifier score, are used as input features for a BRF classifier [18] that delivers the final result. + +Let us denote $K$ as the number of used pixel-wise anomaly detectors and mark them as functions $f_{k}(I):\mathcal{I}\to \mathbb{R}^{W,H},k\in \{1,\ldots ,K\}$ . A supervised classifier is a function $g(I):\mathbb{R}^{W,H}\to \mathbb{R}$ . Next, we introduce a set of $L$ mutually exclusive segments $S = \{s_1,\dots,s_L\}$ , where $s_l\in [0,1]^{W,H}$ . + +After the calculation of anomaly maps for all $f_{k}(I)$ , the masks $s_l$ are used for segment-wise + +extraction of results, followed by the calculation of several statistics. The final feature vector $\mathbf{f} = [g(I),[\mathbf{q}^L ]_{k = 1}^K,[\mathbf{z}^L ]_{k = 1}^K,[\mathbf{c}^L ]_{k = 1}^K,[\mathbf{m}^L ]_{k = 1}^K]$ extracted from one image, is the concatenation of 4 vectors containing different statistics. $\mathbf{q}_k^L$ stands for a vector of $L$ numbers where each stands for a $99.5\%$ quantile from anomaly map $f_{k}(I)$ over pixels where $s_l(\mathbf{x}) = 1$ Similarly, for every anomaly map $f_{k}(I)$ we calculate a vector for skew $\mathbf{z}_k^L$ , kurtosis $\mathbf{c}_k^L$ and mean $\mathbf{m}_k^L$ . The length of the feature vector $\mathbf{f}$ equals to $K*L*4 + 1$ + +For notation simplicity, we keep the supervised classifier score $g(I)$ always present in $\mathbf{f}$ , but it may be omitted, which is clearly defined in every experiment setup. Feature vectors $\mathbf{f}^{(i)}$ are used as inputs for training the final classifier [18], which, at inference time, delivers the final result. + +# 4.1. Training + +Training SegAD requires two separate subsets from the training dataset. The first subset is used to train base models $f_{k}$ , the second subset is used to train the BRF classifier. To create these subsets from one training dataset, we split available good images into two parts. For VAD, we report average results over 5 different splits to show more reliable results. + +Each model $(f_{k}, g, \mathrm{BRF})$ is trained independently. Training can be separated into two stages. In the first stage, anomaly detector(s) $f_{k}$ are trained using the same data. In the same stage, a segmentation map $S$ must be defined. As will be elaborated in more detail in Sec. 5, the segmentation map can be static or produced by a segmentation model, depending on the dataset. A supervised classifier $g$ is trained separately on $80\%$ of the full training set (1600 good and 800 bad parts for VAD) to ensure that the training sets for this model and SegAD are not fully overlapping. The next stage is to train a final BRF classifier in SegAD, which requires both good and bad images, bad images can be replaced with artificial defects in some cases, as shown in Tab. 2. BRF is trained with the output of the models from the previous stage. + +An important thing to note is that SegAD is very fast to train as long as you already have other models available. Extraction of features and training the final classifier takes noticeably less time than training anomaly detectors. The same is true for the inference; SegAD adds a negligible time to processing the image while improving results significantly. + +# 4.2. Setups for SegAD + +SegAD can use any anomaly detector(s) $f_{k}$ as long as it produces an anomaly map showing pixel-level anomaly scores, although, in this paper, we use only one-class anomaly detectors because they require less data to train. Any supervised classifier $g$ that returns a classification score can be + +used. In experiments, we use several setups, which will be described here in detail. SegAD is denoted as Ours. Supervised classifier $g$ is not used unless stated otherwise. + +Anomaly Detector + Ours means $K = 1$ , $f_{1} =$ "Anomaly Detector name". This setup allows us to show improvement for one anomaly detection model in particular and demonstrate that this improvement can be achieved for different types of anomaly detectors. + +All $\mathbf{AD} + \mathbf{Ours}$ denotes that all anomaly detectors from the list $AllAD$ of length $N$ are used to produce anomaly maps. Such a list can be found in the caption of the table. In that case $K = N$ , $f_{1} = AllAD_{1},\ldots ,f_{N} = AllAD_{N}$ . Anomaly detectors might use an ensemble of feature extractors, similarly to [23], which makes them slower but can improve the result. In our case, anomaly detectors themselves can be successfully assembled in a supervised way. + +All AD + Supervised Classifier + Ours denotes a setup in which $g =$ "Supervised Classifier name", and the rest is the same as in the previous setup. A supervised classifier allows us to detect seen defects, which can be invisible to a one-class classifier, similarly to [25]. Such a setup shows the best result for academic purposes, but the speed of inference can be unsatisfying for real-world applications. We expect that, in that case, the number and selection of anomaly detectors can be optimized to fit the problem as well as the time constraints. + +Anomaly Detector + Supervised Classifier + Ours mean a setup in which $g =$ "Supervised Classifier name", $K = 1$ , $f_{1} =$ "Anomaly Detector name". It shows an example of an optimization mentioned in the previous setup. Only one anomaly detector gives a higher inference speed compared to the ensemble of anomaly detectors while giving a satisfying result. + +# 5. Experiments + +Evaluation metrics. We compare results using the following metrics: (1) Area Under the Receiver Operating Characteristic Curve (AUROC) [5], (2) False Positive Rate (FPR) at $95\%$ True Positive Rate (TPR) denoted as FPR@95TPR. Classification AUROC, which describes the performance on the image level for multiple thresholds, denoted by Cl. AUROC. FPR@95TPR shows the percentage of misclassified good parts (false positive) at $95\%$ of bad parts classified correctly (true positive). We report mean results $\pm$ standard deviation. + +Implementation and evaluation details. SegAD uses the output of one-class detectors PatchCore [23], FastFlow [29], Reverse Distillation [7] (referred to as RD4AD), EfficientAD [1] either alone or several anomaly detectors at once. SegAD results are compared to the results of these models alone, as well as supervised anomaly detectors DevNet [21] and DRA [8]. We also compare our results to the supervised classifier Wide ResNet50 [30] (denoted as + +WRN). The final classifier BRF was trained with the same parameters for all datasets using XGBoost library [6]. More details on the implementation can be found in Appendixes A and B. + +Images are resized to $\mathrm{W} = \mathrm{H} = 256$ pixels to make the comparison less biased. The DRA model originally uses a resolution $448 \times 448$ , which shows better results on VAD than $256 \times 256$ , 92.8 CI. AUROC, which is still lower than the top results. The higher resolution also improves results for some of the other models, so we do not use it. $256 \times 256$ is sufficient for defects to remain visible, but it also allows to have a reasonable speed of inference, especially for anomaly detection models. + +For VAD, results averaged for 5 seeds and training runs are reported. These seeds are used for training and splitting the training set for SegAD into two sets, as described in Subsection 4.1. 2000 good parts from the train set were split in half: 1000 for training the base model(s) and 1000 for training SegAD. 0, 1000, or 100 bad parts (depending on the benchmark) were also used to train SegAD. Details on the VisA benchmark can be found in Subsection 5.4. No augmentations were used for training, and existing augmentations were removed from used methods. Augmentations tend to be suitable only for specific tasks (as rotation in DRA or DevNet, removing it improved results by these models on the VAD drastically), and using augmentations makes it hard to compare models themselves. + +We include three benchmarks for the new VAD and one additional for the VisA dataset [32]. We compare results for VAD for one-class, high-shot, and low-shot supervised benchmarks, as described in Subsection 3.1. The benchmark for the VisA dataset is explained in 5.4. The segmen + +![](images/abc22acaf126607a77657c5ce1b4aea02cd5a88f9c5bedd859b46e19176465f4.jpg) +Figure 5. Static segmentation map for objects in VAD dataset, overlayed over the image. The image is separated into $L = 7$ segments: background, outer half-circle, piezo border, solder dots area, wires area, pins area, and piezo in the middle. + +tation maps for VAD is the same static image for every input image from the dataset. The segmentation map can be seen in Fig. 5. Due to the fact that the object in the VAD is al + +ways centered with the constant size and component placement, this is the simplest choice, which may be sub-optimal, but serves the purpose and does not require any manual or model-inferred annotation of the data. It was created to separate the anomaly maps into several meaningful parts, such as piezo border, wire areas, soldering area, etc., which can contain different defects and different levels of falsely high scores on the anomaly map. Another source of segmentation maps can be a specially trained segmentation model or a zero-shot segmentation model. + +# 5.1. VAD, one-class benchmark + +This benchmark allows only good images for training, but SegAD requires bad images as well, so we try to replace them with artificially generated defects. Because SegAD works with anomaly maps, generating realistic defects is unnecessary. For this reason, we have applied several simple augmentations to the 1000 good images available for training SegAD. These augmentations include Gaussian blur or a randomly placed, randomly sized thin rectangle of a random shade of gray. They were applied to randomly selected segments or random parts of the segments; segments are defined with the segmentation map. Code for generating defects can be found in the SegAD GitHub repository. + +
MethodCl. AUROC ↑FPR@95TPR ↓
One-class Anomaly Detection
FastFlow79.1±1.776.2±1.9
RD4AD84.5±0.265.8±2
PatchCore88.3±0.360.4±2.8
EfficientAD89.1±0.844.5±2.2
SegAD (Ours)
FastFlow + Ours83.3±1.371.5±5
RD4AD + Ours87.5±1.261.0±7.9
PatchCore + Ours89.7±0.858.8±3.7
EfficientAD + Ours85.6±1.6100±0
All AD + Ours90.4±0.552.4±4.5
+ +Table 2. One-class benchmark (VAD). The best result is marked in bold. All AD means PatchCore, FastFlow, and RD4AD. + +Results in Table 2 show that even such a naive approach improves results compared to anomaly detection methods alone. Even greater improvement can be achieved by using several anomaly detectors to produce anomaly maps for SegAD. This strategy fails for EfficientAD + Ours, anomaly maps produced by EfficientAD for the generated defects are too different from anomaly maps for real defects, creating a large discrepancy between training and test distributions for SegAD and causing worse results compared to anomaly detector alone. + +# 5.2. VAD, high-shot supervised benchmark + +
MethodCl. AUROC ↑FPR@95TPR ↓
Supervised Anomaly Detection
DevNet86.9±1.266.6±2.7
DRA87.4±0.660.9±4.4
Supervised Classifier
WRN95.0±0.632.8±4.3
SegAD (Ours)
FastFlow + Ours86.9±0.556.9±3.9
RD4AD + Ours90.2±0.346.9±2.6
PatchCore + Ours91.7±0.348.0±2.9
EfficientAD + Ours91.4±0.236.8±1
AllAD+WRN+Ours96.5±0.316.4±1.9
+ +Table 3. High-shot supervised benchmark (VAD). The best result is marked in bold. All AD means PatchCore, FastFlow, RD4AD, and EfficientAD. Improvement calculated compared to base method results in Tab. 2. + +The high-shot benchmark allows training supervised classifier $g$ , which shows competitive results as can be seen in Tab. 3 in the row "WRN", but it won't be able to detect unseen defects. Histograms in Fig. 6 visualize this problem. On the left is the one-class anomaly detector, which can detect unseen defects (in red) but shows a bad separation between good and bad parts. In the center, the supervised classifier $g$ shows a better separation of good and bad parts, but unseen defects distribution is shifted to the left because their scores are relatively low and such defects cannot be detected. On the right, SegAD shows a good separation between classes, but also unseen defects have a similar distribution to the seen defects. We still compare the supervised classifier Wide ResNet50 (WRN) with other methods, but the ability to detect unseen defects is crucial in real-world applications. + +![](images/b6e41dc57d33938baf62ad26303d3a531562e8669653edc7e43b6250cb64a215.jpg) +(a) + +![](images/87b12149746b30f361d2d85e89cd9b02146e549bd0f99c86b0923721ffda84d7.jpg) +(b) +Figure 6. Distributions of scores in the VAD test set. Green color shows good parts, orange bad parts, and red bad parts with unseen defects. (a) shows one-class anomaly detector PatchCore, (b) supervised classifier Wide ResNet, (c) SegAD (PatchCore + WRN + Ours). X axis is score values, Y axis is frequency. + +![](images/6f3e7fed76afdc8428586670d383e9aabc288c58284f41305df1341e99ce7d08.jpg) +(c) + +Combining SegAD, supervised classifier WRN, and all one-class anomaly detection models showed the best results + +presented in Table 3. A more practical setting can be EfficientAD + WRN + Ours, which can achieve Cl. AUROC 96 and FPR@95TPR 19.4 with a much higher inference speed. + +# 5.3. VAD, low-shot supervised benchmark + +This benchmark uses just 100 bad images for training, which can be a relatively low number for a dataset with more than 20 types of defects. However, such a setting is closer to real-world applications, where a limited amount of defective images can be available. In Table 4, WRN shows much lower results due to a smaller training dataset, as well as DevNet and DRA. SegAD also performs worse, than with a high-shot benchmark, but it still can visibly improve the performance of other models. We do not use WRN in SegAD for this benchmark, because it showed low results by itself. + +
MethodCl. AUROC ↑FPR@95TPR ↓
Supervised Anomaly Detection
DevNet73.1±2.388.3±2.3
DRA78.9±1.874.6±1.9
Supervised Classifier
WRN78.7±1.877.9±4.4
SegAD (ours)
FastFlow + Ours84.3±1.464.5±2.5
RD4AD + Ours88.6±0.251.1±0.9
PatchCore + Ours90.8±0.552.7±4.3
EfficientAD + Ours90.6±0.740.1±2.4
All AD + Ours92.7±0.536.8±3.1
+ +Table 4. Low-shot supervised benchmark (VAD). The best result is marked in bold. All AD means PatchCore, FastFlow, RD4AD, and EfficientAD. + +# 5.4. Results on VisA dataset + +VisA contains fewer good images per class for training than VAD (500-1000 in VisA, compared to 2000 in VAD). The training set consists of good images only. For this reason, we use a different strategy to split the training dataset to make sure we have enough images to train the base anomaly detection model with a competitive result. For RD4AD, the training set was divided into $90\%$ of images to train the RD4AD itself and $10\%$ to train SegAD. EfficientAD uses a validation set ( $10\%$ of good images from the train set) to normalize anomaly maps. In our case, the same images were used to train SegAD. The segmentation maps for this dataset were created with the Segment Anything Model (SAM) [12]. For most classes, it is a binary mask for the object and background, and for pcb1-4, static components were mapped on top of it; see Fig. 7. + +![](images/faed8c248e9be65f9aa2750548993ae5d11f7d6f05544f5a94410540b5917c72.jpg) +(a) + +![](images/ce9cb5d1b85db512db2232c5280e543f58cecde19401f5aa932d5dc510f8ca9f.jpg) +(b) + +![](images/a0ea3a90541364f5f291409e93ce9176edf77ffd775758bdf9f5658ed918d6c9.jpg) +(c) +Figure 7. VisA. Examples of segmentation maps created using SAM for (a) cashew, (b) macaroni1, (c) pb2. + +
MethodCl. AUROC↑FPR@95TPR↓
One-class Anomaly Detection
RD4AD94.225.5
EfficientAD98.08.1
Supervised Anomaly Detection
DRA88.942.4
DevNet89.344.0
SegAD (ours)
RD4AD + Ours96.115.5
EfficientAD + Ours98.36.2
All AD + Ours98.47.5
+ +With VisA, we create a new, challenging, supervised benchmark on a public dataset with 12 different classes. In a similar way to the supervised benchmark on MVTec AD [20], we move 10 randomly selected bad images from the test set to the train set, reporting results over 10 different seeds. In Table 5, this benchmark shows how SegAD performs on a different data from VAD and creates a new possibility to compete on a difficult supervised anomaly detection problem. Detailed results per class can be found in Appendix D. + +# 5.5. Ablation study + +We perform an ablation study on the VAD high-shot supervised benchmark. We use the average result of the Patch-Core, FastFlow, RD4AD, and EfficientAD as the baseline in the Table 6, denoted as "An.Det." These anomaly detectors use the maximum value from the anomaly map to calculate the score. On the contrary, we compute our proposed features $\mathbf{f}$ , which are described in Section 4, and use BRF to yield a final score. This shows an improvement of 1.3 CI. AUROC in the row "One Seg." In addition, we evaluate the segmentation map's impact by computing the anomaly map's maximum value per segment. Doing this gives an im + +provement in 2.1 Cl. AUROC, which can be seen in the row "Max." This shows how both features and segmentation maps are important for SegAD. We also show the strength of BRF by comparing it against other algorithms. We compare BRF with BT and RF using a segmentation map and features. The results show that the BRF is better than these methods by 1.7 and 0.3 Cl. AUROC, respectively. + +Table 5. Results on VisA dataset, supervised benchmark. The best result is marked in bold. All AD means RD4AD and EfficientAD. + +
DescriptionFeat. fSeg. SBRFCl. AUROC ↑
An. Det.87.3
One Seg.✓✓88.6
Max.✓✓89.4
BT✓✓88.4
RF✓✓89.8
SegAD✓✓✓90.1
+ +Table 6. Ablation study on the methodical differences between our method (SegAD) and anomaly detection methods. An.Det. = Anomaly Detectors. One Seg. = One Segment. Max. means using maximum value. BT = Boosted Tree. RF = Random Forest. BRF = Boosted Random Forest. SegAD denotes the average result from PatchCore + Ours, FastFlow + Ours, RD4AD + Ours, EfficientAD + Ours. + +Limitations: SegAD (ours) requires to have segmentation maps. Static maps can be used for objects like the product in VAD. Obtaining segmentation maps for unaligned objects can be more difficult, yet SAM [12] or a specially trained segmentation model may be a solution. The performance improvement of SegAD depends on the complexity and structure of the objects. For less complex objects in VisA, improvement is lower than for VAD, as can be seen in Tab. 3 and Tab. 5. Another possible limitation can be the difference in the anomaly maps between training and test data. Due to the nature of anomaly detection, this problem might be impossible to mitigate, because our task is to detect such differences. + +# 6. Conclusion + +This study introduces VAD, a brand new supervised anomaly detection dataset derived from real production, offering challenging benchmarks for the research community to address real-world defect detection. Additionally, we propose SegAD, an innovative supervised anomaly detection method that achieves state-of-the-art performance on VAD as well as on a supervised benchmark on standard VisA dataset. Experimental results reveal the limitations of existing supervised anomaly detection methods in handling complex problems, while the integration of SegAD with top-performing one-class anomaly detection models further enhances their results. + +# References + +[1] Kilian Batzner, Lars Heckler, and Rebecca König. Efficientad: Accurate visual anomaly detection at millisecond-level latencies, 2023. 1, 3, 5 +[2] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad — a comprehensive real-world dataset for unsupervised anomaly detection. In CVPR, pages 9584–9592, 2019. 1, 2, 3 +[3] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. IJCV, 130(4):947-969, 2022. 2, 3 +[4] Jakob Božić, Domen Tabernik, and Danijel Skočaj. Mixed supervision for surface-defect detection: from weakly to fully supervised learning. Computers in Industry, 2021. 2, 3 +[5] Andrew P. Bradley. The use of the area under the roc curve in the evaluation of machine learning algorithms. PR, 30(7): 1145-1159, 1997. 5 +[6] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In SIGKDD, pages 785-794, 2016. 6 +[7] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In CVPR, pages 9737-9746, 2022. 1, 3, 5 +[8] Chouro Ding, Guansong Pang, and Chunhua Shen. Catching both gray and black swans: Open-set supervised anomaly detection. In CVPR, 2022. 3, 5 +[9] Colin Drury and Murray Sinclair. Human and machine performance in an inspection task. Human Factors: The Journal of the Human Factors and Ergonomics Society, 25:391-399, 1983. 1 +[10] Lars Heckler, Rebecca König, and Paul Bergmann. Exploring the importance of pretrained feature extractors for unsupervised anomaly detection and localization. In CVPRW, pages 2917-2926, 2023. 2 +[11] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In ICLR, 2018. 3 +[12] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, pages 4015-4026, 2023. 7, 8 +[13] Alex Krizhevsky et al. Learning multiple layers of features from tiny images. 2009. 3 +[14] Yann LeCun, Corinna Cortes, Chris Burges, et al. Mnist handwritten digit database, 2010. 3 +[15] Jiaqi Liu, Guoyang Xie, Jingbao Wang, Shangnian Li, Chengjie Wang, Feng Zheng, and Yaochu Jin. Deep industrial image anomaly detection: A survey. arXiv e-prints, pages arXiv-2301, 2023. 1, 2 +[16] Muriel Mazzetto, Marcelo Teixeira, Érick Oliveira Rodrigues, and Dalcimar Casanova. Deep learning models for visual inspection on automotive assembling line. arXiv preprint arXiv:2007.01857, 2020. 1 +[17] Declan McIntosh and Alexandra Branzan Albu. Interrealization channels: Unsupervised anomaly detection in images beyond one-class classification, 2023. 2 + +[18] Yohei Mishina, Masamitsu Tsuchiya, and Hironobu Fujiyoshi. Boosted random forest. In VISAPP, pages 594-598, 2014. 4, 5 +[19] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. VT-ADL: A vision transformer network for image anomaly detection and localization. In ISIE, 2021. 3 +[20] Guansong Pang, Chunhua Shen, and Anton van den Hengel. Deep anomaly detection with deviation networks. In SIGKDD, pages 353-362, 2019. 8 +[21] Guansong Pang, Choubo Ding, Chunhua Shen, and Anton van den Hengel. Explainable deep few-shot anomaly detection with deviation networks. arXiv preprint arXiv:2108.00462, 2021. 3, 5 +[22] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In ICML, pages 1530-1538. PMLR, 2015. 3 +[23] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Scholkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In CVPR, pages 14318-14328, 2022. 1, 3, 5 +[24] Lukas Ruff, Robert A Vandermeulen, Nico Gornitz, Alexander Binder, Emmanuel Müller, Klaus-Robert Müller, and Marius Kloft. Deep semi-supervised anomaly detection. In ICLR, 2019. 3 +[25] Maximilian Schwab, Charles Madeline-Dérou, Steffen Klarmann, Nils Thielen, Sven Meier, Jörg Franke, Sandan Chintanippu, and Wilhelm Stork. Multi-model machine learning based industrial vision framework for assembly part quality control. In *ETFA*, pages 1-4, 2022. 5 +[26] Domen Tabernik, Samo Šela, Jure Skvarč, and Danijel Skočaj. Segmentation-based deep-learning approach for surface-defect detection. In Journal of Intelligent Manufacturing, 31(3):759-776, 2020. 2, 3 +[27] Matthias Wieler and Tobias Hahn. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, 2007. 2, 3 +[28] Xincheng Yao, Ruoqi Li, Jing Zhang, Jun Sun, and Chongyang Zhang. Explicit boundary guided semi-push-pull contrastive learning for supervised anomaly detection. In CVPR, pages 24490-24499, 2023. 1, 3 +[29] Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows, 2021. 1, 2, 3, 5 +[30] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC. British Machine Vision Association, 2016. 5 +[31] Hui Zhang, Zuxuan Wu, Zheng Wang, Zhineng Chen, and Yu-Gang Jiang. Prototypical residual networks for anomaly detection and localization. In CVPR, pages 16281-16291, 2023. 1, 3 +[32] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In ECCV, pages 392-408. Springer, 2022. 1, 2, 3, 6 \ No newline at end of file diff --git a/2024/Supervised Anomaly Detection for Complex Industrial Images/images.zip b/2024/Supervised Anomaly Detection for Complex Industrial Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1f23b38f3b958c0e4ba598c1385c46068efea0d3 --- /dev/null +++ b/2024/Supervised Anomaly Detection for Complex Industrial Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a91a515e004338d44a96787ccc15439c29f1f4d296b2900f2f12b98e98a1682 +size 459043 diff --git a/2024/Supervised Anomaly Detection for Complex Industrial Images/layout.json b/2024/Supervised Anomaly Detection for Complex Industrial Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cf14c32a9d897b3ac52b06fc9ba68ae6253729ee --- /dev/null +++ b/2024/Supervised Anomaly Detection for Complex Industrial Images/layout.json @@ -0,0 +1,7515 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 104, + 103, + 489, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 103, + 489, + 121 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 489, + 121 + ], + "type": "text", + "content": "Supervised Anomaly Detection for Complex Industrial Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 142, + 459, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 142, + 459, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 142, + 459, + 158 + ], + "type": "text", + "content": "Aimira Baitieva1, David Hurych2, Victor Besnier2, Olivier Bernard1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 255, + 170, + 337, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 170, + 337, + 184 + ], + "spans": [ + { + "bbox": [ + 255, + 170, + 337, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 255, + 170, + 337, + 184 + ], + "type": "text", + "content": "valeo, " + }, + { + "bbox": [ + 255, + 170, + 337, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 255, + 170, + 337, + 184 + ], + "type": "text", + "content": "valeo.ai" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 95, + 187, + 495, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 187, + 495, + 200 + ], + "spans": [ + { + "bbox": [ + 95, + 187, + 495, + 200 + ], + "type": "text", + "content": "{aimira.baitieva, david.hurych, victor.besnier, olivier.bernard}@valeo.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 252, + 290, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 290, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 290, + 502 + ], + "type": "text", + "content": "Automating visual inspection in industrial production lines is essential for increasing product quality across various industries. Anomaly detection (AD) methods serve as robust tools for this purpose. However, existing public datasets primarily consist of images without anomalies, limiting the practical application of AD methods in production settings. To address this challenge, we present (1) the Valeo Anomaly Dataset (VAD), a novel real-world industrial dataset comprising 5000 images, including 2000 instances of challenging real defects across more than 20 subclasses. Acknowledging that traditional AD methods struggle with this dataset, we introduce (2) Segmentation-based Anomaly Detector (SegAD). First, SegAD leverages anomaly maps as well as segmentation maps to compute local statistics. Next, SegAD uses these statistics and an optional supervised classifier score as input features for a Boosted Random Forest (BRF) classifier, yielding the final anomaly score. Our SegAD achieves state-of-the-art performance on both VAD (+2.1% AUROC) and the VisA dataset (+0.4% AUROC). The code and the models are publicly available1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 528, + 128, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 128, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 128, + 540 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 548, + 287, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 656 + ], + "type": "text", + "content": "Within the manufacturing process, industrial visual inspection plays a crucial role in identifying defects in produced components. This operation holds significant importance in minimizing costs by identifying and removing faulty parts early in the production stages and, more importantly, in preventing the dispatch of defective components to customers. Traditionally, this task has relied on human operators; however, the likelihood of overlooking certain defects can be as high as " + }, + { + "bbox": [ + 46, + 548, + 287, + 656 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 548, + 287, + 656 + ], + "type": "text", + "content": " for specific defect types [9]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 657, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 287, + 693 + ], + "type": "text", + "content": "When the inspected product comprises numerous components, its examination may create a bottleneck in the production process, causing delays across the entire line. While" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 228, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 545, + 276 + ], + "type": "text", + "content": "conventional computer vision methods applied to this task demonstrate superior speed and lower error rates compared to human operators [16], their inflexibility and lack of satisfactory accuracy [15] limit their effectiveness." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 278, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 546, + 445 + ], + "type": "text", + "content": "Industrial deep learning anomaly detection has been an active research field in recent years. Most of anomaly detection methods use only good images for training and try to detect deviations from the training data [1, 7, 23, 29]. Development of the new methods is restrained by publicly available datasets. Recent industrial anomaly detection datasets typically contain approximately one hundred (or even fewer) abnormal images, showcasing defects in the testing set only [2, 32]. This poses a challenge for supervised anomaly detection methods aiming to utilize both normal and defective parts during training. Supervised models often undergo training with just ten abnormal images from the testing set, resulting in overfitting and reduced sensitivity to previously unseen defects [28]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 448, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 547, + 616 + ], + "type": "text", + "content": "In response to the limitations of current datasets, we introduce and publicly release VAD (Valeo Anomaly Dataset), which contains 1000 bad and 2000 good parts in the training set and 1000 bad and 1000 good parts in the test set, see Fig. 1(a). Among the defective parts used for testing, 165 contain specific defect types not present in the training data; these parts are explicitly labeled in the released dataset. All images in VAD are captured from an actual production line, showcasing a diverse range of defects, from highly obvious to extremely subtle. This dataset bridges the gap between the academic community and the industry, offering researchers the chance to advance the performance of methods in tackling more intricate real-world challenges." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": "Current approaches to supervised anomaly detection either yield unsatisfactory results; see Tab. 3, or demand pixel-level labels for defective parts, as seen in [28, 31]. Consequently, we introduce a novel method named SegAD (Segmentation-based Anomaly Detector), which is described in Fig. 1(b). The employed approach eliminates the need for pixel-level labels, requiring only a flag for each image. Anomaly map scores from each segment are used" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 219, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 219, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 219, + 713 + ], + "type": "text", + "content": "1https://github.com/abc-125/segad" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17754" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 75, + 534, + 223 + ], + "blocks": [ + { + "bbox": [ + 65, + 75, + 534, + 223 + ], + "lines": [ + { + "bbox": [ + 65, + 75, + 534, + 223 + ], + "spans": [ + { + "bbox": [ + 65, + 75, + 534, + 223 + ], + "type": "image", + "image_path": "993389f2b9e558e13cbe5c56ca4f5b671efb80b9d3662e8a3e743b2ade7ee9ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 226, + 143, + 236 + ], + "lines": [ + { + "bbox": [ + 133, + 226, + 143, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 226, + 143, + 236 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 352, + 227, + 363, + 237 + ], + "lines": [ + { + "bbox": [ + 352, + 227, + 363, + 237 + ], + "spans": [ + { + "bbox": [ + 352, + 227, + 363, + 237 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 249, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 249, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 545, + 304 + ], + "type": "text", + "content": "Figure 1. Overview of our contributions. (a) VAD, a real-world industrial dataset designed for supervised anomaly detection with complex defects. (b) SegAD, our method that leverages anomaly maps extracted from segmented outputs of one or more anomaly detectors. Higher-level statistical features are computed on these maps, such as skewness, kurtosis, or mean, to generate the local anomaly features. Additionally, our SegAD provides the flexibility to use the output of a supervised classifier score with the local anomaly features, creating input for a final Boosted Random Forest (BRF) classifier that yields the final score." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 326, + 286, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 286, + 396 + ], + "type": "text", + "content": "to delineate the parameters of the distribution, emphasizing relatively low scores in important areas while disregarding higher scores in noisy regions. In this context, the precise position and appearance of the anomaly diminishes in importance as long as it resides within a designated segment. Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 399, + 286, + 578 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 399, + 286, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 286, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 286, + 506 + ], + "type": "text", + "content": "- We propose a supervised anomaly detection dataset with complex objects and a large variety of defects. As an addition, we establish a one-class anomaly detection benchmark, which is more challenging than most of the current ones, a supervised benchmark with a high number of bad images for training (1000 images) and a supervised benchmark with a low number of bad images for training (100 images). We share the dataset freely with the research community." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 506, + 286, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 286, + 578 + ], + "type": "text", + "content": "- We create a novel supervised anomaly detection method called SegAD. This approach performs noticeably better than recent anomaly detectors alone while retaining the ability to detect unknown defects. SegAD reaches SOTA results on VAD as well as on the established VisA dataset [32]." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 591, + 134, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 134, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 134, + 604 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 613, + 286, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 286, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 286, + 660 + ], + "type": "text", + "content": "In this section, we first discuss existing datasets and explain how VAD is different. Next, we give a quick overview of existing anomaly detection methods. We show in Table 1 a global comparison between our VAD and existing datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 670, + 109, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 109, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 109, + 681 + ], + "type": "text", + "content": "2.1. Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "In 2019, the MVTec AD dataset [2] was introduced. This dataset significantly advanced anomaly detection methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 325, + 545, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 432 + ], + "type": "text", + "content": "by providing images with authentic defects and realistic objects, in contrast to earlier datasets predominantly comprising textures and simple defects [4, 26, 27]. MVtec LOCO AD [3] highlighted the challenge of logical defects, such as missing or misplaced parts, necessitating anomaly detection methods capable of capturing the global context of the image. The more recent VisA dataset [32] expands on this by incorporating diverse objects with complex structures, multiple instances, and variations in scale." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 434, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 517 + ], + "type": "text", + "content": "Existing datasets often contain simulated defects [2, 3, 32], creating a domain gap between research and practical applications. This complicates the deployment of anomaly detection methods in real-world settings. To promote supervised anomaly detection, we propose the Valeo Anomaly Dataset (VAD), a real-world industrial dataset with various defects, including logical ones." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 518, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 589 + ], + "type": "text", + "content": "Another notable issue stems from the saturation of solved datasets, where other methods are influenced more by dataset-specific design choices than general applicability. Compared to this, our dataset introduces a broader range of challenging and diverse anomalies and high intra-class variability of good images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 598, + 369, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 369, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 369, + 609 + ], + "type": "text", + "content": "2.2. Methods" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "To define the terms we are working with, \"one-class\" methods are commonly called unsupervised in the context of anomaly detection [10, 15, 29]. We use \"one-class\" as a more accurate name since these methods use labeled images for training. It is important to distinguish such methods from ones that work with unlabelled data, as in [17]. Models that use bad and good images for training are usually called \"supervised\", which is the name we also adopt." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17755" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 112, + 547, + 229 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "type": "text", + "content": "Table 1. Comparison of different anomaly detection datasets. Tr. good and Tr. bad show the average number of images per class for training. tex = texture, obj = object, Log. def. = Logical defects, Pix. labels = pixel-level labels and wl = weakly labelled (ellipse around the defect)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 112, + 547, + 229 + ], + "lines": [ + { + "bbox": [ + 47, + 112, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 47, + 112, + 547, + 229 + ], + "type": "table", + "html": "
DatasetYearClassesImagesTypesDefectsTr. goodTr. badLog. def.Pix. labels
DAGM [27]20071011500texsynthetic50075nowl
KSSD [26]20191399texreal-world11517noyes
MVTec AD [2]2019155354tex, objsimulated2420noyes
KSSD2 [4]202113335texreal-world2085246noyes
BTAD [19]202132830tex, objreal-world6000noyes
LOCO AD [3]202253644objsimulated3540yesyes
VisA [32]20221210821objsimulated7200noyes
VAD (ours)202315000objreal-world20001000yesno
", + "image_path": "a775f8af6adf8393a567976ff47e98ded84b505b7520616782728c004a1beada.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 247, + 289, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 289, + 438 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 289, + 438 + ], + "type": "text", + "content": "One-class Anomaly Detection. PatchCore [23] relies on a pretrained feature extractor model to extract features from the training set into a memory bank and reduce the size of a memory bank using coreset subsampling. Features extracted from the new input are compared to their nearest neighbors in the memory bank. FastFlow [29] uses a similar feature extractor model and maps extracted features to the Gaussian distribution using normalizing flow [22]. RD4AD [7] utilizes a teacher-student architecture, which combines features from several layers of feature extractor to eliminate redundant ones. EfficientAD [1] introduces many innovations, including using an autoencoder to detect logical defects as an addition to a classic teacher-student architecture and their own pretrained feature extractor, which imitates the behavior of a bigger model with a noticeably lesser inference time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 445, + 289, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 289, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 289, + 624 + ], + "type": "text", + "content": "Supervised Anomaly Detection. In theory, leveraging defective parts for training is advantageous for refining class boundaries [24]. However, practical challenges arise, such as the small size of anomalies and the impossibility of collecting all potential defects [31]. Several recent anomaly detection methods use both good and bad images [21, 28, 31], but some of them overfit to seen defects, and others require pixel masks for defects to calculate loss or to generate new defects, which can be problematic in a real-world case. Real-world datasets (as VAD) might contain logical defects that cannot be generated by existing defects generation strategies, which usually cut and paste existing defects with various modifications [28, 31], which is not going to improve results for a wrong wire shape as or other similar types of defects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Very few supervised anomaly detection benchmarks are available; the most popular is Supervised Anomaly Detection on MVTec AD. DevNet [21] is one of the first supervised anomaly detection models that tried to solve industrial MVTec AD dataset, compared to earlier methods which were mostly addressing Out-of-Distribution problem [11, 24] using non-industrial datasets such as MNIST [14]" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 247, + 547, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 547, + 357 + ], + "type": "text", + "content": "or CIFAR-10 [13]. DRA [8] uses several heads to learn both seen and pseudo anomalies, as well as normal examples, which should reduce overfitting on seen anomalies. These methods perform well on MVtec AD but might fail on more complex problems, as shown in Tab. 3 and Tab. 5. Compared to modern supervised anomaly detection methods, SegAD (ours) performs better than current SOTA methods, even on complex problems without pixel-level labels, extensive augmentation, and long training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 369, + 477, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 477, + 382 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 477, + 382 + ], + "type": "text", + "content": "3. Valeo Anomaly Dataset (VAD)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 389, + 547, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 547, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 547, + 593 + ], + "type": "text", + "content": "VAD consists of one class with predefined training and testing sets. The training set contains 1000 bad and 2000 good images, and the testing set contains 1000 bad, 165 of them are unseen bad, and 1000 good images. Unseen bad images in the test dataset refer to several rare defect types that are not present in the training data, several examples of which are shown in Fig. 4. Having such images is important to avoid turning this anomaly detection problem into a supervised classification problem. Some images might have a thin black border at the bottom due to how they were filmed, an example can be seen in Fig. 2. Defects might occur in the whole area of the image. Image-level annotation is provided, but there is no pixel-level annotation due to the complexity of defects and the fact there is no exact position for missing or misplaced components. All images are " + }, + { + "bbox": [ + 304, + 389, + 547, + 593 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 389, + 547, + 593 + ], + "type": "text", + "content": " pixels in PNG format. Examples of all types of defects can be found in Appendix C." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": "Good parts consist of two wires connected to two pins on one side and two soldering dots on the other, placed on a piezo. Piezo means piezoelectric element, a big round area under the other components. Wires should have the right shape, not too straight, too long, or bent too much. Wires should connect the soldering dots as close to their centers as possible. Some amount of deviation is allowed, but if the wire is too close to the side of a solder dot, it is a defect. Soldering should be well-rounded, not too small, and placed in the correct location, not overlapping with the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "17756" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 159, + 176 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 159, + 176 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 159, + 176 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 159, + 176 + ], + "type": "image", + "image_path": "d3105d75fc2c8b7cb52aa9267bc13e366d596b1b232487a26966f68dfde09f67.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 161, + 70, + 266, + 175 + ], + "blocks": [ + { + "bbox": [ + 161, + 70, + 266, + 175 + ], + "lines": [ + { + "bbox": [ + 161, + 70, + 266, + 175 + ], + "spans": [ + { + "bbox": [ + 161, + 70, + 266, + 175 + ], + "type": "image", + "image_path": "37a362d3d40e4339cdb5a9c8d810d436ffd2deb4ff3d73e53db6244eca149be3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 178, + 159, + 283 + ], + "blocks": [ + { + "bbox": [ + 53, + 178, + 159, + 283 + ], + "lines": [ + { + "bbox": [ + 53, + 178, + 159, + 283 + ], + "spans": [ + { + "bbox": [ + 53, + 178, + 159, + 283 + ], + "type": "image", + "image_path": "b551e15ba0358def0273f3b67e614fb5d83867b6726c043a4ab65fc498a63a11.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 307, + 287, + 330 + ], + "lines": [ + { + "bbox": [ + 47, + 307, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 287, + 330 + ], + "type": "text", + "content": "Figure 2. Good parts. Small scratches on the piezo are allowed, as well as wire being closer to the side of the soldering. (VAD)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 160, + 178, + 266, + 283 + ], + "blocks": [ + { + "bbox": [ + 160, + 178, + 266, + 283 + ], + "lines": [ + { + "bbox": [ + 160, + 178, + 266, + 283 + ], + "spans": [ + { + "bbox": [ + 160, + 178, + 266, + 283 + ], + "type": "image", + "image_path": "f451ca90a852a8c1812aab828cdfd6d7cd2e49edb845d9198f7bbbcd8b58dcf1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 350, + 165, + 402 + ], + "blocks": [ + { + "bbox": [ + 58, + 350, + 165, + 402 + ], + "lines": [ + { + "bbox": [ + 58, + 350, + 165, + 402 + ], + "spans": [ + { + "bbox": [ + 58, + 350, + 165, + 402 + ], + "type": "image", + "image_path": "595e8758846c627d9045bb394391f9ac01e448ede99dcb7f376de17dccee8e28.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 403, + 164, + 453 + ], + "blocks": [ + { + "bbox": [ + 58, + 403, + 164, + 453 + ], + "lines": [ + { + "bbox": [ + 58, + 403, + 164, + 453 + ], + "spans": [ + { + "bbox": [ + 58, + 403, + 164, + 453 + ], + "type": "image", + "image_path": "aaf296721c7a7c1d47b3de0bc562f81bc4297747d948a8cd2f52023ccb4d6da5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 464, + 133, + 474 + ], + "lines": [ + { + "bbox": [ + 78, + 464, + 133, + 474 + ], + "spans": [ + { + "bbox": [ + 78, + 464, + 133, + 474 + ], + "type": "text", + "content": "Logical defects" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 486, + 287, + 531 + ], + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 531 + ], + "type": "text", + "content": "Figure 3. Variety of defects: logical, from the top left: wire on the side, wire out of solder dot, missing wire, bad wire shape (bent too much). Structural: soldering paste pollution, crack on a piezo, burned area under the solder dot, broken piezo. (VAD)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 179, + 350, + 277, + 402 + ], + "blocks": [ + { + "bbox": [ + 179, + 350, + 277, + 402 + ], + "lines": [ + { + "bbox": [ + 179, + 350, + 277, + 402 + ], + "spans": [ + { + "bbox": [ + 179, + 350, + 277, + 402 + ], + "type": "image", + "image_path": "04fab07105a0f1a51830e5bd15f8fed0185abd6e05f19e935939e177f395a8ec.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 179, + 403, + 276, + 453 + ], + "blocks": [ + { + "bbox": [ + 179, + 403, + 276, + 453 + ], + "lines": [ + { + "bbox": [ + 179, + 403, + 276, + 453 + ], + "spans": [ + { + "bbox": [ + 179, + 403, + 276, + 453 + ], + "type": "image", + "image_path": "e27f961208d89d5fa335c0ded39c876204df4177a84e7abfe5ba808693a4298e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 464, + 260, + 473 + ], + "lines": [ + { + "bbox": [ + 195, + 464, + 260, + 473 + ], + "spans": [ + { + "bbox": [ + 195, + 464, + 260, + 473 + ], + "type": "text", + "content": "Structural defects" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 555, + 287, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 614 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 614 + ], + "type": "text", + "content": "piezo contours. The wire should be properly connected to the pin without any pieces of it visible on the side of the pin. The piezo can have different textures. Any kind of pollution is considered to be an anomaly, which makes the part defective." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Logical defects result from components being misplaced or misshapen rather than being damaged, e.g. bad wire shape, bad soldering shape, wrong wire position, wrong soldering position, and missing parts. Structural defects include various cracks and broken parts of the piezo, as well as a variety of burns and pollution. For both types of defects, see Fig. 3. The importance of logical defects is often overlooked in the existing datasets (excluding MVTec" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 75, + 393, + 144 + ], + "blocks": [ + { + "bbox": [ + 317, + 75, + 393, + 144 + ], + "lines": [ + { + "bbox": [ + 317, + 75, + 393, + 144 + ], + "spans": [ + { + "bbox": [ + 317, + 75, + 393, + 144 + ], + "type": "image", + "image_path": "4bdb7c339de780f876a09b1a68620c6a1c8aadedf62f2bd1e3ed6978ede888c2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 157, + 545, + 179 + ], + "lines": [ + { + "bbox": [ + 306, + 157, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 157, + 545, + 179 + ], + "type": "text", + "content": "Figure 4. Unseen defects for left to right: burned solder dot, burned wire, extra unnecessary wire. (VAD)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 395, + 76, + 457, + 144 + ], + "blocks": [ + { + "bbox": [ + 395, + 76, + 457, + 144 + ], + "lines": [ + { + "bbox": [ + 395, + 76, + 457, + 144 + ], + "spans": [ + { + "bbox": [ + 395, + 76, + 457, + 144 + ], + "type": "image", + "image_path": "53c3be7cf04fb199446ce0b90d8591a07cbc5fe7ebb1f183b67607b7ad3c31cd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 460, + 76, + 536, + 144 + ], + "blocks": [ + { + "bbox": [ + 460, + 76, + 536, + 144 + ], + "lines": [ + { + "bbox": [ + 460, + 76, + 536, + 144 + ], + "spans": [ + { + "bbox": [ + 460, + 76, + 536, + 144 + ], + "type": "image", + "image_path": "1b815771df7936ec659412c3c916e7270a94a4c06f5193737f11f26d75326185.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 199, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 199, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 199, + 545, + 247 + ], + "type": "text", + "content": "LOCO AD) because they are hard to imitate if the dataset is filmed in the lab, but also hard to detect. VAD provides the research community the opportunity to work with real logical constraints of a complex object." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 254, + 388, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 254, + 388, + 266 + ], + "spans": [ + { + "bbox": [ + 306, + 254, + 388, + 266 + ], + "type": "text", + "content": "3.1.Benchmarks" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 272, + 545, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 272, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 305, + 272, + 545, + 428 + ], + "type": "text", + "content": "We establish three benchmarks for the VAD dataset. The first is one-class anomaly detection. As shown in Tab 2, even SOTA methods struggle to perform well on our dataset compared to other popular datasets, which leaves an opening for further research. For this benchmark, no bad images are used for training. The second benchmark is high-shot supervised anomaly detection, which uses all 1000 bad images from the training set. The last benchmark is low-shot supervised anomaly detection. Only 100 bad images from the training set are used. These images are randomly selected using 5 seeds, with the average result over all seeds being expected. The same 2000 good parts are used for the training in all benchmarks." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 437, + 361, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 361, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 361, + 449 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 457, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 457, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 305, + 457, + 545, + 614 + ], + "type": "text", + "content": "SegAD works with image data at its input, and after inference, it is expected to output a score that describes a probability of anomaly appearance. Let " + }, + { + "bbox": [ + 305, + 457, + 545, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\left[0,\\dots ,255\\right]^{W,H,3}" + }, + { + "bbox": [ + 305, + 457, + 545, + 614 + ], + "type": "text", + "content": " be a space of all images of width " + }, + { + "bbox": [ + 305, + 457, + 545, + 614 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 305, + 457, + 545, + 614 + ], + "type": "text", + "content": ", height " + }, + { + "bbox": [ + 305, + 457, + 545, + 614 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 305, + 457, + 545, + 614 + ], + "type": "text", + "content": " and with 3 color channels. The " + }, + { + "bbox": [ + 305, + 457, + 545, + 614 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 457, + 545, + 614 + ], + "type": "text", + "content": "-th image from a dataset is marked as " + }, + { + "bbox": [ + 305, + 457, + 545, + 614 + ], + "type": "inline_equation", + "content": "I^{(i)}\\in \\mathcal{I}" + }, + { + "bbox": [ + 305, + 457, + 545, + 614 + ], + "type": "text", + "content": ". We denote " + }, + { + "bbox": [ + 305, + 457, + 545, + 614 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 305, + 457, + 545, + 614 + ], + "type": "text", + "content": " to be a 2D pixel coordinate vector. SegAD inference consists of three consecutive stages in a pipeline. First is anomaly map(s) calculation with optional supervised classifier score calculation. Next is the calculation of simple statistics from anomaly maps over individual segments. Lastly, the obtained statistics, and optionally also the classifier score, are used as input features for a BRF classifier [18] that delivers the final result." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 615, + 545, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 615, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 305, + 615, + 545, + 690 + ], + "type": "text", + "content": "Let us denote " + }, + { + "bbox": [ + 305, + 615, + 545, + 690 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 615, + 545, + 690 + ], + "type": "text", + "content": " as the number of used pixel-wise anomaly detectors and mark them as functions " + }, + { + "bbox": [ + 305, + 615, + 545, + 690 + ], + "type": "inline_equation", + "content": "f_{k}(I):\\mathcal{I}\\to \\mathbb{R}^{W,H},k\\in \\{1,\\ldots ,K\\}" + }, + { + "bbox": [ + 305, + 615, + 545, + 690 + ], + "type": "text", + "content": ". A supervised classifier is a function " + }, + { + "bbox": [ + 305, + 615, + 545, + 690 + ], + "type": "inline_equation", + "content": "g(I):\\mathbb{R}^{W,H}\\to \\mathbb{R}" + }, + { + "bbox": [ + 305, + 615, + 545, + 690 + ], + "type": "text", + "content": ". Next, we introduce a set of " + }, + { + "bbox": [ + 305, + 615, + 545, + 690 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 615, + 545, + 690 + ], + "type": "text", + "content": " mutually exclusive segments " + }, + { + "bbox": [ + 305, + 615, + 545, + 690 + ], + "type": "inline_equation", + "content": "S = \\{s_1,\\dots,s_L\\}" + }, + { + "bbox": [ + 305, + 615, + 545, + 690 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 615, + 545, + 690 + ], + "type": "inline_equation", + "content": "s_l\\in [0,1]^{W,H}" + }, + { + "bbox": [ + 305, + 615, + 545, + 690 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 690, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 690, + 545, + 714 + ], + "type": "text", + "content": "After the calculation of anomaly maps for all " + }, + { + "bbox": [ + 306, + 690, + 545, + 714 + ], + "type": "inline_equation", + "content": "f_{k}(I)" + }, + { + "bbox": [ + 306, + 690, + 545, + 714 + ], + "type": "text", + "content": ", the masks " + }, + { + "bbox": [ + 306, + 690, + 545, + 714 + ], + "type": "inline_equation", + "content": "s_l" + }, + { + "bbox": [ + 306, + 690, + 545, + 714 + ], + "type": "text", + "content": " are used for segment-wise" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17757" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "extraction of results, followed by the calculation of several statistics. The final feature vector " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{f} = [g(I),[\\mathbf{q}^L ]_{k = 1}^K,[\\mathbf{z}^L ]_{k = 1}^K,[\\mathbf{c}^L ]_{k = 1}^K,[\\mathbf{m}^L ]_{k = 1}^K]" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " extracted from one image, is the concatenation of 4 vectors containing different statistics. " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_k^L" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " stands for a vector of " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " numbers where each stands for a " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "99.5\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " quantile from anomaly map " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "f_{k}(I)" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " over pixels where " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "s_l(\\mathbf{x}) = 1" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " Similarly, for every anomaly map " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "f_{k}(I)" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " we calculate a vector for skew " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_k^L" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " , kurtosis " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_k^L" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " and mean " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_k^L" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " . The length of the feature vector " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " equals to " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "K*L*4 + 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 192, + 288, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 192, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 288, + 251 + ], + "type": "text", + "content": "For notation simplicity, we keep the supervised classifier score " + }, + { + "bbox": [ + 47, + 192, + 288, + 251 + ], + "type": "inline_equation", + "content": "g(I)" + }, + { + "bbox": [ + 47, + 192, + 288, + 251 + ], + "type": "text", + "content": " always present in " + }, + { + "bbox": [ + 47, + 192, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 47, + 192, + 288, + 251 + ], + "type": "text", + "content": ", but it may be omitted, which is clearly defined in every experiment setup. Feature vectors " + }, + { + "bbox": [ + 47, + 192, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{(i)}" + }, + { + "bbox": [ + 47, + 192, + 288, + 251 + ], + "type": "text", + "content": " are used as inputs for training the final classifier [18], which, at inference time, delivers the final result." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 260, + 111, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 260, + 111, + 273 + ], + "spans": [ + { + "bbox": [ + 47, + 260, + 111, + 273 + ], + "type": "text", + "content": "4.1. Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 279, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 361 + ], + "type": "text", + "content": "Training SegAD requires two separate subsets from the training dataset. The first subset is used to train base models " + }, + { + "bbox": [ + 46, + 279, + 287, + 361 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 46, + 279, + 287, + 361 + ], + "type": "text", + "content": ", the second subset is used to train the BRF classifier. To create these subsets from one training dataset, we split available good images into two parts. For VAD, we report average results over 5 different splits to show more reliable results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 363, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 542 + ], + "type": "text", + "content": "Each model " + }, + { + "bbox": [ + 46, + 363, + 287, + 542 + ], + "type": "inline_equation", + "content": "(f_{k}, g, \\mathrm{BRF})" + }, + { + "bbox": [ + 46, + 363, + 287, + 542 + ], + "type": "text", + "content": " is trained independently. Training can be separated into two stages. In the first stage, anomaly detector(s) " + }, + { + "bbox": [ + 46, + 363, + 287, + 542 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 46, + 363, + 287, + 542 + ], + "type": "text", + "content": " are trained using the same data. In the same stage, a segmentation map " + }, + { + "bbox": [ + 46, + 363, + 287, + 542 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 363, + 287, + 542 + ], + "type": "text", + "content": " must be defined. As will be elaborated in more detail in Sec. 5, the segmentation map can be static or produced by a segmentation model, depending on the dataset. A supervised classifier " + }, + { + "bbox": [ + 46, + 363, + 287, + 542 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 363, + 287, + 542 + ], + "type": "text", + "content": " is trained separately on " + }, + { + "bbox": [ + 46, + 363, + 287, + 542 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 363, + 287, + 542 + ], + "type": "text", + "content": " of the full training set (1600 good and 800 bad parts for VAD) to ensure that the training sets for this model and SegAD are not fully overlapping. The next stage is to train a final BRF classifier in SegAD, which requires both good and bad images, bad images can be replaced with artificial defects in some cases, as shown in Tab. 2. BRF is trained with the output of the models from the previous stage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 543, + 287, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 627 + ], + "type": "text", + "content": "An important thing to note is that SegAD is very fast to train as long as you already have other models available. Extraction of features and training the final classifier takes noticeably less time than training anomaly detectors. The same is true for the inference; SegAD adds a negligible time to processing the image while improving results significantly." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 635, + 152, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 152, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 152, + 647 + ], + "type": "text", + "content": "4.2. Setups for SegAD" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "SegAD can use any anomaly detector(s) " + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": " as long as it produces an anomaly map showing pixel-level anomaly scores, although, in this paper, we use only one-class anomaly detectors because they require less data to train. Any supervised classifier " + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": " that returns a classification score can be" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": "used. In experiments, we use several setups, which will be described here in detail. SegAD is denoted as Ours. Supervised classifier " + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": " is not used unless stated otherwise." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 108, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 167 + ], + "type": "text", + "content": "Anomaly Detector + Ours means " + }, + { + "bbox": [ + 304, + 108, + 545, + 167 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 304, + 108, + 545, + 167 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 108, + 545, + 167 + ], + "type": "inline_equation", + "content": "f_{1} =" + }, + { + "bbox": [ + 304, + 108, + 545, + 167 + ], + "type": "text", + "content": " \"Anomaly Detector name\". This setup allows us to show improvement for one anomaly detection model in particular and demonstrate that this improvement can be achieved for different types of anomaly detectors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 168, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 545, + 263 + ], + "type": "text", + "content": "All " + }, + { + "bbox": [ + 304, + 168, + 545, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{AD} + \\mathbf{Ours}" + }, + { + "bbox": [ + 304, + 168, + 545, + 263 + ], + "type": "text", + "content": " denotes that all anomaly detectors from the list " + }, + { + "bbox": [ + 304, + 168, + 545, + 263 + ], + "type": "inline_equation", + "content": "AllAD" + }, + { + "bbox": [ + 304, + 168, + 545, + 263 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 304, + 168, + 545, + 263 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 168, + 545, + 263 + ], + "type": "text", + "content": " are used to produce anomaly maps. Such a list can be found in the caption of the table. In that case " + }, + { + "bbox": [ + 304, + 168, + 545, + 263 + ], + "type": "inline_equation", + "content": "K = N" + }, + { + "bbox": [ + 304, + 168, + 545, + 263 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 168, + 545, + 263 + ], + "type": "inline_equation", + "content": "f_{1} = AllAD_{1},\\ldots ,f_{N} = AllAD_{N}" + }, + { + "bbox": [ + 304, + 168, + 545, + 263 + ], + "type": "text", + "content": ". Anomaly detectors might use an ensemble of feature extractors, similarly to [23], which makes them slower but can improve the result. In our case, anomaly detectors themselves can be successfully assembled in a supervised way." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 264, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 382 + ], + "type": "text", + "content": "All AD + Supervised Classifier + Ours denotes a setup in which " + }, + { + "bbox": [ + 304, + 264, + 545, + 382 + ], + "type": "inline_equation", + "content": "g =" + }, + { + "bbox": [ + 304, + 264, + 545, + 382 + ], + "type": "text", + "content": " \"Supervised Classifier name\", and the rest is the same as in the previous setup. A supervised classifier allows us to detect seen defects, which can be invisible to a one-class classifier, similarly to [25]. Such a setup shows the best result for academic purposes, but the speed of inference can be unsatisfying for real-world applications. We expect that, in that case, the number and selection of anomaly detectors can be optimized to fit the problem as well as the time constraints." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 384, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 467 + ], + "type": "text", + "content": "Anomaly Detector + Supervised Classifier + Ours mean a setup in which " + }, + { + "bbox": [ + 304, + 384, + 545, + 467 + ], + "type": "inline_equation", + "content": "g =" + }, + { + "bbox": [ + 304, + 384, + 545, + 467 + ], + "type": "text", + "content": " \"Supervised Classifier name\", " + }, + { + "bbox": [ + 304, + 384, + 545, + 467 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 304, + 384, + 545, + 467 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 384, + 545, + 467 + ], + "type": "inline_equation", + "content": "f_{1} =" + }, + { + "bbox": [ + 304, + 384, + 545, + 467 + ], + "type": "text", + "content": " \"Anomaly Detector name\". It shows an example of an optimization mentioned in the previous setup. Only one anomaly detector gives a higher inference speed compared to the ensemble of anomaly detectors while giving a satisfying result." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 478, + 387, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 387, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 387, + 491 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 498, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 616 + ], + "type": "text", + "content": "Evaluation metrics. We compare results using the following metrics: (1) Area Under the Receiver Operating Characteristic Curve (AUROC) [5], (2) False Positive Rate (FPR) at " + }, + { + "bbox": [ + 304, + 498, + 545, + 616 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 616 + ], + "type": "text", + "content": " True Positive Rate (TPR) denoted as FPR@95TPR. Classification AUROC, which describes the performance on the image level for multiple thresholds, denoted by Cl. AUROC. FPR@95TPR shows the percentage of misclassified good parts (false positive) at " + }, + { + "bbox": [ + 304, + 498, + 545, + 616 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 616 + ], + "type": "text", + "content": " of bad parts classified correctly (true positive). We report mean results " + }, + { + "bbox": [ + 304, + 498, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\pm" + }, + { + "bbox": [ + 304, + 498, + 545, + 616 + ], + "type": "text", + "content": " standard deviation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Implementation and evaluation details. SegAD uses the output of one-class detectors PatchCore [23], FastFlow [29], Reverse Distillation [7] (referred to as RD4AD), EfficientAD [1] either alone or several anomaly detectors at once. SegAD results are compared to the results of these models alone, as well as supervised anomaly detectors DevNet [21] and DRA [8]. We also compare our results to the supervised classifier Wide ResNet50 [30] (denoted as" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17758" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": "WRN). The final classifier BRF was trained with the same parameters for all datasets using XGBoost library [6]. More details on the implementation can be found in Appendixes A and B." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "text", + "content": "Images are resized to " + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "inline_equation", + "content": "\\mathrm{W} = \\mathrm{H} = 256" + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "text", + "content": " pixels to make the comparison less biased. The DRA model originally uses a resolution " + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "text", + "content": ", which shows better results on VAD than " + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "text", + "content": ", 92.8 CI. AUROC, which is still lower than the top results. The higher resolution also improves results for some of the other models, so we do not use it. " + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 121, + 286, + 227 + ], + "type": "text", + "content": " is sufficient for defects to remain visible, but it also allows to have a reasonable speed of inference, especially for anomaly detection models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 228, + 286, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 286, + 396 + ], + "type": "text", + "content": "For VAD, results averaged for 5 seeds and training runs are reported. These seeds are used for training and splitting the training set for SegAD into two sets, as described in Subsection 4.1. 2000 good parts from the train set were split in half: 1000 for training the base model(s) and 1000 for training SegAD. 0, 1000, or 100 bad parts (depending on the benchmark) were also used to train SegAD. Details on the VisA benchmark can be found in Subsection 5.4. No augmentations were used for training, and existing augmentations were removed from used methods. Augmentations tend to be suitable only for specific tasks (as rotation in DRA or DevNet, removing it improved results by these models on the VAD drastically), and using augmentations makes it hard to compare models themselves." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 396, + 286, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 286, + 456 + ], + "type": "text", + "content": "We include three benchmarks for the new VAD and one additional for the VisA dataset [32]. We compare results for VAD for one-class, high-shot, and low-shot supervised benchmarks, as described in Subsection 3.1. The benchmark for the VisA dataset is explained in 5.4. The segmen" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 96, + 467, + 239, + 611 + ], + "blocks": [ + { + "bbox": [ + 96, + 467, + 239, + 611 + ], + "lines": [ + { + "bbox": [ + 96, + 467, + 239, + 611 + ], + "spans": [ + { + "bbox": [ + 96, + 467, + 239, + 611 + ], + "type": "image", + "image_path": "abc22acaf126607a77657c5ce1b4aea02cd5a88f9c5bedd859b46e19176465f4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 619, + 286, + 664 + ], + "lines": [ + { + "bbox": [ + 46, + 619, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 619, + 286, + 664 + ], + "type": "text", + "content": "Figure 5. Static segmentation map for objects in VAD dataset, overlayed over the image. The image is separated into " + }, + { + "bbox": [ + 46, + 619, + 286, + 664 + ], + "type": "inline_equation", + "content": "L = 7" + }, + { + "bbox": [ + 46, + 619, + 286, + 664 + ], + "type": "text", + "content": " segments: background, outer half-circle, piezo border, solder dots area, wires area, pins area, and piezo in the middle." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "tation maps for VAD is the same static image for every input image from the dataset. The segmentation map can be seen in Fig. 5. Due to the fact that the object in the VAD is al" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "ways centered with the constant size and component placement, this is the simplest choice, which may be sub-optimal, but serves the purpose and does not require any manual or model-inferred annotation of the data. It was created to separate the anomaly maps into several meaningful parts, such as piezo border, wire areas, soldering area, etc., which can contain different defects and different levels of falsely high scores on the anomaly map. Another source of segmentation maps can be a specially trained segmentation model or a zero-shot segmentation model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 204, + 454, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 204, + 454, + 216 + ], + "spans": [ + { + "bbox": [ + 305, + 204, + 454, + 216 + ], + "type": "text", + "content": "5.1. VAD, one-class benchmark" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 224, + 545, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 545, + 368 + ], + "type": "text", + "content": "This benchmark allows only good images for training, but SegAD requires bad images as well, so we try to replace them with artificially generated defects. Because SegAD works with anomaly maps, generating realistic defects is unnecessary. For this reason, we have applied several simple augmentations to the 1000 good images available for training SegAD. These augmentations include Gaussian blur or a randomly placed, randomly sized thin rectangle of a random shade of gray. They were applied to randomly selected segments or random parts of the segments; segments are defined with the segmentation map. Code for generating defects can be found in the SegAD GitHub repository." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 306, + 380, + 545, + 547 + ], + "blocks": [ + { + "bbox": [ + 306, + 380, + 545, + 547 + ], + "lines": [ + { + "bbox": [ + 306, + 380, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 545, + 547 + ], + "type": "table", + "html": "
MethodCl. AUROC ↑FPR@95TPR ↓
One-class Anomaly Detection
FastFlow79.1±1.776.2±1.9
RD4AD84.5±0.265.8±2
PatchCore88.3±0.360.4±2.8
EfficientAD89.1±0.844.5±2.2
SegAD (Ours)
FastFlow + Ours83.3±1.371.5±5
RD4AD + Ours87.5±1.261.0±7.9
PatchCore + Ours89.7±0.858.8±3.7
EfficientAD + Ours85.6±1.6100±0
All AD + Ours90.4±0.552.4±4.5
", + "image_path": "728cdf57302ac6c3914e9c492b2074778ec529907b7b9a1e1274d480d6215740.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 555, + 545, + 576 + ], + "lines": [ + { + "bbox": [ + 306, + 555, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 555, + 545, + 576 + ], + "type": "text", + "content": "Table 2. One-class benchmark (VAD). The best result is marked in bold. All AD means PatchCore, FastFlow, and RD4AD." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "text", + "content": "Results in Table 2 show that even such a naive approach improves results compared to anomaly detection methods alone. Even greater improvement can be achieved by using several anomaly detectors to produce anomaly maps for SegAD. This strategy fails for EfficientAD + Ours, anomaly maps produced by EfficientAD for the generated defects are too different from anomaly maps for real defects, creating a large discrepancy between training and test distributions for SegAD and causing worse results compared to anomaly detector alone." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17759" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 251, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 251, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 251, + 85 + ], + "type": "text", + "content": "5.2. VAD, high-shot supervised benchmark" + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 98, + 287, + 273 + ], + "blocks": [ + { + "bbox": [ + 48, + 98, + 287, + 273 + ], + "lines": [ + { + "bbox": [ + 48, + 98, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 98, + 287, + 273 + ], + "type": "table", + "html": "
MethodCl. AUROC ↑FPR@95TPR ↓
Supervised Anomaly Detection
DevNet86.9±1.266.6±2.7
DRA87.4±0.660.9±4.4
Supervised Classifier
WRN95.0±0.632.8±4.3
SegAD (Ours)
FastFlow + Ours86.9±0.556.9±3.9
RD4AD + Ours90.2±0.346.9±2.6
PatchCore + Ours91.7±0.348.0±2.9
EfficientAD + Ours91.4±0.236.8±1
AllAD+WRN+Ours96.5±0.316.4±1.9
", + "image_path": "c27cf7d857f1f5bfb13ffa528c27e01b04abbd1aeffe27984edcb56e323b8bbd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 282, + 287, + 327 + ], + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 327 + ], + "type": "text", + "content": "Table 3. High-shot supervised benchmark (VAD). The best result is marked in bold. All AD means PatchCore, FastFlow, RD4AD, and EfficientAD. Improvement calculated compared to base method results in Tab. 2." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 338, + 289, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 289, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 289, + 531 + ], + "type": "text", + "content": "The high-shot benchmark allows training supervised classifier " + }, + { + "bbox": [ + 46, + 338, + 289, + 531 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 338, + 289, + 531 + ], + "type": "text", + "content": ", which shows competitive results as can be seen in Tab. 3 in the row \"WRN\", but it won't be able to detect unseen defects. Histograms in Fig. 6 visualize this problem. On the left is the one-class anomaly detector, which can detect unseen defects (in red) but shows a bad separation between good and bad parts. In the center, the supervised classifier " + }, + { + "bbox": [ + 46, + 338, + 289, + 531 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 338, + 289, + 531 + ], + "type": "text", + "content": " shows a better separation of good and bad parts, but unseen defects distribution is shifted to the left because their scores are relatively low and such defects cannot be detected. On the right, SegAD shows a good separation between classes, but also unseen defects have a similar distribution to the seen defects. We still compare the supervised classifier Wide ResNet50 (WRN) with other methods, but the ability to detect unseen defects is crucial in real-world applications." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 542, + 124, + 601 + ], + "blocks": [ + { + "bbox": [ + 56, + 542, + 124, + 601 + ], + "lines": [ + { + "bbox": [ + 56, + 542, + 124, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 542, + 124, + 601 + ], + "type": "image", + "image_path": "b6e41dc57d33938baf62ad26303d3a531562e8669653edc7e43b6250cb64a215.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 605, + 93, + 613 + ], + "lines": [ + { + "bbox": [ + 85, + 605, + 93, + 613 + ], + "spans": [ + { + "bbox": [ + 85, + 605, + 93, + 613 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 542, + 202, + 601 + ], + "blocks": [ + { + "bbox": [ + 133, + 542, + 202, + 601 + ], + "lines": [ + { + "bbox": [ + 133, + 542, + 202, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 542, + 202, + 601 + ], + "type": "image", + "image_path": "87b12149746b30f361d2d85e89cd9b02146e549bd0f99c86b0923721ffda84d7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 605, + 170, + 613 + ], + "lines": [ + { + "bbox": [ + 161, + 605, + 170, + 613 + ], + "spans": [ + { + "bbox": [ + 161, + 605, + 170, + 613 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 622, + 287, + 679 + ], + "lines": [ + { + "bbox": [ + 46, + 622, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 46, + 622, + 287, + 679 + ], + "type": "text", + "content": "Figure 6. Distributions of scores in the VAD test set. Green color shows good parts, orange bad parts, and red bad parts with unseen defects. (a) shows one-class anomaly detector PatchCore, (b) supervised classifier Wide ResNet, (c) SegAD (PatchCore + WRN + Ours). X axis is score values, Y axis is frequency." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 211, + 542, + 279, + 601 + ], + "blocks": [ + { + "bbox": [ + 211, + 542, + 279, + 601 + ], + "lines": [ + { + "bbox": [ + 211, + 542, + 279, + 601 + ], + "spans": [ + { + "bbox": [ + 211, + 542, + 279, + 601 + ], + "type": "image", + "image_path": "6f3e7fed76afdc8428586670d383e9aabc288c58284f41305df1341e99ce7d08.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 605, + 251, + 613 + ], + "lines": [ + { + "bbox": [ + 244, + 605, + 251, + 613 + ], + "spans": [ + { + "bbox": [ + 244, + 605, + 251, + 613 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Combining SegAD, supervised classifier WRN, and all one-class anomaly detection models showed the best results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "content": "presented in Table 3. A more practical setting can be EfficientAD + WRN + Ours, which can achieve Cl. AUROC 96 and FPR@95TPR 19.4 with a much higher inference speed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 116, + 505, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 116, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 305, + 116, + 505, + 129 + ], + "type": "text", + "content": "5.3. VAD, low-shot supervised benchmark" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 134, + 547, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 134, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 134, + 547, + 266 + ], + "type": "text", + "content": "This benchmark uses just 100 bad images for training, which can be a relatively low number for a dataset with more than 20 types of defects. However, such a setting is closer to real-world applications, where a limited amount of defective images can be available. In Table 4, WRN shows much lower results due to a smaller training dataset, as well as DevNet and DRA. SegAD also performs worse, than with a high-shot benchmark, but it still can visibly improve the performance of other models. We do not use WRN in SegAD for this benchmark, because it showed low results by itself." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 306, + 276, + 545, + 452 + ], + "blocks": [ + { + "bbox": [ + 306, + 276, + 545, + 452 + ], + "lines": [ + { + "bbox": [ + 306, + 276, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 276, + 545, + 452 + ], + "type": "table", + "html": "
MethodCl. AUROC ↑FPR@95TPR ↓
Supervised Anomaly Detection
DevNet73.1±2.388.3±2.3
DRA78.9±1.874.6±1.9
Supervised Classifier
WRN78.7±1.877.9±4.4
SegAD (ours)
FastFlow + Ours84.3±1.464.5±2.5
RD4AD + Ours88.6±0.251.1±0.9
PatchCore + Ours90.8±0.552.7±4.3
EfficientAD + Ours90.6±0.740.1±2.4
All AD + Ours92.7±0.536.8±3.1
", + "image_path": "925ff91b5a4532f4588b82edab18a74c6fb9b056ba564acbde825f843f5f4c4f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 304, + 460, + 546, + 494 + ], + "lines": [ + { + "bbox": [ + 304, + 460, + 546, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 546, + 494 + ], + "type": "text", + "content": "Table 4. Low-shot supervised benchmark (VAD). The best result is marked in bold. All AD means PatchCore, FastFlow, RD4AD, and EfficientAD." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 515, + 440, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 440, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 440, + 528 + ], + "type": "text", + "content": "5.4. Results on VisA dataset" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "VisA contains fewer good images per class for training than VAD (500-1000 in VisA, compared to 2000 in VAD). The training set consists of good images only. For this reason, we use a different strategy to split the training dataset to make sure we have enough images to train the base anomaly detection model with a competitive result. For RD4AD, the training set was divided into " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " of images to train the RD4AD itself and " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " to train SegAD. EfficientAD uses a validation set (" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " of good images from the train set) to normalize anomaly maps. In our case, the same images were used to train SegAD. The segmentation maps for this dataset were created with the Segment Anything Model (SAM) [12]. For most classes, it is a binary mask for the object and background, and for pcb1-4, static components were mapped on top of it; see Fig. 7." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17760" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 74, + 126, + 144 + ], + "blocks": [ + { + "bbox": [ + 55, + 74, + 126, + 144 + ], + "lines": [ + { + "bbox": [ + 55, + 74, + 126, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 74, + 126, + 144 + ], + "type": "image", + "image_path": "faed8c248e9be65f9aa2750548993ae5d11f7d6f05544f5a94410540b5917c72.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 148, + 94, + 157 + ], + "lines": [ + { + "bbox": [ + 85, + 148, + 94, + 157 + ], + "spans": [ + { + "bbox": [ + 85, + 148, + 94, + 157 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 132, + 74, + 202, + 145 + ], + "blocks": [ + { + "bbox": [ + 132, + 74, + 202, + 145 + ], + "lines": [ + { + "bbox": [ + 132, + 74, + 202, + 145 + ], + "spans": [ + { + "bbox": [ + 132, + 74, + 202, + 145 + ], + "type": "image", + "image_path": "ce9cb5d1b85db512db2232c5280e543f58cecde19401f5aa932d5dc510f8ca9f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 148, + 171, + 158 + ], + "lines": [ + { + "bbox": [ + 162, + 148, + 171, + 158 + ], + "spans": [ + { + "bbox": [ + 162, + 148, + 171, + 158 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 208, + 74, + 279, + 144 + ], + "blocks": [ + { + "bbox": [ + 208, + 74, + 279, + 144 + ], + "lines": [ + { + "bbox": [ + 208, + 74, + 279, + 144 + ], + "spans": [ + { + "bbox": [ + 208, + 74, + 279, + 144 + ], + "type": "image", + "image_path": "a0ea3a90541364f5f291409e93ce9176edf77ffd775758bdf9f5658ed918d6c9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 148, + 248, + 157 + ], + "lines": [ + { + "bbox": [ + 239, + 148, + 248, + 157 + ], + "spans": [ + { + "bbox": [ + 239, + 148, + 248, + 157 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 169, + 287, + 192 + ], + "lines": [ + { + "bbox": [ + 47, + 169, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 169, + 287, + 192 + ], + "type": "text", + "content": "Figure 7. VisA. Examples of segmentation maps created using SAM for (a) cashew, (b) macaroni1, (c) pb2." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 47, + 204, + 287, + 367 + ], + "blocks": [ + { + "bbox": [ + 47, + 204, + 287, + 367 + ], + "lines": [ + { + "bbox": [ + 47, + 204, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 204, + 287, + 367 + ], + "type": "table", + "html": "
MethodCl. AUROC↑FPR@95TPR↓
One-class Anomaly Detection
RD4AD94.225.5
EfficientAD98.08.1
Supervised Anomaly Detection
DRA88.942.4
DevNet89.344.0
SegAD (ours)
RD4AD + Ours96.115.5
EfficientAD + Ours98.36.2
All AD + Ours98.47.5
", + "image_path": "38102a446d1d623fce49734ea46f6e219a65b13d25dabe8bdb8ab7d0b1514201.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 433, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 552 + ], + "type": "text", + "content": "With VisA, we create a new, challenging, supervised benchmark on a public dataset with 12 different classes. In a similar way to the supervised benchmark on MVTec AD [20], we move 10 randomly selected bad images from the test set to the train set, reporting results over 10 different seeds. In Table 5, this benchmark shows how SegAD performs on a different data from VAD and creates a new possibility to compete on a difficult supervised anomaly detection problem. Detailed results per class can be found in Appendix D." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 562, + 138, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 138, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 138, + 575 + ], + "type": "text", + "content": "5.5. Ablation study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "We perform an ablation study on the VAD high-shot supervised benchmark. We use the average result of the Patch-Core, FastFlow, RD4AD, and EfficientAD as the baseline in the Table 6, denoted as \"An.Det.\" These anomaly detectors use the maximum value from the anomaly map to calculate the score. On the contrary, we compute our proposed features " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": ", which are described in Section 4, and use BRF to yield a final score. This shows an improvement of 1.3 CI. AUROC in the row \"One Seg.\" In addition, we evaluate the segmentation map's impact by computing the anomaly map's maximum value per segment. Doing this gives an im" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 157 + ], + "type": "text", + "content": "provement in 2.1 Cl. AUROC, which can be seen in the row \"Max.\" This shows how both features and segmentation maps are important for SegAD. We also show the strength of BRF by comparing it against other algorithms. We compare BRF with BT and RF using a segmentation map and features. The results show that the BRF is better than these methods by 1.7 and 0.3 Cl. AUROC, respectively." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 307, + 164, + 545, + 274 + ], + "blocks": [ + { + "bbox": [ + 47, + 376, + 287, + 398 + ], + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 398 + ], + "type": "text", + "content": "Table 5. Results on VisA dataset, supervised benchmark. The best result is marked in bold. All AD means RD4AD and EfficientAD." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 164, + 545, + 274 + ], + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 274 + ], + "type": "table", + "html": "
DescriptionFeat. fSeg. SBRFCl. AUROC ↑
An. Det.87.3
One Seg.✓✓88.6
Max.✓✓89.4
BT✓✓88.4
RF✓✓89.8
SegAD✓✓✓90.1
", + "image_path": "adffd947c55868173cf1c01ba9758807a466da7397157812495609065d0bea5d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 281, + 547, + 358 + ], + "lines": [ + { + "bbox": [ + 304, + 281, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 547, + 358 + ], + "type": "text", + "content": "Table 6. Ablation study on the methodical differences between our method (SegAD) and anomaly detection methods. An.Det. = Anomaly Detectors. One Seg. = One Segment. Max. means using maximum value. BT = Boosted Tree. RF = Random Forest. BRF = Boosted Random Forest. SegAD denotes the average result from PatchCore + Ours, FastFlow + Ours, RD4AD + Ours, EfficientAD + Ours." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 384, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 539 + ], + "type": "text", + "content": "Limitations: SegAD (ours) requires to have segmentation maps. Static maps can be used for objects like the product in VAD. Obtaining segmentation maps for unaligned objects can be more difficult, yet SAM [12] or a specially trained segmentation model may be a solution. The performance improvement of SegAD depends on the complexity and structure of the objects. For less complex objects in VisA, improvement is lower than for VAD, as can be seen in Tab. 3 and Tab. 5. Another possible limitation can be the difference in the anomaly maps between training and test data. Due to the nature of anomaly detection, this problem might be impossible to mitigate, because our task is to detect such differences." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "content": "This study introduces VAD, a brand new supervised anomaly detection dataset derived from real production, offering challenging benchmarks for the research community to address real-world defect detection. Additionally, we propose SegAD, an innovative supervised anomaly detection method that achieves state-of-the-art performance on VAD as well as on a supervised benchmark on standard VisA dataset. Experimental results reveal the limitations of existing supervised anomaly detection methods in handling complex problems, while the integration of SegAD with top-performing one-class anomaly detection models further enhances their results." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17761" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Kilian Batzner, Lars Heckler, and Rebecca König. Efficientad: Accurate visual anomaly detection at millisecond-level latencies, 2023. 1, 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "text", + "content": "[2] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad — a comprehensive real-world dataset for unsupervised anomaly detection. In CVPR, pages 9584–9592, 2019. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 212 + ], + "type": "text", + "content": "[3] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. IJCV, 130(4):947-969, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 246 + ], + "type": "text", + "content": "[4] Jakob Božić, Domen Tabernik, and Danijel Skočaj. Mixed supervision for surface-defect detection: from weakly to fully supervised learning. Computers in Industry, 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 288, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 278 + ], + "type": "text", + "content": "[5] Andrew P. Bradley. The use of the area under the roc curve in the evaluation of machine learning algorithms. PR, 30(7): 1145-1159, 1997. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 280, + 288, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 302 + ], + "type": "text", + "content": "[6] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In SIGKDD, pages 785-794, 2016. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 303, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 288, + 334 + ], + "type": "text", + "content": "[7] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In CVPR, pages 9737-9746, 2022. 1, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "text", + "content": "[8] Chouro Ding, Guansong Pang, and Chunhua Shen. Catching both gray and black swans: Open-set supervised anomaly detection. In CVPR, 2022. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 369, + 288, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 411 + ], + "type": "text", + "content": "[9] Colin Drury and Murray Sinclair. Human and machine performance in an inspection task. Human Factors: The Journal of the Human Factors and Ergonomics Society, 25:391-399, 1983. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 413, + 287, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 413, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 48, + 413, + 287, + 457 + ], + "type": "text", + "content": "[10] Lars Heckler, Rebecca König, and Paul Bergmann. Exploring the importance of pretrained feature extractors for unsupervised anomaly detection and localization. In CVPRW, pages 2917-2926, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 458, + 287, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 287, + 489 + ], + "type": "text", + "content": "[11] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In ICLR, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "type": "text", + "content": "[12] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, pages 4015-4026, 2023. 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 547, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 568 + ], + "type": "text", + "content": "[13] Alex Krizhevsky et al. Learning multiple layers of features from tiny images. 2009. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 569, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 590 + ], + "type": "text", + "content": "[14] Yann LeCun, Corinna Cortes, Chris Burges, et al. Mnist handwritten digit database, 2010. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "text", + "content": "[15] Jiaqi Liu, Guoyang Xie, Jingbao Wang, Shangnian Li, Chengjie Wang, Feng Zheng, and Yaochu Jin. Deep industrial image anomaly detection: A survey. arXiv e-prints, pages arXiv-2301, 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "text", + "content": "[16] Muriel Mazzetto, Marcelo Teixeira, Érick Oliveira Rodrigues, and Dalcimar Casanova. Deep learning models for visual inspection on automotive assembling line. arXiv preprint arXiv:2007.01857, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[17] Declan McIntosh and Alexandra Branzan Albu. Interrealization channels: Unsupervised anomaly detection in images beyond one-class classification, 2023. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 700 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[18] Yohei Mishina, Masamitsu Tsuchiya, and Hironobu Fujiyoshi. Boosted random forest. In VISAPP, pages 594-598, 2014. 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[19] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. VT-ADL: A vision transformer network for image anomaly detection and localization. In ISIE, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "text", + "content": "[20] Guansong Pang, Chunhua Shen, and Anton van den Hengel. Deep anomaly detection with deviation networks. In SIGKDD, pages 353-362, 2019. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "text", + "content": "[21] Guansong Pang, Choubo Ding, Chunhua Shen, and Anton van den Hengel. Explainable deep few-shot anomaly detection with deviation networks. arXiv preprint arXiv:2108.00462, 2021. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "text", + "content": "[22] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In ICML, pages 1530-1538. PMLR, 2015. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "type": "text", + "content": "[23] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Scholkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In CVPR, pages 14318-14328, 2022. 1, 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "type": "text", + "content": "[24] Lukas Ruff, Robert A Vandermeulen, Nico Gornitz, Alexander Binder, Emmanuel Müller, Klaus-Robert Müller, and Marius Kloft. Deep semi-supervised anomaly detection. In ICLR, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "type": "text", + "content": "[25] Maximilian Schwab, Charles Madeline-Dérou, Steffen Klarmann, Nils Thielen, Sven Meier, Jörg Franke, Sandan Chintanippu, and Wilhelm Stork. Multi-model machine learning based industrial vision framework for assembly part quality control. In *ETFA*, pages 1-4, 2022. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "type": "text", + "content": "[26] Domen Tabernik, Samo Šela, Jure Skvarč, and Danijel Skočaj. Segmentation-based deep-learning approach for surface-defect detection. In Journal of Intelligent Manufacturing, 31(3):759-776, 2020. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 455, + 545, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 486 + ], + "type": "text", + "content": "[27] Matthias Wieler and Tobias Hahn. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, 2007. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 488, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 532 + ], + "type": "text", + "content": "[28] Xincheng Yao, Ruoqi Li, Jing Zhang, Jun Sun, and Chongyang Zhang. Explicit boundary guided semi-push-pull contrastive learning for supervised anomaly detection. In CVPR, pages 24490-24499, 2023. 1, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 533, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 576 + ], + "type": "text", + "content": "[29] Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 578, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 610 + ], + "type": "text", + "content": "[30] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC. British Machine Vision Association, 2016. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 612, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 655 + ], + "type": "text", + "content": "[31] Hui Zhang, Zuxuan Wu, Zheng Wang, Zhineng Chen, and Yu-Gang Jiang. Prototypical residual networks for anomaly detection and localization. In CVPR, pages 16281-16291, 2023. 1, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 657, + 545, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 657, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 307, + 657, + 545, + 700 + ], + "type": "text", + "content": "[32] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In ECCV, pages 392-408. Springer, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17762" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/91e475e8-d317-465c-893c-024b420e20a0_content_list.json b/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/91e475e8-d317-465c-893c-024b420e20a0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..383998ef59bd3f4d5b66b4ca2a79d2698d15a593 --- /dev/null +++ b/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/91e475e8-d317-465c-893c-024b420e20a0_content_list.json @@ -0,0 +1,1428 @@ +[ + { + "type": "text", + "text": "Suppress and Rebalance: Towards Generalized Multi-Modal Face Anti-Spoofing", + "text_level": 1, + "bbox": [ + 78, + 130, + 888, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xun Lin $^{1}$ , Shuai Wang $^{1}$ , Rizhao Cai $^{2}$ , Yizhong Liu $^{1}$ , Ying Fu $^{3}$ , Wenzhong Tang $^{1}$ , Zitong Yu $^{4*}$ , Alex Kot $^{2}$ $^{1}$ Beihang University $^{2}$ Nanyang Technological University $^{3}$ Beijing Institute of Technology $^{4}$ Great Bay University", + "bbox": [ + 86, + 179, + 893, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{linxun, wangshuai}@buaa.edu.cn yuzitong@gbu.edu.cn", + "bbox": [ + 250, + 236, + 725, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 312, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Face Anti-Spoofing (FAS) is crucial for securing face recognition systems against presentation attacks. With advancements in sensor manufacture and multi-modal learning techniques, many multi-modal FAS approaches have emerged. However, they face challenges in generalizing to unseen attacks and deployment conditions. These challenges arise from (1) modality unreliability, where some modality sensors like depth and infrared undergo significant domain shifts in varying environments, leading to the spread of unreliable information during cross-modal feature fusion, and (2) modality imbalance, where training overly relies on a dominant modality hinders the convergence of others, reducing effectiveness against attack types that are indistinguishable by sorely using the dominant modality. To address modality unreliability, we propose the Uncertainty-Guided Cross-Adapter (U-Adapter) to recognize unreliable detected regions within each modality and suppress the impact of unreliable regions on other modalities. For modality imbalance, we propose a Rebalanced Modality Gradient Modulation (ReGrad) strategy to rebalance the convergence speed of all modalities by adaptively adjusting their gradients. Besides, we provide the first large-scale benchmark for evaluating multi-modal FAS performance under domain generalization scenarios. Extensive experiments demonstrate that our method outperforms state-of-the-art methods. Source codes and protocols are released on https://github.com/OMGGGGG/mmdg.", + "bbox": [ + 76, + 318, + 473, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 757, + 207, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The remarkable success of deep neural networks has also brought concerns regarding their security vulnerabilities [56, 57, 64]. Especially face recognition (FR) systems, widely used in applications like surveillance and mobile payment [63], are susceptible to various face presentation attacks, including printed photos, video replays, and 3D", + "bbox": [ + 76, + 782, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2cb3603aa2d2e6fa2d850ae99159b9478ae02cbb2ef8f0081681452b21871964.jpg", + "image_caption": [ + "Figure 1. Illustration of DG scenarios in the context of (a) unimodal and (b) multi-modal. (c) DG performance of SSDG [20] on our Protocol 1 (see Sec. 4.1). Though fed with more modalities, SSDG performs worse in multi-modal scenarios compared to unimodal ones. * denotes our re-implemented multi-modal version." + ], + "image_footnote": [], + "bbox": [ + 511, + 287, + 874, + 479 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/e49bef10b176442fa7039eec061b7e7fb1eafc85fac4cc242be3b016d3cc92c8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
(c)
MethodModalityHTER (%) ↓AUC (%) ↑
SSDG(a): RGB24.0783.69
SSDG*(b): RGB+D+I26.09 (+2.02)82.03 (-1.66)
", + "bbox": [ + 511, + 483, + 875, + 536 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wearable masks [3]. These vulnerabilities challenge the security of FR systems, limiting their broader application. As a result, face anti-spoofing (FAS) methods are developed to enhance the security of FR systems.", + "bbox": [ + 496, + 612, + 890, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, deep learning has been widely applied in FAS, particularly in unimodal scenarios based on the RGB spectrum. While these models perform well in intra-dataset evaluations, their generalizability to unseen attacks and deployment environments remains limited [2]. Despite the improvement of robustness via domain generalization (DG) techniques [20, 28, 49, 50, 71], these approaches still tend to over-fit the attacks presented in source domains.", + "bbox": [ + 496, + 674, + 892, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the recent progress of multi-modal learning [31, 52, 54, 68] and the development of sensor manufacture [62], some large-scale multi-modal FAS datasets [15, 34, 45, 67] are proposed, which also drive the advancement in multimodal FAS [13, 14, 33, 36, 58, 60, 61]. By expanding the input from RGB to multi-modal, e.g., RGB, depth, and infrared, richer spoofing traces can be captured for detecting", + "bbox": [ + 496, + 795, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 753, + 45 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "211", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "various presentation attacks [63]. However, existing multimodal methods fail to perform well under unseen deployment conditions. As illustrated in Fig. 1, though the current methods generalize well in unimodal scenarios, their performance abnormally declines when extended and applied to multi-modal FAS. Why using more modalities is not better? We argue that there are two main reasons: (1) Modality unreliability. Deploying in unseen environments or with various kinds of models of imaging sensors within the same modality, especially the depth and infrared shown in Fig. 1 (b), might introduce significant domain shifts [66]. Such shifts can result in unreliable multimodal live/spoof feature extraction [51], misleading mutual modalities through cross-modal fusion. (2) Modality imbalance. In some multi-modal learning scenarios, models tend to overly rely on a dominant modality with the highest convergence speed (i.e. fast modality), which prevents itself from fully exploiting the other dominated modalities with relatively slower convergence speeds (i.e. slow modalities) [43]. In FAS, modality imbalance can cause significant negative impacts, especially when the fast modality is unsuitable or unavailable for detecting attacks in the target domain, relying on slow modalities that have not adequately converged cannot achieve satisfactory performance. This study addresses the cross-domain limitations of multimodal FAS approaches due to modality unreliability and imbalance. As illustrated in Fig. 2, we introduce a MultiModal Domain Generalized (MMDG) framework for FAS. Within MMDG, we develop trainable Uncertainty-guided cross-Adapters (U-Adapter) to fine-tune Vision Transformers (ViT) [9] The U-Adapter address modality unreliability leveraging the uncertainty of each modality to suppress the focus on unreliable tokens during the cross-modal fusion. This prevents the excessive propagation of unreliable spoofing traces which tends to decrease the model's discriminative capability. Meanwhile, to solve modality imbalance, we design Rebalanced modality Gradient modulation strategy (ReGrad). Our ReGrad rebalance the convergence speed of all modalities by dynamically adjusting the backward propagated gradients of all trainable parameters in U-Adapters based on their conflict degree and convergence speed. This ensures that all modalities can be fully leveraged to resist various unseen attacks in the target domain. Our contributions are as follows.", + "bbox": [ + 76, + 90, + 472, + 758 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the MMDG framework to enhance the domain generalizability by addressing the modality unreliability and imbalance. This is the first investigation for multi-modal FAS under multi-source DG scenarios.", + "- Within MMDG, we propose a Uncertainty-Guided Cross-Adapter (U-Adapter) for ViT-based multi-modal FAS, which can complementarily fuse cross-modal features and prevents each modality from being influenced by unreliable information caused by domain shifts." + ], + "bbox": [ + 76, + 763, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We design the Rebalanced Modality Gradient Modulation (ReGrad) for MMDG, a strategy that can monitor and rebalance the convergence speed of each modality by adaptively modulating gradients of U-Adapters.", + "- We build the first large-scale benchmark to evaluate the domain generalizability of multi-modal FAS approaches. Extensive experiments on this benchmark demonstrate the superiority and generalizability of MMDG." + ], + "bbox": [ + 500, + 90, + 890, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 227, + 650, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Domain Generalization for FAS", + "text_level": 1, + "bbox": [ + 500, + 253, + 779, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In FAS, domain generalization focuses on training a model using multiple source domains, with the intention that this model can work effectively on unseen target domains [18, 46]. Previous works verify the effectiveness of adversarial training [22, 37, 65], asymmetric triplet loss [20, 30] and controversial learning [38, 49] in developing a shared feature space across domains. Besides, some studies [50, 70, 71] investigate the style information of different domains, aiming to learn domain-invariant features by minimizing distinct style characteristics. Furthermore, meta-learning-based methods [4, 10, 21, 44, 69] are also discussed to simulate domain shifts during training, fostering the learning of a robust representative feature space to such variations. More recent works focus on parameter-efficient transfer learning [5, 6, 18, 48], enabling a pretrained ViT to efficiently adapt to unseen domains with a lower risk of over-fitting. However, these approaches are proposed for unimodal (i.e. RGB) FAS. Due to the ignorance of modality unreliability, there is a gap in adapting them to multimodal scenarios. Simple multi-modal extensions of these approaches, through early fusion, late fusion, or even the substitution with multi-modal backbones, still exhibit limited performance under multi-modal DG scenarios.", + "bbox": [ + 496, + 276, + 893, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Multi-Modal FAS", + "text_level": 1, + "bbox": [ + 500, + 635, + 671, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-modal FAS refers to the use of multiple spectra, e.g., RGB, depth, and infrared, to reveal live and spoofing traces. Due to the complementary information among different modalities, traces that are undetectable in one modality can be easily captured in others [17, 36]. To fully exploit information from each modality, early methods employ channelwise concatenation of inputted modalities [14, 42], or develop multiple separate branches for modality feature extraction followed by late fusion [25, 26, 47, 58]. Recent works introduce attention-based feature fusion techniques [8, 27] and adaptive cross-modal loss functions [13] to encourage the extraction of complementary information among modalities. Meanwhile, cross-modality translation is discussed [29, 35] to mitigate the semantic gap between different modalities. To further enhance the practical usability of FAS, Yu et al. [60, 61, 62] introduce flexible-modal", + "bbox": [ + 496, + 659, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "212", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5fbb5220c5174dda9f95f70a11e7cd3f3b9b03821544f1fcd2b12dba8c72b718.jpg", + "image_caption": [ + "Figure 2. Overall structure of our MMDG. It consists of ViT-based backbones fine-tuned by the proposed U-Adapters with the modality rebalancing strategy called ReGrad. Each modality is designed with a branch for feature extraction and enables feature interaction and mutual complementarity with other modalities. For simplicity, we illustrate the two-modality scenario." + ], + "image_footnote": [], + "bbox": [ + 106, + 90, + 434, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "benchmarks that force the use of incomplete modalities during training or testing. In the context of flexible-modal FAS, cross-modality attention [33, 36] or multi-modal adapters [60, 61] are proposed for pretrained ViT to learn modality-agnostic live/spoofing features. However, the aforementioned methods fail to perform well in DG scenarios. The lack of sensitivity to modality unreliability and imbalance leads to an insufficient ability to resist domain shifts. Although previous works rebalance modalities by modulating gradients [43] or adjusting modality losses [11], these methods ignore domain shifts and are not suitable for FAS frameworks with complex cross-modal fusion modules.", + "bbox": [ + 75, + 452, + 472, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed MMDG", + "text_level": 1, + "bbox": [ + 76, + 646, + 250, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, our MMDG consists of ViT-based backbones fine-tuned by the proposed U-Adapters and ReGrad strategy, which are respectively designed for solving modality unreliability and imbalance.", + "bbox": [ + 75, + 672, + 468, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Fine-tuning ViT with U-Adapters", + "text_level": 1, + "bbox": [ + 76, + 741, + 369, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In multi-modal FAS, DG scenarios involving significant domain shifts [32], including unseen face presentations attacks [51], different kinds of imaging sensors, and noise from low-quality sensors [66], often lead to the extraction of unreliable live/spoofing features from certain modal inputs. These unreliable features can propagate during cross-modal fusion, adversely affecting other modalities. To this end, we aim to identify unreliable local features within each modality and suppress the interaction of these features", + "bbox": [ + 75, + 763, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with other modalities, ensuring that each modality utilizes reliable information from other modalities as complementary traces. Specifically, we adopt uncertainty estimated via Monte Carlo sampling [12, 23, 40] to evaluate the feature unreliability in each modality, and based on the uncertainty, we reduce their weight in cross-modal attention-based feature fusion, minimizing their negative impact. Inspired by recent success in parameter-efficient transfer learning (PETL) by introducing adapters into pretrained ViT [5, 6, 18, 39, 60, 61], but distinctively, we propose the U-Adapter for fine-tuning, which enables cross-modal feature fusion and recognizes as well as suppresses the propagation of unreliable information from uncertainly detected regions caused by domain shifts. In Fig. 3, we present the process of fine-tuning ViT with the U-Adapters, including adopting an Uncertainty Estimation Module (UEM) to recognize unreliable tokens and employing the U-Adapter for uncertainty-aware cross-modal fusion. The ViT [9], comprising $n$ transformer blocks each with Layer Normalization (LN), Multi-Head Self Attention (MHSA), and Multi-Layer Perceptron (MLP), is fine-tuned at the MLP's output using our U-Adapter. By representing our U-Adapter as $\\mathcal{A}$ and taking RGB (R), depth (D), and infrared (I) as the inputted modalities, the fine-tuning process is formulated as follows.", + "bbox": [ + 496, + 90, + 890, + 452 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {x} _ {R} ^ {\\text {o u t}} = \\mathcal {A} \\left(\\mathbf {x} _ {D} ^ {U}, \\mathbf {x} _ {D} ^ {3}, \\mathbf {x} _ {R} ^ {3}\\right) + \\mathcal {A} \\left(\\mathbf {x} _ {I} ^ {U}, \\mathbf {x} _ {I} ^ {3}, \\mathbf {x} _ {R} ^ {3}\\right) + \\mathbf {x} _ {R} ^ {3} + \\mathbf {x} _ {R} ^ {4}, (1) \\\\ \\mathbf {x} _ {D} ^ {o u t} = \\mathcal {A} \\left(\\mathbf {x} _ {R} ^ {U}, \\mathbf {x} _ {R} ^ {3}, \\mathbf {x} _ {D} ^ {3}\\right) + \\mathbf {x} _ {D} ^ {3} + \\mathbf {x} _ {D} ^ {4}, (2) \\\\ \\mathbf {x} _ {I} ^ {o u t} = \\mathcal {A} \\left(\\mathbf {x} _ {R} ^ {U}, \\mathbf {x} _ {R} ^ {3}, \\mathbf {x} _ {I} ^ {3}\\right) + \\mathbf {x} _ {I} ^ {3} + \\mathbf {x} _ {I} ^ {4}, (3) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 455, + 890, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where for $m \\in \\{R, D, I\\}$ , $\\mathbf{x}_m^U \\in \\mathbb{R}^{B \\times L \\times 1}$ are the estimated token uncertainty in the corresponding modality, and $\\mathbf{x}_m^3, \\mathbf{x}_m^4 \\in \\mathbb{R}^{B \\times L \\times C}$ are the outputs of the last LN and MLP, respectively. The meaning of other vectors is illustrated in Fig. 3. To prevent the modalities that experienced significant domain shifts, i.e. depth and infrared, from misleading each other, we prohibit their direct interaction.", + "bbox": [ + 496, + 505, + 890, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Uncertainty Estimation Module. As discussed above, it is essential to recognize unreliable modal information for FAS. In deep learning, uncertainty reflects the level of confidence a model has in its predictions. When the model is less certain about a prediction, we consider that prediction to be unreliable [12]. When a model takes images with substantial noise caused by significant domain shifts as inputs, uncertainty estimation techniques [12] tend to assign elevated uncertainty values to regions more likely to be misclassified [32]. One well-known approach is the Bayesian Neural Network (BNN), widely applied in various vision tasks [16, 19]. BNN focuses on learning the posterior distribution of the model's weights rather than output a single value. Since the posterior distribution is always not tractable [12], approximate inference methods, such as Monte Carlo dropout (MCD) [7, 40, 53], are employed for estimation. In this work, we also use MCD to construct the UEM as a probabilistic representational model, which imposes a Bernoulli distribution over the weights of a model", + "bbox": [ + 496, + 613, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "213", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6dc01f823a1b95f5cd3632c0351fb3d87275aea4df21dd8f5b7028b03f384e7b.jpg", + "image_caption": [ + "Figure 3. (a1)-(a2) Illustration of fine-tuning ViT with proposed U-Adapters, showcasing the interaction between the RGB (R) and Depth (D) modalities. Note that only parameters of U-Adapters are trainable. (b) Uncertainty Estimation Module (UEM) used for recognizing unreliable tokens. (c) Detailed structure of U-Adapter, which adopts cross-modal fusion and suppresses the interference of unreliable tokens on other modalities. After fusion, discriminative central difference information is integrated for fine-grained spoof representation." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 251, + 253 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a0c3c3d46d0a2cd544e91eb808ae4123941e612adcb06af32fa2e63b5079e7a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 88, + 411, + 253 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/42d110be83df66467e61188eec615a4ae2060dc7db05b6076e73fb1c7f4ca5c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 88, + 500, + 255 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/46fcc5a16d857ea8b5f61b5c7addeca4524310cf0c35685497681953151b4979.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 88, + 736, + 255 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/281db8e4e6573e8d3e06ecc3614d482064112755346ee3f4dc034b4211d0c9eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 88, + 870, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "without introducing extra parameters. As presented in Fig. 3 (b), we perform MCD instead of the vanilla dropout after each MHSA, simulating the effect of sampling from the posterior distribution by randomly dropping out neurons in MHSA. According to Bayesian probability theory, we consider $\\mathcal{T} = \\{\\mathbf{x}_R^{u_1},\\mathbf{x}_R^{u_2},\\dots ,\\mathbf{x}_R^{u_t}\\}$ as empirical samples from the approximate distribution, and use their token-wise variance $\\mathbf{x}_R^U = \\mathrm{Var}(\\mathcal{T})\\in \\mathbb{R}^{B\\times L\\times 1}$ to reflect their unreliability.", + "bbox": [ + 75, + 320, + 468, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Uncertainty-Guided Cross-Adapter. Different from existing cross-modal fusion [33, 36], our U-Adapter, guided by the estimated uncertainty, suppresses the interaction with unreliable tokens from other modalities to ensure the reliability of the fused features. As shown in Fig. 3 (c), U-Adapter interactively fuses features from different modalities and suppresses unreliable information during fusion. For example, in the depth branch (see Fig. 3 (a2)), U-Adapter takes the output of the second LN from another modality $\\mathbf{x}_R^3$ as the query for attention-based feature fusion. The output of the second LN from its own modality $\\mathbf{x}_D^3$ is used as key and value. Moreover, U-Adapter modulates the query tokens from the other modality based on their uncertainty. Since malicious presentation attacks result in subtle traces, we adopt central difference convolution (CDC) [59] to integrate the local details for ViT and explore fine-grained difference information from neighbor tokens, thus making the fused features more discriminative. By ignoring GeLU and vanilla convolution layers, the U-Adapter is formulated as follows.", + "bbox": [ + 75, + 443, + 470, + 743 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} (\\mathbf {x} _ {R} ^ {U}, \\mathbf {x} _ {R} ^ {3}, \\mathbf {x} _ {D} ^ {3}) = \\mathcal {C} \\left(\\mathcal {S} \\left(\\frac {\\mathrm {Q} (\\mathbf {x} _ {R} ^ {3}) \\mathrm {K} (\\mathbf {x} _ {D} ^ {3}) ^ {\\intercal} \\odot \\mathrm {U} (\\mathbf {x} _ {R} ^ {U})}{\\sqrt {n _ {k}}}\\right) \\mathrm {V} (\\mathbf {x} _ {D} ^ {3}))\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 744, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{C}(\\cdot)$ denotes the CDC layer, $S(\\cdot)$ represents the softmax, and $n_k$ is the number of token channels. $\\mathrm{Q}(\\cdot),\\mathrm{K}(\\cdot)$ and $\\mathrm{V}(\\cdot)$ are the linear projections corresponding to the query, key, and value, respectively. $\\odot$ is the Hadamard product. $\\mathrm{U}(\\mathbf{x}_R^U) = \\exp (-r_e\\cdot \\mathbf{x}_R^U\\times \\mathbf{I})$ normalizes the inputted uncertainty using an exponential function, where $r_e$ is a penalty intensity parameter, and $\\mathbf{I} = [1,1,\\dots ,1]_{1\\times L}$ is adopted to expand the dimensions of $x_R^U$ to match the size", + "bbox": [ + 75, + 777, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of $L \\times L$ . We show the meaning of other vectors in Figs. 3 (a1-a2). Note that Eq. (4) only provides the scenario where RGB tokens are fused into depth in Fig. 3 (c). The pairwise fusion among other modalities follows similarly.", + "bbox": [ + 496, + 320, + 893, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Rebalanced Modality Gradient Modulation", + "text_level": 1, + "bbox": [ + 498, + 393, + 870, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Different modalities are suitable for distinct types of attacks and deployment conditions, e.g., the depth modality being less sensitive to light and more apt for detecting replay attacks, and the infrared modality excelling in identifying wearable mask attacks and performing well in low-light conditions. To cope with various unknown attacks and deployment conditions, we aim to ensure that each modality can sufficiently converge, thereby fully leveraging their discriminative capability, rather than over-relying on a specific modality (i.e. fast modality) that shows superior performance in the source domain.", + "bbox": [ + 496, + 417, + 890, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Current modality balancing methods [11, 43], designed for simple late fusion models, are limited in their applicability to FAS. Late fusion networks have been shown to underperform in FAS compared to hybrid fusion networks [33, 36] that allow the fusion both before the final classifier and during the feature extraction. Additionally, the trainable parameters of a late fusion network are only affected by the gradient of its own modality, whereas those of a hybrid fusion network are influenced by the gradients of all modalities. Moreover, gradients from different modalities may even conflict with each other, making the modulation more difficult. To this end, our ReGrad adaptively modulates the gradients of all modalities that simultaneously act on each trainable layer. Moreover, existing modality balancing methods fail to generalize well while balancing each modality due to the ignorance of domain shifts. Drawing inspiration from SSDG [20] and PMR [11], we introduce a single-side prototypical loss to supervise modality branches and monitor their convergence speed. This loss enables each modality to cluster according to domain information through internal impetus, which further mitigates the domi", + "bbox": [ + 496, + 583, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "214", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ec8e154fde5ae0d5369fd1ed6977acf7deb00a96e0c0dd084c7dd6c278411da3.jpg", + "image_caption": [ + "Figure 4. Illustration of gradient modulation via the proposed ReGrad in different scenarios: (Row 1) Non-conflicting (a1) and faster modality $j$ (b1) or $i$ (c1). (Row 2) Conflicting (a2) and faster modality $j$ (b2) or $i$ (c2)." + ], + "image_footnote": [], + "bbox": [ + 117, + 85, + 851, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "nance of faster modalities.", + "bbox": [ + 75, + 258, + 253, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Modulating Gradients for Each Modality. When the optimization direction is dominated by the gradients of the fast modality, the convergence degree of slow modalities is suppressed. The discriminative ability of slow modalities may even be weakened when there is a conflict in the gradient directions between modalities [55]. Inspired by the studies on solving domain conflicts [41] and multi-task conflicts [55] via gradient modulation, our ReGrad adaptively performs gradient modulation based on the conflict and convergence degree of each modality, making it applicable to complex hybrid fusion networks. The criteria of modality convergence speed will be discussed in the following section.", + "bbox": [ + 75, + 276, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 4 (b1), if modality $i$ converges slower than $j$ and their gradients $\\mathbf{g}_i$ and $\\mathbf{g}_j$ are not conflicting, we suppress $\\mathbf{g}_j$ and boost the learning of modality $i$ . This involves decomposing $\\mathbf{g}_j$ into orthogonal $\\mathbf{g}_j^v$ and non-orthogonal $\\mathbf{g}_j^h$ components relative to $\\mathbf{g}_i$ . We drop $\\mathbf{g}_j^v$ to slow down modality $j$ and use $\\mathbf{g}_j^h$ to accelerate modality $i$ . Furthermore, to prevent unreliable modalities that converge rapidly from misleading other modalities, we employ the estimated uncertainty of modality $j$ to further constrain $g_j$ . Conversely, if $\\mathbf{g}_i$ conflicts with $\\mathbf{g}_j$ and modality $i$ remains a slow modal, it is unreasonable to use any part of $\\mathbf{g}_j$ to boost the learning of $\\mathbf{g}_i$ . To prevent $\\mathbf{g}_j$ from seriously slowing down the convergence speed of modality $i$ , as shown in Fig. 4 (b2), we remove the non-orthogonal part of $\\mathbf{g}_j^h$ that is opposite to $\\mathbf{g}_i$ . Meanwhile, we retain its orthogonal part $\\mathbf{g}_j^v$ to allow modality $j$ to learn at a slower speed and reduce $\\mathbf{g}_j$ 's interference with $\\mathbf{g}_i$ . Similarly, the uncertainty-based suppression is adopted. Cases where the modality $i$ is faster are presented in Figs. 4 (c1)-(c2) and Eq. (5).", + "bbox": [ + 75, + 455, + 470, + 748 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R e G r a d} _ {2} \\left(\\mathbf {g} _ {i}, \\mathbf {g} _ {j}\\right) = \\left\\{ \\begin{array}{l l} \\mathbf {g} _ {i} + \\frac {\\mathbf {g} _ {i} \\cdot \\mathbf {g} _ {j}}{\\| \\mathbf {g} _ {i} \\| ^ {2}} \\mathbf {g} _ {i} \\cdot \\mathrm {U} \\left(u _ {j}\\right) & , \\text {i n F i g . 4 (b 1)}, \\\\ \\mathbf {g} _ {i} + \\left(\\mathbf {g} _ {j} - \\frac {\\mathbf {g} _ {i} \\cdot \\mathbf {g} _ {j}}{\\| \\mathbf {g} _ {i} \\| ^ {2}} \\mathbf {g} _ {i}\\right) \\cdot \\mathrm {U} \\left(u _ {j}\\right) & , \\text {i n F i g . 4 (b 2)}, \\\\ \\frac {\\mathbf {g} _ {i} \\cdot \\mathbf {g} _ {j}}{\\| \\mathbf {g} _ {j} \\| ^ {2}} \\mathbf {g} _ {j} \\cdot \\mathrm {U} \\left(u _ {i}\\right) + \\mathbf {g} _ {j} & , \\text {i n F i g . 4 (c 1)}, \\\\ \\left(\\mathbf {g} _ {i} - \\frac {\\mathbf {g} _ {i} \\cdot \\mathbf {g} _ {j}}{\\| \\mathbf {g} _ {j} \\| ^ {2}} \\mathbf {g} _ {j}\\right) \\cdot \\mathrm {U} \\left(u _ {i}\\right) + \\mathbf {g} _ {j} & , \\text {i n F i g . 4 (c 2)}, \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 81, + 752, + 467, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $u_{i}$ and $u_{j}$ are respectively the batch-wise mean uncertainty of the class tokens outputted by the last ViT block", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of the corresponding modality, denoting the modality uncertainty.", + "bbox": [ + 496, + 258, + 890, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Single-Side Prototypical Loss. Prototype learning has been successfully applied in FAS and modality rebalancing. PMR [11] indicates that introducing prototypes to each modality in multi-modal learning to guide the learning process can mitigate modality imbalance. However, vanilla prototype learning is suboptimal under DG scenarios as it neglects the differences in samples across various domains. Inspired by SSDG [20], we propose the Single-Side Prototypical (SSP) loss with the concept of asymmetry representation and prototype learning. Specifically, we introduce prototypes for the live faces of all domains and the fake faces from each domain. Formulate the domain set as $\\mathcal{D} = \\{d_{\\ell}, d_s^1, d_s^2, \\ldots, d_s^{N_s}\\}$ , where $N_s$ is the number of source domains, $d_{\\ell}$ denotes the domain consists of all live faces, and $d_s^i$ represents domain of spoof faces from the $i$ -th dataset. For modality $m$ , the prototype of domain $d \\in \\mathcal{D}$ is the center of all samples belonging to $d$ in the feature space:", + "bbox": [ + 496, + 290, + 892, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {m} ^ {d} = \\frac {1}{N _ {d}} \\sum_ {c _ {m} \\in d} c _ {m}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 551, + 890, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $c_{m}$ is a sample of $d$ in modality $m$ , and $N_{d}$ denotes the total number of samples in $d$ . Based on the prototypes of each domain, we aim to reduce the distance between samples and the prototype of their corresponding domain and increase the distance from other prototypes, thus preventing conflicts when samples from different domains are trained together. Therefore, following the form of cross-entropy, the SSP loss $\\mathcal{L}_{m}^{ssp}$ of modality $m$ is formulated as follows.", + "bbox": [ + 496, + 587, + 890, + 708 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m} ^ {s s p} \\left(c _ {m}, p _ {m} ^ {d _ {g t}}\\right) = - \\log \\frac {\\exp \\left(- \\mathrm {E D} \\left(c _ {m} , p _ {m} ^ {d _ {g t}}\\right)\\right)}{\\sum_ {d \\in \\mathcal {D}} \\exp \\left(- \\mathrm {E D} \\left(c _ {m} , p _ {m} ^ {d}\\right)\\right)}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 714, + 890, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where sample $c_{m}$ belongs to domain $d_{gt}$ . The function $\\mathrm{ED}(\\cdot)$ denotes the Euclidean distance of the input sample and its prototype. Our SSP loss encourages the features of each modality of the samples to be pulled to the corresponding domain's prototype. This optimization of each modality branch is driven by the intrinsic force within each modality. We mark the modality with lower $\\mathcal{L}_m^{ssp}$ as a faster modality. Our final loss is expressed as follows.", + "bbox": [ + 496, + 750, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\text {f i n a l}} = \\mathcal {L} ^ {c e} + \\sum_ {m \\in \\{R, D, I \\}} \\lambda \\cdot \\mathcal {L} _ {m} ^ {s s p}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 875, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "215", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b3611ee8e5d2274e447d745fa4b852fc9f6df25bb4ea87a188f46137d6fabbe0.jpg", + "table_caption": [ + "Table 1. Cross-dataset testing results under the fixed-modal scenarios (Protocol 1) among CASIA-CeFA (C), PADISI (P), CASIA-SURF (S), and WMCA (W). DG, MM, and FM are short for domain-generalized, multi-modal, and flexible-modal, respectively." + ], + "table_footnote": [], + "table_body": "
MethodTypeCPS → WCPW → SCSW → PPSW → CAverage
HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑
SSDG [20]DG26.0982.0328.5075.9141.8260.5640.4862.3134.2270.20
SSAN [50]DG17.7391.6927.9479.0434.4968.8536.4369.2929.1577.22
IADG [71]DG27.0286.5023.0483.1132.0673.8339.2463.6830.3476.78
ViTAF [18]DG20.5885.8229.1677.8030.7573.0339.7563.4430.0675.02
MM-CDCN [58]MM38.9265.3942.9359.7941.3861.5148.1453.7142.8460.10
CMFL [13]MM18.2288.8231.2075.6626.6880.8536.9366.8228.2678.04
ViT + AMA [61]FM17.5688.7427.5080.0021.1885.5147.4855.5628.4377.45
VP-FAS [60]FM16.2691.2224.4281.0721.7685.4639.3566.5525.4581.08
ViT [9]Baseline20.8884.7744.0557.9433.5871.8042.1556.4535.1667.74
MMDGOurs12.7993.8315.3292.8618.9588.6429.9376.5219.2587.96
", + "bbox": [ + 78, + 116, + 890, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L^{ce}$ denotes the cross-entropy loss, and $\\lambda$ is used to balance the trade-off between $\\mathcal{L}_m^{ssp}$ and $\\mathcal{L}^{ce}$ . Here, we utilize $\\mathcal{L}^{ce}$ to ensure the model's performance, while $\\mathcal{L}_m^{ssp}$ is employed to balance all modalities and enhance generalizability. Since $\\mathcal{L}_m^{ssp}$ is not modality-specific, it does not undergo gradient modulation, which we perform to $\\mathcal{L}^{ce}$ only.", + "bbox": [ + 75, + 250, + 468, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experimental Evaluation", + "text_level": 1, + "bbox": [ + 76, + 366, + 312, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Multi-Modal FAS Benchmark in DG Scenarios", + "text_level": 1, + "bbox": [ + 76, + 392, + 468, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets, Protocols, and Performance Metrics. With the increasing popularity of large-scale foundation models [24], the current trend is towards fully leveraging multiple and diverse source datasets/domains [1]. While training with multiple datasets has been widely studied for unimodal FAS [20, 50, 71], the multi-modal FAS community remains in the phase of single-dataset training [13, 33, 36]. Therefore, we propose the first large-scale multi-source benchmark to evaluate the DG performance of multi-modal FAS. This benchmark incorporates four commonly used multimodal datasets: CASIA-CeFA (C) [34], PADISI-Face (P) [45], CASIA-SURF (S) [67], and WMCA (W) [15], each featuring a variety of attack types. For each dataset, we use RGB, depth (D), and infrared (I) modalities as inputs. Following previous works [20], we use Half Total Error Rate (HTER) and Area Under the Receiver Operating Characteristic Curve (AUC) for evaluation. We design three protocols to evaluate the cross-dataset performance under different unseen deployment conditions, i.e., fixed modalities, missing modalities, and limited source domains. For Protocol 1, inspired by the leave-one-out (LOO) protocols designed in unimodal DG scenarios [20], we propose four multi-modal LOO sub-protocols among C, P, S, and W. For example, the sub-protocol CPS $\\rightarrow$ W represents we take C, P, and S as training sets, while W is testing set. In Protocol 2, for each LOO sub-protocol of Protocol 1, we design three test-time missing-modal scenarios, i.e., D is missing, I is missing, and both D and I are missing. In Protocol 3, we limit the number of source domains by proposing two subprotocols, namely CW $\\rightarrow$ PS and PS $\\rightarrow$ CW. Implementation Details. We resize all RGB, depth, and infrared images to $224 \\times 224 \\times 3$ . Each modality's input is divided", + "bbox": [ + 75, + 417, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "into $14 \\times 14$ patch tokens and a class token is attached, collectively serving as the input to the ViT. The hidden size of each token is default set to 768. We employ the Adam optimizer with a learning rate of $5 \\times 10^{-5}$ and a weight decay of $1 \\times 10^{-3}$ to train 70 epochs. The batch size is set to 32. Our classifier is a single-layer fully-connected network that reduces the dimensionality of the class token outputted by the final ViT block from 768 to 2. The number $n$ of ViT blocks in MMDG is set to 12. ViT-b pretrained on ImageNet is adopted as the baseline.", + "bbox": [ + 496, + 250, + 893, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Cross-dataset Testing", + "text_level": 1, + "bbox": [ + 500, + 415, + 700, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method on Protocols 1-3 with three categories of FAS methods: (1) Multi-modal methods, including CMFL [13] and MM-CDCN [58]. (2) Flexible-modal FAS, such as ViT + AMA [61] and VP-FAS [60]. (3) Domain generalized FAS i.e., SSDG [20], SSAN [50], IADG [71], and ViTAF [18]. For methods not designed for unimodal FAS, we re-implement them either by employing a concatenation of multi-modal inputs or by replacing the backbone with our ViT+U-Adapter or vanilla ViT. All these methods are trained on the proposed protocols. We automatically use their hyper-parameters as described in the corresponding reference papers or optimally assign better ones.", + "bbox": [ + 496, + 439, + 890, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Protocol 1: Fixed Modalities. We evaluated the cross-dataset performance of unimodal DG and multi-modal methods in the fixed modalities scenarios. Since unimodal DG methods cannot be directly applied to this protocol, we concatenate the images of the three modalities along the channel dimension. We then reduce the channel number to 3 using a trainable $1 \\times 1$ convolution layer before feeding them to these methods. As shown in the Table 1, our MMDG achieves the best results. Compared to the second-ranked method, i.e. VP-FAS [60], there is an improvement of over $6\\%$ in both HTER and AUC. A higher performance boost occurs when compared to domain-generalized methods. This indicates MMDG's effectiveness in addressing modality unreliability and imbalance.", + "bbox": [ + 496, + 625, + 890, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Protocol 2: Missing Modalities. As shown in Table 2, our MMDG outperforms other algorithms in all three scenarios of modality missing during testing. It's worth noting that ViT+AMA and VP-FAS are specifically designed", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "216", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/593eb446a9764c40a2502b9e16165b4e0ffed382f86aedcff7c2acd01c06aaae.jpg", + "table_caption": [ + "Table 2. Cross-dataset testing results under the test-time missing-modal scenarios (Protocol 2) among CASIA-CeFA (C), PADISI (P), CASIA-SURF (S), and WMCA (W). We report the average HTER (\\%) ↓ and AUC (\\%) ↑ on four sub-protocols, i.e., CPS → W, CPW → S, CSW → P, and PSW → C. DG, MM, and FM are short for domain-generalized, multi-modal, and flexible-modal, respectively." + ], + "table_footnote": [], + "table_body": "
MethodTypeMissing DMissing IMissing D & IAverage
HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑
SSDG [20]DG38.9265.4537.6466.5739.1865.2238.5865.75
SSAN [50]DG36.7769.2141.2061.9233.5273.3837.1668.17
IADG [71]DG40.7258.7242.1761.8337.5066.9040.1362.49
ViTAF [18]DG34.9973.2235.8869.4035.8969.6135.5970.64
MM-CDCN [58]MM44.9055.3543.6058.3844.5455.0844.3556.27
CMFL [13]MM31.3774.6230.5575.4231.8974.2931.2774.78
ViT + AMA [61]FM29.2577.7032.3074.0631.4875.8231.0175.86
VP-FAS [60]FM29.1378.2729.6377.5130.4776.3129.7477.36
ViTBaseline40.0464.6936.7768.1936.2069.0237.6767.30
MMDGOurs24.8982.3923.3983.8225.2681.8624.5182.69
", + "bbox": [ + 80, + 131, + 888, + 255 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6b0884b9d7039c180df00185a86debdb09130d697b79c60f75d23b7d77a10153.jpg", + "table_caption": [ + "Table 3. Cross-dataset testing results under the limited source domain scenarios (Protocol 3) among CeFA-CeFA (C), PADISI-UGC (P), CASIA-SURF (S), and WMCA (W)." + ], + "table_footnote": [], + "table_body": "
MethodTypeCW → PSPS→ CW
HTER (%)AUC (%)HTER (%)AUC (%)
SSDG [20]DG25.3480.1746.9854.29
SSAN [50]DG26.5580.0639.1067.19
IADG [71]DG22.8283.8539.7063.46
ViTAF [18]DG29.6477.3639.9361.31
MM-CDCN [58]MM29.2876.8847.0051.94
CMFL [13]MM31.8672.7539.4363.17
ViT + AMA [61]FM29.2576.8938.0667.64
VP-FAS [60]FM25.9081.7944.3760.83
ViT [9]Baseline42.6657.8042.7560.41
MMDGOurs20.1288.2436.6070.35
", + "bbox": [ + 80, + 308, + 467, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for scenarios of modality missing and have achieved better results compared to other methods. However, under the DG premise, they did not consider the modality imbalance and unreliability issues brought by domain shift, resulting in a difference of over $5\\%$ in average HTER and AUC compared to our method. Another interesting observation is that SSAN, IADG, ViT, and ViT + AMA perform better in the Missing D & I scenario than in Missing D and Missing I. These methods lose more modalities but achieve performance improvement, indicating that unreliable information within the modalities leads to erroneous detection. This reflects the effectiveness of our approach in suppressing the interaction of unreliable information between modalities.", + "bbox": [ + 75, + 438, + 468, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Protocol 3: Limited Source Domains. Compared to protocol 1, the limitation of source domains results in a more significant domain shift. We observe that our proposed MMDG achieves optimal results in both sub-protocols (see Table 3), even in the presence of substantial domain shift in $\\mathbf{PS} \\rightarrow \\mathbf{CW}$ . This further underscores the effectiveness of our U-Adapter and ReGrad in enhancing domain generalizability in unseen multi-modal deployment environments.", + "bbox": [ + 75, + 637, + 468, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study and Discussion", + "text_level": 1, + "bbox": [ + 76, + 768, + 349, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we perform ablation analysis on U-Adapter and ReGrad to verify their individual contribution.", + "bbox": [ + 75, + 792, + 467, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of U-Adapter. To validate the effectiveness of the proposed U-Adapter, we compare the network performance before and after removing the adapter. We also replace the backbone of existing DG methods (i.e., SSDG, SSAN, and IADG) and multi-modal methods (CMFL) to", + "bbox": [ + 75, + 824, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/77cc751d4358442610afe8b2e3ec767b661dbdec20c8f643527b0be757f6b190.jpg", + "image_caption": [ + "Figure 5. Ablation results on our U-Adapter. We report average HTER $\\downarrow$ and AUC $\\uparrow$ on four sub-protocols in Protocol 1." + ], + "image_footnote": [], + "bbox": [ + 500, + 267, + 686, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3df8afb9f8c1b203173ed41cbc16dabb36e89caaff02f38ac14f203e4388e45a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 267, + 888, + 376 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c61e0fd24ebe4659961d4b5134c41984b29037f5386fcba89f7cd24bf3430564.jpg", + "table_caption": [ + "Table 4. Ablation results on the proposed ReGrad. We report the average HTER and AUC on four sub-protocols in Protocol 1." + ], + "table_footnote": [], + "table_body": "
BackboneMethodHTER (%)AUC (%)
ViT [9]-31.1474.81
ViT [9]ReGrad30.5975.94
ViT + U-AdapterSSDG [20]26.1380.31
ViT + U-AdapterSSAN [50]26.9479.52
ViT + U-AdapterIADG [71]24.2382.54
ViT + U-AdapterPMR [11]24.5483.14
ViT + U-AdapterOGM-GE [43]29.3276.54
ViT + U-Adapter-27.3878.88
ViT + U-AdapterReGrad (Conflicted-only)21.9185.65
ViT + U-AdapterReGrad (Unconflicted-only)21.2385.51
ViT + U-AdapterReGrad19.4887.48
", + "bbox": [ + 504, + 441, + 888, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "observe the performance differences when using ViT+U-Adapter as the backbone versus using ViT alone. As shown in Fig. 5, compared to using ViT as the backbone, these methods experience a performance boost when additional U-Adapters are employed. The performance improvement is especially pronounced when no additional method is used, and when our proposed ReGrad is integrated. This demonstrates the effectiveness of U-Adapter and its universality for existing DG methods and multi-modal strategies.", + "bbox": [ + 496, + 579, + 890, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of ReGrad. Here, we try to answer two questions: (1) Whether the idea to address domain shift through modality rebalancing, is more suitable compared to existing DG methods? (2) Is our ReGrad more capable of enhancing the performance of complex cross-modal fusion frameworks like ViT+U-Adapter compared to existing modality balancing methods? From Table 4, we can observe that when ViT+U-Adapter is used as the backbone, our ReGrad surpasses existing DG methods and other modality rebalancing methods. Even for vanilla ViT, our ReGrad can enhance its performance. This indicates the effectiveness of addressing domain shifts in unseen scenarios by balancing", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "217", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc83a24007af5a6d5712c374437e907df9fe01353c6de4f63e2a7ca909eb30af.jpg", + "image_caption": [ + "(a) w/o ReGrad" + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 272, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fabe1805ccfd00ed388beff9bbf28d3441d6ca4374c0d311fcc15b6886a0f9e7.jpg", + "image_caption": [ + "(b) w/ ReGrad", + "Figure 6. Modality balance degree during training. (a) MMDG w/o ReGrad. (b) MMDG w/ ReGrad. The red dashed line in each sub-figure represents the variance of $\\mathcal{L}_R^{ssp}$ , $\\mathcal{L}_D^{ssp}$ , and $\\mathcal{L}_I^{ssp}$ ." + ], + "image_footnote": [], + "bbox": [ + 272, + 88, + 467, + 222 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6ef336f886044346db5b481314e73aa972eac52787f24190ebf828414073aab9.jpg", + "table_caption": [ + "Table 5. Impact of estimated uncertainty in U-Adapters and ReGrad. We report the average HTER $(\\%)$ ↓ and AUC $(\\%)$ ↑ on four sub-protocols in Protocol 1." + ], + "table_footnote": [], + "table_body": "
Using Estimated UncertaintyMetric Value
U-Adapter: Eq. (4)ReGrad: Eq. (5)HTER (%) ↓AUC (%) ↑
××28.4078.65
×✓26.1381.60
✓×22.7584.84
✓✓19.4887.48
", + "bbox": [ + 81, + 330, + 467, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "various modalities. The results in Table 4 also indicate that when compared to the SoTA modality balancing methods (i.e., OGM-GE and PMR), our ReGrad is more suitable for ViT + U-Adapter, which involves hybrid feature fusion and complex cross-modal interactions.", + "bbox": [ + 75, + 429, + 468, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Meanwhile, as illustrated in Fig. 6, the lower variance (right) of $\\mathcal{L}_R^{ssp}$ , $\\mathcal{L}_D^{ssp}$ , and $\\mathcal{L}_I^{ssp}$ demonstrates that ReGrad obviously improves the balance degree of MMDG. We also notice that ReGrad causes slight fluctuations in the loss function values. This is due to the gradient modulation allowing the model to escape from local optima, seeking a direction with a higher degree of modality balance. Despite the fluctuations, MMDG can still converge after about 50 epochs. Additionally, we compare ReGrad with its two variants, which perform modulation only when modality gradients are conflicted and unconflicted, respectively. Although these two variants improve performance compared to not modulating gradients at all, their performance is inferior to modulating gradients in all situations. This validates the necessity of adaptive modulation under various scenarios.", + "bbox": [ + 75, + 505, + 468, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the Uncertainty in U-Adapter and ReGrad. To testify whether the estimated uncertainty can enhance detection performance, we made a comparison by removing the parts that use uncertainty for constraints in U-Adapter and ReGrad. From Table 5, it can be observed that adding uncertainty to constrain the propagation of unreliable information in U-Adapter and the gradient of fast modalities in ReGrad both leads to performance improvement. When uncertainty is added to both modules, HTER and AUC can be further improved by over $3\\%$ . As shown in Fig. 7, MMDG can identify unreliable regions in depth (left)", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/68e22734c7a5fea82f2c054cb969515725eabf94c4f1f355e5717e2b1c50900c.jpg", + "image_caption": [ + "Figure 7. Visual attention and uncertainty maps of our MMDG. Uncertainly predicted regions are highlighted with red boxes. The redder color indicates the higher uncertainty or more attention." + ], + "image_footnote": [], + "bbox": [ + 513, + 88, + 879, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and infrared (right) modalities. Although the U-Adapters of these modalities focus on regions with uncertain predictions, due to our idea of suppressing uncertain regions, these regions fail to mislead mutual modalities, enabling the capture of more reliable spoof traces against domain shifts.", + "bbox": [ + 496, + 282, + 890, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussion. The results in Tables 1-5 and Figs. 5-6 validate the MMDG's generalizability by addressing the problems of modality unreliability and imbalance. This work marks a stride towards practical deployments and can benefit the FAS community. However, our MMDG's performance can still be improved. Continued efforts on this benchmark are necessary for further progress. Meanwhile, our method also has space for improvement, e.g., further enhancing the degree of modality balance and adopting more efficient uncertainty estimation techniques.", + "bbox": [ + 496, + 361, + 890, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 526, + 617, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose the first multi-modal DG framework for FAS, i.e. MMDG, to enhance generalizability by addressing modality unreliability and imbalance. To mitigate the interference of unreliable modalities, we propose U-Adapter to suppress uncertainly predicted regions during cross-modal feature fusion. Besides, we develop ReGrad to adaptively modulate the gradients of each modality branch to achieve modality balance, thus preventing an over-reliance on a dominant modality. We also establish the first large-scale benchmark for multi-modal FAS under DG scenarios. Extensive experiments demonstrate MMDG's effectiveness.", + "bbox": [ + 496, + 551, + 890, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This research was done at the Rapid-Rich Object Search (ROSE) Lab, Nanyang Technological University, and supported by the National Key Research and Development Program of China (Grant No.2022YFB3207700), the National Natural Science Foundation of China (Grants No.62272022, 62306061, and 62331006), the NTU-PKU Joint Research Institute (sponsored by the Ng Teng Fong Charitable Foundation), the Science and Technology Foundation of Guangzhou Huangpu Development District (Grant No.2022GH15), and the Guangdong Provincial Regional Joint Fund - Regional Cultivation Project (Grant No.2023A1515140037)", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "218", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 80, + 90, + 171, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George van den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, Diego de Las Casas, Aurelia Guy, Jacob Menick, Roman Ring, Tom Hennigan, Saffron Huang, Loren Maggiore, Chris Jones, Albin Cassirer, Andy Brock, Michela Paganini, Geoffrey Irving, Oriol Vinyals, Simon Osindero, Karen Simonyan, Jack W. Rae, Erich Elsen, and Laurent Sifre. Improving language models by retrieving from trillions of tokens. In Proceedings of the International Conference on Machine Learning, pages 2206-2240, 2022.", + "[2] Rizhao Cai and Changsheng Chen. Learning deep forest with multi-scale local binary pattern features for face antispoofing. arXiv, abs/1910.03850, 2019.", + "[3] Rizhao Cai, Haoliang Li, Shiqi Wang, Changsheng Chen, and Alex C. Kot. Drl-fas: A novel framework based on deep reinforcement learning for face anti-spoofing. IEEE Transactions on Information Forensics and Security, 16:937-951, 2021.", + "[4] Rizhao Cai, Zhi Li, Renjie Wan, Haoliang Li, Yongjian Hu, and Alex C. Kot. Learning meta pattern for face antispoofing. IEEE Transactions on Information Forensics and Security, 17:1201-1213, 2022.", + "[5] Rizhao Cai, Yawen Cui, Zhi Li, Zitong Yu, Haoliang Li, Yongjian Hu, and Alex C. Kot. Rehearsal-free domain continual face anti-spoofing: Generalize more and forget less. In Proceedings of the IEEE International Conference on Computer Vision, 2023.", + "[6] Rizhao Cai, Zitong Yu, Chenqi Kong, Haoliang Li, Changsheng Chen, Yongjian Hu, and Alex C. Kot. S-adapter: Generalizing vision transformer for face anti-spoofing with statistical tokens. arXiv, abs/2309.04038, 2023.", + "[7] Pamela Carreno, Dana Kulic, and Michael Burke. Adapting neural models with sequential monte carlo dropout. In Conference on Robot Learning, pages 1542-1552, 2022.", + "[8] Pengchao Deng, Chenyang Ge, Xin Qiao, Hao Wei, and Yuan Sun. Attention-aware dual-stream network for multimodal face anti-spoofing. IEEE Transactions on Information Forensics and Security, 18:4258-4271, 2023.", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations, 2021.", + "[10] Zhekai Du, Jingjing Li, Lin Zuo, Lei Zhu, and Ke Lu. Energy-based domain generalization for face anti-spoofing. In Proceedings of the ACM International Conference on Multimedia, pages 1749-1757, 2022.", + "[11] Yunfeng Fan, Wenchao Xu, Haozhao Wang, Junxiao Wang, and Song Guo. PMR: prototypical modal rebalance for multimodal learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 20029-20038, 2023." + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the International Conference on Machine Learning, pages 1050-1059, 2016.", + "[13] Anjith George and Sébastien Marcel. Cross modal focal loss for RGBD face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7882-7891, 2021.", + "[14] Anjith George and Sébastien Marcel. Learning one class representations for face presentation attack detection using multi-channel convolutional neural networks. IEEE Transactions on Information Forensics and Security, 16:361-375, 2021.", + "[15] Anjith George, Zohreh Mostaani, David Geissenbuhler, Olegs Nikisins, André Anjos, and Sébastien Marcel. Biometric face presentation attack detection with multi-channel convolutional neural network. IEEE Transactions on Information Forensics and Security, 15:42-55, 2020.", + "[16] Hongji Guo, Hanjing Wang, and Qiang Ji. Uncertainty-guided probabilistic transformer for complex action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022.", + "[17] Shuangpeng Han, Rizhao Cai, Yawen Cui, Zitong Yu, Yongjian Hu, and Alex C. Kot. Hyperbolic face anti-spoofing. arXiv, abs/2308.09107, 2023.", + "[18] Hsin-Ping Huang, Deqing Sun, Yaojie Liu, Wen-Sheng Chu, Taihong Xiao, Jinwei Yuan, Hartwig Adam, and Ming-Hsuan Yang. Adaptive transformers for robust few-shot cross-domain face anti-spoofing. In Proceedings of the European Conference on Computer Vision, pages 37-54, 2022.", + "[19] Kaixiang Ji, Feng Chen, Xin Guo, Yadong Xu, Jian Wang, and Jingdong Chen. Uncertainty-guided learning for improving image manipulation detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 22456-22465, 2023.", + "[20] Yunpei Jia, Jie Zhang, Shiguang Shan, and Xilin Chen. Single-side domain generalization for face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8481-8490, 2020.", + "[21] Yunpei Jia, Jie Zhang, and Shiguang Shan. Dual-branch meta-learning network with distribution alignment for face anti-spoofing. IEEE Transactions on Information Forensics and Security, 17:138-151, 2022.", + "[22] Fangling Jiang, Qi Li, Pengcheng Liu, Xiang-Dong Zhou, and Zhenan Sun. Adversarial learning domain-invariant conditional features for robust face anti-spoofing. International Journal of Computer Vision, 131(7):1680-1703, 2023.", + "[23] Alex Kendall, Vijay Badrinarayanan, and Roberto Cipolla. Bayesian segnet: Model uncertainty in deep convolutional encoder-decoder architectures for scene understanding. In Proceedings of the British Machine Vision Conference, 2017.", + "[24] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In Proceedings of the IEEE International Conference on Computer Vision, pages 4015-4026, 2023." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "219", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Chenqi Kong, Kexin Zheng, Shiqi Wang, Anderson Rocha, and Haoliang Li. Beyond the pixel world: a novel acoustic-based face anti-spoofing system for smartphones. IEEE Transactions on Information Forensics and Security, 17: 3238-3253, 2022.", + "[26] Chenqi Kong, Kexin Zheng, Yibing Liu, Shiqi Wang, Anderson Rocha, and Haoliang Li. M3fas: An accurate and robust multimodal mobile face anti-spoofing system. arXiv preprint arXiv:2301.12831, 2023.", + "[27] Kaicheng Li, Hongyu Yang, Binghui Chen, Pengyu Li, Biao Wang, and Di Huang. Learning polysemantic spoof trace: A multi-modal disentanglement network for face anti-spoofing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1351-1359, 2023.", + "[28] Zhi Li, Rizhao Cai, Haoliang Li, Kwok-Yan Lam, Yongjian Hu, and Alex C. Kot. One-class knowledge distillation for face presentation attack detection. IEEE Transactions on Information Forensics and Security, 17:2137-2150, 2022.", + "[29] Zhi Li, Haoliang Li, Xin Luo, Yongjian Hu, Kwok-Yan Lam, and Alex C. Kot. Asymmetric modality translation for face presentation attack detection. IEEE Transactions on Multimedia, 25:62-76, 2023.", + "[30] Chen-Hao Liao, Wen-Cheng Chen, Hsuan-Tung Liu, Yi-Ren Yeh, Min-Chun Hu, and Chu-Song Chen. Domain invariant vision transformer learning for face anti-spoofing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6087-6096, 2023.", + "[31] Xun Lin, Shuai Wang, Jiahao Deng, Ying Fu, Xiao Bai, Xinlei Chen, Xiaolei Qu, and Wenzhong Tang. Image manipulation detection by multiple tampering traces and edge artifact enhancement. Pattern Recognition, 133:109026, 2023.", + "[32] Mattia Litrico, Alessio Del Bue, and Pietro Morerio. Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7640-7650, 2023.", + "[33] Ajian Liu and Yanyan Liang. Ma-vit: Modality-agnostic vision transformers for face anti-spoofing. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 1180–1186, 2022.", + "[34] Ajian Liu, Zichang Tan, Jun Wan, Sergio Escalera, Guodong Guo, and Stan Z. Li. CASIA-SURF cefa: A benchmark for multi-modal cross-ethnicity face anti-spoofing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1178–1186, 2021.", + "[35] Ajian Liu, Zichang Tan, Jun Wan, Yanyan Liang, Zhen Lei, Guodong Guo, and Stan Z. Li. Face anti-spoofing via adversarial cross-modality translation. IEEE Transactions on Information Forensics and Security, 16:2759-2772, 2021.", + "[36] Ajian Liu, Zichang Tan, Zitong Yu, Chenxu Zhao, Jun Wan, Yanyan Liang, Zhen Lei, Du Zhang, Stan Z. Li, and Guodong Guo. FM-ViT: Flexible modal vision transformers for face anti-spoofing. IEEE Transactions on Information Forensics and Security, 18:4775-4786, 2023.", + "[37] Yaojie Liu and Xiaoming Liu. Spoof trace disentanglement for generic face anti-spoofing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3813-3830, 2023." + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Yuchen Liu, Yabo Chen, Mengran Gou, Chun-Ting Huang, Yaoming Wang, Wenrui Dai, and Hongkai Xiong. Towards unsupervised domain generalization for face anti-spoofing. In Proceedings of the IEEE International Conference on Computer Vision, pages 20654-20664, 2023.", + "[39] Anwei Luo, Rizhao Cai, Chenqi Kong, Xiangui Kang, Jiwu Huang, and Alex C. Kot. Forgery-aware adaptive vision transformer for face forgery detection. arXiv, abs/2309.11092, 2023.", + "[40] Yuki Mae, Wataru Kumagai, and Takafumi Kanamori. Uncertainty propagation for dropout-based bayesian neural networks. Neural Networks, 144:394-406, 2021.", + "[41] Lucas Mansilla, Rodrigo Echeveste, Diego H. Milone, and Enzo Ferrante. Domain generalization via gradient surgery. In Proceedings of the IEEE International Conference on Computer Vision, pages 6610-6618, 2021.", + "[42] Olegs Nikisins, Anjith George, and Sébastien Marcel. Domain adaptation in multi-channel autoencoder based features for robust face anti-spoofing. In Proceedings of the International Conference on Biometrics, pages 1-8, 2019.", + "[43] Xiaokang Peng, Yake Wei, Andong Deng, Dong Wang, and Di Hu. Balanced multimodal learning via on-the-fly gradient modulation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8228-8237, 2022.", + "[44] Yunxiao Qin, Zitong Yu, Longbin Yan, Zezheng Wang, Chenxu Zhao, and Zhen Lei. Meta-teacher for face antispoofing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6311-6326, 2022.", + "[45] Mohammad Rostami, Leonidas Spinoulas, Mohamed E. Hussein, Joe Mathai, and Wael Abd-Almageed. Detection and continual learning of novel face presentation attacks. In Proceedings of the IEEE International Conference on Computer Vision, pages 14831-14840, 2021.", + "[46] Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C. Yuen. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10023-10031, 2019.", + "[47] Tao Shen, Yuyu Huang, and Zhijun Tong. Facebagnet: Bag-of-local-features model for multi-modal face anti-spoofing. In IEEE Conference on Computer Vision and Pattern Recognition Workshop, pages 1611–1616, 2019.", + "[48] Koushik Srivatsan, Muzammal Naseer, and Karthik Nandakumar. Flip: Cross-domain face anti-spoofing with language guidance. In Proceedings of the IEEE International Conference on Computer Vision, pages 19685-19696, 2023.", + "[49] Yiyou Sun, Yaojie Liu, Xiaoming Liu, Yixuan Li, and WenSheng Chu. Rethinking domain generalization for face antispoofing: Separability and alignment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 24563-24574, 2023.", + "[50] Zhuo Wang, Zezheng Wang, Zitong Yu, Weihong Deng, Jiahong Li, Tingting Gao, and Zhongyuan Wang. Domain generalization via shuffled style assembly for face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4113-4123, 2022." + ], + "bbox": [ + 501, + 92, + 892, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "220", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Junru Wu, Xiang Yu, Buyu Liu, Zhangyang Wang, and Manmohan Chandraker. Uncertainty-aware physically-guided proxy tasks for unseen domain face anti-spoofing. arXiv, abs/2011.14054, 2020.", + "[52] Wuyuan Xie, Kaimin Wang, Yakun Ju, and Miaohui Wang. pmbba: Projection-based blind point cloud quality assessment via multimodal learning. In Proceedings of the ACM International Conference on Multimedia, page 3250-3258, 2023.", + "[53] Lan Xixi, Changchun Zou, Cheng Peng, and Caowei Wu. Uncertainty quantification in intelligent-based electrical resistivity tomography image reconstruction with monte carlo dropout strategy. IEEE Transactions on Geoscience and Remote Sensing, 61:1-16, 2023.", + "[54] Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023.", + "[55] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In Proceedings of the Neural Information Processing Systems, 2020.", + "[56] Yi Yu, Wenhan Yang, Yap-Peng Tan, and Alex C Kot. Towards robust rain removal against adversarial attacks: A comprehensive benchmark analysis and beyond. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6013-6022, 2022.", + "[57] Yi Yu, Yufei Wang, Wenhan Yang, Shijian Lu, Yap-Peng Tan, and Alex C Kot. Backdoor attacks against deep image compression via adaptive frequency trigger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12250-12259, 2023.", + "[58] Zitong Yu, Yunxiao Qin, Xiaobai Li, Zezheng Wang, Chenxu Zhao, Zhen Lei, and Guoying Zhao. Multi-modal face anti-spoofing based on central difference networks. In IEEE Conference on Computer Vision and Pattern Recognition Workshop, pages 2766-2774, 2020.", + "[59] Zitong Yu, Chenxu Zhao, Zezheng Wang, Yunxiao Qin, Zhuo Su, Xiaobai Li, Feng Zhou, and Guoying Zhao. Searching central difference convolutional networks for face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5294-5304, 2020.", + "[60] Zitong Yu, Rizhao Cai, Yawen Cui, Ajian Liu, and Changsheng Chen. Visual prompt flexible-modal face anti-spoofing. arXiv, abs/2307.13958, 2023.", + "[61] Zitong Yu, Rizhao Cai, Yawen Cui, Xin Liu, Yongjian Hu, and Alex C. Kot. Rethinking vision transformer and masked autoencoder in multimodal face anti-spoofing. arXiv, abs/2302.05744, 2023.", + "[62] Zitong Yu, Ajian Liu, Chenxu Zhao, Kevin H. M. Cheng, Xu Cheng, and Guoying Zhao. Flexible-modal face antispoofing: A benchmark. In IEEE Conference on Computer Vision and Pattern Recognition Workshop, pages 6346-6351, 2023.", + "[63] Zitong Yu, Yunxiao Qin, Xiaobai Li, Chenxu Zhao, Zhen Lei, and Guoying Zhao. Deep learning for face anti-spoofing: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5609-5631, 2023." + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Zitong Yu, Rizhao Cai, Zhi Li, Wenhan Yang, Jingang Shi, and Alex C. Kot. Benchmarking joint face spoofing and forgery detection with visual and physiological cues. IEEE Transactions on Dependable and Secure Computing, pages 1-15, 2024.", + "[65] Haixiao Yue, Keyao Wang, Guosheng Zhang, Haocheng Feng, Junyu Han, Errui Ding, and Jingdong Wang. Cyclically disentangled feature translation for face anti-spoofing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3358-3366, 2023.", + "[66] Qingyang Zhang, Haitao Wu, Changqing Zhang, Qinghua Hu, Huazhu Fu, Joey Tianyi Zhou, and Xi Peng. Provable dynamic fusion for low-quality multimodal data. In Proceedings of the International Conference on Machine Learning, pages 41753-41769, 2023.", + "[67] Shifeng Zhang, Ajian Liu, Jun Wan, Yanyan Liang, Guodong Guo, Sergio Escalera, Hugo Jair Escalante, and Stan Z. Li. CASIA-SURF: A large-scale multi-modal benchmark for face anti-spoofing. IEEE transactions on biometrics, behavior, and identity science, 2(2):182-193, 2020.", + "[68] Taoying Zhang, Hesong Li, Qiankun Liu, Xiaoyong Wang, and Ying Fu. Mgt: Modality-guided transformer for infrared and visible image fusion. In Chinese Conference on Pattern Recognition and Computer Vision, pages 321-332, 2023.", + "[69] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Ran Yi, Shouhong Ding, and Lizhuang Ma. Adaptive mixture of experts learning for generalizable face anti-spoofing. In Proceedings of the ACM International Conference on Multimedia, pages 6009-6018, 2022.", + "[70] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Ran Yi, Kekai Sheng, Shouhong Ding, and Lizhuang Ma. Generative domain adaptation for face anti-spoofing. In Proceedings of the European Conference on Computer Vision, pages 335-356, 2022.", + "[71] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Xuequan Lu, Ran Yi, Shouhong Ding, and Lizhuang Ma. Instance-aware domain generalization for face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 20453-20463, 2023." + ], + "bbox": [ + 501, + 90, + 892, + 640 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "221", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/91e475e8-d317-465c-893c-024b420e20a0_model.json b/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/91e475e8-d317-465c-893c-024b420e20a0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0b99ab7c7202054ce94c013ca9069f7076513321 --- /dev/null +++ b/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/91e475e8-d317-465c-893c-024b420e20a0_model.json @@ -0,0 +1,2246 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.754, + 0.046 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.131, + 0.89, + 0.154 + ], + "angle": 0, + "content": "Suppress and Rebalance: Towards Generalized Multi-Modal Face Anti-Spoofing" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.18, + 0.895, + 0.234 + ], + "angle": 0, + "content": "Xun Lin\\(^{1}\\), Shuai Wang\\(^{1}\\), Rizhao Cai\\(^{2}\\), Yizhong Liu\\(^{1}\\), Ying Fu\\(^{3}\\), Wenzhong Tang\\(^{1}\\), Zitong Yu\\(^{4*}\\), Alex Kot\\(^{2}\\) \n\\(^{1}\\)Beihang University \\(^{2}\\)Nanyang Technological University \\(^{3}\\)Beijing Institute of Technology \\(^{4}\\)Great Bay University" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.237, + 0.726, + 0.252 + ], + "angle": 0, + "content": "{linxun, wangshuai}@buaa.edu.cn yuzitong@gbu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.313, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.319, + 0.474, + 0.729 + ], + "angle": 0, + "content": "Face Anti-Spoofing (FAS) is crucial for securing face recognition systems against presentation attacks. With advancements in sensor manufacture and multi-modal learning techniques, many multi-modal FAS approaches have emerged. However, they face challenges in generalizing to unseen attacks and deployment conditions. These challenges arise from (1) modality unreliability, where some modality sensors like depth and infrared undergo significant domain shifts in varying environments, leading to the spread of unreliable information during cross-modal feature fusion, and (2) modality imbalance, where training overly relies on a dominant modality hinders the convergence of others, reducing effectiveness against attack types that are indistinguishable by sorely using the dominant modality. To address modality unreliability, we propose the Uncertainty-Guided Cross-Adapter (U-Adapter) to recognize unreliable detected regions within each modality and suppress the impact of unreliable regions on other modalities. For modality imbalance, we propose a Rebalanced Modality Gradient Modulation (ReGrad) strategy to rebalance the convergence speed of all modalities by adaptively adjusting their gradients. Besides, we provide the first large-scale benchmark for evaluating multi-modal FAS performance under domain generalization scenarios. Extensive experiments demonstrate that our method outperforms state-of-the-art methods. Source codes and protocols are released on https://github.com/OMGGGGG/mmdg." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.758, + 0.208, + 0.773 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.784, + 0.47, + 0.876 + ], + "angle": 0, + "content": "The remarkable success of deep neural networks has also brought concerns regarding their security vulnerabilities [56, 57, 64]. Especially face recognition (FR) systems, widely used in applications like surveillance and mobile payment [63], are susceptible to various face presentation attacks, including printed photos, video replays, and 3D" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.288, + 0.875, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.484, + 0.876, + 0.537 + ], + "angle": 0, + "content": "
(c)
MethodModalityHTER (%) ↓AUC (%) ↑
SSDG(a): RGB24.0783.69
SSDG*(b): RGB+D+I26.09 (+2.02)82.03 (-1.66)
" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.608 + ], + "angle": 0, + "content": "Figure 1. Illustration of DG scenarios in the context of (a) unimodal and (b) multi-modal. (c) DG performance of SSDG [20] on our Protocol 1 (see Sec. 4.1). Though fed with more modalities, SSDG performs worse in multi-modal scenarios compared to unimodal ones. * denotes our re-implemented multi-modal version." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.674 + ], + "angle": 0, + "content": "wearable masks [3]. These vulnerabilities challenge the security of FR systems, limiting their broader application. As a result, face anti-spoofing (FAS) methods are developed to enhance the security of FR systems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.794 + ], + "angle": 0, + "content": "Recently, deep learning has been widely applied in FAS, particularly in unimodal scenarios based on the RGB spectrum. While these models perform well in intra-dataset evaluations, their generalizability to unseen attacks and deployment environments remains limited [2]. Despite the improvement of robustness via domain generalization (DG) techniques [20, 28, 49, 50, 71], these approaches still tend to over-fit the attacks presented in source domains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.902 + ], + "angle": 0, + "content": "With the recent progress of multi-modal learning [31, 52, 54, 68] and the development of sensor manufacture [62], some large-scale multi-modal FAS datasets [15, 34, 45, 67] are proposed, which also drive the advancement in multimodal FAS [13, 14, 33, 36, 58, 60, 61]. By expanding the input from RGB to multi-modal, e.g., RGB, depth, and infrared, richer spoofing traces can be captured for detecting" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "211" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.759 + ], + "angle": 0, + "content": "various presentation attacks [63]. However, existing multimodal methods fail to perform well under unseen deployment conditions. As illustrated in Fig. 1, though the current methods generalize well in unimodal scenarios, their performance abnormally declines when extended and applied to multi-modal FAS. Why using more modalities is not better? We argue that there are two main reasons: (1) Modality unreliability. Deploying in unseen environments or with various kinds of models of imaging sensors within the same modality, especially the depth and infrared shown in Fig. 1 (b), might introduce significant domain shifts [66]. Such shifts can result in unreliable multimodal live/spoof feature extraction [51], misleading mutual modalities through cross-modal fusion. (2) Modality imbalance. In some multi-modal learning scenarios, models tend to overly rely on a dominant modality with the highest convergence speed (i.e. fast modality), which prevents itself from fully exploiting the other dominated modalities with relatively slower convergence speeds (i.e. slow modalities) [43]. In FAS, modality imbalance can cause significant negative impacts, especially when the fast modality is unsuitable or unavailable for detecting attacks in the target domain, relying on slow modalities that have not adequately converged cannot achieve satisfactory performance. This study addresses the cross-domain limitations of multimodal FAS approaches due to modality unreliability and imbalance. As illustrated in Fig. 2, we introduce a MultiModal Domain Generalized (MMDG) framework for FAS. Within MMDG, we develop trainable Uncertainty-guided cross-Adapters (U-Adapter) to fine-tune Vision Transformers (ViT) [9] The U-Adapter address modality unreliability leveraging the uncertainty of each modality to suppress the focus on unreliable tokens during the cross-modal fusion. This prevents the excessive propagation of unreliable spoofing traces which tends to decrease the model's discriminative capability. Meanwhile, to solve modality imbalance, we design Rebalanced modality Gradient modulation strategy (ReGrad). Our ReGrad rebalance the convergence speed of all modalities by dynamically adjusting the backward propagated gradients of all trainable parameters in U-Adapters based on their conflict degree and convergence speed. This ensures that all modalities can be fully leveraged to resist various unseen attacks in the target domain. Our contributions are as follows." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.47, + 0.825 + ], + "angle": 0, + "content": "- We propose the MMDG framework to enhance the domain generalizability by addressing the modality unreliability and imbalance. This is the first investigation for multi-modal FAS under multi-source DG scenarios." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.826, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- Within MMDG, we propose a Uncertainty-Guided Cross-Adapter (U-Adapter) for ViT-based multi-modal FAS, which can complementarily fuse cross-modal features and prevents each modality from being influenced by unreliable information caused by domain shifts." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "- We design the Rebalanced Modality Gradient Modulation (ReGrad) for MMDG, a strategy that can monitor and rebalance the convergence speed of each modality by adaptively modulating gradients of U-Adapters." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.152, + 0.892, + 0.213 + ], + "angle": 0, + "content": "- We build the first large-scale benchmark to evaluate the domain generalizability of multi-modal FAS approaches. Extensive experiments on this benchmark demonstrate the superiority and generalizability of MMDG." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.228, + 0.651, + 0.243 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.254, + 0.78, + 0.269 + ], + "angle": 0, + "content": "2.1. Domain Generalization for FAS" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.895, + 0.626 + ], + "angle": 0, + "content": "In FAS, domain generalization focuses on training a model using multiple source domains, with the intention that this model can work effectively on unseen target domains [18, 46]. Previous works verify the effectiveness of adversarial training [22, 37, 65], asymmetric triplet loss [20, 30] and controversial learning [38, 49] in developing a shared feature space across domains. Besides, some studies [50, 70, 71] investigate the style information of different domains, aiming to learn domain-invariant features by minimizing distinct style characteristics. Furthermore, meta-learning-based methods [4, 10, 21, 44, 69] are also discussed to simulate domain shifts during training, fostering the learning of a robust representative feature space to such variations. More recent works focus on parameter-efficient transfer learning [5, 6, 18, 48], enabling a pretrained ViT to efficiently adapt to unseen domains with a lower risk of over-fitting. However, these approaches are proposed for unimodal (i.e. RGB) FAS. Due to the ignorance of modality unreliability, there is a gap in adapting them to multimodal scenarios. Simple multi-modal extensions of these approaches, through early fusion, late fusion, or even the substitution with multi-modal backbones, still exhibit limited performance under multi-modal DG scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.673, + 0.65 + ], + "angle": 0, + "content": "2.2. Multi-Modal FAS" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Multi-modal FAS refers to the use of multiple spectra, e.g., RGB, depth, and infrared, to reveal live and spoofing traces. Due to the complementary information among different modalities, traces that are undetectable in one modality can be easily captured in others [17, 36]. To fully exploit information from each modality, early methods employ channelwise concatenation of inputted modalities [14, 42], or develop multiple separate branches for modality feature extraction followed by late fusion [25, 26, 47, 58]. Recent works introduce attention-based feature fusion techniques [8, 27] and adaptive cross-modal loss functions [13] to encourage the extraction of complementary information among modalities. Meanwhile, cross-modality translation is discussed [29, 35] to mitigate the semantic gap between different modalities. To further enhance the practical usability of FAS, Yu et al. [60, 61, 62] introduce flexible-modal" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "212" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.092, + 0.436, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.363, + 0.47, + 0.447 + ], + "angle": 0, + "content": "Figure 2. Overall structure of our MMDG. It consists of ViT-based backbones fine-tuned by the proposed U-Adapters with the modality rebalancing strategy called ReGrad. Each modality is designed with a branch for feature extraction and enables feature interaction and mutual complementarity with other modalities. For simplicity, we illustrate the two-modality scenario." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.473, + 0.634 + ], + "angle": 0, + "content": "benchmarks that force the use of incomplete modalities during training or testing. In the context of flexible-modal FAS, cross-modality attention [33, 36] or multi-modal adapters [60, 61] are proposed for pretrained ViT to learn modality-agnostic live/spoofing features. However, the aforementioned methods fail to perform well in DG scenarios. The lack of sensitivity to modality unreliability and imbalance leads to an insufficient ability to resist domain shifts. Although previous works rebalance modalities by modulating gradients [43] or adjusting modality losses [11], these methods ignore domain shifts and are not suitable for FAS frameworks with complex cross-modal fusion modules." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.647, + 0.251, + 0.664 + ], + "angle": 0, + "content": "3. Proposed MMDG" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.735 + ], + "angle": 0, + "content": "As shown in Fig. 2, our MMDG consists of ViT-based backbones fine-tuned by the proposed U-Adapters and ReGrad strategy, which are respectively designed for solving modality unreliability and imbalance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.37, + 0.759 + ], + "angle": 0, + "content": "3.1. Fine-tuning ViT with U-Adapters" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In multi-modal FAS, DG scenarios involving significant domain shifts [32], including unseen face presentations attacks [51], different kinds of imaging sensors, and noise from low-quality sensors [66], often lead to the extraction of unreliable live/spoofing features from certain modal inputs. These unreliable features can propagate during cross-modal fusion, adversely affecting other modalities. To this end, we aim to identify unreliable local features within each modality and suppress the interaction of these features" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.453 + ], + "angle": 0, + "content": "with other modalities, ensuring that each modality utilizes reliable information from other modalities as complementary traces. Specifically, we adopt uncertainty estimated via Monte Carlo sampling [12, 23, 40] to evaluate the feature unreliability in each modality, and based on the uncertainty, we reduce their weight in cross-modal attention-based feature fusion, minimizing their negative impact. Inspired by recent success in parameter-efficient transfer learning (PETL) by introducing adapters into pretrained ViT [5, 6, 18, 39, 60, 61], but distinctively, we propose the U-Adapter for fine-tuning, which enables cross-modal feature fusion and recognizes as well as suppresses the propagation of unreliable information from uncertainly detected regions caused by domain shifts. In Fig. 3, we present the process of fine-tuning ViT with the U-Adapters, including adopting an Uncertainty Estimation Module (UEM) to recognize unreliable tokens and employing the U-Adapter for uncertainty-aware cross-modal fusion. The ViT [9], comprising \\( n \\) transformer blocks each with Layer Normalization (LN), Multi-Head Self Attention (MHSA), and Multi-Layer Perceptron (MLP), is fine-tuned at the MLP's output using our U-Adapter. By representing our U-Adapter as \\( \\mathcal{A} \\) and taking RGB (R), depth (D), and infrared (I) as the inputted modalities, the fine-tuning process is formulated as follows." + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.456, + 0.892, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {x} _ {R} ^ {\\text {o u t}} = \\mathcal {A} \\left(\\mathbf {x} _ {D} ^ {U}, \\mathbf {x} _ {D} ^ {3}, \\mathbf {x} _ {R} ^ {3}\\right) + \\mathcal {A} \\left(\\mathbf {x} _ {I} ^ {U}, \\mathbf {x} _ {I} ^ {3}, \\mathbf {x} _ {R} ^ {3}\\right) + \\mathbf {x} _ {R} ^ {3} + \\mathbf {x} _ {R} ^ {4}, (1) \\\\ \\mathbf {x} _ {D} ^ {o u t} = \\mathcal {A} \\left(\\mathbf {x} _ {R} ^ {U}, \\mathbf {x} _ {R} ^ {3}, \\mathbf {x} _ {D} ^ {3}\\right) + \\mathbf {x} _ {D} ^ {3} + \\mathbf {x} _ {D} ^ {4}, (2) \\\\ \\mathbf {x} _ {I} ^ {o u t} = \\mathcal {A} \\left(\\mathbf {x} _ {R} ^ {U}, \\mathbf {x} _ {R} ^ {3}, \\mathbf {x} _ {I} ^ {3}\\right) + \\mathbf {x} _ {I} ^ {3} + \\mathbf {x} _ {I} ^ {4}, (3) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.892, + 0.611 + ], + "angle": 0, + "content": "where for \\(m \\in \\{R, D, I\\}\\), \\(\\mathbf{x}_m^U \\in \\mathbb{R}^{B \\times L \\times 1}\\) are the estimated token uncertainty in the corresponding modality, and \\(\\mathbf{x}_m^3, \\mathbf{x}_m^4 \\in \\mathbb{R}^{B \\times L \\times C}\\) are the outputs of the last LN and MLP, respectively. The meaning of other vectors is illustrated in Fig. 3. To prevent the modalities that experienced significant domain shifts, i.e. depth and infrared, from misleading each other, we prohibit their direct interaction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Uncertainty Estimation Module. As discussed above, it is essential to recognize unreliable modal information for FAS. In deep learning, uncertainty reflects the level of confidence a model has in its predictions. When the model is less certain about a prediction, we consider that prediction to be unreliable [12]. When a model takes images with substantial noise caused by significant domain shifts as inputs, uncertainty estimation techniques [12] tend to assign elevated uncertainty values to regions more likely to be misclassified [32]. One well-known approach is the Bayesian Neural Network (BNN), widely applied in various vision tasks [16, 19]. BNN focuses on learning the posterior distribution of the model's weights rather than output a single value. Since the posterior distribution is always not tractable [12], approximate inference methods, such as Monte Carlo dropout (MCD) [7, 40, 53], are employed for estimation. In this work, we also use MCD to construct the UEM as a probabilistic representational model, which imposes a Bernoulli distribution over the weights of a model" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "213" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.253, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.089, + 0.412, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.089, + 0.5, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.089, + 0.737, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.089, + 0.871, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.895, + 0.314 + ], + "angle": 0, + "content": "Figure 3. (a1)-(a2) Illustration of fine-tuning ViT with proposed U-Adapters, showcasing the interaction between the RGB (R) and Depth (D) modalities. Note that only parameters of U-Adapters are trainable. (b) Uncertainty Estimation Module (UEM) used for recognizing unreliable tokens. (c) Detailed structure of U-Adapter, which adopts cross-modal fusion and suppresses the interference of unreliable tokens on other modalities. After fusion, discriminative central difference information is integrated for fine-grained spoof representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.321, + 0.47, + 0.444 + ], + "angle": 0, + "content": "without introducing extra parameters. As presented in Fig. 3 (b), we perform MCD instead of the vanilla dropout after each MHSA, simulating the effect of sampling from the posterior distribution by randomly dropping out neurons in MHSA. According to Bayesian probability theory, we consider \\(\\mathcal{T} = \\{\\mathbf{x}_R^{u_1},\\mathbf{x}_R^{u_2},\\dots ,\\mathbf{x}_R^{u_t}\\}\\) as empirical samples from the approximate distribution, and use their token-wise variance \\(\\mathbf{x}_R^U = \\mathrm{Var}(\\mathcal{T})\\in \\mathbb{R}^{B\\times L\\times 1}\\) to reflect their unreliability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.471, + 0.744 + ], + "angle": 0, + "content": "Uncertainty-Guided Cross-Adapter. Different from existing cross-modal fusion [33, 36], our U-Adapter, guided by the estimated uncertainty, suppresses the interaction with unreliable tokens from other modalities to ensure the reliability of the fused features. As shown in Fig. 3 (c), U-Adapter interactively fuses features from different modalities and suppresses unreliable information during fusion. For example, in the depth branch (see Fig. 3 (a2)), U-Adapter takes the output of the second LN from another modality \\(\\mathbf{x}_R^3\\) as the query for attention-based feature fusion. The output of the second LN from its own modality \\(\\mathbf{x}_D^3\\) is used as key and value. Moreover, U-Adapter modulates the query tokens from the other modality based on their uncertainty. Since malicious presentation attacks result in subtle traces, we adopt central difference convolution (CDC) [59] to integrate the local details for ViT and explore fine-grained difference information from neighbor tokens, thus making the fused features more discriminative. By ignoring GeLU and vanilla convolution layers, the U-Adapter is formulated as follows." + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.745, + 0.47, + 0.778 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} (\\mathbf {x} _ {R} ^ {U}, \\mathbf {x} _ {R} ^ {3}, \\mathbf {x} _ {D} ^ {3}) = \\mathcal {C} \\left(\\mathcal {S} \\left(\\frac {\\mathrm {Q} (\\mathbf {x} _ {R} ^ {3}) \\mathrm {K} (\\mathbf {x} _ {D} ^ {3}) ^ {\\intercal} \\odot \\mathrm {U} (\\mathbf {x} _ {R} ^ {U})}{\\sqrt {n _ {k}}}\\right) \\mathrm {V} (\\mathbf {x} _ {D} ^ {3}))\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.471, + 0.904 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}(\\cdot)\\) denotes the CDC layer, \\(S(\\cdot)\\) represents the softmax, and \\(n_k\\) is the number of token channels. \\(\\mathrm{Q}(\\cdot),\\mathrm{K}(\\cdot)\\) and \\(\\mathrm{V}(\\cdot)\\) are the linear projections corresponding to the query, key, and value, respectively. \\(\\odot\\) is the Hadamard product. \\(\\mathrm{U}(\\mathbf{x}_R^U) = \\exp (-r_e\\cdot \\mathbf{x}_R^U\\times \\mathbf{I})\\) normalizes the inputted uncertainty using an exponential function, where \\(r_e\\) is a penalty intensity parameter, and \\(\\mathbf{I} = [1,1,\\dots ,1]_{1\\times L}\\) is adopted to expand the dimensions of \\(x_R^U\\) to match the size" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.895, + 0.382 + ], + "angle": 0, + "content": "of \\( L \\times L \\). We show the meaning of other vectors in Figs. 3 (a1-a2). Note that Eq. (4) only provides the scenario where RGB tokens are fused into depth in Fig. 3 (c). The pairwise fusion among other modalities follows similarly." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.394, + 0.871, + 0.41 + ], + "angle": 0, + "content": "3.2. Rebalanced Modality Gradient Modulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Different modalities are suitable for distinct types of attacks and deployment conditions, e.g., the depth modality being less sensitive to light and more apt for detecting replay attacks, and the infrared modality excelling in identifying wearable mask attacks and performing well in low-light conditions. To cope with various unknown attacks and deployment conditions, we aim to ensure that each modality can sufficiently converge, thereby fully leveraging their discriminative capability, rather than over-relying on a specific modality (i.e. fast modality) that shows superior performance in the source domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Current modality balancing methods [11, 43], designed for simple late fusion models, are limited in their applicability to FAS. Late fusion networks have been shown to underperform in FAS compared to hybrid fusion networks [33, 36] that allow the fusion both before the final classifier and during the feature extraction. Additionally, the trainable parameters of a late fusion network are only affected by the gradient of its own modality, whereas those of a hybrid fusion network are influenced by the gradients of all modalities. Moreover, gradients from different modalities may even conflict with each other, making the modulation more difficult. To this end, our ReGrad adaptively modulates the gradients of all modalities that simultaneously act on each trainable layer. Moreover, existing modality balancing methods fail to generalize well while balancing each modality due to the ignorance of domain shifts. Drawing inspiration from SSDG [20] and PMR [11], we introduce a single-side prototypical loss to supervise modality branches and monitor their convergence speed. This loss enables each modality to cluster according to domain information through internal impetus, which further mitigates the domi" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "214" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.087, + 0.852, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.22, + 0.894, + 0.248 + ], + "angle": 0, + "content": "Figure 4. Illustration of gradient modulation via the proposed ReGrad in different scenarios: (Row 1) Non-conflicting (a1) and faster modality \\( j \\) (b1) or \\( i \\) (c1). (Row 2) Conflicting (a2) and faster modality \\( j \\) (b2) or \\( i \\) (c2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.254, + 0.273 + ], + "angle": 0, + "content": "nance of faster modalities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.47, + 0.457 + ], + "angle": 0, + "content": "Modulating Gradients for Each Modality. When the optimization direction is dominated by the gradients of the fast modality, the convergence degree of slow modalities is suppressed. The discriminative ability of slow modalities may even be weakened when there is a conflict in the gradient directions between modalities [55]. Inspired by the studies on solving domain conflicts [41] and multi-task conflicts [55] via gradient modulation, our ReGrad adaptively performs gradient modulation based on the conflict and convergence degree of each modality, making it applicable to complex hybrid fusion networks. The criteria of modality convergence speed will be discussed in the following section." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.471, + 0.749 + ], + "angle": 0, + "content": "As shown in Fig. 4 (b1), if modality \\(i\\) converges slower than \\(j\\) and their gradients \\(\\mathbf{g}_i\\) and \\(\\mathbf{g}_j\\) are not conflicting, we suppress \\(\\mathbf{g}_j\\) and boost the learning of modality \\(i\\). This involves decomposing \\(\\mathbf{g}_j\\) into orthogonal \\(\\mathbf{g}_j^v\\) and non-orthogonal \\(\\mathbf{g}_j^h\\) components relative to \\(\\mathbf{g}_i\\). We drop \\(\\mathbf{g}_j^v\\) to slow down modality \\(j\\) and use \\(\\mathbf{g}_j^h\\) to accelerate modality \\(i\\). Furthermore, to prevent unreliable modalities that converge rapidly from misleading other modalities, we employ the estimated uncertainty of modality \\(j\\) to further constrain \\(g_j\\). Conversely, if \\(\\mathbf{g}_i\\) conflicts with \\(\\mathbf{g}_j\\) and modality \\(i\\) remains a slow modal, it is unreasonable to use any part of \\(\\mathbf{g}_j\\) to boost the learning of \\(\\mathbf{g}_i\\). To prevent \\(\\mathbf{g}_j\\) from seriously slowing down the convergence speed of modality \\(i\\), as shown in Fig. 4 (b2), we remove the non-orthogonal part of \\(\\mathbf{g}_j^h\\) that is opposite to \\(\\mathbf{g}_i\\). Meanwhile, we retain its orthogonal part \\(\\mathbf{g}_j^v\\) to allow modality \\(j\\) to learn at a slower speed and reduce \\(\\mathbf{g}_j\\)'s interference with \\(\\mathbf{g}_i\\). Similarly, the uncertainty-based suppression is adopted. Cases where the modality \\(i\\) is faster are presented in Figs. 4 (c1)-(c2) and Eq. (5)." + }, + { + "type": "equation", + "bbox": [ + 0.082, + 0.753, + 0.468, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R e G r a d} _ {2} \\left(\\mathbf {g} _ {i}, \\mathbf {g} _ {j}\\right) = \\left\\{ \\begin{array}{l l} \\mathbf {g} _ {i} + \\frac {\\mathbf {g} _ {i} \\cdot \\mathbf {g} _ {j}}{\\| \\mathbf {g} _ {i} \\| ^ {2}} \\mathbf {g} _ {i} \\cdot \\mathrm {U} \\left(u _ {j}\\right) & , \\text {i n F i g . 4 (b 1)}, \\\\ \\mathbf {g} _ {i} + \\left(\\mathbf {g} _ {j} - \\frac {\\mathbf {g} _ {i} \\cdot \\mathbf {g} _ {j}}{\\| \\mathbf {g} _ {i} \\| ^ {2}} \\mathbf {g} _ {i}\\right) \\cdot \\mathrm {U} \\left(u _ {j}\\right) & , \\text {i n F i g . 4 (b 2)}, \\\\ \\frac {\\mathbf {g} _ {i} \\cdot \\mathbf {g} _ {j}}{\\| \\mathbf {g} _ {j} \\| ^ {2}} \\mathbf {g} _ {j} \\cdot \\mathrm {U} \\left(u _ {i}\\right) + \\mathbf {g} _ {j} & , \\text {i n F i g . 4 (c 1)}, \\\\ \\left(\\mathbf {g} _ {i} - \\frac {\\mathbf {g} _ {i} \\cdot \\mathbf {g} _ {j}}{\\| \\mathbf {g} _ {j} \\| ^ {2}} \\mathbf {g} _ {j}\\right) \\cdot \\mathrm {U} \\left(u _ {i}\\right) + \\mathbf {g} _ {j} & , \\text {i n F i g . 4 (c 2)}, \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\( u_{i} \\) and \\( u_{j} \\) are respectively the batch-wise mean uncertainty of the class tokens outputted by the last ViT block" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.26, + 0.892, + 0.289 + ], + "angle": 0, + "content": "of the corresponding modality, denoting the modality uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.893, + 0.549 + ], + "angle": 0, + "content": "Single-Side Prototypical Loss. Prototype learning has been successfully applied in FAS and modality rebalancing. PMR [11] indicates that introducing prototypes to each modality in multi-modal learning to guide the learning process can mitigate modality imbalance. However, vanilla prototype learning is suboptimal under DG scenarios as it neglects the differences in samples across various domains. Inspired by SSDG [20], we propose the Single-Side Prototypical (SSP) loss with the concept of asymmetry representation and prototype learning. Specifically, we introduce prototypes for the live faces of all domains and the fake faces from each domain. Formulate the domain set as \\(\\mathcal{D} = \\{d_{\\ell}, d_s^1, d_s^2, \\ldots, d_s^{N_s}\\}\\), where \\(N_s\\) is the number of source domains, \\(d_{\\ell}\\) denotes the domain consists of all live faces, and \\(d_s^i\\) represents domain of spoof faces from the \\(i\\)-th dataset. For modality \\(m\\), the prototype of domain \\(d \\in \\mathcal{D}\\) is the center of all samples belonging to \\(d\\) in the feature space:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.553, + 0.892, + 0.584 + ], + "angle": 0, + "content": "\\[\np _ {m} ^ {d} = \\frac {1}{N _ {d}} \\sum_ {c _ {m} \\in d} c _ {m}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.588, + 0.892, + 0.709 + ], + "angle": 0, + "content": "where \\( c_{m} \\) is a sample of \\( d \\) in modality \\( m \\), and \\( N_{d} \\) denotes the total number of samples in \\( d \\). Based on the prototypes of each domain, we aim to reduce the distance between samples and the prototype of their corresponding domain and increase the distance from other prototypes, thus preventing conflicts when samples from different domains are trained together. Therefore, following the form of cross-entropy, the SSP loss \\( \\mathcal{L}_{m}^{ssp} \\) of modality \\( m \\) is formulated as follows." + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.715, + 0.892, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m} ^ {s s p} \\left(c _ {m}, p _ {m} ^ {d _ {g t}}\\right) = - \\log \\frac {\\exp \\left(- \\mathrm {E D} \\left(c _ {m} , p _ {m} ^ {d _ {g t}}\\right)\\right)}{\\sum_ {d \\in \\mathcal {D}} \\exp \\left(- \\mathrm {E D} \\left(c _ {m} , p _ {m} ^ {d}\\right)\\right)}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.87 + ], + "angle": 0, + "content": "where sample \\(c_{m}\\) belongs to domain \\(d_{gt}\\). The function \\(\\mathrm{ED}(\\cdot)\\) denotes the Euclidean distance of the input sample and its prototype. Our SSP loss encourages the features of each modality of the samples to be pulled to the corresponding domain's prototype. This optimization of each modality branch is driven by the intrinsic force within each modality. We mark the modality with lower \\(\\mathcal{L}_m^{ssp}\\) as a faster modality. Our final loss is expressed as follows." + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.875, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\text {f i n a l}} = \\mathcal {L} ^ {c e} + \\sum_ {m \\in \\{R, D, I \\}} \\lambda \\cdot \\mathcal {L} _ {m} ^ {s s p}, \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "215" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.089, + 0.894, + 0.117 + ], + "angle": 0, + "content": "Table 1. Cross-dataset testing results under the fixed-modal scenarios (Protocol 1) among CASIA-CeFA (C), PADISI (P), CASIA-SURF (S), and WMCA (W). DG, MM, and FM are short for domain-generalized, multi-modal, and flexible-modal, respectively." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.117, + 0.891, + 0.241 + ], + "angle": 0, + "content": "
MethodTypeCPS → WCPW → SCSW → PPSW → CAverage
HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑
SSDG [20]DG26.0982.0328.5075.9141.8260.5640.4862.3134.2270.20
SSAN [50]DG17.7391.6927.9479.0434.4968.8536.4369.2929.1577.22
IADG [71]DG27.0286.5023.0483.1132.0673.8339.2463.6830.3476.78
ViTAF [18]DG20.5885.8229.1677.8030.7573.0339.7563.4430.0675.02
MM-CDCN [58]MM38.9265.3942.9359.7941.3861.5148.1453.7142.8460.10
CMFL [13]MM18.2288.8231.2075.6626.6880.8536.9366.8228.2678.04
ViT + AMA [61]FM17.5688.7427.5080.0021.1885.5147.4855.5628.4377.45
VP-FAS [60]FM16.2691.2224.4281.0721.7685.4639.3566.5525.4581.08
ViT [9]Baseline20.8884.7744.0557.9433.5871.8042.1556.4535.1667.74
MMDGOurs12.7993.8315.3292.8618.9588.6429.9376.5219.2587.96
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.251, + 0.47, + 0.342 + ], + "angle": 0, + "content": "where \\( L^{ce} \\) denotes the cross-entropy loss, and \\( \\lambda \\) is used to balance the trade-off between \\( \\mathcal{L}_m^{ssp} \\) and \\( \\mathcal{L}^{ce} \\). Here, we utilize \\( \\mathcal{L}^{ce} \\) to ensure the model's performance, while \\( \\mathcal{L}_m^{ssp} \\) is employed to balance all modalities and enhance generalizability. Since \\( \\mathcal{L}_m^{ssp} \\) is not modality-specific, it does not undergo gradient modulation, which we perform to \\( \\mathcal{L}^{ce} \\) only." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.367, + 0.313, + 0.384 + ], + "angle": 0, + "content": "4. Experimental Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.393, + 0.469, + 0.409 + ], + "angle": 0, + "content": "4.1. Multi-Modal FAS Benchmark in DG Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Datasets, Protocols, and Performance Metrics. With the increasing popularity of large-scale foundation models [24], the current trend is towards fully leveraging multiple and diverse source datasets/domains [1]. While training with multiple datasets has been widely studied for unimodal FAS [20, 50, 71], the multi-modal FAS community remains in the phase of single-dataset training [13, 33, 36]. Therefore, we propose the first large-scale multi-source benchmark to evaluate the DG performance of multi-modal FAS. This benchmark incorporates four commonly used multimodal datasets: CASIA-CeFA (C) [34], PADISI-Face (P) [45], CASIA-SURF (S) [67], and WMCA (W) [15], each featuring a variety of attack types. For each dataset, we use RGB, depth (D), and infrared (I) modalities as inputs. Following previous works [20], we use Half Total Error Rate (HTER) and Area Under the Receiver Operating Characteristic Curve (AUC) for evaluation. We design three protocols to evaluate the cross-dataset performance under different unseen deployment conditions, i.e., fixed modalities, missing modalities, and limited source domains. For Protocol 1, inspired by the leave-one-out (LOO) protocols designed in unimodal DG scenarios [20], we propose four multi-modal LOO sub-protocols among C, P, S, and W. For example, the sub-protocol CPS \\(\\rightarrow\\) W represents we take C, P, and S as training sets, while W is testing set. In Protocol 2, for each LOO sub-protocol of Protocol 1, we design three test-time missing-modal scenarios, i.e., D is missing, I is missing, and both D and I are missing. In Protocol 3, we limit the number of source domains by proposing two subprotocols, namely CW \\(\\rightarrow\\) PS and PS \\(\\rightarrow\\) CW. Implementation Details. We resize all RGB, depth, and infrared images to \\(224 \\times 224 \\times 3\\). Each modality's input is divided" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.895, + 0.403 + ], + "angle": 0, + "content": "into \\(14 \\times 14\\) patch tokens and a class token is attached, collectively serving as the input to the ViT. The hidden size of each token is default set to 768. We employ the Adam optimizer with a learning rate of \\(5 \\times 10^{-5}\\) and a weight decay of \\(1 \\times 10^{-3}\\) to train 70 epochs. The batch size is set to 32. Our classifier is a single-layer fully-connected network that reduces the dimensionality of the class token outputted by the final ViT block from 768 to 2. The number \\(n\\) of ViT blocks in MMDG is set to 12. ViT-b pretrained on ImageNet is adopted as the baseline." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.416, + 0.702, + 0.432 + ], + "angle": 0, + "content": "4.2. Cross-dataset Testing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.622 + ], + "angle": 0, + "content": "We compare our method on Protocols 1-3 with three categories of FAS methods: (1) Multi-modal methods, including CMFL [13] and MM-CDCN [58]. (2) Flexible-modal FAS, such as ViT + AMA [61] and VP-FAS [60]. (3) Domain generalized FAS i.e., SSDG [20], SSAN [50], IADG [71], and ViTAF [18]. For methods not designed for unimodal FAS, we re-implement them either by employing a concatenation of multi-modal inputs or by replacing the backbone with our ViT+U-Adapter or vanilla ViT. All these methods are trained on the proposed protocols. We automatically use their hyper-parameters as described in the corresponding reference papers or optimally assign better ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Protocol 1: Fixed Modalities. We evaluated the cross-dataset performance of unimodal DG and multi-modal methods in the fixed modalities scenarios. Since unimodal DG methods cannot be directly applied to this protocol, we concatenate the images of the three modalities along the channel dimension. We then reduce the channel number to 3 using a trainable \\(1 \\times 1\\) convolution layer before feeding them to these methods. As shown in the Table 1, our MMDG achieves the best results. Compared to the second-ranked method, i.e. VP-FAS [60], there is an improvement of over \\(6\\%\\) in both HTER and AUC. A higher performance boost occurs when compared to domain-generalized methods. This indicates MMDG's effectiveness in addressing modality unreliability and imbalance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Protocol 2: Missing Modalities. As shown in Table 2, our MMDG outperforms other algorithms in all three scenarios of modality missing during testing. It's worth noting that ViT+AMA and VP-FAS are specifically designed" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "216" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.131 + ], + "angle": 0, + "content": "Table 2. Cross-dataset testing results under the test-time missing-modal scenarios (Protocol 2) among CASIA-CeFA (C), PADISI (P), CASIA-SURF (S), and WMCA (W). We report the average HTER (\\%) ↓ and AUC (\\%) ↑ on four sub-protocols, i.e., CPS → W, CPW → S, CSW → P, and PSW → C. DG, MM, and FM are short for domain-generalized, multi-modal, and flexible-modal, respectively." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.132, + 0.89, + 0.256 + ], + "angle": 0, + "content": "
MethodTypeMissing DMissing IMissing D & IAverage
HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑
SSDG [20]DG38.9265.4537.6466.5739.1865.2238.5865.75
SSAN [50]DG36.7769.2141.2061.9233.5273.3837.1668.17
IADG [71]DG40.7258.7242.1761.8337.5066.9040.1362.49
ViTAF [18]DG34.9973.2235.8869.4035.8969.6135.5970.64
MM-CDCN [58]MM44.9055.3543.6058.3844.5455.0844.3556.27
CMFL [13]MM31.3774.6230.5575.4231.8974.2931.2774.78
ViT + AMA [61]FM29.2577.7032.3074.0631.4875.8231.0175.86
VP-FAS [60]FM29.1378.2729.6377.5130.4776.3129.7477.36
ViTBaseline40.0464.6936.7768.1936.2069.0237.6767.30
MMDGOurs24.8982.3923.3983.8225.2681.8624.5182.69
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.268, + 0.47, + 0.308 + ], + "angle": 0, + "content": "Table 3. Cross-dataset testing results under the limited source domain scenarios (Protocol 3) among CeFA-CeFA (C), PADISI-UGC (P), CASIA-SURF (S), and WMCA (W)." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.309, + 0.468, + 0.424 + ], + "angle": 0, + "content": "
MethodTypeCW → PSPS→ CW
HTER (%)AUC (%)HTER (%)AUC (%)
SSDG [20]DG25.3480.1746.9854.29
SSAN [50]DG26.5580.0639.1067.19
IADG [71]DG22.8283.8539.7063.46
ViTAF [18]DG29.6477.3639.9361.31
MM-CDCN [58]MM29.2876.8847.0051.94
CMFL [13]MM31.8672.7539.4363.17
ViT + AMA [61]FM29.2576.8938.0667.64
VP-FAS [60]FM25.9081.7944.3760.83
ViT [9]Baseline42.6657.8042.7560.41
MMDGOurs20.1288.2436.6070.35
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.634 + ], + "angle": 0, + "content": "for scenarios of modality missing and have achieved better results compared to other methods. However, under the DG premise, they did not consider the modality imbalance and unreliability issues brought by domain shift, resulting in a difference of over \\(5\\%\\) in average HTER and AUC compared to our method. Another interesting observation is that SSAN, IADG, ViT, and ViT + AMA perform better in the Missing D & I scenario than in Missing D and Missing I. These methods lose more modalities but achieve performance improvement, indicating that unreliable information within the modalities leads to erroneous detection. This reflects the effectiveness of our approach in suppressing the interaction of unreliable information between modalities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.47, + 0.759 + ], + "angle": 0, + "content": "Protocol 3: Limited Source Domains. Compared to protocol 1, the limitation of source domains results in a more significant domain shift. We observe that our proposed MMDG achieves optimal results in both sub-protocols (see Table 3), even in the presence of substantial domain shift in \\(\\mathbf{PS} \\rightarrow \\mathbf{CW}\\). This further underscores the effectiveness of our U-Adapter and ReGrad in enhancing domain generalizability in unseen multi-modal deployment environments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.769, + 0.35, + 0.785 + ], + "angle": 0, + "content": "4.3. Ablation Study and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.468, + 0.822 + ], + "angle": 0, + "content": "In this section, we perform ablation analysis on U-Adapter and ReGrad to verify their individual contribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Effectiveness of U-Adapter. To validate the effectiveness of the proposed U-Adapter, we compare the network performance before and after removing the adapter. We also replace the backbone of existing DG methods (i.e., SSDG, SSAN, and IADG) and multi-modal methods (CMFL) to" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.268, + 0.687, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.268, + 0.889, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.377, + 0.892, + 0.404 + ], + "angle": 0, + "content": "Figure 5. Ablation results on our U-Adapter. We report average HTER \\(\\downarrow\\) and AUC \\(\\uparrow\\) on four sub-protocols in Protocol 1." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.416, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Table 4. Ablation results on the proposed ReGrad. We report the average HTER and AUC on four sub-protocols in Protocol 1." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.442, + 0.89, + 0.569 + ], + "angle": 0, + "content": "
BackboneMethodHTER (%)AUC (%)
ViT [9]-31.1474.81
ViT [9]ReGrad30.5975.94
ViT + U-AdapterSSDG [20]26.1380.31
ViT + U-AdapterSSAN [50]26.9479.52
ViT + U-AdapterIADG [71]24.2382.54
ViT + U-AdapterPMR [11]24.5483.14
ViT + U-AdapterOGM-GE [43]29.3276.54
ViT + U-Adapter-27.3878.88
ViT + U-AdapterReGrad (Conflicted-only)21.9185.65
ViT + U-AdapterReGrad (Unconflicted-only)21.2385.51
ViT + U-AdapterReGrad19.4887.48
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.717 + ], + "angle": 0, + "content": "observe the performance differences when using ViT+U-Adapter as the backbone versus using ViT alone. As shown in Fig. 5, compared to using ViT as the backbone, these methods experience a performance boost when additional U-Adapters are employed. The performance improvement is especially pronounced when no additional method is used, and when our proposed ReGrad is integrated. This demonstrates the effectiveness of U-Adapter and its universality for existing DG methods and multi-modal strategies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of ReGrad. Here, we try to answer two questions: (1) Whether the idea to address domain shift through modality rebalancing, is more suitable compared to existing DG methods? (2) Is our ReGrad more capable of enhancing the performance of complex cross-modal fusion frameworks like ViT+U-Adapter compared to existing modality balancing methods? From Table 4, we can observe that when ViT+U-Adapter is used as the backbone, our ReGrad surpasses existing DG methods and other modality rebalancing methods. Even for vanilla ViT, our ReGrad can enhance its performance. This indicates the effectiveness of addressing domain shifts in unseen scenarios by balancing" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "217" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.273, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.224, + 0.237, + 0.235 + ], + "angle": 0, + "content": "(a) w/o ReGrad" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.089, + 0.468, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.224, + 0.426, + 0.235 + ], + "angle": 0, + "content": "(b) w/ ReGrad" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.236, + 0.47, + 0.28 + ], + "angle": 0, + "content": "Figure 6. Modality balance degree during training. (a) MMDG w/o ReGrad. (b) MMDG w/ ReGrad. The red dashed line in each sub-figure represents the variance of \\(\\mathcal{L}_R^{ssp}\\), \\(\\mathcal{L}_D^{ssp}\\), and \\(\\mathcal{L}_I^{ssp}\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.29, + 0.47, + 0.332 + ], + "angle": 0, + "content": "Table 5. Impact of estimated uncertainty in U-Adapters and ReGrad. We report the average HTER \\((\\%)\\) ↓ and AUC \\((\\%)\\) ↑ on four sub-protocols in Protocol 1." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.332, + 0.468, + 0.411 + ], + "angle": 0, + "content": "
Using Estimated UncertaintyMetric Value
U-Adapter: Eq. (4)ReGrad: Eq. (5)HTER (%) ↓AUC (%) ↑
××28.4078.65
×✓26.1381.60
✓×22.7584.84
✓✓19.4887.48
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.47, + 0.505 + ], + "angle": 0, + "content": "various modalities. The results in Table 4 also indicate that when compared to the SoTA modality balancing methods (i.e., OGM-GE and PMR), our ReGrad is more suitable for ViT + U-Adapter, which involves hybrid feature fusion and complex cross-modal interactions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.732 + ], + "angle": 0, + "content": "Meanwhile, as illustrated in Fig. 6, the lower variance (right) of \\(\\mathcal{L}_R^{ssp}\\), \\(\\mathcal{L}_D^{ssp}\\), and \\(\\mathcal{L}_I^{ssp}\\) demonstrates that ReGrad obviously improves the balance degree of MMDG. We also notice that ReGrad causes slight fluctuations in the loss function values. This is due to the gradient modulation allowing the model to escape from local optima, seeking a direction with a higher degree of modality balance. Despite the fluctuations, MMDG can still converge after about 50 epochs. Additionally, we compare ReGrad with its two variants, which perform modulation only when modality gradients are conflicted and unconflicted, respectively. Although these two variants improve performance compared to not modulating gradients at all, their performance is inferior to modulating gradients in all situations. This validates the necessity of adaptive modulation under various scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Impact of the Uncertainty in U-Adapter and ReGrad. To testify whether the estimated uncertainty can enhance detection performance, we made a comparison by removing the parts that use uncertainty for constraints in U-Adapter and ReGrad. From Table 5, it can be observed that adding uncertainty to constrain the propagation of unreliable information in U-Adapter and the gradient of fast modalities in ReGrad both leads to performance improvement. When uncertainty is added to both modules, HTER and AUC can be further improved by over \\(3\\%\\). As shown in Fig. 7, MMDG can identify unreliable regions in depth (left)" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.089, + 0.88, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.223, + 0.892, + 0.265 + ], + "angle": 0, + "content": "Figure 7. Visual attention and uncertainty maps of our MMDG. Uncertainly predicted regions are highlighted with red boxes. The redder color indicates the higher uncertainty or more attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.284, + 0.892, + 0.36 + ], + "angle": 0, + "content": "and infrared (right) modalities. Although the U-Adapters of these modalities focus on regions with uncertain predictions, due to our idea of suppressing uncertain regions, these regions fail to mislead mutual modalities, enabling the capture of more reliable spoof traces against domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.892, + 0.514 + ], + "angle": 0, + "content": "Discussion. The results in Tables 1-5 and Figs. 5-6 validate the MMDG's generalizability by addressing the problems of modality unreliability and imbalance. This work marks a stride towards practical deployments and can benefit the FAS community. However, our MMDG's performance can still be improved. Continued efforts on this benchmark are necessary for further progress. Meanwhile, our method also has space for improvement, e.g., further enhancing the degree of modality balance and adopting more efficient uncertainty estimation techniques." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.527, + 0.619, + 0.543 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.718 + ], + "angle": 0, + "content": "We propose the first multi-modal DG framework for FAS, i.e. MMDG, to enhance generalizability by addressing modality unreliability and imbalance. To mitigate the interference of unreliable modalities, we propose U-Adapter to suppress uncertainly predicted regions during cross-modal feature fusion. Besides, we develop ReGrad to adaptively modulate the gradients of each modality branch to achieve modality balance, thus preventing an over-reliance on a dominant modality. We also establish the first large-scale benchmark for multi-modal FAS under DG scenarios. Extensive experiments demonstrate MMDG's effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This research was done at the Rapid-Rich Object Search (ROSE) Lab, Nanyang Technological University, and supported by the National Key Research and Development Program of China (Grant No.2022YFB3207700), the National Natural Science Foundation of China (Grants No.62272022, 62306061, and 62331006), the NTU-PKU Joint Research Institute (sponsored by the Ng Teng Fong Charitable Foundation), the Science and Technology Foundation of Guangzhou Huangpu Development District (Grant No.2022GH15), and the Guangdong Provincial Regional Joint Fund - Regional Cultivation Project (Grant No.2023A1515140037)" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "218" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.091, + 0.173, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.268 + ], + "angle": 0, + "content": "[1] Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George van den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, Diego de Las Casas, Aurelia Guy, Jacob Menick, Roman Ring, Tom Hennigan, Saffron Huang, Loren Maggiore, Chris Jones, Albin Cassirer, Andy Brock, Michela Paganini, Geoffrey Irving, Oriol Vinyals, Simon Osindero, Karen Simonyan, Jack W. Rae, Erich Elsen, and Laurent Sifre. Improving language models by retrieving from trillions of tokens. In Proceedings of the International Conference on Machine Learning, pages 2206-2240, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.47, + 0.312 + ], + "angle": 0, + "content": "[2] Rizhao Cai and Changsheng Chen. Learning deep forest with multi-scale local binary pattern features for face antispoofing. arXiv, abs/1910.03850, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.314, + 0.469, + 0.382 + ], + "angle": 0, + "content": "[3] Rizhao Cai, Haoliang Li, Shiqi Wang, Changsheng Chen, and Alex C. Kot. Drl-fas: A novel framework based on deep reinforcement learning for face anti-spoofing. IEEE Transactions on Information Forensics and Security, 16:937-951, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[4] Rizhao Cai, Zhi Li, Renjie Wan, Haoliang Li, Yongjian Hu, and Alex C. Kot. Learning meta pattern for face antispoofing. IEEE Transactions on Information Forensics and Security, 17:1201-1213, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.443, + 0.469, + 0.513 + ], + "angle": 0, + "content": "[5] Rizhao Cai, Yawen Cui, Zhi Li, Zitong Yu, Haoliang Li, Yongjian Hu, and Alex C. Kot. Rehearsal-free domain continual face anti-spoofing: Generalize more and forget less. In Proceedings of the IEEE International Conference on Computer Vision, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.469, + 0.57 + ], + "angle": 0, + "content": "[6] Rizhao Cai, Zitong Yu, Chenqi Kong, Haoliang Li, Changsheng Chen, Yongjian Hu, and Alex C. Kot. S-adapter: Generalizing vision transformer for face anti-spoofing with statistical tokens. arXiv, abs/2309.04038, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.573, + 0.469, + 0.614 + ], + "angle": 0, + "content": "[7] Pamela Carreno, Dana Kulic, and Michael Burke. Adapting neural models with sequential monte carlo dropout. In Conference on Robot Learning, pages 1542-1552, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.617, + 0.469, + 0.671 + ], + "angle": 0, + "content": "[8] Pengchao Deng, Chenyang Ge, Xin Qiao, Hao Wei, and Yuan Sun. Attention-aware dual-stream network for multimodal face anti-spoofing. IEEE Transactions on Information Forensics and Security, 18:4258-4271, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.675, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[10] Zhekai Du, Jingjing Li, Lin Zuo, Lei Zhu, and Ke Lu. Energy-based domain generalization for face anti-spoofing. In Proceedings of the ACM International Conference on Multimedia, pages 1749-1757, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[11] Yunfeng Fan, Wenchao Xu, Haozhao Wang, Junxiao Wang, and Song Guo. PMR: prototypical modal rebalance for multimodal learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 20029-20038, 2023." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[12] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the International Conference on Machine Learning, pages 1050-1059, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[13] Anjith George and Sébastien Marcel. Cross modal focal loss for RGBD face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7882-7891, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[14] Anjith George and Sébastien Marcel. Learning one class representations for face presentation attack detection using multi-channel convolutional neural networks. IEEE Transactions on Information Forensics and Security, 16:361-375, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[15] Anjith George, Zohreh Mostaani, David Geissenbuhler, Olegs Nikisins, André Anjos, and Sébastien Marcel. Biometric face presentation attack detection with multi-channel convolutional neural network. IEEE Transactions on Information Forensics and Security, 15:42-55, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[16] Hongji Guo, Hanjing Wang, and Qiang Ji. Uncertainty-guided probabilistic transformer for complex action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[17] Shuangpeng Han, Rizhao Cai, Yawen Cui, Zitong Yu, Yongjian Hu, and Alex C. Kot. Hyperbolic face anti-spoofing. arXiv, abs/2308.09107, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[18] Hsin-Ping Huang, Deqing Sun, Yaojie Liu, Wen-Sheng Chu, Taihong Xiao, Jinwei Yuan, Hartwig Adam, and Ming-Hsuan Yang. Adaptive transformers for robust few-shot cross-domain face anti-spoofing. In Proceedings of the European Conference on Computer Vision, pages 37-54, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[19] Kaixiang Ji, Feng Chen, Xin Guo, Yadong Xu, Jian Wang, and Jingdong Chen. Uncertainty-guided learning for improving image manipulation detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 22456-22465, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[20] Yunpei Jia, Jie Zhang, Shiguang Shan, and Xilin Chen. Single-side domain generalization for face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8481-8490, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[21] Yunpei Jia, Jie Zhang, and Shiguang Shan. Dual-branch meta-learning network with distribution alignment for face anti-spoofing. IEEE Transactions on Information Forensics and Security, 17:138-151, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[22] Fangling Jiang, Qi Li, Pengcheng Liu, Xiang-Dong Zhou, and Zhenan Sun. Adversarial learning domain-invariant conditional features for robust face anti-spoofing. International Journal of Computer Vision, 131(7):1680-1703, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[23] Alex Kendall, Vijay Badrinarayanan, and Roberto Cipolla. Bayesian segnet: Model uncertainty in deep convolutional encoder-decoder architectures for scene understanding. In Proceedings of the British Machine Vision Conference, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[24] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In Proceedings of the IEEE International Conference on Computer Vision, pages 4015-4026, 2023." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "219" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[25] Chenqi Kong, Kexin Zheng, Shiqi Wang, Anderson Rocha, and Haoliang Li. Beyond the pixel world: a novel acoustic-based face anti-spoofing system for smartphones. IEEE Transactions on Information Forensics and Security, 17: 3238-3253, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[26] Chenqi Kong, Kexin Zheng, Yibing Liu, Shiqi Wang, Anderson Rocha, and Haoliang Li. M3fas: An accurate and robust multimodal mobile face anti-spoofing system. arXiv preprint arXiv:2301.12831, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[27] Kaicheng Li, Hongyu Yang, Binghui Chen, Pengyu Li, Biao Wang, and Di Huang. Learning polysemantic spoof trace: A multi-modal disentanglement network for face anti-spoofing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1351-1359, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.348 + ], + "angle": 0, + "content": "[28] Zhi Li, Rizhao Cai, Haoliang Li, Kwok-Yan Lam, Yongjian Hu, and Alex C. Kot. One-class knowledge distillation for face presentation attack detection. IEEE Transactions on Information Forensics and Security, 17:2137-2150, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.47, + 0.403 + ], + "angle": 0, + "content": "[29] Zhi Li, Haoliang Li, Xin Luo, Yongjian Hu, Kwok-Yan Lam, and Alex C. Kot. Asymmetric modality translation for face presentation attack detection. IEEE Transactions on Multimedia, 25:62-76, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.47, + 0.475 + ], + "angle": 0, + "content": "[30] Chen-Hao Liao, Wen-Cheng Chen, Hsuan-Tung Liu, Yi-Ren Yeh, Min-Chun Hu, and Chu-Song Chen. Domain invariant vision transformer learning for face anti-spoofing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6087-6096, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.47, + 0.532 + ], + "angle": 0, + "content": "[31] Xun Lin, Shuai Wang, Jiahao Deng, Ying Fu, Xiao Bai, Xinlei Chen, Xiaolei Qu, and Wenzhong Tang. Image manipulation detection by multiple tampering traces and edge artifact enhancement. Pattern Recognition, 133:109026, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.603 + ], + "angle": 0, + "content": "[32] Mattia Litrico, Alessio Del Bue, and Pietro Morerio. Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7640-7650, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.472, + 0.659 + ], + "angle": 0, + "content": "[33] Ajian Liu and Yanyan Liang. Ma-vit: Modality-agnostic vision transformers for face anti-spoofing. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 1180–1186, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.472, + 0.731 + ], + "angle": 0, + "content": "[34] Ajian Liu, Zichang Tan, Jun Wan, Sergio Escalera, Guodong Guo, and Stan Z. Li. CASIA-SURF cefa: A benchmark for multi-modal cross-ethnicity face anti-spoofing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1178–1186, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[35] Ajian Liu, Zichang Tan, Jun Wan, Yanyan Liang, Zhen Lei, Guodong Guo, and Stan Z. Li. Face anti-spoofing via adversarial cross-modality translation. IEEE Transactions on Information Forensics and Security, 16:2759-2772, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[36] Ajian Liu, Zichang Tan, Zitong Yu, Chenxu Zhao, Jun Wan, Yanyan Liang, Zhen Lei, Du Zhang, Stan Z. Li, and Guodong Guo. FM-ViT: Flexible modal vision transformers for face anti-spoofing. IEEE Transactions on Information Forensics and Security, 18:4775-4786, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[37] Yaojie Liu and Xiaoming Liu. Spoof trace disentanglement for generic face anti-spoofing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3813-3830, 2023." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[38] Yuchen Liu, Yabo Chen, Mengran Gou, Chun-Ting Huang, Yaoming Wang, Wenrui Dai, and Hongkai Xiong. Towards unsupervised domain generalization for face anti-spoofing. In Proceedings of the IEEE International Conference on Computer Vision, pages 20654-20664, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.219 + ], + "angle": 0, + "content": "[39] Anwei Luo, Rizhao Cai, Chenqi Kong, Xiangui Kang, Jiwu Huang, and Alex C. Kot. Forgery-aware adaptive vision transformer for face forgery detection. arXiv, abs/2309.11092, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.893, + 0.262 + ], + "angle": 0, + "content": "[40] Yuki Mae, Wataru Kumagai, and Takafumi Kanamori. Uncertainty propagation for dropout-based bayesian neural networks. Neural Networks, 144:394-406, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.893, + 0.32 + ], + "angle": 0, + "content": "[41] Lucas Mansilla, Rodrigo Echeveste, Diego H. Milone, and Enzo Ferrante. Domain generalization via gradient surgery. In Proceedings of the IEEE International Conference on Computer Vision, pages 6610-6618, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.893, + 0.377 + ], + "angle": 0, + "content": "[42] Olegs Nikisins, Anjith George, and Sébastien Marcel. Domain adaptation in multi-channel autoencoder based features for robust face anti-spoofing. In Proceedings of the International Conference on Biometrics, pages 1-8, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.893, + 0.447 + ], + "angle": 0, + "content": "[43] Xiaokang Peng, Yake Wei, Andong Deng, Dong Wang, and Di Hu. Balanced multimodal learning via on-the-fly gradient modulation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8228-8237, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.893, + 0.504 + ], + "angle": 0, + "content": "[44] Yunxiao Qin, Zitong Yu, Longbin Yan, Zezheng Wang, Chenxu Zhao, and Zhen Lei. Meta-teacher for face antispoofing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6311-6326, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.893, + 0.575 + ], + "angle": 0, + "content": "[45] Mohammad Rostami, Leonidas Spinoulas, Mohamed E. Hussein, Joe Mathai, and Wael Abd-Almageed. Detection and continual learning of novel face presentation attacks. In Proceedings of the IEEE International Conference on Computer Vision, pages 14831-14840, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.893, + 0.646 + ], + "angle": 0, + "content": "[46] Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C. Yuen. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10023-10031, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.893, + 0.703 + ], + "angle": 0, + "content": "[47] Tao Shen, Yuyu Huang, and Zhijun Tong. Facebagnet: Bag-of-local-features model for multi-modal face anti-spoofing. In IEEE Conference on Computer Vision and Pattern Recognition Workshop, pages 1611–1616, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.893, + 0.761 + ], + "angle": 0, + "content": "[48] Koushik Srivatsan, Muzammal Naseer, and Karthik Nandakumar. Flip: Cross-domain face anti-spoofing with language guidance. In Proceedings of the IEEE International Conference on Computer Vision, pages 19685-19696, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.893, + 0.83 + ], + "angle": 0, + "content": "[49] Yiyou Sun, Yaojie Liu, Xiaoming Liu, Yixuan Li, and WenSheng Chu. Rethinking domain generalization for face antispoofing: Separability and alignment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 24563-24574, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[50] Zhuo Wang, Zezheng Wang, Zitong Yu, Weihong Deng, Jiahong Li, Tingting Gao, and Zhongyuan Wang. Domain generalization via shuffled style assembly for face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4113-4123, 2022." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "220" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[51] Junru Wu, Xiang Yu, Buyu Liu, Zhangyang Wang, and Manmohan Chandraker. Uncertainty-aware physically-guided proxy tasks for unseen domain face anti-spoofing. arXiv, abs/2011.14054, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.472, + 0.215 + ], + "angle": 0, + "content": "[52] Wuyuan Xie, Kaimin Wang, Yakun Ju, and Miaohui Wang. pmbba: Projection-based blind point cloud quality assessment via multimodal learning. In Proceedings of the ACM International Conference on Multimedia, page 3250-3258, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.287 + ], + "angle": 0, + "content": "[53] Lan Xixi, Changchun Zou, Cheng Peng, and Caowei Wu. Uncertainty quantification in intelligent-based electrical resistivity tomography image reconstruction with monte carlo dropout strategy. IEEE Transactions on Geoscience and Remote Sensing, 61:1-16, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[54] Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.47, + 0.385 + ], + "angle": 0, + "content": "[55] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In Proceedings of the Neural Information Processing Systems, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[56] Yi Yu, Wenhan Yang, Yap-Peng Tan, and Alex C Kot. Towards robust rain removal against adversarial attacks: A comprehensive benchmark analysis and beyond. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6013-6022, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.47, + 0.524 + ], + "angle": 0, + "content": "[57] Yi Yu, Yufei Wang, Wenhan Yang, Shijian Lu, Yap-Peng Tan, and Alex C Kot. Backdoor attacks against deep image compression via adaptive frequency trigger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12250-12259, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.47, + 0.594 + ], + "angle": 0, + "content": "[58] Zitong Yu, Yunxiao Qin, Xiaobai Li, Zezheng Wang, Chenxu Zhao, Zhen Lei, and Guoying Zhao. Multi-modal face anti-spoofing based on central difference networks. In IEEE Conference on Computer Vision and Pattern Recognition Workshop, pages 2766-2774, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[59] Zitong Yu, Chenxu Zhao, Zezheng Wang, Yunxiao Qin, Zhuo Su, Xiaobai Li, Feng Zhou, and Guoying Zhao. Searching central difference convolutional networks for face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5294-5304, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[60] Zitong Yu, Rizhao Cai, Yawen Cui, Ajian Liu, and Changsheng Chen. Visual prompt flexible-modal face anti-spoofing. arXiv, abs/2307.13958, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[61] Zitong Yu, Rizhao Cai, Yawen Cui, Xin Liu, Yongjian Hu, and Alex C. Kot. Rethinking vision transformer and masked autoencoder in multimodal face anti-spoofing. arXiv, abs/2302.05744, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[62] Zitong Yu, Ajian Liu, Chenxu Zhao, Kevin H. M. Cheng, Xu Cheng, and Guoying Zhao. Flexible-modal face antispoofing: A benchmark. In IEEE Conference on Computer Vision and Pattern Recognition Workshop, pages 6346-6351, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[63] Zitong Yu, Yunxiao Qin, Xiaobai Li, Chenxu Zhao, Zhen Lei, and Guoying Zhao. Deep learning for face anti-spoofing: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5609-5631, 2023." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.16 + ], + "angle": 0, + "content": "[64] Zitong Yu, Rizhao Cai, Zhi Li, Wenhan Yang, Jingang Shi, and Alex C. Kot. Benchmarking joint face spoofing and forgery detection with visual and physiological cues. IEEE Transactions on Dependable and Secure Computing, pages 1-15, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.893, + 0.232 + ], + "angle": 0, + "content": "[65] Haixiao Yue, Keyao Wang, Guosheng Zhang, Haocheng Feng, Junyu Han, Errui Ding, and Jingdong Wang. Cyclically disentangled feature translation for face anti-spoofing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3358-3366, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.893, + 0.302 + ], + "angle": 0, + "content": "[66] Qingyang Zhang, Haitao Wu, Changqing Zhang, Qinghua Hu, Huazhu Fu, Joey Tianyi Zhou, and Xi Peng. Provable dynamic fusion for low-quality multimodal data. In Proceedings of the International Conference on Machine Learning, pages 41753-41769, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.304, + 0.893, + 0.373 + ], + "angle": 0, + "content": "[67] Shifeng Zhang, Ajian Liu, Jun Wan, Yanyan Liang, Guodong Guo, Sergio Escalera, Hugo Jair Escalante, and Stan Z. Li. CASIA-SURF: A large-scale multi-modal benchmark for face anti-spoofing. IEEE transactions on biometrics, behavior, and identity science, 2(2):182-193, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.374, + 0.893, + 0.43 + ], + "angle": 0, + "content": "[68] Taoying Zhang, Hesong Li, Qiankun Liu, Xiaoyong Wang, and Ying Fu. Mgt: Modality-guided transformer for infrared and visible image fusion. In Chinese Conference on Pattern Recognition and Computer Vision, pages 321-332, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.431, + 0.893, + 0.5 + ], + "angle": 0, + "content": "[69] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Ran Yi, Shouhong Ding, and Lizhuang Ma. Adaptive mixture of experts learning for generalizable face anti-spoofing. In Proceedings of the ACM International Conference on Multimedia, pages 6009-6018, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.501, + 0.893, + 0.569 + ], + "angle": 0, + "content": "[70] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Ran Yi, Kekai Sheng, Shouhong Ding, and Lizhuang Ma. Generative domain adaptation for face anti-spoofing. In Proceedings of the European Conference on Computer Vision, pages 335-356, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.571, + 0.893, + 0.641 + ], + "angle": 0, + "content": "[71] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Xuequan Lu, Ran Yi, Shouhong Ding, and Lizhuang Ma. Instance-aware domain generalization for face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 20453-20463, 2023." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "221" + } + ] +] \ No newline at end of file diff --git a/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/91e475e8-d317-465c-893c-024b420e20a0_origin.pdf b/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/91e475e8-d317-465c-893c-024b420e20a0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..82c54f4d910c6536b76f65dc46234fcaf2d0df2c --- /dev/null +++ b/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/91e475e8-d317-465c-893c-024b420e20a0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8aae172c5030a0133980748c4c29151ede0a9f74f15506262712ecd770a0b36 +size 5159159 diff --git a/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/full.md b/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f1fd677c49e31995bf0e5f7ebdb041f6ff831007 --- /dev/null +++ b/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/full.md @@ -0,0 +1,291 @@ +# Suppress and Rebalance: Towards Generalized Multi-Modal Face Anti-Spoofing + +Xun Lin $^{1}$ , Shuai Wang $^{1}$ , Rizhao Cai $^{2}$ , Yizhong Liu $^{1}$ , Ying Fu $^{3}$ , Wenzhong Tang $^{1}$ , Zitong Yu $^{4*}$ , Alex Kot $^{2}$ $^{1}$ Beihang University $^{2}$ Nanyang Technological University $^{3}$ Beijing Institute of Technology $^{4}$ Great Bay University + +{linxun, wangshuai}@buaa.edu.cn yuzitong@gbu.edu.cn + +# Abstract + +Face Anti-Spoofing (FAS) is crucial for securing face recognition systems against presentation attacks. With advancements in sensor manufacture and multi-modal learning techniques, many multi-modal FAS approaches have emerged. However, they face challenges in generalizing to unseen attacks and deployment conditions. These challenges arise from (1) modality unreliability, where some modality sensors like depth and infrared undergo significant domain shifts in varying environments, leading to the spread of unreliable information during cross-modal feature fusion, and (2) modality imbalance, where training overly relies on a dominant modality hinders the convergence of others, reducing effectiveness against attack types that are indistinguishable by sorely using the dominant modality. To address modality unreliability, we propose the Uncertainty-Guided Cross-Adapter (U-Adapter) to recognize unreliable detected regions within each modality and suppress the impact of unreliable regions on other modalities. For modality imbalance, we propose a Rebalanced Modality Gradient Modulation (ReGrad) strategy to rebalance the convergence speed of all modalities by adaptively adjusting their gradients. Besides, we provide the first large-scale benchmark for evaluating multi-modal FAS performance under domain generalization scenarios. Extensive experiments demonstrate that our method outperforms state-of-the-art methods. Source codes and protocols are released on https://github.com/OMGGGGG/mmdg. + +# 1. Introduction + +The remarkable success of deep neural networks has also brought concerns regarding their security vulnerabilities [56, 57, 64]. Especially face recognition (FR) systems, widely used in applications like surveillance and mobile payment [63], are susceptible to various face presentation attacks, including printed photos, video replays, and 3D + +![](images/2cb3603aa2d2e6fa2d850ae99159b9478ae02cbb2ef8f0081681452b21871964.jpg) +Figure 1. Illustration of DG scenarios in the context of (a) unimodal and (b) multi-modal. (c) DG performance of SSDG [20] on our Protocol 1 (see Sec. 4.1). Though fed with more modalities, SSDG performs worse in multi-modal scenarios compared to unimodal ones. * denotes our re-implemented multi-modal version. + +
(c)
MethodModalityHTER (%) ↓AUC (%) ↑
SSDG(a): RGB24.0783.69
SSDG*(b): RGB+D+I26.09 (+2.02)82.03 (-1.66)
+ +wearable masks [3]. These vulnerabilities challenge the security of FR systems, limiting their broader application. As a result, face anti-spoofing (FAS) methods are developed to enhance the security of FR systems. + +Recently, deep learning has been widely applied in FAS, particularly in unimodal scenarios based on the RGB spectrum. While these models perform well in intra-dataset evaluations, their generalizability to unseen attacks and deployment environments remains limited [2]. Despite the improvement of robustness via domain generalization (DG) techniques [20, 28, 49, 50, 71], these approaches still tend to over-fit the attacks presented in source domains. + +With the recent progress of multi-modal learning [31, 52, 54, 68] and the development of sensor manufacture [62], some large-scale multi-modal FAS datasets [15, 34, 45, 67] are proposed, which also drive the advancement in multimodal FAS [13, 14, 33, 36, 58, 60, 61]. By expanding the input from RGB to multi-modal, e.g., RGB, depth, and infrared, richer spoofing traces can be captured for detecting + +various presentation attacks [63]. However, existing multimodal methods fail to perform well under unseen deployment conditions. As illustrated in Fig. 1, though the current methods generalize well in unimodal scenarios, their performance abnormally declines when extended and applied to multi-modal FAS. Why using more modalities is not better? We argue that there are two main reasons: (1) Modality unreliability. Deploying in unseen environments or with various kinds of models of imaging sensors within the same modality, especially the depth and infrared shown in Fig. 1 (b), might introduce significant domain shifts [66]. Such shifts can result in unreliable multimodal live/spoof feature extraction [51], misleading mutual modalities through cross-modal fusion. (2) Modality imbalance. In some multi-modal learning scenarios, models tend to overly rely on a dominant modality with the highest convergence speed (i.e. fast modality), which prevents itself from fully exploiting the other dominated modalities with relatively slower convergence speeds (i.e. slow modalities) [43]. In FAS, modality imbalance can cause significant negative impacts, especially when the fast modality is unsuitable or unavailable for detecting attacks in the target domain, relying on slow modalities that have not adequately converged cannot achieve satisfactory performance. This study addresses the cross-domain limitations of multimodal FAS approaches due to modality unreliability and imbalance. As illustrated in Fig. 2, we introduce a MultiModal Domain Generalized (MMDG) framework for FAS. Within MMDG, we develop trainable Uncertainty-guided cross-Adapters (U-Adapter) to fine-tune Vision Transformers (ViT) [9] The U-Adapter address modality unreliability leveraging the uncertainty of each modality to suppress the focus on unreliable tokens during the cross-modal fusion. This prevents the excessive propagation of unreliable spoofing traces which tends to decrease the model's discriminative capability. Meanwhile, to solve modality imbalance, we design Rebalanced modality Gradient modulation strategy (ReGrad). Our ReGrad rebalance the convergence speed of all modalities by dynamically adjusting the backward propagated gradients of all trainable parameters in U-Adapters based on their conflict degree and convergence speed. This ensures that all modalities can be fully leveraged to resist various unseen attacks in the target domain. Our contributions are as follows. + +- We propose the MMDG framework to enhance the domain generalizability by addressing the modality unreliability and imbalance. This is the first investigation for multi-modal FAS under multi-source DG scenarios. +- Within MMDG, we propose a Uncertainty-Guided Cross-Adapter (U-Adapter) for ViT-based multi-modal FAS, which can complementarily fuse cross-modal features and prevents each modality from being influenced by unreliable information caused by domain shifts. + +- We design the Rebalanced Modality Gradient Modulation (ReGrad) for MMDG, a strategy that can monitor and rebalance the convergence speed of each modality by adaptively modulating gradients of U-Adapters. +- We build the first large-scale benchmark to evaluate the domain generalizability of multi-modal FAS approaches. Extensive experiments on this benchmark demonstrate the superiority and generalizability of MMDG. + +# 2. Related Works + +# 2.1. Domain Generalization for FAS + +In FAS, domain generalization focuses on training a model using multiple source domains, with the intention that this model can work effectively on unseen target domains [18, 46]. Previous works verify the effectiveness of adversarial training [22, 37, 65], asymmetric triplet loss [20, 30] and controversial learning [38, 49] in developing a shared feature space across domains. Besides, some studies [50, 70, 71] investigate the style information of different domains, aiming to learn domain-invariant features by minimizing distinct style characteristics. Furthermore, meta-learning-based methods [4, 10, 21, 44, 69] are also discussed to simulate domain shifts during training, fostering the learning of a robust representative feature space to such variations. More recent works focus on parameter-efficient transfer learning [5, 6, 18, 48], enabling a pretrained ViT to efficiently adapt to unseen domains with a lower risk of over-fitting. However, these approaches are proposed for unimodal (i.e. RGB) FAS. Due to the ignorance of modality unreliability, there is a gap in adapting them to multimodal scenarios. Simple multi-modal extensions of these approaches, through early fusion, late fusion, or even the substitution with multi-modal backbones, still exhibit limited performance under multi-modal DG scenarios. + +# 2.2. Multi-Modal FAS + +Multi-modal FAS refers to the use of multiple spectra, e.g., RGB, depth, and infrared, to reveal live and spoofing traces. Due to the complementary information among different modalities, traces that are undetectable in one modality can be easily captured in others [17, 36]. To fully exploit information from each modality, early methods employ channelwise concatenation of inputted modalities [14, 42], or develop multiple separate branches for modality feature extraction followed by late fusion [25, 26, 47, 58]. Recent works introduce attention-based feature fusion techniques [8, 27] and adaptive cross-modal loss functions [13] to encourage the extraction of complementary information among modalities. Meanwhile, cross-modality translation is discussed [29, 35] to mitigate the semantic gap between different modalities. To further enhance the practical usability of FAS, Yu et al. [60, 61, 62] introduce flexible-modal + +![](images/5fbb5220c5174dda9f95f70a11e7cd3f3b9b03821544f1fcd2b12dba8c72b718.jpg) +Figure 2. Overall structure of our MMDG. It consists of ViT-based backbones fine-tuned by the proposed U-Adapters with the modality rebalancing strategy called ReGrad. Each modality is designed with a branch for feature extraction and enables feature interaction and mutual complementarity with other modalities. For simplicity, we illustrate the two-modality scenario. + +benchmarks that force the use of incomplete modalities during training or testing. In the context of flexible-modal FAS, cross-modality attention [33, 36] or multi-modal adapters [60, 61] are proposed for pretrained ViT to learn modality-agnostic live/spoofing features. However, the aforementioned methods fail to perform well in DG scenarios. The lack of sensitivity to modality unreliability and imbalance leads to an insufficient ability to resist domain shifts. Although previous works rebalance modalities by modulating gradients [43] or adjusting modality losses [11], these methods ignore domain shifts and are not suitable for FAS frameworks with complex cross-modal fusion modules. + +# 3. Proposed MMDG + +As shown in Fig. 2, our MMDG consists of ViT-based backbones fine-tuned by the proposed U-Adapters and ReGrad strategy, which are respectively designed for solving modality unreliability and imbalance. + +# 3.1. Fine-tuning ViT with U-Adapters + +In multi-modal FAS, DG scenarios involving significant domain shifts [32], including unseen face presentations attacks [51], different kinds of imaging sensors, and noise from low-quality sensors [66], often lead to the extraction of unreliable live/spoofing features from certain modal inputs. These unreliable features can propagate during cross-modal fusion, adversely affecting other modalities. To this end, we aim to identify unreliable local features within each modality and suppress the interaction of these features + +with other modalities, ensuring that each modality utilizes reliable information from other modalities as complementary traces. Specifically, we adopt uncertainty estimated via Monte Carlo sampling [12, 23, 40] to evaluate the feature unreliability in each modality, and based on the uncertainty, we reduce their weight in cross-modal attention-based feature fusion, minimizing their negative impact. Inspired by recent success in parameter-efficient transfer learning (PETL) by introducing adapters into pretrained ViT [5, 6, 18, 39, 60, 61], but distinctively, we propose the U-Adapter for fine-tuning, which enables cross-modal feature fusion and recognizes as well as suppresses the propagation of unreliable information from uncertainly detected regions caused by domain shifts. In Fig. 3, we present the process of fine-tuning ViT with the U-Adapters, including adopting an Uncertainty Estimation Module (UEM) to recognize unreliable tokens and employing the U-Adapter for uncertainty-aware cross-modal fusion. The ViT [9], comprising $n$ transformer blocks each with Layer Normalization (LN), Multi-Head Self Attention (MHSA), and Multi-Layer Perceptron (MLP), is fine-tuned at the MLP's output using our U-Adapter. By representing our U-Adapter as $\mathcal{A}$ and taking RGB (R), depth (D), and infrared (I) as the inputted modalities, the fine-tuning process is formulated as follows. + +$$ +\begin{array}{l} \mathbf {x} _ {R} ^ {\text {o u t}} = \mathcal {A} \left(\mathbf {x} _ {D} ^ {U}, \mathbf {x} _ {D} ^ {3}, \mathbf {x} _ {R} ^ {3}\right) + \mathcal {A} \left(\mathbf {x} _ {I} ^ {U}, \mathbf {x} _ {I} ^ {3}, \mathbf {x} _ {R} ^ {3}\right) + \mathbf {x} _ {R} ^ {3} + \mathbf {x} _ {R} ^ {4}, (1) \\ \mathbf {x} _ {D} ^ {o u t} = \mathcal {A} \left(\mathbf {x} _ {R} ^ {U}, \mathbf {x} _ {R} ^ {3}, \mathbf {x} _ {D} ^ {3}\right) + \mathbf {x} _ {D} ^ {3} + \mathbf {x} _ {D} ^ {4}, (2) \\ \mathbf {x} _ {I} ^ {o u t} = \mathcal {A} \left(\mathbf {x} _ {R} ^ {U}, \mathbf {x} _ {R} ^ {3}, \mathbf {x} _ {I} ^ {3}\right) + \mathbf {x} _ {I} ^ {3} + \mathbf {x} _ {I} ^ {4}, (3) \\ \end{array} +$$ + +where for $m \in \{R, D, I\}$ , $\mathbf{x}_m^U \in \mathbb{R}^{B \times L \times 1}$ are the estimated token uncertainty in the corresponding modality, and $\mathbf{x}_m^3, \mathbf{x}_m^4 \in \mathbb{R}^{B \times L \times C}$ are the outputs of the last LN and MLP, respectively. The meaning of other vectors is illustrated in Fig. 3. To prevent the modalities that experienced significant domain shifts, i.e. depth and infrared, from misleading each other, we prohibit their direct interaction. + +Uncertainty Estimation Module. As discussed above, it is essential to recognize unreliable modal information for FAS. In deep learning, uncertainty reflects the level of confidence a model has in its predictions. When the model is less certain about a prediction, we consider that prediction to be unreliable [12]. When a model takes images with substantial noise caused by significant domain shifts as inputs, uncertainty estimation techniques [12] tend to assign elevated uncertainty values to regions more likely to be misclassified [32]. One well-known approach is the Bayesian Neural Network (BNN), widely applied in various vision tasks [16, 19]. BNN focuses on learning the posterior distribution of the model's weights rather than output a single value. Since the posterior distribution is always not tractable [12], approximate inference methods, such as Monte Carlo dropout (MCD) [7, 40, 53], are employed for estimation. In this work, we also use MCD to construct the UEM as a probabilistic representational model, which imposes a Bernoulli distribution over the weights of a model + +![](images/6dc01f823a1b95f5cd3632c0351fb3d87275aea4df21dd8f5b7028b03f384e7b.jpg) +Figure 3. (a1)-(a2) Illustration of fine-tuning ViT with proposed U-Adapters, showcasing the interaction between the RGB (R) and Depth (D) modalities. Note that only parameters of U-Adapters are trainable. (b) Uncertainty Estimation Module (UEM) used for recognizing unreliable tokens. (c) Detailed structure of U-Adapter, which adopts cross-modal fusion and suppresses the interference of unreliable tokens on other modalities. After fusion, discriminative central difference information is integrated for fine-grained spoof representation. + +![](images/a0c3c3d46d0a2cd544e91eb808ae4123941e612adcb06af32fa2e63b5079e7a7.jpg) + +![](images/42d110be83df66467e61188eec615a4ae2060dc7db05b6076e73fb1c7f4ca5c2.jpg) + +![](images/46fcc5a16d857ea8b5f61b5c7addeca4524310cf0c35685497681953151b4979.jpg) + +![](images/281db8e4e6573e8d3e06ecc3614d482064112755346ee3f4dc034b4211d0c9eb.jpg) + +without introducing extra parameters. As presented in Fig. 3 (b), we perform MCD instead of the vanilla dropout after each MHSA, simulating the effect of sampling from the posterior distribution by randomly dropping out neurons in MHSA. According to Bayesian probability theory, we consider $\mathcal{T} = \{\mathbf{x}_R^{u_1},\mathbf{x}_R^{u_2},\dots ,\mathbf{x}_R^{u_t}\}$ as empirical samples from the approximate distribution, and use their token-wise variance $\mathbf{x}_R^U = \mathrm{Var}(\mathcal{T})\in \mathbb{R}^{B\times L\times 1}$ to reflect their unreliability. + +Uncertainty-Guided Cross-Adapter. Different from existing cross-modal fusion [33, 36], our U-Adapter, guided by the estimated uncertainty, suppresses the interaction with unreliable tokens from other modalities to ensure the reliability of the fused features. As shown in Fig. 3 (c), U-Adapter interactively fuses features from different modalities and suppresses unreliable information during fusion. For example, in the depth branch (see Fig. 3 (a2)), U-Adapter takes the output of the second LN from another modality $\mathbf{x}_R^3$ as the query for attention-based feature fusion. The output of the second LN from its own modality $\mathbf{x}_D^3$ is used as key and value. Moreover, U-Adapter modulates the query tokens from the other modality based on their uncertainty. Since malicious presentation attacks result in subtle traces, we adopt central difference convolution (CDC) [59] to integrate the local details for ViT and explore fine-grained difference information from neighbor tokens, thus making the fused features more discriminative. By ignoring GeLU and vanilla convolution layers, the U-Adapter is formulated as follows. + +$$ +\mathcal {A} (\mathbf {x} _ {R} ^ {U}, \mathbf {x} _ {R} ^ {3}, \mathbf {x} _ {D} ^ {3}) = \mathcal {C} \left(\mathcal {S} \left(\frac {\mathrm {Q} (\mathbf {x} _ {R} ^ {3}) \mathrm {K} (\mathbf {x} _ {D} ^ {3}) ^ {\intercal} \odot \mathrm {U} (\mathbf {x} _ {R} ^ {U})}{\sqrt {n _ {k}}}\right) \mathrm {V} (\mathbf {x} _ {D} ^ {3}))\right), \tag {4} +$$ + +where $\mathcal{C}(\cdot)$ denotes the CDC layer, $S(\cdot)$ represents the softmax, and $n_k$ is the number of token channels. $\mathrm{Q}(\cdot),\mathrm{K}(\cdot)$ and $\mathrm{V}(\cdot)$ are the linear projections corresponding to the query, key, and value, respectively. $\odot$ is the Hadamard product. $\mathrm{U}(\mathbf{x}_R^U) = \exp (-r_e\cdot \mathbf{x}_R^U\times \mathbf{I})$ normalizes the inputted uncertainty using an exponential function, where $r_e$ is a penalty intensity parameter, and $\mathbf{I} = [1,1,\dots ,1]_{1\times L}$ is adopted to expand the dimensions of $x_R^U$ to match the size + +of $L \times L$ . We show the meaning of other vectors in Figs. 3 (a1-a2). Note that Eq. (4) only provides the scenario where RGB tokens are fused into depth in Fig. 3 (c). The pairwise fusion among other modalities follows similarly. + +# 3.2. Rebalanced Modality Gradient Modulation + +Different modalities are suitable for distinct types of attacks and deployment conditions, e.g., the depth modality being less sensitive to light and more apt for detecting replay attacks, and the infrared modality excelling in identifying wearable mask attacks and performing well in low-light conditions. To cope with various unknown attacks and deployment conditions, we aim to ensure that each modality can sufficiently converge, thereby fully leveraging their discriminative capability, rather than over-relying on a specific modality (i.e. fast modality) that shows superior performance in the source domain. + +Current modality balancing methods [11, 43], designed for simple late fusion models, are limited in their applicability to FAS. Late fusion networks have been shown to underperform in FAS compared to hybrid fusion networks [33, 36] that allow the fusion both before the final classifier and during the feature extraction. Additionally, the trainable parameters of a late fusion network are only affected by the gradient of its own modality, whereas those of a hybrid fusion network are influenced by the gradients of all modalities. Moreover, gradients from different modalities may even conflict with each other, making the modulation more difficult. To this end, our ReGrad adaptively modulates the gradients of all modalities that simultaneously act on each trainable layer. Moreover, existing modality balancing methods fail to generalize well while balancing each modality due to the ignorance of domain shifts. Drawing inspiration from SSDG [20] and PMR [11], we introduce a single-side prototypical loss to supervise modality branches and monitor their convergence speed. This loss enables each modality to cluster according to domain information through internal impetus, which further mitigates the domi + +![](images/ec8e154fde5ae0d5369fd1ed6977acf7deb00a96e0c0dd084c7dd6c278411da3.jpg) +Figure 4. Illustration of gradient modulation via the proposed ReGrad in different scenarios: (Row 1) Non-conflicting (a1) and faster modality $j$ (b1) or $i$ (c1). (Row 2) Conflicting (a2) and faster modality $j$ (b2) or $i$ (c2). + +nance of faster modalities. + +Modulating Gradients for Each Modality. When the optimization direction is dominated by the gradients of the fast modality, the convergence degree of slow modalities is suppressed. The discriminative ability of slow modalities may even be weakened when there is a conflict in the gradient directions between modalities [55]. Inspired by the studies on solving domain conflicts [41] and multi-task conflicts [55] via gradient modulation, our ReGrad adaptively performs gradient modulation based on the conflict and convergence degree of each modality, making it applicable to complex hybrid fusion networks. The criteria of modality convergence speed will be discussed in the following section. + +As shown in Fig. 4 (b1), if modality $i$ converges slower than $j$ and their gradients $\mathbf{g}_i$ and $\mathbf{g}_j$ are not conflicting, we suppress $\mathbf{g}_j$ and boost the learning of modality $i$ . This involves decomposing $\mathbf{g}_j$ into orthogonal $\mathbf{g}_j^v$ and non-orthogonal $\mathbf{g}_j^h$ components relative to $\mathbf{g}_i$ . We drop $\mathbf{g}_j^v$ to slow down modality $j$ and use $\mathbf{g}_j^h$ to accelerate modality $i$ . Furthermore, to prevent unreliable modalities that converge rapidly from misleading other modalities, we employ the estimated uncertainty of modality $j$ to further constrain $g_j$ . Conversely, if $\mathbf{g}_i$ conflicts with $\mathbf{g}_j$ and modality $i$ remains a slow modal, it is unreasonable to use any part of $\mathbf{g}_j$ to boost the learning of $\mathbf{g}_i$ . To prevent $\mathbf{g}_j$ from seriously slowing down the convergence speed of modality $i$ , as shown in Fig. 4 (b2), we remove the non-orthogonal part of $\mathbf{g}_j^h$ that is opposite to $\mathbf{g}_i$ . Meanwhile, we retain its orthogonal part $\mathbf{g}_j^v$ to allow modality $j$ to learn at a slower speed and reduce $\mathbf{g}_j$ 's interference with $\mathbf{g}_i$ . Similarly, the uncertainty-based suppression is adopted. Cases where the modality $i$ is faster are presented in Figs. 4 (c1)-(c2) and Eq. (5). + +$$ +\operatorname {R e G r a d} _ {2} \left(\mathbf {g} _ {i}, \mathbf {g} _ {j}\right) = \left\{ \begin{array}{l l} \mathbf {g} _ {i} + \frac {\mathbf {g} _ {i} \cdot \mathbf {g} _ {j}}{\| \mathbf {g} _ {i} \| ^ {2}} \mathbf {g} _ {i} \cdot \mathrm {U} \left(u _ {j}\right) & , \text {i n F i g . 4 (b 1)}, \\ \mathbf {g} _ {i} + \left(\mathbf {g} _ {j} - \frac {\mathbf {g} _ {i} \cdot \mathbf {g} _ {j}}{\| \mathbf {g} _ {i} \| ^ {2}} \mathbf {g} _ {i}\right) \cdot \mathrm {U} \left(u _ {j}\right) & , \text {i n F i g . 4 (b 2)}, \\ \frac {\mathbf {g} _ {i} \cdot \mathbf {g} _ {j}}{\| \mathbf {g} _ {j} \| ^ {2}} \mathbf {g} _ {j} \cdot \mathrm {U} \left(u _ {i}\right) + \mathbf {g} _ {j} & , \text {i n F i g . 4 (c 1)}, \\ \left(\mathbf {g} _ {i} - \frac {\mathbf {g} _ {i} \cdot \mathbf {g} _ {j}}{\| \mathbf {g} _ {j} \| ^ {2}} \mathbf {g} _ {j}\right) \cdot \mathrm {U} \left(u _ {i}\right) + \mathbf {g} _ {j} & , \text {i n F i g . 4 (c 2)}, \end{array} \right. \tag {5} +$$ + +where $u_{i}$ and $u_{j}$ are respectively the batch-wise mean uncertainty of the class tokens outputted by the last ViT block + +of the corresponding modality, denoting the modality uncertainty. + +Single-Side Prototypical Loss. Prototype learning has been successfully applied in FAS and modality rebalancing. PMR [11] indicates that introducing prototypes to each modality in multi-modal learning to guide the learning process can mitigate modality imbalance. However, vanilla prototype learning is suboptimal under DG scenarios as it neglects the differences in samples across various domains. Inspired by SSDG [20], we propose the Single-Side Prototypical (SSP) loss with the concept of asymmetry representation and prototype learning. Specifically, we introduce prototypes for the live faces of all domains and the fake faces from each domain. Formulate the domain set as $\mathcal{D} = \{d_{\ell}, d_s^1, d_s^2, \ldots, d_s^{N_s}\}$ , where $N_s$ is the number of source domains, $d_{\ell}$ denotes the domain consists of all live faces, and $d_s^i$ represents domain of spoof faces from the $i$ -th dataset. For modality $m$ , the prototype of domain $d \in \mathcal{D}$ is the center of all samples belonging to $d$ in the feature space: + +$$ +p _ {m} ^ {d} = \frac {1}{N _ {d}} \sum_ {c _ {m} \in d} c _ {m}, \tag {6} +$$ + +where $c_{m}$ is a sample of $d$ in modality $m$ , and $N_{d}$ denotes the total number of samples in $d$ . Based on the prototypes of each domain, we aim to reduce the distance between samples and the prototype of their corresponding domain and increase the distance from other prototypes, thus preventing conflicts when samples from different domains are trained together. Therefore, following the form of cross-entropy, the SSP loss $\mathcal{L}_{m}^{ssp}$ of modality $m$ is formulated as follows. + +$$ +\mathcal {L} _ {m} ^ {s s p} \left(c _ {m}, p _ {m} ^ {d _ {g t}}\right) = - \log \frac {\exp \left(- \mathrm {E D} \left(c _ {m} , p _ {m} ^ {d _ {g t}}\right)\right)}{\sum_ {d \in \mathcal {D}} \exp \left(- \mathrm {E D} \left(c _ {m} , p _ {m} ^ {d}\right)\right)}, \tag {7} +$$ + +where sample $c_{m}$ belongs to domain $d_{gt}$ . The function $\mathrm{ED}(\cdot)$ denotes the Euclidean distance of the input sample and its prototype. Our SSP loss encourages the features of each modality of the samples to be pulled to the corresponding domain's prototype. This optimization of each modality branch is driven by the intrinsic force within each modality. We mark the modality with lower $\mathcal{L}_m^{ssp}$ as a faster modality. Our final loss is expressed as follows. + +$$ +\mathcal {L} ^ {\text {f i n a l}} = \mathcal {L} ^ {c e} + \sum_ {m \in \{R, D, I \}} \lambda \cdot \mathcal {L} _ {m} ^ {s s p}, \tag {8} +$$ + +Table 1. Cross-dataset testing results under the fixed-modal scenarios (Protocol 1) among CASIA-CeFA (C), PADISI (P), CASIA-SURF (S), and WMCA (W). DG, MM, and FM are short for domain-generalized, multi-modal, and flexible-modal, respectively. + +
MethodTypeCPS → WCPW → SCSW → PPSW → CAverage
HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑
SSDG [20]DG26.0982.0328.5075.9141.8260.5640.4862.3134.2270.20
SSAN [50]DG17.7391.6927.9479.0434.4968.8536.4369.2929.1577.22
IADG [71]DG27.0286.5023.0483.1132.0673.8339.2463.6830.3476.78
ViTAF [18]DG20.5885.8229.1677.8030.7573.0339.7563.4430.0675.02
MM-CDCN [58]MM38.9265.3942.9359.7941.3861.5148.1453.7142.8460.10
CMFL [13]MM18.2288.8231.2075.6626.6880.8536.9366.8228.2678.04
ViT + AMA [61]FM17.5688.7427.5080.0021.1885.5147.4855.5628.4377.45
VP-FAS [60]FM16.2691.2224.4281.0721.7685.4639.3566.5525.4581.08
ViT [9]Baseline20.8884.7744.0557.9433.5871.8042.1556.4535.1667.74
MMDGOurs12.7993.8315.3292.8618.9588.6429.9376.5219.2587.96
+ +where $L^{ce}$ denotes the cross-entropy loss, and $\lambda$ is used to balance the trade-off between $\mathcal{L}_m^{ssp}$ and $\mathcal{L}^{ce}$ . Here, we utilize $\mathcal{L}^{ce}$ to ensure the model's performance, while $\mathcal{L}_m^{ssp}$ is employed to balance all modalities and enhance generalizability. Since $\mathcal{L}_m^{ssp}$ is not modality-specific, it does not undergo gradient modulation, which we perform to $\mathcal{L}^{ce}$ only. + +# 4. Experimental Evaluation + +# 4.1. Multi-Modal FAS Benchmark in DG Scenarios + +Datasets, Protocols, and Performance Metrics. With the increasing popularity of large-scale foundation models [24], the current trend is towards fully leveraging multiple and diverse source datasets/domains [1]. While training with multiple datasets has been widely studied for unimodal FAS [20, 50, 71], the multi-modal FAS community remains in the phase of single-dataset training [13, 33, 36]. Therefore, we propose the first large-scale multi-source benchmark to evaluate the DG performance of multi-modal FAS. This benchmark incorporates four commonly used multimodal datasets: CASIA-CeFA (C) [34], PADISI-Face (P) [45], CASIA-SURF (S) [67], and WMCA (W) [15], each featuring a variety of attack types. For each dataset, we use RGB, depth (D), and infrared (I) modalities as inputs. Following previous works [20], we use Half Total Error Rate (HTER) and Area Under the Receiver Operating Characteristic Curve (AUC) for evaluation. We design three protocols to evaluate the cross-dataset performance under different unseen deployment conditions, i.e., fixed modalities, missing modalities, and limited source domains. For Protocol 1, inspired by the leave-one-out (LOO) protocols designed in unimodal DG scenarios [20], we propose four multi-modal LOO sub-protocols among C, P, S, and W. For example, the sub-protocol CPS $\rightarrow$ W represents we take C, P, and S as training sets, while W is testing set. In Protocol 2, for each LOO sub-protocol of Protocol 1, we design three test-time missing-modal scenarios, i.e., D is missing, I is missing, and both D and I are missing. In Protocol 3, we limit the number of source domains by proposing two subprotocols, namely CW $\rightarrow$ PS and PS $\rightarrow$ CW. Implementation Details. We resize all RGB, depth, and infrared images to $224 \times 224 \times 3$ . Each modality's input is divided + +into $14 \times 14$ patch tokens and a class token is attached, collectively serving as the input to the ViT. The hidden size of each token is default set to 768. We employ the Adam optimizer with a learning rate of $5 \times 10^{-5}$ and a weight decay of $1 \times 10^{-3}$ to train 70 epochs. The batch size is set to 32. Our classifier is a single-layer fully-connected network that reduces the dimensionality of the class token outputted by the final ViT block from 768 to 2. The number $n$ of ViT blocks in MMDG is set to 12. ViT-b pretrained on ImageNet is adopted as the baseline. + +# 4.2. Cross-dataset Testing + +We compare our method on Protocols 1-3 with three categories of FAS methods: (1) Multi-modal methods, including CMFL [13] and MM-CDCN [58]. (2) Flexible-modal FAS, such as ViT + AMA [61] and VP-FAS [60]. (3) Domain generalized FAS i.e., SSDG [20], SSAN [50], IADG [71], and ViTAF [18]. For methods not designed for unimodal FAS, we re-implement them either by employing a concatenation of multi-modal inputs or by replacing the backbone with our ViT+U-Adapter or vanilla ViT. All these methods are trained on the proposed protocols. We automatically use their hyper-parameters as described in the corresponding reference papers or optimally assign better ones. + +Protocol 1: Fixed Modalities. We evaluated the cross-dataset performance of unimodal DG and multi-modal methods in the fixed modalities scenarios. Since unimodal DG methods cannot be directly applied to this protocol, we concatenate the images of the three modalities along the channel dimension. We then reduce the channel number to 3 using a trainable $1 \times 1$ convolution layer before feeding them to these methods. As shown in the Table 1, our MMDG achieves the best results. Compared to the second-ranked method, i.e. VP-FAS [60], there is an improvement of over $6\%$ in both HTER and AUC. A higher performance boost occurs when compared to domain-generalized methods. This indicates MMDG's effectiveness in addressing modality unreliability and imbalance. + +Protocol 2: Missing Modalities. As shown in Table 2, our MMDG outperforms other algorithms in all three scenarios of modality missing during testing. It's worth noting that ViT+AMA and VP-FAS are specifically designed + +Table 2. Cross-dataset testing results under the test-time missing-modal scenarios (Protocol 2) among CASIA-CeFA (C), PADISI (P), CASIA-SURF (S), and WMCA (W). We report the average HTER (\%) ↓ and AUC (\%) ↑ on four sub-protocols, i.e., CPS → W, CPW → S, CSW → P, and PSW → C. DG, MM, and FM are short for domain-generalized, multi-modal, and flexible-modal, respectively. + +
MethodTypeMissing DMissing IMissing D & IAverage
HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑
SSDG [20]DG38.9265.4537.6466.5739.1865.2238.5865.75
SSAN [50]DG36.7769.2141.2061.9233.5273.3837.1668.17
IADG [71]DG40.7258.7242.1761.8337.5066.9040.1362.49
ViTAF [18]DG34.9973.2235.8869.4035.8969.6135.5970.64
MM-CDCN [58]MM44.9055.3543.6058.3844.5455.0844.3556.27
CMFL [13]MM31.3774.6230.5575.4231.8974.2931.2774.78
ViT + AMA [61]FM29.2577.7032.3074.0631.4875.8231.0175.86
VP-FAS [60]FM29.1378.2729.6377.5130.4776.3129.7477.36
ViTBaseline40.0464.6936.7768.1936.2069.0237.6767.30
MMDGOurs24.8982.3923.3983.8225.2681.8624.5182.69
+ +Table 3. Cross-dataset testing results under the limited source domain scenarios (Protocol 3) among CeFA-CeFA (C), PADISI-UGC (P), CASIA-SURF (S), and WMCA (W). + +
MethodTypeCW → PSPS→ CW
HTER (%)AUC (%)HTER (%)AUC (%)
SSDG [20]DG25.3480.1746.9854.29
SSAN [50]DG26.5580.0639.1067.19
IADG [71]DG22.8283.8539.7063.46
ViTAF [18]DG29.6477.3639.9361.31
MM-CDCN [58]MM29.2876.8847.0051.94
CMFL [13]MM31.8672.7539.4363.17
ViT + AMA [61]FM29.2576.8938.0667.64
VP-FAS [60]FM25.9081.7944.3760.83
ViT [9]Baseline42.6657.8042.7560.41
MMDGOurs20.1288.2436.6070.35
+ +for scenarios of modality missing and have achieved better results compared to other methods. However, under the DG premise, they did not consider the modality imbalance and unreliability issues brought by domain shift, resulting in a difference of over $5\%$ in average HTER and AUC compared to our method. Another interesting observation is that SSAN, IADG, ViT, and ViT + AMA perform better in the Missing D & I scenario than in Missing D and Missing I. These methods lose more modalities but achieve performance improvement, indicating that unreliable information within the modalities leads to erroneous detection. This reflects the effectiveness of our approach in suppressing the interaction of unreliable information between modalities. + +Protocol 3: Limited Source Domains. Compared to protocol 1, the limitation of source domains results in a more significant domain shift. We observe that our proposed MMDG achieves optimal results in both sub-protocols (see Table 3), even in the presence of substantial domain shift in $\mathbf{PS} \rightarrow \mathbf{CW}$ . This further underscores the effectiveness of our U-Adapter and ReGrad in enhancing domain generalizability in unseen multi-modal deployment environments. + +# 4.3. Ablation Study and Discussion + +In this section, we perform ablation analysis on U-Adapter and ReGrad to verify their individual contribution. + +Effectiveness of U-Adapter. To validate the effectiveness of the proposed U-Adapter, we compare the network performance before and after removing the adapter. We also replace the backbone of existing DG methods (i.e., SSDG, SSAN, and IADG) and multi-modal methods (CMFL) to + +![](images/77cc751d4358442610afe8b2e3ec767b661dbdec20c8f643527b0be757f6b190.jpg) +Figure 5. Ablation results on our U-Adapter. We report average HTER $\downarrow$ and AUC $\uparrow$ on four sub-protocols in Protocol 1. + +![](images/3df8afb9f8c1b203173ed41cbc16dabb36e89caaff02f38ac14f203e4388e45a.jpg) + +Table 4. Ablation results on the proposed ReGrad. We report the average HTER and AUC on four sub-protocols in Protocol 1. + +
BackboneMethodHTER (%)AUC (%)
ViT [9]-31.1474.81
ViT [9]ReGrad30.5975.94
ViT + U-AdapterSSDG [20]26.1380.31
ViT + U-AdapterSSAN [50]26.9479.52
ViT + U-AdapterIADG [71]24.2382.54
ViT + U-AdapterPMR [11]24.5483.14
ViT + U-AdapterOGM-GE [43]29.3276.54
ViT + U-Adapter-27.3878.88
ViT + U-AdapterReGrad (Conflicted-only)21.9185.65
ViT + U-AdapterReGrad (Unconflicted-only)21.2385.51
ViT + U-AdapterReGrad19.4887.48
+ +observe the performance differences when using ViT+U-Adapter as the backbone versus using ViT alone. As shown in Fig. 5, compared to using ViT as the backbone, these methods experience a performance boost when additional U-Adapters are employed. The performance improvement is especially pronounced when no additional method is used, and when our proposed ReGrad is integrated. This demonstrates the effectiveness of U-Adapter and its universality for existing DG methods and multi-modal strategies. + +Effectiveness of ReGrad. Here, we try to answer two questions: (1) Whether the idea to address domain shift through modality rebalancing, is more suitable compared to existing DG methods? (2) Is our ReGrad more capable of enhancing the performance of complex cross-modal fusion frameworks like ViT+U-Adapter compared to existing modality balancing methods? From Table 4, we can observe that when ViT+U-Adapter is used as the backbone, our ReGrad surpasses existing DG methods and other modality rebalancing methods. Even for vanilla ViT, our ReGrad can enhance its performance. This indicates the effectiveness of addressing domain shifts in unseen scenarios by balancing + +![](images/fc83a24007af5a6d5712c374437e907df9fe01353c6de4f63e2a7ca909eb30af.jpg) +(a) w/o ReGrad + +![](images/fabe1805ccfd00ed388beff9bbf28d3441d6ca4374c0d311fcc15b6886a0f9e7.jpg) +(b) w/ ReGrad +Figure 6. Modality balance degree during training. (a) MMDG w/o ReGrad. (b) MMDG w/ ReGrad. The red dashed line in each sub-figure represents the variance of $\mathcal{L}_R^{ssp}$ , $\mathcal{L}_D^{ssp}$ , and $\mathcal{L}_I^{ssp}$ . + +Table 5. Impact of estimated uncertainty in U-Adapters and ReGrad. We report the average HTER $(\%)$ ↓ and AUC $(\%)$ ↑ on four sub-protocols in Protocol 1. + +
Using Estimated UncertaintyMetric Value
U-Adapter: Eq. (4)ReGrad: Eq. (5)HTER (%) ↓AUC (%) ↑
××28.4078.65
×✓26.1381.60
✓×22.7584.84
✓✓19.4887.48
+ +various modalities. The results in Table 4 also indicate that when compared to the SoTA modality balancing methods (i.e., OGM-GE and PMR), our ReGrad is more suitable for ViT + U-Adapter, which involves hybrid feature fusion and complex cross-modal interactions. + +Meanwhile, as illustrated in Fig. 6, the lower variance (right) of $\mathcal{L}_R^{ssp}$ , $\mathcal{L}_D^{ssp}$ , and $\mathcal{L}_I^{ssp}$ demonstrates that ReGrad obviously improves the balance degree of MMDG. We also notice that ReGrad causes slight fluctuations in the loss function values. This is due to the gradient modulation allowing the model to escape from local optima, seeking a direction with a higher degree of modality balance. Despite the fluctuations, MMDG can still converge after about 50 epochs. Additionally, we compare ReGrad with its two variants, which perform modulation only when modality gradients are conflicted and unconflicted, respectively. Although these two variants improve performance compared to not modulating gradients at all, their performance is inferior to modulating gradients in all situations. This validates the necessity of adaptive modulation under various scenarios. + +Impact of the Uncertainty in U-Adapter and ReGrad. To testify whether the estimated uncertainty can enhance detection performance, we made a comparison by removing the parts that use uncertainty for constraints in U-Adapter and ReGrad. From Table 5, it can be observed that adding uncertainty to constrain the propagation of unreliable information in U-Adapter and the gradient of fast modalities in ReGrad both leads to performance improvement. When uncertainty is added to both modules, HTER and AUC can be further improved by over $3\%$ . As shown in Fig. 7, MMDG can identify unreliable regions in depth (left) + +![](images/68e22734c7a5fea82f2c054cb969515725eabf94c4f1f355e5717e2b1c50900c.jpg) +Figure 7. Visual attention and uncertainty maps of our MMDG. Uncertainly predicted regions are highlighted with red boxes. The redder color indicates the higher uncertainty or more attention. + +and infrared (right) modalities. Although the U-Adapters of these modalities focus on regions with uncertain predictions, due to our idea of suppressing uncertain regions, these regions fail to mislead mutual modalities, enabling the capture of more reliable spoof traces against domain shifts. + +Discussion. The results in Tables 1-5 and Figs. 5-6 validate the MMDG's generalizability by addressing the problems of modality unreliability and imbalance. This work marks a stride towards practical deployments and can benefit the FAS community. However, our MMDG's performance can still be improved. Continued efforts on this benchmark are necessary for further progress. Meanwhile, our method also has space for improvement, e.g., further enhancing the degree of modality balance and adopting more efficient uncertainty estimation techniques. + +# 5. Conclusion + +We propose the first multi-modal DG framework for FAS, i.e. MMDG, to enhance generalizability by addressing modality unreliability and imbalance. To mitigate the interference of unreliable modalities, we propose U-Adapter to suppress uncertainly predicted regions during cross-modal feature fusion. Besides, we develop ReGrad to adaptively modulate the gradients of each modality branch to achieve modality balance, thus preventing an over-reliance on a dominant modality. We also establish the first large-scale benchmark for multi-modal FAS under DG scenarios. Extensive experiments demonstrate MMDG's effectiveness. + +Acknowledgement. This research was done at the Rapid-Rich Object Search (ROSE) Lab, Nanyang Technological University, and supported by the National Key Research and Development Program of China (Grant No.2022YFB3207700), the National Natural Science Foundation of China (Grants No.62272022, 62306061, and 62331006), the NTU-PKU Joint Research Institute (sponsored by the Ng Teng Fong Charitable Foundation), the Science and Technology Foundation of Guangzhou Huangpu Development District (Grant No.2022GH15), and the Guangdong Provincial Regional Joint Fund - Regional Cultivation Project (Grant No.2023A1515140037) + +# References + +[1] Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George van den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, Diego de Las Casas, Aurelia Guy, Jacob Menick, Roman Ring, Tom Hennigan, Saffron Huang, Loren Maggiore, Chris Jones, Albin Cassirer, Andy Brock, Michela Paganini, Geoffrey Irving, Oriol Vinyals, Simon Osindero, Karen Simonyan, Jack W. Rae, Erich Elsen, and Laurent Sifre. Improving language models by retrieving from trillions of tokens. In Proceedings of the International Conference on Machine Learning, pages 2206-2240, 2022. +[2] Rizhao Cai and Changsheng Chen. Learning deep forest with multi-scale local binary pattern features for face antispoofing. arXiv, abs/1910.03850, 2019. +[3] Rizhao Cai, Haoliang Li, Shiqi Wang, Changsheng Chen, and Alex C. Kot. Drl-fas: A novel framework based on deep reinforcement learning for face anti-spoofing. IEEE Transactions on Information Forensics and Security, 16:937-951, 2021. +[4] Rizhao Cai, Zhi Li, Renjie Wan, Haoliang Li, Yongjian Hu, and Alex C. Kot. Learning meta pattern for face antispoofing. IEEE Transactions on Information Forensics and Security, 17:1201-1213, 2022. +[5] Rizhao Cai, Yawen Cui, Zhi Li, Zitong Yu, Haoliang Li, Yongjian Hu, and Alex C. Kot. Rehearsal-free domain continual face anti-spoofing: Generalize more and forget less. In Proceedings of the IEEE International Conference on Computer Vision, 2023. +[6] Rizhao Cai, Zitong Yu, Chenqi Kong, Haoliang Li, Changsheng Chen, Yongjian Hu, and Alex C. Kot. S-adapter: Generalizing vision transformer for face anti-spoofing with statistical tokens. arXiv, abs/2309.04038, 2023. +[7] Pamela Carreno, Dana Kulic, and Michael Burke. Adapting neural models with sequential monte carlo dropout. In Conference on Robot Learning, pages 1542-1552, 2022. +[8] Pengchao Deng, Chenyang Ge, Xin Qiao, Hao Wei, and Yuan Sun. Attention-aware dual-stream network for multimodal face anti-spoofing. IEEE Transactions on Information Forensics and Security, 18:4258-4271, 2023. +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations, 2021. +[10] Zhekai Du, Jingjing Li, Lin Zuo, Lei Zhu, and Ke Lu. Energy-based domain generalization for face anti-spoofing. In Proceedings of the ACM International Conference on Multimedia, pages 1749-1757, 2022. +[11] Yunfeng Fan, Wenchao Xu, Haozhao Wang, Junxiao Wang, and Song Guo. PMR: prototypical modal rebalance for multimodal learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 20029-20038, 2023. + +[12] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the International Conference on Machine Learning, pages 1050-1059, 2016. +[13] Anjith George and Sébastien Marcel. Cross modal focal loss for RGBD face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7882-7891, 2021. +[14] Anjith George and Sébastien Marcel. Learning one class representations for face presentation attack detection using multi-channel convolutional neural networks. IEEE Transactions on Information Forensics and Security, 16:361-375, 2021. +[15] Anjith George, Zohreh Mostaani, David Geissenbuhler, Olegs Nikisins, André Anjos, and Sébastien Marcel. Biometric face presentation attack detection with multi-channel convolutional neural network. IEEE Transactions on Information Forensics and Security, 15:42-55, 2020. +[16] Hongji Guo, Hanjing Wang, and Qiang Ji. Uncertainty-guided probabilistic transformer for complex action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022. +[17] Shuangpeng Han, Rizhao Cai, Yawen Cui, Zitong Yu, Yongjian Hu, and Alex C. Kot. Hyperbolic face anti-spoofing. arXiv, abs/2308.09107, 2023. +[18] Hsin-Ping Huang, Deqing Sun, Yaojie Liu, Wen-Sheng Chu, Taihong Xiao, Jinwei Yuan, Hartwig Adam, and Ming-Hsuan Yang. Adaptive transformers for robust few-shot cross-domain face anti-spoofing. In Proceedings of the European Conference on Computer Vision, pages 37-54, 2022. +[19] Kaixiang Ji, Feng Chen, Xin Guo, Yadong Xu, Jian Wang, and Jingdong Chen. Uncertainty-guided learning for improving image manipulation detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 22456-22465, 2023. +[20] Yunpei Jia, Jie Zhang, Shiguang Shan, and Xilin Chen. Single-side domain generalization for face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8481-8490, 2020. +[21] Yunpei Jia, Jie Zhang, and Shiguang Shan. Dual-branch meta-learning network with distribution alignment for face anti-spoofing. IEEE Transactions on Information Forensics and Security, 17:138-151, 2022. +[22] Fangling Jiang, Qi Li, Pengcheng Liu, Xiang-Dong Zhou, and Zhenan Sun. Adversarial learning domain-invariant conditional features for robust face anti-spoofing. International Journal of Computer Vision, 131(7):1680-1703, 2023. +[23] Alex Kendall, Vijay Badrinarayanan, and Roberto Cipolla. Bayesian segnet: Model uncertainty in deep convolutional encoder-decoder architectures for scene understanding. In Proceedings of the British Machine Vision Conference, 2017. +[24] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In Proceedings of the IEEE International Conference on Computer Vision, pages 4015-4026, 2023. + +[25] Chenqi Kong, Kexin Zheng, Shiqi Wang, Anderson Rocha, and Haoliang Li. Beyond the pixel world: a novel acoustic-based face anti-spoofing system for smartphones. IEEE Transactions on Information Forensics and Security, 17: 3238-3253, 2022. +[26] Chenqi Kong, Kexin Zheng, Yibing Liu, Shiqi Wang, Anderson Rocha, and Haoliang Li. M3fas: An accurate and robust multimodal mobile face anti-spoofing system. arXiv preprint arXiv:2301.12831, 2023. +[27] Kaicheng Li, Hongyu Yang, Binghui Chen, Pengyu Li, Biao Wang, and Di Huang. Learning polysemantic spoof trace: A multi-modal disentanglement network for face anti-spoofing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1351-1359, 2023. +[28] Zhi Li, Rizhao Cai, Haoliang Li, Kwok-Yan Lam, Yongjian Hu, and Alex C. Kot. One-class knowledge distillation for face presentation attack detection. IEEE Transactions on Information Forensics and Security, 17:2137-2150, 2022. +[29] Zhi Li, Haoliang Li, Xin Luo, Yongjian Hu, Kwok-Yan Lam, and Alex C. Kot. Asymmetric modality translation for face presentation attack detection. IEEE Transactions on Multimedia, 25:62-76, 2023. +[30] Chen-Hao Liao, Wen-Cheng Chen, Hsuan-Tung Liu, Yi-Ren Yeh, Min-Chun Hu, and Chu-Song Chen. Domain invariant vision transformer learning for face anti-spoofing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6087-6096, 2023. +[31] Xun Lin, Shuai Wang, Jiahao Deng, Ying Fu, Xiao Bai, Xinlei Chen, Xiaolei Qu, and Wenzhong Tang. Image manipulation detection by multiple tampering traces and edge artifact enhancement. Pattern Recognition, 133:109026, 2023. +[32] Mattia Litrico, Alessio Del Bue, and Pietro Morerio. Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7640-7650, 2023. +[33] Ajian Liu and Yanyan Liang. Ma-vit: Modality-agnostic vision transformers for face anti-spoofing. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 1180–1186, 2022. +[34] Ajian Liu, Zichang Tan, Jun Wan, Sergio Escalera, Guodong Guo, and Stan Z. Li. CASIA-SURF cefa: A benchmark for multi-modal cross-ethnicity face anti-spoofing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1178–1186, 2021. +[35] Ajian Liu, Zichang Tan, Jun Wan, Yanyan Liang, Zhen Lei, Guodong Guo, and Stan Z. Li. Face anti-spoofing via adversarial cross-modality translation. IEEE Transactions on Information Forensics and Security, 16:2759-2772, 2021. +[36] Ajian Liu, Zichang Tan, Zitong Yu, Chenxu Zhao, Jun Wan, Yanyan Liang, Zhen Lei, Du Zhang, Stan Z. Li, and Guodong Guo. FM-ViT: Flexible modal vision transformers for face anti-spoofing. IEEE Transactions on Information Forensics and Security, 18:4775-4786, 2023. +[37] Yaojie Liu and Xiaoming Liu. Spoof trace disentanglement for generic face anti-spoofing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3813-3830, 2023. + +[38] Yuchen Liu, Yabo Chen, Mengran Gou, Chun-Ting Huang, Yaoming Wang, Wenrui Dai, and Hongkai Xiong. Towards unsupervised domain generalization for face anti-spoofing. In Proceedings of the IEEE International Conference on Computer Vision, pages 20654-20664, 2023. +[39] Anwei Luo, Rizhao Cai, Chenqi Kong, Xiangui Kang, Jiwu Huang, and Alex C. Kot. Forgery-aware adaptive vision transformer for face forgery detection. arXiv, abs/2309.11092, 2023. +[40] Yuki Mae, Wataru Kumagai, and Takafumi Kanamori. Uncertainty propagation for dropout-based bayesian neural networks. Neural Networks, 144:394-406, 2021. +[41] Lucas Mansilla, Rodrigo Echeveste, Diego H. Milone, and Enzo Ferrante. Domain generalization via gradient surgery. In Proceedings of the IEEE International Conference on Computer Vision, pages 6610-6618, 2021. +[42] Olegs Nikisins, Anjith George, and Sébastien Marcel. Domain adaptation in multi-channel autoencoder based features for robust face anti-spoofing. In Proceedings of the International Conference on Biometrics, pages 1-8, 2019. +[43] Xiaokang Peng, Yake Wei, Andong Deng, Dong Wang, and Di Hu. Balanced multimodal learning via on-the-fly gradient modulation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8228-8237, 2022. +[44] Yunxiao Qin, Zitong Yu, Longbin Yan, Zezheng Wang, Chenxu Zhao, and Zhen Lei. Meta-teacher for face antispoofing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6311-6326, 2022. +[45] Mohammad Rostami, Leonidas Spinoulas, Mohamed E. Hussein, Joe Mathai, and Wael Abd-Almageed. Detection and continual learning of novel face presentation attacks. In Proceedings of the IEEE International Conference on Computer Vision, pages 14831-14840, 2021. +[46] Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C. Yuen. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10023-10031, 2019. +[47] Tao Shen, Yuyu Huang, and Zhijun Tong. Facebagnet: Bag-of-local-features model for multi-modal face anti-spoofing. In IEEE Conference on Computer Vision and Pattern Recognition Workshop, pages 1611–1616, 2019. +[48] Koushik Srivatsan, Muzammal Naseer, and Karthik Nandakumar. Flip: Cross-domain face anti-spoofing with language guidance. In Proceedings of the IEEE International Conference on Computer Vision, pages 19685-19696, 2023. +[49] Yiyou Sun, Yaojie Liu, Xiaoming Liu, Yixuan Li, and WenSheng Chu. Rethinking domain generalization for face antispoofing: Separability and alignment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 24563-24574, 2023. +[50] Zhuo Wang, Zezheng Wang, Zitong Yu, Weihong Deng, Jiahong Li, Tingting Gao, and Zhongyuan Wang. Domain generalization via shuffled style assembly for face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4113-4123, 2022. + +[51] Junru Wu, Xiang Yu, Buyu Liu, Zhangyang Wang, and Manmohan Chandraker. Uncertainty-aware physically-guided proxy tasks for unseen domain face anti-spoofing. arXiv, abs/2011.14054, 2020. +[52] Wuyuan Xie, Kaimin Wang, Yakun Ju, and Miaohui Wang. pmbba: Projection-based blind point cloud quality assessment via multimodal learning. In Proceedings of the ACM International Conference on Multimedia, page 3250-3258, 2023. +[53] Lan Xixi, Changchun Zou, Cheng Peng, and Caowei Wu. Uncertainty quantification in intelligent-based electrical resistivity tomography image reconstruction with monte carlo dropout strategy. IEEE Transactions on Geoscience and Remote Sensing, 61:1-16, 2023. +[54] Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. +[55] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In Proceedings of the Neural Information Processing Systems, 2020. +[56] Yi Yu, Wenhan Yang, Yap-Peng Tan, and Alex C Kot. Towards robust rain removal against adversarial attacks: A comprehensive benchmark analysis and beyond. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6013-6022, 2022. +[57] Yi Yu, Yufei Wang, Wenhan Yang, Shijian Lu, Yap-Peng Tan, and Alex C Kot. Backdoor attacks against deep image compression via adaptive frequency trigger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12250-12259, 2023. +[58] Zitong Yu, Yunxiao Qin, Xiaobai Li, Zezheng Wang, Chenxu Zhao, Zhen Lei, and Guoying Zhao. Multi-modal face anti-spoofing based on central difference networks. In IEEE Conference on Computer Vision and Pattern Recognition Workshop, pages 2766-2774, 2020. +[59] Zitong Yu, Chenxu Zhao, Zezheng Wang, Yunxiao Qin, Zhuo Su, Xiaobai Li, Feng Zhou, and Guoying Zhao. Searching central difference convolutional networks for face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5294-5304, 2020. +[60] Zitong Yu, Rizhao Cai, Yawen Cui, Ajian Liu, and Changsheng Chen. Visual prompt flexible-modal face anti-spoofing. arXiv, abs/2307.13958, 2023. +[61] Zitong Yu, Rizhao Cai, Yawen Cui, Xin Liu, Yongjian Hu, and Alex C. Kot. Rethinking vision transformer and masked autoencoder in multimodal face anti-spoofing. arXiv, abs/2302.05744, 2023. +[62] Zitong Yu, Ajian Liu, Chenxu Zhao, Kevin H. M. Cheng, Xu Cheng, and Guoying Zhao. Flexible-modal face antispoofing: A benchmark. In IEEE Conference on Computer Vision and Pattern Recognition Workshop, pages 6346-6351, 2023. +[63] Zitong Yu, Yunxiao Qin, Xiaobai Li, Chenxu Zhao, Zhen Lei, and Guoying Zhao. Deep learning for face anti-spoofing: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5609-5631, 2023. + +[64] Zitong Yu, Rizhao Cai, Zhi Li, Wenhan Yang, Jingang Shi, and Alex C. Kot. Benchmarking joint face spoofing and forgery detection with visual and physiological cues. IEEE Transactions on Dependable and Secure Computing, pages 1-15, 2024. +[65] Haixiao Yue, Keyao Wang, Guosheng Zhang, Haocheng Feng, Junyu Han, Errui Ding, and Jingdong Wang. Cyclically disentangled feature translation for face anti-spoofing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3358-3366, 2023. +[66] Qingyang Zhang, Haitao Wu, Changqing Zhang, Qinghua Hu, Huazhu Fu, Joey Tianyi Zhou, and Xi Peng. Provable dynamic fusion for low-quality multimodal data. In Proceedings of the International Conference on Machine Learning, pages 41753-41769, 2023. +[67] Shifeng Zhang, Ajian Liu, Jun Wan, Yanyan Liang, Guodong Guo, Sergio Escalera, Hugo Jair Escalante, and Stan Z. Li. CASIA-SURF: A large-scale multi-modal benchmark for face anti-spoofing. IEEE transactions on biometrics, behavior, and identity science, 2(2):182-193, 2020. +[68] Taoying Zhang, Hesong Li, Qiankun Liu, Xiaoyong Wang, and Ying Fu. Mgt: Modality-guided transformer for infrared and visible image fusion. In Chinese Conference on Pattern Recognition and Computer Vision, pages 321-332, 2023. +[69] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Ran Yi, Shouhong Ding, and Lizhuang Ma. Adaptive mixture of experts learning for generalizable face anti-spoofing. In Proceedings of the ACM International Conference on Multimedia, pages 6009-6018, 2022. +[70] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Ran Yi, Kekai Sheng, Shouhong Ding, and Lizhuang Ma. Generative domain adaptation for face anti-spoofing. In Proceedings of the European Conference on Computer Vision, pages 335-356, 2022. +[71] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Xuequan Lu, Ran Yi, Shouhong Ding, and Lizhuang Ma. Instance-aware domain generalization for face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 20453-20463, 2023. \ No newline at end of file diff --git a/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/images.zip b/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2315e2db3dd3cb1ec9c9b710b6fbfe5ff89eb452 --- /dev/null +++ b/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8641123f51bd1bdec7d9708a08f445a9768c413ffb3c13fc7899ac4f9655261 +size 629620 diff --git a/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/layout.json b/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2654a9f8d20c09027f2cfd16db5845f149780c1e --- /dev/null +++ b/2024/Suppress and Rebalance_ Towards Generalized Multi-Modal Face Anti-Spoofing/layout.json @@ -0,0 +1,9231 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 544, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 544, + 121 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 544, + 121 + ], + "type": "text", + "content": "Suppress and Rebalance: Towards Generalized Multi-Modal Face Anti-Spoofing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "text", + "content": "Xun Lin" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "text", + "content": ", Shuai Wang" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "text", + "content": ", Rizhao Cai" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "text", + "content": ", Yizhong Liu" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "text", + "content": ", Ying Fu" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "text", + "content": ", Wenzhong Tang" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "text", + "content": ", Zitong Yu" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "inline_equation", + "content": "^{4*}" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "text", + "content": ", Alex Kot" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "text", + "content": "Beihang University " + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "text", + "content": "Nanyang Technological University " + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "text", + "content": "Beijing Institute of Technology " + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 53, + 142, + 547, + 185 + ], + "type": "text", + "content": "Great Bay University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 153, + 187, + 444, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 187, + 444, + 199 + ], + "spans": [ + { + "bbox": [ + 153, + 187, + 444, + 199 + ], + "type": "text", + "content": "{linxun, wangshuai}@buaa.edu.cn yuzitong@gbu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 252, + 290, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 290, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 290, + 577 + ], + "type": "text", + "content": "Face Anti-Spoofing (FAS) is crucial for securing face recognition systems against presentation attacks. With advancements in sensor manufacture and multi-modal learning techniques, many multi-modal FAS approaches have emerged. However, they face challenges in generalizing to unseen attacks and deployment conditions. These challenges arise from (1) modality unreliability, where some modality sensors like depth and infrared undergo significant domain shifts in varying environments, leading to the spread of unreliable information during cross-modal feature fusion, and (2) modality imbalance, where training overly relies on a dominant modality hinders the convergence of others, reducing effectiveness against attack types that are indistinguishable by sorely using the dominant modality. To address modality unreliability, we propose the Uncertainty-Guided Cross-Adapter (U-Adapter) to recognize unreliable detected regions within each modality and suppress the impact of unreliable regions on other modalities. For modality imbalance, we propose a Rebalanced Modality Gradient Modulation (ReGrad) strategy to rebalance the convergence speed of all modalities by adaptively adjusting their gradients. Besides, we provide the first large-scale benchmark for evaluating multi-modal FAS performance under domain generalization scenarios. Extensive experiments demonstrate that our method outperforms state-of-the-art methods. Source codes and protocols are released on https://github.com/OMGGGGG/mmdg." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 600, + 127, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 127, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 127, + 612 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 620, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 287, + 693 + ], + "type": "text", + "content": "The remarkable success of deep neural networks has also brought concerns regarding their security vulnerabilities [56, 57, 64]. Especially face recognition (FR) systems, widely used in applications like surveillance and mobile payment [63], are susceptible to various face presentation attacks, including printed photos, video replays, and 3D" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 313, + 228, + 535, + 380 + ], + "blocks": [ + { + "bbox": [ + 313, + 228, + 535, + 380 + ], + "lines": [ + { + "bbox": [ + 313, + 228, + 535, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 535, + 380 + ], + "type": "image", + "image_path": "2cb3603aa2d2e6fa2d850ae99159b9478ae02cbb2ef8f0081681452b21871964.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 426, + 545, + 481 + ], + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 481 + ], + "type": "text", + "content": "Figure 1. Illustration of DG scenarios in the context of (a) unimodal and (b) multi-modal. (c) DG performance of SSDG [20] on our Protocol 1 (see Sec. 4.1). Though fed with more modalities, SSDG performs worse in multi-modal scenarios compared to unimodal ones. * denotes our re-implemented multi-modal version." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 313, + 383, + 536, + 425 + ], + "blocks": [ + { + "bbox": [ + 313, + 383, + 536, + 425 + ], + "lines": [ + { + "bbox": [ + 313, + 383, + 536, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 536, + 425 + ], + "type": "table", + "html": "
(c)
MethodModalityHTER (%) ↓AUC (%) ↑
SSDG(a): RGB24.0783.69
SSDG*(b): RGB+D+I26.09 (+2.02)82.03 (-1.66)
", + "image_path": "e49bef10b176442fa7039eec061b7e7fb1eafc85fac4cc242be3b016d3cc92c8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 485, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 533 + ], + "type": "text", + "content": "wearable masks [3]. These vulnerabilities challenge the security of FR systems, limiting their broader application. As a result, face anti-spoofing (FAS) methods are developed to enhance the security of FR systems." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 534, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 628 + ], + "type": "text", + "content": "Recently, deep learning has been widely applied in FAS, particularly in unimodal scenarios based on the RGB spectrum. While these models perform well in intra-dataset evaluations, their generalizability to unseen attacks and deployment environments remains limited [2]. Despite the improvement of robustness via domain generalization (DG) techniques [20, 28, 49, 50, 71], these approaches still tend to over-fit the attacks presented in source domains." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "content": "With the recent progress of multi-modal learning [31, 52, 54, 68] and the development of sensor manufacture [62], some large-scale multi-modal FAS datasets [15, 34, 45, 67] are proposed, which also drive the advancement in multimodal FAS [13, 14, 33, 36, 58, 60, 61]. By expanding the input from RGB to multi-modal, e.g., RGB, depth, and infrared, richer spoofing traces can be captured for detecting" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "211" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 601 + ], + "type": "text", + "content": "various presentation attacks [63]. However, existing multimodal methods fail to perform well under unseen deployment conditions. As illustrated in Fig. 1, though the current methods generalize well in unimodal scenarios, their performance abnormally declines when extended and applied to multi-modal FAS. Why using more modalities is not better? We argue that there are two main reasons: (1) Modality unreliability. Deploying in unseen environments or with various kinds of models of imaging sensors within the same modality, especially the depth and infrared shown in Fig. 1 (b), might introduce significant domain shifts [66]. Such shifts can result in unreliable multimodal live/spoof feature extraction [51], misleading mutual modalities through cross-modal fusion. (2) Modality imbalance. In some multi-modal learning scenarios, models tend to overly rely on a dominant modality with the highest convergence speed (i.e. fast modality), which prevents itself from fully exploiting the other dominated modalities with relatively slower convergence speeds (i.e. slow modalities) [43]. In FAS, modality imbalance can cause significant negative impacts, especially when the fast modality is unsuitable or unavailable for detecting attacks in the target domain, relying on slow modalities that have not adequately converged cannot achieve satisfactory performance. This study addresses the cross-domain limitations of multimodal FAS approaches due to modality unreliability and imbalance. As illustrated in Fig. 2, we introduce a MultiModal Domain Generalized (MMDG) framework for FAS. Within MMDG, we develop trainable Uncertainty-guided cross-Adapters (U-Adapter) to fine-tune Vision Transformers (ViT) [9] The U-Adapter address modality unreliability leveraging the uncertainty of each modality to suppress the focus on unreliable tokens during the cross-modal fusion. This prevents the excessive propagation of unreliable spoofing traces which tends to decrease the model's discriminative capability. Meanwhile, to solve modality imbalance, we design Rebalanced modality Gradient modulation strategy (ReGrad). Our ReGrad rebalance the convergence speed of all modalities by dynamically adjusting the backward propagated gradients of all trainable parameters in U-Adapters based on their conflict degree and convergence speed. This ensures that all modalities can be fully leveraged to resist various unseen attacks in the target domain. Our contributions are as follows." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 47, + 605, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 287, + 653 + ], + "type": "text", + "content": "- We propose the MMDG framework to enhance the domain generalizability by addressing the modality unreliability and imbalance. This is the first investigation for multi-modal FAS under multi-source DG scenarios." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "text", + "content": "- Within MMDG, we propose a Uncertainty-Guided Cross-Adapter (U-Adapter) for ViT-based multi-modal FAS, which can complementarily fuse cross-modal features and prevents each modality from being influenced by unreliable information caused by domain shifts." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 72, + 545, + 168 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "text", + "content": "- We design the Rebalanced Modality Gradient Modulation (ReGrad) for MMDG, a strategy that can monitor and rebalance the convergence speed of each modality by adaptively modulating gradients of U-Adapters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 120, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 120, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 120, + 545, + 168 + ], + "type": "text", + "content": "- We build the first large-scale benchmark to evaluate the domain generalizability of multi-modal FAS approaches. Extensive experiments on this benchmark demonstrate the superiority and generalizability of MMDG." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 180, + 398, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 398, + 192 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 398, + 192 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 201, + 477, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 201, + 477, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 201, + 477, + 213 + ], + "type": "text", + "content": "2.1. Domain Generalization for FAS" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 219, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 547, + 495 + ], + "type": "text", + "content": "In FAS, domain generalization focuses on training a model using multiple source domains, with the intention that this model can work effectively on unseen target domains [18, 46]. Previous works verify the effectiveness of adversarial training [22, 37, 65], asymmetric triplet loss [20, 30] and controversial learning [38, 49] in developing a shared feature space across domains. Besides, some studies [50, 70, 71] investigate the style information of different domains, aiming to learn domain-invariant features by minimizing distinct style characteristics. Furthermore, meta-learning-based methods [4, 10, 21, 44, 69] are also discussed to simulate domain shifts during training, fostering the learning of a robust representative feature space to such variations. More recent works focus on parameter-efficient transfer learning [5, 6, 18, 48], enabling a pretrained ViT to efficiently adapt to unseen domains with a lower risk of over-fitting. However, these approaches are proposed for unimodal (i.e. RGB) FAS. Due to the ignorance of modality unreliability, there is a gap in adapting them to multimodal scenarios. Simple multi-modal extensions of these approaches, through early fusion, late fusion, or even the substitution with multi-modal backbones, still exhibit limited performance under multi-modal DG scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 503, + 411, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 411, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 411, + 514 + ], + "type": "text", + "content": "2.2. Multi-Modal FAS" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 522, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 715 + ], + "type": "text", + "content": "Multi-modal FAS refers to the use of multiple spectra, e.g., RGB, depth, and infrared, to reveal live and spoofing traces. Due to the complementary information among different modalities, traces that are undetectable in one modality can be easily captured in others [17, 36]. To fully exploit information from each modality, early methods employ channelwise concatenation of inputted modalities [14, 42], or develop multiple separate branches for modality feature extraction followed by late fusion [25, 26, 47, 58]. Recent works introduce attention-based feature fusion techniques [8, 27] and adaptive cross-modal loss functions [13] to encourage the extraction of complementary information among modalities. Meanwhile, cross-modality translation is discussed [29, 35] to mitigate the semantic gap between different modalities. To further enhance the practical usability of FAS, Yu et al. [60, 61, 62] introduce flexible-modal" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "212" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 72, + 266, + 284 + ], + "blocks": [ + { + "bbox": [ + 65, + 72, + 266, + 284 + ], + "lines": [ + { + "bbox": [ + 65, + 72, + 266, + 284 + ], + "spans": [ + { + "bbox": [ + 65, + 72, + 266, + 284 + ], + "type": "image", + "image_path": "5fbb5220c5174dda9f95f70a11e7cd3f3b9b03821544f1fcd2b12dba8c72b718.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 287, + 287, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 287, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 287, + 354 + ], + "type": "text", + "content": "Figure 2. Overall structure of our MMDG. It consists of ViT-based backbones fine-tuned by the proposed U-Adapters with the modality rebalancing strategy called ReGrad. Each modality is designed with a branch for feature extraction and enables feature interaction and mutual complementarity with other modalities. For simplicity, we illustrate the two-modality scenario." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 358, + 289, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 289, + 502 + ], + "type": "text", + "content": "benchmarks that force the use of incomplete modalities during training or testing. In the context of flexible-modal FAS, cross-modality attention [33, 36] or multi-modal adapters [60, 61] are proposed for pretrained ViT to learn modality-agnostic live/spoofing features. However, the aforementioned methods fail to perform well in DG scenarios. The lack of sensitivity to modality unreliability and imbalance leads to an insufficient ability to resist domain shifts. Although previous works rebalance modalities by modulating gradients [43] or adjusting modality losses [11], these methods ignore domain shifts and are not suitable for FAS frameworks with complex cross-modal fusion modules." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 512, + 153, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 153, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 153, + 525 + ], + "type": "text", + "content": "3. Proposed MMDG" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 533, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 582 + ], + "type": "text", + "content": "As shown in Fig. 2, our MMDG consists of ViT-based backbones fine-tuned by the proposed U-Adapters and ReGrad strategy, which are respectively designed for solving modality unreliability and imbalance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 587, + 226, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 226, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 226, + 601 + ], + "type": "text", + "content": "3.1. Fine-tuning ViT with U-Adapters" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": "In multi-modal FAS, DG scenarios involving significant domain shifts [32], including unseen face presentations attacks [51], different kinds of imaging sensors, and noise from low-quality sensors [66], often lead to the extraction of unreliable live/spoofing features from certain modal inputs. These unreliable features can propagate during cross-modal fusion, adversely affecting other modalities. To this end, we aim to identify unreliable local features within each modality and suppress the interaction of these features" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 358 + ], + "type": "text", + "content": "with other modalities, ensuring that each modality utilizes reliable information from other modalities as complementary traces. Specifically, we adopt uncertainty estimated via Monte Carlo sampling [12, 23, 40] to evaluate the feature unreliability in each modality, and based on the uncertainty, we reduce their weight in cross-modal attention-based feature fusion, minimizing their negative impact. Inspired by recent success in parameter-efficient transfer learning (PETL) by introducing adapters into pretrained ViT [5, 6, 18, 39, 60, 61], but distinctively, we propose the U-Adapter for fine-tuning, which enables cross-modal feature fusion and recognizes as well as suppresses the propagation of unreliable information from uncertainly detected regions caused by domain shifts. In Fig. 3, we present the process of fine-tuning ViT with the U-Adapters, including adopting an Uncertainty Estimation Module (UEM) to recognize unreliable tokens and employing the U-Adapter for uncertainty-aware cross-modal fusion. The ViT [9], comprising " + }, + { + "bbox": [ + 304, + 72, + 545, + 358 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 72, + 545, + 358 + ], + "type": "text", + "content": " transformer blocks each with Layer Normalization (LN), Multi-Head Self Attention (MHSA), and Multi-Layer Perceptron (MLP), is fine-tuned at the MLP's output using our U-Adapter. By representing our U-Adapter as " + }, + { + "bbox": [ + 304, + 72, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 72, + 545, + 358 + ], + "type": "text", + "content": " and taking RGB (R), depth (D), and infrared (I) as the inputted modalities, the fine-tuning process is formulated as follows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 328, + 361, + 545, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 361, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 328, + 361, + 545, + 399 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {x} _ {R} ^ {\\text {o u t}} = \\mathcal {A} \\left(\\mathbf {x} _ {D} ^ {U}, \\mathbf {x} _ {D} ^ {3}, \\mathbf {x} _ {R} ^ {3}\\right) + \\mathcal {A} \\left(\\mathbf {x} _ {I} ^ {U}, \\mathbf {x} _ {I} ^ {3}, \\mathbf {x} _ {R} ^ {3}\\right) + \\mathbf {x} _ {R} ^ {3} + \\mathbf {x} _ {R} ^ {4}, (1) \\\\ \\mathbf {x} _ {D} ^ {o u t} = \\mathcal {A} \\left(\\mathbf {x} _ {R} ^ {U}, \\mathbf {x} _ {R} ^ {3}, \\mathbf {x} _ {D} ^ {3}\\right) + \\mathbf {x} _ {D} ^ {3} + \\mathbf {x} _ {D} ^ {4}, (2) \\\\ \\mathbf {x} _ {I} ^ {o u t} = \\mathcal {A} \\left(\\mathbf {x} _ {R} ^ {U}, \\mathbf {x} _ {R} ^ {3}, \\mathbf {x} _ {I} ^ {3}\\right) + \\mathbf {x} _ {I} ^ {3} + \\mathbf {x} _ {I} ^ {4}, (3) \\\\ \\end{array}", + "image_path": "90ea06f6f6b5133d8a7994c375a5b4f3b2a212e9c42dc82af8de8170ad33469c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 400, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 545, + 483 + ], + "type": "text", + "content": "where for " + }, + { + "bbox": [ + 304, + 400, + 545, + 483 + ], + "type": "inline_equation", + "content": "m \\in \\{R, D, I\\}" + }, + { + "bbox": [ + 304, + 400, + 545, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 400, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_m^U \\in \\mathbb{R}^{B \\times L \\times 1}" + }, + { + "bbox": [ + 304, + 400, + 545, + 483 + ], + "type": "text", + "content": " are the estimated token uncertainty in the corresponding modality, and " + }, + { + "bbox": [ + 304, + 400, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_m^3, \\mathbf{x}_m^4 \\in \\mathbb{R}^{B \\times L \\times C}" + }, + { + "bbox": [ + 304, + 400, + 545, + 483 + ], + "type": "text", + "content": " are the outputs of the last LN and MLP, respectively. The meaning of other vectors is illustrated in Fig. 3. To prevent the modalities that experienced significant domain shifts, i.e. depth and infrared, from misleading each other, we prohibit their direct interaction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": "Uncertainty Estimation Module. As discussed above, it is essential to recognize unreliable modal information for FAS. In deep learning, uncertainty reflects the level of confidence a model has in its predictions. When the model is less certain about a prediction, we consider that prediction to be unreliable [12]. When a model takes images with substantial noise caused by significant domain shifts as inputs, uncertainty estimation techniques [12] tend to assign elevated uncertainty values to regions more likely to be misclassified [32]. One well-known approach is the Bayesian Neural Network (BNN), widely applied in various vision tasks [16, 19]. BNN focuses on learning the posterior distribution of the model's weights rather than output a single value. Since the posterior distribution is always not tractable [12], approximate inference methods, such as Monte Carlo dropout (MCD) [7, 40, 53], are employed for estimation. In this work, we also use MCD to construct the UEM as a probabilistic representational model, which imposes a Bernoulli distribution over the weights of a model" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "213" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 154, + 201 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 154, + 201 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 154, + 201 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 154, + 201 + ], + "type": "image", + "image_path": "6dc01f823a1b95f5cd3632c0351fb3d87275aea4df21dd8f5b7028b03f384e7b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 204, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 547, + 248 + ], + "type": "text", + "content": "Figure 3. (a1)-(a2) Illustration of fine-tuning ViT with proposed U-Adapters, showcasing the interaction between the RGB (R) and Depth (D) modalities. Note that only parameters of U-Adapters are trainable. (b) Uncertainty Estimation Module (UEM) used for recognizing unreliable tokens. (c) Detailed structure of U-Adapter, which adopts cross-modal fusion and suppresses the interference of unreliable tokens on other modalities. After fusion, discriminative central difference information is integrated for fine-grained spoof representation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 154, + 70, + 252, + 201 + ], + "blocks": [ + { + "bbox": [ + 154, + 70, + 252, + 201 + ], + "lines": [ + { + "bbox": [ + 154, + 70, + 252, + 201 + ], + "spans": [ + { + "bbox": [ + 154, + 70, + 252, + 201 + ], + "type": "image", + "image_path": "a0c3c3d46d0a2cd544e91eb808ae4123941e612adcb06af32fa2e63b5079e7a7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 253, + 70, + 306, + 202 + ], + "blocks": [ + { + "bbox": [ + 253, + 70, + 306, + 202 + ], + "lines": [ + { + "bbox": [ + 253, + 70, + 306, + 202 + ], + "spans": [ + { + "bbox": [ + 253, + 70, + 306, + 202 + ], + "type": "image", + "image_path": "42d110be83df66467e61188eec615a4ae2060dc7db05b6076e73fb1c7f4ca5c2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 451, + 202 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 451, + 202 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 451, + 202 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 451, + 202 + ], + "type": "image", + "image_path": "46fcc5a16d857ea8b5f61b5c7addeca4524310cf0c35685497681953151b4979.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 451, + 70, + 533, + 202 + ], + "blocks": [ + { + "bbox": [ + 451, + 70, + 533, + 202 + ], + "lines": [ + { + "bbox": [ + 451, + 70, + 533, + 202 + ], + "spans": [ + { + "bbox": [ + 451, + 70, + 533, + 202 + ], + "type": "image", + "image_path": "281db8e4e6573e8d3e06ecc3614d482064112755346ee3f4dc034b4211d0c9eb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 254, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 254, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 287, + 351 + ], + "type": "text", + "content": "without introducing extra parameters. As presented in Fig. 3 (b), we perform MCD instead of the vanilla dropout after each MHSA, simulating the effect of sampling from the posterior distribution by randomly dropping out neurons in MHSA. According to Bayesian probability theory, we consider " + }, + { + "bbox": [ + 46, + 254, + 287, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{\\mathbf{x}_R^{u_1},\\mathbf{x}_R^{u_2},\\dots ,\\mathbf{x}_R^{u_t}\\}" + }, + { + "bbox": [ + 46, + 254, + 287, + 351 + ], + "type": "text", + "content": " as empirical samples from the approximate distribution, and use their token-wise variance " + }, + { + "bbox": [ + 46, + 254, + 287, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_R^U = \\mathrm{Var}(\\mathcal{T})\\in \\mathbb{R}^{B\\times L\\times 1}" + }, + { + "bbox": [ + 46, + 254, + 287, + 351 + ], + "type": "text", + "content": " to reflect their unreliability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 351, + 288, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 288, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 288, + 589 + ], + "type": "text", + "content": "Uncertainty-Guided Cross-Adapter. Different from existing cross-modal fusion [33, 36], our U-Adapter, guided by the estimated uncertainty, suppresses the interaction with unreliable tokens from other modalities to ensure the reliability of the fused features. As shown in Fig. 3 (c), U-Adapter interactively fuses features from different modalities and suppresses unreliable information during fusion. For example, in the depth branch (see Fig. 3 (a2)), U-Adapter takes the output of the second LN from another modality " + }, + { + "bbox": [ + 46, + 351, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_R^3" + }, + { + "bbox": [ + 46, + 351, + 288, + 589 + ], + "type": "text", + "content": " as the query for attention-based feature fusion. The output of the second LN from its own modality " + }, + { + "bbox": [ + 46, + 351, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_D^3" + }, + { + "bbox": [ + 46, + 351, + 288, + 589 + ], + "type": "text", + "content": " is used as key and value. Moreover, U-Adapter modulates the query tokens from the other modality based on their uncertainty. Since malicious presentation attacks result in subtle traces, we adopt central difference convolution (CDC) [59] to integrate the local details for ViT and explore fine-grained difference information from neighbor tokens, thus making the fused features more discriminative. By ignoring GeLU and vanilla convolution layers, the U-Adapter is formulated as follows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 590, + 287, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 590, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 58, + 590, + 287, + 616 + ], + "type": "interline_equation", + "content": "\\mathcal {A} (\\mathbf {x} _ {R} ^ {U}, \\mathbf {x} _ {R} ^ {3}, \\mathbf {x} _ {D} ^ {3}) = \\mathcal {C} \\left(\\mathcal {S} \\left(\\frac {\\mathrm {Q} (\\mathbf {x} _ {R} ^ {3}) \\mathrm {K} (\\mathbf {x} _ {D} ^ {3}) ^ {\\intercal} \\odot \\mathrm {U} (\\mathbf {x} _ {R} ^ {U})}{\\sqrt {n _ {k}}}\\right) \\mathrm {V} (\\mathbf {x} _ {D} ^ {3}))\\right), \\tag {4}", + "image_path": "2b8a000697e16354f66d980dfc3cb738dc7a835301b4a38c54fc5b50682862ff.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\cdot)" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " denotes the CDC layer, " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " represents the softmax, and " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "n_k" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " is the number of token channels. " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{Q}(\\cdot),\\mathrm{K}(\\cdot)" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{V}(\\cdot)" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " are the linear projections corresponding to the query, key, and value, respectively. " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " is the Hadamard product. " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{U}(\\mathbf{x}_R^U) = \\exp (-r_e\\cdot \\mathbf{x}_R^U\\times \\mathbf{I})" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " normalizes the inputted uncertainty using an exponential function, where " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "r_e" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " is a penalty intensity parameter, and " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = [1,1,\\dots ,1]_{1\\times L}" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " is adopted to expand the dimensions of " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "x_R^U" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " to match the size" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 254, + 547, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 547, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 547, + 302 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 304, + 254, + 547, + 302 + ], + "type": "inline_equation", + "content": "L \\times L" + }, + { + "bbox": [ + 304, + 254, + 547, + 302 + ], + "type": "text", + "content": ". We show the meaning of other vectors in Figs. 3 (a1-a2). Note that Eq. (4) only provides the scenario where RGB tokens are fused into depth in Fig. 3 (c). The pairwise fusion among other modalities follows similarly." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 312, + 533, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 533, + 324 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 533, + 324 + ], + "type": "text", + "content": "3.2. Rebalanced Modality Gradient Modulation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 331, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 461 + ], + "type": "text", + "content": "Different modalities are suitable for distinct types of attacks and deployment conditions, e.g., the depth modality being less sensitive to light and more apt for detecting replay attacks, and the infrared modality excelling in identifying wearable mask attacks and performing well in low-light conditions. To cope with various unknown attacks and deployment conditions, we aim to ensure that each modality can sufficiently converge, thereby fully leveraging their discriminative capability, rather than over-relying on a specific modality (i.e. fast modality) that shows superior performance in the source domain." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 546, + 714 + ], + "type": "text", + "content": "Current modality balancing methods [11, 43], designed for simple late fusion models, are limited in their applicability to FAS. Late fusion networks have been shown to underperform in FAS compared to hybrid fusion networks [33, 36] that allow the fusion both before the final classifier and during the feature extraction. Additionally, the trainable parameters of a late fusion network are only affected by the gradient of its own modality, whereas those of a hybrid fusion network are influenced by the gradients of all modalities. Moreover, gradients from different modalities may even conflict with each other, making the modulation more difficult. To this end, our ReGrad adaptively modulates the gradients of all modalities that simultaneously act on each trainable layer. Moreover, existing modality balancing methods fail to generalize well while balancing each modality due to the ignorance of domain shifts. Drawing inspiration from SSDG [20] and PMR [11], we introduce a single-side prototypical loss to supervise modality branches and monitor their convergence speed. This loss enables each modality to cluster according to domain information through internal impetus, which further mitigates the domi" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "214" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 68, + 521, + 173 + ], + "blocks": [ + { + "bbox": [ + 72, + 68, + 521, + 173 + ], + "lines": [ + { + "bbox": [ + 72, + 68, + 521, + 173 + ], + "spans": [ + { + "bbox": [ + 72, + 68, + 521, + 173 + ], + "type": "image", + "image_path": "ec8e154fde5ae0d5369fd1ed6977acf7deb00a96e0c0dd084c7dd6c278411da3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "type": "text", + "content": "Figure 4. Illustration of gradient modulation via the proposed ReGrad in different scenarios: (Row 1) Non-conflicting (a1) and faster modality " + }, + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "type": "text", + "content": " (b1) or " + }, + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "type": "text", + "content": " (c1). (Row 2) Conflicting (a2) and faster modality " + }, + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "type": "text", + "content": " (b2) or " + }, + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "type": "text", + "content": " (c2)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 155, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 155, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 155, + 216 + ], + "type": "text", + "content": "nance of faster modalities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 219, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 287, + 361 + ], + "type": "text", + "content": "Modulating Gradients for Each Modality. When the optimization direction is dominated by the gradients of the fast modality, the convergence degree of slow modalities is suppressed. The discriminative ability of slow modalities may even be weakened when there is a conflict in the gradient directions between modalities [55]. Inspired by the studies on solving domain conflicts [41] and multi-task conflicts [55] via gradient modulation, our ReGrad adaptively performs gradient modulation based on the conflict and convergence degree of each modality, making it applicable to complex hybrid fusion networks. The criteria of modality convergence speed will be discussed in the following section." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": "As shown in Fig. 4 (b1), if modality " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " converges slower than " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " and their gradients " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_i" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_j" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " are not conflicting, we suppress " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_j" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " and boost the learning of modality " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": ". This involves decomposing " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_j" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " into orthogonal " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_j^v" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " and non-orthogonal " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_j^h" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " components relative to " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_i" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": ". We drop " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_j^v" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " to slow down modality " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " and use " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_j^h" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " to accelerate modality " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": ". Furthermore, to prevent unreliable modalities that converge rapidly from misleading other modalities, we employ the estimated uncertainty of modality " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " to further constrain " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "g_j" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": ". Conversely, if " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_i" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " conflicts with " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_j" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " and modality " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " remains a slow modal, it is unreasonable to use any part of " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_j" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " to boost the learning of " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_i" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": ". To prevent " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_j" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " from seriously slowing down the convergence speed of modality " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": ", as shown in Fig. 4 (b2), we remove the non-orthogonal part of " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_j^h" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " that is opposite to " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_i" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": ". Meanwhile, we retain its orthogonal part " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_j^v" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " to allow modality " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " to learn at a slower speed and reduce " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_j" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": "'s interference with " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_i" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": ". Similarly, the uncertainty-based suppression is adopted. Cases where the modality " + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 361, + 288, + 593 + ], + "type": "text", + "content": " is faster are presented in Figs. 4 (c1)-(c2) and Eq. (5)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 50, + 596, + 286, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 596, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 50, + 596, + 286, + 689 + ], + "type": "interline_equation", + "content": "\\operatorname {R e G r a d} _ {2} \\left(\\mathbf {g} _ {i}, \\mathbf {g} _ {j}\\right) = \\left\\{ \\begin{array}{l l} \\mathbf {g} _ {i} + \\frac {\\mathbf {g} _ {i} \\cdot \\mathbf {g} _ {j}}{\\| \\mathbf {g} _ {i} \\| ^ {2}} \\mathbf {g} _ {i} \\cdot \\mathrm {U} \\left(u _ {j}\\right) & , \\text {i n F i g . 4 (b 1)}, \\\\ \\mathbf {g} _ {i} + \\left(\\mathbf {g} _ {j} - \\frac {\\mathbf {g} _ {i} \\cdot \\mathbf {g} _ {j}}{\\| \\mathbf {g} _ {i} \\| ^ {2}} \\mathbf {g} _ {i}\\right) \\cdot \\mathrm {U} \\left(u _ {j}\\right) & , \\text {i n F i g . 4 (b 2)}, \\\\ \\frac {\\mathbf {g} _ {i} \\cdot \\mathbf {g} _ {j}}{\\| \\mathbf {g} _ {j} \\| ^ {2}} \\mathbf {g} _ {j} \\cdot \\mathrm {U} \\left(u _ {i}\\right) + \\mathbf {g} _ {j} & , \\text {i n F i g . 4 (c 1)}, \\\\ \\left(\\mathbf {g} _ {i} - \\frac {\\mathbf {g} _ {i} \\cdot \\mathbf {g} _ {j}}{\\| \\mathbf {g} _ {j} \\| ^ {2}} \\mathbf {g} _ {j}\\right) \\cdot \\mathrm {U} \\left(u _ {i}\\right) + \\mathbf {g} _ {j} & , \\text {i n F i g . 4 (c 2)}, \\end{array} \\right. \\tag {5}", + "image_path": "7130e30df2f4d8dd70442eeae445da081ed5aa5635bd67c2a782e997c3c17969.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "u_{j}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " are respectively the batch-wise mean uncertainty of the class tokens outputted by the last ViT block" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 205, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 228 + ], + "type": "text", + "content": "of the corresponding modality, denoting the modality uncertainty." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": "Single-Side Prototypical Loss. Prototype learning has been successfully applied in FAS and modality rebalancing. PMR [11] indicates that introducing prototypes to each modality in multi-modal learning to guide the learning process can mitigate modality imbalance. However, vanilla prototype learning is suboptimal under DG scenarios as it neglects the differences in samples across various domains. Inspired by SSDG [20], we propose the Single-Side Prototypical (SSP) loss with the concept of asymmetry representation and prototype learning. Specifically, we introduce prototypes for the live faces of all domains and the fake faces from each domain. Formulate the domain set as " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{d_{\\ell}, d_s^1, d_s^2, \\ldots, d_s^{N_s}\\}" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "N_s" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": " is the number of source domains, " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "d_{\\ell}" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": " denotes the domain consists of all live faces, and " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "d_s^i" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": " represents domain of spoof faces from the " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": "-th dataset. For modality " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": ", the prototype of domain " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "d \\in \\mathcal{D}" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": " is the center of all samples belonging to " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": " in the feature space:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 388, + 437, + 545, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 437, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 388, + 437, + 545, + 462 + ], + "type": "interline_equation", + "content": "p _ {m} ^ {d} = \\frac {1}{N _ {d}} \\sum_ {c _ {m} \\in d} c _ {m}, \\tag {6}", + "image_path": "b2c3033474710739db5bbf186cdd1903f71772dc65d191edde822b0db540c7b6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 465, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 465, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 545, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 465, + 545, + 561 + ], + "type": "inline_equation", + "content": "c_{m}" + }, + { + "bbox": [ + 304, + 465, + 545, + 561 + ], + "type": "text", + "content": " is a sample of " + }, + { + "bbox": [ + 304, + 465, + 545, + 561 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 465, + 545, + 561 + ], + "type": "text", + "content": " in modality " + }, + { + "bbox": [ + 304, + 465, + 545, + 561 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 465, + 545, + 561 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 465, + 545, + 561 + ], + "type": "inline_equation", + "content": "N_{d}" + }, + { + "bbox": [ + 304, + 465, + 545, + 561 + ], + "type": "text", + "content": " denotes the total number of samples in " + }, + { + "bbox": [ + 304, + 465, + 545, + 561 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 465, + 545, + 561 + ], + "type": "text", + "content": ". Based on the prototypes of each domain, we aim to reduce the distance between samples and the prototype of their corresponding domain and increase the distance from other prototypes, thus preventing conflicts when samples from different domains are trained together. Therefore, following the form of cross-entropy, the SSP loss " + }, + { + "bbox": [ + 304, + 465, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{m}^{ssp}" + }, + { + "bbox": [ + 304, + 465, + 545, + 561 + ], + "type": "text", + "content": " of modality " + }, + { + "bbox": [ + 304, + 465, + 545, + 561 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 465, + 545, + 561 + ], + "type": "text", + "content": " is formulated as follows." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 334, + 566, + 545, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 566, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 334, + 566, + 545, + 592 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m} ^ {s s p} \\left(c _ {m}, p _ {m} ^ {d _ {g t}}\\right) = - \\log \\frac {\\exp \\left(- \\mathrm {E D} \\left(c _ {m} , p _ {m} ^ {d _ {g t}}\\right)\\right)}{\\sum_ {d \\in \\mathcal {D}} \\exp \\left(- \\mathrm {E D} \\left(c _ {m} , p _ {m} ^ {d}\\right)\\right)}, \\tag {7}", + "image_path": "2d1076532be2e450c1377e7e489bd8daf3627df489996f566cae9f75378396a7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": "where sample " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "c_{m}" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": " belongs to domain " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "d_{gt}" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": ". The function " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{ED}(\\cdot)" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": " denotes the Euclidean distance of the input sample and its prototype. Our SSP loss encourages the features of each modality of the samples to be pulled to the corresponding domain's prototype. This optimization of each modality branch is driven by the intrinsic force within each modality. We mark the modality with lower " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m^{ssp}" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": " as a faster modality. Our final loss is expressed as follows." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 359, + 693, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 693, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 359, + 693, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\text {f i n a l}} = \\mathcal {L} ^ {c e} + \\sum_ {m \\in \\{R, D, I \\}} \\lambda \\cdot \\mathcal {L} _ {m} ^ {s s p}, \\tag {8}", + "image_path": "85e5808e85dce0ba7454cb7e343852297312d94cc6609867dc66d774ef25b1e0.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "215" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 92, + 545, + 190 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 92 + ], + "type": "text", + "content": "Table 1. Cross-dataset testing results under the fixed-modal scenarios (Protocol 1) among CASIA-CeFA (C), PADISI (P), CASIA-SURF (S), and WMCA (W). DG, MM, and FM are short for domain-generalized, multi-modal, and flexible-modal, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 92, + 545, + 190 + ], + "lines": [ + { + "bbox": [ + 48, + 92, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 48, + 92, + 545, + 190 + ], + "type": "table", + "html": "
MethodTypeCPS → WCPW → SCSW → PPSW → CAverage
HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑
SSDG [20]DG26.0982.0328.5075.9141.8260.5640.4862.3134.2270.20
SSAN [50]DG17.7391.6927.9479.0434.4968.8536.4369.2929.1577.22
IADG [71]DG27.0286.5023.0483.1132.0673.8339.2463.6830.3476.78
ViTAF [18]DG20.5885.8229.1677.8030.7573.0339.7563.4430.0675.02
MM-CDCN [58]MM38.9265.3942.9359.7941.3861.5148.1453.7142.8460.10
CMFL [13]MM18.2288.8231.2075.6626.6880.8536.9366.8228.2678.04
ViT + AMA [61]FM17.5688.7427.5080.0021.1885.5147.4855.5628.4377.45
VP-FAS [60]FM16.2691.2224.4281.0721.7685.4639.3566.5525.4581.08
ViT [9]Baseline20.8884.7744.0557.9433.5871.8042.1556.4535.1667.74
MMDGOurs12.7993.8315.3292.8618.9588.6429.9376.5219.2587.96
", + "image_path": "b3611ee8e5d2274e447d745fa4b852fc9f6df25bb4ea87a188f46137d6fabbe0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "inline_equation", + "content": "L^{ce}" + }, + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "text", + "content": " denotes the cross-entropy loss, and " + }, + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "text", + "content": " is used to balance the trade-off between " + }, + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m^{ssp}" + }, + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{ce}" + }, + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "text", + "content": ". Here, we utilize " + }, + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{ce}" + }, + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "text", + "content": " to ensure the model's performance, while " + }, + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m^{ssp}" + }, + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "text", + "content": " is employed to balance all modalities and enhance generalizability. Since " + }, + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m^{ssp}" + }, + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "text", + "content": " is not modality-specific, it does not undergo gradient modulation, which we perform to " + }, + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{ce}" + }, + { + "bbox": [ + 46, + 198, + 287, + 270 + ], + "type": "text", + "content": " only." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 290, + 191, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 290, + 191, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 290, + 191, + 304 + ], + "type": "text", + "content": "4. Experimental Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 311, + 287, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 287, + 323 + ], + "type": "text", + "content": "4.1. Multi-Modal FAS Benchmark in DG Scenarios" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 331, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 288, + 714 + ], + "type": "text", + "content": "Datasets, Protocols, and Performance Metrics. With the increasing popularity of large-scale foundation models [24], the current trend is towards fully leveraging multiple and diverse source datasets/domains [1]. While training with multiple datasets has been widely studied for unimodal FAS [20, 50, 71], the multi-modal FAS community remains in the phase of single-dataset training [13, 33, 36]. Therefore, we propose the first large-scale multi-source benchmark to evaluate the DG performance of multi-modal FAS. This benchmark incorporates four commonly used multimodal datasets: CASIA-CeFA (C) [34], PADISI-Face (P) [45], CASIA-SURF (S) [67], and WMCA (W) [15], each featuring a variety of attack types. For each dataset, we use RGB, depth (D), and infrared (I) modalities as inputs. Following previous works [20], we use Half Total Error Rate (HTER) and Area Under the Receiver Operating Characteristic Curve (AUC) for evaluation. We design three protocols to evaluate the cross-dataset performance under different unseen deployment conditions, i.e., fixed modalities, missing modalities, and limited source domains. For Protocol 1, inspired by the leave-one-out (LOO) protocols designed in unimodal DG scenarios [20], we propose four multi-modal LOO sub-protocols among C, P, S, and W. For example, the sub-protocol CPS " + }, + { + "bbox": [ + 46, + 331, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 331, + 288, + 714 + ], + "type": "text", + "content": " W represents we take C, P, and S as training sets, while W is testing set. In Protocol 2, for each LOO sub-protocol of Protocol 1, we design three test-time missing-modal scenarios, i.e., D is missing, I is missing, and both D and I are missing. In Protocol 3, we limit the number of source domains by proposing two subprotocols, namely CW " + }, + { + "bbox": [ + 46, + 331, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 331, + 288, + 714 + ], + "type": "text", + "content": " PS and PS " + }, + { + "bbox": [ + 46, + 331, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 331, + 288, + 714 + ], + "type": "text", + "content": " CW. Implementation Details. We resize all RGB, depth, and infrared images to " + }, + { + "bbox": [ + 46, + 331, + 288, + 714 + ], + "type": "inline_equation", + "content": "224 \\times 224 \\times 3" + }, + { + "bbox": [ + 46, + 331, + 288, + 714 + ], + "type": "text", + "content": ". Each modality's input is divided" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 198, + 547, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 319 + ], + "type": "text", + "content": "into " + }, + { + "bbox": [ + 304, + 198, + 547, + 319 + ], + "type": "inline_equation", + "content": "14 \\times 14" + }, + { + "bbox": [ + 304, + 198, + 547, + 319 + ], + "type": "text", + "content": " patch tokens and a class token is attached, collectively serving as the input to the ViT. The hidden size of each token is default set to 768. We employ the Adam optimizer with a learning rate of " + }, + { + "bbox": [ + 304, + 198, + 547, + 319 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 198, + 547, + 319 + ], + "type": "text", + "content": " and a weight decay of " + }, + { + "bbox": [ + 304, + 198, + 547, + 319 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 304, + 198, + 547, + 319 + ], + "type": "text", + "content": " to train 70 epochs. The batch size is set to 32. Our classifier is a single-layer fully-connected network that reduces the dimensionality of the class token outputted by the final ViT block from 768 to 2. The number " + }, + { + "bbox": [ + 304, + 198, + 547, + 319 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 198, + 547, + 319 + ], + "type": "text", + "content": " of ViT blocks in MMDG is set to 12. ViT-b pretrained on ImageNet is adopted as the baseline." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 329, + 429, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 329, + 429, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 329, + 429, + 342 + ], + "type": "text", + "content": "4.2. Cross-dataset Testing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 348, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 492 + ], + "type": "text", + "content": "We compare our method on Protocols 1-3 with three categories of FAS methods: (1) Multi-modal methods, including CMFL [13] and MM-CDCN [58]. (2) Flexible-modal FAS, such as ViT + AMA [61] and VP-FAS [60]. (3) Domain generalized FAS i.e., SSDG [20], SSAN [50], IADG [71], and ViTAF [18]. For methods not designed for unimodal FAS, we re-implement them either by employing a concatenation of multi-modal inputs or by replacing the backbone with our ViT+U-Adapter or vanilla ViT. All these methods are trained on the proposed protocols. We automatically use their hyper-parameters as described in the corresponding reference papers or optimally assign better ones." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "text", + "content": "Protocol 1: Fixed Modalities. We evaluated the cross-dataset performance of unimodal DG and multi-modal methods in the fixed modalities scenarios. Since unimodal DG methods cannot be directly applied to this protocol, we concatenate the images of the three modalities along the channel dimension. We then reduce the channel number to 3 using a trainable " + }, + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "text", + "content": " convolution layer before feeding them to these methods. As shown in the Table 1, our MMDG achieves the best results. Compared to the second-ranked method, i.e. VP-FAS [60], there is an improvement of over " + }, + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "text", + "content": " in both HTER and AUC. A higher performance boost occurs when compared to domain-generalized methods. This indicates MMDG's effectiveness in addressing modality unreliability and imbalance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Protocol 2: Missing Modalities. As shown in Table 2, our MMDG outperforms other algorithms in all three scenarios of modality missing during testing. It's worth noting that ViT+AMA and VP-FAS are specifically designed" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "216" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 104, + 544, + 202 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 103 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 103 + ], + "type": "text", + "content": "Table 2. Cross-dataset testing results under the test-time missing-modal scenarios (Protocol 2) among CASIA-CeFA (C), PADISI (P), CASIA-SURF (S), and WMCA (W). We report the average HTER (\\%) ↓ and AUC (\\%) ↑ on four sub-protocols, i.e., CPS → W, CPW → S, CSW → P, and PSW → C. DG, MM, and FM are short for domain-generalized, multi-modal, and flexible-modal, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 104, + 544, + 202 + ], + "lines": [ + { + "bbox": [ + 49, + 104, + 544, + 202 + ], + "spans": [ + { + "bbox": [ + 49, + 104, + 544, + 202 + ], + "type": "table", + "html": "
MethodTypeMissing DMissing IMissing D & IAverage
HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑HTER (%) ↓AUC (%) ↑
SSDG [20]DG38.9265.4537.6466.5739.1865.2238.5865.75
SSAN [50]DG36.7769.2141.2061.9233.5273.3837.1668.17
IADG [71]DG40.7258.7242.1761.8337.5066.9040.1362.49
ViTAF [18]DG34.9973.2235.8869.4035.8969.6135.5970.64
MM-CDCN [58]MM44.9055.3543.6058.3844.5455.0844.3556.27
CMFL [13]MM31.3774.6230.5575.4231.8974.2931.2774.78
ViT + AMA [61]FM29.2577.7032.3074.0631.4875.8231.0175.86
VP-FAS [60]FM29.1378.2729.6377.5130.4776.3129.7477.36
ViTBaseline40.0464.6936.7768.1936.2069.0237.6767.30
MMDGOurs24.8982.3923.3983.8225.2681.8624.5182.69
", + "image_path": "593eb446a9764c40a2502b9e16165b4e0ffed382f86aedcff7c2acd01c06aaae.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 244, + 286, + 335 + ], + "blocks": [ + { + "bbox": [ + 47, + 212, + 287, + 243 + ], + "lines": [ + { + "bbox": [ + 47, + 212, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 212, + 287, + 243 + ], + "type": "text", + "content": "Table 3. Cross-dataset testing results under the limited source domain scenarios (Protocol 3) among CeFA-CeFA (C), PADISI-UGC (P), CASIA-SURF (S), and WMCA (W)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 244, + 286, + 335 + ], + "lines": [ + { + "bbox": [ + 49, + 244, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 49, + 244, + 286, + 335 + ], + "type": "table", + "html": "
MethodTypeCW → PSPS→ CW
HTER (%)AUC (%)HTER (%)AUC (%)
SSDG [20]DG25.3480.1746.9854.29
SSAN [50]DG26.5580.0639.1067.19
IADG [71]DG22.8283.8539.7063.46
ViTAF [18]DG29.6477.3639.9361.31
MM-CDCN [58]MM29.2876.8847.0051.94
CMFL [13]MM31.8672.7539.4363.17
ViT + AMA [61]FM29.2576.8938.0667.64
VP-FAS [60]FM25.9081.7944.3760.83
ViT [9]Baseline42.6657.8042.7560.41
MMDGOurs20.1288.2436.6070.35
", + "image_path": "6b0884b9d7039c180df00185a86debdb09130d697b79c60f75d23b7d77a10153.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 347, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 502 + ], + "type": "text", + "content": "for scenarios of modality missing and have achieved better results compared to other methods. However, under the DG premise, they did not consider the modality imbalance and unreliability issues brought by domain shift, resulting in a difference of over " + }, + { + "bbox": [ + 46, + 347, + 287, + 502 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 347, + 287, + 502 + ], + "type": "text", + "content": " in average HTER and AUC compared to our method. Another interesting observation is that SSAN, IADG, ViT, and ViT + AMA perform better in the Missing D & I scenario than in Missing D and Missing I. These methods lose more modalities but achieve performance improvement, indicating that unreliable information within the modalities leads to erroneous detection. This reflects the effectiveness of our approach in suppressing the interaction of unreliable information between modalities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 505, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 287, + 601 + ], + "type": "text", + "content": "Protocol 3: Limited Source Domains. Compared to protocol 1, the limitation of source domains results in a more significant domain shift. We observe that our proposed MMDG achieves optimal results in both sub-protocols (see Table 3), even in the presence of substantial domain shift in " + }, + { + "bbox": [ + 46, + 505, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{PS} \\rightarrow \\mathbf{CW}" + }, + { + "bbox": [ + 46, + 505, + 287, + 601 + ], + "type": "text", + "content": ". This further underscores the effectiveness of our U-Adapter and ReGrad in enhancing domain generalizability in unseen multi-modal deployment environments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 609, + 214, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 214, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 214, + 621 + ], + "type": "text", + "content": "4.3. Ablation Study and Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 628, + 286, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 286, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 286, + 651 + ], + "type": "text", + "content": "In this section, we perform ablation analysis on U-Adapter and ReGrad to verify their individual contribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": "Effectiveness of U-Adapter. To validate the effectiveness of the proposed U-Adapter, we compare the network performance before and after removing the adapter. We also replace the backbone of existing DG methods (i.e., SSDG, SSAN, and IADG) and multi-modal methods (CMFL) to" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 212, + 420, + 298 + ], + "blocks": [ + { + "bbox": [ + 306, + 212, + 420, + 298 + ], + "lines": [ + { + "bbox": [ + 306, + 212, + 420, + 298 + ], + "spans": [ + { + "bbox": [ + 306, + 212, + 420, + 298 + ], + "type": "image", + "image_path": "77cc751d4358442610afe8b2e3ec767b661dbdec20c8f643527b0be757f6b190.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 298, + 545, + 319 + ], + "lines": [ + { + "bbox": [ + 306, + 298, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 298, + 545, + 319 + ], + "type": "text", + "content": "Figure 5. Ablation results on our U-Adapter. We report average HTER " + }, + { + "bbox": [ + 306, + 298, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 306, + 298, + 545, + 319 + ], + "type": "text", + "content": " and AUC " + }, + { + "bbox": [ + 306, + 298, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 306, + 298, + 545, + 319 + ], + "type": "text", + "content": " on four sub-protocols in Protocol 1." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 429, + 212, + 544, + 298 + ], + "blocks": [ + { + "bbox": [ + 429, + 212, + 544, + 298 + ], + "lines": [ + { + "bbox": [ + 429, + 212, + 544, + 298 + ], + "spans": [ + { + "bbox": [ + 429, + 212, + 544, + 298 + ], + "type": "image", + "image_path": "3df8afb9f8c1b203173ed41cbc16dabb36e89caaff02f38ac14f203e4388e45a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 350, + 544, + 450 + ], + "blocks": [ + { + "bbox": [ + 306, + 329, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 306, + 329, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 329, + 545, + 350 + ], + "type": "text", + "content": "Table 4. Ablation results on the proposed ReGrad. We report the average HTER and AUC on four sub-protocols in Protocol 1." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 350, + 544, + 450 + ], + "lines": [ + { + "bbox": [ + 309, + 350, + 544, + 450 + ], + "spans": [ + { + "bbox": [ + 309, + 350, + 544, + 450 + ], + "type": "table", + "html": "
BackboneMethodHTER (%)AUC (%)
ViT [9]-31.1474.81
ViT [9]ReGrad30.5975.94
ViT + U-AdapterSSDG [20]26.1380.31
ViT + U-AdapterSSAN [50]26.9479.52
ViT + U-AdapterIADG [71]24.2382.54
ViT + U-AdapterPMR [11]24.5483.14
ViT + U-AdapterOGM-GE [43]29.3276.54
ViT + U-Adapter-27.3878.88
ViT + U-AdapterReGrad (Conflicted-only)21.9185.65
ViT + U-AdapterReGrad (Unconflicted-only)21.2385.51
ViT + U-AdapterReGrad19.4887.48
", + "image_path": "c61e0fd24ebe4659961d4b5134c41984b29037f5386fcba89f7cd24bf3430564.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 459, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 567 + ], + "type": "text", + "content": "observe the performance differences when using ViT+U-Adapter as the backbone versus using ViT alone. As shown in Fig. 5, compared to using ViT as the backbone, these methods experience a performance boost when additional U-Adapters are employed. The performance improvement is especially pronounced when no additional method is used, and when our proposed ReGrad is integrated. This demonstrates the effectiveness of U-Adapter and its universality for existing DG methods and multi-modal strategies." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Effectiveness of ReGrad. Here, we try to answer two questions: (1) Whether the idea to address domain shift through modality rebalancing, is more suitable compared to existing DG methods? (2) Is our ReGrad more capable of enhancing the performance of complex cross-modal fusion frameworks like ViT+U-Adapter compared to existing modality balancing methods? From Table 4, we can observe that when ViT+U-Adapter is used as the backbone, our ReGrad surpasses existing DG methods and other modality rebalancing methods. Even for vanilla ViT, our ReGrad can enhance its performance. This indicates the effectiveness of addressing domain shifts in unseen scenarios by balancing" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "217" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 167, + 176 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 167, + 176 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 167, + 176 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 167, + 176 + ], + "type": "image", + "image_path": "fc83a24007af5a6d5712c374437e907df9fe01353c6de4f63e2a7ca909eb30af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 177, + 145, + 186 + ], + "lines": [ + { + "bbox": [ + 89, + 177, + 145, + 186 + ], + "spans": [ + { + "bbox": [ + 89, + 177, + 145, + 186 + ], + "type": "text", + "content": "(a) w/o ReGrad" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 70, + 286, + 176 + ], + "blocks": [ + { + "bbox": [ + 167, + 70, + 286, + 176 + ], + "lines": [ + { + "bbox": [ + 167, + 70, + 286, + 176 + ], + "spans": [ + { + "bbox": [ + 167, + 70, + 286, + 176 + ], + "type": "image", + "image_path": "fabe1805ccfd00ed388beff9bbf28d3441d6ca4374c0d311fcc15b6886a0f9e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 177, + 260, + 186 + ], + "lines": [ + { + "bbox": [ + 208, + 177, + 260, + 186 + ], + "spans": [ + { + "bbox": [ + 208, + 177, + 260, + 186 + ], + "type": "text", + "content": "(b) w/ ReGrad" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 186, + 287, + 221 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 287, + 221 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 287, + 221 + ], + "type": "text", + "content": "Figure 6. Modality balance degree during training. (a) MMDG w/o ReGrad. (b) MMDG w/ ReGrad. The red dashed line in each sub-figure represents the variance of " + }, + { + "bbox": [ + 46, + 186, + 287, + 221 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R^{ssp}" + }, + { + "bbox": [ + 46, + 186, + 287, + 221 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 186, + 287, + 221 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D^{ssp}" + }, + { + "bbox": [ + 46, + 186, + 287, + 221 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 186, + 287, + 221 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_I^{ssp}" + }, + { + "bbox": [ + 46, + 186, + 287, + 221 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 262, + 286, + 325 + ], + "blocks": [ + { + "bbox": [ + 46, + 229, + 287, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 287, + 262 + ], + "type": "text", + "content": "Table 5. Impact of estimated uncertainty in U-Adapters and ReGrad. We report the average HTER " + }, + { + "bbox": [ + 46, + 229, + 287, + 262 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 229, + 287, + 262 + ], + "type": "text", + "content": " ↓ and AUC " + }, + { + "bbox": [ + 46, + 229, + 287, + 262 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 229, + 287, + 262 + ], + "type": "text", + "content": " ↑ on four sub-protocols in Protocol 1." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 262, + 286, + 325 + ], + "lines": [ + { + "bbox": [ + 50, + 262, + 286, + 325 + ], + "spans": [ + { + "bbox": [ + 50, + 262, + 286, + 325 + ], + "type": "table", + "html": "
Using Estimated UncertaintyMetric Value
U-Adapter: Eq. (4)ReGrad: Eq. (5)HTER (%) ↓AUC (%) ↑
××28.4078.65
×✓26.1381.60
✓×22.7584.84
✓✓19.4887.48
", + "image_path": "6ef336f886044346db5b481314e73aa972eac52787f24190ebf828414073aab9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 340, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 399 + ], + "type": "text", + "content": "various modalities. The results in Table 4 also indicate that when compared to the SoTA modality balancing methods (i.e., OGM-GE and PMR), our ReGrad is more suitable for ViT + U-Adapter, which involves hybrid feature fusion and complex cross-modal interactions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 400, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 579 + ], + "type": "text", + "content": "Meanwhile, as illustrated in Fig. 6, the lower variance (right) of " + }, + { + "bbox": [ + 46, + 400, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R^{ssp}" + }, + { + "bbox": [ + 46, + 400, + 287, + 579 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 400, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D^{ssp}" + }, + { + "bbox": [ + 46, + 400, + 287, + 579 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 400, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_I^{ssp}" + }, + { + "bbox": [ + 46, + 400, + 287, + 579 + ], + "type": "text", + "content": " demonstrates that ReGrad obviously improves the balance degree of MMDG. We also notice that ReGrad causes slight fluctuations in the loss function values. This is due to the gradient modulation allowing the model to escape from local optima, seeking a direction with a higher degree of modality balance. Despite the fluctuations, MMDG can still converge after about 50 epochs. Additionally, we compare ReGrad with its two variants, which perform modulation only when modality gradients are conflicted and unconflicted, respectively. Although these two variants improve performance compared to not modulating gradients at all, their performance is inferior to modulating gradients in all situations. This validates the necessity of adaptive modulation under various scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Impact of the Uncertainty in U-Adapter and ReGrad. To testify whether the estimated uncertainty can enhance detection performance, we made a comparison by removing the parts that use uncertainty for constraints in U-Adapter and ReGrad. From Table 5, it can be observed that adding uncertainty to constrain the propagation of unreliable information in U-Adapter and the gradient of fast modalities in ReGrad both leads to performance improvement. When uncertainty is added to both modules, HTER and AUC can be further improved by over " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": ". As shown in Fig. 7, MMDG can identify unreliable regions in depth (left)" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 314, + 70, + 538, + 173 + ], + "blocks": [ + { + "bbox": [ + 314, + 70, + 538, + 173 + ], + "lines": [ + { + "bbox": [ + 314, + 70, + 538, + 173 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 538, + 173 + ], + "type": "image", + "image_path": "68e22734c7a5fea82f2c054cb969515725eabf94c4f1f355e5717e2b1c50900c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 176, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 305, + 176, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 176, + 545, + 209 + ], + "type": "text", + "content": "Figure 7. Visual attention and uncertainty maps of our MMDG. Uncertainly predicted regions are highlighted with red boxes. The redder color indicates the higher uncertainty or more attention." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 224, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 545, + 285 + ], + "type": "text", + "content": "and infrared (right) modalities. Although the U-Adapters of these modalities focus on regions with uncertain predictions, due to our idea of suppressing uncertain regions, these regions fail to mislead mutual modalities, enabling the capture of more reliable spoof traces against domain shifts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "type": "text", + "content": "Discussion. The results in Tables 1-5 and Figs. 5-6 validate the MMDG's generalizability by addressing the problems of modality unreliability and imbalance. This work marks a stride towards practical deployments and can benefit the FAS community. However, our MMDG's performance can still be improved. Continued efforts on this benchmark are necessary for further progress. Meanwhile, our method also has space for improvement, e.g., further enhancing the degree of modality balance and adopting more efficient uncertainty estimation techniques." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 417, + 378, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 417, + 378, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 417, + 378, + 430 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 437, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 568 + ], + "type": "text", + "content": "We propose the first multi-modal DG framework for FAS, i.e. MMDG, to enhance generalizability by addressing modality unreliability and imbalance. To mitigate the interference of unreliable modalities, we propose U-Adapter to suppress uncertainly predicted regions during cross-modal feature fusion. Besides, we develop ReGrad to adaptively modulate the gradients of each modality branch to achieve modality balance, thus preventing an over-reliance on a dominant modality. We also establish the first large-scale benchmark for multi-modal FAS under DG scenarios. Extensive experiments demonstrate MMDG's effectiveness." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This research was done at the Rapid-Rich Object Search (ROSE) Lab, Nanyang Technological University, and supported by the National Key Research and Development Program of China (Grant No.2022YFB3207700), the National Natural Science Foundation of China (Grants No.62272022, 62306061, and 62331006), the NTU-PKU Joint Research Institute (sponsored by the Ng Teng Fong Charitable Foundation), the Science and Technology Foundation of Guangzhou Huangpu Development District (Grant No.2022GH15), and the Guangdong Provincial Regional Joint Fund - Regional Cultivation Project (Grant No.2023A1515140037)" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "218" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 105, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 105, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 105, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 212 + ], + "type": "text", + "content": "[1] Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George van den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, Diego de Las Casas, Aurelia Guy, Jacob Menick, Roman Ring, Tom Hennigan, Saffron Huang, Loren Maggiore, Chris Jones, Albin Cassirer, Andy Brock, Michela Paganini, Geoffrey Irving, Oriol Vinyals, Simon Osindero, Karen Simonyan, Jack W. Rae, Erich Elsen, and Laurent Sifre. Improving language models by retrieving from trillions of tokens. In Proceedings of the International Conference on Machine Learning, pages 2206-2240, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 214, + 287, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 247 + ], + "type": "text", + "content": "[2] Rizhao Cai and Changsheng Chen. Learning deep forest with multi-scale local binary pattern features for face antispoofing. arXiv, abs/1910.03850, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 248, + 287, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 248, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 248, + 287, + 302 + ], + "type": "text", + "content": "[3] Rizhao Cai, Haoliang Li, Shiqi Wang, Changsheng Chen, and Alex C. Kot. Drl-fas: A novel framework based on deep reinforcement learning for face anti-spoofing. IEEE Transactions on Information Forensics and Security, 16:937-951, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 304, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 287, + 348 + ], + "type": "text", + "content": "[4] Rizhao Cai, Zhi Li, Renjie Wan, Haoliang Li, Yongjian Hu, and Alex C. Kot. Learning meta pattern for face antispoofing. IEEE Transactions on Information Forensics and Security, 17:1201-1213, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 350, + 287, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 287, + 406 + ], + "type": "text", + "content": "[5] Rizhao Cai, Yawen Cui, Zhi Li, Zitong Yu, Haoliang Li, Yongjian Hu, and Alex C. Kot. Rehearsal-free domain continual face anti-spoofing: Generalize more and forget less. In Proceedings of the IEEE International Conference on Computer Vision, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 408, + 287, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 287, + 451 + ], + "type": "text", + "content": "[6] Rizhao Cai, Zitong Yu, Chenqi Kong, Haoliang Li, Changsheng Chen, Yongjian Hu, and Alex C. Kot. S-adapter: Generalizing vision transformer for face anti-spoofing with statistical tokens. arXiv, abs/2309.04038, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 453, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 287, + 486 + ], + "type": "text", + "content": "[7] Pamela Carreno, Dana Kulic, and Michael Burke. Adapting neural models with sequential monte carlo dropout. In Conference on Robot Learning, pages 1542-1552, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 488, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 488, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 488, + 287, + 531 + ], + "type": "text", + "content": "[8] Pengchao Deng, Chenyang Ge, Xin Qiao, Hao Wei, and Yuan Sun. Attention-aware dual-stream network for multimodal face anti-spoofing. IEEE Transactions on Information Forensics and Security, 18:4258-4271, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 534, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 534, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 53, + 534, + 287, + 609 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[10] Zhekai Du, Jingjing Li, Lin Zuo, Lei Zhu, and Ke Lu. Energy-based domain generalization for face anti-spoofing. In Proceedings of the ACM International Conference on Multimedia, pages 1749-1757, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[11] Yunfeng Fan, Wenchao Xu, Haozhao Wang, Junxiao Wang, and Song Guo. PMR: prototypical modal rebalance for multimodal learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 20029-20038, 2023." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[12] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the International Conference on Machine Learning, pages 1050-1059, 2016." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[13] Anjith George and Sébastien Marcel. Cross modal focal loss for RGBD face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7882-7891, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "text", + "content": "[14] Anjith George and Sébastien Marcel. Learning one class representations for face presentation attack detection using multi-channel convolutional neural networks. IEEE Transactions on Information Forensics and Security, 16:361-375, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 219, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 274 + ], + "type": "text", + "content": "[15] Anjith George, Zohreh Mostaani, David Geissenbuhler, Olegs Nikisins, André Anjos, and Sébastien Marcel. Biometric face presentation attack detection with multi-channel convolutional neural network. IEEE Transactions on Information Forensics and Security, 15:42-55, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "text", + "content": "[16] Hongji Guo, Hanjing Wang, and Qiang Ji. Uncertainty-guided probabilistic transformer for complex action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 321, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 353 + ], + "type": "text", + "content": "[17] Shuangpeng Han, Rizhao Cai, Yawen Cui, Zitong Yu, Yongjian Hu, and Alex C. Kot. Hyperbolic face anti-spoofing. arXiv, abs/2308.09107, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 355, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 409 + ], + "type": "text", + "content": "[18] Hsin-Ping Huang, Deqing Sun, Yaojie Liu, Wen-Sheng Chu, Taihong Xiao, Jinwei Yuan, Hartwig Adam, and Ming-Hsuan Yang. Adaptive transformers for robust few-shot cross-domain face anti-spoofing. In Proceedings of the European Conference on Computer Vision, pages 37-54, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "text", + "content": "[19] Kaixiang Ji, Feng Chen, Xin Guo, Yadong Xu, Jian Wang, and Jingdong Chen. Uncertainty-guided learning for improving image manipulation detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 22456-22465, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 468, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 510 + ], + "type": "text", + "content": "[20] Yunpei Jia, Jie Zhang, Shiguang Shan, and Xilin Chen. Single-side domain generalization for face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8481-8490, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 513, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 545, + 555 + ], + "type": "text", + "content": "[21] Yunpei Jia, Jie Zhang, and Shiguang Shan. Dual-branch meta-learning network with distribution alignment for face anti-spoofing. IEEE Transactions on Information Forensics and Security, 17:138-151, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 557, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 600 + ], + "type": "text", + "content": "[22] Fangling Jiang, Qi Li, Pengcheng Liu, Xiang-Dong Zhou, and Zhenan Sun. Adversarial learning domain-invariant conditional features for robust face anti-spoofing. International Journal of Computer Vision, 131(7):1680-1703, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 602, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 646 + ], + "type": "text", + "content": "[23] Alex Kendall, Vijay Badrinarayanan, and Roberto Cipolla. Bayesian segnet: Model uncertainty in deep convolutional encoder-decoder architectures for scene understanding. In Proceedings of the British Machine Vision Conference, 2017." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[24] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In Proceedings of the IEEE International Conference on Computer Vision, pages 4015-4026, 2023." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "219" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[25] Chenqi Kong, Kexin Zheng, Shiqi Wang, Anderson Rocha, and Haoliang Li. Beyond the pixel world: a novel acoustic-based face anti-spoofing system for smartphones. IEEE Transactions on Information Forensics and Security, 17: 3238-3253, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "text", + "content": "[26] Chenqi Kong, Kexin Zheng, Yibing Liu, Shiqi Wang, Anderson Rocha, and Haoliang Li. M3fas: An accurate and robust multimodal mobile face anti-spoofing system. arXiv preprint arXiv:2301.12831, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "text", + "content": "[27] Kaicheng Li, Hongyu Yang, Binghui Chen, Pengyu Li, Biao Wang, and Di Huang. Learning polysemantic spoof trace: A multi-modal disentanglement network for face anti-spoofing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1351-1359, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 275 + ], + "type": "text", + "content": "[28] Zhi Li, Rizhao Cai, Haoliang Li, Kwok-Yan Lam, Yongjian Hu, and Alex C. Kot. One-class knowledge distillation for face presentation attack detection. IEEE Transactions on Information Forensics and Security, 17:2137-2150, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "text", + "content": "[29] Zhi Li, Haoliang Li, Xin Luo, Yongjian Hu, Kwok-Yan Lam, and Alex C. Kot. Asymmetric modality translation for face presentation attack detection. IEEE Transactions on Multimedia, 25:62-76, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 376 + ], + "type": "text", + "content": "[30] Chen-Hao Liao, Wen-Cheng Chen, Hsuan-Tung Liu, Yi-Ren Yeh, Min-Chun Hu, and Chu-Song Chen. Domain invariant vision transformer learning for face anti-spoofing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6087-6096, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "type": "text", + "content": "[31] Xun Lin, Shuai Wang, Jiahao Deng, Ying Fu, Xiao Bai, Xinlei Chen, Xiaolei Qu, and Wenzhong Tang. Image manipulation detection by multiple tampering traces and edge artifact enhancement. Pattern Recognition, 133:109026, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 477 + ], + "type": "text", + "content": "[32] Mattia Litrico, Alessio Del Bue, and Pietro Morerio. Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7640-7650, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "type": "text", + "content": "[33] Ajian Liu and Yanyan Liang. Ma-vit: Modality-agnostic vision transformers for face anti-spoofing. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 1180–1186, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 522, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 578 + ], + "type": "text", + "content": "[34] Ajian Liu, Zichang Tan, Jun Wan, Sergio Escalera, Guodong Guo, and Stan Z. Li. CASIA-SURF cefa: A benchmark for multi-modal cross-ethnicity face anti-spoofing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1178–1186, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 578, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 624 + ], + "type": "text", + "content": "[35] Ajian Liu, Zichang Tan, Jun Wan, Yanyan Liang, Zhen Lei, Guodong Guo, and Stan Z. Li. Face anti-spoofing via adversarial cross-modality translation. IEEE Transactions on Information Forensics and Security, 16:2759-2772, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 679 + ], + "type": "text", + "content": "[36] Ajian Liu, Zichang Tan, Zitong Yu, Chenxu Zhao, Jun Wan, Yanyan Liang, Zhen Lei, Du Zhang, Stan Z. Li, and Guodong Guo. FM-ViT: Flexible modal vision transformers for face anti-spoofing. IEEE Transactions on Information Forensics and Security, 18:4775-4786, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[37] Yaojie Liu and Xiaoming Liu. Spoof trace disentanglement for generic face anti-spoofing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3813-3830, 2023." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": "[38] Yuchen Liu, Yabo Chen, Mengran Gou, Chun-Ting Huang, Yaoming Wang, Wenrui Dai, and Hongkai Xiong. Towards unsupervised domain generalization for face anti-spoofing. In Proceedings of the IEEE International Conference on Computer Vision, pages 20654-20664, 2023." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "type": "text", + "content": "[39] Anwei Luo, Rizhao Cai, Chenqi Kong, Xiangui Kang, Jiwu Huang, and Alex C. Kot. Forgery-aware adaptive vision transformer for face forgery detection. arXiv, abs/2309.11092, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 546, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 546, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 546, + 207 + ], + "type": "text", + "content": "[40] Yuki Mae, Wataru Kumagai, and Takafumi Kanamori. Uncertainty propagation for dropout-based bayesian neural networks. Neural Networks, 144:394-406, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 208, + 546, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 546, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 546, + 253 + ], + "type": "text", + "content": "[41] Lucas Mansilla, Rodrigo Echeveste, Diego H. Milone, and Enzo Ferrante. Domain generalization via gradient surgery. In Proceedings of the IEEE International Conference on Computer Vision, pages 6610-6618, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 254, + 546, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 546, + 298 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 546, + 298 + ], + "type": "text", + "content": "[42] Olegs Nikisins, Anjith George, and Sébastien Marcel. Domain adaptation in multi-channel autoencoder based features for robust face anti-spoofing. In Proceedings of the International Conference on Biometrics, pages 1-8, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 299, + 546, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 546, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 546, + 354 + ], + "type": "text", + "content": "[43] Xiaokang Peng, Yake Wei, Andong Deng, Dong Wang, and Di Hu. Balanced multimodal learning via on-the-fly gradient modulation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8228-8237, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 355, + 546, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 546, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 546, + 399 + ], + "type": "text", + "content": "[44] Yunxiao Qin, Zitong Yu, Longbin Yan, Zezheng Wang, Chenxu Zhao, and Zhen Lei. Meta-teacher for face antispoofing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6311-6326, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 399, + 546, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 546, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 546, + 455 + ], + "type": "text", + "content": "[45] Mohammad Rostami, Leonidas Spinoulas, Mohamed E. Hussein, Joe Mathai, and Wael Abd-Almageed. Detection and continual learning of novel face presentation attacks. In Proceedings of the IEEE International Conference on Computer Vision, pages 14831-14840, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 456, + 546, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 546, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 546, + 511 + ], + "type": "text", + "content": "[46] Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C. Yuen. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10023-10031, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 512, + 546, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 546, + 556 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 546, + 556 + ], + "type": "text", + "content": "[47] Tao Shen, Yuyu Huang, and Zhijun Tong. Facebagnet: Bag-of-local-features model for multi-modal face anti-spoofing. In IEEE Conference on Computer Vision and Pattern Recognition Workshop, pages 1611–1616, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 557, + 546, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 546, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 546, + 602 + ], + "type": "text", + "content": "[48] Koushik Srivatsan, Muzammal Naseer, and Karthik Nandakumar. Flip: Cross-domain face anti-spoofing with language guidance. In Proceedings of the IEEE International Conference on Computer Vision, pages 19685-19696, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 602, + 546, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 546, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 546, + 657 + ], + "type": "text", + "content": "[49] Yiyou Sun, Yaojie Liu, Xiaoming Liu, Yixuan Li, and WenSheng Chu. Rethinking domain generalization for face antispoofing: Separability and alignment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 24563-24574, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 546, + 714 + ], + "type": "text", + "content": "[50] Zhuo Wang, Zezheng Wang, Zitong Yu, Weihong Deng, Jiahong Li, Tingting Gao, and Zhongyuan Wang. Domain generalization via shuffled style assembly for face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4113-4123, 2022." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "220" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[51] Junru Wu, Xiang Yu, Buyu Liu, Zhangyang Wang, and Manmohan Chandraker. Uncertainty-aware physically-guided proxy tasks for unseen domain face anti-spoofing. arXiv, abs/2011.14054, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 170 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 170 + ], + "type": "text", + "content": "[52] Wuyuan Xie, Kaimin Wang, Yakun Ju, and Miaohui Wang. pmbba: Projection-based blind point cloud quality assessment via multimodal learning. In Proceedings of the ACM International Conference on Multimedia, page 3250-3258, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "text", + "content": "[53] Lan Xixi, Changchun Zou, Cheng Peng, and Caowei Wu. Uncertainty quantification in intelligent-based electrical resistivity tomography image reconstruction with monte carlo dropout strategy. IEEE Transactions on Geoscience and Remote Sensing, 61:1-16, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 259 + ], + "type": "text", + "content": "[54] Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[55] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In Proceedings of the Neural Information Processing Systems, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "type": "text", + "content": "[56] Yi Yu, Wenhan Yang, Yap-Peng Tan, and Alex C Kot. Towards robust rain removal against adversarial attacks: A comprehensive benchmark analysis and beyond. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6013-6022, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "type": "text", + "content": "[57] Yi Yu, Yufei Wang, Wenhan Yang, Shijian Lu, Yap-Peng Tan, and Alex C Kot. Backdoor attacks against deep image compression via adaptive frequency trigger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12250-12259, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "text", + "content": "[58] Zitong Yu, Yunxiao Qin, Xiaobai Li, Zezheng Wang, Chenxu Zhao, Zhen Lei, and Guoying Zhao. Multi-modal face anti-spoofing based on central difference networks. In IEEE Conference on Computer Vision and Pattern Recognition Workshop, pages 2766-2774, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 535 + ], + "type": "text", + "content": "[59] Zitong Yu, Chenxu Zhao, Zezheng Wang, Yunxiao Qin, Zhuo Su, Xiaobai Li, Feng Zhou, and Guoying Zhao. Searching central difference convolutional networks for face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5294-5304, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "text", + "content": "[60] Zitong Yu, Rizhao Cai, Yawen Cui, Ajian Liu, and Changsheng Chen. Visual prompt flexible-modal face anti-spoofing. arXiv, abs/2307.13958, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "text", + "content": "[61] Zitong Yu, Rizhao Cai, Yawen Cui, Xin Liu, Yongjian Hu, and Alex C. Kot. Rethinking vision transformer and masked autoencoder in multimodal face anti-spoofing. arXiv, abs/2302.05744, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 667 + ], + "type": "text", + "content": "[62] Zitong Yu, Ajian Liu, Chenxu Zhao, Kevin H. M. Cheng, Xu Cheng, and Guoying Zhao. Flexible-modal face antispoofing: A benchmark. In IEEE Conference on Computer Vision and Pattern Recognition Workshop, pages 6346-6351, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[63] Zitong Yu, Yunxiao Qin, Xiaobai Li, Chenxu Zhao, Zhen Lei, and Guoying Zhao. Deep learning for face anti-spoofing: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5609-5631, 2023." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 507 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 126 + ], + "type": "text", + "content": "[64] Zitong Yu, Rizhao Cai, Zhi Li, Wenhan Yang, Jingang Shi, and Alex C. Kot. Benchmarking joint face spoofing and forgery detection with visual and physiological cues. IEEE Transactions on Dependable and Secure Computing, pages 1-15, 2024." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 546, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 183 + ], + "type": "text", + "content": "[65] Haixiao Yue, Keyao Wang, Guosheng Zhang, Haocheng Feng, Junyu Han, Errui Ding, and Jingdong Wang. Cyclically disentangled feature translation for face anti-spoofing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3358-3366, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 185, + 546, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 546, + 239 + ], + "type": "text", + "content": "[66] Qingyang Zhang, Haitao Wu, Changqing Zhang, Qinghua Hu, Huazhu Fu, Joey Tianyi Zhou, and Xi Peng. Provable dynamic fusion for low-quality multimodal data. In Proceedings of the International Conference on Machine Learning, pages 41753-41769, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 240, + 546, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 240, + 546, + 295 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 546, + 295 + ], + "type": "text", + "content": "[67] Shifeng Zhang, Ajian Liu, Jun Wan, Yanyan Liang, Guodong Guo, Sergio Escalera, Hugo Jair Escalante, and Stan Z. Li. CASIA-SURF: A large-scale multi-modal benchmark for face anti-spoofing. IEEE transactions on biometrics, behavior, and identity science, 2(2):182-193, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 296, + 546, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 296, + 546, + 340 + ], + "spans": [ + { + "bbox": [ + 308, + 296, + 546, + 340 + ], + "type": "text", + "content": "[68] Taoying Zhang, Hesong Li, Qiankun Liu, Xiaoyong Wang, and Ying Fu. Mgt: Modality-guided transformer for infrared and visible image fusion. In Chinese Conference on Pattern Recognition and Computer Vision, pages 321-332, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 341, + 546, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 341, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 341, + 546, + 396 + ], + "type": "text", + "content": "[69] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Ran Yi, Shouhong Ding, and Lizhuang Ma. Adaptive mixture of experts learning for generalizable face anti-spoofing. In Proceedings of the ACM International Conference on Multimedia, pages 6009-6018, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 396, + 546, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 396, + 546, + 450 + ], + "spans": [ + { + "bbox": [ + 308, + 396, + 546, + 450 + ], + "type": "text", + "content": "[70] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Ran Yi, Kekai Sheng, Shouhong Ding, and Lizhuang Ma. Generative domain adaptation for face anti-spoofing. In Proceedings of the European Conference on Computer Vision, pages 335-356, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 452, + 546, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 452, + 546, + 507 + ], + "spans": [ + { + "bbox": [ + 308, + 452, + 546, + 507 + ], + "type": "text", + "content": "[71] Qianyu Zhou, Ke-Yue Zhang, Taiping Yao, Xuequan Lu, Ran Yi, Shouhong Ding, and Lizhuang Ma. Instance-aware domain generalization for face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 20453-20463, 2023." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "221" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/6283bde0-3818-458b-8029-7eb693822789_content_list.json b/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/6283bde0-3818-458b-8029-7eb693822789_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fa1deb076dbe356b3397bddf8b9662a276fc86d7 --- /dev/null +++ b/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/6283bde0-3818-458b-8029-7eb693822789_content_list.json @@ -0,0 +1,1458 @@ +[ + { + "type": "text", + "text": "SurMo: Surface-based 4D Motion Modeling for Dynamic Human Rendering", + "text_level": 1, + "bbox": [ + 98, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tao Hu, Fangzhou Hong, Ziwei Liu \nS-Lab, Nanyang Technological University, Singapore", + "bbox": [ + 269, + 167, + 694, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 241, + 313, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dynamic human rendering from video sequences has achieved remarkable progress by formulating the rendering as a mapping from static poses to human images. However, existing methods focus on the human appearance reconstruction of every single frame while the temporal motion relations are not fully explored. In this paper, we propose a new 4D motion modeling paradigm, SurMo, that jointly models the temporal dynamics and human appearances in a unified framework with three key designs: 1) Surface-based motion encoding that models 4D human motions with an efficient compact surface-based triplane. It encodes both spatial and temporal motion relations on the dense surface manifold of a statistical body template, which inherits body topology priors for generalizable novel view synthesis with sparse training observations. 2) Physical motion decoding that is designed to encourage physical motion learning by decoding the motion triplane features at timestep $t$ to predict both spatial derivatives and temporal derivatives at the next timestep $t + 1$ in the training stage. 3) 4D appearance decoding that renders the motion triplanes into images by an efficient volumetric surface-conditioned renderer that focuses on the rendering of body surfaces with motion learning conditioning. Extensive experiments validate the state-of-the-art performance of our new paradigm and illustrate the expressiveness of surface-based motion triplanes for rendering high-fidelity view-consistent humans with fast motions and even motion-dependent shadows. Our project page is at: https://taohuumd.github.io/projects/SurMo.", + "bbox": [ + 75, + 266, + 473, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 709, + 209, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating volumetric videos of human actors is required in many applications such as AR/VR, telepresence, video game and film character creation. Recent neural rendering methods [25, 43, 54, 61] have made great progress in generating free-viewpoint videos of humans from several sparse multi-view videos, which are simple yet effective compared with traditional graphics approaches [2, 3, 62]. At the core of the technique is to model pose- and time-varying appearances of dynamic humans. The motions of dressed humans are often expressed as the movement of body and a sequence of natural secondary motion of clothes, \\( e.g., dy", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/56431c814e4a15490b1d14a0b1ed9f67e6e1f8d110475da3c25b01daf272b112.jpg", + "image_caption": [ + "Figure 1. Given several sparse multi-view video sequences with estimated 3D body meshes, SurMo synthesizes subject-specific appearance. We specifically focus on the synthesis of plausible time-varying appearances by learning an effective 4D motion representation." + ], + "image_footnote": [], + "bbox": [ + 498, + 239, + 893, + 425 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "namic movements of the T-shirt induced by dancing in Fig. 1. The secondary motion of clothes arises from intricate physical interactions with the body, typically changing over time, which leads to challenges for plausible rendering of dynamic humans.", + "bbox": [ + 498, + 516, + 890, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing methods [25, 43, 61] are conditioned on static poses and use a pose-guided generator to synthesize time-varying appearances. However, the appearance of clothed humans undergoes complex geometric transformations induced not only by the static pose but also its dynamics, whereas the modeling of dynamics is often ignored. In addition, these methods are often focused on the 2D appearance reconstruction of every single frame while the temporal motion relations are not fully explored. Due to these issues, they fail to render plausible secondary motions of clothes, i.e., generating the same appearances for fast and slow motions. A key challenge of learning temporal dynamics lies in the requirement of a tremendous amount of training observations to construct a 4D motion volume.", + "bbox": [ + 496, + 595, + 892, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To solve these issues, we propose a new paradigm to learn time-varying appearances of the secondary motion from just several sparse viewpoint video sequences, which is achieved by jointly modeling the temporal motion dynamics and human appearances in a unified rendering framework based on an efficient surface-based motion", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6550", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "representation. At the core of the paradigm is a feature encoder-decoder framework with three key components: 1) surface-based motion encoding; 2) physical motion decoding; and 3) 4D appearance decoding.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Firstly, in contrast to existing pose-guided methods [25, 42, 43, 61] that focus on static poses as a conditional variable, we extract an expressive 4D motion input from the 3D body mesh sequences obtained from training video as our input, which includes both a static pose represented by a spatial 3D mesh and its temporal dynamics. Furthermore, we notice that the non-rigid deformations of garments typically occur around the body surface instead of in a 3D volume, and hence we propose to model human motions on the body surface by projecting the extracted spatial-temporal 4D motion input to the dense 2D surface UV manifold of a clothless body template (e.g., SMPL [28]). To model temporal clothing offsets, a motion encoder is employed to lift the clothless motion features into a motion triplane that encodes both spatial and temporal motion relations in a compact 3D triplane with time-varying dynamics conditioned. The triplane is defined in the surface $\\mathbf{u} - \\mathbf{v} - \\mathbf{h}$ coordinate system, with $\\mathbf{u} - \\mathbf{v}$ to represent the motion of the clothless body template, and an extensional coordinate $\\mathbf{h}$ to represent the secondary motion of clothes, i.e., the temporal clothing offsets are parameterized by a signed distance to the body surface. In this way, 4D motions can be effectively represented by a surface-based triplane.", + "bbox": [ + 75, + 154, + 472, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Secondly, we propose to physically model the spatial and temporal motion dynamics in the rendering network. Specifically, with the surface-based triplane conditioned at time $t$ , a motion decoder is employed to decode the motion triplanes to predict the motion at the next timestep $t + 1$ , i.e., spatial derivative of the motion physically corresponding to surface normal map and temporal derivatives corresponding to surface velocity. We illustrate the physical motion learning significantly improves the rendering quality.", + "bbox": [ + 75, + 502, + 468, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thirdly, the motion triplanes are decoded into high-quality images at two stages: a volumetric surface-conditioned renderer that is focused on the rendering around human body surface and filters the query points far from the body surface for efficient volumetric rendering, and a geometry-aware super-resolution module for efficient high-quality image synthesis.", + "bbox": [ + 75, + 640, + 468, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct a systematical analysis of how human appearances are affected by temporal dynamics, and it was observed that some baseline methods mainly generate pose-dependent appearances instead of time-varying appearances for free-view video generations. In addition, quantitative and qualitative experiments are performed on three datasets with a total of 9 subject sequences, including ZJU-MoCap [43], AIST++ [24], and MPII-RDDC [13], which validate the effectiveness of SurMo in different scenarios.", + "bbox": [ + 75, + 747, + 468, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are:", + "bbox": [ + 96, + 885, + 326, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) A new paradigm for learning dynamic humans from videos that jointly models temporal motions and human appearances in a unified framework, and one of the early works that systematically analyzes how human appearances are affected by temporal dynamics.", + "2) An efficient surface-based triplane that encodes both spatial and temporal motion relations for expressive 4D motion modeling.", + "3) We achieve state-of-the-art results and show that our new paradigm is capable of learning high-fidelity appearances from fast motion sequences (e.g., AIST++ dance videos) or synthesizing motion-dependent shadows in challenging scenarios." + ], + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 301, + 640, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method is closely related to many sub-fields of visual computing, and below we discuss a set of the work.", + "bbox": [ + 496, + 325, + 890, + 356 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Shape Representations. To capture detailed shapes of 3D objects, most recent papers utilize implicit functions [7, 8, 20, 33-35, 40, 41, 48-50, 56, 57, 65, 66] or point clouds [16, 31, 32] due to their topological flexibility. These methods aim to learn geometry from 3D datasets, whereas we synthesize human images of novel poses only from 2D RGB training images.", + "bbox": [ + 496, + 357, + 890, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rendering Humans by 2D GANs. Some existing approaches render human avatars by neural image translation, i.e., they utilize GAN [10] networks to learn a mapping from poses (given in the form of renderings of a skeleton [4, 23, 45, 52, 59, 68], dense mesh [11, 26, 27, 36, 51, 59] or joint position heatmaps [1, 29, 30]) to human images [4, 59]. To improve temporal stability, some methods [17, 44, 46, 55] propose to utilize the SMPL [28] priors for pose-guided generations. However, these methods do not reconstruct geometry explicitly and cannot handle self-occlusions effectively.", + "bbox": [ + 496, + 463, + 890, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rendering Humans by 3D-aware Renderer. For stable view synthesis, recent papers [6, 25, 38, 42, 43, 53] propose to unify geometry reconstruction with view synthesis by volume rendering, which, however, is computationally heavy. To solve this issue, some recent 3D-GAN papers [5, 12, 15, 18, 19, 37, 39, 67] propose a hybrid rendering strategy for efficient geometry-aware rendering, that is, render low-resolution volumetric features for geometry learning, and employ a super-resolution module for high-resolution image synthesis. We adopt this strategy for efficient rendering, whereas ours is distinguished by rendering articulated humans with 4D motion modeling.", + "bbox": [ + 496, + 628, + 890, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "UV-based Pose Representation. Some methods [18, 19, 31, 32, 47, 63] propose to project 3D posed meshes into a 2D positional map for pose encoding, which can be used for different downstream tasks, such as 3D reconstruction and novel view synthesis. [31, 32] rely on 3D supervision to learn 3D reconstructions, and they do not take as input the", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6551", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dynamics. To improve the rendering quality, [18, 47] proposes to utilize additional driving views as input for faithful rendering in telepresence applications. However, they do not explore how to learn temporal dynamics from pose sequences. By taking as input a 3D normal and velocity map, [63] encodes motion dynamics in their rendering network, whereas they do explicitly learn motions, i.e., predicting the motion status at the next timestep, and besides, [63] is a 2D rendering method, and they do not explicitly learn motions.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 242, + 210, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Setup", + "text_level": 1, + "bbox": [ + 76, + 268, + 227, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a sparse multi-view video of a clothed human in motion and corresponding 3D pose estimations $\\{\\mathbf{P}_0, \\dots, \\mathbf{P}_t\\}$ , our goal is to synthesize time-varying appearances of the individual under novel views. Existing methods [25, 42, 43, 61] formulate the problem of human rendering as learning a representation via a feature encoder-decoder framework:", + "bbox": [ + 75, + 292, + 468, + 383 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\mathbf {t}} = \\mathcal {E} _ {\\mathcal {P}} \\left(\\mathbf {P} _ {\\mathbf {t}}, \\mathbf {z} _ {\\mathbf {t}}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 397, + 468, + 421 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {\\mathbf {t}} = \\mathcal {D} _ {\\mathcal {R}} (\\mathbf {f} _ {\\mathbf {t}})\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 417, + 305, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where a pose encoder $\\mathcal{E}_{\\mathcal{P}}$ takes as input a representation of posed body $\\mathbf{P_t}$ (e.g., 2D or 3D keypoints or 3D body mesh vertices) and a timestamp embedding $\\mathbf{z_t}$ at time $t$ , and outputs intermediate pose-dependent features $\\mathbf{f_t}$ that can be rendered by a decoder $\\mathcal{D}_{\\mathcal{R}}$ to reconstruct the appearance images $\\mathbf{I_t}$ of the corresponding pose at time $t$ . The time embedding $\\mathbf{z_t}$ often serves as a residual that is updated passively in backpropagation by image reconstruction loss.", + "bbox": [ + 75, + 441, + 468, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, there are two issues with the above well-adopted paradigm. First, the appearance of clothed humans undergoes complex geometric transformations induced not only by the static pose $\\mathbf{P}_{\\mathbf{t}}$ but also its dynamics, whereas the modeling of dynamics is ignored in E.q. 1, and the residual $\\mathbf{z}_{\\mathbf{t}}$ cannot expressively model physical dynamics neither. Second, existing methods focus on the per-image reconstruction while ignoring the temporal relations of a motion sequence in training, i.e., sampling input pose $\\mathbf{P}_{\\mathbf{t}}$ and timestamp $\\mathbf{z}_{\\mathbf{t}}$ individually, and supervising image reconstructions frame-by-frame, partly because defining temporal supervision in the 2D image space is challenging, especially for articulated humans that often suffer from misalignment problems due to pose estimation errors.", + "bbox": [ + 75, + 564, + 468, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To solve these issues, we propose a new paradigm for learning view synthesis of dynamic humans from video sequences, which is formulated as:", + "bbox": [ + 75, + 776, + 468, + 821 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\mathbf {t}} = \\mathcal {E} _ {\\mathcal {M}} (\\mathbf {P} _ {\\mathbf {t}}, \\mathbf {D} _ {\\mathbf {t}})\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 835, + 380, + 851 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathbf {P} _ {\\mathbf {t} + 1}}{\\partial \\mathbf {x}}, \\frac {\\partial \\mathbf {P} _ {\\mathbf {t} + 1}}{\\partial \\mathbf {t}} = \\mathcal {D} _ {\\mathcal {M}} (\\mathbf {f} _ {\\mathbf {t}}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 854, + 468, + 883 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {\\mathbf {t}} = \\mathcal {D} _ {\\mathcal {R}} (\\mathbf {f} _ {\\mathbf {t}})\n$$\n", + "text_format": "latex", + "bbox": [ + 256, + 887, + 346, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the input is represented by dynamic motions including a static pose $\\mathbf{P}_{\\mathbf{t}}$ and its physical dynamics $\\mathbf{D}_{\\mathbf{t}}$ at time $t$ , and a motion encoder $\\mathcal{E}_{\\mathcal{M}}$ is employed to encode the inputs into motion-dependent features (Sec. 3.2). Besides, in the training stage, a physical motion decoder $\\mathcal{D}_{\\mathcal{M}}$ is introduced to enforce the learning of spatial and temporal relations by decoding the intermediate feature $\\mathbf{f}_{\\mathbf{t}}$ to predict the spatial derivatives at $t + 1$ physically corresponding to surface normal $\\mathbf{N}_{\\mathbf{t} + 1} = \\frac{\\partial\\mathbf{P}_{\\mathbf{t} + 1}}{\\partial\\mathbf{x}}$ , and temporal derivatives at $t + 1$ corresponding to surface velocity $\\mathbf{V}_{\\mathbf{t} + 1} = \\frac{\\partial\\mathbf{P}_{\\mathbf{t} + 1}}{\\partial\\mathbf{t}}$ (Sec. 3.3). $\\mathbf{f}_{\\mathbf{t}}$ is rendered into human images by a decoder $\\mathcal{D}_{\\mathcal{R}}: \\mathcal{G}_2 \\circ \\mathcal{G}_1$ (Sec. 3.5). The framework is depicted in Fig. 2.", + "bbox": [ + 500, + 90, + 890, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notation. In the following parts, we use $\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{space}}$ to denote the features in the pipeline, i.e., $\\mathbf{f}_{\\mathbf{t}}^{3\\mathbf{D}}$ is a 4D representation that is defined in 3D space with temporal $t$ conditioned, and $\\mathbf{M}_{\\mathbf{t}}^{3\\mathbf{D}}$ is 4D motion input defined in 3D space with $t$ conditioned.", + "bbox": [ + 500, + 275, + 890, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Surface-based 4D Motion Encoding", + "text_level": 1, + "bbox": [ + 500, + 364, + 808, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extracting 4D Motions. We first extract 4D motions from a sequence of time-varying parametric posed body (e.g., SMPL) meshes obtained from training video sequences. We describe the motion $\\mathbf{M}_{\\mathbf{t}}^{3\\mathbf{D}}$ at time $t$ as a static skeleton pose $\\mathbf{P}_{\\mathbf{t}}$ and its dynamics $\\mathbf{D}_{\\mathbf{t}}$ . The dynamics at $t$ are physically determined by the current pose, 3D velocity, and motion trajectory of the past several timesteps, which also contribute to the time-varying appearances of the secondary motion. The pose $\\mathbf{P}_{\\mathbf{t}}$ is represented by the 3D vertices of the posed mesh, and dynamics $\\mathbf{D}_{\\mathbf{t}}$ are parameterized by 1) body surface velocity $\\mathbf{V}_{\\mathbf{t}}$ corresponding to the temporal derivatives of the current pose at $t$ , and 2) motion trajectory $\\mathbf{T}_{\\mathbf{t}}$ that aggregates the temporal derivatives over the past several timesteps with a sliding window size of $w$ with weight $\\lambda$ :", + "bbox": [ + 498, + 388, + 890, + 601 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {M} _ {\\mathbf {t}} ^ {\\mathrm {3 D}} = \\left[ \\mathbf {P} _ {\\mathbf {t}}, \\mathbf {D} _ {\\mathbf {t}} \\right], \\quad \\mathbf {D} _ {\\mathbf {t}} = \\left[ \\mathbf {V} _ {\\mathbf {t}}, \\mathbf {T} _ {\\mathbf {t}} \\right] \\\\ \\mathbf {V} _ {\\mathbf {t}} = \\frac {\\partial \\mathbf {P} _ {\\mathbf {t}}}{\\partial \\mathbf {t}}, \\quad \\mathbf {T} _ {\\mathbf {t}} = \\mathbf {P} _ {\\mathbf {t}} + \\frac {1}{\\sum \\lambda_ {i}} \\sum_ {i = 1} ^ {w} \\mathbf {V} _ {\\mathbf {t} - \\mathbf {i}} * \\lambda_ {i} \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 613, + 890, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note the motion trajectory aggregated from several consecutive timesteps makes the motion representation robust to pose estimation errors, i.e., the pose estimations for two consecutive timesteps may be the same due to pose estimation errors in practice. An ablation study of the motion trajectory representation can be found in the supp. mat.", + "bbox": [ + 498, + 688, + 890, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recording 4D Motions on the Surface Manifold. Modeling the motions with temporal dynamics often requires dense observations to construct a 4D motion volume. Instead, we notice that non-rigid deformations of human geometry in motion typically occur around the body surface instead of a 3D volume, and hence we propose to model the motions on the human body surface. Specifically, we project the 4D motion input $\\mathbf{M}_{\\mathbf{t}}^{3\\mathrm{D}}$ including spatial pose", + "bbox": [ + 498, + 779, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6552", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5a621c6cf918e0f0a8ca49a4ed43b02af7ef62e02595360965c875188ed7d426.jpg", + "image_caption": [ + "Figure 2. Framework overview. Given a set of time-varying 3D body meshes $\\{\\mathbf{P}_{\\mathbf{t}},\\dots,\\mathbf{P}_{\\mathbf{t}} - \\mathbf{n}\\}$ obtained from training video sequences, we aim to synthesize high-fidelity appearances of a clothed human in motion via a feature encoder-decoder framework: Motion Encoding, and joint Motion and Appearance Decoding. 1) We take as input an expressive 4D motion representation extracted from the mesh sequences including 3D pose, 3D velocity at time $t$ , and motion trajectory over the past $w$ timesteps that encode both spatial and temporal relations of the motion sequence, which are projected to the spatially aligned UV surface space. A motion encoder $\\mathcal{E}_{\\mathcal{M}}$ is employed to lift the 2D UV-aligned features to a 3D surface-based triplane $\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}$ in an UV-plus-height space with a signed distance height to model temporal clothing offsets. 2) A motion decoder $\\mathcal{D}_{\\mathcal{M}}$ is designed to encourage physical motion learning in training by decoding the triplane features $\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}$ to predict the motion at the next timestep $t + 1$ , i.e. spatial derivatives surface normal $\\mathbf{N}_{\\mathbf{t} + 1}^{\\mathrm{uv}}$ and temporal derivatives surface velocity $\\mathbf{V}_{\\mathbf{t} + 1}^{\\mathrm{uv}}$ in UV space. 3) Finally, given a target camera view, the triplane $\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}$ is rendered into high-quality images by a volumetric surface-conditioned renderer including volumetric low-resolution rendering by $\\mathcal{G}_1$ and an efficient geometry-aware super-resolution by $\\mathcal{G}_2$ ." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 893, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and temporal dynamics from 3D space into a compact spatially aligned UV space using the geometric transformation $\\mathcal{W}$ that is pre-defined by the parametric (e.g., SMPL) body template, which yields $\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uv}} = \\mathcal{W}\\mathbf{M}_{\\mathbf{t}}^{3\\mathrm{D}}$ . The UV-aligned motion feature $\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uv}}$ faithfully preserves the articulated structures of body topology in a compact 2D space.", + "bbox": [ + 75, + 420, + 468, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generating Surface-based Triplanes for Motion Modeling. To model clothed humans, we further employ a motion encoder $\\mathcal{E}_{\\mathcal{M}}$ that lifts the clothless 2D features $\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uv}}$ to a 3D triplane representation $\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}$ that is defined in a $\\mathbf{u} - \\mathbf{v} - \\mathbf{h}$ system, with $\\mathbf{u} - \\mathbf{v}$ to represent the motion of the clothless body template, and an extensional coordinate $\\mathbf{h}$ to represent the secondary motion of clothes, i.e., the temporal cloth offsets, and hence 4D clothed motions can be parameterized by a surface-based triplane by:", + "bbox": [ + 75, + 511, + 470, + 647 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathrm {u v h}} = \\mathcal {E} _ {\\mathcal {M}} \\left(\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathrm {u v}}\\right) = \\mathcal {E} _ {\\mathcal {M}} \\left(\\mathcal {W} \\mathbf {M} _ {\\mathbf {t}} ^ {\\mathbf {3 D}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 657, + 468, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{W}$ is the geometric transformation from 3D space to UV space. $\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}$ consists of three planes $x_{uv}, x_{uh}, x_{hv} \\in \\mathbb{R}^{U \\times V \\times C}$ to form the spatial relationship, where $U, V$ denote spatial resolution and $C$ is the channel number. In contrast to the volumetric triplanes used in EG3D [5] where the three planes represent three vertical planes in the 3D volumetric space, $\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}$ is defined on the human body surface inheriting human topology priors for motion modeling, as illustrated by an unwarped triplane in 3D space in Fig 2.", + "bbox": [ + 75, + 686, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Physical Motion Decoding", + "text_level": 1, + "bbox": [ + 76, + 830, + 316, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a physical motion decoder $\\mathcal{D}_{\\mathcal{M}}$ to learn spatial and temporal features of motion, which is achieved by decoding the intermediate motion feature $\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}$ learned by $\\mathcal{E}_{\\mathcal{M}}$", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to predict the motions at the next timestep $t + 1$ , such as spatial derivatives corresponding to surface normal $\\mathbf{N}_{t+1}^{\\mathrm{uv}}$ , temporal derivatives corresponding to surface velocity $\\mathbf{V}_{t+1}^{\\mathrm{uv}}$ . Note that we model the normal and velocity in the UV space by:", + "bbox": [ + 496, + 420, + 890, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathbf {N} _ {\\mathbf {t} + \\mathbf {1}} ^ {\\mathrm {u v}}, \\mathbf {V} _ {\\mathbf {t} + \\mathbf {1}} ^ {\\mathrm {u v}} \\right\\} = \\mathcal {D} _ {\\mathcal {M}} \\left(\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathrm {u v h}}\\right) = \\mathcal {D} _ {\\mathcal {M}} \\circ \\mathcal {E} _ {\\mathcal {M}} \\left(\\mathcal {W} \\mathbf {M} _ {\\mathbf {t}} ^ {\\mathbf {3 D}}\\right) (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 501, + 890, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. 4D Appearance Decoding", + "text_level": 1, + "bbox": [ + 500, + 534, + 732, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Volumetric Surface-conditioned Rendering. Given a target camera viewpoint, the triplane $\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}$ is first rendered into low-resolution volumetric features $\\mathbf{I}_{\\mathbf{F}} \\in \\mathbb{R}^{H \\times W \\times \\hat{C}}$ by a conditional NeRF $\\mathbb{F}_{\\Phi}$ , where $H$ and $W$ are the spatial resolution and $\\hat{C}$ is the channel number. To be more specific, given a 3D query point $\\pmb{p}_i$ , we transform it into a surface-based local coordinate $\\hat{\\pmb{p}}_i = (\\pmb{u}_i, \\pmb{v}_i, \\pmb{h}_i)$ w.r.t. the tracked body mesh $\\mathbf{P}_{\\mathbf{t}}$ . Here we search the nearest face $\\pmb{f}_i$ of $\\mathbf{P}_{\\mathbf{t}}$ for query point $\\pmb{p}_i$ and $(\\pmb{u}_i, \\pmb{v}_i)$ and $\\pmb{h}_i$ are the barycentric coordinates of the nearest point on $\\pmb{f}_i$ , and the signed distance respectively. We therefore obtain the local feature $\\pmb{z}_i^{\\mathrm{uvh}}$ of the query point $\\pmb{p}_i$ as", + "bbox": [ + 496, + 556, + 890, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {i} ^ {u v h} = \\operatorname {c a t} \\left[ \\Pi \\left(\\boldsymbol {x} _ {u v}; \\boldsymbol {u} _ {i}, \\boldsymbol {v} _ {i}\\right), \\Pi \\left(\\boldsymbol {x} _ {u h}; \\boldsymbol {u} _ {i}, \\boldsymbol {h} _ {i}\\right), \\Pi \\left(\\boldsymbol {x} _ {h v}; \\boldsymbol {h} _ {i}, \\boldsymbol {v} _ {i}\\right) \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 747, + 901, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Pi (\\cdot)$ denotes sampling operation, $\\mathrm{cat}[\\cdot ]$ denotes the concatenation operator.", + "bbox": [ + 496, + 779, + 890, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given camera direction $\\pmb{d}_i$ , the appearance features $\\pmb{c}_i$ and density features $\\pmb{\\sigma}_i$ of point $\\pmb{p}_i$ are predicted by", + "bbox": [ + 500, + 809, + 888, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\boldsymbol {c} _ {i}, \\boldsymbol {\\sigma} _ {i} \\right\\} = \\mathbb {F} _ {\\Phi} \\left(\\boldsymbol {d} _ {i}, \\boldsymbol {z} _ {i} ^ {u v h}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 845, + 890, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We integrate all the radiance features of sampled points into a 2D feature map $I_{F}$ at time $t$ through volume renderer", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6553", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/88c4ec5eef522c9e383384ba91eea32c81698ccdf890a3fe5474f390535bf409.jpg", + "image_caption": [ + "Figure 3. Qualitative comparisons on novel view synthesis on the subject S313 of ZJU-MoCap dataset. Two motion sequences S1 (swing arms left to right) and S2 (raise and lower arms) are shown. We specifically focus on the synthesis of time-varying appearances (especially T-shirt wrinkles), by evaluating the rendering results under similar poses yet with different movement directions, which are marked in the same color, such as the pairs of $①②$ , $③④$ , and $⑤⑥$ . Our method synthesizes high-fidelity time-varying appearances, whereas SOTA HumanNeRF generates almost the same cloth wrinkles." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{G}_1$ [21]", + "bbox": [ + 76, + 662, + 130, + 678 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {\\mathbf {F}} = \\mathcal {G} _ {1} \\left(\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {u v h}}, c a m; \\mathbb {F} _ {\\Phi}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 676, + 468, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_{\\mathrm{t}}^{\\mathrm{uvh}}$ is extracted from Eq. 4.", + "bbox": [ + 76, + 700, + 316, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Efficient Geometry-Aware Super-Resolution. Sampling dense points to render the full-resolution volumetric features is computationally heavy. Instead, we employ a superresolution network $\\mathcal{G}_2$ [5] to render high-resolution images $\\mathbf{I}_{\\mathbf{RGB}}^{+} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times 3}$ :", + "bbox": [ + 75, + 715, + 468, + 795 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {\\mathbf {R G B}} ^ {+} = \\mathcal {G} _ {2} \\circ \\mathcal {G} _ {1} \\left(\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {u v h}}, c a m; \\mathbb {F} _ {\\Phi}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 806, + 468, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Optimization", + "text_level": 1, + "bbox": [ + 76, + 832, + 215, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SurMo is trained end-to-end to optimize $\\mathcal{E}_{\\mathcal{M}}$ , $\\mathcal{D}_{\\mathcal{M}}$ , and renderers $\\mathcal{G}_1$ , $\\mathcal{G}_1$ with 2D image loss. We employ Adversarial Loss and Reconstruction Loss including Pixel Loss,", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Perceptual Loss, Face Identity Loss, Velocity and Normal Loss, and Volume Rendering Loss for supervision in training, with Adam [22] as the optimizer. Refer to the supp. mat. for more details.", + "bbox": [ + 498, + 662, + 890, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 738, + 633, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset and Metrics. We evaluate the novel view synthesis on ZJU-MoCap, MPII-RDDC, and AIST++, with a total of 9 subjects. We use the same camera setup as Neural Body, that is, 4 cameras used for training, and the other 18 for testing. For MPII-RDDC, 18 cameras in training, and 9 for testing. Since ASIT++ only has 9 cameras, we use 6 for training, and the remaining 3 for testing. Refer to more details in the supp. mat. We compare each method on per-pixel metrics including SSIM [60] and PSNR, and", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6554", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0d2e47398f1fb0efc17a0c34bcfbc14ef3610b57d9ac2bff4211c7c03e472755.jpg", + "image_caption": [ + "Neural Body Instant-NVR HumanNeRF Ours GT" + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 486, + 314 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4fb3ce63b4f819b49edb8d72bdfc6828627c391d5ad7c8f53e2ede247011c197.jpg", + "image_caption": [ + "Neural Body Instant-NVR HumanNeRF Ours GT", + "Figure 4. Qualitative comparisons on novel view synthesis on the subject S387, S315 of ZJU-MoCap dataset. Row 1 and 2 show similar poses occurring at different timesteps (not consecutive frames). The results indicate that our method synthesizes time-varying appearances while other methods mainly generate pose-dependent appearances." + ], + "image_footnote": [], + "bbox": [ + 488, + 90, + 890, + 314 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cb31086a7d8f39e001be6c08a9026e21a602c56e1511fa44554197d12abc5c78.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on ZJU-MoCap dataset (averaged on all test views and poses on 6 sequences) for novel-view synthesis. To reduce the influence of the background, all scores are calculated from images cropped to 2D bounding boxes. Note that the perception metrics LPIPS [64] and FID [14] capture human judgment better than per-pixel metrics such as SSIM [60] or PSNR, as stated in [25, 51]." + ], + "table_footnote": [], + "table_body": "
ZJU-S1-6LPIPS ↓FID ↓SSIM ↑PSNR ↑
Neural Body [43].164125.10.70321.438
Instant-NVR [9].202144.99.73821.748
HumanNeRF [61].10688.382.79223.624
Ours.07567.725.83324.815
", + "bbox": [ + 78, + 513, + 468, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "perception metrics including LPIPS [64] and FID [14].", + "bbox": [ + 75, + 625, + 439, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method against SOTA methods including Neural Body [43], HumanNeRF [61], Instant-NVR [9], ARAH [58], DVA [47] and HVTR++ [18]. Neural Body models human poses in 3D space with point cloud as pose representation, HumanNeRF and Instant-NVR model poses in a canonical space with inverse skinning, ARAH is a forward-skinning based method, DVA and HVTR++ encode poses in UV space, and take driving view signals as input for telepresence applications.", + "bbox": [ + 75, + 640, + 468, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparisons to SOTA Methods", + "text_level": 1, + "bbox": [ + 76, + 786, + 356, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparisons We conduct the quantitative comparisons on the three datasets with a total of 9 subjects. The quantitative results on ZJU-MoCap are summarized in Tab. 1, which suggests that our new paradigm outperforms the SOTA by a big margin, and we achieve the best quantitative results on all four metrics. The detailed compar-", + "bbox": [ + 75, + 809, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "isons on each subject of ZJU-MoCap, and quantitative results on MPII-RDDC and AIST++ are listed in the supp. mat. Refer to the comparisons with ARAH [58], DVA [47] and HVTR++ [18] in the supp. mat.", + "bbox": [ + 496, + 407, + 890, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Time-varying Appearances with Dynamics Conditioning on ZJU-MoCap. We compare with baseline methods for novel view synthesis on two motion sequences of subject S313 from the ZJU-MoCap dataset, as shown in Fig. 3. We evaluate the capability of each method to synthesize time-varying appearances. Specifically, two sequences S1 (swing arms left to right) and S2 (raise and lower arms) are evaluated, where similar poses occur at different timesteps with different dynamics (e.g., movement directions, trajectory, or velocity) are marked in the same color, such as the pairs of $①②$ , $③④$ , and $⑤⑥$ . Fig. 3 suggests that Instant-NVR [9] fails to synthesize time-varying high-frequency wrinkles, and Neural Body [43] synthesizes time-varying yet blurry wrinkles. HumanNeRF [61] synthesizes detailed yet static T-shirt wrinkles, i.e., the wrinkles are almost the same for similar poses, such as $①②$ , $③④$ , and $⑤⑥$ . In contrast, our method renders both high-fidelity and time-varying wrinkles. The comparisons on other subjects S387 and S315 are shown in Fig. 4, which illustrate the effectiveness of our proposed paradigm in dynamics learning.", + "bbox": [ + 496, + 472, + 892, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Time-varying Appearances with Motion-dependent Shadows on MPII-RDDC. We compare with baseline methods for novel view synthesis on MPII-RDDC, as shown in Fig. 5. The sequence is captured in a studio with top-down lighting that casts shadows on the human body due to self-occlusions. We notice that the synthesis of shadow can be formulated as learning motion-dependent shadows in the motion representation without the need to", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6555", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bcd6ca47faffd07096384f8eaac4a158e38e6a378315e2a759538a443c396b69.jpg", + "image_caption": [ + "HumanNeRF", + "Ours", + "Ground Truth", + "HumanNeRF", + "Ours", + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 893, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e3ed2af846bde22715ff3cada8a05944270991c71e47028fd8f6166e92449a9.jpg", + "image_caption": [ + "Figure 5. Novel view synthesis of time-varying appearances with both pose and lighting conditioning on MPII-RDDC dataset. The sequence is captured in a studio with top-down lighting that casts shadows on the human performer due to self-occlusion. In Row 1, we specifically focus on synthesizing time-varying shadows (e.g., ① vs. ②, and ③ vs. ④) for different poses with different self-occlusions. In Row 2, we evaluate the synthesis of: 1) time-varying appearances for similar poses occurring in a jump-up-and-down motion sequence, e.g., ⑤ vs. ⑥, 2) shadows ⑦ vs. ⑧, and 3) clothing offsets ⑤ vs. ⑥.", + "Figure 6. Novel view synthesis of fast motions on AIST++." + ], + "image_footnote": [], + "bbox": [ + 78, + 406, + 468, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "explicitly model the lighting. We validate our method from two aspects of the scenario in Fig. 5. 1) In Row 1, Fig. 5 indicates that with the expressive motion representation and physical motion learning, SurMo succeeds in predicting the motion-dependent shadows, such as $①②③④$ , whereas SOTA HumanNeRF renders almost the same T-shirt appearance. 2) In Row 2, we compare the synthesis of time-varying appearances for similar poses occurring in a jump-up-and-down motion sequence i.e., $⑤$ vs. $⑥$ , which shows that SurMo is capable of predicting the clothes offsets under similar pose with different motion trajectories. In addition, we synthesize the shadows $⑦⑧$ in the challenging jump-up-and-down motion sequence. However, HumanNeRF fails to predict the dynamic clothing offsets $⑨$ .", + "bbox": [ + 75, + 536, + 468, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Time-varying Appearances for Fast Motions on AIST++. We also evaluate our methods by rendering humans with fast dance motions (S21 and S13) on AIST++, as shown in Fig. 6, where we compare ours against Neural Body in synthesizing time-varying wrinkles of the T-shirt for different motions. Fig. 6 suggests that Neural Body can only synthesize blurry results partly because Neural Body models pose in a sparse 3D space. Instead, we model motions in a denser body surface space, which enables high-fidelity image synthesis. Note that AIST++ is based", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06d2a2cc665354948ad640d0d01262dfd556dcde6bbbfd86cc42038b4b418641.jpg", + "image_caption": [ + "Figure 8. Ablation study of motion conditioning and learning. We focus on the effect of motion conditioning and learning in Row 1, and whether our method learns to decouple the static pose and dynamics (e.g., velocity) from the motion conditioning." + ], + "image_footnote": [], + "bbox": [ + 501, + 407, + 890, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on SMPL tracking with a scaling factor that affects the inverse LBS used in Instant-NVR or HumanNeRF both relying on inverse LBS, and hence we cannot compare them based on the released official code. Quantitative results can be found in the supp. mat.", + "bbox": [ + 496, + 648, + 890, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 739, + 653, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Surface-based Triplane vs. Volumetric Triplane. We compare the well-adopted volumetric triplane (Vol-Trip) [5] and our proposed surface-based triplane (Surf-Trip) for human modeling on two subjects S313 and S387 from ZJUMoCap, as shown in Fig. 7. Fig. 7 (left) suggests that the Surf-Trip converges faster in training with a smaller reconstruction loss for both sequences, and the performances are further improved with physical motion learning, e.g., 'Surf-Trip + ML'. Fig. 7 (right) compares the rendering re", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6556", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/93fc99a370db7c1e669130b0279e6ed2c28e9220a804220058633a7d8c32d74a.jpg", + "image_caption": [ + "Figure 7. Ablation study of 3D volumetric triplane (Vol-Trip) vs. surface-based triplane (Surf-Trip) for human modeling on subject S313 and S387 from ZJU-MoCap. We focus on the convergence in training (left), and self-occlusion rendering (right). The convergence of S313 is shifted for visualization purposes. Vol-Trip is not effective in handling self-occlusion, e.g., $①②③④$ though it performs well in another viewpoint without self-occlusions $⑥$ . In addition, Vol-Trip cannot synthesize high-quality details for face." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 374, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2d8e22cbf2bb7793f4aa369875a8c53f83a80eda8f3684077af6967db514c0d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 89, + 887, + 277 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/15736cc291c8f4350878bac458b91f024ede26c8902c7001bbef262f6d1fc0fa.jpg", + "table_caption": [ + "Table 2. Ablation study of motion conditioning and learning. $P_{cond}$ and $D_{cond}$ denote the conditioning of pose and dynamics, $V_{pred}$ and $N_{pred}$ denote the prediction of surface velocity and surface normal in motion learning." + ], + "table_footnote": [], + "table_body": "
S313LPIPS ↓FID ↓SSIM ↑PSNR ↑
Pcond.120104.83.79322.781
+ Dcond.08573.674.83424.908
+ Vpred.06962.092.85625.845
+ Npred.06050.170.86926.654
", + "bbox": [ + 101, + 416, + 441, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sults qualitatively, which indicates that Surf-Trip is more effective in handling self-occlusions, whereas Vol-Trip fails $①②③④$ , though it performs well in another viewpoint without self-occlusions $⑥$ . In addition, Surf-Trip generates high-fidelity clothing wrinkles and facial details, whereas the face rendering of Vol-Trip is blurry.", + "bbox": [ + 75, + 532, + 468, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motion Conditioning and Learning. We also analyze the effectiveness of our proposed motion conditioning and learning qualitatively and quantitatively. Tab. 2 summarizes the quantitative results, which suggest that the quantitative results are significantly improved by taking as input the dynamics conditioning $D_{cond}$ . With physical motion learning, i.e., predicting the temporal derivatives surface velocity and spatial derivates normal at the next timestep, the quantitative results are further improved by a big margin.", + "bbox": [ + 75, + 625, + 468, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The qualitative comparisons are shown in Fig. 8, where in Row 1, we observe higher-fidelity appearances with dynamics conditioning, and physical motion learning for both velocity and normal prediction. In Row 2, we evaluate whether our method learns to decouple the temporal dynamics (e.g., velocity) and static poses from motion conditioning, which corresponds to the rendering of clothing wrinkles of secondary motion (T-shirt ①) and tight parts (e.g., shorts ②). We only change the velocity for each variant", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in the ablation study. It suggests that the appearance of the T-shirt varies when the velocity or dynamics are changed, whereas the renderings of tight parts remain the same, such as the head, tight shorts, and shoes. This is consistent with our daily observations. The ablation study illustrates that our method decouples the pose and dynamics, and is capable of generating dynamics-dependent appearances. Note that the appearance changes are small between $V * 1$ and $V * 0.01$ when we reduce the velocity, partly because the method is not sensitive to smaller velocity due to the pose estimation errors in the training dataset where consecutive frames may be estimated with the same pose fitting.", + "bbox": [ + 496, + 361, + 893, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 500, + 570, + 612, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose SurMo, a new paradigm for learning dynamic humans from videos by jointly modeling temporal motions and human appearances in a unified framework, based on an efficient surface-based triplane. We conduct a systematical analysis of how human appearances are affected by temporal dynamics, and extensive experiments validate the expressiveness of the surface-based triplane in rendering fast motions and motion-dependent shadows. Quantitative experiments illustrate that SurMo achieves the SOTA results on different motion sequences.", + "bbox": [ + 496, + 599, + 893, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 779, + 660, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6557", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kfir Aberman, M. Shi, Jing Liao, Dani Lischinski, B. Chen, and D. Cohen-Or. Deep video-based performance cloning. Computer Graphics Forum, 38, 2019. 2", + "[2] George Borshukov, Dan Piponi, Oystein Larsen, J. P. Lewis, and Christina Tempelaar-Lietz. Universal capture: image-based facial animation for \"the matrix reloaded\". In SIGGRAPH '03, 2003. 1", + "[3] Joel Carranza, Christian Theobalt, Marcus A. Magnor, and Hans-Peter Seidel. Free-viewpoint video of human actors. ACM SIGGRAPH 2003 Papers, 2003. 1", + "[4] Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A. Efros. Everybody dance now. ICCV, pages 5932-5941, 2019. 2", + "[5] Eric Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, S. Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. ArXiv, abs/2112.07945, 2021. 2, 4, 5, 7", + "[6] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Linchao Bao, and Huchuan Lu. Animatable neural radiance fields from monocular rgb video. ArXiv, abs/2106.13629, 2021. 2", + "[7] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. CVPR, 2019. 2", + "[8] Boyang Deng, John P Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. Nasa neural articulated shape approximation. In ECCV, 2020. 2", + "[9] Chen Geng, Sida Peng, Zhen Xu, Hujun Bao, and Xiaowei Zhou. Learning neural volumetric representations of dynamic humans in minutes. In CVPR, 2023. 6", + "[10] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014. 2", + "[11] A. K. Grigor'ev, Artem Sevastopolsky, Alexander Vakhitov, and Victor S. Lempitsky. Coordinate-based texture inpainting for pose-guided human image generation. CVPR, pages 12127-12136, 2019. 2", + "[12] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. ArXiv, abs/2110.08985, 2021. 2", + "[13] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics (TOG), 40:1-16, 2021. 2", + "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, 2017. 6", + "[15] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. ArXiv, abs/2112.05637, 2021. 2", + "[16] Tao Hu, Geng Lin, Zhizhong Han, and Matthias Zwicker. Learning to generate dense point clouds with textures on multiple categories. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)," + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pages 2170-2179, January 2021. 2", + "[17] Tao Hu, Kripasindhu Sarkar, Lingjie Liu, Matthias Zwicker, and Christian Theobalt. Egorenderer: Rendering human avatars from egocentric camera images. In ICCV, 2021. 2", + "[18] Tao Hu, Hongyi Xu, Linjie Luo, Tao Yu, Zerong Zheng, He Zhang, Yebin Liu, and Matthias Zwicker. Hvtr++: Image and pose driven human avatars using hybrid volumetric-textural rendering. IEEE Transactions on Visualization and Computer Graphics, pages 1-15, 2023. 2, 3, 6", + "[19] T. Hu, Tao Yu, Zerong Zheng, He Zhang, Yebin Liu, and Matthias Zwicker. Hvtr: Hybrid volumetric-textural rendering for human avatars. 3DV, 2022. 2", + "[20] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. 2020 (CVPR), pages 3090-3099, 2020. 2", + "[21] James T. Kajiya and Brian Von Herzen. Ray tracing volume densities. Proceedings of the 11th annual conference on Computer graphics and interactive techniques, 1984. 5", + "[22] Diederick P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5", + "[23] Bernhard Kratzwald, Zhiwu Huang, Danda Pani Paudel, and Luc Van Gool. Towards an understanding of our world by GANing videos in the wild. arXiv:1711.11453, 2017. 2", + "[24] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Learn to dance with aist++: Music conditioned 3d dance generation, 2021. 2", + "[25] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. TOG, 40, 2021. 1, 2, 3, 6", + "[26] Lingjie Liu, Weipeng Xu, Marc Habermann, Michael Zollhöfer, Florian Bernard, Hyeongwoo Kim, Wenping Wang, and Christian Theobalt. Neural human video rendering by learning dynamic textures and rendering-to-video translation. IEEE Transactions on Visualization and Computer Graphics, 05 2020. 2", + "[27] Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Hyeongwoo Kim, Florian Bernard, Marc Habermann, Wenping Wang, and Christian Theobalt. Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG), 2019. 2", + "[28] M. Loper, Naureen Mahmood, J. Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multi-person linear model. ACM Trans. Graph., 34:248:1-16, 2015. 2", + "[29] Liqian Ma, Xu Jia, Qianru Sun, Bernt Schiele, Tinne Tuytelaars, and Luc Van Gool. Pose guided person image generation. In NeurIPS, pages 405-415, 2017. 2", + "[30] Liqian Ma, Qianru Sun, Stamatios Georgoulis, Luc van Gool, Bernt Schiele, and Mario Fritz. Disentangled person image generation. CVPR, 2018. 2", + "[31] Qianli Ma, Shunsuke Saito, Jinlong Yang, Siyu Tang, and Michael J. Black. Scale: Modeling clothed humans with a surface codec of articulated local elements. In CVPR, 2021. 2", + "[32] Qianli Ma, Jinlong Yang, Siyu Tang, and Michael J Black. The power of points for modeling humans in clothing. In ICCV, 2021. 2", + "[33] Lars M. Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy net" + ], + "bbox": [ + 503, + 92, + 890, + 887 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6558", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "works: Learning 3d reconstruction in function space. CVPR, 2019. 2", + "[34] Mateusz Michalkiewicz, Zhony Kaesemodel Pontes, Dominic Jack, Mahsa Baktash, and Anders P. Eriksson. Deep level sets: Implicit surface representations for 3d shape inference. ArXiv, 2019.", + "[35] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. Leap: Learning articulated occupancy of people. In CVPR, 2021. 2", + "[36] Natalia Neverova, Riza Alp Güler, and Iasonas Kokkinos. Dense pose transfer. ECCV, 2018. 2", + "[37] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. CVPR, pages 11448-11459, 2021. 2", + "[38] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In IEEE/CVF ICCV, 2021. 2", + "[39] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. ArXiv, abs/2112.11427, 2021. 2", + "[40] Pablo Palafox, Aljaž Božić, Justus Thies, Matthias Nießner, and Angela Dai. Npms: Neural parametric models for 3d deformable shapes. In IEEE/CVF ICCV, 2021. 2", + "[41] Jeong Joon Park, Peter R. Florence, Julian Straub, Richard A. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. 2019 (CVPR), pages 165-174, 2019. 2", + "[42] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021, 2, 3", + "[43] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. CVPR, 2021. 1, 2, 3, 6", + "[44] Sergey Prokudin, Michael J. Black, and Javier Romero. Smplpix: Neural avatars from 3d human models. WACV, 2021. 2", + "[45] Albert Pumarola, Antonio Agudo, Alberto Sanfeliu, and Francesc Moreno-Noguer. Unsupervised person image synthesis in arbitrary poses. In CVPR, June 2018. 2", + "[46] Amit Raj, Julian Tanke, James Hays, Minh Vo, Carsten Stoll, and Christoph Lassner. Anr: Articulated neural rendering for virtual avatars. CVPR, pages 3721-3730, 2021. 2", + "[47] Edoardo Remelli, Timur M. Bagautdinov, Shunsuke Saito, Chenglei Wu, Tomas Simon, Shih-En Wei, Kaiwen Guo, Zhe Cao, Fabian Prada, Jason M. Saragih, and Yaser Sheikh. Drivable volumetric avatars using texel-aligned features. ACM SIGGRAPH, 2022. 2, 3, 6", + "[48] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. IEEE/CVF ICCV, pages 2304-2314, 2019. 2", + "[49] Shunsuke Saito, Tomas Simon, Jason M. Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. 2020 (CVPR), pages 81-90, 2020." + ], + "bbox": [ + 78, + 90, + 468, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J. Black. Scintimate: Weakly supervised learning of skinned clothed avatar networks. 2021 (CVPR), pages 2885-2896, 2021. 2", + "[51] Kripasindhu Sarkar, Dushyant Mehta, Weipeng Xu, Vladislav Golyanik, and Christian Theobalt. Neural re- rendering of humans from a single image. In ECCV, 2020. 2,6", + "[52] Aliaksandr Siarohin, Enver Sangineto, Stephane Lathuiliere, and Nicu Sebe. Deformable GANs for pose-based human image generation. In CVPR, 2018. 2", + "[53] Shih-Yang Su, Frank Yu, Michael Zollhoefer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In NeurIPS, 2021. 2", + "[54] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo MartinBrualla, Tomas Simon, Jason M. Saragih, Matthias Nießner, Rohit Pandey, S. Fanello, Gordon Wetzstein, Jun-Yan Zhu, Christian Theobalt, Maneesh Agrawala, Eli Shechtman, Dan B. Goldman, and Michael Zollhofer. State of the art on neural rendering. Computer Graphics Forum, 2020. 1", + "[55] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: image synthesis using neural textures. ACM Transactions on Graphics (TOG), 38, 2019. 2", + "[56] Garvita Tiwari, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Neural-gif: Neural generalized implicit functions for animating people in clothing. In ICCV, 2021. 2", + "[57] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Metaatrix: Learning animatable clothed human models from few depth images. NeurIPS, 2021. 2", + "[58] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In European Conference on Computer Vision, 2022. 6", + "[59] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In NeurIPS, 2018. 2", + "[60] Zhou Wang, A. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13:600-612, 2004. 5, 6", + "[61] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humaner: Free-viewpoint rendering of moving people from monocular video. ArXiv, abs/2201.04127, 2022. 1, 2, 3, 6", + "[62] Feng Xu, Yebin Liu, Carsten Stoll, James Tompkin, Gaurav Bharaj, Qionghai Dai, Hans-Peter Seidel, Jan Kautz, and Christian Theobalt. Video-based characters: creating new human performances from a multi-view video database. ACM SIGGRAPH, 2011. 1", + "[63] Jae Shin Yoon, Duygu Ceylan, Tuanfeng Y. Wang, Jingwan Lu, Jimei Yang, Zhixin Shu, and Hyunjung Park. Learning motion-dependent appearance for high-fidelity rendering of dynamic humans from a single camera. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3397-3407, 2022. 2, 3", + "[64] Richard Zhang, Phillip Isola, Alexei A. Efros, E. Shechtman," + ], + "bbox": [ + 501, + 92, + 890, + 890 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6559", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. CVPR, pages 586-595, 2018. 6", + "[65] Yang Zheng, Ruizhi Shao, Yuxiang Zhang, Tao Yu, Zerong Zheng, Qionghai Dai, and Yebin Liu. Deepmulticap: Performance capture of multiple characters using sparse multiview cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6239-6249, 2021. 2", + "[66] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. TPAMI, PP, 2021. 2", + "[67] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. Cips3d: A 3d-aware generator of gans based on conditionally-independent pixel synthesis. ArXiv, 2021. 2", + "[68] Zhen Zhu, Tengteng Huang, Baoguang Shi, Miao Yu, Bofei Wang, and Xiang Bai. Progressive pose attention transfer for person image generation. In CVPR, pages 2347-2356, 2019. 2" + ], + "bbox": [ + 78, + 90, + 467, + 359 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6560", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/6283bde0-3818-458b-8029-7eb693822789_model.json b/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/6283bde0-3818-458b-8029-7eb693822789_model.json new file mode 100644 index 0000000000000000000000000000000000000000..71bf9606b225e98f30dc867770d07d52a056327c --- /dev/null +++ b/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/6283bde0-3818-458b-8029-7eb693822789_model.json @@ -0,0 +1,2323 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.131, + 0.872, + 0.153 + ], + "angle": 0, + "content": "SurMo: Surface-based 4D Motion Modeling for Dynamic Human Rendering" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.169, + 0.696, + 0.205 + ], + "angle": 0, + "content": "Tao Hu, Fangzhou Hong, Ziwei Liu \nS-Lab, Nanyang Technological University, Singapore" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.242, + 0.314, + 0.258 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.267, + 0.474, + 0.691 + ], + "angle": 0, + "content": "Dynamic human rendering from video sequences has achieved remarkable progress by formulating the rendering as a mapping from static poses to human images. However, existing methods focus on the human appearance reconstruction of every single frame while the temporal motion relations are not fully explored. In this paper, we propose a new 4D motion modeling paradigm, SurMo, that jointly models the temporal dynamics and human appearances in a unified framework with three key designs: 1) Surface-based motion encoding that models 4D human motions with an efficient compact surface-based triplane. It encodes both spatial and temporal motion relations on the dense surface manifold of a statistical body template, which inherits body topology priors for generalizable novel view synthesis with sparse training observations. 2) Physical motion decoding that is designed to encourage physical motion learning by decoding the motion triplane features at timestep \\( t \\) to predict both spatial derivatives and temporal derivatives at the next timestep \\( t + 1 \\) in the training stage. 3) 4D appearance decoding that renders the motion triplanes into images by an efficient volumetric surface-conditioned renderer that focuses on the rendering of body surfaces with motion learning conditioning. Extensive experiments validate the state-of-the-art performance of our new paradigm and illustrate the expressiveness of surface-based motion triplanes for rendering high-fidelity view-consistent humans with fast motions and even motion-dependent shadows. Our project page is at: https://taohuumd.github.io/projects/SurMo." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.71, + 0.21, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Creating volumetric videos of human actors is required in many applications such as AR/VR, telepresence, video game and film character creation. Recent neural rendering methods [25, 43, 54, 61] have made great progress in generating free-viewpoint videos of humans from several sparse multi-view videos, which are simple yet effective compared with traditional graphics approaches [2, 3, 62]. At the core of the technique is to model pose- and time-varying appearances of dynamic humans. The motions of dressed humans are often expressed as the movement of body and a sequence of natural secondary motion of clothes, \\( e.g., dy" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.241, + 0.895, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.427, + 0.895, + 0.495 + ], + "angle": 0, + "content": "Figure 1. Given several sparse multi-view video sequences with estimated 3D body meshes, SurMo synthesizes subject-specific appearance. We specifically focus on the synthesis of plausible time-varying appearances by learning an effective 4D motion representation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.517, + 0.892, + 0.592 + ], + "angle": 0, + "content": "namic movements of the T-shirt induced by dancing in Fig. 1. The secondary motion of clothes arises from intricate physical interactions with the body, typically changing over time, which leads to challenges for plausible rendering of dynamic humans." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.893, + 0.807 + ], + "angle": 0, + "content": "Most existing methods [25, 43, 61] are conditioned on static poses and use a pose-guided generator to synthesize time-varying appearances. However, the appearance of clothed humans undergoes complex geometric transformations induced not only by the static pose but also its dynamics, whereas the modeling of dynamics is often ignored. In addition, these methods are often focused on the 2D appearance reconstruction of every single frame while the temporal motion relations are not fully explored. Due to these issues, they fail to render plausible secondary motions of clothes, i.e., generating the same appearances for fast and slow motions. A key challenge of learning temporal dynamics lies in the requirement of a tremendous amount of training observations to construct a 4D motion volume." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To solve these issues, we propose a new paradigm to learn time-varying appearances of the secondary motion from just several sparse viewpoint video sequences, which is achieved by jointly modeling the temporal motion dynamics and human appearances in a unified rendering framework based on an efficient surface-based motion" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6550" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "representation. At the core of the paradigm is a feature encoder-decoder framework with three key components: 1) surface-based motion encoding; 2) physical motion decoding; and 3) 4D appearance decoding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.473, + 0.501 + ], + "angle": 0, + "content": "Firstly, in contrast to existing pose-guided methods [25, 42, 43, 61] that focus on static poses as a conditional variable, we extract an expressive 4D motion input from the 3D body mesh sequences obtained from training video as our input, which includes both a static pose represented by a spatial 3D mesh and its temporal dynamics. Furthermore, we notice that the non-rigid deformations of garments typically occur around the body surface instead of in a 3D volume, and hence we propose to model human motions on the body surface by projecting the extracted spatial-temporal 4D motion input to the dense 2D surface UV manifold of a clothless body template (e.g., SMPL [28]). To model temporal clothing offsets, a motion encoder is employed to lift the clothless motion features into a motion triplane that encodes both spatial and temporal motion relations in a compact 3D triplane with time-varying dynamics conditioned. The triplane is defined in the surface \\(\\mathbf{u} - \\mathbf{v} - \\mathbf{h}\\) coordinate system, with \\(\\mathbf{u} - \\mathbf{v}\\) to represent the motion of the clothless body template, and an extensional coordinate \\(\\mathbf{h}\\) to represent the secondary motion of clothes, i.e., the temporal clothing offsets are parameterized by a signed distance to the body surface. In this way, 4D motions can be effectively represented by a surface-based triplane." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.47, + 0.639 + ], + "angle": 0, + "content": "Secondly, we propose to physically model the spatial and temporal motion dynamics in the rendering network. Specifically, with the surface-based triplane conditioned at time \\( t \\), a motion decoder is employed to decode the motion triplanes to predict the motion at the next timestep \\( t + 1 \\), i.e., spatial derivative of the motion physically corresponding to surface normal map and temporal derivatives corresponding to surface velocity. We illustrate the physical motion learning significantly improves the rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.47, + 0.746 + ], + "angle": 0, + "content": "Thirdly, the motion triplanes are decoded into high-quality images at two stages: a volumetric surface-conditioned renderer that is focused on the rendering around human body surface and filters the query points far from the body surface for efficient volumetric rendering, and a geometry-aware super-resolution module for efficient high-quality image synthesis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.47, + 0.884 + ], + "angle": 0, + "content": "We conduct a systematical analysis of how human appearances are affected by temporal dynamics, and it was observed that some baseline methods mainly generate pose-dependent appearances instead of time-varying appearances for free-view video generations. In addition, quantitative and qualitative experiments are performed on three datasets with a total of 9 subject sequences, including ZJU-MoCap [43], AIST++ [24], and MPII-RDDC [13], which validate the effectiveness of SurMo in different scenarios." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.327, + 0.901 + ], + "angle": 0, + "content": "In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "1) A new paradigm for learning dynamic humans from videos that jointly models temporal motions and human appearances in a unified framework, and one of the early works that systematically analyzes how human appearances are affected by temporal dynamics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.167, + 0.892, + 0.212 + ], + "angle": 0, + "content": "2) An efficient surface-based triplane that encodes both spatial and temporal motion relations for expressive 4D motion modeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.288 + ], + "angle": 0, + "content": "3) We achieve state-of-the-art results and show that our new paradigm is capable of learning high-fidelity appearances from fast motion sequences (e.g., AIST++ dance videos) or synthesizing motion-dependent shadows in challenging scenarios." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.302, + 0.642, + 0.317 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.357 + ], + "angle": 0, + "content": "Our method is closely related to many sub-fields of visual computing, and below we discuss a set of the work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.463 + ], + "angle": 0, + "content": "3D Shape Representations. To capture detailed shapes of 3D objects, most recent papers utilize implicit functions [7, 8, 20, 33-35, 40, 41, 48-50, 56, 57, 65, 66] or point clouds [16, 31, 32] due to their topological flexibility. These methods aim to learn geometry from 3D datasets, whereas we synthesize human images of novel poses only from 2D RGB training images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.464, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Rendering Humans by 2D GANs. Some existing approaches render human avatars by neural image translation, i.e., they utilize GAN [10] networks to learn a mapping from poses (given in the form of renderings of a skeleton [4, 23, 45, 52, 59, 68], dense mesh [11, 26, 27, 36, 51, 59] or joint position heatmaps [1, 29, 30]) to human images [4, 59]. To improve temporal stability, some methods [17, 44, 46, 55] propose to utilize the SMPL [28] priors for pose-guided generations. However, these methods do not reconstruct geometry explicitly and cannot handle self-occlusions effectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Rendering Humans by 3D-aware Renderer. For stable view synthesis, recent papers [6, 25, 38, 42, 43, 53] propose to unify geometry reconstruction with view synthesis by volume rendering, which, however, is computationally heavy. To solve this issue, some recent 3D-GAN papers [5, 12, 15, 18, 19, 37, 39, 67] propose a hybrid rendering strategy for efficient geometry-aware rendering, that is, render low-resolution volumetric features for geometry learning, and employ a super-resolution module for high-resolution image synthesis. We adopt this strategy for efficient rendering, whereas ours is distinguished by rendering articulated humans with 4D motion modeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "UV-based Pose Representation. Some methods [18, 19, 31, 32, 47, 63] propose to project 3D posed meshes into a 2D positional map for pose encoding, which can be used for different downstream tasks, such as 3D reconstruction and novel view synthesis. [31, 32] rely on 3D supervision to learn 3D reconstructions, and they do not take as input the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6551" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "dynamics. To improve the rendering quality, [18, 47] proposes to utilize additional driving views as input for faithful rendering in telepresence applications. However, they do not explore how to learn temporal dynamics from pose sequences. By taking as input a 3D normal and velocity map, [63] encodes motion dynamics in their rendering network, whereas they do explicitly learn motions, i.e., predicting the motion status at the next timestep, and besides, [63] is a 2D rendering method, and they do not explicitly learn motions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.243, + 0.212, + 0.26 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.269, + 0.228, + 0.285 + ], + "angle": 0, + "content": "3.1. Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.293, + 0.47, + 0.385 + ], + "angle": 0, + "content": "Given a sparse multi-view video of a clothed human in motion and corresponding 3D pose estimations \\(\\{\\mathbf{P}_0, \\dots, \\mathbf{P}_t\\}\\), our goal is to synthesize time-varying appearances of the individual under novel views. Existing methods [25, 42, 43, 61] formulate the problem of human rendering as learning a representation via a feature encoder-decoder framework:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.398, + 0.47, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\mathbf {t}} = \\mathcal {E} _ {\\mathcal {P}} \\left(\\mathbf {P} _ {\\mathbf {t}}, \\mathbf {z} _ {\\mathbf {t}}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.419, + 0.306, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {\\mathbf {t}} = \\mathcal {D} _ {\\mathcal {R}} (\\mathbf {f} _ {\\mathbf {t}})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.47, + 0.564 + ], + "angle": 0, + "content": "where a pose encoder \\(\\mathcal{E}_{\\mathcal{P}}\\) takes as input a representation of posed body \\(\\mathbf{P_t}\\) (e.g., 2D or 3D keypoints or 3D body mesh vertices) and a timestamp embedding \\(\\mathbf{z_t}\\) at time \\(t\\), and outputs intermediate pose-dependent features \\(\\mathbf{f_t}\\) that can be rendered by a decoder \\(\\mathcal{D}_{\\mathcal{R}}\\) to reconstruct the appearance images \\(\\mathbf{I_t}\\) of the corresponding pose at time \\(t\\). The time embedding \\(\\mathbf{z_t}\\) often serves as a residual that is updated passively in backpropagation by image reconstruction loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.47, + 0.775 + ], + "angle": 0, + "content": "However, there are two issues with the above well-adopted paradigm. First, the appearance of clothed humans undergoes complex geometric transformations induced not only by the static pose \\(\\mathbf{P}_{\\mathbf{t}}\\) but also its dynamics, whereas the modeling of dynamics is ignored in E.q. 1, and the residual \\(\\mathbf{z}_{\\mathbf{t}}\\) cannot expressively model physical dynamics neither. Second, existing methods focus on the per-image reconstruction while ignoring the temporal relations of a motion sequence in training, i.e., sampling input pose \\(\\mathbf{P}_{\\mathbf{t}}\\) and timestamp \\(\\mathbf{z}_{\\mathbf{t}}\\) individually, and supervising image reconstructions frame-by-frame, partly because defining temporal supervision in the 2D image space is challenging, especially for articulated humans that often suffer from misalignment problems due to pose estimation errors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.777, + 0.47, + 0.822 + ], + "angle": 0, + "content": "To solve these issues, we propose a new paradigm for learning view synthesis of dynamic humans from video sequences, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.26, + 0.837, + 0.382, + 0.852 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\mathbf {t}} = \\mathcal {E} _ {\\mathcal {M}} (\\mathbf {P} _ {\\mathbf {t}}, \\mathbf {D} _ {\\mathbf {t}})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.856, + 0.469, + 0.885 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathbf {P} _ {\\mathbf {t} + 1}}{\\partial \\mathbf {x}}, \\frac {\\partial \\mathbf {P} _ {\\mathbf {t} + 1}}{\\partial \\mathbf {t}} = \\mathcal {D} _ {\\mathcal {M}} (\\mathbf {f} _ {\\mathbf {t}}) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.258, + 0.888, + 0.347, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {\\mathbf {t}} = \\mathcal {D} _ {\\mathcal {R}} (\\mathbf {f} _ {\\mathbf {t}})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.275 + ], + "angle": 0, + "content": "where the input is represented by dynamic motions including a static pose \\(\\mathbf{P}_{\\mathbf{t}}\\) and its physical dynamics \\(\\mathbf{D}_{\\mathbf{t}}\\) at time \\(t\\), and a motion encoder \\(\\mathcal{E}_{\\mathcal{M}}\\) is employed to encode the inputs into motion-dependent features (Sec. 3.2). Besides, in the training stage, a physical motion decoder \\(\\mathcal{D}_{\\mathcal{M}}\\) is introduced to enforce the learning of spatial and temporal relations by decoding the intermediate feature \\(\\mathbf{f}_{\\mathbf{t}}\\) to predict the spatial derivatives at \\(t + 1\\) physically corresponding to surface normal \\(\\mathbf{N}_{\\mathbf{t} + 1} = \\frac{\\partial\\mathbf{P}_{\\mathbf{t} + 1}}{\\partial\\mathbf{x}}\\), and temporal derivatives at \\(t + 1\\) corresponding to surface velocity \\(\\mathbf{V}_{\\mathbf{t} + 1} = \\frac{\\partial\\mathbf{P}_{\\mathbf{t} + 1}}{\\partial\\mathbf{t}}\\) (Sec. 3.3). \\(\\mathbf{f}_{\\mathbf{t}}\\) is rendered into human images by a decoder \\(\\mathcal{D}_{\\mathcal{R}}: \\mathcal{G}_2 \\circ \\mathcal{G}_1\\) (Sec. 3.5). The framework is depicted in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.276, + 0.892, + 0.351 + ], + "angle": 0, + "content": "Notation. In the following parts, we use \\(\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{space}}\\) to denote the features in the pipeline, i.e., \\(\\mathbf{f}_{\\mathbf{t}}^{3\\mathbf{D}}\\) is a 4D representation that is defined in 3D space with temporal \\(t\\) conditioned, and \\(\\mathbf{M}_{\\mathbf{t}}^{3\\mathbf{D}}\\) is 4D motion input defined in 3D space with \\(t\\) conditioned." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.365, + 0.809, + 0.381 + ], + "angle": 0, + "content": "3.2. Surface-based 4D Motion Encoding" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.389, + 0.892, + 0.602 + ], + "angle": 0, + "content": "Extracting 4D Motions. We first extract 4D motions from a sequence of time-varying parametric posed body (e.g., SMPL) meshes obtained from training video sequences. We describe the motion \\(\\mathbf{M}_{\\mathbf{t}}^{3\\mathbf{D}}\\) at time \\(t\\) as a static skeleton pose \\(\\mathbf{P}_{\\mathbf{t}}\\) and its dynamics \\(\\mathbf{D}_{\\mathbf{t}}\\). The dynamics at \\(t\\) are physically determined by the current pose, 3D velocity, and motion trajectory of the past several timesteps, which also contribute to the time-varying appearances of the secondary motion. The pose \\(\\mathbf{P}_{\\mathbf{t}}\\) is represented by the 3D vertices of the posed mesh, and dynamics \\(\\mathbf{D}_{\\mathbf{t}}\\) are parameterized by 1) body surface velocity \\(\\mathbf{V}_{\\mathbf{t}}\\) corresponding to the temporal derivatives of the current pose at \\(t\\), and 2) motion trajectory \\(\\mathbf{T}_{\\mathbf{t}}\\) that aggregates the temporal derivatives over the past several timesteps with a sliding window size of \\(w\\) with weight \\(\\lambda\\):" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.614, + 0.891, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {M} _ {\\mathbf {t}} ^ {\\mathrm {3 D}} = \\left[ \\mathbf {P} _ {\\mathbf {t}}, \\mathbf {D} _ {\\mathbf {t}} \\right], \\quad \\mathbf {D} _ {\\mathbf {t}} = \\left[ \\mathbf {V} _ {\\mathbf {t}}, \\mathbf {T} _ {\\mathbf {t}} \\right] \\\\ \\mathbf {V} _ {\\mathbf {t}} = \\frac {\\partial \\mathbf {P} _ {\\mathbf {t}}}{\\partial \\mathbf {t}}, \\quad \\mathbf {T} _ {\\mathbf {t}} = \\mathbf {P} _ {\\mathbf {t}} + \\frac {1}{\\sum \\lambda_ {i}} \\sum_ {i = 1} ^ {w} \\mathbf {V} _ {\\mathbf {t} - \\mathbf {i}} * \\lambda_ {i} \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.689, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Note the motion trajectory aggregated from several consecutive timesteps makes the motion representation robust to pose estimation errors, i.e., the pose estimations for two consecutive timesteps may be the same due to pose estimation errors in practice. An ablation study of the motion trajectory representation can be found in the supp. mat." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.78, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Recording 4D Motions on the Surface Manifold. Modeling the motions with temporal dynamics often requires dense observations to construct a 4D motion volume. Instead, we notice that non-rigid deformations of human geometry in motion typically occur around the body surface instead of a 3D volume, and hence we propose to model the motions on the human body surface. Specifically, we project the 4D motion input \\(\\mathbf{M}_{\\mathbf{t}}^{3\\mathrm{D}}\\) including spatial pose" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6552" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.895, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.266, + 0.895, + 0.406 + ], + "angle": 0, + "content": "Figure 2. Framework overview. Given a set of time-varying 3D body meshes \\(\\{\\mathbf{P}_{\\mathbf{t}},\\dots,\\mathbf{P}_{\\mathbf{t}} - \\mathbf{n}\\}\\) obtained from training video sequences, we aim to synthesize high-fidelity appearances of a clothed human in motion via a feature encoder-decoder framework: Motion Encoding, and joint Motion and Appearance Decoding. 1) We take as input an expressive 4D motion representation extracted from the mesh sequences including 3D pose, 3D velocity at time \\(t\\), and motion trajectory over the past \\(w\\) timesteps that encode both spatial and temporal relations of the motion sequence, which are projected to the spatially aligned UV surface space. A motion encoder \\(\\mathcal{E}_{\\mathcal{M}}\\) is employed to lift the 2D UV-aligned features to a 3D surface-based triplane \\(\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}\\) in an UV-plus-height space with a signed distance height to model temporal clothing offsets. 2) A motion decoder \\(\\mathcal{D}_{\\mathcal{M}}\\) is designed to encourage physical motion learning in training by decoding the triplane features \\(\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}\\) to predict the motion at the next timestep \\(t + 1\\), i.e. spatial derivatives surface normal \\(\\mathbf{N}_{\\mathbf{t} + 1}^{\\mathrm{uv}}\\) and temporal derivatives surface velocity \\(\\mathbf{V}_{\\mathbf{t} + 1}^{\\mathrm{uv}}\\) in UV space. 3) Finally, given a target camera view, the triplane \\(\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}\\) is rendered into high-quality images by a volumetric surface-conditioned renderer including volumetric low-resolution rendering by \\(\\mathcal{G}_1\\) and an efficient geometry-aware super-resolution by \\(\\mathcal{G}_2\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.47, + 0.512 + ], + "angle": 0, + "content": "and temporal dynamics from 3D space into a compact spatially aligned UV space using the geometric transformation \\(\\mathcal{W}\\) that is pre-defined by the parametric (e.g., SMPL) body template, which yields \\(\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uv}} = \\mathcal{W}\\mathbf{M}_{\\mathbf{t}}^{3\\mathrm{D}}\\). The UV-aligned motion feature \\(\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uv}}\\) faithfully preserves the articulated structures of body topology in a compact 2D space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.471, + 0.648 + ], + "angle": 0, + "content": "Generating Surface-based Triplanes for Motion Modeling. To model clothed humans, we further employ a motion encoder \\(\\mathcal{E}_{\\mathcal{M}}\\) that lifts the clothless 2D features \\(\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uv}}\\) to a 3D triplane representation \\(\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}\\) that is defined in a \\(\\mathbf{u} - \\mathbf{v} - \\mathbf{h}\\) system, with \\(\\mathbf{u} - \\mathbf{v}\\) to represent the motion of the clothless body template, and an extensional coordinate \\(\\mathbf{h}\\) to represent the secondary motion of clothes, i.e., the temporal cloth offsets, and hence 4D clothed motions can be parameterized by a surface-based triplane by:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.658, + 0.47, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathrm {u v h}} = \\mathcal {E} _ {\\mathcal {M}} \\left(\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathrm {u v}}\\right) = \\mathcal {E} _ {\\mathcal {M}} \\left(\\mathcal {W} \\mathbf {M} _ {\\mathbf {t}} ^ {\\mathbf {3 D}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.47, + 0.825 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}\\) is the geometric transformation from 3D space to UV space. \\(\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}\\) consists of three planes \\(x_{uv}, x_{uh}, x_{hv} \\in \\mathbb{R}^{U \\times V \\times C}\\) to form the spatial relationship, where \\(U, V\\) denote spatial resolution and \\(C\\) is the channel number. In contrast to the volumetric triplanes used in EG3D [5] where the three planes represent three vertical planes in the 3D volumetric space, \\(\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}\\) is defined on the human body surface inheriting human topology priors for motion modeling, as illustrated by an unwarped triplane in 3D space in Fig 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.317, + 0.849 + ], + "angle": 0, + "content": "3.3. Physical Motion Decoding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.903 + ], + "angle": 0, + "content": "We propose a physical motion decoder \\(\\mathcal{D}_{\\mathcal{M}}\\) to learn spatial and temporal features of motion, which is achieved by decoding the intermediate motion feature \\(\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}\\) learned by \\(\\mathcal{E}_{\\mathcal{M}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.892, + 0.496 + ], + "angle": 0, + "content": "to predict the motions at the next timestep \\( t + 1 \\), such as spatial derivatives corresponding to surface normal \\( \\mathbf{N}_{t+1}^{\\mathrm{uv}} \\), temporal derivatives corresponding to surface velocity \\( \\mathbf{V}_{t+1}^{\\mathrm{uv}} \\). Note that we model the normal and velocity in the UV space by:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.502, + 0.892, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathbf {N} _ {\\mathbf {t} + \\mathbf {1}} ^ {\\mathrm {u v}}, \\mathbf {V} _ {\\mathbf {t} + \\mathbf {1}} ^ {\\mathrm {u v}} \\right\\} = \\mathcal {D} _ {\\mathcal {M}} \\left(\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathrm {u v h}}\\right) = \\mathcal {D} _ {\\mathcal {M}} \\circ \\mathcal {E} _ {\\mathcal {M}} \\left(\\mathcal {W} \\mathbf {M} _ {\\mathbf {t}} ^ {\\mathbf {3 D}}\\right) (5)\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.535, + 0.733, + 0.551 + ], + "angle": 0, + "content": "3.4. 4D Appearance Decoding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.892, + 0.743 + ], + "angle": 0, + "content": "Volumetric Surface-conditioned Rendering. Given a target camera viewpoint, the triplane \\(\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}\\) is first rendered into low-resolution volumetric features \\(\\mathbf{I}_{\\mathbf{F}} \\in \\mathbb{R}^{H \\times W \\times \\hat{C}}\\) by a conditional NeRF \\(\\mathbb{F}_{\\Phi}\\), where \\(H\\) and \\(W\\) are the spatial resolution and \\(\\hat{C}\\) is the channel number. To be more specific, given a 3D query point \\(\\pmb{p}_i\\), we transform it into a surface-based local coordinate \\(\\hat{\\pmb{p}}_i = (\\pmb{u}_i, \\pmb{v}_i, \\pmb{h}_i)\\) w.r.t. the tracked body mesh \\(\\mathbf{P}_{\\mathbf{t}}\\). Here we search the nearest face \\(\\pmb{f}_i\\) of \\(\\mathbf{P}_{\\mathbf{t}}\\) for query point \\(\\pmb{p}_i\\) and \\((\\pmb{u}_i, \\pmb{v}_i)\\) and \\(\\pmb{h}_i\\) are the barycentric coordinates of the nearest point on \\(\\pmb{f}_i\\), and the signed distance respectively. We therefore obtain the local feature \\(\\pmb{z}_i^{\\mathrm{uvh}}\\) of the query point \\(\\pmb{p}_i\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.748, + 0.903, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {i} ^ {u v h} = \\operatorname {c a t} \\left[ \\Pi \\left(\\boldsymbol {x} _ {u v}; \\boldsymbol {u} _ {i}, \\boldsymbol {v} _ {i}\\right), \\Pi \\left(\\boldsymbol {x} _ {u h}; \\boldsymbol {u} _ {i}, \\boldsymbol {h} _ {i}\\right), \\Pi \\left(\\boldsymbol {x} _ {h v}; \\boldsymbol {h} _ {i}, \\boldsymbol {v} _ {i}\\right) \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.891, + 0.809 + ], + "angle": 0, + "content": "where \\(\\Pi (\\cdot)\\) denotes sampling operation, \\(\\mathrm{cat}[\\cdot ]\\) denotes the concatenation operator." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.81, + 0.89, + 0.841 + ], + "angle": 0, + "content": "Given camera direction \\( \\pmb{d}_i \\), the appearance features \\( \\pmb{c}_i \\) and density features \\( \\pmb{\\sigma}_i \\) of point \\( \\pmb{p}_i \\) are predicted by" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.846, + 0.891, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\boldsymbol {c} _ {i}, \\boldsymbol {\\sigma} _ {i} \\right\\} = \\mathbb {F} _ {\\Phi} \\left(\\boldsymbol {d} _ {i}, \\boldsymbol {z} _ {i} ^ {u v h}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We integrate all the radiance features of sampled points into a 2D feature map \\( I_{F} \\) at time \\( t \\) through volume renderer" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6553" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.089, + 0.895, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.567, + 0.895, + 0.637 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons on novel view synthesis on the subject S313 of ZJU-MoCap dataset. Two motion sequences S1 (swing arms left to right) and S2 (raise and lower arms) are shown. We specifically focus on the synthesis of time-varying appearances (especially T-shirt wrinkles), by evaluating the rendering results under similar poses yet with different movement directions, which are marked in the same color, such as the pairs of \\(①②\\), \\(③④\\), and \\(⑤⑥\\). Our method synthesizes high-fidelity time-varying appearances, whereas SOTA HumanNeRF generates almost the same cloth wrinkles." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.663, + 0.132, + 0.679 + ], + "angle": 0, + "content": "\\(\\mathcal{G}_1\\) [21]" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.677, + 0.469, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {\\mathbf {F}} = \\mathcal {G} _ {1} \\left(\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {u v h}}, c a m; \\mathbb {F} _ {\\Phi}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.701, + 0.317, + 0.716 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_{\\mathrm{t}}^{\\mathrm{uvh}}\\) is extracted from Eq. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.469, + 0.796 + ], + "angle": 0, + "content": "Efficient Geometry-Aware Super-Resolution. Sampling dense points to render the full-resolution volumetric features is computationally heavy. Instead, we employ a superresolution network \\(\\mathcal{G}_2\\) [5] to render high-resolution images \\(\\mathbf{I}_{\\mathbf{RGB}}^{+} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times 3}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.807, + 0.469, + 0.826 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {\\mathbf {R G B}} ^ {+} = \\mathcal {G} _ {2} \\circ \\mathcal {G} _ {1} \\left(\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {u v h}}, c a m; \\mathbb {F} _ {\\Phi}\\right) \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.833, + 0.217, + 0.849 + ], + "angle": 0, + "content": "3.5. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "SurMo is trained end-to-end to optimize \\(\\mathcal{E}_{\\mathcal{M}}\\), \\(\\mathcal{D}_{\\mathcal{M}}\\), and renderers \\(\\mathcal{G}_1\\), \\(\\mathcal{G}_1\\) with 2D image loss. We employ Adversarial Loss and Reconstruction Loss including Pixel Loss," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.663, + 0.892, + 0.723 + ], + "angle": 0, + "content": "Perceptual Loss, Face Identity Loss, Velocity and Normal Loss, and Volume Rendering Loss for supervision in training, with Adam [22] as the optimizer. Refer to the supp. mat. for more details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.739, + 0.634, + 0.757 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Dataset and Metrics. We evaluate the novel view synthesis on ZJU-MoCap, MPII-RDDC, and AIST++, with a total of 9 subjects. We use the same camera setup as Neural Body, that is, 4 cameras used for training, and the other 18 for testing. For MPII-RDDC, 18 cameras in training, and 9 for testing. Since ASIT++ only has 9 cameras, we use 6 for training, and the remaining 3 for testing. Refer to more details in the supp. mat. We compare each method on per-pixel metrics including SSIM [60] and PSNR, and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6554" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.487, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.316, + 0.445, + 0.328 + ], + "angle": 0, + "content": "Neural Body Instant-NVR HumanNeRF Ours GT" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.091, + 0.892, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.316, + 0.848, + 0.328 + ], + "angle": 0, + "content": "Neural Body Instant-NVR HumanNeRF Ours GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.339, + 0.894, + 0.381 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons on novel view synthesis on the subject S387, S315 of ZJU-MoCap dataset. Row 1 and 2 show similar poses occurring at different timesteps (not consecutive frames). The results indicate that our method synthesizes time-varying appearances while other methods mainly generate pose-dependent appearances." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.406, + 0.47, + 0.504 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on ZJU-MoCap dataset (averaged on all test views and poses on 6 sequences) for novel-view synthesis. To reduce the influence of the background, all scores are calculated from images cropped to 2D bounding boxes. Note that the perception metrics LPIPS [64] and FID [14] capture human judgment better than per-pixel metrics such as SSIM [60] or PSNR, as stated in [25, 51]." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.515, + 0.47, + 0.611 + ], + "angle": 0, + "content": "
ZJU-S1-6LPIPS ↓FID ↓SSIM ↑PSNR ↑
Neural Body [43].164125.10.70321.438
Instant-NVR [9].202144.99.73821.748
HumanNeRF [61].10688.382.79223.624
Ours.07567.725.83324.815
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.44, + 0.641 + ], + "angle": 0, + "content": "perception metrics including LPIPS [64] and FID [14]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.469, + 0.777 + ], + "angle": 0, + "content": "Baselines. We compare our method against SOTA methods including Neural Body [43], HumanNeRF [61], Instant-NVR [9], ARAH [58], DVA [47] and HVTR++ [18]. Neural Body models human poses in 3D space with point cloud as pose representation, HumanNeRF and Instant-NVR model poses in a canonical space with inverse skinning, ARAH is a forward-skinning based method, DVA and HVTR++ encode poses in UV space, and take driving view signals as input for telepresence applications." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.357, + 0.803 + ], + "angle": 0, + "content": "4.1. Comparisons to SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Quantitative comparisons We conduct the quantitative comparisons on the three datasets with a total of 9 subjects. The quantitative results on ZJU-MoCap are summarized in Tab. 1, which suggests that our new paradigm outperforms the SOTA by a big margin, and we achieve the best quantitative results on all four metrics. The detailed compar-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.469 + ], + "angle": 0, + "content": "isons on each subject of ZJU-MoCap, and quantitative results on MPII-RDDC and AIST++ are listed in the supp. mat. Refer to the comparisons with ARAH [58], DVA [47] and HVTR++ [18] in the supp. mat." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.893, + 0.776 + ], + "angle": 0, + "content": "Time-varying Appearances with Dynamics Conditioning on ZJU-MoCap. We compare with baseline methods for novel view synthesis on two motion sequences of subject S313 from the ZJU-MoCap dataset, as shown in Fig. 3. We evaluate the capability of each method to synthesize time-varying appearances. Specifically, two sequences S1 (swing arms left to right) and S2 (raise and lower arms) are evaluated, where similar poses occur at different timesteps with different dynamics (e.g., movement directions, trajectory, or velocity) are marked in the same color, such as the pairs of \\(①②\\), \\(③④\\), and \\(⑤⑥\\). Fig. 3 suggests that Instant-NVR [9] fails to synthesize time-varying high-frequency wrinkles, and Neural Body [43] synthesizes time-varying yet blurry wrinkles. HumanNeRF [61] synthesizes detailed yet static T-shirt wrinkles, i.e., the wrinkles are almost the same for similar poses, such as \\(①②\\), \\(③④\\), and \\(⑤⑥\\). In contrast, our method renders both high-fidelity and time-varying wrinkles. The comparisons on other subjects S387 and S315 are shown in Fig. 4, which illustrate the effectiveness of our proposed paradigm in dynamics learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Time-varying Appearances with Motion-dependent Shadows on MPII-RDDC. We compare with baseline methods for novel view synthesis on MPII-RDDC, as shown in Fig. 5. The sequence is captured in a studio with top-down lighting that casts shadows on the human body due to self-occlusions. We notice that the synthesis of shadow can be formulated as learning motion-dependent shadows in the motion representation without the need to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6555" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.894, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.29, + 0.321, + 0.301 + ], + "angle": 0, + "content": "HumanNeRF" + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.29, + 0.398, + 0.301 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.463, + 0.29, + 0.549, + 0.301 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.29, + 0.662, + 0.301 + ], + "angle": 0, + "content": "HumanNeRF" + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.29, + 0.732, + 0.301 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.79, + 0.29, + 0.875, + 0.301 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.314, + 0.895, + 0.384 + ], + "angle": 0, + "content": "Figure 5. Novel view synthesis of time-varying appearances with both pose and lighting conditioning on MPII-RDDC dataset. The sequence is captured in a studio with top-down lighting that casts shadows on the human performer due to self-occlusion. In Row 1, we specifically focus on synthesizing time-varying shadows (e.g., ① vs. ②, and ③ vs. ④) for different poses with different self-occlusions. In Row 2, we evaluate the synthesis of: 1) time-varying appearances for similar poses occurring in a jump-up-and-down motion sequence, e.g., ⑤ vs. ⑥, 2) shadows ⑦ vs. ⑧, and 3) clothing offsets ⑤ vs. ⑥." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.407, + 0.47, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.497, + 0.45, + 0.511 + ], + "angle": 0, + "content": "Figure 6. Novel view synthesis of fast motions on AIST++." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.749 + ], + "angle": 0, + "content": "explicitly model the lighting. We validate our method from two aspects of the scenario in Fig. 5. 1) In Row 1, Fig. 5 indicates that with the expressive motion representation and physical motion learning, SurMo succeeds in predicting the motion-dependent shadows, such as \\(①②③④\\), whereas SOTA HumanNeRF renders almost the same T-shirt appearance. 2) In Row 2, we compare the synthesis of time-varying appearances for similar poses occurring in a jump-up-and-down motion sequence i.e., \\(⑤\\) vs. \\(⑥\\), which shows that SurMo is capable of predicting the clothes offsets under similar pose with different motion trajectories. In addition, we synthesize the shadows \\(⑦⑧\\) in the challenging jump-up-and-down motion sequence. However, HumanNeRF fails to predict the dynamic clothing offsets \\(⑨\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Time-varying Appearances for Fast Motions on AIST++. We also evaluate our methods by rendering humans with fast dance motions (S21 and S13) on AIST++, as shown in Fig. 6, where we compare ours against Neural Body in synthesizing time-varying wrinkles of the T-shirt for different motions. Fig. 6 suggests that Neural Body can only synthesize blurry results partly because Neural Body models pose in a sparse 3D space. Instead, we model motions in a denser body surface space, which enables high-fidelity image synthesis. Note that AIST++ is based" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.408, + 0.892, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.575, + 0.892, + 0.632 + ], + "angle": 0, + "content": "Figure 8. Ablation study of motion conditioning and learning. We focus on the effect of motion conditioning and learning in Row 1, and whether our method learns to decouple the static pose and dynamics (e.g., velocity) from the motion conditioning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.892, + 0.727 + ], + "angle": 0, + "content": "on SMPL tracking with a scaling factor that affects the inverse LBS used in Instant-NVR or HumanNeRF both relying on inverse LBS, and hence we cannot compare them based on the released official code. Quantitative results can be found in the supp. mat." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.655, + 0.757 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Surface-based Triplane vs. Volumetric Triplane. We compare the well-adopted volumetric triplane (Vol-Trip) [5] and our proposed surface-based triplane (Surf-Trip) for human modeling on two subjects S313 and S387 from ZJUMoCap, as shown in Fig. 7. Fig. 7 (left) suggests that the Surf-Trip converges faster in training with a smaller reconstruction loss for both sequences, and the performances are further improved with physical motion learning, e.g., 'Surf-Trip + ML'. Fig. 7 (right) compares the rendering re" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6556" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.375, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.09, + 0.888, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.28, + 0.894, + 0.337 + ], + "angle": 0, + "content": "Figure 7. Ablation study of 3D volumetric triplane (Vol-Trip) vs. surface-based triplane (Surf-Trip) for human modeling on subject S313 and S387 from ZJU-MoCap. We focus on the convergence in training (left), and self-occlusion rendering (right). The convergence of S313 is shifted for visualization purposes. Vol-Trip is not effective in handling self-occlusion, e.g., \\(①②③④\\) though it performs well in another viewpoint without self-occlusions \\(⑥\\). In addition, Vol-Trip cannot synthesize high-quality details for face." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.36, + 0.47, + 0.416 + ], + "angle": 0, + "content": "Table 2. Ablation study of motion conditioning and learning. \\( P_{cond} \\) and \\( D_{cond} \\) denote the conditioning of pose and dynamics, \\( V_{pred} \\) and \\( N_{pred} \\) denote the prediction of surface velocity and surface normal in motion learning." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.417, + 0.442, + 0.512 + ], + "angle": 0, + "content": "
S313LPIPS ↓FID ↓SSIM ↑PSNR ↑
Pcond.120104.83.79322.781
+ Dcond.08573.674.83424.908
+ Vpred.06962.092.85625.845
+ Npred.06050.170.86926.654
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.47, + 0.625 + ], + "angle": 0, + "content": "sults qualitatively, which indicates that Surf-Trip is more effective in handling self-occlusions, whereas Vol-Trip fails \\(①②③④\\), though it performs well in another viewpoint without self-occlusions \\(⑥\\). In addition, Surf-Trip generates high-fidelity clothing wrinkles and facial details, whereas the face rendering of Vol-Trip is blurry." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Motion Conditioning and Learning. We also analyze the effectiveness of our proposed motion conditioning and learning qualitatively and quantitatively. Tab. 2 summarizes the quantitative results, which suggest that the quantitative results are significantly improved by taking as input the dynamics conditioning \\( D_{cond} \\). With physical motion learning, i.e., predicting the temporal derivatives surface velocity and spatial derivates normal at the next timestep, the quantitative results are further improved by a big margin." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The qualitative comparisons are shown in Fig. 8, where in Row 1, we observe higher-fidelity appearances with dynamics conditioning, and physical motion learning for both velocity and normal prediction. In Row 2, we evaluate whether our method learns to decouple the temporal dynamics (e.g., velocity) and static poses from motion conditioning, which corresponds to the rendering of clothing wrinkles of secondary motion (T-shirt ①) and tight parts (e.g., shorts ②). We only change the velocity for each variant" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.895, + 0.545 + ], + "angle": 0, + "content": "in the ablation study. It suggests that the appearance of the T-shirt varies when the velocity or dynamics are changed, whereas the renderings of tight parts remain the same, such as the head, tight shorts, and shoes. This is consistent with our daily observations. The ablation study illustrates that our method decouples the pose and dynamics, and is capable of generating dynamics-dependent appearances. Note that the appearance changes are small between \\( V * 1 \\) and \\( V * 0.01 \\) when we reduce the velocity, partly because the method is not sensitive to smaller velocity due to the pose estimation errors in the training dataset where consecutive frames may be estimated with the same pose fitting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.571, + 0.614, + 0.587 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.894, + 0.753 + ], + "angle": 0, + "content": "We propose SurMo, a new paradigm for learning dynamic humans from videos by jointly modeling temporal motions and human appearances in a unified framework, based on an efficient surface-based triplane. We conduct a systematical analysis of how human appearances are affected by temporal dynamics, and extensive experiments validate the expressiveness of the surface-based triplane in rendering fast motions and motion-dependent shadows. Quantitative experiments illustrate that SurMo achieves the SOTA results on different motion sequences." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.78, + 0.661, + 0.798 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.903 + ], + "angle": 0, + "content": "This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6557" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Kfir Aberman, M. Shi, Jing Liao, Dani Lischinski, B. Chen, and D. Cohen-Or. Deep video-based performance cloning. Computer Graphics Forum, 38, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.468, + 0.21 + ], + "angle": 0, + "content": "[2] George Borshukov, Dan Piponi, Oystein Larsen, J. P. Lewis, and Christina Tempelaar-Lietz. Universal capture: image-based facial animation for \"the matrix reloaded\". In SIGGRAPH '03, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.21, + 0.468, + 0.251 + ], + "angle": 0, + "content": "[3] Joel Carranza, Christian Theobalt, Marcus A. Magnor, and Hans-Peter Seidel. Free-viewpoint video of human actors. ACM SIGGRAPH 2003 Papers, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.251, + 0.468, + 0.29 + ], + "angle": 0, + "content": "[4] Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A. Efros. Everybody dance now. ICCV, pages 5932-5941, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.291, + 0.468, + 0.37 + ], + "angle": 0, + "content": "[5] Eric Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, S. Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. ArXiv, abs/2112.07945, 2021. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.372, + 0.468, + 0.413 + ], + "angle": 0, + "content": "[6] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Linchao Bao, and Huchuan Lu. Animatable neural radiance fields from monocular rgb video. ArXiv, abs/2106.13629, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.413, + 0.468, + 0.439 + ], + "angle": 0, + "content": "[7] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.468, + 0.493 + ], + "angle": 0, + "content": "[8] Boyang Deng, John P Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. Nasa neural articulated shape approximation. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.493, + 0.468, + 0.533 + ], + "angle": 0, + "content": "[9] Chen Geng, Sida Peng, Zhen Xu, Hujun Bao, and Xiaowei Zhou. Learning neural volumetric representations of dynamic humans in minutes. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.468, + 0.587 + ], + "angle": 0, + "content": "[10] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.468, + 0.642 + ], + "angle": 0, + "content": "[11] A. K. Grigor'ev, Artem Sevastopolsky, Alexander Vakhitov, and Victor S. Lempitsky. Coordinate-based texture inpainting for pose-guided human image generation. CVPR, pages 12127-12136, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.642, + 0.468, + 0.694 + ], + "angle": 0, + "content": "[12] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. ArXiv, abs/2110.08985, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.695, + 0.468, + 0.75 + ], + "angle": 0, + "content": "[13] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics (TOG), 40:1-16, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.75, + 0.468, + 0.804 + ], + "angle": 0, + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.805, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[15] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. ArXiv, abs/2112.05637, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[16] Tao Hu, Geng Lin, Zhizhong Han, and Matthias Zwicker. Learning to generate dense point clouds with textures on multiple categories. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.534, + 0.093, + 0.743, + 0.107 + ], + "angle": 0, + "content": "pages 2170-2179, January 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.107, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[17] Tao Hu, Kripasindhu Sarkar, Lingjie Liu, Matthias Zwicker, and Christian Theobalt. Egorenderer: Rendering human avatars from egocentric camera images. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.147, + 0.892, + 0.215 + ], + "angle": 0, + "content": "[18] Tao Hu, Hongyi Xu, Linjie Luo, Tao Yu, Zerong Zheng, He Zhang, Yebin Liu, and Matthias Zwicker. Hvtr++: Image and pose driven human avatars using hybrid volumetric-textural rendering. IEEE Transactions on Visualization and Computer Graphics, pages 1-15, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.215, + 0.892, + 0.254 + ], + "angle": 0, + "content": "[19] T. Hu, Tao Yu, Zerong Zheng, He Zhang, Yebin Liu, and Matthias Zwicker. Hvtr: Hybrid volumetric-textural rendering for human avatars. 3DV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.254, + 0.892, + 0.295 + ], + "angle": 0, + "content": "[20] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. 2020 (CVPR), pages 3090-3099, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.295, + 0.892, + 0.335 + ], + "angle": 0, + "content": "[21] James T. Kajiya and Brian Von Herzen. Ray tracing volume densities. Proceedings of the 11th annual conference on Computer graphics and interactive techniques, 1984. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[22] Diederick P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[23] Bernhard Kratzwald, Zhiwu Huang, Danda Pani Paudel, and Luc Van Gool. Towards an understanding of our world by GANing videos in the wild. arXiv:1711.11453, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.402, + 0.892, + 0.442 + ], + "angle": 0, + "content": "[24] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Learn to dance with aist++: Music conditioned 3d dance generation, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[25] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. TOG, 40, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.496, + 0.892, + 0.578 + ], + "angle": 0, + "content": "[26] Lingjie Liu, Weipeng Xu, Marc Habermann, Michael Zollhöfer, Florian Bernard, Hyeongwoo Kim, Wenping Wang, and Christian Theobalt. Neural human video rendering by learning dynamic textures and rendering-to-video translation. IEEE Transactions on Visualization and Computer Graphics, 05 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.578, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[27] Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Hyeongwoo Kim, Florian Bernard, Marc Habermann, Wenping Wang, and Christian Theobalt. Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[28] M. Loper, Naureen Mahmood, J. Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multi-person linear model. ACM Trans. Graph., 34:248:1-16, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.686, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[29] Liqian Ma, Xu Jia, Qianru Sun, Bernt Schiele, Tinne Tuytelaars, and Luc Van Gool. Pose guided person image generation. In NeurIPS, pages 405-415, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.727, + 0.892, + 0.767 + ], + "angle": 0, + "content": "[30] Liqian Ma, Qianru Sun, Stamatios Georgoulis, Luc van Gool, Bernt Schiele, and Mario Fritz. Disentangled person image generation. CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.767, + 0.892, + 0.82 + ], + "angle": 0, + "content": "[31] Qianli Ma, Shunsuke Saito, Jinlong Yang, Siyu Tang, and Michael J. Black. Scale: Modeling clothed humans with a surface codec of articulated local elements. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.82, + 0.892, + 0.86 + ], + "angle": 0, + "content": "[32] Qianli Ma, Jinlong Yang, Siyu Tang, and Michael J Black. The power of points for modeling humans in clothing. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.888 + ], + "angle": 0, + "content": "[33] Lars M. Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy net" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6558" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "works: Learning 3d reconstruction in function space. CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.12, + 0.469, + 0.173 + ], + "angle": 0, + "content": "[34] Mateusz Michalkiewicz, Zhony Kaesemodel Pontes, Dominic Jack, Mahsa Baktash, and Anders P. Eriksson. Deep level sets: Implicit surface representations for 3d shape inference. ArXiv, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.174, + 0.469, + 0.213 + ], + "angle": 0, + "content": "[35] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. Leap: Learning articulated occupancy of people. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.215, + 0.468, + 0.24 + ], + "angle": 0, + "content": "[36] Natalia Neverova, Riza Alp Güler, and Iasonas Kokkinos. Dense pose transfer. ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.241, + 0.469, + 0.28 + ], + "angle": 0, + "content": "[37] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. CVPR, pages 11448-11459, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.281, + 0.469, + 0.32 + ], + "angle": 0, + "content": "[38] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In IEEE/CVF ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[39] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. ArXiv, abs/2112.11427, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.469, + 0.415 + ], + "angle": 0, + "content": "[40] Pablo Palafox, Aljaž Božić, Justus Thies, Matthias Nießner, and Angela Dai. Npms: Neural parametric models for 3d deformable shapes. In IEEE/CVF ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.416, + 0.469, + 0.469 + ], + "angle": 0, + "content": "[41] Jeong Joon Park, Peter R. Florence, Julian Straub, Richard A. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. 2019 (CVPR), pages 165-174, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[42] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.524, + 0.469, + 0.59 + ], + "angle": 0, + "content": "[43] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. CVPR, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.629 + ], + "angle": 0, + "content": "[44] Sergey Prokudin, Michael J. Black, and Javier Romero. Smplpix: Neural avatars from 3d human models. WACV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.469, + 0.672 + ], + "angle": 0, + "content": "[45] Albert Pumarola, Antonio Agudo, Alberto Sanfeliu, and Francesc Moreno-Noguer. Unsupervised person image synthesis in arbitrary poses. In CVPR, June 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.469, + 0.712 + ], + "angle": 0, + "content": "[46] Amit Raj, Julian Tanke, James Hays, Minh Vo, Carsten Stoll, and Christoph Lassner. Anr: Articulated neural rendering for virtual avatars. CVPR, pages 3721-3730, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.713, + 0.469, + 0.78 + ], + "angle": 0, + "content": "[47] Edoardo Remelli, Timur M. Bagautdinov, Shunsuke Saito, Chenglei Wu, Tomas Simon, Shih-En Wei, Kaiwen Guo, Zhe Cao, Fabian Prada, Jason M. Saragih, and Yaser Sheikh. Drivable volumetric avatars using texel-aligned features. ACM SIGGRAPH, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.781, + 0.469, + 0.834 + ], + "angle": 0, + "content": "[48] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. IEEE/CVF ICCV, pages 2304-2314, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.835, + 0.469, + 0.889 + ], + "angle": 0, + "content": "[49] Shunsuke Saito, Tomas Simon, Jason M. Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. 2020 (CVPR), pages 81-90, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[50] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J. Black. Scintimate: Weakly supervised learning of skinned clothed avatar networks. 2021 (CVPR), pages 2885-2896, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.148, + 0.892, + 0.2 + ], + "angle": 0, + "content": "[51] Kripasindhu Sarkar, Dushyant Mehta, Weipeng Xu, Vladislav Golyanik, and Christian Theobalt. Neural re- rendering of humans from a single image. In ECCV, 2020. 2,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.202, + 0.892, + 0.242 + ], + "angle": 0, + "content": "[52] Aliaksandr Siarohin, Enver Sangineto, Stephane Lathuiliere, and Nicu Sebe. Deformable GANs for pose-based human image generation. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.243, + 0.892, + 0.295 + ], + "angle": 0, + "content": "[53] Shih-Yang Su, Frank Yu, Michael Zollhoefer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.296, + 0.892, + 0.392 + ], + "angle": 0, + "content": "[54] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo MartinBrualla, Tomas Simon, Jason M. Saragih, Matthias Nießner, Rohit Pandey, S. Fanello, Gordon Wetzstein, Jun-Yan Zhu, Christian Theobalt, Maneesh Agrawala, Eli Shechtman, Dan B. Goldman, and Michael Zollhofer. State of the art on neural rendering. Computer Graphics Forum, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.393, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[55] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: image synthesis using neural textures. ACM Transactions on Graphics (TOG), 38, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[56] Garvita Tiwari, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Neural-gif: Neural generalized implicit functions for animating people in clothing. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.473, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[57] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Metaatrix: Learning animatable clothed human models from few depth images. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.527, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[58] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In European Conference on Computer Vision, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.62 + ], + "angle": 0, + "content": "[59] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In NeurIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.621, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[60] Zhou Wang, A. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13:600-612, 2004. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[61] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humaner: Free-viewpoint rendering of moving people from monocular video. ArXiv, abs/2201.04127, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.729, + 0.892, + 0.796 + ], + "angle": 0, + "content": "[62] Feng Xu, Yebin Liu, Carsten Stoll, James Tompkin, Gaurav Bharaj, Qionghai Dai, Hans-Peter Seidel, Jan Kautz, and Christian Theobalt. Video-based characters: creating new human performances from a multi-view video database. ACM SIGGRAPH, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.797, + 0.892, + 0.878 + ], + "angle": 0, + "content": "[63] Jae Shin Yoon, Duygu Ceylan, Tuanfeng Y. Wang, Jingwan Lu, Jimei Yang, Zhixin Shu, and Hyunjung Park. Learning motion-dependent appearance for high-fidelity rendering of dynamic humans from a single camera. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3397-3407, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.879, + 0.892, + 0.891 + ], + "angle": 0, + "content": "[64] Richard Zhang, Phillip Isola, Alexei A. Efros, E. Shechtman," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6559" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.132 + ], + "angle": 0, + "content": "and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. CVPR, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.468, + 0.214 + ], + "angle": 0, + "content": "[65] Yang Zheng, Ruizhi Shao, Yuxiang Zhang, Tao Yu, Zerong Zheng, Qionghai Dai, and Yebin Liu. Deepmulticap: Performance capture of multiple characters using sparse multiview cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6239-6249, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.216, + 0.468, + 0.268 + ], + "angle": 0, + "content": "[66] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. TPAMI, PP, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.27, + 0.468, + 0.31 + ], + "angle": 0, + "content": "[67] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. Cips3d: A 3d-aware generator of gans based on conditionally-independent pixel synthesis. ArXiv, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.311, + 0.468, + 0.361 + ], + "angle": 0, + "content": "[68] Zhen Zhu, Tengteng Huang, Baoguang Shi, Miao Yu, Bofei Wang, and Xiang Bai. Progressive pose attention transfer for person image generation. In CVPR, pages 2347-2356, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6560" + } + ] +] \ No newline at end of file diff --git a/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/6283bde0-3818-458b-8029-7eb693822789_origin.pdf b/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/6283bde0-3818-458b-8029-7eb693822789_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ac88e9385a310e26532c04c60e3e2aaf5748c8ba --- /dev/null +++ b/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/6283bde0-3818-458b-8029-7eb693822789_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:949ae04583b37dcd79e0b47bb436e616bb387d6567873df500009003895696e8 +size 9519497 diff --git a/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/full.md b/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5552c2d269dc37e491cdf81107d229b924a19981 --- /dev/null +++ b/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/full.md @@ -0,0 +1,320 @@ +# SurMo: Surface-based 4D Motion Modeling for Dynamic Human Rendering + +Tao Hu, Fangzhou Hong, Ziwei Liu +S-Lab, Nanyang Technological University, Singapore + +# Abstract + +Dynamic human rendering from video sequences has achieved remarkable progress by formulating the rendering as a mapping from static poses to human images. However, existing methods focus on the human appearance reconstruction of every single frame while the temporal motion relations are not fully explored. In this paper, we propose a new 4D motion modeling paradigm, SurMo, that jointly models the temporal dynamics and human appearances in a unified framework with three key designs: 1) Surface-based motion encoding that models 4D human motions with an efficient compact surface-based triplane. It encodes both spatial and temporal motion relations on the dense surface manifold of a statistical body template, which inherits body topology priors for generalizable novel view synthesis with sparse training observations. 2) Physical motion decoding that is designed to encourage physical motion learning by decoding the motion triplane features at timestep $t$ to predict both spatial derivatives and temporal derivatives at the next timestep $t + 1$ in the training stage. 3) 4D appearance decoding that renders the motion triplanes into images by an efficient volumetric surface-conditioned renderer that focuses on the rendering of body surfaces with motion learning conditioning. Extensive experiments validate the state-of-the-art performance of our new paradigm and illustrate the expressiveness of surface-based motion triplanes for rendering high-fidelity view-consistent humans with fast motions and even motion-dependent shadows. Our project page is at: https://taohuumd.github.io/projects/SurMo. + +# 1. Introduction + +Creating volumetric videos of human actors is required in many applications such as AR/VR, telepresence, video game and film character creation. Recent neural rendering methods [25, 43, 54, 61] have made great progress in generating free-viewpoint videos of humans from several sparse multi-view videos, which are simple yet effective compared with traditional graphics approaches [2, 3, 62]. At the core of the technique is to model pose- and time-varying appearances of dynamic humans. The motions of dressed humans are often expressed as the movement of body and a sequence of natural secondary motion of clothes, \( e.g., dy + +![](images/56431c814e4a15490b1d14a0b1ed9f67e6e1f8d110475da3c25b01daf272b112.jpg) +Figure 1. Given several sparse multi-view video sequences with estimated 3D body meshes, SurMo synthesizes subject-specific appearance. We specifically focus on the synthesis of plausible time-varying appearances by learning an effective 4D motion representation. + +namic movements of the T-shirt induced by dancing in Fig. 1. The secondary motion of clothes arises from intricate physical interactions with the body, typically changing over time, which leads to challenges for plausible rendering of dynamic humans. + +Most existing methods [25, 43, 61] are conditioned on static poses and use a pose-guided generator to synthesize time-varying appearances. However, the appearance of clothed humans undergoes complex geometric transformations induced not only by the static pose but also its dynamics, whereas the modeling of dynamics is often ignored. In addition, these methods are often focused on the 2D appearance reconstruction of every single frame while the temporal motion relations are not fully explored. Due to these issues, they fail to render plausible secondary motions of clothes, i.e., generating the same appearances for fast and slow motions. A key challenge of learning temporal dynamics lies in the requirement of a tremendous amount of training observations to construct a 4D motion volume. + +To solve these issues, we propose a new paradigm to learn time-varying appearances of the secondary motion from just several sparse viewpoint video sequences, which is achieved by jointly modeling the temporal motion dynamics and human appearances in a unified rendering framework based on an efficient surface-based motion + +representation. At the core of the paradigm is a feature encoder-decoder framework with three key components: 1) surface-based motion encoding; 2) physical motion decoding; and 3) 4D appearance decoding. + +Firstly, in contrast to existing pose-guided methods [25, 42, 43, 61] that focus on static poses as a conditional variable, we extract an expressive 4D motion input from the 3D body mesh sequences obtained from training video as our input, which includes both a static pose represented by a spatial 3D mesh and its temporal dynamics. Furthermore, we notice that the non-rigid deformations of garments typically occur around the body surface instead of in a 3D volume, and hence we propose to model human motions on the body surface by projecting the extracted spatial-temporal 4D motion input to the dense 2D surface UV manifold of a clothless body template (e.g., SMPL [28]). To model temporal clothing offsets, a motion encoder is employed to lift the clothless motion features into a motion triplane that encodes both spatial and temporal motion relations in a compact 3D triplane with time-varying dynamics conditioned. The triplane is defined in the surface $\mathbf{u} - \mathbf{v} - \mathbf{h}$ coordinate system, with $\mathbf{u} - \mathbf{v}$ to represent the motion of the clothless body template, and an extensional coordinate $\mathbf{h}$ to represent the secondary motion of clothes, i.e., the temporal clothing offsets are parameterized by a signed distance to the body surface. In this way, 4D motions can be effectively represented by a surface-based triplane. + +Secondly, we propose to physically model the spatial and temporal motion dynamics in the rendering network. Specifically, with the surface-based triplane conditioned at time $t$ , a motion decoder is employed to decode the motion triplanes to predict the motion at the next timestep $t + 1$ , i.e., spatial derivative of the motion physically corresponding to surface normal map and temporal derivatives corresponding to surface velocity. We illustrate the physical motion learning significantly improves the rendering quality. + +Thirdly, the motion triplanes are decoded into high-quality images at two stages: a volumetric surface-conditioned renderer that is focused on the rendering around human body surface and filters the query points far from the body surface for efficient volumetric rendering, and a geometry-aware super-resolution module for efficient high-quality image synthesis. + +We conduct a systematical analysis of how human appearances are affected by temporal dynamics, and it was observed that some baseline methods mainly generate pose-dependent appearances instead of time-varying appearances for free-view video generations. In addition, quantitative and qualitative experiments are performed on three datasets with a total of 9 subject sequences, including ZJU-MoCap [43], AIST++ [24], and MPII-RDDC [13], which validate the effectiveness of SurMo in different scenarios. + +In summary, our contributions are: + +1) A new paradigm for learning dynamic humans from videos that jointly models temporal motions and human appearances in a unified framework, and one of the early works that systematically analyzes how human appearances are affected by temporal dynamics. +2) An efficient surface-based triplane that encodes both spatial and temporal motion relations for expressive 4D motion modeling. +3) We achieve state-of-the-art results and show that our new paradigm is capable of learning high-fidelity appearances from fast motion sequences (e.g., AIST++ dance videos) or synthesizing motion-dependent shadows in challenging scenarios. + +# 2. Related Work + +Our method is closely related to many sub-fields of visual computing, and below we discuss a set of the work. + +3D Shape Representations. To capture detailed shapes of 3D objects, most recent papers utilize implicit functions [7, 8, 20, 33-35, 40, 41, 48-50, 56, 57, 65, 66] or point clouds [16, 31, 32] due to their topological flexibility. These methods aim to learn geometry from 3D datasets, whereas we synthesize human images of novel poses only from 2D RGB training images. + +Rendering Humans by 2D GANs. Some existing approaches render human avatars by neural image translation, i.e., they utilize GAN [10] networks to learn a mapping from poses (given in the form of renderings of a skeleton [4, 23, 45, 52, 59, 68], dense mesh [11, 26, 27, 36, 51, 59] or joint position heatmaps [1, 29, 30]) to human images [4, 59]. To improve temporal stability, some methods [17, 44, 46, 55] propose to utilize the SMPL [28] priors for pose-guided generations. However, these methods do not reconstruct geometry explicitly and cannot handle self-occlusions effectively. + +Rendering Humans by 3D-aware Renderer. For stable view synthesis, recent papers [6, 25, 38, 42, 43, 53] propose to unify geometry reconstruction with view synthesis by volume rendering, which, however, is computationally heavy. To solve this issue, some recent 3D-GAN papers [5, 12, 15, 18, 19, 37, 39, 67] propose a hybrid rendering strategy for efficient geometry-aware rendering, that is, render low-resolution volumetric features for geometry learning, and employ a super-resolution module for high-resolution image synthesis. We adopt this strategy for efficient rendering, whereas ours is distinguished by rendering articulated humans with 4D motion modeling. + +UV-based Pose Representation. Some methods [18, 19, 31, 32, 47, 63] propose to project 3D posed meshes into a 2D positional map for pose encoding, which can be used for different downstream tasks, such as 3D reconstruction and novel view synthesis. [31, 32] rely on 3D supervision to learn 3D reconstructions, and they do not take as input the + +dynamics. To improve the rendering quality, [18, 47] proposes to utilize additional driving views as input for faithful rendering in telepresence applications. However, they do not explore how to learn temporal dynamics from pose sequences. By taking as input a 3D normal and velocity map, [63] encodes motion dynamics in their rendering network, whereas they do explicitly learn motions, i.e., predicting the motion status at the next timestep, and besides, [63] is a 2D rendering method, and they do not explicitly learn motions. + +# 3. Methodology + +# 3.1. Problem Setup + +Given a sparse multi-view video of a clothed human in motion and corresponding 3D pose estimations $\{\mathbf{P}_0, \dots, \mathbf{P}_t\}$ , our goal is to synthesize time-varying appearances of the individual under novel views. Existing methods [25, 42, 43, 61] formulate the problem of human rendering as learning a representation via a feature encoder-decoder framework: + +$$ +\mathbf {f} _ {\mathbf {t}} = \mathcal {E} _ {\mathcal {P}} \left(\mathbf {P} _ {\mathbf {t}}, \mathbf {z} _ {\mathbf {t}}\right) \tag {1} +$$ + +$$ +\mathbf {I} _ {\mathbf {t}} = \mathcal {D} _ {\mathcal {R}} (\mathbf {f} _ {\mathbf {t}}) +$$ + +where a pose encoder $\mathcal{E}_{\mathcal{P}}$ takes as input a representation of posed body $\mathbf{P_t}$ (e.g., 2D or 3D keypoints or 3D body mesh vertices) and a timestamp embedding $\mathbf{z_t}$ at time $t$ , and outputs intermediate pose-dependent features $\mathbf{f_t}$ that can be rendered by a decoder $\mathcal{D}_{\mathcal{R}}$ to reconstruct the appearance images $\mathbf{I_t}$ of the corresponding pose at time $t$ . The time embedding $\mathbf{z_t}$ often serves as a residual that is updated passively in backpropagation by image reconstruction loss. + +However, there are two issues with the above well-adopted paradigm. First, the appearance of clothed humans undergoes complex geometric transformations induced not only by the static pose $\mathbf{P}_{\mathbf{t}}$ but also its dynamics, whereas the modeling of dynamics is ignored in E.q. 1, and the residual $\mathbf{z}_{\mathbf{t}}$ cannot expressively model physical dynamics neither. Second, existing methods focus on the per-image reconstruction while ignoring the temporal relations of a motion sequence in training, i.e., sampling input pose $\mathbf{P}_{\mathbf{t}}$ and timestamp $\mathbf{z}_{\mathbf{t}}$ individually, and supervising image reconstructions frame-by-frame, partly because defining temporal supervision in the 2D image space is challenging, especially for articulated humans that often suffer from misalignment problems due to pose estimation errors. + +To solve these issues, we propose a new paradigm for learning view synthesis of dynamic humans from video sequences, which is formulated as: + +$$ +\mathbf {f} _ {\mathbf {t}} = \mathcal {E} _ {\mathcal {M}} (\mathbf {P} _ {\mathbf {t}}, \mathbf {D} _ {\mathbf {t}}) +$$ + +$$ +\frac {\partial \mathbf {P} _ {\mathbf {t} + 1}}{\partial \mathbf {x}}, \frac {\partial \mathbf {P} _ {\mathbf {t} + 1}}{\partial \mathbf {t}} = \mathcal {D} _ {\mathcal {M}} (\mathbf {f} _ {\mathbf {t}}) \tag {2} +$$ + +$$ +\mathbf {I} _ {\mathbf {t}} = \mathcal {D} _ {\mathcal {R}} (\mathbf {f} _ {\mathbf {t}}) +$$ + +where the input is represented by dynamic motions including a static pose $\mathbf{P}_{\mathbf{t}}$ and its physical dynamics $\mathbf{D}_{\mathbf{t}}$ at time $t$ , and a motion encoder $\mathcal{E}_{\mathcal{M}}$ is employed to encode the inputs into motion-dependent features (Sec. 3.2). Besides, in the training stage, a physical motion decoder $\mathcal{D}_{\mathcal{M}}$ is introduced to enforce the learning of spatial and temporal relations by decoding the intermediate feature $\mathbf{f}_{\mathbf{t}}$ to predict the spatial derivatives at $t + 1$ physically corresponding to surface normal $\mathbf{N}_{\mathbf{t} + 1} = \frac{\partial\mathbf{P}_{\mathbf{t} + 1}}{\partial\mathbf{x}}$ , and temporal derivatives at $t + 1$ corresponding to surface velocity $\mathbf{V}_{\mathbf{t} + 1} = \frac{\partial\mathbf{P}_{\mathbf{t} + 1}}{\partial\mathbf{t}}$ (Sec. 3.3). $\mathbf{f}_{\mathbf{t}}$ is rendered into human images by a decoder $\mathcal{D}_{\mathcal{R}}: \mathcal{G}_2 \circ \mathcal{G}_1$ (Sec. 3.5). The framework is depicted in Fig. 2. + +Notation. In the following parts, we use $\mathbf{f}_{\mathbf{t}}^{\mathrm{space}}$ to denote the features in the pipeline, i.e., $\mathbf{f}_{\mathbf{t}}^{3\mathbf{D}}$ is a 4D representation that is defined in 3D space with temporal $t$ conditioned, and $\mathbf{M}_{\mathbf{t}}^{3\mathbf{D}}$ is 4D motion input defined in 3D space with $t$ conditioned. + +# 3.2. Surface-based 4D Motion Encoding + +Extracting 4D Motions. We first extract 4D motions from a sequence of time-varying parametric posed body (e.g., SMPL) meshes obtained from training video sequences. We describe the motion $\mathbf{M}_{\mathbf{t}}^{3\mathbf{D}}$ at time $t$ as a static skeleton pose $\mathbf{P}_{\mathbf{t}}$ and its dynamics $\mathbf{D}_{\mathbf{t}}$ . The dynamics at $t$ are physically determined by the current pose, 3D velocity, and motion trajectory of the past several timesteps, which also contribute to the time-varying appearances of the secondary motion. The pose $\mathbf{P}_{\mathbf{t}}$ is represented by the 3D vertices of the posed mesh, and dynamics $\mathbf{D}_{\mathbf{t}}$ are parameterized by 1) body surface velocity $\mathbf{V}_{\mathbf{t}}$ corresponding to the temporal derivatives of the current pose at $t$ , and 2) motion trajectory $\mathbf{T}_{\mathbf{t}}$ that aggregates the temporal derivatives over the past several timesteps with a sliding window size of $w$ with weight $\lambda$ : + +$$ +\begin{array}{l} \mathbf {M} _ {\mathbf {t}} ^ {\mathrm {3 D}} = \left[ \mathbf {P} _ {\mathbf {t}}, \mathbf {D} _ {\mathbf {t}} \right], \quad \mathbf {D} _ {\mathbf {t}} = \left[ \mathbf {V} _ {\mathbf {t}}, \mathbf {T} _ {\mathbf {t}} \right] \\ \mathbf {V} _ {\mathbf {t}} = \frac {\partial \mathbf {P} _ {\mathbf {t}}}{\partial \mathbf {t}}, \quad \mathbf {T} _ {\mathbf {t}} = \mathbf {P} _ {\mathbf {t}} + \frac {1}{\sum \lambda_ {i}} \sum_ {i = 1} ^ {w} \mathbf {V} _ {\mathbf {t} - \mathbf {i}} * \lambda_ {i} \tag {3} \\ \end{array} +$$ + +Note the motion trajectory aggregated from several consecutive timesteps makes the motion representation robust to pose estimation errors, i.e., the pose estimations for two consecutive timesteps may be the same due to pose estimation errors in practice. An ablation study of the motion trajectory representation can be found in the supp. mat. + +Recording 4D Motions on the Surface Manifold. Modeling the motions with temporal dynamics often requires dense observations to construct a 4D motion volume. Instead, we notice that non-rigid deformations of human geometry in motion typically occur around the body surface instead of a 3D volume, and hence we propose to model the motions on the human body surface. Specifically, we project the 4D motion input $\mathbf{M}_{\mathbf{t}}^{3\mathrm{D}}$ including spatial pose + +![](images/5a621c6cf918e0f0a8ca49a4ed43b02af7ef62e02595360965c875188ed7d426.jpg) +Figure 2. Framework overview. Given a set of time-varying 3D body meshes $\{\mathbf{P}_{\mathbf{t}},\dots,\mathbf{P}_{\mathbf{t}} - \mathbf{n}\}$ obtained from training video sequences, we aim to synthesize high-fidelity appearances of a clothed human in motion via a feature encoder-decoder framework: Motion Encoding, and joint Motion and Appearance Decoding. 1) We take as input an expressive 4D motion representation extracted from the mesh sequences including 3D pose, 3D velocity at time $t$ , and motion trajectory over the past $w$ timesteps that encode both spatial and temporal relations of the motion sequence, which are projected to the spatially aligned UV surface space. A motion encoder $\mathcal{E}_{\mathcal{M}}$ is employed to lift the 2D UV-aligned features to a 3D surface-based triplane $\mathbf{f}_{\mathbf{t}}^{\mathrm{uvh}}$ in an UV-plus-height space with a signed distance height to model temporal clothing offsets. 2) A motion decoder $\mathcal{D}_{\mathcal{M}}$ is designed to encourage physical motion learning in training by decoding the triplane features $\mathbf{f}_{\mathbf{t}}^{\mathrm{uvh}}$ to predict the motion at the next timestep $t + 1$ , i.e. spatial derivatives surface normal $\mathbf{N}_{\mathbf{t} + 1}^{\mathrm{uv}}$ and temporal derivatives surface velocity $\mathbf{V}_{\mathbf{t} + 1}^{\mathrm{uv}}$ in UV space. 3) Finally, given a target camera view, the triplane $\mathbf{f}_{\mathbf{t}}^{\mathrm{uvh}}$ is rendered into high-quality images by a volumetric surface-conditioned renderer including volumetric low-resolution rendering by $\mathcal{G}_1$ and an efficient geometry-aware super-resolution by $\mathcal{G}_2$ . + +and temporal dynamics from 3D space into a compact spatially aligned UV space using the geometric transformation $\mathcal{W}$ that is pre-defined by the parametric (e.g., SMPL) body template, which yields $\mathbf{f}_{\mathbf{t}}^{\mathrm{uv}} = \mathcal{W}\mathbf{M}_{\mathbf{t}}^{3\mathrm{D}}$ . The UV-aligned motion feature $\mathbf{f}_{\mathbf{t}}^{\mathrm{uv}}$ faithfully preserves the articulated structures of body topology in a compact 2D space. + +Generating Surface-based Triplanes for Motion Modeling. To model clothed humans, we further employ a motion encoder $\mathcal{E}_{\mathcal{M}}$ that lifts the clothless 2D features $\mathbf{f}_{\mathbf{t}}^{\mathrm{uv}}$ to a 3D triplane representation $\mathbf{f}_{\mathbf{t}}^{\mathrm{uvh}}$ that is defined in a $\mathbf{u} - \mathbf{v} - \mathbf{h}$ system, with $\mathbf{u} - \mathbf{v}$ to represent the motion of the clothless body template, and an extensional coordinate $\mathbf{h}$ to represent the secondary motion of clothes, i.e., the temporal cloth offsets, and hence 4D clothed motions can be parameterized by a surface-based triplane by: + +$$ +\mathbf {f} _ {\mathbf {t}} ^ {\mathrm {u v h}} = \mathcal {E} _ {\mathcal {M}} \left(\mathbf {f} _ {\mathbf {t}} ^ {\mathrm {u v}}\right) = \mathcal {E} _ {\mathcal {M}} \left(\mathcal {W} \mathbf {M} _ {\mathbf {t}} ^ {\mathbf {3 D}}\right), \tag {4} +$$ + +where $\mathcal{W}$ is the geometric transformation from 3D space to UV space. $\mathbf{f}_{\mathbf{t}}^{\mathrm{uvh}}$ consists of three planes $x_{uv}, x_{uh}, x_{hv} \in \mathbb{R}^{U \times V \times C}$ to form the spatial relationship, where $U, V$ denote spatial resolution and $C$ is the channel number. In contrast to the volumetric triplanes used in EG3D [5] where the three planes represent three vertical planes in the 3D volumetric space, $\mathbf{f}_{\mathbf{t}}^{\mathrm{uvh}}$ is defined on the human body surface inheriting human topology priors for motion modeling, as illustrated by an unwarped triplane in 3D space in Fig 2. + +# 3.3. Physical Motion Decoding + +We propose a physical motion decoder $\mathcal{D}_{\mathcal{M}}$ to learn spatial and temporal features of motion, which is achieved by decoding the intermediate motion feature $\mathbf{f}_{\mathbf{t}}^{\mathrm{uvh}}$ learned by $\mathcal{E}_{\mathcal{M}}$ + +to predict the motions at the next timestep $t + 1$ , such as spatial derivatives corresponding to surface normal $\mathbf{N}_{t+1}^{\mathrm{uv}}$ , temporal derivatives corresponding to surface velocity $\mathbf{V}_{t+1}^{\mathrm{uv}}$ . Note that we model the normal and velocity in the UV space by: + +$$ +\left\{\mathbf {N} _ {\mathbf {t} + \mathbf {1}} ^ {\mathrm {u v}}, \mathbf {V} _ {\mathbf {t} + \mathbf {1}} ^ {\mathrm {u v}} \right\} = \mathcal {D} _ {\mathcal {M}} \left(\mathbf {f} _ {\mathbf {t}} ^ {\mathrm {u v h}}\right) = \mathcal {D} _ {\mathcal {M}} \circ \mathcal {E} _ {\mathcal {M}} \left(\mathcal {W} \mathbf {M} _ {\mathbf {t}} ^ {\mathbf {3 D}}\right) (5) +$$ + +# 3.4. 4D Appearance Decoding + +Volumetric Surface-conditioned Rendering. Given a target camera viewpoint, the triplane $\mathbf{f}_{\mathbf{t}}^{\mathrm{uvh}}$ is first rendered into low-resolution volumetric features $\mathbf{I}_{\mathbf{F}} \in \mathbb{R}^{H \times W \times \hat{C}}$ by a conditional NeRF $\mathbb{F}_{\Phi}$ , where $H$ and $W$ are the spatial resolution and $\hat{C}$ is the channel number. To be more specific, given a 3D query point $\pmb{p}_i$ , we transform it into a surface-based local coordinate $\hat{\pmb{p}}_i = (\pmb{u}_i, \pmb{v}_i, \pmb{h}_i)$ w.r.t. the tracked body mesh $\mathbf{P}_{\mathbf{t}}$ . Here we search the nearest face $\pmb{f}_i$ of $\mathbf{P}_{\mathbf{t}}$ for query point $\pmb{p}_i$ and $(\pmb{u}_i, \pmb{v}_i)$ and $\pmb{h}_i$ are the barycentric coordinates of the nearest point on $\pmb{f}_i$ , and the signed distance respectively. We therefore obtain the local feature $\pmb{z}_i^{\mathrm{uvh}}$ of the query point $\pmb{p}_i$ as + +$$ +\boldsymbol {z} _ {i} ^ {u v h} = \operatorname {c a t} \left[ \Pi \left(\boldsymbol {x} _ {u v}; \boldsymbol {u} _ {i}, \boldsymbol {v} _ {i}\right), \Pi \left(\boldsymbol {x} _ {u h}; \boldsymbol {u} _ {i}, \boldsymbol {h} _ {i}\right), \Pi \left(\boldsymbol {x} _ {h v}; \boldsymbol {h} _ {i}, \boldsymbol {v} _ {i}\right) \right], \tag {6} +$$ + +where $\Pi (\cdot)$ denotes sampling operation, $\mathrm{cat}[\cdot ]$ denotes the concatenation operator. + +Given camera direction $\pmb{d}_i$ , the appearance features $\pmb{c}_i$ and density features $\pmb{\sigma}_i$ of point $\pmb{p}_i$ are predicted by + +$$ +\left\{\boldsymbol {c} _ {i}, \boldsymbol {\sigma} _ {i} \right\} = \mathbb {F} _ {\Phi} \left(\boldsymbol {d} _ {i}, \boldsymbol {z} _ {i} ^ {u v h}\right) \tag {7} +$$ + +We integrate all the radiance features of sampled points into a 2D feature map $I_{F}$ at time $t$ through volume renderer + +![](images/88c4ec5eef522c9e383384ba91eea32c81698ccdf890a3fe5474f390535bf409.jpg) +Figure 3. Qualitative comparisons on novel view synthesis on the subject S313 of ZJU-MoCap dataset. Two motion sequences S1 (swing arms left to right) and S2 (raise and lower arms) are shown. We specifically focus on the synthesis of time-varying appearances (especially T-shirt wrinkles), by evaluating the rendering results under similar poses yet with different movement directions, which are marked in the same color, such as the pairs of $①②$ , $③④$ , and $⑤⑥$ . Our method synthesizes high-fidelity time-varying appearances, whereas SOTA HumanNeRF generates almost the same cloth wrinkles. + +$\mathcal{G}_1$ [21] + +$$ +\mathbf {I} _ {\mathbf {F}} = \mathcal {G} _ {1} \left(\mathbf {f} _ {\mathbf {t}} ^ {\mathbf {u v h}}, c a m; \mathbb {F} _ {\Phi}\right) \tag {8} +$$ + +where $\mathbf{f}_{\mathrm{t}}^{\mathrm{uvh}}$ is extracted from Eq. 4. + +Efficient Geometry-Aware Super-Resolution. Sampling dense points to render the full-resolution volumetric features is computationally heavy. Instead, we employ a superresolution network $\mathcal{G}_2$ [5] to render high-resolution images $\mathbf{I}_{\mathbf{RGB}}^{+} \in \mathbb{R}^{\hat{H} \times \hat{W} \times 3}$ : + +$$ +\mathbf {I} _ {\mathbf {R G B}} ^ {+} = \mathcal {G} _ {2} \circ \mathcal {G} _ {1} \left(\mathbf {f} _ {\mathbf {t}} ^ {\mathbf {u v h}}, c a m; \mathbb {F} _ {\Phi}\right) \tag {9} +$$ + +# 3.5. Optimization + +SurMo is trained end-to-end to optimize $\mathcal{E}_{\mathcal{M}}$ , $\mathcal{D}_{\mathcal{M}}$ , and renderers $\mathcal{G}_1$ , $\mathcal{G}_1$ with 2D image loss. We employ Adversarial Loss and Reconstruction Loss including Pixel Loss, + +Perceptual Loss, Face Identity Loss, Velocity and Normal Loss, and Volume Rendering Loss for supervision in training, with Adam [22] as the optimizer. Refer to the supp. mat. for more details. + +# 4. Experiments + +Dataset and Metrics. We evaluate the novel view synthesis on ZJU-MoCap, MPII-RDDC, and AIST++, with a total of 9 subjects. We use the same camera setup as Neural Body, that is, 4 cameras used for training, and the other 18 for testing. For MPII-RDDC, 18 cameras in training, and 9 for testing. Since ASIT++ only has 9 cameras, we use 6 for training, and the remaining 3 for testing. Refer to more details in the supp. mat. We compare each method on per-pixel metrics including SSIM [60] and PSNR, and + +![](images/0d2e47398f1fb0efc17a0c34bcfbc14ef3610b57d9ac2bff4211c7c03e472755.jpg) +Neural Body Instant-NVR HumanNeRF Ours GT + +![](images/4fb3ce63b4f819b49edb8d72bdfc6828627c391d5ad7c8f53e2ede247011c197.jpg) +Neural Body Instant-NVR HumanNeRF Ours GT +Figure 4. Qualitative comparisons on novel view synthesis on the subject S387, S315 of ZJU-MoCap dataset. Row 1 and 2 show similar poses occurring at different timesteps (not consecutive frames). The results indicate that our method synthesizes time-varying appearances while other methods mainly generate pose-dependent appearances. + +Table 1. Quantitative comparisons on ZJU-MoCap dataset (averaged on all test views and poses on 6 sequences) for novel-view synthesis. To reduce the influence of the background, all scores are calculated from images cropped to 2D bounding boxes. Note that the perception metrics LPIPS [64] and FID [14] capture human judgment better than per-pixel metrics such as SSIM [60] or PSNR, as stated in [25, 51]. + +
ZJU-S1-6LPIPS ↓FID ↓SSIM ↑PSNR ↑
Neural Body [43].164125.10.70321.438
Instant-NVR [9].202144.99.73821.748
HumanNeRF [61].10688.382.79223.624
Ours.07567.725.83324.815
+ +perception metrics including LPIPS [64] and FID [14]. + +Baselines. We compare our method against SOTA methods including Neural Body [43], HumanNeRF [61], Instant-NVR [9], ARAH [58], DVA [47] and HVTR++ [18]. Neural Body models human poses in 3D space with point cloud as pose representation, HumanNeRF and Instant-NVR model poses in a canonical space with inverse skinning, ARAH is a forward-skinning based method, DVA and HVTR++ encode poses in UV space, and take driving view signals as input for telepresence applications. + +# 4.1. Comparisons to SOTA Methods + +Quantitative comparisons We conduct the quantitative comparisons on the three datasets with a total of 9 subjects. The quantitative results on ZJU-MoCap are summarized in Tab. 1, which suggests that our new paradigm outperforms the SOTA by a big margin, and we achieve the best quantitative results on all four metrics. The detailed compar- + +isons on each subject of ZJU-MoCap, and quantitative results on MPII-RDDC and AIST++ are listed in the supp. mat. Refer to the comparisons with ARAH [58], DVA [47] and HVTR++ [18] in the supp. mat. + +Time-varying Appearances with Dynamics Conditioning on ZJU-MoCap. We compare with baseline methods for novel view synthesis on two motion sequences of subject S313 from the ZJU-MoCap dataset, as shown in Fig. 3. We evaluate the capability of each method to synthesize time-varying appearances. Specifically, two sequences S1 (swing arms left to right) and S2 (raise and lower arms) are evaluated, where similar poses occur at different timesteps with different dynamics (e.g., movement directions, trajectory, or velocity) are marked in the same color, such as the pairs of $①②$ , $③④$ , and $⑤⑥$ . Fig. 3 suggests that Instant-NVR [9] fails to synthesize time-varying high-frequency wrinkles, and Neural Body [43] synthesizes time-varying yet blurry wrinkles. HumanNeRF [61] synthesizes detailed yet static T-shirt wrinkles, i.e., the wrinkles are almost the same for similar poses, such as $①②$ , $③④$ , and $⑤⑥$ . In contrast, our method renders both high-fidelity and time-varying wrinkles. The comparisons on other subjects S387 and S315 are shown in Fig. 4, which illustrate the effectiveness of our proposed paradigm in dynamics learning. + +Time-varying Appearances with Motion-dependent Shadows on MPII-RDDC. We compare with baseline methods for novel view synthesis on MPII-RDDC, as shown in Fig. 5. The sequence is captured in a studio with top-down lighting that casts shadows on the human body due to self-occlusions. We notice that the synthesis of shadow can be formulated as learning motion-dependent shadows in the motion representation without the need to + +![](images/bcd6ca47faffd07096384f8eaac4a158e38e6a378315e2a759538a443c396b69.jpg) +HumanNeRF +Ours +Ground Truth +HumanNeRF +Ours +Ground Truth + +![](images/2e3ed2af846bde22715ff3cada8a05944270991c71e47028fd8f6166e92449a9.jpg) +Figure 5. Novel view synthesis of time-varying appearances with both pose and lighting conditioning on MPII-RDDC dataset. The sequence is captured in a studio with top-down lighting that casts shadows on the human performer due to self-occlusion. In Row 1, we specifically focus on synthesizing time-varying shadows (e.g., ① vs. ②, and ③ vs. ④) for different poses with different self-occlusions. In Row 2, we evaluate the synthesis of: 1) time-varying appearances for similar poses occurring in a jump-up-and-down motion sequence, e.g., ⑤ vs. ⑥, 2) shadows ⑦ vs. ⑧, and 3) clothing offsets ⑤ vs. ⑥. +Figure 6. Novel view synthesis of fast motions on AIST++. + +explicitly model the lighting. We validate our method from two aspects of the scenario in Fig. 5. 1) In Row 1, Fig. 5 indicates that with the expressive motion representation and physical motion learning, SurMo succeeds in predicting the motion-dependent shadows, such as $①②③④$ , whereas SOTA HumanNeRF renders almost the same T-shirt appearance. 2) In Row 2, we compare the synthesis of time-varying appearances for similar poses occurring in a jump-up-and-down motion sequence i.e., $⑤$ vs. $⑥$ , which shows that SurMo is capable of predicting the clothes offsets under similar pose with different motion trajectories. In addition, we synthesize the shadows $⑦⑧$ in the challenging jump-up-and-down motion sequence. However, HumanNeRF fails to predict the dynamic clothing offsets $⑨$ . + +Time-varying Appearances for Fast Motions on AIST++. We also evaluate our methods by rendering humans with fast dance motions (S21 and S13) on AIST++, as shown in Fig. 6, where we compare ours against Neural Body in synthesizing time-varying wrinkles of the T-shirt for different motions. Fig. 6 suggests that Neural Body can only synthesize blurry results partly because Neural Body models pose in a sparse 3D space. Instead, we model motions in a denser body surface space, which enables high-fidelity image synthesis. Note that AIST++ is based + +![](images/06d2a2cc665354948ad640d0d01262dfd556dcde6bbbfd86cc42038b4b418641.jpg) +Figure 8. Ablation study of motion conditioning and learning. We focus on the effect of motion conditioning and learning in Row 1, and whether our method learns to decouple the static pose and dynamics (e.g., velocity) from the motion conditioning. + +on SMPL tracking with a scaling factor that affects the inverse LBS used in Instant-NVR or HumanNeRF both relying on inverse LBS, and hence we cannot compare them based on the released official code. Quantitative results can be found in the supp. mat. + +# 4.2. Ablation Study + +Surface-based Triplane vs. Volumetric Triplane. We compare the well-adopted volumetric triplane (Vol-Trip) [5] and our proposed surface-based triplane (Surf-Trip) for human modeling on two subjects S313 and S387 from ZJUMoCap, as shown in Fig. 7. Fig. 7 (left) suggests that the Surf-Trip converges faster in training with a smaller reconstruction loss for both sequences, and the performances are further improved with physical motion learning, e.g., 'Surf-Trip + ML'. Fig. 7 (right) compares the rendering re + +![](images/93fc99a370db7c1e669130b0279e6ed2c28e9220a804220058633a7d8c32d74a.jpg) +Figure 7. Ablation study of 3D volumetric triplane (Vol-Trip) vs. surface-based triplane (Surf-Trip) for human modeling on subject S313 and S387 from ZJU-MoCap. We focus on the convergence in training (left), and self-occlusion rendering (right). The convergence of S313 is shifted for visualization purposes. Vol-Trip is not effective in handling self-occlusion, e.g., $①②③④$ though it performs well in another viewpoint without self-occlusions $⑥$ . In addition, Vol-Trip cannot synthesize high-quality details for face. + +![](images/2d8e22cbf2bb7793f4aa369875a8c53f83a80eda8f3684077af6967db514c0d7.jpg) + +Table 2. Ablation study of motion conditioning and learning. $P_{cond}$ and $D_{cond}$ denote the conditioning of pose and dynamics, $V_{pred}$ and $N_{pred}$ denote the prediction of surface velocity and surface normal in motion learning. + +
S313LPIPS ↓FID ↓SSIM ↑PSNR ↑
Pcond.120104.83.79322.781
+ Dcond.08573.674.83424.908
+ Vpred.06962.092.85625.845
+ Npred.06050.170.86926.654
+ +sults qualitatively, which indicates that Surf-Trip is more effective in handling self-occlusions, whereas Vol-Trip fails $①②③④$ , though it performs well in another viewpoint without self-occlusions $⑥$ . In addition, Surf-Trip generates high-fidelity clothing wrinkles and facial details, whereas the face rendering of Vol-Trip is blurry. + +Motion Conditioning and Learning. We also analyze the effectiveness of our proposed motion conditioning and learning qualitatively and quantitatively. Tab. 2 summarizes the quantitative results, which suggest that the quantitative results are significantly improved by taking as input the dynamics conditioning $D_{cond}$ . With physical motion learning, i.e., predicting the temporal derivatives surface velocity and spatial derivates normal at the next timestep, the quantitative results are further improved by a big margin. + +The qualitative comparisons are shown in Fig. 8, where in Row 1, we observe higher-fidelity appearances with dynamics conditioning, and physical motion learning for both velocity and normal prediction. In Row 2, we evaluate whether our method learns to decouple the temporal dynamics (e.g., velocity) and static poses from motion conditioning, which corresponds to the rendering of clothing wrinkles of secondary motion (T-shirt ①) and tight parts (e.g., shorts ②). We only change the velocity for each variant + +in the ablation study. It suggests that the appearance of the T-shirt varies when the velocity or dynamics are changed, whereas the renderings of tight parts remain the same, such as the head, tight shorts, and shoes. This is consistent with our daily observations. The ablation study illustrates that our method decouples the pose and dynamics, and is capable of generating dynamics-dependent appearances. Note that the appearance changes are small between $V * 1$ and $V * 0.01$ when we reduce the velocity, partly because the method is not sensitive to smaller velocity due to the pose estimation errors in the training dataset where consecutive frames may be estimated with the same pose fitting. + +# 5. Discussion + +We propose SurMo, a new paradigm for learning dynamic humans from videos by jointly modeling temporal motions and human appearances in a unified framework, based on an efficient surface-based triplane. We conduct a systematical analysis of how human appearances are affected by temporal dynamics, and extensive experiments validate the expressiveness of the surface-based triplane in rendering fast motions and motion-dependent shadows. Quantitative experiments illustrate that SurMo achieves the SOTA results on different motion sequences. + +# Acknowledgement + +This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +[1] Kfir Aberman, M. Shi, Jing Liao, Dani Lischinski, B. Chen, and D. Cohen-Or. Deep video-based performance cloning. Computer Graphics Forum, 38, 2019. 2 +[2] George Borshukov, Dan Piponi, Oystein Larsen, J. P. Lewis, and Christina Tempelaar-Lietz. Universal capture: image-based facial animation for "the matrix reloaded". In SIGGRAPH '03, 2003. 1 +[3] Joel Carranza, Christian Theobalt, Marcus A. Magnor, and Hans-Peter Seidel. Free-viewpoint video of human actors. ACM SIGGRAPH 2003 Papers, 2003. 1 +[4] Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A. Efros. Everybody dance now. ICCV, pages 5932-5941, 2019. 2 +[5] Eric Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, S. Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. ArXiv, abs/2112.07945, 2021. 2, 4, 5, 7 +[6] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Linchao Bao, and Huchuan Lu. Animatable neural radiance fields from monocular rgb video. ArXiv, abs/2106.13629, 2021. 2 +[7] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. CVPR, 2019. 2 +[8] Boyang Deng, John P Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. Nasa neural articulated shape approximation. In ECCV, 2020. 2 +[9] Chen Geng, Sida Peng, Zhen Xu, Hujun Bao, and Xiaowei Zhou. Learning neural volumetric representations of dynamic humans in minutes. In CVPR, 2023. 6 +[10] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014. 2 +[11] A. K. Grigor'ev, Artem Sevastopolsky, Alexander Vakhitov, and Victor S. Lempitsky. Coordinate-based texture inpainting for pose-guided human image generation. CVPR, pages 12127-12136, 2019. 2 +[12] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. ArXiv, abs/2110.08985, 2021. 2 +[13] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics (TOG), 40:1-16, 2021. 2 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, 2017. 6 +[15] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. ArXiv, abs/2112.05637, 2021. 2 +[16] Tao Hu, Geng Lin, Zhizhong Han, and Matthias Zwicker. Learning to generate dense point clouds with textures on multiple categories. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), + +pages 2170-2179, January 2021. 2 +[17] Tao Hu, Kripasindhu Sarkar, Lingjie Liu, Matthias Zwicker, and Christian Theobalt. Egorenderer: Rendering human avatars from egocentric camera images. In ICCV, 2021. 2 +[18] Tao Hu, Hongyi Xu, Linjie Luo, Tao Yu, Zerong Zheng, He Zhang, Yebin Liu, and Matthias Zwicker. Hvtr++: Image and pose driven human avatars using hybrid volumetric-textural rendering. IEEE Transactions on Visualization and Computer Graphics, pages 1-15, 2023. 2, 3, 6 +[19] T. Hu, Tao Yu, Zerong Zheng, He Zhang, Yebin Liu, and Matthias Zwicker. Hvtr: Hybrid volumetric-textural rendering for human avatars. 3DV, 2022. 2 +[20] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. 2020 (CVPR), pages 3090-3099, 2020. 2 +[21] James T. Kajiya and Brian Von Herzen. Ray tracing volume densities. Proceedings of the 11th annual conference on Computer graphics and interactive techniques, 1984. 5 +[22] Diederick P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5 +[23] Bernhard Kratzwald, Zhiwu Huang, Danda Pani Paudel, and Luc Van Gool. Towards an understanding of our world by GANing videos in the wild. arXiv:1711.11453, 2017. 2 +[24] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Learn to dance with aist++: Music conditioned 3d dance generation, 2021. 2 +[25] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. TOG, 40, 2021. 1, 2, 3, 6 +[26] Lingjie Liu, Weipeng Xu, Marc Habermann, Michael Zollhöfer, Florian Bernard, Hyeongwoo Kim, Wenping Wang, and Christian Theobalt. Neural human video rendering by learning dynamic textures and rendering-to-video translation. IEEE Transactions on Visualization and Computer Graphics, 05 2020. 2 +[27] Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Hyeongwoo Kim, Florian Bernard, Marc Habermann, Wenping Wang, and Christian Theobalt. Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG), 2019. 2 +[28] M. Loper, Naureen Mahmood, J. Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multi-person linear model. ACM Trans. Graph., 34:248:1-16, 2015. 2 +[29] Liqian Ma, Xu Jia, Qianru Sun, Bernt Schiele, Tinne Tuytelaars, and Luc Van Gool. Pose guided person image generation. In NeurIPS, pages 405-415, 2017. 2 +[30] Liqian Ma, Qianru Sun, Stamatios Georgoulis, Luc van Gool, Bernt Schiele, and Mario Fritz. Disentangled person image generation. CVPR, 2018. 2 +[31] Qianli Ma, Shunsuke Saito, Jinlong Yang, Siyu Tang, and Michael J. Black. Scale: Modeling clothed humans with a surface codec of articulated local elements. In CVPR, 2021. 2 +[32] Qianli Ma, Jinlong Yang, Siyu Tang, and Michael J Black. The power of points for modeling humans in clothing. In ICCV, 2021. 2 +[33] Lars M. Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy net + +works: Learning 3d reconstruction in function space. CVPR, 2019. 2 +[34] Mateusz Michalkiewicz, Zhony Kaesemodel Pontes, Dominic Jack, Mahsa Baktash, and Anders P. Eriksson. Deep level sets: Implicit surface representations for 3d shape inference. ArXiv, 2019. +[35] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. Leap: Learning articulated occupancy of people. In CVPR, 2021. 2 +[36] Natalia Neverova, Riza Alp Güler, and Iasonas Kokkinos. Dense pose transfer. ECCV, 2018. 2 +[37] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. CVPR, pages 11448-11459, 2021. 2 +[38] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In IEEE/CVF ICCV, 2021. 2 +[39] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. ArXiv, abs/2112.11427, 2021. 2 +[40] Pablo Palafox, Aljaž Božić, Justus Thies, Matthias Nießner, and Angela Dai. Npms: Neural parametric models for 3d deformable shapes. In IEEE/CVF ICCV, 2021. 2 +[41] Jeong Joon Park, Peter R. Florence, Julian Straub, Richard A. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. 2019 (CVPR), pages 165-174, 2019. 2 +[42] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021, 2, 3 +[43] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. CVPR, 2021. 1, 2, 3, 6 +[44] Sergey Prokudin, Michael J. Black, and Javier Romero. Smplpix: Neural avatars from 3d human models. WACV, 2021. 2 +[45] Albert Pumarola, Antonio Agudo, Alberto Sanfeliu, and Francesc Moreno-Noguer. Unsupervised person image synthesis in arbitrary poses. In CVPR, June 2018. 2 +[46] Amit Raj, Julian Tanke, James Hays, Minh Vo, Carsten Stoll, and Christoph Lassner. Anr: Articulated neural rendering for virtual avatars. CVPR, pages 3721-3730, 2021. 2 +[47] Edoardo Remelli, Timur M. Bagautdinov, Shunsuke Saito, Chenglei Wu, Tomas Simon, Shih-En Wei, Kaiwen Guo, Zhe Cao, Fabian Prada, Jason M. Saragih, and Yaser Sheikh. Drivable volumetric avatars using texel-aligned features. ACM SIGGRAPH, 2022. 2, 3, 6 +[48] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. IEEE/CVF ICCV, pages 2304-2314, 2019. 2 +[49] Shunsuke Saito, Tomas Simon, Jason M. Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. 2020 (CVPR), pages 81-90, 2020. + +[50] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J. Black. Scintimate: Weakly supervised learning of skinned clothed avatar networks. 2021 (CVPR), pages 2885-2896, 2021. 2 +[51] Kripasindhu Sarkar, Dushyant Mehta, Weipeng Xu, Vladislav Golyanik, and Christian Theobalt. Neural re- rendering of humans from a single image. In ECCV, 2020. 2,6 +[52] Aliaksandr Siarohin, Enver Sangineto, Stephane Lathuiliere, and Nicu Sebe. Deformable GANs for pose-based human image generation. In CVPR, 2018. 2 +[53] Shih-Yang Su, Frank Yu, Michael Zollhoefer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In NeurIPS, 2021. 2 +[54] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo MartinBrualla, Tomas Simon, Jason M. Saragih, Matthias Nießner, Rohit Pandey, S. Fanello, Gordon Wetzstein, Jun-Yan Zhu, Christian Theobalt, Maneesh Agrawala, Eli Shechtman, Dan B. Goldman, and Michael Zollhofer. State of the art on neural rendering. Computer Graphics Forum, 2020. 1 +[55] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: image synthesis using neural textures. ACM Transactions on Graphics (TOG), 38, 2019. 2 +[56] Garvita Tiwari, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Neural-gif: Neural generalized implicit functions for animating people in clothing. In ICCV, 2021. 2 +[57] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Metaatrix: Learning animatable clothed human models from few depth images. NeurIPS, 2021. 2 +[58] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In European Conference on Computer Vision, 2022. 6 +[59] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In NeurIPS, 2018. 2 +[60] Zhou Wang, A. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13:600-612, 2004. 5, 6 +[61] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humaner: Free-viewpoint rendering of moving people from monocular video. ArXiv, abs/2201.04127, 2022. 1, 2, 3, 6 +[62] Feng Xu, Yebin Liu, Carsten Stoll, James Tompkin, Gaurav Bharaj, Qionghai Dai, Hans-Peter Seidel, Jan Kautz, and Christian Theobalt. Video-based characters: creating new human performances from a multi-view video database. ACM SIGGRAPH, 2011. 1 +[63] Jae Shin Yoon, Duygu Ceylan, Tuanfeng Y. Wang, Jingwan Lu, Jimei Yang, Zhixin Shu, and Hyunjung Park. Learning motion-dependent appearance for high-fidelity rendering of dynamic humans from a single camera. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3397-3407, 2022. 2, 3 +[64] Richard Zhang, Phillip Isola, Alexei A. Efros, E. Shechtman, + +and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. CVPR, pages 586-595, 2018. 6 +[65] Yang Zheng, Ruizhi Shao, Yuxiang Zhang, Tao Yu, Zerong Zheng, Qionghai Dai, and Yebin Liu. Deepmulticap: Performance capture of multiple characters using sparse multiview cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6239-6249, 2021. 2 +[66] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. TPAMI, PP, 2021. 2 +[67] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. Cips3d: A 3d-aware generator of gans based on conditionally-independent pixel synthesis. ArXiv, 2021. 2 +[68] Zhen Zhu, Tengteng Huang, Baoguang Shi, Miao Yu, Bofei Wang, and Xiang Bai. Progressive pose attention transfer for person image generation. In CVPR, pages 2347-2356, 2019. 2 \ No newline at end of file diff --git a/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/images.zip b/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b4f3fbb4cacdf24ae463d4f31251423b46f5ace2 --- /dev/null +++ b/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:890de70f0ae082a7414a2f6bd6b30b4ff246a9091ad98e4d16504ed6daeb7897 +size 702817 diff --git a/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/layout.json b/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4dab908080ce77464029ec26ed0bfb1a91694e18 --- /dev/null +++ b/2024/SurMo_ Surface-based 4D Motion Modeling for Dynamic Human Rendering/layout.json @@ -0,0 +1,9908 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "type": "text", + "content": "SurMo: Surface-based 4D Motion Modeling for Dynamic Human Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 165, + 133, + 425, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 133, + 425, + 162 + ], + "spans": [ + { + "bbox": [ + 165, + 133, + 425, + 162 + ], + "type": "text", + "content": "Tao Hu, Fangzhou Hong, Ziwei Liu \nS-Lab, Nanyang Technological University, Singapore" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 191, + 192, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 191, + 192, + 204 + ], + "spans": [ + { + "bbox": [ + 143, + 191, + 192, + 204 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 211, + 290, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 211, + 290, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 290, + 547 + ], + "type": "text", + "content": "Dynamic human rendering from video sequences has achieved remarkable progress by formulating the rendering as a mapping from static poses to human images. However, existing methods focus on the human appearance reconstruction of every single frame while the temporal motion relations are not fully explored. In this paper, we propose a new 4D motion modeling paradigm, SurMo, that jointly models the temporal dynamics and human appearances in a unified framework with three key designs: 1) Surface-based motion encoding that models 4D human motions with an efficient compact surface-based triplane. It encodes both spatial and temporal motion relations on the dense surface manifold of a statistical body template, which inherits body topology priors for generalizable novel view synthesis with sparse training observations. 2) Physical motion decoding that is designed to encourage physical motion learning by decoding the motion triplane features at timestep " + }, + { + "bbox": [ + 46, + 211, + 290, + 547 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 211, + 290, + 547 + ], + "type": "text", + "content": " to predict both spatial derivatives and temporal derivatives at the next timestep " + }, + { + "bbox": [ + 46, + 211, + 290, + 547 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 46, + 211, + 290, + 547 + ], + "type": "text", + "content": " in the training stage. 3) 4D appearance decoding that renders the motion triplanes into images by an efficient volumetric surface-conditioned renderer that focuses on the rendering of body surfaces with motion learning conditioning. Extensive experiments validate the state-of-the-art performance of our new paradigm and illustrate the expressiveness of surface-based motion triplanes for rendering high-fidelity view-consistent humans with fast motions and even motion-dependent shadows. Our project page is at: https://taohuumd.github.io/projects/SurMo." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "Creating volumetric videos of human actors is required in many applications such as AR/VR, telepresence, video game and film character creation. Recent neural rendering methods [25, 43, 54, 61] have made great progress in generating free-viewpoint videos of humans from several sparse multi-view videos, which are simple yet effective compared with traditional graphics approaches [2, 3, 62]. At the core of the technique is to model pose- and time-varying appearances of dynamic humans. The motions of dressed humans are often expressed as the movement of body and a sequence of natural secondary motion of clothes, \\( e.g., dy" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 305, + 190, + 547, + 337 + ], + "blocks": [ + { + "bbox": [ + 305, + 190, + 547, + 337 + ], + "lines": [ + { + "bbox": [ + 305, + 190, + 547, + 337 + ], + "spans": [ + { + "bbox": [ + 305, + 190, + 547, + 337 + ], + "type": "image", + "image_path": "56431c814e4a15490b1d14a0b1ed9f67e6e1f8d110475da3c25b01daf272b112.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 338, + 547, + 392 + ], + "lines": [ + { + "bbox": [ + 305, + 338, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 547, + 392 + ], + "type": "text", + "content": "Figure 1. Given several sparse multi-view video sequences with estimated 3D body meshes, SurMo synthesizes subject-specific appearance. We specifically focus on the synthesis of plausible time-varying appearances by learning an effective 4D motion representation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 409, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 409, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 409, + 545, + 468 + ], + "type": "text", + "content": "namic movements of the T-shirt induced by dancing in Fig. 1. The secondary motion of clothes arises from intricate physical interactions with the body, typically changing over time, which leads to challenges for plausible rendering of dynamic humans." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 472, + 546, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 546, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 546, + 639 + ], + "type": "text", + "content": "Most existing methods [25, 43, 61] are conditioned on static poses and use a pose-guided generator to synthesize time-varying appearances. However, the appearance of clothed humans undergoes complex geometric transformations induced not only by the static pose but also its dynamics, whereas the modeling of dynamics is often ignored. In addition, these methods are often focused on the 2D appearance reconstruction of every single frame while the temporal motion relations are not fully explored. Due to these issues, they fail to render plausible secondary motions of clothes, i.e., generating the same appearances for fast and slow motions. A key challenge of learning temporal dynamics lies in the requirement of a tremendous amount of training observations to construct a 4D motion volume." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "To solve these issues, we propose a new paradigm to learn time-varying appearances of the secondary motion from just several sparse viewpoint video sequences, which is achieved by jointly modeling the temporal motion dynamics and human appearances in a unified rendering framework based on an efficient surface-based motion" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6550" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "representation. At the core of the paradigm is a feature encoder-decoder framework with three key components: 1) surface-based motion encoding; 2) physical motion decoding; and 3) 4D appearance decoding." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 289, + 396 + ], + "type": "text", + "content": "Firstly, in contrast to existing pose-guided methods [25, 42, 43, 61] that focus on static poses as a conditional variable, we extract an expressive 4D motion input from the 3D body mesh sequences obtained from training video as our input, which includes both a static pose represented by a spatial 3D mesh and its temporal dynamics. Furthermore, we notice that the non-rigid deformations of garments typically occur around the body surface instead of in a 3D volume, and hence we propose to model human motions on the body surface by projecting the extracted spatial-temporal 4D motion input to the dense 2D surface UV manifold of a clothless body template (e.g., SMPL [28]). To model temporal clothing offsets, a motion encoder is employed to lift the clothless motion features into a motion triplane that encodes both spatial and temporal motion relations in a compact 3D triplane with time-varying dynamics conditioned. The triplane is defined in the surface " + }, + { + "bbox": [ + 46, + 122, + 289, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{u} - \\mathbf{v} - \\mathbf{h}" + }, + { + "bbox": [ + 46, + 122, + 289, + 396 + ], + "type": "text", + "content": " coordinate system, with " + }, + { + "bbox": [ + 46, + 122, + 289, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{u} - \\mathbf{v}" + }, + { + "bbox": [ + 46, + 122, + 289, + 396 + ], + "type": "text", + "content": " to represent the motion of the clothless body template, and an extensional coordinate " + }, + { + "bbox": [ + 46, + 122, + 289, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 46, + 122, + 289, + 396 + ], + "type": "text", + "content": " to represent the secondary motion of clothes, i.e., the temporal clothing offsets are parameterized by a signed distance to the body surface. In this way, 4D motions can be effectively represented by a surface-based triplane." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 398, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 506 + ], + "type": "text", + "content": "Secondly, we propose to physically model the spatial and temporal motion dynamics in the rendering network. Specifically, with the surface-based triplane conditioned at time " + }, + { + "bbox": [ + 46, + 398, + 287, + 506 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 398, + 287, + 506 + ], + "type": "text", + "content": ", a motion decoder is employed to decode the motion triplanes to predict the motion at the next timestep " + }, + { + "bbox": [ + 46, + 398, + 287, + 506 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 46, + 398, + 287, + 506 + ], + "type": "text", + "content": ", i.e., spatial derivative of the motion physically corresponding to surface normal map and temporal derivatives corresponding to surface velocity. We illustrate the physical motion learning significantly improves the rendering quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 507, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 287, + 590 + ], + "type": "text", + "content": "Thirdly, the motion triplanes are decoded into high-quality images at two stages: a volumetric surface-conditioned renderer that is focused on the rendering around human body surface and filters the query points far from the body surface for efficient volumetric rendering, and a geometry-aware super-resolution module for efficient high-quality image synthesis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 592, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 700 + ], + "type": "text", + "content": "We conduct a systematical analysis of how human appearances are affected by temporal dynamics, and it was observed that some baseline methods mainly generate pose-dependent appearances instead of time-varying appearances for free-view video generations. In addition, quantitative and qualitative experiments are performed on three datasets with a total of 9 subject sequences, including ZJU-MoCap [43], AIST++ [24], and MPII-RDDC [13], which validate the effectiveness of SurMo in different scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 200, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 200, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 200, + 713 + ], + "type": "text", + "content": "In summary, our contributions are:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "1) A new paradigm for learning dynamic humans from videos that jointly models temporal motions and human appearances in a unified framework, and one of the early works that systematically analyzes how human appearances are affected by temporal dynamics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 132, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 132, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 132, + 545, + 167 + ], + "type": "text", + "content": "2) An efficient surface-based triplane that encodes both spatial and temporal motion relations for expressive 4D motion modeling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 168, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 545, + 228 + ], + "type": "text", + "content": "3) We achieve state-of-the-art results and show that our new paradigm is capable of learning high-fidelity appearances from fast motion sequences (e.g., AIST++ dance videos) or synthesizing motion-dependent shadows in challenging scenarios." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 239, + 392, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 392, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 392, + 251 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 258, + 545, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 282 + ], + "type": "text", + "content": "Our method is closely related to many sub-fields of visual computing, and below we discuss a set of the work." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 283, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 366 + ], + "type": "text", + "content": "3D Shape Representations. To capture detailed shapes of 3D objects, most recent papers utilize implicit functions [7, 8, 20, 33-35, 40, 41, 48-50, 56, 57, 65, 66] or point clouds [16, 31, 32] due to their topological flexibility. These methods aim to learn geometry from 3D datasets, whereas we synthesize human images of novel poses only from 2D RGB training images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 367, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 367, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 545, + 498 + ], + "type": "text", + "content": "Rendering Humans by 2D GANs. Some existing approaches render human avatars by neural image translation, i.e., they utilize GAN [10] networks to learn a mapping from poses (given in the form of renderings of a skeleton [4, 23, 45, 52, 59, 68], dense mesh [11, 26, 27, 36, 51, 59] or joint position heatmaps [1, 29, 30]) to human images [4, 59]. To improve temporal stability, some methods [17, 44, 46, 55] propose to utilize the SMPL [28] priors for pose-guided generations. However, these methods do not reconstruct geometry explicitly and cannot handle self-occlusions effectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "type": "text", + "content": "Rendering Humans by 3D-aware Renderer. For stable view synthesis, recent papers [6, 25, 38, 42, 43, 53] propose to unify geometry reconstruction with view synthesis by volume rendering, which, however, is computationally heavy. To solve this issue, some recent 3D-GAN papers [5, 12, 15, 18, 19, 37, 39, 67] propose a hybrid rendering strategy for efficient geometry-aware rendering, that is, render low-resolution volumetric features for geometry learning, and employ a super-resolution module for high-resolution image synthesis. We adopt this strategy for efficient rendering, whereas ours is distinguished by rendering articulated humans with 4D motion modeling." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "UV-based Pose Representation. Some methods [18, 19, 31, 32, 47, 63] propose to project 3D posed meshes into a 2D positional map for pose encoding, which can be used for different downstream tasks, such as 3D reconstruction and novel view synthesis. [31, 32] rely on 3D supervision to learn 3D reconstructions, and they do not take as input the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6551" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "dynamics. To improve the rendering quality, [18, 47] proposes to utilize additional driving views as input for faithful rendering in telepresence applications. However, they do not explore how to learn temporal dynamics from pose sequences. By taking as input a 3D normal and velocity map, [63] encodes motion dynamics in their rendering network, whereas they do explicitly learn motions, i.e., predicting the motion status at the next timestep, and besides, [63] is a 2D rendering method, and they do not explicitly learn motions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 192, + 129, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 192, + 129, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 129, + 205 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 213, + 139, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 139, + 225 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 139, + 225 + ], + "type": "text", + "content": "3.1. Problem Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 232, + 287, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 287, + 304 + ], + "type": "text", + "content": "Given a sparse multi-view video of a clothed human in motion and corresponding 3D pose estimations " + }, + { + "bbox": [ + 46, + 232, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_0, \\dots, \\mathbf{P}_t\\}" + }, + { + "bbox": [ + 46, + 232, + 287, + 304 + ], + "type": "text", + "content": ", our goal is to synthesize time-varying appearances of the individual under novel views. Existing methods [25, 42, 43, 61] formulate the problem of human rendering as learning a representation via a feature encoder-decoder framework:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 315, + 287, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 287, + 334 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\mathbf {t}} = \\mathcal {E} _ {\\mathcal {P}} \\left(\\mathbf {P} _ {\\mathbf {t}}, \\mathbf {z} _ {\\mathbf {t}}\\right) \\tag {1}", + "image_path": "6000b3698ce3279eb41447201826e1ba68f0401169b3b890a54afec26ac45b59.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 331, + 187, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 331, + 187, + 344 + ], + "spans": [ + { + "bbox": [ + 133, + 331, + 187, + 344 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {\\mathbf {t}} = \\mathcal {D} _ {\\mathcal {R}} (\\mathbf {f} _ {\\mathbf {t}})", + "image_path": "057d9f3921a83d163bd9c1bcfa7a97b88fbfa365138f06e4225dffdc41b783de.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "text", + "content": "where a pose encoder " + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{P}}" + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "text", + "content": " takes as input a representation of posed body " + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{P_t}" + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "text", + "content": " (e.g., 2D or 3D keypoints or 3D body mesh vertices) and a timestamp embedding " + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{z_t}" + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "text", + "content": ", and outputs intermediate pose-dependent features " + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{f_t}" + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "text", + "content": " that can be rendered by a decoder " + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{R}}" + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "text", + "content": " to reconstruct the appearance images " + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{I_t}" + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "text", + "content": " of the corresponding pose at time " + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "text", + "content": ". The time embedding " + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "inline_equation", + "content": "\\mathbf{z_t}" + }, + { + "bbox": [ + 46, + 350, + 287, + 446 + ], + "type": "text", + "content": " often serves as a residual that is updated passively in backpropagation by image reconstruction loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 447, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 613 + ], + "type": "text", + "content": "However, there are two issues with the above well-adopted paradigm. First, the appearance of clothed humans undergoes complex geometric transformations induced not only by the static pose " + }, + { + "bbox": [ + 46, + 447, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathbf{t}}" + }, + { + "bbox": [ + 46, + 447, + 287, + 613 + ], + "type": "text", + "content": " but also its dynamics, whereas the modeling of dynamics is ignored in E.q. 1, and the residual " + }, + { + "bbox": [ + 46, + 447, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathbf{t}}" + }, + { + "bbox": [ + 46, + 447, + 287, + 613 + ], + "type": "text", + "content": " cannot expressively model physical dynamics neither. Second, existing methods focus on the per-image reconstruction while ignoring the temporal relations of a motion sequence in training, i.e., sampling input pose " + }, + { + "bbox": [ + 46, + 447, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathbf{t}}" + }, + { + "bbox": [ + 46, + 447, + 287, + 613 + ], + "type": "text", + "content": " and timestamp " + }, + { + "bbox": [ + 46, + 447, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathbf{t}}" + }, + { + "bbox": [ + 46, + 447, + 287, + 613 + ], + "type": "text", + "content": " individually, and supervising image reconstructions frame-by-frame, partly because defining temporal supervision in the 2D image space is challenging, especially for articulated humans that often suffer from misalignment problems due to pose estimation errors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 615, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 615, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 287, + 651 + ], + "type": "text", + "content": "To solve these issues, we propose a new paradigm for learning view synthesis of dynamic humans from video sequences, which is formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 662, + 233, + 674 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 662, + 233, + 674 + ], + "spans": [ + { + "bbox": [ + 159, + 662, + 233, + 674 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\mathbf {t}} = \\mathcal {E} _ {\\mathcal {M}} (\\mathbf {P} _ {\\mathbf {t}}, \\mathbf {D} _ {\\mathbf {t}})", + "image_path": "f31fd04682b3f2ace7da8fdf29ab54c2e35cbc9755c2191a1c0c535caf13dda7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 101, + 677, + 287, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 677, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 101, + 677, + 287, + 700 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathbf {P} _ {\\mathbf {t} + 1}}{\\partial \\mathbf {x}}, \\frac {\\partial \\mathbf {P} _ {\\mathbf {t} + 1}}{\\partial \\mathbf {t}} = \\mathcal {D} _ {\\mathcal {M}} (\\mathbf {f} _ {\\mathbf {t}}) \\tag {2}", + "image_path": "cdc47b38c8d6ff6c908d322f3559629d26df84b9930cfdc3796bf715c8b4e586.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 157, + 703, + 212, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 703, + 212, + 715 + ], + "spans": [ + { + "bbox": [ + 157, + 703, + 212, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {\\mathbf {t}} = \\mathcal {D} _ {\\mathcal {R}} (\\mathbf {f} _ {\\mathbf {t}})", + "image_path": "1a3088c5a2ce10c88d75dfe85865565594189f53136a3565a996e827f80e2c55.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "text", + "content": "where the input is represented by dynamic motions including a static pose " + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathbf{t}}" + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "text", + "content": " and its physical dynamics " + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathbf{t}}" + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "text", + "content": ", and a motion encoder " + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{M}}" + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "text", + "content": " is employed to encode the inputs into motion-dependent features (Sec. 3.2). Besides, in the training stage, a physical motion decoder " + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{M}}" + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "text", + "content": " is introduced to enforce the learning of spatial and temporal relations by decoding the intermediate feature " + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{t}}" + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "text", + "content": " to predict the spatial derivatives at " + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "text", + "content": " physically corresponding to surface normal " + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_{\\mathbf{t} + 1} = \\frac{\\partial\\mathbf{P}_{\\mathbf{t} + 1}}{\\partial\\mathbf{x}}" + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "text", + "content": ", and temporal derivatives at " + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "text", + "content": " corresponding to surface velocity " + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathbf{t} + 1} = \\frac{\\partial\\mathbf{P}_{\\mathbf{t} + 1}}{\\partial\\mathbf{t}}" + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "text", + "content": " (Sec. 3.3). " + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{t}}" + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "text", + "content": " is rendered into human images by a decoder " + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{R}}: \\mathcal{G}_2 \\circ \\mathcal{G}_1" + }, + { + "bbox": [ + 306, + 72, + 545, + 217 + ], + "type": "text", + "content": " (Sec. 3.5). The framework is depicted in Fig. 2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 218, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 218, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 218, + 545, + 277 + ], + "type": "text", + "content": "Notation. In the following parts, we use " + }, + { + "bbox": [ + 306, + 218, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{space}}" + }, + { + "bbox": [ + 306, + 218, + 545, + 277 + ], + "type": "text", + "content": " to denote the features in the pipeline, i.e., " + }, + { + "bbox": [ + 306, + 218, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{t}}^{3\\mathbf{D}}" + }, + { + "bbox": [ + 306, + 218, + 545, + 277 + ], + "type": "text", + "content": " is a 4D representation that is defined in 3D space with temporal " + }, + { + "bbox": [ + 306, + 218, + 545, + 277 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 218, + 545, + 277 + ], + "type": "text", + "content": " conditioned, and " + }, + { + "bbox": [ + 306, + 218, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathbf{t}}^{3\\mathbf{D}}" + }, + { + "bbox": [ + 306, + 218, + 545, + 277 + ], + "type": "text", + "content": " is 4D motion input defined in 3D space with " + }, + { + "bbox": [ + 306, + 218, + 545, + 277 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 218, + 545, + 277 + ], + "type": "text", + "content": " conditioned." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 289, + 495, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 289, + 495, + 301 + ], + "spans": [ + { + "bbox": [ + 306, + 289, + 495, + 301 + ], + "type": "text", + "content": "3.2. Surface-based 4D Motion Encoding" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "text", + "content": "Extracting 4D Motions. We first extract 4D motions from a sequence of time-varying parametric posed body (e.g., SMPL) meshes obtained from training video sequences. We describe the motion " + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathbf{t}}^{3\\mathbf{D}}" + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "text", + "content": " as a static skeleton pose " + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathbf{t}}" + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "text", + "content": " and its dynamics " + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathbf{t}}" + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "text", + "content": ". The dynamics at " + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "text", + "content": " are physically determined by the current pose, 3D velocity, and motion trajectory of the past several timesteps, which also contribute to the time-varying appearances of the secondary motion. The pose " + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathbf{t}}" + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "text", + "content": " is represented by the 3D vertices of the posed mesh, and dynamics " + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathbf{t}}" + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "text", + "content": " are parameterized by 1) body surface velocity " + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathbf{t}}" + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "text", + "content": " corresponding to the temporal derivatives of the current pose at " + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "text", + "content": ", and 2) motion trajectory " + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathbf{t}}" + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "text", + "content": " that aggregates the temporal derivatives over the past several timesteps with a sliding window size of " + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "text", + "content": " with weight " + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 308, + 545, + 476 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 486, + 545, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 486, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 315, + 486, + 545, + 544 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {M} _ {\\mathbf {t}} ^ {\\mathrm {3 D}} = \\left[ \\mathbf {P} _ {\\mathbf {t}}, \\mathbf {D} _ {\\mathbf {t}} \\right], \\quad \\mathbf {D} _ {\\mathbf {t}} = \\left[ \\mathbf {V} _ {\\mathbf {t}}, \\mathbf {T} _ {\\mathbf {t}} \\right] \\\\ \\mathbf {V} _ {\\mathbf {t}} = \\frac {\\partial \\mathbf {P} _ {\\mathbf {t}}}{\\partial \\mathbf {t}}, \\quad \\mathbf {T} _ {\\mathbf {t}} = \\mathbf {P} _ {\\mathbf {t}} + \\frac {1}{\\sum \\lambda_ {i}} \\sum_ {i = 1} ^ {w} \\mathbf {V} _ {\\mathbf {t} - \\mathbf {i}} * \\lambda_ {i} \\tag {3} \\\\ \\end{array}", + "image_path": "ef2243143624c788d48bcaafd6a35180577a6abe5160b60c4260877e09752d42.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 545, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 545, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 305, + 545, + 545, + 616 + ], + "type": "text", + "content": "Note the motion trajectory aggregated from several consecutive timesteps makes the motion representation robust to pose estimation errors, i.e., the pose estimations for two consecutive timesteps may be the same due to pose estimation errors in practice. An ablation study of the motion trajectory representation can be found in the supp. mat." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "text", + "content": "Recording 4D Motions on the Surface Manifold. Modeling the motions with temporal dynamics often requires dense observations to construct a 4D motion volume. Instead, we notice that non-rigid deformations of human geometry in motion typically occur around the body surface instead of a 3D volume, and hence we propose to model the motions on the human body surface. Specifically, we project the 4D motion input " + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathbf{t}}^{3\\mathrm{D}}" + }, + { + "bbox": [ + 305, + 617, + 545, + 714 + ], + "type": "text", + "content": " including spatial pose" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6552" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 547, + 203 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 547, + 203 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 547, + 203 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 547, + 203 + ], + "type": "image", + "image_path": "5a621c6cf918e0f0a8ca49a4ed43b02af7ef62e02595360965c875188ed7d426.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "lines": [ + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "text", + "content": "Figure 2. Framework overview. Given a set of time-varying 3D body meshes " + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_{\\mathbf{t}},\\dots,\\mathbf{P}_{\\mathbf{t}} - \\mathbf{n}\\}" + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "text", + "content": " obtained from training video sequences, we aim to synthesize high-fidelity appearances of a clothed human in motion via a feature encoder-decoder framework: Motion Encoding, and joint Motion and Appearance Decoding. 1) We take as input an expressive 4D motion representation extracted from the mesh sequences including 3D pose, 3D velocity at time " + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "text", + "content": ", and motion trajectory over the past " + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "text", + "content": " timesteps that encode both spatial and temporal relations of the motion sequence, which are projected to the spatially aligned UV surface space. A motion encoder " + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{M}}" + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "text", + "content": " is employed to lift the 2D UV-aligned features to a 3D surface-based triplane " + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}" + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "text", + "content": " in an UV-plus-height space with a signed distance height to model temporal clothing offsets. 2) A motion decoder " + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{M}}" + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "text", + "content": " is designed to encourage physical motion learning in training by decoding the triplane features " + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}" + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "text", + "content": " to predict the motion at the next timestep " + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "text", + "content": ", i.e. spatial derivatives surface normal " + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_{\\mathbf{t} + 1}^{\\mathrm{uv}}" + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "text", + "content": " and temporal derivatives surface velocity " + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathbf{t} + 1}^{\\mathrm{uv}}" + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "text", + "content": " in UV space. 3) Finally, given a target camera view, the triplane " + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}" + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "text", + "content": " is rendered into high-quality images by a volumetric surface-conditioned renderer including volumetric low-resolution rendering by " + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_1" + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "text", + "content": " and an efficient geometry-aware super-resolution by " + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_2" + }, + { + "bbox": [ + 46, + 210, + 547, + 321 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 333, + 287, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 287, + 405 + ], + "type": "text", + "content": "and temporal dynamics from 3D space into a compact spatially aligned UV space using the geometric transformation " + }, + { + "bbox": [ + 46, + 333, + 287, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 333, + 287, + 405 + ], + "type": "text", + "content": " that is pre-defined by the parametric (e.g., SMPL) body template, which yields " + }, + { + "bbox": [ + 46, + 333, + 287, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uv}} = \\mathcal{W}\\mathbf{M}_{\\mathbf{t}}^{3\\mathrm{D}}" + }, + { + "bbox": [ + 46, + 333, + 287, + 405 + ], + "type": "text", + "content": ". The UV-aligned motion feature " + }, + { + "bbox": [ + 46, + 333, + 287, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uv}}" + }, + { + "bbox": [ + 46, + 333, + 287, + 405 + ], + "type": "text", + "content": " faithfully preserves the articulated structures of body topology in a compact 2D space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 405, + 288, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 288, + 513 + ], + "type": "text", + "content": "Generating Surface-based Triplanes for Motion Modeling. To model clothed humans, we further employ a motion encoder " + }, + { + "bbox": [ + 46, + 405, + 288, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{M}}" + }, + { + "bbox": [ + 46, + 405, + 288, + 513 + ], + "type": "text", + "content": " that lifts the clothless 2D features " + }, + { + "bbox": [ + 46, + 405, + 288, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uv}}" + }, + { + "bbox": [ + 46, + 405, + 288, + 513 + ], + "type": "text", + "content": " to a 3D triplane representation " + }, + { + "bbox": [ + 46, + 405, + 288, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}" + }, + { + "bbox": [ + 46, + 405, + 288, + 513 + ], + "type": "text", + "content": " that is defined in a " + }, + { + "bbox": [ + 46, + 405, + 288, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{u} - \\mathbf{v} - \\mathbf{h}" + }, + { + "bbox": [ + 46, + 405, + 288, + 513 + ], + "type": "text", + "content": " system, with " + }, + { + "bbox": [ + 46, + 405, + 288, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{u} - \\mathbf{v}" + }, + { + "bbox": [ + 46, + 405, + 288, + 513 + ], + "type": "text", + "content": " to represent the motion of the clothless body template, and an extensional coordinate " + }, + { + "bbox": [ + 46, + 405, + 288, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 46, + 405, + 288, + 513 + ], + "type": "text", + "content": " to represent the secondary motion of clothes, i.e., the temporal cloth offsets, and hence 4D clothed motions can be parameterized by a surface-based triplane by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 521, + 287, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 521, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 93, + 521, + 287, + 536 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathrm {u v h}} = \\mathcal {E} _ {\\mathcal {M}} \\left(\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathrm {u v}}\\right) = \\mathcal {E} _ {\\mathcal {M}} \\left(\\mathcal {W} \\mathbf {M} _ {\\mathbf {t}} ^ {\\mathbf {3 D}}\\right), \\tag {4}", + "image_path": "2a162ac77f17945866f9c8e9fc17d76865e072af97d3468e735e9e04cda981b7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 544, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 544, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 544, + 287, + 653 + ], + "type": "text", + "content": " is the geometric transformation from 3D space to UV space. " + }, + { + "bbox": [ + 46, + 544, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}" + }, + { + "bbox": [ + 46, + 544, + 287, + 653 + ], + "type": "text", + "content": " consists of three planes " + }, + { + "bbox": [ + 46, + 544, + 287, + 653 + ], + "type": "inline_equation", + "content": "x_{uv}, x_{uh}, x_{hv} \\in \\mathbb{R}^{U \\times V \\times C}" + }, + { + "bbox": [ + 46, + 544, + 287, + 653 + ], + "type": "text", + "content": " to form the spatial relationship, where " + }, + { + "bbox": [ + 46, + 544, + 287, + 653 + ], + "type": "inline_equation", + "content": "U, V" + }, + { + "bbox": [ + 46, + 544, + 287, + 653 + ], + "type": "text", + "content": " denote spatial resolution and " + }, + { + "bbox": [ + 46, + 544, + 287, + 653 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 544, + 287, + 653 + ], + "type": "text", + "content": " is the channel number. In contrast to the volumetric triplanes used in EG3D [5] where the three planes represent three vertical planes in the 3D volumetric space, " + }, + { + "bbox": [ + 46, + 544, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}" + }, + { + "bbox": [ + 46, + 544, + 287, + 653 + ], + "type": "text", + "content": " is defined on the human body surface inheriting human topology priors for motion modeling, as illustrated by an unwarped triplane in 3D space in Fig 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 658, + 194, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 194, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 194, + 672 + ], + "type": "text", + "content": "3.3. Physical Motion Decoding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "We propose a physical motion decoder " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{M}}" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " to learn spatial and temporal features of motion, which is achieved by decoding the intermediate motion feature " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " learned by " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{M}}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 333, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 545, + 392 + ], + "type": "text", + "content": "to predict the motions at the next timestep " + }, + { + "bbox": [ + 304, + 333, + 545, + 392 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 304, + 333, + 545, + 392 + ], + "type": "text", + "content": ", such as spatial derivatives corresponding to surface normal " + }, + { + "bbox": [ + 304, + 333, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_{t+1}^{\\mathrm{uv}}" + }, + { + "bbox": [ + 304, + 333, + 545, + 392 + ], + "type": "text", + "content": ", temporal derivatives corresponding to surface velocity " + }, + { + "bbox": [ + 304, + 333, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{t+1}^{\\mathrm{uv}}" + }, + { + "bbox": [ + 304, + 333, + 545, + 392 + ], + "type": "text", + "content": ". Note that we model the normal and velocity in the UV space by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 397, + 545, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 397, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 311, + 397, + 545, + 413 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathbf {N} _ {\\mathbf {t} + \\mathbf {1}} ^ {\\mathrm {u v}}, \\mathbf {V} _ {\\mathbf {t} + \\mathbf {1}} ^ {\\mathrm {u v}} \\right\\} = \\mathcal {D} _ {\\mathcal {M}} \\left(\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathrm {u v h}}\\right) = \\mathcal {D} _ {\\mathcal {M}} \\circ \\mathcal {E} _ {\\mathcal {M}} \\left(\\mathcal {W} \\mathbf {M} _ {\\mathbf {t}} ^ {\\mathbf {3 D}}\\right) (5)", + "image_path": "cf2f9e91ea2ca56ff208812e623e4b55e3e1fd883b0d8d9c29d8e0b41b6b9452.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 423, + 448, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 423, + 448, + 436 + ], + "spans": [ + { + "bbox": [ + 306, + 423, + 448, + 436 + ], + "type": "text", + "content": "3.4. 4D Appearance Decoding" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": "Volumetric Surface-conditioned Rendering. Given a target camera viewpoint, the triplane " + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{t}}^{\\mathrm{uvh}}" + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": " is first rendered into low-resolution volumetric features " + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathbf{F}} \\in \\mathbb{R}^{H \\times W \\times \\hat{C}}" + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": " by a conditional NeRF " + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\mathbb{F}_{\\Phi}" + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": " are the spatial resolution and " + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": " is the channel number. To be more specific, given a 3D query point " + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i" + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": ", we transform it into a surface-based local coordinate " + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}_i = (\\pmb{u}_i, \\pmb{v}_i, \\pmb{h}_i)" + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": " w.r.t. the tracked body mesh " + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathbf{t}}" + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": ". Here we search the nearest face " + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i" + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathbf{t}}" + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": " for query point " + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i" + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "inline_equation", + "content": "(\\pmb{u}_i, \\pmb{v}_i)" + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\pmb{h}_i" + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": " are the barycentric coordinates of the nearest point on " + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i" + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": ", and the signed distance respectively. We therefore obtain the local feature " + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\pmb{z}_i^{\\mathrm{uvh}}" + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": " of the query point " + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i" + }, + { + "bbox": [ + 304, + 441, + 545, + 588 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 592, + 552, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 592, + 552, + 617 + ], + "spans": [ + { + "bbox": [ + 306, + 592, + 552, + 617 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {i} ^ {u v h} = \\operatorname {c a t} \\left[ \\Pi \\left(\\boldsymbol {x} _ {u v}; \\boldsymbol {u} _ {i}, \\boldsymbol {v} _ {i}\\right), \\Pi \\left(\\boldsymbol {x} _ {u h}; \\boldsymbol {u} _ {i}, \\boldsymbol {h} _ {i}\\right), \\Pi \\left(\\boldsymbol {x} _ {h v}; \\boldsymbol {h} _ {i}, \\boldsymbol {v} _ {i}\\right) \\right], \\tag {6}", + "image_path": "1e9b2323396d3404a3334c345a36733ab3ed51c5f31c980bb7817b0e53ebfc73.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 617, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\Pi (\\cdot)" + }, + { + "bbox": [ + 304, + 617, + 545, + 640 + ], + "type": "text", + "content": " denotes sampling operation, " + }, + { + "bbox": [ + 304, + 617, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{cat}[\\cdot ]" + }, + { + "bbox": [ + 304, + 617, + 545, + 640 + ], + "type": "text", + "content": " denotes the concatenation operator." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 641, + 544, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 641, + 544, + 666 + ], + "spans": [ + { + "bbox": [ + 306, + 641, + 544, + 666 + ], + "type": "text", + "content": "Given camera direction " + }, + { + "bbox": [ + 306, + 641, + 544, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{d}_i" + }, + { + "bbox": [ + 306, + 641, + 544, + 666 + ], + "type": "text", + "content": ", the appearance features " + }, + { + "bbox": [ + 306, + 641, + 544, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{c}_i" + }, + { + "bbox": [ + 306, + 641, + 544, + 666 + ], + "type": "text", + "content": " and density features " + }, + { + "bbox": [ + 306, + 641, + 544, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{\\sigma}_i" + }, + { + "bbox": [ + 306, + 641, + 544, + 666 + ], + "type": "text", + "content": " of point " + }, + { + "bbox": [ + 306, + 641, + 544, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i" + }, + { + "bbox": [ + 306, + 641, + 544, + 666 + ], + "type": "text", + "content": " are predicted by" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 374, + 670, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 670, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 374, + 670, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\left\\{\\boldsymbol {c} _ {i}, \\boldsymbol {\\sigma} _ {i} \\right\\} = \\mathbb {F} _ {\\Phi} \\left(\\boldsymbol {d} _ {i}, \\boldsymbol {z} _ {i} ^ {u v h}\\right) \\tag {7}", + "image_path": "8f3e87f1e20d66641b0c14c67027e0a9cd53e5e25e04ae67ba610894a810bb65.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "We integrate all the radiance features of sampled points into a 2D feature map " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{F}" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " through volume renderer" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6553" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 441 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 441 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 441 + ], + "type": "image", + "image_path": "88c4ec5eef522c9e383384ba91eea32c81698ccdf890a3fe5474f390535bf409.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 449, + 547, + 504 + ], + "lines": [ + { + "bbox": [ + 46, + 449, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 547, + 504 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons on novel view synthesis on the subject S313 of ZJU-MoCap dataset. Two motion sequences S1 (swing arms left to right) and S2 (raise and lower arms) are shown. We specifically focus on the synthesis of time-varying appearances (especially T-shirt wrinkles), by evaluating the rendering results under similar poses yet with different movement directions, which are marked in the same color, such as the pairs of " + }, + { + "bbox": [ + 46, + 449, + 547, + 504 + ], + "type": "inline_equation", + "content": "①②" + }, + { + "bbox": [ + 46, + 449, + 547, + 504 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 449, + 547, + 504 + ], + "type": "inline_equation", + "content": "③④" + }, + { + "bbox": [ + 46, + 449, + 547, + 504 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 449, + 547, + 504 + ], + "type": "inline_equation", + "content": "⑤⑥" + }, + { + "bbox": [ + 46, + 449, + 547, + 504 + ], + "type": "text", + "content": ". Our method synthesizes high-fidelity time-varying appearances, whereas SOTA HumanNeRF generates almost the same cloth wrinkles." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 525, + 80, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 80, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 80, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_1" + }, + { + "bbox": [ + 47, + 525, + 80, + 537 + ], + "type": "text", + "content": " [21]" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 536, + 287, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 536, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 115, + 536, + 287, + 550 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {\\mathbf {F}} = \\mathcal {G} _ {1} \\left(\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {u v h}}, c a m; \\mathbb {F} _ {\\Phi}\\right) \\tag {8}", + "image_path": "7667420cfab303731a051c5f5fd027fa9a48b10f45b325b9cc0ffede18c53456.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 555, + 194, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 194, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 194, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 555, + 194, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{t}}^{\\mathrm{uvh}}" + }, + { + "bbox": [ + 47, + 555, + 194, + 567 + ], + "type": "text", + "content": " is extracted from Eq. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 567, + 287, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 630 + ], + "type": "text", + "content": "Efficient Geometry-Aware Super-Resolution. Sampling dense points to render the full-resolution volumetric features is computationally heavy. Instead, we employ a superresolution network " + }, + { + "bbox": [ + 46, + 567, + 287, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_2" + }, + { + "bbox": [ + 46, + 567, + 287, + 630 + ], + "type": "text", + "content": " [5] to render high-resolution images " + }, + { + "bbox": [ + 46, + 567, + 287, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathbf{RGB}}^{+} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times 3}" + }, + { + "bbox": [ + 46, + 567, + 287, + 630 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 639, + 287, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 639, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 97, + 639, + 287, + 654 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {\\mathbf {R G B}} ^ {+} = \\mathcal {G} _ {2} \\circ \\mathcal {G} _ {1} \\left(\\mathbf {f} _ {\\mathbf {t}} ^ {\\mathbf {u v h}}, c a m; \\mathbb {F} _ {\\Phi}\\right) \\tag {9}", + "image_path": "37d3c718e7949767e8808ada6830cc9e27ece55ff38896f556d5ea4047ef4e5b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 659, + 132, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 132, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 132, + 672 + ], + "type": "text", + "content": "3.5. Optimization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "SurMo is trained end-to-end to optimize " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{M}}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{M}}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": ", and renderers " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_1" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_1" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " with 2D image loss. We employ Adversarial Loss and Reconstruction Loss including Pixel Loss," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 525, + 545, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 525, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 305, + 525, + 545, + 572 + ], + "type": "text", + "content": "Perceptual Loss, Face Identity Loss, Velocity and Normal Loss, and Volume Rendering Loss for supervision in training, with Adam [22] as the optimizer. Refer to the supp. mat. for more details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 585, + 388, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 585, + 388, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 585, + 388, + 599 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Dataset and Metrics. We evaluate the novel view synthesis on ZJU-MoCap, MPII-RDDC, and AIST++, with a total of 9 subjects. We use the same camera setup as Neural Body, that is, 4 cameras used for training, and the other 18 for testing. For MPII-RDDC, 18 cameras in training, and 9 for testing. Since ASIT++ only has 9 cameras, we use 6 for training, and the remaining 3 for testing. Refer to more details in the supp. mat. We compare each method on per-pixel metrics including SSIM [60] and PSNR, and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6554" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 298, + 249 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 298, + 249 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 298, + 249 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 298, + 249 + ], + "type": "image", + "image_path": "0d2e47398f1fb0efc17a0c34bcfbc14ef3610b57d9ac2bff4211c7c03e472755.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 250, + 272, + 259 + ], + "lines": [ + { + "bbox": [ + 49, + 250, + 272, + 259 + ], + "spans": [ + { + "bbox": [ + 49, + 250, + 272, + 259 + ], + "type": "text", + "content": "Neural Body Instant-NVR HumanNeRF Ours GT" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 299, + 72, + 545, + 249 + ], + "blocks": [ + { + "bbox": [ + 299, + 72, + 545, + 249 + ], + "lines": [ + { + "bbox": [ + 299, + 72, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 299, + 72, + 545, + 249 + ], + "type": "image", + "image_path": "4fb3ce63b4f819b49edb8d72bdfc6828627c391d5ad7c8f53e2ede247011c197.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 250, + 518, + 259 + ], + "lines": [ + { + "bbox": [ + 304, + 250, + 518, + 259 + ], + "spans": [ + { + "bbox": [ + 304, + 250, + 518, + 259 + ], + "type": "text", + "content": "Neural Body Instant-NVR HumanNeRF Ours GT" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 268, + 547, + 301 + ], + "lines": [ + { + "bbox": [ + 46, + 268, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 547, + 301 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons on novel view synthesis on the subject S387, S315 of ZJU-MoCap dataset. Row 1 and 2 show similar poses occurring at different timesteps (not consecutive frames). The results indicate that our method synthesizes time-varying appearances while other methods mainly generate pose-dependent appearances." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 407, + 287, + 483 + ], + "blocks": [ + { + "bbox": [ + 46, + 321, + 287, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 321, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 287, + 399 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on ZJU-MoCap dataset (averaged on all test views and poses on 6 sequences) for novel-view synthesis. To reduce the influence of the background, all scores are calculated from images cropped to 2D bounding boxes. Note that the perception metrics LPIPS [64] and FID [14] capture human judgment better than per-pixel metrics such as SSIM [60] or PSNR, as stated in [25, 51]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 407, + 287, + 483 + ], + "lines": [ + { + "bbox": [ + 48, + 407, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 407, + 287, + 483 + ], + "type": "table", + "html": "
ZJU-S1-6LPIPS ↓FID ↓SSIM ↑PSNR ↑
Neural Body [43].164125.10.70321.438
Instant-NVR [9].202144.99.73821.748
HumanNeRF [61].10688.382.79223.624
Ours.07567.725.83324.815
", + "image_path": "cb31086a7d8f39e001be6c08a9026e21a602c56e1511fa44554197d12abc5c78.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 495, + 269, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 269, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 269, + 507 + ], + "type": "text", + "content": "perception metrics including LPIPS [64] and FID [14]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 507, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 287, + 615 + ], + "type": "text", + "content": "Baselines. We compare our method against SOTA methods including Neural Body [43], HumanNeRF [61], Instant-NVR [9], ARAH [58], DVA [47] and HVTR++ [18]. Neural Body models human poses in 3D space with point cloud as pose representation, HumanNeRF and Instant-NVR model poses in a canonical space with inverse skinning, ARAH is a forward-skinning based method, DVA and HVTR++ encode poses in UV space, and take driving view signals as input for telepresence applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 623, + 218, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 218, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 218, + 635 + ], + "type": "text", + "content": "4.1. Comparisons to SOTA Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": "Quantitative comparisons We conduct the quantitative comparisons on the three datasets with a total of 9 subjects. The quantitative results on ZJU-MoCap are summarized in Tab. 1, which suggests that our new paradigm outperforms the SOTA by a big margin, and we achieve the best quantitative results on all four metrics. The detailed compar-" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 323, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 371 + ], + "type": "text", + "content": "isons on each subject of ZJU-MoCap, and quantitative results on MPII-RDDC and AIST++ are listed in the supp. mat. Refer to the comparisons with ARAH [58], DVA [47] and HVTR++ [18] in the supp. mat." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 374, + 546, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 546, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 546, + 614 + ], + "type": "text", + "content": "Time-varying Appearances with Dynamics Conditioning on ZJU-MoCap. We compare with baseline methods for novel view synthesis on two motion sequences of subject S313 from the ZJU-MoCap dataset, as shown in Fig. 3. We evaluate the capability of each method to synthesize time-varying appearances. Specifically, two sequences S1 (swing arms left to right) and S2 (raise and lower arms) are evaluated, where similar poses occur at different timesteps with different dynamics (e.g., movement directions, trajectory, or velocity) are marked in the same color, such as the pairs of " + }, + { + "bbox": [ + 304, + 374, + 546, + 614 + ], + "type": "inline_equation", + "content": "①②" + }, + { + "bbox": [ + 304, + 374, + 546, + 614 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 374, + 546, + 614 + ], + "type": "inline_equation", + "content": "③④" + }, + { + "bbox": [ + 304, + 374, + 546, + 614 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 374, + 546, + 614 + ], + "type": "inline_equation", + "content": "⑤⑥" + }, + { + "bbox": [ + 304, + 374, + 546, + 614 + ], + "type": "text", + "content": ". Fig. 3 suggests that Instant-NVR [9] fails to synthesize time-varying high-frequency wrinkles, and Neural Body [43] synthesizes time-varying yet blurry wrinkles. HumanNeRF [61] synthesizes detailed yet static T-shirt wrinkles, i.e., the wrinkles are almost the same for similar poses, such as " + }, + { + "bbox": [ + 304, + 374, + 546, + 614 + ], + "type": "inline_equation", + "content": "①②" + }, + { + "bbox": [ + 304, + 374, + 546, + 614 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 374, + 546, + 614 + ], + "type": "inline_equation", + "content": "③④" + }, + { + "bbox": [ + 304, + 374, + 546, + 614 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 374, + 546, + 614 + ], + "type": "inline_equation", + "content": "⑤⑥" + }, + { + "bbox": [ + 304, + 374, + 546, + 614 + ], + "type": "text", + "content": ". In contrast, our method renders both high-fidelity and time-varying wrinkles. The comparisons on other subjects S387 and S315 are shown in Fig. 4, which illustrate the effectiveness of our proposed paradigm in dynamics learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "Time-varying Appearances with Motion-dependent Shadows on MPII-RDDC. We compare with baseline methods for novel view synthesis on MPII-RDDC, as shown in Fig. 5. The sequence is captured in a studio with top-down lighting that casts shadows on the human body due to self-occlusions. We notice that the synthesis of shadow can be formulated as learning motion-dependent shadows in the motion representation without the need to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6555" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 547, + 227 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 547, + 227 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 547, + 227 + ], + "type": "image", + "image_path": "bcd6ca47faffd07096384f8eaac4a158e38e6a378315e2a759538a443c396b69.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 229, + 196, + 238 + ], + "lines": [ + { + "bbox": [ + 146, + 229, + 196, + 238 + ], + "spans": [ + { + "bbox": [ + 146, + 229, + 196, + 238 + ], + "type": "text", + "content": "HumanNeRF" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 206, + 229, + 243, + 238 + ], + "lines": [ + { + "bbox": [ + 206, + 229, + 243, + 238 + ], + "spans": [ + { + "bbox": [ + 206, + 229, + 243, + 238 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 283, + 229, + 335, + 238 + ], + "lines": [ + { + "bbox": [ + 283, + 229, + 335, + 238 + ], + "spans": [ + { + "bbox": [ + 283, + 229, + 335, + 238 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 354, + 229, + 405, + 238 + ], + "lines": [ + { + "bbox": [ + 354, + 229, + 405, + 238 + ], + "spans": [ + { + "bbox": [ + 354, + 229, + 405, + 238 + ], + "type": "text", + "content": "HumanNeRF" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 414, + 229, + 447, + 238 + ], + "lines": [ + { + "bbox": [ + 414, + 229, + 447, + 238 + ], + "spans": [ + { + "bbox": [ + 414, + 229, + 447, + 238 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 483, + 229, + 535, + 238 + ], + "lines": [ + { + "bbox": [ + 483, + 229, + 535, + 238 + ], + "spans": [ + { + "bbox": [ + 483, + 229, + 535, + 238 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 322, + 287, + 384 + ], + "blocks": [ + { + "bbox": [ + 46, + 248, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 248, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 547, + 304 + ], + "type": "text", + "content": "Figure 5. Novel view synthesis of time-varying appearances with both pose and lighting conditioning on MPII-RDDC dataset. The sequence is captured in a studio with top-down lighting that casts shadows on the human performer due to self-occlusion. In Row 1, we specifically focus on synthesizing time-varying shadows (e.g., ① vs. ②, and ③ vs. ④) for different poses with different self-occlusions. In Row 2, we evaluate the synthesis of: 1) time-varying appearances for similar poses occurring in a jump-up-and-down motion sequence, e.g., ⑤ vs. ⑥, 2) shadows ⑦ vs. ⑧, and 3) clothing offsets ⑤ vs. ⑥." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 322, + 287, + 384 + ], + "lines": [ + { + "bbox": [ + 48, + 322, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 322, + 287, + 384 + ], + "type": "image", + "image_path": "2e3ed2af846bde22715ff3cada8a05944270991c71e47028fd8f6166e92449a9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 393, + 275, + 404 + ], + "lines": [ + { + "bbox": [ + 58, + 393, + 275, + 404 + ], + "spans": [ + { + "bbox": [ + 58, + 393, + 275, + 404 + ], + "type": "text", + "content": "Figure 6. Novel view synthesis of fast motions on AIST++." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 425, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 593 + ], + "type": "text", + "content": "explicitly model the lighting. We validate our method from two aspects of the scenario in Fig. 5. 1) In Row 1, Fig. 5 indicates that with the expressive motion representation and physical motion learning, SurMo succeeds in predicting the motion-dependent shadows, such as " + }, + { + "bbox": [ + 46, + 425, + 287, + 593 + ], + "type": "inline_equation", + "content": "①②③④" + }, + { + "bbox": [ + 46, + 425, + 287, + 593 + ], + "type": "text", + "content": ", whereas SOTA HumanNeRF renders almost the same T-shirt appearance. 2) In Row 2, we compare the synthesis of time-varying appearances for similar poses occurring in a jump-up-and-down motion sequence i.e., " + }, + { + "bbox": [ + 46, + 425, + 287, + 593 + ], + "type": "inline_equation", + "content": "⑤" + }, + { + "bbox": [ + 46, + 425, + 287, + 593 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 46, + 425, + 287, + 593 + ], + "type": "inline_equation", + "content": "⑥" + }, + { + "bbox": [ + 46, + 425, + 287, + 593 + ], + "type": "text", + "content": ", which shows that SurMo is capable of predicting the clothes offsets under similar pose with different motion trajectories. In addition, we synthesize the shadows " + }, + { + "bbox": [ + 46, + 425, + 287, + 593 + ], + "type": "inline_equation", + "content": "⑦⑧" + }, + { + "bbox": [ + 46, + 425, + 287, + 593 + ], + "type": "text", + "content": " in the challenging jump-up-and-down motion sequence. However, HumanNeRF fails to predict the dynamic clothing offsets " + }, + { + "bbox": [ + 46, + 425, + 287, + 593 + ], + "type": "inline_equation", + "content": "⑨" + }, + { + "bbox": [ + 46, + 425, + 287, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Time-varying Appearances for Fast Motions on AIST++. We also evaluate our methods by rendering humans with fast dance motions (S21 and S13) on AIST++, as shown in Fig. 6, where we compare ours against Neural Body in synthesizing time-varying wrinkles of the T-shirt for different motions. Fig. 6 suggests that Neural Body can only synthesize blurry results partly because Neural Body models pose in a sparse 3D space. Instead, we model motions in a denser body surface space, which enables high-fidelity image synthesis. Note that AIST++ is based" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 323, + 545, + 445 + ], + "blocks": [ + { + "bbox": [ + 307, + 323, + 545, + 445 + ], + "lines": [ + { + "bbox": [ + 307, + 323, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 307, + 323, + 545, + 445 + ], + "type": "image", + "image_path": "06d2a2cc665354948ad640d0d01262dfd556dcde6bbbfd86cc42038b4b418641.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 455, + 545, + 500 + ], + "lines": [ + { + "bbox": [ + 305, + 455, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 305, + 455, + 545, + 500 + ], + "type": "text", + "content": "Figure 8. Ablation study of motion conditioning and learning. We focus on the effect of motion conditioning and learning in Row 1, and whether our method learns to decouple the static pose and dynamics (e.g., velocity) from the motion conditioning." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 514, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 575 + ], + "type": "text", + "content": "on SMPL tracking with a scaling factor that affects the inverse LBS used in Instant-NVR or HumanNeRF both relying on inverse LBS, and hence we cannot compare them based on the released official code. Quantitative results can be found in the supp. mat." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 586, + 400, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 400, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 400, + 599 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "Surface-based Triplane vs. Volumetric Triplane. We compare the well-adopted volumetric triplane (Vol-Trip) [5] and our proposed surface-based triplane (Surf-Trip) for human modeling on two subjects S313 and S387 from ZJUMoCap, as shown in Fig. 7. Fig. 7 (left) suggests that the Surf-Trip converges faster in training with a smaller reconstruction loss for both sequences, and the performances are further improved with physical motion learning, e.g., 'Surf-Trip + ML'. Fig. 7 (right) compares the rendering re" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6556" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 229, + 220 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 229, + 220 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 229, + 220 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 229, + 220 + ], + "type": "image", + "image_path": "93fc99a370db7c1e669130b0279e6ed2c28e9220a804220058633a7d8c32d74a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 221, + 547, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 221, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 547, + 266 + ], + "type": "text", + "content": "Figure 7. Ablation study of 3D volumetric triplane (Vol-Trip) vs. surface-based triplane (Surf-Trip) for human modeling on subject S313 and S387 from ZJU-MoCap. We focus on the convergence in training (left), and self-occlusion rendering (right). The convergence of S313 is shifted for visualization purposes. Vol-Trip is not effective in handling self-occlusion, e.g., " + }, + { + "bbox": [ + 46, + 221, + 547, + 266 + ], + "type": "inline_equation", + "content": "①②③④" + }, + { + "bbox": [ + 46, + 221, + 547, + 266 + ], + "type": "text", + "content": " though it performs well in another viewpoint without self-occlusions " + }, + { + "bbox": [ + 46, + 221, + 547, + 266 + ], + "type": "inline_equation", + "content": "⑥" + }, + { + "bbox": [ + 46, + 221, + 547, + 266 + ], + "type": "text", + "content": ". In addition, Vol-Trip cannot synthesize high-quality details for face." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 238, + 71, + 543, + 220 + ], + "blocks": [ + { + "bbox": [ + 238, + 71, + 543, + 220 + ], + "lines": [ + { + "bbox": [ + 238, + 71, + 543, + 220 + ], + "spans": [ + { + "bbox": [ + 238, + 71, + 543, + 220 + ], + "type": "image", + "image_path": "2d8e22cbf2bb7793f4aa369875a8c53f83a80eda8f3684077af6967db514c0d7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 62, + 330, + 270, + 405 + ], + "blocks": [ + { + "bbox": [ + 47, + 285, + 287, + 329 + ], + "lines": [ + { + "bbox": [ + 47, + 285, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 287, + 329 + ], + "type": "text", + "content": "Table 2. Ablation study of motion conditioning and learning. " + }, + { + "bbox": [ + 47, + 285, + 287, + 329 + ], + "type": "inline_equation", + "content": "P_{cond}" + }, + { + "bbox": [ + 47, + 285, + 287, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 285, + 287, + 329 + ], + "type": "inline_equation", + "content": "D_{cond}" + }, + { + "bbox": [ + 47, + 285, + 287, + 329 + ], + "type": "text", + "content": " denote the conditioning of pose and dynamics, " + }, + { + "bbox": [ + 47, + 285, + 287, + 329 + ], + "type": "inline_equation", + "content": "V_{pred}" + }, + { + "bbox": [ + 47, + 285, + 287, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 285, + 287, + 329 + ], + "type": "inline_equation", + "content": "N_{pred}" + }, + { + "bbox": [ + 47, + 285, + 287, + 329 + ], + "type": "text", + "content": " denote the prediction of surface velocity and surface normal in motion learning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 330, + 270, + 405 + ], + "lines": [ + { + "bbox": [ + 62, + 330, + 270, + 405 + ], + "spans": [ + { + "bbox": [ + 62, + 330, + 270, + 405 + ], + "type": "table", + "html": "
S313LPIPS ↓FID ↓SSIM ↑PSNR ↑
Pcond.120104.83.79322.781
+ Dcond.08573.674.83424.908
+ Vpred.06962.092.85625.845
+ Npred.06050.170.86926.654
", + "image_path": "15736cc291c8f4350878bac458b91f024ede26c8902c7001bbef262f6d1fc0fa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 422, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 495 + ], + "type": "text", + "content": "sults qualitatively, which indicates that Surf-Trip is more effective in handling self-occlusions, whereas Vol-Trip fails " + }, + { + "bbox": [ + 46, + 422, + 287, + 495 + ], + "type": "inline_equation", + "content": "①②③④" + }, + { + "bbox": [ + 46, + 422, + 287, + 495 + ], + "type": "text", + "content": ", though it performs well in another viewpoint without self-occlusions " + }, + { + "bbox": [ + 46, + 422, + 287, + 495 + ], + "type": "inline_equation", + "content": "⑥" + }, + { + "bbox": [ + 46, + 422, + 287, + 495 + ], + "type": "text", + "content": ". In addition, Surf-Trip generates high-fidelity clothing wrinkles and facial details, whereas the face rendering of Vol-Trip is blurry." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 495, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 604 + ], + "type": "text", + "content": "Motion Conditioning and Learning. We also analyze the effectiveness of our proposed motion conditioning and learning qualitatively and quantitatively. Tab. 2 summarizes the quantitative results, which suggest that the quantitative results are significantly improved by taking as input the dynamics conditioning " + }, + { + "bbox": [ + 46, + 495, + 287, + 604 + ], + "type": "inline_equation", + "content": "D_{cond}" + }, + { + "bbox": [ + 46, + 495, + 287, + 604 + ], + "type": "text", + "content": ". With physical motion learning, i.e., predicting the temporal derivatives surface velocity and spatial derivates normal at the next timestep, the quantitative results are further improved by a big margin." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "The qualitative comparisons are shown in Fig. 8, where in Row 1, we observe higher-fidelity appearances with dynamics conditioning, and physical motion learning for both velocity and normal prediction. In Row 2, we evaluate whether our method learns to decouple the temporal dynamics (e.g., velocity) and static poses from motion conditioning, which corresponds to the rendering of clothing wrinkles of secondary motion (T-shirt ①) and tight parts (e.g., shorts ②). We only change the velocity for each variant" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 286, + 547, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 547, + 431 + ], + "type": "text", + "content": "in the ablation study. It suggests that the appearance of the T-shirt varies when the velocity or dynamics are changed, whereas the renderings of tight parts remain the same, such as the head, tight shorts, and shoes. This is consistent with our daily observations. The ablation study illustrates that our method decouples the pose and dynamics, and is capable of generating dynamics-dependent appearances. Note that the appearance changes are small between " + }, + { + "bbox": [ + 304, + 286, + 547, + 431 + ], + "type": "inline_equation", + "content": "V * 1" + }, + { + "bbox": [ + 304, + 286, + 547, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 286, + 547, + 431 + ], + "type": "inline_equation", + "content": "V * 0.01" + }, + { + "bbox": [ + 304, + 286, + 547, + 431 + ], + "type": "text", + "content": " when we reduce the velocity, partly because the method is not sensitive to smaller velocity due to the pose estimation errors in the training dataset where consecutive frames may be estimated with the same pose fitting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 452, + 375, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 452, + 375, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 452, + 375, + 464 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 475, + 547, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 547, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 547, + 596 + ], + "type": "text", + "content": "We propose SurMo, a new paradigm for learning dynamic humans from videos by jointly modeling temporal motions and human appearances in a unified framework, based on an efficient surface-based triplane. We conduct a systematical analysis of how human appearances are affected by temporal dynamics, and extensive experiments validate the expressiveness of the surface-based triplane in rendering fast motions and motion-dependent shadows. Quantitative experiments illustrate that SurMo achieves the SOTA results on different motion sequences." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 617, + 404, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 617, + 404, + 632 + ], + "spans": [ + { + "bbox": [ + 306, + 617, + 404, + 632 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6557" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Kfir Aberman, M. Shi, Jing Liao, Dani Lischinski, B. Chen, and D. Cohen-Or. Deep video-based performance cloning. Computer Graphics Forum, 38, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 286, + 166 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 286, + 166 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 286, + 166 + ], + "type": "text", + "content": "[2] George Borshukov, Dan Piponi, Oystein Larsen, J. P. Lewis, and Christina Tempelaar-Lietz. Universal capture: image-based facial animation for \"the matrix reloaded\". In SIGGRAPH '03, 2003. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 166, + 286, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 166, + 286, + 198 + ], + "spans": [ + { + "bbox": [ + 53, + 166, + 286, + 198 + ], + "type": "text", + "content": "[3] Joel Carranza, Christian Theobalt, Marcus A. Magnor, and Hans-Peter Seidel. Free-viewpoint video of human actors. ACM SIGGRAPH 2003 Papers, 2003. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 198, + 286, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 286, + 229 + ], + "type": "text", + "content": "[4] Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A. Efros. Everybody dance now. ICCV, pages 5932-5941, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 230, + 286, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 230, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 230, + 286, + 293 + ], + "type": "text", + "content": "[5] Eric Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, S. Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. ArXiv, abs/2112.07945, 2021. 2, 4, 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 294, + 286, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 286, + 327 + ], + "type": "text", + "content": "[6] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Linchao Bao, and Huchuan Lu. Animatable neural radiance fields from monocular rgb video. ArXiv, abs/2106.13629, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 327, + 286, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 286, + 347 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 286, + 347 + ], + "type": "text", + "content": "[7] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. CVPR, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 347, + 286, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 286, + 390 + ], + "type": "text", + "content": "[8] Boyang Deng, John P Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. Nasa neural articulated shape approximation. In ECCV, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 390, + 286, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 390, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 53, + 390, + 286, + 422 + ], + "type": "text", + "content": "[9] Chen Geng, Sida Peng, Zhen Xu, Hujun Bao, and Xiaowei Zhou. Learning neural volumetric representations of dynamic humans in minutes. In CVPR, 2023. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 422, + 286, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 286, + 464 + ], + "type": "text", + "content": "[10] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 465, + 286, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 286, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 286, + 508 + ], + "type": "text", + "content": "[11] A. K. Grigor'ev, Artem Sevastopolsky, Alexander Vakhitov, and Victor S. Lempitsky. Coordinate-based texture inpainting for pose-guided human image generation. CVPR, pages 12127-12136, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 508, + 286, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 508, + 286, + 549 + ], + "spans": [ + { + "bbox": [ + 48, + 508, + 286, + 549 + ], + "type": "text", + "content": "[12] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. ArXiv, abs/2110.08985, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 550, + 286, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 550, + 286, + 594 + ], + "spans": [ + { + "bbox": [ + 48, + 550, + 286, + 594 + ], + "type": "text", + "content": "[13] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics (TOG), 40:1-16, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 594, + 286, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 594, + 286, + 636 + ], + "spans": [ + { + "bbox": [ + 48, + 594, + 286, + 636 + ], + "type": "text", + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, 2017. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 637, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 637, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 637, + 286, + 668 + ], + "type": "text", + "content": "[15] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. ArXiv, abs/2112.05637, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "type": "text", + "content": "[16] Tao Hu, Geng Lin, Zhizhong Han, and Matthias Zwicker. Learning to generate dense point clouds with textures on multiple categories. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)," + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 703 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 326, + 73, + 454, + 84 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 454, + 84 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 454, + 84 + ], + "type": "text", + "content": "pages 2170-2179, January 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 84, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 84, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 308, + 84, + 545, + 116 + ], + "type": "text", + "content": "[17] Tao Hu, Kripasindhu Sarkar, Lingjie Liu, Matthias Zwicker, and Christian Theobalt. Egorenderer: Rendering human avatars from egocentric camera images. In ICCV, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 116, + 545, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 116, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 308, + 116, + 545, + 170 + ], + "type": "text", + "content": "[18] Tao Hu, Hongyi Xu, Linjie Luo, Tao Yu, Zerong Zheng, He Zhang, Yebin Liu, and Matthias Zwicker. Hvtr++: Image and pose driven human avatars using hybrid volumetric-textural rendering. IEEE Transactions on Visualization and Computer Graphics, pages 1-15, 2023. 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 170, + 545, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 170, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 308, + 170, + 545, + 201 + ], + "type": "text", + "content": "[19] T. Hu, Tao Yu, Zerong Zheng, He Zhang, Yebin Liu, and Matthias Zwicker. Hvtr: Hybrid volumetric-textural rendering for human avatars. 3DV, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 201, + 545, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 201, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 308, + 201, + 545, + 233 + ], + "type": "text", + "content": "[20] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. 2020 (CVPR), pages 3090-3099, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 233, + 545, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 233, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 308, + 233, + 545, + 265 + ], + "type": "text", + "content": "[21] James T. Kajiya and Brian Von Herzen. Ray tracing volume densities. Proceedings of the 11th annual conference on Computer graphics and interactive techniques, 1984. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 265, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 545, + 285 + ], + "type": "text", + "content": "[22] Diederick P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 285, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 545, + 318 + ], + "type": "text", + "content": "[23] Bernhard Kratzwald, Zhiwu Huang, Danda Pani Paudel, and Luc Van Gool. Towards an understanding of our world by GANing videos in the wild. arXiv:1711.11453, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 318, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 318, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 308, + 318, + 545, + 350 + ], + "type": "text", + "content": "[24] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Learn to dance with aist++: Music conditioned 3d dance generation, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[25] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. TOG, 40, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 392, + 545, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 392, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 308, + 392, + 545, + 457 + ], + "type": "text", + "content": "[26] Lingjie Liu, Weipeng Xu, Marc Habermann, Michael Zollhöfer, Florian Bernard, Hyeongwoo Kim, Wenping Wang, and Christian Theobalt. Neural human video rendering by learning dynamic textures and rendering-to-video translation. IEEE Transactions on Visualization and Computer Graphics, 05 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 457, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 457, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 457, + 545, + 510 + ], + "type": "text", + "content": "[27] Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Hyeongwoo Kim, Florian Bernard, Marc Habermann, Wenping Wang, and Christian Theobalt. Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG), 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 511, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 543 + ], + "type": "text", + "content": "[28] M. Loper, Naureen Mahmood, J. Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multi-person linear model. ACM Trans. Graph., 34:248:1-16, 2015. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 543, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 543, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 308, + 543, + 545, + 575 + ], + "type": "text", + "content": "[29] Liqian Ma, Xu Jia, Qianru Sun, Bernt Schiele, Tinne Tuytelaars, and Luc Van Gool. Pose guided person image generation. In NeurIPS, pages 405-415, 2017. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 575, + 545, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 575, + 545, + 607 + ], + "spans": [ + { + "bbox": [ + 308, + 575, + 545, + 607 + ], + "type": "text", + "content": "[30] Liqian Ma, Qianru Sun, Stamatios Georgoulis, Luc van Gool, Bernt Schiele, and Mario Fritz. Disentangled person image generation. CVPR, 2018. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 607, + 545, + 649 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 607, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 308, + 607, + 545, + 649 + ], + "type": "text", + "content": "[31] Qianli Ma, Shunsuke Saito, Jinlong Yang, Siyu Tang, and Michael J. Black. Scale: Modeling clothed humans with a surface codec of articulated local elements. In CVPR, 2021. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 649, + 545, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 649, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 308, + 649, + 545, + 681 + ], + "type": "text", + "content": "[32] Qianli Ma, Jinlong Yang, Siyu Tang, and Michael J Black. The power of points for modeling humans in clothing. In ICCV, 2021. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 681, + 545, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 703 + ], + "type": "text", + "content": "[33] Lars M. Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy net" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6558" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 704 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "text", + "content": "works: Learning 3d reconstruction in function space. CVPR, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 137 + ], + "type": "text", + "content": "[34] Mateusz Michalkiewicz, Zhony Kaesemodel Pontes, Dominic Jack, Mahsa Baktash, and Anders P. Eriksson. Deep level sets: Implicit surface representations for 3d shape inference. ArXiv, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 137, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 137, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 48, + 137, + 287, + 168 + ], + "type": "text", + "content": "[35] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. Leap: Learning articulated occupancy of people. In CVPR, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 170, + 286, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 170, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 48, + 170, + 286, + 190 + ], + "type": "text", + "content": "[36] Natalia Neverova, Riza Alp Güler, and Iasonas Kokkinos. Dense pose transfer. ECCV, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 190, + 287, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 190, + 287, + 221 + ], + "spans": [ + { + "bbox": [ + 48, + 190, + 287, + 221 + ], + "type": "text", + "content": "[37] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. CVPR, pages 11448-11459, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 222, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 222, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 287, + 253 + ], + "type": "text", + "content": "[38] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In IEEE/CVF ICCV, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "text", + "content": "[39] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. ArXiv, abs/2112.11427, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 297, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 328 + ], + "type": "text", + "content": "[40] Pablo Palafox, Aljaž Božić, Justus Thies, Matthias Nießner, and Angela Dai. Npms: Neural parametric models for 3d deformable shapes. In IEEE/CVF ICCV, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 329, + 287, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 329, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 329, + 287, + 371 + ], + "type": "text", + "content": "[41] Jeong Joon Park, Peter R. Florence, Julian Straub, Richard A. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. 2019 (CVPR), pages 165-174, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 372, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 287, + 413 + ], + "type": "text", + "content": "[42] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 415, + 287, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 467 + ], + "type": "text", + "content": "[43] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. CVPR, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 468, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 498 + ], + "type": "text", + "content": "[44] Sergey Prokudin, Michael J. Black, and Javier Romero. Smplpix: Neural avatars from 3d human models. WACV, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 499, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 532 + ], + "type": "text", + "content": "[45] Albert Pumarola, Antonio Agudo, Alberto Sanfeliu, and Francesc Moreno-Noguer. Unsupervised person image synthesis in arbitrary poses. In CVPR, June 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 533, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 563 + ], + "type": "text", + "content": "[46] Amit Raj, Julian Tanke, James Hays, Minh Vo, Carsten Stoll, and Christoph Lassner. Anr: Articulated neural rendering for virtual avatars. CVPR, pages 3721-3730, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 564, + 287, + 617 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 564, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 48, + 564, + 287, + 617 + ], + "type": "text", + "content": "[47] Edoardo Remelli, Timur M. Bagautdinov, Shunsuke Saito, Chenglei Wu, Tomas Simon, Shih-En Wei, Kaiwen Guo, Zhe Cao, Fabian Prada, Jason M. Saragih, and Yaser Sheikh. Drivable volumetric avatars using texel-aligned features. ACM SIGGRAPH, 2022. 2, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 618, + 287, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 618, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 48, + 618, + 287, + 660 + ], + "type": "text", + "content": "[48] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. IEEE/CVF ICCV, pages 2304-2314, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 661, + 287, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 661, + 287, + 704 + ], + "spans": [ + { + "bbox": [ + 48, + 661, + 287, + 704 + ], + "type": "text", + "content": "[49] Shunsuke Saito, Tomas Simon, Jason M. Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. 2020 (CVPR), pages 81-90, 2020." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 705 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[50] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J. Black. Scintimate: Weakly supervised learning of skinned clothed avatar networks. 2021 (CVPR), pages 2885-2896, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 117, + 545, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 117, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 307, + 117, + 545, + 158 + ], + "type": "text", + "content": "[51] Kripasindhu Sarkar, Dushyant Mehta, Weipeng Xu, Vladislav Golyanik, and Christian Theobalt. Neural re- rendering of humans from a single image. In ECCV, 2020. 2,6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 159, + 545, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 159, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 159, + 545, + 191 + ], + "type": "text", + "content": "[52] Aliaksandr Siarohin, Enver Sangineto, Stephane Lathuiliere, and Nicu Sebe. Deformable GANs for pose-based human image generation. In CVPR, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 192, + 545, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 192, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 307, + 192, + 545, + 233 + ], + "type": "text", + "content": "[53] Shih-Yang Su, Frank Yu, Michael Zollhoefer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 234, + 545, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 234, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 545, + 310 + ], + "type": "text", + "content": "[54] Ayush Tewari, Ohad Fried, Justus Thies, Vincent Sitzmann, Stephen Lombardi, Kalyan Sunkavalli, Ricardo MartinBrualla, Tomas Simon, Jason M. Saragih, Matthias Nießner, Rohit Pandey, S. Fanello, Gordon Wetzstein, Jun-Yan Zhu, Christian Theobalt, Maneesh Agrawala, Eli Shechtman, Dan B. Goldman, and Michael Zollhofer. State of the art on neural rendering. Computer Graphics Forum, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 311, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 311, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 311, + 545, + 342 + ], + "type": "text", + "content": "[55] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: image synthesis using neural textures. ACM Transactions on Graphics (TOG), 38, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 342, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 373 + ], + "type": "text", + "content": "[56] Garvita Tiwari, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Neural-gif: Neural generalized implicit functions for animating people in clothing. In ICCV, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 374, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 374, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 374, + 545, + 415 + ], + "type": "text", + "content": "[57] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Metaatrix: Learning animatable clothed human models from few depth images. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 417, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 417, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 307, + 417, + 545, + 458 + ], + "type": "text", + "content": "[58] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In European Conference on Computer Vision, 2022. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 460, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 491 + ], + "type": "text", + "content": "[59] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In NeurIPS, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 491, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 491, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 491, + 545, + 533 + ], + "type": "text", + "content": "[60] Zhou Wang, A. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13:600-612, 2004. 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 534, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 577 + ], + "type": "text", + "content": "[61] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humaner: Free-viewpoint rendering of moving people from monocular video. ArXiv, abs/2201.04127, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 577, + 545, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 577, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 307, + 577, + 545, + 630 + ], + "type": "text", + "content": "[62] Feng Xu, Yebin Liu, Carsten Stoll, James Tompkin, Gaurav Bharaj, Qionghai Dai, Hans-Peter Seidel, Jan Kautz, and Christian Theobalt. Video-based characters: creating new human performances from a multi-view video database. ACM SIGGRAPH, 2011. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 631, + 545, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 631, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 307, + 631, + 545, + 695 + ], + "type": "text", + "content": "[63] Jae Shin Yoon, Duygu Ceylan, Tuanfeng Y. Wang, Jingwan Lu, Jimei Yang, Zhixin Shu, and Hyunjung Park. Learning motion-dependent appearance for high-fidelity rendering of dynamic humans from a single camera. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3397-3407, 2022. 2, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 696, + 545, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 696, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 307, + 696, + 545, + 705 + ], + "type": "text", + "content": "[64] Richard Zhang, Phillip Isola, Alexei A. Efros, E. Shechtman," + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6559" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 285 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 104 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 104 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 104 + ], + "type": "text", + "content": "and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. CVPR, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 286, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 286, + 169 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 286, + 169 + ], + "type": "text", + "content": "[65] Yang Zheng, Ruizhi Shao, Yuxiang Zhang, Tao Yu, Zerong Zheng, Qionghai Dai, and Yebin Liu. Deepmulticap: Performance capture of multiple characters using sparse multiview cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6239-6249, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 171, + 286, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 171, + 286, + 212 + ], + "spans": [ + { + "bbox": [ + 49, + 171, + 286, + 212 + ], + "type": "text", + "content": "[66] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. TPAMI, PP, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 213, + 286, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 213, + 286, + 245 + ], + "spans": [ + { + "bbox": [ + 49, + 213, + 286, + 245 + ], + "type": "text", + "content": "[67] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. Cips3d: A 3d-aware generator of gans based on conditionally-independent pixel synthesis. ArXiv, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 246, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 246, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 246, + 286, + 285 + ], + "type": "text", + "content": "[68] Zhen Zhu, Tengteng Huang, Baoguang Shi, Miao Yu, Bofei Wang, and Xiang Bai. Progressive pose attention transfer for person image generation. In CVPR, pages 2347-2356, 2019. 2" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6560" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/213de3e8-6a97-4d16-96b2-ac03f2f0441a_content_list.json b/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/213de3e8-6a97-4d16-96b2-ac03f2f0441a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dee280eeffab99f640de787574facd7088e6aa37 --- /dev/null +++ b/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/213de3e8-6a97-4d16-96b2-ac03f2f0441a_content_list.json @@ -0,0 +1,1807 @@ +[ + { + "type": "text", + "text": "SurroundSDF: Implicit 3D Scene Understanding Based on Signed Distance Field", + "text_level": 1, + "bbox": [ + 78, + 130, + 890, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lizhe Liu $^{1*}$ Bohua Wang $^{2*§}$ Hongwei Xie $^{1*}$ Daqi Liu $^{1}$ Li Liu $^{1}$ Zhiqiang Tian $^{2}$ Kuiyuan Yang $^{1}$ Bing Wang $^{1\\dagger}$ $^{1}$ Xiaomi EV $^{2}$ School of Software Engineering, Xi'an Jiaotong University", + "bbox": [ + 138, + 178, + 830, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-centric 3D environment understanding is both vital and challenging for autonomous driving systems. Recently, object-free methods have attracted considerable attention. Such methods perceive the world by predicting the semantics of discrete voxel grids but fail to construct continuous and accurate obstacle surfaces. To this end, in this paper, we propose SurroundSDF to implicitly predict the signed distance field (SDF) and semantic field for the continuous perception from surround images. Specifically, we introduce a query-based approach and utilize SDF constrained by the Eikonal formulation to accurately describe the surfaces of obstacles. Furthermore, considering the absence of precise SDF ground truth, we propose a novel weakly supervised paradigm for SDF, referred to as the Sandwich Eikonal formulation, which emphasizes applying correct and dense constraints on both sides of the surface, thereby enhancing the perceptual accuracy of the surface. Experiments suggest that our method achieves SOTA for both occupancy prediction and 3D scene reconstruction tasks on the nuScenes dataset.", + "bbox": [ + 75, + 300, + 473, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 631, + 209, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the recent advancement of 3D object detection algorithms [10, 11, 16, 18, 21, 24], vision-centric autonomous driving system become more practicable. Nevertheless, the persisting challenges related to the long-tail problem and the coarse depiction of the 3D scene underscore its insufficiency. Consequently, a deeper comprehension of 3D geometry and semantics is needed for safety and reliability. This paper delves into a novel vision-centric paradigm of dense and continuous 3D scene understanding.", + "bbox": [ + 75, + 656, + 468, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current approaches of dense 3D prediction can be classified into two categories, 3D reconstruction and 3D perception. Specifically, 3D reconstruction algorithms [6, 33] generate dense point clouds enriched with semantic informa", + "bbox": [ + 75, + 792, + 468, + 853 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/610617bf759b236a5682727c123852d1071a6c659b5b9d9eb314eb7248ad29ab.jpg", + "image_caption": [ + "a. SIREN" + ], + "image_footnote": [], + "bbox": [ + 526, + 270, + 676, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c857769f95962bfe4c30e734d2055cba2c8c6a78a6c87fe0c1fb3734e4b01bfa.jpg", + "image_caption": [ + "b. LODE" + ], + "image_footnote": [], + "bbox": [ + 686, + 268, + 861, + 361 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/77fc15699dcad5e9666cf396005dd220778f551e41a701e1f94cefaa21596f1f.jpg", + "image_caption": [ + "c. Ours", + "Figure 1. The scene perspective results from surround images input. a, b. The result of SIREN and LODE supervision with camera-only input (the original methods are point cloud-based). c. Our result. d. Our result with semantics." + ], + "image_footnote": [], + "bbox": [ + 526, + 375, + 686, + 463 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6c9f094630e34f0067725fc7641387fb142eec39ca21291330ed3000111c7b63.jpg", + "image_caption": [ + "d. Ours with semantic" + ], + "image_footnote": [], + "bbox": [ + 705, + 373, + 867, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion by projecting depth maps and 2D semantic maps into 3D space. 3D perception methods [3, 12, 17, 25, 35, 41, 46] predict the occupied status and the semantic of 3D voxel grids. However, either method proves to be redundant or can only predict coarse-grained discrete grids.", + "bbox": [ + 496, + 537, + 890, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Instead, we develop a vision-centric framework to describe 3D scenes leveraging neural implicit Signed Distance Function (SDF) representation, which we refer to as SurroundSDF. Generally, we aim to: (1) Continuously describe the 3D scene by reconstructing smooth surfaces; (2) Explore the difficulty associated with the utilization of SDF representation and pose an appropriate training strategy; (3) Exploit the strong representation to kill two birds with one stone, addressing 3D semantic segmentation and continuous 3D geometry reconstruction within one framework.", + "bbox": [ + 496, + 613, + 892, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In pursuit of the first and the third objectives, we employ SDF to represent the canonical 3D scene in terms of distance from the surface. Furthermore, we construct an implicit field that encompasses both semantics and geometry by exploiting this strong representation. However, it's still nontrivial to accomplish the SDF modeling, where complete surface and normal vectors, which are crucial supervisions, cannot be accurately computed. To this end, SIREN [34] proposed a weakly supervised implicit per", + "bbox": [ + 496, + 765, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution.", + "bbox": [ + 94, + 862, + 209, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{\\S}$ Work done during an internship at Xiaomi EV.", + "bbox": [ + 96, + 875, + 346, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author.", + "bbox": [ + 96, + 887, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21614", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ception method, which accomplishes 3D mesh completion from point clouds. While LODE [14] proposed to utilize occupancy ground truth (GT) to supervise SDF. However, these methods rely on the input of point clouds, and only alleviate this difficulty by mimicking the surface distance but are either sparse or inaccurate.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the aforementioned challenges, we propose the Sandwich Eikonal formulation, a novel weak supervision paradigm for SDF modeling. Figure 1 gives some reconstruction results and shows the benefit of our method. This method emphasizes applying correct and dense constraints on both sides of the surface to enhance the geometric accuracy and continuity of the surface. Moreover, we revisit current pipelines for training the perception branch and design a novel loss that enhances the integration of geometry and semantics thereby reducing inconsistencies. Our contributions can be summarized as follows:", + "bbox": [ + 75, + 181, + 470, + 347 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a vision-centric implicit semantic SDF perception method, achieving accurate and continuous 3D perception. To the best of our knowledge, we are the first to utilize SDF for surround-view 3D perception.", + "- We introduce the Sandwich Eikonal formulation, a novel weak supervision paradigm for SDF.", + "- We demonstrate how to employ this representation for the reconstruction of continuous 3D geometry with precise semantic information.", + "- We achieve state-of-the-art results on 3D dense semantics perception tasks. Comprehensive experiments on the NuScenes dataset [2] provide extensive validation of our approach." + ], + "bbox": [ + 76, + 349, + 467, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 558, + 217, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Occupancy Prediction Recently, researches [3, 12, 17, 25, 32, 32, 35, 41, 46] on occupancy prediction have demonstrated advantages in 3D scene understanding. Compared to the traditional object detection paradigm [5, 10, 11, 15, 16, 18, 21, 22, 24, 39, 42], occupancy perception has the following advantages. First, it can express dense 3D geometry. Second, it can accurately provide spatial locations for objects beyond predefined categories. Third, it can describe the shapes of irregular obstacles.", + "bbox": [ + 75, + 583, + 467, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on these advantages, occupancy prediction tasks have attracted significant research. These methods predict the occupancy status in the region of interest around the ego vehicle from point clouds or images. Specifically, the space is first divided into voxel grids at a specific resolution. Then the occupancy status and semantics of each grid are estimated. SurroundOcc [41] proposed a surround-view 3D occupancy perception method that utilizes spatial 2D-3D attention to lift image features into 3D space. In addition, to realize the dense occupancy prediction, SurroundOcc designed a pipeline to convert the point cloud to dense occupancy ground truth. VoxFormer [17] employed", + "bbox": [ + 75, + 719, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "an MAE-like [9] approach to achieve camera-based semantic occupancy prediction. FB-OCC [20, 20] proposed a novel forward-backward projection method to compensate for the insufficient BEV feature density of the forward projection method and the large number of mismatches in 2D and 3D space caused by the backward projection. Despite the impressive results, they are limited by the specific resolution of the occupancy annotations, which limit the continuous perception of the scene.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicit Scene Perception Scene reconstruction refers to the task of predicting the 3D geometry structures from some incomplete representations, e.g., images, point clouds, and voxel grids. Implicit neural scene reconstruction methods [1, 4, 14, 26, 30, 34, 45] have demonstrated advantages in accuracy and substantial potential. Generally, they train a neural network to predict a continuous field. Thus, it is possible to query occupancy information for any point within the 3D space. DeepSDF [30] utilizes the SDF to achieve the implicit reconstruction of 3D objects at the instance level. NeRFs [26] and its numerous variants [4, 7, 43] achieve implicit 3D scene reconstruction and novel view synthesis from multi-view images. NeuS [38] and VolSDF [44] introduce the signed distance function into neural rendering, enabling more precise object surface reconstruction. SIREN [34] utilizes the Eikonal equation to achieve semi-supervised SDF reconstruction for point clouds. However, these methods lack generalization to novel scenes. To achieve online implicit scene perception in driving scenes, LODE [14] proposes the Locally Conditioned Eikonal formulation and introduces a dense occupancy ground truth for supervision, which significantly improves the effect of SDF reconstruction in the driving scenes. However, it is limited by the sparsity of input LiDAR points or inaccurate ground truth. To solve these problems, we propose a camera-only implicit scene understanding method, which aims at continuous and accurate surface perception.", + "bbox": [ + 496, + 229, + 890, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Formulation", + "text_level": 1, + "bbox": [ + 500, + 645, + 629, + 660 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section analyzes the SDF supervision paradigm based on the Eikonal formulation with its variations and introduces our Sandwich Eikonal supervision approach.", + "bbox": [ + 496, + 667, + 890, + 713 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Eikonal-based SDF Constraints", + "text_level": 1, + "bbox": [ + 500, + 717, + 777, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a coordinate $\\mathbf{x}$ in the 3D scene, our goal is to construct a function $\\phi$ such that $\\phi(\\mathbf{x})$ provides the SDF value at point $\\mathbf{x}$ . The Eikonal-based optimization objectives [14] is", + "bbox": [ + 496, + 736, + 890, + 781 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {O} _ {E} = \\int_ {\\Omega_ {0}} O _ {0} d \\mathbf {x} + \\int_ {\\Omega_ {1}} O _ {1} d \\mathbf {x} + \\int_ {\\Omega_ {2}} O _ {2} d \\mathbf {x}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 782, + 890, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and", + "bbox": [ + 500, + 806, + 527, + 818 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} O _ {0} = | | \\nabla_ {\\mathbf {x}} \\phi (\\mathbf {x}) | - 1 |, & \\mathbf {x} \\in \\Omega_ {0} \\\\ O _ {1} = | \\nabla_ {\\mathbf {x}} \\phi (\\mathbf {x}) - \\mathbf {n} (\\mathbf {x}) |, & \\mathbf {x} \\in \\Omega_ {1}, \\\\ O _ {2} = | \\phi (\\mathbf {x}) - S D F (\\mathbf {x}) |, & \\mathbf {x} \\in \\Omega_ {2} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 816, + 890, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\Omega_0$ is the set of the whole 3D space of interest, $\\Omega_1$ is the set of points on the surface, $\\Omega_2$ is the set of points", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21615", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/88e186e3c3371d5010599239bd61c556c07e624a95a5fd5bd6be5a9e4ee3e907.jpg", + "image_caption": [ + "a. Ideal supervision" + ], + "image_footnote": [], + "bbox": [ + 161, + 94, + 294, + 196 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b970adbac643898fd614ba27df17c68bef3bf71c1d87bbc0c586eeba0076deee.jpg", + "image_caption": [ + "b. SIREN", + "Figure 2. a. Ideal supervision for SDF where the SDF GT is provided. b. SIREN supervision which uses LiDAR points to supervise the surface. c. LODE supervision based on occupancy GT. d. Our supervision paradigm which combines the LiDAR GT and occupancy GT, it is closer to the ideal supervision." + ], + "image_footnote": [], + "bbox": [ + 295, + 95, + 428, + 196 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3cca8fda26a26d66195b1b41dfa91bbc136a3bb0ab2eca56ce7bb3b06f1afb5c.jpg", + "image_caption": [ + "c.LODE" + ], + "image_footnote": [], + "bbox": [ + 429, + 95, + 563, + 196 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d129ffcfe9fa1e392b0941b20461847b9119f2cf7090edb61335d72f288fc181.jpg", + "image_caption": [ + "d. Ours" + ], + "image_footnote": [ + "Constrain $SDF < 0$", + "Constrain $SDF = 0$", + "Constrain $SDF > 0$", + "Constraint min SDF", + "of one grid $< 0$", + "Surface", + "LiDAR Points", + "Voxel Grid" + ], + "bbox": [ + 566, + 95, + 700, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with SDF GT annotations, $\\nabla_{\\mathbf{x}}\\phi (\\mathbf{x})$ is the gradient on point $\\mathbf{x}$ , $\\mathbf{n}(\\mathbf{x})$ is the normal, and $SDF(\\mathbf{x})$ is the real SDF value at point $\\mathbf{x}$ . One major obstacle to this optimization objective lies in the design of $O_2$ , which provides direct supervision for SDF as shown in Figure 2.a. This objective requires the annotation of SDF values for each anchor point in the scene, which is monumental and challenging.", + "bbox": [ + 75, + 267, + 468, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the absence of precise SDF annotations, SIREN [34] and LODE [14] have implemented weakly supervised paradigm through the following variations of $O_2$ ,", + "bbox": [ + 76, + 373, + 468, + 419 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} O _ {2 - 0} ^ {\\prime} = | \\phi (\\mathbf {x}) |, & \\mathbf {x} \\in \\Omega_ {1} \\\\ O _ {2 - 1} ^ {\\prime} = \\psi (| \\phi (\\mathbf {x}) |), & \\mathbf {x} \\in \\Omega_ {3}, \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 424, + 468, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\psi (\\cdot)$ is a monotonically decreasing function and $\\Omega_3\\subseteq \\Omega_0 / \\Omega_1$ is the whole space except the surface.", + "bbox": [ + 75, + 460, + 468, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The objective $O_{2-0}^{\\prime}$ aims to constrain the SDF values on the surface to 0. Considering the absence of precise supervision, $O_{2-1}^{\\prime}$ adopts a fuzzy constraint that pushes the absolute SDF value away from zero in $\\Omega_3$ . While the constraint in $\\Omega_3$ is loose, it is adequate to capture the geometry information of the surface.", + "bbox": [ + 75, + 491, + 468, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practical implementation, SIREN [34] samples on-surface LiDAR points in $\\Omega_{1}$ and considers the space excluding the LiDAR points as $\\Omega_{3}$ , as is shown in Figure 2.b. When the LiDAR points are dense enough, this strategy can effectively achieve good constraint effects [34]. However, LODE [14] has demonstrated that the sparsity of LiDAR points in the driving scenes results in discontinuities of the estimated mesh. To achieve denser supervision, they use the dense occupancy GT as supervision. This approach treats both the surface and interior of the obstacles as the $\\Omega_{1}$ . Specifically, they sample the $\\Omega_{1}$ points as the center of \"occupied\" grids, and the $\\Omega_{3}$ points as the center of \"free\" grids, as is shown in 2.c. However, this paradigm remains imprecise, as considering the center of the occupied grid to be the surface of objects would introduce errors.", + "bbox": [ + 75, + 582, + 468, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Sandwich Eikonal Formulation", + "text_level": 1, + "bbox": [ + 76, + 816, + 352, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the above concerns, we require a supervision paradigm that has the following characteristics: 1) sampling points as the supervision on the surface must strictly adhere to the surface, 2) for regions beyond the surface, refrain", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "from supervision in uncertain areas to ensure precision, and 3) dense supervision signals. Considering the above factors, we introduce our variants of Eikonal formulation as follows:", + "bbox": [ + 498, + 267, + 890, + 311 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} O _ {2 - 0} = \\phi (\\mathbf {x}), & \\mathbf {x} \\in \\Omega_ {1} \\\\ O _ {2 - 1} = \\psi (- \\min (\\phi (\\mathbf {x}))), & \\mathbf {x} \\in \\Omega_ {\\text {o c c}} ^ {i} \\\\ O _ {2 - 2} = \\psi (\\phi (\\mathbf {x})), & \\mathbf {x} \\in \\Omega_ {\\text {f r e e}} ^ {j}, \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 325, + 890, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Omega_{occ}^{i}$ is the $i$ -th grid of occupied grids and $\\Omega_{free}^{j}$ is the $j$ -th grid of free grids.", + "bbox": [ + 498, + 391, + 890, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we incorporate the LiDAR points and the occupancy GT into the SDF supervision. As is shown in Figure 2.d, for precise sampling points on the surface, we follow SIREN and constrain the SDF value of LiDAR points to be 0. To apply accurate and dense supervision to regions beyond the surface, we introduce the occupancy GT into the supervision, as is shown in optimization objectives $O_{2-1}$ and $O_{2-2}$ . According to the generation process of occupancy GT [35, 36], if a grid is labeled as \"occupied\", it implies that at least part of the grid area falls inside the object. Therefore, as the objective $O_{2-1}$ shows, we constrain the minimum SDF value of each occupied grid to be less than zero and push it to the negative range. On the other hand, if a grid is labeled as \"free\", the entire grid is outside the object and we push the SDF value of free grids to the positive range, as is shown in objective $O_{2-2}$ . We employ a multiframe point cloud fusion strategy to densify the sampling for surface supervision, as described in Section 4.2. Compared to previous supervision paradigms, our formulation emphasizes applying appropriate and dense supervision on both sides of the surface to enhance the geometric constraint based on the Eikonal formulation. Therefore, we name our supervision strategy the \"Sandwich Eikonal\" formulation.", + "bbox": [ + 496, + 425, + 892, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 796, + 589, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given multi-camera images $\\mathbf{I} = \\{I_1, I_2, \\dots, I_n\\}$ , we aim to predict the 3D SDF $\\phi(\\mathbf{x}) : \\mathbb{R}^3 \\Rightarrow \\mathbb{R}$ and semantic field $S(\\mathbf{x}) : \\mathbb{R}^3 \\Rightarrow \\mathbb{R}^s$ , where $s$ indicates the number of classes. The following will introduce our architecture and detailed constraints based on our Sandwich Eikonal formulation.", + "bbox": [ + 498, + 824, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21616", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ef8d1bbee8318de1f774218ef7b57ff1e49655ead98b147e8d2ca998b827fe49.jpg", + "image_caption": [ + "Figure 3. The architecture of our SurroundSDF. Given the surround images as input, an encoder composed of a 2D backbone, LSS module, and BEV backbone, is employed to extract voxel features. We adopt a query-based approach to sample features from the voxel features. Specifically, first, a set of query coordinates in the region of interest is selected. Subsequently, using trilinear interpolation, semantic features are queried from the voxel features. Finally, after concatenation with the positional embeddings from the query coordinates, the features pass through the SDF head and semantic head respectively, yielding SDF and semantic fields. For training, the query coordinates are sampled according to the GT, and the SDF and semantic field are supervised by the losses introduced in Section 4.2. In the inference phase, based on appropriate sampling and post-processing, continuous and accurate scene perception results are obtained (see Section 4.3)." + ], + "image_footnote": [], + "bbox": [ + 148, + 89, + 823, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Architecture", + "text_level": 1, + "bbox": [ + 76, + 407, + 212, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall architecture is shown in Figure 3. First, a CNN-based backbone is used to obtain the image features. Then, the LSS [31] is used to project the image features to BEV space, and a BEV encoder is adopted to obtain the voxel features $V$ of dimension $h \\times w \\times d \\times c$ (c indicates the number of channels). Further, for each point $\\mathbf{x}$ in $\\Omega_0$ , the corresponding features can be obtained by querying $V$ with the trilinear sampling method. Therefore, $\\phi(\\mathbf{x})$ and $S(\\mathbf{x})$ can be expressed as:", + "bbox": [ + 75, + 430, + 470, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi (\\mathbf {x}) = H _ {s d f} \\left(C \\left(Q (V, \\mathbf {x}), P E (\\mathbf {x})\\right), \\theta\\right), \\mathbf {x} \\in \\Omega_ {0}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 574, + 468, + 598 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS (\\mathbf {x}) = H _ {s e m} (C (Q (V, \\mathbf {x}), P E (\\mathbf {x})), \\theta), \\mathbf {x} \\in \\Omega_ {0},\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 593, + 431, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{x} \\in \\mathbb{R}^3$ , $C$ and $Q$ represent the concatenation and query process, respectively. $H_{sdf}$ and $H_{sem}$ indicate the SDF head and semantic head, implemented with three-layer MLPs with sine activation functions. $\\theta$ represents the learnable parameters. $PE$ represents the positional encoding, which aims to capture high-frequency information. Specifically, for a coordinate $\\mathbf{x} \\in \\mathbb{R}^3$ , $PE$ encodes each of its dimensions into a 2n-dimensional vector, it can be expressed", + "bbox": [ + 75, + 619, + 468, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {a s :} \\quad P E (\\mathbf {x}) = (\\zeta (x), \\zeta (y), \\zeta (z)) \\in \\mathbb {R} ^ {6 n}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 744, + 468, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x, y, z$ denote the coordinates, $\\zeta(\\cdot)$ represents an encoding scheme using trigonometric function mapping:", + "bbox": [ + 75, + 765, + 468, + 795 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\zeta (x) = \\left(\\sin \\left(2 ^ {0} \\pi x\\right), \\cos \\left(2 ^ {0} \\pi x\\right), \\dots , \\cos \\left(2 ^ {n - 1} \\pi x\\right)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 800, + 468, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In training, we design a joint supervision method that utilizes both coarse-grained voxel and fine-grained point cloud supervision based on our Sandwich Eikonal formulation. In inference, we query the voxel features to estimate the SDF values and semantic logits of each query point.", + "bbox": [ + 75, + 825, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Supervision", + "text_level": 1, + "bbox": [ + 498, + 407, + 627, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section will elaborate on deriving the SDF loss from our Sandwich Eikonal formulation. Then, we will introduce the semantic loss and our joint supervision strategy.", + "bbox": [ + 496, + 430, + 892, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SDF Supervision Following objectives $O_0$ , $O_1$ in Equation 2 and $O_{2-0}$ to $O_{2-2}$ in Equation 4, we derive the losses for SDF supervision, based on our Sandwich Eikonal formulation. Considering that continuous space supervision is unfeasible, we sample the discrete points in the corresponding space $(\\Omega_0, \\Omega_1, \\Omega_{occ}$ and $\\Omega_{free})$ for supervision.", + "bbox": [ + 496, + 483, + 892, + 575 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/733d3c8523d8de7ff252598d06e9c0b78cd419f865eda5194830af270b6af1e2.jpg", + "image_caption": [ + "a." + ], + "image_footnote": [], + "bbox": [ + 553, + 588, + 697, + 700 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b3fcc936e39af8d4b1a7d4428e8dbe49c23fe53a8f592b113fe12d2e18738be3.jpg", + "image_caption": [ + "b.", + "Figure 4. a. SDF constraints with Sandwich Eikonal formulation in continuous form. b. SDF sampling in discrete form." + ], + "image_footnote": [], + "bbox": [ + 709, + 590, + 836, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The correctness and density of the supervision sampling are of paramount importance. Since our approach is based on the surround cameras, we only sample from the camera-visible space in $\\Omega_0$ (space of interest) [36]. For $\\Omega_{1}$ (surface points), we sample exclusively from the LiDAR points. To enhance the sampling density, temporal point cloud fusion is adopted. Specifically, $\\Omega_{1} = \\{p_{t - k},p_{t - (k - 1)},\\dots ,p_{t + (k - 1)},p_{t + k}\\}$ , where $p_{t + j}$ is the LiDAR points of the $t + j$ frame and $t$ is the current frame. Based on occupancy GT [36], $\\Omega_{occ}$ and $\\Omega_{free}$ are the sets of", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21617", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "voxel grids labeled as \"occupied\" and \"free\", respectively.", + "bbox": [ + 76, + 90, + 460, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, in our objective $O_{2-1}$ , the computation of the minimum SDF within the occupied grid is required. As is shown in Figure 4, we approximate this minimum value through discrete sampling. The minimum value among the SDF values at the centers of each sub-grid is regarded as the minimum SDF value for the grid. Based on Equation 2, 4 and the above sampling strategy, our SDF loss $\\mathcal{L}_{sdf}$ is expressed as follows:", + "bbox": [ + 75, + 106, + 468, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {s d f} = \\gamma_ {1} \\frac {1}{n _ {t o t a l}} \\sum_ {i = 1} ^ {n _ {t o t a l}} \\left(\\left| \\nabla_ {\\mathbf {x}} \\phi (\\mathbf {x} ^ {i}) \\right| - 1\\right) \\\\ + \\gamma_ {2} \\frac {1}{n _ {1 \\cup o c c}} \\sum_ {i = 1} ^ {n _ {1 \\cup o c c}} (\\nabla_ {\\mathbf {x}} \\phi (\\mathbf {x} ^ {i}) - n (\\mathbf {x} ^ {i})) \\\\ + \\gamma_ {3} \\frac {1}{n _ {1}} \\sum_ {i = 1} ^ {n _ {1}} | \\phi (\\mathbf {x} ^ {i}) | \\tag {8} \\\\ + \\gamma_ {4} \\frac {1}{n _ {o c c}} \\sum_ {i = 1} ^ {n _ {o c c}} e ^ {\\alpha \\times m i n (\\phi (\\mathbf {x} ^ {i}), \\mathbf {x} ^ {i} \\in \\Omega_ {o c c} ^ {i})} \\\\ + \\gamma_ {5} \\frac {1}{n _ {f r e e}} \\sum_ {i = 1} ^ {n _ {f r e e}} e ^ {- \\alpha \\times r a n d o m (\\phi (\\mathbf {x} ^ {i}), \\mathbf {x} ^ {i} \\in \\Omega_ {f r e e} ^ {i})}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 229, + 468, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is a hyperparameter that determines the extent of the deviation for the interior and exterior space of the object and we fix $\\alpha$ to 100 in experiments. $\\gamma_{1},\\gamma_{2},\\gamma_{3},\\gamma_{4},\\gamma_{5}$ are loss weights, $n(\\mathbf{x})$ represents the normal vector of the current point on surface. $n_1,n_{occ}$ , and $n_{free}$ represent the number of points or voxel grids sampled from $\\Omega_1,\\Omega_{occ}$ and $\\Omega_{free}$ , respectively. $n_{total}$ is the total number of samples: $n_{total} = n_1 + n_{occ} + n_{free}$ .", + "bbox": [ + 75, + 421, + 468, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic Supervision A widely used cross-entropy loss is adopted to supervise the semantic field as follows:", + "bbox": [ + 75, + 556, + 468, + 587 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s e m} = \\frac {1}{n _ {o c c}} \\sum_ {i = 1} ^ {n _ {o c c}} C E \\left(S (\\mathbf {x} | \\mathbf {x} \\in \\Omega_ {o c c} ^ {i}), \\dot {y} _ {s e m} ^ {i}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 590, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where CE represents cross-entropy loss and $\\hat{y}_{sem}^i$ is the semantic ground truth of i-th voxel grid.", + "bbox": [ + 75, + 613, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Joint Supervision The above losses optimize the SDF and semantic field separately, which leads to a nonnegligible issue of ambiguity between geometric and semantic optimization. To illustrate this issue, we train a model and evaluate the occupancy IoU representing geometric accuracy and the semantic mIoU representing overall semantic accuracy following the occupancy prediction metrics [35, 36]. For each voxel grid, if its minimum SDF is less than a certain threshold, the voxel is considered 'occupied'. As is shown in Figure 5, the optimal thresholds producing the highest peak of these curves are significantly different. This indicates that greater geometric accuracy does not necessarily lead to improved overall semantic accuracy.", + "bbox": [ + 75, + 657, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To mitigate this issue, we propose a joint supervision strategy. First, for the minimum SDF value $\\min(\\phi(\\mathbf{x}))$ of each voxel grid, a SoftMax function is used to convert it to", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ccb7a9fa23c8ade7c27ec8c0d8fcd7de2555818bbb85abc5be7779f509fa86de.jpg", + "image_caption": [ + "Figure 5. Variation of occupancy IoU and semantic mIoU with SDF Threshold. Note that the peak values of these two indicators correspond to different SDF thresholds." + ], + "image_footnote": [], + "bbox": [ + 532, + 89, + 859, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the free probability: $p_{free} = \\text{softmax}(\\beta t, \\beta \\min(\\phi(\\mathbf{x}))$ , where $t$ is fixed to 0.005, $\\beta$ is fixed to 100. Second, we map the free probability $p_{free}$ to logits using an inv-sigmoid function: $l_{free} = \\log(p_{free} / (1 - p_{free}))$ . Finally, $l_{free}$ is combined with semantic logits $l_{sem}$ , and the joint loss $\\mathcal{L}_{joint}$ is expressed as:", + "bbox": [ + 498, + 258, + 892, + 351 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {j o i n t}} = \\operatorname {D i c e} \\left(C \\left(l _ {\\text {s e m}}, l _ {\\text {f r e e}}\\right), C \\left(\\hat {y} _ {\\text {s e m}}, \\hat {y}\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 354, + 890, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C$ is the concatenation operation and $\\hat{y}_{sem}$ is the semantic GT of voxel grid. Dice represents the Dice loss [27]. Finally, the total loss is calculated as:", + "bbox": [ + 498, + 383, + 890, + 429 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\gamma_ {6} \\mathcal {L} _ {s d f} + \\gamma_ {7} \\mathcal {L} _ {s e m} + \\gamma_ {8} \\mathcal {L} _ {j o i n t} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 435, + 890, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma_6, \\gamma_7, \\gamma_8$ are loss weights.", + "bbox": [ + 500, + 455, + 723, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Inference", + "text_level": 1, + "bbox": [ + 500, + 481, + 609, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the output SDF and semantic field, diverse outputs for different tasks can be obtained through different sampling strategies and post-processing methods.", + "bbox": [ + 498, + 503, + 890, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic Mesh Generation To generate the semantic mesh, we uniformly query the SDF value in the space of interest and use the marching cubes algorithm [28] to obtain the mesh. Subsequently, at the vertex positions of each triangular face, we query the corresponding semantic information, resulting in a mesh enriched with semantics.", + "bbox": [ + 498, + 554, + 890, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Occupancy Prediction The objective of the occupancy prediction is to anticipate the semantics of the voxel grid in space. We achieve this through the joint prediction based on the predicted SDF and semantic field as follows:", + "bbox": [ + 498, + 657, + 890, + 717 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {i} = \\operatorname {a r g m a x} \\left(C \\left(l _ {\\text {s e m}} ^ {i}, l _ {f} ^ {i}\\right)\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 722, + 890, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S_{i}$ is the i-th voxel grid, $l_{sem}^{i}$ is the logits of the i-th voxel grid, and $l_{f}^{i}$ the logits value from SDF.", + "bbox": [ + 498, + 752, + 890, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic Scene Reconstruction Our SurroundSDF demonstrates a notable advantage by achieving semantic scene reconstruction concurrently with scene perception. As is shown in Figure 6, we sequentially query the semantic and SDF values for the region of interest frame by frame. The SDF values and semantic logits of static objects are projected onto the sampling points in the world coordinate", + "bbox": [ + 496, + 794, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21618", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ff7020888be29ac63c2182a9e0baf55fbb9d5ea9263c0a3eb359d672b71df0fd.jpg", + "image_caption": [ + "Figure 6. The process of semantic scene reconstruction. SDF values and semantic logits are sampled long the vehicle trajectory into the world coordinate system. For regions where the historical sampling overlaps with the current frame, the historical SDF and semantic logits are updated based on a momentum update strategy." + ], + "image_footnote": [], + "bbox": [ + 117, + 89, + 428, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "system. For the overlapping of the historical sampling area and the current frame sampling area, we employ a momentum update strategy to refine the historical SDF values and semantic logits. Figure 7 shows the final scene reconstruction result with the marching cubes algorithm.", + "bbox": [ + 75, + 310, + 468, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 392, + 209, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 417, + 266, + 434 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/637f3f1f92ad32a744d67d1b8f09af65ed52fd05495c1b1445add806500e7bb2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VersionBackboneImage sizeChannelsFrames
SmallResNet50 [8]256 × 704329
MediumResNet101 [8]640 × 16001286
LargeConvNext-B [23]640 × 16002563
", + "bbox": [ + 102, + 446, + 439, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details We follow BEVStereo [15] to construct our 3D encoder, where the depth net predicts 88 discrete depth values, the voxel feature resolution is $200 \\times 200 \\times 16$ , and the temporal fusion strategy is applied. We pre-train our architecture on semantic segmentation, depth estimation, and 3D object detection tasks on the NuScenes training set. To compare with methods under different settings, we train three versions: Small, Medium, and Large with different backbone, input image size, voxel feature channels, and the number of temporal frames, as is shown in Tabel 1. For the medium and large versions, we reduce the number of temporal frames to save GPU memory. We use the Adam optimizer [13], set the batch size to 32, fix the learning rate to 1e-4, and train for around 32 epochs for each version. The loss weights $\\gamma_{1} \\sim \\gamma_{8}$ are set to 1.0, 1.0, 30.0, 0.05, 0.05, 1.0, 1.0, and 1.0 respectively.", + "bbox": [ + 75, + 551, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset We evaluate our method on Occ3D-nuScenes dataset [35, 36], which is an occupancy dataset with dense voxel grid label based on the nuScenes dataset [2]. It processes dynamic and static objects separately, uses multiframe aggregation to achieve point cloud densification, and then reconstructs the dense point cloud to obtain a 3D occupancy label. It has voxel-level semantic labels of 17 classes,", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56dbb50ed206efffd9f081e546fca7cae165dd5556febb7542d9c206561341aa.jpg", + "image_caption": [ + "a. Scene reconstruction results from a global perspective" + ], + "image_footnote": [], + "bbox": [ + 514, + 89, + 875, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/50ccaa20b9b2f2cd148b6e03ef81513ada1d4cd6b7d48c5ba50ac1be4aadf651.jpg", + "image_caption": [ + "b. Scene reconstruction results from a local perspective", + "Figure 7. Visualization of semantic scene reconstruction." + ], + "image_footnote": [], + "bbox": [ + 514, + 172, + 877, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which contain 16 common classes and an additional general object class. Each frame covers a range of $[-40\\mathrm{m}, -40\\mathrm{m}, -1\\mathrm{m}, 40\\mathrm{m}, 40\\mathrm{m}, 5.4\\mathrm{m}]$ with a voxel size of $0.4\\mathrm{m}$ , resulting in a voxel grid resolution with $200 \\times 200 \\times 16$ .", + "bbox": [ + 498, + 273, + 890, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics To evaluate our performance on dense 3D scene perception, we follow the 3D semantic occupancy prediction task [36] and use the widely adopted mean intersection over union (mIoU).", + "bbox": [ + 498, + 334, + 890, + 393 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nm I o U = \\frac {1}{C} \\sum_ {i = 1} ^ {C} \\frac {T P _ {i}}{T P _ {i} + F P _ {i} + F N _ {i}}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 404, + 890, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $C$ represents the number of categories (set to 17 in the experiment), and TP, FP, and FN represent true positive, false positive, and false negative respectively.", + "bbox": [ + 498, + 446, + 890, + 492 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/06cfbbc1905e340648e344d693ba326c7ea5536c102477da6252b68526468608.jpg", + "table_caption": [ + "Table 1. Experimental settings of different versions." + ], + "table_footnote": [], + "table_body": "
MetricsFormula
Abs Rel1/n∑|d−d*|/d*
Sq Rel1/n∑|d−d*|2/d*
RMSE√1/n∑|d−d*|2
δ<1.251/n∑(max(d/d*, d*/d) < 1.25)
", + "bbox": [ + 513, + 503, + 870, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Depth metrics for 3D scene reconstruction", + "bbox": [ + 542, + 601, + 846, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate our scene reconstruction performance, we render the 2D depth image for the surround cameras based on the reconstructed scene and validate using 2D depth evaluation. As is shown in Table 2, we use the metrics commonly used for depth estimation, which aim to measure the difference between the predicted depth obtained by mesh rendering and GT depth obtained by LiDAR points.", + "bbox": [ + 496, + 627, + 890, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.3D Semantic Occupancy Prediction", + "text_level": 1, + "bbox": [ + 498, + 741, + 807, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the widely adopted experimental setup [19, 36], we conduct the semantic occupancy prediction experiment on the Occ3D-nuScenes dataset. For a fair comparison, in the inference phase, we divide the space into voxel grids with the same size as the other methods (0.4m) and predict the occupancy semantics following Section 4.3. As is shown in Tabel 3, the mIoU score of our Small version using ResNet50 as the backbone and without any test-time-augmentation (TTA) achieves 42.37, surpassing all the", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21619", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "methods with the same image backbone. Additionally, our Medium version significantly outperforms BEVFormer [18] and VoxFormer [17] using the ResNet101 backbone. Moreover, our Large version surpasses UniOCC [29] with the same ConvNext-B backbone by a 1.5 mIoU score, and even outperforms FB-OCC [19, 20], which utilizes a larger image backbone and TTA strategy. These results demonstrated the effectiveness of SurroundSdf in the perception task.", + "bbox": [ + 75, + 90, + 470, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. 3D Scene Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 220, + 305, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the performance of scene reconstruction, we compare our method with state-of-the-art (SOTA) surround-view-based depth estimation methods [33, 40] and occupancy prediction methods [19, 41]. To acquire the surround depth, we utilize the semantic mesh generation steps described in Section 4.3 to generate the mesh and then render the depth maps for each camera based on the camera parameters. To fairly compare with the occupancy-based methods, we follow a similar process to render the depth maps with predicted occupied voxel grids, and all methods use RestNet50 as the image backbone. As is shown in Table 4, our method exhibits significant advantages over occupancy-based approaches across all metrics. Moreover, on the primary metric of absolute relative error, we achieve SOTA results, surpassing SurroundDepth [40] and R3D3 [33], which are the SOTA methods for surround-view depth estimation.", + "bbox": [ + 75, + 239, + 470, + 482 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/19e3f6ca86262c82f6dd7d641c24d9fa954992ef1fb5530f55a3202a36722c49.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsBackboneImage SizemIoU
SurroundOcc [41]ResNet50256 × 70436.32
BEVStereo [15]ResNet50384 × 70439.1
UniOCC [29]ResNet50256 × 70439.7
MiLO [37]ResNet50256 × 70440.49
FB-OCC [19]ResNet50256 × 70440.69
FB-OCC* [19]ResNet50256 × 70442.06
OursResNet50256 × 70442.37
BEVFormer [18]ResNet101900 × 160040.6
VoxFormer [17]ResNet101900 × 160040.6
OursResNet101640 × 160046.0
SurroundOcc [41]InterImage-B640 × 160040.7
BEVFormer [18]InterImage-XL640 × 160043.3
BEVDet [10]Swin-B640 × 160043.1
UniOCC [29]ConvNext-B640 × 160051.5
FB-OCC* [19]InternImage-H960 × 176952.79
OursConvNext-B640 × 160053.01
", + "bbox": [ + 89, + 492, + 455, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 847, + 230, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we study the proposed strategies and the impact of GT quality.", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/287564f119d4700e4adddaf24051e67a874d81df431f47d8496affc2e58575e4.jpg", + "table_caption": [ + "Table 3. Comparison on the Occ3D-nuScenes val set. The superscript * denotes using test time augmentation." + ], + "table_footnote": [], + "table_body": "
MethodsAbs Rel ↓Sq Rel ↓RMSE ↓δ < 1.25 ↑
SurroundOcc0.2742.0725.3270.482
FB-OCC0.3422.0475.9700.290
SurroundDepth0.2242.1024.5730.751
R3D30.2592.3285.5770.583
Ours0.1741.0974.4020.747
", + "bbox": [ + 509, + 89, + 879, + 180 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. 3D scene reconstruction performance. AbsRel is the main metric. “↓” means less is better.", + "bbox": [ + 500, + 185, + 890, + 213 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study on Proposed Strategies We conduct ablation experiments based on the settings of the Small version in Tabel 1 and the results are shown in Tabel 5. These studies aim to analyze the effectiveness of different 3D representations and supervisions. Specifically, we first construct a baseline model (model A) that is supervised by the occupancy GT with the cross-entropy loss and output the occupancy results. Then, to investigate the 3D perception capability of implicit SDF and semantic field, we build the subsequent models based on the architecture in Section 4.1 and output the SDF and semantic field. To analyze the proposed weak supervision paradigm with previous methods, we reproduce the supervision methods of SIREN (model B) and LODE (model C). In model D, we employ our Sandwich Eikonal for SDF supervision. Lastly, model E incorporates the joint supervision strategy in section 4.2.", + "bbox": [ + 496, + 220, + 890, + 462 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparing the results of Model B, and C with A, we observe a significant advantage in the AbsRel metric for both models B and C. This highlights the robust reconstruction capability of SDF-based 3D representation. However, we also observe a significant decline in the semantic mIoU metric for model B, attributed to the sparse supervision with Li-DAR points. LODE mitigates this issue by introducing occupancy supervision, but its mIoU score is still lower than the baseline. Using the proposed Sandwich Eikonal supervision paradigm, the results of model D not only surpass the baseline in semantic mIoU but also outperform models B and C in AbsRel. This indicates the effectiveness of our supervision approach, which successfully integrates the supervision from LiDAR points and occupancy GT. Furthermore, by incorporating of the proposed joint supervision strategy, model E exhibits a 1.85 improvement in semantic mIoU, accompanied by a slight regression of 0.003 in the AbsRel metric. For some long-tail categories, such as bicycle, motorcycle, and traffic cone, our joint supervision strategy yields a noticeable improvement.", + "bbox": [ + 496, + 462, + 892, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study on Ground Truth Quality To further validate the roles of OCC GT and LiDAR GT, we conducted ablation experiments on the resolution of OCC GT and the density of LiDAR points respectively. As is shown in Table 6, the impact of OCC GT resolution variations in our method's OCC GT is mitigated with LiDAR points supervision. When the voxel grid resolution increased from $0.4\\mathrm{m}$ to $1.6\\mathrm{m}$ , the mIoU only decreased by $5.2\\%$ , and Abs-Rel increased by a mere $4.0\\%$ . In comparison, the SOTA OCC", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21620", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d937f8ab06a643e435120747878b7976f66175e08da1158a961d739aa5df78e2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model3D RepresentationSDF SupervisionJoint Supothersbarrierbicyclebuscarvehiclemotorcyclepedestriantraffic conetrailertruckdrivable surfaceother flatsidewalkterrainmannadevegetationmIoUAbs-Rel
AOCC——8.247.07.643.651.425.618.421.722.432.238.783.643.055.357.750.644.638.330.342
BSDFSIREN—8.041.418.625.941.019.519.015.623.019.827.872.338.648.746.536.633.631.510.188
CSDFLODE—8.746.018.638.547.225.322.220.426.323.734.581.244.854.053.840.337.836.660.192
DSDFSandwich—12.347.925.443.051.727.424.924.927.434.039.483.642.954.458.148.343.240.520.171
ESDFSandwich✓13.949.727.844.653.030.029.028.331.135.841.283.644.655.358.949.643.842.370.174
", + "bbox": [ + 107, + 89, + 857, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c19d81ea6bb47b78f02437a3a36b7cbb8822515cb969eb58df349243c9de11ec.jpg", + "image_caption": [ + "Figure 8. Comparison of the 3D scene perspective results with SOTA methods." + ], + "image_footnote": [], + "bbox": [ + 142, + 299, + 831, + 547 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1c07c03c79400ffaaeff061a0f53b70d9a9e03e6b1592f8e3a1b839dc23c1753.jpg", + "table_caption": [ + "Table 5. Results of the ablation experiments. \"3D Representation\" indicates whether the scene is described by occupancy voxels or SDF. \"SDF Supervision\" indicates the supervised paradigm. \"Joint Sup\" indicates whether the joint supervision strategy in Section 4.2 is used." + ], + "table_footnote": [], + "table_body": "
OCC GT size0.4m0.8m1.6m
TargetmIoUAbs-RelmIoUAbs-RelmIoUAbs-Rel
FBOCC40.690.34236.010.37026.960.445
SurroundSDF42.370.17440.740.17640.160.181
", + "bbox": [ + 88, + 579, + 457, + 635 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4a024606e5fc6cdce54255481b241fe3179a3c45b44e69e02f7c5f8cd1e1dcea.jpg", + "table_caption": [ + "Table 6. Results of the experiments on the resolution of OCC GT." + ], + "table_footnote": [], + "table_body": "
Sampling rate10.50.20.1
mIoU42.3742.3742.1641.07
Abs-Rel0.1740.1700.1860.186
", + "bbox": [ + 96, + 662, + 444, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation study on the density of LiDAR points GT.", + "bbox": [ + 94, + 715, + 450, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "prediction method FBOCC experienced a substantial mIoU decrease of $33.7\\%$ , coupled with a $30.1\\%$ increase in Abs-Rel. As shown in Table 7, utilizing only $10\\%$ of the point cloud resulted in a marginal decrease of $3.1\\%$ in mIoU and a modest increase of $6.9\\%$ in Abs-Rel. These experiments further substantiate the robustness and efficacy of the proposed Sandwich Eikonal supervision paradigm.", + "bbox": [ + 75, + 739, + 468, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Comparison of Visual Results", + "text_level": 1, + "bbox": [ + 76, + 854, + 341, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We visualize our 3D scene perception results and compare them with the reproduced SIREN [34] and LODE [14]", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(models B and C in Section 5.4), and the SOTA occupancy prediction method FBOCC [19, 20], as illustrated in Figure 8. Our SurroundSDF demonstrates significant advantages in the continuity and accuracy of the geometric structure.", + "bbox": [ + 498, + 580, + 890, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 652, + 619, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose SurroundSDF, a novel vision-centric 3D scene understanding framework. We introduce SDF to address the continuity and accuracy of perception from surround cameras. Moreover, in the absence of SDF GT, we propose Sandwich Eikonal formulation, a novel SDF supervision paradigm, which emphasizes imposing appropriate constraints on both sides of the surface to enhance geometric accuracy. Furthermore, to alleviate the inconsistency between geometric optimization objectives and semantic optimization objectives, we introduce a joint supervision strategy. Based on the generated SDF and semantic field, different 3D representations can be obtained, including the scene mesh, occupancy voxel grids, and semantic reconstruction mesh of the whole scene. Comprehensive experiments validate the performance of our method.", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21621", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ben Agro, Quinlan Sykora, Sergio Casas, and Raquel Urtasun. Implicit occupancy flow fields for perception and prediction in self-driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1379-1388, 2023. 2", + "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6", + "[3] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1, 2", + "[4] Anh-Quan Cao and Raoul de Charette. Scenerf: Self-supervised monocular 3d scene reconstruction with radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9387–9398, 2023. 2", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2", + "[6] Vitor Guizilini, Rui Hou, Jie Li, Rares Ambrus, and Adrien Gaidon. Semantically-guided representation learning for self-supervised monocular depth. In International Conference on Learning Representations, 2019. 1", + "[7] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. 2023. 2", + "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2", + "[10] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1, 2, 7", + "[11] Peixiang Huang, Li Liu, Renrui Zhang, Song Zhang, Xinli Xu, Baichao Wang, and Guoyi Liu. Tig-bev: Multi-view bev 3d object detection via target inner-geometry learning. 2022. 1, 2", + "[12] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9223-9232, 2023. 1, 2", + "[13] Diederick P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015. 6" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Pengfei Li, Ruowen Zhao, Yongliang Shi, Hao Zhao, Jirui Yuan, Guyue Zhou, and Ya-Qin Zhang. Lode: Locally conditioned eikonal implicit scene completion from sparse lidar. In IEEE International Conference on Robotics and Automation, ICRA 2023, London, UK, May 29 - June 2, 2023, pages 8269-8276. IEEE, 2023. 2, 3, 8", + "[15] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1486-1494, 2023. 2, 6, 7", + "[16] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 1, 2", + "[17] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023. 1, 2, 7", + "[18] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1, 2, 7", + "[19] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 6, 7, 8", + "[20] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M Alvarez. Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6919-6928, 2023. 2, 7, 8", + "[21] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, pages 531-548. Springer, 2022. 1, 2", + "[22] Yingfei Liu, Junjie Yan, Fan Jia, Shuailin Li, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petrv2: A unified framework for 3d perception from multi-camera images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3262-3272, 2023. 2", + "[23] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 6", + "[24] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781. IEEE, 2023. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21622", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. 1, 2", + "[26] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[27] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3D Vision (3DV), 2016 Fourth International Conference on, pages 565-571. IEEE, 2016. 5", + "[28] Timothy S Newman and Hong Yi. A survey of the marching cubes algorithm. Computers & Graphics, 30(5):854-879, 2006. 5", + "[29] Mingjie Pan, Li Liu, Jiaming Liu, Peixiang Huang, Longlong Wang, Shaoqing Xu, Zhiyi Lai, Shanghang Zhang, Kuiyuan Yang, and Xiaomi Car. Uniocc: Unifying vision-centric 3d occupancy prediction with geometric and semantic rendering. 7", + "[30] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2", + "[31] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 4", + "[32] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), 2020. 2", + "[33] Aron Schmied, Tobias Fischer, Martin Danelljan, Marc Pollefeys, and Fisher Yu. R3d3: Dense 3d reconstruction of dynamic scenes from multiple cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3216-3226, 2023. 1, 7", + "[34] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33:7462-7473, 2020. 1, 2, 3, 8", + "[35] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 1, 2, 3, 5, 6", + "[36] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8406-8415, 2023. 3, 4, 5, 6", + "[37] Thang Vu, Jung-Hee Kim, Myeongjin Kim, Seokwoo Jung, and Seong-Gyun Jeong. Milo: Multi-task learning with localization ambiguity suppression for occupancy prediction cvpr 2023 occupancy challenge report. 7" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 34:27171-27183, 2021. 2", + "[39] Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, , and Justin M. Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In The Conference on Robot Learning (CoRL), 2021. 2", + "[40] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Yongming Rao, Guan Huang, Jiwen Lu, and Jie Zhou. Surround-depth: Entangling surrounding views for self-supervised multi-camera depth estimation. In Conference on Robot Learning, pages 539-549. PMLR, 2023. 7", + "[41] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 1, 2, 7", + "[42] Enze Xie, Zhiding Yu, Daquan Zhou, Jonah Philion, Anima Anandkumar, Sanja Fidler, Ping Luo, and Jose M Alvarez. $\\mathbf{M}^2$ bev: Multi-camera joint 3d detection and segmentation with unified birds-eye view representation. arXiv preprint arXiv:2204.05088, 2022. 2", + "[43] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. In The Eleventh International Conference on Learning Representations, 2022. 2", + "[44] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2", + "[45] Weihao Yuan, Xiaodong Gu, Heng Li, Zilong Dong, and Siyu Zhu. Monocular scene reconstruction with 3d sdf transformers. In The Eleventh International Conference on Learning Representations, 2022. 2", + "[46] Yunpeng Zhang, Phigent Robotics, and Zheng Zhu. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9433-9443, 2023. 1, 2" + ], + "bbox": [ + 501, + 90, + 890, + 654 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21623", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/213de3e8-6a97-4d16-96b2-ac03f2f0441a_model.json b/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/213de3e8-6a97-4d16-96b2-ac03f2f0441a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8a7621eeae1ad9e287b86d0e003e3767879f1e64 --- /dev/null +++ b/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/213de3e8-6a97-4d16-96b2-ac03f2f0441a_model.json @@ -0,0 +1,2552 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.131, + 0.892, + 0.155 + ], + "angle": 0, + "content": "SurroundSDF: Implicit 3D Scene Understanding Based on Signed Distance Field" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.179, + 0.831, + 0.24 + ], + "angle": 0, + "content": "Lizhe Liu\\(^{1*}\\) Bohua Wang\\(^{2*§}\\) Hongwei Xie\\(^{1*}\\) Daqi Liu\\(^{1}\\) Li Liu\\(^{1}\\) Zhiqiang Tian\\(^{2}\\) Kuiyuan Yang\\(^{1}\\) Bing Wang\\(^{1\\dagger}\\) \n\\(^{1}\\)Xiaomi EV \\(^{2}\\)School of Software Engineering, Xi'an Jiaotong University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.603 + ], + "angle": 0, + "content": "Vision-centric 3D environment understanding is both vital and challenging for autonomous driving systems. Recently, object-free methods have attracted considerable attention. Such methods perceive the world by predicting the semantics of discrete voxel grids but fail to construct continuous and accurate obstacle surfaces. To this end, in this paper, we propose SurroundSDF to implicitly predict the signed distance field (SDF) and semantic field for the continuous perception from surround images. Specifically, we introduce a query-based approach and utilize SDF constrained by the Eikonal formulation to accurately describe the surfaces of obstacles. Furthermore, considering the absence of precise SDF ground truth, we propose a novel weakly supervised paradigm for SDF, referred to as the Sandwich Eikonal formulation, which emphasizes applying correct and dense constraints on both sides of the surface, thereby enhancing the perceptual accuracy of the surface. Experiments suggest that our method achieves SOTA for both occupancy prediction and 3D scene reconstruction tasks on the nuScenes dataset." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.632, + 0.21, + 0.646 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.793 + ], + "angle": 0, + "content": "With the recent advancement of 3D object detection algorithms [10, 11, 16, 18, 21, 24], vision-centric autonomous driving system become more practicable. Nevertheless, the persisting challenges related to the long-tail problem and the coarse depiction of the 3D scene underscore its insufficiency. Consequently, a deeper comprehension of 3D geometry and semantics is needed for safety and reliability. This paper delves into a novel vision-centric paradigm of dense and continuous 3D scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Current approaches of dense 3D prediction can be classified into two categories, 3D reconstruction and 3D perception. Specifically, 3D reconstruction algorithms [6, 33] generate dense point clouds enriched with semantic informa" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.271, + 0.677, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.361, + 0.633, + 0.372 + ], + "angle": 0, + "content": "a. SIREN" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.269, + 0.862, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.362, + 0.822, + 0.372 + ], + "angle": 0, + "content": "b. LODE" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.376, + 0.687, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.464, + 0.626, + 0.474 + ], + "angle": 0, + "content": "c. Ours" + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.374, + 0.869, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.463, + 0.856, + 0.474 + ], + "angle": 0, + "content": "d. Ours with semantic" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.48, + 0.892, + 0.534 + ], + "angle": 0, + "content": "Figure 1. The scene perspective results from surround images input. a, b. The result of SIREN and LODE supervision with camera-only input (the original methods are point cloud-based). c. Our result. d. Our result with semantics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.613 + ], + "angle": 0, + "content": "tion by projecting depth maps and 2D semantic maps into 3D space. 3D perception methods [3, 12, 17, 25, 35, 41, 46] predict the occupied status and the semantic of 3D voxel grids. However, either method proves to be redundant or can only predict coarse-grained discrete grids." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.765 + ], + "angle": 0, + "content": "Instead, we develop a vision-centric framework to describe 3D scenes leveraging neural implicit Signed Distance Function (SDF) representation, which we refer to as SurroundSDF. Generally, we aim to: (1) Continuously describe the 3D scene by reconstructing smooth surfaces; (2) Explore the difficulty associated with the utilization of SDF representation and pose an appropriate training strategy; (3) Exploit the strong representation to kill two birds with one stone, addressing 3D semantic segmentation and continuous 3D geometry reconstruction within one framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In pursuit of the first and the third objectives, we employ SDF to represent the canonical 3D scene in terms of distance from the surface. Furthermore, we construct an implicit field that encompasses both semantics and geometry by exploiting this strong representation. However, it's still nontrivial to accomplish the SDF modeling, where complete surface and normal vectors, which are crucial supervisions, cannot be accurately computed. To this end, SIREN [34] proposed a weakly supervised implicit per" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.863, + 0.21, + 0.875 + ], + "angle": 0, + "content": "* Equal Contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.876, + 0.348, + 0.888 + ], + "angle": 0, + "content": "\\(^{\\S}\\)Work done during an internship at Xiaomi EV." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.226, + 0.9 + ], + "angle": 0, + "content": "† Corresponding Author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.863, + 0.348, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21614" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "ception method, which accomplishes 3D mesh completion from point clouds. While LODE [14] proposed to utilize occupancy ground truth (GT) to supervise SDF. However, these methods rely on the input of point clouds, and only alleviate this difficulty by mimicking the surface distance but are either sparse or inaccurate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.471, + 0.348 + ], + "angle": 0, + "content": "To overcome the aforementioned challenges, we propose the Sandwich Eikonal formulation, a novel weak supervision paradigm for SDF modeling. Figure 1 gives some reconstruction results and shows the benefit of our method. This method emphasizes applying correct and dense constraints on both sides of the surface to enhance the geometric accuracy and continuity of the surface. Moreover, we revisit current pipelines for training the perception branch and design a novel loss that enhances the integration of geometry and semantics thereby reducing inconsistencies. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.35, + 0.468, + 0.41 + ], + "angle": 0, + "content": "- We propose a vision-centric implicit semantic SDF perception method, achieving accurate and continuous 3D perception. To the best of our knowledge, we are the first to utilize SDF for surround-view 3D perception." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.411, + 0.468, + 0.44 + ], + "angle": 0, + "content": "- We introduce the Sandwich Eikonal formulation, a novel weak supervision paradigm for SDF." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.442, + 0.468, + 0.484 + ], + "angle": 0, + "content": "- We demonstrate how to employ this representation for the reconstruction of continuous 3D geometry with precise semantic information." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.486, + 0.468, + 0.547 + ], + "angle": 0, + "content": "- We achieve state-of-the-art results on 3D dense semantics perception tasks. Comprehensive experiments on the NuScenes dataset [2] provide extensive validation of our approach." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.35, + 0.468, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.559, + 0.218, + 0.573 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.468, + 0.719 + ], + "angle": 0, + "content": "Occupancy Prediction Recently, researches [3, 12, 17, 25, 32, 32, 35, 41, 46] on occupancy prediction have demonstrated advantages in 3D scene understanding. Compared to the traditional object detection paradigm [5, 10, 11, 15, 16, 18, 21, 22, 24, 39, 42], occupancy perception has the following advantages. First, it can express dense 3D geometry. Second, it can accurately provide spatial locations for objects beyond predefined categories. Third, it can describe the shapes of irregular obstacles." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Based on these advantages, occupancy prediction tasks have attracted significant research. These methods predict the occupancy status in the region of interest around the ego vehicle from point clouds or images. Specifically, the space is first divided into voxel grids at a specific resolution. Then the occupancy status and semantics of each grid are estimated. SurroundOcc [41] proposed a surround-view 3D occupancy perception method that utilizes spatial 2D-3D attention to lift image features into 3D space. In addition, to realize the dense occupancy prediction, SurroundOcc designed a pipeline to convert the point cloud to dense occupancy ground truth. VoxFormer [17] employed" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "an MAE-like [9] approach to achieve camera-based semantic occupancy prediction. FB-OCC [20, 20] proposed a novel forward-backward projection method to compensate for the insufficient BEV feature density of the forward projection method and the large number of mismatches in 2D and 3D space caused by the backward projection. Despite the impressive results, they are limited by the specific resolution of the occupancy annotations, which limit the continuous perception of the scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.231, + 0.892, + 0.639 + ], + "angle": 0, + "content": "Implicit Scene Perception Scene reconstruction refers to the task of predicting the 3D geometry structures from some incomplete representations, e.g., images, point clouds, and voxel grids. Implicit neural scene reconstruction methods [1, 4, 14, 26, 30, 34, 45] have demonstrated advantages in accuracy and substantial potential. Generally, they train a neural network to predict a continuous field. Thus, it is possible to query occupancy information for any point within the 3D space. DeepSDF [30] utilizes the SDF to achieve the implicit reconstruction of 3D objects at the instance level. NeRFs [26] and its numerous variants [4, 7, 43] achieve implicit 3D scene reconstruction and novel view synthesis from multi-view images. NeuS [38] and VolSDF [44] introduce the signed distance function into neural rendering, enabling more precise object surface reconstruction. SIREN [34] utilizes the Eikonal equation to achieve semi-supervised SDF reconstruction for point clouds. However, these methods lack generalization to novel scenes. To achieve online implicit scene perception in driving scenes, LODE [14] proposes the Locally Conditioned Eikonal formulation and introduces a dense occupancy ground truth for supervision, which significantly improves the effect of SDF reconstruction in the driving scenes. However, it is limited by the sparsity of input LiDAR points or inaccurate ground truth. To solve these problems, we propose a camera-only implicit scene understanding method, which aims at continuous and accurate surface perception." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.646, + 0.63, + 0.661 + ], + "angle": 0, + "content": "3. Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.668, + 0.892, + 0.714 + ], + "angle": 0, + "content": "This section analyzes the SDF supervision paradigm based on the Eikonal formulation with its variations and introduces our Sandwich Eikonal supervision approach." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.718, + 0.779, + 0.732 + ], + "angle": 0, + "content": "3.1. Eikonal-based SDF Constraints" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.737, + 0.892, + 0.782 + ], + "angle": 0, + "content": "Given a coordinate \\(\\mathbf{x}\\) in the 3D scene, our goal is to construct a function \\(\\phi\\) such that \\(\\phi(\\mathbf{x})\\) provides the SDF value at point \\(\\mathbf{x}\\). The Eikonal-based optimization objectives [14] is" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.784, + 0.891, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathcal {O} _ {E} = \\int_ {\\Omega_ {0}} O _ {0} d \\mathbf {x} + \\int_ {\\Omega_ {1}} O _ {1} d \\mathbf {x} + \\int_ {\\Omega_ {2}} O _ {2} d \\mathbf {x}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.807, + 0.529, + 0.819 + ], + "angle": 0, + "content": "and" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.818, + 0.891, + 0.869 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} O _ {0} = | | \\nabla_ {\\mathbf {x}} \\phi (\\mathbf {x}) | - 1 |, & \\mathbf {x} \\in \\Omega_ {0} \\\\ O _ {1} = | \\nabla_ {\\mathbf {x}} \\phi (\\mathbf {x}) - \\mathbf {n} (\\mathbf {x}) |, & \\mathbf {x} \\in \\Omega_ {1}, \\\\ O _ {2} = | \\phi (\\mathbf {x}) - S D F (\\mathbf {x}) |, & \\mathbf {x} \\in \\Omega_ {2} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "where \\(\\Omega_0\\) is the set of the whole 3D space of interest, \\(\\Omega_1\\) is the set of points on the surface, \\(\\Omega_2\\) is the set of points" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21615" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.162, + 0.095, + 0.295, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.202, + 0.277, + 0.213 + ], + "angle": 0, + "content": "a. Ideal supervision" + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.096, + 0.429, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.202, + 0.386, + 0.212 + ], + "angle": 0, + "content": "b. SIREN" + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.096, + 0.565, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.478, + 0.202, + 0.519, + 0.212 + ], + "angle": 0, + "content": "c.LODE" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.096, + 0.701, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.202, + 0.655, + 0.212 + ], + "angle": 0, + "content": "d. Ours" + }, + { + "type": "image_footnote", + "bbox": [ + 0.704, + 0.099, + 0.81, + 0.109 + ], + "angle": 0, + "content": "Constrain \\(SDF < 0\\)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.705, + 0.117, + 0.807, + 0.126 + ], + "angle": 0, + "content": "Constrain \\(SDF = 0\\)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.705, + 0.133, + 0.807, + 0.142 + ], + "angle": 0, + "content": "Constrain \\(SDF > 0\\)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.705, + 0.148, + 0.81, + 0.156 + ], + "angle": 0, + "content": "Constraint min SDF" + }, + { + "type": "image_footnote", + "bbox": [ + 0.705, + 0.156, + 0.792, + 0.165 + ], + "angle": 0, + "content": "of one grid \\(< 0\\)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.705, + 0.168, + 0.765, + 0.177 + ], + "angle": 0, + "content": "Surface" + }, + { + "type": "image_footnote", + "bbox": [ + 0.705, + 0.183, + 0.791, + 0.192 + ], + "angle": 0, + "content": "LiDAR Points" + }, + { + "type": "image_footnote", + "bbox": [ + 0.705, + 0.195, + 0.776, + 0.206 + ], + "angle": 0, + "content": "Voxel Grid" + }, + { + "type": "list", + "bbox": [ + 0.704, + 0.099, + 0.81, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.217, + 0.894, + 0.259 + ], + "angle": 0, + "content": "Figure 2. a. Ideal supervision for SDF where the SDF GT is provided. b. SIREN supervision which uses LiDAR points to supervise the surface. c. LODE supervision based on occupancy GT. d. Our supervision paradigm which combines the LiDAR GT and occupancy GT, it is closer to the ideal supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.268, + 0.469, + 0.373 + ], + "angle": 0, + "content": "with SDF GT annotations, \\(\\nabla_{\\mathbf{x}}\\phi (\\mathbf{x})\\) is the gradient on point \\(\\mathbf{x}\\), \\(\\mathbf{n}(\\mathbf{x})\\) is the normal, and \\(SDF(\\mathbf{x})\\) is the real SDF value at point \\(\\mathbf{x}\\). One major obstacle to this optimization objective lies in the design of \\(O_2\\), which provides direct supervision for SDF as shown in Figure 2.a. This objective requires the annotation of SDF values for each anchor point in the scene, which is monumental and challenging." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.374, + 0.469, + 0.42 + ], + "angle": 0, + "content": "In the absence of precise SDF annotations, SIREN [34] and LODE [14] have implemented weakly supervised paradigm through the following variations of \\(O_2\\)," + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.425, + 0.469, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} O _ {2 - 0} ^ {\\prime} = | \\phi (\\mathbf {x}) |, & \\mathbf {x} \\in \\Omega_ {1} \\\\ O _ {2 - 1} ^ {\\prime} = \\psi (| \\phi (\\mathbf {x}) |), & \\mathbf {x} \\in \\Omega_ {3}, \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.469, + 0.492 + ], + "angle": 0, + "content": "where \\(\\psi (\\cdot)\\) is a monotonically decreasing function and \\(\\Omega_3\\subseteq \\Omega_0 / \\Omega_1\\) is the whole space except the surface." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.469, + 0.581 + ], + "angle": 0, + "content": "The objective \\(O_{2-0}^{\\prime}\\) aims to constrain the SDF values on the surface to 0. Considering the absence of precise supervision, \\(O_{2-1}^{\\prime}\\) adopts a fuzzy constraint that pushes the absolute SDF value away from zero in \\(\\Omega_3\\). While the constraint in \\(\\Omega_3\\) is loose, it is adequate to capture the geometry information of the surface." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.469, + 0.809 + ], + "angle": 0, + "content": "In practical implementation, SIREN [34] samples on-surface LiDAR points in \\(\\Omega_{1}\\) and considers the space excluding the LiDAR points as \\(\\Omega_{3}\\), as is shown in Figure 2.b. When the LiDAR points are dense enough, this strategy can effectively achieve good constraint effects [34]. However, LODE [14] has demonstrated that the sparsity of LiDAR points in the driving scenes results in discontinuities of the estimated mesh. To achieve denser supervision, they use the dense occupancy GT as supervision. This approach treats both the surface and interior of the obstacles as the \\(\\Omega_{1}\\). Specifically, they sample the \\(\\Omega_{1}\\) points as the center of \"occupied\" grids, and the \\(\\Omega_{3}\\) points as the center of \"free\" grids, as is shown in 2.c. However, this paradigm remains imprecise, as considering the center of the occupied grid to be the surface of objects would introduce errors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.353, + 0.832 + ], + "angle": 0, + "content": "3.2. Sandwich Eikonal Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To address the above concerns, we require a supervision paradigm that has the following characteristics: 1) sampling points as the supervision on the surface must strictly adhere to the surface, 2) for regions beyond the surface, refrain" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.268, + 0.892, + 0.313 + ], + "angle": 0, + "content": "from supervision in uncertain areas to ensure precision, and 3) dense supervision signals. Considering the above factors, we introduce our variants of Eikonal formulation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.327, + 0.892, + 0.379 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} O _ {2 - 0} = \\phi (\\mathbf {x}), & \\mathbf {x} \\in \\Omega_ {1} \\\\ O _ {2 - 1} = \\psi (- \\min (\\phi (\\mathbf {x}))), & \\mathbf {x} \\in \\Omega_ {\\text {o c c}} ^ {i} \\\\ O _ {2 - 2} = \\psi (\\phi (\\mathbf {x})), & \\mathbf {x} \\in \\Omega_ {\\text {f r e e}} ^ {j}, \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.392, + 0.892, + 0.424 + ], + "angle": 0, + "content": "where \\(\\Omega_{occ}^{i}\\) is the \\(i\\)-th grid of occupied grids and \\(\\Omega_{free}^{j}\\) is the \\(j\\)-th grid of free grids." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.893, + 0.775 + ], + "angle": 0, + "content": "To this end, we incorporate the LiDAR points and the occupancy GT into the SDF supervision. As is shown in Figure 2.d, for precise sampling points on the surface, we follow SIREN and constrain the SDF value of LiDAR points to be 0. To apply accurate and dense supervision to regions beyond the surface, we introduce the occupancy GT into the supervision, as is shown in optimization objectives \\( O_{2-1} \\) and \\( O_{2-2} \\). According to the generation process of occupancy GT [35, 36], if a grid is labeled as \"occupied\", it implies that at least part of the grid area falls inside the object. Therefore, as the objective \\( O_{2-1} \\) shows, we constrain the minimum SDF value of each occupied grid to be less than zero and push it to the negative range. On the other hand, if a grid is labeled as \"free\", the entire grid is outside the object and we push the SDF value of free grids to the positive range, as is shown in objective \\( O_{2-2} \\). We employ a multiframe point cloud fusion strategy to densify the sampling for surface supervision, as described in Section 4.2. Compared to previous supervision paradigms, our formulation emphasizes applying appropriate and dense supervision on both sides of the surface to enhance the geometric constraint based on the Eikonal formulation. Therefore, we name our supervision strategy the \"Sandwich Eikonal\" formulation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.797, + 0.591, + 0.812 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Given multi-camera images \\(\\mathbf{I} = \\{I_1, I_2, \\dots, I_n\\}\\), we aim to predict the 3D SDF \\(\\phi(\\mathbf{x}) : \\mathbb{R}^3 \\Rightarrow \\mathbb{R}\\) and semantic field \\(S(\\mathbf{x}) : \\mathbb{R}^3 \\Rightarrow \\mathbb{R}^s\\), where \\(s\\) indicates the number of classes. The following will introduce our architecture and detailed constraints based on our Sandwich Eikonal formulation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21616" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.149, + 0.09, + 0.825, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.303, + 0.895, + 0.402 + ], + "angle": 0, + "content": "Figure 3. The architecture of our SurroundSDF. Given the surround images as input, an encoder composed of a 2D backbone, LSS module, and BEV backbone, is employed to extract voxel features. We adopt a query-based approach to sample features from the voxel features. Specifically, first, a set of query coordinates in the region of interest is selected. Subsequently, using trilinear interpolation, semantic features are queried from the voxel features. Finally, after concatenation with the positional embeddings from the query coordinates, the features pass through the SDF head and semantic head respectively, yielding SDF and semantic fields. For training, the query coordinates are sampled according to the GT, and the SDF and semantic field are supervised by the losses introduced in Section 4.2. In the inference phase, based on appropriate sampling and post-processing, continuous and accurate scene perception results are obtained (see Section 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.409, + 0.213, + 0.423 + ], + "angle": 0, + "content": "4.1. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.471, + 0.567 + ], + "angle": 0, + "content": "The overall architecture is shown in Figure 3. First, a CNN-based backbone is used to obtain the image features. Then, the LSS [31] is used to project the image features to BEV space, and a BEV encoder is adopted to obtain the voxel features \\( V \\) of dimension \\( h \\times w \\times d \\times c \\) (c indicates the number of channels). Further, for each point \\( \\mathbf{x} \\) in \\( \\Omega_0 \\), the corresponding features can be obtained by querying \\( V \\) with the trilinear sampling method. Therefore, \\( \\phi(\\mathbf{x}) \\) and \\( S(\\mathbf{x}) \\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.575, + 0.47, + 0.599 + ], + "angle": 0, + "content": "\\[\n\\phi (\\mathbf {x}) = H _ {s d f} \\left(C \\left(Q (V, \\mathbf {x}), P E (\\mathbf {x})\\right), \\theta\\right), \\mathbf {x} \\in \\Omega_ {0}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.594, + 0.432, + 0.611 + ], + "angle": 0, + "content": "\\[\nS (\\mathbf {x}) = H _ {s e m} (C (Q (V, \\mathbf {x}), P E (\\mathbf {x})), \\theta), \\mathbf {x} \\in \\Omega_ {0},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.47, + 0.743 + ], + "angle": 0, + "content": "where \\(\\mathbf{x} \\in \\mathbb{R}^3\\), \\(C\\) and \\(Q\\) represent the concatenation and query process, respectively. \\(H_{sdf}\\) and \\(H_{sem}\\) indicate the SDF head and semantic head, implemented with three-layer MLPs with sine activation functions. \\(\\theta\\) represents the learnable parameters. \\(PE\\) represents the positional encoding, which aims to capture high-frequency information. Specifically, for a coordinate \\(\\mathbf{x} \\in \\mathbb{R}^3\\), \\(PE\\) encodes each of its dimensions into a 2n-dimensional vector, it can be expressed" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.745, + 0.47, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\text {a s :} \\quad P E (\\mathbf {x}) = (\\zeta (x), \\zeta (y), \\zeta (z)) \\in \\mathbb {R} ^ {6 n}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.469, + 0.796 + ], + "angle": 0, + "content": "where \\(x, y, z\\) denote the coordinates, \\(\\zeta(\\cdot)\\) represents an encoding scheme using trigonometric function mapping:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.801, + 0.47, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\zeta (x) = \\left(\\sin \\left(2 ^ {0} \\pi x\\right), \\cos \\left(2 ^ {0} \\pi x\\right), \\dots , \\cos \\left(2 ^ {n - 1} \\pi x\\right)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In training, we design a joint supervision method that utilizes both coarse-grained voxel and fine-grained point cloud supervision based on our Sandwich Eikonal formulation. In inference, we query the voxel features to estimate the SDF values and semantic logits of each query point." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.409, + 0.629, + 0.424 + ], + "angle": 0, + "content": "4.2. Supervision" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.893, + 0.478 + ], + "angle": 0, + "content": "This section will elaborate on deriving the SDF loss from our Sandwich Eikonal formulation. Then, we will introduce the semantic loss and our joint supervision strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.893, + 0.577 + ], + "angle": 0, + "content": "SDF Supervision Following objectives \\(O_0\\), \\(O_1\\) in Equation 2 and \\(O_{2-0}\\) to \\(O_{2-2}\\) in Equation 4, we derive the losses for SDF supervision, based on our Sandwich Eikonal formulation. Considering that continuous space supervision is unfeasible, we sample the discrete points in the corresponding space \\((\\Omega_0, \\Omega_1, \\Omega_{occ}\\) and \\(\\Omega_{free})\\) for supervision." + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.589, + 0.698, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.704, + 0.619, + 0.713 + ], + "angle": 0, + "content": "a." + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.591, + 0.837, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.704, + 0.779, + 0.713 + ], + "angle": 0, + "content": "b." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.717, + 0.893, + 0.745 + ], + "angle": 0, + "content": "Figure 4. a. SDF constraints with Sandwich Eikonal formulation in continuous form. b. SDF sampling in discrete form." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "The correctness and density of the supervision sampling are of paramount importance. Since our approach is based on the surround cameras, we only sample from the camera-visible space in \\(\\Omega_0\\) (space of interest) [36]. For \\(\\Omega_{1}\\) (surface points), we sample exclusively from the LiDAR points. To enhance the sampling density, temporal point cloud fusion is adopted. Specifically, \\(\\Omega_{1} = \\{p_{t - k},p_{t - (k - 1)},\\dots ,p_{t + (k - 1)},p_{t + k}\\}\\), where \\(p_{t + j}\\) is the LiDAR points of the \\(t + j\\) frame and \\(t\\) is the current frame. Based on occupancy GT [36], \\(\\Omega_{occ}\\) and \\(\\Omega_{free}\\) are the sets of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21617" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.462, + 0.107 + ], + "angle": 0, + "content": "voxel grids labeled as \"occupied\" and \"free\", respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.47, + 0.228 + ], + "angle": 0, + "content": "Additionally, in our objective \\( O_{2-1} \\), the computation of the minimum SDF within the occupied grid is required. As is shown in Figure 4, we approximate this minimum value through discrete sampling. The minimum value among the SDF values at the centers of each sub-grid is regarded as the minimum SDF value for the grid. Based on Equation 2, 4 and the above sampling strategy, our SDF loss \\( \\mathcal{L}_{sdf} \\) is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.23, + 0.47, + 0.418 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {s d f} = \\gamma_ {1} \\frac {1}{n _ {t o t a l}} \\sum_ {i = 1} ^ {n _ {t o t a l}} \\left(\\left| \\nabla_ {\\mathbf {x}} \\phi (\\mathbf {x} ^ {i}) \\right| - 1\\right) \\\\ + \\gamma_ {2} \\frac {1}{n _ {1 \\cup o c c}} \\sum_ {i = 1} ^ {n _ {1 \\cup o c c}} (\\nabla_ {\\mathbf {x}} \\phi (\\mathbf {x} ^ {i}) - n (\\mathbf {x} ^ {i})) \\\\ + \\gamma_ {3} \\frac {1}{n _ {1}} \\sum_ {i = 1} ^ {n _ {1}} | \\phi (\\mathbf {x} ^ {i}) | \\tag {8} \\\\ + \\gamma_ {4} \\frac {1}{n _ {o c c}} \\sum_ {i = 1} ^ {n _ {o c c}} e ^ {\\alpha \\times m i n (\\phi (\\mathbf {x} ^ {i}), \\mathbf {x} ^ {i} \\in \\Omega_ {o c c} ^ {i})} \\\\ + \\gamma_ {5} \\frac {1}{n _ {f r e e}} \\sum_ {i = 1} ^ {n _ {f r e e}} e ^ {- \\alpha \\times r a n d o m (\\phi (\\mathbf {x} ^ {i}), \\mathbf {x} ^ {i} \\in \\Omega_ {f r e e} ^ {i})}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.47, + 0.543 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a hyperparameter that determines the extent of the deviation for the interior and exterior space of the object and we fix \\(\\alpha\\) to 100 in experiments. \\(\\gamma_{1},\\gamma_{2},\\gamma_{3},\\gamma_{4},\\gamma_{5}\\) are loss weights, \\(n(\\mathbf{x})\\) represents the normal vector of the current point on surface. \\(n_1,n_{occ}\\), and \\(n_{free}\\) represent the number of points or voxel grids sampled from \\(\\Omega_1,\\Omega_{occ}\\) and \\(\\Omega_{free}\\), respectively. \\(n_{total}\\) is the total number of samples: \\(n_{total} = n_1 + n_{occ} + n_{free}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.47, + 0.588 + ], + "angle": 0, + "content": "Semantic Supervision A widely used cross-entropy loss is adopted to supervise the semantic field as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.592, + 0.469, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s e m} = \\frac {1}{n _ {o c c}} \\sum_ {i = 1} ^ {n _ {o c c}} C E \\left(S (\\mathbf {x} | \\mathbf {x} \\in \\Omega_ {o c c} ^ {i}), \\dot {y} _ {s e m} ^ {i}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.469, + 0.645 + ], + "angle": 0, + "content": "where CE represents cross-entropy loss and \\(\\hat{y}_{sem}^i\\) is the semantic ground truth of i-th voxel grid." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.469, + 0.856 + ], + "angle": 0, + "content": "Joint Supervision The above losses optimize the SDF and semantic field separately, which leads to a nonnegligible issue of ambiguity between geometric and semantic optimization. To illustrate this issue, we train a model and evaluate the occupancy IoU representing geometric accuracy and the semantic mIoU representing overall semantic accuracy following the occupancy prediction metrics [35, 36]. For each voxel grid, if its minimum SDF is less than a certain threshold, the voxel is considered 'occupied'. As is shown in Figure 5, the optimal thresholds producing the highest peak of these curves are significantly different. This indicates that greater geometric accuracy does not necessarily lead to improved overall semantic accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To mitigate this issue, we propose a joint supervision strategy. First, for the minimum SDF value \\( \\min(\\phi(\\mathbf{x})) \\) of each voxel grid, a SoftMax function is used to convert it to" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.09, + 0.861, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.206, + 0.894, + 0.248 + ], + "angle": 0, + "content": "Figure 5. Variation of occupancy IoU and semantic mIoU with SDF Threshold. Note that the peak values of these two indicators correspond to different SDF thresholds." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.26, + 0.893, + 0.352 + ], + "angle": 0, + "content": "the free probability: \\( p_{free} = \\text{softmax}(\\beta t, \\beta \\min(\\phi(\\mathbf{x})) \\), where \\( t \\) is fixed to 0.005, \\( \\beta \\) is fixed to 100. Second, we map the free probability \\( p_{free} \\) to logits using an inv-sigmoid function: \\( l_{free} = \\log(p_{free} / (1 - p_{free})) \\). Finally, \\( l_{free} \\) is combined with semantic logits \\( l_{sem} \\), and the joint loss \\( \\mathcal{L}_{joint} \\) is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.356, + 0.892, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {j o i n t}} = \\operatorname {D i c e} \\left(C \\left(l _ {\\text {s e m}}, l _ {\\text {f r e e}}\\right), C \\left(\\hat {y} _ {\\text {s e m}}, \\hat {y}\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.385, + 0.892, + 0.43 + ], + "angle": 0, + "content": "where \\(C\\) is the concatenation operation and \\(\\hat{y}_{sem}\\) is the semantic GT of voxel grid. Dice represents the Dice loss [27]. Finally, the total loss is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.436, + 0.892, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\gamma_ {6} \\mathcal {L} _ {s d f} + \\gamma_ {7} \\mathcal {L} _ {s e m} + \\gamma_ {8} \\mathcal {L} _ {j o i n t} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.457, + 0.724, + 0.473 + ], + "angle": 0, + "content": "where \\(\\gamma_6, \\gamma_7, \\gamma_8\\) are loss weights." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.482, + 0.61, + 0.495 + ], + "angle": 0, + "content": "4.3. Inference" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.505, + 0.892, + 0.551 + ], + "angle": 0, + "content": "Based on the output SDF and semantic field, diverse outputs for different tasks can be obtained through different sampling strategies and post-processing methods." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.555, + 0.892, + 0.646 + ], + "angle": 0, + "content": "Semantic Mesh Generation To generate the semantic mesh, we uniformly query the SDF value in the space of interest and use the marching cubes algorithm [28] to obtain the mesh. Subsequently, at the vertex positions of each triangular face, we query the corresponding semantic information, resulting in a mesh enriched with semantics." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.658, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Occupancy Prediction The objective of the occupancy prediction is to anticipate the semantics of the voxel grid in space. We achieve this through the joint prediction based on the predicted SDF and semantic field as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.723, + 0.892, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {i} = \\operatorname {a r g m a x} \\left(C \\left(l _ {\\text {s e m}} ^ {i}, l _ {f} ^ {i}\\right)\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.753, + 0.892, + 0.786 + ], + "angle": 0, + "content": "where \\( S_{i} \\) is the i-th voxel grid, \\( l_{sem}^{i} \\) is the logits of the i-th voxel grid, and \\( l_{f}^{i} \\) the logits value from SDF." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Semantic Scene Reconstruction Our SurroundSDF demonstrates a notable advantage by achieving semantic scene reconstruction concurrently with scene perception. As is shown in Figure 6, we sequentially query the semantic and SDF values for the region of interest frame by frame. The SDF values and semantic logits of static objects are projected onto the sampling points in the world coordinate" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21618" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.09, + 0.429, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.235, + 0.47, + 0.306 + ], + "angle": 0, + "content": "Figure 6. The process of semantic scene reconstruction. SDF values and semantic logits are sampled long the vehicle trajectory into the world coordinate system. For regions where the historical sampling overlaps with the current frame, the historical SDF and semantic logits are updated based on a momentum update strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.469, + 0.386 + ], + "angle": 0, + "content": "system. For the overlapping of the historical sampling area and the current frame sampling area, we employ a momentum update strategy to refine the historical SDF values and semantic logits. Figure 7 shows the final scene reconstruction result with the marching cubes algorithm." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.393, + 0.21, + 0.41 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.418, + 0.267, + 0.435 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.447, + 0.44, + 0.521 + ], + "angle": 0, + "content": "
VersionBackboneImage sizeChannelsFrames
SmallResNet50 [8]256 × 704329
MediumResNet101 [8]640 × 16001286
LargeConvNext-B [23]640 × 16002563
" + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.532, + 0.427, + 0.546 + ], + "angle": 0, + "content": "Table 1. Experimental settings of different versions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.469, + 0.794 + ], + "angle": 0, + "content": "Implementation details We follow BEVStereo [15] to construct our 3D encoder, where the depth net predicts 88 discrete depth values, the voxel feature resolution is \\(200 \\times 200 \\times 16\\), and the temporal fusion strategy is applied. We pre-train our architecture on semantic segmentation, depth estimation, and 3D object detection tasks on the NuScenes training set. To compare with methods under different settings, we train three versions: Small, Medium, and Large with different backbone, input image size, voxel feature channels, and the number of temporal frames, as is shown in Tabel 1. For the medium and large versions, we reduce the number of temporal frames to save GPU memory. We use the Adam optimizer [13], set the batch size to 32, fix the learning rate to 1e-4, and train for around 32 epochs for each version. The loss weights \\(\\gamma_{1} \\sim \\gamma_{8}\\) are set to 1.0, 1.0, 30.0, 0.05, 0.05, 1.0, 1.0, and 1.0 respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Dataset We evaluate our method on Occ3D-nuScenes dataset [35, 36], which is an occupancy dataset with dense voxel grid label based on the nuScenes dataset [2]. It processes dynamic and static objects separately, uses multiframe aggregation to achieve point cloud densification, and then reconstructs the dense point cloud to obtain a 3D occupancy label. It has voxel-level semantic labels of 17 classes," + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.09, + 0.877, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.159, + 0.833, + 0.171 + ], + "angle": 0, + "content": "a. Scene reconstruction results from a global perspective" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.173, + 0.879, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.242, + 0.83, + 0.252 + ], + "angle": 0, + "content": "b. Scene reconstruction results from a local perspective" + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.256, + 0.865, + 0.269 + ], + "angle": 0, + "content": "Figure 7. Visualization of semantic scene reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.274, + 0.892, + 0.334 + ], + "angle": 0, + "content": "which contain 16 common classes and an additional general object class. Each frame covers a range of \\([-40\\mathrm{m}, -40\\mathrm{m}, -1\\mathrm{m}, 40\\mathrm{m}, 40\\mathrm{m}, 5.4\\mathrm{m}]\\) with a voxel size of \\(0.4\\mathrm{m}\\), resulting in a voxel grid resolution with \\(200 \\times 200 \\times 16\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.335, + 0.892, + 0.394 + ], + "angle": 0, + "content": "Evaluation Metrics To evaluate our performance on dense 3D scene perception, we follow the 3D semantic occupancy prediction task [36] and use the widely adopted mean intersection over union (mIoU)." + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.405, + 0.892, + 0.443 + ], + "angle": 0, + "content": "\\[\nm I o U = \\frac {1}{C} \\sum_ {i = 1} ^ {C} \\frac {T P _ {i}}{T P _ {i} + F P _ {i} + F N _ {i}}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.448, + 0.892, + 0.493 + ], + "angle": 0, + "content": "where \\( C \\) represents the number of categories (set to 17 in the experiment), and TP, FP, and FN represent true positive, false positive, and false negative respectively." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.505, + 0.872, + 0.599 + ], + "angle": 0, + "content": "
MetricsFormula
Abs Rel1/n∑|d−d*|/d*
Sq Rel1/n∑|d−d*|2/d*
RMSE√1/n∑|d−d*|2
δ<1.251/n∑(max(d/d*, d*/d) < 1.25)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.543, + 0.602, + 0.848, + 0.616 + ], + "angle": 0, + "content": "Table 2. Depth metrics for 3D scene reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.734 + ], + "angle": 0, + "content": "To evaluate our scene reconstruction performance, we render the 2D depth image for the surround cameras based on the reconstructed scene and validate using 2D depth evaluation. As is shown in Table 2, we use the metrics commonly used for depth estimation, which aim to measure the difference between the predicted depth obtained by mesh rendering and GT depth obtained by LiDAR points." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.742, + 0.808, + 0.759 + ], + "angle": 0, + "content": "5.2.3D Semantic Occupancy Prediction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Following the widely adopted experimental setup [19, 36], we conduct the semantic occupancy prediction experiment on the Occ3D-nuScenes dataset. For a fair comparison, in the inference phase, we divide the space into voxel grids with the same size as the other methods (0.4m) and predict the occupancy semantics following Section 4.3. As is shown in Tabel 3, the mIoU score of our Small version using ResNet50 as the backbone and without any test-time-augmentation (TTA) achieves 42.37, surpassing all the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21619" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.213 + ], + "angle": 0, + "content": "methods with the same image backbone. Additionally, our Medium version significantly outperforms BEVFormer [18] and VoxFormer [17] using the ResNet101 backbone. Moreover, our Large version surpasses UniOCC [29] with the same ConvNext-B backbone by a 1.5 mIoU score, and even outperforms FB-OCC [19, 20], which utilizes a larger image backbone and TTA strategy. These results demonstrated the effectiveness of SurroundSdf in the perception task." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.221, + 0.307, + 0.236 + ], + "angle": 0, + "content": "5.3. 3D Scene Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.24, + 0.471, + 0.483 + ], + "angle": 0, + "content": "To evaluate the performance of scene reconstruction, we compare our method with state-of-the-art (SOTA) surround-view-based depth estimation methods [33, 40] and occupancy prediction methods [19, 41]. To acquire the surround depth, we utilize the semantic mesh generation steps described in Section 4.3 to generate the mesh and then render the depth maps for each camera based on the camera parameters. To fairly compare with the occupancy-based methods, we follow a similar process to render the depth maps with predicted occupied voxel grids, and all methods use RestNet50 as the image backbone. As is shown in Table 4, our method exhibits significant advantages over occupancy-based approaches across all metrics. Moreover, on the primary metric of absolute relative error, we achieve SOTA results, surpassing SurroundDepth [40] and R3D3 [33], which are the SOTA methods for surround-view depth estimation." + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.493, + 0.456, + 0.793 + ], + "angle": 0, + "content": "
MethodsBackboneImage SizemIoU
SurroundOcc [41]ResNet50256 × 70436.32
BEVStereo [15]ResNet50384 × 70439.1
UniOCC [29]ResNet50256 × 70439.7
MiLO [37]ResNet50256 × 70440.49
FB-OCC [19]ResNet50256 × 70440.69
FB-OCC* [19]ResNet50256 × 70442.06
OursResNet50256 × 70442.37
BEVFormer [18]ResNet101900 × 160040.6
VoxFormer [17]ResNet101900 × 160040.6
OursResNet101640 × 160046.0
SurroundOcc [41]InterImage-B640 × 160040.7
BEVFormer [18]InterImage-XL640 × 160043.3
BEVDet [10]Swin-B640 × 160043.1
UniOCC [29]ConvNext-B640 × 160051.5
FB-OCC* [19]InternImage-H960 × 176952.79
OursConvNext-B640 × 160053.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.805, + 0.47, + 0.835 + ], + "angle": 0, + "content": "Table 3. Comparison on the Occ3D-nuScenes val set. The superscript * denotes using test time augmentation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.231, + 0.864 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.903 + ], + "angle": 0, + "content": "In this section, we study the proposed strategies and the impact of GT quality." + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.09, + 0.88, + 0.181 + ], + "angle": 0, + "content": "
MethodsAbs Rel ↓Sq Rel ↓RMSE ↓δ < 1.25 ↑
SurroundOcc0.2742.0725.3270.482
FB-OCC0.3422.0475.9700.290
SurroundDepth0.2242.1024.5730.751
R3D30.2592.3285.5770.583
Ours0.1741.0974.4020.747
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.186, + 0.892, + 0.214 + ], + "angle": 0, + "content": "Table 4. 3D scene reconstruction performance. AbsRel is the main metric. “↓” means less is better." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.221, + 0.892, + 0.463 + ], + "angle": 0, + "content": "Ablation Study on Proposed Strategies We conduct ablation experiments based on the settings of the Small version in Tabel 1 and the results are shown in Tabel 5. These studies aim to analyze the effectiveness of different 3D representations and supervisions. Specifically, we first construct a baseline model (model A) that is supervised by the occupancy GT with the cross-entropy loss and output the occupancy results. Then, to investigate the 3D perception capability of implicit SDF and semantic field, we build the subsequent models based on the architecture in Section 4.1 and output the SDF and semantic field. To analyze the proposed weak supervision paradigm with previous methods, we reproduce the supervision methods of SIREN (model B) and LODE (model C). In model D, we employ our Sandwich Eikonal for SDF supervision. Lastly, model E incorporates the joint supervision strategy in section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.893, + 0.765 + ], + "angle": 0, + "content": "Comparing the results of Model B, and C with A, we observe a significant advantage in the AbsRel metric for both models B and C. This highlights the robust reconstruction capability of SDF-based 3D representation. However, we also observe a significant decline in the semantic mIoU metric for model B, attributed to the sparse supervision with Li-DAR points. LODE mitigates this issue by introducing occupancy supervision, but its mIoU score is still lower than the baseline. Using the proposed Sandwich Eikonal supervision paradigm, the results of model D not only surpass the baseline in semantic mIoU but also outperform models B and C in AbsRel. This indicates the effectiveness of our supervision approach, which successfully integrates the supervision from LiDAR points and occupancy GT. Furthermore, by incorporating of the proposed joint supervision strategy, model E exhibits a 1.85 improvement in semantic mIoU, accompanied by a slight regression of 0.003 in the AbsRel metric. For some long-tail categories, such as bicycle, motorcycle, and traffic cone, our joint supervision strategy yields a noticeable improvement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Ablation Study on Ground Truth Quality To further validate the roles of OCC GT and LiDAR GT, we conducted ablation experiments on the resolution of OCC GT and the density of LiDAR points respectively. As is shown in Table 6, the impact of OCC GT resolution variations in our method's OCC GT is mitigated with LiDAR points supervision. When the voxel grid resolution increased from \\(0.4\\mathrm{m}\\) to \\(1.6\\mathrm{m}\\), the mIoU only decreased by \\(5.2\\%\\), and Abs-Rel increased by a mere \\(4.0\\%\\). In comparison, the SOTA OCC" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21620" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.108, + 0.09, + 0.858, + 0.262 + ], + "angle": 0, + "content": "
Model3D RepresentationSDF SupervisionJoint Supothersbarrierbicyclebuscarvehiclemotorcyclepedestriantraffic conetrailertruckdrivable surfaceother flatsidewalkterrainmannadevegetationmIoUAbs-Rel
AOCC——8.247.07.643.651.425.618.421.722.432.238.783.643.055.357.750.644.638.330.342
BSDFSIREN—8.041.418.625.941.019.519.015.623.019.827.872.338.648.746.536.633.631.510.188
CSDFLODE—8.746.018.638.547.225.322.220.426.323.734.581.244.854.053.840.337.836.660.192
DSDFSandwich—12.347.925.443.051.727.424.924.927.434.039.483.642.954.458.148.343.240.520.171
ESDFSandwich✓13.949.727.844.653.030.029.028.331.135.841.283.644.655.358.949.643.842.370.174
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.264, + 0.892, + 0.293 + ], + "angle": 0, + "content": "Table 5. Results of the ablation experiments. \"3D Representation\" indicates whether the scene is described by occupancy voxels or SDF. \"SDF Supervision\" indicates the supervised paradigm. \"Joint Sup\" indicates whether the joint supervision strategy in Section 4.2 is used." + }, + { + "type": "image", + "bbox": [ + 0.143, + 0.3, + 0.832, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.555, + 0.72, + 0.57 + ], + "angle": 0, + "content": "Figure 8. Comparison of the 3D scene perspective results with SOTA methods." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.58, + 0.458, + 0.636 + ], + "angle": 0, + "content": "
OCC GT size0.4m0.8m1.6m
TargetmIoUAbs-RelmIoUAbs-RelmIoUAbs-Rel
FBOCC40.690.34236.010.37026.960.445
SurroundSDF42.370.17440.740.17640.160.181
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.639, + 0.468, + 0.653 + ], + "angle": 0, + "content": "Table 6. Results of the experiments on the resolution of OCC GT." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.664, + 0.446, + 0.713 + ], + "angle": 0, + "content": "
Sampling rate10.50.20.1
mIoU42.3742.3742.1641.07
Abs-Rel0.1740.1700.1860.186
" + }, + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.716, + 0.452, + 0.73 + ], + "angle": 0, + "content": "Table 7. Ablation study on the density of LiDAR points GT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.74, + 0.47, + 0.846 + ], + "angle": 0, + "content": "prediction method FBOCC experienced a substantial mIoU decrease of \\(33.7\\%\\), coupled with a \\(30.1\\%\\) increase in Abs-Rel. As shown in Table 7, utilizing only \\(10\\%\\) of the point cloud resulted in a marginal decrease of \\(3.1\\%\\) in mIoU and a modest increase of \\(6.9\\%\\) in Abs-Rel. These experiments further substantiate the robustness and efficacy of the proposed Sandwich Eikonal supervision paradigm." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.855, + 0.342, + 0.87 + ], + "angle": 0, + "content": "5.5. Comparison of Visual Results" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We visualize our 3D scene perception results and compare them with the reproduced SIREN [34] and LODE [14]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.582, + 0.892, + 0.643 + ], + "angle": 0, + "content": "(models B and C in Section 5.4), and the SOTA occupancy prediction method FBOCC [19, 20], as illustrated in Figure 8. Our SurroundSDF demonstrates significant advantages in the continuity and accuracy of the geometric structure." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.653, + 0.62, + 0.668 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose SurroundSDF, a novel vision-centric 3D scene understanding framework. We introduce SDF to address the continuity and accuracy of perception from surround cameras. Moreover, in the absence of SDF GT, we propose Sandwich Eikonal formulation, a novel SDF supervision paradigm, which emphasizes imposing appropriate constraints on both sides of the surface to enhance geometric accuracy. Furthermore, to alleviate the inconsistency between geometric optimization objectives and semantic optimization objectives, we introduce a joint supervision strategy. Based on the generated SDF and semantic field, different 3D representations can be obtained, including the scene mesh, occupancy voxel grids, and semantic reconstruction mesh of the whole scene. Comprehensive experiments validate the performance of our method." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21621" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Ben Agro, Quinlan Sykora, Sergio Casas, and Raquel Urtasun. Implicit occupancy flow fields for perception and prediction in self-driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1379-1388, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.472, + 0.325 + ], + "angle": 0, + "content": "[3] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.381 + ], + "angle": 0, + "content": "[4] Anh-Quan Cao and Raoul de Charette. Scenerf: Self-supervised monocular 3d scene reconstruction with radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9387–9398, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.437 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.493 + ], + "angle": 0, + "content": "[6] Vitor Guizilini, Rui Hou, Jie Li, Rares Ambrus, and Adrien Gaidon. Semantically-guided representation learning for self-supervised monocular depth. In International Conference on Learning Representations, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.47, + 0.549 + ], + "angle": 0, + "content": "[7] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.47, + 0.606 + ], + "angle": 0, + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.47, + 0.675 + ], + "angle": 0, + "content": "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[10] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[11] Peixiang Huang, Li Liu, Renrui Zhang, Song Zhang, Xinli Xu, Baichao Wang, and Guoyi Liu. Tig-bev: Multi-view bev 3d object detection via target inner-geometry learning. 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[12] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9223-9232, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Diederick P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[14] Pengfei Li, Ruowen Zhao, Yongliang Shi, Hao Zhao, Jirui Yuan, Guyue Zhou, and Ya-Qin Zhang. Lode: Locally conditioned eikonal implicit scene completion from sparse lidar. In IEEE International Conference on Robotics and Automation, ICRA 2023, London, UK, May 29 - June 2, 2023, pages 8269-8276. IEEE, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[15] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1486-1494, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[16] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[17] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[18] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[19] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[20] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M Alvarez. Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6919-6928, 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[21] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, pages 531-548. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[22] Yingfei Liu, Junjie Yan, Fan Jia, Shuailin Li, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petrv2: A unified framework for 3d perception from multi-camera images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3262-3272, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[23] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[24] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781. IEEE, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21622" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[25] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.469, + 0.202 + ], + "angle": 0, + "content": "[26] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[27] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3D Vision (3DV), 2016 Fourth International Conference on, pages 565-571. IEEE, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[28] Timothy S Newman and Hong Yi. A survey of the marching cubes algorithm. Computers & Graphics, 30(5):854-879, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[29] Mingjie Pan, Li Liu, Jiaming Liu, Peixiang Huang, Longlong Wang, Shaoqing Xu, Zhiyi Lai, Shanghang Zhang, Kuiyuan Yang, and Xiaomi Car. Uniocc: Unifying vision-centric 3d occupancy prediction with geometric and semantic rendering. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[30] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.524 + ], + "angle": 0, + "content": "[31] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[32] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[33] Aron Schmied, Tobias Fischer, Martin Danelljan, Marc Pollefeys, and Fisher Yu. R3d3: Dense 3d reconstruction of dynamic scenes from multiple cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3216-3226, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[34] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33:7462-7473, 2020. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[35] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[36] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8406-8415, 2023. 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] Thang Vu, Jung-Hee Kim, Myeongjin Kim, Seokwoo Jung, and Seong-Gyun Jeong. Milo: Multi-task learning with localization ambiguity suppression for occupancy prediction cvpr 2023 occupancy challenge report. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[38] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 34:27171-27183, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[39] Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, , and Justin M. Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In The Conference on Robot Learning (CoRL), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[40] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Yongming Rao, Guan Huang, Jiwen Lu, and Jie Zhou. Surround-depth: Entangling surrounding views for self-supervised multi-camera depth estimation. In Conference on Robot Learning, pages 539-549. PMLR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.29, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[41] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.36, + 0.892, + 0.429 + ], + "angle": 0, + "content": "[42] Enze Xie, Zhiding Yu, Daquan Zhou, Jonah Philion, Anima Anandkumar, Sanja Fidler, Ping Luo, and Jose M Alvarez. \\(\\mathbf{M}^2\\) bev: Multi-camera joint 3d detection and segmentation with unified birds-eye view representation. arXiv preprint arXiv:2204.05088, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.431, + 0.892, + 0.485 + ], + "angle": 0, + "content": "[43] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. In The Eleventh International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.488, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[44] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.53, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[45] Weihao Yuan, Xiaodong Gu, Heng Li, Zilong Dong, and Siyu Zhu. Monocular scene reconstruction with 3d sdf transformers. In The Eleventh International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[46] Yunpeng Zhang, Phigent Robotics, and Zheng Zhu. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9433-9443, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21623" + } + ] +] \ No newline at end of file diff --git a/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/213de3e8-6a97-4d16-96b2-ac03f2f0441a_origin.pdf b/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/213de3e8-6a97-4d16-96b2-ac03f2f0441a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b144735ef636bed43a21ebf0e61b46677bd8d90d --- /dev/null +++ b/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/213de3e8-6a97-4d16-96b2-ac03f2f0441a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3c03e25716c6c75a63752d57e985642a912fa0bd30a84f4cd235947ca3e01a5 +size 1742734 diff --git a/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/full.md b/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bd141b4d8d91ba88273852f67d9271888c4a5cf1 --- /dev/null +++ b/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/full.md @@ -0,0 +1,370 @@ +# SurroundSDF: Implicit 3D Scene Understanding Based on Signed Distance Field + +Lizhe Liu $^{1*}$ Bohua Wang $^{2*§}$ Hongwei Xie $^{1*}$ Daqi Liu $^{1}$ Li Liu $^{1}$ Zhiqiang Tian $^{2}$ Kuiyuan Yang $^{1}$ Bing Wang $^{1\dagger}$ $^{1}$ Xiaomi EV $^{2}$ School of Software Engineering, Xi'an Jiaotong University + +# Abstract + +Vision-centric 3D environment understanding is both vital and challenging for autonomous driving systems. Recently, object-free methods have attracted considerable attention. Such methods perceive the world by predicting the semantics of discrete voxel grids but fail to construct continuous and accurate obstacle surfaces. To this end, in this paper, we propose SurroundSDF to implicitly predict the signed distance field (SDF) and semantic field for the continuous perception from surround images. Specifically, we introduce a query-based approach and utilize SDF constrained by the Eikonal formulation to accurately describe the surfaces of obstacles. Furthermore, considering the absence of precise SDF ground truth, we propose a novel weakly supervised paradigm for SDF, referred to as the Sandwich Eikonal formulation, which emphasizes applying correct and dense constraints on both sides of the surface, thereby enhancing the perceptual accuracy of the surface. Experiments suggest that our method achieves SOTA for both occupancy prediction and 3D scene reconstruction tasks on the nuScenes dataset. + +# 1. Introduction + +With the recent advancement of 3D object detection algorithms [10, 11, 16, 18, 21, 24], vision-centric autonomous driving system become more practicable. Nevertheless, the persisting challenges related to the long-tail problem and the coarse depiction of the 3D scene underscore its insufficiency. Consequently, a deeper comprehension of 3D geometry and semantics is needed for safety and reliability. This paper delves into a novel vision-centric paradigm of dense and continuous 3D scene understanding. + +Current approaches of dense 3D prediction can be classified into two categories, 3D reconstruction and 3D perception. Specifically, 3D reconstruction algorithms [6, 33] generate dense point clouds enriched with semantic informa + +![](images/610617bf759b236a5682727c123852d1071a6c659b5b9d9eb314eb7248ad29ab.jpg) +a. SIREN + +![](images/c857769f95962bfe4c30e734d2055cba2c8c6a78a6c87fe0c1fb3734e4b01bfa.jpg) +b. LODE + +![](images/77fc15699dcad5e9666cf396005dd220778f551e41a701e1f94cefaa21596f1f.jpg) +c. Ours +Figure 1. The scene perspective results from surround images input. a, b. The result of SIREN and LODE supervision with camera-only input (the original methods are point cloud-based). c. Our result. d. Our result with semantics. + +![](images/6c9f094630e34f0067725fc7641387fb142eec39ca21291330ed3000111c7b63.jpg) +d. Ours with semantic + +tion by projecting depth maps and 2D semantic maps into 3D space. 3D perception methods [3, 12, 17, 25, 35, 41, 46] predict the occupied status and the semantic of 3D voxel grids. However, either method proves to be redundant or can only predict coarse-grained discrete grids. + +Instead, we develop a vision-centric framework to describe 3D scenes leveraging neural implicit Signed Distance Function (SDF) representation, which we refer to as SurroundSDF. Generally, we aim to: (1) Continuously describe the 3D scene by reconstructing smooth surfaces; (2) Explore the difficulty associated with the utilization of SDF representation and pose an appropriate training strategy; (3) Exploit the strong representation to kill two birds with one stone, addressing 3D semantic segmentation and continuous 3D geometry reconstruction within one framework. + +In pursuit of the first and the third objectives, we employ SDF to represent the canonical 3D scene in terms of distance from the surface. Furthermore, we construct an implicit field that encompasses both semantics and geometry by exploiting this strong representation. However, it's still nontrivial to accomplish the SDF modeling, where complete surface and normal vectors, which are crucial supervisions, cannot be accurately computed. To this end, SIREN [34] proposed a weakly supervised implicit per + +ception method, which accomplishes 3D mesh completion from point clouds. While LODE [14] proposed to utilize occupancy ground truth (GT) to supervise SDF. However, these methods rely on the input of point clouds, and only alleviate this difficulty by mimicking the surface distance but are either sparse or inaccurate. + +To overcome the aforementioned challenges, we propose the Sandwich Eikonal formulation, a novel weak supervision paradigm for SDF modeling. Figure 1 gives some reconstruction results and shows the benefit of our method. This method emphasizes applying correct and dense constraints on both sides of the surface to enhance the geometric accuracy and continuity of the surface. Moreover, we revisit current pipelines for training the perception branch and design a novel loss that enhances the integration of geometry and semantics thereby reducing inconsistencies. Our contributions can be summarized as follows: + +- We propose a vision-centric implicit semantic SDF perception method, achieving accurate and continuous 3D perception. To the best of our knowledge, we are the first to utilize SDF for surround-view 3D perception. +- We introduce the Sandwich Eikonal formulation, a novel weak supervision paradigm for SDF. +- We demonstrate how to employ this representation for the reconstruction of continuous 3D geometry with precise semantic information. +- We achieve state-of-the-art results on 3D dense semantics perception tasks. Comprehensive experiments on the NuScenes dataset [2] provide extensive validation of our approach. + +# 2. Related Work + +Occupancy Prediction Recently, researches [3, 12, 17, 25, 32, 32, 35, 41, 46] on occupancy prediction have demonstrated advantages in 3D scene understanding. Compared to the traditional object detection paradigm [5, 10, 11, 15, 16, 18, 21, 22, 24, 39, 42], occupancy perception has the following advantages. First, it can express dense 3D geometry. Second, it can accurately provide spatial locations for objects beyond predefined categories. Third, it can describe the shapes of irregular obstacles. + +Based on these advantages, occupancy prediction tasks have attracted significant research. These methods predict the occupancy status in the region of interest around the ego vehicle from point clouds or images. Specifically, the space is first divided into voxel grids at a specific resolution. Then the occupancy status and semantics of each grid are estimated. SurroundOcc [41] proposed a surround-view 3D occupancy perception method that utilizes spatial 2D-3D attention to lift image features into 3D space. In addition, to realize the dense occupancy prediction, SurroundOcc designed a pipeline to convert the point cloud to dense occupancy ground truth. VoxFormer [17] employed + +an MAE-like [9] approach to achieve camera-based semantic occupancy prediction. FB-OCC [20, 20] proposed a novel forward-backward projection method to compensate for the insufficient BEV feature density of the forward projection method and the large number of mismatches in 2D and 3D space caused by the backward projection. Despite the impressive results, they are limited by the specific resolution of the occupancy annotations, which limit the continuous perception of the scene. + +Implicit Scene Perception Scene reconstruction refers to the task of predicting the 3D geometry structures from some incomplete representations, e.g., images, point clouds, and voxel grids. Implicit neural scene reconstruction methods [1, 4, 14, 26, 30, 34, 45] have demonstrated advantages in accuracy and substantial potential. Generally, they train a neural network to predict a continuous field. Thus, it is possible to query occupancy information for any point within the 3D space. DeepSDF [30] utilizes the SDF to achieve the implicit reconstruction of 3D objects at the instance level. NeRFs [26] and its numerous variants [4, 7, 43] achieve implicit 3D scene reconstruction and novel view synthesis from multi-view images. NeuS [38] and VolSDF [44] introduce the signed distance function into neural rendering, enabling more precise object surface reconstruction. SIREN [34] utilizes the Eikonal equation to achieve semi-supervised SDF reconstruction for point clouds. However, these methods lack generalization to novel scenes. To achieve online implicit scene perception in driving scenes, LODE [14] proposes the Locally Conditioned Eikonal formulation and introduces a dense occupancy ground truth for supervision, which significantly improves the effect of SDF reconstruction in the driving scenes. However, it is limited by the sparsity of input LiDAR points or inaccurate ground truth. To solve these problems, we propose a camera-only implicit scene understanding method, which aims at continuous and accurate surface perception. + +# 3. Formulation + +This section analyzes the SDF supervision paradigm based on the Eikonal formulation with its variations and introduces our Sandwich Eikonal supervision approach. + +# 3.1. Eikonal-based SDF Constraints + +Given a coordinate $\mathbf{x}$ in the 3D scene, our goal is to construct a function $\phi$ such that $\phi(\mathbf{x})$ provides the SDF value at point $\mathbf{x}$ . The Eikonal-based optimization objectives [14] is + +$$ +\mathcal {O} _ {E} = \int_ {\Omega_ {0}} O _ {0} d \mathbf {x} + \int_ {\Omega_ {1}} O _ {1} d \mathbf {x} + \int_ {\Omega_ {2}} O _ {2} d \mathbf {x}, \tag {1} +$$ + +and + +$$ +\left\{ \begin{array}{l l} O _ {0} = | | \nabla_ {\mathbf {x}} \phi (\mathbf {x}) | - 1 |, & \mathbf {x} \in \Omega_ {0} \\ O _ {1} = | \nabla_ {\mathbf {x}} \phi (\mathbf {x}) - \mathbf {n} (\mathbf {x}) |, & \mathbf {x} \in \Omega_ {1}, \\ O _ {2} = | \phi (\mathbf {x}) - S D F (\mathbf {x}) |, & \mathbf {x} \in \Omega_ {2} \end{array} \right. \tag {2} +$$ + +where $\Omega_0$ is the set of the whole 3D space of interest, $\Omega_1$ is the set of points on the surface, $\Omega_2$ is the set of points + +![](images/88e186e3c3371d5010599239bd61c556c07e624a95a5fd5bd6be5a9e4ee3e907.jpg) +a. Ideal supervision + +![](images/b970adbac643898fd614ba27df17c68bef3bf71c1d87bbc0c586eeba0076deee.jpg) +b. SIREN +Figure 2. a. Ideal supervision for SDF where the SDF GT is provided. b. SIREN supervision which uses LiDAR points to supervise the surface. c. LODE supervision based on occupancy GT. d. Our supervision paradigm which combines the LiDAR GT and occupancy GT, it is closer to the ideal supervision. + +![](images/3cca8fda26a26d66195b1b41dfa91bbc136a3bb0ab2eca56ce7bb3b06f1afb5c.jpg) +c.LODE + +d. Ours +![](images/d129ffcfe9fa1e392b0941b20461847b9119f2cf7090edb61335d72f288fc181.jpg) +Constrain $SDF < 0$ +Constrain $SDF = 0$ +Constrain $SDF > 0$ +Constraint min SDF +of one grid $< 0$ +Surface +LiDAR Points +Voxel Grid + +with SDF GT annotations, $\nabla_{\mathbf{x}}\phi (\mathbf{x})$ is the gradient on point $\mathbf{x}$ , $\mathbf{n}(\mathbf{x})$ is the normal, and $SDF(\mathbf{x})$ is the real SDF value at point $\mathbf{x}$ . One major obstacle to this optimization objective lies in the design of $O_2$ , which provides direct supervision for SDF as shown in Figure 2.a. This objective requires the annotation of SDF values for each anchor point in the scene, which is monumental and challenging. + +In the absence of precise SDF annotations, SIREN [34] and LODE [14] have implemented weakly supervised paradigm through the following variations of $O_2$ , + +$$ +\left\{ \begin{array}{l l} O _ {2 - 0} ^ {\prime} = | \phi (\mathbf {x}) |, & \mathbf {x} \in \Omega_ {1} \\ O _ {2 - 1} ^ {\prime} = \psi (| \phi (\mathbf {x}) |), & \mathbf {x} \in \Omega_ {3}, \end{array} \right. \tag {3} +$$ + +where $\psi (\cdot)$ is a monotonically decreasing function and $\Omega_3\subseteq \Omega_0 / \Omega_1$ is the whole space except the surface. + +The objective $O_{2-0}^{\prime}$ aims to constrain the SDF values on the surface to 0. Considering the absence of precise supervision, $O_{2-1}^{\prime}$ adopts a fuzzy constraint that pushes the absolute SDF value away from zero in $\Omega_3$ . While the constraint in $\Omega_3$ is loose, it is adequate to capture the geometry information of the surface. + +In practical implementation, SIREN [34] samples on-surface LiDAR points in $\Omega_{1}$ and considers the space excluding the LiDAR points as $\Omega_{3}$ , as is shown in Figure 2.b. When the LiDAR points are dense enough, this strategy can effectively achieve good constraint effects [34]. However, LODE [14] has demonstrated that the sparsity of LiDAR points in the driving scenes results in discontinuities of the estimated mesh. To achieve denser supervision, they use the dense occupancy GT as supervision. This approach treats both the surface and interior of the obstacles as the $\Omega_{1}$ . Specifically, they sample the $\Omega_{1}$ points as the center of "occupied" grids, and the $\Omega_{3}$ points as the center of "free" grids, as is shown in 2.c. However, this paradigm remains imprecise, as considering the center of the occupied grid to be the surface of objects would introduce errors. + +# 3.2. Sandwich Eikonal Formulation + +To address the above concerns, we require a supervision paradigm that has the following characteristics: 1) sampling points as the supervision on the surface must strictly adhere to the surface, 2) for regions beyond the surface, refrain + +from supervision in uncertain areas to ensure precision, and 3) dense supervision signals. Considering the above factors, we introduce our variants of Eikonal formulation as follows: + +$$ +\left\{ \begin{array}{l l} O _ {2 - 0} = \phi (\mathbf {x}), & \mathbf {x} \in \Omega_ {1} \\ O _ {2 - 1} = \psi (- \min (\phi (\mathbf {x}))), & \mathbf {x} \in \Omega_ {\text {o c c}} ^ {i} \\ O _ {2 - 2} = \psi (\phi (\mathbf {x})), & \mathbf {x} \in \Omega_ {\text {f r e e}} ^ {j}, \end{array} \right. \tag {4} +$$ + +where $\Omega_{occ}^{i}$ is the $i$ -th grid of occupied grids and $\Omega_{free}^{j}$ is the $j$ -th grid of free grids. + +To this end, we incorporate the LiDAR points and the occupancy GT into the SDF supervision. As is shown in Figure 2.d, for precise sampling points on the surface, we follow SIREN and constrain the SDF value of LiDAR points to be 0. To apply accurate and dense supervision to regions beyond the surface, we introduce the occupancy GT into the supervision, as is shown in optimization objectives $O_{2-1}$ and $O_{2-2}$ . According to the generation process of occupancy GT [35, 36], if a grid is labeled as "occupied", it implies that at least part of the grid area falls inside the object. Therefore, as the objective $O_{2-1}$ shows, we constrain the minimum SDF value of each occupied grid to be less than zero and push it to the negative range. On the other hand, if a grid is labeled as "free", the entire grid is outside the object and we push the SDF value of free grids to the positive range, as is shown in objective $O_{2-2}$ . We employ a multiframe point cloud fusion strategy to densify the sampling for surface supervision, as described in Section 4.2. Compared to previous supervision paradigms, our formulation emphasizes applying appropriate and dense supervision on both sides of the surface to enhance the geometric constraint based on the Eikonal formulation. Therefore, we name our supervision strategy the "Sandwich Eikonal" formulation. + +# 4. Method + +Given multi-camera images $\mathbf{I} = \{I_1, I_2, \dots, I_n\}$ , we aim to predict the 3D SDF $\phi(\mathbf{x}) : \mathbb{R}^3 \Rightarrow \mathbb{R}$ and semantic field $S(\mathbf{x}) : \mathbb{R}^3 \Rightarrow \mathbb{R}^s$ , where $s$ indicates the number of classes. The following will introduce our architecture and detailed constraints based on our Sandwich Eikonal formulation. + +![](images/ef8d1bbee8318de1f774218ef7b57ff1e49655ead98b147e8d2ca998b827fe49.jpg) +Figure 3. The architecture of our SurroundSDF. Given the surround images as input, an encoder composed of a 2D backbone, LSS module, and BEV backbone, is employed to extract voxel features. We adopt a query-based approach to sample features from the voxel features. Specifically, first, a set of query coordinates in the region of interest is selected. Subsequently, using trilinear interpolation, semantic features are queried from the voxel features. Finally, after concatenation with the positional embeddings from the query coordinates, the features pass through the SDF head and semantic head respectively, yielding SDF and semantic fields. For training, the query coordinates are sampled according to the GT, and the SDF and semantic field are supervised by the losses introduced in Section 4.2. In the inference phase, based on appropriate sampling and post-processing, continuous and accurate scene perception results are obtained (see Section 4.3). + +# 4.1. Architecture + +The overall architecture is shown in Figure 3. First, a CNN-based backbone is used to obtain the image features. Then, the LSS [31] is used to project the image features to BEV space, and a BEV encoder is adopted to obtain the voxel features $V$ of dimension $h \times w \times d \times c$ (c indicates the number of channels). Further, for each point $\mathbf{x}$ in $\Omega_0$ , the corresponding features can be obtained by querying $V$ with the trilinear sampling method. Therefore, $\phi(\mathbf{x})$ and $S(\mathbf{x})$ can be expressed as: + +$$ +\phi (\mathbf {x}) = H _ {s d f} \left(C \left(Q (V, \mathbf {x}), P E (\mathbf {x})\right), \theta\right), \mathbf {x} \in \Omega_ {0}, \tag {5} +$$ + +$$ +S (\mathbf {x}) = H _ {s e m} (C (Q (V, \mathbf {x}), P E (\mathbf {x})), \theta), \mathbf {x} \in \Omega_ {0}, +$$ + +where $\mathbf{x} \in \mathbb{R}^3$ , $C$ and $Q$ represent the concatenation and query process, respectively. $H_{sdf}$ and $H_{sem}$ indicate the SDF head and semantic head, implemented with three-layer MLPs with sine activation functions. $\theta$ represents the learnable parameters. $PE$ represents the positional encoding, which aims to capture high-frequency information. Specifically, for a coordinate $\mathbf{x} \in \mathbb{R}^3$ , $PE$ encodes each of its dimensions into a 2n-dimensional vector, it can be expressed + +$$ +\text {a s :} \quad P E (\mathbf {x}) = (\zeta (x), \zeta (y), \zeta (z)) \in \mathbb {R} ^ {6 n}, \tag {6} +$$ + +where $x, y, z$ denote the coordinates, $\zeta(\cdot)$ represents an encoding scheme using trigonometric function mapping: + +$$ +\zeta (x) = \left(\sin \left(2 ^ {0} \pi x\right), \cos \left(2 ^ {0} \pi x\right), \dots , \cos \left(2 ^ {n - 1} \pi x\right)\right). \tag {7} +$$ + +In training, we design a joint supervision method that utilizes both coarse-grained voxel and fine-grained point cloud supervision based on our Sandwich Eikonal formulation. In inference, we query the voxel features to estimate the SDF values and semantic logits of each query point. + +# 4.2. Supervision + +This section will elaborate on deriving the SDF loss from our Sandwich Eikonal formulation. Then, we will introduce the semantic loss and our joint supervision strategy. + +SDF Supervision Following objectives $O_0$ , $O_1$ in Equation 2 and $O_{2-0}$ to $O_{2-2}$ in Equation 4, we derive the losses for SDF supervision, based on our Sandwich Eikonal formulation. Considering that continuous space supervision is unfeasible, we sample the discrete points in the corresponding space $(\Omega_0, \Omega_1, \Omega_{occ}$ and $\Omega_{free})$ for supervision. + +![](images/733d3c8523d8de7ff252598d06e9c0b78cd419f865eda5194830af270b6af1e2.jpg) +a. + +![](images/b3fcc936e39af8d4b1a7d4428e8dbe49c23fe53a8f592b113fe12d2e18738be3.jpg) +b. +Figure 4. a. SDF constraints with Sandwich Eikonal formulation in continuous form. b. SDF sampling in discrete form. + +The correctness and density of the supervision sampling are of paramount importance. Since our approach is based on the surround cameras, we only sample from the camera-visible space in $\Omega_0$ (space of interest) [36]. For $\Omega_{1}$ (surface points), we sample exclusively from the LiDAR points. To enhance the sampling density, temporal point cloud fusion is adopted. Specifically, $\Omega_{1} = \{p_{t - k},p_{t - (k - 1)},\dots ,p_{t + (k - 1)},p_{t + k}\}$ , where $p_{t + j}$ is the LiDAR points of the $t + j$ frame and $t$ is the current frame. Based on occupancy GT [36], $\Omega_{occ}$ and $\Omega_{free}$ are the sets of + +voxel grids labeled as "occupied" and "free", respectively. + +Additionally, in our objective $O_{2-1}$ , the computation of the minimum SDF within the occupied grid is required. As is shown in Figure 4, we approximate this minimum value through discrete sampling. The minimum value among the SDF values at the centers of each sub-grid is regarded as the minimum SDF value for the grid. Based on Equation 2, 4 and the above sampling strategy, our SDF loss $\mathcal{L}_{sdf}$ is expressed as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {s d f} = \gamma_ {1} \frac {1}{n _ {t o t a l}} \sum_ {i = 1} ^ {n _ {t o t a l}} \left(\left| \nabla_ {\mathbf {x}} \phi (\mathbf {x} ^ {i}) \right| - 1\right) \\ + \gamma_ {2} \frac {1}{n _ {1 \cup o c c}} \sum_ {i = 1} ^ {n _ {1 \cup o c c}} (\nabla_ {\mathbf {x}} \phi (\mathbf {x} ^ {i}) - n (\mathbf {x} ^ {i})) \\ + \gamma_ {3} \frac {1}{n _ {1}} \sum_ {i = 1} ^ {n _ {1}} | \phi (\mathbf {x} ^ {i}) | \tag {8} \\ + \gamma_ {4} \frac {1}{n _ {o c c}} \sum_ {i = 1} ^ {n _ {o c c}} e ^ {\alpha \times m i n (\phi (\mathbf {x} ^ {i}), \mathbf {x} ^ {i} \in \Omega_ {o c c} ^ {i})} \\ + \gamma_ {5} \frac {1}{n _ {f r e e}} \sum_ {i = 1} ^ {n _ {f r e e}} e ^ {- \alpha \times r a n d o m (\phi (\mathbf {x} ^ {i}), \mathbf {x} ^ {i} \in \Omega_ {f r e e} ^ {i})}, \\ \end{array} +$$ + +where $\alpha$ is a hyperparameter that determines the extent of the deviation for the interior and exterior space of the object and we fix $\alpha$ to 100 in experiments. $\gamma_{1},\gamma_{2},\gamma_{3},\gamma_{4},\gamma_{5}$ are loss weights, $n(\mathbf{x})$ represents the normal vector of the current point on surface. $n_1,n_{occ}$ , and $n_{free}$ represent the number of points or voxel grids sampled from $\Omega_1,\Omega_{occ}$ and $\Omega_{free}$ , respectively. $n_{total}$ is the total number of samples: $n_{total} = n_1 + n_{occ} + n_{free}$ . + +Semantic Supervision A widely used cross-entropy loss is adopted to supervise the semantic field as follows: + +$$ +\mathcal {L} _ {s e m} = \frac {1}{n _ {o c c}} \sum_ {i = 1} ^ {n _ {o c c}} C E \left(S (\mathbf {x} | \mathbf {x} \in \Omega_ {o c c} ^ {i}), \dot {y} _ {s e m} ^ {i}\right), \tag {9} +$$ + +where CE represents cross-entropy loss and $\hat{y}_{sem}^i$ is the semantic ground truth of i-th voxel grid. + +Joint Supervision The above losses optimize the SDF and semantic field separately, which leads to a nonnegligible issue of ambiguity between geometric and semantic optimization. To illustrate this issue, we train a model and evaluate the occupancy IoU representing geometric accuracy and the semantic mIoU representing overall semantic accuracy following the occupancy prediction metrics [35, 36]. For each voxel grid, if its minimum SDF is less than a certain threshold, the voxel is considered 'occupied'. As is shown in Figure 5, the optimal thresholds producing the highest peak of these curves are significantly different. This indicates that greater geometric accuracy does not necessarily lead to improved overall semantic accuracy. + +To mitigate this issue, we propose a joint supervision strategy. First, for the minimum SDF value $\min(\phi(\mathbf{x}))$ of each voxel grid, a SoftMax function is used to convert it to + +![](images/ccb7a9fa23c8ade7c27ec8c0d8fcd7de2555818bbb85abc5be7779f509fa86de.jpg) +Figure 5. Variation of occupancy IoU and semantic mIoU with SDF Threshold. Note that the peak values of these two indicators correspond to different SDF thresholds. + +the free probability: $p_{free} = \text{softmax}(\beta t, \beta \min(\phi(\mathbf{x}))$ , where $t$ is fixed to 0.005, $\beta$ is fixed to 100. Second, we map the free probability $p_{free}$ to logits using an inv-sigmoid function: $l_{free} = \log(p_{free} / (1 - p_{free}))$ . Finally, $l_{free}$ is combined with semantic logits $l_{sem}$ , and the joint loss $\mathcal{L}_{joint}$ is expressed as: + +$$ +\mathcal {L} _ {\text {j o i n t}} = \operatorname {D i c e} \left(C \left(l _ {\text {s e m}}, l _ {\text {f r e e}}\right), C \left(\hat {y} _ {\text {s e m}}, \hat {y}\right)\right), \tag {10} +$$ + +where $C$ is the concatenation operation and $\hat{y}_{sem}$ is the semantic GT of voxel grid. Dice represents the Dice loss [27]. Finally, the total loss is calculated as: + +$$ +\mathcal {L} = \gamma_ {6} \mathcal {L} _ {s d f} + \gamma_ {7} \mathcal {L} _ {s e m} + \gamma_ {8} \mathcal {L} _ {j o i n t} \tag {11} +$$ + +where $\gamma_6, \gamma_7, \gamma_8$ are loss weights. + +# 4.3. Inference + +Based on the output SDF and semantic field, diverse outputs for different tasks can be obtained through different sampling strategies and post-processing methods. + +Semantic Mesh Generation To generate the semantic mesh, we uniformly query the SDF value in the space of interest and use the marching cubes algorithm [28] to obtain the mesh. Subsequently, at the vertex positions of each triangular face, we query the corresponding semantic information, resulting in a mesh enriched with semantics. + +Occupancy Prediction The objective of the occupancy prediction is to anticipate the semantics of the voxel grid in space. We achieve this through the joint prediction based on the predicted SDF and semantic field as follows: + +$$ +\mathcal {S} _ {i} = \operatorname {a r g m a x} \left(C \left(l _ {\text {s e m}} ^ {i}, l _ {f} ^ {i}\right)\right), \tag {12} +$$ + +where $S_{i}$ is the i-th voxel grid, $l_{sem}^{i}$ is the logits of the i-th voxel grid, and $l_{f}^{i}$ the logits value from SDF. + +Semantic Scene Reconstruction Our SurroundSDF demonstrates a notable advantage by achieving semantic scene reconstruction concurrently with scene perception. As is shown in Figure 6, we sequentially query the semantic and SDF values for the region of interest frame by frame. The SDF values and semantic logits of static objects are projected onto the sampling points in the world coordinate + +![](images/ff7020888be29ac63c2182a9e0baf55fbb9d5ea9263c0a3eb359d672b71df0fd.jpg) +Figure 6. The process of semantic scene reconstruction. SDF values and semantic logits are sampled long the vehicle trajectory into the world coordinate system. For regions where the historical sampling overlaps with the current frame, the historical SDF and semantic logits are updated based on a momentum update strategy. + +system. For the overlapping of the historical sampling area and the current frame sampling area, we employ a momentum update strategy to refine the historical SDF values and semantic logits. Figure 7 shows the final scene reconstruction result with the marching cubes algorithm. + +# 5. Experiments + +# 5.1. Experimental Setup + +
VersionBackboneImage sizeChannelsFrames
SmallResNet50 [8]256 × 704329
MediumResNet101 [8]640 × 16001286
LargeConvNext-B [23]640 × 16002563
+ +Implementation details We follow BEVStereo [15] to construct our 3D encoder, where the depth net predicts 88 discrete depth values, the voxel feature resolution is $200 \times 200 \times 16$ , and the temporal fusion strategy is applied. We pre-train our architecture on semantic segmentation, depth estimation, and 3D object detection tasks on the NuScenes training set. To compare with methods under different settings, we train three versions: Small, Medium, and Large with different backbone, input image size, voxel feature channels, and the number of temporal frames, as is shown in Tabel 1. For the medium and large versions, we reduce the number of temporal frames to save GPU memory. We use the Adam optimizer [13], set the batch size to 32, fix the learning rate to 1e-4, and train for around 32 epochs for each version. The loss weights $\gamma_{1} \sim \gamma_{8}$ are set to 1.0, 1.0, 30.0, 0.05, 0.05, 1.0, 1.0, and 1.0 respectively. + +Dataset We evaluate our method on Occ3D-nuScenes dataset [35, 36], which is an occupancy dataset with dense voxel grid label based on the nuScenes dataset [2]. It processes dynamic and static objects separately, uses multiframe aggregation to achieve point cloud densification, and then reconstructs the dense point cloud to obtain a 3D occupancy label. It has voxel-level semantic labels of 17 classes, + +![](images/56dbb50ed206efffd9f081e546fca7cae165dd5556febb7542d9c206561341aa.jpg) +a. Scene reconstruction results from a global perspective + +![](images/50ccaa20b9b2f2cd148b6e03ef81513ada1d4cd6b7d48c5ba50ac1be4aadf651.jpg) +b. Scene reconstruction results from a local perspective +Figure 7. Visualization of semantic scene reconstruction. + +which contain 16 common classes and an additional general object class. Each frame covers a range of $[-40\mathrm{m}, -40\mathrm{m}, -1\mathrm{m}, 40\mathrm{m}, 40\mathrm{m}, 5.4\mathrm{m}]$ with a voxel size of $0.4\mathrm{m}$ , resulting in a voxel grid resolution with $200 \times 200 \times 16$ . + +Evaluation Metrics To evaluate our performance on dense 3D scene perception, we follow the 3D semantic occupancy prediction task [36] and use the widely adopted mean intersection over union (mIoU). + +$$ +m I o U = \frac {1}{C} \sum_ {i = 1} ^ {C} \frac {T P _ {i}}{T P _ {i} + F P _ {i} + F N _ {i}}, \tag {13} +$$ + +where $C$ represents the number of categories (set to 17 in the experiment), and TP, FP, and FN represent true positive, false positive, and false negative respectively. + +Table 1. Experimental settings of different versions. + +
MetricsFormula
Abs Rel1/n∑|d−d*|/d*
Sq Rel1/n∑|d−d*|2/d*
RMSE√1/n∑|d−d*|2
δ<1.251/n∑(max(d/d*, d*/d) < 1.25)
+ +Table 2. Depth metrics for 3D scene reconstruction + +To evaluate our scene reconstruction performance, we render the 2D depth image for the surround cameras based on the reconstructed scene and validate using 2D depth evaluation. As is shown in Table 2, we use the metrics commonly used for depth estimation, which aim to measure the difference between the predicted depth obtained by mesh rendering and GT depth obtained by LiDAR points. + +# 5.2.3D Semantic Occupancy Prediction + +Following the widely adopted experimental setup [19, 36], we conduct the semantic occupancy prediction experiment on the Occ3D-nuScenes dataset. For a fair comparison, in the inference phase, we divide the space into voxel grids with the same size as the other methods (0.4m) and predict the occupancy semantics following Section 4.3. As is shown in Tabel 3, the mIoU score of our Small version using ResNet50 as the backbone and without any test-time-augmentation (TTA) achieves 42.37, surpassing all the + +methods with the same image backbone. Additionally, our Medium version significantly outperforms BEVFormer [18] and VoxFormer [17] using the ResNet101 backbone. Moreover, our Large version surpasses UniOCC [29] with the same ConvNext-B backbone by a 1.5 mIoU score, and even outperforms FB-OCC [19, 20], which utilizes a larger image backbone and TTA strategy. These results demonstrated the effectiveness of SurroundSdf in the perception task. + +# 5.3. 3D Scene Reconstruction + +To evaluate the performance of scene reconstruction, we compare our method with state-of-the-art (SOTA) surround-view-based depth estimation methods [33, 40] and occupancy prediction methods [19, 41]. To acquire the surround depth, we utilize the semantic mesh generation steps described in Section 4.3 to generate the mesh and then render the depth maps for each camera based on the camera parameters. To fairly compare with the occupancy-based methods, we follow a similar process to render the depth maps with predicted occupied voxel grids, and all methods use RestNet50 as the image backbone. As is shown in Table 4, our method exhibits significant advantages over occupancy-based approaches across all metrics. Moreover, on the primary metric of absolute relative error, we achieve SOTA results, surpassing SurroundDepth [40] and R3D3 [33], which are the SOTA methods for surround-view depth estimation. + +
MethodsBackboneImage SizemIoU
SurroundOcc [41]ResNet50256 × 70436.32
BEVStereo [15]ResNet50384 × 70439.1
UniOCC [29]ResNet50256 × 70439.7
MiLO [37]ResNet50256 × 70440.49
FB-OCC [19]ResNet50256 × 70440.69
FB-OCC* [19]ResNet50256 × 70442.06
OursResNet50256 × 70442.37
BEVFormer [18]ResNet101900 × 160040.6
VoxFormer [17]ResNet101900 × 160040.6
OursResNet101640 × 160046.0
SurroundOcc [41]InterImage-B640 × 160040.7
BEVFormer [18]InterImage-XL640 × 160043.3
BEVDet [10]Swin-B640 × 160043.1
UniOCC [29]ConvNext-B640 × 160051.5
FB-OCC* [19]InternImage-H960 × 176952.79
OursConvNext-B640 × 160053.01
+ +# 5.4. Ablation Study + +In this section, we study the proposed strategies and the impact of GT quality. + +Table 3. Comparison on the Occ3D-nuScenes val set. The superscript * denotes using test time augmentation. + +
MethodsAbs Rel ↓Sq Rel ↓RMSE ↓δ < 1.25 ↑
SurroundOcc0.2742.0725.3270.482
FB-OCC0.3422.0475.9700.290
SurroundDepth0.2242.1024.5730.751
R3D30.2592.3285.5770.583
Ours0.1741.0974.4020.747
+ +Table 4. 3D scene reconstruction performance. AbsRel is the main metric. “↓” means less is better. + +Ablation Study on Proposed Strategies We conduct ablation experiments based on the settings of the Small version in Tabel 1 and the results are shown in Tabel 5. These studies aim to analyze the effectiveness of different 3D representations and supervisions. Specifically, we first construct a baseline model (model A) that is supervised by the occupancy GT with the cross-entropy loss and output the occupancy results. Then, to investigate the 3D perception capability of implicit SDF and semantic field, we build the subsequent models based on the architecture in Section 4.1 and output the SDF and semantic field. To analyze the proposed weak supervision paradigm with previous methods, we reproduce the supervision methods of SIREN (model B) and LODE (model C). In model D, we employ our Sandwich Eikonal for SDF supervision. Lastly, model E incorporates the joint supervision strategy in section 4.2. + +Comparing the results of Model B, and C with A, we observe a significant advantage in the AbsRel metric for both models B and C. This highlights the robust reconstruction capability of SDF-based 3D representation. However, we also observe a significant decline in the semantic mIoU metric for model B, attributed to the sparse supervision with Li-DAR points. LODE mitigates this issue by introducing occupancy supervision, but its mIoU score is still lower than the baseline. Using the proposed Sandwich Eikonal supervision paradigm, the results of model D not only surpass the baseline in semantic mIoU but also outperform models B and C in AbsRel. This indicates the effectiveness of our supervision approach, which successfully integrates the supervision from LiDAR points and occupancy GT. Furthermore, by incorporating of the proposed joint supervision strategy, model E exhibits a 1.85 improvement in semantic mIoU, accompanied by a slight regression of 0.003 in the AbsRel metric. For some long-tail categories, such as bicycle, motorcycle, and traffic cone, our joint supervision strategy yields a noticeable improvement. + +Ablation Study on Ground Truth Quality To further validate the roles of OCC GT and LiDAR GT, we conducted ablation experiments on the resolution of OCC GT and the density of LiDAR points respectively. As is shown in Table 6, the impact of OCC GT resolution variations in our method's OCC GT is mitigated with LiDAR points supervision. When the voxel grid resolution increased from $0.4\mathrm{m}$ to $1.6\mathrm{m}$ , the mIoU only decreased by $5.2\%$ , and Abs-Rel increased by a mere $4.0\%$ . In comparison, the SOTA OCC + +
Model3D RepresentationSDF SupervisionJoint Supothersbarrierbicyclebuscarvehiclemotorcyclepedestriantraffic conetrailertruckdrivable surfaceother flatsidewalkterrainmannadevegetationmIoUAbs-Rel
AOCC——8.247.07.643.651.425.618.421.722.432.238.783.643.055.357.750.644.638.330.342
BSDFSIREN—8.041.418.625.941.019.519.015.623.019.827.872.338.648.746.536.633.631.510.188
CSDFLODE—8.746.018.638.547.225.322.220.426.323.734.581.244.854.053.840.337.836.660.192
DSDFSandwich—12.347.925.443.051.727.424.924.927.434.039.483.642.954.458.148.343.240.520.171
ESDFSandwich✓13.949.727.844.653.030.029.028.331.135.841.283.644.655.358.949.643.842.370.174
+ +![](images/c19d81ea6bb47b78f02437a3a36b7cbb8822515cb969eb58df349243c9de11ec.jpg) +Figure 8. Comparison of the 3D scene perspective results with SOTA methods. + +Table 5. Results of the ablation experiments. "3D Representation" indicates whether the scene is described by occupancy voxels or SDF. "SDF Supervision" indicates the supervised paradigm. "Joint Sup" indicates whether the joint supervision strategy in Section 4.2 is used. + +
OCC GT size0.4m0.8m1.6m
TargetmIoUAbs-RelmIoUAbs-RelmIoUAbs-Rel
FBOCC40.690.34236.010.37026.960.445
SurroundSDF42.370.17440.740.17640.160.181
+ +Table 6. Results of the experiments on the resolution of OCC GT. + +
Sampling rate10.50.20.1
mIoU42.3742.3742.1641.07
Abs-Rel0.1740.1700.1860.186
+ +Table 7. Ablation study on the density of LiDAR points GT. + +prediction method FBOCC experienced a substantial mIoU decrease of $33.7\%$ , coupled with a $30.1\%$ increase in Abs-Rel. As shown in Table 7, utilizing only $10\%$ of the point cloud resulted in a marginal decrease of $3.1\%$ in mIoU and a modest increase of $6.9\%$ in Abs-Rel. These experiments further substantiate the robustness and efficacy of the proposed Sandwich Eikonal supervision paradigm. + +# 5.5. Comparison of Visual Results + +We visualize our 3D scene perception results and compare them with the reproduced SIREN [34] and LODE [14] + +(models B and C in Section 5.4), and the SOTA occupancy prediction method FBOCC [19, 20], as illustrated in Figure 8. Our SurroundSDF demonstrates significant advantages in the continuity and accuracy of the geometric structure. + +# 6. Conclusion + +In this work, we propose SurroundSDF, a novel vision-centric 3D scene understanding framework. We introduce SDF to address the continuity and accuracy of perception from surround cameras. Moreover, in the absence of SDF GT, we propose Sandwich Eikonal formulation, a novel SDF supervision paradigm, which emphasizes imposing appropriate constraints on both sides of the surface to enhance geometric accuracy. Furthermore, to alleviate the inconsistency between geometric optimization objectives and semantic optimization objectives, we introduce a joint supervision strategy. Based on the generated SDF and semantic field, different 3D representations can be obtained, including the scene mesh, occupancy voxel grids, and semantic reconstruction mesh of the whole scene. Comprehensive experiments validate the performance of our method. + +# References + +[1] Ben Agro, Quinlan Sykora, Sergio Casas, and Raquel Urtasun. Implicit occupancy flow fields for perception and prediction in self-driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1379-1388, 2023. 2 +[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6 +[3] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1, 2 +[4] Anh-Quan Cao and Raoul de Charette. Scenerf: Self-supervised monocular 3d scene reconstruction with radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9387–9398, 2023. 2 +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2 +[6] Vitor Guizilini, Rui Hou, Jie Li, Rares Ambrus, and Adrien Gaidon. Semantically-guided representation learning for self-supervised monocular depth. In International Conference on Learning Representations, 2019. 1 +[7] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. 2023. 2 +[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2 +[10] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1, 2, 7 +[11] Peixiang Huang, Li Liu, Renrui Zhang, Song Zhang, Xinli Xu, Baichao Wang, and Guoyi Liu. Tig-bev: Multi-view bev 3d object detection via target inner-geometry learning. 2022. 1, 2 +[12] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9223-9232, 2023. 1, 2 +[13] Diederick P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015. 6 + +[14] Pengfei Li, Ruowen Zhao, Yongliang Shi, Hao Zhao, Jirui Yuan, Guyue Zhou, and Ya-Qin Zhang. Lode: Locally conditioned eikonal implicit scene completion from sparse lidar. In IEEE International Conference on Robotics and Automation, ICRA 2023, London, UK, May 29 - June 2, 2023, pages 8269-8276. IEEE, 2023. 2, 3, 8 +[15] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1486-1494, 2023. 2, 6, 7 +[16] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 1, 2 +[17] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023. 1, 2, 7 +[18] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1, 2, 7 +[19] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 6, 7, 8 +[20] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M Alvarez. Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6919-6928, 2023. 2, 7, 8 +[21] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, pages 531-548. Springer, 2022. 1, 2 +[22] Yingfei Liu, Junjie Yan, Fan Jia, Shuailin Li, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petrv2: A unified framework for 3d perception from multi-camera images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3262-3272, 2023. 2 +[23] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 6 +[24] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781. IEEE, 2023. 1, 2 + +[25] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. 1, 2 +[26] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[27] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3D Vision (3DV), 2016 Fourth International Conference on, pages 565-571. IEEE, 2016. 5 +[28] Timothy S Newman and Hong Yi. A survey of the marching cubes algorithm. Computers & Graphics, 30(5):854-879, 2006. 5 +[29] Mingjie Pan, Li Liu, Jiaming Liu, Peixiang Huang, Longlong Wang, Shaoqing Xu, Zhiyi Lai, Shanghang Zhang, Kuiyuan Yang, and Xiaomi Car. Uniocc: Unifying vision-centric 3d occupancy prediction with geometric and semantic rendering. 7 +[30] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2 +[31] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 4 +[32] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), 2020. 2 +[33] Aron Schmied, Tobias Fischer, Martin Danelljan, Marc Pollefeys, and Fisher Yu. R3d3: Dense 3d reconstruction of dynamic scenes from multiple cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3216-3226, 2023. 1, 7 +[34] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33:7462-7473, 2020. 1, 2, 3, 8 +[35] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 1, 2, 3, 5, 6 +[36] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8406-8415, 2023. 3, 4, 5, 6 +[37] Thang Vu, Jung-Hee Kim, Myeongjin Kim, Seokwoo Jung, and Seong-Gyun Jeong. Milo: Multi-task learning with localization ambiguity suppression for occupancy prediction cvpr 2023 occupancy challenge report. 7 + +[38] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 34:27171-27183, 2021. 2 +[39] Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, , and Justin M. Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In The Conference on Robot Learning (CoRL), 2021. 2 +[40] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Yongming Rao, Guan Huang, Jiwen Lu, and Jie Zhou. Surround-depth: Entangling surrounding views for self-supervised multi-camera depth estimation. In Conference on Robot Learning, pages 539-549. PMLR, 2023. 7 +[41] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 1, 2, 7 +[42] Enze Xie, Zhiding Yu, Daquan Zhou, Jonah Philion, Anima Anandkumar, Sanja Fidler, Ping Luo, and Jose M Alvarez. $\mathbf{M}^2$ bev: Multi-camera joint 3d detection and segmentation with unified birds-eye view representation. arXiv preprint arXiv:2204.05088, 2022. 2 +[43] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. In The Eleventh International Conference on Learning Representations, 2022. 2 +[44] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2 +[45] Weihao Yuan, Xiaodong Gu, Heng Li, Zilong Dong, and Siyu Zhu. Monocular scene reconstruction with 3d sdf transformers. In The Eleventh International Conference on Learning Representations, 2022. 2 +[46] Yunpeng Zhang, Phigent Robotics, and Zheng Zhu. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9433-9443, 2023. 1, 2 \ No newline at end of file diff --git a/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/images.zip b/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ef70815b28bb79083c9251d5e43b3a0a9d5085df --- /dev/null +++ b/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f252b4f6d5a9490c55ce91938d3ba0148a9a16b977770305c42b087103a68782 +size 699843 diff --git a/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/layout.json b/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..eb713f5e948413dfe7ac4b59aed1c1bd6d6b2303 --- /dev/null +++ b/2024/SurroundSDF_ Implicit 3D Scene Understanding Based on Signed Distance Field/layout.json @@ -0,0 +1,10342 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 545, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 545, + 122 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 545, + 122 + ], + "type": "text", + "content": "SurroundSDF: Implicit 3D Scene Understanding Based on Signed Distance Field" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "spans": [ + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "text", + "content": "Lizhe Liu" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "text", + "content": " Bohua Wang" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{2*§}" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "text", + "content": " Hongwei Xie" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "text", + "content": " Daqi Liu" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "text", + "content": " Li Liu" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "text", + "content": " Zhiqiang Tian" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "text", + "content": " Kuiyuan Yang" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "text", + "content": " Bing Wang" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "text", + "content": "Xiaomi EV " + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 141, + 508, + 190 + ], + "type": "text", + "content": "School of Software Engineering, Xi'an Jiaotong University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 238, + 290, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 477 + ], + "type": "text", + "content": "Vision-centric 3D environment understanding is both vital and challenging for autonomous driving systems. Recently, object-free methods have attracted considerable attention. Such methods perceive the world by predicting the semantics of discrete voxel grids but fail to construct continuous and accurate obstacle surfaces. To this end, in this paper, we propose SurroundSDF to implicitly predict the signed distance field (SDF) and semantic field for the continuous perception from surround images. Specifically, we introduce a query-based approach and utilize SDF constrained by the Eikonal formulation to accurately describe the surfaces of obstacles. Furthermore, considering the absence of precise SDF ground truth, we propose a novel weakly supervised paradigm for SDF, referred to as the Sandwich Eikonal formulation, which emphasizes applying correct and dense constraints on both sides of the surface, thereby enhancing the perceptual accuracy of the surface. Experiments suggest that our method achieves SOTA for both occupancy prediction and 3D scene reconstruction tasks on the nuScenes dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 500, + 128, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 500, + 128, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 500, + 128, + 511 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 520, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 628 + ], + "type": "text", + "content": "With the recent advancement of 3D object detection algorithms [10, 11, 16, 18, 21, 24], vision-centric autonomous driving system become more practicable. Nevertheless, the persisting challenges related to the long-tail problem and the coarse depiction of the 3D scene underscore its insufficiency. Consequently, a deeper comprehension of 3D geometry and semantics is needed for safety and reliability. This paper delves into a novel vision-centric paradigm of dense and continuous 3D scene understanding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 628, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 676 + ], + "type": "text", + "content": "Current approaches of dense 3D prediction can be classified into two categories, 3D reconstruction and 3D perception. Specifically, 3D reconstruction algorithms [6, 33] generate dense point clouds enriched with semantic informa" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 214, + 414, + 285 + ], + "blocks": [ + { + "bbox": [ + 322, + 214, + 414, + 285 + ], + "lines": [ + { + "bbox": [ + 322, + 214, + 414, + 285 + ], + "spans": [ + { + "bbox": [ + 322, + 214, + 414, + 285 + ], + "type": "image", + "image_path": "610617bf759b236a5682727c123852d1071a6c659b5b9d9eb314eb7248ad29ab.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 285, + 387, + 294 + ], + "lines": [ + { + "bbox": [ + 353, + 285, + 387, + 294 + ], + "spans": [ + { + "bbox": [ + 353, + 285, + 387, + 294 + ], + "type": "text", + "content": "a. SIREN" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 420, + 213, + 527, + 286 + ], + "blocks": [ + { + "bbox": [ + 420, + 213, + 527, + 286 + ], + "lines": [ + { + "bbox": [ + 420, + 213, + 527, + 286 + ], + "spans": [ + { + "bbox": [ + 420, + 213, + 527, + 286 + ], + "type": "image", + "image_path": "c857769f95962bfe4c30e734d2055cba2c8c6a78a6c87fe0c1fb3734e4b01bfa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 286, + 503, + 294 + ], + "lines": [ + { + "bbox": [ + 470, + 286, + 503, + 294 + ], + "spans": [ + { + "bbox": [ + 470, + 286, + 503, + 294 + ], + "type": "text", + "content": "b. LODE" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 322, + 297, + 420, + 367 + ], + "blocks": [ + { + "bbox": [ + 322, + 297, + 420, + 367 + ], + "lines": [ + { + "bbox": [ + 322, + 297, + 420, + 367 + ], + "spans": [ + { + "bbox": [ + 322, + 297, + 420, + 367 + ], + "type": "image", + "image_path": "77fc15699dcad5e9666cf396005dd220778f551e41a701e1f94cefaa21596f1f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 367, + 383, + 375 + ], + "lines": [ + { + "bbox": [ + 358, + 367, + 383, + 375 + ], + "spans": [ + { + "bbox": [ + 358, + 367, + 383, + 375 + ], + "type": "text", + "content": "c. Ours" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 380, + 545, + 422 + ], + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 422 + ], + "type": "text", + "content": "Figure 1. The scene perspective results from surround images input. a, b. The result of SIREN and LODE supervision with camera-only input (the original methods are point cloud-based). c. Our result. d. Our result with semantics." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 432, + 296, + 531, + 365 + ], + "blocks": [ + { + "bbox": [ + 432, + 296, + 531, + 365 + ], + "lines": [ + { + "bbox": [ + 432, + 296, + 531, + 365 + ], + "spans": [ + { + "bbox": [ + 432, + 296, + 531, + 365 + ], + "type": "image", + "image_path": "6c9f094630e34f0067725fc7641387fb142eec39ca21291330ed3000111c7b63.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 450, + 366, + 523, + 375 + ], + "lines": [ + { + "bbox": [ + 450, + 366, + 523, + 375 + ], + "spans": [ + { + "bbox": [ + 450, + 366, + 523, + 375 + ], + "type": "text", + "content": "d. Ours with semantic" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "text", + "content": "tion by projecting depth maps and 2D semantic maps into 3D space. 3D perception methods [3, 12, 17, 25, 35, 41, 46] predict the occupied status and the semantic of 3D voxel grids. However, either method proves to be redundant or can only predict coarse-grained discrete grids." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 486, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 605 + ], + "type": "text", + "content": "Instead, we develop a vision-centric framework to describe 3D scenes leveraging neural implicit Signed Distance Function (SDF) representation, which we refer to as SurroundSDF. Generally, we aim to: (1) Continuously describe the 3D scene by reconstructing smooth surfaces; (2) Explore the difficulty associated with the utilization of SDF representation and pose an appropriate training strategy; (3) Exploit the strong representation to kill two birds with one stone, addressing 3D semantic segmentation and continuous 3D geometry reconstruction within one framework." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 606, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 714 + ], + "type": "text", + "content": "In pursuit of the first and the third objectives, we employ SDF to represent the canonical 3D scene in terms of distance from the surface. Furthermore, we construct an implicit field that encompasses both semantics and geometry by exploiting this strong representation. However, it's still nontrivial to accomplish the SDF modeling, where complete surface and normal vectors, which are crucial supervisions, cannot be accurately computed. To this end, SIREN [34] proposed a weakly supervised implicit per" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 683, + 128, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 683, + 128, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 683, + 128, + 693 + ], + "type": "text", + "content": "* Equal Contribution." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 59, + 693, + 212, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 693, + 212, + 703 + ], + "spans": [ + { + "bbox": [ + 59, + 693, + 212, + 703 + ], + "type": "inline_equation", + "content": "^{\\S}" + }, + { + "bbox": [ + 59, + 693, + 212, + 703 + ], + "type": "text", + "content": "Work done during an internship at Xiaomi EV." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "type": "text", + "content": "† Corresponding Author." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21614" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "ception method, which accomplishes 3D mesh completion from point clouds. While LODE [14] proposed to utilize occupancy ground truth (GT) to supervise SDF. However, these methods rely on the input of point clouds, and only alleviate this difficulty by mimicking the surface distance but are either sparse or inaccurate." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 288, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 275 + ], + "type": "text", + "content": "To overcome the aforementioned challenges, we propose the Sandwich Eikonal formulation, a novel weak supervision paradigm for SDF modeling. Figure 1 gives some reconstruction results and shows the benefit of our method. This method emphasizes applying correct and dense constraints on both sides of the surface to enhance the geometric accuracy and continuity of the surface. Moreover, we revisit current pipelines for training the perception branch and design a novel loss that enhances the integration of geometry and semantics thereby reducing inconsistencies. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 277, + 286, + 433 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 277, + 286, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 286, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 286, + 324 + ], + "type": "text", + "content": "- We propose a vision-centric implicit semantic SDF perception method, achieving accurate and continuous 3D perception. To the best of our knowledge, we are the first to utilize SDF for surround-view 3D perception." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 325, + 286, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 325, + 286, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 325, + 286, + 348 + ], + "type": "text", + "content": "- We introduce the Sandwich Eikonal formulation, a novel weak supervision paradigm for SDF." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 350, + 286, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 286, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 286, + 383 + ], + "type": "text", + "content": "- We demonstrate how to employ this representation for the reconstruction of continuous 3D geometry with precise semantic information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 384, + 286, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 286, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 286, + 433 + ], + "type": "text", + "content": "- We achieve state-of-the-art results on 3D dense semantics perception tasks. Comprehensive experiments on the NuScenes dataset [2] provide extensive validation of our approach." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 442, + 133, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 442, + 133, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 133, + 453 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 462, + 286, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 286, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 286, + 569 + ], + "type": "text", + "content": "Occupancy Prediction Recently, researches [3, 12, 17, 25, 32, 32, 35, 41, 46] on occupancy prediction have demonstrated advantages in 3D scene understanding. Compared to the traditional object detection paradigm [5, 10, 11, 15, 16, 18, 21, 22, 24, 39, 42], occupancy perception has the following advantages. First, it can express dense 3D geometry. Second, it can accurately provide spatial locations for objects beyond predefined categories. Third, it can describe the shapes of irregular obstacles." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": "Based on these advantages, occupancy prediction tasks have attracted significant research. These methods predict the occupancy status in the region of interest around the ego vehicle from point clouds or images. Specifically, the space is first divided into voxel grids at a specific resolution. Then the occupancy status and semantics of each grid are estimated. SurroundOcc [41] proposed a surround-view 3D occupancy perception method that utilizes spatial 2D-3D attention to lift image features into 3D space. In addition, to realize the dense occupancy prediction, SurroundOcc designed a pipeline to convert the point cloud to dense occupancy ground truth. VoxFormer [17] employed" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "an MAE-like [9] approach to achieve camera-based semantic occupancy prediction. FB-OCC [20, 20] proposed a novel forward-backward projection method to compensate for the insufficient BEV feature density of the forward projection method and the large number of mismatches in 2D and 3D space caused by the backward projection. Despite the impressive results, they are limited by the specific resolution of the occupancy annotations, which limit the continuous perception of the scene." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 182, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 506 + ], + "type": "text", + "content": "Implicit Scene Perception Scene reconstruction refers to the task of predicting the 3D geometry structures from some incomplete representations, e.g., images, point clouds, and voxel grids. Implicit neural scene reconstruction methods [1, 4, 14, 26, 30, 34, 45] have demonstrated advantages in accuracy and substantial potential. Generally, they train a neural network to predict a continuous field. Thus, it is possible to query occupancy information for any point within the 3D space. DeepSDF [30] utilizes the SDF to achieve the implicit reconstruction of 3D objects at the instance level. NeRFs [26] and its numerous variants [4, 7, 43] achieve implicit 3D scene reconstruction and novel view synthesis from multi-view images. NeuS [38] and VolSDF [44] introduce the signed distance function into neural rendering, enabling more precise object surface reconstruction. SIREN [34] utilizes the Eikonal equation to achieve semi-supervised SDF reconstruction for point clouds. However, these methods lack generalization to novel scenes. To achieve online implicit scene perception in driving scenes, LODE [14] proposes the Locally Conditioned Eikonal formulation and introduces a dense occupancy ground truth for supervision, which significantly improves the effect of SDF reconstruction in the driving scenes. However, it is limited by the sparsity of input LiDAR points or inaccurate ground truth. To solve these problems, we propose a camera-only implicit scene understanding method, which aims at continuous and accurate surface perception." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 511, + 385, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 511, + 385, + 523 + ], + "spans": [ + { + "bbox": [ + 306, + 511, + 385, + 523 + ], + "type": "text", + "content": "3. Formulation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 529, + 545, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 565 + ], + "type": "text", + "content": "This section analyzes the SDF supervision paradigm based on the Eikonal formulation with its variations and introduces our Sandwich Eikonal supervision approach." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 568, + 476, + 579 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 568, + 476, + 579 + ], + "spans": [ + { + "bbox": [ + 306, + 568, + 476, + 579 + ], + "type": "text", + "content": "3.1. Eikonal-based SDF Constraints" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 583, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 583, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 583, + 545, + 619 + ], + "type": "text", + "content": "Given a coordinate " + }, + { + "bbox": [ + 304, + 583, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 583, + 545, + 619 + ], + "type": "text", + "content": " in the 3D scene, our goal is to construct a function " + }, + { + "bbox": [ + 304, + 583, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 583, + 545, + 619 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 304, + 583, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\phi(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 583, + 545, + 619 + ], + "type": "text", + "content": " provides the SDF value at point " + }, + { + "bbox": [ + 304, + 583, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 583, + 545, + 619 + ], + "type": "text", + "content": ". The Eikonal-based optimization objectives [14] is" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 347, + 620, + 545, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 620, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 347, + 620, + 545, + 634 + ], + "type": "interline_equation", + "content": "\\mathcal {O} _ {E} = \\int_ {\\Omega_ {0}} O _ {0} d \\mathbf {x} + \\int_ {\\Omega_ {1}} O _ {1} d \\mathbf {x} + \\int_ {\\Omega_ {2}} O _ {2} d \\mathbf {x}, \\tag {1}", + "image_path": "39536ba27dcfd67cd68971d0675d658d396de9ddb7f11af3afd5b97dceabd5e5.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 639, + 323, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 323, + 648 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 323, + 648 + ], + "type": "text", + "content": "and" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 354, + 647, + 545, + 688 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 647, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 354, + 647, + 545, + 688 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} O _ {0} = | | \\nabla_ {\\mathbf {x}} \\phi (\\mathbf {x}) | - 1 |, & \\mathbf {x} \\in \\Omega_ {0} \\\\ O _ {1} = | \\nabla_ {\\mathbf {x}} \\phi (\\mathbf {x}) - \\mathbf {n} (\\mathbf {x}) |, & \\mathbf {x} \\in \\Omega_ {1}, \\\\ O _ {2} = | \\phi (\\mathbf {x}) - S D F (\\mathbf {x}) |, & \\mathbf {x} \\in \\Omega_ {2} \\end{array} \\right. \\tag {2}", + "image_path": "7349d42d212e43a3a61856efb59d22c5ca90c8d6ac33f10cb58ab014d82a4b1a.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Omega_0" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " is the set of the whole 3D space of interest, " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Omega_1" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " is the set of points on the surface, " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Omega_2" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " is the set of points" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21615" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 75, + 180, + 156 + ], + "blocks": [ + { + "bbox": [ + 99, + 75, + 180, + 156 + ], + "lines": [ + { + "bbox": [ + 99, + 75, + 180, + 156 + ], + "spans": [ + { + "bbox": [ + 99, + 75, + 180, + 156 + ], + "type": "image", + "image_path": "88e186e3c3371d5010599239bd61c556c07e624a95a5fd5bd6be5a9e4ee3e907.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 159, + 169, + 168 + ], + "lines": [ + { + "bbox": [ + 115, + 159, + 169, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 159, + 169, + 168 + ], + "type": "text", + "content": "a. Ideal supervision" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 76, + 262, + 156 + ], + "blocks": [ + { + "bbox": [ + 181, + 76, + 262, + 156 + ], + "lines": [ + { + "bbox": [ + 181, + 76, + 262, + 156 + ], + "spans": [ + { + "bbox": [ + 181, + 76, + 262, + 156 + ], + "type": "image", + "image_path": "b970adbac643898fd614ba27df17c68bef3bf71c1d87bbc0c586eeba0076deee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 159, + 236, + 167 + ], + "lines": [ + { + "bbox": [ + 208, + 159, + 236, + 167 + ], + "spans": [ + { + "bbox": [ + 208, + 159, + 236, + 167 + ], + "type": "text", + "content": "b. SIREN" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 171, + 547, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 547, + 205 + ], + "type": "text", + "content": "Figure 2. a. Ideal supervision for SDF where the SDF GT is provided. b. SIREN supervision which uses LiDAR points to supervise the surface. c. LODE supervision based on occupancy GT. d. Our supervision paradigm which combines the LiDAR GT and occupancy GT, it is closer to the ideal supervision." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 263, + 76, + 345, + 156 + ], + "blocks": [ + { + "bbox": [ + 263, + 76, + 345, + 156 + ], + "lines": [ + { + "bbox": [ + 263, + 76, + 345, + 156 + ], + "spans": [ + { + "bbox": [ + 263, + 76, + 345, + 156 + ], + "type": "image", + "image_path": "3cca8fda26a26d66195b1b41dfa91bbc136a3bb0ab2eca56ce7bb3b06f1afb5c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 292, + 159, + 317, + 167 + ], + "lines": [ + { + "bbox": [ + 292, + 159, + 317, + 167 + ], + "spans": [ + { + "bbox": [ + 292, + 159, + 317, + 167 + ], + "type": "text", + "content": "c.LODE" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 347, + 76, + 429, + 156 + ], + "blocks": [ + { + "bbox": [ + 347, + 76, + 429, + 156 + ], + "lines": [ + { + "bbox": [ + 347, + 76, + 429, + 156 + ], + "spans": [ + { + "bbox": [ + 347, + 76, + 429, + 156 + ], + "type": "image", + "image_path": "d129ffcfe9fa1e392b0941b20461847b9119f2cf7090edb61335d72f288fc181.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 159, + 400, + 167 + ], + "lines": [ + { + "bbox": [ + 378, + 159, + 400, + 167 + ], + "spans": [ + { + "bbox": [ + 378, + 159, + 400, + 167 + ], + "type": "text", + "content": "d. Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 78, + 495, + 86 + ], + "lines": [ + { + "bbox": [ + 430, + 78, + 495, + 86 + ], + "spans": [ + { + "bbox": [ + 430, + 78, + 495, + 86 + ], + "type": "text", + "content": "Constrain " + }, + { + "bbox": [ + 430, + 78, + 495, + 86 + ], + "type": "inline_equation", + "content": "SDF < 0" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 431, + 92, + 493, + 99 + ], + "lines": [ + { + "bbox": [ + 431, + 92, + 493, + 99 + ], + "spans": [ + { + "bbox": [ + 431, + 92, + 493, + 99 + ], + "type": "text", + "content": "Constrain " + }, + { + "bbox": [ + 431, + 92, + 493, + 99 + ], + "type": "inline_equation", + "content": "SDF = 0" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 431, + 105, + 493, + 112 + ], + "lines": [ + { + "bbox": [ + 431, + 105, + 493, + 112 + ], + "spans": [ + { + "bbox": [ + 431, + 105, + 493, + 112 + ], + "type": "text", + "content": "Constrain " + }, + { + "bbox": [ + 431, + 105, + 493, + 112 + ], + "type": "inline_equation", + "content": "SDF > 0" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 431, + 117, + 495, + 123 + ], + "lines": [ + { + "bbox": [ + 431, + 117, + 495, + 123 + ], + "spans": [ + { + "bbox": [ + 431, + 117, + 495, + 123 + ], + "type": "text", + "content": "Constraint min SDF" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 431, + 123, + 484, + 130 + ], + "lines": [ + { + "bbox": [ + 431, + 123, + 484, + 130 + ], + "spans": [ + { + "bbox": [ + 431, + 123, + 484, + 130 + ], + "type": "text", + "content": "of one grid " + }, + { + "bbox": [ + 431, + 123, + 484, + 130 + ], + "type": "inline_equation", + "content": "< 0" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 431, + 133, + 468, + 140 + ], + "lines": [ + { + "bbox": [ + 431, + 133, + 468, + 140 + ], + "spans": [ + { + "bbox": [ + 431, + 133, + 468, + 140 + ], + "type": "text", + "content": "Surface" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 431, + 144, + 484, + 152 + ], + "lines": [ + { + "bbox": [ + 431, + 144, + 484, + 152 + ], + "spans": [ + { + "bbox": [ + 431, + 144, + 484, + 152 + ], + "type": "text", + "content": "LiDAR Points" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 431, + 154, + 474, + 163 + ], + "lines": [ + { + "bbox": [ + 431, + 154, + 474, + 163 + ], + "spans": [ + { + "bbox": [ + 431, + 154, + 474, + 163 + ], + "type": "text", + "content": "Voxel Grid" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 212, + 287, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 212, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 287, + 295 + ], + "type": "text", + "content": "with SDF GT annotations, " + }, + { + "bbox": [ + 46, + 212, + 287, + 295 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}}\\phi (\\mathbf{x})" + }, + { + "bbox": [ + 46, + 212, + 287, + 295 + ], + "type": "text", + "content": " is the gradient on point " + }, + { + "bbox": [ + 46, + 212, + 287, + 295 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 212, + 287, + 295 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 212, + 287, + 295 + ], + "type": "inline_equation", + "content": "\\mathbf{n}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 212, + 287, + 295 + ], + "type": "text", + "content": " is the normal, and " + }, + { + "bbox": [ + 46, + 212, + 287, + 295 + ], + "type": "inline_equation", + "content": "SDF(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 212, + 287, + 295 + ], + "type": "text", + "content": " is the real SDF value at point " + }, + { + "bbox": [ + 46, + 212, + 287, + 295 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 212, + 287, + 295 + ], + "type": "text", + "content": ". One major obstacle to this optimization objective lies in the design of " + }, + { + "bbox": [ + 46, + 212, + 287, + 295 + ], + "type": "inline_equation", + "content": "O_2" + }, + { + "bbox": [ + 46, + 212, + 287, + 295 + ], + "type": "text", + "content": ", which provides direct supervision for SDF as shown in Figure 2.a. This objective requires the annotation of SDF values for each anchor point in the scene, which is monumental and challenging." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 296, + 287, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 296, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 47, + 296, + 287, + 332 + ], + "type": "text", + "content": "In the absence of precise SDF annotations, SIREN [34] and LODE [14] have implemented weakly supervised paradigm through the following variations of " + }, + { + "bbox": [ + 47, + 296, + 287, + 332 + ], + "type": "inline_equation", + "content": "O_2" + }, + { + "bbox": [ + 47, + 296, + 287, + 332 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 108, + 336, + 287, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 336, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 108, + 336, + 287, + 364 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} O _ {2 - 0} ^ {\\prime} = | \\phi (\\mathbf {x}) |, & \\mathbf {x} \\in \\Omega_ {1} \\\\ O _ {2 - 1} ^ {\\prime} = \\psi (| \\phi (\\mathbf {x}) |), & \\mathbf {x} \\in \\Omega_ {3}, \\end{array} \\right. \\tag {3}", + "image_path": "065202434ec70f46a56320da6a4b4d32692ec4687cb6af42f9c7b0285e5baee1.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 365, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 365, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\psi (\\cdot)" + }, + { + "bbox": [ + 46, + 365, + 287, + 389 + ], + "type": "text", + "content": " is a monotonically decreasing function and " + }, + { + "bbox": [ + 46, + 365, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\Omega_3\\subseteq \\Omega_0 / \\Omega_1" + }, + { + "bbox": [ + 46, + 365, + 287, + 389 + ], + "type": "text", + "content": " is the whole space except the surface." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 389, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 460 + ], + "type": "text", + "content": "The objective " + }, + { + "bbox": [ + 46, + 389, + 287, + 460 + ], + "type": "inline_equation", + "content": "O_{2-0}^{\\prime}" + }, + { + "bbox": [ + 46, + 389, + 287, + 460 + ], + "type": "text", + "content": " aims to constrain the SDF values on the surface to 0. Considering the absence of precise supervision, " + }, + { + "bbox": [ + 46, + 389, + 287, + 460 + ], + "type": "inline_equation", + "content": "O_{2-1}^{\\prime}" + }, + { + "bbox": [ + 46, + 389, + 287, + 460 + ], + "type": "text", + "content": " adopts a fuzzy constraint that pushes the absolute SDF value away from zero in " + }, + { + "bbox": [ + 46, + 389, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\Omega_3" + }, + { + "bbox": [ + 46, + 389, + 287, + 460 + ], + "type": "text", + "content": ". While the constraint in " + }, + { + "bbox": [ + 46, + 389, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\Omega_3" + }, + { + "bbox": [ + 46, + 389, + 287, + 460 + ], + "type": "text", + "content": " is loose, it is adequate to capture the geometry information of the surface." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 461, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 640 + ], + "type": "text", + "content": "In practical implementation, SIREN [34] samples on-surface LiDAR points in " + }, + { + "bbox": [ + 46, + 461, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\Omega_{1}" + }, + { + "bbox": [ + 46, + 461, + 287, + 640 + ], + "type": "text", + "content": " and considers the space excluding the LiDAR points as " + }, + { + "bbox": [ + 46, + 461, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\Omega_{3}" + }, + { + "bbox": [ + 46, + 461, + 287, + 640 + ], + "type": "text", + "content": ", as is shown in Figure 2.b. When the LiDAR points are dense enough, this strategy can effectively achieve good constraint effects [34]. However, LODE [14] has demonstrated that the sparsity of LiDAR points in the driving scenes results in discontinuities of the estimated mesh. To achieve denser supervision, they use the dense occupancy GT as supervision. This approach treats both the surface and interior of the obstacles as the " + }, + { + "bbox": [ + 46, + 461, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\Omega_{1}" + }, + { + "bbox": [ + 46, + 461, + 287, + 640 + ], + "type": "text", + "content": ". Specifically, they sample the " + }, + { + "bbox": [ + 46, + 461, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\Omega_{1}" + }, + { + "bbox": [ + 46, + 461, + 287, + 640 + ], + "type": "text", + "content": " points as the center of \"occupied\" grids, and the " + }, + { + "bbox": [ + 46, + 461, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\Omega_{3}" + }, + { + "bbox": [ + 46, + 461, + 287, + 640 + ], + "type": "text", + "content": " points as the center of \"free\" grids, as is shown in 2.c. However, this paradigm remains imprecise, as considering the center of the occupied grid to be the surface of objects would introduce errors." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 647, + 216, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 216, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 216, + 658 + ], + "type": "text", + "content": "3.2. Sandwich Eikonal Formulation" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "To address the above concerns, we require a supervision paradigm that has the following characteristics: 1) sampling points as the supervision on the surface must strictly adhere to the surface, 2) for regions beyond the surface, refrain" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 212, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 212, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 212, + 545, + 247 + ], + "type": "text", + "content": "from supervision in uncertain areas to ensure precision, and 3) dense supervision signals. Considering the above factors, we introduce our variants of Eikonal formulation as follows:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 349, + 258, + 545, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 258, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 349, + 258, + 545, + 300 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} O _ {2 - 0} = \\phi (\\mathbf {x}), & \\mathbf {x} \\in \\Omega_ {1} \\\\ O _ {2 - 1} = \\psi (- \\min (\\phi (\\mathbf {x}))), & \\mathbf {x} \\in \\Omega_ {\\text {o c c}} ^ {i} \\\\ O _ {2 - 2} = \\psi (\\phi (\\mathbf {x})), & \\mathbf {x} \\in \\Omega_ {\\text {f r e e}} ^ {j}, \\end{array} \\right. \\tag {4}", + "image_path": "dc6f81835cce4092e4897aedd2377cf81af8e3e6bf1ba40bdbfcd3b04f59e601.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 310, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 310, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 310, + 545, + 335 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 310, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\Omega_{occ}^{i}" + }, + { + "bbox": [ + 305, + 310, + 545, + 335 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 305, + 310, + 545, + 335 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 310, + 545, + 335 + ], + "type": "text", + "content": "-th grid of occupied grids and " + }, + { + "bbox": [ + 305, + 310, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\Omega_{free}^{j}" + }, + { + "bbox": [ + 305, + 310, + 545, + 335 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 305, + 310, + 545, + 335 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 310, + 545, + 335 + ], + "type": "text", + "content": "-th grid of free grids." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 337, + 546, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 546, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 546, + 613 + ], + "type": "text", + "content": "To this end, we incorporate the LiDAR points and the occupancy GT into the SDF supervision. As is shown in Figure 2.d, for precise sampling points on the surface, we follow SIREN and constrain the SDF value of LiDAR points to be 0. To apply accurate and dense supervision to regions beyond the surface, we introduce the occupancy GT into the supervision, as is shown in optimization objectives " + }, + { + "bbox": [ + 304, + 337, + 546, + 613 + ], + "type": "inline_equation", + "content": "O_{2-1}" + }, + { + "bbox": [ + 304, + 337, + 546, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 337, + 546, + 613 + ], + "type": "inline_equation", + "content": "O_{2-2}" + }, + { + "bbox": [ + 304, + 337, + 546, + 613 + ], + "type": "text", + "content": ". According to the generation process of occupancy GT [35, 36], if a grid is labeled as \"occupied\", it implies that at least part of the grid area falls inside the object. Therefore, as the objective " + }, + { + "bbox": [ + 304, + 337, + 546, + 613 + ], + "type": "inline_equation", + "content": "O_{2-1}" + }, + { + "bbox": [ + 304, + 337, + 546, + 613 + ], + "type": "text", + "content": " shows, we constrain the minimum SDF value of each occupied grid to be less than zero and push it to the negative range. On the other hand, if a grid is labeled as \"free\", the entire grid is outside the object and we push the SDF value of free grids to the positive range, as is shown in objective " + }, + { + "bbox": [ + 304, + 337, + 546, + 613 + ], + "type": "inline_equation", + "content": "O_{2-2}" + }, + { + "bbox": [ + 304, + 337, + 546, + 613 + ], + "type": "text", + "content": ". We employ a multiframe point cloud fusion strategy to densify the sampling for surface supervision, as described in Section 4.2. Compared to previous supervision paradigms, our formulation emphasizes applying appropriate and dense supervision on both sides of the surface to enhance the geometric constraint based on the Eikonal formulation. Therefore, we name our supervision strategy the \"Sandwich Eikonal\" formulation." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 631, + 361, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 631, + 361, + 643 + ], + "spans": [ + { + "bbox": [ + 306, + 631, + 361, + 643 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "content": "Given multi-camera images " + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = \\{I_1, I_2, \\dots, I_n\\}" + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "content": ", we aim to predict the 3D SDF " + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\phi(\\mathbf{x}) : \\mathbb{R}^3 \\Rightarrow \\mathbb{R}" + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "content": " and semantic field " + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x}) : \\mathbb{R}^3 \\Rightarrow \\mathbb{R}^s" + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "content": " indicates the number of classes. The following will introduce our architecture and detailed constraints based on our Sandwich Eikonal formulation." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21616" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 71, + 504, + 236 + ], + "blocks": [ + { + "bbox": [ + 91, + 71, + 504, + 236 + ], + "lines": [ + { + "bbox": [ + 91, + 71, + 504, + 236 + ], + "spans": [ + { + "bbox": [ + 91, + 71, + 504, + 236 + ], + "type": "image", + "image_path": "ef8d1bbee8318de1f774218ef7b57ff1e49655ead98b147e8d2ca998b827fe49.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 547, + 318 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 547, + 318 + ], + "type": "text", + "content": "Figure 3. The architecture of our SurroundSDF. Given the surround images as input, an encoder composed of a 2D backbone, LSS module, and BEV backbone, is employed to extract voxel features. We adopt a query-based approach to sample features from the voxel features. Specifically, first, a set of query coordinates in the region of interest is selected. Subsequently, using trilinear interpolation, semantic features are queried from the voxel features. Finally, after concatenation with the positional embeddings from the query coordinates, the features pass through the SDF head and semantic head respectively, yielding SDF and semantic fields. For training, the query coordinates are sampled according to the GT, and the SDF and semantic field are supervised by the losses introduced in Section 4.2. In the inference phase, based on appropriate sampling and post-processing, continuous and accurate scene perception results are obtained (see Section 4.3)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 323, + 130, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 130, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 130, + 335 + ], + "type": "text", + "content": "4.1. Architecture" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 341, + 288, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 288, + 449 + ], + "type": "text", + "content": "The overall architecture is shown in Figure 3. First, a CNN-based backbone is used to obtain the image features. Then, the LSS [31] is used to project the image features to BEV space, and a BEV encoder is adopted to obtain the voxel features " + }, + { + "bbox": [ + 46, + 341, + 288, + 449 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 341, + 288, + 449 + ], + "type": "text", + "content": " of dimension " + }, + { + "bbox": [ + 46, + 341, + 288, + 449 + ], + "type": "inline_equation", + "content": "h \\times w \\times d \\times c" + }, + { + "bbox": [ + 46, + 341, + 288, + 449 + ], + "type": "text", + "content": " (c indicates the number of channels). Further, for each point " + }, + { + "bbox": [ + 46, + 341, + 288, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 341, + 288, + 449 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 341, + 288, + 449 + ], + "type": "inline_equation", + "content": "\\Omega_0" + }, + { + "bbox": [ + 46, + 341, + 288, + 449 + ], + "type": "text", + "content": ", the corresponding features can be obtained by querying " + }, + { + "bbox": [ + 46, + 341, + 288, + 449 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 341, + 288, + 449 + ], + "type": "text", + "content": " with the trilinear sampling method. Therefore, " + }, + { + "bbox": [ + 46, + 341, + 288, + 449 + ], + "type": "inline_equation", + "content": "\\phi(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 341, + 288, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 341, + 288, + 449 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 341, + 288, + 449 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 455, + 287, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 455, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 75, + 455, + 287, + 474 + ], + "type": "interline_equation", + "content": "\\phi (\\mathbf {x}) = H _ {s d f} \\left(C \\left(Q (V, \\mathbf {x}), P E (\\mathbf {x})\\right), \\theta\\right), \\mathbf {x} \\in \\Omega_ {0}, \\tag {5}", + "image_path": "85fd738415896bbd50980925867aa4a3f908c9feeced5a313b7a893b5c0f6f22.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 72, + 470, + 264, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 470, + 264, + 483 + ], + "spans": [ + { + "bbox": [ + 72, + 470, + 264, + 483 + ], + "type": "interline_equation", + "content": "S (\\mathbf {x}) = H _ {s e m} (C (Q (V, \\mathbf {x}), P E (\\mathbf {x})), \\theta), \\mathbf {x} \\in \\Omega_ {0},", + "image_path": "08fa552fab4b07800246e0979f0374379f99a75d4ffe2dc9a9562c9f861cbbbb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "text", + "content": " represent the concatenation and query process, respectively. " + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "inline_equation", + "content": "H_{sdf}" + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "inline_equation", + "content": "H_{sem}" + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "text", + "content": " indicate the SDF head and semantic head, implemented with three-layer MLPs with sine activation functions. " + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "text", + "content": " represents the learnable parameters. " + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "inline_equation", + "content": "PE" + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "text", + "content": " represents the positional encoding, which aims to capture high-frequency information. Specifically, for a coordinate " + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "inline_equation", + "content": "PE" + }, + { + "bbox": [ + 46, + 491, + 287, + 588 + ], + "type": "text", + "content": " encodes each of its dimensions into a 2n-dimensional vector, it can be expressed" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 590, + 287, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 287, + 604 + ], + "type": "interline_equation", + "content": "\\text {a s :} \\quad P E (\\mathbf {x}) = (\\zeta (x), \\zeta (y), \\zeta (z)) \\in \\mathbb {R} ^ {6 n}, \\tag {6}", + "image_path": "850066fb7a67fc772d5b3a29048e991f9ed9f43230148570bd146d187ae95fab.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 606, + 287, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 287, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 606, + 287, + 630 + ], + "type": "inline_equation", + "content": "x, y, z" + }, + { + "bbox": [ + 46, + 606, + 287, + 630 + ], + "type": "text", + "content": " denote the coordinates, " + }, + { + "bbox": [ + 46, + 606, + 287, + 630 + ], + "type": "inline_equation", + "content": "\\zeta(\\cdot)" + }, + { + "bbox": [ + 46, + 606, + 287, + 630 + ], + "type": "text", + "content": " represents an encoding scheme using trigonometric function mapping:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 634, + 287, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 634, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 58, + 634, + 287, + 649 + ], + "type": "interline_equation", + "content": "\\zeta (x) = \\left(\\sin \\left(2 ^ {0} \\pi x\\right), \\cos \\left(2 ^ {0} \\pi x\\right), \\dots , \\cos \\left(2 ^ {n - 1} \\pi x\\right)\\right). \\tag {7}", + "image_path": "c730caf23699fdf409528691d5c735b6d992a5013fd7be2a2c73fdf336bb180e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": "In training, we design a joint supervision method that utilizes both coarse-grained voxel and fine-grained point cloud supervision based on our Sandwich Eikonal formulation. In inference, we query the voxel features to estimate the SDF values and semantic logits of each query point." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 323, + 384, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 323, + 384, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 384, + 335 + ], + "type": "text", + "content": "4.2. Supervision" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 341, + 546, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 546, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 546, + 378 + ], + "type": "text", + "content": "This section will elaborate on deriving the SDF loss from our Sandwich Eikonal formulation. Then, we will introduce the semantic loss and our joint supervision strategy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 383, + 546, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 546, + 456 + ], + "type": "text", + "content": "SDF Supervision Following objectives " + }, + { + "bbox": [ + 304, + 383, + 546, + 456 + ], + "type": "inline_equation", + "content": "O_0" + }, + { + "bbox": [ + 304, + 383, + 546, + 456 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 383, + 546, + 456 + ], + "type": "inline_equation", + "content": "O_1" + }, + { + "bbox": [ + 304, + 383, + 546, + 456 + ], + "type": "text", + "content": " in Equation 2 and " + }, + { + "bbox": [ + 304, + 383, + 546, + 456 + ], + "type": "inline_equation", + "content": "O_{2-0}" + }, + { + "bbox": [ + 304, + 383, + 546, + 456 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 383, + 546, + 456 + ], + "type": "inline_equation", + "content": "O_{2-2}" + }, + { + "bbox": [ + 304, + 383, + 546, + 456 + ], + "type": "text", + "content": " in Equation 4, we derive the losses for SDF supervision, based on our Sandwich Eikonal formulation. Considering that continuous space supervision is unfeasible, we sample the discrete points in the corresponding space " + }, + { + "bbox": [ + 304, + 383, + 546, + 456 + ], + "type": "inline_equation", + "content": "(\\Omega_0, \\Omega_1, \\Omega_{occ}" + }, + { + "bbox": [ + 304, + 383, + 546, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 383, + 546, + 456 + ], + "type": "inline_equation", + "content": "\\Omega_{free})" + }, + { + "bbox": [ + 304, + 383, + 546, + 456 + ], + "type": "text", + "content": " for supervision." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 339, + 466, + 427, + 555 + ], + "blocks": [ + { + "bbox": [ + 339, + 466, + 427, + 555 + ], + "lines": [ + { + "bbox": [ + 339, + 466, + 427, + 555 + ], + "spans": [ + { + "bbox": [ + 339, + 466, + 427, + 555 + ], + "type": "image", + "image_path": "733d3c8523d8de7ff252598d06e9c0b78cd419f865eda5194830af270b6af1e2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 557, + 378, + 564 + ], + "lines": [ + { + "bbox": [ + 370, + 557, + 378, + 564 + ], + "spans": [ + { + "bbox": [ + 370, + 557, + 378, + 564 + ], + "type": "text", + "content": "a." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 434, + 468, + 512, + 555 + ], + "blocks": [ + { + "bbox": [ + 434, + 468, + 512, + 555 + ], + "lines": [ + { + "bbox": [ + 434, + 468, + 512, + 555 + ], + "spans": [ + { + "bbox": [ + 434, + 468, + 512, + 555 + ], + "type": "image", + "image_path": "b3fcc936e39af8d4b1a7d4428e8dbe49c23fe53a8f592b113fe12d2e18738be3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 557, + 476, + 564 + ], + "lines": [ + { + "bbox": [ + 468, + 557, + 476, + 564 + ], + "spans": [ + { + "bbox": [ + 468, + 557, + 476, + 564 + ], + "type": "text", + "content": "b." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 567, + 546, + 590 + ], + "lines": [ + { + "bbox": [ + 305, + 567, + 546, + 590 + ], + "spans": [ + { + "bbox": [ + 305, + 567, + 546, + 590 + ], + "type": "text", + "content": "Figure 4. a. SDF constraints with Sandwich Eikonal formulation in continuous form. b. SDF sampling in discrete form." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "The correctness and density of the supervision sampling are of paramount importance. Since our approach is based on the surround cameras, we only sample from the camera-visible space in " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\Omega_0" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " (space of interest) [36]. For " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\Omega_{1}" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " (surface points), we sample exclusively from the LiDAR points. To enhance the sampling density, temporal point cloud fusion is adopted. Specifically, " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\Omega_{1} = \\{p_{t - k},p_{t - (k - 1)},\\dots ,p_{t + (k - 1)},p_{t + k}\\}" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "p_{t + j}" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " is the LiDAR points of the " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "t + j" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " frame and " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " is the current frame. Based on occupancy GT [36], " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\Omega_{occ}" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\Omega_{free}" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " are the sets of" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21617" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 282, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 282, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 282, + 84 + ], + "type": "text", + "content": "voxel grids labeled as \"occupied\" and \"free\", respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 287, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 287, + 180 + ], + "type": "text", + "content": "Additionally, in our objective " + }, + { + "bbox": [ + 46, + 84, + 287, + 180 + ], + "type": "inline_equation", + "content": "O_{2-1}" + }, + { + "bbox": [ + 46, + 84, + 287, + 180 + ], + "type": "text", + "content": ", the computation of the minimum SDF within the occupied grid is required. As is shown in Figure 4, we approximate this minimum value through discrete sampling. The minimum value among the SDF values at the centers of each sub-grid is regarded as the minimum SDF value for the grid. Based on Equation 2, 4 and the above sampling strategy, our SDF loss " + }, + { + "bbox": [ + 46, + 84, + 287, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sdf}" + }, + { + "bbox": [ + 46, + 84, + 287, + 180 + ], + "type": "text", + "content": " is expressed as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 182, + 287, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 182, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 74, + 182, + 287, + 331 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {s d f} = \\gamma_ {1} \\frac {1}{n _ {t o t a l}} \\sum_ {i = 1} ^ {n _ {t o t a l}} \\left(\\left| \\nabla_ {\\mathbf {x}} \\phi (\\mathbf {x} ^ {i}) \\right| - 1\\right) \\\\ + \\gamma_ {2} \\frac {1}{n _ {1 \\cup o c c}} \\sum_ {i = 1} ^ {n _ {1 \\cup o c c}} (\\nabla_ {\\mathbf {x}} \\phi (\\mathbf {x} ^ {i}) - n (\\mathbf {x} ^ {i})) \\\\ + \\gamma_ {3} \\frac {1}{n _ {1}} \\sum_ {i = 1} ^ {n _ {1}} | \\phi (\\mathbf {x} ^ {i}) | \\tag {8} \\\\ + \\gamma_ {4} \\frac {1}{n _ {o c c}} \\sum_ {i = 1} ^ {n _ {o c c}} e ^ {\\alpha \\times m i n (\\phi (\\mathbf {x} ^ {i}), \\mathbf {x} ^ {i} \\in \\Omega_ {o c c} ^ {i})} \\\\ + \\gamma_ {5} \\frac {1}{n _ {f r e e}} \\sum_ {i = 1} ^ {n _ {f r e e}} e ^ {- \\alpha \\times r a n d o m (\\phi (\\mathbf {x} ^ {i}), \\mathbf {x} ^ {i} \\in \\Omega_ {f r e e} ^ {i})}, \\\\ \\end{array}", + "image_path": "57e73dbb4ff5dcb35dc135fff4345cea440067f4086dc0575b425ee76105ab3c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "text", + "content": " is a hyperparameter that determines the extent of the deviation for the interior and exterior space of the object and we fix " + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "text", + "content": " to 100 in experiments. " + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\gamma_{1},\\gamma_{2},\\gamma_{3},\\gamma_{4},\\gamma_{5}" + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "text", + "content": " are loss weights, " + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "inline_equation", + "content": "n(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "text", + "content": " represents the normal vector of the current point on surface. " + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "inline_equation", + "content": "n_1,n_{occ}" + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "inline_equation", + "content": "n_{free}" + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "text", + "content": " represent the number of points or voxel grids sampled from " + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\Omega_1,\\Omega_{occ}" + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\Omega_{free}" + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "inline_equation", + "content": "n_{total}" + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "text", + "content": " is the total number of samples: " + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "inline_equation", + "content": "n_{total} = n_1 + n_{occ} + n_{free}" + }, + { + "bbox": [ + 46, + 334, + 287, + 430 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 441, + 287, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 465 + ], + "type": "text", + "content": "Semantic Supervision A widely used cross-entropy loss is adopted to supervise the semantic field as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 468, + 287, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 468, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 79, + 468, + 287, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s e m} = \\frac {1}{n _ {o c c}} \\sum_ {i = 1} ^ {n _ {o c c}} C E \\left(S (\\mathbf {x} | \\mathbf {x} \\in \\Omega_ {o c c} ^ {i}), \\dot {y} _ {s e m} ^ {i}\\right), \\tag {9}", + "image_path": "e2d32d873936241213923a34fd35a26ed5906edd9843074c76ee5a89ec445707.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 486, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 510 + ], + "type": "text", + "content": "where CE represents cross-entropy loss and " + }, + { + "bbox": [ + 46, + 486, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\hat{y}_{sem}^i" + }, + { + "bbox": [ + 46, + 486, + 287, + 510 + ], + "type": "text", + "content": " is the semantic ground truth of i-th voxel grid." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 521, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 677 + ], + "type": "text", + "content": "Joint Supervision The above losses optimize the SDF and semantic field separately, which leads to a nonnegligible issue of ambiguity between geometric and semantic optimization. To illustrate this issue, we train a model and evaluate the occupancy IoU representing geometric accuracy and the semantic mIoU representing overall semantic accuracy following the occupancy prediction metrics [35, 36]. For each voxel grid, if its minimum SDF is less than a certain threshold, the voxel is considered 'occupied'. As is shown in Figure 5, the optimal thresholds producing the highest peak of these curves are significantly different. This indicates that greater geometric accuracy does not necessarily lead to improved overall semantic accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "To mitigate this issue, we propose a joint supervision strategy. First, for the minimum SDF value " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\min(\\phi(\\mathbf{x}))" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " of each voxel grid, a SoftMax function is used to convert it to" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 326, + 71, + 526, + 160 + ], + "blocks": [ + { + "bbox": [ + 326, + 71, + 526, + 160 + ], + "lines": [ + { + "bbox": [ + 326, + 71, + 526, + 160 + ], + "spans": [ + { + "bbox": [ + 326, + 71, + 526, + 160 + ], + "type": "image", + "image_path": "ccb7a9fa23c8ade7c27ec8c0d8fcd7de2555818bbb85abc5be7779f509fa86de.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 163, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 305, + 163, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 305, + 163, + 547, + 196 + ], + "type": "text", + "content": "Figure 5. Variation of occupancy IoU and semantic mIoU with SDF Threshold. Note that the peak values of these two indicators correspond to different SDF thresholds." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "text", + "content": "the free probability: " + }, + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "inline_equation", + "content": "p_{free} = \\text{softmax}(\\beta t, \\beta \\min(\\phi(\\mathbf{x}))" + }, + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "text", + "content": " is fixed to 0.005, " + }, + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "text", + "content": " is fixed to 100. Second, we map the free probability " + }, + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "inline_equation", + "content": "p_{free}" + }, + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "text", + "content": " to logits using an inv-sigmoid function: " + }, + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "inline_equation", + "content": "l_{free} = \\log(p_{free} / (1 - p_{free}))" + }, + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "inline_equation", + "content": "l_{free}" + }, + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "text", + "content": " is combined with semantic logits " + }, + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "inline_equation", + "content": "l_{sem}" + }, + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "text", + "content": ", and the joint loss " + }, + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{joint}" + }, + { + "bbox": [ + 305, + 205, + 546, + 278 + ], + "type": "text", + "content": " is expressed as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 281, + 545, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 281, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 326, + 281, + 545, + 297 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {j o i n t}} = \\operatorname {D i c e} \\left(C \\left(l _ {\\text {s e m}}, l _ {\\text {f r e e}}\\right), C \\left(\\hat {y} _ {\\text {s e m}}, \\hat {y}\\right)\\right), \\tag {10}", + "image_path": "4590ba1e90278e4ed3a8e94cb293fc43061cf19314d10efd3fb5d69de33445bf.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 304, + 545, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 340 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 304, + 545, + 340 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 304, + 545, + 340 + ], + "type": "text", + "content": " is the concatenation operation and " + }, + { + "bbox": [ + 305, + 304, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\hat{y}_{sem}" + }, + { + "bbox": [ + 305, + 304, + 545, + 340 + ], + "type": "text", + "content": " is the semantic GT of voxel grid. Dice represents the Dice loss [27]. Finally, the total loss is calculated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 354, + 345, + 545, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 345, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 354, + 345, + 545, + 358 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\gamma_ {6} \\mathcal {L} _ {s d f} + \\gamma_ {7} \\mathcal {L} _ {s e m} + \\gamma_ {8} \\mathcal {L} _ {j o i n t} \\tag {11}", + "image_path": "6e30bb4bcca53971fc5ff3fb5d5e1635d070c72132208e03699aaf75bb16d83a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 361, + 443, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 443, + 374 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 443, + 374 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 361, + 443, + 374 + ], + "type": "inline_equation", + "content": "\\gamma_6, \\gamma_7, \\gamma_8" + }, + { + "bbox": [ + 306, + 361, + 443, + 374 + ], + "type": "text", + "content": " are loss weights." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 381, + 373, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 381, + 373, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 381, + 373, + 392 + ], + "type": "text", + "content": "4.3. Inference" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 399, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 545, + 436 + ], + "type": "text", + "content": "Based on the output SDF and semantic field, diverse outputs for different tasks can be obtained through different sampling strategies and post-processing methods." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 439, + 545, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 439, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 305, + 439, + 545, + 511 + ], + "type": "text", + "content": "Semantic Mesh Generation To generate the semantic mesh, we uniformly query the SDF value in the space of interest and use the marching cubes algorithm [28] to obtain the mesh. Subsequently, at the vertex positions of each triangular face, we query the corresponding semantic information, resulting in a mesh enriched with semantics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 521, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 521, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 545, + 568 + ], + "type": "text", + "content": "Occupancy Prediction The objective of the occupancy prediction is to anticipate the semantics of the voxel grid in space. We achieve this through the joint prediction based on the predicted SDF and semantic field as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 366, + 572, + 545, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 572, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 366, + 572, + 545, + 587 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {i} = \\operatorname {a r g m a x} \\left(C \\left(l _ {\\text {s e m}} ^ {i}, l _ {f} ^ {i}\\right)\\right), \\tag {12}", + "image_path": "5d5efa07bb4b498aa01e2ac5772cfb48e5d3ddd8dcb062a6d2a5ea4a8f8db90c.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 596, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 596, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 305, + 596, + 545, + 622 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 596, + 545, + 622 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 305, + 596, + 545, + 622 + ], + "type": "text", + "content": " is the i-th voxel grid, " + }, + { + "bbox": [ + 305, + 596, + 545, + 622 + ], + "type": "inline_equation", + "content": "l_{sem}^{i}" + }, + { + "bbox": [ + 305, + 596, + 545, + 622 + ], + "type": "text", + "content": " is the logits of the i-th voxel grid, and " + }, + { + "bbox": [ + 305, + 596, + 545, + 622 + ], + "type": "inline_equation", + "content": "l_{f}^{i}" + }, + { + "bbox": [ + 305, + 596, + 545, + 622 + ], + "type": "text", + "content": " the logits value from SDF." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "content": "Semantic Scene Reconstruction Our SurroundSDF demonstrates a notable advantage by achieving semantic scene reconstruction concurrently with scene perception. As is shown in Figure 6, we sequentially query the semantic and SDF values for the region of interest frame by frame. The SDF values and semantic logits of static objects are projected onto the sampling points in the world coordinate" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21618" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 71, + 262, + 178 + ], + "blocks": [ + { + "bbox": [ + 72, + 71, + 262, + 178 + ], + "lines": [ + { + "bbox": [ + 72, + 71, + 262, + 178 + ], + "spans": [ + { + "bbox": [ + 72, + 71, + 262, + 178 + ], + "type": "image", + "image_path": "ff7020888be29ac63c2182a9e0baf55fbb9d5ea9263c0a3eb359d672b71df0fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 186, + 287, + 242 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 287, + 242 + ], + "type": "text", + "content": "Figure 6. The process of semantic scene reconstruction. SDF values and semantic logits are sampled long the vehicle trajectory into the world coordinate system. For regions where the historical sampling overlaps with the current frame, the historical SDF and semantic logits are updated based on a momentum update strategy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 246, + 287, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 287, + 305 + ], + "type": "text", + "content": "system. For the overlapping of the historical sampling area and the current frame sampling area, we employ a momentum update strategy to refine the historical SDF values and semantic logits. Figure 7 shows the final scene reconstruction result with the marching cubes algorithm." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 311, + 128, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 128, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 128, + 324 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 331, + 163, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 163, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 163, + 344 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 63, + 354, + 269, + 412 + ], + "blocks": [ + { + "bbox": [ + 63, + 354, + 269, + 412 + ], + "lines": [ + { + "bbox": [ + 63, + 354, + 269, + 412 + ], + "spans": [ + { + "bbox": [ + 63, + 354, + 269, + 412 + ], + "type": "table", + "html": "
VersionBackboneImage sizeChannelsFrames
SmallResNet50 [8]256 × 704329
MediumResNet101 [8]640 × 16001286
LargeConvNext-B [23]640 × 16002563
", + "image_path": "637f3f1f92ad32a744d67d1b8f09af65ed52fd05495c1b1445add806500e7bb2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "content": "Implementation details We follow BEVStereo [15] to construct our 3D encoder, where the depth net predicts 88 discrete depth values, the voxel feature resolution is " + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "inline_equation", + "content": "200 \\times 200 \\times 16" + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "content": ", and the temporal fusion strategy is applied. We pre-train our architecture on semantic segmentation, depth estimation, and 3D object detection tasks on the NuScenes training set. To compare with methods under different settings, we train three versions: Small, Medium, and Large with different backbone, input image size, voxel feature channels, and the number of temporal frames, as is shown in Tabel 1. For the medium and large versions, we reduce the number of temporal frames to save GPU memory. We use the Adam optimizer [13], set the batch size to 32, fix the learning rate to 1e-4, and train for around 32 epochs for each version. The loss weights " + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\gamma_{1} \\sim \\gamma_{8}" + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "content": " are set to 1.0, 1.0, 30.0, 0.05, 0.05, 1.0, 1.0, and 1.0 respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Dataset We evaluate our method on Occ3D-nuScenes dataset [35, 36], which is an occupancy dataset with dense voxel grid label based on the nuScenes dataset [2]. It processes dynamic and static objects separately, uses multiframe aggregation to achieve point cloud densification, and then reconstructs the dense point cloud to obtain a 3D occupancy label. It has voxel-level semantic labels of 17 classes," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 71, + 536, + 125 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 536, + 125 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 536, + 125 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 536, + 125 + ], + "type": "image", + "image_path": "56dbb50ed206efffd9f081e546fca7cae165dd5556febb7542d9c206561341aa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 125, + 509, + 135 + ], + "lines": [ + { + "bbox": [ + 343, + 125, + 509, + 135 + ], + "spans": [ + { + "bbox": [ + 343, + 125, + 509, + 135 + ], + "type": "text", + "content": "a. Scene reconstruction results from a global perspective" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 137, + 537, + 191 + ], + "blocks": [ + { + "bbox": [ + 315, + 137, + 537, + 191 + ], + "lines": [ + { + "bbox": [ + 315, + 137, + 537, + 191 + ], + "spans": [ + { + "bbox": [ + 315, + 137, + 537, + 191 + ], + "type": "image", + "image_path": "50ccaa20b9b2f2cd148b6e03ef81513ada1d4cd6b7d48c5ba50ac1be4aadf651.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 191, + 507, + 199 + ], + "lines": [ + { + "bbox": [ + 343, + 191, + 507, + 199 + ], + "spans": [ + { + "bbox": [ + 343, + 191, + 507, + 199 + ], + "type": "text", + "content": "b. Scene reconstruction results from a local perspective" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 202, + 529, + 213 + ], + "lines": [ + { + "bbox": [ + 321, + 202, + 529, + 213 + ], + "spans": [ + { + "bbox": [ + 321, + 202, + 529, + 213 + ], + "type": "text", + "content": "Figure 7. Visualization of semantic scene reconstruction." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 217, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 217, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 545, + 264 + ], + "type": "text", + "content": "which contain 16 common classes and an additional general object class. Each frame covers a range of " + }, + { + "bbox": [ + 305, + 217, + 545, + 264 + ], + "type": "inline_equation", + "content": "[-40\\mathrm{m}, -40\\mathrm{m}, -1\\mathrm{m}, 40\\mathrm{m}, 40\\mathrm{m}, 5.4\\mathrm{m}]" + }, + { + "bbox": [ + 305, + 217, + 545, + 264 + ], + "type": "text", + "content": " with a voxel size of " + }, + { + "bbox": [ + 305, + 217, + 545, + 264 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{m}" + }, + { + "bbox": [ + 305, + 217, + 545, + 264 + ], + "type": "text", + "content": ", resulting in a voxel grid resolution with " + }, + { + "bbox": [ + 305, + 217, + 545, + 264 + ], + "type": "inline_equation", + "content": "200 \\times 200 \\times 16" + }, + { + "bbox": [ + 305, + 217, + 545, + 264 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 265, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 265, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 265, + 545, + 312 + ], + "type": "text", + "content": "Evaluation Metrics To evaluate our performance on dense 3D scene perception, we follow the 3D semantic occupancy prediction task [36] and use the widely adopted mean intersection over union (mIoU)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 355, + 320, + 545, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 320, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 355, + 320, + 545, + 350 + ], + "type": "interline_equation", + "content": "m I o U = \\frac {1}{C} \\sum_ {i = 1} ^ {C} \\frac {T P _ {i}}{T P _ {i} + F P _ {i} + F N _ {i}}, \\tag {13}", + "image_path": "34c5f9d45160732046981345c78f2c9fff03d9b3b913e98571775c15856f79b3.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 354, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 545, + 390 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 354, + 545, + 390 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 354, + 545, + 390 + ], + "type": "text", + "content": " represents the number of categories (set to 17 in the experiment), and TP, FP, and FN represent true positive, false positive, and false negative respectively." + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 314, + 399, + 533, + 474 + ], + "blocks": [ + { + "bbox": [ + 72, + 421, + 261, + 432 + ], + "lines": [ + { + "bbox": [ + 72, + 421, + 261, + 432 + ], + "spans": [ + { + "bbox": [ + 72, + 421, + 261, + 432 + ], + "type": "text", + "content": "Table 1. Experimental settings of different versions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 399, + 533, + 474 + ], + "lines": [ + { + "bbox": [ + 314, + 399, + 533, + 474 + ], + "spans": [ + { + "bbox": [ + 314, + 399, + 533, + 474 + ], + "type": "table", + "html": "
MetricsFormula
Abs Rel1/n∑|d−d*|/d*
Sq Rel1/n∑|d−d*|2/d*
RMSE√1/n∑|d−d*|2
δ<1.251/n∑(max(d/d*, d*/d) < 1.25)
", + "image_path": "06cfbbc1905e340648e344d693ba326c7ea5536c102477da6252b68526468608.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 332, + 476, + 518, + 487 + ], + "lines": [ + { + "bbox": [ + 332, + 476, + 518, + 487 + ], + "spans": [ + { + "bbox": [ + 332, + 476, + 518, + 487 + ], + "type": "text", + "content": "Table 2. Depth metrics for 3D scene reconstruction" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 497, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 581 + ], + "type": "text", + "content": "To evaluate our scene reconstruction performance, we render the 2D depth image for the surround cameras based on the reconstructed scene and validate using 2D depth evaluation. As is shown in Table 2, we use the metrics commonly used for depth estimation, which aim to measure the difference between the predicted depth obtained by mesh rendering and GT depth obtained by LiDAR points." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 587, + 494, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 587, + 494, + 601 + ], + "spans": [ + { + "bbox": [ + 305, + 587, + 494, + 601 + ], + "type": "text", + "content": "5.2.3D Semantic Occupancy Prediction" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Following the widely adopted experimental setup [19, 36], we conduct the semantic occupancy prediction experiment on the Occ3D-nuScenes dataset. For a fair comparison, in the inference phase, we divide the space into voxel grids with the same size as the other methods (0.4m) and predict the occupancy semantics following Section 4.3. As is shown in Tabel 3, the mIoU score of our Small version using ResNet50 as the backbone and without any test-time-augmentation (TTA) achieves 42.37, surpassing all the" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21619" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "text", + "content": "methods with the same image backbone. Additionally, our Medium version significantly outperforms BEVFormer [18] and VoxFormer [17] using the ResNet101 backbone. Moreover, our Large version surpasses UniOCC [29] with the same ConvNext-B backbone by a 1.5 mIoU score, and even outperforms FB-OCC [19, 20], which utilizes a larger image backbone and TTA strategy. These results demonstrated the effectiveness of SurroundSdf in the perception task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 175, + 187, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 187, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 187, + 186 + ], + "type": "text", + "content": "5.3. 3D Scene Reconstruction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 190, + 288, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 288, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 288, + 382 + ], + "type": "text", + "content": "To evaluate the performance of scene reconstruction, we compare our method with state-of-the-art (SOTA) surround-view-based depth estimation methods [33, 40] and occupancy prediction methods [19, 41]. To acquire the surround depth, we utilize the semantic mesh generation steps described in Section 4.3 to generate the mesh and then render the depth maps for each camera based on the camera parameters. To fairly compare with the occupancy-based methods, we follow a similar process to render the depth maps with predicted occupied voxel grids, and all methods use RestNet50 as the image backbone. As is shown in Table 4, our method exhibits significant advantages over occupancy-based approaches across all metrics. Moreover, on the primary metric of absolute relative error, we achieve SOTA results, surpassing SurroundDepth [40] and R3D3 [33], which are the SOTA methods for surround-view depth estimation." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 55, + 390, + 279, + 628 + ], + "blocks": [ + { + "bbox": [ + 55, + 390, + 279, + 628 + ], + "lines": [ + { + "bbox": [ + 55, + 390, + 279, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 279, + 628 + ], + "type": "table", + "html": "
MethodsBackboneImage SizemIoU
SurroundOcc [41]ResNet50256 × 70436.32
BEVStereo [15]ResNet50384 × 70439.1
UniOCC [29]ResNet50256 × 70439.7
MiLO [37]ResNet50256 × 70440.49
FB-OCC [19]ResNet50256 × 70440.69
FB-OCC* [19]ResNet50256 × 70442.06
OursResNet50256 × 70442.37
BEVFormer [18]ResNet101900 × 160040.6
VoxFormer [17]ResNet101900 × 160040.6
OursResNet101640 × 160046.0
SurroundOcc [41]InterImage-B640 × 160040.7
BEVFormer [18]InterImage-XL640 × 160043.3
BEVDet [10]Swin-B640 × 160043.1
UniOCC [29]ConvNext-B640 × 160051.5
FB-OCC* [19]InternImage-H960 × 176952.79
OursConvNext-B640 × 160053.01
", + "image_path": "19e3f6ca86262c82f6dd7d641c24d9fa954992ef1fb5530f55a3202a36722c49.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 671, + 141, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 141, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 141, + 684 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "In this section, we study the proposed strategies and the impact of GT quality." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 312, + 71, + 538, + 143 + ], + "blocks": [ + { + "bbox": [ + 46, + 637, + 287, + 661 + ], + "lines": [ + { + "bbox": [ + 46, + 637, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 637, + 287, + 661 + ], + "type": "text", + "content": "Table 3. Comparison on the Occ3D-nuScenes val set. The superscript * denotes using test time augmentation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 71, + 538, + 143 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 538, + 143 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 538, + 143 + ], + "type": "table", + "html": "
MethodsAbs Rel ↓Sq Rel ↓RMSE ↓δ < 1.25 ↑
SurroundOcc0.2742.0725.3270.482
FB-OCC0.3422.0475.9700.290
SurroundDepth0.2242.1024.5730.751
R3D30.2592.3285.5770.583
Ours0.1741.0974.4020.747
", + "image_path": "287564f119d4700e4adddaf24051e67a874d81df431f47d8496affc2e58575e4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 147, + 545, + 169 + ], + "lines": [ + { + "bbox": [ + 306, + 147, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 306, + 147, + 545, + 169 + ], + "type": "text", + "content": "Table 4. 3D scene reconstruction performance. AbsRel is the main metric. “↓” means less is better." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 175, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 175, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 545, + 366 + ], + "type": "text", + "content": "Ablation Study on Proposed Strategies We conduct ablation experiments based on the settings of the Small version in Tabel 1 and the results are shown in Tabel 5. These studies aim to analyze the effectiveness of different 3D representations and supervisions. Specifically, we first construct a baseline model (model A) that is supervised by the occupancy GT with the cross-entropy loss and output the occupancy results. Then, to investigate the 3D perception capability of implicit SDF and semantic field, we build the subsequent models based on the architecture in Section 4.1 and output the SDF and semantic field. To analyze the proposed weak supervision paradigm with previous methods, we reproduce the supervision methods of SIREN (model B) and LODE (model C). In model D, we employ our Sandwich Eikonal for SDF supervision. Lastly, model E incorporates the joint supervision strategy in section 4.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 366, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 546, + 605 + ], + "type": "text", + "content": "Comparing the results of Model B, and C with A, we observe a significant advantage in the AbsRel metric for both models B and C. This highlights the robust reconstruction capability of SDF-based 3D representation. However, we also observe a significant decline in the semantic mIoU metric for model B, attributed to the sparse supervision with Li-DAR points. LODE mitigates this issue by introducing occupancy supervision, but its mIoU score is still lower than the baseline. Using the proposed Sandwich Eikonal supervision paradigm, the results of model D not only surpass the baseline in semantic mIoU but also outperform models B and C in AbsRel. This indicates the effectiveness of our supervision approach, which successfully integrates the supervision from LiDAR points and occupancy GT. Furthermore, by incorporating of the proposed joint supervision strategy, model E exhibits a 1.85 improvement in semantic mIoU, accompanied by a slight regression of 0.003 in the AbsRel metric. For some long-tail categories, such as bicycle, motorcycle, and traffic cone, our joint supervision strategy yields a noticeable improvement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Ablation Study on Ground Truth Quality To further validate the roles of OCC GT and LiDAR GT, we conducted ablation experiments on the resolution of OCC GT and the density of LiDAR points respectively. As is shown in Table 6, the impact of OCC GT resolution variations in our method's OCC GT is mitigated with LiDAR points supervision. When the voxel grid resolution increased from " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{m}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "1.6\\mathrm{m}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", the mIoU only decreased by " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", and Abs-Rel increased by a mere " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.0\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". In comparison, the SOTA OCC" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21620" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 71, + 525, + 207 + ], + "blocks": [ + { + "bbox": [ + 66, + 71, + 525, + 207 + ], + "lines": [ + { + "bbox": [ + 66, + 71, + 525, + 207 + ], + "spans": [ + { + "bbox": [ + 66, + 71, + 525, + 207 + ], + "type": "table", + "html": "
Model3D RepresentationSDF SupervisionJoint Supothersbarrierbicyclebuscarvehiclemotorcyclepedestriantraffic conetrailertruckdrivable surfaceother flatsidewalkterrainmannadevegetationmIoUAbs-Rel
AOCC——8.247.07.643.651.425.618.421.722.432.238.783.643.055.357.750.644.638.330.342
BSDFSIREN—8.041.418.625.941.019.519.015.623.019.827.872.338.648.746.536.633.631.510.188
CSDFLODE—8.746.018.638.547.225.322.220.426.323.734.581.244.854.053.840.337.836.660.192
DSDFSandwich—12.347.925.443.051.727.424.924.927.434.039.483.642.954.458.148.343.240.520.171
ESDFSandwich✓13.949.727.844.653.030.029.028.331.135.841.283.644.655.358.949.643.842.370.174
", + "image_path": "d937f8ab06a643e435120747878b7976f66175e08da1158a961d739aa5df78e2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 87, + 237, + 509, + 434 + ], + "blocks": [ + { + "bbox": [ + 87, + 237, + 509, + 434 + ], + "lines": [ + { + "bbox": [ + 87, + 237, + 509, + 434 + ], + "spans": [ + { + "bbox": [ + 87, + 237, + 509, + 434 + ], + "type": "image", + "image_path": "c19d81ea6bb47b78f02437a3a36b7cbb8822515cb969eb58df349243c9de11ec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 439, + 440, + 451 + ], + "lines": [ + { + "bbox": [ + 153, + 439, + 440, + 451 + ], + "spans": [ + { + "bbox": [ + 153, + 439, + 440, + 451 + ], + "type": "text", + "content": "Figure 8. Comparison of the 3D scene perspective results with SOTA methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 54, + 459, + 280, + 503 + ], + "blocks": [ + { + "bbox": [ + 47, + 209, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 47, + 209, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 545, + 232 + ], + "type": "text", + "content": "Table 5. Results of the ablation experiments. \"3D Representation\" indicates whether the scene is described by occupancy voxels or SDF. \"SDF Supervision\" indicates the supervised paradigm. \"Joint Sup\" indicates whether the joint supervision strategy in Section 4.2 is used." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 459, + 280, + 503 + ], + "lines": [ + { + "bbox": [ + 54, + 459, + 280, + 503 + ], + "spans": [ + { + "bbox": [ + 54, + 459, + 280, + 503 + ], + "type": "table", + "html": "
OCC GT size0.4m0.8m1.6m
TargetmIoUAbs-RelmIoUAbs-RelmIoUAbs-Rel
FBOCC40.690.34236.010.37026.960.445
SurroundSDF42.370.17440.740.17640.160.181
", + "image_path": "1c07c03c79400ffaaeff061a0f53b70d9a9e03e6b1592f8e3a1b839dc23c1753.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 59, + 525, + 272, + 564 + ], + "blocks": [ + { + "bbox": [ + 48, + 506, + 286, + 517 + ], + "lines": [ + { + "bbox": [ + 48, + 506, + 286, + 517 + ], + "spans": [ + { + "bbox": [ + 48, + 506, + 286, + 517 + ], + "type": "text", + "content": "Table 6. Results of the experiments on the resolution of OCC GT." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 525, + 272, + 564 + ], + "lines": [ + { + "bbox": [ + 59, + 525, + 272, + 564 + ], + "spans": [ + { + "bbox": [ + 59, + 525, + 272, + 564 + ], + "type": "table", + "html": "
Sampling rate10.50.20.1
mIoU42.3742.3742.1641.07
Abs-Rel0.1740.1700.1860.186
", + "image_path": "4a024606e5fc6cdce54255481b241fe3179a3c45b44e69e02f7c5f8cd1e1dcea.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 567, + 276, + 578 + ], + "lines": [ + { + "bbox": [ + 58, + 567, + 276, + 578 + ], + "spans": [ + { + "bbox": [ + 58, + 567, + 276, + 578 + ], + "type": "text", + "content": "Table 7. Ablation study on the density of LiDAR points GT." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "text", + "content": "prediction method FBOCC experienced a substantial mIoU decrease of " + }, + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "inline_equation", + "content": "33.7\\%" + }, + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "text", + "content": ", coupled with a " + }, + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "inline_equation", + "content": "30.1\\%" + }, + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "text", + "content": " increase in Abs-Rel. As shown in Table 7, utilizing only " + }, + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "text", + "content": " of the point cloud resulted in a marginal decrease of " + }, + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "text", + "content": " in mIoU and a modest increase of " + }, + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "inline_equation", + "content": "6.9\\%" + }, + { + "bbox": [ + 46, + 586, + 287, + 670 + ], + "type": "text", + "content": " in Abs-Rel. These experiments further substantiate the robustness and efficacy of the proposed Sandwich Eikonal supervision paradigm." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 209, + 689 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 209, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 209, + 689 + ], + "type": "text", + "content": "5.5. Comparison of Visual Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "We visualize our 3D scene perception results and compare them with the reproduced SIREN [34] and LODE [14]" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 460, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 460, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 305, + 460, + 545, + 509 + ], + "type": "text", + "content": "(models B and C in Section 5.4), and the SOTA occupancy prediction method FBOCC [19, 20], as illustrated in Figure 8. Our SurroundSDF demonstrates significant advantages in the continuity and accuracy of the geometric structure." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 517, + 379, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 379, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 379, + 529 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "In this work, we propose SurroundSDF, a novel vision-centric 3D scene understanding framework. We introduce SDF to address the continuity and accuracy of perception from surround cameras. Moreover, in the absence of SDF GT, we propose Sandwich Eikonal formulation, a novel SDF supervision paradigm, which emphasizes imposing appropriate constraints on both sides of the surface to enhance geometric accuracy. Furthermore, to alleviate the inconsistency between geometric optimization objectives and semantic optimization objectives, we introduce a joint supervision strategy. Based on the generated SDF and semantic field, different 3D representations can be obtained, including the scene mesh, occupancy voxel grids, and semantic reconstruction mesh of the whole scene. Comprehensive experiments validate the performance of our method." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21621" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Ben Agro, Quinlan Sykora, Sergio Casas, and Raquel Urtasun. Implicit occupancy flow fields for perception and prediction in self-driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1379-1388, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 213 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 257 + ], + "type": "text", + "content": "[3] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "type": "text", + "content": "[4] Anh-Quan Cao and Raoul de Charette. Scenerf: Self-supervised monocular 3d scene reconstruction with radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9387–9398, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 346 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 390 + ], + "type": "text", + "content": "[6] Vitor Guizilini, Rui Hou, Jie Li, Rares Ambrus, and Adrien Gaidon. Semantically-guided representation learning for self-supervised monocular depth. In International Conference on Learning Representations, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "text", + "content": "[7] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 435, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 287, + 479 + ], + "type": "text", + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 534 + ], + "type": "text", + "content": "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "type": "text", + "content": "[10] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1, 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "type": "text", + "content": "[11] Peixiang Huang, Li Liu, Renrui Zhang, Song Zhang, Xinli Xu, Baichao Wang, and Guoyi Liu. Tig-bev: Multi-view bev 3d object detection via target inner-geometry learning. 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 680 + ], + "type": "text", + "content": "[12] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9223-9232, 2023. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[13] Diederick P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "text", + "content": "[14] Pengfei Li, Ruowen Zhao, Yongliang Shi, Hao Zhao, Jirui Yuan, Guyue Zhou, and Ya-Qin Zhang. Lode: Locally conditioned eikonal implicit scene completion from sparse lidar. In IEEE International Conference on Robotics and Automation, ICRA 2023, London, UK, May 29 - June 2, 2023, pages 8269-8276. IEEE, 2023. 2, 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 195 + ], + "type": "text", + "content": "[15] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1486-1494, 2023. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 252 + ], + "type": "text", + "content": "[16] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 253, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 319 + ], + "type": "text", + "content": "[17] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023. 1, 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 320, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 386 + ], + "type": "text", + "content": "[18] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1, 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "type": "text", + "content": "[19] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 433, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 488 + ], + "type": "text", + "content": "[20] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M Alvarez. Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6919-6928, 2023. 2, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "type": "text", + "content": "[21] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision, pages 531-548. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 535, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 590 + ], + "type": "text", + "content": "[22] Yingfei Liu, Junjie Yan, Fan Jia, Shuailin Li, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petrv2: A unified framework for 3d perception from multi-camera images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3262-3272, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "type": "text", + "content": "[23] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[24] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781. IEEE, 2023. 1, 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21622" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[25] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 159 + ], + "type": "text", + "content": "[26] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 287, + 215 + ], + "type": "text", + "content": "[27] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3D Vision (3DV), 2016 Fourth International Conference on, pages 565-571. IEEE, 2016. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 248 + ], + "type": "text", + "content": "[28] Timothy S Newman and Hong Yi. A survey of the marching cubes algorithm. Computers & Graphics, 30(5):854-879, 2006. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 304 + ], + "type": "text", + "content": "[29] Mingjie Pan, Li Liu, Jiaming Liu, Peixiang Huang, Longlong Wang, Shaoqing Xu, Zhiyi Lai, Shanghang Zhang, Kuiyuan Yang, and Xiaomi Car. Uniocc: Unifying vision-centric 3d occupancy prediction with geometric and semantic rendering. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 360 + ], + "type": "text", + "content": "[30] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "type": "text", + "content": "[31] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "type": "text", + "content": "[32] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "text", + "content": "[33] Aron Schmied, Tobias Fischer, Martin Danelljan, Marc Pollefeys, and Fisher Yu. R3d3: Dense 3d reconstruction of dynamic scenes from multiple cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3216-3226, 2023. 1, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "text", + "content": "[34] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33:7462-7473, 2020. 1, 2, 3, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "text", + "content": "[35] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "text", + "content": "[36] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8406-8415, 2023. 3, 4, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[37] Thang Vu, Jung-Hee Kim, Myeongjin Kim, Seokwoo Jung, and Seong-Gyun Jeong. Milo: Multi-task learning with localization ambiguity suppression for occupancy prediction cvpr 2023 occupancy challenge report. 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 518 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 127 + ], + "type": "text", + "content": "[38] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 34:27171-27183, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[39] Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, , and Justin M. Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In The Conference on Robot Learning (CoRL), 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "type": "text", + "content": "[40] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Yongming Rao, Guan Huang, Jiwen Lu, and Jie Zhou. Surround-depth: Entangling surrounding views for self-supervised multi-camera depth estimation. In Conference on Robot Learning, pages 539-549. PMLR, 2023. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 229, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 545, + 284 + ], + "type": "text", + "content": "[41] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 1, 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 285, + 545, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 339 + ], + "type": "text", + "content": "[42] Enze Xie, Zhiding Yu, Daquan Zhou, Jonah Philion, Anima Anandkumar, Sanja Fidler, Ping Luo, and Jose M Alvarez. " + }, + { + "bbox": [ + 307, + 285, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^2" + }, + { + "bbox": [ + 307, + 285, + 545, + 339 + ], + "type": "text", + "content": " bev: Multi-camera joint 3d detection and segmentation with unified birds-eye view representation. arXiv preprint arXiv:2204.05088, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 341, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 341, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 341, + 545, + 384 + ], + "type": "text", + "content": "[43] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. In The Eleventh International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 386, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 386, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 386, + 545, + 418 + ], + "type": "text", + "content": "[44] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 419, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 419, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 419, + 545, + 464 + ], + "type": "text", + "content": "[45] Weihao Yuan, Xiaodong Gu, Heng Li, Zilong Dong, and Siyu Zhu. Monocular scene reconstruction with 3d sdf transformers. In The Eleventh International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 464, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 518 + ], + "type": "text", + "content": "[46] Yunpeng Zhang, Phigent Robotics, and Zheng Zhu. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9433-9443, 2023. 1, 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21623" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/13853f95-872e-49e0-9add-f414db12f2de_content_list.json b/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/13853f95-872e-49e0-9add-f414db12f2de_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9643d9736109844d52229e2b954c7249269b3d64 --- /dev/null +++ b/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/13853f95-872e-49e0-9add-f414db12f2de_content_list.json @@ -0,0 +1,1881 @@ +[ + { + "type": "text", + "text": "SwiftBrush: One-Step Text-to-Image Diffusion Model with Variational Score Distillation", + "text_level": 1, + "bbox": [ + 212, + 130, + 756, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thuan Hoang Nguyen Anh Tran VinAI Research, Vietnam", + "bbox": [ + 343, + 203, + 625, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{v.thuannh5,v.anhtt152}@vinai.io", + "bbox": [ + 341, + 241, + 624, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/abf8a724058c432da20e579aabfe4a3a5081d1395b2a1ea01e03a3f5f260f336.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 263, + 485, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ab8364f8db5c6fac9b15d022b4942be5534d5a9e5868e73ebbf389730da99e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 263, + 888, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f37fd2f3a3fb5ad229907ccb50823ca6a0c03ff4f9a35b15d1ec1d9efe15ab7.jpg", + "image_caption": [ + "Figure 1. Samples generated in a single step by SwiftBrush. Our distillation method can produce images of high-fidelity roughly 20 times faster compared to Stable Diffusion." + ], + "image_footnote": [], + "bbox": [ + 80, + 575, + 483, + 733 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3d95baf6c1b8a72331904f370a5c43577e5b4fc3ff9badb524a970b52e0892aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 575, + 888, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 777, + 313, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite their ability to generate high-resolution and diverse images from text prompts, text-to-image diffusion models often suffer from slow iterative sampling processes. Model distillation is one of the most effective directions to accelerate these models. However, previous distillation methods fail to retain the generation quality while requir", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing a significant amount of images for training, either from real data or synthetically generated by the teacher model. In response to this limitation, we present a novel image-free distillation scheme named SwiftBrush. Drawing inspiration from text-to-3D synthesis, in which a 3D neural radiance field that aligns with the input prompt can be obtained from a 2D text-to-image diffusion prior via a specialized loss without the use of any 3D data ground-truth, our ap", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7807", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "proach re-purposes that same loss for distilling a pretrained multi-step text-to-image model to a student network that can generate high-fidelity images with just a single inference step. In spite of its simplicity, our model stands as one of the first one-step text-to-image generators that can produce images of comparable quality to Stable Diffusion without reliance on any training image data. Remarkably, SwiftBrush achieves an FID score of 16.67 and a CLIP score of 0.29 on the COCO-30K benchmark, achieving competitive results or even substantially surpassing existing state-of-the-art distillation techniques.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 286, + 209, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models are gaining significant attention from the research community as they have achieved remarkable results in various generative tasks, including image generation [11, 26, 40, 41], video synthesis [12], 3D modeling [3, 28, 46], audio generation [17], and text creation [8, 18]. Especially, text-to-image diffusion models, which combine the power of language models and high-quality diffusion models [32, 36, 43], have revolutionized the way that people create visual content based on textual descriptions. Synthetic images indistinguishable from real photos are now possible with only a few clicks. Additionally, with techniques like ControlNet [49] or DreamBooth [35], extra control to the generation outputs can be made to further boost the expressiveness of the text-to-image models and bring them closer to mass users. However, despite their impressive success, these models exhibit a noteworthy drawback of slow inference speed. The primary reason for this lies in the iterative nature of their sampling, presenting significant challenges for deployment on consumer devices.", + "bbox": [ + 75, + 311, + 472, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several approaches have been proposed to improve the inference speed of diffusion-based text-to-image generation. However, the most effective direction is time-step distillation, which aims to reduce the number of sampling steps of the diffusion model with the network backbone unchanged and minimum reduction in generation quality. Starting from the pioneering work [37], more advanced and effective distillation methods have been proposed, and a few recent ones could train efficient student networks that only require one forward step to synthesize images from input captions. In this work, we are interested in the distillation-based approach for one-step text-to-image generation. Particularly, Meng et al. [25] employ a two-stage distillation, where the student tries to first match its output with the classifier-free guided output of the teacher and then bootstrap itself to reduce the inference steps gradually. Such a method is complex and time-consuming, requiring at least 108 A100 GPU days [21]. Based on the work of Song et al. [42], Latent Consistency Model (LCM) [24] enforces the consistency of all points on the ODE trajectory in the", + "bbox": [ + 75, + 599, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "latent space of Stable Diffusion (SD) [43] and unlike [25], they directly combine the two distillation stages into one. Although both Guided Distillation and LCM significantly reduce the number of inference steps to 2 or 4, their 1-step inference yields blurry and unsatisfying results. More sophisticated, Instaflow [21] proposes a technique called \"reflow\" to straighten the sampling trajectory of the teacher before distilling. While their proposed techniques proven to be superior to the original diffusion framework in one-step text-to-image distillation, InstaFlow requires an expensive 4-stage training schedule. Also, it is noteworthy that the effectiveness of these methods relies heavily on a large number of text-image pairs, which is not always the case due to the limited accessibility of such data.", + "bbox": [ + 496, + 90, + 893, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we aim to develop a novel distillation method for one-step text-to-image generation with higher quality and a more approachable training process characterized by an image-free and straightforward mechanism. Building upon the recent advancements in text-to-3D techniques, our work takes inspiration from their ability to generate high-quality 3D Neural Radiance Fields (NeRF) without utilizing 3D ground truth data. Such a remarkable achievement is facilitated by employing a powerful pretrained 2D text-to-image model to assess whether the image rendered from NeRF is realistic or not, akin to the discriminator in GANs. This observation underpins the motivation for our SwiftBrush approach, as we recognize that these principles can be innovatively adapted for text-to-image generation making image supervision unnecessary. SwiftBrush, therefore, emerges as a conceptually novel one-step text-to-image distillation process, uniquely capable of operating without 2D image supervision. This strategic adaptation not only opens up a change of view in the distillation of text-to-image diffusion models into a single-step generators but also points out the potential of blending principles from text-to-3D to text-to-image synthesis.", + "bbox": [ + 496, + 303, + 893, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a result, our method has successfully enabled the generation of high-quality images with remarkable details. Quantitatively, our one-step model achieves promising zero-shot results, 16.6 for the FID score and 0.29 for the CLIP score on the MS COCO 2014 dataset (30K images). Notably, this is the first time a distilled one-step model outperforms the previous approaches without a single image used for training. A demonstration of generated images from our distilled model is shown in Fig. 1.", + "bbox": [ + 496, + 637, + 893, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 791, + 640, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-Image Generation. Previously, text-to-image models relied on GANs and focused on only small-scale object-centric domains such as flowers and birds [33, 48]. Recently, these types of systems have advanced rapidly with the availability of web-scale datasets of text-image pairs", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7808", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "such as LAION5B [39], large language models such as T5 [30], and large vision-language models such as CLIP [29]. DALL-E [31] was the first autoregressive model to demonstrate a remarkable zero-shot ability in creating images from text captions by simply scaling up both the network and dataset size. Following that, a plethora of methods have been used for text-to-image generation, including masked generative transformers [4], GAN-based, and diffusion models. Based on Latent Diffusion Model [34], Stable Diffusion is an open-source diffusion-based generator that has gained widespread popularity among artists as well as researchers. Recently, text-to-image GANs are receiving attention again with notable work like StyleGAN-T [38], or GigaGAN [15]. While they are extremely fast as no iterative sampling is required, they lack behind diffusion models in terms of scalability and realism due to the fact that they suffer from instability and the infamous issue \"mode collapse\". To the best of our knowledge, so far only two above-mentioned work succeeded in applying GANs for text-to-image using tons of intricate techniques and auxiliary modules. Conversely, diffusion models can be easily scale up to generate high-quality and sharp samples, yet they require lengthy sampling process. Here, our objective is to create a large-scale generative model that combines the speed of GANs with the quality of diffusion models.", + "bbox": [ + 76, + 90, + 472, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-3D Generation. Instead of learning a generative model directly from 3D supervision, many research works have utilized the rich knowledge of 2D prior for 3D generation tasks. For instance, one of the earliest works, Dream Fields [14], leverages CLIP [29] to guide the generated images so that rendered images from many camera views highly align with the text caption. On the other hand, pioneer works like DreamFusion [28] and Score Jacobian Chaining [44] (SJC) propose two different but equivalent frameworks where a 2D text-to-image generative prior is used to generate 3D objects from textual descriptions. Subsequent works seek to improve these text-to-3D methods in various ways. Among them, Magic3D [19] and ProlificDreamer [45] are two remarkable works that effectively enhance the generation quality of 3D assets. Magic3D consists of two stages where they first obtain a low-resolution 3D representation via a DreamFusion-style optimization and then refine its texture to get the final high-fidelity 3D mesh. Meanwhile, ProlificDreamer proposes Variational Score Distillation (VSD) involving a secondary teacher to bridge the gap between the teacher and the 3D NeRF. Influenced by the latter work, our method adapts the same technique for one-step text-to-image generation. Similarly, Score GAN [6] and Diff-Instruct [7] use a VSD-like loss to distill a pretrained diffusion model to a one-step generator.", + "bbox": [ + 76, + 479, + 472, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge Distillation introduced by Hinton et al. [10] is one of the methods falling under the umbrella of transfer", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "learning. This algorithm draws its main inspiration from the human learning process, where knowledge is conveyed from a more knowledgeable teacher to a less knowledgeable student. In the context of diffusion models, works have been proposed for faster sampling of student models from a pre-trained diffusion teacher. The most straightforward approach involves direct distillation [23], wherein a student model is supervised from the teacher's output after sampling for 50-1000 steps, which can be prohibitive. Recent work [37, 42] avoids the lengthy sampling during distillation by various bootstrapping methods. Afterward, a series of methods [21, 24, 25] adapt these methods into text-to-image settings and inherit their limitation of being image-dependent in training. In contrast, a concurrent work [9] learns a time-conditioned model to predict the outcome of the teacher at any given time step, all without the need for image supervision. Likewise, our approach is similarly image-free yet achieves significantly better results and offers a much simpler distillation design.", + "bbox": [ + 496, + 90, + 893, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 390, + 671, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 500, + 415, + 629, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models consist of two processes: a forward process that gradually adds noise and another reverse process to predict the distribution of gradually denoised data points. The forward process incrementally introduces noise to an initial data point $x_0$ , drawn from the distribution $q_0(x_0)$ , resulting in $q_t(x_t|x_0) = \\mathcal{N}(\\alpha_t x_0, \\sigma_t^2 I)$ , where $\\{(\\alpha_t, \\sigma_t)\\}_{t=1}^T$ is the noise schedule. The data point $x_t$ at each time-step $t$ is drawn from $q_t(x_t|x_0)$ using a noise $\\epsilon \\sim \\mathcal{N}(0, I)$ :", + "bbox": [ + 496, + 445, + 893, + 566 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\alpha_ {t} x _ {0} + \\sigma_ {t} \\epsilon . \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 577, + 890, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After $T$ time steps, the final data point is expected to be pure Gaussian noise $x_{T}\\sim \\mathcal{N}(0,I)$", + "bbox": [ + 496, + 599, + 890, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conversely, the backward process starts from $x_{T} \\sim \\mathcal{N}(0, I)$ and iteratively denoises for $T$ steps where at each step $t$ , noisy variable $x_{t+1}$ is converted to less-noisy variable $x_{t}$ , guided by a model $\\epsilon_{\\psi}$ that predicts the noise $\\epsilon$ in Eq. (1). The network weights $\\psi$ are trained by minimizing:", + "bbox": [ + 496, + 631, + 893, + 708 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {u n c o n d} (\\psi) = \\mathbb {E} _ {t, \\epsilon \\in \\mathcal {N} (0, 1)} \\| \\epsilon_ {\\psi} (x _ {t}, t) - \\epsilon \\| _ {2} ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 720, + 890, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t$ is uniformly samples within $\\{1,\\dots ,T\\}$", + "bbox": [ + 496, + 742, + 821, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While common diffusion models operate in the pixel space, Latent Diffusion Models (LDMs)[34] model the diffusion process in the latent space of a pre-trained regularized autoencoder, which features representations of smaller size, thus enhancing both training and inference efficiency.", + "bbox": [ + 496, + 757, + 890, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-Image Diffusion Models. In contrast to unconditional diffusion models that generate outputs freely, text-conditioned diffusion models steer the sampling process using an extra prompt $y$ . This guides the model to produce", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7809", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "outputs that are not only photorealistic but also closely adhere to the provided text description. The objective to train such models is formulated as follows:", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d i f f} (\\psi) = \\mathbb {E} _ {t, y, \\epsilon \\in \\mathcal {N} (0, 1)} \\| \\epsilon_ {\\psi} (x _ {t}, t, y) - \\epsilon \\| _ {2} ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 146, + 468, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which is slightly different from the unconditional diffusion loss Eq. (2). Due to the prompt conditioning, the model can deliver more controllable generation compared to its unconditional counterparts. Nevertheless, the implementations of many leading methods [1, 32, 36] remain inaccessible to the public. Stable Diffusion, which primarily utilizes the LDM framework, stands out as the first openly available large-scale model, significantly propelling text-to-image synthesis's widespread adoption and versatility.", + "bbox": [ + 75, + 174, + 468, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Score Distillation Sampling (SDS) is a distillation technique effectively applied in generating 3D assets [19, 28, 44]. It utilizes a pre-trained text-to-image diffusion model, which predicts diffusion noise from text condition $y$ , denoted as $\\epsilon_{\\psi}(x_t,t,y)$ . The method optimizes a single 3D NeRF, parameterized by $\\theta$ , to align with a given text prompt. Given camera parameters $c$ , a differentiable rendering function $g(\\cdot ,c)$ is used to render an image at the camera view $c$ from the 3D NeRF. Here, the rendered image $g(\\theta ,c)$ is utilized to optimize the weight $\\theta$ through a loss function whose gradient can be approximated by:", + "bbox": [ + 75, + 316, + 470, + 483 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {S D S} = \\mathbb {E} _ {t, \\epsilon , c} \\left[ w (t) (\\epsilon_ {\\psi} (x _ {t}, t, y) - \\epsilon) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} \\right] (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 492, + 468, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t\\sim \\mathcal{U}(0.02T,0.98T)$ $T$ is the maximum timesteps of the diffusion model, $\\epsilon \\sim \\mathcal{N}(0,I)$ $x_{t} = \\alpha_{t}g(\\theta ,c) + \\sigma_{t}\\epsilon$ $y$ is the input text, and $w(t)$ is a weighting function", + "bbox": [ + 75, + 536, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite the advancements in text-to-3D synthesis, empirical studies [28, 44] indicate that SDS often encounters issues like over-saturation, over-smoothing, and reduced diversity. The same degradations can be observed if we naively apply SDS in our framework, as shown in Sec. 4.3.", + "bbox": [ + 75, + 580, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Variational Score Distillation (VSD) is introduced in ProlificDreamer [45] to address previously mentioned issues of SDS by modifying the loss slightly:", + "bbox": [ + 75, + 662, + 468, + 709 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\theta} \\mathcal {L} _ {V S D} = \\mathbb {E} _ {t, \\epsilon , c} [ w (t) (\\epsilon_ {\\psi} (x _ {t}, t, y) \\\\ \\left. - \\epsilon_ {\\phi} (x _ {t}, t, y, c)) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} \\right] \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 719, + 468, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Again, $x_{t} = \\alpha_{t}g(\\theta ,c) + \\sigma_{t}\\epsilon$ is the noisy observation of the rendered image at camera view $c$ . VSD sets itself apart from SDS by introducing an additional score function tailored to the images rendered from the 3D NeRF at the camera pose $c$ . This score is derived by fine-tuning a diffusion model $\\epsilon_{\\phi}(x_t,t,y,c)$ with the diffusion loss described below:", + "bbox": [ + 75, + 777, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\epsilon_ {\\phi}} \\mathbb {E} _ {t, c, \\epsilon} \\| \\epsilon_ {\\phi} (x _ {t}, t, y, c) - \\epsilon \\| _ {2} ^ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 878, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As proposed in ProlificDreamer, $\\epsilon_{\\phi}$ is parameterized by Low-Rank Adaption [13] (LoRA) and initialized from the pre-trained diffusion model $\\epsilon_{\\psi}$ with some added layers for conditioning the camera view $c$ . Note that at each iteration $i$ of the optimization process, $\\epsilon_{\\phi}$ needs to adapt with the current distribution of $\\theta$ . Therefore, ProlificDreamer interleaves between finetuning $\\epsilon_{\\phi}$ and optimizing $\\theta$ . With these algorithmic enhancements, ProlificDreamer has significantly advanced its capabilities, enabling the generation of NeRFs and the creation of exceptional textured meshes. Such improvement directly inspires us to adapt VSD to the task of one-step text-to-image diffusion-based distillation.", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. SwiftBrush", + "text_level": 1, + "bbox": [ + 500, + 281, + 624, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivation. While SDS and VSD are designed explicitly for the text-to-3D generation task, they loosely connect to that goal via the rendered image $g(\\theta, c)$ of the 3D NeRF. As a matter of fact, instead of using NeRF rendering, we can replace it with any function that outputs 2D images to suit our needs. Inspired by this motivation, we propose to substitute the NeRF rendering $g(\\theta, c)$ with a text-to-image generator that can directly synthesize a text-guided image in one step, effectively converting the text-to-3D generation training into one-step diffusion model distillation.", + "bbox": [ + 496, + 311, + 890, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Design Space. We adopt the same approach as [45], with modifications to the design space to suit our task better. To begin with, we make use of two teacher models: a pretrained text-to-image teacher $\\epsilon_{\\psi}$ and one additional LoRA teacher $\\epsilon_{\\phi}$ . Also, we remove the conditioning on camera view $c$ from the LoRA teacher as it is unnecessary in our case, and we use classifier-free guidance for both teachers. Then, we replace the NeRF, which is overfitted for a specific user-provided prompt in a text-to-3D setting, with a generalized one-step text-to-image student model $f_{\\theta}(z,y)$ . Our student model $f_{\\theta}$ takes a random Gaussian noise $z$ and a text prompt $y$ as inputs. Both the LoRA teacher and student model are initialized with the weight of the text-to-image teacher. Next, we train both the student model and the LoRA teacher alternately using Eq. (5) and Eq. (6) while freezing the text-to-image teacher. A pseudo-code and system figure can be viewed in Alg. 1 and Fig. 2.", + "bbox": [ + 496, + 469, + 890, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Student Parameterization. Given a pretrained text-to-image diffusion model $\\epsilon_{\\theta}$ , it is possible to directly use its output for the student model, i.e., $f_{\\theta}(z,y) = \\epsilon_{\\theta}(z,T,y)$ , where $T$ is the maximum timestep of the pretrained model. However, in our case, the pretrained model of choice is Stable Diffusion, which is inherently designed to predict the added noise $\\epsilon$ . In contrast, our objective is to refine the student model so that it predicts a clean and noise-free $x_0$ . Therefore, such a naive method results in a large domain gap between what we want the student to learn and the output of the student. For the sake of easy training, we empiri", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7810", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 SwiftBrush Distillation", + "text_level": 1, + "bbox": [ + 78, + 90, + 323, + 104 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Require: a pretrained text-to-image teacher $\\epsilon_{\\psi}$ , a LoRA teacher $\\epsilon_{\\phi}$ , a student model $f_{\\theta}$ , two learning rates $\\eta_{1}$ and $\\eta_{2}$ , a weighting function $\\omega$ , a prompts dataset $Y$ , the maximum number of time steps $T$ and the noise schedule $\\{(\\alpha_t, \\sigma_t)\\}_{t=1}^T$ of the teacher model", + "2: Initialize: $\\phi \\gets \\psi, \\theta \\gets \\psi$", + "3: while not converged do", + "4: Sample input noise $z\\sim \\mathcal{N}(0,I)$", + "5: Sample text caption input $y\\sim Y$", + "6: Compute student output $\\hat{x}_0 = f_\\theta (z,y)$", + "7: Sample timestep $t \\sim \\mathcal{U}(0.02T, 0.98T)$", + "8: Sample added noise $\\epsilon \\sim \\mathcal{N}(0,I)$", + "9: Compute noisy sample $\\hat{x}_t = \\alpha_t\\hat{x}_0 + \\sigma_t\\epsilon$", + "10: $\\theta \\gets \\theta -\\eta_{1}\\left[\\omega (t)\\left(\\epsilon_{\\psi}(\\hat{x}_{t},t,y) - \\epsilon_{\\phi}(\\hat{x}_{t},t,y)\\right)\\frac{\\partial\\hat{x}_{0}}{\\partial\\theta}\\right]$", + "11: Sample timestep $t^{\\prime}\\sim \\mathcal{U}(0,T)$", + "12: Sample added noise $\\epsilon^{\\prime}\\sim \\mathcal{N}(0,I)$", + "13: Compute noisy sample $\\hat{x}_{t'} = \\alpha_{t'}\\hat{x}_0 + \\sigma_{t'}\\epsilon'$", + "14: $\\phi \\gets \\phi -\\eta_{2}\\nabla_{\\phi}\\| \\epsilon_{\\phi}(\\hat{x}_{t^{\\prime}},t^{\\prime},y) - \\epsilon^{\\prime}\\|^{2}$", + "15: end while", + "16: return trained student model $f_{\\theta}$" + ], + "bbox": [ + 81, + 109, + 468, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cally re-parameterize the student output as follows:", + "bbox": [ + 76, + 441, + 415, + 458 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} (z, y) = \\frac {z - \\sigma_ {T} \\epsilon_ {\\theta} (z , T , y)}{\\alpha_ {T}}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 465, + 372, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which is a realization of Eq. (1) if we set $t = T$ , $x_{t} = z$ , $\\epsilon \\approx \\epsilon_{\\theta}(z,T,y)$ and $x_0 \\approx f_\\theta (z,y)$ . So basically, this reparameterization transforms the noise-prediction output of the pretrained model into the \"predicted $x_0$ \" form, which is empirically demonstrated in Sec. 4.3 to be helpful for the student model to learn.", + "bbox": [ + 75, + 507, + 468, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 611, + 209, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 636, + 267, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We conduct our evaluation process by assessing the performance of SwiftBrush against several well-established methods, such as Guided Distillation [25], Instaflow [21], LCM [24], and BOOT [9], across two zero-shot text-to-image benchmarks: COCO 2014 [20] and Human Preference Score v2 [47] (HPSv2). On the COCO 2014, we follow the conventional evaluation protocol [15, 34, 36, 38], computing FID and CLIP score on a 30K subset of its validation set as our primary metrics. Here, we use ViT-G/14 as the backbone for evaluating CLIP scores. In the case of HPSv2, we adopt their evaluation procedure to assess text-image alignment across four different aspects: anime, concept-art, paintings, and photos.", + "bbox": [ + 75, + 666, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experiment Protocols. In these experiments, we rely on Stable Diffusion v2.1 [43] for our two teacher models, and", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "whenever possible, we directly utilize the reported values from other research. As for InstaFlow and LCM, we use the their provided pretrained models to generate images for evaluation. Conversely, we are compelled to re-implement BOOT since they did not provide any metrics for their text-to-image models, and their codebase remains entirely closed.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Similar to prior research, we use student models with the same architectures as those of the teachers and initialize all of the student parameters with the teacher model. During training, a guidance scale of 4.5 is used for both teachers. For the LoRA teacher model, we apply a learning rate of 1e-3, a LoRA rank of 64, and an alpha value of 108. Meanwhile, the student model's learning rate is set at 1e-6 with exponential moving average (EMA) used every iteration. Adam optimizer [16] and a batch size of 64 (16 per GPU) are used to train both the student and LoRA teacher for around 65K iterations. Thanks to the image-free nature of SwiftBrush, we do not require access to any image training set, and only the prompts are needed to distill the generator. Here, we utilize 4M captions from the large-scale text-image dataset JourneyDB [27], which we subsequently deduplicated to 2M. Regarding BOOT, we closely follow their implementation guidelines to the best of our abilities and train the student model until its convergence, except that we use Stable Diffusion v2.1 instead of DeepFloyd IF [5] as the teacher. Here, its blurry results (Fig. 4) compared to the original work could be due to the different teacher.", + "bbox": [ + 496, + 203, + 892, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 500, + 530, + 591, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Results. We first evaluate the proposed method on standard text-to-image generation benchmarks. The quantitative comparison with the previous distillation, both image-dependent and image-free, and the original Stable Diffusion teacher is shown in Tab. 1. Despite being trained with only text captions, SwiftBrush significantly improves the quality of 1-step inference and achieves better photo-realism as well as text-image alignment against other methods, with exceptions of InstaFlow and Stable Diffusion (25 sampling steps) in FID score. However, it is worth noting that Instaflow is so complex, requiring a 4-stage distillation process, while ours is end-to-end and straightforward. We further conduct a quantitative evaluation with the HPSv2 score, a metric trained to capture human preference. Again, we obtain the best score in all aspects among all mentioned works, which is only a slight degradation compared to the original Stable Diffusion teacher (Tab. 2). Moreover, according to the anonymous survey carried out by us (Fig. 3), our method outperforms Instaflow in terms of image fidelity and text-to-image accuracy.", + "bbox": [ + 496, + 560, + 890, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative Results. We report the visual comparison between our method and the mentioned baseline methods", + "bbox": [ + 498, + 869, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7811", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7a3b1c07f625c04d4463492843c84e7923ad2df0f8e49b8c07e897915371bc37.jpg", + "image_caption": [ + "Figure 2. SwiftBrush overview. Our system allows training a one-step text-to-image student network from a frozen pretrained teacher with the help of an additional trainable LORA teacher. The student network takes inputs a text prompt and random noise. The output of the student is then added noise and sent to two teachers, alongside the prompt and a randomly drawn timestep, to compute the gradient of VSD loss, which backpropagates back to the student. Apart from that, the LoRA teacher is also updated using diffusion loss. Similar to ProlificDreamer [45], we update the LoRA teacher (green flow) and the student (red flow) alternatively." + ], + "image_footnote": [], + "bbox": [ + 101, + 85, + 867, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/812107473a718a74917a5075987c817a44a80465a3c8221981a22789f003d793.jpg", + "image_caption": [ + "Figure 3. User Study. We ask 400 users to compare the generated images of InstaFlow and our method using 30 random text prompts. We report the rate at which each method is selected as better for image quality and text-to-image accuracy." + ], + "image_footnote": [], + "bbox": [ + 91, + 380, + 455, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for the one-step text-to-image generation task. As seen in Fig. 4, LCM outputs blurry and over-smoothed images in the 1-step regime. BOOT faithfully produces images that match well with the text caption, but its results still fall short in terms of fidelity. In contrast, a much better quality can be observed in the image produced by Instaflow and SwiftBrush. InstaFlow, nevertheless, cannot handle complex or unusual input prompts properly, which can be noticed in the third and fifth images at the third column of Fig. 4. On top of that, it occasionally produces non-realistic results even for simple prompts like in Fig. 4 at the fourth image of the third column. When, in fact, the outputs generated by SwiftBrush are not only of exceptional quality but also exhibit a remarkable alignment between text descriptions and the corresponding images. Moreover, our results are comparable with those sampled from Stable Diffusion with 25 steps; whereas it is understandable that with only one step, Stable Diffusion cannot create outputs of good quality.", + "bbox": [ + 75, + 527, + 468, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Analysis", + "text_level": 1, + "bbox": [ + 76, + 809, + 181, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The importance of LoRA teacher. In Fig. 6, the importance of including the LoRA teacher in the training process is clearly highlighted. We compare the student outputs sampled with different input noises after 10K iteration when us", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1af7892897cf09d0c3720e0bc8c98a03f171e5f4aee1fba4111afd353d3727e8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodStepsFID-30K ↓CLIP-30K ↑
Guided Distillation†137.30.27
LCM†135.560.24
Instaflow113.10†0.28§
BOOT‡148.200.26
Ours116.670.29
SD 2.1*2513.450.30
SD 2.1*1202.140.08
", + "bbox": [ + 501, + 378, + 903, + 507 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3c337de8ad7dfb648225cd8bf1e422bed0a1d94106395a25543df60faf0620b3.jpg", + "table_caption": [ + "Table 1. Comparison of our method against other works based on FID metric and CLIP score on the COCO 2014 dataset. $^\\dagger$ means that we obtain the numbers from the corresponding papers. $\\S$ means that we obtain the numbers using the provided pretrained models of the corresponding papers. $^\\ddagger$ means that we re-implement the work and report the numbers. * means that we report the number using the pretrained models. We use DPM-Solver [22] with guidance scale 7.5 to sample for SD. Bold and underlined numbers are best and second best for one-step models." + ], + "table_footnote": [], + "table_body": "
ModelsHuman Preference Score v2 ↑
AnimePhotoConcept ArtPaintings
LCM†22.6122.7122.7422.91
InstaFlow†25.9826.3225.7925.93
BOOT‡25.2925.1624.4024.61
Ours26.9127.2126.3226.37
SD 2.1*27.4826.8926.8627.46
", + "bbox": [ + 501, + 643, + 898, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of our method against other works based on HPSv2 score in 1-step regime. $\\dagger$ means that we obtain the score using the provided pretrained models of the corresponding papers. $\\ddagger$ means that we re-implement the work and report the score by ourselves. * means that we obtain the score from [47]. Bold and underlined numbers are best and second best, respectively.", + "bbox": [ + 496, + 760, + 893, + 844 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing the same text prompts. With the guidance of the LoRA teacher, the student progressively learns to create realistic images ( $1^{st}$ column). In contrast, in the absence of an ef", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7812", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5afe2db2e4630a5b55287ad4601b7e8c615c01c2cde9c70e5d47e925dbb659a8.jpg", + "image_caption": [ + "\"A photo of one ice cream ball in a luxurious plate, bokeh\"", + "LCM" + ], + "image_footnote": [], + "bbox": [ + 228, + 107, + 328, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/91101967545bb2a4f16931bb2f2845e6a46114454cc147105779e5bae158bbc8.jpg", + "image_caption": [ + "BOOT" + ], + "image_footnote": [], + "bbox": [ + 339, + 107, + 437, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a1c23d3cae446d089bd37a27efa62ef901c65cc4a7f65d8cf4537df000a7b91.jpg", + "image_caption": [ + "InstaFlow" + ], + "image_footnote": [], + "bbox": [ + 447, + 107, + 545, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/953b8d02bc1827b355999659e5b319eef44e1c04682236372358b482a5e0f3ee.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 555, + 107, + 653, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d42bd44d1942e79d7fa32065bfe84469810d77bc08bf624df4d2f4557bc0981.jpg", + "image_caption": [ + "SD*" + ], + "image_footnote": [], + "bbox": [ + 665, + 107, + 763, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b30057df9197efae4d798cd083105401b0d39043f567c50dfaa4d2fef30979d.jpg", + "image_caption": [ + "SD†" + ], + "image_footnote": [], + "bbox": [ + 774, + 107, + 870, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/198cf500f3369f314c6f9b73aa59f9eb38bd5ca99dfbbdd48f7d2e304377e5b7.jpg", + "image_caption": [ + "\"A hyperrealistic photo of fox astronaut, perfect face, artstation\"" + ], + "image_footnote": [], + "bbox": [ + 230, + 189, + 328, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5a05e9c607af55c062d85492723da54c305015052f89a6bf3ae261e01a9ebb3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 190, + 436, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2802d2ad5aa3e0613f924267ed3b1df9599a4d16aa0be8173d4fb14cb7f9467e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 190, + 545, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d88a1d3d373a34ed133ddf1ab1d5e01f9e623ddc8d1f09871d904e20dd39def7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 190, + 653, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee112dac7cd86a08f1cb8f84b4bd08294a42b04bba0984daf3394ec9cea4aaa7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 190, + 763, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cffa696eb135c8a6bd8d9db1eece800e019afdd9d895ae23e1d1e4fecc1de0f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 190, + 870, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/88aed34cc8fce6a13270bb3647dfedd284471c814722c2d73aec0fcd80b6cd3f.jpg", + "image_caption": [ + "\"A poodle wearing a baseball cap and holding a dictionary in hand\"" + ], + "image_footnote": [], + "bbox": [ + 230, + 273, + 326, + 348 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c61dc697f40e133c0948744834072aee2afea166f1f4d9da81fe21b2043b054a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 273, + 436, + 348 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6c2be08869aafead182b3a99722a5ef6bd8e67f5ebe1eb59ea17c56305dec90b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 273, + 545, + 348 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c20f5069bcc119d1bfcd1e24eac80ab6fe68622566b6cef6d01f718c9b6a10e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 273, + 653, + 348 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9494333f178d2054b331dfa7029502b30b69c4af76bf11d1d06c4caee5e4b0e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 273, + 763, + 348 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d5e8727c06927e472018ad6682530a3c88ca24151b1bcc443f1632709c988736.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 273, + 870, + 348 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/767c986b23b50af3ec467e281e6e182be2e8721010c92fe0fae17f817a5a5573.jpg", + "image_caption": [ + "\"A DSLR photo of a cat jumping over a fence, high-res\"" + ], + "image_footnote": [], + "bbox": [ + 230, + 356, + 326, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f3de0616d99be0965264fb0e9cea175eb4943ba181db7cae8d9abfbed8ad8ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 356, + 436, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c756f0536b9274f20d230c3f0f7fe509f8f73ed556e416083f64b2ed3ac1fc32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 356, + 545, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6b8d3915dd4d9f1db9226db95f51a357b3116d9e48b5a162e55af7da33b72ba9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 356, + 653, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/718d2356631f7e89b9989d6b342bfb19e0948cacbf0ca8dd52db759d94a101d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 356, + 763, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e430d7675b11c63f1599edbe97b9d0fe9aa5132118181f753077f2b085def92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 356, + 870, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/72200b3d12a6c4589659236250a96de56ff83b8891cdc74da9a74c2c17546696.jpg", + "image_caption": [ + "\"A blue Porsche 356 parked in front of a brick wall\"" + ], + "image_footnote": [], + "bbox": [ + 230, + 439, + 326, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9558855b7e62142bff08d080aea99866561af9c5b2b5ea0e00b7bbb1235c8cb7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 440, + 436, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/53a9506e177150ca358fe4b03b7d7cb1f4866e98c6eb9a1fb3a0d93a47cbdde2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 440, + 545, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ecfab9d3c6bad519d4d60292f46e964fce8b74b78a416c021f3a62ac1effc549.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 440, + 653, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/34439a0b6d48491b9a1e85c35db6f26833ff6dafec1e78a021612ee91ef0ee1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 440, + 763, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b4d7b8b131cb3ff407055a3d828c64017c9078e698eb248914072f26e0aa17e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 440, + 870, + 505 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51bc833455ced53b6bae2fb5d9b801db20508affd9b6e752b004ef46b25c8862.jpg", + "image_caption": [ + "\"A photograph of an astronaut riding a horse\"", + "Figure 4. Sample images generated by LCM, BOOT, Instaflow, Ours, SD with 25 sampling steps $(\\mathrm{SD}^{*})$ and SD with 1 sampling step $(\\mathrm{SD}^{\\dagger})$ . Images at the same row are sampled from the same text prompt, while images at the same column are from the same model. We use DPM-Solver [22] with guidance scale 7.5 to sample for Stable Diffusion" + ], + "image_footnote": [], + "bbox": [ + 230, + 522, + 326, + 597 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3fc4ba236e59f17378472457c5054115c817e078fd9ec3a2b9014b3e46d2365a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 523, + 436, + 597 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/47edaade28b7b17e8d421056103dedf98db42e9834571288a1aed386337e8fe8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 523, + 545, + 597 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9fe2b24dfa901decdc8f8c8780d92f25b2722760e3f782db62e48432d56c1774.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 523, + 653, + 597 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2cce683dbc060cbeca3d012aa87879dde8ec52aad7db6fb69ced25cef37bbb15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 523, + 763, + 597 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2944e461664cfdd2b50223ffdc22521eba30a97ae369d4114c25efe8f99cf33c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 522, + 870, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fective LoRA teacher, the student's early training outputs, despite being highly aligned with the text prompt, are overly saturated and unrealistic ( $2^{nd}$ column). These results indicate a fundamental breakdown in the student model's ability to mimic the teacher model's distribution, as can be further demonstrated based on the flat red curve in both FID and CLIP score in Fig. 5 (a) and (b).", + "bbox": [ + 75, + 656, + 468, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, even when employing the LoRA teacher, it is crucial to use a sufficiently large rank in the LoRA weight for stable distillation of the student model. Without this, the model tends to undergo severe mode collapse despite initially producing reasonably good results. The green curves in Fig. 5 clearly illustrate that, while the FID score of a student model with a 4-rank LoRA weight initially decreases, it deteriorates starting from 8K steps, resulting in lousy images in Fig. 6 ( $3^{rd}$ column). Additionally, the correspond", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing CLIP score behaves similarly where there is a slight rise before it worsens in the end. Besides, we also provide extra visuals for this ablative study in Supplementary.", + "bbox": [ + 498, + 656, + 890, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The importance of student parameterization. Apart from the LoRA teacher, the re-parameterization of the student (Sec. 3) also plays an important role in the SwiftBrush training. Without it, the student is unable to create realistic images and it only produces noisy results, as can be noticed from the $4^{th}$ column of Fig. 6. This can be further backed up by the orange curve in both FID and CLIP score in Fig. 5 (a) and (b), where student without re-parameterization performs consistently worst among all configurations.", + "bbox": [ + 498, + 710, + 892, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Interpolating Latent. We conduct a controlled experiment where the noise input $z$ remains the same while varying the input prompt $y$ (Fig. 7). Interestingly, the SwiftBrush-", + "bbox": [ + 498, + 854, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7813", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4e54008329a3726d42c9a4db0721ed675e41c16246872bb9b899ca1c7ed200af.jpg", + "image_caption": [ + "(a) FID curves of four configurations" + ], + "image_footnote": [], + "bbox": [ + 120, + 89, + 444, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a96d66bfdb07dec0ab90d570494aa5022ce35a4bc6cace3c35b17cfad03e28b6.jpg", + "image_caption": [ + "(b) CLIP curves of four configurations" + ], + "image_footnote": [], + "bbox": [ + 526, + 90, + 848, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56374b2b8e00c8ae4ea1f812a4d152dde9a6d323c8ede91b9f1df4384c491b87.jpg", + "image_caption": [ + "Figure 5. Ablation study on the impact of the LoRA teacher as well as the student parameterization. We train all four configurations with same hyperparameters described in Sec. 4 but only with 10K iterations to observe their training dynamics at the early phase. Then, we evaluate FID and CLIP score on COCO 2014 but with only a 5K subset of images. Here, Full means all of our proposed techniques are used. No Parameterization means we exclude the re-parameterization trick. Small-Rank LoRA means we use only 4-rank LoRA teacher. SDS means we remove the LoRA teacher and use only SDS loss." + ], + "image_footnote": [], + "bbox": [ + 81, + 366, + 171, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1278c181d97f8233cc8a83baf20c28b234c0760582307bbfbd2a6290a296d298.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 366, + 269, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2676c2a63ed719b00ee0eb3086fd3ef3ee531072e70fc188d81486a705c3f73a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 366, + 367, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c72a8c976337dcccbcb2e2bfc92d78297b8a2ad8238dd10a3159f5719f2a07dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 367, + 465, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65c964b6513a2a9cf10c08970cef03cfbaa8e50cb16615299a775255e533f260.jpg", + "image_caption": [ + "Full" + ], + "image_footnote": [], + "bbox": [ + 81, + 444, + 171, + 513 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b33d8b218629f196ece9c81e315ec48c3c28a93e540bcb0e6746ff3d2a515109.jpg", + "image_caption": [ + "SDS" + ], + "image_footnote": [], + "bbox": [ + 179, + 444, + 269, + 513 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1903b1873a776c9f4c50dd198a5a1976d0613f21da996bf24b71ee80195a624b.jpg", + "image_caption": [ + "Small-Rank LoRA", + "Figure 6. Visual results of the ablative study, where the first two rows are generated with the input prompt \"An oil painting of a train\", whereas the remaining two are sampled from \"An DSLR photo of a tiger in the city\"." + ], + "image_footnote": [], + "bbox": [ + 277, + 444, + 367, + 513 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95281ca8eb89ba6bc26561b51b94cf8cfbbdb31cf356cd8c8e163eea047a2551.jpg", + "image_caption": [ + "No Parameterization" + ], + "image_footnote": [], + "bbox": [ + 375, + 443, + 465, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "distilled student seemingly features enhanced control and editability compared to the teacher model due to its one-step nature. Additionally, we present further results of latent space interpolation in Supplementary. In this process, the student model, distilled from the Stable Diffusion teacher, demonstrates a seamless transition in its image generation.", + "bbox": [ + 75, + 599, + 467, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 76, + 707, + 323, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. In this work, we have introduced a new image-free distillation scheme, SwiftBrush, which leverages insights from text-to-3D synthesis to speed up text-to-image generation. Empirically, the results presented in this paper demonstrate the effectiveness of our approach in making text-to-image generation more efficient and accessible. We hope our work will encourage broader use of text-to-image generation systems by greatly lowering the inference cost.", + "bbox": [ + 75, + 741, + 467, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussion. SwiftBrush inevitably produces lower quality samples compared to the teacher which utilizes multi-", + "bbox": [ + 75, + 869, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b5a1e2da3fb42bb5a9fe0f7bf16b49d51850457dbe776a8108349ef52a90bdf5.jpg", + "image_caption": [ + "Figure 7. Results of interpolating the input prompts. The prompts used here are selected from a standard template: 'A DSLR photo of a {animal} {action}' Here, 'animal' is one of following: owl, raccoon, cat, or panda, while 'action' is one of following: reading a book, drinking a latte or eating a pizza." + ], + "image_footnote": [], + "bbox": [ + 504, + 366, + 694, + 588 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/df50ac732abc1f17f8bb88fe3977c5f8de8d64c5468da9cdab41da267283253a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 367, + 887, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "step sampling for inference. Another drawback is that the current design only focuses on single-step student model and cannot support few-step generation as did in Luo et al. [24] to improve image fidelity. In future work, we aim to investigate the feasibility of extending SwiftBrush to a few-step generation so that we can trade computation for quality. Furthermore, we also find it compelling to explore SwiftBrush-like training where only one teacher is required. Such an exploration could provide tremendous benefits in scenarios where the computational resource for training is limited. Finally, it is intriguing to see if techniques such as DreamBooth [35], ControlNet [49] or InstructPix2Pix[2] can be integrated with SwiftBrush, allowing instant generation of various applications.", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7814", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. eDiff-I: Text-to-Image Diffusion Models with an Ensemble of Expert Denoisers. arXiv preprint arXiv:2211.01324, 2022. 4", + "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. InstructPix2Pix: Learning to Follow Image Editing Instructions. IEEE Computer Vision and Pattern Recognition Conference, 2023. 8", + "[3] Eric R. Chan, Koki Nagano, Matthew A. Chan, Alexander W. Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. GeNVS: Generative Novel View Synthesis with 3D-Aware Diffusion Models. International Conference on Computer Vision, 2023. 2", + "[4] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T. Freeman, Michael Rubinstein, Yuzhen Li, and Dilip Krishnan. Muse: Text-To-Image Generation via Masked Generative Transformers. arXiv preprint arXiv:2301.00704, 2023. 3", + "[5] DeepFloyd. DeepFloyd IF. https://www deepfloyd.ai/deepfloyd-if, 2023.5", + "[6] Jean-Yves Franceschi, Mike Gartrell, Ludovic Dos Santos, Thibaut Issenhuth, Emmanuel de Bezenac, Mickael Chen, and Alain Rakotomamonjy. Unifying GANs and Score-Based Diffusion as Generative Particle Models. Advances in Neural Information Processing Systems, 2023. 3", + "[7] Jean-Yves Franceschi, Mike Gartrell, Ludovic Dos Santos, Thibaut Issenhuth, Emmanuel de Bezenac, Mickael Chen, and Alain Rakotomamonjy. Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models. Advances in Neural Information Processing Systems, 2023. 3", + "[8] Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and Lingpeng Kong. DiffuSeq: Sequence to Sequence Text Generation with Diffusion Models. International Conference on Learning Representations, 2023. 2", + "[9] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Lingjie Liu, and Josh Susskind. BOOT: Data-free Distillation of Denoising Diffusion Models with Bootstrapping. arXiv preprint arXiv:2306.05544, 2023. 3, 5", + "[10] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the Knowledge in a Neural Network. NIPS Deep Learning Workshop, 2014. 3", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. Advances in Neural Information Processing Systems, 2020. 2", + "[12] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video Diffusion Models. Advances in Neural Information Processing Systems, 2022. 2", + "[13] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Models. International Conference on Learning Representations, 2022. 4", + "[14] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-Shot Text-Guided Object Generation with Dream Fields. IEEE Computer Vision and Pattern Recognition Conference, 2022. 3", + "[15] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up GANs for Text-to-Image Synthesis. IEEE Computer Vision and Pattern Recognition Conference, 2023. 3, 5", + "[16] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. International Conference on Learning Representations, 2015. 5", + "[17] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A Versatile Diffusion Model for Audio Synthesis. International Conference on Learning Representations, 2021. 2", + "[18] Xiang Lisa Li, John Thickstun, Ishaan Gulrajani, Percy Liang, and Tatsunori Hashimoto. Diffusion-LM: Improves Controllable Text Generation. Advances in Neural Information Processing Systems, 2022. 2", + "[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-Resolution Text-to-3D Content Creation. IEEE Computer Vision and Pattern Recognition Conference, 2023. 3, 4", + "[20] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. European Conference on Computer Vision, 2014. 5", + "[21] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation. arXiv preprint arXiv:2309.06380, 2023. 2, 3, 5", + "[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps. Advances in Neural Information Processing Systems, 2022. 6, 7", + "[23] Eric Luhman and Troy Luhman. Knowledge Distillation in Iterative Generative Models for Improved Sampling Speed. arXiv preprint arXiv:2101.02388, 2021. 3", + "[24] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2, 3, 5, 8", + "[25] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik P. Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On Distillation of Guided Diffusion Models. IEEE Computer Vision and Pattern Recognition Conference, 2023. 2, 3, 5", + "[26] Alexander Quinn Nichol and Prafulla Dhariwal. Improved Denoising Diffusion Probabilistic Models. International Conference on Machine Learning, 2021. 2", + "[27] Junting Pan, Keqiang Sun, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. JourneyDB: A Benchmark for Generative Image Understanding." + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7815", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Advances in Neural Information Processing Systems, 2023. 5", + "[28] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion. International Conference on Learning Representations, 2023. 2, 3, 4", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning Transferable Visual Models from Natural Language Supervision. International Conference on Machine Learning, 2021. 3", + "[30] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 3", + "[31] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-Shot Text-to-Image Generation. International Conference on Machine Learning, 2021. 3", + "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv preprint arXiv:2204.06125, 2022. 2, 4", + "[33] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative Adversarial Text to Image Synthesis. International Conference on Machine Learning, 2016. 2", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. IEEE Computer Vision and Pattern Recognition Conference, 2022. 3, 5", + "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-image Diffusion Models for Subject-Driven Generation. IEEE Computer Vision and Pattern Recognition Conference, 2023. 2, 8", + "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 2022. 2, 4, 5", + "[37] Tim Salimans and Jonathan Ho. Progressive Distillation for Fast Sampling of Diffusion Models. International Conference on Learning Representations, 2022. 2, 3", + "[38] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. StyleGAN-T: Unlocking the Power of GANs for Fast Large-Scale Text-to-Image Synthesis. International Conference on Machine Learning, 2023. 3, 5", + "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jitsev. LAION-5B: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 2022. 3", + "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep Unsupervised Learning using Nonequilibrium Thermodynamics. International Conference on Machine Learning, 2015. 2", + "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. International Conference on Learning Representations, 2021. 2", + "[42] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency Models. International Conference on Machine Learning, 2023. 2, 3", + "[43] StabilityAI. Stable Diffusion. https://github.com/Stability-AI/stablediffusion, 2021.2.5", + "[44] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score Jacobian Chaining: Lifting Pretrained 2D Diffusion Models for 3D Generation. IEEE Computer Vision and Pattern Recognition Conference, 2023. 3, 4", + "[45] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-Fidelity and Diverse Text-to-3D Generation with Variational Score Distillation. Advances in Neural Information Processing Systems, 2023. 3, 4, 6", + "[46] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel View Synthesis with Diffusion Models. International Conference on Learning Representations, 2023. 2", + "[47] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis. arXiv preprint arXiv:2306.09341, 2023. 5, 6", + "[48] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris Metaxas. StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks. International Conference on Computer Vision, 2017. 2", + "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. International Conference on Computer Vision, 2023. 2, 8" + ], + "bbox": [ + 503, + 92, + 890, + 727 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7816", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/13853f95-872e-49e0-9add-f414db12f2de_model.json b/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/13853f95-872e-49e0-9add-f414db12f2de_model.json new file mode 100644 index 0000000000000000000000000000000000000000..784847e3f50126ae44278be12b3e84a4c2a7fdc6 --- /dev/null +++ b/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/13853f95-872e-49e0-9add-f414db12f2de_model.json @@ -0,0 +1,2673 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.131, + 0.757, + 0.175 + ], + "angle": 0, + "content": "SwiftBrush: One-Step Text-to-Image Diffusion Model with Variational Score Distillation" + }, + { + "type": "text", + "bbox": [ + 0.344, + 0.204, + 0.626, + 0.238 + ], + "angle": 0, + "content": "Thuan Hoang Nguyen Anh Tran VinAI Research, Vietnam" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.242, + 0.625, + 0.257 + ], + "angle": 0, + "content": "{v.thuannh5,v.anhtt152}@vinai.io" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.264, + 0.486, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.265, + 0.89, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.577, + 0.485, + 0.734 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.577, + 0.89, + 0.734 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.735, + 0.895, + 0.765 + ], + "angle": 0, + "content": "Figure 1. Samples generated in a single step by SwiftBrush. Our distillation method can produce images of high-fidelity roughly 20 times faster compared to Stable Diffusion." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.779, + 0.314, + 0.794 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Despite their ability to generate high-resolution and diverse images from text prompts, text-to-image diffusion models often suffer from slow iterative sampling processes. Model distillation is one of the most effective directions to accelerate these models. However, previous distillation methods fail to retain the generation quality while requir" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "ing a significant amount of images for training, either from real data or synthetically generated by the teacher model. In response to this limitation, we present a novel image-free distillation scheme named SwiftBrush. Drawing inspiration from text-to-3D synthesis, in which a 3D neural radiance field that aligns with the input prompt can be obtained from a 2D text-to-image diffusion prior via a specialized loss without the use of any 3D data ground-truth, our ap" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7807" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "proach re-purposes that same loss for distilling a pretrained multi-step text-to-image model to a student network that can generate high-fidelity images with just a single inference step. In spite of its simplicity, our model stands as one of the first one-step text-to-image generators that can produce images of comparable quality to Stable Diffusion without reliance on any training image data. Remarkably, SwiftBrush achieves an FID score of 16.67 and a CLIP score of 0.29 on the COCO-30K benchmark, achieving competitive results or even substantially surpassing existing state-of-the-art distillation techniques." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.287, + 0.21, + 0.303 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.473, + 0.599 + ], + "angle": 0, + "content": "Diffusion models are gaining significant attention from the research community as they have achieved remarkable results in various generative tasks, including image generation [11, 26, 40, 41], video synthesis [12], 3D modeling [3, 28, 46], audio generation [17], and text creation [8, 18]. Especially, text-to-image diffusion models, which combine the power of language models and high-quality diffusion models [32, 36, 43], have revolutionized the way that people create visual content based on textual descriptions. Synthetic images indistinguishable from real photos are now possible with only a few clicks. Additionally, with techniques like ControlNet [49] or DreamBooth [35], extra control to the generation outputs can be made to further boost the expressiveness of the text-to-image models and bring them closer to mass users. However, despite their impressive success, these models exhibit a noteworthy drawback of slow inference speed. The primary reason for this lies in the iterative nature of their sampling, presenting significant challenges for deployment on consumer devices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Several approaches have been proposed to improve the inference speed of diffusion-based text-to-image generation. However, the most effective direction is time-step distillation, which aims to reduce the number of sampling steps of the diffusion model with the network backbone unchanged and minimum reduction in generation quality. Starting from the pioneering work [37], more advanced and effective distillation methods have been proposed, and a few recent ones could train efficient student networks that only require one forward step to synthesize images from input captions. In this work, we are interested in the distillation-based approach for one-step text-to-image generation. Particularly, Meng et al. [25] employ a two-stage distillation, where the student tries to first match its output with the classifier-free guided output of the teacher and then bootstrap itself to reduce the inference steps gradually. Such a method is complex and time-consuming, requiring at least 108 A100 GPU days [21]. Based on the work of Song et al. [42], Latent Consistency Model (LCM) [24] enforces the consistency of all points on the ODE trajectory in the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.304 + ], + "angle": 0, + "content": "latent space of Stable Diffusion (SD) [43] and unlike [25], they directly combine the two distillation stages into one. Although both Guided Distillation and LCM significantly reduce the number of inference steps to 2 or 4, their 1-step inference yields blurry and unsatisfying results. More sophisticated, Instaflow [21] proposes a technique called \"reflow\" to straighten the sampling trajectory of the teacher before distilling. While their proposed techniques proven to be superior to the original diffusion framework in one-step text-to-image distillation, InstaFlow requires an expensive 4-stage training schedule. Also, it is noteworthy that the effectiveness of these methods relies heavily on a large number of text-image pairs, which is not always the case due to the limited accessibility of such data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.895, + 0.637 + ], + "angle": 0, + "content": "In this work, we aim to develop a novel distillation method for one-step text-to-image generation with higher quality and a more approachable training process characterized by an image-free and straightforward mechanism. Building upon the recent advancements in text-to-3D techniques, our work takes inspiration from their ability to generate high-quality 3D Neural Radiance Fields (NeRF) without utilizing 3D ground truth data. Such a remarkable achievement is facilitated by employing a powerful pretrained 2D text-to-image model to assess whether the image rendered from NeRF is realistic or not, akin to the discriminator in GANs. This observation underpins the motivation for our SwiftBrush approach, as we recognize that these principles can be innovatively adapted for text-to-image generation making image supervision unnecessary. SwiftBrush, therefore, emerges as a conceptually novel one-step text-to-image distillation process, uniquely capable of operating without 2D image supervision. This strategic adaptation not only opens up a change of view in the distillation of text-to-image diffusion models into a single-step generators but also points out the potential of blending principles from text-to-3D to text-to-image synthesis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.895, + 0.775 + ], + "angle": 0, + "content": "As a result, our method has successfully enabled the generation of high-quality images with remarkable details. Quantitatively, our one-step model achieves promising zero-shot results, 16.6 for the FID score and 0.29 for the CLIP score on the MS COCO 2014 dataset (30K images). Notably, this is the first time a distilled one-step model outperforms the previous approaches without a single image used for training. A demonstration of generated images from our distilled model is shown in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.792, + 0.642, + 0.807 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Text-to-Image Generation. Previously, text-to-image models relied on GANs and focused on only small-scale object-centric domains such as flowers and birds [33, 48]. Recently, these types of systems have advanced rapidly with the availability of web-scale datasets of text-image pairs" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7808" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.471 + ], + "angle": 0, + "content": "such as LAION5B [39], large language models such as T5 [30], and large vision-language models such as CLIP [29]. DALL-E [31] was the first autoregressive model to demonstrate a remarkable zero-shot ability in creating images from text captions by simply scaling up both the network and dataset size. Following that, a plethora of methods have been used for text-to-image generation, including masked generative transformers [4], GAN-based, and diffusion models. Based on Latent Diffusion Model [34], Stable Diffusion is an open-source diffusion-based generator that has gained widespread popularity among artists as well as researchers. Recently, text-to-image GANs are receiving attention again with notable work like StyleGAN-T [38], or GigaGAN [15]. While they are extremely fast as no iterative sampling is required, they lack behind diffusion models in terms of scalability and realism due to the fact that they suffer from instability and the infamous issue \"mode collapse\". To the best of our knowledge, so far only two above-mentioned work succeeded in applying GANs for text-to-image using tons of intricate techniques and auxiliary modules. Conversely, diffusion models can be easily scale up to generate high-quality and sharp samples, yet they require lengthy sampling process. Here, our objective is to create a large-scale generative model that combines the speed of GANs with the quality of diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.481, + 0.473, + 0.862 + ], + "angle": 0, + "content": "Text-to-3D Generation. Instead of learning a generative model directly from 3D supervision, many research works have utilized the rich knowledge of 2D prior for 3D generation tasks. For instance, one of the earliest works, Dream Fields [14], leverages CLIP [29] to guide the generated images so that rendered images from many camera views highly align with the text caption. On the other hand, pioneer works like DreamFusion [28] and Score Jacobian Chaining [44] (SJC) propose two different but equivalent frameworks where a 2D text-to-image generative prior is used to generate 3D objects from textual descriptions. Subsequent works seek to improve these text-to-3D methods in various ways. Among them, Magic3D [19] and ProlificDreamer [45] are two remarkable works that effectively enhance the generation quality of 3D assets. Magic3D consists of two stages where they first obtain a low-resolution 3D representation via a DreamFusion-style optimization and then refine its texture to get the final high-fidelity 3D mesh. Meanwhile, ProlificDreamer proposes Variational Score Distillation (VSD) involving a secondary teacher to bridge the gap between the teacher and the 3D NeRF. Influenced by the latter work, our method adapts the same technique for one-step text-to-image generation. Similarly, Score GAN [6] and Diff-Instruct [7] use a VSD-like loss to distill a pretrained diffusion model to a one-step generator." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Knowledge Distillation introduced by Hinton et al. [10] is one of the methods falling under the umbrella of transfer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.381 + ], + "angle": 0, + "content": "learning. This algorithm draws its main inspiration from the human learning process, where knowledge is conveyed from a more knowledgeable teacher to a less knowledgeable student. In the context of diffusion models, works have been proposed for faster sampling of student models from a pre-trained diffusion teacher. The most straightforward approach involves direct distillation [23], wherein a student model is supervised from the teacher's output after sampling for 50-1000 steps, which can be prohibitive. Recent work [37, 42] avoids the lengthy sampling during distillation by various bootstrapping methods. Afterward, a series of methods [21, 24, 25] adapt these methods into text-to-image settings and inherit their limitation of being image-dependent in training. In contrast, a concurrent work [9] learns a time-conditioned model to predict the outcome of the teacher at any given time step, all without the need for image supervision. Likewise, our approach is similarly image-free yet achieves significantly better results and offers a much simpler distillation design." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.391, + 0.673, + 0.408 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.416, + 0.63, + 0.433 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.895, + 0.568 + ], + "angle": 0, + "content": "Diffusion Models consist of two processes: a forward process that gradually adds noise and another reverse process to predict the distribution of gradually denoised data points. The forward process incrementally introduces noise to an initial data point \\( x_0 \\), drawn from the distribution \\( q_0(x_0) \\), resulting in \\( q_t(x_t|x_0) = \\mathcal{N}(\\alpha_t x_0, \\sigma_t^2 I) \\), where \\( \\{(\\alpha_t, \\sigma_t)\\}_{t=1}^T \\) is the noise schedule. The data point \\( x_t \\) at each time-step \\( t \\) is drawn from \\( q_t(x_t|x_0) \\) using a noise \\( \\epsilon \\sim \\mathcal{N}(0, I) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.578, + 0.892, + 0.593 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\alpha_ {t} x _ {0} + \\sigma_ {t} \\epsilon . \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.892, + 0.631 + ], + "angle": 0, + "content": "After \\(T\\) time steps, the final data point is expected to be pure Gaussian noise \\(x_{T}\\sim \\mathcal{N}(0,I)\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.895, + 0.709 + ], + "angle": 0, + "content": "Conversely, the backward process starts from \\( x_{T} \\sim \\mathcal{N}(0, I) \\) and iteratively denoises for \\( T \\) steps where at each step \\( t \\), noisy variable \\( x_{t+1} \\) is converted to less-noisy variable \\( x_{t} \\), guided by a model \\( \\epsilon_{\\psi} \\) that predicts the noise \\( \\epsilon \\) in Eq. (1). The network weights \\( \\psi \\) are trained by minimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.721, + 0.892, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {u n c o n d} (\\psi) = \\mathbb {E} _ {t, \\epsilon \\in \\mathcal {N} (0, 1)} \\| \\epsilon_ {\\psi} (x _ {t}, t) - \\epsilon \\| _ {2} ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.743, + 0.822, + 0.758 + ], + "angle": 0, + "content": "where \\(t\\) is uniformly samples within \\(\\{1,\\dots ,T\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.758, + 0.892, + 0.835 + ], + "angle": 0, + "content": "While common diffusion models operate in the pixel space, Latent Diffusion Models (LDMs)[34] model the diffusion process in the latent space of a pre-trained regularized autoencoder, which features representations of smaller size, thus enhancing both training and inference efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Text-to-Image Diffusion Models. In contrast to unconditional diffusion models that generate outputs freely, text-conditioned diffusion models steer the sampling process using an extra prompt \\( y \\). This guides the model to produce" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7809" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "outputs that are not only photorealistic but also closely adhere to the provided text description. The objective to train such models is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.147, + 0.469, + 0.166 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d i f f} (\\psi) = \\mathbb {E} _ {t, y, \\epsilon \\in \\mathcal {N} (0, 1)} \\| \\epsilon_ {\\psi} (x _ {t}, t, y) - \\epsilon \\| _ {2} ^ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.175, + 0.47, + 0.31 + ], + "angle": 0, + "content": "which is slightly different from the unconditional diffusion loss Eq. (2). Due to the prompt conditioning, the model can deliver more controllable generation compared to its unconditional counterparts. Nevertheless, the implementations of many leading methods [1, 32, 36] remain inaccessible to the public. Stable Diffusion, which primarily utilizes the LDM framework, stands out as the first openly available large-scale model, significantly propelling text-to-image synthesis's widespread adoption and versatility." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.471, + 0.484 + ], + "angle": 0, + "content": "Score Distillation Sampling (SDS) is a distillation technique effectively applied in generating 3D assets [19, 28, 44]. It utilizes a pre-trained text-to-image diffusion model, which predicts diffusion noise from text condition \\(y\\), denoted as \\(\\epsilon_{\\psi}(x_t,t,y)\\). The method optimizes a single 3D NeRF, parameterized by \\(\\theta\\), to align with a given text prompt. Given camera parameters \\(c\\), a differentiable rendering function \\(g(\\cdot ,c)\\) is used to render an image at the camera view \\(c\\) from the 3D NeRF. Here, the rendered image \\(g(\\theta ,c)\\) is utilized to optimize the weight \\(\\theta\\) through a loss function whose gradient can be approximated by:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.493, + 0.47, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {S D S} = \\mathbb {E} _ {t, \\epsilon , c} \\left[ w (t) (\\epsilon_ {\\psi} (x _ {t}, t, y) - \\epsilon) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} \\right] (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.469, + 0.582 + ], + "angle": 0, + "content": "where \\(t\\sim \\mathcal{U}(0.02T,0.98T)\\) \\(T\\) is the maximum timesteps of the diffusion model, \\(\\epsilon \\sim \\mathcal{N}(0,I)\\) \\(x_{t} = \\alpha_{t}g(\\theta ,c) + \\sigma_{t}\\epsilon\\) \\(y\\) is the input text, and \\(w(t)\\) is a weighting function" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.469, + 0.658 + ], + "angle": 0, + "content": "Despite the advancements in text-to-3D synthesis, empirical studies [28, 44] indicate that SDS often encounters issues like over-saturation, over-smoothing, and reduced diversity. The same degradations can be observed if we naively apply SDS in our framework, as shown in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.664, + 0.47, + 0.71 + ], + "angle": 0, + "content": "Variational Score Distillation (VSD) is introduced in ProlificDreamer [45] to address previously mentioned issues of SDS by modifying the loss slightly:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.72, + 0.469, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\theta} \\mathcal {L} _ {V S D} = \\mathbb {E} _ {t, \\epsilon , c} [ w (t) (\\epsilon_ {\\psi} (x _ {t}, t, y) \\\\ \\left. - \\epsilon_ {\\phi} (x _ {t}, t, y, c)) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} \\right] \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Again, \\(x_{t} = \\alpha_{t}g(\\theta ,c) + \\sigma_{t}\\epsilon\\) is the noisy observation of the rendered image at camera view \\(c\\). VSD sets itself apart from SDS by introducing an additional score function tailored to the images rendered from the 3D NeRF at the camera pose \\(c\\). This score is derived by fine-tuning a diffusion model \\(\\epsilon_{\\phi}(x_t,t,y,c)\\) with the diffusion loss described below:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.88, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\epsilon_ {\\phi}} \\mathbb {E} _ {t, c, \\epsilon} \\| \\epsilon_ {\\phi} (x _ {t}, t, y, c) - \\epsilon \\| _ {2} ^ {2} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "As proposed in ProlificDreamer, \\(\\epsilon_{\\phi}\\) is parameterized by Low-Rank Adaption [13] (LoRA) and initialized from the pre-trained diffusion model \\(\\epsilon_{\\psi}\\) with some added layers for conditioning the camera view \\(c\\). Note that at each iteration \\(i\\) of the optimization process, \\(\\epsilon_{\\phi}\\) needs to adapt with the current distribution of \\(\\theta\\). Therefore, ProlificDreamer interleaves between finetuning \\(\\epsilon_{\\phi}\\) and optimizing \\(\\theta\\). With these algorithmic enhancements, ProlificDreamer has significantly advanced its capabilities, enabling the generation of NeRFs and the creation of exceptional textured meshes. Such improvement directly inspires us to adapt VSD to the task of one-step text-to-image diffusion-based distillation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.282, + 0.625, + 0.297 + ], + "angle": 0, + "content": "3.2. SwiftBrush" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.892, + 0.464 + ], + "angle": 0, + "content": "Motivation. While SDS and VSD are designed explicitly for the text-to-3D generation task, they loosely connect to that goal via the rendered image \\( g(\\theta, c) \\) of the 3D NeRF. As a matter of fact, instead of using NeRF rendering, we can replace it with any function that outputs 2D images to suit our needs. Inspired by this motivation, we propose to substitute the NeRF rendering \\( g(\\theta, c) \\) with a text-to-image generator that can directly synthesize a text-guided image in one step, effectively converting the text-to-3D generation training into one-step diffusion model distillation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.729 + ], + "angle": 0, + "content": "Design Space. We adopt the same approach as [45], with modifications to the design space to suit our task better. To begin with, we make use of two teacher models: a pretrained text-to-image teacher \\(\\epsilon_{\\psi}\\) and one additional LoRA teacher \\(\\epsilon_{\\phi}\\). Also, we remove the conditioning on camera view \\(c\\) from the LoRA teacher as it is unnecessary in our case, and we use classifier-free guidance for both teachers. Then, we replace the NeRF, which is overfitted for a specific user-provided prompt in a text-to-3D setting, with a generalized one-step text-to-image student model \\(f_{\\theta}(z,y)\\). Our student model \\(f_{\\theta}\\) takes a random Gaussian noise \\(z\\) and a text prompt \\(y\\) as inputs. Both the LoRA teacher and student model are initialized with the weight of the text-to-image teacher. Next, we train both the student model and the LoRA teacher alternately using Eq. (5) and Eq. (6) while freezing the text-to-image teacher. A pseudo-code and system figure can be viewed in Alg. 1 and Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Student Parameterization. Given a pretrained text-to-image diffusion model \\(\\epsilon_{\\theta}\\), it is possible to directly use its output for the student model, i.e., \\(f_{\\theta}(z,y) = \\epsilon_{\\theta}(z,T,y)\\), where \\(T\\) is the maximum timestep of the pretrained model. However, in our case, the pretrained model of choice is Stable Diffusion, which is inherently designed to predict the added noise \\(\\epsilon\\). In contrast, our objective is to refine the student model so that it predicts a clean and noise-free \\(x_0\\). Therefore, such a naive method results in a large domain gap between what we want the student to learn and the output of the student. For the sake of easy training, we empiri" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7810" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.092, + 0.324, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 SwiftBrush Distillation" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.111, + 0.47, + 0.187 + ], + "angle": 0, + "content": "1: Require: a pretrained text-to-image teacher \\(\\epsilon_{\\psi}\\), a LoRA teacher \\(\\epsilon_{\\phi}\\), a student model \\(f_{\\theta}\\), two learning rates \\(\\eta_{1}\\) and \\(\\eta_{2}\\), a weighting function \\(\\omega\\), a prompts dataset \\(Y\\), the maximum number of time steps \\(T\\) and the noise schedule \\(\\{(\\alpha_t, \\sigma_t)\\}_{t=1}^T\\) of the teacher model" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.187, + 0.282, + 0.2 + ], + "angle": 0, + "content": "2: Initialize: \\(\\phi \\gets \\psi, \\theta \\gets \\psi\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.202, + 0.265, + 0.216 + ], + "angle": 0, + "content": "3: while not converged do" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.217, + 0.345, + 0.232 + ], + "angle": 0, + "content": "4: Sample input noise \\(z\\sim \\mathcal{N}(0,I)\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.233, + 0.348, + 0.246 + ], + "angle": 0, + "content": "5: Sample text caption input \\(y\\sim Y\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.247, + 0.384, + 0.262 + ], + "angle": 0, + "content": "6: Compute student output \\(\\hat{x}_0 = f_\\theta (z,y)\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.263, + 0.385, + 0.277 + ], + "angle": 0, + "content": "7: Sample timestep \\(t \\sim \\mathcal{U}(0.02T, 0.98T)\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.278, + 0.349, + 0.292 + ], + "angle": 0, + "content": "8: Sample added noise \\(\\epsilon \\sim \\mathcal{N}(0,I)\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.294, + 0.399, + 0.307 + ], + "angle": 0, + "content": "9: Compute noisy sample \\(\\hat{x}_t = \\alpha_t\\hat{x}_0 + \\sigma_t\\epsilon\\)" + }, + { + "type": "text", + "bbox": [ + 0.083, + 0.307, + 0.464, + 0.324 + ], + "angle": 0, + "content": "10: \\(\\theta \\gets \\theta -\\eta_{1}\\left[\\omega (t)\\left(\\epsilon_{\\psi}(\\hat{x}_{t},t,y) - \\epsilon_{\\phi}(\\hat{x}_{t},t,y)\\right)\\frac{\\partial\\hat{x}_{0}}{\\partial\\theta}\\right]\\)" + }, + { + "type": "text", + "bbox": [ + 0.083, + 0.324, + 0.329, + 0.337 + ], + "angle": 0, + "content": "11: Sample timestep \\(t^{\\prime}\\sim \\mathcal{U}(0,T)\\)" + }, + { + "type": "text", + "bbox": [ + 0.083, + 0.338, + 0.354, + 0.352 + ], + "angle": 0, + "content": "12: Sample added noise \\(\\epsilon^{\\prime}\\sim \\mathcal{N}(0,I)\\)" + }, + { + "type": "text", + "bbox": [ + 0.083, + 0.353, + 0.417, + 0.367 + ], + "angle": 0, + "content": "13: Compute noisy sample \\(\\hat{x}_{t'} = \\alpha_{t'}\\hat{x}_0 + \\sigma_{t'}\\epsilon'\\)" + }, + { + "type": "text", + "bbox": [ + 0.083, + 0.367, + 0.372, + 0.383 + ], + "angle": 0, + "content": "14: \\(\\phi \\gets \\phi -\\eta_{2}\\nabla_{\\phi}\\| \\epsilon_{\\phi}(\\hat{x}_{t^{\\prime}},t^{\\prime},y) - \\epsilon^{\\prime}\\|^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.083, + 0.384, + 0.177, + 0.396 + ], + "angle": 0, + "content": "15: end while" + }, + { + "type": "text", + "bbox": [ + 0.083, + 0.398, + 0.319, + 0.413 + ], + "angle": 0, + "content": "16: return trained student model \\( f_{\\theta} \\)" + }, + { + "type": "list", + "bbox": [ + 0.083, + 0.111, + 0.47, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.442, + 0.416, + 0.459 + ], + "angle": 0, + "content": "cally re-parameterize the student output as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.467, + 0.374, + 0.5 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} (z, y) = \\frac {z - \\sigma_ {T} \\epsilon_ {\\theta} (z , T , y)}{\\alpha_ {T}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.598 + ], + "angle": 0, + "content": "which is a realization of Eq. (1) if we set \\( t = T \\), \\( x_{t} = z \\), \\( \\epsilon \\approx \\epsilon_{\\theta}(z,T,y) \\) and \\( x_0 \\approx f_\\theta (z,y) \\). So basically, this reparameterization transforms the noise-prediction output of the pretrained model into the \"predicted \\( x_0 \\)\" form, which is empirically demonstrated in Sec. 4.3 to be helpful for the student model to learn." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.612, + 0.21, + 0.629 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.637, + 0.268, + 0.654 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.47, + 0.865 + ], + "angle": 0, + "content": "Evaluation Metrics. We conduct our evaluation process by assessing the performance of SwiftBrush against several well-established methods, such as Guided Distillation [25], Instaflow [21], LCM [24], and BOOT [9], across two zero-shot text-to-image benchmarks: COCO 2014 [20] and Human Preference Score v2 [47] (HPSv2). On the COCO 2014, we follow the conventional evaluation protocol [15, 34, 36, 38], computing FID and CLIP score on a 30K subset of its validation set as our primary metrics. Here, we use ViT-G/14 as the backbone for evaluating CLIP scores. In the case of HPSv2, we adopt their evaluation procedure to assess text-image alignment across four different aspects: anime, concept-art, paintings, and photos." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Experiment Protocols. In these experiments, we rely on Stable Diffusion v2.1 [43] for our two teacher models, and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "whenever possible, we directly utilize the reported values from other research. As for InstaFlow and LCM, we use the their provided pretrained models to generate images for evaluation. Conversely, we are compelled to re-implement BOOT since they did not provide any metrics for their text-to-image models, and their codebase remains entirely closed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.204, + 0.893, + 0.522 + ], + "angle": 0, + "content": "Implementation Details. Similar to prior research, we use student models with the same architectures as those of the teachers and initialize all of the student parameters with the teacher model. During training, a guidance scale of 4.5 is used for both teachers. For the LoRA teacher model, we apply a learning rate of 1e-3, a LoRA rank of 64, and an alpha value of 108. Meanwhile, the student model's learning rate is set at 1e-6 with exponential moving average (EMA) used every iteration. Adam optimizer [16] and a batch size of 64 (16 per GPU) are used to train both the student and LoRA teacher for around 65K iterations. Thanks to the image-free nature of SwiftBrush, we do not require access to any image training set, and only the prompts are needed to distill the generator. Here, we utilize 4M captions from the large-scale text-image dataset JourneyDB [27], which we subsequently deduplicated to 2M. Regarding BOOT, we closely follow their implementation guidelines to the best of our abilities and train the student model until its convergence, except that we use Stable Diffusion v2.1 instead of DeepFloyd IF [5] as the teacher. Here, its blurry results (Fig. 4) compared to the original work could be due to the different teacher." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.531, + 0.593, + 0.546 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.892, + 0.865 + ], + "angle": 0, + "content": "Quantitative Results. We first evaluate the proposed method on standard text-to-image generation benchmarks. The quantitative comparison with the previous distillation, both image-dependent and image-free, and the original Stable Diffusion teacher is shown in Tab. 1. Despite being trained with only text captions, SwiftBrush significantly improves the quality of 1-step inference and achieves better photo-realism as well as text-image alignment against other methods, with exceptions of InstaFlow and Stable Diffusion (25 sampling steps) in FID score. However, it is worth noting that Instaflow is so complex, requiring a 4-stage distillation process, while ours is end-to-end and straightforward. We further conduct a quantitative evaluation with the HPSv2 score, a metric trained to capture human preference. Again, we obtain the best score in all aspects among all mentioned works, which is only a slight degradation compared to the original Stable Diffusion teacher (Tab. 2). Moreover, according to the anonymous survey carried out by us (Fig. 3), our method outperforms Instaflow in terms of image fidelity and text-to-image accuracy." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Qualitative Results. We report the visual comparison between our method and the mentioned baseline methods" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7811" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.087, + 0.868, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.293, + 0.896, + 0.366 + ], + "angle": 0, + "content": "Figure 2. SwiftBrush overview. Our system allows training a one-step text-to-image student network from a frozen pretrained teacher with the help of an additional trainable LORA teacher. The student network takes inputs a text prompt and random noise. The output of the student is then added noise and sent to two teachers, alongside the prompt and a randomly drawn timestep, to compute the gradient of VSD loss, which backpropagates back to the student. Apart from that, the LoRA teacher is also updated using diffusion loss. Similar to ProlificDreamer [45], we update the LoRA teacher (green flow) and the student (red flow) alternatively." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.381, + 0.456, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.459, + 0.471, + 0.515 + ], + "angle": 0, + "content": "Figure 3. User Study. We ask 400 users to compare the generated images of InstaFlow and our method using 30 random text prompts. We report the rate at which each method is selected as better for image quality and text-to-image accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.528, + 0.47, + 0.801 + ], + "angle": 0, + "content": "for the one-step text-to-image generation task. As seen in Fig. 4, LCM outputs blurry and over-smoothed images in the 1-step regime. BOOT faithfully produces images that match well with the text caption, but its results still fall short in terms of fidelity. In contrast, a much better quality can be observed in the image produced by Instaflow and SwiftBrush. InstaFlow, nevertheless, cannot handle complex or unusual input prompts properly, which can be noticed in the third and fifth images at the third column of Fig. 4. On top of that, it occasionally produces non-realistic results even for simple prompts like in Fig. 4 at the fourth image of the third column. When, in fact, the outputs generated by SwiftBrush are not only of exceptional quality but also exhibit a remarkable alignment between text descriptions and the corresponding images. Moreover, our results are comparable with those sampled from Stable Diffusion with 25 steps; whereas it is understandable that with only one step, Stable Diffusion cannot create outputs of good quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.81, + 0.182, + 0.827 + ], + "angle": 0, + "content": "4.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The importance of LoRA teacher. In Fig. 6, the importance of including the LoRA teacher in the training process is clearly highlighted. We compare the student outputs sampled with different input noises after 10K iteration when us" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.38, + 0.905, + 0.508 + ], + "angle": 0, + "content": "
MethodStepsFID-30K ↓CLIP-30K ↑
Guided Distillation†137.30.27
LCM†135.560.24
Instaflow113.10†0.28§
BOOT‡148.200.26
Ours116.670.29
SD 2.1*2513.450.30
SD 2.1*1202.140.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.513, + 0.895, + 0.639 + ], + "angle": 0, + "content": "Table 1. Comparison of our method against other works based on FID metric and CLIP score on the COCO 2014 dataset. \\(^\\dagger\\) means that we obtain the numbers from the corresponding papers. \\(\\S\\) means that we obtain the numbers using the provided pretrained models of the corresponding papers. \\(^\\ddagger\\) means that we re-implement the work and report the numbers. * means that we report the number using the pretrained models. We use DPM-Solver [22] with guidance scale 7.5 to sample for SD. Bold and underlined numbers are best and second best for one-step models." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.645, + 0.899, + 0.757 + ], + "angle": 0, + "content": "
ModelsHuman Preference Score v2 ↑
AnimePhotoConcept ArtPaintings
LCM†22.6122.7122.7422.91
InstaFlow†25.9826.3225.7925.93
BOOT‡25.2925.1624.4024.61
Ours26.9127.2126.3226.37
SD 2.1*27.4826.8926.8627.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.761, + 0.895, + 0.845 + ], + "angle": 0, + "content": "Table 2. Comparison of our method against other works based on HPSv2 score in 1-step regime. \\(\\dagger\\) means that we obtain the score using the provided pretrained models of the corresponding papers. \\(\\ddagger\\) means that we re-implement the work and report the score by ourselves. * means that we obtain the score from [47]. Bold and underlined numbers are best and second best, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "ing the same text prompts. With the guidance of the LoRA teacher, the student progressively learns to create realistic images (\\(1^{st}\\) column). In contrast, in the absence of an ef" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7812" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.135, + 0.223, + 0.16 + ], + "angle": 0, + "content": "\"A photo of one ice cream ball in a luxurious plate, bokeh\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.093, + 0.292, + 0.104 + ], + "angle": 0, + "content": "LCM" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.108, + 0.329, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.094, + 0.407, + 0.104 + ], + "angle": 0, + "content": "BOOT" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.108, + 0.438, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.093, + 0.526, + 0.104 + ], + "angle": 0, + "content": "InstaFlow" + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.108, + 0.547, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.094, + 0.619, + 0.105 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.108, + 0.655, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.093, + 0.725, + 0.104 + ], + "angle": 0, + "content": "SD*" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.108, + 0.764, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.809, + 0.093, + 0.834, + 0.104 + ], + "angle": 0, + "content": "SD†" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.108, + 0.872, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.212, + 0.201, + 0.249 + ], + "angle": 0, + "content": "\"A hyperrealistic photo of fox astronaut, perfect face, artstation\"" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.19, + 0.329, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.191, + 0.437, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.191, + 0.546, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.191, + 0.655, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.191, + 0.764, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.191, + 0.872, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.297, + 0.214, + 0.332 + ], + "angle": 0, + "content": "\"A poodle wearing a baseball cap and holding a dictionary in hand\"" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.274, + 0.328, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.274, + 0.437, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.274, + 0.546, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.274, + 0.655, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.274, + 0.764, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.274, + 0.872, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.384, + 0.225, + 0.41 + ], + "angle": 0, + "content": "\"A DSLR photo of a cat jumping over a fence, high-res\"" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.357, + 0.328, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.357, + 0.437, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.357, + 0.546, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.357, + 0.655, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.357, + 0.764, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.357, + 0.872, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.47, + 0.216, + 0.494 + ], + "angle": 0, + "content": "\"A blue Porsche 356 parked in front of a brick wall\"" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.44, + 0.328, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.441, + 0.437, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.441, + 0.546, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.441, + 0.655, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.441, + 0.764, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.441, + 0.872, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.551, + 0.199, + 0.575 + ], + "angle": 0, + "content": "\"A photograph of an astronaut riding a horse\"" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.523, + 0.328, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.524, + 0.437, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.524, + 0.546, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.524, + 0.655, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.524, + 0.764, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.523, + 0.872, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.602, + 0.893, + 0.644 + ], + "angle": 0, + "content": "Figure 4. Sample images generated by LCM, BOOT, Instaflow, Ours, SD with 25 sampling steps \\((\\mathrm{SD}^{*})\\) and SD with 1 sampling step \\((\\mathrm{SD}^{\\dagger})\\). Images at the same row are sampled from the same text prompt, while images at the same column are from the same model. We use DPM-Solver [22] with guidance scale 7.5 to sample for Stable Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.469, + 0.763 + ], + "angle": 0, + "content": "fective LoRA teacher, the student's early training outputs, despite being highly aligned with the text prompt, are overly saturated and unrealistic (\\(2^{nd}\\) column). These results indicate a fundamental breakdown in the student model's ability to mimic the teacher model's distribution, as can be further demonstrated based on the flat red curve in both FID and CLIP score in Fig. 5 (a) and (b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In addition, even when employing the LoRA teacher, it is crucial to use a sufficiently large rank in the LoRA weight for stable distillation of the student model. Without this, the model tends to undergo severe mode collapse despite initially producing reasonably good results. The green curves in Fig. 5 clearly illustrate that, while the FID score of a student model with a 4-rank LoRA weight initially decreases, it deteriorates starting from 8K steps, resulting in lousy images in Fig. 6 (\\(3^{rd}\\) column). Additionally, the correspond" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.657, + 0.892, + 0.703 + ], + "angle": 0, + "content": "ing CLIP score behaves similarly where there is a slight rise before it worsens in the end. Besides, we also provide extra visuals for this ablative study in Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.711, + 0.893, + 0.848 + ], + "angle": 0, + "content": "The importance of student parameterization. Apart from the LoRA teacher, the re-parameterization of the student (Sec. 3) also plays an important role in the SwiftBrush training. Without it, the student is unable to create realistic images and it only produces noisy results, as can be noticed from the \\(4^{th}\\) column of Fig. 6. This can be further backed up by the orange curve in both FID and CLIP score in Fig. 5 (a) and (b), where student without re-parameterization performs consistently worst among all configurations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Interpolating Latent. We conduct a controlled experiment where the noise input \\( z \\) remains the same while varying the input prompt \\( y \\) (Fig. 7). Interestingly, the SwiftBrush-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7813" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.09, + 0.445, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.255, + 0.381, + 0.267 + ], + "angle": 0, + "content": "(a) FID curves of four configurations" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.091, + 0.849, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.255, + 0.789, + 0.267 + ], + "angle": 0, + "content": "(b) CLIP curves of four configurations" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.284, + 0.892, + 0.354 + ], + "angle": 0, + "content": "Figure 5. Ablation study on the impact of the LoRA teacher as well as the student parameterization. We train all four configurations with same hyperparameters described in Sec. 4 but only with 10K iterations to observe their training dynamics at the early phase. Then, we evaluate FID and CLIP score on COCO 2014 but with only a 5K subset of images. Here, Full means all of our proposed techniques are used. No Parameterization means we exclude the re-parameterization trick. Small-Rank LoRA means we use only 4-rank LoRA teacher. SDS means we remove the LoRA teacher and use only SDS loss." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.367, + 0.172, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.367, + 0.27, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.367, + 0.369, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.368, + 0.466, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.445, + 0.172, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.516, + 0.133, + 0.523 + ], + "angle": 0, + "content": "Full" + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.445, + 0.27, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.517, + 0.231, + 0.524 + ], + "angle": 0, + "content": "SDS" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.445, + 0.368, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.516, + 0.356, + 0.524 + ], + "angle": 0, + "content": "Small-Rank LoRA" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.444, + 0.466, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.516, + 0.458, + 0.524 + ], + "angle": 0, + "content": "No Parameterization" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.529, + 0.47, + 0.585 + ], + "angle": 0, + "content": "Figure 6. Visual results of the ablative study, where the first two rows are generated with the input prompt \"An oil painting of a train\", whereas the remaining two are sampled from \"An DSLR photo of a tiger in the city\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.468, + 0.69 + ], + "angle": 0, + "content": "distilled student seemingly features enhanced control and editability compared to the teacher model due to its one-step nature. Additionally, we present further results of latent space interpolation in Supplementary. In this process, the student model, distilled from the Stable Diffusion teacher, demonstrates a seamless transition in its image generation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.708, + 0.325, + 0.723 + ], + "angle": 0, + "content": "5. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.468, + 0.862 + ], + "angle": 0, + "content": "Conclusion. In this work, we have introduced a new image-free distillation scheme, SwiftBrush, which leverages insights from text-to-3D synthesis to speed up text-to-image generation. Empirically, the results presented in this paper demonstrate the effectiveness of our approach in making text-to-image generation more efficient and accessible. We hope our work will encourage broader use of text-to-image generation systems by greatly lowering the inference cost." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Discussion. SwiftBrush inevitably produces lower quality samples compared to the teacher which utilizes multi-" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.367, + 0.695, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.368, + 0.888, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.594, + 0.892, + 0.664 + ], + "angle": 0, + "content": "Figure 7. Results of interpolating the input prompts. The prompts used here are selected from a standard template: 'A DSLR photo of a {animal} {action}' Here, 'animal' is one of following: owl, raccoon, cat, or panda, while 'action' is one of following: reading a book, drinking a latte or eating a pizza." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "step sampling for inference. Another drawback is that the current design only focuses on single-step student model and cannot support few-step generation as did in Luo et al. [24] to improve image fidelity. In future work, we aim to investigate the feasibility of extending SwiftBrush to a few-step generation so that we can trade computation for quality. Furthermore, we also find it compelling to explore SwiftBrush-like training where only one teacher is required. Such an exploration could provide tremendous benefits in scenarios where the computational resource for training is limited. Finally, it is intriguing to see if techniques such as DreamBooth [35], ControlNet [49] or InstructPix2Pix[2] can be integrated with SwiftBrush, allowing instant generation of various applications." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7814" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. eDiff-I: Text-to-Image Diffusion Models with an Ensemble of Expert Denoisers. arXiv preprint arXiv:2211.01324, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. InstructPix2Pix: Learning to Follow Image Editing Instructions. IEEE Computer Vision and Pattern Recognition Conference, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.472, + 0.338 + ], + "angle": 0, + "content": "[3] Eric R. Chan, Koki Nagano, Matthew A. Chan, Alexander W. Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. GeNVS: Generative Novel View Synthesis with 3D-Aware Diffusion Models. International Conference on Computer Vision, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.472, + 0.422 + ], + "angle": 0, + "content": "[4] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T. Freeman, Michael Rubinstein, Yuzhen Li, and Dilip Krishnan. Muse: Text-To-Image Generation via Masked Generative Transformers. arXiv preprint arXiv:2301.00704, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.468, + 0.451 + ], + "angle": 0, + "content": "[5] DeepFloyd. DeepFloyd IF. https://www deepfloyd.ai/deepfloyd-if, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.47, + 0.522 + ], + "angle": 0, + "content": "[6] Jean-Yves Franceschi, Mike Gartrell, Ludovic Dos Santos, Thibaut Issenhuth, Emmanuel de Bezenac, Mickael Chen, and Alain Rakotomamonjy. Unifying GANs and Score-Based Diffusion as Generative Particle Models. Advances in Neural Information Processing Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.47, + 0.605 + ], + "angle": 0, + "content": "[7] Jean-Yves Franceschi, Mike Gartrell, Ludovic Dos Santos, Thibaut Issenhuth, Emmanuel de Bezenac, Mickael Chen, and Alain Rakotomamonjy. Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models. Advances in Neural Information Processing Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.47, + 0.661 + ], + "angle": 0, + "content": "[8] Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and Lingpeng Kong. DiffuSeq: Sequence to Sequence Text Generation with Diffusion Models. International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.663, + 0.47, + 0.717 + ], + "angle": 0, + "content": "[9] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Lingjie Liu, and Josh Susskind. BOOT: Data-free Distillation of Denoising Diffusion Models with Bootstrapping. arXiv preprint arXiv:2306.05544, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.719, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[10] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the Knowledge in a Neural Network. NIPS Deep Learning Workshop, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.761, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. Advances in Neural Information Processing Systems, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.804, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video Diffusion Models. Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.86, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Models. International Conference on Learning Representations, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[14] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-Shot Text-Guided Object Generation with Dream Fields. IEEE Computer Vision and Pattern Recognition Conference, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[15] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up GANs for Text-to-Image Synthesis. IEEE Computer Vision and Pattern Recognition Conference, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[16] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. International Conference on Learning Representations, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[17] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A Versatile Diffusion Model for Audio Synthesis. International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[18] Xiang Lisa Li, John Thickstun, Ishaan Gulrajani, Percy Liang, and Tatsunori Hashimoto. Diffusion-LM: Improves Controllable Text Generation. Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-Resolution Text-to-3D Content Creation. IEEE Computer Vision and Pattern Recognition Conference, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. European Conference on Computer Vision, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[21] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation. arXiv preprint arXiv:2309.06380, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.634 + ], + "angle": 0, + "content": "[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps. Advances in Neural Information Processing Systems, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[23] Eric Luhman and Troy Luhman. Knowledge Distillation in Iterative Generative Models for Improved Sampling Speed. arXiv preprint arXiv:2101.02388, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[24] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[25] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik P. Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On Distillation of Guided Diffusion Models. IEEE Computer Vision and Pattern Recognition Conference, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[26] Alexander Quinn Nichol and Prafulla Dhariwal. Improved Denoising Diffusion Probabilistic Models. International Conference on Machine Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[27] Junting Pan, Keqiang Sun, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. JourneyDB: A Benchmark for Generative Image Understanding." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7815" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "Advances in Neural Information Processing Systems, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[28] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion. International Conference on Learning Representations, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.468, + 0.261 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning Transferable Visual Models from Natural Language Supervision. International Conference on Machine Learning, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.468, + 0.332 + ], + "angle": 0, + "content": "[30] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.468, + 0.39 + ], + "angle": 0, + "content": "[31] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-Shot Text-to-Image Generation. International Conference on Machine Learning, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.468, + 0.446 + ], + "angle": 0, + "content": "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv preprint arXiv:2204.06125, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.468, + 0.504 + ], + "angle": 0, + "content": "[33] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative Adversarial Text to Image Synthesis. International Conference on Machine Learning, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.468, + 0.56 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. IEEE Computer Vision and Pattern Recognition Conference, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.468, + 0.631 + ], + "angle": 0, + "content": "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-image Diffusion Models for Subject-Driven Generation. IEEE Computer Vision and Pattern Recognition Conference, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.468, + 0.73 + ], + "angle": 0, + "content": "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[37] Tim Salimans and Jonathan Ho. Progressive Distillation for Fast Sampling of Diffusion Models. International Conference on Learning Representations, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.468, + 0.83 + ], + "angle": 0, + "content": "[38] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. StyleGAN-T: Unlocking the Power of GANs for Fast Large-Scale Text-to-Image Synthesis. International Conference on Machine Learning, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Jitsev. LAION-5B: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep Unsupervised Learning using Nonequilibrium Thermodynamics. International Conference on Machine Learning, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[42] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency Models. International Conference on Machine Learning, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.334 + ], + "angle": 0, + "content": "[43] StabilityAI. Stable Diffusion. https://github.com/Stability-AI/stablediffusion, 2021.2.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[44] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score Jacobian Chaining: Lifting Pretrained 2D Diffusion Models for 3D Generation. IEEE Computer Vision and Pattern Recognition Conference, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[45] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-Fidelity and Diverse Text-to-3D Generation with Variational Score Distillation. Advances in Neural Information Processing Systems, 2023. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.475, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[46] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel View Synthesis with Diffusion Models. International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.546, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[47] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis. arXiv preprint arXiv:2306.09341, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.616, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[48] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris Metaxas. StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks. International Conference on Computer Vision, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.687, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. International Conference on Computer Vision, 2023. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.728 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7816" + } + ] +] \ No newline at end of file diff --git a/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/13853f95-872e-49e0-9add-f414db12f2de_origin.pdf b/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/13853f95-872e-49e0-9add-f414db12f2de_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d5b5f7cf4219205da06f90d93b877d5e9a425321 --- /dev/null +++ b/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/13853f95-872e-49e0-9add-f414db12f2de_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a04877e736724dba29bf78c9521e080e68595edad14b06c99ab5e6f40558aff +size 4049998 diff --git a/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/full.md b/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..460d89c2815fcbae96f9f08518774c29d49cecf0 --- /dev/null +++ b/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/full.md @@ -0,0 +1,373 @@ +# SwiftBrush: One-Step Text-to-Image Diffusion Model with Variational Score Distillation + +Thuan Hoang Nguyen Anh Tran VinAI Research, Vietnam + +{v.thuannh5,v.anhtt152}@vinai.io + +![](images/abf8a724058c432da20e579aabfe4a3a5081d1395b2a1ea01e03a3f5f260f336.jpg) + +![](images/7ab8364f8db5c6fac9b15d022b4942be5534d5a9e5868e73ebbf389730da99e3.jpg) + +![](images/6f37fd2f3a3fb5ad229907ccb50823ca6a0c03ff4f9a35b15d1ec1d9efe15ab7.jpg) +Figure 1. Samples generated in a single step by SwiftBrush. Our distillation method can produce images of high-fidelity roughly 20 times faster compared to Stable Diffusion. + +![](images/3d95baf6c1b8a72331904f370a5c43577e5b4fc3ff9badb524a970b52e0892aa.jpg) + +# Abstract + +Despite their ability to generate high-resolution and diverse images from text prompts, text-to-image diffusion models often suffer from slow iterative sampling processes. Model distillation is one of the most effective directions to accelerate these models. However, previous distillation methods fail to retain the generation quality while requir + +ing a significant amount of images for training, either from real data or synthetically generated by the teacher model. In response to this limitation, we present a novel image-free distillation scheme named SwiftBrush. Drawing inspiration from text-to-3D synthesis, in which a 3D neural radiance field that aligns with the input prompt can be obtained from a 2D text-to-image diffusion prior via a specialized loss without the use of any 3D data ground-truth, our ap + +proach re-purposes that same loss for distilling a pretrained multi-step text-to-image model to a student network that can generate high-fidelity images with just a single inference step. In spite of its simplicity, our model stands as one of the first one-step text-to-image generators that can produce images of comparable quality to Stable Diffusion without reliance on any training image data. Remarkably, SwiftBrush achieves an FID score of 16.67 and a CLIP score of 0.29 on the COCO-30K benchmark, achieving competitive results or even substantially surpassing existing state-of-the-art distillation techniques. + +# 1. Introduction + +Diffusion models are gaining significant attention from the research community as they have achieved remarkable results in various generative tasks, including image generation [11, 26, 40, 41], video synthesis [12], 3D modeling [3, 28, 46], audio generation [17], and text creation [8, 18]. Especially, text-to-image diffusion models, which combine the power of language models and high-quality diffusion models [32, 36, 43], have revolutionized the way that people create visual content based on textual descriptions. Synthetic images indistinguishable from real photos are now possible with only a few clicks. Additionally, with techniques like ControlNet [49] or DreamBooth [35], extra control to the generation outputs can be made to further boost the expressiveness of the text-to-image models and bring them closer to mass users. However, despite their impressive success, these models exhibit a noteworthy drawback of slow inference speed. The primary reason for this lies in the iterative nature of their sampling, presenting significant challenges for deployment on consumer devices. + +Several approaches have been proposed to improve the inference speed of diffusion-based text-to-image generation. However, the most effective direction is time-step distillation, which aims to reduce the number of sampling steps of the diffusion model with the network backbone unchanged and minimum reduction in generation quality. Starting from the pioneering work [37], more advanced and effective distillation methods have been proposed, and a few recent ones could train efficient student networks that only require one forward step to synthesize images from input captions. In this work, we are interested in the distillation-based approach for one-step text-to-image generation. Particularly, Meng et al. [25] employ a two-stage distillation, where the student tries to first match its output with the classifier-free guided output of the teacher and then bootstrap itself to reduce the inference steps gradually. Such a method is complex and time-consuming, requiring at least 108 A100 GPU days [21]. Based on the work of Song et al. [42], Latent Consistency Model (LCM) [24] enforces the consistency of all points on the ODE trajectory in the + +latent space of Stable Diffusion (SD) [43] and unlike [25], they directly combine the two distillation stages into one. Although both Guided Distillation and LCM significantly reduce the number of inference steps to 2 or 4, their 1-step inference yields blurry and unsatisfying results. More sophisticated, Instaflow [21] proposes a technique called "reflow" to straighten the sampling trajectory of the teacher before distilling. While their proposed techniques proven to be superior to the original diffusion framework in one-step text-to-image distillation, InstaFlow requires an expensive 4-stage training schedule. Also, it is noteworthy that the effectiveness of these methods relies heavily on a large number of text-image pairs, which is not always the case due to the limited accessibility of such data. + +In this work, we aim to develop a novel distillation method for one-step text-to-image generation with higher quality and a more approachable training process characterized by an image-free and straightforward mechanism. Building upon the recent advancements in text-to-3D techniques, our work takes inspiration from their ability to generate high-quality 3D Neural Radiance Fields (NeRF) without utilizing 3D ground truth data. Such a remarkable achievement is facilitated by employing a powerful pretrained 2D text-to-image model to assess whether the image rendered from NeRF is realistic or not, akin to the discriminator in GANs. This observation underpins the motivation for our SwiftBrush approach, as we recognize that these principles can be innovatively adapted for text-to-image generation making image supervision unnecessary. SwiftBrush, therefore, emerges as a conceptually novel one-step text-to-image distillation process, uniquely capable of operating without 2D image supervision. This strategic adaptation not only opens up a change of view in the distillation of text-to-image diffusion models into a single-step generators but also points out the potential of blending principles from text-to-3D to text-to-image synthesis. + +As a result, our method has successfully enabled the generation of high-quality images with remarkable details. Quantitatively, our one-step model achieves promising zero-shot results, 16.6 for the FID score and 0.29 for the CLIP score on the MS COCO 2014 dataset (30K images). Notably, this is the first time a distilled one-step model outperforms the previous approaches without a single image used for training. A demonstration of generated images from our distilled model is shown in Fig. 1. + +# 2. Related Work + +Text-to-Image Generation. Previously, text-to-image models relied on GANs and focused on only small-scale object-centric domains such as flowers and birds [33, 48]. Recently, these types of systems have advanced rapidly with the availability of web-scale datasets of text-image pairs + +such as LAION5B [39], large language models such as T5 [30], and large vision-language models such as CLIP [29]. DALL-E [31] was the first autoregressive model to demonstrate a remarkable zero-shot ability in creating images from text captions by simply scaling up both the network and dataset size. Following that, a plethora of methods have been used for text-to-image generation, including masked generative transformers [4], GAN-based, and diffusion models. Based on Latent Diffusion Model [34], Stable Diffusion is an open-source diffusion-based generator that has gained widespread popularity among artists as well as researchers. Recently, text-to-image GANs are receiving attention again with notable work like StyleGAN-T [38], or GigaGAN [15]. While they are extremely fast as no iterative sampling is required, they lack behind diffusion models in terms of scalability and realism due to the fact that they suffer from instability and the infamous issue "mode collapse". To the best of our knowledge, so far only two above-mentioned work succeeded in applying GANs for text-to-image using tons of intricate techniques and auxiliary modules. Conversely, diffusion models can be easily scale up to generate high-quality and sharp samples, yet they require lengthy sampling process. Here, our objective is to create a large-scale generative model that combines the speed of GANs with the quality of diffusion models. + +Text-to-3D Generation. Instead of learning a generative model directly from 3D supervision, many research works have utilized the rich knowledge of 2D prior for 3D generation tasks. For instance, one of the earliest works, Dream Fields [14], leverages CLIP [29] to guide the generated images so that rendered images from many camera views highly align with the text caption. On the other hand, pioneer works like DreamFusion [28] and Score Jacobian Chaining [44] (SJC) propose two different but equivalent frameworks where a 2D text-to-image generative prior is used to generate 3D objects from textual descriptions. Subsequent works seek to improve these text-to-3D methods in various ways. Among them, Magic3D [19] and ProlificDreamer [45] are two remarkable works that effectively enhance the generation quality of 3D assets. Magic3D consists of two stages where they first obtain a low-resolution 3D representation via a DreamFusion-style optimization and then refine its texture to get the final high-fidelity 3D mesh. Meanwhile, ProlificDreamer proposes Variational Score Distillation (VSD) involving a secondary teacher to bridge the gap between the teacher and the 3D NeRF. Influenced by the latter work, our method adapts the same technique for one-step text-to-image generation. Similarly, Score GAN [6] and Diff-Instruct [7] use a VSD-like loss to distill a pretrained diffusion model to a one-step generator. + +Knowledge Distillation introduced by Hinton et al. [10] is one of the methods falling under the umbrella of transfer + +learning. This algorithm draws its main inspiration from the human learning process, where knowledge is conveyed from a more knowledgeable teacher to a less knowledgeable student. In the context of diffusion models, works have been proposed for faster sampling of student models from a pre-trained diffusion teacher. The most straightforward approach involves direct distillation [23], wherein a student model is supervised from the teacher's output after sampling for 50-1000 steps, which can be prohibitive. Recent work [37, 42] avoids the lengthy sampling during distillation by various bootstrapping methods. Afterward, a series of methods [21, 24, 25] adapt these methods into text-to-image settings and inherit their limitation of being image-dependent in training. In contrast, a concurrent work [9] learns a time-conditioned model to predict the outcome of the teacher at any given time step, all without the need for image supervision. Likewise, our approach is similarly image-free yet achieves significantly better results and offers a much simpler distillation design. + +# 3. Proposed Method + +# 3.1. Preliminary + +Diffusion Models consist of two processes: a forward process that gradually adds noise and another reverse process to predict the distribution of gradually denoised data points. The forward process incrementally introduces noise to an initial data point $x_0$ , drawn from the distribution $q_0(x_0)$ , resulting in $q_t(x_t|x_0) = \mathcal{N}(\alpha_t x_0, \sigma_t^2 I)$ , where $\{(\alpha_t, \sigma_t)\}_{t=1}^T$ is the noise schedule. The data point $x_t$ at each time-step $t$ is drawn from $q_t(x_t|x_0)$ using a noise $\epsilon \sim \mathcal{N}(0, I)$ : + +$$ +x _ {t} = \alpha_ {t} x _ {0} + \sigma_ {t} \epsilon . \tag {1} +$$ + +After $T$ time steps, the final data point is expected to be pure Gaussian noise $x_{T}\sim \mathcal{N}(0,I)$ + +Conversely, the backward process starts from $x_{T} \sim \mathcal{N}(0, I)$ and iteratively denoises for $T$ steps where at each step $t$ , noisy variable $x_{t+1}$ is converted to less-noisy variable $x_{t}$ , guided by a model $\epsilon_{\psi}$ that predicts the noise $\epsilon$ in Eq. (1). The network weights $\psi$ are trained by minimizing: + +$$ +\mathcal {L} _ {u n c o n d} (\psi) = \mathbb {E} _ {t, \epsilon \in \mathcal {N} (0, 1)} \| \epsilon_ {\psi} (x _ {t}, t) - \epsilon \| _ {2} ^ {2} \tag {2} +$$ + +where $t$ is uniformly samples within $\{1,\dots ,T\}$ + +While common diffusion models operate in the pixel space, Latent Diffusion Models (LDMs)[34] model the diffusion process in the latent space of a pre-trained regularized autoencoder, which features representations of smaller size, thus enhancing both training and inference efficiency. + +Text-to-Image Diffusion Models. In contrast to unconditional diffusion models that generate outputs freely, text-conditioned diffusion models steer the sampling process using an extra prompt $y$ . This guides the model to produce + +outputs that are not only photorealistic but also closely adhere to the provided text description. The objective to train such models is formulated as follows: + +$$ +\mathcal {L} _ {d i f f} (\psi) = \mathbb {E} _ {t, y, \epsilon \in \mathcal {N} (0, 1)} \| \epsilon_ {\psi} (x _ {t}, t, y) - \epsilon \| _ {2} ^ {2} \tag {3} +$$ + +which is slightly different from the unconditional diffusion loss Eq. (2). Due to the prompt conditioning, the model can deliver more controllable generation compared to its unconditional counterparts. Nevertheless, the implementations of many leading methods [1, 32, 36] remain inaccessible to the public. Stable Diffusion, which primarily utilizes the LDM framework, stands out as the first openly available large-scale model, significantly propelling text-to-image synthesis's widespread adoption and versatility. + +Score Distillation Sampling (SDS) is a distillation technique effectively applied in generating 3D assets [19, 28, 44]. It utilizes a pre-trained text-to-image diffusion model, which predicts diffusion noise from text condition $y$ , denoted as $\epsilon_{\psi}(x_t,t,y)$ . The method optimizes a single 3D NeRF, parameterized by $\theta$ , to align with a given text prompt. Given camera parameters $c$ , a differentiable rendering function $g(\cdot ,c)$ is used to render an image at the camera view $c$ from the 3D NeRF. Here, the rendered image $g(\theta ,c)$ is utilized to optimize the weight $\theta$ through a loss function whose gradient can be approximated by: + +$$ +\nabla_ {\theta} \mathcal {L} _ {S D S} = \mathbb {E} _ {t, \epsilon , c} \left[ w (t) (\epsilon_ {\psi} (x _ {t}, t, y) - \epsilon) \frac {\partial g (\theta , c)}{\partial \theta} \right] (4) +$$ + +where $t\sim \mathcal{U}(0.02T,0.98T)$ $T$ is the maximum timesteps of the diffusion model, $\epsilon \sim \mathcal{N}(0,I)$ $x_{t} = \alpha_{t}g(\theta ,c) + \sigma_{t}\epsilon$ $y$ is the input text, and $w(t)$ is a weighting function + +Despite the advancements in text-to-3D synthesis, empirical studies [28, 44] indicate that SDS often encounters issues like over-saturation, over-smoothing, and reduced diversity. The same degradations can be observed if we naively apply SDS in our framework, as shown in Sec. 4.3. + +Variational Score Distillation (VSD) is introduced in ProlificDreamer [45] to address previously mentioned issues of SDS by modifying the loss slightly: + +$$ +\begin{array}{l} \nabla_ {\theta} \mathcal {L} _ {V S D} = \mathbb {E} _ {t, \epsilon , c} [ w (t) (\epsilon_ {\psi} (x _ {t}, t, y) \\ \left. - \epsilon_ {\phi} (x _ {t}, t, y, c)) \frac {\partial g (\theta , c)}{\partial \theta} \right] \tag {5} \\ \end{array} +$$ + +Again, $x_{t} = \alpha_{t}g(\theta ,c) + \sigma_{t}\epsilon$ is the noisy observation of the rendered image at camera view $c$ . VSD sets itself apart from SDS by introducing an additional score function tailored to the images rendered from the 3D NeRF at the camera pose $c$ . This score is derived by fine-tuning a diffusion model $\epsilon_{\phi}(x_t,t,y,c)$ with the diffusion loss described below: + +$$ +\min _ {\epsilon_ {\phi}} \mathbb {E} _ {t, c, \epsilon} \| \epsilon_ {\phi} (x _ {t}, t, y, c) - \epsilon \| _ {2} ^ {2} \tag {6} +$$ + +As proposed in ProlificDreamer, $\epsilon_{\phi}$ is parameterized by Low-Rank Adaption [13] (LoRA) and initialized from the pre-trained diffusion model $\epsilon_{\psi}$ with some added layers for conditioning the camera view $c$ . Note that at each iteration $i$ of the optimization process, $\epsilon_{\phi}$ needs to adapt with the current distribution of $\theta$ . Therefore, ProlificDreamer interleaves between finetuning $\epsilon_{\phi}$ and optimizing $\theta$ . With these algorithmic enhancements, ProlificDreamer has significantly advanced its capabilities, enabling the generation of NeRFs and the creation of exceptional textured meshes. Such improvement directly inspires us to adapt VSD to the task of one-step text-to-image diffusion-based distillation. + +# 3.2. SwiftBrush + +Motivation. While SDS and VSD are designed explicitly for the text-to-3D generation task, they loosely connect to that goal via the rendered image $g(\theta, c)$ of the 3D NeRF. As a matter of fact, instead of using NeRF rendering, we can replace it with any function that outputs 2D images to suit our needs. Inspired by this motivation, we propose to substitute the NeRF rendering $g(\theta, c)$ with a text-to-image generator that can directly synthesize a text-guided image in one step, effectively converting the text-to-3D generation training into one-step diffusion model distillation. + +Design Space. We adopt the same approach as [45], with modifications to the design space to suit our task better. To begin with, we make use of two teacher models: a pretrained text-to-image teacher $\epsilon_{\psi}$ and one additional LoRA teacher $\epsilon_{\phi}$ . Also, we remove the conditioning on camera view $c$ from the LoRA teacher as it is unnecessary in our case, and we use classifier-free guidance for both teachers. Then, we replace the NeRF, which is overfitted for a specific user-provided prompt in a text-to-3D setting, with a generalized one-step text-to-image student model $f_{\theta}(z,y)$ . Our student model $f_{\theta}$ takes a random Gaussian noise $z$ and a text prompt $y$ as inputs. Both the LoRA teacher and student model are initialized with the weight of the text-to-image teacher. Next, we train both the student model and the LoRA teacher alternately using Eq. (5) and Eq. (6) while freezing the text-to-image teacher. A pseudo-code and system figure can be viewed in Alg. 1 and Fig. 2. + +Student Parameterization. Given a pretrained text-to-image diffusion model $\epsilon_{\theta}$ , it is possible to directly use its output for the student model, i.e., $f_{\theta}(z,y) = \epsilon_{\theta}(z,T,y)$ , where $T$ is the maximum timestep of the pretrained model. However, in our case, the pretrained model of choice is Stable Diffusion, which is inherently designed to predict the added noise $\epsilon$ . In contrast, our objective is to refine the student model so that it predicts a clean and noise-free $x_0$ . Therefore, such a naive method results in a large domain gap between what we want the student to learn and the output of the student. For the sake of easy training, we empiri + +# Algorithm 1 SwiftBrush Distillation + +1: Require: a pretrained text-to-image teacher $\epsilon_{\psi}$ , a LoRA teacher $\epsilon_{\phi}$ , a student model $f_{\theta}$ , two learning rates $\eta_{1}$ and $\eta_{2}$ , a weighting function $\omega$ , a prompts dataset $Y$ , the maximum number of time steps $T$ and the noise schedule $\{(\alpha_t, \sigma_t)\}_{t=1}^T$ of the teacher model +2: Initialize: $\phi \gets \psi, \theta \gets \psi$ +3: while not converged do +4: Sample input noise $z\sim \mathcal{N}(0,I)$ +5: Sample text caption input $y\sim Y$ +6: Compute student output $\hat{x}_0 = f_\theta (z,y)$ +7: Sample timestep $t \sim \mathcal{U}(0.02T, 0.98T)$ +8: Sample added noise $\epsilon \sim \mathcal{N}(0,I)$ +9: Compute noisy sample $\hat{x}_t = \alpha_t\hat{x}_0 + \sigma_t\epsilon$ +10: $\theta \gets \theta -\eta_{1}\left[\omega (t)\left(\epsilon_{\psi}(\hat{x}_{t},t,y) - \epsilon_{\phi}(\hat{x}_{t},t,y)\right)\frac{\partial\hat{x}_{0}}{\partial\theta}\right]$ +11: Sample timestep $t^{\prime}\sim \mathcal{U}(0,T)$ +12: Sample added noise $\epsilon^{\prime}\sim \mathcal{N}(0,I)$ +13: Compute noisy sample $\hat{x}_{t'} = \alpha_{t'}\hat{x}_0 + \sigma_{t'}\epsilon'$ +14: $\phi \gets \phi -\eta_{2}\nabla_{\phi}\| \epsilon_{\phi}(\hat{x}_{t^{\prime}},t^{\prime},y) - \epsilon^{\prime}\|^{2}$ +15: end while +16: return trained student model $f_{\theta}$ + +cally re-parameterize the student output as follows: + +$$ +f _ {\theta} (z, y) = \frac {z - \sigma_ {T} \epsilon_ {\theta} (z , T , y)}{\alpha_ {T}} +$$ + +which is a realization of Eq. (1) if we set $t = T$ , $x_{t} = z$ , $\epsilon \approx \epsilon_{\theta}(z,T,y)$ and $x_0 \approx f_\theta (z,y)$ . So basically, this reparameterization transforms the noise-prediction output of the pretrained model into the "predicted $x_0$ " form, which is empirically demonstrated in Sec. 4.3 to be helpful for the student model to learn. + +# 4. Experiments + +# 4.1. Experimental Setup + +Evaluation Metrics. We conduct our evaluation process by assessing the performance of SwiftBrush against several well-established methods, such as Guided Distillation [25], Instaflow [21], LCM [24], and BOOT [9], across two zero-shot text-to-image benchmarks: COCO 2014 [20] and Human Preference Score v2 [47] (HPSv2). On the COCO 2014, we follow the conventional evaluation protocol [15, 34, 36, 38], computing FID and CLIP score on a 30K subset of its validation set as our primary metrics. Here, we use ViT-G/14 as the backbone for evaluating CLIP scores. In the case of HPSv2, we adopt their evaluation procedure to assess text-image alignment across four different aspects: anime, concept-art, paintings, and photos. + +Experiment Protocols. In these experiments, we rely on Stable Diffusion v2.1 [43] for our two teacher models, and + +whenever possible, we directly utilize the reported values from other research. As for InstaFlow and LCM, we use the their provided pretrained models to generate images for evaluation. Conversely, we are compelled to re-implement BOOT since they did not provide any metrics for their text-to-image models, and their codebase remains entirely closed. + +Implementation Details. Similar to prior research, we use student models with the same architectures as those of the teachers and initialize all of the student parameters with the teacher model. During training, a guidance scale of 4.5 is used for both teachers. For the LoRA teacher model, we apply a learning rate of 1e-3, a LoRA rank of 64, and an alpha value of 108. Meanwhile, the student model's learning rate is set at 1e-6 with exponential moving average (EMA) used every iteration. Adam optimizer [16] and a batch size of 64 (16 per GPU) are used to train both the student and LoRA teacher for around 65K iterations. Thanks to the image-free nature of SwiftBrush, we do not require access to any image training set, and only the prompts are needed to distill the generator. Here, we utilize 4M captions from the large-scale text-image dataset JourneyDB [27], which we subsequently deduplicated to 2M. Regarding BOOT, we closely follow their implementation guidelines to the best of our abilities and train the student model until its convergence, except that we use Stable Diffusion v2.1 instead of DeepFloyd IF [5] as the teacher. Here, its blurry results (Fig. 4) compared to the original work could be due to the different teacher. + +# 4.2. Results + +Quantitative Results. We first evaluate the proposed method on standard text-to-image generation benchmarks. The quantitative comparison with the previous distillation, both image-dependent and image-free, and the original Stable Diffusion teacher is shown in Tab. 1. Despite being trained with only text captions, SwiftBrush significantly improves the quality of 1-step inference and achieves better photo-realism as well as text-image alignment against other methods, with exceptions of InstaFlow and Stable Diffusion (25 sampling steps) in FID score. However, it is worth noting that Instaflow is so complex, requiring a 4-stage distillation process, while ours is end-to-end and straightforward. We further conduct a quantitative evaluation with the HPSv2 score, a metric trained to capture human preference. Again, we obtain the best score in all aspects among all mentioned works, which is only a slight degradation compared to the original Stable Diffusion teacher (Tab. 2). Moreover, according to the anonymous survey carried out by us (Fig. 3), our method outperforms Instaflow in terms of image fidelity and text-to-image accuracy. + +Qualitative Results. We report the visual comparison between our method and the mentioned baseline methods + +![](images/7a3b1c07f625c04d4463492843c84e7923ad2df0f8e49b8c07e897915371bc37.jpg) +Figure 2. SwiftBrush overview. Our system allows training a one-step text-to-image student network from a frozen pretrained teacher with the help of an additional trainable LORA teacher. The student network takes inputs a text prompt and random noise. The output of the student is then added noise and sent to two teachers, alongside the prompt and a randomly drawn timestep, to compute the gradient of VSD loss, which backpropagates back to the student. Apart from that, the LoRA teacher is also updated using diffusion loss. Similar to ProlificDreamer [45], we update the LoRA teacher (green flow) and the student (red flow) alternatively. + +![](images/812107473a718a74917a5075987c817a44a80465a3c8221981a22789f003d793.jpg) +Figure 3. User Study. We ask 400 users to compare the generated images of InstaFlow and our method using 30 random text prompts. We report the rate at which each method is selected as better for image quality and text-to-image accuracy. + +for the one-step text-to-image generation task. As seen in Fig. 4, LCM outputs blurry and over-smoothed images in the 1-step regime. BOOT faithfully produces images that match well with the text caption, but its results still fall short in terms of fidelity. In contrast, a much better quality can be observed in the image produced by Instaflow and SwiftBrush. InstaFlow, nevertheless, cannot handle complex or unusual input prompts properly, which can be noticed in the third and fifth images at the third column of Fig. 4. On top of that, it occasionally produces non-realistic results even for simple prompts like in Fig. 4 at the fourth image of the third column. When, in fact, the outputs generated by SwiftBrush are not only of exceptional quality but also exhibit a remarkable alignment between text descriptions and the corresponding images. Moreover, our results are comparable with those sampled from Stable Diffusion with 25 steps; whereas it is understandable that with only one step, Stable Diffusion cannot create outputs of good quality. + +# 4.3. Analysis + +The importance of LoRA teacher. In Fig. 6, the importance of including the LoRA teacher in the training process is clearly highlighted. We compare the student outputs sampled with different input noises after 10K iteration when us + +
MethodStepsFID-30K ↓CLIP-30K ↑
Guided Distillation†137.30.27
LCM†135.560.24
Instaflow113.10†0.28§
BOOT‡148.200.26
Ours116.670.29
SD 2.1*2513.450.30
SD 2.1*1202.140.08
+ +Table 1. Comparison of our method against other works based on FID metric and CLIP score on the COCO 2014 dataset. $^\dagger$ means that we obtain the numbers from the corresponding papers. $\S$ means that we obtain the numbers using the provided pretrained models of the corresponding papers. $^\ddagger$ means that we re-implement the work and report the numbers. * means that we report the number using the pretrained models. We use DPM-Solver [22] with guidance scale 7.5 to sample for SD. Bold and underlined numbers are best and second best for one-step models. + +
ModelsHuman Preference Score v2 ↑
AnimePhotoConcept ArtPaintings
LCM†22.6122.7122.7422.91
InstaFlow†25.9826.3225.7925.93
BOOT‡25.2925.1624.4024.61
Ours26.9127.2126.3226.37
SD 2.1*27.4826.8926.8627.46
+ +Table 2. Comparison of our method against other works based on HPSv2 score in 1-step regime. $\dagger$ means that we obtain the score using the provided pretrained models of the corresponding papers. $\ddagger$ means that we re-implement the work and report the score by ourselves. * means that we obtain the score from [47]. Bold and underlined numbers are best and second best, respectively. + +ing the same text prompts. With the guidance of the LoRA teacher, the student progressively learns to create realistic images ( $1^{st}$ column). In contrast, in the absence of an ef + +![](images/5afe2db2e4630a5b55287ad4601b7e8c615c01c2cde9c70e5d47e925dbb659a8.jpg) +"A photo of one ice cream ball in a luxurious plate, bokeh" +LCM + +![](images/91101967545bb2a4f16931bb2f2845e6a46114454cc147105779e5bae158bbc8.jpg) +BOOT + +![](images/4a1c23d3cae446d089bd37a27efa62ef901c65cc4a7f65d8cf4537df000a7b91.jpg) +InstaFlow + +![](images/953b8d02bc1827b355999659e5b319eef44e1c04682236372358b482a5e0f3ee.jpg) +Ours + +![](images/3d42bd44d1942e79d7fa32065bfe84469810d77bc08bf624df4d2f4557bc0981.jpg) +SD* + +![](images/3b30057df9197efae4d798cd083105401b0d39043f567c50dfaa4d2fef30979d.jpg) +SD† + +![](images/198cf500f3369f314c6f9b73aa59f9eb38bd5ca99dfbbdd48f7d2e304377e5b7.jpg) +"A hyperrealistic photo of fox astronaut, perfect face, artstation" + +![](images/5a05e9c607af55c062d85492723da54c305015052f89a6bf3ae261e01a9ebb3e.jpg) + +![](images/2802d2ad5aa3e0613f924267ed3b1df9599a4d16aa0be8173d4fb14cb7f9467e.jpg) + +![](images/d88a1d3d373a34ed133ddf1ab1d5e01f9e623ddc8d1f09871d904e20dd39def7.jpg) + +![](images/ee112dac7cd86a08f1cb8f84b4bd08294a42b04bba0984daf3394ec9cea4aaa7.jpg) + +![](images/cffa696eb135c8a6bd8d9db1eece800e019afdd9d895ae23e1d1e4fecc1de0f2.jpg) + +![](images/88aed34cc8fce6a13270bb3647dfedd284471c814722c2d73aec0fcd80b6cd3f.jpg) +"A poodle wearing a baseball cap and holding a dictionary in hand" + +![](images/c61dc697f40e133c0948744834072aee2afea166f1f4d9da81fe21b2043b054a.jpg) + +![](images/6c2be08869aafead182b3a99722a5ef6bd8e67f5ebe1eb59ea17c56305dec90b.jpg) + +![](images/1c20f5069bcc119d1bfcd1e24eac80ab6fe68622566b6cef6d01f718c9b6a10e.jpg) + +![](images/9494333f178d2054b331dfa7029502b30b69c4af76bf11d1d06c4caee5e4b0e7.jpg) + +![](images/d5e8727c06927e472018ad6682530a3c88ca24151b1bcc443f1632709c988736.jpg) + +![](images/767c986b23b50af3ec467e281e6e182be2e8721010c92fe0fae17f817a5a5573.jpg) +"A DSLR photo of a cat jumping over a fence, high-res" + +![](images/7f3de0616d99be0965264fb0e9cea175eb4943ba181db7cae8d9abfbed8ad8ef.jpg) + +![](images/c756f0536b9274f20d230c3f0f7fe509f8f73ed556e416083f64b2ed3ac1fc32.jpg) + +![](images/6b8d3915dd4d9f1db9226db95f51a357b3116d9e48b5a162e55af7da33b72ba9.jpg) + +![](images/718d2356631f7e89b9989d6b342bfb19e0948cacbf0ca8dd52db759d94a101d2.jpg) + +![](images/9e430d7675b11c63f1599edbe97b9d0fe9aa5132118181f753077f2b085def92.jpg) + +![](images/72200b3d12a6c4589659236250a96de56ff83b8891cdc74da9a74c2c17546696.jpg) +"A blue Porsche 356 parked in front of a brick wall" + +![](images/9558855b7e62142bff08d080aea99866561af9c5b2b5ea0e00b7bbb1235c8cb7.jpg) + +![](images/53a9506e177150ca358fe4b03b7d7cb1f4866e98c6eb9a1fb3a0d93a47cbdde2.jpg) + +![](images/ecfab9d3c6bad519d4d60292f46e964fce8b74b78a416c021f3a62ac1effc549.jpg) + +![](images/34439a0b6d48491b9a1e85c35db6f26833ff6dafec1e78a021612ee91ef0ee1e.jpg) + +![](images/b4d7b8b131cb3ff407055a3d828c64017c9078e698eb248914072f26e0aa17e2.jpg) + +![](images/51bc833455ced53b6bae2fb5d9b801db20508affd9b6e752b004ef46b25c8862.jpg) +"A photograph of an astronaut riding a horse" +Figure 4. Sample images generated by LCM, BOOT, Instaflow, Ours, SD with 25 sampling steps $(\mathrm{SD}^{*})$ and SD with 1 sampling step $(\mathrm{SD}^{\dagger})$ . Images at the same row are sampled from the same text prompt, while images at the same column are from the same model. We use DPM-Solver [22] with guidance scale 7.5 to sample for Stable Diffusion + +![](images/3fc4ba236e59f17378472457c5054115c817e078fd9ec3a2b9014b3e46d2365a.jpg) + +![](images/47edaade28b7b17e8d421056103dedf98db42e9834571288a1aed386337e8fe8.jpg) + +![](images/9fe2b24dfa901decdc8f8c8780d92f25b2722760e3f782db62e48432d56c1774.jpg) + +![](images/2cce683dbc060cbeca3d012aa87879dde8ec52aad7db6fb69ced25cef37bbb15.jpg) + +![](images/2944e461664cfdd2b50223ffdc22521eba30a97ae369d4114c25efe8f99cf33c.jpg) + +fective LoRA teacher, the student's early training outputs, despite being highly aligned with the text prompt, are overly saturated and unrealistic ( $2^{nd}$ column). These results indicate a fundamental breakdown in the student model's ability to mimic the teacher model's distribution, as can be further demonstrated based on the flat red curve in both FID and CLIP score in Fig. 5 (a) and (b). + +In addition, even when employing the LoRA teacher, it is crucial to use a sufficiently large rank in the LoRA weight for stable distillation of the student model. Without this, the model tends to undergo severe mode collapse despite initially producing reasonably good results. The green curves in Fig. 5 clearly illustrate that, while the FID score of a student model with a 4-rank LoRA weight initially decreases, it deteriorates starting from 8K steps, resulting in lousy images in Fig. 6 ( $3^{rd}$ column). Additionally, the correspond + +ing CLIP score behaves similarly where there is a slight rise before it worsens in the end. Besides, we also provide extra visuals for this ablative study in Supplementary. + +The importance of student parameterization. Apart from the LoRA teacher, the re-parameterization of the student (Sec. 3) also plays an important role in the SwiftBrush training. Without it, the student is unable to create realistic images and it only produces noisy results, as can be noticed from the $4^{th}$ column of Fig. 6. This can be further backed up by the orange curve in both FID and CLIP score in Fig. 5 (a) and (b), where student without re-parameterization performs consistently worst among all configurations. + +Interpolating Latent. We conduct a controlled experiment where the noise input $z$ remains the same while varying the input prompt $y$ (Fig. 7). Interestingly, the SwiftBrush- + +![](images/4e54008329a3726d42c9a4db0721ed675e41c16246872bb9b899ca1c7ed200af.jpg) +(a) FID curves of four configurations + +![](images/a96d66bfdb07dec0ab90d570494aa5022ce35a4bc6cace3c35b17cfad03e28b6.jpg) +(b) CLIP curves of four configurations + +![](images/56374b2b8e00c8ae4ea1f812a4d152dde9a6d323c8ede91b9f1df4384c491b87.jpg) +Figure 5. Ablation study on the impact of the LoRA teacher as well as the student parameterization. We train all four configurations with same hyperparameters described in Sec. 4 but only with 10K iterations to observe their training dynamics at the early phase. Then, we evaluate FID and CLIP score on COCO 2014 but with only a 5K subset of images. Here, Full means all of our proposed techniques are used. No Parameterization means we exclude the re-parameterization trick. Small-Rank LoRA means we use only 4-rank LoRA teacher. SDS means we remove the LoRA teacher and use only SDS loss. + +![](images/1278c181d97f8233cc8a83baf20c28b234c0760582307bbfbd2a6290a296d298.jpg) + +![](images/2676c2a63ed719b00ee0eb3086fd3ef3ee531072e70fc188d81486a705c3f73a.jpg) + +![](images/c72a8c976337dcccbcb2e2bfc92d78297b8a2ad8238dd10a3159f5719f2a07dc.jpg) + +![](images/65c964b6513a2a9cf10c08970cef03cfbaa8e50cb16615299a775255e533f260.jpg) +Full + +![](images/b33d8b218629f196ece9c81e315ec48c3c28a93e540bcb0e6746ff3d2a515109.jpg) +SDS + +![](images/1903b1873a776c9f4c50dd198a5a1976d0613f21da996bf24b71ee80195a624b.jpg) +Small-Rank LoRA +Figure 6. Visual results of the ablative study, where the first two rows are generated with the input prompt "An oil painting of a train", whereas the remaining two are sampled from "An DSLR photo of a tiger in the city". + +![](images/95281ca8eb89ba6bc26561b51b94cf8cfbbdb31cf356cd8c8e163eea047a2551.jpg) +No Parameterization + +distilled student seemingly features enhanced control and editability compared to the teacher model due to its one-step nature. Additionally, we present further results of latent space interpolation in Supplementary. In this process, the student model, distilled from the Stable Diffusion teacher, demonstrates a seamless transition in its image generation. + +# 5. Conclusion and Discussion + +Conclusion. In this work, we have introduced a new image-free distillation scheme, SwiftBrush, which leverages insights from text-to-3D synthesis to speed up text-to-image generation. Empirically, the results presented in this paper demonstrate the effectiveness of our approach in making text-to-image generation more efficient and accessible. We hope our work will encourage broader use of text-to-image generation systems by greatly lowering the inference cost. + +Discussion. SwiftBrush inevitably produces lower quality samples compared to the teacher which utilizes multi- + +![](images/b5a1e2da3fb42bb5a9fe0f7bf16b49d51850457dbe776a8108349ef52a90bdf5.jpg) +Figure 7. Results of interpolating the input prompts. The prompts used here are selected from a standard template: 'A DSLR photo of a {animal} {action}' Here, 'animal' is one of following: owl, raccoon, cat, or panda, while 'action' is one of following: reading a book, drinking a latte or eating a pizza. + +![](images/df50ac732abc1f17f8bb88fe3977c5f8de8d64c5468da9cdab41da267283253a.jpg) + +step sampling for inference. Another drawback is that the current design only focuses on single-step student model and cannot support few-step generation as did in Luo et al. [24] to improve image fidelity. In future work, we aim to investigate the feasibility of extending SwiftBrush to a few-step generation so that we can trade computation for quality. Furthermore, we also find it compelling to explore SwiftBrush-like training where only one teacher is required. Such an exploration could provide tremendous benefits in scenarios where the computational resource for training is limited. Finally, it is intriguing to see if techniques such as DreamBooth [35], ControlNet [49] or InstructPix2Pix[2] can be integrated with SwiftBrush, allowing instant generation of various applications. + +# References + +[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. eDiff-I: Text-to-Image Diffusion Models with an Ensemble of Expert Denoisers. arXiv preprint arXiv:2211.01324, 2022. 4 +[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. InstructPix2Pix: Learning to Follow Image Editing Instructions. IEEE Computer Vision and Pattern Recognition Conference, 2023. 8 +[3] Eric R. Chan, Koki Nagano, Matthew A. Chan, Alexander W. Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. GeNVS: Generative Novel View Synthesis with 3D-Aware Diffusion Models. International Conference on Computer Vision, 2023. 2 +[4] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T. Freeman, Michael Rubinstein, Yuzhen Li, and Dilip Krishnan. Muse: Text-To-Image Generation via Masked Generative Transformers. arXiv preprint arXiv:2301.00704, 2023. 3 +[5] DeepFloyd. DeepFloyd IF. https://www deepfloyd.ai/deepfloyd-if, 2023.5 +[6] Jean-Yves Franceschi, Mike Gartrell, Ludovic Dos Santos, Thibaut Issenhuth, Emmanuel de Bezenac, Mickael Chen, and Alain Rakotomamonjy. Unifying GANs and Score-Based Diffusion as Generative Particle Models. Advances in Neural Information Processing Systems, 2023. 3 +[7] Jean-Yves Franceschi, Mike Gartrell, Ludovic Dos Santos, Thibaut Issenhuth, Emmanuel de Bezenac, Mickael Chen, and Alain Rakotomamonjy. Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models. Advances in Neural Information Processing Systems, 2023. 3 +[8] Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and Lingpeng Kong. DiffuSeq: Sequence to Sequence Text Generation with Diffusion Models. International Conference on Learning Representations, 2023. 2 +[9] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Lingjie Liu, and Josh Susskind. BOOT: Data-free Distillation of Denoising Diffusion Models with Bootstrapping. arXiv preprint arXiv:2306.05544, 2023. 3, 5 +[10] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the Knowledge in a Neural Network. NIPS Deep Learning Workshop, 2014. 3 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. Advances in Neural Information Processing Systems, 2020. 2 +[12] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video Diffusion Models. Advances in Neural Information Processing Systems, 2022. 2 +[13] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language + +Models. International Conference on Learning Representations, 2022. 4 +[14] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-Shot Text-Guided Object Generation with Dream Fields. IEEE Computer Vision and Pattern Recognition Conference, 2022. 3 +[15] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up GANs for Text-to-Image Synthesis. IEEE Computer Vision and Pattern Recognition Conference, 2023. 3, 5 +[16] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. International Conference on Learning Representations, 2015. 5 +[17] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A Versatile Diffusion Model for Audio Synthesis. International Conference on Learning Representations, 2021. 2 +[18] Xiang Lisa Li, John Thickstun, Ishaan Gulrajani, Percy Liang, and Tatsunori Hashimoto. Diffusion-LM: Improves Controllable Text Generation. Advances in Neural Information Processing Systems, 2022. 2 +[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-Resolution Text-to-3D Content Creation. IEEE Computer Vision and Pattern Recognition Conference, 2023. 3, 4 +[20] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. European Conference on Computer Vision, 2014. 5 +[21] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation. arXiv preprint arXiv:2309.06380, 2023. 2, 3, 5 +[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps. Advances in Neural Information Processing Systems, 2022. 6, 7 +[23] Eric Luhman and Troy Luhman. Knowledge Distillation in Iterative Generative Models for Improved Sampling Speed. arXiv preprint arXiv:2101.02388, 2021. 3 +[24] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2, 3, 5, 8 +[25] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik P. Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On Distillation of Guided Diffusion Models. IEEE Computer Vision and Pattern Recognition Conference, 2023. 2, 3, 5 +[26] Alexander Quinn Nichol and Prafulla Dhariwal. Improved Denoising Diffusion Probabilistic Models. International Conference on Machine Learning, 2021. 2 +[27] Junting Pan, Keqiang Sun, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. JourneyDB: A Benchmark for Generative Image Understanding. + +Advances in Neural Information Processing Systems, 2023. 5 +[28] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion. International Conference on Learning Representations, 2023. 2, 3, 4 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning Transferable Visual Models from Natural Language Supervision. International Conference on Machine Learning, 2021. 3 +[30] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 3 +[31] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-Shot Text-to-Image Generation. International Conference on Machine Learning, 2021. 3 +[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv preprint arXiv:2204.06125, 2022. 2, 4 +[33] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative Adversarial Text to Image Synthesis. International Conference on Machine Learning, 2016. 2 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. IEEE Computer Vision and Pattern Recognition Conference, 2022. 3, 5 +[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-image Diffusion Models for Subject-Driven Generation. IEEE Computer Vision and Pattern Recognition Conference, 2023. 2, 8 +[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 2022. 2, 4, 5 +[37] Tim Salimans and Jonathan Ho. Progressive Distillation for Fast Sampling of Diffusion Models. International Conference on Learning Representations, 2022. 2, 3 +[38] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. StyleGAN-T: Unlocking the Power of GANs for Fast Large-Scale Text-to-Image Synthesis. International Conference on Machine Learning, 2023. 3, 5 +[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia + +Jitsev. LAION-5B: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 2022. 3 +[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep Unsupervised Learning using Nonequilibrium Thermodynamics. International Conference on Machine Learning, 2015. 2 +[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. International Conference on Learning Representations, 2021. 2 +[42] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency Models. International Conference on Machine Learning, 2023. 2, 3 +[43] StabilityAI. Stable Diffusion. https://github.com/Stability-AI/stablediffusion, 2021.2.5 +[44] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score Jacobian Chaining: Lifting Pretrained 2D Diffusion Models for 3D Generation. IEEE Computer Vision and Pattern Recognition Conference, 2023. 3, 4 +[45] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-Fidelity and Diverse Text-to-3D Generation with Variational Score Distillation. Advances in Neural Information Processing Systems, 2023. 3, 4, 6 +[46] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel View Synthesis with Diffusion Models. International Conference on Learning Representations, 2023. 2 +[47] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis. arXiv preprint arXiv:2306.09341, 2023. 5, 6 +[48] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris Metaxas. StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks. International Conference on Computer Vision, 2017. 2 +[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. International Conference on Computer Vision, 2023. 2, 8 \ No newline at end of file diff --git a/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/images.zip b/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..36dd6d78102592c9b38adee2712af014232c576b --- /dev/null +++ b/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd9728a68738162bfd4f6689685df16cc8468a8f228975aace0b3d6f57526a5e +size 757143 diff --git a/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/layout.json b/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1daa7b89c8b0dd35233774a90454348ecf3b8a58 --- /dev/null +++ b/2024/SwiftBrush_ One-Step Text-to-Image Diffusion Model with Variational Score Distillation/layout.json @@ -0,0 +1,10495 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 130, + 103, + 463, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 103, + 463, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 103, + 463, + 138 + ], + "type": "text", + "content": "SwiftBrush: One-Step Text-to-Image Diffusion Model with Variational Score Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 210, + 161, + 383, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 161, + 383, + 188 + ], + "spans": [ + { + "bbox": [ + 210, + 161, + 383, + 188 + ], + "type": "text", + "content": "Thuan Hoang Nguyen Anh Tran VinAI Research, Vietnam" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 209, + 191, + 382, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 191, + 382, + 203 + ], + "spans": [ + { + "bbox": [ + 209, + 191, + 382, + 203 + ], + "type": "text", + "content": "{v.thuannh5,v.anhtt152}@vinai.io" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 49, + 209, + 297, + 456 + ], + "blocks": [ + { + "bbox": [ + 49, + 209, + 297, + 456 + ], + "lines": [ + { + "bbox": [ + 49, + 209, + 297, + 456 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 297, + 456 + ], + "type": "image", + "image_path": "abf8a724058c432da20e579aabfe4a3a5081d1395b2a1ea01e03a3f5f260f336.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 297, + 209, + 544, + 456 + ], + "blocks": [ + { + "bbox": [ + 297, + 209, + 544, + 456 + ], + "lines": [ + { + "bbox": [ + 297, + 209, + 544, + 456 + ], + "spans": [ + { + "bbox": [ + 297, + 209, + 544, + 456 + ], + "type": "image", + "image_path": "7ab8364f8db5c6fac9b15d022b4942be5534d5a9e5868e73ebbf389730da99e3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 49, + 456, + 296, + 581 + ], + "blocks": [ + { + "bbox": [ + 49, + 456, + 296, + 581 + ], + "lines": [ + { + "bbox": [ + 49, + 456, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 49, + 456, + 296, + 581 + ], + "type": "image", + "image_path": "6f37fd2f3a3fb5ad229907ccb50823ca6a0c03ff4f9a35b15d1ec1d9efe15ab7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 582, + 547, + 605 + ], + "lines": [ + { + "bbox": [ + 46, + 582, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 547, + 605 + ], + "type": "text", + "content": "Figure 1. Samples generated in a single step by SwiftBrush. Our distillation method can produce images of high-fidelity roughly 20 times faster compared to Stable Diffusion." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 296, + 456, + 544, + 581 + ], + "blocks": [ + { + "bbox": [ + 296, + 456, + 544, + 581 + ], + "lines": [ + { + "bbox": [ + 296, + 456, + 544, + 581 + ], + "spans": [ + { + "bbox": [ + 296, + 456, + 544, + 581 + ], + "type": "image", + "image_path": "3d95baf6c1b8a72331904f370a5c43577e5b4fc3ff9badb524a970b52e0892aa.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 616, + 192, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 616, + 192, + 628 + ], + "spans": [ + { + "bbox": [ + 143, + 616, + 192, + 628 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "Despite their ability to generate high-resolution and diverse images from text prompts, text-to-image diffusion models often suffer from slow iterative sampling processes. Model distillation is one of the most effective directions to accelerate these models. However, previous distillation methods fail to retain the generation quality while requir" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "ing a significant amount of images for training, either from real data or synthetically generated by the teacher model. In response to this limitation, we present a novel image-free distillation scheme named SwiftBrush. Drawing inspiration from text-to-3D synthesis, in which a 3D neural radiance field that aligns with the input prompt can be obtained from a 2D text-to-image diffusion prior via a specialized loss without the use of any 3D data ground-truth, our ap" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7807" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "proach re-purposes that same loss for distilling a pretrained multi-step text-to-image model to a student network that can generate high-fidelity images with just a single inference step. In spite of its simplicity, our model stands as one of the first one-step text-to-image generators that can produce images of comparable quality to Stable Diffusion without reliance on any training image data. Remarkably, SwiftBrush achieves an FID score of 16.67 and a CLIP score of 0.29 on the COCO-30K benchmark, achieving competitive results or even substantially surpassing existing state-of-the-art distillation techniques." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 227, + 128, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 227, + 128, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 227, + 128, + 239 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 247, + 289, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 289, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 289, + 474 + ], + "type": "text", + "content": "Diffusion models are gaining significant attention from the research community as they have achieved remarkable results in various generative tasks, including image generation [11, 26, 40, 41], video synthesis [12], 3D modeling [3, 28, 46], audio generation [17], and text creation [8, 18]. Especially, text-to-image diffusion models, which combine the power of language models and high-quality diffusion models [32, 36, 43], have revolutionized the way that people create visual content based on textual descriptions. Synthetic images indistinguishable from real photos are now possible with only a few clicks. Additionally, with techniques like ControlNet [49] or DreamBooth [35], extra control to the generation outputs can be made to further boost the expressiveness of the text-to-image models and bring them closer to mass users. However, despite their impressive success, these models exhibit a noteworthy drawback of slow inference speed. The primary reason for this lies in the iterative nature of their sampling, presenting significant challenges for deployment on consumer devices." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 475, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 289, + 714 + ], + "type": "text", + "content": "Several approaches have been proposed to improve the inference speed of diffusion-based text-to-image generation. However, the most effective direction is time-step distillation, which aims to reduce the number of sampling steps of the diffusion model with the network backbone unchanged and minimum reduction in generation quality. Starting from the pioneering work [37], more advanced and effective distillation methods have been proposed, and a few recent ones could train efficient student networks that only require one forward step to synthesize images from input captions. In this work, we are interested in the distillation-based approach for one-step text-to-image generation. Particularly, Meng et al. [25] employ a two-stage distillation, where the student tries to first match its output with the classifier-free guided output of the teacher and then bootstrap itself to reduce the inference steps gradually. Such a method is complex and time-consuming, requiring at least 108 A100 GPU days [21]. Based on the work of Song et al. [42], Latent Consistency Model (LCM) [24] enforces the consistency of all points on the ODE trajectory in the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": "latent space of Stable Diffusion (SD) [43] and unlike [25], they directly combine the two distillation stages into one. Although both Guided Distillation and LCM significantly reduce the number of inference steps to 2 or 4, their 1-step inference yields blurry and unsatisfying results. More sophisticated, Instaflow [21] proposes a technique called \"reflow\" to straighten the sampling trajectory of the teacher before distilling. While their proposed techniques proven to be superior to the original diffusion framework in one-step text-to-image distillation, InstaFlow requires an expensive 4-stage training schedule. Also, it is noteworthy that the effectiveness of these methods relies heavily on a large number of text-image pairs, which is not always the case due to the limited accessibility of such data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 240, + 547, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 547, + 504 + ], + "type": "text", + "content": "In this work, we aim to develop a novel distillation method for one-step text-to-image generation with higher quality and a more approachable training process characterized by an image-free and straightforward mechanism. Building upon the recent advancements in text-to-3D techniques, our work takes inspiration from their ability to generate high-quality 3D Neural Radiance Fields (NeRF) without utilizing 3D ground truth data. Such a remarkable achievement is facilitated by employing a powerful pretrained 2D text-to-image model to assess whether the image rendered from NeRF is realistic or not, akin to the discriminator in GANs. This observation underpins the motivation for our SwiftBrush approach, as we recognize that these principles can be innovatively adapted for text-to-image generation making image supervision unnecessary. SwiftBrush, therefore, emerges as a conceptually novel one-step text-to-image distillation process, uniquely capable of operating without 2D image supervision. This strategic adaptation not only opens up a change of view in the distillation of text-to-image diffusion models into a single-step generators but also points out the potential of blending principles from text-to-3D to text-to-image synthesis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 505, + 547, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 547, + 613 + ], + "type": "text", + "content": "As a result, our method has successfully enabled the generation of high-quality images with remarkable details. Quantitatively, our one-step model achieves promising zero-shot results, 16.6 for the FID score and 0.29 for the CLIP score on the MS COCO 2014 dataset (30K images). Notably, this is the first time a distilled one-step model outperforms the previous approaches without a single image used for training. A demonstration of generated images from our distilled model is shown in Fig. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 627, + 392, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 627, + 392, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 627, + 392, + 639 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Text-to-Image Generation. Previously, text-to-image models relied on GANs and focused on only small-scale object-centric domains such as flowers and birds [33, 48]. Recently, these types of systems have advanced rapidly with the availability of web-scale datasets of text-image pairs" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7808" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 373 + ], + "type": "text", + "content": "such as LAION5B [39], large language models such as T5 [30], and large vision-language models such as CLIP [29]. DALL-E [31] was the first autoregressive model to demonstrate a remarkable zero-shot ability in creating images from text captions by simply scaling up both the network and dataset size. Following that, a plethora of methods have been used for text-to-image generation, including masked generative transformers [4], GAN-based, and diffusion models. Based on Latent Diffusion Model [34], Stable Diffusion is an open-source diffusion-based generator that has gained widespread popularity among artists as well as researchers. Recently, text-to-image GANs are receiving attention again with notable work like StyleGAN-T [38], or GigaGAN [15]. While they are extremely fast as no iterative sampling is required, they lack behind diffusion models in terms of scalability and realism due to the fact that they suffer from instability and the infamous issue \"mode collapse\". To the best of our knowledge, so far only two above-mentioned work succeeded in applying GANs for text-to-image using tons of intricate techniques and auxiliary modules. Conversely, diffusion models can be easily scale up to generate high-quality and sharp samples, yet they require lengthy sampling process. Here, our objective is to create a large-scale generative model that combines the speed of GANs with the quality of diffusion models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 380, + 289, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 289, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 289, + 682 + ], + "type": "text", + "content": "Text-to-3D Generation. Instead of learning a generative model directly from 3D supervision, many research works have utilized the rich knowledge of 2D prior for 3D generation tasks. For instance, one of the earliest works, Dream Fields [14], leverages CLIP [29] to guide the generated images so that rendered images from many camera views highly align with the text caption. On the other hand, pioneer works like DreamFusion [28] and Score Jacobian Chaining [44] (SJC) propose two different but equivalent frameworks where a 2D text-to-image generative prior is used to generate 3D objects from textual descriptions. Subsequent works seek to improve these text-to-3D methods in various ways. Among them, Magic3D [19] and ProlificDreamer [45] are two remarkable works that effectively enhance the generation quality of 3D assets. Magic3D consists of two stages where they first obtain a low-resolution 3D representation via a DreamFusion-style optimization and then refine its texture to get the final high-fidelity 3D mesh. Meanwhile, ProlificDreamer proposes Variational Score Distillation (VSD) involving a secondary teacher to bridge the gap between the teacher and the 3D NeRF. Influenced by the latter work, our method adapts the same technique for one-step text-to-image generation. Similarly, Score GAN [6] and Diff-Instruct [7] use a VSD-like loss to distill a pretrained diffusion model to a one-step generator." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Knowledge Distillation introduced by Hinton et al. [10] is one of the methods falling under the umbrella of transfer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 301 + ], + "type": "text", + "content": "learning. This algorithm draws its main inspiration from the human learning process, where knowledge is conveyed from a more knowledgeable teacher to a less knowledgeable student. In the context of diffusion models, works have been proposed for faster sampling of student models from a pre-trained diffusion teacher. The most straightforward approach involves direct distillation [23], wherein a student model is supervised from the teacher's output after sampling for 50-1000 steps, which can be prohibitive. Recent work [37, 42] avoids the lengthy sampling during distillation by various bootstrapping methods. Afterward, a series of methods [21, 24, 25] adapt these methods into text-to-image settings and inherit their limitation of being image-dependent in training. In contrast, a concurrent work [9] learns a time-conditioned model to predict the outcome of the teacher at any given time step, all without the need for image supervision. Likewise, our approach is similarly image-free yet achieves significantly better results and offers a much simpler distillation design." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 309, + 411, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 309, + 411, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 309, + 411, + 323 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 329, + 385, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 329, + 385, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 329, + 385, + 342 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "text", + "content": "Diffusion Models consist of two processes: a forward process that gradually adds noise and another reverse process to predict the distribution of gradually denoised data points. The forward process incrementally introduces noise to an initial data point " + }, + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "text", + "content": ", drawn from the distribution " + }, + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "inline_equation", + "content": "q_0(x_0)" + }, + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "inline_equation", + "content": "q_t(x_t|x_0) = \\mathcal{N}(\\alpha_t x_0, \\sigma_t^2 I)" + }, + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "inline_equation", + "content": "\\{(\\alpha_t, \\sigma_t)\\}_{t=1}^T" + }, + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "text", + "content": " is the noise schedule. The data point " + }, + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "text", + "content": " at each time-step " + }, + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "text", + "content": " is drawn from " + }, + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "inline_equation", + "content": "q_t(x_t|x_0)" + }, + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "text", + "content": " using a noise " + }, + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 304, + 353, + 547, + 449 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 388, + 457, + 545, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 457, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 388, + 457, + 545, + 469 + ], + "type": "interline_equation", + "content": "x _ {t} = \\alpha_ {t} x _ {0} + \\sigma_ {t} \\epsilon . \\tag {1}", + "image_path": "2c645eb7c0c094a67988c664b53e8cb182b681079203f9b5995d10a2173ec671.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 475, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 499 + ], + "type": "text", + "content": "After " + }, + { + "bbox": [ + 304, + 475, + 545, + 499 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 475, + 545, + 499 + ], + "type": "text", + "content": " time steps, the final data point is expected to be pure Gaussian noise " + }, + { + "bbox": [ + 304, + 475, + 545, + 499 + ], + "type": "inline_equation", + "content": "x_{T}\\sim \\mathcal{N}(0,I)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "text", + "content": "Conversely, the backward process starts from " + }, + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "inline_equation", + "content": "x_{T} \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "text", + "content": " and iteratively denoises for " + }, + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "text", + "content": " steps where at each step " + }, + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "text", + "content": ", noisy variable " + }, + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "inline_equation", + "content": "x_{t+1}" + }, + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "text", + "content": " is converted to less-noisy variable " + }, + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "text", + "content": ", guided by a model " + }, + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}" + }, + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "text", + "content": " that predicts the noise " + }, + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "text", + "content": " in Eq. (1). The network weights " + }, + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 500, + 547, + 561 + ], + "type": "text", + "content": " are trained by minimizing:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 337, + 571, + 545, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 571, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 337, + 571, + 545, + 585 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {u n c o n d} (\\psi) = \\mathbb {E} _ {t, \\epsilon \\in \\mathcal {N} (0, 1)} \\| \\epsilon_ {\\psi} (x _ {t}, t) - \\epsilon \\| _ {2} ^ {2} \\tag {2}", + "image_path": "3a7e67f228d8f27cf74a948a0b78b81602b106abaf1c2f6943668f29954fadce.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 588, + 503, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 503, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 503, + 600 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 588, + 503, + 600 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 588, + 503, + 600 + ], + "type": "text", + "content": " is uniformly samples within " + }, + { + "bbox": [ + 304, + 588, + 503, + 600 + ], + "type": "inline_equation", + "content": "\\{1,\\dots ,T\\}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 600, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 600, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 545, + 661 + ], + "type": "text", + "content": "While common diffusion models operate in the pixel space, Latent Diffusion Models (LDMs)[34] model the diffusion process in the latent space of a pre-trained regularized autoencoder, which features representations of smaller size, thus enhancing both training and inference efficiency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Text-to-Image Diffusion Models. In contrast to unconditional diffusion models that generate outputs freely, text-conditioned diffusion models steer the sampling process using an extra prompt " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": ". This guides the model to produce" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7809" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "outputs that are not only photorealistic but also closely adhere to the provided text description. The objective to train such models is formulated as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 76, + 116, + 287, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 116, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 76, + 116, + 287, + 131 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d i f f} (\\psi) = \\mathbb {E} _ {t, y, \\epsilon \\in \\mathcal {N} (0, 1)} \\| \\epsilon_ {\\psi} (x _ {t}, t, y) - \\epsilon \\| _ {2} ^ {2} \\tag {3}", + "image_path": "0e58a53c14ddc03c763cc82a94060dc8c2659a115036ad6cee7ac127485a221e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 138, + 287, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 138, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 138, + 287, + 245 + ], + "type": "text", + "content": "which is slightly different from the unconditional diffusion loss Eq. (2). Due to the prompt conditioning, the model can deliver more controllable generation compared to its unconditional counterparts. Nevertheless, the implementations of many leading methods [1, 32, 36] remain inaccessible to the public. Stable Diffusion, which primarily utilizes the LDM framework, stands out as the first openly available large-scale model, significantly propelling text-to-image synthesis's widespread adoption and versatility." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "text", + "content": "Score Distillation Sampling (SDS) is a distillation technique effectively applied in generating 3D assets [19, 28, 44]. It utilizes a pre-trained text-to-image diffusion model, which predicts diffusion noise from text condition " + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}(x_t,t,y)" + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "text", + "content": ". The method optimizes a single 3D NeRF, parameterized by " + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "text", + "content": ", to align with a given text prompt. Given camera parameters " + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "text", + "content": ", a differentiable rendering function " + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "inline_equation", + "content": "g(\\cdot ,c)" + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "text", + "content": " is used to render an image at the camera view " + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "text", + "content": " from the 3D NeRF. Here, the rendered image " + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "inline_equation", + "content": "g(\\theta ,c)" + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "text", + "content": " is utilized to optimize the weight " + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "text", + "content": " through a loss function whose gradient can be approximated by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 390, + 287, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 287, + 418 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {S D S} = \\mathbb {E} _ {t, \\epsilon , c} \\left[ w (t) (\\epsilon_ {\\psi} (x _ {t}, t, y) - \\epsilon) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} \\right] (4)", + "image_path": "d44b3998014b41f682360098ca3bc9f212578b88030613ba39453e2cfeed89ae.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "type": "inline_equation", + "content": "t\\sim \\mathcal{U}(0.02T,0.98T)" + }, + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "type": "text", + "content": " is the maximum timesteps of the diffusion model, " + }, + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "type": "inline_equation", + "content": "x_{t} = \\alpha_{t}g(\\theta ,c) + \\sigma_{t}\\epsilon" + }, + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "type": "text", + "content": " is the input text, and " + }, + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 46, + 425, + 287, + 460 + ], + "type": "text", + "content": " is a weighting function" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 460, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 521 + ], + "type": "text", + "content": "Despite the advancements in text-to-3D synthesis, empirical studies [28, 44] indicate that SDS often encounters issues like over-saturation, over-smoothing, and reduced diversity. The same degradations can be observed if we naively apply SDS in our framework, as shown in Sec. 4.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 525, + 287, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 562 + ], + "type": "text", + "content": "Variational Score Distillation (VSD) is introduced in ProlificDreamer [45] to address previously mentioned issues of SDS by modifying the loss slightly:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 75, + 570, + 287, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 570, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 75, + 570, + 287, + 609 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\theta} \\mathcal {L} _ {V S D} = \\mathbb {E} _ {t, \\epsilon , c} [ w (t) (\\epsilon_ {\\psi} (x _ {t}, t, y) \\\\ \\left. - \\epsilon_ {\\phi} (x _ {t}, t, y, c)) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} \\right] \\tag {5} \\\\ \\end{array}", + "image_path": "370423e5509be7e9a862ea691d8519ba7c1c1ebc83b8f477e3d652bb1931303c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": "Again, " + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "inline_equation", + "content": "x_{t} = \\alpha_{t}g(\\theta ,c) + \\sigma_{t}\\epsilon" + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": " is the noisy observation of the rendered image at camera view " + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": ". VSD sets itself apart from SDS by introducing an additional score function tailored to the images rendered from the 3D NeRF at the camera pose " + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": ". This score is derived by fine-tuning a diffusion model " + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(x_t,t,y,c)" + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": " with the diffusion loss described below:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 696, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 696, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 104, + 696, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\min _ {\\epsilon_ {\\phi}} \\mathbb {E} _ {t, c, \\epsilon} \\| \\epsilon_ {\\phi} (x _ {t}, t, y, c) - \\epsilon \\| _ {2} ^ {2} \\tag {6}", + "image_path": "e618783819050bbc9eddc9f008b7217a9fb3fb57e978eae997c6ea9e4b2fe20f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "As proposed in ProlificDreamer, " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " is parameterized by Low-Rank Adaption [13] (LoRA) and initialized from the pre-trained diffusion model " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " with some added layers for conditioning the camera view " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": ". Note that at each iteration " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " of the optimization process, " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " needs to adapt with the current distribution of " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": ". Therefore, ProlificDreamer interleaves between finetuning " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " and optimizing " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": ". With these algorithmic enhancements, ProlificDreamer has significantly advanced its capabilities, enabling the generation of NeRFs and the creation of exceptional textured meshes. Such improvement directly inspires us to adapt VSD to the task of one-step text-to-image diffusion-based distillation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 223, + 382, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 223, + 382, + 235 + ], + "spans": [ + { + "bbox": [ + 306, + 223, + 382, + 235 + ], + "type": "text", + "content": "3.2. SwiftBrush" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 247, + 545, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 367 + ], + "type": "text", + "content": "Motivation. While SDS and VSD are designed explicitly for the text-to-3D generation task, they loosely connect to that goal via the rendered image " + }, + { + "bbox": [ + 304, + 247, + 545, + 367 + ], + "type": "inline_equation", + "content": "g(\\theta, c)" + }, + { + "bbox": [ + 304, + 247, + 545, + 367 + ], + "type": "text", + "content": " of the 3D NeRF. As a matter of fact, instead of using NeRF rendering, we can replace it with any function that outputs 2D images to suit our needs. Inspired by this motivation, we propose to substitute the NeRF rendering " + }, + { + "bbox": [ + 304, + 247, + 545, + 367 + ], + "type": "inline_equation", + "content": "g(\\theta, c)" + }, + { + "bbox": [ + 304, + 247, + 545, + 367 + ], + "type": "text", + "content": " with a text-to-image generator that can directly synthesize a text-guided image in one step, effectively converting the text-to-3D generation training into one-step diffusion model distillation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 372, + 545, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 577 + ], + "type": "text", + "content": "Design Space. We adopt the same approach as [45], with modifications to the design space to suit our task better. To begin with, we make use of two teacher models: a pretrained text-to-image teacher " + }, + { + "bbox": [ + 304, + 372, + 545, + 577 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}" + }, + { + "bbox": [ + 304, + 372, + 545, + 577 + ], + "type": "text", + "content": " and one additional LoRA teacher " + }, + { + "bbox": [ + 304, + 372, + 545, + 577 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 304, + 372, + 545, + 577 + ], + "type": "text", + "content": ". Also, we remove the conditioning on camera view " + }, + { + "bbox": [ + 304, + 372, + 545, + 577 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 372, + 545, + 577 + ], + "type": "text", + "content": " from the LoRA teacher as it is unnecessary in our case, and we use classifier-free guidance for both teachers. Then, we replace the NeRF, which is overfitted for a specific user-provided prompt in a text-to-3D setting, with a generalized one-step text-to-image student model " + }, + { + "bbox": [ + 304, + 372, + 545, + 577 + ], + "type": "inline_equation", + "content": "f_{\\theta}(z,y)" + }, + { + "bbox": [ + 304, + 372, + 545, + 577 + ], + "type": "text", + "content": ". Our student model " + }, + { + "bbox": [ + 304, + 372, + 545, + 577 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 372, + 545, + 577 + ], + "type": "text", + "content": " takes a random Gaussian noise " + }, + { + "bbox": [ + 304, + 372, + 545, + 577 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 372, + 545, + 577 + ], + "type": "text", + "content": " and a text prompt " + }, + { + "bbox": [ + 304, + 372, + 545, + 577 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 372, + 545, + 577 + ], + "type": "text", + "content": " as inputs. Both the LoRA teacher and student model are initialized with the weight of the text-to-image teacher. Next, we train both the student model and the LoRA teacher alternately using Eq. (5) and Eq. (6) while freezing the text-to-image teacher. A pseudo-code and system figure can be viewed in Alg. 1 and Fig. 2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Student Parameterization. Given a pretrained text-to-image diffusion model " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", it is possible to directly use its output for the student model, i.e., " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}(z,y) = \\epsilon_{\\theta}(z,T,y)" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " is the maximum timestep of the pretrained model. However, in our case, the pretrained model of choice is Stable Diffusion, which is inherently designed to predict the added noise " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ". In contrast, our objective is to refine the student model so that it predicts a clean and noise-free " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ". Therefore, such a naive method results in a large domain gap between what we want the student to learn and the output of the student. For the sake of easy training, we empiri" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7810" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 198, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 198, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 198, + 83 + ], + "type": "text", + "content": "Algorithm 1 SwiftBrush Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 50, + 87, + 287, + 327 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "text", + "content": "1: Require: a pretrained text-to-image teacher " + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}" + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "text", + "content": ", a LoRA teacher " + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "text", + "content": ", a student model " + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "text", + "content": ", two learning rates " + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "inline_equation", + "content": "\\eta_{1}" + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "inline_equation", + "content": "\\eta_{2}" + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "text", + "content": ", a weighting function " + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "text", + "content": ", a prompts dataset " + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "text", + "content": ", the maximum number of time steps " + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "text", + "content": " and the noise schedule " + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "inline_equation", + "content": "\\{(\\alpha_t, \\sigma_t)\\}_{t=1}^T" + }, + { + "bbox": [ + 53, + 87, + 287, + 148 + ], + "type": "text", + "content": " of the teacher model" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 172, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 172, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 172, + 158 + ], + "type": "text", + "content": "2: Initialize: " + }, + { + "bbox": [ + 53, + 148, + 172, + 158 + ], + "type": "inline_equation", + "content": "\\phi \\gets \\psi, \\theta \\gets \\psi" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 162, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 162, + 171 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 162, + 171 + ], + "type": "text", + "content": "3: while not converged do" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 171, + 211, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 211, + 183 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 211, + 183 + ], + "type": "text", + "content": "4: Sample input noise " + }, + { + "bbox": [ + 53, + 171, + 211, + 183 + ], + "type": "inline_equation", + "content": "z\\sim \\mathcal{N}(0,I)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 184, + 212, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 212, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 212, + 194 + ], + "type": "text", + "content": "5: Sample text caption input " + }, + { + "bbox": [ + 53, + 184, + 212, + 194 + ], + "type": "inline_equation", + "content": "y\\sim Y" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 195, + 235, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 235, + 207 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 235, + 207 + ], + "type": "text", + "content": "6: Compute student output " + }, + { + "bbox": [ + 53, + 195, + 235, + 207 + ], + "type": "inline_equation", + "content": "\\hat{x}_0 = f_\\theta (z,y)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 208, + 235, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 208, + 235, + 219 + ], + "spans": [ + { + "bbox": [ + 53, + 208, + 235, + 219 + ], + "type": "text", + "content": "7: Sample timestep " + }, + { + "bbox": [ + 53, + 208, + 235, + 219 + ], + "type": "inline_equation", + "content": "t \\sim \\mathcal{U}(0.02T, 0.98T)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 220, + 213, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 213, + 231 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 213, + 231 + ], + "type": "text", + "content": "8: Sample added noise " + }, + { + "bbox": [ + 53, + 220, + 213, + 231 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 232, + 244, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 232, + 244, + 243 + ], + "spans": [ + { + "bbox": [ + 53, + 232, + 244, + 243 + ], + "type": "text", + "content": "9: Compute noisy sample " + }, + { + "bbox": [ + 53, + 232, + 244, + 243 + ], + "type": "inline_equation", + "content": "\\hat{x}_t = \\alpha_t\\hat{x}_0 + \\sigma_t\\epsilon" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 50, + 243, + 283, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 243, + 283, + 256 + ], + "spans": [ + { + "bbox": [ + 50, + 243, + 283, + 256 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 50, + 243, + 283, + 256 + ], + "type": "inline_equation", + "content": "\\theta \\gets \\theta -\\eta_{1}\\left[\\omega (t)\\left(\\epsilon_{\\psi}(\\hat{x}_{t},t,y) - \\epsilon_{\\phi}(\\hat{x}_{t},t,y)\\right)\\frac{\\partial\\hat{x}_{0}}{\\partial\\theta}\\right]" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 50, + 256, + 201, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 256, + 201, + 266 + ], + "spans": [ + { + "bbox": [ + 50, + 256, + 201, + 266 + ], + "type": "text", + "content": "11: Sample timestep " + }, + { + "bbox": [ + 50, + 256, + 201, + 266 + ], + "type": "inline_equation", + "content": "t^{\\prime}\\sim \\mathcal{U}(0,T)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 50, + 267, + 216, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 267, + 216, + 278 + ], + "spans": [ + { + "bbox": [ + 50, + 267, + 216, + 278 + ], + "type": "text", + "content": "12: Sample added noise " + }, + { + "bbox": [ + 50, + 267, + 216, + 278 + ], + "type": "inline_equation", + "content": "\\epsilon^{\\prime}\\sim \\mathcal{N}(0,I)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 50, + 279, + 255, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 279, + 255, + 290 + ], + "spans": [ + { + "bbox": [ + 50, + 279, + 255, + 290 + ], + "type": "text", + "content": "13: Compute noisy sample " + }, + { + "bbox": [ + 50, + 279, + 255, + 290 + ], + "type": "inline_equation", + "content": "\\hat{x}_{t'} = \\alpha_{t'}\\hat{x}_0 + \\sigma_{t'}\\epsilon'" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 50, + 290, + 227, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 290, + 227, + 303 + ], + "spans": [ + { + "bbox": [ + 50, + 290, + 227, + 303 + ], + "type": "text", + "content": "14: " + }, + { + "bbox": [ + 50, + 290, + 227, + 303 + ], + "type": "inline_equation", + "content": "\\phi \\gets \\phi -\\eta_{2}\\nabla_{\\phi}\\| \\epsilon_{\\phi}(\\hat{x}_{t^{\\prime}},t^{\\prime},y) - \\epsilon^{\\prime}\\|^{2}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 50, + 304, + 108, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 304, + 108, + 313 + ], + "spans": [ + { + "bbox": [ + 50, + 304, + 108, + 313 + ], + "type": "text", + "content": "15: end while" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 50, + 315, + 195, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 315, + 195, + 327 + ], + "spans": [ + { + "bbox": [ + 50, + 315, + 195, + 327 + ], + "type": "text", + "content": "16: return trained student model " + }, + { + "bbox": [ + 50, + 315, + 195, + 327 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 350, + 254, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 254, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 254, + 363 + ], + "type": "text", + "content": "cally re-parameterize the student output as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 369, + 228, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 369, + 228, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 228, + 396 + ], + "type": "interline_equation", + "content": "f _ {\\theta} (z, y) = \\frac {z - \\sigma_ {T} \\epsilon_ {\\theta} (z , T , y)}{\\alpha_ {T}}", + "image_path": "2adbba6c729e25bc8f68721bc392b652a83070ad86d29b0ab01d8558ad166077.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 402, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 473 + ], + "type": "text", + "content": "which is a realization of Eq. (1) if we set " + }, + { + "bbox": [ + 46, + 402, + 287, + 473 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 46, + 402, + 287, + 473 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 402, + 287, + 473 + ], + "type": "inline_equation", + "content": "x_{t} = z" + }, + { + "bbox": [ + 46, + 402, + 287, + 473 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 402, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\epsilon \\approx \\epsilon_{\\theta}(z,T,y)" + }, + { + "bbox": [ + 46, + 402, + 287, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 402, + 287, + 473 + ], + "type": "inline_equation", + "content": "x_0 \\approx f_\\theta (z,y)" + }, + { + "bbox": [ + 46, + 402, + 287, + 473 + ], + "type": "text", + "content": ". So basically, this reparameterization transforms the noise-prediction output of the pretrained model into the \"predicted " + }, + { + "bbox": [ + 46, + 402, + 287, + 473 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 402, + 287, + 473 + ], + "type": "text", + "content": "\" form, which is empirically demonstrated in Sec. 4.3 to be helpful for the student model to learn." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 484, + 128, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 128, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 128, + 498 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 504, + 164, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 164, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 164, + 517 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 528, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 685 + ], + "type": "text", + "content": "Evaluation Metrics. We conduct our evaluation process by assessing the performance of SwiftBrush against several well-established methods, such as Guided Distillation [25], Instaflow [21], LCM [24], and BOOT [9], across two zero-shot text-to-image benchmarks: COCO 2014 [20] and Human Preference Score v2 [47] (HPSv2). On the COCO 2014, we follow the conventional evaluation protocol [15, 34, 36, 38], computing FID and CLIP score on a 30K subset of its validation set as our primary metrics. Here, we use ViT-G/14 as the backbone for evaluating CLIP scores. In the case of HPSv2, we adopt their evaluation procedure to assess text-image alignment across four different aspects: anime, concept-art, paintings, and photos." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Experiment Protocols. In these experiments, we rely on Stable Diffusion v2.1 [43] for our two teacher models, and" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "whenever possible, we directly utilize the reported values from other research. As for InstaFlow and LCM, we use the their provided pretrained models to generate images for evaluation. Conversely, we are compelled to re-implement BOOT since they did not provide any metrics for their text-to-image models, and their codebase remains entirely closed." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 161, + 546, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 161, + 546, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 161, + 546, + 413 + ], + "type": "text", + "content": "Implementation Details. Similar to prior research, we use student models with the same architectures as those of the teachers and initialize all of the student parameters with the teacher model. During training, a guidance scale of 4.5 is used for both teachers. For the LoRA teacher model, we apply a learning rate of 1e-3, a LoRA rank of 64, and an alpha value of 108. Meanwhile, the student model's learning rate is set at 1e-6 with exponential moving average (EMA) used every iteration. Adam optimizer [16] and a batch size of 64 (16 per GPU) are used to train both the student and LoRA teacher for around 65K iterations. Thanks to the image-free nature of SwiftBrush, we do not require access to any image training set, and only the prompts are needed to distill the generator. Here, we utilize 4M captions from the large-scale text-image dataset JourneyDB [27], which we subsequently deduplicated to 2M. Regarding BOOT, we closely follow their implementation guidelines to the best of our abilities and train the student model until its convergence, except that we use Stable Diffusion v2.1 instead of DeepFloyd IF [5] as the teacher. Here, its blurry results (Fig. 4) compared to the original work could be due to the different teacher." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 420, + 362, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 362, + 432 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 362, + 432 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 444, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 545, + 685 + ], + "type": "text", + "content": "Quantitative Results. We first evaluate the proposed method on standard text-to-image generation benchmarks. The quantitative comparison with the previous distillation, both image-dependent and image-free, and the original Stable Diffusion teacher is shown in Tab. 1. Despite being trained with only text captions, SwiftBrush significantly improves the quality of 1-step inference and achieves better photo-realism as well as text-image alignment against other methods, with exceptions of InstaFlow and Stable Diffusion (25 sampling steps) in FID score. However, it is worth noting that Instaflow is so complex, requiring a 4-stage distillation process, while ours is end-to-end and straightforward. We further conduct a quantitative evaluation with the HPSv2 score, a metric trained to capture human preference. Again, we obtain the best score in all aspects among all mentioned works, which is only a slight degradation compared to the original Stable Diffusion teacher (Tab. 2). Moreover, according to the anonymous survey carried out by us (Fig. 3), our method outperforms Instaflow in terms of image fidelity and text-to-image accuracy." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "content": "Qualitative Results. We report the visual comparison between our method and the mentioned baseline methods" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7811" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 68, + 531, + 228 + ], + "blocks": [ + { + "bbox": [ + 62, + 68, + 531, + 228 + ], + "lines": [ + { + "bbox": [ + 62, + 68, + 531, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 68, + 531, + 228 + ], + "type": "image", + "image_path": "7a3b1c07f625c04d4463492843c84e7923ad2df0f8e49b8c07e897915371bc37.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 232, + 548, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 548, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 548, + 289 + ], + "type": "text", + "content": "Figure 2. SwiftBrush overview. Our system allows training a one-step text-to-image student network from a frozen pretrained teacher with the help of an additional trainable LORA teacher. The student network takes inputs a text prompt and random noise. The output of the student is then added noise and sent to two teachers, alongside the prompt and a randomly drawn timestep, to compute the gradient of VSD loss, which backpropagates back to the student. Apart from that, the LoRA teacher is also updated using diffusion loss. Similar to ProlificDreamer [45], we update the LoRA teacher (green flow) and the student (red flow) alternatively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 301, + 279, + 361 + ], + "blocks": [ + { + "bbox": [ + 56, + 301, + 279, + 361 + ], + "lines": [ + { + "bbox": [ + 56, + 301, + 279, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 301, + 279, + 361 + ], + "type": "image", + "image_path": "812107473a718a74917a5075987c817a44a80465a3c8221981a22789f003d793.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 363, + 288, + 407 + ], + "lines": [ + { + "bbox": [ + 46, + 363, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 288, + 407 + ], + "type": "text", + "content": "Figure 3. User Study. We ask 400 users to compare the generated images of InstaFlow and our method using 30 random text prompts. We report the rate at which each method is selected as better for image quality and text-to-image accuracy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 418, + 287, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 634 + ], + "type": "text", + "content": "for the one-step text-to-image generation task. As seen in Fig. 4, LCM outputs blurry and over-smoothed images in the 1-step regime. BOOT faithfully produces images that match well with the text caption, but its results still fall short in terms of fidelity. In contrast, a much better quality can be observed in the image produced by Instaflow and SwiftBrush. InstaFlow, nevertheless, cannot handle complex or unusual input prompts properly, which can be noticed in the third and fifth images at the third column of Fig. 4. On top of that, it occasionally produces non-realistic results even for simple prompts like in Fig. 4 at the fourth image of the third column. When, in fact, the outputs generated by SwiftBrush are not only of exceptional quality but also exhibit a remarkable alignment between text descriptions and the corresponding images. Moreover, our results are comparable with those sampled from Stable Diffusion with 25 steps; whereas it is understandable that with only one step, Stable Diffusion cannot create outputs of good quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 641, + 111, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 111, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 111, + 654 + ], + "type": "text", + "content": "4.3. Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "The importance of LoRA teacher. In Fig. 6, the importance of including the LoRA teacher in the training process is clearly highlighted. We compare the student outputs sampled with different input noises after 10K iteration when us" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 300, + 553, + 402 + ], + "blocks": [ + { + "bbox": [ + 307, + 300, + 553, + 402 + ], + "lines": [ + { + "bbox": [ + 307, + 300, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 307, + 300, + 553, + 402 + ], + "type": "table", + "html": "
MethodStepsFID-30K ↓CLIP-30K ↑
Guided Distillation†137.30.27
LCM†135.560.24
Instaflow113.10†0.28§
BOOT‡148.200.26
Ours116.670.29
SD 2.1*2513.450.30
SD 2.1*1202.140.08
", + "image_path": "1af7892897cf09d0c3720e0bc8c98a03f171e5f4aee1fba4111afd353d3727e8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 510, + 550, + 599 + ], + "blocks": [ + { + "bbox": [ + 304, + 406, + 547, + 506 + ], + "lines": [ + { + "bbox": [ + 304, + 406, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 547, + 506 + ], + "type": "text", + "content": "Table 1. Comparison of our method against other works based on FID metric and CLIP score on the COCO 2014 dataset. " + }, + { + "bbox": [ + 304, + 406, + 547, + 506 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 304, + 406, + 547, + 506 + ], + "type": "text", + "content": " means that we obtain the numbers from the corresponding papers. " + }, + { + "bbox": [ + 304, + 406, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\S" + }, + { + "bbox": [ + 304, + 406, + 547, + 506 + ], + "type": "text", + "content": " means that we obtain the numbers using the provided pretrained models of the corresponding papers. " + }, + { + "bbox": [ + 304, + 406, + 547, + 506 + ], + "type": "inline_equation", + "content": "^\\ddagger" + }, + { + "bbox": [ + 304, + 406, + 547, + 506 + ], + "type": "text", + "content": " means that we re-implement the work and report the numbers. * means that we report the number using the pretrained models. We use DPM-Solver [22] with guidance scale 7.5 to sample for SD. Bold and underlined numbers are best and second best for one-step models." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 510, + 550, + 599 + ], + "lines": [ + { + "bbox": [ + 307, + 510, + 550, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 550, + 599 + ], + "type": "table", + "html": "
ModelsHuman Preference Score v2 ↑
AnimePhotoConcept ArtPaintings
LCM†22.6122.7122.7422.91
InstaFlow†25.9826.3225.7925.93
BOOT‡25.2925.1624.4024.61
Ours26.9127.2126.3226.37
SD 2.1*27.4826.8926.8627.46
", + "image_path": "3c337de8ad7dfb648225cd8bf1e422bed0a1d94106395a25543df60faf0620b3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 602, + 547, + 669 + ], + "lines": [ + { + "bbox": [ + 304, + 602, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 547, + 669 + ], + "type": "text", + "content": "Table 2. Comparison of our method against other works based on HPSv2 score in 1-step regime. " + }, + { + "bbox": [ + 304, + 602, + 547, + 669 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 602, + 547, + 669 + ], + "type": "text", + "content": " means that we obtain the score using the provided pretrained models of the corresponding papers. " + }, + { + "bbox": [ + 304, + 602, + 547, + 669 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 304, + 602, + 547, + 669 + ], + "type": "text", + "content": " means that we re-implement the work and report the score by ourselves. * means that we obtain the score from [47]. Bold and underlined numbers are best and second best, respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "ing the same text prompts. With the guidance of the LoRA teacher, the student progressively learns to create realistic images (" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " column). In contrast, in the absence of an ef" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7812" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 85, + 201, + 145 + ], + "blocks": [ + { + "bbox": [ + 49, + 106, + 136, + 126 + ], + "lines": [ + { + "bbox": [ + 49, + 106, + 136, + 126 + ], + "spans": [ + { + "bbox": [ + 49, + 106, + 136, + 126 + ], + "type": "text", + "content": "\"A photo of one ice cream ball in a luxurious plate, bokeh\"" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 73, + 178, + 82 + ], + "lines": [ + { + "bbox": [ + 159, + 73, + 178, + 82 + ], + "spans": [ + { + "bbox": [ + 159, + 73, + 178, + 82 + ], + "type": "text", + "content": "LCM" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 140, + 85, + 201, + 145 + ], + "lines": [ + { + "bbox": [ + 140, + 85, + 201, + 145 + ], + "spans": [ + { + "bbox": [ + 140, + 85, + 201, + 145 + ], + "type": "image", + "image_path": "5afe2db2e4630a5b55287ad4601b7e8c615c01c2cde9c70e5d47e925dbb659a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 208, + 85, + 268, + 145 + ], + "blocks": [ + { + "bbox": [ + 225, + 74, + 249, + 82 + ], + "lines": [ + { + "bbox": [ + 225, + 74, + 249, + 82 + ], + "spans": [ + { + "bbox": [ + 225, + 74, + 249, + 82 + ], + "type": "text", + "content": "BOOT" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 208, + 85, + 268, + 145 + ], + "lines": [ + { + "bbox": [ + 208, + 85, + 268, + 145 + ], + "spans": [ + { + "bbox": [ + 208, + 85, + 268, + 145 + ], + "type": "image", + "image_path": "91101967545bb2a4f16931bb2f2845e6a46114454cc147105779e5bae158bbc8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 274, + 85, + 334, + 145 + ], + "blocks": [ + { + "bbox": [ + 284, + 73, + 321, + 82 + ], + "lines": [ + { + "bbox": [ + 284, + 73, + 321, + 82 + ], + "spans": [ + { + "bbox": [ + 284, + 73, + 321, + 82 + ], + "type": "text", + "content": "InstaFlow" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 274, + 85, + 334, + 145 + ], + "lines": [ + { + "bbox": [ + 274, + 85, + 334, + 145 + ], + "spans": [ + { + "bbox": [ + 274, + 85, + 334, + 145 + ], + "type": "image", + "image_path": "4a1c23d3cae446d089bd37a27efa62ef901c65cc4a7f65d8cf4537df000a7b91.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 340, + 85, + 400, + 145 + ], + "blocks": [ + { + "bbox": [ + 359, + 74, + 378, + 83 + ], + "lines": [ + { + "bbox": [ + 359, + 74, + 378, + 83 + ], + "spans": [ + { + "bbox": [ + 359, + 74, + 378, + 83 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 85, + 400, + 145 + ], + "lines": [ + { + "bbox": [ + 340, + 85, + 400, + 145 + ], + "spans": [ + { + "bbox": [ + 340, + 85, + 400, + 145 + ], + "type": "image", + "image_path": "953b8d02bc1827b355999659e5b319eef44e1c04682236372358b482a5e0f3ee.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 407, + 85, + 467, + 144 + ], + "blocks": [ + { + "bbox": [ + 428, + 73, + 443, + 82 + ], + "lines": [ + { + "bbox": [ + 428, + 73, + 443, + 82 + ], + "spans": [ + { + "bbox": [ + 428, + 73, + 443, + 82 + ], + "type": "text", + "content": "SD*" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 407, + 85, + 467, + 144 + ], + "lines": [ + { + "bbox": [ + 407, + 85, + 467, + 144 + ], + "spans": [ + { + "bbox": [ + 407, + 85, + 467, + 144 + ], + "type": "image", + "image_path": "3d42bd44d1942e79d7fa32065bfe84469810d77bc08bf624df4d2f4557bc0981.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 474, + 85, + 533, + 144 + ], + "blocks": [ + { + "bbox": [ + 495, + 73, + 510, + 82 + ], + "lines": [ + { + "bbox": [ + 495, + 73, + 510, + 82 + ], + "spans": [ + { + "bbox": [ + 495, + 73, + 510, + 82 + ], + "type": "text", + "content": "SD†" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 474, + 85, + 533, + 144 + ], + "lines": [ + { + "bbox": [ + 474, + 85, + 533, + 144 + ], + "spans": [ + { + "bbox": [ + 474, + 85, + 533, + 144 + ], + "type": "image", + "image_path": "3b30057df9197efae4d798cd083105401b0d39043f567c50dfaa4d2fef30979d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 141, + 150, + 201, + 210 + ], + "blocks": [ + { + "bbox": [ + 49, + 167, + 123, + 197 + ], + "lines": [ + { + "bbox": [ + 49, + 167, + 123, + 197 + ], + "spans": [ + { + "bbox": [ + 49, + 167, + 123, + 197 + ], + "type": "text", + "content": "\"A hyperrealistic photo of fox astronaut, perfect face, artstation\"" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 150, + 201, + 210 + ], + "lines": [ + { + "bbox": [ + 141, + 150, + 201, + 210 + ], + "spans": [ + { + "bbox": [ + 141, + 150, + 201, + 210 + ], + "type": "image", + "image_path": "198cf500f3369f314c6f9b73aa59f9eb38bd5ca99dfbbdd48f7d2e304377e5b7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 208, + 151, + 267, + 210 + ], + "blocks": [ + { + "bbox": [ + 208, + 151, + 267, + 210 + ], + "lines": [ + { + "bbox": [ + 208, + 151, + 267, + 210 + ], + "spans": [ + { + "bbox": [ + 208, + 151, + 267, + 210 + ], + "type": "image", + "image_path": "5a05e9c607af55c062d85492723da54c305015052f89a6bf3ae261e01a9ebb3e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 274, + 151, + 334, + 210 + ], + "blocks": [ + { + "bbox": [ + 274, + 151, + 334, + 210 + ], + "lines": [ + { + "bbox": [ + 274, + 151, + 334, + 210 + ], + "spans": [ + { + "bbox": [ + 274, + 151, + 334, + 210 + ], + "type": "image", + "image_path": "2802d2ad5aa3e0613f924267ed3b1df9599a4d16aa0be8173d4fb14cb7f9467e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 340, + 151, + 400, + 210 + ], + "blocks": [ + { + "bbox": [ + 340, + 151, + 400, + 210 + ], + "lines": [ + { + "bbox": [ + 340, + 151, + 400, + 210 + ], + "spans": [ + { + "bbox": [ + 340, + 151, + 400, + 210 + ], + "type": "image", + "image_path": "d88a1d3d373a34ed133ddf1ab1d5e01f9e623ddc8d1f09871d904e20dd39def7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 407, + 151, + 467, + 210 + ], + "blocks": [ + { + "bbox": [ + 407, + 151, + 467, + 210 + ], + "lines": [ + { + "bbox": [ + 407, + 151, + 467, + 210 + ], + "spans": [ + { + "bbox": [ + 407, + 151, + 467, + 210 + ], + "type": "image", + "image_path": "ee112dac7cd86a08f1cb8f84b4bd08294a42b04bba0984daf3394ec9cea4aaa7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 474, + 151, + 533, + 210 + ], + "blocks": [ + { + "bbox": [ + 474, + 151, + 533, + 210 + ], + "lines": [ + { + "bbox": [ + 474, + 151, + 533, + 210 + ], + "spans": [ + { + "bbox": [ + 474, + 151, + 533, + 210 + ], + "type": "image", + "image_path": "cffa696eb135c8a6bd8d9db1eece800e019afdd9d895ae23e1d1e4fecc1de0f2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 141, + 217, + 200, + 276 + ], + "blocks": [ + { + "bbox": [ + 49, + 235, + 130, + 262 + ], + "lines": [ + { + "bbox": [ + 49, + 235, + 130, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 235, + 130, + 262 + ], + "type": "text", + "content": "\"A poodle wearing a baseball cap and holding a dictionary in hand\"" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 217, + 200, + 276 + ], + "lines": [ + { + "bbox": [ + 141, + 217, + 200, + 276 + ], + "spans": [ + { + "bbox": [ + 141, + 217, + 200, + 276 + ], + "type": "image", + "image_path": "88aed34cc8fce6a13270bb3647dfedd284471c814722c2d73aec0fcd80b6cd3f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 208, + 217, + 267, + 276 + ], + "blocks": [ + { + "bbox": [ + 208, + 217, + 267, + 276 + ], + "lines": [ + { + "bbox": [ + 208, + 217, + 267, + 276 + ], + "spans": [ + { + "bbox": [ + 208, + 217, + 267, + 276 + ], + "type": "image", + "image_path": "c61dc697f40e133c0948744834072aee2afea166f1f4d9da81fe21b2043b054a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 274, + 217, + 334, + 276 + ], + "blocks": [ + { + "bbox": [ + 274, + 217, + 334, + 276 + ], + "lines": [ + { + "bbox": [ + 274, + 217, + 334, + 276 + ], + "spans": [ + { + "bbox": [ + 274, + 217, + 334, + 276 + ], + "type": "image", + "image_path": "6c2be08869aafead182b3a99722a5ef6bd8e67f5ebe1eb59ea17c56305dec90b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 340, + 217, + 400, + 276 + ], + "blocks": [ + { + "bbox": [ + 340, + 217, + 400, + 276 + ], + "lines": [ + { + "bbox": [ + 340, + 217, + 400, + 276 + ], + "spans": [ + { + "bbox": [ + 340, + 217, + 400, + 276 + ], + "type": "image", + "image_path": "1c20f5069bcc119d1bfcd1e24eac80ab6fe68622566b6cef6d01f718c9b6a10e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 407, + 217, + 467, + 276 + ], + "blocks": [ + { + "bbox": [ + 407, + 217, + 467, + 276 + ], + "lines": [ + { + "bbox": [ + 407, + 217, + 467, + 276 + ], + "spans": [ + { + "bbox": [ + 407, + 217, + 467, + 276 + ], + "type": "image", + "image_path": "9494333f178d2054b331dfa7029502b30b69c4af76bf11d1d06c4caee5e4b0e7.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 474, + 217, + 533, + 276 + ], + "blocks": [ + { + "bbox": [ + 474, + 217, + 533, + 276 + ], + "lines": [ + { + "bbox": [ + 474, + 217, + 533, + 276 + ], + "spans": [ + { + "bbox": [ + 474, + 217, + 533, + 276 + ], + "type": "image", + "image_path": "d5e8727c06927e472018ad6682530a3c88ca24151b1bcc443f1632709c988736.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 141, + 282, + 200, + 342 + ], + "blocks": [ + { + "bbox": [ + 49, + 304, + 137, + 324 + ], + "lines": [ + { + "bbox": [ + 49, + 304, + 137, + 324 + ], + "spans": [ + { + "bbox": [ + 49, + 304, + 137, + 324 + ], + "type": "text", + "content": "\"A DSLR photo of a cat jumping over a fence, high-res\"" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 282, + 200, + 342 + ], + "lines": [ + { + "bbox": [ + 141, + 282, + 200, + 342 + ], + "spans": [ + { + "bbox": [ + 141, + 282, + 200, + 342 + ], + "type": "image", + "image_path": "767c986b23b50af3ec467e281e6e182be2e8721010c92fe0fae17f817a5a5573.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 208, + 282, + 267, + 342 + ], + "blocks": [ + { + "bbox": [ + 208, + 282, + 267, + 342 + ], + "lines": [ + { + "bbox": [ + 208, + 282, + 267, + 342 + ], + "spans": [ + { + "bbox": [ + 208, + 282, + 267, + 342 + ], + "type": "image", + "image_path": "7f3de0616d99be0965264fb0e9cea175eb4943ba181db7cae8d9abfbed8ad8ef.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 274, + 282, + 334, + 342 + ], + "blocks": [ + { + "bbox": [ + 274, + 282, + 334, + 342 + ], + "lines": [ + { + "bbox": [ + 274, + 282, + 334, + 342 + ], + "spans": [ + { + "bbox": [ + 274, + 282, + 334, + 342 + ], + "type": "image", + "image_path": "c756f0536b9274f20d230c3f0f7fe509f8f73ed556e416083f64b2ed3ac1fc32.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 340, + 282, + 400, + 342 + ], + "blocks": [ + { + "bbox": [ + 340, + 282, + 400, + 342 + ], + "lines": [ + { + "bbox": [ + 340, + 282, + 400, + 342 + ], + "spans": [ + { + "bbox": [ + 340, + 282, + 400, + 342 + ], + "type": "image", + "image_path": "6b8d3915dd4d9f1db9226db95f51a357b3116d9e48b5a162e55af7da33b72ba9.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 407, + 282, + 467, + 342 + ], + "blocks": [ + { + "bbox": [ + 407, + 282, + 467, + 342 + ], + "lines": [ + { + "bbox": [ + 407, + 282, + 467, + 342 + ], + "spans": [ + { + "bbox": [ + 407, + 282, + 467, + 342 + ], + "type": "image", + "image_path": "718d2356631f7e89b9989d6b342bfb19e0948cacbf0ca8dd52db759d94a101d2.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 474, + 282, + 533, + 342 + ], + "blocks": [ + { + "bbox": [ + 474, + 282, + 533, + 342 + ], + "lines": [ + { + "bbox": [ + 474, + 282, + 533, + 342 + ], + "spans": [ + { + "bbox": [ + 474, + 282, + 533, + 342 + ], + "type": "image", + "image_path": "9e430d7675b11c63f1599edbe97b9d0fe9aa5132118181f753077f2b085def92.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 141, + 348, + 200, + 407 + ], + "blocks": [ + { + "bbox": [ + 49, + 372, + 132, + 391 + ], + "lines": [ + { + "bbox": [ + 49, + 372, + 132, + 391 + ], + "spans": [ + { + "bbox": [ + 49, + 372, + 132, + 391 + ], + "type": "text", + "content": "\"A blue Porsche 356 parked in front of a brick wall\"" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 348, + 200, + 407 + ], + "lines": [ + { + "bbox": [ + 141, + 348, + 200, + 407 + ], + "spans": [ + { + "bbox": [ + 141, + 348, + 200, + 407 + ], + "type": "image", + "image_path": "72200b3d12a6c4589659236250a96de56ff83b8891cdc74da9a74c2c17546696.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 208, + 349, + 267, + 407 + ], + "blocks": [ + { + "bbox": [ + 208, + 349, + 267, + 407 + ], + "lines": [ + { + "bbox": [ + 208, + 349, + 267, + 407 + ], + "spans": [ + { + "bbox": [ + 208, + 349, + 267, + 407 + ], + "type": "image", + "image_path": "9558855b7e62142bff08d080aea99866561af9c5b2b5ea0e00b7bbb1235c8cb7.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 274, + 349, + 334, + 407 + ], + "blocks": [ + { + "bbox": [ + 274, + 349, + 334, + 407 + ], + "lines": [ + { + "bbox": [ + 274, + 349, + 334, + 407 + ], + "spans": [ + { + "bbox": [ + 274, + 349, + 334, + 407 + ], + "type": "image", + "image_path": "53a9506e177150ca358fe4b03b7d7cb1f4866e98c6eb9a1fb3a0d93a47cbdde2.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 340, + 349, + 400, + 407 + ], + "blocks": [ + { + "bbox": [ + 340, + 349, + 400, + 407 + ], + "lines": [ + { + "bbox": [ + 340, + 349, + 400, + 407 + ], + "spans": [ + { + "bbox": [ + 340, + 349, + 400, + 407 + ], + "type": "image", + "image_path": "ecfab9d3c6bad519d4d60292f46e964fce8b74b78a416c021f3a62ac1effc549.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 407, + 349, + 467, + 407 + ], + "blocks": [ + { + "bbox": [ + 407, + 349, + 467, + 407 + ], + "lines": [ + { + "bbox": [ + 407, + 349, + 467, + 407 + ], + "spans": [ + { + "bbox": [ + 407, + 349, + 467, + 407 + ], + "type": "image", + "image_path": "34439a0b6d48491b9a1e85c35db6f26833ff6dafec1e78a021612ee91ef0ee1e.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 474, + 349, + 533, + 400 + ], + "blocks": [ + { + "bbox": [ + 474, + 349, + 533, + 400 + ], + "lines": [ + { + "bbox": [ + 474, + 349, + 533, + 400 + ], + "spans": [ + { + "bbox": [ + 474, + 349, + 533, + 400 + ], + "type": "image", + "image_path": "b4d7b8b131cb3ff407055a3d828c64017c9078e698eb248914072f26e0aa17e2.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 141, + 414, + 200, + 473 + ], + "blocks": [ + { + "bbox": [ + 49, + 436, + 121, + 455 + ], + "lines": [ + { + "bbox": [ + 49, + 436, + 121, + 455 + ], + "spans": [ + { + "bbox": [ + 49, + 436, + 121, + 455 + ], + "type": "text", + "content": "\"A photograph of an astronaut riding a horse\"" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 414, + 200, + 473 + ], + "lines": [ + { + "bbox": [ + 141, + 414, + 200, + 473 + ], + "spans": [ + { + "bbox": [ + 141, + 414, + 200, + 473 + ], + "type": "image", + "image_path": "51bc833455ced53b6bae2fb5d9b801db20508affd9b6e752b004ef46b25c8862.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 476, + 546, + 510 + ], + "lines": [ + { + "bbox": [ + 46, + 476, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 546, + 510 + ], + "type": "text", + "content": "Figure 4. Sample images generated by LCM, BOOT, Instaflow, Ours, SD with 25 sampling steps " + }, + { + "bbox": [ + 46, + 476, + 546, + 510 + ], + "type": "inline_equation", + "content": "(\\mathrm{SD}^{*})" + }, + { + "bbox": [ + 46, + 476, + 546, + 510 + ], + "type": "text", + "content": " and SD with 1 sampling step " + }, + { + "bbox": [ + 46, + 476, + 546, + 510 + ], + "type": "inline_equation", + "content": "(\\mathrm{SD}^{\\dagger})" + }, + { + "bbox": [ + 46, + 476, + 546, + 510 + ], + "type": "text", + "content": ". Images at the same row are sampled from the same text prompt, while images at the same column are from the same model. We use DPM-Solver [22] with guidance scale 7.5 to sample for Stable Diffusion" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 208, + 415, + 267, + 473 + ], + "blocks": [ + { + "bbox": [ + 208, + 415, + 267, + 473 + ], + "lines": [ + { + "bbox": [ + 208, + 415, + 267, + 473 + ], + "spans": [ + { + "bbox": [ + 208, + 415, + 267, + 473 + ], + "type": "image", + "image_path": "3fc4ba236e59f17378472457c5054115c817e078fd9ec3a2b9014b3e46d2365a.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 274, + 415, + 334, + 473 + ], + "blocks": [ + { + "bbox": [ + 274, + 415, + 334, + 473 + ], + "lines": [ + { + "bbox": [ + 274, + 415, + 334, + 473 + ], + "spans": [ + { + "bbox": [ + 274, + 415, + 334, + 473 + ], + "type": "image", + "image_path": "47edaade28b7b17e8d421056103dedf98db42e9834571288a1aed386337e8fe8.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 340, + 415, + 400, + 473 + ], + "blocks": [ + { + "bbox": [ + 340, + 415, + 400, + 473 + ], + "lines": [ + { + "bbox": [ + 340, + 415, + 400, + 473 + ], + "spans": [ + { + "bbox": [ + 340, + 415, + 400, + 473 + ], + "type": "image", + "image_path": "9fe2b24dfa901decdc8f8c8780d92f25b2722760e3f782db62e48432d56c1774.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 407, + 415, + 467, + 473 + ], + "blocks": [ + { + "bbox": [ + 407, + 415, + 467, + 473 + ], + "lines": [ + { + "bbox": [ + 407, + 415, + 467, + 473 + ], + "spans": [ + { + "bbox": [ + 407, + 415, + 467, + 473 + ], + "type": "image", + "image_path": "2cce683dbc060cbeca3d012aa87879dde8ec52aad7db6fb69ced25cef37bbb15.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 474, + 414, + 533, + 473 + ], + "blocks": [ + { + "bbox": [ + 474, + 414, + 533, + 473 + ], + "lines": [ + { + "bbox": [ + 474, + 414, + 533, + 473 + ], + "spans": [ + { + "bbox": [ + 474, + 414, + 533, + 473 + ], + "type": "image", + "image_path": "2944e461664cfdd2b50223ffdc22521eba30a97ae369d4114c25efe8f99cf33c.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "type": "text", + "content": "fective LoRA teacher, the student's early training outputs, despite being highly aligned with the text prompt, are overly saturated and unrealistic (" + }, + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 46, + 520, + 287, + 604 + ], + "type": "text", + "content": " column). These results indicate a fundamental breakdown in the student model's ability to mimic the teacher model's distribution, as can be further demonstrated based on the flat red curve in both FID and CLIP score in Fig. 5 (a) and (b)." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "In addition, even when employing the LoRA teacher, it is crucial to use a sufficiently large rank in the LoRA weight for stable distillation of the student model. Without this, the model tends to undergo severe mode collapse despite initially producing reasonably good results. The green curves in Fig. 5 clearly illustrate that, while the FID score of a student model with a 4-rank LoRA weight initially decreases, it deteriorates starting from 8K steps, resulting in lousy images in Fig. 6 (" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " column). Additionally, the correspond" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 305, + 520, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 520, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 305, + 520, + 545, + 556 + ], + "type": "text", + "content": "ing CLIP score behaves similarly where there is a slight rise before it worsens in the end. Besides, we also provide extra visuals for this ablative study in Supplementary." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 305, + 563, + 546, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 546, + 671 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 546, + 671 + ], + "type": "text", + "content": "The importance of student parameterization. Apart from the LoRA teacher, the re-parameterization of the student (Sec. 3) also plays an important role in the SwiftBrush training. Without it, the student is unable to create realistic images and it only produces noisy results, as can be noticed from the " + }, + { + "bbox": [ + 305, + 563, + 546, + 671 + ], + "type": "inline_equation", + "content": "4^{th}" + }, + { + "bbox": [ + 305, + 563, + 546, + 671 + ], + "type": "text", + "content": " column of Fig. 6. This can be further backed up by the orange curve in both FID and CLIP score in Fig. 5 (a) and (b), where student without re-parameterization performs consistently worst among all configurations." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "type": "text", + "content": "Interpolating Latent. We conduct a controlled experiment where the noise input " + }, + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "type": "text", + "content": " remains the same while varying the input prompt " + }, + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "type": "text", + "content": " (Fig. 7). Interestingly, the SwiftBrush-" + } + ] + } + ], + "index": 53 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7813" + } + ] + } + ], + "index": 54 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 71, + 272, + 198 + ], + "blocks": [ + { + "bbox": [ + 74, + 71, + 272, + 198 + ], + "lines": [ + { + "bbox": [ + 74, + 71, + 272, + 198 + ], + "spans": [ + { + "bbox": [ + 74, + 71, + 272, + 198 + ], + "type": "image", + "image_path": "4e54008329a3726d42c9a4db0721ed675e41c16246872bb9b899ca1c7ed200af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 201, + 233, + 211 + ], + "lines": [ + { + "bbox": [ + 113, + 201, + 233, + 211 + ], + "spans": [ + { + "bbox": [ + 113, + 201, + 233, + 211 + ], + "type": "text", + "content": "(a) FID curves of four configurations" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 519, + 198 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 519, + 198 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 519, + 198 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 519, + 198 + ], + "type": "image", + "image_path": "a96d66bfdb07dec0ab90d570494aa5022ce35a4bc6cace3c35b17cfad03e28b6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 201, + 482, + 211 + ], + "lines": [ + { + "bbox": [ + 358, + 201, + 482, + 211 + ], + "spans": [ + { + "bbox": [ + 358, + 201, + 482, + 211 + ], + "type": "text", + "content": "(b) CLIP curves of four configurations" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 290, + 105, + 346 + ], + "blocks": [ + { + "bbox": [ + 46, + 224, + 545, + 280 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 545, + 280 + ], + "type": "text", + "content": "Figure 5. Ablation study on the impact of the LoRA teacher as well as the student parameterization. We train all four configurations with same hyperparameters described in Sec. 4 but only with 10K iterations to observe their training dynamics at the early phase. Then, we evaluate FID and CLIP score on COCO 2014 but with only a 5K subset of images. Here, Full means all of our proposed techniques are used. No Parameterization means we exclude the re-parameterization trick. Small-Rank LoRA means we use only 4-rank LoRA teacher. SDS means we remove the LoRA teacher and use only SDS loss." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 290, + 105, + 346 + ], + "lines": [ + { + "bbox": [ + 50, + 290, + 105, + 346 + ], + "spans": [ + { + "bbox": [ + 50, + 290, + 105, + 346 + ], + "type": "image", + "image_path": "56374b2b8e00c8ae4ea1f812a4d152dde9a6d323c8ede91b9f1df4384c491b87.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 110, + 290, + 165, + 346 + ], + "blocks": [ + { + "bbox": [ + 110, + 290, + 165, + 346 + ], + "lines": [ + { + "bbox": [ + 110, + 290, + 165, + 346 + ], + "spans": [ + { + "bbox": [ + 110, + 290, + 165, + 346 + ], + "type": "image", + "image_path": "1278c181d97f8233cc8a83baf20c28b234c0760582307bbfbd2a6290a296d298.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 170, + 290, + 225, + 346 + ], + "blocks": [ + { + "bbox": [ + 170, + 290, + 225, + 346 + ], + "lines": [ + { + "bbox": [ + 170, + 290, + 225, + 346 + ], + "spans": [ + { + "bbox": [ + 170, + 290, + 225, + 346 + ], + "type": "image", + "image_path": "2676c2a63ed719b00ee0eb3086fd3ef3ee531072e70fc188d81486a705c3f73a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 230, + 291, + 285, + 346 + ], + "blocks": [ + { + "bbox": [ + 230, + 291, + 285, + 346 + ], + "lines": [ + { + "bbox": [ + 230, + 291, + 285, + 346 + ], + "spans": [ + { + "bbox": [ + 230, + 291, + 285, + 346 + ], + "type": "image", + "image_path": "c72a8c976337dcccbcb2e2bfc92d78297b8a2ad8238dd10a3159f5719f2a07dc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 50, + 352, + 105, + 407 + ], + "blocks": [ + { + "bbox": [ + 50, + 352, + 105, + 407 + ], + "lines": [ + { + "bbox": [ + 50, + 352, + 105, + 407 + ], + "spans": [ + { + "bbox": [ + 50, + 352, + 105, + 407 + ], + "type": "image", + "image_path": "65c964b6513a2a9cf10c08970cef03cfbaa8e50cb16615299a775255e533f260.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 408, + 81, + 414 + ], + "lines": [ + { + "bbox": [ + 71, + 408, + 81, + 414 + ], + "spans": [ + { + "bbox": [ + 71, + 408, + 81, + 414 + ], + "type": "text", + "content": "Full" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 110, + 352, + 165, + 407 + ], + "blocks": [ + { + "bbox": [ + 110, + 352, + 165, + 407 + ], + "lines": [ + { + "bbox": [ + 110, + 352, + 165, + 407 + ], + "spans": [ + { + "bbox": [ + 110, + 352, + 165, + 407 + ], + "type": "image", + "image_path": "b33d8b218629f196ece9c81e315ec48c3c28a93e540bcb0e6746ff3d2a515109.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 409, + 141, + 415 + ], + "lines": [ + { + "bbox": [ + 130, + 409, + 141, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 141, + 415 + ], + "type": "text", + "content": "SDS" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 170, + 352, + 225, + 407 + ], + "blocks": [ + { + "bbox": [ + 170, + 352, + 225, + 407 + ], + "lines": [ + { + "bbox": [ + 170, + 352, + 225, + 407 + ], + "spans": [ + { + "bbox": [ + 170, + 352, + 225, + 407 + ], + "type": "image", + "image_path": "1903b1873a776c9f4c50dd198a5a1976d0613f21da996bf24b71ee80195a624b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 408, + 217, + 415 + ], + "lines": [ + { + "bbox": [ + 179, + 408, + 217, + 415 + ], + "spans": [ + { + "bbox": [ + 179, + 408, + 217, + 415 + ], + "type": "text", + "content": "Small-Rank LoRA" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 418, + 287, + 463 + ], + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 463 + ], + "type": "text", + "content": "Figure 6. Visual results of the ablative study, where the first two rows are generated with the input prompt \"An oil painting of a train\", whereas the remaining two are sampled from \"An DSLR photo of a tiger in the city\"." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 230, + 351, + 285, + 407 + ], + "blocks": [ + { + "bbox": [ + 230, + 351, + 285, + 407 + ], + "lines": [ + { + "bbox": [ + 230, + 351, + 285, + 407 + ], + "spans": [ + { + "bbox": [ + 230, + 351, + 285, + 407 + ], + "type": "image", + "image_path": "95281ca8eb89ba6bc26561b51b94cf8cfbbdb31cf356cd8c8e163eea047a2551.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 408, + 280, + 415 + ], + "lines": [ + { + "bbox": [ + 235, + 408, + 280, + 415 + ], + "spans": [ + { + "bbox": [ + 235, + 408, + 280, + 415 + ], + "type": "text", + "content": "No Parameterization" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 475, + 286, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 286, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 286, + 546 + ], + "type": "text", + "content": "distilled student seemingly features enhanced control and editability compared to the teacher model due to its one-step nature. Additionally, we present further results of latent space interpolation in Supplementary. In this process, the student model, distilled from the Stable Diffusion teacher, demonstrates a seamless transition in its image generation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 560, + 198, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 198, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 198, + 572 + ], + "type": "text", + "content": "5. Conclusion and Discussion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 587, + 286, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 286, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 286, + 682 + ], + "type": "text", + "content": "Conclusion. In this work, we have introduced a new image-free distillation scheme, SwiftBrush, which leverages insights from text-to-3D synthesis to speed up text-to-image generation. Empirically, the results presented in this paper demonstrate the effectiveness of our approach in making text-to-image generation more efficient and accessible. We hope our work will encourage broader use of text-to-image generation systems by greatly lowering the inference cost." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "type": "text", + "content": "Discussion. SwiftBrush inevitably produces lower quality samples compared to the teacher which utilizes multi-" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 309, + 290, + 425, + 466 + ], + "blocks": [ + { + "bbox": [ + 309, + 290, + 425, + 466 + ], + "lines": [ + { + "bbox": [ + 309, + 290, + 425, + 466 + ], + "spans": [ + { + "bbox": [ + 309, + 290, + 425, + 466 + ], + "type": "image", + "image_path": "b5a1e2da3fb42bb5a9fe0f7bf16b49d51850457dbe776a8108349ef52a90bdf5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 470, + 545, + 525 + ], + "lines": [ + { + "bbox": [ + 305, + 470, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 470, + 545, + 525 + ], + "type": "text", + "content": "Figure 7. Results of interpolating the input prompts. The prompts used here are selected from a standard template: 'A DSLR photo of a {animal} {action}' Here, 'animal' is one of following: owl, raccoon, cat, or panda, while 'action' is one of following: reading a book, drinking a latte or eating a pizza." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 425, + 291, + 543, + 466 + ], + "blocks": [ + { + "bbox": [ + 425, + 291, + 543, + 466 + ], + "lines": [ + { + "bbox": [ + 425, + 291, + 543, + 466 + ], + "spans": [ + { + "bbox": [ + 425, + 291, + 543, + 466 + ], + "type": "image", + "image_path": "df50ac732abc1f17f8bb88fe3977c5f8de8d64c5468da9cdab41da267283253a.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "step sampling for inference. Another drawback is that the current design only focuses on single-step student model and cannot support few-step generation as did in Luo et al. [24] to improve image fidelity. In future work, we aim to investigate the feasibility of extending SwiftBrush to a few-step generation so that we can trade computation for quality. Furthermore, we also find it compelling to explore SwiftBrush-like training where only one teacher is required. Such an exploration could provide tremendous benefits in scenarios where the computational resource for training is limited. Finally, it is intriguing to see if techniques such as DreamBooth [35], ControlNet [49] or InstructPix2Pix[2] can be integrated with SwiftBrush, allowing instant generation of various applications." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7814" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. eDiff-I: Text-to-Image Diffusion Models with an Ensemble of Expert Denoisers. arXiv preprint arXiv:2211.01324, 2022. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "text", + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. InstructPix2Pix: Learning to Follow Image Editing Instructions. IEEE Computer Vision and Pattern Recognition Conference, 2023. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 267 + ], + "type": "text", + "content": "[3] Eric R. Chan, Koki Nagano, Matthew A. Chan, Alexander W. Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. GeNVS: Generative Novel View Synthesis with 3D-Aware Diffusion Models. International Conference on Computer Vision, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 269, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 334 + ], + "type": "text", + "content": "[4] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T. Freeman, Michael Rubinstein, Yuzhen Li, and Dilip Krishnan. Muse: Text-To-Image Generation via Masked Generative Transformers. arXiv preprint arXiv:2301.00704, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 335, + 286, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 286, + 357 + ], + "type": "text", + "content": "[5] DeepFloyd. DeepFloyd IF. https://www deepfloyd.ai/deepfloyd-if, 2023.5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 358, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 413 + ], + "type": "text", + "content": "[6] Jean-Yves Franceschi, Mike Gartrell, Ludovic Dos Santos, Thibaut Issenhuth, Emmanuel de Bezenac, Mickael Chen, and Alain Rakotomamonjy. Unifying GANs and Score-Based Diffusion as Generative Particle Models. Advances in Neural Information Processing Systems, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 414, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 479 + ], + "type": "text", + "content": "[7] Jean-Yves Franceschi, Mike Gartrell, Ludovic Dos Santos, Thibaut Issenhuth, Emmanuel de Bezenac, Mickael Chen, and Alain Rakotomamonjy. Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models. Advances in Neural Information Processing Systems, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 480, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 523 + ], + "type": "text", + "content": "[8] Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and Lingpeng Kong. DiffuSeq: Sequence to Sequence Text Generation with Diffusion Models. International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 525, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 287, + 567 + ], + "type": "text", + "content": "[9] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Lingjie Liu, and Josh Susskind. BOOT: Data-free Distillation of Denoising Diffusion Models with Bootstrapping. arXiv preprint arXiv:2306.05544, 2023. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "type": "text", + "content": "[10] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the Knowledge in a Neural Network. NIPS Deep Learning Workshop, 2014. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. Advances in Neural Information Processing Systems, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "text", + "content": "[12] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video Diffusion Models. Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 714 + ], + "type": "text", + "content": "[13] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Models. International Conference on Learning Representations, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "text", + "content": "[14] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-Shot Text-Guided Object Generation with Dream Fields. IEEE Computer Vision and Pattern Recognition Conference, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "text", + "content": "[15] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up GANs for Text-to-Image Synthesis. IEEE Computer Vision and Pattern Recognition Conference, 2023. 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 217 + ], + "type": "text", + "content": "[16] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. International Conference on Learning Representations, 2015. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "text", + "content": "[17] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A Versatile Diffusion Model for Audio Synthesis. International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "text", + "content": "[18] Xiang Lisa Li, John Thickstun, Ishaan Gulrajani, Percy Liang, and Tatsunori Hashimoto. Diffusion-LM: Improves Controllable Text Generation. Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 305, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 360 + ], + "type": "text", + "content": "[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-Resolution Text-to-3D Content Creation. IEEE Computer Vision and Pattern Recognition Conference, 2023. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 361, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 404 + ], + "type": "text", + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. European Conference on Computer Vision, 2014. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 405, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 448 + ], + "type": "text", + "content": "[21] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation. arXiv preprint arXiv:2309.06380, 2023. 2, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "text", + "content": "[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps. Advances in Neural Information Processing Systems, 2022. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "text", + "content": "[23] Eric Luhman and Troy Luhman. Knowledge Distillation in Iterative Generative Models for Improved Sampling Speed. arXiv preprint arXiv:2101.02388, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "text", + "content": "[24] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2, 3, 5, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "text", + "content": "[25] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik P. Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On Distillation of Guided Diffusion Models. IEEE Computer Vision and Pattern Recognition Conference, 2023. 2, 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "type": "text", + "content": "[26] Alexander Quinn Nichol and Prafulla Dhariwal. Improved Denoising Diffusion Probabilistic Models. International Conference on Machine Learning, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "type": "text", + "content": "[27] Junting Pan, Keqiang Sun, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. JourneyDB: A Benchmark for Generative Image Understanding." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7815" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "Advances in Neural Information Processing Systems, 2023. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[28] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion. International Conference on Learning Representations, 2023. 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 286, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 286, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 286, + 206 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning Transferable Visual Models from Natural Language Supervision. International Conference on Machine Learning, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 286, + 262 + ], + "type": "text", + "content": "[30] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 286, + 308 + ], + "type": "text", + "content": "[31] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-Shot Text-to-Image Generation. International Conference on Machine Learning, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 310, + 286, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 286, + 353 + ], + "type": "text", + "content": "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv preprint arXiv:2204.06125, 2022. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 286, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 286, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 286, + 399 + ], + "type": "text", + "content": "[33] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative Adversarial Text to Image Synthesis. International Conference on Machine Learning, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 286, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 286, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 286, + 443 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. IEEE Computer Vision and Pattern Recognition Conference, 2022. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 286, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 286, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 286, + 499 + ], + "type": "text", + "content": "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-image Diffusion Models for Subject-Driven Generation. IEEE Computer Vision and Pattern Recognition Conference, 2023. 2, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 501, + 286, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 286, + 578 + ], + "type": "text", + "content": "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 2022. 2, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 286, + 612 + ], + "type": "text", + "content": "[37] Tim Salimans and Jonathan Ho. Progressive Distillation for Fast Sampling of Diffusion Models. International Conference on Learning Representations, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 286, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 657 + ], + "type": "text", + "content": "[38] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. StyleGAN-T: Unlocking the Power of GANs for Fast Large-Scale Text-to-Image Synthesis. International Conference on Machine Learning, 2023. 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "type": "text", + "content": "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 576 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Jitsev. LAION-5B: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "text", + "content": "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep Unsupervised Learning using Nonequilibrium Thermodynamics. International Conference on Machine Learning, 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 152, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 205 + ], + "type": "text", + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 208, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 240 + ], + "type": "text", + "content": "[42] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency Models. International Conference on Machine Learning, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 241, + 545, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 264 + ], + "type": "text", + "content": "[43] StabilityAI. Stable Diffusion. https://github.com/Stability-AI/stablediffusion, 2021.2.5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "text", + "content": "[44] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score Jacobian Chaining: Lifting Pretrained 2D Diffusion Models for 3D Generation. IEEE Computer Vision and Pattern Recognition Conference, 2023. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 320, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 374 + ], + "type": "text", + "content": "[45] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-Fidelity and Diverse Text-to-3D Generation with Variational Score Distillation. Advances in Neural Information Processing Systems, 2023. 3, 4, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 376, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 430 + ], + "type": "text", + "content": "[46] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel View Synthesis with Diffusion Models. International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 432, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 432, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 432, + 545, + 485 + ], + "type": "text", + "content": "[47] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis. arXiv preprint arXiv:2306.09341, 2023. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 487, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 487, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 308, + 487, + 545, + 542 + ], + "type": "text", + "content": "[48] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris Metaxas. StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks. International Conference on Computer Vision, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 544, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 576 + ], + "type": "text", + "content": "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. International Conference on Computer Vision, 2023. 2, 8" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7816" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/ce409cac-fd5a-4614-9e6d-87e790ca611e_content_list.json b/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/ce409cac-fd5a-4614-9e6d-87e790ca611e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..40c89c5629141d3a1b599f7b1a95146d1003b0ce --- /dev/null +++ b/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/ce409cac-fd5a-4614-9e6d-87e790ca611e_content_list.json @@ -0,0 +1,1659 @@ +[ + { + "type": "text", + "text": "SwitchLight: Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting", + "text_level": 1, + "bbox": [ + 186, + 130, + 782, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hoon Kim1 Minje Jang1 Wonjun Yoon1 Jisoo Lee1 Donghyun Na1 Sanghyun Woo2", + "bbox": [ + 106, + 202, + 861, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Beeble AI $^{2}$ New York University", + "bbox": [ + 315, + 237, + 648, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c3a9957805d3c58bc988053a24741f13d3fa4a2761520550a6037eb132557f0e.jpg", + "image_caption": [ + "Input RGB", + "Normal", + "Albedo", + "Roughness", + "Reflectivity", + "Figure 1. Be Anywhere at Any Time. SwitchLight processes a human portrait by decomposing it into detailed intrinsic components, and re-renders the image under a designated target illumination, ensuring a seamless composition of the subject into any new environment." + ], + "image_footnote": [], + "bbox": [ + 81, + 304, + 890, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 696, + 313, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce a co-designed approach for human portrait relighting that combines a physics-guided architecture with a pre-training framework. Drawing on the Cook-Torrance reflectance model, we have meticulously configured the architecture design to precisely simulate light-surface interactions. Furthermore, to overcome the limitation of scarce high-quality lightstage data, we have developed a self-supervised pre-training strategy. This novel combination of accurate physical modeling and expanded training dataset establishes a new benchmark in relighting realism.", + "bbox": [ + 75, + 715, + 470, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 694, + 630, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Relighting is more than an aesthetic tool; it unlocks infinite narrative possibilities and enables seamless integration of subjects into diverse environments (see Fig. 1). This advancement resonates with our innate desire to transcend the physical constraints of space and time, while also providing tangible solutions to practical challenges in digital content creation. It is particularly transformative in virtual (VR) and augmented reality (AR) applications, where relighting facilitates the real-time adaptation of lighting, ensuring that users and digital elements coexist naturally within any environment, offering a next level of telepresence.", + "bbox": [ + 496, + 719, + 890, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we focus on human portrait relighting.", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* All authors contributed equally to this work", + "bbox": [ + 94, + 875, + 333, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*https://www.beeble.ai", + "bbox": [ + 96, + 887, + 218, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25096", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While the relighting task fundamentally demands an in-depth understanding of geometry, material properties, and illumination, the challenge is more compounded when addressing human subjects, due to the unique characteristics of skin surfaces as well as the diverse textures and reflectance properties of a wide array of clothing, hairstyles, and accessories. These elements interact in complex ways, necessitating advanced algorithms capable of simulating the subtle interplay of light with these varied surfaces.", + "bbox": [ + 75, + 90, + 470, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Currently, the most promising approach involves the use of deep neural networks trained on pairs of high-quality relict portrait images and their corresponding intrinsic attributes, which are sourced from a light stage setup [10]. Initial efforts approached the relighting process as a 'black box' [44, 47], without delving into the underlying mechanisms. Later advancements adopted a physics-guided model design, incorporating the explicit modeling of image intrinsics and image formation physics [31]. Pandey et al. [33] proposed the Total Relight (TR) architecture, also physics-guided, which decomposes an input image into surface normals and albedo maps, and performs relighting based on the Phong specular reflectance model. The TR architecture has become foundational model for image relighting, with most recent and advanced architectures building upon its principle [22, 30, 51].", + "bbox": [ + 75, + 234, + 470, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following the physics-guided approach, our contribution lies in a co-design of architecture with a self-supervised pre-training framework. First, our architecture evolves towards a more accurate physical model by integrating the Cook-Torrance specular reflectance model [8], representing a notable advancement from the empirical Phong specular model [36] employed in the Total Relight architecture. The Cook-Torrance model adeptly simulates light interactions with surface microfacets, accounting for spatially varying roughness and reflectivity. Secondly, our pretraining framework scales the learning process beyond the typically hard-to-obtain lightstage data. By revisiting the masked autoencoder (MAE) framework [18], we adept it for the task of relighting. These modifications are crafted to address the unique challenges posed by this task, enabling our model to learn from unlabelled data and refine its ability to produce realistic relit portraits during fine-tuning. To the best of our knowledge, this is the first time applying self-supervised pre-training specifically to the relighting task.", + "bbox": [ + 75, + 484, + 468, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contribution is twofold. Firstly, by enhancing the physical reflectance model, we have introduced a new level of realism in the output. Secondly, by adopting self-supervised learning, we have expanded the scale of the training data and enhanced the expression of lighting in diverse real-world scenarios. Collectively, these advancements have led SwitchLight framework to achieve a new state-of-the-art in human portrait relighting.", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Portrait Relighting is an ill-posed problem due to its under-constrained nature. To tackle this, earlier methods incorporated 3D facial priors [43], exploited image intrinsics [3, 39], or framed the task as a style transfer [42]. Light stage techniques [48] offer a more powerful solution by recording subject's reflectance fields under varying lighting conditions [10, 13], though they are labor-intensive and require specialized equipment. A promising alternative has emerged with deep learning, utilizing neural networks trained on light stage data. Sun et al. [44] pioneered this approach, but their method had limitations in representing non-Lambertian effects. This was improved upon by Nestmeyer et al. [31], who integrated rendering physics into network design, albeit limited to directional light. Building upon this, Pandey et al. [33] incorporated the Phong reflection model and a high dynamic range (HDR) lighting map [9] into their network, enabling a more accurate representation of global illumination. Simultaneously, efforts have been made to explore portrait relighting without light stage data [19, 20, 41, 46, 54]. Moreover, introduction of NeRF [7] and diffusion-based [37] models has opened new avenues in the field. However, networks trained with light-stage data maintain superior accuracy and realism, thanks to physics-based composited relight image training pairs and precise ground truth image intrinsics [55].", + "bbox": [ + 496, + 121, + 890, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work furthers this domain by integrating the Cook-Torrance model into our network design, shifting from the empirical Phong model to a more physics-based approach, thereby enhancing the realism and detail in relit images.", + "bbox": [ + 496, + 503, + 890, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised Pre-training has become a standard training scheme in the development of large language models like BERT [11] and GPT [38], and is increasingly influential in vision models, aiming to replicate the 'BERT moment'. This approach typically involves pre-training on extensive unlabeled data, followed by fine-tuning on specific tasks. While early efforts in vision models focused on simple pretext tasks [12, 16, 32, 35, 52], the field has evolved through stages like contrastive learning [5, 17] and masked image modeling [2, 18, 49]. However, the primary focus has remained on visual recognition, with less attention to other domains. Exceptions include low-level image processing tasks [4, 6, 26, 29] using the vision transformer [14].", + "bbox": [ + 496, + 577, + 890, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our research takes a different route, focusing on human portrait relighting—a complex challenge of manipulating illumination in the image. This direction is crucial because acquiring accurate ground truth data, especially from light stage, is both expensive and difficult. We modify the MAE framework [18], previously successful in robust image representation learning and developing locality biases [34], to suit the unique requirements of effective relighting.", + "bbox": [ + 496, + 780, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25097", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. SwitchLight", + "text_level": 1, + "bbox": [ + 76, + 89, + 205, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce SwitchLight, a state-of-the-art framework for human portrait relighting, with its architectural overview presented in Fig. 2. We first provide foundational concepts in Sec. 3.1, and define the problem in Sec. 3.2. This is followed by detailing the architecture in Sec. 3.3, and lastly, we describe the loss functions used in Sec. 3.4.", + "bbox": [ + 75, + 114, + 468, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 214, + 217, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, vectors $\\mathbf{n}$ , $\\mathbf{v}$ , $\\mathbf{l}$ , and $\\mathbf{h}$ are denoted as unit vectors. Specifically, $\\mathbf{n}$ represents the surface normal, $\\mathbf{v}$ is the view direction, $\\mathbf{l}$ is the incident light direction, and $\\mathbf{h}$ is the half-vector computed from $\\mathbf{l}$ and $\\mathbf{v}$ . The dot product is clamped between [0..1], indicated by $\\langle \\cdot \\rangle$ .", + "bbox": [ + 75, + 237, + 468, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Rendering. The primary goal of image rendering is to create a visual representation that accurately simulates the interactions between light and surfaces. These complex interactions are encapsulated by the rendering equation:", + "bbox": [ + 75, + 330, + 468, + 391 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {o} (\\mathbf {v}) = \\int_ {\\Omega} f (\\mathbf {v}, \\mathbf {l}) L _ {i} (\\mathbf {l}) \\langle \\mathbf {n} \\cdot \\mathbf {l} \\rangle d \\mathbf {l} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 400, + 468, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $L_{o}(\\mathbf{v})$ denotes the radiance, or the light intensity perceived by the observer in direction $\\mathbf{v}$ . It is the cumulative result of incident lights $L_{i}(\\mathbf{l})$ from all possible directions over the hemisphere, $\\Omega$ , centered around the surface normal, denoted as $\\mathbf{n}$ . Central to this equation lies the Bidirectional Reflectance Distribution Function (BRDF), denoted as $f(\\mathbf{v},\\mathbf{l})$ , describing the surface's reflection characteristics.", + "bbox": [ + 75, + 439, + 468, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "BRDF Composition. The BRDF, represented by $f(\\mathbf{v},\\mathbf{l})$ , describes how light is reflected at an opaque surface. It is composed of two major components: diffuse reflection ( $f_{d}$ ) and specular reflection ( $f_{s}$ ):", + "bbox": [ + 75, + 563, + 468, + 625 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf (\\mathbf {v}, \\mathbf {l}) = f _ {d} (\\mathbf {v}, \\mathbf {l}) + f _ {s} (\\mathbf {v}, \\mathbf {l}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 633, + 468, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A surface intrinsically exhibits both diffuse and specular reflections. The diffuse component uniformly scatters light, ensuring consistent illumination regardless of the viewing angle. In contrast, the specular component is viewing angle-dependent, producing shiny highlights that are crucial for achieving a photorealistic effect.", + "bbox": [ + 75, + 659, + 468, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lambertian Diffuse Reflectance. Lambertian reflectance, a standard model for diffuse reflection, describes a uniform light scatter irrespective of the viewing angle. This ensures a consistent appearance in brightness:", + "bbox": [ + 75, + 768, + 468, + 828 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {d} (\\mathbf {v}, \\mathbf {l}) = \\frac {\\sigma}{\\pi} [ \\text {c o n s t .} ] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 835, + 468, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\sigma$ is the albedo, indicating the intrinsic color and brightness of the surface.", + "bbox": [ + 75, + 869, + 468, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cook-Torrance Specular Reflectance. The Cook-Torrance model, based on microfacet theory, represents surfaces as a myriad of tiny, mirror-like facets. It incorporates a roughness parameter $\\alpha$ , which allows precise rendering of surface specular reflectance:", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {s} (\\mathbf {v}, \\mathbf {l}) = \\frac {D (\\mathbf {h} , \\alpha) G (\\mathbf {v} , \\mathbf {l} , \\alpha) F (\\mathbf {v} , \\mathbf {h} , f _ {0})}{4 \\langle \\mathbf {n} \\cdot \\mathbf {l} \\rangle \\langle \\mathbf {n} \\cdot \\mathbf {v} \\rangle} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 175, + 890, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this model, $D$ is the microfacet distribution function, describing the orientation of the microfacets relative to the half-vector $h$ , $G$ is the geometric attenuation factor, accounting for the shadowing and masking of microfacets, and $F$ is the Fresnel term, calculating the reflectance variation depending on the viewing angle, where $f_{0}$ is the surface Fresnel reflectivity at normal incidence. A lower $\\alpha$ value implies a smoother surface with sharper specular highlights, whereas a higher $\\alpha$ value indicates a rougher surface, resulting in more diffused reflections. By adjusting $\\alpha$ , the Cook-Torrance model can depict a range of specular reflections.", + "bbox": [ + 498, + 218, + 890, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Formation. Upon the base rendering equation, we include the diffuse and specular components of the BRDF and derive a unified formula:", + "bbox": [ + 498, + 402, + 890, + 446 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {o} (\\mathbf {v}) = \\int_ {\\Omega} \\left(f _ {d} (\\mathbf {v}, \\mathbf {l}) + f _ {s} (\\mathbf {v}, \\mathbf {l})\\right) E (\\mathbf {l}) \\langle \\mathbf {n} \\cdot \\mathbf {l} \\rangle d \\mathbf {l} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 455, + 890, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $E(1)$ denotes the incident environmental lighting. This formula represents the core principle that an image is a product of interplay between the BRDF and lighting. To further clarify this concept, we introduce a rendering function $R$ , succinctly modeling the process of image formation:", + "bbox": [ + 498, + 498, + 890, + 574 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI = R (\\underbrace {\\mathbf {n} , \\sigma , \\alpha , f _ {0}} _ {\\text {s u r f a c e a t t r i b u t e s l i g h t i n g}}, \\underbrace {E} _ {\\text {l i g h t i n g}}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 587, + 890, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is important to note that since the BRDF is a function of surface properties, as detailed in Eqn. 3 and 4, we can now clearly understand that image formation is essentially governed by the interaction of surface attributes and lighting.", + "bbox": [ + 498, + 628, + 890, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Problem Formulation", + "text_level": 1, + "bbox": [ + 500, + 698, + 700, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Relighting. Given the image formation model above, our goal is to manipulate the lighting of an existing image. This involves two main steps: inverse rendering and re-rendering under target illumination, both driven by neural networks. For a given source image $I_{\\mathrm{src}}$ and target illumination $E_{\\mathrm{tgt}}$ , the process is delineated as:", + "bbox": [ + 498, + 720, + 890, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {I n v e r s e} (\\mathbf {n}, \\sigma , \\alpha , f _ {0}, E _ {\\mathrm {s r c}}) = U (I _ {\\mathrm {s r c}})\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 824, + 885, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rendering with Target Light. $I_{\\mathrm{tgt}} = R(\\mathbf{n},\\sigma ,\\alpha ,f_0,E_{\\mathrm{tgt}})$", + "bbox": [ + 501, + 842, + 870, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During the inverse rendering step, the function $U$ unravels the intrinsic properties of $I_{\\mathrm{src}}$ . In the subsequent relighting", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25098", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3aef3794bddc2d79a3b52b533ce87d0ccbbbcaa0c5c7d51a523e59572face98c.jpg", + "image_caption": [ + "Figure 2. SwitchLight Architecture. The input source image is decomposed into normal map, lighting, diffuse and specular components. Given these intrinsic, images are re-rendered under target lighting. The architecture integrates the Cook-Torrance reflection model; the final output combines physically-based predictions with neural network enhancements for realistic portrait relighting." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 890, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "step, the derived intrinsic properties along with new illumination $E_{\\mathrm{tgt}}$ are employed by the rendering function $R$ to generate the target relit image $I_{\\mathrm{tgt}}$ .", + "bbox": [ + 75, + 328, + 470, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Architecture", + "text_level": 1, + "bbox": [ + 76, + 383, + 210, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our architecture systematically executes the two primary stages outlined in our problem formulation. The first stage involves extracting intrinsic properties from the source image $I_{\\mathrm{src}}$ . For this purpose, we employ a matting network [25, 27, 40] to accurately separate the foreground. This extracted image is then processed by our inverse rendering network $U$ , which infers normal, albedo, roughness, reflectivity, and source lighting. Subsequently, the second stage involves re-rendering the image under new target lighting conditions. To achieve this, the acquired intrinsics, along with the target lighting $E_{\\mathrm{tgt}}$ , are fed into our relighting network $R$ , producing the relit image $I_{\\mathrm{tgt}}$ .", + "bbox": [ + 75, + 406, + 468, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Normal Net. The network takes the source image $I_{\\mathrm{src}} \\in \\mathbb{R}^{\\mathrm{H} \\times \\mathrm{W} \\times 3}$ and generates a normal map $\\hat{\\mathbf{N}}$ . Each pixel in this map contains a unit normal vector $\\hat{\\mathbf{n}}$ , indicating the orientation of the corresponding surface point.", + "bbox": [ + 75, + 604, + 468, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Illum Net. The network infers the lighting conditions in the given image captured in an HDRI format. Specifically, it computes the convolved HDRIs:", + "bbox": [ + 75, + 684, + 468, + 729 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\mathrm {s r c}} ^ {p} \\left(\\mathbf {l} ^ {\\prime}\\right) = \\int_ {\\Omega} \\underbrace {E _ {\\mathrm {s r c}} (\\mathbf {l})} _ {\\text {H D R I}} \\underbrace {\\langle \\mathbf {l} ^ {\\prime} \\cdot \\mathbf {l} \\rangle^ {p}} _ {\\text {P h o n g l o b e}} d \\mathbf {l} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 739, + 468, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this equation, $E_{\\mathrm{src}} \\in \\mathbb{R}^{\\mathrm{H}_{\\mathrm{HDRI}} \\times \\mathrm{W}_{\\mathrm{HDRI}} \\times 3}$ is the original source HDRI map, with 1 indicating spherical directions in the HDRI space $\\mathbb{R}^{\\mathrm{H}_{\\mathrm{HDRI}} \\times \\mathrm{W}_{\\mathrm{HDRI}}}$ . The term $\\langle \\mathbf{l}' \\cdot \\mathbf{l} \\rangle^{p}$ represents the Phong reflectance lobe with shininess exponents $p \\in \\{1, 16, 32, 64\\}$ , which incorporates various specular terms. Consequently, it is expressed in a multi-dimensional tensor form as $\\mathbb{R}^{4 \\times \\mathrm{H}_{\\mathrm{HDRI}} \\times \\mathrm{W}_{\\mathrm{HDRI}} \\times \\mathrm{H}_{\\mathrm{HDRI}} \\times \\mathrm{W}_{\\mathrm{HDRI}}}$ . Finally, $E_{\\mathrm{src}}^{p} \\in$", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbb{R}^{4\\times \\mathrm{H_{cHDRI}}\\times \\mathrm{W_{cHDRI}}\\times 3}$ is the convolved HDRI. In this work, we set the resolution of HDRI and convolved HDRI at $32\\times 64$ and $64\\times 128$ , respectively, and we apply convolution on light source coordinates.", + "bbox": [ + 496, + 327, + 892, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The network employs a cross-attention mechanism at its core, where predefined Phong reflectance lobes serve as queries, and the original image acts as both keys and values. Within this setup, the convolved HDRI maps are synthesized by integrating image information into the Phong reflectance lobe representation. Specifically, our model utilizes bottleneck features from the Normal Net as a compact image representation. Our approach simplifies the complex task of HDRI reconstruction by instead focusing on estimating interactions with known surface reflective properties.", + "bbox": [ + 496, + 388, + 892, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffuse Net. Estimating albedo is challenging due to the ambiguities in surface color and material properties, further complicated by shadow effects. To address this, we prioritize the inference of source diffuse render, $I_{\\mathrm{src,diff}}$ :", + "bbox": [ + 496, + 556, + 890, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {s r c}, o _ {\\mathrm {d i f f}}} (\\mathbf {v}) = \\underbrace {\\frac {\\sigma}{\\pi}} _ {\\text {d i f f u s e B R D F}} \\underbrace {\\int_ {\\Omega} E _ {\\mathrm {s r c}} (\\mathbf {l}) \\left\\langle \\mathbf {n} \\cdot \\mathbf {l} \\right\\rangle d \\mathbf {l}} _ {\\text {d i f f u s e s h a d i n g}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 623, + 890, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our key insight is that the diffuse render closely resembles the original image, which simplifies the model learning process. It captures surface color after removing specular reflections, such as shine or gloss, contrasting with albedo that represents the true surface color unaffected by lighting and shadows. The network takes a source image $I_{\\mathrm{src}}$ , concatenated with its diffuse shading, to produce the diffuse render. As in Eqn. 7, the diffuse shading, $\\hat{E}_{\\mathrm{src}}^{1}(\\hat{\\mathbf{n}})$ , is derived using the predicted normals, $\\hat{\\mathbf{n}}$ , and the predicted lighting map, $\\hat{E}_{\\mathrm{src}}^{1}$ , with a Phong exponent of 1 for the diffuse term. The albedo map $\\hat{\\mathbf{A}}$ is then computed by dividing the predicted diffuse render by its diffuse shading:", + "bbox": [ + 496, + 680, + 892, + 861 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\hat {\\sigma}}{\\pi} = \\frac {\\hat {L} _ {\\mathrm {s r c} , o _ {\\mathrm {d i f f}}} (\\mathbf {v})}{\\hat {E} _ {\\mathrm {s r c}} ^ {1} (\\hat {\\mathbf {n}})} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 867, + 890, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25099", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6bb454551f13b8fc93664e680abe94d1c2f018ff65f717cb84723e3772b4fe3e.jpg", + "image_caption": [ + "Figure 3. Render Net Overview. Utilizing extracted image intrinsics, it employs the Cook-Torrance model for initial relighting and a neural network for enhanced refinement, producing high-fidelity relit images through a synergistic computational approach." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 472, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We have empirically validated that it significantly enhances albedo prediction across a range of real-world scenarios.", + "bbox": [ + 76, + 303, + 468, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specular Net. The network infers surface attributes associated with the Cook-Torrance specular elements, specifically, the roughness $\\alpha$ and Fresnel reflectivity $f_{0}$ . It uses a source image, predicted normal, and albedo maps as inputs.", + "bbox": [ + 75, + 352, + 468, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Render Net. The network utilizes extracted intrinsic surface attributes to produce the target relit images. It generates two types of relit images, as shown in Fig. 3. The first type adheres to the physically-based rendering (PBR) principles of the Cook-Torrance model. This involves computing diffuse and specular renders under the target illumination using Eqn. 3 and Eqn. 4. These renders are combined to form the PBR render, $\\hat{I}_{\\mathrm{tgt}}^{\\mathrm{PBR}}$ , as:", + "bbox": [ + 75, + 430, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {L} _ {\\mathrm {t g t}, o} ^ {\\mathrm {P B R}} (\\mathbf {v}) = \\hat {L} _ {\\mathrm {t g t}, o _ {\\text {d i f f}}} (\\mathbf {v}) + \\hat {L} _ {\\mathrm {t g t}, o _ {\\text {s p e c}}} (\\mathbf {v}) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 563, + 468, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The second type of relit image is the result of a neural network process. This enhances the PBR render, capturing finer details that the Cook-Torrance model might miss. It employs the albedo, along with the diffuse and specular renders from the Cook-Torrance model, to infer a more refined target relit image, termed the neural render, $\\hat{I}_{\\mathrm{tgt}}^{\\mathrm{Neural}}$ . The qualitative improvements achieved through this neural enhancement are illustrated in Fig. 4.", + "bbox": [ + 75, + 592, + 468, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Objectives", + "text_level": 1, + "bbox": [ + 76, + 722, + 192, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We supervise both intrinsic image attributes and relit images using their corresponding ground truths, obtained from the lightstage. We employ a combination of reconstruction, perceptual [23], adversarial [21], and specular [33] losses.", + "bbox": [ + 75, + 744, + 468, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reconstruction Loss $\\mathcal{L} = \\ell_1(\\mathbf{X},\\hat{\\mathbf{X}})$ . It measures the pixel-level differences between the ground truth $\\mathbf{X}$ and its prediction $\\hat{\\mathbf{X}}$ . This loss is applied across different attributes, including normal map $\\hat{\\mathbf{N}}$ , convolved source HDRI maps $\\hat{E}_{\\mathrm{src}}^p$ , source diffuse render $\\hat{I}_{\\mathrm{src,diff}}$ , albedo map $\\hat{\\mathbf{A}}$ , and both", + "bbox": [ + 76, + 824, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/38f6fd31e8f47999f2ffdde60969933648706529164eba58e1d6b671effee889.jpg", + "image_caption": [ + "Figure 4. Neural Render Enhancement. Using the Cook-Torrance model, diffuse and specular renders are computed, which are then composited into a physically-based rendering. Subsequently, a neural network enhances this PBR render, improving aspects such as brightness and specular details." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 890, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "types of target relit images $\\hat{I}_{\\mathrm{tgt}}^{\\mathrm{PBR}}$ and $\\hat{I}_{\\mathrm{tgt}}^{\\mathrm{Neural}}$ . The attributes like the roughness $\\alpha$ and Fresnel reflectivity $f_{0}$ are also implicitly learned when this loss applied to the PBR render.", + "bbox": [ + 496, + 446, + 890, + 493 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Perceptual Loss $\\mathcal{L}_{\\mathrm{vgg}} = \\ell_2(\\mathbf{VGG}(\\mathbf{X}),\\mathbf{VGG}(\\hat{\\mathbf{X}}))$ . It captures high-level feature differences based on a VGGnetwork feature comparison. We apply this loss to the source diffuse render, albedo, and target relit images.", + "bbox": [ + 496, + 510, + 890, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adversarial Loss $\\mathcal{L}_{\\mathrm{adv}} = \\mathrm{GAN}(\\mathbf{X},\\hat{\\mathbf{X}})$ . It promotes realism in the generated images by fooling a discriminator network. This loss is applied to the target relit images. We employ a PatchGAN architecture, with detailed specifications provided in the supplementary material.", + "bbox": [ + 496, + 588, + 890, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specular Loss $\\mathcal{L}_{\\mathrm{spec}} = \\ell_1(\\mathbf{X} \\odot \\hat{\\mathbf{S}}, \\hat{\\mathbf{X}} \\odot \\hat{\\mathbf{S}})$ . It enhances the specular highlights in the relit images. Specifically, we utilize the predicted specular render $\\hat{\\mathbf{S}} := \\hat{I}_{\\mathrm{tgt},\\mathrm{spec}}^{\\mathrm{PBR}}$ derived from the Cook-Torrance physical model, to weigh the $\\ell_1$ reconstruction loss. Here, $\\odot$ denotes the element-wise multiplication. This loss is applied to the neural render.", + "bbox": [ + 496, + 681, + 890, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Final Loss. The SwitchLight is trained in an end-to-end manner using the weighted sum of the above losses:", + "bbox": [ + 496, + 791, + 890, + 821 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {r e l i g h t}} = 1 0 \\cdot \\mathcal {L} _ {\\text {n o r m a l}} + 1 0 \\cdot \\mathcal {L} _ {\\text {s r c - H D R I}} + 0. 2 \\cdot \\mathcal {L} _ {\\text {s r c - d i f f}} \\\\ + 0. 2 \\cdot \\mathcal {L} _ {\\text {a l b e d o}} + 0. 2 \\cdot \\mathcal {L} _ {\\text {P B R}} + 0. 2 \\cdot \\mathcal {L} _ {\\text {N e u r a l}} \\tag {11} \\\\ + \\mathcal {L} _ {\\mathrm {v g g} _ {\\text {s r c . d i f f}}} + \\mathcal {L} _ {\\mathrm {v g g} _ {\\text {a l b e d o}}} + \\mathcal {L} _ {\\mathrm {v g g} _ {\\mathrm {P B R}}} + \\mathcal {L} _ {\\mathrm {v g g} _ {\\text {N e u r a l}}} \\\\ + \\mathcal {L} _ {\\mathrm {a d v} _ {\\mathrm {P B R}}} + \\mathcal {L} _ {\\mathrm {a d v} _ {\\mathrm {N e u r a l}}} + 0. 2 \\cdot \\mathcal {L} _ {\\mathrm {s p e c} _ {\\mathrm {N e u r a l}}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 829, + 890, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25100", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/032dec24f947d1a38a99291244c0abcfd6d0d46af8cdf75294c55d89ec87bf49.jpg", + "image_caption": [ + "Figure 5. Dynamic Masking Strategies. We have generalized the MAE masks to include overlapping patches of varying sizes, as well as outpainting and free-form masks." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 176, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ef7513b61de6a217ad8a36fdada2ab8dcd0f0604be62383edb9886b319259fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 89, + 272, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d4ef083b15cd653dd0d380e71b555f4c3e49705032698bca3c113325a261ae4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 89, + 369, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/51234a997d19a0a4b7a616e62ee849c52f28158a3af14f5ad85d0a9d599b427d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 90, + 465, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We empirically determined the weighting coefficients.", + "bbox": [ + 96, + 217, + 455, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Multi-Masked Autoencoder Pre-training", + "text_level": 1, + "bbox": [ + 76, + 247, + 441, + 265 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We introduce the Multi-Masked Autoencoder (MMAE), a self-supervised pre-training framework designed to enhance feature representations in relighting models. It aims to improve output quality without relying on additional, costly light stage data. Building upon the MAE framework [18], MMAE capitalizes on the inherent learning of crucial image features like structure, color, and texture, which are essential for relighting tasks. However, adapting MAE to our specific needs poses several non-trivial challenges. Firstly, MAE is primarily designed for vision transformers [14], while our focus is on a UNet, a convolution-based architecture. This convolutional structure, with its hierarchical nature and aggressive pooling, is known to simplify the MAE task, necessitating careful adaptation [49]. Further, the hyperparameters of MAE, particularly the fixed mask size and ratio, are also specific to vision transformers. These factors could introduce biases during training and hinder the model to recognize image features at various scales. Moreover, MAE relies solely on masked region reconstruction loss, limiting the model to understand the global coherence of the reconstructed region in relation to its visible context.", + "bbox": [ + 75, + 272, + 468, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address these challenges effectively, we have developed two key strategies within the MMAE framework:", + "bbox": [ + 76, + 590, + 468, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dynamic Masking. MMAE eliminates two key hyperparameters, mask size and ratio, by introducing a variety of mask types to generalize the MAE. These types, which include overlapping patches of various sizes, outpainting masks [45], and free-form masks [28] (see Fig.5), each contribute to the model's versatility. With the ability to handle challenging masked regions, MMAE achieves a more comprehensive understanding of image properties.", + "bbox": [ + 75, + 625, + 468, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generative Target. In addition to its structural advancements, MMAE incorporates a new loss function strategy. We have adopted perceptual [23] and adversarial losses [21], along with the original reconstruction loss. As a result, MMAE is equipped not only to reconstruct missing image parts but also to ensure synthesis capabilities and their seamless integration with the original context. In practice, the weights for the three losses are equally set.", + "bbox": [ + 75, + 744, + 468, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We pre-train the entire UNet architecture using MMAE, and, unlike MAE, we retain the decoder and fine-tune the", + "bbox": [ + 76, + 869, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "entire model on relighting ground truths.", + "bbox": [ + 500, + 90, + 769, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Data", + "text_level": 1, + "bbox": [ + 500, + 119, + 566, + 135 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We constructed the OLAT (One Light at a Time) dataset using a light stage [10, 48] equipped with 137 programmable LED lights and 7 frontal-viewing cameras. Our dataset comprises images of 287 subjects, with each subject being captured in up to 15 different poses, resulting in a total of 29,705 OLAT sequences. We sourced HDRI dataset from several publicly available archives. Specifically, we acquired 559 HDRI maps from Polyhaven, 76 from Noah Witchell, 364 from HDRMAPS, 129 from iHDRI, and 34 from eisklotz. In addition, we incorporated synthetic HDRIs created using the method proposed in [30]. During training, HDRIs are randomly selected with equal probability from either real-world or synthetic collections.", + "bbox": [ + 496, + 143, + 890, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We produced training pairs by projecting the sampled source and target lighting maps onto the reflectance fields of the OLAT images [10]. To derive the ground truth intrinsics, we applied the photometric stereo method [50] and obtained normal and albedo maps.", + "bbox": [ + 496, + 340, + 890, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 500, + 430, + 632, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section details our experimental results. We begin with a comparative evaluation of our method against state-of-the-art approaches using the OLAT dataset. We also employ images from the FFHQ-test [24] for user studies. For qualitative analysis, we utilize copyright-free portrait images from Pexels [1]. Additionally, we conduct ablation studies to validate the efficacy of our pre-training framework and architectural design choice. Subsequently, we detail the additional features and conclude by discussing its limitations. Our evaluation uses the OLAT test set, comprising 35 subjects and 11 lighting environments, ensuring no overlap with the train set.", + "bbox": [ + 496, + 455, + 890, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We employ several key metrics for evaluating the prediction accuracy; Mean Absolute Error (MAE), Mean Squared Error (MSE), Structural Similarity Index Measure (SSIM) and Learned Perceptual Image Patch Similarity (LPIPS). While these metrics offer valuable quantitative insights, they do not fully capture the subtleties of visual quality enhancement. Therefore, we emphasize the importance of qualitative evaluations to gain a comprehensive understanding of model performance.", + "bbox": [ + 496, + 640, + 890, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compared our approach with three state-of-the-art baselines: Single Image Portrait Relighting (SIPR) [44], which uses a single neural network for relighting; Total Relight (TR) [33], employing multiple neural networks that incorporate the Phong reflectance model; and Lumos [51], a TR adaptation for synthetic datasets. Due to the lack of publicly available code or model from these methods, we either integrated their techniques into our", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25101", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5bbeb8fdc73c2602473733c8e6b518d8c4c7c826f6a2099029acedb4ef2806eb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MAE ↓MSE ↓SSIM ↑LPIPS ↓
SIPR [44]0.17150.07480.84323.648
TR [33]0.16430.06580.84653.425
Ours0.10230.02750.90022.137
Ours (w. MMAE)0.09330.02350.90512.059
", + "bbox": [ + 89, + 88, + 457, + 184 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bb169db9c6e07b3884260a1ab25b3dadec745b1e40fe85da8a05812f093c004e.jpg", + "table_caption": [ + "Table 1. Quantitative Evaluation on the OLAT test set." + ], + "table_footnote": [], + "table_body": "
Lumos [51]TR [33]Ours
Consistent Lighting0.04780.18520.7671
Facial Details0.20220.26020.5376
Similar Identity0.17410.24400.5819
", + "bbox": [ + 88, + 210, + 459, + 290 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. User Study on the FFHQ test set.", + "bbox": [ + 145, + 300, + 398, + 314 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8cba26b02b2c8544c446226b80440753c45bb31480af2d73c51fe9345f89cff7.jpg", + "image_caption": [ + "Figure 6. Impact of Pre-training. The fine details such as specular highlights, skin tones, and shadows are notably improved." + ], + "image_footnote": [], + "bbox": [ + 80, + 319, + 465, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "framework or requested the original authors to process our inputs with their models and share the results.", + "bbox": [ + 76, + 590, + 468, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparisons. The results in Table. 1 show our method outperforming SIPR and TR baselines, demonstrating the significance of incorporating advanced rendering physics and reflectance models. The transition from SIPR to TR emphasizes the value of physics-based design, while the shift from TR to our approach underscores the importance of transitioning from the empirical Phong model to the more accurate Cook-Torrance model. Additionally, pretraining contributes to further enhancements, as evidenced by the improved image details, depicted in Fig 6.", + "bbox": [ + 75, + 619, + 468, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparisons. Our relighting method exhibits several key advantages over previous approaches, as showcased in Fig. 7. It effectively harmonizes light direction and softness, avoiding harsh highlights and inaccurate lighting that are commonly observed in other methods. A notable strength of our approach lies in its ability to capture high-frequency details like specular highlights and hard shadows. Additionally, as shown in the second row, it preserves fa", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b4143c5ca3289a0076e8b0113ea0f9cac2360741a92a669f01f36dc29f0b7ae.jpg", + "image_caption": [ + "Figure 7. Qualitative Comparison on the Pexels images [1]." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 890, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cial details and identity, ensuring high fidelity to the subject's original features and mitigating common distortions seen in previous approaches. Moreover, our approach excels in handling skin tones, producing natural and accurate results under various lighting conditions. This is clearly demonstrated in the fourth row, where our method contrasts sharply with the over-saturated or pale tones from previous methods. Finally, the nuanced treatment of hair is highlighted in the sixth row, where our approach maintains luster and detail, unlike the flattened effect typical in other methods. More qualitative results are available in our supplementary video demonstration.", + "bbox": [ + 496, + 439, + 890, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. We conducted a human subjective test to evaluate the visual quality of relighting results, summarized in Table. 2. In each test case, workers were presented with an input image and an environment map. They were asked to compare the relighting results from three methods—Ours, Lumos, and TR—based on three criteria: 1) consistency of lighting with the environment map, 2) preservation of facial details, and 3) maintenance of the original identity. To ensure unbiased evaluations, the order of the methods presented was randomized. To aid in understanding the concept of consistent lighting, relit balls were displayed alongside the images. The study included a total of 256 images, consisting of 32 portraits each illuminated with 8 different HDRIs. Each worker was tasked with selecting the best image for each specific criterion, randomly assessing 30 samples. A total of 47 workers participated in the study. The results indicate a strong preference for our results over the baseline methods across all evaluated metrics.", + "bbox": [ + 496, + 628, + 892, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25102", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/af3af130d1500456a1f39eead8507dbd26b9ec67d4ae3f8dfc567d19afa43333.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pre-trainingMethodMAE ↓MSE ↓SSIM ↑LPIPS ↓
MAE [18]0.09520.02420.90072.096
MMAE0.09330.02350.90512.059
DiffuseNetAlbedo0.10530.02950.89852.197
Diff Render0.10230.02750.90022.137
", + "bbox": [ + 83, + 88, + 460, + 184 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation Studies on the OLAT test set.", + "bbox": [ + 127, + 195, + 416, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/175df5cf85a16b03ce2ceaf0fbc83572d83762d5e469848d15a98a9a8d44c001.jpg", + "image_caption": [ + "Figure 8. Ablation on DiffuseNet. Our approach successfully infers the albedo on various surfaces (skin, teeth, and accessories)." + ], + "image_footnote": [], + "bbox": [ + 80, + 219, + 460, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Studies. We analyze our two major design choices in Table. 3: the MMAE pre-training framework and DiffuseNet. The MMAE, which integrates dynamic masking with generative objectives, outperforms MAE. This superiority is mainly due to the incorporation of challenging masks and global coherence objectives, enabling the model to learn richer features during pre-training. Furthermore, our method of predicting diffuse render demonstrates superiority over direct albedo prediction. Firstly, we see it simplifies the learning process, as predicting diffuse render is more closely related to the original image. Secondly, our approach effectively distinguishes between the influences of illumination (diffuse shading) and surface properties (diffuse render). This distinction is crucial for accurately modeling the intrinsic color of surfaces, as it enables independent and precise evaluation of each element (see Eqn. 9). In contrast, methods that predict albedo directly often struggle to differentiate between these factors, leading to significant inaccuracies in color constancy, as shown in Fig. 8.", + "bbox": [ + 76, + 573, + 468, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Applications. We present two applications using predicted intrinsics in Fig. 9. First, real-time PBR via Cook-Torrance", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bcae20feb2c5d47d383108bc8c144e7c5a968d4f0c475f664f06ab3158f89093.jpg", + "image_caption": [ + "(a) Real-Time PBR Rendering" + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 638, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/09ba99ded5c01df2b6056d56f86a4009f651521d91b532638eb6be62d3bd1203.jpg", + "image_caption": [ + "(b) Copy Light" + ], + "image_footnote": [], + "bbox": [ + 653, + 88, + 890, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9ab99f19c87c386009f828589ac64c646f84a386eceb0fe0fdddcaa4467461f7.jpg", + "image_caption": [ + "Figure 9. Applications. We showcase additional features of SwitchLight, powered by the diverse intrinsics features.", + "Figure 10. Limitations. The model faces challenges in removing strong shadows, misinterpreting reflective surfaces like sunglasses, and inaccurately predicting albedo for face paint." + ], + "image_footnote": [], + "bbox": [ + 501, + 214, + 890, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "components in Three.js graphics library. Second, switching the lighting environment between the source and reference images. Further details are in the supplementary video.", + "bbox": [ + 500, + 513, + 890, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. We identified a few failure cases in Fig. 10. First, our model struggles with neutralizing strong shadows, which leads to inaccurate facial geometry and residual shadow artifacts in both albedo and relit images. Incorporating shadow augmentation techniques [15, 53] during training could mitigate this issue. Second, the model incorrectly interprets reflective surfaces, such as sunglasses, as opaque objects in the normal image. This error prevents the model from properly removing reflective highlights in the albedo and relit images. Lastly, the model inaccurately predicts the albedo for face paint. Implementing a semantic mask [51] to distinguish different semantic regions separately from the skin could help resolving these issues.", + "bbox": [ + 500, + 566, + 890, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 768, + 617, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce SwitchLight, an architecture based on Cook-Torrance rendering physics, enhanced with a self-supervised pre-training framework. This co-designed approach significantly outperforms previous models. Our future plans include scaling the current model beyond images to encompass video and 3D data. We hope our proposal serve as a new foundational model for relighting tasks.", + "bbox": [ + 500, + 794, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25103", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Pexels. https://www.pexels.com.6,7", + "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2", + "[3] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2014. 2", + "[4] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 2", + "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2", + "[6] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22367-22377, 2023. 2", + "[7] Zhaoxi Chen and Ziwei Liu. Relighting4d: Neural relightable human from videos. In European Conference on Computer Vision, pages 606-623. Springer, 2022. 2", + "[8] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM Transactions on Graphics (ToG), 1(1):7-24, 1982. 2", + "[9] Paul Debevec. Rendering synthetic objects into real scenes: Bridging traditional and image-based graphics with global illumination and high dynamic range photography. In Acm siggraph 2008 classes, pages 1-10. 2008. 2", + "[10] Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. Acquiring the reflectance field of a human face. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 145-156, 2000. 2, 6", + "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2", + "[12] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422-1430, 2015. 2", + "[13] Julie Dorsey, James Arvo, and Donald Greenberg. Interactive design of complex time dependent lighting. IEEE Computer Graphics and Applications, 15(2):26-36, 1995. 2", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 6", + "[15] David Futschik, Kelvin Ritland, James Vecore, Sean Fanello, Sergio Orts-Escalano, Brian Curless, Daniel Sykora, and Ro" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "hit Pandey. Controllable light diffusion for portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8412-8421, 2023. 8", + "[16] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. 2", + "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2", + "[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 6, 8", + "[19] Andrew Hou, Ze Zhang, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Towards high fidelity face relighting with realistic shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14719-14728, 2021. 2", + "[20] Andrew Hou, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Face relighting with geometrically consistent shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4217-4226, 2022. 2", + "[21] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 5, 6", + "[22] Chaonan Ji, Tao Yu, Kaiwen Guo, Jingxin Liu, and Yebin Liu. Geometry-aware single-image full-body human relighting. In European Conference on Computer Vision, pages 388-405. Springer, 2022. 2", + "[23] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5, 6", + "[24] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6", + "[25] Taehun Kim, Kunhee Kim, Joonyeong Lee, Dongmin Cha, Jiho Lee, and Daijin Kim. Revisiting image pyramid structure for high resolution salient object detection. In Proceedings of the Asian Conference on Computer Vision, pages 108-124, 2022. 4", + "[26] Wenbo Li, Xin Lu, Shengju Qian, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer-based image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2", + "[27] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting." + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25104", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 4", + "[28] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European conference on computer vision (ECCV), pages 85-100, 2018. 6", + "[29] Yihao Liu, Jingwen He, Jinjin Gu, Xiangtao Kong, Yu Qiao, and Chao Dong. Degae: A new pretraining paradigm for low-level vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23292-23303, 2023. 2", + "[30] Yiqun Mei, He Zhang, Xuaner Zhang, Jianming Zhang, Zhixin Shu, Yilin Wang, Zijun Wei, Shi Yan, HyunJoon Jung, and Vishal M Patel. Lightpainter: Interactive portrait relighting with freehand scribble. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 195-205, 2023. 2, 6", + "[31] Thomas Nestmeyer, Jean-François Lalonde, Iain Matthews, and Andreas Lehrmann. Learning physics-guided face relighting under directional light. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5124-5133, 2020. 2", + "[32] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 2", + "[33] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 2, 5, 6, 7", + "[34] Namuk Park, Wonjae Kim, Byeongho Heo, Taekyung Kim, and Sangdoo Yun. What do self-supervised vision transformers learn? arXiv preprint arXiv:2305.00729, 2023. 2", + "[35] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 2", + "[36] Bui Tuong Phong. Illumination for computer generated pictures. In Seminal graphics: pioneering efforts that shaped the field, pages 95-101. 1998. 2", + "[37] Puntawat Ponglertnapakorn, Nontawat Tritrong, and Supasorn Suwajanakorn. Difareli: Diffusion face relighting. arXiv preprint arXiv:2304.09479, 2023. 2", + "[38] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2", + "[39] Soumyadip Sengupta, Angjoo Kanazawa, Carlos D Castillo, and David W Jacobs. Sfsnet: Learning shape, reflectance and illuminance of facesin the wild'. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6296-6305, 2018. 2", + "[40] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Pro-" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2291-2300, 2020. 4", + "[41] Soumyadip Sengupta, Brian Curless, Ira Kemelmacher-Shlizerman, and Steven M Seitz. A light stage on every desk. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2420-2429, 2021. 2", + "[42] YiChang Shih, Sylvain Paris, Connelly Barnes, William T Freeman, and Fredo Durand. Style transfer for headshot portraits. 2014. 2", + "[43] Zhixin Shu, Sunil Hadap, Eli Shechtman, Kalyan Sunkavalli, Sylvain Paris, and Dimitris Samaras. Portrait lighting transfer using a mass transport approach. ACM Transactions on Graphics (TOG), 36(4):1, 2017. 2", + "[44] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 2, 6, 7", + "[45] Piotr Teterwak, Aaron Sarna, Dilip Krishnan, Aaron Maschinot, David Belanger, Ce Liu, and William T Freeman. Boundless: Generative adversarial networks for image extension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10521-10530, 2019. 6", + "[46] Yifan Wang, Aleksander Holynski, Xiuming Zhang, and Xuaner Zhang. Sunstage: Portrait reconstruction and relighting using the sun as a light stage. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20792-20802, 2023. 2", + "[47] Zhibo Wang, Xin Yu, Ming Lu, Quan Wang, Chen Qian, and Feng Xu. Single image portrait relighting via explicit multiple reflectance channel modeling. ACM Transactions on Graphics (TOG), 39(6):1-13, 2020. 2", + "[48] Andreas Wenger, Andrew Gardner, Chris Tchou, Jonas Unger, Tim Hawkins, and Paul Debevec. Performance relighting and reflectance transformation with timemultiplexed illumination. ACM Transactions on Graphics (TOG), 24(3):756-764, 2005. 2, 6", + "[49] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133-16142, 2023. 2, 6", + "[50] Robert J Woodham. Photometric method for determining surface orientation from multiple images. Optical engineering, 19(1):139-144, 1980. 6", + "[51] Yu-Ying Yeh, Koki Nagano, Sameh Khamis, Jan Kautz, Ming-Yu Liu, and Ting-Chun Wang. Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. ACM Transactions on Graphics (TOG), 41(6): 1-21, 2022. 2, 6, 7, 8", + "[52] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 649-666. Springer, 2016. 2" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25105", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Xuaner Zhang, Jonathan T Barron, Yun-Ta Tsai, Rohit Pandey, Xiuming Zhang, Ren Ng, and David E Jacobs. Portrait shadow manipulation. ACM Transactions on Graphics (TOG), 39(4):78-1, 2020. 8", + "[54] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7194-7202, 2019. 2", + "[55] Taotao Zhou, Kai He, Di Wu, Teng Xu, Qixuan Zhang, Kuixiang Shao, Wenzheng Chen, Lan Xu, and Jingyi Yu. Relightable neural human assets from multi-view gradient illuminations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4315-4327, 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 287 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25106", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/ce409cac-fd5a-4614-9e6d-87e790ca611e_model.json b/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/ce409cac-fd5a-4614-9e6d-87e790ca611e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3bb20958b6438c2bd182f217adaf443a17fb5ce4 --- /dev/null +++ b/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/ce409cac-fd5a-4614-9e6d-87e790ca611e_model.json @@ -0,0 +1,2367 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.131, + 0.784, + 0.177 + ], + "angle": 0, + "content": "SwitchLight: Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.203, + 0.862, + 0.224 + ], + "angle": 0, + "content": "Hoon Kim1 Minje Jang1 Wonjun Yoon1 Jisoo Lee1 Donghyun Na1 Sanghyun Woo2" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.238, + 0.65, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)Beeble AI \\(^{2}\\)New York University" + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.289, + 0.203, + 0.304 + ], + "angle": 0, + "content": "Input RGB" + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.289, + 0.346, + 0.303 + ], + "angle": 0, + "content": "Normal" + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.289, + 0.513, + 0.303 + ], + "angle": 0, + "content": "Albedo" + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.289, + 0.687, + 0.304 + ], + "angle": 0, + "content": "Roughness" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.289, + 0.85, + 0.304 + ], + "angle": 0, + "content": "Reflectivity" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.305, + 0.891, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.657, + 0.894, + 0.686 + ], + "angle": 0, + "content": "Figure 1. Be Anywhere at Any Time. SwitchLight processes a human portrait by decomposing it into detailed intrinsic components, and re-renders the image under a designated target illumination, ensuring a seamless composition of the subject into any new environment." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.698, + 0.314, + 0.712 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.471, + 0.867 + ], + "angle": 0, + "content": "We introduce a co-designed approach for human portrait relighting that combines a physics-guided architecture with a pre-training framework. Drawing on the Cook-Torrance reflectance model, we have meticulously configured the architecture design to precisely simulate light-surface interactions. Furthermore, to overcome the limitation of scarce high-quality lightstage data, we have developed a self-supervised pre-training strategy. This novel combination of accurate physical modeling and expanded training dataset establishes a new benchmark in relighting realism." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.695, + 0.631, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.885 + ], + "angle": 0, + "content": "Relighting is more than an aesthetic tool; it unlocks infinite narrative possibilities and enables seamless integration of subjects into diverse environments (see Fig. 1). This advancement resonates with our innate desire to transcend the physical constraints of space and time, while also providing tangible solutions to practical challenges in digital content creation. It is particularly transformative in virtual (VR) and augmented reality (AR) applications, where relighting facilitates the real-time adaptation of lighting, ensuring that users and digital elements coexist naturally within any environment, offering a next level of telepresence." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.891, + 0.902 + ], + "angle": 0, + "content": "In this work, we focus on human portrait relighting." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.334, + 0.888 + ], + "angle": 0, + "content": "* All authors contributed equally to this work" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.22, + 0.9 + ], + "angle": 0, + "content": "\\*https://www.beeble.ai" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.334, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25096" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.227 + ], + "angle": 0, + "content": "While the relighting task fundamentally demands an in-depth understanding of geometry, material properties, and illumination, the challenge is more compounded when addressing human subjects, due to the unique characteristics of skin surfaces as well as the diverse textures and reflectance properties of a wide array of clothing, hairstyles, and accessories. These elements interact in complex ways, necessitating advanced algorithms capable of simulating the subtle interplay of light with these varied surfaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.236, + 0.471, + 0.478 + ], + "angle": 0, + "content": "Currently, the most promising approach involves the use of deep neural networks trained on pairs of high-quality relict portrait images and their corresponding intrinsic attributes, which are sourced from a light stage setup [10]. Initial efforts approached the relighting process as a 'black box' [44, 47], without delving into the underlying mechanisms. Later advancements adopted a physics-guided model design, incorporating the explicit modeling of image intrinsics and image formation physics [31]. Pandey et al. [33] proposed the Total Relight (TR) architecture, also physics-guided, which decomposes an input image into surface normals and albedo maps, and performs relighting based on the Phong specular reflectance model. The TR architecture has become foundational model for image relighting, with most recent and advanced architectures building upon its principle [22, 30, 51]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.47, + 0.773 + ], + "angle": 0, + "content": "Following the physics-guided approach, our contribution lies in a co-design of architecture with a self-supervised pre-training framework. First, our architecture evolves towards a more accurate physical model by integrating the Cook-Torrance specular reflectance model [8], representing a notable advancement from the empirical Phong specular model [36] employed in the Total Relight architecture. The Cook-Torrance model adeptly simulates light interactions with surface microfacets, accounting for spatially varying roughness and reflectivity. Secondly, our pretraining framework scales the learning process beyond the typically hard-to-obtain lightstage data. By revisiting the masked autoencoder (MAE) framework [18], we adept it for the task of relighting. These modifications are crafted to address the unique challenges posed by this task, enabling our model to learn from unlabelled data and refine its ability to produce realistic relit portraits during fine-tuning. To the best of our knowledge, this is the first time applying self-supervised pre-training specifically to the relighting task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To summarize, our contribution is twofold. Firstly, by enhancing the physical reflectance model, we have introduced a new level of realism in the output. Secondly, by adopting self-supervised learning, we have expanded the scale of the training data and enhanced the expression of lighting in diverse real-world scenarios. Collectively, these advancements have led SwitchLight framework to achieve a new state-of-the-art in human portrait relighting." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Human Portrait Relighting is an ill-posed problem due to its under-constrained nature. To tackle this, earlier methods incorporated 3D facial priors [43], exploited image intrinsics [3, 39], or framed the task as a style transfer [42]. Light stage techniques [48] offer a more powerful solution by recording subject's reflectance fields under varying lighting conditions [10, 13], though they are labor-intensive and require specialized equipment. A promising alternative has emerged with deep learning, utilizing neural networks trained on light stage data. Sun et al. [44] pioneered this approach, but their method had limitations in representing non-Lambertian effects. This was improved upon by Nestmeyer et al. [31], who integrated rendering physics into network design, albeit limited to directional light. Building upon this, Pandey et al. [33] incorporated the Phong reflection model and a high dynamic range (HDR) lighting map [9] into their network, enabling a more accurate representation of global illumination. Simultaneously, efforts have been made to explore portrait relighting without light stage data [19, 20, 41, 46, 54]. Moreover, introduction of NeRF [7] and diffusion-based [37] models has opened new avenues in the field. However, networks trained with light-stage data maintain superior accuracy and realism, thanks to physics-based composited relight image training pairs and precise ground truth image intrinsics [55]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.567 + ], + "angle": 0, + "content": "Our work furthers this domain by integrating the Cook-Torrance model into our network design, shifting from the empirical Phong model to a more physics-based approach, thereby enhancing the realism and detail in relit images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.775 + ], + "angle": 0, + "content": "Self-supervised Pre-training has become a standard training scheme in the development of large language models like BERT [11] and GPT [38], and is increasingly influential in vision models, aiming to replicate the 'BERT moment'. This approach typically involves pre-training on extensive unlabeled data, followed by fine-tuning on specific tasks. While early efforts in vision models focused on simple pretext tasks [12, 16, 32, 35, 52], the field has evolved through stages like contrastive learning [5, 17] and masked image modeling [2, 18, 49]. However, the primary focus has remained on visual recognition, with less attention to other domains. Exceptions include low-level image processing tasks [4, 6, 26, 29] using the vision transformer [14]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Our research takes a different route, focusing on human portrait relighting—a complex challenge of manipulating illumination in the image. This direction is crucial because acquiring accurate ground truth data, especially from light stage, is both expensive and difficult. We modify the MAE framework [18], previously successful in robust image representation learning and developing locality biases [34], to suit the unique requirements of effective relighting." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "25097" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.206, + 0.108 + ], + "angle": 0, + "content": "3. SwitchLight" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.206 + ], + "angle": 0, + "content": "We introduce SwitchLight, a state-of-the-art framework for human portrait relighting, with its architectural overview presented in Fig. 2. We first provide foundational concepts in Sec. 3.1, and define the problem in Sec. 3.2. This is followed by detailing the architecture in Sec. 3.3, and lastly, we describe the loss functions used in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.215, + 0.218, + 0.229 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.238, + 0.469, + 0.314 + ], + "angle": 0, + "content": "In this section, vectors \\(\\mathbf{n}\\), \\(\\mathbf{v}\\), \\(\\mathbf{l}\\), and \\(\\mathbf{h}\\) are denoted as unit vectors. Specifically, \\(\\mathbf{n}\\) represents the surface normal, \\(\\mathbf{v}\\) is the view direction, \\(\\mathbf{l}\\) is the incident light direction, and \\(\\mathbf{h}\\) is the half-vector computed from \\(\\mathbf{l}\\) and \\(\\mathbf{v}\\). The dot product is clamped between [0..1], indicated by \\(\\langle \\cdot \\rangle\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.469, + 0.392 + ], + "angle": 0, + "content": "Image Rendering. The primary goal of image rendering is to create a visual representation that accurately simulates the interactions between light and surfaces. These complex interactions are encapsulated by the rendering equation:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.401, + 0.469, + 0.432 + ], + "angle": 0, + "content": "\\[\nL _ {o} (\\mathbf {v}) = \\int_ {\\Omega} f (\\mathbf {v}, \\mathbf {l}) L _ {i} (\\mathbf {l}) \\langle \\mathbf {n} \\cdot \\mathbf {l} \\rangle d \\mathbf {l} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.469, + 0.546 + ], + "angle": 0, + "content": "where \\(L_{o}(\\mathbf{v})\\) denotes the radiance, or the light intensity perceived by the observer in direction \\(\\mathbf{v}\\). It is the cumulative result of incident lights \\(L_{i}(\\mathbf{l})\\) from all possible directions over the hemisphere, \\(\\Omega\\), centered around the surface normal, denoted as \\(\\mathbf{n}\\). Central to this equation lies the Bidirectional Reflectance Distribution Function (BRDF), denoted as \\(f(\\mathbf{v},\\mathbf{l})\\), describing the surface's reflection characteristics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.469, + 0.625 + ], + "angle": 0, + "content": "BRDF Composition. The BRDF, represented by \\( f(\\mathbf{v},\\mathbf{l}) \\), describes how light is reflected at an opaque surface. It is composed of two major components: diffuse reflection (\\( f_{d} \\)) and specular reflection (\\( f_{s} \\)):" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.634, + 0.469, + 0.651 + ], + "angle": 0, + "content": "\\[\nf (\\mathbf {v}, \\mathbf {l}) = f _ {d} (\\mathbf {v}, \\mathbf {l}) + f _ {s} (\\mathbf {v}, \\mathbf {l}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.75 + ], + "angle": 0, + "content": "A surface intrinsically exhibits both diffuse and specular reflections. The diffuse component uniformly scatters light, ensuring consistent illumination regardless of the viewing angle. In contrast, the specular component is viewing angle-dependent, producing shiny highlights that are crucial for achieving a photorealistic effect." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.769, + 0.469, + 0.829 + ], + "angle": 0, + "content": "Lambertian Diffuse Reflectance. Lambertian reflectance, a standard model for diffuse reflection, describes a uniform light scatter irrespective of the viewing angle. This ensures a consistent appearance in brightness:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.836, + 0.469, + 0.864 + ], + "angle": 0, + "content": "\\[\nf _ {d} (\\mathbf {v}, \\mathbf {l}) = \\frac {\\sigma}{\\pi} [ \\text {c o n s t .} ] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Here, \\(\\sigma\\) is the albedo, indicating the intrinsic color and brightness of the surface." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "Cook-Torrance Specular Reflectance. The Cook-Torrance model, based on microfacet theory, represents surfaces as a myriad of tiny, mirror-like facets. It incorporates a roughness parameter \\(\\alpha\\), which allows precise rendering of surface specular reflectance:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.176, + 0.892, + 0.211 + ], + "angle": 0, + "content": "\\[\nf _ {s} (\\mathbf {v}, \\mathbf {l}) = \\frac {D (\\mathbf {h} , \\alpha) G (\\mathbf {v} , \\mathbf {l} , \\alpha) F (\\mathbf {v} , \\mathbf {h} , f _ {0})}{4 \\langle \\mathbf {n} \\cdot \\mathbf {l} \\rangle \\langle \\mathbf {n} \\cdot \\mathbf {v} \\rangle} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.219, + 0.892, + 0.386 + ], + "angle": 0, + "content": "In this model, \\(D\\) is the microfacet distribution function, describing the orientation of the microfacets relative to the half-vector \\(h\\), \\(G\\) is the geometric attenuation factor, accounting for the shadowing and masking of microfacets, and \\(F\\) is the Fresnel term, calculating the reflectance variation depending on the viewing angle, where \\(f_{0}\\) is the surface Fresnel reflectivity at normal incidence. A lower \\(\\alpha\\) value implies a smoother surface with sharper specular highlights, whereas a higher \\(\\alpha\\) value indicates a rougher surface, resulting in more diffused reflections. By adjusting \\(\\alpha\\), the Cook-Torrance model can depict a range of specular reflections." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.404, + 0.892, + 0.448 + ], + "angle": 0, + "content": "Image Formation. Upon the base rendering equation, we include the diffuse and specular components of the BRDF and derive a unified formula:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.457, + 0.892, + 0.489 + ], + "angle": 0, + "content": "\\[\nL _ {o} (\\mathbf {v}) = \\int_ {\\Omega} \\left(f _ {d} (\\mathbf {v}, \\mathbf {l}) + f _ {s} (\\mathbf {v}, \\mathbf {l})\\right) E (\\mathbf {l}) \\langle \\mathbf {n} \\cdot \\mathbf {l} \\rangle d \\mathbf {l} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.499, + 0.892, + 0.575 + ], + "angle": 0, + "content": "where \\( E(1) \\) denotes the incident environmental lighting. This formula represents the core principle that an image is a product of interplay between the BRDF and lighting. To further clarify this concept, we introduce a rendering function \\( R \\), succinctly modeling the process of image formation:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.588, + 0.891, + 0.624 + ], + "angle": 0, + "content": "\\[\nI = R (\\underbrace {\\mathbf {n} , \\sigma , \\alpha , f _ {0}} _ {\\text {s u r f a c e a t t r i b u t e s l i g h t i n g}}, \\underbrace {E} _ {\\text {l i g h t i n g}}) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.63, + 0.892, + 0.69 + ], + "angle": 0, + "content": "It is important to note that since the BRDF is a function of surface properties, as detailed in Eqn. 3 and 4, we can now clearly understand that image formation is essentially governed by the interaction of surface attributes and lighting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.699, + 0.702, + 0.714 + ], + "angle": 0, + "content": "3.2. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.722, + 0.892, + 0.813 + ], + "angle": 0, + "content": "Image Relighting. Given the image formation model above, our goal is to manipulate the lighting of an existing image. This involves two main steps: inverse rendering and re-rendering under target illumination, both driven by neural networks. For a given source image \\( I_{\\mathrm{src}} \\) and target illumination \\( E_{\\mathrm{tgt}} \\), the process is delineated as:" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.825, + 0.886, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\text {I n v e r s e} (\\mathbf {n}, \\sigma , \\alpha , f _ {0}, E _ {\\mathrm {s r c}}) = U (I _ {\\mathrm {s r c}})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.843, + 0.871, + 0.86 + ], + "angle": 0, + "content": "Rendering with Target Light. \\(I_{\\mathrm{tgt}} = R(\\mathbf{n},\\sigma ,\\alpha ,f_0,E_{\\mathrm{tgt}})\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "During the inverse rendering step, the function \\(U\\) unravels the intrinsic properties of \\(I_{\\mathrm{src}}\\). In the subsequent relighting" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25098" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.891, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.268, + 0.893, + 0.311 + ], + "angle": 0, + "content": "Figure 2. SwitchLight Architecture. The input source image is decomposed into normal map, lighting, diffuse and specular components. Given these intrinsic, images are re-rendered under target lighting. The architecture integrates the Cook-Torrance reflection model; the final output combines physically-based predictions with neural network enhancements for realistic portrait relighting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.329, + 0.471, + 0.377 + ], + "angle": 0, + "content": "step, the derived intrinsic properties along with new illumination \\( E_{\\mathrm{tgt}} \\) are employed by the rendering function \\( R \\) to generate the target relit image \\( I_{\\mathrm{tgt}} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.384, + 0.212, + 0.398 + ], + "angle": 0, + "content": "3.3. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.47, + 0.59 + ], + "angle": 0, + "content": "Our architecture systematically executes the two primary stages outlined in our problem formulation. The first stage involves extracting intrinsic properties from the source image \\( I_{\\mathrm{src}} \\). For this purpose, we employ a matting network [25, 27, 40] to accurately separate the foreground. This extracted image is then processed by our inverse rendering network \\( U \\), which infers normal, albedo, roughness, reflectivity, and source lighting. Subsequently, the second stage involves re-rendering the image under new target lighting conditions. To achieve this, the acquired intrinsics, along with the target lighting \\( E_{\\mathrm{tgt}} \\), are fed into our relighting network \\( R \\), producing the relit image \\( I_{\\mathrm{tgt}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.47, + 0.669 + ], + "angle": 0, + "content": "Normal Net. The network takes the source image \\( I_{\\mathrm{src}} \\in \\mathbb{R}^{\\mathrm{H} \\times \\mathrm{W} \\times 3} \\) and generates a normal map \\( \\hat{\\mathbf{N}} \\). Each pixel in this map contains a unit normal vector \\( \\hat{\\mathbf{n}} \\), indicating the orientation of the corresponding surface point." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.47, + 0.731 + ], + "angle": 0, + "content": "Illum Net. The network infers the lighting conditions in the given image captured in an HDRI format. Specifically, it computes the convolved HDRIs:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.74, + 0.47, + 0.784 + ], + "angle": 0, + "content": "\\[\nE _ {\\mathrm {s r c}} ^ {p} \\left(\\mathbf {l} ^ {\\prime}\\right) = \\int_ {\\Omega} \\underbrace {E _ {\\mathrm {s r c}} (\\mathbf {l})} _ {\\text {H D R I}} \\underbrace {\\langle \\mathbf {l} ^ {\\prime} \\cdot \\mathbf {l} \\rangle^ {p}} _ {\\text {P h o n g l o b e}} d \\mathbf {l} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this equation, \\( E_{\\mathrm{src}} \\in \\mathbb{R}^{\\mathrm{H}_{\\mathrm{HDRI}} \\times \\mathrm{W}_{\\mathrm{HDRI}} \\times 3} \\) is the original source HDRI map, with 1 indicating spherical directions in the HDRI space \\( \\mathbb{R}^{\\mathrm{H}_{\\mathrm{HDRI}} \\times \\mathrm{W}_{\\mathrm{HDRI}}} \\). The term \\( \\langle \\mathbf{l}' \\cdot \\mathbf{l} \\rangle^{p} \\) represents the Phong reflectance lobe with shininess exponents \\( p \\in \\{1, 16, 32, 64\\} \\), which incorporates various specular terms. Consequently, it is expressed in a multi-dimensional tensor form as \\( \\mathbb{R}^{4 \\times \\mathrm{H}_{\\mathrm{HDRI}} \\times \\mathrm{W}_{\\mathrm{HDRI}} \\times \\mathrm{H}_{\\mathrm{HDRI}} \\times \\mathrm{W}_{\\mathrm{HDRI}}} \\). Finally, \\( E_{\\mathrm{src}}^{p} \\in \\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.893, + 0.389 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^{4\\times \\mathrm{H_{cHDRI}}\\times \\mathrm{W_{cHDRI}}\\times 3}\\) is the convolved HDRI. In this work, we set the resolution of HDRI and convolved HDRI at \\(32\\times 64\\) and \\(64\\times 128\\), respectively, and we apply convolution on light source coordinates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.893, + 0.542 + ], + "angle": 0, + "content": "The network employs a cross-attention mechanism at its core, where predefined Phong reflectance lobes serve as queries, and the original image acts as both keys and values. Within this setup, the convolved HDRI maps are synthesized by integrating image information into the Phong reflectance lobe representation. Specifically, our model utilizes bottleneck features from the Normal Net as a compact image representation. Our approach simplifies the complex task of HDRI reconstruction by instead focusing on estimating interactions with known surface reflective properties." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.892, + 0.62 + ], + "angle": 0, + "content": "Diffuse Net. Estimating albedo is challenging due to the ambiguities in surface color and material properties, further complicated by shadow effects. To address this, we prioritize the inference of source diffuse render, \\( I_{\\mathrm{src,diff}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.624, + 0.892, + 0.676 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {s r c}, o _ {\\mathrm {d i f f}}} (\\mathbf {v}) = \\underbrace {\\frac {\\sigma}{\\pi}} _ {\\text {d i f f u s e B R D F}} \\underbrace {\\int_ {\\Omega} E _ {\\mathrm {s r c}} (\\mathbf {l}) \\left\\langle \\mathbf {n} \\cdot \\mathbf {l} \\right\\rangle d \\mathbf {l}} _ {\\text {d i f f u s e s h a d i n g}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.681, + 0.893, + 0.862 + ], + "angle": 0, + "content": "Our key insight is that the diffuse render closely resembles the original image, which simplifies the model learning process. It captures surface color after removing specular reflections, such as shine or gloss, contrasting with albedo that represents the true surface color unaffected by lighting and shadows. The network takes a source image \\( I_{\\mathrm{src}} \\), concatenated with its diffuse shading, to produce the diffuse render. As in Eqn. 7, the diffuse shading, \\( \\hat{E}_{\\mathrm{src}}^{1}(\\hat{\\mathbf{n}}) \\), is derived using the predicted normals, \\( \\hat{\\mathbf{n}} \\), and the predicted lighting map, \\( \\hat{E}_{\\mathrm{src}}^{1} \\), with a Phong exponent of 1 for the diffuse term. The albedo map \\( \\hat{\\mathbf{A}} \\) is then computed by dividing the predicted diffuse render by its diffuse shading:" + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.868, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\frac {\\hat {\\sigma}}{\\pi} = \\frac {\\hat {L} _ {\\mathrm {s r c} , o _ {\\mathrm {d i f f}}} (\\mathbf {v})}{\\hat {E} _ {\\mathrm {s r c}} ^ {1} (\\hat {\\mathbf {n}})} \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25099" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.473, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.233, + 0.47, + 0.289 + ], + "angle": 0, + "content": "Figure 3. Render Net Overview. Utilizing extracted image intrinsics, it employs the Cook-Torrance model for initial relighting and a neural network for enhanced refinement, producing high-fidelity relit images through a synergistic computational approach." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.304, + 0.469, + 0.334 + ], + "angle": 0, + "content": "We have empirically validated that it significantly enhances albedo prediction across a range of real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.469, + 0.413 + ], + "angle": 0, + "content": "Specular Net. The network infers surface attributes associated with the Cook-Torrance specular elements, specifically, the roughness \\(\\alpha\\) and Fresnel reflectivity \\(f_{0}\\). It uses a source image, predicted normal, and albedo maps as inputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.469, + 0.553 + ], + "angle": 0, + "content": "Render Net. The network utilizes extracted intrinsic surface attributes to produce the target relit images. It generates two types of relit images, as shown in Fig. 3. The first type adheres to the physically-based rendering (PBR) principles of the Cook-Torrance model. This involves computing diffuse and specular renders under the target illumination using Eqn. 3 and Eqn. 4. These renders are combined to form the PBR render, \\(\\hat{I}_{\\mathrm{tgt}}^{\\mathrm{PBR}}\\), as:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.564, + 0.469, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\hat {L} _ {\\mathrm {t g t}, o} ^ {\\mathrm {P B R}} (\\mathbf {v}) = \\hat {L} _ {\\mathrm {t g t}, o _ {\\text {d i f f}}} (\\mathbf {v}) + \\hat {L} _ {\\mathrm {t g t}, o _ {\\text {s p e c}}} (\\mathbf {v}) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.469, + 0.714 + ], + "angle": 0, + "content": "The second type of relit image is the result of a neural network process. This enhances the PBR render, capturing finer details that the Cook-Torrance model might miss. It employs the albedo, along with the diffuse and specular renders from the Cook-Torrance model, to infer a more refined target relit image, termed the neural render, \\(\\hat{I}_{\\mathrm{tgt}}^{\\mathrm{Neural}}\\). The qualitative improvements achieved through this neural enhancement are illustrated in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.723, + 0.194, + 0.739 + ], + "angle": 0, + "content": "3.4. Objectives" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.469, + 0.807 + ], + "angle": 0, + "content": "We supervise both intrinsic image attributes and relit images using their corresponding ground truths, obtained from the lightstage. We employ a combination of reconstruction, perceptual [23], adversarial [21], and specular [33] losses." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Reconstruction Loss \\(\\mathcal{L} = \\ell_1(\\mathbf{X},\\hat{\\mathbf{X}})\\). It measures the pixel-level differences between the ground truth \\(\\mathbf{X}\\) and its prediction \\(\\hat{\\mathbf{X}}\\). This loss is applied across different attributes, including normal map \\(\\hat{\\mathbf{N}}\\), convolved source HDRI maps \\(\\hat{E}_{\\mathrm{src}}^p\\), source diffuse render \\(\\hat{I}_{\\mathrm{src,diff}}\\), albedo map \\(\\hat{\\mathbf{A}}\\), and both" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.892, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.435 + ], + "angle": 0, + "content": "Figure 4. Neural Render Enhancement. Using the Cook-Torrance model, diffuse and specular renders are computed, which are then composited into a physically-based rendering. Subsequently, a neural network enhances this PBR render, improving aspects such as brightness and specular details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.494 + ], + "angle": 0, + "content": "types of target relit images \\(\\hat{I}_{\\mathrm{tgt}}^{\\mathrm{PBR}}\\) and \\(\\hat{I}_{\\mathrm{tgt}}^{\\mathrm{Neural}}\\). The attributes like the roughness \\(\\alpha\\) and Fresnel reflectivity \\(f_{0}\\) are also implicitly learned when this loss applied to the PBR render." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.892, + 0.572 + ], + "angle": 0, + "content": "Perceptual Loss \\(\\mathcal{L}_{\\mathrm{vgg}} = \\ell_2(\\mathbf{VGG}(\\mathbf{X}),\\mathbf{VGG}(\\hat{\\mathbf{X}}))\\) . It captures high-level feature differences based on a VGGnetwork feature comparison. We apply this loss to the source diffuse render, albedo, and target relit images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.892, + 0.665 + ], + "angle": 0, + "content": "Adversarial Loss \\(\\mathcal{L}_{\\mathrm{adv}} = \\mathrm{GAN}(\\mathbf{X},\\hat{\\mathbf{X}})\\). It promotes realism in the generated images by fooling a discriminator network. This loss is applied to the target relit images. We employ a PatchGAN architecture, with detailed specifications provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.892, + 0.774 + ], + "angle": 0, + "content": "Specular Loss \\(\\mathcal{L}_{\\mathrm{spec}} = \\ell_1(\\mathbf{X} \\odot \\hat{\\mathbf{S}}, \\hat{\\mathbf{X}} \\odot \\hat{\\mathbf{S}})\\). It enhances the specular highlights in the relit images. Specifically, we utilize the predicted specular render \\(\\hat{\\mathbf{S}} := \\hat{I}_{\\mathrm{tgt},\\mathrm{spec}}^{\\mathrm{PBR}}\\) derived from the Cook-Torrance physical model, to weigh the \\(\\ell_1\\) reconstruction loss. Here, \\(\\odot\\) denotes the element-wise multiplication. This loss is applied to the neural render." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.792, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Final Loss. The SwitchLight is trained in an end-to-end manner using the weighted sum of the above losses:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.83, + 0.891, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {r e l i g h t}} = 1 0 \\cdot \\mathcal {L} _ {\\text {n o r m a l}} + 1 0 \\cdot \\mathcal {L} _ {\\text {s r c - H D R I}} + 0. 2 \\cdot \\mathcal {L} _ {\\text {s r c - d i f f}} \\\\ + 0. 2 \\cdot \\mathcal {L} _ {\\text {a l b e d o}} + 0. 2 \\cdot \\mathcal {L} _ {\\text {P B R}} + 0. 2 \\cdot \\mathcal {L} _ {\\text {N e u r a l}} \\tag {11} \\\\ + \\mathcal {L} _ {\\mathrm {v g g} _ {\\text {s r c . d i f f}}} + \\mathcal {L} _ {\\mathrm {v g g} _ {\\text {a l b e d o}}} + \\mathcal {L} _ {\\mathrm {v g g} _ {\\mathrm {P B R}}} + \\mathcal {L} _ {\\mathrm {v g g} _ {\\text {N e u r a l}}} \\\\ + \\mathcal {L} _ {\\mathrm {a d v} _ {\\mathrm {P B R}}} + \\mathcal {L} _ {\\mathrm {a d v} _ {\\mathrm {N e u r a l}}} + 0. 2 \\cdot \\mathcal {L} _ {\\mathrm {s p e c} _ {\\mathrm {N e u r a l}}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25100" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.177, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.09, + 0.273, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.09, + 0.37, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.091, + 0.467, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.17, + 0.47, + 0.212 + ], + "angle": 0, + "content": "Figure 5. Dynamic Masking Strategies. We have generalized the MAE masks to include overlapping patches of varying sizes, as well as outpainting and free-form masks." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.218, + 0.456, + 0.233 + ], + "angle": 0, + "content": "We empirically determined the weighting coefficients." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.248, + 0.442, + 0.266 + ], + "angle": 0, + "content": "4. Multi-Masked Autoencoder Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.47, + 0.59 + ], + "angle": 0, + "content": "We introduce the Multi-Masked Autoencoder (MMAE), a self-supervised pre-training framework designed to enhance feature representations in relighting models. It aims to improve output quality without relying on additional, costly light stage data. Building upon the MAE framework [18], MMAE capitalizes on the inherent learning of crucial image features like structure, color, and texture, which are essential for relighting tasks. However, adapting MAE to our specific needs poses several non-trivial challenges. Firstly, MAE is primarily designed for vision transformers [14], while our focus is on a UNet, a convolution-based architecture. This convolutional structure, with its hierarchical nature and aggressive pooling, is known to simplify the MAE task, necessitating careful adaptation [49]. Further, the hyperparameters of MAE, particularly the fixed mask size and ratio, are also specific to vision transformers. These factors could introduce biases during training and hinder the model to recognize image features at various scales. Moreover, MAE relies solely on masked region reconstruction loss, limiting the model to understand the global coherence of the reconstructed region in relation to its visible context." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.591, + 0.469, + 0.621 + ], + "angle": 0, + "content": "To address these challenges effectively, we have developed two key strategies within the MMAE framework:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.469, + 0.746 + ], + "angle": 0, + "content": "Dynamic Masking. MMAE eliminates two key hyperparameters, mask size and ratio, by introducing a variety of mask types to generalize the MAE. These types, which include overlapping patches of various sizes, outpainting masks [45], and free-form masks [28] (see Fig.5), each contribute to the model's versatility. With the ability to handle challenging masked regions, MMAE achieves a more comprehensive understanding of image properties." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Generative Target. In addition to its structural advancements, MMAE incorporates a new loss function strategy. We have adopted perceptual [23] and adversarial losses [21], along with the original reconstruction loss. As a result, MMAE is equipped not only to reconstruct missing image parts but also to ensure synthesis capabilities and their seamless integration with the original context. In practice, the weights for the three losses are equally set." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.9 + ], + "angle": 0, + "content": "We pre-train the entire UNet architecture using MMAE, and, unlike MAE, we retain the decoder and fine-tune the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.771, + 0.107 + ], + "angle": 0, + "content": "entire model on relighting ground truths." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.12, + 0.567, + 0.136 + ], + "angle": 0, + "content": "5. Data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.145, + 0.892, + 0.341 + ], + "angle": 0, + "content": "We constructed the OLAT (One Light at a Time) dataset using a light stage [10, 48] equipped with 137 programmable LED lights and 7 frontal-viewing cameras. Our dataset comprises images of 287 subjects, with each subject being captured in up to 15 different poses, resulting in a total of 29,705 OLAT sequences. We sourced HDRI dataset from several publicly available archives. Specifically, we acquired 559 HDRI maps from Polyhaven, 76 from Noah Witchell, 364 from HDRMAPS, 129 from iHDRI, and 34 from eisklotz. In addition, we incorporated synthetic HDRIs created using the method proposed in [30]. During training, HDRIs are randomly selected with equal probability from either real-world or synthetic collections." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.417 + ], + "angle": 0, + "content": "We produced training pairs by projecting the sampled source and target lighting maps onto the reflectance fields of the OLAT images [10]. To derive the ground truth intrinsics, we applied the photometric stereo method [50] and obtained normal and albedo maps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.431, + 0.633, + 0.448 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.892, + 0.637 + ], + "angle": 0, + "content": "This section details our experimental results. We begin with a comparative evaluation of our method against state-of-the-art approaches using the OLAT dataset. We also employ images from the FFHQ-test [24] for user studies. For qualitative analysis, we utilize copyright-free portrait images from Pexels [1]. Additionally, we conduct ablation studies to validate the efficacy of our pre-training framework and architectural design choice. Subsequently, we detail the additional features and conclude by discussing its limitations. Our evaluation uses the OLAT test set, comprising 35 subjects and 11 lighting environments, ensuring no overlap with the train set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.892, + 0.777 + ], + "angle": 0, + "content": "Evaluation metrics. We employ several key metrics for evaluating the prediction accuracy; Mean Absolute Error (MAE), Mean Squared Error (MSE), Structural Similarity Index Measure (SSIM) and Learned Perceptual Image Patch Similarity (LPIPS). While these metrics offer valuable quantitative insights, they do not fully capture the subtleties of visual quality enhancement. Therefore, we emphasize the importance of qualitative evaluations to gain a comprehensive understanding of model performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Baselines. We compared our approach with three state-of-the-art baselines: Single Image Portrait Relighting (SIPR) [44], which uses a single neural network for relighting; Total Relight (TR) [33], employing multiple neural networks that incorporate the Phong reflectance model; and Lumos [51], a TR adaptation for synthetic datasets. Due to the lack of publicly available code or model from these methods, we either integrated their techniques into our" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25101" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.09, + 0.089, + 0.458, + 0.185 + ], + "angle": 0, + "content": "
MAE ↓MSE ↓SSIM ↑LPIPS ↓
SIPR [44]0.17150.07480.84323.648
TR [33]0.16430.06580.84653.425
Ours0.10230.02750.90022.137
Ours (w. MMAE)0.09330.02350.90512.059
" + }, + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.195, + 0.441, + 0.209 + ], + "angle": 0, + "content": "Table 1. Quantitative Evaluation on the OLAT test set." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.212, + 0.46, + 0.291 + ], + "angle": 0, + "content": "
Lumos [51]TR [33]Ours
Consistent Lighting0.04780.18520.7671
Facial Details0.20220.26020.5376
Similar Identity0.17410.24400.5819
" + }, + { + "type": "table_caption", + "bbox": [ + 0.146, + 0.301, + 0.4, + 0.315 + ], + "angle": 0, + "content": "Table 2. User Study on the FFHQ test set." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.32, + 0.467, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.556, + 0.469, + 0.584 + ], + "angle": 0, + "content": "Figure 6. Impact of Pre-training. The fine details such as specular highlights, skin tones, and shadows are notably improved." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.591, + 0.469, + 0.619 + ], + "angle": 0, + "content": "framework or requested the original authors to process our inputs with their models and share the results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.469, + 0.773 + ], + "angle": 0, + "content": "Quantitative Comparisons. The results in Table. 1 show our method outperforming SIPR and TR baselines, demonstrating the significance of incorporating advanced rendering physics and reflectance models. The transition from SIPR to TR emphasizes the value of physics-based design, while the shift from TR to our approach underscores the importance of transitioning from the empirical Phong model to the more accurate Cook-Torrance model. Additionally, pretraining contributes to further enhancements, as evidenced by the improved image details, depicted in Fig 6." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Qualitative Comparisons. Our relighting method exhibits several key advantages over previous approaches, as showcased in Fig. 7. It effectively harmonizes light direction and softness, avoiding harsh highlights and inaccurate lighting that are commonly observed in other methods. A notable strength of our approach lies in its ability to capture high-frequency details like specular highlights and hard shadows. Additionally, as shown in the second row, it preserves fa" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.892, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.421, + 0.878, + 0.436 + ], + "angle": 0, + "content": "Figure 7. Qualitative Comparison on the Pexels images [1]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.622 + ], + "angle": 0, + "content": "cial details and identity, ensuring high fidelity to the subject's original features and mitigating common distortions seen in previous approaches. Moreover, our approach excels in handling skin tones, producing natural and accurate results under various lighting conditions. This is clearly demonstrated in the fourth row, where our method contrasts sharply with the over-saturated or pale tones from previous methods. Finally, the nuanced treatment of hair is highlighted in the sixth row, where our approach maintains luster and detail, unlike the flattened effect typical in other methods. More qualitative results are available in our supplementary video demonstration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.893, + 0.9 + ], + "angle": 0, + "content": "User Study. We conducted a human subjective test to evaluate the visual quality of relighting results, summarized in Table. 2. In each test case, workers were presented with an input image and an environment map. They were asked to compare the relighting results from three methods—Ours, Lumos, and TR—based on three criteria: 1) consistency of lighting with the environment map, 2) preservation of facial details, and 3) maintenance of the original identity. To ensure unbiased evaluations, the order of the methods presented was randomized. To aid in understanding the concept of consistent lighting, relit balls were displayed alongside the images. The study included a total of 256 images, consisting of 32 portraits each illuminated with 8 different HDRIs. Each worker was tasked with selecting the best image for each specific criterion, randomly assessing 30 samples. A total of 47 workers participated in the study. The results indicate a strong preference for our results over the baseline methods across all evaluated metrics." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25102" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.462, + 0.185 + ], + "angle": 0, + "content": "
Pre-trainingMethodMAE ↓MSE ↓SSIM ↑LPIPS ↓
MAE [18]0.09520.02420.90072.096
MMAE0.09330.02350.90512.059
DiffuseNetAlbedo0.10530.02950.89852.197
Diff Render0.10230.02750.90022.137
" + }, + { + "type": "table_caption", + "bbox": [ + 0.129, + 0.196, + 0.417, + 0.209 + ], + "angle": 0, + "content": "Table 3. Ablation Studies on the OLAT test set." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.22, + 0.462, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.532, + 0.468, + 0.559 + ], + "angle": 0, + "content": "Figure 8. Ablation on DiffuseNet. Our approach successfully infers the albedo on various surfaces (skin, teeth, and accessories)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.574, + 0.47, + 0.862 + ], + "angle": 0, + "content": "Ablation Studies. We analyze our two major design choices in Table. 3: the MMAE pre-training framework and DiffuseNet. The MMAE, which integrates dynamic masking with generative objectives, outperforms MAE. This superiority is mainly due to the incorporation of challenging masks and global coherence objectives, enabling the model to learn richer features during pre-training. Furthermore, our method of predicting diffuse render demonstrates superiority over direct albedo prediction. Firstly, we see it simplifies the learning process, as predicting diffuse render is more closely related to the original image. Secondly, our approach effectively distinguishes between the influences of illumination (diffuse shading) and surface properties (diffuse render). This distinction is crucial for accurately modeling the intrinsic color of surfaces, as it enables independent and precise evaluation of each element (see Eqn. 9). In contrast, methods that predict albedo directly often struggle to differentiate between these factors, leading to significant inaccuracies in color constancy, as shown in Fig. 8." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Applications. We present two applications using predicted intrinsics in Fig. 9. First, real-time PBR via Cook-Torrance" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.64, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.171, + 0.629, + 0.18 + ], + "angle": 0, + "content": "(a) Real-Time PBR Rendering" + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.089, + 0.892, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.17, + 0.802, + 0.18 + ], + "angle": 0, + "content": "(b) Copy Light" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.183, + 0.894, + 0.21 + ], + "angle": 0, + "content": "Figure 9. Applications. We showcase additional features of SwitchLight, powered by the diverse intrinsics features." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.215, + 0.892, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.465, + 0.892, + 0.507 + ], + "angle": 0, + "content": "Figure 10. Limitations. The model faces challenges in removing strong shadows, misinterpreting reflective surfaces like sunglasses, and inaccurately predicting albedo for face paint." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.515, + 0.891, + 0.56 + ], + "angle": 0, + "content": "components in Three.js graphics library. Second, switching the lighting environment between the source and reference images. Further details are in the supplementary video." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.567, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Limitations. We identified a few failure cases in Fig. 10. First, our model struggles with neutralizing strong shadows, which leads to inaccurate facial geometry and residual shadow artifacts in both albedo and relit images. Incorporating shadow augmentation techniques [15, 53] during training could mitigate this issue. Second, the model incorrectly interprets reflective surfaces, such as sunglasses, as opaque objects in the normal image. This error prevents the model from properly removing reflective highlights in the albedo and relit images. Lastly, the model inaccurately predicts the albedo for face paint. Implementing a semantic mask [51] to distinguish different semantic regions separately from the skin could help resolving these issues." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.77, + 0.619, + 0.785 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduce SwitchLight, an architecture based on Cook-Torrance rendering physics, enhanced with a self-supervised pre-training framework. This co-designed approach significantly outperforms previous models. Our future plans include scaling the current model beyond images to encompass video and 3D data. We hope our proposal serve as a new foundational model for relighting tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25103" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.116, + 0.386, + 0.129 + ], + "angle": 0, + "content": "[1] Pexels. https://www.pexels.com.6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.131, + 0.469, + 0.17 + ], + "angle": 0, + "content": "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.173, + 0.469, + 0.225 + ], + "angle": 0, + "content": "[3] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.229, + 0.469, + 0.297 + ], + "angle": 0, + "content": "[4] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.299, + 0.469, + 0.354 + ], + "angle": 0, + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.355, + 0.469, + 0.423 + ], + "angle": 0, + "content": "[6] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22367-22377, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.425, + 0.469, + 0.466 + ], + "angle": 0, + "content": "[7] Zhaoxi Chen and Ziwei Liu. Relighting4d: Neural relightable human from videos. In European Conference on Computer Vision, pages 606-623. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.467, + 0.469, + 0.508 + ], + "angle": 0, + "content": "[8] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM Transactions on Graphics (ToG), 1(1):7-24, 1982. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.51, + 0.469, + 0.564 + ], + "angle": 0, + "content": "[9] Paul Debevec. Rendering synthetic objects into real scenes: Bridging traditional and image-based graphics with global illumination and high dynamic range photography. In Acm siggraph 2008 classes, pages 1-10. 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.566, + 0.469, + 0.634 + ], + "angle": 0, + "content": "[10] Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. Acquiring the reflectance field of a human face. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 145-156, 2000. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[12] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422-1430, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[13] Julie Dorsey, James Arvo, and Donald Greenberg. Interactive design of complex time dependent lighting. IEEE Computer Graphics and Applications, 15(2):26-36, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] David Futschik, Kelvin Ritland, James Vecore, Sean Fanello, Sergio Orts-Escalano, Brian Curless, Daniel Sykora, and Ro" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "hit Pandey. Controllable light diffusion for portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8412-8421, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[16] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[19] Andrew Hou, Ze Zhang, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Towards high fidelity face relighting with realistic shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14719-14728, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[20] Andrew Hou, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Face relighting with geometrically consistent shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4217-4226, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[21] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[22] Chaonan Ji, Tao Yu, Kaiwen Guo, Jingxin Liu, and Yebin Liu. Geometry-aware single-image full-body human relighting. In European Conference on Computer Vision, pages 388-405. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[23] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[24] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[25] Taehun Kim, Kunhee Kim, Joonyeong Lee, Dongmin Cha, Jiho Lee, and Daijin Kim. Revisiting image pyramid structure for high resolution salient object detection. In Proceedings of the Asian Conference on Computer Vision, pages 108-124, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[26] Wenbo Li, Xin Lu, Shengju Qian, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer-based image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25104" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.19 + ], + "angle": 0, + "content": "[28] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European conference on computer vision (ECCV), pages 85-100, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.261 + ], + "angle": 0, + "content": "[29] Yihao Liu, Jingwen He, Jinjin Gu, Xiangtao Kong, Yu Qiao, and Chao Dong. Degae: A new pretraining paradigm for low-level vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23292-23303, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.472, + 0.346 + ], + "angle": 0, + "content": "[30] Yiqun Mei, He Zhang, Xuaner Zhang, Jianming Zhang, Zhixin Shu, Yilin Wang, Zijun Wei, Shi Yan, HyunJoon Jung, and Vishal M Patel. Lightpainter: Interactive portrait relighting with freehand scribble. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 195-205, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.471, + 0.417 + ], + "angle": 0, + "content": "[31] Thomas Nestmeyer, Jean-François Lalonde, Iain Matthews, and Andreas Lehrmann. Learning physics-guided face relighting under directional light. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5124-5133, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.471, + 0.473 + ], + "angle": 0, + "content": "[32] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.475, + 0.471, + 0.545 + ], + "angle": 0, + "content": "[33] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.471, + 0.587 + ], + "angle": 0, + "content": "[34] Namuk Park, Wonjae Kim, Byeongho Heo, Taekyung Kim, and Sangdoo Yun. What do self-supervised vision transformers learn? arXiv preprint arXiv:2305.00729, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[35] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[36] Bui Tuong Phong. Illumination for computer generated pictures. In Seminal graphics: pioneering efforts that shaped the field, pages 95-101. 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.471, + 0.744 + ], + "angle": 0, + "content": "[37] Puntawat Ponglertnapakorn, Nontawat Tritrong, and Supasorn Suwajanakorn. Difareli: Diffusion face relighting. arXiv preprint arXiv:2304.09479, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[38] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[39] Soumyadip Sengupta, Angjoo Kanazawa, Carlos D Castillo, and David W Jacobs. Sfsnet: Learning shape, reflectance and illuminance of facesin the wild'. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6296-6305, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[40] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Pro-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2291-2300, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[41] Soumyadip Sengupta, Brian Curless, Ira Kemelmacher-Shlizerman, and Steven M Seitz. A light stage on every desk. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2420-2429, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[42] YiChang Shih, Sylvain Paris, Connelly Barnes, William T Freeman, and Fredo Durand. Style transfer for headshot portraits. 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[43] Zhixin Shu, Sunil Hadap, Eli Shechtman, Kalyan Sunkavalli, Sylvain Paris, and Dimitris Samaras. Portrait lighting transfer using a mass transport approach. ACM Transactions on Graphics (TOG), 36(4):1, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.279, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[44] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.35, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[45] Piotr Teterwak, Aaron Sarna, Dilip Krishnan, Aaron Maschinot, David Belanger, Ce Liu, and William T Freeman. Boundless: Generative adversarial networks for image extension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10521-10530, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.435, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[46] Yifan Wang, Aleksander Holynski, Xiuming Zhang, and Xuaner Zhang. Sunstage: Portrait reconstruction and relighting using the sun as a light stage. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20792-20802, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.561 + ], + "angle": 0, + "content": "[47] Zhibo Wang, Xin Yu, Ming Lu, Quan Wang, Chen Qian, and Feng Xu. Single image portrait relighting via explicit multiple reflectance channel modeling. ACM Transactions on Graphics (TOG), 39(6):1-13, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.563, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[48] Andreas Wenger, Andrew Gardner, Chris Tchou, Jonas Unger, Tim Hawkins, and Paul Debevec. Performance relighting and reflectance transformation with timemultiplexed illumination. ACM Transactions on Graphics (TOG), 24(3):756-764, 2005. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[49] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133-16142, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[50] Robert J Woodham. Photometric method for determining surface orientation from multiple images. Optical engineering, 19(1):139-144, 1980. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[51] Yu-Ying Yeh, Koki Nagano, Sameh Khamis, Jan Kautz, Ming-Yu Liu, and Ting-Chun Wang. Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. ACM Transactions on Graphics (TOG), 41(6): 1-21, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[52] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 649-666. Springer, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25105" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[53] Xuaner Zhang, Jonathan T Barron, Yun-Ta Tsai, Rohit Pandey, Xiuming Zhang, Ren Ng, and David E Jacobs. Portrait shadow manipulation. ACM Transactions on Graphics (TOG), 39(4):78-1, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[54] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7194-7202, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[55] Taotao Zhou, Kai He, Di Wu, Teng Xu, Qixuan Zhang, Kuixiang Shao, Wenzheng Chen, Lan Xu, and Jingyi Yu. Relightable neural human assets from multi-view gradient illuminations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4315-4327, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25106" + } + ] +] \ No newline at end of file diff --git a/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/ce409cac-fd5a-4614-9e6d-87e790ca611e_origin.pdf b/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/ce409cac-fd5a-4614-9e6d-87e790ca611e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2d6a352e7687fa98311fc6b50b0f9e66febe8c2b --- /dev/null +++ b/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/ce409cac-fd5a-4614-9e6d-87e790ca611e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:06a4a72f4113b05e792cc79cd688d9b50450485f8f6b3891562c2ecac901e5fc +size 5258159 diff --git a/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/full.md b/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f255409260d320e3713cbb8fec028ab9e5c4fd40 --- /dev/null +++ b/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/full.md @@ -0,0 +1,337 @@ +# SwitchLight: Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting + +Hoon Kim1 Minje Jang1 Wonjun Yoon1 Jisoo Lee1 Donghyun Na1 Sanghyun Woo2 + +$^{1}$ Beeble AI $^{2}$ New York University + +![](images/c3a9957805d3c58bc988053a24741f13d3fa4a2761520550a6037eb132557f0e.jpg) +Input RGB +Normal +Albedo +Roughness +Reflectivity +Figure 1. Be Anywhere at Any Time. SwitchLight processes a human portrait by decomposing it into detailed intrinsic components, and re-renders the image under a designated target illumination, ensuring a seamless composition of the subject into any new environment. + +# Abstract + +We introduce a co-designed approach for human portrait relighting that combines a physics-guided architecture with a pre-training framework. Drawing on the Cook-Torrance reflectance model, we have meticulously configured the architecture design to precisely simulate light-surface interactions. Furthermore, to overcome the limitation of scarce high-quality lightstage data, we have developed a self-supervised pre-training strategy. This novel combination of accurate physical modeling and expanded training dataset establishes a new benchmark in relighting realism. + +# 1. Introduction + +Relighting is more than an aesthetic tool; it unlocks infinite narrative possibilities and enables seamless integration of subjects into diverse environments (see Fig. 1). This advancement resonates with our innate desire to transcend the physical constraints of space and time, while also providing tangible solutions to practical challenges in digital content creation. It is particularly transformative in virtual (VR) and augmented reality (AR) applications, where relighting facilitates the real-time adaptation of lighting, ensuring that users and digital elements coexist naturally within any environment, offering a next level of telepresence. + +In this work, we focus on human portrait relighting. + +While the relighting task fundamentally demands an in-depth understanding of geometry, material properties, and illumination, the challenge is more compounded when addressing human subjects, due to the unique characteristics of skin surfaces as well as the diverse textures and reflectance properties of a wide array of clothing, hairstyles, and accessories. These elements interact in complex ways, necessitating advanced algorithms capable of simulating the subtle interplay of light with these varied surfaces. + +Currently, the most promising approach involves the use of deep neural networks trained on pairs of high-quality relict portrait images and their corresponding intrinsic attributes, which are sourced from a light stage setup [10]. Initial efforts approached the relighting process as a 'black box' [44, 47], without delving into the underlying mechanisms. Later advancements adopted a physics-guided model design, incorporating the explicit modeling of image intrinsics and image formation physics [31]. Pandey et al. [33] proposed the Total Relight (TR) architecture, also physics-guided, which decomposes an input image into surface normals and albedo maps, and performs relighting based on the Phong specular reflectance model. The TR architecture has become foundational model for image relighting, with most recent and advanced architectures building upon its principle [22, 30, 51]. + +Following the physics-guided approach, our contribution lies in a co-design of architecture with a self-supervised pre-training framework. First, our architecture evolves towards a more accurate physical model by integrating the Cook-Torrance specular reflectance model [8], representing a notable advancement from the empirical Phong specular model [36] employed in the Total Relight architecture. The Cook-Torrance model adeptly simulates light interactions with surface microfacets, accounting for spatially varying roughness and reflectivity. Secondly, our pretraining framework scales the learning process beyond the typically hard-to-obtain lightstage data. By revisiting the masked autoencoder (MAE) framework [18], we adept it for the task of relighting. These modifications are crafted to address the unique challenges posed by this task, enabling our model to learn from unlabelled data and refine its ability to produce realistic relit portraits during fine-tuning. To the best of our knowledge, this is the first time applying self-supervised pre-training specifically to the relighting task. + +To summarize, our contribution is twofold. Firstly, by enhancing the physical reflectance model, we have introduced a new level of realism in the output. Secondly, by adopting self-supervised learning, we have expanded the scale of the training data and enhanced the expression of lighting in diverse real-world scenarios. Collectively, these advancements have led SwitchLight framework to achieve a new state-of-the-art in human portrait relighting. + +# 2. Related Work + +Human Portrait Relighting is an ill-posed problem due to its under-constrained nature. To tackle this, earlier methods incorporated 3D facial priors [43], exploited image intrinsics [3, 39], or framed the task as a style transfer [42]. Light stage techniques [48] offer a more powerful solution by recording subject's reflectance fields under varying lighting conditions [10, 13], though they are labor-intensive and require specialized equipment. A promising alternative has emerged with deep learning, utilizing neural networks trained on light stage data. Sun et al. [44] pioneered this approach, but their method had limitations in representing non-Lambertian effects. This was improved upon by Nestmeyer et al. [31], who integrated rendering physics into network design, albeit limited to directional light. Building upon this, Pandey et al. [33] incorporated the Phong reflection model and a high dynamic range (HDR) lighting map [9] into their network, enabling a more accurate representation of global illumination. Simultaneously, efforts have been made to explore portrait relighting without light stage data [19, 20, 41, 46, 54]. Moreover, introduction of NeRF [7] and diffusion-based [37] models has opened new avenues in the field. However, networks trained with light-stage data maintain superior accuracy and realism, thanks to physics-based composited relight image training pairs and precise ground truth image intrinsics [55]. + +Our work furthers this domain by integrating the Cook-Torrance model into our network design, shifting from the empirical Phong model to a more physics-based approach, thereby enhancing the realism and detail in relit images. + +Self-supervised Pre-training has become a standard training scheme in the development of large language models like BERT [11] and GPT [38], and is increasingly influential in vision models, aiming to replicate the 'BERT moment'. This approach typically involves pre-training on extensive unlabeled data, followed by fine-tuning on specific tasks. While early efforts in vision models focused on simple pretext tasks [12, 16, 32, 35, 52], the field has evolved through stages like contrastive learning [5, 17] and masked image modeling [2, 18, 49]. However, the primary focus has remained on visual recognition, with less attention to other domains. Exceptions include low-level image processing tasks [4, 6, 26, 29] using the vision transformer [14]. + +Our research takes a different route, focusing on human portrait relighting—a complex challenge of manipulating illumination in the image. This direction is crucial because acquiring accurate ground truth data, especially from light stage, is both expensive and difficult. We modify the MAE framework [18], previously successful in robust image representation learning and developing locality biases [34], to suit the unique requirements of effective relighting. + +# 3. SwitchLight + +We introduce SwitchLight, a state-of-the-art framework for human portrait relighting, with its architectural overview presented in Fig. 2. We first provide foundational concepts in Sec. 3.1, and define the problem in Sec. 3.2. This is followed by detailing the architecture in Sec. 3.3, and lastly, we describe the loss functions used in Sec. 3.4. + +# 3.1. Preliminaries + +In this section, vectors $\mathbf{n}$ , $\mathbf{v}$ , $\mathbf{l}$ , and $\mathbf{h}$ are denoted as unit vectors. Specifically, $\mathbf{n}$ represents the surface normal, $\mathbf{v}$ is the view direction, $\mathbf{l}$ is the incident light direction, and $\mathbf{h}$ is the half-vector computed from $\mathbf{l}$ and $\mathbf{v}$ . The dot product is clamped between [0..1], indicated by $\langle \cdot \rangle$ . + +Image Rendering. The primary goal of image rendering is to create a visual representation that accurately simulates the interactions between light and surfaces. These complex interactions are encapsulated by the rendering equation: + +$$ +L _ {o} (\mathbf {v}) = \int_ {\Omega} f (\mathbf {v}, \mathbf {l}) L _ {i} (\mathbf {l}) \langle \mathbf {n} \cdot \mathbf {l} \rangle d \mathbf {l} \tag {1} +$$ + +where $L_{o}(\mathbf{v})$ denotes the radiance, or the light intensity perceived by the observer in direction $\mathbf{v}$ . It is the cumulative result of incident lights $L_{i}(\mathbf{l})$ from all possible directions over the hemisphere, $\Omega$ , centered around the surface normal, denoted as $\mathbf{n}$ . Central to this equation lies the Bidirectional Reflectance Distribution Function (BRDF), denoted as $f(\mathbf{v},\mathbf{l})$ , describing the surface's reflection characteristics. + +BRDF Composition. The BRDF, represented by $f(\mathbf{v},\mathbf{l})$ , describes how light is reflected at an opaque surface. It is composed of two major components: diffuse reflection ( $f_{d}$ ) and specular reflection ( $f_{s}$ ): + +$$ +f (\mathbf {v}, \mathbf {l}) = f _ {d} (\mathbf {v}, \mathbf {l}) + f _ {s} (\mathbf {v}, \mathbf {l}) \tag {2} +$$ + +A surface intrinsically exhibits both diffuse and specular reflections. The diffuse component uniformly scatters light, ensuring consistent illumination regardless of the viewing angle. In contrast, the specular component is viewing angle-dependent, producing shiny highlights that are crucial for achieving a photorealistic effect. + +Lambertian Diffuse Reflectance. Lambertian reflectance, a standard model for diffuse reflection, describes a uniform light scatter irrespective of the viewing angle. This ensures a consistent appearance in brightness: + +$$ +f _ {d} (\mathbf {v}, \mathbf {l}) = \frac {\sigma}{\pi} [ \text {c o n s t .} ] \tag {3} +$$ + +Here, $\sigma$ is the albedo, indicating the intrinsic color and brightness of the surface. + +Cook-Torrance Specular Reflectance. The Cook-Torrance model, based on microfacet theory, represents surfaces as a myriad of tiny, mirror-like facets. It incorporates a roughness parameter $\alpha$ , which allows precise rendering of surface specular reflectance: + +$$ +f _ {s} (\mathbf {v}, \mathbf {l}) = \frac {D (\mathbf {h} , \alpha) G (\mathbf {v} , \mathbf {l} , \alpha) F (\mathbf {v} , \mathbf {h} , f _ {0})}{4 \langle \mathbf {n} \cdot \mathbf {l} \rangle \langle \mathbf {n} \cdot \mathbf {v} \rangle} \tag {4} +$$ + +In this model, $D$ is the microfacet distribution function, describing the orientation of the microfacets relative to the half-vector $h$ , $G$ is the geometric attenuation factor, accounting for the shadowing and masking of microfacets, and $F$ is the Fresnel term, calculating the reflectance variation depending on the viewing angle, where $f_{0}$ is the surface Fresnel reflectivity at normal incidence. A lower $\alpha$ value implies a smoother surface with sharper specular highlights, whereas a higher $\alpha$ value indicates a rougher surface, resulting in more diffused reflections. By adjusting $\alpha$ , the Cook-Torrance model can depict a range of specular reflections. + +Image Formation. Upon the base rendering equation, we include the diffuse and specular components of the BRDF and derive a unified formula: + +$$ +L _ {o} (\mathbf {v}) = \int_ {\Omega} \left(f _ {d} (\mathbf {v}, \mathbf {l}) + f _ {s} (\mathbf {v}, \mathbf {l})\right) E (\mathbf {l}) \langle \mathbf {n} \cdot \mathbf {l} \rangle d \mathbf {l} \tag {5} +$$ + +where $E(1)$ denotes the incident environmental lighting. This formula represents the core principle that an image is a product of interplay between the BRDF and lighting. To further clarify this concept, we introduce a rendering function $R$ , succinctly modeling the process of image formation: + +$$ +I = R (\underbrace {\mathbf {n} , \sigma , \alpha , f _ {0}} _ {\text {s u r f a c e a t t r i b u t e s l i g h t i n g}}, \underbrace {E} _ {\text {l i g h t i n g}}) \tag {6} +$$ + +It is important to note that since the BRDF is a function of surface properties, as detailed in Eqn. 3 and 4, we can now clearly understand that image formation is essentially governed by the interaction of surface attributes and lighting. + +# 3.2. Problem Formulation + +Image Relighting. Given the image formation model above, our goal is to manipulate the lighting of an existing image. This involves two main steps: inverse rendering and re-rendering under target illumination, both driven by neural networks. For a given source image $I_{\mathrm{src}}$ and target illumination $E_{\mathrm{tgt}}$ , the process is delineated as: + +$$ +\text {I n v e r s e} (\mathbf {n}, \sigma , \alpha , f _ {0}, E _ {\mathrm {s r c}}) = U (I _ {\mathrm {s r c}}) +$$ + +Rendering with Target Light. $I_{\mathrm{tgt}} = R(\mathbf{n},\sigma ,\alpha ,f_0,E_{\mathrm{tgt}})$ + +During the inverse rendering step, the function $U$ unravels the intrinsic properties of $I_{\mathrm{src}}$ . In the subsequent relighting + +![](images/3aef3794bddc2d79a3b52b533ce87d0ccbbbcaa0c5c7d51a523e59572face98c.jpg) +Figure 2. SwitchLight Architecture. The input source image is decomposed into normal map, lighting, diffuse and specular components. Given these intrinsic, images are re-rendered under target lighting. The architecture integrates the Cook-Torrance reflection model; the final output combines physically-based predictions with neural network enhancements for realistic portrait relighting. + +step, the derived intrinsic properties along with new illumination $E_{\mathrm{tgt}}$ are employed by the rendering function $R$ to generate the target relit image $I_{\mathrm{tgt}}$ . + +# 3.3. Architecture + +Our architecture systematically executes the two primary stages outlined in our problem formulation. The first stage involves extracting intrinsic properties from the source image $I_{\mathrm{src}}$ . For this purpose, we employ a matting network [25, 27, 40] to accurately separate the foreground. This extracted image is then processed by our inverse rendering network $U$ , which infers normal, albedo, roughness, reflectivity, and source lighting. Subsequently, the second stage involves re-rendering the image under new target lighting conditions. To achieve this, the acquired intrinsics, along with the target lighting $E_{\mathrm{tgt}}$ , are fed into our relighting network $R$ , producing the relit image $I_{\mathrm{tgt}}$ . + +Normal Net. The network takes the source image $I_{\mathrm{src}} \in \mathbb{R}^{\mathrm{H} \times \mathrm{W} \times 3}$ and generates a normal map $\hat{\mathbf{N}}$ . Each pixel in this map contains a unit normal vector $\hat{\mathbf{n}}$ , indicating the orientation of the corresponding surface point. + +Illum Net. The network infers the lighting conditions in the given image captured in an HDRI format. Specifically, it computes the convolved HDRIs: + +$$ +E _ {\mathrm {s r c}} ^ {p} \left(\mathbf {l} ^ {\prime}\right) = \int_ {\Omega} \underbrace {E _ {\mathrm {s r c}} (\mathbf {l})} _ {\text {H D R I}} \underbrace {\langle \mathbf {l} ^ {\prime} \cdot \mathbf {l} \rangle^ {p}} _ {\text {P h o n g l o b e}} d \mathbf {l} \tag {7} +$$ + +In this equation, $E_{\mathrm{src}} \in \mathbb{R}^{\mathrm{H}_{\mathrm{HDRI}} \times \mathrm{W}_{\mathrm{HDRI}} \times 3}$ is the original source HDRI map, with 1 indicating spherical directions in the HDRI space $\mathbb{R}^{\mathrm{H}_{\mathrm{HDRI}} \times \mathrm{W}_{\mathrm{HDRI}}}$ . The term $\langle \mathbf{l}' \cdot \mathbf{l} \rangle^{p}$ represents the Phong reflectance lobe with shininess exponents $p \in \{1, 16, 32, 64\}$ , which incorporates various specular terms. Consequently, it is expressed in a multi-dimensional tensor form as $\mathbb{R}^{4 \times \mathrm{H}_{\mathrm{HDRI}} \times \mathrm{W}_{\mathrm{HDRI}} \times \mathrm{H}_{\mathrm{HDRI}} \times \mathrm{W}_{\mathrm{HDRI}}}$ . Finally, $E_{\mathrm{src}}^{p} \in$ + +$\mathbb{R}^{4\times \mathrm{H_{cHDRI}}\times \mathrm{W_{cHDRI}}\times 3}$ is the convolved HDRI. In this work, we set the resolution of HDRI and convolved HDRI at $32\times 64$ and $64\times 128$ , respectively, and we apply convolution on light source coordinates. + +The network employs a cross-attention mechanism at its core, where predefined Phong reflectance lobes serve as queries, and the original image acts as both keys and values. Within this setup, the convolved HDRI maps are synthesized by integrating image information into the Phong reflectance lobe representation. Specifically, our model utilizes bottleneck features from the Normal Net as a compact image representation. Our approach simplifies the complex task of HDRI reconstruction by instead focusing on estimating interactions with known surface reflective properties. + +Diffuse Net. Estimating albedo is challenging due to the ambiguities in surface color and material properties, further complicated by shadow effects. To address this, we prioritize the inference of source diffuse render, $I_{\mathrm{src,diff}}$ : + +$$ +L _ {\mathrm {s r c}, o _ {\mathrm {d i f f}}} (\mathbf {v}) = \underbrace {\frac {\sigma}{\pi}} _ {\text {d i f f u s e B R D F}} \underbrace {\int_ {\Omega} E _ {\mathrm {s r c}} (\mathbf {l}) \left\langle \mathbf {n} \cdot \mathbf {l} \right\rangle d \mathbf {l}} _ {\text {d i f f u s e s h a d i n g}} \tag {8} +$$ + +Our key insight is that the diffuse render closely resembles the original image, which simplifies the model learning process. It captures surface color after removing specular reflections, such as shine or gloss, contrasting with albedo that represents the true surface color unaffected by lighting and shadows. The network takes a source image $I_{\mathrm{src}}$ , concatenated with its diffuse shading, to produce the diffuse render. As in Eqn. 7, the diffuse shading, $\hat{E}_{\mathrm{src}}^{1}(\hat{\mathbf{n}})$ , is derived using the predicted normals, $\hat{\mathbf{n}}$ , and the predicted lighting map, $\hat{E}_{\mathrm{src}}^{1}$ , with a Phong exponent of 1 for the diffuse term. The albedo map $\hat{\mathbf{A}}$ is then computed by dividing the predicted diffuse render by its diffuse shading: + +$$ +\frac {\hat {\sigma}}{\pi} = \frac {\hat {L} _ {\mathrm {s r c} , o _ {\mathrm {d i f f}}} (\mathbf {v})}{\hat {E} _ {\mathrm {s r c}} ^ {1} (\hat {\mathbf {n}})} \tag {9} +$$ + +![](images/6bb454551f13b8fc93664e680abe94d1c2f018ff65f717cb84723e3772b4fe3e.jpg) +Figure 3. Render Net Overview. Utilizing extracted image intrinsics, it employs the Cook-Torrance model for initial relighting and a neural network for enhanced refinement, producing high-fidelity relit images through a synergistic computational approach. + +We have empirically validated that it significantly enhances albedo prediction across a range of real-world scenarios. + +Specular Net. The network infers surface attributes associated with the Cook-Torrance specular elements, specifically, the roughness $\alpha$ and Fresnel reflectivity $f_{0}$ . It uses a source image, predicted normal, and albedo maps as inputs. + +Render Net. The network utilizes extracted intrinsic surface attributes to produce the target relit images. It generates two types of relit images, as shown in Fig. 3. The first type adheres to the physically-based rendering (PBR) principles of the Cook-Torrance model. This involves computing diffuse and specular renders under the target illumination using Eqn. 3 and Eqn. 4. These renders are combined to form the PBR render, $\hat{I}_{\mathrm{tgt}}^{\mathrm{PBR}}$ , as: + +$$ +\hat {L} _ {\mathrm {t g t}, o} ^ {\mathrm {P B R}} (\mathbf {v}) = \hat {L} _ {\mathrm {t g t}, o _ {\text {d i f f}}} (\mathbf {v}) + \hat {L} _ {\mathrm {t g t}, o _ {\text {s p e c}}} (\mathbf {v}) \tag {10} +$$ + +The second type of relit image is the result of a neural network process. This enhances the PBR render, capturing finer details that the Cook-Torrance model might miss. It employs the albedo, along with the diffuse and specular renders from the Cook-Torrance model, to infer a more refined target relit image, termed the neural render, $\hat{I}_{\mathrm{tgt}}^{\mathrm{Neural}}$ . The qualitative improvements achieved through this neural enhancement are illustrated in Fig. 4. + +# 3.4. Objectives + +We supervise both intrinsic image attributes and relit images using their corresponding ground truths, obtained from the lightstage. We employ a combination of reconstruction, perceptual [23], adversarial [21], and specular [33] losses. + +Reconstruction Loss $\mathcal{L} = \ell_1(\mathbf{X},\hat{\mathbf{X}})$ . It measures the pixel-level differences between the ground truth $\mathbf{X}$ and its prediction $\hat{\mathbf{X}}$ . This loss is applied across different attributes, including normal map $\hat{\mathbf{N}}$ , convolved source HDRI maps $\hat{E}_{\mathrm{src}}^p$ , source diffuse render $\hat{I}_{\mathrm{src,diff}}$ , albedo map $\hat{\mathbf{A}}$ , and both + +![](images/38f6fd31e8f47999f2ffdde60969933648706529164eba58e1d6b671effee889.jpg) +Figure 4. Neural Render Enhancement. Using the Cook-Torrance model, diffuse and specular renders are computed, which are then composited into a physically-based rendering. Subsequently, a neural network enhances this PBR render, improving aspects such as brightness and specular details. + +types of target relit images $\hat{I}_{\mathrm{tgt}}^{\mathrm{PBR}}$ and $\hat{I}_{\mathrm{tgt}}^{\mathrm{Neural}}$ . The attributes like the roughness $\alpha$ and Fresnel reflectivity $f_{0}$ are also implicitly learned when this loss applied to the PBR render. + +Perceptual Loss $\mathcal{L}_{\mathrm{vgg}} = \ell_2(\mathbf{VGG}(\mathbf{X}),\mathbf{VGG}(\hat{\mathbf{X}}))$ . It captures high-level feature differences based on a VGGnetwork feature comparison. We apply this loss to the source diffuse render, albedo, and target relit images. + +Adversarial Loss $\mathcal{L}_{\mathrm{adv}} = \mathrm{GAN}(\mathbf{X},\hat{\mathbf{X}})$ . It promotes realism in the generated images by fooling a discriminator network. This loss is applied to the target relit images. We employ a PatchGAN architecture, with detailed specifications provided in the supplementary material. + +Specular Loss $\mathcal{L}_{\mathrm{spec}} = \ell_1(\mathbf{X} \odot \hat{\mathbf{S}}, \hat{\mathbf{X}} \odot \hat{\mathbf{S}})$ . It enhances the specular highlights in the relit images. Specifically, we utilize the predicted specular render $\hat{\mathbf{S}} := \hat{I}_{\mathrm{tgt},\mathrm{spec}}^{\mathrm{PBR}}$ derived from the Cook-Torrance physical model, to weigh the $\ell_1$ reconstruction loss. Here, $\odot$ denotes the element-wise multiplication. This loss is applied to the neural render. + +Final Loss. The SwitchLight is trained in an end-to-end manner using the weighted sum of the above losses: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {r e l i g h t}} = 1 0 \cdot \mathcal {L} _ {\text {n o r m a l}} + 1 0 \cdot \mathcal {L} _ {\text {s r c - H D R I}} + 0. 2 \cdot \mathcal {L} _ {\text {s r c - d i f f}} \\ + 0. 2 \cdot \mathcal {L} _ {\text {a l b e d o}} + 0. 2 \cdot \mathcal {L} _ {\text {P B R}} + 0. 2 \cdot \mathcal {L} _ {\text {N e u r a l}} \tag {11} \\ + \mathcal {L} _ {\mathrm {v g g} _ {\text {s r c . d i f f}}} + \mathcal {L} _ {\mathrm {v g g} _ {\text {a l b e d o}}} + \mathcal {L} _ {\mathrm {v g g} _ {\mathrm {P B R}}} + \mathcal {L} _ {\mathrm {v g g} _ {\text {N e u r a l}}} \\ + \mathcal {L} _ {\mathrm {a d v} _ {\mathrm {P B R}}} + \mathcal {L} _ {\mathrm {a d v} _ {\mathrm {N e u r a l}}} + 0. 2 \cdot \mathcal {L} _ {\mathrm {s p e c} _ {\mathrm {N e u r a l}}}. \\ \end{array} +$$ + +![](images/032dec24f947d1a38a99291244c0abcfd6d0d46af8cdf75294c55d89ec87bf49.jpg) +Figure 5. Dynamic Masking Strategies. We have generalized the MAE masks to include overlapping patches of varying sizes, as well as outpainting and free-form masks. + +![](images/1ef7513b61de6a217ad8a36fdada2ab8dcd0f0604be62383edb9886b319259fc.jpg) + +![](images/d4ef083b15cd653dd0d380e71b555f4c3e49705032698bca3c113325a261ae4b.jpg) + +![](images/51234a997d19a0a4b7a616e62ee849c52f28158a3af14f5ad85d0a9d599b427d.jpg) + +We empirically determined the weighting coefficients. + +# 4. Multi-Masked Autoencoder Pre-training + +We introduce the Multi-Masked Autoencoder (MMAE), a self-supervised pre-training framework designed to enhance feature representations in relighting models. It aims to improve output quality without relying on additional, costly light stage data. Building upon the MAE framework [18], MMAE capitalizes on the inherent learning of crucial image features like structure, color, and texture, which are essential for relighting tasks. However, adapting MAE to our specific needs poses several non-trivial challenges. Firstly, MAE is primarily designed for vision transformers [14], while our focus is on a UNet, a convolution-based architecture. This convolutional structure, with its hierarchical nature and aggressive pooling, is known to simplify the MAE task, necessitating careful adaptation [49]. Further, the hyperparameters of MAE, particularly the fixed mask size and ratio, are also specific to vision transformers. These factors could introduce biases during training and hinder the model to recognize image features at various scales. Moreover, MAE relies solely on masked region reconstruction loss, limiting the model to understand the global coherence of the reconstructed region in relation to its visible context. + +To address these challenges effectively, we have developed two key strategies within the MMAE framework: + +Dynamic Masking. MMAE eliminates two key hyperparameters, mask size and ratio, by introducing a variety of mask types to generalize the MAE. These types, which include overlapping patches of various sizes, outpainting masks [45], and free-form masks [28] (see Fig.5), each contribute to the model's versatility. With the ability to handle challenging masked regions, MMAE achieves a more comprehensive understanding of image properties. + +Generative Target. In addition to its structural advancements, MMAE incorporates a new loss function strategy. We have adopted perceptual [23] and adversarial losses [21], along with the original reconstruction loss. As a result, MMAE is equipped not only to reconstruct missing image parts but also to ensure synthesis capabilities and their seamless integration with the original context. In practice, the weights for the three losses are equally set. + +We pre-train the entire UNet architecture using MMAE, and, unlike MAE, we retain the decoder and fine-tune the + +entire model on relighting ground truths. + +# 5. Data + +We constructed the OLAT (One Light at a Time) dataset using a light stage [10, 48] equipped with 137 programmable LED lights and 7 frontal-viewing cameras. Our dataset comprises images of 287 subjects, with each subject being captured in up to 15 different poses, resulting in a total of 29,705 OLAT sequences. We sourced HDRI dataset from several publicly available archives. Specifically, we acquired 559 HDRI maps from Polyhaven, 76 from Noah Witchell, 364 from HDRMAPS, 129 from iHDRI, and 34 from eisklotz. In addition, we incorporated synthetic HDRIs created using the method proposed in [30]. During training, HDRIs are randomly selected with equal probability from either real-world or synthetic collections. + +We produced training pairs by projecting the sampled source and target lighting maps onto the reflectance fields of the OLAT images [10]. To derive the ground truth intrinsics, we applied the photometric stereo method [50] and obtained normal and albedo maps. + +# 6. Experiments + +This section details our experimental results. We begin with a comparative evaluation of our method against state-of-the-art approaches using the OLAT dataset. We also employ images from the FFHQ-test [24] for user studies. For qualitative analysis, we utilize copyright-free portrait images from Pexels [1]. Additionally, we conduct ablation studies to validate the efficacy of our pre-training framework and architectural design choice. Subsequently, we detail the additional features and conclude by discussing its limitations. Our evaluation uses the OLAT test set, comprising 35 subjects and 11 lighting environments, ensuring no overlap with the train set. + +Evaluation metrics. We employ several key metrics for evaluating the prediction accuracy; Mean Absolute Error (MAE), Mean Squared Error (MSE), Structural Similarity Index Measure (SSIM) and Learned Perceptual Image Patch Similarity (LPIPS). While these metrics offer valuable quantitative insights, they do not fully capture the subtleties of visual quality enhancement. Therefore, we emphasize the importance of qualitative evaluations to gain a comprehensive understanding of model performance. + +Baselines. We compared our approach with three state-of-the-art baselines: Single Image Portrait Relighting (SIPR) [44], which uses a single neural network for relighting; Total Relight (TR) [33], employing multiple neural networks that incorporate the Phong reflectance model; and Lumos [51], a TR adaptation for synthetic datasets. Due to the lack of publicly available code or model from these methods, we either integrated their techniques into our + +
MAE ↓MSE ↓SSIM ↑LPIPS ↓
SIPR [44]0.17150.07480.84323.648
TR [33]0.16430.06580.84653.425
Ours0.10230.02750.90022.137
Ours (w. MMAE)0.09330.02350.90512.059
+ +Table 1. Quantitative Evaluation on the OLAT test set. + +
Lumos [51]TR [33]Ours
Consistent Lighting0.04780.18520.7671
Facial Details0.20220.26020.5376
Similar Identity0.17410.24400.5819
+ +Table 2. User Study on the FFHQ test set. + +![](images/8cba26b02b2c8544c446226b80440753c45bb31480af2d73c51fe9345f89cff7.jpg) +Figure 6. Impact of Pre-training. The fine details such as specular highlights, skin tones, and shadows are notably improved. + +framework or requested the original authors to process our inputs with their models and share the results. + +Quantitative Comparisons. The results in Table. 1 show our method outperforming SIPR and TR baselines, demonstrating the significance of incorporating advanced rendering physics and reflectance models. The transition from SIPR to TR emphasizes the value of physics-based design, while the shift from TR to our approach underscores the importance of transitioning from the empirical Phong model to the more accurate Cook-Torrance model. Additionally, pretraining contributes to further enhancements, as evidenced by the improved image details, depicted in Fig 6. + +Qualitative Comparisons. Our relighting method exhibits several key advantages over previous approaches, as showcased in Fig. 7. It effectively harmonizes light direction and softness, avoiding harsh highlights and inaccurate lighting that are commonly observed in other methods. A notable strength of our approach lies in its ability to capture high-frequency details like specular highlights and hard shadows. Additionally, as shown in the second row, it preserves fa + +![](images/5b4143c5ca3289a0076e8b0113ea0f9cac2360741a92a669f01f36dc29f0b7ae.jpg) +Figure 7. Qualitative Comparison on the Pexels images [1]. + +cial details and identity, ensuring high fidelity to the subject's original features and mitigating common distortions seen in previous approaches. Moreover, our approach excels in handling skin tones, producing natural and accurate results under various lighting conditions. This is clearly demonstrated in the fourth row, where our method contrasts sharply with the over-saturated or pale tones from previous methods. Finally, the nuanced treatment of hair is highlighted in the sixth row, where our approach maintains luster and detail, unlike the flattened effect typical in other methods. More qualitative results are available in our supplementary video demonstration. + +User Study. We conducted a human subjective test to evaluate the visual quality of relighting results, summarized in Table. 2. In each test case, workers were presented with an input image and an environment map. They were asked to compare the relighting results from three methods—Ours, Lumos, and TR—based on three criteria: 1) consistency of lighting with the environment map, 2) preservation of facial details, and 3) maintenance of the original identity. To ensure unbiased evaluations, the order of the methods presented was randomized. To aid in understanding the concept of consistent lighting, relit balls were displayed alongside the images. The study included a total of 256 images, consisting of 32 portraits each illuminated with 8 different HDRIs. Each worker was tasked with selecting the best image for each specific criterion, randomly assessing 30 samples. A total of 47 workers participated in the study. The results indicate a strong preference for our results over the baseline methods across all evaluated metrics. + +
Pre-trainingMethodMAE ↓MSE ↓SSIM ↑LPIPS ↓
MAE [18]0.09520.02420.90072.096
MMAE0.09330.02350.90512.059
DiffuseNetAlbedo0.10530.02950.89852.197
Diff Render0.10230.02750.90022.137
+ +Table 3. Ablation Studies on the OLAT test set. + +![](images/175df5cf85a16b03ce2ceaf0fbc83572d83762d5e469848d15a98a9a8d44c001.jpg) +Figure 8. Ablation on DiffuseNet. Our approach successfully infers the albedo on various surfaces (skin, teeth, and accessories). + +Ablation Studies. We analyze our two major design choices in Table. 3: the MMAE pre-training framework and DiffuseNet. The MMAE, which integrates dynamic masking with generative objectives, outperforms MAE. This superiority is mainly due to the incorporation of challenging masks and global coherence objectives, enabling the model to learn richer features during pre-training. Furthermore, our method of predicting diffuse render demonstrates superiority over direct albedo prediction. Firstly, we see it simplifies the learning process, as predicting diffuse render is more closely related to the original image. Secondly, our approach effectively distinguishes between the influences of illumination (diffuse shading) and surface properties (diffuse render). This distinction is crucial for accurately modeling the intrinsic color of surfaces, as it enables independent and precise evaluation of each element (see Eqn. 9). In contrast, methods that predict albedo directly often struggle to differentiate between these factors, leading to significant inaccuracies in color constancy, as shown in Fig. 8. + +Applications. We present two applications using predicted intrinsics in Fig. 9. First, real-time PBR via Cook-Torrance + +![](images/bcae20feb2c5d47d383108bc8c144e7c5a968d4f0c475f664f06ab3158f89093.jpg) +(a) Real-Time PBR Rendering + +![](images/09ba99ded5c01df2b6056d56f86a4009f651521d91b532638eb6be62d3bd1203.jpg) +(b) Copy Light + +![](images/9ab99f19c87c386009f828589ac64c646f84a386eceb0fe0fdddcaa4467461f7.jpg) +Figure 9. Applications. We showcase additional features of SwitchLight, powered by the diverse intrinsics features. +Figure 10. Limitations. The model faces challenges in removing strong shadows, misinterpreting reflective surfaces like sunglasses, and inaccurately predicting albedo for face paint. + +components in Three.js graphics library. Second, switching the lighting environment between the source and reference images. Further details are in the supplementary video. + +Limitations. We identified a few failure cases in Fig. 10. First, our model struggles with neutralizing strong shadows, which leads to inaccurate facial geometry and residual shadow artifacts in both albedo and relit images. Incorporating shadow augmentation techniques [15, 53] during training could mitigate this issue. Second, the model incorrectly interprets reflective surfaces, such as sunglasses, as opaque objects in the normal image. This error prevents the model from properly removing reflective highlights in the albedo and relit images. Lastly, the model inaccurately predicts the albedo for face paint. Implementing a semantic mask [51] to distinguish different semantic regions separately from the skin could help resolving these issues. + +# 7. Conclusion + +We introduce SwitchLight, an architecture based on Cook-Torrance rendering physics, enhanced with a self-supervised pre-training framework. This co-designed approach significantly outperforms previous models. Our future plans include scaling the current model beyond images to encompass video and 3D data. We hope our proposal serve as a new foundational model for relighting tasks. + +# References + +[1] Pexels. https://www.pexels.com.6,7 +[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2 +[3] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2014. 2 +[4] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 2 +[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2 +[6] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22367-22377, 2023. 2 +[7] Zhaoxi Chen and Ziwei Liu. Relighting4d: Neural relightable human from videos. In European Conference on Computer Vision, pages 606-623. Springer, 2022. 2 +[8] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM Transactions on Graphics (ToG), 1(1):7-24, 1982. 2 +[9] Paul Debevec. Rendering synthetic objects into real scenes: Bridging traditional and image-based graphics with global illumination and high dynamic range photography. In Acm siggraph 2008 classes, pages 1-10. 2008. 2 +[10] Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. Acquiring the reflectance field of a human face. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 145-156, 2000. 2, 6 +[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2 +[12] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422-1430, 2015. 2 +[13] Julie Dorsey, James Arvo, and Donald Greenberg. Interactive design of complex time dependent lighting. IEEE Computer Graphics and Applications, 15(2):26-36, 1995. 2 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 6 +[15] David Futschik, Kelvin Ritland, James Vecore, Sean Fanello, Sergio Orts-Escalano, Brian Curless, Daniel Sykora, and Ro + +hit Pandey. Controllable light diffusion for portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8412-8421, 2023. 8 +[16] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. 2 +[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2 +[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 6, 8 +[19] Andrew Hou, Ze Zhang, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Towards high fidelity face relighting with realistic shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14719-14728, 2021. 2 +[20] Andrew Hou, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Face relighting with geometrically consistent shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4217-4226, 2022. 2 +[21] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 5, 6 +[22] Chaonan Ji, Tao Yu, Kaiwen Guo, Jingxin Liu, and Yebin Liu. Geometry-aware single-image full-body human relighting. In European Conference on Computer Vision, pages 388-405. Springer, 2022. 2 +[23] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5, 6 +[24] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6 +[25] Taehun Kim, Kunhee Kim, Joonyeong Lee, Dongmin Cha, Jiho Lee, and Daijin Kim. Revisiting image pyramid structure for high resolution salient object detection. In Proceedings of the Asian Conference on Computer Vision, pages 108-124, 2022. 4 +[26] Wenbo Li, Xin Lu, Shengju Qian, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer-based image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2 +[27] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. + +In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 4 +[28] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European conference on computer vision (ECCV), pages 85-100, 2018. 6 +[29] Yihao Liu, Jingwen He, Jinjin Gu, Xiangtao Kong, Yu Qiao, and Chao Dong. Degae: A new pretraining paradigm for low-level vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23292-23303, 2023. 2 +[30] Yiqun Mei, He Zhang, Xuaner Zhang, Jianming Zhang, Zhixin Shu, Yilin Wang, Zijun Wei, Shi Yan, HyunJoon Jung, and Vishal M Patel. Lightpainter: Interactive portrait relighting with freehand scribble. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 195-205, 2023. 2, 6 +[31] Thomas Nestmeyer, Jean-François Lalonde, Iain Matthews, and Andreas Lehrmann. Learning physics-guided face relighting under directional light. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5124-5133, 2020. 2 +[32] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 2 +[33] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 2, 5, 6, 7 +[34] Namuk Park, Wonjae Kim, Byeongho Heo, Taekyung Kim, and Sangdoo Yun. What do self-supervised vision transformers learn? arXiv preprint arXiv:2305.00729, 2023. 2 +[35] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 2 +[36] Bui Tuong Phong. Illumination for computer generated pictures. In Seminal graphics: pioneering efforts that shaped the field, pages 95-101. 1998. 2 +[37] Puntawat Ponglertnapakorn, Nontawat Tritrong, and Supasorn Suwajanakorn. Difareli: Diffusion face relighting. arXiv preprint arXiv:2304.09479, 2023. 2 +[38] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2 +[39] Soumyadip Sengupta, Angjoo Kanazawa, Carlos D Castillo, and David W Jacobs. Sfsnet: Learning shape, reflectance and illuminance of facesin the wild'. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6296-6305, 2018. 2 +[40] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Pro- + +ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2291-2300, 2020. 4 +[41] Soumyadip Sengupta, Brian Curless, Ira Kemelmacher-Shlizerman, and Steven M Seitz. A light stage on every desk. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2420-2429, 2021. 2 +[42] YiChang Shih, Sylvain Paris, Connelly Barnes, William T Freeman, and Fredo Durand. Style transfer for headshot portraits. 2014. 2 +[43] Zhixin Shu, Sunil Hadap, Eli Shechtman, Kalyan Sunkavalli, Sylvain Paris, and Dimitris Samaras. Portrait lighting transfer using a mass transport approach. ACM Transactions on Graphics (TOG), 36(4):1, 2017. 2 +[44] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 2, 6, 7 +[45] Piotr Teterwak, Aaron Sarna, Dilip Krishnan, Aaron Maschinot, David Belanger, Ce Liu, and William T Freeman. Boundless: Generative adversarial networks for image extension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10521-10530, 2019. 6 +[46] Yifan Wang, Aleksander Holynski, Xiuming Zhang, and Xuaner Zhang. Sunstage: Portrait reconstruction and relighting using the sun as a light stage. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20792-20802, 2023. 2 +[47] Zhibo Wang, Xin Yu, Ming Lu, Quan Wang, Chen Qian, and Feng Xu. Single image portrait relighting via explicit multiple reflectance channel modeling. ACM Transactions on Graphics (TOG), 39(6):1-13, 2020. 2 +[48] Andreas Wenger, Andrew Gardner, Chris Tchou, Jonas Unger, Tim Hawkins, and Paul Debevec. Performance relighting and reflectance transformation with timemultiplexed illumination. ACM Transactions on Graphics (TOG), 24(3):756-764, 2005. 2, 6 +[49] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133-16142, 2023. 2, 6 +[50] Robert J Woodham. Photometric method for determining surface orientation from multiple images. Optical engineering, 19(1):139-144, 1980. 6 +[51] Yu-Ying Yeh, Koki Nagano, Sameh Khamis, Jan Kautz, Ming-Yu Liu, and Ting-Chun Wang. Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. ACM Transactions on Graphics (TOG), 41(6): 1-21, 2022. 2, 6, 7, 8 +[52] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 649-666. Springer, 2016. 2 + +[53] Xuaner Zhang, Jonathan T Barron, Yun-Ta Tsai, Rohit Pandey, Xiuming Zhang, Ren Ng, and David E Jacobs. Portrait shadow manipulation. ACM Transactions on Graphics (TOG), 39(4):78-1, 2020. 8 +[54] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7194-7202, 2019. 2 +[55] Taotao Zhou, Kai He, Di Wu, Teng Xu, Qixuan Zhang, Kuixiang Shao, Wenzheng Chen, Lan Xu, and Jingyi Yu. Relightable neural human assets from multi-view gradient illuminations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4315-4327, 2023. 2 \ No newline at end of file diff --git a/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/images.zip b/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7268c5ae2b5d4670e6dce6c349f55a4935ec500e --- /dev/null +++ b/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b0d1aa4d8445c7bd479b30ca6878be0f7c99e33f3d2599d01290d12ce182869 +size 691332 diff --git a/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/layout.json b/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ad8f2b159b94a58635893445cb8493933ba675fc --- /dev/null +++ b/2024/SwitchLight_ Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting/layout.json @@ -0,0 +1,8927 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 103, + 479, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 103, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 479, + 140 + ], + "type": "text", + "content": "SwitchLight: Co-design of Physics-driven Architecture and Pre-training Framework for Human Portrait Relighting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 160, + 527, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 160, + 527, + 177 + ], + "spans": [ + { + "bbox": [ + 65, + 160, + 527, + 177 + ], + "type": "text", + "content": "Hoon Kim1 Minje Jang1 Wonjun Yoon1 Jisoo Lee1 Donghyun Na1 Sanghyun Woo2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 193, + 188, + 397, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 188, + 397, + 204 + ], + "spans": [ + { + "bbox": [ + 193, + 188, + 397, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 193, + 188, + 397, + 204 + ], + "type": "text", + "content": "Beeble AI " + }, + { + "bbox": [ + 193, + 188, + 397, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 193, + 188, + 397, + 204 + ], + "type": "text", + "content": "New York University" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 241, + 545, + 518 + ], + "blocks": [ + { + "bbox": [ + 76, + 228, + 124, + 240 + ], + "lines": [ + { + "bbox": [ + 76, + 228, + 124, + 240 + ], + "spans": [ + { + "bbox": [ + 76, + 228, + 124, + 240 + ], + "type": "text", + "content": "Input RGB" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 228, + 211, + 239 + ], + "lines": [ + { + "bbox": [ + 176, + 228, + 211, + 239 + ], + "spans": [ + { + "bbox": [ + 176, + 228, + 211, + 239 + ], + "type": "text", + "content": "Normal" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 280, + 228, + 313, + 239 + ], + "lines": [ + { + "bbox": [ + 280, + 228, + 313, + 239 + ], + "spans": [ + { + "bbox": [ + 280, + 228, + 313, + 239 + ], + "type": "text", + "content": "Albedo" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 371, + 228, + 420, + 240 + ], + "lines": [ + { + "bbox": [ + 371, + 228, + 420, + 240 + ], + "spans": [ + { + "bbox": [ + 371, + 228, + 420, + 240 + ], + "type": "text", + "content": "Roughness" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 228, + 520, + 240 + ], + "lines": [ + { + "bbox": [ + 469, + 228, + 520, + 240 + ], + "spans": [ + { + "bbox": [ + 469, + 228, + 520, + 240 + ], + "type": "text", + "content": "Reflectivity" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 241, + 545, + 518 + ], + "lines": [ + { + "bbox": [ + 50, + 241, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 50, + 241, + 545, + 518 + ], + "type": "image", + "image_path": "c3a9957805d3c58bc988053a24741f13d3fa4a2761520550a6037eb132557f0e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 520, + 547, + 543 + ], + "lines": [ + { + "bbox": [ + 46, + 520, + 547, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 547, + 543 + ], + "type": "text", + "content": "Figure 1. Be Anywhere at Any Time. SwitchLight processes a human portrait by decomposing it into detailed intrinsic components, and re-renders the image under a designated target illumination, ensuring a seamless composition of the subject into any new environment." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 552, + 192, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 552, + 192, + 563 + ], + "spans": [ + { + "bbox": [ + 143, + 552, + 192, + 563 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 567, + 288, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 288, + 686 + ], + "type": "text", + "content": "We introduce a co-designed approach for human portrait relighting that combines a physics-guided architecture with a pre-training framework. Drawing on the Cook-Torrance reflectance model, we have meticulously configured the architecture design to precisely simulate light-surface interactions. Furthermore, to overcome the limitation of scarce high-quality lightstage data, we have developed a self-supervised pre-training strategy. This novel combination of accurate physical modeling and expanded training dataset establishes a new benchmark in relighting realism." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 550, + 386, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 550, + 386, + 562 + ], + "spans": [ + { + "bbox": [ + 307, + 550, + 386, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "content": "Relighting is more than an aesthetic tool; it unlocks infinite narrative possibilities and enables seamless integration of subjects into diverse environments (see Fig. 1). This advancement resonates with our innate desire to transcend the physical constraints of space and time, while also providing tangible solutions to practical challenges in digital content creation. It is particularly transformative in virtual (VR) and augmented reality (AR) applications, where relighting facilitates the real-time adaptation of lighting, ensuring that users and digital elements coexist naturally within any environment, offering a next level of telepresence." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "In this work, we focus on human portrait relighting." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 204, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 204, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 204, + 703 + ], + "type": "text", + "content": "* All authors contributed equally to this work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 59, + 703, + 134, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 134, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 134, + 712 + ], + "type": "text", + "content": "\\*https://www.beeble.ai" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25096" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 179 + ], + "type": "text", + "content": "While the relighting task fundamentally demands an in-depth understanding of geometry, material properties, and illumination, the challenge is more compounded when addressing human subjects, due to the unique characteristics of skin surfaces as well as the diverse textures and reflectance properties of a wide array of clothing, hairstyles, and accessories. These elements interact in complex ways, necessitating advanced algorithms capable of simulating the subtle interplay of light with these varied surfaces." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 186, + 288, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 288, + 378 + ], + "type": "text", + "content": "Currently, the most promising approach involves the use of deep neural networks trained on pairs of high-quality relict portrait images and their corresponding intrinsic attributes, which are sourced from a light stage setup [10]. Initial efforts approached the relighting process as a 'black box' [44, 47], without delving into the underlying mechanisms. Later advancements adopted a physics-guided model design, incorporating the explicit modeling of image intrinsics and image formation physics [31]. Pandey et al. [33] proposed the Total Relight (TR) architecture, also physics-guided, which decomposes an input image into surface normals and albedo maps, and performs relighting based on the Phong specular reflectance model. The TR architecture has become foundational model for image relighting, with most recent and advanced architectures building upon its principle [22, 30, 51]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 384, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 612 + ], + "type": "text", + "content": "Following the physics-guided approach, our contribution lies in a co-design of architecture with a self-supervised pre-training framework. First, our architecture evolves towards a more accurate physical model by integrating the Cook-Torrance specular reflectance model [8], representing a notable advancement from the empirical Phong specular model [36] employed in the Total Relight architecture. The Cook-Torrance model adeptly simulates light interactions with surface microfacets, accounting for spatially varying roughness and reflectivity. Secondly, our pretraining framework scales the learning process beyond the typically hard-to-obtain lightstage data. By revisiting the masked autoencoder (MAE) framework [18], we adept it for the task of relighting. These modifications are crafted to address the unique challenges posed by this task, enabling our model to learn from unlabelled data and refine its ability to produce realistic relit portraits during fine-tuning. To the best of our knowledge, this is the first time applying self-supervised pre-training specifically to the relighting task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "To summarize, our contribution is twofold. Firstly, by enhancing the physical reflectance model, we have introduced a new level of realism in the output. Secondly, by adopting self-supervised learning, we have expanded the scale of the training data and enhanced the expression of lighting in diverse real-world scenarios. Collectively, these advancements have led SwitchLight framework to achieve a new state-of-the-art in human portrait relighting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 96, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 545, + 396 + ], + "type": "text", + "content": "Human Portrait Relighting is an ill-posed problem due to its under-constrained nature. To tackle this, earlier methods incorporated 3D facial priors [43], exploited image intrinsics [3, 39], or framed the task as a style transfer [42]. Light stage techniques [48] offer a more powerful solution by recording subject's reflectance fields under varying lighting conditions [10, 13], though they are labor-intensive and require specialized equipment. A promising alternative has emerged with deep learning, utilizing neural networks trained on light stage data. Sun et al. [44] pioneered this approach, but their method had limitations in representing non-Lambertian effects. This was improved upon by Nestmeyer et al. [31], who integrated rendering physics into network design, albeit limited to directional light. Building upon this, Pandey et al. [33] incorporated the Phong reflection model and a high dynamic range (HDR) lighting map [9] into their network, enabling a more accurate representation of global illumination. Simultaneously, efforts have been made to explore portrait relighting without light stage data [19, 20, 41, 46, 54]. Moreover, introduction of NeRF [7] and diffusion-based [37] models has opened new avenues in the field. However, networks trained with light-stage data maintain superior accuracy and realism, thanks to physics-based composited relight image training pairs and precise ground truth image intrinsics [55]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 399, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 449 + ], + "type": "text", + "content": "Our work furthers this domain by integrating the Cook-Torrance model into our network design, shifting from the empirical Phong model to a more physics-based approach, thereby enhancing the realism and detail in relit images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 457, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 613 + ], + "type": "text", + "content": "Self-supervised Pre-training has become a standard training scheme in the development of large language models like BERT [11] and GPT [38], and is increasingly influential in vision models, aiming to replicate the 'BERT moment'. This approach typically involves pre-training on extensive unlabeled data, followed by fine-tuning on specific tasks. While early efforts in vision models focused on simple pretext tasks [12, 16, 32, 35, 52], the field has evolved through stages like contrastive learning [5, 17] and masked image modeling [2, 18, 49]. However, the primary focus has remained on visual recognition, with less attention to other domains. Exceptions include low-level image processing tasks [4, 6, 26, 29] using the vision transformer [14]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 618, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 715 + ], + "type": "text", + "content": "Our research takes a different route, focusing on human portrait relighting—a complex challenge of manipulating illumination in the image. This direction is crucial because acquiring accurate ground truth data, especially from light stage, is both expensive and difficult. We modify the MAE framework [18], previously successful in robust image representation learning and developing locality biases [34], to suit the unique requirements of effective relighting." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "25097" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 126, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 126, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 126, + 85 + ], + "type": "text", + "content": "3. SwitchLight" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 163 + ], + "type": "text", + "content": "We introduce SwitchLight, a state-of-the-art framework for human portrait relighting, with its architectural overview presented in Fig. 2. We first provide foundational concepts in Sec. 3.1, and define the problem in Sec. 3.2. This is followed by detailing the architecture in Sec. 3.3, and lastly, we describe the loss functions used in Sec. 3.4." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 170, + 133, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 170, + 133, + 181 + ], + "spans": [ + { + "bbox": [ + 47, + 170, + 133, + 181 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "text", + "content": "In this section, vectors " + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{l}" + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "text", + "content": " are denoted as unit vectors. Specifically, " + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "text", + "content": " represents the surface normal, " + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "text", + "content": " is the view direction, " + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{l}" + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "text", + "content": " is the incident light direction, and " + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "text", + "content": " is the half-vector computed from " + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{l}" + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "text", + "content": ". The dot product is clamped between [0..1], indicated by " + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "inline_equation", + "content": "\\langle \\cdot \\rangle" + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 262, + 287, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 310 + ], + "type": "text", + "content": "Image Rendering. The primary goal of image rendering is to create a visual representation that accurately simulates the interactions between light and surfaces. These complex interactions are encapsulated by the rendering equation:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 317, + 287, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 317, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 97, + 317, + 287, + 342 + ], + "type": "interline_equation", + "content": "L _ {o} (\\mathbf {v}) = \\int_ {\\Omega} f (\\mathbf {v}, \\mathbf {l}) L _ {i} (\\mathbf {l}) \\langle \\mathbf {n} \\cdot \\mathbf {l} \\rangle d \\mathbf {l} \\tag {1}", + "image_path": "ae85cfda43a94ed240149c8d82012c49ad702e2803364aeaf344d2887977abdb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "inline_equation", + "content": "L_{o}(\\mathbf{v})" + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "text", + "content": " denotes the radiance, or the light intensity perceived by the observer in direction " + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "text", + "content": ". It is the cumulative result of incident lights " + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "inline_equation", + "content": "L_{i}(\\mathbf{l})" + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "text", + "content": " from all possible directions over the hemisphere, " + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "text", + "content": ", centered around the surface normal, denoted as " + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "text", + "content": ". Central to this equation lies the Bidirectional Reflectance Distribution Function (BRDF), denoted as " + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "inline_equation", + "content": "f(\\mathbf{v},\\mathbf{l})" + }, + { + "bbox": [ + 46, + 348, + 287, + 432 + ], + "type": "text", + "content": ", describing the surface's reflection characteristics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 446, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 287, + 495 + ], + "type": "text", + "content": "BRDF Composition. The BRDF, represented by " + }, + { + "bbox": [ + 46, + 446, + 287, + 495 + ], + "type": "inline_equation", + "content": "f(\\mathbf{v},\\mathbf{l})" + }, + { + "bbox": [ + 46, + 446, + 287, + 495 + ], + "type": "text", + "content": ", describes how light is reflected at an opaque surface. It is composed of two major components: diffuse reflection (" + }, + { + "bbox": [ + 46, + 446, + 287, + 495 + ], + "type": "inline_equation", + "content": "f_{d}" + }, + { + "bbox": [ + 46, + 446, + 287, + 495 + ], + "type": "text", + "content": ") and specular reflection (" + }, + { + "bbox": [ + 46, + 446, + 287, + 495 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 46, + 446, + 287, + 495 + ], + "type": "text", + "content": "):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 502, + 287, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 502, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 107, + 502, + 287, + 515 + ], + "type": "interline_equation", + "content": "f (\\mathbf {v}, \\mathbf {l}) = f _ {d} (\\mathbf {v}, \\mathbf {l}) + f _ {s} (\\mathbf {v}, \\mathbf {l}) \\tag {2}", + "image_path": "e471012007bec7b8c902936f15f0af76fcb87e79ff95a42e5f90fbb44dbfc3fe.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 594 + ], + "type": "text", + "content": "A surface intrinsically exhibits both diffuse and specular reflections. The diffuse component uniformly scatters light, ensuring consistent illumination regardless of the viewing angle. In contrast, the specular component is viewing angle-dependent, producing shiny highlights that are crucial for achieving a photorealistic effect." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 609, + 287, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 656 + ], + "type": "text", + "content": "Lambertian Diffuse Reflectance. Lambertian reflectance, a standard model for diffuse reflection, describes a uniform light scatter irrespective of the viewing angle. This ensures a consistent appearance in brightness:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 121, + 662, + 287, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 662, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 121, + 662, + 287, + 684 + ], + "type": "interline_equation", + "content": "f _ {d} (\\mathbf {v}, \\mathbf {l}) = \\frac {\\sigma}{\\pi} [ \\text {c o n s t .} ] \\tag {3}", + "image_path": "8cae3f6e620f877d8335baa7bbd2ef3d28ce64ef2a45852477566b8c394b1e14.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": " is the albedo, indicating the intrinsic color and brightness of the surface." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "Cook-Torrance Specular Reflectance. The Cook-Torrance model, based on microfacet theory, represents surfaces as a myriad of tiny, mirror-like facets. It incorporates a roughness parameter " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": ", which allows precise rendering of surface specular reflectance:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 340, + 139, + 545, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 139, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 340, + 139, + 545, + 167 + ], + "type": "interline_equation", + "content": "f _ {s} (\\mathbf {v}, \\mathbf {l}) = \\frac {D (\\mathbf {h} , \\alpha) G (\\mathbf {v} , \\mathbf {l} , \\alpha) F (\\mathbf {v} , \\mathbf {h} , f _ {0})}{4 \\langle \\mathbf {n} \\cdot \\mathbf {l} \\rangle \\langle \\mathbf {n} \\cdot \\mathbf {v} \\rangle} \\tag {4}", + "image_path": "a4623dbfce8a999ecd95742f825528b54a48fa73105074eb634b0ca9ae799092.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "text", + "content": "In this model, " + }, + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "text", + "content": " is the microfacet distribution function, describing the orientation of the microfacets relative to the half-vector " + }, + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "text", + "content": " is the geometric attenuation factor, accounting for the shadowing and masking of microfacets, and " + }, + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "text", + "content": " is the Fresnel term, calculating the reflectance variation depending on the viewing angle, where " + }, + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "inline_equation", + "content": "f_{0}" + }, + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "text", + "content": " is the surface Fresnel reflectivity at normal incidence. A lower " + }, + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "text", + "content": " value implies a smoother surface with sharper specular highlights, whereas a higher " + }, + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "text", + "content": " value indicates a rougher surface, resulting in more diffused reflections. By adjusting " + }, + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 173, + 545, + 305 + ], + "type": "text", + "content": ", the Cook-Torrance model can depict a range of specular reflections." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 319, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 545, + 354 + ], + "type": "text", + "content": "Image Formation. Upon the base rendering equation, we include the diffuse and specular components of the BRDF and derive a unified formula:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 329, + 361, + 545, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 361, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 329, + 361, + 545, + 387 + ], + "type": "interline_equation", + "content": "L _ {o} (\\mathbf {v}) = \\int_ {\\Omega} \\left(f _ {d} (\\mathbf {v}, \\mathbf {l}) + f _ {s} (\\mathbf {v}, \\mathbf {l})\\right) E (\\mathbf {l}) \\langle \\mathbf {n} \\cdot \\mathbf {l} \\rangle d \\mathbf {l} \\tag {5}", + "image_path": "e8009bafff23144856f9c9011deb4359e1b31d0562e70b57a5f873f5d646d945.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 395, + 545, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 395, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 305, + 395, + 545, + 455 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 395, + 545, + 455 + ], + "type": "inline_equation", + "content": "E(1)" + }, + { + "bbox": [ + 305, + 395, + 545, + 455 + ], + "type": "text", + "content": " denotes the incident environmental lighting. This formula represents the core principle that an image is a product of interplay between the BRDF and lighting. To further clarify this concept, we introduce a rendering function " + }, + { + "bbox": [ + 305, + 395, + 545, + 455 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 395, + 545, + 455 + ], + "type": "text", + "content": ", succinctly modeling the process of image formation:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 370, + 465, + 545, + 494 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 465, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 370, + 465, + 545, + 494 + ], + "type": "interline_equation", + "content": "I = R (\\underbrace {\\mathbf {n} , \\sigma , \\alpha , f _ {0}} _ {\\text {s u r f a c e a t t r i b u t e s l i g h t i n g}}, \\underbrace {E} _ {\\text {l i g h t i n g}}) \\tag {6}", + "image_path": "cc436ef8b819cecfd300d1c8dc60bd70c131086ae3758653b0efd963a4427c23.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 498, + 545, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 545, + 546 + ], + "type": "text", + "content": "It is important to note that since the BRDF is a function of surface properties, as detailed in Eqn. 3 and 4, we can now clearly understand that image formation is essentially governed by the interaction of surface attributes and lighting." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 553, + 429, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 553, + 429, + 565 + ], + "spans": [ + { + "bbox": [ + 306, + 553, + 429, + 565 + ], + "type": "text", + "content": "3.2. Problem Formulation" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 571, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 571, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 305, + 571, + 545, + 643 + ], + "type": "text", + "content": "Image Relighting. Given the image formation model above, our goal is to manipulate the lighting of an existing image. This involves two main steps: inverse rendering and re-rendering under target illumination, both driven by neural networks. For a given source image " + }, + { + "bbox": [ + 305, + 571, + 545, + 643 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{src}}" + }, + { + "bbox": [ + 305, + 571, + 545, + 643 + ], + "type": "text", + "content": " and target illumination " + }, + { + "bbox": [ + 305, + 571, + 545, + 643 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{tgt}}" + }, + { + "bbox": [ + 305, + 571, + 545, + 643 + ], + "type": "text", + "content": ", the process is delineated as:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 348, + 653, + 542, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 653, + 542, + 666 + ], + "spans": [ + { + "bbox": [ + 348, + 653, + 542, + 666 + ], + "type": "interline_equation", + "content": "\\text {I n v e r s e} (\\mathbf {n}, \\sigma , \\alpha , f _ {0}, E _ {\\mathrm {s r c}}) = U (I _ {\\mathrm {s r c}})", + "image_path": "87addaae50c67689a2242c86d79a8e75b364e39dc05f9c0966e1fde20a2ed2ec.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 667, + 533, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 667, + 533, + 681 + ], + "spans": [ + { + "bbox": [ + 307, + 667, + 533, + 681 + ], + "type": "text", + "content": "Rendering with Target Light. " + }, + { + "bbox": [ + 307, + 667, + 533, + 681 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{tgt}} = R(\\mathbf{n},\\sigma ,\\alpha ,f_0,E_{\\mathrm{tgt}})" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "During the inverse rendering step, the function " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " unravels the intrinsic properties of " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{src}}" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": ". In the subsequent relighting" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25098" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 545, + 205 + ], + "type": "image", + "image_path": "3aef3794bddc2d79a3b52b533ce87d0ccbbbcaa0c5c7d51a523e59572face98c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 212, + 546, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 546, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 546, + 246 + ], + "type": "text", + "content": "Figure 2. SwitchLight Architecture. The input source image is decomposed into normal map, lighting, diffuse and specular components. Given these intrinsic, images are re-rendered under target lighting. The architecture integrates the Cook-Torrance reflection model; the final output combines physically-based predictions with neural network enhancements for realistic portrait relighting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 260, + 288, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 260, + 288, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 288, + 298 + ], + "type": "text", + "content": "step, the derived intrinsic properties along with new illumination " + }, + { + "bbox": [ + 46, + 260, + 288, + 298 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{tgt}}" + }, + { + "bbox": [ + 46, + 260, + 288, + 298 + ], + "type": "text", + "content": " are employed by the rendering function " + }, + { + "bbox": [ + 46, + 260, + 288, + 298 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 260, + 288, + 298 + ], + "type": "text", + "content": " to generate the target relit image " + }, + { + "bbox": [ + 46, + 260, + 288, + 298 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{tgt}}" + }, + { + "bbox": [ + 46, + 260, + 288, + 298 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 304, + 129, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 129, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 129, + 315 + ], + "type": "text", + "content": "3.3. Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 322, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 287, + 467 + ], + "type": "text", + "content": "Our architecture systematically executes the two primary stages outlined in our problem formulation. The first stage involves extracting intrinsic properties from the source image " + }, + { + "bbox": [ + 46, + 322, + 287, + 467 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{src}}" + }, + { + "bbox": [ + 46, + 322, + 287, + 467 + ], + "type": "text", + "content": ". For this purpose, we employ a matting network [25, 27, 40] to accurately separate the foreground. This extracted image is then processed by our inverse rendering network " + }, + { + "bbox": [ + 46, + 322, + 287, + 467 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 46, + 322, + 287, + 467 + ], + "type": "text", + "content": ", which infers normal, albedo, roughness, reflectivity, and source lighting. Subsequently, the second stage involves re-rendering the image under new target lighting conditions. To achieve this, the acquired intrinsics, along with the target lighting " + }, + { + "bbox": [ + 46, + 322, + 287, + 467 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{tgt}}" + }, + { + "bbox": [ + 46, + 322, + 287, + 467 + ], + "type": "text", + "content": ", are fed into our relighting network " + }, + { + "bbox": [ + 46, + 322, + 287, + 467 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 322, + 287, + 467 + ], + "type": "text", + "content": ", producing the relit image " + }, + { + "bbox": [ + 46, + 322, + 287, + 467 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{tgt}}" + }, + { + "bbox": [ + 46, + 322, + 287, + 467 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 479, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 529 + ], + "type": "text", + "content": "Normal Net. The network takes the source image " + }, + { + "bbox": [ + 46, + 479, + 287, + 529 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{src}} \\in \\mathbb{R}^{\\mathrm{H} \\times \\mathrm{W} \\times 3}" + }, + { + "bbox": [ + 46, + 479, + 287, + 529 + ], + "type": "text", + "content": " and generates a normal map " + }, + { + "bbox": [ + 46, + 479, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{N}}" + }, + { + "bbox": [ + 46, + 479, + 287, + 529 + ], + "type": "text", + "content": ". Each pixel in this map contains a unit normal vector " + }, + { + "bbox": [ + 46, + 479, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}" + }, + { + "bbox": [ + 46, + 479, + 287, + 529 + ], + "type": "text", + "content": ", indicating the orientation of the corresponding surface point." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 542, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 287, + 578 + ], + "type": "text", + "content": "Illum Net. The network infers the lighting conditions in the given image captured in an HDRI format. Specifically, it computes the convolved HDRIs:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 586, + 287, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 586, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 102, + 586, + 287, + 620 + ], + "type": "interline_equation", + "content": "E _ {\\mathrm {s r c}} ^ {p} \\left(\\mathbf {l} ^ {\\prime}\\right) = \\int_ {\\Omega} \\underbrace {E _ {\\mathrm {s r c}} (\\mathbf {l})} _ {\\text {H D R I}} \\underbrace {\\langle \\mathbf {l} ^ {\\prime} \\cdot \\mathbf {l} \\rangle^ {p}} _ {\\text {P h o n g l o b e}} d \\mathbf {l} \\tag {7}", + "image_path": "c72e0f216ce5cbae57aaa3f8aa6e4748b8f4e262be5db5cc37f2468924380be8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "In this equation, " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{src}} \\in \\mathbb{R}^{\\mathrm{H}_{\\mathrm{HDRI}} \\times \\mathrm{W}_{\\mathrm{HDRI}} \\times 3}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " is the original source HDRI map, with 1 indicating spherical directions in the HDRI space " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{\\mathrm{H}_{\\mathrm{HDRI}} \\times \\mathrm{W}_{\\mathrm{HDRI}}}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": ". The term " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\langle \\mathbf{l}' \\cdot \\mathbf{l} \\rangle^{p}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " represents the Phong reflectance lobe with shininess exponents " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "p \\in \\{1, 16, 32, 64\\}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": ", which incorporates various specular terms. Consequently, it is expressed in a multi-dimensional tensor form as " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{4 \\times \\mathrm{H}_{\\mathrm{HDRI}} \\times \\mathrm{W}_{\\mathrm{HDRI}} \\times \\mathrm{H}_{\\mathrm{HDRI}} \\times \\mathrm{W}_{\\mathrm{HDRI}}}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{src}}^{p} \\in" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 259, + 546, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 546, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 546, + 308 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{4\\times \\mathrm{H_{cHDRI}}\\times \\mathrm{W_{cHDRI}}\\times 3}" + }, + { + "bbox": [ + 304, + 259, + 546, + 308 + ], + "type": "text", + "content": " is the convolved HDRI. In this work, we set the resolution of HDRI and convolved HDRI at " + }, + { + "bbox": [ + 304, + 259, + 546, + 308 + ], + "type": "inline_equation", + "content": "32\\times 64" + }, + { + "bbox": [ + 304, + 259, + 546, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 259, + 546, + 308 + ], + "type": "inline_equation", + "content": "64\\times 128" + }, + { + "bbox": [ + 304, + 259, + 546, + 308 + ], + "type": "text", + "content": ", respectively, and we apply convolution on light source coordinates." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 308, + 546, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 546, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 546, + 429 + ], + "type": "text", + "content": "The network employs a cross-attention mechanism at its core, where predefined Phong reflectance lobes serve as queries, and the original image acts as both keys and values. Within this setup, the convolved HDRI maps are synthesized by integrating image information into the Phong reflectance lobe representation. Specifically, our model utilizes bottleneck features from the Normal Net as a compact image representation. Our approach simplifies the complex task of HDRI reconstruction by instead focusing on estimating interactions with known surface reflective properties." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 441, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 491 + ], + "type": "text", + "content": "Diffuse Net. Estimating albedo is challenging due to the ambiguities in surface color and material properties, further complicated by shadow effects. To address this, we prioritize the inference of source diffuse render, " + }, + { + "bbox": [ + 304, + 441, + 545, + 491 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{src,diff}}" + }, + { + "bbox": [ + 304, + 441, + 545, + 491 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 338, + 494, + 545, + 535 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 494, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 338, + 494, + 545, + 535 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {s r c}, o _ {\\mathrm {d i f f}}} (\\mathbf {v}) = \\underbrace {\\frac {\\sigma}{\\pi}} _ {\\text {d i f f u s e B R D F}} \\underbrace {\\int_ {\\Omega} E _ {\\mathrm {s r c}} (\\mathbf {l}) \\left\\langle \\mathbf {n} \\cdot \\mathbf {l} \\right\\rangle d \\mathbf {l}} _ {\\text {d i f f u s e s h a d i n g}} \\tag {8}", + "image_path": "056d1def0739a669ad3392aa438fa15e02a44dcc1dcfb6351a1ecd61f2f713fd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 539, + 546, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 546, + 682 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 546, + 682 + ], + "type": "text", + "content": "Our key insight is that the diffuse render closely resembles the original image, which simplifies the model learning process. It captures surface color after removing specular reflections, such as shine or gloss, contrasting with albedo that represents the true surface color unaffected by lighting and shadows. The network takes a source image " + }, + { + "bbox": [ + 304, + 539, + 546, + 682 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{src}}" + }, + { + "bbox": [ + 304, + 539, + 546, + 682 + ], + "type": "text", + "content": ", concatenated with its diffuse shading, to produce the diffuse render. As in Eqn. 7, the diffuse shading, " + }, + { + "bbox": [ + 304, + 539, + 546, + 682 + ], + "type": "inline_equation", + "content": "\\hat{E}_{\\mathrm{src}}^{1}(\\hat{\\mathbf{n}})" + }, + { + "bbox": [ + 304, + 539, + 546, + 682 + ], + "type": "text", + "content": ", is derived using the predicted normals, " + }, + { + "bbox": [ + 304, + 539, + 546, + 682 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}" + }, + { + "bbox": [ + 304, + 539, + 546, + 682 + ], + "type": "text", + "content": ", and the predicted lighting map, " + }, + { + "bbox": [ + 304, + 539, + 546, + 682 + ], + "type": "inline_equation", + "content": "\\hat{E}_{\\mathrm{src}}^{1}" + }, + { + "bbox": [ + 304, + 539, + 546, + 682 + ], + "type": "text", + "content": ", with a Phong exponent of 1 for the diffuse term. The albedo map " + }, + { + "bbox": [ + 304, + 539, + 546, + 682 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}" + }, + { + "bbox": [ + 304, + 539, + 546, + 682 + ], + "type": "text", + "content": " is then computed by dividing the predicted diffuse render by its diffuse shading:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 391, + 687, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 687, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 391, + 687, + 545, + 717 + ], + "type": "interline_equation", + "content": "\\frac {\\hat {\\sigma}}{\\pi} = \\frac {\\hat {L} _ {\\mathrm {s r c} , o _ {\\mathrm {d i f f}}} (\\mathbf {v})}{\\hat {E} _ {\\mathrm {s r c}} ^ {1} (\\hat {\\mathbf {n}})} \\tag {9}", + "image_path": "33a6be0d8cee7e3c550568a206eb71858f5f2b120b4fb161935578513785beb3.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25099" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 289, + 178 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 289, + 178 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 289, + 178 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 289, + 178 + ], + "type": "image", + "image_path": "6bb454551f13b8fc93664e680abe94d1c2f018ff65f717cb84723e3772b4fe3e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 184, + 287, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 184, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 184, + 287, + 228 + ], + "type": "text", + "content": "Figure 3. Render Net Overview. Utilizing extracted image intrinsics, it employs the Cook-Torrance model for initial relighting and a neural network for enhanced refinement, producing high-fidelity relit images through a synergistic computational approach." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 240, + 287, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 240, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 287, + 264 + ], + "type": "text", + "content": "We have empirically validated that it significantly enhances albedo prediction across a range of real-world scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 279, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 327 + ], + "type": "text", + "content": "Specular Net. The network infers surface attributes associated with the Cook-Torrance specular elements, specifically, the roughness " + }, + { + "bbox": [ + 46, + 279, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 279, + 287, + 327 + ], + "type": "text", + "content": " and Fresnel reflectivity " + }, + { + "bbox": [ + 46, + 279, + 287, + 327 + ], + "type": "inline_equation", + "content": "f_{0}" + }, + { + "bbox": [ + 46, + 279, + 287, + 327 + ], + "type": "text", + "content": ". It uses a source image, predicted normal, and albedo maps as inputs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 341, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 437 + ], + "type": "text", + "content": "Render Net. The network utilizes extracted intrinsic surface attributes to produce the target relit images. It generates two types of relit images, as shown in Fig. 3. The first type adheres to the physically-based rendering (PBR) principles of the Cook-Torrance model. This involves computing diffuse and specular renders under the target illumination using Eqn. 3 and Eqn. 4. These renders are combined to form the PBR render, " + }, + { + "bbox": [ + 46, + 341, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\hat{I}_{\\mathrm{tgt}}^{\\mathrm{PBR}}" + }, + { + "bbox": [ + 46, + 341, + 287, + 437 + ], + "type": "text", + "content": ", as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 446, + 287, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 446, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 91, + 446, + 287, + 462 + ], + "type": "interline_equation", + "content": "\\hat {L} _ {\\mathrm {t g t}, o} ^ {\\mathrm {P B R}} (\\mathbf {v}) = \\hat {L} _ {\\mathrm {t g t}, o _ {\\text {d i f f}}} (\\mathbf {v}) + \\hat {L} _ {\\mathrm {t g t}, o _ {\\text {s p e c}}} (\\mathbf {v}) \\tag {10}", + "image_path": "c07a67c49067200cf92aad58fbb2b646f218fae73cdf157c992763a4078a08b8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 469, + 287, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 565 + ], + "type": "text", + "content": "The second type of relit image is the result of a neural network process. This enhances the PBR render, capturing finer details that the Cook-Torrance model might miss. It employs the albedo, along with the diffuse and specular renders from the Cook-Torrance model, to infer a more refined target relit image, termed the neural render, " + }, + { + "bbox": [ + 46, + 469, + 287, + 565 + ], + "type": "inline_equation", + "content": "\\hat{I}_{\\mathrm{tgt}}^{\\mathrm{Neural}}" + }, + { + "bbox": [ + 46, + 469, + 287, + 565 + ], + "type": "text", + "content": ". The qualitative improvements achieved through this neural enhancement are illustrated in Fig. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 572, + 118, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 572, + 118, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 572, + 118, + 585 + ], + "type": "text", + "content": "3.4. Objectives" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "content": "We supervise both intrinsic image attributes and relit images using their corresponding ground truths, obtained from the lightstage. We employ a combination of reconstruction, perceptual [23], adversarial [21], and specular [33] losses." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": "Reconstruction Loss " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\ell_1(\\mathbf{X},\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": ". It measures the pixel-level differences between the ground truth " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": " and its prediction " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": ". This loss is applied across different attributes, including normal map " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{N}}" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": ", convolved source HDRI maps " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{E}_{\\mathrm{src}}^p" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": ", source diffuse render " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{I}_{\\mathrm{src,diff}}" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": ", albedo map " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": ", and both" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 545, + 284 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 284 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 284 + ], + "type": "image", + "image_path": "38f6fd31e8f47999f2ffdde60969933648706529164eba58e1d6b671effee889.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 289, + 545, + 344 + ], + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 344 + ], + "type": "text", + "content": "Figure 4. Neural Render Enhancement. Using the Cook-Torrance model, diffuse and specular renders are computed, which are then composited into a physically-based rendering. Subsequently, a neural network enhances this PBR render, improving aspects such as brightness and specular details." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 354, + 545, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 391 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 391 + ], + "type": "text", + "content": "types of target relit images " + }, + { + "bbox": [ + 304, + 354, + 545, + 391 + ], + "type": "inline_equation", + "content": "\\hat{I}_{\\mathrm{tgt}}^{\\mathrm{PBR}}" + }, + { + "bbox": [ + 304, + 354, + 545, + 391 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 354, + 545, + 391 + ], + "type": "inline_equation", + "content": "\\hat{I}_{\\mathrm{tgt}}^{\\mathrm{Neural}}" + }, + { + "bbox": [ + 304, + 354, + 545, + 391 + ], + "type": "text", + "content": ". The attributes like the roughness " + }, + { + "bbox": [ + 304, + 354, + 545, + 391 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 354, + 545, + 391 + ], + "type": "text", + "content": " and Fresnel reflectivity " + }, + { + "bbox": [ + 304, + 354, + 545, + 391 + ], + "type": "inline_equation", + "content": "f_{0}" + }, + { + "bbox": [ + 304, + 354, + 545, + 391 + ], + "type": "text", + "content": " are also implicitly learned when this loss applied to the PBR render." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 404, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 545, + 453 + ], + "type": "text", + "content": "Perceptual Loss " + }, + { + "bbox": [ + 304, + 404, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vgg}} = \\ell_2(\\mathbf{VGG}(\\mathbf{X}),\\mathbf{VGG}(\\hat{\\mathbf{X}}))" + }, + { + "bbox": [ + 304, + 404, + 545, + 453 + ], + "type": "text", + "content": " . It captures high-level feature differences based on a VGGnetwork feature comparison. We apply this loss to the source diffuse render, albedo, and target relit images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 466, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 545, + 526 + ], + "type": "text", + "content": "Adversarial Loss " + }, + { + "bbox": [ + 304, + 466, + 545, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{adv}} = \\mathrm{GAN}(\\mathbf{X},\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 304, + 466, + 545, + 526 + ], + "type": "text", + "content": ". It promotes realism in the generated images by fooling a discriminator network. This loss is applied to the target relit images. We employ a PatchGAN architecture, with detailed specifications provided in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 540, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 613 + ], + "type": "text", + "content": "Specular Loss " + }, + { + "bbox": [ + 304, + 540, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{spec}} = \\ell_1(\\mathbf{X} \\odot \\hat{\\mathbf{S}}, \\hat{\\mathbf{X}} \\odot \\hat{\\mathbf{S}})" + }, + { + "bbox": [ + 304, + 540, + 545, + 613 + ], + "type": "text", + "content": ". It enhances the specular highlights in the relit images. Specifically, we utilize the predicted specular render " + }, + { + "bbox": [ + 304, + 540, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}} := \\hat{I}_{\\mathrm{tgt},\\mathrm{spec}}^{\\mathrm{PBR}}" + }, + { + "bbox": [ + 304, + 540, + 545, + 613 + ], + "type": "text", + "content": " derived from the Cook-Torrance physical model, to weigh the " + }, + { + "bbox": [ + 304, + 540, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 304, + 540, + 545, + 613 + ], + "type": "text", + "content": " reconstruction loss. Here, " + }, + { + "bbox": [ + 304, + 540, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 540, + 545, + 613 + ], + "type": "text", + "content": " denotes the element-wise multiplication. This loss is applied to the neural render." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 627, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 627, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 627, + 545, + 651 + ], + "type": "text", + "content": "Final Loss. The SwitchLight is trained in an end-to-end manner using the weighted sum of the above losses:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 657, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 657, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 312, + 657, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {r e l i g h t}} = 1 0 \\cdot \\mathcal {L} _ {\\text {n o r m a l}} + 1 0 \\cdot \\mathcal {L} _ {\\text {s r c - H D R I}} + 0. 2 \\cdot \\mathcal {L} _ {\\text {s r c - d i f f}} \\\\ + 0. 2 \\cdot \\mathcal {L} _ {\\text {a l b e d o}} + 0. 2 \\cdot \\mathcal {L} _ {\\text {P B R}} + 0. 2 \\cdot \\mathcal {L} _ {\\text {N e u r a l}} \\tag {11} \\\\ + \\mathcal {L} _ {\\mathrm {v g g} _ {\\text {s r c . d i f f}}} + \\mathcal {L} _ {\\mathrm {v g g} _ {\\text {a l b e d o}}} + \\mathcal {L} _ {\\mathrm {v g g} _ {\\mathrm {P B R}}} + \\mathcal {L} _ {\\mathrm {v g g} _ {\\text {N e u r a l}}} \\\\ + \\mathcal {L} _ {\\mathrm {a d v} _ {\\mathrm {P B R}}} + \\mathcal {L} _ {\\mathrm {a d v} _ {\\mathrm {N e u r a l}}} + 0. 2 \\cdot \\mathcal {L} _ {\\mathrm {s p e c} _ {\\mathrm {N e u r a l}}}. \\\\ \\end{array}", + "image_path": "53abb11f0a6b92deae0bcd6be08715517b637d45bb3e132ca32f007fda6a58ee.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25100" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 108, + 129 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 108, + 129 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 108, + 129 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 108, + 129 + ], + "type": "image", + "image_path": "032dec24f947d1a38a99291244c0abcfd6d0d46af8cdf75294c55d89ec87bf49.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 134, + 287, + 167 + ], + "lines": [ + { + "bbox": [ + 47, + 134, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 134, + 287, + 167 + ], + "type": "text", + "content": "Figure 5. Dynamic Masking Strategies. We have generalized the MAE masks to include overlapping patches of varying sizes, as well as outpainting and free-form masks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 110, + 71, + 167, + 129 + ], + "blocks": [ + { + "bbox": [ + 110, + 71, + 167, + 129 + ], + "lines": [ + { + "bbox": [ + 110, + 71, + 167, + 129 + ], + "spans": [ + { + "bbox": [ + 110, + 71, + 167, + 129 + ], + "type": "image", + "image_path": "1ef7513b61de6a217ad8a36fdada2ab8dcd0f0604be62383edb9886b319259fc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 168, + 71, + 226, + 129 + ], + "blocks": [ + { + "bbox": [ + 168, + 71, + 226, + 129 + ], + "lines": [ + { + "bbox": [ + 168, + 71, + 226, + 129 + ], + "spans": [ + { + "bbox": [ + 168, + 71, + 226, + 129 + ], + "type": "image", + "image_path": "d4ef083b15cd653dd0d380e71b555f4c3e49705032698bca3c113325a261ae4b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 72, + 285, + 129 + ], + "blocks": [ + { + "bbox": [ + 228, + 72, + 285, + 129 + ], + "lines": [ + { + "bbox": [ + 228, + 72, + 285, + 129 + ], + "spans": [ + { + "bbox": [ + 228, + 72, + 285, + 129 + ], + "type": "image", + "image_path": "51234a997d19a0a4b7a616e62ee849c52f28158a3af14f5ad85d0a9d599b427d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 172, + 279, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 172, + 279, + 184 + ], + "spans": [ + { + "bbox": [ + 59, + 172, + 279, + 184 + ], + "type": "text", + "content": "We empirically determined the weighting coefficients." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 196, + 270, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 196, + 270, + 210 + ], + "spans": [ + { + "bbox": [ + 47, + 196, + 270, + 210 + ], + "type": "text", + "content": "4. Multi-Masked Autoencoder Pre-training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 216, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 287, + 467 + ], + "type": "text", + "content": "We introduce the Multi-Masked Autoencoder (MMAE), a self-supervised pre-training framework designed to enhance feature representations in relighting models. It aims to improve output quality without relying on additional, costly light stage data. Building upon the MAE framework [18], MMAE capitalizes on the inherent learning of crucial image features like structure, color, and texture, which are essential for relighting tasks. However, adapting MAE to our specific needs poses several non-trivial challenges. Firstly, MAE is primarily designed for vision transformers [14], while our focus is on a UNet, a convolution-based architecture. This convolutional structure, with its hierarchical nature and aggressive pooling, is known to simplify the MAE task, necessitating careful adaptation [49]. Further, the hyperparameters of MAE, particularly the fixed mask size and ratio, are also specific to vision transformers. These factors could introduce biases during training and hinder the model to recognize image features at various scales. Moreover, MAE relies solely on masked region reconstruction loss, limiting the model to understand the global coherence of the reconstructed region in relation to its visible context." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 468, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 287, + 491 + ], + "type": "text", + "content": "To address these challenges effectively, we have developed two key strategies within the MMAE framework:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 495, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 590 + ], + "type": "text", + "content": "Dynamic Masking. MMAE eliminates two key hyperparameters, mask size and ratio, by introducing a variety of mask types to generalize the MAE. These types, which include overlapping patches of various sizes, outpainting masks [45], and free-form masks [28] (see Fig.5), each contribute to the model's versatility. With the ability to handle challenging masked regions, MMAE achieves a more comprehensive understanding of image properties." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 590, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 686 + ], + "type": "text", + "content": "Generative Target. In addition to its structural advancements, MMAE incorporates a new loss function strategy. We have adopted perceptual [23] and adversarial losses [21], along with the original reconstruction loss. As a result, MMAE is equipped not only to reconstruct missing image parts but also to ensure synthesis capabilities and their seamless integration with the original context. In practice, the weights for the three losses are equally set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "content": "We pre-train the entire UNet architecture using MMAE, and, unlike MAE, we retain the decoder and fine-tune the" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 471, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 471, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 471, + 84 + ], + "type": "text", + "content": "entire model on relighting ground truths." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 95, + 347, + 107 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 95, + 347, + 107 + ], + "spans": [ + { + "bbox": [ + 306, + 95, + 347, + 107 + ], + "type": "text", + "content": "5. Data" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 114, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 114, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 545, + 270 + ], + "type": "text", + "content": "We constructed the OLAT (One Light at a Time) dataset using a light stage [10, 48] equipped with 137 programmable LED lights and 7 frontal-viewing cameras. Our dataset comprises images of 287 subjects, with each subject being captured in up to 15 different poses, resulting in a total of 29,705 OLAT sequences. We sourced HDRI dataset from several publicly available archives. Specifically, we acquired 559 HDRI maps from Polyhaven, 76 from Noah Witchell, 364 from HDRMAPS, 129 from iHDRI, and 34 from eisklotz. In addition, we incorporated synthetic HDRIs created using the method proposed in [30]. During training, HDRIs are randomly selected with equal probability from either real-world or synthetic collections." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 270, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 330 + ], + "type": "text", + "content": "We produced training pairs by projecting the sampled source and target lighting maps onto the reflectance fields of the OLAT images [10]. To derive the ground truth intrinsics, we applied the photometric stereo method [50] and obtained normal and albedo maps." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 341, + 387, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 341, + 387, + 354 + ], + "spans": [ + { + "bbox": [ + 306, + 341, + 387, + 354 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 361, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 504 + ], + "type": "text", + "content": "This section details our experimental results. We begin with a comparative evaluation of our method against state-of-the-art approaches using the OLAT dataset. We also employ images from the FFHQ-test [24] for user studies. For qualitative analysis, we utilize copyright-free portrait images from Pexels [1]. Additionally, we conduct ablation studies to validate the efficacy of our pre-training framework and architectural design choice. Subsequently, we detail the additional features and conclude by discussing its limitations. Our evaluation uses the OLAT test set, comprising 35 subjects and 11 lighting environments, ensuring no overlap with the train set." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 507, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 545, + 615 + ], + "type": "text", + "content": "Evaluation metrics. We employ several key metrics for evaluating the prediction accuracy; Mean Absolute Error (MAE), Mean Squared Error (MSE), Structural Similarity Index Measure (SSIM) and Learned Perceptual Image Patch Similarity (LPIPS). While these metrics offer valuable quantitative insights, they do not fully capture the subtleties of visual quality enhancement. Therefore, we emphasize the importance of qualitative evaluations to gain a comprehensive understanding of model performance." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Baselines. We compared our approach with three state-of-the-art baselines: Single Image Portrait Relighting (SIPR) [44], which uses a single neural network for relighting; Total Relight (TR) [33], employing multiple neural networks that incorporate the Phong reflectance model; and Lumos [51], a TR adaptation for synthetic datasets. Due to the lack of publicly available code or model from these methods, we either integrated their techniques into our" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25101" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 280, + 146 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 280, + 146 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 280, + 146 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 280, + 146 + ], + "type": "table", + "html": "
MAE ↓MSE ↓SSIM ↑LPIPS ↓
SIPR [44]0.17150.07480.84323.648
TR [33]0.16430.06580.84653.425
Ours0.10230.02750.90022.137
Ours (w. MMAE)0.09330.02350.90512.059
", + "image_path": "5bbeb8fdc73c2602473733c8e6b518d8c4c7c826f6a2099029acedb4ef2806eb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 54, + 167, + 281, + 230 + ], + "blocks": [ + { + "bbox": [ + 64, + 154, + 269, + 165 + ], + "lines": [ + { + "bbox": [ + 64, + 154, + 269, + 165 + ], + "spans": [ + { + "bbox": [ + 64, + 154, + 269, + 165 + ], + "type": "text", + "content": "Table 1. Quantitative Evaluation on the OLAT test set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 167, + 281, + 230 + ], + "lines": [ + { + "bbox": [ + 54, + 167, + 281, + 230 + ], + "spans": [ + { + "bbox": [ + 54, + 167, + 281, + 230 + ], + "type": "table", + "html": "
Lumos [51]TR [33]Ours
Consistent Lighting0.04780.18520.7671
Facial Details0.20220.26020.5376
Similar Identity0.17410.24400.5819
", + "image_path": "bb169db9c6e07b3884260a1ab25b3dadec745b1e40fe85da8a05812f093c004e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 238, + 244, + 249 + ], + "lines": [ + { + "bbox": [ + 89, + 238, + 244, + 249 + ], + "spans": [ + { + "bbox": [ + 89, + 238, + 244, + 249 + ], + "type": "text", + "content": "Table 2. User Study on the FFHQ test set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 49, + 253, + 285, + 438 + ], + "blocks": [ + { + "bbox": [ + 49, + 253, + 285, + 438 + ], + "lines": [ + { + "bbox": [ + 49, + 253, + 285, + 438 + ], + "spans": [ + { + "bbox": [ + 49, + 253, + 285, + 438 + ], + "type": "image", + "image_path": "8cba26b02b2c8544c446226b80440753c45bb31480af2d73c51fe9345f89cff7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 440, + 287, + 462 + ], + "lines": [ + { + "bbox": [ + 47, + 440, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 440, + 287, + 462 + ], + "type": "text", + "content": "Figure 6. Impact of Pre-training. The fine details such as specular highlights, skin tones, and shadows are notably improved." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 468, + 287, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 287, + 490 + ], + "type": "text", + "content": "framework or requested the original authors to process our inputs with their models and share the results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 491, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 612 + ], + "type": "text", + "content": "Quantitative Comparisons. The results in Table. 1 show our method outperforming SIPR and TR baselines, demonstrating the significance of incorporating advanced rendering physics and reflectance models. The transition from SIPR to TR emphasizes the value of physics-based design, while the shift from TR to our approach underscores the importance of transitioning from the empirical Phong model to the more accurate Cook-Torrance model. Additionally, pretraining contributes to further enhancements, as evidenced by the improved image details, depicted in Fig 6." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Qualitative Comparisons. Our relighting method exhibits several key advantages over previous approaches, as showcased in Fig. 7. It effectively harmonizes light direction and softness, avoiding harsh highlights and inaccurate lighting that are commonly observed in other methods. A notable strength of our approach lies in its ability to capture high-frequency details like specular highlights and hard shadows. Additionally, as shown in the second row, it preserves fa" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 545, + 327 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 327 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 327 + ], + "type": "image", + "image_path": "5b4143c5ca3289a0076e8b0113ea0f9cac2360741a92a669f01f36dc29f0b7ae.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 333, + 537, + 345 + ], + "lines": [ + { + "bbox": [ + 313, + 333, + 537, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 537, + 345 + ], + "type": "text", + "content": "Figure 7. Qualitative Comparison on the Pexels images [1]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 348, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 492 + ], + "type": "text", + "content": "cial details and identity, ensuring high fidelity to the subject's original features and mitigating common distortions seen in previous approaches. Moreover, our approach excels in handling skin tones, producing natural and accurate results under various lighting conditions. This is clearly demonstrated in the fourth row, where our method contrasts sharply with the over-saturated or pale tones from previous methods. Finally, the nuanced treatment of hair is highlighted in the sixth row, where our approach maintains luster and detail, unlike the flattened effect typical in other methods. More qualitative results are available in our supplementary video demonstration." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 712 + ], + "type": "text", + "content": "User Study. We conducted a human subjective test to evaluate the visual quality of relighting results, summarized in Table. 2. In each test case, workers were presented with an input image and an environment map. They were asked to compare the relighting results from three methods—Ours, Lumos, and TR—based on three criteria: 1) consistency of lighting with the environment map, 2) preservation of facial details, and 3) maintenance of the original identity. To ensure unbiased evaluations, the order of the methods presented was randomized. To aid in understanding the concept of consistent lighting, relit balls were displayed alongside the images. The study included a total of 256 images, consisting of 32 portraits each illuminated with 8 different HDRIs. Each worker was tasked with selecting the best image for each specific criterion, randomly assessing 30 samples. A total of 47 workers participated in the study. The results indicate a strong preference for our results over the baseline methods across all evaluated metrics." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25102" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 282, + 146 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 282, + 146 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 282, + 146 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 282, + 146 + ], + "type": "table", + "html": "
Pre-trainingMethodMAE ↓MSE ↓SSIM ↑LPIPS ↓
MAE [18]0.09520.02420.90072.096
MMAE0.09330.02350.90512.059
DiffuseNetAlbedo0.10530.02950.89852.197
Diff Render0.10230.02750.90022.137
", + "image_path": "af3af130d1500456a1f39eead8507dbd26b9ec67d4ae3f8dfc567d19afa43333.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 78, + 155, + 255, + 165 + ], + "lines": [ + { + "bbox": [ + 78, + 155, + 255, + 165 + ], + "spans": [ + { + "bbox": [ + 78, + 155, + 255, + 165 + ], + "type": "text", + "content": "Table 3. Ablation Studies on the OLAT test set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 49, + 174, + 282, + 417 + ], + "blocks": [ + { + "bbox": [ + 49, + 174, + 282, + 417 + ], + "lines": [ + { + "bbox": [ + 49, + 174, + 282, + 417 + ], + "spans": [ + { + "bbox": [ + 49, + 174, + 282, + 417 + ], + "type": "image", + "image_path": "175df5cf85a16b03ce2ceaf0fbc83572d83762d5e469848d15a98a9a8d44c001.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 421, + 286, + 442 + ], + "lines": [ + { + "bbox": [ + 47, + 421, + 286, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 286, + 442 + ], + "type": "text", + "content": "Figure 8. Ablation on DiffuseNet. Our approach successfully infers the albedo on various surfaces (skin, teeth, and accessories)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 454, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 287, + 682 + ], + "type": "text", + "content": "Ablation Studies. We analyze our two major design choices in Table. 3: the MMAE pre-training framework and DiffuseNet. The MMAE, which integrates dynamic masking with generative objectives, outperforms MAE. This superiority is mainly due to the incorporation of challenging masks and global coherence objectives, enabling the model to learn richer features during pre-training. Furthermore, our method of predicting diffuse render demonstrates superiority over direct albedo prediction. Firstly, we see it simplifies the learning process, as predicting diffuse render is more closely related to the original image. Secondly, our approach effectively distinguishes between the influences of illumination (diffuse shading) and surface properties (diffuse render). This distinction is crucial for accurately modeling the intrinsic color of surfaces, as it enables independent and precise evaluation of each element (see Eqn. 9). In contrast, methods that predict albedo directly often struggle to differentiate between these factors, leading to significant inaccuracies in color constancy, as shown in Fig. 8." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Applications. We present two applications using predicted intrinsics in Fig. 9. First, real-time PBR via Cook-Torrance" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 391, + 134 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 391, + 134 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 391, + 134 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 391, + 134 + ], + "type": "image", + "image_path": "bcae20feb2c5d47d383108bc8c144e7c5a968d4f0c475f664f06ab3158f89093.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 135, + 384, + 142 + ], + "lines": [ + { + "bbox": [ + 315, + 135, + 384, + 142 + ], + "spans": [ + { + "bbox": [ + 315, + 135, + 384, + 142 + ], + "type": "text", + "content": "(a) Real-Time PBR Rendering" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 400, + 70, + 545, + 134 + ], + "blocks": [ + { + "bbox": [ + 400, + 70, + 545, + 134 + ], + "lines": [ + { + "bbox": [ + 400, + 70, + 545, + 134 + ], + "spans": [ + { + "bbox": [ + 400, + 70, + 545, + 134 + ], + "type": "image", + "image_path": "09ba99ded5c01df2b6056d56f86a4009f651521d91b532638eb6be62d3bd1203.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 134, + 490, + 142 + ], + "lines": [ + { + "bbox": [ + 455, + 134, + 490, + 142 + ], + "spans": [ + { + "bbox": [ + 455, + 134, + 490, + 142 + ], + "type": "text", + "content": "(b) Copy Light" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 170, + 545, + 364 + ], + "blocks": [ + { + "bbox": [ + 306, + 144, + 547, + 166 + ], + "lines": [ + { + "bbox": [ + 306, + 144, + 547, + 166 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 547, + 166 + ], + "type": "text", + "content": "Figure 9. Applications. We showcase additional features of SwitchLight, powered by the diverse intrinsics features." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 170, + 545, + 364 + ], + "lines": [ + { + "bbox": [ + 307, + 170, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 170, + 545, + 364 + ], + "type": "image", + "image_path": "9ab99f19c87c386009f828589ac64c646f84a386eceb0fe0fdddcaa4467461f7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 368, + 545, + 401 + ], + "lines": [ + { + "bbox": [ + 306, + 368, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 306, + 368, + 545, + 401 + ], + "type": "text", + "content": "Figure 10. Limitations. The model faces challenges in removing strong shadows, misinterpreting reflective surfaces like sunglasses, and inaccurately predicting albedo for face paint." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 407, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 545, + 443 + ], + "type": "text", + "content": "components in Three.js graphics library. Second, switching the lighting environment between the source and reference images. Further details are in the supplementary video." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 449, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 545, + 605 + ], + "type": "text", + "content": "Limitations. We identified a few failure cases in Fig. 10. First, our model struggles with neutralizing strong shadows, which leads to inaccurate facial geometry and residual shadow artifacts in both albedo and relit images. Incorporating shadow augmentation techniques [15, 53] during training could mitigate this issue. Second, the model incorrectly interprets reflective surfaces, such as sunglasses, as opaque objects in the normal image. This error prevents the model from properly removing reflective highlights in the albedo and relit images. Lastly, the model inaccurately predicts the albedo for face paint. Implementing a semantic mask [51] to distinguish different semantic regions separately from the skin could help resolving these issues." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 609, + 378, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 378, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 378, + 621 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 629, + 545, + 713 + ], + "type": "text", + "content": "We introduce SwitchLight, an architecture based on Cook-Torrance rendering physics, enhanced with a self-supervised pre-training framework. This co-designed approach significantly outperforms previous models. Our future plans include scaling the current model beyond images to encompass video and 3D data. We hope our proposal serve as a new foundational model for relighting tasks." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25103" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 236, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 236, + 102 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 236, + 102 + ], + "type": "text", + "content": "[1] Pexels. https://www.pexels.com.6,7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 103, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 287, + 134 + ], + "type": "text", + "content": "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 137, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 137, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 54, + 137, + 287, + 178 + ], + "type": "text", + "content": "[3] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2014. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 181, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 181, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 54, + 181, + 287, + 235 + ], + "type": "text", + "content": "[4] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 236, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 236, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 287, + 280 + ], + "type": "text", + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 281, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 287, + 335 + ], + "type": "text", + "content": "[6] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22367-22377, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 336, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 287, + 369 + ], + "type": "text", + "content": "[7] Zhaoxi Chen and Ziwei Liu. Relighting4d: Neural relightable human from videos. In European Conference on Computer Vision, pages 606-623. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "type": "text", + "content": "[8] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM Transactions on Graphics (ToG), 1(1):7-24, 1982. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "text", + "content": "[9] Paul Debevec. Rendering synthetic objects into real scenes: Bridging traditional and image-based graphics with global illumination and high dynamic range photography. In Acm siggraph 2008 classes, pages 1-10. 2008. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 448, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 448, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 448, + 287, + 502 + ], + "type": "text", + "content": "[10] Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. Acquiring the reflectance field of a human face. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 145-156, 2000. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "type": "text", + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "type": "text", + "content": "[12] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422-1430, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "type": "text", + "content": "[13] Julie Dorsey, James Arvo, and Donald Greenberg. Interactive design of complex time dependent lighting. IEEE Computer Graphics and Applications, 15(2):26-36, 1995. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[15] David Futschik, Kelvin Ritland, James Vecore, Sean Fanello, Sergio Orts-Escalano, Brian Curless, Daniel Sykora, and Ro" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "hit Pandey. Controllable light diffusion for portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8412-8421, 2023. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 140 + ], + "type": "text", + "content": "[16] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "text", + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "type": "text", + "content": "[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "text", + "content": "[19] Andrew Hou, Ze Zhang, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Towards high fidelity face relighting with realistic shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14719-14728, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 310, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 363 + ], + "type": "text", + "content": "[20] Andrew Hou, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Face relighting with geometrically consistent shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4217-4226, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 365, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 420 + ], + "type": "text", + "content": "[21] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "text", + "content": "[22] Chaonan Ji, Tao Yu, Kaiwen Guo, Jingxin Liu, and Yebin Liu. Geometry-aware single-image full-body human relighting. In European Conference on Computer Vision, pages 388-405. Springer, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 468, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 531 + ], + "type": "text", + "content": "[23] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "text", + "content": "[24] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "type": "text", + "content": "[25] Taehun Kim, Kunhee Kim, Joonyeong Lee, Dongmin Cha, Jiho Lee, and Daijin Kim. Revisiting image pyramid structure for high resolution salient object detection. In Proceedings of the Asian Conference on Computer Vision, pages 108-124, 2022. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "text", + "content": "[26] Wenbo Li, Xin Lu, Shengju Qian, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer-based image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[27] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25104" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 150 + ], + "type": "text", + "content": "[28] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European conference on computer vision (ECCV), pages 85-100, 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 206 + ], + "type": "text", + "content": "[29] Yihao Liu, Jingwen He, Jinjin Gu, Xiangtao Kong, Yu Qiao, and Chao Dong. Degae: A new pretraining paradigm for low-level vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23292-23303, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 288, + 274 + ], + "type": "text", + "content": "[30] Yiqun Mei, He Zhang, Xuaner Zhang, Jianming Zhang, Zhixin Shu, Yilin Wang, Zijun Wei, Shi Yan, HyunJoon Jung, and Vishal M Patel. Lightpainter: Interactive portrait relighting with freehand scribble. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 195-205, 2023. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 288, + 330 + ], + "type": "text", + "content": "[31] Thomas Nestmeyer, Jean-François Lalonde, Iain Matthews, and Andreas Lehrmann. Learning physics-guided face relighting under directional light. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5124-5133, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 374 + ], + "type": "text", + "content": "[32] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 288, + 431 + ], + "type": "text", + "content": "[33] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 433, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 288, + 464 + ], + "type": "text", + "content": "[34] Namuk Park, Wonjae Kim, Byeongho Heo, Taekyung Kim, and Sangdoo Yun. What do self-supervised vision transformers learn? arXiv preprint arXiv:2305.00729, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 466, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 288, + 520 + ], + "type": "text", + "content": "[35] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 522, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 555 + ], + "type": "text", + "content": "[36] Bui Tuong Phong. Illumination for computer generated pictures. In Seminal graphics: pioneering efforts that shaped the field, pages 95-101. 1998. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 556, + 288, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 288, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 288, + 589 + ], + "type": "text", + "content": "[37] Puntawat Ponglertnapakorn, Nontawat Tritrong, and Supasorn Suwajanakorn. Difareli: Diffusion face relighting. arXiv preprint arXiv:2304.09479, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 624 + ], + "type": "text", + "content": "[38] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "text", + "content": "[39] Soumyadip Sengupta, Angjoo Kanazawa, Carlos D Castillo, and David W Jacobs. Sfsnet: Learning shape, reflectance and illuminance of facesin the wild'. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6296-6305, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[40] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In Pro-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2291-2300, 2020. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[41] Soumyadip Sengupta, Brian Curless, Ira Kemelmacher-Shlizerman, and Steven M Seitz. A light stage on every desk. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2420-2429, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[42] YiChang Shih, Sylvain Paris, Connelly Barnes, William T Freeman, and Fredo Durand. Style transfer for headshot portraits. 2014. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "type": "text", + "content": "[43] Zhixin Shu, Sunil Hadap, Eli Shechtman, Kalyan Sunkavalli, Sylvain Paris, and Dimitris Samaras. Portrait lighting transfer using a mass transport approach. ACM Transactions on Graphics (TOG), 36(4):1, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 220, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 275 + ], + "type": "text", + "content": "[44] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 277, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 342 + ], + "type": "text", + "content": "[45] Piotr Teterwak, Aaron Sarna, Dilip Krishnan, Aaron Maschinot, David Belanger, Ce Liu, and William T Freeman. Boundless: Generative adversarial networks for image extension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10521-10530, 2019. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 344, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 344, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 344, + 545, + 399 + ], + "type": "text", + "content": "[46] Yifan Wang, Aleksander Holynski, Xiuming Zhang, and Xuaner Zhang. Sunstage: Portrait reconstruction and relighting using the sun as a light stage. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20792-20802, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 399, + 545, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 444 + ], + "type": "text", + "content": "[47] Zhibo Wang, Xin Yu, Ming Lu, Quan Wang, Chen Qian, and Feng Xu. Single image portrait relighting via explicit multiple reflectance channel modeling. ACM Transactions on Graphics (TOG), 39(6):1-13, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 445, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 499 + ], + "type": "text", + "content": "[48] Andreas Wenger, Andrew Gardner, Chris Tchou, Jonas Unger, Tim Hawkins, and Paul Debevec. Performance relighting and reflectance transformation with timemultiplexed illumination. ACM Transactions on Graphics (TOG), 24(3):756-764, 2005. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 501, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 567 + ], + "type": "text", + "content": "[49] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133-16142, 2023. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 568, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 601 + ], + "type": "text", + "content": "[50] Robert J Woodham. Photometric method for determining surface orientation from multiple images. Optical engineering, 19(1):139-144, 1980. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 602, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 656 + ], + "type": "text", + "content": "[51] Yu-Ying Yeh, Koki Nagano, Sameh Khamis, Jan Kautz, Ming-Yu Liu, and Ting-Chun Wang. Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. ACM Transactions on Graphics (TOG), 41(6): 1-21, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 714 + ], + "type": "text", + "content": "[52] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 649-666. Springer, 2016. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25105" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 228 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[53] Xuaner Zhang, Jonathan T Barron, Yun-Ta Tsai, Rohit Pandey, Xiuming Zhang, Ren Ng, and David E Jacobs. Portrait shadow manipulation. ACM Transactions on Graphics (TOG), 39(4):78-1, 2020. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[54] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7194-7202, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 228 + ], + "type": "text", + "content": "[55] Taotao Zhou, Kai He, Di Wu, Teng Xu, Qixuan Zhang, Kuixiang Shao, Wenzheng Chen, Lan Xu, and Jingyi Yu. Relightable neural human assets from multi-view gradient illuminations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4315-4327, 2023. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25106" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/ba02978b-283d-434b-af7a-d4b604064e5e_content_list.json b/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/ba02978b-283d-434b-af7a-d4b604064e5e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..34c05c39bbad37fd5d0c9e70f0ba7518f5ce8de1 --- /dev/null +++ b/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/ba02978b-283d-434b-af7a-d4b604064e5e_content_list.json @@ -0,0 +1,1633 @@ +[ + { + "type": "text", + "text": "Symphonize 3D Semantic Scene Completion with Contextual Instance Queries", + "text_level": 1, + "bbox": [ + 89, + 130, + 879, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoyi Jiang $^{1*}$ Tianheng Cheng $^{1*}$ Naiyu Gao $^{2}$ Haoyang Zhang $^{2}$ Tianwei Lin $^{2}$ Wenyu Liu $^{1}$ Xinggang Wang $^{1\\dagger}$ $^{1}$ School of EIC, Huazhong University of Science & Technology \n $^{2}$ Horizon Robotics \n{haoyi_jiang,thch,liuwy,xgwang}@hust.edu.cn \n{naiyu.gao,haoyang.zhang,tianwei.lin}@horizon.cc", + "bbox": [ + 145, + 178, + 823, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Semantic Scene Completion (SSC) has emerged as a nascent and pivotal undertaking in autonomous driving, aiming to predict the voxel occupancy within volumetric scenes. However, prevailing methodologies primarily focus on voxel-wise feature aggregation, while neglecting instance semantics and scene context. In this paper, we present a novel paradigm termed Symphonies (Scene-from-Insts), that delves into the integration of instance queries to orchestrate 2D-to-3D reconstruction and 3D scene modeling. Leveraging our proposed Serial Instance-Propagated Attentions, Symphonies dynamically encodes instance-centric semantics, facilitating intricate interactions between the image and volumetric domains. Simultaneously, Symphonies fosters holistic scene comprehension by capturing context through the efficient fusion of instance queries, alleviating geometric ambiguities such as occlusion and perspective errors through contextual scene reasoning. Experimental results demonstrate that Symphonies achieves state-of-the-art performance on the challenging SemanticKITTI and SSCBench-KITTI-360 benchmarks, yielding remarkable mIoU scores of 15.04 and 18.58, respectively. These results showcase the promising advancements of our paradigm. The code for our method is available at https://github.com/hustvl/Symphonies.", + "bbox": [ + 75, + 353, + 473, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 762, + 209, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The advent of autonomous driving has brought forth novel challenges in the realm of 3D perception. In the pursuit of safe navigation and obstacle avoidance, autonomous vehicles must be equipped with the ability to accurately predict", + "bbox": [ + 75, + 787, + 470, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the occupancy of their immediate surroundings. This task, however, is not a facile endeavor, given the inherent complexities of the real world, characterized by clutter, ambiguity, and rapid evolution.", + "bbox": [ + 500, + 323, + 890, + 382 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Semantic Scene Completion (SSC) formulates this challenge as the reconstruction of occupancy and semantics for every voxel grid within a 3D scene. Recent advancements in vision-based solutions, such as MonoScene [3] and OccDepth [30], adopt 3D convolutional networks to elevate 2D image features to 3D volumes. TPVFormer [16], OccFormer [46], and CTF-Occ [38] explore decomposing 3D volumes into multiple coarse view representations and enhancing voxel interactions using Transformer [23, 39, 48] architectures.", + "bbox": [ + 496, + 383, + 892, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these advancements, contemporary approaches tend to prioritize voxel-wise modeling for 3D scenes resorting to pixel-voxel projection [6, 23, 31, 32] for the feature promotion from 2D to 3D. While focusing on these localized representations, they inadvertently neglect higher-level instance semantics, leading to vulnerability to geometric ambiguities arising from occlusion and perspective errors. Humans, in contrast, naturally perceive and comprehend through the concept of \"instance\", rather than isolated pixels or voxels, each imbued with semantic significance and cohesively contributing to the contextual whole of a scene. In light of these limitations, a fundamental question arises: How can we leverage the notion of instances to steer 3D scene modeling and 2D-to-3D reconstruction?", + "bbox": [ + 496, + 537, + 893, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Drawing inspiration from this notion, we propose Symphonies (Scene-from-Insts), a novel method that leverages contextual instance queries derived from image inputs to enhance scene modeling, exploiting inherent instance semantics and scene context. Stemming from this basis, we propose Serial Instance-Propagated Attentions to intricately interact with image and voxel features, deformably aggregating instance-centric semantics. This seamless interaction bridges the gap between low-level pixel or voxel representations and high-level semantics, facilitating feature promo", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. Work done during Haoyi Jiang's internship at Horizon Robotics. \n†Corresponding author.", + "bbox": [ + 76, + 862, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20258", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/630672cb301e7db0fb7131c638c4c3eb3edc5d57345031a59388111516a40c3e.jpg", + "image_caption": [ + "Figure 1. Comparison between voxel-wise modeling (a) and Symphonies (b). Conventional methods primarily depend on Inverse Perspective Mapping (IPM)-based voxel-pixel projection and voxel-wise feature aggregation, resulting in geometric ambiguities and computational redundancy. In contrast, Symphonies leverages instance queries as intermediaries to engage with image and voxel features, thus exploiting instance semantics and enhancing the contextual comprehension of the scene." + ], + "image_footnote": [], + "bbox": [ + 204, + 87, + 767, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion and scene modeling, as illustrated in Fig. 1. Furthermore, the synergistic fusion of multiple instance queries collectively enriches broader contextual information necessary for robust scene reasoning, contributing to the alleviation of geometric ambiguities. In tandem, we introduce the Depth-Rectified Voxel Proposal Layer to refine the initial geometry, elevating 2D image features to the implicit surface of the scene.", + "bbox": [ + 75, + 396, + 470, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our method, extensive experiments are conducted on SemanticKITTI [1] and SSCBench-KITTI-360 [21, 24] datasets. Symphonies achieves a remarkable state-of-the-art performance of 15.04 and $18.58\\mathrm{mIoU}$ , respectively, significantly outperforming previous vision-based methods by a substantial margin. Ablation experiments further underscore the promising advancements of our approach in the field of SSC. In summary, our contributions involve:", + "bbox": [ + 75, + 518, + 468, + 654 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce Symphonies, a pioneering paradigm for 3D Semantic Scene Completion (SSC), which delves into modeling instance-centric semantics with our proposed Serial Instance-Propagated Attentions, facilitating interactions between image features and scene modeling.", + "- Symphonies captures global context of scenes through the fusion of instance queries, thereby alleviating geometric ambiguities. To further enhance robustness, we introduce the Depth-Rectified Voxel Proposal Layer which explicitly refines the initial geometry with estimated depth information.", + "- Our proposed method achieves state-of-the-art performance on the SemanticKITTI and SSCBench-KITTI-360 benchmarks, highlighting the substantial potential of this paradigm in advancing autonomous driving and 3D scene understanding." + ], + "bbox": [ + 76, + 657, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 393, + 650, + 410 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Semantic Scene Completion. 3D Semantic Scene Completion (SSC) entails predicting occupancy and semantics for each voxel within a 3D scene, which was initially introduced by SSCNet [37]. Subsequent methods can be broadly categorized based on their model architectures and input modalities. Volume networks [19, 45] predominantly utilize Truncated Signed Distance Function (TSDF) features generated from depth data, and process them through 3D convolutional networks. On the other hand, view-volume networks [11, 20, 25, 33, 41] extract RGB or depth features before lifting them to 3D volumes. For a more in-depth overview of SSC, we refer readers to the survey by Roldão et al. [34].", + "bbox": [ + 496, + 422, + 890, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, camera-based SSC has garnered increasing attention for its immense applications in autonomous driving. MonoScene [3] presents the first purely visual solution, sampling RGB features along the line of sight and adapting a 3D UNet architecture. TPVFormer [16] introduces a Tri-Perspective View (TPV) representation to decompose voxels onto multiple view planes for efficient scene encoding. VoxFormer [22] proposes a two-stage framework that diffuses the global scene from proposed voxel features, resembling the Masked Autoencoder (MAE) [13]. OccFormer [46] applies a mask-wise prediction paradigm akin to MaskFormer [8, 9]. OccDepth [30] and NDC-Scene [44] improves 2D-to-3D geometric projection by leveraging stereo depth and Normalized Device Coordinates (NDC). OccNet [36] further envisions occupancy as a general scene descriptor for a wide scope of driving tasks.", + "bbox": [ + 496, + 625, + 892, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to prior works, our proposed Symphonies differs by integrating instance queries to enhance scene model", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20259", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/62ce4504b4438294c21678721d69642d5d90c74d7e505649b0386f905f7b0ad2.jpg", + "image_caption": [ + "Figure 2. Overview of Symphonies. The Symphonies framework commences with extracting multi-scale image features through the image backbone and Instance-Aware Image Encoder. The Depth-Rectified Voxel Proposal Layer generates initial voxel features estimating the implicit surface. Subsequently, the Symphonies Decoder Layers, which consist of Serial Instance-Propagated Attentions, facilitate continuous interactions among the image, instances, and the scene, iterated $N$ times. The Segmentation Head upsamples voxel features to the designated resolution and predicts class logits for each voxel." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 893, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing with instance semantics and enriched contextual awareness, mitigating geometric ambiguities arising from voxelwise modeling.", + "bbox": [ + 76, + 383, + 470, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camera-Based 3D Perception. The surge in autonomous driving applications has rekindled interest in camera-based 3D perception, owing to its cost-effectiveness and alignment with human visual perception. Early 3D object detection methods, such as FCOS3D [40] and DETR3D [42], straightforwardly extend 2D detectors to predict additional 3D bounding boxes. Among subsequent Transformer-based approaches, BEVFormer [23] and BEVDet [15] adopt the BEV space to align multi-frame features, while PolarDETR [7] establishes explicit correlations between image patterns. In addition, PETR [26] and PETRv2 [27] utilize 3D position embeddings to encode 2D features.", + "bbox": [ + 75, + 450, + 468, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "BEV segmentation, which is beneficial for representation learning and route planning, has also been explored. Approaches such as OFT [32], Lift-Splat [31], and FIERY [14] transform the camera plane into BEV via Inverse Perspective Mapping (IPM). PolarBEV [28] uses angle-specific and radius-specific embeddings to rasterize BEV features. BEVFormer [23] and CVT [47] aggregate BEV queries through cross-attention layers, while GKT [6] optimizes computational efficiency by constraining local attention calculations.", + "bbox": [ + 75, + 632, + 468, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These developments are closely related to our work in SSC, where techniques like Deformable Attention [48] inspire our methodology to enhance 3D scene completion.", + "bbox": [ + 76, + 784, + 470, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Scene from Instances", + "text_level": 1, + "bbox": [ + 76, + 844, + 279, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section presents a comprehensive elaboration of our proposed Symphonies method, commencing with an archi", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tectural overview in Sec. 3.1. Subsequently, it proceeds to detail the Depth-Rectified Voxel Proposal Layer in Sec. 3.2 and the Symphonies Decoder Layer in Sec. 3.3, shedding light on their synergistic contributions. Further insights into training losses are discussed in Sec. 3.4.", + "bbox": [ + 498, + 383, + 893, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 478, + 609, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The architectural details of our proposed Symphonies are illustrated in Fig. 2. In essence, Symphonies exclusively takes RGB images as input and extracts multi-scale 2D features $F^{2D}$ through a ResNet-50 [12] image backbone and an Instance-Aware Deformable Transformer [48] Encoder, enhancing both global and instance semantics on the image plane. In the Symphonies Decoder, instance queries $q_{ins} \\in \\mathbb{R}^{N \\times C}$ and the volumetric scene representation $q_{vox} \\in \\mathbb{R}^{C \\times X \\times Y \\times Z}$ are initialized with learnable embeddings. Here, $C$ signifies embedding dimensions, $N$ denotes the number of instance queries, while $X$ , $Y$ , and $Z$ indicate the scene grid dimensions.", + "bbox": [ + 496, + 503, + 892, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The subsequent \"scene-from-instances\" process commences with the Depth-Rectified Voxel Proposal Layer initializing voxel proposals $q_{p}$ with image features on the implicit surface. Multi-scale image features $F^{2D}$ , scene features $q_{vox}$ , and instance queries $q_{ins}$ are then passed through our proposed Serial Instance-Propagated Attentions within the Symphonies Decoder Layers. This iterative process continuously propagates image features $F^{2D}$ to scene features $q_{vox}$ guided by instance queries $q_{ins}$ , while simultaneously aggregating instance semantics from both modalities. The Segmentation Head then upsamples the scene features to the desired resolution, and predicts per-voxel class logits with a single linear layer after an Atrous Spatial Pyramid Pooling (ASPP) [5] module.", + "bbox": [ + 496, + 688, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20260", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Depth Estimator. The depth prediction, obtained from a pre-trained depth estimator, is not explicitly illustrated in the diagram for clarity. It is employed to infer the implicit surface within the Voxel Proposal Layer and compute instance reference points in the scene volume. Specifically, we adopt the pre-trained Mobilestereonet [35] as the depth estimator, aligning with VoxFormer [22].", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instance-Aware Image Encoder. The Instance-Aware Image Encoder, vital for integrating instance semantics in the absence of direct instance-level supervision, employs a Deformable Transformer [48] adept at capturing long-range dependencies around diverse instances by attending to deformable reference points. Additionally, it is augmented by utilizing the pre-trained weight of MaskDINO [18] from panoptic segmentation [17], to enrich its instance awareness.", + "bbox": [ + 75, + 214, + 472, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Depth-Rectified Voxel Proposal Layer", + "text_level": 1, + "bbox": [ + 76, + 358, + 406, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Depth-Rectified Voxel Proposal Layer generates initial scene features for voxels located on the implicit surface, known as voxel proposals, which establishes coarse geometry awareness for subsequent instance-level aggregations. The implicit surface is computed through the conversion of camera coordinates to world coordinates using depth estimation, described as follows:", + "bbox": [ + 75, + 381, + 472, + 486 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {C} = \\mathcal {K} ^ {- 1} \\cdot \\left(z _ {c} \\odot x ^ {I}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 494, + 468, + 512 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {W} = [ R, T ] ^ {- 1} \\cdot x ^ {C} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 515, + 468, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x^{I}, x^{C}$ , and $x^{W}$ represent homogeneous coordinates of pixels, camera frustum, and the world, respectively. $\\odot$ denotes the element-wise multiplication. The intrinsic matrix $\\mathcal{K}$ encompasses camera parameters, while the extrinsic matrix is composed of the rotation matrix $R$ and the translation vector $T$ . $z_{c}$ corresponds to the $z$ -coordinate of the camera, i.e., the depth estimation.", + "bbox": [ + 75, + 541, + 468, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the camera-to-world transformation, the positions $V_{p}$ of voxel proposals are determined by mapping image points $x^{I}$ to their corresponding world coordinates $x^{W}$ , confined within the volume $V$ :", + "bbox": [ + 75, + 648, + 468, + 708 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} V _ {p} = \\{x ^ {W} \\mid x ^ {W} = \\mathcal {T} ^ {I W} (x ^ {I}, z _ {c}), \\\\ \\forall x ^ {I} \\in I \\text {s u c h} x ^ {W} \\in V \\} \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 715, + 468, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathcal{T}^{IW}$ refers to the camera-to-world transformation, $I$ represents image pixels, and $V$ represents voxel grids.", + "bbox": [ + 75, + 763, + 468, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, the determined voxel features are initialized by aggregating multi-scale image features using Deformable Attention [48]. This process involves selecting the proposed voxels $q_{p}$ associated with the positions $V_{p}$ from scene volume $q_{vox}$ , along with corresponding pixel positions $p_{I}$ and 2D image features $F^{2D}$ . This process is expressed as $q_{p} = \\mathrm{DeformAttn}(q_{p}, p_{I}, F^{2D})$ .", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f0947edef9d5cb0fac912803284fd8a506f61ca7d9ed52824dc1f4c2319a4997.jpg", + "image_caption": [ + "Figure 3. Illustration of the Depth-Rectified Voxel Proposal Layer." + ], + "image_footnote": [], + "bbox": [ + 532, + 89, + 859, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Deformable Attention operation, denoted as DeformAttn, dynamically aggregates query features $q$ from features $x$ with deformable reference points $p_q$ . The mathematical expression is given by:", + "bbox": [ + 498, + 308, + 893, + 369 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {D e f o r m A t t n} (q, p _ {q}, x) = \\sum_ {k = 1} ^ {K} A _ {q k} W x \\left(p _ {q} + \\Delta p _ {q k}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 378, + 890, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $K$ represents the number of sampling points, and $A_{qk}$ stands for the learnable attention weight at sampling point $k$ deformable based on queries $q$ . The term $\\Delta p_{qk}$ denotes the offset applied to $p_q$ , and $W$ denotes the projection weight. The computation of multi-heads is omitted for brevity.", + "bbox": [ + 498, + 430, + 890, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to the Query Proposal in VoxFormer [22], which employs an extra occupancy network [33] for generating coarse occupancy features, we refrain from it as it introduces additional geometric ambiguities in occlusion regions.", + "bbox": [ + 498, + 506, + 890, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Symphonies Decoder Layer", + "text_level": 1, + "bbox": [ + 500, + 590, + 748, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Symphonies Decoder Layer seamlessly integrates image features, instance queries, and voxel proposals, as depicted in Fig. 2. It orchestrates a dynamic flow of information, where instance queries serve as intermediaries to propagate extracted instance semantics to the broader scene representations. The process initiates with deformable cross-attention modules, which attend to the corresponding instance positions within image and scene features to extract instance-centric semantics. Subsequently, the instance self-attention and scene-instance cross-attention modules strengthen the internal cohesion of instances and aggregate scene context from instance queries. The scene self-attention mechanism further diffuses voxel features throughout the scene, especially for the occluded regions, as only visible surfaces are initially proposed.", + "bbox": [ + 496, + 613, + 890, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The following paragraphs present a detailed explanation of the computations involved in their exact order of operation. To streamline the explanation, detailed discussions on certain components, including Feed-Forward Networks", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20261", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(FFN), Layer Norms (LN), and identity connections, have been omitted.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Deformable Instance-Image Cross-Attention. For each instance query $q_{ins}$ , deformable attention extracts surrounding features from multi-scale image features $F^{2D}$ using learnable 2D reference points $p_{ins}^{2D}$ , denoted as $q_{ins} = \\mathrm{DeformAttn}(q_{ins}, p_{ins}^{2D}, F^{2D})$ .", + "bbox": [ + 76, + 148, + 467, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scene-Instance Cross-Attention. This attention mechanism aggregates scene features $q_{vox}$ from instance queries, formulated as $q_{vox}^{\\in FOV} = \\text{CrossAttn}(q_{vox}^{\\in FOV}, q_{ins}, q_{ins})$ , where $FOV$ refers to the \"field of view\" which is precomputed based on world-to-camera transformation excluding invisible voxels, reducing computational redundancy.", + "bbox": [ + 76, + 250, + 467, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Deformable Scene Self-Attention. The scene self-attention enables feature propagation across the scene, where voxels attend to their neighbors: $q_{vox}^{\\in FOV} = \\mathrm{DeformAttn}(q_{vox}^{\\in FOV}, p_V, q_{vox})$ . Here, $p_V$ represents voxels' relative coordinates in the scene.", + "bbox": [ + 76, + 367, + 467, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Deformable Instance-Scene Cross-Attention. Instance semantics are enhanced by integrating refined information from the reconstructed voxel features $q_{vox}$ . Through coordinate transformation applied to 2D reference points, 3D reference points are derived as $p_{ins}^{3D} = \\mathcal{T}^{IW}(p_{ins}^{2D})$ . The instance-scene cross-attention is then formulated as $q_{ins} = \\mathrm{DeformAttn}(q_{ins}, p_{ins}^{3D}, q_{vox})$ .", + "bbox": [ + 76, + 469, + 467, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instance Self-Attention. The instance self-attention captures internal relations and global context within instance queries, expressed as $q_{ins} = \\mathrm{SelfAttn}(q_{ins})$ .", + "bbox": [ + 76, + 602, + 467, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Losses", + "text_level": 1, + "bbox": [ + 76, + 661, + 163, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the Symphonies framework, we adopt the Scene-Class Affinity Loss $L_{scal}$ from MonoScene [3] to optimize precision, recall, and specificity concurrently. The Scene-Class Affinity Loss is applied to semantic and geometric predictions, in conjunction with the cross-entropy loss weighted by class frequencies. The overall loss function is formulated as follows:", + "bbox": [ + 76, + 686, + 467, + 790 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {g e o}} + \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {s e m}} + \\mathcal {L} _ {\\text {c e}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 806, + 467, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the DETR series [4], auxiliary losses are applied after each Symphonies Decoder Layer for enhanced supervision, following the same formulation as $\\mathcal{L}$ but scaled by a factor of 0.5.", + "bbox": [ + 76, + 839, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 90, + 632, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present the evaluation results of our proposed Symphonies on SemanticKITTI [1] and SSCBenchKITTI-360 [21] datasets. The comparative analysis positioning Symphonies against existing approaches is detailed in Sec. 4.3. Additionally, comprehensive ablation studies are conducted in Sec. 4.4 to shed light on the thorough understanding of Symphonies.", + "bbox": [ + 498, + 114, + 890, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset and Metric", + "text_level": 1, + "bbox": [ + 500, + 232, + 684, + 246 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The evaluation is performed on SemanticKITTI [1] and SSCBench-KITTI-360 [21] datasets, both providing densely annotated urban driving scene sequences, 22 and 9 respectively, from the KITTI Odometry Benchmark [10]. These datasets voxelize point clouds and label scenes measuring $51.2m \\times 51.2m \\times 64m$ , with voxel grids of $256 \\times 256 \\times 32$ and a voxel size of $0.2m$ . SemanticKITTI comprises 10 sequences for training, 1 sequence for validation, and 11 sequences for testing. It furnishes RGB images with shapes of $1226 \\times 370$ as inputs and encompasses 20 semantic classes. SSCBench-KITTI-360 provides 7 sequences for training, 1 sequence for validation, and 1 sequence for testing, with 19 semantic classes and RGB images of size $1408 \\times 376$ . For our camera-based approach, we exclusively adopt RGB images as input, and report the intersection over union (IoU) and mean IoU (mIoU) metrics aligned with standard practices. The IoU metric assesses the binary classification of empty versus occupied voxels, reflecting the performance of geometric reconstruction. Conversely, the mIoU metric provides a comprehensive assessment of semantic understanding, making it the primary metric in most benchmarks.", + "bbox": [ + 498, + 255, + 890, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 599, + 715, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In line with prior studies [3, 16, 22], we train Symphonies for 30 epochs on 4 NVIDIA 3090 GPUs, with a batch size of 4 images. We apply random horizontal flip augmentation and employ the AdamW [29] optimizer with an initial learning rate of 2e-4 and a weight decay of 1e-4. Learning rate reduction occurs by a factor of 0.1 at the 25th epoch. The ResNet-50 [12] backbone and Image Encoder are initialized with pre-trained MaskDINO [18] weights.", + "bbox": [ + 498, + 623, + 890, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 500, + 755, + 638, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct a comprehensive comparison of Symphonies with the latest state-of-the-art camera-based methodologies on the SemanticKITTI and SSCBench-KITTI-360 datasets. The results, outlined in Tab. 1 and Tab. 2, establish the superior performance of Symphonies with substantial improvements of 2.72 and $4.77\\mathrm{mIoU}$ on SemanticKITTI and SSCBench-KITTI-360, respectively. Specifically, Symphonies showcases particular excellence in instance classes,", + "bbox": [ + 498, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20262", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/525ec6fd579f39e7a61154488b9192d8648537d0fcfc67de209360326d1c8bf5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodIoUmIoUroadsidewalkparkingother-grd.buildingcartruckbicyclemotorcycleother-veh.vegetationtrunkterrainpersonbicyclistmotorcyclistfencepoletraf-sign
(15.30%)(11.15%)(12%)(0.56%)(14.1%)(3.92%)(16%)(0.03%)(0.03%)(0.03%)(0.20%)(3.3%)(0.51%)(17%)(0.07%)(0.07%)(0.05%)(3.90%)(0.29%)(0.08%)
LMSCNet†[33]31.387.0746.7019.5013.503.1010.3014.300.300.000.000.0010.800.0010.400.000.000.005.400.000.00
AICNet†[20]23.937.0939.3018.3019.801.609.6015.300.700.000.000.009.601.9013.500.000.000.005.000.100.00
JS3C-Net†[43]34.008.9747.3021.7019.902.8012.7020.100.800.000.004.1014.203.1012.400.000.200.208.701.900.30
MonoScene* [3]34.1611.0854.7027.1024.805.7014.4018.803.300.500.704.4014.902.4019.501.001.400.4011.103.302.10
TPVFormer [16]34.2511.2655.1027.2027.406.5014.8019.203.701.000.502.3013.902.6020.401.102.400.3011.002.901.50
VoxFormer [22]42.9512.2053.9025.3021.105.6019.8020.803.501.000.703.7022.407.5021.301.402.600.2011.105.104.90
OccFormer [46]34.5312.3255.9030.3031.506.5015.7021.601.201.501.703.2016.803.9021.302.201.100.2011.903.803.70
Symphonies42.1915.0458.4029.3026.9011.7024.7023.603.203.602.605.6024.2010.0023.103.201.902.0016.107.708.00
", + "bbox": [ + 83, + 89, + 888, + 287 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e794c0fea0579c084cb3c00dd8016681afbc2b37eb07f939e0b498c09573ce00.jpg", + "table_caption": [ + "Table 1. Quantitative results on SemanticKITTI test. ${}^{ \\dagger }$ denotes the results provided by [3]. * represents the reproduced results in [16, 46]. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
MethodIoUPrec.Rec.mIoUcar(0.89%)bicycle(0.01%)motorcycle(0.03%)truck(0.16%)other-veh.(5.79%)person(0.02%)road(14.98%)parking(2.31%)sidewalk(6.43%)other-grd.(2.08%)building(15.67%)fence(0.96%)vegetation(41.99%)terrain(7.10%)pole(0.22%)traf.-sign(0.06%)other-struct.(4.33%)other-object.(0.28%)
LiDAR-based methods
SSCNet [37]53.5869.6369.9216.9531.950.000.1710.290.000.0765.7017.3341.243.2244.416.7743.7228.870.780.758.690.67
LMSCNet [33]47.5372.7757.5513.6520.910.000.000.260.580.0062.9513.5133.510.2043.670.3340.0126.800.000.003.630.00
Camera-based methods
MonoScene [3]37.8756.7353.2612.3119.340.430.588.022.030.8648.3511.3828.133.3232.893.5326.1516.756.925.674.203.09
TPVFormer [16]40.2259.3255.5413.6421.561.091.378.062.572.3852.9911.9931.073.7834.834.8030.0817.527.465.865.482.70
VoxFormer [22]38.7658.5253.4411.9117.841.160.894.562.061.6347.019.6727.212.8931.184.9728.9914.696.516.923.792.43
OccFormer [46]40.2759.7055.3113.8122.580.660.269.893.822.7754.3013.4431.533.5536.424.8031.0019.517.778.516.954.60
Symphonies44.1269.2454.8818.5830.021.855.9025.0712.068.2054.9413.8332.766.9335.118.5838.3311.5214.019.5714.4411.28
", + "bbox": [ + 84, + 340, + 885, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative results on SSCBench-KITTI-360 test. The results for counterparts are provided in [21]. The best results among all methods are in bold, and the best results for camera-based methods are underlined.", + "bbox": [ + 75, + 559, + 892, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "e.g., buildings, cars, etc. This underscores its prowess in capturing and modeling intricate instance semantics. While VoxFormer attains a marginally higher IoU on SemanticKITTI, its adoption of two-stage training and extra occupancy prediction network disrupts end-to-end training and introduces additional geometric ambiguities. This complexity hampers its robustness, especially on KITTI-360.", + "bbox": [ + 75, + 614, + 468, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The superiority of Symphonies becomes more pronounced on SSCBench-KITTI-360, which can be attributed to the ample data samples and high-quality annotations. Moreover, Symphonies even outperforms LiDAR-based methods in terms of mIoU, despite LiDAR's inherent advantage in IoU owing to its more precise position awareness, particularly at a distance.", + "bbox": [ + 75, + 722, + 468, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 844, + 241, + 859 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The ablation analysis is conducted on the SemanticKITTI validation set from four key perspectives: overall architec", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tural components, the Symphonies Decoder, the Voxel Proposal Layer, and the Image Encoder.", + "bbox": [ + 498, + 614, + 890, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation on architectural components. Tab. 3 presents the breakdown analysis of various architectural components within Symphonies. The vanilla baseline can be considered as a light-weight alternative to MonoScene, composed of a ResNet-50 backbone, an Image Encoder without pretrained weight, a 2D-to-3D projection via FLoSP [3], and a single 3D ASPP layer as the 3D decoder. Pre-training the Image Encoder leads to a notable improvement of 2.15 mIoU, emphasizing the effectiveness of instance awareness brought by 2D segmentation pre-training. Further, the proposed Depth-Rectified Voxel Proposal Layer improves performance by 0.75 mIoU through more accurate geometry. The Symphonies Decoder significantly boosts performance by 5.38 IoU, attributed to its dynamic instance modeling and context-capturing capabilities. In summary, the analy", + "bbox": [ + 496, + 674, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20263", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e2d6d20f85406d32a79dcab535155d6201fdf5019b29f65def0209c595906696.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodIoUmIoUParams (M)FLOPs (G)
Baseline34.0610.4457.22529.20
+ Pre-trained Encoder35.97 (+1.91)12.59 (+2.15)57.22529.20
+ Voxel Proposal Layer36.54 (+0.57)13.34 (+0.75)57.42535.84
+ Symphonies Decoder41.92 (+5.38)14.89 (+1.55)59.31611.89
", + "bbox": [ + 191, + 88, + 777, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sis in Tab. 3 affirms the effectiveness of the proposed components in Symphonies.", + "bbox": [ + 75, + 232, + 470, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on the Symphonies Decoder. To gain insights into the functionality of contextual instance queries, we assess the modular interactions within the Symphonies Decoder Layer. As depicted in Tab. 4, the incorporation of instance queries with either instance-image or instance-scene cross-attention considerably enhances performance by over 5 IoU and about $1\\mathrm{mIoU}$ . This substantiates the significance of instance queries for adaptive aggregation of instance semantics. Among them, the instance-image cross-attention brings less improvement, suggesting that original image features have already been adequately captured in the preceding Voxel Proposal Layer. The instance self-attention further improve the performance, underlining the contextual effectiveness of efficient fusion among instance queries. Besides, though Scene SA may seem to only marginally improve the performance, its omission reveals significant fluctuations during training, attributed to the sparse interactions within instance deformable attentions. This underscores the pivotal role of Scene SA in fostering the generation of consistent features.", + "bbox": [ + 75, + 281, + 472, + 583 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/889861f53b4b22d62f8f373e086e95073927cd3d09285189f5c78aa5b6c1e26b.jpg", + "table_caption": [ + "Table 3. Ablation study on architectural components in Symphonies." + ], + "table_footnote": [], + "table_body": "
Scn.-Inst. CAInst.-Img. CAInst.-Scn. CAInst. SAScn. SAIoUmIoU
35.9713.34
✓✓41.3614.02
✓✓41.4414.37
✓✓✓41.3514.63
✓✓✓✓41.7514.73
✓✓✓✓✓41.9214.89
", + "bbox": [ + 81, + 594, + 468, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on the Voxel Proposal Layer. Comparing the Depth-Rectified Voxel Proposal Layer (VPL) with FLoSP from MonoScene [3] casting pixels to voxels along the line of sight, as well as the mono VPL using monocular depth estimator AdaBins [2] (0.058 REL on KITTI), we note significant occupancy prediction improvements using", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the stereo VPL based on MobileStereoNet [35] (0.66 EPE on KITTI 2015), as shown in Tab. 5. This indicates that the rectification of more precise depth estimation contributes to mitigating geometric ambiguities, aligning with the findings in VoxFormer.", + "bbox": [ + 496, + 232, + 890, + 306 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1dcf5c3440af20887cc19ba77ab330ec3668baf0f8e6004b7add1bc661380477.jpg", + "table_caption": [ + "Table 4. Ablation study on Symphonies Decoder. SA: Self-Attention, CA: Cross-Attention." + ], + "table_footnote": [], + "table_body": "
2D-to-3D ProjectionIoUmIoU
FLoSP [3]36.0211.96
VPL (mono)38.3712.20
VPL (stereo)41.9214.89
", + "bbox": [ + 557, + 316, + 834, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on the Instance-Aware Image Encoder. Tab. 6 evidently showcases the synergistic effects of utilizing pre-trained weights for the Image Encoder with instance queries. Solely utilizing pre-trained weights from MaskDINO [18] contributes an additional improvement of $0.48\\mathrm{mIoU}$ . Moreover, incorporating instance queries with the pre-trained encoder yields a significant improvement of $1.36\\mathrm{mIoU}$ , implying that the proposed instance queries benefit from the enhanced instance awareness of the encoder.", + "bbox": [ + 496, + 453, + 890, + 603 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1c7275d58122db4f6ddbf7f8c3debafff9ed303c25a87727f56c0ac665fa4436.jpg", + "table_caption": [ + "Table 5. Ablation on the Depth-Rectified Voxel Proposal Layer." + ], + "table_footnote": [], + "table_body": "
Pre-trained EncoderInst. QueriesIoUmIoU
41.0913.53
✓41.4213.32
✓41.1814.01
✓✓41.9214.89
", + "bbox": [ + 509, + 614, + 883, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablation study on the Instance-Aware Image Encoder.", + "bbox": [ + 500, + 715, + 890, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Visualizations", + "text_level": 1, + "bbox": [ + 500, + 756, + 643, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 4 presents the visualizations of Symphonies on SemanticKITTI val, in comparison to the counterpart MonoScene. Symphonies generate more detailed predictions for instance-centric classes such as cars and trunks, as well as preserves clear and coherent layouts for structures like buildings and vegetation, attributed to the enriched instance semantics and contextual information provided by instance queries. In contrast, MonoScene", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20264", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d125133181974605ba41e81768fed2c41d55e725119c7d980555382e68ecb1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 158, + 97, + 339, + 146 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d80d39c1169085adc21f138a11d94fdb390b242ed226bc7f7a31dab032d2a834.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 95, + 496, + 156 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5d086f8d2607486274a2e6b6f0e8a054f6ec998aefba9130a931773540a3f12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 94, + 642, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/874ff1dd621234463cc660c3e0abc3f05dea52661a496c44c33d3fc54e5a5686.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 93, + 803, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aeb6af285715e91e4bf615cdf13bb2a9ce3558efc0ab239191bf82264d159a54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 158, + 179, + 338, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f0d6f135247d721faf5164791f1ba8ebaf11e8d1f9e62930483a99f57687011.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 170, + 496, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9e0c1293852789933f2c00d3608686483e558fd6d78314b907cc107c294ddbec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 170, + 640, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/42d247771d3edc90d32153779ad5fd0c102c650d86e78e94a9254f78efada1e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 169, + 805, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c8c12316824ed975ca94e4514e2d8a8b5d2726ad87b4543883aac8da378b7f24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 158, + 257, + 336, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/977752bda281ea6d7792c4385dda1021ea78b96cc3a08742be44429ed6e5a353.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 250, + 491, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3d616aaf9ffc708c8cce5ef02b07e7f6e9c55ae48011db097993f180cfb9b92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 256, + 640, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9646790bfd3561b68d44d152ea33a363e9d00a311420173515c98ed6342402c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 250, + 795, + 313 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2db1dcff58e97e7f1da12db3b21e3e4758a38d2eb5fa94ee7a13e490bcce3017.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 158, + 335, + 336, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5486a991f32fd983fa5d82609ce9a60cfa431bd15c7e71f8d0300840adcef9b.jpg", + "image_caption": [ + "MonoScene" + ], + "image_footnote": [], + "bbox": [ + 356, + 324, + 491, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a59bd518684cea3d24af0b9635cd34b7e98e8d0002b2bb32bef6ddd576bef73c.jpg", + "image_caption": [ + "Symphonies" + ], + "image_footnote": [], + "bbox": [ + 509, + 323, + 637, + 380 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c98e999f12b80536673b01ebcd39c1a6cf14591a7f4796d2021342f72de41452.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 658, + 316, + 805, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "produces vague predictions with a radial shape, which is indicative of the aforementioned ambiguous geometry. These results underscore the superior capability of Symphonies in capturing fine-grained scene representations and enhancing overall scene understanding.", + "bbox": [ + 75, + 477, + 468, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Attention Map Analysis. The attention map analysis in Fig. 5 provides insights into the mechanisms of the Serial Instance-Propagated Attentions within Symphonies layers. Notably, instance queries exhibit selective attention to corresponding regions in both the image and the scene. Additionally, they activate the semantically related regions within the scene-instance cross-attention. This observation validates the effect of our claimed instance-centric semantics in facilitating effective scene modeling.", + "bbox": [ + 75, + 573, + 468, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 723, + 194, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced Symphonies, a novel paradigm for 3D Semantic Scene Completion. Symphonies effectively integrates instance-centric semantics and scene context from both images and volumes, addressing the limitations posed by geometric ambiguity in prior voxel-wise modeling methods. Extensive experiments demonstrate the superiority of our approach over existing methods. We anticipate Symphonies to inspire future research and contribute to advancements in autonomous driving and 3D perception.", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/630d16c9f1c967ae06bf8293c98243d69b046a7a71a30ff338e2fc4199ade2d6.jpg", + "image_caption": [ + "Figure 4. Qualitative visualizations on SemanticKITTI va1. Symphonies consistently produces detailed predictions for objects such as cars and trunks, while maintaining coherent layouts for structures like buildings and vegetation." + ], + "image_footnote": [], + "bbox": [ + 501, + 474, + 746, + 534 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/db37814759ec8622a1ec5a5752cadba636b31f820fe1b68e9f028a388357e04a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 476, + 885, + 534 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/519e8fba28247ac876fe658f51762bf4b9687dfa6a85b7a3bc4b5dd9a7a048e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 544, + 885, + 602 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ff1209a9176fee41c965e16ee4755887c423f9bc9dc8cc5520491605e22d94c2.jpg", + "image_caption": [ + "(c) Attention Maps for SceneInst. Cross-Attention." + ], + "image_footnote": [], + "bbox": [ + 763, + 613, + 885, + 671 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bd5af940c866a1d9da648039955057db242f30594779b19299162508106fcdbf.jpg", + "image_caption": [ + "(a) Reference points for Inst.-Image Cross-Attention.", + "(b) Reference points for Inst.-Scene Cross-Attention.", + "Figure 5. Analysis of attention maps within Symphonies." + ], + "image_footnote": [], + "bbox": [ + 501, + 551, + 746, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was partially supported by the National Natural Science Foundation of China (NSFC) under Grant No. 62276108.", + "bbox": [ + 498, + 748, + 890, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 820, + 596, + 834 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Int. Conf. Comput. Vis., pages 9296-9306,", + "bbox": [ + 506, + 844, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20265", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2019. 2, 5", + "[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 7", + "[3] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3981–3991, 2022. 1, 2, 5, 6, 7", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Eur. Conf. Comput. Vis., pages 213-229, 2020. 5", + "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., pages 834-848, 2018. 3", + "[6] Shaoyu Chen, Tianheng Cheng, Xinggang Wang, Wenming Meng, Qian Zhang, and Wenyu Liu. Efficient and robust 2d-to-bev representation learning via geometry-guided kernel transformer. arXiv preprint arXiv:2206.04584, 2022. 1, 3", + "[7] Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Chang Huang, and Wenyu Liu. Polar parametrization for vision-based surround-view 3d detection. arXiv preprint arXiv:2206.10965, 2022. 3", + "[8] Bowen Cheng, Alexander G. Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. In Adv. Neural Inform. Process. Syst., pages 17864-17875, 2021. 2", + "[9] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1280–1289, 2022. 2", + "[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In IEEE Conf. Comput. Vis. Pattern Recog., 2012. 5", + "[11] Yuxiao Guo and Xin Tong. View-volume network for semantic scene completion from a single depth image. In IJCAI, pages 726-732, 2018. 2", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 3, 5", + "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE Conf. Comput. Vis. Pattern Recog., pages 15979-15988, 2022. 2", + "[14] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. FIERY: future instance prediction in bird's-eye view from surround monocular cameras. In Int. Conf. Comput. Vis., pages 15253-15262, 2021. 3", + "[15] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 3", + "[16] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9223-9232, 2023. 1, 2, 5, 6" + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Alexander Kirillov, Ross B. Girshick, Kaiming He, and Piotr Dólar. Panoptic feature pyramid networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 4", + "[18] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M. Ni, and Heung-Yeung Shum. Mask DINO: towards A unified transformer-based framework for object detection and segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3041-3050, 2023. 4, 5, 7", + "[19] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian D. Reid. RGBD based dimensional decomposition residual network for 3d semantic scene completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7693-7702, 2019. 2", + "[20] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3348-3356, 2020. 2, 6", + "[21] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. arXiv preprint arXiv:2306.09001, 2023. 2, 5, 6", + "[22] Yiming Li, Zhiding Yu, Christopher B. Choy, Chaowei Xiao, José M. Álvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9087-9098, 2023. 2, 4, 5, 6", + "[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Eur. Conf. Comput. Vis., pages 1-18, 2022. 1, 3", + "[24] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Trans. Pattern Anal. Mach. Intell., pages 3292–3310, 2023. 2", + "[25] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. In Adv. Neural Inform. Process. Syst., pages 261-272, 2018. 2", + "[26] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETR: position embedding transformation for multi-view 3d object detection. In Eur. Conf. Comput. Vis., pages 531-548, 2022. 3", + "[27] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. 3", + "[28] Zhi Liu, Shaoyu Chen, Xiaojie Guo, Xinggang Wang, Tianheng Cheng, Hongmei Zhu, Qian Zhang, Wenyu Liu, and Yi Zhang. Vision-based uneven BEV representation learning with polar rasterization and surface estimation. In Conference on Robot Learning, pages 437-446, 2022. 3", + "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Int. Conf. Learn. Represent., 2019. 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20266", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 1, 2", + "[31] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Eur. Conf. Comput. Vis., pages 194-210, 2020. 1, 3", + "[32] Thomas Roddick, Alex Kendall, and Roberto Cipolla. Orthographic feature transform for monocular 3d object detection. In Brit. Mach. Vis. Conf., page 285, 2019. 1, 3", + "[33] Luis Roldão, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 3DV, pages 111–119, 2020. 2, 4, 6", + "[34] Luis Roldão, Raoul de Charette, and Anne Verroust-Blondet. 3d semantic scene completion: A survey. Int. J. Comput. Vis., 130(8):1978-2005, 2022. 2", + "[35] Faranak Shamsafar, Samuel Woerz, Rafia Rahim, and Andreas Zell. Mobilestereonet: Towards lightweight deep networks for stereo matching. In WACV, pages 677-686, 2022. 4, 7", + "[36] Chonghao Sima, Wenwen Tong, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, and Hongyang Li. Scene as occupancy. arXiv preprint arXiv:2306.02851, 2023. 2", + "[37] Shuran Song, Fisher Yu, Andy Zeng, Angel X. Chang, Manolis Savva, and Thomas A. Funkhouser. Semantic scene completion from a single depth image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 190-198, 2017. 2, 6", + "[38] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 1", + "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Adv. Neural Inform. Process. Syst., pages 5998-6008, 2017. 1", + "[40] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. FCOS3D: fully convolutional one-stage monocular 3d object detection. In ICCVW, pages 913-922, 2021. 3", + "[41] Yida Wang, David Joseph Tan, Nassir Navab, and Federico Tombari. Forknet: Multi-branch volumetric semantic completion from a single depth image. In Int. Conf. Comput. Vis., pages 8607–8616, 2019. 2", + "[42] Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. DETR3D: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191, 2021. 3", + "[43] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 6", + "[44] Jiawei Yao, Chuming Li, Keqiang Sun, Yingjie Cai, Hao Li, Wanli Ouyang, and Hongsheng Li. Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. arXiv preprint arXiv:2309.14616, 2023.2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Pingping Zhang, Wei Liu, Yinjie Lei, Huchuan Lu, and Xiaoyun Yang. Cascaded context pyramid for full-resolution 3d semantic scene completion. In Int. Conf. Comput. Vis., pages 7800-7809, 2019. 2", + "[46] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2304.05316, 2023. 1, 2, 6", + "[47] Brady Zhou and Philipp Kähenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13750-13759, 2022. 3", + "[48] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In Int. Conf. Learn. Represent., 2021. 1, 3, 4" + ], + "bbox": [ + 501, + 92, + 890, + 315 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20267", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/ba02978b-283d-434b-af7a-d4b604064e5e_model.json b/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/ba02978b-283d-434b-af7a-d4b604064e5e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7500507acb97efc5922324cd6a211a3fef0d3b39 --- /dev/null +++ b/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/ba02978b-283d-434b-af7a-d4b604064e5e_model.json @@ -0,0 +1,2211 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.131, + 0.88, + 0.154 + ], + "angle": 0, + "content": "Symphonize 3D Semantic Scene Completion with Contextual Instance Queries" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.179, + 0.825, + 0.288 + ], + "angle": 0, + "content": "Haoyi Jiang\\(^{1*}\\) Tianheng Cheng\\(^{1*}\\) Naiyu Gao\\(^{2}\\) Haoyang Zhang\\(^{2}\\) Tianwei Lin\\(^{2}\\) Wenyu Liu\\(^{1}\\) Xinggang Wang\\(^{1\\dagger}\\) \n\\(^{1}\\) School of EIC, Huazhong University of Science & Technology \n\\(^{2}\\) Horizon Robotics \n{haoyi_jiang,thch,liuwy,xgwang}@hust.edu.cn \n{naiyu.gao,haoyang.zhang,tianwei.lin}@horizon.cc" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.474, + 0.735 + ], + "angle": 0, + "content": "3D Semantic Scene Completion (SSC) has emerged as a nascent and pivotal undertaking in autonomous driving, aiming to predict the voxel occupancy within volumetric scenes. However, prevailing methodologies primarily focus on voxel-wise feature aggregation, while neglecting instance semantics and scene context. In this paper, we present a novel paradigm termed Symphonies (Scene-from-Insts), that delves into the integration of instance queries to orchestrate 2D-to-3D reconstruction and 3D scene modeling. Leveraging our proposed Serial Instance-Propagated Attentions, Symphonies dynamically encodes instance-centric semantics, facilitating intricate interactions between the image and volumetric domains. Simultaneously, Symphonies fosters holistic scene comprehension by capturing context through the efficient fusion of instance queries, alleviating geometric ambiguities such as occlusion and perspective errors through contextual scene reasoning. Experimental results demonstrate that Symphonies achieves state-of-the-art performance on the challenging SemanticKITTI and SSCBench-KITTI-360 benchmarks, yielding remarkable mIoU scores of 15.04 and 18.58, respectively. These results showcase the promising advancements of our paradigm. The code for our method is available at https://github.com/hustvl/Symphonies." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.763, + 0.21, + 0.78 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.789, + 0.471, + 0.852 + ], + "angle": 0, + "content": "The advent of autonomous driving has brought forth novel challenges in the realm of 3D perception. In the pursuit of safe navigation and obstacle avoidance, autonomous vehicles must be equipped with the ability to accurately predict" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.324, + 0.892, + 0.383 + ], + "angle": 0, + "content": "the occupancy of their immediate surroundings. This task, however, is not a facile endeavor, given the inherent complexities of the real world, characterized by clutter, ambiguity, and rapid evolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.893, + 0.534 + ], + "angle": 0, + "content": "3D Semantic Scene Completion (SSC) formulates this challenge as the reconstruction of occupancy and semantics for every voxel grid within a 3D scene. Recent advancements in vision-based solutions, such as MonoScene [3] and OccDepth [30], adopt 3D convolutional networks to elevate 2D image features to 3D volumes. TPVFormer [16], OccFormer [46], and CTF-Occ [38] explore decomposing 3D volumes into multiple coarse view representations and enhancing voxel interactions using Transformer [23, 39, 48] architectures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.895, + 0.749 + ], + "angle": 0, + "content": "Despite these advancements, contemporary approaches tend to prioritize voxel-wise modeling for 3D scenes resorting to pixel-voxel projection [6, 23, 31, 32] for the feature promotion from 2D to 3D. While focusing on these localized representations, they inadvertently neglect higher-level instance semantics, leading to vulnerability to geometric ambiguities arising from occlusion and perspective errors. Humans, in contrast, naturally perceive and comprehend through the concept of \"instance\", rather than isolated pixels or voxels, each imbued with semantic significance and cohesively contributing to the contextual whole of a scene. In light of these limitations, a fundamental question arises: How can we leverage the notion of instances to steer 3D scene modeling and 2D-to-3D reconstruction?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Drawing inspiration from this notion, we propose Symphonies (Scene-from-Insts), a novel method that leverages contextual instance queries derived from image inputs to enhance scene modeling, exploiting inherent instance semantics and scene context. Stemming from this basis, we propose Serial Instance-Propagated Attentions to intricately interact with image and voxel features, deformably aggregating instance-centric semantics. This seamless interaction bridges the gap between low-level pixel or voxel representations and high-level semantics, facilitating feature promo" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.863, + 0.471, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. Work done during Haoyi Jiang's internship at Horizon Robotics. \n†Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20258" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.205, + 0.088, + 0.768, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.314, + 0.895, + 0.371 + ], + "angle": 0, + "content": "Figure 1. Comparison between voxel-wise modeling (a) and Symphonies (b). Conventional methods primarily depend on Inverse Perspective Mapping (IPM)-based voxel-pixel projection and voxel-wise feature aggregation, resulting in geometric ambiguities and computational redundancy. In contrast, Symphonies leverages instance queries as intermediaries to engage with image and voxel features, thus exploiting instance semantics and enhancing the contextual comprehension of the scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.471, + 0.517 + ], + "angle": 0, + "content": "tion and scene modeling, as illustrated in Fig. 1. Furthermore, the synergistic fusion of multiple instance queries collectively enriches broader contextual information necessary for robust scene reasoning, contributing to the alleviation of geometric ambiguities. In tandem, we introduce the Depth-Rectified Voxel Proposal Layer to refine the initial geometry, elevating 2D image features to the implicit surface of the scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.655 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our method, extensive experiments are conducted on SemanticKITTI [1] and SSCBench-KITTI-360 [21, 24] datasets. Symphonies achieves a remarkable state-of-the-art performance of 15.04 and \\(18.58\\mathrm{mIoU}\\), respectively, significantly outperforming previous vision-based methods by a substantial margin. Ablation experiments further underscore the promising advancements of our approach in the field of SSC. In summary, our contributions involve:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.659, + 0.47, + 0.735 + ], + "angle": 0, + "content": "- We introduce Symphonies, a pioneering paradigm for 3D Semantic Scene Completion (SSC), which delves into modeling instance-centric semantics with our proposed Serial Instance-Propagated Attentions, facilitating interactions between image features and scene modeling." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.736, + 0.47, + 0.824 + ], + "angle": 0, + "content": "- Symphonies captures global context of scenes through the fusion of instance queries, thereby alleviating geometric ambiguities. To further enhance robustness, we introduce the Depth-Rectified Voxel Proposal Layer which explicitly refines the initial geometry with estimated depth information." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- Our proposed method achieves state-of-the-art performance on the SemanticKITTI and SSCBench-KITTI-360 benchmarks, highlighting the substantial potential of this paradigm in advancing autonomous driving and 3D scene understanding." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.659, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.395, + 0.651, + 0.411 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.621 + ], + "angle": 0, + "content": "3D Semantic Scene Completion. 3D Semantic Scene Completion (SSC) entails predicting occupancy and semantics for each voxel within a 3D scene, which was initially introduced by SSCNet [37]. Subsequent methods can be broadly categorized based on their model architectures and input modalities. Volume networks [19, 45] predominantly utilize Truncated Signed Distance Function (TSDF) features generated from depth data, and process them through 3D convolutional networks. On the other hand, view-volume networks [11, 20, 25, 33, 41] extract RGB or depth features before lifting them to 3D volumes. For a more in-depth overview of SSC, we refer readers to the survey by Roldão et al. [34]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.893, + 0.868 + ], + "angle": 0, + "content": "Recently, camera-based SSC has garnered increasing attention for its immense applications in autonomous driving. MonoScene [3] presents the first purely visual solution, sampling RGB features along the line of sight and adapting a 3D UNet architecture. TPVFormer [16] introduces a Tri-Perspective View (TPV) representation to decompose voxels onto multiple view planes for efficient scene encoding. VoxFormer [22] proposes a two-stage framework that diffuses the global scene from proposed voxel features, resembling the Masked Autoencoder (MAE) [13]. OccFormer [46] applies a mask-wise prediction paradigm akin to MaskFormer [8, 9]. OccDepth [30] and NDC-Scene [44] improves 2D-to-3D geometric projection by leveraging stereo depth and Normalized Device Coordinates (NDC). OccNet [36] further envisions occupancy as a general scene descriptor for a wide scope of driving tasks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In contrast to prior works, our proposed Symphonies differs by integrating instance queries to enhance scene model" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20259" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.087, + 0.895, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.289, + 0.893, + 0.361 + ], + "angle": 0, + "content": "Figure 2. Overview of Symphonies. The Symphonies framework commences with extracting multi-scale image features through the image backbone and Instance-Aware Image Encoder. The Depth-Rectified Voxel Proposal Layer generates initial voxel features estimating the implicit surface. Subsequently, the Symphonies Decoder Layers, which consist of Serial Instance-Propagated Attentions, facilitate continuous interactions among the image, instances, and the scene, iterated \\(N\\) times. The Segmentation Head upsamples voxel features to the designated resolution and predicts class logits for each voxel." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.385, + 0.471, + 0.431 + ], + "angle": 0, + "content": "ing with instance semantics and enriched contextual awareness, mitigating geometric ambiguities arising from voxelwise modeling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.451, + 0.47, + 0.633 + ], + "angle": 0, + "content": "Camera-Based 3D Perception. The surge in autonomous driving applications has rekindled interest in camera-based 3D perception, owing to its cost-effectiveness and alignment with human visual perception. Early 3D object detection methods, such as FCOS3D [40] and DETR3D [42], straightforwardly extend 2D detectors to predict additional 3D bounding boxes. Among subsequent Transformer-based approaches, BEVFormer [23] and BEVDet [15] adopt the BEV space to align multi-frame features, while PolarDETR [7] establishes explicit correlations between image patterns. In addition, PETR [26] and PETRv2 [27] utilize 3D position embeddings to encode 2D features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.47, + 0.784 + ], + "angle": 0, + "content": "BEV segmentation, which is beneficial for representation learning and route planning, has also been explored. Approaches such as OFT [32], Lift-Splat [31], and FIERY [14] transform the camera plane into BEV via Inverse Perspective Mapping (IPM). PolarBEV [28] uses angle-specific and radius-specific embeddings to rasterize BEV features. BEVFormer [23] and CVT [47] aggregate BEV queries through cross-attention layers, while GKT [6] optimizes computational efficiency by constraining local attention calculations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.785, + 0.471, + 0.83 + ], + "angle": 0, + "content": "These developments are closely related to our work in SSC, where techniques like Deformable Attention [48] inspire our methodology to enhance 3D scene completion." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.281, + 0.86 + ], + "angle": 0, + "content": "3. Scene from Instances" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "This section presents a comprehensive elaboration of our proposed Symphonies method, commencing with an archi" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.385, + 0.895, + 0.462 + ], + "angle": 0, + "content": "tectural overview in Sec. 3.1. Subsequently, it proceeds to detail the Depth-Rectified Voxel Proposal Layer in Sec. 3.2 and the Symphonies Decoder Layer in Sec. 3.3, shedding light on their synergistic contributions. Further insights into training losses are discussed in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.479, + 0.611, + 0.493 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.893, + 0.687 + ], + "angle": 0, + "content": "The architectural details of our proposed Symphonies are illustrated in Fig. 2. In essence, Symphonies exclusively takes RGB images as input and extracts multi-scale 2D features \\(F^{2D}\\) through a ResNet-50 [12] image backbone and an Instance-Aware Deformable Transformer [48] Encoder, enhancing both global and instance semantics on the image plane. In the Symphonies Decoder, instance queries \\(q_{ins} \\in \\mathbb{R}^{N \\times C}\\) and the volumetric scene representation \\(q_{vox} \\in \\mathbb{R}^{C \\times X \\times Y \\times Z}\\) are initialized with learnable embeddings. Here, \\(C\\) signifies embedding dimensions, \\(N\\) denotes the number of instance queries, while \\(X\\), \\(Y\\), and \\(Z\\) indicate the scene grid dimensions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.902 + ], + "angle": 0, + "content": "The subsequent \"scene-from-instances\" process commences with the Depth-Rectified Voxel Proposal Layer initializing voxel proposals \\( q_{p} \\) with image features on the implicit surface. Multi-scale image features \\( F^{2D} \\), scene features \\( q_{vox} \\), and instance queries \\( q_{ins} \\) are then passed through our proposed Serial Instance-Propagated Attentions within the Symphonies Decoder Layers. This iterative process continuously propagates image features \\( F^{2D} \\) to scene features \\( q_{vox} \\) guided by instance queries \\( q_{ins} \\), while simultaneously aggregating instance semantics from both modalities. The Segmentation Head then upsamples the scene features to the desired resolution, and predicts per-voxel class logits with a single linear layer after an Atrous Spatial Pyramid Pooling (ASPP) [5] module." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20260" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "Depth Estimator. The depth prediction, obtained from a pre-trained depth estimator, is not explicitly illustrated in the diagram for clarity. It is employed to infer the implicit surface within the Voxel Proposal Layer and compute instance reference points in the scene volume. Specifically, we adopt the pre-trained Mobilestereonet [35] as the depth estimator, aligning with VoxFormer [22]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.473, + 0.351 + ], + "angle": 0, + "content": "Instance-Aware Image Encoder. The Instance-Aware Image Encoder, vital for integrating instance semantics in the absence of direct instance-level supervision, employs a Deformable Transformer [48] adept at capturing long-range dependencies around diverse instances by attending to deformable reference points. Additionally, it is augmented by utilizing the pre-trained weight of MaskDINO [18] from panoptic segmentation [17], to enrich its instance awareness." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.359, + 0.407, + 0.377 + ], + "angle": 0, + "content": "3.2. Depth-Rectified Voxel Proposal Layer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.473, + 0.487 + ], + "angle": 0, + "content": "The Depth-Rectified Voxel Proposal Layer generates initial scene features for voxels located on the implicit surface, known as voxel proposals, which establishes coarse geometry awareness for subsequent instance-level aggregations. The implicit surface is computed through the conversion of camera coordinates to world coordinates using depth estimation, described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.495, + 0.469, + 0.513 + ], + "angle": 0, + "content": "\\[\nx ^ {C} = \\mathcal {K} ^ {- 1} \\cdot \\left(z _ {c} \\odot x ^ {I}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.516, + 0.469, + 0.535 + ], + "angle": 0, + "content": "\\[\nx ^ {W} = [ R, T ] ^ {- 1} \\cdot x ^ {C} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.649 + ], + "angle": 0, + "content": "where \\( x^{I}, x^{C} \\), and \\( x^{W} \\) represent homogeneous coordinates of pixels, camera frustum, and the world, respectively. \\( \\odot \\) denotes the element-wise multiplication. The intrinsic matrix \\( \\mathcal{K} \\) encompasses camera parameters, while the extrinsic matrix is composed of the rotation matrix \\( R \\) and the translation vector \\( T \\). \\( z_{c} \\) corresponds to the \\( z \\)-coordinate of the camera, i.e., the depth estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.65, + 0.47, + 0.709 + ], + "angle": 0, + "content": "Based on the camera-to-world transformation, the positions \\( V_{p} \\) of voxel proposals are determined by mapping image points \\( x^{I} \\) to their corresponding world coordinates \\( x^{W} \\), confined within the volume \\( V \\):" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.716, + 0.469, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} V _ {p} = \\{x ^ {W} \\mid x ^ {W} = \\mathcal {T} ^ {I W} (x ^ {I}, z _ {c}), \\\\ \\forall x ^ {I} \\in I \\text {s u c h} x ^ {W} \\in V \\} \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.47, + 0.796 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{T}^{IW}\\) refers to the camera-to-world transformation, \\(I\\) represents image pixels, and \\(V\\) represents voxel grids." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, the determined voxel features are initialized by aggregating multi-scale image features using Deformable Attention [48]. This process involves selecting the proposed voxels \\( q_{p} \\) associated with the positions \\( V_{p} \\) from scene volume \\( q_{vox} \\), along with corresponding pixel positions \\( p_{I} \\) and 2D image features \\( F^{2D} \\). This process is expressed as \\( q_{p} = \\mathrm{DeformAttn}(q_{p}, p_{I}, F^{2D}) \\)." + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.09, + 0.861, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.255, + 0.894, + 0.283 + ], + "angle": 0, + "content": "Figure 3. Illustration of the Depth-Rectified Voxel Proposal Layer." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.309, + 0.894, + 0.37 + ], + "angle": 0, + "content": "The Deformable Attention operation, denoted as DeformAttn, dynamically aggregates query features \\( q \\) from features \\( x \\) with deformable reference points \\( p_q \\). The mathematical expression is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.38, + 0.892, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\operatorname {D e f o r m A t t n} (q, p _ {q}, x) = \\sum_ {k = 1} ^ {K} A _ {q k} W x \\left(p _ {q} + \\Delta p _ {q k}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.431, + 0.892, + 0.506 + ], + "angle": 0, + "content": "Here, \\( K \\) represents the number of sampling points, and \\( A_{qk} \\) stands for the learnable attention weight at sampling point \\( k \\) deformable based on queries \\( q \\). The term \\( \\Delta p_{qk} \\) denotes the offset applied to \\( p_q \\), and \\( W \\) denotes the projection weight. The computation of multi-heads is omitted for brevity." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.507, + 0.892, + 0.582 + ], + "angle": 0, + "content": "In contrast to the Query Proposal in VoxFormer [22], which employs an extra occupancy network [33] for generating coarse occupancy features, we refrain from it as it introduces additional geometric ambiguities in occlusion regions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.591, + 0.749, + 0.607 + ], + "angle": 0, + "content": "3.3. Symphonies Decoder Layer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.84 + ], + "angle": 0, + "content": "The Symphonies Decoder Layer seamlessly integrates image features, instance queries, and voxel proposals, as depicted in Fig. 2. It orchestrates a dynamic flow of information, where instance queries serve as intermediaries to propagate extracted instance semantics to the broader scene representations. The process initiates with deformable cross-attention modules, which attend to the corresponding instance positions within image and scene features to extract instance-centric semantics. Subsequently, the instance self-attention and scene-instance cross-attention modules strengthen the internal cohesion of instances and aggregate scene context from instance queries. The scene self-attention mechanism further diffuses voxel features throughout the scene, especially for the occluded regions, as only visible surfaces are initially proposed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The following paragraphs present a detailed explanation of the computations involved in their exact order of operation. To streamline the explanation, detailed discussions on certain components, including Feed-Forward Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20261" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "(FFN), Layer Norms (LN), and identity connections, have been omitted." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.149, + 0.468, + 0.226 + ], + "angle": 0, + "content": "Deformable Instance-Image Cross-Attention. For each instance query \\( q_{ins} \\), deformable attention extracts surrounding features from multi-scale image features \\( F^{2D} \\) using learnable 2D reference points \\( p_{ins}^{2D} \\), denoted as \\( q_{ins} = \\mathrm{DeformAttn}(q_{ins}, p_{ins}^{2D}, F^{2D}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.251, + 0.468, + 0.342 + ], + "angle": 0, + "content": "Scene-Instance Cross-Attention. This attention mechanism aggregates scene features \\( q_{vox} \\) from instance queries, formulated as \\( q_{vox}^{\\in FOV} = \\text{CrossAttn}(q_{vox}^{\\in FOV}, q_{ins}, q_{ins}) \\), where \\( FOV \\) refers to the \"field of view\" which is precomputed based on world-to-camera transformation excluding invisible voxels, reducing computational redundancy." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.368, + 0.468, + 0.443 + ], + "angle": 0, + "content": "Deformable Scene Self-Attention. The scene self-attention enables feature propagation across the scene, where voxels attend to their neighbors: \\( q_{vox}^{\\in FOV} = \\mathrm{DeformAttn}(q_{vox}^{\\in FOV}, p_V, q_{vox}) \\). Here, \\( p_V \\) represents voxels' relative coordinates in the scene." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.47, + 0.468, + 0.577 + ], + "angle": 0, + "content": "Deformable Instance-Scene Cross-Attention. Instance semantics are enhanced by integrating refined information from the reconstructed voxel features \\( q_{vox} \\). Through coordinate transformation applied to 2D reference points, 3D reference points are derived as \\( p_{ins}^{3D} = \\mathcal{T}^{IW}(p_{ins}^{2D}) \\). The instance-scene cross-attention is then formulated as \\( q_{ins} = \\mathrm{DeformAttn}(q_{ins}, p_{ins}^{3D}, q_{vox}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.603, + 0.468, + 0.649 + ], + "angle": 0, + "content": "Instance Self-Attention. The instance self-attention captures internal relations and global context within instance queries, expressed as \\( q_{ins} = \\mathrm{SelfAttn}(q_{ins}) \\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.662, + 0.165, + 0.677 + ], + "angle": 0, + "content": "3.4. Losses" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.687, + 0.468, + 0.791 + ], + "angle": 0, + "content": "In the Symphonies framework, we adopt the Scene-Class Affinity Loss \\( L_{scal} \\) from MonoScene [3] to optimize precision, recall, and specificity concurrently. The Scene-Class Affinity Loss is applied to semantic and geometric predictions, in conjunction with the cross-entropy loss weighted by class frequencies. The overall loss function is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.807, + 0.468, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {g e o}} + \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {s e m}} + \\mathcal {L} _ {\\text {c e}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Following the DETR series [4], auxiliary losses are applied after each Symphonies Decoder Layer for enhanced supervision, following the same formulation as \\(\\mathcal{L}\\) but scaled by a factor of 0.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.633, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.116, + 0.892, + 0.222 + ], + "angle": 0, + "content": "In this section, we present the evaluation results of our proposed Symphonies on SemanticKITTI [1] and SSCBenchKITTI-360 [21] datasets. The comparative analysis positioning Symphonies against existing approaches is detailed in Sec. 4.3. Additionally, comprehensive ablation studies are conducted in Sec. 4.4 to shed light on the thorough understanding of Symphonies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.233, + 0.685, + 0.247 + ], + "angle": 0, + "content": "4.1. Dataset and Metric" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.256, + 0.892, + 0.589 + ], + "angle": 0, + "content": "The evaluation is performed on SemanticKITTI [1] and SSCBench-KITTI-360 [21] datasets, both providing densely annotated urban driving scene sequences, 22 and 9 respectively, from the KITTI Odometry Benchmark [10]. These datasets voxelize point clouds and label scenes measuring \\(51.2m \\times 51.2m \\times 64m\\), with voxel grids of \\(256 \\times 256 \\times 32\\) and a voxel size of \\(0.2m\\). SemanticKITTI comprises 10 sequences for training, 1 sequence for validation, and 11 sequences for testing. It furnishes RGB images with shapes of \\(1226 \\times 370\\) as inputs and encompasses 20 semantic classes. SSCBench-KITTI-360 provides 7 sequences for training, 1 sequence for validation, and 1 sequence for testing, with 19 semantic classes and RGB images of size \\(1408 \\times 376\\). For our camera-based approach, we exclusively adopt RGB images as input, and report the intersection over union (IoU) and mean IoU (mIoU) metrics aligned with standard practices. The IoU metric assesses the binary classification of empty versus occupied voxels, reflecting the performance of geometric reconstruction. Conversely, the mIoU metric provides a comprehensive assessment of semantic understanding, making it the primary metric in most benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.6, + 0.716, + 0.616 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.624, + 0.892, + 0.746 + ], + "angle": 0, + "content": "In line with prior studies [3, 16, 22], we train Symphonies for 30 epochs on 4 NVIDIA 3090 GPUs, with a batch size of 4 images. We apply random horizontal flip augmentation and employ the AdamW [29] optimizer with an initial learning rate of 2e-4 and a weight decay of 1e-4. Learning rate reduction occurs by a factor of 0.1 at the 25th epoch. The ResNet-50 [12] backbone and Image Encoder are initialized with pre-trained MaskDINO [18] weights." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.639, + 0.77 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We conduct a comprehensive comparison of Symphonies with the latest state-of-the-art camera-based methodologies on the SemanticKITTI and SSCBench-KITTI-360 datasets. The results, outlined in Tab. 1 and Tab. 2, establish the superior performance of Symphonies with substantial improvements of 2.72 and \\(4.77\\mathrm{mIoU}\\) on SemanticKITTI and SSCBench-KITTI-360, respectively. Specifically, Symphonies showcases particular excellence in instance classes," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20262" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.09, + 0.89, + 0.288 + ], + "angle": 0, + "content": "
MethodIoUmIoUroadsidewalkparkingother-grd.buildingcartruckbicyclemotorcycleother-veh.vegetationtrunkterrainpersonbicyclistmotorcyclistfencepoletraf-sign
(15.30%)(11.15%)(12%)(0.56%)(14.1%)(3.92%)(16%)(0.03%)(0.03%)(0.03%)(0.20%)(3.3%)(0.51%)(17%)(0.07%)(0.07%)(0.05%)(3.90%)(0.29%)(0.08%)
LMSCNet†[33]31.387.0746.7019.5013.503.1010.3014.300.300.000.000.0010.800.0010.400.000.000.005.400.000.00
AICNet†[20]23.937.0939.3018.3019.801.609.6015.300.700.000.000.009.601.9013.500.000.000.005.000.100.00
JS3C-Net†[43]34.008.9747.3021.7019.902.8012.7020.100.800.000.004.1014.203.1012.400.000.200.208.701.900.30
MonoScene* [3]34.1611.0854.7027.1024.805.7014.4018.803.300.500.704.4014.902.4019.501.001.400.4011.103.302.10
TPVFormer [16]34.2511.2655.1027.2027.406.5014.8019.203.701.000.502.3013.902.6020.401.102.400.3011.002.901.50
VoxFormer [22]42.9512.2053.9025.3021.105.6019.8020.803.501.000.703.7022.407.5021.301.402.600.2011.105.104.90
OccFormer [46]34.5312.3255.9030.3031.506.5015.7021.601.201.501.703.2016.803.9021.302.201.100.2011.903.803.70
Symphonies42.1915.0458.4029.3026.9011.7024.7023.603.203.602.605.6024.2010.0023.103.201.902.0016.107.708.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.299, + 0.893, + 0.327 + ], + "angle": 0, + "content": "Table 1. Quantitative results on SemanticKITTI test. \\( {}^{ \\dagger } \\) denotes the results provided by [3]. * represents the reproduced results in [16, 46]. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.342, + 0.887, + 0.551 + ], + "angle": 0, + "content": "
MethodIoUPrec.Rec.mIoUcar(0.89%)bicycle(0.01%)motorcycle(0.03%)truck(0.16%)other-veh.(5.79%)person(0.02%)road(14.98%)parking(2.31%)sidewalk(6.43%)other-grd.(2.08%)building(15.67%)fence(0.96%)vegetation(41.99%)terrain(7.10%)pole(0.22%)traf.-sign(0.06%)other-struct.(4.33%)other-object.(0.28%)
LiDAR-based methods
SSCNet [37]53.5869.6369.9216.9531.950.000.1710.290.000.0765.7017.3341.243.2244.416.7743.7228.870.780.758.690.67
LMSCNet [33]47.5372.7757.5513.6520.910.000.000.260.580.0062.9513.5133.510.2043.670.3340.0126.800.000.003.630.00
Camera-based methods
MonoScene [3]37.8756.7353.2612.3119.340.430.588.022.030.8648.3511.3828.133.3232.893.5326.1516.756.925.674.203.09
TPVFormer [16]40.2259.3255.5413.6421.561.091.378.062.572.3852.9911.9931.073.7834.834.8030.0817.527.465.865.482.70
VoxFormer [22]38.7658.5253.4411.9117.841.160.894.562.061.6347.019.6727.212.8931.184.9728.9914.696.516.923.792.43
OccFormer [46]40.2759.7055.3113.8122.580.660.269.893.822.7754.3013.4431.533.5536.424.8031.0019.517.778.516.954.60
Symphonies44.1269.2454.8818.5830.021.855.9025.0712.068.2054.9413.8332.766.9335.118.5838.3311.5214.019.5714.4411.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.56, + 0.893, + 0.589 + ], + "angle": 0, + "content": "Table 2. Quantitative results on SSCBench-KITTI-360 test. The results for counterparts are provided in [21]. The best results among all methods are in bold, and the best results for camera-based methods are underlined." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.47, + 0.72 + ], + "angle": 0, + "content": "e.g., buildings, cars, etc. This underscores its prowess in capturing and modeling intricate instance semantics. While VoxFormer attains a marginally higher IoU on SemanticKITTI, its adoption of two-stage training and extra occupancy prediction network disrupts end-to-end training and introduces additional geometric ambiguities. This complexity hampers its robustness, especially on KITTI-360." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.47, + 0.83 + ], + "angle": 0, + "content": "The superiority of Symphonies becomes more pronounced on SSCBench-KITTI-360, which can be attributed to the ample data samples and high-quality annotations. Moreover, Symphonies even outperforms LiDAR-based methods in terms of mIoU, despite LiDAR's inherent advantage in IoU owing to its more precise position awareness, particularly at a distance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.242, + 0.86 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The ablation analysis is conducted on the SemanticKITTI validation set from four key perspectives: overall architec" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.615, + 0.892, + 0.647 + ], + "angle": 0, + "content": "tural components, the Symphonies Decoder, the Voxel Proposal Layer, and the Image Encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Ablation on architectural components. Tab. 3 presents the breakdown analysis of various architectural components within Symphonies. The vanilla baseline can be considered as a light-weight alternative to MonoScene, composed of a ResNet-50 backbone, an Image Encoder without pretrained weight, a 2D-to-3D projection via FLoSP [3], and a single 3D ASPP layer as the 3D decoder. Pre-training the Image Encoder leads to a notable improvement of 2.15 mIoU, emphasizing the effectiveness of instance awareness brought by 2D segmentation pre-training. Further, the proposed Depth-Rectified Voxel Proposal Layer improves performance by 0.75 mIoU through more accurate geometry. The Symphonies Decoder significantly boosts performance by 5.38 IoU, attributed to its dynamic instance modeling and context-capturing capabilities. In summary, the analy" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20263" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.192, + 0.089, + 0.778, + 0.183 + ], + "angle": 0, + "content": "
MethodIoUmIoUParams (M)FLOPs (G)
Baseline34.0610.4457.22529.20
+ Pre-trained Encoder35.97 (+1.91)12.59 (+2.15)57.22529.20
+ Voxel Proposal Layer36.54 (+0.57)13.34 (+0.75)57.42535.84
+ Symphonies Decoder41.92 (+5.38)14.89 (+1.55)59.31611.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.269, + 0.192, + 0.7, + 0.207 + ], + "angle": 0, + "content": "Table 3. Ablation study on architectural components in Symphonies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.233, + 0.471, + 0.264 + ], + "angle": 0, + "content": "sis in Tab. 3 affirms the effectiveness of the proposed components in Symphonies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.282, + 0.473, + 0.584 + ], + "angle": 0, + "content": "Ablation on the Symphonies Decoder. To gain insights into the functionality of contextual instance queries, we assess the modular interactions within the Symphonies Decoder Layer. As depicted in Tab. 4, the incorporation of instance queries with either instance-image or instance-scene cross-attention considerably enhances performance by over 5 IoU and about \\(1\\mathrm{mIoU}\\). This substantiates the significance of instance queries for adaptive aggregation of instance semantics. Among them, the instance-image cross-attention brings less improvement, suggesting that original image features have already been adequately captured in the preceding Voxel Proposal Layer. The instance self-attention further improve the performance, underlining the contextual effectiveness of efficient fusion among instance queries. Besides, though Scene SA may seem to only marginally improve the performance, its omission reveals significant fluctuations during training, attributed to the sparse interactions within instance deformable attentions. This underscores the pivotal role of Scene SA in fostering the generation of consistent features." + }, + { + "type": "table", + "bbox": [ + 0.083, + 0.595, + 0.47, + 0.734 + ], + "angle": 0, + "content": "
Scn.-Inst. CAInst.-Img. CAInst.-Scn. CAInst. SAScn. SAIoUmIoU
35.9713.34
✓✓41.3614.02
✓✓41.4414.37
✓✓✓41.3514.63
✓✓✓✓41.7514.73
✓✓✓✓✓41.9214.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.743, + 0.47, + 0.772 + ], + "angle": 0, + "content": "Table 4. Ablation study on Symphonies Decoder. SA: Self-Attention, CA: Cross-Attention." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Ablation on the Voxel Proposal Layer. Comparing the Depth-Rectified Voxel Proposal Layer (VPL) with FLoSP from MonoScene [3] casting pixels to voxels along the line of sight, as well as the mono VPL using monocular depth estimator AdaBins [2] (0.058 REL on KITTI), we note significant occupancy prediction improvements using" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.233, + 0.892, + 0.308 + ], + "angle": 0, + "content": "the stereo VPL based on MobileStereoNet [35] (0.66 EPE on KITTI 2015), as shown in Tab. 5. This indicates that the rectification of more precise depth estimation contributes to mitigating geometric ambiguities, aligning with the findings in VoxFormer." + }, + { + "type": "table", + "bbox": [ + 0.558, + 0.317, + 0.836, + 0.395 + ], + "angle": 0, + "content": "
2D-to-3D ProjectionIoUmIoU
FLoSP [3]36.0211.96
VPL (mono)38.3712.20
VPL (stereo)41.9214.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.405, + 0.892, + 0.42 + ], + "angle": 0, + "content": "Table 5. Ablation on the Depth-Rectified Voxel Proposal Layer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.454, + 0.892, + 0.604 + ], + "angle": 0, + "content": "Ablation on the Instance-Aware Image Encoder. Tab. 6 evidently showcases the synergistic effects of utilizing pre-trained weights for the Image Encoder with instance queries. Solely utilizing pre-trained weights from MaskDINO [18] contributes an additional improvement of \\(0.48\\mathrm{mIoU}\\). Moreover, incorporating instance queries with the pre-trained encoder yields a significant improvement of \\(1.36\\mathrm{mIoU}\\), implying that the proposed instance queries benefit from the enhanced instance awareness of the encoder." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.615, + 0.885, + 0.707 + ], + "angle": 0, + "content": "
Pre-trained EncoderInst. QueriesIoUmIoU
41.0913.53
✓41.4213.32
✓41.1814.01
✓✓41.9214.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.717, + 0.892, + 0.731 + ], + "angle": 0, + "content": "Table 6. Ablation study on the Instance-Aware Image Encoder." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.644, + 0.772 + ], + "angle": 0, + "content": "4.5. Visualizations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 4 presents the visualizations of Symphonies on SemanticKITTI val, in comparison to the counterpart MonoScene. Symphonies generate more detailed predictions for instance-centric classes such as cars and trunks, as well as preserves clear and coherent layouts for structures like buildings and vegetation, attributed to the enriched instance semantics and contextual information provided by instance queries. In contrast, MonoScene" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20264" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.098, + 0.34, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.096, + 0.498, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.095, + 0.643, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.094, + 0.805, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.18, + 0.339, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.171, + 0.498, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.171, + 0.642, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.17, + 0.806, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.258, + 0.338, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.25, + 0.492, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.257, + 0.641, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.25, + 0.797, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.337, + 0.338, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.393, + 0.266, + 0.405 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.325, + 0.492, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.393, + 0.459, + 0.404 + ], + "angle": 0, + "content": "MonoScene" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.324, + 0.638, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.393, + 0.61, + 0.405 + ], + "angle": 0, + "content": "Symphonies" + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.317, + 0.806, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.393, + 0.762, + 0.404 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.424, + 0.892, + 0.451 + ], + "angle": 0, + "content": "Figure 4. Qualitative visualizations on SemanticKITTI va1. Symphonies consistently produces detailed predictions for objects such as cars and trunks, while maintaining coherent layouts for structures like buildings and vegetation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.469, + 0.554 + ], + "angle": 0, + "content": "produces vague predictions with a radial shape, which is indicative of the aforementioned ambiguous geometry. These results underscore the superior capability of Symphonies in capturing fine-grained scene representations and enhancing overall scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.574, + 0.469, + 0.709 + ], + "angle": 0, + "content": "Attention Map Analysis. The attention map analysis in Fig. 5 provides insights into the mechanisms of the Serial Instance-Propagated Attentions within Symphonies layers. Notably, instance queries exhibit selective attention to corresponding regions in both the image and the scene. Additionally, they activate the semantically related regions within the scene-instance cross-attention. This observation validates the effect of our claimed instance-centric semantics in facilitating effective scene modeling." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.724, + 0.196, + 0.74 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduced Symphonies, a novel paradigm for 3D Semantic Scene Completion. Symphonies effectively integrates instance-centric semantics and scene context from both images and volumes, addressing the limitations posed by geometric ambiguity in prior voxel-wise modeling methods. Extensive experiments demonstrate the superiority of our approach over existing methods. We anticipate Symphonies to inspire future research and contribute to advancements in autonomous driving and 3D perception." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.475, + 0.747, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.537, + 0.727, + 0.546 + ], + "angle": 0, + "content": "(a) Reference points for Inst.-Image Cross-Attention." + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.477, + 0.887, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.545, + 0.887, + 0.603 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.614, + 0.887, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.77, + 0.677, + 0.887, + 0.693 + ], + "angle": 0, + "content": "(c) Attention Maps for SceneInst. Cross-Attention." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.552, + 0.747, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.676, + 0.727, + 0.685 + ], + "angle": 0, + "content": "(b) Reference points for Inst.-Scene Cross-Attention." + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.707, + 0.873, + 0.721 + ], + "angle": 0, + "content": "Figure 5. Analysis of attention maps within Symphonies." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.749, + 0.892, + 0.791 + ], + "angle": 0, + "content": "Acknowledgement. This work was partially supported by the National Natural Science Foundation of China (NSFC) under Grant No. 62276108." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.821, + 0.597, + 0.835 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Int. Conf. Comput. Vis., pages 9296-9306," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20265" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.093, + 0.175, + 0.105 + ], + "angle": 0, + "content": "2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.108, + 0.469, + 0.147 + ], + "angle": 0, + "content": "[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.149, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[3] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3981–3991, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.191, + 0.469, + 0.246 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Eur. Conf. Comput. Vis., pages 213-229, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.247, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., pages 834-848, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.317, + 0.469, + 0.383 + ], + "angle": 0, + "content": "[6] Shaoyu Chen, Tianheng Cheng, Xinggang Wang, Wenming Meng, Qian Zhang, and Wenyu Liu. Efficient and robust 2d-to-bev representation learning via geometry-guided kernel transformer. arXiv preprint arXiv:2206.04584, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.386, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[7] Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Chang Huang, and Wenyu Liu. Polar parametrization for vision-based surround-view 3d detection. arXiv preprint arXiv:2206.10965, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.442, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[8] Bowen Cheng, Alexander G. Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. In Adv. Neural Inform. Process. Syst., pages 17864-17875, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.498, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[9] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1280–1289, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.554, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In IEEE Conf. Comput. Vis. Pattern Recog., 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.596, + 0.469, + 0.636 + ], + "angle": 0, + "content": "[11] Yuxiao Guo and Xin Tong. View-volume network for semantic scene completion from a single depth image. In IJCAI, pages 726-732, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.638, + 0.469, + 0.678 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE Conf. Comput. Vis. Pattern Recog., pages 15979-15988, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[14] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. FIERY: future instance prediction in bird's-eye view from surround monocular cameras. In Int. Conf. Comput. Vis., pages 15253-15262, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.805, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[15] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[16] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9223-9232, 2023. 1, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[17] Alexander Kirillov, Ross B. Girshick, Kaiming He, and Piotr Dólar. Panoptic feature pyramid networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[18] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M. Ni, and Heung-Yeung Shum. Mask DINO: towards A unified transformer-based framework for object detection and segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3041-3050, 2023. 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[19] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian D. Reid. RGBD based dimensional decomposition residual network for 3d semantic scene completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7693-7702, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[20] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3348-3356, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[21] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. arXiv preprint arXiv:2306.09001, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.501 + ], + "angle": 0, + "content": "[22] Yiming Li, Zhiding Yu, Christopher B. Choy, Chaowei Xiao, José M. Álvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9087-9098, 2023. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Eur. Conf. Comput. Vis., pages 1-18, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[24] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Trans. Pattern Anal. Mach. Intell., pages 3292–3310, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[25] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. In Adv. Neural Inform. Process. Syst., pages 261-272, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[26] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETR: position embedding transformation for multi-view 3d object detection. In Eur. Conf. Comput. Vis., pages 531-548, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[27] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[28] Zhi Liu, Shaoyu Chen, Xiaojie Guo, Xinggang Wang, Tianheng Cheng, Hongmei Zhu, Qian Zhang, Wenyu Liu, and Yi Zhang. Vision-based uneven BEV representation learning with polar rasterization and surface estimation. In Conference on Robot Learning, pages 437-446, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Int. Conf. Learn. Represent., 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20266" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.147 + ], + "angle": 0, + "content": "[30] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.151, + 0.469, + 0.192 + ], + "angle": 0, + "content": "[31] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Eur. Conf. Comput. Vis., pages 194-210, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.194, + 0.469, + 0.235 + ], + "angle": 0, + "content": "[32] Thomas Roddick, Alex Kendall, and Roberto Cipolla. Orthographic feature transform for monocular 3d object detection. In Brit. Mach. Vis. Conf., page 285, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.239, + 0.468, + 0.279 + ], + "angle": 0, + "content": "[33] Luis Roldão, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 3DV, pages 111–119, 2020. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.282, + 0.468, + 0.322 + ], + "angle": 0, + "content": "[34] Luis Roldão, Raoul de Charette, and Anne Verroust-Blondet. 3d semantic scene completion: A survey. Int. J. Comput. Vis., 130(8):1978-2005, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.326, + 0.468, + 0.379 + ], + "angle": 0, + "content": "[35] Faranak Shamsafar, Samuel Woerz, Rafia Rahim, and Andreas Zell. Mobilestereonet: Towards lightweight deep networks for stereo matching. In WACV, pages 677-686, 2022. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.384, + 0.469, + 0.437 + ], + "angle": 0, + "content": "[36] Chonghao Sima, Wenwen Tong, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, and Hongyang Li. Scene as occupancy. arXiv preprint arXiv:2306.02851, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.442, + 0.468, + 0.496 + ], + "angle": 0, + "content": "[37] Shuran Song, Fisher Yu, Andy Zeng, Angel X. Chang, Manolis Savva, and Thomas A. Funkhouser. Semantic scene completion from a single depth image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 190-198, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.5, + 0.468, + 0.553 + ], + "angle": 0, + "content": "[38] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.557, + 0.468, + 0.611 + ], + "angle": 0, + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Adv. Neural Inform. Process. Syst., pages 5998-6008, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.615, + 0.468, + 0.655 + ], + "angle": 0, + "content": "[40] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. FCOS3D: fully convolutional one-stage monocular 3d object detection. In ICCVW, pages 913-922, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.659, + 0.468, + 0.713 + ], + "angle": 0, + "content": "[41] Yida Wang, David Joseph Tan, Nassir Navab, and Federico Tombari. Forknet: Multi-branch volumetric semantic completion from a single depth image. In Int. Conf. Comput. Vis., pages 8607–8616, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.716, + 0.468, + 0.771 + ], + "angle": 0, + "content": "[42] Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. DETR3D: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.774, + 0.468, + 0.829 + ], + "angle": 0, + "content": "[43] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.832, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[44] Jiawei Yao, Chuming Li, Keqiang Sun, Yingjie Cai, Hao Li, Wanli Ouyang, and Hongsheng Li. Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. arXiv preprint arXiv:2309.14616, 2023.2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[45] Pingping Zhang, Wei Liu, Yinjie Lei, Huchuan Lu, and Xiaoyun Yang. Cascaded context pyramid for full-resolution 3d semantic scene completion. In Int. Conf. Comput. Vis., pages 7800-7809, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[46] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2304.05316, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[47] Brady Zhou and Philipp Kähenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13750-13759, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[48] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In Int. Conf. Learn. Represent., 2021. 1, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20267" + } + ] +] \ No newline at end of file diff --git a/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/ba02978b-283d-434b-af7a-d4b604064e5e_origin.pdf b/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/ba02978b-283d-434b-af7a-d4b604064e5e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2692bfb88628c83866f2cc6b381a47d740f054d5 --- /dev/null +++ b/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/ba02978b-283d-434b-af7a-d4b604064e5e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:349edc15d2635a4dce65b012b04cdf3a5d806152723bb2dee86c3eb3d2fb5776 +size 9356726 diff --git a/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/full.md b/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eea96490aec2da91bca0b9edf11a909a585e512d --- /dev/null +++ b/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/full.md @@ -0,0 +1,316 @@ +# Symphonize 3D Semantic Scene Completion with Contextual Instance Queries + +Haoyi Jiang $^{1*}$ Tianheng Cheng $^{1*}$ Naiyu Gao $^{2}$ Haoyang Zhang $^{2}$ Tianwei Lin $^{2}$ Wenyu Liu $^{1}$ Xinggang Wang $^{1\dagger}$ $^{1}$ School of EIC, Huazhong University of Science & Technology + $^{2}$ Horizon Robotics +{haoyi_jiang,thch,liuwy,xgwang}@hust.edu.cn +{naiyu.gao,haoyang.zhang,tianwei.lin}@horizon.cc + +# Abstract + +3D Semantic Scene Completion (SSC) has emerged as a nascent and pivotal undertaking in autonomous driving, aiming to predict the voxel occupancy within volumetric scenes. However, prevailing methodologies primarily focus on voxel-wise feature aggregation, while neglecting instance semantics and scene context. In this paper, we present a novel paradigm termed Symphonies (Scene-from-Insts), that delves into the integration of instance queries to orchestrate 2D-to-3D reconstruction and 3D scene modeling. Leveraging our proposed Serial Instance-Propagated Attentions, Symphonies dynamically encodes instance-centric semantics, facilitating intricate interactions between the image and volumetric domains. Simultaneously, Symphonies fosters holistic scene comprehension by capturing context through the efficient fusion of instance queries, alleviating geometric ambiguities such as occlusion and perspective errors through contextual scene reasoning. Experimental results demonstrate that Symphonies achieves state-of-the-art performance on the challenging SemanticKITTI and SSCBench-KITTI-360 benchmarks, yielding remarkable mIoU scores of 15.04 and 18.58, respectively. These results showcase the promising advancements of our paradigm. The code for our method is available at https://github.com/hustvl/Symphonies. + +# 1. Introduction + +The advent of autonomous driving has brought forth novel challenges in the realm of 3D perception. In the pursuit of safe navigation and obstacle avoidance, autonomous vehicles must be equipped with the ability to accurately predict + +the occupancy of their immediate surroundings. This task, however, is not a facile endeavor, given the inherent complexities of the real world, characterized by clutter, ambiguity, and rapid evolution. + +3D Semantic Scene Completion (SSC) formulates this challenge as the reconstruction of occupancy and semantics for every voxel grid within a 3D scene. Recent advancements in vision-based solutions, such as MonoScene [3] and OccDepth [30], adopt 3D convolutional networks to elevate 2D image features to 3D volumes. TPVFormer [16], OccFormer [46], and CTF-Occ [38] explore decomposing 3D volumes into multiple coarse view representations and enhancing voxel interactions using Transformer [23, 39, 48] architectures. + +Despite these advancements, contemporary approaches tend to prioritize voxel-wise modeling for 3D scenes resorting to pixel-voxel projection [6, 23, 31, 32] for the feature promotion from 2D to 3D. While focusing on these localized representations, they inadvertently neglect higher-level instance semantics, leading to vulnerability to geometric ambiguities arising from occlusion and perspective errors. Humans, in contrast, naturally perceive and comprehend through the concept of "instance", rather than isolated pixels or voxels, each imbued with semantic significance and cohesively contributing to the contextual whole of a scene. In light of these limitations, a fundamental question arises: How can we leverage the notion of instances to steer 3D scene modeling and 2D-to-3D reconstruction? + +Drawing inspiration from this notion, we propose Symphonies (Scene-from-Insts), a novel method that leverages contextual instance queries derived from image inputs to enhance scene modeling, exploiting inherent instance semantics and scene context. Stemming from this basis, we propose Serial Instance-Propagated Attentions to intricately interact with image and voxel features, deformably aggregating instance-centric semantics. This seamless interaction bridges the gap between low-level pixel or voxel representations and high-level semantics, facilitating feature promo + +![](images/630672cb301e7db0fb7131c638c4c3eb3edc5d57345031a59388111516a40c3e.jpg) +Figure 1. Comparison between voxel-wise modeling (a) and Symphonies (b). Conventional methods primarily depend on Inverse Perspective Mapping (IPM)-based voxel-pixel projection and voxel-wise feature aggregation, resulting in geometric ambiguities and computational redundancy. In contrast, Symphonies leverages instance queries as intermediaries to engage with image and voxel features, thus exploiting instance semantics and enhancing the contextual comprehension of the scene. + +tion and scene modeling, as illustrated in Fig. 1. Furthermore, the synergistic fusion of multiple instance queries collectively enriches broader contextual information necessary for robust scene reasoning, contributing to the alleviation of geometric ambiguities. In tandem, we introduce the Depth-Rectified Voxel Proposal Layer to refine the initial geometry, elevating 2D image features to the implicit surface of the scene. + +To evaluate the effectiveness of our method, extensive experiments are conducted on SemanticKITTI [1] and SSCBench-KITTI-360 [21, 24] datasets. Symphonies achieves a remarkable state-of-the-art performance of 15.04 and $18.58\mathrm{mIoU}$ , respectively, significantly outperforming previous vision-based methods by a substantial margin. Ablation experiments further underscore the promising advancements of our approach in the field of SSC. In summary, our contributions involve: + +- We introduce Symphonies, a pioneering paradigm for 3D Semantic Scene Completion (SSC), which delves into modeling instance-centric semantics with our proposed Serial Instance-Propagated Attentions, facilitating interactions between image features and scene modeling. +- Symphonies captures global context of scenes through the fusion of instance queries, thereby alleviating geometric ambiguities. To further enhance robustness, we introduce the Depth-Rectified Voxel Proposal Layer which explicitly refines the initial geometry with estimated depth information. +- Our proposed method achieves state-of-the-art performance on the SemanticKITTI and SSCBench-KITTI-360 benchmarks, highlighting the substantial potential of this paradigm in advancing autonomous driving and 3D scene understanding. + +# 2. Related Works + +3D Semantic Scene Completion. 3D Semantic Scene Completion (SSC) entails predicting occupancy and semantics for each voxel within a 3D scene, which was initially introduced by SSCNet [37]. Subsequent methods can be broadly categorized based on their model architectures and input modalities. Volume networks [19, 45] predominantly utilize Truncated Signed Distance Function (TSDF) features generated from depth data, and process them through 3D convolutional networks. On the other hand, view-volume networks [11, 20, 25, 33, 41] extract RGB or depth features before lifting them to 3D volumes. For a more in-depth overview of SSC, we refer readers to the survey by Roldão et al. [34]. + +Recently, camera-based SSC has garnered increasing attention for its immense applications in autonomous driving. MonoScene [3] presents the first purely visual solution, sampling RGB features along the line of sight and adapting a 3D UNet architecture. TPVFormer [16] introduces a Tri-Perspective View (TPV) representation to decompose voxels onto multiple view planes for efficient scene encoding. VoxFormer [22] proposes a two-stage framework that diffuses the global scene from proposed voxel features, resembling the Masked Autoencoder (MAE) [13]. OccFormer [46] applies a mask-wise prediction paradigm akin to MaskFormer [8, 9]. OccDepth [30] and NDC-Scene [44] improves 2D-to-3D geometric projection by leveraging stereo depth and Normalized Device Coordinates (NDC). OccNet [36] further envisions occupancy as a general scene descriptor for a wide scope of driving tasks. + +In contrast to prior works, our proposed Symphonies differs by integrating instance queries to enhance scene model + +![](images/62ce4504b4438294c21678721d69642d5d90c74d7e505649b0386f905f7b0ad2.jpg) +Figure 2. Overview of Symphonies. The Symphonies framework commences with extracting multi-scale image features through the image backbone and Instance-Aware Image Encoder. The Depth-Rectified Voxel Proposal Layer generates initial voxel features estimating the implicit surface. Subsequently, the Symphonies Decoder Layers, which consist of Serial Instance-Propagated Attentions, facilitate continuous interactions among the image, instances, and the scene, iterated $N$ times. The Segmentation Head upsamples voxel features to the designated resolution and predicts class logits for each voxel. + +ing with instance semantics and enriched contextual awareness, mitigating geometric ambiguities arising from voxelwise modeling. + +Camera-Based 3D Perception. The surge in autonomous driving applications has rekindled interest in camera-based 3D perception, owing to its cost-effectiveness and alignment with human visual perception. Early 3D object detection methods, such as FCOS3D [40] and DETR3D [42], straightforwardly extend 2D detectors to predict additional 3D bounding boxes. Among subsequent Transformer-based approaches, BEVFormer [23] and BEVDet [15] adopt the BEV space to align multi-frame features, while PolarDETR [7] establishes explicit correlations between image patterns. In addition, PETR [26] and PETRv2 [27] utilize 3D position embeddings to encode 2D features. + +BEV segmentation, which is beneficial for representation learning and route planning, has also been explored. Approaches such as OFT [32], Lift-Splat [31], and FIERY [14] transform the camera plane into BEV via Inverse Perspective Mapping (IPM). PolarBEV [28] uses angle-specific and radius-specific embeddings to rasterize BEV features. BEVFormer [23] and CVT [47] aggregate BEV queries through cross-attention layers, while GKT [6] optimizes computational efficiency by constraining local attention calculations. + +These developments are closely related to our work in SSC, where techniques like Deformable Attention [48] inspire our methodology to enhance 3D scene completion. + +# 3. Scene from Instances + +This section presents a comprehensive elaboration of our proposed Symphonies method, commencing with an archi + +tectural overview in Sec. 3.1. Subsequently, it proceeds to detail the Depth-Rectified Voxel Proposal Layer in Sec. 3.2 and the Symphonies Decoder Layer in Sec. 3.3, shedding light on their synergistic contributions. Further insights into training losses are discussed in Sec. 3.4. + +# 3.1. Overview + +The architectural details of our proposed Symphonies are illustrated in Fig. 2. In essence, Symphonies exclusively takes RGB images as input and extracts multi-scale 2D features $F^{2D}$ through a ResNet-50 [12] image backbone and an Instance-Aware Deformable Transformer [48] Encoder, enhancing both global and instance semantics on the image plane. In the Symphonies Decoder, instance queries $q_{ins} \in \mathbb{R}^{N \times C}$ and the volumetric scene representation $q_{vox} \in \mathbb{R}^{C \times X \times Y \times Z}$ are initialized with learnable embeddings. Here, $C$ signifies embedding dimensions, $N$ denotes the number of instance queries, while $X$ , $Y$ , and $Z$ indicate the scene grid dimensions. + +The subsequent "scene-from-instances" process commences with the Depth-Rectified Voxel Proposal Layer initializing voxel proposals $q_{p}$ with image features on the implicit surface. Multi-scale image features $F^{2D}$ , scene features $q_{vox}$ , and instance queries $q_{ins}$ are then passed through our proposed Serial Instance-Propagated Attentions within the Symphonies Decoder Layers. This iterative process continuously propagates image features $F^{2D}$ to scene features $q_{vox}$ guided by instance queries $q_{ins}$ , while simultaneously aggregating instance semantics from both modalities. The Segmentation Head then upsamples the scene features to the desired resolution, and predicts per-voxel class logits with a single linear layer after an Atrous Spatial Pyramid Pooling (ASPP) [5] module. + +Depth Estimator. The depth prediction, obtained from a pre-trained depth estimator, is not explicitly illustrated in the diagram for clarity. It is employed to infer the implicit surface within the Voxel Proposal Layer and compute instance reference points in the scene volume. Specifically, we adopt the pre-trained Mobilestereonet [35] as the depth estimator, aligning with VoxFormer [22]. + +Instance-Aware Image Encoder. The Instance-Aware Image Encoder, vital for integrating instance semantics in the absence of direct instance-level supervision, employs a Deformable Transformer [48] adept at capturing long-range dependencies around diverse instances by attending to deformable reference points. Additionally, it is augmented by utilizing the pre-trained weight of MaskDINO [18] from panoptic segmentation [17], to enrich its instance awareness. + +# 3.2. Depth-Rectified Voxel Proposal Layer + +The Depth-Rectified Voxel Proposal Layer generates initial scene features for voxels located on the implicit surface, known as voxel proposals, which establishes coarse geometry awareness for subsequent instance-level aggregations. The implicit surface is computed through the conversion of camera coordinates to world coordinates using depth estimation, described as follows: + +$$ +x ^ {C} = \mathcal {K} ^ {- 1} \cdot \left(z _ {c} \odot x ^ {I}\right) \tag {1} +$$ + +$$ +x ^ {W} = [ R, T ] ^ {- 1} \cdot x ^ {C} \tag {2} +$$ + +where $x^{I}, x^{C}$ , and $x^{W}$ represent homogeneous coordinates of pixels, camera frustum, and the world, respectively. $\odot$ denotes the element-wise multiplication. The intrinsic matrix $\mathcal{K}$ encompasses camera parameters, while the extrinsic matrix is composed of the rotation matrix $R$ and the translation vector $T$ . $z_{c}$ corresponds to the $z$ -coordinate of the camera, i.e., the depth estimation. + +Based on the camera-to-world transformation, the positions $V_{p}$ of voxel proposals are determined by mapping image points $x^{I}$ to their corresponding world coordinates $x^{W}$ , confined within the volume $V$ : + +$$ +\begin{array}{l} V _ {p} = \{x ^ {W} \mid x ^ {W} = \mathcal {T} ^ {I W} (x ^ {I}, z _ {c}), \\ \forall x ^ {I} \in I \text {s u c h} x ^ {W} \in V \} \tag {3} \\ \end{array} +$$ + +Here, $\mathcal{T}^{IW}$ refers to the camera-to-world transformation, $I$ represents image pixels, and $V$ represents voxel grids. + +As illustrated in Fig. 3, the determined voxel features are initialized by aggregating multi-scale image features using Deformable Attention [48]. This process involves selecting the proposed voxels $q_{p}$ associated with the positions $V_{p}$ from scene volume $q_{vox}$ , along with corresponding pixel positions $p_{I}$ and 2D image features $F^{2D}$ . This process is expressed as $q_{p} = \mathrm{DeformAttn}(q_{p}, p_{I}, F^{2D})$ . + +![](images/f0947edef9d5cb0fac912803284fd8a506f61ca7d9ed52824dc1f4c2319a4997.jpg) +Figure 3. Illustration of the Depth-Rectified Voxel Proposal Layer. + +The Deformable Attention operation, denoted as DeformAttn, dynamically aggregates query features $q$ from features $x$ with deformable reference points $p_q$ . The mathematical expression is given by: + +$$ +\operatorname {D e f o r m A t t n} (q, p _ {q}, x) = \sum_ {k = 1} ^ {K} A _ {q k} W x \left(p _ {q} + \Delta p _ {q k}\right) \tag {4} +$$ + +Here, $K$ represents the number of sampling points, and $A_{qk}$ stands for the learnable attention weight at sampling point $k$ deformable based on queries $q$ . The term $\Delta p_{qk}$ denotes the offset applied to $p_q$ , and $W$ denotes the projection weight. The computation of multi-heads is omitted for brevity. + +In contrast to the Query Proposal in VoxFormer [22], which employs an extra occupancy network [33] for generating coarse occupancy features, we refrain from it as it introduces additional geometric ambiguities in occlusion regions. + +# 3.3. Symphonies Decoder Layer + +The Symphonies Decoder Layer seamlessly integrates image features, instance queries, and voxel proposals, as depicted in Fig. 2. It orchestrates a dynamic flow of information, where instance queries serve as intermediaries to propagate extracted instance semantics to the broader scene representations. The process initiates with deformable cross-attention modules, which attend to the corresponding instance positions within image and scene features to extract instance-centric semantics. Subsequently, the instance self-attention and scene-instance cross-attention modules strengthen the internal cohesion of instances and aggregate scene context from instance queries. The scene self-attention mechanism further diffuses voxel features throughout the scene, especially for the occluded regions, as only visible surfaces are initially proposed. + +The following paragraphs present a detailed explanation of the computations involved in their exact order of operation. To streamline the explanation, detailed discussions on certain components, including Feed-Forward Networks + +(FFN), Layer Norms (LN), and identity connections, have been omitted. + +Deformable Instance-Image Cross-Attention. For each instance query $q_{ins}$ , deformable attention extracts surrounding features from multi-scale image features $F^{2D}$ using learnable 2D reference points $p_{ins}^{2D}$ , denoted as $q_{ins} = \mathrm{DeformAttn}(q_{ins}, p_{ins}^{2D}, F^{2D})$ . + +Scene-Instance Cross-Attention. This attention mechanism aggregates scene features $q_{vox}$ from instance queries, formulated as $q_{vox}^{\in FOV} = \text{CrossAttn}(q_{vox}^{\in FOV}, q_{ins}, q_{ins})$ , where $FOV$ refers to the "field of view" which is precomputed based on world-to-camera transformation excluding invisible voxels, reducing computational redundancy. + +Deformable Scene Self-Attention. The scene self-attention enables feature propagation across the scene, where voxels attend to their neighbors: $q_{vox}^{\in FOV} = \mathrm{DeformAttn}(q_{vox}^{\in FOV}, p_V, q_{vox})$ . Here, $p_V$ represents voxels' relative coordinates in the scene. + +Deformable Instance-Scene Cross-Attention. Instance semantics are enhanced by integrating refined information from the reconstructed voxel features $q_{vox}$ . Through coordinate transformation applied to 2D reference points, 3D reference points are derived as $p_{ins}^{3D} = \mathcal{T}^{IW}(p_{ins}^{2D})$ . The instance-scene cross-attention is then formulated as $q_{ins} = \mathrm{DeformAttn}(q_{ins}, p_{ins}^{3D}, q_{vox})$ . + +Instance Self-Attention. The instance self-attention captures internal relations and global context within instance queries, expressed as $q_{ins} = \mathrm{SelfAttn}(q_{ins})$ . + +# 3.4. Losses + +In the Symphonies framework, we adopt the Scene-Class Affinity Loss $L_{scal}$ from MonoScene [3] to optimize precision, recall, and specificity concurrently. The Scene-Class Affinity Loss is applied to semantic and geometric predictions, in conjunction with the cross-entropy loss weighted by class frequencies. The overall loss function is formulated as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {s c a l}} ^ {\text {g e o}} + \mathcal {L} _ {\text {s c a l}} ^ {\text {s e m}} + \mathcal {L} _ {\text {c e}} \tag {5} +$$ + +Following the DETR series [4], auxiliary losses are applied after each Symphonies Decoder Layer for enhanced supervision, following the same formulation as $\mathcal{L}$ but scaled by a factor of 0.5. + +# 4. Experiments + +In this section, we present the evaluation results of our proposed Symphonies on SemanticKITTI [1] and SSCBenchKITTI-360 [21] datasets. The comparative analysis positioning Symphonies against existing approaches is detailed in Sec. 4.3. Additionally, comprehensive ablation studies are conducted in Sec. 4.4 to shed light on the thorough understanding of Symphonies. + +# 4.1. Dataset and Metric + +The evaluation is performed on SemanticKITTI [1] and SSCBench-KITTI-360 [21] datasets, both providing densely annotated urban driving scene sequences, 22 and 9 respectively, from the KITTI Odometry Benchmark [10]. These datasets voxelize point clouds and label scenes measuring $51.2m \times 51.2m \times 64m$ , with voxel grids of $256 \times 256 \times 32$ and a voxel size of $0.2m$ . SemanticKITTI comprises 10 sequences for training, 1 sequence for validation, and 11 sequences for testing. It furnishes RGB images with shapes of $1226 \times 370$ as inputs and encompasses 20 semantic classes. SSCBench-KITTI-360 provides 7 sequences for training, 1 sequence for validation, and 1 sequence for testing, with 19 semantic classes and RGB images of size $1408 \times 376$ . For our camera-based approach, we exclusively adopt RGB images as input, and report the intersection over union (IoU) and mean IoU (mIoU) metrics aligned with standard practices. The IoU metric assesses the binary classification of empty versus occupied voxels, reflecting the performance of geometric reconstruction. Conversely, the mIoU metric provides a comprehensive assessment of semantic understanding, making it the primary metric in most benchmarks. + +# 4.2. Implementation Details + +In line with prior studies [3, 16, 22], we train Symphonies for 30 epochs on 4 NVIDIA 3090 GPUs, with a batch size of 4 images. We apply random horizontal flip augmentation and employ the AdamW [29] optimizer with an initial learning rate of 2e-4 and a weight decay of 1e-4. Learning rate reduction occurs by a factor of 0.1 at the 25th epoch. The ResNet-50 [12] backbone and Image Encoder are initialized with pre-trained MaskDINO [18] weights. + +# 4.3. Main Results + +We conduct a comprehensive comparison of Symphonies with the latest state-of-the-art camera-based methodologies on the SemanticKITTI and SSCBench-KITTI-360 datasets. The results, outlined in Tab. 1 and Tab. 2, establish the superior performance of Symphonies with substantial improvements of 2.72 and $4.77\mathrm{mIoU}$ on SemanticKITTI and SSCBench-KITTI-360, respectively. Specifically, Symphonies showcases particular excellence in instance classes, + +
MethodIoUmIoUroadsidewalkparkingother-grd.buildingcartruckbicyclemotorcycleother-veh.vegetationtrunkterrainpersonbicyclistmotorcyclistfencepoletraf-sign
(15.30%)(11.15%)(12%)(0.56%)(14.1%)(3.92%)(16%)(0.03%)(0.03%)(0.03%)(0.20%)(3.3%)(0.51%)(17%)(0.07%)(0.07%)(0.05%)(3.90%)(0.29%)(0.08%)
LMSCNet†[33]31.387.0746.7019.5013.503.1010.3014.300.300.000.000.0010.800.0010.400.000.000.005.400.000.00
AICNet†[20]23.937.0939.3018.3019.801.609.6015.300.700.000.000.009.601.9013.500.000.000.005.000.100.00
JS3C-Net†[43]34.008.9747.3021.7019.902.8012.7020.100.800.000.004.1014.203.1012.400.000.200.208.701.900.30
MonoScene* [3]34.1611.0854.7027.1024.805.7014.4018.803.300.500.704.4014.902.4019.501.001.400.4011.103.302.10
TPVFormer [16]34.2511.2655.1027.2027.406.5014.8019.203.701.000.502.3013.902.6020.401.102.400.3011.002.901.50
VoxFormer [22]42.9512.2053.9025.3021.105.6019.8020.803.501.000.703.7022.407.5021.301.402.600.2011.105.104.90
OccFormer [46]34.5312.3255.9030.3031.506.5015.7021.601.201.501.703.2016.803.9021.302.201.100.2011.903.803.70
Symphonies42.1915.0458.4029.3026.9011.7024.7023.603.203.602.605.6024.2010.0023.103.201.902.0016.107.708.00
+ +Table 1. Quantitative results on SemanticKITTI test. ${}^{ \dagger }$ denotes the results provided by [3]. * represents the reproduced results in [16, 46]. The best results are in bold. + +
MethodIoUPrec.Rec.mIoUcar(0.89%)bicycle(0.01%)motorcycle(0.03%)truck(0.16%)other-veh.(5.79%)person(0.02%)road(14.98%)parking(2.31%)sidewalk(6.43%)other-grd.(2.08%)building(15.67%)fence(0.96%)vegetation(41.99%)terrain(7.10%)pole(0.22%)traf.-sign(0.06%)other-struct.(4.33%)other-object.(0.28%)
LiDAR-based methods
SSCNet [37]53.5869.6369.9216.9531.950.000.1710.290.000.0765.7017.3341.243.2244.416.7743.7228.870.780.758.690.67
LMSCNet [33]47.5372.7757.5513.6520.910.000.000.260.580.0062.9513.5133.510.2043.670.3340.0126.800.000.003.630.00
Camera-based methods
MonoScene [3]37.8756.7353.2612.3119.340.430.588.022.030.8648.3511.3828.133.3232.893.5326.1516.756.925.674.203.09
TPVFormer [16]40.2259.3255.5413.6421.561.091.378.062.572.3852.9911.9931.073.7834.834.8030.0817.527.465.865.482.70
VoxFormer [22]38.7658.5253.4411.9117.841.160.894.562.061.6347.019.6727.212.8931.184.9728.9914.696.516.923.792.43
OccFormer [46]40.2759.7055.3113.8122.580.660.269.893.822.7754.3013.4431.533.5536.424.8031.0019.517.778.516.954.60
Symphonies44.1269.2454.8818.5830.021.855.9025.0712.068.2054.9413.8332.766.9335.118.5838.3311.5214.019.5714.4411.28
+ +Table 2. Quantitative results on SSCBench-KITTI-360 test. The results for counterparts are provided in [21]. The best results among all methods are in bold, and the best results for camera-based methods are underlined. + +e.g., buildings, cars, etc. This underscores its prowess in capturing and modeling intricate instance semantics. While VoxFormer attains a marginally higher IoU on SemanticKITTI, its adoption of two-stage training and extra occupancy prediction network disrupts end-to-end training and introduces additional geometric ambiguities. This complexity hampers its robustness, especially on KITTI-360. + +The superiority of Symphonies becomes more pronounced on SSCBench-KITTI-360, which can be attributed to the ample data samples and high-quality annotations. Moreover, Symphonies even outperforms LiDAR-based methods in terms of mIoU, despite LiDAR's inherent advantage in IoU owing to its more precise position awareness, particularly at a distance. + +# 4.4. Ablation Studies + +The ablation analysis is conducted on the SemanticKITTI validation set from four key perspectives: overall architec + +tural components, the Symphonies Decoder, the Voxel Proposal Layer, and the Image Encoder. + +Ablation on architectural components. Tab. 3 presents the breakdown analysis of various architectural components within Symphonies. The vanilla baseline can be considered as a light-weight alternative to MonoScene, composed of a ResNet-50 backbone, an Image Encoder without pretrained weight, a 2D-to-3D projection via FLoSP [3], and a single 3D ASPP layer as the 3D decoder. Pre-training the Image Encoder leads to a notable improvement of 2.15 mIoU, emphasizing the effectiveness of instance awareness brought by 2D segmentation pre-training. Further, the proposed Depth-Rectified Voxel Proposal Layer improves performance by 0.75 mIoU through more accurate geometry. The Symphonies Decoder significantly boosts performance by 5.38 IoU, attributed to its dynamic instance modeling and context-capturing capabilities. In summary, the analy + +
MethodIoUmIoUParams (M)FLOPs (G)
Baseline34.0610.4457.22529.20
+ Pre-trained Encoder35.97 (+1.91)12.59 (+2.15)57.22529.20
+ Voxel Proposal Layer36.54 (+0.57)13.34 (+0.75)57.42535.84
+ Symphonies Decoder41.92 (+5.38)14.89 (+1.55)59.31611.89
+ +sis in Tab. 3 affirms the effectiveness of the proposed components in Symphonies. + +Ablation on the Symphonies Decoder. To gain insights into the functionality of contextual instance queries, we assess the modular interactions within the Symphonies Decoder Layer. As depicted in Tab. 4, the incorporation of instance queries with either instance-image or instance-scene cross-attention considerably enhances performance by over 5 IoU and about $1\mathrm{mIoU}$ . This substantiates the significance of instance queries for adaptive aggregation of instance semantics. Among them, the instance-image cross-attention brings less improvement, suggesting that original image features have already been adequately captured in the preceding Voxel Proposal Layer. The instance self-attention further improve the performance, underlining the contextual effectiveness of efficient fusion among instance queries. Besides, though Scene SA may seem to only marginally improve the performance, its omission reveals significant fluctuations during training, attributed to the sparse interactions within instance deformable attentions. This underscores the pivotal role of Scene SA in fostering the generation of consistent features. + +Table 3. Ablation study on architectural components in Symphonies. + +
Scn.-Inst. CAInst.-Img. CAInst.-Scn. CAInst. SAScn. SAIoUmIoU
35.9713.34
✓✓41.3614.02
✓✓41.4414.37
✓✓✓41.3514.63
✓✓✓✓41.7514.73
✓✓✓✓✓41.9214.89
+ +Ablation on the Voxel Proposal Layer. Comparing the Depth-Rectified Voxel Proposal Layer (VPL) with FLoSP from MonoScene [3] casting pixels to voxels along the line of sight, as well as the mono VPL using monocular depth estimator AdaBins [2] (0.058 REL on KITTI), we note significant occupancy prediction improvements using + +the stereo VPL based on MobileStereoNet [35] (0.66 EPE on KITTI 2015), as shown in Tab. 5. This indicates that the rectification of more precise depth estimation contributes to mitigating geometric ambiguities, aligning with the findings in VoxFormer. + +Table 4. Ablation study on Symphonies Decoder. SA: Self-Attention, CA: Cross-Attention. + +
2D-to-3D ProjectionIoUmIoU
FLoSP [3]36.0211.96
VPL (mono)38.3712.20
VPL (stereo)41.9214.89
+ +Ablation on the Instance-Aware Image Encoder. Tab. 6 evidently showcases the synergistic effects of utilizing pre-trained weights for the Image Encoder with instance queries. Solely utilizing pre-trained weights from MaskDINO [18] contributes an additional improvement of $0.48\mathrm{mIoU}$ . Moreover, incorporating instance queries with the pre-trained encoder yields a significant improvement of $1.36\mathrm{mIoU}$ , implying that the proposed instance queries benefit from the enhanced instance awareness of the encoder. + +Table 5. Ablation on the Depth-Rectified Voxel Proposal Layer. + +
Pre-trained EncoderInst. QueriesIoUmIoU
41.0913.53
✓41.4213.32
✓41.1814.01
✓✓41.9214.89
+ +Table 6. Ablation study on the Instance-Aware Image Encoder. + +# 4.5. Visualizations + +Qualitative Results. Fig. 4 presents the visualizations of Symphonies on SemanticKITTI val, in comparison to the counterpart MonoScene. Symphonies generate more detailed predictions for instance-centric classes such as cars and trunks, as well as preserves clear and coherent layouts for structures like buildings and vegetation, attributed to the enriched instance semantics and contextual information provided by instance queries. In contrast, MonoScene + +![](images/2d125133181974605ba41e81768fed2c41d55e725119c7d980555382e68ecb1e.jpg) + +![](images/d80d39c1169085adc21f138a11d94fdb390b242ed226bc7f7a31dab032d2a834.jpg) + +![](images/e5d086f8d2607486274a2e6b6f0e8a054f6ec998aefba9130a931773540a3f12.jpg) + +![](images/874ff1dd621234463cc660c3e0abc3f05dea52661a496c44c33d3fc54e5a5686.jpg) + +![](images/aeb6af285715e91e4bf615cdf13bb2a9ce3558efc0ab239191bf82264d159a54.jpg) + +![](images/5f0d6f135247d721faf5164791f1ba8ebaf11e8d1f9e62930483a99f57687011.jpg) + +![](images/9e0c1293852789933f2c00d3608686483e558fd6d78314b907cc107c294ddbec.jpg) + +![](images/42d247771d3edc90d32153779ad5fd0c102c650d86e78e94a9254f78efada1e3.jpg) + +![](images/c8c12316824ed975ca94e4514e2d8a8b5d2726ad87b4543883aac8da378b7f24.jpg) + +![](images/977752bda281ea6d7792c4385dda1021ea78b96cc3a08742be44429ed6e5a353.jpg) + +![](images/f3d616aaf9ffc708c8cce5ef02b07e7f6e9c55ae48011db097993f180cfb9b92.jpg) + +![](images/9646790bfd3561b68d44d152ea33a363e9d00a311420173515c98ed6342402c3.jpg) + +![](images/2db1dcff58e97e7f1da12db3b21e3e4758a38d2eb5fa94ee7a13e490bcce3017.jpg) +Input + +![](images/e5486a991f32fd983fa5d82609ce9a60cfa431bd15c7e71f8d0300840adcef9b.jpg) +MonoScene + +![](images/a59bd518684cea3d24af0b9635cd34b7e98e8d0002b2bb32bef6ddd576bef73c.jpg) +Symphonies + +![](images/c98e999f12b80536673b01ebcd39c1a6cf14591a7f4796d2021342f72de41452.jpg) +Ground Truth + +produces vague predictions with a radial shape, which is indicative of the aforementioned ambiguous geometry. These results underscore the superior capability of Symphonies in capturing fine-grained scene representations and enhancing overall scene understanding. + +Attention Map Analysis. The attention map analysis in Fig. 5 provides insights into the mechanisms of the Serial Instance-Propagated Attentions within Symphonies layers. Notably, instance queries exhibit selective attention to corresponding regions in both the image and the scene. Additionally, they activate the semantically related regions within the scene-instance cross-attention. This observation validates the effect of our claimed instance-centric semantics in facilitating effective scene modeling. + +# 5. Conclusion + +In this paper, we introduced Symphonies, a novel paradigm for 3D Semantic Scene Completion. Symphonies effectively integrates instance-centric semantics and scene context from both images and volumes, addressing the limitations posed by geometric ambiguity in prior voxel-wise modeling methods. Extensive experiments demonstrate the superiority of our approach over existing methods. We anticipate Symphonies to inspire future research and contribute to advancements in autonomous driving and 3D perception. + +![](images/630d16c9f1c967ae06bf8293c98243d69b046a7a71a30ff338e2fc4199ade2d6.jpg) +Figure 4. Qualitative visualizations on SemanticKITTI va1. Symphonies consistently produces detailed predictions for objects such as cars and trunks, while maintaining coherent layouts for structures like buildings and vegetation. + +![](images/db37814759ec8622a1ec5a5752cadba636b31f820fe1b68e9f028a388357e04a.jpg) + +![](images/519e8fba28247ac876fe658f51762bf4b9687dfa6a85b7a3bc4b5dd9a7a048e2.jpg) + +![](images/ff1209a9176fee41c965e16ee4755887c423f9bc9dc8cc5520491605e22d94c2.jpg) +(c) Attention Maps for SceneInst. Cross-Attention. + +![](images/bd5af940c866a1d9da648039955057db242f30594779b19299162508106fcdbf.jpg) +(a) Reference points for Inst.-Image Cross-Attention. +(b) Reference points for Inst.-Scene Cross-Attention. +Figure 5. Analysis of attention maps within Symphonies. + +Acknowledgement. This work was partially supported by the National Natural Science Foundation of China (NSFC) under Grant No. 62276108. + +# References + +[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Int. Conf. Comput. Vis., pages 9296-9306, + +2019. 2, 5 +[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 7 +[3] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3981–3991, 2022. 1, 2, 5, 6, 7 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Eur. Conf. Comput. Vis., pages 213-229, 2020. 5 +[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., pages 834-848, 2018. 3 +[6] Shaoyu Chen, Tianheng Cheng, Xinggang Wang, Wenming Meng, Qian Zhang, and Wenyu Liu. Efficient and robust 2d-to-bev representation learning via geometry-guided kernel transformer. arXiv preprint arXiv:2206.04584, 2022. 1, 3 +[7] Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Chang Huang, and Wenyu Liu. Polar parametrization for vision-based surround-view 3d detection. arXiv preprint arXiv:2206.10965, 2022. 3 +[8] Bowen Cheng, Alexander G. Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. In Adv. Neural Inform. Process. Syst., pages 17864-17875, 2021. 2 +[9] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1280–1289, 2022. 2 +[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In IEEE Conf. Comput. Vis. Pattern Recog., 2012. 5 +[11] Yuxiao Guo and Xin Tong. View-volume network for semantic scene completion from a single depth image. In IJCAI, pages 726-732, 2018. 2 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 3, 5 +[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE Conf. Comput. Vis. Pattern Recog., pages 15979-15988, 2022. 2 +[14] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. FIERY: future instance prediction in bird's-eye view from surround monocular cameras. In Int. Conf. Comput. Vis., pages 15253-15262, 2021. 3 +[15] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 3 +[16] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9223-9232, 2023. 1, 2, 5, 6 + +[17] Alexander Kirillov, Ross B. Girshick, Kaiming He, and Piotr Dólar. Panoptic feature pyramid networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 4 +[18] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M. Ni, and Heung-Yeung Shum. Mask DINO: towards A unified transformer-based framework for object detection and segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3041-3050, 2023. 4, 5, 7 +[19] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian D. Reid. RGBD based dimensional decomposition residual network for 3d semantic scene completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7693-7702, 2019. 2 +[20] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3348-3356, 2020. 2, 6 +[21] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. arXiv preprint arXiv:2306.09001, 2023. 2, 5, 6 +[22] Yiming Li, Zhiding Yu, Christopher B. Choy, Chaowei Xiao, José M. Álvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9087-9098, 2023. 2, 4, 5, 6 +[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Eur. Conf. Comput. Vis., pages 1-18, 2022. 1, 3 +[24] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Trans. Pattern Anal. Mach. Intell., pages 3292–3310, 2023. 2 +[25] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. In Adv. Neural Inform. Process. Syst., pages 261-272, 2018. 2 +[26] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETR: position embedding transformation for multi-view 3d object detection. In Eur. Conf. Comput. Vis., pages 531-548, 2022. 3 +[27] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. 3 +[28] Zhi Liu, Shaoyu Chen, Xiaojie Guo, Xinggang Wang, Tianheng Cheng, Hongmei Zhu, Qian Zhang, Wenyu Liu, and Yi Zhang. Vision-based uneven BEV representation learning with polar rasterization and surface estimation. In Conference on Robot Learning, pages 437-446, 2022. 3 +[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Int. Conf. Learn. Represent., 2019. 5 + +[30] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 1, 2 +[31] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Eur. Conf. Comput. Vis., pages 194-210, 2020. 1, 3 +[32] Thomas Roddick, Alex Kendall, and Roberto Cipolla. Orthographic feature transform for monocular 3d object detection. In Brit. Mach. Vis. Conf., page 285, 2019. 1, 3 +[33] Luis Roldão, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 3DV, pages 111–119, 2020. 2, 4, 6 +[34] Luis Roldão, Raoul de Charette, and Anne Verroust-Blondet. 3d semantic scene completion: A survey. Int. J. Comput. Vis., 130(8):1978-2005, 2022. 2 +[35] Faranak Shamsafar, Samuel Woerz, Rafia Rahim, and Andreas Zell. Mobilestereonet: Towards lightweight deep networks for stereo matching. In WACV, pages 677-686, 2022. 4, 7 +[36] Chonghao Sima, Wenwen Tong, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, and Hongyang Li. Scene as occupancy. arXiv preprint arXiv:2306.02851, 2023. 2 +[37] Shuran Song, Fisher Yu, Andy Zeng, Angel X. Chang, Manolis Savva, and Thomas A. Funkhouser. Semantic scene completion from a single depth image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 190-198, 2017. 2, 6 +[38] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 1 +[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Adv. Neural Inform. Process. Syst., pages 5998-6008, 2017. 1 +[40] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. FCOS3D: fully convolutional one-stage monocular 3d object detection. In ICCVW, pages 913-922, 2021. 3 +[41] Yida Wang, David Joseph Tan, Nassir Navab, and Federico Tombari. Forknet: Multi-branch volumetric semantic completion from a single depth image. In Int. Conf. Comput. Vis., pages 8607–8616, 2019. 2 +[42] Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. DETR3D: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191, 2021. 3 +[43] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 6 +[44] Jiawei Yao, Chuming Li, Keqiang Sun, Yingjie Cai, Hao Li, Wanli Ouyang, and Hongsheng Li. Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. arXiv preprint arXiv:2309.14616, 2023.2 + +[45] Pingping Zhang, Wei Liu, Yinjie Lei, Huchuan Lu, and Xiaoyun Yang. Cascaded context pyramid for full-resolution 3d semantic scene completion. In Int. Conf. Comput. Vis., pages 7800-7809, 2019. 2 +[46] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2304.05316, 2023. 1, 2, 6 +[47] Brady Zhou and Philipp Kähenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13750-13759, 2022. 3 +[48] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In Int. Conf. Learn. Represent., 2021. 1, 3, 4 \ No newline at end of file diff --git a/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/images.zip b/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..449362eb271f4ed362015617fee6a5367c2fd66d --- /dev/null +++ b/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e971f05a7beefca91c76416025006913658fbcda1149c8a9ca5b19cc3fc634a7 +size 672234 diff --git a/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/layout.json b/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f3a7467f67e4c05a0de8b11100cc785a6e01297e --- /dev/null +++ b/2024/Symphonize 3D Semantic Scene Completion with Contextual Instance Queries/layout.json @@ -0,0 +1,8589 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 55, + 103, + 538, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 538, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 538, + 121 + ], + "type": "text", + "content": "Symphonize 3D Semantic Scene Completion with Contextual Instance Queries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "spans": [ + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "text", + "content": "Haoyi Jiang" + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "text", + "content": " Tianheng Cheng" + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "text", + "content": " Naiyu Gao" + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "text", + "content": " Haoyang Zhang" + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "text", + "content": " Tianwei Lin" + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "text", + "content": " Wenyu Liu" + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "text", + "content": " Xinggang Wang" + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "text", + "content": " School of EIC, Huazhong University of Science & Technology \n" + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 89, + 141, + 504, + 228 + ], + "type": "text", + "content": " Horizon Robotics \n{haoyi_jiang,thch,liuwy,xgwang}@hust.edu.cn \n{naiyu.gao,haoyang.zhang,tianwei.lin}@horizon.cc" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 280, + 290, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 290, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 290, + 582 + ], + "type": "text", + "content": "3D Semantic Scene Completion (SSC) has emerged as a nascent and pivotal undertaking in autonomous driving, aiming to predict the voxel occupancy within volumetric scenes. However, prevailing methodologies primarily focus on voxel-wise feature aggregation, while neglecting instance semantics and scene context. In this paper, we present a novel paradigm termed Symphonies (Scene-from-Insts), that delves into the integration of instance queries to orchestrate 2D-to-3D reconstruction and 3D scene modeling. Leveraging our proposed Serial Instance-Propagated Attentions, Symphonies dynamically encodes instance-centric semantics, facilitating intricate interactions between the image and volumetric domains. Simultaneously, Symphonies fosters holistic scene comprehension by capturing context through the efficient fusion of instance queries, alleviating geometric ambiguities such as occlusion and perspective errors through contextual scene reasoning. Experimental results demonstrate that Symphonies achieves state-of-the-art performance on the challenging SemanticKITTI and SSCBench-KITTI-360 benchmarks, yielding remarkable mIoU scores of 15.04 and 18.58, respectively. These results showcase the promising advancements of our paradigm. The code for our method is available at https://github.com/hustvl/Symphonies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 604, + 128, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 128, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 128, + 617 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 624, + 288, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 288, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 288, + 674 + ], + "type": "text", + "content": "The advent of autonomous driving has brought forth novel challenges in the realm of 3D perception. In the pursuit of safe navigation and obstacle avoidance, autonomous vehicles must be equipped with the ability to accurately predict" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 256, + 545, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 256, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 306, + 256, + 545, + 303 + ], + "type": "text", + "content": "the occupancy of their immediate surroundings. This task, however, is not a facile endeavor, given the inherent complexities of the real world, characterized by clutter, ambiguity, and rapid evolution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 304, + 546, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 546, + 422 + ], + "type": "text", + "content": "3D Semantic Scene Completion (SSC) formulates this challenge as the reconstruction of occupancy and semantics for every voxel grid within a 3D scene. Recent advancements in vision-based solutions, such as MonoScene [3] and OccDepth [30], adopt 3D convolutional networks to elevate 2D image features to 3D volumes. TPVFormer [16], OccFormer [46], and CTF-Occ [38] explore decomposing 3D volumes into multiple coarse view representations and enhancing voxel interactions using Transformer [23, 39, 48] architectures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 426, + 547, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 593 + ], + "type": "text", + "content": "Despite these advancements, contemporary approaches tend to prioritize voxel-wise modeling for 3D scenes resorting to pixel-voxel projection [6, 23, 31, 32] for the feature promotion from 2D to 3D. While focusing on these localized representations, they inadvertently neglect higher-level instance semantics, leading to vulnerability to geometric ambiguities arising from occlusion and perspective errors. Humans, in contrast, naturally perceive and comprehend through the concept of \"instance\", rather than isolated pixels or voxels, each imbued with semantic significance and cohesively contributing to the contextual whole of a scene. In light of these limitations, a fundamental question arises: How can we leverage the notion of instances to steer 3D scene modeling and 2D-to-3D reconstruction?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Drawing inspiration from this notion, we propose Symphonies (Scene-from-Insts), a novel method that leverages contextual instance queries derived from image inputs to enhance scene modeling, exploiting inherent instance semantics and scene context. Stemming from this basis, we propose Serial Instance-Propagated Attentions to intricately interact with image and voxel features, deformably aggregating instance-centric semantics. This seamless interaction bridges the gap between low-level pixel or voxel representations and high-level semantics, facilitating feature promo" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 683, + 288, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 288, + 713 + ], + "type": "text", + "content": "*Equal contribution. Work done during Haoyi Jiang's internship at Horizon Robotics. \n†Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20258" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 125, + 69, + 470, + 241 + ], + "blocks": [ + { + "bbox": [ + 125, + 69, + 470, + 241 + ], + "lines": [ + { + "bbox": [ + 125, + 69, + 470, + 241 + ], + "spans": [ + { + "bbox": [ + 125, + 69, + 470, + 241 + ], + "type": "image", + "image_path": "630672cb301e7db0fb7131c638c4c3eb3edc5d57345031a59388111516a40c3e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 248, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 248, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 547, + 293 + ], + "type": "text", + "content": "Figure 1. Comparison between voxel-wise modeling (a) and Symphonies (b). Conventional methods primarily depend on Inverse Perspective Mapping (IPM)-based voxel-pixel projection and voxel-wise feature aggregation, resulting in geometric ambiguities and computational redundancy. In contrast, Symphonies leverages instance queries as intermediaries to engage with image and voxel features, thus exploiting instance semantics and enhancing the contextual comprehension of the scene." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 314, + 288, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 288, + 409 + ], + "type": "text", + "content": "tion and scene modeling, as illustrated in Fig. 1. Furthermore, the synergistic fusion of multiple instance queries collectively enriches broader contextual information necessary for robust scene reasoning, contributing to the alleviation of geometric ambiguities. In tandem, we introduce the Depth-Rectified Voxel Proposal Layer to refine the initial geometry, elevating 2D image features to the implicit surface of the scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 411, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 518 + ], + "type": "text", + "content": "To evaluate the effectiveness of our method, extensive experiments are conducted on SemanticKITTI [1] and SSCBench-KITTI-360 [21, 24] datasets. Symphonies achieves a remarkable state-of-the-art performance of 15.04 and " + }, + { + "bbox": [ + 46, + 411, + 287, + 518 + ], + "type": "inline_equation", + "content": "18.58\\mathrm{mIoU}" + }, + { + "bbox": [ + 46, + 411, + 287, + 518 + ], + "type": "text", + "content": ", respectively, significantly outperforming previous vision-based methods by a substantial margin. Ablation experiments further underscore the promising advancements of our approach in the field of SSC. In summary, our contributions involve:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 521, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 287, + 582 + ], + "type": "text", + "content": "- We introduce Symphonies, a pioneering paradigm for 3D Semantic Scene Completion (SSC), which delves into modeling instance-centric semantics with our proposed Serial Instance-Propagated Attentions, facilitating interactions between image features and scene modeling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 582, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 652 + ], + "type": "text", + "content": "- Symphonies captures global context of scenes through the fusion of instance queries, thereby alleviating geometric ambiguities. To further enhance robustness, we introduce the Depth-Rectified Voxel Proposal Layer which explicitly refines the initial geometry with estimated depth information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "text", + "content": "- Our proposed method achieves state-of-the-art performance on the SemanticKITTI and SSCBench-KITTI-360 benchmarks, highlighting the substantial potential of this paradigm in advancing autonomous driving and 3D scene understanding." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 312, + 398, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 398, + 325 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 398, + 325 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 335, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 491 + ], + "type": "text", + "content": "3D Semantic Scene Completion. 3D Semantic Scene Completion (SSC) entails predicting occupancy and semantics for each voxel within a 3D scene, which was initially introduced by SSCNet [37]. Subsequent methods can be broadly categorized based on their model architectures and input modalities. Volume networks [19, 45] predominantly utilize Truncated Signed Distance Function (TSDF) features generated from depth data, and process them through 3D convolutional networks. On the other hand, view-volume networks [11, 20, 25, 33, 41] extract RGB or depth features before lifting them to 3D volumes. For a more in-depth overview of SSC, we refer readers to the survey by Roldão et al. [34]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 495, + 546, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 546, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 546, + 687 + ], + "type": "text", + "content": "Recently, camera-based SSC has garnered increasing attention for its immense applications in autonomous driving. MonoScene [3] presents the first purely visual solution, sampling RGB features along the line of sight and adapting a 3D UNet architecture. TPVFormer [16] introduces a Tri-Perspective View (TPV) representation to decompose voxels onto multiple view planes for efficient scene encoding. VoxFormer [22] proposes a two-stage framework that diffuses the global scene from proposed voxel features, resembling the Masked Autoencoder (MAE) [13]. OccFormer [46] applies a mask-wise prediction paradigm akin to MaskFormer [8, 9]. OccDepth [30] and NDC-Scene [44] improves 2D-to-3D geometric projection by leveraging stereo depth and Normalized Device Coordinates (NDC). OccNet [36] further envisions occupancy as a general scene descriptor for a wide scope of driving tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "In contrast to prior works, our proposed Symphonies differs by integrating instance queries to enhance scene model" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20259" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 547, + 222 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 547, + 222 + ], + "type": "image", + "image_path": "62ce4504b4438294c21678721d69642d5d90c74d7e505649b0386f905f7b0ad2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 228, + 546, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 546, + 285 + ], + "type": "text", + "content": "Figure 2. Overview of Symphonies. The Symphonies framework commences with extracting multi-scale image features through the image backbone and Instance-Aware Image Encoder. The Depth-Rectified Voxel Proposal Layer generates initial voxel features estimating the implicit surface. Subsequently, the Symphonies Decoder Layers, which consist of Serial Instance-Propagated Attentions, facilitate continuous interactions among the image, instances, and the scene, iterated " + }, + { + "bbox": [ + 46, + 228, + 546, + 285 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 228, + 546, + 285 + ], + "type": "text", + "content": " times. The Segmentation Head upsamples voxel features to the designated resolution and predicts class logits for each voxel." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 304, + 288, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 288, + 341 + ], + "type": "text", + "content": "ing with instance semantics and enriched contextual awareness, mitigating geometric ambiguities arising from voxelwise modeling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 357, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 501 + ], + "type": "text", + "content": "Camera-Based 3D Perception. The surge in autonomous driving applications has rekindled interest in camera-based 3D perception, owing to its cost-effectiveness and alignment with human visual perception. Early 3D object detection methods, such as FCOS3D [40] and DETR3D [42], straightforwardly extend 2D detectors to predict additional 3D bounding boxes. Among subsequent Transformer-based approaches, BEVFormer [23] and BEVDet [15] adopt the BEV space to align multi-frame features, while PolarDETR [7] establishes explicit correlations between image patterns. In addition, PETR [26] and PETRv2 [27] utilize 3D position embeddings to encode 2D features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 501, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 620 + ], + "type": "text", + "content": "BEV segmentation, which is beneficial for representation learning and route planning, has also been explored. Approaches such as OFT [32], Lift-Splat [31], and FIERY [14] transform the camera plane into BEV via Inverse Perspective Mapping (IPM). PolarBEV [28] uses angle-specific and radius-specific embeddings to rasterize BEV features. BEVFormer [23] and CVT [47] aggregate BEV queries through cross-attention layers, while GKT [6] optimizes computational efficiency by constraining local attention calculations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 621, + 288, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 288, + 657 + ], + "type": "text", + "content": "These developments are closely related to our work in SSC, where techniques like Deformable Attention [48] inspire our methodology to enhance 3D scene completion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 669, + 171, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 171, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 171, + 681 + ], + "type": "text", + "content": "3. Scene from Instances" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "This section presents a comprehensive elaboration of our proposed Symphonies method, commencing with an archi" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 304, + 547, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 547, + 365 + ], + "type": "text", + "content": "tectural overview in Sec. 3.1. Subsequently, it proceeds to detail the Depth-Rectified Voxel Proposal Layer in Sec. 3.2 and the Symphonies Decoder Layer in Sec. 3.3, shedding light on their synergistic contributions. Further insights into training losses are discussed in Sec. 3.4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 379, + 373, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 379, + 373, + 390 + ], + "spans": [ + { + "bbox": [ + 306, + 379, + 373, + 390 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "text", + "content": "The architectural details of our proposed Symphonies are illustrated in Fig. 2. In essence, Symphonies exclusively takes RGB images as input and extracts multi-scale 2D features " + }, + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "inline_equation", + "content": "F^{2D}" + }, + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "text", + "content": " through a ResNet-50 [12] image backbone and an Instance-Aware Deformable Transformer [48] Encoder, enhancing both global and instance semantics on the image plane. In the Symphonies Decoder, instance queries " + }, + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "inline_equation", + "content": "q_{ins} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "text", + "content": " and the volumetric scene representation " + }, + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "inline_equation", + "content": "q_{vox} \\in \\mathbb{R}^{C \\times X \\times Y \\times Z}" + }, + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "text", + "content": " are initialized with learnable embeddings. Here, " + }, + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "text", + "content": " signifies embedding dimensions, " + }, + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "text", + "content": " denotes the number of instance queries, while " + }, + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 399, + 546, + 544 + ], + "type": "text", + "content": " indicate the scene grid dimensions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "content": "The subsequent \"scene-from-instances\" process commences with the Depth-Rectified Voxel Proposal Layer initializing voxel proposals " + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "inline_equation", + "content": "q_{p}" + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "content": " with image features on the implicit surface. Multi-scale image features " + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "inline_equation", + "content": "F^{2D}" + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "content": ", scene features " + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "inline_equation", + "content": "q_{vox}" + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "content": ", and instance queries " + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "inline_equation", + "content": "q_{ins}" + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "content": " are then passed through our proposed Serial Instance-Propagated Attentions within the Symphonies Decoder Layers. This iterative process continuously propagates image features " + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "inline_equation", + "content": "F^{2D}" + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "content": " to scene features " + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "inline_equation", + "content": "q_{vox}" + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "content": " guided by instance queries " + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "inline_equation", + "content": "q_{ins}" + }, + { + "bbox": [ + 304, + 545, + 547, + 714 + ], + "type": "text", + "content": ", while simultaneously aggregating instance semantics from both modalities. The Segmentation Head then upsamples the scene features to the desired resolution, and predicts per-voxel class logits with a single linear layer after an Atrous Spatial Pyramid Pooling (ASPP) [5] module." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20260" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "Depth Estimator. The depth prediction, obtained from a pre-trained depth estimator, is not explicitly illustrated in the diagram for clarity. It is employed to infer the implicit surface within the Voxel Proposal Layer and compute instance reference points in the scene volume. Specifically, we adopt the pre-trained Mobilestereonet [35] as the depth estimator, aligning with VoxFormer [22]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 289, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 289, + 277 + ], + "type": "text", + "content": "Instance-Aware Image Encoder. The Instance-Aware Image Encoder, vital for integrating instance semantics in the absence of direct instance-level supervision, employs a Deformable Transformer [48] adept at capturing long-range dependencies around diverse instances by attending to deformable reference points. Additionally, it is augmented by utilizing the pre-trained weight of MaskDINO [18] from panoptic segmentation [17], to enrich its instance awareness." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 284, + 249, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 249, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 249, + 298 + ], + "type": "text", + "content": "3.2. Depth-Rectified Voxel Proposal Layer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 302, + 289, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 289, + 385 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 289, + 385 + ], + "type": "text", + "content": "The Depth-Rectified Voxel Proposal Layer generates initial scene features for voxels located on the implicit surface, known as voxel proposals, which establishes coarse geometry awareness for subsequent instance-level aggregations. The implicit surface is computed through the conversion of camera coordinates to world coordinates using depth estimation, described as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 392, + 287, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 392, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 118, + 392, + 287, + 406 + ], + "type": "interline_equation", + "content": "x ^ {C} = \\mathcal {K} ^ {- 1} \\cdot \\left(z _ {c} \\odot x ^ {I}\\right) \\tag {1}", + "image_path": "f90b8cd6c448572ee42e31debfb22bac1d210d7a26841a7b288a01df0f312590.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 118, + 408, + 287, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 408, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 118, + 408, + 287, + 423 + ], + "type": "interline_equation", + "content": "x ^ {W} = [ R, T ] ^ {- 1} \\cdot x ^ {C} \\tag {2}", + "image_path": "09b5b6b40dd0fc5db40e16d82b0495ff607574a038f446f9d8db1bb5e06f3b01.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "inline_equation", + "content": "x^{I}, x^{C}" + }, + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "inline_equation", + "content": "x^{W}" + }, + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "text", + "content": " represent homogeneous coordinates of pixels, camera frustum, and the world, respectively. " + }, + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "text", + "content": " denotes the element-wise multiplication. The intrinsic matrix " + }, + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "text", + "content": " encompasses camera parameters, while the extrinsic matrix is composed of the rotation matrix " + }, + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "text", + "content": " and the translation vector " + }, + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "inline_equation", + "content": "z_{c}" + }, + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "text", + "content": " corresponds to the " + }, + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 429, + 287, + 514 + ], + "type": "text", + "content": "-coordinate of the camera, i.e., the depth estimation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 514, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 561 + ], + "type": "text", + "content": "Based on the camera-to-world transformation, the positions " + }, + { + "bbox": [ + 46, + 514, + 287, + 561 + ], + "type": "inline_equation", + "content": "V_{p}" + }, + { + "bbox": [ + 46, + 514, + 287, + 561 + ], + "type": "text", + "content": " of voxel proposals are determined by mapping image points " + }, + { + "bbox": [ + 46, + 514, + 287, + 561 + ], + "type": "inline_equation", + "content": "x^{I}" + }, + { + "bbox": [ + 46, + 514, + 287, + 561 + ], + "type": "text", + "content": " to their corresponding world coordinates " + }, + { + "bbox": [ + 46, + 514, + 287, + 561 + ], + "type": "inline_equation", + "content": "x^{W}" + }, + { + "bbox": [ + 46, + 514, + 287, + 561 + ], + "type": "text", + "content": ", confined within the volume " + }, + { + "bbox": [ + 46, + 514, + 287, + 561 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 514, + 287, + 561 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 81, + 567, + 287, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 567, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 81, + 567, + 287, + 597 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} V _ {p} = \\{x ^ {W} \\mid x ^ {W} = \\mathcal {T} ^ {I W} (x ^ {I}, z _ {c}), \\\\ \\forall x ^ {I} \\in I \\text {s u c h} x ^ {W} \\in V \\} \\tag {3} \\\\ \\end{array}", + "image_path": "00154518cef3c5d134470ee7badba45b6773d2eea82961e89ce7f3227eb17724.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 287, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 630 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 605, + 287, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{IW}" + }, + { + "bbox": [ + 46, + 605, + 287, + 630 + ], + "type": "text", + "content": " refers to the camera-to-world transformation, " + }, + { + "bbox": [ + 46, + 605, + 287, + 630 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 605, + 287, + 630 + ], + "type": "text", + "content": " represents image pixels, and " + }, + { + "bbox": [ + 46, + 605, + 287, + 630 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 605, + 287, + 630 + ], + "type": "text", + "content": " represents voxel grids." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "As illustrated in Fig. 3, the determined voxel features are initialized by aggregating multi-scale image features using Deformable Attention [48]. This process involves selecting the proposed voxels " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "q_{p}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " associated with the positions " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "V_{p}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " from scene volume " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "q_{vox}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ", along with corresponding pixel positions " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "p_{I}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " and 2D image features " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "F^{2D}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ". This process is expressed as " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "q_{p} = \\mathrm{DeformAttn}(q_{p}, p_{I}, F^{2D})" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 326, + 71, + 526, + 194 + ], + "blocks": [ + { + "bbox": [ + 326, + 71, + 526, + 194 + ], + "lines": [ + { + "bbox": [ + 326, + 71, + 526, + 194 + ], + "spans": [ + { + "bbox": [ + 326, + 71, + 526, + 194 + ], + "type": "image", + "image_path": "f0947edef9d5cb0fac912803284fd8a506f61ca7d9ed52824dc1f4c2319a4997.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 201, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 305, + 201, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 547, + 224 + ], + "type": "text", + "content": "Figure 3. Illustration of the Depth-Rectified Voxel Proposal Layer." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 244, + 547, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 244, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 244, + 547, + 293 + ], + "type": "text", + "content": "The Deformable Attention operation, denoted as DeformAttn, dynamically aggregates query features " + }, + { + "bbox": [ + 305, + 244, + 547, + 293 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 305, + 244, + 547, + 293 + ], + "type": "text", + "content": " from features " + }, + { + "bbox": [ + 305, + 244, + 547, + 293 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 244, + 547, + 293 + ], + "type": "text", + "content": " with deformable reference points " + }, + { + "bbox": [ + 305, + 244, + 547, + 293 + ], + "type": "inline_equation", + "content": "p_q" + }, + { + "bbox": [ + 305, + 244, + 547, + 293 + ], + "type": "text", + "content": ". The mathematical expression is given by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 300, + 545, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 300, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 318, + 300, + 545, + 334 + ], + "type": "interline_equation", + "content": "\\operatorname {D e f o r m A t t n} (q, p _ {q}, x) = \\sum_ {k = 1} ^ {K} A _ {q k} W x \\left(p _ {q} + \\Delta p _ {q k}\\right) \\tag {4}", + "image_path": "63698db3e811f2e7e11cd5a5cf28620ec6417cf732e4882836fb7167f7ccc7e9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 341, + 545, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 341, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 305, + 341, + 545, + 400 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 305, + 341, + 545, + 400 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 341, + 545, + 400 + ], + "type": "text", + "content": " represents the number of sampling points, and " + }, + { + "bbox": [ + 305, + 341, + 545, + 400 + ], + "type": "inline_equation", + "content": "A_{qk}" + }, + { + "bbox": [ + 305, + 341, + 545, + 400 + ], + "type": "text", + "content": " stands for the learnable attention weight at sampling point " + }, + { + "bbox": [ + 305, + 341, + 545, + 400 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 341, + 545, + 400 + ], + "type": "text", + "content": " deformable based on queries " + }, + { + "bbox": [ + 305, + 341, + 545, + 400 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 305, + 341, + 545, + 400 + ], + "type": "text", + "content": ". The term " + }, + { + "bbox": [ + 305, + 341, + 545, + 400 + ], + "type": "inline_equation", + "content": "\\Delta p_{qk}" + }, + { + "bbox": [ + 305, + 341, + 545, + 400 + ], + "type": "text", + "content": " denotes the offset applied to " + }, + { + "bbox": [ + 305, + 341, + 545, + 400 + ], + "type": "inline_equation", + "content": "p_q" + }, + { + "bbox": [ + 305, + 341, + 545, + 400 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 341, + 545, + 400 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 305, + 341, + 545, + 400 + ], + "type": "text", + "content": " denotes the projection weight. The computation of multi-heads is omitted for brevity." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 401, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 401, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 401, + 545, + 460 + ], + "type": "text", + "content": "In contrast to the Query Proposal in VoxFormer [22], which employs an extra occupancy network [33] for generating coarse occupancy features, we refrain from it as it introduces additional geometric ambiguities in occlusion regions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 468, + 458, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 458, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 458, + 480 + ], + "type": "text", + "content": "3.3. Symphonies Decoder Layer" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 486, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 665 + ], + "type": "text", + "content": "The Symphonies Decoder Layer seamlessly integrates image features, instance queries, and voxel proposals, as depicted in Fig. 2. It orchestrates a dynamic flow of information, where instance queries serve as intermediaries to propagate extracted instance semantics to the broader scene representations. The process initiates with deformable cross-attention modules, which attend to the corresponding instance positions within image and scene features to extract instance-centric semantics. Subsequently, the instance self-attention and scene-instance cross-attention modules strengthen the internal cohesion of instances and aggregate scene context from instance queries. The scene self-attention mechanism further diffuses voxel features throughout the scene, especially for the occluded regions, as only visible surfaces are initially proposed." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "The following paragraphs present a detailed explanation of the computations involved in their exact order of operation. To streamline the explanation, detailed discussions on certain components, including Feed-Forward Networks" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20261" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "(FFN), Layer Norms (LN), and identity connections, have been omitted." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 118, + 286, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 118, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 47, + 118, + 286, + 178 + ], + "type": "text", + "content": "Deformable Instance-Image Cross-Attention. For each instance query " + }, + { + "bbox": [ + 47, + 118, + 286, + 178 + ], + "type": "inline_equation", + "content": "q_{ins}" + }, + { + "bbox": [ + 47, + 118, + 286, + 178 + ], + "type": "text", + "content": ", deformable attention extracts surrounding features from multi-scale image features " + }, + { + "bbox": [ + 47, + 118, + 286, + 178 + ], + "type": "inline_equation", + "content": "F^{2D}" + }, + { + "bbox": [ + 47, + 118, + 286, + 178 + ], + "type": "text", + "content": " using learnable 2D reference points " + }, + { + "bbox": [ + 47, + 118, + 286, + 178 + ], + "type": "inline_equation", + "content": "p_{ins}^{2D}" + }, + { + "bbox": [ + 47, + 118, + 286, + 178 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 47, + 118, + 286, + 178 + ], + "type": "inline_equation", + "content": "q_{ins} = \\mathrm{DeformAttn}(q_{ins}, p_{ins}^{2D}, F^{2D})" + }, + { + "bbox": [ + 47, + 118, + 286, + 178 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 198, + 286, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 286, + 270 + ], + "type": "text", + "content": "Scene-Instance Cross-Attention. This attention mechanism aggregates scene features " + }, + { + "bbox": [ + 47, + 198, + 286, + 270 + ], + "type": "inline_equation", + "content": "q_{vox}" + }, + { + "bbox": [ + 47, + 198, + 286, + 270 + ], + "type": "text", + "content": " from instance queries, formulated as " + }, + { + "bbox": [ + 47, + 198, + 286, + 270 + ], + "type": "inline_equation", + "content": "q_{vox}^{\\in FOV} = \\text{CrossAttn}(q_{vox}^{\\in FOV}, q_{ins}, q_{ins})" + }, + { + "bbox": [ + 47, + 198, + 286, + 270 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 198, + 286, + 270 + ], + "type": "inline_equation", + "content": "FOV" + }, + { + "bbox": [ + 47, + 198, + 286, + 270 + ], + "type": "text", + "content": " refers to the \"field of view\" which is precomputed based on world-to-camera transformation excluding invisible voxels, reducing computational redundancy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 291, + 286, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 286, + 350 + ], + "type": "text", + "content": "Deformable Scene Self-Attention. The scene self-attention enables feature propagation across the scene, where voxels attend to their neighbors: " + }, + { + "bbox": [ + 47, + 291, + 286, + 350 + ], + "type": "inline_equation", + "content": "q_{vox}^{\\in FOV} = \\mathrm{DeformAttn}(q_{vox}^{\\in FOV}, p_V, q_{vox})" + }, + { + "bbox": [ + 47, + 291, + 286, + 350 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 47, + 291, + 286, + 350 + ], + "type": "inline_equation", + "content": "p_V" + }, + { + "bbox": [ + 47, + 291, + 286, + 350 + ], + "type": "text", + "content": " represents voxels' relative coordinates in the scene." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 372, + 286, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 286, + 456 + ], + "type": "text", + "content": "Deformable Instance-Scene Cross-Attention. Instance semantics are enhanced by integrating refined information from the reconstructed voxel features " + }, + { + "bbox": [ + 47, + 372, + 286, + 456 + ], + "type": "inline_equation", + "content": "q_{vox}" + }, + { + "bbox": [ + 47, + 372, + 286, + 456 + ], + "type": "text", + "content": ". Through coordinate transformation applied to 2D reference points, 3D reference points are derived as " + }, + { + "bbox": [ + 47, + 372, + 286, + 456 + ], + "type": "inline_equation", + "content": "p_{ins}^{3D} = \\mathcal{T}^{IW}(p_{ins}^{2D})" + }, + { + "bbox": [ + 47, + 372, + 286, + 456 + ], + "type": "text", + "content": ". The instance-scene cross-attention is then formulated as " + }, + { + "bbox": [ + 47, + 372, + 286, + 456 + ], + "type": "inline_equation", + "content": "q_{ins} = \\mathrm{DeformAttn}(q_{ins}, p_{ins}^{3D}, q_{vox})" + }, + { + "bbox": [ + 47, + 372, + 286, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 477, + 286, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 477, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 286, + 514 + ], + "type": "text", + "content": "Instance Self-Attention. The instance self-attention captures internal relations and global context within instance queries, expressed as " + }, + { + "bbox": [ + 47, + 477, + 286, + 514 + ], + "type": "inline_equation", + "content": "q_{ins} = \\mathrm{SelfAttn}(q_{ins})" + }, + { + "bbox": [ + 47, + 477, + 286, + 514 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 524, + 100, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 524, + 100, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 524, + 100, + 536 + ], + "type": "text", + "content": "3.4. Losses" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 544, + 286, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 286, + 626 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 286, + 626 + ], + "type": "text", + "content": "In the Symphonies framework, we adopt the Scene-Class Affinity Loss " + }, + { + "bbox": [ + 47, + 544, + 286, + 626 + ], + "type": "inline_equation", + "content": "L_{scal}" + }, + { + "bbox": [ + 47, + 544, + 286, + 626 + ], + "type": "text", + "content": " from MonoScene [3] to optimize precision, recall, and specificity concurrently. The Scene-Class Affinity Loss is applied to semantic and geometric predictions, in conjunction with the cross-entropy loss weighted by class frequencies. The overall loss function is formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 113, + 639, + 286, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 639, + 286, + 653 + ], + "spans": [ + { + "bbox": [ + 113, + 639, + 286, + 653 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {g e o}} + \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {s e m}} + \\mathcal {L} _ {\\text {c e}} \\tag {5}", + "image_path": "f65290bca09d792083d2aace1f60d285f34f7eb75b22799edc320b18e1b2f6f0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": "Following the DETR series [4], auxiliary losses are applied after each Symphonies Decoder Layer for enhanced supervision, following the same formulation as " + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": " but scaled by a factor of 0.5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 387, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 387, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 387, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 91, + 545, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 545, + 175 + ], + "type": "text", + "content": "In this section, we present the evaluation results of our proposed Symphonies on SemanticKITTI [1] and SSCBenchKITTI-360 [21] datasets. The comparative analysis positioning Symphonies against existing approaches is detailed in Sec. 4.3. Additionally, comprehensive ablation studies are conducted in Sec. 4.4 to shed light on the thorough understanding of Symphonies." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 184, + 419, + 195 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 184, + 419, + 195 + ], + "spans": [ + { + "bbox": [ + 306, + 184, + 419, + 195 + ], + "type": "text", + "content": "4.1. Dataset and Metric" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 202, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 202, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 305, + 202, + 545, + 466 + ], + "type": "text", + "content": "The evaluation is performed on SemanticKITTI [1] and SSCBench-KITTI-360 [21] datasets, both providing densely annotated urban driving scene sequences, 22 and 9 respectively, from the KITTI Odometry Benchmark [10]. These datasets voxelize point clouds and label scenes measuring " + }, + { + "bbox": [ + 305, + 202, + 545, + 466 + ], + "type": "inline_equation", + "content": "51.2m \\times 51.2m \\times 64m" + }, + { + "bbox": [ + 305, + 202, + 545, + 466 + ], + "type": "text", + "content": ", with voxel grids of " + }, + { + "bbox": [ + 305, + 202, + 545, + 466 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 32" + }, + { + "bbox": [ + 305, + 202, + 545, + 466 + ], + "type": "text", + "content": " and a voxel size of " + }, + { + "bbox": [ + 305, + 202, + 545, + 466 + ], + "type": "inline_equation", + "content": "0.2m" + }, + { + "bbox": [ + 305, + 202, + 545, + 466 + ], + "type": "text", + "content": ". SemanticKITTI comprises 10 sequences for training, 1 sequence for validation, and 11 sequences for testing. It furnishes RGB images with shapes of " + }, + { + "bbox": [ + 305, + 202, + 545, + 466 + ], + "type": "inline_equation", + "content": "1226 \\times 370" + }, + { + "bbox": [ + 305, + 202, + 545, + 466 + ], + "type": "text", + "content": " as inputs and encompasses 20 semantic classes. SSCBench-KITTI-360 provides 7 sequences for training, 1 sequence for validation, and 1 sequence for testing, with 19 semantic classes and RGB images of size " + }, + { + "bbox": [ + 305, + 202, + 545, + 466 + ], + "type": "inline_equation", + "content": "1408 \\times 376" + }, + { + "bbox": [ + 305, + 202, + 545, + 466 + ], + "type": "text", + "content": ". For our camera-based approach, we exclusively adopt RGB images as input, and report the intersection over union (IoU) and mean IoU (mIoU) metrics aligned with standard practices. The IoU metric assesses the binary classification of empty versus occupied voxels, reflecting the performance of geometric reconstruction. Conversely, the mIoU metric provides a comprehensive assessment of semantic understanding, making it the primary metric in most benchmarks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 475, + 438, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 475, + 438, + 487 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 438, + 487 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 494, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 494, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 305, + 494, + 545, + 590 + ], + "type": "text", + "content": "In line with prior studies [3, 16, 22], we train Symphonies for 30 epochs on 4 NVIDIA 3090 GPUs, with a batch size of 4 images. We apply random horizontal flip augmentation and employ the AdamW [29] optimizer with an initial learning rate of 2e-4 and a weight decay of 1e-4. Learning rate reduction occurs by a factor of 0.1 at the 25th epoch. The ResNet-50 [12] backbone and Image Encoder are initialized with pre-trained MaskDINO [18] weights." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 598, + 391, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 391, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 391, + 609 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "text", + "content": "We conduct a comprehensive comparison of Symphonies with the latest state-of-the-art camera-based methodologies on the SemanticKITTI and SSCBench-KITTI-360 datasets. The results, outlined in Tab. 1 and Tab. 2, establish the superior performance of Symphonies with substantial improvements of 2.72 and " + }, + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.77\\mathrm{mIoU}" + }, + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "text", + "content": " on SemanticKITTI and SSCBench-KITTI-360, respectively. Specifically, Symphonies showcases particular excellence in instance classes," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20262" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 71, + 544, + 228 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 544, + 228 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 544, + 228 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 544, + 228 + ], + "type": "table", + "html": "
MethodIoUmIoUroadsidewalkparkingother-grd.buildingcartruckbicyclemotorcycleother-veh.vegetationtrunkterrainpersonbicyclistmotorcyclistfencepoletraf-sign
(15.30%)(11.15%)(12%)(0.56%)(14.1%)(3.92%)(16%)(0.03%)(0.03%)(0.03%)(0.20%)(3.3%)(0.51%)(17%)(0.07%)(0.07%)(0.05%)(3.90%)(0.29%)(0.08%)
LMSCNet†[33]31.387.0746.7019.5013.503.1010.3014.300.300.000.000.0010.800.0010.400.000.000.005.400.000.00
AICNet†[20]23.937.0939.3018.3019.801.609.6015.300.700.000.000.009.601.9013.500.000.000.005.000.100.00
JS3C-Net†[43]34.008.9747.3021.7019.902.8012.7020.100.800.000.004.1014.203.1012.400.000.200.208.701.900.30
MonoScene* [3]34.1611.0854.7027.1024.805.7014.4018.803.300.500.704.4014.902.4019.501.001.400.4011.103.302.10
TPVFormer [16]34.2511.2655.1027.2027.406.5014.8019.203.701.000.502.3013.902.6020.401.102.400.3011.002.901.50
VoxFormer [22]42.9512.2053.9025.3021.105.6019.8020.803.501.000.703.7022.407.5021.301.402.600.2011.105.104.90
OccFormer [46]34.5312.3255.9030.3031.506.5015.7021.601.201.501.703.2016.803.9021.302.201.100.2011.903.803.70
Symphonies42.1915.0458.4029.3026.9011.7024.7023.603.203.602.605.6024.2010.0023.103.201.902.0016.107.708.00
", + "image_path": "525ec6fd579f39e7a61154488b9192d8648537d0fcfc67de209360326d1c8bf5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 270, + 542, + 436 + ], + "blocks": [ + { + "bbox": [ + 47, + 236, + 546, + 258 + ], + "lines": [ + { + "bbox": [ + 47, + 236, + 546, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 546, + 258 + ], + "type": "text", + "content": "Table 1. Quantitative results on SemanticKITTI test. " + }, + { + "bbox": [ + 47, + 236, + 546, + 258 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 47, + 236, + 546, + 258 + ], + "type": "text", + "content": " denotes the results provided by [3]. * represents the reproduced results in [16, 46]. The best results are in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 270, + 542, + 436 + ], + "lines": [ + { + "bbox": [ + 52, + 270, + 542, + 436 + ], + "spans": [ + { + "bbox": [ + 52, + 270, + 542, + 436 + ], + "type": "table", + "html": "
MethodIoUPrec.Rec.mIoUcar(0.89%)bicycle(0.01%)motorcycle(0.03%)truck(0.16%)other-veh.(5.79%)person(0.02%)road(14.98%)parking(2.31%)sidewalk(6.43%)other-grd.(2.08%)building(15.67%)fence(0.96%)vegetation(41.99%)terrain(7.10%)pole(0.22%)traf.-sign(0.06%)other-struct.(4.33%)other-object.(0.28%)
LiDAR-based methods
SSCNet [37]53.5869.6369.9216.9531.950.000.1710.290.000.0765.7017.3341.243.2244.416.7743.7228.870.780.758.690.67
LMSCNet [33]47.5372.7757.5513.6520.910.000.000.260.580.0062.9513.5133.510.2043.670.3340.0126.800.000.003.630.00
Camera-based methods
MonoScene [3]37.8756.7353.2612.3119.340.430.588.022.030.8648.3511.3828.133.3232.893.5326.1516.756.925.674.203.09
TPVFormer [16]40.2259.3255.5413.6421.561.091.378.062.572.3852.9911.9931.073.7834.834.8030.0817.527.465.865.482.70
VoxFormer [22]38.7658.5253.4411.9117.841.160.894.562.061.6347.019.6727.212.8931.184.9728.9914.696.516.923.792.43
OccFormer [46]40.2759.7055.3113.8122.580.660.269.893.822.7754.3013.4431.533.5536.424.8031.0019.517.778.516.954.60
Symphonies44.1269.2454.8818.5830.021.855.9025.0712.068.2054.9413.8332.766.9335.118.5838.3311.5214.019.5714.4411.28
", + "image_path": "e794c0fea0579c084cb3c00dd8016681afbc2b37eb07f939e0b498c09573ce00.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 443, + 546, + 466 + ], + "lines": [ + { + "bbox": [ + 46, + 443, + 546, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 546, + 466 + ], + "type": "text", + "content": "Table 2. Quantitative results on SSCBench-KITTI-360 test. The results for counterparts are provided in [21]. The best results among all methods are in bold, and the best results for camera-based methods are underlined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 487, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 570 + ], + "type": "text", + "content": "e.g., buildings, cars, etc. This underscores its prowess in capturing and modeling intricate instance semantics. While VoxFormer attains a marginally higher IoU on SemanticKITTI, its adoption of two-stage training and extra occupancy prediction network disrupts end-to-end training and introduces additional geometric ambiguities. This complexity hampers its robustness, especially on KITTI-360." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 572, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 572, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 287, + 657 + ], + "type": "text", + "content": "The superiority of Symphonies becomes more pronounced on SSCBench-KITTI-360, which can be attributed to the ample data samples and high-quality annotations. Moreover, Symphonies even outperforms LiDAR-based methods in terms of mIoU, despite LiDAR's inherent advantage in IoU owing to its more precise position awareness, particularly at a distance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 669, + 148, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 148, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 148, + 681 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "The ablation analysis is conducted on the SemanticKITTI validation set from four key perspectives: overall architec" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 487, + 545, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 487, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 305, + 487, + 545, + 512 + ], + "type": "text", + "content": "tural components, the Symphonies Decoder, the Voxel Proposal Layer, and the Image Encoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "type": "text", + "content": "Ablation on architectural components. Tab. 3 presents the breakdown analysis of various architectural components within Symphonies. The vanilla baseline can be considered as a light-weight alternative to MonoScene, composed of a ResNet-50 backbone, an Image Encoder without pretrained weight, a 2D-to-3D projection via FLoSP [3], and a single 3D ASPP layer as the 3D decoder. Pre-training the Image Encoder leads to a notable improvement of 2.15 mIoU, emphasizing the effectiveness of instance awareness brought by 2D segmentation pre-training. Further, the proposed Depth-Rectified Voxel Proposal Layer improves performance by 0.75 mIoU through more accurate geometry. The Symphonies Decoder significantly boosts performance by 5.38 IoU, attributed to its dynamic instance modeling and context-capturing capabilities. In summary, the analy" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20263" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 117, + 70, + 476, + 144 + ], + "blocks": [ + { + "bbox": [ + 117, + 70, + 476, + 144 + ], + "lines": [ + { + "bbox": [ + 117, + 70, + 476, + 144 + ], + "spans": [ + { + "bbox": [ + 117, + 70, + 476, + 144 + ], + "type": "table", + "html": "
MethodIoUmIoUParams (M)FLOPs (G)
Baseline34.0610.4457.22529.20
+ Pre-trained Encoder35.97 (+1.91)12.59 (+2.15)57.22529.20
+ Voxel Proposal Layer36.54 (+0.57)13.34 (+0.75)57.42535.84
+ Symphonies Decoder41.92 (+5.38)14.89 (+1.55)59.31611.89
", + "image_path": "e2d6d20f85406d32a79dcab535155d6201fdf5019b29f65def0209c595906696.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 184, + 288, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 184, + 288, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 184, + 288, + 209 + ], + "type": "text", + "content": "sis in Tab. 3 affirms the effectiveness of the proposed components in Symphonies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 223, + 289, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 223, + 289, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 289, + 462 + ], + "type": "text", + "content": "Ablation on the Symphonies Decoder. To gain insights into the functionality of contextual instance queries, we assess the modular interactions within the Symphonies Decoder Layer. As depicted in Tab. 4, the incorporation of instance queries with either instance-image or instance-scene cross-attention considerably enhances performance by over 5 IoU and about " + }, + { + "bbox": [ + 46, + 223, + 289, + 462 + ], + "type": "inline_equation", + "content": "1\\mathrm{mIoU}" + }, + { + "bbox": [ + 46, + 223, + 289, + 462 + ], + "type": "text", + "content": ". This substantiates the significance of instance queries for adaptive aggregation of instance semantics. Among them, the instance-image cross-attention brings less improvement, suggesting that original image features have already been adequately captured in the preceding Voxel Proposal Layer. The instance self-attention further improve the performance, underlining the contextual effectiveness of efficient fusion among instance queries. Besides, though Scene SA may seem to only marginally improve the performance, its omission reveals significant fluctuations during training, attributed to the sparse interactions within instance deformable attentions. This underscores the pivotal role of Scene SA in fostering the generation of consistent features." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 471, + 287, + 581 + ], + "blocks": [ + { + "bbox": [ + 164, + 152, + 428, + 163 + ], + "lines": [ + { + "bbox": [ + 164, + 152, + 428, + 163 + ], + "spans": [ + { + "bbox": [ + 164, + 152, + 428, + 163 + ], + "type": "text", + "content": "Table 3. Ablation study on architectural components in Symphonies." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 471, + 287, + 581 + ], + "lines": [ + { + "bbox": [ + 50, + 471, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 50, + 471, + 287, + 581 + ], + "type": "table", + "html": "
Scn.-Inst. CAInst.-Img. CAInst.-Scn. CAInst. SAScn. SAIoUmIoU
35.9713.34
✓✓41.3614.02
✓✓41.4414.37
✓✓✓41.3514.63
✓✓✓✓41.7514.73
✓✓✓✓✓41.9214.89
", + "image_path": "889861f53b4b22d62f8f373e086e95073927cd3d09285189f5c78aa5b6c1e26b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Ablation on the Voxel Proposal Layer. Comparing the Depth-Rectified Voxel Proposal Layer (VPL) with FLoSP from MonoScene [3] casting pixels to voxels along the line of sight, as well as the mono VPL using monocular depth estimator AdaBins [2] (0.058 REL on KITTI), we note significant occupancy prediction improvements using" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 184, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 184, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 184, + 545, + 243 + ], + "type": "text", + "content": "the stereo VPL based on MobileStereoNet [35] (0.66 EPE on KITTI 2015), as shown in Tab. 5. This indicates that the rectification of more precise depth estimation contributes to mitigating geometric ambiguities, aligning with the findings in VoxFormer." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 341, + 251, + 511, + 312 + ], + "blocks": [ + { + "bbox": [ + 46, + 588, + 287, + 611 + ], + "lines": [ + { + "bbox": [ + 46, + 588, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 287, + 611 + ], + "type": "text", + "content": "Table 4. Ablation study on Symphonies Decoder. SA: Self-Attention, CA: Cross-Attention." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 341, + 251, + 511, + 312 + ], + "lines": [ + { + "bbox": [ + 341, + 251, + 511, + 312 + ], + "spans": [ + { + "bbox": [ + 341, + 251, + 511, + 312 + ], + "type": "table", + "html": "
2D-to-3D ProjectionIoUmIoU
FLoSP [3]36.0211.96
VPL (mono)38.3712.20
VPL (stereo)41.9214.89
", + "image_path": "1dcf5c3440af20887cc19ba77ab330ec3668baf0f8e6004b7add1bc661380477.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 359, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 359, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 359, + 545, + 478 + ], + "type": "text", + "content": "Ablation on the Instance-Aware Image Encoder. Tab. 6 evidently showcases the synergistic effects of utilizing pre-trained weights for the Image Encoder with instance queries. Solely utilizing pre-trained weights from MaskDINO [18] contributes an additional improvement of " + }, + { + "bbox": [ + 304, + 359, + 545, + 478 + ], + "type": "inline_equation", + "content": "0.48\\mathrm{mIoU}" + }, + { + "bbox": [ + 304, + 359, + 545, + 478 + ], + "type": "text", + "content": ". Moreover, incorporating instance queries with the pre-trained encoder yields a significant improvement of " + }, + { + "bbox": [ + 304, + 359, + 545, + 478 + ], + "type": "inline_equation", + "content": "1.36\\mathrm{mIoU}" + }, + { + "bbox": [ + 304, + 359, + 545, + 478 + ], + "type": "text", + "content": ", implying that the proposed instance queries benefit from the enhanced instance awareness of the encoder." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 312, + 487, + 541, + 559 + ], + "blocks": [ + { + "bbox": [ + 306, + 320, + 545, + 332 + ], + "lines": [ + { + "bbox": [ + 306, + 320, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 306, + 320, + 545, + 332 + ], + "type": "text", + "content": "Table 5. Ablation on the Depth-Rectified Voxel Proposal Layer." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 487, + 541, + 559 + ], + "lines": [ + { + "bbox": [ + 312, + 487, + 541, + 559 + ], + "spans": [ + { + "bbox": [ + 312, + 487, + 541, + 559 + ], + "type": "table", + "html": "
Pre-trained EncoderInst. QueriesIoUmIoU
41.0913.53
✓41.4213.32
✓41.1814.01
✓✓41.9214.89
", + "image_path": "1c7275d58122db4f6ddbf7f8c3debafff9ed303c25a87727f56c0ac665fa4436.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 567, + 545, + 578 + ], + "lines": [ + { + "bbox": [ + 306, + 567, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 545, + 578 + ], + "type": "text", + "content": "Table 6. Ablation study on the Instance-Aware Image Encoder." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 599, + 394, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 394, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 394, + 611 + ], + "type": "text", + "content": "4.5. Visualizations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Qualitative Results. Fig. 4 presents the visualizations of Symphonies on SemanticKITTI val, in comparison to the counterpart MonoScene. Symphonies generate more detailed predictions for instance-centric classes such as cars and trunks, as well as preserves clear and coherent layouts for structures like buildings and vegetation, attributed to the enriched instance semantics and contextual information provided by instance queries. In contrast, MonoScene" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20264" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 77, + 208, + 116 + ], + "blocks": [ + { + "bbox": [ + 97, + 77, + 208, + 116 + ], + "lines": [ + { + "bbox": [ + 97, + 77, + 208, + 116 + ], + "spans": [ + { + "bbox": [ + 97, + 77, + 208, + 116 + ], + "type": "image", + "image_path": "2d125133181974605ba41e81768fed2c41d55e725119c7d980555382e68ecb1e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 215, + 76, + 304, + 124 + ], + "blocks": [ + { + "bbox": [ + 215, + 76, + 304, + 124 + ], + "lines": [ + { + "bbox": [ + 215, + 76, + 304, + 124 + ], + "spans": [ + { + "bbox": [ + 215, + 76, + 304, + 124 + ], + "type": "image", + "image_path": "d80d39c1169085adc21f138a11d94fdb390b242ed226bc7f7a31dab032d2a834.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 311, + 75, + 393, + 128 + ], + "blocks": [ + { + "bbox": [ + 311, + 75, + 393, + 128 + ], + "lines": [ + { + "bbox": [ + 311, + 75, + 393, + 128 + ], + "spans": [ + { + "bbox": [ + 311, + 75, + 393, + 128 + ], + "type": "image", + "image_path": "e5d086f8d2607486274a2e6b6f0e8a054f6ec998aefba9130a931773540a3f12.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 396, + 74, + 492, + 132 + ], + "blocks": [ + { + "bbox": [ + 396, + 74, + 492, + 132 + ], + "lines": [ + { + "bbox": [ + 396, + 74, + 492, + 132 + ], + "spans": [ + { + "bbox": [ + 396, + 74, + 492, + 132 + ], + "type": "image", + "image_path": "874ff1dd621234463cc660c3e0abc3f05dea52661a496c44c33d3fc54e5a5686.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 97, + 142, + 207, + 177 + ], + "blocks": [ + { + "bbox": [ + 97, + 142, + 207, + 177 + ], + "lines": [ + { + "bbox": [ + 97, + 142, + 207, + 177 + ], + "spans": [ + { + "bbox": [ + 97, + 142, + 207, + 177 + ], + "type": "image", + "image_path": "aeb6af285715e91e4bf615cdf13bb2a9ce3558efc0ab239191bf82264d159a54.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 217, + 135, + 304, + 187 + ], + "blocks": [ + { + "bbox": [ + 217, + 135, + 304, + 187 + ], + "lines": [ + { + "bbox": [ + 217, + 135, + 304, + 187 + ], + "spans": [ + { + "bbox": [ + 217, + 135, + 304, + 187 + ], + "type": "image", + "image_path": "5f0d6f135247d721faf5164791f1ba8ebaf11e8d1f9e62930483a99f57687011.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 312, + 135, + 392, + 194 + ], + "blocks": [ + { + "bbox": [ + 312, + 135, + 392, + 194 + ], + "lines": [ + { + "bbox": [ + 312, + 135, + 392, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 135, + 392, + 194 + ], + "type": "image", + "image_path": "9e0c1293852789933f2c00d3608686483e558fd6d78314b907cc107c294ddbec.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 398, + 134, + 493, + 194 + ], + "blocks": [ + { + "bbox": [ + 398, + 134, + 493, + 194 + ], + "lines": [ + { + "bbox": [ + 398, + 134, + 493, + 194 + ], + "spans": [ + { + "bbox": [ + 398, + 134, + 493, + 194 + ], + "type": "image", + "image_path": "42d247771d3edc90d32153779ad5fd0c102c650d86e78e94a9254f78efada1e3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 97, + 204, + 206, + 239 + ], + "blocks": [ + { + "bbox": [ + 97, + 204, + 206, + 239 + ], + "lines": [ + { + "bbox": [ + 97, + 204, + 206, + 239 + ], + "spans": [ + { + "bbox": [ + 97, + 204, + 206, + 239 + ], + "type": "image", + "image_path": "c8c12316824ed975ca94e4514e2d8a8b5d2726ad87b4543883aac8da378b7f24.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 219, + 198, + 301, + 249 + ], + "blocks": [ + { + "bbox": [ + 219, + 198, + 301, + 249 + ], + "lines": [ + { + "bbox": [ + 219, + 198, + 301, + 249 + ], + "spans": [ + { + "bbox": [ + 219, + 198, + 301, + 249 + ], + "type": "image", + "image_path": "977752bda281ea6d7792c4385dda1021ea78b96cc3a08742be44429ed6e5a353.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 203, + 392, + 247 + ], + "blocks": [ + { + "bbox": [ + 312, + 203, + 392, + 247 + ], + "lines": [ + { + "bbox": [ + 312, + 203, + 392, + 247 + ], + "spans": [ + { + "bbox": [ + 312, + 203, + 392, + 247 + ], + "type": "image", + "image_path": "f3d616aaf9ffc708c8cce5ef02b07e7f6e9c55ae48011db097993f180cfb9b92.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 399, + 198, + 487, + 248 + ], + "blocks": [ + { + "bbox": [ + 399, + 198, + 487, + 248 + ], + "lines": [ + { + "bbox": [ + 399, + 198, + 487, + 248 + ], + "spans": [ + { + "bbox": [ + 399, + 198, + 487, + 248 + ], + "type": "image", + "image_path": "9646790bfd3561b68d44d152ea33a363e9d00a311420173515c98ed6342402c3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 97, + 266, + 206, + 300 + ], + "blocks": [ + { + "bbox": [ + 97, + 266, + 206, + 300 + ], + "lines": [ + { + "bbox": [ + 97, + 266, + 206, + 300 + ], + "spans": [ + { + "bbox": [ + 97, + 266, + 206, + 300 + ], + "type": "image", + "image_path": "2db1dcff58e97e7f1da12db3b21e3e4758a38d2eb5fa94ee7a13e490bcce3017.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 142, + 311, + 162, + 320 + ], + "lines": [ + { + "bbox": [ + 142, + 311, + 162, + 320 + ], + "spans": [ + { + "bbox": [ + 142, + 311, + 162, + 320 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 218, + 257, + 301, + 304 + ], + "blocks": [ + { + "bbox": [ + 218, + 257, + 301, + 304 + ], + "lines": [ + { + "bbox": [ + 218, + 257, + 301, + 304 + ], + "spans": [ + { + "bbox": [ + 218, + 257, + 301, + 304 + ], + "type": "image", + "image_path": "e5486a991f32fd983fa5d82609ce9a60cfa431bd15c7e71f8d0300840adcef9b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 311, + 280, + 319 + ], + "lines": [ + { + "bbox": [ + 238, + 311, + 280, + 319 + ], + "spans": [ + { + "bbox": [ + 238, + 311, + 280, + 319 + ], + "type": "text", + "content": "MonoScene" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 312, + 256, + 390, + 301 + ], + "blocks": [ + { + "bbox": [ + 312, + 256, + 390, + 301 + ], + "lines": [ + { + "bbox": [ + 312, + 256, + 390, + 301 + ], + "spans": [ + { + "bbox": [ + 312, + 256, + 390, + 301 + ], + "type": "image", + "image_path": "a59bd518684cea3d24af0b9635cd34b7e98e8d0002b2bb32bef6ddd576bef73c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 311, + 373, + 320 + ], + "lines": [ + { + "bbox": [ + 329, + 311, + 373, + 320 + ], + "spans": [ + { + "bbox": [ + 329, + 311, + 373, + 320 + ], + "type": "text", + "content": "Symphonies" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 403, + 251, + 493, + 309 + ], + "blocks": [ + { + "bbox": [ + 403, + 251, + 493, + 309 + ], + "lines": [ + { + "bbox": [ + 403, + 251, + 493, + 309 + ], + "spans": [ + { + "bbox": [ + 403, + 251, + 493, + 309 + ], + "type": "image", + "image_path": "c98e999f12b80536673b01ebcd39c1a6cf14591a7f4796d2021342f72de41452.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 311, + 466, + 319 + ], + "lines": [ + { + "bbox": [ + 417, + 311, + 466, + 319 + ], + "spans": [ + { + "bbox": [ + 417, + 311, + 466, + 319 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 378, + 287, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 438 + ], + "type": "text", + "content": "produces vague predictions with a radial shape, which is indicative of the aforementioned ambiguous geometry. These results underscore the superior capability of Symphonies in capturing fine-grained scene representations and enhancing overall scene understanding." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 454, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 454, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 287, + 561 + ], + "type": "text", + "content": "Attention Map Analysis. The attention map analysis in Fig. 5 provides insights into the mechanisms of the Serial Instance-Propagated Attentions within Symphonies layers. Notably, instance queries exhibit selective attention to corresponding regions in both the image and the scene. Additionally, they activate the semantically related regions within the scene-instance cross-attention. This observation validates the effect of our claimed instance-centric semantics in facilitating effective scene modeling." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 573, + 119, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 119, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 119, + 586 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "In this paper, we introduced Symphonies, a novel paradigm for 3D Semantic Scene Completion. Symphonies effectively integrates instance-centric semantics and scene context from both images and volumes, addressing the limitations posed by geometric ambiguity in prior voxel-wise modeling methods. Extensive experiments demonstrate the superiority of our approach over existing methods. We anticipate Symphonies to inspire future research and contribute to advancements in autonomous driving and 3D perception." + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 307, + 376, + 457, + 423 + ], + "blocks": [ + { + "bbox": [ + 46, + 335, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 335, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 545, + 357 + ], + "type": "text", + "content": "Figure 4. Qualitative visualizations on SemanticKITTI va1. Symphonies consistently produces detailed predictions for objects such as cars and trunks, while maintaining coherent layouts for structures like buildings and vegetation." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 376, + 457, + 423 + ], + "lines": [ + { + "bbox": [ + 307, + 376, + 457, + 423 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 457, + 423 + ], + "type": "image", + "image_path": "630d16c9f1c967ae06bf8293c98243d69b046a7a71a30ff338e2fc4199ade2d6.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 467, + 377, + 542, + 423 + ], + "blocks": [ + { + "bbox": [ + 467, + 377, + 542, + 423 + ], + "lines": [ + { + "bbox": [ + 467, + 377, + 542, + 423 + ], + "spans": [ + { + "bbox": [ + 467, + 377, + 542, + 423 + ], + "type": "image", + "image_path": "db37814759ec8622a1ec5a5752cadba636b31f820fe1b68e9f028a388357e04a.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 467, + 431, + 542, + 477 + ], + "blocks": [ + { + "bbox": [ + 467, + 431, + 542, + 477 + ], + "lines": [ + { + "bbox": [ + 467, + 431, + 542, + 477 + ], + "spans": [ + { + "bbox": [ + 467, + 431, + 542, + 477 + ], + "type": "image", + "image_path": "519e8fba28247ac876fe658f51762bf4b9687dfa6a85b7a3bc4b5dd9a7a048e2.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 467, + 486, + 542, + 532 + ], + "blocks": [ + { + "bbox": [ + 467, + 486, + 542, + 532 + ], + "lines": [ + { + "bbox": [ + 467, + 486, + 542, + 532 + ], + "spans": [ + { + "bbox": [ + 467, + 486, + 542, + 532 + ], + "type": "image", + "image_path": "ff1209a9176fee41c965e16ee4755887c423f9bc9dc8cc5520491605e22d94c2.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 536, + 542, + 548 + ], + "lines": [ + { + "bbox": [ + 471, + 536, + 542, + 548 + ], + "spans": [ + { + "bbox": [ + 471, + 536, + 542, + 548 + ], + "type": "text", + "content": "(c) Attention Maps for SceneInst. Cross-Attention." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 307, + 437, + 457, + 532 + ], + "blocks": [ + { + "bbox": [ + 318, + 425, + 444, + 432 + ], + "lines": [ + { + "bbox": [ + 318, + 425, + 444, + 432 + ], + "spans": [ + { + "bbox": [ + 318, + 425, + 444, + 432 + ], + "type": "text", + "content": "(a) Reference points for Inst.-Image Cross-Attention." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 437, + 457, + 532 + ], + "lines": [ + { + "bbox": [ + 307, + 437, + 457, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 457, + 532 + ], + "type": "image", + "image_path": "bd5af940c866a1d9da648039955057db242f30594779b19299162508106fcdbf.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 535, + 444, + 542 + ], + "lines": [ + { + "bbox": [ + 318, + 535, + 444, + 542 + ], + "spans": [ + { + "bbox": [ + 318, + 535, + 444, + 542 + ], + "type": "text", + "content": "(b) Reference points for Inst.-Scene Cross-Attention." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 559, + 534, + 571 + ], + "lines": [ + { + "bbox": [ + 317, + 559, + 534, + 571 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 534, + 571 + ], + "type": "text", + "content": "Figure 5. Analysis of attention maps within Symphonies." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "bbox": [ + 305, + 593, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 593, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 305, + 593, + 545, + 626 + ], + "type": "text", + "content": "Acknowledgement. This work was partially supported by the National Natural Science Foundation of China (NSFC) under Grant No. 62276108." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 306, + 650, + 365, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 650, + 365, + 661 + ], + "spans": [ + { + "bbox": [ + 306, + 650, + 365, + 661 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 310, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 310, + 669, + 545, + 713 + ], + "type": "text", + "content": "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Int. Conf. Comput. Vis., pages 9296-9306," + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20265" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 68, + 73, + 107, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 73, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 73, + 107, + 83 + ], + "type": "text", + "content": "2019. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 85, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 85, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 53, + 85, + 287, + 116 + ], + "type": "text", + "content": "[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 118, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 118, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 54, + 118, + 286, + 150 + ], + "type": "text", + "content": "[3] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3981–3991, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 151, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 151, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 54, + 151, + 287, + 194 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Eur. Conf. Comput. Vis., pages 213-229, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 195, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 287, + 249 + ], + "type": "text", + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., pages 834-848, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 287, + 303 + ], + "type": "text", + "content": "[6] Shaoyu Chen, Tianheng Cheng, Xinggang Wang, Wenming Meng, Qian Zhang, and Wenyu Liu. Efficient and robust 2d-to-bev representation learning via geometry-guided kernel transformer. arXiv preprint arXiv:2206.04584, 2022. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 305, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 305, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 53, + 305, + 287, + 348 + ], + "type": "text", + "content": "[7] Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Chang Huang, and Wenyu Liu. Polar parametrization for vision-based surround-view 3d detection. arXiv preprint arXiv:2206.10965, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 350, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 287, + 392 + ], + "type": "text", + "content": "[8] Bowen Cheng, Alexander G. Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. In Adv. Neural Inform. Process. Syst., pages 17864-17875, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 394, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 394, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 394, + 287, + 437 + ], + "type": "text", + "content": "[9] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1280–1289, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 438, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 438, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 438, + 287, + 470 + ], + "type": "text", + "content": "[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In IEEE Conf. Comput. Vis. Pattern Recog., 2012. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 472, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 472, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 287, + 503 + ], + "type": "text", + "content": "[11] Yuxiao Guo and Xin Tong. View-volume network for semantic scene completion from a single depth image. In IJCAI, pages 726-732, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 505, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 505, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 505, + 287, + 536 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "type": "text", + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE Conf. Comput. Vis. Pattern Recog., pages 15979-15988, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "type": "text", + "content": "[14] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. FIERY: future instance prediction in bird's-eye view from surround monocular cameras. In Int. Conf. Comput. Vis., pages 15253-15262, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 637, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 637, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 637, + 287, + 669 + ], + "type": "text", + "content": "[15] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[16] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9223-9232, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[17] Alexander Kirillov, Ross B. Girshick, Kaiming He, and Piotr Dólar. Panoptic feature pyramid networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "type": "text", + "content": "[18] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M. Ni, and Heung-Yeung Shum. Mask DINO: towards A unified transformer-based framework for object detection and segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3041-3050, 2023. 4, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "text", + "content": "[19] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian D. Reid. RGBD based dimensional decomposition residual network for 3d semantic scene completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7693-7702, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "type": "text", + "content": "[20] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3348-3356, 2020. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 264, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 330 + ], + "type": "text", + "content": "[21] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. arXiv preprint arXiv:2306.09001, 2023. 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 331, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 396 + ], + "type": "text", + "content": "[22] Yiming Li, Zhiding Yu, Christopher B. Choy, Chaowei Xiao, José M. Álvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9087-9098, 2023. 2, 4, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "text", + "content": "[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Eur. Conf. Comput. Vis., pages 1-18, 2022. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "text", + "content": "[24] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Trans. Pattern Anal. Mach. Intell., pages 3292–3310, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 500, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 544 + ], + "type": "text", + "content": "[25] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. In Adv. Neural Inform. Process. Syst., pages 261-272, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 545, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 587 + ], + "type": "text", + "content": "[26] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETR: position embedding transformation for multi-view 3d object detection. In Eur. Conf. Comput. Vis., pages 531-548, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "text", + "content": "[27] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 635, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 690 + ], + "type": "text", + "content": "[28] Zhi Liu, Shaoyu Chen, Xiaojie Guo, Xinggang Wang, Tianheng Cheng, Hongmei Zhu, Qian Zhang, Wenyu Liu, and Yi Zhang. Vision-based uneven BEV representation learning with polar rasterization and surface estimation. In Conference on Robot Learning, pages 437-446, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Int. Conf. Learn. Represent., 2019. 5" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20266" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[30] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 119, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 119, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 49, + 119, + 287, + 152 + ], + "type": "text", + "content": "[31] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Eur. Conf. Comput. Vis., pages 194-210, 2020. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 153, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 153, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 49, + 153, + 287, + 186 + ], + "type": "text", + "content": "[32] Thomas Roddick, Alex Kendall, and Roberto Cipolla. Orthographic feature transform for monocular 3d object detection. In Brit. Mach. Vis. Conf., page 285, 2019. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 189, + 286, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 189, + 286, + 220 + ], + "spans": [ + { + "bbox": [ + 49, + 189, + 286, + 220 + ], + "type": "text", + "content": "[33] Luis Roldão, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 3DV, pages 111–119, 2020. 2, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 223, + 286, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 223, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 49, + 223, + 286, + 255 + ], + "type": "text", + "content": "[34] Luis Roldão, Raoul de Charette, and Anne Verroust-Blondet. 3d semantic scene completion: A survey. Int. J. Comput. Vis., 130(8):1978-2005, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 258, + 286, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 258, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 49, + 258, + 286, + 300 + ], + "type": "text", + "content": "[35] Faranak Shamsafar, Samuel Woerz, Rafia Rahim, and Andreas Zell. Mobilestereonet: Towards lightweight deep networks for stereo matching. In WACV, pages 677-686, 2022. 4, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 304, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 304, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 49, + 304, + 287, + 346 + ], + "type": "text", + "content": "[36] Chonghao Sima, Wenwen Tong, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, and Hongyang Li. Scene as occupancy. arXiv preprint arXiv:2306.02851, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 350, + 286, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 350, + 286, + 392 + ], + "spans": [ + { + "bbox": [ + 49, + 350, + 286, + 392 + ], + "type": "text", + "content": "[37] Shuran Song, Fisher Yu, Andy Zeng, Angel X. Chang, Manolis Savva, and Thomas A. Funkhouser. Semantic scene completion from a single depth image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 190-198, 2017. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 396, + 286, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 396, + 286, + 437 + ], + "spans": [ + { + "bbox": [ + 49, + 396, + 286, + 437 + ], + "type": "text", + "content": "[38] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 441, + 286, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 441, + 286, + 483 + ], + "spans": [ + { + "bbox": [ + 49, + 441, + 286, + 483 + ], + "type": "text", + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Adv. Neural Inform. Process. Syst., pages 5998-6008, 2017. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 487, + 286, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 487, + 286, + 518 + ], + "spans": [ + { + "bbox": [ + 49, + 487, + 286, + 518 + ], + "type": "text", + "content": "[40] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. FCOS3D: fully convolutional one-stage monocular 3d object detection. In ICCVW, pages 913-922, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 521, + 286, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 521, + 286, + 564 + ], + "spans": [ + { + "bbox": [ + 49, + 521, + 286, + 564 + ], + "type": "text", + "content": "[41] Yida Wang, David Joseph Tan, Nassir Navab, and Federico Tombari. Forknet: Multi-branch volumetric semantic completion from a single depth image. In Int. Conf. Comput. Vis., pages 8607–8616, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 567, + 286, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 567, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 49, + 567, + 286, + 610 + ], + "type": "text", + "content": "[42] Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. DETR3D: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 613, + 286, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 613, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 49, + 613, + 286, + 656 + ], + "type": "text", + "content": "[43] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 658, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 286, + 712 + ], + "type": "text", + "content": "[44] Jiawei Yao, Chuming Li, Keqiang Sun, Yingjie Cai, Hao Li, Wanli Ouyang, and Hongsheng Li. Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. arXiv preprint arXiv:2309.14616, 2023.2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 250 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[45] Pingping Zhang, Wei Liu, Yinjie Lei, Huchuan Lu, and Xiaoyun Yang. Cascaded context pyramid for full-resolution 3d semantic scene completion. In Int. Conf. Comput. Vis., pages 7800-7809, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[46] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2304.05316, 2023. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "text", + "content": "[47] Brady Zhou and Philipp Kähenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13750-13759, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 208, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 250 + ], + "type": "text", + "content": "[48] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In Int. Conf. Learn. Represent., 2021. 1, 3, 4" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20267" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/30117727-713c-4f8f-a215-de682288f1dd_content_list.json b/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/30117727-713c-4f8f-a215-de682288f1dd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e6643ef259918bdeef553c4ec8a5a01f778c086b --- /dev/null +++ b/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/30117727-713c-4f8f-a215-de682288f1dd_content_list.json @@ -0,0 +1,1810 @@ +[ + { + "type": "text", + "text": "SynFog: A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving", + "text_level": 1, + "bbox": [ + 89, + 130, + 880, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiming Xie $^{1*}$ , Henglu Wei $^{1*}$ , Zhenyi Liu $^{2}$ , Xiaoyu Wang $^{1}$ , Xiangyang Ji $^{1}$ $^{1}$ Tsinghua University, $^{2}$ Stanford University \n{xieym22@mails., weihenglu@mail., xiaoyu-w17@mails., xyji@}tsinghua.edu.cn, {zhenyiliu@}stanford.edu", + "bbox": [ + 155, + 202, + 815, + 281 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8e1ff6fee21aab43dc8c8bc95515cc43b9b31ed14ba4d64a7dd2e191067e2632.jpg", + "image_caption": [ + "Figure 1. The SynFog dataset comprises images under different lighting, with the first row captured in natural sky light and the second with extra light sources like street lamps and automotive lighting. This dataset covers three levels of fog density across each scene, corresponding visibility of $600\\mathrm{m}$ , $300\\mathrm{m}$ , $150\\mathrm{m}$ . Additionally, pixel-accurate depth data and segmentation labels for each scene are also provided." + ], + "image_footnote": [], + "bbox": [ + 88, + 316, + 883, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 520, + 313, + 536 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To advance research in learning-based defogging algorithms, various synthetic fog datasets have been developed. However, existing datasets created using the Atmospheric Scattering Model (ASM) or real-time rendering engines often struggle to produce photo-realistic foggy images that accurately mimic the actual imaging process. This limitation hinders the effective generalization of models from synthetic to real data. In this paper, we introduce an end-to-end simulation pipeline designed to generate photorealistic foggy images. This pipeline comprehensively considers the entire physically-based foggy scene imaging process, closely aligning with real-world image capture methods. Based on this pipeline, we present a new synthetic fog dataset named SynFog, which features both sky light and active lighting conditions, as well as three levels of fog density. Experimental results demonstrate that models trained on SynFog exhibit superior performance in visual perception and detection accuracy compared to others when applied to real-world foggy images.", + "bbox": [ + 75, + 542, + 472, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 842, + 209, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fog, due to its light scattering and attenuation properties, poses a significant challenge for outdoor vision-based sys", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tems such as remote sensing, surveillance, autonomous driving, and fire rescue [6, 10, 52]. It causes severe degradation in object appearance and contrast, resulting in a notable impact on people's daily lives. Consequently, there has been a growing research interest in image defogging task to develop robust outdoor vision systems.", + "bbox": [ + 496, + 522, + 892, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent learning-based defogging algorithms depend on sets of foggy images and corresponding clear images for model training. While several datasets have been collected from real-world scenarios [5, 40], they are typically limited to foggy images only, as it is almost impossible to obtain pixel-accurate clear images with consistent scene contents and environmental illumination. Studies in [2-4, 35] attempt to capture real-world foggy and clear image pairs using smoke machines. However, the scale of these datasets is inadequate for training of a high-performing model. Given the challenge of obtaining paired foggy and clear images in real world, various synthetic fog datasets have been proposed as an alternative [25, 40, 55]. The most widely used method is based on the theory proposed by McCartney [33], also known as the Atmospheric Scattering Model (ASM). However, this model fails to consider accurate global illumination and the actual imaging process that occurs in real-world captures as detailed in Sec. 3.1. These limitations significantly impact the model's overall applicability, par", + "bbox": [ + 496, + 613, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21763", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/67fb9f7b4c55f8898211c4c1c3a14228300936503c99f3768c41d39632ef0019.jpg", + "image_caption": [ + "Figure 2. Relationship between noise level and fog density in real-world foggy images. Using regional contrast stretching for image enhancement, it can be observed that the noise level increases with the concentration of fog." + ], + "image_footnote": [], + "bbox": [ + 84, + 85, + 480, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ticularly in nighttime scenarios, and can lead to a disparity between synthetic and real-world foggy images. As a result, models trained on synthetic foggy images may exhibit limited robustness when applied to real-world ones.", + "bbox": [ + 75, + 338, + 468, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we propose an end-to-end foggy image simulation pipeline that encompasses the entire imaging process starting from 3D virtual scenes to final photorealistic foggy images. Our approach involves foggy scene rendering using volumetric path tracing and imaging through a realistic camera model. As shown in Fig. 3b, fog is treated as a spatial volume, and the spectral scene radiance is rendered using volumetric path tracing [38]. By defining the phase function of fog medium and bidirectional reflectance distribution function (BRDF) of the scene components, we are able to accurately describe the scattering process with multiple light sources under foggy conditions. The foggy scene radiance data is further processed through a physically-based camera model, which comprises optics, sensor and image processing, to faithfully replicate real camera devices. Based on the proposed pipeline, we develop a new extensive synthetic fog dataset called SynFog. This dataset contains 500 unique outdoor scenes, each featuring two kinds of lighting conditions and three levels of fog density. Extensive experiments have been conducted to qualitatively and quantitatively validate the authenticity of our simulation pipeline and the SynFog dataset.", + "bbox": [ + 75, + 398, + 468, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 94, + 732, + 434, + 744 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a physically-based end-to-end foggy image simulation pipeline that incorporates accurate light transportation in scattering medium and physical characteristics of optics and sensor into the synthesized images.", + "- We develop a photo-realistic synthetic fog dataset, SynFog, which includes both sky light and active lighting conditions, as well as three levels of fog density.", + "- We demonstrate the authenticity and effectiveness of our simulation pipeline and the superior realism of SynFog dataset compared to other synthetic fog datasets." + ], + "bbox": [ + 76, + 747, + 467, + 896 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the challenge of acquiring paired foggy and clear images in the real world for training learning-based defogging methods, researchers have proposed many synthetic fog datasets as an alternative. In this section, we will offer a thorough review and analysis of foggy image simulation techniques and their associated datasets.", + "bbox": [ + 496, + 114, + 890, + 204 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on Atmospheric Scattering Model (ASM). This method involves rendering fog onto clean images along with their depth information [1, 19, 25, 26, 40, 53, 55]. Depth information can be acquired through game engines, depth cameras or depth estimation methods. In an earlier study conducted by Tarel et al. [46], FRIDA was constructed using a total of 90 synthetic images, with clear images and depth maps generated using $SiVIC^{TM}$ software. Subsequently, in their later work [47], more synthetic images were added, leading to an upgraded version named FRIDA2, which includes 330 synthetic images. However, the scale of these datasets is insufficient for modern data-driven defogging methods, and the foggy images appear unrealistic. In more recent works [26, 53], a similar approach has been employed to acquire both the clear images and their corresponding depth maps from game engines. In [25], a general dehazing dataset called RESIDE was introduced. The indoor training set (ITS) of RESIDE was constructed based on existing indoor RGBD datasets NYUv2 [42] and Middleburry stereo [41]. For the outdoor training set (OTS), 2071 real-world outdoor images were collected, and the corresponding depth information was estimated using a learning-based technique [27]. It is important to note that the accuracy of depth estimation based on learning methods is still unsatisfactory, as single-image-based depth estimation is inherently an ill-posed problem. Furthermore, Foggy Cityscapes was introduced in [40], utilizing images from the Cityscapes dataset [9]. Despite efforts to enhance the original depth maps in Cityscapes, artifacts and discrepancies still persist, resulting in synthetic foggy images that appear unnatural.", + "bbox": [ + 496, + 205, + 892, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on Real-time Rendering Engines. Foggy images generated utilizing real-time rendering engines like Unity and Unreal have been discussed in works such as [18, 44]. Among them, Virtual KITTI 2 [18] stands out as a notable example. However, these foggy images often lack realism and exhibit limited diversity in terms of fog density.", + "bbox": [ + 496, + 674, + 890, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on Physical Means. Fog/haze machines are employed in real-world settings to generate artificial fog, as discussed in works such as [2-4, 35]. However, these datasets are generally limited in scale and may not offer sufficient supervision for learning-based defogging methods. Moreover, ensuring consistent content between the clear and foggy image captures can be challenging.", + "bbox": [ + 496, + 765, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are also several real-world fog datasets available, such as Foggy Driving [40], Foggy Zurich [11], and Seeing", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21764", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/09614b1fb99b25afb9a90a4e90e07fad931f0d2f43194ac75c63869194cddd1a.jpg", + "image_caption": [ + "(a) Atmospheric Scattering Model (ASM) [34]." + ], + "image_footnote": [], + "bbox": [ + 117, + 94, + 428, + 248 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d5eb05dde2831ddd5af86a71a9db1f88f3a3f7617d3e89d78f062ff002a2d09e.jpg", + "image_caption": [ + "(b) Foggy scene rendering based on volumetric path tracing.", + "Figure 3. Foggy scene simulation methods." + ], + "image_footnote": [], + "bbox": [ + 114, + 282, + 429, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Through Fog [5]. However, due to the absence of corresponding clear reference images, these datasets can only be used for testing purposes rather than for training. In this paper, we propose a physically-based end-to-end foggy image simulation pipeline to improve the authenticity of synthetic foggy images and construct a photo-realistic fog dataset to facilitate future research on defogging algorithms.", + "bbox": [ + 75, + 474, + 468, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. End-to-end Foggy Image Simulation", + "text_level": 1, + "bbox": [ + 76, + 599, + 405, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose an end-to-end approach for generating photorealistic foggy images that incorporates accurate light transportation in scattering medium and physical characteristics of optics and sensor into the synthesized images, as illustrated in Fig. 4. We first introduce a foggy scene imaging model in Sec. 3.1. Based on this model, we establish a simulation pipeline consisting of two components: a) Foggy scene radiance is rendered using volumetric path tracing [38]. b) The radiance data is processed through a physically-based camera model, which comprises optics, sensor and image processing to faithfully replicate real camera devices [17]. These components are further detailed in Sec. 3.2.1 and Sec. 3.2.2, respectively.", + "bbox": [ + 75, + 628, + 468, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Foggy Scene Imaging Model", + "text_level": 1, + "bbox": [ + 76, + 844, + 331, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Atmospheric Scattering Model (ASM) [33, 34] has been widely used for the generation of foggy images [19,", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "25,40,53]:", + "bbox": [ + 500, + 90, + 578, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE (d, \\lambda) = \\underbrace {E _ {o} (\\lambda) e ^ {- \\beta (\\lambda) d}} _ {\\text {t r a n s m i s s i o n}} + \\underbrace {L _ {h} (\\infty , \\lambda) \\left(1 - e ^ {- \\beta (\\lambda) d}\\right)} _ {\\text {a i r l i g h t}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 119, + 890, + 178 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $E$ is the irradiance reaching the film under foggy conditions, $E_{o}$ is the irradiance under clear conditions, $\\beta$ is the scattering coefficient characterizing the scatter ability of the medium, $d$ is the distance between the object and the camera, $L_{h}(\\infty ,\\lambda)$ denotes the radiance of airlight at infinity and $\\lambda$ refers to wavelength. Fig. 3a provides a more intuitive description of this process. Foggy images can be generated with pre-defined $\\beta$ and $L_{h}$ values by directly using the clear images as $E_0$ and the corresponding depth maps as $d$ .", + "bbox": [ + 496, + 179, + 890, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, there are two issues with this scattering model for generating foggy images: (1) The definition of environmental illumination fails to consider the contribution of indirect illumination from the scene to the scattering medium, resulting in an unfaithful representation of the actual illumination. (2) The current formulation of airlight is insufficient for accurately modeling the lighting conditions that involve multiple point light sources, such as street lamps and automotive lights. This limitation restricts the model's applicability in situations like nighttime [29, 51].", + "bbox": [ + 496, + 316, + 892, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, ASM only captures the irradiance reaching the sensor plane. Consequently, foggy images generated using ASM fail to accurately incorporate the realistic physical characteristics of optics and sensor in actual imaging process. In Fig. 2, we depict the relationship between noise level and fog density after the enhancement of regional contrast stretching [48] on a real captured image [8]. It can be seen that the noise level increases with higher fog density, emphasizing the critical need for accurate noise modeling in foggy image simulation.", + "bbox": [ + 496, + 469, + 890, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enhance the consistency between synthetic images and real-world captures, we develop a comprehensive foggy scene imaging model that incorporates the entire imaging process, referred to as:", + "bbox": [ + 496, + 623, + 890, + 683 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {o u t} = f _ {I S P} \\left(f _ {\\text {s e n s o r}} \\left(f _ {\\text {o p t i c s}} \\left(L _ {\\text {s c e n e}}\\right)\\right) + n\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 700, + 890, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $L_{\\text{scene}}$ is the scene radiance obtained using volumetric path tracing [38], as described in Sec. 3.2.1; $\\{f_{\\text{optics}}, f_{\\text{sensor}}, f_{\\text{ISP}}\\}$ denote the optics, sensor and ISP processes respectively, as described in Sec. 3.2.2, and the noise term introduced by the sensor is denoted as $n$ .", + "bbox": [ + 496, + 732, + 890, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By employing volumetric path tracing to obtain the radiance of foggy scene, we can precisely model the light scattering process under global illumination and active lighting. By incorporating the complete imaging process described in Eq. (2), we are able to faithfully replicate the authentic camera characteristics in foggy image simulation.", + "bbox": [ + 496, + 810, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21765", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/367edc5b7d3434f72977b0fa071f665ea871022e136cb73b9d7fa4a9bb9a9005.jpg", + "image_caption": [ + "Figure 4. End-to-end foggy image simulation pipeline. The spectral radiance data is rendered using volumetric path tracing and passing through a realistic optics model before reaching the sensor plane. Subsequently, the irradiance is converted into a digital image through the sensor model. The raw image from the sensor is then processed by an ISP module to generate the final foggy image." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 326, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5ec3364dd38c1088a85fc476b4df2f784def9bf0ec4bb3340002c6aafce770cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 89, + 493, + 167 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4bc42c76256e1ae821b6a302cb41af5c096e23803feae3c13329c018a3a23a79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 89, + 679, + 169 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ef7d7da76639afea1f4855a58d7e8e835c60e548f57ae5c705c1b11b202f0798.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 89, + 890, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. End-to-end Simulation Pipeline", + "text_level": 1, + "bbox": [ + 76, + 241, + 354, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Foggy Scene Rendering", + "text_level": 1, + "bbox": [ + 76, + 265, + 295, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We utilize procedural modeling techniques, similar to the approach employed in [31, 49], to create large-scale and diverse driving scenes. This involves defining adjustable parameters and using a set of predefined rules to generate explicit scene definitions. Our model library incorporates a diverse range of pedestrians, bicyclists and cars, each with well-defined material and texture properties. Furthermore, we carefully select 16 sky maps with cloudy conditions to ensure realistic lighting for foggy scenes, following the established guidelines from previous works [2-4].", + "bbox": [ + 75, + 289, + 468, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our approach, fog is treated as a spatial volume, and the spectral scene radiance is rendered using volumetric path tracing [38]. To accurately model the scattering process of the fog medium, we employ Henyey and Greenstein phase function [21, 38] to model its particle scattering characteristics, denoted as:", + "bbox": [ + 75, + 440, + 468, + 530 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\mathrm {H G}} (\\cos \\theta) = \\frac {1}{4 \\pi} \\frac {1 - g ^ {2}}{\\left(1 + g ^ {2} + 2 g (\\cos \\theta)\\right) ^ {3 / 2}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 540, + 468, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By empirically setting the asymmetry parameter $g = 0.87$ , we can accurately simulate the probability distributions for scattering in a specific direction at a given point within the fog medium, as shown in Fig. 3b.", + "bbox": [ + 75, + 587, + 468, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The rendering equation consists of two parts that account for reduction of radiance caused by out-scattering and increase of radiance due to in-scattering at point $p$ in direction $\\omega$ , represented by Eq. (4) and Eq. (5), respectively [38].", + "bbox": [ + 75, + 648, + 468, + 709 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {i}} \\left(\\mathrm {p} ^ {\\prime}, - \\omega\\right) = \\mathrm {e} ^ {- \\int_ {0} ^ {d} \\sigma_ {\\mathrm {s}} (\\mathrm {p} + t \\omega , \\omega) \\mathrm {d} t} L _ {\\mathrm {o}} (\\mathrm {p}, \\omega). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 719, + 468, + 739 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {s}} (\\mathrm {p}, \\omega) = L _ {\\mathrm {e}} (\\mathrm {p}, \\omega) + \\sigma_ {\\mathrm {s}} (\\mathrm {p}, \\omega) \\int_ {\\mathcal {S} ^ {2}} p (\\mathrm {p}, \\omega_ {\\mathrm {i}}, \\omega) L _ {\\mathrm {i}} (\\mathrm {p}, \\omega_ {\\mathrm {i}}) \\mathrm {d} \\omega_ {\\mathrm {i}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 750, + 482, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\sigma_{s}$ represents the scattering probability per unit distance, $\\{L_i,L_o,L_e,L_s\\}$ are the {incident,exitant,emission,total added} radiance. In our implementation, we employ the Monte Carlo algorithm to solve the above integration.", + "bbox": [ + 76, + 795, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key distinction between our physically-based foggy scene rendering technique and ASM lies in the modeling of airlight, as represented by Eq. (5). Our approach offers a", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0bb284ef9b1207033f17d9117b63a43bdbfc1743e83004c8e544cfd5880e2e90.jpg", + "image_caption": [ + "Figure 5. Noise flow and sensor simulation." + ], + "image_footnote": [], + "bbox": [ + 503, + 241, + 893, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "more accurate depiction of light scattering process through the utilization of specific phase function, and enables incorporation of global illumination and multiple light sources within the scene, producing authentic scene radiance to the camera model described in the next section.", + "bbox": [ + 498, + 373, + 890, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Imaging through Realistic Camera Simulation", + "text_level": 1, + "bbox": [ + 500, + 464, + 875, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining the scene radiance, we employ ISETCam [15-17, 30, 32] to simulate the realistic digital camera system that encompasses optics, sensor and ISP components. Specifically, we use a diffraction-limited optics model [7] with consideration of off-axis irradiance attenuation based on the cosine-fourth law [24] to convert scene radiance $L_{o}$ to optical image irradiance $E_{i}$ , denoted as:", + "bbox": [ + 498, + 487, + 890, + 593 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {i} = \\pi \\cdot L _ {o} \\frac {1}{1 + 4 (f / \\# (1 - m)) ^ {2}} (\\cos \\phi) ^ {4}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 602, + 890, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f / \\#$ is the f-number, $m$ represents the magnification defined as the negative ratio of image distance and object distance. In this optics model, blur mainly depends on the f-number value but little else.", + "bbox": [ + 498, + 643, + 890, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We model the sensor process including photovoltaic conversion, analog amplifier, and analog-to-digital conversion as described in Fig. 4 and simulate the real noise characteristics depicted in Fig. 5. Among the various noise sources, shot noise arising from the photovoltaic conversion emerges as the dominant type of noise in foggy images which follows the Poisson distribution. To ensure the credibility of our sensor simulation, we incorporate a set of calibrated sensor parameters from the Mi10Pro cellphone, encompassing the CFA and infrared filter data.", + "bbox": [ + 496, + 704, + 890, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we process the RAW image output from the sensor with a simple ISP module, including black level compensation, white balance, demosaicing, color correction and", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21766", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/418731bc065c0512b2fe6a12ab590ad7a143741a689d3a9fcd5ab45751d9bb17.jpg", + "table_caption": [ + "Table 1. Comparisons between SynFog and existing synthetic fog datasets in the driving field." + ], + "table_footnote": [], + "table_body": "
DatasetsGlobal IlluminationOptics and Sensor Characteristics
Foggy Cityscapes××
Virtual KITTI 2××
SynFog (Ours)✓✓
", + "bbox": [ + 81, + 128, + 460, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "gamma correction, to get the final synthetic image that mimics the real camera characteristics.", + "bbox": [ + 76, + 229, + 467, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. SynFog Dataset", + "text_level": 1, + "bbox": [ + 76, + 275, + 233, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By utilizing the proposed pipeline, we develop a new extensive synthetic fog dataset called SynFog. Example images are shown in Fig. 1. We first generate 500 large-scale, diverse driving scenes using procedural modeling techniques. For each scene, we generate two sets of images under different lighting conditions: one set with sky light only and the other set with both sky light and active light sources, such as street lamps and automotive lighting. Within each group, we create a set of four images: one clear image and three foggy images with varying fog levels defined by the scattering coefficient $\\sigma_{\\mathrm{s}}$ . According to McCartney's work [33], the scattering coefficient remains independent of the wavelength for fog [13, 34]. Therefore, we use a consistent scattering coefficient for wavelengths within the visible range. To ensure realistic simulation of fog, we employ the meteorological optical range (MOR) [23] or visibility as a reference to select appropriate scattering coefficients, described as:", + "bbox": [ + 75, + 300, + 468, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM O R = 2. 9 9 6 / \\sigma_ {s}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 573, + 468, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on this equation, we select scattering coefficient values of 0.005, 0.01, 0.02, which correspond to visibility range of $600\\mathrm{m}$ , $300\\mathrm{m}$ and $150\\mathrm{m}$ , respectively. According to National Standard of China [36], these visibility ranges are classified as heavy fog, thick fog, and dense fog, thereby the simulated foggy images are diversified enough to reflect real-world complicated foggy conditions.", + "bbox": [ + 75, + 598, + 468, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare SynFog with existing fog datasets in the field of autonomous driving, including the extensively used Foggy Cityscapes [40] and Virtual KITTI 2 [18]. The comparison is presented in Tab. 1. In terms of foggy image generation, methods that rely on ASM [34] require accurate depth maps to calculate the correct transmission maps. However, obtaining accurate depth maps for outdoor images can be challenging, leading to artifacts such as holes and discontinuities in synthetic foggy images, similar to what is observed in Foggy Cityscapes [40]. On the other hand, foggy images generated using Game Engines like Unity or Unreal [18] typically employs rasterization techniques in the rendering process. While effective for creating", + "bbox": [ + 75, + 704, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "scenes, rasterization techniques may not accurately model light transportation as precisely as path tracing, which is employed in SynFog. Moreover, SynFog takes the realistic imaging process into account, resulting in highly authentic foggy images with realistic physical characteristics. The superior realism of SynFog is further supported by the experimental results.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, our data simulation pipeline offers exceptional flexibility and control, allowing us to generate not only the final RGB images but also other valuable data modalities, including RAW images. This capability opens up avenues for research on defogging algorithms in the RAW domain, which has not been explored due to the lack of paired foggy RAW images and clear RGB images necessary for training deep defogging networks.", + "bbox": [ + 496, + 198, + 890, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 498, + 334, + 633, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Fog Chamber Validation", + "text_level": 1, + "bbox": [ + 498, + 359, + 725, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We establish an indoor fog chamber to validate the fidelity of our fog simulation pipeline. The chamber is constructed as a capped glass cube with dimensions of $1.5m\\times 0.5m\\times$ $0.5m$ . To reproduce the dark night environment that has directional light sources in the fog chamber, we position 4 light bars on each side of the front face of the cuboid fog chamber and enclose it with black blackout cloth from the inside. Fog inside the chamber is generated using a 1.7MHz ceramic atomizer that emulates water mist. To ensure uniformity of fog concentration within the chamber, we incorporate dual fans to promote even distribution. A visibility measurement module is designed to quantify the fog density. This module is comprised of a PM-160 wireless power meter from Thorlabs, an 852 narrowband laser, an aperture and a narrowband filter adapted to the laser band. The aperture serves to filter out stray light and the narrowband filter plays the role in blocking interference caused by non-laser light sources in the scene. The principle of visibility (scattering coefficient $\\sigma_{s}$ ) measurement is expressed as [37]:", + "bbox": [ + 496, + 383, + 890, + 671 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {s} = \\frac {\\ln \\left(\\frac {P _ {0}}{P _ {u}}\\right)}{\\mathrm {u}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 681, + 890, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $P_0$ is the laser power measured without fog and $P_u$ is the measured laser power after the propagation of $u$ distance under foggy conditions.", + "bbox": [ + 496, + 733, + 890, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then use Blender to create a virtual fog chamber that can reproduce the real-world counterpart. Our process involves calibrating the light conditions, object information and camera properties prior to capture, resulting in a one-to-one reproduction of the real fog chamber. We select the standard Macbeth color checker as our test object and use Mi10Pro cellphone to capture images. We evaluate the matching degree between the simulation and reality un", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21767", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c81e9100ee41046608c476a51ec4d25e06c83552ef5149f70d7059835f36f1b4.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 472, + 275 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/36ff337738896c144f5d81762896a635365349fab6bfa047015b55d8e3a8ec4a.jpg", + "image_caption": [ + "Figure 6. Color analysis under foggy conditions. (a) Visual comparison between real captures and simulated images under different fog densities in linear RGB space. We increase the brightness of the images $(\\times 2)$ for visual purpose. (b) Analyzing the trend of color variation across different fog densities." + ], + "image_footnote": [], + "bbox": [ + 84, + 299, + 467, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "der clear conditions by conducting a color analysis on the checker, which is available in the supplementary material.", + "bbox": [ + 75, + 534, + 468, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After ensuring the consistency of initial capture conditions, we proceed to capture images under foggy conditions. We first record the laser power under clear condition as $P_0$ , and then proceed to generate fog until the object is no longer visible. During the fog dispersion process, we capture images at regular intervals of 1 second and record the laser energy $P_u$ to obtain real-time fog density measurements. We utilize the measured fog density parameter $\\sigma_s$ to simulate corresponding foggy counterparts based on the proposed foggy image simulation pipeline.", + "bbox": [ + 75, + 566, + 468, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 6a, we showcase both the real-captured foggy images and simulated ones generated using our pipeline and the ASM [33]. It can be seen that images simulated through our pipeline exhibit a closer resemblance to real-captured foggy images. We also conduct a color analysis on 24 color patches under different fog densities as depicted in Fig. 6b. The results demonstrate that, as fog density increases, the trend of each channel in simulated images generated by our pipeline consistently follows that of the real captures. In contrast, the simulated images based on ASM do not exhibit this trend. These findings provide a quantitative validation of the realism of our foggy image simulation method.", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Transferability across the Real-to-Virtual Gap", + "text_level": 1, + "bbox": [ + 498, + 90, + 890, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate 7 representative state-of-the-art defogging algorithms on SynFog dataset: DCP [20], CAP [57], Grid-DehazeNet [28], MSBDN [12], FFA-Net [39], AECR-Net [50], and DehazeFormer [43]. We directly apply the models trained on SynFog to real foggy images sourced from Foggy Zurich [11] and Seeing Through Fog [5] datasets, without undergoing any model fine-tuning or domain adaptation. During the training process, we adhere to the official settings of these methods, and all of them demonstrate convergence by the end of the training. The results are shown in Fig. 7. Despite DCP [20] and CAP [57] producing excessively dark outputs, models trained on SynFog can produce natural defogged images with minimal artifacts.", + "bbox": [ + 496, + 113, + 890, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To underscore the superior realism of SynFog compared to existing synthetic fog datasets, we further enhance our evaluation by training the top-performing models (AECR-Net [50], DehazeFormer [43]) on two prominent synthetic fog datasets relevant to autonomous driving (Foggy Cityscapes [40], Virtual KITTI [18]). Subsequently, these models are evaluated on real fog datasets for a comparative analysis. Additionally, we utilize ASM [34] to create a SynFog- $\\beta$ dataset with clear images from SynFog and test it using DehazeFormer [43].", + "bbox": [ + 496, + 310, + 890, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with existing synthetic fog datasets. To mitigate any potential impact arising from model training issues, we initially conduct tests for each model using the corresponding test set of the training data. Test results are available in the supplementary material. Both AECR-Net [50] and DehazeFormer [43] exhibit high performance, confirming the correctness of the model training process. Subsequently, the trained models are directly applied to real-world test foggy images. Results are shown in Fig. 9 and Tab. 2. Owing to the high authenticity of the SynFog dataset, models trained on it can effectively generalize to real foggy images and produce naturally colored defogged images. Conversely, models trained on the other two datasets display poorer generalization to real data, resulting in artifacts and color distortion in defogged images. We also evaluate the defogged images using a pre-trained YOLOv8 [22] model for detection, showcasing how SynFog enhances machine vision capabilities (See Tab. 3). More results can be found in the supplementary material.", + "bbox": [ + 496, + 462, + 892, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It is noteworthy that the training set of Foggy Cityscapes [40] comprises 8925 data pairs, and Virtual KITTI [18] includes 3452 data pairs, both exceeding the 1350 data pairs in SynFog. However, models trained on SynFog demonstrate superior generalization capabilities on real-world data compared to them, highlighting the crucial role of data quality for researchers.", + "bbox": [ + 496, + 750, + 890, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with SynFog- $\\beta$ synthesized by ASM. With the ability to obtain pixel-accurate depth information for each image, we can eliminate the impact of inaccurate", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21768", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2d9813091332f8ef1b3240e58ad25bd87afeddbf364c97e3f6c1039f7bcd2686.jpg", + "table_caption": [ + "Table 2. Defogging evaluations with AECRNet on real-world fog datasets." + ], + "table_footnote": [], + "table_body": "
Training SetO-Haze [2]Foggy Zurich [11]Foggy Driving [40]BeDDE [56]
PSNR ↑SSIM ↑DHQI [14] ↑DHQI [14] ↑DHQI [14] ↑DHQI [14] ↑
Foggy Cityscapes14.460.573743.4052.0651.5536.07
Virtual KITTI13.900.531542.8050.9447.4633.42
SynFog15.430.611644.4654.1652.0743.28
", + "bbox": [ + 94, + 107, + 874, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b1f5812d7c12de482978a6d936c91211746e0aaf2e213207e94c07117207ccc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 196, + 289, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/126cb07242ae6d47cf3790a8c695d911a9a8fea06ed4c553736c30aab29ee76d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 196, + 480, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/10e1190dce571374cd3edcdb719870ab82d265ec7c2d06a95ff0acbd8f2d05d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 196, + 674, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f0756365c306c6245fecd1237ea401ea9b08a33a0a0d9c36e3539f915cc2c0c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 196, + 867, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6db9bf03d46b68b982a95663e4d27900ac73d9cee60428f91afed4f2a037f6fa.jpg", + "image_caption": [ + "Input", + "MSBDN [12]" + ], + "image_footnote": [], + "bbox": [ + 96, + 292, + 289, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a55c76112416daaba3cc54d9256c52f43eac2eb55e2ccdbb63cd2a1c656758a.jpg", + "image_caption": [ + "DCP [20]", + "FFA-Net [39]" + ], + "image_footnote": [], + "bbox": [ + 292, + 292, + 480, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/897d10a502d991a9126b94861160f0e1e4bc21e7dfb3980962aca42d1af80fdb.jpg", + "image_caption": [ + "CAP [57]", + "AECR-Net [50]" + ], + "image_footnote": [], + "bbox": [ + 483, + 292, + 674, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a9fd4e27f47dc4bb3b876ff7c1d8bdcb239cba1982ff2d99dc728e7ba7301ec2.jpg", + "image_caption": [ + "GridDehazeNet [28]", + "DehazeFormer [43]" + ], + "image_footnote": [], + "bbox": [ + 676, + 292, + 867, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/515484acf83bc484a7a14c0ca17a927119d75ae605e026bbce7219533cbde948.jpg", + "image_caption": [ + "Figure 7. Transferability across real-world fog datasets. Test results of models trained on SynFog.", + "Figure 8. Ablation of sensor noise simulation." + ], + "image_footnote": [], + "bbox": [ + 78, + 420, + 470, + 465 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/81fc69b965ae4f30773a69e11d6a7be90968653c5ebb42aa40da06a2e9e8f2e3.jpg", + "table_caption": [ + "Table 3. Detection evaluations on real-world fog datasets. \"FZ\" refers to Foggy Zurich. \"STF\" represents Seeing Through Fog." + ], + "table_footnote": [], + "table_body": "
MethodTraining SetFZ [11]STF [5]
mAP (%)mAP (%)
AECRNetFoggy Cityscapes69.754.8
Virtual KITTI68.953.3
SynFog71.555.5
DehazeFormerFoggy Cityscapes67.954.9
Virtual KITTI68.553.1
SynFog-β59.755.3
SynFog69.755.3
", + "bbox": [ + 80, + 529, + 467, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "depth and focus solely on the fog simulation method itself. The foggy image is synthesized from its clear counterpart in SynFog according to Eq. (1). Specifically, we employ the same airlight estimation method used in the construction of Foggy Cityscapes [40], which involves computing the median of all the $0.1\\%$ pixels with the largest dark channel values, as proposed in [45]. We use DehazeFormer [43] to train on both SynFog and SynFog- $\\beta$ , followed by testing the trained model on real-world foggy images. As shown in Fig. 9, image output by model trained on SynFog exhibits a better visual appearance compared to that of SynFog- $\\beta$", + "bbox": [ + 75, + 669, + 468, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study of Sensor Noise", + "text_level": 1, + "bbox": [ + 76, + 845, + 352, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the importance of sensor noise simulation in our pipeline, we create SynFog- $\\alpha$ dataset. During the", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/75844c7d8da63cc82c2f6a7525d4b618fc92bf10940ce8412e21833bb043ba98.jpg", + "table_caption": [ + "Table 4. Impact of sensor noise on defogging performance." + ], + "table_footnote": [], + "table_body": "
Experimental settingO-Haze [2]
PSNR ↑SSIM ↑DHQI ↑
AECRNet+SynFog(w/o noise)15.160.579537.53
AECRNet+SynFog(w/ noise)15.430.611644.46
", + "bbox": [ + 501, + 439, + 905, + 500 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4f2dab9c2e31615eb6c3a919d70cc94867f4a94311bee688fa82498b6281bd1f.jpg", + "table_caption": [ + "Table 5. Impact of sensor noise on detection performance. \"FZ\" refers to Foggy Zurich. \"STF\" represents Seeing Through Fog." + ], + "table_footnote": [], + "table_body": "
Experimental settingFZ [11]STF [5]
mAP (%)mAP (%)
AECRNet+SynFog(w/o noise)69.554.7
AECRNet+SynFog(w/ noise)71.555.5
", + "bbox": [ + 516, + 541, + 874, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "data generation process, the sensor-induced noise is set to zero, while all other settings remain consistent with those of the original SynFog dataset. We compare the defogging and detection performance of models trained on SynFog and SynFog- $\\alpha$ datasets, as shown in Fig. 8 and Tab. 4-5. The model trained on SynFog that incorporates sensor noise produces more natural results with fewer artifacts and higher detection accuracy.", + "bbox": [ + 496, + 611, + 890, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Algorithm Benchmarking", + "text_level": 1, + "bbox": [ + 500, + 739, + 735, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use SynFog test set to assess the aforementioned representative defogging methods, employing both full-reference and no-reference metrics to evaluate the defogged results. To illustrate the challenges presented by SynFog, we also provide evaluation results from the widely-used RESIDE dataset [25]. The full-scale quantitative comparison results are available in the supplementary material.", + "bbox": [ + 496, + 763, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generally, learning-based methods consistently outperform earlier algorithms based on natural or statistical priors", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21769", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e07ca0b11678860c79835939bbf540d43c3deca8b839c233bd62950de2c71362.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 246, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c5c2cd1ee05d201650cec21122bf3748f4a806a7c6dc809b08ec2dd3cca7b52.jpg", + "image_caption": [ + "Foggy Cityscapes [40]" + ], + "image_footnote": [], + "bbox": [ + 246, + 88, + 405, + 228 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4997464cd8e9ce6b3a37cac63b0fd43b0009085fb91597498fef1b0d1c38cff1.jpg", + "image_caption": [ + "Virtual KITTI [18]" + ], + "image_footnote": [], + "bbox": [ + 406, + 89, + 563, + 228 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fdba5094e0c1b731c62ccd2b5f282dbd6aaffd1363bb521d28053f4e9c24d86a.jpg", + "image_caption": [ + "SynFog-β" + ], + "image_footnote": [], + "bbox": [ + 565, + 89, + 722, + 228 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/94829d71ccfa1bb7c332a075e0a98a297a365025a39765727253e536c8eda625.jpg", + "image_caption": [ + "SynFog (Ours)" + ], + "image_footnote": [], + "bbox": [ + 723, + 89, + 882, + 228 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/116bde5ef6776d2eb1f4ac1357261707c837680ad51c734e1eb86a4bb38ad9d9.jpg", + "image_caption": [ + "Figure 9. Visual and detection comparisons on real-world foggy images using DehazeFormer models trained on SynFog, SynFog- $\\beta$ , Foggy Cityscapes [40] and Virtual KITTI [18]." + ], + "image_footnote": [], + "bbox": [ + 86, + 295, + 176, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd12f59177431e7fb1b49ca9b57a17c01cc7913056c61742ff142a7b53984d87.jpg", + "image_caption": [ + "Figure 10. Qualitative comparisons of SOTA defogging methods on SynFog dataset with different fog levels. Zoom in for better view." + ], + "image_footnote": [], + "bbox": [ + 176, + 296, + 264, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/69bf9067694c72d7790432942b68d68295dd3f305ddf59c4984c8cdf7c2c6222.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 296, + 351, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5ab8230ca5c33c9d8b8bad43cd490f928f71e3dbb9adb5c99b19c84771782179.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 296, + 439, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f1c4b5bc22508ce06538c71a12e8dba023c246d35663e717f8807ec49582fd3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 441, + 296, + 527, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cffffc0a4cf8cd4aa56d546671d0c40aa1c05ace67aecea7326ccb43a6561ede.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 296, + 616, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e4c33cea8ea00fe764c834af25cde06857646179ee7679837bb8cb3a525c0e3c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 296, + 705, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/204d86f1104f4224248e1368d9d5cbd9ef2b9168f4bcbf77c9eb84b50f023896.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 296, + 792, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e97b2a3b5d69573f987a245210488ab488a0de8f144e894c9796becfb9f29f2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 794, + 296, + 882, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in terms of PSNR, SSIM and LPIPS [54]. In particular, DehazeFormer [43] attains the highest PSNR (27.96) and SSIM (0.923) values, while GridDehazeNet [28] achieves the best LPIPS (0.163) value on SynFog test set. Although the overall performance trend of the algorithms on SynFog is similar to that on RESIDE [25], the evaluation scores are noticeably lower, highlighting the challenges posed by SynFog. Fig. 10 depicts the qualitative comparison among these methods across different fog levels. Images restored by DCP [20] and CAP [57] tend to appear excessively dark and noisy due to inaccurate airlight estimation and the lack of effective denoising capabilities. GridDehazeNet [28] lacks color recovery ability in dense fog, while the outputs of FFA-Net [39] and DehazeFormer [43] seem smooth. This comparison further indicates the substantial room for improvement in existing defogging methods, especially when applied to more realistic fog datasets.", + "bbox": [ + 75, + 505, + 472, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Discussion and Limitations", + "text_level": 1, + "bbox": [ + 76, + 771, + 321, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experimental results reveal that models trained on SynFog, incorporating realistic physical characteristics of optics and sensors, exhibit a more robust performance on real-word fog datasets in terms of visual perception and detection accuracy. This outcome implies that integrating a physically-based end-to-end imaging process replicating real capture methods can significantly improve the generalization of", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "trained models. However, given the time-consuming nature of the physically-based volumetric path tracing technique, the scale of SynFog is not as extensive as other synthetic fog datasets. Nonetheless, we are committed to expanding the scene contents in the future and exploring a broader range of scattering mediums beyond fog.", + "bbox": [ + 496, + 506, + 890, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 604, + 617, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present an end-to-end foggy image simulation pipeline. Our approach utilizes volumetric path tracing to model a more precise light scattering process with global illumination. By incorporating a physically-based camera processing pipeline that includes optics, sensor and image processing, we can closely mimic the authentic capture process under foggy conditions. Additionally, we develop a new synthetic fog dataset, SynFog, to facilitate the research on defogging. Comprehensive experiments have validated the authenticity and reliability of the SynFog dataset.", + "bbox": [ + 496, + 631, + 890, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This work is supported by 1) National Natural Science Foundation of China (Grant No. 62131011 and 62375233); 2) NSFC Young Scientists Fund (Grant No. 62302423). We thank Professor Chuxi Yang and Jiayue Xie for their help in preparing the scene creation. We also thank Professor Brian Wandell, Dr. Devesh Upadhyay, Dr. Alireza Rahimpour, and Devesh Shah for their valuable advice at the early stage of this project.", + "bbox": [ + 496, + 787, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21770", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 118, + 174, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Cosmin Ancuti, Codruta O Ancuti, and Christophe De Vleeschouwer. D-hazy: A dataset to evaluate quantitatively dehazing algorithms. In IEEE International Conference on Image Processing, pages 2226-2230, 2016. 2", + "[2] Codruta O Ancuti, Cosmin Ancuti, Radu Timofte, and Christophe De Vleeschouwer. O-haze: a dehazing benchmark with real hazy and haze-free outdoor images. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 754-762, 2018. 1, 2, 4, 7", + "[3] Codruta O Ancuti, Cosmin Ancuti, Mateu Sbert, and Radu Timofte. Dense-haze: A benchmark for image dehazing with dense-haze and haze-free images. In IEEE International Conference on Image Processing, pages 1014-1018, 2019.", + "[4] Codruta O Ancuti, Cosmin Ancuti, and Radu Timofte. Nh-haze: An image dehazing benchmark with nonhomogeneous hazy and haze-free images. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 444-445, 2020. 1, 2, 4", + "[5] Mario Bijelic, Tobias Gruber, Fahim Mannan, Florian Kraus, Werner Ritter, Klaus Dietmayer, and Felix Heide. Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11682-11692, 2020. 1, 3, 6, 7", + "[6] Norbert Buch, Sergio A Velastin, and James Orwell. A review of computer vision techniques for the analysis of urban traffic. IEEE Transactions on Intelligent Transportation Systems, 12(3):920-939, 2011. 1", + "[7] Ting Chen. Digital camera system simulator and applications. Stanford University, 2003. 4", + "[8] Lark Kwon Choi, Jaehee You, and Alan Conrad Bovik. Referenceless prediction of perceptual fog density and perceptual image defogging. IEEE Transactions on Image Processing, 24(11):3888-3901, 2015. 3", + "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3213-3223, 2016. 2", + "[10] Dengxin Dai and Wen Yang. Satellite image classification via two-layer sparse coding with biased image representation. IEEE Geoscience and Remote Sensing Letters, 8(1): 173-176, 2010. 1", + "[11] Dengxin Dai, Christos Sakaridis, Simon Hecker, and Luc Van Gool. Curriculum model adaptation with synthetic and real data for semantic foggy scene understanding. International Journal of Computer Vision, 128:1182-1204, 2020. 2, 6, 7", + "[12] Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming-Hsuan Yang. Multi-scale boosted dehazing network with dense feature fusion. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2157-2167, 2020. 6" + ], + "bbox": [ + 78, + 143, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] J.B. Farrow et al. Influence of the atmosphere on optical systems. Optica Acta, 1970. 5", + "[14] X. Min et al. Objective quality evaluation of dehazed images. TITS, 2018. 7", + "[15] Joyce E Farrell and Brian A Wandell. Image systems simulation. Handbook of Digital Imaging, 1:373-400, 2015. 4", + "[16] Joyce E Farrell, Feng Xiao, Peter Bert Catrysse, and Brian A Wandell. A simulation tool for evaluating digital camera image quality. In Image Quality and System Performance, pages 124-131. SPIE, 2003.", + "[17] Joyce E Farrell, Peter B Catrysse, and Brian A Wandell. Digital camera simulation. Applied Optics, 51(4):A80-A90, 2012. 3, 4", + "[18] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4340-4349, 2016. 2, 5, 6, 8", + "[19] Martin Hahner, Dengxin Dai, Christos Sakaridis, Jan-Nico Zaech, and Luc Van Gool. Semantic understanding of foggy scenes with purely synthetic data. In IEEE Intelligent Transportation Systems Conference (ITSC), pages 3675-3681, 2019. 2, 3", + "[20] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(12):2341-2353, 2010. 6, 8", + "[21] Louis G Henyey and Jesse Leonard Greenstein. Diffuse radiation in the galaxy. Astrophysical Journal, vol. 93, p. 70-83 (1941), 93:70-83, 1941. 4", + "[22] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolo, 2023. 6", + "[23] Kyung Won Kim. The comparison of visibility measurement between image-based visual range, human eye-based visual range, and meteorological optical range. Atmospheric Environment, 190:74-86, 2018. 5", + "[24] Miles V Klein and Thomas E Furtak. Optik. Springer-Verlag, 2013. 4", + "[25] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2018. 1, 2, 3, 7, 8", + "[26] Ruoteng Li, Xiaoyi Zhang, Shaodi You, and Yu Li. Realistic large-scale fine-depth dehazing dataset from 3d videos. arXiv preprint arXiv:2004.08554, 2020. 2", + "[27] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(10):2024-2039, 2015. 2", + "[28] Xiaohong Liu, Yongrui Ma, Zhihao Shi, and Jun Chen. Griddehazenet: Attention-based multi-scale network for image dehazing. In IEEE International Conference on Computer Vision, pages 7314-7323, 2019. 6, 8", + "[29] Yun Liu, Zhongsheng Yan, Aimin Wu, Tian Ye, and Yuche Li. Nighttime image dehazing based on variational decomposition model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 640-649, 2022. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21771", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Zhenyi Liu, Trisha Lian, Joyce Farrell, and Brian Wandell. Soft prototyping camera designs for car detection based on a convolutional neural network. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 4", + "[31] Zhenyi Liu, Minghao Shen, Jiaqi Zhang, Shuangting Liu, Henryk Blasinski, Trisha Lian, and Brian Wandell. A system for generating complex physically accurate sensor images for automotive applications. arXiv preprint arXiv:1902.04258, 2019. 4", + "[32] Zheng Lyu, Thomas Goossens, Brian A Wandell, and Joyce Farrell. Validation of physics-based image systems simulation with 3-d scenes. IEEE Sensors Journal, 22(20):19400-19410, 2022. 4", + "[33] Earl J McCartney. Optics of the atmosphere: scattering by molecules and particles. New York, 1976. 1, 3, 5, 6", + "[34] Srinivasa G Narasimhan and Shree K Nayar. Vision and the atmosphere. International Journal of Computer Vision, 48: 233-254, 2002. 3, 5, 6", + "[35] Priya Narayanan, Xin Hu, Zhenyu Wu, Matthew D Thielke, John G Rogers, Andre V Harrison, John A D'Agostino, James D Brown, Long P Quang, James R Uplinger, et al. A multi-purpose realistic haze benchmark with quantifiable haze levels and ground truth. IEEE Transactions on Image Processing, 2023. 1, 2", + "[36] National Standards of People's Republic of China. Grade of fog forecast, 2011. 5", + "[37] Roberto Nebuloni. Empirical relationships between extinction coefficient and visibility in fog. Applied Optics, 44(18): 3795-3804, 2005. 5", + "[38] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically based rendering: From theory to implementation. MIT Press, 2023. 2, 3, 4", + "[39] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In AAAI, pages 11908-11915, 2020. 6, 8", + "[40] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Semantic foggy scene understanding with synthetic data. International Journal of Computer Vision, 126:973-992, 2018. 1, 2, 3, 5, 6, 7, 8", + "[41] Daniel Scharstein and Richard Szeliski. High-accuracy stereo depth maps using structured light. In IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages I-I. IEEE, 2003. 2", + "[42] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In European Conference on Computer Vision, pages 746-760. Springer, 2012. 2", + "[43] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927-1941, 2023. 6, 7, 8", + "[44] Tao Sun, Mattia Segu, Janis Postels, Yuxuan Wang, Luc Van Gool, Bernt Schiele, Federico Tombari, and Fisher Yu. Shift: a synthetic driving dataset for continuous multi-task domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 21371-21382, 2022. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Ketan Tang, Jianchao Yang, and Jue Wang. Investigating haze-relevant features in a learning framework for image dehazing. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2995-3000, 2014. 7", + "[46] Jean-Philippe Tarel, Nicolas Hautiere, Aurélien Cord, Dominique Gruyer, and Houssam Halmaoui. Improved visibility of road scene images under heterogeneous fog. In IEEE Intelligent Vehicles Symposium, pages 478–485, 2010. 2", + "[47] Jean-Philippe Tarel, Nicholas Hautiere, Laurent Caraffa, Aurélien Cord, Houssam Halmaoui, and Dominique Gruyer. Vision enhancement in homogeneous and heterogeneous fog. IEEE Intelligent Transportation Systems Magazine, 4 (2):6-20, 2012. 2", + "[48] Tali Treibitz and Yoav Y Schechner. Resolution loss without imaging blur. JOSA A, 29(8):1516-1528, 2012. 3", + "[49] Apostolia Tsirikoglou, Joel Kronander, Magnus Wrenninger, and Jonas Unger. Procedural modeling and physically based rendering for synthetic data generation in automotive applications. arXiv preprint arXiv:1710.06270, 2017. 4", + "[50] Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Contrastive learning for compact single image dehazing. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10551-10560, 2021. 6", + "[51] Jing Zhang, Yang Cao, Zheng-Jun Zha, and Dacheng Tao. Nighttime dehazing with a synthetic benchmark. In Proceedings of the ACM International Conference on Multimedia, pages 2355-2363, 2020. 3", + "[52] Lin Zhang, Junhao Huang, Xiyuan Li, and Lu Xiong. Vision-based parking-slot detection: A dcnn-based approach and a large-scale benchmark dataset. IEEE Transactions on Image Processing, 27(11):5350-5364, 2018. 1", + "[53] Lin Zhang, Anqi Zhu, Shiyu Zhao, and Yicong Zhou. Simulation of atmospheric visibility impairment. IEEE Transactions on Image Processing, 30:8713-8726, 2021. 2, 3", + "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 8", + "[55] Yanfu Zhang, Li Ding, and Gaurav Sharma. Hazerd: an outdoor scene dataset and benchmark for single image dehazing. In IEEE International Conference on Image Processing, pages 3205-3209, 2017. 1, 2", + "[56] Shiyu Zhao, Lin Zhang, Shuaiyi Huang, Ying Shen, and Shengjie Zhao. Dehazing evaluation: Real-world benchmark datasets, criteria, and baselines. IEEE Transactions on Image Processing, 29:6947-6962, 2020. 7", + "[57] Qingsong Zhu, Jiaming Mai, and Ling Shao. A fast single image haze removal algorithm using color attenuation prior. IEEE Transactions on Image Processing, 24(11):3522-3533, 2015. 6, 8" + ], + "bbox": [ + 501, + 92, + 890, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21772", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/30117727-713c-4f8f-a215-de682288f1dd_model.json b/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/30117727-713c-4f8f-a215-de682288f1dd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..891a890516e028f92e8b78783ec5aa5f85e6fea0 --- /dev/null +++ b/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/30117727-713c-4f8f-a215-de682288f1dd_model.json @@ -0,0 +1,2596 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.131, + 0.882, + 0.177 + ], + "angle": 0, + "content": "SynFog: A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.203, + 0.816, + 0.282 + ], + "angle": 0, + "content": "Yiming Xie\\(^{1*}\\), Henglu Wei\\(^{1*}\\), Zhenyi Liu\\(^{2}\\), Xiaoyu Wang\\(^{1}\\), Xiangyang Ji\\(^{1}\\) \n\\(^{1}\\)Tsinghua University, \\(^{2}\\)Stanford University \n{xieym22@mails., weihenglu@mail., xiaoyu-w17@mails., xyji@}tsinghua.edu.cn, {zhenyiliu@}stanford.edu" + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.317, + 0.885, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.471, + 0.893, + 0.514 + ], + "angle": 0, + "content": "Figure 1. The SynFog dataset comprises images under different lighting, with the first row captured in natural sky light and the second with extra light sources like street lamps and automotive lighting. This dataset covers three levels of fog density across each scene, corresponding visibility of \\(600\\mathrm{m}\\), \\(300\\mathrm{m}\\), \\(150\\mathrm{m}\\). Additionally, pixel-accurate depth data and segmentation labels for each scene are also provided." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.521, + 0.314, + 0.537 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.473, + 0.832 + ], + "angle": 0, + "content": "To advance research in learning-based defogging algorithms, various synthetic fog datasets have been developed. However, existing datasets created using the Atmospheric Scattering Model (ASM) or real-time rendering engines often struggle to produce photo-realistic foggy images that accurately mimic the actual imaging process. This limitation hinders the effective generalization of models from synthetic to real data. In this paper, we introduce an end-to-end simulation pipeline designed to generate photorealistic foggy images. This pipeline comprehensively considers the entire physically-based foggy scene imaging process, closely aligning with real-world image capture methods. Based on this pipeline, we present a new synthetic fog dataset named SynFog, which features both sky light and active lighting conditions, as well as three levels of fog density. Experimental results demonstrate that models trained on SynFog exhibit superior performance in visual perception and detection accuracy compared to others when applied to real-world foggy images." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.843, + 0.21, + 0.859 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Fog, due to its light scattering and attenuation properties, poses a significant challenge for outdoor vision-based sys" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.893, + 0.613 + ], + "angle": 0, + "content": "tems such as remote sensing, surveillance, autonomous driving, and fire rescue [6, 10, 52]. It causes severe degradation in object appearance and contrast, resulting in a notable impact on people's daily lives. Consequently, there has been a growing research interest in image defogging task to develop robust outdoor vision systems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Recent learning-based defogging algorithms depend on sets of foggy images and corresponding clear images for model training. While several datasets have been collected from real-world scenarios [5, 40], they are typically limited to foggy images only, as it is almost impossible to obtain pixel-accurate clear images with consistent scene contents and environmental illumination. Studies in [2-4, 35] attempt to capture real-world foggy and clear image pairs using smoke machines. However, the scale of these datasets is inadequate for training of a high-performing model. Given the challenge of obtaining paired foggy and clear images in real world, various synthetic fog datasets have been proposed as an alternative [25, 40, 55]. The most widely used method is based on the theory proposed by McCartney [33], also known as the Atmospheric Scattering Model (ASM). However, this model fails to consider accurate global illumination and the actual imaging process that occurs in real-world captures as detailed in Sec. 3.1. These limitations significantly impact the model's overall applicability, par" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21763" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.087, + 0.482, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.47, + 0.326 + ], + "angle": 0, + "content": "Figure 2. Relationship between noise level and fog density in real-world foggy images. Using regional contrast stretching for image enhancement, it can be observed that the noise level increases with the concentration of fog." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.469, + 0.399 + ], + "angle": 0, + "content": "ticularly in nighttime scenarios, and can lead to a disparity between synthetic and real-world foggy images. As a result, models trained on synthetic foggy images may exhibit limited robustness when applied to real-world ones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.47, + 0.732 + ], + "angle": 0, + "content": "To address these issues, we propose an end-to-end foggy image simulation pipeline that encompasses the entire imaging process starting from 3D virtual scenes to final photorealistic foggy images. Our approach involves foggy scene rendering using volumetric path tracing and imaging through a realistic camera model. As shown in Fig. 3b, fog is treated as a spatial volume, and the spectral scene radiance is rendered using volumetric path tracing [38]. By defining the phase function of fog medium and bidirectional reflectance distribution function (BRDF) of the scene components, we are able to accurately describe the scattering process with multiple light sources under foggy conditions. The foggy scene radiance data is further processed through a physically-based camera model, which comprises optics, sensor and image processing, to faithfully replicate real camera devices. Based on the proposed pipeline, we develop a new extensive synthetic fog dataset called SynFog. This dataset contains 500 unique outdoor scenes, each featuring two kinds of lighting conditions and three levels of fog density. Extensive experiments have been conducted to qualitatively and quantitatively validate the authenticity of our simulation pipeline and the SynFog dataset." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.733, + 0.435, + 0.746 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.748, + 0.468, + 0.806 + ], + "angle": 0, + "content": "- We propose a physically-based end-to-end foggy image simulation pipeline that incorporates accurate light transportation in scattering medium and physical characteristics of optics and sensor into the synthesized images." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.808, + 0.468, + 0.851 + ], + "angle": 0, + "content": "- We develop a photo-realistic synthetic fog dataset, SynFog, which includes both sky light and active lighting conditions, as well as three levels of fog density." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.854, + 0.468, + 0.897 + ], + "angle": 0, + "content": "- We demonstrate the authenticity and effectiveness of our simulation pipeline and the superior realism of SynFog dataset compared to other synthetic fog datasets." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.748, + 0.468, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.642, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.205 + ], + "angle": 0, + "content": "Given the challenge of acquiring paired foggy and clear images in the real world for training learning-based defogging methods, researchers have proposed many synthetic fog datasets as an alternative. In this section, we will offer a thorough review and analysis of foggy image simulation techniques and their associated datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.206, + 0.893, + 0.674 + ], + "angle": 0, + "content": "Based on Atmospheric Scattering Model (ASM). This method involves rendering fog onto clean images along with their depth information [1, 19, 25, 26, 40, 53, 55]. Depth information can be acquired through game engines, depth cameras or depth estimation methods. In an earlier study conducted by Tarel et al. [46], FRIDA was constructed using a total of 90 synthetic images, with clear images and depth maps generated using \\( SiVIC^{TM} \\) software. Subsequently, in their later work [47], more synthetic images were added, leading to an upgraded version named FRIDA2, which includes 330 synthetic images. However, the scale of these datasets is insufficient for modern data-driven defogging methods, and the foggy images appear unrealistic. In more recent works [26, 53], a similar approach has been employed to acquire both the clear images and their corresponding depth maps from game engines. In [25], a general dehazing dataset called RESIDE was introduced. The indoor training set (ITS) of RESIDE was constructed based on existing indoor RGBD datasets NYUv2 [42] and Middleburry stereo [41]. For the outdoor training set (OTS), 2071 real-world outdoor images were collected, and the corresponding depth information was estimated using a learning-based technique [27]. It is important to note that the accuracy of depth estimation based on learning methods is still unsatisfactory, as single-image-based depth estimation is inherently an ill-posed problem. Furthermore, Foggy Cityscapes was introduced in [40], utilizing images from the Cityscapes dataset [9]. Despite efforts to enhance the original depth maps in Cityscapes, artifacts and discrepancies still persist, resulting in synthetic foggy images that appear unnatural." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Based on Real-time Rendering Engines. Foggy images generated utilizing real-time rendering engines like Unity and Unreal have been discussed in works such as [18, 44]. Among them, Virtual KITTI 2 [18] stands out as a notable example. However, these foggy images often lack realism and exhibit limited diversity in terms of fog density." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Based on Physical Means. Fog/haze machines are employed in real-world settings to generate artificial fog, as discussed in works such as [2-4, 35]. However, these datasets are generally limited in scale and may not offer sufficient supervision for learning-based defogging methods. Moreover, ensuring consistent content between the clear and foggy image captures can be challenging." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "There are also several real-world fog datasets available, such as Foggy Driving [40], Foggy Zurich [11], and Seeing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21764" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.095, + 0.429, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.258, + 0.399, + 0.272 + ], + "angle": 0, + "content": "(a) Atmospheric Scattering Model (ASM) [34]." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.283, + 0.431, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.421, + 0.432, + 0.435 + ], + "angle": 0, + "content": "(b) Foggy scene rendering based on volumetric path tracing." + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.446, + 0.402, + 0.461 + ], + "angle": 0, + "content": "Figure 3. Foggy scene simulation methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.584 + ], + "angle": 0, + "content": "Through Fog [5]. However, due to the absence of corresponding clear reference images, these datasets can only be used for testing purposes rather than for training. In this paper, we propose a physically-based end-to-end foggy image simulation pipeline to improve the authenticity of synthetic foggy images and construct a photo-realistic fog dataset to facilitate future research on defogging algorithms." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.6, + 0.406, + 0.618 + ], + "angle": 0, + "content": "3. End-to-end Foggy Image Simulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.826 + ], + "angle": 0, + "content": "We propose an end-to-end approach for generating photorealistic foggy images that incorporates accurate light transportation in scattering medium and physical characteristics of optics and sensor into the synthesized images, as illustrated in Fig. 4. We first introduce a foggy scene imaging model in Sec. 3.1. Based on this model, we establish a simulation pipeline consisting of two components: a) Foggy scene radiance is rendered using volumetric path tracing [38]. b) The radiance data is processed through a physically-based camera model, which comprises optics, sensor and image processing to faithfully replicate real camera devices [17]. These components are further detailed in Sec. 3.2.1 and Sec. 3.2.2, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.332, + 0.861 + ], + "angle": 0, + "content": "3.1. Foggy Scene Imaging Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.903 + ], + "angle": 0, + "content": "The Atmospheric Scattering Model (ASM) [33, 34] has been widely used for the generation of foggy images [19," + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.092, + 0.58, + 0.107 + ], + "angle": 0, + "content": "25,40,53]:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.121, + 0.892, + 0.179 + ], + "angle": 0, + "content": "\\[\nE (d, \\lambda) = \\underbrace {E _ {o} (\\lambda) e ^ {- \\beta (\\lambda) d}} _ {\\text {t r a n s m i s s i o n}} + \\underbrace {L _ {h} (\\infty , \\lambda) \\left(1 - e ^ {- \\beta (\\lambda) d}\\right)} _ {\\text {a i r l i g h t}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.18, + 0.892, + 0.317 + ], + "angle": 0, + "content": "where \\(E\\) is the irradiance reaching the film under foggy conditions, \\(E_{o}\\) is the irradiance under clear conditions, \\(\\beta\\) is the scattering coefficient characterizing the scatter ability of the medium, \\(d\\) is the distance between the object and the camera, \\(L_{h}(\\infty ,\\lambda)\\) denotes the radiance of airlight at infinity and \\(\\lambda\\) refers to wavelength. Fig. 3a provides a more intuitive description of this process. Foggy images can be generated with pre-defined \\(\\beta\\) and \\(L_{h}\\) values by directly using the clear images as \\(E_0\\) and the corresponding depth maps as \\(d\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.893, + 0.469 + ], + "angle": 0, + "content": "However, there are two issues with this scattering model for generating foggy images: (1) The definition of environmental illumination fails to consider the contribution of indirect illumination from the scene to the scattering medium, resulting in an unfaithful representation of the actual illumination. (2) The current formulation of airlight is insufficient for accurately modeling the lighting conditions that involve multiple point light sources, such as street lamps and automotive lights. This limitation restricts the model's applicability in situations like nighttime [29, 51]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.622 + ], + "angle": 0, + "content": "Furthermore, ASM only captures the irradiance reaching the sensor plane. Consequently, foggy images generated using ASM fail to accurately incorporate the realistic physical characteristics of optics and sensor in actual imaging process. In Fig. 2, we depict the relationship between noise level and fog density after the enhancement of regional contrast stretching [48] on a real captured image [8]. It can be seen that the noise level increases with higher fog density, emphasizing the critical need for accurate noise modeling in foggy image simulation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.892, + 0.684 + ], + "angle": 0, + "content": "To enhance the consistency between synthetic images and real-world captures, we develop a comprehensive foggy scene imaging model that incorporates the entire imaging process, referred to as:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.701, + 0.891, + 0.718 + ], + "angle": 0, + "content": "\\[\nI _ {o u t} = f _ {I S P} \\left(f _ {\\text {s e n s o r}} \\left(f _ {\\text {o p t i c s}} \\left(L _ {\\text {s c e n e}}\\right)\\right) + n\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.809 + ], + "angle": 0, + "content": "where \\( L_{\\text{scene}} \\) is the scene radiance obtained using volumetric path tracing [38], as described in Sec. 3.2.1; \\( \\{f_{\\text{optics}}, f_{\\text{sensor}}, f_{\\text{ISP}}\\} \\) denote the optics, sensor and ISP processes respectively, as described in Sec. 3.2.2, and the noise term introduced by the sensor is denoted as \\( n \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.902 + ], + "angle": 0, + "content": "By employing volumetric path tracing to obtain the radiance of foggy scene, we can precisely model the light scattering process under global illumination and active lighting. By incorporating the complete imaging process described in Eq. (2), we are able to faithfully replicate the authentic camera characteristics in foggy image simulation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21765" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.328, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.09, + 0.495, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.09, + 0.681, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.09, + 0.891, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.185, + 0.894, + 0.228 + ], + "angle": 0, + "content": "Figure 4. End-to-end foggy image simulation pipeline. The spectral radiance data is rendered using volumetric path tracing and passing through a realistic optics model before reaching the sensor plane. Subsequently, the irradiance is converted into a digital image through the sensor model. The raw image from the sensor is then processed by an ISP module to generate the final foggy image." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.242, + 0.356, + 0.259 + ], + "angle": 0, + "content": "3.2. End-to-end Simulation Pipeline" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.266, + 0.296, + 0.282 + ], + "angle": 0, + "content": "3.2.1 Foggy Scene Rendering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.47, + 0.44 + ], + "angle": 0, + "content": "We utilize procedural modeling techniques, similar to the approach employed in [31, 49], to create large-scale and diverse driving scenes. This involves defining adjustable parameters and using a set of predefined rules to generate explicit scene definitions. Our model library incorporates a diverse range of pedestrians, bicyclists and cars, each with well-defined material and texture properties. Furthermore, we carefully select 16 sky maps with cloudy conditions to ensure realistic lighting for foggy scenes, following the established guidelines from previous works [2-4]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.469, + 0.531 + ], + "angle": 0, + "content": "In our approach, fog is treated as a spatial volume, and the spectral scene radiance is rendered using volumetric path tracing [38]. To accurately model the scattering process of the fog medium, we employ Henyey and Greenstein phase function [21, 38] to model its particle scattering characteristics, denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.541, + 0.469, + 0.579 + ], + "angle": 0, + "content": "\\[\np _ {\\mathrm {H G}} (\\cos \\theta) = \\frac {1}{4 \\pi} \\frac {1 - g ^ {2}}{\\left(1 + g ^ {2} + 2 g (\\cos \\theta)\\right) ^ {3 / 2}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.588, + 0.469, + 0.649 + ], + "angle": 0, + "content": "By empirically setting the asymmetry parameter \\( g = 0.87 \\), we can accurately simulate the probability distributions for scattering in a specific direction at a given point within the fog medium, as shown in Fig. 3b." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.649, + 0.469, + 0.71 + ], + "angle": 0, + "content": "The rendering equation consists of two parts that account for reduction of radiance caused by out-scattering and increase of radiance due to in-scattering at point \\( p \\) in direction \\( \\omega \\), represented by Eq. (4) and Eq. (5), respectively [38]." + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.72, + 0.469, + 0.74 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {i}} \\left(\\mathrm {p} ^ {\\prime}, - \\omega\\right) = \\mathrm {e} ^ {- \\int_ {0} ^ {d} \\sigma_ {\\mathrm {s}} (\\mathrm {p} + t \\omega , \\omega) \\mathrm {d} t} L _ {\\mathrm {o}} (\\mathrm {p}, \\omega). \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.751, + 0.483, + 0.794 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {s}} (\\mathrm {p}, \\omega) = L _ {\\mathrm {e}} (\\mathrm {p}, \\omega) + \\sigma_ {\\mathrm {s}} (\\mathrm {p}, \\omega) \\int_ {\\mathcal {S} ^ {2}} p (\\mathrm {p}, \\omega_ {\\mathrm {i}}, \\omega) L _ {\\mathrm {i}} (\\mathrm {p}, \\omega_ {\\mathrm {i}}) \\mathrm {d} \\omega_ {\\mathrm {i}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.796, + 0.469, + 0.856 + ], + "angle": 0, + "content": "\\(\\sigma_{s}\\) represents the scattering probability per unit distance, \\(\\{L_i,L_o,L_e,L_s\\}\\) are the {incident,exitant,emission,total added} radiance. In our implementation, we employ the Monte Carlo algorithm to solve the above integration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The key distinction between our physically-based foggy scene rendering technique and ASM lies in the modeling of airlight, as represented by Eq. (5). Our approach offers a" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.242, + 0.894, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.345, + 0.826, + 0.358 + ], + "angle": 0, + "content": "Figure 5. Noise flow and sensor simulation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.374, + 0.892, + 0.449 + ], + "angle": 0, + "content": "more accurate depiction of light scattering process through the utilization of specific phase function, and enables incorporation of global illumination and multiple light sources within the scene, producing authentic scene radiance to the camera model described in the next section." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.465, + 0.877, + 0.48 + ], + "angle": 0, + "content": "3.2.2 Imaging through Realistic Camera Simulation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.488, + 0.892, + 0.594 + ], + "angle": 0, + "content": "After obtaining the scene radiance, we employ ISETCam [15-17, 30, 32] to simulate the realistic digital camera system that encompasses optics, sensor and ISP components. Specifically, we use a diffraction-limited optics model [7] with consideration of off-axis irradiance attenuation based on the cosine-fourth law [24] to convert scene radiance \\( L_{o} \\) to optical image irradiance \\( E_{i} \\), denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.603, + 0.891, + 0.636 + ], + "angle": 0, + "content": "\\[\nE _ {i} = \\pi \\cdot L _ {o} \\frac {1}{1 + 4 (f / \\# (1 - m)) ^ {2}} (\\cos \\phi) ^ {4}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.644, + 0.891, + 0.703 + ], + "angle": 0, + "content": "where \\( f / \\# \\) is the f-number, \\( m \\) represents the magnification defined as the negative ratio of image distance and object distance. In this optics model, blur mainly depends on the f-number value but little else." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.854 + ], + "angle": 0, + "content": "We model the sensor process including photovoltaic conversion, analog amplifier, and analog-to-digital conversion as described in Fig. 4 and simulate the real noise characteristics depicted in Fig. 5. Among the various noise sources, shot noise arising from the photovoltaic conversion emerges as the dominant type of noise in foggy images which follows the Poisson distribution. To ensure the credibility of our sensor simulation, we incorporate a set of calibrated sensor parameters from the Mi10Pro cellphone, encompassing the CFA and infrared filter data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Finally, we process the RAW image output from the sensor with a simple ISP module, including black level compensation, white balance, demosaicing, color correction and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21766" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.118 + ], + "angle": 0, + "content": "Table 1. Comparisons between SynFog and existing synthetic fog datasets in the driving field." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.13, + 0.462, + 0.213 + ], + "angle": 0, + "content": "
DatasetsGlobal IlluminationOptics and Sensor Characteristics
Foggy Cityscapes××
Virtual KITTI 2××
SynFog (Ours)✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.231, + 0.468, + 0.26 + ], + "angle": 0, + "content": "gamma correction, to get the final synthetic image that mimics the real camera characteristics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.276, + 0.234, + 0.292 + ], + "angle": 0, + "content": "4. SynFog Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.47, + 0.572 + ], + "angle": 0, + "content": "By utilizing the proposed pipeline, we develop a new extensive synthetic fog dataset called SynFog. Example images are shown in Fig. 1. We first generate 500 large-scale, diverse driving scenes using procedural modeling techniques. For each scene, we generate two sets of images under different lighting conditions: one set with sky light only and the other set with both sky light and active light sources, such as street lamps and automotive lighting. Within each group, we create a set of four images: one clear image and three foggy images with varying fog levels defined by the scattering coefficient \\(\\sigma_{\\mathrm{s}}\\). According to McCartney's work [33], the scattering coefficient remains independent of the wavelength for fog [13, 34]. Therefore, we use a consistent scattering coefficient for wavelengths within the visible range. To ensure realistic simulation of fog, we employ the meteorological optical range (MOR) [23] or visibility as a reference to select appropriate scattering coefficients, described as:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.574, + 0.469, + 0.59 + ], + "angle": 0, + "content": "\\[\nM O R = 2. 9 9 6 / \\sigma_ {s}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.469, + 0.704 + ], + "angle": 0, + "content": "Based on this equation, we select scattering coefficient values of 0.005, 0.01, 0.02, which correspond to visibility range of \\(600\\mathrm{m}\\), \\(300\\mathrm{m}\\) and \\(150\\mathrm{m}\\), respectively. According to National Standard of China [36], these visibility ranges are classified as heavy fog, thick fog, and dense fog, thereby the simulated foggy images are diversified enough to reflect real-world complicated foggy conditions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We compare SynFog with existing fog datasets in the field of autonomous driving, including the extensively used Foggy Cityscapes [40] and Virtual KITTI 2 [18]. The comparison is presented in Tab. 1. In terms of foggy image generation, methods that rely on ASM [34] require accurate depth maps to calculate the correct transmission maps. However, obtaining accurate depth maps for outdoor images can be challenging, leading to artifacts such as holes and discontinuities in synthetic foggy images, similar to what is observed in Foggy Cityscapes [40]. On the other hand, foggy images generated using Game Engines like Unity or Unreal [18] typically employs rasterization techniques in the rendering process. While effective for creating" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "scenes, rasterization techniques may not accurately model light transportation as precisely as path tracing, which is employed in SynFog. Moreover, SynFog takes the realistic imaging process into account, resulting in highly authentic foggy images with realistic physical characteristics. The superior realism of SynFog is further supported by the experimental results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.892, + 0.319 + ], + "angle": 0, + "content": "Furthermore, our data simulation pipeline offers exceptional flexibility and control, allowing us to generate not only the final RGB images but also other valuable data modalities, including RAW images. This capability opens up avenues for research on defogging algorithms in the RAW domain, which has not been explored due to the lack of paired foggy RAW images and clear RGB images necessary for training deep defogging networks." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.335, + 0.634, + 0.351 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.36, + 0.726, + 0.376 + ], + "angle": 0, + "content": "5.1. Fog Chamber Validation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.892, + 0.672 + ], + "angle": 0, + "content": "We establish an indoor fog chamber to validate the fidelity of our fog simulation pipeline. The chamber is constructed as a capped glass cube with dimensions of \\(1.5m\\times 0.5m\\times\\) \\(0.5m\\). To reproduce the dark night environment that has directional light sources in the fog chamber, we position 4 light bars on each side of the front face of the cuboid fog chamber and enclose it with black blackout cloth from the inside. Fog inside the chamber is generated using a 1.7MHz ceramic atomizer that emulates water mist. To ensure uniformity of fog concentration within the chamber, we incorporate dual fans to promote even distribution. A visibility measurement module is designed to quantify the fog density. This module is comprised of a PM-160 wireless power meter from Thorlabs, an 852 narrowband laser, an aperture and a narrowband filter adapted to the laser band. The aperture serves to filter out stray light and the narrowband filter plays the role in blocking interference caused by non-laser light sources in the scene. The principle of visibility (scattering coefficient \\(\\sigma_{s}\\)) measurement is expressed as [37]:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.683, + 0.892, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {s} = \\frac {\\ln \\left(\\frac {P _ {0}}{P _ {u}}\\right)}{\\mathrm {u}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.779 + ], + "angle": 0, + "content": "where \\(P_0\\) is the laser power measured without fog and \\(P_u\\) is the measured laser power after the propagation of \\(u\\) distance under foggy conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We then use Blender to create a virtual fog chamber that can reproduce the real-world counterpart. Our process involves calibrating the light conditions, object information and camera properties prior to capture, resulting in a one-to-one reproduction of the real fog chamber. We select the standard Macbeth color checker as our test object and use Mi10Pro cellphone to capture images. We evaluate the matching degree between the simulation and reality un" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21767" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.473, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.28, + 0.284, + 0.292 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.3, + 0.468, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.452, + 0.47, + 0.522 + ], + "angle": 0, + "content": "Figure 6. Color analysis under foggy conditions. (a) Visual comparison between real captures and simulated images under different fog densities in linear RGB space. We increase the brightness of the images \\((\\times 2)\\) for visual purpose. (b) Analyzing the trend of color variation across different fog densities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.566 + ], + "angle": 0, + "content": "der clear conditions by conducting a color analysis on the checker, which is available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.469, + 0.718 + ], + "angle": 0, + "content": "After ensuring the consistency of initial capture conditions, we proceed to capture images under foggy conditions. We first record the laser power under clear condition as \\( P_0 \\), and then proceed to generate fog until the object is no longer visible. During the fog dispersion process, we capture images at regular intervals of 1 second and record the laser energy \\( P_u \\) to obtain real-time fog density measurements. We utilize the measured fog density parameter \\( \\sigma_s \\) to simulate corresponding foggy counterparts based on the proposed foggy image simulation pipeline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In Fig. 6a, we showcase both the real-captured foggy images and simulated ones generated using our pipeline and the ASM [33]. It can be seen that images simulated through our pipeline exhibit a closer resemblance to real-captured foggy images. We also conduct a color analysis on 24 color patches under different fog densities as depicted in Fig. 6b. The results demonstrate that, as fog density increases, the trend of each channel in simulated images generated by our pipeline consistently follows that of the real captures. In contrast, the simulated images based on ASM do not exhibit this trend. These findings provide a quantitative validation of the realism of our foggy image simulation method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.108 + ], + "angle": 0, + "content": "5.2. Transferability across the Real-to-Virtual Gap" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.31 + ], + "angle": 0, + "content": "We evaluate 7 representative state-of-the-art defogging algorithms on SynFog dataset: DCP [20], CAP [57], Grid-DehazeNet [28], MSBDN [12], FFA-Net [39], AECR-Net [50], and DehazeFormer [43]. We directly apply the models trained on SynFog to real foggy images sourced from Foggy Zurich [11] and Seeing Through Fog [5] datasets, without undergoing any model fine-tuning or domain adaptation. During the training process, we adhere to the official settings of these methods, and all of them demonstrate convergence by the end of the training. The results are shown in Fig. 7. Despite DCP [20] and CAP [57] producing excessively dark outputs, models trained on SynFog can produce natural defogged images with minimal artifacts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.892, + 0.462 + ], + "angle": 0, + "content": "To underscore the superior realism of SynFog compared to existing synthetic fog datasets, we further enhance our evaluation by training the top-performing models (AECR-Net [50], DehazeFormer [43]) on two prominent synthetic fog datasets relevant to autonomous driving (Foggy Cityscapes [40], Virtual KITTI [18]). Subsequently, these models are evaluated on real fog datasets for a comparative analysis. Additionally, we utilize ASM [34] to create a SynFog-\\(\\beta\\) dataset with clear images from SynFog and test it using DehazeFormer [43]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.893, + 0.749 + ], + "angle": 0, + "content": "Comparison with existing synthetic fog datasets. To mitigate any potential impact arising from model training issues, we initially conduct tests for each model using the corresponding test set of the training data. Test results are available in the supplementary material. Both AECR-Net [50] and DehazeFormer [43] exhibit high performance, confirming the correctness of the model training process. Subsequently, the trained models are directly applied to real-world test foggy images. Results are shown in Fig. 9 and Tab. 2. Owing to the high authenticity of the SynFog dataset, models trained on it can effectively generalize to real foggy images and produce naturally colored defogged images. Conversely, models trained on the other two datasets display poorer generalization to real data, resulting in artifacts and color distortion in defogged images. We also evaluate the defogged images using a pre-trained YOLOv8 [22] model for detection, showcasing how SynFog enhances machine vision capabilities (See Tab. 3). More results can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.854 + ], + "angle": 0, + "content": "It is noteworthy that the training set of Foggy Cityscapes [40] comprises 8925 data pairs, and Virtual KITTI [18] includes 3452 data pairs, both exceeding the 1350 data pairs in SynFog. However, models trained on SynFog demonstrate superior generalization capabilities on real-world data compared to them, highlighting the crucial role of data quality for researchers." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparison with SynFog- \\(\\beta\\) synthesized by ASM. With the ability to obtain pixel-accurate depth information for each image, we can eliminate the impact of inaccurate" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21768" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.264, + 0.089, + 0.706, + 0.104 + ], + "angle": 0, + "content": "Table 2. Defogging evaluations with AECRNet on real-world fog datasets." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.108, + 0.875, + 0.189 + ], + "angle": 0, + "content": "
Training SetO-Haze [2]Foggy Zurich [11]Foggy Driving [40]BeDDE [56]
PSNR ↑SSIM ↑DHQI [14] ↑DHQI [14] ↑DHQI [14] ↑DHQI [14] ↑
Foggy Cityscapes14.460.573743.4052.0651.5536.07
Virtual KITTI13.900.531542.8050.9447.4633.42
SynFog15.430.611644.4654.1652.0743.28
" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.197, + 0.29, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.283, + 0.207, + 0.293 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.197, + 0.482, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.282, + 0.411, + 0.293 + ], + "angle": 0, + "content": "DCP [20]" + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.197, + 0.675, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.282, + 0.604, + 0.294 + ], + "angle": 0, + "content": "CAP [57]" + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.197, + 0.869, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.282, + 0.822, + 0.293 + ], + "angle": 0, + "content": "GridDehazeNet [28]" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.293, + 0.29, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.378, + 0.23, + 0.39 + ], + "angle": 0, + "content": "MSBDN [12]" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.293, + 0.482, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.355, + 0.378, + 0.422, + 0.39 + ], + "angle": 0, + "content": "FFA-Net [39]" + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.293, + 0.675, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.378, + 0.619, + 0.39 + ], + "angle": 0, + "content": "AECR-Net [50]" + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.293, + 0.869, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.378, + 0.819, + 0.39 + ], + "angle": 0, + "content": "DehazeFormer [43]" + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.399, + 0.774, + 0.413 + ], + "angle": 0, + "content": "Figure 7. Transferability across real-world fog datasets. Test results of models trained on SynFog." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.421, + 0.472, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.474, + 0.411, + 0.487 + ], + "angle": 0, + "content": "Figure 8. Ablation of sensor noise simulation." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.497, + 0.469, + 0.525 + ], + "angle": 0, + "content": "Table 3. Detection evaluations on real-world fog datasets. \"FZ\" refers to Foggy Zurich. \"STF\" represents Seeing Through Fog." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.53, + 0.468, + 0.658 + ], + "angle": 0, + "content": "
MethodTraining SetFZ [11]STF [5]
mAP (%)mAP (%)
AECRNetFoggy Cityscapes69.754.8
Virtual KITTI68.953.3
SynFog71.555.5
DehazeFormerFoggy Cityscapes67.954.9
Virtual KITTI68.553.1
SynFog-β59.755.3
SynFog69.755.3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.47, + 0.837 + ], + "angle": 0, + "content": "depth and focus solely on the fog simulation method itself. The foggy image is synthesized from its clear counterpart in SynFog according to Eq. (1). Specifically, we employ the same airlight estimation method used in the construction of Foggy Cityscapes [40], which involves computing the median of all the \\(0.1\\%\\) pixels with the largest dark channel values, as proposed in [45]. We use DehazeFormer [43] to train on both SynFog and SynFog- \\(\\beta\\) , followed by testing the trained model on real-world foggy images. As shown in Fig. 9, image output by model trained on SynFog exhibits a better visual appearance compared to that of SynFog- \\(\\beta\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.353, + 0.863 + ], + "angle": 0, + "content": "5.3. Ablation Study of Sensor Noise" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To demonstrate the importance of sensor noise simulation in our pipeline, we create SynFog- \\(\\alpha\\) dataset. During the" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.422, + 0.871, + 0.437 + ], + "angle": 0, + "content": "Table 4. Impact of sensor noise on defogging performance." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.44, + 0.906, + 0.501 + ], + "angle": 0, + "content": "
Experimental settingO-Haze [2]
PSNR ↑SSIM ↑DHQI ↑
AECRNet+SynFog(w/o noise)15.160.579537.53
AECRNet+SynFog(w/ noise)15.430.611644.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.51, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Table 5. Impact of sensor noise on detection performance. \"FZ\" refers to Foggy Zurich. \"STF\" represents Seeing Through Fog." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.542, + 0.875, + 0.603 + ], + "angle": 0, + "content": "
Experimental settingFZ [11]STF [5]
mAP (%)mAP (%)
AECRNet+SynFog(w/o noise)69.554.7
AECRNet+SynFog(w/ noise)71.555.5
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.733 + ], + "angle": 0, + "content": "data generation process, the sensor-induced noise is set to zero, while all other settings remain consistent with those of the original SynFog dataset. We compare the defogging and detection performance of models trained on SynFog and SynFog- \\(\\alpha\\) datasets, as shown in Fig. 8 and Tab. 4-5. The model trained on SynFog that incorporates sensor noise produces more natural results with fewer artifacts and higher detection accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.736, + 0.757 + ], + "angle": 0, + "content": "5.4. Algorithm Benchmarking" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.87 + ], + "angle": 0, + "content": "We use SynFog test set to assess the aforementioned representative defogging methods, employing both full-reference and no-reference metrics to evaluate the defogged results. To illustrate the challenges presented by SynFog, we also provide evaluation results from the widely-used RESIDE dataset [25]. The full-scale quantitative comparison results are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Generally, learning-based methods consistently outperform earlier algorithms based on natural or statistical priors" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21769" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.247, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.23, + 0.18, + 0.24 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.089, + 0.406, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.231, + 0.378, + 0.242 + ], + "angle": 0, + "content": "Foggy Cityscapes [40]" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.09, + 0.565, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.231, + 0.529, + 0.242 + ], + "angle": 0, + "content": "Virtual KITTI [18]" + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.09, + 0.723, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.231, + 0.668, + 0.242 + ], + "angle": 0, + "content": "SynFog-β" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.09, + 0.883, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.231, + 0.838, + 0.242 + ], + "angle": 0, + "content": "SynFog (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.892, + 0.286 + ], + "angle": 0, + "content": "Figure 9. Visual and detection comparisons on real-world foggy images using DehazeFormer models trained on SynFog, SynFog- \\(\\beta\\), Foggy Cityscapes [40] and Virtual KITTI [18]." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.296, + 0.177, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.297, + 0.265, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.297, + 0.352, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.297, + 0.441, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.297, + 0.528, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.297, + 0.617, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.297, + 0.706, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.297, + 0.794, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.297, + 0.883, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.483, + 0.88, + 0.498 + ], + "angle": 0, + "content": "Figure 10. Qualitative comparisons of SOTA defogging methods on SynFog dataset with different fog levels. Zoom in for better view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.473, + 0.765 + ], + "angle": 0, + "content": "in terms of PSNR, SSIM and LPIPS [54]. In particular, DehazeFormer [43] attains the highest PSNR (27.96) and SSIM (0.923) values, while GridDehazeNet [28] achieves the best LPIPS (0.163) value on SynFog test set. Although the overall performance trend of the algorithms on SynFog is similar to that on RESIDE [25], the evaluation scores are noticeably lower, highlighting the challenges posed by SynFog. Fig. 10 depicts the qualitative comparison among these methods across different fog levels. Images restored by DCP [20] and CAP [57] tend to appear excessively dark and noisy due to inaccurate airlight estimation and the lack of effective denoising capabilities. GridDehazeNet [28] lacks color recovery ability in dense fog, while the outputs of FFA-Net [39] and DehazeFormer [43] seem smooth. This comparison further indicates the substantial room for improvement in existing defogging methods, especially when applied to more realistic fog datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.322, + 0.787 + ], + "angle": 0, + "content": "5.5. Discussion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Experimental results reveal that models trained on SynFog, incorporating realistic physical characteristics of optics and sensors, exhibit a more robust performance on real-word fog datasets in terms of visual perception and detection accuracy. This outcome implies that integrating a physically-based end-to-end imaging process replicating real capture methods can significantly improve the generalization of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.598 + ], + "angle": 0, + "content": "trained models. However, given the time-consuming nature of the physically-based volumetric path tracing technique, the scale of SynFog is not as extensive as other synthetic fog datasets. Nonetheless, we are committed to expanding the scene contents in the future and exploring a broader range of scattering mediums beyond fog." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.619, + 0.621 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.892, + 0.783 + ], + "angle": 0, + "content": "In this paper, we present an end-to-end foggy image simulation pipeline. Our approach utilizes volumetric path tracing to model a more precise light scattering process with global illumination. By incorporating a physically-based camera processing pipeline that includes optics, sensor and image processing, we can closely mimic the authentic capture process under foggy conditions. Additionally, we develop a new synthetic fog dataset, SynFog, to facilitate the research on defogging. Comprehensive experiments have validated the authenticity and reliability of the SynFog dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Acknowledgements This work is supported by 1) National Natural Science Foundation of China (Grant No. 62131011 and 62375233); 2) NSFC Young Scientists Fund (Grant No. 62302423). We thank Professor Chuxi Yang and Jiayue Xie for their help in preparing the scene creation. We also thank Professor Brian Wandell, Dr. Devesh Upadhyay, Dr. Alireza Rahimpour, and Devesh Shah for their valuable advice at the early stage of this project." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21770" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.119, + 0.176, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.47, + 0.2 + ], + "angle": 0, + "content": "[1] Cosmin Ancuti, Codruta O Ancuti, and Christophe De Vleeschouwer. D-hazy: A dataset to evaluate quantitatively dehazing algorithms. In IEEE International Conference on Image Processing, pages 2226-2230, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.27 + ], + "angle": 0, + "content": "[2] Codruta O Ancuti, Cosmin Ancuti, Radu Timofte, and Christophe De Vleeschouwer. O-haze: a dehazing benchmark with real hazy and haze-free outdoor images. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 754-762, 2018. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.47, + 0.327 + ], + "angle": 0, + "content": "[3] Codruta O Ancuti, Cosmin Ancuti, Mateu Sbert, and Radu Timofte. Dense-haze: A benchmark for image dehazing with dense-haze and haze-free images. In IEEE International Conference on Image Processing, pages 1014-1018, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.328, + 0.47, + 0.396 + ], + "angle": 0, + "content": "[4] Codruta O Ancuti, Cosmin Ancuti, and Radu Timofte. Nh-haze: An image dehazing benchmark with nonhomogeneous hazy and haze-free images. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 444-445, 2020. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.398, + 0.47, + 0.479 + ], + "angle": 0, + "content": "[5] Mario Bijelic, Tobias Gruber, Fahim Mannan, Florian Kraus, Werner Ritter, Klaus Dietmayer, and Felix Heide. Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11682-11692, 2020. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.47, + 0.536 + ], + "angle": 0, + "content": "[6] Norbert Buch, Sergio A Velastin, and James Orwell. A review of computer vision techniques for the analysis of urban traffic. IEEE Transactions on Intelligent Transportation Systems, 12(3):920-939, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[7] Ting Chen. Digital camera system simulator and applications. Stanford University, 2003. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[8] Lark Kwon Choi, Jaehee You, and Alan Conrad Bovik. Referenceless prediction of perceptual fog density and perceptual image defogging. IEEE Transactions on Image Processing, 24(11):3888-3901, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.622, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3213-3223, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.706, + 0.47, + 0.76 + ], + "angle": 0, + "content": "[10] Dengxin Dai and Wen Yang. Satellite image classification via two-layer sparse coding with biased image representation. IEEE Geoscience and Remote Sensing Letters, 8(1): 173-176, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.761, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[11] Dengxin Dai, Christos Sakaridis, Simon Hecker, and Luc Van Gool. Curriculum model adaptation with synthetic and real data for semantic foggy scene understanding. International Journal of Computer Vision, 128:1182-1204, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.831, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming-Hsuan Yang. Multi-scale boosted dehazing network with dense feature fusion. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2157-2167, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.145, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[13] J.B. Farrow et al. Influence of the atmosphere on optical systems. Optica Acta, 1970. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] X. Min et al. Objective quality evaluation of dehazed images. TITS, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[15] Joyce E Farrell and Brian A Wandell. Image systems simulation. Handbook of Digital Imaging, 1:373-400, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[16] Joyce E Farrell, Feng Xiao, Peter Bert Catrysse, and Brian A Wandell. A simulation tool for evaluating digital camera image quality. In Image Quality and System Performance, pages 124-131. SPIE, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[17] Joyce E Farrell, Peter B Catrysse, and Brian A Wandell. Digital camera simulation. Applied Optics, 51(4):A80-A90, 2012. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[18] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4340-4349, 2016. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[19] Martin Hahner, Dengxin Dai, Christos Sakaridis, Jan-Nico Zaech, and Luc Van Gool. Semantic understanding of foggy scenes with purely synthetic data. In IEEE Intelligent Transportation Systems Conference (ITSC), pages 3675-3681, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[20] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(12):2341-2353, 2010. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[21] Louis G Henyey and Jesse Leonard Greenstein. Diffuse radiation in the galaxy. Astrophysical Journal, vol. 93, p. 70-83 (1941), 93:70-83, 1941. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[22] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolo, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[23] Kyung Won Kim. The comparison of visibility measurement between image-based visual range, human eye-based visual range, and meteorological optical range. Atmospheric Environment, 190:74-86, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.58, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[24] Miles V Klein and Thomas E Furtak. Optik. Springer-Verlag, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[25] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2018. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.706 + ], + "angle": 0, + "content": "[26] Ruoteng Li, Xiaoyi Zhang, Shaodi You, and Yu Li. Realistic large-scale fine-depth dehazing dataset from 3d videos. arXiv preprint arXiv:2004.08554, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[27] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(10):2024-2039, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[28] Xiaohong Liu, Yongrui Ma, Zhihao Shi, and Jun Chen. Griddehazenet: Attention-based multi-scale network for image dehazing. In IEEE International Conference on Computer Vision, pages 7314-7323, 2019. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[29] Yun Liu, Zhongsheng Yan, Aimin Wu, Tian Ye, and Yuche Li. Nighttime image dehazing based on variational decomposition model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 640-649, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21771" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[30] Zhenyi Liu, Trisha Lian, Joyce Farrell, and Brian Wandell. Soft prototyping camera designs for car detection based on a convolutional neural network. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[31] Zhenyi Liu, Minghao Shen, Jiaqi Zhang, Shuangting Liu, Henryk Blasinski, Trisha Lian, and Brian Wandell. A system for generating complex physically accurate sensor images for automotive applications. arXiv preprint arXiv:1902.04258, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.472, + 0.288 + ], + "angle": 0, + "content": "[32] Zheng Lyu, Thomas Goossens, Brian A Wandell, and Joyce Farrell. Validation of physics-based image systems simulation with 3-d scenes. IEEE Sensors Journal, 22(20):19400-19410, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.29, + 0.472, + 0.318 + ], + "angle": 0, + "content": "[33] Earl J McCartney. Optics of the atmosphere: scattering by molecules and particles. New York, 1976. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.472, + 0.36 + ], + "angle": 0, + "content": "[34] Srinivasa G Narasimhan and Shree K Nayar. Vision and the atmosphere. International Journal of Computer Vision, 48: 233-254, 2002. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.472, + 0.445 + ], + "angle": 0, + "content": "[35] Priya Narayanan, Xin Hu, Zhenyu Wu, Matthew D Thielke, John G Rogers, Andre V Harrison, John A D'Agostino, James D Brown, Long P Quang, James R Uplinger, et al. A multi-purpose realistic haze benchmark with quantifiable haze levels and ground truth. IEEE Transactions on Image Processing, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.472, + 0.474 + ], + "angle": 0, + "content": "[36] National Standards of People's Republic of China. Grade of fog forecast, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.472, + 0.517 + ], + "angle": 0, + "content": "[37] Roberto Nebuloni. Empirical relationships between extinction coefficient and visibility in fog. Applied Optics, 44(18): 3795-3804, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.472, + 0.56 + ], + "angle": 0, + "content": "[38] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically based rendering: From theory to implementation. MIT Press, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.472, + 0.616 + ], + "angle": 0, + "content": "[39] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In AAAI, pages 11908-11915, 2020. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.472, + 0.673 + ], + "angle": 0, + "content": "[40] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Semantic foggy scene understanding with synthetic data. International Journal of Computer Vision, 126:973-992, 2018. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.472, + 0.731 + ], + "angle": 0, + "content": "[41] Daniel Scharstein and Richard Szeliski. High-accuracy stereo depth maps using structured light. In IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages I-I. IEEE, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.472, + 0.788 + ], + "angle": 0, + "content": "[42] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In European Conference on Computer Vision, pages 746-760. Springer, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.472, + 0.83 + ], + "angle": 0, + "content": "[43] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927-1941, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[44] Tao Sun, Mattia Segu, Janis Postels, Yuxuan Wang, Luc Van Gool, Bernt Schiele, Federico Tombari, and Fisher Yu. Shift: a synthetic driving dataset for continuous multi-task domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 21371-21382, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[45] Ketan Tang, Jianchao Yang, and Jue Wang. Investigating haze-relevant features in a learning framework for image dehazing. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2995-3000, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[46] Jean-Philippe Tarel, Nicolas Hautiere, Aurélien Cord, Dominique Gruyer, and Houssam Halmaoui. Improved visibility of road scene images under heterogeneous fog. In IEEE Intelligent Vehicles Symposium, pages 478–485, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[47] Jean-Philippe Tarel, Nicholas Hautiere, Laurent Caraffa, Aurélien Cord, Houssam Halmaoui, and Dominique Gruyer. Vision enhancement in homogeneous and heterogeneous fog. IEEE Intelligent Transportation Systems Magazine, 4 (2):6-20, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[48] Tali Treibitz and Yoav Y Schechner. Resolution loss without imaging blur. JOSA A, 29(8):1516-1528, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[49] Apostolia Tsirikoglou, Joel Kronander, Magnus Wrenninger, and Jonas Unger. Procedural modeling and physically based rendering for synthetic data generation in automotive applications. arXiv preprint arXiv:1710.06270, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[50] Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Contrastive learning for compact single image dehazing. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10551-10560, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[51] Jing Zhang, Yang Cao, Zheng-Jun Zha, and Dacheng Tao. Nighttime dehazing with a synthetic benchmark. In Proceedings of the ACM International Conference on Multimedia, pages 2355-2363, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[52] Lin Zhang, Junhao Huang, Xiyuan Li, and Lu Xiong. Vision-based parking-slot detection: A dcnn-based approach and a large-scale benchmark dataset. IEEE Transactions on Image Processing, 27(11):5350-5364, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[53] Lin Zhang, Anqi Zhu, Shiyu Zhao, and Yicong Zhou. Simulation of atmospheric visibility impairment. IEEE Transactions on Image Processing, 30:8713-8726, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.656 + ], + "angle": 0, + "content": "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.659, + 0.892, + 0.714 + ], + "angle": 0, + "content": "[55] Yanfu Zhang, Li Ding, and Gaurav Sharma. Hazerd: an outdoor scene dataset and benchmark for single image dehazing. In IEEE International Conference on Image Processing, pages 3205-3209, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[56] Shiyu Zhao, Lin Zhang, Shuaiyi Huang, Ying Shen, and Shengjie Zhao. Dehazing evaluation: Real-world benchmark datasets, criteria, and baselines. IEEE Transactions on Image Processing, 29:6947-6962, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[57] Qingsong Zhu, Jiaming Mai, and Ling Shao. A fast single image haze removal algorithm using color attenuation prior. IEEE Transactions on Image Processing, 24(11):3522-3533, 2015. 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.827 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21772" + } + ] +] \ No newline at end of file diff --git a/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/30117727-713c-4f8f-a215-de682288f1dd_origin.pdf b/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/30117727-713c-4f8f-a215-de682288f1dd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fa1f26d953d58f41d6ed25a25dffa3f302a005d7 --- /dev/null +++ b/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/30117727-713c-4f8f-a215-de682288f1dd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2125ef3ef15b79b1fb5be9ac5828df79b500020cd2b7f5677aec28c0cf6cca50 +size 5673647 diff --git a/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/full.md b/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5d6fd44a60e1330be107d88100861cef5d28efa0 --- /dev/null +++ b/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/full.md @@ -0,0 +1,369 @@ +# SynFog: A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving + +Yiming Xie $^{1*}$ , Henglu Wei $^{1*}$ , Zhenyi Liu $^{2}$ , Xiaoyu Wang $^{1}$ , Xiangyang Ji $^{1}$ $^{1}$ Tsinghua University, $^{2}$ Stanford University +{xieym22@mails., weihenglu@mail., xiaoyu-w17@mails., xyji@}tsinghua.edu.cn, {zhenyiliu@}stanford.edu + +![](images/8e1ff6fee21aab43dc8c8bc95515cc43b9b31ed14ba4d64a7dd2e191067e2632.jpg) +Figure 1. The SynFog dataset comprises images under different lighting, with the first row captured in natural sky light and the second with extra light sources like street lamps and automotive lighting. This dataset covers three levels of fog density across each scene, corresponding visibility of $600\mathrm{m}$ , $300\mathrm{m}$ , $150\mathrm{m}$ . Additionally, pixel-accurate depth data and segmentation labels for each scene are also provided. + +# Abstract + +To advance research in learning-based defogging algorithms, various synthetic fog datasets have been developed. However, existing datasets created using the Atmospheric Scattering Model (ASM) or real-time rendering engines often struggle to produce photo-realistic foggy images that accurately mimic the actual imaging process. This limitation hinders the effective generalization of models from synthetic to real data. In this paper, we introduce an end-to-end simulation pipeline designed to generate photorealistic foggy images. This pipeline comprehensively considers the entire physically-based foggy scene imaging process, closely aligning with real-world image capture methods. Based on this pipeline, we present a new synthetic fog dataset named SynFog, which features both sky light and active lighting conditions, as well as three levels of fog density. Experimental results demonstrate that models trained on SynFog exhibit superior performance in visual perception and detection accuracy compared to others when applied to real-world foggy images. + +# 1. Introduction + +Fog, due to its light scattering and attenuation properties, poses a significant challenge for outdoor vision-based sys + +tems such as remote sensing, surveillance, autonomous driving, and fire rescue [6, 10, 52]. It causes severe degradation in object appearance and contrast, resulting in a notable impact on people's daily lives. Consequently, there has been a growing research interest in image defogging task to develop robust outdoor vision systems. + +Recent learning-based defogging algorithms depend on sets of foggy images and corresponding clear images for model training. While several datasets have been collected from real-world scenarios [5, 40], they are typically limited to foggy images only, as it is almost impossible to obtain pixel-accurate clear images with consistent scene contents and environmental illumination. Studies in [2-4, 35] attempt to capture real-world foggy and clear image pairs using smoke machines. However, the scale of these datasets is inadequate for training of a high-performing model. Given the challenge of obtaining paired foggy and clear images in real world, various synthetic fog datasets have been proposed as an alternative [25, 40, 55]. The most widely used method is based on the theory proposed by McCartney [33], also known as the Atmospheric Scattering Model (ASM). However, this model fails to consider accurate global illumination and the actual imaging process that occurs in real-world captures as detailed in Sec. 3.1. These limitations significantly impact the model's overall applicability, par + +![](images/67fb9f7b4c55f8898211c4c1c3a14228300936503c99f3768c41d39632ef0019.jpg) +Figure 2. Relationship between noise level and fog density in real-world foggy images. Using regional contrast stretching for image enhancement, it can be observed that the noise level increases with the concentration of fog. + +ticularly in nighttime scenarios, and can lead to a disparity between synthetic and real-world foggy images. As a result, models trained on synthetic foggy images may exhibit limited robustness when applied to real-world ones. + +To address these issues, we propose an end-to-end foggy image simulation pipeline that encompasses the entire imaging process starting from 3D virtual scenes to final photorealistic foggy images. Our approach involves foggy scene rendering using volumetric path tracing and imaging through a realistic camera model. As shown in Fig. 3b, fog is treated as a spatial volume, and the spectral scene radiance is rendered using volumetric path tracing [38]. By defining the phase function of fog medium and bidirectional reflectance distribution function (BRDF) of the scene components, we are able to accurately describe the scattering process with multiple light sources under foggy conditions. The foggy scene radiance data is further processed through a physically-based camera model, which comprises optics, sensor and image processing, to faithfully replicate real camera devices. Based on the proposed pipeline, we develop a new extensive synthetic fog dataset called SynFog. This dataset contains 500 unique outdoor scenes, each featuring two kinds of lighting conditions and three levels of fog density. Extensive experiments have been conducted to qualitatively and quantitatively validate the authenticity of our simulation pipeline and the SynFog dataset. + +Our main contributions are summarized as follows: + +- We propose a physically-based end-to-end foggy image simulation pipeline that incorporates accurate light transportation in scattering medium and physical characteristics of optics and sensor into the synthesized images. +- We develop a photo-realistic synthetic fog dataset, SynFog, which includes both sky light and active lighting conditions, as well as three levels of fog density. +- We demonstrate the authenticity and effectiveness of our simulation pipeline and the superior realism of SynFog dataset compared to other synthetic fog datasets. + +# 2. Related Work + +Given the challenge of acquiring paired foggy and clear images in the real world for training learning-based defogging methods, researchers have proposed many synthetic fog datasets as an alternative. In this section, we will offer a thorough review and analysis of foggy image simulation techniques and their associated datasets. + +Based on Atmospheric Scattering Model (ASM). This method involves rendering fog onto clean images along with their depth information [1, 19, 25, 26, 40, 53, 55]. Depth information can be acquired through game engines, depth cameras or depth estimation methods. In an earlier study conducted by Tarel et al. [46], FRIDA was constructed using a total of 90 synthetic images, with clear images and depth maps generated using $SiVIC^{TM}$ software. Subsequently, in their later work [47], more synthetic images were added, leading to an upgraded version named FRIDA2, which includes 330 synthetic images. However, the scale of these datasets is insufficient for modern data-driven defogging methods, and the foggy images appear unrealistic. In more recent works [26, 53], a similar approach has been employed to acquire both the clear images and their corresponding depth maps from game engines. In [25], a general dehazing dataset called RESIDE was introduced. The indoor training set (ITS) of RESIDE was constructed based on existing indoor RGBD datasets NYUv2 [42] and Middleburry stereo [41]. For the outdoor training set (OTS), 2071 real-world outdoor images were collected, and the corresponding depth information was estimated using a learning-based technique [27]. It is important to note that the accuracy of depth estimation based on learning methods is still unsatisfactory, as single-image-based depth estimation is inherently an ill-posed problem. Furthermore, Foggy Cityscapes was introduced in [40], utilizing images from the Cityscapes dataset [9]. Despite efforts to enhance the original depth maps in Cityscapes, artifacts and discrepancies still persist, resulting in synthetic foggy images that appear unnatural. + +Based on Real-time Rendering Engines. Foggy images generated utilizing real-time rendering engines like Unity and Unreal have been discussed in works such as [18, 44]. Among them, Virtual KITTI 2 [18] stands out as a notable example. However, these foggy images often lack realism and exhibit limited diversity in terms of fog density. + +Based on Physical Means. Fog/haze machines are employed in real-world settings to generate artificial fog, as discussed in works such as [2-4, 35]. However, these datasets are generally limited in scale and may not offer sufficient supervision for learning-based defogging methods. Moreover, ensuring consistent content between the clear and foggy image captures can be challenging. + +There are also several real-world fog datasets available, such as Foggy Driving [40], Foggy Zurich [11], and Seeing + +![](images/09614b1fb99b25afb9a90a4e90e07fad931f0d2f43194ac75c63869194cddd1a.jpg) +(a) Atmospheric Scattering Model (ASM) [34]. + +![](images/d5eb05dde2831ddd5af86a71a9db1f88f3a3f7617d3e89d78f062ff002a2d09e.jpg) +(b) Foggy scene rendering based on volumetric path tracing. +Figure 3. Foggy scene simulation methods. + +Through Fog [5]. However, due to the absence of corresponding clear reference images, these datasets can only be used for testing purposes rather than for training. In this paper, we propose a physically-based end-to-end foggy image simulation pipeline to improve the authenticity of synthetic foggy images and construct a photo-realistic fog dataset to facilitate future research on defogging algorithms. + +# 3. End-to-end Foggy Image Simulation + +We propose an end-to-end approach for generating photorealistic foggy images that incorporates accurate light transportation in scattering medium and physical characteristics of optics and sensor into the synthesized images, as illustrated in Fig. 4. We first introduce a foggy scene imaging model in Sec. 3.1. Based on this model, we establish a simulation pipeline consisting of two components: a) Foggy scene radiance is rendered using volumetric path tracing [38]. b) The radiance data is processed through a physically-based camera model, which comprises optics, sensor and image processing to faithfully replicate real camera devices [17]. These components are further detailed in Sec. 3.2.1 and Sec. 3.2.2, respectively. + +# 3.1. Foggy Scene Imaging Model + +The Atmospheric Scattering Model (ASM) [33, 34] has been widely used for the generation of foggy images [19, + +25,40,53]: + +$$ +E (d, \lambda) = \underbrace {E _ {o} (\lambda) e ^ {- \beta (\lambda) d}} _ {\text {t r a n s m i s s i o n}} + \underbrace {L _ {h} (\infty , \lambda) \left(1 - e ^ {- \beta (\lambda) d}\right)} _ {\text {a i r l i g h t}}, \tag {1} +$$ + +where $E$ is the irradiance reaching the film under foggy conditions, $E_{o}$ is the irradiance under clear conditions, $\beta$ is the scattering coefficient characterizing the scatter ability of the medium, $d$ is the distance between the object and the camera, $L_{h}(\infty ,\lambda)$ denotes the radiance of airlight at infinity and $\lambda$ refers to wavelength. Fig. 3a provides a more intuitive description of this process. Foggy images can be generated with pre-defined $\beta$ and $L_{h}$ values by directly using the clear images as $E_0$ and the corresponding depth maps as $d$ . + +However, there are two issues with this scattering model for generating foggy images: (1) The definition of environmental illumination fails to consider the contribution of indirect illumination from the scene to the scattering medium, resulting in an unfaithful representation of the actual illumination. (2) The current formulation of airlight is insufficient for accurately modeling the lighting conditions that involve multiple point light sources, such as street lamps and automotive lights. This limitation restricts the model's applicability in situations like nighttime [29, 51]. + +Furthermore, ASM only captures the irradiance reaching the sensor plane. Consequently, foggy images generated using ASM fail to accurately incorporate the realistic physical characteristics of optics and sensor in actual imaging process. In Fig. 2, we depict the relationship between noise level and fog density after the enhancement of regional contrast stretching [48] on a real captured image [8]. It can be seen that the noise level increases with higher fog density, emphasizing the critical need for accurate noise modeling in foggy image simulation. + +To enhance the consistency between synthetic images and real-world captures, we develop a comprehensive foggy scene imaging model that incorporates the entire imaging process, referred to as: + +$$ +I _ {o u t} = f _ {I S P} \left(f _ {\text {s e n s o r}} \left(f _ {\text {o p t i c s}} \left(L _ {\text {s c e n e}}\right)\right) + n\right), \tag {2} +$$ + +where $L_{\text{scene}}$ is the scene radiance obtained using volumetric path tracing [38], as described in Sec. 3.2.1; $\{f_{\text{optics}}, f_{\text{sensor}}, f_{\text{ISP}}\}$ denote the optics, sensor and ISP processes respectively, as described in Sec. 3.2.2, and the noise term introduced by the sensor is denoted as $n$ . + +By employing volumetric path tracing to obtain the radiance of foggy scene, we can precisely model the light scattering process under global illumination and active lighting. By incorporating the complete imaging process described in Eq. (2), we are able to faithfully replicate the authentic camera characteristics in foggy image simulation. + +![](images/367edc5b7d3434f72977b0fa071f665ea871022e136cb73b9d7fa4a9bb9a9005.jpg) +Figure 4. End-to-end foggy image simulation pipeline. The spectral radiance data is rendered using volumetric path tracing and passing through a realistic optics model before reaching the sensor plane. Subsequently, the irradiance is converted into a digital image through the sensor model. The raw image from the sensor is then processed by an ISP module to generate the final foggy image. + +![](images/5ec3364dd38c1088a85fc476b4df2f784def9bf0ec4bb3340002c6aafce770cd.jpg) + +![](images/4bc42c76256e1ae821b6a302cb41af5c096e23803feae3c13329c018a3a23a79.jpg) + +![](images/ef7d7da76639afea1f4855a58d7e8e835c60e548f57ae5c705c1b11b202f0798.jpg) + +# 3.2. End-to-end Simulation Pipeline + +# 3.2.1 Foggy Scene Rendering + +We utilize procedural modeling techniques, similar to the approach employed in [31, 49], to create large-scale and diverse driving scenes. This involves defining adjustable parameters and using a set of predefined rules to generate explicit scene definitions. Our model library incorporates a diverse range of pedestrians, bicyclists and cars, each with well-defined material and texture properties. Furthermore, we carefully select 16 sky maps with cloudy conditions to ensure realistic lighting for foggy scenes, following the established guidelines from previous works [2-4]. + +In our approach, fog is treated as a spatial volume, and the spectral scene radiance is rendered using volumetric path tracing [38]. To accurately model the scattering process of the fog medium, we employ Henyey and Greenstein phase function [21, 38] to model its particle scattering characteristics, denoted as: + +$$ +p _ {\mathrm {H G}} (\cos \theta) = \frac {1}{4 \pi} \frac {1 - g ^ {2}}{\left(1 + g ^ {2} + 2 g (\cos \theta)\right) ^ {3 / 2}}. \tag {3} +$$ + +By empirically setting the asymmetry parameter $g = 0.87$ , we can accurately simulate the probability distributions for scattering in a specific direction at a given point within the fog medium, as shown in Fig. 3b. + +The rendering equation consists of two parts that account for reduction of radiance caused by out-scattering and increase of radiance due to in-scattering at point $p$ in direction $\omega$ , represented by Eq. (4) and Eq. (5), respectively [38]. + +$$ +L _ {\mathrm {i}} \left(\mathrm {p} ^ {\prime}, - \omega\right) = \mathrm {e} ^ {- \int_ {0} ^ {d} \sigma_ {\mathrm {s}} (\mathrm {p} + t \omega , \omega) \mathrm {d} t} L _ {\mathrm {o}} (\mathrm {p}, \omega). \tag {4} +$$ + +$$ +L _ {\mathrm {s}} (\mathrm {p}, \omega) = L _ {\mathrm {e}} (\mathrm {p}, \omega) + \sigma_ {\mathrm {s}} (\mathrm {p}, \omega) \int_ {\mathcal {S} ^ {2}} p (\mathrm {p}, \omega_ {\mathrm {i}}, \omega) L _ {\mathrm {i}} (\mathrm {p}, \omega_ {\mathrm {i}}) \mathrm {d} \omega_ {\mathrm {i}}. \tag {5} +$$ + +$\sigma_{s}$ represents the scattering probability per unit distance, $\{L_i,L_o,L_e,L_s\}$ are the {incident,exitant,emission,total added} radiance. In our implementation, we employ the Monte Carlo algorithm to solve the above integration. + +The key distinction between our physically-based foggy scene rendering technique and ASM lies in the modeling of airlight, as represented by Eq. (5). Our approach offers a + +![](images/0bb284ef9b1207033f17d9117b63a43bdbfc1743e83004c8e544cfd5880e2e90.jpg) +Figure 5. Noise flow and sensor simulation. + +more accurate depiction of light scattering process through the utilization of specific phase function, and enables incorporation of global illumination and multiple light sources within the scene, producing authentic scene radiance to the camera model described in the next section. + +# 3.2.2 Imaging through Realistic Camera Simulation + +After obtaining the scene radiance, we employ ISETCam [15-17, 30, 32] to simulate the realistic digital camera system that encompasses optics, sensor and ISP components. Specifically, we use a diffraction-limited optics model [7] with consideration of off-axis irradiance attenuation based on the cosine-fourth law [24] to convert scene radiance $L_{o}$ to optical image irradiance $E_{i}$ , denoted as: + +$$ +E _ {i} = \pi \cdot L _ {o} \frac {1}{1 + 4 (f / \# (1 - m)) ^ {2}} (\cos \phi) ^ {4}, \tag {6} +$$ + +where $f / \#$ is the f-number, $m$ represents the magnification defined as the negative ratio of image distance and object distance. In this optics model, blur mainly depends on the f-number value but little else. + +We model the sensor process including photovoltaic conversion, analog amplifier, and analog-to-digital conversion as described in Fig. 4 and simulate the real noise characteristics depicted in Fig. 5. Among the various noise sources, shot noise arising from the photovoltaic conversion emerges as the dominant type of noise in foggy images which follows the Poisson distribution. To ensure the credibility of our sensor simulation, we incorporate a set of calibrated sensor parameters from the Mi10Pro cellphone, encompassing the CFA and infrared filter data. + +Finally, we process the RAW image output from the sensor with a simple ISP module, including black level compensation, white balance, demosaicing, color correction and + +Table 1. Comparisons between SynFog and existing synthetic fog datasets in the driving field. + +
DatasetsGlobal IlluminationOptics and Sensor Characteristics
Foggy Cityscapes××
Virtual KITTI 2××
SynFog (Ours)✓✓
+ +gamma correction, to get the final synthetic image that mimics the real camera characteristics. + +# 4. SynFog Dataset + +By utilizing the proposed pipeline, we develop a new extensive synthetic fog dataset called SynFog. Example images are shown in Fig. 1. We first generate 500 large-scale, diverse driving scenes using procedural modeling techniques. For each scene, we generate two sets of images under different lighting conditions: one set with sky light only and the other set with both sky light and active light sources, such as street lamps and automotive lighting. Within each group, we create a set of four images: one clear image and three foggy images with varying fog levels defined by the scattering coefficient $\sigma_{\mathrm{s}}$ . According to McCartney's work [33], the scattering coefficient remains independent of the wavelength for fog [13, 34]. Therefore, we use a consistent scattering coefficient for wavelengths within the visible range. To ensure realistic simulation of fog, we employ the meteorological optical range (MOR) [23] or visibility as a reference to select appropriate scattering coefficients, described as: + +$$ +M O R = 2. 9 9 6 / \sigma_ {s}. \tag {7} +$$ + +Based on this equation, we select scattering coefficient values of 0.005, 0.01, 0.02, which correspond to visibility range of $600\mathrm{m}$ , $300\mathrm{m}$ and $150\mathrm{m}$ , respectively. According to National Standard of China [36], these visibility ranges are classified as heavy fog, thick fog, and dense fog, thereby the simulated foggy images are diversified enough to reflect real-world complicated foggy conditions. + +We compare SynFog with existing fog datasets in the field of autonomous driving, including the extensively used Foggy Cityscapes [40] and Virtual KITTI 2 [18]. The comparison is presented in Tab. 1. In terms of foggy image generation, methods that rely on ASM [34] require accurate depth maps to calculate the correct transmission maps. However, obtaining accurate depth maps for outdoor images can be challenging, leading to artifacts such as holes and discontinuities in synthetic foggy images, similar to what is observed in Foggy Cityscapes [40]. On the other hand, foggy images generated using Game Engines like Unity or Unreal [18] typically employs rasterization techniques in the rendering process. While effective for creating + +scenes, rasterization techniques may not accurately model light transportation as precisely as path tracing, which is employed in SynFog. Moreover, SynFog takes the realistic imaging process into account, resulting in highly authentic foggy images with realistic physical characteristics. The superior realism of SynFog is further supported by the experimental results. + +Furthermore, our data simulation pipeline offers exceptional flexibility and control, allowing us to generate not only the final RGB images but also other valuable data modalities, including RAW images. This capability opens up avenues for research on defogging algorithms in the RAW domain, which has not been explored due to the lack of paired foggy RAW images and clear RGB images necessary for training deep defogging networks. + +# 5. Experiments + +# 5.1. Fog Chamber Validation + +We establish an indoor fog chamber to validate the fidelity of our fog simulation pipeline. The chamber is constructed as a capped glass cube with dimensions of $1.5m\times 0.5m\times$ $0.5m$ . To reproduce the dark night environment that has directional light sources in the fog chamber, we position 4 light bars on each side of the front face of the cuboid fog chamber and enclose it with black blackout cloth from the inside. Fog inside the chamber is generated using a 1.7MHz ceramic atomizer that emulates water mist. To ensure uniformity of fog concentration within the chamber, we incorporate dual fans to promote even distribution. A visibility measurement module is designed to quantify the fog density. This module is comprised of a PM-160 wireless power meter from Thorlabs, an 852 narrowband laser, an aperture and a narrowband filter adapted to the laser band. The aperture serves to filter out stray light and the narrowband filter plays the role in blocking interference caused by non-laser light sources in the scene. The principle of visibility (scattering coefficient $\sigma_{s}$ ) measurement is expressed as [37]: + +$$ +\sigma_ {s} = \frac {\ln \left(\frac {P _ {0}}{P _ {u}}\right)}{\mathrm {u}}, \tag {8} +$$ + +where $P_0$ is the laser power measured without fog and $P_u$ is the measured laser power after the propagation of $u$ distance under foggy conditions. + +We then use Blender to create a virtual fog chamber that can reproduce the real-world counterpart. Our process involves calibrating the light conditions, object information and camera properties prior to capture, resulting in a one-to-one reproduction of the real fog chamber. We select the standard Macbeth color checker as our test object and use Mi10Pro cellphone to capture images. We evaluate the matching degree between the simulation and reality un + +![](images/c81e9100ee41046608c476a51ec4d25e06c83552ef5149f70d7059835f36f1b4.jpg) +(a) + +![](images/36ff337738896c144f5d81762896a635365349fab6bfa047015b55d8e3a8ec4a.jpg) +Figure 6. Color analysis under foggy conditions. (a) Visual comparison between real captures and simulated images under different fog densities in linear RGB space. We increase the brightness of the images $(\times 2)$ for visual purpose. (b) Analyzing the trend of color variation across different fog densities. + +der clear conditions by conducting a color analysis on the checker, which is available in the supplementary material. + +After ensuring the consistency of initial capture conditions, we proceed to capture images under foggy conditions. We first record the laser power under clear condition as $P_0$ , and then proceed to generate fog until the object is no longer visible. During the fog dispersion process, we capture images at regular intervals of 1 second and record the laser energy $P_u$ to obtain real-time fog density measurements. We utilize the measured fog density parameter $\sigma_s$ to simulate corresponding foggy counterparts based on the proposed foggy image simulation pipeline. + +In Fig. 6a, we showcase both the real-captured foggy images and simulated ones generated using our pipeline and the ASM [33]. It can be seen that images simulated through our pipeline exhibit a closer resemblance to real-captured foggy images. We also conduct a color analysis on 24 color patches under different fog densities as depicted in Fig. 6b. The results demonstrate that, as fog density increases, the trend of each channel in simulated images generated by our pipeline consistently follows that of the real captures. In contrast, the simulated images based on ASM do not exhibit this trend. These findings provide a quantitative validation of the realism of our foggy image simulation method. + +# 5.2. Transferability across the Real-to-Virtual Gap + +We evaluate 7 representative state-of-the-art defogging algorithms on SynFog dataset: DCP [20], CAP [57], Grid-DehazeNet [28], MSBDN [12], FFA-Net [39], AECR-Net [50], and DehazeFormer [43]. We directly apply the models trained on SynFog to real foggy images sourced from Foggy Zurich [11] and Seeing Through Fog [5] datasets, without undergoing any model fine-tuning or domain adaptation. During the training process, we adhere to the official settings of these methods, and all of them demonstrate convergence by the end of the training. The results are shown in Fig. 7. Despite DCP [20] and CAP [57] producing excessively dark outputs, models trained on SynFog can produce natural defogged images with minimal artifacts. + +To underscore the superior realism of SynFog compared to existing synthetic fog datasets, we further enhance our evaluation by training the top-performing models (AECR-Net [50], DehazeFormer [43]) on two prominent synthetic fog datasets relevant to autonomous driving (Foggy Cityscapes [40], Virtual KITTI [18]). Subsequently, these models are evaluated on real fog datasets for a comparative analysis. Additionally, we utilize ASM [34] to create a SynFog- $\beta$ dataset with clear images from SynFog and test it using DehazeFormer [43]. + +Comparison with existing synthetic fog datasets. To mitigate any potential impact arising from model training issues, we initially conduct tests for each model using the corresponding test set of the training data. Test results are available in the supplementary material. Both AECR-Net [50] and DehazeFormer [43] exhibit high performance, confirming the correctness of the model training process. Subsequently, the trained models are directly applied to real-world test foggy images. Results are shown in Fig. 9 and Tab. 2. Owing to the high authenticity of the SynFog dataset, models trained on it can effectively generalize to real foggy images and produce naturally colored defogged images. Conversely, models trained on the other two datasets display poorer generalization to real data, resulting in artifacts and color distortion in defogged images. We also evaluate the defogged images using a pre-trained YOLOv8 [22] model for detection, showcasing how SynFog enhances machine vision capabilities (See Tab. 3). More results can be found in the supplementary material. + +It is noteworthy that the training set of Foggy Cityscapes [40] comprises 8925 data pairs, and Virtual KITTI [18] includes 3452 data pairs, both exceeding the 1350 data pairs in SynFog. However, models trained on SynFog demonstrate superior generalization capabilities on real-world data compared to them, highlighting the crucial role of data quality for researchers. + +Comparison with SynFog- $\beta$ synthesized by ASM. With the ability to obtain pixel-accurate depth information for each image, we can eliminate the impact of inaccurate + +Table 2. Defogging evaluations with AECRNet on real-world fog datasets. + +
Training SetO-Haze [2]Foggy Zurich [11]Foggy Driving [40]BeDDE [56]
PSNR ↑SSIM ↑DHQI [14] ↑DHQI [14] ↑DHQI [14] ↑DHQI [14] ↑
Foggy Cityscapes14.460.573743.4052.0651.5536.07
Virtual KITTI13.900.531542.8050.9447.4633.42
SynFog15.430.611644.4654.1652.0743.28
+ +![](images/b1f5812d7c12de482978a6d936c91211746e0aaf2e213207e94c07117207ccc4.jpg) + +![](images/126cb07242ae6d47cf3790a8c695d911a9a8fea06ed4c553736c30aab29ee76d.jpg) + +![](images/10e1190dce571374cd3edcdb719870ab82d265ec7c2d06a95ff0acbd8f2d05d1.jpg) + +![](images/f0756365c306c6245fecd1237ea401ea9b08a33a0a0d9c36e3539f915cc2c0c6.jpg) + +![](images/6db9bf03d46b68b982a95663e4d27900ac73d9cee60428f91afed4f2a037f6fa.jpg) +Input +MSBDN [12] + +![](images/4a55c76112416daaba3cc54d9256c52f43eac2eb55e2ccdbb63cd2a1c656758a.jpg) +DCP [20] +FFA-Net [39] + +![](images/897d10a502d991a9126b94861160f0e1e4bc21e7dfb3980962aca42d1af80fdb.jpg) +CAP [57] +AECR-Net [50] + +![](images/a9fd4e27f47dc4bb3b876ff7c1d8bdcb239cba1982ff2d99dc728e7ba7301ec2.jpg) +GridDehazeNet [28] +DehazeFormer [43] + +![](images/515484acf83bc484a7a14c0ca17a927119d75ae605e026bbce7219533cbde948.jpg) +Figure 7. Transferability across real-world fog datasets. Test results of models trained on SynFog. +Figure 8. Ablation of sensor noise simulation. + +Table 3. Detection evaluations on real-world fog datasets. "FZ" refers to Foggy Zurich. "STF" represents Seeing Through Fog. + +
MethodTraining SetFZ [11]STF [5]
mAP (%)mAP (%)
AECRNetFoggy Cityscapes69.754.8
Virtual KITTI68.953.3
SynFog71.555.5
DehazeFormerFoggy Cityscapes67.954.9
Virtual KITTI68.553.1
SynFog-β59.755.3
SynFog69.755.3
+ +depth and focus solely on the fog simulation method itself. The foggy image is synthesized from its clear counterpart in SynFog according to Eq. (1). Specifically, we employ the same airlight estimation method used in the construction of Foggy Cityscapes [40], which involves computing the median of all the $0.1\%$ pixels with the largest dark channel values, as proposed in [45]. We use DehazeFormer [43] to train on both SynFog and SynFog- $\beta$ , followed by testing the trained model on real-world foggy images. As shown in Fig. 9, image output by model trained on SynFog exhibits a better visual appearance compared to that of SynFog- $\beta$ + +# 5.3. Ablation Study of Sensor Noise + +To demonstrate the importance of sensor noise simulation in our pipeline, we create SynFog- $\alpha$ dataset. During the + +Table 4. Impact of sensor noise on defogging performance. + +
Experimental settingO-Haze [2]
PSNR ↑SSIM ↑DHQI ↑
AECRNet+SynFog(w/o noise)15.160.579537.53
AECRNet+SynFog(w/ noise)15.430.611644.46
+ +Table 5. Impact of sensor noise on detection performance. "FZ" refers to Foggy Zurich. "STF" represents Seeing Through Fog. + +
Experimental settingFZ [11]STF [5]
mAP (%)mAP (%)
AECRNet+SynFog(w/o noise)69.554.7
AECRNet+SynFog(w/ noise)71.555.5
+ +data generation process, the sensor-induced noise is set to zero, while all other settings remain consistent with those of the original SynFog dataset. We compare the defogging and detection performance of models trained on SynFog and SynFog- $\alpha$ datasets, as shown in Fig. 8 and Tab. 4-5. The model trained on SynFog that incorporates sensor noise produces more natural results with fewer artifacts and higher detection accuracy. + +# 5.4. Algorithm Benchmarking + +We use SynFog test set to assess the aforementioned representative defogging methods, employing both full-reference and no-reference metrics to evaluate the defogged results. To illustrate the challenges presented by SynFog, we also provide evaluation results from the widely-used RESIDE dataset [25]. The full-scale quantitative comparison results are available in the supplementary material. + +Generally, learning-based methods consistently outperform earlier algorithms based on natural or statistical priors + +![](images/e07ca0b11678860c79835939bbf540d43c3deca8b839c233bd62950de2c71362.jpg) +Input + +![](images/4c5c2cd1ee05d201650cec21122bf3748f4a806a7c6dc809b08ec2dd3cca7b52.jpg) +Foggy Cityscapes [40] + +![](images/4997464cd8e9ce6b3a37cac63b0fd43b0009085fb91597498fef1b0d1c38cff1.jpg) +Virtual KITTI [18] + +![](images/fdba5094e0c1b731c62ccd2b5f282dbd6aaffd1363bb521d28053f4e9c24d86a.jpg) +SynFog-β + +![](images/94829d71ccfa1bb7c332a075e0a98a297a365025a39765727253e536c8eda625.jpg) +SynFog (Ours) + +![](images/116bde5ef6776d2eb1f4ac1357261707c837680ad51c734e1eb86a4bb38ad9d9.jpg) +Figure 9. Visual and detection comparisons on real-world foggy images using DehazeFormer models trained on SynFog, SynFog- $\beta$ , Foggy Cityscapes [40] and Virtual KITTI [18]. + +![](images/cd12f59177431e7fb1b49ca9b57a17c01cc7913056c61742ff142a7b53984d87.jpg) +Figure 10. Qualitative comparisons of SOTA defogging methods on SynFog dataset with different fog levels. Zoom in for better view. + +![](images/69bf9067694c72d7790432942b68d68295dd3f305ddf59c4984c8cdf7c2c6222.jpg) + +![](images/5ab8230ca5c33c9d8b8bad43cd490f928f71e3dbb9adb5c99b19c84771782179.jpg) + +![](images/5f1c4b5bc22508ce06538c71a12e8dba023c246d35663e717f8807ec49582fd3.jpg) + +![](images/cffffc0a4cf8cd4aa56d546671d0c40aa1c05ace67aecea7326ccb43a6561ede.jpg) + +![](images/e4c33cea8ea00fe764c834af25cde06857646179ee7679837bb8cb3a525c0e3c.jpg) + +![](images/204d86f1104f4224248e1368d9d5cbd9ef2b9168f4bcbf77c9eb84b50f023896.jpg) + +![](images/e97b2a3b5d69573f987a245210488ab488a0de8f144e894c9796becfb9f29f2e.jpg) + +in terms of PSNR, SSIM and LPIPS [54]. In particular, DehazeFormer [43] attains the highest PSNR (27.96) and SSIM (0.923) values, while GridDehazeNet [28] achieves the best LPIPS (0.163) value on SynFog test set. Although the overall performance trend of the algorithms on SynFog is similar to that on RESIDE [25], the evaluation scores are noticeably lower, highlighting the challenges posed by SynFog. Fig. 10 depicts the qualitative comparison among these methods across different fog levels. Images restored by DCP [20] and CAP [57] tend to appear excessively dark and noisy due to inaccurate airlight estimation and the lack of effective denoising capabilities. GridDehazeNet [28] lacks color recovery ability in dense fog, while the outputs of FFA-Net [39] and DehazeFormer [43] seem smooth. This comparison further indicates the substantial room for improvement in existing defogging methods, especially when applied to more realistic fog datasets. + +# 5.5. Discussion and Limitations + +Experimental results reveal that models trained on SynFog, incorporating realistic physical characteristics of optics and sensors, exhibit a more robust performance on real-word fog datasets in terms of visual perception and detection accuracy. This outcome implies that integrating a physically-based end-to-end imaging process replicating real capture methods can significantly improve the generalization of + +trained models. However, given the time-consuming nature of the physically-based volumetric path tracing technique, the scale of SynFog is not as extensive as other synthetic fog datasets. Nonetheless, we are committed to expanding the scene contents in the future and exploring a broader range of scattering mediums beyond fog. + +# 6. Conclusion + +In this paper, we present an end-to-end foggy image simulation pipeline. Our approach utilizes volumetric path tracing to model a more precise light scattering process with global illumination. By incorporating a physically-based camera processing pipeline that includes optics, sensor and image processing, we can closely mimic the authentic capture process under foggy conditions. Additionally, we develop a new synthetic fog dataset, SynFog, to facilitate the research on defogging. Comprehensive experiments have validated the authenticity and reliability of the SynFog dataset. + +Acknowledgements This work is supported by 1) National Natural Science Foundation of China (Grant No. 62131011 and 62375233); 2) NSFC Young Scientists Fund (Grant No. 62302423). We thank Professor Chuxi Yang and Jiayue Xie for their help in preparing the scene creation. We also thank Professor Brian Wandell, Dr. Devesh Upadhyay, Dr. Alireza Rahimpour, and Devesh Shah for their valuable advice at the early stage of this project. + +# References + +[1] Cosmin Ancuti, Codruta O Ancuti, and Christophe De Vleeschouwer. D-hazy: A dataset to evaluate quantitatively dehazing algorithms. In IEEE International Conference on Image Processing, pages 2226-2230, 2016. 2 +[2] Codruta O Ancuti, Cosmin Ancuti, Radu Timofte, and Christophe De Vleeschouwer. O-haze: a dehazing benchmark with real hazy and haze-free outdoor images. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 754-762, 2018. 1, 2, 4, 7 +[3] Codruta O Ancuti, Cosmin Ancuti, Mateu Sbert, and Radu Timofte. Dense-haze: A benchmark for image dehazing with dense-haze and haze-free images. In IEEE International Conference on Image Processing, pages 1014-1018, 2019. +[4] Codruta O Ancuti, Cosmin Ancuti, and Radu Timofte. Nh-haze: An image dehazing benchmark with nonhomogeneous hazy and haze-free images. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 444-445, 2020. 1, 2, 4 +[5] Mario Bijelic, Tobias Gruber, Fahim Mannan, Florian Kraus, Werner Ritter, Klaus Dietmayer, and Felix Heide. Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11682-11692, 2020. 1, 3, 6, 7 +[6] Norbert Buch, Sergio A Velastin, and James Orwell. A review of computer vision techniques for the analysis of urban traffic. IEEE Transactions on Intelligent Transportation Systems, 12(3):920-939, 2011. 1 +[7] Ting Chen. Digital camera system simulator and applications. Stanford University, 2003. 4 +[8] Lark Kwon Choi, Jaehee You, and Alan Conrad Bovik. Referenceless prediction of perceptual fog density and perceptual image defogging. IEEE Transactions on Image Processing, 24(11):3888-3901, 2015. 3 +[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3213-3223, 2016. 2 +[10] Dengxin Dai and Wen Yang. Satellite image classification via two-layer sparse coding with biased image representation. IEEE Geoscience and Remote Sensing Letters, 8(1): 173-176, 2010. 1 +[11] Dengxin Dai, Christos Sakaridis, Simon Hecker, and Luc Van Gool. Curriculum model adaptation with synthetic and real data for semantic foggy scene understanding. International Journal of Computer Vision, 128:1182-1204, 2020. 2, 6, 7 +[12] Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming-Hsuan Yang. Multi-scale boosted dehazing network with dense feature fusion. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2157-2167, 2020. 6 + +[13] J.B. Farrow et al. Influence of the atmosphere on optical systems. Optica Acta, 1970. 5 +[14] X. Min et al. Objective quality evaluation of dehazed images. TITS, 2018. 7 +[15] Joyce E Farrell and Brian A Wandell. Image systems simulation. Handbook of Digital Imaging, 1:373-400, 2015. 4 +[16] Joyce E Farrell, Feng Xiao, Peter Bert Catrysse, and Brian A Wandell. A simulation tool for evaluating digital camera image quality. In Image Quality and System Performance, pages 124-131. SPIE, 2003. +[17] Joyce E Farrell, Peter B Catrysse, and Brian A Wandell. Digital camera simulation. Applied Optics, 51(4):A80-A90, 2012. 3, 4 +[18] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4340-4349, 2016. 2, 5, 6, 8 +[19] Martin Hahner, Dengxin Dai, Christos Sakaridis, Jan-Nico Zaech, and Luc Van Gool. Semantic understanding of foggy scenes with purely synthetic data. In IEEE Intelligent Transportation Systems Conference (ITSC), pages 3675-3681, 2019. 2, 3 +[20] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(12):2341-2353, 2010. 6, 8 +[21] Louis G Henyey and Jesse Leonard Greenstein. Diffuse radiation in the galaxy. Astrophysical Journal, vol. 93, p. 70-83 (1941), 93:70-83, 1941. 4 +[22] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolo, 2023. 6 +[23] Kyung Won Kim. The comparison of visibility measurement between image-based visual range, human eye-based visual range, and meteorological optical range. Atmospheric Environment, 190:74-86, 2018. 5 +[24] Miles V Klein and Thomas E Furtak. Optik. Springer-Verlag, 2013. 4 +[25] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2018. 1, 2, 3, 7, 8 +[26] Ruoteng Li, Xiaoyi Zhang, Shaodi You, and Yu Li. Realistic large-scale fine-depth dehazing dataset from 3d videos. arXiv preprint arXiv:2004.08554, 2020. 2 +[27] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(10):2024-2039, 2015. 2 +[28] Xiaohong Liu, Yongrui Ma, Zhihao Shi, and Jun Chen. Griddehazenet: Attention-based multi-scale network for image dehazing. In IEEE International Conference on Computer Vision, pages 7314-7323, 2019. 6, 8 +[29] Yun Liu, Zhongsheng Yan, Aimin Wu, Tian Ye, and Yuche Li. Nighttime image dehazing based on variational decomposition model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 640-649, 2022. 3 + +[30] Zhenyi Liu, Trisha Lian, Joyce Farrell, and Brian Wandell. Soft prototyping camera designs for car detection based on a convolutional neural network. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 4 +[31] Zhenyi Liu, Minghao Shen, Jiaqi Zhang, Shuangting Liu, Henryk Blasinski, Trisha Lian, and Brian Wandell. A system for generating complex physically accurate sensor images for automotive applications. arXiv preprint arXiv:1902.04258, 2019. 4 +[32] Zheng Lyu, Thomas Goossens, Brian A Wandell, and Joyce Farrell. Validation of physics-based image systems simulation with 3-d scenes. IEEE Sensors Journal, 22(20):19400-19410, 2022. 4 +[33] Earl J McCartney. Optics of the atmosphere: scattering by molecules and particles. New York, 1976. 1, 3, 5, 6 +[34] Srinivasa G Narasimhan and Shree K Nayar. Vision and the atmosphere. International Journal of Computer Vision, 48: 233-254, 2002. 3, 5, 6 +[35] Priya Narayanan, Xin Hu, Zhenyu Wu, Matthew D Thielke, John G Rogers, Andre V Harrison, John A D'Agostino, James D Brown, Long P Quang, James R Uplinger, et al. A multi-purpose realistic haze benchmark with quantifiable haze levels and ground truth. IEEE Transactions on Image Processing, 2023. 1, 2 +[36] National Standards of People's Republic of China. Grade of fog forecast, 2011. 5 +[37] Roberto Nebuloni. Empirical relationships between extinction coefficient and visibility in fog. Applied Optics, 44(18): 3795-3804, 2005. 5 +[38] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically based rendering: From theory to implementation. MIT Press, 2023. 2, 3, 4 +[39] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In AAAI, pages 11908-11915, 2020. 6, 8 +[40] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Semantic foggy scene understanding with synthetic data. International Journal of Computer Vision, 126:973-992, 2018. 1, 2, 3, 5, 6, 7, 8 +[41] Daniel Scharstein and Richard Szeliski. High-accuracy stereo depth maps using structured light. In IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages I-I. IEEE, 2003. 2 +[42] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In European Conference on Computer Vision, pages 746-760. Springer, 2012. 2 +[43] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927-1941, 2023. 6, 7, 8 +[44] Tao Sun, Mattia Segu, Janis Postels, Yuxuan Wang, Luc Van Gool, Bernt Schiele, Federico Tombari, and Fisher Yu. Shift: a synthetic driving dataset for continuous multi-task domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 21371-21382, 2022. 2 + +[45] Ketan Tang, Jianchao Yang, and Jue Wang. Investigating haze-relevant features in a learning framework for image dehazing. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2995-3000, 2014. 7 +[46] Jean-Philippe Tarel, Nicolas Hautiere, Aurélien Cord, Dominique Gruyer, and Houssam Halmaoui. Improved visibility of road scene images under heterogeneous fog. In IEEE Intelligent Vehicles Symposium, pages 478–485, 2010. 2 +[47] Jean-Philippe Tarel, Nicholas Hautiere, Laurent Caraffa, Aurélien Cord, Houssam Halmaoui, and Dominique Gruyer. Vision enhancement in homogeneous and heterogeneous fog. IEEE Intelligent Transportation Systems Magazine, 4 (2):6-20, 2012. 2 +[48] Tali Treibitz and Yoav Y Schechner. Resolution loss without imaging blur. JOSA A, 29(8):1516-1528, 2012. 3 +[49] Apostolia Tsirikoglou, Joel Kronander, Magnus Wrenninger, and Jonas Unger. Procedural modeling and physically based rendering for synthetic data generation in automotive applications. arXiv preprint arXiv:1710.06270, 2017. 4 +[50] Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Contrastive learning for compact single image dehazing. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10551-10560, 2021. 6 +[51] Jing Zhang, Yang Cao, Zheng-Jun Zha, and Dacheng Tao. Nighttime dehazing with a synthetic benchmark. In Proceedings of the ACM International Conference on Multimedia, pages 2355-2363, 2020. 3 +[52] Lin Zhang, Junhao Huang, Xiyuan Li, and Lu Xiong. Vision-based parking-slot detection: A dcnn-based approach and a large-scale benchmark dataset. IEEE Transactions on Image Processing, 27(11):5350-5364, 2018. 1 +[53] Lin Zhang, Anqi Zhu, Shiyu Zhao, and Yicong Zhou. Simulation of atmospheric visibility impairment. IEEE Transactions on Image Processing, 30:8713-8726, 2021. 2, 3 +[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 8 +[55] Yanfu Zhang, Li Ding, and Gaurav Sharma. Hazerd: an outdoor scene dataset and benchmark for single image dehazing. In IEEE International Conference on Image Processing, pages 3205-3209, 2017. 1, 2 +[56] Shiyu Zhao, Lin Zhang, Shuaiyi Huang, Ying Shen, and Shengjie Zhao. Dehazing evaluation: Real-world benchmark datasets, criteria, and baselines. IEEE Transactions on Image Processing, 29:6947-6962, 2020. 7 +[57] Qingsong Zhu, Jiaming Mai, and Ling Shao. A fast single image haze removal algorithm using color attenuation prior. IEEE Transactions on Image Processing, 24(11):3522-3533, 2015. 6, 8 \ No newline at end of file diff --git a/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/images.zip b/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1ed16f5b947ace6175c1128befd97483d0ee0d23 --- /dev/null +++ b/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:015a7699c4b490f106177b9bf626d8be95049f7e2863a481d9241eb280418b91 +size 671121 diff --git a/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/layout.json b/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..446019c725ff5dad2df5dd7787deb875c217b4c7 --- /dev/null +++ b/2024/SynFog_ A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving/layout.json @@ -0,0 +1,9284 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 55, + 103, + 539, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 539, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 539, + 140 + ], + "type": "text", + "content": "SynFog: A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 160, + 499, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 160, + 499, + 223 + ], + "spans": [ + { + "bbox": [ + 95, + 160, + 499, + 223 + ], + "type": "text", + "content": "Yiming Xie" + }, + { + "bbox": [ + 95, + 160, + 499, + 223 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 95, + 160, + 499, + 223 + ], + "type": "text", + "content": ", Henglu Wei" + }, + { + "bbox": [ + 95, + 160, + 499, + 223 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 95, + 160, + 499, + 223 + ], + "type": "text", + "content": ", Zhenyi Liu" + }, + { + "bbox": [ + 95, + 160, + 499, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 160, + 499, + 223 + ], + "type": "text", + "content": ", Xiaoyu Wang" + }, + { + "bbox": [ + 95, + 160, + 499, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 160, + 499, + 223 + ], + "type": "text", + "content": ", Xiangyang Ji" + }, + { + "bbox": [ + 95, + 160, + 499, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 160, + 499, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 160, + 499, + 223 + ], + "type": "text", + "content": "Tsinghua University, " + }, + { + "bbox": [ + 95, + 160, + 499, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 160, + 499, + 223 + ], + "type": "text", + "content": "Stanford University \n{xieym22@mails., weihenglu@mail., xiaoyu-w17@mails., xyji@}tsinghua.edu.cn, {zhenyiliu@}stanford.edu" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 54, + 251, + 541, + 363 + ], + "blocks": [ + { + "bbox": [ + 54, + 251, + 541, + 363 + ], + "lines": [ + { + "bbox": [ + 54, + 251, + 541, + 363 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 541, + 363 + ], + "type": "image", + "image_path": "8e1ff6fee21aab43dc8c8bc95515cc43b9b31ed14ba4d64a7dd2e191067e2632.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 373, + 546, + 407 + ], + "lines": [ + { + "bbox": [ + 46, + 373, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 546, + 407 + ], + "type": "text", + "content": "Figure 1. The SynFog dataset comprises images under different lighting, with the first row captured in natural sky light and the second with extra light sources like street lamps and automotive lighting. This dataset covers three levels of fog density across each scene, corresponding visibility of " + }, + { + "bbox": [ + 46, + 373, + 546, + 407 + ], + "type": "inline_equation", + "content": "600\\mathrm{m}" + }, + { + "bbox": [ + 46, + 373, + 546, + 407 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 373, + 546, + 407 + ], + "type": "inline_equation", + "content": "300\\mathrm{m}" + }, + { + "bbox": [ + 46, + 373, + 546, + 407 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 373, + 546, + 407 + ], + "type": "inline_equation", + "content": "150\\mathrm{m}" + }, + { + "bbox": [ + 46, + 373, + 546, + 407 + ], + "type": "text", + "content": ". Additionally, pixel-accurate depth data and segmentation labels for each scene are also provided." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 412, + 192, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 412, + 192, + 425 + ], + "spans": [ + { + "bbox": [ + 143, + 412, + 192, + 425 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 430, + 289, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 289, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 289, + 658 + ], + "type": "text", + "content": "To advance research in learning-based defogging algorithms, various synthetic fog datasets have been developed. However, existing datasets created using the Atmospheric Scattering Model (ASM) or real-time rendering engines often struggle to produce photo-realistic foggy images that accurately mimic the actual imaging process. This limitation hinders the effective generalization of models from synthetic to real data. In this paper, we introduce an end-to-end simulation pipeline designed to generate photorealistic foggy images. This pipeline comprehensively considers the entire physically-based foggy scene imaging process, closely aligning with real-world image capture methods. Based on this pipeline, we present a new synthetic fog dataset named SynFog, which features both sky light and active lighting conditions, as well as three levels of fog density. Experimental results demonstrate that models trained on SynFog exhibit superior performance in visual perception and detection accuracy compared to others when applied to real-world foggy images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 667, + 128, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 667, + 128, + 680 + ], + "spans": [ + { + "bbox": [ + 47, + 667, + 128, + 680 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Fog, due to its light scattering and attenuation properties, poses a significant challenge for outdoor vision-based sys" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 414, + 546, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 546, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 546, + 485 + ], + "type": "text", + "content": "tems such as remote sensing, surveillance, autonomous driving, and fire rescue [6, 10, 52]. It causes severe degradation in object appearance and contrast, resulting in a notable impact on people's daily lives. Consequently, there has been a growing research interest in image defogging task to develop robust outdoor vision systems." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "content": "Recent learning-based defogging algorithms depend on sets of foggy images and corresponding clear images for model training. While several datasets have been collected from real-world scenarios [5, 40], they are typically limited to foggy images only, as it is almost impossible to obtain pixel-accurate clear images with consistent scene contents and environmental illumination. Studies in [2-4, 35] attempt to capture real-world foggy and clear image pairs using smoke machines. However, the scale of these datasets is inadequate for training of a high-performing model. Given the challenge of obtaining paired foggy and clear images in real world, various synthetic fog datasets have been proposed as an alternative [25, 40, 55]. The most widely used method is based on the theory proposed by McCartney [33], also known as the Atmospheric Scattering Model (ASM). However, this model fails to consider accurate global illumination and the actual imaging process that occurs in real-world captures as detailed in Sec. 3.1. These limitations significantly impact the model's overall applicability, par" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21763" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 68, + 294, + 203 + ], + "blocks": [ + { + "bbox": [ + 52, + 68, + 294, + 203 + ], + "lines": [ + { + "bbox": [ + 52, + 68, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 52, + 68, + 294, + 203 + ], + "type": "image", + "image_path": "67fb9f7b4c55f8898211c4c1c3a14228300936503c99f3768c41d39632ef0019.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 287, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 287, + 258 + ], + "type": "text", + "content": "Figure 2. Relationship between noise level and fog density in real-world foggy images. Using regional contrast stretching for image enhancement, it can be observed that the noise level increases with the concentration of fog." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 268, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 287, + 316 + ], + "type": "text", + "content": "ticularly in nighttime scenarios, and can lead to a disparity between synthetic and real-world foggy images. As a result, models trained on synthetic foggy images may exhibit limited robustness when applied to real-world ones." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 579 + ], + "type": "text", + "content": "To address these issues, we propose an end-to-end foggy image simulation pipeline that encompasses the entire imaging process starting from 3D virtual scenes to final photorealistic foggy images. Our approach involves foggy scene rendering using volumetric path tracing and imaging through a realistic camera model. As shown in Fig. 3b, fog is treated as a spatial volume, and the spectral scene radiance is rendered using volumetric path tracing [38]. By defining the phase function of fog medium and bidirectional reflectance distribution function (BRDF) of the scene components, we are able to accurately describe the scattering process with multiple light sources under foggy conditions. The foggy scene radiance data is further processed through a physically-based camera model, which comprises optics, sensor and image processing, to faithfully replicate real camera devices. Based on the proposed pipeline, we develop a new extensive synthetic fog dataset called SynFog. This dataset contains 500 unique outdoor scenes, each featuring two kinds of lighting conditions and three levels of fog density. Extensive experiments have been conducted to qualitatively and quantitatively validate the authenticity of our simulation pipeline and the SynFog dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 580, + 266, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 580, + 266, + 590 + ], + "spans": [ + { + "bbox": [ + 58, + 580, + 266, + 590 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 592, + 286, + 710 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 592, + 286, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 592, + 286, + 638 + ], + "spans": [ + { + "bbox": [ + 47, + 592, + 286, + 638 + ], + "type": "text", + "content": "- We propose a physically-based end-to-end foggy image simulation pipeline that incorporates accurate light transportation in scattering medium and physical characteristics of optics and sensor into the synthesized images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 639, + 286, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 286, + 673 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 286, + 673 + ], + "type": "text", + "content": "- We develop a photo-realistic synthetic fog dataset, SynFog, which includes both sky light and active lighting conditions, as well as three levels of fog density." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 676, + 286, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 676, + 286, + 710 + ], + "spans": [ + { + "bbox": [ + 47, + 676, + 286, + 710 + ], + "type": "text", + "content": "- We demonstrate the authenticity and effectiveness of our simulation pipeline and the superior realism of SynFog dataset compared to other synthetic fog datasets." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 91, + 545, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 162 + ], + "type": "text", + "content": "Given the challenge of acquiring paired foggy and clear images in the real world for training learning-based defogging methods, researchers have proposed many synthetic fog datasets as an alternative. In this section, we will offer a thorough review and analysis of foggy image simulation techniques and their associated datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 163, + 546, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 546, + 533 + ], + "type": "text", + "content": "Based on Atmospheric Scattering Model (ASM). This method involves rendering fog onto clean images along with their depth information [1, 19, 25, 26, 40, 53, 55]. Depth information can be acquired through game engines, depth cameras or depth estimation methods. In an earlier study conducted by Tarel et al. [46], FRIDA was constructed using a total of 90 synthetic images, with clear images and depth maps generated using " + }, + { + "bbox": [ + 304, + 163, + 546, + 533 + ], + "type": "inline_equation", + "content": "SiVIC^{TM}" + }, + { + "bbox": [ + 304, + 163, + 546, + 533 + ], + "type": "text", + "content": " software. Subsequently, in their later work [47], more synthetic images were added, leading to an upgraded version named FRIDA2, which includes 330 synthetic images. However, the scale of these datasets is insufficient for modern data-driven defogging methods, and the foggy images appear unrealistic. In more recent works [26, 53], a similar approach has been employed to acquire both the clear images and their corresponding depth maps from game engines. In [25], a general dehazing dataset called RESIDE was introduced. The indoor training set (ITS) of RESIDE was constructed based on existing indoor RGBD datasets NYUv2 [42] and Middleburry stereo [41]. For the outdoor training set (OTS), 2071 real-world outdoor images were collected, and the corresponding depth information was estimated using a learning-based technique [27]. It is important to note that the accuracy of depth estimation based on learning methods is still unsatisfactory, as single-image-based depth estimation is inherently an ill-posed problem. Furthermore, Foggy Cityscapes was introduced in [40], utilizing images from the Cityscapes dataset [9]. Despite efforts to enhance the original depth maps in Cityscapes, artifacts and discrepancies still persist, resulting in synthetic foggy images that appear unnatural." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 605 + ], + "type": "text", + "content": "Based on Real-time Rendering Engines. Foggy images generated utilizing real-time rendering engines like Unity and Unreal have been discussed in works such as [18, 44]. Among them, Virtual KITTI 2 [18] stands out as a notable example. However, these foggy images often lack realism and exhibit limited diversity in terms of fog density." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "text", + "content": "Based on Physical Means. Fog/haze machines are employed in real-world settings to generate artificial fog, as discussed in works such as [2-4, 35]. However, these datasets are generally limited in scale and may not offer sufficient supervision for learning-based defogging methods. Moreover, ensuring consistent content between the clear and foggy image captures can be challenging." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "There are also several real-world fog datasets available, such as Foggy Driving [40], Foggy Zurich [11], and Seeing" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21764" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 75, + 262, + 197 + ], + "blocks": [ + { + "bbox": [ + 72, + 75, + 262, + 197 + ], + "lines": [ + { + "bbox": [ + 72, + 75, + 262, + 197 + ], + "spans": [ + { + "bbox": [ + 72, + 75, + 262, + 197 + ], + "type": "image", + "image_path": "09614b1fb99b25afb9a90a4e90e07fad931f0d2f43194ac75c63869194cddd1a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 204, + 244, + 215 + ], + "lines": [ + { + "bbox": [ + 90, + 204, + 244, + 215 + ], + "spans": [ + { + "bbox": [ + 90, + 204, + 244, + 215 + ], + "type": "text", + "content": "(a) Atmospheric Scattering Model (ASM) [34]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 70, + 224, + 263, + 327 + ], + "blocks": [ + { + "bbox": [ + 70, + 224, + 263, + 327 + ], + "lines": [ + { + "bbox": [ + 70, + 224, + 263, + 327 + ], + "spans": [ + { + "bbox": [ + 70, + 224, + 263, + 327 + ], + "type": "image", + "image_path": "d5eb05dde2831ddd5af86a71a9db1f88f3a3f7617d3e89d78f062ff002a2d09e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 333, + 264, + 344 + ], + "lines": [ + { + "bbox": [ + 70, + 333, + 264, + 344 + ], + "spans": [ + { + "bbox": [ + 70, + 333, + 264, + 344 + ], + "type": "text", + "content": "(b) Foggy scene rendering based on volumetric path tracing." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 87, + 353, + 246, + 365 + ], + "lines": [ + { + "bbox": [ + 87, + 353, + 246, + 365 + ], + "spans": [ + { + "bbox": [ + 87, + 353, + 246, + 365 + ], + "type": "text", + "content": "Figure 3. Foggy scene simulation methods." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 462 + ], + "type": "text", + "content": "Through Fog [5]. However, due to the absence of corresponding clear reference images, these datasets can only be used for testing purposes rather than for training. In this paper, we propose a physically-based end-to-end foggy image simulation pipeline to improve the authenticity of synthetic foggy images and construct a photo-realistic fog dataset to facilitate future research on defogging algorithms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 475, + 248, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 248, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 248, + 489 + ], + "type": "text", + "content": "3. End-to-end Foggy Image Simulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 498, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 654 + ], + "type": "text", + "content": "We propose an end-to-end approach for generating photorealistic foggy images that incorporates accurate light transportation in scattering medium and physical characteristics of optics and sensor into the synthesized images, as illustrated in Fig. 4. We first introduce a foggy scene imaging model in Sec. 3.1. Based on this model, we establish a simulation pipeline consisting of two components: a) Foggy scene radiance is rendered using volumetric path tracing [38]. b) The radiance data is processed through a physically-based camera model, which comprises optics, sensor and image processing to faithfully replicate real camera devices [17]. These components are further detailed in Sec. 3.2.1 and Sec. 3.2.2, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 669, + 203, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 203, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 203, + 681 + ], + "type": "text", + "content": "3.1. Foggy Scene Imaging Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "The Atmospheric Scattering Model (ASM) [33, 34] has been widely used for the generation of foggy images [19," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 354, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 354, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 354, + 84 + ], + "type": "text", + "content": "25,40,53]:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 95, + 545, + 141 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 95, + 545, + 141 + ], + "spans": [ + { + "bbox": [ + 315, + 95, + 545, + 141 + ], + "type": "interline_equation", + "content": "E (d, \\lambda) = \\underbrace {E _ {o} (\\lambda) e ^ {- \\beta (\\lambda) d}} _ {\\text {t r a n s m i s s i o n}} + \\underbrace {L _ {h} (\\infty , \\lambda) \\left(1 - e ^ {- \\beta (\\lambda) d}\\right)} _ {\\text {a i r l i g h t}}, \\tag {1}", + "image_path": "dd1967f4e83f7435b945ef88b7f34344100bdd62b59ee6de230bfd569fe8c3ce.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "text", + "content": " is the irradiance reaching the film under foggy conditions, " + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "inline_equation", + "content": "E_{o}" + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "text", + "content": " is the irradiance under clear conditions, " + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "text", + "content": " is the scattering coefficient characterizing the scatter ability of the medium, " + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "text", + "content": " is the distance between the object and the camera, " + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "inline_equation", + "content": "L_{h}(\\infty ,\\lambda)" + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "text", + "content": " denotes the radiance of airlight at infinity and " + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "text", + "content": " refers to wavelength. Fig. 3a provides a more intuitive description of this process. Foggy images can be generated with pre-defined " + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "inline_equation", + "content": "L_{h}" + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "text", + "content": " values by directly using the clear images as " + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "inline_equation", + "content": "E_0" + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "text", + "content": " and the corresponding depth maps as " + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 251, + 546, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 546, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 546, + 371 + ], + "type": "text", + "content": "However, there are two issues with this scattering model for generating foggy images: (1) The definition of environmental illumination fails to consider the contribution of indirect illumination from the scene to the scattering medium, resulting in an unfaithful representation of the actual illumination. (2) The current formulation of airlight is insufficient for accurately modeling the lighting conditions that involve multiple point light sources, such as street lamps and automotive lights. This limitation restricts the model's applicability in situations like nighttime [29, 51]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 372, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 492 + ], + "type": "text", + "content": "Furthermore, ASM only captures the irradiance reaching the sensor plane. Consequently, foggy images generated using ASM fail to accurately incorporate the realistic physical characteristics of optics and sensor in actual imaging process. In Fig. 2, we depict the relationship between noise level and fog density after the enhancement of regional contrast stretching [48] on a real captured image [8]. It can be seen that the noise level increases with higher fog density, emphasizing the critical need for accurate noise modeling in foggy image simulation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 494, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 545, + 541 + ], + "type": "text", + "content": "To enhance the consistency between synthetic images and real-world captures, we develop a comprehensive foggy scene imaging model that incorporates the entire imaging process, referred to as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 335, + 555, + 545, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 555, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 335, + 555, + 545, + 568 + ], + "type": "interline_equation", + "content": "I _ {o u t} = f _ {I S P} \\left(f _ {\\text {s e n s o r}} \\left(f _ {\\text {o p t i c s}} \\left(L _ {\\text {s c e n e}}\\right)\\right) + n\\right), \\tag {2}", + "image_path": "6c722ef4b61ad5011eb1d6b8ed6696656335b1fd96ad775ea4fb6cbd4fa01ffc.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "inline_equation", + "content": "L_{\\text{scene}}" + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "content": " is the scene radiance obtained using volumetric path tracing [38], as described in Sec. 3.2.1; " + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\{f_{\\text{optics}}, f_{\\text{sensor}}, f_{\\text{ISP}}\\}" + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "content": " denote the optics, sensor and ISP processes respectively, as described in Sec. 3.2.2, and the noise term introduced by the sensor is denoted as " + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "type": "text", + "content": "By employing volumetric path tracing to obtain the radiance of foggy scene, we can precisely model the light scattering process under global illumination and active lighting. By incorporating the complete imaging process described in Eq. (2), we are able to faithfully replicate the authentic camera characteristics in foggy image simulation." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21765" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 200, + 137 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 200, + 137 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 200, + 137 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 200, + 137 + ], + "type": "image", + "image_path": "367edc5b7d3434f72977b0fa071f665ea871022e136cb73b9d7fa4a9bb9a9005.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 146, + 547, + 180 + ], + "lines": [ + { + "bbox": [ + 46, + 146, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 547, + 180 + ], + "type": "text", + "content": "Figure 4. End-to-end foggy image simulation pipeline. The spectral radiance data is rendered using volumetric path tracing and passing through a realistic optics model before reaching the sensor plane. Subsequently, the irradiance is converted into a digital image through the sensor model. The raw image from the sensor is then processed by an ISP module to generate the final foggy image." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 205, + 71, + 302, + 133 + ], + "blocks": [ + { + "bbox": [ + 205, + 71, + 302, + 133 + ], + "lines": [ + { + "bbox": [ + 205, + 71, + 302, + 133 + ], + "spans": [ + { + "bbox": [ + 205, + 71, + 302, + 133 + ], + "type": "image", + "image_path": "5ec3364dd38c1088a85fc476b4df2f784def9bf0ec4bb3340002c6aafce770cd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 302, + 71, + 416, + 134 + ], + "blocks": [ + { + "bbox": [ + 302, + 71, + 416, + 134 + ], + "lines": [ + { + "bbox": [ + 302, + 71, + 416, + 134 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 416, + 134 + ], + "type": "image", + "image_path": "4bc42c76256e1ae821b6a302cb41af5c096e23803feae3c13329c018a3a23a79.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 417, + 71, + 545, + 137 + ], + "blocks": [ + { + "bbox": [ + 417, + 71, + 545, + 137 + ], + "lines": [ + { + "bbox": [ + 417, + 71, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 417, + 71, + 545, + 137 + ], + "type": "image", + "image_path": "ef7d7da76639afea1f4855a58d7e8e835c60e548f57ae5c705c1b11b202f0798.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 191, + 217, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 191, + 217, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 191, + 217, + 205 + ], + "type": "text", + "content": "3.2. End-to-end Simulation Pipeline" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 210, + 181, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 210, + 181, + 223 + ], + "spans": [ + { + "bbox": [ + 47, + 210, + 181, + 223 + ], + "type": "text", + "content": "3.2.1 Foggy Scene Rendering" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 229, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 287, + 348 + ], + "type": "text", + "content": "We utilize procedural modeling techniques, similar to the approach employed in [31, 49], to create large-scale and diverse driving scenes. This involves defining adjustable parameters and using a set of predefined rules to generate explicit scene definitions. Our model library incorporates a diverse range of pedestrians, bicyclists and cars, each with well-defined material and texture properties. Furthermore, we carefully select 16 sky maps with cloudy conditions to ensure realistic lighting for foggy scenes, following the established guidelines from previous works [2-4]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 349, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 287, + 420 + ], + "type": "text", + "content": "In our approach, fog is treated as a spatial volume, and the spectral scene radiance is rendered using volumetric path tracing [38]. To accurately model the scattering process of the fog medium, we employ Henyey and Greenstein phase function [21, 38] to model its particle scattering characteristics, denoted as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 428, + 287, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 428, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 77, + 428, + 287, + 458 + ], + "type": "interline_equation", + "content": "p _ {\\mathrm {H G}} (\\cos \\theta) = \\frac {1}{4 \\pi} \\frac {1 - g ^ {2}}{\\left(1 + g ^ {2} + 2 g (\\cos \\theta)\\right) ^ {3 / 2}}. \\tag {3}", + "image_path": "80a4c0b04ed8d04743da9931d76115e7d894d7ab522fe137b9850b10adeb2064.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 465, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 465, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 465, + 287, + 514 + ], + "type": "text", + "content": "By empirically setting the asymmetry parameter " + }, + { + "bbox": [ + 46, + 465, + 287, + 514 + ], + "type": "inline_equation", + "content": "g = 0.87" + }, + { + "bbox": [ + 46, + 465, + 287, + 514 + ], + "type": "text", + "content": ", we can accurately simulate the probability distributions for scattering in a specific direction at a given point within the fog medium, as shown in Fig. 3b." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 514, + 287, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 562 + ], + "type": "text", + "content": "The rendering equation consists of two parts that account for reduction of radiance caused by out-scattering and increase of radiance due to in-scattering at point " + }, + { + "bbox": [ + 46, + 514, + 287, + 562 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 514, + 287, + 562 + ], + "type": "text", + "content": " in direction " + }, + { + "bbox": [ + 46, + 514, + 287, + 562 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 514, + 287, + 562 + ], + "type": "text", + "content": ", represented by Eq. (4) and Eq. (5), respectively [38]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 82, + 570, + 287, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 570, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 82, + 570, + 287, + 586 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {i}} \\left(\\mathrm {p} ^ {\\prime}, - \\omega\\right) = \\mathrm {e} ^ {- \\int_ {0} ^ {d} \\sigma_ {\\mathrm {s}} (\\mathrm {p} + t \\omega , \\omega) \\mathrm {d} t} L _ {\\mathrm {o}} (\\mathrm {p}, \\omega). \\tag {4}", + "image_path": "0e595884f26720c9645ce9db5203131b4ef0179234df72272a99ea1645ac630a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 594, + 295, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 295, + 628 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {s}} (\\mathrm {p}, \\omega) = L _ {\\mathrm {e}} (\\mathrm {p}, \\omega) + \\sigma_ {\\mathrm {s}} (\\mathrm {p}, \\omega) \\int_ {\\mathcal {S} ^ {2}} p (\\mathrm {p}, \\omega_ {\\mathrm {i}}, \\omega) L _ {\\mathrm {i}} (\\mathrm {p}, \\omega_ {\\mathrm {i}}) \\mathrm {d} \\omega_ {\\mathrm {i}}. \\tag {5}", + "image_path": "fa4b8b052064c86d2dc5ebb6639bcd9702d2edee354fc94c24223927e8c0de5e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 630, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\sigma_{s}" + }, + { + "bbox": [ + 47, + 630, + 287, + 677 + ], + "type": "text", + "content": " represents the scattering probability per unit distance, " + }, + { + "bbox": [ + 47, + 630, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\{L_i,L_o,L_e,L_s\\}" + }, + { + "bbox": [ + 47, + 630, + 287, + 677 + ], + "type": "text", + "content": " are the {incident,exitant,emission,total added} radiance. In our implementation, we employ the Monte Carlo algorithm to solve the above integration." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "The key distinction between our physically-based foggy scene rendering technique and ASM lies in the modeling of airlight, as represented by Eq. (5). Our approach offers a" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 308, + 191, + 547, + 263 + ], + "blocks": [ + { + "bbox": [ + 308, + 191, + 547, + 263 + ], + "lines": [ + { + "bbox": [ + 308, + 191, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 191, + 547, + 263 + ], + "type": "image", + "image_path": "0bb284ef9b1207033f17d9117b63a43bdbfc1743e83004c8e544cfd5880e2e90.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 273, + 505, + 283 + ], + "lines": [ + { + "bbox": [ + 345, + 273, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 345, + 273, + 505, + 283 + ], + "type": "text", + "content": "Figure 5. Noise flow and sensor simulation." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 296, + 545, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 296, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 305, + 296, + 545, + 355 + ], + "type": "text", + "content": "more accurate depiction of light scattering process through the utilization of specific phase function, and enables incorporation of global illumination and multiple light sources within the scene, producing authentic scene radiance to the camera model described in the next section." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 368, + 536, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 368, + 536, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 368, + 536, + 380 + ], + "type": "text", + "content": "3.2.2 Imaging through Realistic Camera Simulation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 386, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 386, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 305, + 386, + 545, + 470 + ], + "type": "text", + "content": "After obtaining the scene radiance, we employ ISETCam [15-17, 30, 32] to simulate the realistic digital camera system that encompasses optics, sensor and ISP components. Specifically, we use a diffraction-limited optics model [7] with consideration of off-axis irradiance attenuation based on the cosine-fourth law [24] to convert scene radiance " + }, + { + "bbox": [ + 305, + 386, + 545, + 470 + ], + "type": "inline_equation", + "content": "L_{o}" + }, + { + "bbox": [ + 305, + 386, + 545, + 470 + ], + "type": "text", + "content": " to optical image irradiance " + }, + { + "bbox": [ + 305, + 386, + 545, + 470 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 305, + 386, + 545, + 470 + ], + "type": "text", + "content": ", denoted as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 337, + 477, + 545, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 477, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 337, + 477, + 545, + 503 + ], + "type": "interline_equation", + "content": "E _ {i} = \\pi \\cdot L _ {o} \\frac {1}{1 + 4 (f / \\# (1 - m)) ^ {2}} (\\cos \\phi) ^ {4}, \\tag {6}", + "image_path": "57979397963553d95aab9c90ae452dbd65a9adb9e9267510192ceb96a92d1e64.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 510, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 510, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 305, + 510, + 545, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 510, + 545, + 556 + ], + "type": "inline_equation", + "content": "f / \\#" + }, + { + "bbox": [ + 305, + 510, + 545, + 556 + ], + "type": "text", + "content": " is the f-number, " + }, + { + "bbox": [ + 305, + 510, + 545, + 556 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 510, + 545, + 556 + ], + "type": "text", + "content": " represents the magnification defined as the negative ratio of image distance and object distance. In this optics model, blur mainly depends on the f-number value but little else." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 558, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 676 + ], + "type": "text", + "content": "We model the sensor process including photovoltaic conversion, analog amplifier, and analog-to-digital conversion as described in Fig. 4 and simulate the real noise characteristics depicted in Fig. 5. Among the various noise sources, shot noise arising from the photovoltaic conversion emerges as the dominant type of noise in foggy images which follows the Poisson distribution. To ensure the credibility of our sensor simulation, we incorporate a set of calibrated sensor parameters from the Mi10Pro cellphone, encompassing the CFA and infrared filter data." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Finally, we process the RAW image output from the sensor with a simple ISP module, including black level compensation, white balance, demosaicing, color correction and" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21766" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 102, + 282, + 168 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "text", + "content": "Table 1. Comparisons between SynFog and existing synthetic fog datasets in the driving field." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 102, + 282, + 168 + ], + "lines": [ + { + "bbox": [ + 50, + 102, + 282, + 168 + ], + "spans": [ + { + "bbox": [ + 50, + 102, + 282, + 168 + ], + "type": "table", + "html": "
DatasetsGlobal IlluminationOptics and Sensor Characteristics
Foggy Cityscapes××
Virtual KITTI 2××
SynFog (Ours)✓✓
", + "image_path": "418731bc065c0512b2fe6a12ab590ad7a143741a689d3a9fcd5ab45751d9bb17.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 182, + 286, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 182, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 286, + 205 + ], + "type": "text", + "content": "gamma correction, to get the final synthetic image that mimics the real camera characteristics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 218, + 143, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 218, + 143, + 231 + ], + "spans": [ + { + "bbox": [ + 47, + 218, + 143, + 231 + ], + "type": "text", + "content": "4. SynFog Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 238, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 287, + 453 + ], + "type": "text", + "content": "By utilizing the proposed pipeline, we develop a new extensive synthetic fog dataset called SynFog. Example images are shown in Fig. 1. We first generate 500 large-scale, diverse driving scenes using procedural modeling techniques. For each scene, we generate two sets of images under different lighting conditions: one set with sky light only and the other set with both sky light and active light sources, such as street lamps and automotive lighting. Within each group, we create a set of four images: one clear image and three foggy images with varying fog levels defined by the scattering coefficient " + }, + { + "bbox": [ + 46, + 238, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{s}}" + }, + { + "bbox": [ + 46, + 238, + 287, + 453 + ], + "type": "text", + "content": ". According to McCartney's work [33], the scattering coefficient remains independent of the wavelength for fog [13, 34]. Therefore, we use a consistent scattering coefficient for wavelengths within the visible range. To ensure realistic simulation of fog, we employ the meteorological optical range (MOR) [23] or visibility as a reference to select appropriate scattering coefficients, described as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 454, + 287, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 454, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 125, + 454, + 287, + 467 + ], + "type": "interline_equation", + "content": "M O R = 2. 9 9 6 / \\sigma_ {s}. \\tag {7}", + "image_path": "7d81b341c0cfca62aff77e5ac2917f3fba18847c44ef409e4f671236159b3f34.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 474, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 557 + ], + "type": "text", + "content": "Based on this equation, we select scattering coefficient values of 0.005, 0.01, 0.02, which correspond to visibility range of " + }, + { + "bbox": [ + 46, + 474, + 287, + 557 + ], + "type": "inline_equation", + "content": "600\\mathrm{m}" + }, + { + "bbox": [ + 46, + 474, + 287, + 557 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 474, + 287, + 557 + ], + "type": "inline_equation", + "content": "300\\mathrm{m}" + }, + { + "bbox": [ + 46, + 474, + 287, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 474, + 287, + 557 + ], + "type": "inline_equation", + "content": "150\\mathrm{m}" + }, + { + "bbox": [ + 46, + 474, + 287, + 557 + ], + "type": "text", + "content": ", respectively. According to National Standard of China [36], these visibility ranges are classified as heavy fog, thick fog, and dense fog, thereby the simulated foggy images are diversified enough to reflect real-world complicated foggy conditions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": "We compare SynFog with existing fog datasets in the field of autonomous driving, including the extensively used Foggy Cityscapes [40] and Virtual KITTI 2 [18]. The comparison is presented in Tab. 1. In terms of foggy image generation, methods that rely on ASM [34] require accurate depth maps to calculate the correct transmission maps. However, obtaining accurate depth maps for outdoor images can be challenging, leading to artifacts such as holes and discontinuities in synthetic foggy images, similar to what is observed in Foggy Cityscapes [40]. On the other hand, foggy images generated using Game Engines like Unity or Unreal [18] typically employs rasterization techniques in the rendering process. While effective for creating" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "scenes, rasterization techniques may not accurately model light transportation as precisely as path tracing, which is employed in SynFog. Moreover, SynFog takes the realistic imaging process into account, resulting in highly authentic foggy images with realistic physical characteristics. The superior realism of SynFog is further supported by the experimental results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 157, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 545, + 252 + ], + "type": "text", + "content": "Furthermore, our data simulation pipeline offers exceptional flexibility and control, allowing us to generate not only the final RGB images but also other valuable data modalities, including RAW images. This capability opens up avenues for research on defogging algorithms in the RAW domain, which has not been explored due to the lack of paired foggy RAW images and clear RGB images necessary for training deep defogging networks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 265, + 388, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 265, + 388, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 265, + 388, + 277 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 285, + 444, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 444, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 444, + 297 + ], + "type": "text", + "content": "5.1. Fog Chamber Validation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 304, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 532 + ], + "type": "text", + "content": "We establish an indoor fog chamber to validate the fidelity of our fog simulation pipeline. The chamber is constructed as a capped glass cube with dimensions of " + }, + { + "bbox": [ + 304, + 304, + 545, + 532 + ], + "type": "inline_equation", + "content": "1.5m\\times 0.5m\\times" + }, + { + "bbox": [ + 304, + 304, + 545, + 532 + ], + "type": "inline_equation", + "content": "0.5m" + }, + { + "bbox": [ + 304, + 304, + 545, + 532 + ], + "type": "text", + "content": ". To reproduce the dark night environment that has directional light sources in the fog chamber, we position 4 light bars on each side of the front face of the cuboid fog chamber and enclose it with black blackout cloth from the inside. Fog inside the chamber is generated using a 1.7MHz ceramic atomizer that emulates water mist. To ensure uniformity of fog concentration within the chamber, we incorporate dual fans to promote even distribution. A visibility measurement module is designed to quantify the fog density. This module is comprised of a PM-160 wireless power meter from Thorlabs, an 852 narrowband laser, an aperture and a narrowband filter adapted to the laser band. The aperture serves to filter out stray light and the narrowband filter plays the role in blocking interference caused by non-laser light sources in the scene. The principle of visibility (scattering coefficient " + }, + { + "bbox": [ + 304, + 304, + 545, + 532 + ], + "type": "inline_equation", + "content": "\\sigma_{s}" + }, + { + "bbox": [ + 304, + 304, + 545, + 532 + ], + "type": "text", + "content": ") measurement is expressed as [37]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 392, + 540, + 545, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 540, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 392, + 540, + 545, + 572 + ], + "type": "interline_equation", + "content": "\\sigma_ {s} = \\frac {\\ln \\left(\\frac {P _ {0}}{P _ {u}}\\right)}{\\mathrm {u}}, \\tag {8}", + "image_path": "d0986159bfe751dcc101c8cf0e42a28fefcce39dfd4711e67ba45353ae82d7a8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 581, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 581, + 545, + 616 + ], + "type": "inline_equation", + "content": "P_0" + }, + { + "bbox": [ + 304, + 581, + 545, + 616 + ], + "type": "text", + "content": " is the laser power measured without fog and " + }, + { + "bbox": [ + 304, + 581, + 545, + 616 + ], + "type": "inline_equation", + "content": "P_u" + }, + { + "bbox": [ + 304, + 581, + 545, + 616 + ], + "type": "text", + "content": " is the measured laser power after the propagation of " + }, + { + "bbox": [ + 304, + 581, + 545, + 616 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 581, + 545, + 616 + ], + "type": "text", + "content": " distance under foggy conditions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "We then use Blender to create a virtual fog chamber that can reproduce the real-world counterpart. Our process involves calibrating the light conditions, object information and camera properties prior to capture, resulting in a one-to-one reproduction of the real fog chamber. We select the standard Macbeth color checker as our test object and use Mi10Pro cellphone to capture images. We evaluate the matching degree between the simulation and reality un" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21767" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 289, + 218 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 289, + 218 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 289, + 218 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 289, + 218 + ], + "type": "image", + "image_path": "c81e9100ee41046608c476a51ec4d25e06c83552ef5149f70d7059835f36f1b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 221, + 173, + 231 + ], + "lines": [ + { + "bbox": [ + 163, + 221, + 173, + 231 + ], + "spans": [ + { + "bbox": [ + 163, + 221, + 173, + 231 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 237, + 286, + 348 + ], + "blocks": [ + { + "bbox": [ + 52, + 237, + 286, + 348 + ], + "lines": [ + { + "bbox": [ + 52, + 237, + 286, + 348 + ], + "spans": [ + { + "bbox": [ + 52, + 237, + 286, + 348 + ], + "type": "image", + "image_path": "36ff337738896c144f5d81762896a635365349fab6bfa047015b55d8e3a8ec4a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 357, + 287, + 413 + ], + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 413 + ], + "type": "text", + "content": "Figure 6. Color analysis under foggy conditions. (a) Visual comparison between real captures and simulated images under different fog densities in linear RGB space. We increase the brightness of the images " + }, + { + "bbox": [ + 46, + 357, + 287, + 413 + ], + "type": "inline_equation", + "content": "(\\times 2)" + }, + { + "bbox": [ + 46, + 357, + 287, + 413 + ], + "type": "text", + "content": " for visual purpose. (b) Analyzing the trend of color variation across different fog densities." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 423, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 448 + ], + "type": "text", + "content": "der clear conditions by conducting a color analysis on the checker, which is available in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": "After ensuring the consistency of initial capture conditions, we proceed to capture images under foggy conditions. We first record the laser power under clear condition as " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "P_0" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": ", and then proceed to generate fog until the object is no longer visible. During the fog dispersion process, we capture images at regular intervals of 1 second and record the laser energy " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "P_u" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": " to obtain real-time fog density measurements. We utilize the measured fog density parameter " + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\sigma_s" + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": " to simulate corresponding foggy counterparts based on the proposed foggy image simulation pipeline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "In Fig. 6a, we showcase both the real-captured foggy images and simulated ones generated using our pipeline and the ASM [33]. It can be seen that images simulated through our pipeline exhibit a closer resemblance to real-captured foggy images. We also conduct a color analysis on 24 color patches under different fog densities as depicted in Fig. 6b. The results demonstrate that, as fog density increases, the trend of each channel in simulated images generated by our pipeline consistently follows that of the real captures. In contrast, the simulated images based on ASM do not exhibit this trend. These findings provide a quantitative validation of the realism of our foggy image simulation method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 85 + ], + "type": "text", + "content": "5.2. Transferability across the Real-to-Virtual Gap" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 90, + 545, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 245 + ], + "type": "text", + "content": "We evaluate 7 representative state-of-the-art defogging algorithms on SynFog dataset: DCP [20], CAP [57], Grid-DehazeNet [28], MSBDN [12], FFA-Net [39], AECR-Net [50], and DehazeFormer [43]. We directly apply the models trained on SynFog to real foggy images sourced from Foggy Zurich [11] and Seeing Through Fog [5] datasets, without undergoing any model fine-tuning or domain adaptation. During the training process, we adhere to the official settings of these methods, and all of them demonstrate convergence by the end of the training. The results are shown in Fig. 7. Despite DCP [20] and CAP [57] producing excessively dark outputs, models trained on SynFog can produce natural defogged images with minimal artifacts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 246, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 545, + 365 + ], + "type": "text", + "content": "To underscore the superior realism of SynFog compared to existing synthetic fog datasets, we further enhance our evaluation by training the top-performing models (AECR-Net [50], DehazeFormer [43]) on two prominent synthetic fog datasets relevant to autonomous driving (Foggy Cityscapes [40], Virtual KITTI [18]). Subsequently, these models are evaluated on real fog datasets for a comparative analysis. Additionally, we utilize ASM [34] to create a SynFog-" + }, + { + "bbox": [ + 304, + 246, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 246, + 545, + 365 + ], + "type": "text", + "content": " dataset with clear images from SynFog and test it using DehazeFormer [43]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 366, + 546, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 546, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 546, + 593 + ], + "type": "text", + "content": "Comparison with existing synthetic fog datasets. To mitigate any potential impact arising from model training issues, we initially conduct tests for each model using the corresponding test set of the training data. Test results are available in the supplementary material. Both AECR-Net [50] and DehazeFormer [43] exhibit high performance, confirming the correctness of the model training process. Subsequently, the trained models are directly applied to real-world test foggy images. Results are shown in Fig. 9 and Tab. 2. Owing to the high authenticity of the SynFog dataset, models trained on it can effectively generalize to real foggy images and produce naturally colored defogged images. Conversely, models trained on the other two datasets display poorer generalization to real data, resulting in artifacts and color distortion in defogged images. We also evaluate the defogged images using a pre-trained YOLOv8 [22] model for detection, showcasing how SynFog enhances machine vision capabilities (See Tab. 3). More results can be found in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 676 + ], + "type": "text", + "content": "It is noteworthy that the training set of Foggy Cityscapes [40] comprises 8925 data pairs, and Virtual KITTI [18] includes 3452 data pairs, both exceeding the 1350 data pairs in SynFog. However, models trained on SynFog demonstrate superior generalization capabilities on real-world data compared to them, highlighting the crucial role of data quality for researchers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Comparison with SynFog- " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " synthesized by ASM. With the ability to obtain pixel-accurate depth information for each image, we can eliminate the impact of inaccurate" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21768" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 85, + 535, + 149 + ], + "blocks": [ + { + "bbox": [ + 161, + 70, + 432, + 82 + ], + "lines": [ + { + "bbox": [ + 161, + 70, + 432, + 82 + ], + "spans": [ + { + "bbox": [ + 161, + 70, + 432, + 82 + ], + "type": "text", + "content": "Table 2. Defogging evaluations with AECRNet on real-world fog datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 85, + 535, + 149 + ], + "lines": [ + { + "bbox": [ + 58, + 85, + 535, + 149 + ], + "spans": [ + { + "bbox": [ + 58, + 85, + 535, + 149 + ], + "type": "table", + "html": "
Training SetO-Haze [2]Foggy Zurich [11]Foggy Driving [40]BeDDE [56]
PSNR ↑SSIM ↑DHQI [14] ↑DHQI [14] ↑DHQI [14] ↑DHQI [14] ↑
Foggy Cityscapes14.460.573743.4052.0651.5536.07
Virtual KITTI13.900.531542.8050.9447.4633.42
SynFog15.430.611644.4654.1652.0743.28
", + "image_path": "2d9813091332f8ef1b3240e58ad25bd87afeddbf364c97e3f6c1039f7bcd2686.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 60, + 156, + 177, + 222 + ], + "blocks": [ + { + "bbox": [ + 60, + 156, + 177, + 222 + ], + "lines": [ + { + "bbox": [ + 60, + 156, + 177, + 222 + ], + "spans": [ + { + "bbox": [ + 60, + 156, + 177, + 222 + ], + "type": "image", + "image_path": "b1f5812d7c12de482978a6d936c91211746e0aaf2e213207e94c07117207ccc4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 179, + 156, + 294, + 222 + ], + "blocks": [ + { + "bbox": [ + 179, + 156, + 294, + 222 + ], + "lines": [ + { + "bbox": [ + 179, + 156, + 294, + 222 + ], + "spans": [ + { + "bbox": [ + 179, + 156, + 294, + 222 + ], + "type": "image", + "image_path": "126cb07242ae6d47cf3790a8c695d911a9a8fea06ed4c553736c30aab29ee76d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 296, + 156, + 413, + 222 + ], + "blocks": [ + { + "bbox": [ + 296, + 156, + 413, + 222 + ], + "lines": [ + { + "bbox": [ + 296, + 156, + 413, + 222 + ], + "spans": [ + { + "bbox": [ + 296, + 156, + 413, + 222 + ], + "type": "image", + "image_path": "10e1190dce571374cd3edcdb719870ab82d265ec7c2d06a95ff0acbd8f2d05d1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 414, + 156, + 531, + 222 + ], + "blocks": [ + { + "bbox": [ + 414, + 156, + 531, + 222 + ], + "lines": [ + { + "bbox": [ + 414, + 156, + 531, + 222 + ], + "spans": [ + { + "bbox": [ + 414, + 156, + 531, + 222 + ], + "type": "image", + "image_path": "f0756365c306c6245fecd1237ea401ea9b08a33a0a0d9c36e3539f915cc2c0c6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 59, + 232, + 177, + 298 + ], + "blocks": [ + { + "bbox": [ + 110, + 224, + 126, + 232 + ], + "lines": [ + { + "bbox": [ + 110, + 224, + 126, + 232 + ], + "spans": [ + { + "bbox": [ + 110, + 224, + 126, + 232 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 232, + 177, + 298 + ], + "lines": [ + { + "bbox": [ + 59, + 232, + 177, + 298 + ], + "spans": [ + { + "bbox": [ + 59, + 232, + 177, + 298 + ], + "type": "image", + "image_path": "6db9bf03d46b68b982a95663e4d27900ac73d9cee60428f91afed4f2a037f6fa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 299, + 140, + 308 + ], + "lines": [ + { + "bbox": [ + 99, + 299, + 140, + 308 + ], + "spans": [ + { + "bbox": [ + 99, + 299, + 140, + 308 + ], + "type": "text", + "content": "MSBDN [12]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 179, + 232, + 294, + 298 + ], + "blocks": [ + { + "bbox": [ + 223, + 223, + 251, + 232 + ], + "lines": [ + { + "bbox": [ + 223, + 223, + 251, + 232 + ], + "spans": [ + { + "bbox": [ + 223, + 223, + 251, + 232 + ], + "type": "text", + "content": "DCP [20]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 179, + 232, + 294, + 298 + ], + "lines": [ + { + "bbox": [ + 179, + 232, + 294, + 298 + ], + "spans": [ + { + "bbox": [ + 179, + 232, + 294, + 298 + ], + "type": "image", + "image_path": "4a55c76112416daaba3cc54d9256c52f43eac2eb55e2ccdbb63cd2a1c656758a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 299, + 258, + 308 + ], + "lines": [ + { + "bbox": [ + 217, + 299, + 258, + 308 + ], + "spans": [ + { + "bbox": [ + 217, + 299, + 258, + 308 + ], + "type": "text", + "content": "FFA-Net [39]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 296, + 232, + 413, + 298 + ], + "blocks": [ + { + "bbox": [ + 340, + 223, + 369, + 232 + ], + "lines": [ + { + "bbox": [ + 340, + 223, + 369, + 232 + ], + "spans": [ + { + "bbox": [ + 340, + 223, + 369, + 232 + ], + "type": "text", + "content": "CAP [57]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 232, + 413, + 298 + ], + "lines": [ + { + "bbox": [ + 296, + 232, + 413, + 298 + ], + "spans": [ + { + "bbox": [ + 296, + 232, + 413, + 298 + ], + "type": "image", + "image_path": "897d10a502d991a9126b94861160f0e1e4bc21e7dfb3980962aca42d1af80fdb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 299, + 378, + 308 + ], + "lines": [ + { + "bbox": [ + 331, + 299, + 378, + 308 + ], + "spans": [ + { + "bbox": [ + 331, + 299, + 378, + 308 + ], + "type": "text", + "content": "AECR-Net [50]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 414, + 232, + 531, + 298 + ], + "blocks": [ + { + "bbox": [ + 443, + 223, + 503, + 232 + ], + "lines": [ + { + "bbox": [ + 443, + 223, + 503, + 232 + ], + "spans": [ + { + "bbox": [ + 443, + 223, + 503, + 232 + ], + "type": "text", + "content": "GridDehazeNet [28]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 414, + 232, + 531, + 298 + ], + "lines": [ + { + "bbox": [ + 414, + 232, + 531, + 298 + ], + "spans": [ + { + "bbox": [ + 414, + 232, + 531, + 298 + ], + "type": "image", + "image_path": "a9fd4e27f47dc4bb3b876ff7c1d8bdcb239cba1982ff2d99dc728e7ba7301ec2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 299, + 501, + 308 + ], + "lines": [ + { + "bbox": [ + 443, + 299, + 501, + 308 + ], + "spans": [ + { + "bbox": [ + 443, + 299, + 501, + 308 + ], + "type": "text", + "content": "DehazeFormer [43]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 48, + 333, + 288, + 369 + ], + "blocks": [ + { + "bbox": [ + 118, + 316, + 473, + 327 + ], + "lines": [ + { + "bbox": [ + 118, + 316, + 473, + 327 + ], + "spans": [ + { + "bbox": [ + 118, + 316, + 473, + 327 + ], + "type": "text", + "content": "Figure 7. Transferability across real-world fog datasets. Test results of models trained on SynFog." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 333, + 288, + 369 + ], + "lines": [ + { + "bbox": [ + 48, + 333, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 288, + 369 + ], + "type": "image", + "image_path": "515484acf83bc484a7a14c0ca17a927119d75ae605e026bbce7219533cbde948.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 375, + 251, + 385 + ], + "lines": [ + { + "bbox": [ + 83, + 375, + 251, + 385 + ], + "spans": [ + { + "bbox": [ + 83, + 375, + 251, + 385 + ], + "type": "text", + "content": "Figure 8. Ablation of sensor noise simulation." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 49, + 419, + 286, + 521 + ], + "blocks": [ + { + "bbox": [ + 46, + 393, + 287, + 415 + ], + "lines": [ + { + "bbox": [ + 46, + 393, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 287, + 415 + ], + "type": "text", + "content": "Table 3. Detection evaluations on real-world fog datasets. \"FZ\" refers to Foggy Zurich. \"STF\" represents Seeing Through Fog." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 419, + 286, + 521 + ], + "lines": [ + { + "bbox": [ + 49, + 419, + 286, + 521 + ], + "spans": [ + { + "bbox": [ + 49, + 419, + 286, + 521 + ], + "type": "table", + "html": "
MethodTraining SetFZ [11]STF [5]
mAP (%)mAP (%)
AECRNetFoggy Cityscapes69.754.8
Virtual KITTI68.953.3
SynFog71.555.5
DehazeFormerFoggy Cityscapes67.954.9
Virtual KITTI68.553.1
SynFog-β59.755.3
SynFog69.755.3
", + "image_path": "81fc69b965ae4f30773a69e11d6a7be90968653c5ebb42aa40da06a2e9e8f2e3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 530, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 287, + 662 + ], + "type": "text", + "content": "depth and focus solely on the fog simulation method itself. The foggy image is synthesized from its clear counterpart in SynFog according to Eq. (1). Specifically, we employ the same airlight estimation method used in the construction of Foggy Cityscapes [40], which involves computing the median of all the " + }, + { + "bbox": [ + 46, + 530, + 287, + 662 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 46, + 530, + 287, + 662 + ], + "type": "text", + "content": " pixels with the largest dark channel values, as proposed in [45]. We use DehazeFormer [43] to train on both SynFog and SynFog- " + }, + { + "bbox": [ + 46, + 530, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 530, + 287, + 662 + ], + "type": "text", + "content": " , followed by testing the trained model on real-world foggy images. As shown in Fig. 9, image output by model trained on SynFog exhibits a better visual appearance compared to that of SynFog- " + }, + { + "bbox": [ + 46, + 530, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\beta" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 670, + 216, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 216, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 216, + 683 + ], + "type": "text", + "content": "5.3. Ablation Study of Sensor Noise" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "To demonstrate the importance of sensor noise simulation in our pipeline, we create SynFog- " + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": " dataset. During the" + } + ] + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 307, + 348, + 554, + 396 + ], + "blocks": [ + { + "bbox": [ + 318, + 334, + 533, + 346 + ], + "lines": [ + { + "bbox": [ + 318, + 334, + 533, + 346 + ], + "spans": [ + { + "bbox": [ + 318, + 334, + 533, + 346 + ], + "type": "text", + "content": "Table 4. Impact of sensor noise on defogging performance." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 348, + 554, + 396 + ], + "lines": [ + { + "bbox": [ + 307, + 348, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 348, + 554, + 396 + ], + "type": "table", + "html": "
Experimental settingO-Haze [2]
PSNR ↑SSIM ↑DHQI ↑
AECRNet+SynFog(w/o noise)15.160.579537.53
AECRNet+SynFog(w/ noise)15.430.611644.46
", + "image_path": "75844c7d8da63cc82c2f6a7525d4b618fc92bf10940ce8412e21833bb043ba98.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_body" + } + ], + "index": 27 + }, + { + "type": "table", + "bbox": [ + 316, + 429, + 535, + 477 + ], + "blocks": [ + { + "bbox": [ + 306, + 403, + 545, + 426 + ], + "lines": [ + { + "bbox": [ + 306, + 403, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 545, + 426 + ], + "type": "text", + "content": "Table 5. Impact of sensor noise on detection performance. \"FZ\" refers to Foggy Zurich. \"STF\" represents Seeing Through Fog." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 429, + 535, + 477 + ], + "lines": [ + { + "bbox": [ + 316, + 429, + 535, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 429, + 535, + 477 + ], + "type": "table", + "html": "
Experimental settingFZ [11]STF [5]
mAP (%)mAP (%)
AECRNet+SynFog(w/o noise)69.554.7
AECRNet+SynFog(w/ noise)71.555.5
", + "image_path": "4f2dab9c2e31615eb6c3a919d70cc94867f4a94311bee688fa82498b6281bd1f.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 484, + 545, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 580 + ], + "type": "text", + "content": "data generation process, the sensor-induced noise is set to zero, while all other settings remain consistent with those of the original SynFog dataset. We compare the defogging and detection performance of models trained on SynFog and SynFog- " + }, + { + "bbox": [ + 304, + 484, + 545, + 580 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 484, + 545, + 580 + ], + "type": "text", + "content": " datasets, as shown in Fig. 8 and Tab. 4-5. The model trained on SynFog that incorporates sensor noise produces more natural results with fewer artifacts and higher detection accuracy." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 586, + 450, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 450, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 450, + 599 + ], + "type": "text", + "content": "5.4. Algorithm Benchmarking" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 605, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 689 + ], + "type": "text", + "content": "We use SynFog test set to assess the aforementioned representative defogging methods, employing both full-reference and no-reference metrics to evaluate the defogged results. To illustrate the challenges presented by SynFog, we also provide evaluation results from the widely-used RESIDE dataset [25]. The full-scale quantitative comparison results are available in the supplementary material." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Generally, learning-based methods consistently outperform earlier algorithms based on natural or statistical priors" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21769" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 151, + 180 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 151, + 180 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 151, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 151, + 180 + ], + "type": "image", + "image_path": "e07ca0b11678860c79835939bbf540d43c3deca8b839c233bd62950de2c71362.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 182, + 110, + 190 + ], + "lines": [ + { + "bbox": [ + 94, + 182, + 110, + 190 + ], + "spans": [ + { + "bbox": [ + 94, + 182, + 110, + 190 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 151, + 70, + 248, + 181 + ], + "blocks": [ + { + "bbox": [ + 151, + 70, + 248, + 181 + ], + "lines": [ + { + "bbox": [ + 151, + 70, + 248, + 181 + ], + "spans": [ + { + "bbox": [ + 151, + 70, + 248, + 181 + ], + "type": "image", + "image_path": "4c5c2cd1ee05d201650cec21122bf3748f4a806a7c6dc809b08ec2dd3cca7b52.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 182, + 231, + 191 + ], + "lines": [ + { + "bbox": [ + 168, + 182, + 231, + 191 + ], + "spans": [ + { + "bbox": [ + 168, + 182, + 231, + 191 + ], + "type": "text", + "content": "Foggy Cityscapes [40]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 71, + 345, + 181 + ], + "blocks": [ + { + "bbox": [ + 249, + 71, + 345, + 181 + ], + "lines": [ + { + "bbox": [ + 249, + 71, + 345, + 181 + ], + "spans": [ + { + "bbox": [ + 249, + 71, + 345, + 181 + ], + "type": "image", + "image_path": "4997464cd8e9ce6b3a37cac63b0fd43b0009085fb91597498fef1b0d1c38cff1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 182, + 323, + 191 + ], + "lines": [ + { + "bbox": [ + 271, + 182, + 323, + 191 + ], + "spans": [ + { + "bbox": [ + 271, + 182, + 323, + 191 + ], + "type": "text", + "content": "Virtual KITTI [18]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 346, + 71, + 442, + 181 + ], + "blocks": [ + { + "bbox": [ + 346, + 71, + 442, + 181 + ], + "lines": [ + { + "bbox": [ + 346, + 71, + 442, + 181 + ], + "spans": [ + { + "bbox": [ + 346, + 71, + 442, + 181 + ], + "type": "image", + "image_path": "fdba5094e0c1b731c62ccd2b5f282dbd6aaffd1363bb521d28053f4e9c24d86a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 182, + 408, + 191 + ], + "lines": [ + { + "bbox": [ + 380, + 182, + 408, + 191 + ], + "spans": [ + { + "bbox": [ + 380, + 182, + 408, + 191 + ], + "type": "text", + "content": "SynFog-β" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 443, + 71, + 540, + 181 + ], + "blocks": [ + { + "bbox": [ + 443, + 71, + 540, + 181 + ], + "lines": [ + { + "bbox": [ + 443, + 71, + 540, + 181 + ], + "spans": [ + { + "bbox": [ + 443, + 71, + 540, + 181 + ], + "type": "image", + "image_path": "94829d71ccfa1bb7c332a075e0a98a297a365025a39765727253e536c8eda625.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 182, + 512, + 191 + ], + "lines": [ + { + "bbox": [ + 470, + 182, + 512, + 191 + ], + "spans": [ + { + "bbox": [ + 470, + 182, + 512, + 191 + ], + "type": "text", + "content": "SynFog (Ours)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 53, + 234, + 108, + 371 + ], + "blocks": [ + { + "bbox": [ + 46, + 204, + 545, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 545, + 226 + ], + "type": "text", + "content": "Figure 9. Visual and detection comparisons on real-world foggy images using DehazeFormer models trained on SynFog, SynFog- " + }, + { + "bbox": [ + 46, + 204, + 545, + 226 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 204, + 545, + 226 + ], + "type": "text", + "content": ", Foggy Cityscapes [40] and Virtual KITTI [18]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 234, + 108, + 371 + ], + "lines": [ + { + "bbox": [ + 53, + 234, + 108, + 371 + ], + "spans": [ + { + "bbox": [ + 53, + 234, + 108, + 371 + ], + "type": "image", + "image_path": "116bde5ef6776d2eb1f4ac1357261707c837680ad51c734e1eb86a4bb38ad9d9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 108, + 235, + 162, + 371 + ], + "blocks": [ + { + "bbox": [ + 108, + 235, + 162, + 371 + ], + "lines": [ + { + "bbox": [ + 108, + 235, + 162, + 371 + ], + "spans": [ + { + "bbox": [ + 108, + 235, + 162, + 371 + ], + "type": "image", + "image_path": "cd12f59177431e7fb1b49ca9b57a17c01cc7913056c61742ff142a7b53984d87.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 382, + 538, + 394 + ], + "lines": [ + { + "bbox": [ + 53, + 382, + 538, + 394 + ], + "spans": [ + { + "bbox": [ + 53, + 382, + 538, + 394 + ], + "type": "text", + "content": "Figure 10. Qualitative comparisons of SOTA defogging methods on SynFog dataset with different fog levels. Zoom in for better view." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 162, + 235, + 215, + 371 + ], + "blocks": [ + { + "bbox": [ + 162, + 235, + 215, + 371 + ], + "lines": [ + { + "bbox": [ + 162, + 235, + 215, + 371 + ], + "spans": [ + { + "bbox": [ + 162, + 235, + 215, + 371 + ], + "type": "image", + "image_path": "69bf9067694c72d7790432942b68d68295dd3f305ddf59c4984c8cdf7c2c6222.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 216, + 235, + 269, + 371 + ], + "blocks": [ + { + "bbox": [ + 216, + 235, + 269, + 371 + ], + "lines": [ + { + "bbox": [ + 216, + 235, + 269, + 371 + ], + "spans": [ + { + "bbox": [ + 216, + 235, + 269, + 371 + ], + "type": "image", + "image_path": "5ab8230ca5c33c9d8b8bad43cd490f928f71e3dbb9adb5c99b19c84771782179.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 270, + 235, + 323, + 371 + ], + "blocks": [ + { + "bbox": [ + 270, + 235, + 323, + 371 + ], + "lines": [ + { + "bbox": [ + 270, + 235, + 323, + 371 + ], + "spans": [ + { + "bbox": [ + 270, + 235, + 323, + 371 + ], + "type": "image", + "image_path": "5f1c4b5bc22508ce06538c71a12e8dba023c246d35663e717f8807ec49582fd3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 324, + 235, + 377, + 371 + ], + "blocks": [ + { + "bbox": [ + 324, + 235, + 377, + 371 + ], + "lines": [ + { + "bbox": [ + 324, + 235, + 377, + 371 + ], + "spans": [ + { + "bbox": [ + 324, + 235, + 377, + 371 + ], + "type": "image", + "image_path": "cffffc0a4cf8cd4aa56d546671d0c40aa1c05ace67aecea7326ccb43a6561ede.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 378, + 235, + 432, + 371 + ], + "blocks": [ + { + "bbox": [ + 378, + 235, + 432, + 371 + ], + "lines": [ + { + "bbox": [ + 378, + 235, + 432, + 371 + ], + "spans": [ + { + "bbox": [ + 378, + 235, + 432, + 371 + ], + "type": "image", + "image_path": "e4c33cea8ea00fe764c834af25cde06857646179ee7679837bb8cb3a525c0e3c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 433, + 235, + 485, + 371 + ], + "blocks": [ + { + "bbox": [ + 433, + 235, + 485, + 371 + ], + "lines": [ + { + "bbox": [ + 433, + 235, + 485, + 371 + ], + "spans": [ + { + "bbox": [ + 433, + 235, + 485, + 371 + ], + "type": "image", + "image_path": "204d86f1104f4224248e1368d9d5cbd9ef2b9168f4bcbf77c9eb84b50f023896.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 486, + 235, + 540, + 371 + ], + "blocks": [ + { + "bbox": [ + 486, + 235, + 540, + 371 + ], + "lines": [ + { + "bbox": [ + 486, + 235, + 540, + 371 + ], + "spans": [ + { + "bbox": [ + 486, + 235, + 540, + 371 + ], + "type": "image", + "image_path": "e97b2a3b5d69573f987a245210488ab488a0de8f144e894c9796becfb9f29f2e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 400, + 289, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 289, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 289, + 605 + ], + "type": "text", + "content": "in terms of PSNR, SSIM and LPIPS [54]. In particular, DehazeFormer [43] attains the highest PSNR (27.96) and SSIM (0.923) values, while GridDehazeNet [28] achieves the best LPIPS (0.163) value on SynFog test set. Although the overall performance trend of the algorithms on SynFog is similar to that on RESIDE [25], the evaluation scores are noticeably lower, highlighting the challenges posed by SynFog. Fig. 10 depicts the qualitative comparison among these methods across different fog levels. Images restored by DCP [20] and CAP [57] tend to appear excessively dark and noisy due to inaccurate airlight estimation and the lack of effective denoising capabilities. GridDehazeNet [28] lacks color recovery ability in dense fog, while the outputs of FFA-Net [39] and DehazeFormer [43] seem smooth. This comparison further indicates the substantial room for improvement in existing defogging methods, especially when applied to more realistic fog datasets." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 611, + 197, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 197, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 197, + 623 + ], + "type": "text", + "content": "5.5. Discussion and Limitations" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "Experimental results reveal that models trained on SynFog, incorporating realistic physical characteristics of optics and sensors, exhibit a more robust performance on real-word fog datasets in terms of visual perception and detection accuracy. This outcome implies that integrating a physically-based end-to-end imaging process replicating real capture methods can significantly improve the generalization of" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 401, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 473 + ], + "type": "text", + "content": "trained models. However, given the time-consuming nature of the physically-based volumetric path tracing technique, the scale of SynFog is not as extensive as other synthetic fog datasets. Nonetheless, we are committed to expanding the scene contents in the future and exploring a broader range of scattering mediums beyond fog." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 479, + 378, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 378, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 378, + 491 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 500, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 545, + 620 + ], + "type": "text", + "content": "In this paper, we present an end-to-end foggy image simulation pipeline. Our approach utilizes volumetric path tracing to model a more precise light scattering process with global illumination. By incorporating a physically-based camera processing pipeline that includes optics, sensor and image processing, we can closely mimic the authentic capture process under foggy conditions. Additionally, we develop a new synthetic fog dataset, SynFog, to facilitate the research on defogging. Comprehensive experiments have validated the authenticity and reliability of the SynFog dataset." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 624, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 715 + ], + "type": "text", + "content": "Acknowledgements This work is supported by 1) National Natural Science Foundation of China (Grant No. 62131011 and 62375233); 2) NSFC Young Scientists Fund (Grant No. 62302423). We thank Professor Chuxi Yang and Jiayue Xie for their help in preparing the scene creation. We also thank Professor Brian Wandell, Dr. Devesh Upadhyay, Dr. Alireza Rahimpour, and Devesh Shah for their valuable advice at the early stage of this project." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21770" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 94, + 107, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 94, + 107, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 94, + 107, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 114, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 114, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 287, + 158 + ], + "type": "text", + "content": "[1] Cosmin Ancuti, Codruta O Ancuti, and Christophe De Vleeschouwer. D-hazy: A dataset to evaluate quantitatively dehazing algorithms. In IEEE International Conference on Image Processing, pages 2226-2230, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 213 + ], + "type": "text", + "content": "[2] Codruta O Ancuti, Cosmin Ancuti, Radu Timofte, and Christophe De Vleeschouwer. O-haze: a dehazing benchmark with real hazy and haze-free outdoor images. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 754-762, 2018. 1, 2, 4, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 214, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 258 + ], + "type": "text", + "content": "[3] Codruta O Ancuti, Cosmin Ancuti, Mateu Sbert, and Radu Timofte. Dense-haze: A benchmark for image dehazing with dense-haze and haze-free images. In IEEE International Conference on Image Processing, pages 1014-1018, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 259, + 287, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 259, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 259, + 287, + 313 + ], + "type": "text", + "content": "[4] Codruta O Ancuti, Cosmin Ancuti, and Radu Timofte. Nh-haze: An image dehazing benchmark with nonhomogeneous hazy and haze-free images. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 444-445, 2020. 1, 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 315, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 315, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 315, + 287, + 379 + ], + "type": "text", + "content": "[5] Mario Bijelic, Tobias Gruber, Fahim Mannan, Florian Kraus, Werner Ritter, Klaus Dietmayer, and Felix Heide. Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In IEEE Conference on Computer Vision and Pattern Recognition, pages 11682-11692, 2020. 1, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 380, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 424 + ], + "type": "text", + "content": "[6] Norbert Buch, Sergio A Velastin, and James Orwell. A review of computer vision techniques for the analysis of urban traffic. IEEE Transactions on Intelligent Transportation Systems, 12(3):920-939, 2011. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 425, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 287, + 446 + ], + "type": "text", + "content": "[7] Ting Chen. Digital camera system simulator and applications. Stanford University, 2003. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "type": "text", + "content": "[8] Lark Kwon Choi, Jaehee You, and Alan Conrad Bovik. Referenceless prediction of perceptual fog density and perceptual image defogging. IEEE Transactions on Image Processing, 24(11):3888-3901, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 492, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 492, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 492, + 287, + 557 + ], + "type": "text", + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3213-3223, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 601 + ], + "type": "text", + "content": "[10] Dengxin Dai and Wen Yang. Satellite image classification via two-layer sparse coding with biased image representation. IEEE Geoscience and Remote Sensing Letters, 8(1): 173-176, 2010. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "text", + "content": "[11] Dengxin Dai, Christos Sakaridis, Simon Hecker, and Luc Van Gool. Curriculum model adaptation with synthetic and real data for semantic foggy scene understanding. International Journal of Computer Vision, 128:1182-1204, 2020. 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming-Hsuan Yang. Multi-scale boosted dehazing network with dense feature fusion. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2157-2167, 2020. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[13] J.B. Farrow et al. Influence of the atmosphere on optical systems. Optica Acta, 1970. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 95, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 116 + ], + "type": "text", + "content": "[14] X. Min et al. Objective quality evaluation of dehazed images. TITS, 2018. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "type": "text", + "content": "[15] Joyce E Farrell and Brian A Wandell. Image systems simulation. Handbook of Digital Imaging, 1:373-400, 2015. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "text", + "content": "[16] Joyce E Farrell, Feng Xiao, Peter Bert Catrysse, and Brian A Wandell. A simulation tool for evaluating digital camera image quality. In Image Quality and System Performance, pages 124-131. SPIE, 2003." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 216 + ], + "type": "text", + "content": "[17] Joyce E Farrell, Peter B Catrysse, and Brian A Wandell. Digital camera simulation. Applied Optics, 51(4):A80-A90, 2012. 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 217, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 261 + ], + "type": "text", + "content": "[18] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4340-4349, 2016. 2, 5, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "type": "text", + "content": "[19] Martin Hahner, Dengxin Dai, Christos Sakaridis, Jan-Nico Zaech, and Luc Van Gool. Semantic understanding of foggy scenes with purely synthetic data. In IEEE Intelligent Transportation Systems Conference (ITSC), pages 3675-3681, 2019. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "type": "text", + "content": "[20] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(12):2341-2353, 2010. 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 360, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 392 + ], + "type": "text", + "content": "[21] Louis G Henyey and Jesse Leonard Greenstein. Diffuse radiation in the galaxy. Astrophysical Journal, vol. 93, p. 70-83 (1941), 93:70-83, 1941. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 393, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 415 + ], + "type": "text", + "content": "[22] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolo, 2023. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 416, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 459 + ], + "type": "text", + "content": "[23] Kyung Won Kim. The comparison of visibility measurement between image-based visual range, human eye-based visual range, and meteorological optical range. Atmospheric Environment, 190:74-86, 2018. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 459, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 459, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 459, + 545, + 480 + ], + "type": "text", + "content": "[24] Miles V Klein and Thomas E Furtak. Optik. Springer-Verlag, 2013. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "text", + "content": "[25] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2018. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 526, + 545, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 559 + ], + "type": "text", + "content": "[26] Ruoteng Li, Xiaoyi Zhang, Shaodi You, and Yu Li. Realistic large-scale fine-depth dehazing dataset from 3d videos. arXiv preprint arXiv:2004.08554, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "text", + "content": "[27] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(10):2024-2039, 2015. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "text", + "content": "[28] Xiaohong Liu, Yongrui Ma, Zhihao Shi, and Jun Chen. Griddehazenet: Attention-based multi-scale network for image dehazing. In IEEE International Conference on Computer Vision, pages 7314-7323, 2019. 6, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[29] Yun Liu, Zhongsheng Yan, Aimin Wu, Tian Ye, and Yuche Li. Nighttime image dehazing based on variational decomposition model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 640-649, 2022. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21771" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[30] Zhenyi Liu, Trisha Lian, Joyce Farrell, and Brian Wandell. Soft prototyping camera designs for car detection based on a convolutional neural network. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "text", + "content": "[31] Zhenyi Liu, Minghao Shen, Jiaqi Zhang, Shuangting Liu, Henryk Blasinski, Trisha Lian, and Brian Wandell. A system for generating complex physically accurate sensor images for automotive applications. arXiv preprint arXiv:1902.04258, 2019. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "type": "text", + "content": "[32] Zheng Lyu, Thomas Goossens, Brian A Wandell, and Joyce Farrell. Validation of physics-based image systems simulation with 3-d scenes. IEEE Sensors Journal, 22(20):19400-19410, 2022. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 229, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 229, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 229, + 288, + 251 + ], + "type": "text", + "content": "[33] Earl J McCartney. Optics of the atmosphere: scattering by molecules and particles. New York, 1976. 1, 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 288, + 285 + ], + "type": "text", + "content": "[34] Srinivasa G Narasimhan and Shree K Nayar. Vision and the atmosphere. International Journal of Computer Vision, 48: 233-254, 2002. 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 288, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 352 + ], + "type": "text", + "content": "[35] Priya Narayanan, Xin Hu, Zhenyu Wu, Matthew D Thielke, John G Rogers, Andre V Harrison, John A D'Agostino, James D Brown, Long P Quang, James R Uplinger, et al. A multi-purpose realistic haze benchmark with quantifiable haze levels and ground truth. IEEE Transactions on Image Processing, 2023. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 288, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 375 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 375 + ], + "type": "text", + "content": "[36] National Standards of People's Republic of China. Grade of fog forecast, 2011. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 376, + 288, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 288, + 409 + ], + "type": "text", + "content": "[37] Roberto Nebuloni. Empirical relationships between extinction coefficient and visibility in fog. Applied Optics, 44(18): 3795-3804, 2005. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 288, + 443 + ], + "type": "text", + "content": "[38] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically based rendering: From theory to implementation. MIT Press, 2023. 2, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 445, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 288, + 487 + ], + "type": "text", + "content": "[39] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In AAAI, pages 11908-11915, 2020. 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "type": "text", + "content": "[40] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Semantic foggy scene understanding with synthetic data. International Journal of Computer Vision, 126:973-992, 2018. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "type": "text", + "content": "[41] Daniel Scharstein and Richard Szeliski. High-accuracy stereo depth maps using structured light. In IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages I-I. IEEE, 2003. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 579, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 624 + ], + "type": "text", + "content": "[42] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In European Conference on Computer Vision, pages 746-760. Springer, 2012. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "type": "text", + "content": "[43] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927-1941, 2023. 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[44] Tao Sun, Mattia Segu, Janis Postels, Yuxuan Wang, Luc Van Gool, Bernt Schiele, Federico Tombari, and Fisher Yu. Shift: a synthetic driving dataset for continuous multi-task domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 21371-21382, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 654 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[45] Ketan Tang, Jianchao Yang, and Jue Wang. Investigating haze-relevant features in a learning framework for image dehazing. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2995-3000, 2014. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[46] Jean-Philippe Tarel, Nicolas Hautiere, Aurélien Cord, Dominique Gruyer, and Houssam Halmaoui. Improved visibility of road scene images under heterogeneous fog. In IEEE Intelligent Vehicles Symposium, pages 478–485, 2010. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[47] Jean-Philippe Tarel, Nicholas Hautiere, Laurent Caraffa, Aurélien Cord, Houssam Halmaoui, and Dominique Gruyer. Vision enhancement in homogeneous and heterogeneous fog. IEEE Intelligent Transportation Systems Magazine, 4 (2):6-20, 2012. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 219, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 240 + ], + "type": "text", + "content": "[48] Tali Treibitz and Yoav Y Schechner. Resolution loss without imaging blur. JOSA A, 29(8):1516-1528, 2012. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[49] Apostolia Tsirikoglou, Joel Kronander, Magnus Wrenninger, and Jonas Unger. Procedural modeling and physically based rendering for synthetic data generation in automotive applications. arXiv preprint arXiv:1710.06270, 2017. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 286, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 341 + ], + "type": "text", + "content": "[50] Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Contrastive learning for compact single image dehazing. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10551-10560, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "text", + "content": "[51] Jing Zhang, Yang Cao, Zheng-Jun Zha, and Dacheng Tao. Nighttime dehazing with a synthetic benchmark. In Proceedings of the ACM International Conference on Multimedia, pages 2355-2363, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "type": "text", + "content": "[52] Lin Zhang, Junhao Huang, Xiyuan Li, and Lu Xiong. Vision-based parking-slot detection: A dcnn-based approach and a large-scale benchmark dataset. IEEE Transactions on Image Processing, 27(11):5350-5364, 2018. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 432, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 545, + 464 + ], + "type": "text", + "content": "[53] Lin Zhang, Anqi Zhu, Shiyu Zhao, and Yicong Zhou. Simulation of atmospheric visibility impairment. IEEE Transactions on Image Processing, 30:8713-8726, 2021. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 466, + 545, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 519 + ], + "type": "text", + "content": "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 521, + 545, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 521, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 307, + 521, + 545, + 565 + ], + "type": "text", + "content": "[55] Yanfu Zhang, Li Ding, and Gaurav Sharma. Hazerd: an outdoor scene dataset and benchmark for single image dehazing. In IEEE International Conference on Image Processing, pages 3205-3209, 2017. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 567, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 609 + ], + "type": "text", + "content": "[56] Shiyu Zhao, Lin Zhang, Shuaiyi Huang, Ying Shen, and Shengjie Zhao. Dehazing evaluation: Real-world benchmark datasets, criteria, and baselines. IEEE Transactions on Image Processing, 29:6947-6962, 2020. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 611, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 545, + 654 + ], + "type": "text", + "content": "[57] Qingsong Zhu, Jiaming Mai, and Ling Shao. A fast single image haze removal algorithm using color attenuation prior. IEEE Transactions on Image Processing, 24(11):3522-3533, 2015. 6, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21772" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/e165eedf-6d86-4e7b-90c3-ccc6312e093e_content_list.json b/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/e165eedf-6d86-4e7b-90c3-ccc6312e093e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6c2dd7f552b6b4011544b91f71749c2dda9760a1 --- /dev/null +++ b/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/e165eedf-6d86-4e7b-90c3-ccc6312e093e_content_list.json @@ -0,0 +1,1479 @@ +[ + { + "type": "text", + "text": "SynSP: Synergy of Smoothness and Precision in Pose Sequences Refinement", + "text_level": 1, + "bbox": [ + 102, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tao Wang\\*, Lei Jin\\*, Zheng Wang?, Jianshu Li?, Liang Li?, Fang Zhao?, Yu Cheng?, Li Yuan?, Li Zhou?, Junliang Xing?, Jian Zhao?", + "bbox": [ + 186, + 179, + 779, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Beijing University of Posts and Telecommunications, 2Wuhan University, 3Ant Group,", + "bbox": [ + 222, + 218, + 745, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Institute of computing technology Chinese Academy of Sciences, $^{5}$ Nanjing University, $^{6}$ National University of Singapore,", + "bbox": [ + 119, + 234, + 849, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{7}$ Peking University, $^{8}$ Tsinghua University, $^{9}$ China Telecom Institute of AI, $^{10}$ Northwestern Polytechnical University", + "bbox": [ + 140, + 252, + 828, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wangtao@bupt.edu.cn, jinlei@bupt.edu.cn, wangzwhu@whu.edu.cn, jianshu@u.NUS.edu, liang.li@ict.ac.cn, zhaofang0627@gmail.com, e0321276@u.nus.edu, yuanli-ece@pku.edu.cn, zhouli2025@126.com, jlxing@tsinghua.edu.cn, jian.zhao@nwpu.edu.cn", + "bbox": [ + 102, + 271, + 861, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 356, + 313, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Predicting human pose sequences via existing pose estimators often encounters various estimation errors. Motion refinement methods aim to optimize the predicted human pose sequences from pose estimators while ensuring minimal computational overhead and latency. Prior investigations have primarily concentrated on striking a balance between the two objectives, i.e., smoothness and precision, while optimizing the predicted pose sequences. However, it has come to our attention that the tension between these two objectives can provide additional quality cues about the predicted pose sequences. These cues, in turn, are able to aid the network in optimizing lower-quality poses. To leverage this quality information, we propose a motion refinement network, termed SynSP, to achieve a Synergy of Smoothness and Precision in the sequence refinement tasks. Moreover, SynSP can also address multi-view poses of one person simultaneously, fixing inaccuracies in predicted poses through heightened attention to similar poses from other views, thereby amplifying the resultant quality cues and overall performance. Compared with previous methods, SynSP benefits from both pose quality and multi-view information with a much shorter input sequence length, achieving state-of-the-art results among four challenging datasets involving 2D, 3D, and SMPL pose representations in both single-view and multi-view scenes. Github code: https://github.com/InvertedForest/SynSP.", + "bbox": [ + 75, + 388, + 473, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 809, + 209, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human pose estimation [4, 16, 21, 37] has been widely used in automatic driving, human-computer interaction, motion", + "bbox": [ + 75, + 835, + 468, + 866 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/baf8189bb9233acac58ade0ad8229c78e828127b997e6d9c937f35f11d1ca177.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 356, + 890, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/75e2fddf21e0ef126936e467bc2448d0f9de88ec45683547542e40d5670ad282.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 455, + 883, + 556 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/813f6dcadbd43f806ae20c98274f895aa5d0479ba48fbf752b1c879107ae1af6.jpg", + "image_caption": [ + "Figure 1. Illustration of our motivation. (a): The x-axis represents the ratio of weights for Precision and Smoothness loss in the training process of SynSP, while the y-axis represents the value of MPJPE (Precision) and Acceleration error (Smoothness). These two errors exhibit a conflicting relationship. (b): Diff1 denotes the difference between the predicted sequence (from pose estimator) and the ground-truth sequence, and Diff2 denotes the difference between two output sequences from training SynSP with a loss ratio of 0.5 and 5, respectively. We notice that the trends of Diff1 and Diff2 are roughly the same in (b). (c): A larger value of Diff1 indicates lower prediction quality for the predicted sequences. For example, the left ankle of the person in the middle picture of (c) is occluded, making it difficult for the pose estimator to infer." + ], + "image_footnote": [], + "bbox": [ + 504, + 556, + 870, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "analysis, etc. Although current methods have achieved outstanding performance, both image-based and video-based pose estimators still sometimes exhibit noticeable errors in", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Both authors contributed equally to this research. † Lei Jin is corresponding author.", + "bbox": [ + 75, + 875, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1824", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "predicting pose sequences. The pose sequences generated by image-based pose estimators often show significant jitters due to a lack of inter-frame correlation, while video-based pose estimators alleviate jitters and improve accuracy by leveraging temporal information of input videos. However, training video-based pose estimators requires substantial amounts of video data and computational resources [7, 38, 39].", + "bbox": [ + 75, + 90, + 470, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Consequently, researchers propose motion refinement methods [3, 42] as post-processing techniques to optimize the predicted sequences from pose estimators to approximate the ground-truth sequences from human annotation, with two primary objectives: precision and smoothness. Precision objective aims to improve the position accuracy of the predicted joints through the temporal and spatial information, evaluated by Mean Per Joint Position Error (MPJPE) [11, 29]. Smoothness objective does not strive to make the output sequences as smooth as possible but rather to align the acceleration of the output with that of the ground-truth sequences. Both objectives aim to make the output sequences close to the ground-truth sequences. However, focusing on one of them will lead to errors in the other greatly increased, as shown in Fig. 1.", + "bbox": [ + 75, + 215, + 468, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the inherent tension between Precision and Smoothness objectives, previous methods attempt to achieve a satisfactory balance between them. As shown in Fig. 1(a), it can be observed that with an increase in the ratio of weights for Precision and Smoothness loss during SynSP training process, there is a noticeable rise in Acceleration error (Smoothness) while MPJPE (Precision) decreases, which is a common phenomenon in previous methods [42]. However, we note that the quality of the predicted sequences from estimator is related to the difference of outputs from SynSP biased towards the Precision and Smoothness objectives. Specifically, as illustrated in Fig. 1(b), the difference (Diff2) between the two output sequences, which focus on precision and smoothness respectively, is correlated with the quality of the predicted sequences, which is generated by the difference (Diff1) between predicted sequences from pose estimators and ground-truth sequences from human annotation. To further verify this correlation, we calculate the Pearson Correlation Coefficient between Diff1 and Diff2 on the Human3.6M and 3DPW datasets. Based on our experiment, the mean Pearson correlation coefficient [9] is 0.48 (bigger than 0.3 is considered as relevant), indicating that the dissimilarity (Diff2) between the two output sequences can serve as an indicator for the quality of predicted sequences without knowing the ground-truth.", + "bbox": [ + 75, + 444, + 470, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose SynSP to utilize this quality information. Specifically, we employ a transformer network and encode the quality information by Pose Quality Encoding (PQE) module to guide the network's attention towards", + "bbox": [ + 76, + 839, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "low-quality poses in sequences. Additionally, with the support of multi-view dataset: Human3.6M, we further find that multi-view information is effective to enhance the reliability of PQE. Therefore, we introduce the Pose Similarity Encoding (PSE) module to exploit the multi-view information by guide the network's attention towards similar poses from multi-view.", + "bbox": [ + 496, + 90, + 890, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions are summarized as follows:", + "bbox": [ + 517, + 196, + 857, + 210 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We translate the inherent tension between smoothness and precision into the quality cue of predicted sequences with the Pose Quality Encoding module in SynSP. Then we utilize this cue as the weight to help regulate the network's attention on each frame in the sequences.", + "- We further observe that similar poses can effectively rectify erroneous poses. Consequently, we propose Pose Similarity Encoding to encode pose similarity information and refine poses of multiple views simultaneously.", + "- We conduct sufficient experiments on four challenging datasets among 2D, 3D, and SMPL representations with image-based and video-based pose estimators. Notably, compared with previous method SmoothNet, SynSP's MPJPE improves by $16.9\\%$ on Human3.6M dataset with 3D pose estimator FCN, while reducing input sequence length to $25\\%$ and minimizing time delay to just $17\\%$ ." + ], + "bbox": [ + 501, + 214, + 890, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 467, + 640, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Pose Estimation. Human pose estimation task [14-16, 27] is to extract location of human joints from the collected data. There are 2D, 3D, and SMPL representations for the human pose. 2D representation only contains 2D joint coordinates parallel to the camera plane. 3D representation has one more dimension of depth information. SMPL representation [23] is a skinned vertex based model that represents a wide variety of body shapes in natural human poses, while needs more parameters to represent.", + "bbox": [ + 496, + 492, + 890, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human pose estimators can be divided into image-based methods [27] and video-based methods [17]. Pose sequences from image-based methods often show a lot of jitters when these methods are applied to videos, while video-based methods require more data and computing resources for training. 2D pose estimator, Hourglass [27] found an outstanding stacked hourglass networks for human pose estimation; SPIN [19] can learn to reconstruct 3D human pose and SMPL pose via model-fitting in the loop; video-based pose estimators, TCMR [6] and VIBE [17], can generate smooth pose sequences with temporal information. We conduct the experiments based on pose estimators mentioned above.", + "bbox": [ + 496, + 628, + 890, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motion Refinement Methods. The predicted human pose sequences can be further refined by motion refinement methods, making the position and acceleration of each joint in these sequences are closer to those of the ground-truth sequences. The motion refinement methods can be divided to", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1825", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/99289463ddb4e36c6fdff262e505eacb29452fb26849ed0b9a314baa5f55a50b.jpg", + "image_caption": [ + "Figure 2. The architecture of SynSP. FCN represents fully-connected network. In the base stage, with pose sequences of multiple views as input, SynSP first encodes these sequences individually in the Local Correlation Encoder, and merge multiple view's information in the Global Correlation Encoder with the Pose Similarity Encoding as their weight. After obtaining the individual embedding as the query, and incorporating the biased global information as the key and value, we get two outputs through the decoder module. In the refinement stage, we utilize these two detached outputs to generate Pose Quality Encoding and more precise Pose Similarity Encoding. By reusing embedding generated earlier, we achieve superior results using only one decoder with these encodings." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 872, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "low-pass filters and deep learning based methods. The traditional low-pass filters, e.g., Savitzky-Golay [30], Gaussian1d [40], and One-Euro [3], are generally adapted to smooth the predicted pose sequences, and perform well for those abrupt jitters, such as confusion in left and right hips. Although the processed sequences from low-pass filter does look smoother, it is difficult to ensure precision at the same time. Deep learning based method, i.e., SmoothNet [42], takes the two objectives into account and aims to improve both precision and smoothness. Our SynSP is also based on deep learning and can achieve better results by considering the pose sequence quality and multi-view correlation information.", + "bbox": [ + 75, + 426, + 472, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, there are also several human pose prior methods [2, 22, 28, 32], which try to learn a prior distribution of valid human poses. Although time-consuming for their sampling steps, these methods can be applied to the task of pose generation/completion/refinement. Hu-MoR [31] introduces a novel conditional VAE which enables expressive and general motion reconstruction and generation; GAN-based method [10] proposes a simple yet effective prior for SMPL model to bound it to realistic human poses; GFPose [8] employs a novel score-based generative framework to model plausible 3D human poses.", + "bbox": [ + 75, + 623, + 468, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 805, + 166, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 830, + 186, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task Description. Given predicted pose sequences generated by human pose estimators [16], motion refinement methods aim to fit these predicted sequences to the ground", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "truth sequences, making these prediction more precise and smooth. Usually, the predicted pose sequence $\\mathbf{P}$ has a shape of $\\mathbb{R}^{V\\times T\\times (J\\times D)}$ , where $V$ stands for the number of views for one individual; $T$ refers to the frame number of these predicted pose sequences; $J$ is the number of joints associated with datasets; $D$ denotes the spatial dimensions of these joints. We use $P_{(v,t,j)}$ to represent the $j_{th}$ joint's position at $t_{th}$ frame from $v_{th}$ view throughout the paper. The spatial dimension is included in the joint dimension, ignoring the spatial dimensions to simplify it.", + "bbox": [ + 496, + 426, + 890, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall Architecture. We proposed SynSP to use this quality information from the certain tension relationship between the two objectives, i.e., smoothness and precision. We also introduce multi-view information fusion as an optional enhancement for SynSP. As illustrated in Fig. 2, SynSP is mainly based on the transformer structure [34] and composed of two stages. Firstly, under the guidance of Pose Similarity Encoding I, Base Decoder can utilize both individual and global information to generate two branches, which are respectively biased towards smoothness and precision. Next, these two branches can generate Pose Quality Encoding, and a more precise Pose Similarity Encoding II, which further helps Refinement Decoder generate refined results.", + "bbox": [ + 496, + 577, + 892, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Base Stage", + "text_level": 1, + "bbox": [ + 500, + 800, + 619, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the base stage, SynSP takes predicted sequences from pose estimators as input and generates two sequences that focus on smoothness and precision, respectively.", + "bbox": [ + 496, + 824, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pose Similarity Encoding. Similarity for poses from multiview is effective in enhancing the performance of SynSP.", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1826", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On one hand, similar poses from different views can complement each other and improve the robustness of joint position. On the other hand, this encoding minimizes the network's attention on poses with significant distortions to reduce their impact for the refinement task.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pose Similarity Encoding (PSE) is calculated as follows:", + "bbox": [ + 96, + 167, + 467, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P _ {(v, t, j)} ^ {r e l} = \\hat {P} _ {(v, t, j)} - P _ {(v, t, c)}, \\\\ P S E _ {\\left(v _ {1}, v _ {2}, t\\right)} = \\exp \\left(- \\sum_ {j = 1} ^ {J} \\left(P _ {\\left(v _ {1}, t, j\\right)} ^ {r e l} - P _ {\\left(v _ {2}, t, j\\right)} ^ {r e l}\\right) ^ {2}\\right), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 195, + 468, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c$ represents the index of the \"pelvis\" keypoint in the joint list. $v_{1}$ and $v_{2}$ are used only for distinction, and both have the same domain as $v$ . Overall, PSE has a shape of $\\mathbb{R}^{V\\times V\\times T}$ , and we present a visualization example of PSE for one frame in Fig. 3.", + "bbox": [ + 75, + 258, + 468, + 334 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Global Correlation Encoder. In multi-view scenes, the correlation information between multiple views can help SynSP to deduce the location of the joint with jitters globally. Global Correlation Encoder (GCE) module is introduced to produce the keys and values required by the Base Decoder with the global correlation information between multiple views. First, the predicted sequences $\\mathbf{P}$ are projected to embedding $\\mathbf{E}$ through a full connected layer along the joint dimension.", + "bbox": [ + 75, + 335, + 468, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, we use $E_{(v,t,e)}$ to denote an item of projected embedding $\\mathbf{E}$ , where $e$ refers to the index of the embedding dimension. Then, the view (v) and the frame (t) dimensions of $\\mathbf{E}$ are merged, so that pose information from multiple views and temporal information can be sensed by the encoder simultaneously. Afterwards, $E_{((v,t),e)}$ is sent to a encoder module of transformer [34] to interact pose sequences from multiple views and get the global information $E_{((v,t),e)}^g$ . Then this information is added with $PSE_{(v_1,v_2,t)}$ to adjust $v_{1th}$ view's attention to other views' information by broadcast operation: $E_{((v_2,t),e)}^g + PSE_{(v_1,v_2,t)} = E_{(1,(v_2,t),e)}^g + PSE_{(v_1,(v_2,t),1)} = E_{(v_1,(v_2,t),e)}^g$ , which is the final global information from GCE module.", + "bbox": [ + 75, + 473, + 468, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Local Correlation Encoder. Local Correlation Encoder (LCE) module focuses on the temporal dimension with single-view and produces the query required by the Base Decoder. LCE reuses the projected embedding $\\mathbf{E}$ from the full connected layer mentioned in GCE as input, and directly sends it to another standard encoder module of transformer [34] to obtain local information in the frame (t) dimension (take the view (v) dimension as input batch). Finally, for $v_{1th}$ view, $t_{th}$ frame, $e_{th}$ item, LCE produces the local information $E_{(v_1,t,e)}^l$ with exploring the correlation of the time dimension.", + "bbox": [ + 75, + 670, + 468, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Base Decoder. At the end of base stage, we stack 5 standard decoding layers of transformer [34] into Base Decoder. The Base Decoder takes the local information $E_{(v_1,t,e)}^l$ as the query, to interact with global information $E_{(v_1,(v_2,t),e)}^g$ ,", + "bbox": [ + 75, + 837, + 468, + 904 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7d1b9819a56f28f6b928590aa779535d386b062872f317c7fc54809d43c3886e.jpg", + "image_caption": [ + "Pose Similarity Encoding" + ], + "image_footnote": [], + "bbox": [ + 522, + 90, + 671, + 208 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f42e415c8a0ab521219e0ca802c9e523e72a52082e0873c33c6b803d24a5fe9d.jpg", + "image_caption": [ + "Pose Quality Encoding", + "Figure 3. Visualization of the two encodings of one frame. $PSE$ : the diagonal has a constant value of 1. $PQE$ : the value outside the diagonal is constant 0. In this way, the addition of the two encodings cannot affect each other." + ], + "image_footnote": [], + "bbox": [ + 691, + 90, + 870, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which serves as keys and values for the Base Decoder. At the last layer of Base Decoder, we use two decoder layers to generate two branches biased towards smoothness and precision respectively at the same time.", + "bbox": [ + 496, + 313, + 890, + 373 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} K, V = E _ {(v _ {1}, (v _ {2}, t), e)} ^ {g}, Q ^ {0} = E _ {(v _ {1}, t, e)} ^ {l}, \\\\ Q _ {(v _ {1}, t, e)} ^ {d _ {t h}} = \\mathbf {D L a y e r} ^ {d _ {t h}} \\left(Q ^ {(d - 1) _ {t h}}, K, V\\right), d \\in \\{1, 2, 3 \\}, \\tag {2} \\\\ P _ {(v _ {1}, t, j)} ^ {\\text {p o s}} = \\operatorname {F C N} \\left(\\mathbf {D L a y e r} ^ {4 _ {\\mathrm {t h}}} \\left(\\mathrm {Q} ^ {3 _ {\\mathrm {t h}}}, \\mathrm {K}, \\mathrm {V}\\right)\\right), \\\\ P _ {(v _ {1}, t, j)} ^ {a c c} = \\operatorname {F C N} \\left(\\mathbf {D L a y e r} ^ {5 _ {\\mathrm {t h}}} \\left(\\mathrm {Q} ^ {3 _ {\\mathrm {t h}}}, \\mathrm {K}, \\mathrm {V}\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 383, + 890, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q$ , $K$ , and $V$ denote the query, key, and value for Base Decoder respectively, $\\mathbf{DLayer}^{d_{th}}$ denotes the $d_{th}$ decoder layer, $\\mathrm{FCN}(\\cdot)$ is fully connected network. In this way, SynSP can generate two pose squences $P^{pos}$ and $P^{acc}$ from its Acceleration Branch and Position Branch, which are biased towards the precision and smoothness objectives in the final loss function Equ. (4), respectively. Next, we can also further optimize these sequences in the refinement stage, with a small amount of calculation and parameters as the cost.", + "bbox": [ + 496, + 473, + 890, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Refinement Stage", + "text_level": 1, + "bbox": [ + 500, + 633, + 669, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the refinement stage, we aim to seek a better way to fuse Acceleration Branch and Position Branch, which concentrate on different objectives.", + "bbox": [ + 496, + 657, + 890, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Acceleration and Position Branches. The output sequences from Acceleration Branch and Position Branch are detached from computation graph for a stabler training process as the predictions from base stage are independent of the refinement stage. These outputs can not only generate more accurate Pose Similarity Encoding, but also provide the quality information about these pose sequences.", + "bbox": [ + 496, + 704, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pose Quality Encoding. We have empirically observed that when outputs from the two branches show a greater discrepancy, the prediction quality is poorer. Specifically, we computed the Pearson Correlation Coefficient between the prediction quality and the difference between two outputs is about 0.48 (more details in the supplementary materials),", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1827", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which shows that the two have a strong correlation. Hence, we propose the Pose Quality Encoding (PQE) to utilize this correlation to improve the final prediction. Each item of PQE is calculated as follows:", + "bbox": [ + 76, + 90, + 470, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {(v _ {1}, v _ {2}, t, j)} ^ {d i s} = | P _ {(v _ {1}, t, j)} ^ {a c c} - P _ {(v _ {2}, t, j)} ^ {p o s} |,\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 159, + 328, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP Q E _ {(v _ {1}, v _ {2}, t)} ^ {\\prime} = \\frac {1}{J} \\sum_ {j = 1} ^ {J} \\exp \\left(- \\frac {P _ {(v _ {1} , v _ {2} , t , j)} ^ {d i s}}{\\max _ {j \\in J} P _ {(v _ {1} , v _ {2} , t , j)} ^ {d i s} + \\sigma}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 181, + 468, + 222 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP Q E _ {(v _ {1}, v _ {2}, t)} = \\left\\{ \\begin{array}{c c} P Q E _ {(v _ {1}, v _ {2}, t)} ^ {\\prime}, & v _ {1} = v _ {2}; \\\\ 0, & v _ {1} \\neq v _ {2}. \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 224, + 377, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we calculate PQE by the distance $P^{dis}$ between the outputs $P^{acc}$ and $P^{pos}$ , which are generated by Acceleration Branch and Position Branch, respectively. To ensure that extreme values do not interfere with PQE, $P^{dis}$ is normalized by dividing the maximum value in the individual's sequence. We introduce a small constant $\\sigma$ (set to $1 \\times 10^{-6}$ ) to avoid division by zero.", + "bbox": [ + 76, + 266, + 468, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike PSE, which is calculated based on correlations among multiple views on one person, PQE is computed by $P^{dis}$ for a single view across multiple frames. Since PQE and PSE have distinct considerations regarding multiple views versus single view, they can be combined through addition rather than concatenation without mutual influence as illustrated in Fig. 3, thereby enhancing the computational efficiency of SynSP.", + "bbox": [ + 75, + 371, + 468, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Refinement Decoder. The output sequences of the two branches exhibit enhanced accuracy and smoothness compared to the input sequences. A straightforward approach would involve averaging these outputs and replicating the base stage pipeline once again. However, our experimental findings indicate that this approach does not yield significant benefits but doubles the computational cost. Please refer to supplementary material for details. This may be attributed to the fact that while the optimized output is indeed smooth and precise, it lacks crucial positional information from the original jitter input, rendering optimization of an already optimized output meaningless.", + "bbox": [ + 75, + 492, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To circumvent this issue and maintain a lightweight model, we reuse the embedding from the Global Correlation Encoder and Base Decoder. The former embedding retains the original position information, while the latter embedding contains optimization information about the output. In this way, we can produce superior outcomes by using just one additional decoder. We stack 4 standard decoding layers of transformer [34] in Refinement Decoder, and the inner calculation of Refinement Decoder is similar with Base Decoder. As shown in Tab. 5, this approach effectively improves the positional accuracy of pose sequences.", + "bbox": [ + 75, + 674, + 468, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Analysis on the Refinement Stage. In the refinement stage, the difference between the position and acceleration branches serves as a guiding factor for the quality of predicted sequence. Next, within the refinement decoder,", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/09c11a710195fc6c8d00267fc9afcd46a1b0f2068a3a8e3408eec1065cd251c5.jpg", + "image_caption": [ + "Figure 4. Illustration of Sliding Window Average Algorithm (SWAA), which is widely used by motion refinement works. This figure proves that the amount of computation and latency is proportional to the window length." + ], + "image_footnote": [], + "bbox": [ + 506, + 88, + 849, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "greater emphasis is placed on frames with lower pose quality to facilitate improvements in the final performance. Regarding why this difference is associated with pose quality, we posit that the two branches are easy to remain consistent under the normal circumstances, nevertheless, in challenging scenarios such as occlusion or sudden movement, it is more likely for the two branches to exhibit dissimilar behavior. Whether in pose estimator or motion refinement model, pose quality is is obviously related with scenarios.", + "bbox": [ + 498, + 270, + 890, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training and Inference", + "text_level": 1, + "bbox": [ + 500, + 417, + 714, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Different from previous work, we only use the window length $T = 8$ for training and inference. Compared with the window length of 32, this setting reduces the time delay of $(32 - 8) / 12 = 2s$ while the inference speed of pose estimator is 12fps, and have $8 / 32 = 0.25$ times the previous amount of computation as shown in Fig. 4.", + "bbox": [ + 498, + 440, + 890, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. The losses of the three outputs, i.e., Acceleration Branch output, Position Branch output, and final output, share the same calculation method. We have two losses, $\\mathcal{L}_{acc}$ and $\\mathcal{L}_{pos}$ , for each one of the three outputs:", + "bbox": [ + 498, + 532, + 890, + 592 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} A _ {(v, t, j)} = \\left(P _ {(v, t + 2, j)} - P _ {(v, t + 1, j)}\\right) - \\left(P _ {(v, t + 1, j)} - P _ {(v, t, j)}\\right), \\\\ \\mathcal {L} _ {\\mathrm {p o s}} = \\frac {1}{V \\cdot T \\cdot J} \\sum_ {v = 1} ^ {V} \\sum_ {t = 1} ^ {T} \\sum_ {j = 1} ^ {J} | P _ {(v, t, j)} - \\hat {P} _ {(v, t, j)} |, \\\\ \\mathcal {L} _ {\\mathrm {a c c}} = \\frac {1}{V \\cdot T \\cdot J} \\sum_ {v = 1} ^ {V} \\sum_ {t = 1} ^ {T} \\sum_ {j = 1} ^ {J} | A _ {(v, t, j)} - \\hat {A} _ {(v, t, j)} |, \\\\ \\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p o s}} + \\lambda \\mathcal {L} _ {\\text {a c c}}. \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 603, + 890, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments, $\\lambda$ is set as 0, 1, and 0.5 for Position Branch, Acceleration Branch, and the final output, respectively, which is a trade-off between smoothness and precision. For all branches, $\\mathcal{L}_{\\mathrm{pos}}$ is crucial to ensuring training stability. Furthermore, to balance $\\mathcal{L}_{\\mathrm{pos}}$ and $\\mathcal{L}_{\\mathrm{acc}}$ , the parameter $\\lambda$ should be set as 0.5, determined by the absolute value of each element. Thus, for the final output, we have set $\\lambda$ to 0.5, whereas in the Acceleration branch, $\\lambda$ is assigned as 1 to prioritize smoothness.", + "bbox": [ + 496, + 734, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. During inference, we obtain outputs from the refinement stage. Then the outputs go through a", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1828", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/20168a4302eb6b363ecc8f149bc35f57114088967a4b428796b40ed6a98663d0.jpg", + "table_caption": [ + "Table 1. Comparison with related works on Human3.6M, AIST++ datasets with 2D, 3D, and SMPL human pose estimator Hourglass [27], FCN [26], and SPIN [19], respectively. WS denotes the windows size in training, inference, and post-process. * indicates that SynSP is trained and tested in multi-view scenes." + ], + "table_footnote": [], + "table_body": "
MethodWSHuman3.6M / 2DHuman3.6M / 3DAIST++ / SMPL
MPJPE↓PA-MPJPE↓Accel↓MPJPE↓PA-MPJPE↓Accel↓MPJPE↓PA-MPJPE↓Accel↓
InputN/A9.427.641.5454.5542.219.17107.7274.4033.20
One-Euro [3]110.697.980.3455.2042.733.80108.9775.2714.70
Gaussian1d [40]329.377.560.5153.6741.602.43104.8472.1810.05
Savitzky-Golay [30]329.357.550.1753.4841.191.34104.5872.306.07
SmoothNet [42]329.257.570.1552.7240.921.03103.0071.195.72
SynSP88.136.090.1551.3640.131.0284.6359.026.08
SynSP*87.625.640.1541.7833.320.98---
", + "bbox": [ + 102, + 127, + 870, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "post-processing called sliding window average algorithm adopted by most works as shown in Fig. 4 (we have accelerated it by parallel computing to $1\\%$ of the original latency, refer to supplementary material for more details). It can be seen that the delay and computation needs for model and post-processing becomes smaller with shorter window length.", + "bbox": [ + 75, + 306, + 470, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Single-View Scenes. While $V = 1$ in the equations above, SynSP can be directly applied to adapt single-view scenes. In single-view scenes, PSE module is no longer needed. Since multi-view scenes are not considered by current related works and not our main motivation for the proposed SynSP, we conduct sufficient single-view experiments for a comprehensive comparison.", + "bbox": [ + 75, + 412, + 468, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 534, + 209, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 76, + 559, + 269, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. To demonstrate the generalizability and efficiency of our method, we conduct experiments of SynSP in three human pose representations of 2D, 3D, and SMPL for single-view and multi-view scenes on Human3.6M [12], AIST++ [20], 3DPW [35], and CMU-Mocap [33] datasets. For the fairness of the comparison, we use these datasets made by SmoothNet [42] with kinds of pose estimators in the experiments. To compare with pose prior models, we also follow GFPose [8] to add the same degree of Gaussian noise or uniformly distributed noise to Human3.6M and CMU-Mocap datasets as input for the pose denoise task.", + "bbox": [ + 75, + 582, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Following the related works [42], we use Accel error, MPJPE, and PA-MPJPE (Procrustes-aligned MPJPE) [1, 5] to evaluate the acceleration and position proximity between the predicted sequence and the ground-truth sequence, and the lower metrics the smaller error of the predicted sequence.", + "bbox": [ + 75, + 750, + 470, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. SynSP consists of two 4-layer encoder modules (GCE, LCE), a 5-layer decoder module (Base Decoder), and a 4-layer decoder module (Refinement Decoder). We initialize the learning rate as 0.001 and decay", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "it by a factor of 0.95 per epoch. SynSP is trained for 70 epochs with Adam optimizer, requiring approximately two hours and 7GB of GPU memory with a batch size of 4096. The computational setup consists of an Intel(R) Xeon(R) Gold 6271C CPU @ 2.60GHz for the CPU and a Tesla V100 SXM2 32GB for the GPU.", + "bbox": [ + 496, + 306, + 893, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-Of-The-Arts.", + "text_level": 1, + "bbox": [ + 498, + 412, + 818, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our experiments, we first demonstrate that SynSP consistently outperforms previous methods in both single-view and multi-view scenes with several pose estimators. Then we compare with pose prior models on the pose denoise task by smoothing the pose sequences with artificial noisy. (There are also some methods[13, 41] with a super-long window length, and the comparison with them is shown in the supplementary materials.)", + "bbox": [ + 496, + 436, + 893, + 556 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/90ca88a5d09d01e8b49eaf9f37fa88490c3ac12d5b72a896aab14d67fef85d5f.jpg", + "table_caption": [ + "Table 2. Comparison of various video-based pose estimators. + represents the combination of estimators and motion refinement methods. * indicates multi-view inputs." + ], + "table_footnote": [], + "table_body": "
DatasetMethodWSMPJPE↓Accel↓
3DPWMAED [36]1679.00-
MPS-Net [38]1684.30-
TCMR [6]1686.466.75
TCMR+SmoothNet [42]16+3286.506.00
TCMR+SynSP16+886.105.90
PARE [18]179.0025.60
PARE+SmoothNet [42]1+3278.105.91
PARE+SynSP1+876.206.16
AIST++VIBE [17]16106.931.60
VIBE+SmoothNet [42]16+3297.474.15
VIBE+SynSP16+877.004.32
H36MTransFusion* [24]125.52-
PPT* [25]125.1611.60
PPT*+SmoothNet [42]1+3223.971.31
PPT*+SynSP1+821.511.10
", + "bbox": [ + 501, + 627, + 893, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1829", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9b1bd3402fbc2b4d68f4787cb174ce5469c1b1b3c913779335d82f28ca2286d6.jpg", + "table_caption": [ + "Table 3. Comparison on Human3.6M dataset for the denoise task. $N(0,\\sigma)$ represents a Gaussian distribution with a mean of 0 and a variance of $\\sigma$ , and $U(k)$ represents a uniform distribution at $[0,k]$ . $\\mathbf{M}\\downarrow$ and $\\mathrm{A}\\downarrow$ denote MPJPE↓ and Accel↓ respectively." + ], + "table_footnote": [], + "table_body": "
MethodN(0,100)N(0,400)U(50)U(100)
M↓A↓M↓A↓M↓A↓M↓A↓
Noisy input65.6160.3131.1320.594.8231.3189.5462.7
GFpose [8]42.8-64.6-50.9-89.4-
SmoothNet [42]42.410.757.914.249.410.274.316.3
SynSP37.68.856.113.045.610.969.415.7
", + "bbox": [ + 78, + 155, + 470, + 260 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e7dfb1cf1b4ff02a9f1abb2ebd08064c26bf39f71e9513131118b4fe87738823.jpg", + "table_caption": [ + "Table 4. Comparison on CMU-Mocap dataset for the denoise task. $N(0,\\sigma)$ represents a Gaussian distribution with a mean of 0 and a variance of $\\sigma$ , and $U(k)$ represents a uniform distribution at [0,k]. $\\mathbf{M} \\downarrow$ and $\\mathbf{A} \\downarrow$ denote MPJPE $\\downarrow$ and Accel $\\downarrow$ respectively." + ], + "table_footnote": [], + "table_body": "
MethodN(0,100)N(0,400)U(50)U(100)
M↓A↓M↓A↓M↓A↓M↓A↓
Noisy input61.2143.8122.5287.688.7207.9177.4415.8
GFpose [8]40.4-60.0-45.6-84.4-
SmoothNet [42]35.14.341.54.937.74.244.95.1
SynSP20.63.728.04.424.44.132.44.6
", + "bbox": [ + 78, + 340, + 470, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "2D Representation. In Tab. 1, we compare SynSP with other motion refinement methods in Human3.6M dataset with 2D pose estimator Hourglass [27]. With the setting of 8 window length, our SynSP outperforms all the methods while it is about 2 seconds faster (except the One-Euro) according to Fig. 4. Compared with SmoothNet, which is trained on multiple datasets, i.e., AIST-VIBE-3D, 3DPW-SPIN-3D, and Human3.6M-FCN-3D, SynSP can still outperforms it by $16.9\\%$ for MPJPE and $25.2\\%$ for PA-MPJPE with smaller window size. According to the ablation experiment results in Sec. 4.3, the significant accuracy improvement of SynSP is attributed to the mutual patching between the pose quality information (PQE) and multi-view information (PSE).", + "bbox": [ + 75, + 460, + 468, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D Representation. In Tab. 1, we also compare SynSP with other motion refinement methods in Human3.6M dataset with 3D pose estimator FCN [26]. Similar with the comparison in 2D Representation, we can still outperforms SmoothNet by $20.8\\%$ for MPJPE and $18.6\\%$ for PA-MPJPE in Human3.6M dataset. There is also a decrease of $4.9\\%$ for the Accel error. Furthermore, SynSP obtains these results only with 8 window size, while SmoothNet and most filters need 32 frames.", + "bbox": [ + 75, + 672, + 468, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SMPL Representation. Then we compare SynSP with other motion refinement methods in AIST++ dataset with SMPL pose estimator SPIN as shown in Tab. 1. SynSP shows superior performance in SMPL representation.", + "bbox": [ + 75, + 809, + 468, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Video-based Pose Estimators. The video-based estimators already include temporal informa", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion. Compared with the image-based estimators, motion refinement methods have difficult to optimize the sequences from video-based estimators. In Tab. 2, we compare SynSP with SOTA methods and SmoothNet in AIST++ and 3DPW datasets with 3D pose video-based estimator VIBE [17] and TCMR [6], respectively. We can find that the high performance video-based estimators, TCMR, is hard to be refined. SmoothNet improves the smoothness by processing a longer window length than the video-based estimators. SynSP greatly increases precision and smoothness in a shorter window length with its outstanding architecture as shown in Fig. 2. The image-based estimator, PARE [18], has excellent performance in precision, SynSP can also make up for its shortcomings in smoothness and outperform some video-based estimators.", + "bbox": [ + 496, + 90, + 890, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Multi-view Pose Estimators. We further verify the performance of SynSP with the multi-view pose estimator, i.e., PPT [25]. As shown in Tab. 2, we can notice that SynSP can further improve the precision of PPT, which has superiority performance by considering multiview connections. We also present the result of other multiview methods such as TransFusion for comparison. This experiment results further proves the multi-view refinement ability of PSE for video-based estimators.", + "bbox": [ + 496, + 316, + 890, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Pose Priors Methods. To ensure a fair comparison with human pose prior models, we also conducted denoise experiments following the methodology of the SOTA work [8] among pose priors methods. Here, window size of GFPose, SmoothNet, and SynSP are 1, 32, and 8 respectively, while GFPose, as a generative model for human pose prior task, has 1000 sampling steps, i.e., 1000 times of model inference. GFPose has not presented its Accel errors since it does not take temporal information into account. As presented in Tab. 3 and Tab. 4, SynSP can effectively solve MPJPE (Precision) problems. Across noisy Human3.6M and CMU-Mocap datasets, SynSP can improve the pose quality by about $54.5\\%$ in terms of MPJPE and $96.0\\%$ in terms of Accel errors for Gaussian noise with a variance of 100. For Gaussian noise with a large variance of 400, SynSP can improve the pose quality by about $67.2\\%$ in terms of MPJPE and $97.2\\%$ int terms of Accel errors. We also observe consistent improvement in uniform noise as shown in Tab. 4.", + "bbox": [ + 496, + 455, + 890, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study.", + "text_level": 1, + "bbox": [ + 500, + 753, + 656, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To ensure representativeness of ablation experiments, all experiments were performed on the largest dataset, Human3.6M, with 3D estimator FCN.", + "bbox": [ + 496, + 777, + 890, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Modules of SynSP. In order to verify the rationality of the structure of SynSP, we conducted ablation experiments on each module in SynSP. In Tab. 5, the line without a tick represents the results of baseline, which is the network of base stage without PSE I. And this network surpasses Smooth-", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1830", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3986a39a4d6a09b41c382cbaec81b0e0e4643159a63870d336a5659873003bc7.jpg", + "table_caption": [ + "Table 5. Ablation study on PSE I, RS (Refinement Stage), PSE II, and PQE modules. * denotes directly using the output from base stage as input of refinement stage to re-optimize these sequence." + ], + "table_footnote": [], + "table_body": "
PSE IRSPSE IIPQEMPJPEAccel
51.21.09
✓48.71.07
✓*51.41.08
✓49.61.07
✓✓46.71.05
✓✓✓46.41.04
✓✓✓42.50.99
✓✓✓✓41.80.98
", + "bbox": [ + 78, + 142, + 467, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Net in precision by $2.8\\%$ . PSE I, RS, PSE II, and PQE all have gains in results, and the combination of PSE I, RS, and PQE has brought greater benefits. We believe that this benefit mainly comes from the combination of SynSP with global information PSE I, which can help generate more accurate quality information according to the poses from other views. Besides, utilizing the difference between acceleration and Position Branch improves both smoothness and precision of final prediction. It is worth noting that using the average value of the two branches as the input RS does not bring reasonable gains. We suspect that the output of base stage loses the original predicted position information, which affects the accuracy and makes the refinement stage useless compared with the baseline.", + "bbox": [ + 75, + 324, + 468, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Time Efficiency Analysis. We conducted latency experiments on the Human3.6M dataset using FCN [26] as the 3D pose estimator with a batch size of 1416, as shown in Tab. 6. Considering the practical application environment, we report not only the inference latency of the model, as shown in the 3rd column of Tab. 6, but also the latency associated with the Sliding Window Average Algorithm (SWAA), which includes the delay in waiting for pose estimators and the computation cost of SWAA (we have optimized SWAA specifically for parallel computation on GPUs). The bold cell in each row of Tab. 6 represents the primary factor contributing to latency. For video processing speed of 12 frames per second for the pose estimator (based on the general inference speed of various pose estimators), a window size of 8 would require a waiting time of $8/12 = 667\\mathrm{ms}$ ; similarly, a window size of 32 would require a waiting time of $32/12 = 2667\\mathrm{ms}$ . Furthermore, the model inference time for SynSP is only about 10ms. Overall, the main time cost primarily arises from the internal waiting latency for SmoothNet and SynSP. For the pose priors method GFPose, the main time delay is its 1000 sampling steps. Hence, SynSP is more suitable for near real-time pose refinement task.", + "bbox": [ + 75, + 536, + 468, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis on Number of Views. As shown in Tab. 7, as more poses from different views are input at the same time", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/36e3d9001c87066462db8f15c626805ff12a0e949b3bb63ec219b8fa3adb33cc.jpg", + "table_caption": [ + "Table 6. Time efficiency analysis on SynSP including model inference time and time cost for sliding window. Besides, we assume that FPS (Frame Per Second) for pose estimator is 12." + ], + "table_footnote": [], + "table_body": "
MethodsWSModel (ms)SWAATotal (ms)
Wait (ms)Execution (ms)
GFPose13772--3772
SmoothNet321.2126678023470
SynSP811.06670.3678
SynSP*812.36670.3680
", + "bbox": [ + 500, + 142, + 895, + 247 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d9fea97b989168d78f11220127f9b604a9155d29830f318fd7c95abea8c8b192.jpg", + "table_caption": [ + "Table 7. Contrast experiments on the number of input views. $\\dagger$ means that we randomly combine several people together for input during the training process, instead of using multiple views for one person. * means that in the case of $\\dagger$ , several people with similar poses are input at the same time during the inference process." + ], + "table_footnote": [], + "table_body": "
View NumberMPJPEPA-MPJPEAccel
151.440.11.02
247.237.61.02
346.537.11.00
441.833.30.98
4†50.439.51.01
4*49.739.11.01
", + "bbox": [ + 501, + 342, + 890, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "during the training process and inference process, the performance of SynSP is more efficient. It is justifiable due to the introduction of more global information. If number of views is larger than 4, we believe that the performance of our model can be further promoted with only a marginal increase in model complexity.", + "bbox": [ + 496, + 488, + 890, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 593, + 617, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes SynSP, a motion refinement method that utilizes the tension relationship between smoothness and precision through synergistic optimization. Specifically, SynSP outputs Position Branch and Acceleration Branch biased towards precision and smoothness, respectively. Then we employ the discrepancy between the two branches as a quality cue for the predicted sequence, which can be utilized in subsequent stages to further refine the pose sequence. In addition, we introduce Pose Similarity Encoding, achieving simultaneous refinement of multi-view poses and performance enhancement.", + "bbox": [ + 496, + 618, + 890, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 799, + 651, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The paper is supported by National Key R & D Program of China (No. 2021YFB3300100), Natural Science Foundation of China No.62102039, the Fundamental Research Funds for the Central Universities No.500422813, and Natural Science Foundation of China No.6232211.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1831", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] A. Benzine, F. Chabot, B. Luvison, Q. C. Pham, and C. Achard. Pandanet: Anchor-based single-shot multi-person 3d pose estimation. In CVPR, 2020. 6", + "[2] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In ECCV, pages 561-578, 2016. 3", + "[3] Géry Casiez, Nicolas Roussel, and Daniel Vogel. 1€ filter: a simple speed-based low-pass filter for noisy input in interactive systems. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pages 2527-2530, 2012. 2, 3, 6", + "[4] Yu Cheng, Bo Yang, Bo Wang, and Robby T Tan. 3d human pose estimation using spatio-temporal networks with explicit occlusion training. In AAAI, pages 10631-10638, 2020. 1", + "[5] Yu Cheng, Bo Wang, Bo Yang, and Robby T. Tan. Monocular 3d multi-person pose estimation by integrating top-down and bottom-up networks. In CVPR, pages 7649-7659, 2021. 6", + "[6] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consistent 3d human pose and shape from a video. In CVPR, pages 1964-1973, 2021. 2, 6, 7", + "[7] Jiaming Chu, Lei Jin, Xiaojin Fan, Yinglei Teng, Yunchao Wei, Yuqiang Fang, Junliang Xing, and Jian Zhao. Single-stage multi-human parsing via point sets and center-based offsets. In ACM MM, pages 1863-1873, 2023. 2", + "[8] Hai Ci, Mingdong Wu, Wentao Zhu, Xiaoxuan Ma, Hao Dong, Fangwei Zhong, and Yizhou Wang. Gfpose: Learning 3d human pose prior with gradient fields. CVPR, 2023. 3, 6, 7", + "[9] Israel Cohen, Yiteng Huang, Jingdong Chen, Jacob Benesty, Jacob Benesty, Jingdong Chen, Yiteng Huang, and Israel Cohen. Pearson correlation coefficient. Noise reduction in speech processing, pages 1-4, 2009. 2", + "[10] Andrey Davydov, Anastasia Remizova, Victor Constantin, Sina Honari, Mathieu Salzmann, and Pascal Fua. Adversarial parametric pose prior. In CVPR, pages 10997-11005, 2022. 3", + "[11] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Stefano Alletto, and Rita Cucchiara. Compressed volumetric heatmaps for multi-person 3d pose estimation. In CVPR, 2020. 2", + "[12] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE TPAMI, 36(7):1325-1339, 2013. 6", + "[13] Kyung-Min Jin, Byoung-Sung Lim, Gun-Hee Lee, Tae-Kyung Kang, and Seong-Whan Lee. Kinematic-aware hierarchical attention network for human pose estimation in videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5725-5734, 2023. 6", + "[14] Lei Jin, Xiaojuan Wang, Xuecheng Nie, Luoqi Liu, Yandong Guo, and Jian Zhao. Grouping by center: Predicting centripetal offsets for the bottom-up human pose estimation. IEEE TMM, 2022. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Lei Jin, Chenyang Xu, Xiaojuan Wang, Yabo Xiao, Yandong Guo, Xuecheng Nie, and Jian Zhao. Single-stage is enough: Multi-person absolute 3d pose estimation. In CVPR, pages 13086-13095, 2022.", + "[16] Lei Jin, Xiaojuan Wang, Xuecheng Nie, Wendong Wang, Yandong Guo, Shuicheng Yan, and Jian Zhao. Rethinking the person localization for single-stage multi-person pose estimation. IEEE TMM, 2023. 1, 2, 3", + "[17] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In CVPR, pages 5253-5263, 2020. 2, 6, 7", + "[18] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. In ICCV, pages 11127-11137, 2021. 6, 7", + "[19] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 2, 6", + "[20] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In ICCV, pages 13401-13412, 2021. 6", + "[21] Yang Li, Kan Li, Shuai Jiang, Ziyue Zhang, Congzhen-tao Huang, and Richard Yi Da Xu. Geometry-driven self-supervised method for 3d human pose estimation. In AAAI, pages 11442-11449, 2020. 1", + "[22] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Transactions on Graphics (TOG), 39(4):40-1, 2020. 3", + "[23] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM TOG, 34(6):1-16, 2015. 2", + "[24] Haoyu Ma, Liangjian Chen, Deying Kong, Zhe Wang, Xingwei Liu, Hao Tang, Xiangyi Yan, Yusheng Xie, Shih-Yao Lin, and Xiaohui Xie. Transfusion: Cross-view fusion with transformer for 3d human pose estimation. arXiv preprint arXiv:2110.09554, 2021. 6", + "[25] Haoyu Ma, Zhe Wang, Yifei Chen, Deying Kong, Liangjian Chen, Xingwei Liu, Xiangyi Yan, Hao Tang, and Xiaohui Xie. Ppt: Token-pruned pose transformer for monocular and multi-view human pose estimation. In ECCV, 2022. 6, 7", + "[26] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, pages 2640-2649, 2017. 6, 7, 8", + "[27] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In ECCV, pages 483-499, 2016. 2, 6, 7", + "[28] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 3", + "[29] Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with temporal convolutions and semi-supervised training. In CVPR, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1832", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] William H Press and Saul A Teukolsky. Savitzky-golay smoothing filters. Computers in Physics, 4(6):669-672, 1990. 3, 6", + "[31] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In ICCV, pages 11488-11499, 2021. 3", + "[32] Garvita Tiwari, Dimitrije Antić, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In ECCV, pages 572–589, 2022. 3", + "[33] Carnegie Mellon University. Cmu graphics lab motion capture database: http://mocap.cs.cmu.edu/. 6", + "[34] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeuIPS, 30, 2017. 3, 4, 5", + "[35] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, pages 601-617, 2018. 6", + "[36] Ziniu Wan, Zhengjia Li, Maoqing Tian, Jianbo Liu, Shuai Yi, and Hongsheng Li. Encoder-decoder with multi-level attention for 3d human shape and pose estimation. In ICCV, pages 13033-13042, 2021. 6", + "[37] Tao Wang, Lei Jin, Zhang Wang, Xiaojin Fan, Yu Cheng, Yinglei Teng, Junliang Xing, and Jian Zhao. Decenternet: Bottom-up human pose estimation via decentralized pose representation. In ACM MM, pages 1798-1808, 2023. 1", + "[38] Wen-Li Wei, Jen-Chun Lin, Tyng-Luh Liu, and Hong-Yuan Mark Liao. Capturing humans in motion: Temporal-attentive 3d human pose and shape estimation from monocular video. In CVPR, pages 13211-13220, 2022. 2, 6", + "[39] J. Wu, H. Zheng, B. Zhao, Y. Li, B. Yan, R. Liang, W. Wang, S. Zhou, G. Lin, and Y. Fu. Ai challenger: A large-scale dataset for going deeper in image understanding. In ICME, 2017. 2", + "[40] Ian T Young and Lucas J Van Vliet. Recursive implementation of the gaussian filter. Signal processing, 44(2):139-151, 1995. 3, 6", + "[41] Ailing Zeng, Xuan Ju, Lei Yang, Ruiyuan Gao, Xizhou Zhu, Bo Dai, and Qiang Xu. Deciwatch: A simple baseline for $10 \\times$ efficient 2d and 3d pose estimation. In ECCV, pages 607-624. Springer, 2022. 6", + "[42] Ailing Zeng, Lei Yang, Xuan Ju, Jiefeng Li, Jianyi Wang, and Qiang Xu. Smoothnet: A plug-and-play network for refining human poses in videos. In ECCV, pages 625-642, 2022. 2, 3, 6, 7" + ], + "bbox": [ + 78, + 90, + 467, + 770 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1833", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/e165eedf-6d86-4e7b-90c3-ccc6312e093e_model.json b/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/e165eedf-6d86-4e7b-90c3-ccc6312e093e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8cb566af3067907ef401e9428d31d54c7ac1dfc9 --- /dev/null +++ b/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/e165eedf-6d86-4e7b-90c3-ccc6312e093e_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.103, + 0.131, + 0.868, + 0.154 + ], + "angle": 0, + "content": "SynSP: Synergy of Smoothness and Precision in Pose Sequences Refinement" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.18, + 0.78, + 0.217 + ], + "angle": 0, + "content": "Tao Wang\\*, Lei Jin\\*, Zheng Wang?, Jianshu Li?, Liang Li?, Fang Zhao?, Yu Cheng?, Li Yuan?, Li Zhou?, Junliang Xing?, Jian Zhao?" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.219, + 0.746, + 0.234 + ], + "angle": 0, + "content": "1Beijing University of Posts and Telecommunications, 2Wuhan University, 3Ant Group," + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.236, + 0.85, + 0.252 + ], + "angle": 0, + "content": "\\(^{4}\\)Institute of computing technology Chinese Academy of Sciences, \\(^{5}\\)Nanjing University, \\(^{6}\\)National University of Singapore," + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.253, + 0.83, + 0.27 + ], + "angle": 0, + "content": "\\(^{7}\\)Peking University, \\(^{8}\\)Tsinghua University, \\(^{9}\\)China Telecom Institute of AI, \\(^{10}\\)Northwestern Polytechnical University" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.272, + 0.862, + 0.323 + ], + "angle": 0, + "content": "wangtao@bupt.edu.cn, jinlei@bupt.edu.cn, wangzwhu@whu.edu.cn, jianshu@u.NUS.edu, liang.li@ict.ac.cn, zhaofang0627@gmail.com, e0321276@u.nus.edu, yuanli-ece@pku.edu.cn, zhouli2025@126.com, jlxing@tsinghua.edu.cn, jian.zhao@nwpu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.357, + 0.314, + 0.373 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.474, + 0.782 + ], + "angle": 0, + "content": "Predicting human pose sequences via existing pose estimators often encounters various estimation errors. Motion refinement methods aim to optimize the predicted human pose sequences from pose estimators while ensuring minimal computational overhead and latency. Prior investigations have primarily concentrated on striking a balance between the two objectives, i.e., smoothness and precision, while optimizing the predicted pose sequences. However, it has come to our attention that the tension between these two objectives can provide additional quality cues about the predicted pose sequences. These cues, in turn, are able to aid the network in optimizing lower-quality poses. To leverage this quality information, we propose a motion refinement network, termed SynSP, to achieve a Synergy of Smoothness and Precision in the sequence refinement tasks. Moreover, SynSP can also address multi-view poses of one person simultaneously, fixing inaccuracies in predicted poses through heightened attention to similar poses from other views, thereby amplifying the resultant quality cues and overall performance. Compared with previous methods, SynSP benefits from both pose quality and multi-view information with a much shorter input sequence length, achieving state-of-the-art results among four challenging datasets involving 2D, 3D, and SMPL pose representations in both single-view and multi-view scenes. Github code: https://github.com/InvertedForest/SynSP." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.81, + 0.21, + 0.826 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.836, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Human pose estimation [4, 16, 21, 37] has been widely used in automatic driving, human-computer interaction, motion" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.357, + 0.891, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.457, + 0.885, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.557, + 0.871, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.65, + 0.895, + 0.832 + ], + "angle": 0, + "content": "Figure 1. Illustration of our motivation. (a): The x-axis represents the ratio of weights for Precision and Smoothness loss in the training process of SynSP, while the y-axis represents the value of MPJPE (Precision) and Acceleration error (Smoothness). These two errors exhibit a conflicting relationship. (b): Diff1 denotes the difference between the predicted sequence (from pose estimator) and the ground-truth sequence, and Diff2 denotes the difference between two output sequences from training SynSP with a loss ratio of 0.5 and 5, respectively. We notice that the trends of Diff1 and Diff2 are roughly the same in (b). (c): A larger value of Diff1 indicates lower prediction quality for the predicted sequences. For example, the left ankle of the person in the middle picture of (c) is occluded, making it difficult for the pose estimator to infer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "analysis, etc. Although current methods have achieved outstanding performance, both image-based and video-based pose estimators still sometimes exhibit noticeable errors in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*Both authors contributed equally to this research. † Lei Jin is corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1824" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.212 + ], + "angle": 0, + "content": "predicting pose sequences. The pose sequences generated by image-based pose estimators often show significant jitters due to a lack of inter-frame correlation, while video-based pose estimators alleviate jitters and improve accuracy by leveraging temporal information of input videos. However, training video-based pose estimators requires substantial amounts of video data and computational resources [7, 38, 39]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.216, + 0.47, + 0.442 + ], + "angle": 0, + "content": "Consequently, researchers propose motion refinement methods [3, 42] as post-processing techniques to optimize the predicted sequences from pose estimators to approximate the ground-truth sequences from human annotation, with two primary objectives: precision and smoothness. Precision objective aims to improve the position accuracy of the predicted joints through the temporal and spatial information, evaluated by Mean Per Joint Position Error (MPJPE) [11, 29]. Smoothness objective does not strive to make the output sequences as smooth as possible but rather to align the acceleration of the output with that of the ground-truth sequences. Both objectives aim to make the output sequences close to the ground-truth sequences. However, focusing on one of them will lead to errors in the other greatly increased, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.471, + 0.837 + ], + "angle": 0, + "content": "Due to the inherent tension between Precision and Smoothness objectives, previous methods attempt to achieve a satisfactory balance between them. As shown in Fig. 1(a), it can be observed that with an increase in the ratio of weights for Precision and Smoothness loss during SynSP training process, there is a noticeable rise in Acceleration error (Smoothness) while MPJPE (Precision) decreases, which is a common phenomenon in previous methods [42]. However, we note that the quality of the predicted sequences from estimator is related to the difference of outputs from SynSP biased towards the Precision and Smoothness objectives. Specifically, as illustrated in Fig. 1(b), the difference (Diff2) between the two output sequences, which focus on precision and smoothness respectively, is correlated with the quality of the predicted sequences, which is generated by the difference (Diff1) between predicted sequences from pose estimators and ground-truth sequences from human annotation. To further verify this correlation, we calculate the Pearson Correlation Coefficient between Diff1 and Diff2 on the Human3.6M and 3DPW datasets. Based on our experiment, the mean Pearson correlation coefficient [9] is 0.48 (bigger than 0.3 is considered as relevant), indicating that the dissimilarity (Diff2) between the two output sequences can serve as an indicator for the quality of predicted sequences without knowing the ground-truth." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose SynSP to utilize this quality information. Specifically, we employ a transformer network and encode the quality information by Pose Quality Encoding (PQE) module to guide the network's attention towards" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.196 + ], + "angle": 0, + "content": "low-quality poses in sequences. Additionally, with the support of multi-view dataset: Human3.6M, we further find that multi-view information is effective to enhance the reliability of PQE. Therefore, we introduce the Pose Similarity Encoding (PSE) module to exploit the multi-view information by guide the network's attention towards similar poses from multi-view." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.197, + 0.858, + 0.212 + ], + "angle": 0, + "content": "The main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.215, + 0.891, + 0.289 + ], + "angle": 0, + "content": "- We translate the inherent tension between smoothness and precision into the quality cue of predicted sequences with the Pose Quality Encoding module in SynSP. Then we utilize this cue as the weight to help regulate the network's attention on each frame in the sequences." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.29, + 0.891, + 0.349 + ], + "angle": 0, + "content": "- We further observe that similar poses can effectively rectify erroneous poses. Consequently, we propose Pose Similarity Encoding to encode pose similarity information and refine poses of multiple views simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.351, + 0.891, + 0.455 + ], + "angle": 0, + "content": "- We conduct sufficient experiments on four challenging datasets among 2D, 3D, and SMPL representations with image-based and video-based pose estimators. Notably, compared with previous method SmoothNet, SynSP's MPJPE improves by \\(16.9\\%\\) on Human3.6M dataset with 3D pose estimator FCN, while reducing input sequence length to \\(25\\%\\) and minimizing time delay to just \\(17\\%\\)." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.215, + 0.891, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.468, + 0.642, + 0.484 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Human Pose Estimation. Human pose estimation task [14-16, 27] is to extract location of human joints from the collected data. There are 2D, 3D, and SMPL representations for the human pose. 2D representation only contains 2D joint coordinates parallel to the camera plane. 3D representation has one more dimension of depth information. SMPL representation [23] is a skinned vertex based model that represents a wide variety of body shapes in natural human poses, while needs more parameters to represent." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Human pose estimators can be divided into image-based methods [27] and video-based methods [17]. Pose sequences from image-based methods often show a lot of jitters when these methods are applied to videos, while video-based methods require more data and computing resources for training. 2D pose estimator, Hourglass [27] found an outstanding stacked hourglass networks for human pose estimation; SPIN [19] can learn to reconstruct 3D human pose and SMPL pose via model-fitting in the loop; video-based pose estimators, TCMR [6] and VIBE [17], can generate smooth pose sequences with temporal information. We conduct the experiments based on pose estimators mentioned above." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Motion Refinement Methods. The predicted human pose sequences can be further refined by motion refinement methods, making the position and acceleration of each joint in these sequences are closer to those of the ground-truth sequences. The motion refinement methods can be divided to" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1825" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.873, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.317, + 0.896, + 0.403 + ], + "angle": 0, + "content": "Figure 2. The architecture of SynSP. FCN represents fully-connected network. In the base stage, with pose sequences of multiple views as input, SynSP first encodes these sequences individually in the Local Correlation Encoder, and merge multiple view's information in the Global Correlation Encoder with the Pose Similarity Encoding as their weight. After obtaining the individual embedding as the query, and incorporating the biased global information as the key and value, we get two outputs through the decoder module. In the refinement stage, we utilize these two detached outputs to generate Pose Quality Encoding and more precise Pose Similarity Encoding. By reusing embedding generated earlier, we achieve superior results using only one decoder with these encodings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.473, + 0.622 + ], + "angle": 0, + "content": "low-pass filters and deep learning based methods. The traditional low-pass filters, e.g., Savitzky-Golay [30], Gaussian1d [40], and One-Euro [3], are generally adapted to smooth the predicted pose sequences, and perform well for those abrupt jitters, such as confusion in left and right hips. Although the processed sequences from low-pass filter does look smoother, it is difficult to ensure precision at the same time. Deep learning based method, i.e., SmoothNet [42], takes the two objectives into account and aims to improve both precision and smoothness. Our SynSP is also based on deep learning and can achieve better results by considering the pose sequence quality and multi-view correlation information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.47, + 0.79 + ], + "angle": 0, + "content": "Additionally, there are also several human pose prior methods [2, 22, 28, 32], which try to learn a prior distribution of valid human poses. Although time-consuming for their sampling steps, these methods can be applied to the task of pose generation/completion/refinement. Hu-MoR [31] introduces a novel conditional VAE which enables expressive and general motion reconstruction and generation; GAN-based method [10] proposes a simple yet effective prior for SMPL model to bound it to realistic human poses; GFPose [8] employs a novel score-based generative framework to model plausible 3D human poses." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.806, + 0.168, + 0.822 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.832, + 0.187, + 0.846 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Task Description. Given predicted pose sequences generated by human pose estimators [16], motion refinement methods aim to fit these predicted sequences to the ground" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.892, + 0.577 + ], + "angle": 0, + "content": "truth sequences, making these prediction more precise and smooth. Usually, the predicted pose sequence \\(\\mathbf{P}\\) has a shape of \\(\\mathbb{R}^{V\\times T\\times (J\\times D)}\\), where \\(V\\) stands for the number of views for one individual; \\(T\\) refers to the frame number of these predicted pose sequences; \\(J\\) is the number of joints associated with datasets; \\(D\\) denotes the spatial dimensions of these joints. We use \\(P_{(v,t,j)}\\) to represent the \\(j_{th}\\) joint's position at \\(t_{th}\\) frame from \\(v_{th}\\) view throughout the paper. The spatial dimension is included in the joint dimension, ignoring the spatial dimensions to simplify it." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.893, + 0.789 + ], + "angle": 0, + "content": "Overall Architecture. We proposed SynSP to use this quality information from the certain tension relationship between the two objectives, i.e., smoothness and precision. We also introduce multi-view information fusion as an optional enhancement for SynSP. As illustrated in Fig. 2, SynSP is mainly based on the transformer structure [34] and composed of two stages. Firstly, under the guidance of Pose Similarity Encoding I, Base Decoder can utilize both individual and global information to generate two branches, which are respectively biased towards smoothness and precision. Next, these two branches can generate Pose Quality Encoding, and a more precise Pose Similarity Encoding II, which further helps Refinement Decoder generate refined results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.62, + 0.817 + ], + "angle": 0, + "content": "3.2. Base Stage" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.87 + ], + "angle": 0, + "content": "In the base stage, SynSP takes predicted sequences from pose estimators as input and generates two sequences that focus on smoothness and precision, respectively." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Pose Similarity Encoding. Similarity for poses from multiview is effective in enhancing the performance of SynSP." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1826" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "On one hand, similar poses from different views can complement each other and improve the robustness of joint position. On the other hand, this encoding minimizes the network's attention on poses with significant distortions to reduce their impact for the refinement task." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.168, + 0.468, + 0.183 + ], + "angle": 0, + "content": "Pose Similarity Encoding (PSE) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.196, + 0.469, + 0.255 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P _ {(v, t, j)} ^ {r e l} = \\hat {P} _ {(v, t, j)} - P _ {(v, t, c)}, \\\\ P S E _ {\\left(v _ {1}, v _ {2}, t\\right)} = \\exp \\left(- \\sum_ {j = 1} ^ {J} \\left(P _ {\\left(v _ {1}, t, j\\right)} ^ {r e l} - P _ {\\left(v _ {2}, t, j\\right)} ^ {r e l}\\right) ^ {2}\\right), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.469, + 0.335 + ], + "angle": 0, + "content": "where \\( c \\) represents the index of the \"pelvis\" keypoint in the joint list. \\( v_{1} \\) and \\( v_{2} \\) are used only for distinction, and both have the same domain as \\( v \\). Overall, PSE has a shape of \\( \\mathbb{R}^{V\\times V\\times T} \\), and we present a visualization example of PSE for one frame in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.469, + 0.472 + ], + "angle": 0, + "content": "Global Correlation Encoder. In multi-view scenes, the correlation information between multiple views can help SynSP to deduce the location of the joint with jitters globally. Global Correlation Encoder (GCE) module is introduced to produce the keys and values required by the Base Decoder with the global correlation information between multiple views. First, the predicted sequences \\(\\mathbf{P}\\) are projected to embedding \\(\\mathbf{E}\\) through a full connected layer along the joint dimension." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.469, + 0.669 + ], + "angle": 0, + "content": "Here, we use \\( E_{(v,t,e)} \\) to denote an item of projected embedding \\( \\mathbf{E} \\), where \\( e \\) refers to the index of the embedding dimension. Then, the view (v) and the frame (t) dimensions of \\( \\mathbf{E} \\) are merged, so that pose information from multiple views and temporal information can be sensed by the encoder simultaneously. Afterwards, \\( E_{((v,t),e)} \\) is sent to a encoder module of transformer [34] to interact pose sequences from multiple views and get the global information \\( E_{((v,t),e)}^g \\). Then this information is added with \\( PSE_{(v_1,v_2,t)} \\) to adjust \\( v_{1th} \\) view's attention to other views' information by broadcast operation: \\( E_{((v_2,t),e)}^g + PSE_{(v_1,v_2,t)} = E_{(1,(v_2,t),e)}^g + PSE_{(v_1,(v_2,t),1)} = E_{(v_1,(v_2,t),e)}^g \\), which is the final global information from GCE module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.469, + 0.835 + ], + "angle": 0, + "content": "Local Correlation Encoder. Local Correlation Encoder (LCE) module focuses on the temporal dimension with single-view and produces the query required by the Base Decoder. LCE reuses the projected embedding \\(\\mathbf{E}\\) from the full connected layer mentioned in GCE as input, and directly sends it to another standard encoder module of transformer [34] to obtain local information in the frame (t) dimension (take the view (v) dimension as input batch). Finally, for \\(v_{1th}\\) view, \\(t_{th}\\) frame, \\(e_{th}\\) item, LCE produces the local information \\(E_{(v_1,t,e)}^l\\) with exploring the correlation of the time dimension." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.838, + 0.469, + 0.905 + ], + "angle": 0, + "content": "Base Decoder. At the end of base stage, we stack 5 standard decoding layers of transformer [34] into Base Decoder. The Base Decoder takes the local information \\( E_{(v_1,t,e)}^l \\) as the query, to interact with global information \\( E_{(v_1,(v_2,t),e)}^g \\)," + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.091, + 0.673, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.21, + 0.664, + 0.221 + ], + "angle": 0, + "content": "Pose Similarity Encoding" + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.092, + 0.871, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.21, + 0.858, + 0.221 + ], + "angle": 0, + "content": "Pose Quality Encoding" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.23, + 0.892, + 0.286 + ], + "angle": 0, + "content": "Figure 3. Visualization of the two encodings of one frame. \\(PSE\\): the diagonal has a constant value of 1. \\(PQE\\): the value outside the diagonal is constant 0. In this way, the addition of the two encodings cannot affect each other." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.892, + 0.374 + ], + "angle": 0, + "content": "which serves as keys and values for the Base Decoder. At the last layer of Base Decoder, we use two decoder layers to generate two branches biased towards smoothness and precision respectively at the same time." + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.384, + 0.891, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} K, V = E _ {(v _ {1}, (v _ {2}, t), e)} ^ {g}, Q ^ {0} = E _ {(v _ {1}, t, e)} ^ {l}, \\\\ Q _ {(v _ {1}, t, e)} ^ {d _ {t h}} = \\mathbf {D L a y e r} ^ {d _ {t h}} \\left(Q ^ {(d - 1) _ {t h}}, K, V\\right), d \\in \\{1, 2, 3 \\}, \\tag {2} \\\\ P _ {(v _ {1}, t, j)} ^ {\\text {p o s}} = \\operatorname {F C N} \\left(\\mathbf {D L a y e r} ^ {4 _ {\\mathrm {t h}}} \\left(\\mathrm {Q} ^ {3 _ {\\mathrm {t h}}}, \\mathrm {K}, \\mathrm {V}\\right)\\right), \\\\ P _ {(v _ {1}, t, j)} ^ {a c c} = \\operatorname {F C N} \\left(\\mathbf {D L a y e r} ^ {5 _ {\\mathrm {t h}}} \\left(\\mathrm {Q} ^ {3 _ {\\mathrm {t h}}}, \\mathrm {K}, \\mathrm {V}\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.892, + 0.623 + ], + "angle": 0, + "content": "where \\(Q\\), \\(K\\), and \\(V\\) denote the query, key, and value for Base Decoder respectively, \\(\\mathbf{DLayer}^{d_{th}}\\) denotes the \\(d_{th}\\) decoder layer, \\(\\mathrm{FCN}(\\cdot)\\) is fully connected network. In this way, SynSP can generate two pose squences \\(P^{pos}\\) and \\(P^{acc}\\) from its Acceleration Branch and Position Branch, which are biased towards the precision and smoothness objectives in the final loss function Equ. (4), respectively. Next, we can also further optimize these sequences in the refinement stage, with a small amount of calculation and parameters as the cost." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.635, + 0.671, + 0.651 + ], + "angle": 0, + "content": "3.3. Refinement Stage" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.891, + 0.703 + ], + "angle": 0, + "content": "In the refinement stage, we aim to seek a better way to fuse Acceleration Branch and Position Branch, which concentrate on different objectives." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Acceleration and Position Branches. The output sequences from Acceleration Branch and Position Branch are detached from computation graph for a stabler training process as the predictions from base stage are independent of the refinement stage. These outputs can not only generate more accurate Pose Similarity Encoding, but also provide the quality information about these pose sequences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Pose Quality Encoding. We have empirically observed that when outputs from the two branches show a greater discrepancy, the prediction quality is poorer. Specifically, we computed the Pearson Correlation Coefficient between the prediction quality and the difference between two outputs is about 0.48 (more details in the supplementary materials)," + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1827" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.471, + 0.152 + ], + "angle": 0, + "content": "which shows that the two have a strong correlation. Hence, we propose the Pose Quality Encoding (PQE) to utilize this correlation to improve the final prediction. Each item of PQE is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.16, + 0.329, + 0.182 + ], + "angle": 0, + "content": "\\[\nP _ {(v _ {1}, v _ {2}, t, j)} ^ {d i s} = | P _ {(v _ {1}, t, j)} ^ {a c c} - P _ {(v _ {2}, t, j)} ^ {p o s} |,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.183, + 0.469, + 0.223 + ], + "angle": 0, + "content": "\\[\nP Q E _ {(v _ {1}, v _ {2}, t)} ^ {\\prime} = \\frac {1}{J} \\sum_ {j = 1} ^ {J} \\exp \\left(- \\frac {P _ {(v _ {1} , v _ {2} , t , j)} ^ {d i s}}{\\max _ {j \\in J} P _ {(v _ {1} , v _ {2} , t , j)} ^ {d i s} + \\sigma}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.226, + 0.378, + 0.256 + ], + "angle": 0, + "content": "\\[\nP Q E _ {(v _ {1}, v _ {2}, t)} = \\left\\{ \\begin{array}{c c} P Q E _ {(v _ {1}, v _ {2}, t)} ^ {\\prime}, & v _ {1} = v _ {2}; \\\\ 0, & v _ {1} \\neq v _ {2}. \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.267, + 0.469, + 0.372 + ], + "angle": 0, + "content": "Here, we calculate PQE by the distance \\( P^{dis} \\) between the outputs \\( P^{acc} \\) and \\( P^{pos} \\), which are generated by Acceleration Branch and Position Branch, respectively. To ensure that extreme values do not interfere with PQE, \\( P^{dis} \\) is normalized by dividing the maximum value in the individual's sequence. We introduce a small constant \\( \\sigma \\) (set to \\( 1 \\times 10^{-6} \\)) to avoid division by zero." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.469, + 0.493 + ], + "angle": 0, + "content": "Unlike PSE, which is calculated based on correlations among multiple views on one person, PQE is computed by \\( P^{dis} \\) for a single view across multiple frames. Since PQE and PSE have distinct considerations regarding multiple views versus single view, they can be combined through addition rather than concatenation without mutual influence as illustrated in Fig. 3, thereby enhancing the computational efficiency of SynSP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.469, + 0.674 + ], + "angle": 0, + "content": "Refinement Decoder. The output sequences of the two branches exhibit enhanced accuracy and smoothness compared to the input sequences. A straightforward approach would involve averaging these outputs and replicating the base stage pipeline once again. However, our experimental findings indicate that this approach does not yield significant benefits but doubles the computational cost. Please refer to supplementary material for details. This may be attributed to the fact that while the optimized output is indeed smooth and precise, it lacks crucial positional information from the original jitter input, rendering optimization of an already optimized output meaningless." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.469, + 0.84 + ], + "angle": 0, + "content": "To circumvent this issue and maintain a lightweight model, we reuse the embedding from the Global Correlation Encoder and Base Decoder. The former embedding retains the original position information, while the latter embedding contains optimization information about the output. In this way, we can produce superior outcomes by using just one additional decoder. We stack 4 standard decoding layers of transformer [34] in Refinement Decoder, and the inner calculation of Refinement Decoder is similar with Base Decoder. As shown in Tab. 5, this approach effectively improves the positional accuracy of pose sequences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Analysis on the Refinement Stage. In the refinement stage, the difference between the position and acceleration branches serves as a guiding factor for the quality of predicted sequence. Next, within the refinement decoder," + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.089, + 0.85, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.188, + 0.892, + 0.244 + ], + "angle": 0, + "content": "Figure 4. Illustration of Sliding Window Average Algorithm (SWAA), which is widely used by motion refinement works. This figure proves that the amount of computation and latency is proportional to the window length." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.271, + 0.892, + 0.408 + ], + "angle": 0, + "content": "greater emphasis is placed on frames with lower pose quality to facilitate improvements in the final performance. Regarding why this difference is associated with pose quality, we posit that the two branches are easy to remain consistent under the normal circumstances, nevertheless, in challenging scenarios such as occlusion or sudden movement, it is more likely for the two branches to exhibit dissimilar behavior. Whether in pose estimator or motion refinement model, pose quality is is obviously related with scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.418, + 0.715, + 0.433 + ], + "angle": 0, + "content": "3.4. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.441, + 0.892, + 0.532 + ], + "angle": 0, + "content": "Different from previous work, we only use the window length \\( T = 8 \\) for training and inference. Compared with the window length of 32, this setting reduces the time delay of \\( (32 - 8) / 12 = 2s \\) while the inference speed of pose estimator is 12fps, and have \\( 8 / 32 = 0.25 \\) times the previous amount of computation as shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.533, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Training. The losses of the three outputs, i.e., Acceleration Branch output, Position Branch output, and final output, share the same calculation method. We have two losses, \\(\\mathcal{L}_{acc}\\) and \\(\\mathcal{L}_{pos}\\), for each one of the three outputs:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.604, + 0.891, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} A _ {(v, t, j)} = \\left(P _ {(v, t + 2, j)} - P _ {(v, t + 1, j)}\\right) - \\left(P _ {(v, t + 1, j)} - P _ {(v, t, j)}\\right), \\\\ \\mathcal {L} _ {\\mathrm {p o s}} = \\frac {1}{V \\cdot T \\cdot J} \\sum_ {v = 1} ^ {V} \\sum_ {t = 1} ^ {T} \\sum_ {j = 1} ^ {J} | P _ {(v, t, j)} - \\hat {P} _ {(v, t, j)} |, \\\\ \\mathcal {L} _ {\\mathrm {a c c}} = \\frac {1}{V \\cdot T \\cdot J} \\sum_ {v = 1} ^ {V} \\sum_ {t = 1} ^ {T} \\sum_ {j = 1} ^ {J} | A _ {(v, t, j)} - \\hat {A} _ {(v, t, j)} |, \\\\ \\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p o s}} + \\lambda \\mathcal {L} _ {\\text {a c c}}. \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.87 + ], + "angle": 0, + "content": "In our experiments, \\(\\lambda\\) is set as 0, 1, and 0.5 for Position Branch, Acceleration Branch, and the final output, respectively, which is a trade-off between smoothness and precision. For all branches, \\(\\mathcal{L}_{\\mathrm{pos}}\\) is crucial to ensuring training stability. Furthermore, to balance \\(\\mathcal{L}_{\\mathrm{pos}}\\) and \\(\\mathcal{L}_{\\mathrm{acc}}\\), the parameter \\(\\lambda\\) should be set as 0.5, determined by the absolute value of each element. Thus, for the final output, we have set \\(\\lambda\\) to 0.5, whereas in the Acceleration branch, \\(\\lambda\\) is assigned as 1 to prioritize smoothness." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Inference. During inference, we obtain outputs from the refinement stage. Then the outputs go through a" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1828" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.075, + 0.895, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison with related works on Human3.6M, AIST++ datasets with 2D, 3D, and SMPL human pose estimator Hourglass [27], FCN [26], and SPIN [19], respectively. WS denotes the windows size in training, inference, and post-process. * indicates that SynSP is trained and tested in multi-view scenes." + }, + { + "type": "table", + "bbox": [ + 0.103, + 0.128, + 0.872, + 0.284 + ], + "angle": 0, + "content": "
MethodWSHuman3.6M / 2DHuman3.6M / 3DAIST++ / SMPL
MPJPE↓PA-MPJPE↓Accel↓MPJPE↓PA-MPJPE↓Accel↓MPJPE↓PA-MPJPE↓Accel↓
InputN/A9.427.641.5454.5542.219.17107.7274.4033.20
One-Euro [3]110.697.980.3455.2042.733.80108.9775.2714.70
Gaussian1d [40]329.377.560.5153.6741.602.43104.8472.1810.05
Savitzky-Golay [30]329.357.550.1753.4841.191.34104.5872.306.07
SmoothNet [42]329.257.570.1552.7240.921.03103.0071.195.72
SynSP88.136.090.1551.3640.131.0284.6359.026.08
SynSP*87.625.640.1541.7833.320.98---
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.471, + 0.413 + ], + "angle": 0, + "content": "post-processing called sliding window average algorithm adopted by most works as shown in Fig. 4 (we have accelerated it by parallel computing to \\(1\\%\\) of the original latency, refer to supplementary material for more details). It can be seen that the delay and computation needs for model and post-processing becomes smaller with shorter window length." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.47, + 0.52 + ], + "angle": 0, + "content": "Single-View Scenes. While \\( V = 1 \\) in the equations above, SynSP can be directly applied to adapt single-view scenes. In single-view scenes, PSE module is no longer needed. Since multi-view scenes are not considered by current related works and not our main motivation for the proposed SynSP, we conduct sufficient single-view experiments for a comprehensive comparison." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.535, + 0.21, + 0.551 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.271, + 0.577 + ], + "angle": 0, + "content": "4.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Dataset. To demonstrate the generalizability and efficiency of our method, we conduct experiments of SynSP in three human pose representations of 2D, 3D, and SMPL for single-view and multi-view scenes on Human3.6M [12], AIST++ [20], 3DPW [35], and CMU-Mocap [33] datasets. For the fairness of the comparison, we use these datasets made by SmoothNet [42] with kinds of pose estimators in the experiments. To compare with pose prior models, we also follow GFPose [8] to add the same degree of Gaussian noise or uniformly distributed noise to Human3.6M and CMU-Mocap datasets as input for the pose denoise task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.84 + ], + "angle": 0, + "content": "Metrics. Following the related works [42], we use Accel error, MPJPE, and PA-MPJPE (Procrustes-aligned MPJPE) [1, 5] to evaluate the acceleration and position proximity between the predicted sequence and the ground-truth sequence, and the lower metrics the smaller error of the predicted sequence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. SynSP consists of two 4-layer encoder modules (GCE, LCE), a 5-layer decoder module (Base Decoder), and a 4-layer decoder module (Refinement Decoder). We initialize the learning rate as 0.001 and decay" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.895, + 0.4 + ], + "angle": 0, + "content": "it by a factor of 0.95 per epoch. SynSP is trained for 70 epochs with Adam optimizer, requiring approximately two hours and 7GB of GPU memory with a batch size of 4096. The computational setup consists of an Intel(R) Xeon(R) Gold 6271C CPU @ 2.60GHz for the CPU and a Tesla V100 SXM2 32GB for the GPU." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.413, + 0.82, + 0.429 + ], + "angle": 0, + "content": "4.2. Comparison with State-Of-The-Arts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.894, + 0.558 + ], + "angle": 0, + "content": "In our experiments, we first demonstrate that SynSP consistently outperforms previous methods in both single-view and multi-view scenes with several pose estimators. Then we compare with pose prior models on the pose denoise task by smoothing the pose sequences with artificial noisy. (There are also some methods[13, 41] with a super-long window length, and the comparison with them is shown in the supplementary materials.)" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.574, + 0.894, + 0.617 + ], + "angle": 0, + "content": "Table 2. Comparison of various video-based pose estimators. + represents the combination of estimators and motion refinement methods. * indicates multi-view inputs." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.628, + 0.894, + 0.899 + ], + "angle": 0, + "content": "
DatasetMethodWSMPJPE↓Accel↓
3DPWMAED [36]1679.00-
MPS-Net [38]1684.30-
TCMR [6]1686.466.75
TCMR+SmoothNet [42]16+3286.506.00
TCMR+SynSP16+886.105.90
PARE [18]179.0025.60
PARE+SmoothNet [42]1+3278.105.91
PARE+SynSP1+876.206.16
AIST++VIBE [17]16106.931.60
VIBE+SmoothNet [42]16+3297.474.15
VIBE+SynSP16+877.004.32
H36MTransFusion* [24]125.52-
PPT* [25]125.1611.60
PPT*+SmoothNet [42]1+3223.971.31
PPT*+SynSP1+821.511.10
" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1829" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.146 + ], + "angle": 0, + "content": "Table 3. Comparison on Human3.6M dataset for the denoise task. \\(N(0,\\sigma)\\) represents a Gaussian distribution with a mean of 0 and a variance of \\(\\sigma\\), and \\(U(k)\\) represents a uniform distribution at \\([0,k]\\). \\(\\mathbf{M}\\downarrow\\) and \\(\\mathrm{A}\\downarrow\\) denote MPJPE↓ and Accel↓ respectively." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.156, + 0.472, + 0.261 + ], + "angle": 0, + "content": "
MethodN(0,100)N(0,400)U(50)U(100)
M↓A↓M↓A↓M↓A↓M↓A↓
Noisy input65.6160.3131.1320.594.8231.3189.5462.7
GFpose [8]42.8-64.6-50.9-89.4-
SmoothNet [42]42.410.757.914.249.410.274.316.3
SynSP37.68.856.113.045.610.969.415.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.274, + 0.471, + 0.332 + ], + "angle": 0, + "content": "Table 4. Comparison on CMU-Mocap dataset for the denoise task. \\( N(0,\\sigma) \\) represents a Gaussian distribution with a mean of 0 and a variance of \\( \\sigma \\), and \\( U(k) \\) represents a uniform distribution at [0,k]. \\( \\mathbf{M} \\downarrow \\) and \\( \\mathbf{A} \\downarrow \\) denote MPJPE \\( \\downarrow \\) and Accel \\( \\downarrow \\) respectively." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.342, + 0.472, + 0.446 + ], + "angle": 0, + "content": "
MethodN(0,100)N(0,400)U(50)U(100)
M↓A↓M↓A↓M↓A↓M↓A↓
Noisy input61.2143.8122.5287.688.7207.9177.4415.8
GFpose [8]40.4-60.0-45.6-84.4-
SmoothNet [42]35.14.341.54.937.74.244.95.1
SynSP20.63.728.04.424.44.132.44.6
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.672 + ], + "angle": 0, + "content": "2D Representation. In Tab. 1, we compare SynSP with other motion refinement methods in Human3.6M dataset with 2D pose estimator Hourglass [27]. With the setting of 8 window length, our SynSP outperforms all the methods while it is about 2 seconds faster (except the One-Euro) according to Fig. 4. Compared with SmoothNet, which is trained on multiple datasets, i.e., AIST-VIBE-3D, 3DPW-SPIN-3D, and Human3.6M-FCN-3D, SynSP can still outperforms it by \\(16.9\\%\\) for MPJPE and \\(25.2\\%\\) for PA-MPJPE with smaller window size. According to the ablation experiment results in Sec. 4.3, the significant accuracy improvement of SynSP is attributed to the mutual patching between the pose quality information (PQE) and multi-view information (PSE)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.809 + ], + "angle": 0, + "content": "3D Representation. In Tab. 1, we also compare SynSP with other motion refinement methods in Human3.6M dataset with 3D pose estimator FCN [26]. Similar with the comparison in 2D Representation, we can still outperforms SmoothNet by \\(20.8\\%\\) for MPJPE and \\(18.6\\%\\) for PA-MPJPE in Human3.6M dataset. There is also a decrease of \\(4.9\\%\\) for the Accel error. Furthermore, SynSP obtains these results only with 8 window size, while SmoothNet and most filters need 32 frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.87 + ], + "angle": 0, + "content": "SMPL Representation. Then we compare SynSP with other motion refinement methods in AIST++ dataset with SMPL pose estimator SPIN as shown in Tab. 1. SynSP shows superior performance in SMPL representation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Comparison with Video-based Pose Estimators. The video-based estimators already include temporal informa" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.317 + ], + "angle": 0, + "content": "tion. Compared with the image-based estimators, motion refinement methods have difficult to optimize the sequences from video-based estimators. In Tab. 2, we compare SynSP with SOTA methods and SmoothNet in AIST++ and 3DPW datasets with 3D pose video-based estimator VIBE [17] and TCMR [6], respectively. We can find that the high performance video-based estimators, TCMR, is hard to be refined. SmoothNet improves the smoothness by processing a longer window length than the video-based estimators. SynSP greatly increases precision and smoothness in a shorter window length with its outstanding architecture as shown in Fig. 2. The image-based estimator, PARE [18], has excellent performance in precision, SynSP can also make up for its shortcomings in smoothness and outperform some video-based estimators." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.455 + ], + "angle": 0, + "content": "Comparison with Multi-view Pose Estimators. We further verify the performance of SynSP with the multi-view pose estimator, i.e., PPT [25]. As shown in Tab. 2, we can notice that SynSP can further improve the precision of PPT, which has superiority performance by considering multiview connections. We also present the result of other multiview methods such as TransFusion for comparison. This experiment results further proves the multi-view refinement ability of PSE for video-based estimators." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.892, + 0.743 + ], + "angle": 0, + "content": "Comparison with Pose Priors Methods. To ensure a fair comparison with human pose prior models, we also conducted denoise experiments following the methodology of the SOTA work [8] among pose priors methods. Here, window size of GFPose, SmoothNet, and SynSP are 1, 32, and 8 respectively, while GFPose, as a generative model for human pose prior task, has 1000 sampling steps, i.e., 1000 times of model inference. GFPose has not presented its Accel errors since it does not take temporal information into account. As presented in Tab. 3 and Tab. 4, SynSP can effectively solve MPJPE (Precision) problems. Across noisy Human3.6M and CMU-Mocap datasets, SynSP can improve the pose quality by about \\(54.5\\%\\) in terms of MPJPE and \\(96.0\\%\\) in terms of Accel errors for Gaussian noise with a variance of 100. For Gaussian noise with a large variance of 400, SynSP can improve the pose quality by about \\(67.2\\%\\) in terms of MPJPE and \\(97.2\\%\\) int terms of Accel errors. We also observe consistent improvement in uniform noise as shown in Tab. 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.657, + 0.771 + ], + "angle": 0, + "content": "4.3. Ablation Study." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.824 + ], + "angle": 0, + "content": "To ensure representativeness of ablation experiments, all experiments were performed on the largest dataset, Human3.6M, with 3D estimator FCN." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Modules of SynSP. In order to verify the rationality of the structure of SynSP, we conducted ablation experiments on each module in SynSP. In Tab. 5, the line without a tick represents the results of baseline, which is the network of base stage without PSE I. And this network surpasses Smooth-" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1830" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.089, + 0.472, + 0.133 + ], + "angle": 0, + "content": "Table 5. Ablation study on PSE I, RS (Refinement Stage), PSE II, and PQE modules. * denotes directly using the output from base stage as input of refinement stage to re-optimize these sequence." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.143, + 0.468, + 0.307 + ], + "angle": 0, + "content": "
PSE IRSPSE IIPQEMPJPEAccel
51.21.09
✓48.71.07
✓*51.41.08
✓49.61.07
✓✓46.71.05
✓✓✓46.41.04
✓✓✓42.50.99
✓✓✓✓41.80.98
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.47, + 0.536 + ], + "angle": 0, + "content": "Net in precision by \\(2.8\\%\\). PSE I, RS, PSE II, and PQE all have gains in results, and the combination of PSE I, RS, and PQE has brought greater benefits. We believe that this benefit mainly comes from the combination of SynSP with global information PSE I, which can help generate more accurate quality information according to the poses from other views. Besides, utilizing the difference between acceleration and Position Branch improves both smoothness and precision of final prediction. It is worth noting that using the average value of the two branches as the input RS does not bring reasonable gains. We suspect that the output of base stage loses the original predicted position information, which affects the accuracy and makes the refinement stage useless compared with the baseline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.869 + ], + "angle": 0, + "content": "Time Efficiency Analysis. We conducted latency experiments on the Human3.6M dataset using FCN [26] as the 3D pose estimator with a batch size of 1416, as shown in Tab. 6. Considering the practical application environment, we report not only the inference latency of the model, as shown in the 3rd column of Tab. 6, but also the latency associated with the Sliding Window Average Algorithm (SWAA), which includes the delay in waiting for pose estimators and the computation cost of SWAA (we have optimized SWAA specifically for parallel computation on GPUs). The bold cell in each row of Tab. 6 represents the primary factor contributing to latency. For video processing speed of 12 frames per second for the pose estimator (based on the general inference speed of various pose estimators), a window size of 8 would require a waiting time of \\(8/12 = 667\\mathrm{ms}\\); similarly, a window size of 32 would require a waiting time of \\(32/12 = 2667\\mathrm{ms}\\). Furthermore, the model inference time for SynSP is only about 10ms. Overall, the main time cost primarily arises from the internal waiting latency for SmoothNet and SynSP. For the pose priors method GFPose, the main time delay is its 1000 sampling steps. Hence, SynSP is more suitable for near real-time pose refinement task." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Analysis on Number of Views. As shown in Tab. 7, as more poses from different views are input at the same time" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.132 + ], + "angle": 0, + "content": "Table 6. Time efficiency analysis on SynSP including model inference time and time cost for sliding window. Besides, we assume that FPS (Frame Per Second) for pose estimator is 12." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.143, + 0.897, + 0.248 + ], + "angle": 0, + "content": "
MethodsWSModel (ms)SWAATotal (ms)
Wait (ms)Execution (ms)
GFPose13772--3772
SmoothNet321.2126678023470
SynSP811.06670.3678
SynSP*812.36670.3680
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.261, + 0.892, + 0.332 + ], + "angle": 0, + "content": "Table 7. Contrast experiments on the number of input views. \\(\\dagger\\) means that we randomly combine several people together for input during the training process, instead of using multiple views for one person. * means that in the case of \\(\\dagger\\), several people with similar poses are input at the same time during the inference process." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.343, + 0.892, + 0.464 + ], + "angle": 0, + "content": "
View NumberMPJPEPA-MPJPEAccel
151.440.11.02
247.237.61.02
346.537.11.00
441.833.30.98
4†50.439.51.01
4*49.739.11.01
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.581 + ], + "angle": 0, + "content": "during the training process and inference process, the performance of SynSP is more efficient. It is justifiable due to the introduction of more global information. If number of views is larger than 4, we believe that the performance of our model can be further promoted with only a marginal increase in model complexity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.594, + 0.618, + 0.61 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.619, + 0.892, + 0.786 + ], + "angle": 0, + "content": "This paper proposes SynSP, a motion refinement method that utilizes the tension relationship between smoothness and precision through synergistic optimization. Specifically, SynSP outputs Position Branch and Acceleration Branch biased towards precision and smoothness, respectively. Then we employ the discrepancy between the two branches as a quality cue for the predicted sequence, which can be utilized in subsequent stages to further refine the pose sequence. In addition, we introduce Pose Similarity Encoding, achieving simultaneous refinement of multi-view poses and performance enhancement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.652, + 0.817 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The paper is supported by National Key R & D Program of China (No. 2021YFB3300100), Natural Science Foundation of China No.62102039, the Fundamental Research Funds for the Central Universities No.500422813, and Natural Science Foundation of China No.6232211." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1831" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] A. Benzine, F. Chabot, B. Luvison, Q. C. Pham, and C. Achard. Pandanet: Anchor-based single-shot multi-person 3d pose estimation. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.158, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In ECCV, pages 561-578, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.47, + 0.282 + ], + "angle": 0, + "content": "[3] Géry Casiez, Nicolas Roussel, and Daniel Vogel. 1€ filter: a simple speed-based low-pass filter for noisy input in interactive systems. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pages 2527-2530, 2012. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.47, + 0.325 + ], + "angle": 0, + "content": "[4] Yu Cheng, Bo Yang, Bo Wang, and Robby T Tan. 3d human pose estimation using spatio-temporal networks with explicit occlusion training. In AAAI, pages 10631-10638, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.47, + 0.38 + ], + "angle": 0, + "content": "[5] Yu Cheng, Bo Wang, Bo Yang, and Robby T. Tan. Monocular 3d multi-person pose estimation by integrating top-down and bottom-up networks. In CVPR, pages 7649-7659, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.436 + ], + "angle": 0, + "content": "[6] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consistent 3d human pose and shape from a video. In CVPR, pages 1964-1973, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.47, + 0.493 + ], + "angle": 0, + "content": "[7] Jiaming Chu, Lei Jin, Xiaojin Fan, Yinglei Teng, Yunchao Wei, Yuqiang Fang, Junliang Xing, and Jian Zhao. Single-stage multi-human parsing via point sets and center-based offsets. In ACM MM, pages 1863-1873, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.47, + 0.548 + ], + "angle": 0, + "content": "[8] Hai Ci, Mingdong Wu, Wentao Zhu, Xiaoxuan Ma, Hao Dong, Fangwei Zhong, and Yizhou Wang. Gfpose: Learning 3d human pose prior with gradient fields. CVPR, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.468, + 0.606 + ], + "angle": 0, + "content": "[9] Israel Cohen, Yiteng Huang, Jingdong Chen, Jacob Benesty, Jacob Benesty, Jingdong Chen, Yiteng Huang, and Israel Cohen. Pearson correlation coefficient. Noise reduction in speech processing, pages 1-4, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.468, + 0.661 + ], + "angle": 0, + "content": "[10] Andrey Davydov, Anastasia Remizova, Victor Constantin, Sina Honari, Mathieu Salzmann, and Pascal Fua. Adversarial parametric pose prior. In CVPR, pages 10997-11005, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[11] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Stefano Alletto, and Rita Cucchiara. Compressed volumetric heatmaps for multi-person 3d pose estimation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.468, + 0.76 + ], + "angle": 0, + "content": "[12] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE TPAMI, 36(7):1325-1339, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[13] Kyung-Min Jin, Byoung-Sung Lim, Gun-Hee Lee, Tae-Kyung Kang, and Seong-Whan Lee. Kinematic-aware hierarchical attention network for human pose estimation in videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5725-5734, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[14] Lei Jin, Xiaojuan Wang, Xuecheng Nie, Luoqi Liu, Yandong Guo, and Jian Zhao. Grouping by center: Predicting centripetal offsets for the bottom-up human pose estimation. IEEE TMM, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Lei Jin, Chenyang Xu, Xiaojuan Wang, Yabo Xiao, Yandong Guo, Xuecheng Nie, and Jian Zhao. Single-stage is enough: Multi-person absolute 3d pose estimation. In CVPR, pages 13086-13095, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[16] Lei Jin, Xiaojuan Wang, Xuecheng Nie, Wendong Wang, Yandong Guo, Shuicheng Yan, and Jian Zhao. Rethinking the person localization for single-stage multi-person pose estimation. IEEE TMM, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[17] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In CVPR, pages 5253-5263, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[18] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. In ICCV, pages 11127-11137, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[19] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[20] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In ICCV, pages 13401-13412, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[21] Yang Li, Kan Li, Shuai Jiang, Ziyue Zhang, Congzhen-tao Huang, and Richard Yi Da Xu. Geometry-driven self-supervised method for 3d human pose estimation. In AAAI, pages 11442-11449, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[22] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Transactions on Graphics (TOG), 39(4):40-1, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[23] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM TOG, 34(6):1-16, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[24] Haoyu Ma, Liangjian Chen, Deying Kong, Zhe Wang, Xingwei Liu, Hao Tang, Xiangyi Yan, Yusheng Xie, Shih-Yao Lin, and Xiaohui Xie. Transfusion: Cross-view fusion with transformer for 3d human pose estimation. arXiv preprint arXiv:2110.09554, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[25] Haoyu Ma, Zhe Wang, Yifei Chen, Deying Kong, Liangjian Chen, Xingwei Liu, Xiangyi Yan, Hao Tang, and Xiaohui Xie. Ppt: Token-pruned pose transformer for monocular and multi-view human pose estimation. In ECCV, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[26] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, pages 2640-2649, 2017. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[27] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In ECCV, pages 483-499, 2016. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[28] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[29] Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with temporal convolutions and semi-supervised training. In CVPR, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1832" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.133 + ], + "angle": 0, + "content": "[30] William H Press and Saul A Teukolsky. Savitzky-golay smoothing filters. Computers in Physics, 4(6):669-672, 1990. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.189 + ], + "angle": 0, + "content": "[31] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In ICCV, pages 11488-11499, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.193, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[32] Garvita Tiwari, Dimitrije Antić, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In ECCV, pages 572–589, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.468, + 0.275 + ], + "angle": 0, + "content": "[33] Carnegie Mellon University. Cmu graphics lab motion capture database: http://mocap.cs.cmu.edu/. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.468, + 0.33 + ], + "angle": 0, + "content": "[34] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeuIPS, 30, 2017. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.335, + 0.468, + 0.389 + ], + "angle": 0, + "content": "[35] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, pages 601-617, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.392, + 0.468, + 0.445 + ], + "angle": 0, + "content": "[36] Ziniu Wan, Zhengjia Li, Maoqing Tian, Jianbo Liu, Shuai Yi, and Hongsheng Li. Encoder-decoder with multi-level attention for 3d human shape and pose estimation. In ICCV, pages 13033-13042, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.448, + 0.468, + 0.502 + ], + "angle": 0, + "content": "[37] Tao Wang, Lei Jin, Zhang Wang, Xiaojin Fan, Yu Cheng, Yinglei Teng, Junliang Xing, and Jian Zhao. Decenternet: Bottom-up human pose estimation via decentralized pose representation. In ACM MM, pages 1798-1808, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.505, + 0.468, + 0.559 + ], + "angle": 0, + "content": "[38] Wen-Li Wei, Jen-Chun Lin, Tyng-Luh Liu, and Hong-Yuan Mark Liao. Capturing humans in motion: Temporal-attentive 3d human pose and shape estimation from monocular video. In CVPR, pages 13211-13220, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.562, + 0.468, + 0.614 + ], + "angle": 0, + "content": "[39] J. Wu, H. Zheng, B. Zhao, Y. Li, B. Yan, R. Liang, W. Wang, S. Zhou, G. Lin, and Y. Fu. Ai challenger: A large-scale dataset for going deeper in image understanding. In ICME, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.618, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[40] Ian T Young and Lucas J Van Vliet. Recursive implementation of the gaussian filter. Signal processing, 44(2):139-151, 1995. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.661, + 0.468, + 0.714 + ], + "angle": 0, + "content": "[41] Ailing Zeng, Xuan Ju, Lei Yang, Ruiyuan Gao, Xizhou Zhu, Bo Dai, and Qiang Xu. Deciwatch: A simple baseline for \\(10 \\times\\) efficient 2d and 3d pose estimation. In ECCV, pages 607-624. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.718, + 0.468, + 0.771 + ], + "angle": 0, + "content": "[42] Ailing Zeng, Lei Yang, Xuan Ju, Jiefeng Li, Jianyi Wang, and Qiang Xu. Smoothnet: A plug-and-play network for refining human poses in videos. In ECCV, pages 625-642, 2022. 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.771 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "1833" + } + ] +] \ No newline at end of file diff --git a/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/e165eedf-6d86-4e7b-90c3-ccc6312e093e_origin.pdf b/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/e165eedf-6d86-4e7b-90c3-ccc6312e093e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..90fa40db7b282ae038777f480fcfc772e32db134 --- /dev/null +++ b/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/e165eedf-6d86-4e7b-90c3-ccc6312e093e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ac241dd8497e8899844e850b9088a33c257beb5e783de1115ccc7ee0c0c6bfc +size 4862502 diff --git a/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/full.md b/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c995a8a495cfff0f60795be5eb3d906d5b124e2b --- /dev/null +++ b/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/full.md @@ -0,0 +1,290 @@ +# SynSP: Synergy of Smoothness and Precision in Pose Sequences Refinement + +Tao Wang\*, Lei Jin\*, Zheng Wang?, Jianshu Li?, Liang Li?, Fang Zhao?, Yu Cheng?, Li Yuan?, Li Zhou?, Junliang Xing?, Jian Zhao? + +1Beijing University of Posts and Telecommunications, 2Wuhan University, 3Ant Group, + +$^{4}$ Institute of computing technology Chinese Academy of Sciences, $^{5}$ Nanjing University, $^{6}$ National University of Singapore, + +$^{7}$ Peking University, $^{8}$ Tsinghua University, $^{9}$ China Telecom Institute of AI, $^{10}$ Northwestern Polytechnical University + +wangtao@bupt.edu.cn, jinlei@bupt.edu.cn, wangzwhu@whu.edu.cn, jianshu@u.NUS.edu, liang.li@ict.ac.cn, zhaofang0627@gmail.com, e0321276@u.nus.edu, yuanli-ece@pku.edu.cn, zhouli2025@126.com, jlxing@tsinghua.edu.cn, jian.zhao@nwpu.edu.cn + +# Abstract + +Predicting human pose sequences via existing pose estimators often encounters various estimation errors. Motion refinement methods aim to optimize the predicted human pose sequences from pose estimators while ensuring minimal computational overhead and latency. Prior investigations have primarily concentrated on striking a balance between the two objectives, i.e., smoothness and precision, while optimizing the predicted pose sequences. However, it has come to our attention that the tension between these two objectives can provide additional quality cues about the predicted pose sequences. These cues, in turn, are able to aid the network in optimizing lower-quality poses. To leverage this quality information, we propose a motion refinement network, termed SynSP, to achieve a Synergy of Smoothness and Precision in the sequence refinement tasks. Moreover, SynSP can also address multi-view poses of one person simultaneously, fixing inaccuracies in predicted poses through heightened attention to similar poses from other views, thereby amplifying the resultant quality cues and overall performance. Compared with previous methods, SynSP benefits from both pose quality and multi-view information with a much shorter input sequence length, achieving state-of-the-art results among four challenging datasets involving 2D, 3D, and SMPL pose representations in both single-view and multi-view scenes. Github code: https://github.com/InvertedForest/SynSP. + +# 1. Introduction + +Human pose estimation [4, 16, 21, 37] has been widely used in automatic driving, human-computer interaction, motion + +![](images/baf8189bb9233acac58ade0ad8229c78e828127b997e6d9c937f35f11d1ca177.jpg) + +![](images/75e2fddf21e0ef126936e467bc2448d0f9de88ec45683547542e40d5670ad282.jpg) + +![](images/813f6dcadbd43f806ae20c98274f895aa5d0479ba48fbf752b1c879107ae1af6.jpg) +Figure 1. Illustration of our motivation. (a): The x-axis represents the ratio of weights for Precision and Smoothness loss in the training process of SynSP, while the y-axis represents the value of MPJPE (Precision) and Acceleration error (Smoothness). These two errors exhibit a conflicting relationship. (b): Diff1 denotes the difference between the predicted sequence (from pose estimator) and the ground-truth sequence, and Diff2 denotes the difference between two output sequences from training SynSP with a loss ratio of 0.5 and 5, respectively. We notice that the trends of Diff1 and Diff2 are roughly the same in (b). (c): A larger value of Diff1 indicates lower prediction quality for the predicted sequences. For example, the left ankle of the person in the middle picture of (c) is occluded, making it difficult for the pose estimator to infer. + +analysis, etc. Although current methods have achieved outstanding performance, both image-based and video-based pose estimators still sometimes exhibit noticeable errors in + +predicting pose sequences. The pose sequences generated by image-based pose estimators often show significant jitters due to a lack of inter-frame correlation, while video-based pose estimators alleviate jitters and improve accuracy by leveraging temporal information of input videos. However, training video-based pose estimators requires substantial amounts of video data and computational resources [7, 38, 39]. + +Consequently, researchers propose motion refinement methods [3, 42] as post-processing techniques to optimize the predicted sequences from pose estimators to approximate the ground-truth sequences from human annotation, with two primary objectives: precision and smoothness. Precision objective aims to improve the position accuracy of the predicted joints through the temporal and spatial information, evaluated by Mean Per Joint Position Error (MPJPE) [11, 29]. Smoothness objective does not strive to make the output sequences as smooth as possible but rather to align the acceleration of the output with that of the ground-truth sequences. Both objectives aim to make the output sequences close to the ground-truth sequences. However, focusing on one of them will lead to errors in the other greatly increased, as shown in Fig. 1. + +Due to the inherent tension between Precision and Smoothness objectives, previous methods attempt to achieve a satisfactory balance between them. As shown in Fig. 1(a), it can be observed that with an increase in the ratio of weights for Precision and Smoothness loss during SynSP training process, there is a noticeable rise in Acceleration error (Smoothness) while MPJPE (Precision) decreases, which is a common phenomenon in previous methods [42]. However, we note that the quality of the predicted sequences from estimator is related to the difference of outputs from SynSP biased towards the Precision and Smoothness objectives. Specifically, as illustrated in Fig. 1(b), the difference (Diff2) between the two output sequences, which focus on precision and smoothness respectively, is correlated with the quality of the predicted sequences, which is generated by the difference (Diff1) between predicted sequences from pose estimators and ground-truth sequences from human annotation. To further verify this correlation, we calculate the Pearson Correlation Coefficient between Diff1 and Diff2 on the Human3.6M and 3DPW datasets. Based on our experiment, the mean Pearson correlation coefficient [9] is 0.48 (bigger than 0.3 is considered as relevant), indicating that the dissimilarity (Diff2) between the two output sequences can serve as an indicator for the quality of predicted sequences without knowing the ground-truth. + +In this paper, we propose SynSP to utilize this quality information. Specifically, we employ a transformer network and encode the quality information by Pose Quality Encoding (PQE) module to guide the network's attention towards + +low-quality poses in sequences. Additionally, with the support of multi-view dataset: Human3.6M, we further find that multi-view information is effective to enhance the reliability of PQE. Therefore, we introduce the Pose Similarity Encoding (PSE) module to exploit the multi-view information by guide the network's attention towards similar poses from multi-view. + +The main contributions are summarized as follows: + +- We translate the inherent tension between smoothness and precision into the quality cue of predicted sequences with the Pose Quality Encoding module in SynSP. Then we utilize this cue as the weight to help regulate the network's attention on each frame in the sequences. +- We further observe that similar poses can effectively rectify erroneous poses. Consequently, we propose Pose Similarity Encoding to encode pose similarity information and refine poses of multiple views simultaneously. +- We conduct sufficient experiments on four challenging datasets among 2D, 3D, and SMPL representations with image-based and video-based pose estimators. Notably, compared with previous method SmoothNet, SynSP's MPJPE improves by $16.9\%$ on Human3.6M dataset with 3D pose estimator FCN, while reducing input sequence length to $25\%$ and minimizing time delay to just $17\%$ . + +# 2. Related Work + +Human Pose Estimation. Human pose estimation task [14-16, 27] is to extract location of human joints from the collected data. There are 2D, 3D, and SMPL representations for the human pose. 2D representation only contains 2D joint coordinates parallel to the camera plane. 3D representation has one more dimension of depth information. SMPL representation [23] is a skinned vertex based model that represents a wide variety of body shapes in natural human poses, while needs more parameters to represent. + +Human pose estimators can be divided into image-based methods [27] and video-based methods [17]. Pose sequences from image-based methods often show a lot of jitters when these methods are applied to videos, while video-based methods require more data and computing resources for training. 2D pose estimator, Hourglass [27] found an outstanding stacked hourglass networks for human pose estimation; SPIN [19] can learn to reconstruct 3D human pose and SMPL pose via model-fitting in the loop; video-based pose estimators, TCMR [6] and VIBE [17], can generate smooth pose sequences with temporal information. We conduct the experiments based on pose estimators mentioned above. + +Motion Refinement Methods. The predicted human pose sequences can be further refined by motion refinement methods, making the position and acceleration of each joint in these sequences are closer to those of the ground-truth sequences. The motion refinement methods can be divided to + +![](images/99289463ddb4e36c6fdff262e505eacb29452fb26849ed0b9a314baa5f55a50b.jpg) +Figure 2. The architecture of SynSP. FCN represents fully-connected network. In the base stage, with pose sequences of multiple views as input, SynSP first encodes these sequences individually in the Local Correlation Encoder, and merge multiple view's information in the Global Correlation Encoder with the Pose Similarity Encoding as their weight. After obtaining the individual embedding as the query, and incorporating the biased global information as the key and value, we get two outputs through the decoder module. In the refinement stage, we utilize these two detached outputs to generate Pose Quality Encoding and more precise Pose Similarity Encoding. By reusing embedding generated earlier, we achieve superior results using only one decoder with these encodings. + +low-pass filters and deep learning based methods. The traditional low-pass filters, e.g., Savitzky-Golay [30], Gaussian1d [40], and One-Euro [3], are generally adapted to smooth the predicted pose sequences, and perform well for those abrupt jitters, such as confusion in left and right hips. Although the processed sequences from low-pass filter does look smoother, it is difficult to ensure precision at the same time. Deep learning based method, i.e., SmoothNet [42], takes the two objectives into account and aims to improve both precision and smoothness. Our SynSP is also based on deep learning and can achieve better results by considering the pose sequence quality and multi-view correlation information. + +Additionally, there are also several human pose prior methods [2, 22, 28, 32], which try to learn a prior distribution of valid human poses. Although time-consuming for their sampling steps, these methods can be applied to the task of pose generation/completion/refinement. Hu-MoR [31] introduces a novel conditional VAE which enables expressive and general motion reconstruction and generation; GAN-based method [10] proposes a simple yet effective prior for SMPL model to bound it to realistic human poses; GFPose [8] employs a novel score-based generative framework to model plausible 3D human poses. + +# 3. Method + +# 3.1. Overview + +Task Description. Given predicted pose sequences generated by human pose estimators [16], motion refinement methods aim to fit these predicted sequences to the ground + +truth sequences, making these prediction more precise and smooth. Usually, the predicted pose sequence $\mathbf{P}$ has a shape of $\mathbb{R}^{V\times T\times (J\times D)}$ , where $V$ stands for the number of views for one individual; $T$ refers to the frame number of these predicted pose sequences; $J$ is the number of joints associated with datasets; $D$ denotes the spatial dimensions of these joints. We use $P_{(v,t,j)}$ to represent the $j_{th}$ joint's position at $t_{th}$ frame from $v_{th}$ view throughout the paper. The spatial dimension is included in the joint dimension, ignoring the spatial dimensions to simplify it. + +Overall Architecture. We proposed SynSP to use this quality information from the certain tension relationship between the two objectives, i.e., smoothness and precision. We also introduce multi-view information fusion as an optional enhancement for SynSP. As illustrated in Fig. 2, SynSP is mainly based on the transformer structure [34] and composed of two stages. Firstly, under the guidance of Pose Similarity Encoding I, Base Decoder can utilize both individual and global information to generate two branches, which are respectively biased towards smoothness and precision. Next, these two branches can generate Pose Quality Encoding, and a more precise Pose Similarity Encoding II, which further helps Refinement Decoder generate refined results. + +# 3.2. Base Stage + +In the base stage, SynSP takes predicted sequences from pose estimators as input and generates two sequences that focus on smoothness and precision, respectively. + +Pose Similarity Encoding. Similarity for poses from multiview is effective in enhancing the performance of SynSP. + +On one hand, similar poses from different views can complement each other and improve the robustness of joint position. On the other hand, this encoding minimizes the network's attention on poses with significant distortions to reduce their impact for the refinement task. + +Pose Similarity Encoding (PSE) is calculated as follows: + +$$ +\begin{array}{l} P _ {(v, t, j)} ^ {r e l} = \hat {P} _ {(v, t, j)} - P _ {(v, t, c)}, \\ P S E _ {\left(v _ {1}, v _ {2}, t\right)} = \exp \left(- \sum_ {j = 1} ^ {J} \left(P _ {\left(v _ {1}, t, j\right)} ^ {r e l} - P _ {\left(v _ {2}, t, j\right)} ^ {r e l}\right) ^ {2}\right), \tag {1} \\ \end{array} +$$ + +where $c$ represents the index of the "pelvis" keypoint in the joint list. $v_{1}$ and $v_{2}$ are used only for distinction, and both have the same domain as $v$ . Overall, PSE has a shape of $\mathbb{R}^{V\times V\times T}$ , and we present a visualization example of PSE for one frame in Fig. 3. + +Global Correlation Encoder. In multi-view scenes, the correlation information between multiple views can help SynSP to deduce the location of the joint with jitters globally. Global Correlation Encoder (GCE) module is introduced to produce the keys and values required by the Base Decoder with the global correlation information between multiple views. First, the predicted sequences $\mathbf{P}$ are projected to embedding $\mathbf{E}$ through a full connected layer along the joint dimension. + +Here, we use $E_{(v,t,e)}$ to denote an item of projected embedding $\mathbf{E}$ , where $e$ refers to the index of the embedding dimension. Then, the view (v) and the frame (t) dimensions of $\mathbf{E}$ are merged, so that pose information from multiple views and temporal information can be sensed by the encoder simultaneously. Afterwards, $E_{((v,t),e)}$ is sent to a encoder module of transformer [34] to interact pose sequences from multiple views and get the global information $E_{((v,t),e)}^g$ . Then this information is added with $PSE_{(v_1,v_2,t)}$ to adjust $v_{1th}$ view's attention to other views' information by broadcast operation: $E_{((v_2,t),e)}^g + PSE_{(v_1,v_2,t)} = E_{(1,(v_2,t),e)}^g + PSE_{(v_1,(v_2,t),1)} = E_{(v_1,(v_2,t),e)}^g$ , which is the final global information from GCE module. + +Local Correlation Encoder. Local Correlation Encoder (LCE) module focuses on the temporal dimension with single-view and produces the query required by the Base Decoder. LCE reuses the projected embedding $\mathbf{E}$ from the full connected layer mentioned in GCE as input, and directly sends it to another standard encoder module of transformer [34] to obtain local information in the frame (t) dimension (take the view (v) dimension as input batch). Finally, for $v_{1th}$ view, $t_{th}$ frame, $e_{th}$ item, LCE produces the local information $E_{(v_1,t,e)}^l$ with exploring the correlation of the time dimension. + +Base Decoder. At the end of base stage, we stack 5 standard decoding layers of transformer [34] into Base Decoder. The Base Decoder takes the local information $E_{(v_1,t,e)}^l$ as the query, to interact with global information $E_{(v_1,(v_2,t),e)}^g$ , + +![](images/7d1b9819a56f28f6b928590aa779535d386b062872f317c7fc54809d43c3886e.jpg) +Pose Similarity Encoding + +![](images/f42e415c8a0ab521219e0ca802c9e523e72a52082e0873c33c6b803d24a5fe9d.jpg) +Pose Quality Encoding +Figure 3. Visualization of the two encodings of one frame. $PSE$ : the diagonal has a constant value of 1. $PQE$ : the value outside the diagonal is constant 0. In this way, the addition of the two encodings cannot affect each other. + +which serves as keys and values for the Base Decoder. At the last layer of Base Decoder, we use two decoder layers to generate two branches biased towards smoothness and precision respectively at the same time. + +$$ +\begin{array}{l} K, V = E _ {(v _ {1}, (v _ {2}, t), e)} ^ {g}, Q ^ {0} = E _ {(v _ {1}, t, e)} ^ {l}, \\ Q _ {(v _ {1}, t, e)} ^ {d _ {t h}} = \mathbf {D L a y e r} ^ {d _ {t h}} \left(Q ^ {(d - 1) _ {t h}}, K, V\right), d \in \{1, 2, 3 \}, \tag {2} \\ P _ {(v _ {1}, t, j)} ^ {\text {p o s}} = \operatorname {F C N} \left(\mathbf {D L a y e r} ^ {4 _ {\mathrm {t h}}} \left(\mathrm {Q} ^ {3 _ {\mathrm {t h}}}, \mathrm {K}, \mathrm {V}\right)\right), \\ P _ {(v _ {1}, t, j)} ^ {a c c} = \operatorname {F C N} \left(\mathbf {D L a y e r} ^ {5 _ {\mathrm {t h}}} \left(\mathrm {Q} ^ {3 _ {\mathrm {t h}}}, \mathrm {K}, \mathrm {V}\right)\right), \\ \end{array} +$$ + +where $Q$ , $K$ , and $V$ denote the query, key, and value for Base Decoder respectively, $\mathbf{DLayer}^{d_{th}}$ denotes the $d_{th}$ decoder layer, $\mathrm{FCN}(\cdot)$ is fully connected network. In this way, SynSP can generate two pose squences $P^{pos}$ and $P^{acc}$ from its Acceleration Branch and Position Branch, which are biased towards the precision and smoothness objectives in the final loss function Equ. (4), respectively. Next, we can also further optimize these sequences in the refinement stage, with a small amount of calculation and parameters as the cost. + +# 3.3. Refinement Stage + +In the refinement stage, we aim to seek a better way to fuse Acceleration Branch and Position Branch, which concentrate on different objectives. + +Acceleration and Position Branches. The output sequences from Acceleration Branch and Position Branch are detached from computation graph for a stabler training process as the predictions from base stage are independent of the refinement stage. These outputs can not only generate more accurate Pose Similarity Encoding, but also provide the quality information about these pose sequences. + +Pose Quality Encoding. We have empirically observed that when outputs from the two branches show a greater discrepancy, the prediction quality is poorer. Specifically, we computed the Pearson Correlation Coefficient between the prediction quality and the difference between two outputs is about 0.48 (more details in the supplementary materials), + +which shows that the two have a strong correlation. Hence, we propose the Pose Quality Encoding (PQE) to utilize this correlation to improve the final prediction. Each item of PQE is calculated as follows: + +$$ +P _ {(v _ {1}, v _ {2}, t, j)} ^ {d i s} = | P _ {(v _ {1}, t, j)} ^ {a c c} - P _ {(v _ {2}, t, j)} ^ {p o s} |, +$$ + +$$ +P Q E _ {(v _ {1}, v _ {2}, t)} ^ {\prime} = \frac {1}{J} \sum_ {j = 1} ^ {J} \exp \left(- \frac {P _ {(v _ {1} , v _ {2} , t , j)} ^ {d i s}}{\max _ {j \in J} P _ {(v _ {1} , v _ {2} , t , j)} ^ {d i s} + \sigma}\right), \tag {3} +$$ + +$$ +P Q E _ {(v _ {1}, v _ {2}, t)} = \left\{ \begin{array}{c c} P Q E _ {(v _ {1}, v _ {2}, t)} ^ {\prime}, & v _ {1} = v _ {2}; \\ 0, & v _ {1} \neq v _ {2}. \end{array} \right. +$$ + +Here, we calculate PQE by the distance $P^{dis}$ between the outputs $P^{acc}$ and $P^{pos}$ , which are generated by Acceleration Branch and Position Branch, respectively. To ensure that extreme values do not interfere with PQE, $P^{dis}$ is normalized by dividing the maximum value in the individual's sequence. We introduce a small constant $\sigma$ (set to $1 \times 10^{-6}$ ) to avoid division by zero. + +Unlike PSE, which is calculated based on correlations among multiple views on one person, PQE is computed by $P^{dis}$ for a single view across multiple frames. Since PQE and PSE have distinct considerations regarding multiple views versus single view, they can be combined through addition rather than concatenation without mutual influence as illustrated in Fig. 3, thereby enhancing the computational efficiency of SynSP. + +Refinement Decoder. The output sequences of the two branches exhibit enhanced accuracy and smoothness compared to the input sequences. A straightforward approach would involve averaging these outputs and replicating the base stage pipeline once again. However, our experimental findings indicate that this approach does not yield significant benefits but doubles the computational cost. Please refer to supplementary material for details. This may be attributed to the fact that while the optimized output is indeed smooth and precise, it lacks crucial positional information from the original jitter input, rendering optimization of an already optimized output meaningless. + +To circumvent this issue and maintain a lightweight model, we reuse the embedding from the Global Correlation Encoder and Base Decoder. The former embedding retains the original position information, while the latter embedding contains optimization information about the output. In this way, we can produce superior outcomes by using just one additional decoder. We stack 4 standard decoding layers of transformer [34] in Refinement Decoder, and the inner calculation of Refinement Decoder is similar with Base Decoder. As shown in Tab. 5, this approach effectively improves the positional accuracy of pose sequences. + +Analysis on the Refinement Stage. In the refinement stage, the difference between the position and acceleration branches serves as a guiding factor for the quality of predicted sequence. Next, within the refinement decoder, + +![](images/09c11a710195fc6c8d00267fc9afcd46a1b0f2068a3a8e3408eec1065cd251c5.jpg) +Figure 4. Illustration of Sliding Window Average Algorithm (SWAA), which is widely used by motion refinement works. This figure proves that the amount of computation and latency is proportional to the window length. + +greater emphasis is placed on frames with lower pose quality to facilitate improvements in the final performance. Regarding why this difference is associated with pose quality, we posit that the two branches are easy to remain consistent under the normal circumstances, nevertheless, in challenging scenarios such as occlusion or sudden movement, it is more likely for the two branches to exhibit dissimilar behavior. Whether in pose estimator or motion refinement model, pose quality is is obviously related with scenarios. + +# 3.4. Training and Inference + +Different from previous work, we only use the window length $T = 8$ for training and inference. Compared with the window length of 32, this setting reduces the time delay of $(32 - 8) / 12 = 2s$ while the inference speed of pose estimator is 12fps, and have $8 / 32 = 0.25$ times the previous amount of computation as shown in Fig. 4. + +Training. The losses of the three outputs, i.e., Acceleration Branch output, Position Branch output, and final output, share the same calculation method. We have two losses, $\mathcal{L}_{acc}$ and $\mathcal{L}_{pos}$ , for each one of the three outputs: + +$$ +\begin{array}{l} A _ {(v, t, j)} = \left(P _ {(v, t + 2, j)} - P _ {(v, t + 1, j)}\right) - \left(P _ {(v, t + 1, j)} - P _ {(v, t, j)}\right), \\ \mathcal {L} _ {\mathrm {p o s}} = \frac {1}{V \cdot T \cdot J} \sum_ {v = 1} ^ {V} \sum_ {t = 1} ^ {T} \sum_ {j = 1} ^ {J} | P _ {(v, t, j)} - \hat {P} _ {(v, t, j)} |, \\ \mathcal {L} _ {\mathrm {a c c}} = \frac {1}{V \cdot T \cdot J} \sum_ {v = 1} ^ {V} \sum_ {t = 1} ^ {T} \sum_ {j = 1} ^ {J} | A _ {(v, t, j)} - \hat {A} _ {(v, t, j)} |, \\ \mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {p o s}} + \lambda \mathcal {L} _ {\text {a c c}}. \tag {4} \\ \end{array} +$$ + +In our experiments, $\lambda$ is set as 0, 1, and 0.5 for Position Branch, Acceleration Branch, and the final output, respectively, which is a trade-off between smoothness and precision. For all branches, $\mathcal{L}_{\mathrm{pos}}$ is crucial to ensuring training stability. Furthermore, to balance $\mathcal{L}_{\mathrm{pos}}$ and $\mathcal{L}_{\mathrm{acc}}$ , the parameter $\lambda$ should be set as 0.5, determined by the absolute value of each element. Thus, for the final output, we have set $\lambda$ to 0.5, whereas in the Acceleration branch, $\lambda$ is assigned as 1 to prioritize smoothness. + +Inference. During inference, we obtain outputs from the refinement stage. Then the outputs go through a + +Table 1. Comparison with related works on Human3.6M, AIST++ datasets with 2D, 3D, and SMPL human pose estimator Hourglass [27], FCN [26], and SPIN [19], respectively. WS denotes the windows size in training, inference, and post-process. * indicates that SynSP is trained and tested in multi-view scenes. + +
MethodWSHuman3.6M / 2DHuman3.6M / 3DAIST++ / SMPL
MPJPE↓PA-MPJPE↓Accel↓MPJPE↓PA-MPJPE↓Accel↓MPJPE↓PA-MPJPE↓Accel↓
InputN/A9.427.641.5454.5542.219.17107.7274.4033.20
One-Euro [3]110.697.980.3455.2042.733.80108.9775.2714.70
Gaussian1d [40]329.377.560.5153.6741.602.43104.8472.1810.05
Savitzky-Golay [30]329.357.550.1753.4841.191.34104.5872.306.07
SmoothNet [42]329.257.570.1552.7240.921.03103.0071.195.72
SynSP88.136.090.1551.3640.131.0284.6359.026.08
SynSP*87.625.640.1541.7833.320.98---
+ +post-processing called sliding window average algorithm adopted by most works as shown in Fig. 4 (we have accelerated it by parallel computing to $1\%$ of the original latency, refer to supplementary material for more details). It can be seen that the delay and computation needs for model and post-processing becomes smaller with shorter window length. + +Single-View Scenes. While $V = 1$ in the equations above, SynSP can be directly applied to adapt single-view scenes. In single-view scenes, PSE module is no longer needed. Since multi-view scenes are not considered by current related works and not our main motivation for the proposed SynSP, we conduct sufficient single-view experiments for a comprehensive comparison. + +# 4. Experiments + +# 4.1. Experiment Settings + +Dataset. To demonstrate the generalizability and efficiency of our method, we conduct experiments of SynSP in three human pose representations of 2D, 3D, and SMPL for single-view and multi-view scenes on Human3.6M [12], AIST++ [20], 3DPW [35], and CMU-Mocap [33] datasets. For the fairness of the comparison, we use these datasets made by SmoothNet [42] with kinds of pose estimators in the experiments. To compare with pose prior models, we also follow GFPose [8] to add the same degree of Gaussian noise or uniformly distributed noise to Human3.6M and CMU-Mocap datasets as input for the pose denoise task. + +Metrics. Following the related works [42], we use Accel error, MPJPE, and PA-MPJPE (Procrustes-aligned MPJPE) [1, 5] to evaluate the acceleration and position proximity between the predicted sequence and the ground-truth sequence, and the lower metrics the smaller error of the predicted sequence. + +Implementation Details. SynSP consists of two 4-layer encoder modules (GCE, LCE), a 5-layer decoder module (Base Decoder), and a 4-layer decoder module (Refinement Decoder). We initialize the learning rate as 0.001 and decay + +it by a factor of 0.95 per epoch. SynSP is trained for 70 epochs with Adam optimizer, requiring approximately two hours and 7GB of GPU memory with a batch size of 4096. The computational setup consists of an Intel(R) Xeon(R) Gold 6271C CPU @ 2.60GHz for the CPU and a Tesla V100 SXM2 32GB for the GPU. + +# 4.2. Comparison with State-Of-The-Arts. + +In our experiments, we first demonstrate that SynSP consistently outperforms previous methods in both single-view and multi-view scenes with several pose estimators. Then we compare with pose prior models on the pose denoise task by smoothing the pose sequences with artificial noisy. (There are also some methods[13, 41] with a super-long window length, and the comparison with them is shown in the supplementary materials.) + +Table 2. Comparison of various video-based pose estimators. + represents the combination of estimators and motion refinement methods. * indicates multi-view inputs. + +
DatasetMethodWSMPJPE↓Accel↓
3DPWMAED [36]1679.00-
MPS-Net [38]1684.30-
TCMR [6]1686.466.75
TCMR+SmoothNet [42]16+3286.506.00
TCMR+SynSP16+886.105.90
PARE [18]179.0025.60
PARE+SmoothNet [42]1+3278.105.91
PARE+SynSP1+876.206.16
AIST++VIBE [17]16106.931.60
VIBE+SmoothNet [42]16+3297.474.15
VIBE+SynSP16+877.004.32
H36MTransFusion* [24]125.52-
PPT* [25]125.1611.60
PPT*+SmoothNet [42]1+3223.971.31
PPT*+SynSP1+821.511.10
+ +Table 3. Comparison on Human3.6M dataset for the denoise task. $N(0,\sigma)$ represents a Gaussian distribution with a mean of 0 and a variance of $\sigma$ , and $U(k)$ represents a uniform distribution at $[0,k]$ . $\mathbf{M}\downarrow$ and $\mathrm{A}\downarrow$ denote MPJPE↓ and Accel↓ respectively. + +
MethodN(0,100)N(0,400)U(50)U(100)
M↓A↓M↓A↓M↓A↓M↓A↓
Noisy input65.6160.3131.1320.594.8231.3189.5462.7
GFpose [8]42.8-64.6-50.9-89.4-
SmoothNet [42]42.410.757.914.249.410.274.316.3
SynSP37.68.856.113.045.610.969.415.7
+ +Table 4. Comparison on CMU-Mocap dataset for the denoise task. $N(0,\sigma)$ represents a Gaussian distribution with a mean of 0 and a variance of $\sigma$ , and $U(k)$ represents a uniform distribution at [0,k]. $\mathbf{M} \downarrow$ and $\mathbf{A} \downarrow$ denote MPJPE $\downarrow$ and Accel $\downarrow$ respectively. + +
MethodN(0,100)N(0,400)U(50)U(100)
M↓A↓M↓A↓M↓A↓M↓A↓
Noisy input61.2143.8122.5287.688.7207.9177.4415.8
GFpose [8]40.4-60.0-45.6-84.4-
SmoothNet [42]35.14.341.54.937.74.244.95.1
SynSP20.63.728.04.424.44.132.44.6
+ +2D Representation. In Tab. 1, we compare SynSP with other motion refinement methods in Human3.6M dataset with 2D pose estimator Hourglass [27]. With the setting of 8 window length, our SynSP outperforms all the methods while it is about 2 seconds faster (except the One-Euro) according to Fig. 4. Compared with SmoothNet, which is trained on multiple datasets, i.e., AIST-VIBE-3D, 3DPW-SPIN-3D, and Human3.6M-FCN-3D, SynSP can still outperforms it by $16.9\%$ for MPJPE and $25.2\%$ for PA-MPJPE with smaller window size. According to the ablation experiment results in Sec. 4.3, the significant accuracy improvement of SynSP is attributed to the mutual patching between the pose quality information (PQE) and multi-view information (PSE). + +3D Representation. In Tab. 1, we also compare SynSP with other motion refinement methods in Human3.6M dataset with 3D pose estimator FCN [26]. Similar with the comparison in 2D Representation, we can still outperforms SmoothNet by $20.8\%$ for MPJPE and $18.6\%$ for PA-MPJPE in Human3.6M dataset. There is also a decrease of $4.9\%$ for the Accel error. Furthermore, SynSP obtains these results only with 8 window size, while SmoothNet and most filters need 32 frames. + +SMPL Representation. Then we compare SynSP with other motion refinement methods in AIST++ dataset with SMPL pose estimator SPIN as shown in Tab. 1. SynSP shows superior performance in SMPL representation. + +Comparison with Video-based Pose Estimators. The video-based estimators already include temporal informa + +tion. Compared with the image-based estimators, motion refinement methods have difficult to optimize the sequences from video-based estimators. In Tab. 2, we compare SynSP with SOTA methods and SmoothNet in AIST++ and 3DPW datasets with 3D pose video-based estimator VIBE [17] and TCMR [6], respectively. We can find that the high performance video-based estimators, TCMR, is hard to be refined. SmoothNet improves the smoothness by processing a longer window length than the video-based estimators. SynSP greatly increases precision and smoothness in a shorter window length with its outstanding architecture as shown in Fig. 2. The image-based estimator, PARE [18], has excellent performance in precision, SynSP can also make up for its shortcomings in smoothness and outperform some video-based estimators. + +Comparison with Multi-view Pose Estimators. We further verify the performance of SynSP with the multi-view pose estimator, i.e., PPT [25]. As shown in Tab. 2, we can notice that SynSP can further improve the precision of PPT, which has superiority performance by considering multiview connections. We also present the result of other multiview methods such as TransFusion for comparison. This experiment results further proves the multi-view refinement ability of PSE for video-based estimators. + +Comparison with Pose Priors Methods. To ensure a fair comparison with human pose prior models, we also conducted denoise experiments following the methodology of the SOTA work [8] among pose priors methods. Here, window size of GFPose, SmoothNet, and SynSP are 1, 32, and 8 respectively, while GFPose, as a generative model for human pose prior task, has 1000 sampling steps, i.e., 1000 times of model inference. GFPose has not presented its Accel errors since it does not take temporal information into account. As presented in Tab. 3 and Tab. 4, SynSP can effectively solve MPJPE (Precision) problems. Across noisy Human3.6M and CMU-Mocap datasets, SynSP can improve the pose quality by about $54.5\%$ in terms of MPJPE and $96.0\%$ in terms of Accel errors for Gaussian noise with a variance of 100. For Gaussian noise with a large variance of 400, SynSP can improve the pose quality by about $67.2\%$ in terms of MPJPE and $97.2\%$ int terms of Accel errors. We also observe consistent improvement in uniform noise as shown in Tab. 4. + +# 4.3. Ablation Study. + +To ensure representativeness of ablation experiments, all experiments were performed on the largest dataset, Human3.6M, with 3D estimator FCN. + +Modules of SynSP. In order to verify the rationality of the structure of SynSP, we conducted ablation experiments on each module in SynSP. In Tab. 5, the line without a tick represents the results of baseline, which is the network of base stage without PSE I. And this network surpasses Smooth- + +Table 5. Ablation study on PSE I, RS (Refinement Stage), PSE II, and PQE modules. * denotes directly using the output from base stage as input of refinement stage to re-optimize these sequence. + +
PSE IRSPSE IIPQEMPJPEAccel
51.21.09
✓48.71.07
✓*51.41.08
✓49.61.07
✓✓46.71.05
✓✓✓46.41.04
✓✓✓42.50.99
✓✓✓✓41.80.98
+ +Net in precision by $2.8\%$ . PSE I, RS, PSE II, and PQE all have gains in results, and the combination of PSE I, RS, and PQE has brought greater benefits. We believe that this benefit mainly comes from the combination of SynSP with global information PSE I, which can help generate more accurate quality information according to the poses from other views. Besides, utilizing the difference between acceleration and Position Branch improves both smoothness and precision of final prediction. It is worth noting that using the average value of the two branches as the input RS does not bring reasonable gains. We suspect that the output of base stage loses the original predicted position information, which affects the accuracy and makes the refinement stage useless compared with the baseline. + +Time Efficiency Analysis. We conducted latency experiments on the Human3.6M dataset using FCN [26] as the 3D pose estimator with a batch size of 1416, as shown in Tab. 6. Considering the practical application environment, we report not only the inference latency of the model, as shown in the 3rd column of Tab. 6, but also the latency associated with the Sliding Window Average Algorithm (SWAA), which includes the delay in waiting for pose estimators and the computation cost of SWAA (we have optimized SWAA specifically for parallel computation on GPUs). The bold cell in each row of Tab. 6 represents the primary factor contributing to latency. For video processing speed of 12 frames per second for the pose estimator (based on the general inference speed of various pose estimators), a window size of 8 would require a waiting time of $8/12 = 667\mathrm{ms}$ ; similarly, a window size of 32 would require a waiting time of $32/12 = 2667\mathrm{ms}$ . Furthermore, the model inference time for SynSP is only about 10ms. Overall, the main time cost primarily arises from the internal waiting latency for SmoothNet and SynSP. For the pose priors method GFPose, the main time delay is its 1000 sampling steps. Hence, SynSP is more suitable for near real-time pose refinement task. + +Analysis on Number of Views. As shown in Tab. 7, as more poses from different views are input at the same time + +Table 6. Time efficiency analysis on SynSP including model inference time and time cost for sliding window. Besides, we assume that FPS (Frame Per Second) for pose estimator is 12. + +
MethodsWSModel (ms)SWAATotal (ms)
Wait (ms)Execution (ms)
GFPose13772--3772
SmoothNet321.2126678023470
SynSP811.06670.3678
SynSP*812.36670.3680
+ +Table 7. Contrast experiments on the number of input views. $\dagger$ means that we randomly combine several people together for input during the training process, instead of using multiple views for one person. * means that in the case of $\dagger$ , several people with similar poses are input at the same time during the inference process. + +
View NumberMPJPEPA-MPJPEAccel
151.440.11.02
247.237.61.02
346.537.11.00
441.833.30.98
4†50.439.51.01
4*49.739.11.01
+ +during the training process and inference process, the performance of SynSP is more efficient. It is justifiable due to the introduction of more global information. If number of views is larger than 4, we believe that the performance of our model can be further promoted with only a marginal increase in model complexity. + +# 5. Conclusion + +This paper proposes SynSP, a motion refinement method that utilizes the tension relationship between smoothness and precision through synergistic optimization. Specifically, SynSP outputs Position Branch and Acceleration Branch biased towards precision and smoothness, respectively. Then we employ the discrepancy between the two branches as a quality cue for the predicted sequence, which can be utilized in subsequent stages to further refine the pose sequence. In addition, we introduce Pose Similarity Encoding, achieving simultaneous refinement of multi-view poses and performance enhancement. + +# Acknowledgment + +The paper is supported by National Key R & D Program of China (No. 2021YFB3300100), Natural Science Foundation of China No.62102039, the Fundamental Research Funds for the Central Universities No.500422813, and Natural Science Foundation of China No.6232211. + +# References + +[1] A. Benzine, F. Chabot, B. Luvison, Q. C. Pham, and C. Achard. Pandanet: Anchor-based single-shot multi-person 3d pose estimation. In CVPR, 2020. 6 +[2] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In ECCV, pages 561-578, 2016. 3 +[3] Géry Casiez, Nicolas Roussel, and Daniel Vogel. 1€ filter: a simple speed-based low-pass filter for noisy input in interactive systems. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pages 2527-2530, 2012. 2, 3, 6 +[4] Yu Cheng, Bo Yang, Bo Wang, and Robby T Tan. 3d human pose estimation using spatio-temporal networks with explicit occlusion training. In AAAI, pages 10631-10638, 2020. 1 +[5] Yu Cheng, Bo Wang, Bo Yang, and Robby T. Tan. Monocular 3d multi-person pose estimation by integrating top-down and bottom-up networks. In CVPR, pages 7649-7659, 2021. 6 +[6] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consistent 3d human pose and shape from a video. In CVPR, pages 1964-1973, 2021. 2, 6, 7 +[7] Jiaming Chu, Lei Jin, Xiaojin Fan, Yinglei Teng, Yunchao Wei, Yuqiang Fang, Junliang Xing, and Jian Zhao. Single-stage multi-human parsing via point sets and center-based offsets. In ACM MM, pages 1863-1873, 2023. 2 +[8] Hai Ci, Mingdong Wu, Wentao Zhu, Xiaoxuan Ma, Hao Dong, Fangwei Zhong, and Yizhou Wang. Gfpose: Learning 3d human pose prior with gradient fields. CVPR, 2023. 3, 6, 7 +[9] Israel Cohen, Yiteng Huang, Jingdong Chen, Jacob Benesty, Jacob Benesty, Jingdong Chen, Yiteng Huang, and Israel Cohen. Pearson correlation coefficient. Noise reduction in speech processing, pages 1-4, 2009. 2 +[10] Andrey Davydov, Anastasia Remizova, Victor Constantin, Sina Honari, Mathieu Salzmann, and Pascal Fua. Adversarial parametric pose prior. In CVPR, pages 10997-11005, 2022. 3 +[11] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Stefano Alletto, and Rita Cucchiara. Compressed volumetric heatmaps for multi-person 3d pose estimation. In CVPR, 2020. 2 +[12] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE TPAMI, 36(7):1325-1339, 2013. 6 +[13] Kyung-Min Jin, Byoung-Sung Lim, Gun-Hee Lee, Tae-Kyung Kang, and Seong-Whan Lee. Kinematic-aware hierarchical attention network for human pose estimation in videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5725-5734, 2023. 6 +[14] Lei Jin, Xiaojuan Wang, Xuecheng Nie, Luoqi Liu, Yandong Guo, and Jian Zhao. Grouping by center: Predicting centripetal offsets for the bottom-up human pose estimation. IEEE TMM, 2022. 2 + +[15] Lei Jin, Chenyang Xu, Xiaojuan Wang, Yabo Xiao, Yandong Guo, Xuecheng Nie, and Jian Zhao. Single-stage is enough: Multi-person absolute 3d pose estimation. In CVPR, pages 13086-13095, 2022. +[16] Lei Jin, Xiaojuan Wang, Xuecheng Nie, Wendong Wang, Yandong Guo, Shuicheng Yan, and Jian Zhao. Rethinking the person localization for single-stage multi-person pose estimation. IEEE TMM, 2023. 1, 2, 3 +[17] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In CVPR, pages 5253-5263, 2020. 2, 6, 7 +[18] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. In ICCV, pages 11127-11137, 2021. 6, 7 +[19] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 2, 6 +[20] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In ICCV, pages 13401-13412, 2021. 6 +[21] Yang Li, Kan Li, Shuai Jiang, Ziyue Zhang, Congzhen-tao Huang, and Richard Yi Da Xu. Geometry-driven self-supervised method for 3d human pose estimation. In AAAI, pages 11442-11449, 2020. 1 +[22] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Transactions on Graphics (TOG), 39(4):40-1, 2020. 3 +[23] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM TOG, 34(6):1-16, 2015. 2 +[24] Haoyu Ma, Liangjian Chen, Deying Kong, Zhe Wang, Xingwei Liu, Hao Tang, Xiangyi Yan, Yusheng Xie, Shih-Yao Lin, and Xiaohui Xie. Transfusion: Cross-view fusion with transformer for 3d human pose estimation. arXiv preprint arXiv:2110.09554, 2021. 6 +[25] Haoyu Ma, Zhe Wang, Yifei Chen, Deying Kong, Liangjian Chen, Xingwei Liu, Xiangyi Yan, Hao Tang, and Xiaohui Xie. Ppt: Token-pruned pose transformer for monocular and multi-view human pose estimation. In ECCV, 2022. 6, 7 +[26] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, pages 2640-2649, 2017. 6, 7, 8 +[27] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In ECCV, pages 483-499, 2016. 2, 6, 7 +[28] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 3 +[29] Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with temporal convolutions and semi-supervised training. In CVPR, 2019. 2 + +[30] William H Press and Saul A Teukolsky. Savitzky-golay smoothing filters. Computers in Physics, 4(6):669-672, 1990. 3, 6 +[31] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In ICCV, pages 11488-11499, 2021. 3 +[32] Garvita Tiwari, Dimitrije Antić, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In ECCV, pages 572–589, 2022. 3 +[33] Carnegie Mellon University. Cmu graphics lab motion capture database: http://mocap.cs.cmu.edu/. 6 +[34] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeuIPS, 30, 2017. 3, 4, 5 +[35] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, pages 601-617, 2018. 6 +[36] Ziniu Wan, Zhengjia Li, Maoqing Tian, Jianbo Liu, Shuai Yi, and Hongsheng Li. Encoder-decoder with multi-level attention for 3d human shape and pose estimation. In ICCV, pages 13033-13042, 2021. 6 +[37] Tao Wang, Lei Jin, Zhang Wang, Xiaojin Fan, Yu Cheng, Yinglei Teng, Junliang Xing, and Jian Zhao. Decenternet: Bottom-up human pose estimation via decentralized pose representation. In ACM MM, pages 1798-1808, 2023. 1 +[38] Wen-Li Wei, Jen-Chun Lin, Tyng-Luh Liu, and Hong-Yuan Mark Liao. Capturing humans in motion: Temporal-attentive 3d human pose and shape estimation from monocular video. In CVPR, pages 13211-13220, 2022. 2, 6 +[39] J. Wu, H. Zheng, B. Zhao, Y. Li, B. Yan, R. Liang, W. Wang, S. Zhou, G. Lin, and Y. Fu. Ai challenger: A large-scale dataset for going deeper in image understanding. In ICME, 2017. 2 +[40] Ian T Young and Lucas J Van Vliet. Recursive implementation of the gaussian filter. Signal processing, 44(2):139-151, 1995. 3, 6 +[41] Ailing Zeng, Xuan Ju, Lei Yang, Ruiyuan Gao, Xizhou Zhu, Bo Dai, and Qiang Xu. Deciwatch: A simple baseline for $10 \times$ efficient 2d and 3d pose estimation. In ECCV, pages 607-624. Springer, 2022. 6 +[42] Ailing Zeng, Lei Yang, Xuan Ju, Jiefeng Li, Jianyi Wang, and Qiang Xu. Smoothnet: A plug-and-play network for refining human poses in videos. In ECCV, pages 625-642, 2022. 2, 3, 6, 7 \ No newline at end of file diff --git a/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/images.zip b/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8b4623b3f77b33757ae34b2435abd2456dcf1c60 --- /dev/null +++ b/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed9431c2e903d9097198f98a284791a1df75fb864211fde777d1a75765a86659 +size 448128 diff --git a/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/layout.json b/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ebb1fef641efa598a06a4716170b66187e1d0222 --- /dev/null +++ b/2024/SynSP_ Synergy of Smoothness and Precision in Pose Sequences Refinement/layout.json @@ -0,0 +1,8181 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 63, + 103, + 531, + 121 + ], + "type": "text", + "content": "SynSP: Synergy of Smoothness and Precision in Pose Sequences Refinement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 142, + 477, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 142, + 477, + 171 + ], + "spans": [ + { + "bbox": [ + 114, + 142, + 477, + 171 + ], + "type": "text", + "content": "Tao Wang\\*, Lei Jin\\*, Zheng Wang?, Jianshu Li?, Liang Li?, Fang Zhao?, Yu Cheng?, Li Yuan?, Li Zhou?, Junliang Xing?, Jian Zhao?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 173, + 456, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 173, + 456, + 185 + ], + "spans": [ + { + "bbox": [ + 136, + 173, + 456, + 185 + ], + "type": "text", + "content": "1Beijing University of Posts and Telecommunications, 2Wuhan University, 3Ant Group," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 73, + 186, + 520, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 186, + 520, + 199 + ], + "spans": [ + { + "bbox": [ + 73, + 186, + 520, + 199 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 73, + 186, + 520, + 199 + ], + "type": "text", + "content": "Institute of computing technology Chinese Academy of Sciences, " + }, + { + "bbox": [ + 73, + 186, + 520, + 199 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 73, + 186, + 520, + 199 + ], + "type": "text", + "content": "Nanjing University, " + }, + { + "bbox": [ + 73, + 186, + 520, + 199 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 73, + 186, + 520, + 199 + ], + "type": "text", + "content": "National University of Singapore," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 200, + 507, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 200, + 507, + 213 + ], + "spans": [ + { + "bbox": [ + 86, + 200, + 507, + 213 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 86, + 200, + 507, + 213 + ], + "type": "text", + "content": "Peking University, " + }, + { + "bbox": [ + 86, + 200, + 507, + 213 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 86, + 200, + 507, + 213 + ], + "type": "text", + "content": "Tsinghua University, " + }, + { + "bbox": [ + 86, + 200, + 507, + 213 + ], + "type": "inline_equation", + "content": "^{9}" + }, + { + "bbox": [ + 86, + 200, + 507, + 213 + ], + "type": "text", + "content": "China Telecom Institute of AI, " + }, + { + "bbox": [ + 86, + 200, + 507, + 213 + ], + "type": "inline_equation", + "content": "^{10}" + }, + { + "bbox": [ + 86, + 200, + 507, + 213 + ], + "type": "text", + "content": "Northwestern Polytechnical University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 215, + 527, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 215, + 527, + 255 + ], + "spans": [ + { + "bbox": [ + 63, + 215, + 527, + 255 + ], + "type": "text", + "content": "wangtao@bupt.edu.cn, jinlei@bupt.edu.cn, wangzwhu@whu.edu.cn, jianshu@u.NUS.edu, liang.li@ict.ac.cn, zhaofang0627@gmail.com, e0321276@u.nus.edu, yuanli-ece@pku.edu.cn, zhouli2025@126.com, jlxing@tsinghua.edu.cn, jian.zhao@nwpu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "spans": [ + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 308, + 290, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 290, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 290, + 619 + ], + "type": "text", + "content": "Predicting human pose sequences via existing pose estimators often encounters various estimation errors. Motion refinement methods aim to optimize the predicted human pose sequences from pose estimators while ensuring minimal computational overhead and latency. Prior investigations have primarily concentrated on striking a balance between the two objectives, i.e., smoothness and precision, while optimizing the predicted pose sequences. However, it has come to our attention that the tension between these two objectives can provide additional quality cues about the predicted pose sequences. These cues, in turn, are able to aid the network in optimizing lower-quality poses. To leverage this quality information, we propose a motion refinement network, termed SynSP, to achieve a Synergy of Smoothness and Precision in the sequence refinement tasks. Moreover, SynSP can also address multi-view poses of one person simultaneously, fixing inaccuracies in predicted poses through heightened attention to similar poses from other views, thereby amplifying the resultant quality cues and overall performance. Compared with previous methods, SynSP benefits from both pose quality and multi-view information with a much shorter input sequence length, achieving state-of-the-art results among four challenging datasets involving 2D, 3D, and SMPL pose representations in both single-view and multi-view scenes. Github code: https://github.com/InvertedForest/SynSP." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 641, + 128, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 128, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 128, + 654 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 662, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 662, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 662, + 287, + 686 + ], + "type": "text", + "content": "Human pose estimation [4, 16, 21, 37] has been widely used in automatic driving, human-computer interaction, motion" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 282, + 545, + 361 + ], + "blocks": [ + { + "bbox": [ + 307, + 282, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 307, + 282, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 545, + 361 + ], + "type": "image", + "image_path": "baf8189bb9233acac58ade0ad8229c78e828127b997e6d9c937f35f11d1ca177.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 308, + 361, + 541, + 441 + ], + "blocks": [ + { + "bbox": [ + 308, + 361, + 541, + 441 + ], + "lines": [ + { + "bbox": [ + 308, + 361, + 541, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 541, + 441 + ], + "type": "image", + "image_path": "75e2fddf21e0ef126936e467bc2448d0f9de88ec45683547542e40d5670ad282.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 441, + 533, + 510 + ], + "blocks": [ + { + "bbox": [ + 309, + 441, + 533, + 510 + ], + "lines": [ + { + "bbox": [ + 309, + 441, + 533, + 510 + ], + "spans": [ + { + "bbox": [ + 309, + 441, + 533, + 510 + ], + "type": "image", + "image_path": "813f6dcadbd43f806ae20c98274f895aa5d0479ba48fbf752b1c879107ae1af6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 514, + 547, + 658 + ], + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 658 + ], + "type": "text", + "content": "Figure 1. Illustration of our motivation. (a): The x-axis represents the ratio of weights for Precision and Smoothness loss in the training process of SynSP, while the y-axis represents the value of MPJPE (Precision) and Acceleration error (Smoothness). These two errors exhibit a conflicting relationship. (b): Diff1 denotes the difference between the predicted sequence (from pose estimator) and the ground-truth sequence, and Diff2 denotes the difference between two output sequences from training SynSP with a loss ratio of 0.5 and 5, respectively. We notice that the trends of Diff1 and Diff2 are roughly the same in (b). (c): A larger value of Diff1 indicates lower prediction quality for the predicted sequences. For example, the left ankle of the person in the middle picture of (c) is occluded, making it difficult for the pose estimator to infer." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "analysis, etc. Although current methods have achieved outstanding performance, both image-based and video-based pose estimators still sometimes exhibit noticeable errors in" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "text", + "content": "*Both authors contributed equally to this research. † Lei Jin is corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1824" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "content": "predicting pose sequences. The pose sequences generated by image-based pose estimators often show significant jitters due to a lack of inter-frame correlation, while video-based pose estimators alleviate jitters and improve accuracy by leveraging temporal information of input videos. However, training video-based pose estimators requires substantial amounts of video data and computational resources [7, 38, 39]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 171, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 350 + ], + "type": "text", + "content": "Consequently, researchers propose motion refinement methods [3, 42] as post-processing techniques to optimize the predicted sequences from pose estimators to approximate the ground-truth sequences from human annotation, with two primary objectives: precision and smoothness. Precision objective aims to improve the position accuracy of the predicted joints through the temporal and spatial information, evaluated by Mean Per Joint Position Error (MPJPE) [11, 29]. Smoothness objective does not strive to make the output sequences as smooth as possible but rather to align the acceleration of the output with that of the ground-truth sequences. Both objectives aim to make the output sequences close to the ground-truth sequences. However, focusing on one of them will lead to errors in the other greatly increased, as shown in Fig. 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 352, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 288, + 662 + ], + "type": "text", + "content": "Due to the inherent tension between Precision and Smoothness objectives, previous methods attempt to achieve a satisfactory balance between them. As shown in Fig. 1(a), it can be observed that with an increase in the ratio of weights for Precision and Smoothness loss during SynSP training process, there is a noticeable rise in Acceleration error (Smoothness) while MPJPE (Precision) decreases, which is a common phenomenon in previous methods [42]. However, we note that the quality of the predicted sequences from estimator is related to the difference of outputs from SynSP biased towards the Precision and Smoothness objectives. Specifically, as illustrated in Fig. 1(b), the difference (Diff2) between the two output sequences, which focus on precision and smoothness respectively, is correlated with the quality of the predicted sequences, which is generated by the difference (Diff1) between predicted sequences from pose estimators and ground-truth sequences from human annotation. To further verify this correlation, we calculate the Pearson Correlation Coefficient between Diff1 and Diff2 on the Human3.6M and 3DPW datasets. Based on our experiment, the mean Pearson correlation coefficient [9] is 0.48 (bigger than 0.3 is considered as relevant), indicating that the dissimilarity (Diff2) between the two output sequences can serve as an indicator for the quality of predicted sequences without knowing the ground-truth." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": "In this paper, we propose SynSP to utilize this quality information. Specifically, we employ a transformer network and encode the quality information by Pose Quality Encoding (PQE) module to guide the network's attention towards" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 155 + ], + "type": "text", + "content": "low-quality poses in sequences. Additionally, with the support of multi-view dataset: Human3.6M, we further find that multi-view information is effective to enhance the reliability of PQE. Therefore, we introduce the Pose Similarity Encoding (PSE) module to exploit the multi-view information by guide the network's attention towards similar poses from multi-view." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 156, + 525, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 156, + 525, + 167 + ], + "spans": [ + { + "bbox": [ + 317, + 156, + 525, + 167 + ], + "type": "text", + "content": "The main contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 170, + 545, + 360 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 307, + 170, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 170, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 170, + 545, + 228 + ], + "type": "text", + "content": "- We translate the inherent tension between smoothness and precision into the quality cue of predicted sequences with the Pose Quality Encoding module in SynSP. Then we utilize this cue as the weight to help regulate the network's attention on each frame in the sequences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 229, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 545, + 276 + ], + "type": "text", + "content": "- We further observe that similar poses can effectively rectify erroneous poses. Consequently, we propose Pose Similarity Encoding to encode pose similarity information and refine poses of multiple views simultaneously." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 277, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 360 + ], + "type": "text", + "content": "- We conduct sufficient experiments on four challenging datasets among 2D, 3D, and SMPL representations with image-based and video-based pose estimators. Notably, compared with previous method SmoothNet, SynSP's MPJPE improves by " + }, + { + "bbox": [ + 307, + 277, + 545, + 360 + ], + "type": "inline_equation", + "content": "16.9\\%" + }, + { + "bbox": [ + 307, + 277, + 545, + 360 + ], + "type": "text", + "content": " on Human3.6M dataset with 3D pose estimator FCN, while reducing input sequence length to " + }, + { + "bbox": [ + 307, + 277, + 545, + 360 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 307, + 277, + 545, + 360 + ], + "type": "text", + "content": " and minimizing time delay to just " + }, + { + "bbox": [ + 307, + 277, + 545, + 360 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 307, + 277, + 545, + 360 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 370, + 392, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 392, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 392, + 383 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 390, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 498 + ], + "type": "text", + "content": "Human Pose Estimation. Human pose estimation task [14-16, 27] is to extract location of human joints from the collected data. There are 2D, 3D, and SMPL representations for the human pose. 2D representation only contains 2D joint coordinates parallel to the camera plane. 3D representation has one more dimension of depth information. SMPL representation [23] is a skinned vertex based model that represents a wide variety of body shapes in natural human poses, while needs more parameters to represent." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 652 + ], + "type": "text", + "content": "Human pose estimators can be divided into image-based methods [27] and video-based methods [17]. Pose sequences from image-based methods often show a lot of jitters when these methods are applied to videos, while video-based methods require more data and computing resources for training. 2D pose estimator, Hourglass [27] found an outstanding stacked hourglass networks for human pose estimation; SPIN [19] can learn to reconstruct 3D human pose and SMPL pose via model-fitting in the loop; video-based pose estimators, TCMR [6] and VIBE [17], can generate smooth pose sequences with temporal information. We conduct the experiments based on pose estimators mentioned above." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Motion Refinement Methods. The predicted human pose sequences can be further refined by motion refinement methods, making the position and acceleration of each joint in these sequences are closer to those of the ground-truth sequences. The motion refinement methods can be divided to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1825" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 534, + 238 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 534, + 238 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 534, + 238 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 534, + 238 + ], + "type": "image", + "image_path": "99289463ddb4e36c6fdff262e505eacb29452fb26849ed0b9a314baa5f55a50b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 548, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 548, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 548, + 319 + ], + "type": "text", + "content": "Figure 2. The architecture of SynSP. FCN represents fully-connected network. In the base stage, with pose sequences of multiple views as input, SynSP first encodes these sequences individually in the Local Correlation Encoder, and merge multiple view's information in the Global Correlation Encoder with the Pose Similarity Encoding as their weight. After obtaining the individual embedding as the query, and incorporating the biased global information as the key and value, we get two outputs through the decoder module. In the refinement stage, we utilize these two detached outputs to generate Pose Quality Encoding and more precise Pose Similarity Encoding. By reusing embedding generated earlier, we achieve superior results using only one decoder with these encodings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 338, + 289, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 289, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 289, + 492 + ], + "type": "text", + "content": "low-pass filters and deep learning based methods. The traditional low-pass filters, e.g., Savitzky-Golay [30], Gaussian1d [40], and One-Euro [3], are generally adapted to smooth the predicted pose sequences, and perform well for those abrupt jitters, such as confusion in left and right hips. Although the processed sequences from low-pass filter does look smoother, it is difficult to ensure precision at the same time. Deep learning based method, i.e., SmoothNet [42], takes the two objectives into account and aims to improve both precision and smoothness. Our SynSP is also based on deep learning and can achieve better results by considering the pose sequence quality and multi-view correlation information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 494, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 287, + 625 + ], + "type": "text", + "content": "Additionally, there are also several human pose prior methods [2, 22, 28, 32], which try to learn a prior distribution of valid human poses. Although time-consuming for their sampling steps, these methods can be applied to the task of pose generation/completion/refinement. Hu-MoR [31] introduces a novel conditional VAE which enables expressive and general motion reconstruction and generation; GAN-based method [10] proposes a simple yet effective prior for SMPL model to bound it to realistic human poses; GFPose [8] employs a novel score-based generative framework to model plausible 3D human poses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 638, + 102, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 102, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 102, + 651 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 114, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 114, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 114, + 670 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Task Description. Given predicted pose sequences generated by human pose estimators [16], motion refinement methods aim to fit these predicted sequences to the ground" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "text", + "content": "truth sequences, making these prediction more precise and smooth. Usually, the predicted pose sequence " + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "text", + "content": " has a shape of " + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{V\\times T\\times (J\\times D)}" + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "text", + "content": " stands for the number of views for one individual; " + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "text", + "content": " refers to the frame number of these predicted pose sequences; " + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "text", + "content": " is the number of joints associated with datasets; " + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "text", + "content": " denotes the spatial dimensions of these joints. We use " + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "inline_equation", + "content": "P_{(v,t,j)}" + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "text", + "content": " to represent the " + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "inline_equation", + "content": "j_{th}" + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "text", + "content": " joint's position at " + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "inline_equation", + "content": "t_{th}" + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "text", + "content": " frame from " + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "inline_equation", + "content": "v_{th}" + }, + { + "bbox": [ + 304, + 338, + 545, + 456 + ], + "type": "text", + "content": " view throughout the paper. The spatial dimension is included in the joint dimension, ignoring the spatial dimensions to simplify it." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 457, + 546, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 546, + 624 + ], + "type": "text", + "content": "Overall Architecture. We proposed SynSP to use this quality information from the certain tension relationship between the two objectives, i.e., smoothness and precision. We also introduce multi-view information fusion as an optional enhancement for SynSP. As illustrated in Fig. 2, SynSP is mainly based on the transformer structure [34] and composed of two stages. Firstly, under the guidance of Pose Similarity Encoding I, Base Decoder can utilize both individual and global information to generate two branches, which are respectively biased towards smoothness and precision. Next, these two branches can generate Pose Quality Encoding, and a more precise Pose Similarity Encoding II, which further helps Refinement Decoder generate refined results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 634, + 379, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 379, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 379, + 647 + ], + "type": "text", + "content": "3.2. Base Stage" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": "In the base stage, SynSP takes predicted sequences from pose estimators as input and generates two sequences that focus on smoothness and precision, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Pose Similarity Encoding. Similarity for poses from multiview is effective in enhancing the performance of SynSP." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1826" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "On one hand, similar poses from different views can complement each other and improve the robustness of joint position. On the other hand, this encoding minimizes the network's attention on poses with significant distortions to reduce their impact for the refinement task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 133, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 133, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 59, + 133, + 286, + 144 + ], + "type": "text", + "content": "Pose Similarity Encoding (PSE) is calculated as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 155, + 287, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 155, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 69, + 155, + 287, + 201 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P _ {(v, t, j)} ^ {r e l} = \\hat {P} _ {(v, t, j)} - P _ {(v, t, c)}, \\\\ P S E _ {\\left(v _ {1}, v _ {2}, t\\right)} = \\exp \\left(- \\sum_ {j = 1} ^ {J} \\left(P _ {\\left(v _ {1}, t, j\\right)} ^ {r e l} - P _ {\\left(v _ {2}, t, j\\right)} ^ {r e l}\\right) ^ {2}\\right), \\tag {1} \\\\ \\end{array}", + "image_path": "6fe313851de9c2d05f13068ecc6d9370a39a6aa3f779f898e51cd2c5f210b758.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 205, + 287, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 265 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 205, + 287, + 265 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 205, + 287, + 265 + ], + "type": "text", + "content": " represents the index of the \"pelvis\" keypoint in the joint list. " + }, + { + "bbox": [ + 46, + 205, + 287, + 265 + ], + "type": "inline_equation", + "content": "v_{1}" + }, + { + "bbox": [ + 46, + 205, + 287, + 265 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 205, + 287, + 265 + ], + "type": "inline_equation", + "content": "v_{2}" + }, + { + "bbox": [ + 46, + 205, + 287, + 265 + ], + "type": "text", + "content": " are used only for distinction, and both have the same domain as " + }, + { + "bbox": [ + 46, + 205, + 287, + 265 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 205, + 287, + 265 + ], + "type": "text", + "content": ". Overall, PSE has a shape of " + }, + { + "bbox": [ + 46, + 205, + 287, + 265 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{V\\times V\\times T}" + }, + { + "bbox": [ + 46, + 205, + 287, + 265 + ], + "type": "text", + "content": ", and we present a visualization example of PSE for one frame in Fig. 3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "text", + "content": "Global Correlation Encoder. In multi-view scenes, the correlation information between multiple views can help SynSP to deduce the location of the joint with jitters globally. Global Correlation Encoder (GCE) module is introduced to produce the keys and values required by the Base Decoder with the global correlation information between multiple views. First, the predicted sequences " + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "text", + "content": " are projected to embedding " + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 46, + 266, + 287, + 373 + ], + "type": "text", + "content": " through a full connected layer along the joint dimension." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "text", + "content": "Here, we use " + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "inline_equation", + "content": "E_{(v,t,e)}" + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "text", + "content": " to denote an item of projected embedding " + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "text", + "content": " refers to the index of the embedding dimension. Then, the view (v) and the frame (t) dimensions of " + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "text", + "content": " are merged, so that pose information from multiple views and temporal information can be sensed by the encoder simultaneously. Afterwards, " + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "inline_equation", + "content": "E_{((v,t),e)}" + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "text", + "content": " is sent to a encoder module of transformer [34] to interact pose sequences from multiple views and get the global information " + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "inline_equation", + "content": "E_{((v,t),e)}^g" + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "text", + "content": ". Then this information is added with " + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "inline_equation", + "content": "PSE_{(v_1,v_2,t)}" + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "text", + "content": " to adjust " + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "inline_equation", + "content": "v_{1th}" + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "text", + "content": " view's attention to other views' information by broadcast operation: " + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "inline_equation", + "content": "E_{((v_2,t),e)}^g + PSE_{(v_1,v_2,t)} = E_{(1,(v_2,t),e)}^g + PSE_{(v_1,(v_2,t),1)} = E_{(v_1,(v_2,t),e)}^g" + }, + { + "bbox": [ + 46, + 375, + 287, + 529 + ], + "type": "text", + "content": ", which is the final global information from GCE module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 531, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 661 + ], + "type": "text", + "content": "Local Correlation Encoder. Local Correlation Encoder (LCE) module focuses on the temporal dimension with single-view and produces the query required by the Base Decoder. LCE reuses the projected embedding " + }, + { + "bbox": [ + 46, + 531, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 46, + 531, + 287, + 661 + ], + "type": "text", + "content": " from the full connected layer mentioned in GCE as input, and directly sends it to another standard encoder module of transformer [34] to obtain local information in the frame (t) dimension (take the view (v) dimension as input batch). Finally, for " + }, + { + "bbox": [ + 46, + 531, + 287, + 661 + ], + "type": "inline_equation", + "content": "v_{1th}" + }, + { + "bbox": [ + 46, + 531, + 287, + 661 + ], + "type": "text", + "content": " view, " + }, + { + "bbox": [ + 46, + 531, + 287, + 661 + ], + "type": "inline_equation", + "content": "t_{th}" + }, + { + "bbox": [ + 46, + 531, + 287, + 661 + ], + "type": "text", + "content": " frame, " + }, + { + "bbox": [ + 46, + 531, + 287, + 661 + ], + "type": "inline_equation", + "content": "e_{th}" + }, + { + "bbox": [ + 46, + 531, + 287, + 661 + ], + "type": "text", + "content": " item, LCE produces the local information " + }, + { + "bbox": [ + 46, + 531, + 287, + 661 + ], + "type": "inline_equation", + "content": "E_{(v_1,t,e)}^l" + }, + { + "bbox": [ + 46, + 531, + 287, + 661 + ], + "type": "text", + "content": " with exploring the correlation of the time dimension." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 663, + 287, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 663, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 46, + 663, + 287, + 716 + ], + "type": "text", + "content": "Base Decoder. At the end of base stage, we stack 5 standard decoding layers of transformer [34] into Base Decoder. The Base Decoder takes the local information " + }, + { + "bbox": [ + 46, + 663, + 287, + 716 + ], + "type": "inline_equation", + "content": "E_{(v_1,t,e)}^l" + }, + { + "bbox": [ + 46, + 663, + 287, + 716 + ], + "type": "text", + "content": " as the query, to interact with global information " + }, + { + "bbox": [ + 46, + 663, + 287, + 716 + ], + "type": "inline_equation", + "content": "E_{(v_1,(v_2,t),e)}^g" + }, + { + "bbox": [ + 46, + 663, + 287, + 716 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 320, + 72, + 411, + 165 + ], + "blocks": [ + { + "bbox": [ + 320, + 72, + 411, + 165 + ], + "lines": [ + { + "bbox": [ + 320, + 72, + 411, + 165 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 411, + 165 + ], + "type": "image", + "image_path": "7d1b9819a56f28f6b928590aa779535d386b062872f317c7fc54809d43c3886e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 166, + 406, + 175 + ], + "lines": [ + { + "bbox": [ + 329, + 166, + 406, + 175 + ], + "spans": [ + { + "bbox": [ + 329, + 166, + 406, + 175 + ], + "type": "text", + "content": "Pose Similarity Encoding" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 423, + 72, + 533, + 165 + ], + "blocks": [ + { + "bbox": [ + 423, + 72, + 533, + 165 + ], + "lines": [ + { + "bbox": [ + 423, + 72, + 533, + 165 + ], + "spans": [ + { + "bbox": [ + 423, + 72, + 533, + 165 + ], + "type": "image", + "image_path": "f42e415c8a0ab521219e0ca802c9e523e72a52082e0873c33c6b803d24a5fe9d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 166, + 525, + 175 + ], + "lines": [ + { + "bbox": [ + 455, + 166, + 525, + 175 + ], + "spans": [ + { + "bbox": [ + 455, + 166, + 525, + 175 + ], + "type": "text", + "content": "Pose Quality Encoding" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 182, + 545, + 226 + ], + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 226 + ], + "type": "text", + "content": "Figure 3. Visualization of the two encodings of one frame. " + }, + { + "bbox": [ + 304, + 182, + 545, + 226 + ], + "type": "inline_equation", + "content": "PSE" + }, + { + "bbox": [ + 304, + 182, + 545, + 226 + ], + "type": "text", + "content": ": the diagonal has a constant value of 1. " + }, + { + "bbox": [ + 304, + 182, + 545, + 226 + ], + "type": "inline_equation", + "content": "PQE" + }, + { + "bbox": [ + 304, + 182, + 545, + 226 + ], + "type": "text", + "content": ": the value outside the diagonal is constant 0. In this way, the addition of the two encodings cannot affect each other." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 248, + 545, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 545, + 296 + ], + "type": "text", + "content": "which serves as keys and values for the Base Decoder. At the last layer of Base Decoder, we use two decoder layers to generate two branches biased towards smoothness and precision respectively at the same time." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 304, + 545, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 304, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 318, + 304, + 545, + 368 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} K, V = E _ {(v _ {1}, (v _ {2}, t), e)} ^ {g}, Q ^ {0} = E _ {(v _ {1}, t, e)} ^ {l}, \\\\ Q _ {(v _ {1}, t, e)} ^ {d _ {t h}} = \\mathbf {D L a y e r} ^ {d _ {t h}} \\left(Q ^ {(d - 1) _ {t h}}, K, V\\right), d \\in \\{1, 2, 3 \\}, \\tag {2} \\\\ P _ {(v _ {1}, t, j)} ^ {\\text {p o s}} = \\operatorname {F C N} \\left(\\mathbf {D L a y e r} ^ {4 _ {\\mathrm {t h}}} \\left(\\mathrm {Q} ^ {3 _ {\\mathrm {t h}}}, \\mathrm {K}, \\mathrm {V}\\right)\\right), \\\\ P _ {(v _ {1}, t, j)} ^ {a c c} = \\operatorname {F C N} \\left(\\mathbf {D L a y e r} ^ {5 _ {\\mathrm {t h}}} \\left(\\mathrm {Q} ^ {3 _ {\\mathrm {t h}}}, \\mathrm {K}, \\mathrm {V}\\right)\\right), \\\\ \\end{array}", + "image_path": "782d98eaa973b4b7f0890c8a2fce262b11cf2193174a18b38a6a610cd55a532d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "text", + "content": " denote the query, key, and value for Base Decoder respectively, " + }, + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{DLayer}^{d_{th}}" + }, + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "inline_equation", + "content": "d_{th}" + }, + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "text", + "content": " decoder layer, " + }, + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\mathrm{FCN}(\\cdot)" + }, + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "text", + "content": " is fully connected network. In this way, SynSP can generate two pose squences " + }, + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "inline_equation", + "content": "P^{pos}" + }, + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "inline_equation", + "content": "P^{acc}" + }, + { + "bbox": [ + 304, + 375, + 545, + 493 + ], + "type": "text", + "content": " from its Acceleration Branch and Position Branch, which are biased towards the precision and smoothness objectives in the final loss function Equ. (4), respectively. Next, we can also further optimize these sequences in the refinement stage, with a small amount of calculation and parameters as the cost." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 502, + 410, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 410, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 410, + 515 + ], + "type": "text", + "content": "3.3. Refinement Stage" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 521, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 556 + ], + "type": "text", + "content": "In the refinement stage, we aim to seek a better way to fuse Acceleration Branch and Position Branch, which concentrate on different objectives." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 641 + ], + "type": "text", + "content": "Acceleration and Position Branches. The output sequences from Acceleration Branch and Position Branch are detached from computation graph for a stabler training process as the predictions from base stage are independent of the refinement stage. These outputs can not only generate more accurate Pose Similarity Encoding, but also provide the quality information about these pose sequences." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Pose Quality Encoding. We have empirically observed that when outputs from the two branches show a greater discrepancy, the prediction quality is poorer. Specifically, we computed the Pearson Correlation Coefficient between the prediction quality and the difference between two outputs is about 0.48 (more details in the supplementary materials)," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1827" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "type": "text", + "content": "which shows that the two have a strong correlation. Hence, we propose the Pose Quality Encoding (PQE) to utilize this correlation to improve the final prediction. Each item of PQE is calculated as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 126, + 201, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 126, + 201, + 144 + ], + "spans": [ + { + "bbox": [ + 61, + 126, + 201, + 144 + ], + "type": "interline_equation", + "content": "P _ {(v _ {1}, v _ {2}, t, j)} ^ {d i s} = | P _ {(v _ {1}, t, j)} ^ {a c c} - P _ {(v _ {2}, t, j)} ^ {p o s} |,", + "image_path": "4d81cc3f6836f4b46af614b3e9aca8fac880383205a9a4d1e3d2cc97400854d5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 144, + 287, + 176 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 144, + 287, + 176 + ], + "spans": [ + { + "bbox": [ + 62, + 144, + 287, + 176 + ], + "type": "interline_equation", + "content": "P Q E _ {(v _ {1}, v _ {2}, t)} ^ {\\prime} = \\frac {1}{J} \\sum_ {j = 1} ^ {J} \\exp \\left(- \\frac {P _ {(v _ {1} , v _ {2} , t , j)} ^ {d i s}}{\\max _ {j \\in J} P _ {(v _ {1} , v _ {2} , t , j)} ^ {d i s} + \\sigma}\\right), \\tag {3}", + "image_path": "d5fdd79caa5d71327a6876263ddf912032a662d3430904f95a4a4368a39851b2.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 178, + 231, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 178, + 231, + 202 + ], + "spans": [ + { + "bbox": [ + 62, + 178, + 231, + 202 + ], + "type": "interline_equation", + "content": "P Q E _ {(v _ {1}, v _ {2}, t)} = \\left\\{ \\begin{array}{c c} P Q E _ {(v _ {1}, v _ {2}, t)} ^ {\\prime}, & v _ {1} = v _ {2}; \\\\ 0, & v _ {1} \\neq v _ {2}. \\end{array} \\right.", + "image_path": "341ff20969319f1680cdab5eb8a752eb0e01d3d672303c53b675ff3444472d79.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 211, + 287, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 211, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 47, + 211, + 287, + 294 + ], + "type": "text", + "content": "Here, we calculate PQE by the distance " + }, + { + "bbox": [ + 47, + 211, + 287, + 294 + ], + "type": "inline_equation", + "content": "P^{dis}" + }, + { + "bbox": [ + 47, + 211, + 287, + 294 + ], + "type": "text", + "content": " between the outputs " + }, + { + "bbox": [ + 47, + 211, + 287, + 294 + ], + "type": "inline_equation", + "content": "P^{acc}" + }, + { + "bbox": [ + 47, + 211, + 287, + 294 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 211, + 287, + 294 + ], + "type": "inline_equation", + "content": "P^{pos}" + }, + { + "bbox": [ + 47, + 211, + 287, + 294 + ], + "type": "text", + "content": ", which are generated by Acceleration Branch and Position Branch, respectively. To ensure that extreme values do not interfere with PQE, " + }, + { + "bbox": [ + 47, + 211, + 287, + 294 + ], + "type": "inline_equation", + "content": "P^{dis}" + }, + { + "bbox": [ + 47, + 211, + 287, + 294 + ], + "type": "text", + "content": " is normalized by dividing the maximum value in the individual's sequence. We introduce a small constant " + }, + { + "bbox": [ + 47, + 211, + 287, + 294 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 211, + 287, + 294 + ], + "type": "text", + "content": " (set to " + }, + { + "bbox": [ + 47, + 211, + 287, + 294 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-6}" + }, + { + "bbox": [ + 47, + 211, + 287, + 294 + ], + "type": "text", + "content": ") to avoid division by zero." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 294, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 390 + ], + "type": "text", + "content": "Unlike PSE, which is calculated based on correlations among multiple views on one person, PQE is computed by " + }, + { + "bbox": [ + 46, + 294, + 287, + 390 + ], + "type": "inline_equation", + "content": "P^{dis}" + }, + { + "bbox": [ + 46, + 294, + 287, + 390 + ], + "type": "text", + "content": " for a single view across multiple frames. Since PQE and PSE have distinct considerations regarding multiple views versus single view, they can be combined through addition rather than concatenation without mutual influence as illustrated in Fig. 3, thereby enhancing the computational efficiency of SynSP." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 390, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 533 + ], + "type": "text", + "content": "Refinement Decoder. The output sequences of the two branches exhibit enhanced accuracy and smoothness compared to the input sequences. A straightforward approach would involve averaging these outputs and replicating the base stage pipeline once again. However, our experimental findings indicate that this approach does not yield significant benefits but doubles the computational cost. Please refer to supplementary material for details. This may be attributed to the fact that while the optimized output is indeed smooth and precise, it lacks crucial positional information from the original jitter input, rendering optimization of an already optimized output meaningless." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 534, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 665 + ], + "type": "text", + "content": "To circumvent this issue and maintain a lightweight model, we reuse the embedding from the Global Correlation Encoder and Base Decoder. The former embedding retains the original position information, while the latter embedding contains optimization information about the output. In this way, we can produce superior outcomes by using just one additional decoder. We stack 4 standard decoding layers of transformer [34] in Refinement Decoder, and the inner calculation of Refinement Decoder is similar with Base Decoder. As shown in Tab. 5, this approach effectively improves the positional accuracy of pose sequences." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Analysis on the Refinement Stage. In the refinement stage, the difference between the position and acceleration branches serves as a guiding factor for the quality of predicted sequence. Next, within the refinement decoder," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 520, + 148 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 520, + 148 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 520, + 148 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 520, + 148 + ], + "type": "image", + "image_path": "09c11a710195fc6c8d00267fc9afcd46a1b0f2068a3a8e3408eec1065cd251c5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 148, + 545, + 193 + ], + "lines": [ + { + "bbox": [ + 305, + 148, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 305, + 148, + 545, + 193 + ], + "type": "text", + "content": "Figure 4. Illustration of Sliding Window Average Algorithm (SWAA), which is widely used by motion refinement works. This figure proves that the amount of computation and latency is proportional to the window length." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 214, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 214, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 214, + 545, + 323 + ], + "type": "text", + "content": "greater emphasis is placed on frames with lower pose quality to facilitate improvements in the final performance. Regarding why this difference is associated with pose quality, we posit that the two branches are easy to remain consistent under the normal circumstances, nevertheless, in challenging scenarios such as occlusion or sudden movement, it is more likely for the two branches to exhibit dissimilar behavior. Whether in pose estimator or motion refinement model, pose quality is is obviously related with scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 331, + 437, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 331, + 437, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 437, + 342 + ], + "type": "text", + "content": "3.4. Training and Inference" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 349, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 349, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 305, + 349, + 545, + 421 + ], + "type": "text", + "content": "Different from previous work, we only use the window length " + }, + { + "bbox": [ + 305, + 349, + 545, + 421 + ], + "type": "inline_equation", + "content": "T = 8" + }, + { + "bbox": [ + 305, + 349, + 545, + 421 + ], + "type": "text", + "content": " for training and inference. Compared with the window length of 32, this setting reduces the time delay of " + }, + { + "bbox": [ + 305, + 349, + 545, + 421 + ], + "type": "inline_equation", + "content": "(32 - 8) / 12 = 2s" + }, + { + "bbox": [ + 305, + 349, + 545, + 421 + ], + "type": "text", + "content": " while the inference speed of pose estimator is 12fps, and have " + }, + { + "bbox": [ + 305, + 349, + 545, + 421 + ], + "type": "inline_equation", + "content": "8 / 32 = 0.25" + }, + { + "bbox": [ + 305, + 349, + 545, + 421 + ], + "type": "text", + "content": " times the previous amount of computation as shown in Fig. 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 422, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 422, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 545, + 469 + ], + "type": "text", + "content": "Training. The losses of the three outputs, i.e., Acceleration Branch output, Position Branch output, and final output, share the same calculation method. We have two losses, " + }, + { + "bbox": [ + 305, + 422, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{acc}" + }, + { + "bbox": [ + 305, + 422, + 545, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 422, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pos}" + }, + { + "bbox": [ + 305, + 422, + 545, + 469 + ], + "type": "text", + "content": ", for each one of the three outputs:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 478, + 545, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 478, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 310, + 478, + 545, + 581 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} A _ {(v, t, j)} = \\left(P _ {(v, t + 2, j)} - P _ {(v, t + 1, j)}\\right) - \\left(P _ {(v, t + 1, j)} - P _ {(v, t, j)}\\right), \\\\ \\mathcal {L} _ {\\mathrm {p o s}} = \\frac {1}{V \\cdot T \\cdot J} \\sum_ {v = 1} ^ {V} \\sum_ {t = 1} ^ {T} \\sum_ {j = 1} ^ {J} | P _ {(v, t, j)} - \\hat {P} _ {(v, t, j)} |, \\\\ \\mathcal {L} _ {\\mathrm {a c c}} = \\frac {1}{V \\cdot T \\cdot J} \\sum_ {v = 1} ^ {V} \\sum_ {t = 1} ^ {T} \\sum_ {j = 1} ^ {J} | A _ {(v, t, j)} - \\hat {A} _ {(v, t, j)} |, \\\\ \\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p o s}} + \\lambda \\mathcal {L} _ {\\text {a c c}}. \\tag {4} \\\\ \\end{array}", + "image_path": "d639747a7232a6076574ab11316f4ff6ccd45a2a332e7ec55806cc4a18bd49f2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": "In our experiments, " + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": " is set as 0, 1, and 0.5 for Position Branch, Acceleration Branch, and the final output, respectively, which is a trade-off between smoothness and precision. For all branches, " + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": " is crucial to ensuring training stability. Furthermore, to balance " + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{acc}}" + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": ", the parameter " + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": " should be set as 0.5, determined by the absolute value of each element. Thus, for the final output, we have set " + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": " to 0.5, whereas in the Acceleration branch, " + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": " is assigned as 1 to prioritize smoothness." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Inference. During inference, we obtain outputs from the refinement stage. Then the outputs go through a" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1828" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 101, + 533, + 224 + ], + "blocks": [ + { + "bbox": [ + 46, + 59, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 46, + 59, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 46, + 59, + 547, + 94 + ], + "type": "text", + "content": "Table 1. Comparison with related works on Human3.6M, AIST++ datasets with 2D, 3D, and SMPL human pose estimator Hourglass [27], FCN [26], and SPIN [19], respectively. WS denotes the windows size in training, inference, and post-process. * indicates that SynSP is trained and tested in multi-view scenes." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 101, + 533, + 224 + ], + "lines": [ + { + "bbox": [ + 63, + 101, + 533, + 224 + ], + "spans": [ + { + "bbox": [ + 63, + 101, + 533, + 224 + ], + "type": "table", + "html": "
MethodWSHuman3.6M / 2DHuman3.6M / 3DAIST++ / SMPL
MPJPE↓PA-MPJPE↓Accel↓MPJPE↓PA-MPJPE↓Accel↓MPJPE↓PA-MPJPE↓Accel↓
InputN/A9.427.641.5454.5542.219.17107.7274.4033.20
One-Euro [3]110.697.980.3455.2042.733.80108.9775.2714.70
Gaussian1d [40]329.377.560.5153.6741.602.43104.8472.1810.05
Savitzky-Golay [30]329.357.550.1753.4841.191.34104.5872.306.07
SmoothNet [42]329.257.570.1552.7240.921.03103.0071.195.72
SynSP88.136.090.1551.3640.131.0284.6359.026.08
SynSP*87.625.640.1541.7833.320.98---
", + "image_path": "20168a4302eb6b363ecc8f149bc35f57114088967a4b428796b40ed6a98663d0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 243, + 288, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 288, + 327 + ], + "type": "text", + "content": "post-processing called sliding window average algorithm adopted by most works as shown in Fig. 4 (we have accelerated it by parallel computing to " + }, + { + "bbox": [ + 46, + 243, + 288, + 327 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 243, + 288, + 327 + ], + "type": "text", + "content": " of the original latency, refer to supplementary material for more details). It can be seen that the delay and computation needs for model and post-processing becomes smaller with shorter window length." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "text", + "content": "Single-View Scenes. While " + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "inline_equation", + "content": "V = 1" + }, + { + "bbox": [ + 46, + 327, + 287, + 411 + ], + "type": "text", + "content": " in the equations above, SynSP can be directly applied to adapt single-view scenes. In single-view scenes, PSE module is no longer needed. Since multi-view scenes are not considered by current related works and not our main motivation for the proposed SynSP, we conduct sufficient single-view experiments for a comprehensive comparison." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 423, + 128, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 128, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 128, + 436 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 443, + 165, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 165, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 165, + 456 + ], + "type": "text", + "content": "4.1. Experiment Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 461, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 593 + ], + "type": "text", + "content": "Dataset. To demonstrate the generalizability and efficiency of our method, we conduct experiments of SynSP in three human pose representations of 2D, 3D, and SMPL for single-view and multi-view scenes on Human3.6M [12], AIST++ [20], 3DPW [35], and CMU-Mocap [33] datasets. For the fairness of the comparison, we use these datasets made by SmoothNet [42] with kinds of pose estimators in the experiments. To compare with pose prior models, we also follow GFPose [8] to add the same degree of Gaussian noise or uniformly distributed noise to Human3.6M and CMU-Mocap datasets as input for the pose denoise task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 288, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 665 + ], + "type": "text", + "content": "Metrics. Following the related works [42], we use Accel error, MPJPE, and PA-MPJPE (Procrustes-aligned MPJPE) [1, 5] to evaluate the acceleration and position proximity between the predicted sequence and the ground-truth sequence, and the lower metrics the smaller error of the predicted sequence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Implementation Details. SynSP consists of two 4-layer encoder modules (GCE, LCE), a 5-layer decoder module (Base Decoder), and a 4-layer decoder module (Refinement Decoder). We initialize the learning rate as 0.001 and decay" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 243, + 547, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 547, + 316 + ], + "type": "text", + "content": "it by a factor of 0.95 per epoch. SynSP is trained for 70 epochs with Adam optimizer, requiring approximately two hours and 7GB of GPU memory with a batch size of 4096. The computational setup consists of an Intel(R) Xeon(R) Gold 6271C CPU @ 2.60GHz for the CPU and a Tesla V100 SXM2 32GB for the GPU." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 327, + 501, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 501, + 339 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 501, + 339 + ], + "type": "text", + "content": "4.2. Comparison with State-Of-The-Arts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 346, + 547, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 547, + 441 + ], + "type": "text", + "content": "In our experiments, we first demonstrate that SynSP consistently outperforms previous methods in both single-view and multi-view scenes with several pose estimators. Then we compare with pose prior models on the pose denoise task by smoothing the pose sequences with artificial noisy. (There are also some methods[13, 41] with a super-long window length, and the comparison with them is shown in the supplementary materials.)" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 497, + 547, + 712 + ], + "blocks": [ + { + "bbox": [ + 304, + 454, + 547, + 488 + ], + "lines": [ + { + "bbox": [ + 304, + 454, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 547, + 488 + ], + "type": "text", + "content": "Table 2. Comparison of various video-based pose estimators. + represents the combination of estimators and motion refinement methods. * indicates multi-view inputs." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 497, + 547, + 712 + ], + "lines": [ + { + "bbox": [ + 307, + 497, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 497, + 547, + 712 + ], + "type": "table", + "html": "
DatasetMethodWSMPJPE↓Accel↓
3DPWMAED [36]1679.00-
MPS-Net [38]1684.30-
TCMR [6]1686.466.75
TCMR+SmoothNet [42]16+3286.506.00
TCMR+SynSP16+886.105.90
PARE [18]179.0025.60
PARE+SmoothNet [42]1+3278.105.91
PARE+SynSP1+876.206.16
AIST++VIBE [17]16106.931.60
VIBE+SmoothNet [42]16+3297.474.15
VIBE+SynSP16+877.004.32
H36MTransFusion* [24]125.52-
PPT* [25]125.1611.60
PPT*+SmoothNet [42]1+3223.971.31
PPT*+SynSP1+821.511.10
", + "image_path": "90ca88a5d09d01e8b49eaf9f37fa88490c3ac12d5b72a896aab14d67fef85d5f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1829" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 123, + 288, + 206 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": "Table 3. Comparison on Human3.6M dataset for the denoise task. " + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "inline_equation", + "content": "N(0,\\sigma)" + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": " represents a Gaussian distribution with a mean of 0 and a variance of " + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "inline_equation", + "content": "U(k)" + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": " represents a uniform distribution at " + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "inline_equation", + "content": "[0,k]" + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "inline_equation", + "content": "\\mathbf{M}\\downarrow" + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "inline_equation", + "content": "\\mathrm{A}\\downarrow" + }, + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": " denote MPJPE↓ and Accel↓ respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 123, + 288, + 206 + ], + "lines": [ + { + "bbox": [ + 48, + 123, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 123, + 288, + 206 + ], + "type": "table", + "html": "
MethodN(0,100)N(0,400)U(50)U(100)
M↓A↓M↓A↓M↓A↓M↓A↓
Noisy input65.6160.3131.1320.594.8231.3189.5462.7
GFpose [8]42.8-64.6-50.9-89.4-
SmoothNet [42]42.410.757.914.249.410.274.316.3
SynSP37.68.856.113.045.610.969.415.7
", + "image_path": "9b1bd3402fbc2b4d68f4787cb174ce5469c1b1b3c913779335d82f28ca2286d6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 270, + 288, + 353 + ], + "blocks": [ + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "type": "text", + "content": "Table 4. Comparison on CMU-Mocap dataset for the denoise task. " + }, + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "type": "inline_equation", + "content": "N(0,\\sigma)" + }, + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "type": "text", + "content": " represents a Gaussian distribution with a mean of 0 and a variance of " + }, + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "type": "inline_equation", + "content": "U(k)" + }, + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "type": "text", + "content": " represents a uniform distribution at [0,k]. " + }, + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\downarrow" + }, + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\downarrow" + }, + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "type": "text", + "content": " denote MPJPE " + }, + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "type": "text", + "content": " and Accel " + }, + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 217, + 288, + 262 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 270, + 288, + 353 + ], + "lines": [ + { + "bbox": [ + 48, + 270, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 270, + 288, + 353 + ], + "type": "table", + "html": "
MethodN(0,100)N(0,400)U(50)U(100)
M↓A↓M↓A↓M↓A↓M↓A↓
Noisy input61.2143.8122.5287.688.7207.9177.4415.8
GFpose [8]40.4-60.0-45.6-84.4-
SmoothNet [42]35.14.341.54.937.74.244.95.1
SynSP20.63.728.04.424.44.132.44.6
", + "image_path": "e7dfb1cf1b4ff02a9f1abb2ebd08064c26bf39f71e9513131118b4fe87738823.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 365, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 532 + ], + "type": "text", + "content": "2D Representation. In Tab. 1, we compare SynSP with other motion refinement methods in Human3.6M dataset with 2D pose estimator Hourglass [27]. With the setting of 8 window length, our SynSP outperforms all the methods while it is about 2 seconds faster (except the One-Euro) according to Fig. 4. Compared with SmoothNet, which is trained on multiple datasets, i.e., AIST-VIBE-3D, 3DPW-SPIN-3D, and Human3.6M-FCN-3D, SynSP can still outperforms it by " + }, + { + "bbox": [ + 46, + 365, + 287, + 532 + ], + "type": "inline_equation", + "content": "16.9\\%" + }, + { + "bbox": [ + 46, + 365, + 287, + 532 + ], + "type": "text", + "content": " for MPJPE and " + }, + { + "bbox": [ + 46, + 365, + 287, + 532 + ], + "type": "inline_equation", + "content": "25.2\\%" + }, + { + "bbox": [ + 46, + 365, + 287, + 532 + ], + "type": "text", + "content": " for PA-MPJPE with smaller window size. According to the ablation experiment results in Sec. 4.3, the significant accuracy improvement of SynSP is attributed to the mutual patching between the pose quality information (PQE) and multi-view information (PSE)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "content": "3D Representation. In Tab. 1, we also compare SynSP with other motion refinement methods in Human3.6M dataset with 3D pose estimator FCN [26]. Similar with the comparison in 2D Representation, we can still outperforms SmoothNet by " + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "inline_equation", + "content": "20.8\\%" + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "content": " for MPJPE and " + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "inline_equation", + "content": "18.6\\%" + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "content": " for PA-MPJPE in Human3.6M dataset. There is also a decrease of " + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "content": " for the Accel error. Furthermore, SynSP obtains these results only with 8 window size, while SmoothNet and most filters need 32 frames." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": "SMPL Representation. Then we compare SynSP with other motion refinement methods in AIST++ dataset with SMPL pose estimator SPIN as shown in Tab. 1. SynSP shows superior performance in SMPL representation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Comparison with Video-based Pose Estimators. The video-based estimators already include temporal informa" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": "tion. Compared with the image-based estimators, motion refinement methods have difficult to optimize the sequences from video-based estimators. In Tab. 2, we compare SynSP with SOTA methods and SmoothNet in AIST++ and 3DPW datasets with 3D pose video-based estimator VIBE [17] and TCMR [6], respectively. We can find that the high performance video-based estimators, TCMR, is hard to be refined. SmoothNet improves the smoothness by processing a longer window length than the video-based estimators. SynSP greatly increases precision and smoothness in a shorter window length with its outstanding architecture as shown in Fig. 2. The image-based estimator, PARE [18], has excellent performance in precision, SynSP can also make up for its shortcomings in smoothness and outperform some video-based estimators." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "text", + "content": "Comparison with Multi-view Pose Estimators. We further verify the performance of SynSP with the multi-view pose estimator, i.e., PPT [25]. As shown in Tab. 2, we can notice that SynSP can further improve the precision of PPT, which has superiority performance by considering multiview connections. We also present the result of other multiview methods such as TransFusion for comparison. This experiment results further proves the multi-view refinement ability of PSE for video-based estimators." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 361, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 588 + ], + "type": "text", + "content": "Comparison with Pose Priors Methods. To ensure a fair comparison with human pose prior models, we also conducted denoise experiments following the methodology of the SOTA work [8] among pose priors methods. Here, window size of GFPose, SmoothNet, and SynSP are 1, 32, and 8 respectively, while GFPose, as a generative model for human pose prior task, has 1000 sampling steps, i.e., 1000 times of model inference. GFPose has not presented its Accel errors since it does not take temporal information into account. As presented in Tab. 3 and Tab. 4, SynSP can effectively solve MPJPE (Precision) problems. Across noisy Human3.6M and CMU-Mocap datasets, SynSP can improve the pose quality by about " + }, + { + "bbox": [ + 304, + 361, + 545, + 588 + ], + "type": "inline_equation", + "content": "54.5\\%" + }, + { + "bbox": [ + 304, + 361, + 545, + 588 + ], + "type": "text", + "content": " in terms of MPJPE and " + }, + { + "bbox": [ + 304, + 361, + 545, + 588 + ], + "type": "inline_equation", + "content": "96.0\\%" + }, + { + "bbox": [ + 304, + 361, + 545, + 588 + ], + "type": "text", + "content": " in terms of Accel errors for Gaussian noise with a variance of 100. For Gaussian noise with a large variance of 400, SynSP can improve the pose quality by about " + }, + { + "bbox": [ + 304, + 361, + 545, + 588 + ], + "type": "inline_equation", + "content": "67.2\\%" + }, + { + "bbox": [ + 304, + 361, + 545, + 588 + ], + "type": "text", + "content": " in terms of MPJPE and " + }, + { + "bbox": [ + 304, + 361, + 545, + 588 + ], + "type": "inline_equation", + "content": "97.2\\%" + }, + { + "bbox": [ + 304, + 361, + 545, + 588 + ], + "type": "text", + "content": " int terms of Accel errors. We also observe consistent improvement in uniform noise as shown in Tab. 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 597, + 402, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 402, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 402, + 610 + ], + "type": "text", + "content": "4.3. Ablation Study." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 616, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 652 + ], + "type": "text", + "content": "To ensure representativeness of ablation experiments, all experiments were performed on the largest dataset, Human3.6M, with 3D estimator FCN." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Modules of SynSP. In order to verify the rationality of the structure of SynSP, we conducted ablation experiments on each module in SynSP. In Tab. 5, the line without a tick represents the results of baseline, which is the network of base stage without PSE I. And this network surpasses Smooth-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1830" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 113, + 286, + 243 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 105 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 105 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 105 + ], + "type": "text", + "content": "Table 5. Ablation study on PSE I, RS (Refinement Stage), PSE II, and PQE modules. * denotes directly using the output from base stage as input of refinement stage to re-optimize these sequence." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 113, + 286, + 243 + ], + "lines": [ + { + "bbox": [ + 48, + 113, + 286, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 113, + 286, + 243 + ], + "type": "table", + "html": "
PSE IRSPSE IIPQEMPJPEAccel
51.21.09
✓48.71.07
✓*51.41.08
✓49.61.07
✓✓46.71.05
✓✓✓46.41.04
✓✓✓42.50.99
✓✓✓✓41.80.98
", + "image_path": "3986a39a4d6a09b41c382cbaec81b0e0e4643159a63870d336a5659873003bc7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "type": "text", + "content": "Net in precision by " + }, + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "type": "text", + "content": ". PSE I, RS, PSE II, and PQE all have gains in results, and the combination of PSE I, RS, and PQE has brought greater benefits. We believe that this benefit mainly comes from the combination of SynSP with global information PSE I, which can help generate more accurate quality information according to the poses from other views. Besides, utilizing the difference between acceleration and Position Branch improves both smoothness and precision of final prediction. It is worth noting that using the average value of the two branches as the input RS does not bring reasonable gains. We suspect that the output of base stage loses the original predicted position information, which affects the accuracy and makes the refinement stage useless compared with the baseline." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 425, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 688 + ], + "type": "text", + "content": "Time Efficiency Analysis. We conducted latency experiments on the Human3.6M dataset using FCN [26] as the 3D pose estimator with a batch size of 1416, as shown in Tab. 6. Considering the practical application environment, we report not only the inference latency of the model, as shown in the 3rd column of Tab. 6, but also the latency associated with the Sliding Window Average Algorithm (SWAA), which includes the delay in waiting for pose estimators and the computation cost of SWAA (we have optimized SWAA specifically for parallel computation on GPUs). The bold cell in each row of Tab. 6 represents the primary factor contributing to latency. For video processing speed of 12 frames per second for the pose estimator (based on the general inference speed of various pose estimators), a window size of 8 would require a waiting time of " + }, + { + "bbox": [ + 46, + 425, + 287, + 688 + ], + "type": "inline_equation", + "content": "8/12 = 667\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 425, + 287, + 688 + ], + "type": "text", + "content": "; similarly, a window size of 32 would require a waiting time of " + }, + { + "bbox": [ + 46, + 425, + 287, + 688 + ], + "type": "inline_equation", + "content": "32/12 = 2667\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 425, + 287, + 688 + ], + "type": "text", + "content": ". Furthermore, the model inference time for SynSP is only about 10ms. Overall, the main time cost primarily arises from the internal waiting latency for SmoothNet and SynSP. For the pose priors method GFPose, the main time delay is its 1000 sampling steps. Hence, SynSP is more suitable for near real-time pose refinement task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Analysis on Number of Views. As shown in Tab. 7, as more poses from different views are input at the same time" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 306, + 113, + 548, + 196 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "type": "text", + "content": "Table 6. Time efficiency analysis on SynSP including model inference time and time cost for sliding window. Besides, we assume that FPS (Frame Per Second) for pose estimator is 12." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 113, + 548, + 196 + ], + "lines": [ + { + "bbox": [ + 306, + 113, + 548, + 196 + ], + "spans": [ + { + "bbox": [ + 306, + 113, + 548, + 196 + ], + "type": "table", + "html": "
MethodsWSModel (ms)SWAATotal (ms)
Wait (ms)Execution (ms)
GFPose13772--3772
SmoothNet321.2126678023470
SynSP811.06670.3678
SynSP*812.36670.3680
", + "image_path": "36e3d9001c87066462db8f15c626805ff12a0e949b3bb63ec219b8fa3adb33cc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 271, + 545, + 367 + ], + "blocks": [ + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "type": "text", + "content": "Table 7. Contrast experiments on the number of input views. " + }, + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "type": "text", + "content": " means that we randomly combine several people together for input during the training process, instead of using multiple views for one person. * means that in the case of " + }, + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 305, + 206, + 545, + 262 + ], + "type": "text", + "content": ", several people with similar poses are input at the same time during the inference process." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 271, + 545, + 367 + ], + "lines": [ + { + "bbox": [ + 307, + 271, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 307, + 271, + 545, + 367 + ], + "type": "table", + "html": "
View NumberMPJPEPA-MPJPEAccel
151.440.11.02
247.237.61.02
346.537.11.00
441.833.30.98
4†50.439.51.01
4*49.739.11.01
", + "image_path": "d9fea97b989168d78f11220127f9b604a9155d29830f318fd7c95abea8c8b192.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 387, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 460 + ], + "type": "text", + "content": "during the training process and inference process, the performance of SynSP is more efficient. It is justifiable due to the introduction of more global information. If number of views is larger than 4, we believe that the performance of our model can be further promoted with only a marginal increase in model complexity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 470, + 378, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 470, + 378, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 470, + 378, + 483 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 490, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 490, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 490, + 545, + 622 + ], + "type": "text", + "content": "This paper proposes SynSP, a motion refinement method that utilizes the tension relationship between smoothness and precision through synergistic optimization. Specifically, SynSP outputs Position Branch and Acceleration Branch biased towards precision and smoothness, respectively. Then we employ the discrepancy between the two branches as a quality cue for the predicted sequence, which can be utilized in subsequent stages to further refine the pose sequence. In addition, we introduce Pose Similarity Encoding, achieving simultaneous refinement of multi-view poses and performance enhancement." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 633, + 399, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 399, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 399, + 647 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "The paper is supported by National Key R & D Program of China (No. 2021YFB3300100), Natural Science Foundation of China No.62102039, the Fundamental Research Funds for the Central Universities No.500422813, and Natural Science Foundation of China No.6232211." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1831" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] A. Benzine, F. Chabot, B. Luvison, Q. C. Pham, and C. Achard. Pandanet: Anchor-based single-shot multi-person 3d pose estimation. In CVPR, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "text", + "content": "[2] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In ECCV, pages 561-578, 2016. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "type": "text", + "content": "[3] Géry Casiez, Nicolas Roussel, and Daniel Vogel. 1€ filter: a simple speed-based low-pass filter for noisy input in interactive systems. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pages 2527-2530, 2012. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "type": "text", + "content": "[4] Yu Cheng, Bo Yang, Bo Wang, and Robby T Tan. 3d human pose estimation using spatio-temporal networks with explicit occlusion training. In AAAI, pages 10631-10638, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "type": "text", + "content": "[5] Yu Cheng, Bo Wang, Bo Yang, and Robby T. Tan. Monocular 3d multi-person pose estimation by integrating top-down and bottom-up networks. In CVPR, pages 7649-7659, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 345 + ], + "type": "text", + "content": "[6] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consistent 3d human pose and shape from a video. In CVPR, pages 1964-1973, 2021. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 347, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 287, + 390 + ], + "type": "text", + "content": "[7] Jiaming Chu, Lei Jin, Xiaojin Fan, Yinglei Teng, Yunchao Wei, Yuqiang Fang, Junliang Xing, and Jian Zhao. Single-stage multi-human parsing via point sets and center-based offsets. In ACM MM, pages 1863-1873, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 392, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 287, + 434 + ], + "type": "text", + "content": "[8] Hai Ci, Mingdong Wu, Wentao Zhu, Xiaoxuan Ma, Hao Dong, Fangwei Zhong, and Yizhou Wang. Gfpose: Learning 3d human pose prior with gradient fields. CVPR, 2023. 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 436, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 286, + 479 + ], + "type": "text", + "content": "[9] Israel Cohen, Yiteng Huang, Jingdong Chen, Jacob Benesty, Jacob Benesty, Jingdong Chen, Yiteng Huang, and Israel Cohen. Pearson correlation coefficient. Noise reduction in speech processing, pages 1-4, 2009. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 481, + 286, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 286, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 286, + 523 + ], + "type": "text", + "content": "[10] Andrey Davydov, Anastasia Remizova, Victor Constantin, Sina Honari, Mathieu Salzmann, and Pascal Fua. Adversarial parametric pose prior. In CVPR, pages 10997-11005, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "type": "text", + "content": "[11] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Stefano Alletto, and Rita Cucchiara. Compressed volumetric heatmaps for multi-person 3d pose estimation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 559, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 286, + 601 + ], + "type": "text", + "content": "[12] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE TPAMI, 36(7):1325-1339, 2013. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 603, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 286, + 667 + ], + "type": "text", + "content": "[13] Kyung-Min Jin, Byoung-Sung Lim, Gun-Hee Lee, Tae-Kyung Kang, and Seong-Whan Lee. Kinematic-aware hierarchical attention network for human pose estimation in videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5725-5734, 2023. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[14] Lei Jin, Xiaojuan Wang, Xuecheng Nie, Luoqi Liu, Yandong Guo, and Jian Zhao. Grouping by center: Predicting centripetal offsets for the bottom-up human pose estimation. IEEE TMM, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Lei Jin, Chenyang Xu, Xiaojuan Wang, Yabo Xiao, Yandong Guo, Xuecheng Nie, and Jian Zhao. Single-stage is enough: Multi-person absolute 3d pose estimation. In CVPR, pages 13086-13095, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[16] Lei Jin, Xiaojuan Wang, Xuecheng Nie, Wendong Wang, Yandong Guo, Shuicheng Yan, and Jian Zhao. Rethinking the person localization for single-stage multi-person pose estimation. IEEE TMM, 2023. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "text", + "content": "[17] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In CVPR, pages 5253-5263, 2020. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 198, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 239 + ], + "type": "text", + "content": "[18] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. In ICCV, pages 11127-11137, 2021. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[19] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "text", + "content": "[20] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In ICCV, pages 13401-13412, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 331, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 375 + ], + "type": "text", + "content": "[21] Yang Li, Kan Li, Shuai Jiang, Ziyue Zhang, Congzhen-tao Huang, and Richard Yi Da Xu. Geometry-driven self-supervised method for 3d human pose estimation. In AAAI, pages 11442-11449, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 410 + ], + "type": "text", + "content": "[22] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Transactions on Graphics (TOG), 39(4):40-1, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "type": "text", + "content": "[23] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM TOG, 34(6):1-16, 2015. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "text", + "content": "[24] Haoyu Ma, Liangjian Chen, Deying Kong, Zhe Wang, Xingwei Liu, Hao Tang, Xiangyi Yan, Yusheng Xie, Shih-Yao Lin, and Xiaohui Xie. Transfusion: Cross-view fusion with transformer for 3d human pose estimation. arXiv preprint arXiv:2110.09554, 2021. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "type": "text", + "content": "[25] Haoyu Ma, Zhe Wang, Yifei Chen, Deying Kong, Liangjian Chen, Xingwei Liu, Xiangyi Yan, Hao Tang, and Xiaohui Xie. Ppt: Token-pruned pose transformer for monocular and multi-view human pose estimation. In ECCV, 2022. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "text", + "content": "[26] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, pages 2640-2649, 2017. 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 579, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 612 + ], + "type": "text", + "content": "[27] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In ECCV, pages 483-499, 2016. 2, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "text", + "content": "[28] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[29] Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with temporal convolutions and semi-supervised training. In CVPR, 2019. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1832" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 610 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "text", + "content": "[30] William H Press and Saul A Teukolsky. Savitzky-golay smoothing filters. Computers in Physics, 4(6):669-672, 1990. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 149 + ], + "type": "text", + "content": "[31] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In ICCV, pages 11488-11499, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "type": "text", + "content": "[32] Garvita Tiwari, Dimitrije Antić, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. Pose-ndf: Modeling human pose manifolds with neural distance fields. In ECCV, pages 572–589, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 286, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 286, + 217 + ], + "type": "text", + "content": "[33] Carnegie Mellon University. Cmu graphics lab motion capture database: http://mocap.cs.cmu.edu/. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 286, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 286, + 261 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 286, + 261 + ], + "type": "text", + "content": "[34] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeuIPS, 30, 2017. 3, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 265, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 265, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 265, + 286, + 308 + ], + "type": "text", + "content": "[35] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, pages 601-617, 2018. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 310, + 286, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 310, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 49, + 310, + 286, + 352 + ], + "type": "text", + "content": "[36] Ziniu Wan, Zhengjia Li, Maoqing Tian, Jianbo Liu, Shuai Yi, and Hongsheng Li. Encoder-decoder with multi-level attention for 3d human shape and pose estimation. In ICCV, pages 13033-13042, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 354, + 286, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 354, + 286, + 397 + ], + "spans": [ + { + "bbox": [ + 49, + 354, + 286, + 397 + ], + "type": "text", + "content": "[37] Tao Wang, Lei Jin, Zhang Wang, Xiaojin Fan, Yu Cheng, Yinglei Teng, Junliang Xing, and Jian Zhao. Decenternet: Bottom-up human pose estimation via decentralized pose representation. In ACM MM, pages 1798-1808, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 399, + 286, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 286, + 442 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 286, + 442 + ], + "type": "text", + "content": "[38] Wen-Li Wei, Jen-Chun Lin, Tyng-Luh Liu, and Hong-Yuan Mark Liao. Capturing humans in motion: Temporal-attentive 3d human pose and shape estimation from monocular video. In CVPR, pages 13211-13220, 2022. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 445, + 286, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 445, + 286, + 486 + ], + "spans": [ + { + "bbox": [ + 49, + 445, + 286, + 486 + ], + "type": "text", + "content": "[39] J. Wu, H. Zheng, B. Zhao, Y. Li, B. Yan, R. Liang, W. Wang, S. Zhou, G. Lin, and Y. Fu. Ai challenger: A large-scale dataset for going deeper in image understanding. In ICME, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 489, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 489, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 49, + 489, + 286, + 520 + ], + "type": "text", + "content": "[40] Ian T Young and Lucas J Van Vliet. Recursive implementation of the gaussian filter. Signal processing, 44(2):139-151, 1995. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 523, + 286, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 523, + 286, + 565 + ], + "spans": [ + { + "bbox": [ + 49, + 523, + 286, + 565 + ], + "type": "text", + "content": "[41] Ailing Zeng, Xuan Ju, Lei Yang, Ruiyuan Gao, Xizhou Zhu, Bo Dai, and Qiang Xu. Deciwatch: A simple baseline for " + }, + { + "bbox": [ + 49, + 523, + 286, + 565 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 49, + 523, + 286, + 565 + ], + "type": "text", + "content": " efficient 2d and 3d pose estimation. In ECCV, pages 607-624. Springer, 2022. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 568, + 286, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 568, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 49, + 568, + 286, + 610 + ], + "type": "text", + "content": "[42] Ailing Zeng, Lei Yang, Xuan Ju, Jiefeng Li, Jianyi Wang, and Qiang Xu. Smoothnet: A plug-and-play network for refining human poses in videos. In ECCV, pages 625-642, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1833" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/92bc41a7-a993-4d23-a197-14e6dd1e0509_content_list.json b/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/92bc41a7-a993-4d23-a197-14e6dd1e0509_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a778f37b64dc1da47997b206bb71d36814f58446 --- /dev/null +++ b/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/92bc41a7-a993-4d23-a197-14e6dd1e0509_content_list.json @@ -0,0 +1,1988 @@ +[ + { + "type": "text", + "text": "SyncMask: Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining", + "text_level": 1, + "bbox": [ + 197, + 181, + 776, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chull Hwan Song $^{1*}$ Taebaek Hwang $^{1*}$ Jooyoung Yoon $^{1}$ Shunghyun Choi $^{1}$ Yeong Hyeon $\\mathrm{Gu}^{2\\dagger}$", + "bbox": [ + 145, + 243, + 833, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Dealicious Inc. $^{2}$ Sejong University", + "bbox": [ + 379, + 261, + 594, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 271, + 305, + 339, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language models (VLMs) have made significant strides in cross-modal understanding through large-scale paired datasets. However, in fashion domain, datasets often exhibit a disparity between the information conveyed in image and text. This issue stems from datasets containing multiple images of a single fashion item all paired with one text, leading to cases where some textual details are not visible in individual images. This mismatch, particularly when non-co-occurring elements are masked, undermines the training of conventional VLM objectives like Masked Language Modeling and Masked Image Modeling, thereby hindering the model's ability to accurately align fine-grained visual and textual features. Addressing this problem, we propose Synchronized attentional Masking (SyncMask), which generate masks that pinpoint the image patches and word tokens where the information co-occur in both image and text. This synchronization is accomplished by harnessing cross-attentional features obtained from a momentum model, ensuring a precise alignment between the two modalities. Additionally, we enhance grouped batch sampling with semihard negatives, effectively mitigating false negative issues in Image-Text Matching and Image-Text Contrastive learning objectives within fashion datasets. Our experiments demonstrate the effectiveness of the proposed approach, outperforming existing methods in three downstream tasks.", + "bbox": [ + 133, + 333, + 475, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 135, + 684, + 250, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, there has been rapid progress in developing Vision-Language Pretraining (VLP) [32, 38, 6, 39, 19, 18, 42, 23, 36, 20, 27, 3], paving the way to bridge the gap between visual and textual features. These VLP methods, trained on extensive image-text datasets, have enabled a deeper understanding of semantic alignment across different modalities. By fine-tuning these pretrained models for specific tasks, particularly in data-scarce areas like image-", + "bbox": [ + 133, + 707, + 475, + 811 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6be5b5d0e6cd21fae3e8adc905e60aefed227250a53ee29eae28a238a1fcc05a.jpg", + "image_caption": [ + "[images, caption]", + "Notched lapel collar. Padded shoulders....", + "Four-button cuffs. Two vents at back...", + "Figure 1. Example of misaligned masks in the MLM task." + ], + "image_footnote": [], + "bbox": [ + 508, + 329, + 552, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/596396104a1e803272469db78a90aba6efd435297d72c420f22fa425f3907645.jpg", + "image_caption": [ + "Dataset", + "Long sleeve blazer in deep navy.", + "(a)" + ], + "image_footnote": [], + "bbox": [ + 553, + 330, + 653, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4d929c6dfd94b47404609a70da777ed04be62b257c307440ec8bfa4276f7d8b9.jpg", + "image_caption": [ + "Input of MLM", + "[image, masked caption]", + "Long sleeve blazer [MASK] deep navy.", + "Notched lapel collar. [MASK] shoulders...", + "Four-button [MASK]. Two veins at [MASK].", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 736, + 330, + 779, + 402 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "text retrieval, notable performance improvements have been observed. In the pretraining phase, various factors such as model architecture, training objectives, and batch sampling techniques play a crucial role in effectively harnessing the joint representation of multi-modal data.", + "bbox": [ + 496, + 501, + 836, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, there are some issues in applying conventional generic VLP approaches to task-specific domains such as fashion. Fashion VLP models [10, 45, 13, 14, 15] typically employ objectives such as Masked Language Modeling (MLM) and Masked Image Modeling (MIM). These methods mask elements like text words or image patches, leveraging surrounding context for prediction or reconstruction. They boost cross-modality by making models infer masked text tokens or image patches from aligned features. However, existing MLM and MIM often suffer from inherent misalignment limitations because the masks are generated randomly, often leading to unmatched elements being masked.", + "bbox": [ + 496, + 568, + 838, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To illustrate these limitations, consider Figure 1 (a) shows a single description associated with four images from the FashionGen [37] dataset. In Figure 1 (b), a random masking scenario is shown where the blue [MASK] might lead the model to predict the masked word using only the text context, thereby not incorporating the image information. Similarly, the two red [MASK] tokens in (b) lack relevance to the accompanying image, thus hindering the model's ability to connecting between visual and textual features. Fur-", + "bbox": [ + 496, + 727, + 839, + 844 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 138, + 45, + 207, + 80 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 259, + 42, + 781, + 56 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 338, + 56, + 704, + 70 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 308, + 71, + 732, + 84 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 151, + 821, + 243, + 833 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 155, + 833, + 258, + 843 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13948", + "bbox": [ + 480, + 878, + 517, + 891 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "thermore, in the MIM task, random masking may inadvertently cover parts of the image that contain fashion items not described in the text, leading to mismatches during the training of the alignment of cross-modal features. To solve these problem, we propose SyncMask, selecting masks that represent synchronously co-occurring features utilizing the vision-language cross-attention map.", + "bbox": [ + 133, + 148, + 473, + 238 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, compared to general domains, the fashion domain often suffers from smaller dataset sizes and less variance in data distribution. This suggests that using standard VLP methods may not adequately distinguish the fine-grained features vital for fashion-related tasks. Thus, we pay attention to a grouped batch sampling technique [3] that similar samples are gradually collected within the batch as the training progresses, impacting the pretraining objectives of Image-Text Contrastive Learning (ITC) and Image-Text Matching (ITM). When similar samples exist within the same batch, it becomes more challenging to differentiate positives and negatives during the training of ITC and ITM compared to using random samples. This encourages the model to more intensely focus on learning fine-grained differences, even with less training.", + "bbox": [ + 133, + 239, + 475, + 433 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The existing grouped batch sampling method uses the output features of two uni-modal encoders to find the most similar sample. However, as shown in Figure 1, there are many data that have the same caption for multiple images as for the fashion domain. Therefore, a false negative problem that causes actual positive samples in the same batch to be wrongly labelled as negative when learning ITC and ITM arises if the existing methods are used without changes. To overcome this limitation, we propose a semi-hard negative sampling technique with lower similarities between samples that constitute the batch while removing the false negative.", + "bbox": [ + 133, + 434, + 473, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contributions of this study are:", + "bbox": [ + 150, + 577, + 454, + 589 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Synchronized Attentional Masking: We introduce SyncMask, which replaces random mask in MLM and MIM with targeted mask of co-occurring segments in image-text pairs. By utilizing cross-attention features from a momentum model to generate these masks, this method effectively addresses the problem of mismatched image-text inputs, thereby enhancing fine-grained alignment of cross-modal features.", + "2. Refined Grouped Batch Sampling: Our method incorporates semi-hard negative sampling to tackle data scarcity and distribution disparities in domain-specific datasets, thereby reducing false negatives." + ], + "bbox": [ + 145, + 589, + 473, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 135, + 756, + 264, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision and Language (VL) Model Recently, VLMs have focused on enhancing model architecture and designing objectives to integrate visual and textual features effectively. Early studies [32, 38, 39, 6] have used object detectors for extracting visual features as an input for a multi-modal trans", + "bbox": [ + 133, + 777, + 475, + 844 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/4e7e6204d8167c461444feb991e82751efaba6503449f8983911ebfe6f8f6228.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
METHODMMMLMMIMATMAVMONUP
DMAE [1]✓
MaskDistill [34]✓✓✓
AttMask [21]✓✓✓
ALBEF [27]✓✓
MaskVLM [24]✓✓✓
MAMO [44]✓✓✓
FashionBERT [10]✓✓✓
Kaleido-BERT [45]✓✓✓✓✓
FashionViL [13]✓✓✓
FashionSAP [15]✓✓
Ours✓✓✓✓✓✓
", + "bbox": [ + 498, + 145, + 838, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Related works vs. Ours on Masked Modeling. MM:MultiModal. MLM:Masked Language Modeling. MIM:Masked Image Modeling. ATM:Attentional Textual Mask. AVM:Attentional Visual Mask. OnUp:Online Update for Attentional Masking", + "bbox": [ + 498, + 320, + 838, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "former along with textual features. The objective for training models extends vanilla BERT [7] to use MLM, MIM, and ITM losses; however, the object detection module incurs a high computational cost for training and inference. Therefore, there have been attempts to replace it with CNN [19, 18] or linear projection [23]. These studies commonly train models with MLM and ITM, tailoring MIM to their specific architectures. Concurrently, CLIP [36], ALIGN [20] propose models comprising only two unimodal encoders, demonstrating the outstanding representation embedding capabilities of contrastive learning. ALBEF [27] add a contrastive learning objective to the previous multi-modal transformer structure for aligning the two modalities before fusion. Based on this model, GRIT-VLP [3] demonstrate improved learning efficiency when configuring batches with hard negative samples.", + "bbox": [ + 496, + 377, + 839, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "FashionVL Model In recent years, various studies [10, 45, 13, 14, 15] attempted to capture the finer details of images, building upon established models and pre-training objectives from the general VL task. FashionBERT [10] integrates patch-based image features and BERT-based text representations for addressing the limitations of region of interests (RoIs) in capturing fine-grained details. Kaleido-BERT [45] improves fine-grained fashion cross-modality representations through alignment guided masking compared to random masking. FashionViL [13] employs a versatile VLP framework, leveraging two pre-training tasks for capturing the rich fine-grained information of fashion data. FashionSAP [15] employs abstract fashion symbols and an attributes prompt technique for effectively modeling multi-modal fashion attributes. We clarify the importance of the attentional masking technique that employs alignment between MIM and MLM building upon preceding methods. In addition, we underscore a previously unaddressed need for grouped batch sampling within the fashion domain.", + "bbox": [ + 496, + 597, + 839, + 845 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13949", + "bbox": [ + 480, + 880, + 517, + 891 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9ca3eb8e0a9ba93e86ada5bc6bee0afaa51d79d29825dc4d30198641bcc90e07.jpg", + "image_caption": [ + "Figure 2. Overview of masking strategies using a teacher-student distillation framework. 1) Uni-modal models: (a) random masking, (b) teacher-guided attentional masking. 2) Multi-modal models: (c) random text masking, (d) random image/text masking, (e) teacher-guided cross-attentional masking (Ours)." + ], + "image_footnote": [], + "bbox": [ + 138, + 147, + 558, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attention-guided Masked Modeling MIM and MLM leverage unmasked contextual clues to predict masked visual and textual elements, respectively. As shown in the upper section of Table 1 and Figure 2 (a, b), in the MIM task, prior studies have evolved from random masking [16, 2, 41, 1] to attention-guided masking [29, 34, 21], demonstrating that targeting highly-attended patches with teacher-student distillation framework improves masked modeling outcomes. Masked modeling also has been pivotal in advancing cross-modal alignment within VLMs, spanning both general [38, 32, 27, 24, 44] and fashion-specific [10, 45, 13, 15] domains. This is briefly illustrated in Figure 2 (c, d) and delineated in the middle and lower sections of Table 1. These enhance the model's proficiency aligning co-occurring visual and textual representations. However, challenges emerge when masking non-co-occurring elements, which hinders the accurate pairing of visual and textual features. KaleidoBERT [45] improve fine-grained cross-modality representations through text-image alignment-guided masking. This requires additional components for an attention-based alignment generator, increasing computational demands and potential model overfitting on fixed text-image mask pairs. Our approach overcomes these limitations by progressively tailoring masks during end-to-end training.", + "bbox": [ + 138, + 316, + 475, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 135, + 638, + 220, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We provide an overview of the preliminary aspects, which includes the model architecture and two well-established training objectives for VLP. Subsequently, we present a detailed explanation of the proposed methods, which are synchronized attentional masked modeling and grouped batch sampling with semi-hard negatives.", + "bbox": [ + 133, + 661, + 473, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 135, + 744, + 256, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For an image-text pair, we refer input sequences as follows: tokenized text embeddings are denoted as $T = \\left[\\mathbf{t}_{[\\mathrm{CLS}]} \\right]$ , $\\mathbf{t}_1, \\ldots, \\mathbf{t}_N \\in \\mathbb{R}^{(N+1) \\times D}$ and visual patch embeddings represented as $V = \\left[\\mathbf{v}_{[\\mathrm{CLS}]} \\right]$ , $\\mathbf{v}_1, \\ldots, \\mathbf{v}_{N'} \\in \\mathbb{R}^{(N'+1) \\times D}$ . Here, $D, N$ , and $N'$ refer to the transformer dimension, the number of text tokens, and the number of", + "bbox": [ + 135, + 766, + 475, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "image patches, respectively. Additionally, $\\mathbf{t}_{[\\mathrm{CLS}]}$ and $\\mathbf{v}_{[\\mathrm{CLS}]}$ specifically reference the [CLS] embeddings.", + "bbox": [ + 498, + 318, + 834, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The model consists of two components: a teacher model, referred to as the momentum model, and a student model. The student model, denoted as $f_{\\theta}$ , is parameterized by $\\theta$ which includes a textual encoder $f_{\\theta}^{T}(T) \\in \\mathbb{R}^{(N + 1)\\times D}$ , visual encoder $f_{\\theta}^{I}(V) \\in \\mathbb{R}^{(N' + 1)\\times D}$ , and multi-modal encoder $f_{\\theta}^{M}(f_{\\theta}^{T}(T),f_{\\theta}^{I}(V)) \\in \\mathbb{R}^{(N + 1)\\times D}$ . For the momentum model $f_{\\theta'}$ , the training parameters are updated by the exponential moving average method, $\\theta' \\gets \\beta \\theta' + (1 - \\beta) \\theta$ where $\\beta$ represents a hyperparameter.", + "bbox": [ + 496, + 345, + 838, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-Text Contrastive Learning (ITC) At the front of the multi-modal encoder, ITC pre-aligns the joint latent space of the textual encoder and visual encoder. This objective have proved its effectiveness in VLMs [36, 27, 26, 3, 13, 15]. We also adopt the ITC loss framework proposed by [27, 26, 15], which incorporates a momentum encoder for utilizing soft labels as ITC training targets, thereby addressing potential positive instances within negative pairs.", + "bbox": [ + 496, + 484, + 838, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-Text Matching (ITM) For the ITM loss, the model classifies image-text pairs as either matched (positive) or not matched (negative) using a joint representation obtained from the [CLS] token output embedding of the multi-modal encoder. This vector is passed through an FC layer and softmax for binary prediction. Like ALBEF [27], we exploit hard negatives in the ITM task, identifying pairs that share similar semantics but differ in fine-grained details, using in-batch contrastive similarity from ITC.", + "bbox": [ + 496, + 606, + 836, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Synchronized Attentional Masked Modeling", + "text_level": 1, + "bbox": [ + 500, + 732, + 821, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We extend the use of momentum model, a self-supervised tool for momentum distillation outlined in MoCo [17] and ALBEF [27], beyond its conventional role of generating pseudo-labels. We employ its multi-modal encoder, which calculate the cross-attention map to identify patches and tokens where image and text features strongly correlate. These elements, indicated by heightened attention weights, are then", + "bbox": [ + 496, + 753, + 836, + 844 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13950", + "bbox": [ + 480, + 878, + 517, + 891 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2f3d9987da1aecb5d1a5c28fae82dfa1134df5e1d4c2025f6767f256364433d6.jpg", + "image_caption": [ + "Figure 3. A schematic overview of the SyncMask process: Leveraging cross-attention features from the teacher (momentum) model to generate informative masks for both MIM and MLM tasks. It is important to note that the input for MLM consists of unmasked image paired with masked text." + ], + "image_footnote": [], + "bbox": [ + 155, + 145, + 823, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "selectively masked. This attentional masking, targeting synchronously co-occurring features, enhances cross-modality over traditional random masking methods in the MLM and MIM phases. Moreover, the momentum model's output features provide enhanced labels for these masked regions, offering a depth of information beyond conventional discrete labels. We will explore this method further to understand its full potential in capturing intricate multi-modal interactions.", + "bbox": [ + 133, + 409, + 475, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision-Language Synchronized Attentional Masking MLM and MIM elevate the alignment of visual and textual representations in models. However, when non-co-occurring elements are masked, these techniques face limitations that restrict the model's ability to accurately match multi-modal features. To address this issue, we propose a Synchronized attentional Masking (SyncMask) strategy into masked multimodal modeling objectives. As depicted in Figure 3, we extracted two sets of synchronized attention weights from the cross-attention module of the multi-modal encoder's last layer in the momentum model. The module enables the model to fuse image-text features using a cross-attention mechanism that processes a query $(Q^T)$ , key $(K^I)$ , and value $(V^I)$ , as represented by the following equation:", + "bbox": [ + 135, + 527, + 473, + 710 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {A t t e n t i o n} \\left(Q _ {i} ^ {T}, K _ {i} ^ {I}, V _ {i} ^ {I}\\right) = \\alpha \\left(Q _ {i} ^ {T}, K _ {i} ^ {I}\\right) \\odot V _ {i} ^ {I} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 724, + 473, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $1 \\leq i \\leq H$ , with $H$ denoting the number of heads in the MHA, and $\\odot$ representing the Hadamard product. In this context, $\\alpha$ refers the cross-attention function, which can be expressed as:", + "bbox": [ + 133, + 750, + 473, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha \\left(Q _ {i} ^ {T}, K _ {i} ^ {I}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {i} ^ {T} \\left(K _ {i} ^ {I}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\in \\mathbb {R} ^ {(N ^ {\\prime} + 1) \\times (N + 1)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 803, + 472, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The function $\\alpha(Q_i^T, K_i^I)$ computes the attention weights for the image from the perspective of the text. Similarly, by altering the query $(Q)$ and key $(K)$ , we calculate the text attention weight from the image perspective, as represented by the following equation:", + "bbox": [ + 496, + 407, + 836, + 473 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha \\left(Q _ {i} ^ {I}, K _ {i} ^ {T}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {i} ^ {I} \\left(K _ {i} ^ {T}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\in \\mathbb {R} ^ {(N + 1) \\times \\left(N ^ {\\prime} + 1\\right)} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 474, + 836, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Utilizing Equation 2 and Equation 3, we derive two synchronized textual-visual cross-attention weights. These weights, $\\mathbf{o}^T\\in \\mathbb{R}^N$ $\\mathbf{o}^I\\in \\mathbb{R}^{N'}$ , are obtained by averaging the patch tokens of the last layer, excluding the [CLS] token. This process allows us to map each word in the sentence sequence to its corresponding attention in $\\mathbf{o}^T$ .Further, $\\mathbf{o}^I$ can be reshaped to $\\mathbb{R}^{P\\times P}$ , which aligns with the image patches.", + "bbox": [ + 496, + 515, + 838, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {I d x} ^ {T} = \\text {s h u f f l e} \\left(\\operatorname {s o r t} ^ {\\text {d e s c}} \\left(\\mathbf {o} ^ {T}\\right) [ \\geq L ]\\right) [ \\geq K ] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 619, + 836, + 636 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {I d x} ^ {I} = \\text {s h u f f l e} \\left(\\operatorname {s o r t} ^ {\\text {d e s c}} \\left(\\mathbf {o} ^ {I}\\right) [ \\geq L ^ {\\prime} ]\\right) [ \\geq K ^ {\\prime} ] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 651, + 836, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Equation 4 and Equation 5, we first sort the attention weights $\\mathbf{o}^T$ and $\\mathbf{o}^I$ in descending order (sort desc), and then extract their indices (Idx). Equation 4 focuses on indices corresponding to the top $L$ values of $\\mathbf{o}^T$ , where $L$ is a threshold greater than the actual mask size $K$ . These indices are randomly shuffled (shuffle) to introduce randomness, and ultimately, only those indices that satisfy the condition $K \\leq L$ are retained. A similar approach is applied to $K'$ and $L'$ in Equation 5. The parameters $K, K', L$ , and $L'$ are defined based on a mask ratio $r \\in [0,1]$ .", + "bbox": [ + 496, + 674, + 838, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final attention masks for textual and visual components are represented by the vectors $\\mathbf{m}^T$ and $\\mathbf{m}^I$ , respectively. The computation of these masks utilizes the indices", + "bbox": [ + 496, + 805, + 838, + 844 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13951", + "bbox": [ + 480, + 878, + 514, + 891 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e19aee941cafce9f5e9026208f9f0a8216253559363566a22c879e9481486e16.jpg", + "image_caption": [ + "Figure 4. Selection phase of the SyncMask" + ], + "image_footnote": [], + "bbox": [ + 138, + 146, + 472, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{I}\\mathrm{dx}^T$ for text and $\\mathbf{I}\\mathrm{dx}^I$ for images, as formulated in the equation below:", + "bbox": [ + 135, + 397, + 473, + 424 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} ^ {T} \\left[ \\mathrm {I d} \\mathbf {x} ^ {T} \\right] \\leftarrow 1, \\quad \\mathbf {m} ^ {I} \\left[ \\mathrm {I d} \\mathbf {x} ^ {I} \\right] \\leftarrow 1 \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 431, + 472, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initially, the vectors $\\mathbf{m}^T$ and $\\mathbf{m}^I$ are zero-initialized with dimensions $N$ (for text) and $N'$ (for images). They are then updated to binary attention masks, which are integral in capturing the synchronized interaction between textual and visual elements, as illustrated in Figure 4. Subsequently, these masks are employed in the masked modeling processes for text and images (detailed in Equation 7 and Equation 8).", + "bbox": [ + 133, + 454, + 473, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building upon this foundation, the text mask vector $\\mathbf{m}^T$ consists of elements $[m_1^t,\\dots,m_N^t ]\\in \\{0,1\\} ^N$ , while the image mask vector $\\mathbf{m}^I$ is composed of $[m_1^i,\\dots,m_{N'}^i ]\\in$ $\\{0,1\\}^{N^{\\prime}}$ . For each tokenized text vector, the masked version $\\tilde{\\mathbf{t}}_i$ is determined by:", + "bbox": [ + 135, + 545, + 473, + 612 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {t}} _ {j} = \\left(1 - m _ {j} ^ {t}\\right) \\cdot \\mathbf {t} _ {j} + m _ {j} ^ {t} \\cdot \\mathbf {t} _ {\\text {m a s k}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 619, + 472, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this equation, $1 \\leq j \\leq N$ , and $\\mathbf{t}_{\\mathrm{mask}}$ denotes the special token used for textual masking. In the context of Masked Image Modeling (MIM), each image patch is processed with a learnable mask, resulting in the masked image vector $\\tilde{\\mathbf{v}}_k$ :", + "bbox": [ + 135, + 642, + 472, + 695 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {v}} _ {k} = \\left(1 - m _ {k} ^ {i}\\right) \\cdot \\mathbf {v} _ {k} + m _ {k} ^ {i} \\cdot \\mathbf {v} _ {\\text {m a s k}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 702, + 472, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $1 \\leq k \\leq N'$ , and $\\mathbf{v}_{\\mathrm{mask}}$ represents the learnable mask embedding [4]. The masked tokenized inputs are thus represented as $\\tilde{T} = [\\mathbf{t}_{\\mathrm{cls}}; \\tilde{\\mathbf{t}}_1; \\ldots; \\tilde{\\mathbf{t}}_N]$ for text and $\\tilde{V} = [\\mathbf{v}_{\\mathrm{cls}}; \\tilde{\\mathbf{v}}_1; \\ldots; \\tilde{\\mathbf{v}}_{N'}]$ for image. These masked inputs are processed using the proposed SyncMask $\\mathbf{m}^T$ and $\\mathbf{m}^I$ .", + "bbox": [ + 135, + 724, + 473, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synchronized Attentional Masked Language Modeling MLM predicts masked words based on the surrounding contextual text and image. Many existing VLMs applied the", + "bbox": [ + 135, + 805, + 473, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MLM method proposed in BERT [7], randomly masking words with a probability of $15\\%$ . However, this approach may not be suitable for vision-language datasets with short caption lengths, especially for nonstandard datasets such as fashion, demanding a thorough understanding of fine-grained attributes. We leverage previous works that addressed these issues by increasing masking probabilities [3] and employing attribute prompts [15]. Building upon these methods, we employ masks generated by SyncMask which is contextually attuned to the corresponding image.", + "bbox": [ + 496, + 148, + 836, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $h(V, \\tilde{T})$ denote the model's predicted probability for a masked token. $\\tilde{y}$ denote a one-hot vocabulary distribution in which the ground-truth token is assigned a probability of 1. MLM minimize cross-entropy loss, described as follows:", + "bbox": [ + 496, + 277, + 836, + 330 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M L M}} = \\mathbb {E} _ {(V, \\hat {T}) \\sim D} [ \\operatorname {C E} (\\tilde {\\boldsymbol {y}}, \\boldsymbol {h} (V, \\hat {T}) ], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 335, + 834, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{CE}(\\cdot ,\\cdot)$ refers the cross-entropy between two vectors.", + "bbox": [ + 496, + 358, + 836, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synchronized Attentional Masked Image Modeling In the distillation-based MIM [21, 29, 34], a teacher encoder sees the full image, whereas the student encoder, seeing the masked image, tackles the reconstruction objective. Our method adopts a similar objective framework, but with a key difference: our masks, generated through SyncMask, are designed to reflect text-informed elements in masked image. To calculate the MIM loss, the following approach is used:", + "bbox": [ + 496, + 387, + 838, + 491 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {D i s t} \\left(\\boldsymbol {f} _ {\\theta^ {\\prime}} ^ {I} (V), \\boldsymbol {f} _ {\\theta} ^ {I} (\\tilde {V})\\right) \\\\ = \\frac {1}{\\Omega (\\mathbf {m} ^ {I})} \\sum_ {k = 1} ^ {N ^ {\\prime}} \\mathbf {m} _ {k} ^ {I} \\cdot \\ell_ {1} ^ {S m o o t h} \\left(\\boldsymbol {f} _ {\\theta^ {\\prime}} ^ {I} (V) _ {k}, \\boldsymbol {f} _ {\\theta} ^ {I} (\\tilde {V}) _ {k}\\right) \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 503, + 836, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{f}_{\\theta^{\\prime}}^{I}(\\cdot)_{k}$ and $\\pmb{f}_{\\theta}^{I}(\\cdot)_{k}$ represent the output feature of teacher and student model for the $k$ -th image patch, respectively. $\\Omega (\\cdot)$ means the number of elements with a value of 1 in a vector.", + "bbox": [ + 496, + 570, + 838, + 621 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {1} ^ {\\text {S m o o t h}} (a, b) = \\left\\{ \\begin{array}{l l} 0. 5 (a - b) ^ {2} & \\text {i f} | a - b | < \\gamma \\\\ | a - b | - 0. 5 & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 625, + 836, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\ell_1^{Smooth}$ [11] represents a robust L1 loss less sensitive to outliers than the L2 loss and $\\gamma$ is a hyperparameter set to 1. Conclusively, the training objective of MIM can be formulated as:", + "bbox": [ + 496, + 667, + 836, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M I M}} = \\mathbb {E} _ {(V, \\hat {V}) \\sim D} [ \\operatorname {D i s t} \\left(\\boldsymbol {f} _ {\\theta^ {\\prime}} ^ {I} (V), \\boldsymbol {f} _ {\\theta} ^ {I} (\\tilde {V})\\right) ] \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 732, + 836, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final loss $(\\mathcal{L})$ is given as:", + "bbox": [ + 500, + 757, + 666, + 771 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M I M}} + \\mathcal {L} _ {\\mathrm {M L M}} + \\mathcal {L} _ {\\mathrm {I T C}} + \\mathcal {L} _ {\\mathrm {I T M}} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 781, + 836, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{ITC}}$ and $\\mathcal{L}_{\\mathrm{ITM}}$ denote ITC and ITM, respectively. Due to space constraints, detailed formulations of these two losses are elaborated in the Appendix.", + "bbox": [ + 496, + 805, + 836, + 844 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13952", + "bbox": [ + 480, + 880, + 517, + 891 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7a5b45d914818f866db2f60af7e33792daedf1a9b9645d491d726f23c84928f4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
METHODSI2T*T2I*MEAN*I2TT2IMEAN
R@1R@1R@1R@1
VSE++ [9]4.594.604.60---
VL-BERT [38]19.2622.6320.95---
ViLBERT [32]20.9721.1221.05---
Image-BERT [35]22.7624.7823.77---
OSCAR [28]23.3925.1024.25---
FashionBERT [10]23.9626.7525.36---
Kaleido-BERT [45]27.9933.8830.94---
CommerceMM [43]41.6039.6062.75---
EI-CLIP [33]38.7040.0639.3825.7028.4027.05
ALBEF [27]63.9760.5262.2041.6850.9546.32
FashionViL [13]65.5461.8863.7142.8851.3447.11
FashionSAP [15]73.1470.1271.6354.4362.8258.63
Ours75.0071.0073.0055.3964.0659.73
", + "bbox": [ + 205, + 145, + 767, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Cross-modal retrieval result on FashionGen [37] in the sub/full set of evaluation following previous work. $\\star$ sub set.", + "bbox": [ + 205, + 368, + 767, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Grouped Batch with Semi-hard Negatives", + "text_level": 1, + "bbox": [ + 135, + 397, + 442, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the generic domain, [3] proposed the GRIT strategy for enhancing training effectiveness by forming mini-batches with similar examples. In this strategy, the grouping based on similarity phase plays a crucial role. During this phase, similarity calculations are performed in both directions, utilizing the [CLS] outputs from the unimodal encoders. For each example, the algorithm iteratively identifies the index with the highest similarity, alternating between the image-to-text and text-to-image directions in a sequential manner. These highly similar indices are grouped together within the mini-batch, ensuring that each mini-batch consists of examples exhibiting the highest possible similarity.", + "bbox": [ + 133, + 417, + 475, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, in fashion datasets, this strategy leads to a false-negative problem in ITM and ITC, where true positives are mislabeled as negatives within mini-batches. To address this problem, we opt to group semi-hard negatives with relatively lower similarity (the $s^{th}$ highest pairwise similarity) instead of the highest $(1^{st})$ during the grouping based on similarity phase, where $s$ represents a predefined hyperparameter that is greater than 1. In addition, we prevent true positive samples from grouping by considering the item indices of the samples. Through this approach, the model is trained with negatives that are similar but exhibit meaningful differences, thereby enabling the effective learning of fine-grained distinctions with a limited dataset. More details are in the Appendix.", + "bbox": [ + 133, + 575, + 473, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 135, + 760, + 250, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 135, + 782, + 321, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The foundational architecture of the proposed model is aligned with prior works for demonstrating the effectiveness of the proposed approach [27, 3, 15]. The image encoder", + "bbox": [ + 133, + 805, + 473, + 844 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "adopt the architecture of ViT-B16 [8], whereas the text encoder comprises the first six blocks of the BERT-bas3 [7]. The multi-modal encoder extends the self-attention layers of the last six blocks of BERT with the cross-attention layers. The proposed model is initialized with pre-trained ALBEF [27] same as the Fashion-SAP to ensure a fair comparison [15]. In addition, we employ the same data augmentation and prompt input strategies as FashionSAP [15]. During pretraining, we conduct experiments using 8 RTX 3090 GPUs each with a batch size of 8 for 30 epochs. We adopt a momentum queue size of 48,000 to facilitate grouped batch sampling, which is consistent with GRIT-VLP [3]. The input image size is set to $256 \\times 256$ . We apply the AdamW [31] optimizer with a learning rate of 6e-5.", + "bbox": [ + 496, + 397, + 839, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Datasets", + "text_level": 1, + "bbox": [ + 500, + 588, + 588, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FashionGen [37] FashionGen comprises 320K text-image pairs and 40K unique fashion items, each represented by multiple images from different angles. For pre-training, we employ the FashionGen train set, which contains approximately 260.5K text-image pairs. In addition, FashionGen supports various downstream tasks, including text-to-image retrieval, image-to-text retrieval, category recognition, and subcategory recognition.", + "bbox": [ + 496, + 608, + 838, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FashionIQ [40] FashionIQ encompasses 77K unique fashion items and includes 18K training triplets (i.e., query image, modified text, target image) and 6K validation datasets for a text-guided image retrieval task. It contains three different categories: Dress, Toptee, and Shirt.", + "bbox": [ + 496, + 722, + 838, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Downstream Tasks", + "text_level": 1, + "bbox": [ + 500, + 797, + 656, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-modal Retrieval We evaluate a cross-modal retrieval task that includes image-text retrieval (ITR) and text", + "bbox": [ + 500, + 816, + 838, + 844 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13953", + "bbox": [ + 480, + 878, + 517, + 891 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/08d1cb3eaa6aae78202f92b240280ee5c838d122718473275bac5f7f09dbc1f0.jpg", + "image_caption": [ + "is light blue in color and has more blue" + ], + "image_footnote": [], + "bbox": [ + 189, + 147, + 233, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23e59536dbf2dded67e37b69ff33c9ae716f43c10a29aa995c31d072cea5f6eb.jpg", + "image_caption": [ + "is yellow and has short sleeves and is beige and shorter sleeves" + ], + "image_footnote": [], + "bbox": [ + 200, + 204, + 228, + 248 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8170751b04763694d4eb3f1fb6cf964f3bba9ef3b41a6b3fe97abdbf08155f2d.jpg", + "image_caption": [ + "is light blue in color and has more blue" + ], + "image_footnote": [], + "bbox": [ + 191, + 256, + 230, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c31d8ede41a43e06be8211a0ab6e007994f76763cb64cde4e1e17e6dff8bf8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 148, + 372, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/94e388a82e2ff592d1c2ed228c55692492ed9e48b26c2c48d8ce49541fae0a44.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 148, + 410, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5bcb4a08e5c54e53f3c8e507e123a9e76ea93cfe83871d4f2a6193b53e97e3a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 150, + 462, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a370e0610b30d30aebc21b72d677fe590e829d50020d476a887a20935a21e1ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 472, + 155, + 501, + 193 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/624a08638b624bf5cdd6f7fe9b1d7fc67fe107b88de443fc6c06085d9c16a586.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 155, + 549, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b441dc7c3ee79658ef78c52d3f4f26c07008d9df38600cae675a55bddb212663.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 568, + 148, + 594, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea28ef8ad5e218aa8874f36d2750a73727f9d914f8a59d4c0b1404fb1907a1e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 151, + 635, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/318f0906cf0a562e3f0ab193a6e7cbfd9d053ca25270046b1e14cca38f87233a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 148, + 691, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4e4f78659e3e372f60ac4ebfcd3d872c018fbb1d3b02775f877a33de274ae55e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 150, + 727, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83d40df87bede9d3578974a9d082cd47a8d877f484bc3e954fc2c69a162e7be6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 151, + 774, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d8266be9434b5fd28869d4c2b931cfd73e10b0a8bbcc29a28c8081c3ea0476b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 202, + 364, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20fe4e6dcf5048f0bc320a9f9665d414fa03e49b246d94907b01027548d45152.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 203, + 416, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c15649a5457f0cbd6ad1496a51f15875a5a034642c7bb201bd9976bf3aafdf3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 203, + 454, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/282e09030c25ecbc8e69764919c3f6d773171f6e66d53e73d5b84c44d08592d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 472, + 203, + 500, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/57acc5e9ccd96d50c72998ae79e895bd874d46fc43ce9151d5eb12b372570bdb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 205, + 550, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/86dfa22cf839e7351b28ef3ac4c3bfa84481d65078ee0c143c708268ec723bce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 204, + 594, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/acc07d0ffaf8df3e76cdca01693354eeaa3812b6364c8d49ceb76598e5a189b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 203, + 637, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0255ecdf2a25cf15b2f15b9bc4053e116e64d4f910ed649a074361432cf273cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 202, + 687, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1cd8d6951877adcb2f4d504332429f3a65d0fbaac4ade7446264282245d4678f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 202, + 728, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec91c71b5b9074869a80b19f3bc33ae407edfa3c5270b3dad56a50bfad5be857.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 203, + 776, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1baab022be951c854d8b399acd73fc441aaf27c07c3a2bde1e497970d08f80cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 253, + 372, + 300 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6af5f3a7e30792093bf7b4762ab058ca532e094b8cd9ce891869b0d0b9ddd038.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 255, + 410, + 300 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a16f39e1dfd022b3e9a9f2b58e7ce60f75514a3b3a7a46aafa05eb061129530.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 253, + 462, + 300 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6f6d12ec51ad75be41402fc4fc4d7fff212048e7b508ca136e94c60c66adb3aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 472, + 260, + 501, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e7e624a202a285f28ed5079046737346ea5fe036c71c57602fc3a8ca5eb53f5.jpg", + "image_caption": [ + "Figure 5. The top-10 TGIR results of the SyncMask model on the FashionIQ dataset. On the left, the reference images paired with their guided descriptions are shown, while the right side presents the model's predicted images ranked by descending scores. Ground truth images are distinctly outlined with a green bounding box. It is worth mentioning that the set of predictions includes other images that also qualify as suitable matches." + ], + "image_footnote": [], + "bbox": [ + 508, + 253, + 555, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c1dc7a10fe31403f843fb4318d3da5750bdaf9375a3cfcdb912c16598fcd73b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 260, + 599, + 299 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f4b841bad0e88fd318f474c0906466606356288703f40b4925a13a89294664e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 253, + 645, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/572004d130ae93c223711f78f570eae81443ae4042a9188c228a01f1ba518f49.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 253, + 694, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c38f01af8724c542738866a5529290a45d23d75f1df949fe4de0c19dab3d5c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 253, + 733, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8942212fffe07cc177c1440480bb860a806223e831294f54e57d47ec97fb1486.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 255, + 784, + 299 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/17c2bf65f4e306b1735e90c7b078686e2142be17114d8c475cedaaa93c6b9918.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
METHODSDRESSTOPTEESHIRTMEAN
R@10R@50R@10R@50R@10R@50R@10R@50
CIRR [30]17.4540.4121.6445.3817.5338.8118.8741.53
VAL [5]22.5344.0027.5351.6822.3844.1524.1546.61
CosMo [25]25.6450.3029.2157.4624.9049.1826.5852.31
DCNet [22]28.9556.730.4458.2923.9547.327.7854.10
FashionVLP [12]32.4260.2938.5168.7931.8958.4434.2762.51
FashionViL [13]33.4759.9434.9860.7925.1750.3931.2157.04
FashionSAP [15]33.7160.4341.9170.9333.1761.3336.2664.23
Ours33.7661.2344.8272.0635.8262.1238.1365.14
", + "bbox": [ + 163, + 354, + 805, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "image retrieval (TIR). Further, ITR focuses on finding relevant textual descriptions for a given image query. TIR, the inverse task, retrieves pertinent images based on a textual query. These tasks assess the effectiveness of the model in capturing cross-modal relationships between text and images within retrieval scenarios. Following the previous works [13, 15], we evaluate cross-modal retrieval not only on the subset with 1K retrievals but also on the full dataset of FashionGen [37]. The results, including R@1 scores for both subset and the full set, are presented in Table Table 2, demonstrating an improved performance compared to that of the previous results.", + "bbox": [ + 133, + 540, + 475, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text-guided Image Retrieval This task aims to retrieve target images by considering query pairs that reference image and modified descriptions; this is more challenging than traditional retrievals. Therefore, we need to select the target image for identifying minor differences in the changes in description while maintaining the characteristics of the reference image. For a fair comparison, we adopt a similar fine-tuning as outlined in the FashionSAP [15]. In the actual dataset, there are many images that match with the query pairs in addition to the target image referred to as the real correct answer, as shown in Figure 5. Thus, a qualitative", + "bbox": [ + 133, + 700, + 475, + 844 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cc2b7a59bd5387cc75c8fed1b4a1268b1506881d0c301f8b750edaffecdcfd79.jpg", + "table_caption": [ + "Table 3. Text-guided image retrieval performance in FashionIQ [40]" + ], + "table_footnote": [], + "table_body": "
METHODSCRSCR
AccMacro-FAccMacro-F
F-BERT [10]91.2570.5085.2762.00
K-BERT [45]95.0771.4088.0763.60
F-ViL [13]97.4888.6092.2383.02
FashionSAP [15]98.3489.8494.3387.67
Ours98.4190.3194.2187.83
", + "bbox": [ + 524, + 537, + 808, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. CR and SCR results on FashionGen [37].", + "bbox": [ + 553, + 625, + 782, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "evaluation can be conducted in that the model finds similar images well in addition to the actual correct answer. As shown in Table 3, the proposed model surpasses previous models and demonstrates state-of-the-art performance.", + "bbox": [ + 498, + 662, + 836, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Category/Subcategory Recognition In this downstream task, the objective is to classify the category and subcategory of fashion items using the textual and visual information provided. In line with earlier studies [10, 45, 13, 15], we simply attach a linear layer to the [CLS] token, which serves as the fusion feature, for task label prediction. As indicated in Table 4, our proposed model exhibits competitive performance compared to existing models.", + "bbox": [ + 496, + 739, + 836, + 844 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13954", + "bbox": [ + 480, + 880, + 517, + 891 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b38fe64ecf743820b70109df9a7c0010bce8a8afd65f64ac4c0ada4d69883d6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 145, + 460, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/83a69f65308f388961b7250158c55abbf440375d89af6fc9e0c4c1ee3b7be5d9.jpg", + "image_caption": [ + "(a) Unmasked input", + "(b) Random mask", + "(c) SyncMask", + "Figure 6. Examples of random masking (b) and SyncMask (c) for MIM and MLM. The latter applies masks to pertinent features in both the input image and text (a), offering a more context-sensitive selection than the former." + ], + "image_footnote": [], + "bbox": [ + 143, + 270, + 460, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 135, + 464, + 267, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Random vs. Attentional Masking We ablate the SyncMask with other non-synchronized masking methods. Table 5 lists the comparison results for the combination of random and attentional masking. In this experiment, we evaluated five experiments for I2T (R@1), T2I (R@1), CR (Macro-F), SCR (Macro-F), and TMIR (R@10), which were evaluated in the previous experiments. In these experiments, the results of the three experiments with at least one attentional masking were higher than those with no attentional masking (all random). This can be considered an indication of the efficacy of attentional masking. Further, the synergistic performance of the proposed synchronized textual-visual attentional masking (+2.14) is greater than that of only one attentional masking (+0.4, +0.66), which shows the superiority of the SyncMask approach. Figure 6 displays the distinction between applying random masks (b) and SyncMask (c) to unmasked image-text pairs (a) from the FashionGen [37]. SyncMask strategically places masks on image patches that correlate with the text and applies masks to the text informed by the image details, unlike random masking.", + "bbox": [ + 133, + 484, + 475, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Random vs. Hardest vs. Semi-hard Negative Sampling Table 6 compare the effectiveness of various grouped minibatch sampling strategies across five downstream tasks. The methods evaluated encompassed four scenarios: random grouping, hardest grouping (+1.90), hardest grouping with exclusion of false negatives (+1.98), and semi-hard grouping", + "bbox": [ + 135, + 765, + 473, + 844 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2c90a8c5606be799d997609ca0f0e498d2927d3ffb33bb9febcb27f2881f560e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RTRVATAVMEANGAINI2T R@1T21 R@1CR Macro-FSCR Macro-FTMIR R@10
✓✓70.3173.3069.8086.2186.1636.08
✓✓70.71+0.4074.8070.1086.4986.3135.84
✓✓70.97+0.6674.3070.8085.1286.7837.87
✓✓72.45+2.1475.0071.0090.3187.8338.13
", + "bbox": [ + 500, + 146, + 834, + 210 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/76f6f20e9b568ae8c846fbd50bc360737dabe47e89ee0996f64ca1d62361fcf3.jpg", + "table_caption": [ + "Table 5. Ablation study results for Random vs. Attention Masked Modeling on five downstream tasks. Rt:Random Textual masking. Rv:Random Visual masking. At:Attentional Textual masking. Av:Attentional Visual masking." + ], + "table_footnote": [], + "table_body": "
GROUPEFNMEANGAIN12T R@1T2I R@1CR Macro-FSCR Macro-FTMIR R@10
Random69.9972.3069.0086.2185.4037.06
Hardest71.89+1.9074.7070.5090.2086.6437.40
Hardest✓71.97+1.9874.3071.0090.1087.1537.28
Semi-hard✓72.45+2.4675.0071.0090.3187.8338.13
", + "bbox": [ + 500, + 257, + 834, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study results comparing the Grouped Batch Sampling Strategy across five downstream tasks. GROUP: Strategy for grouping phase of GRIT. EFN: Exclude False Negative in a grouping phase using index.", + "bbox": [ + 498, + 324, + 836, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "also excluding false negatives (+2.46). These findings suggest that grouping similar samples in a mini-batch is more beneficial for learning than composing batches with random samples. However, given that the fashion dataset often has multiple captions per image or vice versa, performance gains were observed when systematically preventing the grouping of false negatives by using indexing. Nevertheless, due to the prevalence of inherently similar samples that cannot be systematically excluded, we opted for grouping semi-hard negatives instead of the hardest ones, which resulted in a significant performance boost. This highlights the importance of further research from a data perspective, not just in terms of model architecture or loss functions.", + "bbox": [ + 496, + 378, + 838, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 559, + 602, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced Synchronized attentional Masking for enhanced masked modeling in fashion-centric VLMs. Leveraging cross-attention features of a momentum model, our method tailors the random mask into a targeted mask for synchronously co-occurring segments in image-text pairs in MLM and MIM objectives. This approach effectively resolves misaligned image-text input issues and improving fine-grained cross-modal representation. Additionally, we addressed data scarcity and distribution challenges in fashion datasets, refining grouped batch sampling with semi-hard negatives for ITM and ITC losses. The experimental results showed our methods outperformed established benchmarks in multiple downstream tasks.", + "bbox": [ + 496, + 580, + 838, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 761, + 630, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) under the metaverse support program to nurture the best talents (IITP-2024-RS-2023-00254529) grant funded by the Korea government (MSIT).", + "bbox": [ + 496, + 777, + 836, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13955", + "bbox": [ + 480, + 880, + 517, + 891 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 137, + 146, + 218, + 159 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yutong Bai, Zeyu Wang, Junfei Xiao, Chen Wei, Huiyu Wang, Alan L Yuille, Yuyin Zhou, and Cihang Xie. Masked autoencoders enable efficient knowledge distillers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24256-24265, 2023.", + "[2] Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT pretraining of image transformers. In ICLR, 2022.", + "[3] Jaeseok Byun, Taebaek Hwang, Jianlong Fu, and Taesup Moon. Grit-vlp: Grouped mini-batch sampling for efficient vision and language pre-training. In European Conference on Computer Vision, pages 395-412. Springer, 2022.", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021.", + "[5] Yanbei Chen, Shaogang Gong, and Loris Bazzani. Image search with text feedback by visiolinguistic attention learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3001-3011, 2020.", + "[6] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020.", + "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, 2019.", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.", + "[9] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017.", + "[10] Dehong Gao, Linbo Jin, Ben Chen, Minghui Qiu, Peng Li, Yi Wei, Yi Hu, and Hao Wang. Fashionbert: Text and image matching with adaptive loss for cross-modal retrieval. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2251-2260, 2020.", + "[11] Ross Girshick. Fast R-CNN. In ICCV, 2015.", + "[12] Sonam Goenka, Zhaoheng Zheng, Ayush Jaiswal, Rakesh Chada, Yue Wu, Varsha Hedau, and Pradeep Natarajan. Fashionvlp: Vision language transformer for fashion retrieval with feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14105-14115, 2022.", + "[13] Xiao Han, Licheng Yu, Xiatian Zhu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Fashionvil: Fashion-focused vision-and-language representation learning. In ECCV. Springer, 2022." + ], + "bbox": [ + 138, + 167, + 473, + 843 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Xiao Han, Xiatian Zhu, Licheng Yu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Fame-vil: Multi-tasking vision-language model for heterogeneous fashion tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2669-2680, 2023.", + "[15] Yunpeng Han, Lisai Zhang, Qingcai Chen, Zhijian Chen, Zhonghua Li, Jianxin Yang, and Zhao Cao. Fashionsap: Symbols and attributes prompt for fine-grained fashion vision-language pre-training. In CVPR, 2023.", + "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In arXiv:2111.06377, 2021.", + "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020.", + "[18] Zhicheng Huang, Zhaoyang Zeng, Yupan Huang, Bei Liu, Dongmei Fu, and Jianlong Fu. Seeing out of the box: End-to-end pre-training for vision-language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12976-12985, 2021.", + "[19] Zhicheng Huang, Zhaoyang Zeng, Bei Liu, Dongmei Fu, and Jianlong Fu. Pixel-bert: Aligning image pixels with text by deep multi-modal transformers. 2020.", + "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021.", + "[21] Ioannis Kakogeorgiou, Spyros Gidaris, Bill Psomas, Yannis Avrithis, Andrei Bursuc, Konstantinos Karantzalos, and Nikos Komodakis. What to hide from your students: Attention-guided masked image modeling. In ECCV, 2022.", + "[22] Jongseok Kim, Youngjae Yu, Hoeseong Kim, and Gunhee Kim. Dual compositional learning in interactive image retrieval. In AAAI, pages 1771-1779, 2021.", + "[23] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021.", + "[24] Gukyeong Kwon, Zhaowei Cai, Avinash Ravichandran, Erhan Bas, Rahul Bhotika, and Stefano Soatto. Masked vision and language modeling for multi-modal representation learning. In ICLR, 2023.", + "[25] Seungmin Lee, Dongwan Kim, and Bohyung Han. Cosmo: Content-style modulation for image retrieval with text feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 802-812, 2021.", + "[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022.", + "[27] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with" + ], + "bbox": [ + 501, + 148, + 836, + 843 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13956", + "bbox": [ + 480, + 880, + 517, + 891 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021.", + "[28] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020.", + "[29] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, et al. MST: Masked self-supervised transformer for visual representation. In NeurIPS, volume 34, 2021.", + "[30] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2125-2134, 2021.", + "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017.", + "[32] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019.", + "[33] Haoyu Ma, Handong Zhao, Zhe Lin, Ajinkya Kale, Zhangyang Wang, Tong Yu, Jiumiang Gu, Sunav Choudhary, and Xiaohui Xie. Ei-clip: Entity-aware interventional contrastive learning for e-commerce cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18051-18061, 2022.", + "[34] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. A unified view of masked image modeling. In arXiv preprint arXiv:2210.10615, 2022.", + "[35] Di Qi, Lin Su, Jia Song, Edward Cui, Taroon Bharti, and Arun Sacheti. Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966, 2020.", + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021.", + "[37] Negar Rostamzadeh, Seyedarian Hosseini, Thomas Boquet, Wojciech Stokowiec, Ying Zhang, Christian Jauvin, and Chris Pal. Fashion-gen: The generative fashion dataset and challenge. arXiv preprint arXiv:1806.08317, 2018.", + "[38] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. In International Conference on Learning Representations, 2019.", + "[39] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. 2019.", + "[40] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11307-11317, 2021." + ], + "bbox": [ + 137, + 148, + 473, + 842 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022.", + "[42] Hongwei Xue, Yupan Huang, Bei Liu, Houwen Peng, Jianlong Fu, Houqiang Li, and Jiebo Luo. Probing inter-modality: Visual parsing with self-attention for vision-and-language pre-training. volume 34, pages 4514-4528, 2021.", + "[43] Licheng Yu, Jun Chen, Animesh Sinha, Mengjiao Wang, Yu Chen, Tamara L Berg, and Ning Zhang. Commercemm: Large-scale commerce multimodal representation learning with omni retrieval. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4433-4442, 2022.", + "[44] Zijia Zhao, Longteng Guo, Xingjian He, Shuai Shao, Zehuan Yuan, and Jing Liu. Mamo: Fine-grained vision-language representations learning with masked multimodal modeling. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1528-1538, 2023.", + "[45] Mingchen Zhuge, Dehong Gao, Deng-Ping Fan, Linbo Jin, Ben Chen, Haoming Zhou, Minghui Qiu, and Ling Shao. Kaleido-bert: Vision-language pre-training on fashion domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12647–12657, 2021." + ], + "bbox": [ + 501, + 148, + 836, + 462 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13957", + "bbox": [ + 480, + 880, + 516, + 890 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/92bc41a7-a993-4d23-a197-14e6dd1e0509_model.json b/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/92bc41a7-a993-4d23-a197-14e6dd1e0509_model.json new file mode 100644 index 0000000000000000000000000000000000000000..30598b91132cb7ce33f47cb2539b12c51001d1b9 --- /dev/null +++ b/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/92bc41a7-a993-4d23-a197-14e6dd1e0509_model.json @@ -0,0 +1,2607 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.14, + 0.046, + 0.209, + 0.082 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.26, + 0.044, + 0.782, + 0.058 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.339, + 0.058, + 0.705, + 0.071 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.31, + 0.072, + 0.733, + 0.085 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.199, + 0.182, + 0.777, + 0.222 + ], + "angle": 0, + "content": "SyncMask: Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.244, + 0.834, + 0.261 + ], + "angle": 0, + "content": "Chull Hwan Song\\(^{1*}\\) Taebaek Hwang\\(^{1*}\\) Jooyoung Yoon\\(^{1}\\) Shunghyun Choi\\(^{1}\\) Yeong Hyeon \\(\\mathrm{Gu}^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.38, + 0.262, + 0.595, + 0.276 + ], + "angle": 0, + "content": "\\(^{1}\\)Dealicious Inc. \\(^{2}\\)Sejong University" + }, + { + "type": "title", + "bbox": [ + 0.272, + 0.306, + 0.34, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.334, + 0.477, + 0.659 + ], + "angle": 0, + "content": "Vision-language models (VLMs) have made significant strides in cross-modal understanding through large-scale paired datasets. However, in fashion domain, datasets often exhibit a disparity between the information conveyed in image and text. This issue stems from datasets containing multiple images of a single fashion item all paired with one text, leading to cases where some textual details are not visible in individual images. This mismatch, particularly when non-co-occurring elements are masked, undermines the training of conventional VLM objectives like Masked Language Modeling and Masked Image Modeling, thereby hindering the model's ability to accurately align fine-grained visual and textual features. Addressing this problem, we propose Synchronized attentional Masking (SyncMask), which generate masks that pinpoint the image patches and word tokens where the information co-occur in both image and text. This synchronization is accomplished by harnessing cross-attentional features obtained from a momentum model, ensuring a precise alignment between the two modalities. Additionally, we enhance grouped batch sampling with semihard negatives, effectively mitigating false negative issues in Image-Text Matching and Image-Text Contrastive learning objectives within fashion datasets. Our experiments demonstrate the effectiveness of the proposed approach, outperforming existing methods in three downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.685, + 0.251, + 0.699 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.708, + 0.476, + 0.812 + ], + "angle": 0, + "content": "Recently, there has been rapid progress in developing Vision-Language Pretraining (VLP) [32, 38, 6, 39, 19, 18, 42, 23, 36, 20, 27, 3], paving the way to bridge the gap between visual and textual features. These VLP methods, trained on extensive image-text datasets, have enabled a deeper understanding of semantic alignment across different modalities. By fine-tuning these pretrained models for specific tasks, particularly in data-scarce areas like image-" + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.305, + 0.597, + 0.314 + ], + "angle": 0, + "content": "Dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.315, + 0.612, + 0.323 + ], + "angle": 0, + "content": "[images, caption]" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.33, + 0.553, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.332, + 0.655, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.411, + 0.627, + 0.42 + ], + "angle": 0, + "content": "Long sleeve blazer in deep navy." + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.422, + 0.661, + 0.43 + ], + "angle": 0, + "content": "Notched lapel collar. Padded shoulders...." + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.433, + 0.645, + 0.442 + ], + "angle": 0, + "content": "Four-button cuffs. Two vents at back..." + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.451, + 0.589, + 0.46 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.306, + 0.785, + 0.314 + ], + "angle": 0, + "content": "Input of MLM" + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.315, + 0.798, + 0.323 + ], + "angle": 0, + "content": "[image, masked caption]" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.331, + 0.781, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.411, + 0.822, + 0.421 + ], + "angle": 0, + "content": "Long sleeve blazer [MASK] deep navy." + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.422, + 0.835, + 0.431 + ], + "angle": 0, + "content": "Notched lapel collar. [MASK] shoulders..." + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.433, + 0.836, + 0.442 + ], + "angle": 0, + "content": "Four-button [MASK]. Two veins at [MASK]." + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.451, + 0.767, + 0.459 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.473, + 0.8, + 0.485 + ], + "angle": 0, + "content": "Figure 1. Example of misaligned masks in the MLM task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.837, + 0.567 + ], + "angle": 0, + "content": "text retrieval, notable performance improvements have been observed. In the pretraining phase, various factors such as model architecture, training objectives, and batch sampling techniques play a crucial role in effectively harnessing the joint representation of multi-modal data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.839, + 0.726 + ], + "angle": 0, + "content": "However, there are some issues in applying conventional generic VLP approaches to task-specific domains such as fashion. Fashion VLP models [10, 45, 13, 14, 15] typically employ objectives such as Masked Language Modeling (MLM) and Masked Image Modeling (MIM). These methods mask elements like text words or image patches, leveraging surrounding context for prediction or reconstruction. They boost cross-modality by making models infer masked text tokens or image patches from aligned features. However, existing MLM and MIM often suffer from inherent misalignment limitations because the masks are generated randomly, often leading to unmatched elements being masked." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.84, + 0.845 + ], + "angle": 0, + "content": "To illustrate these limitations, consider Figure 1 (a) shows a single description associated with four images from the FashionGen [37] dataset. In Figure 1 (b), a random masking scenario is shown where the blue [MASK] might lead the model to predict the masked word using only the text context, thereby not incorporating the image information. Similarly, the two red [MASK] tokens in (b) lack relevance to the accompanying image, thus hindering the model's ability to connecting between visual and textual features. Fur-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.152, + 0.823, + 0.245, + 0.834 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.156, + 0.834, + 0.259, + 0.844 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.152, + 0.823, + 0.259, + 0.844 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.88, + 0.518, + 0.892 + ], + "angle": 0, + "content": "13948" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.135, + 0.149, + 0.475, + 0.239 + ], + "angle": 0, + "content": "thermore, in the MIM task, random masking may inadvertently cover parts of the image that contain fashion items not described in the text, leading to mismatches during the training of the alignment of cross-modal features. To solve these problem, we propose SyncMask, selecting masks that represent synchronously co-occurring features utilizing the vision-language cross-attention map." + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.24, + 0.476, + 0.434 + ], + "angle": 0, + "content": "In addition, compared to general domains, the fashion domain often suffers from smaller dataset sizes and less variance in data distribution. This suggests that using standard VLP methods may not adequately distinguish the fine-grained features vital for fashion-related tasks. Thus, we pay attention to a grouped batch sampling technique [3] that similar samples are gradually collected within the batch as the training progresses, impacting the pretraining objectives of Image-Text Contrastive Learning (ITC) and Image-Text Matching (ITM). When similar samples exist within the same batch, it becomes more challenging to differentiate positives and negatives during the training of ITC and ITM compared to using random samples. This encourages the model to more intensely focus on learning fine-grained differences, even with less training." + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.435, + 0.475, + 0.577 + ], + "angle": 0, + "content": "The existing grouped batch sampling method uses the output features of two uni-modal encoders to find the most similar sample. However, as shown in Figure 1, there are many data that have the same caption for multiple images as for the fashion domain. Therefore, a false negative problem that causes actual positive samples in the same batch to be wrongly labelled as negative when learning ITC and ITM arises if the existing methods are used without changes. To overcome this limitation, we propose a semi-hard negative sampling technique with lower similarities between samples that constitute the batch while removing the false negative." + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.578, + 0.455, + 0.59 + ], + "angle": 0, + "content": "In summary, the main contributions of this study are:" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.59, + 0.475, + 0.693 + ], + "angle": 0, + "content": "1. Synchronized Attentional Masking: We introduce SyncMask, which replaces random mask in MLM and MIM with targeted mask of co-occurring segments in image-text pairs. By utilizing cross-attention features from a momentum model to generate these masks, this method effectively addresses the problem of mismatched image-text inputs, thereby enhancing fine-grained alignment of cross-modal features." + }, + { + "type": "text", + "bbox": [ + 0.147, + 0.696, + 0.475, + 0.747 + ], + "angle": 0, + "content": "2. Refined Grouped Batch Sampling: Our method incorporates semi-hard negative sampling to tackle data scarcity and distribution disparities in domain-specific datasets, thereby reducing false negatives." + }, + { + "type": "list", + "bbox": [ + 0.147, + 0.59, + 0.475, + 0.747 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.757, + 0.265, + 0.77 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.779, + 0.476, + 0.845 + ], + "angle": 0, + "content": "Vision and Language (VL) Model Recently, VLMs have focused on enhancing model architecture and designing objectives to integrate visual and textual features effectively. Early studies [32, 38, 39, 6] have used object detectors for extracting visual features as an input for a multi-modal trans" + }, + { + "type": "table", + "bbox": [ + 0.499, + 0.146, + 0.839, + 0.32 + ], + "angle": 0, + "content": "
METHODMMMLMMIMATMAVMONUP
DMAE [1]✓
MaskDistill [34]✓✓✓
AttMask [21]✓✓✓
ALBEF [27]✓✓
MaskVLM [24]✓✓✓
MAMO [44]✓✓✓
FashionBERT [10]✓✓✓
Kaleido-BERT [45]✓✓✓✓✓
FashionViL [13]✓✓✓
FashionSAP [15]✓✓
Ours✓✓✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.321, + 0.839, + 0.364 + ], + "angle": 0, + "content": "Table 1. Related works vs. Ours on Masked Modeling. MM:MultiModal. MLM:Masked Language Modeling. MIM:Masked Image Modeling. ATM:Attentional Textual Mask. AVM:Attentional Visual Mask. OnUp:Online Update for Attentional Masking" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.84, + 0.574 + ], + "angle": 0, + "content": "former along with textual features. The objective for training models extends vanilla BERT [7] to use MLM, MIM, and ITM losses; however, the object detection module incurs a high computational cost for training and inference. Therefore, there have been attempts to replace it with CNN [19, 18] or linear projection [23]. These studies commonly train models with MLM and ITM, tailoring MIM to their specific architectures. Concurrently, CLIP [36], ALIGN [20] propose models comprising only two unimodal encoders, demonstrating the outstanding representation embedding capabilities of contrastive learning. ALBEF [27] add a contrastive learning objective to the previous multi-modal transformer structure for aligning the two modalities before fusion. Based on this model, GRIT-VLP [3] demonstrate improved learning efficiency when configuring batches with hard negative samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.84, + 0.846 + ], + "angle": 0, + "content": "FashionVL Model In recent years, various studies [10, 45, 13, 14, 15] attempted to capture the finer details of images, building upon established models and pre-training objectives from the general VL task. FashionBERT [10] integrates patch-based image features and BERT-based text representations for addressing the limitations of region of interests (RoIs) in capturing fine-grained details. Kaleido-BERT [45] improves fine-grained fashion cross-modality representations through alignment guided masking compared to random masking. FashionViL [13] employs a versatile VLP framework, leveraging two pre-training tasks for capturing the rich fine-grained information of fashion data. FashionSAP [15] employs abstract fashion symbols and an attributes prompt technique for effectively modeling multi-modal fashion attributes. We clarify the importance of the attentional masking technique that employs alignment between MIM and MLM building upon preceding methods. In addition, we underscore a previously unaddressed need for grouped batch sampling within the fashion domain." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.881, + 0.518, + 0.892 + ], + "angle": 0, + "content": "13949" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.148, + 0.56, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.285, + 0.839, + 0.308 + ], + "angle": 0, + "content": "Figure 2. Overview of masking strategies using a teacher-student distillation framework. 1) Uni-modal models: (a) random masking, (b) teacher-guided attentional masking. 2) Multi-modal models: (c) random text masking, (d) random image/text masking, (e) teacher-guided cross-attentional masking (Ours)." + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.318, + 0.476, + 0.631 + ], + "angle": 0, + "content": "Attention-guided Masked Modeling MIM and MLM leverage unmasked contextual clues to predict masked visual and textual elements, respectively. As shown in the upper section of Table 1 and Figure 2 (a, b), in the MIM task, prior studies have evolved from random masking [16, 2, 41, 1] to attention-guided masking [29, 34, 21], demonstrating that targeting highly-attended patches with teacher-student distillation framework improves masked modeling outcomes. Masked modeling also has been pivotal in advancing cross-modal alignment within VLMs, spanning both general [38, 32, 27, 24, 44] and fashion-specific [10, 45, 13, 15] domains. This is briefly illustrated in Figure 2 (c, d) and delineated in the middle and lower sections of Table 1. These enhance the model's proficiency aligning co-occurring visual and textual representations. However, challenges emerge when masking non-co-occurring elements, which hinders the accurate pairing of visual and textual features. KaleidoBERT [45] improve fine-grained cross-modality representations through text-image alignment-guided masking. This requires additional components for an attention-based alignment generator, increasing computational demands and potential model overfitting on fixed text-image mask pairs. Our approach overcomes these limitations by progressively tailoring masks during end-to-end training." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.64, + 0.221, + 0.654 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.662, + 0.475, + 0.74 + ], + "angle": 0, + "content": "We provide an overview of the preliminary aspects, which includes the model architecture and two well-established training objectives for VLP. Subsequently, we present a detailed explanation of the proposed methods, which are synchronized attentional masked modeling and grouped batch sampling with semi-hard negatives." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.746, + 0.258, + 0.759 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.767, + 0.476, + 0.844 + ], + "angle": 0, + "content": "For an image-text pair, we refer input sequences as follows: tokenized text embeddings are denoted as \\( T = \\left[\\mathbf{t}_{[\\mathrm{CLS}]} \\right] \\), \\( \\mathbf{t}_1, \\ldots, \\mathbf{t}_N \\in \\mathbb{R}^{(N+1) \\times D} \\) and visual patch embeddings represented as \\( V = \\left[\\mathbf{v}_{[\\mathrm{CLS}]} \\right] \\), \\( \\mathbf{v}_1, \\ldots, \\mathbf{v}_{N'} \\in \\mathbb{R}^{(N'+1) \\times D} \\). Here, \\( D, N \\), and \\( N' \\) refer to the transformer dimension, the number of text tokens, and the number of" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.319, + 0.835, + 0.345 + ], + "angle": 0, + "content": "image patches, respectively. Additionally, \\(\\mathbf{t}_{[\\mathrm{CLS}]}\\) and \\(\\mathbf{v}_{[\\mathrm{CLS}]}\\) specifically reference the [CLS] embeddings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.346, + 0.839, + 0.468 + ], + "angle": 0, + "content": "The model consists of two components: a teacher model, referred to as the momentum model, and a student model. The student model, denoted as \\( f_{\\theta} \\), is parameterized by \\( \\theta \\) which includes a textual encoder \\( f_{\\theta}^{T}(T) \\in \\mathbb{R}^{(N + 1)\\times D} \\), visual encoder \\( f_{\\theta}^{I}(V) \\in \\mathbb{R}^{(N' + 1)\\times D} \\), and multi-modal encoder \\( f_{\\theta}^{M}(f_{\\theta}^{T}(T),f_{\\theta}^{I}(V)) \\in \\mathbb{R}^{(N + 1)\\times D} \\). For the momentum model \\( f_{\\theta'} \\), the training parameters are updated by the exponential moving average method, \\( \\theta' \\gets \\beta \\theta' + (1 - \\beta) \\theta \\) where \\( \\beta \\) represents a hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.839, + 0.59 + ], + "angle": 0, + "content": "Image-Text Contrastive Learning (ITC) At the front of the multi-modal encoder, ITC pre-aligns the joint latent space of the textual encoder and visual encoder. This objective have proved its effectiveness in VLMs [36, 27, 26, 3, 13, 15]. We also adopt the ITC loss framework proposed by [27, 26, 15], which incorporates a momentum encoder for utilizing soft labels as ITC training targets, thereby addressing potential positive instances within negative pairs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.838, + 0.724 + ], + "angle": 0, + "content": "Image-Text Matching (ITM) For the ITM loss, the model classifies image-text pairs as either matched (positive) or not matched (negative) using a joint representation obtained from the [CLS] token output embedding of the multi-modal encoder. This vector is passed through an FC layer and softmax for binary prediction. Like ALBEF [27], we exploit hard negatives in the ITM task, identifying pairs that share similar semantics but differ in fine-grained details, using in-batch contrastive similarity from ITC." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.733, + 0.822, + 0.747 + ], + "angle": 0, + "content": "3.2. Synchronized Attentional Masked Modeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.754, + 0.838, + 0.845 + ], + "angle": 0, + "content": "We extend the use of momentum model, a self-supervised tool for momentum distillation outlined in MoCo [17] and ALBEF [27], beyond its conventional role of generating pseudo-labels. We employ its multi-modal encoder, which calculate the cross-attention map to identify patches and tokens where image and text features strongly correlate. These elements, indicated by heightened attention weights, are then" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.88, + 0.518, + 0.892 + ], + "angle": 0, + "content": "13950" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.146, + 0.825, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.374, + 0.839, + 0.397 + ], + "angle": 0, + "content": "Figure 3. A schematic overview of the SyncMask process: Leveraging cross-attention features from the teacher (momentum) model to generate informative masks for both MIM and MLM tasks. It is important to note that the input for MLM consists of unmasked image paired with masked text." + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.41, + 0.477, + 0.513 + ], + "angle": 0, + "content": "selectively masked. This attentional masking, targeting synchronously co-occurring features, enhances cross-modality over traditional random masking methods in the MLM and MIM phases. Moreover, the momentum model's output features provide enhanced labels for these masked regions, offering a depth of information beyond conventional discrete labels. We will explore this method further to understand its full potential in capturing intricate multi-modal interactions." + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.529, + 0.475, + 0.711 + ], + "angle": 0, + "content": "Vision-Language Synchronized Attentional Masking MLM and MIM elevate the alignment of visual and textual representations in models. However, when non-co-occurring elements are masked, these techniques face limitations that restrict the model's ability to accurately match multi-modal features. To address this issue, we propose a Synchronized attentional Masking (SyncMask) strategy into masked multimodal modeling objectives. As depicted in Figure 3, we extracted two sets of synchronized attention weights from the cross-attention module of the multi-modal encoder's last layer in the momentum model. The module enables the model to fuse image-text features using a cross-attention mechanism that processes a query \\((Q^T)\\), key \\((K^I)\\), and value \\((V^I)\\), as represented by the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.726, + 0.474, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\text {A t t e n t i o n} \\left(Q _ {i} ^ {T}, K _ {i} ^ {I}, V _ {i} ^ {I}\\right) = \\alpha \\left(Q _ {i} ^ {T}, K _ {i} ^ {I}\\right) \\odot V _ {i} ^ {I} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.75, + 0.474, + 0.802 + ], + "angle": 0, + "content": "where \\(1 \\leq i \\leq H\\), with \\(H\\) denoting the number of heads in the MHA, and \\(\\odot\\) representing the Hadamard product. In this context, \\(\\alpha\\) refers the cross-attention function, which can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.804, + 0.473, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\alpha \\left(Q _ {i} ^ {T}, K _ {i} ^ {I}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {i} ^ {T} \\left(K _ {i} ^ {I}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\in \\mathbb {R} ^ {(N ^ {\\prime} + 1) \\times (N + 1)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.409, + 0.837, + 0.474 + ], + "angle": 0, + "content": "The function \\(\\alpha(Q_i^T, K_i^I)\\) computes the attention weights for the image from the perspective of the text. Similarly, by altering the query \\((Q)\\) and key \\((K)\\), we calculate the text attention weight from the image perspective, as represented by the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.475, + 0.837, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\alpha \\left(Q _ {i} ^ {I}, K _ {i} ^ {T}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {i} ^ {I} \\left(K _ {i} ^ {T}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\in \\mathbb {R} ^ {(N + 1) \\times \\left(N ^ {\\prime} + 1\\right)} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.839, + 0.62 + ], + "angle": 0, + "content": "Utilizing Equation 2 and Equation 3, we derive two synchronized textual-visual cross-attention weights. These weights, \\(\\mathbf{o}^T\\in \\mathbb{R}^N\\) \\(\\mathbf{o}^I\\in \\mathbb{R}^{N'}\\) , are obtained by averaging the patch tokens of the last layer, excluding the [CLS] token. This process allows us to map each word in the sentence sequence to its corresponding attention in \\(\\mathbf{o}^T\\) .Further, \\(\\mathbf{o}^I\\) can be reshaped to \\(\\mathbb{R}^{P\\times P}\\) , which aligns with the image patches." + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.62, + 0.837, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\operatorname {I d x} ^ {T} = \\text {s h u f f l e} \\left(\\operatorname {s o r t} ^ {\\text {d e s c}} \\left(\\mathbf {o} ^ {T}\\right) [ \\geq L ]\\right) [ \\geq K ] \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.652, + 0.837, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\operatorname {I d x} ^ {I} = \\text {s h u f f l e} \\left(\\operatorname {s o r t} ^ {\\text {d e s c}} \\left(\\mathbf {o} ^ {I}\\right) [ \\geq L ^ {\\prime} ]\\right) [ \\geq K ^ {\\prime} ] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.839, + 0.805 + ], + "angle": 0, + "content": "In Equation 4 and Equation 5, we first sort the attention weights \\(\\mathbf{o}^T\\) and \\(\\mathbf{o}^I\\) in descending order (sort desc), and then extract their indices (Idx). Equation 4 focuses on indices corresponding to the top \\(L\\) values of \\(\\mathbf{o}^T\\), where \\(L\\) is a threshold greater than the actual mask size \\(K\\). These indices are randomly shuffled (shuffle) to introduce randomness, and ultimately, only those indices that satisfy the condition \\(K \\leq L\\) are retained. A similar approach is applied to \\(K'\\) and \\(L'\\) in Equation 5. The parameters \\(K, K', L\\), and \\(L'\\) are defined based on a mask ratio \\(r \\in [0,1]\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.806, + 0.839, + 0.845 + ], + "angle": 0, + "content": "The final attention masks for textual and visual components are represented by the vectors \\(\\mathbf{m}^T\\) and \\(\\mathbf{m}^I\\), respectively. The computation of these masks utilizes the indices" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.88, + 0.516, + 0.892 + ], + "angle": 0, + "content": "13951" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.147, + 0.473, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.367, + 0.404, + 0.379 + ], + "angle": 0, + "content": "Figure 4. Selection phase of the SyncMask" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.398, + 0.474, + 0.425 + ], + "angle": 0, + "content": "\\(\\mathbf{I}\\mathrm{dx}^T\\) for text and \\(\\mathbf{I}\\mathrm{dx}^I\\) for images, as formulated in the equation below:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.432, + 0.473, + 0.449 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} ^ {T} \\left[ \\mathrm {I d} \\mathbf {x} ^ {T} \\right] \\leftarrow 1, \\quad \\mathbf {m} ^ {I} \\left[ \\mathrm {I d} \\mathbf {x} ^ {I} \\right] \\leftarrow 1 \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.455, + 0.475, + 0.546 + ], + "angle": 0, + "content": "Initially, the vectors \\(\\mathbf{m}^T\\) and \\(\\mathbf{m}^I\\) are zero-initialized with dimensions \\(N\\) (for text) and \\(N'\\) (for images). They are then updated to binary attention masks, which are integral in capturing the synchronized interaction between textual and visual elements, as illustrated in Figure 4. Subsequently, these masks are employed in the masked modeling processes for text and images (detailed in Equation 7 and Equation 8)." + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.546, + 0.474, + 0.613 + ], + "angle": 0, + "content": "Building upon this foundation, the text mask vector \\(\\mathbf{m}^T\\) consists of elements \\([m_1^t,\\dots,m_N^t ]\\in \\{0,1\\} ^N\\) , while the image mask vector \\(\\mathbf{m}^I\\) is composed of \\([m_1^i,\\dots,m_{N'}^i ]\\in\\) \\(\\{0,1\\}^{N^{\\prime}}\\) . For each tokenized text vector, the masked version \\(\\tilde{\\mathbf{t}}_i\\) is determined by:" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.62, + 0.473, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {t}} _ {j} = \\left(1 - m _ {j} ^ {t}\\right) \\cdot \\mathbf {t} _ {j} + m _ {j} ^ {t} \\cdot \\mathbf {t} _ {\\text {m a s k}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.643, + 0.473, + 0.696 + ], + "angle": 0, + "content": "In this equation, \\(1 \\leq j \\leq N\\), and \\(\\mathbf{t}_{\\mathrm{mask}}\\) denotes the special token used for textual masking. In the context of Masked Image Modeling (MIM), each image patch is processed with a learnable mask, resulting in the masked image vector \\(\\tilde{\\mathbf{v}}_k\\):" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.703, + 0.473, + 0.718 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {v}} _ {k} = \\left(1 - m _ {k} ^ {i}\\right) \\cdot \\mathbf {v} _ {k} + m _ {k} ^ {i} \\cdot \\mathbf {v} _ {\\text {m a s k}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.725, + 0.475, + 0.79 + ], + "angle": 0, + "content": "Here, \\(1 \\leq k \\leq N'\\), and \\(\\mathbf{v}_{\\mathrm{mask}}\\) represents the learnable mask embedding [4]. The masked tokenized inputs are thus represented as \\(\\tilde{T} = [\\mathbf{t}_{\\mathrm{cls}}; \\tilde{\\mathbf{t}}_1; \\ldots; \\tilde{\\mathbf{t}}_N]\\) for text and \\(\\tilde{V} = [\\mathbf{v}_{\\mathrm{cls}}; \\tilde{\\mathbf{v}}_1; \\ldots; \\tilde{\\mathbf{v}}_{N'}]\\) for image. These masked inputs are processed using the proposed SyncMask \\(\\mathbf{m}^T\\) and \\(\\mathbf{m}^I\\)." + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.806, + 0.475, + 0.845 + ], + "angle": 0, + "content": "Synchronized Attentional Masked Language Modeling MLM predicts masked words based on the surrounding contextual text and image. Many existing VLMs applied the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.149, + 0.837, + 0.278 + ], + "angle": 0, + "content": "MLM method proposed in BERT [7], randomly masking words with a probability of \\(15\\%\\). However, this approach may not be suitable for vision-language datasets with short caption lengths, especially for nonstandard datasets such as fashion, demanding a thorough understanding of fine-grained attributes. We leverage previous works that addressed these issues by increasing masking probabilities [3] and employing attribute prompts [15]. Building upon these methods, we employ masks generated by SyncMask which is contextually attuned to the corresponding image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.837, + 0.331 + ], + "angle": 0, + "content": "Let \\( h(V, \\tilde{T}) \\) denote the model's predicted probability for a masked token. \\( \\tilde{y} \\) denote a one-hot vocabulary distribution in which the ground-truth token is assigned a probability of 1. MLM minimize cross-entropy loss, described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.337, + 0.836, + 0.354 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M L M}} = \\mathbb {E} _ {(V, \\hat {T}) \\sim D} [ \\operatorname {C E} (\\tilde {\\boldsymbol {y}}, \\boldsymbol {h} (V, \\hat {T}) ], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.837, + 0.372 + ], + "angle": 0, + "content": "where \\(\\mathrm{CE}(\\cdot ,\\cdot)\\) refers the cross-entropy between two vectors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.839, + 0.492 + ], + "angle": 0, + "content": "Synchronized Attentional Masked Image Modeling In the distillation-based MIM [21, 29, 34], a teacher encoder sees the full image, whereas the student encoder, seeing the masked image, tackles the reconstruction objective. Our method adopts a similar objective framework, but with a key difference: our masks, generated through SyncMask, are designed to reflect text-informed elements in masked image. To calculate the MIM loss, the following approach is used:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.505, + 0.837, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {D i s t} \\left(\\boldsymbol {f} _ {\\theta^ {\\prime}} ^ {I} (V), \\boldsymbol {f} _ {\\theta} ^ {I} (\\tilde {V})\\right) \\\\ = \\frac {1}{\\Omega (\\mathbf {m} ^ {I})} \\sum_ {k = 1} ^ {N ^ {\\prime}} \\mathbf {m} _ {k} ^ {I} \\cdot \\ell_ {1} ^ {S m o o t h} \\left(\\boldsymbol {f} _ {\\theta^ {\\prime}} ^ {I} (V) _ {k}, \\boldsymbol {f} _ {\\theta} ^ {I} (\\tilde {V}) _ {k}\\right) \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.839, + 0.622 + ], + "angle": 0, + "content": "where \\( \\pmb{f}_{\\theta^{\\prime}}^{I}(\\cdot)_{k} \\) and \\( \\pmb{f}_{\\theta}^{I}(\\cdot)_{k} \\) represent the output feature of teacher and student model for the \\( k \\)-th image patch, respectively. \\( \\Omega (\\cdot) \\) means the number of elements with a value of 1 in a vector." + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.626, + 0.837, + 0.659 + ], + "angle": 0, + "content": "\\[\n\\ell_ {1} ^ {\\text {S m o o t h}} (a, b) = \\left\\{ \\begin{array}{l l} 0. 5 (a - b) ^ {2} & \\text {i f} | a - b | < \\gamma \\\\ | a - b | - 0. 5 & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.838, + 0.721 + ], + "angle": 0, + "content": "where \\(\\ell_1^{Smooth}\\) [11] represents a robust L1 loss less sensitive to outliers than the L2 loss and \\(\\gamma\\) is a hyperparameter set to 1. Conclusively, the training objective of MIM can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.733, + 0.837, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M I M}} = \\mathbb {E} _ {(V, \\hat {V}) \\sim D} [ \\operatorname {D i s t} \\left(\\boldsymbol {f} _ {\\theta^ {\\prime}} ^ {I} (V), \\boldsymbol {f} _ {\\theta} ^ {I} (\\tilde {V})\\right) ] \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.758, + 0.668, + 0.772 + ], + "angle": 0, + "content": "The final loss \\((\\mathcal{L})\\) is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.782, + 0.837, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M I M}} + \\mathcal {L} _ {\\mathrm {M L M}} + \\mathcal {L} _ {\\mathrm {I T C}} + \\mathcal {L} _ {\\mathrm {I T M}} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.806, + 0.838, + 0.845 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{ITC}}\\) and \\(\\mathcal{L}_{\\mathrm{ITM}}\\) denote ITC and ITM, respectively. Due to space constraints, detailed formulations of these two losses are elaborated in the Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.881, + 0.518, + 0.892 + ], + "angle": 0, + "content": "13952" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.206, + 0.146, + 0.768, + 0.365 + ], + "angle": 0, + "content": "
METHODSI2T*T2I*MEAN*I2TT2IMEAN
R@1R@1R@1R@1
VSE++ [9]4.594.604.60---
VL-BERT [38]19.2622.6320.95---
ViLBERT [32]20.9721.1221.05---
Image-BERT [35]22.7624.7823.77---
OSCAR [28]23.3925.1024.25---
FashionBERT [10]23.9626.7525.36---
Kaleido-BERT [45]27.9933.8830.94---
CommerceMM [43]41.6039.6062.75---
EI-CLIP [33]38.7040.0639.3825.7028.4027.05
ALBEF [27]63.9760.5262.2041.6850.9546.32
FashionViL [13]65.5461.8863.7142.8851.3447.11
FashionSAP [15]73.1470.1271.6354.4362.8258.63
Ours75.0071.0073.0055.3964.0659.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.206, + 0.369, + 0.768, + 0.382 + ], + "angle": 0, + "content": "Table 2. Cross-modal retrieval result on FashionGen [37] in the sub/full set of evaluation following previous work. \\(\\star\\) sub set." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.398, + 0.444, + 0.412 + ], + "angle": 0, + "content": "3.3. Grouped Batch with Semi-hard Negatives" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.419, + 0.477, + 0.575 + ], + "angle": 0, + "content": "In the generic domain, [3] proposed the GRIT strategy for enhancing training effectiveness by forming mini-batches with similar examples. In this strategy, the grouping based on similarity phase plays a crucial role. During this phase, similarity calculations are performed in both directions, utilizing the [CLS] outputs from the unimodal encoders. For each example, the algorithm iteratively identifies the index with the highest similarity, alternating between the image-to-text and text-to-image directions in a sequential manner. These highly similar indices are grouped together within the mini-batch, ensuring that each mini-batch consists of examples exhibiting the highest possible similarity." + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.576, + 0.475, + 0.745 + ], + "angle": 0, + "content": "However, in fashion datasets, this strategy leads to a false-negative problem in ITM and ITC, where true positives are mislabeled as negatives within mini-batches. To address this problem, we opt to group semi-hard negatives with relatively lower similarity (the \\( s^{th} \\) highest pairwise similarity) instead of the highest \\( (1^{st}) \\) during the grouping based on similarity phase, where \\( s \\) represents a predefined hyperparameter that is greater than 1. In addition, we prevent true positive samples from grouping by considering the item indices of the samples. Through this approach, the model is trained with negatives that are similar but exhibit meaningful differences, thereby enabling the effective learning of fine-grained distinctions with a limited dataset. More details are in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.761, + 0.251, + 0.777 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.784, + 0.323, + 0.798 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.806, + 0.475, + 0.845 + ], + "angle": 0, + "content": "The foundational architecture of the proposed model is aligned with prior works for demonstrating the effectiveness of the proposed approach [27, 3, 15]. The image encoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.841, + 0.58 + ], + "angle": 0, + "content": "adopt the architecture of ViT-B16 [8], whereas the text encoder comprises the first six blocks of the BERT-bas3 [7]. The multi-modal encoder extends the self-attention layers of the last six blocks of BERT with the cross-attention layers. The proposed model is initialized with pre-trained ALBEF [27] same as the Fashion-SAP to ensure a fair comparison [15]. In addition, we employ the same data augmentation and prompt input strategies as FashionSAP [15]. During pretraining, we conduct experiments using 8 RTX 3090 GPUs each with a batch size of 8 for 30 epochs. We adopt a momentum queue size of 48,000 to facilitate grouped batch sampling, which is consistent with GRIT-VLP [3]. The input image size is set to \\(256 \\times 256\\). We apply the AdamW [31] optimizer with a learning rate of 6e-5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.589, + 0.589, + 0.602 + ], + "angle": 0, + "content": "4.2. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.839, + 0.715 + ], + "angle": 0, + "content": "FashionGen [37] FashionGen comprises 320K text-image pairs and 40K unique fashion items, each represented by multiple images from different angles. For pre-training, we employ the FashionGen train set, which contains approximately 260.5K text-image pairs. In addition, FashionGen supports various downstream tasks, including text-to-image retrieval, image-to-text retrieval, category recognition, and subcategory recognition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.723, + 0.839, + 0.789 + ], + "angle": 0, + "content": "FashionIQ [40] FashionIQ encompasses 77K unique fashion items and includes 18K training triplets (i.e., query image, modified text, target image) and 6K validation datasets for a text-guided image retrieval task. It contains three different categories: Dress, Toptee, and Shirt." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.798, + 0.657, + 0.81 + ], + "angle": 0, + "content": "4.3. Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.818, + 0.839, + 0.845 + ], + "angle": 0, + "content": "Cross-modal Retrieval We evaluate a cross-modal retrieval task that includes image-text retrieval (ITR) and text" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.88, + 0.518, + 0.892 + ], + "angle": 0, + "content": "13953" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.191, + 0.148, + 0.235, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.163, + 0.301, + 0.187 + ], + "angle": 0, + "content": "is light blue in color and has more blue" + }, + { + "type": "image", + "bbox": [ + 0.201, + 0.205, + 0.23, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.21, + 0.313, + 0.244 + ], + "angle": 0, + "content": "is yellow and has short sleeves and is beige and shorter sleeves" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.257, + 0.231, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.266, + 0.3, + 0.289 + ], + "angle": 0, + "content": "is light blue in color and has more blue" + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.149, + 0.373, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.15, + 0.411, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.151, + 0.464, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.156, + 0.503, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.156, + 0.55, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.15, + 0.595, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.152, + 0.637, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.15, + 0.692, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.151, + 0.728, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.152, + 0.776, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.203, + 0.365, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.204, + 0.417, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.204, + 0.455, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.204, + 0.501, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.206, + 0.551, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.205, + 0.596, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.204, + 0.638, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.203, + 0.689, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.203, + 0.73, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.204, + 0.777, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.255, + 0.373, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.256, + 0.411, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.255, + 0.464, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.261, + 0.502, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.254, + 0.556, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.261, + 0.6, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.255, + 0.647, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.254, + 0.695, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.255, + 0.735, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.256, + 0.785, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.313, + 0.838, + 0.345 + ], + "angle": 0, + "content": "Figure 5. The top-10 TGIR results of the SyncMask model on the FashionIQ dataset. On the left, the reference images paired with their guided descriptions are shown, while the right side presents the model's predicted images ranked by descending scores. Ground truth images are distinctly outlined with a green bounding box. It is worth mentioning that the set of predictions includes other images that also qualify as suitable matches." + }, + { + "type": "table", + "bbox": [ + 0.165, + 0.356, + 0.807, + 0.51 + ], + "angle": 0, + "content": "
METHODSDRESSTOPTEESHIRTMEAN
R@10R@50R@10R@50R@10R@50R@10R@50
CIRR [30]17.4540.4121.6445.3817.5338.8118.8741.53
VAL [5]22.5344.0027.5351.6822.3844.1524.1546.61
CosMo [25]25.6450.3029.2157.4624.9049.1826.5852.31
DCNet [22]28.9556.730.4458.2923.9547.327.7854.10
FashionVLP [12]32.4260.2938.5168.7931.8958.4434.2762.51
FashionViL [13]33.4759.9434.9860.7925.1750.3931.2157.04
FashionSAP [15]33.7160.4341.9170.9333.1761.3336.2664.23
Ours33.7661.2344.8272.0635.8262.1238.1365.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.332, + 0.512, + 0.642, + 0.524 + ], + "angle": 0, + "content": "Table 3. Text-guided image retrieval performance in FashionIQ [40]" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.541, + 0.477, + 0.684 + ], + "angle": 0, + "content": "image retrieval (TIR). Further, ITR focuses on finding relevant textual descriptions for a given image query. TIR, the inverse task, retrieves pertinent images based on a textual query. These tasks assess the effectiveness of the model in capturing cross-modal relationships between text and images within retrieval scenarios. Following the previous works [13, 15], we evaluate cross-modal retrieval not only on the subset with 1K retrievals but also on the full dataset of FashionGen [37]. The results, including R@1 scores for both subset and the full set, are presented in Table Table 2, demonstrating an improved performance compared to that of the previous results." + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.702, + 0.476, + 0.845 + ], + "angle": 0, + "content": "Text-guided Image Retrieval This task aims to retrieve target images by considering query pairs that reference image and modified descriptions; this is more challenging than traditional retrievals. Therefore, we need to select the target image for identifying minor differences in the changes in description while maintaining the characteristics of the reference image. For a fair comparison, we adopt a similar fine-tuning as outlined in the FashionSAP [15]. In the actual dataset, there are many images that match with the query pairs in addition to the target image referred to as the real correct answer, as shown in Figure 5. Thus, a qualitative" + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.538, + 0.81, + 0.624 + ], + "angle": 0, + "content": "
METHODSCRSCR
AccMacro-FAccMacro-F
F-BERT [10]91.2570.5085.2762.00
K-BERT [45]95.0771.4088.0763.60
F-ViL [13]97.4888.6092.2383.02
FashionSAP [15]98.3489.8494.3387.67
Ours98.4190.3194.2187.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.554, + 0.625, + 0.783, + 0.636 + ], + "angle": 0, + "content": "Table 4. CR and SCR results on FashionGen [37]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.663, + 0.838, + 0.715 + ], + "angle": 0, + "content": "evaluation can be conducted in that the model finds similar images well in addition to the actual correct answer. As shown in Table 3, the proposed model surpasses previous models and demonstrates state-of-the-art performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.741, + 0.838, + 0.845 + ], + "angle": 0, + "content": "Category/Subcategory Recognition In this downstream task, the objective is to classify the category and subcategory of fashion items using the textual and visual information provided. In line with earlier studies [10, 45, 13, 15], we simply attach a linear layer to the [CLS] token, which serves as the fusion feature, for task label prediction. As indicated in Table 4, our proposed model exhibits competitive performance compared to existing models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.881, + 0.518, + 0.892 + ], + "angle": 0, + "content": "13954" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.146, + 0.462, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.271, + 0.462, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.4, + 0.232, + 0.409 + ], + "angle": 0, + "content": "(a) Unmasked input" + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.4, + 0.341, + 0.409 + ], + "angle": 0, + "content": "(b) Random mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.4, + 0.436, + 0.409 + ], + "angle": 0, + "content": "(c) SyncMask" + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.41, + 0.474, + 0.441 + ], + "angle": 0, + "content": "Figure 6. Examples of random masking (b) and SyncMask (c) for MIM and MLM. The latter applies masks to pertinent features in both the input image and text (a), offering a more context-sensitive selection than the former." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.465, + 0.269, + 0.48 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.486, + 0.477, + 0.747 + ], + "angle": 0, + "content": "Random vs. Attentional Masking We ablate the SyncMask with other non-synchronized masking methods. Table 5 lists the comparison results for the combination of random and attentional masking. In this experiment, we evaluated five experiments for I2T (R@1), T2I (R@1), CR (Macro-F), SCR (Macro-F), and TMIR (R@10), which were evaluated in the previous experiments. In these experiments, the results of the three experiments with at least one attentional masking were higher than those with no attentional masking (all random). This can be considered an indication of the efficacy of attentional masking. Further, the synergistic performance of the proposed synchronized textual-visual attentional masking (+2.14) is greater than that of only one attentional masking (+0.4, +0.66), which shows the superiority of the SyncMask approach. Figure 6 displays the distinction between applying random masks (b) and SyncMask (c) to unmasked image-text pairs (a) from the FashionGen [37]. SyncMask strategically places masks on image patches that correlate with the text and applies masks to the text informed by the image details, unlike random masking." + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.766, + 0.474, + 0.845 + ], + "angle": 0, + "content": "Random vs. Hardest vs. Semi-hard Negative Sampling Table 6 compare the effectiveness of various grouped minibatch sampling strategies across five downstream tasks. The methods evaluated encompassed four scenarios: random grouping, hardest grouping (+1.90), hardest grouping with exclusion of false negatives (+1.98), and semi-hard grouping" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.147, + 0.836, + 0.212 + ], + "angle": 0, + "content": "
RTRVATAVMEANGAINI2T R@1T21 R@1CR Macro-FSCR Macro-FTMIR R@10
✓✓70.3173.3069.8086.2186.1636.08
✓✓70.71+0.4074.8070.1086.4986.3135.84
✓✓70.97+0.6674.3070.8085.1286.7837.87
✓✓72.45+2.1475.0071.0090.3187.8338.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.214, + 0.837, + 0.247 + ], + "angle": 0, + "content": "Table 5. Ablation study results for Random vs. Attention Masked Modeling on five downstream tasks. Rt:Random Textual masking. Rv:Random Visual masking. At:Attentional Textual masking. Av:Attentional Visual masking." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.258, + 0.835, + 0.322 + ], + "angle": 0, + "content": "
GROUPEFNMEANGAIN12T R@1T2I R@1CR Macro-FSCR Macro-FTMIR R@10
Random69.9972.3069.0086.2185.4037.06
Hardest71.89+1.9074.7070.5090.2086.6437.40
Hardest✓71.97+1.9874.3071.0090.1087.1537.28
Semi-hard✓72.45+2.4675.0071.0090.3187.8338.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.325, + 0.838, + 0.358 + ], + "angle": 0, + "content": "Table 6. Ablation study results comparing the Grouped Batch Sampling Strategy across five downstream tasks. GROUP: Strategy for grouping phase of GRIT. EFN: Exclude False Negative in a grouping phase using index." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.839, + 0.548 + ], + "angle": 0, + "content": "also excluding false negatives (+2.46). These findings suggest that grouping similar samples in a mini-batch is more beneficial for learning than composing batches with random samples. However, given that the fashion dataset often has multiple captions per image or vice versa, performance gains were observed when systematically preventing the grouping of false negatives by using indexing. Nevertheless, due to the prevalence of inherently similar samples that cannot be systematically excluded, we opted for grouping semi-hard negatives instead of the hardest ones, which resulted in a significant performance boost. This highlights the importance of further research from a data perspective, not just in terms of model architecture or loss functions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.603, + 0.573 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.839, + 0.751 + ], + "angle": 0, + "content": "We introduced Synchronized attentional Masking for enhanced masked modeling in fashion-centric VLMs. Leveraging cross-attention features of a momentum model, our method tailors the random mask into a targeted mask for synchronously co-occurring segments in image-text pairs in MLM and MIM objectives. This approach effectively resolves misaligned image-text input issues and improving fine-grained cross-modal representation. Additionally, we addressed data scarcity and distribution challenges in fashion datasets, refining grouped batch sampling with semi-hard negatives for ITM and ITC losses. The experimental results showed our methods outperformed established benchmarks in multiple downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.762, + 0.631, + 0.777 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.838, + 0.845 + ], + "angle": 0, + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) under the metaverse support program to nurture the best talents (IITP-2024-RS-2023-00254529) grant funded by the Korea government (MSIT)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.881, + 0.518, + 0.892 + ], + "angle": 0, + "content": "13955" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.138, + 0.147, + 0.22, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.145, + 0.169, + 0.474, + 0.228 + ], + "angle": 0, + "content": "[1] Yutong Bai, Zeyu Wang, Junfei Xiao, Chen Wei, Huiyu Wang, Alan L Yuille, Yuyin Zhou, and Cihang Xie. Masked autoencoders enable efficient knowledge distillers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24256-24265, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.145, + 0.23, + 0.475, + 0.253 + ], + "angle": 0, + "content": "[2] Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT pretraining of image transformers. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.145, + 0.255, + 0.473, + 0.302 + ], + "angle": 0, + "content": "[3] Jaeseok Byun, Taebaek Hwang, Jianlong Fu, and Taesup Moon. Grit-vlp: Grouped mini-batch sampling for efficient vision and language pre-training. In European Conference on Computer Vision, pages 395-412. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.146, + 0.303, + 0.475, + 0.349 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.145, + 0.351, + 0.474, + 0.398 + ], + "angle": 0, + "content": "[5] Yanbei Chen, Shaogang Gong, and Loris Bazzani. Image search with text feedback by visiolinguistic attention learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3001-3011, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.146, + 0.4, + 0.474, + 0.457 + ], + "angle": 0, + "content": "[6] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.146, + 0.46, + 0.474, + 0.541 + ], + "angle": 0, + "content": "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.146, + 0.544, + 0.474, + 0.613 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.146, + 0.615, + 0.473, + 0.651 + ], + "angle": 0, + "content": "[9] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.14, + 0.652, + 0.473, + 0.722 + ], + "angle": 0, + "content": "[10] Dehong Gao, Linbo Jin, Ben Chen, Minghui Qiu, Peng Li, Yi Wei, Yi Hu, and Hao Wang. Fashionbert: Text and image matching with adaptive loss for cross-modal retrieval. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2251-2260, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.139, + 0.724, + 0.395, + 0.735 + ], + "angle": 0, + "content": "[11] Ross Girshick. Fast R-CNN. In ICCV, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.139, + 0.737, + 0.474, + 0.807 + ], + "angle": 0, + "content": "[12] Sonam Goenka, Zhaoheng Zheng, Ayush Jaiswal, Rakesh Chada, Yue Wu, Varsha Hedau, and Pradeep Natarajan. Fashionvlp: Vision language transformer for fashion retrieval with feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14105-14115, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.139, + 0.809, + 0.475, + 0.844 + ], + "angle": 0, + "content": "[13] Xiao Han, Licheng Yu, Xiatian Zhu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Fashionvil: Fashion-focused vision-and-language representation learning. In ECCV. Springer, 2022." + }, + { + "type": "list", + "bbox": [ + 0.139, + 0.169, + 0.475, + 0.844 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.838, + 0.209 + ], + "angle": 0, + "content": "[14] Xiao Han, Xiatian Zhu, Licheng Yu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Fame-vil: Multi-tasking vision-language model for heterogeneous fashion tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2669-2680, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.211, + 0.838, + 0.259 + ], + "angle": 0, + "content": "[15] Yunpeng Han, Lisai Zhang, Qingcai Chen, Zhijian Chen, Zhonghua Li, Jianxin Yang, and Zhao Cao. Fashionsap: Symbols and attributes prompt for fine-grained fashion vision-language pre-training. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.837, + 0.295 + ], + "angle": 0, + "content": "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In arXiv:2111.06377, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.297, + 0.838, + 0.354 + ], + "angle": 0, + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.838, + 0.417 + ], + "angle": 0, + "content": "[18] Zhicheng Huang, Zhaoyang Zeng, Yupan Huang, Bei Liu, Dongmei Fu, and Jianlong Fu. Seeing out of the box: End-to-end pre-training for vision-language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12976-12985, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.837, + 0.453 + ], + "angle": 0, + "content": "[19] Zhicheng Huang, Zhaoyang Zeng, Bei Liu, Dongmei Fu, and Jianlong Fu. Pixel-bert: Aligning image pixels with text by deep multi-modal transformers. 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.838, + 0.525 + ], + "angle": 0, + "content": "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.528, + 0.838, + 0.575 + ], + "angle": 0, + "content": "[21] Ioannis Kakogeorgiou, Spyros Gidaris, Bill Psomas, Yannis Avrithis, Andrei Bursuc, Konstantinos Karantzalos, and Nikos Komodakis. What to hide from your students: Attention-guided masked image modeling. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.838, + 0.612 + ], + "angle": 0, + "content": "[22] Jongseok Kim, Youngjae Yu, Hoeseong Kim, and Gunhee Kim. Dual compositional learning in interactive image retrieval. In AAAI, pages 1771-1779, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.614, + 0.838, + 0.66 + ], + "angle": 0, + "content": "[23] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.662, + 0.838, + 0.709 + ], + "angle": 0, + "content": "[24] Gukyeong Kwon, Zhaowei Cai, Avinash Ravichandran, Erhan Bas, Rahul Bhotika, and Stefano Soatto. Masked vision and language modeling for multi-modal representation learning. In ICLR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.712, + 0.838, + 0.759 + ], + "angle": 0, + "content": "[25] Seungmin Lee, Dongwan Kim, and Bohyung Han. Cosmo: Content-style modulation for image retrieval with text feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 802-812, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.838, + 0.806 + ], + "angle": 0, + "content": "[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.808, + 0.838, + 0.844 + ], + "angle": 0, + "content": "[27] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.15, + 0.838, + 0.844 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.881, + 0.518, + 0.892 + ], + "angle": 0, + "content": "13956" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.165, + 0.15, + 0.475, + 0.173 + ], + "angle": 0, + "content": "momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.139, + 0.175, + 0.474, + 0.234 + ], + "angle": 0, + "content": "[28] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.138, + 0.236, + 0.474, + 0.283 + ], + "angle": 0, + "content": "[29] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, et al. MST: Masked self-supervised transformer for visual representation. In NeurIPS, volume 34, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.139, + 0.285, + 0.475, + 0.343 + ], + "angle": 0, + "content": "[30] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2125-2134, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.139, + 0.346, + 0.473, + 0.369 + ], + "angle": 0, + "content": "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.139, + 0.371, + 0.474, + 0.417 + ], + "angle": 0, + "content": "[32] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.139, + 0.42, + 0.475, + 0.49 + ], + "angle": 0, + "content": "[33] Haoyu Ma, Handong Zhao, Zhe Lin, Ajinkya Kale, Zhangyang Wang, Tong Yu, Jiumiang Gu, Sunav Choudhary, and Xiaohui Xie. Ei-clip: Entity-aware interventional contrastive learning for e-commerce cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18051-18061, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.139, + 0.492, + 0.473, + 0.527 + ], + "angle": 0, + "content": "[34] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. A unified view of masked image modeling. In arXiv preprint arXiv:2210.10615, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.139, + 0.529, + 0.473, + 0.575 + ], + "angle": 0, + "content": "[35] Di Qi, Lin Su, Jia Song, Edward Cui, Taroon Bharti, and Arun Sacheti. Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.139, + 0.578, + 0.474, + 0.648 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.139, + 0.651, + 0.474, + 0.697 + ], + "angle": 0, + "content": "[37] Negar Rostamzadeh, Seyedarian Hosseini, Thomas Boquet, Wojciech Stokowiec, Ying Zhang, Christian Jauvin, and Chris Pal. Fashion-gen: The generative fashion dataset and challenge. arXiv preprint arXiv:1806.08317, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.139, + 0.7, + 0.474, + 0.746 + ], + "angle": 0, + "content": "[38] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. In International Conference on Learning Representations, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.139, + 0.748, + 0.473, + 0.771 + ], + "angle": 0, + "content": "[39] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.139, + 0.773, + 0.475, + 0.843 + ], + "angle": 0, + "content": "[40] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11307-11317, 2021." + }, + { + "type": "list", + "bbox": [ + 0.138, + 0.15, + 0.475, + 0.843 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.838, + 0.209 + ], + "angle": 0, + "content": "[41] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.212, + 0.838, + 0.258 + ], + "angle": 0, + "content": "[42] Hongwei Xue, Yupan Huang, Bei Liu, Houwen Peng, Jianlong Fu, Houqiang Li, and Jiebo Luo. Probing inter-modality: Visual parsing with self-attention for vision-and-language pre-training. volume 34, pages 4514-4528, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.26, + 0.838, + 0.328 + ], + "angle": 0, + "content": "[43] Licheng Yu, Jun Chen, Animesh Sinha, Mengjiao Wang, Yu Chen, Tamara L Berg, and Ning Zhang. Commercemm: Large-scale commerce multimodal representation learning with omni retrieval. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4433-4442, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.332, + 0.838, + 0.403 + ], + "angle": 0, + "content": "[44] Zijia Zhao, Longteng Guo, Xingjian He, Shuai Shao, Zehuan Yuan, and Jing Liu. Mamo: Fine-grained vision-language representations learning with masked multimodal modeling. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1528-1538, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.838, + 0.463 + ], + "angle": 0, + "content": "[45] Mingchen Zhuge, Dehong Gao, Deng-Ping Fan, Linbo Jin, Ben Chen, Haoming Zhou, Minghui Qiu, and Ling Shao. Kaleido-bert: Vision-language pre-training on fashion domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12647–12657, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.15, + 0.838, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.881, + 0.517, + 0.891 + ], + "angle": 0, + "content": "13957" + } + ] +] \ No newline at end of file diff --git a/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/92bc41a7-a993-4d23-a197-14e6dd1e0509_origin.pdf b/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/92bc41a7-a993-4d23-a197-14e6dd1e0509_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2b6940b1fe1b8cefd1dc8789e69e262676788054 --- /dev/null +++ b/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/92bc41a7-a993-4d23-a197-14e6dd1e0509_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:378f3ae2a7b05ab0aefe3a55e0ed420264493d880b587107f34dc6f48daa21d8 +size 1093308 diff --git a/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/full.md b/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eba8940291ca403cbebf7af5fc71555b5e7b98bf --- /dev/null +++ b/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/full.md @@ -0,0 +1,384 @@ +# SyncMask: Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining + +Chull Hwan Song $^{1*}$ Taebaek Hwang $^{1*}$ Jooyoung Yoon $^{1}$ Shunghyun Choi $^{1}$ Yeong Hyeon $\mathrm{Gu}^{2\dagger}$ + +$^{1}$ Dealicious Inc. $^{2}$ Sejong University + +# Abstract + +Vision-language models (VLMs) have made significant strides in cross-modal understanding through large-scale paired datasets. However, in fashion domain, datasets often exhibit a disparity between the information conveyed in image and text. This issue stems from datasets containing multiple images of a single fashion item all paired with one text, leading to cases where some textual details are not visible in individual images. This mismatch, particularly when non-co-occurring elements are masked, undermines the training of conventional VLM objectives like Masked Language Modeling and Masked Image Modeling, thereby hindering the model's ability to accurately align fine-grained visual and textual features. Addressing this problem, we propose Synchronized attentional Masking (SyncMask), which generate masks that pinpoint the image patches and word tokens where the information co-occur in both image and text. This synchronization is accomplished by harnessing cross-attentional features obtained from a momentum model, ensuring a precise alignment between the two modalities. Additionally, we enhance grouped batch sampling with semihard negatives, effectively mitigating false negative issues in Image-Text Matching and Image-Text Contrastive learning objectives within fashion datasets. Our experiments demonstrate the effectiveness of the proposed approach, outperforming existing methods in three downstream tasks. + +# 1. Introduction + +Recently, there has been rapid progress in developing Vision-Language Pretraining (VLP) [32, 38, 6, 39, 19, 18, 42, 23, 36, 20, 27, 3], paving the way to bridge the gap between visual and textual features. These VLP methods, trained on extensive image-text datasets, have enabled a deeper understanding of semantic alignment across different modalities. By fine-tuning these pretrained models for specific tasks, particularly in data-scarce areas like image- + +![](images/6be5b5d0e6cd21fae3e8adc905e60aefed227250a53ee29eae28a238a1fcc05a.jpg) +[images, caption] +Notched lapel collar. Padded shoulders.... +Four-button cuffs. Two vents at back... +Figure 1. Example of misaligned masks in the MLM task. + +![](images/596396104a1e803272469db78a90aba6efd435297d72c420f22fa425f3907645.jpg) +Dataset +Long sleeve blazer in deep navy. +(a) + +![](images/4d929c6dfd94b47404609a70da777ed04be62b257c307440ec8bfa4276f7d8b9.jpg) +Input of MLM +[image, masked caption] +Long sleeve blazer [MASK] deep navy. +Notched lapel collar. [MASK] shoulders... +Four-button [MASK]. Two veins at [MASK]. +(b) + +text retrieval, notable performance improvements have been observed. In the pretraining phase, various factors such as model architecture, training objectives, and batch sampling techniques play a crucial role in effectively harnessing the joint representation of multi-modal data. + +However, there are some issues in applying conventional generic VLP approaches to task-specific domains such as fashion. Fashion VLP models [10, 45, 13, 14, 15] typically employ objectives such as Masked Language Modeling (MLM) and Masked Image Modeling (MIM). These methods mask elements like text words or image patches, leveraging surrounding context for prediction or reconstruction. They boost cross-modality by making models infer masked text tokens or image patches from aligned features. However, existing MLM and MIM often suffer from inherent misalignment limitations because the masks are generated randomly, often leading to unmatched elements being masked. + +To illustrate these limitations, consider Figure 1 (a) shows a single description associated with four images from the FashionGen [37] dataset. In Figure 1 (b), a random masking scenario is shown where the blue [MASK] might lead the model to predict the masked word using only the text context, thereby not incorporating the image information. Similarly, the two red [MASK] tokens in (b) lack relevance to the accompanying image, thus hindering the model's ability to connecting between visual and textual features. Fur- + +thermore, in the MIM task, random masking may inadvertently cover parts of the image that contain fashion items not described in the text, leading to mismatches during the training of the alignment of cross-modal features. To solve these problem, we propose SyncMask, selecting masks that represent synchronously co-occurring features utilizing the vision-language cross-attention map. + +In addition, compared to general domains, the fashion domain often suffers from smaller dataset sizes and less variance in data distribution. This suggests that using standard VLP methods may not adequately distinguish the fine-grained features vital for fashion-related tasks. Thus, we pay attention to a grouped batch sampling technique [3] that similar samples are gradually collected within the batch as the training progresses, impacting the pretraining objectives of Image-Text Contrastive Learning (ITC) and Image-Text Matching (ITM). When similar samples exist within the same batch, it becomes more challenging to differentiate positives and negatives during the training of ITC and ITM compared to using random samples. This encourages the model to more intensely focus on learning fine-grained differences, even with less training. + +The existing grouped batch sampling method uses the output features of two uni-modal encoders to find the most similar sample. However, as shown in Figure 1, there are many data that have the same caption for multiple images as for the fashion domain. Therefore, a false negative problem that causes actual positive samples in the same batch to be wrongly labelled as negative when learning ITC and ITM arises if the existing methods are used without changes. To overcome this limitation, we propose a semi-hard negative sampling technique with lower similarities between samples that constitute the batch while removing the false negative. + +In summary, the main contributions of this study are: + +1. Synchronized Attentional Masking: We introduce SyncMask, which replaces random mask in MLM and MIM with targeted mask of co-occurring segments in image-text pairs. By utilizing cross-attention features from a momentum model to generate these masks, this method effectively addresses the problem of mismatched image-text inputs, thereby enhancing fine-grained alignment of cross-modal features. +2. Refined Grouped Batch Sampling: Our method incorporates semi-hard negative sampling to tackle data scarcity and distribution disparities in domain-specific datasets, thereby reducing false negatives. + +# 2. Related Works + +Vision and Language (VL) Model Recently, VLMs have focused on enhancing model architecture and designing objectives to integrate visual and textual features effectively. Early studies [32, 38, 39, 6] have used object detectors for extracting visual features as an input for a multi-modal trans + +
METHODMMMLMMIMATMAVMONUP
DMAE [1]✓
MaskDistill [34]✓✓✓
AttMask [21]✓✓✓
ALBEF [27]✓✓
MaskVLM [24]✓✓✓
MAMO [44]✓✓✓
FashionBERT [10]✓✓✓
Kaleido-BERT [45]✓✓✓✓✓
FashionViL [13]✓✓✓
FashionSAP [15]✓✓
Ours✓✓✓✓✓✓
+ +Table 1. Related works vs. Ours on Masked Modeling. MM:MultiModal. MLM:Masked Language Modeling. MIM:Masked Image Modeling. ATM:Attentional Textual Mask. AVM:Attentional Visual Mask. OnUp:Online Update for Attentional Masking + +former along with textual features. The objective for training models extends vanilla BERT [7] to use MLM, MIM, and ITM losses; however, the object detection module incurs a high computational cost for training and inference. Therefore, there have been attempts to replace it with CNN [19, 18] or linear projection [23]. These studies commonly train models with MLM and ITM, tailoring MIM to their specific architectures. Concurrently, CLIP [36], ALIGN [20] propose models comprising only two unimodal encoders, demonstrating the outstanding representation embedding capabilities of contrastive learning. ALBEF [27] add a contrastive learning objective to the previous multi-modal transformer structure for aligning the two modalities before fusion. Based on this model, GRIT-VLP [3] demonstrate improved learning efficiency when configuring batches with hard negative samples. + +FashionVL Model In recent years, various studies [10, 45, 13, 14, 15] attempted to capture the finer details of images, building upon established models and pre-training objectives from the general VL task. FashionBERT [10] integrates patch-based image features and BERT-based text representations for addressing the limitations of region of interests (RoIs) in capturing fine-grained details. Kaleido-BERT [45] improves fine-grained fashion cross-modality representations through alignment guided masking compared to random masking. FashionViL [13] employs a versatile VLP framework, leveraging two pre-training tasks for capturing the rich fine-grained information of fashion data. FashionSAP [15] employs abstract fashion symbols and an attributes prompt technique for effectively modeling multi-modal fashion attributes. We clarify the importance of the attentional masking technique that employs alignment between MIM and MLM building upon preceding methods. In addition, we underscore a previously unaddressed need for grouped batch sampling within the fashion domain. + +![](images/9ca3eb8e0a9ba93e86ada5bc6bee0afaa51d79d29825dc4d30198641bcc90e07.jpg) +Figure 2. Overview of masking strategies using a teacher-student distillation framework. 1) Uni-modal models: (a) random masking, (b) teacher-guided attentional masking. 2) Multi-modal models: (c) random text masking, (d) random image/text masking, (e) teacher-guided cross-attentional masking (Ours). + +Attention-guided Masked Modeling MIM and MLM leverage unmasked contextual clues to predict masked visual and textual elements, respectively. As shown in the upper section of Table 1 and Figure 2 (a, b), in the MIM task, prior studies have evolved from random masking [16, 2, 41, 1] to attention-guided masking [29, 34, 21], demonstrating that targeting highly-attended patches with teacher-student distillation framework improves masked modeling outcomes. Masked modeling also has been pivotal in advancing cross-modal alignment within VLMs, spanning both general [38, 32, 27, 24, 44] and fashion-specific [10, 45, 13, 15] domains. This is briefly illustrated in Figure 2 (c, d) and delineated in the middle and lower sections of Table 1. These enhance the model's proficiency aligning co-occurring visual and textual representations. However, challenges emerge when masking non-co-occurring elements, which hinders the accurate pairing of visual and textual features. KaleidoBERT [45] improve fine-grained cross-modality representations through text-image alignment-guided masking. This requires additional components for an attention-based alignment generator, increasing computational demands and potential model overfitting on fixed text-image mask pairs. Our approach overcomes these limitations by progressively tailoring masks during end-to-end training. + +# 3. Methods + +We provide an overview of the preliminary aspects, which includes the model architecture and two well-established training objectives for VLP. Subsequently, we present a detailed explanation of the proposed methods, which are synchronized attentional masked modeling and grouped batch sampling with semi-hard negatives. + +# 3.1. Preliminaries + +For an image-text pair, we refer input sequences as follows: tokenized text embeddings are denoted as $T = \left[\mathbf{t}_{[\mathrm{CLS}]} \right]$ , $\mathbf{t}_1, \ldots, \mathbf{t}_N \in \mathbb{R}^{(N+1) \times D}$ and visual patch embeddings represented as $V = \left[\mathbf{v}_{[\mathrm{CLS}]} \right]$ , $\mathbf{v}_1, \ldots, \mathbf{v}_{N'} \in \mathbb{R}^{(N'+1) \times D}$ . Here, $D, N$ , and $N'$ refer to the transformer dimension, the number of text tokens, and the number of + +image patches, respectively. Additionally, $\mathbf{t}_{[\mathrm{CLS}]}$ and $\mathbf{v}_{[\mathrm{CLS}]}$ specifically reference the [CLS] embeddings. + +The model consists of two components: a teacher model, referred to as the momentum model, and a student model. The student model, denoted as $f_{\theta}$ , is parameterized by $\theta$ which includes a textual encoder $f_{\theta}^{T}(T) \in \mathbb{R}^{(N + 1)\times D}$ , visual encoder $f_{\theta}^{I}(V) \in \mathbb{R}^{(N' + 1)\times D}$ , and multi-modal encoder $f_{\theta}^{M}(f_{\theta}^{T}(T),f_{\theta}^{I}(V)) \in \mathbb{R}^{(N + 1)\times D}$ . For the momentum model $f_{\theta'}$ , the training parameters are updated by the exponential moving average method, $\theta' \gets \beta \theta' + (1 - \beta) \theta$ where $\beta$ represents a hyperparameter. + +Image-Text Contrastive Learning (ITC) At the front of the multi-modal encoder, ITC pre-aligns the joint latent space of the textual encoder and visual encoder. This objective have proved its effectiveness in VLMs [36, 27, 26, 3, 13, 15]. We also adopt the ITC loss framework proposed by [27, 26, 15], which incorporates a momentum encoder for utilizing soft labels as ITC training targets, thereby addressing potential positive instances within negative pairs. + +Image-Text Matching (ITM) For the ITM loss, the model classifies image-text pairs as either matched (positive) or not matched (negative) using a joint representation obtained from the [CLS] token output embedding of the multi-modal encoder. This vector is passed through an FC layer and softmax for binary prediction. Like ALBEF [27], we exploit hard negatives in the ITM task, identifying pairs that share similar semantics but differ in fine-grained details, using in-batch contrastive similarity from ITC. + +# 3.2. Synchronized Attentional Masked Modeling + +We extend the use of momentum model, a self-supervised tool for momentum distillation outlined in MoCo [17] and ALBEF [27], beyond its conventional role of generating pseudo-labels. We employ its multi-modal encoder, which calculate the cross-attention map to identify patches and tokens where image and text features strongly correlate. These elements, indicated by heightened attention weights, are then + +![](images/2f3d9987da1aecb5d1a5c28fae82dfa1134df5e1d4c2025f6767f256364433d6.jpg) +Figure 3. A schematic overview of the SyncMask process: Leveraging cross-attention features from the teacher (momentum) model to generate informative masks for both MIM and MLM tasks. It is important to note that the input for MLM consists of unmasked image paired with masked text. + +selectively masked. This attentional masking, targeting synchronously co-occurring features, enhances cross-modality over traditional random masking methods in the MLM and MIM phases. Moreover, the momentum model's output features provide enhanced labels for these masked regions, offering a depth of information beyond conventional discrete labels. We will explore this method further to understand its full potential in capturing intricate multi-modal interactions. + +Vision-Language Synchronized Attentional Masking MLM and MIM elevate the alignment of visual and textual representations in models. However, when non-co-occurring elements are masked, these techniques face limitations that restrict the model's ability to accurately match multi-modal features. To address this issue, we propose a Synchronized attentional Masking (SyncMask) strategy into masked multimodal modeling objectives. As depicted in Figure 3, we extracted two sets of synchronized attention weights from the cross-attention module of the multi-modal encoder's last layer in the momentum model. The module enables the model to fuse image-text features using a cross-attention mechanism that processes a query $(Q^T)$ , key $(K^I)$ , and value $(V^I)$ , as represented by the following equation: + +$$ +\text {A t t e n t i o n} \left(Q _ {i} ^ {T}, K _ {i} ^ {I}, V _ {i} ^ {I}\right) = \alpha \left(Q _ {i} ^ {T}, K _ {i} ^ {I}\right) \odot V _ {i} ^ {I} \tag {1} +$$ + +where $1 \leq i \leq H$ , with $H$ denoting the number of heads in the MHA, and $\odot$ representing the Hadamard product. In this context, $\alpha$ refers the cross-attention function, which can be expressed as: + +$$ +\alpha \left(Q _ {i} ^ {T}, K _ {i} ^ {I}\right) = \operatorname {s o f t m a x} \left(\frac {Q _ {i} ^ {T} \left(K _ {i} ^ {I}\right) ^ {\top}}{\sqrt {d}}\right) \in \mathbb {R} ^ {(N ^ {\prime} + 1) \times (N + 1)} \tag {2} +$$ + +The function $\alpha(Q_i^T, K_i^I)$ computes the attention weights for the image from the perspective of the text. Similarly, by altering the query $(Q)$ and key $(K)$ , we calculate the text attention weight from the image perspective, as represented by the following equation: + +$$ +\alpha \left(Q _ {i} ^ {I}, K _ {i} ^ {T}\right) = \operatorname {s o f t m a x} \left(\frac {Q _ {i} ^ {I} \left(K _ {i} ^ {T}\right) ^ {\top}}{\sqrt {d}}\right) \in \mathbb {R} ^ {(N + 1) \times \left(N ^ {\prime} + 1\right)} \tag {3} +$$ + +Utilizing Equation 2 and Equation 3, we derive two synchronized textual-visual cross-attention weights. These weights, $\mathbf{o}^T\in \mathbb{R}^N$ $\mathbf{o}^I\in \mathbb{R}^{N'}$ , are obtained by averaging the patch tokens of the last layer, excluding the [CLS] token. This process allows us to map each word in the sentence sequence to its corresponding attention in $\mathbf{o}^T$ .Further, $\mathbf{o}^I$ can be reshaped to $\mathbb{R}^{P\times P}$ , which aligns with the image patches. + +$$ +\operatorname {I d x} ^ {T} = \text {s h u f f l e} \left(\operatorname {s o r t} ^ {\text {d e s c}} \left(\mathbf {o} ^ {T}\right) [ \geq L ]\right) [ \geq K ] \tag {4} +$$ + +$$ +\operatorname {I d x} ^ {I} = \text {s h u f f l e} \left(\operatorname {s o r t} ^ {\text {d e s c}} \left(\mathbf {o} ^ {I}\right) [ \geq L ^ {\prime} ]\right) [ \geq K ^ {\prime} ] \tag {5} +$$ + +In Equation 4 and Equation 5, we first sort the attention weights $\mathbf{o}^T$ and $\mathbf{o}^I$ in descending order (sort desc), and then extract their indices (Idx). Equation 4 focuses on indices corresponding to the top $L$ values of $\mathbf{o}^T$ , where $L$ is a threshold greater than the actual mask size $K$ . These indices are randomly shuffled (shuffle) to introduce randomness, and ultimately, only those indices that satisfy the condition $K \leq L$ are retained. A similar approach is applied to $K'$ and $L'$ in Equation 5. The parameters $K, K', L$ , and $L'$ are defined based on a mask ratio $r \in [0,1]$ . + +The final attention masks for textual and visual components are represented by the vectors $\mathbf{m}^T$ and $\mathbf{m}^I$ , respectively. The computation of these masks utilizes the indices + +![](images/e19aee941cafce9f5e9026208f9f0a8216253559363566a22c879e9481486e16.jpg) +Figure 4. Selection phase of the SyncMask + +$\mathbf{I}\mathrm{dx}^T$ for text and $\mathbf{I}\mathrm{dx}^I$ for images, as formulated in the equation below: + +$$ +\mathbf {m} ^ {T} \left[ \mathrm {I d} \mathbf {x} ^ {T} \right] \leftarrow 1, \quad \mathbf {m} ^ {I} \left[ \mathrm {I d} \mathbf {x} ^ {I} \right] \leftarrow 1 \tag {6} +$$ + +Initially, the vectors $\mathbf{m}^T$ and $\mathbf{m}^I$ are zero-initialized with dimensions $N$ (for text) and $N'$ (for images). They are then updated to binary attention masks, which are integral in capturing the synchronized interaction between textual and visual elements, as illustrated in Figure 4. Subsequently, these masks are employed in the masked modeling processes for text and images (detailed in Equation 7 and Equation 8). + +Building upon this foundation, the text mask vector $\mathbf{m}^T$ consists of elements $[m_1^t,\dots,m_N^t ]\in \{0,1\} ^N$ , while the image mask vector $\mathbf{m}^I$ is composed of $[m_1^i,\dots,m_{N'}^i ]\in$ $\{0,1\}^{N^{\prime}}$ . For each tokenized text vector, the masked version $\tilde{\mathbf{t}}_i$ is determined by: + +$$ +\tilde {\mathbf {t}} _ {j} = \left(1 - m _ {j} ^ {t}\right) \cdot \mathbf {t} _ {j} + m _ {j} ^ {t} \cdot \mathbf {t} _ {\text {m a s k}} \tag {7} +$$ + +In this equation, $1 \leq j \leq N$ , and $\mathbf{t}_{\mathrm{mask}}$ denotes the special token used for textual masking. In the context of Masked Image Modeling (MIM), each image patch is processed with a learnable mask, resulting in the masked image vector $\tilde{\mathbf{v}}_k$ : + +$$ +\tilde {\mathbf {v}} _ {k} = \left(1 - m _ {k} ^ {i}\right) \cdot \mathbf {v} _ {k} + m _ {k} ^ {i} \cdot \mathbf {v} _ {\text {m a s k}} \tag {8} +$$ + +Here, $1 \leq k \leq N'$ , and $\mathbf{v}_{\mathrm{mask}}$ represents the learnable mask embedding [4]. The masked tokenized inputs are thus represented as $\tilde{T} = [\mathbf{t}_{\mathrm{cls}}; \tilde{\mathbf{t}}_1; \ldots; \tilde{\mathbf{t}}_N]$ for text and $\tilde{V} = [\mathbf{v}_{\mathrm{cls}}; \tilde{\mathbf{v}}_1; \ldots; \tilde{\mathbf{v}}_{N'}]$ for image. These masked inputs are processed using the proposed SyncMask $\mathbf{m}^T$ and $\mathbf{m}^I$ . + +Synchronized Attentional Masked Language Modeling MLM predicts masked words based on the surrounding contextual text and image. Many existing VLMs applied the + +MLM method proposed in BERT [7], randomly masking words with a probability of $15\%$ . However, this approach may not be suitable for vision-language datasets with short caption lengths, especially for nonstandard datasets such as fashion, demanding a thorough understanding of fine-grained attributes. We leverage previous works that addressed these issues by increasing masking probabilities [3] and employing attribute prompts [15]. Building upon these methods, we employ masks generated by SyncMask which is contextually attuned to the corresponding image. + +Let $h(V, \tilde{T})$ denote the model's predicted probability for a masked token. $\tilde{y}$ denote a one-hot vocabulary distribution in which the ground-truth token is assigned a probability of 1. MLM minimize cross-entropy loss, described as follows: + +$$ +\mathcal {L} _ {\mathrm {M L M}} = \mathbb {E} _ {(V, \hat {T}) \sim D} [ \operatorname {C E} (\tilde {\boldsymbol {y}}, \boldsymbol {h} (V, \hat {T}) ], \tag {9} +$$ + +where $\mathrm{CE}(\cdot ,\cdot)$ refers the cross-entropy between two vectors. + +Synchronized Attentional Masked Image Modeling In the distillation-based MIM [21, 29, 34], a teacher encoder sees the full image, whereas the student encoder, seeing the masked image, tackles the reconstruction objective. Our method adopts a similar objective framework, but with a key difference: our masks, generated through SyncMask, are designed to reflect text-informed elements in masked image. To calculate the MIM loss, the following approach is used: + +$$ +\begin{array}{l} \operatorname {D i s t} \left(\boldsymbol {f} _ {\theta^ {\prime}} ^ {I} (V), \boldsymbol {f} _ {\theta} ^ {I} (\tilde {V})\right) \\ = \frac {1}{\Omega (\mathbf {m} ^ {I})} \sum_ {k = 1} ^ {N ^ {\prime}} \mathbf {m} _ {k} ^ {I} \cdot \ell_ {1} ^ {S m o o t h} \left(\boldsymbol {f} _ {\theta^ {\prime}} ^ {I} (V) _ {k}, \boldsymbol {f} _ {\theta} ^ {I} (\tilde {V}) _ {k}\right) \tag {10} \\ \end{array} +$$ + +where $\pmb{f}_{\theta^{\prime}}^{I}(\cdot)_{k}$ and $\pmb{f}_{\theta}^{I}(\cdot)_{k}$ represent the output feature of teacher and student model for the $k$ -th image patch, respectively. $\Omega (\cdot)$ means the number of elements with a value of 1 in a vector. + +$$ +\ell_ {1} ^ {\text {S m o o t h}} (a, b) = \left\{ \begin{array}{l l} 0. 5 (a - b) ^ {2} & \text {i f} | a - b | < \gamma \\ | a - b | - 0. 5 & \text {o t h e r w i s e ,} \end{array} \right. \tag {11} +$$ + +where $\ell_1^{Smooth}$ [11] represents a robust L1 loss less sensitive to outliers than the L2 loss and $\gamma$ is a hyperparameter set to 1. Conclusively, the training objective of MIM can be formulated as: + +$$ +\mathcal {L} _ {\mathrm {M I M}} = \mathbb {E} _ {(V, \hat {V}) \sim D} [ \operatorname {D i s t} \left(\boldsymbol {f} _ {\theta^ {\prime}} ^ {I} (V), \boldsymbol {f} _ {\theta} ^ {I} (\tilde {V})\right) ] \tag {12} +$$ + +The final loss $(\mathcal{L})$ is given as: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {M I M}} + \mathcal {L} _ {\mathrm {M L M}} + \mathcal {L} _ {\mathrm {I T C}} + \mathcal {L} _ {\mathrm {I T M}} \tag {13} +$$ + +where $\mathcal{L}_{\mathrm{ITC}}$ and $\mathcal{L}_{\mathrm{ITM}}$ denote ITC and ITM, respectively. Due to space constraints, detailed formulations of these two losses are elaborated in the Appendix. + +
METHODSI2T*T2I*MEAN*I2TT2IMEAN
R@1R@1R@1R@1
VSE++ [9]4.594.604.60---
VL-BERT [38]19.2622.6320.95---
ViLBERT [32]20.9721.1221.05---
Image-BERT [35]22.7624.7823.77---
OSCAR [28]23.3925.1024.25---
FashionBERT [10]23.9626.7525.36---
Kaleido-BERT [45]27.9933.8830.94---
CommerceMM [43]41.6039.6062.75---
EI-CLIP [33]38.7040.0639.3825.7028.4027.05
ALBEF [27]63.9760.5262.2041.6850.9546.32
FashionViL [13]65.5461.8863.7142.8851.3447.11
FashionSAP [15]73.1470.1271.6354.4362.8258.63
Ours75.0071.0073.0055.3964.0659.73
+ +Table 2. Cross-modal retrieval result on FashionGen [37] in the sub/full set of evaluation following previous work. $\star$ sub set. + +# 3.3. Grouped Batch with Semi-hard Negatives + +In the generic domain, [3] proposed the GRIT strategy for enhancing training effectiveness by forming mini-batches with similar examples. In this strategy, the grouping based on similarity phase plays a crucial role. During this phase, similarity calculations are performed in both directions, utilizing the [CLS] outputs from the unimodal encoders. For each example, the algorithm iteratively identifies the index with the highest similarity, alternating between the image-to-text and text-to-image directions in a sequential manner. These highly similar indices are grouped together within the mini-batch, ensuring that each mini-batch consists of examples exhibiting the highest possible similarity. + +However, in fashion datasets, this strategy leads to a false-negative problem in ITM and ITC, where true positives are mislabeled as negatives within mini-batches. To address this problem, we opt to group semi-hard negatives with relatively lower similarity (the $s^{th}$ highest pairwise similarity) instead of the highest $(1^{st})$ during the grouping based on similarity phase, where $s$ represents a predefined hyperparameter that is greater than 1. In addition, we prevent true positive samples from grouping by considering the item indices of the samples. Through this approach, the model is trained with negatives that are similar but exhibit meaningful differences, thereby enabling the effective learning of fine-grained distinctions with a limited dataset. More details are in the Appendix. + +# 4. Experiments + +# 4.1. Implementation Details + +The foundational architecture of the proposed model is aligned with prior works for demonstrating the effectiveness of the proposed approach [27, 3, 15]. The image encoder + +adopt the architecture of ViT-B16 [8], whereas the text encoder comprises the first six blocks of the BERT-bas3 [7]. The multi-modal encoder extends the self-attention layers of the last six blocks of BERT with the cross-attention layers. The proposed model is initialized with pre-trained ALBEF [27] same as the Fashion-SAP to ensure a fair comparison [15]. In addition, we employ the same data augmentation and prompt input strategies as FashionSAP [15]. During pretraining, we conduct experiments using 8 RTX 3090 GPUs each with a batch size of 8 for 30 epochs. We adopt a momentum queue size of 48,000 to facilitate grouped batch sampling, which is consistent with GRIT-VLP [3]. The input image size is set to $256 \times 256$ . We apply the AdamW [31] optimizer with a learning rate of 6e-5. + +# 4.2. Datasets + +FashionGen [37] FashionGen comprises 320K text-image pairs and 40K unique fashion items, each represented by multiple images from different angles. For pre-training, we employ the FashionGen train set, which contains approximately 260.5K text-image pairs. In addition, FashionGen supports various downstream tasks, including text-to-image retrieval, image-to-text retrieval, category recognition, and subcategory recognition. + +FashionIQ [40] FashionIQ encompasses 77K unique fashion items and includes 18K training triplets (i.e., query image, modified text, target image) and 6K validation datasets for a text-guided image retrieval task. It contains three different categories: Dress, Toptee, and Shirt. + +# 4.3. Downstream Tasks + +Cross-modal Retrieval We evaluate a cross-modal retrieval task that includes image-text retrieval (ITR) and text + +![](images/08d1cb3eaa6aae78202f92b240280ee5c838d122718473275bac5f7f09dbc1f0.jpg) +is light blue in color and has more blue + +![](images/23e59536dbf2dded67e37b69ff33c9ae716f43c10a29aa995c31d072cea5f6eb.jpg) +is yellow and has short sleeves and is beige and shorter sleeves + +![](images/8170751b04763694d4eb3f1fb6cf964f3bba9ef3b41a6b3fe97abdbf08155f2d.jpg) +is light blue in color and has more blue + +![](images/8c31d8ede41a43e06be8211a0ab6e007994f76763cb64cde4e1e17e6dff8bf8d.jpg) + +![](images/94e388a82e2ff592d1c2ed228c55692492ed9e48b26c2c48d8ce49541fae0a44.jpg) + +![](images/5bcb4a08e5c54e53f3c8e507e123a9e76ea93cfe83871d4f2a6193b53e97e3a0.jpg) + +![](images/a370e0610b30d30aebc21b72d677fe590e829d50020d476a887a20935a21e1ea.jpg) + +![](images/624a08638b624bf5cdd6f7fe9b1d7fc67fe107b88de443fc6c06085d9c16a586.jpg) + +![](images/b441dc7c3ee79658ef78c52d3f4f26c07008d9df38600cae675a55bddb212663.jpg) + +![](images/ea28ef8ad5e218aa8874f36d2750a73727f9d914f8a59d4c0b1404fb1907a1e4.jpg) + +![](images/318f0906cf0a562e3f0ab193a6e7cbfd9d053ca25270046b1e14cca38f87233a.jpg) + +![](images/4e4f78659e3e372f60ac4ebfcd3d872c018fbb1d3b02775f877a33de274ae55e.jpg) + +![](images/83d40df87bede9d3578974a9d082cd47a8d877f484bc3e954fc2c69a162e7be6.jpg) + +![](images/d8266be9434b5fd28869d4c2b931cfd73e10b0a8bbcc29a28c8081c3ea0476b2.jpg) + +![](images/20fe4e6dcf5048f0bc320a9f9665d414fa03e49b246d94907b01027548d45152.jpg) + +![](images/2c15649a5457f0cbd6ad1496a51f15875a5a034642c7bb201bd9976bf3aafdf3.jpg) + +![](images/282e09030c25ecbc8e69764919c3f6d773171f6e66d53e73d5b84c44d08592d9.jpg) + +![](images/57acc5e9ccd96d50c72998ae79e895bd874d46fc43ce9151d5eb12b372570bdb.jpg) + +![](images/86dfa22cf839e7351b28ef3ac4c3bfa84481d65078ee0c143c708268ec723bce.jpg) + +![](images/acc07d0ffaf8df3e76cdca01693354eeaa3812b6364c8d49ceb76598e5a189b5.jpg) + +![](images/0255ecdf2a25cf15b2f15b9bc4053e116e64d4f910ed649a074361432cf273cf.jpg) + +![](images/1cd8d6951877adcb2f4d504332429f3a65d0fbaac4ade7446264282245d4678f.jpg) + +![](images/ec91c71b5b9074869a80b19f3bc33ae407edfa3c5270b3dad56a50bfad5be857.jpg) + +![](images/1baab022be951c854d8b399acd73fc441aaf27c07c3a2bde1e497970d08f80cd.jpg) + +![](images/6af5f3a7e30792093bf7b4762ab058ca532e094b8cd9ce891869b0d0b9ddd038.jpg) + +![](images/0a16f39e1dfd022b3e9a9f2b58e7ce60f75514a3b3a7a46aafa05eb061129530.jpg) + +![](images/6f6d12ec51ad75be41402fc4fc4d7fff212048e7b508ca136e94c60c66adb3aa.jpg) + +![](images/8e7e624a202a285f28ed5079046737346ea5fe036c71c57602fc3a8ca5eb53f5.jpg) +Figure 5. The top-10 TGIR results of the SyncMask model on the FashionIQ dataset. On the left, the reference images paired with their guided descriptions are shown, while the right side presents the model's predicted images ranked by descending scores. Ground truth images are distinctly outlined with a green bounding box. It is worth mentioning that the set of predictions includes other images that also qualify as suitable matches. + +![](images/9c1dc7a10fe31403f843fb4318d3da5750bdaf9375a3cfcdb912c16598fcd73b.jpg) + +![](images/f4b841bad0e88fd318f474c0906466606356288703f40b4925a13a89294664e7.jpg) + +![](images/572004d130ae93c223711f78f570eae81443ae4042a9188c228a01f1ba518f49.jpg) + +![](images/3c38f01af8724c542738866a5529290a45d23d75f1df949fe4de0c19dab3d5c4.jpg) + +![](images/8942212fffe07cc177c1440480bb860a806223e831294f54e57d47ec97fb1486.jpg) + +
METHODSDRESSTOPTEESHIRTMEAN
R@10R@50R@10R@50R@10R@50R@10R@50
CIRR [30]17.4540.4121.6445.3817.5338.8118.8741.53
VAL [5]22.5344.0027.5351.6822.3844.1524.1546.61
CosMo [25]25.6450.3029.2157.4624.9049.1826.5852.31
DCNet [22]28.9556.730.4458.2923.9547.327.7854.10
FashionVLP [12]32.4260.2938.5168.7931.8958.4434.2762.51
FashionViL [13]33.4759.9434.9860.7925.1750.3931.2157.04
FashionSAP [15]33.7160.4341.9170.9333.1761.3336.2664.23
Ours33.7661.2344.8272.0635.8262.1238.1365.14
+ +image retrieval (TIR). Further, ITR focuses on finding relevant textual descriptions for a given image query. TIR, the inverse task, retrieves pertinent images based on a textual query. These tasks assess the effectiveness of the model in capturing cross-modal relationships between text and images within retrieval scenarios. Following the previous works [13, 15], we evaluate cross-modal retrieval not only on the subset with 1K retrievals but also on the full dataset of FashionGen [37]. The results, including R@1 scores for both subset and the full set, are presented in Table Table 2, demonstrating an improved performance compared to that of the previous results. + +Text-guided Image Retrieval This task aims to retrieve target images by considering query pairs that reference image and modified descriptions; this is more challenging than traditional retrievals. Therefore, we need to select the target image for identifying minor differences in the changes in description while maintaining the characteristics of the reference image. For a fair comparison, we adopt a similar fine-tuning as outlined in the FashionSAP [15]. In the actual dataset, there are many images that match with the query pairs in addition to the target image referred to as the real correct answer, as shown in Figure 5. Thus, a qualitative + +Table 3. Text-guided image retrieval performance in FashionIQ [40] + +
METHODSCRSCR
AccMacro-FAccMacro-F
F-BERT [10]91.2570.5085.2762.00
K-BERT [45]95.0771.4088.0763.60
F-ViL [13]97.4888.6092.2383.02
FashionSAP [15]98.3489.8494.3387.67
Ours98.4190.3194.2187.83
+ +Table 4. CR and SCR results on FashionGen [37]. + +evaluation can be conducted in that the model finds similar images well in addition to the actual correct answer. As shown in Table 3, the proposed model surpasses previous models and demonstrates state-of-the-art performance. + +Category/Subcategory Recognition In this downstream task, the objective is to classify the category and subcategory of fashion items using the textual and visual information provided. In line with earlier studies [10, 45, 13, 15], we simply attach a linear layer to the [CLS] token, which serves as the fusion feature, for task label prediction. As indicated in Table 4, our proposed model exhibits competitive performance compared to existing models. + +![](images/b38fe64ecf743820b70109df9a7c0010bce8a8afd65f64ac4c0ada4d69883d6d.jpg) + +![](images/83a69f65308f388961b7250158c55abbf440375d89af6fc9e0c4c1ee3b7be5d9.jpg) +(a) Unmasked input +(b) Random mask +(c) SyncMask +Figure 6. Examples of random masking (b) and SyncMask (c) for MIM and MLM. The latter applies masks to pertinent features in both the input image and text (a), offering a more context-sensitive selection than the former. + +# 4.4. Ablation Study + +Random vs. Attentional Masking We ablate the SyncMask with other non-synchronized masking methods. Table 5 lists the comparison results for the combination of random and attentional masking. In this experiment, we evaluated five experiments for I2T (R@1), T2I (R@1), CR (Macro-F), SCR (Macro-F), and TMIR (R@10), which were evaluated in the previous experiments. In these experiments, the results of the three experiments with at least one attentional masking were higher than those with no attentional masking (all random). This can be considered an indication of the efficacy of attentional masking. Further, the synergistic performance of the proposed synchronized textual-visual attentional masking (+2.14) is greater than that of only one attentional masking (+0.4, +0.66), which shows the superiority of the SyncMask approach. Figure 6 displays the distinction between applying random masks (b) and SyncMask (c) to unmasked image-text pairs (a) from the FashionGen [37]. SyncMask strategically places masks on image patches that correlate with the text and applies masks to the text informed by the image details, unlike random masking. + +Random vs. Hardest vs. Semi-hard Negative Sampling Table 6 compare the effectiveness of various grouped minibatch sampling strategies across five downstream tasks. The methods evaluated encompassed four scenarios: random grouping, hardest grouping (+1.90), hardest grouping with exclusion of false negatives (+1.98), and semi-hard grouping + +
RTRVATAVMEANGAINI2T R@1T21 R@1CR Macro-FSCR Macro-FTMIR R@10
✓✓70.3173.3069.8086.2186.1636.08
✓✓70.71+0.4074.8070.1086.4986.3135.84
✓✓70.97+0.6674.3070.8085.1286.7837.87
✓✓72.45+2.1475.0071.0090.3187.8338.13
+ +Table 5. Ablation study results for Random vs. Attention Masked Modeling on five downstream tasks. Rt:Random Textual masking. Rv:Random Visual masking. At:Attentional Textual masking. Av:Attentional Visual masking. + +
GROUPEFNMEANGAIN12T R@1T2I R@1CR Macro-FSCR Macro-FTMIR R@10
Random69.9972.3069.0086.2185.4037.06
Hardest71.89+1.9074.7070.5090.2086.6437.40
Hardest✓71.97+1.9874.3071.0090.1087.1537.28
Semi-hard✓72.45+2.4675.0071.0090.3187.8338.13
+ +Table 6. Ablation study results comparing the Grouped Batch Sampling Strategy across five downstream tasks. GROUP: Strategy for grouping phase of GRIT. EFN: Exclude False Negative in a grouping phase using index. + +also excluding false negatives (+2.46). These findings suggest that grouping similar samples in a mini-batch is more beneficial for learning than composing batches with random samples. However, given that the fashion dataset often has multiple captions per image or vice versa, performance gains were observed when systematically preventing the grouping of false negatives by using indexing. Nevertheless, due to the prevalence of inherently similar samples that cannot be systematically excluded, we opted for grouping semi-hard negatives instead of the hardest ones, which resulted in a significant performance boost. This highlights the importance of further research from a data perspective, not just in terms of model architecture or loss functions. + +# 5. Conclusion + +We introduced Synchronized attentional Masking for enhanced masked modeling in fashion-centric VLMs. Leveraging cross-attention features of a momentum model, our method tailors the random mask into a targeted mask for synchronously co-occurring segments in image-text pairs in MLM and MIM objectives. This approach effectively resolves misaligned image-text input issues and improving fine-grained cross-modal representation. Additionally, we addressed data scarcity and distribution challenges in fashion datasets, refining grouped batch sampling with semi-hard negatives for ITM and ITC losses. The experimental results showed our methods outperformed established benchmarks in multiple downstream tasks. + +# Acknowledgment + +This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) under the metaverse support program to nurture the best talents (IITP-2024-RS-2023-00254529) grant funded by the Korea government (MSIT). + +# References + +[1] Yutong Bai, Zeyu Wang, Junfei Xiao, Chen Wei, Huiyu Wang, Alan L Yuille, Yuyin Zhou, and Cihang Xie. Masked autoencoders enable efficient knowledge distillers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24256-24265, 2023. +[2] Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT pretraining of image transformers. In ICLR, 2022. +[3] Jaeseok Byun, Taebaek Hwang, Jianlong Fu, and Taesup Moon. Grit-vlp: Grouped mini-batch sampling for efficient vision and language pre-training. In European Conference on Computer Vision, pages 395-412. Springer, 2022. +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. +[5] Yanbei Chen, Shaogang Gong, and Loris Bazzani. Image search with text feedback by visiolinguistic attention learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3001-3011, 2020. +[6] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. +[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, 2019. +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. +[9] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017. +[10] Dehong Gao, Linbo Jin, Ben Chen, Minghui Qiu, Peng Li, Yi Wei, Yi Hu, and Hao Wang. Fashionbert: Text and image matching with adaptive loss for cross-modal retrieval. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2251-2260, 2020. +[11] Ross Girshick. Fast R-CNN. In ICCV, 2015. +[12] Sonam Goenka, Zhaoheng Zheng, Ayush Jaiswal, Rakesh Chada, Yue Wu, Varsha Hedau, and Pradeep Natarajan. Fashionvlp: Vision language transformer for fashion retrieval with feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14105-14115, 2022. +[13] Xiao Han, Licheng Yu, Xiatian Zhu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Fashionvil: Fashion-focused vision-and-language representation learning. In ECCV. Springer, 2022. + +[14] Xiao Han, Xiatian Zhu, Licheng Yu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Fame-vil: Multi-tasking vision-language model for heterogeneous fashion tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2669-2680, 2023. +[15] Yunpeng Han, Lisai Zhang, Qingcai Chen, Zhijian Chen, Zhonghua Li, Jianxin Yang, and Zhao Cao. Fashionsap: Symbols and attributes prompt for fine-grained fashion vision-language pre-training. In CVPR, 2023. +[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In arXiv:2111.06377, 2021. +[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. +[18] Zhicheng Huang, Zhaoyang Zeng, Yupan Huang, Bei Liu, Dongmei Fu, and Jianlong Fu. Seeing out of the box: End-to-end pre-training for vision-language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12976-12985, 2021. +[19] Zhicheng Huang, Zhaoyang Zeng, Bei Liu, Dongmei Fu, and Jianlong Fu. Pixel-bert: Aligning image pixels with text by deep multi-modal transformers. 2020. +[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. +[21] Ioannis Kakogeorgiou, Spyros Gidaris, Bill Psomas, Yannis Avrithis, Andrei Bursuc, Konstantinos Karantzalos, and Nikos Komodakis. What to hide from your students: Attention-guided masked image modeling. In ECCV, 2022. +[22] Jongseok Kim, Youngjae Yu, Hoeseong Kim, and Gunhee Kim. Dual compositional learning in interactive image retrieval. In AAAI, pages 1771-1779, 2021. +[23] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. +[24] Gukyeong Kwon, Zhaowei Cai, Avinash Ravichandran, Erhan Bas, Rahul Bhotika, and Stefano Soatto. Masked vision and language modeling for multi-modal representation learning. In ICLR, 2023. +[25] Seungmin Lee, Dongwan Kim, and Bohyung Han. Cosmo: Content-style modulation for image retrieval with text feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 802-812, 2021. +[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. +[27] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with + +momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. +[28] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020. +[29] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, et al. MST: Masked self-supervised transformer for visual representation. In NeurIPS, volume 34, 2021. +[30] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2125-2134, 2021. +[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. +[32] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. +[33] Haoyu Ma, Handong Zhao, Zhe Lin, Ajinkya Kale, Zhangyang Wang, Tong Yu, Jiumiang Gu, Sunav Choudhary, and Xiaohui Xie. Ei-clip: Entity-aware interventional contrastive learning for e-commerce cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18051-18061, 2022. +[34] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. A unified view of masked image modeling. In arXiv preprint arXiv:2210.10615, 2022. +[35] Di Qi, Lin Su, Jia Song, Edward Cui, Taroon Bharti, and Arun Sacheti. Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966, 2020. +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. +[37] Negar Rostamzadeh, Seyedarian Hosseini, Thomas Boquet, Wojciech Stokowiec, Ying Zhang, Christian Jauvin, and Chris Pal. Fashion-gen: The generative fashion dataset and challenge. arXiv preprint arXiv:1806.08317, 2018. +[38] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. In International Conference on Learning Representations, 2019. +[39] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. 2019. +[40] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11307-11317, 2021. + +[41] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022. +[42] Hongwei Xue, Yupan Huang, Bei Liu, Houwen Peng, Jianlong Fu, Houqiang Li, and Jiebo Luo. Probing inter-modality: Visual parsing with self-attention for vision-and-language pre-training. volume 34, pages 4514-4528, 2021. +[43] Licheng Yu, Jun Chen, Animesh Sinha, Mengjiao Wang, Yu Chen, Tamara L Berg, and Ning Zhang. Commercemm: Large-scale commerce multimodal representation learning with omni retrieval. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4433-4442, 2022. +[44] Zijia Zhao, Longteng Guo, Xingjian He, Shuai Shao, Zehuan Yuan, and Jing Liu. Mamo: Fine-grained vision-language representations learning with masked multimodal modeling. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1528-1538, 2023. +[45] Mingchen Zhuge, Dehong Gao, Deng-Ping Fan, Linbo Jin, Ben Chen, Haoming Zhou, Minghui Qiu, and Ling Shao. Kaleido-bert: Vision-language pre-training on fashion domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12647–12657, 2021. \ No newline at end of file diff --git a/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/images.zip b/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3e6909aa32a1830a66e5b7e0e643ab1392cbc960 --- /dev/null +++ b/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfbe5d1167b800f721671a13e4dbc9088d27598ac957406fc6f21d9a3af5e978 +size 480833 diff --git a/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/layout.json b/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ac5663eab870119e38b84fc6ef56a5e794119115 --- /dev/null +++ b/2024/SyncMask_ Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining/layout.json @@ -0,0 +1,10096 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 144, + 475, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 144, + 475, + 175 + ], + "spans": [ + { + "bbox": [ + 121, + 144, + 475, + 175 + ], + "type": "text", + "content": "SyncMask: Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 193, + 510, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 193, + 510, + 206 + ], + "spans": [ + { + "bbox": [ + 89, + 193, + 510, + 206 + ], + "type": "text", + "content": "Chull Hwan Song" + }, + { + "bbox": [ + 89, + 193, + 510, + 206 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 89, + 193, + 510, + 206 + ], + "type": "text", + "content": " Taebaek Hwang" + }, + { + "bbox": [ + 89, + 193, + 510, + 206 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 89, + 193, + 510, + 206 + ], + "type": "text", + "content": " Jooyoung Yoon" + }, + { + "bbox": [ + 89, + 193, + 510, + 206 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 193, + 510, + 206 + ], + "type": "text", + "content": " Shunghyun Choi" + }, + { + "bbox": [ + 89, + 193, + 510, + 206 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 193, + 510, + 206 + ], + "type": "text", + "content": " Yeong Hyeon " + }, + { + "bbox": [ + 89, + 193, + 510, + 206 + ], + "type": "inline_equation", + "content": "\\mathrm{Gu}^{2\\dagger}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 232, + 207, + 364, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 207, + 364, + 218 + ], + "spans": [ + { + "bbox": [ + 232, + 207, + 364, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 232, + 207, + 364, + 218 + ], + "type": "text", + "content": "Dealicious Inc. " + }, + { + "bbox": [ + 232, + 207, + 364, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 232, + 207, + 364, + 218 + ], + "type": "text", + "content": "Sejong University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 166, + 242, + 208, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 242, + 208, + 253 + ], + "spans": [ + { + "bbox": [ + 166, + 242, + 208, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 82, + 264, + 291, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 264, + 291, + 521 + ], + "spans": [ + { + "bbox": [ + 82, + 264, + 291, + 521 + ], + "type": "text", + "content": "Vision-language models (VLMs) have made significant strides in cross-modal understanding through large-scale paired datasets. However, in fashion domain, datasets often exhibit a disparity between the information conveyed in image and text. This issue stems from datasets containing multiple images of a single fashion item all paired with one text, leading to cases where some textual details are not visible in individual images. This mismatch, particularly when non-co-occurring elements are masked, undermines the training of conventional VLM objectives like Masked Language Modeling and Masked Image Modeling, thereby hindering the model's ability to accurately align fine-grained visual and textual features. Addressing this problem, we propose Synchronized attentional Masking (SyncMask), which generate masks that pinpoint the image patches and word tokens where the information co-occur in both image and text. This synchronization is accomplished by harnessing cross-attentional features obtained from a momentum model, ensuring a precise alignment between the two modalities. Additionally, we enhance grouped batch sampling with semihard negatives, effectively mitigating false negative issues in Image-Text Matching and Image-Text Contrastive learning objectives within fashion datasets. Our experiments demonstrate the effectiveness of the proposed approach, outperforming existing methods in three downstream tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 83, + 542, + 153, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 542, + 153, + 553 + ], + "spans": [ + { + "bbox": [ + 83, + 542, + 153, + 553 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 82, + 560, + 291, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 560, + 291, + 643 + ], + "spans": [ + { + "bbox": [ + 82, + 560, + 291, + 643 + ], + "type": "text", + "content": "Recently, there has been rapid progress in developing Vision-Language Pretraining (VLP) [32, 38, 6, 39, 19, 18, 42, 23, 36, 20, 27, 3], paving the way to bridge the gap between visual and textual features. These VLP methods, trained on extensive image-text datasets, have enabled a deeper understanding of semantic alignment across different modalities. By fine-tuning these pretrained models for specific tasks, particularly in data-scarce areas like image-" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 311, + 261, + 338, + 319 + ], + "blocks": [ + { + "bbox": [ + 337, + 249, + 374, + 255 + ], + "lines": [ + { + "bbox": [ + 337, + 249, + 374, + 255 + ], + "spans": [ + { + "bbox": [ + 337, + 249, + 374, + 255 + ], + "type": "text", + "content": "[images, caption]" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 311, + 261, + 338, + 319 + ], + "lines": [ + { + "bbox": [ + 311, + 261, + 338, + 319 + ], + "spans": [ + { + "bbox": [ + 311, + 261, + 338, + 319 + ], + "type": "image", + "image_path": "6be5b5d0e6cd21fae3e8adc905e60aefed227250a53ee29eae28a238a1fcc05a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 334, + 404, + 340 + ], + "lines": [ + { + "bbox": [ + 306, + 334, + 404, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 334, + 404, + 340 + ], + "type": "text", + "content": "Notched lapel collar. Padded shoulders...." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 342, + 394, + 350 + ], + "lines": [ + { + "bbox": [ + 306, + 342, + 394, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 394, + 350 + ], + "type": "text", + "content": "Four-button cuffs. Two vents at back..." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 374, + 489, + 384 + ], + "lines": [ + { + "bbox": [ + 328, + 374, + 489, + 384 + ], + "spans": [ + { + "bbox": [ + 328, + 374, + 489, + 384 + ], + "type": "text", + "content": "Figure 1. Example of misaligned masks in the MLM task." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 339, + 262, + 400, + 319 + ], + "blocks": [ + { + "bbox": [ + 347, + 241, + 365, + 248 + ], + "lines": [ + { + "bbox": [ + 347, + 241, + 365, + 248 + ], + "spans": [ + { + "bbox": [ + 347, + 241, + 365, + 248 + ], + "type": "text", + "content": "Dataset" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 339, + 262, + 400, + 319 + ], + "lines": [ + { + "bbox": [ + 339, + 262, + 400, + 319 + ], + "spans": [ + { + "bbox": [ + 339, + 262, + 400, + 319 + ], + "type": "image", + "image_path": "596396104a1e803272469db78a90aba6efd435297d72c420f22fa425f3907645.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 325, + 383, + 332 + ], + "lines": [ + { + "bbox": [ + 306, + 325, + 383, + 332 + ], + "spans": [ + { + "bbox": [ + 306, + 325, + 383, + 332 + ], + "type": "text", + "content": "Long sleeve blazer in deep navy." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 352, + 357, + 360, + 364 + ], + "lines": [ + { + "bbox": [ + 352, + 357, + 360, + 364 + ], + "spans": [ + { + "bbox": [ + 352, + 357, + 360, + 364 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 451, + 262, + 477, + 319 + ], + "blocks": [ + { + "bbox": [ + 446, + 242, + 480, + 248 + ], + "lines": [ + { + "bbox": [ + 446, + 242, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 446, + 242, + 480, + 248 + ], + "type": "text", + "content": "Input of MLM" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 437, + 249, + 488, + 255 + ], + "lines": [ + { + "bbox": [ + 437, + 249, + 488, + 255 + ], + "spans": [ + { + "bbox": [ + 437, + 249, + 488, + 255 + ], + "type": "text", + "content": "[image, masked caption]" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 451, + 262, + 477, + 319 + ], + "lines": [ + { + "bbox": [ + 451, + 262, + 477, + 319 + ], + "spans": [ + { + "bbox": [ + 451, + 262, + 477, + 319 + ], + "type": "image", + "image_path": "4d929c6dfd94b47404609a70da777ed04be62b257c307440ec8bfa4276f7d8b9.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 325, + 503, + 333 + ], + "lines": [ + { + "bbox": [ + 413, + 325, + 503, + 333 + ], + "spans": [ + { + "bbox": [ + 413, + 325, + 503, + 333 + ], + "type": "text", + "content": "Long sleeve blazer [MASK] deep navy." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 413, + 334, + 511, + 341 + ], + "lines": [ + { + "bbox": [ + 413, + 334, + 511, + 341 + ], + "spans": [ + { + "bbox": [ + 413, + 334, + 511, + 341 + ], + "type": "text", + "content": "Notched lapel collar. [MASK] shoulders..." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 414, + 342, + 511, + 350 + ], + "lines": [ + { + "bbox": [ + 414, + 342, + 511, + 350 + ], + "spans": [ + { + "bbox": [ + 414, + 342, + 511, + 350 + ], + "type": "text", + "content": "Four-button [MASK]. Two veins at [MASK]." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 460, + 357, + 469, + 363 + ], + "lines": [ + { + "bbox": [ + 460, + 357, + 469, + 363 + ], + "spans": [ + { + "bbox": [ + 460, + 357, + 469, + 363 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 397, + 512, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 512, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 512, + 449 + ], + "type": "text", + "content": "text retrieval, notable performance improvements have been observed. In the pretraining phase, various factors such as model architecture, training objectives, and batch sampling techniques play a crucial role in effectively harnessing the joint representation of multi-modal data." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 450, + 513, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 513, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 513, + 574 + ], + "type": "text", + "content": "However, there are some issues in applying conventional generic VLP approaches to task-specific domains such as fashion. Fashion VLP models [10, 45, 13, 14, 15] typically employ objectives such as Masked Language Modeling (MLM) and Masked Image Modeling (MIM). These methods mask elements like text words or image patches, leveraging surrounding context for prediction or reconstruction. They boost cross-modality by making models infer masked text tokens or image patches from aligned features. However, existing MLM and MIM often suffer from inherent misalignment limitations because the masks are generated randomly, often leading to unmatched elements being masked." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 576, + 514, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 576, + 514, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 576, + 514, + 669 + ], + "type": "text", + "content": "To illustrate these limitations, consider Figure 1 (a) shows a single description associated with four images from the FashionGen [37] dataset. In Figure 1 (b), a random masking scenario is shown where the blue [MASK] might lead the model to predict the masked word using only the text context, thereby not incorporating the image information. Similarly, the two red [MASK] tokens in (b) lack relevance to the accompanying image, thus hindering the model's ability to connecting between visual and textual features. Fur-" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "spans": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "spans": [ + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 207, + 45, + 431, + 56 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 45, + 431, + 56 + ], + "spans": [ + { + "bbox": [ + 207, + 45, + 431, + 56 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 189, + 57, + 448, + 67 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 57, + 448, + 67 + ], + "spans": [ + { + "bbox": [ + 189, + 57, + 448, + 67 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 651, + 149, + 660 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 651, + 149, + 660 + ], + "spans": [ + { + "bbox": [ + 93, + 651, + 149, + 660 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 95, + 660, + 158, + 668 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 660, + 158, + 668 + ], + "spans": [ + { + "bbox": [ + 95, + 660, + 158, + 668 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 294, + 696, + 317, + 706 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 696, + 317, + 706 + ], + "spans": [ + { + "bbox": [ + 294, + 696, + 317, + 706 + ], + "type": "text", + "content": "13948" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 82, + 118, + 290, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 118, + 290, + 189 + ], + "spans": [ + { + "bbox": [ + 82, + 118, + 290, + 189 + ], + "type": "text", + "content": "thermore, in the MIM task, random masking may inadvertently cover parts of the image that contain fashion items not described in the text, leading to mismatches during the training of the alignment of cross-modal features. To solve these problem, we propose SyncMask, selecting masks that represent synchronously co-occurring features utilizing the vision-language cross-attention map." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 82, + 190, + 291, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 190, + 291, + 343 + ], + "spans": [ + { + "bbox": [ + 82, + 190, + 291, + 343 + ], + "type": "text", + "content": "In addition, compared to general domains, the fashion domain often suffers from smaller dataset sizes and less variance in data distribution. This suggests that using standard VLP methods may not adequately distinguish the fine-grained features vital for fashion-related tasks. Thus, we pay attention to a grouped batch sampling technique [3] that similar samples are gradually collected within the batch as the training progresses, impacting the pretraining objectives of Image-Text Contrastive Learning (ITC) and Image-Text Matching (ITM). When similar samples exist within the same batch, it becomes more challenging to differentiate positives and negatives during the training of ITC and ITM compared to using random samples. This encourages the model to more intensely focus on learning fine-grained differences, even with less training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 82, + 344, + 290, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 344, + 290, + 456 + ], + "spans": [ + { + "bbox": [ + 82, + 344, + 290, + 456 + ], + "type": "text", + "content": "The existing grouped batch sampling method uses the output features of two uni-modal encoders to find the most similar sample. However, as shown in Figure 1, there are many data that have the same caption for multiple images as for the fashion domain. Therefore, a false negative problem that causes actual positive samples in the same batch to be wrongly labelled as negative when learning ITC and ITM arises if the existing methods are used without changes. To overcome this limitation, we propose a semi-hard negative sampling technique with lower similarities between samples that constitute the batch while removing the false negative." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 457, + 278, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 457, + 278, + 467 + ], + "spans": [ + { + "bbox": [ + 92, + 457, + 278, + 467 + ], + "type": "text", + "content": "In summary, the main contributions of this study are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 467, + 290, + 591 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 91, + 467, + 290, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 467, + 290, + 548 + ], + "spans": [ + { + "bbox": [ + 91, + 467, + 290, + 548 + ], + "type": "text", + "content": "1. Synchronized Attentional Masking: We introduce SyncMask, which replaces random mask in MLM and MIM with targeted mask of co-occurring segments in image-text pairs. By utilizing cross-attention features from a momentum model to generate these masks, this method effectively addresses the problem of mismatched image-text inputs, thereby enhancing fine-grained alignment of cross-modal features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 551, + 290, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 551, + 290, + 591 + ], + "spans": [ + { + "bbox": [ + 89, + 551, + 290, + 591 + ], + "type": "text", + "content": "2. Refined Grouped Batch Sampling: Our method incorporates semi-hard negative sampling to tackle data scarcity and distribution disparities in domain-specific datasets, thereby reducing false negatives." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 83, + 599, + 162, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 599, + 162, + 609 + ], + "spans": [ + { + "bbox": [ + 83, + 599, + 162, + 609 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 82, + 616, + 291, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 616, + 291, + 669 + ], + "spans": [ + { + "bbox": [ + 82, + 616, + 291, + 669 + ], + "type": "text", + "content": "Vision and Language (VL) Model Recently, VLMs have focused on enhancing model architecture and designing objectives to integrate visual and textual features effectively. Early studies [32, 38, 39, 6] have used object detectors for extracting visual features as an input for a multi-modal trans" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 305, + 115, + 513, + 253 + ], + "blocks": [ + { + "bbox": [ + 305, + 115, + 513, + 253 + ], + "lines": [ + { + "bbox": [ + 305, + 115, + 513, + 253 + ], + "spans": [ + { + "bbox": [ + 305, + 115, + 513, + 253 + ], + "type": "table", + "html": "
METHODMMMLMMIMATMAVMONUP
DMAE [1]✓
MaskDistill [34]✓✓✓
AttMask [21]✓✓✓
ALBEF [27]✓✓
MaskVLM [24]✓✓✓
MAMO [44]✓✓✓
FashionBERT [10]✓✓✓
Kaleido-BERT [45]✓✓✓✓✓
FashionViL [13]✓✓✓
FashionSAP [15]✓✓
Ours✓✓✓✓✓✓
", + "image_path": "4e7e6204d8167c461444feb991e82751efaba6503449f8983911ebfe6f8f6228.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 254, + 513, + 288 + ], + "lines": [ + { + "bbox": [ + 305, + 254, + 513, + 288 + ], + "spans": [ + { + "bbox": [ + 305, + 254, + 513, + 288 + ], + "type": "text", + "content": "Table 1. Related works vs. Ours on Masked Modeling. MM:MultiModal. MLM:Masked Language Modeling. MIM:Masked Image Modeling. ATM:Attentional Textual Mask. AVM:Attentional Visual Mask. OnUp:Online Update for Attentional Masking" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 299, + 514, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 514, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 514, + 454 + ], + "type": "text", + "content": "former along with textual features. The objective for training models extends vanilla BERT [7] to use MLM, MIM, and ITM losses; however, the object detection module incurs a high computational cost for training and inference. Therefore, there have been attempts to replace it with CNN [19, 18] or linear projection [23]. These studies commonly train models with MLM and ITM, tailoring MIM to their specific architectures. Concurrently, CLIP [36], ALIGN [20] propose models comprising only two unimodal encoders, demonstrating the outstanding representation embedding capabilities of contrastive learning. ALBEF [27] add a contrastive learning objective to the previous multi-modal transformer structure for aligning the two modalities before fusion. Based on this model, GRIT-VLP [3] demonstrate improved learning efficiency when configuring batches with hard negative samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 473, + 514, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 514, + 670 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 514, + 670 + ], + "type": "text", + "content": "FashionVL Model In recent years, various studies [10, 45, 13, 14, 15] attempted to capture the finer details of images, building upon established models and pre-training objectives from the general VL task. FashionBERT [10] integrates patch-based image features and BERT-based text representations for addressing the limitations of region of interests (RoIs) in capturing fine-grained details. Kaleido-BERT [45] improves fine-grained fashion cross-modality representations through alignment guided masking compared to random masking. FashionViL [13] employs a versatile VLP framework, leveraging two pre-training tasks for capturing the rich fine-grained information of fashion data. FashionSAP [15] employs abstract fashion symbols and an attributes prompt technique for effectively modeling multi-modal fashion attributes. We clarify the importance of the attentional masking technique that employs alignment between MIM and MLM building upon preceding methods. In addition, we underscore a previously unaddressed need for grouped batch sampling within the fashion domain." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 697, + 317, + 706 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 697, + 317, + 706 + ], + "spans": [ + { + "bbox": [ + 294, + 697, + 317, + 706 + ], + "type": "text", + "content": "13949" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 117, + 342, + 218 + ], + "blocks": [ + { + "bbox": [ + 85, + 117, + 342, + 218 + ], + "lines": [ + { + "bbox": [ + 85, + 117, + 342, + 218 + ], + "spans": [ + { + "bbox": [ + 85, + 117, + 342, + 218 + ], + "type": "image", + "image_path": "9ca3eb8e0a9ba93e86ada5bc6bee0afaa51d79d29825dc4d30198641bcc90e07.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 225, + 513, + 243 + ], + "lines": [ + { + "bbox": [ + 82, + 225, + 513, + 243 + ], + "spans": [ + { + "bbox": [ + 82, + 225, + 513, + 243 + ], + "type": "text", + "content": "Figure 2. Overview of masking strategies using a teacher-student distillation framework. 1) Uni-modal models: (a) random masking, (b) teacher-guided attentional masking. 2) Multi-modal models: (c) random text masking, (d) random image/text masking, (e) teacher-guided cross-attentional masking (Ours)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 85, + 251, + 291, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 251, + 291, + 499 + ], + "spans": [ + { + "bbox": [ + 85, + 251, + 291, + 499 + ], + "type": "text", + "content": "Attention-guided Masked Modeling MIM and MLM leverage unmasked contextual clues to predict masked visual and textual elements, respectively. As shown in the upper section of Table 1 and Figure 2 (a, b), in the MIM task, prior studies have evolved from random masking [16, 2, 41, 1] to attention-guided masking [29, 34, 21], demonstrating that targeting highly-attended patches with teacher-student distillation framework improves masked modeling outcomes. Masked modeling also has been pivotal in advancing cross-modal alignment within VLMs, spanning both general [38, 32, 27, 24, 44] and fashion-specific [10, 45, 13, 15] domains. This is briefly illustrated in Figure 2 (c, d) and delineated in the middle and lower sections of Table 1. These enhance the model's proficiency aligning co-occurring visual and textual representations. However, challenges emerge when masking non-co-occurring elements, which hinders the accurate pairing of visual and textual features. KaleidoBERT [45] improve fine-grained cross-modality representations through text-image alignment-guided masking. This requires additional components for an attention-based alignment generator, increasing computational demands and potential model overfitting on fixed text-image mask pairs. Our approach overcomes these limitations by progressively tailoring masks during end-to-end training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 506, + 135, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 506, + 135, + 517 + ], + "spans": [ + { + "bbox": [ + 83, + 506, + 135, + 517 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 524, + 290, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 524, + 290, + 586 + ], + "spans": [ + { + "bbox": [ + 82, + 524, + 290, + 586 + ], + "type": "text", + "content": "We provide an overview of the preliminary aspects, which includes the model architecture and two well-established training objectives for VLP. Subsequently, we present a detailed explanation of the proposed methods, which are synchronized attentional masked modeling and grouped batch sampling with semi-hard negatives." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 590, + 157, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 590, + 157, + 601 + ], + "spans": [ + { + "bbox": [ + 83, + 590, + 157, + 601 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 607, + 291, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 607, + 291, + 668 + ], + "spans": [ + { + "bbox": [ + 83, + 607, + 291, + 668 + ], + "type": "text", + "content": "For an image-text pair, we refer input sequences as follows: tokenized text embeddings are denoted as " + }, + { + "bbox": [ + 83, + 607, + 291, + 668 + ], + "type": "inline_equation", + "content": "T = \\left[\\mathbf{t}_{[\\mathrm{CLS}]} \\right]" + }, + { + "bbox": [ + 83, + 607, + 291, + 668 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 83, + 607, + 291, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_1, \\ldots, \\mathbf{t}_N \\in \\mathbb{R}^{(N+1) \\times D}" + }, + { + "bbox": [ + 83, + 607, + 291, + 668 + ], + "type": "text", + "content": " and visual patch embeddings represented as " + }, + { + "bbox": [ + 83, + 607, + 291, + 668 + ], + "type": "inline_equation", + "content": "V = \\left[\\mathbf{v}_{[\\mathrm{CLS}]} \\right]" + }, + { + "bbox": [ + 83, + 607, + 291, + 668 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 83, + 607, + 291, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_1, \\ldots, \\mathbf{v}_{N'} \\in \\mathbb{R}^{(N'+1) \\times D}" + }, + { + "bbox": [ + 83, + 607, + 291, + 668 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 83, + 607, + 291, + 668 + ], + "type": "inline_equation", + "content": "D, N" + }, + { + "bbox": [ + 83, + 607, + 291, + 668 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 83, + 607, + 291, + 668 + ], + "type": "inline_equation", + "content": "N'" + }, + { + "bbox": [ + 83, + 607, + 291, + 668 + ], + "type": "text", + "content": " refer to the transformer dimension, the number of text tokens, and the number of" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 252, + 511, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 252, + 511, + 273 + ], + "spans": [ + { + "bbox": [ + 305, + 252, + 511, + 273 + ], + "type": "text", + "content": "image patches, respectively. Additionally, " + }, + { + "bbox": [ + 305, + 252, + 511, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{[\\mathrm{CLS}]}" + }, + { + "bbox": [ + 305, + 252, + 511, + 273 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 252, + 511, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{[\\mathrm{CLS}]}" + }, + { + "bbox": [ + 305, + 252, + 511, + 273 + ], + "type": "text", + "content": " specifically reference the [CLS] embeddings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "text", + "content": "The model consists of two components: a teacher model, referred to as the momentum model, and a student model. The student model, denoted as " + }, + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "text", + "content": ", is parameterized by " + }, + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "text", + "content": " which includes a textual encoder " + }, + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{T}(T) \\in \\mathbb{R}^{(N + 1)\\times D}" + }, + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "text", + "content": ", visual encoder " + }, + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{I}(V) \\in \\mathbb{R}^{(N' + 1)\\times D}" + }, + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "text", + "content": ", and multi-modal encoder " + }, + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{M}(f_{\\theta}^{T}(T),f_{\\theta}^{I}(V)) \\in \\mathbb{R}^{(N + 1)\\times D}" + }, + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "text", + "content": ". For the momentum model " + }, + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "inline_equation", + "content": "f_{\\theta'}" + }, + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "text", + "content": ", the training parameters are updated by the exponential moving average method, " + }, + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "inline_equation", + "content": "\\theta' \\gets \\beta \\theta' + (1 - \\beta) \\theta" + }, + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 274, + 513, + 370 + ], + "type": "text", + "content": " represents a hyperparameter." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 384, + 513, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 513, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 513, + 467 + ], + "type": "text", + "content": "Image-Text Contrastive Learning (ITC) At the front of the multi-modal encoder, ITC pre-aligns the joint latent space of the textual encoder and visual encoder. This objective have proved its effectiveness in VLMs [36, 27, 26, 3, 13, 15]. We also adopt the ITC loss framework proposed by [27, 26, 15], which incorporates a momentum encoder for utilizing soft labels as ITC training targets, thereby addressing potential positive instances within negative pairs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 480, + 512, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 512, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 512, + 573 + ], + "type": "text", + "content": "Image-Text Matching (ITM) For the ITM loss, the model classifies image-text pairs as either matched (positive) or not matched (negative) using a joint representation obtained from the [CLS] token output embedding of the multi-modal encoder. This vector is passed through an FC layer and softmax for binary prediction. Like ALBEF [27], we exploit hard negatives in the ITM task, identifying pairs that share similar semantics but differ in fine-grained details, using in-batch contrastive similarity from ITC." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 580, + 503, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 580, + 503, + 591 + ], + "spans": [ + { + "bbox": [ + 306, + 580, + 503, + 591 + ], + "type": "text", + "content": "3.2. Synchronized Attentional Masked Modeling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 597, + 512, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 512, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 512, + 669 + ], + "type": "text", + "content": "We extend the use of momentum model, a self-supervised tool for momentum distillation outlined in MoCo [17] and ALBEF [27], beyond its conventional role of generating pseudo-labels. We employ its multi-modal encoder, which calculate the cross-attention map to identify patches and tokens where image and text features strongly correlate. These elements, indicated by heightened attention weights, are then" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 696, + 317, + 706 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 696, + 317, + 706 + ], + "spans": [ + { + "bbox": [ + 294, + 696, + 317, + 706 + ], + "type": "text", + "content": "13950" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 115, + 504, + 290 + ], + "blocks": [ + { + "bbox": [ + 95, + 115, + 504, + 290 + ], + "lines": [ + { + "bbox": [ + 95, + 115, + 504, + 290 + ], + "spans": [ + { + "bbox": [ + 95, + 115, + 504, + 290 + ], + "type": "image", + "image_path": "2f3d9987da1aecb5d1a5c28fae82dfa1134df5e1d4c2025f6767f256364433d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 296, + 513, + 314 + ], + "lines": [ + { + "bbox": [ + 82, + 296, + 513, + 314 + ], + "spans": [ + { + "bbox": [ + 82, + 296, + 513, + 314 + ], + "type": "text", + "content": "Figure 3. A schematic overview of the SyncMask process: Leveraging cross-attention features from the teacher (momentum) model to generate informative masks for both MIM and MLM tasks. It is important to note that the input for MLM consists of unmasked image paired with masked text." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 82, + 324, + 291, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 324, + 291, + 406 + ], + "spans": [ + { + "bbox": [ + 82, + 324, + 291, + 406 + ], + "type": "text", + "content": "selectively masked. This attentional masking, targeting synchronously co-occurring features, enhances cross-modality over traditional random masking methods in the MLM and MIM phases. Moreover, the momentum model's output features provide enhanced labels for these masked regions, offering a depth of information beyond conventional discrete labels. We will explore this method further to understand its full potential in capturing intricate multi-modal interactions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 418, + 290, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 418, + 290, + 563 + ], + "spans": [ + { + "bbox": [ + 83, + 418, + 290, + 563 + ], + "type": "text", + "content": "Vision-Language Synchronized Attentional Masking MLM and MIM elevate the alignment of visual and textual representations in models. However, when non-co-occurring elements are masked, these techniques face limitations that restrict the model's ability to accurately match multi-modal features. To address this issue, we propose a Synchronized attentional Masking (SyncMask) strategy into masked multimodal modeling objectives. As depicted in Figure 3, we extracted two sets of synchronized attention weights from the cross-attention module of the multi-modal encoder's last layer in the momentum model. The module enables the model to fuse image-text features using a cross-attention mechanism that processes a query " + }, + { + "bbox": [ + 83, + 418, + 290, + 563 + ], + "type": "inline_equation", + "content": "(Q^T)" + }, + { + "bbox": [ + 83, + 418, + 290, + 563 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 83, + 418, + 290, + 563 + ], + "type": "inline_equation", + "content": "(K^I)" + }, + { + "bbox": [ + 83, + 418, + 290, + 563 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 83, + 418, + 290, + 563 + ], + "type": "inline_equation", + "content": "(V^I)" + }, + { + "bbox": [ + 83, + 418, + 290, + 563 + ], + "type": "text", + "content": ", as represented by the following equation:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 574, + 290, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 574, + 290, + 587 + ], + "spans": [ + { + "bbox": [ + 104, + 574, + 290, + 587 + ], + "type": "interline_equation", + "content": "\\text {A t t e n t i o n} \\left(Q _ {i} ^ {T}, K _ {i} ^ {I}, V _ {i} ^ {I}\\right) = \\alpha \\left(Q _ {i} ^ {T}, K _ {i} ^ {I}\\right) \\odot V _ {i} ^ {I} \\tag {1}", + "image_path": "9c67bdf46ec0d2e59633984e826ff6d49e8dc3033b19bd81a626d988bbdc53e0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 594, + 290, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 594, + 290, + 635 + ], + "spans": [ + { + "bbox": [ + 82, + 594, + 290, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 82, + 594, + 290, + 635 + ], + "type": "inline_equation", + "content": "1 \\leq i \\leq H" + }, + { + "bbox": [ + 82, + 594, + 290, + 635 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 82, + 594, + 290, + 635 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 82, + 594, + 290, + 635 + ], + "type": "text", + "content": " denoting the number of heads in the MHA, and " + }, + { + "bbox": [ + 82, + 594, + 290, + 635 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 82, + 594, + 290, + 635 + ], + "type": "text", + "content": " representing the Hadamard product. In this context, " + }, + { + "bbox": [ + 82, + 594, + 290, + 635 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 82, + 594, + 290, + 635 + ], + "type": "text", + "content": " refers the cross-attention function, which can be expressed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 636, + 289, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 636, + 289, + 669 + ], + "spans": [ + { + "bbox": [ + 88, + 636, + 289, + 669 + ], + "type": "interline_equation", + "content": "\\alpha \\left(Q _ {i} ^ {T}, K _ {i} ^ {I}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {i} ^ {T} \\left(K _ {i} ^ {I}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\in \\mathbb {R} ^ {(N ^ {\\prime} + 1) \\times (N + 1)} \\tag {2}", + "image_path": "4120c82facd85a49ca202d10d79f0798e40005650cef0b3dc040f01504044e69.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 323, + 512, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 512, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 512, + 375 + ], + "type": "text", + "content": "The function " + }, + { + "bbox": [ + 304, + 323, + 512, + 375 + ], + "type": "inline_equation", + "content": "\\alpha(Q_i^T, K_i^I)" + }, + { + "bbox": [ + 304, + 323, + 512, + 375 + ], + "type": "text", + "content": " computes the attention weights for the image from the perspective of the text. Similarly, by altering the query " + }, + { + "bbox": [ + 304, + 323, + 512, + 375 + ], + "type": "inline_equation", + "content": "(Q)" + }, + { + "bbox": [ + 304, + 323, + 512, + 375 + ], + "type": "text", + "content": " and key " + }, + { + "bbox": [ + 304, + 323, + 512, + 375 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 304, + 323, + 512, + 375 + ], + "type": "text", + "content": ", we calculate the text attention weight from the image perspective, as represented by the following equation:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 310, + 376, + 512, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 376, + 512, + 409 + ], + "spans": [ + { + "bbox": [ + 310, + 376, + 512, + 409 + ], + "type": "interline_equation", + "content": "\\alpha \\left(Q _ {i} ^ {I}, K _ {i} ^ {T}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {i} ^ {I} \\left(K _ {i} ^ {T}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\in \\mathbb {R} ^ {(N + 1) \\times \\left(N ^ {\\prime} + 1\\right)} \\tag {3}", + "image_path": "53594e3b26347334414fd43b675dbd73ae96763e52827e3a70d4d6da031f61ab.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 408, + 513, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 513, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 513, + 491 + ], + "type": "text", + "content": "Utilizing Equation 2 and Equation 3, we derive two synchronized textual-visual cross-attention weights. These weights, " + }, + { + "bbox": [ + 304, + 408, + 513, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{o}^T\\in \\mathbb{R}^N" + }, + { + "bbox": [ + 304, + 408, + 513, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{o}^I\\in \\mathbb{R}^{N'}" + }, + { + "bbox": [ + 304, + 408, + 513, + 491 + ], + "type": "text", + "content": " , are obtained by averaging the patch tokens of the last layer, excluding the [CLS] token. This process allows us to map each word in the sentence sequence to its corresponding attention in " + }, + { + "bbox": [ + 304, + 408, + 513, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{o}^T" + }, + { + "bbox": [ + 304, + 408, + 513, + 491 + ], + "type": "text", + "content": " .Further, " + }, + { + "bbox": [ + 304, + 408, + 513, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{o}^I" + }, + { + "bbox": [ + 304, + 408, + 513, + 491 + ], + "type": "text", + "content": " can be reshaped to " + }, + { + "bbox": [ + 304, + 408, + 513, + 491 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{P\\times P}" + }, + { + "bbox": [ + 304, + 408, + 513, + 491 + ], + "type": "text", + "content": " , which aligns with the image patches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 328, + 491, + 512, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 491, + 512, + 504 + ], + "spans": [ + { + "bbox": [ + 328, + 491, + 512, + 504 + ], + "type": "interline_equation", + "content": "\\operatorname {I d x} ^ {T} = \\text {s h u f f l e} \\left(\\operatorname {s o r t} ^ {\\text {d e s c}} \\left(\\mathbf {o} ^ {T}\\right) [ \\geq L ]\\right) [ \\geq K ] \\tag {4}", + "image_path": "4dbe6c07d28788b976c9105236899f654bedf906399c1b22b0581af974c3d024.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 327, + 516, + 512, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 516, + 512, + 529 + ], + "spans": [ + { + "bbox": [ + 327, + 516, + 512, + 529 + ], + "type": "interline_equation", + "content": "\\operatorname {I d x} ^ {I} = \\text {s h u f f l e} \\left(\\operatorname {s o r t} ^ {\\text {d e s c}} \\left(\\mathbf {o} ^ {I}\\right) [ \\geq L ^ {\\prime} ]\\right) [ \\geq K ^ {\\prime} ] \\tag {5}", + "image_path": "b0dd30e392f954ebe61d1680ac476da945be445fc8a92a66ef684d4b15498948.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "text", + "content": "In Equation 4 and Equation 5, we first sort the attention weights " + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{o}^T" + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{o}^I" + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "text", + "content": " in descending order (sort desc), and then extract their indices (Idx). Equation 4 focuses on indices corresponding to the top " + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "text", + "content": " values of " + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{o}^T" + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "text", + "content": " is a threshold greater than the actual mask size " + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "text", + "content": ". These indices are randomly shuffled (shuffle) to introduce randomness, and ultimately, only those indices that satisfy the condition " + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "inline_equation", + "content": "K \\leq L" + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "text", + "content": " are retained. A similar approach is applied to " + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "inline_equation", + "content": "K'" + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "inline_equation", + "content": "L'" + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "text", + "content": " in Equation 5. The parameters " + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "inline_equation", + "content": "K, K', L" + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "inline_equation", + "content": "L'" + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "text", + "content": " are defined based on a mask ratio " + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "inline_equation", + "content": "r \\in [0,1]" + }, + { + "bbox": [ + 304, + 534, + 513, + 637 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 638, + 513, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 638, + 513, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 638, + 513, + 669 + ], + "type": "text", + "content": "The final attention masks for textual and visual components are represented by the vectors " + }, + { + "bbox": [ + 304, + 638, + 513, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^T" + }, + { + "bbox": [ + 304, + 638, + 513, + 669 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 638, + 513, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^I" + }, + { + "bbox": [ + 304, + 638, + 513, + 669 + ], + "type": "text", + "content": ", respectively. The computation of these masks utilizes the indices" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 696, + 315, + 706 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 696, + 315, + 706 + ], + "spans": [ + { + "bbox": [ + 294, + 696, + 315, + 706 + ], + "type": "text", + "content": "13951" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 116, + 289, + 279 + ], + "blocks": [ + { + "bbox": [ + 85, + 116, + 289, + 279 + ], + "lines": [ + { + "bbox": [ + 85, + 116, + 289, + 279 + ], + "spans": [ + { + "bbox": [ + 85, + 116, + 289, + 279 + ], + "type": "image", + "image_path": "e19aee941cafce9f5e9026208f9f0a8216253559363566a22c879e9481486e16.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 290, + 247, + 300 + ], + "lines": [ + { + "bbox": [ + 126, + 290, + 247, + 300 + ], + "spans": [ + { + "bbox": [ + 126, + 290, + 247, + 300 + ], + "type": "text", + "content": "Figure 4. Selection phase of the SyncMask" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 83, + 315, + 290, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 315, + 290, + 336 + ], + "spans": [ + { + "bbox": [ + 83, + 315, + 290, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{I}\\mathrm{dx}^T" + }, + { + "bbox": [ + 83, + 315, + 290, + 336 + ], + "type": "text", + "content": " for text and " + }, + { + "bbox": [ + 83, + 315, + 290, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{I}\\mathrm{dx}^I" + }, + { + "bbox": [ + 83, + 315, + 290, + 336 + ], + "type": "text", + "content": " for images, as formulated in the equation below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 125, + 342, + 289, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 342, + 289, + 355 + ], + "spans": [ + { + "bbox": [ + 125, + 342, + 289, + 355 + ], + "type": "interline_equation", + "content": "\\mathbf {m} ^ {T} \\left[ \\mathrm {I d} \\mathbf {x} ^ {T} \\right] \\leftarrow 1, \\quad \\mathbf {m} ^ {I} \\left[ \\mathrm {I d} \\mathbf {x} ^ {I} \\right] \\leftarrow 1 \\tag {6}", + "image_path": "254ff7be34409a35ce41190076c39b0957f8a626ebc2241e54577955de23a41c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 360, + 290, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 360, + 290, + 432 + ], + "spans": [ + { + "bbox": [ + 82, + 360, + 290, + 432 + ], + "type": "text", + "content": "Initially, the vectors " + }, + { + "bbox": [ + 82, + 360, + 290, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^T" + }, + { + "bbox": [ + 82, + 360, + 290, + 432 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 82, + 360, + 290, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^I" + }, + { + "bbox": [ + 82, + 360, + 290, + 432 + ], + "type": "text", + "content": " are zero-initialized with dimensions " + }, + { + "bbox": [ + 82, + 360, + 290, + 432 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 82, + 360, + 290, + 432 + ], + "type": "text", + "content": " (for text) and " + }, + { + "bbox": [ + 82, + 360, + 290, + 432 + ], + "type": "inline_equation", + "content": "N'" + }, + { + "bbox": [ + 82, + 360, + 290, + 432 + ], + "type": "text", + "content": " (for images). They are then updated to binary attention masks, which are integral in capturing the synchronized interaction between textual and visual elements, as illustrated in Figure 4. Subsequently, these masks are employed in the masked modeling processes for text and images (detailed in Equation 7 and Equation 8)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 432, + 290, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 432, + 290, + 485 + ], + "spans": [ + { + "bbox": [ + 83, + 432, + 290, + 485 + ], + "type": "text", + "content": "Building upon this foundation, the text mask vector " + }, + { + "bbox": [ + 83, + 432, + 290, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^T" + }, + { + "bbox": [ + 83, + 432, + 290, + 485 + ], + "type": "text", + "content": " consists of elements " + }, + { + "bbox": [ + 83, + 432, + 290, + 485 + ], + "type": "inline_equation", + "content": "[m_1^t,\\dots,m_N^t ]\\in \\{0,1\\} ^N" + }, + { + "bbox": [ + 83, + 432, + 290, + 485 + ], + "type": "text", + "content": " , while the image mask vector " + }, + { + "bbox": [ + 83, + 432, + 290, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^I" + }, + { + "bbox": [ + 83, + 432, + 290, + 485 + ], + "type": "text", + "content": " is composed of " + }, + { + "bbox": [ + 83, + 432, + 290, + 485 + ], + "type": "inline_equation", + "content": "[m_1^i,\\dots,m_{N'}^i ]\\in" + }, + { + "bbox": [ + 83, + 432, + 290, + 485 + ], + "type": "inline_equation", + "content": "\\{0,1\\}^{N^{\\prime}}" + }, + { + "bbox": [ + 83, + 432, + 290, + 485 + ], + "type": "text", + "content": " . For each tokenized text vector, the masked version " + }, + { + "bbox": [ + 83, + 432, + 290, + 485 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{t}}_i" + }, + { + "bbox": [ + 83, + 432, + 290, + 485 + ], + "type": "text", + "content": " is determined by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 491, + 289, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 491, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 129, + 491, + 289, + 504 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {t}} _ {j} = \\left(1 - m _ {j} ^ {t}\\right) \\cdot \\mathbf {t} _ {j} + m _ {j} ^ {t} \\cdot \\mathbf {t} _ {\\text {m a s k}} \\tag {7}", + "image_path": "692a8bca1f75d1e81cab65b5429a421c052ca7ff1c7e7e38e68ce9be6327d726.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 83, + 509, + 289, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 509, + 289, + 551 + ], + "spans": [ + { + "bbox": [ + 83, + 509, + 289, + 551 + ], + "type": "text", + "content": "In this equation, " + }, + { + "bbox": [ + 83, + 509, + 289, + 551 + ], + "type": "inline_equation", + "content": "1 \\leq j \\leq N" + }, + { + "bbox": [ + 83, + 509, + 289, + 551 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 83, + 509, + 289, + 551 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 83, + 509, + 289, + 551 + ], + "type": "text", + "content": " denotes the special token used for textual masking. In the context of Masked Image Modeling (MIM), each image patch is processed with a learnable mask, resulting in the masked image vector " + }, + { + "bbox": [ + 83, + 509, + 289, + 551 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{v}}_k" + }, + { + "bbox": [ + 83, + 509, + 289, + 551 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 556, + 289, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 556, + 289, + 568 + ], + "spans": [ + { + "bbox": [ + 126, + 556, + 289, + 568 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {v}} _ {k} = \\left(1 - m _ {k} ^ {i}\\right) \\cdot \\mathbf {v} _ {k} + m _ {k} ^ {i} \\cdot \\mathbf {v} _ {\\text {m a s k}} \\tag {8}", + "image_path": "71da9a1847600f7bbf256b8fd7990e921cd86c52c31ee72c27836dc8554e1878.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 83, + 574, + 290, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 574, + 290, + 625 + ], + "spans": [ + { + "bbox": [ + 83, + 574, + 290, + 625 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 83, + 574, + 290, + 625 + ], + "type": "inline_equation", + "content": "1 \\leq k \\leq N'" + }, + { + "bbox": [ + 83, + 574, + 290, + 625 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 83, + 574, + 290, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 83, + 574, + 290, + 625 + ], + "type": "text", + "content": " represents the learnable mask embedding [4]. The masked tokenized inputs are thus represented as " + }, + { + "bbox": [ + 83, + 574, + 290, + 625 + ], + "type": "inline_equation", + "content": "\\tilde{T} = [\\mathbf{t}_{\\mathrm{cls}}; \\tilde{\\mathbf{t}}_1; \\ldots; \\tilde{\\mathbf{t}}_N]" + }, + { + "bbox": [ + 83, + 574, + 290, + 625 + ], + "type": "text", + "content": " for text and " + }, + { + "bbox": [ + 83, + 574, + 290, + 625 + ], + "type": "inline_equation", + "content": "\\tilde{V} = [\\mathbf{v}_{\\mathrm{cls}}; \\tilde{\\mathbf{v}}_1; \\ldots; \\tilde{\\mathbf{v}}_{N'}]" + }, + { + "bbox": [ + 83, + 574, + 290, + 625 + ], + "type": "text", + "content": " for image. These masked inputs are processed using the proposed SyncMask " + }, + { + "bbox": [ + 83, + 574, + 290, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^T" + }, + { + "bbox": [ + 83, + 574, + 290, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 83, + 574, + 290, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^I" + }, + { + "bbox": [ + 83, + 574, + 290, + 625 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 83, + 638, + 290, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 638, + 290, + 669 + ], + "spans": [ + { + "bbox": [ + 83, + 638, + 290, + 669 + ], + "type": "text", + "content": "Synchronized Attentional Masked Language Modeling MLM predicts masked words based on the surrounding contextual text and image. Many existing VLMs applied the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 118, + 512, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 118, + 512, + 220 + ], + "spans": [ + { + "bbox": [ + 304, + 118, + 512, + 220 + ], + "type": "text", + "content": "MLM method proposed in BERT [7], randomly masking words with a probability of " + }, + { + "bbox": [ + 304, + 118, + 512, + 220 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 118, + 512, + 220 + ], + "type": "text", + "content": ". However, this approach may not be suitable for vision-language datasets with short caption lengths, especially for nonstandard datasets such as fashion, demanding a thorough understanding of fine-grained attributes. We leverage previous works that addressed these issues by increasing masking probabilities [3] and employing attribute prompts [15]. Building upon these methods, we employ masks generated by SyncMask which is contextually attuned to the corresponding image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 220, + 512, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 512, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 512, + 262 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 220, + 512, + 262 + ], + "type": "inline_equation", + "content": "h(V, \\tilde{T})" + }, + { + "bbox": [ + 304, + 220, + 512, + 262 + ], + "type": "text", + "content": " denote the model's predicted probability for a masked token. " + }, + { + "bbox": [ + 304, + 220, + 512, + 262 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 304, + 220, + 512, + 262 + ], + "type": "text", + "content": " denote a one-hot vocabulary distribution in which the ground-truth token is assigned a probability of 1. MLM minimize cross-entropy loss, described as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 343, + 266, + 511, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 266, + 511, + 280 + ], + "spans": [ + { + "bbox": [ + 343, + 266, + 511, + 280 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M L M}} = \\mathbb {E} _ {(V, \\hat {T}) \\sim D} [ \\operatorname {C E} (\\tilde {\\boldsymbol {y}}, \\boldsymbol {h} (V, \\hat {T}) ], \\tag {9}", + "image_path": "6f1890e34a427575f4453da51347c583fc755a6372a302eb6f08abda0ee5d8c1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 284, + 512, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 512, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 512, + 294 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 284, + 512, + 294 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 284, + 512, + 294 + ], + "type": "text", + "content": " refers the cross-entropy between two vectors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 307, + 513, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 513, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 513, + 389 + ], + "type": "text", + "content": "Synchronized Attentional Masked Image Modeling In the distillation-based MIM [21, 29, 34], a teacher encoder sees the full image, whereas the student encoder, seeing the masked image, tackles the reconstruction objective. Our method adopts a similar objective framework, but with a key difference: our masks, generated through SyncMask, are designed to reflect text-informed elements in masked image. To calculate the MIM loss, the following approach is used:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 399, + 512, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 399, + 512, + 444 + ], + "spans": [ + { + "bbox": [ + 314, + 399, + 512, + 444 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {D i s t} \\left(\\boldsymbol {f} _ {\\theta^ {\\prime}} ^ {I} (V), \\boldsymbol {f} _ {\\theta} ^ {I} (\\tilde {V})\\right) \\\\ = \\frac {1}{\\Omega (\\mathbf {m} ^ {I})} \\sum_ {k = 1} ^ {N ^ {\\prime}} \\mathbf {m} _ {k} ^ {I} \\cdot \\ell_ {1} ^ {S m o o t h} \\left(\\boldsymbol {f} _ {\\theta^ {\\prime}} ^ {I} (V) _ {k}, \\boldsymbol {f} _ {\\theta} ^ {I} (\\tilde {V}) _ {k}\\right) \\tag {10} \\\\ \\end{array}", + "image_path": "e1245f85a3904a7908e258e80e8733e1a9e13de1555b96d69fed1d89ddc952b8.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 452, + 513, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 513, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 513, + 492 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 452, + 513, + 492 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\theta^{\\prime}}^{I}(\\cdot)_{k}" + }, + { + "bbox": [ + 304, + 452, + 513, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 452, + 513, + 492 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\theta}^{I}(\\cdot)_{k}" + }, + { + "bbox": [ + 304, + 452, + 513, + 492 + ], + "type": "text", + "content": " represent the output feature of teacher and student model for the " + }, + { + "bbox": [ + 304, + 452, + 513, + 492 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 452, + 513, + 492 + ], + "type": "text", + "content": "-th image patch, respectively. " + }, + { + "bbox": [ + 304, + 452, + 513, + 492 + ], + "type": "inline_equation", + "content": "\\Omega (\\cdot)" + }, + { + "bbox": [ + 304, + 452, + 513, + 492 + ], + "type": "text", + "content": " means the number of elements with a value of 1 in a vector." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 495, + 512, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 495, + 512, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 495, + 512, + 521 + ], + "type": "interline_equation", + "content": "\\ell_ {1} ^ {\\text {S m o o t h}} (a, b) = \\left\\{ \\begin{array}{l l} 0. 5 (a - b) ^ {2} & \\text {i f} | a - b | < \\gamma \\\\ | a - b | - 0. 5 & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {11}", + "image_path": "459baf595929e05655fbd5d67efb3ac725edfe80596d4bcc6fcf4ac6c050b88d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 529, + 512, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 512, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 512, + 571 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 529, + 512, + 571 + ], + "type": "inline_equation", + "content": "\\ell_1^{Smooth}" + }, + { + "bbox": [ + 304, + 529, + 512, + 571 + ], + "type": "text", + "content": " [11] represents a robust L1 loss less sensitive to outliers than the L2 loss and " + }, + { + "bbox": [ + 304, + 529, + 512, + 571 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 529, + 512, + 571 + ], + "type": "text", + "content": " is a hyperparameter set to 1. Conclusively, the training objective of MIM can be formulated as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 326, + 580, + 512, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 580, + 512, + 594 + ], + "spans": [ + { + "bbox": [ + 326, + 580, + 512, + 594 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M I M}} = \\mathbb {E} _ {(V, \\hat {V}) \\sim D} [ \\operatorname {D i s t} \\left(\\boldsymbol {f} _ {\\theta^ {\\prime}} ^ {I} (V), \\boldsymbol {f} _ {\\theta} ^ {I} (\\tilde {V})\\right) ] \\tag {12}", + "image_path": "f84d8aab80044432f8baa308471a33bb859d2c09a87f493897025488231159c9.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 600, + 408, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 600, + 408, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 600, + 408, + 611 + ], + "type": "text", + "content": "The final loss " + }, + { + "bbox": [ + 306, + 600, + 408, + 611 + ], + "type": "inline_equation", + "content": "(\\mathcal{L})" + }, + { + "bbox": [ + 306, + 600, + 408, + 611 + ], + "type": "text", + "content": " is given as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 347, + 619, + 512, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 619, + 512, + 630 + ], + "spans": [ + { + "bbox": [ + 347, + 619, + 512, + 630 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M I M}} + \\mathcal {L} _ {\\mathrm {M L M}} + \\mathcal {L} _ {\\mathrm {I T C}} + \\mathcal {L} _ {\\mathrm {I T M}} \\tag {13}", + "image_path": "626a1f503586eed1008805989c3b5e3539e2f442d02f57c2ee114bf9aedaf141.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 638, + 512, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 638, + 512, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 638, + 512, + 669 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 638, + 512, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ITC}}" + }, + { + "bbox": [ + 304, + 638, + 512, + 669 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 638, + 512, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ITM}}" + }, + { + "bbox": [ + 304, + 638, + 512, + 669 + ], + "type": "text", + "content": " denote ITC and ITM, respectively. Due to space constraints, detailed formulations of these two losses are elaborated in the Appendix." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 697, + 317, + 706 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 697, + 317, + 706 + ], + "spans": [ + { + "bbox": [ + 294, + 697, + 317, + 706 + ], + "type": "text", + "content": "13952" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 126, + 115, + 470, + 289 + ], + "blocks": [ + { + "bbox": [ + 126, + 115, + 470, + 289 + ], + "lines": [ + { + "bbox": [ + 126, + 115, + 470, + 289 + ], + "spans": [ + { + "bbox": [ + 126, + 115, + 470, + 289 + ], + "type": "table", + "html": "
METHODSI2T*T2I*MEAN*I2TT2IMEAN
R@1R@1R@1R@1
VSE++ [9]4.594.604.60---
VL-BERT [38]19.2622.6320.95---
ViLBERT [32]20.9721.1221.05---
Image-BERT [35]22.7624.7823.77---
OSCAR [28]23.3925.1024.25---
FashionBERT [10]23.9626.7525.36---
Kaleido-BERT [45]27.9933.8830.94---
CommerceMM [43]41.6039.6062.75---
EI-CLIP [33]38.7040.0639.3825.7028.4027.05
ALBEF [27]63.9760.5262.2041.6850.9546.32
FashionViL [13]65.5461.8863.7142.8851.3447.11
FashionSAP [15]73.1470.1271.6354.4362.8258.63
Ours75.0071.0073.0055.3964.0659.73
", + "image_path": "7a5b45d914818f866db2f60af7e33792daedf1a9b9645d491d726f23c84928f4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 292, + 470, + 302 + ], + "lines": [ + { + "bbox": [ + 126, + 292, + 470, + 302 + ], + "spans": [ + { + "bbox": [ + 126, + 292, + 470, + 302 + ], + "type": "text", + "content": "Table 2. Cross-modal retrieval result on FashionGen [37] in the sub/full set of evaluation following previous work. " + }, + { + "bbox": [ + 126, + 292, + 470, + 302 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 126, + 292, + 470, + 302 + ], + "type": "text", + "content": " sub set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 83, + 315, + 271, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 315, + 271, + 326 + ], + "spans": [ + { + "bbox": [ + 83, + 315, + 271, + 326 + ], + "type": "text", + "content": "3.3. Grouped Batch with Semi-hard Negatives" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 331, + 291, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 331, + 291, + 455 + ], + "spans": [ + { + "bbox": [ + 82, + 331, + 291, + 455 + ], + "type": "text", + "content": "In the generic domain, [3] proposed the GRIT strategy for enhancing training effectiveness by forming mini-batches with similar examples. In this strategy, the grouping based on similarity phase plays a crucial role. During this phase, similarity calculations are performed in both directions, utilizing the [CLS] outputs from the unimodal encoders. For each example, the algorithm iteratively identifies the index with the highest similarity, alternating between the image-to-text and text-to-image directions in a sequential manner. These highly similar indices are grouped together within the mini-batch, ensuring that each mini-batch consists of examples exhibiting the highest possible similarity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 456, + 290, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 456, + 290, + 590 + ], + "spans": [ + { + "bbox": [ + 82, + 456, + 290, + 590 + ], + "type": "text", + "content": "However, in fashion datasets, this strategy leads to a false-negative problem in ITM and ITC, where true positives are mislabeled as negatives within mini-batches. To address this problem, we opt to group semi-hard negatives with relatively lower similarity (the " + }, + { + "bbox": [ + 82, + 456, + 290, + 590 + ], + "type": "inline_equation", + "content": "s^{th}" + }, + { + "bbox": [ + 82, + 456, + 290, + 590 + ], + "type": "text", + "content": " highest pairwise similarity) instead of the highest " + }, + { + "bbox": [ + 82, + 456, + 290, + 590 + ], + "type": "inline_equation", + "content": "(1^{st})" + }, + { + "bbox": [ + 82, + 456, + 290, + 590 + ], + "type": "text", + "content": " during the grouping based on similarity phase, where " + }, + { + "bbox": [ + 82, + 456, + 290, + 590 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 82, + 456, + 290, + 590 + ], + "type": "text", + "content": " represents a predefined hyperparameter that is greater than 1. In addition, we prevent true positive samples from grouping by considering the item indices of the samples. Through this approach, the model is trained with negatives that are similar but exhibit meaningful differences, thereby enabling the effective learning of fine-grained distinctions with a limited dataset. More details are in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 602, + 153, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 602, + 153, + 615 + ], + "spans": [ + { + "bbox": [ + 83, + 602, + 153, + 615 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 620, + 197, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 620, + 197, + 632 + ], + "spans": [ + { + "bbox": [ + 83, + 620, + 197, + 632 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 82, + 638, + 290, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 638, + 290, + 669 + ], + "spans": [ + { + "bbox": [ + 82, + 638, + 290, + 669 + ], + "type": "text", + "content": "The foundational architecture of the proposed model is aligned with prior works for demonstrating the effectiveness of the proposed approach [27, 3, 15]. The image encoder" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 315, + 514, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 514, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 514, + 459 + ], + "type": "text", + "content": "adopt the architecture of ViT-B16 [8], whereas the text encoder comprises the first six blocks of the BERT-bas3 [7]. The multi-modal encoder extends the self-attention layers of the last six blocks of BERT with the cross-attention layers. The proposed model is initialized with pre-trained ALBEF [27] same as the Fashion-SAP to ensure a fair comparison [15]. In addition, we employ the same data augmentation and prompt input strategies as FashionSAP [15]. During pretraining, we conduct experiments using 8 RTX 3090 GPUs each with a batch size of 8 for 30 epochs. We adopt a momentum queue size of 48,000 to facilitate grouped batch sampling, which is consistent with GRIT-VLP [3]. The input image size is set to " + }, + { + "bbox": [ + 304, + 315, + 514, + 459 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 315, + 514, + 459 + ], + "type": "text", + "content": ". We apply the AdamW [31] optimizer with a learning rate of 6e-5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 466, + 360, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 360, + 476 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 360, + 476 + ], + "type": "text", + "content": "4.2. Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 482, + 513, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 513, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 513, + 566 + ], + "type": "text", + "content": "FashionGen [37] FashionGen comprises 320K text-image pairs and 40K unique fashion items, each represented by multiple images from different angles. For pre-training, we employ the FashionGen train set, which contains approximately 260.5K text-image pairs. In addition, FashionGen supports various downstream tasks, including text-to-image retrieval, image-to-text retrieval, category recognition, and subcategory recognition." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 572, + 513, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 572, + 513, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 572, + 513, + 624 + ], + "type": "text", + "content": "FashionIQ [40] FashionIQ encompasses 77K unique fashion items and includes 18K training triplets (i.e., query image, modified text, target image) and 6K validation datasets for a text-guided image retrieval task. It contains three different categories: Dress, Toptee, and Shirt." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 632, + 402, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 632, + 402, + 641 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 402, + 641 + ], + "type": "text", + "content": "4.3. Downstream Tasks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 647, + 513, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 513, + 669 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 513, + 669 + ], + "type": "text", + "content": "Cross-modal Retrieval We evaluate a cross-modal retrieval task that includes image-text retrieval (ITR) and text" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 696, + 317, + 706 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 696, + 317, + 706 + ], + "spans": [ + { + "bbox": [ + 294, + 696, + 317, + 706 + ], + "type": "text", + "content": "13953" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 117, + 143, + 157 + ], + "blocks": [ + { + "bbox": [ + 116, + 117, + 143, + 157 + ], + "lines": [ + { + "bbox": [ + 116, + 117, + 143, + 157 + ], + "spans": [ + { + "bbox": [ + 116, + 117, + 143, + 157 + ], + "type": "image", + "image_path": "08d1cb3eaa6aae78202f92b240280ee5c838d122718473275bac5f7f09dbc1f0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 129, + 184, + 148 + ], + "lines": [ + { + "bbox": [ + 144, + 129, + 184, + 148 + ], + "spans": [ + { + "bbox": [ + 144, + 129, + 184, + 148 + ], + "type": "text", + "content": "is light blue in color and has more blue" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 123, + 162, + 140, + 197 + ], + "blocks": [ + { + "bbox": [ + 123, + 162, + 140, + 197 + ], + "lines": [ + { + "bbox": [ + 123, + 162, + 140, + 197 + ], + "spans": [ + { + "bbox": [ + 123, + 162, + 140, + 197 + ], + "type": "image", + "image_path": "23e59536dbf2dded67e37b69ff33c9ae716f43c10a29aa995c31d072cea5f6eb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 166, + 191, + 193 + ], + "lines": [ + { + "bbox": [ + 144, + 166, + 191, + 193 + ], + "spans": [ + { + "bbox": [ + 144, + 166, + 191, + 193 + ], + "type": "text", + "content": "is yellow and has short sleeves and is beige and shorter sleeves" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 117, + 203, + 141, + 236 + ], + "blocks": [ + { + "bbox": [ + 117, + 203, + 141, + 236 + ], + "lines": [ + { + "bbox": [ + 117, + 203, + 141, + 236 + ], + "spans": [ + { + "bbox": [ + 117, + 203, + 141, + 236 + ], + "type": "image", + "image_path": "8170751b04763694d4eb3f1fb6cf964f3bba9ef3b41a6b3fe97abdbf08155f2d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 210, + 183, + 228 + ], + "lines": [ + { + "bbox": [ + 144, + 210, + 183, + 228 + ], + "spans": [ + { + "bbox": [ + 144, + 210, + 183, + 228 + ], + "type": "text", + "content": "is light blue in color and has more blue" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 200, + 118, + 228, + 159 + ], + "blocks": [ + { + "bbox": [ + 200, + 118, + 228, + 159 + ], + "lines": [ + { + "bbox": [ + 200, + 118, + 228, + 159 + ], + "spans": [ + { + "bbox": [ + 200, + 118, + 228, + 159 + ], + "type": "image", + "image_path": "8c31d8ede41a43e06be8211a0ab6e007994f76763cb64cde4e1e17e6dff8bf8d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 233, + 118, + 251, + 158 + ], + "blocks": [ + { + "bbox": [ + 233, + 118, + 251, + 158 + ], + "lines": [ + { + "bbox": [ + 233, + 118, + 251, + 158 + ], + "spans": [ + { + "bbox": [ + 233, + 118, + 251, + 158 + ], + "type": "image", + "image_path": "94e388a82e2ff592d1c2ed228c55692492ed9e48b26c2c48d8ce49541fae0a44.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 256, + 119, + 283, + 158 + ], + "blocks": [ + { + "bbox": [ + 256, + 119, + 283, + 158 + ], + "lines": [ + { + "bbox": [ + 256, + 119, + 283, + 158 + ], + "spans": [ + { + "bbox": [ + 256, + 119, + 283, + 158 + ], + "type": "image", + "image_path": "5bcb4a08e5c54e53f3c8e507e123a9e76ea93cfe83871d4f2a6193b53e97e3a0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 289, + 123, + 307, + 153 + ], + "blocks": [ + { + "bbox": [ + 289, + 123, + 307, + 153 + ], + "lines": [ + { + "bbox": [ + 289, + 123, + 307, + 153 + ], + "spans": [ + { + "bbox": [ + 289, + 123, + 307, + 153 + ], + "type": "image", + "image_path": "a370e0610b30d30aebc21b72d677fe590e829d50020d476a887a20935a21e1ea.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 123, + 336, + 154 + ], + "blocks": [ + { + "bbox": [ + 316, + 123, + 336, + 154 + ], + "lines": [ + { + "bbox": [ + 316, + 123, + 336, + 154 + ], + "spans": [ + { + "bbox": [ + 316, + 123, + 336, + 154 + ], + "type": "image", + "image_path": "624a08638b624bf5cdd6f7fe9b1d7fc67fe107b88de443fc6c06085d9c16a586.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 348, + 118, + 364, + 157 + ], + "blocks": [ + { + "bbox": [ + 348, + 118, + 364, + 157 + ], + "lines": [ + { + "bbox": [ + 348, + 118, + 364, + 157 + ], + "spans": [ + { + "bbox": [ + 348, + 118, + 364, + 157 + ], + "type": "image", + "image_path": "b441dc7c3ee79658ef78c52d3f4f26c07008d9df38600cae675a55bddb212663.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 374, + 120, + 389, + 156 + ], + "blocks": [ + { + "bbox": [ + 374, + 120, + 389, + 156 + ], + "lines": [ + { + "bbox": [ + 374, + 120, + 389, + 156 + ], + "spans": [ + { + "bbox": [ + 374, + 120, + 389, + 156 + ], + "type": "image", + "image_path": "ea28ef8ad5e218aa8874f36d2750a73727f9d914f8a59d4c0b1404fb1907a1e4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 399, + 118, + 423, + 156 + ], + "blocks": [ + { + "bbox": [ + 399, + 118, + 423, + 156 + ], + "lines": [ + { + "bbox": [ + 399, + 118, + 423, + 156 + ], + "spans": [ + { + "bbox": [ + 399, + 118, + 423, + 156 + ], + "type": "image", + "image_path": "318f0906cf0a562e3f0ab193a6e7cbfd9d053ca25270046b1e14cca38f87233a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 428, + 119, + 445, + 156 + ], + "blocks": [ + { + "bbox": [ + 428, + 119, + 445, + 156 + ], + "lines": [ + { + "bbox": [ + 428, + 119, + 445, + 156 + ], + "spans": [ + { + "bbox": [ + 428, + 119, + 445, + 156 + ], + "type": "image", + "image_path": "4e4f78659e3e372f60ac4ebfcd3d872c018fbb1d3b02775f877a33de274ae55e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 457, + 120, + 474, + 156 + ], + "blocks": [ + { + "bbox": [ + 457, + 120, + 474, + 156 + ], + "lines": [ + { + "bbox": [ + 457, + 120, + 474, + 156 + ], + "spans": [ + { + "bbox": [ + 457, + 120, + 474, + 156 + ], + "type": "image", + "image_path": "83d40df87bede9d3578974a9d082cd47a8d877f484bc3e954fc2c69a162e7be6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 206, + 160, + 223, + 200 + ], + "blocks": [ + { + "bbox": [ + 206, + 160, + 223, + 200 + ], + "lines": [ + { + "bbox": [ + 206, + 160, + 223, + 200 + ], + "spans": [ + { + "bbox": [ + 206, + 160, + 223, + 200 + ], + "type": "image", + "image_path": "d8266be9434b5fd28869d4c2b931cfd73e10b0a8bbcc29a28c8081c3ea0476b2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 223, + 161, + 255, + 200 + ], + "blocks": [ + { + "bbox": [ + 223, + 161, + 255, + 200 + ], + "lines": [ + { + "bbox": [ + 223, + 161, + 255, + 200 + ], + "spans": [ + { + "bbox": [ + 223, + 161, + 255, + 200 + ], + "type": "image", + "image_path": "20fe4e6dcf5048f0bc320a9f9665d414fa03e49b246d94907b01027548d45152.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 259, + 161, + 278, + 200 + ], + "blocks": [ + { + "bbox": [ + 259, + 161, + 278, + 200 + ], + "lines": [ + { + "bbox": [ + 259, + 161, + 278, + 200 + ], + "spans": [ + { + "bbox": [ + 259, + 161, + 278, + 200 + ], + "type": "image", + "image_path": "2c15649a5457f0cbd6ad1496a51f15875a5a034642c7bb201bd9976bf3aafdf3.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 289, + 161, + 306, + 200 + ], + "blocks": [ + { + "bbox": [ + 289, + 161, + 306, + 200 + ], + "lines": [ + { + "bbox": [ + 289, + 161, + 306, + 200 + ], + "spans": [ + { + "bbox": [ + 289, + 161, + 306, + 200 + ], + "type": "image", + "image_path": "282e09030c25ecbc8e69764919c3f6d773171f6e66d53e73d5b84c44d08592d9.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 312, + 163, + 337, + 198 + ], + "blocks": [ + { + "bbox": [ + 312, + 163, + 337, + 198 + ], + "lines": [ + { + "bbox": [ + 312, + 163, + 337, + 198 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 337, + 198 + ], + "type": "image", + "image_path": "57acc5e9ccd96d50c72998ae79e895bd874d46fc43ce9151d5eb12b372570bdb.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 342, + 162, + 364, + 198 + ], + "blocks": [ + { + "bbox": [ + 342, + 162, + 364, + 198 + ], + "lines": [ + { + "bbox": [ + 342, + 162, + 364, + 198 + ], + "spans": [ + { + "bbox": [ + 342, + 162, + 364, + 198 + ], + "type": "image", + "image_path": "86dfa22cf839e7351b28ef3ac4c3bfa84481d65078ee0c143c708268ec723bce.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 374, + 161, + 390, + 198 + ], + "blocks": [ + { + "bbox": [ + 374, + 161, + 390, + 198 + ], + "lines": [ + { + "bbox": [ + 374, + 161, + 390, + 198 + ], + "spans": [ + { + "bbox": [ + 374, + 161, + 390, + 198 + ], + "type": "image", + "image_path": "acc07d0ffaf8df3e76cdca01693354eeaa3812b6364c8d49ceb76598e5a189b5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 399, + 160, + 421, + 198 + ], + "blocks": [ + { + "bbox": [ + 399, + 160, + 421, + 198 + ], + "lines": [ + { + "bbox": [ + 399, + 160, + 421, + 198 + ], + "spans": [ + { + "bbox": [ + 399, + 160, + 421, + 198 + ], + "type": "image", + "image_path": "0255ecdf2a25cf15b2f15b9bc4053e116e64d4f910ed649a074361432cf273cf.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 428, + 160, + 446, + 198 + ], + "blocks": [ + { + "bbox": [ + 428, + 160, + 446, + 198 + ], + "lines": [ + { + "bbox": [ + 428, + 160, + 446, + 198 + ], + "spans": [ + { + "bbox": [ + 428, + 160, + 446, + 198 + ], + "type": "image", + "image_path": "1cd8d6951877adcb2f4d504332429f3a65d0fbaac4ade7446264282245d4678f.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 457, + 161, + 475, + 198 + ], + "blocks": [ + { + "bbox": [ + 457, + 161, + 475, + 198 + ], + "lines": [ + { + "bbox": [ + 457, + 161, + 475, + 198 + ], + "spans": [ + { + "bbox": [ + 457, + 161, + 475, + 198 + ], + "type": "image", + "image_path": "ec91c71b5b9074869a80b19f3bc33ae407edfa3c5270b3dad56a50bfad5be857.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 200, + 201, + 228, + 238 + ], + "blocks": [ + { + "bbox": [ + 200, + 201, + 228, + 238 + ], + "lines": [ + { + "bbox": [ + 200, + 201, + 228, + 238 + ], + "spans": [ + { + "bbox": [ + 200, + 201, + 228, + 238 + ], + "type": "image", + "image_path": "1baab022be951c854d8b399acd73fc441aaf27c07c3a2bde1e497970d08f80cd.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 229, + 202, + 251, + 238 + ], + "blocks": [ + { + "bbox": [ + 229, + 202, + 251, + 238 + ], + "lines": [ + { + "bbox": [ + 229, + 202, + 251, + 238 + ], + "spans": [ + { + "bbox": [ + 229, + 202, + 251, + 238 + ], + "type": "image", + "image_path": "6af5f3a7e30792093bf7b4762ab058ca532e094b8cd9ce891869b0d0b9ddd038.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 256, + 201, + 283, + 238 + ], + "blocks": [ + { + "bbox": [ + 256, + 201, + 283, + 238 + ], + "lines": [ + { + "bbox": [ + 256, + 201, + 283, + 238 + ], + "spans": [ + { + "bbox": [ + 256, + 201, + 283, + 238 + ], + "type": "image", + "image_path": "0a16f39e1dfd022b3e9a9f2b58e7ce60f75514a3b3a7a46aafa05eb061129530.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 289, + 206, + 307, + 236 + ], + "blocks": [ + { + "bbox": [ + 289, + 206, + 307, + 236 + ], + "lines": [ + { + "bbox": [ + 289, + 206, + 307, + 236 + ], + "spans": [ + { + "bbox": [ + 289, + 206, + 307, + 236 + ], + "type": "image", + "image_path": "6f6d12ec51ad75be41402fc4fc4d7fff212048e7b508ca136e94c60c66adb3aa.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 311, + 201, + 340, + 239 + ], + "blocks": [ + { + "bbox": [ + 311, + 201, + 340, + 239 + ], + "lines": [ + { + "bbox": [ + 311, + 201, + 340, + 239 + ], + "spans": [ + { + "bbox": [ + 311, + 201, + 340, + 239 + ], + "type": "image", + "image_path": "8e7e624a202a285f28ed5079046737346ea5fe036c71c57602fc3a8ca5eb53f5.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 247, + 512, + 273 + ], + "lines": [ + { + "bbox": [ + 82, + 247, + 512, + 273 + ], + "spans": [ + { + "bbox": [ + 82, + 247, + 512, + 273 + ], + "type": "text", + "content": "Figure 5. The top-10 TGIR results of the SyncMask model on the FashionIQ dataset. On the left, the reference images paired with their guided descriptions are shown, while the right side presents the model's predicted images ranked by descending scores. Ground truth images are distinctly outlined with a green bounding box. It is worth mentioning that the set of predictions includes other images that also qualify as suitable matches." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 342, + 206, + 367, + 237 + ], + "blocks": [ + { + "bbox": [ + 342, + 206, + 367, + 237 + ], + "lines": [ + { + "bbox": [ + 342, + 206, + 367, + 237 + ], + "spans": [ + { + "bbox": [ + 342, + 206, + 367, + 237 + ], + "type": "image", + "image_path": "9c1dc7a10fe31403f843fb4318d3da5750bdaf9375a3cfcdb912c16598fcd73b.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 368, + 201, + 395, + 239 + ], + "blocks": [ + { + "bbox": [ + 368, + 201, + 395, + 239 + ], + "lines": [ + { + "bbox": [ + 368, + 201, + 395, + 239 + ], + "spans": [ + { + "bbox": [ + 368, + 201, + 395, + 239 + ], + "type": "image", + "image_path": "f4b841bad0e88fd318f474c0906466606356288703f40b4925a13a89294664e7.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 410, + 201, + 425, + 239 + ], + "blocks": [ + { + "bbox": [ + 410, + 201, + 425, + 239 + ], + "lines": [ + { + "bbox": [ + 410, + 201, + 425, + 239 + ], + "spans": [ + { + "bbox": [ + 410, + 201, + 425, + 239 + ], + "type": "image", + "image_path": "572004d130ae93c223711f78f570eae81443ae4042a9188c228a01f1ba518f49.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 426, + 201, + 449, + 239 + ], + "blocks": [ + { + "bbox": [ + 426, + 201, + 449, + 239 + ], + "lines": [ + { + "bbox": [ + 426, + 201, + 449, + 239 + ], + "spans": [ + { + "bbox": [ + 426, + 201, + 449, + 239 + ], + "type": "image", + "image_path": "3c38f01af8724c542738866a5529290a45d23d75f1df949fe4de0c19dab3d5c4.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 452, + 202, + 480, + 237 + ], + "blocks": [ + { + "bbox": [ + 452, + 202, + 480, + 237 + ], + "lines": [ + { + "bbox": [ + 452, + 202, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 452, + 202, + 480, + 237 + ], + "type": "image", + "image_path": "8942212fffe07cc177c1440480bb860a806223e831294f54e57d47ec97fb1486.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "table", + "bbox": [ + 100, + 281, + 493, + 403 + ], + "blocks": [ + { + "bbox": [ + 100, + 281, + 493, + 403 + ], + "lines": [ + { + "bbox": [ + 100, + 281, + 493, + 403 + ], + "spans": [ + { + "bbox": [ + 100, + 281, + 493, + 403 + ], + "type": "table", + "html": "
METHODSDRESSTOPTEESHIRTMEAN
R@10R@50R@10R@50R@10R@50R@10R@50
CIRR [30]17.4540.4121.6445.3817.5338.8118.8741.53
VAL [5]22.5344.0027.5351.6822.3844.1524.1546.61
CosMo [25]25.6450.3029.2157.4624.9049.1826.5852.31
DCNet [22]28.9556.730.4458.2923.9547.327.7854.10
FashionVLP [12]32.4260.2938.5168.7931.8958.4434.2762.51
FashionViL [13]33.4759.9434.9860.7925.1750.3931.2157.04
FashionSAP [15]33.7160.4341.9170.9333.1761.3336.2664.23
Ours33.7661.2344.8272.0635.8262.1238.1365.14
", + "image_path": "17c2bf65f4e306b1735e90c7b078686e2142be17114d8c475cedaaa93c6b9918.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "table_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 82, + 428, + 291, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 428, + 291, + 541 + ], + "spans": [ + { + "bbox": [ + 82, + 428, + 291, + 541 + ], + "type": "text", + "content": "image retrieval (TIR). Further, ITR focuses on finding relevant textual descriptions for a given image query. TIR, the inverse task, retrieves pertinent images based on a textual query. These tasks assess the effectiveness of the model in capturing cross-modal relationships between text and images within retrieval scenarios. Following the previous works [13, 15], we evaluate cross-modal retrieval not only on the subset with 1K retrievals but also on the full dataset of FashionGen [37]. The results, including R@1 scores for both subset and the full set, are presented in Table Table 2, demonstrating an improved performance compared to that of the previous results." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 82, + 555, + 291, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 555, + 291, + 669 + ], + "spans": [ + { + "bbox": [ + 82, + 555, + 291, + 669 + ], + "type": "text", + "content": "Text-guided Image Retrieval This task aims to retrieve target images by considering query pairs that reference image and modified descriptions; this is more challenging than traditional retrievals. Therefore, we need to select the target image for identifying minor differences in the changes in description while maintaining the characteristics of the reference image. For a fair comparison, we adopt a similar fine-tuning as outlined in the FashionSAP [15]. In the actual dataset, there are many images that match with the query pairs in addition to the target image referred to as the real correct answer, as shown in Figure 5. Thus, a qualitative" + } + ] + } + ], + "index": 40 + }, + { + "type": "table", + "bbox": [ + 321, + 426, + 495, + 494 + ], + "blocks": [ + { + "bbox": [ + 203, + 405, + 392, + 415 + ], + "lines": [ + { + "bbox": [ + 203, + 405, + 392, + 415 + ], + "spans": [ + { + "bbox": [ + 203, + 405, + 392, + 415 + ], + "type": "text", + "content": "Table 3. Text-guided image retrieval performance in FashionIQ [40]" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 426, + 495, + 494 + ], + "lines": [ + { + "bbox": [ + 321, + 426, + 495, + 494 + ], + "spans": [ + { + "bbox": [ + 321, + 426, + 495, + 494 + ], + "type": "table", + "html": "
METHODSCRSCR
AccMacro-FAccMacro-F
F-BERT [10]91.2570.5085.2762.00
K-BERT [45]95.0771.4088.0763.60
F-ViL [13]97.4888.6092.2383.02
FashionSAP [15]98.3489.8494.3387.67
Ours98.4190.3194.2187.83
", + "image_path": "cc2b7a59bd5387cc75c8fed1b4a1268b1506881d0c301f8b750edaffecdcfd79.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "table_body" + } + ], + "index": 41 + }, + { + "bbox": [ + 339, + 495, + 479, + 503 + ], + "lines": [ + { + "bbox": [ + 339, + 495, + 479, + 503 + ], + "spans": [ + { + "bbox": [ + 339, + 495, + 479, + 503 + ], + "type": "text", + "content": "Table 4. CR and SCR results on FashionGen [37]." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 525, + 512, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 525, + 512, + 566 + ], + "spans": [ + { + "bbox": [ + 305, + 525, + 512, + 566 + ], + "type": "text", + "content": "evaluation can be conducted in that the model finds similar images well in addition to the actual correct answer. As shown in Table 3, the proposed model surpasses previous models and demonstrates state-of-the-art performance." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 304, + 586, + 512, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 512, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 512, + 669 + ], + "type": "text", + "content": "Category/Subcategory Recognition In this downstream task, the objective is to classify the category and subcategory of fashion items using the textual and visual information provided. In line with earlier studies [10, 45, 13, 15], we simply attach a linear layer to the [CLS] token, which serves as the fusion feature, for task label prediction. As indicated in Table 4, our proposed model exhibits competitive performance compared to existing models." + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 697, + 317, + 706 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 697, + 317, + 706 + ], + "spans": [ + { + "bbox": [ + 294, + 697, + 317, + 706 + ], + "type": "text", + "content": "13954" + } + ] + } + ], + "index": 45 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 115, + 282, + 209 + ], + "blocks": [ + { + "bbox": [ + 88, + 115, + 282, + 209 + ], + "lines": [ + { + "bbox": [ + 88, + 115, + 282, + 209 + ], + "spans": [ + { + "bbox": [ + 88, + 115, + 282, + 209 + ], + "type": "image", + "image_path": "b38fe64ecf743820b70109df9a7c0010bce8a8afd65f64ac4c0ada4d69883d6d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 88, + 214, + 282, + 310 + ], + "blocks": [ + { + "bbox": [ + 88, + 214, + 282, + 310 + ], + "lines": [ + { + "bbox": [ + 88, + 214, + 282, + 310 + ], + "spans": [ + { + "bbox": [ + 88, + 214, + 282, + 310 + ], + "type": "image", + "image_path": "83a69f65308f388961b7250158c55abbf440375d89af6fc9e0c4c1ee3b7be5d9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 95, + 316, + 141, + 323 + ], + "lines": [ + { + "bbox": [ + 95, + 316, + 141, + 323 + ], + "spans": [ + { + "bbox": [ + 95, + 316, + 141, + 323 + ], + "type": "text", + "content": "(a) Unmasked input" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 166, + 316, + 208, + 323 + ], + "lines": [ + { + "bbox": [ + 166, + 316, + 208, + 323 + ], + "spans": [ + { + "bbox": [ + 166, + 316, + 208, + 323 + ], + "type": "text", + "content": "(b) Random mask" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 235, + 316, + 266, + 323 + ], + "lines": [ + { + "bbox": [ + 235, + 316, + 266, + 323 + ], + "spans": [ + { + "bbox": [ + 235, + 316, + 266, + 323 + ], + "type": "text", + "content": "(c) SyncMask" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 83, + 324, + 290, + 349 + ], + "lines": [ + { + "bbox": [ + 83, + 324, + 290, + 349 + ], + "spans": [ + { + "bbox": [ + 83, + 324, + 290, + 349 + ], + "type": "text", + "content": "Figure 6. Examples of random masking (b) and SyncMask (c) for MIM and MLM. The latter applies masks to pertinent features in both the input image and text (a), offering a more context-sensitive selection than the former." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 83, + 368, + 164, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 368, + 164, + 380 + ], + "spans": [ + { + "bbox": [ + 83, + 368, + 164, + 380 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 82, + 384, + 291, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 384, + 291, + 591 + ], + "spans": [ + { + "bbox": [ + 82, + 384, + 291, + 591 + ], + "type": "text", + "content": "Random vs. Attentional Masking We ablate the SyncMask with other non-synchronized masking methods. Table 5 lists the comparison results for the combination of random and attentional masking. In this experiment, we evaluated five experiments for I2T (R@1), T2I (R@1), CR (Macro-F), SCR (Macro-F), and TMIR (R@10), which were evaluated in the previous experiments. In these experiments, the results of the three experiments with at least one attentional masking were higher than those with no attentional masking (all random). This can be considered an indication of the efficacy of attentional masking. Further, the synergistic performance of the proposed synchronized textual-visual attentional masking (+2.14) is greater than that of only one attentional masking (+0.4, +0.66), which shows the superiority of the SyncMask approach. Figure 6 displays the distinction between applying random masks (b) and SyncMask (c) to unmasked image-text pairs (a) from the FashionGen [37]. SyncMask strategically places masks on image patches that correlate with the text and applies masks to the text informed by the image details, unlike random masking." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 83, + 606, + 290, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 606, + 290, + 669 + ], + "spans": [ + { + "bbox": [ + 83, + 606, + 290, + 669 + ], + "type": "text", + "content": "Random vs. Hardest vs. Semi-hard Negative Sampling Table 6 compare the effectiveness of various grouped minibatch sampling strategies across five downstream tasks. The methods evaluated encompassed four scenarios: random grouping, hardest grouping (+1.90), hardest grouping with exclusion of false negatives (+1.98), and semi-hard grouping" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 306, + 116, + 511, + 167 + ], + "blocks": [ + { + "bbox": [ + 306, + 116, + 511, + 167 + ], + "lines": [ + { + "bbox": [ + 306, + 116, + 511, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 116, + 511, + 167 + ], + "type": "table", + "html": "
RTRVATAVMEANGAINI2T R@1T21 R@1CR Macro-FSCR Macro-FTMIR R@10
✓✓70.3173.3069.8086.2186.1636.08
✓✓70.71+0.4074.8070.1086.4986.3135.84
✓✓70.97+0.6674.3070.8085.1286.7837.87
✓✓72.45+2.1475.0071.0090.3187.8338.13
", + "image_path": "2c90a8c5606be799d997609ca0f0e498d2927d3ffb33bb9febcb27f2881f560e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 306, + 204, + 511, + 255 + ], + "blocks": [ + { + "bbox": [ + 305, + 169, + 512, + 195 + ], + "lines": [ + { + "bbox": [ + 305, + 169, + 512, + 195 + ], + "spans": [ + { + "bbox": [ + 305, + 169, + 512, + 195 + ], + "type": "text", + "content": "Table 5. Ablation study results for Random vs. Attention Masked Modeling on five downstream tasks. Rt:Random Textual masking. Rv:Random Visual masking. At:Attentional Textual masking. Av:Attentional Visual masking." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 204, + 511, + 255 + ], + "lines": [ + { + "bbox": [ + 306, + 204, + 511, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 204, + 511, + 255 + ], + "type": "table", + "html": "
GROUPEFNMEANGAIN12T R@1T2I R@1CR Macro-FSCR Macro-FTMIR R@10
Random69.9972.3069.0086.2185.4037.06
Hardest71.89+1.9074.7070.5090.2086.6437.40
Hardest✓71.97+1.9874.3071.0090.1087.1537.28
Semi-hard✓72.45+2.4675.0071.0090.3187.8338.13
", + "image_path": "76f6f20e9b568ae8c846fbd50bc360737dabe47e89ee0996f64ca1d62361fcf3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 257, + 512, + 283 + ], + "lines": [ + { + "bbox": [ + 305, + 257, + 512, + 283 + ], + "spans": [ + { + "bbox": [ + 305, + 257, + 512, + 283 + ], + "type": "text", + "content": "Table 6. Ablation study results comparing the Grouped Batch Sampling Strategy across five downstream tasks. GROUP: Strategy for grouping phase of GRIT. EFN: Exclude False Negative in a grouping phase using index." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 300, + 513, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 513, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 513, + 434 + ], + "type": "text", + "content": "also excluding false negatives (+2.46). These findings suggest that grouping similar samples in a mini-batch is more beneficial for learning than composing batches with random samples. However, given that the fashion dataset often has multiple captions per image or vice versa, performance gains were observed when systematically preventing the grouping of false negatives by using indexing. Nevertheless, due to the prevalence of inherently similar samples that cannot be systematically excluded, we opted for grouping semi-hard negatives instead of the hardest ones, which resulted in a significant performance boost. This highlights the importance of further research from a data perspective, not just in terms of model architecture or loss functions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 443, + 369, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 369, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 369, + 453 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 460, + 513, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 513, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 513, + 594 + ], + "type": "text", + "content": "We introduced Synchronized attentional Masking for enhanced masked modeling in fashion-centric VLMs. Leveraging cross-attention features of a momentum model, our method tailors the random mask into a targeted mask for synchronously co-occurring segments in image-text pairs in MLM and MIM objectives. This approach effectively resolves misaligned image-text input issues and improving fine-grained cross-modal representation. Additionally, we addressed data scarcity and distribution challenges in fashion datasets, refining grouped batch sampling with semi-hard negatives for ITM and ITC losses. The experimental results showed our methods outperformed established benchmarks in multiple downstream tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 603, + 386, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 603, + 386, + 615 + ], + "spans": [ + { + "bbox": [ + 306, + 603, + 386, + 615 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 616, + 512, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 512, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 512, + 669 + ], + "type": "text", + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) under the metaverse support program to nurture the best talents (IITP-2024-RS-2023-00254529) grant funded by the Korea government (MSIT)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 697, + 317, + 706 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 697, + 317, + 706 + ], + "spans": [ + { + "bbox": [ + 294, + 697, + 317, + 706 + ], + "type": "text", + "content": "13955" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 84, + 116, + 134, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 116, + 134, + 126 + ], + "spans": [ + { + "bbox": [ + 84, + 116, + 134, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 85, + 133, + 290, + 668 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 88, + 133, + 290, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 133, + 290, + 180 + ], + "spans": [ + { + "bbox": [ + 88, + 133, + 290, + 180 + ], + "type": "text", + "content": "[1] Yutong Bai, Zeyu Wang, Junfei Xiao, Chen Wei, Huiyu Wang, Alan L Yuille, Yuyin Zhou, and Cihang Xie. Masked autoencoders enable efficient knowledge distillers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24256-24265, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 88, + 182, + 290, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 182, + 290, + 200 + ], + "spans": [ + { + "bbox": [ + 88, + 182, + 290, + 200 + ], + "type": "text", + "content": "[2] Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT pretraining of image transformers. In ICLR, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 201, + 289, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 201, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 88, + 201, + 289, + 239 + ], + "type": "text", + "content": "[3] Jaeseok Byun, Taebaek Hwang, Jianlong Fu, and Taesup Moon. Grit-vlp: Grouped mini-batch sampling for efficient vision and language pre-training. In European Conference on Computer Vision, pages 395-412. Springer, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 239, + 290, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 239, + 290, + 276 + ], + "spans": [ + { + "bbox": [ + 89, + 239, + 290, + 276 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 277, + 290, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 277, + 290, + 315 + ], + "spans": [ + { + "bbox": [ + 88, + 277, + 290, + 315 + ], + "type": "text", + "content": "[5] Yanbei Chen, Shaogang Gong, and Loris Bazzani. Image search with text feedback by visiolinguistic attention learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3001-3011, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 316, + 290, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 316, + 290, + 361 + ], + "spans": [ + { + "bbox": [ + 89, + 316, + 290, + 361 + ], + "type": "text", + "content": "[6] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 364, + 290, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 364, + 290, + 428 + ], + "spans": [ + { + "bbox": [ + 89, + 364, + 290, + 428 + ], + "type": "text", + "content": "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 430, + 290, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 430, + 290, + 485 + ], + "spans": [ + { + "bbox": [ + 89, + 430, + 290, + 485 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 487, + 289, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 487, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 89, + 487, + 289, + 515 + ], + "type": "text", + "content": "[9] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 85, + 516, + 289, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 516, + 289, + 571 + ], + "spans": [ + { + "bbox": [ + 85, + 516, + 289, + 571 + ], + "type": "text", + "content": "[10] Dehong Gao, Linbo Jin, Ben Chen, Minghui Qiu, Peng Li, Yi Wei, Yi Hu, and Hao Wang. Fashionbert: Text and image matching with adaptive loss for cross-modal retrieval. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2251-2260, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 85, + 573, + 241, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 573, + 241, + 582 + ], + "spans": [ + { + "bbox": [ + 85, + 573, + 241, + 582 + ], + "type": "text", + "content": "[11] Ross Girshick. Fast R-CNN. In ICCV, 2015." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 85, + 583, + 290, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 583, + 290, + 639 + ], + "spans": [ + { + "bbox": [ + 85, + 583, + 290, + 639 + ], + "type": "text", + "content": "[12] Sonam Goenka, Zhaoheng Zheng, Ayush Jaiswal, Rakesh Chada, Yue Wu, Varsha Hedau, and Pradeep Natarajan. Fashionvlp: Vision language transformer for fashion retrieval with feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14105-14115, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 85, + 640, + 290, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 640, + 290, + 668 + ], + "spans": [ + { + "bbox": [ + 85, + 640, + 290, + 668 + ], + "type": "text", + "content": "[13] Xiao Han, Licheng Yu, Xiatian Zhu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Fashionvil: Fashion-focused vision-and-language representation learning. In ECCV. Springer, 2022." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 118, + 512, + 668 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 118, + 512, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 512, + 165 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 512, + 165 + ], + "type": "text", + "content": "[14] Xiao Han, Xiatian Zhu, Licheng Yu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Fame-vil: Multi-tasking vision-language model for heterogeneous fashion tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2669-2680, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 167, + 512, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 167, + 512, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 167, + 512, + 205 + ], + "type": "text", + "content": "[15] Yunpeng Han, Lisai Zhang, Qingcai Chen, Zhijian Chen, Zhonghua Li, Jianxin Yang, and Zhao Cao. Fashionsap: Symbols and attributes prompt for fine-grained fashion vision-language pre-training. In CVPR, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 206, + 512, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 512, + 233 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 512, + 233 + ], + "type": "text", + "content": "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In arXiv:2111.06377, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 235, + 512, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 235, + 512, + 280 + ], + "spans": [ + { + "bbox": [ + 307, + 235, + 512, + 280 + ], + "type": "text", + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 282, + 512, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 512, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 512, + 330 + ], + "type": "text", + "content": "[18] Zhicheng Huang, Zhaoyang Zeng, Yupan Huang, Bei Liu, Dongmei Fu, and Jianlong Fu. Seeing out of the box: End-to-end pre-training for vision-language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12976-12985, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 331, + 512, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 512, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 512, + 358 + ], + "type": "text", + "content": "[19] Zhicheng Huang, Zhaoyang Zeng, Bei Liu, Dongmei Fu, and Jianlong Fu. Pixel-bert: Aligning image pixels with text by deep multi-modal transformers. 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 360, + 512, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 512, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 512, + 415 + ], + "type": "text", + "content": "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 418, + 512, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 418, + 512, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 512, + 455 + ], + "type": "text", + "content": "[21] Ioannis Kakogeorgiou, Spyros Gidaris, Bill Psomas, Yannis Avrithis, Andrei Bursuc, Konstantinos Karantzalos, and Nikos Komodakis. What to hide from your students: Attention-guided masked image modeling. In ECCV, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 456, + 512, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 512, + 484 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 512, + 484 + ], + "type": "text", + "content": "[22] Jongseok Kim, Youngjae Yu, Hoeseong Kim, and Gunhee Kim. Dual compositional learning in interactive image retrieval. In AAAI, pages 1771-1779, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 486, + 512, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 486, + 512, + 522 + ], + "spans": [ + { + "bbox": [ + 307, + 486, + 512, + 522 + ], + "type": "text", + "content": "[23] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 524, + 512, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 524, + 512, + 561 + ], + "spans": [ + { + "bbox": [ + 307, + 524, + 512, + 561 + ], + "type": "text", + "content": "[24] Gukyeong Kwon, Zhaowei Cai, Avinash Ravichandran, Erhan Bas, Rahul Bhotika, and Stefano Soatto. Masked vision and language modeling for multi-modal representation learning. In ICLR, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 563, + 512, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 563, + 512, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 563, + 512, + 601 + ], + "type": "text", + "content": "[25] Seungmin Lee, Dongwan Kim, and Bohyung Han. Cosmo: Content-style modulation for image retrieval with text feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 802-812, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 602, + 512, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 512, + 638 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 512, + 638 + ], + "type": "text", + "content": "[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 639, + 512, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 639, + 512, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 639, + 512, + 668 + ], + "type": "text", + "content": "[27] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 697, + 317, + 706 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 697, + 317, + 706 + ], + "spans": [ + { + "bbox": [ + 294, + 697, + 317, + 706 + ], + "type": "text", + "content": "13956" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 84, + 118, + 290, + 667 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 100, + 118, + 290, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 118, + 290, + 137 + ], + "spans": [ + { + "bbox": [ + 100, + 118, + 290, + 137 + ], + "type": "text", + "content": "momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 85, + 138, + 290, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 138, + 290, + 185 + ], + "spans": [ + { + "bbox": [ + 85, + 138, + 290, + 185 + ], + "type": "text", + "content": "[28] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 84, + 186, + 290, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 186, + 290, + 224 + ], + "spans": [ + { + "bbox": [ + 84, + 186, + 290, + 224 + ], + "type": "text", + "content": "[29] Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, et al. MST: Masked self-supervised transformer for visual representation. In NeurIPS, volume 34, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 225, + 290, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 225, + 290, + 271 + ], + "spans": [ + { + "bbox": [ + 85, + 225, + 290, + 271 + ], + "type": "text", + "content": "[30] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2125-2134, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 274, + 289, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 274, + 289, + 292 + ], + "spans": [ + { + "bbox": [ + 85, + 274, + 289, + 292 + ], + "type": "text", + "content": "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 293, + 290, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 293, + 290, + 330 + ], + "spans": [ + { + "bbox": [ + 85, + 293, + 290, + 330 + ], + "type": "text", + "content": "[32] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 332, + 290, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 332, + 290, + 388 + ], + "spans": [ + { + "bbox": [ + 85, + 332, + 290, + 388 + ], + "type": "text", + "content": "[33] Haoyu Ma, Handong Zhao, Zhe Lin, Ajinkya Kale, Zhangyang Wang, Tong Yu, Jiumiang Gu, Sunav Choudhary, and Xiaohui Xie. Ei-clip: Entity-aware interventional contrastive learning for e-commerce cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18051-18061, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 85, + 389, + 289, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 389, + 289, + 417 + ], + "spans": [ + { + "bbox": [ + 85, + 389, + 289, + 417 + ], + "type": "text", + "content": "[34] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. A unified view of masked image modeling. In arXiv preprint arXiv:2210.10615, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 85, + 418, + 289, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 418, + 289, + 455 + ], + "spans": [ + { + "bbox": [ + 85, + 418, + 289, + 455 + ], + "type": "text", + "content": "[35] Di Qi, Lin Su, Jia Song, Edward Cui, Taroon Bharti, and Arun Sacheti. Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 85, + 457, + 290, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 457, + 290, + 513 + ], + "spans": [ + { + "bbox": [ + 85, + 457, + 290, + 513 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 85, + 515, + 290, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 515, + 290, + 552 + ], + "spans": [ + { + "bbox": [ + 85, + 515, + 290, + 552 + ], + "type": "text", + "content": "[37] Negar Rostamzadeh, Seyedarian Hosseini, Thomas Boquet, Wojciech Stokowiec, Ying Zhang, Christian Jauvin, and Chris Pal. Fashion-gen: The generative fashion dataset and challenge. arXiv preprint arXiv:1806.08317, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 85, + 554, + 290, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 554, + 290, + 590 + ], + "spans": [ + { + "bbox": [ + 85, + 554, + 290, + 590 + ], + "type": "text", + "content": "[38] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. In International Conference on Learning Representations, 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 85, + 592, + 289, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 592, + 289, + 610 + ], + "spans": [ + { + "bbox": [ + 85, + 592, + 289, + 610 + ], + "type": "text", + "content": "[39] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 85, + 612, + 290, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 612, + 290, + 667 + ], + "spans": [ + { + "bbox": [ + 85, + 612, + 290, + 667 + ], + "type": "text", + "content": "[40] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11307-11317, 2021." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 118, + 512, + 366 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 118, + 512, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 512, + 165 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 512, + 165 + ], + "type": "text", + "content": "[41] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653-9663, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 167, + 512, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 167, + 512, + 204 + ], + "spans": [ + { + "bbox": [ + 307, + 167, + 512, + 204 + ], + "type": "text", + "content": "[42] Hongwei Xue, Yupan Huang, Bei Liu, Houwen Peng, Jianlong Fu, Houqiang Li, and Jiebo Luo. Probing inter-modality: Visual parsing with self-attention for vision-and-language pre-training. volume 34, pages 4514-4528, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 205, + 512, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 205, + 512, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 205, + 512, + 259 + ], + "type": "text", + "content": "[43] Licheng Yu, Jun Chen, Animesh Sinha, Mengjiao Wang, Yu Chen, Tamara L Berg, and Ning Zhang. Commercemm: Large-scale commerce multimodal representation learning with omni retrieval. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4433-4442, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 262, + 512, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 512, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 512, + 319 + ], + "type": "text", + "content": "[44] Zijia Zhao, Longteng Guo, Xingjian He, Shuai Shao, Zehuan Yuan, and Jing Liu. Mamo: Fine-grained vision-language representations learning with masked multimodal modeling. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1528-1538, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 320, + 512, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 512, + 366 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 512, + 366 + ], + "type": "text", + "content": "[45] Mingchen Zhuge, Dehong Gao, Deng-Ping Fan, Linbo Jin, Ben Chen, Haoming Zhou, Minghui Qiu, and Ling Shao. Kaleido-bert: Vision-language pre-training on fashion domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12647–12657, 2021." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 697, + 316, + 705 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 697, + 316, + 705 + ], + "spans": [ + { + "bbox": [ + 294, + 697, + 316, + 705 + ], + "type": "text", + "content": "13957" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/89645a31-8829-4d8f-939e-0d17fcee95a5_content_list.json b/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/89645a31-8829-4d8f-939e-0d17fcee95a5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c05f2c462fff7fc91ab0eec6f59c2a8491b8da5e --- /dev/null +++ b/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/89645a31-8829-4d8f-939e-0d17fcee95a5_content_list.json @@ -0,0 +1,1572 @@ +[ + { + "type": "text", + "text": "SyncTalk: The Devil is in the Synchronization for Talking Head Synthesis", + "text_level": 1, + "bbox": [ + 112, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziqiao Peng $^{1}$ Wentao Hu $^{2}$ Yue Shi $^{3}$ Xiangyu Zhu $^{4}$ Xiaomei Zhang $^{4}$ Hao Zhao $^{5}$ Jun He $^{1}$ Hongyan Liu $^{5*}$ Zhaoxin Fan $^{1*}$", + "bbox": [ + 129, + 178, + 839, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Renmin University of China $^{2}$ Beijing University of Posts and Telecommunications $^{3}$ Psyche AI Inc. $^{4}$ Chinese Academy of Sciences $^{5}$ Tsinghua University", + "bbox": [ + 148, + 215, + 821, + 253 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/785227c2823cf3d039b62977e4be371e69d0f11158259451da2e2a7fca027ef6.jpg", + "image_caption": [ + "Figure 1. The proposed SyncTalk synthesizes synchronized talking head videos, employing tri-plane hash representations to maintain subject identity. It can generate synchronized lip movements, facial expressions, and stable head poses, and restores hair details to create high-resolution videos." + ], + "image_footnote": [], + "bbox": [ + 94, + 268, + 872, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 589, + 313, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Achieving high synchronization in the synthesis of realistic, speech-driven talking head videos presents a significant challenge. Traditional Generative Adversarial Networks (GAN) struggle to maintain consistent facial identity, while Neural Radiance Fields (NeRF) methods, although they can address this issue, often produce mismatched lip movements, inadequate facial expressions, and unstable head poses. A lifelike talking head requires synchronized coordination of subject identity, lip movements, facial expressions, and head poses. The absence of these synchronizations is a fundamental flaw, leading to unrealistic and artificial outcomes. To address the critical issue of synchronization, identified as the \"devil\" in creating realistic talking heads, we introduce SyncTalk. This NeRF-based method effectively maintains subject identity, enhancing synchronization and realism in talking head synthesis. SyncTalk employs a Face-Sync Controller to align lip movements", + "bbox": [ + 75, + 621, + 473, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "with speech and innovatively uses a 3D facial blendshape model to capture accurate facial expressions. Our Head-Sync Stabilizer optimizes head poses, achieving more natural head movements. The Portrait-Sync Generator restores hair details and blends the generated head with the torso for a seamless visual experience. Extensive experiments and user studies demonstrate that SyncTalk outperforms state-of-the-art methods in synchronization and realism. We recommend watching the supplementary video: https://ziqiaopeng.github.io/synctalk", + "bbox": [ + 500, + 590, + 890, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 768, + 632, + 786 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The quest to generate dynamic and realistic speech-driven talking heads has intensified, driven by expanding applications in digital assistants [39], virtual reality [32], and film-making [18]. The ultimate goal is to enhance the realism of synthetic videos to align with human perceptual expectations. However, a fundamental challenge that persists across existing methods is the need for synchroniza", + "bbox": [ + 496, + 794, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 20 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 723, + 31 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 99, + 887, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "666", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion. Traditional methods based on Generative Adversarial Networks (GAN) [13, 41, 50, 53], while adept at modeling the lip movements of speakers, often produce inconsistent identities across different frames, leading to issues such as different tooth sizes and fluctuating lip thickness. Similarly, emerging technologies based on Neural Radiance Fields (NeRF) [14, 21, 35, 43, 44] excel in maintaining identity consistency and preserving facial details. However, they struggle with mismatches in lip movements, challenging facial expression control, and unstable head poses, thereby diminishing the overall realism of the video.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we find that the \"devil\" is in the synchronization. Existing methods need more synchronization in four key areas: subject identity, lip movements, facial expressions, and head poses. Firstly, in GAN-based methods, maintaining the subject's identity in the video is challenging due to the instability of features in consecutive frames and the use of only a few frames as references for facial reconstruction [34]. Secondly, lip movements are not synchronized with speech. In NeRF-based methods, audio features trained only on a 5-minute speech dataset struggle to generalize to different speech inputs [44]. Thirdly, there is a lack of facial expression control, with most methods only producing lip movements or controlling blinking, resulting in unnatural facial actions [12]. Fourthly, the head pose is not synchronized. Previous methods relied on sparse landmarks to compute projection error, but jitter and inaccuracy in these landmarks lead to unstable head poses [49], as shown in Fig. 1. These synchronization issues introduce artifacts and significantly reduce realism.", + "bbox": [ + 75, + 260, + 472, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these synchronization challenges, we introduce SyncTalk, a NeRF-based method focused on highly synchronized, realistic, speech-driven talking head synthesis, employing tri-plane hash representations to maintain subject identity. Through the Face-Sync Controller and Head-Sync Stabilizer, SyncTalk significantly enhances the synchronization and visual quality of the synthesized videos. Visual quality is further improved by the Portrait-Sync Generator, which meticulously refines visual details. The entire rendering process can achieve 50 FPS and outputs high-resolution videos.", + "bbox": [ + 75, + 550, + 472, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Within the Face-Sync Controller, we pre-train an audio-visual encoder on the 2D audio-visual dataset, resulting in a generalized representation that ensures synchronized lip movements across different speech samples. For controlling facial expressions, we employ a semantically enriched 3D facial blendshape model [33]. This model is distinguished by its ability to control specific facial expression regions through 52 parameters. Regarding the Head-Sync Stabilizer, we use a head motion tracker [14] to infer the head's rough rotation and translation parameters. Due to the instability of the rough parameters, inspired by Simultaneous Localization and Mapping (SLAM), we incorporate a head", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "point tracker to track dense keypoints and integrate a bundle adjustment method to optimize the head pose, thereby achieving stable and continuous head motion. To further enhance the visual fidelity of SyncTalk, we have designed a Portrait-Sync Generator. This module repairs artifacts in the NeRF modeling, particularly the intricate details of hair and background flaws, and outputs high-resolution video.", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contributions of our work are as follows:", + "bbox": [ + 496, + 200, + 893, + 228 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a Face-Sync Controller that utilizes an Audio-Visual Encoder in conjunction with a Facial Animation Capturer, ensuring accurate lip synchronization and dynamic facial expression rendering.", + "- We introduce a Head-Sync Stabilizer that tracks head rotation and facial movement keypoints. Utilizing the bundle adjustment method, this Stabilizer guarantees smooth and synchronous head motion.", + "- We design a Portrait-Sync Generator that improves visual fidelity by repairing artifacts in NeRF modeling and refining intricate details like hair and background in high-resolution videos." + ], + "bbox": [ + 496, + 232, + 890, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 434, + 640, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. GAN-based Talking Head Synthesis", + "text_level": 1, + "bbox": [ + 500, + 460, + 810, + 478 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, GAN-based talking head synthesis [5, 6, 9, 20, 26, 36, 40, 54, 55] has emerged as an essential research area in computer vision, but they struggle to maintain the identity of the subject in videos consistently. A cluster of noteworthy techniques, including [13, 34, 37, 41, 50, 53], mainly focus on generating video streams for the lip region. These methods create new visuals for talking head portraits by changing the lip region. For instance, Wav2Lip [34] introduces a lip sync expert for supervising lip movements. However, due to the use of five frames from the reference frame to reconstruct the lip, it struggles to maintain the subject's identity. In contrast, methods like [6, 24, 42, 56] perform full-face synthesis but struggle to ensure sync between facial expressions and head poses.", + "bbox": [ + 496, + 487, + 893, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Apart from video stream techniques, efforts have also been made to enable a single image to \"speak\" using speech, as demonstrated in [17, 45, 48]. For example, SadTalker [48] can generate videos of a person speaking from a single image. However, such methods fail to generate natural head poses and facial expressions and struggle to maintain the subject's identity, affecting the sync effect and leading to an unrealistic visual perception.", + "bbox": [ + 496, + 700, + 893, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to these methods, SyncTalk uses NeRF to perform three-dimensional modeling of the face. Its capability to represent continuous 3D scenes in canonical spaces translates to exceptional performance in maintaining subject identity consistency and detail preservation.", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "667", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/22c07d340564f4ba4e434acfa89e395806ce39fb6903ee6560081d815f96a98e.jpg", + "image_caption": [ + "Figure 2. Overview of SyncTalk. Given a cropped reference video of a talking head and the corresponding speech, SyncTalk can extract the Lip Feature $f_{l}$ , Expression Feature $f_{e}$ , and Head Pose $(R, T)$ through two synchronization modules $(a)$ and $(b)$ . The Tri-Plane Hash Representation then models the head, outputting a rough speech-driven video. The Portrait-Sync Generator further restores details such as hair and background, ultimately producing a high-resolution talking head video." + ], + "image_footnote": [], + "bbox": [ + 81, + 77, + 890, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. NeRF-based Talking Head Synthesis", + "text_level": 1, + "bbox": [ + 76, + 411, + 393, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the recent rise of NeRF, numerous fields have begun to utilize it to tackle related challenges [11, 25]. Previous work [14, 22, 43, 44] has integrated NeRF into the task of synthesizing talking heads and has used audio as the driving signal, but these methods are all based on the vanilla NeRF model. For instance, AD-NeRF [14] requires approximately 10 seconds to render a single image. RAD-NeRF [38] aims for real-time video generation and employs a NeRF based on Instant-NGP [30]. ER-NeRF [21] innovatively introduces triple-plane hash encoders to trim the empty spatial regions, advocating for a compact and accelerated rendering approach. GeneFace [44] attempts to reduce NeRF artifacts by translating speech features into facial landmarks, but this often results in inaccurate lip movements. Attempts to create character avatars with NeRF-based methods, such as [10, 51, 52, 57], cannot be directly driven by speech. These methods only use audio as a condition, without a clear concept of sync, and usually result in average lip movement. Additionally, previous methods lack control over facial expressions, being limited to controlling blinking only, and cannot model actions like raising eyebrows or frowning. Furthermore, these methods have a significant issue with unstable head poses, leading to the separation between the head and the torso.", + "bbox": [ + 75, + 441, + 472, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, we use the Face-Sync Controller to model the relationship between audio and lip movements, thereby enhancing the synchronization of lip movements and expressions, and the Head-Sync Stabilizer to stabilize head poses. By addressing these synchronization \"devils\", our method improves visual quality.", + "bbox": [ + 75, + 810, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 410, + 589, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 435, + 609, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the proposed SyncTalk, as shown in Fig. 2. SyncTalk mainly consists of 3 parts: a) lip movements and facial expressions controlled by the FaceSync Controller, b) stable head pose provided by the HeadSync Stabilizer, and c) high-synchronization facial frames rendered by the Dynamic Portrait Renderer. We will describe the content of these three parts in detail in the following subsections.", + "bbox": [ + 496, + 459, + 890, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Face-Sync Controller", + "text_level": 1, + "bbox": [ + 500, + 590, + 699, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Audio-Visual Encoder. Existing methods based on NeRF utilize pre-trained models such as DeepSpeech [2], Wav2Vec 2.0 [3], or HuBERT [16]. These are audio feature extraction methods designed for speech recognition tasks. Using an audio encoder designed for Automatic Speech Recognition (ASR) tasks does not truly reflect lip movements. This is because the pre-trained model is based on the distribution of features from audio to text, whereas we need the feature distribution from audio to lip movements.", + "bbox": [ + 496, + 613, + 890, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering the above, we use an audio and visual synchronization audio encoder trained on the 2D audio-visual synchronization dataset LRS2 [1]. This ensures that the audio features extracted by our method and lip movements have the same feature distribution. The specific implementation method is as follows: We use a pre-trained lip synchronization discriminator [8]. It can give confidence for the lip synchronization effect of the video. The lip synchronization discriminator takes as input a continuous face window $F$ and the corresponding audio frame $A$ . They are", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "668", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "judged as positive samples (with label $y = 1$ ) if they overlap entirely. Otherwise, they are judged as negative samples (with label $y = 0$ ). The discriminator calculates the cosine similarity between these sequences as:", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s i m} (F, A) = \\frac {F \\cdot A}{\\| F \\| _ {2} \\| A \\| _ {2}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 161, + 468, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and then uses binary cross-entropy loss:", + "bbox": [ + 76, + 198, + 341, + 213 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {s y n c}} = - (y \\log (\\operatorname {s i m} (F, A)) + (1 - y) \\log (1 - \\operatorname {s i m} (F, A))), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 226, + 483, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to minimize the distance for synchronized samples and maximize the distance for non-synchronized samples.", + "bbox": [ + 75, + 257, + 468, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Under the supervision of the lip synchronization discriminator, we pre-train a highly synchronized audio-visual feature extractor related to lip movements. First, we use convolutional networks to obtain audio features $\\mathrm{Conv}(A)$ and encode facial features $\\mathrm{Conv}(F)$ . These features are then concatenated. In the decoding phase, we use stacked convolutional layers to restore facial frames using the operation $\\mathrm{Dec}(\\mathrm{Conv}(A) \\oplus \\mathrm{Conv}(F))$ . The $L_{1}$ reconstruction loss during training is given by:", + "bbox": [ + 75, + 287, + 468, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {r e c o n}} = \\| F - \\operatorname {D e c} (\\operatorname {C o n v} (A) \\oplus \\operatorname {C o n v} (F)) \\| _ {1}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 436, + 468, + 453 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Simultaneously, we sample synchronized and nonsynchronized segments using lip movement discriminators, and employ a same sync loss as Eq. 2. By minimizing both losses, we train a facial generation network related to audio. After training, we use $\\mathrm{Conv}(A)$ as the lip space extracted from the audio. Ultimately, we obtain our highly synchronized audio-visual encoder related to lip movements.", + "bbox": [ + 75, + 462, + 468, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Facial Animation Capturer. It is observed that previous methods based on NeRF [14, 21, 44] could only change blinking and could not model facial expressions accurately. This leads to issues such as rigid facial expressions and incorrect facial details if the trained character has significant facial movements, like squinting, raising eyebrows, or frowning. Considering the need for more synchronized and realistic facial expressions, we add an expression synchronization control module. Specifically, we introduce a 3D facial prior using 52 semantically facial blendshape coefficients [33] represented by $B$ to model the face, as shown in Fig. 3. Because the 3D face model can retain the structure information of face motion, it can reflect the content of facial movements well without causing facial structural distortion. During the training, we first use a sophisticated facial blendshape capture module to capture facial expressions as $E(B)$ , and select seven core facial expression control coefficients to control the eyebrow, forehead, and eye areas. They are highly correlated with expression and independent of lip movements. We can synchronize the speaker's expression during the inference process because the facial coefficients are semantically informed.", + "bbox": [ + 75, + 568, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5d13645c920df6df94817562e51c9a4750f4c4cebed23dfd82ba2e57d0f91e33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 89, + 694, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/851fad13038e29b3b0cad2a1260d21d3a60c0ebea9fe827e47c45e19b40d2e2e.jpg", + "image_caption": [ + "Figure 3. Facial Animation Capturer. We use 3D facial blend-shape coefficients to control the expressions of characters." + ], + "image_footnote": [], + "bbox": [ + 504, + 172, + 692, + 256 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2611530a845c6638ed5c75306e89a915a09d63b9f136ed59ab86fac44b732a0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 89, + 888, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/40f9df9661c8a9dde93ad542361784c8ffb56f9e4ccc38e9561f9c732720c7f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 172, + 888, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Facial-Aware Masked-Attention. To reduce the mutual interference between lip features and expression features during training, we introduce the Facial-Aware Disentangle Attention module. Building on the region attention vector $V$ [21], we add masks $M_{\\mathrm{lip}}$ and $M_{\\mathrm{exp}}$ to the attention areas for lips and expressions, respectively. Specifically, the new attention mechanisms are given by:", + "bbox": [ + 498, + 306, + 890, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {\\text {l i p}} = V \\odot M _ {\\text {l i p}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 426, + 890, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {\\mathrm {e x p}} = V \\odot M _ {\\mathrm {e x p}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 446, + 756, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These formulations allow the attention mechanisms to focus solely on their respective parts, thereby reducing entanglement between them. Before the disentanglement, lip movements might induce blinking tendencies and affect hair volume. By introducing the mask module, the attention mechanism can focus on either expressions or lips without affecting other areas, thereby reducing the artifact caused by coupling. Finally, we obtain the disentangled lip feature $f_{l} = f_{\\mathrm{lip}} \\odot V_{\\mathrm{lip}}$ and expression feature $f_{e} = f_{\\mathrm{exp}} \\odot V_{\\mathrm{exp}}$ .", + "bbox": [ + 496, + 468, + 890, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Head-Sync Stabilizer", + "text_level": 1, + "bbox": [ + 500, + 612, + 697, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Head Motion Tracker. The head pose, denoted as $p$ , refers to the rotation angle of a person's head in 3D space and is defined by a rotation $R$ and a translation $T$ . An unstable head pose can lead to head jitter. To obtain a rough estimate of the head pose, initially, the best focal length is determined through $i$ iterations within a predetermined range. For each focal length candidate, $f_{i}$ , the system re-initializes the rotation and translation values. The objective is to minimize the error between the projected landmarks from the 3D Morphable Models (3DMM) [31] and the actual landmarks in the video frame. Formally, the optimal focal length $f_{\\mathrm{opt}}$ is given by:", + "bbox": [ + 496, + 635, + 890, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\text {o p t}} = \\arg \\min _ {f _ {i}} E _ {i} \\left(L _ {2 D}, L _ {3 D} \\left(f _ {i}, R _ {i}, T _ {i}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 825, + 890, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $E_{i}$ represents the Mean Squared Error (MSE) between these landmarks, $L_{3D}(f_i,R_i,T_i)$ represents the projected landmarks from the 3DMM for a given focal length", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "669", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$f_{i}$ , the corresponding rotation and translation parameters $R_{i}$ and $T_{i}$ , $L_{2D}$ are the actual landmarks from the video frame. Subsequently, leveraging the optimal focal length $f_{\\mathrm{opt}}$ , the system refines the rotation $R$ and translation $T$ parameters for all frames to better align the model's projected landmarks with the actual video landmarks. This refinement process can be mathematically represented as:", + "bbox": [ + 76, + 90, + 472, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(R _ {\\text {o p t}}, T _ {\\text {o p t}}\\right) = \\arg \\min _ {R, T} E \\left(L _ {2 D}, L _ {3 D} \\left(f _ {\\text {o p t}}, R, T\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 210, + 468, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $E$ denotes the MSE metric, between the 3D model's projected landmarks $L_{3D}$ for the optimal focal length $f_{\\mathrm{opt}}$ , and the actual 2D landmarks $L_{2D}$ in the video frame. The optimized rotation $R_{\\mathrm{opt}}$ and translation $T_{\\mathrm{opt}}$ are obtained by minimizing this error across all frames.", + "bbox": [ + 75, + 247, + 468, + 323 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Head Points Tracker. Considering methods based on NeRF and their requirements for inputting head rotation $R$ and translation $T$ , previous methods utilize 3DMM-based techniques to extract head poses and generate an inaccurate result. To improve the precision of $R$ and $T$ , We use an optical flow estimation model from [43] to track facial keypoints $K$ . Specifically, using a pre-trained optical flow estimation model, after obtaining the facial motion optical flow, we use the Laplacian filter to select the keypoints where the most significant flow changes are located and track the motion trajectories of these keypoints in the flow sequence. Through this module, our method ensures a more precise and consistent facial keypoint alignment across all frames, enhancing the accuracy of head pose parameters.", + "bbox": [ + 75, + 324, + 468, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Bundle Adjustment.** Given the keypoints and the rough head pose, we introduce a two-stage optimization framework from [14] to enhance the accuracy of keypoints and head pose estimations. In the first stage, we randomly initialize the 3D coordinates of $j$ keypoints and optimize their positions to align with the tracked keypoints on the image plane. This process involves minimizing a loss function $L_{\\mathrm{init}}$ , which captures the discrepancy between projected keypoints $P$ and the tracked keypoints $K$ , as given by:", + "bbox": [ + 75, + 536, + 470, + 674 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {i n i t}} = \\sum_ {j} \\| P _ {j} - K _ {j} \\| _ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 688, + 468, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, in the second stage, we embark on a more comprehensive optimization to refine the 3D keypoints and the associated head jointly pose parameters. Through the Adam Optimization [19], the algorithm adjust the spatial coordinates, rotation angles $R$ , and translations $T$ to minimize the alignment error $L_{\\mathrm{sec}}$ , expressed as:", + "bbox": [ + 75, + 726, + 468, + 818 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\sec} = \\sum_ {j} \\| P _ {j} (R, T) - K _ {j} \\| _ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 833, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After these optimizations, the resultant head pose and translation parameters are observed to be smooth and stable.", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Dynamic Portrait Rendering", + "text_level": 1, + "bbox": [ + 500, + 90, + 750, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tri-Plane Hash Representation. Using a collection of multi-perspective images along with corresponding camera poses, NeRF [27] harnesses these resources to manifest a 3D static scene. This representation utilizes an implicit function $\\mathcal{F}$ , delineated as $\\mathcal{F}:(\\mathrm{x},\\mathrm{d})\\to (\\mathrm{c},\\sigma)$ , where the 3D spatial location is given by $\\mathrm{x} = (x,y,z)$ , and the direction of viewing is characterized by $\\mathrm{d} = (\\theta ,\\phi)$ . The resultant values, $\\mathrm{c} = (r,g,b)$ and $\\sigma$ , represent the radiance and density, respectively. The predicted pixel color, denoted by $\\hat{C} (\\mathbf{r})$ , intertwined with ray $\\mathrm{r}(t) = \\mathrm{o} + t\\mathrm{d}$ originating from the camera's core position $\\mathrm{o}$ , is derivable using:", + "bbox": [ + 498, + 113, + 893, + 281 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} (\\mathrm {r}) = \\int_ {t _ {n}} ^ {t _ {f}} \\sigma (\\mathrm {r} (t)) \\cdot \\mathrm {c} (\\mathrm {r} (t), \\mathrm {d}) \\cdot T (t) d t, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 290, + 890, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t_n$ and $t_f$ are the near and far bounds, and $T(t)$ is the accumulated transmittance. Addressing the challenges of hash collisions and optimizing audio feature processing, we incorporate three uniquely oriented 2D hash grids [21]. A coordinate, given by $\\mathbf{x} = (x,y,z) \\in \\mathbb{R}^{\\mathrm{XYZ}}$ , undergoes an encoding transformation for its projected values via three individual 2D-multiresolution hash encoders [30]:", + "bbox": [ + 498, + 329, + 890, + 435 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} ^ {\\mathrm {A B}}: (a, b) \\rightarrow \\mathrm {f} _ {a b} ^ {\\mathrm {A B}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 443, + 890, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the output $\\mathsf{f}_{ab}^{\\mathrm{AB}} \\in \\mathbb{R}^{LD}$ , with the number of levels $L$ and feature dimensions per entry $D$ , signifies the planar geometric feature corresponding to the projected coordinate $(a,b)$ and $\\mathcal{H}^{\\mathrm{AB}}$ denotes the multiresolution hash encoder for plane $\\mathbb{R}^{\\mathrm{AB}}$ . By fusing the outcomes, the conclusive geometric feature $\\mathsf{fg} \\in \\mathbb{R}^{3 \\times LD}$ is derived as:", + "bbox": [ + 498, + 468, + 890, + 560 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {f} _ {\\mathrm {x}} = \\mathcal {H} ^ {\\mathrm {X Y}} (x, y) \\oplus \\mathcal {H} ^ {\\mathrm {Y Z}} (y, z) \\oplus \\mathcal {H} ^ {\\mathrm {X Z}} (x, z), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 574, + 890, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the concatenation of features is symbolized by $\\oplus$ , resulting in a $3 \\times LD$ -channel vector. Employing $\\mathrm{f_x}$ , the perspective direction d, the lip feature $f_l$ , and the expression feature $f_e$ , the tri-plane hash's implicit function is defined as:", + "bbox": [ + 498, + 599, + 890, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} ^ {\\mathcal {H}}: (\\mathrm {x}, \\mathrm {d}, f _ {l}, f _ {e}; \\mathcal {H} ^ {3}) \\rightarrow (\\mathrm {c}, \\sigma), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 683, + 890, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{H}^3$ amalgamates the triad of planar hash encoders, as illustrated in Eq. 10.", + "bbox": [ + 498, + 709, + 888, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our training employs a two-step coarse-to-fine strategy, initially using MSE loss to assess the difference between predicted $\\hat{C} (\\mathbf{r})$ and actual image colors $C(\\mathbf{r})$ . Recognizing MSE's limitations in detail capture, we advance to a refinement stage, incorporating LPIPS loss for enhanced detail, similar to ER-NeRF [21]. We extract random patches $\\mathcal{P}$ from the image, combining LPIPS (weighted by $\\lambda$ ) with MSE for improved detail representation, as shown:", + "bbox": [ + 496, + 739, + 890, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\sum_ {\\mathrm {r}} \\| C (\\mathrm {r}) - \\hat {C} (\\mathrm {r}) \\| _ {2} + \\lambda \\times \\mathcal {L} _ {\\mathrm {L P I P S}} (\\hat {\\mathcal {P}}, \\mathcal {P}). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 872, + 890, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "670", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b88f1d17a9a426abb6f5a59c4c6b08dced8705a4b4a7c2131677b97e5b246d63.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPSNR ↑LPIPS ↓MS-SSIM ↑FID ↓NIQE ↓BRISQUE ↓LMD ↓AUE ↓LSE-C ↑
GANWav2Lip (ACM MM 20 [34])33.43850.06970.978116.022814.536744.26594.96302.90299.2387
VideoReTalking (SIGGRAPH Asia 22 [7])31.79230.04880.96809.206314.241043.04655.85753.33087.9683
DINet (AAAI 23 [50])31.64750.04430.96409.430014.685040.36504.37253.68756.5653
TalkLip (CVPR 23 [41])32.51540.07820.969718.499714.638546.67175.86052.95795.9472
IP-LAP (CVPR 23 [53])35.15250.04430.98038.212514.640042.07503.33502.84004.9541
NeRFAD-NeRF (ICCV 21 [14])26.72910.15360.911128.986214.909155.46672.99955.54814.4996
RAD-NeRF (arXiv 22 [38])31.77540.07780.94528.657013.443344.68922.91155.09585.5219
GeneFace (ICLR 23 [44])24.81650.11780.875321.708413.335346.50614.28595.45275.1950
ER-NeRF (ICCV 23 [21])32.52160.03340.95015.293613.704834.73612.81374.18735.7749
SyncTalk (w/o Portrait)35.35420.02350.97693.924713.133333.29542.57142.57968.1331
SyncTalk (Portrait)37.40170.01130.98412.707014.216537.30422.50433.20748.0263
", + "bbox": [ + 81, + 77, + 893, + 291 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Portrait-Sync Generator. During the training process, to address NeRF's limitations in capturing fine details like hair strands and dynamic backgrounds, we introduce a Portrait-Sync Generator with two key sections. First, NeRF renders the face area $(F_r)$ , creates $G(F_r)$ through Gaussian blur, and then uses our synchronized head pose to be able to merge with the original image $(F_o)$ to enhance hair detail fidelity. Second, when the head and torso are combined, if the character in the source video speaks while the generated face is silent, a dark gap area might appear, as shown in Fig. 5 (b). We fill these areas with the average neck color $(C_n)$ . This approach results in more realistic details and improved visual quality through the Portrait-Sync Generator.", + "bbox": [ + 75, + 343, + 472, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 559, + 209, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 585, + 284, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. To ensure a fair comparison, we use the same well-edited video sequences from [14, 21, 44], including English and French. The average length of these videos is approximately 8,843 frames, and each video is recorded at 25 FPS. Except for the video from AD-NeRF [14], which has a resolution of $450 \\times 450$ , all other videos have a resolution of $512 \\times 512$ , with the character-centered.", + "bbox": [ + 75, + 609, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison Baselines. We compare our method with five GAN-based methods, including Wav2Lip [34], Video-ReTalking [7], DINet [50], TalkLip [41], and IP-LAP [53], and NeRF-based methods such as AD-NeRF [14], RAD-NeRF [38], GeneFace [44], and ER-NeRF [21].", + "bbox": [ + 75, + 717, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. In the coarse stage, the portrait head is trained for 100,000 iterations and 25,000 in the fine stage, sampling $256^2$ rays per iteration using a 2D hash encoder ( $L = 14$ , $F = 1$ ). We employ the AdamW optimizer [23], with learning rates of 0.01 for the hash encoder and 0.001 for other modules. Total training time is approximately 2 hours on an NVIDIA RTX 3090 GPU.", + "bbox": [ + 75, + 795, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/12d62942c6b37fe3355ce4596ebccc24d3a79b633ea42f4b18bed8dfdc0cfb45.jpg", + "table_caption": [ + "Table 1. The quantitative results of the head reconstruction. \"Portrait\" refers to the use of the Portrait-Sync Generator. We achieve state-of-the-art performance on most metrics. We highlight best and second-best results." + ], + "table_footnote": [], + "table_body": "
MethodsAudio AAudio B
LSE-D ↓LSE-C ↑LSE-D ↓LSE-C ↑
DINet (AAAI 23 [50])8.50315.69568.20385.1134
TalkLip (CVPR 23 [41])8.76155.74498.70195.5359
IP-LAP (CVPR 23 [53])9.80373.85789.11024.389
GeneFace (ICLR 23 [44])9.54514.29339.66753.7342
ER-NeRF (ICCV 23 [21])11.8132.407610.73383.0242
SyncTalk (Ours)7.72116.66598.02486.2596
", + "bbox": [ + 503, + 340, + 893, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. The quantitative results of the lip synchronization. We use two different audio samples to drive the same subject, then highlight best and second-best results.", + "bbox": [ + 498, + 473, + 892, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 529, + 720, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Full Reference Quality Assessment. In terms of image quality, we use full reference metrics such as Peak Signal-to-Noise Ratio (PSNR), Learned Perceptual Image Patch Similarity (LPIPS) [46], Multi-Scale Structure Similarity (MS-SSIM), and Frechet Inception Distance (FID) [15] as evaluation metrics.", + "bbox": [ + 496, + 551, + 890, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "No Reference Quality Assessment. In high PSNR images, texture details may not align with human visual perception [47]. For more precise output definition and comparison, we use two No Reference methods: the Natural Image Quality Evaluator (NIQE)[29] and the Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE)[28].", + "bbox": [ + 496, + 643, + 890, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synchronization Assessment. For synchronization, we use landmark distance (LMD) to measure the synchronicity of facial movements, action units error (AUE) [4] to assess the accuracy of facial movements, and introduce Lip Sync Error Confidence (LSE-C), consistent with Wav2Lip [34], to evaluate the synchronization between lip movements and audio.", + "bbox": [ + 496, + 734, + 890, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Results. The evaluation results of the head reconstruction are shown in Tab. 1. We compare recent methods based on GAN and NeRF. It can be observed that our image quality is superior to other methods in all aspects. In terms of synchronization, our results surpass most methods.", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "671", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f7d0ae1aceef51e6c939e47d81b1e346580de459ee560e4411f40a69f7591ab0.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison of facial synthesis by different methods. Our method has the best visual effect on lip movements and facial expressions without the problem of separation of head and torso. Please zoom in for better visualization." + ], + "image_footnote": [], + "bbox": [ + 76, + 73, + 893, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the two output modes of SyncTalk, one processed through the Portrait-Sync Generator and one without it. After processing through the Portrait-Sync Generator, hair details are restored, and the image quality is improved. Since we can maintain the subject's identity well, we surpass GAN-based methods in image quality. Thanks to the sync of lip, expression, and pose, we also outperform NeRF-based methods in image quality. Especially in terms of the LPIPS metric, our method has three times improvement compared to the previous state-of-the-art method ER-NeRF [21]. We compare the latest SOTA method drivers using out-of-distribution (OOD) audio, and the results are shown in Tab. 2. We introduce Lip Synchronization Error Distance (LSE-D) and Confidence (LSEC) for lip-audio sync evaluation, aligning with [34]. Our method shows state-of-the-art lip synchronization, overcoming small-sample NeRF limitations by incorporating a pre-trained audio-visual encoder for lip modeling.", + "bbox": [ + 75, + 517, + 472, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also test the rendering speed. On an NVIDIA RTX 3090 GPU, and having preloaded the data onto the GPU, the head is outputted at 52 FPS with a resolution of $512 \\times 512$ . In Portrait mode, using the Portrait-Sync Generator, we achieve rendering speeds of 50 FPS with our CUDA acceleration. This far exceeds the video input speed of 25 FPS, allowing for real-time generation of video streams.", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 517, + 709, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Results. To more intuitively evaluate image quality, we display a comparison between our method and other methods in Fig. 4. In this figure, it can be observed that SyncTalk demonstrates more precise and more accurate facial details. Compared to Wav2Lip [34], our method better preserves the subject's identity while offering higher fidelity and resolution. Against IP-LAP [53], our method excels in lip shape synchronization, primarily due to the audio-visual consistency brought by the audio-visual encoder. Compared to GeneFace [44], our method can accurately reproduce actions such as blinking and eyebrow-raising through expression sync. In contrast to ER-NeRF [21], our method avoids the separation between the head and body through the Pose-Sync Stabilizer and generates more accurate lip shapes. Our method achieves the best overall visual effect; we recommend watching the supplementary video for comparison.", + "bbox": [ + 496, + 546, + 890, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. To provide a more comprehensive evaluation of the proposed model, we design an exhaustive user study questionnaire. We extract 24 video clips, each lasting more than 10 seconds, which include head poses, facial expressions, and lip movements. Each method is represented by three of these clips. We invite 35 participants", + "bbox": [ + 496, + 809, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "672", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fe64f0830fe8272cad41bbcefd7800e80cb8b51541f9ea9b253302b109e6ea71.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Wav2Lip [34]DINet [50]TalkLip [41]IP-LAP [53]AD-NeRF [14]GeneFace [44]ER-NeRF [21]SyncTalk
Lip-sync Accuracy3.8393.6962.8933.1612.6962.9823.1894.304
Exp-sync Accuracy3.5363.4822.6073.4112.2503.0362.9464.036
Pose-sync Accuracy3.5713.5712.8753.6962.2322.9292.6073.980
Image Quality2.5002.6962.0543.5712.4643.4823.0364.054
Video Realness2.9292.4292.4293.1612.0362.7322.5184.018
", + "bbox": [ + 80, + 89, + 885, + 188 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a3232178b6751e92d9e9df513fad385c645650286e1f0fe520d7d6ffe6e49853.jpg", + "table_caption": [ + "Table 3. User Study. Rating is on a scale of 1-5; the higher the better. The term \"Exp-sync Accuracy\" is an abbreviation for \"Expression-sync Accuracy\". We highlight best and second-best results." + ], + "table_footnote": [], + "table_body": "
PSNR ↑LPIPS ↓LMD ↓
Ours37.3110.01212.8032
replace Audio-Visual Encoder with Hubert [16]33.5160.02763.3961
replace Facial Animation Capture with ER-NeRF [21]'s30.2730.04153.0516
w/o Facial-Aware Masked-Attention36.5360.01652.9139
w/o Head-Sync Stabilizer28.9840.06343.5373
w/o Portrait-Sync Generator32.2390.03952.8154
", + "bbox": [ + 80, + 239, + 465, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study for our components. We show the PSNR, LPIPS, and LMD in different cases.", + "bbox": [ + 75, + 426, + 468, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to provide scores. The questionnaire is designed using the Mean Opinion Score (MOS) scoring protocol, asking participants to rate the generated videos from five perspectives: (1) Lip-sync Accuracy, (2) Expression-sync Accuracy, (3) Pose-sync Accuracy, (4) Image Quality, and (5) Video Realness. On average, participants take 19 minutes to complete the questionnaire, with a standardized Cronbach $\\alpha$ coefficient of 0.96. The results of the User Study are displayed in Tab. 3. SyncTalk surpasses previous methods in all evaluations. Furthermore, we achieve the highest score in video authenticity, surpassing the second-place method, IP-LAP [53], by a margin of $20\\%$ . User studies indicate that our method can generate visually excellent quality as perceived by humans, achieving high realism.", + "bbox": [ + 75, + 470, + 468, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 695, + 230, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct an ablation study to examine the contributions of different components in our model. We select three core metrics for evaluation: PSNR, LPIPS, and LMD. We select a subject named \"May\" for testing, and the results are presented in Tab. 4.", + "bbox": [ + 75, + 718, + 468, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Audio-Visual Encoder provides the primary lip sync information. When this module is replaced, all three metrics become worse, with the LMD error increasing by $21.15\\%$ in particular, indicating decreased lip motion synchronization, as shown in Fig. 5 (a), and showing that our audio-visual encoders can extract accurate lip features. Replacing Facial Animation Capture with ER-NeRF [21]'s blink module", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/612756322144cdd4323c7c0cd0c4bfcfef36b696637d5a02aa5d813042a41744.jpg", + "image_caption": [ + "w/o Audio-Visual Encoder w/ Audio-Visual Encoder w/o Head-Sync Stabilizer w/ Head-Sync Stabilizer", + "Figure 5. Ablation Study on Audio-Visual Encoder and Head-Sync Stabilizer. Removing them will lead to $(a)$ and $(b)$ ." + ], + "image_footnote": [], + "bbox": [ + 504, + 251, + 890, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "affects eyebrow movements and image quality.", + "bbox": [ + 500, + 388, + 808, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Facial-Aware Masked-Attention mainly alleviates motion entanglement between the lips and the rest of the face, slightly affecting the image quality after removal. Without the Head-Sync Stabilizer, all metrics significantly declined, notably LPIPS, leading to head pose jitter and head and torso separation, as shown in Fig. 5 (b). The Portrait-Sync Generator restores details like hair. Removing this module impacts the restoration of details like hair, leading to noticeable segmentation boundaries.", + "bbox": [ + 496, + 404, + 890, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 558, + 617, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce SyncTalk, a highly synchronized NeRF-based method for realistic speech-driven talking head synthesis. Our framework includes a Facial Sync Controller, Head Sync Stabilizer, and Portrait Sync Generator, which maintain subject identity and generate synchronized lip movements, facial expressions, and stable head poses. Through extensive evaluation, SyncTalk demonstrates superior performance in creating realistic and synchronized talking head videos compared to existing methods. We expect that SyncTalk will not only enhance various applications but also inspire further innovation in the field of talking head synthesis.", + "bbox": [ + 496, + 584, + 890, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 784, + 658, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by National Natural Science Foundation of China (NSFC) under Grant Nos. 62172421 and 62072459, the Outstanding Innovative Talents Cultivation Funded Programs 2023 of Renmin University of China, and the Public Computing Cloud at Renmin University of China.", + "bbox": [ + 496, + 809, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "673", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence, 44(12):8717-8727, 2018. 3", + "[2] Dario Amodei, Sundaram Ananthanarayanan, Rishita Anubhai, Jingliang Bai, Eric Battenberg, Carl Case, Jared Casper, Bryan Catanzaro, Qiang Cheng, Guoliang Chen, et al. Deep speech 2: End-to-end speech recognition in english and mandarin. In International conference on machine learning, pages 173-182. PMLR, 2016. 3", + "[3] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 3", + "[4] Tadas Baltrusaitis, Marwa Mahmoud, and Peter Robinson. Cross-dataset learning and person-specific normalisation for automatic action unit detection. In 2015 11th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-6. IEEE, 2015. 6", + "[5] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European conference on computer vision (ECCV), pages 520-535, 2018. 2", + "[6] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7832-7841, 2019. 2", + "[7] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 6", + "[8] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 3", + "[9] Dipanjan Das, Sandika Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 408-424. Springer, 2020. 2", + "[10] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3", + "[11] Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379, 2022.3", + "[12] Shreyank N Gowda, Dheeraj Pandey, and Shashank Narayana Gowda. From pixels to portraits:" + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "A comprehensive survey of talking head generation techniques and applications. arXiv preprint arXiv:2308.16041, 2023. 2", + "[13] Jiazhi Guan, Zhanwang Zhang, Hang Zhou, Tianshu Hu, Kaisiyuan Wang, Dongliang He, Haocheng Feng, Jingtuo Liu, Errui Ding, Ziwei Liu, et al. Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1505-1515, 2023. 2", + "[14] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. 2, 3, 4, 5, 6, 8", + "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[16] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460, 2021. 3, 8", + "[17] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2", + "[18] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM transactions on graphics (TOG), 37(4): 1-14, 2018. 1", + "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[20] Prajwal KR, Rudrabha Mukhopadhyay, Jerin Philip, Abhishek Jha, Vinay Namboodiri, and CV Jawahar. Towards automatic face-to-face translation. In Proceedings of the 27th ACM international conference on multimedia, pages 1428-1436, 2019. 2", + "[21] Jiahe Li, Jiawei Zhang, Xiao Bai, Jun Zhou, and Lin Gu. Efficient region-aware neural radiance fields for high-fidelity talking portrait synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7568-7578, 2023. 2, 3, 4, 5, 6, 7, 8", + "[22] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. In European Conference on Computer Vision, pages 106-125. Springer, 2022. 3", + "[23] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[24] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "674", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 3", + "[26] Moustafa Meshry, Saksham Suri, Larry S Davis, and Abhinav Shrivastava. Learned spatial representations for few-shot talking-head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13829-13838, 2021. 2", + "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 5", + "[28] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing, 21(12): 4695-4708, 2012. 6", + "[29] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6", + "[30] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3, 5", + "[31] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 4", + "[32] Ziqiao Peng, Yihao Luo, Yue Shi, Hao Xu, Xiangyu Zhu, Hongyan Liu, Jun He, and Zhaoxin Fan. Selftalk: A self-supervised commutative training diagram to comprehend 3d talking faces. In Proceedings of the 31st ACM International Conference on Multimedia, pages 5292-5301, 2023. 1", + "[33] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20687-20697, 2023. 2, 4", + "[34] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 2, 6, 7, 8", + "[35] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In European Conference on Computer Vision, pages 666-682. Springer, 2022. 2", + "[36] Linsen Song, Wayne Wu, Chen Qian, Ran He, and Chen Change Loy. Everybody's talkin': Let me talk as you want. IEEE Transactions on Information Forensics and Security, 17:585-598, 2022. 2", + "[37] Yasheng Sun, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Zhibin Hong, Jingtuo Liu, Errui Ding, Jingdong Wang, Zwei Liu, and Koike Hideki. Masked lip-sync prediction by" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "audio-visual contextual exploitation in transformers. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2", + "[38] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 3, 6", + "[39] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVI 16, pages 716–731. Springer, 2020. 1", + "[40] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. International Journal of Computer Vision, 128:1398-1413, 2020. 2", + "[41] Jiadong Wang, Xinyuan Qian, Malu Zhang, Robby T Tan, and Haizhou Li. Seeing what you said: Talking face generation guided by a lip reading expert. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14653-14662, 2023. 2, 6, 8", + "[42] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2", + "[43] Shunyu Yao, RuiZhe Zhong, Yichao Yan, Guangtao Zhai, and Xiaokang Yang. Dfa-nerf: Personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791, 2022. 2, 3, 5", + "[44] Zhenhui Ye, Ziyue Jiang, Yi Ren, Jinglin Liu, JinZheng He, and Zhou Zhao. Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430, 2023. 2, 3, 4, 6, 7, 8", + "[45] Chenxu Zhang, Yifan Zhao, Yifei Huang, Ming Zeng, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Facial: Synthesizing dynamic talking face with implicit attribute learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3867-3876, 2021. 2", + "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[47] Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. Ranksrgan: Generative adversarial networks with ranker for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3096-3105, 2019. 6", + "[48] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2", + "[49] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "675", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661-3670, 2021. 2", + "[50] Zhimeng Zhang, Zhipeng Hu, Wenjin Deng, Changjie Fan, Tangjie Lv, and Yu Ding. Dinet: Deformation inpainting network for realistic face visually dubbing on high resolution video. arXiv preprint arXiv:2303.03988, 2023. 2, 6, 8", + "[51] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 3", + "[52] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21057-21067, 2023. 3", + "[53] Weizhi Zhong, Chaowei Fang, Yinqi Cai, Pengxu Wei, Gangming Zhao, Liang Lin, and Guanbin Li. Identity-preserving talking face generation with landmark and appearance priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2023. 2, 6, 7, 8", + "[54] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI conference on artificial intelligence, pages 9299–9306, 2019. 2", + "[55] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 2", + "[56] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG), 39(6):1-15, 2020. 2", + "[57] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4574-4584, 2023. 3" + ], + "bbox": [ + 78, + 92, + 468, + 656 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "676", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/89645a31-8829-4d8f-939e-0d17fcee95a5_model.json b/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/89645a31-8829-4d8f-939e-0d17fcee95a5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..63db1135e0734a7bb55f2ee8a4542cc87a810f1d --- /dev/null +++ b/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/89645a31-8829-4d8f-939e-0d17fcee95a5_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.021 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.725, + 0.032 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.756, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.131, + 0.856, + 0.154 + ], + "angle": 0, + "content": "SyncTalk: The Devil is in the Synchronization for Talking Head Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.179, + 0.841, + 0.217 + ], + "angle": 0, + "content": "Ziqiao Peng\\(^{1}\\) Wentao Hu\\(^{2}\\) Yue Shi\\(^{3}\\) Xiangyu Zhu\\(^{4}\\) Xiaomei Zhang\\(^{4}\\) Hao Zhao\\(^{5}\\) Jun He\\(^{1}\\) Hongyan Liu\\(^{5*}\\) Zhaoxin Fan\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.217, + 0.823, + 0.254 + ], + "angle": 0, + "content": "\\(^{1}\\)Renmin University of China \\(^{2}\\)Beijing University of Posts and Telecommunications \\(^{3}\\)Psyche AI Inc. \\(^{4}\\)Chinese Academy of Sciences \\(^{5}\\)Tsinghua University" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.269, + 0.873, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.535, + 0.895, + 0.579 + ], + "angle": 0, + "content": "Figure 1. The proposed SyncTalk synthesizes synchronized talking head videos, employing tri-plane hash representations to maintain subject identity. It can generate synchronized lip movements, facial expressions, and stable head poses, and restores hair details to create high-resolution videos." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.59, + 0.314, + 0.607 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.474, + 0.879 + ], + "angle": 0, + "content": "Achieving high synchronization in the synthesis of realistic, speech-driven talking head videos presents a significant challenge. Traditional Generative Adversarial Networks (GAN) struggle to maintain consistent facial identity, while Neural Radiance Fields (NeRF) methods, although they can address this issue, often produce mismatched lip movements, inadequate facial expressions, and unstable head poses. A lifelike talking head requires synchronized coordination of subject identity, lip movements, facial expressions, and head poses. The absence of these synchronizations is a fundamental flaw, leading to unrealistic and artificial outcomes. To address the critical issue of synchronization, identified as the \"devil\" in creating realistic talking heads, we introduce SyncTalk. This NeRF-based method effectively maintains subject identity, enhancing synchronization and realism in talking head synthesis. SyncTalk employs a Face-Sync Controller to align lip movements" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.592, + 0.892, + 0.743 + ], + "angle": 0, + "content": "with speech and innovatively uses a 3D facial blendshape model to capture accurate facial expressions. Our Head-Sync Stabilizer optimizes head poses, achieving more natural head movements. The Portrait-Sync Generator restores hair details and blends the generated head with the torso for a seamless visual experience. Extensive experiments and user studies demonstrate that SyncTalk outperforms state-of-the-art methods in synchronization and realism. We recommend watching the supplementary video: https://ziqiaopeng.github.io/synctalk" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.77, + 0.633, + 0.787 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.893, + 0.903 + ], + "angle": 0, + "content": "The quest to generate dynamic and realistic speech-driven talking heads has intensified, driven by expanding applications in digital assistants [39], virtual reality [32], and film-making [18]. The ultimate goal is to enhance the realism of synthetic videos to align with human perceptual expectations. However, a fundamental challenge that persists across existing methods is the need for synchroniza" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.888, + 0.235, + 0.901 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "666" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "tion. Traditional methods based on Generative Adversarial Networks (GAN) [13, 41, 50, 53], while adept at modeling the lip movements of speakers, often produce inconsistent identities across different frames, leading to issues such as different tooth sizes and fluctuating lip thickness. Similarly, emerging technologies based on Neural Radiance Fields (NeRF) [14, 21, 35, 43, 44] excel in maintaining identity consistency and preserving facial details. However, they struggle with mismatches in lip movements, challenging facial expression control, and unstable head poses, thereby diminishing the overall realism of the video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.473, + 0.548 + ], + "angle": 0, + "content": "In this paper, we find that the \"devil\" is in the synchronization. Existing methods need more synchronization in four key areas: subject identity, lip movements, facial expressions, and head poses. Firstly, in GAN-based methods, maintaining the subject's identity in the video is challenging due to the instability of features in consecutive frames and the use of only a few frames as references for facial reconstruction [34]. Secondly, lip movements are not synchronized with speech. In NeRF-based methods, audio features trained only on a 5-minute speech dataset struggle to generalize to different speech inputs [44]. Thirdly, there is a lack of facial expression control, with most methods only producing lip movements or controlling blinking, resulting in unnatural facial actions [12]. Fourthly, the head pose is not synchronized. Previous methods relied on sparse landmarks to compute projection error, but jitter and inaccuracy in these landmarks lead to unstable head poses [49], as shown in Fig. 1. These synchronization issues introduce artifacts and significantly reduce realism." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.473, + 0.718 + ], + "angle": 0, + "content": "To address these synchronization challenges, we introduce SyncTalk, a NeRF-based method focused on highly synchronized, realistic, speech-driven talking head synthesis, employing tri-plane hash representations to maintain subject identity. Through the Face-Sync Controller and Head-Sync Stabilizer, SyncTalk significantly enhances the synchronization and visual quality of the synthesized videos. Visual quality is further improved by the Portrait-Sync Generator, which meticulously refines visual details. The entire rendering process can achieve 50 FPS and outputs high-resolution videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Within the Face-Sync Controller, we pre-train an audio-visual encoder on the 2D audio-visual dataset, resulting in a generalized representation that ensures synchronized lip movements across different speech samples. For controlling facial expressions, we employ a semantically enriched 3D facial blendshape model [33]. This model is distinguished by its ability to control specific facial expression regions through 52 parameters. Regarding the Head-Sync Stabilizer, we use a head motion tracker [14] to infer the head's rough rotation and translation parameters. Due to the instability of the rough parameters, inspired by Simultaneous Localization and Mapping (SLAM), we incorporate a head" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.895, + 0.198 + ], + "angle": 0, + "content": "point tracker to track dense keypoints and integrate a bundle adjustment method to optimize the head pose, thereby achieving stable and continuous head motion. To further enhance the visual fidelity of SyncTalk, we have designed a Portrait-Sync Generator. This module repairs artifacts in the NeRF modeling, particularly the intricate details of hair and background flaws, and outputs high-resolution video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.895, + 0.229 + ], + "angle": 0, + "content": "In summary, the main contributions of our work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.233, + 0.892, + 0.293 + ], + "angle": 0, + "content": "- We present a Face-Sync Controller that utilizes an Audio-Visual Encoder in conjunction with a Facial Animation Capturer, ensuring accurate lip synchronization and dynamic facial expression rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.293, + 0.892, + 0.352 + ], + "angle": 0, + "content": "- We introduce a Head-Sync Stabilizer that tracks head rotation and facial movement keypoints. Utilizing the bundle adjustment method, this Stabilizer guarantees smooth and synchronous head motion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.892, + 0.412 + ], + "angle": 0, + "content": "- We design a Portrait-Sync Generator that improves visual fidelity by repairing artifacts in NeRF modeling and refining intricate details like hair and background in high-resolution videos." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.233, + 0.892, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.435, + 0.642, + 0.451 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.462, + 0.812, + 0.479 + ], + "angle": 0, + "content": "2.1. GAN-based Talking Head Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.488, + 0.895, + 0.7 + ], + "angle": 0, + "content": "Recently, GAN-based talking head synthesis [5, 6, 9, 20, 26, 36, 40, 54, 55] has emerged as an essential research area in computer vision, but they struggle to maintain the identity of the subject in videos consistently. A cluster of noteworthy techniques, including [13, 34, 37, 41, 50, 53], mainly focus on generating video streams for the lip region. These methods create new visuals for talking head portraits by changing the lip region. For instance, Wav2Lip [34] introduces a lip sync expert for supervising lip movements. However, due to the use of five frames from the reference frame to reconstruct the lip, it struggles to maintain the subject's identity. In contrast, methods like [6, 24, 42, 56] perform full-face synthesis but struggle to ensure sync between facial expressions and head poses." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.702, + 0.895, + 0.824 + ], + "angle": 0, + "content": "Apart from video stream techniques, efforts have also been made to enable a single image to \"speak\" using speech, as demonstrated in [17, 45, 48]. For example, SadTalker [48] can generate videos of a person speaking from a single image. However, such methods fail to generate natural head poses and facial expressions and struggle to maintain the subject's identity, affecting the sync effect and leading to an unrealistic visual perception." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Compared to these methods, SyncTalk uses NeRF to perform three-dimensional modeling of the face. Its capability to represent continuous 3D scenes in canonical spaces translates to exceptional performance in maintaining subject identity consistency and detail preservation." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "667" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.078, + 0.891, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.337, + 0.895, + 0.395 + ], + "angle": 0, + "content": "Figure 2. Overview of SyncTalk. Given a cropped reference video of a talking head and the corresponding speech, SyncTalk can extract the Lip Feature \\( f_{l} \\), Expression Feature \\( f_{e} \\), and Head Pose \\( (R, T) \\) through two synchronization modules \\( (a) \\) and \\( (b) \\). The Tri-Plane Hash Representation then models the head, outputting a rough speech-driven video. The Portrait-Sync Generator further restores details such as hair and background, ultimately producing a high-resolution talking head video." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.412, + 0.395, + 0.429 + ], + "angle": 0, + "content": "2.2. NeRF-based Talking Head Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.473, + 0.804 + ], + "angle": 0, + "content": "With the recent rise of NeRF, numerous fields have begun to utilize it to tackle related challenges [11, 25]. Previous work [14, 22, 43, 44] has integrated NeRF into the task of synthesizing talking heads and has used audio as the driving signal, but these methods are all based on the vanilla NeRF model. For instance, AD-NeRF [14] requires approximately 10 seconds to render a single image. RAD-NeRF [38] aims for real-time video generation and employs a NeRF based on Instant-NGP [30]. ER-NeRF [21] innovatively introduces triple-plane hash encoders to trim the empty spatial regions, advocating for a compact and accelerated rendering approach. GeneFace [44] attempts to reduce NeRF artifacts by translating speech features into facial landmarks, but this often results in inaccurate lip movements. Attempts to create character avatars with NeRF-based methods, such as [10, 51, 52, 57], cannot be directly driven by speech. These methods only use audio as a condition, without a clear concept of sync, and usually result in average lip movement. Additionally, previous methods lack control over facial expressions, being limited to controlling blinking only, and cannot model actions like raising eyebrows or frowning. Furthermore, these methods have a significant issue with unstable head poses, leading to the separation between the head and the torso." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In contrast, we use the Face-Sync Controller to model the relationship between audio and lip movements, thereby enhancing the synchronization of lip movements and expressions, and the Head-Sync Stabilizer to stabilize head poses. By addressing these synchronization \"devils\", our method improves visual quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.411, + 0.591, + 0.427 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.436, + 0.611, + 0.451 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.581 + ], + "angle": 0, + "content": "In this section, we introduce the proposed SyncTalk, as shown in Fig. 2. SyncTalk mainly consists of 3 parts: a) lip movements and facial expressions controlled by the FaceSync Controller, b) stable head pose provided by the HeadSync Stabilizer, and c) high-synchronization facial frames rendered by the Dynamic Portrait Renderer. We will describe the content of these three parts in detail in the following subsections." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.591, + 0.7, + 0.607 + ], + "angle": 0, + "content": "3.2. Face-Sync Controller" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Audio-Visual Encoder. Existing methods based on NeRF utilize pre-trained models such as DeepSpeech [2], Wav2Vec 2.0 [3], or HuBERT [16]. These are audio feature extraction methods designed for speech recognition tasks. Using an audio encoder designed for Automatic Speech Recognition (ASR) tasks does not truly reflect lip movements. This is because the pre-trained model is based on the distribution of features from audio to text, whereas we need the feature distribution from audio to lip movements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Considering the above, we use an audio and visual synchronization audio encoder trained on the 2D audio-visual synchronization dataset LRS2 [1]. This ensures that the audio features extracted by our method and lip movements have the same feature distribution. The specific implementation method is as follows: We use a pre-trained lip synchronization discriminator [8]. It can give confidence for the lip synchronization effect of the video. The lip synchronization discriminator takes as input a continuous face window \\( F \\) and the corresponding audio frame \\( A \\). They are" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "668" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "judged as positive samples (with label \\( y = 1 \\)) if they overlap entirely. Otherwise, they are judged as negative samples (with label \\( y = 0 \\)). The discriminator calculates the cosine similarity between these sequences as:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.162, + 0.469, + 0.196 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s i m} (F, A) = \\frac {F \\cdot A}{\\| F \\| _ {2} \\| A \\| _ {2}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.199, + 0.343, + 0.214 + ], + "angle": 0, + "content": "and then uses binary cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.227, + 0.484, + 0.257 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {s y n c}} = - (y \\log (\\operatorname {s i m} (F, A)) + (1 - y) \\log (1 - \\operatorname {s i m} (F, A))), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.469, + 0.287 + ], + "angle": 0, + "content": "to minimize the distance for synchronized samples and maximize the distance for non-synchronized samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.469, + 0.424 + ], + "angle": 0, + "content": "Under the supervision of the lip synchronization discriminator, we pre-train a highly synchronized audio-visual feature extractor related to lip movements. First, we use convolutional networks to obtain audio features \\(\\mathrm{Conv}(A)\\) and encode facial features \\(\\mathrm{Conv}(F)\\). These features are then concatenated. In the decoding phase, we use stacked convolutional layers to restore facial frames using the operation \\(\\mathrm{Dec}(\\mathrm{Conv}(A) \\oplus \\mathrm{Conv}(F))\\). The \\(L_{1}\\) reconstruction loss during training is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.437, + 0.469, + 0.454 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {r e c o n}} = \\| F - \\operatorname {D e c} (\\operatorname {C o n v} (A) \\oplus \\operatorname {C o n v} (F)) \\| _ {1}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.469, + 0.569 + ], + "angle": 0, + "content": "Simultaneously, we sample synchronized and nonsynchronized segments using lip movement discriminators, and employ a same sync loss as Eq. 2. By minimizing both losses, we train a facial generation network related to audio. After training, we use \\(\\mathrm{Conv}(A)\\) as the lip space extracted from the audio. Ultimately, we obtain our highly synchronized audio-visual encoder related to lip movements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Facial Animation Capturer. It is observed that previous methods based on NeRF [14, 21, 44] could only change blinking and could not model facial expressions accurately. This leads to issues such as rigid facial expressions and incorrect facial details if the trained character has significant facial movements, like squinting, raising eyebrows, or frowning. Considering the need for more synchronized and realistic facial expressions, we add an expression synchronization control module. Specifically, we introduce a 3D facial prior using 52 semantically facial blendshape coefficients [33] represented by \\( B \\) to model the face, as shown in Fig. 3. Because the 3D face model can retain the structure information of face motion, it can reflect the content of facial movements well without causing facial structural distortion. During the training, we first use a sophisticated facial blendshape capture module to capture facial expressions as \\( E(B) \\), and select seven core facial expression control coefficients to control the eyebrow, forehead, and eye areas. They are highly correlated with expression and independent of lip movements. We can synchronize the speaker's expression during the inference process because the facial coefficients are semantically informed." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.695, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.174, + 0.694, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.09, + 0.89, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.174, + 0.89, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.269, + 0.892, + 0.297 + ], + "angle": 0, + "content": "Figure 3. Facial Animation Capturer. We use 3D facial blend-shape coefficients to control the expressions of characters." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.308, + 0.892, + 0.414 + ], + "angle": 0, + "content": "Facial-Aware Masked-Attention. To reduce the mutual interference between lip features and expression features during training, we introduce the Facial-Aware Disentangle Attention module. Building on the region attention vector \\( V \\) [21], we add masks \\( M_{\\mathrm{lip}} \\) and \\( M_{\\mathrm{exp}} \\) to the attention areas for lips and expressions, respectively. Specifically, the new attention mechanisms are given by:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.428, + 0.891, + 0.451 + ], + "angle": 0, + "content": "\\[\nV _ {\\text {l i p}} = V \\odot M _ {\\text {l i p}}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.448, + 0.758, + 0.464 + ], + "angle": 0, + "content": "\\[\nV _ {\\mathrm {e x p}} = V \\odot M _ {\\mathrm {e x p}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.892, + 0.607 + ], + "angle": 0, + "content": "These formulations allow the attention mechanisms to focus solely on their respective parts, thereby reducing entanglement between them. Before the disentanglement, lip movements might induce blinking tendencies and affect hair volume. By introducing the mask module, the attention mechanism can focus on either expressions or lips without affecting other areas, thereby reducing the artifact caused by coupling. Finally, we obtain the disentangled lip feature \\( f_{l} = f_{\\mathrm{lip}} \\odot V_{\\mathrm{lip}} \\) and expression feature \\( f_{e} = f_{\\mathrm{exp}} \\odot V_{\\mathrm{exp}} \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.613, + 0.698, + 0.629 + ], + "angle": 0, + "content": "3.3. Head-Sync Stabilizer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.818 + ], + "angle": 0, + "content": "Head Motion Tracker. The head pose, denoted as \\( p \\), refers to the rotation angle of a person's head in 3D space and is defined by a rotation \\( R \\) and a translation \\( T \\). An unstable head pose can lead to head jitter. To obtain a rough estimate of the head pose, initially, the best focal length is determined through \\( i \\) iterations within a predetermined range. For each focal length candidate, \\( f_{i} \\), the system re-initializes the rotation and translation values. The objective is to minimize the error between the projected landmarks from the 3D Morphable Models (3DMM) [31] and the actual landmarks in the video frame. Formally, the optimal focal length \\( f_{\\mathrm{opt}} \\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.826, + 0.891, + 0.849 + ], + "angle": 0, + "content": "\\[\nf _ {\\text {o p t}} = \\arg \\min _ {f _ {i}} E _ {i} \\left(L _ {2 D}, L _ {3 D} \\left(f _ {i}, R _ {i}, T _ {i}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\( E_{i} \\) represents the Mean Squared Error (MSE) between these landmarks, \\( L_{3D}(f_i,R_i,T_i) \\) represents the projected landmarks from the 3DMM for a given focal length" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "669" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "\\(f_{i}\\), the corresponding rotation and translation parameters \\(R_{i}\\) and \\(T_{i}\\), \\(L_{2D}\\) are the actual landmarks from the video frame. Subsequently, leveraging the optimal focal length \\(f_{\\mathrm{opt}}\\), the system refines the rotation \\(R\\) and translation \\(T\\) parameters for all frames to better align the model's projected landmarks with the actual video landmarks. This refinement process can be mathematically represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.211, + 0.47, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\left(R _ {\\text {o p t}}, T _ {\\text {o p t}}\\right) = \\arg \\min _ {R, T} E \\left(L _ {2 D}, L _ {3 D} \\left(f _ {\\text {o p t}}, R, T\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.248, + 0.47, + 0.324 + ], + "angle": 0, + "content": "where \\( E \\) denotes the MSE metric, between the 3D model's projected landmarks \\( L_{3D} \\) for the optimal focal length \\( f_{\\mathrm{opt}} \\), and the actual 2D landmarks \\( L_{2D} \\) in the video frame. The optimized rotation \\( R_{\\mathrm{opt}} \\) and translation \\( T_{\\mathrm{opt}} \\) are obtained by minimizing this error across all frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.47, + 0.536 + ], + "angle": 0, + "content": "Head Points Tracker. Considering methods based on NeRF and their requirements for inputting head rotation \\( R \\) and translation \\( T \\), previous methods utilize 3DMM-based techniques to extract head poses and generate an inaccurate result. To improve the precision of \\( R \\) and \\( T \\), We use an optical flow estimation model from [43] to track facial keypoints \\( K \\). Specifically, using a pre-trained optical flow estimation model, after obtaining the facial motion optical flow, we use the Laplacian filter to select the keypoints where the most significant flow changes are located and track the motion trajectories of these keypoints in the flow sequence. Through this module, our method ensures a more precise and consistent facial keypoint alignment across all frames, enhancing the accuracy of head pose parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.471, + 0.675 + ], + "angle": 0, + "content": "**Bundle Adjustment.** Given the keypoints and the rough head pose, we introduce a two-stage optimization framework from [14] to enhance the accuracy of keypoints and head pose estimations. In the first stage, we randomly initialize the 3D coordinates of \\(j\\) keypoints and optimize their positions to align with the tracked keypoints on the image plane. This process involves minimizing a loss function \\(L_{\\mathrm{init}}\\), which captures the discrepancy between projected keypoints \\(P\\) and the tracked keypoints \\(K\\), as given by:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.689, + 0.47, + 0.723 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {i n i t}} = \\sum_ {j} \\| P _ {j} - K _ {j} \\| _ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.47, + 0.819 + ], + "angle": 0, + "content": "Subsequently, in the second stage, we embark on a more comprehensive optimization to refine the 3D keypoints and the associated head jointly pose parameters. Through the Adam Optimization [19], the algorithm adjust the spatial coordinates, rotation angles \\( R \\), and translations \\( T \\) to minimize the alignment error \\( L_{\\mathrm{sec}} \\), expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.834, + 0.469, + 0.867 + ], + "angle": 0, + "content": "\\[\nL _ {\\sec} = \\sum_ {j} \\| P _ {j} (R, T) - K _ {j} \\| _ {2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "After these optimizations, the resultant head pose and translation parameters are observed to be smooth and stable." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.75, + 0.108 + ], + "angle": 0, + "content": "3.4. Dynamic Portrait Rendering" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.114, + 0.894, + 0.282 + ], + "angle": 0, + "content": "Tri-Plane Hash Representation. Using a collection of multi-perspective images along with corresponding camera poses, NeRF [27] harnesses these resources to manifest a 3D static scene. This representation utilizes an implicit function \\(\\mathcal{F}\\), delineated as \\(\\mathcal{F}:(\\mathrm{x},\\mathrm{d})\\to (\\mathrm{c},\\sigma)\\), where the 3D spatial location is given by \\(\\mathrm{x} = (x,y,z)\\), and the direction of viewing is characterized by \\(\\mathrm{d} = (\\theta ,\\phi)\\). The resultant values, \\(\\mathrm{c} = (r,g,b)\\) and \\(\\sigma\\), represent the radiance and density, respectively. The predicted pixel color, denoted by \\(\\hat{C} (\\mathbf{r})\\), intertwined with ray \\(\\mathrm{r}(t) = \\mathrm{o} + t\\mathrm{d}\\) originating from the camera's core position \\(\\mathrm{o}\\), is derivable using:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.291, + 0.892, + 0.327 + ], + "angle": 0, + "content": "\\[\n\\hat {C} (\\mathrm {r}) = \\int_ {t _ {n}} ^ {t _ {f}} \\sigma (\\mathrm {r} (t)) \\cdot \\mathrm {c} (\\mathrm {r} (t), \\mathrm {d}) \\cdot T (t) d t, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.33, + 0.892, + 0.436 + ], + "angle": 0, + "content": "where \\( t_n \\) and \\( t_f \\) are the near and far bounds, and \\( T(t) \\) is the accumulated transmittance. Addressing the challenges of hash collisions and optimizing audio feature processing, we incorporate three uniquely oriented 2D hash grids [21]. A coordinate, given by \\( \\mathbf{x} = (x,y,z) \\in \\mathbb{R}^{\\mathrm{XYZ}} \\), undergoes an encoding transformation for its projected values via three individual 2D-multiresolution hash encoders [30]:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.444, + 0.892, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} ^ {\\mathrm {A B}}: (a, b) \\rightarrow \\mathrm {f} _ {a b} ^ {\\mathrm {A B}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.469, + 0.892, + 0.561 + ], + "angle": 0, + "content": "where the output \\( \\mathsf{f}_{ab}^{\\mathrm{AB}} \\in \\mathbb{R}^{LD} \\), with the number of levels \\( L \\) and feature dimensions per entry \\( D \\), signifies the planar geometric feature corresponding to the projected coordinate \\( (a,b) \\) and \\( \\mathcal{H}^{\\mathrm{AB}} \\) denotes the multiresolution hash encoder for plane \\( \\mathbb{R}^{\\mathrm{AB}} \\). By fusing the outcomes, the conclusive geometric feature \\( \\mathsf{fg} \\in \\mathbb{R}^{3 \\times LD} \\) is derived as:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.575, + 0.892, + 0.594 + ], + "angle": 0, + "content": "\\[\n\\mathrm {f} _ {\\mathrm {x}} = \\mathcal {H} ^ {\\mathrm {X Y}} (x, y) \\oplus \\mathcal {H} ^ {\\mathrm {Y Z}} (y, z) \\oplus \\mathcal {H} ^ {\\mathrm {X Z}} (x, z), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.6, + 0.892, + 0.674 + ], + "angle": 0, + "content": "where the concatenation of features is symbolized by \\(\\oplus\\), resulting in a \\(3 \\times LD\\)-channel vector. Employing \\(\\mathrm{f_x}\\), the perspective direction d, the lip feature \\(f_l\\), and the expression feature \\(f_e\\), the tri-plane hash's implicit function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.684, + 0.892, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} ^ {\\mathcal {H}}: (\\mathrm {x}, \\mathrm {d}, f _ {l}, f _ {e}; \\mathcal {H} ^ {3}) \\rightarrow (\\mathrm {c}, \\sigma), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.71, + 0.89, + 0.74 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}^3\\) amalgamates the triad of planar hash encoders, as illustrated in Eq. 10." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.74, + 0.892, + 0.861 + ], + "angle": 0, + "content": "Our training employs a two-step coarse-to-fine strategy, initially using MSE loss to assess the difference between predicted \\(\\hat{C} (\\mathbf{r})\\) and actual image colors \\(C(\\mathbf{r})\\). Recognizing MSE's limitations in detail capture, we advance to a refinement stage, incorporating LPIPS loss for enhanced detail, similar to ER-NeRF [21]. We extract random patches \\(\\mathcal{P}\\) from the image, combining LPIPS (weighted by \\(\\lambda\\)) with MSE for improved detail representation, as shown:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.873, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\sum_ {\\mathrm {r}} \\| C (\\mathrm {r}) - \\hat {C} (\\mathrm {r}) \\| _ {2} + \\lambda \\times \\mathcal {L} _ {\\mathrm {L P I P S}} (\\hat {\\mathcal {P}}, \\mathcal {P}). \\tag {13}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "670" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.078, + 0.894, + 0.292 + ], + "angle": 0, + "content": "
MethodsPSNR ↑LPIPS ↓MS-SSIM ↑FID ↓NIQE ↓BRISQUE ↓LMD ↓AUE ↓LSE-C ↑
GANWav2Lip (ACM MM 20 [34])33.43850.06970.978116.022814.536744.26594.96302.90299.2387
VideoReTalking (SIGGRAPH Asia 22 [7])31.79230.04880.96809.206314.241043.04655.85753.33087.9683
DINet (AAAI 23 [50])31.64750.04430.96409.430014.685040.36504.37253.68756.5653
TalkLip (CVPR 23 [41])32.51540.07820.969718.499714.638546.67175.86052.95795.9472
IP-LAP (CVPR 23 [53])35.15250.04430.98038.212514.640042.07503.33502.84004.9541
NeRFAD-NeRF (ICCV 21 [14])26.72910.15360.911128.986214.909155.46672.99955.54814.4996
RAD-NeRF (arXiv 22 [38])31.77540.07780.94528.657013.443344.68922.91155.09585.5219
GeneFace (ICLR 23 [44])24.81650.11780.875321.708413.335346.50614.28595.45275.1950
ER-NeRF (ICCV 23 [21])32.52160.03340.95015.293613.704834.73612.81374.18735.7749
SyncTalk (w/o Portrait)35.35420.02350.97693.924713.133333.29542.57142.57968.1331
SyncTalk (Portrait)37.40170.01130.98412.707014.216537.30422.50433.20748.0263
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.302, + 0.893, + 0.331 + ], + "angle": 0, + "content": "Table 1. The quantitative results of the head reconstruction. \"Portrait\" refers to the use of the Portrait-Sync Generator. We achieve state-of-the-art performance on most metrics. We highlight best and second-best results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.344, + 0.473, + 0.542 + ], + "angle": 0, + "content": "Portrait-Sync Generator. During the training process, to address NeRF's limitations in capturing fine details like hair strands and dynamic backgrounds, we introduce a Portrait-Sync Generator with two key sections. First, NeRF renders the face area \\((F_r)\\), creates \\(G(F_r)\\) through Gaussian blur, and then uses our synchronized head pose to be able to merge with the original image \\((F_o)\\) to enhance hair detail fidelity. Second, when the head and torso are combined, if the character in the source video speaks while the generated face is silent, a dark gap area might appear, as shown in Fig. 5 (b). We fill these areas with the average neck color \\((C_n)\\). This approach results in more realistic details and improved visual quality through the Portrait-Sync Generator." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.21, + 0.577 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.586, + 0.285, + 0.603 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.47, + 0.716 + ], + "angle": 0, + "content": "Dataset. To ensure a fair comparison, we use the same well-edited video sequences from [14, 21, 44], including English and French. The average length of these videos is approximately 8,843 frames, and each video is recorded at 25 FPS. Except for the video from AD-NeRF [14], which has a resolution of \\(450 \\times 450\\), all other videos have a resolution of \\(512 \\times 512\\), with the character-centered." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Comparison Baselines. We compare our method with five GAN-based methods, including Wav2Lip [34], Video-ReTalking [7], DINet [50], TalkLip [41], and IP-LAP [53], and NeRF-based methods such as AD-NeRF [14], RAD-NeRF [38], GeneFace [44], and ER-NeRF [21]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. In the coarse stage, the portrait head is trained for 100,000 iterations and 25,000 in the fine stage, sampling \\(256^2\\) rays per iteration using a 2D hash encoder (\\(L = 14\\), \\(F = 1\\)). We employ the AdamW optimizer [23], with learning rates of 0.01 for the hash encoder and 0.001 for other modules. Total training time is approximately 2 hours on an NVIDIA RTX 3090 GPU." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.341, + 0.895, + 0.466 + ], + "angle": 0, + "content": "
MethodsAudio AAudio B
LSE-D ↓LSE-C ↑LSE-D ↓LSE-C ↑
DINet (AAAI 23 [50])8.50315.69568.20385.1134
TalkLip (CVPR 23 [41])8.76155.74498.70195.5359
IP-LAP (CVPR 23 [53])9.80373.85789.11024.389
GeneFace (ICLR 23 [44])9.54514.29339.66753.7342
ER-NeRF (ICCV 23 [21])11.8132.407610.73383.0242
SyncTalk (Ours)7.72116.66598.02486.2596
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.474, + 0.893, + 0.517 + ], + "angle": 0, + "content": "Table 2. The quantitative results of the lip synchronization. We use two different audio samples to drive the same subject, then highlight best and second-best results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.53, + 0.722, + 0.546 + ], + "angle": 0, + "content": "4.2. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Full Reference Quality Assessment. In terms of image quality, we use full reference metrics such as Peak Signal-to-Noise Ratio (PSNR), Learned Perceptual Image Patch Similarity (LPIPS) [46], Multi-Scale Structure Similarity (MS-SSIM), and Frechet Inception Distance (FID) [15] as evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.735 + ], + "angle": 0, + "content": "No Reference Quality Assessment. In high PSNR images, texture details may not align with human visual perception [47]. For more precise output definition and comparison, we use two No Reference methods: the Natural Image Quality Evaluator (NIQE)[29] and the Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE)[28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.736, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Synchronization Assessment. For synchronization, we use landmark distance (LMD) to measure the synchronicity of facial movements, action units error (AUE) [4] to assess the accuracy of facial movements, and introduce Lip Sync Error Confidence (LSE-C), consistent with Wav2Lip [34], to evaluate the synchronization between lip movements and audio." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Evaluation Results. The evaluation results of the head reconstruction are shown in Tab. 1. We compare recent methods based on GAN and NeRF. It can be observed that our image quality is superior to other methods in all aspects. In terms of synchronization, our results surpass most methods." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "671" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.074, + 0.895, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.476, + 0.893, + 0.504 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of facial synthesis by different methods. Our method has the best visual effect on lip movements and facial expressions without the problem of separation of head and torso. Please zoom in for better visualization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.473, + 0.79 + ], + "angle": 0, + "content": "We compare the two output modes of SyncTalk, one processed through the Portrait-Sync Generator and one without it. After processing through the Portrait-Sync Generator, hair details are restored, and the image quality is improved. Since we can maintain the subject's identity well, we surpass GAN-based methods in image quality. Thanks to the sync of lip, expression, and pose, we also outperform NeRF-based methods in image quality. Especially in terms of the LPIPS metric, our method has three times improvement compared to the previous state-of-the-art method ER-NeRF [21]. We compare the latest SOTA method drivers using out-of-distribution (OOD) audio, and the results are shown in Tab. 2. We introduce Lip Synchronization Error Distance (LSE-D) and Confidence (LSEC) for lip-audio sync evaluation, aligning with [34]. Our method shows state-of-the-art lip synchronization, overcoming small-sample NeRF limitations by incorporating a pre-trained audio-visual encoder for lip modeling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We also test the rendering speed. On an NVIDIA RTX 3090 GPU, and having preloaded the data onto the GPU, the head is outputted at 52 FPS with a resolution of \\(512 \\times 512\\). In Portrait mode, using the Portrait-Sync Generator, we achieve rendering speeds of 50 FPS with our CUDA acceleration. This far exceeds the video input speed of 25 FPS, allowing for real-time generation of video streams." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.518, + 0.71, + 0.533 + ], + "angle": 0, + "content": "4.3. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.804 + ], + "angle": 0, + "content": "Evaluation Results. To more intuitively evaluate image quality, we display a comparison between our method and other methods in Fig. 4. In this figure, it can be observed that SyncTalk demonstrates more precise and more accurate facial details. Compared to Wav2Lip [34], our method better preserves the subject's identity while offering higher fidelity and resolution. Against IP-LAP [53], our method excels in lip shape synchronization, primarily due to the audio-visual consistency brought by the audio-visual encoder. Compared to GeneFace [44], our method can accurately reproduce actions such as blinking and eyebrow-raising through expression sync. In contrast to ER-NeRF [21], our method avoids the separation between the head and body through the Pose-Sync Stabilizer and generates more accurate lip shapes. Our method achieves the best overall visual effect; we recommend watching the supplementary video for comparison." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.893, + 0.903 + ], + "angle": 0, + "content": "User Study. To provide a more comprehensive evaluation of the proposed model, we design an exhaustive user study questionnaire. We extract 24 video clips, each lasting more than 10 seconds, which include head poses, facial expressions, and lip movements. Each method is represented by three of these clips. We invite 35 participants" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "672" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.09, + 0.887, + 0.189 + ], + "angle": 0, + "content": "
Wav2Lip [34]DINet [50]TalkLip [41]IP-LAP [53]AD-NeRF [14]GeneFace [44]ER-NeRF [21]SyncTalk
Lip-sync Accuracy3.8393.6962.8933.1612.6962.9823.1894.304
Exp-sync Accuracy3.5363.4822.6073.4112.2503.0362.9464.036
Pose-sync Accuracy3.5713.5712.8753.6962.2322.9292.6073.980
Image Quality2.5002.6962.0543.5712.4643.4823.0364.054
Video Realness2.9292.4292.4293.1612.0362.7322.5184.018
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.2, + 0.893, + 0.228 + ], + "angle": 0, + "content": "Table 3. User Study. Rating is on a scale of 1-5; the higher the better. The term \"Exp-sync Accuracy\" is an abbreviation for \"Expression-sync Accuracy\". We highlight best and second-best results." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.24, + 0.466, + 0.417 + ], + "angle": 0, + "content": "
PSNR ↑LPIPS ↓LMD ↓
Ours37.3110.01212.8032
replace Audio-Visual Encoder with Hubert [16]33.5160.02763.3961
replace Facial Animation Capture with ER-NeRF [21]'s30.2730.04153.0516
w/o Facial-Aware Masked-Attention36.5360.01652.9139
w/o Head-Sync Stabilizer28.9840.06343.5373
w/o Portrait-Sync Generator32.2390.03952.8154
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.428, + 0.47, + 0.456 + ], + "angle": 0, + "content": "Table 4. Ablation study for our components. We show the PSNR, LPIPS, and LMD in different cases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.684 + ], + "angle": 0, + "content": "to provide scores. The questionnaire is designed using the Mean Opinion Score (MOS) scoring protocol, asking participants to rate the generated videos from five perspectives: (1) Lip-sync Accuracy, (2) Expression-sync Accuracy, (3) Pose-sync Accuracy, (4) Image Quality, and (5) Video Realness. On average, participants take 19 minutes to complete the questionnaire, with a standardized Cronbach \\(\\alpha\\) coefficient of 0.96. The results of the User Study are displayed in Tab. 3. SyncTalk surpasses previous methods in all evaluations. Furthermore, we achieve the highest score in video authenticity, surpassing the second-place method, IP-LAP [53], by a margin of \\(20\\%\\). User studies indicate that our method can generate visually excellent quality as perceived by humans, achieving high realism." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.231, + 0.712 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.469, + 0.794 + ], + "angle": 0, + "content": "We conduct an ablation study to examine the contributions of different components in our model. We select three core metrics for evaluation: PSNR, LPIPS, and LMD. We select a subject named \"May\" for testing, and the results are presented in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The Audio-Visual Encoder provides the primary lip sync information. When this module is replaced, all three metrics become worse, with the LMD error increasing by \\(21.15\\%\\) in particular, indicating decreased lip motion synchronization, as shown in Fig. 5 (a), and showing that our audio-visual encoders can extract accurate lip features. Replacing Facial Animation Capture with ER-NeRF [21]'s blink module" + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.24, + 0.89, + 0.25 + ], + "angle": 0, + "content": "w/o Audio-Visual Encoder w/ Audio-Visual Encoder w/o Head-Sync Stabilizer w/ Head-Sync Stabilizer" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.252, + 0.892, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.343, + 0.892, + 0.372 + ], + "angle": 0, + "content": "Figure 5. Ablation Study on Audio-Visual Encoder and Head-Sync Stabilizer. Removing them will lead to \\((a)\\) and \\((b)\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.389, + 0.81, + 0.404 + ], + "angle": 0, + "content": "affects eyebrow movements and image quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.405, + 0.892, + 0.541 + ], + "angle": 0, + "content": "The Facial-Aware Masked-Attention mainly alleviates motion entanglement between the lips and the rest of the face, slightly affecting the image quality after removal. Without the Head-Sync Stabilizer, all metrics significantly declined, notably LPIPS, leading to head pose jitter and head and torso separation, as shown in Fig. 5 (b). The Portrait-Sync Generator restores details like hair. Removing this module impacts the restoration of details like hair, leading to noticeable segmentation boundaries." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.559, + 0.619, + 0.574 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.766 + ], + "angle": 0, + "content": "In this paper, we introduce SyncTalk, a highly synchronized NeRF-based method for realistic speech-driven talking head synthesis. Our framework includes a Facial Sync Controller, Head Sync Stabilizer, and Portrait Sync Generator, which maintain subject identity and generate synchronized lip movements, facial expressions, and stable head poses. Through extensive evaluation, SyncTalk demonstrates superior performance in creating realistic and synchronized talking head videos compared to existing methods. We expect that SyncTalk will not only enhance various applications but also inspire further innovation in the field of talking head synthesis." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.659, + 0.801 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported by National Natural Science Foundation of China (NSFC) under Grant Nos. 62172421 and 62072459, the Outstanding Innovative Talents Cultivation Funded Programs 2023 of Renmin University of China, and the Public Computing Cloud at Renmin University of China." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "673" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence, 44(12):8717-8727, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Dario Amodei, Sundaram Ananthanarayanan, Rishita Anubhai, Jingliang Bai, Eric Battenberg, Carl Case, Jared Casper, Bryan Catanzaro, Qiang Cheng, Guoliang Chen, et al. Deep speech 2: End-to-end speech recognition in english and mandarin. In International conference on machine learning, pages 173-182. PMLR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.314 + ], + "angle": 0, + "content": "[3] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.471, + 0.386 + ], + "angle": 0, + "content": "[4] Tadas Baltrusaitis, Marwa Mahmoud, and Peter Robinson. Cross-dataset learning and person-specific normalisation for automatic action unit detection. In 2015 11th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-6. IEEE, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.471, + 0.443 + ], + "angle": 0, + "content": "[5] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European conference on computer vision (ECCV), pages 520-535, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.471, + 0.514 + ], + "angle": 0, + "content": "[6] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7832-7841, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.471, + 0.586 + ], + "angle": 0, + "content": "[7] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.588, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[8] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.659, + 0.471, + 0.742 + ], + "angle": 0, + "content": "[9] Dipanjan Das, Sandika Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 408-424. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.745, + 0.471, + 0.814 + ], + "angle": 0, + "content": "[10] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.816, + 0.471, + 0.87 + ], + "angle": 0, + "content": "[11] Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[12] Shreyank N Gowda, Dheeraj Pandey, and Shashank Narayana Gowda. From pixels to portraits:" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "A comprehensive survey of talking head generation techniques and applications. arXiv preprint arXiv:2308.16041, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.894, + 0.231 + ], + "angle": 0, + "content": "[13] Jiazhi Guan, Zhanwang Zhang, Hang Zhou, Tianshu Hu, Kaisiyuan Wang, Dongliang He, Haocheng Feng, Jingtuo Liu, Errui Ding, Ziwei Liu, et al. Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1505-1515, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.895, + 0.302 + ], + "angle": 0, + "content": "[14] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.894, + 0.37 + ], + "angle": 0, + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[16] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[17] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[18] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM transactions on graphics (TOG), 37(4): 1-14, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[20] Prajwal KR, Rudrabha Mukhopadhyay, Jerin Philip, Abhishek Jha, Vinay Namboodiri, and CV Jawahar. Towards automatic face-to-face translation. In Proceedings of the 27th ACM international conference on multimedia, pages 1428-1436, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[21] Jiahe Li, Jiawei Zhang, Xiao Bai, Jun Zhou, and Lin Gu. Efficient region-aware neural radiance fields for high-fidelity talking portrait synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7568-7578, 2023. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[22] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. In European Conference on Computer Vision, pages 106-125. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[23] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "674" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[25] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[26] Moustafa Meshry, Saksham Suri, Larry S Davis, and Abhinav Shrivastava. Learned spatial representations for few-shot talking-head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13829-13838, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.471, + 0.313 + ], + "angle": 0, + "content": "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.471, + 0.37 + ], + "angle": 0, + "content": "[28] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing, 21(12): 4695-4708, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.471, + 0.412 + ], + "angle": 0, + "content": "[29] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[30] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.471, + 0.538 + ], + "angle": 0, + "content": "[31] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[32] Ziqiao Peng, Yihao Luo, Yue Shi, Hao Xu, Xiangyu Zhu, Hongyan Liu, Jun He, and Zhaoxin Fan. Selftalk: A self-supervised commutative training diagram to comprehend 3d talking faces. In Proceedings of the 31st ACM International Conference on Multimedia, pages 5292-5301, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.471, + 0.678 + ], + "angle": 0, + "content": "[33] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20687-20697, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.471, + 0.747 + ], + "angle": 0, + "content": "[34] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[35] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In European Conference on Computer Vision, pages 666-682. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.805, + 0.471, + 0.859 + ], + "angle": 0, + "content": "[36] Linsen Song, Wayne Wu, Chen Qian, Ran He, and Chen Change Loy. Everybody's talkin': Let me talk as you want. IEEE Transactions on Information Forensics and Security, 17:585-598, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[37] Yasheng Sun, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Zhibin Hong, Jingtuo Liu, Errui Ding, Jingdong Wang, Zwei Liu, and Koike Hideki. Masked lip-sync prediction by" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "audio-visual contextual exploitation in transformers. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[38] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.894, + 0.276 + ], + "angle": 0, + "content": "[39] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVI 16, pages 716–731. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[40] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. International Journal of Computer Vision, 128:1398-1413, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.894, + 0.39 + ], + "angle": 0, + "content": "[41] Jiadong Wang, Xinyuan Qian, Malu Zhang, Robby T Tan, and Haizhou Li. Seeing what you said: Talking face generation guided by a lip reading expert. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14653-14662, 2023. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[42] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[43] Shunyu Yao, RuiZhe Zhong, Yichao Yan, Guangtao Zhai, and Xiaokang Yang. Dfa-nerf: Personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[44] Zhenhui Ye, Ziyue Jiang, Yi Ren, Jinglin Liu, JinZheng He, and Zhou Zhao. Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430, 2023. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.563, + 0.892, + 0.632 + ], + "angle": 0, + "content": "[45] Chenxu Zhang, Yifan Zhao, Yifei Huang, Ming Zeng, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Facial: Synthesizing dynamic talking face with implicit attribute learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3867-3876, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[47] Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. Ranksrgan: Generative adversarial networks with ranker for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3096-3105, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[48] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[49] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "675" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661-3670, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[50] Zhimeng Zhang, Zhipeng Hu, Wenjin Deng, Changjie Fan, Tangjie Lv, and Yu Ding. Dinet: Deformation inpainting network for realistic face visually dubbing on high resolution video. arXiv preprint arXiv:2303.03988, 2023. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[51] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[52] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21057-21067, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.401 + ], + "angle": 0, + "content": "[53] Weizhi Zhong, Chaowei Fang, Yinqi Cai, Pengxu Wei, Gangming Zhao, Liang Lin, and Guanbin Li. Identity-preserving talking face generation with landmark and appearance priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2023. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[54] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI conference on artificial intelligence, pages 9299–9306, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.469, + 0.542 + ], + "angle": 0, + "content": "[55] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.545, + 0.469, + 0.599 + ], + "angle": 0, + "content": "[56] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG), 39(6):1-15, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[57] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4574-4584, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "676" + } + ] +] \ No newline at end of file diff --git a/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/89645a31-8829-4d8f-939e-0d17fcee95a5_origin.pdf b/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/89645a31-8829-4d8f-939e-0d17fcee95a5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a209502395d461d901fac9391bd755a5c7eaefa1 --- /dev/null +++ b/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/89645a31-8829-4d8f-939e-0d17fcee95a5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:422dbb0b84eff7d06a4929c88488db383c3931aa59f1b48e6415a730f203df1d +size 2510146 diff --git a/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/full.md b/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e6c64524ac196fb232c971377195c891ac880908 --- /dev/null +++ b/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/full.md @@ -0,0 +1,320 @@ +# SyncTalk: The Devil is in the Synchronization for Talking Head Synthesis + +Ziqiao Peng $^{1}$ Wentao Hu $^{2}$ Yue Shi $^{3}$ Xiangyu Zhu $^{4}$ Xiaomei Zhang $^{4}$ Hao Zhao $^{5}$ Jun He $^{1}$ Hongyan Liu $^{5*}$ Zhaoxin Fan $^{1*}$ + +$^{1}$ Renmin University of China $^{2}$ Beijing University of Posts and Telecommunications $^{3}$ Psyche AI Inc. $^{4}$ Chinese Academy of Sciences $^{5}$ Tsinghua University + +![](images/785227c2823cf3d039b62977e4be371e69d0f11158259451da2e2a7fca027ef6.jpg) +Figure 1. The proposed SyncTalk synthesizes synchronized talking head videos, employing tri-plane hash representations to maintain subject identity. It can generate synchronized lip movements, facial expressions, and stable head poses, and restores hair details to create high-resolution videos. + +# Abstract + +Achieving high synchronization in the synthesis of realistic, speech-driven talking head videos presents a significant challenge. Traditional Generative Adversarial Networks (GAN) struggle to maintain consistent facial identity, while Neural Radiance Fields (NeRF) methods, although they can address this issue, often produce mismatched lip movements, inadequate facial expressions, and unstable head poses. A lifelike talking head requires synchronized coordination of subject identity, lip movements, facial expressions, and head poses. The absence of these synchronizations is a fundamental flaw, leading to unrealistic and artificial outcomes. To address the critical issue of synchronization, identified as the "devil" in creating realistic talking heads, we introduce SyncTalk. This NeRF-based method effectively maintains subject identity, enhancing synchronization and realism in talking head synthesis. SyncTalk employs a Face-Sync Controller to align lip movements + +with speech and innovatively uses a 3D facial blendshape model to capture accurate facial expressions. Our Head-Sync Stabilizer optimizes head poses, achieving more natural head movements. The Portrait-Sync Generator restores hair details and blends the generated head with the torso for a seamless visual experience. Extensive experiments and user studies demonstrate that SyncTalk outperforms state-of-the-art methods in synchronization and realism. We recommend watching the supplementary video: https://ziqiaopeng.github.io/synctalk + +# 1. Introduction + +The quest to generate dynamic and realistic speech-driven talking heads has intensified, driven by expanding applications in digital assistants [39], virtual reality [32], and film-making [18]. The ultimate goal is to enhance the realism of synthetic videos to align with human perceptual expectations. However, a fundamental challenge that persists across existing methods is the need for synchroniza + +tion. Traditional methods based on Generative Adversarial Networks (GAN) [13, 41, 50, 53], while adept at modeling the lip movements of speakers, often produce inconsistent identities across different frames, leading to issues such as different tooth sizes and fluctuating lip thickness. Similarly, emerging technologies based on Neural Radiance Fields (NeRF) [14, 21, 35, 43, 44] excel in maintaining identity consistency and preserving facial details. However, they struggle with mismatches in lip movements, challenging facial expression control, and unstable head poses, thereby diminishing the overall realism of the video. + +In this paper, we find that the "devil" is in the synchronization. Existing methods need more synchronization in four key areas: subject identity, lip movements, facial expressions, and head poses. Firstly, in GAN-based methods, maintaining the subject's identity in the video is challenging due to the instability of features in consecutive frames and the use of only a few frames as references for facial reconstruction [34]. Secondly, lip movements are not synchronized with speech. In NeRF-based methods, audio features trained only on a 5-minute speech dataset struggle to generalize to different speech inputs [44]. Thirdly, there is a lack of facial expression control, with most methods only producing lip movements or controlling blinking, resulting in unnatural facial actions [12]. Fourthly, the head pose is not synchronized. Previous methods relied on sparse landmarks to compute projection error, but jitter and inaccuracy in these landmarks lead to unstable head poses [49], as shown in Fig. 1. These synchronization issues introduce artifacts and significantly reduce realism. + +To address these synchronization challenges, we introduce SyncTalk, a NeRF-based method focused on highly synchronized, realistic, speech-driven talking head synthesis, employing tri-plane hash representations to maintain subject identity. Through the Face-Sync Controller and Head-Sync Stabilizer, SyncTalk significantly enhances the synchronization and visual quality of the synthesized videos. Visual quality is further improved by the Portrait-Sync Generator, which meticulously refines visual details. The entire rendering process can achieve 50 FPS and outputs high-resolution videos. + +Within the Face-Sync Controller, we pre-train an audio-visual encoder on the 2D audio-visual dataset, resulting in a generalized representation that ensures synchronized lip movements across different speech samples. For controlling facial expressions, we employ a semantically enriched 3D facial blendshape model [33]. This model is distinguished by its ability to control specific facial expression regions through 52 parameters. Regarding the Head-Sync Stabilizer, we use a head motion tracker [14] to infer the head's rough rotation and translation parameters. Due to the instability of the rough parameters, inspired by Simultaneous Localization and Mapping (SLAM), we incorporate a head + +point tracker to track dense keypoints and integrate a bundle adjustment method to optimize the head pose, thereby achieving stable and continuous head motion. To further enhance the visual fidelity of SyncTalk, we have designed a Portrait-Sync Generator. This module repairs artifacts in the NeRF modeling, particularly the intricate details of hair and background flaws, and outputs high-resolution video. + +In summary, the main contributions of our work are as follows: + +- We present a Face-Sync Controller that utilizes an Audio-Visual Encoder in conjunction with a Facial Animation Capturer, ensuring accurate lip synchronization and dynamic facial expression rendering. +- We introduce a Head-Sync Stabilizer that tracks head rotation and facial movement keypoints. Utilizing the bundle adjustment method, this Stabilizer guarantees smooth and synchronous head motion. +- We design a Portrait-Sync Generator that improves visual fidelity by repairing artifacts in NeRF modeling and refining intricate details like hair and background in high-resolution videos. + +# 2. Related Work + +# 2.1. GAN-based Talking Head Synthesis + +Recently, GAN-based talking head synthesis [5, 6, 9, 20, 26, 36, 40, 54, 55] has emerged as an essential research area in computer vision, but they struggle to maintain the identity of the subject in videos consistently. A cluster of noteworthy techniques, including [13, 34, 37, 41, 50, 53], mainly focus on generating video streams for the lip region. These methods create new visuals for talking head portraits by changing the lip region. For instance, Wav2Lip [34] introduces a lip sync expert for supervising lip movements. However, due to the use of five frames from the reference frame to reconstruct the lip, it struggles to maintain the subject's identity. In contrast, methods like [6, 24, 42, 56] perform full-face synthesis but struggle to ensure sync between facial expressions and head poses. + +Apart from video stream techniques, efforts have also been made to enable a single image to "speak" using speech, as demonstrated in [17, 45, 48]. For example, SadTalker [48] can generate videos of a person speaking from a single image. However, such methods fail to generate natural head poses and facial expressions and struggle to maintain the subject's identity, affecting the sync effect and leading to an unrealistic visual perception. + +Compared to these methods, SyncTalk uses NeRF to perform three-dimensional modeling of the face. Its capability to represent continuous 3D scenes in canonical spaces translates to exceptional performance in maintaining subject identity consistency and detail preservation. + +![](images/22c07d340564f4ba4e434acfa89e395806ce39fb6903ee6560081d815f96a98e.jpg) +Figure 2. Overview of SyncTalk. Given a cropped reference video of a talking head and the corresponding speech, SyncTalk can extract the Lip Feature $f_{l}$ , Expression Feature $f_{e}$ , and Head Pose $(R, T)$ through two synchronization modules $(a)$ and $(b)$ . The Tri-Plane Hash Representation then models the head, outputting a rough speech-driven video. The Portrait-Sync Generator further restores details such as hair and background, ultimately producing a high-resolution talking head video. + +# 2.2. NeRF-based Talking Head Synthesis + +With the recent rise of NeRF, numerous fields have begun to utilize it to tackle related challenges [11, 25]. Previous work [14, 22, 43, 44] has integrated NeRF into the task of synthesizing talking heads and has used audio as the driving signal, but these methods are all based on the vanilla NeRF model. For instance, AD-NeRF [14] requires approximately 10 seconds to render a single image. RAD-NeRF [38] aims for real-time video generation and employs a NeRF based on Instant-NGP [30]. ER-NeRF [21] innovatively introduces triple-plane hash encoders to trim the empty spatial regions, advocating for a compact and accelerated rendering approach. GeneFace [44] attempts to reduce NeRF artifacts by translating speech features into facial landmarks, but this often results in inaccurate lip movements. Attempts to create character avatars with NeRF-based methods, such as [10, 51, 52, 57], cannot be directly driven by speech. These methods only use audio as a condition, without a clear concept of sync, and usually result in average lip movement. Additionally, previous methods lack control over facial expressions, being limited to controlling blinking only, and cannot model actions like raising eyebrows or frowning. Furthermore, these methods have a significant issue with unstable head poses, leading to the separation between the head and the torso. + +In contrast, we use the Face-Sync Controller to model the relationship between audio and lip movements, thereby enhancing the synchronization of lip movements and expressions, and the Head-Sync Stabilizer to stabilize head poses. By addressing these synchronization "devils", our method improves visual quality. + +# 3. Method + +# 3.1. Overview + +In this section, we introduce the proposed SyncTalk, as shown in Fig. 2. SyncTalk mainly consists of 3 parts: a) lip movements and facial expressions controlled by the FaceSync Controller, b) stable head pose provided by the HeadSync Stabilizer, and c) high-synchronization facial frames rendered by the Dynamic Portrait Renderer. We will describe the content of these three parts in detail in the following subsections. + +# 3.2. Face-Sync Controller + +Audio-Visual Encoder. Existing methods based on NeRF utilize pre-trained models such as DeepSpeech [2], Wav2Vec 2.0 [3], or HuBERT [16]. These are audio feature extraction methods designed for speech recognition tasks. Using an audio encoder designed for Automatic Speech Recognition (ASR) tasks does not truly reflect lip movements. This is because the pre-trained model is based on the distribution of features from audio to text, whereas we need the feature distribution from audio to lip movements. + +Considering the above, we use an audio and visual synchronization audio encoder trained on the 2D audio-visual synchronization dataset LRS2 [1]. This ensures that the audio features extracted by our method and lip movements have the same feature distribution. The specific implementation method is as follows: We use a pre-trained lip synchronization discriminator [8]. It can give confidence for the lip synchronization effect of the video. The lip synchronization discriminator takes as input a continuous face window $F$ and the corresponding audio frame $A$ . They are + +judged as positive samples (with label $y = 1$ ) if they overlap entirely. Otherwise, they are judged as negative samples (with label $y = 0$ ). The discriminator calculates the cosine similarity between these sequences as: + +$$ +\operatorname {s i m} (F, A) = \frac {F \cdot A}{\| F \| _ {2} \| A \| _ {2}}, \tag {1} +$$ + +and then uses binary cross-entropy loss: + +$$ +L _ {\text {s y n c}} = - (y \log (\operatorname {s i m} (F, A)) + (1 - y) \log (1 - \operatorname {s i m} (F, A))), \tag {2} +$$ + +to minimize the distance for synchronized samples and maximize the distance for non-synchronized samples. + +Under the supervision of the lip synchronization discriminator, we pre-train a highly synchronized audio-visual feature extractor related to lip movements. First, we use convolutional networks to obtain audio features $\mathrm{Conv}(A)$ and encode facial features $\mathrm{Conv}(F)$ . These features are then concatenated. In the decoding phase, we use stacked convolutional layers to restore facial frames using the operation $\mathrm{Dec}(\mathrm{Conv}(A) \oplus \mathrm{Conv}(F))$ . The $L_{1}$ reconstruction loss during training is given by: + +$$ +L _ {\text {r e c o n}} = \| F - \operatorname {D e c} (\operatorname {C o n v} (A) \oplus \operatorname {C o n v} (F)) \| _ {1}. \tag {3} +$$ + +Simultaneously, we sample synchronized and nonsynchronized segments using lip movement discriminators, and employ a same sync loss as Eq. 2. By minimizing both losses, we train a facial generation network related to audio. After training, we use $\mathrm{Conv}(A)$ as the lip space extracted from the audio. Ultimately, we obtain our highly synchronized audio-visual encoder related to lip movements. + +Facial Animation Capturer. It is observed that previous methods based on NeRF [14, 21, 44] could only change blinking and could not model facial expressions accurately. This leads to issues such as rigid facial expressions and incorrect facial details if the trained character has significant facial movements, like squinting, raising eyebrows, or frowning. Considering the need for more synchronized and realistic facial expressions, we add an expression synchronization control module. Specifically, we introduce a 3D facial prior using 52 semantically facial blendshape coefficients [33] represented by $B$ to model the face, as shown in Fig. 3. Because the 3D face model can retain the structure information of face motion, it can reflect the content of facial movements well without causing facial structural distortion. During the training, we first use a sophisticated facial blendshape capture module to capture facial expressions as $E(B)$ , and select seven core facial expression control coefficients to control the eyebrow, forehead, and eye areas. They are highly correlated with expression and independent of lip movements. We can synchronize the speaker's expression during the inference process because the facial coefficients are semantically informed. + +![](images/5d13645c920df6df94817562e51c9a4750f4c4cebed23dfd82ba2e57d0f91e33.jpg) + +![](images/851fad13038e29b3b0cad2a1260d21d3a60c0ebea9fe827e47c45e19b40d2e2e.jpg) +Figure 3. Facial Animation Capturer. We use 3D facial blend-shape coefficients to control the expressions of characters. + +![](images/2611530a845c6638ed5c75306e89a915a09d63b9f136ed59ab86fac44b732a0f.jpg) + +![](images/40f9df9661c8a9dde93ad542361784c8ffb56f9e4ccc38e9561f9c732720c7f7.jpg) + +Facial-Aware Masked-Attention. To reduce the mutual interference between lip features and expression features during training, we introduce the Facial-Aware Disentangle Attention module. Building on the region attention vector $V$ [21], we add masks $M_{\mathrm{lip}}$ and $M_{\mathrm{exp}}$ to the attention areas for lips and expressions, respectively. Specifically, the new attention mechanisms are given by: + +$$ +V _ {\text {l i p}} = V \odot M _ {\text {l i p}}, \tag {4} +$$ + +$$ +V _ {\mathrm {e x p}} = V \odot M _ {\mathrm {e x p}}. +$$ + +These formulations allow the attention mechanisms to focus solely on their respective parts, thereby reducing entanglement between them. Before the disentanglement, lip movements might induce blinking tendencies and affect hair volume. By introducing the mask module, the attention mechanism can focus on either expressions or lips without affecting other areas, thereby reducing the artifact caused by coupling. Finally, we obtain the disentangled lip feature $f_{l} = f_{\mathrm{lip}} \odot V_{\mathrm{lip}}$ and expression feature $f_{e} = f_{\mathrm{exp}} \odot V_{\mathrm{exp}}$ . + +# 3.3. Head-Sync Stabilizer + +Head Motion Tracker. The head pose, denoted as $p$ , refers to the rotation angle of a person's head in 3D space and is defined by a rotation $R$ and a translation $T$ . An unstable head pose can lead to head jitter. To obtain a rough estimate of the head pose, initially, the best focal length is determined through $i$ iterations within a predetermined range. For each focal length candidate, $f_{i}$ , the system re-initializes the rotation and translation values. The objective is to minimize the error between the projected landmarks from the 3D Morphable Models (3DMM) [31] and the actual landmarks in the video frame. Formally, the optimal focal length $f_{\mathrm{opt}}$ is given by: + +$$ +f _ {\text {o p t}} = \arg \min _ {f _ {i}} E _ {i} \left(L _ {2 D}, L _ {3 D} \left(f _ {i}, R _ {i}, T _ {i}\right)\right), \tag {5} +$$ + +where $E_{i}$ represents the Mean Squared Error (MSE) between these landmarks, $L_{3D}(f_i,R_i,T_i)$ represents the projected landmarks from the 3DMM for a given focal length + +$f_{i}$ , the corresponding rotation and translation parameters $R_{i}$ and $T_{i}$ , $L_{2D}$ are the actual landmarks from the video frame. Subsequently, leveraging the optimal focal length $f_{\mathrm{opt}}$ , the system refines the rotation $R$ and translation $T$ parameters for all frames to better align the model's projected landmarks with the actual video landmarks. This refinement process can be mathematically represented as: + +$$ +\left(R _ {\text {o p t}}, T _ {\text {o p t}}\right) = \arg \min _ {R, T} E \left(L _ {2 D}, L _ {3 D} \left(f _ {\text {o p t}}, R, T\right)\right), \tag {6} +$$ + +where $E$ denotes the MSE metric, between the 3D model's projected landmarks $L_{3D}$ for the optimal focal length $f_{\mathrm{opt}}$ , and the actual 2D landmarks $L_{2D}$ in the video frame. The optimized rotation $R_{\mathrm{opt}}$ and translation $T_{\mathrm{opt}}$ are obtained by minimizing this error across all frames. + +Head Points Tracker. Considering methods based on NeRF and their requirements for inputting head rotation $R$ and translation $T$ , previous methods utilize 3DMM-based techniques to extract head poses and generate an inaccurate result. To improve the precision of $R$ and $T$ , We use an optical flow estimation model from [43] to track facial keypoints $K$ . Specifically, using a pre-trained optical flow estimation model, after obtaining the facial motion optical flow, we use the Laplacian filter to select the keypoints where the most significant flow changes are located and track the motion trajectories of these keypoints in the flow sequence. Through this module, our method ensures a more precise and consistent facial keypoint alignment across all frames, enhancing the accuracy of head pose parameters. + +**Bundle Adjustment.** Given the keypoints and the rough head pose, we introduce a two-stage optimization framework from [14] to enhance the accuracy of keypoints and head pose estimations. In the first stage, we randomly initialize the 3D coordinates of $j$ keypoints and optimize their positions to align with the tracked keypoints on the image plane. This process involves minimizing a loss function $L_{\mathrm{init}}$ , which captures the discrepancy between projected keypoints $P$ and the tracked keypoints $K$ , as given by: + +$$ +L _ {\text {i n i t}} = \sum_ {j} \| P _ {j} - K _ {j} \| _ {2}. \tag {7} +$$ + +Subsequently, in the second stage, we embark on a more comprehensive optimization to refine the 3D keypoints and the associated head jointly pose parameters. Through the Adam Optimization [19], the algorithm adjust the spatial coordinates, rotation angles $R$ , and translations $T$ to minimize the alignment error $L_{\mathrm{sec}}$ , expressed as: + +$$ +L _ {\sec} = \sum_ {j} \| P _ {j} (R, T) - K _ {j} \| _ {2}. \tag {8} +$$ + +After these optimizations, the resultant head pose and translation parameters are observed to be smooth and stable. + +# 3.4. Dynamic Portrait Rendering + +Tri-Plane Hash Representation. Using a collection of multi-perspective images along with corresponding camera poses, NeRF [27] harnesses these resources to manifest a 3D static scene. This representation utilizes an implicit function $\mathcal{F}$ , delineated as $\mathcal{F}:(\mathrm{x},\mathrm{d})\to (\mathrm{c},\sigma)$ , where the 3D spatial location is given by $\mathrm{x} = (x,y,z)$ , and the direction of viewing is characterized by $\mathrm{d} = (\theta ,\phi)$ . The resultant values, $\mathrm{c} = (r,g,b)$ and $\sigma$ , represent the radiance and density, respectively. The predicted pixel color, denoted by $\hat{C} (\mathbf{r})$ , intertwined with ray $\mathrm{r}(t) = \mathrm{o} + t\mathrm{d}$ originating from the camera's core position $\mathrm{o}$ , is derivable using: + +$$ +\hat {C} (\mathrm {r}) = \int_ {t _ {n}} ^ {t _ {f}} \sigma (\mathrm {r} (t)) \cdot \mathrm {c} (\mathrm {r} (t), \mathrm {d}) \cdot T (t) d t, \tag {9} +$$ + +where $t_n$ and $t_f$ are the near and far bounds, and $T(t)$ is the accumulated transmittance. Addressing the challenges of hash collisions and optimizing audio feature processing, we incorporate three uniquely oriented 2D hash grids [21]. A coordinate, given by $\mathbf{x} = (x,y,z) \in \mathbb{R}^{\mathrm{XYZ}}$ , undergoes an encoding transformation for its projected values via three individual 2D-multiresolution hash encoders [30]: + +$$ +\mathcal {H} ^ {\mathrm {A B}}: (a, b) \rightarrow \mathrm {f} _ {a b} ^ {\mathrm {A B}}, \tag {10} +$$ + +where the output $\mathsf{f}_{ab}^{\mathrm{AB}} \in \mathbb{R}^{LD}$ , with the number of levels $L$ and feature dimensions per entry $D$ , signifies the planar geometric feature corresponding to the projected coordinate $(a,b)$ and $\mathcal{H}^{\mathrm{AB}}$ denotes the multiresolution hash encoder for plane $\mathbb{R}^{\mathrm{AB}}$ . By fusing the outcomes, the conclusive geometric feature $\mathsf{fg} \in \mathbb{R}^{3 \times LD}$ is derived as: + +$$ +\mathrm {f} _ {\mathrm {x}} = \mathcal {H} ^ {\mathrm {X Y}} (x, y) \oplus \mathcal {H} ^ {\mathrm {Y Z}} (y, z) \oplus \mathcal {H} ^ {\mathrm {X Z}} (x, z), \tag {11} +$$ + +where the concatenation of features is symbolized by $\oplus$ , resulting in a $3 \times LD$ -channel vector. Employing $\mathrm{f_x}$ , the perspective direction d, the lip feature $f_l$ , and the expression feature $f_e$ , the tri-plane hash's implicit function is defined as: + +$$ +\mathcal {F} ^ {\mathcal {H}}: (\mathrm {x}, \mathrm {d}, f _ {l}, f _ {e}; \mathcal {H} ^ {3}) \rightarrow (\mathrm {c}, \sigma), \tag {12} +$$ + +where $\mathcal{H}^3$ amalgamates the triad of planar hash encoders, as illustrated in Eq. 10. + +Our training employs a two-step coarse-to-fine strategy, initially using MSE loss to assess the difference between predicted $\hat{C} (\mathbf{r})$ and actual image colors $C(\mathbf{r})$ . Recognizing MSE's limitations in detail capture, we advance to a refinement stage, incorporating LPIPS loss for enhanced detail, similar to ER-NeRF [21]. We extract random patches $\mathcal{P}$ from the image, combining LPIPS (weighted by $\lambda$ ) with MSE for improved detail representation, as shown: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \sum_ {\mathrm {r}} \| C (\mathrm {r}) - \hat {C} (\mathrm {r}) \| _ {2} + \lambda \times \mathcal {L} _ {\mathrm {L P I P S}} (\hat {\mathcal {P}}, \mathcal {P}). \tag {13} +$$ + +
MethodsPSNR ↑LPIPS ↓MS-SSIM ↑FID ↓NIQE ↓BRISQUE ↓LMD ↓AUE ↓LSE-C ↑
GANWav2Lip (ACM MM 20 [34])33.43850.06970.978116.022814.536744.26594.96302.90299.2387
VideoReTalking (SIGGRAPH Asia 22 [7])31.79230.04880.96809.206314.241043.04655.85753.33087.9683
DINet (AAAI 23 [50])31.64750.04430.96409.430014.685040.36504.37253.68756.5653
TalkLip (CVPR 23 [41])32.51540.07820.969718.499714.638546.67175.86052.95795.9472
IP-LAP (CVPR 23 [53])35.15250.04430.98038.212514.640042.07503.33502.84004.9541
NeRFAD-NeRF (ICCV 21 [14])26.72910.15360.911128.986214.909155.46672.99955.54814.4996
RAD-NeRF (arXiv 22 [38])31.77540.07780.94528.657013.443344.68922.91155.09585.5219
GeneFace (ICLR 23 [44])24.81650.11780.875321.708413.335346.50614.28595.45275.1950
ER-NeRF (ICCV 23 [21])32.52160.03340.95015.293613.704834.73612.81374.18735.7749
SyncTalk (w/o Portrait)35.35420.02350.97693.924713.133333.29542.57142.57968.1331
SyncTalk (Portrait)37.40170.01130.98412.707014.216537.30422.50433.20748.0263
+ +Portrait-Sync Generator. During the training process, to address NeRF's limitations in capturing fine details like hair strands and dynamic backgrounds, we introduce a Portrait-Sync Generator with two key sections. First, NeRF renders the face area $(F_r)$ , creates $G(F_r)$ through Gaussian blur, and then uses our synchronized head pose to be able to merge with the original image $(F_o)$ to enhance hair detail fidelity. Second, when the head and torso are combined, if the character in the source video speaks while the generated face is silent, a dark gap area might appear, as shown in Fig. 5 (b). We fill these areas with the average neck color $(C_n)$ . This approach results in more realistic details and improved visual quality through the Portrait-Sync Generator. + +# 4. Experiments + +# 4.1. Experimental Settings + +Dataset. To ensure a fair comparison, we use the same well-edited video sequences from [14, 21, 44], including English and French. The average length of these videos is approximately 8,843 frames, and each video is recorded at 25 FPS. Except for the video from AD-NeRF [14], which has a resolution of $450 \times 450$ , all other videos have a resolution of $512 \times 512$ , with the character-centered. + +Comparison Baselines. We compare our method with five GAN-based methods, including Wav2Lip [34], Video-ReTalking [7], DINet [50], TalkLip [41], and IP-LAP [53], and NeRF-based methods such as AD-NeRF [14], RAD-NeRF [38], GeneFace [44], and ER-NeRF [21]. + +Implementation Details. In the coarse stage, the portrait head is trained for 100,000 iterations and 25,000 in the fine stage, sampling $256^2$ rays per iteration using a 2D hash encoder ( $L = 14$ , $F = 1$ ). We employ the AdamW optimizer [23], with learning rates of 0.01 for the hash encoder and 0.001 for other modules. Total training time is approximately 2 hours on an NVIDIA RTX 3090 GPU. + +Table 1. The quantitative results of the head reconstruction. "Portrait" refers to the use of the Portrait-Sync Generator. We achieve state-of-the-art performance on most metrics. We highlight best and second-best results. + +
MethodsAudio AAudio B
LSE-D ↓LSE-C ↑LSE-D ↓LSE-C ↑
DINet (AAAI 23 [50])8.50315.69568.20385.1134
TalkLip (CVPR 23 [41])8.76155.74498.70195.5359
IP-LAP (CVPR 23 [53])9.80373.85789.11024.389
GeneFace (ICLR 23 [44])9.54514.29339.66753.7342
ER-NeRF (ICCV 23 [21])11.8132.407610.73383.0242
SyncTalk (Ours)7.72116.66598.02486.2596
+ +Table 2. The quantitative results of the lip synchronization. We use two different audio samples to drive the same subject, then highlight best and second-best results. + +# 4.2. Quantitative Evaluation + +Full Reference Quality Assessment. In terms of image quality, we use full reference metrics such as Peak Signal-to-Noise Ratio (PSNR), Learned Perceptual Image Patch Similarity (LPIPS) [46], Multi-Scale Structure Similarity (MS-SSIM), and Frechet Inception Distance (FID) [15] as evaluation metrics. + +No Reference Quality Assessment. In high PSNR images, texture details may not align with human visual perception [47]. For more precise output definition and comparison, we use two No Reference methods: the Natural Image Quality Evaluator (NIQE)[29] and the Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE)[28]. + +Synchronization Assessment. For synchronization, we use landmark distance (LMD) to measure the synchronicity of facial movements, action units error (AUE) [4] to assess the accuracy of facial movements, and introduce Lip Sync Error Confidence (LSE-C), consistent with Wav2Lip [34], to evaluate the synchronization between lip movements and audio. + +Evaluation Results. The evaluation results of the head reconstruction are shown in Tab. 1. We compare recent methods based on GAN and NeRF. It can be observed that our image quality is superior to other methods in all aspects. In terms of synchronization, our results surpass most methods. + +![](images/f7d0ae1aceef51e6c939e47d81b1e346580de459ee560e4411f40a69f7591ab0.jpg) +Figure 4. Qualitative comparison of facial synthesis by different methods. Our method has the best visual effect on lip movements and facial expressions without the problem of separation of head and torso. Please zoom in for better visualization. + +We compare the two output modes of SyncTalk, one processed through the Portrait-Sync Generator and one without it. After processing through the Portrait-Sync Generator, hair details are restored, and the image quality is improved. Since we can maintain the subject's identity well, we surpass GAN-based methods in image quality. Thanks to the sync of lip, expression, and pose, we also outperform NeRF-based methods in image quality. Especially in terms of the LPIPS metric, our method has three times improvement compared to the previous state-of-the-art method ER-NeRF [21]. We compare the latest SOTA method drivers using out-of-distribution (OOD) audio, and the results are shown in Tab. 2. We introduce Lip Synchronization Error Distance (LSE-D) and Confidence (LSEC) for lip-audio sync evaluation, aligning with [34]. Our method shows state-of-the-art lip synchronization, overcoming small-sample NeRF limitations by incorporating a pre-trained audio-visual encoder for lip modeling. + +We also test the rendering speed. On an NVIDIA RTX 3090 GPU, and having preloaded the data onto the GPU, the head is outputted at 52 FPS with a resolution of $512 \times 512$ . In Portrait mode, using the Portrait-Sync Generator, we achieve rendering speeds of 50 FPS with our CUDA acceleration. This far exceeds the video input speed of 25 FPS, allowing for real-time generation of video streams. + +# 4.3. Qualitative Evaluation + +Evaluation Results. To more intuitively evaluate image quality, we display a comparison between our method and other methods in Fig. 4. In this figure, it can be observed that SyncTalk demonstrates more precise and more accurate facial details. Compared to Wav2Lip [34], our method better preserves the subject's identity while offering higher fidelity and resolution. Against IP-LAP [53], our method excels in lip shape synchronization, primarily due to the audio-visual consistency brought by the audio-visual encoder. Compared to GeneFace [44], our method can accurately reproduce actions such as blinking and eyebrow-raising through expression sync. In contrast to ER-NeRF [21], our method avoids the separation between the head and body through the Pose-Sync Stabilizer and generates more accurate lip shapes. Our method achieves the best overall visual effect; we recommend watching the supplementary video for comparison. + +User Study. To provide a more comprehensive evaluation of the proposed model, we design an exhaustive user study questionnaire. We extract 24 video clips, each lasting more than 10 seconds, which include head poses, facial expressions, and lip movements. Each method is represented by three of these clips. We invite 35 participants + +
Wav2Lip [34]DINet [50]TalkLip [41]IP-LAP [53]AD-NeRF [14]GeneFace [44]ER-NeRF [21]SyncTalk
Lip-sync Accuracy3.8393.6962.8933.1612.6962.9823.1894.304
Exp-sync Accuracy3.5363.4822.6073.4112.2503.0362.9464.036
Pose-sync Accuracy3.5713.5712.8753.6962.2322.9292.6073.980
Image Quality2.5002.6962.0543.5712.4643.4823.0364.054
Video Realness2.9292.4292.4293.1612.0362.7322.5184.018
+ +Table 3. User Study. Rating is on a scale of 1-5; the higher the better. The term "Exp-sync Accuracy" is an abbreviation for "Expression-sync Accuracy". We highlight best and second-best results. + +
PSNR ↑LPIPS ↓LMD ↓
Ours37.3110.01212.8032
replace Audio-Visual Encoder with Hubert [16]33.5160.02763.3961
replace Facial Animation Capture with ER-NeRF [21]'s30.2730.04153.0516
w/o Facial-Aware Masked-Attention36.5360.01652.9139
w/o Head-Sync Stabilizer28.9840.06343.5373
w/o Portrait-Sync Generator32.2390.03952.8154
+ +Table 4. Ablation study for our components. We show the PSNR, LPIPS, and LMD in different cases. + +to provide scores. The questionnaire is designed using the Mean Opinion Score (MOS) scoring protocol, asking participants to rate the generated videos from five perspectives: (1) Lip-sync Accuracy, (2) Expression-sync Accuracy, (3) Pose-sync Accuracy, (4) Image Quality, and (5) Video Realness. On average, participants take 19 minutes to complete the questionnaire, with a standardized Cronbach $\alpha$ coefficient of 0.96. The results of the User Study are displayed in Tab. 3. SyncTalk surpasses previous methods in all evaluations. Furthermore, we achieve the highest score in video authenticity, surpassing the second-place method, IP-LAP [53], by a margin of $20\%$ . User studies indicate that our method can generate visually excellent quality as perceived by humans, achieving high realism. + +# 4.4. Ablation Study + +We conduct an ablation study to examine the contributions of different components in our model. We select three core metrics for evaluation: PSNR, LPIPS, and LMD. We select a subject named "May" for testing, and the results are presented in Tab. 4. + +The Audio-Visual Encoder provides the primary lip sync information. When this module is replaced, all three metrics become worse, with the LMD error increasing by $21.15\%$ in particular, indicating decreased lip motion synchronization, as shown in Fig. 5 (a), and showing that our audio-visual encoders can extract accurate lip features. Replacing Facial Animation Capture with ER-NeRF [21]'s blink module + +![](images/612756322144cdd4323c7c0cd0c4bfcfef36b696637d5a02aa5d813042a41744.jpg) +w/o Audio-Visual Encoder w/ Audio-Visual Encoder w/o Head-Sync Stabilizer w/ Head-Sync Stabilizer +Figure 5. Ablation Study on Audio-Visual Encoder and Head-Sync Stabilizer. Removing them will lead to $(a)$ and $(b)$ . + +affects eyebrow movements and image quality. + +The Facial-Aware Masked-Attention mainly alleviates motion entanglement between the lips and the rest of the face, slightly affecting the image quality after removal. Without the Head-Sync Stabilizer, all metrics significantly declined, notably LPIPS, leading to head pose jitter and head and torso separation, as shown in Fig. 5 (b). The Portrait-Sync Generator restores details like hair. Removing this module impacts the restoration of details like hair, leading to noticeable segmentation boundaries. + +# 5. Conclusion + +In this paper, we introduce SyncTalk, a highly synchronized NeRF-based method for realistic speech-driven talking head synthesis. Our framework includes a Facial Sync Controller, Head Sync Stabilizer, and Portrait Sync Generator, which maintain subject identity and generate synchronized lip movements, facial expressions, and stable head poses. Through extensive evaluation, SyncTalk demonstrates superior performance in creating realistic and synchronized talking head videos compared to existing methods. We expect that SyncTalk will not only enhance various applications but also inspire further innovation in the field of talking head synthesis. + +# Acknowledgments + +This work was supported by National Natural Science Foundation of China (NSFC) under Grant Nos. 62172421 and 62072459, the Outstanding Innovative Talents Cultivation Funded Programs 2023 of Renmin University of China, and the Public Computing Cloud at Renmin University of China. + +# References + +[1] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence, 44(12):8717-8727, 2018. 3 +[2] Dario Amodei, Sundaram Ananthanarayanan, Rishita Anubhai, Jingliang Bai, Eric Battenberg, Carl Case, Jared Casper, Bryan Catanzaro, Qiang Cheng, Guoliang Chen, et al. Deep speech 2: End-to-end speech recognition in english and mandarin. In International conference on machine learning, pages 173-182. PMLR, 2016. 3 +[3] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 3 +[4] Tadas Baltrusaitis, Marwa Mahmoud, and Peter Robinson. Cross-dataset learning and person-specific normalisation for automatic action unit detection. In 2015 11th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-6. IEEE, 2015. 6 +[5] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European conference on computer vision (ECCV), pages 520-535, 2018. 2 +[6] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7832-7841, 2019. 2 +[7] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 6 +[8] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 3 +[9] Dipanjan Das, Sandika Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 408-424. Springer, 2020. 2 +[10] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3 +[11] Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379, 2022.3 +[12] Shreyank N Gowda, Dheeraj Pandey, and Shashank Narayana Gowda. From pixels to portraits: + +A comprehensive survey of talking head generation techniques and applications. arXiv preprint arXiv:2308.16041, 2023. 2 +[13] Jiazhi Guan, Zhanwang Zhang, Hang Zhou, Tianshu Hu, Kaisiyuan Wang, Dongliang He, Haocheng Feng, Jingtuo Liu, Errui Ding, Ziwei Liu, et al. Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1505-1515, 2023. 2 +[14] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. 2, 3, 4, 5, 6, 8 +[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[16] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460, 2021. 3, 8 +[17] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2 +[18] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM transactions on graphics (TOG), 37(4): 1-14, 2018. 1 +[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[20] Prajwal KR, Rudrabha Mukhopadhyay, Jerin Philip, Abhishek Jha, Vinay Namboodiri, and CV Jawahar. Towards automatic face-to-face translation. In Proceedings of the 27th ACM international conference on multimedia, pages 1428-1436, 2019. 2 +[21] Jiahe Li, Jiawei Zhang, Xiao Bai, Jun Zhou, and Lin Gu. Efficient region-aware neural radiance fields for high-fidelity talking portrait synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7568-7578, 2023. 2, 3, 4, 5, 6, 7, 8 +[22] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. In European Conference on Computer Vision, pages 106-125. Springer, 2022. 3 +[23] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[24] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2 + +[25] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 3 +[26] Moustafa Meshry, Saksham Suri, Larry S Davis, and Abhinav Shrivastava. Learned spatial representations for few-shot talking-head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13829-13838, 2021. 2 +[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 5 +[28] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing, 21(12): 4695-4708, 2012. 6 +[29] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6 +[30] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3, 5 +[31] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 4 +[32] Ziqiao Peng, Yihao Luo, Yue Shi, Hao Xu, Xiangyu Zhu, Hongyan Liu, Jun He, and Zhaoxin Fan. Selftalk: A self-supervised commutative training diagram to comprehend 3d talking faces. In Proceedings of the 31st ACM International Conference on Multimedia, pages 5292-5301, 2023. 1 +[33] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20687-20697, 2023. 2, 4 +[34] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 2, 6, 7, 8 +[35] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In European Conference on Computer Vision, pages 666-682. Springer, 2022. 2 +[36] Linsen Song, Wayne Wu, Chen Qian, Ran He, and Chen Change Loy. Everybody's talkin': Let me talk as you want. IEEE Transactions on Information Forensics and Security, 17:585-598, 2022. 2 +[37] Yasheng Sun, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Zhibin Hong, Jingtuo Liu, Errui Ding, Jingdong Wang, Zwei Liu, and Koike Hideki. Masked lip-sync prediction by + +audio-visual contextual exploitation in transformers. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2 +[38] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 3, 6 +[39] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVI 16, pages 716–731. Springer, 2020. 1 +[40] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. International Journal of Computer Vision, 128:1398-1413, 2020. 2 +[41] Jiadong Wang, Xinyuan Qian, Malu Zhang, Robby T Tan, and Haizhou Li. Seeing what you said: Talking face generation guided by a lip reading expert. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14653-14662, 2023. 2, 6, 8 +[42] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2 +[43] Shunyu Yao, RuiZhe Zhong, Yichao Yan, Guangtao Zhai, and Xiaokang Yang. Dfa-nerf: Personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791, 2022. 2, 3, 5 +[44] Zhenhui Ye, Ziyue Jiang, Yi Ren, Jinglin Liu, JinZheng He, and Zhou Zhao. Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430, 2023. 2, 3, 4, 6, 7, 8 +[45] Chenxu Zhang, Yifan Zhao, Yifei Huang, Ming Zeng, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Facial: Synthesizing dynamic talking face with implicit attribute learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3867-3876, 2021. 2 +[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[47] Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. Ranksrgan: Generative adversarial networks with ranker for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3096-3105, 2019. 6 +[48] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2 +[49] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661-3670, 2021. 2 +[50] Zhimeng Zhang, Zhipeng Hu, Wenjin Deng, Changjie Fan, Tangjie Lv, and Yu Ding. Dinet: Deformation inpainting network for realistic face visually dubbing on high resolution video. arXiv preprint arXiv:2303.03988, 2023. 2, 6, 8 +[51] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 3 +[52] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21057-21067, 2023. 3 +[53] Weizhi Zhong, Chaowei Fang, Yinqi Cai, Pengxu Wei, Gangming Zhao, Liang Lin, and Guanbin Li. Identity-preserving talking face generation with landmark and appearance priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2023. 2, 6, 7, 8 +[54] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI conference on artificial intelligence, pages 9299–9306, 2019. 2 +[55] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 2 +[56] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG), 39(6):1-15, 2020. 2 +[57] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4574-4584, 2023. 3 \ No newline at end of file diff --git a/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/images.zip b/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..41e465a4785db9cdddddeb62df293275e9b0942a --- /dev/null +++ b/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa6d19915ca317693754e9be4342d87ae25b6e3a4b53b571b928b65796552176 +size 763021 diff --git a/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/layout.json b/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d4965c4e8a73970efe252f0b2071fc41f9b196c9 --- /dev/null +++ b/2024/SyncTalk_ The Devil is in the Synchronization for Talking Head Synthesis/layout.json @@ -0,0 +1,9005 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 523, + 121 + ], + "type": "text", + "content": "SyncTalk: The Devil is in the Synchronization for Talking Head Synthesis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "spans": [ + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "text", + "content": "Ziqiao Peng" + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "text", + "content": " Wentao Hu" + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "text", + "content": " Yue Shi" + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "text", + "content": " Xiangyu Zhu" + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "text", + "content": " Xiaomei Zhang" + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "text", + "content": " Hao Zhao" + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "text", + "content": " Jun He" + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "text", + "content": " Hongyan Liu" + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "inline_equation", + "content": "^{5*}" + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "text", + "content": " Zhaoxin Fan" + }, + { + "bbox": [ + 79, + 141, + 514, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 171, + 503, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 171, + 503, + 201 + ], + "spans": [ + { + "bbox": [ + 91, + 171, + 503, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 171, + 503, + 201 + ], + "type": "text", + "content": "Renmin University of China " + }, + { + "bbox": [ + 91, + 171, + 503, + 201 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 171, + 503, + 201 + ], + "type": "text", + "content": "Beijing University of Posts and Telecommunications " + }, + { + "bbox": [ + 91, + 171, + 503, + 201 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 91, + 171, + 503, + 201 + ], + "type": "text", + "content": "Psyche AI Inc. " + }, + { + "bbox": [ + 91, + 171, + 503, + 201 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 91, + 171, + 503, + 201 + ], + "type": "text", + "content": "Chinese Academy of Sciences " + }, + { + "bbox": [ + 91, + 171, + 503, + 201 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 91, + 171, + 503, + 201 + ], + "type": "text", + "content": "Tsinghua University" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 213, + 534, + 418 + ], + "blocks": [ + { + "bbox": [ + 58, + 213, + 534, + 418 + ], + "lines": [ + { + "bbox": [ + 58, + 213, + 534, + 418 + ], + "spans": [ + { + "bbox": [ + 58, + 213, + 534, + 418 + ], + "type": "image", + "image_path": "785227c2823cf3d039b62977e4be371e69d0f11158259451da2e2a7fca027ef6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 423, + 547, + 458 + ], + "lines": [ + { + "bbox": [ + 45, + 423, + 547, + 458 + ], + "spans": [ + { + "bbox": [ + 45, + 423, + 547, + 458 + ], + "type": "text", + "content": "Figure 1. The proposed SyncTalk synthesizes synchronized talking head videos, employing tri-plane hash representations to maintain subject identity. It can generate synchronized lip movements, facial expressions, and stable head poses, and restores hair details to create high-resolution videos." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 467, + 192, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 467, + 192, + 480 + ], + "spans": [ + { + "bbox": [ + 143, + 467, + 192, + 480 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 492, + 290, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 290, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 290, + 696 + ], + "type": "text", + "content": "Achieving high synchronization in the synthesis of realistic, speech-driven talking head videos presents a significant challenge. Traditional Generative Adversarial Networks (GAN) struggle to maintain consistent facial identity, while Neural Radiance Fields (NeRF) methods, although they can address this issue, often produce mismatched lip movements, inadequate facial expressions, and unstable head poses. A lifelike talking head requires synchronized coordination of subject identity, lip movements, facial expressions, and head poses. The absence of these synchronizations is a fundamental flaw, leading to unrealistic and artificial outcomes. To address the critical issue of synchronization, identified as the \"devil\" in creating realistic talking heads, we introduce SyncTalk. This NeRF-based method effectively maintains subject identity, enhancing synchronization and realism in talking head synthesis. SyncTalk employs a Face-Sync Controller to align lip movements" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 468, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 545, + 588 + ], + "type": "text", + "content": "with speech and innovatively uses a 3D facial blendshape model to capture accurate facial expressions. Our Head-Sync Stabilizer optimizes head poses, achieving more natural head movements. The Portrait-Sync Generator restores hair details and blends the generated head with the torso for a seamless visual experience. Extensive experiments and user studies demonstrate that SyncTalk outperforms state-of-the-art methods in synchronization and realism. We recommend watching the supplementary video: https://ziqiaopeng.github.io/synctalk" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 609, + 387, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 387, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 387, + 623 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "text", + "content": "The quest to generate dynamic and realistic speech-driven talking heads has intensified, driven by expanding applications in digital assistants [39], virtual reality [32], and film-making [18]. The ultimate goal is to enhance the realism of synthetic videos to align with human perceptual expectations. However, a fundamental challenge that persists across existing methods is the need for synchroniza" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 703, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 143, + 713 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "666" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "tion. Traditional methods based on Generative Adversarial Networks (GAN) [13, 41, 50, 53], while adept at modeling the lip movements of speakers, often produce inconsistent identities across different frames, leading to issues such as different tooth sizes and fluctuating lip thickness. Similarly, emerging technologies based on Neural Radiance Fields (NeRF) [14, 21, 35, 43, 44] excel in maintaining identity consistency and preserving facial details. However, they struggle with mismatches in lip movements, challenging facial expression control, and unstable head poses, thereby diminishing the overall realism of the video." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 206, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 289, + 434 + ], + "type": "text", + "content": "In this paper, we find that the \"devil\" is in the synchronization. Existing methods need more synchronization in four key areas: subject identity, lip movements, facial expressions, and head poses. Firstly, in GAN-based methods, maintaining the subject's identity in the video is challenging due to the instability of features in consecutive frames and the use of only a few frames as references for facial reconstruction [34]. Secondly, lip movements are not synchronized with speech. In NeRF-based methods, audio features trained only on a 5-minute speech dataset struggle to generalize to different speech inputs [44]. Thirdly, there is a lack of facial expression control, with most methods only producing lip movements or controlling blinking, resulting in unnatural facial actions [12]. Fourthly, the head pose is not synchronized. Previous methods relied on sparse landmarks to compute projection error, but jitter and inaccuracy in these landmarks lead to unstable head poses [49], as shown in Fig. 1. These synchronization issues introduce artifacts and significantly reduce realism." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 436, + 289, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 289, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 289, + 568 + ], + "type": "text", + "content": "To address these synchronization challenges, we introduce SyncTalk, a NeRF-based method focused on highly synchronized, realistic, speech-driven talking head synthesis, employing tri-plane hash representations to maintain subject identity. Through the Face-Sync Controller and Head-Sync Stabilizer, SyncTalk significantly enhances the synchronization and visual quality of the synthesized videos. Visual quality is further improved by the Portrait-Sync Generator, which meticulously refines visual details. The entire rendering process can achieve 50 FPS and outputs high-resolution videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "Within the Face-Sync Controller, we pre-train an audio-visual encoder on the 2D audio-visual dataset, resulting in a generalized representation that ensures synchronized lip movements across different speech samples. For controlling facial expressions, we employ a semantically enriched 3D facial blendshape model [33]. This model is distinguished by its ability to control specific facial expression regions through 52 parameters. Regarding the Head-Sync Stabilizer, we use a head motion tracker [14] to infer the head's rough rotation and translation parameters. Due to the instability of the rough parameters, inspired by Simultaneous Localization and Mapping (SLAM), we incorporate a head" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "point tracker to track dense keypoints and integrate a bundle adjustment method to optimize the head pose, thereby achieving stable and continuous head motion. To further enhance the visual fidelity of SyncTalk, we have designed a Portrait-Sync Generator. This module repairs artifacts in the NeRF modeling, particularly the intricate details of hair and background flaws, and outputs high-resolution video." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 159, + 547, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 547, + 181 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 547, + 181 + ], + "type": "text", + "content": "In summary, the main contributions of our work are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 184, + 545, + 326 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 304, + 184, + 545, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 184, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 304, + 184, + 545, + 232 + ], + "type": "text", + "content": "- We present a Face-Sync Controller that utilizes an Audio-Visual Encoder in conjunction with a Facial Animation Capturer, ensuring accurate lip synchronization and dynamic facial expression rendering." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 232, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 278 + ], + "type": "text", + "content": "- We introduce a Head-Sync Stabilizer that tracks head rotation and facial movement keypoints. Utilizing the bundle adjustment method, this Stabilizer guarantees smooth and synchronous head motion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 280, + 545, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 545, + 326 + ], + "type": "text", + "content": "- We design a Portrait-Sync Generator that improves visual fidelity by repairing artifacts in NeRF modeling and refining intricate details like hair and background in high-resolution videos." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 344, + 392, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 344, + 392, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 344, + 392, + 357 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 365, + 496, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 496, + 379 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 496, + 379 + ], + "type": "text", + "content": "2.1. GAN-based Talking Head Synthesis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 386, + 547, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 547, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 547, + 554 + ], + "type": "text", + "content": "Recently, GAN-based talking head synthesis [5, 6, 9, 20, 26, 36, 40, 54, 55] has emerged as an essential research area in computer vision, but they struggle to maintain the identity of the subject in videos consistently. A cluster of noteworthy techniques, including [13, 34, 37, 41, 50, 53], mainly focus on generating video streams for the lip region. These methods create new visuals for talking head portraits by changing the lip region. For instance, Wav2Lip [34] introduces a lip sync expert for supervising lip movements. However, due to the use of five frames from the reference frame to reconstruct the lip, it struggles to maintain the subject's identity. In contrast, methods like [6, 24, 42, 56] perform full-face synthesis but struggle to ensure sync between facial expressions and head poses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 555, + 547, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 547, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 547, + 652 + ], + "type": "text", + "content": "Apart from video stream techniques, efforts have also been made to enable a single image to \"speak\" using speech, as demonstrated in [17, 45, 48]. For example, SadTalker [48] can generate videos of a person speaking from a single image. However, such methods fail to generate natural head poses and facial expressions and struggle to maintain the subject's identity, affecting the sync effect and leading to an unrealistic visual perception." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Compared to these methods, SyncTalk uses NeRF to perform three-dimensional modeling of the face. Its capability to represent continuous 3D scenes in canonical spaces translates to exceptional performance in maintaining subject identity consistency and detail preservation." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "667" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 61, + 545, + 257 + ], + "blocks": [ + { + "bbox": [ + 50, + 61, + 545, + 257 + ], + "lines": [ + { + "bbox": [ + 50, + 61, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 50, + 61, + 545, + 257 + ], + "type": "image", + "image_path": "22c07d340564f4ba4e434acfa89e395806ce39fb6903ee6560081d815f96a98e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "type": "text", + "content": "Figure 2. Overview of SyncTalk. Given a cropped reference video of a talking head and the corresponding speech, SyncTalk can extract the Lip Feature " + }, + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "type": "inline_equation", + "content": "f_{l}" + }, + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "type": "text", + "content": ", Expression Feature " + }, + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "type": "text", + "content": ", and Head Pose " + }, + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "type": "inline_equation", + "content": "(R, T)" + }, + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "type": "text", + "content": " through two synchronization modules " + }, + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "type": "inline_equation", + "content": "(a)" + }, + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "type": "inline_equation", + "content": "(b)" + }, + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "type": "text", + "content": ". The Tri-Plane Hash Representation then models the head, outputting a rough speech-driven video. The Portrait-Sync Generator further restores details such as hair and background, ultimately producing a high-resolution talking head video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 326, + 241, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 326, + 241, + 339 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 241, + 339 + ], + "type": "text", + "content": "2.2. NeRF-based Talking Head Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 289, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 289, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 289, + 636 + ], + "type": "text", + "content": "With the recent rise of NeRF, numerous fields have begun to utilize it to tackle related challenges [11, 25]. Previous work [14, 22, 43, 44] has integrated NeRF into the task of synthesizing talking heads and has used audio as the driving signal, but these methods are all based on the vanilla NeRF model. For instance, AD-NeRF [14] requires approximately 10 seconds to render a single image. RAD-NeRF [38] aims for real-time video generation and employs a NeRF based on Instant-NGP [30]. ER-NeRF [21] innovatively introduces triple-plane hash encoders to trim the empty spatial regions, advocating for a compact and accelerated rendering approach. GeneFace [44] attempts to reduce NeRF artifacts by translating speech features into facial landmarks, but this often results in inaccurate lip movements. Attempts to create character avatars with NeRF-based methods, such as [10, 51, 52, 57], cannot be directly driven by speech. These methods only use audio as a condition, without a clear concept of sync, and usually result in average lip movement. Additionally, previous methods lack control over facial expressions, being limited to controlling blinking only, and cannot model actions like raising eyebrows or frowning. Furthermore, these methods have a significant issue with unstable head poses, leading to the separation between the head and the torso." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": "In contrast, we use the Face-Sync Controller to model the relationship between audio and lip movements, thereby enhancing the synchronization of lip movements and expressions, and the Head-Sync Stabilizer to stabilize head poses. By addressing these synchronization \"devils\", our method improves visual quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 325, + 361, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 325, + 361, + 338 + ], + "spans": [ + { + "bbox": [ + 306, + 325, + 361, + 338 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 345, + 373, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 345, + 373, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 373, + 357 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 364, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 460 + ], + "type": "text", + "content": "In this section, we introduce the proposed SyncTalk, as shown in Fig. 2. SyncTalk mainly consists of 3 parts: a) lip movements and facial expressions controlled by the FaceSync Controller, b) stable head pose provided by the HeadSync Stabilizer, and c) high-synchronization facial frames rendered by the Dynamic Portrait Renderer. We will describe the content of these three parts in detail in the following subsections." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 468, + 428, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 428, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 428, + 480 + ], + "type": "text", + "content": "3.2. Face-Sync Controller" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "content": "Audio-Visual Encoder. Existing methods based on NeRF utilize pre-trained models such as DeepSpeech [2], Wav2Vec 2.0 [3], or HuBERT [16]. These are audio feature extraction methods designed for speech recognition tasks. Using an audio encoder designed for Automatic Speech Recognition (ASR) tasks does not truly reflect lip movements. This is because the pre-trained model is based on the distribution of features from audio to text, whereas we need the feature distribution from audio to lip movements." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "Considering the above, we use an audio and visual synchronization audio encoder trained on the 2D audio-visual synchronization dataset LRS2 [1]. This ensures that the audio features extracted by our method and lip movements have the same feature distribution. The specific implementation method is as follows: We use a pre-trained lip synchronization discriminator [8]. It can give confidence for the lip synchronization effect of the video. The lip synchronization discriminator takes as input a continuous face window " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": " and the corresponding audio frame " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": ". They are" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "668" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "judged as positive samples (with label " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "y = 1" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ") if they overlap entirely. Otherwise, they are judged as negative samples (with label " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "y = 0" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "). The discriminator calculates the cosine similarity between these sequences as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 113, + 128, + 287, + 155 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 128, + 287, + 155 + ], + "spans": [ + { + "bbox": [ + 113, + 128, + 287, + 155 + ], + "type": "interline_equation", + "content": "\\operatorname {s i m} (F, A) = \\frac {F \\cdot A}{\\| F \\| _ {2} \\| A \\| _ {2}}, \\tag {1}", + "image_path": "b715f969c775f71607d67d8853cf389fac8a737e9c47a02246acd24d029ec504.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 157, + 209, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 157, + 209, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 157, + 209, + 169 + ], + "type": "text", + "content": "and then uses binary cross-entropy loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 179, + 296, + 203 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 179, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 47, + 179, + 296, + 203 + ], + "type": "interline_equation", + "content": "L _ {\\text {s y n c}} = - (y \\log (\\operatorname {s i m} (F, A)) + (1 - y) \\log (1 - \\operatorname {s i m} (F, A))), \\tag {2}", + "image_path": "86618bf9b0a57730a8299603a856252b2af7bceeace4bd3601d6bcde2c0c6a3d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 204, + 287, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 287, + 227 + ], + "type": "text", + "content": "to minimize the distance for synchronized samples and maximize the distance for non-synchronized samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 228, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 335 + ], + "type": "text", + "content": "Under the supervision of the lip synchronization discriminator, we pre-train a highly synchronized audio-visual feature extractor related to lip movements. First, we use convolutional networks to obtain audio features " + }, + { + "bbox": [ + 46, + 228, + 287, + 335 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}(A)" + }, + { + "bbox": [ + 46, + 228, + 287, + 335 + ], + "type": "text", + "content": " and encode facial features " + }, + { + "bbox": [ + 46, + 228, + 287, + 335 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}(F)" + }, + { + "bbox": [ + 46, + 228, + 287, + 335 + ], + "type": "text", + "content": ". These features are then concatenated. In the decoding phase, we use stacked convolutional layers to restore facial frames using the operation " + }, + { + "bbox": [ + 46, + 228, + 287, + 335 + ], + "type": "inline_equation", + "content": "\\mathrm{Dec}(\\mathrm{Conv}(A) \\oplus \\mathrm{Conv}(F))" + }, + { + "bbox": [ + 46, + 228, + 287, + 335 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 228, + 287, + 335 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 228, + 287, + 335 + ], + "type": "text", + "content": " reconstruction loss during training is given by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 346, + 287, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 346, + 287, + 359 + ], + "spans": [ + { + "bbox": [ + 74, + 346, + 287, + 359 + ], + "type": "interline_equation", + "content": "L _ {\\text {r e c o n}} = \\| F - \\operatorname {D e c} (\\operatorname {C o n v} (A) \\oplus \\operatorname {C o n v} (F)) \\| _ {1}. \\tag {3}", + "image_path": "e5764eb17bf4999aaee02168eac69e6c7da53754f77fb3391a1e4800126b7b7c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 366, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 450 + ], + "type": "text", + "content": "Simultaneously, we sample synchronized and nonsynchronized segments using lip movement discriminators, and employ a same sync loss as Eq. 2. By minimizing both losses, we train a facial generation network related to audio. After training, we use " + }, + { + "bbox": [ + 46, + 366, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}(A)" + }, + { + "bbox": [ + 46, + 366, + 287, + 450 + ], + "type": "text", + "content": " as the lip space extracted from the audio. Ultimately, we obtain our highly synchronized audio-visual encoder related to lip movements." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": "Facial Animation Capturer. It is observed that previous methods based on NeRF [14, 21, 44] could only change blinking and could not model facial expressions accurately. This leads to issues such as rigid facial expressions and incorrect facial details if the trained character has significant facial movements, like squinting, raising eyebrows, or frowning. Considering the need for more synchronized and realistic facial expressions, we add an expression synchronization control module. Specifically, we introduce a 3D facial prior using 52 semantically facial blendshape coefficients [33] represented by " + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": " to model the face, as shown in Fig. 3. Because the 3D face model can retain the structure information of face motion, it can reflect the content of facial movements well without causing facial structural distortion. During the training, we first use a sophisticated facial blendshape capture module to capture facial expressions as " + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "E(B)" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": ", and select seven core facial expression control coefficients to control the eyebrow, forehead, and eye areas. They are highly correlated with expression and independent of lip movements. We can synchronize the speaker's expression during the inference process because the facial coefficients are semantically informed." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 425, + 137 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 425, + 137 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 425, + 137 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 425, + 137 + ], + "type": "image", + "image_path": "5d13645c920df6df94817562e51c9a4750f4c4cebed23dfd82ba2e57d0f91e33.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 137, + 424, + 203 + ], + "blocks": [ + { + "bbox": [ + 309, + 137, + 424, + 203 + ], + "lines": [ + { + "bbox": [ + 309, + 137, + 424, + 203 + ], + "spans": [ + { + "bbox": [ + 309, + 137, + 424, + 203 + ], + "type": "image", + "image_path": "851fad13038e29b3b0cad2a1260d21d3a60c0ebea9fe827e47c45e19b40d2e2e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 213, + 545, + 235 + ], + "lines": [ + { + "bbox": [ + 305, + 213, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 545, + 235 + ], + "type": "text", + "content": "Figure 3. Facial Animation Capturer. We use 3D facial blend-shape coefficients to control the expressions of characters." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 71, + 544, + 137 + ], + "blocks": [ + { + "bbox": [ + 425, + 71, + 544, + 137 + ], + "lines": [ + { + "bbox": [ + 425, + 71, + 544, + 137 + ], + "spans": [ + { + "bbox": [ + 425, + 71, + 544, + 137 + ], + "type": "image", + "image_path": "2611530a845c6638ed5c75306e89a915a09d63b9f136ed59ab86fac44b732a0f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 425, + 137, + 544, + 203 + ], + "blocks": [ + { + "bbox": [ + 425, + 137, + 544, + 203 + ], + "lines": [ + { + "bbox": [ + 425, + 137, + 544, + 203 + ], + "spans": [ + { + "bbox": [ + 425, + 137, + 544, + 203 + ], + "type": "image", + "image_path": "40f9df9661c8a9dde93ad542361784c8ffb56f9e4ccc38e9561f9c732720c7f7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 243, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 327 + ], + "type": "text", + "content": "Facial-Aware Masked-Attention. To reduce the mutual interference between lip features and expression features during training, we introduce the Facial-Aware Disentangle Attention module. Building on the region attention vector " + }, + { + "bbox": [ + 305, + 243, + 545, + 327 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 305, + 243, + 545, + 327 + ], + "type": "text", + "content": " [21], we add masks " + }, + { + "bbox": [ + 305, + 243, + 545, + 327 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{lip}}" + }, + { + "bbox": [ + 305, + 243, + 545, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 243, + 545, + 327 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{exp}}" + }, + { + "bbox": [ + 305, + 243, + 545, + 327 + ], + "type": "text", + "content": " to the attention areas for lips and expressions, respectively. Specifically, the new attention mechanisms are given by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 390, + 338, + 545, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 338, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 390, + 338, + 545, + 357 + ], + "type": "interline_equation", + "content": "V _ {\\text {l i p}} = V \\odot M _ {\\text {l i p}}, \\tag {4}", + "image_path": "603e4f7e135af54102aa2ca203282ca9ffe8e24e39f471a1169a38b7405f8e1b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 389, + 354, + 463, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 354, + 463, + 367 + ], + "spans": [ + { + "bbox": [ + 389, + 354, + 463, + 367 + ], + "type": "interline_equation", + "content": "V _ {\\mathrm {e x p}} = V \\odot M _ {\\mathrm {e x p}}.", + "image_path": "f9aa3c30cb652fc1a2d3656c0f270015d3b5daa8b2fbc33fe42acae22de6f4ed.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 371, + 545, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 545, + 480 + ], + "type": "text", + "content": "These formulations allow the attention mechanisms to focus solely on their respective parts, thereby reducing entanglement between them. Before the disentanglement, lip movements might induce blinking tendencies and affect hair volume. By introducing the mask module, the attention mechanism can focus on either expressions or lips without affecting other areas, thereby reducing the artifact caused by coupling. Finally, we obtain the disentangled lip feature " + }, + { + "bbox": [ + 304, + 371, + 545, + 480 + ], + "type": "inline_equation", + "content": "f_{l} = f_{\\mathrm{lip}} \\odot V_{\\mathrm{lip}}" + }, + { + "bbox": [ + 304, + 371, + 545, + 480 + ], + "type": "text", + "content": " and expression feature " + }, + { + "bbox": [ + 304, + 371, + 545, + 480 + ], + "type": "inline_equation", + "content": "f_{e} = f_{\\mathrm{exp}} \\odot V_{\\mathrm{exp}}" + }, + { + "bbox": [ + 304, + 371, + 545, + 480 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 485, + 427, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 485, + 427, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 485, + 427, + 498 + ], + "type": "text", + "content": "3.3. Head-Sync Stabilizer" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "text", + "content": "Head Motion Tracker. The head pose, denoted as " + }, + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "text", + "content": ", refers to the rotation angle of a person's head in 3D space and is defined by a rotation " + }, + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "text", + "content": " and a translation " + }, + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "text", + "content": ". An unstable head pose can lead to head jitter. To obtain a rough estimate of the head pose, initially, the best focal length is determined through " + }, + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "text", + "content": " iterations within a predetermined range. For each focal length candidate, " + }, + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "text", + "content": ", the system re-initializes the rotation and translation values. The objective is to minimize the error between the projected landmarks from the 3D Morphable Models (3DMM) [31] and the actual landmarks in the video frame. Formally, the optimal focal length " + }, + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{opt}}" + }, + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 339, + 654, + 545, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 654, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 339, + 654, + 545, + 672 + ], + "type": "interline_equation", + "content": "f _ {\\text {o p t}} = \\arg \\min _ {f _ {i}} E _ {i} \\left(L _ {2 D}, L _ {3 D} \\left(f _ {i}, R _ {i}, T _ {i}\\right)\\right), \\tag {5}", + "image_path": "ac3ccbf2447b2b14c8a8979723b67776afd2ac6fc3e41f895eaed5c7233a6523.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " represents the Mean Squared Error (MSE) between these landmarks, " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "L_{3D}(f_i,R_i,T_i)" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " represents the projected landmarks from the 3DMM for a given focal length" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "669" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": ", the corresponding rotation and translation parameters " + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "L_{2D}" + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": " are the actual landmarks from the video frame. Subsequently, leveraging the optimal focal length " + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{opt}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": ", the system refines the rotation " + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": " parameters for all frames to better align the model's projected landmarks with the actual video landmarks. This refinement process can be mathematically represented as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 167, + 287, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 167, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 167, + 287, + 186 + ], + "type": "interline_equation", + "content": "\\left(R _ {\\text {o p t}}, T _ {\\text {o p t}}\\right) = \\arg \\min _ {R, T} E \\left(L _ {2 D}, L _ {3 D} \\left(f _ {\\text {o p t}}, R, T\\right)\\right), \\tag {6}", + "image_path": "a283db43216f367c40d33bb3423c99b4a141154ae8e3afa4af18c1d1d8f7bda3.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "type": "text", + "content": " denotes the MSE metric, between the 3D model's projected landmarks " + }, + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "type": "inline_equation", + "content": "L_{3D}" + }, + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "type": "text", + "content": " for the optimal focal length " + }, + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{opt}}" + }, + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "type": "text", + "content": ", and the actual 2D landmarks " + }, + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "type": "inline_equation", + "content": "L_{2D}" + }, + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "type": "text", + "content": " in the video frame. The optimized rotation " + }, + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{opt}}" + }, + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{opt}}" + }, + { + "bbox": [ + 46, + 196, + 287, + 256 + ], + "type": "text", + "content": " are obtained by minimizing this error across all frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "type": "text", + "content": "Head Points Tracker. Considering methods based on NeRF and their requirements for inputting head rotation " + }, + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "type": "text", + "content": ", previous methods utilize 3DMM-based techniques to extract head poses and generate an inaccurate result. To improve the precision of " + }, + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "type": "text", + "content": ", We use an optical flow estimation model from [43] to track facial keypoints " + }, + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 257, + 287, + 424 + ], + "type": "text", + "content": ". Specifically, using a pre-trained optical flow estimation model, after obtaining the facial motion optical flow, we use the Laplacian filter to select the keypoints where the most significant flow changes are located and track the motion trajectories of these keypoints in the flow sequence. Through this module, our method ensures a more precise and consistent facial keypoint alignment across all frames, enhancing the accuracy of head pose parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 425, + 288, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 288, + 534 + ], + "type": "text", + "content": "**Bundle Adjustment.** Given the keypoints and the rough head pose, we introduce a two-stage optimization framework from [14] to enhance the accuracy of keypoints and head pose estimations. In the first stage, we randomly initialize the 3D coordinates of " + }, + { + "bbox": [ + 46, + 425, + 288, + 534 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 425, + 288, + 534 + ], + "type": "text", + "content": " keypoints and optimize their positions to align with the tracked keypoints on the image plane. This process involves minimizing a loss function " + }, + { + "bbox": [ + 46, + 425, + 288, + 534 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{init}}" + }, + { + "bbox": [ + 46, + 425, + 288, + 534 + ], + "type": "text", + "content": ", which captures the discrepancy between projected keypoints " + }, + { + "bbox": [ + 46, + 425, + 288, + 534 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 425, + 288, + 534 + ], + "type": "text", + "content": " and the tracked keypoints " + }, + { + "bbox": [ + 46, + 425, + 288, + 534 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 425, + 288, + 534 + ], + "type": "text", + "content": ", as given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 545, + 287, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 545, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 116, + 545, + 287, + 572 + ], + "type": "interline_equation", + "content": "L _ {\\text {i n i t}} = \\sum_ {j} \\| P _ {j} - K _ {j} \\| _ {2}. \\tag {7}", + "image_path": "e015a831b4ae4848a0fadc7961c27d99c6196663c5d87042a74f89ecbdc9dd31.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 575, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 287, + 648 + ], + "type": "text", + "content": "Subsequently, in the second stage, we embark on a more comprehensive optimization to refine the 3D keypoints and the associated head jointly pose parameters. Through the Adam Optimization [19], the algorithm adjust the spatial coordinates, rotation angles " + }, + { + "bbox": [ + 46, + 575, + 287, + 648 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 575, + 287, + 648 + ], + "type": "text", + "content": ", and translations " + }, + { + "bbox": [ + 46, + 575, + 287, + 648 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 575, + 287, + 648 + ], + "type": "text", + "content": " to minimize the alignment error " + }, + { + "bbox": [ + 46, + 575, + 287, + 648 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{sec}}" + }, + { + "bbox": [ + 46, + 575, + 287, + 648 + ], + "type": "text", + "content": ", expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 660, + 287, + 686 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 660, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 104, + 660, + 287, + 686 + ], + "type": "interline_equation", + "content": "L _ {\\sec} = \\sum_ {j} \\| P _ {j} (R, T) - K _ {j} \\| _ {2}. \\tag {8}", + "image_path": "455ef9ab1a15ae80a3d841d70d8d53afa3f2c1ba791837151a5b905631287699.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "After these optimizations, the resultant head pose and translation parameters are observed to be smooth and stable." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 459, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 459, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 459, + 85 + ], + "type": "text", + "content": "3.4. Dynamic Portrait Rendering" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "spans": [ + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "text", + "content": "Tri-Plane Hash Representation. Using a collection of multi-perspective images along with corresponding camera poses, NeRF [27] harnesses these resources to manifest a 3D static scene. This representation utilizes an implicit function " + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "text", + "content": ", delineated as " + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{F}:(\\mathrm{x},\\mathrm{d})\\to (\\mathrm{c},\\sigma)" + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "text", + "content": ", where the 3D spatial location is given by " + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "inline_equation", + "content": "\\mathrm{x} = (x,y,z)" + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "text", + "content": ", and the direction of viewing is characterized by " + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "inline_equation", + "content": "\\mathrm{d} = (\\theta ,\\phi)" + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "text", + "content": ". The resultant values, " + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "inline_equation", + "content": "\\mathrm{c} = (r,g,b)" + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "text", + "content": ", represent the radiance and density, respectively. The predicted pixel color, denoted by " + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "inline_equation", + "content": "\\hat{C} (\\mathbf{r})" + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "text", + "content": ", intertwined with ray " + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "inline_equation", + "content": "\\mathrm{r}(t) = \\mathrm{o} + t\\mathrm{d}" + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "text", + "content": " originating from the camera's core position " + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "inline_equation", + "content": "\\mathrm{o}" + }, + { + "bbox": [ + 305, + 90, + 547, + 223 + ], + "type": "text", + "content": ", is derivable using:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 342, + 230, + 545, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 230, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 342, + 230, + 545, + 258 + ], + "type": "interline_equation", + "content": "\\hat {C} (\\mathrm {r}) = \\int_ {t _ {n}} ^ {t _ {f}} \\sigma (\\mathrm {r} (t)) \\cdot \\mathrm {c} (\\mathrm {r} (t), \\mathrm {d}) \\cdot T (t) d t, \\tag {9}", + "image_path": "efc6539bffa87788f14537997b12efb7f6d066c83b2b1c373469c8a0913d261c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 261, + 545, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 261, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 305, + 261, + 545, + 345 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 261, + 545, + 345 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 305, + 261, + 545, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 261, + 545, + 345 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 305, + 261, + 545, + 345 + ], + "type": "text", + "content": " are the near and far bounds, and " + }, + { + "bbox": [ + 305, + 261, + 545, + 345 + ], + "type": "inline_equation", + "content": "T(t)" + }, + { + "bbox": [ + 305, + 261, + 545, + 345 + ], + "type": "text", + "content": " is the accumulated transmittance. Addressing the challenges of hash collisions and optimizing audio feature processing, we incorporate three uniquely oriented 2D hash grids [21]. A coordinate, given by " + }, + { + "bbox": [ + 305, + 261, + 545, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = (x,y,z) \\in \\mathbb{R}^{\\mathrm{XYZ}}" + }, + { + "bbox": [ + 305, + 261, + 545, + 345 + ], + "type": "text", + "content": ", undergoes an encoding transformation for its projected values via three individual 2D-multiresolution hash encoders [30]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 381, + 351, + 545, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 351, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 381, + 351, + 545, + 366 + ], + "type": "interline_equation", + "content": "\\mathcal {H} ^ {\\mathrm {A B}}: (a, b) \\rightarrow \\mathrm {f} _ {a b} ^ {\\mathrm {A B}}, \\tag {10}", + "image_path": "19705fdde27b2dc6c3eb3cf59fe7b05daf696f8a5baf59f54758a576d9a15b75.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 371, + 545, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 371, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 305, + 371, + 545, + 444 + ], + "type": "text", + "content": "where the output " + }, + { + "bbox": [ + 305, + 371, + 545, + 444 + ], + "type": "inline_equation", + "content": "\\mathsf{f}_{ab}^{\\mathrm{AB}} \\in \\mathbb{R}^{LD}" + }, + { + "bbox": [ + 305, + 371, + 545, + 444 + ], + "type": "text", + "content": ", with the number of levels " + }, + { + "bbox": [ + 305, + 371, + 545, + 444 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 371, + 545, + 444 + ], + "type": "text", + "content": " and feature dimensions per entry " + }, + { + "bbox": [ + 305, + 371, + 545, + 444 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 371, + 545, + 444 + ], + "type": "text", + "content": ", signifies the planar geometric feature corresponding to the projected coordinate " + }, + { + "bbox": [ + 305, + 371, + 545, + 444 + ], + "type": "inline_equation", + "content": "(a,b)" + }, + { + "bbox": [ + 305, + 371, + 545, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 371, + 545, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^{\\mathrm{AB}}" + }, + { + "bbox": [ + 305, + 371, + 545, + 444 + ], + "type": "text", + "content": " denotes the multiresolution hash encoder for plane " + }, + { + "bbox": [ + 305, + 371, + 545, + 444 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{\\mathrm{AB}}" + }, + { + "bbox": [ + 305, + 371, + 545, + 444 + ], + "type": "text", + "content": ". By fusing the outcomes, the conclusive geometric feature " + }, + { + "bbox": [ + 305, + 371, + 545, + 444 + ], + "type": "inline_equation", + "content": "\\mathsf{fg} \\in \\mathbb{R}^{3 \\times LD}" + }, + { + "bbox": [ + 305, + 371, + 545, + 444 + ], + "type": "text", + "content": " is derived as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 326, + 455, + 545, + 470 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 455, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 326, + 455, + 545, + 470 + ], + "type": "interline_equation", + "content": "\\mathrm {f} _ {\\mathrm {x}} = \\mathcal {H} ^ {\\mathrm {X Y}} (x, y) \\oplus \\mathcal {H} ^ {\\mathrm {Y Z}} (y, z) \\oplus \\mathcal {H} ^ {\\mathrm {X Z}} (x, z), \\tag {11}", + "image_path": "d4e2b8f595d06f1d79167845d32bb4cca467319b4c4b659aaf31fe80f98c52aa.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 475, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 475, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 475, + 545, + 533 + ], + "type": "text", + "content": "where the concatenation of features is symbolized by " + }, + { + "bbox": [ + 305, + 475, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 305, + 475, + 545, + 533 + ], + "type": "text", + "content": ", resulting in a " + }, + { + "bbox": [ + 305, + 475, + 545, + 533 + ], + "type": "inline_equation", + "content": "3 \\times LD" + }, + { + "bbox": [ + 305, + 475, + 545, + 533 + ], + "type": "text", + "content": "-channel vector. Employing " + }, + { + "bbox": [ + 305, + 475, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{f_x}" + }, + { + "bbox": [ + 305, + 475, + 545, + 533 + ], + "type": "text", + "content": ", the perspective direction d, the lip feature " + }, + { + "bbox": [ + 305, + 475, + 545, + 533 + ], + "type": "inline_equation", + "content": "f_l" + }, + { + "bbox": [ + 305, + 475, + 545, + 533 + ], + "type": "text", + "content": ", and the expression feature " + }, + { + "bbox": [ + 305, + 475, + 545, + 533 + ], + "type": "inline_equation", + "content": "f_e" + }, + { + "bbox": [ + 305, + 475, + 545, + 533 + ], + "type": "text", + "content": ", the tri-plane hash's implicit function is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 358, + 541, + 545, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 541, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 358, + 541, + 545, + 556 + ], + "type": "interline_equation", + "content": "\\mathcal {F} ^ {\\mathcal {H}}: (\\mathrm {x}, \\mathrm {d}, f _ {l}, f _ {e}; \\mathcal {H} ^ {3}) \\rightarrow (\\mathrm {c}, \\sigma), \\tag {12}", + "image_path": "086b801da625d4fc92d005dc96df557e6a93cc9a1b898472c00bc10bdfcd519b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 562, + 544, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 562, + 544, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 562, + 544, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 562, + 544, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^3" + }, + { + "bbox": [ + 305, + 562, + 544, + 586 + ], + "type": "text", + "content": " amalgamates the triad of planar hash encoders, as illustrated in Eq. 10." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "text", + "content": "Our training employs a two-step coarse-to-fine strategy, initially using MSE loss to assess the difference between predicted " + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\hat{C} (\\mathbf{r})" + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "text", + "content": " and actual image colors " + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "inline_equation", + "content": "C(\\mathbf{r})" + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "text", + "content": ". Recognizing MSE's limitations in detail capture, we advance to a refinement stage, incorporating LPIPS loss for enhanced detail, similar to ER-NeRF [21]. We extract random patches " + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "text", + "content": " from the image, combining LPIPS (weighted by " + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "text", + "content": ") with MSE for improved detail representation, as shown:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 691, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 691, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 315, + 691, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\sum_ {\\mathrm {r}} \\| C (\\mathrm {r}) - \\hat {C} (\\mathrm {r}) \\| _ {2} + \\lambda \\times \\mathcal {L} _ {\\mathrm {L P I P S}} (\\hat {\\mathcal {P}}, \\mathcal {P}). \\tag {13}", + "image_path": "fbe0ff78aec55bc51a4d13b0e95ad24620fba0cec5cfa7e11f38b689b321450b.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "670" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 61, + 547, + 231 + ], + "blocks": [ + { + "bbox": [ + 50, + 61, + 547, + 231 + ], + "lines": [ + { + "bbox": [ + 50, + 61, + 547, + 231 + ], + "spans": [ + { + "bbox": [ + 50, + 61, + 547, + 231 + ], + "type": "table", + "html": "
MethodsPSNR ↑LPIPS ↓MS-SSIM ↑FID ↓NIQE ↓BRISQUE ↓LMD ↓AUE ↓LSE-C ↑
GANWav2Lip (ACM MM 20 [34])33.43850.06970.978116.022814.536744.26594.96302.90299.2387
VideoReTalking (SIGGRAPH Asia 22 [7])31.79230.04880.96809.206314.241043.04655.85753.33087.9683
DINet (AAAI 23 [50])31.64750.04430.96409.430014.685040.36504.37253.68756.5653
TalkLip (CVPR 23 [41])32.51540.07820.969718.499714.638546.67175.86052.95795.9472
IP-LAP (CVPR 23 [53])35.15250.04430.98038.212514.640042.07503.33502.84004.9541
NeRFAD-NeRF (ICCV 21 [14])26.72910.15360.911128.986214.909155.46672.99955.54814.4996
RAD-NeRF (arXiv 22 [38])31.77540.07780.94528.657013.443344.68922.91155.09585.5219
GeneFace (ICLR 23 [44])24.81650.11780.875321.708413.335346.50614.28595.45275.1950
ER-NeRF (ICCV 23 [21])32.52160.03340.95015.293613.704834.73612.81374.18735.7749
SyncTalk (w/o Portrait)35.35420.02350.97693.924713.133333.29542.57142.57968.1331
SyncTalk (Portrait)37.40170.01130.98412.707014.216537.30422.50433.20748.0263
", + "image_path": "b88f1d17a9a426abb6f5a59c4c6b08dced8705a4b4a7c2131677b97e5b246d63.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 272, + 289, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 272, + 289, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 289, + 429 + ], + "type": "text", + "content": "Portrait-Sync Generator. During the training process, to address NeRF's limitations in capturing fine details like hair strands and dynamic backgrounds, we introduce a Portrait-Sync Generator with two key sections. First, NeRF renders the face area " + }, + { + "bbox": [ + 46, + 272, + 289, + 429 + ], + "type": "inline_equation", + "content": "(F_r)" + }, + { + "bbox": [ + 46, + 272, + 289, + 429 + ], + "type": "text", + "content": ", creates " + }, + { + "bbox": [ + 46, + 272, + 289, + 429 + ], + "type": "inline_equation", + "content": "G(F_r)" + }, + { + "bbox": [ + 46, + 272, + 289, + 429 + ], + "type": "text", + "content": " through Gaussian blur, and then uses our synchronized head pose to be able to merge with the original image " + }, + { + "bbox": [ + 46, + 272, + 289, + 429 + ], + "type": "inline_equation", + "content": "(F_o)" + }, + { + "bbox": [ + 46, + 272, + 289, + 429 + ], + "type": "text", + "content": " to enhance hair detail fidelity. Second, when the head and torso are combined, if the character in the source video speaks while the generated face is silent, a dark gap area might appear, as shown in Fig. 5 (b). We fill these areas with the average neck color " + }, + { + "bbox": [ + 46, + 272, + 289, + 429 + ], + "type": "inline_equation", + "content": "(C_n)" + }, + { + "bbox": [ + 46, + 272, + 289, + 429 + ], + "type": "text", + "content": ". This approach results in more realistic details and improved visual quality through the Portrait-Sync Generator." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 443, + 128, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 128, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 128, + 456 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 464, + 174, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 174, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 174, + 477 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": "Dataset. To ensure a fair comparison, we use the same well-edited video sequences from [14, 21, 44], including English and French. The average length of these videos is approximately 8,843 frames, and each video is recorded at 25 FPS. Except for the video from AD-NeRF [14], which has a resolution of " + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "450 \\times 450" + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": ", all other videos have a resolution of " + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": ", with the character-centered." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 568, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 628 + ], + "type": "text", + "content": "Comparison Baselines. We compare our method with five GAN-based methods, including Wav2Lip [34], Video-ReTalking [7], DINet [50], TalkLip [41], and IP-LAP [53], and NeRF-based methods such as AD-NeRF [14], RAD-NeRF [38], GeneFace [44], and ER-NeRF [21]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": "Implementation Details. In the coarse stage, the portrait head is trained for 100,000 iterations and 25,000 in the fine stage, sampling " + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "inline_equation", + "content": "256^2" + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": " rays per iteration using a 2D hash encoder (" + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "inline_equation", + "content": "L = 14" + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "inline_equation", + "content": "F = 1" + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": "). We employ the AdamW optimizer [23], with learning rates of 0.01 for the hash encoder and 0.001 for other modules. Total training time is approximately 2 hours on an NVIDIA RTX 3090 GPU." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 270, + 547, + 369 + ], + "blocks": [ + { + "bbox": [ + 46, + 239, + 546, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 546, + 262 + ], + "type": "text", + "content": "Table 1. The quantitative results of the head reconstruction. \"Portrait\" refers to the use of the Portrait-Sync Generator. We achieve state-of-the-art performance on most metrics. We highlight best and second-best results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 270, + 547, + 369 + ], + "lines": [ + { + "bbox": [ + 308, + 270, + 547, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 270, + 547, + 369 + ], + "type": "table", + "html": "
MethodsAudio AAudio B
LSE-D ↓LSE-C ↑LSE-D ↓LSE-C ↑
DINet (AAAI 23 [50])8.50315.69568.20385.1134
TalkLip (CVPR 23 [41])8.76155.74498.70195.5359
IP-LAP (CVPR 23 [53])9.80373.85789.11024.389
GeneFace (ICLR 23 [44])9.54514.29339.66753.7342
ER-NeRF (ICCV 23 [21])11.8132.407610.73383.0242
SyncTalk (Ours)7.72116.66598.02486.2596
", + "image_path": "12d62942c6b37fe3355ce4596ebccc24d3a79b633ea42f4b18bed8dfdc0cfb45.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 375, + 546, + 409 + ], + "lines": [ + { + "bbox": [ + 305, + 375, + 546, + 409 + ], + "spans": [ + { + "bbox": [ + 305, + 375, + 546, + 409 + ], + "type": "text", + "content": "Table 2. The quantitative results of the lip synchronization. We use two different audio samples to drive the same subject, then highlight best and second-best results." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 419, + 441, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 419, + 441, + 432 + ], + "spans": [ + { + "bbox": [ + 306, + 419, + 441, + 432 + ], + "type": "text", + "content": "4.2. Quantitative Evaluation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 510 + ], + "type": "text", + "content": "Full Reference Quality Assessment. In terms of image quality, we use full reference metrics such as Peak Signal-to-Noise Ratio (PSNR), Learned Perceptual Image Patch Similarity (LPIPS) [46], Multi-Scale Structure Similarity (MS-SSIM), and Frechet Inception Distance (FID) [15] as evaluation metrics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "text", + "content": "No Reference Quality Assessment. In high PSNR images, texture details may not align with human visual perception [47]. For more precise output definition and comparison, we use two No Reference methods: the Natural Image Quality Evaluator (NIQE)[29] and the Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE)[28]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "text", + "content": "Synchronization Assessment. For synchronization, we use landmark distance (LMD) to measure the synchronicity of facial movements, action units error (AUE) [4] to assess the accuracy of facial movements, and introduce Lip Sync Error Confidence (LSE-C), consistent with Wav2Lip [34], to evaluate the synchronization between lip movements and audio." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "Evaluation Results. The evaluation results of the head reconstruction are shown in Tab. 1. We compare recent methods based on GAN and NeRF. It can be observed that our image quality is superior to other methods in all aspects. In terms of synchronization, our results surpass most methods." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "671" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 58, + 547, + 370 + ], + "blocks": [ + { + "bbox": [ + 47, + 58, + 547, + 370 + ], + "lines": [ + { + "bbox": [ + 47, + 58, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 58, + 547, + 370 + ], + "type": "image", + "image_path": "f7d0ae1aceef51e6c939e47d81b1e346580de459ee560e4411f40a69f7591ab0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 376, + 546, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 376, + 546, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 546, + 399 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of facial synthesis by different methods. Our method has the best visual effect on lip movements and facial expressions without the problem of separation of head and torso. Please zoom in for better visualization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 410, + 289, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 289, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 289, + 625 + ], + "type": "text", + "content": "We compare the two output modes of SyncTalk, one processed through the Portrait-Sync Generator and one without it. After processing through the Portrait-Sync Generator, hair details are restored, and the image quality is improved. Since we can maintain the subject's identity well, we surpass GAN-based methods in image quality. Thanks to the sync of lip, expression, and pose, we also outperform NeRF-based methods in image quality. Especially in terms of the LPIPS metric, our method has three times improvement compared to the previous state-of-the-art method ER-NeRF [21]. We compare the latest SOTA method drivers using out-of-distribution (OOD) audio, and the results are shown in Tab. 2. We introduce Lip Synchronization Error Distance (LSE-D) and Confidence (LSEC) for lip-audio sync evaluation, aligning with [34]. Our method shows state-of-the-art lip synchronization, overcoming small-sample NeRF limitations by incorporating a pre-trained audio-visual encoder for lip modeling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "We also test the rendering speed. On an NVIDIA RTX 3090 GPU, and having preloaded the data onto the GPU, the head is outputted at 52 FPS with a resolution of " + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": ". In Portrait mode, using the Portrait-Sync Generator, we achieve rendering speeds of 50 FPS with our CUDA acceleration. This far exceeds the video input speed of 25 FPS, allowing for real-time generation of video streams." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 410, + 434, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 410, + 434, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 410, + 434, + 422 + ], + "type": "text", + "content": "4.3. Qualitative Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 636 + ], + "type": "text", + "content": "Evaluation Results. To more intuitively evaluate image quality, we display a comparison between our method and other methods in Fig. 4. In this figure, it can be observed that SyncTalk demonstrates more precise and more accurate facial details. Compared to Wav2Lip [34], our method better preserves the subject's identity while offering higher fidelity and resolution. Against IP-LAP [53], our method excels in lip shape synchronization, primarily due to the audio-visual consistency brought by the audio-visual encoder. Compared to GeneFace [44], our method can accurately reproduce actions such as blinking and eyebrow-raising through expression sync. In contrast to ER-NeRF [21], our method avoids the separation between the head and body through the Pose-Sync Stabilizer and generates more accurate lip shapes. Our method achieves the best overall visual effect; we recommend watching the supplementary video for comparison." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "text", + "content": "User Study. To provide a more comprehensive evaluation of the proposed model, we design an exhaustive user study questionnaire. We extract 24 video clips, each lasting more than 10 seconds, which include head poses, facial expressions, and lip movements. Each method is represented by three of these clips. We invite 35 participants" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "672" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 71, + 542, + 149 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 542, + 149 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 542, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 542, + 149 + ], + "type": "table", + "html": "
Wav2Lip [34]DINet [50]TalkLip [41]IP-LAP [53]AD-NeRF [14]GeneFace [44]ER-NeRF [21]SyncTalk
Lip-sync Accuracy3.8393.6962.8933.1612.6962.9823.1894.304
Exp-sync Accuracy3.5363.4822.6073.4112.2503.0362.9464.036
Pose-sync Accuracy3.5713.5712.8753.6962.2322.9292.6073.980
Image Quality2.5002.6962.0543.5712.4643.4823.0364.054
Video Realness2.9292.4292.4293.1612.0362.7322.5184.018
", + "image_path": "fe64f0830fe8272cad41bbcefd7800e80cb8b51541f9ea9b253302b109e6ea71.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 190, + 285, + 330 + ], + "blocks": [ + { + "bbox": [ + 46, + 158, + 546, + 180 + ], + "lines": [ + { + "bbox": [ + 46, + 158, + 546, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 546, + 180 + ], + "type": "text", + "content": "Table 3. User Study. Rating is on a scale of 1-5; the higher the better. The term \"Exp-sync Accuracy\" is an abbreviation for \"Expression-sync Accuracy\". We highlight best and second-best results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 190, + 285, + 330 + ], + "lines": [ + { + "bbox": [ + 49, + 190, + 285, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 190, + 285, + 330 + ], + "type": "table", + "html": "
PSNR ↑LPIPS ↓LMD ↓
Ours37.3110.01212.8032
replace Audio-Visual Encoder with Hubert [16]33.5160.02763.3961
replace Facial Animation Capture with ER-NeRF [21]'s30.2730.04153.0516
w/o Facial-Aware Masked-Attention36.5360.01652.9139
w/o Head-Sync Stabilizer28.9840.06343.5373
w/o Portrait-Sync Generator32.2390.03952.8154
", + "image_path": "a3232178b6751e92d9e9df513fad385c645650286e1f0fe520d7d6ffe6e49853.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 338, + 287, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 361 + ], + "type": "text", + "content": "Table 4. Ablation study for our components. We show the PSNR, LPIPS, and LMD in different cases." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 373, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 541 + ], + "type": "text", + "content": "to provide scores. The questionnaire is designed using the Mean Opinion Score (MOS) scoring protocol, asking participants to rate the generated videos from five perspectives: (1) Lip-sync Accuracy, (2) Expression-sync Accuracy, (3) Pose-sync Accuracy, (4) Image Quality, and (5) Video Realness. On average, participants take 19 minutes to complete the questionnaire, with a standardized Cronbach " + }, + { + "bbox": [ + 46, + 373, + 287, + 541 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 373, + 287, + 541 + ], + "type": "text", + "content": " coefficient of 0.96. The results of the User Study are displayed in Tab. 3. SyncTalk surpasses previous methods in all evaluations. Furthermore, we achieve the highest score in video authenticity, surpassing the second-place method, IP-LAP [53], by a margin of " + }, + { + "bbox": [ + 46, + 373, + 287, + 541 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 373, + 287, + 541 + ], + "type": "text", + "content": ". User studies indicate that our method can generate visually excellent quality as perceived by humans, achieving high realism." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 551, + 141, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 141, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 141, + 563 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 569, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 628 + ], + "type": "text", + "content": "We conduct an ablation study to examine the contributions of different components in our model. We select three core metrics for evaluation: PSNR, LPIPS, and LMD. We select a subject named \"May\" for testing, and the results are presented in Tab. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "The Audio-Visual Encoder provides the primary lip sync information. When this module is replaced, all three metrics become worse, with the LMD error increasing by " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "21.15\\%" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " in particular, indicating decreased lip motion synchronization, as shown in Fig. 5 (a), and showing that our audio-visual encoders can extract accurate lip features. Replacing Facial Animation Capture with ER-NeRF [21]'s blink module" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 199, + 545, + 263 + ], + "blocks": [ + { + "bbox": [ + 307, + 190, + 544, + 198 + ], + "lines": [ + { + "bbox": [ + 307, + 190, + 544, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 190, + 544, + 198 + ], + "type": "text", + "content": "w/o Audio-Visual Encoder w/ Audio-Visual Encoder w/o Head-Sync Stabilizer w/ Head-Sync Stabilizer" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 199, + 545, + 263 + ], + "lines": [ + { + "bbox": [ + 309, + 199, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 309, + 199, + 545, + 263 + ], + "type": "image", + "image_path": "612756322144cdd4323c7c0cd0c4bfcfef36b696637d5a02aa5d813042a41744.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 271, + 545, + 294 + ], + "lines": [ + { + "bbox": [ + 305, + 271, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 305, + 271, + 545, + 294 + ], + "type": "text", + "content": "Figure 5. Ablation Study on Audio-Visual Encoder and Head-Sync Stabilizer. Removing them will lead to " + }, + { + "bbox": [ + 305, + 271, + 545, + 294 + ], + "type": "inline_equation", + "content": "(a)" + }, + { + "bbox": [ + 305, + 271, + 545, + 294 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 271, + 545, + 294 + ], + "type": "inline_equation", + "content": "(b)" + }, + { + "bbox": [ + 305, + 271, + 545, + 294 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 308, + 495, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 495, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 495, + 319 + ], + "type": "text", + "content": "affects eyebrow movements and image quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 320, + 545, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 320, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 304, + 320, + 545, + 428 + ], + "type": "text", + "content": "The Facial-Aware Masked-Attention mainly alleviates motion entanglement between the lips and the rest of the face, slightly affecting the image quality after removal. Without the Head-Sync Stabilizer, all metrics significantly declined, notably LPIPS, leading to head pose jitter and head and torso separation, as shown in Fig. 5 (b). The Portrait-Sync Generator restores details like hair. Removing this module impacts the restoration of details like hair, leading to noticeable segmentation boundaries." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 442, + 378, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 442, + 378, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 378, + 454 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 463, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 606 + ], + "type": "text", + "content": "In this paper, we introduce SyncTalk, a highly synchronized NeRF-based method for realistic speech-driven talking head synthesis. Our framework includes a Facial Sync Controller, Head Sync Stabilizer, and Portrait Sync Generator, which maintain subject identity and generate synchronized lip movements, facial expressions, and stable head poses. Through extensive evaluation, SyncTalk demonstrates superior performance in creating realistic and synchronized talking head videos compared to existing methods. We expect that SyncTalk will not only enhance various applications but also inspire further innovation in the field of talking head synthesis." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 621, + 403, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 403, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 403, + 634 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": "This work was supported by National Natural Science Foundation of China (NSFC) under Grant Nos. 62172421 and 62072459, the Outstanding Innovative Talents Cultivation Funded Programs 2023 of Renmin University of China, and the Public Computing Cloud at Renmin University of China." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "673" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence, 44(12):8717-8727, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 203 + ], + "type": "text", + "content": "[2] Dario Amodei, Sundaram Ananthanarayanan, Rishita Anubhai, Jingliang Bai, Eric Battenberg, Carl Case, Jared Casper, Bryan Catanzaro, Qiang Cheng, Guoliang Chen, et al. Deep speech 2: End-to-end speech recognition in english and mandarin. In International conference on machine learning, pages 173-182. PMLR, 2016. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "type": "text", + "content": "[3] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 305 + ], + "type": "text", + "content": "[4] Tadas Baltrusaitis, Marwa Mahmoud, and Peter Robinson. Cross-dataset learning and person-specific normalisation for automatic action unit detection. In 2015 11th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-6. IEEE, 2015. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 306, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 288, + 350 + ], + "type": "text", + "content": "[5] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European conference on computer vision (ECCV), pages 520-535, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 352, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 288, + 407 + ], + "type": "text", + "content": "[6] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7832-7841, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 409, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 288, + 464 + ], + "type": "text", + "content": "[7] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 465, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 288, + 520 + ], + "type": "text", + "content": "[8] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 521, + 288, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 521, + 288, + 587 + ], + "spans": [ + { + "bbox": [ + 53, + 521, + 288, + 587 + ], + "type": "text", + "content": "[9] Dipanjan Das, Sandika Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 408-424. Springer, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "type": "text", + "content": "[10] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[11] Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379, 2022.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[12] Shreyank N Gowda, Dheeraj Pandey, and Shashank Narayana Gowda. From pixels to portraits:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "text", + "content": "A comprehensive survey of talking head generation techniques and applications. arXiv preprint arXiv:2308.16041, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 106, + 547, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 547, + 182 + ], + "type": "text", + "content": "[13] Jiazhi Guan, Zhanwang Zhang, Hang Zhou, Tianshu Hu, Kaisiyuan Wang, Dongliang He, Haocheng Feng, Jingtuo Liu, Errui Ding, Ziwei Liu, et al. Stylesync: High-fidelity generalized and personalized lip sync in style-based generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1505-1515, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 184, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 547, + 239 + ], + "type": "text", + "content": "[14] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 239, + 547, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 547, + 293 + ], + "type": "text", + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 294, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 358 + ], + "type": "text", + "content": "[16] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460, 2021. 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 360, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 414 + ], + "type": "text", + "content": "[17] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 415, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 469 + ], + "type": "text", + "content": "[18] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM transactions on graphics (TOG), 37(4): 1-14, 2018. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 471, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 502 + ], + "type": "text", + "content": "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "text", + "content": "[20] Prajwal KR, Rudrabha Mukhopadhyay, Jerin Philip, Abhishek Jha, Vinay Namboodiri, and CV Jawahar. Towards automatic face-to-face translation. In Proceedings of the 27th ACM international conference on multimedia, pages 1428-1436, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "text", + "content": "[21] Jiahe Li, Jiawei Zhang, Xiao Bai, Jun Zhou, and Lin Gu. Efficient region-aware neural radiance fields for high-fidelity talking portrait synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7568-7578, 2023. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "text", + "content": "[22] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. In European Conference on Computer Vision, pages 106-125. Springer, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 658, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 680 + ], + "type": "text", + "content": "[23] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[24] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "674" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "text", + "content": "[25] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 194 + ], + "type": "text", + "content": "[26] Moustafa Meshry, Saksham Suri, Larry S Davis, and Abhinav Shrivastava. Learned spatial representations for few-shot talking-head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13829-13838, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 195, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 288, + 247 + ], + "type": "text", + "content": "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 250, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 288, + 293 + ], + "type": "text", + "content": "[28] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing, 21(12): 4695-4708, 2012. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 294, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 288, + 326 + ], + "type": "text", + "content": "[29] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "type": "text", + "content": "[30] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 372, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 288, + 426 + ], + "type": "text", + "content": "[31] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "type": "text", + "content": "[32] Ziqiao Peng, Yihao Luo, Yue Shi, Hao Xu, Xiangyu Zhu, Hongyan Liu, Jun He, and Zhaoxin Fan. Selftalk: A self-supervised commutative training diagram to comprehend 3d talking faces. In Proceedings of the 31st ACM International Conference on Multimedia, pages 5292-5301, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 482, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 288, + 536 + ], + "type": "text", + "content": "[33] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20687-20697, 2023. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 537, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 591 + ], + "type": "text", + "content": "[34] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 2, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "type": "text", + "content": "[35] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In European Conference on Computer Vision, pages 666-682. Springer, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 637, + 288, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 637, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 637, + 288, + 680 + ], + "type": "text", + "content": "[36] Linsen Song, Wayne Wu, Chen Qian, Ran He, and Chen Change Loy. Everybody's talkin': Let me talk as you want. IEEE Transactions on Information Forensics and Security, 17:585-598, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "type": "text", + "content": "[37] Yasheng Sun, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Zhibin Hong, Jingtuo Liu, Errui Ding, Jingdong Wang, Zwei Liu, and Koike Hideki. Masked lip-sync prediction by" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "audio-visual contextual exploitation in transformers. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 97, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 547, + 152 + ], + "type": "text", + "content": "[38] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 153, + 547, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 547, + 218 + ], + "type": "text", + "content": "[39] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVI 16, pages 716–731. Springer, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "type": "text", + "content": "[40] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. International Journal of Computer Vision, 128:1398-1413, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 254, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 547, + 308 + ], + "type": "text", + "content": "[41] Jiadong Wang, Xinyuan Qian, Malu Zhang, Robby T Tan, and Haizhou Li. Seeing what you said: Talking face generation guided by a lip reading expert. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14653-14662, 2023. 2, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 310, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 353 + ], + "type": "text", + "content": "[42] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "type": "text", + "content": "[43] Shunyu Yao, RuiZhe Zhong, Yichao Yan, Guangtao Zhai, and Xiaokang Yang. Dfa-nerf: Personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791, 2022. 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[44] Zhenhui Ye, Ziyue Jiang, Yi Ren, Jinglin Liu, JinZheng He, and Zhou Zhao. Geneface: Generalized and high-fidelity audio-driven 3d talking face synthesis. arXiv preprint arXiv:2301.13430, 2023. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 445, + 545, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 500 + ], + "type": "text", + "content": "[45] Chenxu Zhang, Yifan Zhao, Yifei Huang, Ming Zeng, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Facial: Synthesizing dynamic talking face with implicit attribute learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3867-3876, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "text", + "content": "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 557, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 611 + ], + "type": "text", + "content": "[47] Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. Ranksrgan: Generative adversarial networks with ranker for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3096-3105, 2019. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 613, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 679 + ], + "type": "text", + "content": "[48] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 681, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 547, + 714 + ], + "type": "text", + "content": "[49] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "675" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 520 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661-3670, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[50] Zhimeng Zhang, Zhipeng Hu, Wenjin Deng, Changjie Fan, Tangjie Lv, and Yu Ding. Dinet: Deformation inpainting network for realistic face visually dubbing on high resolution video. arXiv preprint arXiv:2303.03988, 2023. 2, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "type": "text", + "content": "[51] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "type": "text", + "content": "[52] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21057-21067, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 317 + ], + "type": "text", + "content": "[53] Weizhi Zhong, Chaowei Fang, Yinqi Cai, Pengxu Wei, Gangming Zhao, Liang Lin, and Guanbin Li. Identity-preserving talking face generation with landmark and appearance priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2023. 2, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 363 + ], + "type": "text", + "content": "[54] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI conference on artificial intelligence, pages 9299–9306, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 429 + ], + "type": "text", + "content": "[55] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 431, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 431, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 431, + 287, + 474 + ], + "type": "text", + "content": "[56] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG), 39(6):1-15, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 476, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 287, + 520 + ], + "type": "text", + "content": "[57] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4574-4584, 2023. 3" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "676" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/908ee619-1331-452b-9ccb-4e5b4d116ac0_content_list.json b/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/908ee619-1331-452b-9ccb-4e5b4d116ac0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..671092c7432bcba2c983b12c2a08b1a69b12dadc --- /dev/null +++ b/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/908ee619-1331-452b-9ccb-4e5b4d116ac0_content_list.json @@ -0,0 +1,1685 @@ +[ + { + "type": "text", + "text": "Synergistic Global-space Camera and Human Reconstruction from Videos", + "text_level": 1, + "bbox": [ + 109, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yizhou Zhao $^{1}$ ; Tuanfeng Yang Wang $^{2}$ , Bhiksha Raj $^{1}$ , Min Xu $^{1}$ , Jimei Yang $^{2}$ , Chun-Hao Paul Huang $^{2}$ $^{1}$ Carnegie Mellon University \n $^{2}$ Adobe Research", + "bbox": [ + 86, + 179, + 880, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yizhouz,bhiksha,mxul}@cs.cmu.edu,{yangtwan,jimyang,chunhaoh}@adobe.com", + "bbox": [ + 165, + 218, + 797, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7af274f60e1e4a33d3c56170353e17ad29edf8edff3f22c7ddc7e0ec95f4dab1.jpg", + "image_caption": [ + "Figure 1. Our SynCHMR recovers metric-scale camera trajectories (color pyramids), human meshes, and scene point clouds in a global coordinate from casual videos by joining forces of Human Mesh Recovery (HMR) and Simultaneous Localization and Mapping (SLAM)." + ], + "image_footnote": [], + "bbox": [ + 81, + 250, + 890, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 551, + 313, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Remarkable strides have been made in reconstructing static scenes or human bodies from monocular videos. Yet, the two problems have largely been approached independently, without much synergy. Most visual SLAM methods can only reconstruct camera trajectories and scene structures up to scale, while most HMR methods reconstruct human meshes in metric scale but fall short in reasoning with cameras and scenes. This work introduces Synergistic Camera and Human Reconstruction (SynCHMR) to marry the best of both worlds. Specifically, we design Human-aware Metric SLAM to reconstruct metric-scale camera poses and scene point clouds using camera-frame HMR as a strong prior, addressing depth, scale, and dynamic ambiguities. Conditioning on the dense scene recovered, we further learn a Scene-aware SMPL Denoiser to enhance world-frame HMR by incorporating spatio-temporal coherency and dynamic scene constraints. Together, they lead to consistent reconstructions of camera trajectories, human meshes, and dense scene point clouds in a common world frame.", + "bbox": [ + 75, + 575, + 472, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 551, + 630, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Physically plausible 3D human motion reconstruction from monocular videos is a long-standing problem in computer vision and graphics and has many applications in character animation, VFX, video games, sports, and healthcare. It requires estimating 3D humans across video frames in a common coordinate even with a moving camera. While human mesh recovery (HMR) has made significant progress recently [55], most existing methods typically estimate 3D humans in the camera coordinate by one frame at a time and fail to disambiguate camera motion. It calls for methods to jointly reconstruct 3D human and camera motion in a consistent global coordinate system from monocular videos. In other words, taking a video captured by a moving camera as input, the method should recover both temporally and spatially coherent movements of human bodies and cameras.", + "bbox": [ + 496, + 575, + 890, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Intuitively, if the accurate camera motion is given, one can transform the bodies from individual camera frames to a common world frame by multiplying the inverse of camera extrinsic matrices. In practice, with humans moving in the scene, estimating the camera motion of a video is still an open challenge in monocular SLAM [1]. It not only", + "bbox": [ + 496, + 810, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Part of this work was done when interned at Adobe Research.", + "bbox": [ + 94, + 886, + 426, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1216", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b62b71ad1330bd3eee757cc9f16d644b2d4915e10b4103900824a204b2ef7dbe.jpg", + "image_caption": [ + "(a) Depth ambiguity" + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 467, + 169 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a785b7734ae7970d02717379d3d9844643c75e2a39dd7bb860031a3eaa85d7e7.jpg", + "image_caption": [ + "(b) Scale ambiguity" + ], + "image_footnote": [], + "bbox": [ + 86, + 186, + 465, + 234 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8c84ab4284bfe3a986e9a31ab134b8832a8e42e81327e4ad39cd17f17622c302.jpg", + "image_caption": [ + "(c) Dynamic ambiguity", + "Figure 2. Illustration of three types of ambiguities in visual SLAM. We show SLAM reconstruction results from DROID-SLAM [54]. (a) Depth ambiguity occurs when there are only minor camera translations between different views. This can lead to geometric failures in reconstruction such as the folded back corridor in the side view. (b) Scale ambiguity is inherent in monocular SLAM systems and requires additional reference for disambiguation. (c) Dynamic ambiguity gets pronounced when moving foregrounds dominate frames. Over-reliance on foreground key points will result in incorrect camera trajectories." + ], + "image_footnote": [], + "bbox": [ + 86, + 253, + 467, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "falls short in capturing accurate depths on views with small camera translations but more crucially, only estimates scene structures and camera trajectories up to scale. The human motion also breaks the static key point assumption in the bundle adjustment. As a result, one needs additional reference to disambiguate the depth, the scale, and the dynamic as illustrated in Fig. 2.", + "bbox": [ + 75, + 521, + 468, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To leverage SLAM results in HMR pipelines, current world-frame HMR methods often refine camera poses by integrating either partial camera parameters, such as a global scale of the translation [62], or full extrinsic matrices [15, 30, 46] in an optimization-based framework. However, their optimization-based nature leads to complex multi-stage schemes, making the overall pipeline unnecessarily slow and easy to break.", + "bbox": [ + 75, + 628, + 468, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we explore a fundamentally different way to marry the best of HMR and SLAM. A 2D object can first be lifted from the image plane to the camera frame and then transformed into a common 3D space. This two-step process coincides with the combination of camera-frame HMR, which brings imaged 2D humans to 3D camera frames, and SLAM, which estimates the camera-to-world transformation. Noticing these correspondences, we leverage camera-frame HMR as a strong prior to bridge from the image plane to the camera frame for disambiguating SLAM, and utilize", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SLAM reconstructions to constrain the transformation of human meshes from individual camera frames to a common global space. The overall pipeline thus results in a better synergy of the two, which we dub Synergistic Camera and Human Reconstruction (SynCHMR).", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We design SynCHMR based on several insights. First, despite camera-frame HMR methods cannot reconstruct humans in a coherent global frame, the estimated body dimension and location still provide cues to disambiguate SLAM. Unlike SLAHMR [62] which applies SLAM out of the box and corrects the scale afterward, we endow the SLAM process with human meshes from camera-frame HMR to address ambiguities. To this end, we capitalize on estimated absolute depths to provide pseudo-RGB-D inputs for SLAM [54] and confine the bundle adjustment to static backgrounds. Since current depth estimation methods [2, 43] predicts either relative depth maps or depths with data biases, we propose to calibrate their outputs by aligning with estimated human bodies in the camera frame [9]. With these priors, SLAM knows the depth, scale, and dynamic information from HMR and consequently estimates less ambiguous scene structures and camera poses.", + "bbox": [ + 496, + 169, + 890, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Next, we place human meshes in the common coordinate recovered by SLAM. The gap between human tracks transformed from camera frames and their real plausible world-frame motions stems from two sources of error: noise induced by camera-frame HMR and by SLAM. Motion prior models [14, 44, 68] can be used for denoising purposes as they contribute to the temporal coherence of worldframe human tracks. However, their exclusive focus on human modeling either leaves them agnostic to the underlying scenes [14] or assumes the scene is a simple ground plane [44, 68]. Our intuition is that when placing a human, static elements of the scene, such as the ground, and dynamic components like moving objects are both possible to be in contact with the human, thereby providing clues for placing the body coherently and compatibly with the scene. We hence introduce a Scene-aware SMPL Denoiser that learns to denoise the transformed human tracks by considering both temporal consistencies of moving humans and implicit constraints from dynamic scenes. This global awareness makes it more flexible for in-the-wild videos.", + "bbox": [ + 496, + 428, + 892, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summed up as follows:", + "bbox": [ + 517, + 732, + 839, + 747 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a novel pipeline, SynCHMR, that takes a monocular video as input and reconstructs human motions, camera trajectory and dense scene point clouds all in one global coordinate, as shown in Fig. 1, whereas current world-frame HMR methods [30, 62, 65] can recover only an estimated or pre-defined ground plane.", + "- We propose a novel Human-aware Metric SLAM process to robustly calibrate estimated depth with estimated human meshes, resulting in metric-scale camera pose estimation and metric-scale scene reconstruction." + ], + "bbox": [ + 500, + 750, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1217", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We present Scene-aware SMPL Denoising that enforces spatiotemporal coherencies and applies dynamic scene constraints on world-frame human meshes. Notably, this is achieved without requiring extra annotations or heuristic designs to decide which part of a human should be interacting with the scene [49] and which region in the scene is most likely to be in contact with humans [38, 62].", + "bbox": [ + 76, + 90, + 468, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 214, + 218, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There is considerable prior arts of HMR. We briefly discuss how they adopt different camera models and refer the readers to [1] for a more comprehensive review.", + "bbox": [ + 76, + 241, + 467, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "HMR from a single image. State-of-the-art (SOTA) methods use parametric body models [21, 37, 41, 59] and estimate the parameters either by fitting to detected image features [3, 41, 58] or by regressing directly from pixels with deep neural networks [9, 11, 19, 22, 23, 28, 31, 32, 35, 45, 50, 60, 66, 67]. These approaches assume weak perspective/orthographic projection or pre-define the focal length as a large constant for all images. Kissos et al. [26] show that replacing focal length with a constant closer to ground truth alleviate the body tilting problem. SPEC [29] and Zolly [57] estimate focal length to account for perspective distortion. CIUFF [34] takes into account the location of humans in images to regress better poses in the camera coordinates.", + "bbox": [ + 75, + 291, + 467, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Many of these camera-frame HMR methods assume zero camera rotation, which entangles body rotation and camera rotation. When applied on video data, they fail to reconstruct humans in a coherent global space since they operate in a per-frame manner and hence cannot reason about how the camera moves across frames.", + "bbox": [ + 75, + 489, + 467, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "HMR from videos aims to regress a series of body parameters from a temporal sequence. It opens up new problems such as whether the reconstructed bodies are in a common global coordinate or not. Some temporal methods consider a static camera [38, 44, 68], which makes the camera space a natural choice of the common coordinate. The challenge of coherent global space emerges when the camera moves. Early methods [5, 24, 27] show promising results on videos of dynamic cameras. Despite the reconstructed human meshes look great when overlaid on images, they do not share a common coordinate in 3D.", + "bbox": [ + 75, + 582, + 467, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent HMR methods capitalize on human motion prior to constrain the global trajectories in the world space, which in turn implicitly disentangles human movement from camera movement. GLAMR [65] consider a data-driven prior models learned on large-scale MoCap database e.g. AMASS [39], while D&D [33] and Yu [64] consider physic-inspired prior. These world-frame HMR methods often struggle on noise in local poses caused by partial occlusions, which is very common in in-the-wild videos with close-up shots and crowd scenes. Kaufmann et al. [25] and", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "BodySLAM++ [16] circumvent this problem by employing IMU sensors to provide more robust body estimates but require extra sensory devices. To fully disentangle human and camera motion, another line of work [15, 30, 36, 46] leverages state-of-the-art SLAM techniques, e.g. [47, 54, 70], to explicitly estimate camera motion from the input video and infer the body parameters in the world coordinate of SLAM. Closest to us is SLAHMR [62] which solves for a global scale to connect the pre-computed SLAM results and body trajectories. To carefully guide the optimization process, these methods tend to have complex, multi-stage optimization schemes, making the overall pipeline easy to break and unnecessarily slow.", + "bbox": [ + 496, + 90, + 890, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that in stark contrast to the methods above, which either assume or estimate a simple ground plane as scene representation, SynCHMR reconstructs dense scenes from in-the-wild videos without pre-scanning with extra devices a priori like in [6, 7, 12, 13, 17, 61, 69]. We provide detailed comparisons with these world-frame HMR in Supp. Mat.", + "bbox": [ + 496, + 287, + 890, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 390, + 589, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Taking as input an RGB video $\\{\\mathbf{I}_t\\in \\mathbb{R}^{H\\times W\\times 3}\\}_{t = 1}^T$ with $T$ frames and $N$ people in the scene, we aim to recover human meshes $\\{\\mathbf{V}_{nt}^{\\mathrm{w}}\\in \\mathbb{R}^{3\\times 6890}\\}_{n = 1,t = 1}^{N,T}$ , dynamic scene point clouds $\\{\\mathbf{P}_t^{\\mathrm{wm}}\\in \\mathbb{R}^{H\\times W\\times 3}\\}_{t = 1}^T$ , and corresponding camera poses $\\{\\mathbf{G}_t^{\\mathrm{m}}\\in \\mathrm{SE}(3)\\}_{t = 1}^T$ in a common world coordinate system. The superscripts ${}^{\\mathrm{W}},{}^{\\mathrm{C}}$ , and ${}^{\\mathrm{m}}$ denote the world frame, the camera frame, and the metric scale, respectively. To this aim, we propose a two-phase alternative conditioning pipeline as depicted in Fig. 3. In the first phase, we calibrate camera motion by injecting a camera-frame human prior to SLAM. This resolves depth, scale, and dynamic ambiguities, yielding metric-scale camera poses and dynamic point clouds. Subsequently, in the second phase, we transform the camera-frame human tracks into the world frame and utilize the dynamic point clouds obtained in the first phase for conditional denoising.", + "bbox": [ + 496, + 414, + 890, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 667, + 638, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 SLAM", + "text_level": 1, + "bbox": [ + 500, + 691, + 601, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a monocular RGB video $\\{\\mathbf{I}_t\\}_{t = 1}^T$ DROID-SLAM [54] solves a dense bundle adjustment for a set of camera poses $\\{\\mathbf{G}_t\\in \\mathrm{SE}(3)\\}_{t = 1}^T$ and inverse depths $\\{\\mathbf{d}_t\\in \\mathbb{R}_+^{H\\times W}\\}_{t = 1}^T$ . To update these estimations, it first computes a dense correspondence field $\\mathbf{p}_{ij}\\in \\mathbb{R}^{H\\times W\\times 2}$ based on reprojection for each pair of frames $(i,j)$ ..", + "bbox": [ + 496, + 713, + 890, + 806 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {i j} = \\Pi \\left(\\mathbf {G} _ {i j} \\circ \\Pi^ {- 1} \\left(\\mathbf {p} _ {i}, \\frac {1}{\\mathbf {d} _ {j}}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 814, + 890, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_i\\in \\mathbb{R}^{H\\times W\\times 2}$ is a grid of pixel coordinates in frame $i$ $\\mathbf{G}_{ij} = \\mathbf{G}_j\\circ \\mathbf{G}_i^{-1}$ is the relative pose, and $\\Pi$ and $\\Pi^{-1}$ are the camera projection and inverse projection functions.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1218", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0a5ef57b61c9a023648d8f342766ae6edf7f543a99cd003fd11a52004ae42588.jpg", + "image_caption": [ + "Figure 3. The architecture of SynCHMR. Our pipeline comprises two phases. The first phase, Human-aware Metric SLAM (Sec. 3.2), infers metric-scale camera poses and metric-scale point clouds by exploiting the camera-frame human prior. The second phase, Scene-aware SMPL Denoising (Sec. 3.3), involves the conditional denoising of world-frame noisy SMPL parameters. These parameters, initialized by transforming from the camera frame, get refined through conditioning on the dynamic point clouds obtained in the first phase. The whole pipeline thus reconstructs humans, scene point clouds, and cameras harmoniously in a common world frame." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 897, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then with a learned neural network, the system predicts a revision flow field $\\mathbf{r}_{ij} \\in \\mathbb{R}^{H \\times W \\times 2}$ and associated confidence map $\\mathbf{w}_{ij} \\in \\mathbb{R}_+^{H \\times W \\times 2}$ to construct the cost function", + "bbox": [ + 75, + 398, + 470, + 446 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\Sigma} = \\sum_ {(i, j)} \\left\\| \\mathbf {p} _ {i j} ^ {*} - \\Pi \\left(\\mathbf {G} _ {i j} ^ {\\prime} \\circ \\Pi^ {- 1} \\left(\\mathbf {p} _ {i}, \\frac {1}{\\mathbf {d} _ {i} ^ {\\prime}}\\right)\\right) \\right\\| _ {\\Sigma_ {i j}} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 460, + 468, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_{ij}^{*} = \\mathbf{r}_{ij} + \\mathbf{p}_{ij}$ is the corrected correspondence, $\\| \\cdot \\|_{\\Sigma}$ is the Mahalanobis distance which weighs the error terms with $\\Sigma_{ij} = \\mathrm{diag}\\mathbf{w}_{ij}$ , and $\\mathbf{G}'$ and $\\mathbf{d}'$ are updated poses and inverse depths. Upon this objective, DROID-SLAM considers an additional term that penalizes the squared distance between the measured and predicted depth if the input is with an extra sensor depth channel $\\{\\mathbf{D}_t\\}_{t=1}^T$ .", + "bbox": [ + 75, + 508, + 470, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2 HMR", + "text_level": 1, + "bbox": [ + 76, + 626, + 171, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ 4DHumans [9] for reconstructing camera-frame human meshes from an in-the-wild video. Specifically, it performs per-frame human mesh recovery with an end-to-end transformer architecture and associates them to form human tracks. Each tracked human $n$ in frame $t$ is represented by SMPL [37] parameters as $\\{\\Phi_{nt},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}\\}$ , including global orientation $\\Phi_{nt}\\in \\mathbb{R}^{3\\times 3}$ , body pose $\\theta_{nt}\\in \\mathbb{R}^{22\\times 3\\times 3}$ , shape $\\beta_{nt}\\in \\mathbb{R}^{10}$ , and root translation $\\Gamma_{nt}\\in \\mathbb{R}^3$ . Then the parametric SMPL model can use these parameters to recover a human mesh with vertices $\\mathbf{V}_{nt}\\in \\mathbb{R}^{3\\times 6890}$ in metric scale: $\\mathbf{V}_{nt} = \\mathrm{SMPL}(\\Phi_{nt},\\theta_{nt},\\beta_{nt}) + \\mathbf{\\Gamma}_{nt}$ .", + "bbox": [ + 75, + 650, + 468, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Human-aware Metric SLAM", + "text_level": 1, + "bbox": [ + 76, + 823, + 336, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Preprocessing", + "text_level": 1, + "bbox": [ + 76, + 847, + 230, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To start off, we estimate per-frame depth maps $\\{\\mathbf{D}_t\\}$ with an off-the-shelf depth estimator, ZoeDepth [2] and", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "predict per-frame human instance segmentation masks $\\{\\mathbf{M}_{nt}\\}$ with an image instance segmentation network, Mask2Former [4]. We adapt ZoeDepth for video-consistent depth estimation by choosing a per-video metric head from the majority vote of per-frame routers, for which we dub ZoeDepth $^+$ . While ZoeDepth claims to estimate metric depths, we observe a domain gap when inference on new datasets. Consequently, we only treat its output as up-to-affine depths that need to be further aligned with the metric scale. To aid our optimization with human awareness, we use camera-frame human meshes $\\{\\mathbf{V}_{nt}^{\\mathrm{c}}\\}$ recovered by 4DHumans [9] to introduce a metric prior.", + "bbox": [ + 496, + 398, + 893, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Calibrating Depth with Human Prior", + "text_level": 1, + "bbox": [ + 498, + 590, + 815, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We calibrate the per-frame depths with human meshes in Human-aware Depth Calibration. This involves optimizing two parameters, a world scale $s$ and a world offset $o$ , shared across all frames. During optimization, we linearly transform $\\mathbf{D}_t$ to $\\mathbf{D}_t' = s\\mathbf{D}_t + o$ and unproject these depth maps to camera-frame point clouds $\\{\\mathbf{P}_t^c\\}$ with $\\mathbf{P}_t^c = \\Pi^{-1}(\\mathbf{p}_t, \\mathbf{D}_t')$ . Our intuition is to align the human point cloud $\\mathbf{P}_{nt}^c = \\mathbf{M}_{nt} \\odot \\mathbf{P}_t^c$ with the camera-frame human mesh vertices $\\mathbf{V}_{nt}^c$ in terms of absolute depth and size. To achieve pixel-wise alignment, we use a depth term to pull points on the human point cloud toward their corresponding human mesh vertices along the z-axis", + "bbox": [ + 496, + 614, + 893, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\text {d e p t h}} = \\frac {\\sum_ {n , t} \\| \\mathbf {S} _ {n t} \\odot [ z \\left(\\mathbf {V} _ {n t} ^ {\\mathrm {c}}\\right) - z \\left(\\mathbf {P} _ {n t} ^ {\\mathrm {c}}\\right) ] \\| _ {2} ^ {2}}{\\sum_ {n , t} \\| \\mathbf {S} _ {n t} \\| _ {0}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 805, + 890, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{S}_{nt} = \\rho (\\mathbf{V}_{nt}^{\\mathrm{c}})\\cap \\mathbf{M}_{nt}$ is the intersection of the rasterized human mesh mask $\\rho (\\mathbf{V}_{nt}^{\\mathrm{c}})$ and the instance segmentation mask $\\mathbf{M}_{nt}$ , $z(\\cdot)$ is the rasterized depth, and $\\| \\cdot \\| _0$ is the", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1219", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/441dac30c7eb42a434600800fbdd161d93bfc451e56c5004a14ff0e74d4c3252.jpg", + "image_caption": [ + "Figure 4. The architecture of Scene-aware SMPL Denoiser. World-frame noisy SMPL parameters $\\{\\Phi_{nt}^{\\mathrm{w}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{w}}\\}_{0}$ are first projected by a linear layer and summed with temporal positional embeddings (TPE) to get initial latent humans $\\{\\mathbf{z}_{nt,0}^{\\mathrm{SMPL}}\\}$ . Per-frame point clouds are aggregated to $\\mathbf{x}_{\\mathrm{scene}}$ and encoded with the point encoder $\\mathcal{E}$ . Then we query the encoded scene $\\mathcal{E}(\\mathbf{x}^{\\mathrm{scene}})$ with latent humans $\\{\\mathbf{z}_{nt,0}^{\\mathrm{SMPL}}\\}$ in the scene-conditioned denoiser $\\mathcal{D}$ and feed the result $\\{\\mathbf{z}_{nt,1}^{\\mathrm{SMPL}}\\}$ to prediction heads $\\{\\mathcal{P}_{\\Phi},\\mathcal{P}_{\\theta},\\mathcal{P}_{\\beta},\\mathcal{P}_{\\Gamma}\\}$ to obtain denoised SMPL parameters $\\{\\Phi_{nt}^{\\mathrm{w}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{w}}\\}_{1}$ ." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 472, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "0-norm indicating the number of non-zero pixels on a mask.", + "bbox": [ + 76, + 377, + 467, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the recovered human meshes can be noisy in depth but still have a stable body dimension, we also adopt a size term to leverage the relative position of mesh vertices", + "bbox": [ + 76, + 392, + 468, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\mathrm {d x}} = \\frac {\\sum_ {n , t} \\| \\Delta_ {x} \\left(\\mathbf {V} _ {n t} ^ {\\mathrm {c}} , \\mathbf {S} _ {n t}\\right) - \\Delta_ {x} \\left(\\mathbf {P} _ {n t} ^ {\\mathrm {c}} , \\mathbf {S} _ {n t}\\right) \\| _ {2} ^ {2}}{N T}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 446, + 468, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define $E_{\\mathrm{dy}}$ similarly as $E_{\\mathrm{dx}}$ , where", + "bbox": [ + 76, + 488, + 333, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {*} (\\mathbf {X}, \\mathbf {Y}) = \\left(\\max _ {*} - \\min _ {*} \\right) \\left[ \\Pi^ {- 1} (\\mathbf {Y} \\odot \\Pi (\\mathbf {X}), z (\\mathbf {X})) \\right] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 512, + 468, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $(\\max_{*} - \\min_{*})$ denotes the difference between the maximum value and the minimum value on coordinate $^*$", + "bbox": [ + 76, + 545, + 468, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we have the calibrated depths with optimization", + "bbox": [ + 96, + 575, + 444, + 589 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(s ^ {\\mathrm {m}}, o ^ {\\mathrm {m}}\\right) = \\operatorname {a r g m i n} _ {s, o} \\left(E _ {\\text {d e p t h}} + \\lambda E _ {\\text {s i z e}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 599, + 468, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {t} ^ {\\mathrm {m}} = s ^ {\\mathrm {m}} \\mathbf {D} _ {t} + o ^ {\\mathrm {m}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 619, + 468, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $E_{\\mathrm{size}} = E_{\\mathrm{dx}} + E_{\\mathrm{dy}}$ , and $\\lambda$ is a hyperparameter to balance two energy terms with a default value of 1.", + "bbox": [ + 76, + 647, + 468, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3 Disambiguating SLAM with Calibrated Depth", + "text_level": 1, + "bbox": [ + 76, + 695, + 452, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While DROID-SLAM [54] originally supports RGB-D input mode where the D channel stands for sensor depth, one cannot trivially access sensor depths from in-the-wild videos. Our insight is that an estimated absolute depth can be utilized as a depth prior, albeit noisy. So we combine the original RGB video and the calibrated depth as pseudoRGB-D inputs $\\{\\mathbf{I}_t,\\mathbf{D}_t^{\\mathrm{m}}\\}$ to disambiguate depth and scale. Furthermore, we modify the cost function Eq. (2) to resolve the dynamic ambiguity by masking out dynamic foregrounds in confidence maps", + "bbox": [ + 75, + 718, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma_ {i j} ^ {\\prime} = \\operatorname {d i a g} \\mathbf {w} _ {i j} ^ {\\prime} = \\operatorname {d i a g} \\left(\\left(1 - \\left[ \\mathbf {M} _ {i}, \\mathbf {M} _ {j} \\right]\\right) \\odot \\mathbf {w} _ {i j}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 878, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{M}_i = \\bigcup_n\\mathbf{M}_{ni}$ and $\\mathbf{M}_j = \\bigcup_n\\mathbf{M}_{nj}$ are the union of all human instance masks on their corresponding frame, and $[\\cdot ,\\cdot ]$ is the concatenation operation. As a result, we obtain metric-scale camera poses $\\{\\mathbf{G}_t^{\\mathrm{m}}\\}$ and metric-scale point clouds $\\{\\mathbf{P}_t^{\\mathrm{wm}}\\}$ by disambiguating SLAM with calibrated metric depths", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathbf {G} _ {t} ^ {\\mathrm {m}}, \\mathbf {d} _ {t} ^ {\\mathrm {m}} \\right\\} = \\operatorname {a r g m i n} _ {\\left\\{\\mathbf {G} _ {i j} ^ {\\prime}, \\mathbf {d} _ {i j} ^ {\\prime} \\right\\}} E _ {\\Sigma^ {\\prime}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 194, + 890, + 215 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {t} ^ {\\mathrm {w m}} = \\mathbf {G} _ {t} ^ {\\mathrm {m}} \\circ \\Pi^ {- 1} \\left(\\mathbf {p} _ {t}, \\frac {1}{\\mathbf {d} _ {t} ^ {\\mathrm {m}}}\\right)). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 217, + 890, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Scene-aware SMPL Denoising", + "text_level": 1, + "bbox": [ + 500, + 260, + 767, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1 Initializing Humans with Metric Cameras", + "text_level": 1, + "bbox": [ + 500, + 284, + 844, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To put humans properly in the scene recovered by SLAM, we initialize them by transforming estimated camera-frame SMPL parameters $\\{\\Phi_{nt}^{\\mathrm{c}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{c}}\\}$ to the world frame with camera-to-world transforms $\\{\\mathbf{G}_t^{\\mathrm{m}} = [\\mathbf{R}_t|\\mathbf{t}_t^{\\mathrm{m}}]\\}$ . Given the pelvis as the center of global orientation $\\Phi$ , we have:", + "bbox": [ + 496, + 309, + 890, + 383 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Phi} _ {n t} ^ {\\mathrm {w}} = \\mathbf {R} _ {t} \\boldsymbol {\\Phi} _ {n t} ^ {\\mathrm {c}}, \\quad \\boldsymbol {\\Gamma} _ {n t} ^ {\\mathrm {w}} = \\mathbf {R} _ {t} \\left(\\boldsymbol {\\Gamma} _ {n t} ^ {\\mathrm {c}} + \\mathbf {c}\\right) + \\mathbf {t} _ {t} ^ {\\mathrm {m}} - \\mathbf {c}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 412, + 890, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{c} = \\mathbf{c}(\\beta_{nt})$ is the pelvis location in the shape blend body mesh. Note that we do not need to introduce an extra camera scale as SLAHMR [62] since the camera poses have already been in the metric scale. The root-relative poses $\\theta_{nt}$ and the shapes $\\beta_{nt}$ stay unchanged as in the camera frame. We denote the initialized and the denoised parameters with a suffix 0 and 1 respectively, i.e. $\\{\\Phi_{nt}^{\\mathrm{w}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{w}}\\}_{0,1}$ .", + "bbox": [ + 496, + 441, + 890, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Constraining Humans with Dynamic Scenes", + "text_level": 1, + "bbox": [ + 500, + 569, + 859, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Different from existing works [30, 36, 62, 64] that incorporate energy terms in optimization to apply explicit scene constraints, we propose to learn implicit scene constraints with a Scene-aware SMPL denoiser shown in Fig. 4. The noisy initial SMPL parameters $\\{\\Phi_{nt}^{\\mathrm{w}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{w}}\\}_{0}$ are first projected to a latent space, where it gets further updated by conditioning on implicit scene constraints", + "bbox": [ + 496, + 593, + 890, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {n t, 0} ^ {\\mathrm {S M P L}} = \\operatorname {F C} \\left(\\left[ \\boldsymbol {\\Phi} _ {n t, 0} ^ {\\mathrm {w}}, \\boldsymbol {\\theta} _ {n t, 0}, \\boldsymbol {\\beta} _ {n t, 0}, \\boldsymbol {\\Gamma} _ {n t, 0} ^ {\\mathrm {w}} \\right]\\right) + \\text {T P E}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 710, + 890, + 729 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathbf {z} _ {n t, 1} ^ {\\text {S M P L}} \\right\\} _ {t = 1} ^ {T} = \\mathcal {D} \\left(\\left\\{\\mathbf {z} _ {n t, 0} ^ {\\text {S M P L}} \\right\\} _ {t = 1} ^ {T}, \\mathcal {E} \\left(\\mathbf {x} ^ {\\text {s c e n e}}\\right) + \\text {T P E}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 732, + 890, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where FC is a shared linear layer, TPE is shared temporal positional embeddings, $\\{\\mathbf{z}_{nt,*}^{\\mathrm{SMPL}}\\}_{t = 1}^{T}\\in \\mathbb{R}^{T\\times D}$ is the $D$ dimensional latent for human $n$ , and $\\mathbf{x}^{\\mathrm{scene}}\\in \\mathbb{R}^{L\\times C}$ is the $C$ -channel dynamic scene point clouds with a total number of points $L$ . $\\mathcal{E}$ and $\\mathcal{D}$ refer to the scene encoder and the scene-conditioned denoiser, respectively. We set $C = 7$ which is the concatenation of point coordinates $\\{\\mathbf{P}_t^{\\mathrm{wm}}\\}$ , colors $\\{\\mathbf{I}_t\\}$ , and estimated human semantic segmentation masks $\\{\\mathbf{M}_t = \\bigcup_n\\mathbf{M}_{nt}\\}$ . Following [9], the updated latent", + "bbox": [ + 496, + 762, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1220", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "are decoded with different prediction head $\\mathcal{P}_{(\\cdot)}$ to regress the residual for each SMPL parameter:", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Phi} _ {n t, 1} ^ {\\mathrm {w}} = \\mathcal {P} _ {\\boldsymbol {\\Phi}} \\left(\\mathbf {z} _ {n t, 1} ^ {\\mathrm {S M P L}}\\right) \\boldsymbol {\\Phi} _ {n t, 0} ^ {\\mathrm {w}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 131, + 467, + 150 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} _ {n t, 1} = \\mathcal {P} _ {\\boldsymbol {\\theta}} \\left(\\mathbf {z} _ {n t, 1} ^ {\\mathrm {S M P L}}\\right) \\boldsymbol {\\theta} _ {n t, 0}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 152, + 467, + 170 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\beta} _ {n t, 1} = \\mathcal {P} _ {\\boldsymbol {\\beta}} \\left(\\mathbf {z} _ {n t, 1} ^ {\\mathrm {S M P L}}\\right) + \\boldsymbol {\\beta} _ {n t, 0}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 172, + 467, + 191 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Gamma} _ {n t, 1} ^ {\\mathrm {w}} = \\mathcal {P} _ {\\boldsymbol {\\Gamma}} \\left(\\mathbf {z} _ {n t, 1} ^ {\\mathrm {S M P L}}\\right) + \\boldsymbol {\\Gamma} _ {n t, 0} ^ {\\mathrm {w}}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 194, + 467, + 213 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We apply direct supervision on $\\{\\Phi_{nt}^{\\mathrm{w}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{w}}\\}_{1}$ which is common in the literature. Please see Supp. Mat. for the details of the full training objectives.", + "bbox": [ + 76, + 224, + 467, + 270 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 282, + 209, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 76, + 306, + 277, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We assess the performance of SynCHMR primarily for global human motion estimation but also report the accuracy of estimated camera trajectories. Traditional video datasets in HMR literature are typically captured by static cameras, e.g. [13, 18, 20, 40, 63], hence not suitable for our purpose. Standard SLAM benchmarks such as [48, 51] do not meet our needs either as there is often no human moving in the scene. We consider the following datasets.", + "bbox": [ + 75, + 334, + 467, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3DPW [56] is an in-the-wild dataset captured with iPhones. The ground truth bodies are not in coherent world frames so we use it to supervise root relative poses and for evaluation. EgoBody [69] has ground-truth poses captured by multiple Kinects and egocentric-view sequences recorded by a head-mounted device, whose trajectories are further registered in the world space of Kinect array. We use it for training the SMPL denoiser in Sec. 3.3 and for evaluation (on both body and camera estimation). For HMR evaluation, unlike [30, 62] considering only the validation set, we additionally report results on its completely withheld test set.", + "bbox": [ + 75, + 455, + 467, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EMDB [25] is a new dataset providing SMPL poses from IMU sensors and global camera trajectories. We include it for training the SMPL denoiser to enrich the diversity and use the camera trajectories to evaluate the quality of SLAM.", + "bbox": [ + 75, + 621, + 467, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For HMR evaluation, we report common PA-MPJPE, which measures the quality of rootrelative poses. For datasets that have ground-truth poses in a world coordinate, we follow [62] and consider WA-MPJPE and FA-MPJPE. The former measures the error after aligning the entire trajectories of the prediction and ground truth with Procrustes Alignment [10], while the latter aligns only with the first frame. We also report acceleration errors. For SLAM, we consider absolute trajectory error (ATE) for camera trajectory evaluation as well as the threshold accuracy $(\\delta_{n})$ , the absolute relative error (REL), and the root mean squared error (RMSE) for scene depth evaluation [2].", + "bbox": [ + 75, + 685, + 467, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. In Human-aware Depth Calibration, we use the L-BFGS algorithm with learning rate 1 to", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2fd701643bc4a5d7d5b61166d002d5916d09324bb3126e627c16677311f60848.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Camera ModelHuman ModelPA ↓
DROID-SLAM [54]SLAHMR [62] w/ PHALP+55.9
DROID-SLAM [54]SLAHMR [62] w/ 4DHumans [9]57.4
Human-aware Metric SLAM (ours)4DHumans [9]52.9
Human-aware Metric SLAM (ours)Scene-aware SMPL Denoiser (ours)52.4
", + "bbox": [ + 500, + 88, + 890, + 157 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison results on 3DPW-Test. The row in gray is the full pipeline of SynCHMR. We abbreviate PA-MPJPE as PA, with the same below for FA-MPJPE (FA) and WA-MPJPE (WA).", + "bbox": [ + 498, + 167, + 890, + 209 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "optimize for a maximum of 30 iterations. As for the Scene-aware SMPL Denoiser, we train it on the union of 3DPW-Train, EgoBody-Train, and EMDB for 100k steps with an AdamW optimizer, a batch size of 16, and a learning rate of 1e-5. For camera-frame SMPL ground truths like in 3DPW, we only incorporate body shapes $\\beta$ and poses $\\theta$ in training. We train the denoising process by randomly sampling a temporal window size $T$ spanning 64 to 128 and inference with $T = 100$ . The scene-conditioned denoiser $\\mathcal{D}$ is parameterized with a 6-layer Transformer Decoder. For the scene encoder $\\mathcal{E}$ , we consider ViT and SPVCNN in Tab. 4 and report results for SPVCNN in Tabs. 1 and 2. Before inputting the world-frame noisy SMPL parameters to the denoiser, we first interpolate $\\Phi_{nt,0}^{\\mathrm{w}}$ and $\\theta_{nt,0}$ on SO(3), $\\beta_{nt,0}$ on $\\mathbb{R}^{10}$ , and $\\Gamma_{nt,0}^{\\mathrm{w}}$ on $\\mathbb{R}^3$ when there are missing observations.", + "bbox": [ + 498, + 228, + 890, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison Results", + "text_level": 1, + "bbox": [ + 500, + 468, + 691, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluate the estimated local poses with PA-MPJPE on 3DPW, which is common in the literature. In Tab. 1, we show that placing the bodies from 4DHumans already leads to lower error than SLAHMR. Passing them through the denoiser further reduces the error. We note that PA-MPJPE only measures local pose accuracy not the quality of global trajectories. Since 3DPW does not support any world metrics, Tab. 1 only aims to show that SynCHMR produces reasonable local poses on a common dataset.", + "bbox": [ + 496, + 491, + 890, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we assess the quality of global motion estimation, which is essentially a more challenging task. Tab. 2 shows the results on EgoBody. Note that current optimization-based methods [30, 62] report the error of the validation set. For fairness and completeness, we report results on both validation and test sets and run state-of-the-art methods on the test set when the code is available. In Tab. 2, we see that the proposed SynCHMR has the overall lowest PA-MPJPE, FA-MPJPE, and WA-MPJPE (gray rows). Comparing it with the row above (4DHumans) confirms the benefit of our scene-conditioned denoiser. For a fair comparison, we also initialize the global optimization of SLAHMR with 4DHumans, which is more accurate than $\\mathrm{PHALP}^+$ in SLAHMR, but we do not observe improvement. Notably, despite the concurrent work PACE [30] has a tightly integrated SLAM and body fitting objective, it still uses native DROID-SLAM to initialize the camera parameters like SLAHMR does. This is arguably sub-optimal as the initial", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1221", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4057fb5c19ef539edc02255b51a2763e9f7728ceca9cb83a62eeb630a31af605.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SubsetCamera ModelHuman ModelPA-MPJPE (mm) ↓FA-MPJPE (mm) ↓WA-MPJPE (mm) ↓Acc Error (mm/frame2) ↓Runtime/100 ims
Val-GLAMR [65]114.3416.1239.0173.54 min
DROID-SLAM [54]PACE [30]66.5147.9101.06.71 min
DROID-SLAM [54]SLAHMR [62] w/ PHALP+79.1141.1101.225.840 min
DROID-SLAM [54]SLAHMR [62] w/ 4DHumans [9]79.3273.0144.779.440 min
Human-aware Metric SLAM (ours)4DHumans [9]73.0164.4106.7127.05 min
Human-aware Metric SLAM (ours)Scene-aware SMPL Denoiser (ours)57.7115.181.164.85 min
Test-GLAMR [65]112.8351.4216.3105.94 min
DROID-SLAM [54]SLAHMR [62] w/ PHALP+63.1163.999.431.740 min
DROID-SLAM [54]SLAHMR [62] w/ 4DHumans [9]69.3185.8113.045.740 min
Human-aware Metric SLAM (ours)4DHumans [9]75.4160.0108.1138.85 min
Human-aware Metric SLAM (ours)Scene-aware SMPL Denoiser (ours)61.3122.184.669.45 min
", + "bbox": [ + 81, + 88, + 890, + 218 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/94fcffe239d8adfb946a4f5da01632b9168fc5eab21d286bd2d82eead6ccdc9a.jpg", + "table_caption": [ + "Table 2. Comparison results with state-of-the-art approaches on EgoBody. The row in gray is the full pipeline of SynCHMR." + ], + "table_footnote": [], + "table_body": "
RGBDepthMaskEgoBodyEMDB\nATE ↓
ATE ↓δ1 ↑REL ↓RMSE ↓
✓XX80.90.08514.5901617.361400.3
✓XMask2Former [4]81.60.0638.5301009.127385.8
✓ZoeDepth+X35.00.5620.30815.360456.8
✓ZoeDepth+Mask2Former [4]28.60.5640.30710.852389.6
✓ZoeDepth+ + Cal.Mask2Former [4]26.40.7970.27410.452107.0
", + "bbox": [ + 81, + 250, + 465, + 329 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d5f2fa350721696515b3c30675db92123bad2cc54e7e16f147e518b4e7bc40c0.jpg", + "table_caption": [ + "Table 3. Ablation study for SLAM configurations in terms of optimized camera trajectories and scene depths. ZoeDepth+ denotes our video-adapted ZoeDepth [2]." + ], + "table_footnote": [], + "table_body": "
StageBackboneRGBXYZMaskPA ↓FA ↓WA ↓Acc Error ↓
Init.-XXX73.7120.893.1127.1
Pred.-XXX63.398.877.275.2
Pred.ViT [8]✓XX63.994.976.743.3
Pred.ViT [8]✓✓X64.597.377.745.6
Pred.ViT [8]✓X✓66.896.578.644.7
Pred.ViT [8]✓✓✓69.3100.982.046.4
Pred.SPVCNN [53]X✓X62.995.176.072.6
Pred.SPVCNN [53]✓✓X61.093.474.367.7
Pred.SPVCNN [53]X✓✓62.093.975.369.9
Pred.SPVCNN [53]✓✓✓61.391.973.664.8
", + "bbox": [ + 81, + 387, + 465, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study for different scene encoders and features regarding world-frame HMR. Init. and Pred. refer to before and after SMPL denoising, respectively.", + "bbox": [ + 76, + 525, + 468, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ization is not aware of body information, which can lead to errors that cannot be corrected in the global optimization stage. Consequently, it also has higher world-space errors. Optimization methods often employ a zero velocity term to smooth out human motion, which explains the lower acceleration error. However, we do not observe a big difference in jittery between our results. Please refer to Supp. Mat. for more details.", + "bbox": [ + 75, + 580, + 468, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 710, + 230, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We ablate the design choices in SynCHMR. In Tab. 3, we evaluate SLAM-optimized camera trajectories and scene depths with EgoBody and EMDB. We see that directly including un-calibrated monocular depths does not guarantee more accurate estimations ( $3^{\\mathrm{rd}}$ vs. $1^{\\mathrm{st}}$ and $4^{\\mathrm{th}}$ vs. $2^{\\mathrm{nd}}$ row). Precluding the dynamic foreground pixels with Mask2Former [4] generally improves performance. We empirically find that our depth calibration with human prior works the best when using it with foreground masking, which has the lowest error in both datasets. More SLAM evaluation and discussion can be found in Supp. Mat.", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 4, we verify the benefit of scene conditioning for the SMPL denoiser. We train it with EgoBody-train in different conditioning schemes and report the $T = 32$ results on EgoBody-val. First, placing the predicted bodies from 4DHuman in the global space directly with estimated camera extrinsics has the highest error (1 $^{\\text{st}}$ row). When conditioning on a constant zero tensor, the denoiser behaves like a motion prior and reduces the error (2 $^{\\text{nd}}$ row). To encode the appearance and geometry information of the scene, we consider ViT [8] or SPVCNN [53] as the encoder $\\mathcal{E}$ and try varied combinations of appearance features (RGB), geometry features (XYZ) and aggregated subject masks (Mask). When using ViT to encode the scene, adding XYZ features or masks does not reduce the error. In contrast, when using SPVCNN, adding RGB information or conditioning on masks does improve performance. Overall, SPVCNN yields lower errors than ViT and enabling all conditioning leads to the lowest world-space error measure.", + "bbox": [ + 496, + 253, + 893, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative Analysis and Discussion", + "text_level": 1, + "bbox": [ + 498, + 540, + 808, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the first two rows of Fig. 5, we visualize the results of 3DPW and EgoBody in a global space. Despite occlusions, our SynCHMR estimates human meshes reliably and places them in a dense scene point cloud, whereas the scenes in GLAMR [65] and SLAHMR [62] consist of only a simple ground plane. Applying scene constraints with such an overly simplified scene can result in erroneous estimation, e.g., incorrect human trajectories as shown in the top view of the $1^{\\text{st}}$ row, and the vertically shortened human bodies in the $4^{\\text{th}}$ row of (d). Note that since TRACE [52] is scene agnostic, the ground plane in (c) is only for visualization, not necessarily indicating scene penetration.", + "bbox": [ + 496, + 566, + 890, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also test on more in-the-wild DAVIS [42] videos containing human subjects. Since DAVIS provides no ground-truth human meshes nor camera trajectories, we show only the visual comparison. The $3^{\\mathrm{rd}}$ row shows that we can handle multi-person cases as well as SLAHMR, while GLAMR often fails when multiple humans and dynamic cameras both occur. In a challenging scenario where the subject is taking selfies (the $4^{\\mathrm{th}}$ row), both GLAMR and SLAHMR are confused by the foreground human dominating the frames and reconstruct an almost static global trajectory, failing to", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1222", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d1ff7543f2ddc4f0a56d27c65a6b48963f0baaf3a8c307f5d0ab13b2a563ff0.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison among world-frame HMR approaches. We show (b) GLAMR [65] and (c) TRACE [52] results with their pre-defined ground planes, (d) SLAHMR [62] outputs with its estimated ground plane, and (e) our SynCHMR outputs with dense scenes. In the first row, we also demonstrate top-view human trajectories within circles. See supplementary for video results." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 893, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "disentangle the camera and the human motions due to the dynamic ambiguity. TRACE fails to produce results due to severe frame truncation. In contrast, SynCHMR still successfully provides reasonable trajectories.", + "bbox": [ + 75, + 570, + 468, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitation Discussion", + "text_level": 1, + "bbox": [ + 76, + 642, + 284, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As SynCHMR focuses on disentangling camera and human movements, we follow SLAHMR to approximate the focal length as $\\frac{W + H}{2}$ . When the subject has a shape that the body model cannot explain well, e.g., children or obese people, calibrating depth with the estimated bodies is less ideal. As we develop and validate SynCHMR on real videos, its accuracy on composed or generated videos remains an open question. Finally, since SynCHMR handles dynamic scenes with moving subjects, it does not require an a priori scanned static scene. This opens up new challenges, such as incorporating dynamic point clouds as scene constraints.", + "bbox": [ + 75, + 667, + 468, + 833 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 845, + 194, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present SynCHMR, a method that reconstructs camera trajectories, human bodies, and dense scenes from in-the", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "wild videos all in one global coordinate. SynCHMR has two core innovations. First, it leverages monocular depth estimation and uses the dimension and location of human meshes to calibrate the range of depth. This allows SLAM to better resolve the inherent scale ambiguity problem as shown in the experiment. Second, we train a data-driven motion denoiser and condition it with the scene in the same global coordinate, which is the first such scene-conditioned motion prior. Combining the two, the full SynCHMR pipeline uses human bodies to improve SLAM, and the better estimated scene and camera trajectory, in turn, provide better constraints for feed-forward human motion denoising. It achieves SOTA results on common benchmarks compared with existing optimization-based approaches.", + "bbox": [ + 496, + 569, + 890, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 799, + 650, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We appreciate constructive comments from Duygu Ceylan. This project was partially supported by the NIH under contracts R01GM134020 and P41GM103712, and by the NSF under contracts DBI-1949629, DBI-2238093, IIS-2007595, IIS-2211597, and MCB-2205148.", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1223", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iman Abaspur Kazerouni, Luke Fitzgerald, Gerard Dooly, and Daniel Toal. A survey of state-of-the-art on visual SLAM. Expert Systems with Applications, 205:117734, 2022. 1, 3", + "[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. ZoeDepth: Zero-shot transfer by combining relative and metric depth. arXiv, 2023. 2, 4, 6, 7, 1", + "[3] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European Conference on Computer Vision (ECCV), 2016. 3", + "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Computer Vision and Pattern Recognition (CVPR), pages 1290–1299, 2022. 4, 7, 2", + "[5] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consistent 3d human pose and shape from a video. In Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[6] Yudi Dai, Yitai Lin, Chenglu Wen, Siqi Shen, Lan Xu, Jingyi Yu, Yuexin Ma, and Cheng Wang. HSC4D: Human-centered 4D scene capture in large-scale indoor-outdoor space using wearable imus and LiDAR. In Computer Vision and Pattern Recognition (CVPR), pages 6792-6802, 2022. 3", + "[7] Yudi Dai, Yitai Lin, Xiping Lin, Chenglu Wen, Lan Xu, Hongwei Yi, Siqi Shen, Yuexin Ma, and Cheng Wang. Sloper4d: A scene-aware dataset for global 4d human pose estimation in urban environments. In Computer Vision and Pattern Recognition (CVPR), pages 682-692, 2023. 3", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 7", + "[9] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 5, 6, 7, 1", + "[10] John C Gower. Generalized procrustes analysis. Psychometrika, 40(1):33-51, 1975. 6", + "[11] Riza Alp Guler and Iasonas Kokkinos. HoloPose: Holistic 3D human reconstruction in-the-wild. In Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[12] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human positioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In Computer Vision and Pattern Recognition (CVPR). IEEE, 2021. 3", + "[13] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambigu" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ities with 3D scene constraints. In International Conference on Computer Vision (ICCV), pages 2282-2292, 2019. 3, 6, 1", + "[14] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. Advances in Neural Information Processing Systems, 35:4244-4256, 2022. 2", + "[15] Dorian F Henning, Tristan Laidlow, and Stefan Leutenegger. BodySLAM: joint camera localisation, mapping, and human motion tracking. In European Conference on Computer Vision (ECCV), pages 656-673. Springer, 2022. 2, 3, 1", + "[16] Dorian F Henning, Christopher Choi, Simon Schaefer, and Stefan Leutenegger. BodySLAM++: Fast and tightly-coupled visual-inertial camera and human motion tracking. In International Conference on Intelligent Robots and Systems (IROS), pages 3781-3788. IEEE, 2023. 3", + "[17] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In Computer Vision and Pattern Recognition (CVPR), pages 13274-13285, 2022. 3, 1", + "[18] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6M: Large scale datasets and predictive methods for 3D human sensing in natural environments. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 36(7):1325-1339, 2014. 6", + "[19] Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[20] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In International Conference on Computer Vision (ICCV), 2015. 6", + "[21] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3D deformation model for tracking faces, hands, and bodies. In Computer Vision and Pattern Recognition (CVPR), pages 8320-8329, 2018. 3", + "[22] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human pose fitting towards in-the-wild 3D human pose estimation. In International Conference on 3D Vision (3DV), pages 42-52, 2021. 3", + "[23] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018. 3", + "[24] Angjoo Kanazawa, Jason Y. Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. In Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[25] Manuel Kaufmann, Jie Song, Chen Guo, Kaiyue Shen, Tianjian Jiang, Chengcheng Tang, Juan Jose Zárate, and Otmar Hilliges. EMDB: The Electromagnetic Database of Global 3D Human Pose and Shape in the Wild. In International Conference on Computer Vision (ICCV), 2023. 3, 6", + "[26] Imry Kissos, Lior Fritz, Matan Goldman, Omer Meir, Edward Oks, and Mark Kliger. Beyond weak perspective for" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1224", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "monocular 3d human pose estimation. In European Conference on Computer Vision Workshops (ECCVw), pages 541-554, Cham, 2020. Springer International Publishing. 3", + "[27] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[28] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In International Conference on Computer Vision (ICCV). IEEE, 2021. 3", + "[29] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J. Black. SPEC: Seeing people in the wild with an estimated camera. In International Conference on Computer Vision (ICCV), pages 11035-11045, 2021. 3", + "[30] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J. Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. PACE: Human and motion estimation from in-the-wild videos. In International Conference on 3D Vision (3DV), 2024. 2, 3, 5, 6, 7, 1", + "[31] Nikos Kolotouros, Georgios Pavlakos, Michael J. Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In International Conference on Computer Vision (ICCV), 2019. 3", + "[32] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), pages 3383-3393, 2021. 3", + "[33] Jiefeng Li, Siyuan Bian, Chao Xu, Gang Liu, Gang Yu, and Cewu Lu. D&D: Learning human dynamics from dynamic camera. In European Conference on Computer Vision (ECCV), 2022. 3, 1, 2", + "[34] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In European Conference on Computer Vision (ECCV), pages 590-606, 2022. 3", + "[35] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In Computer Vision and Pattern Recognition (CVPR), pages 21159-21168, 2023. 3", + "[36] Miao Liu, Dexin Yang, Yan Zhang, Zhaopeng Cui, James M Rehg, and Siyu Tang. 4D human body capture from egocentric video via 3D scene grounding. In International Conference on 3D Vision (3DV), pages 930-939. IEEE, 2021. 3, 5, 1, 2", + "[37] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. Transactions on Graphics (TOG), 34 (6):248:1-248:16, 2015. 3, 4", + "[38] Diogo Luvizon, Marc Habermann, Vladislav Golyanik, Adam Kortylewski, and Christian Theobalt. Scene-Aware 3D Multi-Human Motion Capture from a Single Camera. Computer Graphics Forum (CGF), 42(2):371-383, 2023. 3", + "[39] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of" + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "motion capture as surface shapes. In International Conference on Computer Vision (ICCV), pages 5441-5450, 2019. 3", + "[40] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3D pose estimation from monocular RGB. In International Conference on 3D Vision (3DV), 2018. 6", + "[41] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[42] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Computer Vision and Pattern Recognition (CVPR), 2016. 7, 2", + "[43] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(3):1623-1637, 2020. 2", + "[44] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. HuMoR: 3D human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), pages 11468-11479, 2021. 2, 3", + "[45] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. FrankMocap: Fast monocular 3d hand and body motion capture by regression and integration. In International Conference on Computer Vision Workshops (ICCVw), 2021. 3", + "[46] Nitin Saini, Chun-Hao P Huang, Michael J Black, and Aamir Ahmad. SmartMocap: Joint estimation of human and camera motion using uncalibrated rgb cameras. IEEE Robotics and Automation Letters, 2023. 2, 3, 1", + "[47] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Computer Vision and Pattern Recognition (CVPR), 2016. 3, 1, 2", + "[48] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Computer Vision and Pattern Recognition (CVPR), pages 134-144, 2019. 6", + "[49] Zehong Shen, Zhi Cen, Sida Peng, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Learning human mesh recovery in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17038-17047, 2023. 3", + "[50] Jie Song, Xu Chen, and Otmar Hilliges. Human body model fitting by learned gradient descent. In European Conference on Computer Vision (ECCV), 2020. 3", + "[51] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A benchmark for the evaluation of rgb-d slam systems. In International Conference on Intelligent Robots and Systems (IROS), 2012. 6, 1, 3", + "[52] Yu Sun, Qian Bao, Wu Liu, Tao Mei, and Michael J. Black. TRACE: 5D temporal regression of avatars with dynamic cameras in 3d environments. In Computer Vision and Pat" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1225", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tern Recognition (CVPR), pages 8856-8866, 2023. 7, 8, 1, 2", + "[53] Haotian* Tang, Zhijian* Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European Conference on Computer Vision, 2020. 7", + "[54] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2, 3, 5, 6, 7, 1", + "[55] Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3d human mesh from monocular images: A survey. arXiv preprint arXiv:2203.01923, 2022. 1", + "[56] Timo von Marcard, Roberto Henschel, Michael J. Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3D human pose in the wild using IMUs and a moving camera. In European Conference on Computer Vision (ECCV), pages 614-631, 2018. 6", + "[57] Wenjia Wang, Yongtao Ge, Haiyi Mei, Zhongang Cai, Qingping Sun, Yanjun Wang, Chunhua Shen, Lei Yang, and Taku Komura. Zolly: Zoom focal length correctly for perspective-distorted human mesh reconstruction. In International Conference on Computer Vision (ICCV), 2023. 3", + "[58] Donglai Xiang, Hanbyul Joo, and Yaser Sheikh. Monocular total capture: Posing face, body, and hands in the wild. In Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[59] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In Computer Vision and Pattern Recognition (CVPR), pages 6183-6192, 2020. 3", + "[60] Yuanlu Xu, Song-Chun Zhu, and Tony Tung. DenseRaC: Joint 3d pose and shape estimation by dense render-and compare. In International Conference on Computer Vision (ICCV), 2019. 3", + "[61] Ming Yan, Xin Wang, Yudi Dai, Siqi Shen, Chenglu Wen, Lan Xu, Yuexin Ma, and Cheng Wang. Cimi4d: A large multimodal climbing motion dataset under human-scene interactions. In Computer Vision and Pattern Recognition (CVPR), pages 12977-12988, 2023. 3", + "[62] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In Computer Vision and Pattern Recognition (CVPR), pages 21222-21232, 2023. 2, 3, 5, 6, 7, 8, 1", + "[63] Jae Shin Yoon, Zhixuan Yu, Jaesik Park, and Hyun Soo Park. HUMBI: A large multiview dataset of human body expressions and benchmark challenge. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(1):623-640, 2021. 6", + "[64] Ri Yu, Hwangpil Park, and Jehee Lee. Human dynamics from monocular video with dynamic camera movements. Transactions on Graphics (TOG), 40(6), 2021. 3, 5, 1, 2", + "[65] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. GLAMR: Global occlusion-aware human mesh recovery with dynamic cameras. In Computer Vision and Pattern Recognition (CVPR), pages 11028-11039, 2022. 2, 3, 7, 8, 1" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Andrei Zanfir, Eduard Gabriel Bazavan, Hongyi Xu, Bill Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Weakly supervised 3d human pose and shape reconstruction with normalizing flows. In European Conference on Computer Vision (ECCV), 2020. 3", + "[67] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In International Conference on Computer Vision (ICCV), 2021. 3", + "[68] Siwei Zhang, Yan Zhang, Federica Bogo, Marc Pollefeys, and Siyu Tang. Learning motion priors for 4D human body capture in 3D scenes. In International Conference on Computer Vision (ICCV), 2021. 2, 3", + "[69] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. EgoBody: Human body shape and motion of interacting people from head-mounted devices. In European Conference on Computer Vision (ECCV), 2022. 3, 6", + "[70] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snively, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision (ECCV), pages 20-37. Springer, 2022. 3" + ], + "bbox": [ + 501, + 92, + 890, + 415 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1226", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/908ee619-1331-452b-9ccb-4e5b4d116ac0_model.json b/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/908ee619-1331-452b-9ccb-4e5b4d116ac0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..314d8a0b8653731dade66d7d5fa8813b04e4a9a0 --- /dev/null +++ b/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/908ee619-1331-452b-9ccb-4e5b4d116ac0_model.json @@ -0,0 +1,2477 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.131, + 0.861, + 0.154 + ], + "angle": 0, + "content": "Synergistic Global-space Camera and Human Reconstruction from Videos" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.18, + 0.882, + 0.217 + ], + "angle": 0, + "content": "Yizhou Zhao\\(^{1}\\); Tuanfeng Yang Wang\\(^{2}\\), Bhiksha Raj\\(^{1}\\), Min Xu\\(^{1}\\), Jimei Yang\\(^{2}\\), Chun-Hao Paul Huang\\(^{2}\\) \n\\(^{1}\\)Carnegie Mellon University \n\\(^{2}\\)Adobe Research" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.219, + 0.799, + 0.234 + ], + "angle": 0, + "content": "{yizhouz,bhiksha,mxul}@cs.cmu.edu,{yangtwan,jimyang,chunhaoh}@adobe.com" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.251, + 0.891, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.511, + 0.893, + 0.54 + ], + "angle": 0, + "content": "Figure 1. Our SynCHMR recovers metric-scale camera trajectories (color pyramids), human meshes, and scene point clouds in a global coordinate from casual videos by joining forces of Human Mesh Recovery (HMR) and Simultaneous Localization and Mapping (SLAM)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.552, + 0.314, + 0.567 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.473, + 0.864 + ], + "angle": 0, + "content": "Remarkable strides have been made in reconstructing static scenes or human bodies from monocular videos. Yet, the two problems have largely been approached independently, without much synergy. Most visual SLAM methods can only reconstruct camera trajectories and scene structures up to scale, while most HMR methods reconstruct human meshes in metric scale but fall short in reasoning with cameras and scenes. This work introduces Synergistic Camera and Human Reconstruction (SynCHMR) to marry the best of both worlds. Specifically, we design Human-aware Metric SLAM to reconstruct metric-scale camera poses and scene point clouds using camera-frame HMR as a strong prior, addressing depth, scale, and dynamic ambiguities. Conditioning on the dense scene recovered, we further learn a Scene-aware SMPL Denoiser to enhance world-frame HMR by incorporating spatio-temporal coherency and dynamic scene constraints. Together, they lead to consistent reconstructions of camera trajectories, human meshes, and dense scene point clouds in a common world frame." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.552, + 0.631, + 0.567 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.803 + ], + "angle": 0, + "content": "Physically plausible 3D human motion reconstruction from monocular videos is a long-standing problem in computer vision and graphics and has many applications in character animation, VFX, video games, sports, and healthcare. It requires estimating 3D humans across video frames in a common coordinate even with a moving camera. While human mesh recovery (HMR) has made significant progress recently [55], most existing methods typically estimate 3D humans in the camera coordinate by one frame at a time and fail to disambiguate camera motion. It calls for methods to jointly reconstruct 3D human and camera motion in a consistent global coordinate system from monocular videos. In other words, taking a video captured by a moving camera as input, the method should recover both temporally and spatially coherent movements of human bodies and cameras." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Intuitively, if the accurate camera motion is given, one can transform the bodies from individual camera frames to a common world frame by multiplying the inverse of camera extrinsic matrices. In practice, with humans moving in the scene, estimating the camera motion of a video is still an open challenge in monocular SLAM [1]. It not only" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.428, + 0.9 + ], + "angle": 0, + "content": "*Part of this work was done when interned at Adobe Research." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1216" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.468, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.174, + 0.321, + 0.185 + ], + "angle": 0, + "content": "(a) Depth ambiguity" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.188, + 0.466, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.24, + 0.321, + 0.252 + ], + "angle": 0, + "content": "(b) Scale ambiguity" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.254, + 0.468, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.356, + 0.328, + 0.367 + ], + "angle": 0, + "content": "(c) Dynamic ambiguity" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.373, + 0.47, + 0.511 + ], + "angle": 0, + "content": "Figure 2. Illustration of three types of ambiguities in visual SLAM. We show SLAM reconstruction results from DROID-SLAM [54]. (a) Depth ambiguity occurs when there are only minor camera translations between different views. This can lead to geometric failures in reconstruction such as the folded back corridor in the side view. (b) Scale ambiguity is inherent in monocular SLAM systems and requires additional reference for disambiguation. (c) Dynamic ambiguity gets pronounced when moving foregrounds dominate frames. Over-reliance on foreground key points will result in incorrect camera trajectories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.469, + 0.627 + ], + "angle": 0, + "content": "falls short in capturing accurate depths on views with small camera translations but more crucially, only estimates scene structures and camera trajectories up to scale. The human motion also breaks the static key point assumption in the bundle adjustment. As a result, one needs additional reference to disambiguate the depth, the scale, and the dynamic as illustrated in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.748 + ], + "angle": 0, + "content": "To leverage SLAM results in HMR pipelines, current world-frame HMR methods often refine camera poses by integrating either partial camera parameters, such as a global scale of the translation [62], or full extrinsic matrices [15, 30, 46] in an optimization-based framework. However, their optimization-based nature leads to complex multi-stage schemes, making the overall pipeline unnecessarily slow and easy to break." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this work, we explore a fundamentally different way to marry the best of HMR and SLAM. A 2D object can first be lifted from the image plane to the camera frame and then transformed into a common 3D space. This two-step process coincides with the combination of camera-frame HMR, which brings imaged 2D humans to 3D camera frames, and SLAM, which estimates the camera-to-world transformation. Noticing these correspondences, we leverage camera-frame HMR as a strong prior to bridge from the image plane to the camera frame for disambiguating SLAM, and utilize" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "SLAM reconstructions to constrain the transformation of human meshes from individual camera frames to a common global space. The overall pipeline thus results in a better synergy of the two, which we dub Synergistic Camera and Human Reconstruction (SynCHMR)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.17, + 0.892, + 0.426 + ], + "angle": 0, + "content": "We design SynCHMR based on several insights. First, despite camera-frame HMR methods cannot reconstruct humans in a coherent global frame, the estimated body dimension and location still provide cues to disambiguate SLAM. Unlike SLAHMR [62] which applies SLAM out of the box and corrects the scale afterward, we endow the SLAM process with human meshes from camera-frame HMR to address ambiguities. To this end, we capitalize on estimated absolute depths to provide pseudo-RGB-D inputs for SLAM [54] and confine the bundle adjustment to static backgrounds. Since current depth estimation methods [2, 43] predicts either relative depth maps or depths with data biases, we propose to calibrate their outputs by aligning with estimated human bodies in the camera frame [9]. With these priors, SLAM knows the depth, scale, and dynamic information from HMR and consequently estimates less ambiguous scene structures and camera poses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.893, + 0.73 + ], + "angle": 0, + "content": "Next, we place human meshes in the common coordinate recovered by SLAM. The gap between human tracks transformed from camera frames and their real plausible world-frame motions stems from two sources of error: noise induced by camera-frame HMR and by SLAM. Motion prior models [14, 44, 68] can be used for denoising purposes as they contribute to the temporal coherence of worldframe human tracks. However, their exclusive focus on human modeling either leaves them agnostic to the underlying scenes [14] or assumes the scene is a simple ground plane [44, 68]. Our intuition is that when placing a human, static elements of the scene, such as the ground, and dynamic components like moving objects are both possible to be in contact with the human, thereby providing clues for placing the body coherently and compatibly with the scene. We hence introduce a Scene-aware SMPL Denoiser that learns to denoise the transformed human tracks by considering both temporal consistencies of moving humans and implicit constraints from dynamic scenes. This global awareness makes it more flexible for in-the-wild videos." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.733, + 0.84, + 0.748 + ], + "angle": 0, + "content": "Our contributions can be summed up as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.75, + 0.892, + 0.841 + ], + "angle": 0, + "content": "- We present a novel pipeline, SynCHMR, that takes a monocular video as input and reconstructs human motions, camera trajectory and dense scene point clouds all in one global coordinate, as shown in Fig. 1, whereas current world-frame HMR methods [30, 62, 65] can recover only an estimated or pre-defined ground plane." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.841, + 0.892, + 0.9 + ], + "angle": 0, + "content": "- We propose a novel Human-aware Metric SLAM process to robustly calibrate estimated depth with estimated human meshes, resulting in metric-scale camera pose estimation and metric-scale scene reconstruction." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.75, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1217" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "- We present Scene-aware SMPL Denoising that enforces spatiotemporal coherencies and applies dynamic scene constraints on world-frame human meshes. Notably, this is achieved without requiring extra annotations or heuristic designs to decide which part of a human should be interacting with the scene [49] and which region in the scene is most likely to be in contact with humans [38, 62]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.215, + 0.22, + 0.231 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.242, + 0.468, + 0.288 + ], + "angle": 0, + "content": "There is considerable prior arts of HMR. We briefly discuss how they adopt different camera models and refer the readers to [1] for a more comprehensive review." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.468, + 0.488 + ], + "angle": 0, + "content": "HMR from a single image. State-of-the-art (SOTA) methods use parametric body models [21, 37, 41, 59] and estimate the parameters either by fitting to detected image features [3, 41, 58] or by regressing directly from pixels with deep neural networks [9, 11, 19, 22, 23, 28, 31, 32, 35, 45, 50, 60, 66, 67]. These approaches assume weak perspective/orthographic projection or pre-define the focal length as a large constant for all images. Kissos et al. [26] show that replacing focal length with a constant closer to ground truth alleviate the body tilting problem. SPEC [29] and Zolly [57] estimate focal length to account for perspective distortion. CIUFF [34] takes into account the location of humans in images to regress better poses in the camera coordinates." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.468, + 0.579 + ], + "angle": 0, + "content": "Many of these camera-frame HMR methods assume zero camera rotation, which entangles body rotation and camera rotation. When applied on video data, they fail to reconstruct humans in a coherent global space since they operate in a per-frame manner and hence cannot reason about how the camera moves across frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.468, + 0.747 + ], + "angle": 0, + "content": "HMR from videos aims to regress a series of body parameters from a temporal sequence. It opens up new problems such as whether the reconstructed bodies are in a common global coordinate or not. Some temporal methods consider a static camera [38, 44, 68], which makes the camera space a natural choice of the common coordinate. The challenge of coherent global space emerges when the camera moves. Early methods [5, 24, 27] show promising results on videos of dynamic cameras. Despite the reconstructed human meshes look great when overlaid on images, they do not share a common coordinate in 3D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Recent HMR methods capitalize on human motion prior to constrain the global trajectories in the world space, which in turn implicitly disentangles human movement from camera movement. GLAMR [65] consider a data-driven prior models learned on large-scale MoCap database e.g. AMASS [39], while D&D [33] and Yu [64] consider physic-inspired prior. These world-frame HMR methods often struggle on noise in local poses caused by partial occlusions, which is very common in in-the-wild videos with close-up shots and crowd scenes. Kaufmann et al. [25] and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.287 + ], + "angle": 0, + "content": "BodySLAM++ [16] circumvent this problem by employing IMU sensors to provide more robust body estimates but require extra sensory devices. To fully disentangle human and camera motion, another line of work [15, 30, 36, 46] leverages state-of-the-art SLAM techniques, e.g. [47, 54, 70], to explicitly estimate camera motion from the input video and infer the body parameters in the world coordinate of SLAM. Closest to us is SLAHMR [62] which solves for a global scale to connect the pre-computed SLAM results and body trajectories. To carefully guide the optimization process, these methods tend to have complex, multi-stage optimization schemes, making the overall pipeline easy to break and unnecessarily slow." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Note that in stark contrast to the methods above, which either assume or estimate a simple ground plane as scene representation, SynCHMR reconstructs dense scenes from in-the-wild videos without pre-scanning with extra devices a priori like in [6, 7, 12, 13, 17, 61, 69]. We provide detailed comparisons with these world-frame HMR in Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.391, + 0.59, + 0.405 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.66 + ], + "angle": 0, + "content": "Taking as input an RGB video \\(\\{\\mathbf{I}_t\\in \\mathbb{R}^{H\\times W\\times 3}\\}_{t = 1}^T\\) with \\(T\\) frames and \\(N\\) people in the scene, we aim to recover human meshes \\(\\{\\mathbf{V}_{nt}^{\\mathrm{w}}\\in \\mathbb{R}^{3\\times 6890}\\}_{n = 1,t = 1}^{N,T}\\), dynamic scene point clouds \\(\\{\\mathbf{P}_t^{\\mathrm{wm}}\\in \\mathbb{R}^{H\\times W\\times 3}\\}_{t = 1}^T\\), and corresponding camera poses \\(\\{\\mathbf{G}_t^{\\mathrm{m}}\\in \\mathrm{SE}(3)\\}_{t = 1}^T\\) in a common world coordinate system. The superscripts \\({}^{\\mathrm{W}},{}^{\\mathrm{C}}\\), and \\({}^{\\mathrm{m}}\\) denote the world frame, the camera frame, and the metric scale, respectively. To this aim, we propose a two-phase alternative conditioning pipeline as depicted in Fig. 3. In the first phase, we calibrate camera motion by injecting a camera-frame human prior to SLAM. This resolves depth, scale, and dynamic ambiguities, yielding metric-scale camera poses and dynamic point clouds. Subsequently, in the second phase, we transform the camera-frame human tracks into the world frame and utilize the dynamic point clouds obtained in the first phase for conditional denoising." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.669, + 0.64, + 0.683 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.692, + 0.602, + 0.705 + ], + "angle": 0, + "content": "3.1.1 SLAM" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Given a monocular RGB video \\(\\{\\mathbf{I}_t\\}_{t = 1}^T\\) DROID-SLAM [54] solves a dense bundle adjustment for a set of camera poses \\(\\{\\mathbf{G}_t\\in \\mathrm{SE}(3)\\}_{t = 1}^T\\) and inverse depths \\(\\{\\mathbf{d}_t\\in \\mathbb{R}_+^{H\\times W}\\}_{t = 1}^T\\) . To update these estimations, it first computes a dense correspondence field \\(\\mathbf{p}_{ij}\\in \\mathbb{R}^{H\\times W\\times 2}\\) based on reprojection for each pair of frames \\((i,j)\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.815, + 0.892, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {i j} = \\Pi \\left(\\mathbf {G} _ {i j} \\circ \\Pi^ {- 1} \\left(\\mathbf {p} _ {i}, \\frac {1}{\\mathbf {d} _ {j}}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.855, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_i\\in \\mathbb{R}^{H\\times W\\times 2}\\) is a grid of pixel coordinates in frame \\(i\\) \\(\\mathbf{G}_{ij} = \\mathbf{G}_j\\circ \\mathbf{G}_i^{-1}\\) is the relative pose, and \\(\\Pi\\) and \\(\\Pi^{-1}\\) are the camera projection and inverse projection functions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1218" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.898, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.315, + 0.895, + 0.388 + ], + "angle": 0, + "content": "Figure 3. The architecture of SynCHMR. Our pipeline comprises two phases. The first phase, Human-aware Metric SLAM (Sec. 3.2), infers metric-scale camera poses and metric-scale point clouds by exploiting the camera-frame human prior. The second phase, Scene-aware SMPL Denoising (Sec. 3.3), involves the conditional denoising of world-frame noisy SMPL parameters. These parameters, initialized by transforming from the camera frame, get refined through conditioning on the dynamic point clouds obtained in the first phase. The whole pipeline thus reconstructs humans, scene point clouds, and cameras harmoniously in a common world frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.471, + 0.448 + ], + "angle": 0, + "content": "Then with a learned neural network, the system predicts a revision flow field \\(\\mathbf{r}_{ij} \\in \\mathbb{R}^{H \\times W \\times 2}\\) and associated confidence map \\(\\mathbf{w}_{ij} \\in \\mathbb{R}_+^{H \\times W \\times 2}\\) to construct the cost function" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.462, + 0.47, + 0.505 + ], + "angle": 0, + "content": "\\[\nE _ {\\Sigma} = \\sum_ {(i, j)} \\left\\| \\mathbf {p} _ {i j} ^ {*} - \\Pi \\left(\\mathbf {G} _ {i j} ^ {\\prime} \\circ \\Pi^ {- 1} \\left(\\mathbf {p} _ {i}, \\frac {1}{\\mathbf {d} _ {i} ^ {\\prime}}\\right)\\right) \\right\\| _ {\\Sigma_ {i j}} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.471, + 0.618 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_{ij}^{*} = \\mathbf{r}_{ij} + \\mathbf{p}_{ij}\\) is the corrected correspondence, \\(\\| \\cdot \\|_{\\Sigma}\\) is the Mahalanobis distance which weighs the error terms with \\(\\Sigma_{ij} = \\mathrm{diag}\\mathbf{w}_{ij}\\), and \\(\\mathbf{G}'\\) and \\(\\mathbf{d}'\\) are updated poses and inverse depths. Upon this objective, DROID-SLAM considers an additional term that penalizes the squared distance between the measured and predicted depth if the input is with an extra sensor depth channel \\(\\{\\mathbf{D}_t\\}_{t=1}^T\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.627, + 0.173, + 0.643 + ], + "angle": 0, + "content": "3.1.2 HMR" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.47, + 0.819 + ], + "angle": 0, + "content": "We employ 4DHumans [9] for reconstructing camera-frame human meshes from an in-the-wild video. Specifically, it performs per-frame human mesh recovery with an end-to-end transformer architecture and associates them to form human tracks. Each tracked human \\( n \\) in frame \\( t \\) is represented by SMPL [37] parameters as \\( \\{\\Phi_{nt},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}\\} \\), including global orientation \\( \\Phi_{nt}\\in \\mathbb{R}^{3\\times 3} \\), body pose \\( \\theta_{nt}\\in \\mathbb{R}^{22\\times 3\\times 3} \\), shape \\( \\beta_{nt}\\in \\mathbb{R}^{10} \\), and root translation \\( \\Gamma_{nt}\\in \\mathbb{R}^3 \\). Then the parametric SMPL model can use these parameters to recover a human mesh with vertices \\( \\mathbf{V}_{nt}\\in \\mathbb{R}^{3\\times 6890} \\) in metric scale: \\( \\mathbf{V}_{nt} = \\mathrm{SMPL}(\\Phi_{nt},\\theta_{nt},\\beta_{nt}) + \\mathbf{\\Gamma}_{nt} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.824, + 0.337, + 0.839 + ], + "angle": 0, + "content": "3.2. Human-aware Metric SLAM" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.231, + 0.864 + ], + "angle": 0, + "content": "3.2.1 Preprocessing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.87, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To start off, we estimate per-frame depth maps \\(\\{\\mathbf{D}_t\\}\\) with an off-the-shelf depth estimator, ZoeDepth [2] and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.895, + 0.583 + ], + "angle": 0, + "content": "predict per-frame human instance segmentation masks \\(\\{\\mathbf{M}_{nt}\\}\\) with an image instance segmentation network, Mask2Former [4]. We adapt ZoeDepth for video-consistent depth estimation by choosing a per-video metric head from the majority vote of per-frame routers, for which we dub ZoeDepth\\(^+\\). While ZoeDepth claims to estimate metric depths, we observe a domain gap when inference on new datasets. Consequently, we only treat its output as up-to-affine depths that need to be further aligned with the metric scale. To aid our optimization with human awareness, we use camera-frame human meshes \\(\\{\\mathbf{V}_{nt}^{\\mathrm{c}}\\}\\) recovered by 4DHumans [9] to introduce a metric prior." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.592, + 0.816, + 0.608 + ], + "angle": 0, + "content": "3.2.2 Calibrating Depth with Human Prior" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.616, + 0.894, + 0.798 + ], + "angle": 0, + "content": "We calibrate the per-frame depths with human meshes in Human-aware Depth Calibration. This involves optimizing two parameters, a world scale \\(s\\) and a world offset \\(o\\), shared across all frames. During optimization, we linearly transform \\(\\mathbf{D}_t\\) to \\(\\mathbf{D}_t' = s\\mathbf{D}_t + o\\) and unproject these depth maps to camera-frame point clouds \\(\\{\\mathbf{P}_t^c\\}\\) with \\(\\mathbf{P}_t^c = \\Pi^{-1}(\\mathbf{p}_t, \\mathbf{D}_t')\\). Our intuition is to align the human point cloud \\(\\mathbf{P}_{nt}^c = \\mathbf{M}_{nt} \\odot \\mathbf{P}_t^c\\) with the camera-frame human mesh vertices \\(\\mathbf{V}_{nt}^c\\) in terms of absolute depth and size. To achieve pixel-wise alignment, we use a depth term to pull points on the human point cloud toward their corresponding human mesh vertices along the z-axis" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.806, + 0.892, + 0.845 + ], + "angle": 0, + "content": "\\[\nE _ {\\text {d e p t h}} = \\frac {\\sum_ {n , t} \\| \\mathbf {S} _ {n t} \\odot [ z \\left(\\mathbf {V} _ {n t} ^ {\\mathrm {c}}\\right) - z \\left(\\mathbf {P} _ {n t} ^ {\\mathrm {c}}\\right) ] \\| _ {2} ^ {2}}{\\sum_ {n , t} \\| \\mathbf {S} _ {n t} \\| _ {0}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathbf{S}_{nt} = \\rho (\\mathbf{V}_{nt}^{\\mathrm{c}})\\cap \\mathbf{M}_{nt}\\) is the intersection of the rasterized human mesh mask \\(\\rho (\\mathbf{V}_{nt}^{\\mathrm{c}})\\) and the instance segmentation mask \\(\\mathbf{M}_{nt}\\), \\(z(\\cdot)\\) is the rasterized depth, and \\(\\| \\cdot \\| _0\\) is the" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "1219" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.473, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.24, + 0.47, + 0.366 + ], + "angle": 0, + "content": "Figure 4. The architecture of Scene-aware SMPL Denoiser. World-frame noisy SMPL parameters \\(\\{\\Phi_{nt}^{\\mathrm{w}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{w}}\\}_{0}\\) are first projected by a linear layer and summed with temporal positional embeddings (TPE) to get initial latent humans \\(\\{\\mathbf{z}_{nt,0}^{\\mathrm{SMPL}}\\}\\). Per-frame point clouds are aggregated to \\(\\mathbf{x}_{\\mathrm{scene}}\\) and encoded with the point encoder \\(\\mathcal{E}\\). Then we query the encoded scene \\(\\mathcal{E}(\\mathbf{x}^{\\mathrm{scene}})\\) with latent humans \\(\\{\\mathbf{z}_{nt,0}^{\\mathrm{SMPL}}\\}\\) in the scene-conditioned denoiser \\(\\mathcal{D}\\) and feed the result \\(\\{\\mathbf{z}_{nt,1}^{\\mathrm{SMPL}}\\}\\) to prediction heads \\(\\{\\mathcal{P}_{\\Phi},\\mathcal{P}_{\\theta},\\mathcal{P}_{\\beta},\\mathcal{P}_{\\Gamma}\\}\\) to obtain denoised SMPL parameters \\(\\{\\Phi_{nt}^{\\mathrm{w}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{w}}\\}_{1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.378, + 0.468, + 0.392 + ], + "angle": 0, + "content": "0-norm indicating the number of non-zero pixels on a mask." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.393, + 0.469, + 0.438 + ], + "angle": 0, + "content": "As the recovered human meshes can be noisy in depth but still have a stable body dimension, we also adopt a size term to leverage the relative position of mesh vertices" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.447, + 0.47, + 0.482 + ], + "angle": 0, + "content": "\\[\nE _ {\\mathrm {d x}} = \\frac {\\sum_ {n , t} \\| \\Delta_ {x} \\left(\\mathbf {V} _ {n t} ^ {\\mathrm {c}} , \\mathbf {S} _ {n t}\\right) - \\Delta_ {x} \\left(\\mathbf {P} _ {n t} ^ {\\mathrm {c}} , \\mathbf {S} _ {n t}\\right) \\| _ {2} ^ {2}}{N T}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.489, + 0.334, + 0.505 + ], + "angle": 0, + "content": "We define \\(E_{\\mathrm{dy}}\\) similarly as \\(E_{\\mathrm{dx}}\\), where" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.513, + 0.469, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {*} (\\mathbf {X}, \\mathbf {Y}) = \\left(\\max _ {*} - \\min _ {*} \\right) \\left[ \\Pi^ {- 1} (\\mathbf {Y} \\odot \\Pi (\\mathbf {X}), z (\\mathbf {X})) \\right] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.546, + 0.469, + 0.574 + ], + "angle": 0, + "content": "and \\((\\max_{*} - \\min_{*})\\) denotes the difference between the maximum value and the minimum value on coordinate \\(^*\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.576, + 0.446, + 0.59 + ], + "angle": 0, + "content": "Then we have the calibrated depths with optimization" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.6, + 0.469, + 0.618 + ], + "angle": 0, + "content": "\\[\n\\left(s ^ {\\mathrm {m}}, o ^ {\\mathrm {m}}\\right) = \\operatorname {a r g m i n} _ {s, o} \\left(E _ {\\text {d e p t h}} + \\lambda E _ {\\text {s i z e}}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.62, + 0.469, + 0.638 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {t} ^ {\\mathrm {m}} = s ^ {\\mathrm {m}} \\mathbf {D} _ {t} + o ^ {\\mathrm {m}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.648, + 0.469, + 0.678 + ], + "angle": 0, + "content": "where \\( E_{\\mathrm{size}} = E_{\\mathrm{dx}} + E_{\\mathrm{dy}} \\), and \\( \\lambda \\) is a hyperparameter to balance two energy terms with a default value of 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.454, + 0.711 + ], + "angle": 0, + "content": "3.2.3 Disambiguating SLAM with Calibrated Depth" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.469, + 0.871 + ], + "angle": 0, + "content": "While DROID-SLAM [54] originally supports RGB-D input mode where the D channel stands for sensor depth, one cannot trivially access sensor depths from in-the-wild videos. Our insight is that an estimated absolute depth can be utilized as a depth prior, albeit noisy. So we combine the original RGB video and the calibrated depth as pseudoRGB-D inputs \\(\\{\\mathbf{I}_t,\\mathbf{D}_t^{\\mathrm{m}}\\}\\) to disambiguate depth and scale. Furthermore, we modify the cost function Eq. (2) to resolve the dynamic ambiguity by masking out dynamic foregrounds in confidence maps" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.88, + 0.469, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\Sigma_ {i j} ^ {\\prime} = \\operatorname {d i a g} \\mathbf {w} _ {i j} ^ {\\prime} = \\operatorname {d i a g} \\left(\\left(1 - \\left[ \\mathbf {M} _ {i}, \\mathbf {M} _ {j} \\right]\\right) \\odot \\mathbf {w} _ {i j}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.183 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}_i = \\bigcup_n\\mathbf{M}_{ni}\\) and \\(\\mathbf{M}_j = \\bigcup_n\\mathbf{M}_{nj}\\) are the union of all human instance masks on their corresponding frame, and \\([\\cdot ,\\cdot ]\\) is the concatenation operation. As a result, we obtain metric-scale camera poses \\(\\{\\mathbf{G}_t^{\\mathrm{m}}\\}\\) and metric-scale point clouds \\(\\{\\mathbf{P}_t^{\\mathrm{wm}}\\}\\) by disambiguating SLAM with calibrated metric depths" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.195, + 0.892, + 0.216 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathbf {G} _ {t} ^ {\\mathrm {m}}, \\mathbf {d} _ {t} ^ {\\mathrm {m}} \\right\\} = \\operatorname {a r g m i n} _ {\\left\\{\\mathbf {G} _ {i j} ^ {\\prime}, \\mathbf {d} _ {i j} ^ {\\prime} \\right\\}} E _ {\\Sigma^ {\\prime}}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.218, + 0.892, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {t} ^ {\\mathrm {w m}} = \\mathbf {G} _ {t} ^ {\\mathrm {m}} \\circ \\Pi^ {- 1} \\left(\\mathbf {p} _ {t}, \\frac {1}{\\mathbf {d} _ {t} ^ {\\mathrm {m}}}\\right)). \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.261, + 0.768, + 0.277 + ], + "angle": 0, + "content": "3.3. Scene-aware SMPL Denoising" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.285, + 0.846, + 0.3 + ], + "angle": 0, + "content": "3.3.1 Initializing Humans with Metric Cameras" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.892, + 0.385 + ], + "angle": 0, + "content": "To put humans properly in the scene recovered by SLAM, we initialize them by transforming estimated camera-frame SMPL parameters \\(\\{\\Phi_{nt}^{\\mathrm{c}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{c}}\\}\\) to the world frame with camera-to-world transforms \\(\\{\\mathbf{G}_t^{\\mathrm{m}} = [\\mathbf{R}_t|\\mathbf{t}_t^{\\mathrm{m}}]\\}\\). Given the pelvis as the center of global orientation \\(\\Phi\\), we have:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.413, + 0.892, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Phi} _ {n t} ^ {\\mathrm {w}} = \\mathbf {R} _ {t} \\boldsymbol {\\Phi} _ {n t} ^ {\\mathrm {c}}, \\quad \\boldsymbol {\\Gamma} _ {n t} ^ {\\mathrm {w}} = \\mathbf {R} _ {t} \\left(\\boldsymbol {\\Gamma} _ {n t} ^ {\\mathrm {c}} + \\mathbf {c}\\right) + \\mathbf {t} _ {t} ^ {\\mathrm {m}} - \\mathbf {c}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.55 + ], + "angle": 0, + "content": "where \\(\\mathbf{c} = \\mathbf{c}(\\beta_{nt})\\) is the pelvis location in the shape blend body mesh. Note that we do not need to introduce an extra camera scale as SLAHMR [62] since the camera poses have already been in the metric scale. The root-relative poses \\(\\theta_{nt}\\) and the shapes \\(\\beta_{nt}\\) stay unchanged as in the camera frame. We denote the initialized and the denoised parameters with a suffix 0 and 1 respectively, i.e. \\(\\{\\Phi_{nt}^{\\mathrm{w}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{w}}\\}_{0,1}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.57, + 0.86, + 0.585 + ], + "angle": 0, + "content": "3.3.2 Constraining Humans with Dynamic Scenes" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Different from existing works [30, 36, 62, 64] that incorporate energy terms in optimization to apply explicit scene constraints, we propose to learn implicit scene constraints with a Scene-aware SMPL denoiser shown in Fig. 4. The noisy initial SMPL parameters \\(\\{\\Phi_{nt}^{\\mathrm{w}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{w}}\\}_{0}\\) are first projected to a latent space, where it gets further updated by conditioning on implicit scene constraints" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.712, + 0.891, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {n t, 0} ^ {\\mathrm {S M P L}} = \\operatorname {F C} \\left(\\left[ \\boldsymbol {\\Phi} _ {n t, 0} ^ {\\mathrm {w}}, \\boldsymbol {\\theta} _ {n t, 0}, \\boldsymbol {\\beta} _ {n t, 0}, \\boldsymbol {\\Gamma} _ {n t, 0} ^ {\\mathrm {w}} \\right]\\right) + \\text {T P E}, \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.733, + 0.891, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathbf {z} _ {n t, 1} ^ {\\text {S M P L}} \\right\\} _ {t = 1} ^ {T} = \\mathcal {D} \\left(\\left\\{\\mathbf {z} _ {n t, 0} ^ {\\text {S M P L}} \\right\\} _ {t = 1} ^ {T}, \\mathcal {E} \\left(\\mathbf {x} ^ {\\text {s c e n e}}\\right) + \\text {T P E}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.903 + ], + "angle": 0, + "content": "where FC is a shared linear layer, TPE is shared temporal positional embeddings, \\(\\{\\mathbf{z}_{nt,*}^{\\mathrm{SMPL}}\\}_{t = 1}^{T}\\in \\mathbb{R}^{T\\times D}\\) is the \\(D\\) dimensional latent for human \\(n\\), and \\(\\mathbf{x}^{\\mathrm{scene}}\\in \\mathbb{R}^{L\\times C}\\) is the \\(C\\) -channel dynamic scene point clouds with a total number of points \\(L\\). \\(\\mathcal{E}\\) and \\(\\mathcal{D}\\) refer to the scene encoder and the scene-conditioned denoiser, respectively. We set \\(C = 7\\) which is the concatenation of point coordinates \\(\\{\\mathbf{P}_t^{\\mathrm{wm}}\\}\\), colors \\(\\{\\mathbf{I}_t\\}\\), and estimated human semantic segmentation masks \\(\\{\\mathbf{M}_t = \\bigcup_n\\mathbf{M}_{nt}\\}\\). Following [9], the updated latent" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1220" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "are decoded with different prediction head \\(\\mathcal{P}_{(\\cdot)}\\) to regress the residual for each SMPL parameter:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.132, + 0.468, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Phi} _ {n t, 1} ^ {\\mathrm {w}} = \\mathcal {P} _ {\\boldsymbol {\\Phi}} \\left(\\mathbf {z} _ {n t, 1} ^ {\\mathrm {S M P L}}\\right) \\boldsymbol {\\Phi} _ {n t, 0} ^ {\\mathrm {w}}, \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.153, + 0.468, + 0.171 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} _ {n t, 1} = \\mathcal {P} _ {\\boldsymbol {\\theta}} \\left(\\mathbf {z} _ {n t, 1} ^ {\\mathrm {S M P L}}\\right) \\boldsymbol {\\theta} _ {n t, 0}, \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.174, + 0.468, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\beta} _ {n t, 1} = \\mathcal {P} _ {\\boldsymbol {\\beta}} \\left(\\mathbf {z} _ {n t, 1} ^ {\\mathrm {S M P L}}\\right) + \\boldsymbol {\\beta} _ {n t, 0}, \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.195, + 0.468, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Gamma} _ {n t, 1} ^ {\\mathrm {w}} = \\mathcal {P} _ {\\boldsymbol {\\Gamma}} \\left(\\mathbf {z} _ {n t, 1} ^ {\\mathrm {S M P L}}\\right) + \\boldsymbol {\\Gamma} _ {n t, 0} ^ {\\mathrm {w}}. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.225, + 0.468, + 0.271 + ], + "angle": 0, + "content": "We apply direct supervision on \\(\\{\\Phi_{nt}^{\\mathrm{w}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{w}}\\}_{1}\\) which is common in the literature. Please see Supp. Mat. for the details of the full training objectives." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.284, + 0.21, + 0.301 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.308, + 0.278, + 0.325 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.468, + 0.456 + ], + "angle": 0, + "content": "Datasets. We assess the performance of SynCHMR primarily for global human motion estimation but also report the accuracy of estimated camera trajectories. Traditional video datasets in HMR literature are typically captured by static cameras, e.g. [13, 18, 20, 40, 63], hence not suitable for our purpose. Standard SLAM benchmarks such as [48, 51] do not meet our needs either as there is often no human moving in the scene. We consider the following datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.468, + 0.621 + ], + "angle": 0, + "content": "3DPW [56] is an in-the-wild dataset captured with iPhones. The ground truth bodies are not in coherent world frames so we use it to supervise root relative poses and for evaluation. EgoBody [69] has ground-truth poses captured by multiple Kinects and egocentric-view sequences recorded by a head-mounted device, whose trajectories are further registered in the world space of Kinect array. We use it for training the SMPL denoiser in Sec. 3.3 and for evaluation (on both body and camera estimation). For HMR evaluation, unlike [30, 62] considering only the validation set, we additionally report results on its completely withheld test set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.468, + 0.682 + ], + "angle": 0, + "content": "EMDB [25] is a new dataset providing SMPL poses from IMU sensors and global camera trajectories. We include it for training the SMPL denoiser to enrich the diversity and use the camera trajectories to evaluate the quality of SLAM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.468, + 0.867 + ], + "angle": 0, + "content": "Evaluation Metrics. For HMR evaluation, we report common PA-MPJPE, which measures the quality of rootrelative poses. For datasets that have ground-truth poses in a world coordinate, we follow [62] and consider WA-MPJPE and FA-MPJPE. The former measures the error after aligning the entire trajectories of the prediction and ground truth with Procrustes Alignment [10], while the latter aligns only with the first frame. We also report acceleration errors. For SLAM, we consider absolute trajectory error (ATE) for camera trajectory evaluation as well as the threshold accuracy \\((\\delta_{n})\\) , the absolute relative error (REL), and the root mean squared error (RMSE) for scene depth evaluation [2]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. In Human-aware Depth Calibration, we use the L-BFGS algorithm with learning rate 1 to" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.892, + 0.158 + ], + "angle": 0, + "content": "
Camera ModelHuman ModelPA ↓
DROID-SLAM [54]SLAHMR [62] w/ PHALP+55.9
DROID-SLAM [54]SLAHMR [62] w/ 4DHumans [9]57.4
Human-aware Metric SLAM (ours)4DHumans [9]52.9
Human-aware Metric SLAM (ours)Scene-aware SMPL Denoiser (ours)52.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.168, + 0.892, + 0.21 + ], + "angle": 0, + "content": "Table 1. Comparison results on 3DPW-Test. The row in gray is the full pipeline of SynCHMR. We abbreviate PA-MPJPE as PA, with the same below for FA-MPJPE (FA) and WA-MPJPE (WA)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.229, + 0.892, + 0.457 + ], + "angle": 0, + "content": "optimize for a maximum of 30 iterations. As for the Scene-aware SMPL Denoiser, we train it on the union of 3DPW-Train, EgoBody-Train, and EMDB for 100k steps with an AdamW optimizer, a batch size of 16, and a learning rate of 1e-5. For camera-frame SMPL ground truths like in 3DPW, we only incorporate body shapes \\(\\beta\\) and poses \\(\\theta\\) in training. We train the denoising process by randomly sampling a temporal window size \\(T\\) spanning 64 to 128 and inference with \\(T = 100\\). The scene-conditioned denoiser \\(\\mathcal{D}\\) is parameterized with a 6-layer Transformer Decoder. For the scene encoder \\(\\mathcal{E}\\), we consider ViT and SPVCNN in Tab. 4 and report results for SPVCNN in Tabs. 1 and 2. Before inputting the world-frame noisy SMPL parameters to the denoiser, we first interpolate \\(\\Phi_{nt,0}^{\\mathrm{w}}\\) and \\(\\theta_{nt,0}\\) on SO(3), \\(\\beta_{nt,0}\\) on \\(\\mathbb{R}^{10}\\), and \\(\\Gamma_{nt,0}^{\\mathrm{w}}\\) on \\(\\mathbb{R}^3\\) when there are missing observations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.469, + 0.692, + 0.484 + ], + "angle": 0, + "content": "4.2. Comparison Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.628 + ], + "angle": 0, + "content": "We first evaluate the estimated local poses with PA-MPJPE on 3DPW, which is common in the literature. In Tab. 1, we show that placing the bodies from 4DHumans already leads to lower error than SLAHMR. Passing them through the denoiser further reduces the error. We note that PA-MPJPE only measures local pose accuracy not the quality of global trajectories. Since 3DPW does not support any world metrics, Tab. 1 only aims to show that SynCHMR produces reasonable local poses on a common dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Next, we assess the quality of global motion estimation, which is essentially a more challenging task. Tab. 2 shows the results on EgoBody. Note that current optimization-based methods [30, 62] report the error of the validation set. For fairness and completeness, we report results on both validation and test sets and run state-of-the-art methods on the test set when the code is available. In Tab. 2, we see that the proposed SynCHMR has the overall lowest PA-MPJPE, FA-MPJPE, and WA-MPJPE (gray rows). Comparing it with the row above (4DHumans) confirms the benefit of our scene-conditioned denoiser. For a fair comparison, we also initialize the global optimization of SLAHMR with 4DHumans, which is more accurate than \\(\\mathrm{PHALP}^+\\) in SLAHMR, but we do not observe improvement. Notably, despite the concurrent work PACE [30] has a tightly integrated SLAM and body fitting objective, it still uses native DROID-SLAM to initialize the camera parameters like SLAHMR does. This is arguably sub-optimal as the initial" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1221" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.219 + ], + "angle": 0, + "content": "
SubsetCamera ModelHuman ModelPA-MPJPE (mm) ↓FA-MPJPE (mm) ↓WA-MPJPE (mm) ↓Acc Error (mm/frame2) ↓Runtime/100 ims
Val-GLAMR [65]114.3416.1239.0173.54 min
DROID-SLAM [54]PACE [30]66.5147.9101.06.71 min
DROID-SLAM [54]SLAHMR [62] w/ PHALP+79.1141.1101.225.840 min
DROID-SLAM [54]SLAHMR [62] w/ 4DHumans [9]79.3273.0144.779.440 min
Human-aware Metric SLAM (ours)4DHumans [9]73.0164.4106.7127.05 min
Human-aware Metric SLAM (ours)Scene-aware SMPL Denoiser (ours)57.7115.181.164.85 min
Test-GLAMR [65]112.8351.4216.3105.94 min
DROID-SLAM [54]SLAHMR [62] w/ PHALP+63.1163.999.431.740 min
DROID-SLAM [54]SLAHMR [62] w/ 4DHumans [9]69.3185.8113.045.740 min
Human-aware Metric SLAM (ours)4DHumans [9]75.4160.0108.1138.85 min
Human-aware Metric SLAM (ours)Scene-aware SMPL Denoiser (ours)61.3122.184.669.45 min
" + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.226, + 0.866, + 0.24 + ], + "angle": 0, + "content": "Table 2. Comparison results with state-of-the-art approaches on EgoBody. The row in gray is the full pipeline of SynCHMR." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.251, + 0.466, + 0.33 + ], + "angle": 0, + "content": "
RGBDepthMaskEgoBodyEMDB\nATE ↓
ATE ↓δ1 ↑REL ↓RMSE ↓
✓XX80.90.08514.5901617.361400.3
✓XMask2Former [4]81.60.0638.5301009.127385.8
✓ZoeDepth+X35.00.5620.30815.360456.8
✓ZoeDepth+Mask2Former [4]28.60.5640.30710.852389.6
✓ZoeDepth+ + Cal.Mask2Former [4]26.40.7970.27410.452107.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.337, + 0.47, + 0.379 + ], + "angle": 0, + "content": "Table 3. Ablation study for SLAM configurations in terms of optimized camera trajectories and scene depths. ZoeDepth+ denotes our video-adapted ZoeDepth [2]." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.388, + 0.466, + 0.519 + ], + "angle": 0, + "content": "
StageBackboneRGBXYZMaskPA ↓FA ↓WA ↓Acc Error ↓
Init.-XXX73.7120.893.1127.1
Pred.-XXX63.398.877.275.2
Pred.ViT [8]✓XX63.994.976.743.3
Pred.ViT [8]✓✓X64.597.377.745.6
Pred.ViT [8]✓X✓66.896.578.644.7
Pred.ViT [8]✓✓✓69.3100.982.046.4
Pred.SPVCNN [53]X✓X62.995.176.072.6
Pred.SPVCNN [53]✓✓X61.093.474.367.7
Pred.SPVCNN [53]X✓✓62.093.975.369.9
Pred.SPVCNN [53]✓✓✓61.391.973.664.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.526, + 0.47, + 0.569 + ], + "angle": 0, + "content": "Table 4. Ablation study for different scene encoders and features regarding world-frame HMR. Init. and Pred. refer to before and after SMPL denoising, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.47, + 0.702 + ], + "angle": 0, + "content": "ization is not aware of body information, which can lead to errors that cannot be corrected in the global optimization stage. Consequently, it also has higher world-space errors. Optimization methods often employ a zero velocity term to smooth out human motion, which explains the lower acceleration error. However, we do not observe a big difference in jittery between our results. Please refer to Supp. Mat. for more details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.231, + 0.728 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We ablate the design choices in SynCHMR. In Tab. 3, we evaluate SLAM-optimized camera trajectories and scene depths with EgoBody and EMDB. We see that directly including un-calibrated monocular depths does not guarantee more accurate estimations (\\(3^{\\mathrm{rd}}\\) vs. \\(1^{\\mathrm{st}}\\) and \\(4^{\\mathrm{th}}\\) vs. \\(2^{\\mathrm{nd}}\\) row). Precluding the dynamic foreground pixels with Mask2Former [4] generally improves performance. We empirically find that our depth calibration with human prior works the best when using it with foreground masking, which has the lowest error in both datasets. More SLAM evaluation and discussion can be found in Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.254, + 0.895, + 0.526 + ], + "angle": 0, + "content": "In Tab. 4, we verify the benefit of scene conditioning for the SMPL denoiser. We train it with EgoBody-train in different conditioning schemes and report the \\( T = 32 \\) results on EgoBody-val. First, placing the predicted bodies from 4DHuman in the global space directly with estimated camera extrinsics has the highest error (1\\(^{\\text{st}}\\) row). When conditioning on a constant zero tensor, the denoiser behaves like a motion prior and reduces the error (2\\(^{\\text{nd}}\\) row). To encode the appearance and geometry information of the scene, we consider ViT [8] or SPVCNN [53] as the encoder \\( \\mathcal{E} \\) and try varied combinations of appearance features (RGB), geometry features (XYZ) and aggregated subject masks (Mask). When using ViT to encode the scene, adding XYZ features or masks does not reduce the error. In contrast, when using SPVCNN, adding RGB information or conditioning on masks does improve performance. Overall, SPVCNN yields lower errors than ViT and enabling all conditioning leads to the lowest world-space error measure." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.541, + 0.81, + 0.558 + ], + "angle": 0, + "content": "4.4. Qualitative Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.748 + ], + "angle": 0, + "content": "In the first two rows of Fig. 5, we visualize the results of 3DPW and EgoBody in a global space. Despite occlusions, our SynCHMR estimates human meshes reliably and places them in a dense scene point cloud, whereas the scenes in GLAMR [65] and SLAHMR [62] consist of only a simple ground plane. Applying scene constraints with such an overly simplified scene can result in erroneous estimation, e.g., incorrect human trajectories as shown in the top view of the \\(1^{\\text{st}}\\) row, and the vertically shortened human bodies in the \\(4^{\\text{th}}\\) row of (d). Note that since TRACE [52] is scene agnostic, the ground plane in (c) is only for visualization, not necessarily indicating scene penetration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "We also test on more in-the-wild DAVIS [42] videos containing human subjects. Since DAVIS provides no ground-truth human meshes nor camera trajectories, we show only the visual comparison. The \\(3^{\\mathrm{rd}}\\) row shows that we can handle multi-person cases as well as SLAHMR, while GLAMR often fails when multiple humans and dynamic cameras both occur. In a challenging scenario where the subject is taking selfies (the \\(4^{\\mathrm{th}}\\) row), both GLAMR and SLAHMR are confused by the foreground human dominating the frames and reconstruct an almost static global trajectory, failing to" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1222" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.895, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.515, + 0.893, + 0.556 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison among world-frame HMR approaches. We show (b) GLAMR [65] and (c) TRACE [52] results with their pre-defined ground planes, (d) SLAHMR [62] outputs with its estimated ground plane, and (e) our SynCHMR outputs with dense scenes. In the first row, we also demonstrate top-view human trajectories within circles. See supplementary for video results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.571, + 0.47, + 0.632 + ], + "angle": 0, + "content": "disentangle the camera and the human motions due to the dynamic ambiguity. TRACE fails to produce results due to severe frame truncation. In contrast, SynCHMR still successfully provides reasonable trajectories." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.643, + 0.285, + 0.658 + ], + "angle": 0, + "content": "5. Limitation Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.47, + 0.834 + ], + "angle": 0, + "content": "As SynCHMR focuses on disentangling camera and human movements, we follow SLAHMR to approximate the focal length as \\(\\frac{W + H}{2}\\). When the subject has a shape that the body model cannot explain well, e.g., children or obese people, calibrating depth with the estimated bodies is less ideal. As we develop and validate SynCHMR on real videos, its accuracy on composed or generated videos remains an open question. Finally, since SynCHMR handles dynamic scenes with moving subjects, it does not require an a priori scanned static scene. This opens up new challenges, such as incorporating dynamic point clouds as scene constraints." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.195, + 0.861 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We present SynCHMR, a method that reconstructs camera trajectories, human bodies, and dense scenes from in-the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.892, + 0.783 + ], + "angle": 0, + "content": "wild videos all in one global coordinate. SynCHMR has two core innovations. First, it leverages monocular depth estimation and uses the dimension and location of human meshes to calibrate the range of depth. This allows SLAM to better resolve the inherent scale ambiguity problem as shown in the experiment. Second, we train a data-driven motion denoiser and condition it with the scene in the same global coordinate, which is the first such scene-conditioned motion prior. Combining the two, the full SynCHMR pipeline uses human bodies to improve SLAM, and the better estimated scene and camera trajectory, in turn, provide better constraints for feed-forward human motion denoising. It achieves SOTA results on common benchmarks compared with existing optimization-based approaches." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.651, + 0.816 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We appreciate constructive comments from Duygu Ceylan. This project was partially supported by the NIH under contracts R01GM134020 and P41GM103712, and by the NSF under contracts DBI-1949629, DBI-2238093, IIS-2007595, IIS-2211597, and MCB-2205148." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1223" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Iman Abaspur Kazerouni, Luke Fitzgerald, Gerard Dooly, and Daniel Toal. A survey of state-of-the-art on visual SLAM. Expert Systems with Applications, 205:117734, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.226 + ], + "angle": 0, + "content": "[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. ZoeDepth: Zero-shot transfer by combining relative and metric depth. arXiv, 2023. 2, 4, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European Conference on Computer Vision (ECCV), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Computer Vision and Pattern Recognition (CVPR), pages 1290–1299, 2022. 4, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.424 + ], + "angle": 0, + "content": "[5] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consistent 3d human pose and shape from a video. In Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[6] Yudi Dai, Yitai Lin, Chenglu Wen, Siqi Shen, Lan Xu, Jingyi Yu, Yuexin Ma, and Cheng Wang. HSC4D: Human-centered 4D scene capture in large-scale indoor-outdoor space using wearable imus and LiDAR. In Computer Vision and Pattern Recognition (CVPR), pages 6792-6802, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[7] Yudi Dai, Yitai Lin, Xiping Lin, Chenglu Wen, Lan Xu, Hongwei Yi, Siqi Shen, Yuexin Ma, and Cheng Wang. Sloper4d: A scene-aware dataset for global 4d human pose estimation in urban environments. In Computer Vision and Pattern Recognition (CVPR), pages 682-692, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.661 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.662, + 0.471, + 0.731 + ], + "angle": 0, + "content": "[9] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 5, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[10] John C Gower. Generalized procrustes analysis. Psychometrika, 40(1):33-51, 1975. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[11] Riza Alp Guler and Iasonas Kokkinos. HoloPose: Holistic 3D human reconstruction in-the-wild. In Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[12] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human positioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In Computer Vision and Pattern Recognition (CVPR). IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambigu" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "ities with 3D scene constraints. In International Conference on Computer Vision (ICCV), pages 2282-2292, 2019. 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[14] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. Advances in Neural Information Processing Systems, 35:4244-4256, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[15] Dorian F Henning, Tristan Laidlow, and Stefan Leutenegger. BodySLAM: joint camera localisation, mapping, and human motion tracking. In European Conference on Computer Vision (ECCV), pages 656-673. Springer, 2022. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[16] Dorian F Henning, Christopher Choi, Simon Schaefer, and Stefan Leutenegger. BodySLAM++: Fast and tightly-coupled visual-inertial camera and human motion tracking. In International Conference on Intelligent Robots and Systems (IROS), pages 3781-3788. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[17] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In Computer Vision and Pattern Recognition (CVPR), pages 13274-13285, 2022. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[18] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6M: Large scale datasets and predictive methods for 3D human sensing in natural environments. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 36(7):1325-1339, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[19] Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[20] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In International Conference on Computer Vision (ICCV), 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[21] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3D deformation model for tracking faces, hands, and bodies. In Computer Vision and Pattern Recognition (CVPR), pages 8320-8329, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[22] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human pose fitting towards in-the-wild 3D human pose estimation. In International Conference on 3D Vision (3DV), pages 42-52, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[23] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[24] Angjoo Kanazawa, Jason Y. Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. In Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[25] Manuel Kaufmann, Jie Song, Chen Guo, Kaiyue Shen, Tianjian Jiang, Chengcheng Tang, Juan Jose Zárate, and Otmar Hilliges. EMDB: The Electromagnetic Database of Global 3D Human Pose and Shape in the Wild. In International Conference on Computer Vision (ICCV), 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[26] Imry Kissos, Lior Fritz, Matan Goldman, Omer Meir, Edward Oks, and Mark Kliger. Beyond weak perspective for" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1224" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.47, + 0.134 + ], + "angle": 0, + "content": "monocular 3d human pose estimation. In European Conference on Computer Vision Workshops (ECCVw), pages 541-554, Cham, 2020. Springer International Publishing. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[27] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[28] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In International Conference on Computer Vision (ICCV). IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[29] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J. Black. SPEC: Seeing people in the wild with an estimated camera. In International Conference on Computer Vision (ICCV), pages 11035-11045, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[30] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J. Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. PACE: Human and motion estimation from in-the-wild videos. In International Conference on 3D Vision (3DV), 2024. 2, 3, 5, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[31] Nikos Kolotouros, Georgios Pavlakos, Michael J. Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In International Conference on Computer Vision (ICCV), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[32] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), pages 3383-3393, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[33] Jiefeng Li, Siyuan Bian, Chao Xu, Gang Liu, Gang Yu, and Cewu Lu. D&D: Learning human dynamics from dynamic camera. In European Conference on Computer Vision (ECCV), 2022. 3, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[34] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In European Conference on Computer Vision (ECCV), pages 590-606, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.692 + ], + "angle": 0, + "content": "[35] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In Computer Vision and Pattern Recognition (CVPR), pages 21159-21168, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[36] Miao Liu, Dexin Yang, Yan Zhang, Zhaopeng Cui, James M Rehg, and Siyu Tang. 4D human body capture from egocentric video via 3D scene grounding. In International Conference on 3D Vision (3DV), pages 930-939. IEEE, 2021. 3, 5, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[37] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. Transactions on Graphics (TOG), 34 (6):248:1-248:16, 2015. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.873 + ], + "angle": 0, + "content": "[38] Diogo Luvizon, Marc Habermann, Vladislav Golyanik, Adam Kortylewski, and Christian Theobalt. Scene-Aware 3D Multi-Human Motion Capture from a Single Camera. Computer Graphics Forum (CGF), 42(2):371-383, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[39] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "motion capture as surface shapes. In International Conference on Computer Vision (ICCV), pages 5441-5450, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[40] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3D pose estimation from monocular RGB. In International Conference on 3D Vision (3DV), 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[41] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[42] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Computer Vision and Pattern Recognition (CVPR), 2016. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.347, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[43] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(3):1623-1637, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[44] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. HuMoR: 3D human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), pages 11468-11479, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[45] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. FrankMocap: Fast monocular 3d hand and body motion capture by regression and integration. In International Conference on Computer Vision Workshops (ICCVw), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[46] Nitin Saini, Chun-Hao P Huang, Michael J Black, and Aamir Ahmad. SmartMocap: Joint estimation of human and camera motion using uncalibrated rgb cameras. IEEE Robotics and Automation Letters, 2023. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[47] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Computer Vision and Pattern Recognition (CVPR), 2016. 3, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[48] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Computer Vision and Pattern Recognition (CVPR), pages 134-144, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[49] Zehong Shen, Zhi Cen, Sida Peng, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Learning human mesh recovery in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17038-17047, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[50] Jie Song, Xu Chen, and Otmar Hilliges. Human body model fitting by learned gradient descent. In European Conference on Computer Vision (ECCV), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[51] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A benchmark for the evaluation of rgb-d slam systems. In International Conference on Intelligent Robots and Systems (IROS), 2012. 6, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[52] Yu Sun, Qian Bao, Wu Liu, Tao Mei, and Michael J. Black. TRACE: 5D temporal regression of avatars with dynamic cameras in 3d environments. In Computer Vision and Pat" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1225" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "tern Recognition (CVPR), pages 8856-8866, 2023. 7, 8, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[53] Haotian* Tang, Zhijian* Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European Conference on Computer Vision, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[54] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2, 3, 5, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[55] Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3d human mesh from monocular images: A survey. arXiv preprint arXiv:2203.01923, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.275, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[56] Timo von Marcard, Roberto Henschel, Michael J. Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3D human pose in the wild using IMUs and a moving camera. In European Conference on Computer Vision (ECCV), pages 614-631, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.413 + ], + "angle": 0, + "content": "[57] Wenjia Wang, Yongtao Ge, Haiyi Mei, Zhongang Cai, Qingping Sun, Yanjun Wang, Chunhua Shen, Lei Yang, and Taku Komura. Zolly: Zoom focal length correctly for perspective-distorted human mesh reconstruction. In International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[58] Donglai Xiang, Hanbyul Joo, and Yaser Sheikh. Monocular total capture: Posing face, body, and hands in the wild. In Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.457, + 0.469, + 0.525 + ], + "angle": 0, + "content": "[59] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In Computer Vision and Pattern Recognition (CVPR), pages 6183-6192, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[60] Yuanlu Xu, Song-Chun Zhu, and Tony Tung. DenseRaC: Joint 3d pose and shape estimation by dense render-and compare. In International Conference on Computer Vision (ICCV), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.65 + ], + "angle": 0, + "content": "[61] Ming Yan, Xin Wang, Yudi Dai, Siqi Shen, Chenglu Wen, Lan Xu, Yuexin Ma, and Cheng Wang. Cimi4d: A large multimodal climbing motion dataset under human-scene interactions. In Computer Vision and Pattern Recognition (CVPR), pages 12977-12988, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[62] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In Computer Vision and Pattern Recognition (CVPR), pages 21222-21232, 2023. 2, 3, 5, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[63] Jae Shin Yoon, Zhixuan Yu, Jaesik Park, and Hyun Soo Park. HUMBI: A large multiview dataset of human body expressions and benchmark challenge. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(1):623-640, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[64] Ri Yu, Hwangpil Park, and Jehee Lee. Human dynamics from monocular video with dynamic camera movements. Transactions on Graphics (TOG), 40(6), 2021. 3, 5, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[65] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. GLAMR: Global occlusion-aware human mesh recovery with dynamic cameras. In Computer Vision and Pattern Recognition (CVPR), pages 11028-11039, 2022. 2, 3, 7, 8, 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[66] Andrei Zanfir, Eduard Gabriel Bazavan, Hongyi Xu, Bill Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Weakly supervised 3d human pose and shape reconstruction with normalizing flows. In European Conference on Computer Vision (ECCV), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[67] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[68] Siwei Zhang, Yan Zhang, Federica Bogo, Marc Pollefeys, and Siyu Tang. Learning motion priors for 4D human body capture in 3D scenes. In International Conference on Computer Vision (ICCV), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.29, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[69] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. EgoBody: Human body shape and motion of interacting people from head-mounted devices. In European Conference on Computer Vision (ECCV), 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.36, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[70] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snively, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision (ECCV), pages 20-37. Springer, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1226" + } + ] +] \ No newline at end of file diff --git a/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/908ee619-1331-452b-9ccb-4e5b4d116ac0_origin.pdf b/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/908ee619-1331-452b-9ccb-4e5b4d116ac0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c8ae8dd62ec66a3d9159dadccc11c0ebdeb04d74 --- /dev/null +++ b/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/908ee619-1331-452b-9ccb-4e5b4d116ac0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ed2fbbf7894819ddf26a956b2eac9a258b9851eb00bb54d3edb4af150423405 +size 2766550 diff --git a/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/full.md b/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..69be6e44fdec145390bf55b5187b5453b3afecc3 --- /dev/null +++ b/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/full.md @@ -0,0 +1,362 @@ +# Synergistic Global-space Camera and Human Reconstruction from Videos + +Yizhou Zhao $^{1}$ ; Tuanfeng Yang Wang $^{2}$ , Bhiksha Raj $^{1}$ , Min Xu $^{1}$ , Jimei Yang $^{2}$ , Chun-Hao Paul Huang $^{2}$ $^{1}$ Carnegie Mellon University + $^{2}$ Adobe Research + +{yizhouz,bhiksha,mxul}@cs.cmu.edu,{yangtwan,jimyang,chunhaoh}@adobe.com + +![](images/7af274f60e1e4a33d3c56170353e17ad29edf8edff3f22c7ddc7e0ec95f4dab1.jpg) +Figure 1. Our SynCHMR recovers metric-scale camera trajectories (color pyramids), human meshes, and scene point clouds in a global coordinate from casual videos by joining forces of Human Mesh Recovery (HMR) and Simultaneous Localization and Mapping (SLAM). + +# Abstract + +Remarkable strides have been made in reconstructing static scenes or human bodies from monocular videos. Yet, the two problems have largely been approached independently, without much synergy. Most visual SLAM methods can only reconstruct camera trajectories and scene structures up to scale, while most HMR methods reconstruct human meshes in metric scale but fall short in reasoning with cameras and scenes. This work introduces Synergistic Camera and Human Reconstruction (SynCHMR) to marry the best of both worlds. Specifically, we design Human-aware Metric SLAM to reconstruct metric-scale camera poses and scene point clouds using camera-frame HMR as a strong prior, addressing depth, scale, and dynamic ambiguities. Conditioning on the dense scene recovered, we further learn a Scene-aware SMPL Denoiser to enhance world-frame HMR by incorporating spatio-temporal coherency and dynamic scene constraints. Together, they lead to consistent reconstructions of camera trajectories, human meshes, and dense scene point clouds in a common world frame. + +# 1. Introduction + +Physically plausible 3D human motion reconstruction from monocular videos is a long-standing problem in computer vision and graphics and has many applications in character animation, VFX, video games, sports, and healthcare. It requires estimating 3D humans across video frames in a common coordinate even with a moving camera. While human mesh recovery (HMR) has made significant progress recently [55], most existing methods typically estimate 3D humans in the camera coordinate by one frame at a time and fail to disambiguate camera motion. It calls for methods to jointly reconstruct 3D human and camera motion in a consistent global coordinate system from monocular videos. In other words, taking a video captured by a moving camera as input, the method should recover both temporally and spatially coherent movements of human bodies and cameras. + +Intuitively, if the accurate camera motion is given, one can transform the bodies from individual camera frames to a common world frame by multiplying the inverse of camera extrinsic matrices. In practice, with humans moving in the scene, estimating the camera motion of a video is still an open challenge in monocular SLAM [1]. It not only + +![](images/b62b71ad1330bd3eee757cc9f16d644b2d4915e10b4103900824a204b2ef7dbe.jpg) +(a) Depth ambiguity + +![](images/a785b7734ae7970d02717379d3d9844643c75e2a39dd7bb860031a3eaa85d7e7.jpg) +(b) Scale ambiguity + +![](images/8c84ab4284bfe3a986e9a31ab134b8832a8e42e81327e4ad39cd17f17622c302.jpg) +(c) Dynamic ambiguity +Figure 2. Illustration of three types of ambiguities in visual SLAM. We show SLAM reconstruction results from DROID-SLAM [54]. (a) Depth ambiguity occurs when there are only minor camera translations between different views. This can lead to geometric failures in reconstruction such as the folded back corridor in the side view. (b) Scale ambiguity is inherent in monocular SLAM systems and requires additional reference for disambiguation. (c) Dynamic ambiguity gets pronounced when moving foregrounds dominate frames. Over-reliance on foreground key points will result in incorrect camera trajectories. + +falls short in capturing accurate depths on views with small camera translations but more crucially, only estimates scene structures and camera trajectories up to scale. The human motion also breaks the static key point assumption in the bundle adjustment. As a result, one needs additional reference to disambiguate the depth, the scale, and the dynamic as illustrated in Fig. 2. + +To leverage SLAM results in HMR pipelines, current world-frame HMR methods often refine camera poses by integrating either partial camera parameters, such as a global scale of the translation [62], or full extrinsic matrices [15, 30, 46] in an optimization-based framework. However, their optimization-based nature leads to complex multi-stage schemes, making the overall pipeline unnecessarily slow and easy to break. + +In this work, we explore a fundamentally different way to marry the best of HMR and SLAM. A 2D object can first be lifted from the image plane to the camera frame and then transformed into a common 3D space. This two-step process coincides with the combination of camera-frame HMR, which brings imaged 2D humans to 3D camera frames, and SLAM, which estimates the camera-to-world transformation. Noticing these correspondences, we leverage camera-frame HMR as a strong prior to bridge from the image plane to the camera frame for disambiguating SLAM, and utilize + +SLAM reconstructions to constrain the transformation of human meshes from individual camera frames to a common global space. The overall pipeline thus results in a better synergy of the two, which we dub Synergistic Camera and Human Reconstruction (SynCHMR). + +We design SynCHMR based on several insights. First, despite camera-frame HMR methods cannot reconstruct humans in a coherent global frame, the estimated body dimension and location still provide cues to disambiguate SLAM. Unlike SLAHMR [62] which applies SLAM out of the box and corrects the scale afterward, we endow the SLAM process with human meshes from camera-frame HMR to address ambiguities. To this end, we capitalize on estimated absolute depths to provide pseudo-RGB-D inputs for SLAM [54] and confine the bundle adjustment to static backgrounds. Since current depth estimation methods [2, 43] predicts either relative depth maps or depths with data biases, we propose to calibrate their outputs by aligning with estimated human bodies in the camera frame [9]. With these priors, SLAM knows the depth, scale, and dynamic information from HMR and consequently estimates less ambiguous scene structures and camera poses. + +Next, we place human meshes in the common coordinate recovered by SLAM. The gap between human tracks transformed from camera frames and their real plausible world-frame motions stems from two sources of error: noise induced by camera-frame HMR and by SLAM. Motion prior models [14, 44, 68] can be used for denoising purposes as they contribute to the temporal coherence of worldframe human tracks. However, their exclusive focus on human modeling either leaves them agnostic to the underlying scenes [14] or assumes the scene is a simple ground plane [44, 68]. Our intuition is that when placing a human, static elements of the scene, such as the ground, and dynamic components like moving objects are both possible to be in contact with the human, thereby providing clues for placing the body coherently and compatibly with the scene. We hence introduce a Scene-aware SMPL Denoiser that learns to denoise the transformed human tracks by considering both temporal consistencies of moving humans and implicit constraints from dynamic scenes. This global awareness makes it more flexible for in-the-wild videos. + +Our contributions can be summed up as follows: + +- We present a novel pipeline, SynCHMR, that takes a monocular video as input and reconstructs human motions, camera trajectory and dense scene point clouds all in one global coordinate, as shown in Fig. 1, whereas current world-frame HMR methods [30, 62, 65] can recover only an estimated or pre-defined ground plane. +- We propose a novel Human-aware Metric SLAM process to robustly calibrate estimated depth with estimated human meshes, resulting in metric-scale camera pose estimation and metric-scale scene reconstruction. + +- We present Scene-aware SMPL Denoising that enforces spatiotemporal coherencies and applies dynamic scene constraints on world-frame human meshes. Notably, this is achieved without requiring extra annotations or heuristic designs to decide which part of a human should be interacting with the scene [49] and which region in the scene is most likely to be in contact with humans [38, 62]. + +# 2. Related Work + +There is considerable prior arts of HMR. We briefly discuss how they adopt different camera models and refer the readers to [1] for a more comprehensive review. + +HMR from a single image. State-of-the-art (SOTA) methods use parametric body models [21, 37, 41, 59] and estimate the parameters either by fitting to detected image features [3, 41, 58] or by regressing directly from pixels with deep neural networks [9, 11, 19, 22, 23, 28, 31, 32, 35, 45, 50, 60, 66, 67]. These approaches assume weak perspective/orthographic projection or pre-define the focal length as a large constant for all images. Kissos et al. [26] show that replacing focal length with a constant closer to ground truth alleviate the body tilting problem. SPEC [29] and Zolly [57] estimate focal length to account for perspective distortion. CIUFF [34] takes into account the location of humans in images to regress better poses in the camera coordinates. + +Many of these camera-frame HMR methods assume zero camera rotation, which entangles body rotation and camera rotation. When applied on video data, they fail to reconstruct humans in a coherent global space since they operate in a per-frame manner and hence cannot reason about how the camera moves across frames. + +HMR from videos aims to regress a series of body parameters from a temporal sequence. It opens up new problems such as whether the reconstructed bodies are in a common global coordinate or not. Some temporal methods consider a static camera [38, 44, 68], which makes the camera space a natural choice of the common coordinate. The challenge of coherent global space emerges when the camera moves. Early methods [5, 24, 27] show promising results on videos of dynamic cameras. Despite the reconstructed human meshes look great when overlaid on images, they do not share a common coordinate in 3D. + +Recent HMR methods capitalize on human motion prior to constrain the global trajectories in the world space, which in turn implicitly disentangles human movement from camera movement. GLAMR [65] consider a data-driven prior models learned on large-scale MoCap database e.g. AMASS [39], while D&D [33] and Yu [64] consider physic-inspired prior. These world-frame HMR methods often struggle on noise in local poses caused by partial occlusions, which is very common in in-the-wild videos with close-up shots and crowd scenes. Kaufmann et al. [25] and + +BodySLAM++ [16] circumvent this problem by employing IMU sensors to provide more robust body estimates but require extra sensory devices. To fully disentangle human and camera motion, another line of work [15, 30, 36, 46] leverages state-of-the-art SLAM techniques, e.g. [47, 54, 70], to explicitly estimate camera motion from the input video and infer the body parameters in the world coordinate of SLAM. Closest to us is SLAHMR [62] which solves for a global scale to connect the pre-computed SLAM results and body trajectories. To carefully guide the optimization process, these methods tend to have complex, multi-stage optimization schemes, making the overall pipeline easy to break and unnecessarily slow. + +Note that in stark contrast to the methods above, which either assume or estimate a simple ground plane as scene representation, SynCHMR reconstructs dense scenes from in-the-wild videos without pre-scanning with extra devices a priori like in [6, 7, 12, 13, 17, 61, 69]. We provide detailed comparisons with these world-frame HMR in Supp. Mat. + +# 3. Method + +Taking as input an RGB video $\{\mathbf{I}_t\in \mathbb{R}^{H\times W\times 3}\}_{t = 1}^T$ with $T$ frames and $N$ people in the scene, we aim to recover human meshes $\{\mathbf{V}_{nt}^{\mathrm{w}}\in \mathbb{R}^{3\times 6890}\}_{n = 1,t = 1}^{N,T}$ , dynamic scene point clouds $\{\mathbf{P}_t^{\mathrm{wm}}\in \mathbb{R}^{H\times W\times 3}\}_{t = 1}^T$ , and corresponding camera poses $\{\mathbf{G}_t^{\mathrm{m}}\in \mathrm{SE}(3)\}_{t = 1}^T$ in a common world coordinate system. The superscripts ${}^{\mathrm{W}},{}^{\mathrm{C}}$ , and ${}^{\mathrm{m}}$ denote the world frame, the camera frame, and the metric scale, respectively. To this aim, we propose a two-phase alternative conditioning pipeline as depicted in Fig. 3. In the first phase, we calibrate camera motion by injecting a camera-frame human prior to SLAM. This resolves depth, scale, and dynamic ambiguities, yielding metric-scale camera poses and dynamic point clouds. Subsequently, in the second phase, we transform the camera-frame human tracks into the world frame and utilize the dynamic point clouds obtained in the first phase for conditional denoising. + +# 3.1. Preliminaries + +# 3.1.1 SLAM + +Given a monocular RGB video $\{\mathbf{I}_t\}_{t = 1}^T$ DROID-SLAM [54] solves a dense bundle adjustment for a set of camera poses $\{\mathbf{G}_t\in \mathrm{SE}(3)\}_{t = 1}^T$ and inverse depths $\{\mathbf{d}_t\in \mathbb{R}_+^{H\times W}\}_{t = 1}^T$ . To update these estimations, it first computes a dense correspondence field $\mathbf{p}_{ij}\in \mathbb{R}^{H\times W\times 2}$ based on reprojection for each pair of frames $(i,j)$ .. + +$$ +\mathbf {p} _ {i j} = \Pi \left(\mathbf {G} _ {i j} \circ \Pi^ {- 1} \left(\mathbf {p} _ {i}, \frac {1}{\mathbf {d} _ {j}}\right)\right), \tag {1} +$$ + +where $\mathbf{p}_i\in \mathbb{R}^{H\times W\times 2}$ is a grid of pixel coordinates in frame $i$ $\mathbf{G}_{ij} = \mathbf{G}_j\circ \mathbf{G}_i^{-1}$ is the relative pose, and $\Pi$ and $\Pi^{-1}$ are the camera projection and inverse projection functions. + +![](images/0a5ef57b61c9a023648d8f342766ae6edf7f543a99cd003fd11a52004ae42588.jpg) +Figure 3. The architecture of SynCHMR. Our pipeline comprises two phases. The first phase, Human-aware Metric SLAM (Sec. 3.2), infers metric-scale camera poses and metric-scale point clouds by exploiting the camera-frame human prior. The second phase, Scene-aware SMPL Denoising (Sec. 3.3), involves the conditional denoising of world-frame noisy SMPL parameters. These parameters, initialized by transforming from the camera frame, get refined through conditioning on the dynamic point clouds obtained in the first phase. The whole pipeline thus reconstructs humans, scene point clouds, and cameras harmoniously in a common world frame. + +Then with a learned neural network, the system predicts a revision flow field $\mathbf{r}_{ij} \in \mathbb{R}^{H \times W \times 2}$ and associated confidence map $\mathbf{w}_{ij} \in \mathbb{R}_+^{H \times W \times 2}$ to construct the cost function + +$$ +E _ {\Sigma} = \sum_ {(i, j)} \left\| \mathbf {p} _ {i j} ^ {*} - \Pi \left(\mathbf {G} _ {i j} ^ {\prime} \circ \Pi^ {- 1} \left(\mathbf {p} _ {i}, \frac {1}{\mathbf {d} _ {i} ^ {\prime}}\right)\right) \right\| _ {\Sigma_ {i j}} ^ {2}, \tag {2} +$$ + +where $\mathbf{p}_{ij}^{*} = \mathbf{r}_{ij} + \mathbf{p}_{ij}$ is the corrected correspondence, $\| \cdot \|_{\Sigma}$ is the Mahalanobis distance which weighs the error terms with $\Sigma_{ij} = \mathrm{diag}\mathbf{w}_{ij}$ , and $\mathbf{G}'$ and $\mathbf{d}'$ are updated poses and inverse depths. Upon this objective, DROID-SLAM considers an additional term that penalizes the squared distance between the measured and predicted depth if the input is with an extra sensor depth channel $\{\mathbf{D}_t\}_{t=1}^T$ . + +# 3.1.2 HMR + +We employ 4DHumans [9] for reconstructing camera-frame human meshes from an in-the-wild video. Specifically, it performs per-frame human mesh recovery with an end-to-end transformer architecture and associates them to form human tracks. Each tracked human $n$ in frame $t$ is represented by SMPL [37] parameters as $\{\Phi_{nt},\theta_{nt},\beta_{nt},\Gamma_{nt}\}$ , including global orientation $\Phi_{nt}\in \mathbb{R}^{3\times 3}$ , body pose $\theta_{nt}\in \mathbb{R}^{22\times 3\times 3}$ , shape $\beta_{nt}\in \mathbb{R}^{10}$ , and root translation $\Gamma_{nt}\in \mathbb{R}^3$ . Then the parametric SMPL model can use these parameters to recover a human mesh with vertices $\mathbf{V}_{nt}\in \mathbb{R}^{3\times 6890}$ in metric scale: $\mathbf{V}_{nt} = \mathrm{SMPL}(\Phi_{nt},\theta_{nt},\beta_{nt}) + \mathbf{\Gamma}_{nt}$ . + +# 3.2. Human-aware Metric SLAM + +# 3.2.1 Preprocessing + +To start off, we estimate per-frame depth maps $\{\mathbf{D}_t\}$ with an off-the-shelf depth estimator, ZoeDepth [2] and + +predict per-frame human instance segmentation masks $\{\mathbf{M}_{nt}\}$ with an image instance segmentation network, Mask2Former [4]. We adapt ZoeDepth for video-consistent depth estimation by choosing a per-video metric head from the majority vote of per-frame routers, for which we dub ZoeDepth $^+$ . While ZoeDepth claims to estimate metric depths, we observe a domain gap when inference on new datasets. Consequently, we only treat its output as up-to-affine depths that need to be further aligned with the metric scale. To aid our optimization with human awareness, we use camera-frame human meshes $\{\mathbf{V}_{nt}^{\mathrm{c}}\}$ recovered by 4DHumans [9] to introduce a metric prior. + +# 3.2.2 Calibrating Depth with Human Prior + +We calibrate the per-frame depths with human meshes in Human-aware Depth Calibration. This involves optimizing two parameters, a world scale $s$ and a world offset $o$ , shared across all frames. During optimization, we linearly transform $\mathbf{D}_t$ to $\mathbf{D}_t' = s\mathbf{D}_t + o$ and unproject these depth maps to camera-frame point clouds $\{\mathbf{P}_t^c\}$ with $\mathbf{P}_t^c = \Pi^{-1}(\mathbf{p}_t, \mathbf{D}_t')$ . Our intuition is to align the human point cloud $\mathbf{P}_{nt}^c = \mathbf{M}_{nt} \odot \mathbf{P}_t^c$ with the camera-frame human mesh vertices $\mathbf{V}_{nt}^c$ in terms of absolute depth and size. To achieve pixel-wise alignment, we use a depth term to pull points on the human point cloud toward their corresponding human mesh vertices along the z-axis + +$$ +E _ {\text {d e p t h}} = \frac {\sum_ {n , t} \| \mathbf {S} _ {n t} \odot [ z \left(\mathbf {V} _ {n t} ^ {\mathrm {c}}\right) - z \left(\mathbf {P} _ {n t} ^ {\mathrm {c}}\right) ] \| _ {2} ^ {2}}{\sum_ {n , t} \| \mathbf {S} _ {n t} \| _ {0}}, \tag {3} +$$ + +where $\mathbf{S}_{nt} = \rho (\mathbf{V}_{nt}^{\mathrm{c}})\cap \mathbf{M}_{nt}$ is the intersection of the rasterized human mesh mask $\rho (\mathbf{V}_{nt}^{\mathrm{c}})$ and the instance segmentation mask $\mathbf{M}_{nt}$ , $z(\cdot)$ is the rasterized depth, and $\| \cdot \| _0$ is the + +![](images/441dac30c7eb42a434600800fbdd161d93bfc451e56c5004a14ff0e74d4c3252.jpg) +Figure 4. The architecture of Scene-aware SMPL Denoiser. World-frame noisy SMPL parameters $\{\Phi_{nt}^{\mathrm{w}},\theta_{nt},\beta_{nt},\Gamma_{nt}^{\mathrm{w}}\}_{0}$ are first projected by a linear layer and summed with temporal positional embeddings (TPE) to get initial latent humans $\{\mathbf{z}_{nt,0}^{\mathrm{SMPL}}\}$ . Per-frame point clouds are aggregated to $\mathbf{x}_{\mathrm{scene}}$ and encoded with the point encoder $\mathcal{E}$ . Then we query the encoded scene $\mathcal{E}(\mathbf{x}^{\mathrm{scene}})$ with latent humans $\{\mathbf{z}_{nt,0}^{\mathrm{SMPL}}\}$ in the scene-conditioned denoiser $\mathcal{D}$ and feed the result $\{\mathbf{z}_{nt,1}^{\mathrm{SMPL}}\}$ to prediction heads $\{\mathcal{P}_{\Phi},\mathcal{P}_{\theta},\mathcal{P}_{\beta},\mathcal{P}_{\Gamma}\}$ to obtain denoised SMPL parameters $\{\Phi_{nt}^{\mathrm{w}},\theta_{nt},\beta_{nt},\Gamma_{nt}^{\mathrm{w}}\}_{1}$ . + +0-norm indicating the number of non-zero pixels on a mask. + +As the recovered human meshes can be noisy in depth but still have a stable body dimension, we also adopt a size term to leverage the relative position of mesh vertices + +$$ +E _ {\mathrm {d x}} = \frac {\sum_ {n , t} \| \Delta_ {x} \left(\mathbf {V} _ {n t} ^ {\mathrm {c}} , \mathbf {S} _ {n t}\right) - \Delta_ {x} \left(\mathbf {P} _ {n t} ^ {\mathrm {c}} , \mathbf {S} _ {n t}\right) \| _ {2} ^ {2}}{N T}. \tag {4} +$$ + +We define $E_{\mathrm{dy}}$ similarly as $E_{\mathrm{dx}}$ , where + +$$ +\Delta_ {*} (\mathbf {X}, \mathbf {Y}) = \left(\max _ {*} - \min _ {*} \right) \left[ \Pi^ {- 1} (\mathbf {Y} \odot \Pi (\mathbf {X}), z (\mathbf {X})) \right] \tag {5} +$$ + +and $(\max_{*} - \min_{*})$ denotes the difference between the maximum value and the minimum value on coordinate $^*$ + +Then we have the calibrated depths with optimization + +$$ +\left(s ^ {\mathrm {m}}, o ^ {\mathrm {m}}\right) = \operatorname {a r g m i n} _ {s, o} \left(E _ {\text {d e p t h}} + \lambda E _ {\text {s i z e}}\right), \tag {6} +$$ + +$$ +\mathbf {D} _ {t} ^ {\mathrm {m}} = s ^ {\mathrm {m}} \mathbf {D} _ {t} + o ^ {\mathrm {m}}, \tag {7} +$$ + +where $E_{\mathrm{size}} = E_{\mathrm{dx}} + E_{\mathrm{dy}}$ , and $\lambda$ is a hyperparameter to balance two energy terms with a default value of 1. + +# 3.2.3 Disambiguating SLAM with Calibrated Depth + +While DROID-SLAM [54] originally supports RGB-D input mode where the D channel stands for sensor depth, one cannot trivially access sensor depths from in-the-wild videos. Our insight is that an estimated absolute depth can be utilized as a depth prior, albeit noisy. So we combine the original RGB video and the calibrated depth as pseudoRGB-D inputs $\{\mathbf{I}_t,\mathbf{D}_t^{\mathrm{m}}\}$ to disambiguate depth and scale. Furthermore, we modify the cost function Eq. (2) to resolve the dynamic ambiguity by masking out dynamic foregrounds in confidence maps + +$$ +\Sigma_ {i j} ^ {\prime} = \operatorname {d i a g} \mathbf {w} _ {i j} ^ {\prime} = \operatorname {d i a g} \left(\left(1 - \left[ \mathbf {M} _ {i}, \mathbf {M} _ {j} \right]\right) \odot \mathbf {w} _ {i j}\right), \tag {8} +$$ + +where $\mathbf{M}_i = \bigcup_n\mathbf{M}_{ni}$ and $\mathbf{M}_j = \bigcup_n\mathbf{M}_{nj}$ are the union of all human instance masks on their corresponding frame, and $[\cdot ,\cdot ]$ is the concatenation operation. As a result, we obtain metric-scale camera poses $\{\mathbf{G}_t^{\mathrm{m}}\}$ and metric-scale point clouds $\{\mathbf{P}_t^{\mathrm{wm}}\}$ by disambiguating SLAM with calibrated metric depths + +$$ +\left\{\mathbf {G} _ {t} ^ {\mathrm {m}}, \mathbf {d} _ {t} ^ {\mathrm {m}} \right\} = \operatorname {a r g m i n} _ {\left\{\mathbf {G} _ {i j} ^ {\prime}, \mathbf {d} _ {i j} ^ {\prime} \right\}} E _ {\Sigma^ {\prime}}, \tag {9} +$$ + +$$ +\mathbf {P} _ {t} ^ {\mathrm {w m}} = \mathbf {G} _ {t} ^ {\mathrm {m}} \circ \Pi^ {- 1} \left(\mathbf {p} _ {t}, \frac {1}{\mathbf {d} _ {t} ^ {\mathrm {m}}}\right)). \tag {10} +$$ + +# 3.3. Scene-aware SMPL Denoising + +# 3.3.1 Initializing Humans with Metric Cameras + +To put humans properly in the scene recovered by SLAM, we initialize them by transforming estimated camera-frame SMPL parameters $\{\Phi_{nt}^{\mathrm{c}},\theta_{nt},\beta_{nt},\Gamma_{nt}^{\mathrm{c}}\}$ to the world frame with camera-to-world transforms $\{\mathbf{G}_t^{\mathrm{m}} = [\mathbf{R}_t|\mathbf{t}_t^{\mathrm{m}}]\}$ . Given the pelvis as the center of global orientation $\Phi$ , we have: + +$$ +\boldsymbol {\Phi} _ {n t} ^ {\mathrm {w}} = \mathbf {R} _ {t} \boldsymbol {\Phi} _ {n t} ^ {\mathrm {c}}, \quad \boldsymbol {\Gamma} _ {n t} ^ {\mathrm {w}} = \mathbf {R} _ {t} \left(\boldsymbol {\Gamma} _ {n t} ^ {\mathrm {c}} + \mathbf {c}\right) + \mathbf {t} _ {t} ^ {\mathrm {m}} - \mathbf {c}, \tag {11} +$$ + +where $\mathbf{c} = \mathbf{c}(\beta_{nt})$ is the pelvis location in the shape blend body mesh. Note that we do not need to introduce an extra camera scale as SLAHMR [62] since the camera poses have already been in the metric scale. The root-relative poses $\theta_{nt}$ and the shapes $\beta_{nt}$ stay unchanged as in the camera frame. We denote the initialized and the denoised parameters with a suffix 0 and 1 respectively, i.e. $\{\Phi_{nt}^{\mathrm{w}},\theta_{nt},\beta_{nt},\Gamma_{nt}^{\mathrm{w}}\}_{0,1}$ . + +# 3.3.2 Constraining Humans with Dynamic Scenes + +Different from existing works [30, 36, 62, 64] that incorporate energy terms in optimization to apply explicit scene constraints, we propose to learn implicit scene constraints with a Scene-aware SMPL denoiser shown in Fig. 4. The noisy initial SMPL parameters $\{\Phi_{nt}^{\mathrm{w}},\theta_{nt},\beta_{nt},\Gamma_{nt}^{\mathrm{w}}\}_{0}$ are first projected to a latent space, where it gets further updated by conditioning on implicit scene constraints + +$$ +\mathbf {z} _ {n t, 0} ^ {\mathrm {S M P L}} = \operatorname {F C} \left(\left[ \boldsymbol {\Phi} _ {n t, 0} ^ {\mathrm {w}}, \boldsymbol {\theta} _ {n t, 0}, \boldsymbol {\beta} _ {n t, 0}, \boldsymbol {\Gamma} _ {n t, 0} ^ {\mathrm {w}} \right]\right) + \text {T P E}, \tag {12} +$$ + +$$ +\left\{\mathbf {z} _ {n t, 1} ^ {\text {S M P L}} \right\} _ {t = 1} ^ {T} = \mathcal {D} \left(\left\{\mathbf {z} _ {n t, 0} ^ {\text {S M P L}} \right\} _ {t = 1} ^ {T}, \mathcal {E} \left(\mathbf {x} ^ {\text {s c e n e}}\right) + \text {T P E}\right), \tag {13} +$$ + +where FC is a shared linear layer, TPE is shared temporal positional embeddings, $\{\mathbf{z}_{nt,*}^{\mathrm{SMPL}}\}_{t = 1}^{T}\in \mathbb{R}^{T\times D}$ is the $D$ dimensional latent for human $n$ , and $\mathbf{x}^{\mathrm{scene}}\in \mathbb{R}^{L\times C}$ is the $C$ -channel dynamic scene point clouds with a total number of points $L$ . $\mathcal{E}$ and $\mathcal{D}$ refer to the scene encoder and the scene-conditioned denoiser, respectively. We set $C = 7$ which is the concatenation of point coordinates $\{\mathbf{P}_t^{\mathrm{wm}}\}$ , colors $\{\mathbf{I}_t\}$ , and estimated human semantic segmentation masks $\{\mathbf{M}_t = \bigcup_n\mathbf{M}_{nt}\}$ . Following [9], the updated latent + +are decoded with different prediction head $\mathcal{P}_{(\cdot)}$ to regress the residual for each SMPL parameter: + +$$ +\boldsymbol {\Phi} _ {n t, 1} ^ {\mathrm {w}} = \mathcal {P} _ {\boldsymbol {\Phi}} \left(\mathbf {z} _ {n t, 1} ^ {\mathrm {S M P L}}\right) \boldsymbol {\Phi} _ {n t, 0} ^ {\mathrm {w}}, \tag {14} +$$ + +$$ +\boldsymbol {\theta} _ {n t, 1} = \mathcal {P} _ {\boldsymbol {\theta}} \left(\mathbf {z} _ {n t, 1} ^ {\mathrm {S M P L}}\right) \boldsymbol {\theta} _ {n t, 0}, \tag {15} +$$ + +$$ +\boldsymbol {\beta} _ {n t, 1} = \mathcal {P} _ {\boldsymbol {\beta}} \left(\mathbf {z} _ {n t, 1} ^ {\mathrm {S M P L}}\right) + \boldsymbol {\beta} _ {n t, 0}, \tag {16} +$$ + +$$ +\boldsymbol {\Gamma} _ {n t, 1} ^ {\mathrm {w}} = \mathcal {P} _ {\boldsymbol {\Gamma}} \left(\mathbf {z} _ {n t, 1} ^ {\mathrm {S M P L}}\right) + \boldsymbol {\Gamma} _ {n t, 0} ^ {\mathrm {w}}. \tag {17} +$$ + +We apply direct supervision on $\{\Phi_{nt}^{\mathrm{w}},\theta_{nt},\beta_{nt},\Gamma_{nt}^{\mathrm{w}}\}_{1}$ which is common in the literature. Please see Supp. Mat. for the details of the full training objectives. + +# 4. Experiments + +# 4.1. Experimental Setting + +Datasets. We assess the performance of SynCHMR primarily for global human motion estimation but also report the accuracy of estimated camera trajectories. Traditional video datasets in HMR literature are typically captured by static cameras, e.g. [13, 18, 20, 40, 63], hence not suitable for our purpose. Standard SLAM benchmarks such as [48, 51] do not meet our needs either as there is often no human moving in the scene. We consider the following datasets. + +3DPW [56] is an in-the-wild dataset captured with iPhones. The ground truth bodies are not in coherent world frames so we use it to supervise root relative poses and for evaluation. EgoBody [69] has ground-truth poses captured by multiple Kinects and egocentric-view sequences recorded by a head-mounted device, whose trajectories are further registered in the world space of Kinect array. We use it for training the SMPL denoiser in Sec. 3.3 and for evaluation (on both body and camera estimation). For HMR evaluation, unlike [30, 62] considering only the validation set, we additionally report results on its completely withheld test set. + +EMDB [25] is a new dataset providing SMPL poses from IMU sensors and global camera trajectories. We include it for training the SMPL denoiser to enrich the diversity and use the camera trajectories to evaluate the quality of SLAM. + +Evaluation Metrics. For HMR evaluation, we report common PA-MPJPE, which measures the quality of rootrelative poses. For datasets that have ground-truth poses in a world coordinate, we follow [62] and consider WA-MPJPE and FA-MPJPE. The former measures the error after aligning the entire trajectories of the prediction and ground truth with Procrustes Alignment [10], while the latter aligns only with the first frame. We also report acceleration errors. For SLAM, we consider absolute trajectory error (ATE) for camera trajectory evaluation as well as the threshold accuracy $(\delta_{n})$ , the absolute relative error (REL), and the root mean squared error (RMSE) for scene depth evaluation [2]. + +Implementation Details. In Human-aware Depth Calibration, we use the L-BFGS algorithm with learning rate 1 to + +
Camera ModelHuman ModelPA ↓
DROID-SLAM [54]SLAHMR [62] w/ PHALP+55.9
DROID-SLAM [54]SLAHMR [62] w/ 4DHumans [9]57.4
Human-aware Metric SLAM (ours)4DHumans [9]52.9
Human-aware Metric SLAM (ours)Scene-aware SMPL Denoiser (ours)52.4
+ +Table 1. Comparison results on 3DPW-Test. The row in gray is the full pipeline of SynCHMR. We abbreviate PA-MPJPE as PA, with the same below for FA-MPJPE (FA) and WA-MPJPE (WA). + +optimize for a maximum of 30 iterations. As for the Scene-aware SMPL Denoiser, we train it on the union of 3DPW-Train, EgoBody-Train, and EMDB for 100k steps with an AdamW optimizer, a batch size of 16, and a learning rate of 1e-5. For camera-frame SMPL ground truths like in 3DPW, we only incorporate body shapes $\beta$ and poses $\theta$ in training. We train the denoising process by randomly sampling a temporal window size $T$ spanning 64 to 128 and inference with $T = 100$ . The scene-conditioned denoiser $\mathcal{D}$ is parameterized with a 6-layer Transformer Decoder. For the scene encoder $\mathcal{E}$ , we consider ViT and SPVCNN in Tab. 4 and report results for SPVCNN in Tabs. 1 and 2. Before inputting the world-frame noisy SMPL parameters to the denoiser, we first interpolate $\Phi_{nt,0}^{\mathrm{w}}$ and $\theta_{nt,0}$ on SO(3), $\beta_{nt,0}$ on $\mathbb{R}^{10}$ , and $\Gamma_{nt,0}^{\mathrm{w}}$ on $\mathbb{R}^3$ when there are missing observations. + +# 4.2. Comparison Results + +We first evaluate the estimated local poses with PA-MPJPE on 3DPW, which is common in the literature. In Tab. 1, we show that placing the bodies from 4DHumans already leads to lower error than SLAHMR. Passing them through the denoiser further reduces the error. We note that PA-MPJPE only measures local pose accuracy not the quality of global trajectories. Since 3DPW does not support any world metrics, Tab. 1 only aims to show that SynCHMR produces reasonable local poses on a common dataset. + +Next, we assess the quality of global motion estimation, which is essentially a more challenging task. Tab. 2 shows the results on EgoBody. Note that current optimization-based methods [30, 62] report the error of the validation set. For fairness and completeness, we report results on both validation and test sets and run state-of-the-art methods on the test set when the code is available. In Tab. 2, we see that the proposed SynCHMR has the overall lowest PA-MPJPE, FA-MPJPE, and WA-MPJPE (gray rows). Comparing it with the row above (4DHumans) confirms the benefit of our scene-conditioned denoiser. For a fair comparison, we also initialize the global optimization of SLAHMR with 4DHumans, which is more accurate than $\mathrm{PHALP}^+$ in SLAHMR, but we do not observe improvement. Notably, despite the concurrent work PACE [30] has a tightly integrated SLAM and body fitting objective, it still uses native DROID-SLAM to initialize the camera parameters like SLAHMR does. This is arguably sub-optimal as the initial + +
SubsetCamera ModelHuman ModelPA-MPJPE (mm) ↓FA-MPJPE (mm) ↓WA-MPJPE (mm) ↓Acc Error (mm/frame2) ↓Runtime/100 ims
Val-GLAMR [65]114.3416.1239.0173.54 min
DROID-SLAM [54]PACE [30]66.5147.9101.06.71 min
DROID-SLAM [54]SLAHMR [62] w/ PHALP+79.1141.1101.225.840 min
DROID-SLAM [54]SLAHMR [62] w/ 4DHumans [9]79.3273.0144.779.440 min
Human-aware Metric SLAM (ours)4DHumans [9]73.0164.4106.7127.05 min
Human-aware Metric SLAM (ours)Scene-aware SMPL Denoiser (ours)57.7115.181.164.85 min
Test-GLAMR [65]112.8351.4216.3105.94 min
DROID-SLAM [54]SLAHMR [62] w/ PHALP+63.1163.999.431.740 min
DROID-SLAM [54]SLAHMR [62] w/ 4DHumans [9]69.3185.8113.045.740 min
Human-aware Metric SLAM (ours)4DHumans [9]75.4160.0108.1138.85 min
Human-aware Metric SLAM (ours)Scene-aware SMPL Denoiser (ours)61.3122.184.669.45 min
+ +Table 2. Comparison results with state-of-the-art approaches on EgoBody. The row in gray is the full pipeline of SynCHMR. + +
RGBDepthMaskEgoBodyEMDB +ATE ↓
ATE ↓δ1 ↑REL ↓RMSE ↓
✓XX80.90.08514.5901617.361400.3
✓XMask2Former [4]81.60.0638.5301009.127385.8
✓ZoeDepth+X35.00.5620.30815.360456.8
✓ZoeDepth+Mask2Former [4]28.60.5640.30710.852389.6
✓ZoeDepth+ + Cal.Mask2Former [4]26.40.7970.27410.452107.0
+ +Table 3. Ablation study for SLAM configurations in terms of optimized camera trajectories and scene depths. ZoeDepth+ denotes our video-adapted ZoeDepth [2]. + +
StageBackboneRGBXYZMaskPA ↓FA ↓WA ↓Acc Error ↓
Init.-XXX73.7120.893.1127.1
Pred.-XXX63.398.877.275.2
Pred.ViT [8]✓XX63.994.976.743.3
Pred.ViT [8]✓✓X64.597.377.745.6
Pred.ViT [8]✓X✓66.896.578.644.7
Pred.ViT [8]✓✓✓69.3100.982.046.4
Pred.SPVCNN [53]X✓X62.995.176.072.6
Pred.SPVCNN [53]✓✓X61.093.474.367.7
Pred.SPVCNN [53]X✓✓62.093.975.369.9
Pred.SPVCNN [53]✓✓✓61.391.973.664.8
+ +Table 4. Ablation study for different scene encoders and features regarding world-frame HMR. Init. and Pred. refer to before and after SMPL denoising, respectively. + +ization is not aware of body information, which can lead to errors that cannot be corrected in the global optimization stage. Consequently, it also has higher world-space errors. Optimization methods often employ a zero velocity term to smooth out human motion, which explains the lower acceleration error. However, we do not observe a big difference in jittery between our results. Please refer to Supp. Mat. for more details. + +# 4.3. Ablation Study + +We ablate the design choices in SynCHMR. In Tab. 3, we evaluate SLAM-optimized camera trajectories and scene depths with EgoBody and EMDB. We see that directly including un-calibrated monocular depths does not guarantee more accurate estimations ( $3^{\mathrm{rd}}$ vs. $1^{\mathrm{st}}$ and $4^{\mathrm{th}}$ vs. $2^{\mathrm{nd}}$ row). Precluding the dynamic foreground pixels with Mask2Former [4] generally improves performance. We empirically find that our depth calibration with human prior works the best when using it with foreground masking, which has the lowest error in both datasets. More SLAM evaluation and discussion can be found in Supp. Mat. + +In Tab. 4, we verify the benefit of scene conditioning for the SMPL denoiser. We train it with EgoBody-train in different conditioning schemes and report the $T = 32$ results on EgoBody-val. First, placing the predicted bodies from 4DHuman in the global space directly with estimated camera extrinsics has the highest error (1 $^{\text{st}}$ row). When conditioning on a constant zero tensor, the denoiser behaves like a motion prior and reduces the error (2 $^{\text{nd}}$ row). To encode the appearance and geometry information of the scene, we consider ViT [8] or SPVCNN [53] as the encoder $\mathcal{E}$ and try varied combinations of appearance features (RGB), geometry features (XYZ) and aggregated subject masks (Mask). When using ViT to encode the scene, adding XYZ features or masks does not reduce the error. In contrast, when using SPVCNN, adding RGB information or conditioning on masks does improve performance. Overall, SPVCNN yields lower errors than ViT and enabling all conditioning leads to the lowest world-space error measure. + +# 4.4. Qualitative Analysis and Discussion + +In the first two rows of Fig. 5, we visualize the results of 3DPW and EgoBody in a global space. Despite occlusions, our SynCHMR estimates human meshes reliably and places them in a dense scene point cloud, whereas the scenes in GLAMR [65] and SLAHMR [62] consist of only a simple ground plane. Applying scene constraints with such an overly simplified scene can result in erroneous estimation, e.g., incorrect human trajectories as shown in the top view of the $1^{\text{st}}$ row, and the vertically shortened human bodies in the $4^{\text{th}}$ row of (d). Note that since TRACE [52] is scene agnostic, the ground plane in (c) is only for visualization, not necessarily indicating scene penetration. + +We also test on more in-the-wild DAVIS [42] videos containing human subjects. Since DAVIS provides no ground-truth human meshes nor camera trajectories, we show only the visual comparison. The $3^{\mathrm{rd}}$ row shows that we can handle multi-person cases as well as SLAHMR, while GLAMR often fails when multiple humans and dynamic cameras both occur. In a challenging scenario where the subject is taking selfies (the $4^{\mathrm{th}}$ row), both GLAMR and SLAHMR are confused by the foreground human dominating the frames and reconstruct an almost static global trajectory, failing to + +![](images/2d1ff7543f2ddc4f0a56d27c65a6b48963f0baaf3a8c307f5d0ab13b2a563ff0.jpg) +Figure 5. Qualitative comparison among world-frame HMR approaches. We show (b) GLAMR [65] and (c) TRACE [52] results with their pre-defined ground planes, (d) SLAHMR [62] outputs with its estimated ground plane, and (e) our SynCHMR outputs with dense scenes. In the first row, we also demonstrate top-view human trajectories within circles. See supplementary for video results. + +disentangle the camera and the human motions due to the dynamic ambiguity. TRACE fails to produce results due to severe frame truncation. In contrast, SynCHMR still successfully provides reasonable trajectories. + +# 5. Limitation Discussion + +As SynCHMR focuses on disentangling camera and human movements, we follow SLAHMR to approximate the focal length as $\frac{W + H}{2}$ . When the subject has a shape that the body model cannot explain well, e.g., children or obese people, calibrating depth with the estimated bodies is less ideal. As we develop and validate SynCHMR on real videos, its accuracy on composed or generated videos remains an open question. Finally, since SynCHMR handles dynamic scenes with moving subjects, it does not require an a priori scanned static scene. This opens up new challenges, such as incorporating dynamic point clouds as scene constraints. + +# 6. Conclusion + +We present SynCHMR, a method that reconstructs camera trajectories, human bodies, and dense scenes from in-the + +wild videos all in one global coordinate. SynCHMR has two core innovations. First, it leverages monocular depth estimation and uses the dimension and location of human meshes to calibrate the range of depth. This allows SLAM to better resolve the inherent scale ambiguity problem as shown in the experiment. Second, we train a data-driven motion denoiser and condition it with the scene in the same global coordinate, which is the first such scene-conditioned motion prior. Combining the two, the full SynCHMR pipeline uses human bodies to improve SLAM, and the better estimated scene and camera trajectory, in turn, provide better constraints for feed-forward human motion denoising. It achieves SOTA results on common benchmarks compared with existing optimization-based approaches. + +# Acknowledgment + +We appreciate constructive comments from Duygu Ceylan. This project was partially supported by the NIH under contracts R01GM134020 and P41GM103712, and by the NSF under contracts DBI-1949629, DBI-2238093, IIS-2007595, IIS-2211597, and MCB-2205148. + +# References + +[1] Iman Abaspur Kazerouni, Luke Fitzgerald, Gerard Dooly, and Daniel Toal. A survey of state-of-the-art on visual SLAM. Expert Systems with Applications, 205:117734, 2022. 1, 3 +[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. ZoeDepth: Zero-shot transfer by combining relative and metric depth. arXiv, 2023. 2, 4, 6, 7, 1 +[3] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European Conference on Computer Vision (ECCV), 2016. 3 +[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Computer Vision and Pattern Recognition (CVPR), pages 1290–1299, 2022. 4, 7, 2 +[5] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consistent 3d human pose and shape from a video. In Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[6] Yudi Dai, Yitai Lin, Chenglu Wen, Siqi Shen, Lan Xu, Jingyi Yu, Yuexin Ma, and Cheng Wang. HSC4D: Human-centered 4D scene capture in large-scale indoor-outdoor space using wearable imus and LiDAR. In Computer Vision and Pattern Recognition (CVPR), pages 6792-6802, 2022. 3 +[7] Yudi Dai, Yitai Lin, Xiping Lin, Chenglu Wen, Lan Xu, Hongwei Yi, Siqi Shen, Yuexin Ma, and Cheng Wang. Sloper4d: A scene-aware dataset for global 4d human pose estimation in urban environments. In Computer Vision and Pattern Recognition (CVPR), pages 682-692, 2023. 3 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 7 +[9] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 5, 6, 7, 1 +[10] John C Gower. Generalized procrustes analysis. Psychometrika, 40(1):33-51, 1975. 6 +[11] Riza Alp Guler and Iasonas Kokkinos. HoloPose: Holistic 3D human reconstruction in-the-wild. In Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[12] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human positioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In Computer Vision and Pattern Recognition (CVPR). IEEE, 2021. 3 +[13] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambigu + +ities with 3D scene constraints. In International Conference on Computer Vision (ICCV), pages 2282-2292, 2019. 3, 6, 1 +[14] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. Advances in Neural Information Processing Systems, 35:4244-4256, 2022. 2 +[15] Dorian F Henning, Tristan Laidlow, and Stefan Leutenegger. BodySLAM: joint camera localisation, mapping, and human motion tracking. In European Conference on Computer Vision (ECCV), pages 656-673. Springer, 2022. 2, 3, 1 +[16] Dorian F Henning, Christopher Choi, Simon Schaefer, and Stefan Leutenegger. BodySLAM++: Fast and tightly-coupled visual-inertial camera and human motion tracking. In International Conference on Intelligent Robots and Systems (IROS), pages 3781-3788. IEEE, 2023. 3 +[17] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In Computer Vision and Pattern Recognition (CVPR), pages 13274-13285, 2022. 3, 1 +[18] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6M: Large scale datasets and predictive methods for 3D human sensing in natural environments. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 36(7):1325-1339, 2014. 6 +[19] Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[20] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In International Conference on Computer Vision (ICCV), 2015. 6 +[21] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3D deformation model for tracking faces, hands, and bodies. In Computer Vision and Pattern Recognition (CVPR), pages 8320-8329, 2018. 3 +[22] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human pose fitting towards in-the-wild 3D human pose estimation. In International Conference on 3D Vision (3DV), pages 42-52, 2021. 3 +[23] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018. 3 +[24] Angjoo Kanazawa, Jason Y. Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. In Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[25] Manuel Kaufmann, Jie Song, Chen Guo, Kaiyue Shen, Tianjian Jiang, Chengcheng Tang, Juan Jose Zárate, and Otmar Hilliges. EMDB: The Electromagnetic Database of Global 3D Human Pose and Shape in the Wild. In International Conference on Computer Vision (ICCV), 2023. 3, 6 +[26] Imry Kissos, Lior Fritz, Matan Goldman, Omer Meir, Edward Oks, and Mark Kliger. Beyond weak perspective for + +monocular 3d human pose estimation. In European Conference on Computer Vision Workshops (ECCVw), pages 541-554, Cham, 2020. Springer International Publishing. 3 +[27] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[28] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In International Conference on Computer Vision (ICCV). IEEE, 2021. 3 +[29] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J. Black. SPEC: Seeing people in the wild with an estimated camera. In International Conference on Computer Vision (ICCV), pages 11035-11045, 2021. 3 +[30] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J. Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. PACE: Human and motion estimation from in-the-wild videos. In International Conference on 3D Vision (3DV), 2024. 2, 3, 5, 6, 7, 1 +[31] Nikos Kolotouros, Georgios Pavlakos, Michael J. Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In International Conference on Computer Vision (ICCV), 2019. 3 +[32] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), pages 3383-3393, 2021. 3 +[33] Jiefeng Li, Siyuan Bian, Chao Xu, Gang Liu, Gang Yu, and Cewu Lu. D&D: Learning human dynamics from dynamic camera. In European Conference on Computer Vision (ECCV), 2022. 3, 1, 2 +[34] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In European Conference on Computer Vision (ECCV), pages 590-606, 2022. 3 +[35] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In Computer Vision and Pattern Recognition (CVPR), pages 21159-21168, 2023. 3 +[36] Miao Liu, Dexin Yang, Yan Zhang, Zhaopeng Cui, James M Rehg, and Siyu Tang. 4D human body capture from egocentric video via 3D scene grounding. In International Conference on 3D Vision (3DV), pages 930-939. IEEE, 2021. 3, 5, 1, 2 +[37] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. Transactions on Graphics (TOG), 34 (6):248:1-248:16, 2015. 3, 4 +[38] Diogo Luvizon, Marc Habermann, Vladislav Golyanik, Adam Kortylewski, and Christian Theobalt. Scene-Aware 3D Multi-Human Motion Capture from a Single Camera. Computer Graphics Forum (CGF), 42(2):371-383, 2023. 3 +[39] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of + +motion capture as surface shapes. In International Conference on Computer Vision (ICCV), pages 5441-5450, 2019. 3 +[40] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3D pose estimation from monocular RGB. In International Conference on 3D Vision (3DV), 2018. 6 +[41] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[42] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Computer Vision and Pattern Recognition (CVPR), 2016. 7, 2 +[43] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(3):1623-1637, 2020. 2 +[44] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. HuMoR: 3D human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), pages 11468-11479, 2021. 2, 3 +[45] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. FrankMocap: Fast monocular 3d hand and body motion capture by regression and integration. In International Conference on Computer Vision Workshops (ICCVw), 2021. 3 +[46] Nitin Saini, Chun-Hao P Huang, Michael J Black, and Aamir Ahmad. SmartMocap: Joint estimation of human and camera motion using uncalibrated rgb cameras. IEEE Robotics and Automation Letters, 2023. 2, 3, 1 +[47] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Computer Vision and Pattern Recognition (CVPR), 2016. 3, 1, 2 +[48] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Computer Vision and Pattern Recognition (CVPR), pages 134-144, 2019. 6 +[49] Zehong Shen, Zhi Cen, Sida Peng, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Learning human mesh recovery in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17038-17047, 2023. 3 +[50] Jie Song, Xu Chen, and Otmar Hilliges. Human body model fitting by learned gradient descent. In European Conference on Computer Vision (ECCV), 2020. 3 +[51] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A benchmark for the evaluation of rgb-d slam systems. In International Conference on Intelligent Robots and Systems (IROS), 2012. 6, 1, 3 +[52] Yu Sun, Qian Bao, Wu Liu, Tao Mei, and Michael J. Black. TRACE: 5D temporal regression of avatars with dynamic cameras in 3d environments. In Computer Vision and Pat + +tern Recognition (CVPR), pages 8856-8866, 2023. 7, 8, 1, 2 +[53] Haotian* Tang, Zhijian* Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European Conference on Computer Vision, 2020. 7 +[54] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2, 3, 5, 6, 7, 1 +[55] Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3d human mesh from monocular images: A survey. arXiv preprint arXiv:2203.01923, 2022. 1 +[56] Timo von Marcard, Roberto Henschel, Michael J. Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3D human pose in the wild using IMUs and a moving camera. In European Conference on Computer Vision (ECCV), pages 614-631, 2018. 6 +[57] Wenjia Wang, Yongtao Ge, Haiyi Mei, Zhongang Cai, Qingping Sun, Yanjun Wang, Chunhua Shen, Lei Yang, and Taku Komura. Zolly: Zoom focal length correctly for perspective-distorted human mesh reconstruction. In International Conference on Computer Vision (ICCV), 2023. 3 +[58] Donglai Xiang, Hanbyul Joo, and Yaser Sheikh. Monocular total capture: Posing face, body, and hands in the wild. In Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[59] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In Computer Vision and Pattern Recognition (CVPR), pages 6183-6192, 2020. 3 +[60] Yuanlu Xu, Song-Chun Zhu, and Tony Tung. DenseRaC: Joint 3d pose and shape estimation by dense render-and compare. In International Conference on Computer Vision (ICCV), 2019. 3 +[61] Ming Yan, Xin Wang, Yudi Dai, Siqi Shen, Chenglu Wen, Lan Xu, Yuexin Ma, and Cheng Wang. Cimi4d: A large multimodal climbing motion dataset under human-scene interactions. In Computer Vision and Pattern Recognition (CVPR), pages 12977-12988, 2023. 3 +[62] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In Computer Vision and Pattern Recognition (CVPR), pages 21222-21232, 2023. 2, 3, 5, 6, 7, 8, 1 +[63] Jae Shin Yoon, Zhixuan Yu, Jaesik Park, and Hyun Soo Park. HUMBI: A large multiview dataset of human body expressions and benchmark challenge. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(1):623-640, 2021. 6 +[64] Ri Yu, Hwangpil Park, and Jehee Lee. Human dynamics from monocular video with dynamic camera movements. Transactions on Graphics (TOG), 40(6), 2021. 3, 5, 1, 2 +[65] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. GLAMR: Global occlusion-aware human mesh recovery with dynamic cameras. In Computer Vision and Pattern Recognition (CVPR), pages 11028-11039, 2022. 2, 3, 7, 8, 1 + +[66] Andrei Zanfir, Eduard Gabriel Bazavan, Hongyi Xu, Bill Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Weakly supervised 3d human pose and shape reconstruction with normalizing flows. In European Conference on Computer Vision (ECCV), 2020. 3 +[67] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In International Conference on Computer Vision (ICCV), 2021. 3 +[68] Siwei Zhang, Yan Zhang, Federica Bogo, Marc Pollefeys, and Siyu Tang. Learning motion priors for 4D human body capture in 3D scenes. In International Conference on Computer Vision (ICCV), 2021. 2, 3 +[69] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. EgoBody: Human body shape and motion of interacting people from head-mounted devices. In European Conference on Computer Vision (ECCV), 2022. 3, 6 +[70] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snively, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision (ECCV), pages 20-37. Springer, 2022. 3 \ No newline at end of file diff --git a/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/images.zip b/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4e66491b39bad76d82d23ca78c11e1e2bdda4a17 --- /dev/null +++ b/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4eb93ba8f322e44d7af2e730906369fdde58e7b3dcdf5e4862f0d1bdb47e67f6 +size 772272 diff --git a/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/layout.json b/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2b760a08f99e3f2a7e7b8dca88a0fabbb31b77cf --- /dev/null +++ b/2024/Synergistic Global-space Camera and Human Reconstruction from Videos/layout.json @@ -0,0 +1,10096 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "type": "text", + "content": "Synergistic Global-space Camera and Human Reconstruction from Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "spans": [ + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "type": "text", + "content": "Yizhou Zhao" + }, + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "type": "text", + "content": "; Tuanfeng Yang Wang" + }, + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "type": "text", + "content": ", Bhiksha Raj" + }, + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "type": "text", + "content": ", Min Xu" + }, + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "type": "text", + "content": ", Jimei Yang" + }, + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "type": "text", + "content": ", Chun-Hao Paul Huang" + }, + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "type": "text", + "content": "Carnegie Mellon University \n" + }, + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 539, + 171 + ], + "type": "text", + "content": "Adobe Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 101, + 173, + 488, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 173, + 488, + 185 + ], + "spans": [ + { + "bbox": [ + 101, + 173, + 488, + 185 + ], + "type": "text", + "content": "{yizhouz,bhiksha,mxul}@cs.cmu.edu,{yangtwan,jimyang,chunhaoh}@adobe.com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 198, + 545, + 399 + ], + "blocks": [ + { + "bbox": [ + 50, + 198, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 50, + 198, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 50, + 198, + 545, + 399 + ], + "type": "image", + "image_path": "7af274f60e1e4a33d3c56170353e17ad29edf8edff3f22c7ddc7e0ec95f4dab1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 404, + 546, + 427 + ], + "lines": [ + { + "bbox": [ + 46, + 404, + 546, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 546, + 427 + ], + "type": "text", + "content": "Figure 1. Our SynCHMR recovers metric-scale camera trajectories (color pyramids), human meshes, and scene point clouds in a global coordinate from casual videos by joining forces of Human Mesh Recovery (HMR) and Simultaneous Localization and Mapping (SLAM)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 437, + 192, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 437, + 192, + 449 + ], + "spans": [ + { + "bbox": [ + 143, + 437, + 192, + 449 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 456, + 289, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 289, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 289, + 684 + ], + "type": "text", + "content": "Remarkable strides have been made in reconstructing static scenes or human bodies from monocular videos. Yet, the two problems have largely been approached independently, without much synergy. Most visual SLAM methods can only reconstruct camera trajectories and scene structures up to scale, while most HMR methods reconstruct human meshes in metric scale but fall short in reasoning with cameras and scenes. This work introduces Synergistic Camera and Human Reconstruction (SynCHMR) to marry the best of both worlds. Specifically, we design Human-aware Metric SLAM to reconstruct metric-scale camera poses and scene point clouds using camera-frame HMR as a strong prior, addressing depth, scale, and dynamic ambiguities. Conditioning on the dense scene recovered, we further learn a Scene-aware SMPL Denoiser to enhance world-frame HMR by incorporating spatio-temporal coherency and dynamic scene constraints. Together, they lead to consistent reconstructions of camera trajectories, human meshes, and dense scene point clouds in a common world frame." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 437, + 386, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 386, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 386, + 449 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 456, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 635 + ], + "type": "text", + "content": "Physically plausible 3D human motion reconstruction from monocular videos is a long-standing problem in computer vision and graphics and has many applications in character animation, VFX, video games, sports, and healthcare. It requires estimating 3D humans across video frames in a common coordinate even with a moving camera. While human mesh recovery (HMR) has made significant progress recently [55], most existing methods typically estimate 3D humans in the camera coordinate by one frame at a time and fail to disambiguate camera motion. It calls for methods to jointly reconstruct 3D human and camera motion in a consistent global coordinate system from monocular videos. In other words, taking a video captured by a moving camera as input, the method should recover both temporally and spatially coherent movements of human bodies and cameras." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "content": "Intuitively, if the accurate camera motion is given, one can transform the bodies from individual camera frames to a common world frame by multiplying the inverse of camera extrinsic matrices. In practice, with humans moving in the scene, estimating the camera motion of a video is still an open challenge in monocular SLAM [1]. It not only" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 261, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 261, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 261, + 712 + ], + "type": "text", + "content": "*Part of this work was done when interned at Adobe Research." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1216" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 286, + 134 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 286, + 134 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 286, + 134 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 286, + 134 + ], + "type": "image", + "image_path": "b62b71ad1330bd3eee757cc9f16d644b2d4915e10b4103900824a204b2ef7dbe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 137, + 196, + 146 + ], + "lines": [ + { + "bbox": [ + 138, + 137, + 196, + 146 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 196, + 146 + ], + "type": "text", + "content": "(a) Depth ambiguity" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 148, + 285, + 186 + ], + "blocks": [ + { + "bbox": [ + 53, + 148, + 285, + 186 + ], + "lines": [ + { + "bbox": [ + 53, + 148, + 285, + 186 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 285, + 186 + ], + "type": "image", + "image_path": "a785b7734ae7970d02717379d3d9844643c75e2a39dd7bb860031a3eaa85d7e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 190, + 196, + 199 + ], + "lines": [ + { + "bbox": [ + 139, + 190, + 196, + 199 + ], + "spans": [ + { + "bbox": [ + 139, + 190, + 196, + 199 + ], + "type": "text", + "content": "(b) Scale ambiguity" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 53, + 201, + 286, + 278 + ], + "blocks": [ + { + "bbox": [ + 53, + 201, + 286, + 278 + ], + "lines": [ + { + "bbox": [ + 53, + 201, + 286, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 201, + 286, + 278 + ], + "type": "image", + "image_path": "8c84ab4284bfe3a986e9a31ab134b8832a8e42e81327e4ad39cd17f17622c302.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 281, + 200, + 290 + ], + "lines": [ + { + "bbox": [ + 134, + 281, + 200, + 290 + ], + "spans": [ + { + "bbox": [ + 134, + 281, + 200, + 290 + ], + "type": "text", + "content": "(c) Dynamic ambiguity" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 295, + 287, + 404 + ], + "lines": [ + { + "bbox": [ + 46, + 295, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 287, + 404 + ], + "type": "text", + "content": "Figure 2. Illustration of three types of ambiguities in visual SLAM. We show SLAM reconstruction results from DROID-SLAM [54]. (a) Depth ambiguity occurs when there are only minor camera translations between different views. This can lead to geometric failures in reconstruction such as the folded back corridor in the side view. (b) Scale ambiguity is inherent in monocular SLAM systems and requires additional reference for disambiguation. (c) Dynamic ambiguity gets pronounced when moving foregrounds dominate frames. Over-reliance on foreground key points will result in incorrect camera trajectories." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "type": "text", + "content": "falls short in capturing accurate depths on views with small camera translations but more crucially, only estimates scene structures and camera trajectories up to scale. The human motion also breaks the static key point assumption in the bundle adjustment. As a result, one needs additional reference to disambiguate the depth, the scale, and the dynamic as illustrated in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 498, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 592 + ], + "type": "text", + "content": "To leverage SLAM results in HMR pipelines, current world-frame HMR methods often refine camera poses by integrating either partial camera parameters, such as a global scale of the translation [62], or full extrinsic matrices [15, 30, 46] in an optimization-based framework. However, their optimization-based nature leads to complex multi-stage schemes, making the overall pipeline unnecessarily slow and easy to break." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "In this work, we explore a fundamentally different way to marry the best of HMR and SLAM. A 2D object can first be lifted from the image plane to the camera frame and then transformed into a common 3D space. This two-step process coincides with the combination of camera-frame HMR, which brings imaged 2D humans to 3D camera frames, and SLAM, which estimates the camera-to-world transformation. Noticing these correspondences, we leverage camera-frame HMR as a strong prior to bridge from the image plane to the camera frame for disambiguating SLAM, and utilize" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "SLAM reconstructions to constrain the transformation of human meshes from individual camera frames to a common global space. The overall pipeline thus results in a better synergy of the two, which we dub Synergistic Camera and Human Reconstruction (SynCHMR)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 134, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 134, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 134, + 545, + 337 + ], + "type": "text", + "content": "We design SynCHMR based on several insights. First, despite camera-frame HMR methods cannot reconstruct humans in a coherent global frame, the estimated body dimension and location still provide cues to disambiguate SLAM. Unlike SLAHMR [62] which applies SLAM out of the box and corrects the scale afterward, we endow the SLAM process with human meshes from camera-frame HMR to address ambiguities. To this end, we capitalize on estimated absolute depths to provide pseudo-RGB-D inputs for SLAM [54] and confine the bundle adjustment to static backgrounds. Since current depth estimation methods [2, 43] predicts either relative depth maps or depths with data biases, we propose to calibrate their outputs by aligning with estimated human bodies in the camera frame [9]. With these priors, SLAM knows the depth, scale, and dynamic information from HMR and consequently estimates less ambiguous scene structures and camera poses." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 339, + 546, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 546, + 578 + ], + "type": "text", + "content": "Next, we place human meshes in the common coordinate recovered by SLAM. The gap between human tracks transformed from camera frames and their real plausible world-frame motions stems from two sources of error: noise induced by camera-frame HMR and by SLAM. Motion prior models [14, 44, 68] can be used for denoising purposes as they contribute to the temporal coherence of worldframe human tracks. However, their exclusive focus on human modeling either leaves them agnostic to the underlying scenes [14] or assumes the scene is a simple ground plane [44, 68]. Our intuition is that when placing a human, static elements of the scene, such as the ground, and dynamic components like moving objects are both possible to be in contact with the human, thereby providing clues for placing the body coherently and compatibly with the scene. We hence introduce a Scene-aware SMPL Denoiser that learns to denoise the transformed human tracks by considering both temporal consistencies of moving humans and implicit constraints from dynamic scenes. This global awareness makes it more flexible for in-the-wild videos." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 580, + 514, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 514, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 514, + 592 + ], + "type": "text", + "content": "Our contributions can be summed up as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 594, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 306, + 594, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 594, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 306, + 594, + 545, + 666 + ], + "type": "text", + "content": "- We present a novel pipeline, SynCHMR, that takes a monocular video as input and reconstructs human motions, camera trajectory and dense scene point clouds all in one global coordinate, as shown in Fig. 1, whereas current world-frame HMR methods [30, 62, 65] can recover only an estimated or pre-defined ground plane." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 666, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 666, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 666, + 545, + 712 + ], + "type": "text", + "content": "- We propose a novel Human-aware Metric SLAM process to robustly calibrate estimated depth with estimated human meshes, resulting in metric-scale camera pose estimation and metric-scale scene reconstruction." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1217" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 156 + ], + "type": "text", + "content": "- We present Scene-aware SMPL Denoising that enforces spatiotemporal coherencies and applies dynamic scene constraints on world-frame human meshes. Notably, this is achieved without requiring extra annotations or heuristic designs to decide which part of a human should be interacting with the scene [49] and which region in the scene is most likely to be in contact with humans [38, 62]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 170, + 134, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 170, + 134, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 170, + 134, + 182 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 191, + 286, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 191, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 191, + 286, + 228 + ], + "type": "text", + "content": "There is considerable prior arts of HMR. We briefly discuss how they adopt different camera models and refer the readers to [1] for a more comprehensive review." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 231, + 286, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 286, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 286, + 386 + ], + "type": "text", + "content": "HMR from a single image. State-of-the-art (SOTA) methods use parametric body models [21, 37, 41, 59] and estimate the parameters either by fitting to detected image features [3, 41, 58] or by regressing directly from pixels with deep neural networks [9, 11, 19, 22, 23, 28, 31, 32, 35, 45, 50, 60, 66, 67]. These approaches assume weak perspective/orthographic projection or pre-define the focal length as a large constant for all images. Kissos et al. [26] show that replacing focal length with a constant closer to ground truth alleviate the body tilting problem. SPEC [29] and Zolly [57] estimate focal length to account for perspective distortion. CIUFF [34] takes into account the location of humans in images to regress better poses in the camera coordinates." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 388, + 286, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 286, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 286, + 458 + ], + "type": "text", + "content": "Many of these camera-frame HMR methods assume zero camera rotation, which entangles body rotation and camera rotation. When applied on video data, they fail to reconstruct humans in a coherent global space since they operate in a per-frame manner and hence cannot reason about how the camera moves across frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 461, + 286, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 286, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 286, + 591 + ], + "type": "text", + "content": "HMR from videos aims to regress a series of body parameters from a temporal sequence. It opens up new problems such as whether the reconstructed bodies are in a common global coordinate or not. Some temporal methods consider a static camera [38, 44, 68], which makes the camera space a natural choice of the common coordinate. The challenge of coherent global space emerges when the camera moves. Early methods [5, 24, 27] show promising results on videos of dynamic cameras. Despite the reconstructed human meshes look great when overlaid on images, they do not share a common coordinate in 3D." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "Recent HMR methods capitalize on human motion prior to constrain the global trajectories in the world space, which in turn implicitly disentangles human movement from camera movement. GLAMR [65] consider a data-driven prior models learned on large-scale MoCap database e.g. AMASS [39], while D&D [33] and Yu [64] consider physic-inspired prior. These world-frame HMR methods often struggle on noise in local poses caused by partial occlusions, which is very common in in-the-wild videos with close-up shots and crowd scenes. Kaufmann et al. [25] and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": "BodySLAM++ [16] circumvent this problem by employing IMU sensors to provide more robust body estimates but require extra sensory devices. To fully disentangle human and camera motion, another line of work [15, 30, 36, 46] leverages state-of-the-art SLAM techniques, e.g. [47, 54, 70], to explicitly estimate camera motion from the input video and infer the body parameters in the world coordinate of SLAM. Closest to us is SLAHMR [62] which solves for a global scale to connect the pre-computed SLAM results and body trajectories. To carefully guide the optimization process, these methods tend to have complex, multi-stage optimization schemes, making the overall pipeline easy to break and unnecessarily slow." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 228, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 300 + ], + "type": "text", + "content": "Note that in stark contrast to the methods above, which either assume or estimate a simple ground plane as scene representation, SynCHMR reconstructs dense scenes from in-the-wild videos without pre-scanning with extra devices a priori like in [6, 7, 12, 13, 17, 61, 69]. We provide detailed comparisons with these world-frame HMR in Supp. Mat." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 309, + 361, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 309, + 361, + 320 + ], + "spans": [ + { + "bbox": [ + 306, + 309, + 361, + 320 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "text", + "content": "Taking as input an RGB video " + }, + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_t\\in \\mathbb{R}^{H\\times W\\times 3}\\}_{t = 1}^T" + }, + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "text", + "content": " frames and " + }, + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "text", + "content": " people in the scene, we aim to recover human meshes " + }, + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{V}_{nt}^{\\mathrm{w}}\\in \\mathbb{R}^{3\\times 6890}\\}_{n = 1,t = 1}^{N,T}" + }, + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "text", + "content": ", dynamic scene point clouds " + }, + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_t^{\\mathrm{wm}}\\in \\mathbb{R}^{H\\times W\\times 3}\\}_{t = 1}^T" + }, + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "text", + "content": ", and corresponding camera poses " + }, + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{G}_t^{\\mathrm{m}}\\in \\mathrm{SE}(3)\\}_{t = 1}^T" + }, + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "text", + "content": " in a common world coordinate system. The superscripts " + }, + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "inline_equation", + "content": "{}^{\\mathrm{W}},{}^{\\mathrm{C}}" + }, + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "inline_equation", + "content": "{}^{\\mathrm{m}}" + }, + { + "bbox": [ + 304, + 328, + 545, + 522 + ], + "type": "text", + "content": " denote the world frame, the camera frame, and the metric scale, respectively. To this aim, we propose a two-phase alternative conditioning pipeline as depicted in Fig. 3. In the first phase, we calibrate camera motion by injecting a camera-frame human prior to SLAM. This resolves depth, scale, and dynamic ambiguities, yielding metric-scale camera poses and dynamic point clouds. Subsequently, in the second phase, we transform the camera-frame human tracks into the world frame and utilize the dynamic point clouds obtained in the first phase for conditional denoising." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 529, + 391, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 391, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 391, + 540 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 548, + 368, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 368, + 558 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 368, + 558 + ], + "type": "text", + "content": "3.1.1 SLAM" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 565, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 639 + ], + "type": "text", + "content": "Given a monocular RGB video " + }, + { + "bbox": [ + 304, + 565, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_t\\}_{t = 1}^T" + }, + { + "bbox": [ + 304, + 565, + 545, + 639 + ], + "type": "text", + "content": " DROID-SLAM [54] solves a dense bundle adjustment for a set of camera poses " + }, + { + "bbox": [ + 304, + 565, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{G}_t\\in \\mathrm{SE}(3)\\}_{t = 1}^T" + }, + { + "bbox": [ + 304, + 565, + 545, + 639 + ], + "type": "text", + "content": " and inverse depths " + }, + { + "bbox": [ + 304, + 565, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{d}_t\\in \\mathbb{R}_+^{H\\times W}\\}_{t = 1}^T" + }, + { + "bbox": [ + 304, + 565, + 545, + 639 + ], + "type": "text", + "content": " . To update these estimations, it first computes a dense correspondence field " + }, + { + "bbox": [ + 304, + 565, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{ij}\\in \\mathbb{R}^{H\\times W\\times 2}" + }, + { + "bbox": [ + 304, + 565, + 545, + 639 + ], + "type": "text", + "content": " based on reprojection for each pair of frames " + }, + { + "bbox": [ + 304, + 565, + 545, + 639 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 304, + 565, + 545, + 639 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 362, + 645, + 545, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 645, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 362, + 645, + 545, + 669 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {i j} = \\Pi \\left(\\mathbf {G} _ {i j} \\circ \\Pi^ {- 1} \\left(\\mathbf {p} _ {i}, \\frac {1}{\\mathbf {d} _ {j}}\\right)\\right), \\tag {1}", + "image_path": "022485be776aaf65772579b747803abec79173e0f2e9ce440beeebecedf97005.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i\\in \\mathbb{R}^{H\\times W\\times 2}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " is a grid of pixel coordinates in frame " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{ij} = \\mathbf{G}_j\\circ \\mathbf{G}_i^{-1}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " is the relative pose, and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Pi^{-1}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " are the camera projection and inverse projection functions." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1218" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 549, + 242 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 549, + 242 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 549, + 242 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 549, + 242 + ], + "type": "image", + "image_path": "0a5ef57b61c9a023648d8f342766ae6edf7f543a99cd003fd11a52004ae42588.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 249, + 547, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 249, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 547, + 307 + ], + "type": "text", + "content": "Figure 3. The architecture of SynCHMR. Our pipeline comprises two phases. The first phase, Human-aware Metric SLAM (Sec. 3.2), infers metric-scale camera poses and metric-scale point clouds by exploiting the camera-frame human prior. The second phase, Scene-aware SMPL Denoising (Sec. 3.3), involves the conditional denoising of world-frame noisy SMPL parameters. These parameters, initialized by transforming from the camera frame, get refined through conditioning on the dynamic point clouds obtained in the first phase. The whole pipeline thus reconstructs humans, scene point clouds, and cameras harmoniously in a common world frame." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 316, + 288, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 354 + ], + "type": "text", + "content": "Then with a learned neural network, the system predicts a revision flow field " + }, + { + "bbox": [ + 46, + 316, + 288, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{ij} \\in \\mathbb{R}^{H \\times W \\times 2}" + }, + { + "bbox": [ + 46, + 316, + 288, + 354 + ], + "type": "text", + "content": " and associated confidence map " + }, + { + "bbox": [ + 46, + 316, + 288, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{ij} \\in \\mathbb{R}_+^{H \\times W \\times 2}" + }, + { + "bbox": [ + 46, + 316, + 288, + 354 + ], + "type": "text", + "content": " to construct the cost function" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 365, + 287, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 365, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 64, + 365, + 287, + 399 + ], + "type": "interline_equation", + "content": "E _ {\\Sigma} = \\sum_ {(i, j)} \\left\\| \\mathbf {p} _ {i j} ^ {*} - \\Pi \\left(\\mathbf {G} _ {i j} ^ {\\prime} \\circ \\Pi^ {- 1} \\left(\\mathbf {p} _ {i}, \\frac {1}{\\mathbf {d} _ {i} ^ {\\prime}}\\right)\\right) \\right\\| _ {\\Sigma_ {i j}} ^ {2}, \\tag {2}", + "image_path": "8e6f4c51cc99d56153f2665bc9f1dfff78e3cd5b3fe21caec142c30f5c3bb5c9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 403, + 288, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 288, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 288, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 403, + 288, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{ij}^{*} = \\mathbf{r}_{ij} + \\mathbf{p}_{ij}" + }, + { + "bbox": [ + 46, + 403, + 288, + 489 + ], + "type": "text", + "content": " is the corrected correspondence, " + }, + { + "bbox": [ + 46, + 403, + 288, + 489 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_{\\Sigma}" + }, + { + "bbox": [ + 46, + 403, + 288, + 489 + ], + "type": "text", + "content": " is the Mahalanobis distance which weighs the error terms with " + }, + { + "bbox": [ + 46, + 403, + 288, + 489 + ], + "type": "inline_equation", + "content": "\\Sigma_{ij} = \\mathrm{diag}\\mathbf{w}_{ij}" + }, + { + "bbox": [ + 46, + 403, + 288, + 489 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 403, + 288, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{G}'" + }, + { + "bbox": [ + 46, + 403, + 288, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 403, + 288, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{d}'" + }, + { + "bbox": [ + 46, + 403, + 288, + 489 + ], + "type": "text", + "content": " are updated poses and inverse depths. Upon this objective, DROID-SLAM considers an additional term that penalizes the squared distance between the measured and predicted depth if the input is with an extra sensor depth channel " + }, + { + "bbox": [ + 46, + 403, + 288, + 489 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{D}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 46, + 403, + 288, + 489 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 496, + 105, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 496, + 105, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 496, + 105, + 509 + ], + "type": "text", + "content": "3.1.2 HMR" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "text", + "content": "We employ 4DHumans [9] for reconstructing camera-frame human meshes from an in-the-wild video. Specifically, it performs per-frame human mesh recovery with an end-to-end transformer architecture and associates them to form human tracks. Each tracked human " + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "text", + "content": " is represented by SMPL [37] parameters as " + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\{\\Phi_{nt},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}\\}" + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "text", + "content": ", including global orientation " + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\Phi_{nt}\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "text", + "content": ", body pose " + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\theta_{nt}\\in \\mathbb{R}^{22\\times 3\\times 3}" + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "text", + "content": ", shape " + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\beta_{nt}\\in \\mathbb{R}^{10}" + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "text", + "content": ", and root translation " + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\Gamma_{nt}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "text", + "content": ". Then the parametric SMPL model can use these parameters to recover a human mesh with vertices " + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{nt}\\in \\mathbb{R}^{3\\times 6890}" + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "text", + "content": " in metric scale: " + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{nt} = \\mathrm{SMPL}(\\Phi_{nt},\\theta_{nt},\\beta_{nt}) + \\mathbf{\\Gamma}_{nt}" + }, + { + "bbox": [ + 46, + 515, + 287, + 648 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 652, + 206, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 652, + 206, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 652, + 206, + 664 + ], + "type": "text", + "content": "3.2. Human-aware Metric SLAM" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 671, + 141, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 141, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 141, + 684 + ], + "type": "text", + "content": "3.2.1 Preprocessing" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "To start off, we estimate per-frame depth maps " + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{D}_t\\}" + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": " with an off-the-shelf depth estimator, ZoeDepth [2] and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 316, + 547, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 461 + ], + "type": "text", + "content": "predict per-frame human instance segmentation masks " + }, + { + "bbox": [ + 304, + 316, + 547, + 461 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_{nt}\\}" + }, + { + "bbox": [ + 304, + 316, + 547, + 461 + ], + "type": "text", + "content": " with an image instance segmentation network, Mask2Former [4]. We adapt ZoeDepth for video-consistent depth estimation by choosing a per-video metric head from the majority vote of per-frame routers, for which we dub ZoeDepth" + }, + { + "bbox": [ + 304, + 316, + 547, + 461 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 304, + 316, + 547, + 461 + ], + "type": "text", + "content": ". While ZoeDepth claims to estimate metric depths, we observe a domain gap when inference on new datasets. Consequently, we only treat its output as up-to-affine depths that need to be further aligned with the metric scale. To aid our optimization with human awareness, we use camera-frame human meshes " + }, + { + "bbox": [ + 304, + 316, + 547, + 461 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{V}_{nt}^{\\mathrm{c}}\\}" + }, + { + "bbox": [ + 304, + 316, + 547, + 461 + ], + "type": "text", + "content": " recovered by 4DHumans [9] to introduce a metric prior." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 468, + 499, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 468, + 499, + 481 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 499, + 481 + ], + "type": "text", + "content": "3.2.2 Calibrating Depth with Human Prior" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "text", + "content": "We calibrate the per-frame depths with human meshes in Human-aware Depth Calibration. This involves optimizing two parameters, a world scale " + }, + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "text", + "content": " and a world offset " + }, + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "text", + "content": ", shared across all frames. During optimization, we linearly transform " + }, + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_t" + }, + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_t' = s\\mathbf{D}_t + o" + }, + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "text", + "content": " and unproject these depth maps to camera-frame point clouds " + }, + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_t^c\\}" + }, + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_t^c = \\Pi^{-1}(\\mathbf{p}_t, \\mathbf{D}_t')" + }, + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "text", + "content": ". Our intuition is to align the human point cloud " + }, + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{nt}^c = \\mathbf{M}_{nt} \\odot \\mathbf{P}_t^c" + }, + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "text", + "content": " with the camera-frame human mesh vertices " + }, + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{nt}^c" + }, + { + "bbox": [ + 304, + 487, + 547, + 632 + ], + "type": "text", + "content": " in terms of absolute depth and size. To achieve pixel-wise alignment, we use a depth term to pull points on the human point cloud toward their corresponding human mesh vertices along the z-axis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 332, + 638, + 545, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 638, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 332, + 638, + 545, + 669 + ], + "type": "interline_equation", + "content": "E _ {\\text {d e p t h}} = \\frac {\\sum_ {n , t} \\| \\mathbf {S} _ {n t} \\odot [ z \\left(\\mathbf {V} _ {n t} ^ {\\mathrm {c}}\\right) - z \\left(\\mathbf {P} _ {n t} ^ {\\mathrm {c}}\\right) ] \\| _ {2} ^ {2}}{\\sum_ {n , t} \\| \\mathbf {S} _ {n t} \\| _ {0}}, \\tag {3}", + "image_path": "d8071436746e168bdb7e3b821de72950a262875cc52c26266fe919f35f862b74.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{nt} = \\rho (\\mathbf{V}_{nt}^{\\mathrm{c}})\\cap \\mathbf{M}_{nt}" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " is the intersection of the rasterized human mesh mask " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\rho (\\mathbf{V}_{nt}^{\\mathrm{c}})" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " and the instance segmentation mask " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{nt}" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "z(\\cdot)" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " is the rasterized depth, and " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _0" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " is the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1219" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 289, + 183 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 289, + 183 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 289, + 183 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 289, + 183 + ], + "type": "image", + "image_path": "441dac30c7eb42a434600800fbdd161d93bfc451e56c5004a14ff0e74d4c3252.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "lines": [ + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "text", + "content": "Figure 4. The architecture of Scene-aware SMPL Denoiser. World-frame noisy SMPL parameters " + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\{\\Phi_{nt}^{\\mathrm{w}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{w}}\\}_{0}" + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "text", + "content": " are first projected by a linear layer and summed with temporal positional embeddings (TPE) to get initial latent humans " + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_{nt,0}^{\\mathrm{SMPL}}\\}" + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "text", + "content": ". Per-frame point clouds are aggregated to " + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{scene}}" + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "text", + "content": " and encoded with the point encoder " + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "text", + "content": ". Then we query the encoded scene " + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\mathbf{x}^{\\mathrm{scene}})" + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "text", + "content": " with latent humans " + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_{nt,0}^{\\mathrm{SMPL}}\\}" + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "text", + "content": " in the scene-conditioned denoiser " + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "text", + "content": " and feed the result " + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_{nt,1}^{\\mathrm{SMPL}}\\}" + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "text", + "content": " to prediction heads " + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{P}_{\\Phi},\\mathcal{P}_{\\theta},\\mathcal{P}_{\\beta},\\mathcal{P}_{\\Gamma}\\}" + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "text", + "content": " to obtain denoised SMPL parameters " + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\{\\Phi_{nt}^{\\mathrm{w}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{w}}\\}_{1}" + }, + { + "bbox": [ + 47, + 190, + 287, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 299, + 286, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 299, + 286, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 299, + 286, + 310 + ], + "type": "text", + "content": "0-norm indicating the number of non-zero pixels on a mask." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 311, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 287, + 346 + ], + "type": "text", + "content": "As the recovered human meshes can be noisy in depth but still have a stable body dimension, we also adopt a size term to leverage the relative position of mesh vertices" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 354, + 287, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 354, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 61, + 354, + 287, + 381 + ], + "type": "interline_equation", + "content": "E _ {\\mathrm {d x}} = \\frac {\\sum_ {n , t} \\| \\Delta_ {x} \\left(\\mathbf {V} _ {n t} ^ {\\mathrm {c}} , \\mathbf {S} _ {n t}\\right) - \\Delta_ {x} \\left(\\mathbf {P} _ {n t} ^ {\\mathrm {c}} , \\mathbf {S} _ {n t}\\right) \\| _ {2} ^ {2}}{N T}. \\tag {4}", + "image_path": "2d9979fcae6a2d7b2ad89342f9ec666565bd3d7b4280b5fd553926ab8eb9980c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 387, + 204, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 204, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 204, + 399 + ], + "type": "text", + "content": "We define " + }, + { + "bbox": [ + 47, + 387, + 204, + 399 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{dy}}" + }, + { + "bbox": [ + 47, + 387, + 204, + 399 + ], + "type": "text", + "content": " similarly as " + }, + { + "bbox": [ + 47, + 387, + 204, + 399 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{dx}}" + }, + { + "bbox": [ + 47, + 387, + 204, + 399 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 406, + 287, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 406, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 52, + 406, + 287, + 423 + ], + "type": "interline_equation", + "content": "\\Delta_ {*} (\\mathbf {X}, \\mathbf {Y}) = \\left(\\max _ {*} - \\min _ {*} \\right) \\left[ \\Pi^ {- 1} (\\mathbf {Y} \\odot \\Pi (\\mathbf {X}), z (\\mathbf {X})) \\right] \\tag {5}", + "image_path": "75885f36c79454e6afef5ce5aaef70cee35505371f16491c561434d525f7b875.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 432, + 287, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 287, + 454 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 47, + 432, + 287, + 454 + ], + "type": "inline_equation", + "content": "(\\max_{*} - \\min_{*})" + }, + { + "bbox": [ + 47, + 432, + 287, + 454 + ], + "type": "text", + "content": " denotes the difference between the maximum value and the minimum value on coordinate " + }, + { + "bbox": [ + 47, + 432, + 287, + 454 + ], + "type": "inline_equation", + "content": "^*" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 456, + 272, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 456, + 272, + 467 + ], + "spans": [ + { + "bbox": [ + 59, + 456, + 272, + 467 + ], + "type": "text", + "content": "Then we have the calibrated depths with optimization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 85, + 475, + 287, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 475, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 85, + 475, + 287, + 489 + ], + "type": "interline_equation", + "content": "\\left(s ^ {\\mathrm {m}}, o ^ {\\mathrm {m}}\\right) = \\operatorname {a r g m i n} _ {s, o} \\left(E _ {\\text {d e p t h}} + \\lambda E _ {\\text {s i z e}}\\right), \\tag {6}", + "image_path": "39b45cb0f98ead1c5272941a9fecedd1ff830ae47bb334e0decea8fda6ae54cd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 491, + 287, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 491, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 287, + 505 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {t} ^ {\\mathrm {m}} = s ^ {\\mathrm {m}} \\mathbf {D} _ {t} + o ^ {\\mathrm {m}}, \\tag {7}", + "image_path": "307f1614b83296b877e167a238f1426787f177f7c0503f3361b1d539c7f6a1e1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 513, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 287, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 513, + 287, + 536 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{size}} = E_{\\mathrm{dx}} + E_{\\mathrm{dy}}" + }, + { + "bbox": [ + 47, + 513, + 287, + 536 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 513, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 513, + 287, + 536 + ], + "type": "text", + "content": " is a hyperparameter to balance two energy terms with a default value of 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 551, + 277, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 277, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 277, + 563 + ], + "type": "text", + "content": "3.2.3 Disambiguating SLAM with Calibrated Depth" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 569, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 689 + ], + "type": "text", + "content": "While DROID-SLAM [54] originally supports RGB-D input mode where the D channel stands for sensor depth, one cannot trivially access sensor depths from in-the-wild videos. Our insight is that an estimated absolute depth can be utilized as a depth prior, albeit noisy. So we combine the original RGB video and the calibrated depth as pseudoRGB-D inputs " + }, + { + "bbox": [ + 46, + 569, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_t,\\mathbf{D}_t^{\\mathrm{m}}\\}" + }, + { + "bbox": [ + 46, + 569, + 287, + 689 + ], + "type": "text", + "content": " to disambiguate depth and scale. Furthermore, we modify the cost function Eq. (2) to resolve the dynamic ambiguity by masking out dynamic foregrounds in confidence maps" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 696, + 287, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 696, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 696, + 287, + 712 + ], + "type": "interline_equation", + "content": "\\Sigma_ {i j} ^ {\\prime} = \\operatorname {d i a g} \\mathbf {w} _ {i j} ^ {\\prime} = \\operatorname {d i a g} \\left(\\left(1 - \\left[ \\mathbf {M} _ {i}, \\mathbf {M} _ {j} \\right]\\right) \\odot \\mathbf {w} _ {i j}\\right), \\tag {8}", + "image_path": "47a143ded0e4bbacf7a79ce14a7966fc149ecedeb5342f83554982863a497006.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i = \\bigcup_n\\mathbf{M}_{ni}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_j = \\bigcup_n\\mathbf{M}_{nj}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " are the union of all human instance masks on their corresponding frame, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "[\\cdot ,\\cdot ]" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " is the concatenation operation. As a result, we obtain metric-scale camera poses " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{G}_t^{\\mathrm{m}}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " and metric-scale point clouds " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_t^{\\mathrm{wm}}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " by disambiguating SLAM with calibrated metric depths" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 353, + 154, + 545, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 154, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 353, + 154, + 545, + 171 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathbf {G} _ {t} ^ {\\mathrm {m}}, \\mathbf {d} _ {t} ^ {\\mathrm {m}} \\right\\} = \\operatorname {a r g m i n} _ {\\left\\{\\mathbf {G} _ {i j} ^ {\\prime}, \\mathbf {d} _ {i j} ^ {\\prime} \\right\\}} E _ {\\Sigma^ {\\prime}}, \\tag {9}", + "image_path": "4fd7e417810cc1cdc6315c7eba43d63e5222fde646862fbe0b2c67755f2d2ffe.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 376, + 172, + 545, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 172, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 376, + 172, + 545, + 196 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {t} ^ {\\mathrm {w m}} = \\mathbf {G} _ {t} ^ {\\mathrm {m}} \\circ \\Pi^ {- 1} \\left(\\mathbf {p} _ {t}, \\frac {1}{\\mathbf {d} _ {t} ^ {\\mathrm {m}}}\\right)). \\tag {10}", + "image_path": "97436c9815446e041b02c74d058794cc1f868d50955036c6d557e184d1114903.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 206, + 470, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 206, + 470, + 219 + ], + "spans": [ + { + "bbox": [ + 306, + 206, + 470, + 219 + ], + "type": "text", + "content": "3.3. Scene-aware SMPL Denoising" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 225, + 517, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 225, + 517, + 237 + ], + "spans": [ + { + "bbox": [ + 306, + 225, + 517, + 237 + ], + "type": "text", + "content": "3.3.1 Initializing Humans with Metric Cameras" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 245, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 545, + 304 + ], + "type": "text", + "content": "To put humans properly in the scene recovered by SLAM, we initialize them by transforming estimated camera-frame SMPL parameters " + }, + { + "bbox": [ + 304, + 245, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\{\\Phi_{nt}^{\\mathrm{c}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{c}}\\}" + }, + { + "bbox": [ + 304, + 245, + 545, + 304 + ], + "type": "text", + "content": " to the world frame with camera-to-world transforms " + }, + { + "bbox": [ + 304, + 245, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{G}_t^{\\mathrm{m}} = [\\mathbf{R}_t|\\mathbf{t}_t^{\\mathrm{m}}]\\}" + }, + { + "bbox": [ + 304, + 245, + 545, + 304 + ], + "type": "text", + "content": ". Given the pelvis as the center of global orientation " + }, + { + "bbox": [ + 304, + 245, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 245, + 545, + 304 + ], + "type": "text", + "content": ", we have:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 327, + 545, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 545, + 340 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Phi} _ {n t} ^ {\\mathrm {w}} = \\mathbf {R} _ {t} \\boldsymbol {\\Phi} _ {n t} ^ {\\mathrm {c}}, \\quad \\boldsymbol {\\Gamma} _ {n t} ^ {\\mathrm {w}} = \\mathbf {R} _ {t} \\left(\\boldsymbol {\\Gamma} _ {n t} ^ {\\mathrm {c}} + \\mathbf {c}\\right) + \\mathbf {t} _ {t} ^ {\\mathrm {m}} - \\mathbf {c}, \\tag {11}", + "image_path": "c01905bbef0061c9c445356e77400aa995a9f7a8b35f6154ecb64cba270f44a2.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 350, + 545, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 435 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 350, + 545, + 435 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = \\mathbf{c}(\\beta_{nt})" + }, + { + "bbox": [ + 304, + 350, + 545, + 435 + ], + "type": "text", + "content": " is the pelvis location in the shape blend body mesh. Note that we do not need to introduce an extra camera scale as SLAHMR [62] since the camera poses have already been in the metric scale. The root-relative poses " + }, + { + "bbox": [ + 304, + 350, + 545, + 435 + ], + "type": "inline_equation", + "content": "\\theta_{nt}" + }, + { + "bbox": [ + 304, + 350, + 545, + 435 + ], + "type": "text", + "content": " and the shapes " + }, + { + "bbox": [ + 304, + 350, + 545, + 435 + ], + "type": "inline_equation", + "content": "\\beta_{nt}" + }, + { + "bbox": [ + 304, + 350, + 545, + 435 + ], + "type": "text", + "content": " stay unchanged as in the camera frame. We denote the initialized and the denoised parameters with a suffix 0 and 1 respectively, i.e. " + }, + { + "bbox": [ + 304, + 350, + 545, + 435 + ], + "type": "inline_equation", + "content": "\\{\\Phi_{nt}^{\\mathrm{w}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{w}}\\}_{0,1}" + }, + { + "bbox": [ + 304, + 350, + 545, + 435 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 451, + 526, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 451, + 526, + 463 + ], + "spans": [ + { + "bbox": [ + 306, + 451, + 526, + 463 + ], + "type": "text", + "content": "3.3.2 Constraining Humans with Dynamic Scenes" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 470, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 545, + 555 + ], + "type": "text", + "content": "Different from existing works [30, 36, 62, 64] that incorporate energy terms in optimization to apply explicit scene constraints, we propose to learn implicit scene constraints with a Scene-aware SMPL denoiser shown in Fig. 4. The noisy initial SMPL parameters " + }, + { + "bbox": [ + 304, + 470, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\{\\Phi_{nt}^{\\mathrm{w}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{w}}\\}_{0}" + }, + { + "bbox": [ + 304, + 470, + 545, + 555 + ], + "type": "text", + "content": " are first projected to a latent space, where it gets further updated by conditioning on implicit scene constraints" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 318, + 563, + 545, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 563, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 318, + 563, + 545, + 578 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {n t, 0} ^ {\\mathrm {S M P L}} = \\operatorname {F C} \\left(\\left[ \\boldsymbol {\\Phi} _ {n t, 0} ^ {\\mathrm {w}}, \\boldsymbol {\\theta} _ {n t, 0}, \\boldsymbol {\\beta} _ {n t, 0}, \\boldsymbol {\\Gamma} _ {n t, 0} ^ {\\mathrm {w}} \\right]\\right) + \\text {T P E}, \\tag {12}", + "image_path": "f1a6894fb8b33e18d6fb2524970725965c3894b8a8b0b0025725b3ad0b2ac851.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 580, + 545, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 580, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 315, + 580, + 545, + 594 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathbf {z} _ {n t, 1} ^ {\\text {S M P L}} \\right\\} _ {t = 1} ^ {T} = \\mathcal {D} \\left(\\left\\{\\mathbf {z} _ {n t, 0} ^ {\\text {S M P L}} \\right\\} _ {t = 1} ^ {T}, \\mathcal {E} \\left(\\mathbf {x} ^ {\\text {s c e n e}}\\right) + \\text {T P E}\\right), \\tag {13}", + "image_path": "cc8048a0479d647831dc1d429f54fd57280e0f952828d15631cd36b0e419aa51.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "text", + "content": "where FC is a shared linear layer, TPE is shared temporal positional embeddings, " + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_{nt,*}^{\\mathrm{SMPL}}\\}_{t = 1}^{T}\\in \\mathbb{R}^{T\\times D}" + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "text", + "content": " dimensional latent for human " + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{scene}}\\in \\mathbb{R}^{L\\times C}" + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "text", + "content": " -channel dynamic scene point clouds with a total number of points " + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "text", + "content": " refer to the scene encoder and the scene-conditioned denoiser, respectively. We set " + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "inline_equation", + "content": "C = 7" + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "text", + "content": " which is the concatenation of point coordinates " + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_t^{\\mathrm{wm}}\\}" + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "text", + "content": ", colors " + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_t\\}" + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "text", + "content": ", and estimated human semantic segmentation masks " + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_t = \\bigcup_n\\mathbf{M}_{nt}\\}" + }, + { + "bbox": [ + 304, + 604, + 545, + 715 + ], + "type": "text", + "content": ". Following [9], the updated latent" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1220" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "are decoded with different prediction head " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{(\\cdot)}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " to regress the residual for each SMPL parameter:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 104, + 286, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 104, + 286, + 119 + ], + "spans": [ + { + "bbox": [ + 107, + 104, + 286, + 119 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Phi} _ {n t, 1} ^ {\\mathrm {w}} = \\mathcal {P} _ {\\boldsymbol {\\Phi}} \\left(\\mathbf {z} _ {n t, 1} ^ {\\mathrm {S M P L}}\\right) \\boldsymbol {\\Phi} _ {n t, 0} ^ {\\mathrm {w}}, \\tag {14}", + "image_path": "a2f981f2c63f4f5b62ece8f8522e564e86bba89d02c02a72fe5bbe2f73e79afe.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 110, + 121, + 286, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 121, + 286, + 135 + ], + "spans": [ + { + "bbox": [ + 110, + 121, + 286, + 135 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} _ {n t, 1} = \\mathcal {P} _ {\\boldsymbol {\\theta}} \\left(\\mathbf {z} _ {n t, 1} ^ {\\mathrm {S M P L}}\\right) \\boldsymbol {\\theta} _ {n t, 0}, \\tag {15}", + "image_path": "20192aff172fcb341dcd52babb06bf00e763706ffccaf11718fa3619d56883ce.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 137, + 286, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 137, + 286, + 152 + ], + "spans": [ + { + "bbox": [ + 109, + 137, + 286, + 152 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\beta} _ {n t, 1} = \\mathcal {P} _ {\\boldsymbol {\\beta}} \\left(\\mathbf {z} _ {n t, 1} ^ {\\mathrm {S M P L}}\\right) + \\boldsymbol {\\beta} _ {n t, 0}, \\tag {16}", + "image_path": "08368cbc712005686e678c8d16a9e0994b9abbac1105b07f1f214d6c9d8db6a8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 154, + 286, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 154, + 286, + 169 + ], + "spans": [ + { + "bbox": [ + 108, + 154, + 286, + 169 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Gamma} _ {n t, 1} ^ {\\mathrm {w}} = \\mathcal {P} _ {\\boldsymbol {\\Gamma}} \\left(\\mathbf {z} _ {n t, 1} ^ {\\mathrm {S M P L}}\\right) + \\boldsymbol {\\Gamma} _ {n t, 0} ^ {\\mathrm {w}}. \\tag {17}", + "image_path": "dd7cc59d4fcb6d527d3836e25e4d139e58d5eba992c7207c63c51adc9c015f40.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 178, + 286, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 178, + 286, + 214 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 286, + 214 + ], + "type": "text", + "content": "We apply direct supervision on " + }, + { + "bbox": [ + 47, + 178, + 286, + 214 + ], + "type": "inline_equation", + "content": "\\{\\Phi_{nt}^{\\mathrm{w}},\\theta_{nt},\\beta_{nt},\\Gamma_{nt}^{\\mathrm{w}}\\}_{1}" + }, + { + "bbox": [ + 47, + 178, + 286, + 214 + ], + "type": "text", + "content": " which is common in the literature. Please see Supp. Mat. for the details of the full training objectives." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 224, + 128, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 224, + 128, + 238 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 128, + 238 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 243, + 170, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 243, + 170, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 170, + 257 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 286, + 361 + ], + "type": "text", + "content": "Datasets. We assess the performance of SynCHMR primarily for global human motion estimation but also report the accuracy of estimated camera trajectories. Traditional video datasets in HMR literature are typically captured by static cameras, e.g. [13, 18, 20, 40, 63], hence not suitable for our purpose. Standard SLAM benchmarks such as [48, 51] do not meet our needs either as there is often no human moving in the scene. We consider the following datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 361, + 286, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 286, + 491 + ], + "type": "text", + "content": "3DPW [56] is an in-the-wild dataset captured with iPhones. The ground truth bodies are not in coherent world frames so we use it to supervise root relative poses and for evaluation. EgoBody [69] has ground-truth poses captured by multiple Kinects and egocentric-view sequences recorded by a head-mounted device, whose trajectories are further registered in the world space of Kinect array. We use it for training the SMPL denoiser in Sec. 3.3 and for evaluation (on both body and camera estimation). For HMR evaluation, unlike [30, 62] considering only the validation set, we additionally report results on its completely withheld test set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 492, + 286, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 286, + 540 + ], + "type": "text", + "content": "EMDB [25] is a new dataset providing SMPL poses from IMU sensors and global camera trajectories. We include it for training the SMPL denoiser to enrich the diversity and use the camera trajectories to evaluate the quality of SLAM." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 543, + 286, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 286, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 286, + 686 + ], + "type": "text", + "content": "Evaluation Metrics. For HMR evaluation, we report common PA-MPJPE, which measures the quality of rootrelative poses. For datasets that have ground-truth poses in a world coordinate, we follow [62] and consider WA-MPJPE and FA-MPJPE. The former measures the error after aligning the entire trajectories of the prediction and ground truth with Procrustes Alignment [10], while the latter aligns only with the first frame. We also report acceleration errors. For SLAM, we consider absolute trajectory error (ATE) for camera trajectory evaluation as well as the threshold accuracy " + }, + { + "bbox": [ + 46, + 543, + 286, + 686 + ], + "type": "inline_equation", + "content": "(\\delta_{n})" + }, + { + "bbox": [ + 46, + 543, + 286, + 686 + ], + "type": "text", + "content": " , the absolute relative error (REL), and the root mean squared error (RMSE) for scene depth evaluation [2]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "Implementation Details. In Human-aware Depth Calibration, we use the L-BFGS algorithm with learning rate 1 to" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 545, + 125 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 545, + 125 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 125 + ], + "type": "table", + "html": "
Camera ModelHuman ModelPA ↓
DROID-SLAM [54]SLAHMR [62] w/ PHALP+55.9
DROID-SLAM [54]SLAHMR [62] w/ 4DHumans [9]57.4
Human-aware Metric SLAM (ours)4DHumans [9]52.9
Human-aware Metric SLAM (ours)Scene-aware SMPL Denoiser (ours)52.4
", + "image_path": "2fd701643bc4a5d7d5b61166d002d5916d09324bb3126e627c16677311f60848.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 133, + 545, + 166 + ], + "lines": [ + { + "bbox": [ + 305, + 133, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 545, + 166 + ], + "type": "text", + "content": "Table 1. Comparison results on 3DPW-Test. The row in gray is the full pipeline of SynCHMR. We abbreviate PA-MPJPE as PA, with the same below for FA-MPJPE (FA) and WA-MPJPE (WA)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "text", + "content": "optimize for a maximum of 30 iterations. As for the Scene-aware SMPL Denoiser, we train it on the union of 3DPW-Train, EgoBody-Train, and EMDB for 100k steps with an AdamW optimizer, a batch size of 16, and a learning rate of 1e-5. For camera-frame SMPL ground truths like in 3DPW, we only incorporate body shapes " + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "text", + "content": " and poses " + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "text", + "content": " in training. We train the denoising process by randomly sampling a temporal window size " + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "text", + "content": " spanning 64 to 128 and inference with " + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "inline_equation", + "content": "T = 100" + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "text", + "content": ". The scene-conditioned denoiser " + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "text", + "content": " is parameterized with a 6-layer Transformer Decoder. For the scene encoder " + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "text", + "content": ", we consider ViT and SPVCNN in Tab. 4 and report results for SPVCNN in Tabs. 1 and 2. Before inputting the world-frame noisy SMPL parameters to the denoiser, we first interpolate " + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\Phi_{nt,0}^{\\mathrm{w}}" + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\theta_{nt,0}" + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "text", + "content": " on SO(3), " + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\beta_{nt,0}" + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{10}" + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\Gamma_{nt,0}^{\\mathrm{w}}" + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 181, + 545, + 361 + ], + "type": "text", + "content": " when there are missing observations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 371, + 423, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 423, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 423, + 383 + ], + "type": "text", + "content": "4.2. Comparison Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 389, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 497 + ], + "type": "text", + "content": "We first evaluate the estimated local poses with PA-MPJPE on 3DPW, which is common in the literature. In Tab. 1, we show that placing the bodies from 4DHumans already leads to lower error than SLAHMR. Passing them through the denoiser further reduces the error. We note that PA-MPJPE only measures local pose accuracy not the quality of global trajectories. Since 3DPW does not support any world metrics, Tab. 1 only aims to show that SynCHMR produces reasonable local poses on a common dataset." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Next, we assess the quality of global motion estimation, which is essentially a more challenging task. Tab. 2 shows the results on EgoBody. Note that current optimization-based methods [30, 62] report the error of the validation set. For fairness and completeness, we report results on both validation and test sets and run state-of-the-art methods on the test set when the code is available. In Tab. 2, we see that the proposed SynCHMR has the overall lowest PA-MPJPE, FA-MPJPE, and WA-MPJPE (gray rows). Comparing it with the row above (4DHumans) confirms the benefit of our scene-conditioned denoiser. For a fair comparison, we also initialize the global optimization of SLAHMR with 4DHumans, which is more accurate than " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{PHALP}^+" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " in SLAHMR, but we do not observe improvement. Notably, despite the concurrent work PACE [30] has a tightly integrated SLAM and body fitting objective, it still uses native DROID-SLAM to initialize the camera parameters like SLAHMR does. This is arguably sub-optimal as the initial" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1221" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 173 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 173 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 173 + ], + "type": "table", + "html": "
SubsetCamera ModelHuman ModelPA-MPJPE (mm) ↓FA-MPJPE (mm) ↓WA-MPJPE (mm) ↓Acc Error (mm/frame2) ↓Runtime/100 ims
Val-GLAMR [65]114.3416.1239.0173.54 min
DROID-SLAM [54]PACE [30]66.5147.9101.06.71 min
DROID-SLAM [54]SLAHMR [62] w/ PHALP+79.1141.1101.225.840 min
DROID-SLAM [54]SLAHMR [62] w/ 4DHumans [9]79.3273.0144.779.440 min
Human-aware Metric SLAM (ours)4DHumans [9]73.0164.4106.7127.05 min
Human-aware Metric SLAM (ours)Scene-aware SMPL Denoiser (ours)57.7115.181.164.85 min
Test-GLAMR [65]112.8351.4216.3105.94 min
DROID-SLAM [54]SLAHMR [62] w/ PHALP+63.1163.999.431.740 min
DROID-SLAM [54]SLAHMR [62] w/ 4DHumans [9]69.3185.8113.045.740 min
Human-aware Metric SLAM (ours)4DHumans [9]75.4160.0108.1138.85 min
Human-aware Metric SLAM (ours)Scene-aware SMPL Denoiser (ours)61.3122.184.669.45 min
", + "image_path": "4057fb5c19ef539edc02255b51a2763e9f7728ceca9cb83a62eeb630a31af605.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 198, + 285, + 261 + ], + "blocks": [ + { + "bbox": [ + 62, + 178, + 529, + 190 + ], + "lines": [ + { + "bbox": [ + 62, + 178, + 529, + 190 + ], + "spans": [ + { + "bbox": [ + 62, + 178, + 529, + 190 + ], + "type": "text", + "content": "Table 2. Comparison results with state-of-the-art approaches on EgoBody. The row in gray is the full pipeline of SynCHMR." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 198, + 285, + 261 + ], + "lines": [ + { + "bbox": [ + 50, + 198, + 285, + 261 + ], + "spans": [ + { + "bbox": [ + 50, + 198, + 285, + 261 + ], + "type": "table", + "html": "
RGBDepthMaskEgoBodyEMDB\nATE ↓
ATE ↓δ1 ↑REL ↓RMSE ↓
✓XX80.90.08514.5901617.361400.3
✓XMask2Former [4]81.60.0638.5301009.127385.8
✓ZoeDepth+X35.00.5620.30815.360456.8
✓ZoeDepth+Mask2Former [4]28.60.5640.30710.852389.6
✓ZoeDepth+ + Cal.Mask2Former [4]26.40.7970.27410.452107.0
", + "image_path": "94fcffe239d8adfb946a4f5da01632b9168fc5eab21d286bd2d82eead6ccdc9a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 307, + 285, + 411 + ], + "blocks": [ + { + "bbox": [ + 47, + 266, + 287, + 300 + ], + "lines": [ + { + "bbox": [ + 47, + 266, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 287, + 300 + ], + "type": "text", + "content": "Table 3. Ablation study for SLAM configurations in terms of optimized camera trajectories and scene depths. ZoeDepth+ denotes our video-adapted ZoeDepth [2]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 307, + 285, + 411 + ], + "lines": [ + { + "bbox": [ + 50, + 307, + 285, + 411 + ], + "spans": [ + { + "bbox": [ + 50, + 307, + 285, + 411 + ], + "type": "table", + "html": "
StageBackboneRGBXYZMaskPA ↓FA ↓WA ↓Acc Error ↓
Init.-XXX73.7120.893.1127.1
Pred.-XXX63.398.877.275.2
Pred.ViT [8]✓XX63.994.976.743.3
Pred.ViT [8]✓✓X64.597.377.745.6
Pred.ViT [8]✓X✓66.896.578.644.7
Pred.ViT [8]✓✓✓69.3100.982.046.4
Pred.SPVCNN [53]X✓X62.995.176.072.6
Pred.SPVCNN [53]✓✓X61.093.474.367.7
Pred.SPVCNN [53]X✓✓62.093.975.369.9
Pred.SPVCNN [53]✓✓✓61.391.973.664.8
", + "image_path": "d5f2fa350721696515b3c30675db92123bad2cc54e7e16f147e518b4e7bc40c0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 416, + 287, + 450 + ], + "lines": [ + { + "bbox": [ + 47, + 416, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 287, + 450 + ], + "type": "text", + "content": "Table 4. Ablation study for different scene encoders and features regarding world-frame HMR. Init. and Pred. refer to before and after SMPL denoising, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "type": "text", + "content": "ization is not aware of body information, which can lead to errors that cannot be corrected in the global optimization stage. Consequently, it also has higher world-space errors. Optimization methods often employ a zero velocity term to smooth out human motion, which explains the lower acceleration error. However, we do not observe a big difference in jittery between our results. Please refer to Supp. Mat. for more details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 563, + 141, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 141, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 141, + 576 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "We ablate the design choices in SynCHMR. In Tab. 3, we evaluate SLAM-optimized camera trajectories and scene depths with EgoBody and EMDB. We see that directly including un-calibrated monocular depths does not guarantee more accurate estimations (" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "4^{\\mathrm{th}}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " row). Precluding the dynamic foreground pixels with Mask2Former [4] generally improves performance. We empirically find that our depth calibration with human prior works the best when using it with foreground masking, which has the lowest error in both datasets. More SLAM evaluation and discussion can be found in Supp. Mat." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 201, + 547, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 547, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 547, + 416 + ], + "type": "text", + "content": "In Tab. 4, we verify the benefit of scene conditioning for the SMPL denoiser. We train it with EgoBody-train in different conditioning schemes and report the " + }, + { + "bbox": [ + 304, + 201, + 547, + 416 + ], + "type": "inline_equation", + "content": "T = 32" + }, + { + "bbox": [ + 304, + 201, + 547, + 416 + ], + "type": "text", + "content": " results on EgoBody-val. First, placing the predicted bodies from 4DHuman in the global space directly with estimated camera extrinsics has the highest error (1" + }, + { + "bbox": [ + 304, + 201, + 547, + 416 + ], + "type": "inline_equation", + "content": "^{\\text{st}}" + }, + { + "bbox": [ + 304, + 201, + 547, + 416 + ], + "type": "text", + "content": " row). When conditioning on a constant zero tensor, the denoiser behaves like a motion prior and reduces the error (2" + }, + { + "bbox": [ + 304, + 201, + 547, + 416 + ], + "type": "inline_equation", + "content": "^{\\text{nd}}" + }, + { + "bbox": [ + 304, + 201, + 547, + 416 + ], + "type": "text", + "content": " row). To encode the appearance and geometry information of the scene, we consider ViT [8] or SPVCNN [53] as the encoder " + }, + { + "bbox": [ + 304, + 201, + 547, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 201, + 547, + 416 + ], + "type": "text", + "content": " and try varied combinations of appearance features (RGB), geometry features (XYZ) and aggregated subject masks (Mask). When using ViT to encode the scene, adding XYZ features or masks does not reduce the error. In contrast, when using SPVCNN, adding RGB information or conditioning on masks does improve performance. Overall, SPVCNN yields lower errors than ViT and enabling all conditioning leads to the lowest world-space error measure." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 428, + 495, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 428, + 495, + 441 + ], + "spans": [ + { + "bbox": [ + 305, + 428, + 495, + 441 + ], + "type": "text", + "content": "4.4. Qualitative Analysis and Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 449, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 592 + ], + "type": "text", + "content": "In the first two rows of Fig. 5, we visualize the results of 3DPW and EgoBody in a global space. Despite occlusions, our SynCHMR estimates human meshes reliably and places them in a dense scene point cloud, whereas the scenes in GLAMR [65] and SLAHMR [62] consist of only a simple ground plane. Applying scene constraints with such an overly simplified scene can result in erroneous estimation, e.g., incorrect human trajectories as shown in the top view of the " + }, + { + "bbox": [ + 304, + 449, + 545, + 592 + ], + "type": "inline_equation", + "content": "1^{\\text{st}}" + }, + { + "bbox": [ + 304, + 449, + 545, + 592 + ], + "type": "text", + "content": " row, and the vertically shortened human bodies in the " + }, + { + "bbox": [ + 304, + 449, + 545, + 592 + ], + "type": "inline_equation", + "content": "4^{\\text{th}}" + }, + { + "bbox": [ + 304, + 449, + 545, + 592 + ], + "type": "text", + "content": " row of (d). Note that since TRACE [52] is scene agnostic, the ground plane in (c) is only for visualization, not necessarily indicating scene penetration." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "We also test on more in-the-wild DAVIS [42] videos containing human subjects. Since DAVIS provides no ground-truth human meshes nor camera trajectories, we show only the visual comparison. The " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": " row shows that we can handle multi-person cases as well as SLAHMR, while GLAMR often fails when multiple humans and dynamic cameras both occur. In a challenging scenario where the subject is taking selfies (the " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "4^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": " row), both GLAMR and SLAHMR are confused by the foreground human dominating the frames and reconstruct an almost static global trajectory, failing to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1222" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 547, + 401 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 547, + 401 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 547, + 401 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 547, + 401 + ], + "type": "image", + "image_path": "2d1ff7543f2ddc4f0a56d27c65a6b48963f0baaf3a8c307f5d0ab13b2a563ff0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 407, + 546, + 440 + ], + "lines": [ + { + "bbox": [ + 46, + 407, + 546, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 546, + 440 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison among world-frame HMR approaches. We show (b) GLAMR [65] and (c) TRACE [52] results with their pre-defined ground planes, (d) SLAHMR [62] outputs with its estimated ground plane, and (e) our SynCHMR outputs with dense scenes. In the first row, we also demonstrate top-view human trajectories within circles. See supplementary for video results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 452, + 287, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 452, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 46, + 452, + 287, + 500 + ], + "type": "text", + "content": "disentangle the camera and the human motions due to the dynamic ambiguity. TRACE fails to produce results due to severe frame truncation. In contrast, SynCHMR still successfully provides reasonable trajectories." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 509, + 174, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 509, + 174, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 509, + 174, + 521 + ], + "type": "text", + "content": "5. Limitation Discussion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 529, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 660 + ], + "type": "text", + "content": "As SynCHMR focuses on disentangling camera and human movements, we follow SLAHMR to approximate the focal length as " + }, + { + "bbox": [ + 46, + 529, + 287, + 660 + ], + "type": "inline_equation", + "content": "\\frac{W + H}{2}" + }, + { + "bbox": [ + 46, + 529, + 287, + 660 + ], + "type": "text", + "content": ". When the subject has a shape that the body model cannot explain well, e.g., children or obese people, calibrating depth with the estimated bodies is less ideal. As we develop and validate SynCHMR on real videos, its accuracy on composed or generated videos remains an open question. Finally, since SynCHMR handles dynamic scenes with moving subjects, it does not require an a priori scanned static scene. This opens up new challenges, such as incorporating dynamic point clouds as scene constraints." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 670, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 119, + 681 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "We present SynCHMR, a method that reconstructs camera trajectories, human bodies, and dense scenes from in-the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 451, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 545, + 620 + ], + "type": "text", + "content": "wild videos all in one global coordinate. SynCHMR has two core innovations. First, it leverages monocular depth estimation and uses the dimension and location of human meshes to calibrate the range of depth. This allows SLAM to better resolve the inherent scale ambiguity problem as shown in the experiment. Second, we train a data-driven motion denoiser and condition it with the scene in the same global coordinate, which is the first such scene-conditioned motion prior. Combining the two, the full SynCHMR pipeline uses human bodies to improve SLAM, and the better estimated scene and camera trajectory, in turn, provide better constraints for feed-forward human motion denoising. It achieves SOTA results on common benchmarks compared with existing optimization-based approaches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 633, + 398, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 398, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 398, + 646 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "We appreciate constructive comments from Duygu Ceylan. This project was partially supported by the NIH under contracts R01GM134020 and P41GM103712, and by the NSF under contracts DBI-1949629, DBI-2238093, IIS-2007595, IIS-2211597, and MCB-2205148." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1223" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Iman Abaspur Kazerouni, Luke Fitzgerald, Gerard Dooly, and Daniel Toal. A survey of state-of-the-art on visual SLAM. Expert Systems with Applications, 205:117734, 2022. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "text", + "content": "[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. ZoeDepth: Zero-shot transfer by combining relative and metric depth. arXiv, 2023. 2, 4, 6, 7, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "type": "text", + "content": "[3] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European Conference on Computer Vision (ECCV), 2016. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "type": "text", + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Computer Vision and Pattern Recognition (CVPR), pages 1290–1299, 2022. 4, 7, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "text", + "content": "[5] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consistent 3d human pose and shape from a video. In Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 336, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 288, + 390 + ], + "type": "text", + "content": "[6] Yudi Dai, Yitai Lin, Chenglu Wen, Siqi Shen, Lan Xu, Jingyi Yu, Yuexin Ma, and Cheng Wang. HSC4D: Human-centered 4D scene capture in large-scale indoor-outdoor space using wearable imus and LiDAR. In Computer Vision and Pattern Recognition (CVPR), pages 6792-6802, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 446 + ], + "type": "text", + "content": "[7] Yudi Dai, Yitai Lin, Xiping Lin, Chenglu Wen, Lan Xu, Hongwei Yi, Siqi Shen, Yuexin Ma, and Cheng Wang. Sloper4d: A scene-aware dataset for global 4d human pose estimation in urban environments. In Computer Vision and Pattern Recognition (CVPR), pages 682-692, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 288, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 523 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 524, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 524, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 524, + 288, + 578 + ], + "type": "text", + "content": "[9] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023. 2, 3, 4, 5, 6, 7, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 601 + ], + "type": "text", + "content": "[10] John C Gower. Generalized procrustes analysis. Psychometrika, 40(1):33-51, 1975. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 635 + ], + "type": "text", + "content": "[11] Riza Alp Guler and Iasonas Kokkinos. HoloPose: Holistic 3D human reconstruction in-the-wild. In Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "type": "text", + "content": "[12] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human positioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In Computer Vision and Pattern Recognition (CVPR). IEEE, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[13] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambigu" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ities with 3D scene constraints. In International Conference on Computer Vision (ICCV), pages 2282-2292, 2019. 3, 6, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "text", + "content": "[14] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. Advances in Neural Information Processing Systems, 35:4244-4256, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "text", + "content": "[15] Dorian F Henning, Tristan Laidlow, and Stefan Leutenegger. BodySLAM: joint camera localisation, mapping, and human motion tracking. In European Conference on Computer Vision (ECCV), pages 656-673. Springer, 2022. 2, 3, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 241 + ], + "type": "text", + "content": "[16] Dorian F Henning, Christopher Choi, Simon Schaefer, and Stefan Leutenegger. BodySLAM++: Fast and tightly-coupled visual-inertial camera and human motion tracking. In International Conference on Intelligent Robots and Systems (IROS), pages 3781-3788. IEEE, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 242, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 308 + ], + "type": "text", + "content": "[17] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In Computer Vision and Pattern Recognition (CVPR), pages 13274-13285, 2022. 3, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 309, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 364 + ], + "type": "text", + "content": "[18] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6M: Large scale datasets and predictive methods for 3D human sensing in natural environments. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 36(7):1325-1339, 2014. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "text", + "content": "[19] Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 410, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 464 + ], + "type": "text", + "content": "[20] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In International Conference on Computer Vision (ICCV), 2015. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 466, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 510 + ], + "type": "text", + "content": "[21] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3D deformation model for tracking faces, hands, and bodies. In Computer Vision and Pattern Recognition (CVPR), pages 8320-8329, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 510, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 555 + ], + "type": "text", + "content": "[22] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human pose fitting towards in-the-wild 3D human pose estimation. In International Conference on 3D Vision (3DV), pages 42-52, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 555, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 599 + ], + "type": "text", + "content": "[23] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "text", + "content": "[24] Angjoo Kanazawa, Jason Y. Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. In Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[25] Manuel Kaufmann, Jie Song, Chen Guo, Kaiyue Shen, Tianjian Jiang, Chengcheng Tang, Juan Jose Zárate, and Otmar Hilliges. EMDB: The Electromagnetic Database of Global 3D Human Pose and Shape in the Wild. In International Conference on Computer Vision (ICCV), 2023. 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "text", + "content": "[26] Imry Kissos, Lior Fritz, Matan Goldman, Omer Meir, Edward Oks, and Mark Kliger. Beyond weak perspective for" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1224" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 287, + 106 + ], + "type": "text", + "content": "monocular 3d human pose estimation. In European Conference on Computer Vision Workshops (ECCVw), pages 541-554, Cham, 2020. Springer International Publishing. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "text", + "content": "[27] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "text", + "content": "[28] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In International Conference on Computer Vision (ICCV). IEEE, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 287, + 248 + ], + "type": "text", + "content": "[29] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J. Black. SPEC: Seeing people in the wild with an estimated camera. In International Conference on Computer Vision (ICCV), pages 11035-11045, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 304 + ], + "type": "text", + "content": "[30] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J. Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. PACE: Human and motion estimation from in-the-wild videos. In International Conference on 3D Vision (3DV), 2024. 2, 3, 5, 6, 7, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "type": "text", + "content": "[31] Nikos Kolotouros, Georgios Pavlakos, Michael J. Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In International Conference on Computer Vision (ICCV), 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "type": "text", + "content": "[32] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Computer Vision and Pattern Recognition (CVPR), pages 3383-3393, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "text", + "content": "[33] Jiefeng Li, Siyuan Bian, Chao Xu, Gang Liu, Gang Yu, and Cewu Lu. D&D: Learning human dynamics from dynamic camera. In European Conference on Computer Vision (ECCV), 2022. 3, 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "text", + "content": "[34] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In European Conference on Computer Vision (ECCV), pages 590-606, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 548 + ], + "type": "text", + "content": "[35] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In Computer Vision and Pattern Recognition (CVPR), pages 21159-21168, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 601 + ], + "type": "text", + "content": "[36] Miao Liu, Dexin Yang, Yan Zhang, Zhaopeng Cui, James M Rehg, and Siyu Tang. 4D human body capture from egocentric video via 3D scene grounding. In International Conference on 3D Vision (3DV), pages 930-939. IEEE, 2021. 3, 5, 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "text", + "content": "[37] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. Transactions on Graphics (TOG), 34 (6):248:1-248:16, 2015. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "type": "text", + "content": "[38] Diogo Luvizon, Marc Habermann, Vladislav Golyanik, Adam Kortylewski, and Christian Theobalt. Scene-Aware 3D Multi-Human Motion Capture from a Single Camera. Computer Graphics Forum (CGF), 42(2):371-383, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "text", + "content": "[39] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "text", + "content": "motion capture as surface shapes. In International Conference on Computer Vision (ICCV), pages 5441-5450, 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 106, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 162 + ], + "type": "text", + "content": "[40] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3D pose estimation from monocular RGB. In International Conference on 3D Vision (3DV), 2018. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[41] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 219, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 272 + ], + "type": "text", + "content": "[42] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Computer Vision and Pattern Recognition (CVPR), 2016. 7, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 274, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 329 + ], + "type": "text", + "content": "[43] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(3):1623-1637, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 331, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 385 + ], + "type": "text", + "content": "[44] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. HuMoR: 3D human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), pages 11468-11479, 2021. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "type": "text", + "content": "[45] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. FrankMocap: Fast monocular 3d hand and body motion capture by regression and integration. In International Conference on Computer Vision Workshops (ICCVw), 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 432, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 545, + 475 + ], + "type": "text", + "content": "[46] Nitin Saini, Chun-Hao P Huang, Michael J Black, and Aamir Ahmad. SmartMocap: Joint estimation of human and camera motion using uncalibrated rgb cameras. IEEE Robotics and Automation Letters, 2023. 2, 3, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "type": "text", + "content": "[47] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Computer Vision and Pattern Recognition (CVPR), 2016. 3, 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 510, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 544 + ], + "type": "text", + "content": "[48] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Computer Vision and Pattern Recognition (CVPR), pages 134-144, 2019. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 545, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 599 + ], + "type": "text", + "content": "[49] Zehong Shen, Zhi Cen, Sida Peng, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Learning human mesh recovery in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17038-17047, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "type": "text", + "content": "[50] Jie Song, Xu Chen, and Otmar Hilliges. Human body model fitting by learned gradient descent. In European Conference on Computer Vision (ECCV), 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[51] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A benchmark for the evaluation of rgb-d slam systems. In International Conference on Intelligent Robots and Systems (IROS), 2012. 6, 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[52] Yu Sun, Qian Bao, Wu Liu, Tao Mei, and Michael J. Black. TRACE: 5D temporal regression of avatars with dynamic cameras in 3d environments. In Computer Vision and Pat" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1225" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "text", + "content": "tern Recognition (CVPR), pages 8856-8866, 2023. 7, 8, 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "text", + "content": "[53] Haotian* Tang, Zhijian* Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In European Conference on Computer Vision, 2020. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[54] Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2, 3, 5, 6, 7, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "type": "text", + "content": "[55] Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3d human mesh from monocular images: A survey. arXiv preprint arXiv:2203.01923, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 272 + ], + "type": "text", + "content": "[56] Timo von Marcard, Roberto Henschel, Michael J. Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3D human pose in the wild using IMUs and a moving camera. In European Conference on Computer Vision (ECCV), pages 614-631, 2018. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 327 + ], + "type": "text", + "content": "[57] Wenjia Wang, Yongtao Ge, Haiyi Mei, Zhongang Cai, Qingping Sun, Yanjun Wang, Chunhua Shen, Lei Yang, and Taku Komura. Zolly: Zoom focal length correctly for perspective-distorted human mesh reconstruction. In International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 360 + ], + "type": "text", + "content": "[58] Donglai Xiang, Hanbyul Joo, and Yaser Sheikh. Monocular total capture: Posing face, body, and hands in the wild. In Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 415 + ], + "type": "text", + "content": "[59] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In Computer Vision and Pattern Recognition (CVPR), pages 6183-6192, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 416, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 459 + ], + "type": "text", + "content": "[60] Yuanlu Xu, Song-Chun Zhu, and Tony Tung. DenseRaC: Joint 3d pose and shape estimation by dense render-and compare. In International Conference on Computer Vision (ICCV), 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "text", + "content": "[61] Ming Yan, Xin Wang, Yudi Dai, Siqi Shen, Chenglu Wen, Lan Xu, Yuexin Ma, and Cheng Wang. Cimi4d: A large multimodal climbing motion dataset under human-scene interactions. In Computer Vision and Pattern Recognition (CVPR), pages 12977-12988, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 515, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 568 + ], + "type": "text", + "content": "[62] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In Computer Vision and Pattern Recognition (CVPR), pages 21222-21232, 2023. 2, 3, 5, 6, 7, 8, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "text", + "content": "[63] Jae Shin Yoon, Zhixuan Yu, Jaesik Park, and Hyun Soo Park. HUMBI: A large multiview dataset of human body expressions and benchmark challenge. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(1):623-640, 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 658 + ], + "type": "text", + "content": "[64] Ri Yu, Hwangpil Park, and Jehee Lee. Human dynamics from monocular video with dynamic camera movements. Transactions on Graphics (TOG), 40(6), 2021. 3, 5, 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[65] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. GLAMR: Global occlusion-aware human mesh recovery with dynamic cameras. In Computer Vision and Pattern Recognition (CVPR), pages 11028-11039, 2022. 2, 3, 7, 8, 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 329 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[66] Andrei Zanfir, Eduard Gabriel Bazavan, Hongyi Xu, Bill Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Weakly supervised 3d human pose and shape reconstruction with normalizing flows. In European Conference on Computer Vision (ECCV), 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[67] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 184, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 228 + ], + "type": "text", + "content": "[68] Siwei Zhang, Yan Zhang, Federica Bogo, Marc Pollefeys, and Siyu Tang. Learning motion priors for 4D human body capture in 3D scenes. In International Conference on Computer Vision (ICCV), 2021. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 229, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 229, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 229, + 545, + 284 + ], + "type": "text", + "content": "[69] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. EgoBody: Human body shape and motion of interacting people from head-mounted devices. In European Conference on Computer Vision (ECCV), 2022. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 285, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 545, + 329 + ], + "type": "text", + "content": "[70] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snively, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision (ECCV), pages 20-37. Springer, 2022. 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1226" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/eca2ceba-84d3-4751-b38a-1ba7e74c92ba_content_list.json b/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/eca2ceba-84d3-4751-b38a-1ba7e74c92ba_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1637a603225aa6fc6653bb652657b4690eec0a22 --- /dev/null +++ b/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/eca2ceba-84d3-4751-b38a-1ba7e74c92ba_content_list.json @@ -0,0 +1,1729 @@ +[ + { + "type": "text", + "text": "Synthesize, Diagnose, and Optimize: Towards Fine-Grained Vision-Language Understanding", + "text_level": 1, + "bbox": [ + 181, + 128, + 789, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wujian Peng $^{1,2}$ Sicheng Xie $^{1,2}$ Zuyao You $^{1,2}$ Shiyi Lan $^{3}$ Zuxuan Wu $^{1,2\\dagger}$", + "bbox": [ + 178, + 200, + 787, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University $^{2}$ Shanghai Collaborative Innovation Center of Intelligent Visual Computing $^{3}$ NVIDIA", + "bbox": [ + 174, + 228, + 795, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 305, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision language models (VLM) have demonstrated remarkable performance across various downstream tasks. However, understanding fine-grained visual-linguistic concepts, such as attributes and inter-object relationships, remains a significant challenge. While several benchmarks aim to evaluate VLMs in finer granularity, their primary focus remains on the linguistic aspect, neglecting the visual dimension. Here, we highlight the importance of evaluating VLMs from both a textual and visual perspective. We introduce a progressive pipeline to synthesize images that vary in a specific attribute while ensuring consistency in all other aspects. Utilizing this data engine, we carefully design a benchmark, SPEC, to diagnose the comprehension of object size, position, existence, and count. Subsequently, we conduct a thorough evaluation of four leading VLMs on SPEC. Surprisingly, their performance is close to random guess, revealing significant limitations. With this in mind, we propose a simple yet effective approach to optimize VLMs in fine-grained understanding, achieving significant improvements on SPEC without compromising the zero-shot performance. Results on two additional fine-grained benchmarks also show consistent improvements, further validating the transferability of our approach. Code and data are available at https://github.com/wjpoom/SPEC.", + "bbox": [ + 75, + 337, + 473, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 729, + 209, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision and Language Foundation Models (VLMs) pretrained on large-scale image-text data [13, 17, 23, 29, 41] have consistently demonstrated impressive performance across a wide range of well-established evaluating tasks, i.e. image classification [5], image captioning [1, 18], visual question answering [3] and cross-modal image-text retrieval [18, 38]. Their remarkable performance is gradually convincing the community that these currently avail", + "bbox": [ + 75, + 753, + 468, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/081244d44c780210bb529ca213ef1e18fcdb4ba519bbcf54f44a0ff5b1800355.jpg", + "image_caption": [ + "(a) Image to Text Matching" + ], + "image_footnote": [], + "bbox": [ + 506, + 304, + 890, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f2783ff56bdb731df68494df20710ac98437830a11fb60c3a08bca652b7c8fd3.jpg", + "image_caption": [ + "(b) Text to Image Matching", + "Figure 1. We conduct a symmetrical assessment of VLMs in fine-grained comprehension, considering both linguistic and visual perspectives. The bars in $\\square$ represent the image-text matching scores for CLIP [23] and our method, respectively. It is evident that CLIP struggles with tasks related to quantity comprehension, whereas our method significantly enhances the model in understanding fine-grained details." + ], + "image_footnote": [], + "bbox": [ + 503, + 419, + 890, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "able VLMs are almost robust and powerful enough to be transferred to a broad spectrum of downstream tasks, either through finetuning or even in a zero-shot manner.", + "bbox": [ + 496, + 669, + 892, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, recent research has shattered this captivating illusion, revealing that even state-of-the-art VLMs [17, 23, 29, 41] exhibit significant limitations in understanding visual-linguistic concepts that require fine-grained compositional reasoning, especially in tasks involving object attributes or inter-object relationships [21, 31, 33, 40, 43]. This raises a crucial question: to what extent and in what aspects are VLMs excelling or struggling? To answer this, previous effort evaluates fine-grained capabilities of VLMs through the image-to-text matching task, as shown in Fig. 1(a). This involves providing a query image and retrieving the matching text from a set of confusing candi", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 101, + 886, + 230, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13279", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dates, differing subtly in texts. For example, when assessing the counting ability, it is crucial to ensure that quantity is the unique variable and other clues are kept the same. Therefore, a straightforward way to do so is modifying quantity and adjective words used in the texts. While manipulating texts to construct confusing candidate sets has been well-studied due to the sparsity of the text space and advancements in Large Language Models (LLMs) [19, 21, 40, 43], the visual side remains relatively under explored, largely due to the complexity of visual signals and the absence of powerful tools. We posit that exploring the visual dimension in a fine-grained manner is also essential for a comprehensive understanding of VLMs.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the great progress achieved in generative models, we present an effective framework for generating high-quality image candidates that are suitable for evaluating the performance of VLMs. This framework ensures that images within the same candidate set only differ in the specified property of interest, while all other properties remain consistent. We break down this task into several simple and manageable steps. As illustrated in Fig. 2, we start by utilizing a text-to-image model [22, 24] to generate images featuring a single object. Then, a segmenter [14, 15] is employed to separate the objects from their backgrounds, yielding a library of foreground instance spanning various categories. From there, we select instances and arrange them on a blank canvas (manipulating attributes such as size, position, existence, and quantity of a specific object at this stage is straightforward). Finally, we use an inpainting model [22, 24] to fill the missing background portions, producing an photo-realistic image. It is worth noting that, during the inpainting process, we design a progressive background filling strategy, effectively ensuring consistency in the background across all images in the same candidate set.", + "bbox": [ + 75, + 291, + 472, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Empowered by this data construction pipeline, we carefully develop a new benchmark, named as SPEC, to evaluate the proficiency of VLMs in comprehending fine-grained concepts including Size, Position, Existence and Counting. We systematical test four VLMs [17, 23, 29, 41] on this newly created test bed. Surprisingly, even state-of-the-art models perform at chance-level, exposing significant performance deficiencies. Following this, we implement a straightforward approach to remedy this by incorporating hard negative examples (i.e., confusing images or texts within the same candidate set) into the same training batch. This encourages the model to discern subtle differences among candidate examples, leading to a significant improvement in performance on SEPC while preserving the original zero-shot capability. Furthermore, to demonstrate the model's generalization ability, we conduct additional tests on two existing datasets [33, 40], which also focus on compositional reasoning. The consistent improvement further validates that our method effectively guide the model", + "bbox": [ + 75, + 613, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to acquire essential and transferable comprehending abilities at a finer granularity. Our main contributions are:", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. A progressive data constructing pipeline. We present a progressive data construction pipeline designed for creating a candidate image set. Within each candidate set, images vary exclusively in a specified attribute while ensuring consistency across other aspects. Such data are valuable for conducting text-to-image matching tasks, as depicted in Fig. 1(b), offering a visual perspective for evaluating VLMs.", + "2. A carefully curated benchmark: SPEC. We meticulously craft a novel benchmark, SPEC, with a specific focus on evaluating VLMs' understanding of fine-grained visual-linguistic concepts, encompassing object size, position, existence, and count. The introduction of SPEC enables a symmetrical evaluation of VLMs from both image and text perspectives, addressing the previous lack of image-centric testing data.", + "3. A simple and effective remedy. We evaluate four VLMs on SPEC, revealing significant limitations. In response, we propose a method to enhance the understanding of fine-grained visual-linguistic concepts. Experimental results indicate notable improvement not only on SPEC but also consistent results on two additional datasets, while preserving zero-shot capability." + ], + "bbox": [ + 500, + 121, + 893, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 479, + 640, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision and Language Models (VLMs). Models such as CLIP [23], ALIGN [13], CyCLIP [10] CoCa [39], OmniVL [32] and Open-VCLIP [34, 35] have demonstrated impressive performance across a wide range of downstream tasks. These models include two separate unimodal encoders, each designed to extract representations for visual and textual input. To achieve alignment between the two modalities, they typically employ a huge number of visual-textual pairs for contrastive learning. By pretraining on 400M noisy data, CLIP [23] achieves a top-1 accuracy on ImageNet-1K [5] comparable to that of ResNet-50 [11], even though it is never specifically trained on ImageNet and is evaluated in a zero-shot manner. However, as highlighted in recent work [31, 40], these advancements are primarily attributed to the simplicity of evaluation tasks which requires no reasoning or compositional capabilities. The performance of these models are limited on tasks that require fine-grained understanding [21, 40].", + "bbox": [ + 496, + 515, + 890, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Benchmarking VLMs in Finer Granularity. To assess the model's understanding of nuanced visual-linguistic concepts, several new benchmarks have been proposed. Winoground [31] is curated by experts with a focus on compositional understanding. VALID [21] and VL-Checklist [43] investigate several linguistic phenomena by transforming real captions into confusing alternatives.", + "bbox": [ + 496, + 794, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13280", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/098b3614aa297badaebef7d1fb29a1ea41a5cdd78623dd20599ba4c251485aaf.jpg", + "image_caption": [ + "Figure 2. The overall illustration of our data progressive construction pipeline. We initiate the process by generating a batch of images containing a single object. Subsequently, we extract the object from the background in the images. Following that, we arrange the background-free images on a blank canvas according to specifications (with control over attributes). Finally, we meticulously fill in the missing background, ensuring consistency across candidates." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 893, + 308 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ARO [40] diagnoses VLMs in attribution, relation and ordering. Eqben [33] assesses whether the model is sensitive to visual semantic changes. Most existing benchmarks solely focus on subtle textual changes [12, 19, 21, 40, 43], as crafting confusing text candidates is straightforward that can be achieved through either LLMs or simple rules. Winoground [40] and Eqben [33] are most relevant to us, since we focus on minimal semantic changes in both images and texts, enabling a more comprehensive evaluation across modalities. However, the scale of Winoground is restricted by its costly curation, and the image diversity of Eqben is limited by virtual engines. In contrast, our data construction pipeline is scalable and can produce diverse images.", + "bbox": [ + 75, + 385, + 472, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Enhancing VLMs for Fine-grained Understanding. To mitigate challenges in fine-grained recognition, various approaches have been explored. Syn-CLIP [2] utilize data synthesized by 3D simulation engines to enhance the model's understanding of concepts beyond nouns. EQSIM [33] incorporates an additional regularization loss to generalize VLMs to nuanced multimodal compositions. TSVLC [7] and ViLEM [4] introduce negative texts generated by LLMs [6, 26] to inject fine-grained knowledge. Construct-VL [30] addresses these challenges from a continual learning perspective. These methods compel the model to focus on subtle differences by introducing confusing texts as hard negatives. However, we argue that the absence of visual hard negatives limits its performance. Thus, we introduce hard negatives for both modalities, simultaneously enhancing the visual and textual encoders.", + "bbox": [ + 75, + 589, + 472, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Synthesize: Data Construction Pipeline", + "text_level": 1, + "bbox": [ + 76, + 845, + 431, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to build a set of perplexing image, wherein each image differs solely in a specified attribute while ensuring", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2ea48f0364ac828571b2f925099d178a4c131109275b05698d3a7d15ecaa7ab7.jpg", + "image_caption": [ + "Figure 3. Ensuring consistency among candidates is crucial to avoid ambiguity. The images above not only differ in quantity but also show a significant variation in the appearance of the objects. Consequently, attributing the model's correctness or errors solely to the understanding of quantity is not convincing." + ], + "image_footnote": [], + "bbox": [ + 558, + 383, + 834, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "consistency in all other aspects. We first emphasize the importance of preserving consistency among candidate images for effective evaluation (Sec. 3.1). To address this, we break down this problem and introduce a progressive data construction pipeline (Sec. 3.2). Then, we carefully devise a benchmark that centers on evaluating VLMs' grasp of fine-grained visual-linguistic concepts (Sec. 3.3).", + "bbox": [ + 496, + 619, + 893, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Importance of Candidate Consistency", + "text_level": 1, + "bbox": [ + 498, + 739, + 826, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, when conducting a matching task with the textual query \"a photo of two dogs\", the model might mistakenly choose the image on the right (which actually contains three dogs). However, attributing this error solely to counting difficulty is not convincing. The model might select the right-side image due to its more photorealistic appearance or better alignment with the word \"dog\", as the left-side image has a cartoon style. Conversely, if the model correctly selects the right-side image", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13281", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for the query \"a photo of three dogs\", we also cannot assert that the model is proficient in counting. These obscure ambiguities arise because there is no guarantee of the uniqueness of changing factors among candidate images during evaluation. Therefore, ensuring consistency among candidates in all aspects except the one under investigation is crucial. To this end, we propose a progressive data construction method, which will be elaborated in detail as follows.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Progressive Data Construction", + "text_level": 1, + "bbox": [ + 76, + 220, + 349, + 237 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The data construction framework is illustrated in Fig. 2, comprising four progressive steps. Initially, we generate images featuring a single prominent object. Then, we isolate the foreground from the background, resulting in a library of foreground objects spanning various categories. Subsequently, we select objects from this library and arrange them on a blank canvas, adjusting their attributes and relationships. Lastly, we carefully fill in the missing background, ensuring consistency among different candidates.", + "bbox": [ + 75, + 244, + 472, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Generating Images with Single Objects", + "text_level": 1, + "bbox": [ + 76, + 398, + 405, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We initiate the process by utilizing a generation model to obtain a collection of images, each featuring a single and prominent object corresponding to a specific category. Due to the progress in visual generation models [24, 25, 36, 37], these images display a high level of photo-realistic and diverse content. In practice, we use Stable-Diffusion-XL 1.0 [22] as our generator, and prompt it with \"a photo of a single and fully visible [class name]\". The emphasis on \"single and fully visible\" is crucial, as the model might otherwise generate images with multiple objects or encounter occlusion issues, as observed in [42]. The [class name] represents a specific category from 80 classes of COCO [18].", + "bbox": [ + 75, + 422, + 472, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Isolating Objects from the Background", + "text_level": 1, + "bbox": [ + 76, + 623, + 403, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For ease in subsequent processes, we need to separate the objects from the backgrounds where they are embedded. To accomplish this, we first utilize an open-set detector, Grounding-DINO [20] to outline the regions containing the objects. Subsequently, we prompt SAM [15] with this bounding box as to obtain the final segmentation results. Thus far, we have established a library containing instances from various categories. These instances are background-free, allowing for composition on a blank canvas, while their attributes and relations can be controlled.", + "bbox": [ + 75, + 647, + 472, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3 Arranging Objects on the Canvas", + "text_level": 1, + "bbox": [ + 76, + 816, + 366, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recall that our goal is to manipulate a specific visual-linguistic concept of an image. Currently, this task appears straightforward when we exclude the background from consideration. We can retrieve instances from the library and", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "arrange them on a blank canvas according to our specifications. For example, we can flexibly control the quantity of an object through duplication operations, modify their sizes via resizing, and determine whether an object exists and specify the position of an existing object. This process resembles the concept of copy-paste [9, 42], with a notable distinction: we paste objects onto a blank background and placing emphasis on controlling properties such as the size, position, existence, and quantity of each individual object.", + "bbox": [ + 496, + 90, + 893, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.4 Infilling the Missing Background", + "text_level": 1, + "bbox": [ + 498, + 244, + 784, + 260 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We have constructed images with differences in specified attributes, however, they currently lack a suitable background. To fill the missing area, we employ a inpainting model [22] which demonstrates proficiency in filling large holes. It is worth noting that generating backgrounds individually for each image would result in significant differences in the backgrounds, posing a challenge to maintaining consistency among candidate images. To overcome this, we introduce a strategy where images within the same candidate set share a common and consistent background during the inpainting process. As depicted in the upper part of Fig. 4, to present the giraffe in different positions, we start by surrounding it with an initial background. Then, we relocate this initial image on the canvas as required, and fill the remaining blank space. Similarly, in the lower part, we first embed the zebra into a reasonable environment. Following that, we expand the scene horizontally, introduce the elephant, and fill the blank areas through inpainting. In summary, we begin the process by generating an initial background, which is then expanded to the surroundings, effectively ensuring consistency among the candidate images. Additional examples, such as adjusting the size or quantity of an object, can be found in the supplementary.", + "bbox": [ + 496, + 268, + 893, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. SPEC Benchmark", + "text_level": 1, + "bbox": [ + 500, + 625, + 678, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Utilizing the data engine outlined in Sec. 3.2, we carefully devise the SPEC benchmark with the goal of assessing the performance of VLMs in comprehending object size, position, existence and count. An overview of the SPEC benchmark is presented in Fig. 5. SPEC contains six subsets, which will be elaborated as follows:", + "bbox": [ + 496, + 647, + 893, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Absolute Size reflects how large an object is in comparison to the entire image. We categorize this into three levels: large, medium, or small, and define them following:", + "bbox": [ + 496, + 750, + 893, + 796 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S i z e} _ {a b s.} (x) = \\left\\{ \\begin{array}{l l} \\text {s m a l l}, & P \\leq 0. 2 \\\\ \\text {m e d i u m}, & 0. 4 \\leq P \\leq 0. 6 \\\\ \\text {l a r g e}, & P \\geq 0. 8, \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 804, + 890, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P$ denotes the proportion of the space occupied by the object relative to the area of the entire image. A safety", + "bbox": [ + 498, + 869, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13282", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b05e55f7ee605039c607c86b96c6f89711416fdd85bfa7870ef808f2b2ba3646.jpg", + "image_caption": [ + "Figure 4. Consistent background inpainting strategy. We first generate an initial background shared by all candidate images. Then, we expand around this region, ensuring consistency in the backgrounds of different images." + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 468, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "threshold is deliberately introduced between these three levels to prevent ambiguity.", + "bbox": [ + 75, + 406, + 468, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Relative Size focuses on the size relationship between two objects. It is categorized as follows: object A is [smaller than, equal to, larger than] object B, and we measure this following:", + "bbox": [ + 75, + 446, + 472, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S i z e} _ {r e a.} (A, B) = \\left\\{ \\begin{array}{l l} A \\text {s m a l l e r t h a n} B, & R \\leq 0. 5 \\\\ A \\text {e q u a l t o} B, & 0. 9 \\leq R \\leq 1. 1 \\\\ A \\text {l a r g e r t h a n} B, & R \\geq 2, \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 516, + 486, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R = \\frac{S_A}{S_B}$ is the ratio of the areas of object A and object B.", + "bbox": [ + 75, + 585, + 472, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Absolute Position signifies the location of an object relative to the image. We partition the image into a $3 \\times 3$ grid, defining nine possible positions: top-left, top, top-right, left, center, right, bottom-left, bottom, and bottom-right. The absolute position of an object is determined based on the grid in which its center point resides.", + "bbox": [ + 75, + 626, + 468, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Relative Position describes the spatial relationship between two objects. We consider four common spatial relationships: A is [to the left of, to the right of, above, below] B. The position relationship between objects is defined based on the relative positions of their center points.", + "bbox": [ + 75, + 728, + 470, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existence indicates whether an object appears in a given image, expressed using existential quantifiers: There is [no, at least one] object in the image.", + "bbox": [ + 75, + 814, + 470, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Count represents the number of occurrences of an object, providing a metric for the model's quantitative understand-", + "bbox": [ + 75, + 869, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/20d9acfd11a90b1187e5f182e6f9beaffa0be2b83f9a304b60e8cf48f0dc612d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BenchmarksVisual HNScalabilityText-realisticPhoto-realistic#Candidates
FALSE [21]X✓✓✓2
ARO [40]X✓X✓2
Winoground [31]✓X✓✓2
Eqben [33]✓✓✓✓2
SPEC(Ours)✓✓✓✓2-9
", + "bbox": [ + 504, + 88, + 893, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparison of SPEC with other fine-grained benchmarks.", + "bbox": [ + 498, + 176, + 893, + 203 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing. Due to potential occlusion issues with a large number of objects, we restrict our consideration to the range of 1 to 9: “there are [one, two, ..., nine] object(s) in the image”.", + "bbox": [ + 498, + 228, + 893, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Format. The basic unit of SPEC is an individual test case, wherein each test case comprises two components: an image candidate set, which differs only in certain visual aspects, and a text candidate set, which differs only in the corresponding language descriptions. We formally represent a test case as:", + "bbox": [ + 498, + 282, + 893, + 375 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {T}: \\left(\\mathcal {I} = \\left\\{I _ {1}, \\dots , I _ {K} \\right\\}, \\mathcal {T} = \\left\\{T _ {1}, \\dots , T _ {K} \\right\\}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 382, + 890, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{I}$ and $\\mathcal{T}$ represent the image and text candidate sets, respectively. The i-th image $I_{i}$ is paired with the i-th text $T_{i}$ , i.e., they mutually describe each other. $K$ is the semantic cardinality of the test case, determined by the definition of each subset. For instance, if a test case belongs to the absolute size subset, then $K = 3$ (representing the three semantics: large, medium, small). In Fig. 5, we present examplar test cases for each subset, and more examples can be found in the supplementary.", + "bbox": [ + 498, + 406, + 893, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparing SPEC with other benchmarks. In Tab. 1, we conduct a comprehensive comparison of SPEC with four similar benchmarks. Visual HN indicates the presence of image hard negatives, which are essential for the text-to-image matching task. Notably, VALID [21] and ARO [40] concentrate solely on text hard negatives, neglecting their visual counterparts. Scalability assesses whether the data construction method can be scaled up. For instance, Winoground [31] is limited to 400 examples due to high costs for manual collection. Additionally, Text-realistic evaluates the grammatical correctness of the texts, where ARO involves directly swapping the positions of two words without considering grammar. Similarly, Image-realistic indicates the realism of the images. Eqben [33] incorporates images rendered using a virtual engine, compromising their quality. In contrast, images from SPEC, while synthesized, leverage advanced image generation models and our effective data construction pipeline, resulting in photo-realistic images. Finally, we compare the number of Candidates in each test case. In all datasets except SPEC, each example features only two candidates, i.e., identifying the correct item from two candidates, which is relatively straightforward. In contrast, tasks in SPEC are more challenging,", + "bbox": [ + 496, + 551, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13283", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Absolute Size", + "text_level": 1, + "bbox": [ + 89, + 88, + 184, + 101 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2a5cca9bedece61d7b999815b34ce7ac2dea1412d8bb1daae44e397ddc16950e.jpg", + "image_caption": [ + "500 test cases, 3 candidates each", + "the sheep is large in the image \nthe sheep is medium-sized in the image \nthe sheep is small in the image" + ], + "image_footnote": [], + "bbox": [ + 81, + 116, + 326, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Relative Size", + "text_level": 1, + "bbox": [ + 89, + 220, + 181, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/db313293c0402a989307f4edaa6c3e30780b1a41d711493d1f76ea959442cd28.jpg", + "image_caption": [ + "500 test cases, 3 candidates each", + "the pizza is smaller than the sandwich \nthe pizza is equal-size with the sandwich \nthe pizza is bigger than the sandwich" + ], + "image_footnote": [], + "bbox": [ + 81, + 250, + 326, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Absolute Position", + "text_level": 1, + "bbox": [ + 369, + 89, + 493, + 102 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/13e47e4bb7989688a9307c7f2827199f1e8543b9fd20f37027c3240fd228bc92.jpg", + "image_caption": [ + "500 test cases, 9 candidates each" + ], + "image_footnote": [], + "bbox": [ + 361, + 117, + 442, + 178 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c3d9f019e69011da1f4102204c48986e13edea0b2e6f3f61ebe4f430a25b3d30.jpg", + "image_caption": [ + "the kite is in the top-left corner \nthe kite is in the center of the image \nthe kite is in the bottom-left corner" + ], + "image_footnote": [], + "bbox": [ + 442, + 118, + 522, + 179 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56f20d2194fb1fde36d7314976f430b56dc15e7c459540efda1f20b16698c026.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 118, + 604, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Relative Position", + "text_level": 1, + "bbox": [ + 370, + 223, + 488, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c21fcae67e0124aa3a071d61b520e21550e1ffe7c385e9170a2f9633e7906b2f.jpg", + "image_caption": [ + "500 test cases, 4 candidates each", + "the toilet is to the left of the backpack the toilet is on top of the backpack the toilet is below the backpack" + ], + "image_footnote": [], + "bbox": [ + 361, + 250, + 442, + 313 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/274c2c6cc8881b0357cadbaca0405e14ca0c8627ba32bed87dc9ac4408d7dffe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 251, + 526, + 313 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8036839ee38258c4555b0bf141d748f6937890ba953253a1530c0e996a1a4639.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 251, + 607, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Count", + "text_level": 1, + "bbox": [ + 651, + 89, + 697, + 102 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cb3b518742e3132392c0f1f50a8edc610c0d64f9f051b5ccf8828be52e02c186.jpg", + "image_caption": [ + "500 test cases, 9 candidates each", + "a photo of one snowboard a photo of two snowboards a photo of three snowboards" + ], + "image_footnote": [], + "bbox": [ + 643, + 116, + 723, + 179 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d8dc1b298fccc219b7c60a7330ed00f5d2bc8720ff5aed3997cc369c765fc6ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 117, + 805, + 179 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b685079b70d99267165d64f1c466b33ccfebc6ff8953471ce928dd0bb8fef1fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 117, + 890, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Existence", + "text_level": 1, + "bbox": [ + 645, + 223, + 715, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fc0c3b73e183f4c03a7f8d157c1073d5e08b73822d12e4c437da95d27788534e.jpg", + "image_caption": [ + "500 test cases, 2 candidates each", + "there is no giraffe in the image", + "there is a giraffe in the image" + ], + "image_footnote": [], + "bbox": [ + 638, + 252, + 720, + 314 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d2a8c259b75cedb5576255e8975b8ee82f1201bbcbca5679a9d71900fe4b5d87.jpg", + "image_caption": [ + "Figure 5. An overview of the SPEC benchmark. SPEC consists of six distinct subsets, distributed across the dimensions of Size, Position, Existence and Count. Each test case consists of an image candidate set, which differs only in certain visual concept, and a text candidate set, which differs only in corresponding language concept. Due to space constraints, we present a maximum of three images and texts here, however, more comprehensive test cases are available in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 805, + 252, + 885, + 315 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the semantic space covered by the candidate set is more comprehensive. For example, in the case of relative position, the candidates include all four semantic directions (up, down, left, right), and in absolute position and count, it extends to nine candidates. As will be discussed below, more confusing candidates can be readily used as hard negatives to improve current VLMs.", + "bbox": [ + 75, + 440, + 468, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Diagnose: Probing VLMs on SPEC", + "text_level": 1, + "bbox": [ + 76, + 559, + 397, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct a systematical evaluation of four state-of-the-art VLMs: CLIP [23], BLIP [17], FLAVA [29] and CoCa [39] using our newly proposed SPEC benchmark, aiming to uncover that to what extent and in what aspect are they excelling or suffering.", + "bbox": [ + 75, + 584, + 468, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Evaluation Task and Protocol", + "text_level": 1, + "bbox": [ + 76, + 670, + 339, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Symmetric image-text matching task. We conduct evaluations using the image-text matching task, where each test case $\\mathbb{T} = (\\mathcal{I},\\mathcal{T})$ comprises an image set $\\mathcal{I}$ and a corresponding text set $\\mathcal{T}$ , as outlined in Eq. (3). The evaluation process is symmetric, considering both visual and textual perspectives. In the image-to-text matching task, when querying with an image $I_{i}$ , the model is required to accurately identify $T_{i}$ from the candidate set $\\mathcal{T}$ . Similarly, in the text-to-image matching task, the goal is to find $I_{i}$ for a given query text $T_{i}$ . In practice, we accomplish the matching process using the similarity $s(I,T)$ between a given image $I$ and text $T$ . A candidate will be selected if its similarity with the query ranks first among the entire candidate set.", + "bbox": [ + 75, + 703, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation protocols. We measure the performance on SPEC using two metrics: I2TACC and T2IACC, representing the accuracy of image-to-text and text-to-image matching task, respectively:", + "bbox": [ + 498, + 440, + 892, + 500 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI 2 \\mathrm {T} _ {\\text {A C C}} = \\frac {1}{| D |} \\sum_ {\\left(\\mathcal {I} _ {i}, \\mathcal {T} _ {i}\\right) \\in D} \\frac {1}{\\left| \\mathcal {I} _ {i} \\right|} \\sum_ {I _ {j} \\in \\mathcal {I} _ {i}} h \\left(I _ {j}, \\mathcal {T} _ {i}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 521, + 890, + 560 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {T 2 I A C C} = \\frac {1}{| D |} \\sum_ {\\left(\\mathcal {I} _ {i}, \\mathcal {T} _ {i}\\right) \\in D} \\frac {1}{\\left| \\mathcal {T} _ {i} \\right|} \\sum_ {T _ {j} \\in \\mathcal {T} _ {i}} g \\left(T _ {j}, \\mathcal {I} _ {i}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 582, + 890, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $D$ contains $|D|$ test cases, $h(I_j,\\mathcal{T}_i)$ equals to 1 if and only if $I_{j}$ correctly find its matched text $T_{j}$ from the candidate set $\\mathcal{T}_i$ , otherwise it is set to 0. Similarly, $g(T_j,\\mathcal{I}_i)$ equals 1 if and only if $T_{j}$ correctly finds its matched image $I_{j}$ from the candidate set $\\mathcal{I}_i$ , otherwise, it is set to 0.", + "bbox": [ + 498, + 630, + 893, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Key Insights from SPEC Results", + "text_level": 1, + "bbox": [ + 500, + 715, + 784, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate four VLMs using the SPEC benchmark, and their results are summarized in Tab. 2. We find that all the models exhibit a limited accuracy close to random chance, from which we gain the following insights:", + "bbox": [ + 498, + 738, + 890, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Even state-of-the-art VLMs perform at chance level. From the results, we surprisingly find that even the most advanced VLMs achieve only a marginal advantage compared to random chance, which sharply contrasts with their impressive performance on common tasks. For instance, CLIP [23] demonstrates a mere $33.4\\%$ T2IACC for relative", + "bbox": [ + 496, + 809, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13284", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/06f022482399cac0ffb05fc251a64a6481b42748aad0f5d08735f48c843e0e8e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Absolute SizeRelative SizeAbsolute PositionRelative PositionExistenceCount
I2TT2ICLSI2TT2ICLSI2TT2ICLSI2TT2ICLSI2TT2ICLSI2TT2ICLS
Random33.333.31.333.333.32.511.111.11.325.025.02.550.050.05.011.111.11.3
FLAVA [29]37.337.289.332.932.384.113.115.788.825.526.784.057.951.974.414.421.275.8
BLIP [17]43.342.797.733.232.598.212.211.097.930.529.797.655.450.196.037.437.193.8
CoCa [39]39.136.195.933.633.395.411.711.896.130.328.892.250.950.083.736.535.689.6
CLIP [23]42.536.192.134.433.494.912.612.392.828.026.690.158.351.284.025.123.283.9
Ours68.960.796.340.344.197.330.634.296.946.646.996.283.453.192.555.657.892.5
", + "bbox": [ + 81, + 88, + 893, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Evaluation results on SPEC. We extensively benchmark four state-of-the-art VLMs on SPEC to investigate their comprehension of fine-grained visual-linguistic concepts. Our evaluation employs two metrics, I2TACC and T2IACC, and we report the detailed performance on the each subset. We also report the classification accuracy, CLS, to highlight the capability in recognizing object categories. The first row indicates accuracy at chance level, serving as a baseline for comparison.", + "bbox": [ + 75, + 244, + 893, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "size recognition, while the chance-level accuracy is $33.3\\%$ . While BLIP performs reasonably well in absolute size, surpassing random level by around $9.7\\%$ , the i2TACC on relative size is $0.7\\%$ lagged behind. CoCa [39] and FLAVA [29] also exhibit significant weaknesses in performance. The last row presents the performance of our improved model, demonstrating significant advancements across all metrics (as will be introduced in Sec. 5).", + "bbox": [ + 75, + 327, + 470, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The challenge arises from the task itself, not the data. One might attribute the poor performance of VLMs to the data quality or distribution. To address this concern, we conduct an additional experiment. Specifically, we perform classification experiments using the SPEC dataset to assess the model's understanding of nouns or object categories. In this context, the models exhibit impressive performance, achieving approximately $90\\%$ Top-1 accuracy in the 80-class classification task (denoted as CLS in Tab. 2). This aligns well with earlier findings [40] that VLMs struggle in compositional reasoning while excelling in object category recognition. The remarkable accuracy of the models in the object classification task confirms the high quality of SPEC data. This also validates that the challenges faced by VLMs stem from the tasks which require fine-grained recognition rather than issues with the data itself.", + "bbox": [ + 75, + 459, + 468, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Discussion on Model Limitations", + "text_level": 1, + "bbox": [ + 76, + 710, + 366, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We attribute the poor performance of vision and language models on SPEC to their pretraining methods, specifically, the inherent limitation in standard contrastive loss. The conventional contrastive learning involves randomly sampling batches of images and texts, requiring the model to identify matching pairs within the batch. This task is intended to facilitate alignment between text and image spaces. However, as highlighted in prior studies [2, 40], the substantial differences between items in a randomly sampled batch allows the model to effortlessly complete this task. It can easily achieve this by focusing solely on nouns", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in the text and object categories in the images through a shortcut [8]. This leads the model biased towards noun concepts, neglecting other finer-grained concepts. This is the reason why these models demonstrate poor performance on fine-grained tasks that demanding understanding concepts beyond nouns.", + "bbox": [ + 496, + 327, + 893, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Optimize: A Simple but Effective Remedy", + "text_level": 1, + "bbox": [ + 498, + 429, + 875, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We experiment with CLIP [23] and propose a remedy to enhance its performance in fine-grained understanding.", + "bbox": [ + 496, + 455, + 893, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Method", + "text_level": 1, + "bbox": [ + 500, + 492, + 596, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CLIP [23] consists of an visual encoder to extract image embedding: $e_{I} = \\mathcal{E}_{I}(I)$ and a textual encoder to extract text embedding: $e_{T} = \\mathcal{E}_{T}(T)$ . The similarity score between an image $I$ and a text $T$ is computed following:", + "bbox": [ + 496, + 516, + 893, + 575 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ns (I, T) = \\exp \\left(\\frac {\\tau e _ {I} ^ {T} e _ {T}}{\\| e _ {I} \\| ^ {2} \\| e _ {T} \\| ^ {2}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 582, + 890, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\tau$ is a learnable temperature.", + "bbox": [ + 500, + 622, + 735, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to guide CLIP to focus on fine-grained visual-linguistic concepts, we incorporate confusing images and text as hard negatives within the same batch. This requires CLIP to pull positives closer and push hard negatives away, thereby enhancing its ability to discern nuanced visual and textual differences. Specifically, we introduce an hard negative aware contrastive loss $\\mathcal{L}_{hn} = \\mathcal{L}_{hn}^{12\\mathrm{T}} + \\mathcal{L}_{hn}^{\\mathrm{T}2\\mathrm{I}}$ , which comprises an image-to-text and a text-to-image term:", + "bbox": [ + 496, + 637, + 893, + 758 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {h n} ^ {\\mathrm {l 2 T}} = - \\sum_ {i} \\log \\frac {s \\left(I _ {i} , T _ {i}\\right)}{\\sum_ {T _ {j} \\in \\mathcal {T}} s \\left(I _ {i} , T _ {j}\\right) + \\sum_ {T _ {k} ^ {h n} \\in \\mathcal {T} ^ {h n}} s \\left(I _ {i} , T _ {k} ^ {h n}\\right)} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 762, + 890, + 821 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {h n} ^ {\\mathrm {T 2 I}} = - \\sum_ {i} \\log \\frac {s \\left(I _ {i} , T _ {i}\\right)}{\\sum_ {I _ {j} \\in \\mathcal {I}} s \\left(I _ {j} , T _ {i}\\right) + \\sum_ {I _ {k} ^ {h n} \\in \\mathcal {I} ^ {h n}} s \\left(I _ {k} ^ {h n} , T _ {i}\\right)}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13285", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{I}$ and $\\mathcal{T}$ represent the trivial images and texts, respectively, while $\\mathcal{I}^{hn}$ and $\\mathcal{T}^{hn}$ denote non-trivial hard negatives. In our implementation, these hard negative examples are constructed using the data pipeline described in Sec. 3.2, and more details are in the supplementary.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To preserve the inherent zero-shot capability of CLIP, we also leverage the conventional image-text pairs from LAION-400M [27]. We apply the standard contrastive loss of CLIP [23] to these data, introducing an additional loss term $\\mathcal{L}_{clip}$ . The overall loss consists of two terms:", + "bbox": [ + 75, + 167, + 468, + 243 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {c l i p}} + \\lambda \\mathcal {L} _ {\\text {h n}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 252, + 468, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyperparameter that balancing these two terms. Training on this multi-task loss enables improving the performance of CLIP in fine-grained understanding while maintaining its zero-shot capability.", + "bbox": [ + 75, + 277, + 468, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Experiments", + "text_level": 1, + "bbox": [ + 76, + 347, + 210, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training details. We experiment with the ViT-B/32 variant of CLIP, and resume from the OpenAI pretrained checkpoint [23]. We finetune for 1,000 steps using a cosine schedule with an initial learning rate of $1e - 6$ and use 800 steps for warm up. The batch size of LAION data is set to 2048, and the batch size of hard negative data is set to 768. The weight $\\lambda$ of the hard negative aware loss is set to 0.2.", + "bbox": [ + 75, + 380, + 468, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Main results. We utilize the SPEC benchmark to assess the understanding of model in fine-grained concepts. In Tab. 3, we present the average I2TACC and T2IACC on all subsets of SPEC. To assess the general performance of the model, we also utilize the toolkit from ELEVATER [16] to evaluate the zero-shot performance on 9 classification and retrieval datasets, and report the average accuracy. Compared to CLIP [23], our model demonstrates remarkable advancements with a $19.8\\%$ boost in I2TACC, an $18.9\\%$ improvement in T2IACC on SPEC, and a noteworthy $1.2\\%$ enhancement in zero-shot accuracy. We also conduct ablation on different training configurations. From the results in Tab. 3, it can be observed that the introducing of $\\mathcal{L}_{hn}$ significantly improves the performance on SPEC. Moreover, the $\\mathcal{L}_{clip}$ plays a crucial role in preserving zero-shot performance. Without it, we observe a decline in accuracy by $5.1\\%$ . With the combination of these two losses, we achieve substantial improvement in SPEC while maintaining the original zero-shot capability.", + "bbox": [ + 75, + 496, + 468, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Validation on other fine-grained benchmarks. To assess whether our approach aids the model in acquiring fundamental visual-linguistic understanding or merely leads to overfitting on SPEC, we conduct evaluations on two additional benchmarks which also focus on the assessment of fine-grained concepts. ARO [40] explores three aspects of vision-language understanding: object attributes,", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/feb3bac01ed06e163f0e601f91fe1860f60a998f86c764a18d4837e928fd4e10.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ConfigSPECZero-shot Accuracy
LclipLhnI2TT2I
CLIP33.530.567.5
+ Lhn✓64.360.862.4
+ Lclip✓32.231.569.4
+ Lhn+Lclip (ours)✓✓53.349.468.7
", + "bbox": [ + 514, + 88, + 883, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e7fd3ebc1eb3b558b7d3405d2141784aa5355205f3f044fec1885f3f51ff5726.jpg", + "table_caption": [ + "Table 3. Main Results: The first row represents the pretrained checkpoint without fine-tuning. We sequentially introduce two additional loss terms to investigate their impact for the performance." + ], + "table_footnote": [], + "table_body": "
EqbenARO
ImageTextGroupAttributeRelationOrder
CLIP17.621.410.163.263.953.3
CLIPFT18.123.711.165.168.054.1
Ours19.524.011.766.473.760.7
", + "bbox": [ + 504, + 266, + 890, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Cross-dataset evaluation results on Eqben [33] and ARO [40]. To demonstrate that the improvement comes from the negative loss, rather than training on more data, we also report $\\mathrm{CLIP}_{FT}$ , which is also finetuned but without negative samples.", + "bbox": [ + 498, + 372, + 890, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "inter-object relations, and word ordering. Eqben [33] focuses on minimal visual semantic changes, aiming to diagnose VLMs in understanding fine-grained concepts such as counting and location. In Tab. 4, we present the experimental results, demonstrating a clear improvement compared to CLIP [23] on both datasets. For example, compared to CLIP, our method shows an average improvement of $2\\%$ on Eqben and respective enhancements of $3.2\\%$ , $9.8\\%$ , and $7.4\\%$ on the three subsets of ARO. The consistent improvement on these datasets demonstrates that our approach has facilitated the model in acquiring transferable fine-grained understanding capabilities.", + "bbox": [ + 496, + 454, + 890, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 648, + 617, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we explored the comprehension abilities of Visual Language Models (VLMs) with respect to fine-grained visual-linguistic concepts. We first established an efficient pipeline to synthesize candidate images that exclusively differ in a particular visual attribute. Leveraging this pipeline, we created the SPEC benchmark to diagnose the comprehension proficiency of VLMs in terms of object size, position, existence, and count. Upon evaluating four leading VLMs using SPEC, we uncovered substantial performance limitations. To address this, we introduced an enhancement strategy that effectively optimizes the model for fine-grained understanding, while maintaining its original zero-shot capability.", + "bbox": [ + 496, + 672, + 890, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This project was supported by NSFC under Grant No. 62102092.", + "bbox": [ + 498, + 869, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13286", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In ICCV, 2019. 1", + "[2] Paola Cascante-Bonilla, Khaled Shehada, James Smith, Sivan Doveh, Donghyun Kim, Rameswar Panda, Gül Varol, Aude Oliva, Vicente Ordonez, Rogério Schmidt Feris, and Leonid Karlinsky. Going beyond nouns with vision & language models using synthetic data. In ICCV, 2023. 3, 7", + "[3] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel M. Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model. In ICLR, 2022. 1", + "[4] Yuxin Chen, Zongyang Ma, Ziqi Zhang, Zhongang Qi, Chunfen Yuan, Ying Shan, Bing Li, Weiming Hu, Xiaohu Qie, and Jianping Wu. Vilem: Visual-language error modeling for image-text retrieval. In CVPR, 2023. 3", + "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1, 2", + "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 3", + "[7] Sivan Doveh, Assaf Arbelle, Sivan Harary, Rameswar Panda, Roei Herzig, Eli Schwartz, Donghyun Kim, Raja Giryes, Rogério Schmidt Feris, Shimon Ullman, and Leonid Karlinsky. Teaching structured vision & language concepts to vision & language models. In CVPR, 2023. 3", + "[8] Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard S. Zemel, Wieland Brendel, Matthias Bethge, and Felix Wichmann. Shortcut learning in deep neural networks. Nat. Mach. Intell., 2020. 7", + "[9] Golnaz Ghiasi, Yin Cui, A. Srinivas, Rui Qian, Tsung-Yi Lin, Ekin Dogus Cubuk, Quoc V. Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. CVPR, 2020. 4", + "[10] Shashank Goel, Hritik Bansal, Sumit Kaur Bhatia, Ryan A. Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. In NeurIPS, 2022. 2", + "[11] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2015. 2", + "[12] Lisa Anne Hendricks and Aida Nematzadeh. Probing image-language transformers for verb understanding. In Findings, 2021. 3", + "[13] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1, 2", + "[14] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing" + ], + "bbox": [ + 78, + 114, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. ArXiv, 2023. 2", + "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In ICCV, 2023. 2, 4", + "[16] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Yong Jae Lee, Houdong Hu, Zicheng Liu, and Jianfeng Gao. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. In NeurIPS, 2022. 8", + "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 1, 2, 6, 7, 3", + "[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 4, 3", + "[19] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. TACL, 2023. 2, 3", + "[20] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. ArXiv, 2023. 4", + "[21] Letitia Parcalabescu, Michele Cafagna, Lilitta Muradjan, Anette Frank, Iacer Calixto, and Albert Gatt. Valse: A task-independent benchmark for vision and language models centered on linguistic phenomena. In ACL, 2021. 1, 2, 3, 5", + "[22] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. ArXiv, 2023. 2, 4", + "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 6, 7, 8, 3", + "[24] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In 2021. 2, 4", + "[25] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 4", + "[26] Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Galle, et al. Bloom: A 176b-parameter open-access multilingual language model. ArXiv, 2022. 3", + "[27] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. ArXiv, 2021. 8" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13287", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 3", + "[29] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022. 1, 2, 6, 7, 3", + "[30] James Smith, Paola Cascante-Bonilla, Assaf Arbelle, Donghyun Kim, Rameswar Panda, David Cox, Diyi Yang, Zsolt Kira, Rogério Schmidt Feris, and Leonid Karlinsky. Construct-vl: Data-free continual structured vl concepts learning*. In CVPR, 2023. 3", + "[31] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In CVPR, 2022. 1, 2, 5", + "[32] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnivl: One foundation model for image-language and video-language tasks. In NeurIPS, 2022. 2", + "[33] Tan Wang, Kevin Lin, Linjie Li, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Equivariant similarity for vision-language foundation models. In ICCV, 2023. 1, 2, 3, 5, 8", + "[34] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization. In ICML, 2023. 2", + "[35] Zuxuan Wu, Zejia Weng, Wujian Peng, Xitong Yang, Ang Li, Larry S Davis, and Yu-Gang Jiang. Building an open-vocabulary video clip model with better architectures, optimization and data. TPAMI, 2024. 2", + "[36] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. arXiv preprint arXiv:2310.10647, 2023. 4", + "[37] Zhen Xing, Qi Dai, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Simda: Simple diffusion adapter for efficient video generation. In CVPR, 2024. 4", + "[38] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 1", + "[39] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Trans. Mach. Learn. Res., 2022. 2, 6, 7, 3", + "[40] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bag-of-words models, and what to do about it? ICLR, 2023. 1, 2, 3, 5, 7, 8", + "[41] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. In ICML, 2021. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Hanqing Zhao, Dianmo Sheng, Jianmin Bao, Dongdong Chen, Dong Chen, Fang Wen, Lu Yuan, Ce Liu, Wenbo Zhou, Qi Chu, Weiming Zhang, and Neng H. Yu. X-paste: Revisiting scalable copy-paste for instance segmentation using clip and stablediffusion. In ICML, 2022. 4", + "[43] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. ArXiv, 2022. 1, 2, 3" + ], + "bbox": [ + 501, + 90, + 890, + 229 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13288", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/eca2ceba-84d3-4751-b38a-1ba7e74c92ba_model.json b/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/eca2ceba-84d3-4751-b38a-1ba7e74c92ba_model.json new file mode 100644 index 0000000000000000000000000000000000000000..24863a7dda1eaca4c6f68aa57c049f521a94d11e --- /dev/null +++ b/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/eca2ceba-84d3-4751-b38a-1ba7e74c92ba_model.json @@ -0,0 +1,2321 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.182, + 0.13, + 0.79, + 0.177 + ], + "angle": 0, + "content": "Synthesize, Diagnose, and Optimize: Towards Fine-Grained Vision-Language Understanding" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.202, + 0.788, + 0.223 + ], + "angle": 0, + "content": "Wujian Peng\\(^{1,2}\\) Sicheng Xie\\(^{1,2}\\) Zuyao You\\(^{1,2}\\) Shiyi Lan\\(^{3}\\) Zuxuan Wu\\(^{1,2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.229, + 0.796, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University \\(^{2}\\)Shanghai Collaborative Innovation Center of Intelligent Visual Computing \\(^{3}\\)NVIDIA" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.306, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.474, + 0.701 + ], + "angle": 0, + "content": "Vision language models (VLM) have demonstrated remarkable performance across various downstream tasks. However, understanding fine-grained visual-linguistic concepts, such as attributes and inter-object relationships, remains a significant challenge. While several benchmarks aim to evaluate VLMs in finer granularity, their primary focus remains on the linguistic aspect, neglecting the visual dimension. Here, we highlight the importance of evaluating VLMs from both a textual and visual perspective. We introduce a progressive pipeline to synthesize images that vary in a specific attribute while ensuring consistency in all other aspects. Utilizing this data engine, we carefully design a benchmark, SPEC, to diagnose the comprehension of object size, position, existence, and count. Subsequently, we conduct a thorough evaluation of four leading VLMs on SPEC. Surprisingly, their performance is close to random guess, revealing significant limitations. With this in mind, we propose a simple yet effective approach to optimize VLMs in fine-grained understanding, achieving significant improvements on SPEC without compromising the zero-shot performance. Results on two additional fine-grained benchmarks also show consistent improvements, further validating the transferability of our approach. Code and data are available at https://github.com/wjpoom/SPEC." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.731, + 0.21, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.755, + 0.47, + 0.878 + ], + "angle": 0, + "content": "Vision and Language Foundation Models (VLMs) pretrained on large-scale image-text data [13, 17, 23, 29, 41] have consistently demonstrated impressive performance across a wide range of well-established evaluating tasks, i.e. image classification [5], image captioning [1, 18], visual question answering [3] and cross-modal image-text retrieval [18, 38]. Their remarkable performance is gradually convincing the community that these currently avail" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.305, + 0.891, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.389, + 0.785, + 0.404 + ], + "angle": 0, + "content": "(a) Image to Text Matching" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.42, + 0.891, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.632, + 0.531, + 0.782, + 0.545 + ], + "angle": 0, + "content": "(b) Text to Image Matching" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.555, + 0.894, + 0.653 + ], + "angle": 0, + "content": "Figure 1. We conduct a symmetrical assessment of VLMs in fine-grained comprehension, considering both linguistic and visual perspectives. The bars in \\(\\square\\) represent the image-text matching scores for CLIP [23] and our method, respectively. It is evident that CLIP struggles with tasks related to quantity comprehension, whereas our method significantly enhances the model in understanding fine-grained details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.893, + 0.716 + ], + "angle": 0, + "content": "able VLMs are almost robust and powerful enough to be transferred to a broad spectrum of downstream tasks, either through finetuning or even in a zero-shot manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "However, recent research has shattered this captivating illusion, revealing that even state-of-the-art VLMs [17, 23, 29, 41] exhibit significant limitations in understanding visual-linguistic concepts that require fine-grained compositional reasoning, especially in tasks involving object attributes or inter-object relationships [21, 31, 33, 40, 43]. This raises a crucial question: to what extent and in what aspects are VLMs excelling or struggling? To answer this, previous effort evaluates fine-grained capabilities of VLMs through the image-to-text matching task, as shown in Fig. 1(a). This involves providing a query image and retrieving the matching text from a set of confusing candi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.887, + 0.232, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13279" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "dates, differing subtly in texts. For example, when assessing the counting ability, it is crucial to ensure that quantity is the unique variable and other clues are kept the same. Therefore, a straightforward way to do so is modifying quantity and adjective words used in the texts. While manipulating texts to construct confusing candidate sets has been well-studied due to the sparsity of the text space and advancements in Large Language Models (LLMs) [19, 21, 40, 43], the visual side remains relatively under explored, largely due to the complexity of visual signals and the absence of powerful tools. We posit that exploring the visual dimension in a fine-grained manner is also essential for a comprehensive understanding of VLMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.473, + 0.611 + ], + "angle": 0, + "content": "Motivated by the great progress achieved in generative models, we present an effective framework for generating high-quality image candidates that are suitable for evaluating the performance of VLMs. This framework ensures that images within the same candidate set only differ in the specified property of interest, while all other properties remain consistent. We break down this task into several simple and manageable steps. As illustrated in Fig. 2, we start by utilizing a text-to-image model [22, 24] to generate images featuring a single object. Then, a segmenter [14, 15] is employed to separate the objects from their backgrounds, yielding a library of foreground instance spanning various categories. From there, we select instances and arrange them on a blank canvas (manipulating attributes such as size, position, existence, and quantity of a specific object at this stage is straightforward). Finally, we use an inpainting model [22, 24] to fill the missing background portions, producing an photo-realistic image. It is worth noting that, during the inpainting process, we design a progressive background filling strategy, effectively ensuring consistency in the background across all images in the same candidate set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Empowered by this data construction pipeline, we carefully develop a new benchmark, named as SPEC, to evaluate the proficiency of VLMs in comprehending fine-grained concepts including Size, Position, Existence and Counting. We systematical test four VLMs [17, 23, 29, 41] on this newly created test bed. Surprisingly, even state-of-the-art models perform at chance-level, exposing significant performance deficiencies. Following this, we implement a straightforward approach to remedy this by incorporating hard negative examples (i.e., confusing images or texts within the same candidate set) into the same training batch. This encourages the model to discern subtle differences among candidate examples, leading to a significant improvement in performance on SEPC while preserving the original zero-shot capability. Furthermore, to demonstrate the model's generalization ability, we conduct additional tests on two existing datasets [33, 40], which also focus on compositional reasoning. The consistent improvement further validates that our method effectively guide the model" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "to acquire essential and transferable comprehending abilities at a finer granularity. Our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.122, + 0.894, + 0.242 + ], + "angle": 0, + "content": "1. A progressive data constructing pipeline. We present a progressive data construction pipeline designed for creating a candidate image set. Within each candidate set, images vary exclusively in a specified attribute while ensuring consistency across other aspects. Such data are valuable for conducting text-to-image matching tasks, as depicted in Fig. 1(b), offering a visual perspective for evaluating VLMs." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.243, + 0.895, + 0.363 + ], + "angle": 0, + "content": "2. A carefully curated benchmark: SPEC. We meticulously craft a novel benchmark, SPEC, with a specific focus on evaluating VLMs' understanding of fine-grained visual-linguistic concepts, encompassing object size, position, existence, and count. The introduction of SPEC enables a symmetrical evaluation of VLMs from both image and text perspectives, addressing the previous lack of image-centric testing data." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.363, + 0.895, + 0.47 + ], + "angle": 0, + "content": "3. A simple and effective remedy. We evaluate four VLMs on SPEC, revealing significant limitations. In response, we propose a method to enhance the understanding of fine-grained visual-linguistic concepts. Experimental results indicate notable improvement not only on SPEC but also consistent results on two additional datasets, while preserving zero-shot capability." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.122, + 0.895, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.48, + 0.642, + 0.496 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.892, + 0.789 + ], + "angle": 0, + "content": "Vision and Language Models (VLMs). Models such as CLIP [23], ALIGN [13], CyCLIP [10] CoCa [39], OmniVL [32] and Open-VCLIP [34, 35] have demonstrated impressive performance across a wide range of downstream tasks. These models include two separate unimodal encoders, each designed to extract representations for visual and textual input. To achieve alignment between the two modalities, they typically employ a huge number of visual-textual pairs for contrastive learning. By pretraining on 400M noisy data, CLIP [23] achieves a top-1 accuracy on ImageNet-1K [5] comparable to that of ResNet-50 [11], even though it is never specifically trained on ImageNet and is evaluated in a zero-shot manner. However, as highlighted in recent work [31, 40], these advancements are primarily attributed to the simplicity of evaluation tasks which requires no reasoning or compositional capabilities. The performance of these models are limited on tasks that require fine-grained understanding [21, 40]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Benchmarking VLMs in Finer Granularity. To assess the model's understanding of nuanced visual-linguistic concepts, several new benchmarks have been proposed. Winoground [31] is curated by experts with a focus on compositional understanding. VALID [21] and VL-Checklist [43] investigate several linguistic phenomena by transforming real captions into confusing alternatives." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13280" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.895, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.318, + 0.897, + 0.377 + ], + "angle": 0, + "content": "Figure 2. The overall illustration of our data progressive construction pipeline. We initiate the process by generating a batch of images containing a single object. Subsequently, we extract the object from the background in the images. Following that, we arrange the background-free images on a blank canvas according to specifications (with control over attributes). Finally, we meticulously fill in the missing background, ensuring consistency across candidates." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.473, + 0.584 + ], + "angle": 0, + "content": "ARO [40] diagnoses VLMs in attribution, relation and ordering. Eqben [33] assesses whether the model is sensitive to visual semantic changes. Most existing benchmarks solely focus on subtle textual changes [12, 19, 21, 40, 43], as crafting confusing text candidates is straightforward that can be achieved through either LLMs or simple rules. Winoground [40] and Eqben [33] are most relevant to us, since we focus on minimal semantic changes in both images and texts, enabling a more comprehensive evaluation across modalities. However, the scale of Winoground is restricted by its costly curation, and the image diversity of Eqben is limited by virtual engines. In contrast, our data construction pipeline is scalable and can produce diverse images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.473, + 0.834 + ], + "angle": 0, + "content": "Enhancing VLMs for Fine-grained Understanding. To mitigate challenges in fine-grained recognition, various approaches have been explored. Syn-CLIP [2] utilize data synthesized by 3D simulation engines to enhance the model's understanding of concepts beyond nouns. EQSIM [33] incorporates an additional regularization loss to generalize VLMs to nuanced multimodal compositions. TSVLC [7] and ViLEM [4] introduce negative texts generated by LLMs [6, 26] to inject fine-grained knowledge. Construct-VL [30] addresses these challenges from a continual learning perspective. These methods compel the model to focus on subtle differences by introducing confusing texts as hard negatives. However, we argue that the absence of visual hard negatives limits its performance. Thus, we introduce hard negatives for both modalities, simultaneously enhancing the visual and textual encoders." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.433, + 0.865 + ], + "angle": 0, + "content": "3. Synthesize: Data Construction Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Our goal is to build a set of perplexing image, wherein each image differs solely in a specified attribute while ensuring" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.384, + 0.835, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.533, + 0.895, + 0.605 + ], + "angle": 0, + "content": "Figure 3. Ensuring consistency among candidates is crucial to avoid ambiguity. The images above not only differ in quantity but also show a significant variation in the appearance of the objects. Consequently, attributing the model's correctness or errors solely to the understanding of quantity is not convincing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.895, + 0.73 + ], + "angle": 0, + "content": "consistency in all other aspects. We first emphasize the importance of preserving consistency among candidate images for effective evaluation (Sec. 3.1). To address this, we break down this problem and introduce a progressive data construction pipeline (Sec. 3.2). Then, we carefully devise a benchmark that centers on evaluating VLMs' grasp of fine-grained visual-linguistic concepts (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.741, + 0.828, + 0.759 + ], + "angle": 0, + "content": "3.1. Importance of Candidate Consistency" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, when conducting a matching task with the textual query \"a photo of two dogs\", the model might mistakenly choose the image on the right (which actually contains three dogs). However, attributing this error solely to counting difficulty is not convincing. The model might select the right-side image due to its more photorealistic appearance or better alignment with the word \"dog\", as the left-side image has a cartoon style. Conversely, if the model correctly selects the right-side image" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "13281" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "for the query \"a photo of three dogs\", we also cannot assert that the model is proficient in counting. These obscure ambiguities arise because there is no guarantee of the uniqueness of changing factors among candidate images during evaluation. Therefore, ensuring consistency among candidates in all aspects except the one under investigation is crucial. To this end, we propose a progressive data construction method, which will be elaborated in detail as follows." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.222, + 0.35, + 0.238 + ], + "angle": 0, + "content": "3.2. Progressive Data Construction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.245, + 0.473, + 0.382 + ], + "angle": 0, + "content": "The data construction framework is illustrated in Fig. 2, comprising four progressive steps. Initially, we generate images featuring a single prominent object. Then, we isolate the foreground from the background, resulting in a library of foreground objects spanning various categories. Subsequently, we select objects from this library and arrange them on a blank canvas, adjusting their attributes and relationships. Lastly, we carefully fill in the missing background, ensuring consistency among different candidates." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.4, + 0.406, + 0.416 + ], + "angle": 0, + "content": "3.2.1 Generating Images with Single Objects" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.473, + 0.606 + ], + "angle": 0, + "content": "We initiate the process by utilizing a generation model to obtain a collection of images, each featuring a single and prominent object corresponding to a specific category. Due to the progress in visual generation models [24, 25, 36, 37], these images display a high level of photo-realistic and diverse content. In practice, we use Stable-Diffusion-XL 1.0 [22] as our generator, and prompt it with \"a photo of a single and fully visible [class name]\". The emphasis on \"single and fully visible\" is crucial, as the model might otherwise generate images with multiple objects or encounter occlusion issues, as observed in [42]. The [class name] represents a specific category from 80 classes of COCO [18]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.624, + 0.405, + 0.64 + ], + "angle": 0, + "content": "3.2.2 Isolating Objects from the Background" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.648, + 0.473, + 0.799 + ], + "angle": 0, + "content": "For ease in subsequent processes, we need to separate the objects from the backgrounds where they are embedded. To accomplish this, we first utilize an open-set detector, Grounding-DINO [20] to outline the regions containing the objects. Subsequently, we prompt SAM [15] with this bounding box as to obtain the final segmentation results. Thus far, we have established a library containing instances from various categories. These instances are background-free, allowing for composition on a blank canvas, while their attributes and relations can be controlled." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.367, + 0.833 + ], + "angle": 0, + "content": "3.2.3 Arranging Objects on the Canvas" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Recall that our goal is to manipulate a specific visual-linguistic concept of an image. Currently, this task appears straightforward when we exclude the background from consideration. We can retrieve instances from the library and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.23 + ], + "angle": 0, + "content": "arrange them on a blank canvas according to our specifications. For example, we can flexibly control the quantity of an object through duplication operations, modify their sizes via resizing, and determine whether an object exists and specify the position of an existing object. This process resembles the concept of copy-paste [9, 42], with a notable distinction: we paste objects onto a blank background and placing emphasis on controlling properties such as the size, position, existence, and quantity of each individual object." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.246, + 0.785, + 0.261 + ], + "angle": 0, + "content": "3.2.4 Infilling the Missing Background" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.269, + 0.895, + 0.618 + ], + "angle": 0, + "content": "We have constructed images with differences in specified attributes, however, they currently lack a suitable background. To fill the missing area, we employ a inpainting model [22] which demonstrates proficiency in filling large holes. It is worth noting that generating backgrounds individually for each image would result in significant differences in the backgrounds, posing a challenge to maintaining consistency among candidate images. To overcome this, we introduce a strategy where images within the same candidate set share a common and consistent background during the inpainting process. As depicted in the upper part of Fig. 4, to present the giraffe in different positions, we start by surrounding it with an initial background. Then, we relocate this initial image on the canvas as required, and fill the remaining blank space. Similarly, in the lower part, we first embed the zebra into a reasonable environment. Following that, we expand the scene horizontally, introduce the elephant, and fill the blank areas through inpainting. In summary, we begin the process by generating an initial background, which is then expanded to the surroundings, effectively ensuring consistency among the candidate images. Additional examples, such as adjusting the size or quantity of an object, can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.625, + 0.679, + 0.64 + ], + "angle": 0, + "content": "3.3. SPEC Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.894, + 0.74 + ], + "angle": 0, + "content": "Utilizing the data engine outlined in Sec. 3.2, we carefully devise the SPEC benchmark with the goal of assessing the performance of VLMs in comprehending object size, position, existence and count. An overview of the SPEC benchmark is presented in Fig. 5. SPEC contains six subsets, which will be elaborated as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.797 + ], + "angle": 0, + "content": "Absolute Size reflects how large an object is in comparison to the entire image. We categorize this into three levels: large, medium, or small, and define them following:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.805, + 0.892, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S i z e} _ {a b s.} (x) = \\left\\{ \\begin{array}{l l} \\text {s m a l l}, & P \\leq 0. 2 \\\\ \\text {m e d i u m}, & 0. 4 \\leq P \\leq 0. 6 \\\\ \\text {l a r g e}, & P \\geq 0. 8, \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where \\( P \\) denotes the proportion of the space occupied by the object relative to the area of the entire image. A safety" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13282" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.088, + 0.469, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.326, + 0.473, + 0.385 + ], + "angle": 0, + "content": "Figure 4. Consistent background inpainting strategy. We first generate an initial background shared by all candidate images. Then, we expand around this region, ensuring consistency in the backgrounds of different images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.47, + 0.439 + ], + "angle": 0, + "content": "threshold is deliberately introduced between these three levels to prevent ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.473, + 0.51 + ], + "angle": 0, + "content": "Relative Size focuses on the size relationship between two objects. It is categorized as follows: object A is [smaller than, equal to, larger than] object B, and we measure this following:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.517, + 0.487, + 0.586 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S i z e} _ {r e a.} (A, B) = \\left\\{ \\begin{array}{l l} A \\text {s m a l l e r t h a n} B, & R \\leq 0. 5 \\\\ A \\text {e q u a l t o} B, & 0. 9 \\leq R \\leq 1. 1 \\\\ A \\text {l a r g e r t h a n} B, & R \\geq 2, \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.586, + 0.473, + 0.618 + ], + "angle": 0, + "content": "where \\( R = \\frac{S_A}{S_B} \\) is the ratio of the areas of object A and object B." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.721 + ], + "angle": 0, + "content": "Absolute Position signifies the location of an object relative to the image. We partition the image into a \\(3 \\times 3\\) grid, defining nine possible positions: top-left, top, top-right, left, center, right, bottom-left, bottom, and bottom-right. The absolute position of an object is determined based on the grid in which its center point resides." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.729, + 0.471, + 0.807 + ], + "angle": 0, + "content": "Relative Position describes the spatial relationship between two objects. We consider four common spatial relationships: A is [to the left of, to the right of, above, below] B. The position relationship between objects is defined based on the relative positions of their center points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.815, + 0.471, + 0.863 + ], + "angle": 0, + "content": "Existence indicates whether an object appears in a given image, expressed using existential quantifiers: There is [no, at least one] object in the image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Count represents the number of occurrences of an object, providing a metric for the model's quantitative understand-" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.895, + 0.168 + ], + "angle": 0, + "content": "
BenchmarksVisual HNScalabilityText-realisticPhoto-realistic#Candidates
FALSE [21]X✓✓✓2
ARO [40]X✓X✓2
Winoground [31]✓X✓✓2
Eqben [33]✓✓✓✓2
SPEC(Ours)✓✓✓✓2-9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.177, + 0.895, + 0.204 + ], + "angle": 0, + "content": "Table 1. Comparison of SPEC with other fine-grained benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.229, + 0.895, + 0.275 + ], + "angle": 0, + "content": "ing. Due to potential occlusion issues with a large number of objects, we restrict our consideration to the range of 1 to 9: “there are [one, two, ..., nine] object(s) in the image”." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.284, + 0.895, + 0.375 + ], + "angle": 0, + "content": "Data Format. The basic unit of SPEC is an individual test case, wherein each test case comprises two components: an image candidate set, which differs only in certain visual aspects, and a text candidate set, which differs only in the corresponding language descriptions. We formally represent a test case as:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.383, + 0.892, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\mathbb {T}: \\left(\\mathcal {I} = \\left\\{I _ {1}, \\dots , I _ {K} \\right\\}, \\mathcal {T} = \\left\\{T _ {1}, \\dots , T _ {K} \\right\\}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.407, + 0.895, + 0.545 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}\\) and \\(\\mathcal{T}\\) represent the image and text candidate sets, respectively. The i-th image \\(I_{i}\\) is paired with the i-th text \\(T_{i}\\), i.e., they mutually describe each other. \\(K\\) is the semantic cardinality of the test case, determined by the definition of each subset. For instance, if a test case belongs to the absolute size subset, then \\(K = 3\\) (representing the three semantics: large, medium, small). In Fig. 5, we present examplar test cases for each subset, and more examples can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Comparing SPEC with other benchmarks. In Tab. 1, we conduct a comprehensive comparison of SPEC with four similar benchmarks. Visual HN indicates the presence of image hard negatives, which are essential for the text-to-image matching task. Notably, VALID [21] and ARO [40] concentrate solely on text hard negatives, neglecting their visual counterparts. Scalability assesses whether the data construction method can be scaled up. For instance, Winoground [31] is limited to 400 examples due to high costs for manual collection. Additionally, Text-realistic evaluates the grammatical correctness of the texts, where ARO involves directly swapping the positions of two words without considering grammar. Similarly, Image-realistic indicates the realism of the images. Eqben [33] incorporates images rendered using a virtual engine, compromising their quality. In contrast, images from SPEC, while synthesized, leverage advanced image generation models and our effective data construction pipeline, resulting in photo-realistic images. Finally, we compare the number of Candidates in each test case. In all datasets except SPEC, each example features only two candidates, i.e., identifying the correct item from two candidates, which is relatively straightforward. In contrast, tasks in SPEC are more challenging," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13283" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.09, + 0.089, + 0.186, + 0.102 + ], + "angle": 0, + "content": "Absolute Size" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.103, + 0.277, + 0.115 + ], + "angle": 0, + "content": "500 test cases, 3 candidates each" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.117, + 0.328, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.18, + 0.275, + 0.214 + ], + "angle": 0, + "content": "the sheep is large in the image \nthe sheep is medium-sized in the image \nthe sheep is small in the image" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.222, + 0.182, + 0.234 + ], + "angle": 0, + "content": "Relative Size" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.236, + 0.277, + 0.248 + ], + "angle": 0, + "content": "500 test cases, 3 candidates each" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.251, + 0.327, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.314, + 0.285, + 0.347 + ], + "angle": 0, + "content": "the pizza is smaller than the sandwich \nthe pizza is equal-size with the sandwich \nthe pizza is bigger than the sandwich" + }, + { + "type": "title", + "bbox": [ + 0.37, + 0.09, + 0.494, + 0.103 + ], + "angle": 0, + "content": "Absolute Position" + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.104, + 0.558, + 0.116 + ], + "angle": 0, + "content": "500 test cases, 9 candidates each" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.118, + 0.443, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.444, + 0.119, + 0.524, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.119, + 0.606, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.18, + 0.544, + 0.214 + ], + "angle": 0, + "content": "the kite is in the top-left corner \nthe kite is in the center of the image \nthe kite is in the bottom-left corner" + }, + { + "type": "title", + "bbox": [ + 0.371, + 0.224, + 0.489, + 0.236 + ], + "angle": 0, + "content": "Relative Position" + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.238, + 0.558, + 0.249 + ], + "angle": 0, + "content": "500 test cases, 4 candidates each" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.251, + 0.444, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.252, + 0.527, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.252, + 0.609, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.315, + 0.551, + 0.347 + ], + "angle": 0, + "content": "the toilet is to the left of the backpack the toilet is on top of the backpack the toilet is below the backpack" + }, + { + "type": "title", + "bbox": [ + 0.652, + 0.09, + 0.699, + 0.103 + ], + "angle": 0, + "content": "Count" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.104, + 0.839, + 0.116 + ], + "angle": 0, + "content": "500 test cases, 9 candidates each" + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.117, + 0.725, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.118, + 0.807, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.118, + 0.892, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.181, + 0.788, + 0.214 + ], + "angle": 0, + "content": "a photo of one snowboard a photo of two snowboards a photo of three snowboards" + }, + { + "type": "title", + "bbox": [ + 0.647, + 0.224, + 0.716, + 0.236 + ], + "angle": 0, + "content": "Existence" + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.238, + 0.835, + 0.25 + ], + "angle": 0, + "content": "500 test cases, 2 candidates each" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.253, + 0.721, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.253, + 0.887, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.315, + 0.793, + 0.326 + ], + "angle": 0, + "content": "there is no giraffe in the image" + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.336, + 0.786, + 0.348 + ], + "angle": 0, + "content": "there is a giraffe in the image" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.358, + 0.894, + 0.415 + ], + "angle": 0, + "content": "Figure 5. An overview of the SPEC benchmark. SPEC consists of six distinct subsets, distributed across the dimensions of Size, Position, Existence and Count. Each test case consists of an image candidate set, which differs only in certain visual concept, and a text candidate set, which differs only in corresponding language concept. Due to space constraints, we present a maximum of three images and texts here, however, more comprehensive test cases are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.47, + 0.546 + ], + "angle": 0, + "content": "and the semantic space covered by the candidate set is more comprehensive. For example, in the case of relative position, the candidates include all four semantic directions (up, down, left, right), and in absolute position and count, it extends to nine candidates. As will be discussed below, more confusing candidates can be readily used as hard negatives to improve current VLMs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.398, + 0.577 + ], + "angle": 0, + "content": "4. Diagnose: Probing VLMs on SPEC" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.47, + 0.661 + ], + "angle": 0, + "content": "We conduct a systematical evaluation of four state-of-the-art VLMs: CLIP [23], BLIP [17], FLAVA [29] and CoCa [39] using our newly proposed SPEC benchmark, aiming to uncover that to what extent and in what aspect are they excelling or suffering." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.671, + 0.341, + 0.685 + ], + "angle": 0, + "content": "4.1. Evaluation Task and Protocol" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Symmetric image-text matching task. We conduct evaluations using the image-text matching task, where each test case \\(\\mathbb{T} = (\\mathcal{I},\\mathcal{T})\\) comprises an image set \\(\\mathcal{I}\\) and a corresponding text set \\(\\mathcal{T}\\), as outlined in Eq. (3). The evaluation process is symmetric, considering both visual and textual perspectives. In the image-to-text matching task, when querying with an image \\(I_{i}\\), the model is required to accurately identify \\(T_{i}\\) from the candidate set \\(\\mathcal{T}\\). Similarly, in the text-to-image matching task, the goal is to find \\(I_{i}\\) for a given query text \\(T_{i}\\). In practice, we accomplish the matching process using the similarity \\(s(I,T)\\) between a given image \\(I\\) and text \\(T\\). A candidate will be selected if its similarity with the query ranks first among the entire candidate set." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.441, + 0.893, + 0.501 + ], + "angle": 0, + "content": "Evaluation protocols. We measure the performance on SPEC using two metrics: I2TACC and T2IACC, representing the accuracy of image-to-text and text-to-image matching task, respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.522, + 0.892, + 0.561 + ], + "angle": 0, + "content": "\\[\nI 2 \\mathrm {T} _ {\\text {A C C}} = \\frac {1}{| D |} \\sum_ {\\left(\\mathcal {I} _ {i}, \\mathcal {T} _ {i}\\right) \\in D} \\frac {1}{\\left| \\mathcal {I} _ {i} \\right|} \\sum_ {I _ {j} \\in \\mathcal {I} _ {i}} h \\left(I _ {j}, \\mathcal {T} _ {i}\\right) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.583, + 0.892, + 0.622 + ], + "angle": 0, + "content": "\\[\n\\mathrm {T 2 I A C C} = \\frac {1}{| D |} \\sum_ {\\left(\\mathcal {I} _ {i}, \\mathcal {T} _ {i}\\right) \\in D} \\frac {1}{\\left| \\mathcal {T} _ {i} \\right|} \\sum_ {T _ {j} \\in \\mathcal {T} _ {i}} g \\left(T _ {j}, \\mathcal {I} _ {i}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.631, + 0.894, + 0.708 + ], + "angle": 0, + "content": "where \\(D\\) contains \\(|D|\\) test cases, \\(h(I_j,\\mathcal{T}_i)\\) equals to 1 if and only if \\(I_{j}\\) correctly find its matched text \\(T_{j}\\) from the candidate set \\(\\mathcal{T}_i\\), otherwise it is set to 0. Similarly, \\(g(T_j,\\mathcal{I}_i)\\) equals 1 if and only if \\(T_{j}\\) correctly finds its matched image \\(I_{j}\\) from the candidate set \\(\\mathcal{I}_i\\), otherwise, it is set to 0." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.716, + 0.785, + 0.732 + ], + "angle": 0, + "content": "4.2. Key Insights from SPEC Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.739, + 0.892, + 0.8 + ], + "angle": 0, + "content": "We evaluate four VLMs using the SPEC benchmark, and their results are summarized in Tab. 2. We find that all the models exhibit a limited accuracy close to random chance, from which we gain the following insights:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Even state-of-the-art VLMs perform at chance level. From the results, we surprisingly find that even the most advanced VLMs achieve only a marginal advantage compared to random chance, which sharply contrasts with their impressive performance on common tasks. For instance, CLIP [23] demonstrates a mere \\(33.4\\%\\) T2IACC for relative" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13284" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.231 + ], + "angle": 0, + "content": "
Absolute SizeRelative SizeAbsolute PositionRelative PositionExistenceCount
I2TT2ICLSI2TT2ICLSI2TT2ICLSI2TT2ICLSI2TT2ICLSI2TT2ICLS
Random33.333.31.333.333.32.511.111.11.325.025.02.550.050.05.011.111.11.3
FLAVA [29]37.337.289.332.932.384.113.115.788.825.526.784.057.951.974.414.421.275.8
BLIP [17]43.342.797.733.232.598.212.211.097.930.529.797.655.450.196.037.437.193.8
CoCa [39]39.136.195.933.633.395.411.711.896.130.328.892.250.950.083.736.535.689.6
CLIP [23]42.536.192.134.433.494.912.612.392.828.026.690.158.351.284.025.123.283.9
Ours68.960.796.340.344.197.330.634.296.946.646.996.283.453.192.555.657.892.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.245, + 0.895, + 0.304 + ], + "angle": 0, + "content": "Table 2. Evaluation results on SPEC. We extensively benchmark four state-of-the-art VLMs on SPEC to investigate their comprehension of fine-grained visual-linguistic concepts. Our evaluation employs two metrics, I2TACC and T2IACC, and we report the detailed performance on the each subset. We also report the classification accuracy, CLS, to highlight the capability in recognizing object categories. The first row indicates accuracy at chance level, serving as a baseline for comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.471, + 0.449 + ], + "angle": 0, + "content": "size recognition, while the chance-level accuracy is \\(33.3\\%\\). While BLIP performs reasonably well in absolute size, surpassing random level by around \\(9.7\\%\\), the i2TACC on relative size is \\(0.7\\%\\) lagged behind. CoCa [39] and FLAVA [29] also exhibit significant weaknesses in performance. The last row presents the performance of our improved model, demonstrating significant advancements across all metrics (as will be introduced in Sec. 5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.47, + 0.701 + ], + "angle": 0, + "content": "The challenge arises from the task itself, not the data. One might attribute the poor performance of VLMs to the data quality or distribution. To address this concern, we conduct an additional experiment. Specifically, we perform classification experiments using the SPEC dataset to assess the model's understanding of nouns or object categories. In this context, the models exhibit impressive performance, achieving approximately \\(90\\%\\) Top-1 accuracy in the 80-class classification task (denoted as CLS in Tab. 2). This aligns well with earlier findings [40] that VLMs struggle in compositional reasoning while excelling in object category recognition. The remarkable accuracy of the models in the object classification task confirms the high quality of SPEC data. This also validates that the challenges faced by VLMs stem from the tasks which require fine-grained recognition rather than issues with the data itself." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.367, + 0.727 + ], + "angle": 0, + "content": "4.3. Discussion on Model Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We attribute the poor performance of vision and language models on SPEC to their pretraining methods, specifically, the inherent limitation in standard contrastive loss. The conventional contrastive learning involves randomly sampling batches of images and texts, requiring the model to identify matching pairs within the batch. This task is intended to facilitate alignment between text and image spaces. However, as highlighted in prior studies [2, 40], the substantial differences between items in a randomly sampled batch allows the model to effortlessly complete this task. It can easily achieve this by focusing solely on nouns" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.895, + 0.42 + ], + "angle": 0, + "content": "in the text and object categories in the images through a shortcut [8]. This leads the model biased towards noun concepts, neglecting other finer-grained concepts. This is the reason why these models demonstrate poor performance on fine-grained tasks that demanding understanding concepts beyond nouns." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.43, + 0.877, + 0.449 + ], + "angle": 0, + "content": "5. Optimize: A Simple but Effective Remedy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.894, + 0.487 + ], + "angle": 0, + "content": "We experiment with CLIP [23] and propose a remedy to enhance its performance in fine-grained understanding." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.493, + 0.598, + 0.508 + ], + "angle": 0, + "content": "5.1. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.894, + 0.577 + ], + "angle": 0, + "content": "CLIP [23] consists of an visual encoder to extract image embedding: \\( e_{I} = \\mathcal{E}_{I}(I) \\) and a textual encoder to extract text embedding: \\( e_{T} = \\mathcal{E}_{T}(T) \\). The similarity score between an image \\( I \\) and a text \\( T \\) is computed following:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.583, + 0.892, + 0.619 + ], + "angle": 0, + "content": "\\[\ns (I, T) = \\exp \\left(\\frac {\\tau e _ {I} ^ {T} e _ {T}}{\\| e _ {I} \\| ^ {2} \\| e _ {T} \\| ^ {2}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.623, + 0.736, + 0.637 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a learnable temperature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.894, + 0.759 + ], + "angle": 0, + "content": "In order to guide CLIP to focus on fine-grained visual-linguistic concepts, we incorporate confusing images and text as hard negatives within the same batch. This requires CLIP to pull positives closer and push hard negatives away, thereby enhancing its ability to discern nuanced visual and textual differences. Specifically, we introduce an hard negative aware contrastive loss \\(\\mathcal{L}_{hn} = \\mathcal{L}_{hn}^{12\\mathrm{T}} + \\mathcal{L}_{hn}^{\\mathrm{T}2\\mathrm{I}}\\), which comprises an image-to-text and a text-to-image term:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.763, + 0.892, + 0.823 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {h n} ^ {\\mathrm {l 2 T}} = - \\sum_ {i} \\log \\frac {s \\left(I _ {i} , T _ {i}\\right)}{\\sum_ {T _ {j} \\in \\mathcal {T}} s \\left(I _ {i} , T _ {j}\\right) + \\sum_ {T _ {k} ^ {h n} \\in \\mathcal {T} ^ {h n}} s \\left(I _ {i} , T _ {k} ^ {h n}\\right)} \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {h n} ^ {\\mathrm {T 2 I}} = - \\sum_ {i} \\log \\frac {s \\left(I _ {i} , T _ {i}\\right)}{\\sum_ {I _ {j} \\in \\mathcal {I}} s \\left(I _ {j} , T _ {i}\\right) + \\sum_ {I _ {k} ^ {h n} \\in \\mathcal {I} ^ {h n}} s \\left(I _ {k} ^ {h n} , T _ {i}\\right)}, \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13285" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}\\) and \\(\\mathcal{T}\\) represent the trivial images and texts, respectively, while \\(\\mathcal{I}^{hn}\\) and \\(\\mathcal{T}^{hn}\\) denote non-trivial hard negatives. In our implementation, these hard negative examples are constructed using the data pipeline described in Sec. 3.2, and more details are in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.47, + 0.244 + ], + "angle": 0, + "content": "To preserve the inherent zero-shot capability of CLIP, we also leverage the conventional image-text pairs from LAION-400M [27]. We apply the standard contrastive loss of CLIP [23] to these data, introducing an additional loss term \\(\\mathcal{L}_{clip}\\). The overall loss consists of two terms:" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.253, + 0.469, + 0.27 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {c l i p}} + \\lambda \\mathcal {L} _ {\\text {h n}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.279, + 0.469, + 0.34 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyperparameter that balancing these two terms. Training on this multi-task loss enables improving the performance of CLIP in fine-grained understanding while maintaining its zero-shot capability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.348, + 0.212, + 0.364 + ], + "angle": 0, + "content": "5.2. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.469, + 0.487 + ], + "angle": 0, + "content": "Training details. We experiment with the ViT-B/32 variant of CLIP, and resume from the OpenAI pretrained checkpoint [23]. We finetune for 1,000 steps using a cosine schedule with an initial learning rate of \\(1e - 6\\) and use 800 steps for warm up. The batch size of LAION data is set to 2048, and the batch size of hard negative data is set to 768. The weight \\(\\lambda\\) of the hard negative aware loss is set to 0.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.469, + 0.786 + ], + "angle": 0, + "content": "Main results. We utilize the SPEC benchmark to assess the understanding of model in fine-grained concepts. In Tab. 3, we present the average I2TACC and T2IACC on all subsets of SPEC. To assess the general performance of the model, we also utilize the toolkit from ELEVATER [16] to evaluate the zero-shot performance on 9 classification and retrieval datasets, and report the average accuracy. Compared to CLIP [23], our model demonstrates remarkable advancements with a \\(19.8\\%\\) boost in I2TACC, an \\(18.9\\%\\) improvement in T2IACC on SPEC, and a noteworthy \\(1.2\\%\\) enhancement in zero-shot accuracy. We also conduct ablation on different training configurations. From the results in Tab. 3, it can be observed that the introducing of \\(\\mathcal{L}_{hn}\\) significantly improves the performance on SPEC. Moreover, the \\(\\mathcal{L}_{clip}\\) plays a crucial role in preserving zero-shot performance. Without it, we observe a decline in accuracy by \\(5.1\\%\\). With the combination of these two losses, we achieve substantial improvement in SPEC while maintaining the original zero-shot capability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Validation on other fine-grained benchmarks. To assess whether our approach aids the model in acquiring fundamental visual-linguistic understanding or merely leads to overfitting on SPEC, we conduct evaluations on two additional benchmarks which also focus on the assessment of fine-grained concepts. ARO [40] explores three aspects of vision-language understanding: object attributes," + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.885, + 0.201 + ], + "angle": 0, + "content": "
ConfigSPECZero-shot Accuracy
LclipLhnI2TT2I
CLIP33.530.567.5
+ Lhn✓64.360.862.4
+ Lclip✓32.231.569.4
+ Lhn+Lclip (ours)✓✓53.349.468.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.212, + 0.892, + 0.255 + ], + "angle": 0, + "content": "Table 3. Main Results: The first row represents the pretrained checkpoint without fine-tuning. We sequentially introduce two additional loss terms to investigate their impact for the performance." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.267, + 0.892, + 0.363 + ], + "angle": 0, + "content": "
EqbenARO
ImageTextGroupAttributeRelationOrder
CLIP17.621.410.163.263.953.3
CLIPFT18.123.711.165.168.054.1
Ours19.524.011.766.473.760.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.373, + 0.892, + 0.43 + ], + "angle": 0, + "content": "Table 4. Cross-dataset evaluation results on Eqben [33] and ARO [40]. To demonstrate that the improvement comes from the negative loss, rather than training on more data, we also report \\(\\mathrm{CLIP}_{FT}\\), which is also finetuned but without negative samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.636 + ], + "angle": 0, + "content": "inter-object relations, and word ordering. Eqben [33] focuses on minimal visual semantic changes, aiming to diagnose VLMs in understanding fine-grained concepts such as counting and location. In Tab. 4, we present the experimental results, demonstrating a clear improvement compared to CLIP [23] on both datasets. For example, compared to CLIP, our method shows an average improvement of \\(2\\%\\) on Eqben and respective enhancements of \\(3.2\\%\\), \\(9.8\\%\\), and \\(7.4\\%\\) on the three subsets of ARO. The consistent improvement on these datasets demonstrates that our approach has facilitated the model in acquiring transferable fine-grained understanding capabilities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.619, + 0.665 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.87 + ], + "angle": 0, + "content": "In this study, we explored the comprehension abilities of Visual Language Models (VLMs) with respect to fine-grained visual-linguistic concepts. We first established an efficient pipeline to synthesize candidate images that exclusively differ in a particular visual attribute. Leveraging this pipeline, we created the SPEC benchmark to diagnose the comprehension proficiency of VLMs in terms of object size, position, existence, and count. Upon evaluating four leading VLMs using SPEC, we uncovered substantial performance limitations. To address this, we introduced an enhancement strategy that effectively optimizes the model for fine-grained understanding, while maintaining its original zero-shot capability." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement This project was supported by NSFC under Grant No. 62102092." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13286" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Paola Cascante-Bonilla, Khaled Shehada, James Smith, Sivan Doveh, Donghyun Kim, Rameswar Panda, Gül Varol, Aude Oliva, Vicente Ordonez, Rogério Schmidt Feris, and Leonid Karlinsky. Going beyond nouns with vision & language models using synthetic data. In ICCV, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.473, + 0.383 + ], + "angle": 0, + "content": "[3] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel M. Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model. In ICLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.384, + 0.472, + 0.44 + ], + "angle": 0, + "content": "[4] Yuxin Chen, Zongyang Ma, Ziqi Zhang, Zhongang Qi, Chunfen Yuan, Ying Shan, Bing Li, Weiming Hu, Xiaohu Qie, and Jianping Wu. Vilem: Visual-language error modeling for image-text retrieval. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.442, + 0.472, + 0.482 + ], + "angle": 0, + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.485, + 0.472, + 0.526 + ], + "angle": 0, + "content": "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.528, + 0.472, + 0.597 + ], + "angle": 0, + "content": "[7] Sivan Doveh, Assaf Arbelle, Sivan Harary, Rameswar Panda, Roei Herzig, Eli Schwartz, Donghyun Kim, Raja Giryes, Rogério Schmidt Feris, Shimon Ullman, and Leonid Karlinsky. Teaching structured vision & language concepts to vision & language models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.599, + 0.472, + 0.655 + ], + "angle": 0, + "content": "[8] Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard S. Zemel, Wieland Brendel, Matthias Bethge, and Felix Wichmann. Shortcut learning in deep neural networks. Nat. Mach. Intell., 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.657, + 0.472, + 0.712 + ], + "angle": 0, + "content": "[9] Golnaz Ghiasi, Yin Cui, A. Srinivas, Rui Qian, Tsung-Yi Lin, Ekin Dogus Cubuk, Quoc V. Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. CVPR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.472, + 0.756 + ], + "angle": 0, + "content": "[10] Shashank Goel, Hritik Bansal, Sumit Kaur Bhatia, Ryan A. Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.472, + 0.785 + ], + "angle": 0, + "content": "[11] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.472, + 0.828 + ], + "angle": 0, + "content": "[12] Lisa Anne Hendricks and Aida Nematzadeh. Probing image-language transformers for verb understanding. In Findings, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.472, + 0.885 + ], + "angle": 0, + "content": "[13] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.887, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[14] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. ArXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In ICCV, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[16] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Yong Jae Lee, Houdong Hu, Zicheng Liu, and Jianfeng Gao. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. In NeurIPS, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 1, 2, 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 4, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[19] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. TACL, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.407, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[20] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. ArXiv, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.464, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[21] Letitia Parcalabescu, Michele Cafagna, Lilitta Muradjan, Anette Frank, Iacer Calixto, and Albert Gatt. Valse: A task-independent benchmark for vision and language models centered on linguistic phenomena. In ACL, 2021. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.521, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[22] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. ArXiv, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.577, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 6, 7, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.649, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[24] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.691, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[25] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[26] Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Galle, et al. Bloom: A 176b-parameter open-access multilingual language model. ArXiv, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. ArXiv, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13287" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[28] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[29] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022. 1, 2, 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[30] James Smith, Paola Cascante-Bonilla, Assaf Arbelle, Donghyun Kim, Rameswar Panda, David Cox, Diyi Yang, Zsolt Kira, Rogério Schmidt Feris, and Leonid Karlinsky. Construct-vl: Data-free continual structured vl concepts learning*. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[31] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In CVPR, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.469, + 0.438 + ], + "angle": 0, + "content": "[32] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnivl: One foundation model for image-language and video-language tasks. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[33] Tan Wang, Kevin Lin, Linjie Li, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Equivariant similarity for vision-language foundation models. In ICCV, 2023. 1, 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[34] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[35] Zuxuan Wu, Zejia Weng, Wujian Peng, Xitong Yang, Ang Li, Larry S Davis, and Yu-Gang Jiang. Building an open-vocabulary video clip model with better architectures, optimization and data. TPAMI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.65 + ], + "angle": 0, + "content": "[36] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. arXiv preprint arXiv:2310.10647, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.692 + ], + "angle": 0, + "content": "[37] Zhen Xing, Qi Dai, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Simda: Simple diffusion adapter for efficient video generation. In CVPR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[38] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[39] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Trans. Mach. Learn. Res., 2022. 2, 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[40] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bag-of-words models, and what to do about it? ICLR, 2023. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[41] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. In ICML, 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[42] Hanqing Zhao, Dianmo Sheng, Jianmin Bao, Dongdong Chen, Dong Chen, Fang Wen, Lu Yuan, Ce Liu, Wenbo Zhou, Qi Chu, Weiming Zhang, and Neng H. Yu. X-paste: Revisiting scalable copy-paste for instance segmentation using clip and stablediffusion. In ICML, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.23 + ], + "angle": 0, + "content": "[43] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. ArXiv, 2022. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13288" + } + ] +] \ No newline at end of file diff --git a/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/eca2ceba-84d3-4751-b38a-1ba7e74c92ba_origin.pdf b/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/eca2ceba-84d3-4751-b38a-1ba7e74c92ba_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..353fb8f711f80251488c0a80764932641c6e866e --- /dev/null +++ b/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/eca2ceba-84d3-4751-b38a-1ba7e74c92ba_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b673a08dc30a60a813c34a2147e14953393d365709fb079088e4f55bb9cbf69 +size 2015116 diff --git a/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/full.md b/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ff9a635ca29fa19e7e96a5044e372f4a783f5d9a --- /dev/null +++ b/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/full.md @@ -0,0 +1,343 @@ +# Synthesize, Diagnose, and Optimize: Towards Fine-Grained Vision-Language Understanding + +Wujian Peng $^{1,2}$ Sicheng Xie $^{1,2}$ Zuyao You $^{1,2}$ Shiyi Lan $^{3}$ Zuxuan Wu $^{1,2\dagger}$ + +$^{1}$ Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University $^{2}$ Shanghai Collaborative Innovation Center of Intelligent Visual Computing $^{3}$ NVIDIA + +# Abstract + +Vision language models (VLM) have demonstrated remarkable performance across various downstream tasks. However, understanding fine-grained visual-linguistic concepts, such as attributes and inter-object relationships, remains a significant challenge. While several benchmarks aim to evaluate VLMs in finer granularity, their primary focus remains on the linguistic aspect, neglecting the visual dimension. Here, we highlight the importance of evaluating VLMs from both a textual and visual perspective. We introduce a progressive pipeline to synthesize images that vary in a specific attribute while ensuring consistency in all other aspects. Utilizing this data engine, we carefully design a benchmark, SPEC, to diagnose the comprehension of object size, position, existence, and count. Subsequently, we conduct a thorough evaluation of four leading VLMs on SPEC. Surprisingly, their performance is close to random guess, revealing significant limitations. With this in mind, we propose a simple yet effective approach to optimize VLMs in fine-grained understanding, achieving significant improvements on SPEC without compromising the zero-shot performance. Results on two additional fine-grained benchmarks also show consistent improvements, further validating the transferability of our approach. Code and data are available at https://github.com/wjpoom/SPEC. + +# 1. Introduction + +Vision and Language Foundation Models (VLMs) pretrained on large-scale image-text data [13, 17, 23, 29, 41] have consistently demonstrated impressive performance across a wide range of well-established evaluating tasks, i.e. image classification [5], image captioning [1, 18], visual question answering [3] and cross-modal image-text retrieval [18, 38]. Their remarkable performance is gradually convincing the community that these currently avail + +![](images/081244d44c780210bb529ca213ef1e18fcdb4ba519bbcf54f44a0ff5b1800355.jpg) +(a) Image to Text Matching + +![](images/f2783ff56bdb731df68494df20710ac98437830a11fb60c3a08bca652b7c8fd3.jpg) +(b) Text to Image Matching +Figure 1. We conduct a symmetrical assessment of VLMs in fine-grained comprehension, considering both linguistic and visual perspectives. The bars in $\square$ represent the image-text matching scores for CLIP [23] and our method, respectively. It is evident that CLIP struggles with tasks related to quantity comprehension, whereas our method significantly enhances the model in understanding fine-grained details. + +able VLMs are almost robust and powerful enough to be transferred to a broad spectrum of downstream tasks, either through finetuning or even in a zero-shot manner. + +However, recent research has shattered this captivating illusion, revealing that even state-of-the-art VLMs [17, 23, 29, 41] exhibit significant limitations in understanding visual-linguistic concepts that require fine-grained compositional reasoning, especially in tasks involving object attributes or inter-object relationships [21, 31, 33, 40, 43]. This raises a crucial question: to what extent and in what aspects are VLMs excelling or struggling? To answer this, previous effort evaluates fine-grained capabilities of VLMs through the image-to-text matching task, as shown in Fig. 1(a). This involves providing a query image and retrieving the matching text from a set of confusing candi + +dates, differing subtly in texts. For example, when assessing the counting ability, it is crucial to ensure that quantity is the unique variable and other clues are kept the same. Therefore, a straightforward way to do so is modifying quantity and adjective words used in the texts. While manipulating texts to construct confusing candidate sets has been well-studied due to the sparsity of the text space and advancements in Large Language Models (LLMs) [19, 21, 40, 43], the visual side remains relatively under explored, largely due to the complexity of visual signals and the absence of powerful tools. We posit that exploring the visual dimension in a fine-grained manner is also essential for a comprehensive understanding of VLMs. + +Motivated by the great progress achieved in generative models, we present an effective framework for generating high-quality image candidates that are suitable for evaluating the performance of VLMs. This framework ensures that images within the same candidate set only differ in the specified property of interest, while all other properties remain consistent. We break down this task into several simple and manageable steps. As illustrated in Fig. 2, we start by utilizing a text-to-image model [22, 24] to generate images featuring a single object. Then, a segmenter [14, 15] is employed to separate the objects from their backgrounds, yielding a library of foreground instance spanning various categories. From there, we select instances and arrange them on a blank canvas (manipulating attributes such as size, position, existence, and quantity of a specific object at this stage is straightforward). Finally, we use an inpainting model [22, 24] to fill the missing background portions, producing an photo-realistic image. It is worth noting that, during the inpainting process, we design a progressive background filling strategy, effectively ensuring consistency in the background across all images in the same candidate set. + +Empowered by this data construction pipeline, we carefully develop a new benchmark, named as SPEC, to evaluate the proficiency of VLMs in comprehending fine-grained concepts including Size, Position, Existence and Counting. We systematical test four VLMs [17, 23, 29, 41] on this newly created test bed. Surprisingly, even state-of-the-art models perform at chance-level, exposing significant performance deficiencies. Following this, we implement a straightforward approach to remedy this by incorporating hard negative examples (i.e., confusing images or texts within the same candidate set) into the same training batch. This encourages the model to discern subtle differences among candidate examples, leading to a significant improvement in performance on SEPC while preserving the original zero-shot capability. Furthermore, to demonstrate the model's generalization ability, we conduct additional tests on two existing datasets [33, 40], which also focus on compositional reasoning. The consistent improvement further validates that our method effectively guide the model + +to acquire essential and transferable comprehending abilities at a finer granularity. Our main contributions are: + +1. A progressive data constructing pipeline. We present a progressive data construction pipeline designed for creating a candidate image set. Within each candidate set, images vary exclusively in a specified attribute while ensuring consistency across other aspects. Such data are valuable for conducting text-to-image matching tasks, as depicted in Fig. 1(b), offering a visual perspective for evaluating VLMs. +2. A carefully curated benchmark: SPEC. We meticulously craft a novel benchmark, SPEC, with a specific focus on evaluating VLMs' understanding of fine-grained visual-linguistic concepts, encompassing object size, position, existence, and count. The introduction of SPEC enables a symmetrical evaluation of VLMs from both image and text perspectives, addressing the previous lack of image-centric testing data. +3. A simple and effective remedy. We evaluate four VLMs on SPEC, revealing significant limitations. In response, we propose a method to enhance the understanding of fine-grained visual-linguistic concepts. Experimental results indicate notable improvement not only on SPEC but also consistent results on two additional datasets, while preserving zero-shot capability. + +# 2. Related Work + +Vision and Language Models (VLMs). Models such as CLIP [23], ALIGN [13], CyCLIP [10] CoCa [39], OmniVL [32] and Open-VCLIP [34, 35] have demonstrated impressive performance across a wide range of downstream tasks. These models include two separate unimodal encoders, each designed to extract representations for visual and textual input. To achieve alignment between the two modalities, they typically employ a huge number of visual-textual pairs for contrastive learning. By pretraining on 400M noisy data, CLIP [23] achieves a top-1 accuracy on ImageNet-1K [5] comparable to that of ResNet-50 [11], even though it is never specifically trained on ImageNet and is evaluated in a zero-shot manner. However, as highlighted in recent work [31, 40], these advancements are primarily attributed to the simplicity of evaluation tasks which requires no reasoning or compositional capabilities. The performance of these models are limited on tasks that require fine-grained understanding [21, 40]. + +Benchmarking VLMs in Finer Granularity. To assess the model's understanding of nuanced visual-linguistic concepts, several new benchmarks have been proposed. Winoground [31] is curated by experts with a focus on compositional understanding. VALID [21] and VL-Checklist [43] investigate several linguistic phenomena by transforming real captions into confusing alternatives. + +![](images/098b3614aa297badaebef7d1fb29a1ea41a5cdd78623dd20599ba4c251485aaf.jpg) +Figure 2. The overall illustration of our data progressive construction pipeline. We initiate the process by generating a batch of images containing a single object. Subsequently, we extract the object from the background in the images. Following that, we arrange the background-free images on a blank canvas according to specifications (with control over attributes). Finally, we meticulously fill in the missing background, ensuring consistency across candidates. + +ARO [40] diagnoses VLMs in attribution, relation and ordering. Eqben [33] assesses whether the model is sensitive to visual semantic changes. Most existing benchmarks solely focus on subtle textual changes [12, 19, 21, 40, 43], as crafting confusing text candidates is straightforward that can be achieved through either LLMs or simple rules. Winoground [40] and Eqben [33] are most relevant to us, since we focus on minimal semantic changes in both images and texts, enabling a more comprehensive evaluation across modalities. However, the scale of Winoground is restricted by its costly curation, and the image diversity of Eqben is limited by virtual engines. In contrast, our data construction pipeline is scalable and can produce diverse images. + +Enhancing VLMs for Fine-grained Understanding. To mitigate challenges in fine-grained recognition, various approaches have been explored. Syn-CLIP [2] utilize data synthesized by 3D simulation engines to enhance the model's understanding of concepts beyond nouns. EQSIM [33] incorporates an additional regularization loss to generalize VLMs to nuanced multimodal compositions. TSVLC [7] and ViLEM [4] introduce negative texts generated by LLMs [6, 26] to inject fine-grained knowledge. Construct-VL [30] addresses these challenges from a continual learning perspective. These methods compel the model to focus on subtle differences by introducing confusing texts as hard negatives. However, we argue that the absence of visual hard negatives limits its performance. Thus, we introduce hard negatives for both modalities, simultaneously enhancing the visual and textual encoders. + +# 3. Synthesize: Data Construction Pipeline + +Our goal is to build a set of perplexing image, wherein each image differs solely in a specified attribute while ensuring + +![](images/2ea48f0364ac828571b2f925099d178a4c131109275b05698d3a7d15ecaa7ab7.jpg) +Figure 3. Ensuring consistency among candidates is crucial to avoid ambiguity. The images above not only differ in quantity but also show a significant variation in the appearance of the objects. Consequently, attributing the model's correctness or errors solely to the understanding of quantity is not convincing. + +consistency in all other aspects. We first emphasize the importance of preserving consistency among candidate images for effective evaluation (Sec. 3.1). To address this, we break down this problem and introduce a progressive data construction pipeline (Sec. 3.2). Then, we carefully devise a benchmark that centers on evaluating VLMs' grasp of fine-grained visual-linguistic concepts (Sec. 3.3). + +# 3.1. Importance of Candidate Consistency + +As illustrated in Fig. 3, when conducting a matching task with the textual query "a photo of two dogs", the model might mistakenly choose the image on the right (which actually contains three dogs). However, attributing this error solely to counting difficulty is not convincing. The model might select the right-side image due to its more photorealistic appearance or better alignment with the word "dog", as the left-side image has a cartoon style. Conversely, if the model correctly selects the right-side image + +for the query "a photo of three dogs", we also cannot assert that the model is proficient in counting. These obscure ambiguities arise because there is no guarantee of the uniqueness of changing factors among candidate images during evaluation. Therefore, ensuring consistency among candidates in all aspects except the one under investigation is crucial. To this end, we propose a progressive data construction method, which will be elaborated in detail as follows. + +# 3.2. Progressive Data Construction + +The data construction framework is illustrated in Fig. 2, comprising four progressive steps. Initially, we generate images featuring a single prominent object. Then, we isolate the foreground from the background, resulting in a library of foreground objects spanning various categories. Subsequently, we select objects from this library and arrange them on a blank canvas, adjusting their attributes and relationships. Lastly, we carefully fill in the missing background, ensuring consistency among different candidates. + +# 3.2.1 Generating Images with Single Objects + +We initiate the process by utilizing a generation model to obtain a collection of images, each featuring a single and prominent object corresponding to a specific category. Due to the progress in visual generation models [24, 25, 36, 37], these images display a high level of photo-realistic and diverse content. In practice, we use Stable-Diffusion-XL 1.0 [22] as our generator, and prompt it with "a photo of a single and fully visible [class name]". The emphasis on "single and fully visible" is crucial, as the model might otherwise generate images with multiple objects or encounter occlusion issues, as observed in [42]. The [class name] represents a specific category from 80 classes of COCO [18]. + +# 3.2.2 Isolating Objects from the Background + +For ease in subsequent processes, we need to separate the objects from the backgrounds where they are embedded. To accomplish this, we first utilize an open-set detector, Grounding-DINO [20] to outline the regions containing the objects. Subsequently, we prompt SAM [15] with this bounding box as to obtain the final segmentation results. Thus far, we have established a library containing instances from various categories. These instances are background-free, allowing for composition on a blank canvas, while their attributes and relations can be controlled. + +# 3.2.3 Arranging Objects on the Canvas + +Recall that our goal is to manipulate a specific visual-linguistic concept of an image. Currently, this task appears straightforward when we exclude the background from consideration. We can retrieve instances from the library and + +arrange them on a blank canvas according to our specifications. For example, we can flexibly control the quantity of an object through duplication operations, modify their sizes via resizing, and determine whether an object exists and specify the position of an existing object. This process resembles the concept of copy-paste [9, 42], with a notable distinction: we paste objects onto a blank background and placing emphasis on controlling properties such as the size, position, existence, and quantity of each individual object. + +# 3.2.4 Infilling the Missing Background + +We have constructed images with differences in specified attributes, however, they currently lack a suitable background. To fill the missing area, we employ a inpainting model [22] which demonstrates proficiency in filling large holes. It is worth noting that generating backgrounds individually for each image would result in significant differences in the backgrounds, posing a challenge to maintaining consistency among candidate images. To overcome this, we introduce a strategy where images within the same candidate set share a common and consistent background during the inpainting process. As depicted in the upper part of Fig. 4, to present the giraffe in different positions, we start by surrounding it with an initial background. Then, we relocate this initial image on the canvas as required, and fill the remaining blank space. Similarly, in the lower part, we first embed the zebra into a reasonable environment. Following that, we expand the scene horizontally, introduce the elephant, and fill the blank areas through inpainting. In summary, we begin the process by generating an initial background, which is then expanded to the surroundings, effectively ensuring consistency among the candidate images. Additional examples, such as adjusting the size or quantity of an object, can be found in the supplementary. + +# 3.3. SPEC Benchmark + +Utilizing the data engine outlined in Sec. 3.2, we carefully devise the SPEC benchmark with the goal of assessing the performance of VLMs in comprehending object size, position, existence and count. An overview of the SPEC benchmark is presented in Fig. 5. SPEC contains six subsets, which will be elaborated as follows: + +Absolute Size reflects how large an object is in comparison to the entire image. We categorize this into three levels: large, medium, or small, and define them following: + +$$ +\operatorname {S i z e} _ {a b s.} (x) = \left\{ \begin{array}{l l} \text {s m a l l}, & P \leq 0. 2 \\ \text {m e d i u m}, & 0. 4 \leq P \leq 0. 6 \\ \text {l a r g e}, & P \geq 0. 8, \end{array} \right. \tag {1} +$$ + +where $P$ denotes the proportion of the space occupied by the object relative to the area of the entire image. A safety + +![](images/b05e55f7ee605039c607c86b96c6f89711416fdd85bfa7870ef808f2b2ba3646.jpg) +Figure 4. Consistent background inpainting strategy. We first generate an initial background shared by all candidate images. Then, we expand around this region, ensuring consistency in the backgrounds of different images. + +threshold is deliberately introduced between these three levels to prevent ambiguity. + +Relative Size focuses on the size relationship between two objects. It is categorized as follows: object A is [smaller than, equal to, larger than] object B, and we measure this following: + +$$ +\operatorname {S i z e} _ {r e a.} (A, B) = \left\{ \begin{array}{l l} A \text {s m a l l e r t h a n} B, & R \leq 0. 5 \\ A \text {e q u a l t o} B, & 0. 9 \leq R \leq 1. 1 \\ A \text {l a r g e r t h a n} B, & R \geq 2, \end{array} \right. \tag {2} +$$ + +where $R = \frac{S_A}{S_B}$ is the ratio of the areas of object A and object B. + +Absolute Position signifies the location of an object relative to the image. We partition the image into a $3 \times 3$ grid, defining nine possible positions: top-left, top, top-right, left, center, right, bottom-left, bottom, and bottom-right. The absolute position of an object is determined based on the grid in which its center point resides. + +Relative Position describes the spatial relationship between two objects. We consider four common spatial relationships: A is [to the left of, to the right of, above, below] B. The position relationship between objects is defined based on the relative positions of their center points. + +Existence indicates whether an object appears in a given image, expressed using existential quantifiers: There is [no, at least one] object in the image. + +Count represents the number of occurrences of an object, providing a metric for the model's quantitative understand- + +
BenchmarksVisual HNScalabilityText-realisticPhoto-realistic#Candidates
FALSE [21]X✓✓✓2
ARO [40]X✓X✓2
Winoground [31]✓X✓✓2
Eqben [33]✓✓✓✓2
SPEC(Ours)✓✓✓✓2-9
+ +Table 1. Comparison of SPEC with other fine-grained benchmarks. + +ing. Due to potential occlusion issues with a large number of objects, we restrict our consideration to the range of 1 to 9: “there are [one, two, ..., nine] object(s) in the image”. + +Data Format. The basic unit of SPEC is an individual test case, wherein each test case comprises two components: an image candidate set, which differs only in certain visual aspects, and a text candidate set, which differs only in the corresponding language descriptions. We formally represent a test case as: + +$$ +\mathbb {T}: \left(\mathcal {I} = \left\{I _ {1}, \dots , I _ {K} \right\}, \mathcal {T} = \left\{T _ {1}, \dots , T _ {K} \right\}\right), \tag {3} +$$ + +where $\mathcal{I}$ and $\mathcal{T}$ represent the image and text candidate sets, respectively. The i-th image $I_{i}$ is paired with the i-th text $T_{i}$ , i.e., they mutually describe each other. $K$ is the semantic cardinality of the test case, determined by the definition of each subset. For instance, if a test case belongs to the absolute size subset, then $K = 3$ (representing the three semantics: large, medium, small). In Fig. 5, we present examplar test cases for each subset, and more examples can be found in the supplementary. + +Comparing SPEC with other benchmarks. In Tab. 1, we conduct a comprehensive comparison of SPEC with four similar benchmarks. Visual HN indicates the presence of image hard negatives, which are essential for the text-to-image matching task. Notably, VALID [21] and ARO [40] concentrate solely on text hard negatives, neglecting their visual counterparts. Scalability assesses whether the data construction method can be scaled up. For instance, Winoground [31] is limited to 400 examples due to high costs for manual collection. Additionally, Text-realistic evaluates the grammatical correctness of the texts, where ARO involves directly swapping the positions of two words without considering grammar. Similarly, Image-realistic indicates the realism of the images. Eqben [33] incorporates images rendered using a virtual engine, compromising their quality. In contrast, images from SPEC, while synthesized, leverage advanced image generation models and our effective data construction pipeline, resulting in photo-realistic images. Finally, we compare the number of Candidates in each test case. In all datasets except SPEC, each example features only two candidates, i.e., identifying the correct item from two candidates, which is relatively straightforward. In contrast, tasks in SPEC are more challenging, + +# Absolute Size + +![](images/2a5cca9bedece61d7b999815b34ce7ac2dea1412d8bb1daae44e397ddc16950e.jpg) +500 test cases, 3 candidates each +the sheep is large in the image +the sheep is medium-sized in the image +the sheep is small in the image + +# Relative Size + +![](images/db313293c0402a989307f4edaa6c3e30780b1a41d711493d1f76ea959442cd28.jpg) +500 test cases, 3 candidates each +the pizza is smaller than the sandwich +the pizza is equal-size with the sandwich +the pizza is bigger than the sandwich + +# Absolute Position + +![](images/13e47e4bb7989688a9307c7f2827199f1e8543b9fd20f37027c3240fd228bc92.jpg) +500 test cases, 9 candidates each + +![](images/c3d9f019e69011da1f4102204c48986e13edea0b2e6f3f61ebe4f430a25b3d30.jpg) +the kite is in the top-left corner +the kite is in the center of the image +the kite is in the bottom-left corner + +![](images/56f20d2194fb1fde36d7314976f430b56dc15e7c459540efda1f20b16698c026.jpg) + +# Relative Position + +![](images/c21fcae67e0124aa3a071d61b520e21550e1ffe7c385e9170a2f9633e7906b2f.jpg) +500 test cases, 4 candidates each +the toilet is to the left of the backpack the toilet is on top of the backpack the toilet is below the backpack + +![](images/274c2c6cc8881b0357cadbaca0405e14ca0c8627ba32bed87dc9ac4408d7dffe.jpg) + +![](images/8036839ee38258c4555b0bf141d748f6937890ba953253a1530c0e996a1a4639.jpg) + +# Count + +![](images/cb3b518742e3132392c0f1f50a8edc610c0d64f9f051b5ccf8828be52e02c186.jpg) +500 test cases, 9 candidates each +a photo of one snowboard a photo of two snowboards a photo of three snowboards + +![](images/d8dc1b298fccc219b7c60a7330ed00f5d2bc8720ff5aed3997cc369c765fc6ff.jpg) + +![](images/b685079b70d99267165d64f1c466b33ccfebc6ff8953471ce928dd0bb8fef1fb.jpg) + +# Existence + +![](images/fc0c3b73e183f4c03a7f8d157c1073d5e08b73822d12e4c437da95d27788534e.jpg) +500 test cases, 2 candidates each +there is no giraffe in the image +there is a giraffe in the image + +![](images/d2a8c259b75cedb5576255e8975b8ee82f1201bbcbca5679a9d71900fe4b5d87.jpg) +Figure 5. An overview of the SPEC benchmark. SPEC consists of six distinct subsets, distributed across the dimensions of Size, Position, Existence and Count. Each test case consists of an image candidate set, which differs only in certain visual concept, and a text candidate set, which differs only in corresponding language concept. Due to space constraints, we present a maximum of three images and texts here, however, more comprehensive test cases are available in the supplementary material. + +and the semantic space covered by the candidate set is more comprehensive. For example, in the case of relative position, the candidates include all four semantic directions (up, down, left, right), and in absolute position and count, it extends to nine candidates. As will be discussed below, more confusing candidates can be readily used as hard negatives to improve current VLMs. + +# 4. Diagnose: Probing VLMs on SPEC + +We conduct a systematical evaluation of four state-of-the-art VLMs: CLIP [23], BLIP [17], FLAVA [29] and CoCa [39] using our newly proposed SPEC benchmark, aiming to uncover that to what extent and in what aspect are they excelling or suffering. + +# 4.1. Evaluation Task and Protocol + +Symmetric image-text matching task. We conduct evaluations using the image-text matching task, where each test case $\mathbb{T} = (\mathcal{I},\mathcal{T})$ comprises an image set $\mathcal{I}$ and a corresponding text set $\mathcal{T}$ , as outlined in Eq. (3). The evaluation process is symmetric, considering both visual and textual perspectives. In the image-to-text matching task, when querying with an image $I_{i}$ , the model is required to accurately identify $T_{i}$ from the candidate set $\mathcal{T}$ . Similarly, in the text-to-image matching task, the goal is to find $I_{i}$ for a given query text $T_{i}$ . In practice, we accomplish the matching process using the similarity $s(I,T)$ between a given image $I$ and text $T$ . A candidate will be selected if its similarity with the query ranks first among the entire candidate set. + +Evaluation protocols. We measure the performance on SPEC using two metrics: I2TACC and T2IACC, representing the accuracy of image-to-text and text-to-image matching task, respectively: + +$$ +I 2 \mathrm {T} _ {\text {A C C}} = \frac {1}{| D |} \sum_ {\left(\mathcal {I} _ {i}, \mathcal {T} _ {i}\right) \in D} \frac {1}{\left| \mathcal {I} _ {i} \right|} \sum_ {I _ {j} \in \mathcal {I} _ {i}} h \left(I _ {j}, \mathcal {T} _ {i}\right) \tag {4} +$$ + +$$ +\mathrm {T 2 I A C C} = \frac {1}{| D |} \sum_ {\left(\mathcal {I} _ {i}, \mathcal {T} _ {i}\right) \in D} \frac {1}{\left| \mathcal {T} _ {i} \right|} \sum_ {T _ {j} \in \mathcal {T} _ {i}} g \left(T _ {j}, \mathcal {I} _ {i}\right), \tag {5} +$$ + +where $D$ contains $|D|$ test cases, $h(I_j,\mathcal{T}_i)$ equals to 1 if and only if $I_{j}$ correctly find its matched text $T_{j}$ from the candidate set $\mathcal{T}_i$ , otherwise it is set to 0. Similarly, $g(T_j,\mathcal{I}_i)$ equals 1 if and only if $T_{j}$ correctly finds its matched image $I_{j}$ from the candidate set $\mathcal{I}_i$ , otherwise, it is set to 0. + +# 4.2. Key Insights from SPEC Results + +We evaluate four VLMs using the SPEC benchmark, and their results are summarized in Tab. 2. We find that all the models exhibit a limited accuracy close to random chance, from which we gain the following insights: + +Even state-of-the-art VLMs perform at chance level. From the results, we surprisingly find that even the most advanced VLMs achieve only a marginal advantage compared to random chance, which sharply contrasts with their impressive performance on common tasks. For instance, CLIP [23] demonstrates a mere $33.4\%$ T2IACC for relative + +
Absolute SizeRelative SizeAbsolute PositionRelative PositionExistenceCount
I2TT2ICLSI2TT2ICLSI2TT2ICLSI2TT2ICLSI2TT2ICLSI2TT2ICLS
Random33.333.31.333.333.32.511.111.11.325.025.02.550.050.05.011.111.11.3
FLAVA [29]37.337.289.332.932.384.113.115.788.825.526.784.057.951.974.414.421.275.8
BLIP [17]43.342.797.733.232.598.212.211.097.930.529.797.655.450.196.037.437.193.8
CoCa [39]39.136.195.933.633.395.411.711.896.130.328.892.250.950.083.736.535.689.6
CLIP [23]42.536.192.134.433.494.912.612.392.828.026.690.158.351.284.025.123.283.9
Ours68.960.796.340.344.197.330.634.296.946.646.996.283.453.192.555.657.892.5
+ +Table 2. Evaluation results on SPEC. We extensively benchmark four state-of-the-art VLMs on SPEC to investigate their comprehension of fine-grained visual-linguistic concepts. Our evaluation employs two metrics, I2TACC and T2IACC, and we report the detailed performance on the each subset. We also report the classification accuracy, CLS, to highlight the capability in recognizing object categories. The first row indicates accuracy at chance level, serving as a baseline for comparison. + +size recognition, while the chance-level accuracy is $33.3\%$ . While BLIP performs reasonably well in absolute size, surpassing random level by around $9.7\%$ , the i2TACC on relative size is $0.7\%$ lagged behind. CoCa [39] and FLAVA [29] also exhibit significant weaknesses in performance. The last row presents the performance of our improved model, demonstrating significant advancements across all metrics (as will be introduced in Sec. 5). + +The challenge arises from the task itself, not the data. One might attribute the poor performance of VLMs to the data quality or distribution. To address this concern, we conduct an additional experiment. Specifically, we perform classification experiments using the SPEC dataset to assess the model's understanding of nouns or object categories. In this context, the models exhibit impressive performance, achieving approximately $90\%$ Top-1 accuracy in the 80-class classification task (denoted as CLS in Tab. 2). This aligns well with earlier findings [40] that VLMs struggle in compositional reasoning while excelling in object category recognition. The remarkable accuracy of the models in the object classification task confirms the high quality of SPEC data. This also validates that the challenges faced by VLMs stem from the tasks which require fine-grained recognition rather than issues with the data itself. + +# 4.3. Discussion on Model Limitations + +We attribute the poor performance of vision and language models on SPEC to their pretraining methods, specifically, the inherent limitation in standard contrastive loss. The conventional contrastive learning involves randomly sampling batches of images and texts, requiring the model to identify matching pairs within the batch. This task is intended to facilitate alignment between text and image spaces. However, as highlighted in prior studies [2, 40], the substantial differences between items in a randomly sampled batch allows the model to effortlessly complete this task. It can easily achieve this by focusing solely on nouns + +in the text and object categories in the images through a shortcut [8]. This leads the model biased towards noun concepts, neglecting other finer-grained concepts. This is the reason why these models demonstrate poor performance on fine-grained tasks that demanding understanding concepts beyond nouns. + +# 5. Optimize: A Simple but Effective Remedy + +We experiment with CLIP [23] and propose a remedy to enhance its performance in fine-grained understanding. + +# 5.1. Method + +CLIP [23] consists of an visual encoder to extract image embedding: $e_{I} = \mathcal{E}_{I}(I)$ and a textual encoder to extract text embedding: $e_{T} = \mathcal{E}_{T}(T)$ . The similarity score between an image $I$ and a text $T$ is computed following: + +$$ +s (I, T) = \exp \left(\frac {\tau e _ {I} ^ {T} e _ {T}}{\| e _ {I} \| ^ {2} \| e _ {T} \| ^ {2}}\right), \tag {6} +$$ + +where $\tau$ is a learnable temperature. + +In order to guide CLIP to focus on fine-grained visual-linguistic concepts, we incorporate confusing images and text as hard negatives within the same batch. This requires CLIP to pull positives closer and push hard negatives away, thereby enhancing its ability to discern nuanced visual and textual differences. Specifically, we introduce an hard negative aware contrastive loss $\mathcal{L}_{hn} = \mathcal{L}_{hn}^{12\mathrm{T}} + \mathcal{L}_{hn}^{\mathrm{T}2\mathrm{I}}$ , which comprises an image-to-text and a text-to-image term: + +$$ +\mathcal {L} _ {h n} ^ {\mathrm {l 2 T}} = - \sum_ {i} \log \frac {s \left(I _ {i} , T _ {i}\right)}{\sum_ {T _ {j} \in \mathcal {T}} s \left(I _ {i} , T _ {j}\right) + \sum_ {T _ {k} ^ {h n} \in \mathcal {T} ^ {h n}} s \left(I _ {i} , T _ {k} ^ {h n}\right)} \tag {7} +$$ + +$$ +\mathcal {L} _ {h n} ^ {\mathrm {T 2 I}} = - \sum_ {i} \log \frac {s \left(I _ {i} , T _ {i}\right)}{\sum_ {I _ {j} \in \mathcal {I}} s \left(I _ {j} , T _ {i}\right) + \sum_ {I _ {k} ^ {h n} \in \mathcal {I} ^ {h n}} s \left(I _ {k} ^ {h n} , T _ {i}\right)}, \tag {8} +$$ + +where $\mathcal{I}$ and $\mathcal{T}$ represent the trivial images and texts, respectively, while $\mathcal{I}^{hn}$ and $\mathcal{T}^{hn}$ denote non-trivial hard negatives. In our implementation, these hard negative examples are constructed using the data pipeline described in Sec. 3.2, and more details are in the supplementary. + +To preserve the inherent zero-shot capability of CLIP, we also leverage the conventional image-text pairs from LAION-400M [27]. We apply the standard contrastive loss of CLIP [23] to these data, introducing an additional loss term $\mathcal{L}_{clip}$ . The overall loss consists of two terms: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {c l i p}} + \lambda \mathcal {L} _ {\text {h n}}, \tag {9} +$$ + +where $\lambda$ is a hyperparameter that balancing these two terms. Training on this multi-task loss enables improving the performance of CLIP in fine-grained understanding while maintaining its zero-shot capability. + +# 5.2. Experiments + +Training details. We experiment with the ViT-B/32 variant of CLIP, and resume from the OpenAI pretrained checkpoint [23]. We finetune for 1,000 steps using a cosine schedule with an initial learning rate of $1e - 6$ and use 800 steps for warm up. The batch size of LAION data is set to 2048, and the batch size of hard negative data is set to 768. The weight $\lambda$ of the hard negative aware loss is set to 0.2. + +Main results. We utilize the SPEC benchmark to assess the understanding of model in fine-grained concepts. In Tab. 3, we present the average I2TACC and T2IACC on all subsets of SPEC. To assess the general performance of the model, we also utilize the toolkit from ELEVATER [16] to evaluate the zero-shot performance on 9 classification and retrieval datasets, and report the average accuracy. Compared to CLIP [23], our model demonstrates remarkable advancements with a $19.8\%$ boost in I2TACC, an $18.9\%$ improvement in T2IACC on SPEC, and a noteworthy $1.2\%$ enhancement in zero-shot accuracy. We also conduct ablation on different training configurations. From the results in Tab. 3, it can be observed that the introducing of $\mathcal{L}_{hn}$ significantly improves the performance on SPEC. Moreover, the $\mathcal{L}_{clip}$ plays a crucial role in preserving zero-shot performance. Without it, we observe a decline in accuracy by $5.1\%$ . With the combination of these two losses, we achieve substantial improvement in SPEC while maintaining the original zero-shot capability. + +Validation on other fine-grained benchmarks. To assess whether our approach aids the model in acquiring fundamental visual-linguistic understanding or merely leads to overfitting on SPEC, we conduct evaluations on two additional benchmarks which also focus on the assessment of fine-grained concepts. ARO [40] explores three aspects of vision-language understanding: object attributes, + +
ConfigSPECZero-shot Accuracy
LclipLhnI2TT2I
CLIP33.530.567.5
+ Lhn✓64.360.862.4
+ Lclip✓32.231.569.4
+ Lhn+Lclip (ours)✓✓53.349.468.7
+ +Table 3. Main Results: The first row represents the pretrained checkpoint without fine-tuning. We sequentially introduce two additional loss terms to investigate their impact for the performance. + +
EqbenARO
ImageTextGroupAttributeRelationOrder
CLIP17.621.410.163.263.953.3
CLIPFT18.123.711.165.168.054.1
Ours19.524.011.766.473.760.7
+ +Table 4. Cross-dataset evaluation results on Eqben [33] and ARO [40]. To demonstrate that the improvement comes from the negative loss, rather than training on more data, we also report $\mathrm{CLIP}_{FT}$ , which is also finetuned but without negative samples. + +inter-object relations, and word ordering. Eqben [33] focuses on minimal visual semantic changes, aiming to diagnose VLMs in understanding fine-grained concepts such as counting and location. In Tab. 4, we present the experimental results, demonstrating a clear improvement compared to CLIP [23] on both datasets. For example, compared to CLIP, our method shows an average improvement of $2\%$ on Eqben and respective enhancements of $3.2\%$ , $9.8\%$ , and $7.4\%$ on the three subsets of ARO. The consistent improvement on these datasets demonstrates that our approach has facilitated the model in acquiring transferable fine-grained understanding capabilities. + +# 6. Conclusion + +In this study, we explored the comprehension abilities of Visual Language Models (VLMs) with respect to fine-grained visual-linguistic concepts. We first established an efficient pipeline to synthesize candidate images that exclusively differ in a particular visual attribute. Leveraging this pipeline, we created the SPEC benchmark to diagnose the comprehension proficiency of VLMs in terms of object size, position, existence, and count. Upon evaluating four leading VLMs using SPEC, we uncovered substantial performance limitations. To address this, we introduced an enhancement strategy that effectively optimizes the model for fine-grained understanding, while maintaining its original zero-shot capability. + +Acknowledgement This project was supported by NSFC under Grant No. 62102092. + +# References + +[1] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In ICCV, 2019. 1 +[2] Paola Cascante-Bonilla, Khaled Shehada, James Smith, Sivan Doveh, Donghyun Kim, Rameswar Panda, Gül Varol, Aude Oliva, Vicente Ordonez, Rogério Schmidt Feris, and Leonid Karlinsky. Going beyond nouns with vision & language models using synthetic data. In ICCV, 2023. 3, 7 +[3] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel M. Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model. In ICLR, 2022. 1 +[4] Yuxin Chen, Zongyang Ma, Ziqi Zhang, Zhongang Qi, Chunfen Yuan, Ying Shan, Bing Li, Weiming Hu, Xiaohu Qie, and Jianping Wu. Vilem: Visual-language error modeling for image-text retrieval. In CVPR, 2023. 3 +[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1, 2 +[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 3 +[7] Sivan Doveh, Assaf Arbelle, Sivan Harary, Rameswar Panda, Roei Herzig, Eli Schwartz, Donghyun Kim, Raja Giryes, Rogério Schmidt Feris, Shimon Ullman, and Leonid Karlinsky. Teaching structured vision & language concepts to vision & language models. In CVPR, 2023. 3 +[8] Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard S. Zemel, Wieland Brendel, Matthias Bethge, and Felix Wichmann. Shortcut learning in deep neural networks. Nat. Mach. Intell., 2020. 7 +[9] Golnaz Ghiasi, Yin Cui, A. Srinivas, Rui Qian, Tsung-Yi Lin, Ekin Dogus Cubuk, Quoc V. Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. CVPR, 2020. 4 +[10] Shashank Goel, Hritik Bansal, Sumit Kaur Bhatia, Ryan A. Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. In NeurIPS, 2022. 2 +[11] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2015. 2 +[12] Lisa Anne Hendricks and Aida Nematzadeh. Probing image-language transformers for verb understanding. In Findings, 2021. 3 +[13] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1, 2 +[14] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing + +Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. ArXiv, 2023. 2 +[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In ICCV, 2023. 2, 4 +[16] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Yong Jae Lee, Houdong Hu, Zicheng Liu, and Jianfeng Gao. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. In NeurIPS, 2022. 8 +[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 1, 2, 6, 7, 3 +[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 4, 3 +[19] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. TACL, 2023. 2, 3 +[20] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. ArXiv, 2023. 4 +[21] Letitia Parcalabescu, Michele Cafagna, Lilitta Muradjan, Anette Frank, Iacer Calixto, and Albert Gatt. Valse: A task-independent benchmark for vision and language models centered on linguistic phenomena. In ACL, 2021. 1, 2, 3, 5 +[22] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. ArXiv, 2023. 2, 4 +[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 6, 7, 8, 3 +[24] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In 2021. 2, 4 +[25] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 4 +[26] Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Galle, et al. Bloom: A 176b-parameter open-access multilingual language model. ArXiv, 2022. 3 +[27] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. ArXiv, 2021. 8 + +[28] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 3 +[29] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022. 1, 2, 6, 7, 3 +[30] James Smith, Paola Cascante-Bonilla, Assaf Arbelle, Donghyun Kim, Rameswar Panda, David Cox, Diyi Yang, Zsolt Kira, Rogério Schmidt Feris, and Leonid Karlinsky. Construct-vl: Data-free continual structured vl concepts learning*. In CVPR, 2023. 3 +[31] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In CVPR, 2022. 1, 2, 5 +[32] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnivl: One foundation model for image-language and video-language tasks. In NeurIPS, 2022. 2 +[33] Tan Wang, Kevin Lin, Linjie Li, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Equivariant similarity for vision-language foundation models. In ICCV, 2023. 1, 2, 3, 5, 8 +[34] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization. In ICML, 2023. 2 +[35] Zuxuan Wu, Zejia Weng, Wujian Peng, Xitong Yang, Ang Li, Larry S Davis, and Yu-Gang Jiang. Building an open-vocabulary video clip model with better architectures, optimization and data. TPAMI, 2024. 2 +[36] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. arXiv preprint arXiv:2310.10647, 2023. 4 +[37] Zhen Xing, Qi Dai, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Simda: Simple diffusion adapter for efficient video generation. In CVPR, 2024. 4 +[38] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 1 +[39] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Trans. Mach. Learn. Res., 2022. 2, 6, 7, 3 +[40] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bag-of-words models, and what to do about it? ICLR, 2023. 1, 2, 3, 5, 7, 8 +[41] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. In ICML, 2021. 1, 2 + +[42] Hanqing Zhao, Dianmo Sheng, Jianmin Bao, Dongdong Chen, Dong Chen, Fang Wen, Lu Yuan, Ce Liu, Wenbo Zhou, Qi Chu, Weiming Zhang, and Neng H. Yu. X-paste: Revisiting scalable copy-paste for instance segmentation using clip and stablediffusion. In ICML, 2022. 4 +[43] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. ArXiv, 2022. 1, 2, 3 \ No newline at end of file diff --git a/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/images.zip b/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..89bc40eb6608c152e71c7a48bc3c177330e8e30f --- /dev/null +++ b/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d94f0904b79d55913d6d88a56fc8f3560819e564957dffcf71b896d622f23318 +size 480361 diff --git a/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/layout.json b/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..94f42427a8cc17654f305947821451b20a00665d --- /dev/null +++ b/2024/Synthesize Diagnose and Optimize_ Towards Fine-Grained Vision-Language Understanding/layout.json @@ -0,0 +1,8555 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 102, + 483, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 102, + 483, + 140 + ], + "spans": [ + { + "bbox": [ + 111, + 102, + 483, + 140 + ], + "type": "text", + "content": "Synthesize, Diagnose, and Optimize: Towards Fine-Grained Vision-Language Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 159, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 159, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 109, + 159, + 482, + 176 + ], + "type": "text", + "content": "Wujian Peng" + }, + { + "bbox": [ + 109, + 159, + 482, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 109, + 159, + 482, + 176 + ], + "type": "text", + "content": " Sicheng Xie" + }, + { + "bbox": [ + 109, + 159, + 482, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 109, + 159, + 482, + 176 + ], + "type": "text", + "content": " Zuyao You" + }, + { + "bbox": [ + 109, + 159, + 482, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 109, + 159, + 482, + 176 + ], + "type": "text", + "content": " Shiyi Lan" + }, + { + "bbox": [ + 109, + 159, + 482, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 109, + 159, + 482, + 176 + ], + "type": "text", + "content": " Zuxuan Wu" + }, + { + "bbox": [ + 109, + 159, + 482, + 176 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 181, + 487, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 181, + 487, + 223 + ], + "spans": [ + { + "bbox": [ + 107, + 181, + 487, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 181, + 487, + 223 + ], + "type": "text", + "content": "Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University " + }, + { + "bbox": [ + 107, + 181, + 487, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 181, + 487, + 223 + ], + "type": "text", + "content": "Shanghai Collaborative Innovation Center of Intelligent Visual Computing " + }, + { + "bbox": [ + 107, + 181, + 487, + 223 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 181, + 487, + 223 + ], + "type": "text", + "content": "NVIDIA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 242, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 242, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 242, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 267, + 290, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 290, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 290, + 555 + ], + "type": "text", + "content": "Vision language models (VLM) have demonstrated remarkable performance across various downstream tasks. However, understanding fine-grained visual-linguistic concepts, such as attributes and inter-object relationships, remains a significant challenge. While several benchmarks aim to evaluate VLMs in finer granularity, their primary focus remains on the linguistic aspect, neglecting the visual dimension. Here, we highlight the importance of evaluating VLMs from both a textual and visual perspective. We introduce a progressive pipeline to synthesize images that vary in a specific attribute while ensuring consistency in all other aspects. Utilizing this data engine, we carefully design a benchmark, SPEC, to diagnose the comprehension of object size, position, existence, and count. Subsequently, we conduct a thorough evaluation of four leading VLMs on SPEC. Surprisingly, their performance is close to random guess, revealing significant limitations. With this in mind, we propose a simple yet effective approach to optimize VLMs in fine-grained understanding, achieving significant improvements on SPEC without compromising the zero-shot performance. Results on two additional fine-grained benchmarks also show consistent improvements, further validating the transferability of our approach. Code and data are available at https://github.com/wjpoom/SPEC." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 597, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 287, + 695 + ], + "type": "text", + "content": "Vision and Language Foundation Models (VLMs) pretrained on large-scale image-text data [13, 17, 23, 29, 41] have consistently demonstrated impressive performance across a wide range of well-established evaluating tasks, i.e. image classification [5], image captioning [1, 18], visual question answering [3] and cross-modal image-text retrieval [18, 38]. Their remarkable performance is gradually convincing the community that these currently avail" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 241, + 545, + 304 + ], + "blocks": [ + { + "bbox": [ + 310, + 241, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 310, + 241, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 310, + 241, + 545, + 304 + ], + "type": "image", + "image_path": "081244d44c780210bb529ca213ef1e18fcdb4ba519bbcf54f44a0ff5b1800355.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 308, + 480, + 319 + ], + "lines": [ + { + "bbox": [ + 389, + 308, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 389, + 308, + 480, + 319 + ], + "type": "text", + "content": "(a) Image to Text Matching" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 308, + 332, + 545, + 415 + ], + "blocks": [ + { + "bbox": [ + 308, + 332, + 545, + 415 + ], + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 415 + ], + "type": "image", + "image_path": "f2783ff56bdb731df68494df20710ac98437830a11fb60c3a08bca652b7c8fd3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 420, + 478, + 431 + ], + "lines": [ + { + "bbox": [ + 386, + 420, + 478, + 431 + ], + "spans": [ + { + "bbox": [ + 386, + 420, + 478, + 431 + ], + "type": "text", + "content": "(b) Text to Image Matching" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 439, + 547, + 517 + ], + "lines": [ + { + "bbox": [ + 304, + 439, + 547, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 439, + 547, + 517 + ], + "type": "text", + "content": "Figure 1. We conduct a symmetrical assessment of VLMs in fine-grained comprehension, considering both linguistic and visual perspectives. The bars in " + }, + { + "bbox": [ + 304, + 439, + 547, + 517 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 304, + 439, + 547, + 517 + ], + "type": "text", + "content": " represent the image-text matching scores for CLIP [23] and our method, respectively. It is evident that CLIP struggles with tasks related to quantity comprehension, whereas our method significantly enhances the model in understanding fine-grained details." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 530, + 546, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 546, + 567 + ], + "type": "text", + "content": "able VLMs are almost robust and powerful enough to be transferred to a broad spectrum of downstream tasks, either through finetuning or even in a zero-shot manner." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "However, recent research has shattered this captivating illusion, revealing that even state-of-the-art VLMs [17, 23, 29, 41] exhibit significant limitations in understanding visual-linguistic concepts that require fine-grained compositional reasoning, especially in tasks involving object attributes or inter-object relationships [21, 31, 33, 40, 43]. This raises a crucial question: to what extent and in what aspects are VLMs excelling or struggling? To answer this, previous effort evaluates fine-grained capabilities of VLMs through the image-to-text matching task, as shown in Fig. 1(a). This involves providing a query image and retrieving the matching text from a set of confusing candi" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 702, + 141, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 702, + 141, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 702, + 141, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13279" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "dates, differing subtly in texts. For example, when assessing the counting ability, it is crucial to ensure that quantity is the unique variable and other clues are kept the same. Therefore, a straightforward way to do so is modifying quantity and adjective words used in the texts. While manipulating texts to construct confusing candidate sets has been well-studied due to the sparsity of the text space and advancements in Large Language Models (LLMs) [19, 21, 40, 43], the visual side remains relatively under explored, largely due to the complexity of visual signals and the absence of powerful tools. We posit that exploring the visual dimension in a fine-grained manner is also essential for a comprehensive understanding of VLMs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 231, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 289, + 483 + ], + "type": "text", + "content": "Motivated by the great progress achieved in generative models, we present an effective framework for generating high-quality image candidates that are suitable for evaluating the performance of VLMs. This framework ensures that images within the same candidate set only differ in the specified property of interest, while all other properties remain consistent. We break down this task into several simple and manageable steps. As illustrated in Fig. 2, we start by utilizing a text-to-image model [22, 24] to generate images featuring a single object. Then, a segmenter [14, 15] is employed to separate the objects from their backgrounds, yielding a library of foreground instance spanning various categories. From there, we select instances and arrange them on a blank canvas (manipulating attributes such as size, position, existence, and quantity of a specific object at this stage is straightforward). Finally, we use an inpainting model [22, 24] to fill the missing background portions, producing an photo-realistic image. It is worth noting that, during the inpainting process, we design a progressive background filling strategy, effectively ensuring consistency in the background across all images in the same candidate set." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": "Empowered by this data construction pipeline, we carefully develop a new benchmark, named as SPEC, to evaluate the proficiency of VLMs in comprehending fine-grained concepts including Size, Position, Existence and Counting. We systematical test four VLMs [17, 23, 29, 41] on this newly created test bed. Surprisingly, even state-of-the-art models perform at chance-level, exposing significant performance deficiencies. Following this, we implement a straightforward approach to remedy this by incorporating hard negative examples (i.e., confusing images or texts within the same candidate set) into the same training batch. This encourages the model to discern subtle differences among candidate examples, leading to a significant improvement in performance on SEPC while preserving the original zero-shot capability. Furthermore, to demonstrate the model's generalization ability, we conduct additional tests on two existing datasets [33, 40], which also focus on compositional reasoning. The consistent improvement further validates that our method effectively guide the model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "to acquire essential and transferable comprehending abilities at a finer granularity. Our main contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 96, + 547, + 372 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 307, + 96, + 547, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 191 + ], + "type": "text", + "content": "1. A progressive data constructing pipeline. We present a progressive data construction pipeline designed for creating a candidate image set. Within each candidate set, images vary exclusively in a specified attribute while ensuring consistency across other aspects. Such data are valuable for conducting text-to-image matching tasks, as depicted in Fig. 1(b), offering a visual perspective for evaluating VLMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 192, + 547, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 192, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 306, + 192, + 547, + 287 + ], + "type": "text", + "content": "2. A carefully curated benchmark: SPEC. We meticulously craft a novel benchmark, SPEC, with a specific focus on evaluating VLMs' understanding of fine-grained visual-linguistic concepts, encompassing object size, position, existence, and count. The introduction of SPEC enables a symmetrical evaluation of VLMs from both image and text perspectives, addressing the previous lack of image-centric testing data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 287, + 547, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 287, + 547, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 287, + 547, + 372 + ], + "type": "text", + "content": "3. A simple and effective remedy. We evaluate four VLMs on SPEC, revealing significant limitations. In response, we propose a method to enhance the understanding of fine-grained visual-linguistic concepts. Experimental results indicate notable improvement not only on SPEC but also consistent results on two additional datasets, while preserving zero-shot capability." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 380, + 392, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 392, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 392, + 392 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 408, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 545, + 624 + ], + "type": "text", + "content": "Vision and Language Models (VLMs). Models such as CLIP [23], ALIGN [13], CyCLIP [10] CoCa [39], OmniVL [32] and Open-VCLIP [34, 35] have demonstrated impressive performance across a wide range of downstream tasks. These models include two separate unimodal encoders, each designed to extract representations for visual and textual input. To achieve alignment between the two modalities, they typically employ a huge number of visual-textual pairs for contrastive learning. By pretraining on 400M noisy data, CLIP [23] achieves a top-1 accuracy on ImageNet-1K [5] comparable to that of ResNet-50 [11], even though it is never specifically trained on ImageNet and is evaluated in a zero-shot manner. However, as highlighted in recent work [31, 40], these advancements are primarily attributed to the simplicity of evaluation tasks which requires no reasoning or compositional capabilities. The performance of these models are limited on tasks that require fine-grained understanding [21, 40]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": "Benchmarking VLMs in Finer Granularity. To assess the model's understanding of nuanced visual-linguistic concepts, several new benchmarks have been proposed. Winoground [31] is curated by experts with a focus on compositional understanding. VALID [21] and VL-Checklist [43] investigate several linguistic phenomena by transforming real captions into confusing alternatives." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13280" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 547, + 244 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 547, + 244 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 547, + 244 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 547, + 244 + ], + "type": "image", + "image_path": "098b3614aa297badaebef7d1fb29a1ea41a5cdd78623dd20599ba4c251485aaf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 548, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 548, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 548, + 298 + ], + "type": "text", + "content": "Figure 2. The overall illustration of our data progressive construction pipeline. We initiate the process by generating a batch of images containing a single object. Subsequently, we extract the object from the background in the images. Following that, we arrange the background-free images on a blank canvas according to specifications (with control over attributes). Finally, we meticulously fill in the missing background, ensuring consistency across candidates." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 305, + 289, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 289, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 289, + 462 + ], + "type": "text", + "content": "ARO [40] diagnoses VLMs in attribution, relation and ordering. Eqben [33] assesses whether the model is sensitive to visual semantic changes. Most existing benchmarks solely focus on subtle textual changes [12, 19, 21, 40, 43], as crafting confusing text candidates is straightforward that can be achieved through either LLMs or simple rules. Winoground [40] and Eqben [33] are most relevant to us, since we focus on minimal semantic changes in both images and texts, enabling a more comprehensive evaluation across modalities. However, the scale of Winoground is restricted by its costly curation, and the image diversity of Eqben is limited by virtual engines. In contrast, our data construction pipeline is scalable and can produce diverse images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 467, + 289, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 289, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 289, + 660 + ], + "type": "text", + "content": "Enhancing VLMs for Fine-grained Understanding. To mitigate challenges in fine-grained recognition, various approaches have been explored. Syn-CLIP [2] utilize data synthesized by 3D simulation engines to enhance the model's understanding of concepts beyond nouns. EQSIM [33] incorporates an additional regularization loss to generalize VLMs to nuanced multimodal compositions. TSVLC [7] and ViLEM [4] introduce negative texts generated by LLMs [6, 26] to inject fine-grained knowledge. Construct-VL [30] addresses these challenges from a continual learning perspective. These methods compel the model to focus on subtle differences by introducing confusing texts as hard negatives. However, we argue that the absence of visual hard negatives limits its performance. Thus, we introduce hard negatives for both modalities, simultaneously enhancing the visual and textual encoders." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 670, + 264, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 264, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 264, + 685 + ], + "type": "text", + "content": "3. Synthesize: Data Construction Pipeline" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Our goal is to build a set of perplexing image, wherein each image differs solely in a specified attribute while ensuring" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 342, + 304, + 511, + 413 + ], + "blocks": [ + { + "bbox": [ + 342, + 304, + 511, + 413 + ], + "lines": [ + { + "bbox": [ + 342, + 304, + 511, + 413 + ], + "spans": [ + { + "bbox": [ + 342, + 304, + 511, + 413 + ], + "type": "image", + "image_path": "2ea48f0364ac828571b2f925099d178a4c131109275b05698d3a7d15ecaa7ab7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 422, + 547, + 479 + ], + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 479 + ], + "type": "text", + "content": "Figure 3. Ensuring consistency among candidates is crucial to avoid ambiguity. The images above not only differ in quantity but also show a significant variation in the appearance of the objects. Consequently, attributing the model's correctness or errors solely to the understanding of quantity is not convincing." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 491, + 547, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 547, + 578 + ], + "type": "text", + "content": "consistency in all other aspects. We first emphasize the importance of preserving consistency among candidate images for effective evaluation (Sec. 3.1). To address this, we break down this problem and introduce a progressive data construction pipeline (Sec. 3.2). Then, we carefully devise a benchmark that centers on evaluating VLMs' grasp of fine-grained visual-linguistic concepts (Sec. 3.3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 586, + 506, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 506, + 601 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 506, + 601 + ], + "type": "text", + "content": "3.1. Importance of Candidate Consistency" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "As illustrated in Fig. 3, when conducting a matching task with the textual query \"a photo of two dogs\", the model might mistakenly choose the image on the right (which actually contains three dogs). However, attributing this error solely to counting difficulty is not convincing. The model might select the right-side image due to its more photorealistic appearance or better alignment with the word \"dog\", as the left-side image has a cartoon style. Conversely, if the model correctly selects the right-side image" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "13281" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "for the query \"a photo of three dogs\", we also cannot assert that the model is proficient in counting. These obscure ambiguities arise because there is no guarantee of the uniqueness of changing factors among candidate images during evaluation. Therefore, ensuring consistency among candidates in all aspects except the one under investigation is crucial. To this end, we propose a progressive data construction method, which will be elaborated in detail as follows." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 175, + 214, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 214, + 188 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 214, + 188 + ], + "type": "text", + "content": "3.2. Progressive Data Construction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 194, + 289, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 302 + ], + "type": "text", + "content": "The data construction framework is illustrated in Fig. 2, comprising four progressive steps. Initially, we generate images featuring a single prominent object. Then, we isolate the foreground from the background, resulting in a library of foreground objects spanning various categories. Subsequently, we select objects from this library and arrange them on a blank canvas, adjusting their attributes and relationships. Lastly, we carefully fill in the missing background, ensuring consistency among different candidates." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 316, + 248, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 248, + 329 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 248, + 329 + ], + "type": "text", + "content": "3.2.1 Generating Images with Single Objects" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 335, + 289, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 289, + 479 + ], + "type": "text", + "content": "We initiate the process by utilizing a generation model to obtain a collection of images, each featuring a single and prominent object corresponding to a specific category. Due to the progress in visual generation models [24, 25, 36, 37], these images display a high level of photo-realistic and diverse content. In practice, we use Stable-Diffusion-XL 1.0 [22] as our generator, and prompt it with \"a photo of a single and fully visible [class name]\". The emphasis on \"single and fully visible\" is crucial, as the model might otherwise generate images with multiple objects or encounter occlusion issues, as observed in [42]. The [class name] represents a specific category from 80 classes of COCO [18]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 494, + 247, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 494, + 247, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 494, + 247, + 506 + ], + "type": "text", + "content": "3.2.2 Isolating Objects from the Background" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 513, + 289, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 513, + 289, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 513, + 289, + 632 + ], + "type": "text", + "content": "For ease in subsequent processes, we need to separate the objects from the backgrounds where they are embedded. To accomplish this, we first utilize an open-set detector, Grounding-DINO [20] to outline the regions containing the objects. Subsequently, we prompt SAM [15] with this bounding box as to obtain the final segmentation results. Thus far, we have established a library containing instances from various categories. These instances are background-free, allowing for composition on a blank canvas, while their attributes and relations can be controlled." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 647, + 224, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 224, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 224, + 659 + ], + "type": "text", + "content": "3.2.3 Arranging Objects on the Canvas" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Recall that our goal is to manipulate a specific visual-linguistic concept of an image. Currently, this task appears straightforward when we exclude the background from consideration. We can retrieve instances from the library and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "content": "arrange them on a blank canvas according to our specifications. For example, we can flexibly control the quantity of an object through duplication operations, modify their sizes via resizing, and determine whether an object exists and specify the position of an existing object. This process resembles the concept of copy-paste [9, 42], with a notable distinction: we paste objects onto a blank background and placing emphasis on controlling properties such as the size, position, existence, and quantity of each individual object." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 194, + 480, + 206 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 194, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 305, + 194, + 480, + 206 + ], + "type": "text", + "content": "3.2.4 Infilling the Missing Background" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 213, + 547, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 489 + ], + "type": "text", + "content": "We have constructed images with differences in specified attributes, however, they currently lack a suitable background. To fill the missing area, we employ a inpainting model [22] which demonstrates proficiency in filling large holes. It is worth noting that generating backgrounds individually for each image would result in significant differences in the backgrounds, posing a challenge to maintaining consistency among candidate images. To overcome this, we introduce a strategy where images within the same candidate set share a common and consistent background during the inpainting process. As depicted in the upper part of Fig. 4, to present the giraffe in different positions, we start by surrounding it with an initial background. Then, we relocate this initial image on the canvas as required, and fill the remaining blank space. Similarly, in the lower part, we first embed the zebra into a reasonable environment. Following that, we expand the scene horizontally, introduce the elephant, and fill the blank areas through inpainting. In summary, we begin the process by generating an initial background, which is then expanded to the surroundings, effectively ensuring consistency among the candidate images. Additional examples, such as adjusting the size or quantity of an object, can be found in the supplementary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 495, + 415, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 415, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 415, + 506 + ], + "type": "text", + "content": "3.3. SPEC Benchmark" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 513, + 547, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 547, + 586 + ], + "type": "text", + "content": "Utilizing the data engine outlined in Sec. 3.2, we carefully devise the SPEC benchmark with the goal of assessing the performance of VLMs in comprehending object size, position, existence and count. An overview of the SPEC benchmark is presented in Fig. 5. SPEC contains six subsets, which will be elaborated as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 547, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 631 + ], + "type": "text", + "content": "Absolute Size reflects how large an object is in comparison to the entire image. We categorize this into three levels: large, medium, or small, and define them following:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 335, + 637, + 545, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 637, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 335, + 637, + 545, + 682 + ], + "type": "interline_equation", + "content": "\\operatorname {S i z e} _ {a b s.} (x) = \\left\\{ \\begin{array}{l l} \\text {s m a l l}, & P \\leq 0. 2 \\\\ \\text {m e d i u m}, & 0. 4 \\leq P \\leq 0. 6 \\\\ \\text {l a r g e}, & P \\geq 0. 8, \\end{array} \\right. \\tag {1}", + "image_path": "33bceb37f7df3842eef8a1efccb871f81a4a19502e72476186df32e32d2e04f1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": " denotes the proportion of the space occupied by the object relative to the area of the entire image. A safety" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13282" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 287, + 251 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 287, + 251 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 287, + 251 + ], + "type": "image", + "image_path": "b05e55f7ee605039c607c86b96c6f89711416fdd85bfa7870ef808f2b2ba3646.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 258, + 289, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 258, + 289, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 289, + 304 + ], + "type": "text", + "content": "Figure 4. Consistent background inpainting strategy. We first generate an initial background shared by all candidate images. Then, we expand around this region, ensuring consistency in the backgrounds of different images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 322, + 287, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 287, + 347 + ], + "type": "text", + "content": "threshold is deliberately introduced between these three levels to prevent ambiguity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 354, + 289, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 403 + ], + "type": "text", + "content": "Relative Size focuses on the size relationship between two objects. It is categorized as follows: object A is [smaller than, equal to, larger than] object B, and we measure this following:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 409, + 298, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 409, + 298, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 409, + 298, + 464 + ], + "type": "interline_equation", + "content": "\\operatorname {S i z e} _ {r e a.} (A, B) = \\left\\{ \\begin{array}{l l} A \\text {s m a l l e r t h a n} B, & R \\leq 0. 5 \\\\ A \\text {e q u a l t o} B, & 0. 9 \\leq R \\leq 1. 1 \\\\ A \\text {l a r g e r t h a n} B, & R \\geq 2, \\end{array} \\right. \\tag {2}", + "image_path": "40e588127d38268a627aab4bb3527a478e1fcf182c3ff01364f8e57ee09a42dc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 464, + 289, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 289, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 289, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 464, + 289, + 489 + ], + "type": "inline_equation", + "content": "R = \\frac{S_A}{S_B}" + }, + { + "bbox": [ + 46, + 464, + 289, + 489 + ], + "type": "text", + "content": " is the ratio of the areas of object A and object B." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 496, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 571 + ], + "type": "text", + "content": "Absolute Position signifies the location of an object relative to the image. We partition the image into a " + }, + { + "bbox": [ + 46, + 496, + 287, + 571 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 496, + 287, + 571 + ], + "type": "text", + "content": " grid, defining nine possible positions: top-left, top, top-right, left, center, right, bottom-left, bottom, and bottom-right. The absolute position of an object is determined based on the grid in which its center point resides." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 577, + 288, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 577, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 577, + 288, + 639 + ], + "type": "text", + "content": "Relative Position describes the spatial relationship between two objects. We consider four common spatial relationships: A is [to the left of, to the right of, above, below] B. The position relationship between objects is defined based on the relative positions of their center points." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 645, + 288, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 645, + 288, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 645, + 288, + 683 + ], + "type": "text", + "content": "Existence indicates whether an object appears in a given image, expressed using existential quantifiers: There is [no, at least one] object in the image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "text", + "content": "Count represents the number of occurrences of an object, providing a metric for the model's quantitative understand-" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 547, + 133 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 547, + 133 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 547, + 133 + ], + "type": "table", + "html": "
BenchmarksVisual HNScalabilityText-realisticPhoto-realistic#Candidates
FALSE [21]X✓✓✓2
ARO [40]X✓X✓2
Winoground [31]✓X✓✓2
Eqben [33]✓✓✓✓2
SPEC(Ours)✓✓✓✓2-9
", + "image_path": "20d9acfd11a90b1187e5f182e6f9beaffa0be2b83f9a304b60e8cf48f0dc612d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 140, + 547, + 161 + ], + "lines": [ + { + "bbox": [ + 305, + 140, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 547, + 161 + ], + "type": "text", + "content": "Table 1. Comparison of SPEC with other fine-grained benchmarks." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 181, + 547, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 181, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 305, + 181, + 547, + 217 + ], + "type": "text", + "content": "ing. Due to potential occlusion issues with a large number of objects, we restrict our consideration to the range of 1 to 9: “there are [one, two, ..., nine] object(s) in the image”." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 224, + 547, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 224, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 547, + 297 + ], + "type": "text", + "content": "Data Format. The basic unit of SPEC is an individual test case, wherein each test case comprises two components: an image candidate set, which differs only in certain visual aspects, and a text candidate set, which differs only in the corresponding language descriptions. We formally represent a test case as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 331, + 303, + 545, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 303, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 331, + 303, + 545, + 317 + ], + "type": "interline_equation", + "content": "\\mathbb {T}: \\left(\\mathcal {I} = \\left\\{I _ {1}, \\dots , I _ {K} \\right\\}, \\mathcal {T} = \\left\\{T _ {1}, \\dots , T _ {K} \\right\\}\\right), \\tag {3}", + "image_path": "a8cebd5d33571d858a597cfc7565a4fc09db0d34409f0672b91d27f484754a80.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 322, + 547, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 322, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 322, + 547, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 322, + 547, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 305, + 322, + 547, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 322, + 547, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 305, + 322, + 547, + 431 + ], + "type": "text", + "content": " represent the image and text candidate sets, respectively. The i-th image " + }, + { + "bbox": [ + 305, + 322, + 547, + 431 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 305, + 322, + 547, + 431 + ], + "type": "text", + "content": " is paired with the i-th text " + }, + { + "bbox": [ + 305, + 322, + 547, + 431 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 305, + 322, + 547, + 431 + ], + "type": "text", + "content": ", i.e., they mutually describe each other. " + }, + { + "bbox": [ + 305, + 322, + 547, + 431 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 322, + 547, + 431 + ], + "type": "text", + "content": " is the semantic cardinality of the test case, determined by the definition of each subset. For instance, if a test case belongs to the absolute size subset, then " + }, + { + "bbox": [ + 305, + 322, + 547, + 431 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 305, + 322, + 547, + 431 + ], + "type": "text", + "content": " (representing the three semantics: large, medium, small). In Fig. 5, we present examplar test cases for each subset, and more examples can be found in the supplementary." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 437, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 715 + ], + "type": "text", + "content": "Comparing SPEC with other benchmarks. In Tab. 1, we conduct a comprehensive comparison of SPEC with four similar benchmarks. Visual HN indicates the presence of image hard negatives, which are essential for the text-to-image matching task. Notably, VALID [21] and ARO [40] concentrate solely on text hard negatives, neglecting their visual counterparts. Scalability assesses whether the data construction method can be scaled up. For instance, Winoground [31] is limited to 400 examples due to high costs for manual collection. Additionally, Text-realistic evaluates the grammatical correctness of the texts, where ARO involves directly swapping the positions of two words without considering grammar. Similarly, Image-realistic indicates the realism of the images. Eqben [33] incorporates images rendered using a virtual engine, compromising their quality. In contrast, images from SPEC, while synthesized, leverage advanced image generation models and our effective data construction pipeline, resulting in photo-realistic images. Finally, we compare the number of Candidates in each test case. In all datasets except SPEC, each example features only two candidates, i.e., identifying the correct item from two candidates, which is relatively straightforward. In contrast, tasks in SPEC are more challenging," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13283" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 70, + 113, + 80 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 70, + 113, + 80 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 113, + 80 + ], + "type": "text", + "content": "Absolute Size" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 92, + 200, + 142 + ], + "blocks": [ + { + "bbox": [ + 55, + 81, + 169, + 91 + ], + "lines": [ + { + "bbox": [ + 55, + 81, + 169, + 91 + ], + "spans": [ + { + "bbox": [ + 55, + 81, + 169, + 91 + ], + "type": "text", + "content": "500 test cases, 3 candidates each" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 92, + 200, + 142 + ], + "lines": [ + { + "bbox": [ + 50, + 92, + 200, + 142 + ], + "spans": [ + { + "bbox": [ + 50, + 92, + 200, + 142 + ], + "type": "image", + "image_path": "2a5cca9bedece61d7b999815b34ce7ac2dea1412d8bb1daae44e397ddc16950e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 142, + 168, + 169 + ], + "lines": [ + { + "bbox": [ + 50, + 142, + 168, + 169 + ], + "spans": [ + { + "bbox": [ + 50, + 142, + 168, + 169 + ], + "type": "text", + "content": "the sheep is large in the image \nthe sheep is medium-sized in the image \nthe sheep is small in the image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 175, + 111, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 111, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 111, + 185 + ], + "type": "text", + "content": "Relative Size" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 198, + 200, + 248 + ], + "blocks": [ + { + "bbox": [ + 55, + 186, + 169, + 196 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 169, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 169, + 196 + ], + "type": "text", + "content": "500 test cases, 3 candidates each" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 198, + 200, + 248 + ], + "lines": [ + { + "bbox": [ + 50, + 198, + 200, + 248 + ], + "spans": [ + { + "bbox": [ + 50, + 198, + 200, + 248 + ], + "type": "image", + "image_path": "db313293c0402a989307f4edaa6c3e30780b1a41d711493d1f76ea959442cd28.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 51, + 248, + 174, + 274 + ], + "lines": [ + { + "bbox": [ + 51, + 248, + 174, + 274 + ], + "spans": [ + { + "bbox": [ + 51, + 248, + 174, + 274 + ], + "type": "text", + "content": "the pizza is smaller than the sandwich \nthe pizza is equal-size with the sandwich \nthe pizza is bigger than the sandwich" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 226, + 71, + 302, + 81 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 71, + 302, + 81 + ], + "spans": [ + { + "bbox": [ + 226, + 71, + 302, + 81 + ], + "type": "text", + "content": "Absolute Position" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 221, + 93, + 271, + 141 + ], + "blocks": [ + { + "bbox": [ + 226, + 82, + 341, + 91 + ], + "lines": [ + { + "bbox": [ + 226, + 82, + 341, + 91 + ], + "spans": [ + { + "bbox": [ + 226, + 82, + 341, + 91 + ], + "type": "text", + "content": "500 test cases, 9 candidates each" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 221, + 93, + 271, + 141 + ], + "lines": [ + { + "bbox": [ + 221, + 93, + 271, + 141 + ], + "spans": [ + { + "bbox": [ + 221, + 93, + 271, + 141 + ], + "type": "image", + "image_path": "13e47e4bb7989688a9307c7f2827199f1e8543b9fd20f37027c3240fd228bc92.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 271, + 94, + 320, + 142 + ], + "blocks": [ + { + "bbox": [ + 271, + 94, + 320, + 142 + ], + "lines": [ + { + "bbox": [ + 271, + 94, + 320, + 142 + ], + "spans": [ + { + "bbox": [ + 271, + 94, + 320, + 142 + ], + "type": "image", + "image_path": "c3d9f019e69011da1f4102204c48986e13edea0b2e6f3f61ebe4f430a25b3d30.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 142, + 332, + 169 + ], + "lines": [ + { + "bbox": [ + 223, + 142, + 332, + 169 + ], + "spans": [ + { + "bbox": [ + 223, + 142, + 332, + 169 + ], + "type": "text", + "content": "the kite is in the top-left corner \nthe kite is in the center of the image \nthe kite is in the bottom-left corner" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 321, + 94, + 370, + 142 + ], + "blocks": [ + { + "bbox": [ + 321, + 94, + 370, + 142 + ], + "lines": [ + { + "bbox": [ + 321, + 94, + 370, + 142 + ], + "spans": [ + { + "bbox": [ + 321, + 94, + 370, + 142 + ], + "type": "image", + "image_path": "56f20d2194fb1fde36d7314976f430b56dc15e7c459540efda1f20b16698c026.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 227, + 177, + 299, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 177, + 299, + 186 + ], + "spans": [ + { + "bbox": [ + 227, + 177, + 299, + 186 + ], + "type": "text", + "content": "Relative Position" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 221, + 198, + 271, + 248 + ], + "blocks": [ + { + "bbox": [ + 227, + 188, + 341, + 197 + ], + "lines": [ + { + "bbox": [ + 227, + 188, + 341, + 197 + ], + "spans": [ + { + "bbox": [ + 227, + 188, + 341, + 197 + ], + "type": "text", + "content": "500 test cases, 4 candidates each" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 221, + 198, + 271, + 248 + ], + "lines": [ + { + "bbox": [ + 221, + 198, + 271, + 248 + ], + "spans": [ + { + "bbox": [ + 221, + 198, + 271, + 248 + ], + "type": "image", + "image_path": "c21fcae67e0124aa3a071d61b520e21550e1ffe7c385e9170a2f9633e7906b2f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 249, + 337, + 274 + ], + "lines": [ + { + "bbox": [ + 223, + 249, + 337, + 274 + ], + "spans": [ + { + "bbox": [ + 223, + 249, + 337, + 274 + ], + "type": "text", + "content": "the toilet is to the left of the backpack the toilet is on top of the backpack the toilet is below the backpack" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 272, + 199, + 322, + 248 + ], + "blocks": [ + { + "bbox": [ + 272, + 199, + 322, + 248 + ], + "lines": [ + { + "bbox": [ + 272, + 199, + 322, + 248 + ], + "spans": [ + { + "bbox": [ + 272, + 199, + 322, + 248 + ], + "type": "image", + "image_path": "274c2c6cc8881b0357cadbaca0405e14ca0c8627ba32bed87dc9ac4408d7dffe.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 323, + 199, + 372, + 248 + ], + "blocks": [ + { + "bbox": [ + 323, + 199, + 372, + 248 + ], + "lines": [ + { + "bbox": [ + 323, + 199, + 372, + 248 + ], + "spans": [ + { + "bbox": [ + 323, + 199, + 372, + 248 + ], + "type": "image", + "image_path": "8036839ee38258c4555b0bf141d748f6937890ba953253a1530c0e996a1a4639.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 399, + 71, + 427, + 81 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 71, + 427, + 81 + ], + "spans": [ + { + "bbox": [ + 399, + 71, + 427, + 81 + ], + "type": "text", + "content": "Count" + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 394, + 92, + 443, + 142 + ], + "blocks": [ + { + "bbox": [ + 399, + 82, + 513, + 91 + ], + "lines": [ + { + "bbox": [ + 399, + 82, + 513, + 91 + ], + "spans": [ + { + "bbox": [ + 399, + 82, + 513, + 91 + ], + "type": "text", + "content": "500 test cases, 9 candidates each" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 394, + 92, + 443, + 142 + ], + "lines": [ + { + "bbox": [ + 394, + 92, + 443, + 142 + ], + "spans": [ + { + "bbox": [ + 394, + 92, + 443, + 142 + ], + "type": "image", + "image_path": "cb3b518742e3132392c0f1f50a8edc610c0d64f9f051b5ccf8828be52e02c186.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 143, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 395, + 143, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 395, + 143, + 482, + 169 + ], + "type": "text", + "content": "a photo of one snowboard a photo of two snowboards a photo of three snowboards" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 444, + 93, + 493, + 142 + ], + "blocks": [ + { + "bbox": [ + 444, + 93, + 493, + 142 + ], + "lines": [ + { + "bbox": [ + 444, + 93, + 493, + 142 + ], + "spans": [ + { + "bbox": [ + 444, + 93, + 493, + 142 + ], + "type": "image", + "image_path": "d8dc1b298fccc219b7c60a7330ed00f5d2bc8720ff5aed3997cc369c765fc6ff.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 495, + 93, + 545, + 142 + ], + "blocks": [ + { + "bbox": [ + 495, + 93, + 545, + 142 + ], + "lines": [ + { + "bbox": [ + 495, + 93, + 545, + 142 + ], + "spans": [ + { + "bbox": [ + 495, + 93, + 545, + 142 + ], + "type": "image", + "image_path": "b685079b70d99267165d64f1c466b33ccfebc6ff8953471ce928dd0bb8fef1fb.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 395, + 177, + 438, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 177, + 438, + 186 + ], + "spans": [ + { + "bbox": [ + 395, + 177, + 438, + 186 + ], + "type": "text", + "content": "Existence" + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 391, + 200, + 441, + 249 + ], + "blocks": [ + { + "bbox": [ + 395, + 188, + 511, + 198 + ], + "lines": [ + { + "bbox": [ + 395, + 188, + 511, + 198 + ], + "spans": [ + { + "bbox": [ + 395, + 188, + 511, + 198 + ], + "type": "text", + "content": "500 test cases, 2 candidates each" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 200, + 441, + 249 + ], + "lines": [ + { + "bbox": [ + 391, + 200, + 441, + 249 + ], + "spans": [ + { + "bbox": [ + 391, + 200, + 441, + 249 + ], + "type": "image", + "image_path": "fc0c3b73e183f4c03a7f8d157c1073d5e08b73822d12e4c437da95d27788534e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 249, + 485, + 258 + ], + "lines": [ + { + "bbox": [ + 392, + 249, + 485, + 258 + ], + "spans": [ + { + "bbox": [ + 392, + 249, + 485, + 258 + ], + "type": "text", + "content": "there is no giraffe in the image" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 392, + 266, + 481, + 275 + ], + "lines": [ + { + "bbox": [ + 392, + 266, + 481, + 275 + ], + "spans": [ + { + "bbox": [ + 392, + 266, + 481, + 275 + ], + "type": "text", + "content": "there is a giraffe in the image" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 493, + 200, + 542, + 250 + ], + "blocks": [ + { + "bbox": [ + 493, + 200, + 542, + 250 + ], + "lines": [ + { + "bbox": [ + 493, + 200, + 542, + 250 + ], + "spans": [ + { + "bbox": [ + 493, + 200, + 542, + 250 + ], + "type": "image", + "image_path": "d2a8c259b75cedb5576255e8975b8ee82f1201bbcbca5679a9d71900fe4b5d87.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 283, + 547, + 328 + ], + "lines": [ + { + "bbox": [ + 46, + 283, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 547, + 328 + ], + "type": "text", + "content": "Figure 5. An overview of the SPEC benchmark. SPEC consists of six distinct subsets, distributed across the dimensions of Size, Position, Existence and Count. Each test case consists of an image candidate set, which differs only in certain visual concept, and a text candidate set, which differs only in corresponding language concept. Due to space constraints, we present a maximum of three images and texts here, however, more comprehensive test cases are available in the supplementary material." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 349, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 287, + 432 + ], + "type": "text", + "content": "and the semantic space covered by the candidate set is more comprehensive. For example, in the case of relative position, the candidates include all four semantic directions (up, down, left, right), and in absolute position and count, it extends to nine candidates. As will be discussed below, more confusing candidates can be readily used as hard negatives to improve current VLMs." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 47, + 443, + 243, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 243, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 243, + 456 + ], + "type": "text", + "content": "4. Diagnose: Probing VLMs on SPEC" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 46, + 463, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 523 + ], + "type": "text", + "content": "We conduct a systematical evaluation of four state-of-the-art VLMs: CLIP [23], BLIP [17], FLAVA [29] and CoCa [39] using our newly proposed SPEC benchmark, aiming to uncover that to what extent and in what aspect are they excelling or suffering." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 47, + 531, + 208, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 208, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 208, + 542 + ], + "type": "text", + "content": "4.1. Evaluation Task and Protocol" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": "Symmetric image-text matching task. We conduct evaluations using the image-text matching task, where each test case " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{T} = (\\mathcal{I},\\mathcal{T})" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": " comprises an image set " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": " and a corresponding text set " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": ", as outlined in Eq. (3). The evaluation process is symmetric, considering both visual and textual perspectives. In the image-to-text matching task, when querying with an image " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": ", the model is required to accurately identify " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": " from the candidate set " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": ". Similarly, in the text-to-image matching task, the goal is to find " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": " for a given query text " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": ". In practice, we accomplish the matching process using the similarity " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "s(I,T)" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": " between a given image " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": " and text " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": ". A candidate will be selected if its similarity with the query ranks first among the entire candidate set." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 305, + 349, + 546, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 349, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 349, + 546, + 396 + ], + "type": "text", + "content": "Evaluation protocols. We measure the performance on SPEC using two metrics: I2TACC and T2IACC, representing the accuracy of image-to-text and text-to-image matching task, respectively:" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 335, + 413, + 545, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 413, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 335, + 413, + 545, + 444 + ], + "type": "interline_equation", + "content": "I 2 \\mathrm {T} _ {\\text {A C C}} = \\frac {1}{| D |} \\sum_ {\\left(\\mathcal {I} _ {i}, \\mathcal {T} _ {i}\\right) \\in D} \\frac {1}{\\left| \\mathcal {I} _ {i} \\right|} \\sum_ {I _ {j} \\in \\mathcal {I} _ {i}} h \\left(I _ {j}, \\mathcal {T} _ {i}\\right) \\tag {4}", + "image_path": "26c03d8d6a08ebe7c5a952b5f0b77d887c3c12b8a632d95ec099761ba04247bd.jpg" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 332, + 461, + 545, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 461, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 332, + 461, + 545, + 492 + ], + "type": "interline_equation", + "content": "\\mathrm {T 2 I A C C} = \\frac {1}{| D |} \\sum_ {\\left(\\mathcal {I} _ {i}, \\mathcal {T} _ {i}\\right) \\in D} \\frac {1}{\\left| \\mathcal {T} _ {i} \\right|} \\sum_ {T _ {j} \\in \\mathcal {T} _ {i}} g \\left(T _ {j}, \\mathcal {I} _ {i}\\right), \\tag {5}", + "image_path": "c3112428ef94442b22b8b6db2431c98141fe09bbff0e1fd4f86fad8c3c7fb258.jpg" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "spans": [ + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "inline_equation", + "content": "|D|" + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "text", + "content": " test cases, " + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "inline_equation", + "content": "h(I_j,\\mathcal{T}_i)" + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "text", + "content": " equals to 1 if and only if " + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "text", + "content": " correctly find its matched text " + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "text", + "content": " from the candidate set " + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "text", + "content": ", otherwise it is set to 0. Similarly, " + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "inline_equation", + "content": "g(T_j,\\mathcal{I}_i)" + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "text", + "content": " equals 1 if and only if " + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "text", + "content": " correctly finds its matched image " + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "text", + "content": " from the candidate set " + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_i" + }, + { + "bbox": [ + 305, + 499, + 547, + 560 + ], + "type": "text", + "content": ", otherwise, it is set to 0." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 306, + 567, + 480, + 579 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 480, + 579 + ], + "type": "text", + "content": "4.2. Key Insights from SPEC Results" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 305, + 585, + 545, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 585, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 305, + 585, + 545, + 633 + ], + "type": "text", + "content": "We evaluate four VLMs using the SPEC benchmark, and their results are summarized in Tab. 2. We find that all the models exhibit a limited accuracy close to random chance, from which we gain the following insights:" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "content": "Even state-of-the-art VLMs perform at chance level. From the results, we surprisingly find that even the most advanced VLMs achieve only a marginal advantage compared to random chance, which sharply contrasts with their impressive performance on common tasks. For instance, CLIP [23] demonstrates a mere " + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "inline_equation", + "content": "33.4\\%" + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "content": " T2IACC for relative" + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13284" + } + ] + } + ], + "index": 45 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 182 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 182 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 182 + ], + "type": "table", + "html": "
Absolute SizeRelative SizeAbsolute PositionRelative PositionExistenceCount
I2TT2ICLSI2TT2ICLSI2TT2ICLSI2TT2ICLSI2TT2ICLSI2TT2ICLS
Random33.333.31.333.333.32.511.111.11.325.025.02.550.050.05.011.111.11.3
FLAVA [29]37.337.289.332.932.384.113.115.788.825.526.784.057.951.974.414.421.275.8
BLIP [17]43.342.797.733.232.598.212.211.097.930.529.797.655.450.196.037.437.193.8
CoCa [39]39.136.195.933.633.395.411.711.896.130.328.892.250.950.083.736.535.689.6
CLIP [23]42.536.192.134.433.494.912.612.392.828.026.690.158.351.284.025.123.283.9
Ours68.960.796.340.344.197.330.634.296.946.646.996.283.453.192.555.657.892.5
", + "image_path": "06f022482399cac0ffb05fc251a64a6481b42748aad0f5d08735f48c843e0e8e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "text", + "content": "Table 2. Evaluation results on SPEC. We extensively benchmark four state-of-the-art VLMs on SPEC to investigate their comprehension of fine-grained visual-linguistic concepts. Our evaluation employs two metrics, I2TACC and T2IACC, and we report the detailed performance on the each subset. We also report the classification accuracy, CLS, to highlight the capability in recognizing object categories. The first row indicates accuracy at chance level, serving as a baseline for comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 259, + 288, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 288, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 288, + 355 + ], + "type": "text", + "content": "size recognition, while the chance-level accuracy is " + }, + { + "bbox": [ + 46, + 259, + 288, + 355 + ], + "type": "inline_equation", + "content": "33.3\\%" + }, + { + "bbox": [ + 46, + 259, + 288, + 355 + ], + "type": "text", + "content": ". While BLIP performs reasonably well in absolute size, surpassing random level by around " + }, + { + "bbox": [ + 46, + 259, + 288, + 355 + ], + "type": "inline_equation", + "content": "9.7\\%" + }, + { + "bbox": [ + 46, + 259, + 288, + 355 + ], + "type": "text", + "content": ", the i2TACC on relative size is " + }, + { + "bbox": [ + 46, + 259, + 288, + 355 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 46, + 259, + 288, + 355 + ], + "type": "text", + "content": " lagged behind. CoCa [39] and FLAVA [29] also exhibit significant weaknesses in performance. The last row presents the performance of our improved model, demonstrating significant advancements across all metrics (as will be introduced in Sec. 5)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 364, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 287, + 555 + ], + "type": "text", + "content": "The challenge arises from the task itself, not the data. One might attribute the poor performance of VLMs to the data quality or distribution. To address this concern, we conduct an additional experiment. Specifically, we perform classification experiments using the SPEC dataset to assess the model's understanding of nouns or object categories. In this context, the models exhibit impressive performance, achieving approximately " + }, + { + "bbox": [ + 46, + 364, + 287, + 555 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 364, + 287, + 555 + ], + "type": "text", + "content": " Top-1 accuracy in the 80-class classification task (denoted as CLS in Tab. 2). This aligns well with earlier findings [40] that VLMs struggle in compositional reasoning while excelling in object category recognition. The remarkable accuracy of the models in the object classification task confirms the high quality of SPEC data. This also validates that the challenges faced by VLMs stem from the tasks which require fine-grained recognition rather than issues with the data itself." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 563, + 224, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 224, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 224, + 575 + ], + "type": "text", + "content": "4.3. Discussion on Model Limitations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "We attribute the poor performance of vision and language models on SPEC to their pretraining methods, specifically, the inherent limitation in standard contrastive loss. The conventional contrastive learning involves randomly sampling batches of images and texts, requiring the model to identify matching pairs within the batch. This task is intended to facilitate alignment between text and image spaces. However, as highlighted in prior studies [2, 40], the substantial differences between items in a randomly sampled batch allows the model to effortlessly complete this task. It can easily achieve this by focusing solely on nouns" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 259, + 547, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 547, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 547, + 332 + ], + "type": "text", + "content": "in the text and object categories in the images through a shortcut [8]. This leads the model biased towards noun concepts, neglecting other finer-grained concepts. This is the reason why these models demonstrate poor performance on fine-grained tasks that demanding understanding concepts beyond nouns." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 340, + 536, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 340, + 536, + 355 + ], + "spans": [ + { + "bbox": [ + 305, + 340, + 536, + 355 + ], + "type": "text", + "content": "5. Optimize: A Simple but Effective Remedy" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 361, + 547, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 547, + 385 + ], + "type": "text", + "content": "We experiment with CLIP [23] and propose a remedy to enhance its performance in fine-grained understanding." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 390, + 365, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 390, + 365, + 402 + ], + "spans": [ + { + "bbox": [ + 306, + 390, + 365, + 402 + ], + "type": "text", + "content": "5.1. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 409, + 547, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 547, + 456 + ], + "type": "text", + "content": "CLIP [23] consists of an visual encoder to extract image embedding: " + }, + { + "bbox": [ + 304, + 409, + 547, + 456 + ], + "type": "inline_equation", + "content": "e_{I} = \\mathcal{E}_{I}(I)" + }, + { + "bbox": [ + 304, + 409, + 547, + 456 + ], + "type": "text", + "content": " and a textual encoder to extract text embedding: " + }, + { + "bbox": [ + 304, + 409, + 547, + 456 + ], + "type": "inline_equation", + "content": "e_{T} = \\mathcal{E}_{T}(T)" + }, + { + "bbox": [ + 304, + 409, + 547, + 456 + ], + "type": "text", + "content": ". The similarity score between an image " + }, + { + "bbox": [ + 304, + 409, + 547, + 456 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 409, + 547, + 456 + ], + "type": "text", + "content": " and a text " + }, + { + "bbox": [ + 304, + 409, + 547, + 456 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 409, + 547, + 456 + ], + "type": "text", + "content": " is computed following:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 359, + 461, + 545, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 461, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 359, + 461, + 545, + 490 + ], + "type": "interline_equation", + "content": "s (I, T) = \\exp \\left(\\frac {\\tau e _ {I} ^ {T} e _ {T}}{\\| e _ {I} \\| ^ {2} \\| e _ {T} \\| ^ {2}}\\right), \\tag {6}", + "image_path": "e51bafe00e78e4312914c5e51faf5390a58ad4829b2b4b987d52c6ca2dca3193.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 493, + 450, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 493, + 450, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 493, + 450, + 504 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 493, + 450, + 504 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 306, + 493, + 450, + 504 + ], + "type": "text", + "content": " is a learnable temperature." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 505, + 547, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 547, + 601 + ], + "type": "text", + "content": "In order to guide CLIP to focus on fine-grained visual-linguistic concepts, we incorporate confusing images and text as hard negatives within the same batch. This requires CLIP to pull positives closer and push hard negatives away, thereby enhancing its ability to discern nuanced visual and textual differences. Specifically, we introduce an hard negative aware contrastive loss " + }, + { + "bbox": [ + 304, + 505, + 547, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{hn} = \\mathcal{L}_{hn}^{12\\mathrm{T}} + \\mathcal{L}_{hn}^{\\mathrm{T}2\\mathrm{I}}" + }, + { + "bbox": [ + 304, + 505, + 547, + 601 + ], + "type": "text", + "content": ", which comprises an image-to-text and a text-to-image term:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 604, + 545, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 604, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 315, + 604, + 545, + 651 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {h n} ^ {\\mathrm {l 2 T}} = - \\sum_ {i} \\log \\frac {s \\left(I _ {i} , T _ {i}\\right)}{\\sum_ {T _ {j} \\in \\mathcal {T}} s \\left(I _ {i} , T _ {j}\\right) + \\sum_ {T _ {k} ^ {h n} \\in \\mathcal {T} ^ {h n}} s \\left(I _ {i} , T _ {k} ^ {h n}\\right)} \\tag {7}", + "image_path": "d0ef41108f9c7296867f0e61b814e3a75d34129b130318dbe2f2ccd3df1ffff1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 665, + 545, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 665, + 545, + 713 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {h n} ^ {\\mathrm {T 2 I}} = - \\sum_ {i} \\log \\frac {s \\left(I _ {i} , T _ {i}\\right)}{\\sum_ {I _ {j} \\in \\mathcal {I}} s \\left(I _ {j} , T _ {i}\\right) + \\sum_ {I _ {k} ^ {h n} \\in \\mathcal {I} ^ {h n}} s \\left(I _ {k} ^ {h n} , T _ {i}\\right)}, \\tag {8}", + "image_path": "0722f4784d24e0de7c418dfca378cffb7cf09f73283d7e92ba373d4571f18264.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13285" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " represent the trivial images and texts, respectively, while " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{hn}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{hn}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " denote non-trivial hard negatives. In our implementation, these hard negative examples are constructed using the data pipeline described in Sec. 3.2, and more details are in the supplementary." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 193 + ], + "type": "text", + "content": "To preserve the inherent zero-shot capability of CLIP, we also leverage the conventional image-text pairs from LAION-400M [27]. We apply the standard contrastive loss of CLIP [23] to these data, introducing an additional loss term " + }, + { + "bbox": [ + 46, + 133, + 287, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{clip}" + }, + { + "bbox": [ + 46, + 133, + 287, + 193 + ], + "type": "text", + "content": ". The overall loss consists of two terms:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 200, + 287, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 200, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 126, + 200, + 287, + 213 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {c l i p}} + \\lambda \\mathcal {L} _ {\\text {h n}}, \\tag {9}", + "image_path": "7636acfeca38c9d080853bf51bad97b058e5d4dca35c363a98353672224b7c0f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 220, + 287, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 269 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 220, + 287, + 269 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 220, + 287, + 269 + ], + "type": "text", + "content": " is a hyperparameter that balancing these two terms. Training on this multi-task loss enables improving the performance of CLIP in fine-grained understanding while maintaining its zero-shot capability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 275, + 129, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 129, + 288 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 129, + 288 + ], + "type": "text", + "content": "5.2. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 301, + 287, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 385 + ], + "type": "text", + "content": "Training details. We experiment with the ViT-B/32 variant of CLIP, and resume from the OpenAI pretrained checkpoint [23]. We finetune for 1,000 steps using a cosine schedule with an initial learning rate of " + }, + { + "bbox": [ + 46, + 301, + 287, + 385 + ], + "type": "inline_equation", + "content": "1e - 6" + }, + { + "bbox": [ + 46, + 301, + 287, + 385 + ], + "type": "text", + "content": " and use 800 steps for warm up. The batch size of LAION data is set to 2048, and the batch size of hard negative data is set to 768. The weight " + }, + { + "bbox": [ + 46, + 301, + 287, + 385 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 301, + 287, + 385 + ], + "type": "text", + "content": " of the hard negative aware loss is set to 0.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 393, + 287, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 287, + 622 + ], + "type": "text", + "content": "Main results. We utilize the SPEC benchmark to assess the understanding of model in fine-grained concepts. In Tab. 3, we present the average I2TACC and T2IACC on all subsets of SPEC. To assess the general performance of the model, we also utilize the toolkit from ELEVATER [16] to evaluate the zero-shot performance on 9 classification and retrieval datasets, and report the average accuracy. Compared to CLIP [23], our model demonstrates remarkable advancements with a " + }, + { + "bbox": [ + 46, + 393, + 287, + 622 + ], + "type": "inline_equation", + "content": "19.8\\%" + }, + { + "bbox": [ + 46, + 393, + 287, + 622 + ], + "type": "text", + "content": " boost in I2TACC, an " + }, + { + "bbox": [ + 46, + 393, + 287, + 622 + ], + "type": "inline_equation", + "content": "18.9\\%" + }, + { + "bbox": [ + 46, + 393, + 287, + 622 + ], + "type": "text", + "content": " improvement in T2IACC on SPEC, and a noteworthy " + }, + { + "bbox": [ + 46, + 393, + 287, + 622 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 46, + 393, + 287, + 622 + ], + "type": "text", + "content": " enhancement in zero-shot accuracy. We also conduct ablation on different training configurations. From the results in Tab. 3, it can be observed that the introducing of " + }, + { + "bbox": [ + 46, + 393, + 287, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{hn}" + }, + { + "bbox": [ + 46, + 393, + 287, + 622 + ], + "type": "text", + "content": " significantly improves the performance on SPEC. Moreover, the " + }, + { + "bbox": [ + 46, + 393, + 287, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{clip}" + }, + { + "bbox": [ + 46, + 393, + 287, + 622 + ], + "type": "text", + "content": " plays a crucial role in preserving zero-shot performance. Without it, we observe a decline in accuracy by " + }, + { + "bbox": [ + 46, + 393, + 287, + 622 + ], + "type": "inline_equation", + "content": "5.1\\%" + }, + { + "bbox": [ + 46, + 393, + 287, + 622 + ], + "type": "text", + "content": ". With the combination of these two losses, we achieve substantial improvement in SPEC while maintaining the original zero-shot capability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "Validation on other fine-grained benchmarks. To assess whether our approach aids the model in acquiring fundamental visual-linguistic understanding or merely leads to overfitting on SPEC, we conduct evaluations on two additional benchmarks which also focus on the assessment of fine-grained concepts. ARO [40] explores three aspects of vision-language understanding: object attributes," + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 541, + 159 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 541, + 159 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 541, + 159 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 541, + 159 + ], + "type": "table", + "html": "
ConfigSPECZero-shot Accuracy
LclipLhnI2TT2I
CLIP33.530.567.5
+ Lhn✓64.360.862.4
+ Lclip✓32.231.569.4
+ Lhn+Lclip (ours)✓✓53.349.468.7
", + "image_path": "feb3bac01ed06e163f0e601f91fe1860f60a998f86c764a18d4837e928fd4e10.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 211, + 545, + 287 + ], + "blocks": [ + { + "bbox": [ + 305, + 167, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 305, + 167, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 167, + 545, + 201 + ], + "type": "text", + "content": "Table 3. Main Results: The first row represents the pretrained checkpoint without fine-tuning. We sequentially introduce two additional loss terms to investigate their impact for the performance." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 211, + 545, + 287 + ], + "lines": [ + { + "bbox": [ + 309, + 211, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 309, + 211, + 545, + 287 + ], + "type": "table", + "html": "
EqbenARO
ImageTextGroupAttributeRelationOrder
CLIP17.621.410.163.263.953.3
CLIPFT18.123.711.165.168.054.1
Ours19.524.011.766.473.760.7
", + "image_path": "e7fd3ebc1eb3b558b7d3405d2141784aa5355205f3f044fec1885f3f51ff5726.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 295, + 545, + 340 + ], + "lines": [ + { + "bbox": [ + 305, + 295, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 305, + 295, + 545, + 340 + ], + "type": "text", + "content": "Table 4. Cross-dataset evaluation results on Eqben [33] and ARO [40]. To demonstrate that the improvement comes from the negative loss, rather than training on more data, we also report " + }, + { + "bbox": [ + 305, + 295, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{FT}" + }, + { + "bbox": [ + 305, + 295, + 545, + 340 + ], + "type": "text", + "content": ", which is also finetuned but without negative samples." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "text", + "content": "inter-object relations, and word ordering. Eqben [33] focuses on minimal visual semantic changes, aiming to diagnose VLMs in understanding fine-grained concepts such as counting and location. In Tab. 4, we present the experimental results, demonstrating a clear improvement compared to CLIP [23] on both datasets. For example, compared to CLIP, our method shows an average improvement of " + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "text", + "content": " on Eqben and respective enhancements of " + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "inline_equation", + "content": "9.8\\%" + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "text", + "content": " on the three subsets of ARO. The consistent improvement on these datasets demonstrates that our approach has facilitated the model in acquiring transferable fine-grained understanding capabilities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 514, + 378, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 378, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 378, + 526 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 533, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 689 + ], + "type": "text", + "content": "In this study, we explored the comprehension abilities of Visual Language Models (VLMs) with respect to fine-grained visual-linguistic concepts. We first established an efficient pipeline to synthesize candidate images that exclusively differ in a particular visual attribute. Leveraging this pipeline, we created the SPEC benchmark to diagnose the comprehension proficiency of VLMs in terms of object size, position, existence, and count. Upon evaluating four leading VLMs using SPEC, we uncovered substantial performance limitations. To address this, we introduced an enhancement strategy that effectively optimizes the model for fine-grained understanding, while maintaining its original zero-shot capability." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement This project was supported by NSFC under Grant No. 62102092." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13286" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In ICCV, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Paola Cascante-Bonilla, Khaled Shehada, James Smith, Sivan Doveh, Donghyun Kim, Rameswar Panda, Gül Varol, Aude Oliva, Vicente Ordonez, Rogério Schmidt Feris, and Leonid Karlinsky. Going beyond nouns with vision & language models using synthetic data. In ICCV, 2023. 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 289, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 289, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 289, + 303 + ], + "type": "text", + "content": "[3] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel M. Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model. In ICLR, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 304, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 288, + 348 + ], + "type": "text", + "content": "[4] Yuxin Chen, Zongyang Ma, Ziqi Zhang, Zhongang Qi, Chunfen Yuan, Ying Shan, Bing Li, Weiming Hu, Xiaohu Qie, and Jianping Wu. Vilem: Visual-language error modeling for image-text retrieval. In CVPR, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 350, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 288, + 381 + ], + "type": "text", + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 384, + 288, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 384, + 288, + 416 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 288, + 416 + ], + "type": "text", + "content": "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 418, + 288, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 418, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 53, + 418, + 288, + 472 + ], + "type": "text", + "content": "[7] Sivan Doveh, Assaf Arbelle, Sivan Harary, Rameswar Panda, Roei Herzig, Eli Schwartz, Donghyun Kim, Raja Giryes, Rogério Schmidt Feris, Shimon Ullman, and Leonid Karlinsky. Teaching structured vision & language concepts to vision & language models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 474, + 288, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 474, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 53, + 474, + 288, + 518 + ], + "type": "text", + "content": "[8] Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard S. Zemel, Wieland Brendel, Matthias Bethge, and Felix Wichmann. Shortcut learning in deep neural networks. Nat. Mach. Intell., 2020. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 520, + 288, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 520, + 288, + 563 + ], + "spans": [ + { + "bbox": [ + 53, + 520, + 288, + 563 + ], + "type": "text", + "content": "[9] Golnaz Ghiasi, Yin Cui, A. Srinivas, Rui Qian, Tsung-Yi Lin, Ekin Dogus Cubuk, Quoc V. Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. CVPR, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 565, + 288, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 565, + 288, + 598 + ], + "spans": [ + { + "bbox": [ + 48, + 565, + 288, + 598 + ], + "type": "text", + "content": "[10] Shashank Goel, Hritik Bansal, Sumit Kaur Bhatia, Ryan A. Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 600, + 288, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 288, + 621 + ], + "type": "text", + "content": "[11] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 623, + 288, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 288, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 288, + 655 + ], + "type": "text", + "content": "[12] Lisa Anne Hendricks and Aida Nematzadeh. Probing image-language transformers for verb understanding. In Findings, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 657, + 288, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 288, + 700 + ], + "type": "text", + "content": "[13] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "type": "text", + "content": "[14] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan Liu, Yu-Wing" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. ArXiv, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 97, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 547, + 150 + ], + "type": "text", + "content": "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In ICCV, 2023. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "type": "text", + "content": "[16] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Yong Jae Lee, Houdong Hu, Zicheng Liu, and Jianfeng Gao. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. In NeurIPS, 2022. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "type": "text", + "content": "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 1, 2, 6, 7, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "type": "text", + "content": "[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 4, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 298, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 319 + ], + "type": "text", + "content": "[19] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. TACL, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 322, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 322, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 322, + 545, + 365 + ], + "type": "text", + "content": "[20] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. ArXiv, 2023. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 367, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 367, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 367, + 545, + 410 + ], + "type": "text", + "content": "[21] Letitia Parcalabescu, Michele Cafagna, Lilitta Muradjan, Anette Frank, Iacer Calixto, and Albert Gatt. Valse: A task-independent benchmark for vision and language models centered on linguistic phenomena. In ACL, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 412, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 412, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 308, + 412, + 545, + 455 + ], + "type": "text", + "content": "[22] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. ArXiv, 2023. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 456, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 511 + ], + "type": "text", + "content": "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 6, 7, 8, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 514, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 545 + ], + "type": "text", + "content": "[24] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In 2021. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 547, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 547, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 547, + 545, + 602 + ], + "type": "text", + "content": "[25] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "type": "text", + "content": "[26] Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Galle, et al. Bloom: A 176b-parameter open-access multilingual language model. ArXiv, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[27] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. ArXiv, 2021. 8" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13287" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 149 + ], + "type": "text", + "content": "[28] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "text", + "content": "[29] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022. 1, 2, 6, 7, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 194, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 287, + 248 + ], + "type": "text", + "content": "[30] James Smith, Paola Cascante-Bonilla, Assaf Arbelle, Donghyun Kim, Rameswar Panda, David Cox, Diyi Yang, Zsolt Kira, Rogério Schmidt Feris, and Leonid Karlinsky. Construct-vl: Data-free continual structured vl concepts learning*. In CVPR, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "type": "text", + "content": "[31] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In CVPR, 2022. 1, 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 293, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 346 + ], + "type": "text", + "content": "[32] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnivl: One foundation model for image-language and video-language tasks. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "type": "text", + "content": "[33] Tan Wang, Kevin Lin, Linjie Li, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Equivariant similarity for vision-language foundation models. In ICCV, 2023. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 393, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 435 + ], + "type": "text", + "content": "[34] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization. In ICML, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "text", + "content": "[35] Zuxuan Wu, Zejia Weng, Wujian Peng, Xitong Yang, Ang Li, Larry S Davis, and Yu-Gang Jiang. Building an open-vocabulary video clip model with better architectures, optimization and data. TPAMI, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "type": "text", + "content": "[36] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. arXiv preprint arXiv:2310.10647, 2023. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 548 + ], + "type": "text", + "content": "[37] Zhen Xing, Qi Dai, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Simda: Simple diffusion adapter for efficient video generation. In CVPR, 2024. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "text", + "content": "[38] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "text", + "content": "[39] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Trans. Mach. Learn. Res., 2022. 2, 6, 7, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "text", + "content": "[40] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bag-of-words models, and what to do about it? ICLR, 2023. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[41] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. In ICML, 2021. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 182 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 128 + ], + "type": "text", + "content": "[42] Hanqing Zhao, Dianmo Sheng, Jianmin Bao, Dongdong Chen, Dong Chen, Fang Wen, Lu Yuan, Ce Liu, Wenbo Zhou, Qi Chu, Weiming Zhang, and Neng H. Yu. X-paste: Revisiting scalable copy-paste for instance segmentation using clip and stablediffusion. In ICML, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "type": "text", + "content": "[43] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. ArXiv, 2022. 1, 2, 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13288" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/58e8331e-f572-44c9-b724-02ca5cd604a2_content_list.json b/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/58e8331e-f572-44c9-b724-02ca5cd604a2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4b9c760f8c9ac029cee73f77a148ff6d556d0868 --- /dev/null +++ b/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/58e8331e-f572-44c9-b724-02ca5cd604a2_content_list.json @@ -0,0 +1,1494 @@ +[ + { + "type": "text", + "text": "Synthesize Step-by-Step: Tools, Templates and LLMs as Data Generators for Reasoning-Based Chart VQA", + "text_level": 1, + "bbox": [ + 181, + 130, + 787, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhuowan Li $^{1}$ * Bhavan Jasani $^{2}$ * Peng Tang $^{2}$ Shabnam Ghadar $^{2}$", + "bbox": [ + 210, + 202, + 754, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Johns Hopkins University 2 AWS AI Labs", + "bbox": [ + 302, + 220, + 668, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zili110@jhu.edu {bjasani, tangpen, shabnam}@amazon.com", + "bbox": [ + 250, + 241, + 715, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding data visualizations like charts and plots requires reasoning about both visual elements and numerics. Although strong in extractive questions, current chart visual question answering (chart VQA) models suffer on complex reasoning questions. In this work, we address the lack of reasoning ability by data augmentation. We leverage Large Language Models (LLMs), which have shown to have strong reasoning ability, as an automatic data annotator that generates question-answer annotations for chart images. The key innovation in our method lies in the Synthesize Step-by-Step strategy: our LLM-based data generator learns to decompose the complex question into step-by-step sub-questions (rationales), which are then used to derive the final answer using external tools, i.e. Python. This step-wise generation procedure is trained on synthetic data generated using a template-based QA generation pipeline. Experimental results highlight the significance of the proposed step-by-step generation. By training with the LLM-augmented data (LAMENDA), we significantly enhance the chart VQA models, achieving the state-of-the-art accuracy on the ChartQA and PlotQA datasets. In particular, our approach improves the accuracy of the previous state-of-the-art approach from $38\\%$ to $54\\%$ on the human-written questions in the ChartQA dataset, which needs strong reasoning. We hope our work underscores the potential of synthetic data and encourages further exploration of data augmentation using LLMs for reasoning-heavy tasks.", + "bbox": [ + 76, + 323, + 473, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 762, + 209, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Data visualizations like charts and plots play an important role in real-world data analysis applications. Unlike natural images, chart are text-heavy and data-driven, thus requiring better visual perception (e.g. OCR) and cognitive reasoning (e.g. math calculation, matching the legends with bars). For", + "bbox": [ + 75, + 787, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d6864ec69a86d47cd65f2a00a46dc0c667af30eb534d8990db4d35a3bc1694eb.jpg", + "image_caption": [ + "Figure 1. Existing chart VQA models struggle with complex reasoning questions. We attribute this to limited reasoning questions in existing datasets and address it by data augmentation. We fine-tune an LLM-based data generator that automatically generates question-answer annotations given a chart image. Our key innovation is Synthesize Step-by-Step, which breaks complex questions down into easy steps that could be solved using external tools. We use templates to train the LLM. Training with LLM-augmented data, LAMENDA, greatly enhances the chart VQA models." + ], + "image_footnote": [], + "bbox": [ + 503, + 279, + 890, + 459 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "example, to answer \"what is the total of Democrats and Republicans in 2010\" in Fig. 1, the model needs to recognize the texts, match the legends of \"democrats\" and \"republicans\" to the bars, then calculate the sum. This is a challenging task requiring multimodal perception and reasoning.", + "bbox": [ + 496, + 612, + 890, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compounding the complexity, the inherent difficulty of chart reasoning also amplifies the cost of collecting human annotations. Most existing datasets are synthetically generated using simple rules and heuristics, posing a risk of model overfitting. Among the datasets, only ChartQA [40] contains complex human-annotated question-answer pairs, although there are very few of those per image. For example, for the image in Fig. 1, while a lot more questions can be asked, only two questions are annotated (What is the total of Republicans and Democrats in 2010? Which color represents Republican?). Importantly, these human-written questions pose great challenges for contemporary models. For example, the state-of-the-art model [35] achieves only $38\\%$ accuracy when confronted with human-written ques", + "bbox": [ + 496, + 689, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*indicates equal contributions. The work was done when Zhuowan Li was an intern at Amazon.", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13613", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tions that require multi-step complex reasoning. Approximately half of the errors, as reported in [35], are due to reasoning failures.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose to leverage large language models (LLMs) as automatic data annotators for chart reasoning. Recently, LLMs have demonstrated strong reasoning ability for complex tasks [27, 58, 59, 63, 68, 69]. Our motivation is to take advantage of this strong reasoning ability and generate question-answer pairs using LLMs. This LLM-generated data can be used to train the chart VQA models, thus improving their reasoning ability.", + "bbox": [ + 75, + 136, + 468, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key innovation in our method lies in the Synthesize Step-by-Step strategy, which leverages LLMs to generate data in a step-by-step manner. Similar to human annotators who are trained to break down complex tasks into easier subtasks, our LLM-based data generator generates step-by-step. Instead of generating question-answer all at once, the generator learns to generate the question first, then decompose the complex question into step-by-step sub-questions (rationales) and answer them one by one. Finally, the answers of the sub-questions are combined to derive the final answer using external tools like calculators or Python code. Specifically, we feed the image features extracted with a ViT [12] backbone into an LLM using a simple trainable projection layer. To supervise the training of this projection layer, we construct a template-based training corpus generated using a template-based QA generation pipeline. Importantly, as we will show in the paper, the step-by-step synthesizing, trained using these template-based rationales, is critical for generating good-quality data.", + "bbox": [ + 75, + 257, + 470, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments showcase the effectiveness of Synthesize Step-by-Step. Our LLM-augmented Data, LaMenDa, is used to train the downstream chart VQA models and its effectiveness is shown on two challenging datasets, ChartQA and PlotQA with the state-of-the-art accuracy. Notably, we significantly improve the performance from $38\\%$ to $54\\%$ on the challenging human-written questions in ChartQA dataset, which requires complex reasoning. We also demonstrate the advantage of step-wise generation over straightforward generation. With the results, we hope our work underscores the potential of synthetic data and encourages further exploration of data augmentation using LLMs for reasoning-heavy tasks.", + "bbox": [ + 75, + 545, + 468, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 753, + 218, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal models for text-heavy images fall into two categories: OCR-dependent or OCR-free. Models like LayoutLM [61], PaLI(-X) [7, 8], ChartBERT [1], DocFormerv2 [4], MQMA [53], DEED [54] rely on external OCR systems like ChartOCR [38] as either model input or training objectives. Recent top performing models, like Donut [26], Dessurt [10], Pix2struct [28], shift away from OCR thanks to vision transformers [11]. Pix2Struct [28] is a trans", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "former encoder-decoder model for multiple VQA tasks, pretrained with an HTML-masked prediction objective, using a large amount of web page screenshots. MATCHA [35] and DEPLOT [34] are two recent follow-up works, tuning Pix2struct with chart-specific tasks like chart de-rendering and math reasoning, which are used in our work.", + "bbox": [ + 496, + 90, + 890, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Datasets for chart VQA. Earlier datasets like DVQA [22], FigureQA [23], LeafQA [5] are based on synthetically generated images and templatic questions, which are easy to solve and prior models achieve $>95\\%$ accuracy. PlotQA [41] is also a synthetic dataset but contains open-vocabulary questions. ChartQA [40] is a recent dataset containing real images with human-written questions, which is much more challenging. Our work is mainly based on ChartQA and PlotQA, following the settings in [34, 35].", + "bbox": [ + 496, + 181, + 890, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLMs for multimodal tasks. There is a surge of interest in extending LLMs into the vision-and-language domain. Recent works align the image features into the input space of the pretrained LLMs using a lightweight module, e.g. Frozen [56], Flamingo [2], LLaVA [36], Otter [29], MiniGPT-4 [70], BLIP-2 [30], AdaptorV2 [13], etc. For example, LLaVA [36] uses a simple projection layer to align the modalities and build a visually-finetuned multimodal model; LLaVAR [67] extends it onto text-heavy images.", + "bbox": [ + 496, + 316, + 890, + 453 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reasoning step-by-step has been studied for a long time in the multimodal field. Compositional models like neural modular networks [3, 15, 19] or neural symbolic reasoning [32, 33, 39, 64] decompose complex visual questions into reasoning steps, and compose modules to solve the steps. For LLMs, chain-of-thought prompting is a simple yet effective method to boost the reasoning performance [27, 58, 59, 63, 68, 69]. Recent works like VisProg [16] and ViperGPT [51] utilize LLMs as the planner to compose domain-specific models [37, 43, 62, 66] and external tools or APIs [25, 42, 45, 48, 50] for solving complex tasks.", + "bbox": [ + 496, + 454, + 890, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLMs for data generation has been recently studied for language tasks [9, 49, 65] or multimodal tasks [36, 70]. This relates to knowledge distillation [17], which teaches a smaller model by learning from a larger teacher model, benefiting semi-supervised learning [6, 57]. Distillation from LLMs [14, 18, 31, 44, 60] can be completed by either learning from the logits or from the generated data directly. Our work is related but different from these works in that we focus on step-by-step generation, combining LLMs with templates and tools, for the chart reasoning task.", + "bbox": [ + 496, + 619, + 890, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 784, + 589, + 799 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a chart plot, our goal is to build a data generator that generates complex reasoning questions with answers. The generated data can be used for training the downstream chart VQA models.", + "bbox": [ + 496, + 809, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The overview of our LLM-based data generator is shown in Fig. 2. In this section, we describe the data generator in a", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13614", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3f5454cc676b243f2de2603e5cdfdc1b54d499dfe70a8af81e7d9896c8561793.jpg", + "image_caption": [ + "(b) straight-forward generation", + "(c) synthesize step-by-step", + "LLM Output :", + "Figure 2. Method overview. (a) The architecture of the LLM-based question-answer generator. Image features, projected by a linear layer concatenated with the predicted data table and a prompt, are fed into LLM for QA generation. (b) Straight-forward generation, where questions with answers are generated in a straightforward way. (c) Synthesizing step-by-step, which breaks the question down into rationale programs and derives the answer by executing the program." + ], + "image_footnote": [ + "Prompt : Ask a question about the chart plot and answer it", + "LLM Output : What is the average of the two largest categories?13%", + "Prompt: Ask a question about the chart plot and think step-by-step to answer it.", + "Question: What is the average of the two largest categories?", + "Rationale: ans_0 = VQA(\"What is the value of 2002?\") # 16.2%", + "ans_1 = VQA(\"What is the value of 2003?\") # 16.7%", + "ans $=$ avg(ans_0, ans_1)", + "Code execution: ans_0 = 16.2%, ans_1 = 16.7%, ans = 16.5%" + ], + "bbox": [ + 120, + 80, + 478, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "progressive manner. We first introduce an intuitive, straightforward method for data generation. Then, we introduce Synthesize Step-by-Step, which enhances the straightforward method with step-wise rationales. Finally, we introduce the settings for training the downstream chart VQA models using our generated QA data.", + "bbox": [ + 75, + 388, + 468, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Architecture", + "text_level": 1, + "bbox": [ + 76, + 486, + 210, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The architecture of our LLM-based data generator is shown in Fig. 2 (a). We leverage MPT-7B [55] as the LLM for data generation. The inputs $X$ to LLM is a concatenation of three components: the projected image features $F$ , the underlying data table for the chart image $T$ , and a natural language prompt $P$ :", + "bbox": [ + 75, + 508, + 468, + 599 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nX = \\left[ F; T; P \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 606, + 323, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To extract the image features, we consider the pretrained visual transformer (ViT) [11] in CLIP [46]. The ViT takes in the image patches, and extracts features for each image patch. Then we use a trainable linear layer to project the extracted image features into the input embedding space of LLM. This architecture, i.e. linear layer for visual feature alignment, has been shown effective by recent multimodal large models like LLaVA [36].", + "bbox": [ + 75, + 628, + 468, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The predicted data table is critical to encode the textual information in the image. Because the pretrained ViT (in CLIP [46]) is not specifically trained for text-heavy images, the image features have limited OCR ability. Therefore, we use DEPLOT [34], which is designed for information extraction from charts, to predict the underlying data table for the given chart image. The linearized data table, which is a sequence of word tokens, is fed into the LLM.", + "bbox": [ + 75, + 750, + 470, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, the projected image features, the data table, and a prompt like \"Ask a question about the given chart plot and", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "answer it\" (rephrased in several different ways), are provided as the input to LLM, for question generation.", + "bbox": [ + 498, + 388, + 890, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Straight-forward data generation", + "text_level": 1, + "bbox": [ + 500, + 429, + 792, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the above inputs, a straightforward manner for data generation is to have the LLM generate question and answer directly, as shown in Fig. 2 (b). The output $Y$ is the generated question $Q$ , followed by the corresponding answer $Q$ :", + "bbox": [ + 498, + 453, + 890, + 513 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nY = [ Q; A ]\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 526, + 736, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This straightforward data generator can be trained using the existing chart VQA datasets, which contains chart images paired with question and answers. The model is trained with a next-token prediction loss. Note that the ViT and LLM are pretrained and kept frozen during training. The only trainable parameters are the linear projection layer1.", + "bbox": [ + 496, + 555, + 890, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As our experiments will show, this straightforward generation method yields reasonable results, generating questions and answers that are helpful for training the chart VQA models. However, the generated answers are noisy, which is improved using step-by-step generation, described next.", + "bbox": [ + 496, + 647, + 890, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Synthesize step-by-step", + "text_level": 1, + "bbox": [ + 500, + 750, + 714, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here we introduce Synthesize Step-by-Step, which shares the same model architecture as straight-forward generation, but performs data synthesizing in a step-by-step manner.", + "bbox": [ + 496, + 772, + 890, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of having the model generate answers directly, we train the LLM to generate them in a step-by-step manner. The LLM generates the question $Q$ with the corresponding", + "bbox": [ + 496, + 818, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1We also tried a two-stage training procedure, i.e. finetune the whole model after training the projection layer, but did not see performance gain.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13615", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/21ba88fae6548e22fc6795757509dad0e68187b39a4a53bfe5e086d0243165df.jpg", + "image_caption": [ + "Figure 3. Example of template-based question generation. Given a manually created template with defined program, by querying the image SVG annotations, templatic questions and answers can be generated, with rationales." + ], + "image_footnote": [], + "bbox": [ + 163, + 80, + 808, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rationales $R$ . Then the answer $A$ is derived by executing the rationales using python scripts:", + "bbox": [ + 75, + 324, + 470, + 356 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY = [ Q; R ], A = \\mathrm {E x e c u t e} (R)\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 369, + 375, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the example shows in Fig. 2 (c), the LLM outputs the generated question and a decomposition of this complex reasoning question into a series of atom reasoning steps, i.e. rationales. The rationales $R$ a step-by-step executable program, where each step in the program can be either an atom extractive VQA questions (e.g. ans_0 = VQA (\"What is the value of 2002?\")) or a Python-like program call for math calculation, counting, comparison, etc. (e.g. ans = avg(ans_0, ans_1)). The rationale program is then parsed and executed using Python in order to derive the final answer. During execution, there are two ways to answer the atom extractive VQA questions: we can either use an off-the-shelf VQA model, or directly predict the answer using the LLM data generator itself. We experiment with both options and find the latter option to be more effective, as well as more simple, without requiring additional VQA models.", + "bbox": [ + 75, + 400, + 468, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training with rationales. To train the step-by-step generator, data with rationales is required. This is different from the straightforward generation, which only requires question-answer pairs to train. Because existing datasets do not contain the rationale annotations for each question, we introduce a template-based QA generation pipeline, which will be described in Sec. 3.4, to generate synthetic questions with rationales. With the template-generated data as the training corpus, we tune the generator for step-by-step generation. An additional advantage of training with template-generated data is that, in inference time, the generator can be prompted with different prompts for controllable generation. The prompts may specify the format or the type of the questions, e.g. \"the question should start with how many\", \"the question should ask about the colors in the image\", etc. A complete list of the prompts can be found in Appendix.", + "bbox": [ + 75, + 659, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Post hoc filtering. We take an additional post hoc filtering step to filter out the noisy questions. We use the decoding score (for the whole generated sequence including question and rationales) as an indicator and filter out the questions with decoding scores lower than a threshold. Better filtering strategies, like using LLMs for data filtering, can be explored, which we will leave as future work.", + "bbox": [ + 496, + 324, + 890, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Tuning corpus: template-based QA generation", + "text_level": 1, + "bbox": [ + 498, + 439, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here we describe how we use a template-based pipeline to generate questions with rationales for training the LLM-based generator. The method is motivated by the success of synthetic compositional questions in VQA datasets like CLEVR [21], Super-CLEVR [33] and GQA [20].", + "bbox": [ + 496, + 462, + 890, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3 illustrates the template-based question generation method. We first manually create question templates, then instantiate the templates into questions and answers by querying the rich SVG annotations of the chart image (as shown in the dotted boxes in the figure). For example, the template \"What is the average of the $\\langle N\\rangle ?$ can be instantiated into \"what is the average of the more category?\". The argument $\\langle N\\rangle$ is instantiated into \"the more category\" according to the SVG annotations, which includes the category names, values, and visual elements like colors and bounding boxes. Moreover, for each template, we define a reasoning program (in a domain-specific language) containing a series of reasoning operations (e.g. select_color, query_value) that specifies the reasoning process needed to answer the question. The program is then instantiated into a step-by-step rationale.", + "bbox": [ + 496, + 537, + 892, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that there are limitations to this template-based QA generation pipeline. The pipeline relies on the rich annotations extracted automatically from the source SVG files, and thus cannot generalize to unlabeled images. Moreover, even though some existing datasets like ChartQA provide such rich annotations crawled from websources, they are inevitably noisy, containing missing or incorrect values. In contrast, the LLM generator is more generalizable, being", + "bbox": [ + 496, + 780, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13616", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "able to generate data for images in the wild. Therefore, in our work, these template-based questions are used as the training corpus of the LLM-based generator, which is more generalizable and scalable.", + "bbox": [ + 75, + 90, + 470, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Evaluation of generated data", + "text_level": 1, + "bbox": [ + 76, + 160, + 334, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To measure the usefulness of the generated QA, we use the generated data to train the chart VQA models. The accuracy improvement of chart VQA models, after training with the generated QA, is taken as the indicator of the data quality. We select MATCHA [35], which is a transformer encoder-decoder model pretrained on large-scale text-heavy images, as our primary evaluation model, considering its modern architecture and state-of-the-art performance on standard datasets2.", + "bbox": [ + 75, + 183, + 468, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Summary. Our data generation procedure contains the following steps: first, we generate templatic QA corpus with rationales using the rich SVG annotations, which is then used to tune the LLM-based data generator; next, the LLM-based generator generates questions with step-by-step rationales; then we use Python to execute the rationales and derive the answers; finally, this LLM-augmented data, LAMENDA, is used to train the downstream VQA model MATCHA. As the experiments will show, the augmented data leads to significant performance improvements.", + "bbox": [ + 75, + 319, + 470, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 483, + 200, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Set up", + "text_level": 1, + "bbox": [ + 76, + 508, + 161, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. Following [34, 35], we run experiments on two chart VQA datasets: ChartQA and PlotQA $^{3}$ . ChartQA [40] contains 21k real-world chart images crawled from four web sources, with 9k human-annotated (Human split) and 23k machine-augmented (Augmented split) QA pairs. The augmented questions are mostly extractive questions, while the human-annotated questions are free-form and require complex reasoning. PlotQA [41] is a large synthetic dataset containing 224k chart images generated from real-world online data, with 37M synthetic questions and answers divided into two splits: V1 (8M) and V2 (29M). In addition, to show the generalizability of our LLM data generator, we generate QA for images from additional chart captioning datasets, which do not provide QAs or rich SVG annotations. We use images from three datasets: Chart-to-Text [24], VisText [52] and ChartSumm [47]. Chart-to-Text contains 35k images from \"Statistica\" and 9k images from \"Pew\"; VisText contains 9k synthetic charts of various styles; ChartSumm contains 84k images from \"Statistica\"", + "bbox": [ + 75, + 531, + 468, + 819 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2While there are two-stage models of better performance, which extracts information from charts then prompts LLMs like Codex to answer questions, MATCHA is the best single-stage model.", + "3There are other chart VQA datasets like FigureQA, DVQA, etc., but since they are out-dated and models reach $>95\\%$ accuracy, we focus on more challenging settings." + ], + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and \"Knoema\". Note that we only use the images from the chart captioning datasets without using the captions.", + "bbox": [ + 496, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generated data. Tab. 1 shows the statistics of the generated QA. To construct the template-based tuning corpus, we use 28 templates to generate $357\\mathrm{k}$ question-answer pairs for the ChartQA training images. The template list and the template QA can be found in the Appendix. We train our LLM data generator using the template-based corpus, and generate QAs for images in ChartQA, PlotQA, and the three chart captioning datasets. For chartQA training images, we generate $403\\mathrm{k}$ QAs, which reduces to $326\\mathrm{k}$ after step-by-step execution (dropping ones that cannot be executed) and filtering based on the decoding scores using a threshold of -10. For PlotQA training images, we generate 3M QAs, which reduces to $1.7\\mathrm{M}$ after execution and filtering. For the $137\\mathrm{k}$ chart images in the chart captioning datasets, we generate $1.6\\mathrm{M}$ QAs.", + "bbox": [ + 496, + 121, + 892, + 348 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7987d56c136995a3b661d8dba5a735174134c110bc29252ed6a5a346aa4ac967.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetImagesGeneratedFiltered
TemplateChartQA18,317356,606-
LAMENDAChartQA18,317402,974326,160
PlotQA157,0703,455,5391,726,822
ChartCap137,2812,987,7181,635,364
", + "bbox": [ + 504, + 353, + 890, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Statistics of generated data, before and after filtering.", + "bbox": [ + 509, + 439, + 879, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. Using the 357k template QA, we tune the data generator for 5 epochs with a batch size of 128. We use a cosine learning rate schedule with a max learning rate of 2e-3. The training takes about 12 hours using 8 A100 GPUs. The trainable projection layer is initialized with the first stage checkpoint of LLaVA [36]. We select the best checkpoint at 14k iterations as our best generator. For MATCHA training, we take the pretrained checkpoints (base size) from HuggingFace and finetune the models with a combination of our generated questions and the original ChartQA or PlotQA training questions. The training setting follows [28, 35]. We use a cosine scheduler with 500 warm-up steps and a max learning rate of 0.0001. The batch size is 256. Without special description, the models are trained for 10k iterations for ChartQA (combining human and augmented splits) and 20k iterations for PlotQA because it is much larger (V1 and V2 are trained separately). We set the max number of image patches as 4096. The ChartQA training takes 3 days on 8 32GB V100 GPUs. Considering the training time, in some experiments, we reduce the image patch number to 1024 for quick experiments in 12 hours with the same setting.", + "bbox": [ + 496, + 455, + 890, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. We suggest that both strict accuracy and relaxed accuracy should be reported, especially for the ChartQA dataset. While existing works [34, 35, 40, 41] evaluate answers are evaluated using relaxed accuracy, which allows $5\\%$ error for numerical answers, we find that this relaxed accuracy metric is problematic for ChartQA.", + "bbox": [ + 496, + 787, + 890, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "4https://huggingface.co/google/matcha-base", + "bbox": [ + 514, + 886, + 849, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13617", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e68ba9bca98d05b7dd395a34a448fb324ac34be0531b483d770ed28e713e4f76.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AccuracyRelaxed Accuracy
avghumanaugmentavghumanaugment
0baseline (MATCHA [35])47.7630.2165.3158.5439.5877.50
1syna48.8028.6568.9659.4338.6580.21
2synat w/ VQA answer47.7628.4467.0857.9737.8178.12
3synat w/ step-VQA answer52.5537.2967.8164.3848.8579.90
4synat w/ synat answer53.7537.9269.5865.1048.8581.35
5synat w/ step-synat answer55.2140.2170.2166.4151.3581.46
6synat w/ step-synat answer + tpl.57.3442.6072.0867.8653.1282.60
7synat w/ step-synat answer + tpl. + addi.58.6545.6271.6769.8456.5683.12
", + "bbox": [ + 142, + 82, + 834, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Main results on ChartQA val split. The data generator trained on Annotated questions is denoted as $\\mathrm{syn}_a$ ; the data generator trained on both Annotated questions and our Template-generated questions $\\mathrm{syn}_{at}$ is denoted as $\\mathrm{syn}_{at}$ . Step-by-step generation is better than straight-forward generation (row 1-5). Training with a combination of step-by-step generated QA, template-generated QA (+mpl.), and additional images from chart captioning datasets (+addi.), leads to the best model. See Sec. 4.2 for more details.", + "bbox": [ + 75, + 243, + 893, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For example, many questions in ChartQA ask about years (e.g. 1996, 2012), for which allowing $5\\%$ numerical errors (i.e. around 100 years) will result in incorrectly high results.", + "bbox": [ + 75, + 313, + 470, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main results", + "text_level": 1, + "bbox": [ + 76, + 372, + 210, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We study our data generation method with different settings on the ChartQA dataset, with MATCHA as our baseline. We first generate QA data using our method, and then use the generated data to train the MATCHA model. The strict accuracy and relaxed accuracy reported on the ChartQA validation split are shown in Tab. 2.", + "bbox": [ + 75, + 397, + 468, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The baseline model MATCHA (row-0) achieves $47.76\\%$ accuracy and $58.54\\%$ relaxed accuracy. For this baseline, we finetune the pretrained MATCHA checkpoint for 10k iterations with the ChartQA training split (combing human and augmented questions), using 1024 image patches. This setting follows [35], except that a smaller number of image patches are used for faster experiments and less GPU memory requirement. We validate the correctness of our implementation by showing that the model reaches $63.7\\%$ relaxed accuracy using the original resolution, which matches the numbers reported in [35] (as shown in Tab. 3).", + "bbox": [ + 75, + 489, + 468, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Row-1 shows the results for the straight-forward data generation, where we train the data generator with QAs from the ChartQA training split and use the trained generator to directly generate questions and answers. This straightforward data generation improves the accuracy by $1\\%$ (from $47.76\\%$ to $48.80\\%$ ), suggesting that LLMs can be helpful in data augmentation for chart reasoning.", + "bbox": [ + 75, + 657, + 468, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Row-2 to row-5 shows the advantage of synthesizing step-by-step over straight-forward generation. We train the data generator with our template-generated QA, but the answers are generated with different methods. In row-2, we feed in the generated complex question into an additional VQA model (i.e. baseline MATCHA) to get the answers; in row-3, we generate the step-by-step rationales, and the stepwise questions are fed into the VQA model to get step-wise answers, then the final answer is derived using the step", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "wise answers. As can be seen, the step-wise answer generation (52.55%) is significantly better than the direct answer generation (47.76%), which suggests the importance of decomposing the complex questions into atom sub-questions. Row-4 and row-5 follow a similar setting, except that we use the LLM data generator to generate the answers instead of the additional VQA model. Two findings can be drawn: (a) LLM-generated answers are better than VQA-predicted answers, possibly because the LLM data generator has access to the predicted data tables as inputs; (b) step-by-step synthesizing outperforms direct generation, for both VQA-predicted and LLM-generated answers.", + "bbox": [ + 496, + 313, + 893, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In row-6, we combine the step-by-step generated data with the templated-generated QAs and show that training with both leads to further better performance (57.34%). In row-7, we generate QAs for the chart images from additional chart captioning datasets. The final model, trained with a combination of all the synthetic data, gets 58.65% accuracy. Compared with the baseline, training with the augmented data leads to a 10.89% accuracy improvement. Moreover, the improvement is more significant on the human-written questions (+15.41%), which requires complex reasoning, than the augmented extractive questions (+6.36%). The results suggest that our synthetic data improves both the extractive ability and the reasoning ability, with the latter being more significant. We show examples in the Appendix.", + "bbox": [ + 496, + 497, + 893, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with SoTA", + "text_level": 1, + "bbox": [ + 500, + 739, + 718, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To compare our method with the best performing models, we apply our generated data onto two models, Pix2Struct [28] and MATCHA [35], and report the results on the ChartQA test split in Tab. 3. For faster training, we first train the model (for both Pix2struct and MATCHA) with a max of 1024 image patches using our best setting (i.e. row-7 in Tab. 2), then finetune with 4096 image patches to get the final result. We report both the 1024-resolution results and the 4096-resolution results.", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13618", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f8fcd1e1be9f068e4cc8199e870c95e59fb6e44dd176642ef15f0a83d9a29b1a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
paramsaccuracyRelaxed Accuracy
avghumanaugmentavghumanaugment
VL-T5-OCR [40]----41.56--
VisionTapas-OCR [40]----45.52--
Pix2Struct4096 [28]282M---56.030.581.6
MATCHA4096 [35]282M---64.238.290.2
DEPLOT+FlanPaLM+Codex [34]540B---79.367.691.0
PaLI-X [7]55B---72.3--
Pix2Struct4096 (reimpl.)282M50.4424.6476.2458.2432.0084.48
Pix2Struct1024 + LAMENDA282M51.7734.6968.8563.0246.2579.79
Pix2Struct4096 + LAMENDA282M55.8340.1071.5666.8251.6781.98
MATCHA4096 (reimpl.)282M55.6831.0480.3263.7237.7689.68
MATCHA1024 + LAMENDA282M60.2841.9278.6469.8851.7688.00
MATCHA4096 + LAMENDA282M63.3644.8881.8472.6453.9291.36
", + "bbox": [ + 150, + 75, + 823, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As can be seen, our generated data improves the overall relaxed accuracy of Pix2Struct by $8.58\\%$ and MATCHA by $8.92\\%$ . Moreover, the improvement is much more significant on human-written questions, e.g. the relaxed accuracy of MATCHA on human-written questions improves from $38\\%$ to $54\\%$ . Interestingly, with our augmented data, the low-resolution (1024) models can already outperform their high-resolution (4096) counterparts without data augmentation. The significant performance gain indicates that learning with synthetic data is extremely helpful for better reasoning ability with human-written queries.", + "bbox": [ + 75, + 334, + 470, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our best performing MATCHA achieves $72.6\\%$ relaxed accuracy, which beats all existing end-to-end chart reasoning models. Note that DEPLOT+FlanPaLM+Codex is a two-stage model that first extracts information from the charts, then prompts external LLMs like FlanPalm (540B) Codex (12B) for question answering. The method relies on much larger models of billions of parameters and thus is not comparable with our method, which is much smaller and not rely on external LLMs for inference.", + "bbox": [ + 75, + 500, + 468, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Generalization", + "text_level": 1, + "bbox": [ + 76, + 647, + 227, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To show the generalization ability of the proposed data generation method, we run experiments on the PlotQA dataset. We directly use our data generator (trained on template questions on ChartQA) to generate QA for the PlotQA training images. The results are shown in Tab. 4. As can be seen, our method, when applied to MATCHA, achieves state-of-the-art performance (92.89% relaxed accuracy) on PlotQA. Despite the synthetic nature of this dataset and the saturated model performance, our augmented data can still help on both low-resolution (+3.46%) and high-resolution (+0.92%) settings.", + "bbox": [ + 75, + 670, + 468, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablations and analysis", + "text_level": 1, + "bbox": [ + 76, + 845, + 285, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative analysis and failure cases. Fig. 5 shows examples of our LLM-generated QA, including good and bad ex", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "samples. Most generated questions are specific to the given image, asking about the categories plotted, such as “2019” or “Tablet 2018”, indicating that the model is able to utilize the information of the input. Moreover, various questions are generated, including counting, colors, average, etc. In the failure example, the question requires comparing the “silver bar” with the “midnightblue” bar, with reasonable rationales. However, there are multiple bars in the specified colors in the image, resulting in ambiguous/invalid questions. When looking at more failure cases, we find that there are several typical failure modes. For example, questions are sometimes invalid/ambiguous (e.g. asking about the top/bottom bar for a vertical bar plot); questions can be hallucinated or irrelevant (e.g. asking about a category that is not shown in the figure); rationales and answers are not correct, etc., which opens up opportunities for future works. Before and after LAMENDA. To gain insights on how LAMENDA helps model training, we conduct a comparison analysis of MATCHA “before” and “after” training with our LLM-generated data, i.e. the baseline model versus our best model. According to the predictions of these two models, we divide the 960 human-written questions from ChartQA val split into 4 groups: both model answers correctly, both models answers wrongly, the before model answers correctly, but the after model answers wrongly, and", + "bbox": [ + 496, + 334, + 892, + 713 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0e84780e527b9d0462e722f1750db71eb9e6b731e6e9cf2f962d4fc81bbb3486.jpg", + "table_caption": [ + "Table 3. Comparison with SoTAs on ChartQA test split. Subscripts (1024, 4096) means image token numbers (resolution). With our generated data, both Pix2Struct and MATCHA achieve significantly better performance, even comparable to much larger models." + ], + "table_footnote": [], + "table_body": "
Relaxed Accuracy
avgV1V2
VisionTapas-OCR [40]53.9065.3042.50
VL-T5-OCR [40]65.9675.9056.02
DEPLOT+FlanPaLM+Codex [34]66.662.271.0
Pix2Struct [28]72.573.271.9
MATCHA4096 [35]91.592.390.7
MATCHA1024 (reimpl.)74.9573.8876.02
MATCHA1024 + LAMENDA78.4178.4478.38
MATCHA4096 (reimpl.)91.9792.6491.30
MATCHA4096 + LAMENDA92.8993.9491.84
", + "bbox": [ + 513, + 719, + 883, + 882 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparison with SoTAs on PlotQA test split. With our generated data, MATCHA achieves the SoTA performance.", + "bbox": [ + 498, + 887, + 890, + 915 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13619", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ac7fdad08d09825866a1179e3b4f14522bf0b5f0bfdc1c6bfe0db7d70be31ba.jpg", + "image_caption": [ + "Figure 4. Analysis of model predictions before and after training with LAMENDA. In I (left), ChartQA val questions are divided into four categories according to the model prediction correctness (before/after). II (right) further decomposes each category into different question types. LAMENDA improves the model prediction, and the improvement is most significant on reasoning questions." + ], + "image_footnote": [], + "bbox": [ + 166, + 85, + 803, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vice versa. The statistics of these four groups are shown in Fig. 4. From Fig. 4 I. (left side), we see that LAMENDA is helpful for $22\\%$ of the questions, while there are still $45\\%$ of the questions where both models cannot answer correctly. In Fig. 4 II. (right side), we randomly sample 30 questions from each of the four groups and manually classify them into six question types: extractive, count, spatial, color, reasoning, and others. As shown in (a), i.e. the questions where the \"before\" model answers wrongly and our model answers correctly, most $(70\\%)$ of the questions belong to the reasoning category, indicating that our data helps most on the reasoning questions. Our data also helps with color questions $(10\\%)$ and count questions $(17\\%)$ . However, our data does not help with extractive questions $(0\\%)$ . In addition, comparing the distribution of (c) both correct with (d) both wrong, we can see that extractive questions are easy for the models, while reasoning questions are very challenging.", + "bbox": [ + 75, + 285, + 472, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, Fig. 4 shows the effectiveness of our data for the reasoning questions over other question types. However, reasoning questions are still a big challenge for current models, which requires further exploration.", + "bbox": [ + 76, + 542, + 470, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different question types. In Tab. 5, we study the improve", + "bbox": [ + 76, + 606, + 468, + 622 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/709dbcf39525bede908916656b06832d45d92e989ca0be34efb3600a1540b0ac.jpg", + "image_caption": [ + "Figure 5. Examples of generated QA, 3 good examples and 1 failure case. Rationales are only shown for 2 examples for brevity." + ], + "image_footnote": [], + "bbox": [ + 138, + 638, + 406, + 873 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ments brought by each type of the generated questions. The total 326k generated QAs are grouped into 7 different groups, including questions about color understanding, spatial reasoning, counting, minimum/maximum reasoning, average, comparison (e.g. larger than, smaller than), and other math calculations. This grouping is done by different prompting of the data generation LLM, as specified in the Appendix. The improvements from different question types are distinct and can be accumulated. As we see, each of the seven types improves the model performance by less than $3\\%$ , while combining all of them leads to $7.03\\%$ improvement. Moreover, among the seven types, the \"math calculation\" questions brings the most improvements $(2.76\\%)$ .", + "bbox": [ + 496, + 285, + 893, + 482 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aac68200a983c4d629bee3a6f603efb4ec0f6b8eb071b81d79948471e4aa9e75.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#QAavghumanaug
baseline28,29947.7630.2165.31
+colors+46,51247.8129.9065.73
+spatial+49,38749.2730.6367.92
+count+36,70547.4029.1765.63
+minmax+71,78849.5831.8767.29
+average+50,71748.9631.3566.56
+compare+14,69047.6628.8566.46
+calculation+56,36150.5233.8567.19
+all326,16054.7939.2770.31
", + "bbox": [ + 560, + 489, + 834, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. A detailed look at the effect for different question types. Strict accuracy on ChartQA val split is shown. Each type helps and combining all of them leads to a further performance gain.", + "bbox": [ + 498, + 621, + 893, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 678, + 619, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose Synthesize Step-by-Step, which utilizes LLMs to augment the training data for the chart reasoning task. We show that the step-by-step generation, which decomposes the complex reasoning questions into easier sub-questions, is critical for synthesizing good-quality data. We run experiments on two standard benchmarks and show significant improvements by training with the augmented data. Specifically, the relaxed accuracy on the human-written questions is improved from $37.8\\%$ to $53.9\\%$ on ChartQA. The large improvement indicates the promise of boosting the model's reasoning ability by data generation using LLMs, from which the strong step-by-step reasoning ability can be distilled into downstream models.", + "bbox": [ + 496, + 703, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13620", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mubashara Akhtar, Oana Cocarascu, and Elena Simperl. Reading and reasoning over chart images for evidence-based automated fact-checking. arXiv preprint arXiv:2301.11843, 2023. 2", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2", + "[3] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 39-48, 2016. 2", + "[4] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R Manmatha. Docformerv2: Local features for document understanding. arXiv preprint arXiv:2306.01733, 2023. 2", + "[5] Ritwick Chaudhry, Sumit Shekhar, Utkarsh Gupta, Pranav Maneriker, Prann Bansal, and Ajay Joshi. Leaf-qa: Locate, encode & attend for figure question answering. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3512–3521, 2020. 2", + "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 2", + "[7] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 2, 7", + "[8] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794, 2022. 2", + "[9] John Joon Young Chung, Ece Kamar, and Saleema Amershi. Increasing diversity while maintaining accuracy: Text data generation with large language models and human interventions. arXiv preprint arXiv:2306.04140, 2023. 2", + "[10] Brian Davis, Bryan Morse, Brian Price, Chris Tensmeyer, Curtis Wigington, and Vlad Morariu. End-to-end document recognition and understanding with dessurt. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 2", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at" + ], + "bbox": [ + 78, + 114, + 470, + 890 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "scale, 2021. 2", + "[13] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 2", + "[14] Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. Knowledge distillation of large language models. arXiv preprint arXiv:2306.08543, 2023. 2", + "[15] Nitish Gupta, Kevin Lin, Dan Roth, Sameer Singh, and Matt Gardner. Neural module networks for reasoning over text. arXiv preprint arXiv:1912.04971, 2019. 2", + "[16] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 2", + "[17] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2", + "[18] Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes. arXiv preprint arXiv:2305.02301, 2023. 2", + "[19] Ronghang Hu, Jacob Andreas, Trevor Darrell, and Kate Saenko. Explainable neural computation via stack neural module networks. In Proceedings of the European conference on computer vision (ECCV), pages 53-69, 2018. 2", + "[20] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 4", + "[21] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. 4", + "[22] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering, 2018. 2", + "[23] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Akos Kadar, Adam Trischler, and Yoshua Bengio. Figureqa: An annotated figure dataset for visual reasoning, 2018. 2", + "[24] Shankar Kantharaj, Rixie Tiffany Leong, Xiang Lin, Ahmed Masry, Megh Thakkar, Enamul Hoque, and Shafiq Joty. Chart-to-text: A large-scale benchmark for chart summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4005–4023, 2022. 5", + "[25] Tushar Khot, Harsh Trivedi, Matthew Finlayson, Yao Fu, Kyle Richardson, Peter Clark, and Ashish Sabharwal. Decomposed prompting: A modular approach for solving complex tasks. arXiv preprint arXiv:2210.02406, 2022. 2", + "[26] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang" + ], + "bbox": [ + 503, + 92, + 890, + 892 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13621", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "doo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 2", + "[27] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 2", + "[28] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In International Conference on Machine Learning, pages 18893–18912. PMLR, 2023. 2, 5, 6, 7", + "[29] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 2", + "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2", + "[31] Liunian Harold Li, Jack Hessel, Youngjae Yu, Xiang Ren, Kai-Wei Chang, and Yejin Choi. Symbolic chain-of-thought distillation: Small models can also\" think\" step-by-step. arXiv preprint arXiv:2306.14050, 2023. 2", + "[32] Zhuowan Li, Elias Stengel-Eskin, Yixiao Zhang, Cihang Xie, Quan Hung Tran, Benjamin Van Durme, and Alan Yuille. Calibrating concepts and operations: Towards symbolic reasoning on real images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14910-14919, 2021. 2", + "[33] Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Kortylewski, Wufei Ma, Benjamin Van Durme, and Alan L Yuille. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14963-14973, 2023. 2, 4", + "[34] Fangyu Liu, Julian Martin Eisenschlos, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Wenhu Chen, Nigel Collier, and Yasemin Altun. Deplot: One-shot visual language reasoning by plot-to-table translation. arXiv preprint arXiv:2212.10505, 2022. 2, 3, 5, 7", + "[35] Fangyu Liu, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Yasemin Altun, Nigel Collier, and Julian Martin Eisenschlos. Matcha: Enhancing visual language pretraining with math reasoning and chart derendering. arXiv preprint arXiv:2212.09662, 2022. 1, 2, 5, 6, 7", + "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 3, 5", + "[37] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842, 2023. 2", + "[38] Junyu Luo, Zekun Li, Jinpeng Wang, and Chin-Yew Lin." + ], + "bbox": [ + 78, + 90, + 468, + 893 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chartocr: Data extraction from charts images via a deep hybrid framework. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1917-1925, 2021. 2", + "[39] Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B Tenenbaum, and Jiajun Wu. The neuro-symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision. arXiv preprint arXiv:1904.12584, 2019. 2", + "[40] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 1, 2, 5, 7", + "[41] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2020. 2, 5", + "[42] Aaron Parisi, Yao Zhao, and Noah Fiedel. Talm: Tool augmented language models. arXiv preprint arXiv:2205.12255, 2022. 2", + "[43] Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334, 2023. 2", + "[44] Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). Association for Computational Linguistics, 2019. 2", + "[45] Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, et al. Toolllm: Facilitating large language models to master 16000+ real-world apis. arXiv preprint arXiv:2307.16789, 2023. 2", + "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[47] Raian Rahman, Rizvi Hasan, Abdullah Al Farhad, Md Tahmid Rahman Laskar, Md Hamjajul Ashmafee, and Abu Raihan Mostofa Kamal. Chartsumm: A comprehensive benchmark for automatic chart summarization of long and short summaries. arXiv preprint arXiv:2304.13620, 2023. 5", + "[48] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. arXiv preprint arXiv:2302.04761, 2023. 2", + "[49] Timo Schick and Hinrich Schütze. Generating datasets with pretrained language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6943-6951, 2021. 2", + "[50] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580, 2023. 2", + "[51] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt:" + ], + "bbox": [ + 501, + 92, + 890, + 892 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13622", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128, 2023. 2", + "[52] Benny J Tang, Angie Boggust, and Arvind Satyanarayan. Vistext: A benchmark for semantically rich chart captioning. arXiv preprint arXiv:2307.05356, 2023. 5", + "[53] Peng Tang, Srikar Appalaraju, R Manmatha, Yusheng Xie, and Vijay Mahadevan. Multiple-question multiple-answer text-vqa. arXiv preprint arXiv:2311.08622, 2023. 2", + "[54] Peng Tang, Pengkai Zhu, Tian Li, Srikar Appalaraju, Vijay Mahadevan, and R Manmatha. Deed: Dynamic early exit on decoder for accelerating encoder-decoder transformer models. arXiv preprint arXiv:2311.08623, 2023. 2", + "[55] MN Team et al. Introducing mpt-7b: A new standard for open-source, commercially usable llms, 2023. 3", + "[56] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 2", + "[57] Shuohang Wang, Yang Liu, Yichong Xu, Chenguang Zhu, and Michael Zeng. Want to reduce labeling cost? gpt-3 can help. arXiv preprint arXiv:2108.13487, 2021. 2", + "[58] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171, 2022. 2", + "[59] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 2", + "[60] Peter West, Chandra Bhagavatula, Jack Hessel, Jena D Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, and Yejin Choi. Symbolic knowledge distillation: from general language models to commonsense models. arXiv preprint arXiv:2110.07178, 2021. 2", + "[61] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1192-1200, 2020. 2", + "[62] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023. 2", + "[63] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601, 2023. 2", + "[64] Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Josh Tenenbaum. Neural-symbolic vqa: Disentangling reasoning from vision and language understanding. Advances in neural information processing systems, 31, 2018. 2", + "[65] Yue Yu, Yuchen Zhuang, Jieyu Zhang, Yu Meng, Alexander Ratner, Ranjay Krishna, Jiaming Shen, and Chao Zhang. Large language model as attributed training data generator: A tale of diversity and bias. In Thirty-seventh Conference on" + ], + "bbox": [ + 78, + 90, + 468, + 891 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2", + "[66] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 2", + "[67] Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 2", + "[68] Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. Automatic chain of thought prompting in large language models. arXiv preprint arXiv:2210.03493, 2022. 2", + "[69] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 2", + "[70] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 390 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13623", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/58e8331e-f572-44c9-b724-02ca5cd604a2_model.json b/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/58e8331e-f572-44c9-b724-02ca5cd604a2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7760e933a77317a084e3c083ec5b0c461b6dda79 --- /dev/null +++ b/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/58e8331e-f572-44c9-b724-02ca5cd604a2_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.182, + 0.131, + 0.788, + 0.177 + ], + "angle": 0, + "content": "Synthesize Step-by-Step: Tools, Templates and LLMs as Data Generators for Reasoning-Based Chart VQA" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.203, + 0.756, + 0.221 + ], + "angle": 0, + "content": "Zhuowan Li\\(^{1}\\) * Bhavan Jasani\\(^{2}\\) * Peng Tang\\(^{2}\\) Shabnam Ghadar\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.303, + 0.222, + 0.669, + 0.24 + ], + "angle": 0, + "content": "1 Johns Hopkins University 2 AWS AI Labs" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.242, + 0.717, + 0.256 + ], + "angle": 0, + "content": "zili110@jhu.edu {bjasani, tangpen, shabnam}@amazon.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.324, + 0.474, + 0.734 + ], + "angle": 0, + "content": "Understanding data visualizations like charts and plots requires reasoning about both visual elements and numerics. Although strong in extractive questions, current chart visual question answering (chart VQA) models suffer on complex reasoning questions. In this work, we address the lack of reasoning ability by data augmentation. We leverage Large Language Models (LLMs), which have shown to have strong reasoning ability, as an automatic data annotator that generates question-answer annotations for chart images. The key innovation in our method lies in the Synthesize Step-by-Step strategy: our LLM-based data generator learns to decompose the complex question into step-by-step sub-questions (rationales), which are then used to derive the final answer using external tools, i.e. Python. This step-wise generation procedure is trained on synthetic data generated using a template-based QA generation pipeline. Experimental results highlight the significance of the proposed step-by-step generation. By training with the LLM-augmented data (LAMENDA), we significantly enhance the chart VQA models, achieving the state-of-the-art accuracy on the ChartQA and PlotQA datasets. In particular, our approach improves the accuracy of the previous state-of-the-art approach from \\(38\\%\\) to \\(54\\%\\) on the human-written questions in the ChartQA dataset, which needs strong reasoning. We hope our work underscores the potential of synthetic data and encourages further exploration of data augmentation using LLMs for reasoning-heavy tasks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.763, + 0.21, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.788, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Data visualizations like charts and plots play an important role in real-world data analysis applications. Unlike natural images, chart are text-heavy and data-driven, thus requiring better visual perception (e.g. OCR) and cognitive reasoning (e.g. math calculation, matching the legends with bars). For" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.28, + 0.891, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.472, + 0.895, + 0.6 + ], + "angle": 0, + "content": "Figure 1. Existing chart VQA models struggle with complex reasoning questions. We attribute this to limited reasoning questions in existing datasets and address it by data augmentation. We fine-tune an LLM-based data generator that automatically generates question-answer annotations given a chart image. Our key innovation is Synthesize Step-by-Step, which breaks complex questions down into easy steps that could be solved using external tools. We use templates to train the LLM. Training with LLM-augmented data, LAMENDA, greatly enhances the chart VQA models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.689 + ], + "angle": 0, + "content": "example, to answer \"what is the total of Democrats and Republicans in 2010\" in Fig. 1, the model needs to recognize the texts, match the legends of \"democrats\" and \"republicans\" to the bars, then calculate the sum. This is a challenging task requiring multimodal perception and reasoning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Compounding the complexity, the inherent difficulty of chart reasoning also amplifies the cost of collecting human annotations. Most existing datasets are synthetically generated using simple rules and heuristics, posing a risk of model overfitting. Among the datasets, only ChartQA [40] contains complex human-annotated question-answer pairs, although there are very few of those per image. For example, for the image in Fig. 1, while a lot more questions can be asked, only two questions are annotated (What is the total of Republicans and Democrats in 2010? Which color represents Republican?). Importantly, these human-written questions pose great challenges for contemporary models. For example, the state-of-the-art model [35] achieves only \\(38\\%\\) accuracy when confronted with human-written ques" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.47, + 0.9 + ], + "angle": 0, + "content": "*indicates equal contributions. The work was done when Zhuowan Li was an intern at Amazon." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13613" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "tions that require multi-step complex reasoning. Approximately half of the errors, as reported in [35], are due to reasoning failures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.47, + 0.258 + ], + "angle": 0, + "content": "In this work, we propose to leverage large language models (LLMs) as automatic data annotators for chart reasoning. Recently, LLMs have demonstrated strong reasoning ability for complex tasks [27, 58, 59, 63, 68, 69]. Our motivation is to take advantage of this strong reasoning ability and generate question-answer pairs using LLMs. This LLM-generated data can be used to train the chart VQA models, thus improving their reasoning ability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.471, + 0.544 + ], + "angle": 0, + "content": "The key innovation in our method lies in the Synthesize Step-by-Step strategy, which leverages LLMs to generate data in a step-by-step manner. Similar to human annotators who are trained to break down complex tasks into easier subtasks, our LLM-based data generator generates step-by-step. Instead of generating question-answer all at once, the generator learns to generate the question first, then decompose the complex question into step-by-step sub-questions (rationales) and answer them one by one. Finally, the answers of the sub-questions are combined to derive the final answer using external tools like calculators or Python code. Specifically, we feed the image features extracted with a ViT [12] backbone into an LLM using a simple trainable projection layer. To supervise the training of this projection layer, we construct a template-based training corpus generated using a template-based QA generation pipeline. Importantly, as we will show in the paper, the step-by-step synthesizing, trained using these template-based rationales, is critical for generating good-quality data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.742 + ], + "angle": 0, + "content": "Experiments showcase the effectiveness of Synthesize Step-by-Step. Our LLM-augmented Data, LaMenDa, is used to train the downstream chart VQA models and its effectiveness is shown on two challenging datasets, ChartQA and PlotQA with the state-of-the-art accuracy. Notably, we significantly improve the performance from \\(38\\%\\) to \\(54\\%\\) on the challenging human-written questions in ChartQA dataset, which requires complex reasoning. We also demonstrate the advantage of step-wise generation over straightforward generation. With the results, we hope our work underscores the potential of synthetic data and encourages further exploration of data augmentation using LLMs for reasoning-heavy tasks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.22, + 0.77 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Multimodal models for text-heavy images fall into two categories: OCR-dependent or OCR-free. Models like LayoutLM [61], PaLI(-X) [7, 8], ChartBERT [1], DocFormerv2 [4], MQMA [53], DEED [54] rely on external OCR systems like ChartOCR [38] as either model input or training objectives. Recent top performing models, like Donut [26], Dessurt [10], Pix2struct [28], shift away from OCR thanks to vision transformers [11]. Pix2Struct [28] is a trans" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.181 + ], + "angle": 0, + "content": "former encoder-decoder model for multiple VQA tasks, pretrained with an HTML-masked prediction objective, using a large amount of web page screenshots. MATCHA [35] and DEPLOT [34] are two recent follow-up works, tuning Pix2struct with chart-specific tasks like chart de-rendering and math reasoning, which are used in our work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.182, + 0.892, + 0.318 + ], + "angle": 0, + "content": "Datasets for chart VQA. Earlier datasets like DVQA [22], FigureQA [23], LeafQA [5] are based on synthetically generated images and templatic questions, which are easy to solve and prior models achieve \\(>95\\%\\) accuracy. PlotQA [41] is also a synthetic dataset but contains open-vocabulary questions. ChartQA [40] is a recent dataset containing real images with human-written questions, which is much more challenging. Our work is mainly based on ChartQA and PlotQA, following the settings in [34, 35]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.454 + ], + "angle": 0, + "content": "LLMs for multimodal tasks. There is a surge of interest in extending LLMs into the vision-and-language domain. Recent works align the image features into the input space of the pretrained LLMs using a lightweight module, e.g. Frozen [56], Flamingo [2], LLaVA [36], Otter [29], MiniGPT-4 [70], BLIP-2 [30], AdaptorV2 [13], etc. For example, LLaVA [36] uses a simple projection layer to align the modalities and build a visually-finetuned multimodal model; LLaVAR [67] extends it onto text-heavy images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.62 + ], + "angle": 0, + "content": "Reasoning step-by-step has been studied for a long time in the multimodal field. Compositional models like neural modular networks [3, 15, 19] or neural symbolic reasoning [32, 33, 39, 64] decompose complex visual questions into reasoning steps, and compose modules to solve the steps. For LLMs, chain-of-thought prompting is a simple yet effective method to boost the reasoning performance [27, 58, 59, 63, 68, 69]. Recent works like VisProg [16] and ViperGPT [51] utilize LLMs as the planner to compose domain-specific models [37, 43, 62, 66] and external tools or APIs [25, 42, 45, 48, 50] for solving complex tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.772 + ], + "angle": 0, + "content": "LLMs for data generation has been recently studied for language tasks [9, 49, 65] or multimodal tasks [36, 70]. This relates to knowledge distillation [17], which teaches a smaller model by learning from a larger teacher model, benefiting semi-supervised learning [6, 57]. Distillation from LLMs [14, 18, 31, 44, 60] can be completed by either learning from the logits or from the generated data directly. Our work is related but different from these works in that we focus on step-by-step generation, combining LLMs with templates and tools, for the chart reasoning task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.591, + 0.8 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Given a chart plot, our goal is to build a data generator that generates complex reasoning questions with answers. The generated data can be used for training the downstream chart VQA models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The overview of our LLM-based data generator is shown in Fig. 2. In this section, we describe the data generator in a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13614" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.082, + 0.48, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.084, + 0.738, + 0.101 + ], + "angle": 0, + "content": "(b) straight-forward generation" + }, + { + "type": "image_footnote", + "bbox": [ + 0.499, + 0.109, + 0.777, + 0.122 + ], + "angle": 0, + "content": "Prompt : Ask a question about the chart plot and answer it" + }, + { + "type": "image_footnote", + "bbox": [ + 0.499, + 0.131, + 0.848, + 0.146 + ], + "angle": 0, + "content": "LLM Output : What is the average of the two largest categories?13%" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.16, + 0.706, + 0.176 + ], + "angle": 0, + "content": "(c) synthesize step-by-step" + }, + { + "type": "image_footnote", + "bbox": [ + 0.499, + 0.18, + 0.834, + 0.204 + ], + "angle": 0, + "content": "Prompt: Ask a question about the chart plot and think step-by-step to answer it." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.213, + 0.575, + 0.225 + ], + "angle": 0, + "content": "LLM Output :" + }, + { + "type": "image_footnote", + "bbox": [ + 0.499, + 0.226, + 0.781, + 0.237 + ], + "angle": 0, + "content": "Question: What is the average of the two largest categories?" + }, + { + "type": "image_footnote", + "bbox": [ + 0.5, + 0.238, + 0.801, + 0.248 + ], + "angle": 0, + "content": "Rationale: ans_0 = VQA(\"What is the value of 2002?\") # 16.2%" + }, + { + "type": "image_footnote", + "bbox": [ + 0.55, + 0.25, + 0.8, + 0.261 + ], + "angle": 0, + "content": "ans_1 = VQA(\"What is the value of 2003?\") # 16.7%" + }, + { + "type": "image_footnote", + "bbox": [ + 0.55, + 0.263, + 0.67, + 0.275 + ], + "angle": 0, + "content": "ans \\(=\\) avg(ans_0, ans_1)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.499, + 0.284, + 0.842, + 0.297 + ], + "angle": 0, + "content": "Code execution: ans_0 = 16.2%, ans_1 = 16.7%, ans = 16.5%" + }, + { + "type": "list", + "bbox": [ + 0.499, + 0.213, + 0.842, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.313, + 0.894, + 0.369 + ], + "angle": 0, + "content": "Figure 2. Method overview. (a) The architecture of the LLM-based question-answer generator. Image features, projected by a linear layer concatenated with the predicted data table and a prompt, are fed into LLM for QA generation. (b) Straight-forward generation, where questions with answers are generated in a straightforward way. (c) Synthesizing step-by-step, which breaks the question down into rationale programs and derives the answer by executing the program." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.47, + 0.48 + ], + "angle": 0, + "content": "progressive manner. We first introduce an intuitive, straightforward method for data generation. Then, we introduce Synthesize Step-by-Step, which enhances the straightforward method with step-wise rationales. Finally, we introduce the settings for training the downstream chart VQA models using our generated QA data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.487, + 0.212, + 0.501 + ], + "angle": 0, + "content": "3.1. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.47, + 0.601 + ], + "angle": 0, + "content": "The architecture of our LLM-based data generator is shown in Fig. 2 (a). We leverage MPT-7B [55] as the LLM for data generation. The inputs \\( X \\) to LLM is a concatenation of three components: the projected image features \\( F \\), the underlying data table for the chart image \\( T \\), and a natural language prompt \\( P \\):" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.607, + 0.325, + 0.623 + ], + "angle": 0, + "content": "\\[\nX = \\left[ F; T; P \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.75 + ], + "angle": 0, + "content": "To extract the image features, we consider the pretrained visual transformer (ViT) [11] in CLIP [46]. The ViT takes in the image patches, and extracts features for each image patch. Then we use a trainable linear layer to project the extracted image features into the input embedding space of LLM. This architecture, i.e. linear layer for visual feature alignment, has been shown effective by recent multimodal large models like LLaVA [36]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.87 + ], + "angle": 0, + "content": "The predicted data table is critical to encode the textual information in the image. Because the pretrained ViT (in CLIP [46]) is not specifically trained for text-heavy images, the image features have limited OCR ability. Therefore, we use DEPLOT [34], which is designed for information extraction from charts, to predict the underlying data table for the given chart image. The linearized data table, which is a sequence of word tokens, is fed into the LLM." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Finally, the projected image features, the data table, and a prompt like \"Ask a question about the given chart plot and" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.389, + 0.892, + 0.42 + ], + "angle": 0, + "content": "answer it\" (rephrased in several different ways), are provided as the input to LLM, for question generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.43, + 0.794, + 0.446 + ], + "angle": 0, + "content": "3.2. Straight-forward data generation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.454, + 0.892, + 0.515 + ], + "angle": 0, + "content": "Given the above inputs, a straightforward manner for data generation is to have the LLM generate question and answer directly, as shown in Fig. 2 (b). The output \\( Y \\) is the generated question \\( Q \\), followed by the corresponding answer \\( Q \\):" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.527, + 0.737, + 0.544 + ], + "angle": 0, + "content": "\\[\nY = [ Q; A ]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.892, + 0.648 + ], + "angle": 0, + "content": "This straightforward data generator can be trained using the existing chart VQA datasets, which contains chart images paired with question and answers. The model is trained with a next-token prediction loss. Note that the ViT and LLM are pretrained and kept frozen during training. The only trainable parameters are the linear projection layer1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.892, + 0.737 + ], + "angle": 0, + "content": "As our experiments will show, this straightforward generation method yields reasonable results, generating questions and answers that are helpful for training the chart VQA models. However, the generated answers are noisy, which is improved using step-by-step generation, described next." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.75, + 0.715, + 0.766 + ], + "angle": 0, + "content": "3.3. Synthesize step-by-step" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.773, + 0.892, + 0.818 + ], + "angle": 0, + "content": "Here we introduce Synthesize Step-by-Step, which shares the same model architecture as straight-forward generation, but performs data synthesizing in a step-by-step manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.819, + 0.892, + 0.865 + ], + "angle": 0, + "content": "Instead of having the model generate answers directly, we train the LLM to generate them in a step-by-step manner. The LLM generates the question \\( Q \\) with the corresponding" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.876, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1We also tried a two-stage training procedure, i.e. finetune the whole model after training the projection layer, but did not see performance gain." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13615" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.164, + 0.082, + 0.81, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.277, + 0.895, + 0.307 + ], + "angle": 0, + "content": "Figure 3. Example of template-based question generation. Given a manually created template with defined program, by querying the image SVG annotations, templatic questions and answers can be generated, with rationales." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.471, + 0.357 + ], + "angle": 0, + "content": "rationales \\( R \\). Then the answer \\( A \\) is derived by executing the rationales using python scripts:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.37, + 0.377, + 0.389 + ], + "angle": 0, + "content": "\\[\nY = [ Q; R ], A = \\mathrm {E x e c u t e} (R)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.47, + 0.657 + ], + "angle": 0, + "content": "As the example shows in Fig. 2 (c), the LLM outputs the generated question and a decomposition of this complex reasoning question into a series of atom reasoning steps, i.e. rationales. The rationales \\( R \\) a step-by-step executable program, where each step in the program can be either an atom extractive VQA questions (e.g. ans_0 = VQA (\"What is the value of 2002?\")) or a Python-like program call for math calculation, counting, comparison, etc. (e.g. ans = avg(ans_0, ans_1)). The rationale program is then parsed and executed using Python in order to derive the final answer. During execution, there are two ways to answer the atom extractive VQA questions: we can either use an off-the-shelf VQA model, or directly predict the answer using the LLM data generator itself. We experiment with both options and find the latter option to be more effective, as well as more simple, without requiring additional VQA models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Training with rationales. To train the step-by-step generator, data with rationales is required. This is different from the straightforward generation, which only requires question-answer pairs to train. Because existing datasets do not contain the rationale annotations for each question, we introduce a template-based QA generation pipeline, which will be described in Sec. 3.4, to generate synthetic questions with rationales. With the template-generated data as the training corpus, we tune the generator for step-by-step generation. An additional advantage of training with template-generated data is that, in inference time, the generator can be prompted with different prompts for controllable generation. The prompts may specify the format or the type of the questions, e.g. \"the question should start with how many\", \"the question should ask about the colors in the image\", etc. A complete list of the prompts can be found in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.432 + ], + "angle": 0, + "content": "Post hoc filtering. We take an additional post hoc filtering step to filter out the noisy questions. We use the decoding score (for the whole generated sequence including question and rationales) as an indicator and filter out the questions with decoding scores lower than a threshold. Better filtering strategies, like using LLMs for data filtering, can be explored, which we will leave as future work." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.44, + 0.892, + 0.457 + ], + "angle": 0, + "content": "3.4. Tuning corpus: template-based QA generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Here we describe how we use a template-based pipeline to generate questions with rationales for training the LLM-based generator. The method is motivated by the success of synthetic compositional questions in VQA datasets like CLEVR [21], Super-CLEVR [33] and GQA [20]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.893, + 0.78 + ], + "angle": 0, + "content": "Fig. 3 illustrates the template-based question generation method. We first manually create question templates, then instantiate the templates into questions and answers by querying the rich SVG annotations of the chart image (as shown in the dotted boxes in the figure). For example, the template \"What is the average of the \\(\\langle N\\rangle ?\\) can be instantiated into \"what is the average of the more category?\". The argument \\(\\langle N\\rangle\\) is instantiated into \"the more category\" according to the SVG annotations, which includes the category names, values, and visual elements like colors and bounding boxes. Moreover, for each template, we define a reasoning program (in a domain-specific language) containing a series of reasoning operations (e.g. select_color, query_value) that specifies the reasoning process needed to answer the question. The program is then instantiated into a step-by-step rationale." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Note that there are limitations to this template-based QA generation pipeline. The pipeline relies on the rich annotations extracted automatically from the source SVG files, and thus cannot generalize to unlabeled images. Moreover, even though some existing datasets like ChartQA provide such rich annotations crawled from websources, they are inevitably noisy, containing missing or incorrect values. In contrast, the LLM generator is more generalizable, being" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13616" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.152 + ], + "angle": 0, + "content": "able to generate data for images in the wild. Therefore, in our work, these template-based questions are used as the training corpus of the LLM-based generator, which is more generalizable and scalable." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.161, + 0.336, + 0.178 + ], + "angle": 0, + "content": "3.5. Evaluation of generated data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.47, + 0.319 + ], + "angle": 0, + "content": "To measure the usefulness of the generated QA, we use the generated data to train the chart VQA models. The accuracy improvement of chart VQA models, after training with the generated QA, is taken as the indicator of the data quality. We select MATCHA [35], which is a transformer encoder-decoder model pretrained on large-scale text-heavy images, as our primary evaluation model, considering its modern architecture and state-of-the-art performance on standard datasets2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.471, + 0.472 + ], + "angle": 0, + "content": "Summary. Our data generation procedure contains the following steps: first, we generate templatic QA corpus with rationales using the rich SVG annotations, which is then used to tune the LLM-based data generator; next, the LLM-based generator generates questions with step-by-step rationales; then we use Python to execute the rationales and derive the answers; finally, this LLM-augmented data, LAMENDA, is used to train the downstream VQA model MATCHA. As the experiments will show, the augmented data leads to significant performance improvements." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.484, + 0.202, + 0.501 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.509, + 0.163, + 0.525 + ], + "angle": 0, + "content": "4.1. Set up" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.47, + 0.82 + ], + "angle": 0, + "content": "Dataset. Following [34, 35], we run experiments on two chart VQA datasets: ChartQA and PlotQA \\(^{3}\\). ChartQA [40] contains 21k real-world chart images crawled from four web sources, with 9k human-annotated (Human split) and 23k machine-augmented (Augmented split) QA pairs. The augmented questions are mostly extractive questions, while the human-annotated questions are free-form and require complex reasoning. PlotQA [41] is a large synthetic dataset containing 224k chart images generated from real-world online data, with 37M synthetic questions and answers divided into two splits: V1 (8M) and V2 (29M). In addition, to show the generalizability of our LLM data generator, we generate QA for images from additional chart captioning datasets, which do not provide QAs or rich SVG annotations. We use images from three datasets: Chart-to-Text [24], VisText [52] and ChartSumm [47]. Chart-to-Text contains 35k images from \"Statistica\" and 9k images from \"Pew\"; VisText contains 9k synthetic charts of various styles; ChartSumm contains 84k images from \"Statistica\"" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.827, + 0.47, + 0.864 + ], + "angle": 0, + "content": "2While there are two-stage models of better performance, which extracts information from charts then prompts LLMs like Codex to answer questions, MATCHA is the best single-stage model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.864, + 0.469, + 0.901 + ], + "angle": 0, + "content": "3There are other chart VQA datasets like FigureQA, DVQA, etc., but since they are out-dated and models reach \\(>95\\%\\) accuracy, we focus on more challenging settings." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.827, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "and \"Knoema\". Note that we only use the images from the chart captioning datasets without using the captions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.893, + 0.349 + ], + "angle": 0, + "content": "Generated data. Tab. 1 shows the statistics of the generated QA. To construct the template-based tuning corpus, we use 28 templates to generate \\(357\\mathrm{k}\\) question-answer pairs for the ChartQA training images. The template list and the template QA can be found in the Appendix. We train our LLM data generator using the template-based corpus, and generate QAs for images in ChartQA, PlotQA, and the three chart captioning datasets. For chartQA training images, we generate \\(403\\mathrm{k}\\) QAs, which reduces to \\(326\\mathrm{k}\\) after step-by-step execution (dropping ones that cannot be executed) and filtering based on the decoding scores using a threshold of -10. For PlotQA training images, we generate 3M QAs, which reduces to \\(1.7\\mathrm{M}\\) after execution and filtering. For the \\(137\\mathrm{k}\\) chart images in the chart captioning datasets, we generate \\(1.6\\mathrm{M}\\) QAs." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.354, + 0.892, + 0.436 + ], + "angle": 0, + "content": "
DatasetImagesGeneratedFiltered
TemplateChartQA18,317356,606-
LAMENDAChartQA18,317402,974326,160
PlotQA157,0703,455,5391,726,822
ChartCap137,2812,987,7181,635,364
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.44, + 0.88, + 0.455 + ], + "angle": 0, + "content": "Table 1. Statistics of generated data, before and after filtering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.789 + ], + "angle": 0, + "content": "Implementation details. Using the 357k template QA, we tune the data generator for 5 epochs with a batch size of 128. We use a cosine learning rate schedule with a max learning rate of 2e-3. The training takes about 12 hours using 8 A100 GPUs. The trainable projection layer is initialized with the first stage checkpoint of LLaVA [36]. We select the best checkpoint at 14k iterations as our best generator. For MATCHA training, we take the pretrained checkpoints (base size) from HuggingFace and finetune the models with a combination of our generated questions and the original ChartQA or PlotQA training questions. The training setting follows [28, 35]. We use a cosine scheduler with 500 warm-up steps and a max learning rate of 0.0001. The batch size is 256. Without special description, the models are trained for 10k iterations for ChartQA (combining human and augmented splits) and 20k iterations for PlotQA because it is much larger (V1 and V2 are trained separately). We set the max number of image patches as 4096. The ChartQA training takes 3 days on 8 32GB V100 GPUs. Considering the training time, in some experiments, we reduce the image patch number to 1024 for quick experiments in 12 hours with the same setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.892, + 0.88 + ], + "angle": 0, + "content": "Evaluation metrics. We suggest that both strict accuracy and relaxed accuracy should be reported, especially for the ChartQA dataset. While existing works [34, 35, 40, 41] evaluate answers are evaluated using relaxed accuracy, which allows \\(5\\%\\) error for numerical answers, we find that this relaxed accuracy metric is problematic for ChartQA." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.851, + 0.9 + ], + "angle": 0, + "content": "4https://huggingface.co/google/matcha-base" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13617" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.143, + 0.083, + 0.835, + 0.24 + ], + "angle": 0, + "content": "
AccuracyRelaxed Accuracy
avghumanaugmentavghumanaugment
0baseline (MATCHA [35])47.7630.2165.3158.5439.5877.50
1syna48.8028.6568.9659.4338.6580.21
2synat w/ VQA answer47.7628.4467.0857.9737.8178.12
3synat w/ step-VQA answer52.5537.2967.8164.3848.8579.90
4synat w/ synat answer53.7537.9269.5865.1048.8581.35
5synat w/ step-synat answer55.2140.2170.2166.4151.3581.46
6synat w/ step-synat answer + tpl.57.3442.6072.0867.8653.1282.60
7synat w/ step-synat answer + tpl. + addi.58.6545.6271.6769.8456.5683.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.244, + 0.895, + 0.3 + ], + "angle": 0, + "content": "Table 2. Main results on ChartQA val split. The data generator trained on Annotated questions is denoted as \\( \\mathrm{syn}_a \\); the data generator trained on both Annotated questions and our Template-generated questions \\( \\mathrm{syn}_{at} \\) is denoted as \\( \\mathrm{syn}_{at} \\). Step-by-step generation is better than straight-forward generation (row 1-5). Training with a combination of step-by-step generated QA, template-generated QA (+mpl.), and additional images from chart captioning datasets (+addi.), leads to the best model. See Sec. 4.2 for more details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.471, + 0.36 + ], + "angle": 0, + "content": "For example, many questions in ChartQA ask about years (e.g. 1996, 2012), for which allowing \\(5\\%\\) numerical errors (i.e. around 100 years) will result in incorrectly high results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.373, + 0.212, + 0.387 + ], + "angle": 0, + "content": "4.2. Main results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.47, + 0.489 + ], + "angle": 0, + "content": "We study our data generation method with different settings on the ChartQA dataset, with MATCHA as our baseline. We first generate QA data using our method, and then use the generated data to train the MATCHA model. The strict accuracy and relaxed accuracy reported on the ChartQA validation split are shown in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.47, + 0.656 + ], + "angle": 0, + "content": "The baseline model MATCHA (row-0) achieves \\(47.76\\%\\) accuracy and \\(58.54\\%\\) relaxed accuracy. For this baseline, we finetune the pretrained MATCHA checkpoint for 10k iterations with the ChartQA training split (combing human and augmented questions), using 1024 image patches. This setting follows [35], except that a smaller number of image patches are used for faster experiments and less GPU memory requirement. We validate the correctness of our implementation by showing that the model reaches \\(63.7\\%\\) relaxed accuracy using the original resolution, which matches the numbers reported in [35] (as shown in Tab. 3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.764 + ], + "angle": 0, + "content": "Row-1 shows the results for the straight-forward data generation, where we train the data generator with QAs from the ChartQA training split and use the trained generator to directly generate questions and answers. This straightforward data generation improves the accuracy by \\(1\\%\\) (from \\(47.76\\%\\) to \\(48.80\\%\\)), suggesting that LLMs can be helpful in data augmentation for chart reasoning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Row-2 to row-5 shows the advantage of synthesizing step-by-step over straight-forward generation. We train the data generator with our template-generated QA, but the answers are generated with different methods. In row-2, we feed in the generated complex question into an additional VQA model (i.e. baseline MATCHA) to get the answers; in row-3, we generate the step-by-step rationales, and the stepwise questions are fed into the VQA model to get step-wise answers, then the final answer is derived using the step" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.895, + 0.495 + ], + "angle": 0, + "content": "wise answers. As can be seen, the step-wise answer generation (52.55%) is significantly better than the direct answer generation (47.76%), which suggests the importance of decomposing the complex questions into atom sub-questions. Row-4 and row-5 follow a similar setting, except that we use the LLM data generator to generate the answers instead of the additional VQA model. Two findings can be drawn: (a) LLM-generated answers are better than VQA-predicted answers, possibly because the LLM data generator has access to the predicted data tables as inputs; (b) step-by-step synthesizing outperforms direct generation, for both VQA-predicted and LLM-generated answers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.895, + 0.725 + ], + "angle": 0, + "content": "In row-6, we combine the step-by-step generated data with the templated-generated QAs and show that training with both leads to further better performance (57.34%). In row-7, we generate QAs for the chart images from additional chart captioning datasets. The final model, trained with a combination of all the synthetic data, gets 58.65% accuracy. Compared with the baseline, training with the augmented data leads to a 10.89% accuracy improvement. Moreover, the improvement is more significant on the human-written questions (+15.41%), which requires complex reasoning, than the augmented extractive questions (+6.36%). The results suggest that our synthetic data improves both the extractive ability and the reasoning ability, with the latter being more significant. We show examples in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.719, + 0.756 + ], + "angle": 0, + "content": "4.3. Comparison with SoTA" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To compare our method with the best performing models, we apply our generated data onto two models, Pix2Struct [28] and MATCHA [35], and report the results on the ChartQA test split in Tab. 3. For faster training, we first train the model (for both Pix2struct and MATCHA) with a max of 1024 image patches using our best setting (i.e. row-7 in Tab. 2), then finetune with 4096 image patches to get the final result. We report both the 1024-resolution results and the 4096-resolution results." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13618" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.151, + 0.077, + 0.825, + 0.288 + ], + "angle": 0, + "content": "
paramsaccuracyRelaxed Accuracy
avghumanaugmentavghumanaugment
VL-T5-OCR [40]----41.56--
VisionTapas-OCR [40]----45.52--
Pix2Struct4096 [28]282M---56.030.581.6
MATCHA4096 [35]282M---64.238.290.2
DEPLOT+FlanPaLM+Codex [34]540B---79.367.691.0
PaLI-X [7]55B---72.3--
Pix2Struct4096 (reimpl.)282M50.4424.6476.2458.2432.0084.48
Pix2Struct1024 + LAMENDA282M51.7734.6968.8563.0246.2579.79
Pix2Struct4096 + LAMENDA282M55.8340.1071.5666.8251.6781.98
MATCHA4096 (reimpl.)282M55.6831.0480.3263.7237.7689.68
MATCHA1024 + LAMENDA282M60.2841.9278.6469.8851.7688.00
MATCHA4096 + LAMENDA282M63.3644.8881.8472.6453.9291.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.292, + 0.893, + 0.322 + ], + "angle": 0, + "content": "Table 3. Comparison with SoTAs on ChartQA test split. Subscripts (1024, 4096) means image token numbers (resolution). With our generated data, both Pix2Struct and MATCHA achieve significantly better performance, even comparable to much larger models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.471, + 0.501 + ], + "angle": 0, + "content": "As can be seen, our generated data improves the overall relaxed accuracy of Pix2Struct by \\(8.58\\%\\) and MATCHA by \\(8.92\\%\\). Moreover, the improvement is much more significant on human-written questions, e.g. the relaxed accuracy of MATCHA on human-written questions improves from \\(38\\%\\) to \\(54\\%\\). Interestingly, with our augmented data, the low-resolution (1024) models can already outperform their high-resolution (4096) counterparts without data augmentation. The significant performance gain indicates that learning with synthetic data is extremely helpful for better reasoning ability with human-written queries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.47, + 0.637 + ], + "angle": 0, + "content": "Our best performing MATCHA achieves \\(72.6\\%\\) relaxed accuracy, which beats all existing end-to-end chart reasoning models. Note that DEPLOT+FlanPaLM+Codex is a two-stage model that first extracts information from the charts, then prompts external LLMs like FlanPalm (540B) Codex (12B) for question answering. The method relies on much larger models of billions of parameters and thus is not comparable with our method, which is much smaller and not rely on external LLMs for inference." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.228, + 0.663 + ], + "angle": 0, + "content": "4.4. Generalization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.837 + ], + "angle": 0, + "content": "To show the generalization ability of the proposed data generation method, we run experiments on the PlotQA dataset. We directly use our data generator (trained on template questions on ChartQA) to generate QA for the PlotQA training images. The results are shown in Tab. 4. As can be seen, our method, when applied to MATCHA, achieves state-of-the-art performance (92.89% relaxed accuracy) on PlotQA. Despite the synthetic nature of this dataset and the saturated model performance, our augmented data can still help on both low-resolution (+3.46%) and high-resolution (+0.92%) settings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.287, + 0.863 + ], + "angle": 0, + "content": "4.5. Ablations and analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Qualitative analysis and failure cases. Fig. 5 shows examples of our LLM-generated QA, including good and bad ex" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.893, + 0.714 + ], + "angle": 0, + "content": "samples. Most generated questions are specific to the given image, asking about the categories plotted, such as “2019” or “Tablet 2018”, indicating that the model is able to utilize the information of the input. Moreover, various questions are generated, including counting, colors, average, etc. In the failure example, the question requires comparing the “silver bar” with the “midnightblue” bar, with reasonable rationales. However, there are multiple bars in the specified colors in the image, resulting in ambiguous/invalid questions. When looking at more failure cases, we find that there are several typical failure modes. For example, questions are sometimes invalid/ambiguous (e.g. asking about the top/bottom bar for a vertical bar plot); questions can be hallucinated or irrelevant (e.g. asking about a category that is not shown in the figure); rationales and answers are not correct, etc., which opens up opportunities for future works. Before and after LAMENDA. To gain insights on how LAMENDA helps model training, we conduct a comparison analysis of MATCHA “before” and “after” training with our LLM-generated data, i.e. the baseline model versus our best model. According to the predictions of these two models, we divide the 960 human-written questions from ChartQA val split into 4 groups: both model answers correctly, both models answers wrongly, the before model answers correctly, but the after model answers wrongly, and" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.72, + 0.884, + 0.883 + ], + "angle": 0, + "content": "
Relaxed Accuracy
avgV1V2
VisionTapas-OCR [40]53.9065.3042.50
VL-T5-OCR [40]65.9675.9056.02
DEPLOT+FlanPaLM+Codex [34]66.662.271.0
Pix2Struct [28]72.573.271.9
MATCHA4096 [35]91.592.390.7
MATCHA1024 (reimpl.)74.9573.8876.02
MATCHA1024 + LAMENDA78.4178.4478.38
MATCHA4096 (reimpl.)91.9792.6491.30
MATCHA4096 + LAMENDA92.8993.9491.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.888, + 0.892, + 0.916 + ], + "angle": 0, + "content": "Table 4. Comparison with SoTAs on PlotQA test split. With our generated data, MATCHA achieves the SoTA performance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13619" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.168, + 0.086, + 0.805, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.229, + 0.895, + 0.273 + ], + "angle": 0, + "content": "Figure 4. Analysis of model predictions before and after training with LAMENDA. In I (left), ChartQA val questions are divided into four categories according to the model prediction correctness (before/after). II (right) further decomposes each category into different question types. LAMENDA improves the model prediction, and the improvement is most significant on reasoning questions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.286, + 0.473, + 0.543 + ], + "angle": 0, + "content": "vice versa. The statistics of these four groups are shown in Fig. 4. From Fig. 4 I. (left side), we see that LAMENDA is helpful for \\(22\\%\\) of the questions, while there are still \\(45\\%\\) of the questions where both models cannot answer correctly. In Fig. 4 II. (right side), we randomly sample 30 questions from each of the four groups and manually classify them into six question types: extractive, count, spatial, color, reasoning, and others. As shown in (a), i.e. the questions where the \"before\" model answers wrongly and our model answers correctly, most \\((70\\%)\\) of the questions belong to the reasoning category, indicating that our data helps most on the reasoning questions. Our data also helps with color questions \\((10\\%)\\) and count questions \\((17\\%)\\). However, our data does not help with extractive questions \\((0\\%)\\). In addition, comparing the distribution of (c) both correct with (d) both wrong, we can see that extractive questions are easy for the models, while reasoning questions are very challenging." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.544, + 0.471, + 0.605 + ], + "angle": 0, + "content": "In summary, Fig. 4 shows the effectiveness of our data for the reasoning questions over other question types. However, reasoning questions are still a big challenge for current models, which requires further exploration." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.607, + 0.47, + 0.623 + ], + "angle": 0, + "content": "Different question types. In Tab. 5, we study the improve" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.639, + 0.407, + 0.874 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.875, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Figure 5. Examples of generated QA, 3 good examples and 1 failure case. Rationales are only shown for 2 examples for brevity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.286, + 0.895, + 0.483 + ], + "angle": 0, + "content": "ments brought by each type of the generated questions. The total 326k generated QAs are grouped into 7 different groups, including questions about color understanding, spatial reasoning, counting, minimum/maximum reasoning, average, comparison (e.g. larger than, smaller than), and other math calculations. This grouping is done by different prompting of the data generation LLM, as specified in the Appendix. The improvements from different question types are distinct and can be accumulated. As we see, each of the seven types improves the model performance by less than \\(3\\%\\), while combining all of them leads to \\(7.03\\%\\) improvement. Moreover, among the seven types, the \"math calculation\" questions brings the most improvements \\((2.76\\%)\\)." + }, + { + "type": "table", + "bbox": [ + 0.562, + 0.49, + 0.836, + 0.617 + ], + "angle": 0, + "content": "
#QAavghumanaug
baseline28,29947.7630.2165.31
+colors+46,51247.8129.9065.73
+spatial+49,38749.2730.6367.92
+count+36,70547.4029.1765.63
+minmax+71,78849.5831.8767.29
+average+50,71748.9631.3566.56
+compare+14,69047.6628.8566.46
+calculation+56,36150.5233.8567.19
+all326,16054.7939.2770.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.622, + 0.894, + 0.665 + ], + "angle": 0, + "content": "Table 5. A detailed look at the effect for different question types. Strict accuracy on ChartQA val split is shown. Each type helps and combining all of them leads to a further performance gain." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.62, + 0.694 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We propose Synthesize Step-by-Step, which utilizes LLMs to augment the training data for the chart reasoning task. We show that the step-by-step generation, which decomposes the complex reasoning questions into easier sub-questions, is critical for synthesizing good-quality data. We run experiments on two standard benchmarks and show significant improvements by training with the augmented data. Specifically, the relaxed accuracy on the human-written questions is improved from \\(37.8\\%\\) to \\(53.9\\%\\) on ChartQA. The large improvement indicates the promise of boosting the model's reasoning ability by data generation using LLMs, from which the strong step-by-step reasoning ability can be distilled into downstream models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13620" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Mubashara Akhtar, Oana Cocarascu, and Elena Simperl. Reading and reasoning over chart images for evidence-based automated fact-checking. arXiv preprint arXiv:2301.11843, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.17, + 0.472, + 0.25 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.251, + 0.471, + 0.306 + ], + "angle": 0, + "content": "[3] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 39-48, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.306, + 0.471, + 0.359 + ], + "angle": 0, + "content": "[4] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R Manmatha. Docformerv2: Local features for document understanding. arXiv preprint arXiv:2306.01733, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.471, + 0.428 + ], + "angle": 0, + "content": "[5] Ritwick Chaudhry, Sumit Shekhar, Utkarsh Gupta, Pranav Maneriker, Prann Bansal, and Ajay Joshi. Leaf-qa: Locate, encode & attend for figure question answering. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3512–3521, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.429, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.483, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[7] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.471, + 0.618 + ], + "angle": 0, + "content": "[8] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.618, + 0.471, + 0.672 + ], + "angle": 0, + "content": "[9] John Joon Young Chung, Ece Kamar, and Saleema Amershi. Increasing diversity while maintaining accuracy: Text data generation with large language models and human interventions. arXiv preprint arXiv:2306.04140, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.672, + 0.471, + 0.738 + ], + "angle": 0, + "content": "[10] Brian Davis, Bryan Morse, Brian Price, Chris Tensmeyer, Curtis Wigington, and Vlad Morariu. End-to-end document recognition and understanding with dessurt. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.739, + 0.471, + 0.821 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.821, + 0.471, + 0.891 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.621, + 0.106 + ], + "angle": 0, + "content": "scale, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.106, + 0.892, + 0.173 + ], + "angle": 0, + "content": "[13] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.173, + 0.892, + 0.214 + ], + "angle": 0, + "content": "[14] Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. Knowledge distillation of large language models. arXiv preprint arXiv:2306.08543, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.214, + 0.892, + 0.255 + ], + "angle": 0, + "content": "[15] Nitish Gupta, Kevin Lin, Dan Roth, Sameer Singh, and Matt Gardner. Neural module networks for reasoning over text. arXiv preprint arXiv:1912.04971, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.255, + 0.892, + 0.309 + ], + "angle": 0, + "content": "[16] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.309, + 0.892, + 0.349 + ], + "angle": 0, + "content": "[17] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[18] Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes. arXiv preprint arXiv:2305.02301, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.43, + 0.892, + 0.485 + ], + "angle": 0, + "content": "[19] Ronghang Hu, Jacob Andreas, Trevor Darrell, and Kate Saenko. Explainable neural computation via stack neural module networks. In Proceedings of the European conference on computer vision (ECCV), pages 53-69, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.485, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[20] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[21] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.634, + 0.892, + 0.675 + ], + "angle": 0, + "content": "[22] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[23] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Akos Kadar, Adam Trischler, and Yoshua Bengio. Figureqa: An annotated figure dataset for visual reasoning, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.729, + 0.892, + 0.811 + ], + "angle": 0, + "content": "[24] Shankar Kantharaj, Rixie Tiffany Leong, Xiang Lin, Ahmed Masry, Megh Thakkar, Enamul Hoque, and Shafiq Joty. Chart-to-text: A large-scale benchmark for chart summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4005–4023, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.811, + 0.892, + 0.865 + ], + "angle": 0, + "content": "[25] Tushar Khot, Harsh Trivedi, Matthew Finlayson, Yao Fu, Kyle Richardson, Peter Clark, and Ashish Sabharwal. Decomposed prompting: A modular approach for solving complex tasks. arXiv preprint arXiv:2210.02406, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.865, + 0.892, + 0.893 + ], + "angle": 0, + "content": "[26] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13621" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "doo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.147, + 0.47, + 0.202 + ], + "angle": 0, + "content": "[27] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.203, + 0.469, + 0.295 + ], + "angle": 0, + "content": "[28] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In International Conference on Machine Learning, pages 18893–18912. PMLR, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.297, + 0.469, + 0.35 + ], + "angle": 0, + "content": "[29] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.351, + 0.469, + 0.405 + ], + "angle": 0, + "content": "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[31] Liunian Harold Li, Jack Hessel, Youngjae Yu, Xiang Ren, Kai-Wei Chang, and Yejin Choi. Symbolic chain-of-thought distillation: Small models can also\" think\" step-by-step. arXiv preprint arXiv:2306.14050, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.46, + 0.469, + 0.541 + ], + "angle": 0, + "content": "[32] Zhuowan Li, Elias Stengel-Eskin, Yixiao Zhang, Cihang Xie, Quan Hung Tran, Benjamin Van Durme, and Alan Yuille. Calibrating concepts and operations: Towards symbolic reasoning on real images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14910-14919, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.542, + 0.47, + 0.623 + ], + "angle": 0, + "content": "[33] Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Kortylewski, Wufei Ma, Benjamin Van Durme, and Alan L Yuille. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14963-14973, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.624, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[34] Fangyu Liu, Julian Martin Eisenschlos, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Wenhu Chen, Nigel Collier, and Yasemin Altun. Deplot: One-shot visual language reasoning by plot-to-table translation. arXiv preprint arXiv:2212.10505, 2022. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[35] Fangyu Liu, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Yasemin Altun, Nigel Collier, and Julian Martin Eisenschlos. Matcha: Enhancing visual language pretraining with math reasoning and chart derendering. arXiv preprint arXiv:2212.09662, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.814, + 0.469, + 0.88 + ], + "angle": 0, + "content": "[37] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.881, + 0.469, + 0.894 + ], + "angle": 0, + "content": "[38] Junyu Luo, Zekun Li, Jinpeng Wang, and Chin-Yew Lin." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.894 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Chartocr: Data extraction from charts images via a deep hybrid framework. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1917-1925, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.148, + 0.892, + 0.202 + ], + "angle": 0, + "content": "[39] Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B Tenenbaum, and Jiajun Wu. The neuro-symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision. arXiv preprint arXiv:1904.12584, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.203, + 0.892, + 0.256 + ], + "angle": 0, + "content": "[40] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.256, + 0.892, + 0.31 + ], + "angle": 0, + "content": "[41] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.311, + 0.892, + 0.349 + ], + "angle": 0, + "content": "[42] Aaron Parisi, Yao Zhao, and Noah Fiedel. Talm: Tool augmented language models. arXiv preprint arXiv:2205.12255, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.35, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[43] Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.485 + ], + "angle": 0, + "content": "[44] Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). Association for Computational Linguistics, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.486, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[45] Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, et al. Toolllm: Facilitating large language models to master 16000+ real-world apis. arXiv preprint arXiv:2307.16789, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.554, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.636, + 0.892, + 0.703 + ], + "angle": 0, + "content": "[47] Raian Rahman, Rizvi Hasan, Abdullah Al Farhad, Md Tahmid Rahman Laskar, Md Hamjajul Ashmafee, and Abu Raihan Mostofa Kamal. Chartsumm: A comprehensive benchmark for automatic chart summarization of long and short summaries. arXiv preprint arXiv:2304.13620, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[48] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. arXiv preprint arXiv:2302.04761, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.771, + 0.892, + 0.826 + ], + "angle": 0, + "content": "[49] Timo Schick and Hinrich Schütze. Generating datasets with pretrained language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6943-6951, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.826, + 0.892, + 0.88 + ], + "angle": 0, + "content": "[50] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.88, + 0.892, + 0.893 + ], + "angle": 0, + "content": "[51] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt:" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13622" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.12, + 0.468, + 0.16 + ], + "angle": 0, + "content": "[52] Benny J Tang, Angie Boggust, and Arvind Satyanarayan. Vistext: A benchmark for semantically rich chart captioning. arXiv preprint arXiv:2307.05356, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.16, + 0.469, + 0.2 + ], + "angle": 0, + "content": "[53] Peng Tang, Srikar Appalaraju, R Manmatha, Yusheng Xie, and Vijay Mahadevan. Multiple-question multiple-answer text-vqa. arXiv preprint arXiv:2311.08622, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.2, + 0.469, + 0.254 + ], + "angle": 0, + "content": "[54] Peng Tang, Pengkai Zhu, Tian Li, Srikar Appalaraju, Vijay Mahadevan, and R Manmatha. Deed: Dynamic early exit on decoder for accelerating encoder-decoder transformer models. arXiv preprint arXiv:2311.08623, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.254, + 0.469, + 0.281 + ], + "angle": 0, + "content": "[55] MN Team et al. Introducing mpt-7b: A new standard for open-source, commercially usable llms, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.281, + 0.469, + 0.334 + ], + "angle": 0, + "content": "[56] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.376 + ], + "angle": 0, + "content": "[57] Shuohang Wang, Yang Liu, Yichong Xu, Chenguang Zhu, and Michael Zeng. Want to reduce labeling cost? gpt-3 can help. arXiv preprint arXiv:2108.13487, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.469, + 0.442 + ], + "angle": 0, + "content": "[58] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[59] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[60] Peter West, Chandra Bhagavatula, Jack Hessel, Jena D Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, and Yejin Choi. Symbolic knowledge distillation: from general language models to commonsense models. arXiv preprint arXiv:2110.07178, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.579, + 0.469, + 0.648 + ], + "angle": 0, + "content": "[61] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1192-1200, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[62] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.469, + 0.769 + ], + "angle": 0, + "content": "[63] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.769, + 0.469, + 0.837 + ], + "angle": 0, + "content": "[64] Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Josh Tenenbaum. Neural-symbolic vqa: Disentangling reasoning from vision and language understanding. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.837, + 0.469, + 0.892 + ], + "angle": 0, + "content": "[65] Yue Yu, Yuchen Zhuang, Jieyu Zhang, Yu Meng, Alexander Ratner, Ranjay Krishna, Jiaming Shen, and Chao Zhang. Large language model as attributed training data generator: A tale of diversity and bias. In Thirty-seventh Conference on" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.12, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[66] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.188, + 0.892, + 0.242 + ], + "angle": 0, + "content": "[67] Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.242, + 0.892, + 0.282 + ], + "angle": 0, + "content": "[68] Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. Automatic chain of thought prompting in large language models. arXiv preprint arXiv:2210.03493, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.282, + 0.892, + 0.335 + ], + "angle": 0, + "content": "[69] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.391 + ], + "angle": 0, + "content": "[70] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13623" + } + ] +] \ No newline at end of file diff --git a/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/58e8331e-f572-44c9-b724-02ca5cd604a2_origin.pdf b/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/58e8331e-f572-44c9-b724-02ca5cd604a2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..107637411687238f8abc8a89fbcbfe7b15908bfb --- /dev/null +++ b/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/58e8331e-f572-44c9-b724-02ca5cd604a2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4a31b077ee2d793c007c6d8dfea662479a89512e40051e27277f0e78649a929 +size 1546733 diff --git a/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/full.md b/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bf0ea0a4fe6534336005424f2a1acba017a15315 --- /dev/null +++ b/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/full.md @@ -0,0 +1,303 @@ +# Synthesize Step-by-Step: Tools, Templates and LLMs as Data Generators for Reasoning-Based Chart VQA + +Zhuowan Li $^{1}$ * Bhavan Jasani $^{2}$ * Peng Tang $^{2}$ Shabnam Ghadar $^{2}$ + +1 Johns Hopkins University 2 AWS AI Labs + +zili110@jhu.edu {bjasani, tangpen, shabnam}@amazon.com + +# Abstract + +Understanding data visualizations like charts and plots requires reasoning about both visual elements and numerics. Although strong in extractive questions, current chart visual question answering (chart VQA) models suffer on complex reasoning questions. In this work, we address the lack of reasoning ability by data augmentation. We leverage Large Language Models (LLMs), which have shown to have strong reasoning ability, as an automatic data annotator that generates question-answer annotations for chart images. The key innovation in our method lies in the Synthesize Step-by-Step strategy: our LLM-based data generator learns to decompose the complex question into step-by-step sub-questions (rationales), which are then used to derive the final answer using external tools, i.e. Python. This step-wise generation procedure is trained on synthetic data generated using a template-based QA generation pipeline. Experimental results highlight the significance of the proposed step-by-step generation. By training with the LLM-augmented data (LAMENDA), we significantly enhance the chart VQA models, achieving the state-of-the-art accuracy on the ChartQA and PlotQA datasets. In particular, our approach improves the accuracy of the previous state-of-the-art approach from $38\%$ to $54\%$ on the human-written questions in the ChartQA dataset, which needs strong reasoning. We hope our work underscores the potential of synthetic data and encourages further exploration of data augmentation using LLMs for reasoning-heavy tasks. + +# 1. Introduction + +Data visualizations like charts and plots play an important role in real-world data analysis applications. Unlike natural images, chart are text-heavy and data-driven, thus requiring better visual perception (e.g. OCR) and cognitive reasoning (e.g. math calculation, matching the legends with bars). For + +![](images/d6864ec69a86d47cd65f2a00a46dc0c667af30eb534d8990db4d35a3bc1694eb.jpg) +Figure 1. Existing chart VQA models struggle with complex reasoning questions. We attribute this to limited reasoning questions in existing datasets and address it by data augmentation. We fine-tune an LLM-based data generator that automatically generates question-answer annotations given a chart image. Our key innovation is Synthesize Step-by-Step, which breaks complex questions down into easy steps that could be solved using external tools. We use templates to train the LLM. Training with LLM-augmented data, LAMENDA, greatly enhances the chart VQA models. + +example, to answer "what is the total of Democrats and Republicans in 2010" in Fig. 1, the model needs to recognize the texts, match the legends of "democrats" and "republicans" to the bars, then calculate the sum. This is a challenging task requiring multimodal perception and reasoning. + +Compounding the complexity, the inherent difficulty of chart reasoning also amplifies the cost of collecting human annotations. Most existing datasets are synthetically generated using simple rules and heuristics, posing a risk of model overfitting. Among the datasets, only ChartQA [40] contains complex human-annotated question-answer pairs, although there are very few of those per image. For example, for the image in Fig. 1, while a lot more questions can be asked, only two questions are annotated (What is the total of Republicans and Democrats in 2010? Which color represents Republican?). Importantly, these human-written questions pose great challenges for contemporary models. For example, the state-of-the-art model [35] achieves only $38\%$ accuracy when confronted with human-written ques + +tions that require multi-step complex reasoning. Approximately half of the errors, as reported in [35], are due to reasoning failures. + +In this work, we propose to leverage large language models (LLMs) as automatic data annotators for chart reasoning. Recently, LLMs have demonstrated strong reasoning ability for complex tasks [27, 58, 59, 63, 68, 69]. Our motivation is to take advantage of this strong reasoning ability and generate question-answer pairs using LLMs. This LLM-generated data can be used to train the chart VQA models, thus improving their reasoning ability. + +The key innovation in our method lies in the Synthesize Step-by-Step strategy, which leverages LLMs to generate data in a step-by-step manner. Similar to human annotators who are trained to break down complex tasks into easier subtasks, our LLM-based data generator generates step-by-step. Instead of generating question-answer all at once, the generator learns to generate the question first, then decompose the complex question into step-by-step sub-questions (rationales) and answer them one by one. Finally, the answers of the sub-questions are combined to derive the final answer using external tools like calculators or Python code. Specifically, we feed the image features extracted with a ViT [12] backbone into an LLM using a simple trainable projection layer. To supervise the training of this projection layer, we construct a template-based training corpus generated using a template-based QA generation pipeline. Importantly, as we will show in the paper, the step-by-step synthesizing, trained using these template-based rationales, is critical for generating good-quality data. + +Experiments showcase the effectiveness of Synthesize Step-by-Step. Our LLM-augmented Data, LaMenDa, is used to train the downstream chart VQA models and its effectiveness is shown on two challenging datasets, ChartQA and PlotQA with the state-of-the-art accuracy. Notably, we significantly improve the performance from $38\%$ to $54\%$ on the challenging human-written questions in ChartQA dataset, which requires complex reasoning. We also demonstrate the advantage of step-wise generation over straightforward generation. With the results, we hope our work underscores the potential of synthetic data and encourages further exploration of data augmentation using LLMs for reasoning-heavy tasks. + +# 2. Related Work + +Multimodal models for text-heavy images fall into two categories: OCR-dependent or OCR-free. Models like LayoutLM [61], PaLI(-X) [7, 8], ChartBERT [1], DocFormerv2 [4], MQMA [53], DEED [54] rely on external OCR systems like ChartOCR [38] as either model input or training objectives. Recent top performing models, like Donut [26], Dessurt [10], Pix2struct [28], shift away from OCR thanks to vision transformers [11]. Pix2Struct [28] is a trans + +former encoder-decoder model for multiple VQA tasks, pretrained with an HTML-masked prediction objective, using a large amount of web page screenshots. MATCHA [35] and DEPLOT [34] are two recent follow-up works, tuning Pix2struct with chart-specific tasks like chart de-rendering and math reasoning, which are used in our work. + +Datasets for chart VQA. Earlier datasets like DVQA [22], FigureQA [23], LeafQA [5] are based on synthetically generated images and templatic questions, which are easy to solve and prior models achieve $>95\%$ accuracy. PlotQA [41] is also a synthetic dataset but contains open-vocabulary questions. ChartQA [40] is a recent dataset containing real images with human-written questions, which is much more challenging. Our work is mainly based on ChartQA and PlotQA, following the settings in [34, 35]. + +LLMs for multimodal tasks. There is a surge of interest in extending LLMs into the vision-and-language domain. Recent works align the image features into the input space of the pretrained LLMs using a lightweight module, e.g. Frozen [56], Flamingo [2], LLaVA [36], Otter [29], MiniGPT-4 [70], BLIP-2 [30], AdaptorV2 [13], etc. For example, LLaVA [36] uses a simple projection layer to align the modalities and build a visually-finetuned multimodal model; LLaVAR [67] extends it onto text-heavy images. + +Reasoning step-by-step has been studied for a long time in the multimodal field. Compositional models like neural modular networks [3, 15, 19] or neural symbolic reasoning [32, 33, 39, 64] decompose complex visual questions into reasoning steps, and compose modules to solve the steps. For LLMs, chain-of-thought prompting is a simple yet effective method to boost the reasoning performance [27, 58, 59, 63, 68, 69]. Recent works like VisProg [16] and ViperGPT [51] utilize LLMs as the planner to compose domain-specific models [37, 43, 62, 66] and external tools or APIs [25, 42, 45, 48, 50] for solving complex tasks. + +LLMs for data generation has been recently studied for language tasks [9, 49, 65] or multimodal tasks [36, 70]. This relates to knowledge distillation [17], which teaches a smaller model by learning from a larger teacher model, benefiting semi-supervised learning [6, 57]. Distillation from LLMs [14, 18, 31, 44, 60] can be completed by either learning from the logits or from the generated data directly. Our work is related but different from these works in that we focus on step-by-step generation, combining LLMs with templates and tools, for the chart reasoning task. + +# 3. Method + +Given a chart plot, our goal is to build a data generator that generates complex reasoning questions with answers. The generated data can be used for training the downstream chart VQA models. + +The overview of our LLM-based data generator is shown in Fig. 2. In this section, we describe the data generator in a + +(b) straight-forward generation +(c) synthesize step-by-step +LLM Output : +Figure 2. Method overview. (a) The architecture of the LLM-based question-answer generator. Image features, projected by a linear layer concatenated with the predicted data table and a prompt, are fed into LLM for QA generation. (b) Straight-forward generation, where questions with answers are generated in a straightforward way. (c) Synthesizing step-by-step, which breaks the question down into rationale programs and derives the answer by executing the program. +![](images/3f5454cc676b243f2de2603e5cdfdc1b54d499dfe70a8af81e7d9896c8561793.jpg) +Prompt : Ask a question about the chart plot and answer it +LLM Output : What is the average of the two largest categories?13% +Prompt: Ask a question about the chart plot and think step-by-step to answer it. +Question: What is the average of the two largest categories? +Rationale: ans_0 = VQA("What is the value of 2002?") # 16.2% +ans_1 = VQA("What is the value of 2003?") # 16.7% +ans $=$ avg(ans_0, ans_1) +Code execution: ans_0 = 16.2%, ans_1 = 16.7%, ans = 16.5% + +progressive manner. We first introduce an intuitive, straightforward method for data generation. Then, we introduce Synthesize Step-by-Step, which enhances the straightforward method with step-wise rationales. Finally, we introduce the settings for training the downstream chart VQA models using our generated QA data. + +# 3.1. Architecture + +The architecture of our LLM-based data generator is shown in Fig. 2 (a). We leverage MPT-7B [55] as the LLM for data generation. The inputs $X$ to LLM is a concatenation of three components: the projected image features $F$ , the underlying data table for the chart image $T$ , and a natural language prompt $P$ : + +$$ +X = \left[ F; T; P \right] +$$ + +To extract the image features, we consider the pretrained visual transformer (ViT) [11] in CLIP [46]. The ViT takes in the image patches, and extracts features for each image patch. Then we use a trainable linear layer to project the extracted image features into the input embedding space of LLM. This architecture, i.e. linear layer for visual feature alignment, has been shown effective by recent multimodal large models like LLaVA [36]. + +The predicted data table is critical to encode the textual information in the image. Because the pretrained ViT (in CLIP [46]) is not specifically trained for text-heavy images, the image features have limited OCR ability. Therefore, we use DEPLOT [34], which is designed for information extraction from charts, to predict the underlying data table for the given chart image. The linearized data table, which is a sequence of word tokens, is fed into the LLM. + +Finally, the projected image features, the data table, and a prompt like "Ask a question about the given chart plot and + +answer it" (rephrased in several different ways), are provided as the input to LLM, for question generation. + +# 3.2. Straight-forward data generation + +Given the above inputs, a straightforward manner for data generation is to have the LLM generate question and answer directly, as shown in Fig. 2 (b). The output $Y$ is the generated question $Q$ , followed by the corresponding answer $Q$ : + +$$ +Y = [ Q; A ] +$$ + +This straightforward data generator can be trained using the existing chart VQA datasets, which contains chart images paired with question and answers. The model is trained with a next-token prediction loss. Note that the ViT and LLM are pretrained and kept frozen during training. The only trainable parameters are the linear projection layer1. + +As our experiments will show, this straightforward generation method yields reasonable results, generating questions and answers that are helpful for training the chart VQA models. However, the generated answers are noisy, which is improved using step-by-step generation, described next. + +# 3.3. Synthesize step-by-step + +Here we introduce Synthesize Step-by-Step, which shares the same model architecture as straight-forward generation, but performs data synthesizing in a step-by-step manner. + +Instead of having the model generate answers directly, we train the LLM to generate them in a step-by-step manner. The LLM generates the question $Q$ with the corresponding + +![](images/21ba88fae6548e22fc6795757509dad0e68187b39a4a53bfe5e086d0243165df.jpg) +Figure 3. Example of template-based question generation. Given a manually created template with defined program, by querying the image SVG annotations, templatic questions and answers can be generated, with rationales. + +rationales $R$ . Then the answer $A$ is derived by executing the rationales using python scripts: + +$$ +Y = [ Q; R ], A = \mathrm {E x e c u t e} (R) +$$ + +As the example shows in Fig. 2 (c), the LLM outputs the generated question and a decomposition of this complex reasoning question into a series of atom reasoning steps, i.e. rationales. The rationales $R$ a step-by-step executable program, where each step in the program can be either an atom extractive VQA questions (e.g. ans_0 = VQA ("What is the value of 2002?")) or a Python-like program call for math calculation, counting, comparison, etc. (e.g. ans = avg(ans_0, ans_1)). The rationale program is then parsed and executed using Python in order to derive the final answer. During execution, there are two ways to answer the atom extractive VQA questions: we can either use an off-the-shelf VQA model, or directly predict the answer using the LLM data generator itself. We experiment with both options and find the latter option to be more effective, as well as more simple, without requiring additional VQA models. + +Training with rationales. To train the step-by-step generator, data with rationales is required. This is different from the straightforward generation, which only requires question-answer pairs to train. Because existing datasets do not contain the rationale annotations for each question, we introduce a template-based QA generation pipeline, which will be described in Sec. 3.4, to generate synthetic questions with rationales. With the template-generated data as the training corpus, we tune the generator for step-by-step generation. An additional advantage of training with template-generated data is that, in inference time, the generator can be prompted with different prompts for controllable generation. The prompts may specify the format or the type of the questions, e.g. "the question should start with how many", "the question should ask about the colors in the image", etc. A complete list of the prompts can be found in Appendix. + +Post hoc filtering. We take an additional post hoc filtering step to filter out the noisy questions. We use the decoding score (for the whole generated sequence including question and rationales) as an indicator and filter out the questions with decoding scores lower than a threshold. Better filtering strategies, like using LLMs for data filtering, can be explored, which we will leave as future work. + +# 3.4. Tuning corpus: template-based QA generation + +Here we describe how we use a template-based pipeline to generate questions with rationales for training the LLM-based generator. The method is motivated by the success of synthetic compositional questions in VQA datasets like CLEVR [21], Super-CLEVR [33] and GQA [20]. + +Fig. 3 illustrates the template-based question generation method. We first manually create question templates, then instantiate the templates into questions and answers by querying the rich SVG annotations of the chart image (as shown in the dotted boxes in the figure). For example, the template "What is the average of the $\langle N\rangle ?$ can be instantiated into "what is the average of the more category?". The argument $\langle N\rangle$ is instantiated into "the more category" according to the SVG annotations, which includes the category names, values, and visual elements like colors and bounding boxes. Moreover, for each template, we define a reasoning program (in a domain-specific language) containing a series of reasoning operations (e.g. select_color, query_value) that specifies the reasoning process needed to answer the question. The program is then instantiated into a step-by-step rationale. + +Note that there are limitations to this template-based QA generation pipeline. The pipeline relies on the rich annotations extracted automatically from the source SVG files, and thus cannot generalize to unlabeled images. Moreover, even though some existing datasets like ChartQA provide such rich annotations crawled from websources, they are inevitably noisy, containing missing or incorrect values. In contrast, the LLM generator is more generalizable, being + +able to generate data for images in the wild. Therefore, in our work, these template-based questions are used as the training corpus of the LLM-based generator, which is more generalizable and scalable. + +# 3.5. Evaluation of generated data + +To measure the usefulness of the generated QA, we use the generated data to train the chart VQA models. The accuracy improvement of chart VQA models, after training with the generated QA, is taken as the indicator of the data quality. We select MATCHA [35], which is a transformer encoder-decoder model pretrained on large-scale text-heavy images, as our primary evaluation model, considering its modern architecture and state-of-the-art performance on standard datasets2. + +Summary. Our data generation procedure contains the following steps: first, we generate templatic QA corpus with rationales using the rich SVG annotations, which is then used to tune the LLM-based data generator; next, the LLM-based generator generates questions with step-by-step rationales; then we use Python to execute the rationales and derive the answers; finally, this LLM-augmented data, LAMENDA, is used to train the downstream VQA model MATCHA. As the experiments will show, the augmented data leads to significant performance improvements. + +# 4. Experiment + +# 4.1. Set up + +Dataset. Following [34, 35], we run experiments on two chart VQA datasets: ChartQA and PlotQA $^{3}$ . ChartQA [40] contains 21k real-world chart images crawled from four web sources, with 9k human-annotated (Human split) and 23k machine-augmented (Augmented split) QA pairs. The augmented questions are mostly extractive questions, while the human-annotated questions are free-form and require complex reasoning. PlotQA [41] is a large synthetic dataset containing 224k chart images generated from real-world online data, with 37M synthetic questions and answers divided into two splits: V1 (8M) and V2 (29M). In addition, to show the generalizability of our LLM data generator, we generate QA for images from additional chart captioning datasets, which do not provide QAs or rich SVG annotations. We use images from three datasets: Chart-to-Text [24], VisText [52] and ChartSumm [47]. Chart-to-Text contains 35k images from "Statistica" and 9k images from "Pew"; VisText contains 9k synthetic charts of various styles; ChartSumm contains 84k images from "Statistica" + +2While there are two-stage models of better performance, which extracts information from charts then prompts LLMs like Codex to answer questions, MATCHA is the best single-stage model. +3There are other chart VQA datasets like FigureQA, DVQA, etc., but since they are out-dated and models reach $>95\%$ accuracy, we focus on more challenging settings. + +and "Knoema". Note that we only use the images from the chart captioning datasets without using the captions. + +Generated data. Tab. 1 shows the statistics of the generated QA. To construct the template-based tuning corpus, we use 28 templates to generate $357\mathrm{k}$ question-answer pairs for the ChartQA training images. The template list and the template QA can be found in the Appendix. We train our LLM data generator using the template-based corpus, and generate QAs for images in ChartQA, PlotQA, and the three chart captioning datasets. For chartQA training images, we generate $403\mathrm{k}$ QAs, which reduces to $326\mathrm{k}$ after step-by-step execution (dropping ones that cannot be executed) and filtering based on the decoding scores using a threshold of -10. For PlotQA training images, we generate 3M QAs, which reduces to $1.7\mathrm{M}$ after execution and filtering. For the $137\mathrm{k}$ chart images in the chart captioning datasets, we generate $1.6\mathrm{M}$ QAs. + +
DatasetImagesGeneratedFiltered
TemplateChartQA18,317356,606-
LAMENDAChartQA18,317402,974326,160
PlotQA157,0703,455,5391,726,822
ChartCap137,2812,987,7181,635,364
+ +Table 1. Statistics of generated data, before and after filtering. + +Implementation details. Using the 357k template QA, we tune the data generator for 5 epochs with a batch size of 128. We use a cosine learning rate schedule with a max learning rate of 2e-3. The training takes about 12 hours using 8 A100 GPUs. The trainable projection layer is initialized with the first stage checkpoint of LLaVA [36]. We select the best checkpoint at 14k iterations as our best generator. For MATCHA training, we take the pretrained checkpoints (base size) from HuggingFace and finetune the models with a combination of our generated questions and the original ChartQA or PlotQA training questions. The training setting follows [28, 35]. We use a cosine scheduler with 500 warm-up steps and a max learning rate of 0.0001. The batch size is 256. Without special description, the models are trained for 10k iterations for ChartQA (combining human and augmented splits) and 20k iterations for PlotQA because it is much larger (V1 and V2 are trained separately). We set the max number of image patches as 4096. The ChartQA training takes 3 days on 8 32GB V100 GPUs. Considering the training time, in some experiments, we reduce the image patch number to 1024 for quick experiments in 12 hours with the same setting. + +Evaluation metrics. We suggest that both strict accuracy and relaxed accuracy should be reported, especially for the ChartQA dataset. While existing works [34, 35, 40, 41] evaluate answers are evaluated using relaxed accuracy, which allows $5\%$ error for numerical answers, we find that this relaxed accuracy metric is problematic for ChartQA. + +
AccuracyRelaxed Accuracy
avghumanaugmentavghumanaugment
0baseline (MATCHA [35])47.7630.2165.3158.5439.5877.50
1syna48.8028.6568.9659.4338.6580.21
2synat w/ VQA answer47.7628.4467.0857.9737.8178.12
3synat w/ step-VQA answer52.5537.2967.8164.3848.8579.90
4synat w/ synat answer53.7537.9269.5865.1048.8581.35
5synat w/ step-synat answer55.2140.2170.2166.4151.3581.46
6synat w/ step-synat answer + tpl.57.3442.6072.0867.8653.1282.60
7synat w/ step-synat answer + tpl. + addi.58.6545.6271.6769.8456.5683.12
+ +Table 2. Main results on ChartQA val split. The data generator trained on Annotated questions is denoted as $\mathrm{syn}_a$ ; the data generator trained on both Annotated questions and our Template-generated questions $\mathrm{syn}_{at}$ is denoted as $\mathrm{syn}_{at}$ . Step-by-step generation is better than straight-forward generation (row 1-5). Training with a combination of step-by-step generated QA, template-generated QA (+mpl.), and additional images from chart captioning datasets (+addi.), leads to the best model. See Sec. 4.2 for more details. + +For example, many questions in ChartQA ask about years (e.g. 1996, 2012), for which allowing $5\%$ numerical errors (i.e. around 100 years) will result in incorrectly high results. + +# 4.2. Main results + +We study our data generation method with different settings on the ChartQA dataset, with MATCHA as our baseline. We first generate QA data using our method, and then use the generated data to train the MATCHA model. The strict accuracy and relaxed accuracy reported on the ChartQA validation split are shown in Tab. 2. + +The baseline model MATCHA (row-0) achieves $47.76\%$ accuracy and $58.54\%$ relaxed accuracy. For this baseline, we finetune the pretrained MATCHA checkpoint for 10k iterations with the ChartQA training split (combing human and augmented questions), using 1024 image patches. This setting follows [35], except that a smaller number of image patches are used for faster experiments and less GPU memory requirement. We validate the correctness of our implementation by showing that the model reaches $63.7\%$ relaxed accuracy using the original resolution, which matches the numbers reported in [35] (as shown in Tab. 3). + +Row-1 shows the results for the straight-forward data generation, where we train the data generator with QAs from the ChartQA training split and use the trained generator to directly generate questions and answers. This straightforward data generation improves the accuracy by $1\%$ (from $47.76\%$ to $48.80\%$ ), suggesting that LLMs can be helpful in data augmentation for chart reasoning. + +Row-2 to row-5 shows the advantage of synthesizing step-by-step over straight-forward generation. We train the data generator with our template-generated QA, but the answers are generated with different methods. In row-2, we feed in the generated complex question into an additional VQA model (i.e. baseline MATCHA) to get the answers; in row-3, we generate the step-by-step rationales, and the stepwise questions are fed into the VQA model to get step-wise answers, then the final answer is derived using the step + +wise answers. As can be seen, the step-wise answer generation (52.55%) is significantly better than the direct answer generation (47.76%), which suggests the importance of decomposing the complex questions into atom sub-questions. Row-4 and row-5 follow a similar setting, except that we use the LLM data generator to generate the answers instead of the additional VQA model. Two findings can be drawn: (a) LLM-generated answers are better than VQA-predicted answers, possibly because the LLM data generator has access to the predicted data tables as inputs; (b) step-by-step synthesizing outperforms direct generation, for both VQA-predicted and LLM-generated answers. + +In row-6, we combine the step-by-step generated data with the templated-generated QAs and show that training with both leads to further better performance (57.34%). In row-7, we generate QAs for the chart images from additional chart captioning datasets. The final model, trained with a combination of all the synthetic data, gets 58.65% accuracy. Compared with the baseline, training with the augmented data leads to a 10.89% accuracy improvement. Moreover, the improvement is more significant on the human-written questions (+15.41%), which requires complex reasoning, than the augmented extractive questions (+6.36%). The results suggest that our synthetic data improves both the extractive ability and the reasoning ability, with the latter being more significant. We show examples in the Appendix. + +# 4.3. Comparison with SoTA + +To compare our method with the best performing models, we apply our generated data onto two models, Pix2Struct [28] and MATCHA [35], and report the results on the ChartQA test split in Tab. 3. For faster training, we first train the model (for both Pix2struct and MATCHA) with a max of 1024 image patches using our best setting (i.e. row-7 in Tab. 2), then finetune with 4096 image patches to get the final result. We report both the 1024-resolution results and the 4096-resolution results. + +
paramsaccuracyRelaxed Accuracy
avghumanaugmentavghumanaugment
VL-T5-OCR [40]----41.56--
VisionTapas-OCR [40]----45.52--
Pix2Struct4096 [28]282M---56.030.581.6
MATCHA4096 [35]282M---64.238.290.2
DEPLOT+FlanPaLM+Codex [34]540B---79.367.691.0
PaLI-X [7]55B---72.3--
Pix2Struct4096 (reimpl.)282M50.4424.6476.2458.2432.0084.48
Pix2Struct1024 + LAMENDA282M51.7734.6968.8563.0246.2579.79
Pix2Struct4096 + LAMENDA282M55.8340.1071.5666.8251.6781.98
MATCHA4096 (reimpl.)282M55.6831.0480.3263.7237.7689.68
MATCHA1024 + LAMENDA282M60.2841.9278.6469.8851.7688.00
MATCHA4096 + LAMENDA282M63.3644.8881.8472.6453.9291.36
+ +As can be seen, our generated data improves the overall relaxed accuracy of Pix2Struct by $8.58\%$ and MATCHA by $8.92\%$ . Moreover, the improvement is much more significant on human-written questions, e.g. the relaxed accuracy of MATCHA on human-written questions improves from $38\%$ to $54\%$ . Interestingly, with our augmented data, the low-resolution (1024) models can already outperform their high-resolution (4096) counterparts without data augmentation. The significant performance gain indicates that learning with synthetic data is extremely helpful for better reasoning ability with human-written queries. + +Our best performing MATCHA achieves $72.6\%$ relaxed accuracy, which beats all existing end-to-end chart reasoning models. Note that DEPLOT+FlanPaLM+Codex is a two-stage model that first extracts information from the charts, then prompts external LLMs like FlanPalm (540B) Codex (12B) for question answering. The method relies on much larger models of billions of parameters and thus is not comparable with our method, which is much smaller and not rely on external LLMs for inference. + +# 4.4. Generalization + +To show the generalization ability of the proposed data generation method, we run experiments on the PlotQA dataset. We directly use our data generator (trained on template questions on ChartQA) to generate QA for the PlotQA training images. The results are shown in Tab. 4. As can be seen, our method, when applied to MATCHA, achieves state-of-the-art performance (92.89% relaxed accuracy) on PlotQA. Despite the synthetic nature of this dataset and the saturated model performance, our augmented data can still help on both low-resolution (+3.46%) and high-resolution (+0.92%) settings. + +# 4.5. Ablations and analysis + +Qualitative analysis and failure cases. Fig. 5 shows examples of our LLM-generated QA, including good and bad ex + +samples. Most generated questions are specific to the given image, asking about the categories plotted, such as “2019” or “Tablet 2018”, indicating that the model is able to utilize the information of the input. Moreover, various questions are generated, including counting, colors, average, etc. In the failure example, the question requires comparing the “silver bar” with the “midnightblue” bar, with reasonable rationales. However, there are multiple bars in the specified colors in the image, resulting in ambiguous/invalid questions. When looking at more failure cases, we find that there are several typical failure modes. For example, questions are sometimes invalid/ambiguous (e.g. asking about the top/bottom bar for a vertical bar plot); questions can be hallucinated or irrelevant (e.g. asking about a category that is not shown in the figure); rationales and answers are not correct, etc., which opens up opportunities for future works. Before and after LAMENDA. To gain insights on how LAMENDA helps model training, we conduct a comparison analysis of MATCHA “before” and “after” training with our LLM-generated data, i.e. the baseline model versus our best model. According to the predictions of these two models, we divide the 960 human-written questions from ChartQA val split into 4 groups: both model answers correctly, both models answers wrongly, the before model answers correctly, but the after model answers wrongly, and + +Table 3. Comparison with SoTAs on ChartQA test split. Subscripts (1024, 4096) means image token numbers (resolution). With our generated data, both Pix2Struct and MATCHA achieve significantly better performance, even comparable to much larger models. + +
Relaxed Accuracy
avgV1V2
VisionTapas-OCR [40]53.9065.3042.50
VL-T5-OCR [40]65.9675.9056.02
DEPLOT+FlanPaLM+Codex [34]66.662.271.0
Pix2Struct [28]72.573.271.9
MATCHA4096 [35]91.592.390.7
MATCHA1024 (reimpl.)74.9573.8876.02
MATCHA1024 + LAMENDA78.4178.4478.38
MATCHA4096 (reimpl.)91.9792.6491.30
MATCHA4096 + LAMENDA92.8993.9491.84
+ +Table 4. Comparison with SoTAs on PlotQA test split. With our generated data, MATCHA achieves the SoTA performance. + +![](images/2ac7fdad08d09825866a1179e3b4f14522bf0b5f0bfdc1c6bfe0db7d70be31ba.jpg) +Figure 4. Analysis of model predictions before and after training with LAMENDA. In I (left), ChartQA val questions are divided into four categories according to the model prediction correctness (before/after). II (right) further decomposes each category into different question types. LAMENDA improves the model prediction, and the improvement is most significant on reasoning questions. + +vice versa. The statistics of these four groups are shown in Fig. 4. From Fig. 4 I. (left side), we see that LAMENDA is helpful for $22\%$ of the questions, while there are still $45\%$ of the questions where both models cannot answer correctly. In Fig. 4 II. (right side), we randomly sample 30 questions from each of the four groups and manually classify them into six question types: extractive, count, spatial, color, reasoning, and others. As shown in (a), i.e. the questions where the "before" model answers wrongly and our model answers correctly, most $(70\%)$ of the questions belong to the reasoning category, indicating that our data helps most on the reasoning questions. Our data also helps with color questions $(10\%)$ and count questions $(17\%)$ . However, our data does not help with extractive questions $(0\%)$ . In addition, comparing the distribution of (c) both correct with (d) both wrong, we can see that extractive questions are easy for the models, while reasoning questions are very challenging. + +In summary, Fig. 4 shows the effectiveness of our data for the reasoning questions over other question types. However, reasoning questions are still a big challenge for current models, which requires further exploration. + +Different question types. In Tab. 5, we study the improve + +![](images/709dbcf39525bede908916656b06832d45d92e989ca0be34efb3600a1540b0ac.jpg) +Figure 5. Examples of generated QA, 3 good examples and 1 failure case. Rationales are only shown for 2 examples for brevity. + +ments brought by each type of the generated questions. The total 326k generated QAs are grouped into 7 different groups, including questions about color understanding, spatial reasoning, counting, minimum/maximum reasoning, average, comparison (e.g. larger than, smaller than), and other math calculations. This grouping is done by different prompting of the data generation LLM, as specified in the Appendix. The improvements from different question types are distinct and can be accumulated. As we see, each of the seven types improves the model performance by less than $3\%$ , while combining all of them leads to $7.03\%$ improvement. Moreover, among the seven types, the "math calculation" questions brings the most improvements $(2.76\%)$ . + +
#QAavghumanaug
baseline28,29947.7630.2165.31
+colors+46,51247.8129.9065.73
+spatial+49,38749.2730.6367.92
+count+36,70547.4029.1765.63
+minmax+71,78849.5831.8767.29
+average+50,71748.9631.3566.56
+compare+14,69047.6628.8566.46
+calculation+56,36150.5233.8567.19
+all326,16054.7939.2770.31
+ +Table 5. A detailed look at the effect for different question types. Strict accuracy on ChartQA val split is shown. Each type helps and combining all of them leads to a further performance gain. + +# 5. Conclusion + +We propose Synthesize Step-by-Step, which utilizes LLMs to augment the training data for the chart reasoning task. We show that the step-by-step generation, which decomposes the complex reasoning questions into easier sub-questions, is critical for synthesizing good-quality data. We run experiments on two standard benchmarks and show significant improvements by training with the augmented data. Specifically, the relaxed accuracy on the human-written questions is improved from $37.8\%$ to $53.9\%$ on ChartQA. The large improvement indicates the promise of boosting the model's reasoning ability by data generation using LLMs, from which the strong step-by-step reasoning ability can be distilled into downstream models. + +# References + +[1] Mubashara Akhtar, Oana Cocarascu, and Elena Simperl. Reading and reasoning over chart images for evidence-based automated fact-checking. arXiv preprint arXiv:2301.11843, 2023. 2 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2 +[3] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 39-48, 2016. 2 +[4] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R Manmatha. Docformerv2: Local features for document understanding. arXiv preprint arXiv:2306.01733, 2023. 2 +[5] Ritwick Chaudhry, Sumit Shekhar, Utkarsh Gupta, Pranav Maneriker, Prann Bansal, and Ajay Joshi. Leaf-qa: Locate, encode & attend for figure question answering. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3512–3521, 2020. 2 +[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 2 +[7] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 2, 7 +[8] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794, 2022. 2 +[9] John Joon Young Chung, Ece Kamar, and Saleema Amershi. Increasing diversity while maintaining accuracy: Text data generation with large language models and human interventions. arXiv preprint arXiv:2306.04140, 2023. 2 +[10] Brian Davis, Bryan Morse, Brian Price, Chris Tensmeyer, Curtis Wigington, and Vlad Morariu. End-to-end document recognition and understanding with dessurt. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 2 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at + +scale, 2021. 2 +[13] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 2 +[14] Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. Knowledge distillation of large language models. arXiv preprint arXiv:2306.08543, 2023. 2 +[15] Nitish Gupta, Kevin Lin, Dan Roth, Sameer Singh, and Matt Gardner. Neural module networks for reasoning over text. arXiv preprint arXiv:1912.04971, 2019. 2 +[16] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 2 +[17] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2 +[18] Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes. arXiv preprint arXiv:2305.02301, 2023. 2 +[19] Ronghang Hu, Jacob Andreas, Trevor Darrell, and Kate Saenko. Explainable neural computation via stack neural module networks. In Proceedings of the European conference on computer vision (ECCV), pages 53-69, 2018. 2 +[20] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 4 +[21] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. 4 +[22] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering, 2018. 2 +[23] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Akos Kadar, Adam Trischler, and Yoshua Bengio. Figureqa: An annotated figure dataset for visual reasoning, 2018. 2 +[24] Shankar Kantharaj, Rixie Tiffany Leong, Xiang Lin, Ahmed Masry, Megh Thakkar, Enamul Hoque, and Shafiq Joty. Chart-to-text: A large-scale benchmark for chart summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4005–4023, 2022. 5 +[25] Tushar Khot, Harsh Trivedi, Matthew Finlayson, Yao Fu, Kyle Richardson, Peter Clark, and Ashish Sabharwal. Decomposed prompting: A modular approach for solving complex tasks. arXiv preprint arXiv:2210.02406, 2022. 2 +[26] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang + +doo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 2 +[27] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 2 +[28] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In International Conference on Machine Learning, pages 18893–18912. PMLR, 2023. 2, 5, 6, 7 +[29] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 2 +[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2 +[31] Liunian Harold Li, Jack Hessel, Youngjae Yu, Xiang Ren, Kai-Wei Chang, and Yejin Choi. Symbolic chain-of-thought distillation: Small models can also" think" step-by-step. arXiv preprint arXiv:2306.14050, 2023. 2 +[32] Zhuowan Li, Elias Stengel-Eskin, Yixiao Zhang, Cihang Xie, Quan Hung Tran, Benjamin Van Durme, and Alan Yuille. Calibrating concepts and operations: Towards symbolic reasoning on real images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14910-14919, 2021. 2 +[33] Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Kortylewski, Wufei Ma, Benjamin Van Durme, and Alan L Yuille. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14963-14973, 2023. 2, 4 +[34] Fangyu Liu, Julian Martin Eisenschlos, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Wenhu Chen, Nigel Collier, and Yasemin Altun. Deplot: One-shot visual language reasoning by plot-to-table translation. arXiv preprint arXiv:2212.10505, 2022. 2, 3, 5, 7 +[35] Fangyu Liu, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Yasemin Altun, Nigel Collier, and Julian Martin Eisenschlos. Matcha: Enhancing visual language pretraining with math reasoning and chart derendering. arXiv preprint arXiv:2212.09662, 2022. 1, 2, 5, 6, 7 +[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 3, 5 +[37] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842, 2023. 2 +[38] Junyu Luo, Zekun Li, Jinpeng Wang, and Chin-Yew Lin. + +Chartocr: Data extraction from charts images via a deep hybrid framework. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1917-1925, 2021. 2 +[39] Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B Tenenbaum, and Jiajun Wu. The neuro-symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision. arXiv preprint arXiv:1904.12584, 2019. 2 +[40] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 1, 2, 5, 7 +[41] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2020. 2, 5 +[42] Aaron Parisi, Yao Zhao, and Noah Fiedel. Talm: Tool augmented language models. arXiv preprint arXiv:2205.12255, 2022. 2 +[43] Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334, 2023. 2 +[44] Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). Association for Computational Linguistics, 2019. 2 +[45] Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, et al. Toolllm: Facilitating large language models to master 16000+ real-world apis. arXiv preprint arXiv:2307.16789, 2023. 2 +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[47] Raian Rahman, Rizvi Hasan, Abdullah Al Farhad, Md Tahmid Rahman Laskar, Md Hamjajul Ashmafee, and Abu Raihan Mostofa Kamal. Chartsumm: A comprehensive benchmark for automatic chart summarization of long and short summaries. arXiv preprint arXiv:2304.13620, 2023. 5 +[48] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. arXiv preprint arXiv:2302.04761, 2023. 2 +[49] Timo Schick and Hinrich Schütze. Generating datasets with pretrained language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6943-6951, 2021. 2 +[50] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580, 2023. 2 +[51] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: + +Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128, 2023. 2 +[52] Benny J Tang, Angie Boggust, and Arvind Satyanarayan. Vistext: A benchmark for semantically rich chart captioning. arXiv preprint arXiv:2307.05356, 2023. 5 +[53] Peng Tang, Srikar Appalaraju, R Manmatha, Yusheng Xie, and Vijay Mahadevan. Multiple-question multiple-answer text-vqa. arXiv preprint arXiv:2311.08622, 2023. 2 +[54] Peng Tang, Pengkai Zhu, Tian Li, Srikar Appalaraju, Vijay Mahadevan, and R Manmatha. Deed: Dynamic early exit on decoder for accelerating encoder-decoder transformer models. arXiv preprint arXiv:2311.08623, 2023. 2 +[55] MN Team et al. Introducing mpt-7b: A new standard for open-source, commercially usable llms, 2023. 3 +[56] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 2 +[57] Shuohang Wang, Yang Liu, Yichong Xu, Chenguang Zhu, and Michael Zeng. Want to reduce labeling cost? gpt-3 can help. arXiv preprint arXiv:2108.13487, 2021. 2 +[58] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171, 2022. 2 +[59] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 2 +[60] Peter West, Chandra Bhagavatula, Jack Hessel, Jena D Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, and Yejin Choi. Symbolic knowledge distillation: from general language models to commonsense models. arXiv preprint arXiv:2110.07178, 2021. 2 +[61] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1192-1200, 2020. 2 +[62] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023. 2 +[63] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601, 2023. 2 +[64] Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Josh Tenenbaum. Neural-symbolic vqa: Disentangling reasoning from vision and language understanding. Advances in neural information processing systems, 31, 2018. 2 +[65] Yue Yu, Yuchen Zhuang, Jieyu Zhang, Yu Meng, Alexander Ratner, Ranjay Krishna, Jiaming Shen, and Chao Zhang. Large language model as attributed training data generator: A tale of diversity and bias. In Thirty-seventh Conference on + +Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2 +[66] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 2 +[67] Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 2 +[68] Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. Automatic chain of thought prompting in large language models. arXiv preprint arXiv:2210.03493, 2022. 2 +[69] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 2 +[70] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2 \ No newline at end of file diff --git a/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/images.zip b/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6cfb33656ddfe5674439973d72fba1bdab35afdb --- /dev/null +++ b/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a190c8addf71eef5ff597933b310911e92ef57bce05e8bc493b752db097b5ec7 +size 470399 diff --git a/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/layout.json b/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..03990222284a0bf4400731989ba6e6dbcf51511e --- /dev/null +++ b/2024/Synthesize Step-by-Step_ Tools Templates and LLMs as Data Generators for Reasoning-Based Chart VQA/layout.json @@ -0,0 +1,8387 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 482, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 482, + 140 + ], + "type": "text", + "content": "Synthesize Step-by-Step: Tools, Templates and LLMs as Data Generators for Reasoning-Based Chart VQA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 160, + 462, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 160, + 462, + 175 + ], + "spans": [ + { + "bbox": [ + 129, + 160, + 462, + 175 + ], + "type": "text", + "content": "Zhuowan Li" + }, + { + "bbox": [ + 129, + 160, + 462, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 160, + 462, + 175 + ], + "type": "text", + "content": " * Bhavan Jasani" + }, + { + "bbox": [ + 129, + 160, + 462, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 129, + 160, + 462, + 175 + ], + "type": "text", + "content": " * Peng Tang" + }, + { + "bbox": [ + 129, + 160, + 462, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 129, + 160, + 462, + 175 + ], + "type": "text", + "content": " Shabnam Ghadar" + }, + { + "bbox": [ + 129, + 160, + 462, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 185, + 175, + 409, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 175, + 409, + 190 + ], + "spans": [ + { + "bbox": [ + 185, + 175, + 409, + 190 + ], + "type": "text", + "content": "1 Johns Hopkins University 2 AWS AI Labs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 153, + 191, + 438, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 191, + 438, + 202 + ], + "spans": [ + { + "bbox": [ + 153, + 191, + 438, + 202 + ], + "type": "text", + "content": "zili110@jhu.edu {bjasani, tangpen, shabnam}@amazon.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 256, + 290, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 256, + 290, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 256, + 290, + 581 + ], + "type": "text", + "content": "Understanding data visualizations like charts and plots requires reasoning about both visual elements and numerics. Although strong in extractive questions, current chart visual question answering (chart VQA) models suffer on complex reasoning questions. In this work, we address the lack of reasoning ability by data augmentation. We leverage Large Language Models (LLMs), which have shown to have strong reasoning ability, as an automatic data annotator that generates question-answer annotations for chart images. The key innovation in our method lies in the Synthesize Step-by-Step strategy: our LLM-based data generator learns to decompose the complex question into step-by-step sub-questions (rationales), which are then used to derive the final answer using external tools, i.e. Python. This step-wise generation procedure is trained on synthetic data generated using a template-based QA generation pipeline. Experimental results highlight the significance of the proposed step-by-step generation. By training with the LLM-augmented data (LAMENDA), we significantly enhance the chart VQA models, achieving the state-of-the-art accuracy on the ChartQA and PlotQA datasets. In particular, our approach improves the accuracy of the previous state-of-the-art approach from " + }, + { + "bbox": [ + 47, + 256, + 290, + 581 + ], + "type": "inline_equation", + "content": "38\\%" + }, + { + "bbox": [ + 47, + 256, + 290, + 581 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 256, + 290, + 581 + ], + "type": "inline_equation", + "content": "54\\%" + }, + { + "bbox": [ + 47, + 256, + 290, + 581 + ], + "type": "text", + "content": " on the human-written questions in the ChartQA dataset, which needs strong reasoning. We hope our work underscores the potential of synthetic data and encourages further exploration of data augmentation using LLMs for reasoning-heavy tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 624, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 287, + 685 + ], + "type": "text", + "content": "Data visualizations like charts and plots play an important role in real-world data analysis applications. Unlike natural images, chart are text-heavy and data-driven, thus requiring better visual perception (e.g. OCR) and cognitive reasoning (e.g. math calculation, matching the legends with bars). For" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 308, + 221, + 545, + 364 + ], + "blocks": [ + { + "bbox": [ + 308, + 221, + 545, + 364 + ], + "lines": [ + { + "bbox": [ + 308, + 221, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 221, + 545, + 364 + ], + "type": "image", + "image_path": "d6864ec69a86d47cd65f2a00a46dc0c667af30eb534d8990db4d35a3bc1694eb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 373, + 547, + 475 + ], + "lines": [ + { + "bbox": [ + 304, + 373, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 547, + 475 + ], + "type": "text", + "content": "Figure 1. Existing chart VQA models struggle with complex reasoning questions. We attribute this to limited reasoning questions in existing datasets and address it by data augmentation. We fine-tune an LLM-based data generator that automatically generates question-answer annotations given a chart image. Our key innovation is Synthesize Step-by-Step, which breaks complex questions down into easy steps that could be solved using external tools. We use templates to train the LLM. Training with LLM-augmented data, LAMENDA, greatly enhances the chart VQA models." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 485, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 545 + ], + "type": "text", + "content": "example, to answer \"what is the total of Democrats and Republicans in 2010\" in Fig. 1, the model needs to recognize the texts, match the legends of \"democrats\" and \"republicans\" to the bars, then calculate the sum. This is a challenging task requiring multimodal perception and reasoning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": "Compounding the complexity, the inherent difficulty of chart reasoning also amplifies the cost of collecting human annotations. Most existing datasets are synthetically generated using simple rules and heuristics, posing a risk of model overfitting. Among the datasets, only ChartQA [40] contains complex human-annotated question-answer pairs, although there are very few of those per image. For example, for the image in Fig. 1, while a lot more questions can be asked, only two questions are annotated (What is the total of Republicans and Democrats in 2010? Which color represents Republican?). Importantly, these human-written questions pose great challenges for contemporary models. For example, the state-of-the-art model [35] achieves only " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "38\\%" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": " accuracy when confronted with human-written ques" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": "*indicates equal contributions. The work was done when Zhuowan Li was an intern at Amazon." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13613" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "tions that require multi-step complex reasoning. Approximately half of the errors, as reported in [35], are due to reasoning failures." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "type": "text", + "content": "In this work, we propose to leverage large language models (LLMs) as automatic data annotators for chart reasoning. Recently, LLMs have demonstrated strong reasoning ability for complex tasks [27, 58, 59, 63, 68, 69]. Our motivation is to take advantage of this strong reasoning ability and generate question-answer pairs using LLMs. This LLM-generated data can be used to train the chart VQA models, thus improving their reasoning ability." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 204, + 288, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 288, + 430 + ], + "type": "text", + "content": "The key innovation in our method lies in the Synthesize Step-by-Step strategy, which leverages LLMs to generate data in a step-by-step manner. Similar to human annotators who are trained to break down complex tasks into easier subtasks, our LLM-based data generator generates step-by-step. Instead of generating question-answer all at once, the generator learns to generate the question first, then decompose the complex question into step-by-step sub-questions (rationales) and answer them one by one. Finally, the answers of the sub-questions are combined to derive the final answer using external tools like calculators or Python code. Specifically, we feed the image features extracted with a ViT [12] backbone into an LLM using a simple trainable projection layer. To supervise the training of this projection layer, we construct a template-based training corpus generated using a template-based QA generation pipeline. Importantly, as we will show in the paper, the step-by-step synthesizing, trained using these template-based rationales, is critical for generating good-quality data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "type": "text", + "content": "Experiments showcase the effectiveness of Synthesize Step-by-Step. Our LLM-augmented Data, LaMenDa, is used to train the downstream chart VQA models and its effectiveness is shown on two challenging datasets, ChartQA and PlotQA with the state-of-the-art accuracy. Notably, we significantly improve the performance from " + }, + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "type": "inline_equation", + "content": "38\\%" + }, + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "type": "inline_equation", + "content": "54\\%" + }, + { + "bbox": [ + 46, + 432, + 287, + 587 + ], + "type": "text", + "content": " on the challenging human-written questions in ChartQA dataset, which requires complex reasoning. We also demonstrate the advantage of step-wise generation over straightforward generation. With the results, we hope our work underscores the potential of synthetic data and encourages further exploration of data augmentation using LLMs for reasoning-heavy tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 597, + 134, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 134, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 134, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Multimodal models for text-heavy images fall into two categories: OCR-dependent or OCR-free. Models like LayoutLM [61], PaLI(-X) [7, 8], ChartBERT [1], DocFormerv2 [4], MQMA [53], DEED [54] rely on external OCR systems like ChartOCR [38] as either model input or training objectives. Recent top performing models, like Donut [26], Dessurt [10], Pix2struct [28], shift away from OCR thanks to vision transformers [11]. Pix2Struct [28] is a trans" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": "former encoder-decoder model for multiple VQA tasks, pretrained with an HTML-masked prediction objective, using a large amount of web page screenshots. MATCHA [35] and DEPLOT [34] are two recent follow-up works, tuning Pix2struct with chart-specific tasks like chart de-rendering and math reasoning, which are used in our work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 144, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 251 + ], + "type": "text", + "content": "Datasets for chart VQA. Earlier datasets like DVQA [22], FigureQA [23], LeafQA [5] are based on synthetically generated images and templatic questions, which are easy to solve and prior models achieve " + }, + { + "bbox": [ + 304, + 144, + 545, + 251 + ], + "type": "inline_equation", + "content": ">95\\%" + }, + { + "bbox": [ + 304, + 144, + 545, + 251 + ], + "type": "text", + "content": " accuracy. PlotQA [41] is also a synthetic dataset but contains open-vocabulary questions. ChartQA [40] is a recent dataset containing real images with human-written questions, which is much more challenging. Our work is mainly based on ChartQA and PlotQA, following the settings in [34, 35]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 251, + 545, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 359 + ], + "type": "text", + "content": "LLMs for multimodal tasks. There is a surge of interest in extending LLMs into the vision-and-language domain. Recent works align the image features into the input space of the pretrained LLMs using a lightweight module, e.g. Frozen [56], Flamingo [2], LLaVA [36], Otter [29], MiniGPT-4 [70], BLIP-2 [30], AdaptorV2 [13], etc. For example, LLaVA [36] uses a simple projection layer to align the modalities and build a visually-finetuned multimodal model; LLaVAR [67] extends it onto text-heavy images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 360, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 491 + ], + "type": "text", + "content": "Reasoning step-by-step has been studied for a long time in the multimodal field. Compositional models like neural modular networks [3, 15, 19] or neural symbolic reasoning [32, 33, 39, 64] decompose complex visual questions into reasoning steps, and compose modules to solve the steps. For LLMs, chain-of-thought prompting is a simple yet effective method to boost the reasoning performance [27, 58, 59, 63, 68, 69]. Recent works like VisProg [16] and ViperGPT [51] utilize LLMs as the planner to compose domain-specific models [37, 43, 62, 66] and external tools or APIs [25, 42, 45, 48, 50] for solving complex tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 491, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 611 + ], + "type": "text", + "content": "LLMs for data generation has been recently studied for language tasks [9, 49, 65] or multimodal tasks [36, 70]. This relates to knowledge distillation [17], which teaches a smaller model by learning from a larger teacher model, benefiting semi-supervised learning [6, 57]. Distillation from LLMs [14, 18, 31, 44, 60] can be completed by either learning from the logits or from the generated data directly. Our work is related but different from these works in that we focus on step-by-step generation, combining LLMs with templates and tools, for the chart reasoning task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 621, + 361, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 361, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 361, + 633 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": "Given a chart plot, our goal is to build a data generator that generates complex reasoning questions with answers. The generated data can be used for training the downstream chart VQA models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "The overview of our LLM-based data generator is shown in Fig. 2. In this section, we describe the data generator in a" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13614" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 64, + 293, + 240 + ], + "blocks": [ + { + "bbox": [ + 74, + 64, + 293, + 240 + ], + "lines": [ + { + "bbox": [ + 74, + 64, + 293, + 240 + ], + "spans": [ + { + "bbox": [ + 74, + 64, + 293, + 240 + ], + "type": "image", + "image_path": "3f5454cc676b243f2de2603e5cdfdc1b54d499dfe70a8af81e7d9896c8561793.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 66, + 451, + 79 + ], + "lines": [ + { + "bbox": [ + 302, + 66, + 451, + 79 + ], + "spans": [ + { + "bbox": [ + 302, + 66, + 451, + 79 + ], + "type": "text", + "content": "(b) straight-forward generation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 86, + 475, + 96 + ], + "lines": [ + { + "bbox": [ + 305, + 86, + 475, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 86, + 475, + 96 + ], + "type": "text", + "content": "Prompt : Ask a question about the chart plot and answer it" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 305, + 103, + 518, + 115 + ], + "lines": [ + { + "bbox": [ + 305, + 103, + 518, + 115 + ], + "spans": [ + { + "bbox": [ + 305, + 103, + 518, + 115 + ], + "type": "text", + "content": "LLM Output : What is the average of the two largest categories?13%" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 305, + 126, + 432, + 139 + ], + "lines": [ + { + "bbox": [ + 305, + 126, + 432, + 139 + ], + "spans": [ + { + "bbox": [ + 305, + 126, + 432, + 139 + ], + "type": "text", + "content": "(c) synthesize step-by-step" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 142, + 510, + 161 + ], + "lines": [ + { + "bbox": [ + 305, + 142, + 510, + 161 + ], + "spans": [ + { + "bbox": [ + 305, + 142, + 510, + 161 + ], + "type": "text", + "content": "Prompt: Ask a question about the chart plot and think step-by-step to answer it." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 305, + 168, + 351, + 178 + ], + "lines": [ + { + "bbox": [ + 305, + 168, + 351, + 178 + ], + "spans": [ + { + "bbox": [ + 305, + 168, + 351, + 178 + ], + "type": "text", + "content": "LLM Output :" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 178, + 477, + 187 + ], + "lines": [ + { + "bbox": [ + 305, + 178, + 477, + 187 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 477, + 187 + ], + "type": "text", + "content": "Question: What is the average of the two largest categories?" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 306, + 188, + 490, + 196 + ], + "lines": [ + { + "bbox": [ + 306, + 188, + 490, + 196 + ], + "spans": [ + { + "bbox": [ + 306, + 188, + 490, + 196 + ], + "type": "text", + "content": "Rationale: ans_0 = VQA(\"What is the value of 2002?\") # 16.2%" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 336, + 198, + 489, + 206 + ], + "lines": [ + { + "bbox": [ + 336, + 198, + 489, + 206 + ], + "spans": [ + { + "bbox": [ + 336, + 198, + 489, + 206 + ], + "type": "text", + "content": "ans_1 = VQA(\"What is the value of 2003?\") # 16.7%" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 336, + 208, + 410, + 217 + ], + "lines": [ + { + "bbox": [ + 336, + 208, + 410, + 217 + ], + "spans": [ + { + "bbox": [ + 336, + 208, + 410, + 217 + ], + "type": "text", + "content": "ans " + }, + { + "bbox": [ + 336, + 208, + 410, + 217 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 336, + 208, + 410, + 217 + ], + "type": "text", + "content": " avg(ans_0, ans_1)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 305, + 224, + 515, + 235 + ], + "lines": [ + { + "bbox": [ + 305, + 224, + 515, + 235 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 515, + 235 + ], + "type": "text", + "content": "Code execution: ans_0 = 16.2%, ans_1 = 16.7%, ans = 16.5%" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 247, + 547, + 292 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 547, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 547, + 292 + ], + "type": "text", + "content": "Figure 2. Method overview. (a) The architecture of the LLM-based question-answer generator. Image features, projected by a linear layer concatenated with the predicted data table and a prompt, are fed into LLM for QA generation. (b) Straight-forward generation, where questions with answers are generated in a straightforward way. (c) Synthesizing step-by-step, which breaks the question down into rationale programs and derives the answer by executing the program." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 308, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 380 + ], + "type": "text", + "content": "progressive manner. We first introduce an intuitive, straightforward method for data generation. Then, we introduce Synthesize Step-by-Step, which enhances the straightforward method with step-wise rationales. Finally, we introduce the settings for training the downstream chart VQA models using our generated QA data." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 385, + 129, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 385, + 129, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 385, + 129, + 396 + ], + "type": "text", + "content": "3.1. Architecture" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 403, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 475 + ], + "type": "text", + "content": "The architecture of our LLM-based data generator is shown in Fig. 2 (a). We leverage MPT-7B [55] as the LLM for data generation. The inputs " + }, + { + "bbox": [ + 46, + 403, + 287, + 475 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 403, + 287, + 475 + ], + "type": "text", + "content": " to LLM is a concatenation of three components: the projected image features " + }, + { + "bbox": [ + 46, + 403, + 287, + 475 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 403, + 287, + 475 + ], + "type": "text", + "content": ", the underlying data table for the chart image " + }, + { + "bbox": [ + 46, + 403, + 287, + 475 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 403, + 287, + 475 + ], + "type": "text", + "content": ", and a natural language prompt " + }, + { + "bbox": [ + 46, + 403, + 287, + 475 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 403, + 287, + 475 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 135, + 480, + 198, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 480, + 198, + 493 + ], + "spans": [ + { + "bbox": [ + 135, + 480, + 198, + 493 + ], + "type": "interline_equation", + "content": "X = \\left[ F; T; P \\right]", + "image_path": "be385204e52f57a8778aff1703d9ba15b52a01de937415b7869cdf33819a1e72.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "type": "text", + "content": "To extract the image features, we consider the pretrained visual transformer (ViT) [11] in CLIP [46]. The ViT takes in the image patches, and extracts features for each image patch. Then we use a trainable linear layer to project the extracted image features into the input embedding space of LLM. This architecture, i.e. linear layer for visual feature alignment, has been shown effective by recent multimodal large models like LLaVA [36]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 594, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 689 + ], + "type": "text", + "content": "The predicted data table is critical to encode the textual information in the image. Because the pretrained ViT (in CLIP [46]) is not specifically trained for text-heavy images, the image features have limited OCR ability. Therefore, we use DEPLOT [34], which is designed for information extraction from charts, to predict the underlying data table for the given chart image. The linearized data table, which is a sequence of word tokens, is fed into the LLM." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Finally, the projected image features, the data table, and a prompt like \"Ask a question about the given chart plot and" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 308, + 545, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 308, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 545, + 332 + ], + "type": "text", + "content": "answer it\" (rephrased in several different ways), are provided as the input to LLM, for question generation." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 340, + 485, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 340, + 485, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 340, + 485, + 353 + ], + "type": "text", + "content": "3.2. Straight-forward data generation" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 359, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 359, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 359, + 545, + 407 + ], + "type": "text", + "content": "Given the above inputs, a straightforward manner for data generation is to have the LLM generate question and answer directly, as shown in Fig. 2 (b). The output " + }, + { + "bbox": [ + 305, + 359, + 545, + 407 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 305, + 359, + 545, + 407 + ], + "type": "text", + "content": " is the generated question " + }, + { + "bbox": [ + 305, + 359, + 545, + 407 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 305, + 359, + 545, + 407 + ], + "type": "text", + "content": ", followed by the corresponding answer " + }, + { + "bbox": [ + 305, + 359, + 545, + 407 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 305, + 359, + 545, + 407 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 400, + 417, + 451, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 417, + 451, + 430 + ], + "spans": [ + { + "bbox": [ + 400, + 417, + 451, + 430 + ], + "type": "interline_equation", + "content": "Y = [ Q; A ]", + "image_path": "d529e73ea73453b423cdf037485e2c641a1977924a40d5cdc2c87d06e6e4c1d8.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 440, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 545, + 513 + ], + "type": "text", + "content": "This straightforward data generator can be trained using the existing chart VQA datasets, which contains chart images paired with question and answers. The model is trained with a next-token prediction loss. Note that the ViT and LLM are pretrained and kept frozen during training. The only trainable parameters are the linear projection layer1." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 513, + 545, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 545, + 583 + ], + "type": "text", + "content": "As our experiments will show, this straightforward generation method yields reasonable results, generating questions and answers that are helpful for training the chart VQA models. However, the generated answers are noisy, which is improved using step-by-step generation, described next." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 594, + 437, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 594, + 437, + 606 + ], + "spans": [ + { + "bbox": [ + 306, + 594, + 437, + 606 + ], + "type": "text", + "content": "3.3. Synthesize step-by-step" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 612, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 612, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 545, + 647 + ], + "type": "text", + "content": "Here we introduce Synthesize Step-by-Step, which shares the same model architecture as straight-forward generation, but performs data synthesizing in a step-by-step manner." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 648, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 648, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 648, + 545, + 685 + ], + "type": "text", + "content": "Instead of having the model generate answers directly, we train the LLM to generate them in a step-by-step manner. The LLM generates the question " + }, + { + "bbox": [ + 304, + 648, + 545, + 685 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 648, + 545, + 685 + ], + "type": "text", + "content": " with the corresponding" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "1We also tried a two-stage training procedure, i.e. finetune the whole model after training the projection layer, but did not see performance gain." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13615" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 100, + 64, + 495, + 213 + ], + "blocks": [ + { + "bbox": [ + 100, + 64, + 495, + 213 + ], + "lines": [ + { + "bbox": [ + 100, + 64, + 495, + 213 + ], + "spans": [ + { + "bbox": [ + 100, + 64, + 495, + 213 + ], + "type": "image", + "image_path": "21ba88fae6548e22fc6795757509dad0e68187b39a4a53bfe5e086d0243165df.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 219, + 547, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 219, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 547, + 243 + ], + "type": "text", + "content": "Figure 3. Example of template-based question generation. Given a manually created template with defined program, by querying the image SVG annotations, templatic questions and answers can be generated, with rationales." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 257, + 288, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 288, + 282 + ], + "type": "text", + "content": "rationales " + }, + { + "bbox": [ + 46, + 257, + 288, + 282 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 257, + 288, + 282 + ], + "type": "text", + "content": ". Then the answer " + }, + { + "bbox": [ + 46, + 257, + 288, + 282 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 257, + 288, + 282 + ], + "type": "text", + "content": " is derived by executing the rationales using python scripts:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 293, + 230, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 293, + 230, + 308 + ], + "spans": [ + { + "bbox": [ + 104, + 293, + 230, + 308 + ], + "type": "interline_equation", + "content": "Y = [ Q; R ], A = \\mathrm {E x e c u t e} (R)", + "image_path": "2e46b94dc21fdcdc8ea5eae5f769194ebd707f7c8381838ce545c708dfa7d31c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 317, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 287, + 520 + ], + "type": "text", + "content": "As the example shows in Fig. 2 (c), the LLM outputs the generated question and a decomposition of this complex reasoning question into a series of atom reasoning steps, i.e. rationales. The rationales " + }, + { + "bbox": [ + 46, + 317, + 287, + 520 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 317, + 287, + 520 + ], + "type": "text", + "content": " a step-by-step executable program, where each step in the program can be either an atom extractive VQA questions (e.g. ans_0 = VQA (\"What is the value of 2002?\")) or a Python-like program call for math calculation, counting, comparison, etc. (e.g. ans = avg(ans_0, ans_1)). The rationale program is then parsed and executed using Python in order to derive the final answer. During execution, there are two ways to answer the atom extractive VQA questions: we can either use an off-the-shelf VQA model, or directly predict the answer using the LLM data generator itself. We experiment with both options and find the latter option to be more effective, as well as more simple, without requiring additional VQA models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "content": "Training with rationales. To train the step-by-step generator, data with rationales is required. This is different from the straightforward generation, which only requires question-answer pairs to train. Because existing datasets do not contain the rationale annotations for each question, we introduce a template-based QA generation pipeline, which will be described in Sec. 3.4, to generate synthetic questions with rationales. With the template-generated data as the training corpus, we tune the generator for step-by-step generation. An additional advantage of training with template-generated data is that, in inference time, the generator can be prompted with different prompts for controllable generation. The prompts may specify the format or the type of the questions, e.g. \"the question should start with how many\", \"the question should ask about the colors in the image\", etc. A complete list of the prompts can be found in Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 257, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 342 + ], + "type": "text", + "content": "Post hoc filtering. We take an additional post hoc filtering step to filter out the noisy questions. We use the decoding score (for the whole generated sequence including question and rationales) as an indicator and filter out the questions with decoding scores lower than a threshold. Better filtering strategies, like using LLMs for data filtering, can be explored, which we will leave as future work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 348, + 545, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 348, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 545, + 361 + ], + "type": "text", + "content": "3.4. Tuning corpus: template-based QA generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 366, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 426 + ], + "type": "text", + "content": "Here we describe how we use a template-based pipeline to generate questions with rationales for training the LLM-based generator. The method is motivated by the success of synthetic compositional questions in VQA datasets like CLEVR [21], Super-CLEVR [33] and GQA [20]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 426, + 546, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 546, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 546, + 617 + ], + "type": "text", + "content": "Fig. 3 illustrates the template-based question generation method. We first manually create question templates, then instantiate the templates into questions and answers by querying the rich SVG annotations of the chart image (as shown in the dotted boxes in the figure). For example, the template \"What is the average of the " + }, + { + "bbox": [ + 304, + 426, + 546, + 617 + ], + "type": "inline_equation", + "content": "\\langle N\\rangle ?" + }, + { + "bbox": [ + 304, + 426, + 546, + 617 + ], + "type": "text", + "content": " can be instantiated into \"what is the average of the more category?\". The argument " + }, + { + "bbox": [ + 304, + 426, + 546, + 617 + ], + "type": "inline_equation", + "content": "\\langle N\\rangle" + }, + { + "bbox": [ + 304, + 426, + 546, + 617 + ], + "type": "text", + "content": " is instantiated into \"the more category\" according to the SVG annotations, which includes the category names, values, and visual elements like colors and bounding boxes. Moreover, for each template, we define a reasoning program (in a domain-specific language) containing a series of reasoning operations (e.g. select_color, query_value) that specifies the reasoning process needed to answer the question. The program is then instantiated into a step-by-step rationale." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": "Note that there are limitations to this template-based QA generation pipeline. The pipeline relies on the rich annotations extracted automatically from the source SVG files, and thus cannot generalize to unlabeled images. Moreover, even though some existing datasets like ChartQA provide such rich annotations crawled from websources, they are inevitably noisy, containing missing or incorrect values. In contrast, the LLM generator is more generalizable, being" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13616" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": "able to generate data for images in the wild. Therefore, in our work, these template-based questions are used as the training corpus of the LLM-based generator, which is more generalizable and scalable." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 127, + 205, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 127, + 205, + 140 + ], + "spans": [ + { + "bbox": [ + 47, + 127, + 205, + 140 + ], + "type": "text", + "content": "3.5. Evaluation of generated data" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 145, + 287, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 287, + 252 + ], + "type": "text", + "content": "To measure the usefulness of the generated QA, we use the generated data to train the chart VQA models. The accuracy improvement of chart VQA models, after training with the generated QA, is taken as the indicator of the data quality. We select MATCHA [35], which is a transformer encoder-decoder model pretrained on large-scale text-heavy images, as our primary evaluation model, considering its modern architecture and state-of-the-art performance on standard datasets2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 253, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 288, + 373 + ], + "type": "text", + "content": "Summary. Our data generation procedure contains the following steps: first, we generate templatic QA corpus with rationales using the rich SVG annotations, which is then used to tune the LLM-based data generator; next, the LLM-based generator generates questions with step-by-step rationales; then we use Python to execute the rationales and derive the answers; finally, this LLM-augmented data, LAMENDA, is used to train the downstream VQA model MATCHA. As the experiments will show, the augmented data leads to significant performance improvements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 383, + 123, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 383, + 123, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 383, + 123, + 396 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 403, + 99, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 99, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 99, + 415 + ], + "type": "text", + "content": "4.1. Set up" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 421, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 287, + 649 + ], + "type": "text", + "content": "Dataset. Following [34, 35], we run experiments on two chart VQA datasets: ChartQA and PlotQA " + }, + { + "bbox": [ + 46, + 421, + 287, + 649 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 46, + 421, + 287, + 649 + ], + "type": "text", + "content": ". ChartQA [40] contains 21k real-world chart images crawled from four web sources, with 9k human-annotated (Human split) and 23k machine-augmented (Augmented split) QA pairs. The augmented questions are mostly extractive questions, while the human-annotated questions are free-form and require complex reasoning. PlotQA [41] is a large synthetic dataset containing 224k chart images generated from real-world online data, with 37M synthetic questions and answers divided into two splits: V1 (8M) and V2 (29M). In addition, to show the generalizability of our LLM data generator, we generate QA for images from additional chart captioning datasets, which do not provide QAs or rich SVG annotations. We use images from three datasets: Chart-to-Text [24], VisText [52] and ChartSumm [47]. Chart-to-Text contains 35k images from \"Statistica\" and 9k images from \"Pew\"; VisText contains 9k synthetic charts of various styles; ChartSumm contains 84k images from \"Statistica\"" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 46, + 654, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 684 + ], + "type": "text", + "content": "2While there are two-stage models of better performance, which extracts information from charts then prompts LLMs like Codex to answer questions, MATCHA is the best single-stage model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 684, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 684, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 684, + 287, + 713 + ], + "type": "text", + "content": "3There are other chart VQA datasets like FigureQA, DVQA, etc., but since they are out-dated and models reach " + }, + { + "bbox": [ + 46, + 684, + 287, + 713 + ], + "type": "inline_equation", + "content": ">95\\%" + }, + { + "bbox": [ + 46, + 684, + 287, + 713 + ], + "type": "text", + "content": " accuracy, we focus on more challenging settings." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": "and \"Knoema\". Note that we only use the images from the chart captioning datasets without using the captions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 96, + 546, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 546, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 546, + 276 + ], + "type": "text", + "content": "Generated data. Tab. 1 shows the statistics of the generated QA. To construct the template-based tuning corpus, we use 28 templates to generate " + }, + { + "bbox": [ + 304, + 96, + 546, + 276 + ], + "type": "inline_equation", + "content": "357\\mathrm{k}" + }, + { + "bbox": [ + 304, + 96, + 546, + 276 + ], + "type": "text", + "content": " question-answer pairs for the ChartQA training images. The template list and the template QA can be found in the Appendix. We train our LLM data generator using the template-based corpus, and generate QAs for images in ChartQA, PlotQA, and the three chart captioning datasets. For chartQA training images, we generate " + }, + { + "bbox": [ + 304, + 96, + 546, + 276 + ], + "type": "inline_equation", + "content": "403\\mathrm{k}" + }, + { + "bbox": [ + 304, + 96, + 546, + 276 + ], + "type": "text", + "content": " QAs, which reduces to " + }, + { + "bbox": [ + 304, + 96, + 546, + 276 + ], + "type": "inline_equation", + "content": "326\\mathrm{k}" + }, + { + "bbox": [ + 304, + 96, + 546, + 276 + ], + "type": "text", + "content": " after step-by-step execution (dropping ones that cannot be executed) and filtering based on the decoding scores using a threshold of -10. For PlotQA training images, we generate 3M QAs, which reduces to " + }, + { + "bbox": [ + 304, + 96, + 546, + 276 + ], + "type": "inline_equation", + "content": "1.7\\mathrm{M}" + }, + { + "bbox": [ + 304, + 96, + 546, + 276 + ], + "type": "text", + "content": " after execution and filtering. For the " + }, + { + "bbox": [ + 304, + 96, + 546, + 276 + ], + "type": "inline_equation", + "content": "137\\mathrm{k}" + }, + { + "bbox": [ + 304, + 96, + 546, + 276 + ], + "type": "text", + "content": " chart images in the chart captioning datasets, we generate " + }, + { + "bbox": [ + 304, + 96, + 546, + 276 + ], + "type": "inline_equation", + "content": "1.6\\mathrm{M}" + }, + { + "bbox": [ + 304, + 96, + 546, + 276 + ], + "type": "text", + "content": " QAs." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 309, + 280, + 545, + 345 + ], + "blocks": [ + { + "bbox": [ + 309, + 280, + 545, + 345 + ], + "lines": [ + { + "bbox": [ + 309, + 280, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 309, + 280, + 545, + 345 + ], + "type": "table", + "html": "
DatasetImagesGeneratedFiltered
TemplateChartQA18,317356,606-
LAMENDAChartQA18,317402,974326,160
PlotQA157,0703,455,5391,726,822
ChartCap137,2812,987,7181,635,364
", + "image_path": "7987d56c136995a3b661d8dba5a735174134c110bc29252ed6a5a346aa4ac967.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 348, + 538, + 360 + ], + "lines": [ + { + "bbox": [ + 312, + 348, + 538, + 360 + ], + "spans": [ + { + "bbox": [ + 312, + 348, + 538, + 360 + ], + "type": "text", + "content": "Table 1. Statistics of generated data, before and after filtering." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 361, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 624 + ], + "type": "text", + "content": "Implementation details. Using the 357k template QA, we tune the data generator for 5 epochs with a batch size of 128. We use a cosine learning rate schedule with a max learning rate of 2e-3. The training takes about 12 hours using 8 A100 GPUs. The trainable projection layer is initialized with the first stage checkpoint of LLaVA [36]. We select the best checkpoint at 14k iterations as our best generator. For MATCHA training, we take the pretrained checkpoints (base size) from HuggingFace and finetune the models with a combination of our generated questions and the original ChartQA or PlotQA training questions. The training setting follows [28, 35]. We use a cosine scheduler with 500 warm-up steps and a max learning rate of 0.0001. The batch size is 256. Without special description, the models are trained for 10k iterations for ChartQA (combining human and augmented splits) and 20k iterations for PlotQA because it is much larger (V1 and V2 are trained separately). We set the max number of image patches as 4096. The ChartQA training takes 3 days on 8 32GB V100 GPUs. Considering the training time, in some experiments, we reduce the image patch number to 1024 for quick experiments in 12 hours with the same setting." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 624, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 696 + ], + "type": "text", + "content": "Evaluation metrics. We suggest that both strict accuracy and relaxed accuracy should be reported, especially for the ChartQA dataset. While existing works [34, 35, 40, 41] evaluate answers are evaluated using relaxed accuracy, which allows " + }, + { + "bbox": [ + 304, + 624, + 545, + 696 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 624, + 545, + 696 + ], + "type": "text", + "content": " error for numerical answers, we find that this relaxed accuracy metric is problematic for ChartQA." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 520, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 520, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 520, + 712 + ], + "type": "text", + "content": "4https://huggingface.co/google/matcha-base" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13617" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 87, + 65, + 511, + 190 + ], + "blocks": [ + { + "bbox": [ + 87, + 65, + 511, + 190 + ], + "lines": [ + { + "bbox": [ + 87, + 65, + 511, + 190 + ], + "spans": [ + { + "bbox": [ + 87, + 65, + 511, + 190 + ], + "type": "table", + "html": "
AccuracyRelaxed Accuracy
avghumanaugmentavghumanaugment
0baseline (MATCHA [35])47.7630.2165.3158.5439.5877.50
1syna48.8028.6568.9659.4338.6580.21
2synat w/ VQA answer47.7628.4467.0857.9737.8178.12
3synat w/ step-VQA answer52.5537.2967.8164.3848.8579.90
4synat w/ synat answer53.7537.9269.5865.1048.8581.35
5synat w/ step-synat answer55.2140.2170.2166.4151.3581.46
6synat w/ step-synat answer + tpl.57.3442.6072.0867.8653.1282.60
7synat w/ step-synat answer + tpl. + addi.58.6545.6271.6769.8456.5683.12
", + "image_path": "e68ba9bca98d05b7dd395a34a448fb324ac34be0531b483d770ed28e713e4f76.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "text", + "content": "Table 2. Main results on ChartQA val split. The data generator trained on Annotated questions is denoted as " + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "inline_equation", + "content": "\\mathrm{syn}_a" + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "text", + "content": "; the data generator trained on both Annotated questions and our Template-generated questions " + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "inline_equation", + "content": "\\mathrm{syn}_{at}" + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "inline_equation", + "content": "\\mathrm{syn}_{at}" + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "text", + "content": ". Step-by-step generation is better than straight-forward generation (row 1-5). Training with a combination of step-by-step generated QA, template-generated QA (+mpl.), and additional images from chart captioning datasets (+addi.), leads to the best model. See Sec. 4.2 for more details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 248, + 288, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 288, + 285 + ], + "type": "text", + "content": "For example, many questions in ChartQA ask about years (e.g. 1996, 2012), for which allowing " + }, + { + "bbox": [ + 46, + 248, + 288, + 285 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 248, + 288, + 285 + ], + "type": "text", + "content": " numerical errors (i.e. around 100 years) will result in incorrectly high results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 295, + 129, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 295, + 129, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 295, + 129, + 306 + ], + "type": "text", + "content": "4.2. Main results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 315, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 287, + 387 + ], + "type": "text", + "content": "We study our data generation method with different settings on the ChartQA dataset, with MATCHA as our baseline. We first generate QA data using our method, and then use the generated data to train the MATCHA model. The strict accuracy and relaxed accuracy reported on the ChartQA validation split are shown in Tab. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 388, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 519 + ], + "type": "text", + "content": "The baseline model MATCHA (row-0) achieves " + }, + { + "bbox": [ + 46, + 388, + 287, + 519 + ], + "type": "inline_equation", + "content": "47.76\\%" + }, + { + "bbox": [ + 46, + 388, + 287, + 519 + ], + "type": "text", + "content": " accuracy and " + }, + { + "bbox": [ + 46, + 388, + 287, + 519 + ], + "type": "inline_equation", + "content": "58.54\\%" + }, + { + "bbox": [ + 46, + 388, + 287, + 519 + ], + "type": "text", + "content": " relaxed accuracy. For this baseline, we finetune the pretrained MATCHA checkpoint for 10k iterations with the ChartQA training split (combing human and augmented questions), using 1024 image patches. This setting follows [35], except that a smaller number of image patches are used for faster experiments and less GPU memory requirement. We validate the correctness of our implementation by showing that the model reaches " + }, + { + "bbox": [ + 46, + 388, + 287, + 519 + ], + "type": "inline_equation", + "content": "63.7\\%" + }, + { + "bbox": [ + 46, + 388, + 287, + 519 + ], + "type": "text", + "content": " relaxed accuracy using the original resolution, which matches the numbers reported in [35] (as shown in Tab. 3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "content": "Row-1 shows the results for the straight-forward data generation, where we train the data generator with QAs from the ChartQA training split and use the trained generator to directly generate questions and answers. This straightforward data generation improves the accuracy by " + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "inline_equation", + "content": "47.76\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "inline_equation", + "content": "48.80\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "content": "), suggesting that LLMs can be helpful in data augmentation for chart reasoning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Row-2 to row-5 shows the advantage of synthesizing step-by-step over straight-forward generation. We train the data generator with our template-generated QA, but the answers are generated with different methods. In row-2, we feed in the generated complex question into an additional VQA model (i.e. baseline MATCHA) to get the answers; in row-3, we generate the step-by-step rationales, and the stepwise questions are fed into the VQA model to get step-wise answers, then the final answer is derived using the step" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 248, + 547, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 547, + 392 + ], + "type": "text", + "content": "wise answers. As can be seen, the step-wise answer generation (52.55%) is significantly better than the direct answer generation (47.76%), which suggests the importance of decomposing the complex questions into atom sub-questions. Row-4 and row-5 follow a similar setting, except that we use the LLM data generator to generate the answers instead of the additional VQA model. Two findings can be drawn: (a) LLM-generated answers are better than VQA-predicted answers, possibly because the LLM data generator has access to the predicted data tables as inputs; (b) step-by-step synthesizing outperforms direct generation, for both VQA-predicted and LLM-generated answers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 394, + 547, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 547, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 547, + 574 + ], + "type": "text", + "content": "In row-6, we combine the step-by-step generated data with the templated-generated QAs and show that training with both leads to further better performance (57.34%). In row-7, we generate QAs for the chart images from additional chart captioning datasets. The final model, trained with a combination of all the synthetic data, gets 58.65% accuracy. Compared with the baseline, training with the augmented data leads to a 10.89% accuracy improvement. Moreover, the improvement is more significant on the human-written questions (+15.41%), which requires complex reasoning, than the augmented extractive questions (+6.36%). The results suggest that our synthetic data improves both the extractive ability and the reasoning ability, with the latter being more significant. We show examples in the Appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 586, + 440, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 440, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 440, + 598 + ], + "type": "text", + "content": "4.3. Comparison with SoTA" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "To compare our method with the best performing models, we apply our generated data onto two models, Pix2Struct [28] and MATCHA [35], and report the results on the ChartQA test split in Tab. 3. For faster training, we first train the model (for both Pix2struct and MATCHA) with a max of 1024 image patches using our best setting (i.e. row-7 in Tab. 2), then finetune with 4096 image patches to get the final result. We report both the 1024-resolution results and the 4096-resolution results." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13618" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 92, + 60, + 504, + 228 + ], + "blocks": [ + { + "bbox": [ + 92, + 60, + 504, + 228 + ], + "lines": [ + { + "bbox": [ + 92, + 60, + 504, + 228 + ], + "spans": [ + { + "bbox": [ + 92, + 60, + 504, + 228 + ], + "type": "table", + "html": "
paramsaccuracyRelaxed Accuracy
avghumanaugmentavghumanaugment
VL-T5-OCR [40]----41.56--
VisionTapas-OCR [40]----45.52--
Pix2Struct4096 [28]282M---56.030.581.6
MATCHA4096 [35]282M---64.238.290.2
DEPLOT+FlanPaLM+Codex [34]540B---79.367.691.0
PaLI-X [7]55B---72.3--
Pix2Struct4096 (reimpl.)282M50.4424.6476.2458.2432.0084.48
Pix2Struct1024 + LAMENDA282M51.7734.6968.8563.0246.2579.79
Pix2Struct4096 + LAMENDA282M55.8340.1071.5666.8251.6781.98
MATCHA4096 (reimpl.)282M55.6831.0480.3263.7237.7689.68
MATCHA1024 + LAMENDA282M60.2841.9278.6469.8851.7688.00
MATCHA4096 + LAMENDA282M63.3644.8881.8472.6453.9291.36
", + "image_path": "f8fcd1e1be9f068e4cc8199e870c95e59fb6e44dd176642ef15f0a83d9a29b1a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 265, + 288, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 288, + 396 + ], + "type": "text", + "content": "As can be seen, our generated data improves the overall relaxed accuracy of Pix2Struct by " + }, + { + "bbox": [ + 46, + 265, + 288, + 396 + ], + "type": "inline_equation", + "content": "8.58\\%" + }, + { + "bbox": [ + 46, + 265, + 288, + 396 + ], + "type": "text", + "content": " and MATCHA by " + }, + { + "bbox": [ + 46, + 265, + 288, + 396 + ], + "type": "inline_equation", + "content": "8.92\\%" + }, + { + "bbox": [ + 46, + 265, + 288, + 396 + ], + "type": "text", + "content": ". Moreover, the improvement is much more significant on human-written questions, e.g. the relaxed accuracy of MATCHA on human-written questions improves from " + }, + { + "bbox": [ + 46, + 265, + 288, + 396 + ], + "type": "inline_equation", + "content": "38\\%" + }, + { + "bbox": [ + 46, + 265, + 288, + 396 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 265, + 288, + 396 + ], + "type": "inline_equation", + "content": "54\\%" + }, + { + "bbox": [ + 46, + 265, + 288, + 396 + ], + "type": "text", + "content": ". Interestingly, with our augmented data, the low-resolution (1024) models can already outperform their high-resolution (4096) counterparts without data augmentation. The significant performance gain indicates that learning with synthetic data is extremely helpful for better reasoning ability with human-written queries." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 396, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 504 + ], + "type": "text", + "content": "Our best performing MATCHA achieves " + }, + { + "bbox": [ + 46, + 396, + 287, + 504 + ], + "type": "inline_equation", + "content": "72.6\\%" + }, + { + "bbox": [ + 46, + 396, + 287, + 504 + ], + "type": "text", + "content": " relaxed accuracy, which beats all existing end-to-end chart reasoning models. Note that DEPLOT+FlanPaLM+Codex is a two-stage model that first extracts information from the charts, then prompts external LLMs like FlanPalm (540B) Codex (12B) for question answering. The method relies on much larger models of billions of parameters and thus is not comparable with our method, which is much smaller and not rely on external LLMs for inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 513, + 139, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 139, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 139, + 525 + ], + "type": "text", + "content": "4.4. Generalization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 531, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 662 + ], + "type": "text", + "content": "To show the generalization ability of the proposed data generation method, we run experiments on the PlotQA dataset. We directly use our data generator (trained on template questions on ChartQA) to generate QA for the PlotQA training images. The results are shown in Tab. 4. As can be seen, our method, when applied to MATCHA, achieves state-of-the-art performance (92.89% relaxed accuracy) on PlotQA. Despite the synthetic nature of this dataset and the saturated model performance, our augmented data can still help on both low-resolution (+3.46%) and high-resolution (+0.92%) settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 670, + 175, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 175, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 175, + 683 + ], + "type": "text", + "content": "4.5. Ablations and analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Qualitative analysis and failure cases. Fig. 5 shows examples of our LLM-generated QA, including good and bad ex" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 265, + 546, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 546, + 565 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 546, + 565 + ], + "type": "text", + "content": "samples. Most generated questions are specific to the given image, asking about the categories plotted, such as “2019” or “Tablet 2018”, indicating that the model is able to utilize the information of the input. Moreover, various questions are generated, including counting, colors, average, etc. In the failure example, the question requires comparing the “silver bar” with the “midnightblue” bar, with reasonable rationales. However, there are multiple bars in the specified colors in the image, resulting in ambiguous/invalid questions. When looking at more failure cases, we find that there are several typical failure modes. For example, questions are sometimes invalid/ambiguous (e.g. asking about the top/bottom bar for a vertical bar plot); questions can be hallucinated or irrelevant (e.g. asking about a category that is not shown in the figure); rationales and answers are not correct, etc., which opens up opportunities for future works. Before and after LAMENDA. To gain insights on how LAMENDA helps model training, we conduct a comparison analysis of MATCHA “before” and “after” training with our LLM-generated data, i.e. the baseline model versus our best model. According to the predictions of these two models, we divide the 960 human-written questions from ChartQA val split into 4 groups: both model answers correctly, both models answers wrongly, the before model answers correctly, but the after model answers wrongly, and" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 314, + 570, + 541, + 699 + ], + "blocks": [ + { + "bbox": [ + 46, + 231, + 546, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 546, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 546, + 255 + ], + "type": "text", + "content": "Table 3. Comparison with SoTAs on ChartQA test split. Subscripts (1024, 4096) means image token numbers (resolution). With our generated data, both Pix2Struct and MATCHA achieve significantly better performance, even comparable to much larger models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 570, + 541, + 699 + ], + "lines": [ + { + "bbox": [ + 314, + 570, + 541, + 699 + ], + "spans": [ + { + "bbox": [ + 314, + 570, + 541, + 699 + ], + "type": "table", + "html": "
Relaxed Accuracy
avgV1V2
VisionTapas-OCR [40]53.9065.3042.50
VL-T5-OCR [40]65.9675.9056.02
DEPLOT+FlanPaLM+Codex [34]66.662.271.0
Pix2Struct [28]72.573.271.9
MATCHA4096 [35]91.592.390.7
MATCHA1024 (reimpl.)74.9573.8876.02
MATCHA1024 + LAMENDA78.4178.4478.38
MATCHA4096 (reimpl.)91.9792.6491.30
MATCHA4096 + LAMENDA92.8993.9491.84
", + "image_path": "0e84780e527b9d0462e722f1750db71eb9e6b731e6e9cf2f962d4fc81bbb3486.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 703, + 545, + 725 + ], + "lines": [ + { + "bbox": [ + 305, + 703, + 545, + 725 + ], + "spans": [ + { + "bbox": [ + 305, + 703, + 545, + 725 + ], + "type": "text", + "content": "Table 4. Comparison with SoTAs on PlotQA test split. With our generated data, MATCHA achieves the SoTA performance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13619" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 68, + 492, + 175 + ], + "blocks": [ + { + "bbox": [ + 102, + 68, + 492, + 175 + ], + "lines": [ + { + "bbox": [ + 102, + 68, + 492, + 175 + ], + "spans": [ + { + "bbox": [ + 102, + 68, + 492, + 175 + ], + "type": "image", + "image_path": "2ac7fdad08d09825866a1179e3b4f14522bf0b5f0bfdc1c6bfe0db7d70be31ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 181, + 547, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 181, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 547, + 216 + ], + "type": "text", + "content": "Figure 4. Analysis of model predictions before and after training with LAMENDA. In I (left), ChartQA val questions are divided into four categories according to the model prediction correctness (before/after). II (right) further decomposes each category into different question types. LAMENDA improves the model prediction, and the improvement is most significant on reasoning questions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "text", + "content": "vice versa. The statistics of these four groups are shown in Fig. 4. From Fig. 4 I. (left side), we see that LAMENDA is helpful for " + }, + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "text", + "content": " of the questions, while there are still " + }, + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "text", + "content": " of the questions where both models cannot answer correctly. In Fig. 4 II. (right side), we randomly sample 30 questions from each of the four groups and manually classify them into six question types: extractive, count, spatial, color, reasoning, and others. As shown in (a), i.e. the questions where the \"before\" model answers wrongly and our model answers correctly, most " + }, + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "inline_equation", + "content": "(70\\%)" + }, + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "text", + "content": " of the questions belong to the reasoning category, indicating that our data helps most on the reasoning questions. Our data also helps with color questions " + }, + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "inline_equation", + "content": "(10\\%)" + }, + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "text", + "content": " and count questions " + }, + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "inline_equation", + "content": "(17\\%)" + }, + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "text", + "content": ". However, our data does not help with extractive questions " + }, + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "inline_equation", + "content": "(0\\%)" + }, + { + "bbox": [ + 46, + 226, + 289, + 430 + ], + "type": "text", + "content": ". In addition, comparing the distribution of (c) both correct with (d) both wrong, we can see that extractive questions are easy for the models, while reasoning questions are very challenging." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 430, + 288, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 288, + 479 + ], + "type": "text", + "content": "In summary, Fig. 4 shows the effectiveness of our data for the reasoning questions over other question types. However, reasoning questions are still a big challenge for current models, which requires further exploration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 480, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 480, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 287, + 493 + ], + "type": "text", + "content": "Different question types. In Tab. 5, we study the improve" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 85, + 506, + 249, + 692 + ], + "blocks": [ + { + "bbox": [ + 85, + 506, + 249, + 692 + ], + "lines": [ + { + "bbox": [ + 85, + 506, + 249, + 692 + ], + "spans": [ + { + "bbox": [ + 85, + 506, + 249, + 692 + ], + "type": "image", + "image_path": "709dbcf39525bede908916656b06832d45d92e989ca0be34efb3600a1540b0ac.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 693, + 287, + 715 + ], + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 715 + ], + "type": "text", + "content": "Figure 5. Examples of generated QA, 3 good examples and 1 failure case. Rationales are only shown for 2 examples for brevity." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 226, + 547, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 226, + 547, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 226, + 547, + 382 + ], + "type": "text", + "content": "ments brought by each type of the generated questions. The total 326k generated QAs are grouped into 7 different groups, including questions about color understanding, spatial reasoning, counting, minimum/maximum reasoning, average, comparison (e.g. larger than, smaller than), and other math calculations. This grouping is done by different prompting of the data generation LLM, as specified in the Appendix. The improvements from different question types are distinct and can be accumulated. As we see, each of the seven types improves the model performance by less than " + }, + { + "bbox": [ + 304, + 226, + 547, + 382 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 304, + 226, + 547, + 382 + ], + "type": "text", + "content": ", while combining all of them leads to " + }, + { + "bbox": [ + 304, + 226, + 547, + 382 + ], + "type": "inline_equation", + "content": "7.03\\%" + }, + { + "bbox": [ + 304, + 226, + 547, + 382 + ], + "type": "text", + "content": " improvement. Moreover, among the seven types, the \"math calculation\" questions brings the most improvements " + }, + { + "bbox": [ + 304, + 226, + 547, + 382 + ], + "type": "inline_equation", + "content": "(2.76\\%)" + }, + { + "bbox": [ + 304, + 226, + 547, + 382 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 343, + 388, + 511, + 488 + ], + "blocks": [ + { + "bbox": [ + 343, + 388, + 511, + 488 + ], + "lines": [ + { + "bbox": [ + 343, + 388, + 511, + 488 + ], + "spans": [ + { + "bbox": [ + 343, + 388, + 511, + 488 + ], + "type": "table", + "html": "
#QAavghumanaug
baseline28,29947.7630.2165.31
+colors+46,51247.8129.9065.73
+spatial+49,38749.2730.6367.92
+count+36,70547.4029.1765.63
+minmax+71,78849.5831.8767.29
+average+50,71748.9631.3566.56
+compare+14,69047.6628.8566.46
+calculation+56,36150.5233.8567.19
+all326,16054.7939.2770.31
", + "image_path": "aac68200a983c4d629bee3a6f603efb4ec0f6b8eb071b81d79948471e4aa9e75.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 492, + 547, + 526 + ], + "lines": [ + { + "bbox": [ + 305, + 492, + 547, + 526 + ], + "spans": [ + { + "bbox": [ + 305, + 492, + 547, + 526 + ], + "type": "text", + "content": "Table 5. A detailed look at the effect for different question types. Strict accuracy on ChartQA val split is shown. Each type helps and combining all of them leads to a further performance gain." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 537, + 379, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 379, + 549 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 379, + 549 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": "We propose Synthesize Step-by-Step, which utilizes LLMs to augment the training data for the chart reasoning task. We show that the step-by-step generation, which decomposes the complex reasoning questions into easier sub-questions, is critical for synthesizing good-quality data. We run experiments on two standard benchmarks and show significant improvements by training with the augmented data. Specifically, the relaxed accuracy on the human-written questions is improved from " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "37.8\\%" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "53.9\\%" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " on ChartQA. The large improvement indicates the promise of boosting the model's reasoning ability by data generation using LLMs, from which the strong step-by-step reasoning ability can be distilled into downstream models." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13620" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 705 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Mubashara Akhtar, Oana Cocarascu, and Elena Simperl. Reading and reasoning over chart images for evidence-based automated fact-checking. arXiv preprint arXiv:2301.11843, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 134, + 288, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 134, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 53, + 134, + 288, + 198 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 198, + 288, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 288, + 242 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 288, + 242 + ], + "type": "text", + "content": "[3] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 39-48, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 242, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 242, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 53, + 242, + 288, + 284 + ], + "type": "text", + "content": "[4] Srikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran, Yichu Zhou, and R Manmatha. Docformerv2: Local features for document understanding. arXiv preprint arXiv:2306.01733, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 288, + 338 + ], + "type": "text", + "content": "[5] Ritwick Chaudhry, Sumit Shekhar, Utkarsh Gupta, Pranav Maneriker, Prann Bansal, and Ajay Joshi. Leaf-qa: Locate, encode & attend for figure question answering. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3512–3521, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 339, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 339, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 53, + 339, + 288, + 381 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 382, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 382, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 382, + 288, + 435 + ], + "type": "text", + "content": "[7] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 435, + 288, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 288, + 489 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 288, + 489 + ], + "type": "text", + "content": "[8] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 489, + 288, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 489, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 489, + 288, + 532 + ], + "type": "text", + "content": "[9] John Joon Young Chung, Ece Kamar, and Saleema Amershi. Increasing diversity while maintaining accuracy: Text data generation with large language models and human interventions. arXiv preprint arXiv:2306.04140, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 532, + 288, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 288, + 584 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 288, + 584 + ], + "type": "text", + "content": "[10] Brian Davis, Bryan Morse, Brian Price, Chris Tensmeyer, Curtis Wigington, and Vlad Morariu. End-to-end document recognition and understanding with dessurt. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 585, + 288, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 585, + 288, + 650 + ], + "spans": [ + { + "bbox": [ + 48, + 585, + 288, + 650 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 650, + 288, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 650, + 288, + 705 + ], + "spans": [ + { + "bbox": [ + 48, + 650, + 288, + 705 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 707 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 380, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 380, + 83 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 380, + 83 + ], + "type": "text", + "content": "scale, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 83, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 83, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 308, + 83, + 545, + 137 + ], + "type": "text", + "content": "[13] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 137, + 545, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 137, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 308, + 137, + 545, + 169 + ], + "type": "text", + "content": "[14] Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. Knowledge distillation of large language models. arXiv preprint arXiv:2306.08543, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 169, + 545, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 169, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 308, + 169, + 545, + 201 + ], + "type": "text", + "content": "[15] Nitish Gupta, Kevin Lin, Dan Roth, Sameer Singh, and Matt Gardner. Neural module networks for reasoning over text. arXiv preprint arXiv:1912.04971, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 201, + 545, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 201, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 308, + 201, + 545, + 244 + ], + "type": "text", + "content": "[16] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 244, + 545, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 244, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 308, + 244, + 545, + 276 + ], + "type": "text", + "content": "[17] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 276, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 340 + ], + "type": "text", + "content": "[18] Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes. arXiv preprint arXiv:2305.02301, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 340, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 340, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 340, + 545, + 384 + ], + "type": "text", + "content": "[19] Ronghang Hu, Jacob Andreas, Trevor Darrell, and Kate Saenko. Explainable neural computation via stack neural module networks. In Proceedings of the European conference on computer vision (ECCV), pages 53-69, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 384, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 545, + 437 + ], + "type": "text", + "content": "[20] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 437, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 502 + ], + "type": "text", + "content": "[21] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 502, + 545, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 502, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 308, + 502, + 545, + 534 + ], + "type": "text", + "content": "[22] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 534, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 576 + ], + "type": "text", + "content": "[23] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Akos Kadar, Adam Trischler, and Yoshua Bengio. Figureqa: An annotated figure dataset for visual reasoning, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 577, + 545, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 577, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 308, + 577, + 545, + 642 + ], + "type": "text", + "content": "[24] Shankar Kantharaj, Rixie Tiffany Leong, Xiang Lin, Ahmed Masry, Megh Thakkar, Enamul Hoque, and Shafiq Joty. Chart-to-text: A large-scale benchmark for chart summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4005–4023, 2022. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 642, + 545, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 642, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 308, + 642, + 545, + 685 + ], + "type": "text", + "content": "[25] Tushar Khot, Harsh Trivedi, Matthew Finlayson, Yao Fu, Kyle Richardson, Peter Clark, and Ashish Sabharwal. Decomposed prompting: A modular approach for solving complex tasks. arXiv preprint arXiv:2210.02406, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 685, + 545, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 685, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 308, + 685, + 545, + 707 + ], + "type": "text", + "content": "[26] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13621" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 708 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 115 + ], + "type": "text", + "content": "doo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 116, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 116, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 116, + 287, + 159 + ], + "type": "text", + "content": "[27] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 160, + 287, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 160, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 48, + 160, + 287, + 233 + ], + "type": "text", + "content": "[28] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In International Conference on Machine Learning, pages 18893–18912. PMLR, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 235, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 235, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 235, + 287, + 277 + ], + "type": "text", + "content": "[29] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 277, + 287, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 320 + ], + "type": "text", + "content": "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 321, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 363 + ], + "type": "text", + "content": "[31] Liunian Harold Li, Jack Hessel, Youngjae Yu, Xiang Ren, Kai-Wei Chang, and Yejin Choi. Symbolic chain-of-thought distillation: Small models can also\" think\" step-by-step. arXiv preprint arXiv:2306.14050, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 364, + 287, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 364, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 48, + 364, + 287, + 428 + ], + "type": "text", + "content": "[32] Zhuowan Li, Elias Stengel-Eskin, Yixiao Zhang, Cihang Xie, Quan Hung Tran, Benjamin Van Durme, and Alan Yuille. Calibrating concepts and operations: Towards symbolic reasoning on real images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14910-14919, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 429, + 287, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 429, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 48, + 429, + 287, + 493 + ], + "type": "text", + "content": "[33] Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Kortylewski, Wufei Ma, Benjamin Van Durme, and Alan L Yuille. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14963-14973, 2023. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 494, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 494, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 494, + 287, + 547 + ], + "type": "text", + "content": "[34] Fangyu Liu, Julian Martin Eisenschlos, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Wenhu Chen, Nigel Collier, and Yasemin Altun. Deplot: One-shot visual language reasoning by plot-to-table translation. arXiv preprint arXiv:2212.10505, 2022. 2, 3, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 547, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 610 + ], + "type": "text", + "content": "[35] Fangyu Liu, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Yasemin Altun, Nigel Collier, and Julian Martin Eisenschlos. Matcha: Enhancing visual language pretraining with math reasoning and chart derendering. arXiv preprint arXiv:2212.09662, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 612, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 643 + ], + "type": "text", + "content": "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 644, + 287, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 644, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 48, + 644, + 287, + 696 + ], + "type": "text", + "content": "[37] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 697, + 287, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 697, + 287, + 708 + ], + "spans": [ + { + "bbox": [ + 48, + 697, + 287, + 708 + ], + "type": "text", + "content": "[38] Junyu Luo, Zekun Li, Jinpeng Wang, and Chin-Yew Lin." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 707 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "Chartocr: Data extraction from charts images via a deep hybrid framework. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1917-1925, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 117, + 545, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 117, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 307, + 117, + 545, + 159 + ], + "type": "text", + "content": "[39] Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B Tenenbaum, and Jiajun Wu. The neuro-symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision. arXiv preprint arXiv:1904.12584, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 160, + 545, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 160, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 307, + 160, + 545, + 202 + ], + "type": "text", + "content": "[40] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 1, 2, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 202, + 545, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 202, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 307, + 202, + 545, + 245 + ], + "type": "text", + "content": "[41] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2020. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 246, + 545, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 246, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 307, + 246, + 545, + 276 + ], + "type": "text", + "content": "[42] Aaron Parisi, Yao Zhao, and Noah Fiedel. Talm: Tool augmented language models. arXiv preprint arXiv:2205.12255, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 277, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 308 + ], + "type": "text", + "content": "[43] Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 309, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 384 + ], + "type": "text", + "content": "[44] Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). Association for Computational Linguistics, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 384, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 384, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 384, + 545, + 437 + ], + "type": "text", + "content": "[45] Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, et al. Toolllm: Facilitating large language models to master 16000+ real-world apis. arXiv preprint arXiv:2307.16789, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 438, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 438, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 438, + 545, + 502 + ], + "type": "text", + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 503, + 545, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 503, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 307, + 503, + 545, + 556 + ], + "type": "text", + "content": "[47] Raian Rahman, Rizvi Hasan, Abdullah Al Farhad, Md Tahmid Rahman Laskar, Md Hamjajul Ashmafee, and Abu Raihan Mostofa Kamal. Chartsumm: A comprehensive benchmark for automatic chart summarization of long and short summaries. arXiv preprint arXiv:2304.13620, 2023. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 557, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 609 + ], + "type": "text", + "content": "[48] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. arXiv preprint arXiv:2302.04761, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 610, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 610, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 307, + 610, + 545, + 654 + ], + "type": "text", + "content": "[49] Timo Schick and Hinrich Schütze. Generating datasets with pretrained language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6943-6951, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 654, + 545, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 654, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 307, + 654, + 545, + 696 + ], + "type": "text", + "content": "[50] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 696, + 545, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 696, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 307, + 696, + 545, + 707 + ], + "type": "text", + "content": "[51] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt:" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13622" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 706 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 286, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 286, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 286, + 126 + ], + "type": "text", + "content": "[52] Benny J Tang, Angie Boggust, and Arvind Satyanarayan. Vistext: A benchmark for semantically rich chart captioning. arXiv preprint arXiv:2307.05356, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 126, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 126, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 48, + 126, + 287, + 158 + ], + "type": "text", + "content": "[53] Peng Tang, Srikar Appalaraju, R Manmatha, Yusheng Xie, and Vijay Mahadevan. Multiple-question multiple-answer text-vqa. arXiv preprint arXiv:2311.08622, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 158, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 158, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 48, + 158, + 287, + 201 + ], + "type": "text", + "content": "[54] Peng Tang, Pengkai Zhu, Tian Li, Srikar Appalaraju, Vijay Mahadevan, and R Manmatha. Deed: Dynamic early exit on decoder for accelerating encoder-decoder transformer models. arXiv preprint arXiv:2311.08623, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 201, + 287, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 201, + 287, + 222 + ], + "spans": [ + { + "bbox": [ + 48, + 201, + 287, + 222 + ], + "type": "text", + "content": "[55] MN Team et al. Introducing mpt-7b: A new standard for open-source, commercially usable llms, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 222, + 287, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 222, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 287, + 264 + ], + "type": "text", + "content": "[56] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyls, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "text", + "content": "[57] Shuohang Wang, Yang Liu, Yichong Xu, Chenguang Zhu, and Michael Zeng. Want to reduce labeling cost? gpt-3 can help. arXiv preprint arXiv:2108.13487, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 297, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 350 + ], + "type": "text", + "content": "[58] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 350, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 404 + ], + "type": "text", + "content": "[59] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "type": "text", + "content": "[60] Peter West, Chandra Bhagavatula, Jack Hessel, Jena D Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, and Yejin Choi. Symbolic knowledge distillation: from general language models to commonsense models. arXiv preprint arXiv:2110.07178, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 458, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 287, + 513 + ], + "type": "text", + "content": "[61] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1192-1200, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 513, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 287, + 565 + ], + "type": "text", + "content": "[62] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 565, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 565, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 565, + 287, + 609 + ], + "type": "text", + "content": "[63] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 609, + 287, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 609, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 48, + 609, + 287, + 662 + ], + "type": "text", + "content": "[64] Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Josh Tenenbaum. Neural-symbolic vqa: Disentangling reasoning from vision and language understanding. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 662, + 287, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 662, + 287, + 706 + ], + "spans": [ + { + "bbox": [ + 48, + 662, + 287, + 706 + ], + "type": "text", + "content": "[65] Yue Yu, Yuchen Zhuang, Jieyu Zhang, Yu Meng, Alexander Ratner, Ranjay Krishna, Jiaming Shen, and Chao Zhang. Large language model as attributed training data generator: A tale of diversity and bias. In Thirty-seventh Conference on" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 309 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "type": "text", + "content": "[66] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 148, + 545, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 148, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 545, + 191 + ], + "type": "text", + "content": "[67] Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 191, + 545, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 191, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 307, + 191, + 545, + 223 + ], + "type": "text", + "content": "[68] Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. Automatic chain of thought prompting in large language models. arXiv preprint arXiv:2210.03493, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 223, + 545, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 223, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 307, + 223, + 545, + 265 + ], + "type": "text", + "content": "[69] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 265, + 545, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 309 + ], + "type": "text", + "content": "[70] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13623" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/a33fa80f-bfa4-45b2-878d-abb01f3d724e_content_list.json b/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/a33fa80f-bfa4-45b2-878d-abb01f3d724e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..67bc3e87e662e33849261ea37794828f2053b3f5 --- /dev/null +++ b/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/a33fa80f-bfa4-45b2-878d-abb01f3d724e_content_list.json @@ -0,0 +1,1513 @@ +[ + { + "type": "text", + "text": "Systematic comparison of semi-supervised and self-supervised learning for medical image classification", + "text_level": 1, + "bbox": [ + 127, + 130, + 843, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhe Huang* Ruijie Jiang* Shuchin Aeron Michael C. Hughes Tufts University, School of Engineering", + "bbox": [ + 225, + 203, + 743, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{Zhe.Huang, Ruijie.Jiang, Shuchin.Aeron, Michael.Hughes}@tufts.edu", + "bbox": [ + 189, + 242, + 774, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In typical medical image classification problems, labeled data is scarce while unlabeled data is more available. Semi-supervised learning and self-supervised learning are two different research directions that can improve accuracy by learning from extra unlabeled data. Recent methods from both directions have reported significant gains on traditional benchmarks. Yet past benchmarks do not focus on medical tasks and rarely compare self- and semi- methods together on an equal footing. Furthermore, past benchmarks often handle hyperparameter tuning suboptimally. First, they may not tune hyperparameters at all, leading to underfitting. Second, when tuning does occur, it often unrealistically uses a labeled validation set that is much larger than the training set. Therefore currently published rankings might not always corroborate with their practical utility. This study contributes a systematic evaluation of self- and semi- methods with a unified experimental protocol intended to guide a practitioner with scarce overall labeled data and a limited compute budget. We answer two key questions: Can hyperparameter tuning be effective with realistic-sized validation sets? If so, when all methods are tuned well, which self- or semi-supervised methods achieve the best accuracy? Our study compares 13 representative semi- and self-supervised methods to strong labeled-set-only baselines on 4 medical datasets. From $20000+$ GPU hours of computation, we provide valuable best practices to resource-constrained practitioners: hyperparameter tuning is effective, and the semi-supervised method known as MixMatch delivers the most reliable gains across 4 datasets.", + "bbox": [ + 76, + 323, + 472, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. INTRODUCTION", + "text_level": 1, + "bbox": [ + 76, + 806, + 259, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks can deliver exceptional performance on classification tasks when trained with vast labeled", + "bbox": [ + 76, + 834, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "datasets. However, in medical imaging applications assembling a large dataset with appropriate label can be prohibitively costly due to manual effort required by a human expert. In contrast, images alone, without labels, are often readily available in health records databases. In recent years, significant research has focused on developing methods that leverage both the large unlabeled set and a small labeled set to enhance image classifier training, aiming to surpass models trained solely on labeled data.", + "bbox": [ + 500, + 292, + 893, + 430 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Two ways of leveraging unlabeled data are particularly popular: semi-supervised and self-supervised learning, both often abbreviated as SSL. Recent efforts in semisupervised learning [74, 88] usually train deep classifiers jointly [5, 69] using an objective with two loss terms, one favoring labeled-set accuracy and the other favoring label-consistency or label-smoothness. Alternatively, self-supervised methods [65] take a two-stage approach, first training deep representations on the unlabeled set, then fine-tuning a classifier on the labeled set. Exemplars are numerous [10, 17, 19, 20, 37]. Despite the remarkable progress reported in each direction, these two paradigms have been largely developed independently [21]. A direct comparison of the two paradigms has been notably absent. A practitioner building a medical image classifier from limited labeled data may ask, \"Which recent semi- or self-supervised methods are likely to be most effective?\"", + "bbox": [ + 496, + 430, + 893, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Performance can be quite sensitive to hyperparameters for both semi-SL [69, 71] and self-SL [75]. We argue that any careful comparison must consider tuning hyperparameters of all methods in a fair fashion. However, recent work has not handled this well. First, as reviewed in Table 1, recent benchmarks for both semi- and self- paradigms often omit any tuning of hyperparameters (see the Fungi semi-SL experiments of Su et al. [71] or Ericsson et al. [27]'s self-SL benchmark). This practice of using off-the-shelf defaults likely leads to under-performing given a new task, and may impact some methods more than others. An even bigger issue with prevailing semi-SL practice was originally raised by Oliver et al. [62]: many published papers perform hyperparameter tuning on a labeled validation set that", + "bbox": [ + 496, + 688, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Authors Zhe Huang and Ruijie Jiang contributed equally to this work. \nCode: github.com/tufts-ms/SSL-vs-SSL-benchmark [MIT license].", + "bbox": [ + 94, + 875, + 467, + 902 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22282", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "is larger (sometimes much larger) than the limited labeled train set. Such experimental settings are unrealistic. SSL is intended for practitioners without abundant available labeled data. Practitioners that need to make the most of 1000 available labeled images will not elect to put more images in validation than training, and thus are not helped by benchmarks that do. Unfortunately, five years later after [62] we find this issue is still widespread (see Tab. 1). For example, Wang et al. [76]'s semi-SL results on TissueMNIST tune on a validation set over $50\\mathrm{x}$ larger than the labeled train set.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Oliver et al. [62] further cast some doubt on whether effective tuning is possible with small labeled sets, saying \"Extensive hyperparameter tuning may be somewhat futile due to an excessively small collection of held-out data to measure performance on\". There thus remains a pressing question for resource-constrained practitioners: \"Given limited available labeled data and limited compute, is hyperparameter tuning worthwhile?\"", + "bbox": [ + 75, + 242, + 472, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This study makes progress toward answering these questions by delivering a comprehensive comparison of semi- and self-supervised learning algorithms under a resource-constrained scenario that matches how SSL might be used on real-world medical tasks. Our goal is to enable practitioners to tackle new image classification challenges where only modest-sized labeled datasets exist. We target settings with roughly 2-10 class labels of interest, where each class has 30-1000 available labeled images for all model development (including training and validation). We select representative tasks across 4 datasets that span low-resolution (28x28) to moderate-resolution (112x112 and 384x384). On each task, we run careful experiments with representative recent methods from both semi- and self-supervised learning to clarify what gains are possible with unlabeled data and how to achieve them.", + "bbox": [ + 75, + 364, + 472, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We emphasize realism throughout, in 4 distinct ways: (1) using validation set sizes that are never bigger than the train set, avoiding the unrealistically-large validation sets common in past work (Tab. 1); (2) performing hyperparameter search using the same protocol, same compute budget (fixed number of hours)* and same hardware (one NVIDIA A100 GPU) for all algorithms for fair comparison; (3) respecting natural class imbalance; (4) profiling performance over time, to inform labs with smaller runtime budgets.", + "bbox": [ + 75, + 607, + 472, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the contributions of this study are:", + "bbox": [ + 96, + 744, + 410, + 758 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We provide a systematic comparison of semi- and self-supervised methods on equal footing to connect two research directions that have been heretofore separate.", + "2. We adopt a realistic experimental protocol designed to consider the same constraints on available labels and runtime that SSL practitioners face, avoiding the unre" + ], + "bbox": [ + 76, + 758, + 470, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "alistic aspects of past benchmarks, especially the use of far too large validation sets.", + "bbox": [ + 517, + 90, + 890, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. We show that hyperparameter tuning is viable with a realistic-sized validation set, and in many cases necessary to do well at a new classification task with new data.", + "bbox": [ + 500, + 121, + 893, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ultimately, we hope this study guides practitioners with limited data toward successful deployment of semi- and self-supervised methods on real problems.", + "bbox": [ + 500, + 166, + 893, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. BACKGROUND AND METHODS", + "text_level": 1, + "bbox": [ + 500, + 224, + 818, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unified Problem Formulation. Following the recent survey by Chen et al. [21], we adopt a unified perspective for supervised, semi-supervised and self-supervised image classification with limited available labeled data. For model development (including training and hyperparameter selection), we assume there are two available datasets. First, a small labeled dataset $\\mathcal{L}$ of feature-label pairs $(x,y)$ , where each image is represented by a $D$ -dimensional feature vector $x \\in \\mathbb{R}^D$ and its corresponding class label takes one of $C$ possible values: $y \\in \\{1,2,\\dots,C\\}$ . Second, an unlabeled set $\\mathcal{U}$ containing only feature vectors. Typically, we assume the unlabeled set is much larger: $|\\mathcal{U}| \\gg |\\mathcal{L}|$ .", + "bbox": [ + 496, + 250, + 893, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given labeled set $\\mathcal{L}$ and unlabeled set $\\mathcal{U}$ , we wish to train a neural network that can map each input $x$ to a probability vector in the $C$ -dimensional simplex $\\Delta^C$ representing a distribution over $C$ class labels. Let $f_v(\\cdot): \\mathbb{R}^D \\to \\mathbb{R}^F$ denote a backbone neural network with parameters $v$ producing an $F$ -dimensional embedding given any input image. Let $g_w(\\cdot): \\mathbb{R}^F \\to \\Delta^C$ denote a final linear-softmax classification layer with parameters $w$ . The following unified objective can capture all three learning paradigms:", + "bbox": [ + 496, + 431, + 893, + 566 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} v ^ {*}, w ^ {*} \\leftarrow \\arg \\min _ {v, w} \\sum_ {x, y \\in \\mathcal {L}} \\lambda^ {L} \\ell^ {L} (y, g _ {w} (f _ {v} (x))) \\tag {1} \\\\ + \\sum_ {x \\in \\mathcal {U}} \\lambda^ {U} \\ell^ {U} (x, f _ {v}, g _ {w}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 574, + 890, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, $\\ell^L$ represents a labeled-set loss (e.g. multi-class cross-entropy), and $\\ell^U$ represents a unlabeled-set loss. $\\lambda^L, \\lambda^U \\geq 0$ are weights for the corresponding loss terms. The design of $\\ell^U$ is usually what differs substantially across methods. For instance, setting $\\ell^U$ to be cross-entropy computed with pseudo-labels generated from classifier $g$ recovers PseudoLabel [58]; a temperature-scaled instance-similarity contrastive loss recovers SimCLR [17].", + "bbox": [ + 496, + 628, + 890, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "All three learning paradigms that we study optimize Eq. (1), yet differ in the number of phases and in how to set the scalar weights $\\lambda^L, \\lambda^U$ on each loss term. Supervised learning ignores the unlabeled term throughout training ( $\\lambda^U = 0$ ), thus learning parameters using only the labeled set. Semi-supervised methods include both terms in one end-to-end training, keeping both $\\lambda^U > 0$ and $\\lambda^L > 0$ .", + "bbox": [ + 496, + 750, + 893, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised learning has two phases. In phase 1 (\"pretraining\"), the labeled term is omitted ( $\\lambda^L = 0$ , $\\lambda^U = 1$ ) and the focus of learning is an effective representation", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "*Oliver et al. [62] give each method 1000 trials of a cloud hyperparameter tuning service. However, training speed can differ substantially across SSL methods. We argue a fixed wallclock time budget is more fair.", + "bbox": [ + 75, + 862, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22283", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/ad0f42e6e06b398df174d9b3833a92bcbd9166d948a05ec5342b566d927f235f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BenchmarkMethodsUnrealistic ExperimentsLabeled train sizeLabeled val. sizeAcc vs. Time?
Realistic eval. SSL [a]SemiCIFAR-10 (Tab. 1-2, Fig. 2)40005000no
SVHN (Tab. 1-2, Fig. 3-4)10007325
Fine-grained SSL [b]Semi & 1 Self*Semi-Aves59598000no
Semi-Fungi4141None
USB [c]SemiTissueMNIST80/40023640†no
Semi-Aves5959None
Self benchmark [d]SelfImageNet1.28 mil.Noneno
SSL-vs-SSL (ours)Semi & Selfno400-1660no bigger than trainyes
", + "bbox": [ + 78, + 87, + 897, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparison of related benchmarks of semi-supervised and self-supervised learning. Past works either do no hyperparameter tuning at all (val. size = None), or use an unrealistically large validation set (defined as larger than the labeled train set) in some or all experiments. Further, each work almost exclusively looks at methods from one paradigm, either semi- or self- (*: [b] includes one self-supervised method, MOCO). In contrast, in this paper we benchmark 6 semi- and 7 self-supervised algorithms with hyperparameter tuning on realistic validation sets. Acc vs. Time? indicates whether the work analyzes performance over training time. Number marked † confirmed via GitHub comment by USB authors. Citations: a: Oliver et al. [62], b: Su et al. [71] c: Wang et al. [76] d: Ericsson et al. [27].", + "bbox": [ + 75, + 239, + 893, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "layer $f_{v}$ (classifier $g_{w}$ is not included in this phase). In phase 2 (\"fine-tuning\"), we focus on the labeled term and omit the unlabeled term $(\\lambda^{L} = 1, \\lambda^{U} = 0)$ . We fix the representation parameter $v$ and fine-tune the classifier $w$ .", + "bbox": [ + 75, + 345, + 468, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now identify the 16 methods we will evaluate:", + "bbox": [ + 96, + 410, + 428, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Supervised methods. The goal of leveraging unlabeled data $\\mathcal{U}$ is to obtain better performance than what we could obtain using only the labeled set $\\mathcal{L}$ . Therefore, we naturally compare to 3 high-quality supervised baselines that use only the labeled set. First, \"Sup\" denotes a classifier trained with supervised loss $\\ell^L$ set to multi-class cross-entropy. Second, \"MixUp\" trains with cross entropy with the addition of mixup data augmentation [85]. Finally, \"SupCon\" pursues a supervised contrastive learning loss for $\\ell^L$ [49].", + "bbox": [ + 75, + 426, + 468, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semi-supervised methods. We compare 6 semi-supervised methods that train deep classifiers on both labeled and unlabeled data simultaneously as in Eq. (1). To represent the state-of-the-art of semi-supervised image classification, we select Pseudo Label (\"PseudoL\") [58], Mean Teacher (\"MeanTch\") [73], MixMatch [5], FixMatch [69], FlexMatch [84] and CoMatch [59]. These choices cover a reasonably wide spectrum of unlabeled loss design strategy, year of publication, and computation cost. CoMatch represents a recent trend of combining semi- and self-supervision. See App D.2 for a wider literature review.", + "bbox": [ + 75, + 566, + 468, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised methods. We compare 7 self-supervised algorithms: SimCLR [17], MOCO (v2) [20, 37], SwAV [10], BYOL [34], SimSiam [19], DINO [11] and Barlow Twins (\"BarlowTw\") [82]. These algorithms epitomize the field of self-supervised learning as of this writing. SimCLR, MOCO (v2), and SwAV are based on contrastive learning, which learns effective representations when provided with similar and dissimilar samples. BYOL was designed to circumvent the need for dissimilar samples. SimSiam is a simple yet effective Siamese representation learning method. DINO uses a teacher-student network architec", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ture for representation distillation. Barlow Twins preserves information and reduces redundancy by favoring a close-to-identity cross-correlation matrix between augmented pairs of data. See App D.3 for further review.", + "bbox": [ + 496, + 345, + 890, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. RELATED WORK", + "text_level": 1, + "bbox": [ + 498, + 421, + 687, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1 summarizes key attributes of existing major benchmarks for semi-supervised and self-supervised methods. We now discuss how our study situates in this context.", + "bbox": [ + 496, + 446, + 890, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Comparison to Oliver et al. Oliver et al. [62] provide an influential benchmark of deep semi-supervised methods. Like our work, they highlight the pressing issue of using unrealistically large validation sets for hyperparameter tuning. However, most of their experiments (e.g. all their tables and their figures 2-4) still use this unrealistic setting to be comparable to other works; only one subsection (Sec. 4.6) examines validation sets no larger than the train set. Instead, we exclusively use validation set sizes no larger than training set, mimicking what practitioners would face in real applications. Our definition of a \"realistic\" size for a validation set – no larger than the labeled train set – is broader than Oliver et al.'s definition of $10\\%$ of the train set. We favor our definition because Oliver et al. find specifically that \"for validation sets of the same size $(100\\%)$ as the training set, some differentiation between the approaches is possible.\" In contrast, they found reliable differentiation was not always possible with much smaller validation sets.", + "bbox": [ + 496, + 492, + 892, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, the benchmarking results presented by Oliver et al. require each algorithm to complete \"1000 trials of Gaussian Process-based black-box optimization using Google Cloud ML Engine\". This resource-intensive process seems impractical for researchers outside well-funded industrial labs. We use more modest compute budgets and specifically allow each method the same total wallclock runtime for tuning, again obeying practical constraints.", + "bbox": [ + 496, + 763, + 890, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Other semi-supervised benchmarks. TorchSSL [84]", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22284", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "benchmarked eight popular semi-supervised learning algorithm using a unified codebase. The same group later extended that effort to natural language and audio processing in their USB benchmark [76]. Another recent effort for fine-grained classification by Su et al. [71] evaluates semi-supervised learning on datasets that exhibit class imbalance and contains images from novel classes in the unlabeled set, studying the effect of different initializations and the contents of unlabeled data on the performance of semi-supervised methods. While making significant contributions, these works focused on semi-supervision almost exclusively and did not include multiple self-supervised learning methods, thus leaving open the questions \"How do the two paradigms compare? Which methods are best overall?\"", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prior self-supervised benchmarks. Several works have proposed different benchmarks for self-supervised methods. Goyal et al. [32] introduced a benchmark that covers 9 different tasks, such as object detection and visual navigation. Ericsson et al. [27] compared thirteen top self-supervised models on 40 downstream tasks. Da Costa et al. [23] presented a library of self-supervised methods that can be easily plugged into different downstream tasks and datasets. However, these works mainly consider other self-SL methods without direct comparison to the semi-SL paradigm. Moreover, they mostly do not incorporate the use of a validation set for hyperparameter tuning at all. This can lead to suboptimal accuracy at test time and complicate comparisons to methods that do tune.", + "bbox": [ + 75, + 301, + 473, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Self-supervision for medical images. Our work is complementary to recent efforts that assess self-supervised pipelines for medical image classification [2, 3]. They focus primarily on how to design multi-stage transfer learning pipelines and do not comprehensively compare many different self-supervised methods or any semi-supervised methods. Further, many datasets studied by [3] come from proprietary projects conducted at Google. In contrast, all our data and experiments are open and reproducible by others.", + "bbox": [ + 75, + 513, + 473, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Combining semi- and self-supervision. Recent research has explored the fusion of semi-supervised learning and self-supervised learning ideas [50, 59, 83, 87]. Some recent surveys [21, 65] compare semi- and self-supervised learning, but they focus on literature review while we offer realistic and comprehensive benchmarking experiments.", + "bbox": [ + 75, + 650, + 472, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. DATASETS AND TASKS", + "text_level": 1, + "bbox": [ + 76, + 753, + 316, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We study four open-access medical image classification datasets, all with 2D images that are fully deidentified. Table 2 reports statistics for all train/test splits. The exact splits used in our study can be found in our open-source codebase, documented in App. A.", + "bbox": [ + 75, + 779, + 470, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Two datasets - PathMNIST and TissueMNIST - are selected from the MedMNIST collection [80] (criteria in App. B.2). Prior experiments by Yang et al. [80] suggest", + "bbox": [ + 75, + 854, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/bcb04603f2917c0b00b92bd5dec78f9e269192871472c36a5e0f00f75e9a388e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TissueMNISTPathMNIST
LabeledUnlab.LabeledUnlab.
TrainValTestTrainTrainValTestTrain
total40040047280165066total450450718089546
GE151516775851NORM39397417847
DCT191922337795MUC404010357966
POD191922027686ADI474713389319
LEU2828336911761STR47474219354
IE3737440215369BACK48488479461
TAL5959703124549DEB525233910308
PT95951120139108LYM525264310349
CD/CT1281281516552947MUS616159212121
TUM6464123312821
TMED-2AIROGS
LabeledUnlab.LabeledUnlab.
TrainValTestTrainTrainValTestTrain
total16602352019353500total600600600094242
PSAX22350342-Glaucoma6060600-
A2C32528319-No Glauc.5405405400-
A4C46239423-
PLAX650118935-
", + "bbox": [ + 506, + 88, + 890, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Summary statistics of train/validation/test splits for all datasets in our study. Each table's rows are arranged in ascending order based on the number of per-class labeled train images. Full description of class names can be found in App B.3.", + "bbox": [ + 498, + 402, + 893, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "their $28 \\times 28$ resolution is a reasonable choice for rapid prototyping; using larger $224 \\times 224$ resolution does not yield much more accurate classifiers for these two datasets.", + "bbox": [ + 498, + 488, + 890, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Two other datasets represent more moderate resolutions closely tied to contemporary clinical research: the 112x112 Tufts Medical Echocardiogram Dataset (TMED-2) and the 384x384 AIROGS dataset. Further details for each dataset are provided in a dedicated paragraph below.", + "bbox": [ + 496, + 535, + 893, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each dataset, our data splitting strategy closely mirrors the conditions of real SSL applications. First, we let labeled training and validation sets contain a natural distribution of classes even if that may be imbalanced. This reflects how data would likely be affordably collected (by random sampling) and avoids artificially balanced training sets that will not match the population an algorithm would encounter in a deployment. Second, to be realistic we ensure validation sets are never larger than the available labeled train sets. This is in contrast to previous benchmarks: Wang et al. [76]'s TissueMNIST hyperparameter search used a validation set of 23,640 images even though the labeled training set contained only 80 or 400 images. In practice, such a large validation set would almost certainly be re-allocated to improve training, as noted in Oliver et al. [62].", + "bbox": [ + 496, + 612, + 893, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TissueMNIST is a lightweight dataset of 28x28 images of human kidney cortex cells, organized into 8 categories. The original dataset is fully-labeled and distributed with predefined train/val/test split of 165,466/23,640/47,280 im", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22285", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ages with some class imbalance. We assume a total labeling budget of 800 images, evenly split between training and validation (to ensure hyperparameter tuning can differentiate methods). We form a labeled training set of 400 images from the predefined training set, sampling each class by its frequency in the original training set. We form a labeled validation set of 400 images sampled from the predefined validation set. For unlabeled set, we keep all remaining images in the original training split, discarding known labels.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PathMNIST is another lightweight dataset of 28x28 images of patches from colorectal cancer histology slides that comprise of 9 tissue types. The original dataset is fully-labeled and distributed with predefined train/val/test split of 89,996/10,004/7,180 images. Class imbalance is less severe than TissueMNIST. We assume a total labeled data budget of 900 images, again evenly split between training and validation. Labeled train, labeled valid, and unlabeled sets are sampled via the same procedures as in TissueMNIST. For both MedMNIST datasets, we use the predefined test set.", + "bbox": [ + 75, + 234, + 472, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TMED-2 [39, 40] is an open-access dataset of $112 \\times 112$ 2D grayscale images captured from routine echocardiogram scans. Each scan produces dozens of ultrasound images of the heart, captured from multiple acquisition angles (i.e., different anatomic views). In this study, we adopt Huang et al.'s view-classification task: the goal is to classify each image into one of 4 canonical view types (see App. B.3). View classification is clinically important: measurements or diagnoses of heart disease can only be made when looking at the right anatomical view [60, 77]. We used the provided train/validation/test split (id #1), which is naturally imbalanced and matches our criteria for realistic sizing. We further use the provided large unlabeled set of 353,500 images. TMED-2's unlabeled set is both authentic (no true labels are available at all, unlike other benchmarks that \"forget\" known labels) and uncurated [41] (contains images of view types beyond the 4 classes in the labeled task).", + "bbox": [ + 75, + 393, + 472, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AIROGS [24] is a public dataset released for a recent competition where the binary classification task is to decide whether the patient should be referred for glaucoma or not, given a color fundus image of the retina (eye). We selected this dataset because it represents an active research challenge even with all available labels. Furthermore, because labels for this dataset were acquired at great expense (multiple human annotators graded each of over 100k images), we hope our work helps assess how self-/semi-supervision could reduce the annotation costs of future challenges. We selected the $384 \\times 384$ resolution favored by several high-performing competition entries. We chose a labeling budget of 1200 total images, evenly split between train and validation; the rare positive class (9 to 1 imbalance) is representative of many screening tasks. We took the rest of available images as the unlabeled train set.", + "bbox": [ + 75, + 657, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. EXPERIMENTAL DESIGN", + "text_level": 1, + "bbox": [ + 498, + 89, + 764, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Performance metric. We use balanced accuracy [33, 35] as our primary performance metric. For a task with $C \\geq 2$ classes, let $y_{1:N}$ denote true labels for $N$ examples in a test set, and $\\hat{y}_{1:N}$ denote a classifier's predicted labels. Let $\\mathrm{TP}_c$ count true positives for class $c$ (number of correctly classified examples whose true label is $c$ ), and let $N_c$ count the total number of examples with true label $c$ . Then we compute balanced accuracy (BA) as a percentage:", + "bbox": [ + 496, + 114, + 893, + 237 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf{BA}(y_{1:N},\\hat{y}_{1:N}) = \\frac{1}{C}\\sum_{c = 1}^{C}\\frac{\\mathrm{TP}_{c}(y_{1:N},\\hat{y}_{1:N})}{N_{c}(y_{1:N})}\\cdot 100\\% \\tag{2}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 244, + 890, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Balanced accuracy is more suitable than standard accuracy for imbalanced problems when each class matters equally. The expected BA of a uniform random guess is $\\frac{100}{C}\\%$ .", + "bbox": [ + 496, + 296, + 890, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On AIROGS, we also track metrics recommended by its creators [24] for clinical utility: AUROC, partial AUROC ( $>90\\%$ specificity) and sensitivity-at- $95\\%$ -specificity.", + "bbox": [ + 496, + 342, + 890, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architectures. CNN backbones are popular in medical imaging [6, 28-31, 42, 45, 54-56, 77, 78]. We use ResNet-18 [36] on Tissue and Path, and WideResNet-28-2 [81] on TMED-2. We experiment with both ResNet-18 and 50 [36] on AIROGS to assess architectural differences.", + "bbox": [ + 496, + 387, + 890, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training with early stopping. For each training phase, we perform minibatch gradient descent on Eq. (1) with Adam [51] optimizer and a cosine learning rate schedule [69]. Each training phase proceeds for up to 200 epochs, where one epoch represents enough minibatch updates to process the equivalent of the entire combined training set $\\mathcal{L} \\cup \\mathcal{U}$ . After every epoch, we record balanced accuracy on the validation set. If this value plateaus for 20 consecutive epochs, we stop the current training phase early.", + "bbox": [ + 496, + 462, + 890, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hyperparameters. Semi- and self-supervised learning can both be sensitive to hyperparameters [71, 75]. Our evaluations tune both shared variables (e.g. learning rates, weight decay, unlabeled loss weight $\\lambda^U$ ) as well as hyperparameters unique to each algorithm. See App. F.1 for details on all hyperparameters for all methods.", + "bbox": [ + 496, + 598, + 890, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unified procedure for training and hyperparameter tuning. We formulate a unified procedure mindful of realistic hardware and runtime constraints for a non-industrial-scale lab working on a new medical dataset. We assume each algorithm has access to one NVIDIA A100 GPU for a fixed number of hours. Within the allotted compute budget, for each algorithm we execute a serial random search over hyperparameters, sequentially sampling each new configuration and then training until either an early stopping criteria is satisfied or the maximum epoch is reached. We track the best-so-far classifier in terms of validation-set performance every epoch. Algorithm D.1 provides pseudocode for the procedure. Each time a new hyperparameter configuration is needed, we sample each hyperparameter value independ-", + "bbox": [ + 496, + 689, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22286", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dently from a distribution designed to cover its common settings in prior literature (App. F.1). Our choice of random search on a budget is thought to yield better performance than grid search [4], fairly expends the same effort to train all methods regardless of their cost-per-epoch or hyperparameter complexity, and does not assume industrial-scale access to 1000 cloud-computing trials as in Oliver et al. [62]. We find that performance saturates after about 25 hours per 80000 unlabeled examples. We thus allocate for the total time budget 25 hours for PathMNIST, 50 hours for TissueMNIST, and 100 hours for TMED-2. We allow 100 hours for AIROGS due to its larger resolution. Due to limited resources, we select only a few competitive methods to run on AIROGS based on the results on other datasets.", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Self-supervised classification phase. Self-supervised methods by definition do not utilize label information when training representation layer weights $v$ . To enable a proper comparison, for all self-SL methods our Alg. D.1 introduces an additional classification layer, training it anew after each epoch, each time using a 10 trial random search for an L2-penalty regularization hyperparameter. We retain the best performing weights $w$ on the validation set. We emphasize that $w$ does not impact overall self-supervised training of $v$ .", + "bbox": [ + 75, + 301, + 472, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data augmentation. Random flip and crop are used for all semi-supervised and supervised methods. MixMatch also utilizes MixUp [85], while RandAugment [22] is used by FixMatch and FlexMatch. For each self-supervised method as well as SupCon, we apply the same SimCLR augmentation: random flip, crop, color jitter, and grayscale.", + "bbox": [ + 75, + 439, + 472, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multiple trials. We repeat Alg. D.1 for each method across 5 separate trials (distinct random seeds). We record the mean balanced accuracy on valid and test sets of these trials every 60 minutes (every $30\\mathrm{min}$ for Tissue and Path).", + "bbox": [ + 75, + 529, + 472, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. RESULTS & ANALYSIS", + "text_level": 1, + "bbox": [ + 76, + 603, + 313, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In each panel of Fig. 1, we focus on one dataset and architecture, plotting for each algorithm one line showing test set balanced accuracy (averaged across 5 trials) as a function of time spent running Alg. D.1. Variance across trials is visualized in Fig C.4. Extra results on AIROGS showing other performance metrics over time (AUROC, partial AUROC, and sensitivity at $95\\%$ specificity) are found in Fig. C.3. We emphasize that following best practice, each checkpoint represents the best hyperparameters as selected on the validation set; we then report that checkpoint's test set performance. These results help us answer two key questions:", + "bbox": [ + 75, + 628, + 472, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Is hyperparameter tuning worthwhile for SSL methods when validation set sizes cannot be larger than the train set? Across Fig. 1 and C.3, all 16 algorithms show roughly monotonic improvements in test performance over time, despite using a realistic-sized validation set. The final performance of every algorithm and dataset shows improvement due to our unified training and tuning procedure (fur", + "bbox": [ + 75, + 795, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b5c1caa6b876ca95b2dbdd48a6e1f291c10eb6da222038fe4210180b0388b1a9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
nameGain in Bal.Acc. over best labeled-set-only
PathTissueTMEDAIROGSmedianworst
MixMatch12.43.7-0.25.74.7-0.2
MOCO7.94.4-18.8n/a4.4-18.8
SwAV8.44.3-10.1n/a4.3-10.1
SimCLR11.05.9-7.81.63.8-7.8
BYOL12.16.2-6.00.93.5-6.0
BarlowTw11.22.2-1.9n/a2.2-1.9
SimSiam1.92.1-11.6n/a1.9-11.6
FlexMtech11.14.5-1.1-4.21.7-4.2
DINO12.66.5-3.4-3.81.6-3.8
FixMatch9.32.90.2-4.51.6-4.5
CoMatch14.10.2-3.0n/a0.2-3.0
MeanTch-1.6-1.9-1.2n/a-1.6-1.9
PseudoL0.6-1.4-4.4-3.2-2.3-4.4
ref. val.69.833.994.971.5
", + "bbox": [ + 501, + 88, + 895, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Gains from SSL methods over only using labeled set, across 4 datasets. Self (italicized), Semi (normal). We report gains in percentage balanced accuracy (higher is better) in the mean final test set performance (averaged over 5 runs of Alg. D.1) over a reference value that represents the best labeled-set-only run among Sup (minimize cross entropy), SupCon, and MixUp. MixMatch has the highest median gain, and is the only method to never score notably worse ( $>1$ point) than the best labeled set only run.", + "bbox": [ + 496, + 349, + 893, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ther discussion in App. E.1). Therefore, we answer: Yes, realistically-sized validation sets for hyperparameter tuning and checkpoint selection can be effective.", + "bbox": [ + 496, + 489, + 890, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "What are the best SSL methods? Among the many methods we study in Fig. 1, no method clearly outperforms others across all datasets. On Path, CoMatch, MixMatch, DINO, and BYOL perform best. On Tissue, self-supervised methods like DINO, BYOL, and SimCLR score well. On TMED-2, ultimately only FixMatch and MixMatch are competitive with the MixUp baseline. On AIROGS, only MixMatch, SimCLR and BYOL beat the Sup baseline.", + "bbox": [ + 496, + 536, + 893, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each semi- and self- SSL method, Tab. 3 reports its relative gain in balanced accuracy over the best labeled-set-only supervised baseline on each dataset. We conclude that MixMatch represents the best overall choice as it consistently performs at or near the top across all tasks. It is the only method to never deliver results worse by more than 1 percentage point than the best labeled-set-only baseline (see \"worst\" column of Tab. 3). This is in stark contrast to USB [76], where FixMatch and FlexMatch are ranked notably higher than MixMatch. Such differences stress the importance of choosing proper evaluation protocol tailored to specific needs. We do caution that hyperparameter tuning is strongly recommended for MixMatch to succeed on a new dataset (see AIROGS result in Tab. 4).", + "bbox": [ + 496, + 657, + 895, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pretraining vs. from scratch. Plots in App. C show accuracy-over-time profiles using initial weights pretrained", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22287", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d5b0f3d7dae1ceee445e3d403de4cfecd9548f62debb697ce82095b30a09e424.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 88, + 344, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08b5e3edfd6600c308e8c6cf71fd85e8ecee90584020fa583ec438a3695b4844.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 88, + 614, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4cb7f33695ab7fcb2ab4a807491cf415dbc15f16e2411b4392e8fc34b889309e.jpg", + "image_caption": [], + "image_footnote": [ + "semi labeled-set-only self FixMatch SupCon SimCLR", + "-FlexMtech-MixUp BYOL" + ], + "bbox": [ + 620, + 89, + 880, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7fd21f09a046eeba760d39e1bfa78d57b007ab829541543994066279d6e534e3.jpg", + "image_caption": [ + "Figure 1. Balanced accuracy over time profiles of semi- and self-supervised methods across all 4 datasets. At each time, we report the test set bal. acc. of each method (mean over 5 trials of Alg. D.1). Best viewed electronically. Top Row: On these 3 datasets, we compare all 6 semi- and 7 self- methods (see legend in lower right, citations in Sec. 2), to 3 labeled-set-only baselines. Bottom Row: On larger AIROGS dataset, we compare selected methods representative of the best in the top row. Thin lines show final performance to ease comparison. From these charts, we suggest that our unified training and tuning (Alg. D.1) is effective, as all methods show gains over time." + ], + "image_footnote": [], + "bbox": [ + 80, + 227, + 344, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79769279afefefe5ce417a5d014f6f367d2f8df10ac392213341f8c6e4d0f28d.jpg", + "image_caption": [], + "image_footnote": [ + "-MixMatch..Sup -DINO", + "—CoMatch —BarlowTw", + "—MeanTch —SwAV", + "-PseudoL -SimSiam", + "—MOCO" + ], + "bbox": [ + 351, + 227, + 612, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on ImageNet. Compared to training from scratch, we see only slight gains (less than 3 points of BA) in ultimate test-set accuracy when top methods use pretraining, which aligns with observations in [66] but counters the \"large margin\" gains (sometimes over 10 points) reported by Su et al. [71]. Pretraining's benefits still include improved convergence speed and reduced variance across trials.", + "bbox": [ + 75, + 476, + 468, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ResNet-18 vs. ResNet-50. The bottom row of Fig 1 compares ResNet-18 and ResNet-50 architectures on AIROGS. We broadly find that the method rankings are similar. More notably, ResNet-50 is not substantially better than ResNet-18 in this task when compared using the same runtime budget. Profiles of other performance metrics in App C.3 suggest the same conclusions.", + "bbox": [ + 75, + 583, + 470, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tuning vs. Transferring from another dataset. To make the most of limited labeled data, one strategy is to use the entire labeled set for training, reserving no validation set at all. This approach relies on pre-established \"best\" hyperparameters sourced from other datasets and transferred to the target task, obviating the need for data-specific tuning. Su et al. [71] employed this strategy in their benchmark, avoiding tuning due to concerns that small validation sets would yield unreliable selection. However, this claim in their work was not supported by experiments. Here, we directly compare our hyperparameter tuning strategy (Alg. D.1) with Su et al.'s transfer strategy. We train the latter on the combined train + validation sets, using \"best\" hyperparameters from CIFAR-10 and Tissue (details", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in App. F.2). Tab. 4 reports the final test-set balanced accuracy of each strategy on Path, TMED-2 and AIROGS.", + "bbox": [ + 498, + 476, + 890, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our hyperparameter tuning strategy is competitive across all datasets: the \"Tune\" columns of Tab. 4 have the highest fraction of green cells. The alternative transfer strategy is faster and occasionally yield reasonable results. However, there is no guarantee of good performance from hyperparameter transfer, as evidenced by the inferior cases in Tab. 4 highlighted in yellow and red. There is also no consistent distinction between transferring from Tissue (an arguably more related medical dataset) and transferring from CIFAR-10, underscoring the challenge of identifying a \"closely related\" dataset for hyperparameter transfer.", + "bbox": [ + 496, + 537, + 890, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, we recommend researchers pursuing SSL for a new medical classification task should adopt our hyperparameter tuning strategy. Transferring off-the-shelf hyperparameters has highly variable performance and can even result in divergence (see MixMatch on AIROGS in Tab. 4). Even with reasonable hyperparameters, having no validation set makes checkpoint selection challenging. Adjacent epochs that exhibit similar low training losses may differ by over $5\\%$ in test accuracy in our experiments. Improving transfer from existing hyperparameters to a new task is an interesting further direction of research.", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22288", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9110be42a07950ec75ea9743160ae98742d09f7fb3364ae3a35a7ac3ba6a47de.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PathMNISTTMED-2AIROGS
Hyperparam. strategy:TuneBest from TissueBest from CIFAR-10TuneBest from TissueBest from CIFAR-10TuneBest from TissueBest from CIFAR-10
semiMixMatch82.2483.8773.5194.7192.6774.7377.1461.75Diverged
CoMatch83.9580.5186.4291.8792.5192.67
FixMatch79.1182.6479.7395.0695.1295.7366.9266.3165.50
FlexMatch80.9178.7080.2393.7793.9793.7067.2372.8274.52
Pseudo-Label70.4266.1264.7190.5288.7882.1368.2164.4366.93
Mean-Teacher68.2155.9056.1393.7091.2158.26
selfSimCLR81.0579.8680.0587.0484.4187.2473.0167.1567.95
BYOL81.7980.3580.1688.9088.4185.1572.4070.1370.63
SwAV77.9078.2481.3084.7682.1782.20
MoCo77.7379.9579.1276.1478.5679.11
SimSiam71.7868.6270.4483.2480.9380.20
BT80.8572.7980.2592.9691.6390.81
DINO82.5277.5080.6791.4584.8090.0367.6166.4667.13
", + "bbox": [ + 88, + 87, + 883, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Hyperparameter strategy evaluation. Table reports ultimate test-set balanced accuracy as percentage (mean over 5 trials, higher is better) on three target datasets: PathMNIST, TMED-2, and AIROGS. 3 different hyperparameter strategies are compared for each SSL method. Due to computation constraints, we only select a few algorithms to run on AIROGS, explained in Sec. 5. Tune strategy: tuning for best hyperparameters via our proposed Alg. D.1 for 100 hours (TMED-2 and AIROGS) and 25 hours (PathMNIST), using provided train/validation split. Best from Tissue/CIFAR-10 strategy: just one training phase on the target dataset using the \"best\" hyperparameters selected from the named source data. These runs combine train and validation data for fitting following Su et al. [71], and typically take 1-4/5-30/10-40 hours on Path/TMED-2/AIROGS depending on the algorithm and whether early stopping is triggered. To highlight which strategies are most effective, we bold the best strategy within each dataset-specific row, and color cells relative to this bold value as green (within 2.5 of best), yellow (within 5 of best), and red (worse by 5 or more points).", + "bbox": [ + 75, + 345, + 893, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. DISCUSSION & CONCLUSION", + "text_level": 1, + "bbox": [ + 76, + 494, + 380, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have contributed a benchmark that helps practitioners quantify what gains in a classification task are possible from the addition of unlabeled data and which methods help achieve them. We offer a unified approach to training and hyperparameter selection of semi-supervised methods, self-supervised methods, and supervised baselines that is both affordable and realistic for research labs tackling new challenges without industrial-scale resources.", + "bbox": [ + 75, + 522, + 468, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. We deliberately focused on a modest number of labeled examples (30 - 1000) per class, motivated by projects where substantive effort has already gone into labeled data collection. Applications in more scarce-label regimes (e.g. zero-shot or few-shot) may need to also consult other benchmarks, as should those looking at hundreds of fine-grained classes. We also did not specifically study situations where the distribution of labeled and unlabeled data significantly differs [41, 44, 62, 67].", + "bbox": [ + 75, + 643, + 470, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To enable thorough experiments, we focused on moderate ResNet architectures that have proven effective in clinical applications [42, 78]. As ViT-based methods become popular in medical imaging [1, 43], it would be valuable to understand if similar findings hold with vision transformers [26]. We leave this to future work. Ideally, our results would be verified across multiple train-validation splits; however the resources needed remain prohibitive.", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our analysis here is limited the specific metrics of balanced accuracy and sensitivity/specificity (for AIROGS). Other clinically useful metrics, such as AUPRC, calibration, or net benefit, may be needed to decide if a classifier is appropriate for deployment [70]. For medical applications, it is also key to understand fairness across subpopulations [13] to avoid propagating structural disadvantages.", + "bbox": [ + 496, + 497, + 890, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Broader impact. All data analyzed here represent fully-deidentified open-access images approved for widespread use by their creators. We think the benefit of promoting these medical tasks to advance ML research outweighs the slight risk of patient reidentification by a bad actor.", + "bbox": [ + 496, + 608, + 890, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Outlook. Our experiments show that real benefits from the addition of unlabeled data are sometimes possible: our recommended methods see gains of $+5$ points of balanced accuracy on TissueMNIST, $+10$ points on PathMNIST, and $+5$ points on AIROGS against strong labeled-set-only baselines. In contrast, most tested methods do not add significant gains on TMED-2, perhaps due to that benchmark's uncurated nature [41]. We hope our work enables the research community to convert decades of effort on SSL into improved patient outcomes and better scientific understanding of disease and possible treatments. We further hope that our benchmark inspires those that pursue improved methodological contributions to favor realistic evaluation protocols and clinically-relevant datasets.", + "bbox": [ + 496, + 688, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22289", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] N Ahmadi, MY Tsang, AN Gu, TSM Tsang, and P Abolmaesumi. Transformer-based spatio-temporal analysis for classification of aortic stenosis severity from echocardiography cine series. IEEE Transactions on Medical Imaging, 2023. 8", + "[2] Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zachary Beaver, Jan Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, Vivek Natarajan, and Mohammad Norouzi. Big Self-Supervised Models Advance Medical Image Classification. In International Conference on Computer Vision (ICCV). arXiv, 2021. 4", + "[3] Shekoofeh Azizi, Laura Culp, Jan Freyberg, Basil Mustafa, Sebastien Baur, Simon Kornblith, Ting Chen, Nenad Tomasev, Jovana Mitrovic, Patricia Strachan, S. Sara Mahdavi, Ellery Wulczyn, Boris Babenko, Megan Walker, Aaron Loh, Po-Hsuan Cameron Chen, Yuan Liu, Pinal Bavishi, Scott Mayer McKinney, Jim Winkens, Abhijit Guha Roy, Zach Beaver, Fiona Ryan, Justin Krogue, Mozziyar Etemadi, Umesh Telang, Yun Liu, Lily Peng, Greg S. Corrado, Dale R. Webster, David Fleet, Geoffrey Hinton, Neil Houlsby, Alan Karthikesalingam, Mohammad Norouzi, and Vivek Natarajan. Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging. Nature Biomedical Engineering, 7(6):756-779, 2023. 4, 23", + "[4] James Bergstra and Yoshua Bengio. Random search for hyper-parameter optimization. Journal of machine learning research, 13(2), 2012. 6", + "[5] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 1, 3, 21, 23", + "[6] Benjamin Billot, Colin Magdamo, Steven E Arnold, Sudeshna Das, and Juan Eugenio Iglesias. Robust segmentation of brain mri in the wild with hierarchical cnns and no retraining. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 538-548. Springer, 2022. 5", + "[7] Avrim Blum and Tom Mitchell. Combining labeled and unlabeled data with co-training. In Proceedings of the eleventh annual conference on Computational learning theory, pages 92-100, 1998. 21", + "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 21", + "[9] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 21", + "[10] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 1, 3, 21" + ], + "bbox": [ + 78, + 114, + 472, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3", + "[12] Paola Cascante-Bonilla, Fuwen Tan, Yanjun Qi, and Vicente Ordonez. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. In Proceedings of the AAAI Conference on Artificial Intelligence, 2021. 21", + "[13] Leo Anthony Celi, Jacqueline Cellini, Marie-Laure Charpignon, Edward Christopher Dee, Franck Dernoncourt, Rene Eber, William Greig Mitchell, Lama Moukheiber, Julian Schirmer, et al. Sources of bias in artificial intelligence that perpetuate healthcare disparities—A global review. PLOS Digital Health, 1(3), 2022. 8", + "[14] Olivier Chapelle, Bernhard Scholkopf, and Alexander Zien. Semi-supervised learning (chapelle, o. et al., eds.; 2006)[book reviews]. IEEE Transactions on Neural Networks, 20(3):542-542, 2009. 21", + "[15] Chen Chen, Chen Qin, Huaqi Qiu, Cheng Ouyang, Shuo Wang, Liang Chen, Giacomo Tarroni, Wenjia Bai, and Daniel Rueckert. Realistic adversarial data augmentation for mr image segmentation. In Medical Image Computing and Computer Assisted Intervention MICCAI, 2020. 23", + "[16] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 21", + "[17] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 2, 3, 21", + "[18] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 21", + "[19] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 1, 3", + "[20] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 3", + "[21] Yanbei Chen, Massimiliano Mancini, Xiatian Zhu, and Zeynep Akata. Semi-supervised and unsupervised deep visual learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 4", + "[22] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 6, 21", + "[23] Victor Guilherme Turrisi Da Costa, Enrico Fini, Moin Nabi, Nicu Sebe, and Elisa Ricci. solo-learn: A library of self-supervised methods for visual representation learning. J. Mach. Learn. Res., 23(56):1-6, 2022. 4" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22290", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Coen de Vente, Koenraad A. Vermeer, Nicolas Jaccard, He Wang, Hongyi Sun, FIRas Khader, Daniel Truhn, Temirgali Aimshev, Yerkebulan Zhanibekuly, Tien-Dung Le, Adrian Galdran, Miguel Angel Gonzalez Ballester, Gustavo Carneiro, Devika R. G, Hrishikesh P. S, Densen Puthussery, Hong Liu, Zekang Yang, Satoshi Kondo, Satoshi Kasai, Edward Wang, Ashritha Durvasula, Jonathan Heras, Miguel Angel Zapata, Teresa Araujo, Guilherme Aresta, Hrvoje Bogunovic, Mustafa Arikan, Yeong Chan Lee, Hyun Bin Cho, Yoon Ho Choi, Abdul Qayyum, Imran Razzak, Bram van Ginneken, Hans G. Lemij, and Clara I. Sanchez. AIROGS: Artificial Intelligence for RObust Glaucoma Screening Challenge, 2023. 5, 15", + "[25] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 21", + "[26] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.8", + "[27] Linus Ericsson, Henry Gouk, and Timothy M Hospedales. How well do self-supervised models transfer? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5414-5423, 2021. 1, 3, 4", + "[28] Andre Esteva, Brett Kuprel, Roberto A Novoa, Justin Ko, Susan M Swetter, Helen M Blau, and Sebastian Thrun. Dermatologist-level classification of skin cancer with deep neural networks. nature, 542(7639):115-118, 2017. 5", + "[29] Loveleen Gaur, Ujwal Bhatia, NZ Jhanjhi, Ghulam Muhammad, and Mehedi Masud. Medical image-based detection of Covid-19 using deep convolution neural networks. Multimedia systems, 29(3):1729-1738, 2023.", + "[30] Hemant Ghayvat, Muhammad Awais, AK Bashir, Sharnil Pandya, Mohd Zuhair, Mamoon Rashid, and Jamel Nebhen. Ai-enabled radiologist in the loop: novel ai-based framework to augment radiologist performance for Covid-19 chest CT medical image annotation and classification from pneumonia. Neural Computing and Applications, 35(20):14591–14609, 2023.", + "[31] Amirata Ghorbani, David Ouyang, Abubakar Abid, Bryan He, Jonathan H Chen, Robert A Harrington, David H Liang, Euan A Ashley, and James Y Zou. Deep learning interpretation of echocardiograms. NPJ digital medicine, 3(1):10, 2020. 5", + "[32] Priya Goyal, Dhruv Mahajan, Abhinav Gupta, and Ishan Misra. Scaling and benchmarking self-supervised visual representation learning. In Proceedings of the IEEE/cvf International Conference on computer vision, pages 6391-6400, 2019. 4", + "[33] Margherita Grandini, Enrico Bagli, and Giorgio Visani. Metrics for multi-class classification: an overview. arXiv preprint arXiv:2008.05756, 2020. 5", + "[34] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch," + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3", + "[35] Isabelle Guyon, Kristin Bennett, Gavin Cawley, Hugo Jair Escalante, Sergio Escalera, Tin Kam Ho, Núria Macía, Bisakha Ray, Mehreen Saeed, Alexander Statnikov, et al. Design of the 2015 ChaLearn AutoML challenge. In 2015 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2015. 5", + "[36] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[37] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 3, 21", + "[38] Wassily Hoeffding. Probability inequalities for sums of bounded random variables. The collected works of Wassily Hoeffding, pages 409-426, 1994. 23", + "[39] Zhe Huang, Gary Long, Benjamin Wessler, and Michael C Hughes. A new semi-supervised learning benchmark for classifying view and diagnosing aortic stenosis from echocardiograms. In Proceedings of the Machine Learning for Healthcare Conference. PMLR, 2021. 5, 15", + "[40] Zhe Huang, Gary Long, Benjamin S Wessler, and Michael C Hughes. TMED 2: A dataset for semi-supervised classification of echocardiograms. In DataPerf: Benchmarking Data for Data-Centric AI Workshop, 2022. 5, 15", + "[41] Zhe Huang, Mary-Joy Sidhom, Benjamin S Wessler, and Michael C Hughes. Fix-a-step: Semi-supervised learning from uncurated unlabeled data. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics (AISTATS), 2023. 5, 8", + "[42] Zhe Huang, Benjamin S Wessler, and Michael C Hughes. Detecting heart disease from multi-view ultrasound images via supervised attention multiple instance learning. In Machine Learning for Healthcare Conference, pages 285-307. PMLR, 2023. 5, 8", + "[43] Zhe Huang, Xiaowei Yu, Benjamin S Wessler, and Michael C Hughes. Semi-supervised multimodal multi-instance learning for aortic stenosis diagnosis. arXiv preprint arXiv:2403.06024, 2024. 8", + "[44] Zhe Huang, Xiaowei Yu, Dajiang Zhu, and Michael C Hughes. Interlude: Interactions between labeled and unlabeled data to enhance semi-supervised learning. arXiv preprint arXiv:2403.10658, 2024. 8", + "[45] Tongtong Huo, Yi Xie, Ying Fang, Ziyi Wang, Pengran Liu, Yuyu Duan, Jiayao Zhang, Honglin Wang, Mingdi Xue, Songxiang Liu, et al. Deep learning-based algorithm improves radiologists' performance in lung cancer bone metastases detection on computed tomography. Frontiers in Oncology, 13:1125637, 2023. 5", + "[46] Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Label propagation for deep semi-supervised learning." + ], + "bbox": [ + 501, + 90, + 893, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22291", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5070-5079, 2019. 21", + "[47] Longlong Jing and Yingli Tian. Self-supervised visual feature learning with deep neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(11):4037-4058, 2020. 21", + "[48] Jakob Nikolas Kather, Johannes Krisam, Pornpimol Charoentong, Tom Luedde, Esther Herpel, Cleo-Aron Weis, Timo Gaiser, Alexander Marx, Nektarios A Valous, Dyke Ferber, et al. Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study. PLoS medicine, 16(1):e1002730, 2019. 15", + "[49] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 3", + "[50] Byoungjip Kim, Jinho Choo, Yeong-Dae Kwon, Seongho Joe, Seungjai Min, and Youngjune Gwon. Selfmatch: Combining contrastive self-supervision and consistency for semi-supervised learning. arXiv preprint arXiv:2101.06480, 2021. 4", + "[51] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5", + "[52] Durk P Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semi-supervised learning with deep generative models. Advances in neural information processing systems, 27, 2014. 21", + "[53] Abhishek Kumar, Prasanna Sattigeri, and Tom Fletcher. Semi-supervised learning with gans: Manifold invariance with improved inference. Advances in neural information processing systems, 30, 2017. 21", + "[54] Devidas T Kushnure, Shweta Tyagi, and Sanjay N Talbar. Lim-net: Lightweight multi-level multiscale network with deep residual learning for automatic liver segmentation in ct images. Biomedical Signal Processing and Control, 80: 104305, 2023. 5", + "[55] Zhengfeng Lai, Chao Wang, Luca Cerny Oliveira, Brittany N Dugger, Sen-Ching Cheung, and Chen-Nee Chuah. Joint semi-supervised and active learning for segmentation of gigapixel pathology images with cost-effective labeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 591–600, 2021.", + "[56] Zhengfeng Lai, Chao Wang, Henry Gunawan, Sen-Ching S Cheung, and Chen-Nee Chuah. Smoothed adaptive weighting for imbalanced semi-supervised learning: Improve reliability against unknown distribution data. In International Conference on Machine Learning, pages 11828–11843. PMLR, 2022. 5", + "[57] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 21", + "[58] Dong-Hyun Lee. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning at ICML, 2013. 2, 3, 21" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9475-9484, 2021. 3, 4", + "[60] Ali Madani, Remy Arnaout, Mohammad Mofrad, and Rima Arnaout. Fast and accurate view classification of echocardiograms using deep learning. NPJ digital medicine, 1(1):6, 2018. 5", + "[61] Shaobo Min, Xuejin Chen, Hongtao Xie, Zheng-Jun Zha, and Yongdong Zhang. A mutually attentive co-training framework for semi-supervised recognition. IEEE Transactions on Multimedia, 23:899-910, 2020. 21", + "[62] Avital Oliver, Augustus Odena, Colin A Raffel, Ekin Dogus Cubuk, and Ian Goodfellow. Realistic evaluation of deep semi-supervised learning algorithms. Advances in neural information processing systems, 31, 2018. 1, 2, 3, 4, 6, 8, 13, 22, 23", + "[63] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 21", + "[64] Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. the Journal of machine learning research, 12:2825–2830, 2011. 20", + "[65] Guo-Jun Qi and Jiebo Luo. Small data challenges in big data era: A survey of recent progress on unsupervised and semi-supervised methods. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2168-2187, 2020. 1, 4", + "[66] Maithra Raghu, Chiyuan Zhang, Jon Kleinberg, and Samy Bengio. Transfusion: Understanding transfer learning for medical imaging. Advances in neural information processing systems, 32, 2019. 7", + "[67] Kuniaki Saito, Donghyun Kim, and Kate Saenko. Openmatch: Open-set consistency regularization for semi-supervised learning with outliers. arXiv preprint arXiv:2105.14148, 2021. 8", + "[68] Azizi Shekoofeh, Mustafa Basil, Ryan Fiona, Beaver Zachary, Freyberg Jan, Deaton Jonathan, Loh Aaron, Karthikesalingam Alan, Kornblith Simon, Chen Ting, et al. Big self-supervised models advance medical image classification. arXiv preprint arXiv:2101.05224, 2021. 23", + "[69] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 1, 3, 5, 21", + "[70] Ewout W Steyerberg and Yvonne Vergouwe. Towards better clinical prediction models: seven steps for development and an abcd for validation. European Heart Journal, 35(29), 2014. 8", + "[71] Jong-Chyi Su, Zezhou Cheng, and Subhransu Maji. A realistic evaluation of semi-supervised learning for fine-grained classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12966-12975, 2021. 1, 3, 4, 5, 7, 8, 13, 16, 23, 25" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22292", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] Terpeei Suzuki. Consistency regularization for semi-supervised learning with pytorch. https://github.com/perrying/pytorch-consistency-regularization, 2020.13", + "[73] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 3, 21", + "[74] Jesper E Van Engelen and Holger H Hoos. A survey on semi-supervised learning. Machine learning, 109(2):373-440, 2020. 1, 21", + "[75] Diane Wagner, Fabio Ferreira, Danny Stoll, Robin Tibor Schirrmeister, Samuel Müller, and Frank Hutter. On the importance of hyperparameters and data augmentation for self-supervised learning. arXiv preprint arXiv:2207.07875, 2022. 1, 5", + "[76] Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, et al. Usb: A unified semi-supervised learning benchmark for classification. Advances in Neural Information Processing Systems, 35:3938-3961, 2022. 2, 3, 4, 6", + "[77] Benjamin S Wessler, Zhe Huang, Gary M Long Jr, Stefano Pacifici, Nishant Prashar, Samuel Karmiy, Roman A Sandler, Joseph Z Sokol, Daniel B Sokol, Monica M Dehn, et al. Automated detection of aortic stenosis using machine learning. Journal of the American Society of Echocardiography, 36(4): 411-420, 2023. 5", + "[78] Nan Wu, Jason Phang, Jungkyu Park, Yiqiu Shen, Zhe Huang, Masha Zorin, Stanisław Jastrzebski, Thibault Févry, Joe Katsnelson, Eric Kim, et al. Deep neural networks improve radiologists' performance in breast cancer screening. IEEE transactions on medical imaging, 39(4):1184-1194, 2019. 5, 8", + "[79] Jiancheng Yang, Rui Shi, and Bingbing Ni. Medmnist classification decathlon: A lightweight automl benchmark for medical image analysis. In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), pages 191-195. IEEE, 2021. 14", + "[80] Jiancheng Yang, Rui Shi, Donglai Wei, Zequan Liu, Lin Zhao, Bilian Ke, Hanspeter Pfister, and Bingbing Ni. Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data, 10(1):41, 2023. 4, 14, 15", + "[81] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Proceedings of the British Machine Vision Conference (BMVC), 2016. 5", + "[82] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 3", + "[83] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1476-1485, 2019. 4", + "[84] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curricu" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "lum pseudo labeling. Advances in Neural Information Processing Systems, 34:18408-18419, 2021. 3", + "[85] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 3, 6, 21", + "[86] Wenqiao Zhang, Lei Zhu, James Hallinan, Shengyu Zhang, Andrew Makmur, Qingpeng Cai, and Beng Chin Ooi. BoostMIS: Boosting medical image semi-supervised learning with adaptive pseudo labeling and informative active annotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 23", + "[87] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14471-14481, 2022. 4", + "[88] Xiaojin Zhu. Semi-Supervised Learning Literature Survey. Technical Report 1530, Department of Computer Science, University of Wisconsin Madison., 2005. 1, 21", + "[89] Xiaojin Zhu, Zoubin Ghahramani, and John D Lafferty. Semi-supervised learning using gaussian fields and harmonic functions. In Proceedings of the 20th International conference on Machine learning (ICML-03), pages 912–919, 2003. 21", + "[90] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2020. 21" + ], + "bbox": [ + 503, + 92, + 893, + 486 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "22293", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/a33fa80f-bfa4-45b2-878d-abb01f3d724e_model.json b/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/a33fa80f-bfa4-45b2-878d-abb01f3d724e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..28e42d90d4895411370ede9dc607cd1e743b9496 --- /dev/null +++ b/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/a33fa80f-bfa4-45b2-878d-abb01f3d724e_model.json @@ -0,0 +1,2523 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.131, + 0.844, + 0.177 + ], + "angle": 0, + "content": "Systematic comparison of semi-supervised and self-supervised learning for medical image classification" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.204, + 0.744, + 0.24 + ], + "angle": 0, + "content": "Zhe Huang* Ruijie Jiang* Shuchin Aeron Michael C. Hughes Tufts University, School of Engineering" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.243, + 0.776, + 0.258 + ], + "angle": 0, + "content": "{Zhe.Huang, Ruijie.Jiang, Shuchin.Aeron, Michael.Hughes}@tufts.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.324, + 0.473, + 0.779 + ], + "angle": 0, + "content": "In typical medical image classification problems, labeled data is scarce while unlabeled data is more available. Semi-supervised learning and self-supervised learning are two different research directions that can improve accuracy by learning from extra unlabeled data. Recent methods from both directions have reported significant gains on traditional benchmarks. Yet past benchmarks do not focus on medical tasks and rarely compare self- and semi- methods together on an equal footing. Furthermore, past benchmarks often handle hyperparameter tuning suboptimally. First, they may not tune hyperparameters at all, leading to underfitting. Second, when tuning does occur, it often unrealistically uses a labeled validation set that is much larger than the training set. Therefore currently published rankings might not always corroborate with their practical utility. This study contributes a systematic evaluation of self- and semi- methods with a unified experimental protocol intended to guide a practitioner with scarce overall labeled data and a limited compute budget. We answer two key questions: Can hyperparameter tuning be effective with realistic-sized validation sets? If so, when all methods are tuned well, which self- or semi-supervised methods achieve the best accuracy? Our study compares 13 representative semi- and self-supervised methods to strong labeled-set-only baselines on 4 medical datasets. From \\(20000+\\) GPU hours of computation, we provide valuable best practices to resource-constrained practitioners: hyperparameter tuning is effective, and the semi-supervised method known as MixMatch delivers the most reliable gains across 4 datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.808, + 0.26, + 0.825 + ], + "angle": 0, + "content": "1. INTRODUCTION" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.835, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Deep neural networks can deliver exceptional performance on classification tasks when trained with vast labeled" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.293, + 0.895, + 0.431 + ], + "angle": 0, + "content": "datasets. However, in medical imaging applications assembling a large dataset with appropriate label can be prohibitively costly due to manual effort required by a human expert. In contrast, images alone, without labels, are often readily available in health records databases. In recent years, significant research has focused on developing methods that leverage both the large unlabeled set and a small labeled set to enhance image classifier training, aiming to surpass models trained solely on labeled data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.895, + 0.689 + ], + "angle": 0, + "content": "Two ways of leveraging unlabeled data are particularly popular: semi-supervised and self-supervised learning, both often abbreviated as SSL. Recent efforts in semisupervised learning [74, 88] usually train deep classifiers jointly [5, 69] using an objective with two loss terms, one favoring labeled-set accuracy and the other favoring label-consistency or label-smoothness. Alternatively, self-supervised methods [65] take a two-stage approach, first training deep representations on the unlabeled set, then fine-tuning a classifier on the labeled set. Exemplars are numerous [10, 17, 19, 20, 37]. Despite the remarkable progress reported in each direction, these two paradigms have been largely developed independently [21]. A direct comparison of the two paradigms has been notably absent. A practitioner building a medical image classifier from limited labeled data may ask, \"Which recent semi- or self-supervised methods are likely to be most effective?\"" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Performance can be quite sensitive to hyperparameters for both semi-SL [69, 71] and self-SL [75]. We argue that any careful comparison must consider tuning hyperparameters of all methods in a fair fashion. However, recent work has not handled this well. First, as reviewed in Table 1, recent benchmarks for both semi- and self- paradigms often omit any tuning of hyperparameters (see the Fungi semi-SL experiments of Su et al. [71] or Ericsson et al. [27]'s self-SL benchmark). This practice of using off-the-shelf defaults likely leads to under-performing given a new task, and may impact some methods more than others. An even bigger issue with prevailing semi-SL practice was originally raised by Oliver et al. [62]: many published papers perform hyperparameter tuning on a labeled validation set that" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.468, + 0.903 + ], + "angle": 0, + "content": "*Authors Zhe Huang and Ruijie Jiang contributed equally to this work. \nCode: github.com/tufts-ms/SSL-vs-SSL-benchmark [MIT license]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22282" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "is larger (sometimes much larger) than the limited labeled train set. Such experimental settings are unrealistic. SSL is intended for practitioners without abundant available labeled data. Practitioners that need to make the most of 1000 available labeled images will not elect to put more images in validation than training, and thus are not helped by benchmarks that do. Unfortunately, five years later after [62] we find this issue is still widespread (see Tab. 1). For example, Wang et al. [76]'s semi-SL results on TissueMNIST tune on a validation set over \\(50\\mathrm{x}\\) larger than the labeled train set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.365 + ], + "angle": 0, + "content": "Oliver et al. [62] further cast some doubt on whether effective tuning is possible with small labeled sets, saying \"Extensive hyperparameter tuning may be somewhat futile due to an excessively small collection of held-out data to measure performance on\". There thus remains a pressing question for resource-constrained practitioners: \"Given limited available labeled data and limited compute, is hyperparameter tuning worthwhile?\"" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.473, + 0.606 + ], + "angle": 0, + "content": "This study makes progress toward answering these questions by delivering a comprehensive comparison of semi- and self-supervised learning algorithms under a resource-constrained scenario that matches how SSL might be used on real-world medical tasks. Our goal is to enable practitioners to tackle new image classification challenges where only modest-sized labeled datasets exist. We target settings with roughly 2-10 class labels of interest, where each class has 30-1000 available labeled images for all model development (including training and validation). We select representative tasks across 4 datasets that span low-resolution (28x28) to moderate-resolution (112x112 and 384x384). On each task, we run careful experiments with representative recent methods from both semi- and self-supervised learning to clarify what gains are possible with unlabeled data and how to achieve them." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.473, + 0.744 + ], + "angle": 0, + "content": "We emphasize realism throughout, in 4 distinct ways: (1) using validation set sizes that are never bigger than the train set, avoiding the unrealistically-large validation sets common in past work (Tab. 1); (2) performing hyperparameter search using the same protocol, same compute budget (fixed number of hours)* and same hardware (one NVIDIA A100 GPU) for all algorithms for fair comparison; (3) respecting natural class imbalance; (4) profiling performance over time, to inform labs with smaller runtime budgets." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.745, + 0.411, + 0.76 + ], + "angle": 0, + "content": "In summary, the contributions of this study are:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.76, + 0.47, + 0.806 + ], + "angle": 0, + "content": "1. We provide a systematic comparison of semi- and self-supervised methods on equal footing to connect two research directions that have been heretofore separate." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.806, + 0.471, + 0.852 + ], + "angle": 0, + "content": "2. We adopt a realistic experimental protocol designed to consider the same constraints on available labels and runtime that SSL practitioners face, avoiding the unre" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.76, + 0.471, + 0.852 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "alistic aspects of past benchmarks, especially the use of far too large validation sets." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.122, + 0.894, + 0.166 + ], + "angle": 0, + "content": "3. We show that hyperparameter tuning is viable with a realistic-sized validation set, and in many cases necessary to do well at a new classification task with new data." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.167, + 0.894, + 0.213 + ], + "angle": 0, + "content": "Ultimately, we hope this study guides practitioners with limited data toward successful deployment of semi- and self-supervised methods on real problems." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.225, + 0.819, + 0.24 + ], + "angle": 0, + "content": "2. BACKGROUND AND METHODS" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.895, + 0.432 + ], + "angle": 0, + "content": "Unified Problem Formulation. Following the recent survey by Chen et al. [21], we adopt a unified perspective for supervised, semi-supervised and self-supervised image classification with limited available labeled data. For model development (including training and hyperparameter selection), we assume there are two available datasets. First, a small labeled dataset \\(\\mathcal{L}\\) of feature-label pairs \\((x,y)\\), where each image is represented by a \\(D\\)-dimensional feature vector \\(x \\in \\mathbb{R}^D\\) and its corresponding class label takes one of \\(C\\) possible values: \\(y \\in \\{1,2,\\dots,C\\}\\). Second, an unlabeled set \\(\\mathcal{U}\\) containing only feature vectors. Typically, we assume the unlabeled set is much larger: \\(|\\mathcal{U}| \\gg |\\mathcal{L}|\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.894, + 0.568 + ], + "angle": 0, + "content": "Given labeled set \\(\\mathcal{L}\\) and unlabeled set \\(\\mathcal{U}\\), we wish to train a neural network that can map each input \\(x\\) to a probability vector in the \\(C\\)-dimensional simplex \\(\\Delta^C\\) representing a distribution over \\(C\\) class labels. Let \\(f_v(\\cdot): \\mathbb{R}^D \\to \\mathbb{R}^F\\) denote a backbone neural network with parameters \\(v\\) producing an \\(F\\)-dimensional embedding given any input image. Let \\(g_w(\\cdot): \\mathbb{R}^F \\to \\Delta^C\\) denote a final linear-softmax classification layer with parameters \\(w\\). The following unified objective can capture all three learning paradigms:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.575, + 0.892, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} v ^ {*}, w ^ {*} \\leftarrow \\arg \\min _ {v, w} \\sum_ {x, y \\in \\mathcal {L}} \\lambda^ {L} \\ell^ {L} (y, g _ {w} (f _ {v} (x))) \\tag {1} \\\\ + \\sum_ {x \\in \\mathcal {U}} \\lambda^ {U} \\ell^ {U} (x, f _ {v}, g _ {w}). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Here, \\(\\ell^L\\) represents a labeled-set loss (e.g. multi-class cross-entropy), and \\(\\ell^U\\) represents a unlabeled-set loss. \\(\\lambda^L, \\lambda^U \\geq 0\\) are weights for the corresponding loss terms. The design of \\(\\ell^U\\) is usually what differs substantially across methods. For instance, setting \\(\\ell^U\\) to be cross-entropy computed with pseudo-labels generated from classifier \\(g\\) recovers PseudoLabel [58]; a temperature-scaled instance-similarity contrastive loss recovers SimCLR [17]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.856 + ], + "angle": 0, + "content": "All three learning paradigms that we study optimize Eq. (1), yet differ in the number of phases and in how to set the scalar weights \\(\\lambda^L, \\lambda^U\\) on each loss term. Supervised learning ignores the unlabeled term throughout training (\\(\\lambda^U = 0\\)), thus learning parameters using only the labeled set. Semi-supervised methods include both terms in one end-to-end training, keeping both \\(\\lambda^U > 0\\) and \\(\\lambda^L > 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Self-supervised learning has two phases. In phase 1 (\"pretraining\"), the labeled term is omitted (\\(\\lambda^L = 0\\), \\(\\lambda^U = 1\\)) and the focus of learning is an effective representation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.471, + 0.901 + ], + "angle": 0, + "content": "*Oliver et al. [62] give each method 1000 trials of a cloud hyperparameter tuning service. However, training speed can differ substantially across SSL methods. We argue a fixed wallclock time budget is more fair." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22283" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.088, + 0.898, + 0.23 + ], + "angle": 0, + "content": "
BenchmarkMethodsUnrealistic ExperimentsLabeled train sizeLabeled val. sizeAcc vs. Time?
Realistic eval. SSL [a]SemiCIFAR-10 (Tab. 1-2, Fig. 2)40005000no
SVHN (Tab. 1-2, Fig. 3-4)10007325
Fine-grained SSL [b]Semi & 1 Self*Semi-Aves59598000no
Semi-Fungi4141None
USB [c]SemiTissueMNIST80/40023640†no
Semi-Aves5959None
Self benchmark [d]SelfImageNet1.28 mil.Noneno
SSL-vs-SSL (ours)Semi & Selfno400-1660no bigger than trainyes
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.241, + 0.894, + 0.325 + ], + "angle": 0, + "content": "Table 1. Comparison of related benchmarks of semi-supervised and self-supervised learning. Past works either do no hyperparameter tuning at all (val. size = None), or use an unrealistically large validation set (defined as larger than the labeled train set) in some or all experiments. Further, each work almost exclusively looks at methods from one paradigm, either semi- or self- (*: [b] includes one self-supervised method, MOCO). In contrast, in this paper we benchmark 6 semi- and 7 self-supervised algorithms with hyperparameter tuning on realistic validation sets. Acc vs. Time? indicates whether the work analyzes performance over training time. Number marked † confirmed via GitHub comment by USB authors. Citations: a: Oliver et al. [62], b: Su et al. [71] c: Wang et al. [76] d: Ericsson et al. [27]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.347, + 0.47, + 0.408 + ], + "angle": 0, + "content": "layer \\( f_{v} \\) (classifier \\( g_{w} \\) is not included in this phase). In phase 2 (\"fine-tuning\"), we focus on the labeled term and omit the unlabeled term \\( (\\lambda^{L} = 1, \\lambda^{U} = 0) \\). We fix the representation parameter \\( v \\) and fine-tune the classifier \\( w \\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.411, + 0.429, + 0.425 + ], + "angle": 0, + "content": "We now identify the 16 methods we will evaluate:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.47, + 0.563 + ], + "angle": 0, + "content": "Supervised methods. The goal of leveraging unlabeled data \\(\\mathcal{U}\\) is to obtain better performance than what we could obtain using only the labeled set \\(\\mathcal{L}\\). Therefore, we naturally compare to 3 high-quality supervised baselines that use only the labeled set. First, \"Sup\" denotes a classifier trained with supervised loss \\(\\ell^L\\) set to multi-class cross-entropy. Second, \"MixUp\" trains with cross entropy with the addition of mixup data augmentation [85]. Finally, \"SupCon\" pursues a supervised contrastive learning loss for \\(\\ell^L\\) [49]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.732 + ], + "angle": 0, + "content": "Semi-supervised methods. We compare 6 semi-supervised methods that train deep classifiers on both labeled and unlabeled data simultaneously as in Eq. (1). To represent the state-of-the-art of semi-supervised image classification, we select Pseudo Label (\"PseudoL\") [58], Mean Teacher (\"MeanTch\") [73], MixMatch [5], FixMatch [69], FlexMatch [84] and CoMatch [59]. These choices cover a reasonably wide spectrum of unlabeled loss design strategy, year of publication, and computation cost. CoMatch represents a recent trend of combining semi- and self-supervision. See App D.2 for a wider literature review." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Self-supervised methods. We compare 7 self-supervised algorithms: SimCLR [17], MOCO (v2) [20, 37], SwAV [10], BYOL [34], SimSiam [19], DINO [11] and Barlow Twins (\"BarlowTw\") [82]. These algorithms epitomize the field of self-supervised learning as of this writing. SimCLR, MOCO (v2), and SwAV are based on contrastive learning, which learns effective representations when provided with similar and dissimilar samples. BYOL was designed to circumvent the need for dissimilar samples. SimSiam is a simple yet effective Siamese representation learning method. DINO uses a teacher-student network architec" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.347, + 0.892, + 0.408 + ], + "angle": 0, + "content": "ture for representation distillation. Barlow Twins preserves information and reduces redundancy by favoring a close-to-identity cross-correlation matrix between augmented pairs of data. See App D.3 for further review." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.422, + 0.688, + 0.437 + ], + "angle": 0, + "content": "3. RELATED WORK" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Table 1 summarizes key attributes of existing major benchmarks for semi-supervised and self-supervised methods. We now discuss how our study situates in this context." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.893, + 0.764 + ], + "angle": 0, + "content": "Comparison to Oliver et al. Oliver et al. [62] provide an influential benchmark of deep semi-supervised methods. Like our work, they highlight the pressing issue of using unrealistically large validation sets for hyperparameter tuning. However, most of their experiments (e.g. all their tables and their figures 2-4) still use this unrealistic setting to be comparable to other works; only one subsection (Sec. 4.6) examines validation sets no larger than the train set. Instead, we exclusively use validation set sizes no larger than training set, mimicking what practitioners would face in real applications. Our definition of a \"realistic\" size for a validation set – no larger than the labeled train set – is broader than Oliver et al.'s definition of \\(10\\%\\) of the train set. We favor our definition because Oliver et al. find specifically that \"for validation sets of the same size \\((100\\%)\\) as the training set, some differentiation between the approaches is possible.\" In contrast, they found reliable differentiation was not always possible with much smaller validation sets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.885 + ], + "angle": 0, + "content": "Furthermore, the benchmarking results presented by Oliver et al. require each algorithm to complete \"1000 trials of Gaussian Process-based black-box optimization using Google Cloud ML Engine\". This resource-intensive process seems impractical for researchers outside well-funded industrial labs. We use more modest compute budgets and specifically allow each method the same total wallclock runtime for tuning, again obeying practical constraints." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Other semi-supervised benchmarks. TorchSSL [84]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22284" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "benchmarked eight popular semi-supervised learning algorithm using a unified codebase. The same group later extended that effort to natural language and audio processing in their USB benchmark [76]. Another recent effort for fine-grained classification by Su et al. [71] evaluates semi-supervised learning on datasets that exhibit class imbalance and contains images from novel classes in the unlabeled set, studying the effect of different initializations and the contents of unlabeled data on the performance of semi-supervised methods. While making significant contributions, these works focused on semi-supervision almost exclusively and did not include multiple self-supervised learning methods, thus leaving open the questions \"How do the two paradigms compare? Which methods are best overall?\"" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.474, + 0.514 + ], + "angle": 0, + "content": "Prior self-supervised benchmarks. Several works have proposed different benchmarks for self-supervised methods. Goyal et al. [32] introduced a benchmark that covers 9 different tasks, such as object detection and visual navigation. Ericsson et al. [27] compared thirteen top self-supervised models on 40 downstream tasks. Da Costa et al. [23] presented a library of self-supervised methods that can be easily plugged into different downstream tasks and datasets. However, these works mainly consider other self-SL methods without direct comparison to the semi-SL paradigm. Moreover, they mostly do not incorporate the use of a validation set for hyperparameter tuning at all. This can lead to suboptimal accuracy at test time and complicate comparisons to methods that do tune." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.474, + 0.65 + ], + "angle": 0, + "content": "Self-supervision for medical images. Our work is complementary to recent efforts that assess self-supervised pipelines for medical image classification [2, 3]. They focus primarily on how to design multi-stage transfer learning pipelines and do not comprehensively compare many different self-supervised methods or any semi-supervised methods. Further, many datasets studied by [3] come from proprietary projects conducted at Google. In contrast, all our data and experiments are open and reproducible by others." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.473, + 0.742 + ], + "angle": 0, + "content": "Combining semi- and self-supervision. Recent research has explored the fusion of semi-supervised learning and self-supervised learning ideas [50, 59, 83, 87]. Some recent surveys [21, 65] compare semi- and self-supervised learning, but they focus on literature review while we offer realistic and comprehensive benchmarking experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.317, + 0.77 + ], + "angle": 0, + "content": "4. DATASETS AND TASKS" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.856 + ], + "angle": 0, + "content": "We study four open-access medical image classification datasets, all with 2D images that are fully deidentified. Table 2 reports statistics for all train/test splits. The exact splits used in our study can be found in our open-source codebase, documented in App. A." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Two datasets - PathMNIST and TissueMNIST - are selected from the MedMNIST collection [80] (criteria in App. B.2). Prior experiments by Yang et al. [80] suggest" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.089, + 0.891, + 0.392 + ], + "angle": 0, + "content": "
TissueMNISTPathMNIST
LabeledUnlab.LabeledUnlab.
TrainValTestTrainTrainValTestTrain
total40040047280165066total450450718089546
GE151516775851NORM39397417847
DCT191922337795MUC404010357966
POD191922027686ADI474713389319
LEU2828336911761STR47474219354
IE3737440215369BACK48488479461
TAL5959703124549DEB525233910308
PT95951120139108LYM525264310349
CD/CT1281281516552947MUS616159212121
TUM6464123312821
TMED-2AIROGS
LabeledUnlab.LabeledUnlab.
TrainValTestTrainTrainValTestTrain
total16602352019353500total600600600094242
PSAX22350342-Glaucoma6060600-
A2C32528319-No Glauc.5405405400-
A4C46239423-
PLAX650118935-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.404, + 0.895, + 0.462 + ], + "angle": 0, + "content": "Table 2. Summary statistics of train/validation/test splits for all datasets in our study. Each table's rows are arranged in ascending order based on the number of per-class labeled train images. Full description of class names can be found in App B.3." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.489, + 0.892, + 0.534 + ], + "angle": 0, + "content": "their \\(28 \\times 28\\) resolution is a reasonable choice for rapid prototyping; using larger \\(224 \\times 224\\) resolution does not yield much more accurate classifiers for these two datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.894, + 0.612 + ], + "angle": 0, + "content": "Two other datasets represent more moderate resolutions closely tied to contemporary clinical research: the 112x112 Tufts Medical Echocardiogram Dataset (TMED-2) and the 384x384 AIROGS dataset. Further details for each dataset are provided in a dedicated paragraph below." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.895, + 0.839 + ], + "angle": 0, + "content": "For each dataset, our data splitting strategy closely mirrors the conditions of real SSL applications. First, we let labeled training and validation sets contain a natural distribution of classes even if that may be imbalanced. This reflects how data would likely be affordably collected (by random sampling) and avoids artificially balanced training sets that will not match the population an algorithm would encounter in a deployment. Second, to be realistic we ensure validation sets are never larger than the available labeled train sets. This is in contrast to previous benchmarks: Wang et al. [76]'s TissueMNIST hyperparameter search used a validation set of 23,640 images even though the labeled training set contained only 80 or 400 images. In practice, such a large validation set would almost certainly be re-allocated to improve training, as noted in Oliver et al. [62]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "TissueMNIST is a lightweight dataset of 28x28 images of human kidney cortex cells, organized into 8 categories. The original dataset is fully-labeled and distributed with predefined train/val/test split of 165,466/23,640/47,280 im" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22285" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "ages with some class imbalance. We assume a total labeling budget of 800 images, evenly split between training and validation (to ensure hyperparameter tuning can differentiate methods). We form a labeled training set of 400 images from the predefined training set, sampling each class by its frequency in the original training set. We form a labeled validation set of 400 images sampled from the predefined validation set. For unlabeled set, we keep all remaining images in the original training split, discarding known labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.235, + 0.473, + 0.388 + ], + "angle": 0, + "content": "PathMNIST is another lightweight dataset of 28x28 images of patches from colorectal cancer histology slides that comprise of 9 tissue types. The original dataset is fully-labeled and distributed with predefined train/val/test split of 89,996/10,004/7,180 images. Class imbalance is less severe than TissueMNIST. We assume a total labeled data budget of 900 images, again evenly split between training and validation. Labeled train, labeled valid, and unlabeled sets are sampled via the same procedures as in TissueMNIST. For both MedMNIST datasets, we use the predefined test set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.473, + 0.653 + ], + "angle": 0, + "content": "TMED-2 [39, 40] is an open-access dataset of \\(112 \\times 112\\) 2D grayscale images captured from routine echocardiogram scans. Each scan produces dozens of ultrasound images of the heart, captured from multiple acquisition angles (i.e., different anatomic views). In this study, we adopt Huang et al.'s view-classification task: the goal is to classify each image into one of 4 canonical view types (see App. B.3). View classification is clinically important: measurements or diagnoses of heart disease can only be made when looking at the right anatomical view [60, 77]. We used the provided train/validation/test split (id #1), which is naturally imbalanced and matches our criteria for realistic sizing. We further use the provided large unlabeled set of 353,500 images. TMED-2's unlabeled set is both authentic (no true labels are available at all, unlike other benchmarks that \"forget\" known labels) and uncurated [41] (contains images of view types beyond the 4 classes in the labeled task)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.902 + ], + "angle": 0, + "content": "AIROGS [24] is a public dataset released for a recent competition where the binary classification task is to decide whether the patient should be referred for glaucoma or not, given a color fundus image of the retina (eye). We selected this dataset because it represents an active research challenge even with all available labels. Furthermore, because labels for this dataset were acquired at great expense (multiple human annotators graded each of over 100k images), we hope our work helps assess how self-/semi-supervision could reduce the annotation costs of future challenges. We selected the \\(384 \\times 384\\) resolution favored by several high-performing competition entries. We chose a labeling budget of 1200 total images, evenly split between train and validation; the rare positive class (9 to 1 imbalance) is representative of many screening tasks. We took the rest of available images as the unlabeled train set." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.09, + 0.765, + 0.107 + ], + "angle": 0, + "content": "5. EXPERIMENTAL DESIGN" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.238 + ], + "angle": 0, + "content": "Performance metric. We use balanced accuracy [33, 35] as our primary performance metric. For a task with \\( C \\geq 2 \\) classes, let \\( y_{1:N} \\) denote true labels for \\( N \\) examples in a test set, and \\( \\hat{y}_{1:N} \\) denote a classifier's predicted labels. Let \\( \\mathrm{TP}_c \\) count true positives for class \\( c \\) (number of correctly classified examples whose true label is \\( c \\)), and let \\( N_c \\) count the total number of examples with true label \\( c \\). Then we compute balanced accuracy (BA) as a percentage:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.246, + 0.892, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\mathbf{BA}(y_{1:N},\\hat{y}_{1:N}) = \\frac{1}{C}\\sum_{c = 1}^{C}\\frac{\\mathrm{TP}_{c}(y_{1:N},\\hat{y}_{1:N})}{N_{c}(y_{1:N})}\\cdot 100\\% \\tag{2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.892, + 0.342 + ], + "angle": 0, + "content": "Balanced accuracy is more suitable than standard accuracy for imbalanced problems when each class matters equally. The expected BA of a uniform random guess is \\(\\frac{100}{C}\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.892, + 0.387 + ], + "angle": 0, + "content": "On AIROGS, we also track metrics recommended by its creators [24] for clinical utility: AUROC, partial AUROC (\\(>90\\%\\) specificity) and sensitivity-at- \\(95\\%\\) -specificity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.892, + 0.463 + ], + "angle": 0, + "content": "Architectures. CNN backbones are popular in medical imaging [6, 28-31, 42, 45, 54-56, 77, 78]. We use ResNet-18 [36] on Tissue and Path, and WideResNet-28-2 [81] on TMED-2. We experiment with both ResNet-18 and 50 [36] on AIROGS to assess architectural differences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.599 + ], + "angle": 0, + "content": "Training with early stopping. For each training phase, we perform minibatch gradient descent on Eq. (1) with Adam [51] optimizer and a cosine learning rate schedule [69]. Each training phase proceeds for up to 200 epochs, where one epoch represents enough minibatch updates to process the equivalent of the entire combined training set \\(\\mathcal{L} \\cup \\mathcal{U}\\). After every epoch, we record balanced accuracy on the validation set. If this value plateaus for 20 consecutive epochs, we stop the current training phase early." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.689 + ], + "angle": 0, + "content": "Hyperparameters. Semi- and self-supervised learning can both be sensitive to hyperparameters [71, 75]. Our evaluations tune both shared variables (e.g. learning rates, weight decay, unlabeled loss weight \\(\\lambda^U\\)) as well as hyperparameters unique to each algorithm. See App. F.1 for details on all hyperparameters for all methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Unified procedure for training and hyperparameter tuning. We formulate a unified procedure mindful of realistic hardware and runtime constraints for a non-industrial-scale lab working on a new medical dataset. We assume each algorithm has access to one NVIDIA A100 GPU for a fixed number of hours. Within the allotted compute budget, for each algorithm we execute a serial random search over hyperparameters, sequentially sampling each new configuration and then training until either an early stopping criteria is satisfied or the maximum epoch is reached. We track the best-so-far classifier in terms of validation-set performance every epoch. Algorithm D.1 provides pseudocode for the procedure. Each time a new hyperparameter configuration is needed, we sample each hyperparameter value independ-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22286" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "dently from a distribution designed to cover its common settings in prior literature (App. F.1). Our choice of random search on a budget is thought to yield better performance than grid search [4], fairly expends the same effort to train all methods regardless of their cost-per-epoch or hyperparameter complexity, and does not assume industrial-scale access to 1000 cloud-computing trials as in Oliver et al. [62]. We find that performance saturates after about 25 hours per 80000 unlabeled examples. We thus allocate for the total time budget 25 hours for PathMNIST, 50 hours for TissueMNIST, and 100 hours for TMED-2. We allow 100 hours for AIROGS due to its larger resolution. Due to limited resources, we select only a few competitive methods to run on AIROGS based on the results on other datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.473, + 0.439 + ], + "angle": 0, + "content": "Self-supervised classification phase. Self-supervised methods by definition do not utilize label information when training representation layer weights \\( v \\). To enable a proper comparison, for all self-SL methods our Alg. D.1 introduces an additional classification layer, training it anew after each epoch, each time using a 10 trial random search for an L2-penalty regularization hyperparameter. We retain the best performing weights \\( w \\) on the validation set. We emphasize that \\( w \\) does not impact overall self-supervised training of \\( v \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.473, + 0.53 + ], + "angle": 0, + "content": "Data augmentation. Random flip and crop are used for all semi-supervised and supervised methods. MixMatch also utilizes MixUp [85], while RandAugment [22] is used by FixMatch and FlexMatch. For each self-supervised method as well as SupCon, we apply the same SimCLR augmentation: random flip, crop, color jitter, and grayscale." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.473, + 0.591 + ], + "angle": 0, + "content": "Multiple trials. We repeat Alg. D.1 for each method across 5 separate trials (distinct random seeds). We record the mean balanced accuracy on valid and test sets of these trials every 60 minutes (every \\(30\\mathrm{min}\\) for Tissue and Path)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.604, + 0.315, + 0.62 + ], + "angle": 0, + "content": "6. RESULTS & ANALYSIS" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.795 + ], + "angle": 0, + "content": "In each panel of Fig. 1, we focus on one dataset and architecture, plotting for each algorithm one line showing test set balanced accuracy (averaged across 5 trials) as a function of time spent running Alg. D.1. Variance across trials is visualized in Fig C.4. Extra results on AIROGS showing other performance metrics over time (AUROC, partial AUROC, and sensitivity at \\(95\\%\\) specificity) are found in Fig. C.3. We emphasize that following best practice, each checkpoint represents the best hyperparameters as selected on the validation set; we then report that checkpoint's test set performance. These results help us answer two key questions:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Is hyperparameter tuning worthwhile for SSL methods when validation set sizes cannot be larger than the train set? Across Fig. 1 and C.3, all 16 algorithms show roughly monotonic improvements in test performance over time, despite using a realistic-sized validation set. The final performance of every algorithm and dataset shows improvement due to our unified training and tuning procedure (fur" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.089, + 0.897, + 0.34 + ], + "angle": 0, + "content": "
nameGain in Bal.Acc. over best labeled-set-only
PathTissueTMEDAIROGSmedianworst
MixMatch12.43.7-0.25.74.7-0.2
MOCO7.94.4-18.8n/a4.4-18.8
SwAV8.44.3-10.1n/a4.3-10.1
SimCLR11.05.9-7.81.63.8-7.8
BYOL12.16.2-6.00.93.5-6.0
BarlowTw11.22.2-1.9n/a2.2-1.9
SimSiam1.92.1-11.6n/a1.9-11.6
FlexMtech11.14.5-1.1-4.21.7-4.2
DINO12.66.5-3.4-3.81.6-3.8
FixMatch9.32.90.2-4.51.6-4.5
CoMatch14.10.2-3.0n/a0.2-3.0
MeanTch-1.6-1.9-1.2n/a-1.6-1.9
PseudoL0.6-1.4-4.4-3.2-2.3-4.4
ref. val.69.833.994.971.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.35, + 0.895, + 0.463 + ], + "angle": 0, + "content": "Table 3. Gains from SSL methods over only using labeled set, across 4 datasets. Self (italicized), Semi (normal). We report gains in percentage balanced accuracy (higher is better) in the mean final test set performance (averaged over 5 runs of Alg. D.1) over a reference value that represents the best labeled-set-only run among Sup (minimize cross entropy), SupCon, and MixUp. MixMatch has the highest median gain, and is the only method to never score notably worse (\\(>1\\) point) than the best labeled set only run." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.892, + 0.536 + ], + "angle": 0, + "content": "ther discussion in App. E.1). Therefore, we answer: Yes, realistically-sized validation sets for hyperparameter tuning and checkpoint selection can be effective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.895, + 0.657 + ], + "angle": 0, + "content": "What are the best SSL methods? Among the many methods we study in Fig. 1, no method clearly outperforms others across all datasets. On Path, CoMatch, MixMatch, DINO, and BYOL perform best. On Tissue, self-supervised methods like DINO, BYOL, and SimCLR score well. On TMED-2, ultimately only FixMatch and MixMatch are competitive with the MixUp baseline. On AIROGS, only MixMatch, SimCLR and BYOL beat the Sup baseline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.896, + 0.869 + ], + "angle": 0, + "content": "For each semi- and self- SSL method, Tab. 3 reports its relative gain in balanced accuracy over the best labeled-set-only supervised baseline on each dataset. We conclude that MixMatch represents the best overall choice as it consistently performs at or near the top across all tasks. It is the only method to never deliver results worse by more than 1 percentage point than the best labeled-set-only baseline (see \"worst\" column of Tab. 3). This is in stark contrast to USB [76], where FixMatch and FlexMatch are ranked notably higher than MixMatch. Such differences stress the importance of choosing proper evaluation protocol tailored to specific needs. We do caution that hyperparameter tuning is strongly recommended for MixMatch to succeed on a new dataset (see AIROGS result in Tab. 4)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Pretraining vs. from scratch. Plots in App. C show accuracy-over-time profiles using initial weights pretrained" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22287" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.089, + 0.346, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.089, + 0.616, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.09, + 0.882, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.228, + 0.345, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.228, + 0.613, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.669, + 0.23, + 0.862, + 0.251 + ], + "angle": 0, + "content": "semi labeled-set-only self FixMatch SupCon SimCLR" + }, + { + "type": "image_footnote", + "bbox": [ + 0.669, + 0.253, + 0.848, + 0.266 + ], + "angle": 0, + "content": "-FlexMtech-MixUp BYOL" + }, + { + "type": "image_footnote", + "bbox": [ + 0.669, + 0.269, + 0.845, + 0.282 + ], + "angle": 0, + "content": "-MixMatch..Sup -DINO" + }, + { + "type": "image_footnote", + "bbox": [ + 0.669, + 0.286, + 0.869, + 0.298 + ], + "angle": 0, + "content": "—CoMatch —BarlowTw" + }, + { + "type": "image_footnote", + "bbox": [ + 0.669, + 0.302, + 0.848, + 0.313 + ], + "angle": 0, + "content": "—MeanTch —SwAV" + }, + { + "type": "image_footnote", + "bbox": [ + 0.669, + 0.318, + 0.866, + 0.33 + ], + "angle": 0, + "content": "-PseudoL -SimSiam" + }, + { + "type": "image_footnote", + "bbox": [ + 0.669, + 0.333, + 0.854, + 0.345 + ], + "angle": 0, + "content": "—MOCO" + }, + { + "type": "list", + "bbox": [ + 0.669, + 0.23, + 0.869, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.381, + 0.895, + 0.452 + ], + "angle": 0, + "content": "Figure 1. Balanced accuracy over time profiles of semi- and self-supervised methods across all 4 datasets. At each time, we report the test set bal. acc. of each method (mean over 5 trials of Alg. D.1). Best viewed electronically. Top Row: On these 3 datasets, we compare all 6 semi- and 7 self- methods (see legend in lower right, citations in Sec. 2), to 3 labeled-set-only baselines. Bottom Row: On larger AIROGS dataset, we compare selected methods representative of the best in the top row. Thin lines show final performance to ease comparison. From these charts, we suggest that our unified training and tuning (Alg. D.1) is effective, as all methods show gains over time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.47, + 0.583 + ], + "angle": 0, + "content": "on ImageNet. Compared to training from scratch, we see only slight gains (less than 3 points of BA) in ultimate test-set accuracy when top methods use pretraining, which aligns with observations in [66] but counters the \"large margin\" gains (sometimes over 10 points) reported by Su et al. [71]. Pretraining's benefits still include improved convergence speed and reduced variance across trials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.688 + ], + "angle": 0, + "content": "ResNet-18 vs. ResNet-50. The bottom row of Fig 1 compares ResNet-18 and ResNet-50 architectures on AIROGS. We broadly find that the method rankings are similar. More notably, ResNet-50 is not substantially better than ResNet-18 in this task when compared using the same runtime budget. Profiles of other performance metrics in App C.3 suggest the same conclusions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Tuning vs. Transferring from another dataset. To make the most of limited labeled data, one strategy is to use the entire labeled set for training, reserving no validation set at all. This approach relies on pre-established \"best\" hyperparameters sourced from other datasets and transferred to the target task, obviating the need for data-specific tuning. Su et al. [71] employed this strategy in their benchmark, avoiding tuning due to concerns that small validation sets would yield unreliable selection. However, this claim in their work was not supported by experiments. Here, we directly compare our hyperparameter tuning strategy (Alg. D.1) with Su et al.'s transfer strategy. We train the latter on the combined train + validation sets, using \"best\" hyperparameters from CIFAR-10 and Tissue (details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.477, + 0.892, + 0.509 + ], + "angle": 0, + "content": "in App. F.2). Tab. 4 reports the final test-set balanced accuracy of each strategy on Path, TMED-2 and AIROGS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.705 + ], + "angle": 0, + "content": "Our hyperparameter tuning strategy is competitive across all datasets: the \"Tune\" columns of Tab. 4 have the highest fraction of green cells. The alternative transfer strategy is faster and occasionally yield reasonable results. However, there is no guarantee of good performance from hyperparameter transfer, as evidenced by the inferior cases in Tab. 4 highlighted in yellow and red. There is also no consistent distinction between transferring from Tissue (an arguably more related medical dataset) and transferring from CIFAR-10, underscoring the challenge of identifying a \"closely related\" dataset for hyperparameter transfer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Overall, we recommend researchers pursuing SSL for a new medical classification task should adopt our hyperparameter tuning strategy. Transferring off-the-shelf hyperparameters has highly variable performance and can even result in divergence (see MixMatch on AIROGS in Tab. 4). Even with reasonable hyperparameters, having no validation set makes checkpoint selection challenging. Adjacent epochs that exhibit similar low training losses may differ by over \\(5\\%\\) in test accuracy in our experiments. Improving transfer from existing hyperparameters to a new task is an interesting further direction of research." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22288" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.089, + 0.088, + 0.885, + 0.337 + ], + "angle": 0, + "content": "
PathMNISTTMED-2AIROGS
Hyperparam. strategy:TuneBest from TissueBest from CIFAR-10TuneBest from TissueBest from CIFAR-10TuneBest from TissueBest from CIFAR-10
semiMixMatch82.2483.8773.5194.7192.6774.7377.1461.75Diverged
CoMatch83.9580.5186.4291.8792.5192.67
FixMatch79.1182.6479.7395.0695.1295.7366.9266.3165.50
FlexMatch80.9178.7080.2393.7793.9793.7067.2372.8274.52
Pseudo-Label70.4266.1264.7190.5288.7882.1368.2164.4366.93
Mean-Teacher68.2155.9056.1393.7091.2158.26
selfSimCLR81.0579.8680.0587.0484.4187.2473.0167.1567.95
BYOL81.7980.3580.1688.9088.4185.1572.4070.1370.63
SwAV77.9078.2481.3084.7682.1782.20
MoCo77.7379.9579.1276.1478.5679.11
SimSiam71.7868.6270.4483.2480.9380.20
BT80.8572.7980.2592.9691.6390.81
DINO82.5277.5080.6791.4584.8090.0367.6166.4667.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.346, + 0.895, + 0.472 + ], + "angle": 0, + "content": "Table 4. Hyperparameter strategy evaluation. Table reports ultimate test-set balanced accuracy as percentage (mean over 5 trials, higher is better) on three target datasets: PathMNIST, TMED-2, and AIROGS. 3 different hyperparameter strategies are compared for each SSL method. Due to computation constraints, we only select a few algorithms to run on AIROGS, explained in Sec. 5. Tune strategy: tuning for best hyperparameters via our proposed Alg. D.1 for 100 hours (TMED-2 and AIROGS) and 25 hours (PathMNIST), using provided train/validation split. Best from Tissue/CIFAR-10 strategy: just one training phase on the target dataset using the \"best\" hyperparameters selected from the named source data. These runs combine train and validation data for fitting following Su et al. [71], and typically take 1-4/5-30/10-40 hours on Path/TMED-2/AIROGS depending on the algorithm and whether early stopping is triggered. To highlight which strategies are most effective, we bold the best strategy within each dataset-specific row, and color cells relative to this bold value as green (within 2.5 of best), yellow (within 5 of best), and red (worse by 5 or more points)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.496, + 0.381, + 0.512 + ], + "angle": 0, + "content": "7. DISCUSSION & CONCLUSION" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.643 + ], + "angle": 0, + "content": "We have contributed a benchmark that helps practitioners quantify what gains in a classification task are possible from the addition of unlabeled data and which methods help achieve them. We offer a unified approach to training and hyperparameter selection of semi-supervised methods, self-supervised methods, and supervised baselines that is both affordable and realistic for research labs tackling new challenges without industrial-scale resources." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.78 + ], + "angle": 0, + "content": "Limitations. We deliberately focused on a modest number of labeled examples (30 - 1000) per class, motivated by projects where substantive effort has already gone into labeled data collection. Applications in more scarce-label regimes (e.g. zero-shot or few-shot) may need to also consult other benchmarks, as should those looking at hundreds of fine-grained classes. We also did not specifically study situations where the distribution of labeled and unlabeled data significantly differs [41, 44, 62, 67]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To enable thorough experiments, we focused on moderate ResNet architectures that have proven effective in clinical applications [42, 78]. As ViT-based methods become popular in medical imaging [1, 43], it would be valuable to understand if similar findings hold with vision transformers [26]. We leave this to future work. Ideally, our results would be verified across multiple train-validation splits; however the resources needed remain prohibitive." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.892, + 0.606 + ], + "angle": 0, + "content": "Our analysis here is limited the specific metrics of balanced accuracy and sensitivity/specificity (for AIROGS). Other clinically useful metrics, such as AUPRC, calibration, or net benefit, may be needed to decide if a classifier is appropriate for deployment [70]. For medical applications, it is also key to understand fairness across subpopulations [13] to avoid propagating structural disadvantages." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.892, + 0.685 + ], + "angle": 0, + "content": "Broader impact. All data analyzed here represent fully-deidentified open-access images approved for widespread use by their creators. We think the benefit of promoting these medical tasks to advance ML research outweighs the slight risk of patient reidentification by a bad actor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Outlook. Our experiments show that real benefits from the addition of unlabeled data are sometimes possible: our recommended methods see gains of \\(+5\\) points of balanced accuracy on TissueMNIST, \\(+10\\) points on PathMNIST, and \\(+5\\) points on AIROGS against strong labeled-set-only baselines. In contrast, most tested methods do not add significant gains on TMED-2, perhaps due to that benchmark's uncurated nature [41]. We hope our work enables the research community to convert decades of effort on SSL into improved patient outcomes and better scientific understanding of disease and possible treatments. We further hope that our benchmark inspires those that pursue improved methodological contributions to favor realistic evaluation protocols and clinically-relevant datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22289" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] N Ahmadi, MY Tsang, AN Gu, TSM Tsang, and P Abolmaesumi. Transformer-based spatio-temporal analysis for classification of aortic stenosis severity from echocardiography cine series. IEEE Transactions on Medical Imaging, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.266 + ], + "angle": 0, + "content": "[2] Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zachary Beaver, Jan Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, Vivek Natarajan, and Mohammad Norouzi. Big Self-Supervised Models Advance Medical Image Classification. In International Conference on Computer Vision (ICCV). arXiv, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.473, + 0.45 + ], + "angle": 0, + "content": "[3] Shekoofeh Azizi, Laura Culp, Jan Freyberg, Basil Mustafa, Sebastien Baur, Simon Kornblith, Ting Chen, Nenad Tomasev, Jovana Mitrovic, Patricia Strachan, S. Sara Mahdavi, Ellery Wulczyn, Boris Babenko, Megan Walker, Aaron Loh, Po-Hsuan Cameron Chen, Yuan Liu, Pinal Bavishi, Scott Mayer McKinney, Jim Winkens, Abhijit Guha Roy, Zach Beaver, Fiona Ryan, Justin Krogue, Mozziyar Etemadi, Umesh Telang, Yun Liu, Lily Peng, Greg S. Corrado, Dale R. Webster, David Fleet, Geoffrey Hinton, Neil Houlsby, Alan Karthikesalingam, Mohammad Norouzi, and Vivek Natarajan. Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging. Nature Biomedical Engineering, 7(6):756-779, 2023. 4, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.451, + 0.47, + 0.492 + ], + "angle": 0, + "content": "[4] James Bergstra and Yoshua Bengio. Random search for hyper-parameter optimization. Journal of machine learning research, 13(2), 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.493, + 0.472, + 0.561 + ], + "angle": 0, + "content": "[5] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 1, 3, 21, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.564, + 0.472, + 0.647 + ], + "angle": 0, + "content": "[6] Benjamin Billot, Colin Magdamo, Steven E Arnold, Sudeshna Das, and Juan Eugenio Iglesias. Robust segmentation of brain mri in the wild with hierarchical cnns and no retraining. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 538-548. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[7] Avrim Blum and Tom Mitchell. Combining labeled and unlabeled data with co-training. In Proceedings of the eleventh annual conference on Computational learning theory, pages 92-100, 1998. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.704, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.775, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[9] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[10] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 1, 3, 21" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.163 + ], + "angle": 0, + "content": "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.22 + ], + "angle": 0, + "content": "[12] Paola Cascante-Bonilla, Fuwen Tan, Yanjun Qi, and Vicente Ordonez. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. In Proceedings of the AAAI Conference on Artificial Intelligence, 2021. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.894, + 0.306 + ], + "angle": 0, + "content": "[13] Leo Anthony Celi, Jacqueline Cellini, Marie-Laure Charpignon, Edward Christopher Dee, Franck Dernoncourt, Rene Eber, William Greig Mitchell, Lama Moukheiber, Julian Schirmer, et al. Sources of bias in artificial intelligence that perpetuate healthcare disparities—A global review. PLOS Digital Health, 1(3), 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.308, + 0.894, + 0.364 + ], + "angle": 0, + "content": "[14] Olivier Chapelle, Bernhard Scholkopf, and Alexander Zien. Semi-supervised learning (chapelle, o. et al., eds.; 2006)[book reviews]. IEEE Transactions on Neural Networks, 20(3):542-542, 2009. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.366, + 0.894, + 0.436 + ], + "angle": 0, + "content": "[15] Chen Chen, Chen Qin, Huaqi Qiu, Cheng Ouyang, Shuo Wang, Liang Chen, Giacomo Tarroni, Wenjia Bai, and Daniel Rueckert. Realistic adversarial data augmentation for mr image segmentation. In Medical Image Computing and Computer Assisted Intervention MICCAI, 2020. 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.438, + 0.894, + 0.494 + ], + "angle": 0, + "content": "[16] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.496, + 0.894, + 0.552 + ], + "angle": 0, + "content": "[17] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 2, 3, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.894, + 0.611 + ], + "angle": 0, + "content": "[18] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.612, + 0.894, + 0.668 + ], + "angle": 0, + "content": "[19] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.67, + 0.893, + 0.713 + ], + "angle": 0, + "content": "[20] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.714, + 0.894, + 0.771 + ], + "angle": 0, + "content": "[21] Yanbei Chen, Massimiliano Mancini, Xiatian Zhu, and Zeynep Akata. Semi-supervised and unsupervised deep visual learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.772, + 0.895, + 0.843 + ], + "angle": 0, + "content": "[22] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 6, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[23] Victor Guilherme Turrisi Da Costa, Enrico Fini, Moin Nabi, Nicu Sebe, and Elisa Ricci. solo-learn: A library of self-supervised methods for visual representation learning. J. Mach. Learn. Res., 23(56):1-6, 2022. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22290" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.273 + ], + "angle": 0, + "content": "[24] Coen de Vente, Koenraad A. Vermeer, Nicolas Jaccard, He Wang, Hongyi Sun, FIRas Khader, Daniel Truhn, Temirgali Aimshev, Yerkebulan Zhanibekuly, Tien-Dung Le, Adrian Galdran, Miguel Angel Gonzalez Ballester, Gustavo Carneiro, Devika R. G, Hrishikesh P. S, Densen Puthussery, Hong Liu, Zekang Yang, Satoshi Kondo, Satoshi Kasai, Edward Wang, Ashritha Durvasula, Jonathan Heras, Miguel Angel Zapata, Teresa Araujo, Guilherme Aresta, Hrvoje Bogunovic, Mustafa Arikan, Yeong Chan Lee, Hyun Bin Cho, Yoon Ho Choi, Abdul Qayyum, Imran Razzak, Bram van Ginneken, Hans G. Lemij, and Clara I. Sanchez. AIROGS: Artificial Intelligence for RObust Glaucoma Screening Challenge, 2023. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.472, + 0.33 + ], + "angle": 0, + "content": "[25] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.332, + 0.472, + 0.415 + ], + "angle": 0, + "content": "[26] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.417, + 0.472, + 0.473 + ], + "angle": 0, + "content": "[27] Linus Ericsson, Henry Gouk, and Timothy M Hospedales. How well do self-supervised models transfer? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5414-5423, 2021. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.474, + 0.472, + 0.53 + ], + "angle": 0, + "content": "[28] Andre Esteva, Brett Kuprel, Roberto A Novoa, Justin Ko, Susan M Swetter, Helen M Blau, and Sebastian Thrun. Dermatologist-level classification of skin cancer with deep neural networks. nature, 542(7639):115-118, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.472, + 0.587 + ], + "angle": 0, + "content": "[29] Loveleen Gaur, Ujwal Bhatia, NZ Jhanjhi, Ghulam Muhammad, and Mehedi Masud. Medical image-based detection of Covid-19 using deep convolution neural networks. Multimedia systems, 29(3):1729-1738, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.472, + 0.685 + ], + "angle": 0, + "content": "[30] Hemant Ghayvat, Muhammad Awais, AK Bashir, Sharnil Pandya, Mohd Zuhair, Mamoon Rashid, and Jamel Nebhen. Ai-enabled radiologist in the loop: novel ai-based framework to augment radiologist performance for Covid-19 chest CT medical image annotation and classification from pneumonia. Neural Computing and Applications, 35(20):14591–14609, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.472, + 0.757 + ], + "angle": 0, + "content": "[31] Amirata Ghorbani, David Ouyang, Abubakar Abid, Bryan He, Jonathan H Chen, Robert A Harrington, David H Liang, Euan A Ashley, and James Y Zou. Deep learning interpretation of echocardiograms. NPJ digital medicine, 3(1):10, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.472, + 0.827 + ], + "angle": 0, + "content": "[32] Priya Goyal, Dhruv Mahajan, Abhinav Gupta, and Ishan Misra. Scaling and benchmarking self-supervised visual representation learning. In Proceedings of the IEEE/cvf International Conference on computer vision, pages 6391-6400, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.829, + 0.472, + 0.872 + ], + "angle": 0, + "content": "[33] Margherita Grandini, Enrico Bagli, and Giorgio Visani. Metrics for multi-class classification: an overview. arXiv preprint arXiv:2008.05756, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.903 + ], + "angle": 0, + "content": "[34] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.092, + 0.892, + 0.149 + ], + "angle": 0, + "content": "Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[35] Isabelle Guyon, Kristin Bennett, Gavin Cawley, Hugo Jair Escalante, Sergio Escalera, Tin Kam Ho, Núria Macía, Bisakha Ray, Mehreen Saeed, Alexander Statnikov, et al. Design of the 2015 ChaLearn AutoML challenge. In 2015 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.894, + 0.29 + ], + "angle": 0, + "content": "[36] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[37] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 3, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[38] Wassily Hoeffding. Probability inequalities for sums of bounded random variables. The collected works of Wassily Hoeffding, pages 409-426, 1994. 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[39] Zhe Huang, Gary Long, Benjamin Wessler, and Michael C Hughes. A new semi-supervised learning benchmark for classifying view and diagnosing aortic stenosis from echocardiograms. In Proceedings of the Machine Learning for Healthcare Conference. PMLR, 2021. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[40] Zhe Huang, Gary Long, Benjamin S Wessler, and Michael C Hughes. TMED 2: A dataset for semi-supervised classification of echocardiograms. In DataPerf: Benchmarking Data for Data-Centric AI Workshop, 2022. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[41] Zhe Huang, Mary-Joy Sidhom, Benjamin S Wessler, and Michael C Hughes. Fix-a-step: Semi-supervised learning from uncurated unlabeled data. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics (AISTATS), 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[42] Zhe Huang, Benjamin S Wessler, and Michael C Hughes. Detecting heart disease from multi-view ultrasound images via supervised attention multiple instance learning. In Machine Learning for Healthcare Conference, pages 285-307. PMLR, 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[43] Zhe Huang, Xiaowei Yu, Benjamin S Wessler, and Michael C Hughes. Semi-supervised multimodal multi-instance learning for aortic stenosis diagnosis. arXiv preprint arXiv:2403.06024, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[44] Zhe Huang, Xiaowei Yu, Dajiang Zhu, and Michael C Hughes. Interlude: Interactions between labeled and unlabeled data to enhance semi-supervised learning. arXiv preprint arXiv:2403.10658, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[45] Tongtong Huo, Yi Xie, Ying Fang, Ziyi Wang, Pengran Liu, Yuyu Duan, Jiayao Zhang, Honglin Wang, Mingdi Xue, Songxiang Liu, et al. Deep learning-based algorithm improves radiologists' performance in lung cancer bone metastases detection on computed tomography. Frontiers in Oncology, 13:1125637, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[46] Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Label propagation for deep semi-supervised learning." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22291" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5070-5079, 2019. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[47] Longlong Jing and Yingli Tian. Self-supervised visual feature learning with deep neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(11):4037-4058, 2020. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[48] Jakob Nikolas Kather, Johannes Krisam, Pornpimol Charoentong, Tom Luedde, Esther Herpel, Cleo-Aron Weis, Timo Gaiser, Alexander Marx, Nektarios A Valous, Dyke Ferber, et al. Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study. PLoS medicine, 16(1):e1002730, 2019. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[49] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.401 + ], + "angle": 0, + "content": "[50] Byoungjip Kim, Jinho Choo, Yeong-Dae Kwon, Seongho Joe, Seungjai Min, and Youngjune Gwon. Selfmatch: Combining contrastive self-supervision and consistency for semi-supervised learning. arXiv preprint arXiv:2101.06480, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[51] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[52] Durk P Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semi-supervised learning with deep generative models. Advances in neural information processing systems, 27, 2014. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[53] Abhishek Kumar, Prasanna Sattigeri, and Tom Fletcher. Semi-supervised learning with gans: Manifold invariance with improved inference. Advances in neural information processing systems, 30, 2017. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[54] Devidas T Kushnure, Shweta Tyagi, and Sanjay N Talbar. Lim-net: Lightweight multi-level multiscale network with deep residual learning for automatic liver segmentation in ct images. Biomedical Signal Processing and Control, 80: 104305, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[55] Zhengfeng Lai, Chao Wang, Luca Cerny Oliveira, Brittany N Dugger, Sen-Ching Cheung, and Chen-Nee Chuah. Joint semi-supervised and active learning for segmentation of gigapixel pathology images with cost-effective labeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 591–600, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[56] Zhengfeng Lai, Chao Wang, Henry Gunawan, Sen-Ching S Cheung, and Chen-Nee Chuah. Smoothed adaptive weighting for imbalanced semi-supervised learning: Improve reliability against unknown distribution data. In International Conference on Machine Learning, pages 11828–11843. PMLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[57] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[58] Dong-Hyun Lee. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning at ICML, 2013. 2, 3, 21" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[59] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9475-9484, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[60] Ali Madani, Remy Arnaout, Mohammad Mofrad, and Rima Arnaout. Fast and accurate view classification of echocardiograms using deep learning. NPJ digital medicine, 1(1):6, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[61] Shaobo Min, Xuejin Chen, Hongtao Xie, Zheng-Jun Zha, and Yongdong Zhang. A mutually attentive co-training framework for semi-supervised recognition. IEEE Transactions on Multimedia, 23:899-910, 2020. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[62] Avital Oliver, Augustus Odena, Colin A Raffel, Ekin Dogus Cubuk, and Ian Goodfellow. Realistic evaluation of deep semi-supervised learning algorithms. Advances in neural information processing systems, 31, 2018. 1, 2, 3, 4, 6, 8, 13, 22, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[63] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[64] Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. the Journal of machine learning research, 12:2825–2830, 2011. 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[65] Guo-Jun Qi and Jiebo Luo. Small data challenges in big data era: A survey of recent progress on unsupervised and semi-supervised methods. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2168-2187, 2020. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[66] Maithra Raghu, Chiyuan Zhang, Jon Kleinberg, and Samy Bengio. Transfusion: Understanding transfer learning for medical imaging. Advances in neural information processing systems, 32, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.563, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[67] Kuniaki Saito, Donghyun Kim, and Kate Saenko. Openmatch: Open-set consistency regularization for semi-supervised learning with outliers. arXiv preprint arXiv:2105.14148, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.62, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[68] Azizi Shekoofeh, Mustafa Basil, Ryan Fiona, Beaver Zachary, Freyberg Jan, Deaton Jonathan, Loh Aaron, Karthikesalingam Alan, Kornblith Simon, Chen Ting, et al. Big self-supervised models advance medical image classification. arXiv preprint arXiv:2101.05224, 2021. 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[69] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 1, 3, 5, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[70] Ewout W Steyerberg and Yvonne Vergouwe. Towards better clinical prediction models: seven steps for development and an abcd for validation. European Heart Journal, 35(29), 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[71] Jong-Chyi Su, Zezhou Cheng, and Subhransu Maji. A realistic evaluation of semi-supervised learning for fine-grained classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12966-12975, 2021. 1, 3, 4, 5, 7, 8, 13, 16, 23, 25" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22292" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[72] Terpeei Suzuki. Consistency regularization for semi-supervised learning with pytorch. https://github.com/perrying/pytorch-consistency-regularization, 2020.13" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.15, + 0.471, + 0.205 + ], + "angle": 0, + "content": "[73] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 3, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[74] Jesper E Van Engelen and Holger H Hoos. A survey on semi-supervised learning. Machine learning, 109(2):373-440, 2020. 1, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.47, + 0.318 + ], + "angle": 0, + "content": "[75] Diane Wagner, Fabio Ferreira, Danny Stoll, Robin Tibor Schirrmeister, Samuel Müller, and Frank Hutter. On the importance of hyperparameters and data augmentation for self-supervised learning. arXiv preprint arXiv:2207.07875, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.321, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[76] Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, et al. Usb: A unified semi-supervised learning benchmark for classification. Advances in Neural Information Processing Systems, 35:3938-3961, 2022. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.392, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[77] Benjamin S Wessler, Zhe Huang, Gary M Long Jr, Stefano Pacifici, Nishant Prashar, Samuel Karmiy, Roman A Sandler, Joseph Z Sokol, Daniel B Sokol, Monica M Dehn, et al. Automated detection of aortic stenosis using machine learning. Journal of the American Society of Echocardiography, 36(4): 411-420, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.477, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[78] Nan Wu, Jason Phang, Jungkyu Park, Yiqiu Shen, Zhe Huang, Masha Zorin, Stanisław Jastrzebski, Thibault Févry, Joe Katsnelson, Eric Kim, et al. Deep neural networks improve radiologists' performance in breast cancer screening. IEEE transactions on medical imaging, 39(4):1184-1194, 2019. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.562, + 0.469, + 0.629 + ], + "angle": 0, + "content": "[79] Jiancheng Yang, Rui Shi, and Bingbing Ni. Medmnist classification decathlon: A lightweight automl benchmark for medical image analysis. In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), pages 191-195. IEEE, 2021. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.632, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[80] Jiancheng Yang, Rui Shi, Donglai Wei, Zequan Liu, Lin Zhao, Bilian Ke, Hanspeter Pfister, and Bingbing Ni. Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data, 10(1):41, 2023. 4, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.703, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[81] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Proceedings of the British Machine Vision Conference (BMVC), 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.746, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[82] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.804, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[83] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1476-1485, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[84] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curricu" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "lum pseudo labeling. Advances in Neural Information Processing Systems, 34:18408-18419, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[85] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 3, 6, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[86] Wenqiao Zhang, Lei Zhu, James Hallinan, Shengyu Zhang, Andrew Makmur, Qingpeng Cai, and Beng Chin Ooi. BoostMIS: Boosting medical image semi-supervised learning with adaptive pseudo labeling and informative active annotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[87] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14471-14481, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[88] Xiaojin Zhu. Semi-Supervised Learning Literature Survey. Technical Report 1530, Department of Computer Science, University of Wisconsin Madison., 2005. 1, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.429 + ], + "angle": 0, + "content": "[89] Xiaojin Zhu, Zoubin Ghahramani, and John D Lafferty. Semi-supervised learning using gaussian fields and harmonic functions. In Proceedings of the 20th International conference on Machine learning (ICML-03), pages 912–919, 2003. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.894, + 0.487 + ], + "angle": 0, + "content": "[90] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2020. 21" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22293" + } + ] +] \ No newline at end of file diff --git a/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/a33fa80f-bfa4-45b2-878d-abb01f3d724e_origin.pdf b/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/a33fa80f-bfa4-45b2-878d-abb01f3d724e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2f94eacce29e3fad337e128a0c55432ead3dd5d7 --- /dev/null +++ b/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/a33fa80f-bfa4-45b2-878d-abb01f3d724e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f72580df436ea3f4103314d04a807c723e69f36a8faf7eae8ba3574ad2d002e +size 505034 diff --git a/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/full.md b/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5fea05df6b23d0f3f966b1b9ea0680dc0b2e46f3 --- /dev/null +++ b/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/full.md @@ -0,0 +1,311 @@ +# Systematic comparison of semi-supervised and self-supervised learning for medical image classification + +Zhe Huang* Ruijie Jiang* Shuchin Aeron Michael C. Hughes Tufts University, School of Engineering + +{Zhe.Huang, Ruijie.Jiang, Shuchin.Aeron, Michael.Hughes}@tufts.edu + +# Abstract + +In typical medical image classification problems, labeled data is scarce while unlabeled data is more available. Semi-supervised learning and self-supervised learning are two different research directions that can improve accuracy by learning from extra unlabeled data. Recent methods from both directions have reported significant gains on traditional benchmarks. Yet past benchmarks do not focus on medical tasks and rarely compare self- and semi- methods together on an equal footing. Furthermore, past benchmarks often handle hyperparameter tuning suboptimally. First, they may not tune hyperparameters at all, leading to underfitting. Second, when tuning does occur, it often unrealistically uses a labeled validation set that is much larger than the training set. Therefore currently published rankings might not always corroborate with their practical utility. This study contributes a systematic evaluation of self- and semi- methods with a unified experimental protocol intended to guide a practitioner with scarce overall labeled data and a limited compute budget. We answer two key questions: Can hyperparameter tuning be effective with realistic-sized validation sets? If so, when all methods are tuned well, which self- or semi-supervised methods achieve the best accuracy? Our study compares 13 representative semi- and self-supervised methods to strong labeled-set-only baselines on 4 medical datasets. From $20000+$ GPU hours of computation, we provide valuable best practices to resource-constrained practitioners: hyperparameter tuning is effective, and the semi-supervised method known as MixMatch delivers the most reliable gains across 4 datasets. + +# 1. INTRODUCTION + +Deep neural networks can deliver exceptional performance on classification tasks when trained with vast labeled + +datasets. However, in medical imaging applications assembling a large dataset with appropriate label can be prohibitively costly due to manual effort required by a human expert. In contrast, images alone, without labels, are often readily available in health records databases. In recent years, significant research has focused on developing methods that leverage both the large unlabeled set and a small labeled set to enhance image classifier training, aiming to surpass models trained solely on labeled data. + +Two ways of leveraging unlabeled data are particularly popular: semi-supervised and self-supervised learning, both often abbreviated as SSL. Recent efforts in semisupervised learning [74, 88] usually train deep classifiers jointly [5, 69] using an objective with two loss terms, one favoring labeled-set accuracy and the other favoring label-consistency or label-smoothness. Alternatively, self-supervised methods [65] take a two-stage approach, first training deep representations on the unlabeled set, then fine-tuning a classifier on the labeled set. Exemplars are numerous [10, 17, 19, 20, 37]. Despite the remarkable progress reported in each direction, these two paradigms have been largely developed independently [21]. A direct comparison of the two paradigms has been notably absent. A practitioner building a medical image classifier from limited labeled data may ask, "Which recent semi- or self-supervised methods are likely to be most effective?" + +Performance can be quite sensitive to hyperparameters for both semi-SL [69, 71] and self-SL [75]. We argue that any careful comparison must consider tuning hyperparameters of all methods in a fair fashion. However, recent work has not handled this well. First, as reviewed in Table 1, recent benchmarks for both semi- and self- paradigms often omit any tuning of hyperparameters (see the Fungi semi-SL experiments of Su et al. [71] or Ericsson et al. [27]'s self-SL benchmark). This practice of using off-the-shelf defaults likely leads to under-performing given a new task, and may impact some methods more than others. An even bigger issue with prevailing semi-SL practice was originally raised by Oliver et al. [62]: many published papers perform hyperparameter tuning on a labeled validation set that + +is larger (sometimes much larger) than the limited labeled train set. Such experimental settings are unrealistic. SSL is intended for practitioners without abundant available labeled data. Practitioners that need to make the most of 1000 available labeled images will not elect to put more images in validation than training, and thus are not helped by benchmarks that do. Unfortunately, five years later after [62] we find this issue is still widespread (see Tab. 1). For example, Wang et al. [76]'s semi-SL results on TissueMNIST tune on a validation set over $50\mathrm{x}$ larger than the labeled train set. + +Oliver et al. [62] further cast some doubt on whether effective tuning is possible with small labeled sets, saying "Extensive hyperparameter tuning may be somewhat futile due to an excessively small collection of held-out data to measure performance on". There thus remains a pressing question for resource-constrained practitioners: "Given limited available labeled data and limited compute, is hyperparameter tuning worthwhile?" + +This study makes progress toward answering these questions by delivering a comprehensive comparison of semi- and self-supervised learning algorithms under a resource-constrained scenario that matches how SSL might be used on real-world medical tasks. Our goal is to enable practitioners to tackle new image classification challenges where only modest-sized labeled datasets exist. We target settings with roughly 2-10 class labels of interest, where each class has 30-1000 available labeled images for all model development (including training and validation). We select representative tasks across 4 datasets that span low-resolution (28x28) to moderate-resolution (112x112 and 384x384). On each task, we run careful experiments with representative recent methods from both semi- and self-supervised learning to clarify what gains are possible with unlabeled data and how to achieve them. + +We emphasize realism throughout, in 4 distinct ways: (1) using validation set sizes that are never bigger than the train set, avoiding the unrealistically-large validation sets common in past work (Tab. 1); (2) performing hyperparameter search using the same protocol, same compute budget (fixed number of hours)* and same hardware (one NVIDIA A100 GPU) for all algorithms for fair comparison; (3) respecting natural class imbalance; (4) profiling performance over time, to inform labs with smaller runtime budgets. + +In summary, the contributions of this study are: + +1. We provide a systematic comparison of semi- and self-supervised methods on equal footing to connect two research directions that have been heretofore separate. +2. We adopt a realistic experimental protocol designed to consider the same constraints on available labels and runtime that SSL practitioners face, avoiding the unre + +alistic aspects of past benchmarks, especially the use of far too large validation sets. + +3. We show that hyperparameter tuning is viable with a realistic-sized validation set, and in many cases necessary to do well at a new classification task with new data. + +Ultimately, we hope this study guides practitioners with limited data toward successful deployment of semi- and self-supervised methods on real problems. + +# 2. BACKGROUND AND METHODS + +Unified Problem Formulation. Following the recent survey by Chen et al. [21], we adopt a unified perspective for supervised, semi-supervised and self-supervised image classification with limited available labeled data. For model development (including training and hyperparameter selection), we assume there are two available datasets. First, a small labeled dataset $\mathcal{L}$ of feature-label pairs $(x,y)$ , where each image is represented by a $D$ -dimensional feature vector $x \in \mathbb{R}^D$ and its corresponding class label takes one of $C$ possible values: $y \in \{1,2,\dots,C\}$ . Second, an unlabeled set $\mathcal{U}$ containing only feature vectors. Typically, we assume the unlabeled set is much larger: $|\mathcal{U}| \gg |\mathcal{L}|$ . + +Given labeled set $\mathcal{L}$ and unlabeled set $\mathcal{U}$ , we wish to train a neural network that can map each input $x$ to a probability vector in the $C$ -dimensional simplex $\Delta^C$ representing a distribution over $C$ class labels. Let $f_v(\cdot): \mathbb{R}^D \to \mathbb{R}^F$ denote a backbone neural network with parameters $v$ producing an $F$ -dimensional embedding given any input image. Let $g_w(\cdot): \mathbb{R}^F \to \Delta^C$ denote a final linear-softmax classification layer with parameters $w$ . The following unified objective can capture all three learning paradigms: + +$$ +\begin{array}{l} v ^ {*}, w ^ {*} \leftarrow \arg \min _ {v, w} \sum_ {x, y \in \mathcal {L}} \lambda^ {L} \ell^ {L} (y, g _ {w} (f _ {v} (x))) \tag {1} \\ + \sum_ {x \in \mathcal {U}} \lambda^ {U} \ell^ {U} (x, f _ {v}, g _ {w}). \\ \end{array} +$$ + +Here, $\ell^L$ represents a labeled-set loss (e.g. multi-class cross-entropy), and $\ell^U$ represents a unlabeled-set loss. $\lambda^L, \lambda^U \geq 0$ are weights for the corresponding loss terms. The design of $\ell^U$ is usually what differs substantially across methods. For instance, setting $\ell^U$ to be cross-entropy computed with pseudo-labels generated from classifier $g$ recovers PseudoLabel [58]; a temperature-scaled instance-similarity contrastive loss recovers SimCLR [17]. + +All three learning paradigms that we study optimize Eq. (1), yet differ in the number of phases and in how to set the scalar weights $\lambda^L, \lambda^U$ on each loss term. Supervised learning ignores the unlabeled term throughout training ( $\lambda^U = 0$ ), thus learning parameters using only the labeled set. Semi-supervised methods include both terms in one end-to-end training, keeping both $\lambda^U > 0$ and $\lambda^L > 0$ . + +Self-supervised learning has two phases. In phase 1 ("pretraining"), the labeled term is omitted ( $\lambda^L = 0$ , $\lambda^U = 1$ ) and the focus of learning is an effective representation + +
BenchmarkMethodsUnrealistic ExperimentsLabeled train sizeLabeled val. sizeAcc vs. Time?
Realistic eval. SSL [a]SemiCIFAR-10 (Tab. 1-2, Fig. 2)40005000no
SVHN (Tab. 1-2, Fig. 3-4)10007325
Fine-grained SSL [b]Semi & 1 Self*Semi-Aves59598000no
Semi-Fungi4141None
USB [c]SemiTissueMNIST80/40023640†no
Semi-Aves5959None
Self benchmark [d]SelfImageNet1.28 mil.Noneno
SSL-vs-SSL (ours)Semi & Selfno400-1660no bigger than trainyes
+ +Table 1. Comparison of related benchmarks of semi-supervised and self-supervised learning. Past works either do no hyperparameter tuning at all (val. size = None), or use an unrealistically large validation set (defined as larger than the labeled train set) in some or all experiments. Further, each work almost exclusively looks at methods from one paradigm, either semi- or self- (*: [b] includes one self-supervised method, MOCO). In contrast, in this paper we benchmark 6 semi- and 7 self-supervised algorithms with hyperparameter tuning on realistic validation sets. Acc vs. Time? indicates whether the work analyzes performance over training time. Number marked † confirmed via GitHub comment by USB authors. Citations: a: Oliver et al. [62], b: Su et al. [71] c: Wang et al. [76] d: Ericsson et al. [27]. + +layer $f_{v}$ (classifier $g_{w}$ is not included in this phase). In phase 2 ("fine-tuning"), we focus on the labeled term and omit the unlabeled term $(\lambda^{L} = 1, \lambda^{U} = 0)$ . We fix the representation parameter $v$ and fine-tune the classifier $w$ . + +We now identify the 16 methods we will evaluate: + +Supervised methods. The goal of leveraging unlabeled data $\mathcal{U}$ is to obtain better performance than what we could obtain using only the labeled set $\mathcal{L}$ . Therefore, we naturally compare to 3 high-quality supervised baselines that use only the labeled set. First, "Sup" denotes a classifier trained with supervised loss $\ell^L$ set to multi-class cross-entropy. Second, "MixUp" trains with cross entropy with the addition of mixup data augmentation [85]. Finally, "SupCon" pursues a supervised contrastive learning loss for $\ell^L$ [49]. + +Semi-supervised methods. We compare 6 semi-supervised methods that train deep classifiers on both labeled and unlabeled data simultaneously as in Eq. (1). To represent the state-of-the-art of semi-supervised image classification, we select Pseudo Label ("PseudoL") [58], Mean Teacher ("MeanTch") [73], MixMatch [5], FixMatch [69], FlexMatch [84] and CoMatch [59]. These choices cover a reasonably wide spectrum of unlabeled loss design strategy, year of publication, and computation cost. CoMatch represents a recent trend of combining semi- and self-supervision. See App D.2 for a wider literature review. + +Self-supervised methods. We compare 7 self-supervised algorithms: SimCLR [17], MOCO (v2) [20, 37], SwAV [10], BYOL [34], SimSiam [19], DINO [11] and Barlow Twins ("BarlowTw") [82]. These algorithms epitomize the field of self-supervised learning as of this writing. SimCLR, MOCO (v2), and SwAV are based on contrastive learning, which learns effective representations when provided with similar and dissimilar samples. BYOL was designed to circumvent the need for dissimilar samples. SimSiam is a simple yet effective Siamese representation learning method. DINO uses a teacher-student network architec + +ture for representation distillation. Barlow Twins preserves information and reduces redundancy by favoring a close-to-identity cross-correlation matrix between augmented pairs of data. See App D.3 for further review. + +# 3. RELATED WORK + +Table 1 summarizes key attributes of existing major benchmarks for semi-supervised and self-supervised methods. We now discuss how our study situates in this context. + +Comparison to Oliver et al. Oliver et al. [62] provide an influential benchmark of deep semi-supervised methods. Like our work, they highlight the pressing issue of using unrealistically large validation sets for hyperparameter tuning. However, most of their experiments (e.g. all their tables and their figures 2-4) still use this unrealistic setting to be comparable to other works; only one subsection (Sec. 4.6) examines validation sets no larger than the train set. Instead, we exclusively use validation set sizes no larger than training set, mimicking what practitioners would face in real applications. Our definition of a "realistic" size for a validation set – no larger than the labeled train set – is broader than Oliver et al.'s definition of $10\%$ of the train set. We favor our definition because Oliver et al. find specifically that "for validation sets of the same size $(100\%)$ as the training set, some differentiation between the approaches is possible." In contrast, they found reliable differentiation was not always possible with much smaller validation sets. + +Furthermore, the benchmarking results presented by Oliver et al. require each algorithm to complete "1000 trials of Gaussian Process-based black-box optimization using Google Cloud ML Engine". This resource-intensive process seems impractical for researchers outside well-funded industrial labs. We use more modest compute budgets and specifically allow each method the same total wallclock runtime for tuning, again obeying practical constraints. + +Other semi-supervised benchmarks. TorchSSL [84] + +benchmarked eight popular semi-supervised learning algorithm using a unified codebase. The same group later extended that effort to natural language and audio processing in their USB benchmark [76]. Another recent effort for fine-grained classification by Su et al. [71] evaluates semi-supervised learning on datasets that exhibit class imbalance and contains images from novel classes in the unlabeled set, studying the effect of different initializations and the contents of unlabeled data on the performance of semi-supervised methods. While making significant contributions, these works focused on semi-supervision almost exclusively and did not include multiple self-supervised learning methods, thus leaving open the questions "How do the two paradigms compare? Which methods are best overall?" + +Prior self-supervised benchmarks. Several works have proposed different benchmarks for self-supervised methods. Goyal et al. [32] introduced a benchmark that covers 9 different tasks, such as object detection and visual navigation. Ericsson et al. [27] compared thirteen top self-supervised models on 40 downstream tasks. Da Costa et al. [23] presented a library of self-supervised methods that can be easily plugged into different downstream tasks and datasets. However, these works mainly consider other self-SL methods without direct comparison to the semi-SL paradigm. Moreover, they mostly do not incorporate the use of a validation set for hyperparameter tuning at all. This can lead to suboptimal accuracy at test time and complicate comparisons to methods that do tune. + +Self-supervision for medical images. Our work is complementary to recent efforts that assess self-supervised pipelines for medical image classification [2, 3]. They focus primarily on how to design multi-stage transfer learning pipelines and do not comprehensively compare many different self-supervised methods or any semi-supervised methods. Further, many datasets studied by [3] come from proprietary projects conducted at Google. In contrast, all our data and experiments are open and reproducible by others. + +Combining semi- and self-supervision. Recent research has explored the fusion of semi-supervised learning and self-supervised learning ideas [50, 59, 83, 87]. Some recent surveys [21, 65] compare semi- and self-supervised learning, but they focus on literature review while we offer realistic and comprehensive benchmarking experiments. + +# 4. DATASETS AND TASKS + +We study four open-access medical image classification datasets, all with 2D images that are fully deidentified. Table 2 reports statistics for all train/test splits. The exact splits used in our study can be found in our open-source codebase, documented in App. A. + +Two datasets - PathMNIST and TissueMNIST - are selected from the MedMNIST collection [80] (criteria in App. B.2). Prior experiments by Yang et al. [80] suggest + +
TissueMNISTPathMNIST
LabeledUnlab.LabeledUnlab.
TrainValTestTrainTrainValTestTrain
total40040047280165066total450450718089546
GE151516775851NORM39397417847
DCT191922337795MUC404010357966
POD191922027686ADI474713389319
LEU2828336911761STR47474219354
IE3737440215369BACK48488479461
TAL5959703124549DEB525233910308
PT95951120139108LYM525264310349
CD/CT1281281516552947MUS616159212121
TUM6464123312821
TMED-2AIROGS
LabeledUnlab.LabeledUnlab.
TrainValTestTrainTrainValTestTrain
total16602352019353500total600600600094242
PSAX22350342-Glaucoma6060600-
A2C32528319-No Glauc.5405405400-
A4C46239423-
PLAX650118935-
+ +Table 2. Summary statistics of train/validation/test splits for all datasets in our study. Each table's rows are arranged in ascending order based on the number of per-class labeled train images. Full description of class names can be found in App B.3. + +their $28 \times 28$ resolution is a reasonable choice for rapid prototyping; using larger $224 \times 224$ resolution does not yield much more accurate classifiers for these two datasets. + +Two other datasets represent more moderate resolutions closely tied to contemporary clinical research: the 112x112 Tufts Medical Echocardiogram Dataset (TMED-2) and the 384x384 AIROGS dataset. Further details for each dataset are provided in a dedicated paragraph below. + +For each dataset, our data splitting strategy closely mirrors the conditions of real SSL applications. First, we let labeled training and validation sets contain a natural distribution of classes even if that may be imbalanced. This reflects how data would likely be affordably collected (by random sampling) and avoids artificially balanced training sets that will not match the population an algorithm would encounter in a deployment. Second, to be realistic we ensure validation sets are never larger than the available labeled train sets. This is in contrast to previous benchmarks: Wang et al. [76]'s TissueMNIST hyperparameter search used a validation set of 23,640 images even though the labeled training set contained only 80 or 400 images. In practice, such a large validation set would almost certainly be re-allocated to improve training, as noted in Oliver et al. [62]. + +TissueMNIST is a lightweight dataset of 28x28 images of human kidney cortex cells, organized into 8 categories. The original dataset is fully-labeled and distributed with predefined train/val/test split of 165,466/23,640/47,280 im + +ages with some class imbalance. We assume a total labeling budget of 800 images, evenly split between training and validation (to ensure hyperparameter tuning can differentiate methods). We form a labeled training set of 400 images from the predefined training set, sampling each class by its frequency in the original training set. We form a labeled validation set of 400 images sampled from the predefined validation set. For unlabeled set, we keep all remaining images in the original training split, discarding known labels. + +PathMNIST is another lightweight dataset of 28x28 images of patches from colorectal cancer histology slides that comprise of 9 tissue types. The original dataset is fully-labeled and distributed with predefined train/val/test split of 89,996/10,004/7,180 images. Class imbalance is less severe than TissueMNIST. We assume a total labeled data budget of 900 images, again evenly split between training and validation. Labeled train, labeled valid, and unlabeled sets are sampled via the same procedures as in TissueMNIST. For both MedMNIST datasets, we use the predefined test set. + +TMED-2 [39, 40] is an open-access dataset of $112 \times 112$ 2D grayscale images captured from routine echocardiogram scans. Each scan produces dozens of ultrasound images of the heart, captured from multiple acquisition angles (i.e., different anatomic views). In this study, we adopt Huang et al.'s view-classification task: the goal is to classify each image into one of 4 canonical view types (see App. B.3). View classification is clinically important: measurements or diagnoses of heart disease can only be made when looking at the right anatomical view [60, 77]. We used the provided train/validation/test split (id #1), which is naturally imbalanced and matches our criteria for realistic sizing. We further use the provided large unlabeled set of 353,500 images. TMED-2's unlabeled set is both authentic (no true labels are available at all, unlike other benchmarks that "forget" known labels) and uncurated [41] (contains images of view types beyond the 4 classes in the labeled task). + +AIROGS [24] is a public dataset released for a recent competition where the binary classification task is to decide whether the patient should be referred for glaucoma or not, given a color fundus image of the retina (eye). We selected this dataset because it represents an active research challenge even with all available labels. Furthermore, because labels for this dataset were acquired at great expense (multiple human annotators graded each of over 100k images), we hope our work helps assess how self-/semi-supervision could reduce the annotation costs of future challenges. We selected the $384 \times 384$ resolution favored by several high-performing competition entries. We chose a labeling budget of 1200 total images, evenly split between train and validation; the rare positive class (9 to 1 imbalance) is representative of many screening tasks. We took the rest of available images as the unlabeled train set. + +# 5. EXPERIMENTAL DESIGN + +Performance metric. We use balanced accuracy [33, 35] as our primary performance metric. For a task with $C \geq 2$ classes, let $y_{1:N}$ denote true labels for $N$ examples in a test set, and $\hat{y}_{1:N}$ denote a classifier's predicted labels. Let $\mathrm{TP}_c$ count true positives for class $c$ (number of correctly classified examples whose true label is $c$ ), and let $N_c$ count the total number of examples with true label $c$ . Then we compute balanced accuracy (BA) as a percentage: + +$$ +\mathbf{BA}(y_{1:N},\hat{y}_{1:N}) = \frac{1}{C}\sum_{c = 1}^{C}\frac{\mathrm{TP}_{c}(y_{1:N},\hat{y}_{1:N})}{N_{c}(y_{1:N})}\cdot 100\% \tag{2} +$$ + +Balanced accuracy is more suitable than standard accuracy for imbalanced problems when each class matters equally. The expected BA of a uniform random guess is $\frac{100}{C}\%$ . + +On AIROGS, we also track metrics recommended by its creators [24] for clinical utility: AUROC, partial AUROC ( $>90\%$ specificity) and sensitivity-at- $95\%$ -specificity. + +Architectures. CNN backbones are popular in medical imaging [6, 28-31, 42, 45, 54-56, 77, 78]. We use ResNet-18 [36] on Tissue and Path, and WideResNet-28-2 [81] on TMED-2. We experiment with both ResNet-18 and 50 [36] on AIROGS to assess architectural differences. + +Training with early stopping. For each training phase, we perform minibatch gradient descent on Eq. (1) with Adam [51] optimizer and a cosine learning rate schedule [69]. Each training phase proceeds for up to 200 epochs, where one epoch represents enough minibatch updates to process the equivalent of the entire combined training set $\mathcal{L} \cup \mathcal{U}$ . After every epoch, we record balanced accuracy on the validation set. If this value plateaus for 20 consecutive epochs, we stop the current training phase early. + +Hyperparameters. Semi- and self-supervised learning can both be sensitive to hyperparameters [71, 75]. Our evaluations tune both shared variables (e.g. learning rates, weight decay, unlabeled loss weight $\lambda^U$ ) as well as hyperparameters unique to each algorithm. See App. F.1 for details on all hyperparameters for all methods. + +Unified procedure for training and hyperparameter tuning. We formulate a unified procedure mindful of realistic hardware and runtime constraints for a non-industrial-scale lab working on a new medical dataset. We assume each algorithm has access to one NVIDIA A100 GPU for a fixed number of hours. Within the allotted compute budget, for each algorithm we execute a serial random search over hyperparameters, sequentially sampling each new configuration and then training until either an early stopping criteria is satisfied or the maximum epoch is reached. We track the best-so-far classifier in terms of validation-set performance every epoch. Algorithm D.1 provides pseudocode for the procedure. Each time a new hyperparameter configuration is needed, we sample each hyperparameter value independ- + +dently from a distribution designed to cover its common settings in prior literature (App. F.1). Our choice of random search on a budget is thought to yield better performance than grid search [4], fairly expends the same effort to train all methods regardless of their cost-per-epoch or hyperparameter complexity, and does not assume industrial-scale access to 1000 cloud-computing trials as in Oliver et al. [62]. We find that performance saturates after about 25 hours per 80000 unlabeled examples. We thus allocate for the total time budget 25 hours for PathMNIST, 50 hours for TissueMNIST, and 100 hours for TMED-2. We allow 100 hours for AIROGS due to its larger resolution. Due to limited resources, we select only a few competitive methods to run on AIROGS based on the results on other datasets. + +Self-supervised classification phase. Self-supervised methods by definition do not utilize label information when training representation layer weights $v$ . To enable a proper comparison, for all self-SL methods our Alg. D.1 introduces an additional classification layer, training it anew after each epoch, each time using a 10 trial random search for an L2-penalty regularization hyperparameter. We retain the best performing weights $w$ on the validation set. We emphasize that $w$ does not impact overall self-supervised training of $v$ . + +Data augmentation. Random flip and crop are used for all semi-supervised and supervised methods. MixMatch also utilizes MixUp [85], while RandAugment [22] is used by FixMatch and FlexMatch. For each self-supervised method as well as SupCon, we apply the same SimCLR augmentation: random flip, crop, color jitter, and grayscale. + +Multiple trials. We repeat Alg. D.1 for each method across 5 separate trials (distinct random seeds). We record the mean balanced accuracy on valid and test sets of these trials every 60 minutes (every $30\mathrm{min}$ for Tissue and Path). + +# 6. RESULTS & ANALYSIS + +In each panel of Fig. 1, we focus on one dataset and architecture, plotting for each algorithm one line showing test set balanced accuracy (averaged across 5 trials) as a function of time spent running Alg. D.1. Variance across trials is visualized in Fig C.4. Extra results on AIROGS showing other performance metrics over time (AUROC, partial AUROC, and sensitivity at $95\%$ specificity) are found in Fig. C.3. We emphasize that following best practice, each checkpoint represents the best hyperparameters as selected on the validation set; we then report that checkpoint's test set performance. These results help us answer two key questions: + +Is hyperparameter tuning worthwhile for SSL methods when validation set sizes cannot be larger than the train set? Across Fig. 1 and C.3, all 16 algorithms show roughly monotonic improvements in test performance over time, despite using a realistic-sized validation set. The final performance of every algorithm and dataset shows improvement due to our unified training and tuning procedure (fur + +
nameGain in Bal.Acc. over best labeled-set-only
PathTissueTMEDAIROGSmedianworst
MixMatch12.43.7-0.25.74.7-0.2
MOCO7.94.4-18.8n/a4.4-18.8
SwAV8.44.3-10.1n/a4.3-10.1
SimCLR11.05.9-7.81.63.8-7.8
BYOL12.16.2-6.00.93.5-6.0
BarlowTw11.22.2-1.9n/a2.2-1.9
SimSiam1.92.1-11.6n/a1.9-11.6
FlexMtech11.14.5-1.1-4.21.7-4.2
DINO12.66.5-3.4-3.81.6-3.8
FixMatch9.32.90.2-4.51.6-4.5
CoMatch14.10.2-3.0n/a0.2-3.0
MeanTch-1.6-1.9-1.2n/a-1.6-1.9
PseudoL0.6-1.4-4.4-3.2-2.3-4.4
ref. val.69.833.994.971.5
+ +Table 3. Gains from SSL methods over only using labeled set, across 4 datasets. Self (italicized), Semi (normal). We report gains in percentage balanced accuracy (higher is better) in the mean final test set performance (averaged over 5 runs of Alg. D.1) over a reference value that represents the best labeled-set-only run among Sup (minimize cross entropy), SupCon, and MixUp. MixMatch has the highest median gain, and is the only method to never score notably worse ( $>1$ point) than the best labeled set only run. + +ther discussion in App. E.1). Therefore, we answer: Yes, realistically-sized validation sets for hyperparameter tuning and checkpoint selection can be effective. + +What are the best SSL methods? Among the many methods we study in Fig. 1, no method clearly outperforms others across all datasets. On Path, CoMatch, MixMatch, DINO, and BYOL perform best. On Tissue, self-supervised methods like DINO, BYOL, and SimCLR score well. On TMED-2, ultimately only FixMatch and MixMatch are competitive with the MixUp baseline. On AIROGS, only MixMatch, SimCLR and BYOL beat the Sup baseline. + +For each semi- and self- SSL method, Tab. 3 reports its relative gain in balanced accuracy over the best labeled-set-only supervised baseline on each dataset. We conclude that MixMatch represents the best overall choice as it consistently performs at or near the top across all tasks. It is the only method to never deliver results worse by more than 1 percentage point than the best labeled-set-only baseline (see "worst" column of Tab. 3). This is in stark contrast to USB [76], where FixMatch and FlexMatch are ranked notably higher than MixMatch. Such differences stress the importance of choosing proper evaluation protocol tailored to specific needs. We do caution that hyperparameter tuning is strongly recommended for MixMatch to succeed on a new dataset (see AIROGS result in Tab. 4). + +Pretraining vs. from scratch. Plots in App. C show accuracy-over-time profiles using initial weights pretrained + +![](images/d5b0f3d7dae1ceee445e3d403de4cfecd9548f62debb697ce82095b30a09e424.jpg) + +![](images/08b5e3edfd6600c308e8c6cf71fd85e8ecee90584020fa583ec438a3695b4844.jpg) + +![](images/4cb7f33695ab7fcb2ab4a807491cf415dbc15f16e2411b4392e8fc34b889309e.jpg) +semi labeled-set-only self FixMatch SupCon SimCLR +-FlexMtech-MixUp BYOL + +![](images/7fd21f09a046eeba760d39e1bfa78d57b007ab829541543994066279d6e534e3.jpg) +Figure 1. Balanced accuracy over time profiles of semi- and self-supervised methods across all 4 datasets. At each time, we report the test set bal. acc. of each method (mean over 5 trials of Alg. D.1). Best viewed electronically. Top Row: On these 3 datasets, we compare all 6 semi- and 7 self- methods (see legend in lower right, citations in Sec. 2), to 3 labeled-set-only baselines. Bottom Row: On larger AIROGS dataset, we compare selected methods representative of the best in the top row. Thin lines show final performance to ease comparison. From these charts, we suggest that our unified training and tuning (Alg. D.1) is effective, as all methods show gains over time. + +![](images/79769279afefefe5ce417a5d014f6f367d2f8df10ac392213341f8c6e4d0f28d.jpg) +-MixMatch..Sup -DINO +—CoMatch —BarlowTw +—MeanTch —SwAV +-PseudoL -SimSiam +—MOCO + +on ImageNet. Compared to training from scratch, we see only slight gains (less than 3 points of BA) in ultimate test-set accuracy when top methods use pretraining, which aligns with observations in [66] but counters the "large margin" gains (sometimes over 10 points) reported by Su et al. [71]. Pretraining's benefits still include improved convergence speed and reduced variance across trials. + +ResNet-18 vs. ResNet-50. The bottom row of Fig 1 compares ResNet-18 and ResNet-50 architectures on AIROGS. We broadly find that the method rankings are similar. More notably, ResNet-50 is not substantially better than ResNet-18 in this task when compared using the same runtime budget. Profiles of other performance metrics in App C.3 suggest the same conclusions. + +Tuning vs. Transferring from another dataset. To make the most of limited labeled data, one strategy is to use the entire labeled set for training, reserving no validation set at all. This approach relies on pre-established "best" hyperparameters sourced from other datasets and transferred to the target task, obviating the need for data-specific tuning. Su et al. [71] employed this strategy in their benchmark, avoiding tuning due to concerns that small validation sets would yield unreliable selection. However, this claim in their work was not supported by experiments. Here, we directly compare our hyperparameter tuning strategy (Alg. D.1) with Su et al.'s transfer strategy. We train the latter on the combined train + validation sets, using "best" hyperparameters from CIFAR-10 and Tissue (details + +in App. F.2). Tab. 4 reports the final test-set balanced accuracy of each strategy on Path, TMED-2 and AIROGS. + +Our hyperparameter tuning strategy is competitive across all datasets: the "Tune" columns of Tab. 4 have the highest fraction of green cells. The alternative transfer strategy is faster and occasionally yield reasonable results. However, there is no guarantee of good performance from hyperparameter transfer, as evidenced by the inferior cases in Tab. 4 highlighted in yellow and red. There is also no consistent distinction between transferring from Tissue (an arguably more related medical dataset) and transferring from CIFAR-10, underscoring the challenge of identifying a "closely related" dataset for hyperparameter transfer. + +Overall, we recommend researchers pursuing SSL for a new medical classification task should adopt our hyperparameter tuning strategy. Transferring off-the-shelf hyperparameters has highly variable performance and can even result in divergence (see MixMatch on AIROGS in Tab. 4). Even with reasonable hyperparameters, having no validation set makes checkpoint selection challenging. Adjacent epochs that exhibit similar low training losses may differ by over $5\%$ in test accuracy in our experiments. Improving transfer from existing hyperparameters to a new task is an interesting further direction of research. + +
PathMNISTTMED-2AIROGS
Hyperparam. strategy:TuneBest from TissueBest from CIFAR-10TuneBest from TissueBest from CIFAR-10TuneBest from TissueBest from CIFAR-10
semiMixMatch82.2483.8773.5194.7192.6774.7377.1461.75Diverged
CoMatch83.9580.5186.4291.8792.5192.67
FixMatch79.1182.6479.7395.0695.1295.7366.9266.3165.50
FlexMatch80.9178.7080.2393.7793.9793.7067.2372.8274.52
Pseudo-Label70.4266.1264.7190.5288.7882.1368.2164.4366.93
Mean-Teacher68.2155.9056.1393.7091.2158.26
selfSimCLR81.0579.8680.0587.0484.4187.2473.0167.1567.95
BYOL81.7980.3580.1688.9088.4185.1572.4070.1370.63
SwAV77.9078.2481.3084.7682.1782.20
MoCo77.7379.9579.1276.1478.5679.11
SimSiam71.7868.6270.4483.2480.9380.20
BT80.8572.7980.2592.9691.6390.81
DINO82.5277.5080.6791.4584.8090.0367.6166.4667.13
+ +Table 4. Hyperparameter strategy evaluation. Table reports ultimate test-set balanced accuracy as percentage (mean over 5 trials, higher is better) on three target datasets: PathMNIST, TMED-2, and AIROGS. 3 different hyperparameter strategies are compared for each SSL method. Due to computation constraints, we only select a few algorithms to run on AIROGS, explained in Sec. 5. Tune strategy: tuning for best hyperparameters via our proposed Alg. D.1 for 100 hours (TMED-2 and AIROGS) and 25 hours (PathMNIST), using provided train/validation split. Best from Tissue/CIFAR-10 strategy: just one training phase on the target dataset using the "best" hyperparameters selected from the named source data. These runs combine train and validation data for fitting following Su et al. [71], and typically take 1-4/5-30/10-40 hours on Path/TMED-2/AIROGS depending on the algorithm and whether early stopping is triggered. To highlight which strategies are most effective, we bold the best strategy within each dataset-specific row, and color cells relative to this bold value as green (within 2.5 of best), yellow (within 5 of best), and red (worse by 5 or more points). + +# 7. DISCUSSION & CONCLUSION + +We have contributed a benchmark that helps practitioners quantify what gains in a classification task are possible from the addition of unlabeled data and which methods help achieve them. We offer a unified approach to training and hyperparameter selection of semi-supervised methods, self-supervised methods, and supervised baselines that is both affordable and realistic for research labs tackling new challenges without industrial-scale resources. + +Limitations. We deliberately focused on a modest number of labeled examples (30 - 1000) per class, motivated by projects where substantive effort has already gone into labeled data collection. Applications in more scarce-label regimes (e.g. zero-shot or few-shot) may need to also consult other benchmarks, as should those looking at hundreds of fine-grained classes. We also did not specifically study situations where the distribution of labeled and unlabeled data significantly differs [41, 44, 62, 67]. + +To enable thorough experiments, we focused on moderate ResNet architectures that have proven effective in clinical applications [42, 78]. As ViT-based methods become popular in medical imaging [1, 43], it would be valuable to understand if similar findings hold with vision transformers [26]. We leave this to future work. Ideally, our results would be verified across multiple train-validation splits; however the resources needed remain prohibitive. + +Our analysis here is limited the specific metrics of balanced accuracy and sensitivity/specificity (for AIROGS). Other clinically useful metrics, such as AUPRC, calibration, or net benefit, may be needed to decide if a classifier is appropriate for deployment [70]. For medical applications, it is also key to understand fairness across subpopulations [13] to avoid propagating structural disadvantages. + +Broader impact. All data analyzed here represent fully-deidentified open-access images approved for widespread use by their creators. We think the benefit of promoting these medical tasks to advance ML research outweighs the slight risk of patient reidentification by a bad actor. + +Outlook. Our experiments show that real benefits from the addition of unlabeled data are sometimes possible: our recommended methods see gains of $+5$ points of balanced accuracy on TissueMNIST, $+10$ points on PathMNIST, and $+5$ points on AIROGS against strong labeled-set-only baselines. In contrast, most tested methods do not add significant gains on TMED-2, perhaps due to that benchmark's uncurated nature [41]. We hope our work enables the research community to convert decades of effort on SSL into improved patient outcomes and better scientific understanding of disease and possible treatments. We further hope that our benchmark inspires those that pursue improved methodological contributions to favor realistic evaluation protocols and clinically-relevant datasets. + +# References + +[1] N Ahmadi, MY Tsang, AN Gu, TSM Tsang, and P Abolmaesumi. Transformer-based spatio-temporal analysis for classification of aortic stenosis severity from echocardiography cine series. IEEE Transactions on Medical Imaging, 2023. 8 +[2] Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zachary Beaver, Jan Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, Vivek Natarajan, and Mohammad Norouzi. Big Self-Supervised Models Advance Medical Image Classification. In International Conference on Computer Vision (ICCV). arXiv, 2021. 4 +[3] Shekoofeh Azizi, Laura Culp, Jan Freyberg, Basil Mustafa, Sebastien Baur, Simon Kornblith, Ting Chen, Nenad Tomasev, Jovana Mitrovic, Patricia Strachan, S. Sara Mahdavi, Ellery Wulczyn, Boris Babenko, Megan Walker, Aaron Loh, Po-Hsuan Cameron Chen, Yuan Liu, Pinal Bavishi, Scott Mayer McKinney, Jim Winkens, Abhijit Guha Roy, Zach Beaver, Fiona Ryan, Justin Krogue, Mozziyar Etemadi, Umesh Telang, Yun Liu, Lily Peng, Greg S. Corrado, Dale R. Webster, David Fleet, Geoffrey Hinton, Neil Houlsby, Alan Karthikesalingam, Mohammad Norouzi, and Vivek Natarajan. Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging. Nature Biomedical Engineering, 7(6):756-779, 2023. 4, 23 +[4] James Bergstra and Yoshua Bengio. Random search for hyper-parameter optimization. Journal of machine learning research, 13(2), 2012. 6 +[5] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 1, 3, 21, 23 +[6] Benjamin Billot, Colin Magdamo, Steven E Arnold, Sudeshna Das, and Juan Eugenio Iglesias. Robust segmentation of brain mri in the wild with hierarchical cnns and no retraining. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 538-548. Springer, 2022. 5 +[7] Avrim Blum and Tom Mitchell. Combining labeled and unlabeled data with co-training. In Proceedings of the eleventh annual conference on Computational learning theory, pages 92-100, 1998. 21 +[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 21 +[9] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 21 +[10] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 1, 3, 21 + +[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3 +[12] Paola Cascante-Bonilla, Fuwen Tan, Yanjun Qi, and Vicente Ordonez. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. In Proceedings of the AAAI Conference on Artificial Intelligence, 2021. 21 +[13] Leo Anthony Celi, Jacqueline Cellini, Marie-Laure Charpignon, Edward Christopher Dee, Franck Dernoncourt, Rene Eber, William Greig Mitchell, Lama Moukheiber, Julian Schirmer, et al. Sources of bias in artificial intelligence that perpetuate healthcare disparities—A global review. PLOS Digital Health, 1(3), 2022. 8 +[14] Olivier Chapelle, Bernhard Scholkopf, and Alexander Zien. Semi-supervised learning (chapelle, o. et al., eds.; 2006)[book reviews]. IEEE Transactions on Neural Networks, 20(3):542-542, 2009. 21 +[15] Chen Chen, Chen Qin, Huaqi Qiu, Cheng Ouyang, Shuo Wang, Liang Chen, Giacomo Tarroni, Wenjia Bai, and Daniel Rueckert. Realistic adversarial data augmentation for mr image segmentation. In Medical Image Computing and Computer Assisted Intervention MICCAI, 2020. 23 +[16] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 21 +[17] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 2, 3, 21 +[18] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 21 +[19] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 1, 3 +[20] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 3 +[21] Yanbei Chen, Massimiliano Mancini, Xiatian Zhu, and Zeynep Akata. Semi-supervised and unsupervised deep visual learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 4 +[22] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 6, 21 +[23] Victor Guilherme Turrisi Da Costa, Enrico Fini, Moin Nabi, Nicu Sebe, and Elisa Ricci. solo-learn: A library of self-supervised methods for visual representation learning. J. Mach. Learn. Res., 23(56):1-6, 2022. 4 + +[24] Coen de Vente, Koenraad A. Vermeer, Nicolas Jaccard, He Wang, Hongyi Sun, FIRas Khader, Daniel Truhn, Temirgali Aimshev, Yerkebulan Zhanibekuly, Tien-Dung Le, Adrian Galdran, Miguel Angel Gonzalez Ballester, Gustavo Carneiro, Devika R. G, Hrishikesh P. S, Densen Puthussery, Hong Liu, Zekang Yang, Satoshi Kondo, Satoshi Kasai, Edward Wang, Ashritha Durvasula, Jonathan Heras, Miguel Angel Zapata, Teresa Araujo, Guilherme Aresta, Hrvoje Bogunovic, Mustafa Arikan, Yeong Chan Lee, Hyun Bin Cho, Yoon Ho Choi, Abdul Qayyum, Imran Razzak, Bram van Ginneken, Hans G. Lemij, and Clara I. Sanchez. AIROGS: Artificial Intelligence for RObust Glaucoma Screening Challenge, 2023. 5, 15 +[25] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 21 +[26] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.8 +[27] Linus Ericsson, Henry Gouk, and Timothy M Hospedales. How well do self-supervised models transfer? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5414-5423, 2021. 1, 3, 4 +[28] Andre Esteva, Brett Kuprel, Roberto A Novoa, Justin Ko, Susan M Swetter, Helen M Blau, and Sebastian Thrun. Dermatologist-level classification of skin cancer with deep neural networks. nature, 542(7639):115-118, 2017. 5 +[29] Loveleen Gaur, Ujwal Bhatia, NZ Jhanjhi, Ghulam Muhammad, and Mehedi Masud. Medical image-based detection of Covid-19 using deep convolution neural networks. Multimedia systems, 29(3):1729-1738, 2023. +[30] Hemant Ghayvat, Muhammad Awais, AK Bashir, Sharnil Pandya, Mohd Zuhair, Mamoon Rashid, and Jamel Nebhen. Ai-enabled radiologist in the loop: novel ai-based framework to augment radiologist performance for Covid-19 chest CT medical image annotation and classification from pneumonia. Neural Computing and Applications, 35(20):14591–14609, 2023. +[31] Amirata Ghorbani, David Ouyang, Abubakar Abid, Bryan He, Jonathan H Chen, Robert A Harrington, David H Liang, Euan A Ashley, and James Y Zou. Deep learning interpretation of echocardiograms. NPJ digital medicine, 3(1):10, 2020. 5 +[32] Priya Goyal, Dhruv Mahajan, Abhinav Gupta, and Ishan Misra. Scaling and benchmarking self-supervised visual representation learning. In Proceedings of the IEEE/cvf International Conference on computer vision, pages 6391-6400, 2019. 4 +[33] Margherita Grandini, Enrico Bagli, and Giorgio Visani. Metrics for multi-class classification: an overview. arXiv preprint arXiv:2008.05756, 2020. 5 +[34] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, + +Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3 +[35] Isabelle Guyon, Kristin Bennett, Gavin Cawley, Hugo Jair Escalante, Sergio Escalera, Tin Kam Ho, Núria Macía, Bisakha Ray, Mehreen Saeed, Alexander Statnikov, et al. Design of the 2015 ChaLearn AutoML challenge. In 2015 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2015. 5 +[36] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[37] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 3, 21 +[38] Wassily Hoeffding. Probability inequalities for sums of bounded random variables. The collected works of Wassily Hoeffding, pages 409-426, 1994. 23 +[39] Zhe Huang, Gary Long, Benjamin Wessler, and Michael C Hughes. A new semi-supervised learning benchmark for classifying view and diagnosing aortic stenosis from echocardiograms. In Proceedings of the Machine Learning for Healthcare Conference. PMLR, 2021. 5, 15 +[40] Zhe Huang, Gary Long, Benjamin S Wessler, and Michael C Hughes. TMED 2: A dataset for semi-supervised classification of echocardiograms. In DataPerf: Benchmarking Data for Data-Centric AI Workshop, 2022. 5, 15 +[41] Zhe Huang, Mary-Joy Sidhom, Benjamin S Wessler, and Michael C Hughes. Fix-a-step: Semi-supervised learning from uncurated unlabeled data. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics (AISTATS), 2023. 5, 8 +[42] Zhe Huang, Benjamin S Wessler, and Michael C Hughes. Detecting heart disease from multi-view ultrasound images via supervised attention multiple instance learning. In Machine Learning for Healthcare Conference, pages 285-307. PMLR, 2023. 5, 8 +[43] Zhe Huang, Xiaowei Yu, Benjamin S Wessler, and Michael C Hughes. Semi-supervised multimodal multi-instance learning for aortic stenosis diagnosis. arXiv preprint arXiv:2403.06024, 2024. 8 +[44] Zhe Huang, Xiaowei Yu, Dajiang Zhu, and Michael C Hughes. Interlude: Interactions between labeled and unlabeled data to enhance semi-supervised learning. arXiv preprint arXiv:2403.10658, 2024. 8 +[45] Tongtong Huo, Yi Xie, Ying Fang, Ziyi Wang, Pengran Liu, Yuyu Duan, Jiayao Zhang, Honglin Wang, Mingdi Xue, Songxiang Liu, et al. Deep learning-based algorithm improves radiologists' performance in lung cancer bone metastases detection on computed tomography. Frontiers in Oncology, 13:1125637, 2023. 5 +[46] Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Label propagation for deep semi-supervised learning. + +In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5070-5079, 2019. 21 +[47] Longlong Jing and Yingli Tian. Self-supervised visual feature learning with deep neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(11):4037-4058, 2020. 21 +[48] Jakob Nikolas Kather, Johannes Krisam, Pornpimol Charoentong, Tom Luedde, Esther Herpel, Cleo-Aron Weis, Timo Gaiser, Alexander Marx, Nektarios A Valous, Dyke Ferber, et al. Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study. PLoS medicine, 16(1):e1002730, 2019. 15 +[49] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 3 +[50] Byoungjip Kim, Jinho Choo, Yeong-Dae Kwon, Seongho Joe, Seungjai Min, and Youngjune Gwon. Selfmatch: Combining contrastive self-supervision and consistency for semi-supervised learning. arXiv preprint arXiv:2101.06480, 2021. 4 +[51] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5 +[52] Durk P Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semi-supervised learning with deep generative models. Advances in neural information processing systems, 27, 2014. 21 +[53] Abhishek Kumar, Prasanna Sattigeri, and Tom Fletcher. Semi-supervised learning with gans: Manifold invariance with improved inference. Advances in neural information processing systems, 30, 2017. 21 +[54] Devidas T Kushnure, Shweta Tyagi, and Sanjay N Talbar. Lim-net: Lightweight multi-level multiscale network with deep residual learning for automatic liver segmentation in ct images. Biomedical Signal Processing and Control, 80: 104305, 2023. 5 +[55] Zhengfeng Lai, Chao Wang, Luca Cerny Oliveira, Brittany N Dugger, Sen-Ching Cheung, and Chen-Nee Chuah. Joint semi-supervised and active learning for segmentation of gigapixel pathology images with cost-effective labeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 591–600, 2021. +[56] Zhengfeng Lai, Chao Wang, Henry Gunawan, Sen-Ching S Cheung, and Chen-Nee Chuah. Smoothed adaptive weighting for imbalanced semi-supervised learning: Improve reliability against unknown distribution data. In International Conference on Machine Learning, pages 11828–11843. PMLR, 2022. 5 +[57] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 21 +[58] Dong-Hyun Lee. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning at ICML, 2013. 2, 3, 21 + +[59] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9475-9484, 2021. 3, 4 +[60] Ali Madani, Remy Arnaout, Mohammad Mofrad, and Rima Arnaout. Fast and accurate view classification of echocardiograms using deep learning. NPJ digital medicine, 1(1):6, 2018. 5 +[61] Shaobo Min, Xuejin Chen, Hongtao Xie, Zheng-Jun Zha, and Yongdong Zhang. A mutually attentive co-training framework for semi-supervised recognition. IEEE Transactions on Multimedia, 23:899-910, 2020. 21 +[62] Avital Oliver, Augustus Odena, Colin A Raffel, Ekin Dogus Cubuk, and Ian Goodfellow. Realistic evaluation of deep semi-supervised learning algorithms. Advances in neural information processing systems, 31, 2018. 1, 2, 3, 4, 6, 8, 13, 22, 23 +[63] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 21 +[64] Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. the Journal of machine learning research, 12:2825–2830, 2011. 20 +[65] Guo-Jun Qi and Jiebo Luo. Small data challenges in big data era: A survey of recent progress on unsupervised and semi-supervised methods. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2168-2187, 2020. 1, 4 +[66] Maithra Raghu, Chiyuan Zhang, Jon Kleinberg, and Samy Bengio. Transfusion: Understanding transfer learning for medical imaging. Advances in neural information processing systems, 32, 2019. 7 +[67] Kuniaki Saito, Donghyun Kim, and Kate Saenko. Openmatch: Open-set consistency regularization for semi-supervised learning with outliers. arXiv preprint arXiv:2105.14148, 2021. 8 +[68] Azizi Shekoofeh, Mustafa Basil, Ryan Fiona, Beaver Zachary, Freyberg Jan, Deaton Jonathan, Loh Aaron, Karthikesalingam Alan, Kornblith Simon, Chen Ting, et al. Big self-supervised models advance medical image classification. arXiv preprint arXiv:2101.05224, 2021. 23 +[69] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 1, 3, 5, 21 +[70] Ewout W Steyerberg and Yvonne Vergouwe. Towards better clinical prediction models: seven steps for development and an abcd for validation. European Heart Journal, 35(29), 2014. 8 +[71] Jong-Chyi Su, Zezhou Cheng, and Subhransu Maji. A realistic evaluation of semi-supervised learning for fine-grained classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12966-12975, 2021. 1, 3, 4, 5, 7, 8, 13, 16, 23, 25 + +[72] Terpeei Suzuki. Consistency regularization for semi-supervised learning with pytorch. https://github.com/perrying/pytorch-consistency-regularization, 2020.13 +[73] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 3, 21 +[74] Jesper E Van Engelen and Holger H Hoos. A survey on semi-supervised learning. Machine learning, 109(2):373-440, 2020. 1, 21 +[75] Diane Wagner, Fabio Ferreira, Danny Stoll, Robin Tibor Schirrmeister, Samuel Müller, and Frank Hutter. On the importance of hyperparameters and data augmentation for self-supervised learning. arXiv preprint arXiv:2207.07875, 2022. 1, 5 +[76] Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, et al. Usb: A unified semi-supervised learning benchmark for classification. Advances in Neural Information Processing Systems, 35:3938-3961, 2022. 2, 3, 4, 6 +[77] Benjamin S Wessler, Zhe Huang, Gary M Long Jr, Stefano Pacifici, Nishant Prashar, Samuel Karmiy, Roman A Sandler, Joseph Z Sokol, Daniel B Sokol, Monica M Dehn, et al. Automated detection of aortic stenosis using machine learning. Journal of the American Society of Echocardiography, 36(4): 411-420, 2023. 5 +[78] Nan Wu, Jason Phang, Jungkyu Park, Yiqiu Shen, Zhe Huang, Masha Zorin, Stanisław Jastrzebski, Thibault Févry, Joe Katsnelson, Eric Kim, et al. Deep neural networks improve radiologists' performance in breast cancer screening. IEEE transactions on medical imaging, 39(4):1184-1194, 2019. 5, 8 +[79] Jiancheng Yang, Rui Shi, and Bingbing Ni. Medmnist classification decathlon: A lightweight automl benchmark for medical image analysis. In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), pages 191-195. IEEE, 2021. 14 +[80] Jiancheng Yang, Rui Shi, Donglai Wei, Zequan Liu, Lin Zhao, Bilian Ke, Hanspeter Pfister, and Bingbing Ni. Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data, 10(1):41, 2023. 4, 14, 15 +[81] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Proceedings of the British Machine Vision Conference (BMVC), 2016. 5 +[82] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 3 +[83] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1476-1485, 2019. 4 +[84] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curricu + +lum pseudo labeling. Advances in Neural Information Processing Systems, 34:18408-18419, 2021. 3 +[85] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 3, 6, 21 +[86] Wenqiao Zhang, Lei Zhu, James Hallinan, Shengyu Zhang, Andrew Makmur, Qingpeng Cai, and Beng Chin Ooi. BoostMIS: Boosting medical image semi-supervised learning with adaptive pseudo labeling and informative active annotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 23 +[87] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14471-14481, 2022. 4 +[88] Xiaojin Zhu. Semi-Supervised Learning Literature Survey. Technical Report 1530, Department of Computer Science, University of Wisconsin Madison., 2005. 1, 21 +[89] Xiaojin Zhu, Zoubin Ghahramani, and John D Lafferty. Semi-supervised learning using gaussian fields and harmonic functions. In Proceedings of the 20th International conference on Machine learning (ICML-03), pages 912–919, 2003. 21 +[90] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2020. 21 \ No newline at end of file diff --git a/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/images.zip b/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0114a1a27076fbee554f9c565bb8d0c00b32eb1b --- /dev/null +++ b/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d0552493d69aa529455fa8401e046680d3c4da6883edb94a1170ccc4f56813b +size 440865 diff --git a/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/layout.json b/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..35fce572d290a2ecaf433a555c705914e28bd501 --- /dev/null +++ b/2024/Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification/layout.json @@ -0,0 +1,8999 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 516, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 516, + 140 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 516, + 140 + ], + "type": "text", + "content": "Systematic comparison of semi-supervised and self-supervised learning for medical image classification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 161, + 455, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 161, + 455, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 161, + 455, + 190 + ], + "type": "text", + "content": "Zhe Huang* Ruijie Jiang* Shuchin Aeron Michael C. Hughes Tufts University, School of Engineering" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 192, + 474, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 192, + 474, + 204 + ], + "spans": [ + { + "bbox": [ + 116, + 192, + 474, + 204 + ], + "type": "text", + "content": "{Zhe.Huang, Ruijie.Jiang, Shuchin.Aeron, Michael.Hughes}@tufts.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 256, + 289, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 256, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 256, + 289, + 616 + ], + "type": "text", + "content": "In typical medical image classification problems, labeled data is scarce while unlabeled data is more available. Semi-supervised learning and self-supervised learning are two different research directions that can improve accuracy by learning from extra unlabeled data. Recent methods from both directions have reported significant gains on traditional benchmarks. Yet past benchmarks do not focus on medical tasks and rarely compare self- and semi- methods together on an equal footing. Furthermore, past benchmarks often handle hyperparameter tuning suboptimally. First, they may not tune hyperparameters at all, leading to underfitting. Second, when tuning does occur, it often unrealistically uses a labeled validation set that is much larger than the training set. Therefore currently published rankings might not always corroborate with their practical utility. This study contributes a systematic evaluation of self- and semi- methods with a unified experimental protocol intended to guide a practitioner with scarce overall labeled data and a limited compute budget. We answer two key questions: Can hyperparameter tuning be effective with realistic-sized validation sets? If so, when all methods are tuned well, which self- or semi-supervised methods achieve the best accuracy? Our study compares 13 representative semi- and self-supervised methods to strong labeled-set-only baselines on 4 medical datasets. From " + }, + { + "bbox": [ + 47, + 256, + 289, + 616 + ], + "type": "inline_equation", + "content": "20000+" + }, + { + "bbox": [ + 47, + 256, + 289, + 616 + ], + "type": "text", + "content": " GPU hours of computation, we provide valuable best practices to resource-constrained practitioners: hyperparameter tuning is effective, and the semi-supervised method known as MixMatch delivers the most reliable gains across 4 datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 639, + 159, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 159, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 159, + 653 + ], + "type": "text", + "content": "1. INTRODUCTION" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 661, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 661, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 661, + 287, + 685 + ], + "type": "text", + "content": "Deep neural networks can deliver exceptional performance on classification tasks when trained with vast labeled" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 232, + 547, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 547, + 341 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 547, + 341 + ], + "type": "text", + "content": "datasets. However, in medical imaging applications assembling a large dataset with appropriate label can be prohibitively costly due to manual effort required by a human expert. In contrast, images alone, without labels, are often readily available in health records databases. In recent years, significant research has focused on developing methods that leverage both the large unlabeled set and a small labeled set to enhance image classifier training, aiming to surpass models trained solely on labeled data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 341, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 547, + 545 + ], + "type": "text", + "content": "Two ways of leveraging unlabeled data are particularly popular: semi-supervised and self-supervised learning, both often abbreviated as SSL. Recent efforts in semisupervised learning [74, 88] usually train deep classifiers jointly [5, 69] using an objective with two loss terms, one favoring labeled-set accuracy and the other favoring label-consistency or label-smoothness. Alternatively, self-supervised methods [65] take a two-stage approach, first training deep representations on the unlabeled set, then fine-tuning a classifier on the labeled set. Exemplars are numerous [10, 17, 19, 20, 37]. Despite the remarkable progress reported in each direction, these two paradigms have been largely developed independently [21]. A direct comparison of the two paradigms has been notably absent. A practitioner building a medical image classifier from limited labeled data may ask, \"Which recent semi- or self-supervised methods are likely to be most effective?\"" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 548, + 715 + ], + "type": "text", + "content": "Performance can be quite sensitive to hyperparameters for both semi-SL [69, 71] and self-SL [75]. We argue that any careful comparison must consider tuning hyperparameters of all methods in a fair fashion. However, recent work has not handled this well. First, as reviewed in Table 1, recent benchmarks for both semi- and self- paradigms often omit any tuning of hyperparameters (see the Fungi semi-SL experiments of Su et al. [71] or Ericsson et al. [27]'s self-SL benchmark). This practice of using off-the-shelf defaults likely leads to under-performing given a new task, and may impact some methods more than others. An even bigger issue with prevailing semi-SL practice was originally raised by Oliver et al. [62]: many published papers perform hyperparameter tuning on a labeled validation set that" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 286, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 286, + 715 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 286, + 715 + ], + "type": "text", + "content": "*Authors Zhe Huang and Ruijie Jiang contributed equally to this work. \nCode: github.com/tufts-ms/SSL-vs-SSL-benchmark [MIT license]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22282" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "is larger (sometimes much larger) than the limited labeled train set. Such experimental settings are unrealistic. SSL is intended for practitioners without abundant available labeled data. Practitioners that need to make the most of 1000 available labeled images will not elect to put more images in validation than training, and thus are not helped by benchmarks that do. Unfortunately, five years later after [62] we find this issue is still widespread (see Tab. 1). For example, Wang et al. [76]'s semi-SL results on TissueMNIST tune on a validation set over " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "50\\mathrm{x}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " larger than the labeled train set." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 289 + ], + "type": "text", + "content": "Oliver et al. [62] further cast some doubt on whether effective tuning is possible with small labeled sets, saying \"Extensive hyperparameter tuning may be somewhat futile due to an excessively small collection of held-out data to measure performance on\". There thus remains a pressing question for resource-constrained practitioners: \"Given limited available labeled data and limited compute, is hyperparameter tuning worthwhile?\"" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 289, + 289, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 479 + ], + "type": "text", + "content": "This study makes progress toward answering these questions by delivering a comprehensive comparison of semi- and self-supervised learning algorithms under a resource-constrained scenario that matches how SSL might be used on real-world medical tasks. Our goal is to enable practitioners to tackle new image classification challenges where only modest-sized labeled datasets exist. We target settings with roughly 2-10 class labels of interest, where each class has 30-1000 available labeled images for all model development (including training and validation). We select representative tasks across 4 datasets that span low-resolution (28x28) to moderate-resolution (112x112 and 384x384). On each task, we run careful experiments with representative recent methods from both semi- and self-supervised learning to clarify what gains are possible with unlabeled data and how to achieve them." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 481, + 289, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 289, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 289, + 589 + ], + "type": "text", + "content": "We emphasize realism throughout, in 4 distinct ways: (1) using validation set sizes that are never bigger than the train set, avoiding the unrealistically-large validation sets common in past work (Tab. 1); (2) performing hyperparameter search using the same protocol, same compute budget (fixed number of hours)* and same hardware (one NVIDIA A100 GPU) for all algorithms for fair comparison; (3) respecting natural class imbalance; (4) profiling performance over time, to inform labs with smaller runtime budgets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 590, + 251, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 590, + 251, + 601 + ], + "spans": [ + { + "bbox": [ + 59, + 590, + 251, + 601 + ], + "type": "text", + "content": "In summary, the contributions of this study are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 601, + 288, + 674 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 601, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 287, + 638 + ], + "type": "text", + "content": "1. We provide a systematic comparison of semi- and self-supervised methods on equal footing to connect two research directions that have been heretofore separate." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 638, + 288, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 288, + 674 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 288, + 674 + ], + "type": "text", + "content": "2. We adopt a realistic experimental protocol designed to consider the same constraints on available labels and runtime that SSL practitioners face, avoiding the unre" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 317, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 545, + 95 + ], + "type": "text", + "content": "alistic aspects of past benchmarks, especially the use of far too large validation sets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 96, + 547, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 547, + 131 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 547, + 131 + ], + "type": "text", + "content": "3. We show that hyperparameter tuning is viable with a realistic-sized validation set, and in many cases necessary to do well at a new classification task with new data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 132, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 132, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 132, + 547, + 168 + ], + "type": "text", + "content": "Ultimately, we hope this study guides practitioners with limited data toward successful deployment of semi- and self-supervised methods on real problems." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 178, + 501, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 178, + 501, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 501, + 190 + ], + "type": "text", + "content": "2. BACKGROUND AND METHODS" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "text", + "content": "Unified Problem Formulation. Following the recent survey by Chen et al. [21], we adopt a unified perspective for supervised, semi-supervised and self-supervised image classification with limited available labeled data. For model development (including training and hyperparameter selection), we assume there are two available datasets. First, a small labeled dataset " + }, + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "text", + "content": " of feature-label pairs " + }, + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "text", + "content": ", where each image is represented by a " + }, + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "text", + "content": "-dimensional feature vector " + }, + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "text", + "content": " and its corresponding class label takes one of " + }, + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "text", + "content": " possible values: " + }, + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "inline_equation", + "content": "y \\in \\{1,2,\\dots,C\\}" + }, + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "text", + "content": ". Second, an unlabeled set " + }, + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "text", + "content": " containing only feature vectors. Typically, we assume the unlabeled set is much larger: " + }, + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "inline_equation", + "content": "|\\mathcal{U}| \\gg |\\mathcal{L}|" + }, + { + "bbox": [ + 304, + 198, + 547, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "text", + "content": "Given labeled set " + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "text", + "content": " and unlabeled set " + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "text", + "content": ", we wish to train a neural network that can map each input " + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "text", + "content": " to a probability vector in the " + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "text", + "content": "-dimensional simplex " + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "inline_equation", + "content": "\\Delta^C" + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "text", + "content": " representing a distribution over " + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "text", + "content": " class labels. Let " + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "inline_equation", + "content": "f_v(\\cdot): \\mathbb{R}^D \\to \\mathbb{R}^F" + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "text", + "content": " denote a backbone neural network with parameters " + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "text", + "content": " producing an " + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "text", + "content": "-dimensional embedding given any input image. Let " + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "inline_equation", + "content": "g_w(\\cdot): \\mathbb{R}^F \\to \\Delta^C" + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "text", + "content": " denote a final linear-softmax classification layer with parameters " + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 342, + 547, + 449 + ], + "type": "text", + "content": ". The following unified objective can capture all three learning paradigms:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 455, + 545, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 455, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 318, + 455, + 545, + 491 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} v ^ {*}, w ^ {*} \\leftarrow \\arg \\min _ {v, w} \\sum_ {x, y \\in \\mathcal {L}} \\lambda^ {L} \\ell^ {L} (y, g _ {w} (f _ {v} (x))) \\tag {1} \\\\ + \\sum_ {x \\in \\mathcal {U}} \\lambda^ {U} \\ell^ {U} (x, f _ {v}, g _ {w}). \\\\ \\end{array}", + "image_path": "82dc1b74bccc69135438efa5199d4a23bc0a547a826a8cb61241f549f816648d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\ell^L" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": " represents a labeled-set loss (e.g. multi-class cross-entropy), and " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\ell^U" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": " represents a unlabeled-set loss. " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\lambda^L, \\lambda^U \\geq 0" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": " are weights for the corresponding loss terms. The design of " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\ell^U" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": " is usually what differs substantially across methods. For instance, setting " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\ell^U" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": " to be cross-entropy computed with pseudo-labels generated from classifier " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": " recovers PseudoLabel [58]; a temperature-scaled instance-similarity contrastive loss recovers SimCLR [17]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 677 + ], + "type": "text", + "content": "All three learning paradigms that we study optimize Eq. (1), yet differ in the number of phases and in how to set the scalar weights " + }, + { + "bbox": [ + 304, + 594, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\lambda^L, \\lambda^U" + }, + { + "bbox": [ + 304, + 594, + 547, + 677 + ], + "type": "text", + "content": " on each loss term. Supervised learning ignores the unlabeled term throughout training (" + }, + { + "bbox": [ + 304, + 594, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\lambda^U = 0" + }, + { + "bbox": [ + 304, + 594, + 547, + 677 + ], + "type": "text", + "content": "), thus learning parameters using only the labeled set. Semi-supervised methods include both terms in one end-to-end training, keeping both " + }, + { + "bbox": [ + 304, + 594, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\lambda^U > 0" + }, + { + "bbox": [ + 304, + 594, + 547, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\lambda^L > 0" + }, + { + "bbox": [ + 304, + 594, + 547, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Self-supervised learning has two phases. In phase 1 (\"pretraining\"), the labeled term is omitted (" + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\lambda^L = 0" + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\lambda^U = 1" + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": ") and the focus of learning is an effective representation" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "text", + "content": "*Oliver et al. [62] give each method 1000 trials of a cloud hyperparameter tuning service. However, training speed can differ substantially across SSL methods. We argue a fixed wallclock time budget is more fair." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22283" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 69, + 549, + 182 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 549, + 182 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 549, + 182 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 549, + 182 + ], + "type": "table", + "html": "
BenchmarkMethodsUnrealistic ExperimentsLabeled train sizeLabeled val. sizeAcc vs. Time?
Realistic eval. SSL [a]SemiCIFAR-10 (Tab. 1-2, Fig. 2)40005000no
SVHN (Tab. 1-2, Fig. 3-4)10007325
Fine-grained SSL [b]Semi & 1 Self*Semi-Aves59598000no
Semi-Fungi4141None
USB [c]SemiTissueMNIST80/40023640†no
Semi-Aves5959None
Self benchmark [d]SelfImageNet1.28 mil.Noneno
SSL-vs-SSL (ours)Semi & Selfno400-1660no bigger than trainyes
", + "image_path": "ad0f42e6e06b398df174d9b3833a92bcbd9166d948a05ec5342b566d927f235f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 190, + 547, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 257 + ], + "type": "text", + "content": "Table 1. Comparison of related benchmarks of semi-supervised and self-supervised learning. Past works either do no hyperparameter tuning at all (val. size = None), or use an unrealistically large validation set (defined as larger than the labeled train set) in some or all experiments. Further, each work almost exclusively looks at methods from one paradigm, either semi- or self- (*: [b] includes one self-supervised method, MOCO). In contrast, in this paper we benchmark 6 semi- and 7 self-supervised algorithms with hyperparameter tuning on realistic validation sets. Acc vs. Time? indicates whether the work analyzes performance over training time. Number marked † confirmed via GitHub comment by USB authors. Citations: a: Oliver et al. [62], b: Su et al. [71] c: Wang et al. [76] d: Ericsson et al. [27]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "text", + "content": "layer " + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "text", + "content": " (classifier " + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "inline_equation", + "content": "g_{w}" + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "text", + "content": " is not included in this phase). In phase 2 (\"fine-tuning\"), we focus on the labeled term and omit the unlabeled term " + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "inline_equation", + "content": "(\\lambda^{L} = 1, \\lambda^{U} = 0)" + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "text", + "content": ". We fix the representation parameter " + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "text", + "content": " and fine-tune the classifier " + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 274, + 287, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 325, + 262, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 325, + 262, + 336 + ], + "spans": [ + { + "bbox": [ + 59, + 325, + 262, + 336 + ], + "type": "text", + "content": "We now identify the 16 methods we will evaluate:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 338, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 445 + ], + "type": "text", + "content": "Supervised methods. The goal of leveraging unlabeled data " + }, + { + "bbox": [ + 46, + 338, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 46, + 338, + 287, + 445 + ], + "type": "text", + "content": " is to obtain better performance than what we could obtain using only the labeled set " + }, + { + "bbox": [ + 46, + 338, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 338, + 287, + 445 + ], + "type": "text", + "content": ". Therefore, we naturally compare to 3 high-quality supervised baselines that use only the labeled set. First, \"Sup\" denotes a classifier trained with supervised loss " + }, + { + "bbox": [ + 46, + 338, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\ell^L" + }, + { + "bbox": [ + 46, + 338, + 287, + 445 + ], + "type": "text", + "content": " set to multi-class cross-entropy. Second, \"MixUp\" trains with cross entropy with the addition of mixup data augmentation [85]. Finally, \"SupCon\" pursues a supervised contrastive learning loss for " + }, + { + "bbox": [ + 46, + 338, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\ell^L" + }, + { + "bbox": [ + 46, + 338, + 287, + 445 + ], + "type": "text", + "content": " [49]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 579 + ], + "type": "text", + "content": "Semi-supervised methods. We compare 6 semi-supervised methods that train deep classifiers on both labeled and unlabeled data simultaneously as in Eq. (1). To represent the state-of-the-art of semi-supervised image classification, we select Pseudo Label (\"PseudoL\") [58], Mean Teacher (\"MeanTch\") [73], MixMatch [5], FixMatch [69], FlexMatch [84] and CoMatch [59]. These choices cover a reasonably wide spectrum of unlabeled loss design strategy, year of publication, and computation cost. CoMatch represents a recent trend of combining semi- and self-supervision. See App D.2 for a wider literature review." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "Self-supervised methods. We compare 7 self-supervised algorithms: SimCLR [17], MOCO (v2) [20, 37], SwAV [10], BYOL [34], SimSiam [19], DINO [11] and Barlow Twins (\"BarlowTw\") [82]. These algorithms epitomize the field of self-supervised learning as of this writing. SimCLR, MOCO (v2), and SwAV are based on contrastive learning, which learns effective representations when provided with similar and dissimilar samples. BYOL was designed to circumvent the need for dissimilar samples. SimSiam is a simple yet effective Siamese representation learning method. DINO uses a teacher-student network architec" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 274, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 545, + 323 + ], + "type": "text", + "content": "ture for representation distillation. Barlow Twins preserves information and reduces redundancy by favoring a close-to-identity cross-correlation matrix between augmented pairs of data. See App D.3 for further review." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 334, + 421, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 334, + 421, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 421, + 346 + ], + "type": "text", + "content": "3. RELATED WORK" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 354, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 389 + ], + "type": "text", + "content": "Table 1 summarizes key attributes of existing major benchmarks for semi-supervised and self-supervised methods. We now discuss how our study situates in this context." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 390, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 546, + 605 + ], + "type": "text", + "content": "Comparison to Oliver et al. Oliver et al. [62] provide an influential benchmark of deep semi-supervised methods. Like our work, they highlight the pressing issue of using unrealistically large validation sets for hyperparameter tuning. However, most of their experiments (e.g. all their tables and their figures 2-4) still use this unrealistic setting to be comparable to other works; only one subsection (Sec. 4.6) examines validation sets no larger than the train set. Instead, we exclusively use validation set sizes no larger than training set, mimicking what practitioners would face in real applications. Our definition of a \"realistic\" size for a validation set – no larger than the labeled train set – is broader than Oliver et al.'s definition of " + }, + { + "bbox": [ + 304, + 390, + 546, + 605 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 390, + 546, + 605 + ], + "type": "text", + "content": " of the train set. We favor our definition because Oliver et al. find specifically that \"for validation sets of the same size " + }, + { + "bbox": [ + 304, + 390, + 546, + 605 + ], + "type": "inline_equation", + "content": "(100\\%)" + }, + { + "bbox": [ + 304, + 390, + 546, + 605 + ], + "type": "text", + "content": " as the training set, some differentiation between the approaches is possible.\" In contrast, they found reliable differentiation was not always possible with much smaller validation sets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "text", + "content": "Furthermore, the benchmarking results presented by Oliver et al. require each algorithm to complete \"1000 trials of Gaussian Process-based black-box optimization using Google Cloud ML Engine\". This resource-intensive process seems impractical for researchers outside well-funded industrial labs. We use more modest compute budgets and specifically allow each method the same total wallclock runtime for tuning, again obeying practical constraints." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Other semi-supervised benchmarks. TorchSSL [84]" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22284" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "benchmarked eight popular semi-supervised learning algorithm using a unified codebase. The same group later extended that effort to natural language and audio processing in their USB benchmark [76]. Another recent effort for fine-grained classification by Su et al. [71] evaluates semi-supervised learning on datasets that exhibit class imbalance and contains images from novel classes in the unlabeled set, studying the effect of different initializations and the contents of unlabeled data on the performance of semi-supervised methods. While making significant contributions, these works focused on semi-supervision almost exclusively and did not include multiple self-supervised learning methods, thus leaving open the questions \"How do the two paradigms compare? Which methods are best overall?\"" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 290, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 407 + ], + "type": "text", + "content": "Prior self-supervised benchmarks. Several works have proposed different benchmarks for self-supervised methods. Goyal et al. [32] introduced a benchmark that covers 9 different tasks, such as object detection and visual navigation. Ericsson et al. [27] compared thirteen top self-supervised models on 40 downstream tasks. Da Costa et al. [23] presented a library of self-supervised methods that can be easily plugged into different downstream tasks and datasets. However, these works mainly consider other self-SL methods without direct comparison to the semi-SL paradigm. Moreover, they mostly do not incorporate the use of a validation set for hyperparameter tuning at all. This can lead to suboptimal accuracy at test time and complicate comparisons to methods that do tune." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 407, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 290, + 514 + ], + "type": "text", + "content": "Self-supervision for medical images. Our work is complementary to recent efforts that assess self-supervised pipelines for medical image classification [2, 3]. They focus primarily on how to design multi-stage transfer learning pipelines and do not comprehensively compare many different self-supervised methods or any semi-supervised methods. Further, many datasets studied by [3] come from proprietary projects conducted at Google. In contrast, all our data and experiments are open and reproducible by others." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 515, + 289, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 289, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 289, + 587 + ], + "type": "text", + "content": "Combining semi- and self-supervision. Recent research has explored the fusion of semi-supervised learning and self-supervised learning ideas [50, 59, 83, 87]. Some recent surveys [21, 65] compare semi- and self-supervised learning, but they focus on literature review while we offer realistic and comprehensive benchmarking experiments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 597, + 194, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 194, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 194, + 609 + ], + "type": "text", + "content": "4. DATASETS AND TASKS" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 677 + ], + "type": "text", + "content": "We study four open-access medical image classification datasets, all with 2D images that are fully deidentified. Table 2 reports statistics for all train/test splits. The exact splits used in our study can be found in our open-source codebase, documented in App. A." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": "Two datasets - PathMNIST and TissueMNIST - are selected from the MedMNIST collection [80] (criteria in App. B.2). Prior experiments by Yang et al. [80] suggest" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 545, + 310 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 545, + 310 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 545, + 310 + ], + "type": "table", + "html": "
TissueMNISTPathMNIST
LabeledUnlab.LabeledUnlab.
TrainValTestTrainTrainValTestTrain
total40040047280165066total450450718089546
GE151516775851NORM39397417847
DCT191922337795MUC404010357966
POD191922027686ADI474713389319
LEU2828336911761STR47474219354
IE3737440215369BACK48488479461
TAL5959703124549DEB525233910308
PT95951120139108LYM525264310349
CD/CT1281281516552947MUS616159212121
TUM6464123312821
TMED-2AIROGS
LabeledUnlab.LabeledUnlab.
TrainValTestTrainTrainValTestTrain
total16602352019353500total600600600094242
PSAX22350342-Glaucoma6060600-
A2C32528319-No Glauc.5405405400-
A4C46239423-
PLAX650118935-
", + "image_path": "bcb04603f2917c0b00b92bd5dec78f9e269192871472c36a5e0f00f75e9a388e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 319, + 547, + 365 + ], + "lines": [ + { + "bbox": [ + 305, + 319, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 547, + 365 + ], + "type": "text", + "content": "Table 2. Summary statistics of train/validation/test splits for all datasets in our study. Each table's rows are arranged in ascending order based on the number of per-class labeled train images. Full description of class names can be found in App B.3." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 387, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 387, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 387, + 545, + 422 + ], + "type": "text", + "content": "their " + }, + { + "bbox": [ + 305, + 387, + 545, + 422 + ], + "type": "inline_equation", + "content": "28 \\times 28" + }, + { + "bbox": [ + 305, + 387, + 545, + 422 + ], + "type": "text", + "content": " resolution is a reasonable choice for rapid prototyping; using larger " + }, + { + "bbox": [ + 305, + 387, + 545, + 422 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 305, + 387, + 545, + 422 + ], + "type": "text", + "content": " resolution does not yield much more accurate classifiers for these two datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 424, + 547, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 547, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 547, + 484 + ], + "type": "text", + "content": "Two other datasets represent more moderate resolutions closely tied to contemporary clinical research: the 112x112 Tufts Medical Echocardiogram Dataset (TMED-2) and the 384x384 AIROGS dataset. Further details for each dataset are provided in a dedicated paragraph below." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 485, + 547, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 664 + ], + "type": "text", + "content": "For each dataset, our data splitting strategy closely mirrors the conditions of real SSL applications. First, we let labeled training and validation sets contain a natural distribution of classes even if that may be imbalanced. This reflects how data would likely be affordably collected (by random sampling) and avoids artificially balanced training sets that will not match the population an algorithm would encounter in a deployment. Second, to be realistic we ensure validation sets are never larger than the available labeled train sets. This is in contrast to previous benchmarks: Wang et al. [76]'s TissueMNIST hyperparameter search used a validation set of 23,640 images even though the labeled training set contained only 80 or 400 images. In practice, such a large validation set would almost certainly be re-allocated to improve training, as noted in Oliver et al. [62]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "TissueMNIST is a lightweight dataset of 28x28 images of human kidney cortex cells, organized into 8 categories. The original dataset is fully-labeled and distributed with predefined train/val/test split of 165,466/23,640/47,280 im" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22285" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "ages with some class imbalance. We assume a total labeling budget of 800 images, evenly split between training and validation (to ensure hyperparameter tuning can differentiate methods). We form a labeled training set of 400 images from the predefined training set, sampling each class by its frequency in the original training set. We form a labeled validation set of 400 images sampled from the predefined validation set. For unlabeled set, we keep all remaining images in the original training split, discarding known labels." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 186, + 289, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 289, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 289, + 307 + ], + "type": "text", + "content": "PathMNIST is another lightweight dataset of 28x28 images of patches from colorectal cancer histology slides that comprise of 9 tissue types. The original dataset is fully-labeled and distributed with predefined train/val/test split of 89,996/10,004/7,180 images. Class imbalance is less severe than TissueMNIST. We assume a total labeled data budget of 900 images, again evenly split between training and validation. Labeled train, labeled valid, and unlabeled sets are sampled via the same procedures as in TissueMNIST. For both MedMNIST datasets, we use the predefined test set." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 312, + 289, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 289, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 289, + 517 + ], + "type": "text", + "content": "TMED-2 [39, 40] is an open-access dataset of " + }, + { + "bbox": [ + 46, + 312, + 289, + 517 + ], + "type": "inline_equation", + "content": "112 \\times 112" + }, + { + "bbox": [ + 46, + 312, + 289, + 517 + ], + "type": "text", + "content": " 2D grayscale images captured from routine echocardiogram scans. Each scan produces dozens of ultrasound images of the heart, captured from multiple acquisition angles (i.e., different anatomic views). In this study, we adopt Huang et al.'s view-classification task: the goal is to classify each image into one of 4 canonical view types (see App. B.3). View classification is clinically important: measurements or diagnoses of heart disease can only be made when looking at the right anatomical view [60, 77]. We used the provided train/validation/test split (id #1), which is naturally imbalanced and matches our criteria for realistic sizing. We further use the provided large unlabeled set of 353,500 images. TMED-2's unlabeled set is both authentic (no true labels are available at all, unlike other benchmarks that \"forget\" known labels) and uncurated [41] (contains images of view types beyond the 4 classes in the labeled task)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": "AIROGS [24] is a public dataset released for a recent competition where the binary classification task is to decide whether the patient should be referred for glaucoma or not, given a color fundus image of the retina (eye). We selected this dataset because it represents an active research challenge even with all available labels. Furthermore, because labels for this dataset were acquired at great expense (multiple human annotators graded each of over 100k images), we hope our work helps assess how self-/semi-supervision could reduce the annotation costs of future challenges. We selected the " + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": " resolution favored by several high-performing competition entries. We chose a labeling budget of 1200 total images, evenly split between train and validation; the rare positive class (9 to 1 imbalance) is representative of many screening tasks. We took the rest of available images as the unlabeled train set." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 71, + 468, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 468, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 468, + 84 + ], + "type": "text", + "content": "5. EXPERIMENTAL DESIGN" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "text", + "content": "Performance metric. We use balanced accuracy [33, 35] as our primary performance metric. For a task with " + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "inline_equation", + "content": "C \\geq 2" + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "text", + "content": " classes, let " + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "inline_equation", + "content": "y_{1:N}" + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "text", + "content": " denote true labels for " + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "text", + "content": " examples in a test set, and " + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "inline_equation", + "content": "\\hat{y}_{1:N}" + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "text", + "content": " denote a classifier's predicted labels. Let " + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "inline_equation", + "content": "\\mathrm{TP}_c" + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "text", + "content": " count true positives for class " + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "text", + "content": " (number of correctly classified examples whose true label is " + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "text", + "content": "), and let " + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "inline_equation", + "content": "N_c" + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "text", + "content": " count the total number of examples with true label " + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "text", + "content": ". Then we compute balanced accuracy (BA) as a percentage:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 194, + 545, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 545, + 228 + ], + "type": "interline_equation", + "content": "\\mathbf{BA}(y_{1:N},\\hat{y}_{1:N}) = \\frac{1}{C}\\sum_{c = 1}^{C}\\frac{\\mathrm{TP}_{c}(y_{1:N},\\hat{y}_{1:N})}{N_{c}(y_{1:N})}\\cdot 100\\% \\tag{2}", + "image_path": "4e77b67079f17e3ec1d7b85b7dd931b02a16e209f21f6a8ff4ffffbf37e0bc27.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 235, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 545, + 270 + ], + "type": "text", + "content": "Balanced accuracy is more suitable than standard accuracy for imbalanced problems when each class matters equally. The expected BA of a uniform random guess is " + }, + { + "bbox": [ + 304, + 235, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\frac{100}{C}\\%" + }, + { + "bbox": [ + 304, + 235, + 545, + 270 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 271, + 545, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 545, + 306 + ], + "type": "text", + "content": "On AIROGS, we also track metrics recommended by its creators [24] for clinical utility: AUROC, partial AUROC (" + }, + { + "bbox": [ + 304, + 271, + 545, + 306 + ], + "type": "inline_equation", + "content": ">90\\%" + }, + { + "bbox": [ + 304, + 271, + 545, + 306 + ], + "type": "text", + "content": " specificity) and sensitivity-at- " + }, + { + "bbox": [ + 304, + 271, + 545, + 306 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 271, + 545, + 306 + ], + "type": "text", + "content": " -specificity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 307, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 545, + 366 + ], + "type": "text", + "content": "Architectures. CNN backbones are popular in medical imaging [6, 28-31, 42, 45, 54-56, 77, 78]. We use ResNet-18 [36] on Tissue and Path, and WideResNet-28-2 [81] on TMED-2. We experiment with both ResNet-18 and 50 [36] on AIROGS to assess architectural differences." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 366, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 474 + ], + "type": "text", + "content": "Training with early stopping. For each training phase, we perform minibatch gradient descent on Eq. (1) with Adam [51] optimizer and a cosine learning rate schedule [69]. Each training phase proceeds for up to 200 epochs, where one epoch represents enough minibatch updates to process the equivalent of the entire combined training set " + }, + { + "bbox": [ + 304, + 366, + 545, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{L} \\cup \\mathcal{U}" + }, + { + "bbox": [ + 304, + 366, + 545, + 474 + ], + "type": "text", + "content": ". After every epoch, we record balanced accuracy on the validation set. If this value plateaus for 20 consecutive epochs, we stop the current training phase early." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 474, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 545 + ], + "type": "text", + "content": "Hyperparameters. Semi- and self-supervised learning can both be sensitive to hyperparameters [71, 75]. Our evaluations tune both shared variables (e.g. learning rates, weight decay, unlabeled loss weight " + }, + { + "bbox": [ + 304, + 474, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\lambda^U" + }, + { + "bbox": [ + 304, + 474, + 545, + 545 + ], + "type": "text", + "content": ") as well as hyperparameters unique to each algorithm. See App. F.1 for details on all hyperparameters for all methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": "Unified procedure for training and hyperparameter tuning. We formulate a unified procedure mindful of realistic hardware and runtime constraints for a non-industrial-scale lab working on a new medical dataset. We assume each algorithm has access to one NVIDIA A100 GPU for a fixed number of hours. Within the allotted compute budget, for each algorithm we execute a serial random search over hyperparameters, sequentially sampling each new configuration and then training until either an early stopping criteria is satisfied or the maximum epoch is reached. We track the best-so-far classifier in terms of validation-set performance every epoch. Algorithm D.1 provides pseudocode for the procedure. Each time a new hyperparameter configuration is needed, we sample each hyperparameter value independ-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22286" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "dently from a distribution designed to cover its common settings in prior literature (App. F.1). Our choice of random search on a budget is thought to yield better performance than grid search [4], fairly expends the same effort to train all methods regardless of their cost-per-epoch or hyperparameter complexity, and does not assume industrial-scale access to 1000 cloud-computing trials as in Oliver et al. [62]. We find that performance saturates after about 25 hours per 80000 unlabeled examples. We thus allocate for the total time budget 25 hours for PathMNIST, 50 hours for TissueMNIST, and 100 hours for TMED-2. We allow 100 hours for AIROGS due to its larger resolution. Due to limited resources, we select only a few competitive methods to run on AIROGS based on the results on other datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 289, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 347 + ], + "type": "text", + "content": "Self-supervised classification phase. Self-supervised methods by definition do not utilize label information when training representation layer weights " + }, + { + "bbox": [ + 46, + 239, + 289, + 347 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 239, + 289, + 347 + ], + "type": "text", + "content": ". To enable a proper comparison, for all self-SL methods our Alg. D.1 introduces an additional classification layer, training it anew after each epoch, each time using a 10 trial random search for an L2-penalty regularization hyperparameter. We retain the best performing weights " + }, + { + "bbox": [ + 46, + 239, + 289, + 347 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 239, + 289, + 347 + ], + "type": "text", + "content": " on the validation set. We emphasize that " + }, + { + "bbox": [ + 46, + 239, + 289, + 347 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 239, + 289, + 347 + ], + "type": "text", + "content": " does not impact overall self-supervised training of " + }, + { + "bbox": [ + 46, + 239, + 289, + 347 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 239, + 289, + 347 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 348, + 289, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 289, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 289, + 419 + ], + "type": "text", + "content": "Data augmentation. Random flip and crop are used for all semi-supervised and supervised methods. MixMatch also utilizes MixUp [85], while RandAugment [22] is used by FixMatch and FlexMatch. For each self-supervised method as well as SupCon, we apply the same SimCLR augmentation: random flip, crop, color jitter, and grayscale." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 419, + 289, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 289, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 289, + 468 + ], + "type": "text", + "content": "Multiple trials. We repeat Alg. D.1 for each method across 5 separate trials (distinct random seeds). We record the mean balanced accuracy on valid and test sets of these trials every 60 minutes (every " + }, + { + "bbox": [ + 46, + 419, + 289, + 468 + ], + "type": "inline_equation", + "content": "30\\mathrm{min}" + }, + { + "bbox": [ + 46, + 419, + 289, + 468 + ], + "type": "text", + "content": " for Tissue and Path)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 478, + 192, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 192, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 192, + 491 + ], + "type": "text", + "content": "6. RESULTS & ANALYSIS" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 289, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 629 + ], + "type": "text", + "content": "In each panel of Fig. 1, we focus on one dataset and architecture, plotting for each algorithm one line showing test set balanced accuracy (averaged across 5 trials) as a function of time spent running Alg. D.1. Variance across trials is visualized in Fig C.4. Extra results on AIROGS showing other performance metrics over time (AUROC, partial AUROC, and sensitivity at " + }, + { + "bbox": [ + 46, + 498, + 289, + 629 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 498, + 289, + 629 + ], + "type": "text", + "content": " specificity) are found in Fig. C.3. We emphasize that following best practice, each checkpoint represents the best hyperparameters as selected on the validation set; we then report that checkpoint's test set performance. These results help us answer two key questions:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "text", + "content": "Is hyperparameter tuning worthwhile for SSL methods when validation set sizes cannot be larger than the train set? Across Fig. 1 and C.3, all 16 algorithms show roughly monotonic improvements in test performance over time, despite using a realistic-sized validation set. The final performance of every algorithm and dataset shows improvement due to our unified training and tuning procedure (fur" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 548, + 269 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 548, + 269 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 548, + 269 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 548, + 269 + ], + "type": "table", + "html": "
nameGain in Bal.Acc. over best labeled-set-only
PathTissueTMEDAIROGSmedianworst
MixMatch12.43.7-0.25.74.7-0.2
MOCO7.94.4-18.8n/a4.4-18.8
SwAV8.44.3-10.1n/a4.3-10.1
SimCLR11.05.9-7.81.63.8-7.8
BYOL12.16.2-6.00.93.5-6.0
BarlowTw11.22.2-1.9n/a2.2-1.9
SimSiam1.92.1-11.6n/a1.9-11.6
FlexMtech11.14.5-1.1-4.21.7-4.2
DINO12.66.5-3.4-3.81.6-3.8
FixMatch9.32.90.2-4.51.6-4.5
CoMatch14.10.2-3.0n/a0.2-3.0
MeanTch-1.6-1.9-1.2n/a-1.6-1.9
PseudoL0.6-1.4-4.4-3.2-2.3-4.4
ref. val.69.833.994.971.5
", + "image_path": "b5c1caa6b876ca95b2dbdd48a6e1f291c10eb6da222038fe4210180b0388b1a9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 277, + 547, + 366 + ], + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 366 + ], + "type": "text", + "content": "Table 3. Gains from SSL methods over only using labeled set, across 4 datasets. Self (italicized), Semi (normal). We report gains in percentage balanced accuracy (higher is better) in the mean final test set performance (averaged over 5 runs of Alg. D.1) over a reference value that represents the best labeled-set-only run among Sup (minimize cross entropy), SupCon, and MixUp. MixMatch has the highest median gain, and is the only method to never score notably worse (" + }, + { + "bbox": [ + 304, + 277, + 547, + 366 + ], + "type": "inline_equation", + "content": ">1" + }, + { + "bbox": [ + 304, + 277, + 547, + 366 + ], + "type": "text", + "content": " point) than the best labeled set only run." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 388, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 424 + ], + "type": "text", + "content": "ther discussion in App. E.1). Therefore, we answer: Yes, realistically-sized validation sets for hyperparameter tuning and checkpoint selection can be effective." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 425, + 547, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 547, + 520 + ], + "type": "text", + "content": "What are the best SSL methods? Among the many methods we study in Fig. 1, no method clearly outperforms others across all datasets. On Path, CoMatch, MixMatch, DINO, and BYOL perform best. On Tissue, self-supervised methods like DINO, BYOL, and SimCLR score well. On TMED-2, ultimately only FixMatch and MixMatch are competitive with the MixUp baseline. On AIROGS, only MixMatch, SimCLR and BYOL beat the Sup baseline." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 521, + 548, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 548, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 548, + 688 + ], + "type": "text", + "content": "For each semi- and self- SSL method, Tab. 3 reports its relative gain in balanced accuracy over the best labeled-set-only supervised baseline on each dataset. We conclude that MixMatch represents the best overall choice as it consistently performs at or near the top across all tasks. It is the only method to never deliver results worse by more than 1 percentage point than the best labeled-set-only baseline (see \"worst\" column of Tab. 3). This is in stark contrast to USB [76], where FixMatch and FlexMatch are ranked notably higher than MixMatch. Such differences stress the importance of choosing proper evaluation protocol tailored to specific needs. We do caution that hyperparameter tuning is strongly recommended for MixMatch to succeed on a new dataset (see AIROGS result in Tab. 4)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "Pretraining vs. from scratch. Plots in App. C show accuracy-over-time profiles using initial weights pretrained" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22287" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 211, + 179 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 211, + 179 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 211, + 179 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 211, + 179 + ], + "type": "image", + "image_path": "d5b0f3d7dae1ceee445e3d403de4cfecd9548f62debb697ce82095b30a09e424.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 215, + 70, + 376, + 179 + ], + "blocks": [ + { + "bbox": [ + 215, + 70, + 376, + 179 + ], + "lines": [ + { + "bbox": [ + 215, + 70, + 376, + 179 + ], + "spans": [ + { + "bbox": [ + 215, + 70, + 376, + 179 + ], + "type": "image", + "image_path": "08b5e3edfd6600c308e8c6cf71fd85e8ecee90584020fa583ec438a3695b4844.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 380, + 71, + 539, + 178 + ], + "blocks": [ + { + "bbox": [ + 380, + 71, + 539, + 178 + ], + "lines": [ + { + "bbox": [ + 380, + 71, + 539, + 178 + ], + "spans": [ + { + "bbox": [ + 380, + 71, + 539, + 178 + ], + "type": "image", + "image_path": "4cb7f33695ab7fcb2ab4a807491cf415dbc15f16e2411b4392e8fc34b889309e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 182, + 527, + 198 + ], + "lines": [ + { + "bbox": [ + 409, + 182, + 527, + 198 + ], + "spans": [ + { + "bbox": [ + 409, + 182, + 527, + 198 + ], + "type": "text", + "content": "semi labeled-set-only self FixMatch SupCon SimCLR" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 409, + 200, + 518, + 210 + ], + "lines": [ + { + "bbox": [ + 409, + 200, + 518, + 210 + ], + "spans": [ + { + "bbox": [ + 409, + 200, + 518, + 210 + ], + "type": "text", + "content": "-FlexMtech-MixUp BYOL" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 49, + 180, + 211, + 290 + ], + "blocks": [ + { + "bbox": [ + 49, + 180, + 211, + 290 + ], + "lines": [ + { + "bbox": [ + 49, + 180, + 211, + 290 + ], + "spans": [ + { + "bbox": [ + 49, + 180, + 211, + 290 + ], + "type": "image", + "image_path": "7fd21f09a046eeba760d39e1bfa78d57b007ab829541543994066279d6e534e3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 301, + 547, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 301, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 547, + 357 + ], + "type": "text", + "content": "Figure 1. Balanced accuracy over time profiles of semi- and self-supervised methods across all 4 datasets. At each time, we report the test set bal. acc. of each method (mean over 5 trials of Alg. D.1). Best viewed electronically. Top Row: On these 3 datasets, we compare all 6 semi- and 7 self- methods (see legend in lower right, citations in Sec. 2), to 3 labeled-set-only baselines. Bottom Row: On larger AIROGS dataset, we compare selected methods representative of the best in the top row. Thin lines show final performance to ease comparison. From these charts, we suggest that our unified training and tuning (Alg. D.1) is effective, as all methods show gains over time." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 215, + 180, + 375, + 289 + ], + "blocks": [ + { + "bbox": [ + 215, + 180, + 375, + 289 + ], + "lines": [ + { + "bbox": [ + 215, + 180, + 375, + 289 + ], + "spans": [ + { + "bbox": [ + 215, + 180, + 375, + 289 + ], + "type": "image", + "image_path": "79769279afefefe5ce417a5d014f6f367d2f8df10ac392213341f8c6e4d0f28d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 213, + 517, + 223 + ], + "lines": [ + { + "bbox": [ + 409, + 213, + 517, + 223 + ], + "spans": [ + { + "bbox": [ + 409, + 213, + 517, + 223 + ], + "type": "text", + "content": "-MixMatch..Sup -DINO" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 409, + 226, + 531, + 236 + ], + "lines": [ + { + "bbox": [ + 409, + 226, + 531, + 236 + ], + "spans": [ + { + "bbox": [ + 409, + 226, + 531, + 236 + ], + "type": "text", + "content": "—CoMatch —BarlowTw" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 409, + 239, + 518, + 247 + ], + "lines": [ + { + "bbox": [ + 409, + 239, + 518, + 247 + ], + "spans": [ + { + "bbox": [ + 409, + 239, + 518, + 247 + ], + "type": "text", + "content": "—MeanTch —SwAV" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 409, + 251, + 529, + 261 + ], + "lines": [ + { + "bbox": [ + 409, + 251, + 529, + 261 + ], + "spans": [ + { + "bbox": [ + 409, + 251, + 529, + 261 + ], + "type": "text", + "content": "-PseudoL -SimSiam" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 409, + 263, + 522, + 273 + ], + "lines": [ + { + "bbox": [ + 409, + 263, + 522, + 273 + ], + "spans": [ + { + "bbox": [ + 409, + 263, + 522, + 273 + ], + "type": "text", + "content": "—MOCO" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 377, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 461 + ], + "type": "text", + "content": "on ImageNet. Compared to training from scratch, we see only slight gains (less than 3 points of BA) in ultimate test-set accuracy when top methods use pretraining, which aligns with observations in [66] but counters the \"large margin\" gains (sometimes over 10 points) reported by Su et al. [71]. Pretraining's benefits still include improved convergence speed and reduced variance across trials." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 462, + 288, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 544 + ], + "type": "text", + "content": "ResNet-18 vs. ResNet-50. The bottom row of Fig 1 compares ResNet-18 and ResNet-50 architectures on AIROGS. We broadly find that the method rankings are similar. More notably, ResNet-50 is not substantially better than ResNet-18 in this task when compared using the same runtime budget. Profiles of other performance metrics in App C.3 suggest the same conclusions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Tuning vs. Transferring from another dataset. To make the most of limited labeled data, one strategy is to use the entire labeled set for training, reserving no validation set at all. This approach relies on pre-established \"best\" hyperparameters sourced from other datasets and transferred to the target task, obviating the need for data-specific tuning. Su et al. [71] employed this strategy in their benchmark, avoiding tuning due to concerns that small validation sets would yield unreliable selection. However, this claim in their work was not supported by experiments. Here, we directly compare our hyperparameter tuning strategy (Alg. D.1) with Su et al.'s transfer strategy. We train the latter on the combined train + validation sets, using \"best\" hyperparameters from CIFAR-10 and Tissue (details" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 377, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 377, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 305, + 377, + 545, + 403 + ], + "type": "text", + "content": "in App. F.2). Tab. 4 reports the final test-set balanced accuracy of each strategy on Path, TMED-2 and AIROGS." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 426, + 545, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 558 + ], + "type": "text", + "content": "Our hyperparameter tuning strategy is competitive across all datasets: the \"Tune\" columns of Tab. 4 have the highest fraction of green cells. The alternative transfer strategy is faster and occasionally yield reasonable results. However, there is no guarantee of good performance from hyperparameter transfer, as evidenced by the inferior cases in Tab. 4 highlighted in yellow and red. There is also no consistent distinction between transferring from Tissue (an arguably more related medical dataset) and transferring from CIFAR-10, underscoring the challenge of identifying a \"closely related\" dataset for hyperparameter transfer." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "Overall, we recommend researchers pursuing SSL for a new medical classification task should adopt our hyperparameter tuning strategy. Transferring off-the-shelf hyperparameters has highly variable performance and can even result in divergence (see MixMatch on AIROGS in Tab. 4). Even with reasonable hyperparameters, having no validation set makes checkpoint selection challenging. Adjacent epochs that exhibit similar low training losses may differ by over " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " in test accuracy in our experiments. Improving transfer from existing hyperparameters to a new task is an interesting further direction of research." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22288" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 54, + 69, + 541, + 266 + ], + "blocks": [ + { + "bbox": [ + 54, + 69, + 541, + 266 + ], + "lines": [ + { + "bbox": [ + 54, + 69, + 541, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 69, + 541, + 266 + ], + "type": "table", + "html": "
PathMNISTTMED-2AIROGS
Hyperparam. strategy:TuneBest from TissueBest from CIFAR-10TuneBest from TissueBest from CIFAR-10TuneBest from TissueBest from CIFAR-10
semiMixMatch82.2483.8773.5194.7192.6774.7377.1461.75Diverged
CoMatch83.9580.5186.4291.8792.5192.67
FixMatch79.1182.6479.7395.0695.1295.7366.9266.3165.50
FlexMatch80.9178.7080.2393.7793.9793.7067.2372.8274.52
Pseudo-Label70.4266.1264.7190.5288.7882.1368.2164.4366.93
Mean-Teacher68.2155.9056.1393.7091.2158.26
selfSimCLR81.0579.8680.0587.0484.4187.2473.0167.1567.95
BYOL81.7980.3580.1688.9088.4185.1572.4070.1370.63
SwAV77.9078.2481.3084.7682.1782.20
MoCo77.7379.9579.1276.1478.5679.11
SimSiam71.7868.6270.4483.2480.9380.20
BT80.8572.7980.2592.9691.6390.81
DINO82.5277.5080.6791.4584.8090.0367.6166.4667.13
", + "image_path": "9110be42a07950ec75ea9743160ae98742d09f7fb3364ae3a35a7ac3ba6a47de.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 274, + 547, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 547, + 373 + ], + "type": "text", + "content": "Table 4. Hyperparameter strategy evaluation. Table reports ultimate test-set balanced accuracy as percentage (mean over 5 trials, higher is better) on three target datasets: PathMNIST, TMED-2, and AIROGS. 3 different hyperparameter strategies are compared for each SSL method. Due to computation constraints, we only select a few algorithms to run on AIROGS, explained in Sec. 5. Tune strategy: tuning for best hyperparameters via our proposed Alg. D.1 for 100 hours (TMED-2 and AIROGS) and 25 hours (PathMNIST), using provided train/validation split. Best from Tissue/CIFAR-10 strategy: just one training phase on the target dataset using the \"best\" hyperparameters selected from the named source data. These runs combine train and validation data for fitting following Su et al. [71], and typically take 1-4/5-30/10-40 hours on Path/TMED-2/AIROGS depending on the algorithm and whether early stopping is triggered. To highlight which strategies are most effective, we bold the best strategy within each dataset-specific row, and color cells relative to this bold value as green (within 2.5 of best), yellow (within 5 of best), and red (worse by 5 or more points)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 392, + 233, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 233, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 233, + 405 + ], + "type": "text", + "content": "7. DISCUSSION & CONCLUSION" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 414, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 509 + ], + "type": "text", + "content": "We have contributed a benchmark that helps practitioners quantify what gains in a classification task are possible from the addition of unlabeled data and which methods help achieve them. We offer a unified approach to training and hyperparameter selection of semi-supervised methods, self-supervised methods, and supervised baselines that is both affordable and realistic for research labs tackling new challenges without industrial-scale resources." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "text", + "content": "Limitations. We deliberately focused on a modest number of labeled examples (30 - 1000) per class, motivated by projects where substantive effort has already gone into labeled data collection. Applications in more scarce-label regimes (e.g. zero-shot or few-shot) may need to also consult other benchmarks, as should those looking at hundreds of fine-grained classes. We also did not specifically study situations where the distribution of labeled and unlabeled data significantly differs [41, 44, 62, 67]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "To enable thorough experiments, we focused on moderate ResNet architectures that have proven effective in clinical applications [42, 78]. As ViT-based methods become popular in medical imaging [1, 43], it would be valuable to understand if similar findings hold with vision transformers [26]. We leave this to future work. Ideally, our results would be verified across multiple train-validation splits; however the resources needed remain prohibitive." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 394, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 545, + 479 + ], + "type": "text", + "content": "Our analysis here is limited the specific metrics of balanced accuracy and sensitivity/specificity (for AIROGS). Other clinically useful metrics, such as AUPRC, calibration, or net benefit, may be needed to decide if a classifier is appropriate for deployment [70]. For medical applications, it is also key to understand fairness across subpopulations [13] to avoid propagating structural disadvantages." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 482, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 545, + 542 + ], + "type": "text", + "content": "Broader impact. All data analyzed here represent fully-deidentified open-access images approved for widespread use by their creators. We think the benefit of promoting these medical tasks to advance ML research outweighs the slight risk of patient reidentification by a bad actor." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": "Outlook. Our experiments show that real benefits from the addition of unlabeled data are sometimes possible: our recommended methods see gains of " + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "inline_equation", + "content": "+5" + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": " points of balanced accuracy on TissueMNIST, " + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "inline_equation", + "content": "+10" + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": " points on PathMNIST, and " + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "inline_equation", + "content": "+5" + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": " points on AIROGS against strong labeled-set-only baselines. In contrast, most tested methods do not add significant gains on TMED-2, perhaps due to that benchmark's uncurated nature [41]. We hope our work enables the research community to convert decades of effort on SSL into improved patient outcomes and better scientific understanding of disease and possible treatments. We further hope that our benchmark inspires those that pursue improved methodological contributions to favor realistic evaluation protocols and clinically-relevant datasets." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22289" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] N Ahmadi, MY Tsang, AN Gu, TSM Tsang, and P Abolmaesumi. Transformer-based spatio-temporal analysis for classification of aortic stenosis severity from echocardiography cine series. IEEE Transactions on Medical Imaging, 2023. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 210 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 210 + ], + "type": "text", + "content": "[2] Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zachary Beaver, Jan Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, Vivek Natarajan, and Mohammad Norouzi. Big Self-Supervised Models Advance Medical Image Classification. In International Conference on Computer Vision (ICCV). arXiv, 2021. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 289, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 289, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 289, + 356 + ], + "type": "text", + "content": "[3] Shekoofeh Azizi, Laura Culp, Jan Freyberg, Basil Mustafa, Sebastien Baur, Simon Kornblith, Ting Chen, Nenad Tomasev, Jovana Mitrovic, Patricia Strachan, S. Sara Mahdavi, Ellery Wulczyn, Boris Babenko, Megan Walker, Aaron Loh, Po-Hsuan Cameron Chen, Yuan Liu, Pinal Bavishi, Scott Mayer McKinney, Jim Winkens, Abhijit Guha Roy, Zach Beaver, Fiona Ryan, Justin Krogue, Mozziyar Etemadi, Umesh Telang, Yun Liu, Lily Peng, Greg S. Corrado, Dale R. Webster, David Fleet, Geoffrey Hinton, Neil Houlsby, Alan Karthikesalingam, Mohammad Norouzi, and Vivek Natarajan. Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging. Nature Biomedical Engineering, 7(6):756-779, 2023. 4, 23" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 357, + 287, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 389 + ], + "type": "text", + "content": "[4] James Bergstra and Yoshua Bengio. Random search for hyper-parameter optimization. Journal of machine learning research, 13(2), 2012. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 390, + 288, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 390, + 288, + 444 + ], + "spans": [ + { + "bbox": [ + 53, + 390, + 288, + 444 + ], + "type": "text", + "content": "[5] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 1, 3, 21, 23" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 446, + 288, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 446, + 288, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 446, + 288, + 512 + ], + "type": "text", + "content": "[6] Benjamin Billot, Colin Magdamo, Steven E Arnold, Sudeshna Das, and Juan Eugenio Iglesias. Robust segmentation of brain mri in the wild with hierarchical cnns and no retraining. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 538-548. Springer, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 514, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 555 + ], + "type": "text", + "content": "[7] Avrim Blum and Tom Mitchell. Combining labeled and unlabeled data with co-training. In Proceedings of the eleventh annual conference on Computational learning theory, pages 92-100, 1998. 21" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 557, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 557, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 557, + 287, + 612 + ], + "type": "text", + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 21" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 613, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 613, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 53, + 613, + 287, + 657 + ], + "type": "text", + "content": "[9] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 21" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[10] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 1, 3, 21" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 129 + ], + "type": "text", + "content": "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 129, + 547, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 174 + ], + "type": "text", + "content": "[12] Paola Cascante-Bonilla, Fuwen Tan, Yanjun Qi, and Vicente Ordonez. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. In Proceedings of the AAAI Conference on Artificial Intelligence, 2021. 21" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 175, + 547, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 242 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 242 + ], + "type": "text", + "content": "[13] Leo Anthony Celi, Jacqueline Cellini, Marie-Laure Charpignon, Edward Christopher Dee, Franck Dernoncourt, Rene Eber, William Greig Mitchell, Lama Moukheiber, Julian Schirmer, et al. Sources of bias in artificial intelligence that perpetuate healthcare disparities—A global review. PLOS Digital Health, 1(3), 2022. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 243, + 547, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 547, + 288 + ], + "type": "text", + "content": "[14] Olivier Chapelle, Bernhard Scholkopf, and Alexander Zien. Semi-supervised learning (chapelle, o. et al., eds.; 2006)[book reviews]. IEEE Transactions on Neural Networks, 20(3):542-542, 2009. 21" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 289, + 547, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 289, + 547, + 345 + ], + "spans": [ + { + "bbox": [ + 308, + 289, + 547, + 345 + ], + "type": "text", + "content": "[15] Chen Chen, Chen Qin, Huaqi Qiu, Cheng Ouyang, Shuo Wang, Liang Chen, Giacomo Tarroni, Wenjia Bai, and Daniel Rueckert. Realistic adversarial data augmentation for mr image segmentation. In Medical Image Computing and Computer Assisted Intervention MICCAI, 2020. 23" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 346, + 547, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 346, + 547, + 391 + ], + "spans": [ + { + "bbox": [ + 308, + 346, + 547, + 391 + ], + "type": "text", + "content": "[16] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 21" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 392, + 547, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 392, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 392, + 547, + 437 + ], + "type": "text", + "content": "[17] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 2, 3, 21" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 438, + 547, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 547, + 483 + ], + "type": "text", + "content": "[18] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 21" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 484, + 547, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 484, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 484, + 547, + 529 + ], + "type": "text", + "content": "[19] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 530, + 546, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 530, + 546, + 564 + ], + "spans": [ + { + "bbox": [ + 308, + 530, + 546, + 564 + ], + "type": "text", + "content": "[20] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 565, + 547, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 565, + 547, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 565, + 547, + 610 + ], + "type": "text", + "content": "[21] Yanbei Chen, Massimiliano Mancini, Xiatian Zhu, and Zeynep Akata. Semi-supervised and unsupervised deep visual learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 611, + 547, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 611, + 547, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 611, + 547, + 667 + ], + "type": "text", + "content": "[22] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 6, 21" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 669, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 547, + 712 + ], + "type": "text", + "content": "[23] Victor Guilherme Turrisi Da Costa, Enrico Fini, Moin Nabi, Nicu Sebe, and Elisa Ricci. solo-learn: A library of self-supervised methods for visual representation learning. J. Mach. Learn. Res., 23(56):1-6, 2022. 4" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22290" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 216 + ], + "type": "text", + "content": "[24] Coen de Vente, Koenraad A. Vermeer, Nicolas Jaccard, He Wang, Hongyi Sun, FIRas Khader, Daniel Truhn, Temirgali Aimshev, Yerkebulan Zhanibekuly, Tien-Dung Le, Adrian Galdran, Miguel Angel Gonzalez Ballester, Gustavo Carneiro, Devika R. G, Hrishikesh P. S, Densen Puthussery, Hong Liu, Zekang Yang, Satoshi Kondo, Satoshi Kasai, Edward Wang, Ashritha Durvasula, Jonathan Heras, Miguel Angel Zapata, Teresa Araujo, Guilherme Aresta, Hrvoje Bogunovic, Mustafa Arikan, Yeong Chan Lee, Hyun Bin Cho, Yoon Ho Choi, Abdul Qayyum, Imran Razzak, Bram van Ginneken, Hans G. Lemij, and Clara I. Sanchez. AIROGS: Artificial Intelligence for RObust Glaucoma Screening Challenge, 2023. 5, 15" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 217, + 288, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 261 + ], + "type": "text", + "content": "[25] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 21" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 262, + 288, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 262, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 48, + 262, + 288, + 328 + ], + "type": "text", + "content": "[26] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 330, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 330, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 330, + 288, + 374 + ], + "type": "text", + "content": "[27] Linus Ericsson, Henry Gouk, and Timothy M Hospedales. How well do self-supervised models transfer? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5414-5423, 2021. 1, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 375, + 288, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 375, + 288, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 375, + 288, + 419 + ], + "type": "text", + "content": "[28] Andre Esteva, Brett Kuprel, Roberto A Novoa, Justin Ko, Susan M Swetter, Helen M Blau, and Sebastian Thrun. Dermatologist-level classification of skin cancer with deep neural networks. nature, 542(7639):115-118, 2017. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 420, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 288, + 464 + ], + "type": "text", + "content": "[29] Loveleen Gaur, Ujwal Bhatia, NZ Jhanjhi, Ghulam Muhammad, and Mehedi Masud. Medical image-based detection of Covid-19 using deep convolution neural networks. Multimedia systems, 29(3):1729-1738, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 466, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 288, + 542 + ], + "type": "text", + "content": "[30] Hemant Ghayvat, Muhammad Awais, AK Bashir, Sharnil Pandya, Mohd Zuhair, Mamoon Rashid, and Jamel Nebhen. Ai-enabled radiologist in the loop: novel ai-based framework to augment radiologist performance for Covid-19 chest CT medical image annotation and classification from pneumonia. Neural Computing and Applications, 35(20):14591–14609, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 544, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 599 + ], + "type": "text", + "content": "[31] Amirata Ghorbani, David Ouyang, Abubakar Abid, Bryan He, Jonathan H Chen, Robert A Harrington, David H Liang, Euan A Ashley, and James Y Zou. Deep learning interpretation of echocardiograms. NPJ digital medicine, 3(1):10, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 600, + 288, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 288, + 654 + ], + "type": "text", + "content": "[32] Priya Goyal, Dhruv Mahajan, Abhinav Gupta, and Ishan Misra. Scaling and benchmarking self-supervised visual representation learning. In Proceedings of the IEEE/cvf International Conference on computer vision, pages 6391-6400, 2019. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 656, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 656, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 656, + 288, + 690 + ], + "type": "text", + "content": "[33] Margherita Grandini, Enrico Bagli, and Giorgio Visani. Metrics for multi-class classification: an overview. arXiv preprint arXiv:2008.05756, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[34] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch," + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 72, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 118 + ], + "type": "text", + "content": "Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "type": "text", + "content": "[35] Isabelle Guyon, Kristin Bennett, Gavin Cawley, Hugo Jair Escalante, Sergio Escalera, Tin Kam Ho, Núria Macía, Bisakha Ray, Mehreen Saeed, Alexander Statnikov, et al. Design of the 2015 ChaLearn AutoML challenge. In 2015 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2015. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 185, + 547, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 547, + 229 + ], + "type": "text", + "content": "[36] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 230, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 285 + ], + "type": "text", + "content": "[37] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 3, 21" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 285, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 319 + ], + "type": "text", + "content": "[38] Wassily Hoeffding. Probability inequalities for sums of bounded random variables. The collected works of Wassily Hoeffding, pages 409-426, 1994. 23" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 320, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 375 + ], + "type": "text", + "content": "[39] Zhe Huang, Gary Long, Benjamin Wessler, and Michael C Hughes. A new semi-supervised learning benchmark for classifying view and diagnosing aortic stenosis from echocardiograms. In Proceedings of the Machine Learning for Healthcare Conference. PMLR, 2021. 5, 15" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "text", + "content": "[40] Zhe Huang, Gary Long, Benjamin S Wessler, and Michael C Hughes. TMED 2: A dataset for semi-supervised classification of echocardiograms. In DataPerf: Benchmarking Data for Data-Centric AI Workshop, 2022. 5, 15" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 422, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 477 + ], + "type": "text", + "content": "[41] Zhe Huang, Mary-Joy Sidhom, Benjamin S Wessler, and Michael C Hughes. Fix-a-step: Semi-supervised learning from uncurated unlabeled data. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics (AISTATS), 2023. 5, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "type": "text", + "content": "[42] Zhe Huang, Benjamin S Wessler, and Michael C Hughes. Detecting heart disease from multi-view ultrasound images via supervised attention multiple instance learning. In Machine Learning for Healthcare Conference, pages 285-307. PMLR, 2023. 5, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "type": "text", + "content": "[43] Zhe Huang, Xiaowei Yu, Benjamin S Wessler, and Michael C Hughes. Semi-supervised multimodal multi-instance learning for aortic stenosis diagnosis. arXiv preprint arXiv:2403.06024, 2024. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 578, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 624 + ], + "type": "text", + "content": "[44] Zhe Huang, Xiaowei Yu, Dajiang Zhu, and Michael C Hughes. Interlude: Interactions between labeled and unlabeled data to enhance semi-supervised learning. arXiv preprint arXiv:2403.10658, 2024. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 624, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 690 + ], + "type": "text", + "content": "[45] Tongtong Huo, Yi Xie, Ying Fang, Ziyi Wang, Pengran Liu, Yuyu Duan, Jiayao Zhang, Honglin Wang, Mingdi Xue, Songxiang Liu, et al. Deep learning-based algorithm improves radiologists' performance in lung cancer bone metastases detection on computed tomography. Frontiers in Oncology, 13:1125637, 2023. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "text", + "content": "[46] Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Label propagation for deep semi-supervised learning." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22291" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5070-5079, 2019. 21" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[47] Longlong Jing and Yingli Tian. Self-supervised visual feature learning with deep neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(11):4037-4058, 2020. 21" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 207 + ], + "type": "text", + "content": "[48] Jakob Nikolas Kather, Johannes Krisam, Pornpimol Charoentong, Tom Luedde, Esther Herpel, Cleo-Aron Weis, Timo Gaiser, Alexander Marx, Nektarios A Valous, Dyke Ferber, et al. Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study. PLoS medicine, 16(1):e1002730, 2019. 15" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "text", + "content": "[49] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 317 + ], + "type": "text", + "content": "[50] Byoungjip Kim, Jinho Choo, Yeong-Dae Kwon, Seongho Joe, Seungjai Min, and Youngjune Gwon. Selfmatch: Combining contrastive self-supervision and consistency for semi-supervised learning. arXiv preprint arXiv:2101.06480, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 352 + ], + "type": "text", + "content": "[51] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "text", + "content": "[52] Durk P Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semi-supervised learning with deep generative models. Advances in neural information processing systems, 27, 2014. 21" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "text", + "content": "[53] Abhishek Kumar, Prasanna Sattigeri, and Tom Fletcher. Semi-supervised learning with gans: Manifold invariance with improved inference. Advances in neural information processing systems, 30, 2017. 21" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "type": "text", + "content": "[54] Devidas T Kushnure, Shweta Tyagi, and Sanjay N Talbar. Lim-net: Lightweight multi-level multiscale network with deep residual learning for automatic liver segmentation in ct images. Biomedical Signal Processing and Control, 80: 104305, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 501, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 567 + ], + "type": "text", + "content": "[55] Zhengfeng Lai, Chao Wang, Luca Cerny Oliveira, Brittany N Dugger, Sen-Ching Cheung, and Chen-Nee Chuah. Joint semi-supervised and active learning for segmentation of gigapixel pathology images with cost-effective labeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 591–600, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 633 + ], + "type": "text", + "content": "[56] Zhengfeng Lai, Chao Wang, Henry Gunawan, Sen-Ching S Cheung, and Chen-Nee Chuah. Smoothed adaptive weighting for imbalanced semi-supervised learning: Improve reliability against unknown distribution data. In International Conference on Machine Learning, pages 11828–11843. PMLR, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "text", + "content": "[57] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 21" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[58] Dong-Hyun Lee. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning at ICML, 2013. 2, 3, 21" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[59] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9475-9484, 2021. 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 119, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 161 + ], + "type": "text", + "content": "[60] Ali Madani, Remy Arnaout, Mohammad Mofrad, and Rima Arnaout. Fast and accurate view classification of echocardiograms using deep learning. NPJ digital medicine, 1(1):6, 2018. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[61] Shaobo Min, Xuejin Chen, Hongtao Xie, Zheng-Jun Zha, and Yongdong Zhang. A mutually attentive co-training framework for semi-supervised recognition. IEEE Transactions on Multimedia, 23:899-910, 2020. 21" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 209, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 261 + ], + "type": "text", + "content": "[62] Avital Oliver, Augustus Odena, Colin A Raffel, Ekin Dogus Cubuk, and Ian Goodfellow. Realistic evaluation of deep semi-supervised learning algorithms. Advances in neural information processing systems, 31, 2018. 1, 2, 3, 4, 6, 8, 13, 22, 23" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "type": "text", + "content": "[63] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 21" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 299, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 545, + 353 + ], + "type": "text", + "content": "[64] Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. the Journal of machine learning research, 12:2825–2830, 2011. 20" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 355, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 398 + ], + "type": "text", + "content": "[65] Guo-Jun Qi and Jiebo Luo. Small data challenges in big data era: A survey of recent progress on unsupervised and semi-supervised methods. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2168-2187, 2020. 1, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "text", + "content": "[66] Maithra Raghu, Chiyuan Zhang, Jon Kleinberg, and Samy Bengio. Transfusion: Understanding transfer learning for medical imaging. Advances in neural information processing systems, 32, 2019. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "type": "text", + "content": "[67] Kuniaki Saito, Donghyun Kim, and Kate Saenko. Openmatch: Open-set consistency regularization for semi-supervised learning with outliers. arXiv preprint arXiv:2105.14148, 2021. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 491, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 491, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 491, + 545, + 544 + ], + "type": "text", + "content": "[68] Azizi Shekoofeh, Mustafa Basil, Ryan Fiona, Beaver Zachary, Freyberg Jan, Deaton Jonathan, Loh Aaron, Karthikesalingam Alan, Kornblith Simon, Chen Ting, et al. Big self-supervised models advance medical image classification. arXiv preprint arXiv:2101.05224, 2021. 23" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 546, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 545, + 611 + ], + "type": "text", + "content": "[69] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 1, 3, 5, 21" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 656 + ], + "type": "text", + "content": "[70] Ewout W Steyerberg and Yvonne Vergouwe. Towards better clinical prediction models: seven steps for development and an abcd for validation. European Heart Journal, 35(29), 2014. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[71] Jong-Chyi Su, Zezhou Cheng, and Subhransu Maji. A realistic evaluation of semi-supervised learning for fine-grained classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12966-12975, 2021. 1, 3, 4, 5, 7, 8, 13, 16, 23, 25" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22292" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 287, + 116 + ], + "type": "text", + "content": "[72] Terpeei Suzuki. Consistency regularization for semi-supervised learning with pytorch. https://github.com/perrying/pytorch-consistency-regularization, 2020.13" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "text", + "content": "[73] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 3, 21" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 288, + 195 + ], + "type": "text", + "content": "[74] Jesper E Van Engelen and Holger H Hoos. A survey on semi-supervised learning. Machine learning, 109(2):373-440, 2020. 1, 21" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 287, + 251 + ], + "type": "text", + "content": "[75] Diane Wagner, Fabio Ferreira, Danny Stoll, Robin Tibor Schirrmeister, Samuel Müller, and Frank Hutter. On the importance of hyperparameters and data augmentation for self-supervised learning. arXiv preprint arXiv:2207.07875, 2022. 1, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 254, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 254, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 254, + 287, + 308 + ], + "type": "text", + "content": "[76] Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, et al. Usb: A unified semi-supervised learning benchmark for classification. Advances in Neural Information Processing Systems, 35:3938-3961, 2022. 2, 3, 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 310, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 310, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 49, + 310, + 287, + 374 + ], + "type": "text", + "content": "[77] Benjamin S Wessler, Zhe Huang, Gary M Long Jr, Stefano Pacifici, Nishant Prashar, Samuel Karmiy, Roman A Sandler, Joseph Z Sokol, Daniel B Sokol, Monica M Dehn, et al. Automated detection of aortic stenosis using machine learning. Journal of the American Society of Echocardiography, 36(4): 411-420, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 377, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 377, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 49, + 377, + 287, + 441 + ], + "type": "text", + "content": "[78] Nan Wu, Jason Phang, Jungkyu Park, Yiqiu Shen, Zhe Huang, Masha Zorin, Stanisław Jastrzebski, Thibault Févry, Joe Katsnelson, Eric Kim, et al. Deep neural networks improve radiologists' performance in breast cancer screening. IEEE transactions on medical imaging, 39(4):1184-1194, 2019. 5, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 445, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 445, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 445, + 287, + 498 + ], + "type": "text", + "content": "[79] Jiancheng Yang, Rui Shi, and Bingbing Ni. Medmnist classification decathlon: A lightweight automl benchmark for medical image analysis. In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), pages 191-195. IEEE, 2021. 14" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 500, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 500, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 49, + 500, + 287, + 554 + ], + "type": "text", + "content": "[80] Jiancheng Yang, Rui Shi, Donglai Wei, Zequan Liu, Lin Zhao, Bilian Ke, Hanspeter Pfister, and Bingbing Ni. Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data, 10(1):41, 2023. 4, 14, 15" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 556, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 556, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 49, + 556, + 287, + 588 + ], + "type": "text", + "content": "[81] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Proceedings of the British Machine Vision Conference (BMVC), 2016. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 49, + 590, + 287, + 634 + ], + "type": "text", + "content": "[82] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 49, + 636, + 287, + 679 + ], + "type": "text", + "content": "[83] Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1476-1485, 2019. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "type": "text", + "content": "[84] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curricu" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 385 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "lum pseudo labeling. Advances in Neural Information Processing Systems, 34:18408-18419, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[85] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 3, 6, 21" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 130, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 195 + ], + "type": "text", + "content": "[86] Wenqiao Zhang, Lei Zhu, James Hallinan, Shengyu Zhang, Andrew Makmur, Qingpeng Cai, and Beng Chin Ooi. BoostMIS: Boosting medical image semi-supervised learning with adaptive pseudo labeling and informative active annotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 23" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "type": "text", + "content": "[87] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14471-14481, 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 254, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 285 + ], + "type": "text", + "content": "[88] Xiaojin Zhu. Semi-Supervised Learning Literature Survey. Technical Report 1530, Department of Computer Science, University of Wisconsin Madison., 2005. 1, 21" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 287, + 545, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 339 + ], + "type": "text", + "content": "[89] Xiaojin Zhu, Zoubin Ghahramani, and John D Lafferty. Semi-supervised learning using gaussian fields and harmonic functions. In Proceedings of the 20th International conference on Machine learning (ICML-03), pages 912–919, 2003. 21" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 342, + 547, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 547, + 385 + ], + "type": "text", + "content": "[90] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2020. 21" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22293" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/2ac9042c-6511-4205-8b71-c79e7f00a075_content_list.json b/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/2ac9042c-6511-4205-8b71-c79e7f00a075_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bcb791d17754fa763fdec50c4ac6a9134c7ca06a --- /dev/null +++ b/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/2ac9042c-6511-4205-8b71-c79e7f00a075_content_list.json @@ -0,0 +1,1546 @@ +[ + { + "type": "text", + "text": "T-VSL: Text-Guided Visual Sound Source Localization in Mixtures", + "text_level": 1, + "bbox": [ + 147, + 130, + 821, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tanvir Mahmud \nUniversity of Texas at Austin", + "bbox": [ + 96, + 181, + 330, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yapeng Tian \nUniversity of Texas at Dallas", + "bbox": [ + 367, + 181, + 596, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diana Marculescu \nUniversity of Texas at Austin", + "bbox": [ + 633, + 181, + 864, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 312, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual sound source localization poses a significant challenge in identifying the semantic region of each sounding source within a video. Existing self-supervised and weakly supervised source localization methods struggle to accurately distinguish the semantic regions of each sounding object, particularly in multi-source mixtures. These methods often rely on audio-visual correspondence as guidance, which can lead to substantial performance drops in complex multi-source localization scenarios. The lack of access to individual source sounds in multi-source mixtures during training exacerbates the difficulty of learning effective audio-visual correspondence for localization. To address this limitation, in this paper, we propose incorporating the text modality as an intermediate feature guide using tri-modal joint embedding models (e.g., AudioCLIP) to disentangle the semantic audio-visual source correspondence in multi-source mixtures. Our framework, dubbed T-VSL, begins by predicting the class of sounding entities in mixtures. Subsequently, the textual representation of each sounding source is employed as guidance to disentangle fine-grained audio-visual source correspondence from multi-source mixtures, leveraging the tri-modal AudioCLIP embedding. This approach enables our framework to handle a flexible number of sources and exhibits promising zero-shot transferability to unseen classes during test time. Extensive experiments conducted on the MUSIC, VGGSound, and VGGSound-Instruments datasets demonstrate significant performance improvements over state-of-the-art methods. Code is released at https://github.com/encyac-group/T-VSL/tree/main.", + "bbox": [ + 76, + 282, + 472, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 753, + 207, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While observing a conversation between two individuals, we can easily associate the audio signal with the corresponding speaking person in the visual scene. This remarkable ability to perceive audio-visual correspondence stems from our extensive exposure to both single-source sounds and multi-source mixtures in everyday life. Inspired by this human capability, significant research efforts [21, 30, 42, 43] have been dedicated to developing vi", + "bbox": [ + 75, + 779, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/20e68c70fc09ecb7301d28ed1abfe6bc1c6571d92f0c2adb001f53931c8dacef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 251, + 637, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0ef5b245a37750e3ce6a6007a28fcf790de95eb2e2866e5caa80e6fb2ee6df3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 251, + 758, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9a80375f90a17aadea0741648bba785c016c836c1b6789f658918fb35096ae4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 251, + 885, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/41ebede698abbc14ac2ff7d9a8bcf160b57613ddfd316b9ea0c3eee9723b8d56.jpg", + "image_caption": [ + "Figure 1. Comparison of our T-VSL with state-of-the-art methods on single-source (Top Row) and multi-source (Bottom Row) sound localization on VGGSound Sources [7], VGGSoundInstruments [21], and MUSIC [49] benchmark datasets. We use the same setup for the fair comparison." + ], + "image_footnote": [], + "bbox": [ + 501, + 327, + 637, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c8a46b3e47749300327f77158d28188652d7bfbfbe0ed56454aad3a5b525b11b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 327, + 758, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6bdd148357b304515ec2d9b67b73d2cc2bb78dd544c670cc6d6ce73fed5dcde7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 327, + 885, + 405 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sual sound source localization approaches in recent years.", + "bbox": [ + 500, + 488, + 880, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual sound source localization aims to locate the visual regions representing sound sources present in a video. Earlier methods on single-source localization [1, 7, 19, 28, 29, 40, 41] mostly used audio-visual correspondence as guidance for localizing sounding objects in each frame. These approaches, developed for single sound source localization, suffer from significant performance drop in localizing multi-source mixtures. Multi-source localization from mixtures is very challenging, as the model must learn to distinguish each sounding source within a mixture and establish their cross-modal relationships, without access to individual source sounds. Earlier methods [20, 21, 30, 36] on multi-source localization attempt to infer fine-grained cross-modal associations directly from noisy multi-source mixtures, often leading to sub-optimal performance in practice for learning improper source correspondence. In this paper, we tackle the problem by introducing a unified solution for localizing visual sound sources in both single and multi-source mixtures.", + "bbox": [ + 496, + 505, + 890, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The primary challenge of this task stems from the difficulty in disentangling the cross-modal correspondence of each sounding object from natural mixtures, given the absence of one-to-one corresponding single-source pairs. Additionally, the presence of silent visual objects and noise from invisible background sources further complicates this alignment process. To solve this problem, our key idea is to", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26742", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "leverage the text modality as a coarse supervision for disentangling categorical audio-visual correspondence in natural mixtures, which differentiates our method from existing works. Unlike audio and visual modalities that may contain significant noises from different sources present in mixtures, textual representation can explicitly discriminate across multiple sources. However, a critical obstacle to utilizing text guidance in visual sound source localization is grounding fine-grained audio and visual features with their textual representation. Later, CLIP [39] introduced learning visual language grounding from web-scraped 400M imagetext pairs. Recently, AudioCLIP [17] introduced the audio modality in the existing CLIP architecture, thereby generating tri-modal feature grounding through large-scale training. Our approach capitalizes on the tri-modal joint embedding space of AudioCLIP to disentangle one-to-one audio-visual correspondence in natural mixtures.", + "bbox": [ + 75, + 90, + 472, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, in this paper, we propose a novel text-guided multi-source localization framework that can discover fine-grained audio-visual semantic correspondence in multisource mixtures. Initially, we detect the class instances of visual sounding objects in the frame using the noisy mixture features from AudioCLIP image and audio encoders. Then, the text representation of each detected sounding source class instance is extracted with AudioCLIP text encoder, which serves as a coarse guidance for audio and visual feature separation. Specifically, we disentangle the categorical audio and visual features of each sounding source using the coarse text-guidance through additional audio and image conditioning blocks, respectively. Finally, the extracted categorical audio-visual semantic features are further aligned through an audio-visual correspondence block for localizing each sounding source. In comparison with prior multisource baselines, our approach can selectively localize the semantic visual region of each class of sounding objects in mixtures of varying number of sources. Extensive experiments on MUSIC [49], VGGSound [6], and VGGSound-Instruments [21] demonstrate significant performance improvements compared to existing single and multi-source baselines. Moreover, our method shows promising zero-shot transferability on unseen classes during test time. In addition, thorough ablation studies and qualitative analysis vividly showcase the effectiveness of the proposed framework.", + "bbox": [ + 75, + 349, + 472, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 94, + 762, + 457, + 776 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel text-guided multi-source localization framework to disentangle one-to-one audio-visual correspondence from natural mixtures.", + "- Our work is the first to introduce AudioCLIP for utilizing the tri-modal feature grounding from large-scale pretraining in solving multi-source localization problems.", + "- Our proposed framework can operate with a flexible number of sources and shows promising zero-shot perfor" + ], + "bbox": [ + 76, + 780, + 467, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mance on unseen audio-visual classes.", + "bbox": [ + 511, + 92, + 767, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive experiments on benchmark datasets clearly demonstrate the superiority of our proposed framework over other state-of-the-art methods.", + "bbox": [ + 500, + 106, + 890, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 162, + 640, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Sound Source Localization. Visual sound source localization aims to locate the visual regions of sounding objects in video frames. Prior work explored diverse methods for audio-visual feature alignment, including traditional statistical machine learning approaches [12, 18, 23] to latest deep neural nets [1, 7, 19, 28, 29, 36, 40, 41]. However, most of these self-supervised and weakly-supervised methods are mostly designed for localizing single-source sounds. A two-stream architecture is proposed in Attention10k [40] which incorporates cross-attention to locate sounding sources. Later, motion cues present in videos are incorporated for finer localization [1]. Recently, various contrastive learning based methods are explored in LVS [7], EZ-VSL [28], and SLAVC [29] for single-source localization. Despite their promising performance, these single-source localization methods can hardly discriminate audio-visual correspondence in multi-source mixtures.", + "bbox": [ + 496, + 188, + 890, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sounds mostly appear in mixtures in the presence of significant background noises. The main challenges of multisource localization is to disentangle the audio-visual correspondence in mixtures without having access to clean single-source audio-visual pairs. Several recent works have explored multi-source localization with explicit approaches for mixtures [20, 21, 30, 36]. To selectively isolate the silent objects in visual frames, DSOL [20] proposed a two-stage weakly-supervised training method. Mix-and-localize [21] proposed a self-supervised training with contrastive random walk to inherently align each sounding source with its visual representation. Recently, AVGN [30] introduced weakly-supervised grouping of category-aware features with learnable prompts. However, these approaches tackle the audio-visual correspondence in mixtures without any explicit guidance for single-source thereby resulting in sub-optimal performance. In contrast, we leverage text representation of fine-grained sounding sources to disentangle one-to-one audio-visual correspondence from multi-source mixtures. To the best of our knowledge, this work is the first to utilize text representation of sound sources for solving multisource localization.", + "bbox": [ + 496, + 446, + 890, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Audio, Visual, and Text Grounding. Large-scale pretraining on web-scraped data has been used to generate grounded features across modalities. CLIP [39] first introduced vision-language grounding using 400M imagedtext pairs, which has been extensively studied for zero-shot classification [2, 11, 14, 15, 22, 22, 44], text-video retrieval [3, 8, 25], open-vocabulary segmentation [9, 13, 26, 31, 37, 47, 48] and object detection [5, 10, 27, 45]. Re", + "bbox": [ + 496, + 779, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26743", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9bd966ff2e2af2de15ff2660debcb0f5b45a0b47f8d82317f11a42ea62a99279.jpg", + "image_caption": [ + "Figure 2. The proposed text-guided visual sound source localization (T-VSL) framework (for $K = 2$ ). We use the text modality to disentangle the fine-grained audio-visual correspondence from mixtures. Initially, we detect the audio-visual class instances from multisource mixtures using AudioCLIP joint embedding space. Later, categorical text features of each detected $K$ classes are used as coarse guidance in conditioning blocks to extract categorical visual and audio features. Afterwards, cross-modal feature alignment on extracted categorical features is performed in audio-visual correspondence block. Finally, cosine similarity of mean categorical audio features, and aligned visual features are used recursively to extract localization map of each class." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cently, AudioCLIP [17] and Wav2CLIP [46] introduced additional audio grounding in original CLIP model by learning on image, text, and audio triplets. Very recently, a CLIP-based single source localization method [34] is introduced that attempts replacing the text-conditioning branch of a supervised image segmentation baseline, trained with dense supervision, with audio conditioning. In relation with this concurrent work, our focus lies primarily on weakly supervised sound source localization in multi-source mixtures, leveraging the self-supervised AudioCLIP model.", + "bbox": [ + 75, + 468, + 472, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 633, + 169, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a mixture of audio and a video frame, our objective is to spatially localize each sounding object in the frame. We propose a novel text-guided multi-source localization framework to disentangle the fine-grained audio-visual correspondence, which is illustrated in Fig. 2. To suppress the background noises, we initially detect the common class instances in both audio and visual modality (Sec. 3.2). In particular, we leverage tri-modal AudioCLIP encoders to detect $K$ visual-sounding class instances from $N$ classes $(K \\leq N)$ . The primary challenge in multi-source localization lies in establishing one-to-one audio-visual correspondences for individual sources without access to single-source samples. To overcome this, we leverage the text modality to generate coarse representation of the $K$ detected sources in the mixture. Later, by exploiting the multimodal grounding of AudioCLIP, we extract categorical au", + "bbox": [ + 75, + 659, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dio and visual features from multi-source representation using text guidance (Sec. 3.3). Rather than aligning audio-visual mixture features as prior work, we introduce the categorical audio-visual feature alignment to further enhance one-to-one correspondence (Sec. 3.4). Finally, cosine similarity of categorical audio features and aligned visual features is used to extract semantic localization maps.", + "bbox": [ + 496, + 468, + 893, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 498, + 587, + 640, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we formulate the multi-source localization problem, and revisit the AudioCLIP [17] for audio, visual, and text grounding.", + "bbox": [ + 496, + 612, + 893, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation. Let's assume the dataset contains a total of $N$ classes of sounding events across all videos. Given a video frame and an audio mixture, our objective is to localize $K(K \\leq N)$ sound sources within the frame. For each audio-visual mixture pair, we can extract $Y = \\{y_{i}\\}_{i=1}^{N}$ binary ground truth labels, along with categorical text representations for the on-screen sound sources in the mixture. During training, we are limited to video-level sounding class entities, excluding bounding boxes or mask-level annotations, thus constituting weakly-supervised learning.", + "bbox": [ + 496, + 666, + 895, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Revisit AudioCLIP for multi-modal grounding. Audio-CLIP [17] learns a joint embedding space of audio, visual, and text triplets through contrastive learning. Let's consider the dataset $\\mathcal{D} = \\{a,v,t\\}_{i = 1}^{Z}$ containing audio spectrogram $a\\in \\mathbb{R}^{M\\times F}$ , text $t\\in \\mathbb{R}^l$ , and image $v\\in \\mathbb{R}^{C\\times H\\times W}$ triplets.", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26744", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Also, separate audio encoder $(E_{a}:\\mathbb{R}^{M\\times F}\\to \\mathbb{R}^{D})$ , visual encoder $(E_{v}:\\mathbb{R}^{C\\times H\\times W}\\to \\mathbb{R}^{D})$ , and text encoder $(E_{t}:\\mathbb{R}^{l}\\to \\mathbb{R}^{D})$ are used for each modality. Hence, the extracted audio, visual, and text features are represented by $A = E_{a}(a)$ , $V = E_{v}(v)$ , and $T = E_{t}(t)$ , respectively, where $A,V,T\\in \\mathbb{R}^D$ . Later, InfoNCE contrastive loss $(\\mathcal{L}_{\\mathrm{cnt}})$ [33] across each modality pair is used for feature grounding given by:", + "bbox": [ + 75, + 89, + 472, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a c l i p}} = \\mathcal {L} _ {\\mathrm {c n t}} (A, T) + \\mathcal {L} _ {\\mathrm {c n t}} (V, T) + \\mathcal {L} _ {\\mathrm {c n t}} (A, V) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 224, + 468, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This training objective, combined with large-scale pretraining on web-scraped data, establishes a grounded joint embedding space across all three modalities.", + "bbox": [ + 75, + 256, + 468, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Audio-Visual Class Instance Detection", + "text_level": 1, + "bbox": [ + 76, + 313, + 405, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The target sounding objects for localization must be present in both the audio mixture and the corresponding video frame. However, audio mixtures may contain noise from off-screen background sources, while video frames may include non-sounding objects. To address these challenges, we first detect the class instances of visible sounding objects in the frame. Then, we utilize this detection to suppress redundant features in the mixture, ensuring that the localization focuses on the relevant sounding objects.", + "bbox": [ + 75, + 338, + 468, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Token extraction from Uni-modal encoders. Multiple sounding objects can be spatially located in different regions of the reference frame, while the audio mixture may contain the temporally distributed signal of different sources. This makes it challenging to detect multiple class instances with uni-dimensional single-token features generated from AudioCLIP encoders. Instead of a single pooled token representation, we extract numerous patch tokens from audio and visual encoders with simple modifications, such that $\\widehat{E}_a: \\mathbb{R}^{M \\times F} \\to \\mathbb{R}^{n_a \\times D}$ and $\\widehat{E}_v: \\mathbb{R}^{C \\times H \\times W} \\to \\mathbb{R}^{n_v \\times D}$ . Here, $n_a = m \\times f$ and $n_v = h \\times w$ represent the number of audio and visual patch tokens, respectively. Afterwards, we apply linear projectors on these extracted patch tokens, such that $P_a: \\mathbb{R}^{n_a \\times D} \\to \\mathbb{R}^{n_a \\times D}$ and $P_v: \\mathbb{R}^{n_v \\times D} \\to \\mathbb{R}^{n_v \\times D}$ . Thus, extracted audio and visual patch tokens are represented as $A' = P_a(\\widehat{E}_a(a))$ , $V' = P_v(\\widehat{E}_v(v))$ , respectively, where $A' \\in \\mathbb{R}^{n_a \\times D}$ and $V' \\in \\mathbb{R}^{n_v \\times D}$ . Simultaneously, categorical text features $\\mathcal{T} = \\{e_i^t\\}_{i=1}^N \\in \\mathbb{R}^{N \\times D}$ representing all $N$ sound source classes in the dataset are extracted.", + "bbox": [ + 75, + 474, + 472, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-label class instance detection. With the categorical text features $\\mathcal{T}$ , mixture audio $A^{\\prime}$ , and visual $V^{\\prime}$ patch tokens, we detect class entities present in the mixture. Initially, we extract the mean-pooled audio-visual features $X^{\\prime} \\in \\mathbb{R}^{2D}$ by simple concatenation, and then, apply fusion projector $P_{f}: \\mathbb{R}^{2D} \\to \\mathbb{R}^{D}$ to extract mixture audio-visual features $F_{av} \\in \\mathbb{R}^{D}$ . Finally, cosine similarity between $N$ class text features $\\mathcal{T}$ and $F_{av}$ is used to detect mixture class", + "bbox": [ + 75, + 779, + 472, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "entities $\\widehat{Y}\\in \\mathbb{R}^N$ , given by:", + "bbox": [ + 498, + 88, + 681, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F _ {a v} = P _ {f} \\left(X ^ {\\prime}\\right), X ^ {\\prime} = \\left[ \\operatorname {M e a n} \\left(A ^ {\\prime}\\right) \\oplus \\operatorname {M e a n} \\left(V ^ {\\prime}\\right) \\right], \\\\ \\widehat {\\quad} \\end{array} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 114, + 893, + 141 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {Y} = \\operatorname {s i m} (\\mathcal {T}, F _ {a v})\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 137, + 653, + 155 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[\\oplus ]$ denotes the concatenation operator, and $\\operatorname {sim}(X,Y)$ denotes cosine similarity between $X$ and $Y$ . We formulate a multi-label classification objective $\\mathcal{L}_{\\mathrm{mcid}}$ by applying binary cross-entropy loss $\\mathcal{L}_{\\mathrm{bce}}(\\cdot)$ on each prediction using video-level ground-truth labels $Y\\in \\mathbb{R}^{N}$ as:", + "bbox": [ + 498, + 165, + 893, + 242 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {m c i d}} = \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathrm {b c e}} \\left(y _ {i}, \\widehat {y} _ {i}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 253, + 890, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In particular, we train audio, visual, and fusion projectors $(P_{a}, P_{v}, P_{f})$ on noisy multi-source data keeping the AudioCLIP backbones frozen. Hence, these projectors assist in extracting relevant foreground features from noisy audio and visual mixture features extracted from AudioCLIP.", + "bbox": [ + 498, + 299, + 890, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Audio and Visual Conditioning Blocks", + "text_level": 1, + "bbox": [ + 498, + 383, + 828, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Establishing one-to-one correspondences for each source in multi-source mixtures is particularly challenging, as both audio and visual features contain mixed representations of various sources. Prior works [21, 30] on multi-source localization have attempted to learn audio-visual alignment intrinsically without explicit single-source guidance. However, in the absence of single-source data, these methods struggle to adequately disentangle multi-source features. In contrast, we leverage the categorical text features of $K$ classes present in the mixture as coarse guidance for fine-grained feature disentanglement from multi-source mixtures. We use ground truth weak labels $Y$ of representative audios in the mixture to extract text embedding of $K$ visual sounding sources $\\mathcal{T}_K = \\{e_k^t\\}_{k=1}^K \\subseteq \\mathcal{T}$ where $e_k^t \\in \\mathbb{R}^{1 \\times D}$ .", + "bbox": [ + 496, + 407, + 890, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the categorical text features $\\mathcal{T}_K$ of $K$ sources, we independently disentangle the multi-object visual patch tokens $V^{\\prime}$ and mixture audio patch tokens $A^\\prime$ by using crossattention $\\phi^v (\\cdot)$ and $\\phi^a (\\cdot)$ , respectively, as:", + "bbox": [ + 496, + 619, + 890, + 679 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\begin{array}{c} \\widetilde {V} _ {k} = \\phi^ {v} \\left(V ^ {\\prime}, e _ {k} ^ {t}, V ^ {\\prime}\\right), f _ {k} ^ {v} = \\operatorname {M e a n} \\left(\\widetilde {V} _ {k}\\right), \\\\ \\sim \\end{array} \\tag {4} \\\\ \\tilde {f} _ {k} ^ {v} = P _ {v c} \\left(f _ {k} ^ {v}\\right), \\forall k = \\{1, \\dots , K \\} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 688, + 890, + 729 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\underset {\\sim} {\\widetilde {A}} _ {k} = \\phi^ {a} \\left(A ^ {\\prime}, e _ {k} ^ {t}, A ^ {\\prime}\\right), f _ {k} ^ {a} = \\operatorname {M e a n} \\left(\\widetilde {A} _ {k}\\right), (5) \\\\ \\hat {f} _ {k} ^ {a} = P _ {a c} \\left(f _ {k} ^ {a}\\right), \\forall k = \\{1, \\dots , K \\} \\\\ \\phi (A, B, C) = \\left(\\frac {A B ^ {\\top}}{\\| A \\| \\| B \\|}\\right) \\odot C (6) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 741, + 890, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\widetilde{f}_k^a,\\widetilde{f}_k^v,f^a,f^v\\in \\mathbb{R}^{1\\times D},\\widetilde{V}_k\\in \\mathbb{R}^{n_v\\times D},\\widetilde{A}_k\\in$ $\\mathbb{R}^{n_a\\times D}$ , and $P_{ac},P_{vc}:\\mathbb{R}^D\\to \\mathbb{R}^D$ are linear audio and visual projectors on conditioned mean features, respectively, and $\\odot$ is the Hadamard product.", + "bbox": [ + 496, + 837, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26745", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/181c6437b7e4bf9ea3f05518cff3e16243eb88a914805711e57f7607ff2f9841.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVGGSound-SingleVGGSound-InstrumentsMUSIC-Solo
AP(%)IoU@0.5(%)AUC(%)AP(%)IoU@0.3(%)AUC(%)AP(%)IoU@0.5(%)AUC(%)
Current SOTA Based Methods
OTS [4] ECCV1838.943.243.750.956.938.476.257.848.9
CoarsetoFine [36] ECCV2039.443.144.851.157.539.176.358.149.7
LVS [7] CVPR2139.744.645.351.557.939.476.157.149.3
EZ-VSL [28] ECCV2240.645.147.252.358.839.978.458.751.1
Mix-and-Localize [21] CVPR2241.845.647.453.859.341.578.959.952.7
DSOL [20] NeurIPS20-46.847.9-61.643.7-62.754.6
MarginNCE [35] ICASSP2342.546.948.356.961.844.283.163.355.2
FNAC [43] CVPR2343.348.449.157.263.245.384.664.556.4
AVGN [30] CVPR2344.149.649.559.364.746.185.465.856.9
Alignment [42] ICCV2345.350.850.260.465.848.786.166.457.2
CLIP-Based Baseline Methods
AudioCLIP [17] ICASSP2242.847.448.558.362.745.283.863.155.7
Wav2CLIP [46] ICASSP2239.343.644.753.857.242.078.959.951.2
T-VSL (Ours)48.153.752.964.669.551.488.268.560.1
", + "bbox": [ + 99, + 88, + 877, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Performance comparison of single-source localization on VGGSound-single, VGGSound-Instruments, and MUSIC-Solo datasets. For the fair comparison, we use pre-trained AudioCLIP audio and image encoders for baseline methods.", + "bbox": [ + 75, + 328, + 890, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, to guide extracting categorical audio $(\\widetilde{A}_k)$ and visual $(\\widetilde{V}_k)$ patch tokens, we introduce coarse supervision on the output of projected mean patch tokens $\\widetilde{f}_k^a,\\widetilde{f}_k^v$ using $N$ -class text embedding $\\mathcal{T}\\in \\mathbb{R}^{N\\times D}$ . Hence, the class conditioning loss $(\\mathcal{L}_{\\mathrm{cls}})$ is given by:", + "bbox": [ + 75, + 369, + 472, + 449 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {e} _ {k} ^ {v} = \\operatorname {s i m} \\left(\\mathcal {T}, \\widetilde {f _ {k} ^ {v}}\\right); \\mathbf {e} _ {k} ^ {a} = \\operatorname {s i m} \\left(\\mathcal {T}, \\widetilde {f _ {k} ^ {a}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 458, + 385, + 477 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c l s}} = \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathbf {e} _ {k} ^ {v}, \\mathbf {h} _ {k}\\right) + \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathbf {e} _ {k} ^ {a}, \\mathbf {h} _ {k}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 479, + 468, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{h}_k\\in \\mathbb{R}^N$ denotes an one-hot encoding of the class label, and $\\mathbf{e}_k^v\\in \\mathbb{R}^N$ , $\\mathbf{e}_k^a\\in \\mathbb{R}^N$ are visual and audio class predictions for each $k^{th}$ source, respectively.", + "bbox": [ + 76, + 531, + 468, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Audio-Visual Correspondence Block", + "text_level": 1, + "bbox": [ + 76, + 585, + 393, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While text representation provides coarse guidance for disentangling fine-grained spatio-temporal audio and visual features, it has inherent limitations in practice. For example, multiple instances of a sounding object class can appear in the visual frame, including silent instances, which are difficult to distinguish using simple text class representation. To address this challenge, we introduce an audio-visual correspondence block to further refine the alignment of audio and visual features based on extracted categorical features.", + "bbox": [ + 75, + 609, + 468, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the disentangled categorical audio $(\\widetilde{A}_k)$ and visual $(\\widetilde{V}_k)$ patch tokens, we apply cross-attention $\\phi^{av}(\\cdot)$ to enhance audio-visual patch alignment as:", + "bbox": [ + 75, + 744, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng _ {k} ^ {a} = \\operatorname {M e a n} \\left(\\widetilde {A} _ {k}\\right), \\widehat {g} _ {k} ^ {a} = P _ {a v} \\left(g _ {k} ^ {a}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 799, + 354, + 819 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {V} _ {k} = \\phi^ {a v} \\left(\\widetilde {V} _ {k}, \\widehat {g} _ {k} ^ {a}, \\widetilde {V} _ {k}\\right), g _ {k} ^ {v} = \\operatorname {M e a n} \\left(\\widehat {V} _ {k}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 821, + 468, + 840 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {g} _ {k} ^ {v} = P _ {v a} \\left(g _ {k} ^ {v}\\right), \\forall k = \\{1, \\dots , K \\}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 842, + 364, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $P_{av}, P_{va}: \\mathbb{R}^D \\to \\mathbb{R}^D$ are linear projectors. Afterwards, to correspond between projected mean categorical", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "audio and visual features $\\widehat{g}_k^a,\\widehat{g}_k^v\\in \\mathbb{R}^D$ , we apply audiovisual contrastive loss $(\\mathcal{L}_{\\mathrm{av}})$ given by:", + "bbox": [ + 498, + 369, + 890, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a v}} = \\mathcal {L} _ {\\mathrm {c n t}} \\left(\\widehat {g} _ {k} ^ {v}, \\widehat {g} _ {k} ^ {a}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 409, + 890, + 433 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {a v}} + \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {m c i d}}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 430, + 799, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{total}}$ is the total accumulated loss.", + "bbox": [ + 500, + 452, + 782, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference: During inference, we estimate cosine similarity across projected mean audio feature $\\widehat{g}_k^a\\in \\mathbb{R}^D$ and aligned visual patch tokens $(\\widehat{V}_k = \\{\\widehat{v}_k^j\\}_{j = 1}^{n_v},\\widehat{v}_k^j\\in \\mathbb{R}^D)$ . Finally, after reshaping and up-scaling through bilinear interpolation, we generate $K$ -source heat-maps $\\mathbf{H}\\in \\mathbb{R}^{K\\times H\\times W}$ .", + "bbox": [ + 498, + 474, + 890, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results and Discussions", + "text_level": 1, + "bbox": [ + 498, + 556, + 723, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 500, + 575, + 687, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We used MUSIC, VGGSound-Instruments, and VGGSound datasets for the performance evaluation1. MUSIC [49] dataset contains 445 solo music videos of 11 instruments and 142 duet music videos of 8 instruments from YouTube. Following prior works [30], we use MUSIC-Solo for single-source and MUSIC-Duet for multi-source localization. From MUSIC-Solo, we use 350 videos for training and remaining 95 for evaluation. For MUSIC-Duet, we use 120 videos for training and remaining 22 videos for evaluation. VGGSound-Instruments [21] is a subset of VGGSound dataset [6] containing around 32k video clips from 37 music instruments of 10s duration. Apart from the musical instrument datasets, we gather around 150k videos of 10s duration from original VGGSound dataset [6] representing single-source sounds of 221 categories, such as animals, people, nature, instruments, etc. For the single-source evaluation on VGGSound, we use the full test", + "bbox": [ + 496, + 599, + 890, + 856 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1Since many videos are no longer available for public use, these datasets become smaller. We use the same data split for all baselines and reproduce them under the same setting for a fair comparison.", + "bbox": [ + 498, + 862, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26746", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ac0a22cafc4a11ab7c789b44b93195df52f122aa69f2c259e0f41d7f685257c6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVGGSound-DuetVGGSound-InstrumentsMUSIC-Duet
CAP(%)CIoU@0.3(%)AUC(%)CAP(%)CIoU@0.1(%)AUC(%)CAP(%)CIoU@0.3(%)AUC(%)
Current SOTA Single-Source Methods
OTS [4] ECCV1823.926.427.528.777.818.251.229.424.3
CoarseToFine [36] ECCV20-28.427.9-78.819.7-31.824.1
LVS [7] CVPR21-31.830.1-81.123.2-33.126.9
EZ-VSL [28] ECCV22-32.430.6-80.322.9-34.326.7
MarginNCE [35] ICASSP2327.133.231.233.582.324.755.136.128.5
FNAC [43] CVPR2329.835.133.435.484.126.857.337.930.6
Alignment [42] ICCV2330.435.833.935.684.927.157.838.330.9
Current SOTA Multi-Source Methods
Mix-and-Localize [21] CVPR2229.535.634.136.384.526.757.438.130.7
DSOL [20] NeurIPS20-36.934.6-85.928.4-38.831.3
AVGN [30] CVPR2331.937.835.438.186.728.859.239.332.4
CLIP-Based Baseline Methods
AudioCLIP [17] ICASSP2228.434.932.834.783.825.956.136.929.2
Wav2CLIP [46] ICASSP2226.331.428.931.280.322.252.632.227.2
T-VSL (ours)35.740.137.941.889.631.562.943.235.9
", + "bbox": [ + 88, + 88, + 883, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance comparison of multi-source localization on VGGSound-Duet, VGGSound-Instruments, and MUSIC-Duet datasets. For the fair comparison, we use pre-trained AudioCLIP audio and image encoders for baseline methods.", + "bbox": [ + 76, + 333, + 890, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "set of 5158 videos, while on VGGSound-Instruments, we use 446 videos following prior work [21, 30]. For the multi-source training and evaluation in both VGGSound and VGGSound-Instruments, we randomly concatenate two frames from different sounding sources to produce the multi-source input image with a size of $448 \\times 224$ , and mix the single-source audios, following prior work [21, 30].", + "bbox": [ + 75, + 375, + 468, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For evaluating single-source localization performance, we use Intersection over Union (IoU), average precision (AP), and Area Under Curve (AUC) following prior work [21, 30]. For the multi-source localization evaluation, we use Class-aware Average precision (CAP), Class-aware IoU (CIoU), and Area Under Curve (AUC) following prior work [21, 30]. Similarly, for fair comparison with the prior work [30], we use the same thresholds in these metrics: we use IoU@0.5 and CIoU@0.3 for the MUSIC-Solo and MUSIC-Duet datasets, IoU@0.3 and IoU@0.5 for single-source VGGSound-Instruments and VGGSound datasets, and CIoU@0.1 and CIoU@0.3 for multi-source VGGSound-Instruments and VGGSound.", + "bbox": [ + 75, + 484, + 468, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We use the AudioCLIP [17] pretrained encoders containing ResNet-50 [19] image encoder, ESResNeXt [16] audio encoder, and transformer [38] based text encoder. For the fair comparison, we reproduce the result of existing methods with AudioCLIP image and audio encoders, instead of using separately pre-trained encoders. We use $224 \\times 224$ resolution for the single-source input image, $D = 1024$ , $n_v = 49$ for $7 \\times 7(h \\times w)$ spatial maps generated from the visual encoder, and $n_a = 60$ for $10 \\times 6(m \\times f)$ time-frequency map generated from the audio encoder. We use Adam optimizer [24] with a batch size of 256 and with a learning rate of $1e - 4$ .", + "bbox": [ + 75, + 685, + 468, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CLIP-Based Baseline methods. We adopt several CLIP-based baseline methods for the sound source localization", + "bbox": [ + 76, + 869, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "task, which are described as follows:", + "bbox": [ + 498, + 375, + 743, + 388 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- AudioCLIP [17]: We fine-tune pre-trained image and audio encoders of AudioCLIP on the target dataset. We estimate the localization heatmap for the sounding sources using the similar cosine similarity across patch tokens.", + "- Wav2CLIP [46]: Similar to AudioCLIP baseline, we experiment with Wav2CLIP audio and image encoders to extract localization heatmaps of sounding sources." + ], + "bbox": [ + 500, + 392, + 890, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison to SOTA Baselines", + "text_level": 1, + "bbox": [ + 500, + 511, + 774, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Single-source localization: We present the quantitative results on single-source localization in Tab. 1. We apply the AudioCLIP image and audio encoders in all baseline methods for the fair comparison with our methods. We observe some performance improvements with several SOTA methods [30, 42, 43] over the AudioCLIP baseline. However, we also notice some reduction in performance from the AudioCLIP baseline in several methods [4, 21, 28]. We hypothesize that these can be due to the alterations of AudioCLIP grounding with the modified contrastive loss. In contrast, by utilizing all three modalities for sound source localization, the proposed T-VSL achieves the best performance over all CLIP-based baselines as well as SOTA methods. T-VSL outperforms the current SOTA single source localization method Alignment [42] by $+2.9$ , $+3.7$ , $+2.1$ higher IoUs on VGGSound-Single, VGGSound-Instruments, and MUSIC-solo datasets, respectively. Moreover, T-VSL achieves $+6.3$ , $+6.8$ , and $+5.4$ higher IoUs over the AudioCLIP baseline [32]. We hypothesize that the enhanced refinement of noisy audio and visual features with text guidance in T-VSL effectively contributes to performance improvement.", + "bbox": [ + 496, + 536, + 890, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-source localization: We present performance comparison in multi-source sound localization in Tab. 2. As be", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26747", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2e149157b3f9d2717826ed51b1b01f0022c525cc0a32bfcb2cd9c2f5ca105fd6.jpg", + "image_caption": [ + "Figure 3. We present qualitative comparisons on challenging multi-source localization with SOTA single and multi-source baseline methods. Here, blue color represents high-attention values to the sounding object, and red color represents low-attention values. The proposed T-VSL can selectively isolate the sounding regions from the background and generates more precise localization maps for sounding sources." + ], + "image_footnote": [], + "bbox": [ + 122, + 88, + 848, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fore, we apply the AudioCLIP image and audio encoders in all baseline methods for the fair comparison with our methods. We observe some performance improvements, mostly in SOTA multi-source methods [20, 21, 30], over the AudioCLIP baseline. In most single source methods, we observe performance loss in multi-source cases over the baseline. The proposed T-VSL achieves the best performance that outperforms the SOTA multi-source baseline AVGN [30] by $+2.3$ , $+2.9$ , and $+3.9$ CIoUs on multi-source VGGSound-Duet, VGGSound-Instruments, and MUSIC-Duet datasets, respectively. Also, T-VSL achieves $+5.2$ , $+5.8$ , and $+6.3$ higher CIoUs over the AudioCLIP baseline. These significant performance gains on multi-source benchmarks demonstrate the effectiveness of the proposed T-VSL in disentangling multi-source mixtures.", + "bbox": [ + 75, + 455, + 472, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative comparisons: In addition, we present qualitative comparisons on multi-source localization in Fig. 3 among single-source baseline Alignment [42], self-supervised multi-source baseline Mix-and-Localize [21], weakly-supervised multi-source baseline AVGN [30], and proposed T-VSL. Single-source SOTA baseline Alignment [42] struggles in multi-source localization due to the lack of audio-visual correspondence, mostly in the presence of noisy audios and silent visual objects. However, without any guidance on fine-grained source localization, multi-source baselines [21, 30] underperform on localizing challenging multi-source mixtures. Notably, our method generates superior localization maps, which shows the effectiveness of T-VSL in multi-source feature disentanglement.", + "bbox": [ + 75, + 691, + 468, + 905 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation on T-VSL building blocks", + "text_level": 1, + "bbox": [ + 498, + 455, + 803, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present the ablation study of different building blocks of T-VSL in Tab. 3 under both single-source and multisource settings. The vanilla baseline contains AudioCLIP image and audio encoders that directly operate on the input data. By integrating the audio-visual correspondence (AVC) block on extracted mixture patch tokens, we notice considerable performance increase in single-source (by $+1.1$ AP and $+0.7$ IoU@0.5) and multi-source localization (by $+1.3$ CAP and $+1.4$ CIoU@0.3). By integrating the audio and image conditioning blocks following audio-visual instance detection for separating fine-grained class features, we observe additional improvements of $+1.8$ AP and $+2.3$ AP in single-source, and $+3.1$ CAP and $+3.6$ AP in multisource localization, respectively. Finally, the combination of all building blocks results in the best performance that improves the baseline by $+5.3$ AP, and 6.3 IoU@0.5 in single-source and by $+7.3$ CAP and $+5.2$ CIoU@0.3 in multi-source localization. This further demonstrates the effective disentanglement of audio-visual features facilitated by intermediate text guidance.", + "bbox": [ + 496, + 473, + 890, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Zero-shot transfer across datasets", + "text_level": 1, + "bbox": [ + 500, + 777, + 795, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We can perform zero-shot transfer across datasets with proposed T-VSL by simply replacing the $N$ -class text prompts in Fig. 2. Since other weakly-supervised methods cannot perform zero-shot transfer, we present comparisons with self-supervised Mix-and-Localize [21], Alignment [42], and FNAC [43] methods in Tab. 4. To ensure a fair comparison, we conducted uniform training with $50\\% - 50\\%$ class", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26748", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/47df21d6af5219e16d098f94171889e4cc6d7093b0fabd0b7e44764d8360cff3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Audio Cond.Image Cond.AVCVGGSound-SingleVGGSound-Duet
AP(%)IoU@0.5(%)CAP(%)CIoU@0.3(%)
XXX42.847.428.434.9
XX✓43.948.129.736.3
✓X✓45.750.932.838.1
X✓✓46.251.433.338.6
✓✓✓48.153.735.740.1
", + "bbox": [ + 86, + 88, + 464, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/eb7aada93367fb16955008f3ca0606de3bfd9fce8d1f58dc8e291b7da6ef7249.jpg", + "table_caption": [ + "Table 3. Ablation study on audio conditioning block, image-conditioning block, and audio-visual correspondence block (AVC) of T-VSL in single-source and multi-source localization tasks." + ], + "table_footnote": [ + "Table 4. Ablation on zero-shot transfer across single and multi-source datasets with $50\\% -50\\%$ class split. We report IoU@0.5(%) for MUSIC-Solo, CIoU@0.3(%) for MUSIC-Duet, and CIoU@0.1(%) for VGGSound-Instruments. We use same AudioCLIP encoders for all baselines." + ], + "table_body": "
Target DatasetProposed T-VSLMix-and-Localize [21]Alignment [42]FNAC [43]
MUSIC-Duet55.850.951.751.2
MUSIC-Solo80.575.177.677.4
VGGSound-Instr.81.776.876.576.1
", + "bbox": [ + 83, + 247, + 467, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "split for all methods with same AudioCLIP encoders. We note that our method achieves significantly better performance on zero-shot transfer to target datasets than these compared approaches. In particular, we achieve $+4.9$ , $+5.4$ , $+4.9$ improvements on target datasets than the multi-source Mix-and-Localize [21]. Also, we achieve $+4.1$ , $+2.9$ , and $+5.2$ higher scores than the SOTA single source baseline, Alignment [42]. This result suggests that our T-VSL has superior generalization capabilities for zero-shot localization tasks in unseen classes.", + "bbox": [ + 75, + 417, + 468, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Robustness to a higher number of sources", + "text_level": 1, + "bbox": [ + 76, + 573, + 433, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed T-VSL is not limited by the number of sources present in training mixtures unlike prior work [21]. We present comparative analysis with SOTA multi-source AVGN [30], and self-supervised single source methods FNAC [43] and Alignment [42] in Tab. 5 on robustness to higher number of test sources in VGGSound dataset. We train all methods on the VGGSound-Single dataset with same encoders for a fair comparison. We note that our method consistently maintains significantly higher CIOU score irrespective of test scenarios, and also achieves notably smaller relative performance drop for a large number of sources in mixtures. These results further demonstrate the robustness of the proposed T-VSL in disentangling multi-source mixtures.", + "bbox": [ + 75, + 590, + 468, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Use of learnable text prompts", + "text_level": 1, + "bbox": [ + 76, + 806, + 339, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To disentangle multi-source mixtures, we primarily use the class label of each sounding source as text prompts (e.g. dog barking, snake hissing). Inspired by the conditional text prompt learning in [50], we study the use of learnable prompts along with class label representations. Instead of", + "bbox": [ + 75, + 825, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e85c6252a1322cbfc37a4d9e7665240c8f8b92471d8bbe779529c9135cb5da19.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Test Source No.Proposed T-VSLAVGN [30]Alignment [42]FNAC [43]
239.636.835.434.1
335.730.828.927.1
429.522.419.318.8
", + "bbox": [ + 519, + 88, + 874, + 178 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/14fdd797b16167749f5137e8da0a12307e32b1474d99bf651e9a1392ae871ea1.jpg", + "table_caption": [ + "Table 5. Robustness to higher number of sources in VGGSound dataset when trained with single-source data. CIoU@0.3(%) score is reported. Same AudioCLIP encoders are used for all baselines." + ], + "table_footnote": [], + "table_body": "
Prompt LengthVGGSound-SingleVGGSound-Duet
AP(%)IoU@0.5(%)CAP(%)CIoU@0.3(%)
048.153.735.740.1
248.554.136.040.4
449.254.636.841.3
850.355.237.142.5
1650.255.037.442.9
3249.754.537.142.7
", + "bbox": [ + 511, + 232, + 879, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study on the effect of learnable text prompts on VGGSound-Single and VGGSound-Duet datasets. Integration of learnable prompts results in noticeable performance improvements in both single and multi-source localization.", + "bbox": [ + 498, + 373, + 890, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "only using class label representation, a set of learnable embedding is integrated with class label embedding that provides additional flexibility to adapt the text prompt to target objective. We analyze the sound source localization performance with different length of learnable prompts. The performances are reported in Table 6. We notice considerable performance improvements by optimizing learnable prompts in both single and multi-source cases. However, such prompts limit zero-shot use cases to unseen classes.", + "bbox": [ + 496, + 436, + 890, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 579, + 617, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose T-VSL, a novel text-guided multi-source visual sound source localization framework that can disentangle one-to-one audio-visual correspondence from multi-source mixtures. We leverage the text modality to guide fine-grained feature separation and localization, from noisy audio and visual features of mixtures, by exploiting the joint embedding space of AudioCLIP. In comparison with SOTA multi-source baselines, our method shows superior zero-shot transfer across datasets. Moreover, our method demonstrates notable robustness on challenging test mixtures with higher number of sources than training scenarios. In both single-source and multi-source localization, our method largely outperforms existing weakly-supervised and self-supervised baselines on three benchmark datasets.", + "bbox": [ + 496, + 598, + 890, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 818, + 653, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was supported in part by ONR Minerva program, iMAGiNE - the Intelligent Machine Engineering Consortium at UT Austin, and a UT Cockrell School of Engineering Doctoral Fellowship.", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26749", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In Proceedings of European Conference on Computer Vision (ECCV), pages 208–224, 2020. 1, 2", + "[2] Muhammad Ali and Salman Khan. Clip-decoder: Zeroshot multilabel classification using multimodal clip aligned representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4675-4679, 2023. 2", + "[3] Tayfun Alpay, Sven Magg, Philipp Broze, and Daniel Speck. Multimodal video retrieval with clip: a user study. Information Retrieval Journal, 26(1-2):6, 2023. 2", + "[4] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proceedings of the European Conference on Computer Vision (ECCV), pages 435-451, 2018. 5, 6", + "[5] Maria A Bravo, Sudhanshu Mittal, and Thomas Brox. Localized vision-language matching for open-vocabulary object detection. In DAGM German Conference on Pattern Recognition, pages 393-408. Springer, 2022. 2", + "[6] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 2, 5", + "[7] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16867-16876, 2021. 1, 2, 5, 6", + "[8] Chaorui Deng, Qi Chen, Pengda Qin, Da Chen, and Qi Wu. Prompt switch: Efficient clip adaptation for text-video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15648-15658, 2023. 2", + "[9] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary panoptic segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022. 2", + "[10] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 2", + "[11] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model clip. In Proceedings of the AAAI conference on artificial intelligence, pages 6568-6576, 2022. 2", + "[12] John W Fisher III, Trevor Darrell, William Freeman, and Paul Viola. Learning joint statistical models for audio-visual fusion and segregation. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2000. 2", + "[13] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2", + "[14] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models. In Proceed-" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19338-19347, 2023. 2", + "[15] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xupeng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 746-754, 2023. 2", + "[16] Andrey Guzhov, Federico Raue, Jörn Hees, and Andreas Dengel. Esresne (x) t-fbsp: Learning robust time-frequency transformation of audio. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021. 6", + "[17] Andrey Guzhov, Federico Raue, Jorn Hees, and Andreas Dengel. Audioclip: Extending clip to image, text and audio. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 976-980. IEEE, 2022. 2, 3, 5, 6, 1", + "[18] John Hershey and Javier Movellan. Audio vision: Using audio-visual synchrony to locate sounds. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 1999. 2", + "[19] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9248-9257, 2019. 1, 2, 6", + "[20] Di Hu, Rui Qian, Minyue Jiang, Xiao Tan, Shilei Wen, Errui Ding, Weiyao Lin, and Dejing Dou. Discriminative sounding objects localization via self-supervised audiovisual matching. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 10077-10087, 2020. 1, 2, 5, 6, 7", + "[21] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10483-10492, 2022. 1, 2, 4, 5, 6, 7, 8", + "[22] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 2", + "[23] Einat Kidron, Yoav Y Schechner, and Michael Elad. Pixels that sound. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2005. 2", + "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[25] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2", + "[26] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In International Conference on Machine Learning, pages 23033-23044. PMLR, 2023. 2", + "[27] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26750", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In European Conference on Computer Vision, pages 728-755. Springer, 2022. 2", + "[28] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In Proceedings of European Conference on Computer Vision (ECCV), page 218-234, 2022. 1, 2, 5, 6", + "[29] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2", + "[30] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10565–10574, 2023. 1, 2, 4, 5, 6, 7, 8", + "[31] Jishnu Mukhoti, Tsung-Yu Lin, Omid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19413-19423, 2023. 2", + "[32] Jishnu Mukhoti, Tsung-Yu Lin, Amid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19413-19423, 2023. 6", + "[33] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4", + "[34] Sooyoung Park, Arda Senocak, and Joon Son Chung. Can clip help sound source localization? arXiv preprint arXiv:2311.04066, 2023. 3, 1", + "[35] Sooyoung Park, Arda Senocak, and Joon Son Chung. Marginnce: Robust sound localization with a negative margin. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 5, 6", + "[36] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In Proceedings of European Conference on Computer Vision (ECCV), pages 292-308, 2020. 1, 2, 5, 6", + "[37] Jie Qin, Jie Wu, Pengxiang Yan, Ming Li, Ren Yuxi, Xuefeng Xiao, Yitong Wang, Rui Wang, Shilei Wen, Xin Pan, et al. Freeseg: Unified, universal and open-vocabulary image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19446-19455, 2023. 2", + "[38] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 6", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4358-4366, 2018. 1, 2", + "[41] Arda Senocak, Hyeonggon Ryu, Junsik Kim, and In So Kweon. Learning sound localization better from semantically similar samples. In Proceedings of IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022. 1, 2", + "[42] Arda Senocak, Hyeonggon Ryu, Junsik Kim, Tae-Hyun Oh, Hanspeter Pfister, and Joon Son Chung. Sound source localization is all about cross-modal alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7777-7787, 2023. 1, 5, 6, 7, 8", + "[43] Weixuan Sun, Jiayi Zhang, Jianyuan Wang, Zheyuan Liu, Yiran Zhong, Tianpeng Feng, Yandong Guo, Yanhao Zhang, and Nick Barnes. Learning audio-visual source localization via false negative aware contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6420–6429, 2023. 1, 5, 6, 7, 8", + "[44] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1802-1812, 2023. 2", + "[45] Tao Wang. Learning to detect and segment for open vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7051-7060, 2023. 2", + "[46] Ho-Hsiang Wu, Prem Seetharaman, Kundan Kumar, and Juan Pablo Bello. Wav2clip: Learning robust audio representations from clip. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4563-4567. IEEE, 2022. 3, 5, 6", + "[47] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2935-2944, 2023. 2", + "[48] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 2", + "[49] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In Proceedings of the European Conference on Computer Vision (ECCV), pages 570-586, 2018. 1, 2, 5", + "[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 8" + ], + "bbox": [ + 501, + 92, + 892, + 837 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26751", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/2ac9042c-6511-4205-8b71-c79e7f00a075_model.json b/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/2ac9042c-6511-4205-8b71-c79e7f00a075_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6807bb81d5d30234917833d3ad5efa3d050892ed --- /dev/null +++ b/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/2ac9042c-6511-4205-8b71-c79e7f00a075_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.148, + 0.131, + 0.823, + 0.151 + ], + "angle": 0, + "content": "T-VSL: Text-Guided Visual Sound Source Localization in Mixtures" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.182, + 0.331, + 0.216 + ], + "angle": 0, + "content": "Tanvir Mahmud \nUniversity of Texas at Austin" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.182, + 0.597, + 0.216 + ], + "angle": 0, + "content": "Yapeng Tian \nUniversity of Texas at Dallas" + }, + { + "type": "text", + "bbox": [ + 0.635, + 0.182, + 0.866, + 0.216 + ], + "angle": 0, + "content": "Diana Marculescu \nUniversity of Texas at Austin" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.313, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.284, + 0.473, + 0.737 + ], + "angle": 0, + "content": "Visual sound source localization poses a significant challenge in identifying the semantic region of each sounding source within a video. Existing self-supervised and weakly supervised source localization methods struggle to accurately distinguish the semantic regions of each sounding object, particularly in multi-source mixtures. These methods often rely on audio-visual correspondence as guidance, which can lead to substantial performance drops in complex multi-source localization scenarios. The lack of access to individual source sounds in multi-source mixtures during training exacerbates the difficulty of learning effective audio-visual correspondence for localization. To address this limitation, in this paper, we propose incorporating the text modality as an intermediate feature guide using tri-modal joint embedding models (e.g., AudioCLIP) to disentangle the semantic audio-visual source correspondence in multi-source mixtures. Our framework, dubbed T-VSL, begins by predicting the class of sounding entities in mixtures. Subsequently, the textual representation of each sounding source is employed as guidance to disentangle fine-grained audio-visual source correspondence from multi-source mixtures, leveraging the tri-modal AudioCLIP embedding. This approach enables our framework to handle a flexible number of sources and exhibits promising zero-shot transferability to unseen classes during test time. Extensive experiments conducted on the MUSIC, VGGSound, and VGGSound-Instruments datasets demonstrate significant performance improvements over state-of-the-art methods. Code is released at https://github.com/encyac-group/T-VSL/tree/main." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.208, + 0.769 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.902 + ], + "angle": 0, + "content": "While observing a conversation between two individuals, we can easily associate the audio signal with the corresponding speaking person in the visual scene. This remarkable ability to perceive audio-visual correspondence stems from our extensive exposure to both single-source sounds and multi-source mixtures in everyday life. Inspired by this human capability, significant research efforts [21, 30, 42, 43] have been dedicated to developing vi" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.252, + 0.638, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.252, + 0.759, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.252, + 0.886, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.328, + 0.638, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.328, + 0.759, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.328, + 0.886, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.413, + 0.892, + 0.482 + ], + "angle": 0, + "content": "Figure 1. Comparison of our T-VSL with state-of-the-art methods on single-source (Top Row) and multi-source (Bottom Row) sound localization on VGGSound Sources [7], VGGSoundInstruments [21], and MUSIC [49] benchmark datasets. We use the same setup for the fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.489, + 0.882, + 0.504 + ], + "angle": 0, + "content": "sual sound source localization approaches in recent years." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Visual sound source localization aims to locate the visual regions representing sound sources present in a video. Earlier methods on single-source localization [1, 7, 19, 28, 29, 40, 41] mostly used audio-visual correspondence as guidance for localizing sounding objects in each frame. These approaches, developed for single sound source localization, suffer from significant performance drop in localizing multi-source mixtures. Multi-source localization from mixtures is very challenging, as the model must learn to distinguish each sounding source within a mixture and establish their cross-modal relationships, without access to individual source sounds. Earlier methods [20, 21, 30, 36] on multi-source localization attempt to infer fine-grained cross-modal associations directly from noisy multi-source mixtures, often leading to sub-optimal performance in practice for learning improper source correspondence. In this paper, we tackle the problem by introducing a unified solution for localizing visual sound sources in both single and multi-source mixtures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The primary challenge of this task stems from the difficulty in disentangling the cross-modal correspondence of each sounding object from natural mixtures, given the absence of one-to-one corresponding single-source pairs. Additionally, the presence of silent visual objects and noise from invisible background sources further complicates this alignment process. To solve this problem, our key idea is to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26742" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.348 + ], + "angle": 0, + "content": "leverage the text modality as a coarse supervision for disentangling categorical audio-visual correspondence in natural mixtures, which differentiates our method from existing works. Unlike audio and visual modalities that may contain significant noises from different sources present in mixtures, textual representation can explicitly discriminate across multiple sources. However, a critical obstacle to utilizing text guidance in visual sound source localization is grounding fine-grained audio and visual features with their textual representation. Later, CLIP [39] introduced learning visual language grounding from web-scraped 400M imagetext pairs. Recently, AudioCLIP [17] introduced the audio modality in the existing CLIP architecture, thereby generating tri-modal feature grounding through large-scale training. Our approach capitalizes on the tri-modal joint embedding space of AudioCLIP to disentangle one-to-one audio-visual correspondence in natural mixtures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.473, + 0.759 + ], + "angle": 0, + "content": "To this end, in this paper, we propose a novel text-guided multi-source localization framework that can discover fine-grained audio-visual semantic correspondence in multisource mixtures. Initially, we detect the class instances of visual sounding objects in the frame using the noisy mixture features from AudioCLIP image and audio encoders. Then, the text representation of each detected sounding source class instance is extracted with AudioCLIP text encoder, which serves as a coarse guidance for audio and visual feature separation. Specifically, we disentangle the categorical audio and visual features of each sounding source using the coarse text-guidance through additional audio and image conditioning blocks, respectively. Finally, the extracted categorical audio-visual semantic features are further aligned through an audio-visual correspondence block for localizing each sounding source. In comparison with prior multisource baselines, our approach can selectively localize the semantic visual region of each class of sounding objects in mixtures of varying number of sources. Extensive experiments on MUSIC [49], VGGSound [6], and VGGSound-Instruments [21] demonstrate significant performance improvements compared to existing single and multi-source baselines. Moreover, our method shows promising zero-shot transferability on unseen classes during test time. In addition, thorough ablation studies and qualitative analysis vividly showcase the effectiveness of the proposed framework." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.763, + 0.459, + 0.777 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.781, + 0.468, + 0.825 + ], + "angle": 0, + "content": "- We propose a novel text-guided multi-source localization framework to disentangle one-to-one audio-visual correspondence from natural mixtures." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.826, + 0.468, + 0.871 + ], + "angle": 0, + "content": "- Our work is the first to introduce AudioCLIP for utilizing the tri-modal feature grounding from large-scale pretraining in solving multi-source localization problems." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.468, + 0.902 + ], + "angle": 0, + "content": "- Our proposed framework can operate with a flexible number of sources and shows promising zero-shot perfor" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.781, + 0.468, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.093, + 0.768, + 0.106 + ], + "angle": 0, + "content": "mance on unseen audio-visual classes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.107, + 0.892, + 0.151 + ], + "angle": 0, + "content": "- Extensive experiments on benchmark datasets clearly demonstrate the superiority of our proposed framework over other state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.163, + 0.642, + 0.178 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.189, + 0.892, + 0.446 + ], + "angle": 0, + "content": "Visual Sound Source Localization. Visual sound source localization aims to locate the visual regions of sounding objects in video frames. Prior work explored diverse methods for audio-visual feature alignment, including traditional statistical machine learning approaches [12, 18, 23] to latest deep neural nets [1, 7, 19, 28, 29, 36, 40, 41]. However, most of these self-supervised and weakly-supervised methods are mostly designed for localizing single-source sounds. A two-stream architecture is proposed in Attention10k [40] which incorporates cross-attention to locate sounding sources. Later, motion cues present in videos are incorporated for finer localization [1]. Recently, various contrastive learning based methods are explored in LVS [7], EZ-VSL [28], and SLAVC [29] for single-source localization. Despite their promising performance, these single-source localization methods can hardly discriminate audio-visual correspondence in multi-source mixtures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Sounds mostly appear in mixtures in the presence of significant background noises. The main challenges of multisource localization is to disentangle the audio-visual correspondence in mixtures without having access to clean single-source audio-visual pairs. Several recent works have explored multi-source localization with explicit approaches for mixtures [20, 21, 30, 36]. To selectively isolate the silent objects in visual frames, DSOL [20] proposed a two-stage weakly-supervised training method. Mix-and-localize [21] proposed a self-supervised training with contrastive random walk to inherently align each sounding source with its visual representation. Recently, AVGN [30] introduced weakly-supervised grouping of category-aware features with learnable prompts. However, these approaches tackle the audio-visual correspondence in mixtures without any explicit guidance for single-source thereby resulting in sub-optimal performance. In contrast, we leverage text representation of fine-grained sounding sources to disentangle one-to-one audio-visual correspondence from multi-source mixtures. To the best of our knowledge, this work is the first to utilize text representation of sound sources for solving multisource localization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Audio, Visual, and Text Grounding. Large-scale pretraining on web-scraped data has been used to generate grounded features across modalities. CLIP [39] first introduced vision-language grounding using 400M imagedtext pairs, which has been extensively studied for zero-shot classification [2, 11, 14, 15, 22, 22, 44], text-video retrieval [3, 8, 25], open-vocabulary segmentation [9, 13, 26, 31, 37, 47, 48] and object detection [5, 10, 27, 45]. Re" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26743" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.372, + 0.895, + 0.457 + ], + "angle": 0, + "content": "Figure 2. The proposed text-guided visual sound source localization (T-VSL) framework (for \\( K = 2 \\)). We use the text modality to disentangle the fine-grained audio-visual correspondence from mixtures. Initially, we detect the audio-visual class instances from multisource mixtures using AudioCLIP joint embedding space. Later, categorical text features of each detected \\( K \\) classes are used as coarse guidance in conditioning blocks to extract categorical visual and audio features. Afterwards, cross-modal feature alignment on extracted categorical features is performed in audio-visual correspondence block. Finally, cosine similarity of mean categorical audio features, and aligned visual features are used recursively to extract localization map of each class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.473, + 0.622 + ], + "angle": 0, + "content": "cently, AudioCLIP [17] and Wav2CLIP [46] introduced additional audio grounding in original CLIP model by learning on image, text, and audio triplets. Very recently, a CLIP-based single source localization method [34] is introduced that attempts replacing the text-conditioning branch of a supervised image segmentation baseline, trained with dense supervision, with audio conditioning. In relation with this concurrent work, our focus lies primarily on weakly supervised sound source localization in multi-source mixtures, leveraging the self-supervised AudioCLIP model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.17, + 0.65 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Given a mixture of audio and a video frame, our objective is to spatially localize each sounding object in the frame. We propose a novel text-guided multi-source localization framework to disentangle the fine-grained audio-visual correspondence, which is illustrated in Fig. 2. To suppress the background noises, we initially detect the common class instances in both audio and visual modality (Sec. 3.2). In particular, we leverage tri-modal AudioCLIP encoders to detect \\( K \\) visual-sounding class instances from \\( N \\) classes \\( (K \\leq N) \\). The primary challenge in multi-source localization lies in establishing one-to-one audio-visual correspondences for individual sources without access to single-source samples. To overcome this, we leverage the text modality to generate coarse representation of the \\( K \\) detected sources in the mixture. Later, by exploiting the multimodal grounding of AudioCLIP, we extract categorical au" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.895, + 0.577 + ], + "angle": 0, + "content": "dio and visual features from multi-source representation using text guidance (Sec. 3.3). Rather than aligning audio-visual mixture features as prior work, we introduce the categorical audio-visual feature alignment to further enhance one-to-one correspondence (Sec. 3.4). Finally, cosine similarity of categorical audio features and aligned visual features is used to extract semantic localization maps." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.588, + 0.642, + 0.603 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.895, + 0.658 + ], + "angle": 0, + "content": "In this section, we formulate the multi-source localization problem, and revisit the AudioCLIP [17] for audio, visual, and text grounding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.667, + 0.896, + 0.819 + ], + "angle": 0, + "content": "Problem Formulation. Let's assume the dataset contains a total of \\(N\\) classes of sounding events across all videos. Given a video frame and an audio mixture, our objective is to localize \\(K(K \\leq N)\\) sound sources within the frame. For each audio-visual mixture pair, we can extract \\(Y = \\{y_{i}\\}_{i=1}^{N}\\) binary ground truth labels, along with categorical text representations for the on-screen sound sources in the mixture. During training, we are limited to video-level sounding class entities, excluding bounding boxes or mask-level annotations, thus constituting weakly-supervised learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Revisit AudioCLIP for multi-modal grounding. Audio-CLIP [17] learns a joint embedding space of audio, visual, and text triplets through contrastive learning. Let's consider the dataset \\(\\mathcal{D} = \\{a,v,t\\}_{i = 1}^{Z}\\) containing audio spectrogram \\(a\\in \\mathbb{R}^{M\\times F}\\), text \\(t\\in \\mathbb{R}^l\\), and image \\(v\\in \\mathbb{R}^{C\\times H\\times W}\\) triplets." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26744" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.09, + 0.473, + 0.213 + ], + "angle": 0, + "content": "Also, separate audio encoder \\((E_{a}:\\mathbb{R}^{M\\times F}\\to \\mathbb{R}^{D})\\), visual encoder \\((E_{v}:\\mathbb{R}^{C\\times H\\times W}\\to \\mathbb{R}^{D})\\), and text encoder \\((E_{t}:\\mathbb{R}^{l}\\to \\mathbb{R}^{D})\\) are used for each modality. Hence, the extracted audio, visual, and text features are represented by \\(A = E_{a}(a)\\), \\(V = E_{v}(v)\\), and \\(T = E_{t}(t)\\), respectively, where \\(A,V,T\\in \\mathbb{R}^D\\). Later, InfoNCE contrastive loss \\((\\mathcal{L}_{\\mathrm{cnt}})\\) [33] across each modality pair is used for feature grounding given by:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.225, + 0.47, + 0.246 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a c l i p}} = \\mathcal {L} _ {\\mathrm {c n t}} (A, T) + \\mathcal {L} _ {\\mathrm {c n t}} (V, T) + \\mathcal {L} _ {\\mathrm {c n t}} (A, V) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.257, + 0.47, + 0.303 + ], + "angle": 0, + "content": "This training objective, combined with large-scale pretraining on web-scraped data, establishes a grounded joint embedding space across all three modalities." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.314, + 0.406, + 0.33 + ], + "angle": 0, + "content": "3.2. Audio-Visual Class Instance Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.47, + 0.475 + ], + "angle": 0, + "content": "The target sounding objects for localization must be present in both the audio mixture and the corresponding video frame. However, audio mixtures may contain noise from off-screen background sources, while video frames may include non-sounding objects. To address these challenges, we first detect the class instances of visible sounding objects in the frame. Then, we utilize this detection to suppress redundant features in the mixture, ensuring that the localization focuses on the relevant sounding objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.473, + 0.779 + ], + "angle": 0, + "content": "Token extraction from Uni-modal encoders. Multiple sounding objects can be spatially located in different regions of the reference frame, while the audio mixture may contain the temporally distributed signal of different sources. This makes it challenging to detect multiple class instances with uni-dimensional single-token features generated from AudioCLIP encoders. Instead of a single pooled token representation, we extract numerous patch tokens from audio and visual encoders with simple modifications, such that \\(\\widehat{E}_a: \\mathbb{R}^{M \\times F} \\to \\mathbb{R}^{n_a \\times D}\\) and \\(\\widehat{E}_v: \\mathbb{R}^{C \\times H \\times W} \\to \\mathbb{R}^{n_v \\times D}\\). Here, \\(n_a = m \\times f\\) and \\(n_v = h \\times w\\) represent the number of audio and visual patch tokens, respectively. Afterwards, we apply linear projectors on these extracted patch tokens, such that \\(P_a: \\mathbb{R}^{n_a \\times D} \\to \\mathbb{R}^{n_a \\times D}\\) and \\(P_v: \\mathbb{R}^{n_v \\times D} \\to \\mathbb{R}^{n_v \\times D}\\). Thus, extracted audio and visual patch tokens are represented as \\(A' = P_a(\\widehat{E}_a(a))\\), \\(V' = P_v(\\widehat{E}_v(v))\\), respectively, where \\(A' \\in \\mathbb{R}^{n_a \\times D}\\) and \\(V' \\in \\mathbb{R}^{n_v \\times D}\\). Simultaneously, categorical text features \\(\\mathcal{T} = \\{e_i^t\\}_{i=1}^N \\in \\mathbb{R}^{N \\times D}\\) representing all \\(N\\) sound source classes in the dataset are extracted." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Multi-label class instance detection. With the categorical text features \\(\\mathcal{T}\\), mixture audio \\(A^{\\prime}\\), and visual \\(V^{\\prime}\\) patch tokens, we detect class entities present in the mixture. Initially, we extract the mean-pooled audio-visual features \\(X^{\\prime} \\in \\mathbb{R}^{2D}\\) by simple concatenation, and then, apply fusion projector \\(P_{f}: \\mathbb{R}^{2D} \\to \\mathbb{R}^{D}\\) to extract mixture audio-visual features \\(F_{av} \\in \\mathbb{R}^{D}\\). Finally, cosine similarity between \\(N\\) class text features \\(\\mathcal{T}\\) and \\(F_{av}\\) is used to detect mixture class" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.089, + 0.682, + 0.108 + ], + "angle": 0, + "content": "entities \\(\\widehat{Y}\\in \\mathbb{R}^N\\) , given by:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.116, + 0.895, + 0.142 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F _ {a v} = P _ {f} \\left(X ^ {\\prime}\\right), X ^ {\\prime} = \\left[ \\operatorname {M e a n} \\left(A ^ {\\prime}\\right) \\oplus \\operatorname {M e a n} \\left(V ^ {\\prime}\\right) \\right], \\\\ \\widehat {\\quad} \\end{array} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.138, + 0.654, + 0.156 + ], + "angle": 0, + "content": "\\[\n\\widehat {Y} = \\operatorname {s i m} (\\mathcal {T}, F _ {a v})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.166, + 0.894, + 0.243 + ], + "angle": 0, + "content": "where \\([\\oplus ]\\) denotes the concatenation operator, and \\(\\operatorname {sim}(X,Y)\\) denotes cosine similarity between \\(X\\) and \\(Y\\). We formulate a multi-label classification objective \\(\\mathcal{L}_{\\mathrm{mcid}}\\) by applying binary cross-entropy loss \\(\\mathcal{L}_{\\mathrm{bce}}(\\cdot)\\) on each prediction using video-level ground-truth labels \\(Y\\in \\mathbb{R}^{N}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.254, + 0.892, + 0.295 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {m c i d}} = \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathrm {b c e}} \\left(y _ {i}, \\widehat {y} _ {i}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.3, + 0.892, + 0.375 + ], + "angle": 0, + "content": "In particular, we train audio, visual, and fusion projectors \\((P_{a}, P_{v}, P_{f})\\) on noisy multi-source data keeping the AudioCLIP backbones frozen. Hence, these projectors assist in extracting relevant foreground features from noisy audio and visual mixture features extracted from AudioCLIP." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.384, + 0.83, + 0.401 + ], + "angle": 0, + "content": "3.3. Audio and Visual Conditioning Blocks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.62 + ], + "angle": 0, + "content": "Establishing one-to-one correspondences for each source in multi-source mixtures is particularly challenging, as both audio and visual features contain mixed representations of various sources. Prior works [21, 30] on multi-source localization have attempted to learn audio-visual alignment intrinsically without explicit single-source guidance. However, in the absence of single-source data, these methods struggle to adequately disentangle multi-source features. In contrast, we leverage the categorical text features of \\( K \\) classes present in the mixture as coarse guidance for fine-grained feature disentanglement from multi-source mixtures. We use ground truth weak labels \\( Y \\) of representative audios in the mixture to extract text embedding of \\( K \\) visual sounding sources \\( \\mathcal{T}_K = \\{e_k^t\\}_{k=1}^K \\subseteq \\mathcal{T} \\) where \\( e_k^t \\in \\mathbb{R}^{1 \\times D} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.68 + ], + "angle": 0, + "content": "With the categorical text features \\(\\mathcal{T}_K\\) of \\(K\\) sources, we independently disentangle the multi-object visual patch tokens \\(V^{\\prime}\\) and mixture audio patch tokens \\(A^\\prime\\) by using crossattention \\(\\phi^v (\\cdot)\\) and \\(\\phi^a (\\cdot)\\), respectively, as:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.689, + 0.892, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\begin{array}{c} \\widetilde {V} _ {k} = \\phi^ {v} \\left(V ^ {\\prime}, e _ {k} ^ {t}, V ^ {\\prime}\\right), f _ {k} ^ {v} = \\operatorname {M e a n} \\left(\\widetilde {V} _ {k}\\right), \\\\ \\sim \\end{array} \\tag {4} \\\\ \\tilde {f} _ {k} ^ {v} = P _ {v c} \\left(f _ {k} ^ {v}\\right), \\forall k = \\{1, \\dots , K \\} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.742, + 0.892, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\underset {\\sim} {\\widetilde {A}} _ {k} = \\phi^ {a} \\left(A ^ {\\prime}, e _ {k} ^ {t}, A ^ {\\prime}\\right), f _ {k} ^ {a} = \\operatorname {M e a n} \\left(\\widetilde {A} _ {k}\\right), (5) \\\\ \\hat {f} _ {k} ^ {a} = P _ {a c} \\left(f _ {k} ^ {a}\\right), \\forall k = \\{1, \\dots , K \\} \\\\ \\phi (A, B, C) = \\left(\\frac {A B ^ {\\top}}{\\| A \\| \\| B \\|}\\right) \\odot C (6) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.838, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\widetilde{f}_k^a,\\widetilde{f}_k^v,f^a,f^v\\in \\mathbb{R}^{1\\times D},\\widetilde{V}_k\\in \\mathbb{R}^{n_v\\times D},\\widetilde{A}_k\\in\\) \\(\\mathbb{R}^{n_a\\times D}\\) , and \\(P_{ac},P_{vc}:\\mathbb{R}^D\\to \\mathbb{R}^D\\) are linear audio and visual projectors on conditioned mean features, respectively, and \\(\\odot\\) is the Hadamard product." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26745" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.878, + 0.32 + ], + "angle": 0, + "content": "
MethodVGGSound-SingleVGGSound-InstrumentsMUSIC-Solo
AP(%)IoU@0.5(%)AUC(%)AP(%)IoU@0.3(%)AUC(%)AP(%)IoU@0.5(%)AUC(%)
Current SOTA Based Methods
OTS [4] ECCV1838.943.243.750.956.938.476.257.848.9
CoarsetoFine [36] ECCV2039.443.144.851.157.539.176.358.149.7
LVS [7] CVPR2139.744.645.351.557.939.476.157.149.3
EZ-VSL [28] ECCV2240.645.147.252.358.839.978.458.751.1
Mix-and-Localize [21] CVPR2241.845.647.453.859.341.578.959.952.7
DSOL [20] NeurIPS20-46.847.9-61.643.7-62.754.6
MarginNCE [35] ICASSP2342.546.948.356.961.844.283.163.355.2
FNAC [43] CVPR2343.348.449.157.263.245.384.664.556.4
AVGN [30] CVPR2344.149.649.559.364.746.185.465.856.9
Alignment [42] ICCV2345.350.850.260.465.848.786.166.457.2
CLIP-Based Baseline Methods
AudioCLIP [17] ICASSP2242.847.448.558.362.745.283.863.155.7
Wav2CLIP [46] ICASSP2239.343.644.753.857.242.078.959.951.2
T-VSL (Ours)48.153.752.964.669.551.488.268.560.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.329, + 0.892, + 0.358 + ], + "angle": 0, + "content": "Table 1. Performance comparison of single-source localization on VGGSound-single, VGGSound-Instruments, and MUSIC-Solo datasets. For the fair comparison, we use pre-trained AudioCLIP audio and image encoders for baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.473, + 0.45 + ], + "angle": 0, + "content": "Additionally, to guide extracting categorical audio \\((\\widetilde{A}_k)\\) and visual \\((\\widetilde{V}_k)\\) patch tokens, we introduce coarse supervision on the output of projected mean patch tokens \\(\\widetilde{f}_k^a,\\widetilde{f}_k^v\\) using \\(N\\)-class text embedding \\(\\mathcal{T}\\in \\mathbb{R}^{N\\times D}\\). Hence, the class conditioning loss \\((\\mathcal{L}_{\\mathrm{cls}})\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.459, + 0.386, + 0.478 + ], + "angle": 0, + "content": "\\[\n\\mathbf {e} _ {k} ^ {v} = \\operatorname {s i m} \\left(\\mathcal {T}, \\widetilde {f _ {k} ^ {v}}\\right); \\mathbf {e} _ {k} ^ {a} = \\operatorname {s i m} \\left(\\mathcal {T}, \\widetilde {f _ {k} ^ {a}}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.481, + 0.469, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c l s}} = \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathbf {e} _ {k} ^ {v}, \\mathbf {h} _ {k}\\right) + \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathbf {e} _ {k} ^ {a}, \\mathbf {h} _ {k}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.532, + 0.47, + 0.58 + ], + "angle": 0, + "content": "where \\(\\mathbf{h}_k\\in \\mathbb{R}^N\\) denotes an one-hot encoding of the class label, and \\(\\mathbf{e}_k^v\\in \\mathbb{R}^N\\), \\(\\mathbf{e}_k^a\\in \\mathbb{R}^N\\) are visual and audio class predictions for each \\(k^{th}\\) source, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.394, + 0.604 + ], + "angle": 0, + "content": "3.4. Audio-Visual Correspondence Block" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.47, + 0.746 + ], + "angle": 0, + "content": "While text representation provides coarse guidance for disentangling fine-grained spatio-temporal audio and visual features, it has inherent limitations in practice. For example, multiple instances of a sounding object class can appear in the visual frame, including silent instances, which are difficult to distinguish using simple text class representation. To address this challenge, we introduce an audio-visual correspondence block to further refine the alignment of audio and visual features based on extracted categorical features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.47, + 0.793 + ], + "angle": 0, + "content": "With the disentangled categorical audio \\((\\widetilde{A}_k)\\) and visual \\((\\widetilde{V}_k)\\) patch tokens, we apply cross-attention \\(\\phi^{av}(\\cdot)\\) to enhance audio-visual patch alignment as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.8, + 0.356, + 0.82 + ], + "angle": 0, + "content": "\\[\ng _ {k} ^ {a} = \\operatorname {M e a n} \\left(\\widetilde {A} _ {k}\\right), \\widehat {g} _ {k} ^ {a} = P _ {a v} \\left(g _ {k} ^ {a}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.822, + 0.469, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\widehat {V} _ {k} = \\phi^ {a v} \\left(\\widetilde {V} _ {k}, \\widehat {g} _ {k} ^ {a}, \\widetilde {V} _ {k}\\right), g _ {k} ^ {v} = \\operatorname {M e a n} \\left(\\widehat {V} _ {k}\\right) \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.843, + 0.366, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\widehat {g} _ {k} ^ {v} = P _ {v a} \\left(g _ {k} ^ {v}\\right), \\forall k = \\{1, \\dots , K \\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(P_{av}, P_{va}: \\mathbb{R}^D \\to \\mathbb{R}^D\\) are linear projectors. Afterwards, to correspond between projected mean categorical" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.37, + 0.892, + 0.403 + ], + "angle": 0, + "content": "audio and visual features \\(\\widehat{g}_k^a,\\widehat{g}_k^v\\in \\mathbb{R}^D\\) , we apply audiovisual contrastive loss \\((\\mathcal{L}_{\\mathrm{av}})\\) given by:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.41, + 0.891, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a v}} = \\mathcal {L} _ {\\mathrm {c n t}} \\left(\\widehat {g} _ {k} ^ {v}, \\widehat {g} _ {k} ^ {a}\\right) \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.431, + 0.8, + 0.447 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {a v}} + \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {m c i d}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.453, + 0.783, + 0.47 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{total}}\\) is the total accumulated loss." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.475, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Inference: During inference, we estimate cosine similarity across projected mean audio feature \\(\\widehat{g}_k^a\\in \\mathbb{R}^D\\) and aligned visual patch tokens \\((\\widehat{V}_k = \\{\\widehat{v}_k^j\\}_{j = 1}^{n_v},\\widehat{v}_k^j\\in \\mathbb{R}^D)\\). Finally, after reshaping and up-scaling through bilinear interpolation, we generate \\(K\\)-source heat-maps \\(\\mathbf{H}\\in \\mathbb{R}^{K\\times H\\times W}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.558, + 0.724, + 0.574 + ], + "angle": 0, + "content": "4. Results and Discussions" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.577, + 0.689, + 0.593 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.857 + ], + "angle": 0, + "content": "Datasets. We used MUSIC, VGGSound-Instruments, and VGGSound datasets for the performance evaluation1. MUSIC [49] dataset contains 445 solo music videos of 11 instruments and 142 duet music videos of 8 instruments from YouTube. Following prior works [30], we use MUSIC-Solo for single-source and MUSIC-Duet for multi-source localization. From MUSIC-Solo, we use 350 videos for training and remaining 95 for evaluation. For MUSIC-Duet, we use 120 videos for training and remaining 22 videos for evaluation. VGGSound-Instruments [21] is a subset of VGGSound dataset [6] containing around 32k video clips from 37 music instruments of 10s duration. Apart from the musical instrument datasets, we gather around 150k videos of 10s duration from original VGGSound dataset [6] representing single-source sounds of 221 categories, such as animals, people, nature, instruments, etc. For the single-source evaluation on VGGSound, we use the full test" + }, + { + "type": "page_footnote", + "bbox": [ + 0.499, + 0.863, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1Since many videos are no longer available for public use, these datasets become smaller. We use the same data split for all baselines and reproduce them under the same setting for a fair comparison." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26746" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.089, + 0.089, + 0.885, + 0.33 + ], + "angle": 0, + "content": "
MethodVGGSound-DuetVGGSound-InstrumentsMUSIC-Duet
CAP(%)CIoU@0.3(%)AUC(%)CAP(%)CIoU@0.1(%)AUC(%)CAP(%)CIoU@0.3(%)AUC(%)
Current SOTA Single-Source Methods
OTS [4] ECCV1823.926.427.528.777.818.251.229.424.3
CoarseToFine [36] ECCV20-28.427.9-78.819.7-31.824.1
LVS [7] CVPR21-31.830.1-81.123.2-33.126.9
EZ-VSL [28] ECCV22-32.430.6-80.322.9-34.326.7
MarginNCE [35] ICASSP2327.133.231.233.582.324.755.136.128.5
FNAC [43] CVPR2329.835.133.435.484.126.857.337.930.6
Alignment [42] ICCV2330.435.833.935.684.927.157.838.330.9
Current SOTA Multi-Source Methods
Mix-and-Localize [21] CVPR2229.535.634.136.384.526.757.438.130.7
DSOL [20] NeurIPS20-36.934.6-85.928.4-38.831.3
AVGN [30] CVPR2331.937.835.438.186.728.859.239.332.4
CLIP-Based Baseline Methods
AudioCLIP [17] ICASSP2228.434.932.834.783.825.956.136.929.2
Wav2CLIP [46] ICASSP2226.331.428.931.280.322.252.632.227.2
T-VSL (ours)35.740.137.941.889.631.562.943.235.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.334, + 0.892, + 0.361 + ], + "angle": 0, + "content": "Table 2. Performance comparison of multi-source localization on VGGSound-Duet, VGGSound-Instruments, and MUSIC-Duet datasets. For the fair comparison, we use pre-trained AudioCLIP audio and image encoders for baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.376, + 0.47, + 0.482 + ], + "angle": 0, + "content": "set of 5158 videos, while on VGGSound-Instruments, we use 446 videos following prior work [21, 30]. For the multi-source training and evaluation in both VGGSound and VGGSound-Instruments, we randomly concatenate two frames from different sounding sources to produce the multi-source input image with a size of \\(448 \\times 224\\), and mix the single-source audios, following prior work [21, 30]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.47, + 0.681 + ], + "angle": 0, + "content": "Evaluation Metrics. For evaluating single-source localization performance, we use Intersection over Union (IoU), average precision (AP), and Area Under Curve (AUC) following prior work [21, 30]. For the multi-source localization evaluation, we use Class-aware Average precision (CAP), Class-aware IoU (CIoU), and Area Under Curve (AUC) following prior work [21, 30]. Similarly, for fair comparison with the prior work [30], we use the same thresholds in these metrics: we use IoU@0.5 and CIoU@0.3 for the MUSIC-Solo and MUSIC-Duet datasets, IoU@0.3 and IoU@0.5 for single-source VGGSound-Instruments and VGGSound datasets, and CIoU@0.1 and CIoU@0.3 for multi-source VGGSound-Instruments and VGGSound." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Implementation details. We use the AudioCLIP [17] pretrained encoders containing ResNet-50 [19] image encoder, ESResNeXt [16] audio encoder, and transformer [38] based text encoder. For the fair comparison, we reproduce the result of existing methods with AudioCLIP image and audio encoders, instead of using separately pre-trained encoders. We use \\(224 \\times 224\\) resolution for the single-source input image, \\(D = 1024\\), \\(n_v = 49\\) for \\(7 \\times 7(h \\times w)\\) spatial maps generated from the visual encoder, and \\(n_a = 60\\) for \\(10 \\times 6(m \\times f)\\) time-frequency map generated from the audio encoder. We use Adam optimizer [24] with a batch size of 256 and with a learning rate of \\(1e - 4\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.9 + ], + "angle": 0, + "content": "CLIP-Based Baseline methods. We adopt several CLIP-based baseline methods for the sound source localization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.376, + 0.744, + 0.389 + ], + "angle": 0, + "content": "task, which are described as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.393, + 0.891, + 0.451 + ], + "angle": 0, + "content": "- AudioCLIP [17]: We fine-tune pre-trained image and audio encoders of AudioCLIP on the target dataset. We estimate the localization heatmap for the sounding sources using the similar cosine similarity across patch tokens." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.453, + 0.891, + 0.498 + ], + "angle": 0, + "content": "- Wav2CLIP [46]: Similar to AudioCLIP baseline, we experiment with Wav2CLIP audio and image encoders to extract localization heatmaps of sounding sources." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.393, + 0.891, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.512, + 0.776, + 0.527 + ], + "angle": 0, + "content": "4.2. Comparison to SOTA Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Single-source localization: We present the quantitative results on single-source localization in Tab. 1. We apply the AudioCLIP image and audio encoders in all baseline methods for the fair comparison with our methods. We observe some performance improvements with several SOTA methods [30, 42, 43] over the AudioCLIP baseline. However, we also notice some reduction in performance from the AudioCLIP baseline in several methods [4, 21, 28]. We hypothesize that these can be due to the alterations of AudioCLIP grounding with the modified contrastive loss. In contrast, by utilizing all three modalities for sound source localization, the proposed T-VSL achieves the best performance over all CLIP-based baselines as well as SOTA methods. T-VSL outperforms the current SOTA single source localization method Alignment [42] by \\(+2.9\\), \\(+3.7\\), \\(+2.1\\) higher IoUs on VGGSound-Single, VGGSound-Instruments, and MUSIC-solo datasets, respectively. Moreover, T-VSL achieves \\(+6.3\\), \\(+6.8\\), and \\(+5.4\\) higher IoUs over the AudioCLIP baseline [32]. We hypothesize that the enhanced refinement of noisy audio and visual features with text guidance in T-VSL effectively contributes to performance improvement." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multi-source localization: We present performance comparison in multi-source sound localization in Tab. 2. As be" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26747" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.089, + 0.849, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.401, + 0.893, + 0.445 + ], + "angle": 0, + "content": "Figure 3. We present qualitative comparisons on challenging multi-source localization with SOTA single and multi-source baseline methods. Here, blue color represents high-attention values to the sounding object, and red color represents low-attention values. The proposed T-VSL can selectively isolate the sounding regions from the background and generates more precise localization maps for sounding sources." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.473, + 0.684 + ], + "angle": 0, + "content": "fore, we apply the AudioCLIP image and audio encoders in all baseline methods for the fair comparison with our methods. We observe some performance improvements, mostly in SOTA multi-source methods [20, 21, 30], over the AudioCLIP baseline. In most single source methods, we observe performance loss in multi-source cases over the baseline. The proposed T-VSL achieves the best performance that outperforms the SOTA multi-source baseline AVGN [30] by \\(+2.3\\), \\(+2.9\\), and \\(+3.9\\) CIoUs on multi-source VGGSound-Duet, VGGSound-Instruments, and MUSIC-Duet datasets, respectively. Also, T-VSL achieves \\(+5.2\\), \\(+5.8\\), and \\(+6.3\\) higher CIoUs over the AudioCLIP baseline. These significant performance gains on multi-source benchmarks demonstrate the effectiveness of the proposed T-VSL in disentangling multi-source mixtures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.693, + 0.47, + 0.906 + ], + "angle": 0, + "content": "Qualitative comparisons: In addition, we present qualitative comparisons on multi-source localization in Fig. 3 among single-source baseline Alignment [42], self-supervised multi-source baseline Mix-and-Localize [21], weakly-supervised multi-source baseline AVGN [30], and proposed T-VSL. Single-source SOTA baseline Alignment [42] struggles in multi-source localization due to the lack of audio-visual correspondence, mostly in the presence of noisy audios and silent visual objects. However, without any guidance on fine-grained source localization, multi-source baselines [21, 30] underperform on localizing challenging multi-source mixtures. Notably, our method generates superior localization maps, which shows the effectiveness of T-VSL in multi-source feature disentanglement." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.457, + 0.804, + 0.473 + ], + "angle": 0, + "content": "4.3. Ablation on T-VSL building blocks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.892, + 0.775 + ], + "angle": 0, + "content": "We present the ablation study of different building blocks of T-VSL in Tab. 3 under both single-source and multisource settings. The vanilla baseline contains AudioCLIP image and audio encoders that directly operate on the input data. By integrating the audio-visual correspondence (AVC) block on extracted mixture patch tokens, we notice considerable performance increase in single-source (by \\(+1.1\\) AP and \\(+0.7\\) IoU@0.5) and multi-source localization (by \\(+1.3\\) CAP and \\(+1.4\\) CIoU@0.3). By integrating the audio and image conditioning blocks following audio-visual instance detection for separating fine-grained class features, we observe additional improvements of \\(+1.8\\) AP and \\(+2.3\\) AP in single-source, and \\(+3.1\\) CAP and \\(+3.6\\) AP in multisource localization, respectively. Finally, the combination of all building blocks results in the best performance that improves the baseline by \\(+5.3\\) AP, and 6.3 IoU@0.5 in single-source and by \\(+7.3\\) CAP and \\(+5.2\\) CIoU@0.3 in multi-source localization. This further demonstrates the effective disentanglement of audio-visual features facilitated by intermediate text guidance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.779, + 0.796, + 0.793 + ], + "angle": 0, + "content": "4.4. Zero-shot transfer across datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We can perform zero-shot transfer across datasets with proposed T-VSL by simply replacing the \\(N\\)-class text prompts in Fig. 2. Since other weakly-supervised methods cannot perform zero-shot transfer, we present comparisons with self-supervised Mix-and-Localize [21], Alignment [42], and FNAC [43] methods in Tab. 4. To ensure a fair comparison, we conducted uniform training with \\(50\\% - 50\\%\\) class" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26748" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.089, + 0.465, + 0.187 + ], + "angle": 0, + "content": "
Audio Cond.Image Cond.AVCVGGSound-SingleVGGSound-Duet
AP(%)IoU@0.5(%)CAP(%)CIoU@0.3(%)
XXX42.847.428.434.9
XX✓43.948.129.736.3
✓X✓45.750.932.838.1
X✓✓46.251.433.338.6
✓✓✓48.153.735.740.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.197, + 0.47, + 0.24 + ], + "angle": 0, + "content": "Table 3. Ablation study on audio conditioning block, image-conditioning block, and audio-visual correspondence block (AVC) of T-VSL in single-source and multi-source localization tasks." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.248, + 0.468, + 0.329 + ], + "angle": 0, + "content": "
Target DatasetProposed T-VSLMix-and-Localize [21]Alignment [42]FNAC [43]
MUSIC-Duet55.850.951.751.2
MUSIC-Solo80.575.177.677.4
VGGSound-Instr.81.776.876.576.1
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.34, + 0.47, + 0.408 + ], + "angle": 0, + "content": "Table 4. Ablation on zero-shot transfer across single and multi-source datasets with \\(50\\% -50\\%\\) class split. We report IoU@0.5(%) for MUSIC-Solo, CIoU@0.3(%) for MUSIC-Duet, and CIoU@0.1(%) for VGGSound-Instruments. We use same AudioCLIP encoders for all baselines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.47, + 0.569 + ], + "angle": 0, + "content": "split for all methods with same AudioCLIP encoders. We note that our method achieves significantly better performance on zero-shot transfer to target datasets than these compared approaches. In particular, we achieve \\(+4.9\\), \\(+5.4\\), \\(+4.9\\) improvements on target datasets than the multi-source Mix-and-Localize [21]. Also, we achieve \\(+4.1\\), \\(+2.9\\), and \\(+5.2\\) higher scores than the SOTA single source baseline, Alignment [42]. This result suggests that our T-VSL has superior generalization capabilities for zero-shot localization tasks in unseen classes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.574, + 0.434, + 0.59 + ], + "angle": 0, + "content": "4.5. Robustness to a higher number of sources" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.47, + 0.803 + ], + "angle": 0, + "content": "The proposed T-VSL is not limited by the number of sources present in training mixtures unlike prior work [21]. We present comparative analysis with SOTA multi-source AVGN [30], and self-supervised single source methods FNAC [43] and Alignment [42] in Tab. 5 on robustness to higher number of test sources in VGGSound dataset. We train all methods on the VGGSound-Single dataset with same encoders for a fair comparison. We note that our method consistently maintains significantly higher CIOU score irrespective of test scenarios, and also achieves notably smaller relative performance drop for a large number of sources in mixtures. These results further demonstrate the robustness of the proposed T-VSL in disentangling multi-source mixtures." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.808, + 0.34, + 0.825 + ], + "angle": 0, + "content": "4.6. Use of learnable text prompts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To disentangle multi-source mixtures, we primarily use the class label of each sounding source as text prompts (e.g. dog barking, snake hissing). Inspired by the conditional text prompt learning in [50], we study the use of learnable prompts along with class label representations. Instead of" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.089, + 0.875, + 0.179 + ], + "angle": 0, + "content": "
Test Source No.Proposed T-VSLAVGN [30]Alignment [42]FNAC [43]
239.636.835.434.1
335.730.828.927.1
429.522.419.318.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.183, + 0.892, + 0.225 + ], + "angle": 0, + "content": "Table 5. Robustness to higher number of sources in VGGSound dataset when trained with single-source data. CIoU@0.3(%) score is reported. Same AudioCLIP encoders are used for all baselines." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.233, + 0.88, + 0.365 + ], + "angle": 0, + "content": "
Prompt LengthVGGSound-SingleVGGSound-Duet
AP(%)IoU@0.5(%)CAP(%)CIoU@0.3(%)
048.153.735.740.1
248.554.136.040.4
449.254.636.841.3
850.355.237.142.5
1650.255.037.442.9
3249.754.537.142.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.374, + 0.892, + 0.43 + ], + "angle": 0, + "content": "Table 6. Ablation study on the effect of learnable text prompts on VGGSound-Single and VGGSound-Duet datasets. Integration of learnable prompts results in noticeable performance improvements in both single and multi-source localization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.573 + ], + "angle": 0, + "content": "only using class label representation, a set of learnable embedding is integrated with class label embedding that provides additional flexibility to adapt the text prompt to target objective. We analyze the sound source localization performance with different length of learnable prompts. The performances are reported in Table 6. We notice considerable performance improvements by optimizing learnable prompts in both single and multi-source cases. However, such prompts limit zero-shot use cases to unseen classes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.58, + 0.619, + 0.595 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.81 + ], + "angle": 0, + "content": "In this paper, we propose T-VSL, a novel text-guided multi-source visual sound source localization framework that can disentangle one-to-one audio-visual correspondence from multi-source mixtures. We leverage the text modality to guide fine-grained feature separation and localization, from noisy audio and visual features of mixtures, by exploiting the joint embedding space of AudioCLIP. In comparison with SOTA multi-source baselines, our method shows superior zero-shot transfer across datasets. Moreover, our method demonstrates notable robustness on challenging test mixtures with higher number of sources than training scenarios. In both single-source and multi-source localization, our method largely outperforms existing weakly-supervised and self-supervised baselines on three benchmark datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.819, + 0.655, + 0.835 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This research was supported in part by ONR Minerva program, iMAGiNE - the Intelligent Machine Engineering Consortium at UT Austin, and a UT Cockrell School of Engineering Doctoral Fellowship." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26749" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In Proceedings of European Conference on Computer Vision (ECCV), pages 208–224, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Muhammad Ali and Salman Khan. Clip-decoder: Zeroshot multilabel classification using multimodal clip aligned representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4675-4679, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.47, + 0.271 + ], + "angle": 0, + "content": "[3] Tayfun Alpay, Sven Magg, Philipp Broze, and Daniel Speck. Multimodal video retrieval with clip: a user study. Information Retrieval Journal, 26(1-2):6, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.47, + 0.315 + ], + "angle": 0, + "content": "[4] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proceedings of the European Conference on Computer Vision (ECCV), pages 435-451, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.47, + 0.373 + ], + "angle": 0, + "content": "[5] Maria A Bravo, Sudhanshu Mittal, and Thomas Brox. Localized vision-language matching for open-vocabulary object detection. In DAGM German Conference on Pattern Recognition, pages 393-408. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[6] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[7] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16867-16876, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[8] Chaorui Deng, Qi Chen, Pengda Qin, Da Chen, and Qi Wu. Prompt switch: Efficient clip adaptation for text-video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15648-15658, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.47, + 0.615 + ], + "angle": 0, + "content": "[9] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary panoptic segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.47, + 0.686 + ], + "angle": 0, + "content": "[10] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[11] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model clip. In Proceedings of the AAAI conference on artificial intelligence, pages 6568-6576, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[12] John W Fisher III, Trevor Darrell, William Freeman, and Paul Viola. Learning joint statistical models for audio-visual fusion and segregation. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[13] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19338-19347, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[15] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xupeng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 746-754, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[16] Andrey Guzhov, Federico Raue, Jörn Hees, and Andreas Dengel. Esresne (x) t-fbsp: Learning robust time-frequency transformation of audio. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[17] Andrey Guzhov, Federico Raue, Jorn Hees, and Andreas Dengel. Audioclip: Extending clip to image, text and audio. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 976-980. IEEE, 2022. 2, 3, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[18] John Hershey and Javier Movellan. Audio vision: Using audio-visual synchrony to locate sounds. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[19] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9248-9257, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[20] Di Hu, Rui Qian, Minyue Jiang, Xiao Tan, Shilei Wen, Errui Ding, Weiyao Lin, and Dejing Dou. Discriminative sounding objects localization via self-supervised audiovisual matching. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 10077-10087, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[21] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10483-10492, 2022. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[22] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[23] Einat Kidron, Yoav Y Schechner, and Michael Elad. Pixels that sound. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[25] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[26] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In International Conference on Machine Learning, pages 23033-23044. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26750" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In European Conference on Computer Vision, pages 728-755. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[28] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In Proceedings of European Conference on Computer Vision (ECCV), page 218-234, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[29] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.47, + 0.314 + ], + "angle": 0, + "content": "[30] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10565–10574, 2023. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.47, + 0.398 + ], + "angle": 0, + "content": "[31] Jishnu Mukhoti, Tsung-Yu Lin, Omid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19413-19423, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.47, + 0.481 + ], + "angle": 0, + "content": "[32] Jishnu Mukhoti, Tsung-Yu Lin, Amid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19413-19423, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.47, + 0.524 + ], + "angle": 0, + "content": "[33] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.47, + 0.566 + ], + "angle": 0, + "content": "[34] Sooyoung Park, Arda Senocak, and Joon Son Chung. Can clip help sound source localization? arXiv preprint arXiv:2311.04066, 2023. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.47, + 0.635 + ], + "angle": 0, + "content": "[35] Sooyoung Park, Arda Senocak, and Joon Son Chung. Marginnce: Robust sound localization with a negative margin. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.47, + 0.692 + ], + "angle": 0, + "content": "[36] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In Proceedings of European Conference on Computer Vision (ECCV), pages 292-308, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[37] Jie Qin, Jie Wu, Pengxiang Yan, Ming Li, Ren Yuxi, Xuefeng Xiao, Yitong Wang, Rui Wang, Shilei Wen, Xin Pan, et al. Freeseg: Unified, universal and open-vocabulary image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19446-19455, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.817 + ], + "angle": 0, + "content": "[38] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[40] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4358-4366, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.232 + ], + "angle": 0, + "content": "[41] Arda Senocak, Hyeonggon Ryu, Junsik Kim, and In So Kweon. Learning sound localization better from semantically similar samples. In Proceedings of IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.893, + 0.303 + ], + "angle": 0, + "content": "[42] Arda Senocak, Hyeonggon Ryu, Junsik Kim, Tae-Hyun Oh, Hanspeter Pfister, and Joon Son Chung. Sound source localization is all about cross-modal alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7777-7787, 2023. 1, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.893, + 0.387 + ], + "angle": 0, + "content": "[43] Weixuan Sun, Jiayi Zhang, Jianyuan Wang, Zheyuan Liu, Yiran Zhong, Tianpeng Feng, Yandong Guo, Yanhao Zhang, and Nick Barnes. Learning audio-visual source localization via false negative aware contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6420–6429, 2023. 1, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.388, + 0.893, + 0.444 + ], + "angle": 0, + "content": "[44] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1802-1812, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.445, + 0.893, + 0.5 + ], + "angle": 0, + "content": "[45] Tao Wang. Learning to detect and segment for open vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7051-7060, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.501, + 0.893, + 0.572 + ], + "angle": 0, + "content": "[46] Ho-Hsiang Wu, Prem Seetharaman, Kundan Kumar, and Juan Pablo Bello. Wav2clip: Learning robust audio representations from clip. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4563-4567. IEEE, 2022. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.572, + 0.893, + 0.642 + ], + "angle": 0, + "content": "[47] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2935-2944, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.642, + 0.893, + 0.712 + ], + "angle": 0, + "content": "[48] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.712, + 0.893, + 0.768 + ], + "angle": 0, + "content": "[49] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In Proceedings of the European Conference on Computer Vision (ECCV), pages 570-586, 2018. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.769, + 0.893, + 0.838 + ], + "angle": 0, + "content": "[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.838 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26751" + } + ] +] \ No newline at end of file diff --git a/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/2ac9042c-6511-4205-8b71-c79e7f00a075_origin.pdf b/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/2ac9042c-6511-4205-8b71-c79e7f00a075_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..397cf45545b1d8c05142fd36ca3f4ab688d85de7 --- /dev/null +++ b/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/2ac9042c-6511-4205-8b71-c79e7f00a075_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:593d20d9fc2809cee9d8615edd1acb6fef436fd695b35d3f786685d01d002dee +size 5338220 diff --git a/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/full.md b/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bc657a9336237fa87681a0dc2d3555a5e8fdf8c6 --- /dev/null +++ b/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/full.md @@ -0,0 +1,316 @@ +# T-VSL: Text-Guided Visual Sound Source Localization in Mixtures + +Tanvir Mahmud +University of Texas at Austin + +Yapeng Tian +University of Texas at Dallas + +Diana Marculescu +University of Texas at Austin + +# Abstract + +Visual sound source localization poses a significant challenge in identifying the semantic region of each sounding source within a video. Existing self-supervised and weakly supervised source localization methods struggle to accurately distinguish the semantic regions of each sounding object, particularly in multi-source mixtures. These methods often rely on audio-visual correspondence as guidance, which can lead to substantial performance drops in complex multi-source localization scenarios. The lack of access to individual source sounds in multi-source mixtures during training exacerbates the difficulty of learning effective audio-visual correspondence for localization. To address this limitation, in this paper, we propose incorporating the text modality as an intermediate feature guide using tri-modal joint embedding models (e.g., AudioCLIP) to disentangle the semantic audio-visual source correspondence in multi-source mixtures. Our framework, dubbed T-VSL, begins by predicting the class of sounding entities in mixtures. Subsequently, the textual representation of each sounding source is employed as guidance to disentangle fine-grained audio-visual source correspondence from multi-source mixtures, leveraging the tri-modal AudioCLIP embedding. This approach enables our framework to handle a flexible number of sources and exhibits promising zero-shot transferability to unseen classes during test time. Extensive experiments conducted on the MUSIC, VGGSound, and VGGSound-Instruments datasets demonstrate significant performance improvements over state-of-the-art methods. Code is released at https://github.com/encyac-group/T-VSL/tree/main. + +# 1. Introduction + +While observing a conversation between two individuals, we can easily associate the audio signal with the corresponding speaking person in the visual scene. This remarkable ability to perceive audio-visual correspondence stems from our extensive exposure to both single-source sounds and multi-source mixtures in everyday life. Inspired by this human capability, significant research efforts [21, 30, 42, 43] have been dedicated to developing vi + +![](images/20e68c70fc09ecb7301d28ed1abfe6bc1c6571d92f0c2adb001f53931c8dacef.jpg) + +![](images/0ef5b245a37750e3ce6a6007a28fcf790de95eb2e2866e5caa80e6fb2ee6df3b.jpg) + +![](images/9a80375f90a17aadea0741648bba785c016c836c1b6789f658918fb35096ae4e.jpg) + +![](images/41ebede698abbc14ac2ff7d9a8bcf160b57613ddfd316b9ea0c3eee9723b8d56.jpg) +Figure 1. Comparison of our T-VSL with state-of-the-art methods on single-source (Top Row) and multi-source (Bottom Row) sound localization on VGGSound Sources [7], VGGSoundInstruments [21], and MUSIC [49] benchmark datasets. We use the same setup for the fair comparison. + +![](images/c8a46b3e47749300327f77158d28188652d7bfbfbe0ed56454aad3a5b525b11b.jpg) + +![](images/6bdd148357b304515ec2d9b67b73d2cc2bb78dd544c670cc6d6ce73fed5dcde7.jpg) + +sual sound source localization approaches in recent years. + +Visual sound source localization aims to locate the visual regions representing sound sources present in a video. Earlier methods on single-source localization [1, 7, 19, 28, 29, 40, 41] mostly used audio-visual correspondence as guidance for localizing sounding objects in each frame. These approaches, developed for single sound source localization, suffer from significant performance drop in localizing multi-source mixtures. Multi-source localization from mixtures is very challenging, as the model must learn to distinguish each sounding source within a mixture and establish their cross-modal relationships, without access to individual source sounds. Earlier methods [20, 21, 30, 36] on multi-source localization attempt to infer fine-grained cross-modal associations directly from noisy multi-source mixtures, often leading to sub-optimal performance in practice for learning improper source correspondence. In this paper, we tackle the problem by introducing a unified solution for localizing visual sound sources in both single and multi-source mixtures. + +The primary challenge of this task stems from the difficulty in disentangling the cross-modal correspondence of each sounding object from natural mixtures, given the absence of one-to-one corresponding single-source pairs. Additionally, the presence of silent visual objects and noise from invisible background sources further complicates this alignment process. To solve this problem, our key idea is to + +leverage the text modality as a coarse supervision for disentangling categorical audio-visual correspondence in natural mixtures, which differentiates our method from existing works. Unlike audio and visual modalities that may contain significant noises from different sources present in mixtures, textual representation can explicitly discriminate across multiple sources. However, a critical obstacle to utilizing text guidance in visual sound source localization is grounding fine-grained audio and visual features with their textual representation. Later, CLIP [39] introduced learning visual language grounding from web-scraped 400M imagetext pairs. Recently, AudioCLIP [17] introduced the audio modality in the existing CLIP architecture, thereby generating tri-modal feature grounding through large-scale training. Our approach capitalizes on the tri-modal joint embedding space of AudioCLIP to disentangle one-to-one audio-visual correspondence in natural mixtures. + +To this end, in this paper, we propose a novel text-guided multi-source localization framework that can discover fine-grained audio-visual semantic correspondence in multisource mixtures. Initially, we detect the class instances of visual sounding objects in the frame using the noisy mixture features from AudioCLIP image and audio encoders. Then, the text representation of each detected sounding source class instance is extracted with AudioCLIP text encoder, which serves as a coarse guidance for audio and visual feature separation. Specifically, we disentangle the categorical audio and visual features of each sounding source using the coarse text-guidance through additional audio and image conditioning blocks, respectively. Finally, the extracted categorical audio-visual semantic features are further aligned through an audio-visual correspondence block for localizing each sounding source. In comparison with prior multisource baselines, our approach can selectively localize the semantic visual region of each class of sounding objects in mixtures of varying number of sources. Extensive experiments on MUSIC [49], VGGSound [6], and VGGSound-Instruments [21] demonstrate significant performance improvements compared to existing single and multi-source baselines. Moreover, our method shows promising zero-shot transferability on unseen classes during test time. In addition, thorough ablation studies and qualitative analysis vividly showcase the effectiveness of the proposed framework. + +Our main contributions can be summarized as follows: + +- We propose a novel text-guided multi-source localization framework to disentangle one-to-one audio-visual correspondence from natural mixtures. +- Our work is the first to introduce AudioCLIP for utilizing the tri-modal feature grounding from large-scale pretraining in solving multi-source localization problems. +- Our proposed framework can operate with a flexible number of sources and shows promising zero-shot perfor + +mance on unseen audio-visual classes. + +- Extensive experiments on benchmark datasets clearly demonstrate the superiority of our proposed framework over other state-of-the-art methods. + +# 2. Related Work + +Visual Sound Source Localization. Visual sound source localization aims to locate the visual regions of sounding objects in video frames. Prior work explored diverse methods for audio-visual feature alignment, including traditional statistical machine learning approaches [12, 18, 23] to latest deep neural nets [1, 7, 19, 28, 29, 36, 40, 41]. However, most of these self-supervised and weakly-supervised methods are mostly designed for localizing single-source sounds. A two-stream architecture is proposed in Attention10k [40] which incorporates cross-attention to locate sounding sources. Later, motion cues present in videos are incorporated for finer localization [1]. Recently, various contrastive learning based methods are explored in LVS [7], EZ-VSL [28], and SLAVC [29] for single-source localization. Despite their promising performance, these single-source localization methods can hardly discriminate audio-visual correspondence in multi-source mixtures. + +Sounds mostly appear in mixtures in the presence of significant background noises. The main challenges of multisource localization is to disentangle the audio-visual correspondence in mixtures without having access to clean single-source audio-visual pairs. Several recent works have explored multi-source localization with explicit approaches for mixtures [20, 21, 30, 36]. To selectively isolate the silent objects in visual frames, DSOL [20] proposed a two-stage weakly-supervised training method. Mix-and-localize [21] proposed a self-supervised training with contrastive random walk to inherently align each sounding source with its visual representation. Recently, AVGN [30] introduced weakly-supervised grouping of category-aware features with learnable prompts. However, these approaches tackle the audio-visual correspondence in mixtures without any explicit guidance for single-source thereby resulting in sub-optimal performance. In contrast, we leverage text representation of fine-grained sounding sources to disentangle one-to-one audio-visual correspondence from multi-source mixtures. To the best of our knowledge, this work is the first to utilize text representation of sound sources for solving multisource localization. + +Audio, Visual, and Text Grounding. Large-scale pretraining on web-scraped data has been used to generate grounded features across modalities. CLIP [39] first introduced vision-language grounding using 400M imagedtext pairs, which has been extensively studied for zero-shot classification [2, 11, 14, 15, 22, 22, 44], text-video retrieval [3, 8, 25], open-vocabulary segmentation [9, 13, 26, 31, 37, 47, 48] and object detection [5, 10, 27, 45]. Re + +![](images/9bd966ff2e2af2de15ff2660debcb0f5b45a0b47f8d82317f11a42ea62a99279.jpg) +Figure 2. The proposed text-guided visual sound source localization (T-VSL) framework (for $K = 2$ ). We use the text modality to disentangle the fine-grained audio-visual correspondence from mixtures. Initially, we detect the audio-visual class instances from multisource mixtures using AudioCLIP joint embedding space. Later, categorical text features of each detected $K$ classes are used as coarse guidance in conditioning blocks to extract categorical visual and audio features. Afterwards, cross-modal feature alignment on extracted categorical features is performed in audio-visual correspondence block. Finally, cosine similarity of mean categorical audio features, and aligned visual features are used recursively to extract localization map of each class. + +cently, AudioCLIP [17] and Wav2CLIP [46] introduced additional audio grounding in original CLIP model by learning on image, text, and audio triplets. Very recently, a CLIP-based single source localization method [34] is introduced that attempts replacing the text-conditioning branch of a supervised image segmentation baseline, trained with dense supervision, with audio conditioning. In relation with this concurrent work, our focus lies primarily on weakly supervised sound source localization in multi-source mixtures, leveraging the self-supervised AudioCLIP model. + +# 3. Method + +Given a mixture of audio and a video frame, our objective is to spatially localize each sounding object in the frame. We propose a novel text-guided multi-source localization framework to disentangle the fine-grained audio-visual correspondence, which is illustrated in Fig. 2. To suppress the background noises, we initially detect the common class instances in both audio and visual modality (Sec. 3.2). In particular, we leverage tri-modal AudioCLIP encoders to detect $K$ visual-sounding class instances from $N$ classes $(K \leq N)$ . The primary challenge in multi-source localization lies in establishing one-to-one audio-visual correspondences for individual sources without access to single-source samples. To overcome this, we leverage the text modality to generate coarse representation of the $K$ detected sources in the mixture. Later, by exploiting the multimodal grounding of AudioCLIP, we extract categorical au + +dio and visual features from multi-source representation using text guidance (Sec. 3.3). Rather than aligning audio-visual mixture features as prior work, we introduce the categorical audio-visual feature alignment to further enhance one-to-one correspondence (Sec. 3.4). Finally, cosine similarity of categorical audio features and aligned visual features is used to extract semantic localization maps. + +# 3.1. Preliminaries + +In this section, we formulate the multi-source localization problem, and revisit the AudioCLIP [17] for audio, visual, and text grounding. + +Problem Formulation. Let's assume the dataset contains a total of $N$ classes of sounding events across all videos. Given a video frame and an audio mixture, our objective is to localize $K(K \leq N)$ sound sources within the frame. For each audio-visual mixture pair, we can extract $Y = \{y_{i}\}_{i=1}^{N}$ binary ground truth labels, along with categorical text representations for the on-screen sound sources in the mixture. During training, we are limited to video-level sounding class entities, excluding bounding boxes or mask-level annotations, thus constituting weakly-supervised learning. + +Revisit AudioCLIP for multi-modal grounding. Audio-CLIP [17] learns a joint embedding space of audio, visual, and text triplets through contrastive learning. Let's consider the dataset $\mathcal{D} = \{a,v,t\}_{i = 1}^{Z}$ containing audio spectrogram $a\in \mathbb{R}^{M\times F}$ , text $t\in \mathbb{R}^l$ , and image $v\in \mathbb{R}^{C\times H\times W}$ triplets. + +Also, separate audio encoder $(E_{a}:\mathbb{R}^{M\times F}\to \mathbb{R}^{D})$ , visual encoder $(E_{v}:\mathbb{R}^{C\times H\times W}\to \mathbb{R}^{D})$ , and text encoder $(E_{t}:\mathbb{R}^{l}\to \mathbb{R}^{D})$ are used for each modality. Hence, the extracted audio, visual, and text features are represented by $A = E_{a}(a)$ , $V = E_{v}(v)$ , and $T = E_{t}(t)$ , respectively, where $A,V,T\in \mathbb{R}^D$ . Later, InfoNCE contrastive loss $(\mathcal{L}_{\mathrm{cnt}})$ [33] across each modality pair is used for feature grounding given by: + +$$ +\mathcal {L} _ {\mathrm {a c l i p}} = \mathcal {L} _ {\mathrm {c n t}} (A, T) + \mathcal {L} _ {\mathrm {c n t}} (V, T) + \mathcal {L} _ {\mathrm {c n t}} (A, V) \tag {1} +$$ + +This training objective, combined with large-scale pretraining on web-scraped data, establishes a grounded joint embedding space across all three modalities. + +# 3.2. Audio-Visual Class Instance Detection + +The target sounding objects for localization must be present in both the audio mixture and the corresponding video frame. However, audio mixtures may contain noise from off-screen background sources, while video frames may include non-sounding objects. To address these challenges, we first detect the class instances of visible sounding objects in the frame. Then, we utilize this detection to suppress redundant features in the mixture, ensuring that the localization focuses on the relevant sounding objects. + +Token extraction from Uni-modal encoders. Multiple sounding objects can be spatially located in different regions of the reference frame, while the audio mixture may contain the temporally distributed signal of different sources. This makes it challenging to detect multiple class instances with uni-dimensional single-token features generated from AudioCLIP encoders. Instead of a single pooled token representation, we extract numerous patch tokens from audio and visual encoders with simple modifications, such that $\widehat{E}_a: \mathbb{R}^{M \times F} \to \mathbb{R}^{n_a \times D}$ and $\widehat{E}_v: \mathbb{R}^{C \times H \times W} \to \mathbb{R}^{n_v \times D}$ . Here, $n_a = m \times f$ and $n_v = h \times w$ represent the number of audio and visual patch tokens, respectively. Afterwards, we apply linear projectors on these extracted patch tokens, such that $P_a: \mathbb{R}^{n_a \times D} \to \mathbb{R}^{n_a \times D}$ and $P_v: \mathbb{R}^{n_v \times D} \to \mathbb{R}^{n_v \times D}$ . Thus, extracted audio and visual patch tokens are represented as $A' = P_a(\widehat{E}_a(a))$ , $V' = P_v(\widehat{E}_v(v))$ , respectively, where $A' \in \mathbb{R}^{n_a \times D}$ and $V' \in \mathbb{R}^{n_v \times D}$ . Simultaneously, categorical text features $\mathcal{T} = \{e_i^t\}_{i=1}^N \in \mathbb{R}^{N \times D}$ representing all $N$ sound source classes in the dataset are extracted. + +Multi-label class instance detection. With the categorical text features $\mathcal{T}$ , mixture audio $A^{\prime}$ , and visual $V^{\prime}$ patch tokens, we detect class entities present in the mixture. Initially, we extract the mean-pooled audio-visual features $X^{\prime} \in \mathbb{R}^{2D}$ by simple concatenation, and then, apply fusion projector $P_{f}: \mathbb{R}^{2D} \to \mathbb{R}^{D}$ to extract mixture audio-visual features $F_{av} \in \mathbb{R}^{D}$ . Finally, cosine similarity between $N$ class text features $\mathcal{T}$ and $F_{av}$ is used to detect mixture class + +entities $\widehat{Y}\in \mathbb{R}^N$ , given by: + +$$ +\begin{array}{l} F _ {a v} = P _ {f} \left(X ^ {\prime}\right), X ^ {\prime} = \left[ \operatorname {M e a n} \left(A ^ {\prime}\right) \oplus \operatorname {M e a n} \left(V ^ {\prime}\right) \right], \\ \widehat {\quad} \end{array} \tag {2} +$$ + +$$ +\widehat {Y} = \operatorname {s i m} (\mathcal {T}, F _ {a v}) +$$ + +where $[\oplus ]$ denotes the concatenation operator, and $\operatorname {sim}(X,Y)$ denotes cosine similarity between $X$ and $Y$ . We formulate a multi-label classification objective $\mathcal{L}_{\mathrm{mcid}}$ by applying binary cross-entropy loss $\mathcal{L}_{\mathrm{bce}}(\cdot)$ on each prediction using video-level ground-truth labels $Y\in \mathbb{R}^{N}$ as: + +$$ +\mathcal {L} _ {\mathrm {m c i d}} = \sum_ {i = 1} ^ {N} \mathcal {L} _ {\mathrm {b c e}} \left(y _ {i}, \widehat {y} _ {i}\right) \tag {3} +$$ + +In particular, we train audio, visual, and fusion projectors $(P_{a}, P_{v}, P_{f})$ on noisy multi-source data keeping the AudioCLIP backbones frozen. Hence, these projectors assist in extracting relevant foreground features from noisy audio and visual mixture features extracted from AudioCLIP. + +# 3.3. Audio and Visual Conditioning Blocks + +Establishing one-to-one correspondences for each source in multi-source mixtures is particularly challenging, as both audio and visual features contain mixed representations of various sources. Prior works [21, 30] on multi-source localization have attempted to learn audio-visual alignment intrinsically without explicit single-source guidance. However, in the absence of single-source data, these methods struggle to adequately disentangle multi-source features. In contrast, we leverage the categorical text features of $K$ classes present in the mixture as coarse guidance for fine-grained feature disentanglement from multi-source mixtures. We use ground truth weak labels $Y$ of representative audios in the mixture to extract text embedding of $K$ visual sounding sources $\mathcal{T}_K = \{e_k^t\}_{k=1}^K \subseteq \mathcal{T}$ where $e_k^t \in \mathbb{R}^{1 \times D}$ . + +With the categorical text features $\mathcal{T}_K$ of $K$ sources, we independently disentangle the multi-object visual patch tokens $V^{\prime}$ and mixture audio patch tokens $A^\prime$ by using crossattention $\phi^v (\cdot)$ and $\phi^a (\cdot)$ , respectively, as: + +$$ +\begin{array}{l} \begin{array}{c} \widetilde {V} _ {k} = \phi^ {v} \left(V ^ {\prime}, e _ {k} ^ {t}, V ^ {\prime}\right), f _ {k} ^ {v} = \operatorname {M e a n} \left(\widetilde {V} _ {k}\right), \\ \sim \end{array} \tag {4} \\ \tilde {f} _ {k} ^ {v} = P _ {v c} \left(f _ {k} ^ {v}\right), \forall k = \{1, \dots , K \} \\ \end{array} +$$ + +$$ +\begin{array}{l} \underset {\sim} {\widetilde {A}} _ {k} = \phi^ {a} \left(A ^ {\prime}, e _ {k} ^ {t}, A ^ {\prime}\right), f _ {k} ^ {a} = \operatorname {M e a n} \left(\widetilde {A} _ {k}\right), (5) \\ \hat {f} _ {k} ^ {a} = P _ {a c} \left(f _ {k} ^ {a}\right), \forall k = \{1, \dots , K \} \\ \phi (A, B, C) = \left(\frac {A B ^ {\top}}{\| A \| \| B \|}\right) \odot C (6) \\ \end{array} +$$ + +where $\widetilde{f}_k^a,\widetilde{f}_k^v,f^a,f^v\in \mathbb{R}^{1\times D},\widetilde{V}_k\in \mathbb{R}^{n_v\times D},\widetilde{A}_k\in$ $\mathbb{R}^{n_a\times D}$ , and $P_{ac},P_{vc}:\mathbb{R}^D\to \mathbb{R}^D$ are linear audio and visual projectors on conditioned mean features, respectively, and $\odot$ is the Hadamard product. + +
MethodVGGSound-SingleVGGSound-InstrumentsMUSIC-Solo
AP(%)IoU@0.5(%)AUC(%)AP(%)IoU@0.3(%)AUC(%)AP(%)IoU@0.5(%)AUC(%)
Current SOTA Based Methods
OTS [4] ECCV1838.943.243.750.956.938.476.257.848.9
CoarsetoFine [36] ECCV2039.443.144.851.157.539.176.358.149.7
LVS [7] CVPR2139.744.645.351.557.939.476.157.149.3
EZ-VSL [28] ECCV2240.645.147.252.358.839.978.458.751.1
Mix-and-Localize [21] CVPR2241.845.647.453.859.341.578.959.952.7
DSOL [20] NeurIPS20-46.847.9-61.643.7-62.754.6
MarginNCE [35] ICASSP2342.546.948.356.961.844.283.163.355.2
FNAC [43] CVPR2343.348.449.157.263.245.384.664.556.4
AVGN [30] CVPR2344.149.649.559.364.746.185.465.856.9
Alignment [42] ICCV2345.350.850.260.465.848.786.166.457.2
CLIP-Based Baseline Methods
AudioCLIP [17] ICASSP2242.847.448.558.362.745.283.863.155.7
Wav2CLIP [46] ICASSP2239.343.644.753.857.242.078.959.951.2
T-VSL (Ours)48.153.752.964.669.551.488.268.560.1
+ +Table 1. Performance comparison of single-source localization on VGGSound-single, VGGSound-Instruments, and MUSIC-Solo datasets. For the fair comparison, we use pre-trained AudioCLIP audio and image encoders for baseline methods. + +Additionally, to guide extracting categorical audio $(\widetilde{A}_k)$ and visual $(\widetilde{V}_k)$ patch tokens, we introduce coarse supervision on the output of projected mean patch tokens $\widetilde{f}_k^a,\widetilde{f}_k^v$ using $N$ -class text embedding $\mathcal{T}\in \mathbb{R}^{N\times D}$ . Hence, the class conditioning loss $(\mathcal{L}_{\mathrm{cls}})$ is given by: + +$$ +\mathbf {e} _ {k} ^ {v} = \operatorname {s i m} \left(\mathcal {T}, \widetilde {f _ {k} ^ {v}}\right); \mathbf {e} _ {k} ^ {a} = \operatorname {s i m} \left(\mathcal {T}, \widetilde {f _ {k} ^ {a}}\right) +$$ + +$$ +\mathcal {L} _ {\mathrm {c l s}} = \sum_ {k = 1} ^ {K} \mathcal {L} _ {\mathrm {c e}} \left(\mathbf {e} _ {k} ^ {v}, \mathbf {h} _ {k}\right) + \mathcal {L} _ {\mathrm {c e}} \left(\mathbf {e} _ {k} ^ {a}, \mathbf {h} _ {k}\right) \tag {7} +$$ + +where $\mathbf{h}_k\in \mathbb{R}^N$ denotes an one-hot encoding of the class label, and $\mathbf{e}_k^v\in \mathbb{R}^N$ , $\mathbf{e}_k^a\in \mathbb{R}^N$ are visual and audio class predictions for each $k^{th}$ source, respectively. + +# 3.4. Audio-Visual Correspondence Block + +While text representation provides coarse guidance for disentangling fine-grained spatio-temporal audio and visual features, it has inherent limitations in practice. For example, multiple instances of a sounding object class can appear in the visual frame, including silent instances, which are difficult to distinguish using simple text class representation. To address this challenge, we introduce an audio-visual correspondence block to further refine the alignment of audio and visual features based on extracted categorical features. + +With the disentangled categorical audio $(\widetilde{A}_k)$ and visual $(\widetilde{V}_k)$ patch tokens, we apply cross-attention $\phi^{av}(\cdot)$ to enhance audio-visual patch alignment as: + +$$ +g _ {k} ^ {a} = \operatorname {M e a n} \left(\widetilde {A} _ {k}\right), \widehat {g} _ {k} ^ {a} = P _ {a v} \left(g _ {k} ^ {a}\right) +$$ + +$$ +\widehat {V} _ {k} = \phi^ {a v} \left(\widetilde {V} _ {k}, \widehat {g} _ {k} ^ {a}, \widetilde {V} _ {k}\right), g _ {k} ^ {v} = \operatorname {M e a n} \left(\widehat {V} _ {k}\right) \tag {8} +$$ + +$$ +\widehat {g} _ {k} ^ {v} = P _ {v a} \left(g _ {k} ^ {v}\right), \forall k = \{1, \dots , K \} +$$ + +where $P_{av}, P_{va}: \mathbb{R}^D \to \mathbb{R}^D$ are linear projectors. Afterwards, to correspond between projected mean categorical + +audio and visual features $\widehat{g}_k^a,\widehat{g}_k^v\in \mathbb{R}^D$ , we apply audiovisual contrastive loss $(\mathcal{L}_{\mathrm{av}})$ given by: + +$$ +\mathcal {L} _ {\mathrm {a v}} = \mathcal {L} _ {\mathrm {c n t}} \left(\widehat {g} _ {k} ^ {v}, \widehat {g} _ {k} ^ {a}\right) \tag {9} +$$ + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {a v}} + \mathcal {L} _ {\text {c l s}} + \mathcal {L} _ {\text {m c i d}} +$$ + +where $\mathcal{L}_{\mathrm{total}}$ is the total accumulated loss. + +Inference: During inference, we estimate cosine similarity across projected mean audio feature $\widehat{g}_k^a\in \mathbb{R}^D$ and aligned visual patch tokens $(\widehat{V}_k = \{\widehat{v}_k^j\}_{j = 1}^{n_v},\widehat{v}_k^j\in \mathbb{R}^D)$ . Finally, after reshaping and up-scaling through bilinear interpolation, we generate $K$ -source heat-maps $\mathbf{H}\in \mathbb{R}^{K\times H\times W}$ . + +# 4. Results and Discussions + +# 4.1. Experimental setup + +Datasets. We used MUSIC, VGGSound-Instruments, and VGGSound datasets for the performance evaluation1. MUSIC [49] dataset contains 445 solo music videos of 11 instruments and 142 duet music videos of 8 instruments from YouTube. Following prior works [30], we use MUSIC-Solo for single-source and MUSIC-Duet for multi-source localization. From MUSIC-Solo, we use 350 videos for training and remaining 95 for evaluation. For MUSIC-Duet, we use 120 videos for training and remaining 22 videos for evaluation. VGGSound-Instruments [21] is a subset of VGGSound dataset [6] containing around 32k video clips from 37 music instruments of 10s duration. Apart from the musical instrument datasets, we gather around 150k videos of 10s duration from original VGGSound dataset [6] representing single-source sounds of 221 categories, such as animals, people, nature, instruments, etc. For the single-source evaluation on VGGSound, we use the full test + +
MethodVGGSound-DuetVGGSound-InstrumentsMUSIC-Duet
CAP(%)CIoU@0.3(%)AUC(%)CAP(%)CIoU@0.1(%)AUC(%)CAP(%)CIoU@0.3(%)AUC(%)
Current SOTA Single-Source Methods
OTS [4] ECCV1823.926.427.528.777.818.251.229.424.3
CoarseToFine [36] ECCV20-28.427.9-78.819.7-31.824.1
LVS [7] CVPR21-31.830.1-81.123.2-33.126.9
EZ-VSL [28] ECCV22-32.430.6-80.322.9-34.326.7
MarginNCE [35] ICASSP2327.133.231.233.582.324.755.136.128.5
FNAC [43] CVPR2329.835.133.435.484.126.857.337.930.6
Alignment [42] ICCV2330.435.833.935.684.927.157.838.330.9
Current SOTA Multi-Source Methods
Mix-and-Localize [21] CVPR2229.535.634.136.384.526.757.438.130.7
DSOL [20] NeurIPS20-36.934.6-85.928.4-38.831.3
AVGN [30] CVPR2331.937.835.438.186.728.859.239.332.4
CLIP-Based Baseline Methods
AudioCLIP [17] ICASSP2228.434.932.834.783.825.956.136.929.2
Wav2CLIP [46] ICASSP2226.331.428.931.280.322.252.632.227.2
T-VSL (ours)35.740.137.941.889.631.562.943.235.9
+ +Table 2. Performance comparison of multi-source localization on VGGSound-Duet, VGGSound-Instruments, and MUSIC-Duet datasets. For the fair comparison, we use pre-trained AudioCLIP audio and image encoders for baseline methods. + +set of 5158 videos, while on VGGSound-Instruments, we use 446 videos following prior work [21, 30]. For the multi-source training and evaluation in both VGGSound and VGGSound-Instruments, we randomly concatenate two frames from different sounding sources to produce the multi-source input image with a size of $448 \times 224$ , and mix the single-source audios, following prior work [21, 30]. + +Evaluation Metrics. For evaluating single-source localization performance, we use Intersection over Union (IoU), average precision (AP), and Area Under Curve (AUC) following prior work [21, 30]. For the multi-source localization evaluation, we use Class-aware Average precision (CAP), Class-aware IoU (CIoU), and Area Under Curve (AUC) following prior work [21, 30]. Similarly, for fair comparison with the prior work [30], we use the same thresholds in these metrics: we use IoU@0.5 and CIoU@0.3 for the MUSIC-Solo and MUSIC-Duet datasets, IoU@0.3 and IoU@0.5 for single-source VGGSound-Instruments and VGGSound datasets, and CIoU@0.1 and CIoU@0.3 for multi-source VGGSound-Instruments and VGGSound. + +Implementation details. We use the AudioCLIP [17] pretrained encoders containing ResNet-50 [19] image encoder, ESResNeXt [16] audio encoder, and transformer [38] based text encoder. For the fair comparison, we reproduce the result of existing methods with AudioCLIP image and audio encoders, instead of using separately pre-trained encoders. We use $224 \times 224$ resolution for the single-source input image, $D = 1024$ , $n_v = 49$ for $7 \times 7(h \times w)$ spatial maps generated from the visual encoder, and $n_a = 60$ for $10 \times 6(m \times f)$ time-frequency map generated from the audio encoder. We use Adam optimizer [24] with a batch size of 256 and with a learning rate of $1e - 4$ . + +CLIP-Based Baseline methods. We adopt several CLIP-based baseline methods for the sound source localization + +task, which are described as follows: + +- AudioCLIP [17]: We fine-tune pre-trained image and audio encoders of AudioCLIP on the target dataset. We estimate the localization heatmap for the sounding sources using the similar cosine similarity across patch tokens. +- Wav2CLIP [46]: Similar to AudioCLIP baseline, we experiment with Wav2CLIP audio and image encoders to extract localization heatmaps of sounding sources. + +# 4.2. Comparison to SOTA Baselines + +Single-source localization: We present the quantitative results on single-source localization in Tab. 1. We apply the AudioCLIP image and audio encoders in all baseline methods for the fair comparison with our methods. We observe some performance improvements with several SOTA methods [30, 42, 43] over the AudioCLIP baseline. However, we also notice some reduction in performance from the AudioCLIP baseline in several methods [4, 21, 28]. We hypothesize that these can be due to the alterations of AudioCLIP grounding with the modified contrastive loss. In contrast, by utilizing all three modalities for sound source localization, the proposed T-VSL achieves the best performance over all CLIP-based baselines as well as SOTA methods. T-VSL outperforms the current SOTA single source localization method Alignment [42] by $+2.9$ , $+3.7$ , $+2.1$ higher IoUs on VGGSound-Single, VGGSound-Instruments, and MUSIC-solo datasets, respectively. Moreover, T-VSL achieves $+6.3$ , $+6.8$ , and $+5.4$ higher IoUs over the AudioCLIP baseline [32]. We hypothesize that the enhanced refinement of noisy audio and visual features with text guidance in T-VSL effectively contributes to performance improvement. + +Multi-source localization: We present performance comparison in multi-source sound localization in Tab. 2. As be + +![](images/2e149157b3f9d2717826ed51b1b01f0022c525cc0a32bfcb2cd9c2f5ca105fd6.jpg) +Figure 3. We present qualitative comparisons on challenging multi-source localization with SOTA single and multi-source baseline methods. Here, blue color represents high-attention values to the sounding object, and red color represents low-attention values. The proposed T-VSL can selectively isolate the sounding regions from the background and generates more precise localization maps for sounding sources. + +fore, we apply the AudioCLIP image and audio encoders in all baseline methods for the fair comparison with our methods. We observe some performance improvements, mostly in SOTA multi-source methods [20, 21, 30], over the AudioCLIP baseline. In most single source methods, we observe performance loss in multi-source cases over the baseline. The proposed T-VSL achieves the best performance that outperforms the SOTA multi-source baseline AVGN [30] by $+2.3$ , $+2.9$ , and $+3.9$ CIoUs on multi-source VGGSound-Duet, VGGSound-Instruments, and MUSIC-Duet datasets, respectively. Also, T-VSL achieves $+5.2$ , $+5.8$ , and $+6.3$ higher CIoUs over the AudioCLIP baseline. These significant performance gains on multi-source benchmarks demonstrate the effectiveness of the proposed T-VSL in disentangling multi-source mixtures. + +Qualitative comparisons: In addition, we present qualitative comparisons on multi-source localization in Fig. 3 among single-source baseline Alignment [42], self-supervised multi-source baseline Mix-and-Localize [21], weakly-supervised multi-source baseline AVGN [30], and proposed T-VSL. Single-source SOTA baseline Alignment [42] struggles in multi-source localization due to the lack of audio-visual correspondence, mostly in the presence of noisy audios and silent visual objects. However, without any guidance on fine-grained source localization, multi-source baselines [21, 30] underperform on localizing challenging multi-source mixtures. Notably, our method generates superior localization maps, which shows the effectiveness of T-VSL in multi-source feature disentanglement. + +# 4.3. Ablation on T-VSL building blocks + +We present the ablation study of different building blocks of T-VSL in Tab. 3 under both single-source and multisource settings. The vanilla baseline contains AudioCLIP image and audio encoders that directly operate on the input data. By integrating the audio-visual correspondence (AVC) block on extracted mixture patch tokens, we notice considerable performance increase in single-source (by $+1.1$ AP and $+0.7$ IoU@0.5) and multi-source localization (by $+1.3$ CAP and $+1.4$ CIoU@0.3). By integrating the audio and image conditioning blocks following audio-visual instance detection for separating fine-grained class features, we observe additional improvements of $+1.8$ AP and $+2.3$ AP in single-source, and $+3.1$ CAP and $+3.6$ AP in multisource localization, respectively. Finally, the combination of all building blocks results in the best performance that improves the baseline by $+5.3$ AP, and 6.3 IoU@0.5 in single-source and by $+7.3$ CAP and $+5.2$ CIoU@0.3 in multi-source localization. This further demonstrates the effective disentanglement of audio-visual features facilitated by intermediate text guidance. + +# 4.4. Zero-shot transfer across datasets + +We can perform zero-shot transfer across datasets with proposed T-VSL by simply replacing the $N$ -class text prompts in Fig. 2. Since other weakly-supervised methods cannot perform zero-shot transfer, we present comparisons with self-supervised Mix-and-Localize [21], Alignment [42], and FNAC [43] methods in Tab. 4. To ensure a fair comparison, we conducted uniform training with $50\% - 50\%$ class + +
Audio Cond.Image Cond.AVCVGGSound-SingleVGGSound-Duet
AP(%)IoU@0.5(%)CAP(%)CIoU@0.3(%)
XXX42.847.428.434.9
XX✓43.948.129.736.3
✓X✓45.750.932.838.1
X✓✓46.251.433.338.6
✓✓✓48.153.735.740.1
+ +Table 3. Ablation study on audio conditioning block, image-conditioning block, and audio-visual correspondence block (AVC) of T-VSL in single-source and multi-source localization tasks. + +
Target DatasetProposed T-VSLMix-and-Localize [21]Alignment [42]FNAC [43]
MUSIC-Duet55.850.951.751.2
MUSIC-Solo80.575.177.677.4
VGGSound-Instr.81.776.876.576.1
+ +Table 4. Ablation on zero-shot transfer across single and multi-source datasets with $50\% -50\%$ class split. We report IoU@0.5(%) for MUSIC-Solo, CIoU@0.3(%) for MUSIC-Duet, and CIoU@0.1(%) for VGGSound-Instruments. We use same AudioCLIP encoders for all baselines. + +split for all methods with same AudioCLIP encoders. We note that our method achieves significantly better performance on zero-shot transfer to target datasets than these compared approaches. In particular, we achieve $+4.9$ , $+5.4$ , $+4.9$ improvements on target datasets than the multi-source Mix-and-Localize [21]. Also, we achieve $+4.1$ , $+2.9$ , and $+5.2$ higher scores than the SOTA single source baseline, Alignment [42]. This result suggests that our T-VSL has superior generalization capabilities for zero-shot localization tasks in unseen classes. + +# 4.5. Robustness to a higher number of sources + +The proposed T-VSL is not limited by the number of sources present in training mixtures unlike prior work [21]. We present comparative analysis with SOTA multi-source AVGN [30], and self-supervised single source methods FNAC [43] and Alignment [42] in Tab. 5 on robustness to higher number of test sources in VGGSound dataset. We train all methods on the VGGSound-Single dataset with same encoders for a fair comparison. We note that our method consistently maintains significantly higher CIOU score irrespective of test scenarios, and also achieves notably smaller relative performance drop for a large number of sources in mixtures. These results further demonstrate the robustness of the proposed T-VSL in disentangling multi-source mixtures. + +# 4.6. Use of learnable text prompts + +To disentangle multi-source mixtures, we primarily use the class label of each sounding source as text prompts (e.g. dog barking, snake hissing). Inspired by the conditional text prompt learning in [50], we study the use of learnable prompts along with class label representations. Instead of + +
Test Source No.Proposed T-VSLAVGN [30]Alignment [42]FNAC [43]
239.636.835.434.1
335.730.828.927.1
429.522.419.318.8
+ +Table 5. Robustness to higher number of sources in VGGSound dataset when trained with single-source data. CIoU@0.3(%) score is reported. Same AudioCLIP encoders are used for all baselines. + +
Prompt LengthVGGSound-SingleVGGSound-Duet
AP(%)IoU@0.5(%)CAP(%)CIoU@0.3(%)
048.153.735.740.1
248.554.136.040.4
449.254.636.841.3
850.355.237.142.5
1650.255.037.442.9
3249.754.537.142.7
+ +Table 6. Ablation study on the effect of learnable text prompts on VGGSound-Single and VGGSound-Duet datasets. Integration of learnable prompts results in noticeable performance improvements in both single and multi-source localization. + +only using class label representation, a set of learnable embedding is integrated with class label embedding that provides additional flexibility to adapt the text prompt to target objective. We analyze the sound source localization performance with different length of learnable prompts. The performances are reported in Table 6. We notice considerable performance improvements by optimizing learnable prompts in both single and multi-source cases. However, such prompts limit zero-shot use cases to unseen classes. + +# 5. Conclusion + +In this paper, we propose T-VSL, a novel text-guided multi-source visual sound source localization framework that can disentangle one-to-one audio-visual correspondence from multi-source mixtures. We leverage the text modality to guide fine-grained feature separation and localization, from noisy audio and visual features of mixtures, by exploiting the joint embedding space of AudioCLIP. In comparison with SOTA multi-source baselines, our method shows superior zero-shot transfer across datasets. Moreover, our method demonstrates notable robustness on challenging test mixtures with higher number of sources than training scenarios. In both single-source and multi-source localization, our method largely outperforms existing weakly-supervised and self-supervised baselines on three benchmark datasets. + +# Acknowledgements + +This research was supported in part by ONR Minerva program, iMAGiNE - the Intelligent Machine Engineering Consortium at UT Austin, and a UT Cockrell School of Engineering Doctoral Fellowship. + +# References + +[1] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In Proceedings of European Conference on Computer Vision (ECCV), pages 208–224, 2020. 1, 2 +[2] Muhammad Ali and Salman Khan. Clip-decoder: Zeroshot multilabel classification using multimodal clip aligned representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4675-4679, 2023. 2 +[3] Tayfun Alpay, Sven Magg, Philipp Broze, and Daniel Speck. Multimodal video retrieval with clip: a user study. Information Retrieval Journal, 26(1-2):6, 2023. 2 +[4] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proceedings of the European Conference on Computer Vision (ECCV), pages 435-451, 2018. 5, 6 +[5] Maria A Bravo, Sudhanshu Mittal, and Thomas Brox. Localized vision-language matching for open-vocabulary object detection. In DAGM German Conference on Pattern Recognition, pages 393-408. Springer, 2022. 2 +[6] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 2, 5 +[7] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16867-16876, 2021. 1, 2, 5, 6 +[8] Chaorui Deng, Qi Chen, Pengda Qin, Da Chen, and Qi Wu. Prompt switch: Efficient clip adaptation for text-video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15648-15658, 2023. 2 +[9] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary panoptic segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022. 2 +[10] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 2 +[11] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model clip. In Proceedings of the AAAI conference on artificial intelligence, pages 6568-6576, 2022. 2 +[12] John W Fisher III, Trevor Darrell, William Freeman, and Paul Viola. Learning joint statistical models for audio-visual fusion and segregation. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2000. 2 +[13] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2 +[14] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models. In Proceed- + +ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19338-19347, 2023. 2 +[15] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xupeng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 746-754, 2023. 2 +[16] Andrey Guzhov, Federico Raue, Jörn Hees, and Andreas Dengel. Esresne (x) t-fbsp: Learning robust time-frequency transformation of audio. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021. 6 +[17] Andrey Guzhov, Federico Raue, Jorn Hees, and Andreas Dengel. Audioclip: Extending clip to image, text and audio. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 976-980. IEEE, 2022. 2, 3, 5, 6, 1 +[18] John Hershey and Javier Movellan. Audio vision: Using audio-visual synchrony to locate sounds. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 1999. 2 +[19] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9248-9257, 2019. 1, 2, 6 +[20] Di Hu, Rui Qian, Minyue Jiang, Xiao Tan, Shilei Wen, Errui Ding, Weiyao Lin, and Dejing Dou. Discriminative sounding objects localization via self-supervised audiovisual matching. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 10077-10087, 2020. 1, 2, 5, 6, 7 +[21] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10483-10492, 2022. 1, 2, 4, 5, 6, 7, 8 +[22] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 2 +[23] Einat Kidron, Yoav Y Schechner, and Michael Elad. Pixels that sound. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2005. 2 +[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[25] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2 +[26] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In International Conference on Machine Learning, pages 23033-23044. PMLR, 2023. 2 +[27] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh + +Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In European Conference on Computer Vision, pages 728-755. Springer, 2022. 2 +[28] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In Proceedings of European Conference on Computer Vision (ECCV), page 218-234, 2022. 1, 2, 5, 6 +[29] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2 +[30] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10565–10574, 2023. 1, 2, 4, 5, 6, 7, 8 +[31] Jishnu Mukhoti, Tsung-Yu Lin, Omid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19413-19423, 2023. 2 +[32] Jishnu Mukhoti, Tsung-Yu Lin, Amid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19413-19423, 2023. 6 +[33] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4 +[34] Sooyoung Park, Arda Senocak, and Joon Son Chung. Can clip help sound source localization? arXiv preprint arXiv:2311.04066, 2023. 3, 1 +[35] Sooyoung Park, Arda Senocak, and Joon Son Chung. Marginnce: Robust sound localization with a negative margin. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 5, 6 +[36] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In Proceedings of European Conference on Computer Vision (ECCV), pages 292-308, 2020. 1, 2, 5, 6 +[37] Jie Qin, Jie Wu, Pengxiang Yan, Ming Li, Ren Yuxi, Xuefeng Xiao, Yitong Wang, Rui Wang, Shilei Wen, Xin Pan, et al. Freeseg: Unified, universal and open-vocabulary image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19446-19455, 2023. 2 +[38] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 6 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 + +[40] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4358-4366, 2018. 1, 2 +[41] Arda Senocak, Hyeonggon Ryu, Junsik Kim, and In So Kweon. Learning sound localization better from semantically similar samples. In Proceedings of IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022. 1, 2 +[42] Arda Senocak, Hyeonggon Ryu, Junsik Kim, Tae-Hyun Oh, Hanspeter Pfister, and Joon Son Chung. Sound source localization is all about cross-modal alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7777-7787, 2023. 1, 5, 6, 7, 8 +[43] Weixuan Sun, Jiayi Zhang, Jianyuan Wang, Zheyuan Liu, Yiran Zhong, Tianpeng Feng, Yandong Guo, Yanhao Zhang, and Nick Barnes. Learning audio-visual source localization via false negative aware contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6420–6429, 2023. 1, 5, 6, 7, 8 +[44] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1802-1812, 2023. 2 +[45] Tao Wang. Learning to detect and segment for open vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7051-7060, 2023. 2 +[46] Ho-Hsiang Wu, Prem Seetharaman, Kundan Kumar, and Juan Pablo Bello. Wav2clip: Learning robust audio representations from clip. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4563-4567. IEEE, 2022. 3, 5, 6 +[47] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2935-2944, 2023. 2 +[48] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 2 +[49] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In Proceedings of the European Conference on Computer Vision (ECCV), pages 570-586, 2018. 1, 2, 5 +[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 8 \ No newline at end of file diff --git a/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/images.zip b/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7a7fe297e3e0b902a5ed1306d610200bfda32fb4 --- /dev/null +++ b/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd27b21feca3729ec103a50e4ed3bd4e2cdf105abd1833124c004bdf1cec2e8c +size 631400 diff --git a/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/layout.json b/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..36f06d2b913d344d8a979e53c33b6e6975a085b4 --- /dev/null +++ b/2024/T-VSL_ Text-Guided Visual Sound Source Localization in Mixtures/layout.json @@ -0,0 +1,8989 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 90, + 103, + 503, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 103, + 503, + 119 + ], + "spans": [ + { + "bbox": [ + 90, + 103, + 503, + 119 + ], + "type": "text", + "content": "T-VSL: Text-Guided Visual Sound Source Localization in Mixtures" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 144, + 202, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 144, + 202, + 171 + ], + "spans": [ + { + "bbox": [ + 59, + 144, + 202, + 171 + ], + "type": "text", + "content": "Tanvir Mahmud \nUniversity of Texas at Austin" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 225, + 144, + 365, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 144, + 365, + 171 + ], + "spans": [ + { + "bbox": [ + 225, + 144, + 365, + 171 + ], + "type": "text", + "content": "Yapeng Tian \nUniversity of Texas at Dallas" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 388, + 144, + 529, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 144, + 529, + 171 + ], + "spans": [ + { + "bbox": [ + 388, + 144, + 529, + 171 + ], + "type": "text", + "content": "Diana Marculescu \nUniversity of Texas at Austin" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 199, + 191, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 191, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 191, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 224, + 289, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 224, + 289, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 289, + 583 + ], + "type": "text", + "content": "Visual sound source localization poses a significant challenge in identifying the semantic region of each sounding source within a video. Existing self-supervised and weakly supervised source localization methods struggle to accurately distinguish the semantic regions of each sounding object, particularly in multi-source mixtures. These methods often rely on audio-visual correspondence as guidance, which can lead to substantial performance drops in complex multi-source localization scenarios. The lack of access to individual source sounds in multi-source mixtures during training exacerbates the difficulty of learning effective audio-visual correspondence for localization. To address this limitation, in this paper, we propose incorporating the text modality as an intermediate feature guide using tri-modal joint embedding models (e.g., AudioCLIP) to disentangle the semantic audio-visual source correspondence in multi-source mixtures. Our framework, dubbed T-VSL, begins by predicting the class of sounding entities in mixtures. Subsequently, the textual representation of each sounding source is employed as guidance to disentangle fine-grained audio-visual source correspondence from multi-source mixtures, leveraging the tri-modal AudioCLIP embedding. This approach enables our framework to handle a flexible number of sources and exhibits promising zero-shot transferability to unseen classes during test time. Extensive experiments conducted on the MUSIC, VGGSound, and VGGSound-Instruments datasets demonstrate significant performance improvements over state-of-the-art methods. Code is released at https://github.com/encyac-group/T-VSL/tree/main." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 597, + 127, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 127, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 127, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": "While observing a conversation between two individuals, we can easily associate the audio signal with the corresponding speaking person in the visual scene. This remarkable ability to perceive audio-visual correspondence stems from our extensive exposure to both single-source sounds and multi-source mixtures in everyday life. Inspired by this human capability, significant research efforts [21, 30, 42, 43] have been dedicated to developing vi" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 199, + 390, + 258 + ], + "blocks": [ + { + "bbox": [ + 307, + 199, + 390, + 258 + ], + "lines": [ + { + "bbox": [ + 307, + 199, + 390, + 258 + ], + "spans": [ + { + "bbox": [ + 307, + 199, + 390, + 258 + ], + "type": "image", + "image_path": "20e68c70fc09ecb7301d28ed1abfe6bc1c6571d92f0c2adb001f53931c8dacef.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 390, + 199, + 464, + 258 + ], + "blocks": [ + { + "bbox": [ + 390, + 199, + 464, + 258 + ], + "lines": [ + { + "bbox": [ + 390, + 199, + 464, + 258 + ], + "spans": [ + { + "bbox": [ + 390, + 199, + 464, + 258 + ], + "type": "image", + "image_path": "0ef5b245a37750e3ce6a6007a28fcf790de95eb2e2866e5caa80e6fb2ee6df3b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 465, + 199, + 542, + 258 + ], + "blocks": [ + { + "bbox": [ + 465, + 199, + 542, + 258 + ], + "lines": [ + { + "bbox": [ + 465, + 199, + 542, + 258 + ], + "spans": [ + { + "bbox": [ + 465, + 199, + 542, + 258 + ], + "type": "image", + "image_path": "9a80375f90a17aadea0741648bba785c016c836c1b6789f658918fb35096ae4e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 259, + 390, + 321 + ], + "blocks": [ + { + "bbox": [ + 307, + 259, + 390, + 321 + ], + "lines": [ + { + "bbox": [ + 307, + 259, + 390, + 321 + ], + "spans": [ + { + "bbox": [ + 307, + 259, + 390, + 321 + ], + "type": "image", + "image_path": "41ebede698abbc14ac2ff7d9a8bcf160b57613ddfd316b9ea0c3eee9723b8d56.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 327, + 545, + 381 + ], + "lines": [ + { + "bbox": [ + 305, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 545, + 381 + ], + "type": "text", + "content": "Figure 1. Comparison of our T-VSL with state-of-the-art methods on single-source (Top Row) and multi-source (Bottom Row) sound localization on VGGSound Sources [7], VGGSoundInstruments [21], and MUSIC [49] benchmark datasets. We use the same setup for the fair comparison." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 390, + 259, + 464, + 321 + ], + "blocks": [ + { + "bbox": [ + 390, + 259, + 464, + 321 + ], + "lines": [ + { + "bbox": [ + 390, + 259, + 464, + 321 + ], + "spans": [ + { + "bbox": [ + 390, + 259, + 464, + 321 + ], + "type": "image", + "image_path": "c8a46b3e47749300327f77158d28188652d7bfbfbe0ed56454aad3a5b525b11b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 465, + 259, + 542, + 321 + ], + "blocks": [ + { + "bbox": [ + 465, + 259, + 542, + 321 + ], + "lines": [ + { + "bbox": [ + 465, + 259, + 542, + 321 + ], + "spans": [ + { + "bbox": [ + 465, + 259, + 542, + 321 + ], + "type": "image", + "image_path": "6bdd148357b304515ec2d9b67b73d2cc2bb78dd544c670cc6d6ce73fed5dcde7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 387, + 539, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 387, + 539, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 387, + 539, + 399 + ], + "type": "text", + "content": "sual sound source localization approaches in recent years." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 400, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 545, + 628 + ], + "type": "text", + "content": "Visual sound source localization aims to locate the visual regions representing sound sources present in a video. Earlier methods on single-source localization [1, 7, 19, 28, 29, 40, 41] mostly used audio-visual correspondence as guidance for localizing sounding objects in each frame. These approaches, developed for single sound source localization, suffer from significant performance drop in localizing multi-source mixtures. Multi-source localization from mixtures is very challenging, as the model must learn to distinguish each sounding source within a mixture and establish their cross-modal relationships, without access to individual source sounds. Earlier methods [20, 21, 30, 36] on multi-source localization attempt to infer fine-grained cross-modal associations directly from noisy multi-source mixtures, often leading to sub-optimal performance in practice for learning improper source correspondence. In this paper, we tackle the problem by introducing a unified solution for localizing visual sound sources in both single and multi-source mixtures." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "The primary challenge of this task stems from the difficulty in disentangling the cross-modal correspondence of each sounding object from natural mixtures, given the absence of one-to-one corresponding single-source pairs. Additionally, the presence of silent visual objects and noise from invisible background sources further complicates this alignment process. To solve this problem, our key idea is to" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26742" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "leverage the text modality as a coarse supervision for disentangling categorical audio-visual correspondence in natural mixtures, which differentiates our method from existing works. Unlike audio and visual modalities that may contain significant noises from different sources present in mixtures, textual representation can explicitly discriminate across multiple sources. However, a critical obstacle to utilizing text guidance in visual sound source localization is grounding fine-grained audio and visual features with their textual representation. Later, CLIP [39] introduced learning visual language grounding from web-scraped 400M imagetext pairs. Recently, AudioCLIP [17] introduced the audio modality in the existing CLIP architecture, thereby generating tri-modal feature grounding through large-scale training. Our approach capitalizes on the tri-modal joint embedding space of AudioCLIP to disentangle one-to-one audio-visual correspondence in natural mixtures." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 601 + ], + "type": "text", + "content": "To this end, in this paper, we propose a novel text-guided multi-source localization framework that can discover fine-grained audio-visual semantic correspondence in multisource mixtures. Initially, we detect the class instances of visual sounding objects in the frame using the noisy mixture features from AudioCLIP image and audio encoders. Then, the text representation of each detected sounding source class instance is extracted with AudioCLIP text encoder, which serves as a coarse guidance for audio and visual feature separation. Specifically, we disentangle the categorical audio and visual features of each sounding source using the coarse text-guidance through additional audio and image conditioning blocks, respectively. Finally, the extracted categorical audio-visual semantic features are further aligned through an audio-visual correspondence block for localizing each sounding source. In comparison with prior multisource baselines, our approach can selectively localize the semantic visual region of each class of sounding objects in mixtures of varying number of sources. Extensive experiments on MUSIC [49], VGGSound [6], and VGGSound-Instruments [21] demonstrate significant performance improvements compared to existing single and multi-source baselines. Moreover, our method shows promising zero-shot transferability on unseen classes during test time. In addition, thorough ablation studies and qualitative analysis vividly showcase the effectiveness of the proposed framework." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 604, + 280, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 604, + 280, + 615 + ], + "spans": [ + { + "bbox": [ + 58, + 604, + 280, + 615 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 618, + 286, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 618, + 286, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 286, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 286, + 653 + ], + "type": "text", + "content": "- We propose a novel text-guided multi-source localization framework to disentangle one-to-one audio-visual correspondence from natural mixtures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 654, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 286, + 689 + ], + "type": "text", + "content": "- Our work is the first to introduce AudioCLIP for utilizing the tri-modal feature grounding from large-scale pretraining in solving multi-source localization problems." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": "- Our proposed framework can operate with a flexible number of sources and shows promising zero-shot perfor" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 73, + 470, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 73, + 470, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 73, + 470, + 83 + ], + "type": "text", + "content": "mance on unseen audio-visual classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 84, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 84, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 306, + 84, + 545, + 119 + ], + "type": "text", + "content": "- Extensive experiments on benchmark datasets clearly demonstrate the superiority of our proposed framework over other state-of-the-art methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 129, + 392, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 129, + 392, + 140 + ], + "spans": [ + { + "bbox": [ + 306, + 129, + 392, + 140 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 149, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 149, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 149, + 545, + 353 + ], + "type": "text", + "content": "Visual Sound Source Localization. Visual sound source localization aims to locate the visual regions of sounding objects in video frames. Prior work explored diverse methods for audio-visual feature alignment, including traditional statistical machine learning approaches [12, 18, 23] to latest deep neural nets [1, 7, 19, 28, 29, 36, 40, 41]. However, most of these self-supervised and weakly-supervised methods are mostly designed for localizing single-source sounds. A two-stream architecture is proposed in Attention10k [40] which incorporates cross-attention to locate sounding sources. Later, motion cues present in videos are incorporated for finer localization [1]. Recently, various contrastive learning based methods are explored in LVS [7], EZ-VSL [28], and SLAVC [29] for single-source localization. Despite their promising performance, these single-source localization methods can hardly discriminate audio-visual correspondence in multi-source mixtures." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 354, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 616 + ], + "type": "text", + "content": "Sounds mostly appear in mixtures in the presence of significant background noises. The main challenges of multisource localization is to disentangle the audio-visual correspondence in mixtures without having access to clean single-source audio-visual pairs. Several recent works have explored multi-source localization with explicit approaches for mixtures [20, 21, 30, 36]. To selectively isolate the silent objects in visual frames, DSOL [20] proposed a two-stage weakly-supervised training method. Mix-and-localize [21] proposed a self-supervised training with contrastive random walk to inherently align each sounding source with its visual representation. Recently, AVGN [30] introduced weakly-supervised grouping of category-aware features with learnable prompts. However, these approaches tackle the audio-visual correspondence in mixtures without any explicit guidance for single-source thereby resulting in sub-optimal performance. In contrast, we leverage text representation of fine-grained sounding sources to disentangle one-to-one audio-visual correspondence from multi-source mixtures. To the best of our knowledge, this work is the first to utilize text representation of sound sources for solving multisource localization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": "Audio, Visual, and Text Grounding. Large-scale pretraining on web-scraped data has been used to generate grounded features across modalities. CLIP [39] first introduced vision-language grounding using 400M imagedtext pairs, which has been extensively studied for zero-shot classification [2, 11, 14, 15, 22, 22, 44], text-video retrieval [3, 8, 25], open-vocabulary segmentation [9, 13, 26, 31, 37, 47, 48] and object detection [5, 10, 27, 45]. Re" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26743" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 281 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 281 + ], + "type": "image", + "image_path": "9bd966ff2e2af2de15ff2660debcb0f5b45a0b47f8d82317f11a42ea62a99279.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 294, + 547, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 294, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 547, + 361 + ], + "type": "text", + "content": "Figure 2. The proposed text-guided visual sound source localization (T-VSL) framework (for " + }, + { + "bbox": [ + 46, + 294, + 547, + 361 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 46, + 294, + 547, + 361 + ], + "type": "text", + "content": "). We use the text modality to disentangle the fine-grained audio-visual correspondence from mixtures. Initially, we detect the audio-visual class instances from multisource mixtures using AudioCLIP joint embedding space. Later, categorical text features of each detected " + }, + { + "bbox": [ + 46, + 294, + 547, + 361 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 294, + 547, + 361 + ], + "type": "text", + "content": " classes are used as coarse guidance in conditioning blocks to extract categorical visual and audio features. Afterwards, cross-modal feature alignment on extracted categorical features is performed in audio-visual correspondence block. Finally, cosine similarity of mean categorical audio features, and aligned visual features are used recursively to extract localization map of each class." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 371, + 289, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 289, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 289, + 492 + ], + "type": "text", + "content": "cently, AudioCLIP [17] and Wav2CLIP [46] introduced additional audio grounding in original CLIP model by learning on image, text, and audio triplets. Very recently, a CLIP-based single source localization method [34] is introduced that attempts replacing the text-conditioning branch of a supervised image segmentation baseline, trained with dense supervision, with audio conditioning. In relation with this concurrent work, our focus lies primarily on weakly supervised sound source localization in multi-source mixtures, leveraging the self-supervised AudioCLIP model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 502, + 104, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 104, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 104, + 514 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": "Given a mixture of audio and a video frame, our objective is to spatially localize each sounding object in the frame. We propose a novel text-guided multi-source localization framework to disentangle the fine-grained audio-visual correspondence, which is illustrated in Fig. 2. To suppress the background noises, we initially detect the common class instances in both audio and visual modality (Sec. 3.2). In particular, we leverage tri-modal AudioCLIP encoders to detect " + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": " visual-sounding class instances from " + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": " classes " + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "inline_equation", + "content": "(K \\leq N)" + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": ". The primary challenge in multi-source localization lies in establishing one-to-one audio-visual correspondences for individual sources without access to single-source samples. To overcome this, we leverage the text modality to generate coarse representation of the " + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": " detected sources in the mixture. Later, by exploiting the multimodal grounding of AudioCLIP, we extract categorical au" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 371, + 547, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 547, + 456 + ], + "type": "text", + "content": "dio and visual features from multi-source representation using text guidance (Sec. 3.3). Rather than aligning audio-visual mixture features as prior work, we introduce the categorical audio-visual feature alignment to further enhance one-to-one correspondence (Sec. 3.4). Finally, cosine similarity of categorical audio features and aligned visual features is used to extract semantic localization maps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 465, + 392, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 465, + 392, + 477 + ], + "spans": [ + { + "bbox": [ + 305, + 465, + 392, + 477 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 485, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 521 + ], + "type": "text", + "content": "In this section, we formulate the multi-source localization problem, and revisit the AudioCLIP [17] for audio, visual, and text grounding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 528, + 548, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 548, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 548, + 648 + ], + "type": "text", + "content": "Problem Formulation. Let's assume the dataset contains a total of " + }, + { + "bbox": [ + 304, + 528, + 548, + 648 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 528, + 548, + 648 + ], + "type": "text", + "content": " classes of sounding events across all videos. Given a video frame and an audio mixture, our objective is to localize " + }, + { + "bbox": [ + 304, + 528, + 548, + 648 + ], + "type": "inline_equation", + "content": "K(K \\leq N)" + }, + { + "bbox": [ + 304, + 528, + 548, + 648 + ], + "type": "text", + "content": " sound sources within the frame. For each audio-visual mixture pair, we can extract " + }, + { + "bbox": [ + 304, + 528, + 548, + 648 + ], + "type": "inline_equation", + "content": "Y = \\{y_{i}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 304, + 528, + 548, + 648 + ], + "type": "text", + "content": " binary ground truth labels, along with categorical text representations for the on-screen sound sources in the mixture. During training, we are limited to video-level sounding class entities, excluding bounding boxes or mask-level annotations, thus constituting weakly-supervised learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Revisit AudioCLIP for multi-modal grounding. Audio-CLIP [17] learns a joint embedding space of audio, visual, and text triplets through contrastive learning. Let's consider the dataset " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{a,v,t\\}_{i = 1}^{Z}" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": " containing audio spectrogram " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "a\\in \\mathbb{R}^{M\\times F}" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": ", text " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "t\\in \\mathbb{R}^l" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": ", and image " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "v\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": " triplets." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26744" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "content": "Also, separate audio encoder " + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "inline_equation", + "content": "(E_{a}:\\mathbb{R}^{M\\times F}\\to \\mathbb{R}^{D})" + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "content": ", visual encoder " + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "inline_equation", + "content": "(E_{v}:\\mathbb{R}^{C\\times H\\times W}\\to \\mathbb{R}^{D})" + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "content": ", and text encoder " + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "inline_equation", + "content": "(E_{t}:\\mathbb{R}^{l}\\to \\mathbb{R}^{D})" + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "content": " are used for each modality. Hence, the extracted audio, visual, and text features are represented by " + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "inline_equation", + "content": "A = E_{a}(a)" + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "inline_equation", + "content": "V = E_{v}(v)" + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "inline_equation", + "content": "T = E_{t}(t)" + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "inline_equation", + "content": "A,V,T\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "content": ". Later, InfoNCE contrastive loss " + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{cnt}})" + }, + { + "bbox": [ + 46, + 71, + 289, + 168 + ], + "type": "text", + "content": " [33] across each modality pair is used for feature grounding given by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 178, + 287, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 178, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 52, + 178, + 287, + 194 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a c l i p}} = \\mathcal {L} _ {\\mathrm {c n t}} (A, T) + \\mathcal {L} _ {\\mathrm {c n t}} (V, T) + \\mathcal {L} _ {\\mathrm {c n t}} (A, V) \\tag {1}", + "image_path": "aa48db8aee20aafd886b3fb2a0c46a2abac727299f902da2a1498115c2e59440.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 203, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 203, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 287, + 239 + ], + "type": "text", + "content": "This training objective, combined with large-scale pretraining on web-scraped data, establishes a grounded joint embedding space across all three modalities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 248, + 248, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 248, + 261 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 248, + 261 + ], + "type": "text", + "content": "3.2. Audio-Visual Class Instance Detection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 268, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 287, + 376 + ], + "type": "text", + "content": "The target sounding objects for localization must be present in both the audio mixture and the corresponding video frame. However, audio mixtures may contain noise from off-screen background sources, while video frames may include non-sounding objects. To address these challenges, we first detect the class instances of visible sounding objects in the frame. Then, we utilize this detection to suppress redundant features in the mixture, ensuring that the localization focuses on the relevant sounding objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "text", + "content": "Token extraction from Uni-modal encoders. Multiple sounding objects can be spatially located in different regions of the reference frame, while the audio mixture may contain the temporally distributed signal of different sources. This makes it challenging to detect multiple class instances with uni-dimensional single-token features generated from AudioCLIP encoders. Instead of a single pooled token representation, we extract numerous patch tokens from audio and visual encoders with simple modifications, such that " + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "inline_equation", + "content": "\\widehat{E}_a: \\mathbb{R}^{M \\times F} \\to \\mathbb{R}^{n_a \\times D}" + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "inline_equation", + "content": "\\widehat{E}_v: \\mathbb{R}^{C \\times H \\times W} \\to \\mathbb{R}^{n_v \\times D}" + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "inline_equation", + "content": "n_a = m \\times f" + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "inline_equation", + "content": "n_v = h \\times w" + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "text", + "content": " represent the number of audio and visual patch tokens, respectively. Afterwards, we apply linear projectors on these extracted patch tokens, such that " + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "inline_equation", + "content": "P_a: \\mathbb{R}^{n_a \\times D} \\to \\mathbb{R}^{n_a \\times D}" + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "inline_equation", + "content": "P_v: \\mathbb{R}^{n_v \\times D} \\to \\mathbb{R}^{n_v \\times D}" + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "text", + "content": ". Thus, extracted audio and visual patch tokens are represented as " + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "inline_equation", + "content": "A' = P_a(\\widehat{E}_a(a))" + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "inline_equation", + "content": "V' = P_v(\\widehat{E}_v(v))" + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "inline_equation", + "content": "A' \\in \\mathbb{R}^{n_a \\times D}" + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "inline_equation", + "content": "V' \\in \\mathbb{R}^{n_v \\times D}" + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "text", + "content": ". Simultaneously, categorical text features " + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{e_i^t\\}_{i=1}^N \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "text", + "content": " representing all " + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 376, + 289, + 616 + ], + "type": "text", + "content": " sound source classes in the dataset are extracted." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "text", + "content": "Multi-label class instance detection. With the categorical text features " + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "text", + "content": ", mixture audio " + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "inline_equation", + "content": "A^{\\prime}" + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "text", + "content": ", and visual " + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "inline_equation", + "content": "V^{\\prime}" + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "text", + "content": " patch tokens, we detect class entities present in the mixture. Initially, we extract the mean-pooled audio-visual features " + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "inline_equation", + "content": "X^{\\prime} \\in \\mathbb{R}^{2D}" + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "text", + "content": " by simple concatenation, and then, apply fusion projector " + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "inline_equation", + "content": "P_{f}: \\mathbb{R}^{2D} \\to \\mathbb{R}^{D}" + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "text", + "content": " to extract mixture audio-visual features " + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "inline_equation", + "content": "F_{av} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "text", + "content": ". Finally, cosine similarity between " + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "text", + "content": " class text features " + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "inline_equation", + "content": "F_{av}" + }, + { + "bbox": [ + 46, + 617, + 289, + 713 + ], + "type": "text", + "content": " is used to detect mixture class" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 70, + 417, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 70, + 417, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 417, + 85 + ], + "type": "text", + "content": "entities " + }, + { + "bbox": [ + 305, + 70, + 417, + 85 + ], + "type": "inline_equation", + "content": "\\widehat{Y}\\in \\mathbb{R}^N" + }, + { + "bbox": [ + 305, + 70, + 417, + 85 + ], + "type": "text", + "content": " , given by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 322, + 91, + 547, + 112 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 91, + 547, + 112 + ], + "spans": [ + { + "bbox": [ + 322, + 91, + 547, + 112 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F _ {a v} = P _ {f} \\left(X ^ {\\prime}\\right), X ^ {\\prime} = \\left[ \\operatorname {M e a n} \\left(A ^ {\\prime}\\right) \\oplus \\operatorname {M e a n} \\left(V ^ {\\prime}\\right) \\right], \\\\ \\widehat {\\quad} \\end{array} \\tag {2}", + "image_path": "036fef2bcbb6cde0640a4e1bbd4912fd370e3885e7fa4b16f9f763fd93bbb895.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 323, + 109, + 400, + 123 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 109, + 400, + 123 + ], + "spans": [ + { + "bbox": [ + 323, + 109, + 400, + 123 + ], + "type": "interline_equation", + "content": "\\widehat {Y} = \\operatorname {s i m} (\\mathcal {T}, F _ {a v})", + "image_path": "7c79754b376292fee6d250484bf1bbfb8ab5c58195d076466822cc018c591707.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 131, + 547, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 131, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 305, + 131, + 547, + 192 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 131, + 547, + 192 + ], + "type": "inline_equation", + "content": "[\\oplus ]" + }, + { + "bbox": [ + 305, + 131, + 547, + 192 + ], + "type": "text", + "content": " denotes the concatenation operator, and " + }, + { + "bbox": [ + 305, + 131, + 547, + 192 + ], + "type": "inline_equation", + "content": "\\operatorname {sim}(X,Y)" + }, + { + "bbox": [ + 305, + 131, + 547, + 192 + ], + "type": "text", + "content": " denotes cosine similarity between " + }, + { + "bbox": [ + 305, + 131, + 547, + 192 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 131, + 547, + 192 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 131, + 547, + 192 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 305, + 131, + 547, + 192 + ], + "type": "text", + "content": ". We formulate a multi-label classification objective " + }, + { + "bbox": [ + 305, + 131, + 547, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mcid}}" + }, + { + "bbox": [ + 305, + 131, + 547, + 192 + ], + "type": "text", + "content": " by applying binary cross-entropy loss " + }, + { + "bbox": [ + 305, + 131, + 547, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{bce}}(\\cdot)" + }, + { + "bbox": [ + 305, + 131, + 547, + 192 + ], + "type": "text", + "content": " on each prediction using video-level ground-truth labels " + }, + { + "bbox": [ + 305, + 131, + 547, + 192 + ], + "type": "inline_equation", + "content": "Y\\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 305, + 131, + 547, + 192 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 370, + 201, + 545, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 201, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 370, + 201, + 545, + 233 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {m c i d}} = \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\mathrm {b c e}} \\left(y _ {i}, \\widehat {y} _ {i}\\right) \\tag {3}", + "image_path": "6eff00af90f8218b4016be12ef8b99b25817e07e3c6969f96703eb25eb0856e1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 237, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 237, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 237, + 545, + 297 + ], + "type": "text", + "content": "In particular, we train audio, visual, and fusion projectors " + }, + { + "bbox": [ + 305, + 237, + 545, + 297 + ], + "type": "inline_equation", + "content": "(P_{a}, P_{v}, P_{f})" + }, + { + "bbox": [ + 305, + 237, + 545, + 297 + ], + "type": "text", + "content": " on noisy multi-source data keeping the AudioCLIP backbones frozen. Hence, these projectors assist in extracting relevant foreground features from noisy audio and visual mixture features extracted from AudioCLIP." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 304, + 507, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 507, + 317 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 507, + 317 + ], + "type": "text", + "content": "3.3. Audio and Visual Conditioning Blocks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 323, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 491 + ], + "type": "text", + "content": "Establishing one-to-one correspondences for each source in multi-source mixtures is particularly challenging, as both audio and visual features contain mixed representations of various sources. Prior works [21, 30] on multi-source localization have attempted to learn audio-visual alignment intrinsically without explicit single-source guidance. However, in the absence of single-source data, these methods struggle to adequately disentangle multi-source features. In contrast, we leverage the categorical text features of " + }, + { + "bbox": [ + 304, + 323, + 545, + 491 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 323, + 545, + 491 + ], + "type": "text", + "content": " classes present in the mixture as coarse guidance for fine-grained feature disentanglement from multi-source mixtures. We use ground truth weak labels " + }, + { + "bbox": [ + 304, + 323, + 545, + 491 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 323, + 545, + 491 + ], + "type": "text", + "content": " of representative audios in the mixture to extract text embedding of " + }, + { + "bbox": [ + 304, + 323, + 545, + 491 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 323, + 545, + 491 + ], + "type": "text", + "content": " visual sounding sources " + }, + { + "bbox": [ + 304, + 323, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_K = \\{e_k^t\\}_{k=1}^K \\subseteq \\mathcal{T}" + }, + { + "bbox": [ + 304, + 323, + 545, + 491 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 323, + 545, + 491 + ], + "type": "inline_equation", + "content": "e_k^t \\in \\mathbb{R}^{1 \\times D}" + }, + { + "bbox": [ + 304, + 323, + 545, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "text", + "content": "With the categorical text features " + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_K" + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "text", + "content": " sources, we independently disentangle the multi-object visual patch tokens " + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "inline_equation", + "content": "V^{\\prime}" + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "text", + "content": " and mixture audio patch tokens " + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "inline_equation", + "content": "A^\\prime" + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "text", + "content": " by using crossattention " + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\phi^v (\\cdot)" + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\phi^a (\\cdot)" + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "text", + "content": ", respectively, as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 345, + 545, + 545, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 345, + 545, + 545, + 578 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\begin{array}{c} \\widetilde {V} _ {k} = \\phi^ {v} \\left(V ^ {\\prime}, e _ {k} ^ {t}, V ^ {\\prime}\\right), f _ {k} ^ {v} = \\operatorname {M e a n} \\left(\\widetilde {V} _ {k}\\right), \\\\ \\sim \\end{array} \\tag {4} \\\\ \\tilde {f} _ {k} ^ {v} = P _ {v c} \\left(f _ {k} ^ {v}\\right), \\forall k = \\{1, \\dots , K \\} \\\\ \\end{array}", + "image_path": "9c7c05c34a8ceab4acfd36ffa1d2de3ba7a2dd763e51f1db1ca0b2f8b21f5dab.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 345, + 587, + 545, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 587, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 345, + 587, + 545, + 659 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\underset {\\sim} {\\widetilde {A}} _ {k} = \\phi^ {a} \\left(A ^ {\\prime}, e _ {k} ^ {t}, A ^ {\\prime}\\right), f _ {k} ^ {a} = \\operatorname {M e a n} \\left(\\widetilde {A} _ {k}\\right), (5) \\\\ \\hat {f} _ {k} ^ {a} = P _ {a c} \\left(f _ {k} ^ {a}\\right), \\forall k = \\{1, \\dots , K \\} \\\\ \\phi (A, B, C) = \\left(\\frac {A B ^ {\\top}}{\\| A \\| \\| B \\|}\\right) \\odot C (6) \\\\ \\end{array}", + "image_path": "6fa80b27fc387778c990cef3536bbadcd93db8dd15214346a19a3581377bc352.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\widetilde{f}_k^a,\\widetilde{f}_k^v,f^a,f^v\\in \\mathbb{R}^{1\\times D},\\widetilde{V}_k\\in \\mathbb{R}^{n_v\\times D},\\widetilde{A}_k\\in" + }, + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{n_a\\times D}" + }, + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "type": "inline_equation", + "content": "P_{ac},P_{vc}:\\mathbb{R}^D\\to \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "type": "text", + "content": " are linear audio and visual projectors on conditioned mean features, respectively, and " + }, + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 663, + 545, + 713 + ], + "type": "text", + "content": " is the Hadamard product." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26745" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 537, + 253 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 537, + 253 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 537, + 253 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 537, + 253 + ], + "type": "table", + "html": "
MethodVGGSound-SingleVGGSound-InstrumentsMUSIC-Solo
AP(%)IoU@0.5(%)AUC(%)AP(%)IoU@0.3(%)AUC(%)AP(%)IoU@0.5(%)AUC(%)
Current SOTA Based Methods
OTS [4] ECCV1838.943.243.750.956.938.476.257.848.9
CoarsetoFine [36] ECCV2039.443.144.851.157.539.176.358.149.7
LVS [7] CVPR2139.744.645.351.557.939.476.157.149.3
EZ-VSL [28] ECCV2240.645.147.252.358.839.978.458.751.1
Mix-and-Localize [21] CVPR2241.845.647.453.859.341.578.959.952.7
DSOL [20] NeurIPS20-46.847.9-61.643.7-62.754.6
MarginNCE [35] ICASSP2342.546.948.356.961.844.283.163.355.2
FNAC [43] CVPR2343.348.449.157.263.245.384.664.556.4
AVGN [30] CVPR2344.149.649.559.364.746.185.465.856.9
Alignment [42] ICCV2345.350.850.260.465.848.786.166.457.2
CLIP-Based Baseline Methods
AudioCLIP [17] ICASSP2242.847.448.558.362.745.283.863.155.7
Wav2CLIP [46] ICASSP2239.343.644.753.857.242.078.959.951.2
T-VSL (Ours)48.153.752.964.669.551.488.268.560.1
", + "image_path": "181c6437b7e4bf9ea3f05518cff3e16243eb88a914805711e57f7607ff2f9841.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 260, + 545, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 260, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 545, + 283 + ], + "type": "text", + "content": "Table 1. Performance comparison of single-source localization on VGGSound-single, VGGSound-Instruments, and MUSIC-Solo datasets. For the fair comparison, we use pre-trained AudioCLIP audio and image encoders for baseline methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 293, + 289, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 289, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 289, + 356 + ], + "type": "text", + "content": "Additionally, to guide extracting categorical audio " + }, + { + "bbox": [ + 46, + 293, + 289, + 356 + ], + "type": "inline_equation", + "content": "(\\widetilde{A}_k)" + }, + { + "bbox": [ + 46, + 293, + 289, + 356 + ], + "type": "text", + "content": " and visual " + }, + { + "bbox": [ + 46, + 293, + 289, + 356 + ], + "type": "inline_equation", + "content": "(\\widetilde{V}_k)" + }, + { + "bbox": [ + 46, + 293, + 289, + 356 + ], + "type": "text", + "content": " patch tokens, we introduce coarse supervision on the output of projected mean patch tokens " + }, + { + "bbox": [ + 46, + 293, + 289, + 356 + ], + "type": "inline_equation", + "content": "\\widetilde{f}_k^a,\\widetilde{f}_k^v" + }, + { + "bbox": [ + 46, + 293, + 289, + 356 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 46, + 293, + 289, + 356 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 293, + 289, + 356 + ], + "type": "text", + "content": "-class text embedding " + }, + { + "bbox": [ + 46, + 293, + 289, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{T}\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 46, + 293, + 289, + 356 + ], + "type": "text", + "content": ". Hence, the class conditioning loss " + }, + { + "bbox": [ + 46, + 293, + 289, + 356 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{cls}})" + }, + { + "bbox": [ + 46, + 293, + 289, + 356 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 363, + 236, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 363, + 236, + 378 + ], + "spans": [ + { + "bbox": [ + 85, + 363, + 236, + 378 + ], + "type": "interline_equation", + "content": "\\mathbf {e} _ {k} ^ {v} = \\operatorname {s i m} \\left(\\mathcal {T}, \\widetilde {f _ {k} ^ {v}}\\right); \\mathbf {e} _ {k} ^ {a} = \\operatorname {s i m} \\left(\\mathcal {T}, \\widetilde {f _ {k} ^ {a}}\\right)", + "image_path": "28fe5637b2989b3743889832b8fdb27964b58d0e26995a4af2cff15d4a52f065.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 380, + 287, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 380, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 85, + 380, + 287, + 413 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c l s}} = \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathbf {e} _ {k} ^ {v}, \\mathbf {h} _ {k}\\right) + \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathbf {e} _ {k} ^ {a}, \\mathbf {h} _ {k}\\right) \\tag {7}", + "image_path": "a2e0bff0925ab2030094e1b081a9b5e360bed80964757d6c87e7c4c74dccd8e4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 421, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 421, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 287, + 459 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 421, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_k\\in \\mathbb{R}^N" + }, + { + "bbox": [ + 47, + 421, + 287, + 459 + ], + "type": "text", + "content": " denotes an one-hot encoding of the class label, and " + }, + { + "bbox": [ + 47, + 421, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_k^v\\in \\mathbb{R}^N" + }, + { + "bbox": [ + 47, + 421, + 287, + 459 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 421, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_k^a\\in \\mathbb{R}^N" + }, + { + "bbox": [ + 47, + 421, + 287, + 459 + ], + "type": "text", + "content": " are visual and audio class predictions for each " + }, + { + "bbox": [ + 47, + 421, + 287, + 459 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 47, + 421, + 287, + 459 + ], + "type": "text", + "content": " source, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 464, + 241, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 241, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 241, + 478 + ], + "type": "text", + "content": "3.4. Audio-Visual Correspondence Block" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "text", + "content": "While text representation provides coarse guidance for disentangling fine-grained spatio-temporal audio and visual features, it has inherent limitations in practice. For example, multiple instances of a sounding object class can appear in the visual frame, including silent instances, which are difficult to distinguish using simple text class representation. To address this challenge, we introduce an audio-visual correspondence block to further refine the alignment of audio and visual features based on extracted categorical features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 590, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 628 + ], + "type": "text", + "content": "With the disentangled categorical audio " + }, + { + "bbox": [ + 46, + 590, + 287, + 628 + ], + "type": "inline_equation", + "content": "(\\widetilde{A}_k)" + }, + { + "bbox": [ + 46, + 590, + 287, + 628 + ], + "type": "text", + "content": " and visual " + }, + { + "bbox": [ + 46, + 590, + 287, + 628 + ], + "type": "inline_equation", + "content": "(\\widetilde{V}_k)" + }, + { + "bbox": [ + 46, + 590, + 287, + 628 + ], + "type": "text", + "content": " patch tokens, we apply cross-attention " + }, + { + "bbox": [ + 46, + 590, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\phi^{av}(\\cdot)" + }, + { + "bbox": [ + 46, + 590, + 287, + 628 + ], + "type": "text", + "content": " to enhance audio-visual patch alignment as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 85, + 633, + 217, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 633, + 217, + 649 + ], + "spans": [ + { + "bbox": [ + 85, + 633, + 217, + 649 + ], + "type": "interline_equation", + "content": "g _ {k} ^ {a} = \\operatorname {M e a n} \\left(\\widetilde {A} _ {k}\\right), \\widehat {g} _ {k} ^ {a} = P _ {a v} \\left(g _ {k} ^ {a}\\right)", + "image_path": "e3183fc9695773f794e78a4e04c314b6b24a5d97939d741e82e2d617eddaa8d8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 85, + 651, + 287, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 651, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 85, + 651, + 287, + 666 + ], + "type": "interline_equation", + "content": "\\widehat {V} _ {k} = \\phi^ {a v} \\left(\\widetilde {V} _ {k}, \\widehat {g} _ {k} ^ {a}, \\widetilde {V} _ {k}\\right), g _ {k} ^ {v} = \\operatorname {M e a n} \\left(\\widehat {V} _ {k}\\right) \\tag {8}", + "image_path": "fac785d9a865152fe1ce36c2e37f688d203c27462a1bb9e3341fb3d15b56b186.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 85, + 667, + 223, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 667, + 223, + 681 + ], + "spans": [ + { + "bbox": [ + 85, + 667, + 223, + 681 + ], + "type": "interline_equation", + "content": "\\widehat {g} _ {k} ^ {v} = P _ {v a} \\left(g _ {k} ^ {v}\\right), \\forall k = \\{1, \\dots , K \\}", + "image_path": "999403a566af2e8c291e652cfb4c14cdbc5a4848723c62ad94d65c36c6008df9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "P_{av}, P_{va}: \\mathbb{R}^D \\to \\mathbb{R}^D" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " are linear projectors. Afterwards, to correspond between projected mean categorical" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 293, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 319 + ], + "type": "text", + "content": "audio and visual features " + }, + { + "bbox": [ + 305, + 293, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\widehat{g}_k^a,\\widehat{g}_k^v\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 305, + 293, + 545, + 319 + ], + "type": "text", + "content": " , we apply audiovisual contrastive loss " + }, + { + "bbox": [ + 305, + 293, + 545, + 319 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{av}})" + }, + { + "bbox": [ + 305, + 293, + 545, + 319 + ], + "type": "text", + "content": " given by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 324, + 545, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 324, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 361, + 324, + 545, + 343 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a v}} = \\mathcal {L} _ {\\mathrm {c n t}} \\left(\\widehat {g} _ {k} ^ {v}, \\widehat {g} _ {k} ^ {a}\\right) \\tag {9}", + "image_path": "7d450faf249f7cf998d0279b7f2e9045cec6608159f86773650f29926456762b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 362, + 341, + 489, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 341, + 489, + 354 + ], + "spans": [ + { + "bbox": [ + 362, + 341, + 489, + 354 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {a v}} + \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {m c i d}}", + "image_path": "8ebeca68f00a8ab9f7619caa2c25f143fceb17dcde2df3a2825d6f275dfb95e4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 358, + 479, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 358, + 479, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 358, + 479, + 372 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 358, + 479, + 372 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}" + }, + { + "bbox": [ + 306, + 358, + 479, + 372 + ], + "type": "text", + "content": " is the total accumulated loss." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 376, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 376, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 376, + 545, + 437 + ], + "type": "text", + "content": "Inference: During inference, we estimate cosine similarity across projected mean audio feature " + }, + { + "bbox": [ + 305, + 376, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\widehat{g}_k^a\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 305, + 376, + 545, + 437 + ], + "type": "text", + "content": " and aligned visual patch tokens " + }, + { + "bbox": [ + 305, + 376, + 545, + 437 + ], + "type": "inline_equation", + "content": "(\\widehat{V}_k = \\{\\widehat{v}_k^j\\}_{j = 1}^{n_v},\\widehat{v}_k^j\\in \\mathbb{R}^D)" + }, + { + "bbox": [ + 305, + 376, + 545, + 437 + ], + "type": "text", + "content": ". Finally, after reshaping and up-scaling through bilinear interpolation, we generate " + }, + { + "bbox": [ + 305, + 376, + 545, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 376, + 545, + 437 + ], + "type": "text", + "content": "-source heat-maps " + }, + { + "bbox": [ + 305, + 376, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{H}\\in \\mathbb{R}^{K\\times H\\times W}" + }, + { + "bbox": [ + 305, + 376, + 545, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 441, + 443, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 443, + 454 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 443, + 454 + ], + "type": "text", + "content": "4. Results and Discussions" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 456, + 421, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 421, + 469 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 421, + 469 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 475, + 545, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 678 + ], + "type": "text", + "content": "Datasets. We used MUSIC, VGGSound-Instruments, and VGGSound datasets for the performance evaluation1. MUSIC [49] dataset contains 445 solo music videos of 11 instruments and 142 duet music videos of 8 instruments from YouTube. Following prior works [30], we use MUSIC-Solo for single-source and MUSIC-Duet for multi-source localization. From MUSIC-Solo, we use 350 videos for training and remaining 95 for evaluation. For MUSIC-Duet, we use 120 videos for training and remaining 22 videos for evaluation. VGGSound-Instruments [21] is a subset of VGGSound dataset [6] containing around 32k video clips from 37 music instruments of 10s duration. Apart from the musical instrument datasets, we gather around 150k videos of 10s duration from original VGGSound dataset [6] representing single-source sounds of 221 categories, such as animals, people, nature, instruments, etc. For the single-source evaluation on VGGSound, we use the full test" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "type": "text", + "content": "1Since many videos are no longer available for public use, these datasets become smaller. We use the same data split for all baselines and reproduce them under the same setting for a fair comparison." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26746" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 54, + 70, + 541, + 261 + ], + "blocks": [ + { + "bbox": [ + 54, + 70, + 541, + 261 + ], + "lines": [ + { + "bbox": [ + 54, + 70, + 541, + 261 + ], + "spans": [ + { + "bbox": [ + 54, + 70, + 541, + 261 + ], + "type": "table", + "html": "
MethodVGGSound-DuetVGGSound-InstrumentsMUSIC-Duet
CAP(%)CIoU@0.3(%)AUC(%)CAP(%)CIoU@0.1(%)AUC(%)CAP(%)CIoU@0.3(%)AUC(%)
Current SOTA Single-Source Methods
OTS [4] ECCV1823.926.427.528.777.818.251.229.424.3
CoarseToFine [36] ECCV20-28.427.9-78.819.7-31.824.1
LVS [7] CVPR21-31.830.1-81.123.2-33.126.9
EZ-VSL [28] ECCV22-32.430.6-80.322.9-34.326.7
MarginNCE [35] ICASSP2327.133.231.233.582.324.755.136.128.5
FNAC [43] CVPR2329.835.133.435.484.126.857.337.930.6
Alignment [42] ICCV2330.435.833.935.684.927.157.838.330.9
Current SOTA Multi-Source Methods
Mix-and-Localize [21] CVPR2229.535.634.136.384.526.757.438.130.7
DSOL [20] NeurIPS20-36.934.6-85.928.4-38.831.3
AVGN [30] CVPR2331.937.835.438.186.728.859.239.332.4
CLIP-Based Baseline Methods
AudioCLIP [17] ICASSP2228.434.932.834.783.825.956.136.929.2
Wav2CLIP [46] ICASSP2226.331.428.931.280.322.252.632.227.2
T-VSL (ours)35.740.137.941.889.631.562.943.235.9
", + "image_path": "ac0a22cafc4a11ab7c789b44b93195df52f122aa69f2c259e0f41d7f685257c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 264, + 545, + 285 + ], + "lines": [ + { + "bbox": [ + 47, + 264, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 264, + 545, + 285 + ], + "type": "text", + "content": "Table 2. Performance comparison of multi-source localization on VGGSound-Duet, VGGSound-Instruments, and MUSIC-Duet datasets. For the fair comparison, we use pre-trained AudioCLIP audio and image encoders for baseline methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 297, + 287, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 287, + 381 + ], + "type": "text", + "content": "set of 5158 videos, while on VGGSound-Instruments, we use 446 videos following prior work [21, 30]. For the multi-source training and evaluation in both VGGSound and VGGSound-Instruments, we randomly concatenate two frames from different sounding sources to produce the multi-source input image with a size of " + }, + { + "bbox": [ + 46, + 297, + 287, + 381 + ], + "type": "inline_equation", + "content": "448 \\times 224" + }, + { + "bbox": [ + 46, + 297, + 287, + 381 + ], + "type": "text", + "content": ", and mix the single-source audios, following prior work [21, 30]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 384, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 539 + ], + "type": "text", + "content": "Evaluation Metrics. For evaluating single-source localization performance, we use Intersection over Union (IoU), average precision (AP), and Area Under Curve (AUC) following prior work [21, 30]. For the multi-source localization evaluation, we use Class-aware Average precision (CAP), Class-aware IoU (CIoU), and Area Under Curve (AUC) following prior work [21, 30]. Similarly, for fair comparison with the prior work [30], we use the same thresholds in these metrics: we use IoU@0.5 and CIoU@0.3 for the MUSIC-Solo and MUSIC-Duet datasets, IoU@0.3 and IoU@0.5 for single-source VGGSound-Instruments and VGGSound datasets, and CIoU@0.1 and CIoU@0.3 for multi-source VGGSound-Instruments and VGGSound." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": "Implementation details. We use the AudioCLIP [17] pretrained encoders containing ResNet-50 [19] image encoder, ESResNeXt [16] audio encoder, and transformer [38] based text encoder. For the fair comparison, we reproduce the result of existing methods with AudioCLIP image and audio encoders, instead of using separately pre-trained encoders. We use " + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": " resolution for the single-source input image, " + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "inline_equation", + "content": "D = 1024" + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "inline_equation", + "content": "n_v = 49" + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "inline_equation", + "content": "7 \\times 7(h \\times w)" + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": " spatial maps generated from the visual encoder, and " + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "inline_equation", + "content": "n_a = 60" + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "inline_equation", + "content": "10 \\times 6(m \\times f)" + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": " time-frequency map generated from the audio encoder. We use Adam optimizer [24] with a batch size of 256 and with a learning rate of " + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "content": "CLIP-Based Baseline methods. We adopt several CLIP-based baseline methods for the sound source localization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 297, + 455, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 455, + 308 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 455, + 308 + ], + "type": "text", + "content": "task, which are described as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 311, + 545, + 394 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 311, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 311, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 311, + 545, + 357 + ], + "type": "text", + "content": "- AudioCLIP [17]: We fine-tune pre-trained image and audio encoders of AudioCLIP on the target dataset. We estimate the localization heatmap for the sounding sources using the similar cosine similarity across patch tokens." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 358, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 358, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 306, + 358, + 545, + 394 + ], + "type": "text", + "content": "- Wav2CLIP [46]: Similar to AudioCLIP baseline, we experiment with Wav2CLIP audio and image encoders to extract localization heatmaps of sounding sources." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 405, + 474, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 405, + 474, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 474, + 417 + ], + "type": "text", + "content": "4.2. Comparison to SOTA Baselines" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 425, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 688 + ], + "type": "text", + "content": "Single-source localization: We present the quantitative results on single-source localization in Tab. 1. We apply the AudioCLIP image and audio encoders in all baseline methods for the fair comparison with our methods. We observe some performance improvements with several SOTA methods [30, 42, 43] over the AudioCLIP baseline. However, we also notice some reduction in performance from the AudioCLIP baseline in several methods [4, 21, 28]. We hypothesize that these can be due to the alterations of AudioCLIP grounding with the modified contrastive loss. In contrast, by utilizing all three modalities for sound source localization, the proposed T-VSL achieves the best performance over all CLIP-based baselines as well as SOTA methods. T-VSL outperforms the current SOTA single source localization method Alignment [42] by " + }, + { + "bbox": [ + 304, + 425, + 545, + 688 + ], + "type": "inline_equation", + "content": "+2.9" + }, + { + "bbox": [ + 304, + 425, + 545, + 688 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 425, + 545, + 688 + ], + "type": "inline_equation", + "content": "+3.7" + }, + { + "bbox": [ + 304, + 425, + 545, + 688 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 425, + 545, + 688 + ], + "type": "inline_equation", + "content": "+2.1" + }, + { + "bbox": [ + 304, + 425, + 545, + 688 + ], + "type": "text", + "content": " higher IoUs on VGGSound-Single, VGGSound-Instruments, and MUSIC-solo datasets, respectively. Moreover, T-VSL achieves " + }, + { + "bbox": [ + 304, + 425, + 545, + 688 + ], + "type": "inline_equation", + "content": "+6.3" + }, + { + "bbox": [ + 304, + 425, + 545, + 688 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 425, + 545, + 688 + ], + "type": "inline_equation", + "content": "+6.8" + }, + { + "bbox": [ + 304, + 425, + 545, + 688 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 425, + 545, + 688 + ], + "type": "inline_equation", + "content": "+5.4" + }, + { + "bbox": [ + 304, + 425, + 545, + 688 + ], + "type": "text", + "content": " higher IoUs over the AudioCLIP baseline [32]. We hypothesize that the enhanced refinement of noisy audio and visual features with text guidance in T-VSL effectively contributes to performance improvement." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Multi-source localization: We present performance comparison in multi-source sound localization in Tab. 2. As be" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26747" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 70, + 519, + 308 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 519, + 308 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 519, + 308 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 519, + 308 + ], + "type": "image", + "image_path": "2e149157b3f9d2717826ed51b1b01f0022c525cc0a32bfcb2cd9c2f5ca105fd6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 317, + 546, + 352 + ], + "lines": [ + { + "bbox": [ + 46, + 317, + 546, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 546, + 352 + ], + "type": "text", + "content": "Figure 3. We present qualitative comparisons on challenging multi-source localization with SOTA single and multi-source baseline methods. Here, blue color represents high-attention values to the sounding object, and red color represents low-attention values. The proposed T-VSL can selectively isolate the sounding regions from the background and generates more precise localization maps for sounding sources." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 361, + 289, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 541 + ], + "type": "text", + "content": "fore, we apply the AudioCLIP image and audio encoders in all baseline methods for the fair comparison with our methods. We observe some performance improvements, mostly in SOTA multi-source methods [20, 21, 30], over the AudioCLIP baseline. In most single source methods, we observe performance loss in multi-source cases over the baseline. The proposed T-VSL achieves the best performance that outperforms the SOTA multi-source baseline AVGN [30] by " + }, + { + "bbox": [ + 46, + 361, + 289, + 541 + ], + "type": "inline_equation", + "content": "+2.3" + }, + { + "bbox": [ + 46, + 361, + 289, + 541 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 361, + 289, + 541 + ], + "type": "inline_equation", + "content": "+2.9" + }, + { + "bbox": [ + 46, + 361, + 289, + 541 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 361, + 289, + 541 + ], + "type": "inline_equation", + "content": "+3.9" + }, + { + "bbox": [ + 46, + 361, + 289, + 541 + ], + "type": "text", + "content": " CIoUs on multi-source VGGSound-Duet, VGGSound-Instruments, and MUSIC-Duet datasets, respectively. Also, T-VSL achieves " + }, + { + "bbox": [ + 46, + 361, + 289, + 541 + ], + "type": "inline_equation", + "content": "+5.2" + }, + { + "bbox": [ + 46, + 361, + 289, + 541 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 361, + 289, + 541 + ], + "type": "inline_equation", + "content": "+5.8" + }, + { + "bbox": [ + 46, + 361, + 289, + 541 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 361, + 289, + 541 + ], + "type": "inline_equation", + "content": "+6.3" + }, + { + "bbox": [ + 46, + 361, + 289, + 541 + ], + "type": "text", + "content": " higher CIoUs over the AudioCLIP baseline. These significant performance gains on multi-source benchmarks demonstrate the effectiveness of the proposed T-VSL in disentangling multi-source mixtures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 548, + 287, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 717 + ], + "type": "text", + "content": "Qualitative comparisons: In addition, we present qualitative comparisons on multi-source localization in Fig. 3 among single-source baseline Alignment [42], self-supervised multi-source baseline Mix-and-Localize [21], weakly-supervised multi-source baseline AVGN [30], and proposed T-VSL. Single-source SOTA baseline Alignment [42] struggles in multi-source localization due to the lack of audio-visual correspondence, mostly in the presence of noisy audios and silent visual objects. However, without any guidance on fine-grained source localization, multi-source baselines [21, 30] underperform on localizing challenging multi-source mixtures. Notably, our method generates superior localization maps, which shows the effectiveness of T-VSL in multi-source feature disentanglement." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 361, + 492, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 361, + 492, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 492, + 374 + ], + "type": "text", + "content": "4.3. Ablation on T-VSL building blocks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "text", + "content": "We present the ablation study of different building blocks of T-VSL in Tab. 3 under both single-source and multisource settings. The vanilla baseline contains AudioCLIP image and audio encoders that directly operate on the input data. By integrating the audio-visual correspondence (AVC) block on extracted mixture patch tokens, we notice considerable performance increase in single-source (by " + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "inline_equation", + "content": "+1.1" + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "text", + "content": " AP and " + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "inline_equation", + "content": "+0.7" + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "text", + "content": " IoU@0.5) and multi-source localization (by " + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "inline_equation", + "content": "+1.3" + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "text", + "content": " CAP and " + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "inline_equation", + "content": "+1.4" + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "text", + "content": " CIoU@0.3). By integrating the audio and image conditioning blocks following audio-visual instance detection for separating fine-grained class features, we observe additional improvements of " + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "inline_equation", + "content": "+1.8" + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "text", + "content": " AP and " + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "inline_equation", + "content": "+2.3" + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "text", + "content": " AP in single-source, and " + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "inline_equation", + "content": "+3.1" + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "text", + "content": " CAP and " + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "inline_equation", + "content": "+3.6" + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "text", + "content": " AP in multisource localization, respectively. Finally, the combination of all building blocks results in the best performance that improves the baseline by " + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "inline_equation", + "content": "+5.3" + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "text", + "content": " AP, and 6.3 IoU@0.5 in single-source and by " + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "inline_equation", + "content": "+7.3" + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "text", + "content": " CAP and " + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "inline_equation", + "content": "+5.2" + }, + { + "bbox": [ + 304, + 375, + 545, + 613 + ], + "type": "text", + "content": " CIoU@0.3 in multi-source localization. This further demonstrates the effective disentanglement of audio-visual features facilitated by intermediate text guidance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 616, + 487, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 487, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 487, + 628 + ], + "type": "text", + "content": "4.4. Zero-shot transfer across datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "We can perform zero-shot transfer across datasets with proposed T-VSL by simply replacing the " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "-class text prompts in Fig. 2. Since other weakly-supervised methods cannot perform zero-shot transfer, we present comparisons with self-supervised Mix-and-Localize [21], Alignment [42], and FNAC [43] methods in Tab. 4. To ensure a fair comparison, we conducted uniform training with " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "50\\% - 50\\%" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " class" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26748" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 284, + 148 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 284, + 148 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 284, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 284, + 148 + ], + "type": "table", + "html": "
Audio Cond.Image Cond.AVCVGGSound-SingleVGGSound-Duet
AP(%)IoU@0.5(%)CAP(%)CIoU@0.3(%)
XXX42.847.428.434.9
XX✓43.948.129.736.3
✓X✓45.750.932.838.1
X✓✓46.251.433.338.6
✓✓✓48.153.735.740.1
", + "image_path": "47df21d6af5219e16d098f94171889e4cc6d7093b0fabd0b7e44764d8360cff3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 196, + 286, + 260 + ], + "blocks": [ + { + "bbox": [ + 47, + 156, + 287, + 190 + ], + "lines": [ + { + "bbox": [ + 47, + 156, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 287, + 190 + ], + "type": "text", + "content": "Table 3. Ablation study on audio conditioning block, image-conditioning block, and audio-visual correspondence block (AVC) of T-VSL in single-source and multi-source localization tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 196, + 286, + 260 + ], + "lines": [ + { + "bbox": [ + 51, + 196, + 286, + 260 + ], + "spans": [ + { + "bbox": [ + 51, + 196, + 286, + 260 + ], + "type": "table", + "html": "
Target DatasetProposed T-VSLMix-and-Localize [21]Alignment [42]FNAC [43]
MUSIC-Duet55.850.951.751.2
MUSIC-Solo80.575.177.677.4
VGGSound-Instr.81.776.876.576.1
", + "image_path": "eb7aada93367fb16955008f3ca0606de3bfd9fce8d1f58dc8e291b7da6ef7249.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 269, + 287, + 323 + ], + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 323 + ], + "type": "text", + "content": "Table 4. Ablation on zero-shot transfer across single and multi-source datasets with " + }, + { + "bbox": [ + 46, + 269, + 287, + 323 + ], + "type": "inline_equation", + "content": "50\\% -50\\%" + }, + { + "bbox": [ + 46, + 269, + 287, + 323 + ], + "type": "text", + "content": " class split. We report IoU@0.5(%) for MUSIC-Solo, CIoU@0.3(%) for MUSIC-Duet, and CIoU@0.1(%) for VGGSound-Instruments. We use same AudioCLIP encoders for all baselines." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "text", + "content": "split for all methods with same AudioCLIP encoders. We note that our method achieves significantly better performance on zero-shot transfer to target datasets than these compared approaches. In particular, we achieve " + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "inline_equation", + "content": "+4.9" + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "inline_equation", + "content": "+5.4" + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "inline_equation", + "content": "+4.9" + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "text", + "content": " improvements on target datasets than the multi-source Mix-and-Localize [21]. Also, we achieve " + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "inline_equation", + "content": "+4.1" + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "inline_equation", + "content": "+2.9" + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "inline_equation", + "content": "+5.2" + }, + { + "bbox": [ + 46, + 331, + 287, + 450 + ], + "type": "text", + "content": " higher scores than the SOTA single source baseline, Alignment [42]. This result suggests that our T-VSL has superior generalization capabilities for zero-shot localization tasks in unseen classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 454, + 265, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 265, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 265, + 467 + ], + "type": "text", + "content": "4.5. Robustness to a higher number of sources" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 468, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 635 + ], + "type": "text", + "content": "The proposed T-VSL is not limited by the number of sources present in training mixtures unlike prior work [21]. We present comparative analysis with SOTA multi-source AVGN [30], and self-supervised single source methods FNAC [43] and Alignment [42] in Tab. 5 on robustness to higher number of test sources in VGGSound dataset. We train all methods on the VGGSound-Single dataset with same encoders for a fair comparison. We note that our method consistently maintains significantly higher CIOU score irrespective of test scenarios, and also achieves notably smaller relative performance drop for a large number of sources in mixtures. These results further demonstrate the robustness of the proposed T-VSL in disentangling multi-source mixtures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 639, + 208, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 208, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 208, + 653 + ], + "type": "text", + "content": "4.6. Use of learnable text prompts" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "content": "To disentangle multi-source mixtures, we primarily use the class label of each sounding source as text prompts (e.g. dog barking, snake hissing). Inspired by the conditional text prompt learning in [50], we study the use of learnable prompts along with class label representations. Instead of" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 70, + 535, + 141 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 535, + 141 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 535, + 141 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 535, + 141 + ], + "type": "table", + "html": "
Test Source No.Proposed T-VSLAVGN [30]Alignment [42]FNAC [43]
239.636.835.434.1
335.730.828.927.1
429.522.419.318.8
", + "image_path": "e85c6252a1322cbfc37a4d9e7665240c8f8b92471d8bbe779529c9135cb5da19.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 313, + 184, + 538, + 289 + ], + "blocks": [ + { + "bbox": [ + 306, + 144, + 545, + 178 + ], + "lines": [ + { + "bbox": [ + 306, + 144, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 545, + 178 + ], + "type": "text", + "content": "Table 5. Robustness to higher number of sources in VGGSound dataset when trained with single-source data. CIoU@0.3(%) score is reported. Same AudioCLIP encoders are used for all baselines." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 184, + 538, + 289 + ], + "lines": [ + { + "bbox": [ + 313, + 184, + 538, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 538, + 289 + ], + "type": "table", + "html": "
Prompt LengthVGGSound-SingleVGGSound-Duet
AP(%)IoU@0.5(%)CAP(%)CIoU@0.3(%)
048.153.735.740.1
248.554.136.040.4
449.254.636.841.3
850.355.237.142.5
1650.255.037.442.9
3249.754.537.142.7
", + "image_path": "14fdd797b16167749f5137e8da0a12307e32b1474d99bf651e9a1392ae871ea1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 296, + 545, + 340 + ], + "lines": [ + { + "bbox": [ + 305, + 296, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 305, + 296, + 545, + 340 + ], + "type": "text", + "content": "Table 6. Ablation study on the effect of learnable text prompts on VGGSound-Single and VGGSound-Duet datasets. Integration of learnable prompts results in noticeable performance improvements in both single and multi-source localization." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 346, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 453 + ], + "type": "text", + "content": "only using class label representation, a set of learnable embedding is integrated with class label embedding that provides additional flexibility to adapt the text prompt to target objective. We analyze the sound source localization performance with different length of learnable prompts. The performances are reported in Table 6. We notice considerable performance improvements by optimizing learnable prompts in both single and multi-source cases. However, such prompts limit zero-shot use cases to unseen classes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 459, + 378, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 459, + 378, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 459, + 378, + 471 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 474, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 641 + ], + "type": "text", + "content": "In this paper, we propose T-VSL, a novel text-guided multi-source visual sound source localization framework that can disentangle one-to-one audio-visual correspondence from multi-source mixtures. We leverage the text modality to guide fine-grained feature separation and localization, from noisy audio and visual features of mixtures, by exploiting the joint embedding space of AudioCLIP. In comparison with SOTA multi-source baselines, our method shows superior zero-shot transfer across datasets. Moreover, our method demonstrates notable robustness on challenging test mixtures with higher number of sources than training scenarios. In both single-source and multi-source localization, our method largely outperforms existing weakly-supervised and self-supervised baselines on three benchmark datasets." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 648, + 400, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 648, + 400, + 661 + ], + "spans": [ + { + "bbox": [ + 306, + 648, + 400, + 661 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "This research was supported in part by ONR Minerva program, iMAGiNE - the Intelligent Machine Engineering Consortium at UT Austin, and a UT Cockrell School of Engineering Doctoral Fellowship." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26749" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In Proceedings of European Conference on Computer Vision (ECCV), pages 208–224, 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "text", + "content": "[2] Muhammad Ali and Salman Khan. Clip-decoder: Zeroshot multilabel classification using multimodal clip aligned representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4675-4679, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 214 + ], + "type": "text", + "content": "[3] Tayfun Alpay, Sven Magg, Philipp Broze, and Daniel Speck. Multimodal video retrieval with clip: a user study. Information Retrieval Journal, 26(1-2):6, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 287, + 249 + ], + "type": "text", + "content": "[4] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proceedings of the European Conference on Computer Vision (ECCV), pages 435-451, 2018. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 287, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 287, + 295 + ], + "type": "text", + "content": "[5] Maria A Bravo, Sudhanshu Mittal, and Thomas Brox. Localized vision-language matching for open-vocabulary object detection. In DAGM German Conference on Pattern Recognition, pages 393-408. Springer, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 296, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 287, + 350 + ], + "type": "text", + "content": "[6] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "type": "text", + "content": "[7] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16867-16876, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 408, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 287, + 453 + ], + "type": "text", + "content": "[8] Chaorui Deng, Qi Chen, Pengda Qin, Da Chen, and Qi Wu. Prompt switch: Efficient clip adaptation for text-video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15648-15658, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 454, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 487 + ], + "type": "text", + "content": "[9] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary panoptic segmentation with maskclip. arXiv preprint arXiv:2208.08984, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 488, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 287, + 543 + ], + "type": "text", + "content": "[10] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "text", + "content": "[11] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model clip. In Proceedings of the AAAI conference on artificial intelligence, pages 6568-6576, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "text", + "content": "[12] John W Fisher III, Trevor Darrell, William Freeman, and Paul Viola. Learning joint statistical models for audio-visual fusion and segregation. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2000. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "text", + "content": "[13] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[14] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models. In Proceed-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19338-19347, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "type": "text", + "content": "[15] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xupeng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 746-754, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 150, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 204 + ], + "type": "text", + "content": "[16] Andrey Guzhov, Federico Raue, Jörn Hees, and Andreas Dengel. Esresne (x) t-fbsp: Learning robust time-frequency transformation of audio. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "text", + "content": "[17] Andrey Guzhov, Federico Raue, Jorn Hees, and Andreas Dengel. Audioclip: Extending clip to image, text and audio. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 976-980. IEEE, 2022. 2, 3, 5, 6, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "text", + "content": "[18] John Hershey and Javier Movellan. Audio vision: Using audio-visual synchrony to locate sounds. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 1999. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "type": "text", + "content": "[19] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9248-9257, 2019. 1, 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 350, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 414 + ], + "type": "text", + "content": "[20] Di Hu, Rui Qian, Minyue Jiang, Xiao Tan, Shilei Wen, Errui Ding, Weiyao Lin, and Dejing Dou. Discriminative sounding objects localization via self-supervised audiovisual matching. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 10077-10087, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 415, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 469 + ], + "type": "text", + "content": "[21] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10483-10492, 2022. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "text", + "content": "[22] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "type": "text", + "content": "[23] Einat Kidron, Yoav Y Schechner, and Michael Elad. Pixels that sound. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2005. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 559, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 590 + ], + "type": "text", + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "text", + "content": "[25] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "text", + "content": "[26] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In International Conference on Machine Learning, pages 23033-23044. PMLR, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[27] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26750" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "text", + "content": "Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In European Conference on Computer Vision, pages 728-755. Springer, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 150 + ], + "type": "text", + "content": "[28] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In Proceedings of European Conference on Computer Vision (ECCV), page 218-234, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 151, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 287, + 194 + ], + "type": "text", + "content": "[29] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 248 + ], + "type": "text", + "content": "[30] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10565–10574, 2023. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 315 + ], + "type": "text", + "content": "[31] Jishnu Mukhoti, Tsung-Yu Lin, Omid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19413-19423, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 380 + ], + "type": "text", + "content": "[32] Jishnu Mukhoti, Tsung-Yu Lin, Amid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open vocabulary semantic segmentation with patch aligned contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19413-19423, 2023. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "type": "text", + "content": "[33] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "type": "text", + "content": "[34] Sooyoung Park, Arda Senocak, and Joon Son Chung. Can clip help sound source localization? arXiv preprint arXiv:2311.04066, 2023. 3, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "text", + "content": "[35] Sooyoung Park, Arda Senocak, and Joon Son Chung. Marginnce: Robust sound localization with a negative margin. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 504, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 548 + ], + "type": "text", + "content": "[36] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In Proceedings of European Conference on Computer Vision (ECCV), pages 292-308, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "text", + "content": "[37] Jie Qin, Jie Wu, Pengxiang Yan, Ming Li, Ren Yuxi, Xuefeng Xiao, Yitong Wang, Rui Wang, Shilei Wen, Xin Pan, et al. Freeseg: Unified, universal and open-vocabulary image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19446-19455, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 647 + ], + "type": "text", + "content": "[38] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 663 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "text", + "content": "[40] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4358-4366, 2018. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 546, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 183 + ], + "type": "text", + "content": "[41] Arda Senocak, Hyeonggon Ryu, Junsik Kim, and In So Kweon. Learning sound localization better from semantically similar samples. In Proceedings of IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 185, + 546, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 546, + 239 + ], + "type": "text", + "content": "[42] Arda Senocak, Hyeonggon Ryu, Junsik Kim, Tae-Hyun Oh, Hanspeter Pfister, and Joon Son Chung. Sound source localization is all about cross-modal alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7777-7787, 2023. 1, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 240, + 546, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 546, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 546, + 306 + ], + "type": "text", + "content": "[43] Weixuan Sun, Jiayi Zhang, Jianyuan Wang, Zheyuan Liu, Yiran Zhong, Tianpeng Feng, Yandong Guo, Yanhao Zhang, and Nick Barnes. Learning audio-visual source localization via false negative aware contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6420–6429, 2023. 1, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 307, + 546, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 307, + 546, + 351 + ], + "spans": [ + { + "bbox": [ + 307, + 307, + 546, + 351 + ], + "type": "text", + "content": "[44] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1802-1812, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 352, + 546, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 352, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 352, + 546, + 396 + ], + "type": "text", + "content": "[45] Tao Wang. Learning to detect and segment for open vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7051-7060, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 396, + 546, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 396, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 396, + 546, + 453 + ], + "type": "text", + "content": "[46] Ho-Hsiang Wu, Prem Seetharaman, Kundan Kumar, and Juan Pablo Bello. Wav2clip: Learning robust audio representations from clip. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4563-4567. IEEE, 2022. 3, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 453, + 546, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 453, + 546, + 508 + ], + "spans": [ + { + "bbox": [ + 307, + 453, + 546, + 508 + ], + "type": "text", + "content": "[47] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2935-2944, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 508, + 546, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 508, + 546, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 508, + 546, + 563 + ], + "type": "text", + "content": "[48] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 563, + 546, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 563, + 546, + 608 + ], + "spans": [ + { + "bbox": [ + 307, + 563, + 546, + 608 + ], + "type": "text", + "content": "[49] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In Proceedings of the European Conference on Computer Vision (ECCV), pages 570-586, 2018. 1, 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 609, + 546, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 609, + 546, + 663 + ], + "spans": [ + { + "bbox": [ + 307, + 609, + 546, + 663 + ], + "type": "text", + "content": "[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 8" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26751" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/11bec0da-9290-4827-bb16-e062c2aadc7c_content_list.json b/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/11bec0da-9290-4827-bb16-e062c2aadc7c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..500a7949584d6ccf1a3ba0b844fc15e0078de8c1 --- /dev/null +++ b/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/11bec0da-9290-4827-bb16-e062c2aadc7c_content_list.json @@ -0,0 +1,1669 @@ +[ + { + "type": "text", + "text": "T4P: Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory", + "text_level": 1, + "bbox": [ + 107, + 128, + 864, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daehee Park, Jaeseok Jeong, Sung-Hoon Yoon, Jaewoo Jeong, and Kuk-Jin Yoon KAIST", + "bbox": [ + 163, + 203, + 805, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{bag2824, jason. jeong, yoon307, jeong207, kjyoon}@kaist.ac.kr", + "bbox": [ + 228, + 241, + 733, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Trajectory prediction is a challenging problem that requires considering interactions among multiple actors and the surrounding environment. While data-driven approaches have been used to address this complex problem, they suffer from unreliable predictions under distribution shifts during test time. Accordingly, several online learning methods have been proposed using regression loss from the ground truth of observed data leveraging the auto-labeling nature of trajectory prediction task. We mainly tackle the following two issues. First, previous works underfit and overfit as they only optimize the last layer of motion decoder. To this end, we employ the masked autoencoder (MAE) for representation learning to encourage complex interaction modeling in shifted test distribution for updating deeper layers. Second, utilizing the sequential nature of driving data, we propose an actor-specific token memory that enables the test-time learning of actor-wise motion characteristics. Our proposed method has been validated across various challenging cross-dataset distribution shift scenarios including nuScenes, Lyft, Waymo, and Interaction. Our method surpasses the performance of existing state-of-the-art online learning methods in terms of both prediction accuracy and computational efficiency. The code is available at https://github.com/daeheepark/T4P.", + "bbox": [ + 75, + 325, + 473, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 737, + 209, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Trajectory prediction plays a significant role in autonomous systems, enhancing safety and navigation efficiency [28, 31]. Recently, data-driven methods have shown remarkable prediction capabilities [1, 16, 38, 52, 55, 57, 60, 81, 89, 90]; however, they are prone to distribution shift [9, 88]. Trajectory prediction models also produce unreliable output when faced shifts in the data distribution [25], posing significant risks in various real-world applications. This vulnerability stems from the ease with which the trajectory data distri", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/33acb881603285214330de90b83340b0d3fc7c37f8a9239ca2e9d5f9f611682c.jpg", + "image_caption": [ + "Figure 1. Previous methods optimize the last layer of the decoder using regression loss from delayed ground truth. Our method, on the other hand, learns representation via a masked autoencoder, which boosts prediction performance by optimizing deeper layers. In addition, the proposed actor-specific token enables the prediction model to learn actor-wise motion characteristics." + ], + "image_footnote": [], + "bbox": [ + 517, + 292, + 867, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "bution can be altered by numerous factors, such as scene changes and driving habits; i.e. road layout, interaction between agents, driver demographics [6, 51, 82].", + "bbox": [ + 496, + 640, + 893, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this challenge, recent methods have proposed domain adaptation and generalization strategies which aim to anticipate the distribution shifts and accordingly train the model [32, 76, 80]. However, due to the wide variety of factors influencing data distribution, the anticipated shifts may differ significantly from those encountered at test time. As a result, several online learning methods have been developed to dynamically adapt models during test time [36, 73]. Since trajectory prediction serves as an auto-labeling task where trajectory data is obtained from object tracking, the observed past and future trajectory provide both input and corresponding ground truth for supervision $(\\mathcal{L}_{reg})$ , as depicted at the top right of Fig. 1. Nevertheless, updating the entire model may ruin the representation learned from the", + "bbox": [ + 496, + 688, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 1, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15065", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "source data, so only part of the network, such as the batch normalization layer, is updated [27, 40, 61, 83]. Particularly, because regression loss is calculated at delayed timestamps and only a few samples are available during test time, this approach risks deteriorating the model's learned representation [73]. Therefore, previous online prediction methods mainly update only the last layer of the decoder.", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a test-time training (TTT) for trajectory prediction with two key aspects. First, we build a masked autoencoder (MAE) framework to adapt deep features, incorporating good representation that captures complex interactions between agents and road structures. Due to the challenge of existing online learning methods in damaging learned representations when updating deeper layers, we employ a MAE to guide representation learning. Second, we introduce an actor-specific token memory that has significant advantages in real-world driving scenario where data arrives continuously and sequentially. As each actor instance has its own driving habits and the past motion pattern of specific actors can be accessed from the arrived observations, we design a token memory in transformer structure [68] and its training strategy to learn actor-wise motion characteristics. The proposed TTT framework is validated on challenging cross-dataset distribution shift cases between nuScenes, Lyft, Waymo, and INTERACTION, and shows state-of-the-art performance surpassing previous online learning methods. Furthermore, we show the practicality of our TTT framework by evaluating its computational efficiency. We summarize our contributions as below:", + "bbox": [ + 75, + 198, + 473, + 529 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a test time training for trajectory prediction (T4P) by utilizing a masked autoencoder to learn deep feature representations that stably improve prediction performance across entire network layers.", + "- We introduce an actor-specific token memory used to learn the different actor characteristics and habits.", + "- Our method is validated across 4 different datasets as well as different temporal configurations. Ours achieves state-of-the-art performance both in accuracy and efficiency." + ], + "bbox": [ + 76, + 531, + 468, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 684, + 225, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Trajectory Prediction", + "text_level": 1, + "bbox": [ + 76, + 710, + 279, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Trajectory prediction garners attention with the emergence of methods that can enhance perception or planning [11, 12, 44]. Its goal is to predict future trajectories of traffic actors based on their historical trajectories and the context of their environment [3, 17, 19, 20, 39, 58, 63, 78, 91]. Historical trajectories, or tracklets of traffic actors, are sequentially acquired via vehicle detection and tracking systems. Some studies employed this temporal property to enhance prediction via memory replay [35, 42, 59]. In the early stages of trajectory prediction, only the historical trajectory of the actors of interest was considered. However, recent stud", + "bbox": [ + 75, + 734, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ies emphasize the significance of understanding interactions among agents [67, 85] and the rules governed by surrounding environments [72, 77] in improving prediction performance. This has led to the development of models that incorporate multi-head attention or graph-based methods to capture these interactions [24, 26, 42]. Additionally, MAE has been adopted for pretraining to better understand agent interactions [8, 13]. To further refine prediction capabilities, various generative models have been introduced, enabling the generation of future trajectories [15, 41, 71, 79].", + "bbox": [ + 496, + 90, + 890, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Transfer Learning in Trajectory Prediction", + "text_level": 1, + "bbox": [ + 498, + 253, + 867, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With data-driven approaches offering superior performance in trajectory prediction, their effectiveness diminishes under distribution shifts [25, 56]. In response, several studies have adopted domain adaptation or generalization strategy [69, 74]. Some specifically aimed to reduce the domain gap within unique characteristics of trajectory prediction: differences in road structures [84], actor interaction [80], etc. However, these methods rely on anticipating how to cover domain shifts. Yet, given that trajectory data is subject to influence from numerous factors, predicting and accommodating for shifts may not always be sufficient. Consequently, recent developments have introduced adaptation to unseen test sets using online learning [33, 34, 43]. Among them, some methods [36, 73] showed remarkable prediction performance improvement under severe distribution shifts like cross-dataset cases by utilizing regression loss for online learning. These methods exploit the fact that the input and ground truth (GT) labels are provided at test time as tracking history. However, with the limitations of updating with only a few samples in a delayed time, adaptation becomes restricted to the last layer of the decoder.", + "bbox": [ + 496, + 276, + 890, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Test Time Training", + "text_level": 1, + "bbox": [ + 498, + 604, + 684, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Test-time training (TTT) is a method that trains the network on test time data, unseen during training [4, 7, 14, 18, 22, 49, 66, 75]. Unlike domain generalization or adaptation, which are confined to the training phase, TTT extends model adaptation into the test phase by utilizing available test data [47]. TTT methods are categorized into regularization-based approaches for post-hoc regularization of out-of-distribution (OOD) samples [46, 87], and self-supervised approaches that employ pretext tasks on test data for optimal representation learning [7, 9, 45, 50, 54]. Specifically, TTT [65] introduced a Y-shaped network structure consisting of a feature encoder, a pretext branch, and a decoder branch. The decoder branch is fixed, while the encoder and pretext branch are optimized through self-supervision. Adhering to this model, TTT-MAE [23] integrated a MAE in the pretext task. Expanding on this method, we adopt TTT-MAE to the domain of trajectory prediction, leveraging its representation learning capabilities to enhance test-time training.", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15066", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f3b3e43995adeb1299905a2a1f4552e9eb13aeb779576994b464a229953c164b.jpg", + "image_caption": [ + "Figure 2. Overall method. During test-time training, the network trained on source dataset is optimized on target data under online setting. The model is optimized both from regression and reconstruction loss. Both losses utilize the data observed at the delayed time stamp $(t_{\\tau})$ . Actor-specific token is used to learn instance-wise motion pattern during test-time training phase. During online evaluation phase, model and actor-specific token learned from test-time training phase are used." + ], + "image_footnote": [], + "bbox": [ + 84, + 92, + 885, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 410, + 166, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem definition", + "text_level": 1, + "bbox": [ + 76, + 435, + 256, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Trajectory prediction aims to learn the mapping function between the input, consisting of historical trajectory and map information $\\mathbf{X}:\\{\\mathcal{X}_t,\\mathcal{M}_t\\}$ , and the output, consisting of $K$ possible candidates for future trajectory of $N$ traffic actors, $\\mathbf{Y}:\\{\\mathcal{Y}_t^{0:K - 1}\\}$ , at current time $t$ . We predict $C$ different actor classes including vehicle, cyclist, etc. Historical and future trajectories are represented as $\\mathcal{X}_t = \\mathbf{x}_{t - t_h:t}^{0:N - 1}$ and $\\mathcal{Y}_t = \\mathbf{x}_{t:t + t_f}^{0:N - 1}$ where $t_h$ and $t_f$ represent sequence length of input and output trajectory. Here, $\\mathbf{x}_t^n$ represents the spatial location of actor $n$ at time $t$ . For map information $\\mathcal{M}_t$ , we use $L$ segmented lane centerlines around ego-actors which is widely-used in trajectory prediction methods.", + "bbox": [ + 75, + 458, + 468, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We deal with the case when the target data distribution during test time $\\{\\mathbf{X},\\mathbf{Y}\\}^T$ is different from the source data distribution seen during the training phase $\\{\\mathbf{X},\\mathbf{Y}\\}^S$ . We formulate the problem as a online adaptation scenario in which one data sample is given per each time interval as time passes. The test data is consists of multiple distinct scenes. Each scene includes temporally ordered data samples which are captured through real-world driving. Following standard TTT methods, there is no access to the source data at test time. However, thanks to the auto-labeling nature of trajectory prediction, there is access to delayed GT future trajectory $(\\mathbf{x}_{t_{\\tau}:t_{\\tau} + t_f})$ from a previous time window at time $t_\\tau (= t - \\tau)$ as depicted in left lower corner of Fig. 2.", + "bbox": [ + 75, + 642, + 470, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overall method", + "text_level": 1, + "bbox": [ + 76, + 847, + 235, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method, Test-Time Training of Trajectory Prediction (T4P), enhances the online learning method using supervi", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sion from a delayed GT future trajectory with representation learning from MAE and actor-specific token memory. Following standard TTT [65], the overall framework consists of three phases: offline training, test-time training, and online evaluation. Offline training occurs before test time, and test-time training and online evaluation are executed repeatedly and sequentially during test time. We adopt the ForecastMAE [13] backbone consisting of embedding layers $f$ , a shared encoder $E$ , a reconstruction head $R$ and a motion decoder head $D$ , as depicted in the middle of Fig. 2. The detailed methods during each phase are described below:", + "bbox": [ + 496, + 412, + 890, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Offline training", + "text_level": 1, + "bbox": [ + 500, + 587, + 656, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During offline training, the model is trained on source data using both reconstruction loss and regression loss.", + "bbox": [ + 496, + 609, + 890, + 641 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta \\in \\{f, E, R, D \\}} \\mathbb {E} _ {\\mathcal {X}, \\mathcal {Y}, \\mathcal {M} \\in \\{\\mathbf {X}, \\mathbf {Y} \\} ^ {S}} \\left[ \\mathcal {L} _ {\\text {r e c o n}} + \\mathcal {L} _ {\\text {r e g}} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 651, + 890, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this subsection, subscript $t$ is omitted for simplicity. First, all input elements $(\\mathcal{X},\\mathcal{Y},\\mathcal{M})$ are embedded with their respective embedding layer $(f_x,f_y,f_m)$ . Additionally, we define actor class token $\\bar{\\alpha}\\in \\mathbb{R}^{C\\times D}$ that learns different motion patterns of each actor class. The actor class token is implemented as a learnable embedding of a transformer structure. Corresponding actor class token $\\alpha (c)$ is added to trajectory embedding according to the class of each actor.", + "bbox": [ + 496, + 685, + 890, + 806 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {x}, h _ {y}, h _ {m} = f _ {x} (\\mathcal {X}) + \\alpha (c), f _ {y} (\\mathcal {Y}) + \\alpha (c), f _ {m} (\\mathcal {M}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 813, + 890, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For reconstruction, the history/future trajectory and lane embeddings are fed to the encoder to obtain the encodings $F_{x}, F_{y}, F_{m}$ . Then, segments of the encodings are randomly masked and replaced with masking tokens, $M_{x}, M_{y}, M_{m}$ ,", + "bbox": [ + 496, + 840, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15067", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and the other encodings remain unmasked $(F_x', F_y', F_m')$ . Here, we use random masking for the lane centerline and complementary masking strategy for history/future trajectory following previous works [13]. The masking tokens and the unmasked encodings are fed to the reconstructor to reconstruct the masked elements.", + "bbox": [ + 75, + 90, + 472, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {x}, F _ {y}, F _ {m} = E \\left(h _ {x}, h _ {y}, h _ {m}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 188, + 468, + 213 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {X}}, \\hat {\\mathcal {Y}}, \\hat {\\mathcal {M}} = R \\left(M _ {x}, M _ {y}, M _ {m}, F _ {x} ^ {\\prime}, F _ {y} ^ {\\prime}, F _ {m} ^ {\\prime}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 210, + 413, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The encoder and reconstructor both consist of multi-head attention to utilize interaction between history, future and lanes, thus, the reconstruction guides to the model have the capability of interaction reasoning. Finally, a reconstruction loss $\\mathcal{L}_{\\text{recon}}$ is computed as the MSE loss between the ground truth and the reconstructed outputs.", + "bbox": [ + 75, + 234, + 468, + 325 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{N} \\sum_ {n} (\\mathcal {X} - \\hat {\\mathcal {X}}) ^ {2} + \\frac {1}{N} \\sum_ {n} (\\mathcal {Y} - \\hat {\\mathcal {Y}}) ^ {2} + \\frac {1}{L} \\sum_ {l} (\\mathcal {M} - \\hat {\\mathcal {M}}) ^ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 332, + 468, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the decoder head, historical trajectory and lanes embeddings are again fed to the same encoder and the output encodings are passed to the motion decoder, composed of MLP layers. The decoder outputs $\\mathrm{K}$ candidates for trajectory prediction, and the regression loss $\\mathcal{L}_{reg}$ is computed with the widely-used Winner-takes-all (WTA) loss [29, 48].", + "bbox": [ + 75, + 377, + 468, + 467 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {Y}} ^ {0: K - 1} = D \\left(E \\left(h _ {h}, h _ {l}\\right)\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 478, + 318, + 496 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e g} = \\frac {1}{N} \\sum_ {n} \\operatorname {a r g m i n} _ {k \\in K} \\left(\\mathcal {Y} ^ {n} - \\hat {\\mathcal {Y}} ^ {n, k}\\right) ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 500, + 468, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Test-time training", + "text_level": 1, + "bbox": [ + 76, + 542, + 251, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During test-time, a data sample consisting of trajectories and maps arrives sequentially. Therefore, even though we cannot access the GT future trajectory of current time $(\\mathcal{V}_t)$ , we can access both the inputs and GT $(\\mathcal{X}_{t_{\\tau}},\\mathcal{Y}_{t_{\\tau}},\\mathcal{M}_{t_{\\tau}})$ at a previous time $t_\\tau$ . With this data, the model is optimized with the same objective as Eq. 1 with target data distribution instead of source data distribution.", + "bbox": [ + 75, + 565, + 468, + 670 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta \\in \\{f, E, R, D \\}} \\mathbb {E} _ {\\chi_ {t _ {\\tau}}, \\mathcal {Y} _ {t _ {\\tau}}, \\mathcal {M} _ {t _ {\\tau}} \\in \\{\\mathbf {X}, \\mathbf {Y} \\} ^ {T}} \\left[ \\mathcal {L} _ {\\text {r e c o n}} + \\mathcal {L} _ {\\text {r e g}} \\right] \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 679, + 468, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike existing online learning methods that only utilize regression loss, we incorporate an additional reconstruction loss. This enables the model to learn a good representation that considers the complex actor-actor and actor-lane interaction even in the unseen target data distribution. An advantage of representation learning is that the performance stably improves even when the deeper layers are optimized.", + "bbox": [ + 75, + 707, + 468, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4.1 Actor-specific token memory", + "text_level": 1, + "bbox": [ + 76, + 830, + 331, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike during the offline training phase, when the data order is shuffled, data at test-time comes in sequentially. Therefore, at test time, it is possible to keep track of movement", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e535b160ff33fd972055114707520df3b58efa713657fe5868a37432b5d83e17.jpg", + "image_caption": [ + "Figure 3. Actor-specific token memory is colored in gray. It evolves as time passes within a scene. For newborn actors, the corresponding class token is registered. Until the actor disappears, the token is updated through test-time training. At the end of the scene, all tokens are averaged by each class and passed to the next scene as denoted in red arrow and Eq. 7." + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 883, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "patterns of a specific actor instance. Using this, we propose an actor-specific token memory.", + "bbox": [ + 498, + 396, + 890, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall scheme is described in Fig. 3. During offline training, actor class tokens $\\bar{\\alpha}_{train} \\in \\mathbb{R}^{C \\times D}$ are trained to reflect the average motion pattern of each of the $C$ classes. At the beginning of the test-time, scene 0, class tokens are initialized from that of training phase ( $\\bar{\\alpha}_{\\text{scene}(0)} \\gets \\bar{\\alpha}_{\\text{train}}$ ). When a new $n^{th}$ -actor appears at time $t$ , the class token $\\alpha_{n}^{t}(c)$ is cloned from $\\bar{\\alpha}_{\\text{scene}(0)}$ by selecting corresponding class. The newborn tokens are then registered to the actor-specific token memory. The token memory is structured as a dictionary where actor instance ID/corresponding tokens are key/values. At each iteration, the actor-specific tokens are used for both test-time training and online evaluation. As time progresses, the actor-specific tokens evolve and are updated through the reconstruction and regression losses until the actor disappears in the scene. By giving each actor its own specific token that distinguishes it from the sharing of other parts of models with other actors, actor-specific motion patterns can be learned.", + "bbox": [ + 498, + 430, + 890, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the scene changes, scene 1, the actors observed during scene 0 are not to be observed anymore, so we need a new averaged actor class token $\\bar{\\alpha}_{\\text{scene}(1)}$ . For that, we average all the tokens in the memory at the final time step $T_{S}$ of scene 0 after gathering by classes as Eq. 7. Here, $N_{c}$ refers to the number of actors of class $c$ . It is because, with a sufficient number of actors class tokens per each class in memory, their average motion can be a representative motion pattern of actor classes. This is more useful than $\\bar{\\alpha}_{\\text{train}}$ because newly averaged tokens are trained on the target dataset while $\\bar{\\alpha}_{\\text{train}}$ contains motion pattern trained on the source dataset. The averaged class tokens are then passed to the next scene and used to initialize tokens for the", + "bbox": [ + 498, + 704, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15068", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/52282f1f8b365f2b19f0c7c4c3f5d0647264d0561bf35ca9874e4f13495e909e.jpg", + "table_caption": [ + "Table 1. Adaptation results in various distribution shifts. The model is trained on source dataset, and test-time trained and evaluated on target dataset (Source $\\rightarrow$ Target). All metrics are better in lower value. The best and second-best results are marked in bold and underline." + ], + "table_footnote": [], + "table_body": "
mADE6/mFDE6Short-term exp (1/3/0.1)Long-term exp (2/6/0.5)
INTER → nuSINTER → LyftnuS → WayMeannuS → LyftWay → LyftWay → nuSMean
Source Only1.047 / 2.2471.391 / 2.9450.431 / 1.0310.956 / 2.0741.122 / 2.5770.621 / 1.3471.153 / 2.2200.965 / 2.048
Joint Training1.116 / 2.4451.553 / 3.4580.472 / 1.1251.047 / 2.3431.108 / 2.5970.638 / 1.4041.091 / 2.0310.946 / 2.011
DUA1.118 / 2.4551.516 / 3.3520.516 / 1.2941.050 / 2.3671.365 / 3.2570.790 / 1.8681.270 / 2.6341.142 / 2.586
TENT (w/ sup)1.102 / 2.4231.519 / 3.4050.448 / 1.0711.023 / 2.3001.068 / 2.5140.628 / 1.3811.077 / 2.0120.924 / 1.969
MEK (τ = tf/2)1.012 / 2.4451.283 / 3.4580.445 / 1.1250.913 / 2.3431.079 / 2.5970.629 / 1.4041.079 / 2.0310.929 / 2.011
MEK (τ = tf)0.892 / 1.9520.746 / 1.6540.405 / 1.0610.691 / 1.5561.006 / 2.3690.615 / 1.3511.117 / 2.1400.913 / 1.953
AML (K0)2.093 / 4.6972.695 / 6.6771.624 / 2.1392.137 / 4.5041.787 / 3.0671.322 / 2.5711.618 / 2.9991.576 / 2.879
AML (full)1.149 / 2.5501.042 / 2.6160.764 / 1.7910.985 / 2.3191.462 / 2.5730.977 / 2.1841.495 / 2.9781.311 / 2.578
Ours (T4P)0.537 / 1.1370.391 / 0.8240.336 / 0.8070.421 / 0.9230.776 / 1.8200.549 / 1.1710.996 / 1.7840.774 / 1.592
", + "bbox": [ + 80, + 122, + 890, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "newborn actors. Please note that while scene 0 is initialized by actor class tokens from the training phase, subsequent scenes obtain as in Eq. 7. More details of memory evolving strategies can be found on the supplementary material.", + "bbox": [ + 75, + 277, + 470, + 339 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\alpha} _ {s c e n e (i + 1)} \\leftarrow \\left\\{\\frac {1}{N _ {c}} \\sum_ {n} ^ {N _ {c}} \\alpha_ {n} ^ {T} (c) \\right\\} _ {s c e n e (i)} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 344, + 468, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Online evaluation", + "text_level": 1, + "bbox": [ + 76, + 393, + 250, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using the updated model weight and actor-specific token memory during test-time training, online evaluation is executed. With the input data $(\\mathcal{X}_t,\\mathcal{M}_t)$ at current time $t$ , the learned encoder and decoder predict multi-modal trajectory $(\\mathcal{Y}_t)$ for all actors in the sample.", + "bbox": [ + 75, + 417, + 468, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 505, + 202, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 529, + 179, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conducted experiments on well-known datasets, nuScenes [5], Lyft [30], WOMD [21], and INTERACTION [86], to evaluate T4P on various data distribution shifts. These datasets are parsed into the same format using trajdata [37]. Additionally, to verify in various prediction configurations, experiments were conducted with the two most widely used configurations of long-term and short-term prediction. Long-term prediction requires predicting 6 seconds into the future given 2 seconds of the past with a time interval of 0.5s, making the input/output sequence lengths to be 5 and 12, respectively. Short-term prediction requires predicting 3 seconds into the future given 0.9 seconds of the past with a time interval of 0.1s, making the input/output sequence lengths to be 10 and 30, respectively.", + "bbox": [ + 75, + 551, + 468, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 771, + 290, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For actor classes, we use the 5 classes: unknown, vehicle, pedestrian, bicycle and motorcycle. Our method predicts $K = 6$ future candidates for all actors in the sample. We use $\\tau$ as $t_f$ to enable the past GT future to contain full prediction horizon. We train and evaluate our model with a single NVIDIA A6000. Learning rates of model weight and actor-specific parameters are set as 0.01 and 0.5, respectively, and", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "weight decay is set to 0.001 for all. The gradient is clipped by 15. For metrics, widely used $\\mathrm{mADE}_6$ and $\\mathrm{mFDE}_6$ are used. Detailed metric definition, model architecture, and training details are included in the supplementary material.", + "bbox": [ + 496, + 277, + 890, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Baselines", + "text_level": 1, + "bbox": [ + 498, + 349, + 607, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our $T4P$ with several baselines, including unsupervised/supervised test-time-training methods and online learning trajectory prediction methods. All baseline methods are implemented using the same backbone.", + "bbox": [ + 496, + 375, + 890, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Source only refers to the backbone model trained on the source dataset only using regression loss.", + "bbox": [ + 496, + 436, + 890, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Joint training is similar to source only but trained with regression and reconstruction loss jointly.", + "bbox": [ + 496, + 467, + 890, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DUA [53] is an unsupervised post-hoc regularization method only updates batch normalization statistics in a momentum-updating manner without back-propagation.", + "bbox": [ + 496, + 500, + 890, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TENT with supervision is a variant of the original TENT [70] in which regression loss is used to optimize the batch normalization layers instead of entropy minimization loss, as entropy minimization is not applicable.", + "bbox": [ + 496, + 545, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MEK [73] is an online learning trajectory prediction method utilizing the Modified Extended Kalman filter. It uses only regression loss to optimize the last layer of the decoder. As the prediction horizon is different in our experiment from the original paper, we use both $\\frac{1}{2} t_f$ and $t_f$ .", + "bbox": [ + 496, + 606, + 890, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AML [36] is an Adaptive Meta-learning method. Unlike the other methods that use the same backbone, AML replaces the last decoder layer with a Bayesian linear regression layer for adaptive training. The modified version of backbone without adaptive training is denoted as $K_{0}$ , while the full version with adaptive training is denoted as full.", + "bbox": [ + 496, + 683, + 890, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 500, + 790, + 586, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1.Quantitative results", + "text_level": 1, + "bbox": [ + 500, + 815, + 687, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The results of comparing our method with the baselines in various distribution shift scenarios are presented in Tab. 1. We reported three distribution shift scenarios per each time configuration on the table, and other results are included in", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15069", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4fd3c636e118474740e034311324639347cb5a334a62e9040f6797039ff237ae.jpg", + "image_caption": [ + "Figure 4. The first row shows prediction before adaptation, and the second row indicates adaptation results by three methods: ours (blue), TENT w/ sup (orange) and MEK (green). Sky blue and orange boxes refer to surrounding actors and actors to be predicted. We depicted only one actor result and one mode among multi-modal predictions closest to the GT for visual simplicity. Please note that our method is multi-modal prediction for all actors method." + ], + "image_footnote": [], + "bbox": [ + 84, + 85, + 888, + 271 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a52ae0ae537e5b08008adbc411759534ea5d38dd112a24fa1b5c629b8dfbb041.jpg", + "image_caption": [ + "Figure 5. Prediction accuracy and execution time on INTER $\\rightarrow$ nuS (1/3/0.1) experiment. Adjusting update frequency can balance between accuracy and efficiency. Our method significantly outperforms the baseline methods in both accuracy and efficiency." + ], + "image_footnote": [], + "bbox": [ + 120, + 342, + 423, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the supplementary material. Notably, our approach consistently surpassed baseline performance across all scenarios.", + "bbox": [ + 75, + 571, + 468, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DUA consistently exhibits compromised performance across nearly all cases, due to the distinctive characteristics of trajectory data. In contrast to image data, where data is treated as a singular sample, trajectory data involves multiple agents, each exhibiting distinct motion patterns, within a single data. Consequently, holistically updating batch statistics proves to be counterproductive. Similar challenges are encountered by TENT w/ sup. While regression loss prevents performance decline, updating only the batch norm layer has little to no effect on the prediction performance.", + "bbox": [ + 75, + 608, + 468, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MEK exhibited the most competitive performance among the baselines. While performance improved significantly in short-term settings, it brings limited improvement in long-term cases. As the Kalman filter updates based on the number of prediction steps, short-term configurations with 12 update steps show a better improvement than long-term configurations with only 5 update steps.", + "bbox": [ + 75, + 762, + 468, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although AML led to a considerable improvement in the full version, the predictive performance itself was substan", + "bbox": [ + 73, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tially degraded due to the significant performance drop in the modified backbone $(K_0)$ . The limitation of the backbone is due to the Bayesian regression layer being based on probability sampling which is known to be worse than non-probability sampling method of ours [2]. In contrast to all the baseline methods, our method demonstrated state-of-the-art performance in all scenarios featuring various distribution shifts, whether short-term or long-term, showcasing the generalizability of our approach.", + "bbox": [ + 496, + 340, + 890, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.1 Efficiency", + "text_level": 1, + "bbox": [ + 500, + 494, + 624, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As efficiency is a crucial factor in TTT, we evaluate the frame per second (FPS) along with accuracy $(\\mathrm{mADE}_6)$ , shown in Fig. 5. We set the performance of the joint training method w/o adaptation as the benchmark and present MEK and TENT, which demonstrates competitive performance among the baselines. Our approach allows for the adjustment of the update frequency, with a frequency of 1 indicates updating at every opportunity, and 2 means updating every other opportunity. While frequent updates improve prediction performance, they also increase execution time; adjusting the update frequency allows for a balance between efficiency and accuracy. As shown in Fig. 5, our method outperforms in both accuracy and efficiency. Given that the time interval is 0.1 seconds, real-time execution requires a processing speed of at least 10 FPS. Even at the maximum update frequency of 1, our method maintains real-time capability with a superior accuracy of 0.39. When increasing the update frequency to 20, the error increases to 0.81, close to MEK's 0.75. However, the FPS reaches 24.2, demonstrating overwhelmingly faster operation compared to MEK's speed of 3.3 FPS.", + "bbox": [ + 496, + 518, + 890, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Qualitative results", + "text_level": 1, + "bbox": [ + 500, + 847, + 678, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison to the baselines: We compare our results with TENT and MEK in Fig. 4. While all methods, includ", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15070", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d085ef7d88ac5057af959f00a6e97261c08931baa7e5625dc918e537e76a8e66.jpg", + "image_caption": [ + "Figure 6. The first row indicates masked samples, and the row below shows the reconstructed outputs. The blue/red arrows indicate historical/future trajectories. The black arrows refer to the masked trajectories. The white lines are the lane centerlines, and the gray dashed lines are the masked lane centerlines." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 467, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c365f1fd0cd1174f5ab1788d0fbd97f4c6916990322d11398392d02b6c30732a.jpg", + "image_caption": [ + "Figure 7. Multi modal prediction results (blue arrows) before and after adaptation via our method. Ours generates elaborate samples that consider interaction between lane (above) or actors (below) due to representation learning, which cannot be learned from the GT (red arrow) via regression loss." + ], + "image_footnote": [], + "bbox": [ + 78, + 364, + 467, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing ours, perform multi-agent, multi-modal predictions, we only illustrate one actor and the closest mode to the GT for visual simplicity. The first row of the figure represents predictions before adaptation, and below are the results after adaptation using three different methods. MEK and TENT exhibit instances of underfitting or excessive overfitting upon adaptation, whereas our method consistently demonstrates stable and accurate predictions.", + "bbox": [ + 75, + 612, + 468, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reconstruction results: Reconstruction examples are depicted in Fig. 6. For all agents and lanes within a data sample, random masking is applied, as shown in the first row. During test-time training, learning for reconstruction is conducted, resulting in successful reconstruction for data with different distributions, as seen in the second row.", + "bbox": [ + 75, + 734, + 468, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-modal prediction results: As multi-modality is a crucial issue [10, 62, 64], we show that ours can handle multi-modal prediction results in Fig. 7. The adapted predictions showcase diverse yet plausible scenarios, either considering the lane structure (above) or surrounding agents", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6af54dae414e3be7f34aaa98918059a3c222e0d1a9c258ad896fa575d5f84300.jpg", + "table_caption": [ + "Table 2. Effect of type of losses to be optimized. Optimizing all losses shows optimal test-time adaptation performance." + ], + "table_footnote": [], + "table_body": "
Exp.Loss typemADE6 /mFDE6
Actor reconLane reconReg
INTER → Lyft (1/3/0.1)✓1.553 / 3.458
✓✓1.054 / 2.007
✓0.842 / 1.512
0.674 / 1.430
✓✓✓0.391 / 0.824
nuS → Lyft (2/6/0.5)✓1.108 / 2.597
0.987 / 2.304
✓✓0.973 / 2.280
✓0.942 / 2.262
✓✓✓0.776 / 1.820
", + "bbox": [ + 514, + 127, + 875, + 318 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5e4751f96ac34cbaaf31b457c513cea49af647416f9f3835b0ce88ef427a4444.jpg", + "table_caption": [ + "Table 3. Ablation on the depth of optimizing layer according to the loss types. The right side of the table indicates the optimization of deeper layers. Using only $\\mathcal{L}_{reg}$ deteriorates performance when optimizing all layers while ours stably improves as deeper." + ], + "table_footnote": [], + "table_body": "
LossOptimizing layers
DD+ED+E+f h,f,l
Lreg0.864 / 2.0720.840 / 2.0930.942 / 2.262
Lreg + Lrecon0.859 / 2.0600.813 / 1.9230.776 / 1.820
", + "bbox": [ + 504, + 383, + 883, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(below). These elaborated samples, although not present in the observed ground truth (GT) future, are learned through the representation learning from reconstruction loss. In addition, it shows that actor-specific tokens do not induce mode collapse to only one motion.", + "bbox": [ + 498, + 459, + 892, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Ablation", + "text_level": 1, + "bbox": [ + 500, + 550, + 598, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Reconstruction objective", + "text_level": 1, + "bbox": [ + 500, + 574, + 725, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table. 2 shows ablation studies on optimizing different loss types. Both reconstruction and regression losses individually boost prediction performance, with their joint optimization yielding even greater improvements. Table. 3 compares the effects of using only regression loss versus both losses on prediction performance across different layer depths. Updating just the decoder $(D)$ shows similar results in both scenarios, but extending updates to the encoder $(E)$ significantly enhances performance when using both losses. Furthermore, extending updates to the embedding layers $(f_{h,f,l})$ deteriorates performance when only regression loss is optimized. This highlights the importance of incorporating representation learning through the MAE, as relying solely on regression loss can lead to suboptimal adaptation and damage to learned representations.", + "bbox": [ + 496, + 598, + 892, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also conduct ablation studies on the masking ratio for both actors and lane centerlines in Tab. 4. The result is visualized via graphs in Fig. 8 according to lane masking ratio and actor masking ratio, respectively. Around 0.3 of lane and 0.4 of actor masking ratio, tendencies of $\\mathrm{mADE}_6$ follow", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15071", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/af9b92b9a8f8ce965deed41d189599df35e0b81e0728634df127a6c8b1e2f046.jpg", + "table_caption": [ + "Table 4. ${\\mathrm{{mADE}}}_{6}$ according to actor and lane masking ratio." + ], + "table_footnote": [], + "table_body": "
INTER → Lyft (1/3/0.1)Lane Masking Ratio
0.10.30.50.70.9
Actor0.10.4180.4070.4640.4810.446
0.30.4230.4430.4430.4450.390
Masking Ratio0.50.4170.4550.5150.3910.435
0.70.5670.4350.4480.4530.491
0.90.4470.4210.4940.4170.398
", + "bbox": [ + 84, + 107, + 460, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4e415c63103f80aa7a0e4edd6a56f39c46758c3de3bcadd677d9ae131bb984c3.jpg", + "image_caption": [ + "Figure 8. Tendency of $\\mathrm{mADE}_6$ according to actor and lane masking ratio respectively. (INTER $\\rightarrow$ Lyft $(1/3/0.1)$ )" + ], + "image_footnote": [], + "bbox": [ + 80, + 215, + 269, + 339 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/78156583d219c4af1bd5b431da4f43dc19101a43a9b0c1fbef7de4d95e1e9685.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 217, + 460, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a U-shape. In case of too small masking ratio, the reconstruction does not learn sufficient representation from the loss, while large masking ratio interrupt interaction learning due to absence of sufficient information. However, in both lane and actor masking, $\\mathrm{mADE}_6$ gets improved when the masking ratio increases above 0.8. In that case, the reconstruction network is induced to learn scene-specific information. In addition, unlike regression loss which deteriorates performance, reconstruction loss does not harm performance because it induces learning the semantic relationship than direct regression supervision.", + "bbox": [ + 75, + 383, + 467, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Actor-specific token", + "text_level": 1, + "bbox": [ + 76, + 560, + 264, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table. 5 presents results for the baseline without adaptation, our method without actor-specific tokens, and our full method. The second column reveals that even without actor-specific tokens, the prediction performance is 0.581 and 0.931, surpassing MEK's 0.746 and 1.006. However, incorporating actor-specific tokens for instance-aware adaptation yields notable improvements of $32.7\\%$ and $16.7\\%$ for short-term and long-term experiments, respectively. The difference in performance between short-term and long-term is influenced by the scene length in the dataset. The average scene length for short-term data with a 0.1 time interval is 200.04, significantly longer than the average of 32.67 for long-term data with a 0.5 time interval. Intuitively, as scene length increases, the time spent observing previously adapted actors also increases, enhancing the effectiveness of actor-specific tokens. To verify this, Fig. 9 adjusts scene length arbitrarily by skipping to the next scene in data loading once a specific scene length is exceeded. The results confirm that as scene length decreases by skipping scenes earlier, the effectiveness of actor-specific tokens diminishes in both short-term and long-term scenarios.", + "bbox": [ + 75, + 583, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9bf639bafce7eec560c8b5baa67176543b680f12cf0e6779153eb3e814c2ab71.jpg", + "table_caption": [ + "Table 5. Effect of actor-specific token in $\\mathrm{mADE}_6 / \\mathrm{mFDE}_6$ . The proposed method enhances adaptation performance by learning actor-wise motion characteristics." + ], + "table_footnote": [], + "table_body": "
Exp.BaselineOurs w/o Actor-specificOurs (Full)
INTER → Lyft (1/3/0.1)1.553 / 3.4580.581 / 1.1510.391 / 0.824
nuS → Lyft (2/6/0.5)1.108 / 2.5970.932 / 2.2200.776 / 1.820
", + "bbox": [ + 508, + 136, + 883, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/44509efe139cf590a75aa806d5674052ec86ef6891c09f3d29290fc0c408aca5.jpg", + "image_caption": [ + "Figure 9. Effect of scenario length to the effectiveness of actor-specific token. As the scenario length shortens with manual skipping, its effectiveness diminishes because the duration available for the actor-specific token to adapt is reduced. In real-world applications, driving scenarios are continuous, resulting in maximal efficacy of the proposed method." + ], + "image_footnote": [], + "bbox": [ + 508, + 232, + 887, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 522, + 617, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a test-time training method for trajectory prediction by incorporating the MAE and actor-specific token memory. The introduced MAE objective addresses a limitation of conventional online learning, preventing the loss of representations learned from source data. Consequently, our approach enables learning deeper layers, leading to improved representations and enhanced predictions even for out-of-distribution samples. The integration of actor-specific tokens during test-time allows for instance-wise learning of motion patterns, resulting in substantial performance improvements. This approach, particularly effective in continuous real-world autonomous driving scenarios without scene breaks, demonstrates significant efficacy and holds promise for practical applications.", + "bbox": [ + 496, + 547, + 890, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This research was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF2022R1A2B5B03002636) and the Challengeable Future Defense Technology Research and Development Program through the Agency For Defense Development (ADD) funded by the Defense Acquisition Program Administration (DAPA) in 2024 (No.912768601).", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15072", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Görkay Aydemir, Adil Kaan Akan, and Fatma Güney. Adapt: Efficient multi-agent trajectory prediction with adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8295-8305, 2023. 1", + "[2] Inhwan Bae, Jin-Hwi Park, and Hae-Gon Jeon. Nonprobability sampling network for stochastic human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6477-6487, 2022. 6", + "[3] Inhwan Bae, Jean Oh, and Hae-Gon Jeon. Eigentrajectory: Low-rank descriptors for multi-modal trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10017-10029, 2023. 2", + "[4] Collin Burns and Jacob Steinhardt. Limitations of post-hoc feature alignment for robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2525-2533, 2021. 2", + "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5", + "[6] Rohan Chandra, Uttaran Bhattacharya, Aniket Bera, and Dinesh Manocha. Traphic: Trajectory prediction in dense and heterogeneous traffic using weighted interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8483-8492, 2019. 1", + "[7] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 2", + "[8] Hao Chen, Jiaze Wang, Kun Shao, Furui Liu, Jianye Hao, Chenyong Guan, Guangyong Chen, and Pheng-Ann Heng. Traj-mae: Masked autoencoders for trajectory prediction. arXiv preprint arXiv:2303.06697, 2023. 2", + "[9] Liang Chen, Yong Zhang, Yibing Song, Ying Shan, and Lingqiao Liu. Improved test-time adaptation for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24172-24182, 2023. 1, 2", + "[10] Weihuang Chen, Zhigang Yang, Lingyang Xue, Jinghai Duan, Hongbin Sun, and Nanning Zheng. Multimodal pedestrian trajectory prediction using probabilistic proposal network. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 7", + "[11] Xuesong Chen, Shaoshuai Shi, Chao Zhang, Benjamin Zhu, Qiang Wang, Ka Chun Cheung, Simon See, and Hongsheng Li. Trajectoryformer: 3d object tracking transformer with predictive trajectory hypotheses. arXiv preprint arXiv:2306.05888, 2023. 2", + "[12] Yuxiao Chen, Boris Ivanovic, and Marco Pavone. Scept: Scene-consistent, policy-based trajectory predictions for planning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17103-17112, 2022. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Jie Cheng, Xiaodong Mei, and Ming Liu. Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8679-8689, 2023. 2, 3, 4", + "[14] Junhyeong Cho, Gilhyun Nam, Sungyeon Kim, Hunmin Yang, and Suha Kwak. Promptstyler: Prompt-driven style generation for source-free domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15702-15712, 2023. 2", + "[15] Dooseop Choi and KyoungWook Min. Hierarchical latent structure for multi-modal vehicle trajectory forecasting. In European Conference on Computer Vision, pages 129-145. Springer, 2022. 2", + "[16] Sehwan Choi, Jungho Kim, Junyong Yun, and Jun Won Choi. R-pred: Two-stage motion prediction via tube-query attention-based trajectory refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8525-8535, 2023. 1", + "[17] Henggang Cui, Vladan Radosavljevic, Fang-Chieh Chou, Tsung-Han Lin, Thi Nguyen, Tzu-Kuo Huang, Jeff Schneider, and Nemanja Djuric. Multimodal trajectory predictions for autonomous driving using deep convolutional networks. In 2019 International Conference on Robotics and Automation (ICRA), pages 2090-2096. IEEE, 2019. 2", + "[18] Ning Ding, Yixing Xu, Yehui Tang, Chao Xu, Yunhe Wang, and Dacheng Tao. Source-free domain adaptation via distribution estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7212-7222, 2022. 2", + "[19] Nemanja Djuric, Vladan Radosavljevic, Henggang Cui, Thi Nguyen, Fang-Chieh Chou, Tsung-Han Lin, Nitin Singh, and Jeff Schneider. Uncertainty-aware short-term motion prediction of traffic actors for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2095-2104, 2020. 2", + "[20] Yonghao Dong, Le Wang, Sanping Zhou, and Gang Hua. Sparse instance conditioned multimodal trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9763-9772, 2023. 2", + "[21] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 5", + "[22] Francois Fleuret et al. Uncertainty reduction for model adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9613-9623, 2021. 2", + "[23] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. Advances in Neural Information Processing Systems, 35:29374-29385, 2022. 2", + "[24] Thomas Gilles, Stefano Sabatini, Dzmitry Tsishkou, Bogdan Stanciulescu, and Fabien Moutarde. THOMAS: Trajectory heatmap output with learned multi-agent sampling. In In-" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15073", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ternational Conference on Learning Representations, 2022. 2", + "[25] Thomas Gilles, Stefano Sabatini, Dzmitry Tsishkou, Bogdan Stanciulescu, and Fabien Moutarde. Uncertainty estimation for cross-dataset performance in trajectory prediction. arXiv preprint arXiv:2205.07310, 2022. 1, 2", + "[26] Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. In International Conference on Learning Representations, 2022. 2", + "[27] Taesik Gong, Jongheon Jeong, Taewon Kim, Yewon Kim, Jinwoo Shin, and Sung-Ju Lee. NOTE: Robust continual test-time adaptation against temporal correlation. In Advances in Neural Information Processing Systems, 2022. 2", + "[28] Junru Gu, Chenxu Hu, Tianyuan Zhang, Xuanyao Chen, Yilun Wang, Yue Wang, and Hang Zhao. Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5496-5506, 2023. 1", + "[29] Abner Guzman-Rivera, Dhruv Batra, and Pushmeet Kohli. Multiple choice learning: Learning to produce multiple structured outputs. Advances in neural information processing systems, 25, 2012. 4", + "[30] John Houston, Guido Zuidhof, Luca Bergamini, Yawei Ye, Long Chen, Ashesh Jain, Sammy Omari, Vladimir Iglovikov, and Peter Ondruska. One thousand and one hours: Self-driving motion prediction dataset. In Conference on Robot Learning, pages 409-418. PMLR, 2021. 5", + "[31] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1", + "[32] Pingxuan Huang, Zhenhua Cui, Jing Li, Shenghua Gao, bo Hu, and Yanyan Fang. Cross-domain trajectory prediction with ctp-net, 2022. 1", + "[33] Manh Huynh and Gita Alaghband. AOL: Adaptive online learning for human trajectory prediction in dynamic video scenes. arXiv preprint arXiv:2002.06666, 2020. 2", + "[34] Manh Huynh and Gita Alaghband. Online adaptive temporal memory with certainty estimation for human trajectory prediction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 940-949, 2023. 2", + "[35] David Isele and Akansel Cosgun. Selective experience replay for lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2", + "[36] Boris Ivanovic, James Harrison, and Marco Pavone. Expanding the deployment envelope of behavior prediction via adaptive meta-learning. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7786-7793. IEEE, 2023. 1, 2, 5", + "[37] Boris Ivanovic, Guanyu Song, Igor Gilitschenski, and Marco Pavone. trajdata: A unified interface to multiple human trajectory datasets. In Proceedings of the Neural Information" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Processing Systems (NeurIPS) Track on Datasets and Benchmarks, New Orleans, USA, 2023. 5", + "[38] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9644-9653, 2023. 1", + "[39] Ruochen Jiao, Xiangguo Liu, Takami Sato, Qi Alfred Chen, and Qi Zhu. Semi-supervised semantics-guided adversarial training for robust trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8207-8217, 2023. 2", + "[40] Junho Kim, Inwoo Hwang, and Young Min Kim. Ev-tta: Test-time adaptation for event-based object recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17745-17754, 2022. 2", + "[41] Mihee Lee, Samuel S Sohn, Seonghyeon Moon, Sejong Yoon, Mubbasir Kapadia, and Vladimir Pavlovic. Musevae: multi-scale vae for environment-aware long term trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2221-2230, 2022. 2", + "[42] Lihuan Li, Maurice Pagnucco, and Yang Song. Graph-based spatial transformer with memory replay for multifuture pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2231-2241, 2022. 2", + "[43] Maosen Li, Siheng Chen, Yanning Shen, Genjia Liu, Ivor W Tsang, and Ya Zhang. Online multi-agent forecasting with interpretable collaborative graph neural networks. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2", + "[44] Yingwei Li, Charles R Qi, Yin Zhou, Chenxi Liu, and Dragomir Anguelov. Modar: Using motion forecasting for 3d object detection in point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9329-9339, 2023. 2", + "[45] Yushu Li, Xun Xu, Yongyi Su, and Kui Jia. On the robustness of open-world test-time training: Self-training with dynamic prototype expansion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11836-11846, 2023. 2", + "[46] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 2", + "[47] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. arXiv preprint arXiv:2303.15361, 2023. 2", + "[48] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecasting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 541-556. Springer, 2020. 4", + "[49] Hyesu Lim, Byeonggeun Kim, Jaegul Choo, and Sungha Choi. Ttn: A domain-shift aware batch normalization in test" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15074", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "time adaptation. In The Eleventh International Conference on Learning Representations, 2022. 2", + "[50] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808-21820, 2021. 2", + "[51] Yuejiang Liu, Riccardo Cadei, Jonas Schweizer, Sherwin Bahmani, and Alexandre Alahi. Towards robust and adaptive motion forecasting: A causal representation perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17081-17092, 2022. 1", + "[52] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5517-5526, 2023. 1", + "[53] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 5", + "[54] M Jehanzeb Mirza, Inkyu Shin, Wei Lin, Andreas Schriebl, Kunyang Sun, Jaesung Choe, Mateusz Kozinski, Horst Possegger, In So Kweon, Kuk-Jin Yoon, et al. Mate: Masked autoencoders are online 3d test-time learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16709-16718, 2023. 2", + "[55] Jiquan Ngiam, Vijay Vasudevan, Benjamin Caine, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, David J Weiss, Benjamin Sapp, Zhifeng Chen, and Jonathon Shlens. Scene transformer: A unified architecture for predicting future trajectories of multiple agents. In International Conference on Learning Representations, 2022. 1", + "[56] Daehee Park, Jaewoo Jeong, and Kuk-Jin Yoon. Improving transferability for cross-domain trajectory prediction via neural stochastic differential equation. arXiv preprint arXiv:2312.15906, 2023. 2", + "[57] Daehee Park, Hobin Ryu, Yunseo Yang, Jegyeong Cho, Jiwon Kim, and Kuk-Jin Yoon. Leveraging future relationship reasoning for vehicle trajectory prediction. In The Eleventh International Conference on Learning Representations, 2023. 1", + "[58] Mozhgan Pourkeshavarz, Changhe Chen, and Amir Rasouli. Learn tarot with mentor: A meta-learned self-supervised approach for trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8384–8393, 2023. 2", + "[59] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. Advances in neural information processing systems, 32, 2019. 2", + "[60] Luke Rowe, Martin Ethier, Eli-Henry Dykhne, and Krzysztof Czarnecki. Fjmp: Factorized joint multi-agent motion prediction over learned directed acyclic interaction" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13745-13755, 2023. 1", + "[61] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 2", + "[62] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Wei Tang, Nanning Zheng, and Gang Hua. Representing multimodal behaviors with mean location for pedestrian trajectory prediction. IEEE transactions on pattern analysis and machine intelligence, 2023. 7", + "[63] Liushuai Shi, Le Wang, Sanping Zhou, and Gang Hua. Trajectory unified transformer for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9675-9684, 2023. 2", + "[64] Jianhua Sun, Yuxuan Li, Hao-Shu Fang, and Cewu Lu. Three steps to multimodal trajectory prediction: Modality clustering, classification and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13250-13259, 2021. 7", + "[65] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In Proceedings of the 37th International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 2, 3", + "[66] Devavrat Tomar, Guillaume Vray, Behzad Bozortabar, and Jean-Philippe Thiran. Tesla: Test-time self-learning with automatic adversarial augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20341–20350, 2023. 2", + "[67] Li-Wu Tsao, Yan-Kai Wang, Hao-Siang Lin, Hong-Han Shuai, Lai-Kuan Wong, and Wen-Huang Cheng. Socialssl: Self-supervised cross-sequence representation learning based on transformers for multi-agent trajectory prediction. In European Conference on Computer Vision, pages 234-250. Springer, 2022. 2", + "[68] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[69] Chunnan Wang, Xiang Chen, Junzhe Wang, and Hongzhi Wang. Atpfl: Automatic trajectory prediction model design under federated learning framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6563-6572, 2022. 2", + "[70] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 5", + "[71] Eason Wang, Henggang Cui, Sai Yalamanchi, Mohana Moorthy, and Nemanja Djuric. Improving movement predictions of traffic actors in bird's-eye view models using gans and differentiable trajectory rasterization. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2340-2348, 2020. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15075", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] Jingke Wang, Tengju Ye, Ziqing Gu, and Junbo Chen. Ltp: Lane-based trajectory prediction for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17134-17142, 2022. 2", + "[73] Letian Wang, Yeping Hu, and Changliu Liu. Online adaptation of neural network models by modified extended kalman filter for customizable and transferable driving behavior prediction, 2022. 1, 2, 5", + "[74] Letian Wang, Yeping Hu, Liting Sun, Wei Zhan, Masayoshi Tomizuka, and Changliu Liu. Transferable and adaptable driving behavior prediction. arXiv preprint arXiv:2202.05140, 2022. 2", + "[75] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 2", + "[76] Zhibo Wang, Jiayu Guo, Haiqiang Zhang, Ru Wan, Junping Zhang, and Jian Pu. Bridging the gap: Improving domain generalization in trajectory prediction. IEEE Transactions on Intelligent Vehicles, 2023. 1", + "[77] Chenfeng Xu, Tian Li, Chen Tang, Lingfeng Sun, Kurt Keutzer, Masayoshi Tomizuka, Alireza Fathi, and Wei Zhan. Pretram: Self-supervised pre-training via connecting trajectory and map. In European Conference on Computer Vision, pages 34-50. Springer, 2022. 2", + "[78] Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1410-1420, 2023. 2", + "[79] Pei Xu, Jean-Bernard Hayet, and Ioannis Karamouzas. Socialvae: Human trajectory prediction using timewise latents. In European Conference on Computer Vision, pages 511-528. Springer, 2022. 2", + "[80] Yi Xu, Lichen Wang, Yizhou Wang, and Yun Fu. Adaptive trajectory prediction via transferable gnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6520-6531, 2022. 1, 2", + "[81] Yi Xu, Armin Bazarjani, Hyung-gun Chi, Chiho Choi, and Yun Fu. Uncovering the missing pattern: Unified framework towards trajectory imputation and prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9632-9643, 2023. 1", + "[82] Zhijie Yan, Pengfei Li, Zheng Fu, Shaocong Xu, Yongliang Shi, Xiaoxue Chen, Yuhang Zheng, Yang Li, Tianyu Liu, Chuxuan Li, et al. Int2: Interactive trajectory prediction at intersections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8536-8547, 2023. 1", + "[83] Tao Yang, Shenglong Zhou, Yuwang Wang, Yan Lu, and Nanning Zheng. Test-time batch normalization. arXiv preprint arXiv:2205.10210, 2022. 2", + "[84] Luyao Ye, Zikang Zhou, and Jianping Wang. Improving the generalizability of trajectory prediction models with freenet-based domain normalization. 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11562-11568, 2023. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[85] Jiangbei Yue, Dinesh Manocha, and He Wang. Human trajectory prediction via neural social physics. In European Conference on Computer Vision, pages 376-394. Springer, 2022. 2", + "[86] Wei Zhan, Liting Sun, Di Wang, Haojie Shi, Aubrey Clausse, Maximilian Naumann, Julius Kummerle, Hendrik Konigshof, Christoph Stiller, Arnaud de La Fortelle, et al. Interaction dataset: An international, adversarial and cooperative motion dataset in interactive driving scenarios with semantic maps. arXiv preprint arXiv:1910.03088, 2019. 5", + "[87] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems, 35: 38629-38642, 2022. 2", + "[88] Xingxuan Zhang, Peng Cui, Renzhe Xu, Linjun Zhou, Yue He, and Zheyan Shen. Deep stable learning for out-of-distribution generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5372-5382, 2021. 1", + "[89] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863-17873, 2023. 1", + "[90] Dekai Zhu, Guangyao Zhai, Yan Di, Fabian Manhardt, Hendrik Berkemeyer, Tuan Tran, Nassir Navab, Federico Tombari, and Benjamin Busam. Ipcc-tp: Utilizing incremental pearson correlation coefficient for joint multi-agent trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5507-5516, 2023. 1", + "[91] Yiyao Zhu, Di Luan, and Shaojie Shen. Biff: Bi-level future fusion with polyline-based coordinate for interactive trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8260-8271, 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 582 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "15076", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/11bec0da-9290-4827-bb16-e062c2aadc7c_model.json b/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/11bec0da-9290-4827-bb16-e062c2aadc7c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e71365a61a13ce82a66fa2f94ed73a7212370e60 --- /dev/null +++ b/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/11bec0da-9290-4827-bb16-e062c2aadc7c_model.json @@ -0,0 +1,2710 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.002, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.13, + 0.865, + 0.177 + ], + "angle": 0, + "content": "T4P: Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory" + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.204, + 0.807, + 0.238 + ], + "angle": 0, + "content": "Daehee Park, Jaeseok Jeong, Sung-Hoon Yoon, Jaewoo Jeong, and Kuk-Jin Yoon KAIST" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.242, + 0.735, + 0.258 + ], + "angle": 0, + "content": "{bag2824, jason. jeong, yoon307, jeong207, kjyoon}@kaist.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.474, + 0.705 + ], + "angle": 0, + "content": "Trajectory prediction is a challenging problem that requires considering interactions among multiple actors and the surrounding environment. While data-driven approaches have been used to address this complex problem, they suffer from unreliable predictions under distribution shifts during test time. Accordingly, several online learning methods have been proposed using regression loss from the ground truth of observed data leveraging the auto-labeling nature of trajectory prediction task. We mainly tackle the following two issues. First, previous works underfit and overfit as they only optimize the last layer of motion decoder. To this end, we employ the masked autoencoder (MAE) for representation learning to encourage complex interaction modeling in shifted test distribution for updating deeper layers. Second, utilizing the sequential nature of driving data, we propose an actor-specific token memory that enables the test-time learning of actor-wise motion characteristics. Our proposed method has been validated across various challenging cross-dataset distribution shift scenarios including nuScenes, Lyft, Waymo, and Interaction. Our method surpasses the performance of existing state-of-the-art online learning methods in terms of both prediction accuracy and computational efficiency. The code is available at https://github.com/daeheepark/T4P." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.738, + 0.21, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Trajectory prediction plays a significant role in autonomous systems, enhancing safety and navigation efficiency [28, 31]. Recently, data-driven methods have shown remarkable prediction capabilities [1, 16, 38, 52, 55, 57, 60, 81, 89, 90]; however, they are prone to distribution shift [9, 88]. Trajectory prediction models also produce unreliable output when faced shifts in the data distribution [25], posing significant risks in various real-world applications. This vulnerability stems from the ease with which the trajectory data distri" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.294, + 0.868, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.538, + 0.895, + 0.622 + ], + "angle": 0, + "content": "Figure 1. Previous methods optimize the last layer of the decoder using regression loss from delayed ground truth. Our method, on the other hand, learns representation via a masked autoencoder, which boosts prediction performance by optimizing deeper layers. In addition, the proposed actor-specific token enables the prediction model to learn actor-wise motion characteristics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.895, + 0.687 + ], + "angle": 0, + "content": "bution can be altered by numerous factors, such as scene changes and driving habits; i.e. road layout, interaction between agents, driver demographics [6, 51, 82]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.896, + 0.901 + ], + "angle": 0, + "content": "To address this challenge, recent methods have proposed domain adaptation and generalization strategies which aim to anticipate the distribution shifts and accordingly train the model [32, 76, 80]. However, due to the wide variety of factors influencing data distribution, the anticipated shifts may differ significantly from those encountered at test time. As a result, several online learning methods have been developed to dynamically adapt models during test time [36, 73]. Since trajectory prediction serves as an auto-labeling task where trajectory data is obtained from object tracking, the observed past and future trajectory provide both input and corresponding ground truth for supervision \\((\\mathcal{L}_{reg})\\), as depicted at the top right of Fig. 1. Nevertheless, updating the entire model may ruin the representation learned from the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15065" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "source data, so only part of the network, such as the batch normalization layer, is updated [27, 40, 61, 83]. Particularly, because regression loss is calculated at delayed timestamps and only a few samples are available during test time, this approach risks deteriorating the model's learned representation [73]. Therefore, previous online prediction methods mainly update only the last layer of the decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.474, + 0.53 + ], + "angle": 0, + "content": "In this work, we propose a test-time training (TTT) for trajectory prediction with two key aspects. First, we build a masked autoencoder (MAE) framework to adapt deep features, incorporating good representation that captures complex interactions between agents and road structures. Due to the challenge of existing online learning methods in damaging learned representations when updating deeper layers, we employ a MAE to guide representation learning. Second, we introduce an actor-specific token memory that has significant advantages in real-world driving scenario where data arrives continuously and sequentially. As each actor instance has its own driving habits and the past motion pattern of specific actors can be accessed from the arrived observations, we design a token memory in transformer structure [68] and its training strategy to learn actor-wise motion characteristics. The proposed TTT framework is validated on challenging cross-dataset distribution shift cases between nuScenes, Lyft, Waymo, and INTERACTION, and shows state-of-the-art performance surpassing previous online learning methods. Furthermore, we show the practicality of our TTT framework by evaluating its computational efficiency. We summarize our contributions as below:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.532, + 0.47, + 0.591 + ], + "angle": 0, + "content": "- We propose a test time training for trajectory prediction (T4P) by utilizing a masked autoencoder to learn deep feature representations that stably improve prediction performance across entire network layers." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.593, + 0.47, + 0.621 + ], + "angle": 0, + "content": "- We introduce an actor-specific token memory used to learn the different actor characteristics and habits." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.622, + 0.469, + 0.667 + ], + "angle": 0, + "content": "- Our method is validated across 4 different datasets as well as different temporal configurations. Ours achieves state-of-the-art performance both in accuracy and efficiency." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.532, + 0.47, + 0.667 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.685, + 0.227, + 0.7 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.28, + 0.727 + ], + "angle": 0, + "content": "2.1. Trajectory Prediction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Trajectory prediction garners attention with the emergence of methods that can enhance perception or planning [11, 12, 44]. Its goal is to predict future trajectories of traffic actors based on their historical trajectories and the context of their environment [3, 17, 19, 20, 39, 58, 63, 78, 91]. Historical trajectories, or tracklets of traffic actors, are sequentially acquired via vehicle detection and tracking systems. Some studies employed this temporal property to enhance prediction via memory replay [35, 42, 59]. In the early stages of trajectory prediction, only the historical trajectory of the actors of interest was considered. However, recent stud" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.244 + ], + "angle": 0, + "content": "ies emphasize the significance of understanding interactions among agents [67, 85] and the rules governed by surrounding environments [72, 77] in improving prediction performance. This has led to the development of models that incorporate multi-head attention or graph-based methods to capture these interactions [24, 26, 42]. Additionally, MAE has been adopted for pretraining to better understand agent interactions [8, 13]. To further refine prediction capabilities, various generative models have been introduced, enabling the generation of future trajectories [15, 41, 71, 79]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.254, + 0.868, + 0.27 + ], + "angle": 0, + "content": "2.2. Transfer Learning in Trajectory Prediction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.595 + ], + "angle": 0, + "content": "With data-driven approaches offering superior performance in trajectory prediction, their effectiveness diminishes under distribution shifts [25, 56]. In response, several studies have adopted domain adaptation or generalization strategy [69, 74]. Some specifically aimed to reduce the domain gap within unique characteristics of trajectory prediction: differences in road structures [84], actor interaction [80], etc. However, these methods rely on anticipating how to cover domain shifts. Yet, given that trajectory data is subject to influence from numerous factors, predicting and accommodating for shifts may not always be sufficient. Consequently, recent developments have introduced adaptation to unseen test sets using online learning [33, 34, 43]. Among them, some methods [36, 73] showed remarkable prediction performance improvement under severe distribution shifts like cross-dataset cases by utilizing regression loss for online learning. These methods exploit the fact that the input and ground truth (GT) labels are provided at test time as tracking history. However, with the limitations of updating with only a few samples in a delayed time, adaptation becomes restricted to the last layer of the decoder." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.606, + 0.685, + 0.621 + ], + "angle": 0, + "content": "2.3. Test Time Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Test-time training (TTT) is a method that trains the network on test time data, unseen during training [4, 7, 14, 18, 22, 49, 66, 75]. Unlike domain generalization or adaptation, which are confined to the training phase, TTT extends model adaptation into the test phase by utilizing available test data [47]. TTT methods are categorized into regularization-based approaches for post-hoc regularization of out-of-distribution (OOD) samples [46, 87], and self-supervised approaches that employ pretext tasks on test data for optimal representation learning [7, 9, 45, 50, 54]. Specifically, TTT [65] introduced a Y-shaped network structure consisting of a feature encoder, a pretext branch, and a decoder branch. The decoder branch is fixed, while the encoder and pretext branch are optimized through self-supervision. Adhering to this model, TTT-MAE [23] integrated a MAE in the pretext task. Expanding on this method, we adopt TTT-MAE to the domain of trajectory prediction, leveraging its representation learning capabilities to enhance test-time training." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15066" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.093, + 0.887, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.339, + 0.892, + 0.396 + ], + "angle": 0, + "content": "Figure 2. Overall method. During test-time training, the network trained on source dataset is optimized on target data under online setting. The model is optimized both from regression and reconstruction loss. Both losses utilize the data observed at the delayed time stamp \\((t_{\\tau})\\). Actor-specific token is used to learn instance-wise motion pattern during test-time training phase. During online evaluation phase, model and actor-specific token learned from test-time training phase are used." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.411, + 0.168, + 0.426 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.436, + 0.258, + 0.451 + ], + "angle": 0, + "content": "3.1. Problem definition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.642 + ], + "angle": 0, + "content": "Trajectory prediction aims to learn the mapping function between the input, consisting of historical trajectory and map information \\(\\mathbf{X}:\\{\\mathcal{X}_t,\\mathcal{M}_t\\}\\), and the output, consisting of \\(K\\) possible candidates for future trajectory of \\(N\\) traffic actors, \\(\\mathbf{Y}:\\{\\mathcal{Y}_t^{0:K - 1}\\}\\), at current time \\(t\\). We predict \\(C\\) different actor classes including vehicle, cyclist, etc. Historical and future trajectories are represented as \\(\\mathcal{X}_t = \\mathbf{x}_{t - t_h:t}^{0:N - 1}\\) and \\(\\mathcal{Y}_t = \\mathbf{x}_{t:t + t_f}^{0:N - 1}\\) where \\(t_h\\) and \\(t_f\\) represent sequence length of input and output trajectory. Here, \\(\\mathbf{x}_t^n\\) represents the spatial location of actor \\(n\\) at time \\(t\\). For map information \\(\\mathcal{M}_t\\), we use \\(L\\) segmented lane centerlines around ego-actors which is widely-used in trajectory prediction methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.471, + 0.841 + ], + "angle": 0, + "content": "We deal with the case when the target data distribution during test time \\(\\{\\mathbf{X},\\mathbf{Y}\\}^T\\) is different from the source data distribution seen during the training phase \\(\\{\\mathbf{X},\\mathbf{Y}\\}^S\\). We formulate the problem as a online adaptation scenario in which one data sample is given per each time interval as time passes. The test data is consists of multiple distinct scenes. Each scene includes temporally ordered data samples which are captured through real-world driving. Following standard TTT methods, there is no access to the source data at test time. However, thanks to the auto-labeling nature of trajectory prediction, there is access to delayed GT future trajectory \\((\\mathbf{x}_{t_{\\tau}:t_{\\tau} + t_f})\\) from a previous time window at time \\(t_\\tau (= t - \\tau)\\) as depicted in left lower corner of Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.236, + 0.862 + ], + "angle": 0, + "content": "3.2. Overall method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Our method, Test-Time Training of Trajectory Prediction (T4P), enhances the online learning method using supervi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.892, + 0.58 + ], + "angle": 0, + "content": "sion from a delayed GT future trajectory with representation learning from MAE and actor-specific token memory. Following standard TTT [65], the overall framework consists of three phases: offline training, test-time training, and online evaluation. Offline training occurs before test time, and test-time training and online evaluation are executed repeatedly and sequentially during test time. We adopt the ForecastMAE [13] backbone consisting of embedding layers \\( f \\), a shared encoder \\( E \\), a reconstruction head \\( R \\) and a motion decoder head \\( D \\), as depicted in the middle of Fig. 2. The detailed methods during each phase are described below:" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.588, + 0.658, + 0.604 + ], + "angle": 0, + "content": "3.3. Offline training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.891, + 0.642 + ], + "angle": 0, + "content": "During offline training, the model is trained on source data using both reconstruction loss and regression loss." + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.652, + 0.891, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta \\in \\{f, E, R, D \\}} \\mathbb {E} _ {\\mathcal {X}, \\mathcal {Y}, \\mathcal {M} \\in \\{\\mathbf {X}, \\mathbf {Y} \\} ^ {S}} \\left[ \\mathcal {L} _ {\\text {r e c o n}} + \\mathcal {L} _ {\\text {r e g}} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.807 + ], + "angle": 0, + "content": "In this subsection, subscript \\(t\\) is omitted for simplicity. First, all input elements \\((\\mathcal{X},\\mathcal{Y},\\mathcal{M})\\) are embedded with their respective embedding layer \\((f_x,f_y,f_m)\\). Additionally, we define actor class token \\(\\bar{\\alpha}\\in \\mathbb{R}^{C\\times D}\\) that learns different motion patterns of each actor class. The actor class token is implemented as a learnable embedding of a transformer structure. Corresponding actor class token \\(\\alpha (c)\\) is added to trajectory embedding according to the class of each actor." + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.814, + 0.891, + 0.831 + ], + "angle": 0, + "content": "\\[\nh _ {x}, h _ {y}, h _ {m} = f _ {x} (\\mathcal {X}) + \\alpha (c), f _ {y} (\\mathcal {Y}) + \\alpha (c), f _ {m} (\\mathcal {M}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.903 + ], + "angle": 0, + "content": "For reconstruction, the history/future trajectory and lane embeddings are fed to the encoder to obtain the encodings \\( F_{x}, F_{y}, F_{m} \\). Then, segments of the encodings are randomly masked and replaced with masking tokens, \\( M_{x}, M_{y}, M_{m} \\)," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15067" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.182 + ], + "angle": 0, + "content": "and the other encodings remain unmasked \\((F_x', F_y', F_m')\\). Here, we use random masking for the lane centerline and complementary masking strategy for history/future trajectory following previous works [13]. The masking tokens and the unmasked encodings are fed to the reconstructor to reconstruct the masked elements." + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.189, + 0.47, + 0.214 + ], + "angle": 0, + "content": "\\[\nF _ {x}, F _ {y}, F _ {m} = E \\left(h _ {x}, h _ {y}, h _ {m}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.211, + 0.415, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {X}}, \\hat {\\mathcal {Y}}, \\hat {\\mathcal {M}} = R \\left(M _ {x}, M _ {y}, M _ {m}, F _ {x} ^ {\\prime}, F _ {y} ^ {\\prime}, F _ {m} ^ {\\prime}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.236, + 0.47, + 0.327 + ], + "angle": 0, + "content": "The encoder and reconstructor both consist of multi-head attention to utilize interaction between history, future and lanes, thus, the reconstruction guides to the model have the capability of interaction reasoning. Finally, a reconstruction loss \\(\\mathcal{L}_{\\text{recon}}\\) is computed as the MSE loss between the ground truth and the reconstructed outputs." + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.333, + 0.469, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{N} \\sum_ {n} (\\mathcal {X} - \\hat {\\mathcal {X}}) ^ {2} + \\frac {1}{N} \\sum_ {n} (\\mathcal {Y} - \\hat {\\mathcal {Y}}) ^ {2} + \\frac {1}{L} \\sum_ {l} (\\mathcal {M} - \\hat {\\mathcal {M}}) ^ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.378, + 0.47, + 0.468 + ], + "angle": 0, + "content": "For the decoder head, historical trajectory and lanes embeddings are again fed to the same encoder and the output encodings are passed to the motion decoder, composed of MLP layers. The decoder outputs \\( \\mathrm{K} \\) candidates for trajectory prediction, and the regression loss \\( \\mathcal{L}_{reg} \\) is computed with the widely-used Winner-takes-all (WTA) loss [29, 48]." + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.479, + 0.319, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {Y}} ^ {0: K - 1} = D \\left(E \\left(h _ {h}, h _ {l}\\right)\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.5, + 0.469, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e g} = \\frac {1}{N} \\sum_ {n} \\operatorname {a r g m i n} _ {k \\in K} \\left(\\mathcal {Y} ^ {n} - \\hat {\\mathcal {Y}} ^ {n, k}\\right) ^ {2} \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.543, + 0.252, + 0.559 + ], + "angle": 0, + "content": "3.4. Test-time training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.47, + 0.671 + ], + "angle": 0, + "content": "During test-time, a data sample consisting of trajectories and maps arrives sequentially. Therefore, even though we cannot access the GT future trajectory of current time \\((\\mathcal{V}_t)\\), we can access both the inputs and GT \\((\\mathcal{X}_{t_{\\tau}},\\mathcal{Y}_{t_{\\tau}},\\mathcal{M}_{t_{\\tau}})\\) at a previous time \\(t_\\tau\\). With this data, the model is optimized with the same objective as Eq. 1 with target data distribution instead of source data distribution." + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.68, + 0.469, + 0.702 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta \\in \\{f, E, R, D \\}} \\mathbb {E} _ {\\chi_ {t _ {\\tau}}, \\mathcal {Y} _ {t _ {\\tau}}, \\mathcal {M} _ {t _ {\\tau}} \\in \\{\\mathbf {X}, \\mathbf {Y} \\} ^ {T}} \\left[ \\mathcal {L} _ {\\text {r e c o n}} + \\mathcal {L} _ {\\text {r e g}} \\right] \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.708, + 0.469, + 0.814 + ], + "angle": 0, + "content": "Unlike existing online learning methods that only utilize regression loss, we incorporate an additional reconstruction loss. This enables the model to learn a good representation that considers the complex actor-actor and actor-lane interaction even in the unseen target data distribution. An advantage of representation learning is that the performance stably improves even when the deeper layers are optimized." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.332, + 0.848 + ], + "angle": 0, + "content": "3.4.1 Actor-specific token memory" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Unlike during the offline training phase, when the data order is shuffled, data at test-time comes in sequentially. Therefore, at test time, it is possible to keep track of movement" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.093, + 0.885, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.3, + 0.892, + 0.384 + ], + "angle": 0, + "content": "Figure 3. Actor-specific token memory is colored in gray. It evolves as time passes within a scene. For newborn actors, the corresponding class token is registered. Until the actor disappears, the token is updated through test-time training. At the end of the scene, all tokens are averaged by each class and passed to the next scene as denoted in red arrow and Eq. 7." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.397, + 0.891, + 0.427 + ], + "angle": 0, + "content": "patterns of a specific actor instance. Using this, we propose an actor-specific token memory." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.431, + 0.892, + 0.704 + ], + "angle": 0, + "content": "The overall scheme is described in Fig. 3. During offline training, actor class tokens \\(\\bar{\\alpha}_{train} \\in \\mathbb{R}^{C \\times D}\\) are trained to reflect the average motion pattern of each of the \\(C\\) classes. At the beginning of the test-time, scene 0, class tokens are initialized from that of training phase (\\(\\bar{\\alpha}_{\\text{scene}(0)} \\gets \\bar{\\alpha}_{\\text{train}}\\)). When a new \\(n^{th}\\)-actor appears at time \\(t\\), the class token \\(\\alpha_{n}^{t}(c)\\) is cloned from \\(\\bar{\\alpha}_{\\text{scene}(0)}\\) by selecting corresponding class. The newborn tokens are then registered to the actor-specific token memory. The token memory is structured as a dictionary where actor instance ID/corresponding tokens are key/values. At each iteration, the actor-specific tokens are used for both test-time training and online evaluation. As time progresses, the actor-specific tokens evolve and are updated through the reconstruction and regression losses until the actor disappears in the scene. By giving each actor its own specific token that distinguishes it from the sharing of other parts of models with other actors, actor-specific motion patterns can be learned." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "When the scene changes, scene 1, the actors observed during scene 0 are not to be observed anymore, so we need a new averaged actor class token \\(\\bar{\\alpha}_{\\text{scene}(1)}\\). For that, we average all the tokens in the memory at the final time step \\(T_{S}\\) of scene 0 after gathering by classes as Eq. 7. Here, \\(N_{c}\\) refers to the number of actors of class \\(c\\). It is because, with a sufficient number of actors class tokens per each class in memory, their average motion can be a representative motion pattern of actor classes. This is more useful than \\(\\bar{\\alpha}_{\\text{train}}\\) because newly averaged tokens are trained on the target dataset while \\(\\bar{\\alpha}_{\\text{train}}\\) contains motion pattern trained on the source dataset. The averaged class tokens are then passed to the next scene and used to initialize tokens for the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15068" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.894, + 0.119 + ], + "angle": 0, + "content": "Table 1. Adaptation results in various distribution shifts. The model is trained on source dataset, and test-time trained and evaluated on target dataset (Source \\(\\rightarrow\\) Target). All metrics are better in lower value. The best and second-best results are marked in bold and underline." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.123, + 0.892, + 0.266 + ], + "angle": 0, + "content": "
mADE6/mFDE6Short-term exp (1/3/0.1)Long-term exp (2/6/0.5)
INTER → nuSINTER → LyftnuS → WayMeannuS → LyftWay → LyftWay → nuSMean
Source Only1.047 / 2.2471.391 / 2.9450.431 / 1.0310.956 / 2.0741.122 / 2.5770.621 / 1.3471.153 / 2.2200.965 / 2.048
Joint Training1.116 / 2.4451.553 / 3.4580.472 / 1.1251.047 / 2.3431.108 / 2.5970.638 / 1.4041.091 / 2.0310.946 / 2.011
DUA1.118 / 2.4551.516 / 3.3520.516 / 1.2941.050 / 2.3671.365 / 3.2570.790 / 1.8681.270 / 2.6341.142 / 2.586
TENT (w/ sup)1.102 / 2.4231.519 / 3.4050.448 / 1.0711.023 / 2.3001.068 / 2.5140.628 / 1.3811.077 / 2.0120.924 / 1.969
MEK (τ = tf/2)1.012 / 2.4451.283 / 3.4580.445 / 1.1250.913 / 2.3431.079 / 2.5970.629 / 1.4041.079 / 2.0310.929 / 2.011
MEK (τ = tf)0.892 / 1.9520.746 / 1.6540.405 / 1.0610.691 / 1.5561.006 / 2.3690.615 / 1.3511.117 / 2.1400.913 / 1.953
AML (K0)2.093 / 4.6972.695 / 6.6771.624 / 2.1392.137 / 4.5041.787 / 3.0671.322 / 2.5711.618 / 2.9991.576 / 2.879
AML (full)1.149 / 2.5501.042 / 2.6160.764 / 1.7910.985 / 2.3191.462 / 2.5730.977 / 2.1841.495 / 2.9781.311 / 2.578
Ours (T4P)0.537 / 1.1370.391 / 0.8240.336 / 0.8070.421 / 0.9230.776 / 1.8200.549 / 1.1710.996 / 1.7840.774 / 1.592
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.279, + 0.471, + 0.34 + ], + "angle": 0, + "content": "newborn actors. Please note that while scene 0 is initialized by actor class tokens from the training phase, subsequent scenes obtain as in Eq. 7. More details of memory evolving strategies can be found on the supplementary material." + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.345, + 0.47, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\bar {\\alpha} _ {s c e n e (i + 1)} \\leftarrow \\left\\{\\frac {1}{N _ {c}} \\sum_ {n} ^ {N _ {c}} \\alpha_ {n} ^ {T} (c) \\right\\} _ {s c e n e (i)} \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.395, + 0.251, + 0.41 + ], + "angle": 0, + "content": "3.5. Online evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.47, + 0.495 + ], + "angle": 0, + "content": "Using the updated model weight and actor-specific token memory during test-time training, online evaluation is executed. With the input data \\((\\mathcal{X}_t,\\mathcal{M}_t)\\) at current time \\(t\\), the learned encoder and decoder predict multi-modal trajectory \\((\\mathcal{Y}_t)\\) for all actors in the sample." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.506, + 0.203, + 0.523 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.53, + 0.181, + 0.545 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.766 + ], + "angle": 0, + "content": "We conducted experiments on well-known datasets, nuScenes [5], Lyft [30], WOMD [21], and INTERACTION [86], to evaluate T4P on various data distribution shifts. These datasets are parsed into the same format using trajdata [37]. Additionally, to verify in various prediction configurations, experiments were conducted with the two most widely used configurations of long-term and short-term prediction. Long-term prediction requires predicting 6 seconds into the future given 2 seconds of the past with a time interval of 0.5s, making the input/output sequence lengths to be 5 and 12, respectively. Short-term prediction requires predicting 3 seconds into the future given 0.9 seconds of the past with a time interval of 0.1s, making the input/output sequence lengths to be 10 and 30, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.292, + 0.789 + ], + "angle": 0, + "content": "4.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "For actor classes, we use the 5 classes: unknown, vehicle, pedestrian, bicycle and motorcycle. Our method predicts \\( K = 6 \\) future candidates for all actors in the sample. We use \\( \\tau \\) as \\( t_f \\) to enable the past GT future to contain full prediction horizon. We train and evaluate our model with a single NVIDIA A6000. Learning rates of model weight and actor-specific parameters are set as 0.01 and 0.5, respectively, and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.892, + 0.34 + ], + "angle": 0, + "content": "weight decay is set to 0.001 for all. The gradient is clipped by 15. For metrics, widely used \\(\\mathrm{mADE}_6\\) and \\(\\mathrm{mFDE}_6\\) are used. Detailed metric definition, model architecture, and training details are included in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.351, + 0.608, + 0.367 + ], + "angle": 0, + "content": "4.3. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.376, + 0.892, + 0.436 + ], + "angle": 0, + "content": "We compare our \\(T4P\\) with several baselines, including unsupervised/supervised test-time-training methods and online learning trajectory prediction methods. All baseline methods are implemented using the same backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.468 + ], + "angle": 0, + "content": "Source only refers to the backbone model trained on the source dataset only using regression loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.892, + 0.499 + ], + "angle": 0, + "content": "Joint training is similar to source only but trained with regression and reconstruction loss jointly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.545 + ], + "angle": 0, + "content": "DUA [53] is an unsupervised post-hoc regularization method only updates batch normalization statistics in a momentum-updating manner without back-propagation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.606 + ], + "angle": 0, + "content": "TENT with supervision is a variant of the original TENT [70] in which regression loss is used to optimize the batch normalization layers instead of entropy minimization loss, as entropy minimization is not applicable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.892, + 0.685 + ], + "angle": 0, + "content": "MEK [73] is an online learning trajectory prediction method utilizing the Modified Extended Kalman filter. It uses only regression loss to optimize the last layer of the decoder. As the prediction horizon is different in our experiment from the original paper, we use both \\(\\frac{1}{2} t_f\\) and \\(t_f\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.684, + 0.892, + 0.775 + ], + "angle": 0, + "content": "AML [36] is an Adaptive Meta-learning method. Unlike the other methods that use the same backbone, AML replaces the last decoder layer with a Bayesian linear regression layer for adaptive training. The modified version of backbone without adaptive training is denoted as \\( K_{0} \\), while the full version with adaptive training is denoted as full." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.791, + 0.587, + 0.807 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.688, + 0.832 + ], + "angle": 0, + "content": "5.1.Quantitative results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The results of comparing our method with the baselines in various distribution shift scenarios are presented in Tab. 1. We reported three distribution shift scenarios per each time configuration on the table, and other results are included in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15069" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.087, + 0.889, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.276, + 0.893, + 0.332 + ], + "angle": 0, + "content": "Figure 4. The first row shows prediction before adaptation, and the second row indicates adaptation results by three methods: ours (blue), TENT w/ sup (orange) and MEK (green). Sky blue and orange boxes refer to surrounding actors and actors to be predicted. We depicted only one actor result and one mode among multi-modal predictions closest to the GT for visual simplicity. Please note that our method is multi-modal prediction for all actors method." + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.343, + 0.424, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.502, + 0.47, + 0.558 + ], + "angle": 0, + "content": "Figure 5. Prediction accuracy and execution time on INTER \\(\\rightarrow\\) nuS (1/3/0.1) experiment. Adjusting update frequency can balance between accuracy and efficiency. Our method significantly outperforms the baseline methods in both accuracy and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.47, + 0.603 + ], + "angle": 0, + "content": "the supplementary material. Notably, our approach consistently surpassed baseline performance across all scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.761 + ], + "angle": 0, + "content": "DUA consistently exhibits compromised performance across nearly all cases, due to the distinctive characteristics of trajectory data. In contrast to image data, where data is treated as a singular sample, trajectory data involves multiple agents, each exhibiting distinct motion patterns, within a single data. Consequently, holistically updating batch statistics proves to be counterproductive. Similar challenges are encountered by TENT w/ sup. While regression loss prevents performance decline, updating only the batch norm layer has little to no effect on the prediction performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.47, + 0.868 + ], + "angle": 0, + "content": "MEK exhibited the most competitive performance among the baselines. While performance improved significantly in short-term settings, it brings limited improvement in long-term cases. As the Kalman filter updates based on the number of prediction steps, short-term configurations with 12 update steps show a better improvement than long-term configurations with only 5 update steps." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Although AML led to a considerable improvement in the full version, the predictive performance itself was substan" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.341, + 0.892, + 0.478 + ], + "angle": 0, + "content": "tially degraded due to the significant performance drop in the modified backbone \\((K_0)\\). The limitation of the backbone is due to the Bayesian regression layer being based on probability sampling which is known to be worse than non-probability sampling method of ours [2]. In contrast to all the baseline methods, our method demonstrated state-of-the-art performance in all scenarios featuring various distribution shifts, whether short-term or long-term, showcasing the generalizability of our approach." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.496, + 0.625, + 0.512 + ], + "angle": 0, + "content": "5.1.1 Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.837 + ], + "angle": 0, + "content": "As efficiency is a crucial factor in TTT, we evaluate the frame per second (FPS) along with accuracy \\((\\mathrm{mADE}_6)\\), shown in Fig. 5. We set the performance of the joint training method w/o adaptation as the benchmark and present MEK and TENT, which demonstrates competitive performance among the baselines. Our approach allows for the adjustment of the update frequency, with a frequency of 1 indicates updating at every opportunity, and 2 means updating every other opportunity. While frequent updates improve prediction performance, they also increase execution time; adjusting the update frequency allows for a balance between efficiency and accuracy. As shown in Fig. 5, our method outperforms in both accuracy and efficiency. Given that the time interval is 0.1 seconds, real-time execution requires a processing speed of at least 10 FPS. Even at the maximum update frequency of 1, our method maintains real-time capability with a superior accuracy of 0.39. When increasing the update frequency to 20, the error increases to 0.81, close to MEK's 0.75. However, the FPS reaches 24.2, demonstrating overwhelmingly faster operation compared to MEK's speed of 3.3 FPS." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.679, + 0.864 + ], + "angle": 0, + "content": "5.2. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparison to the baselines: We compare our results with TENT and MEK in Fig. 4. While all methods, includ" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15070" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.468, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.289, + 0.472, + 0.358 + ], + "angle": 0, + "content": "Figure 6. The first row indicates masked samples, and the row below shows the reconstructed outputs. The blue/red arrows indicate historical/future trajectories. The black arrows refer to the masked trajectories. The white lines are the lane centerlines, and the gray dashed lines are the masked lane centerlines." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.366, + 0.468, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.527, + 0.471, + 0.597 + ], + "angle": 0, + "content": "Figure 7. Multi modal prediction results (blue arrows) before and after adaptation via our method. Ours generates elaborate samples that consider interaction between lane (above) or actors (below) due to representation learning, which cannot be learned from the GT (red arrow) via regression loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.734 + ], + "angle": 0, + "content": "ing ours, perform multi-agent, multi-modal predictions, we only illustrate one actor and the closest mode to the GT for visual simplicity. The first row of the figure represents predictions before adaptation, and below are the results after adaptation using three different methods. MEK and TENT exhibit instances of underfitting or excessive overfitting upon adaptation, whereas our method consistently demonstrates stable and accurate predictions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.824 + ], + "angle": 0, + "content": "Reconstruction results: Reconstruction examples are depicted in Fig. 6. For all agents and lanes within a data sample, random masking is applied, as shown in the first row. During test-time training, learning for reconstruction is conducted, resulting in successful reconstruction for data with different distributions, as seen in the second row." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Multi-modal prediction results: As multi-modality is a crucial issue [10, 62, 64], we show that ours can handle multi-modal prediction results in Fig. 7. The adapted predictions showcase diverse yet plausible scenarios, either considering the lane structure (above) or surrounding agents" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.894, + 0.119 + ], + "angle": 0, + "content": "Table 2. Effect of type of losses to be optimized. Optimizing all losses shows optimal test-time adaptation performance." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.128, + 0.877, + 0.319 + ], + "angle": 0, + "content": "
Exp.Loss typemADE6 /mFDE6
Actor reconLane reconReg
INTER → Lyft (1/3/0.1)✓1.553 / 3.458
✓✓1.054 / 2.007
✓0.842 / 1.512
0.674 / 1.430
✓✓✓0.391 / 0.824
nuS → Lyft (2/6/0.5)✓1.108 / 2.597
0.987 / 2.304
✓✓0.973 / 2.280
✓0.942 / 2.262
✓✓✓0.776 / 1.820
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.326, + 0.894, + 0.383 + ], + "angle": 0, + "content": "Table 3. Ablation on the depth of optimizing layer according to the loss types. The right side of the table indicates the optimization of deeper layers. Using only \\(\\mathcal{L}_{reg}\\) deteriorates performance when optimizing all layers while ours stably improves as deeper." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.385, + 0.885, + 0.445 + ], + "angle": 0, + "content": "
LossOptimizing layers
DD+ED+E+f h,f,l
Lreg0.864 / 2.0720.840 / 2.0930.942 / 2.262
Lreg + Lrecon0.859 / 2.0600.813 / 1.9230.776 / 1.820
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.46, + 0.893, + 0.537 + ], + "angle": 0, + "content": "(below). These elaborated samples, although not present in the observed ground truth (GT) future, are learned through the representation learning from reconstruction loss. In addition, it shows that actor-specific tokens do not induce mode collapse to only one motion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.551, + 0.599, + 0.566 + ], + "angle": 0, + "content": "6. Ablation" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.727, + 0.592 + ], + "angle": 0, + "content": "6.1. Reconstruction objective" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.893, + 0.825 + ], + "angle": 0, + "content": "Table. 2 shows ablation studies on optimizing different loss types. Both reconstruction and regression losses individually boost prediction performance, with their joint optimization yielding even greater improvements. Table. 3 compares the effects of using only regression loss versus both losses on prediction performance across different layer depths. Updating just the decoder \\((D)\\) shows similar results in both scenarios, but extending updates to the encoder \\((E)\\) significantly enhances performance when using both losses. Furthermore, extending updates to the embedding layers \\((f_{h,f,l})\\) deteriorates performance when only regression loss is optimized. This highlights the importance of incorporating representation learning through the MAE, as relying solely on regression loss can lead to suboptimal adaptation and damage to learned representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We also conduct ablation studies on the masking ratio for both actors and lane centerlines in Tab. 4. The result is visualized via graphs in Fig. 8 according to lane masking ratio and actor masking ratio, respectively. Around 0.3 of lane and 0.4 of actor masking ratio, tendencies of \\(\\mathrm{mADE}_6\\) follow" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15071" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.096, + 0.09, + 0.45, + 0.104 + ], + "angle": 0, + "content": "Table 4. \\( {\\mathrm{{mADE}}}_{6} \\) according to actor and lane masking ratio." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.108, + 0.462, + 0.214 + ], + "angle": 0, + "content": "
INTER → Lyft (1/3/0.1)Lane Masking Ratio
0.10.30.50.70.9
Actor0.10.4180.4070.4640.4810.446
0.30.4230.4430.4430.4450.390
Masking Ratio0.50.4170.4550.5150.3910.435
0.70.5670.4350.4480.4530.491
0.90.4470.4210.4940.4170.398
" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.217, + 0.27, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.218, + 0.462, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.344, + 0.468, + 0.372 + ], + "angle": 0, + "content": "Figure 8. Tendency of \\(\\mathrm{mADE}_6\\) according to actor and lane masking ratio respectively. (INTER \\(\\rightarrow\\) Lyft \\((1/3/0.1)\\))" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.468, + 0.552 + ], + "angle": 0, + "content": "a U-shape. In case of too small masking ratio, the reconstruction does not learn sufficient representation from the loss, while large masking ratio interrupt interaction learning due to absence of sufficient information. However, in both lane and actor masking, \\(\\mathrm{mADE}_6\\) gets improved when the masking ratio increases above 0.8. In that case, the reconstruction network is induced to learn scene-specific information. In addition, unlike regression loss which deteriorates performance, reconstruction loss does not harm performance because it induces learning the semantic relationship than direct regression supervision." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.561, + 0.266, + 0.576 + ], + "angle": 0, + "content": "6.2. Actor-specific token" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Table. 5 presents results for the baseline without adaptation, our method without actor-specific tokens, and our full method. The second column reveals that even without actor-specific tokens, the prediction performance is 0.581 and 0.931, surpassing MEK's 0.746 and 1.006. However, incorporating actor-specific tokens for instance-aware adaptation yields notable improvements of \\(32.7\\%\\) and \\(16.7\\%\\) for short-term and long-term experiments, respectively. The difference in performance between short-term and long-term is influenced by the scene length in the dataset. The average scene length for short-term data with a 0.1 time interval is 200.04, significantly longer than the average of 32.67 for long-term data with a 0.5 time interval. Intuitively, as scene length increases, the time spent observing previously adapted actors also increases, enhancing the effectiveness of actor-specific tokens. To verify this, Fig. 9 adjusts scene length arbitrarily by skipping to the next scene in data loading once a specific scene length is exceeded. The results confirm that as scene length decreases by skipping scenes earlier, the effectiveness of actor-specific tokens diminishes in both short-term and long-term scenarios." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.13 + ], + "angle": 0, + "content": "Table 5. Effect of actor-specific token in \\( \\mathrm{mADE}_6 / \\mathrm{mFDE}_6 \\). The proposed method enhances adaptation performance by learning actor-wise motion characteristics." + }, + { + "type": "table", + "bbox": [ + 0.509, + 0.137, + 0.885, + 0.22 + ], + "angle": 0, + "content": "
Exp.BaselineOurs w/o Actor-specificOurs (Full)
INTER → Lyft (1/3/0.1)1.553 / 3.4580.581 / 1.1510.391 / 0.824
nuS → Lyft (2/6/0.5)1.108 / 2.5970.932 / 2.2200.776 / 1.820
" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.233, + 0.888, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.509 + ], + "angle": 0, + "content": "Figure 9. Effect of scenario length to the effectiveness of actor-specific token. As the scenario length shortens with manual skipping, its effectiveness diminishes because the duration available for the actor-specific token to adapt is reduced. In real-world applications, driving scenarios are continuous, resulting in maximal efficacy of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.523, + 0.618, + 0.538 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.761 + ], + "angle": 0, + "content": "We propose a test-time training method for trajectory prediction by incorporating the MAE and actor-specific token memory. The introduced MAE objective addresses a limitation of conventional online learning, preventing the loss of representations learned from source data. Consequently, our approach enables learning deeper layers, leading to improved representations and enhanced predictions even for out-of-distribution samples. The integration of actor-specific tokens during test-time allows for instance-wise learning of motion patterns, resulting in substantial performance improvements. This approach, particularly effective in continuous real-world autonomous driving scenarios without scene breaks, demonstrates significant efficacy and holds promise for practical applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements This research was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF2022R1A2B5B03002636) and the Challengeable Future Defense Technology Research and Development Program through the Agency For Defense Development (ADD) funded by the Defense Acquisition Program Administration (DAPA) in 2024 (No.912768601)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15072" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Görkay Aydemir, Adil Kaan Akan, and Fatma Güney. Adapt: Efficient multi-agent trajectory prediction with adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8295-8305, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Inhwan Bae, Jin-Hwi Park, and Hae-Gon Jeon. Nonprobability sampling network for stochastic human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6477-6487, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Inhwan Bae, Jean Oh, and Hae-Gon Jeon. Eigentrajectory: Low-rank descriptors for multi-modal trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10017-10029, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.472, + 0.354 + ], + "angle": 0, + "content": "[4] Collin Burns and Jacob Steinhardt. Limitations of post-hoc feature alignment for robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2525-2533, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.472, + 0.437 + ], + "angle": 0, + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[6] Rohan Chandra, Uttaran Bhattacharya, Aniket Bera, and Dinesh Manocha. Traphic: Trajectory prediction in dense and heterogeneous traffic using weighted interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8483-8492, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[7] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[8] Hao Chen, Jiaze Wang, Kun Shao, Furui Liu, Jianye Hao, Chenyong Guan, Guangyong Chen, and Pheng-Ann Heng. Traj-mae: Masked autoencoders for trajectory prediction. arXiv preprint arXiv:2303.06697, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.471, + 0.689 + ], + "angle": 0, + "content": "[9] Liang Chen, Yong Zhang, Yibing Song, Ying Shan, and Lingqiao Liu. Improved test-time adaptation for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24172-24182, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[10] Weihuang Chen, Zhigang Yang, Lingyang Xue, Jinghai Duan, Hongbin Sun, and Nanning Zheng. Multimodal pedestrian trajectory prediction using probabilistic proposal network. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[11] Xuesong Chen, Shaoshuai Shi, Chao Zhang, Benjamin Zhu, Qiang Wang, Ka Chun Cheung, Simon See, and Hongsheng Li. Trajectoryformer: 3d object tracking transformer with predictive trajectory hypotheses. arXiv preprint arXiv:2306.05888, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Yuxiao Chen, Boris Ivanovic, and Marco Pavone. Scept: Scene-consistent, policy-based trajectory predictions for planning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17103-17112, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[13] Jie Cheng, Xiaodong Mei, and Ming Liu. Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8679-8689, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[14] Junhyeong Cho, Gilhyun Nam, Sungyeon Kim, Hunmin Yang, and Suha Kwak. Promptstyler: Prompt-driven style generation for source-free domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15702-15712, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.893, + 0.29 + ], + "angle": 0, + "content": "[15] Dooseop Choi and KyoungWook Min. Hierarchical latent structure for multi-modal vehicle trajectory forecasting. In European Conference on Computer Vision, pages 129-145. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.893, + 0.36 + ], + "angle": 0, + "content": "[16] Sehwan Choi, Jungho Kim, Junyong Yun, and Jun Won Choi. R-pred: Two-stage motion prediction via tube-query attention-based trajectory refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8525-8535, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.894, + 0.446 + ], + "angle": 0, + "content": "[17] Henggang Cui, Vladan Radosavljevic, Fang-Chieh Chou, Tsung-Han Lin, Thi Nguyen, Tzu-Kuo Huang, Jeff Schneider, and Nemanja Djuric. Multimodal trajectory predictions for autonomous driving using deep convolutional networks. In 2019 International Conference on Robotics and Automation (ICRA), pages 2090-2096. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.447, + 0.893, + 0.515 + ], + "angle": 0, + "content": "[18] Ning Ding, Yixing Xu, Yehui Tang, Chao Xu, Yunhe Wang, and Dacheng Tao. Source-free domain adaptation via distribution estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7212-7222, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.518, + 0.893, + 0.602 + ], + "angle": 0, + "content": "[19] Nemanja Djuric, Vladan Radosavljevic, Henggang Cui, Thi Nguyen, Fang-Chieh Chou, Tsung-Han Lin, Nitin Singh, and Jeff Schneider. Uncertainty-aware short-term motion prediction of traffic actors for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2095-2104, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.603, + 0.893, + 0.659 + ], + "angle": 0, + "content": "[20] Yonghao Dong, Le Wang, Sanping Zhou, and Gang Hua. Sparse instance conditioned multimodal trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9763-9772, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.894, + 0.744 + ], + "angle": 0, + "content": "[21] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.745, + 0.894, + 0.801 + ], + "angle": 0, + "content": "[22] Francois Fleuret et al. Uncertainty reduction for model adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9613-9623, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.893, + 0.857 + ], + "angle": 0, + "content": "[23] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. Advances in Neural Information Processing Systems, 35:29374-29385, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.858, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[24] Thomas Gilles, Stefano Sabatini, Dzmitry Tsishkou, Bogdan Stanciulescu, and Fabien Moutarde. THOMAS: Trajectory heatmap output with learned multi-agent sampling. In In-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15073" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "ternational Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[25] Thomas Gilles, Stefano Sabatini, Dzmitry Tsishkou, Bogdan Stanciulescu, and Fabien Moutarde. Uncertainty estimation for cross-dataset performance in trajectory prediction. arXiv preprint arXiv:2205.07310, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[26] Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[27] Taesik Gong, Jongheon Jeong, Taewon Kim, Yewon Kim, Jinwoo Shin, and Sung-Ju Lee. NOTE: Robust continual test-time adaptation against temporal correlation. In Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[28] Junru Gu, Chenxu Hu, Tianyuan Zhang, Xuanyao Chen, Yilun Wang, Yue Wang, and Hang Zhao. Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5496-5506, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[29] Abner Guzman-Rivera, Dhruv Batra, and Pushmeet Kohli. Multiple choice learning: Learning to produce multiple structured outputs. Advances in neural information processing systems, 25, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[30] John Houston, Guido Zuidhof, Luca Bergamini, Yawei Ye, Long Chen, Ashesh Jain, Sammy Omari, Vladimir Iglovikov, and Peter Ondruska. One thousand and one hours: Self-driving motion prediction dataset. In Conference on Robot Learning, pages 409-418. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[31] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[32] Pingxuan Huang, Zhenhua Cui, Jing Li, Shenghua Gao, bo Hu, and Yanyan Fang. Cross-domain trajectory prediction with ctp-net, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[33] Manh Huynh and Gita Alaghband. AOL: Adaptive online learning for human trajectory prediction in dynamic video scenes. arXiv preprint arXiv:2002.06666, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[34] Manh Huynh and Gita Alaghband. Online adaptive temporal memory with certainty estimation for human trajectory prediction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 940-949, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[35] David Isele and Akansel Cosgun. Selective experience replay for lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[36] Boris Ivanovic, James Harrison, and Marco Pavone. Expanding the deployment envelope of behavior prediction via adaptive meta-learning. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7786-7793. IEEE, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] Boris Ivanovic, Guanyu Song, Igor Gilitschenski, and Marco Pavone. trajdata: A unified interface to multiple human trajectory datasets. In Proceedings of the Neural Information" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.891, + 0.119 + ], + "angle": 0, + "content": "Processing Systems (NeurIPS) Track on Datasets and Benchmarks, New Orleans, USA, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[38] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9644-9653, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[39] Ruochen Jiao, Xiangguo Liu, Takami Sato, Qi Alfred Chen, and Qi Zhu. Semi-supervised semantics-guided adversarial training for robust trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8207-8217, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[40] Junho Kim, Inwoo Hwang, and Young Min Kim. Ev-tta: Test-time adaptation for event-based object recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17745-17754, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[41] Mihee Lee, Samuel S Sohn, Seonghyeon Moon, Sejong Yoon, Mubbasir Kapadia, and Vladimir Pavlovic. Musevae: multi-scale vae for environment-aware long term trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2221-2230, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[42] Lihuan Li, Maurice Pagnucco, and Yang Song. Graph-based spatial transformer with memory replay for multifuture pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2231-2241, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[43] Maosen Li, Siheng Chen, Yanning Shen, Genjia Liu, Ivor W Tsang, and Ya Zhang. Online multi-agent forecasting with interpretable collaborative graph neural networks. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[44] Yingwei Li, Charles R Qi, Yin Zhou, Chenxi Liu, and Dragomir Anguelov. Modar: Using motion forecasting for 3d object detection in point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9329-9339, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[45] Yushu Li, Xun Xu, Yongyi Su, and Kui Jia. On the robustness of open-world test-time training: Self-training with dynamic prototype expansion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11836-11846, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[46] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[47] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. arXiv preprint arXiv:2303.15361, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[48] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecasting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 541-556. Springer, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[49] Hyesu Lim, Byeonggeun Kim, Jaegul Choo, and Sungha Choi. Ttn: A domain-shift aware batch normalization in test" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15074" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "time adaptation. In The Eleventh International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[50] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808-21820, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.47, + 0.274 + ], + "angle": 0, + "content": "[51] Yuejiang Liu, Riccardo Cadei, Jonas Schweizer, Sherwin Bahmani, and Alexandre Alahi. Towards robust and adaptive motion forecasting: A causal representation perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17081-17092, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.346 + ], + "angle": 0, + "content": "[52] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5517-5526, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[53] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[54] M Jehanzeb Mirza, Inkyu Shin, Wei Lin, Andreas Schriebl, Kunyang Sun, Jaesung Choe, Mateusz Kozinski, Horst Possegger, In So Kweon, Kuk-Jin Yoon, et al. Mate: Masked autoencoders are online 3d test-time learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16709-16718, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.601 + ], + "angle": 0, + "content": "[55] Jiquan Ngiam, Vijay Vasudevan, Benjamin Caine, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, David J Weiss, Benjamin Sapp, Zhifeng Chen, and Jonathon Shlens. Scene transformer: A unified architecture for predicting future trajectories of multiple agents. In International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[56] Daehee Park, Jaewoo Jeong, and Kuk-Jin Yoon. Improving transferability for cross-domain trajectory prediction via neural stochastic differential equation. arXiv preprint arXiv:2312.15906, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[57] Daehee Park, Hobin Ryu, Yunseo Yang, Jegyeong Cho, Jiwon Kim, and Kuk-Jin Yoon. Leveraging future relationship reasoning for vehicle trajectory prediction. In The Eleventh International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[58] Mozhgan Pourkeshavarz, Changhe Chen, and Amir Rasouli. Learn tarot with mentor: A meta-learned self-supervised approach for trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8384–8393, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[59] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[60] Luke Rowe, Martin Ethier, Eli-Henry Dykhne, and Krzysztof Czarnecki. Fjmp: Factorized joint multi-agent motion prediction over learned directed acyclic interaction" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13745-13755, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[61] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[62] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Wei Tang, Nanning Zheng, and Gang Hua. Representing multimodal behaviors with mean location for pedestrian trajectory prediction. IEEE transactions on pattern analysis and machine intelligence, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.279, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[63] Liushuai Shi, Le Wang, Sanping Zhou, and Gang Hua. Trajectory unified transformer for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9675-9684, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[64] Jianhua Sun, Yuxuan Li, Hao-Shu Fang, and Cewu Lu. Three steps to multimodal trajectory prediction: Modality clustering, classification and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13250-13259, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.407, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[65] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In Proceedings of the 37th International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.478, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[66] Devavrat Tomar, Guillaume Vray, Behzad Bozortabar, and Jean-Philippe Thiran. Tesla: Test-time self-learning with automatic adversarial augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20341–20350, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.549, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[67] Li-Wu Tsao, Yan-Kai Wang, Hao-Siang Lin, Hong-Han Shuai, Lai-Kuan Wong, and Wen-Huang Cheng. Socialssl: Self-supervised cross-sequence representation learning based on transformers for multi-agent trajectory prediction. In European Conference on Computer Vision, pages 234-250. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[68] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[69] Chunnan Wang, Xiang Chen, Junzhe Wang, and Hongzhi Wang. Atpfl: Automatic trajectory prediction model design under federated learning framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6563-6572, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[70] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[71] Eason Wang, Henggang Cui, Sai Yalamanchi, Mohana Moorthy, and Nemanja Djuric. Improving movement predictions of traffic actors in bird's-eye view models using gans and differentiable trajectory rasterization. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2340-2348, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15075" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[72] Jingke Wang, Tengju Ye, Ziqing Gu, and Junbo Chen. Ltp: Lane-based trajectory prediction for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17134-17142, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[73] Letian Wang, Yeping Hu, and Changliu Liu. Online adaptation of neural network models by modified extended kalman filter for customizable and transferable driving behavior prediction, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[74] Letian Wang, Yeping Hu, Liting Sun, Wei Zhan, Masayoshi Tomizuka, and Changliu Liu. Transferable and adaptable driving behavior prediction. arXiv preprint arXiv:2202.05140, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.472, + 0.319 + ], + "angle": 0, + "content": "[75] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.376 + ], + "angle": 0, + "content": "[76] Zhibo Wang, Jiayu Guo, Haiqiang Zhang, Ru Wan, Junping Zhang, and Jian Pu. Bridging the gap: Improving domain generalization in trajectory prediction. IEEE Transactions on Intelligent Vehicles, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.47, + 0.447 + ], + "angle": 0, + "content": "[77] Chenfeng Xu, Tian Li, Chen Tang, Lingfeng Sun, Kurt Keutzer, Masayoshi Tomizuka, Alireza Fathi, and Wei Zhan. Pretram: Self-supervised pre-training via connecting trajectory and map. In European Conference on Computer Vision, pages 34-50. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.47, + 0.532 + ], + "angle": 0, + "content": "[78] Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1410-1420, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.588 + ], + "angle": 0, + "content": "[79] Pei Xu, Jean-Bernard Hayet, and Ioannis Karamouzas. Socialvae: Human trajectory prediction using timewise latents. In European Conference on Computer Vision, pages 511-528. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.471, + 0.646 + ], + "angle": 0, + "content": "[80] Yi Xu, Lichen Wang, Yizhou Wang, and Yun Fu. Adaptive trajectory prediction via transferable gnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6520-6531, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.717 + ], + "angle": 0, + "content": "[81] Yi Xu, Armin Bazarjani, Hyung-gun Chi, Chiho Choi, and Yun Fu. Uncovering the missing pattern: Unified framework towards trajectory imputation and prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9632-9643, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[82] Zhijie Yan, Pengfei Li, Zheng Fu, Shaocong Xu, Yongliang Shi, Xiaoxue Chen, Yuhang Zheng, Yang Li, Tianyu Liu, Chuxuan Li, et al. Int2: Interactive trajectory prediction at intersections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8536-8547, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[83] Tao Yang, Shenglong Zhou, Yuwang Wang, Yan Lu, and Nanning Zheng. Test-time batch normalization. arXiv preprint arXiv:2205.10210, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[84] Luyao Ye, Zikang Zhou, and Jianping Wang. Improving the generalizability of trajectory prediction models with freenet-based domain normalization. 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11562-11568, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[85] Jiangbei Yue, Dinesh Manocha, and He Wang. Human trajectory prediction via neural social physics. In European Conference on Computer Vision, pages 376-394. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[86] Wei Zhan, Liting Sun, Di Wang, Haojie Shi, Aubrey Clausse, Maximilian Naumann, Julius Kummerle, Hendrik Konigshof, Christoph Stiller, Arnaud de La Fortelle, et al. Interaction dataset: An international, adversarial and cooperative motion dataset in interactive driving scenarios with semantic maps. arXiv preprint arXiv:1910.03088, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[87] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems, 35: 38629-38642, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.29, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[88] Xingxuan Zhang, Peng Cui, Renzhe Xu, Linjun Zhou, Yue He, and Zheyan Shen. Deep stable learning for out-of-distribution generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5372-5382, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[89] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863-17873, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.892, + 0.513 + ], + "angle": 0, + "content": "[90] Dekai Zhu, Guangyao Zhai, Yan Di, Fabian Manhardt, Hendrik Berkemeyer, Tuan Tran, Nassir Navab, Federico Tombari, and Benjamin Busam. Ipcc-tp: Utilizing incremental pearson correlation coefficient for joint multi-agent trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5507-5516, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.515, + 0.892, + 0.583 + ], + "angle": 0, + "content": "[91] Yiyao Zhu, Di Luan, and Shaojie Shen. Biff: Bi-level future fusion with polyline-based coordinate for interactive trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8260-8271, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15076" + } + ] +] \ No newline at end of file diff --git a/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/11bec0da-9290-4827-bb16-e062c2aadc7c_origin.pdf b/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/11bec0da-9290-4827-bb16-e062c2aadc7c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ab306f98dcf9c5ea5b8239b36d15166c0abebde9 --- /dev/null +++ b/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/11bec0da-9290-4827-bb16-e062c2aadc7c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:652826353df4fd5ca3b3b279c442952a958aba0169127e17936e1002ab83bc6a +size 4371310 diff --git a/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/full.md b/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/full.md new file mode 100644 index 0000000000000000000000000000000000000000..044cc8823881a32f6dc70b6cd013365019da3623 --- /dev/null +++ b/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/full.md @@ -0,0 +1,361 @@ +# T4P: Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory + +Daehee Park, Jaeseok Jeong, Sung-Hoon Yoon, Jaewoo Jeong, and Kuk-Jin Yoon KAIST + +{bag2824, jason. jeong, yoon307, jeong207, kjyoon}@kaist.ac.kr + +# Abstract + +Trajectory prediction is a challenging problem that requires considering interactions among multiple actors and the surrounding environment. While data-driven approaches have been used to address this complex problem, they suffer from unreliable predictions under distribution shifts during test time. Accordingly, several online learning methods have been proposed using regression loss from the ground truth of observed data leveraging the auto-labeling nature of trajectory prediction task. We mainly tackle the following two issues. First, previous works underfit and overfit as they only optimize the last layer of motion decoder. To this end, we employ the masked autoencoder (MAE) for representation learning to encourage complex interaction modeling in shifted test distribution for updating deeper layers. Second, utilizing the sequential nature of driving data, we propose an actor-specific token memory that enables the test-time learning of actor-wise motion characteristics. Our proposed method has been validated across various challenging cross-dataset distribution shift scenarios including nuScenes, Lyft, Waymo, and Interaction. Our method surpasses the performance of existing state-of-the-art online learning methods in terms of both prediction accuracy and computational efficiency. The code is available at https://github.com/daeheepark/T4P. + +# 1. Introduction + +Trajectory prediction plays a significant role in autonomous systems, enhancing safety and navigation efficiency [28, 31]. Recently, data-driven methods have shown remarkable prediction capabilities [1, 16, 38, 52, 55, 57, 60, 81, 89, 90]; however, they are prone to distribution shift [9, 88]. Trajectory prediction models also produce unreliable output when faced shifts in the data distribution [25], posing significant risks in various real-world applications. This vulnerability stems from the ease with which the trajectory data distri + +![](images/33acb881603285214330de90b83340b0d3fc7c37f8a9239ca2e9d5f9f611682c.jpg) +Figure 1. Previous methods optimize the last layer of the decoder using regression loss from delayed ground truth. Our method, on the other hand, learns representation via a masked autoencoder, which boosts prediction performance by optimizing deeper layers. In addition, the proposed actor-specific token enables the prediction model to learn actor-wise motion characteristics. + +bution can be altered by numerous factors, such as scene changes and driving habits; i.e. road layout, interaction between agents, driver demographics [6, 51, 82]. + +To address this challenge, recent methods have proposed domain adaptation and generalization strategies which aim to anticipate the distribution shifts and accordingly train the model [32, 76, 80]. However, due to the wide variety of factors influencing data distribution, the anticipated shifts may differ significantly from those encountered at test time. As a result, several online learning methods have been developed to dynamically adapt models during test time [36, 73]. Since trajectory prediction serves as an auto-labeling task where trajectory data is obtained from object tracking, the observed past and future trajectory provide both input and corresponding ground truth for supervision $(\mathcal{L}_{reg})$ , as depicted at the top right of Fig. 1. Nevertheless, updating the entire model may ruin the representation learned from the + +source data, so only part of the network, such as the batch normalization layer, is updated [27, 40, 61, 83]. Particularly, because regression loss is calculated at delayed timestamps and only a few samples are available during test time, this approach risks deteriorating the model's learned representation [73]. Therefore, previous online prediction methods mainly update only the last layer of the decoder. + +In this work, we propose a test-time training (TTT) for trajectory prediction with two key aspects. First, we build a masked autoencoder (MAE) framework to adapt deep features, incorporating good representation that captures complex interactions between agents and road structures. Due to the challenge of existing online learning methods in damaging learned representations when updating deeper layers, we employ a MAE to guide representation learning. Second, we introduce an actor-specific token memory that has significant advantages in real-world driving scenario where data arrives continuously and sequentially. As each actor instance has its own driving habits and the past motion pattern of specific actors can be accessed from the arrived observations, we design a token memory in transformer structure [68] and its training strategy to learn actor-wise motion characteristics. The proposed TTT framework is validated on challenging cross-dataset distribution shift cases between nuScenes, Lyft, Waymo, and INTERACTION, and shows state-of-the-art performance surpassing previous online learning methods. Furthermore, we show the practicality of our TTT framework by evaluating its computational efficiency. We summarize our contributions as below: + +- We propose a test time training for trajectory prediction (T4P) by utilizing a masked autoencoder to learn deep feature representations that stably improve prediction performance across entire network layers. +- We introduce an actor-specific token memory used to learn the different actor characteristics and habits. +- Our method is validated across 4 different datasets as well as different temporal configurations. Ours achieves state-of-the-art performance both in accuracy and efficiency. + +# 2. Related Works + +# 2.1. Trajectory Prediction + +Trajectory prediction garners attention with the emergence of methods that can enhance perception or planning [11, 12, 44]. Its goal is to predict future trajectories of traffic actors based on their historical trajectories and the context of their environment [3, 17, 19, 20, 39, 58, 63, 78, 91]. Historical trajectories, or tracklets of traffic actors, are sequentially acquired via vehicle detection and tracking systems. Some studies employed this temporal property to enhance prediction via memory replay [35, 42, 59]. In the early stages of trajectory prediction, only the historical trajectory of the actors of interest was considered. However, recent stud + +ies emphasize the significance of understanding interactions among agents [67, 85] and the rules governed by surrounding environments [72, 77] in improving prediction performance. This has led to the development of models that incorporate multi-head attention or graph-based methods to capture these interactions [24, 26, 42]. Additionally, MAE has been adopted for pretraining to better understand agent interactions [8, 13]. To further refine prediction capabilities, various generative models have been introduced, enabling the generation of future trajectories [15, 41, 71, 79]. + +# 2.2. Transfer Learning in Trajectory Prediction + +With data-driven approaches offering superior performance in trajectory prediction, their effectiveness diminishes under distribution shifts [25, 56]. In response, several studies have adopted domain adaptation or generalization strategy [69, 74]. Some specifically aimed to reduce the domain gap within unique characteristics of trajectory prediction: differences in road structures [84], actor interaction [80], etc. However, these methods rely on anticipating how to cover domain shifts. Yet, given that trajectory data is subject to influence from numerous factors, predicting and accommodating for shifts may not always be sufficient. Consequently, recent developments have introduced adaptation to unseen test sets using online learning [33, 34, 43]. Among them, some methods [36, 73] showed remarkable prediction performance improvement under severe distribution shifts like cross-dataset cases by utilizing regression loss for online learning. These methods exploit the fact that the input and ground truth (GT) labels are provided at test time as tracking history. However, with the limitations of updating with only a few samples in a delayed time, adaptation becomes restricted to the last layer of the decoder. + +# 2.3. Test Time Training + +Test-time training (TTT) is a method that trains the network on test time data, unseen during training [4, 7, 14, 18, 22, 49, 66, 75]. Unlike domain generalization or adaptation, which are confined to the training phase, TTT extends model adaptation into the test phase by utilizing available test data [47]. TTT methods are categorized into regularization-based approaches for post-hoc regularization of out-of-distribution (OOD) samples [46, 87], and self-supervised approaches that employ pretext tasks on test data for optimal representation learning [7, 9, 45, 50, 54]. Specifically, TTT [65] introduced a Y-shaped network structure consisting of a feature encoder, a pretext branch, and a decoder branch. The decoder branch is fixed, while the encoder and pretext branch are optimized through self-supervision. Adhering to this model, TTT-MAE [23] integrated a MAE in the pretext task. Expanding on this method, we adopt TTT-MAE to the domain of trajectory prediction, leveraging its representation learning capabilities to enhance test-time training. + +![](images/f3b3e43995adeb1299905a2a1f4552e9eb13aeb779576994b464a229953c164b.jpg) +Figure 2. Overall method. During test-time training, the network trained on source dataset is optimized on target data under online setting. The model is optimized both from regression and reconstruction loss. Both losses utilize the data observed at the delayed time stamp $(t_{\tau})$ . Actor-specific token is used to learn instance-wise motion pattern during test-time training phase. During online evaluation phase, model and actor-specific token learned from test-time training phase are used. + +# 3. Method + +# 3.1. Problem definition + +Trajectory prediction aims to learn the mapping function between the input, consisting of historical trajectory and map information $\mathbf{X}:\{\mathcal{X}_t,\mathcal{M}_t\}$ , and the output, consisting of $K$ possible candidates for future trajectory of $N$ traffic actors, $\mathbf{Y}:\{\mathcal{Y}_t^{0:K - 1}\}$ , at current time $t$ . We predict $C$ different actor classes including vehicle, cyclist, etc. Historical and future trajectories are represented as $\mathcal{X}_t = \mathbf{x}_{t - t_h:t}^{0:N - 1}$ and $\mathcal{Y}_t = \mathbf{x}_{t:t + t_f}^{0:N - 1}$ where $t_h$ and $t_f$ represent sequence length of input and output trajectory. Here, $\mathbf{x}_t^n$ represents the spatial location of actor $n$ at time $t$ . For map information $\mathcal{M}_t$ , we use $L$ segmented lane centerlines around ego-actors which is widely-used in trajectory prediction methods. + +We deal with the case when the target data distribution during test time $\{\mathbf{X},\mathbf{Y}\}^T$ is different from the source data distribution seen during the training phase $\{\mathbf{X},\mathbf{Y}\}^S$ . We formulate the problem as a online adaptation scenario in which one data sample is given per each time interval as time passes. The test data is consists of multiple distinct scenes. Each scene includes temporally ordered data samples which are captured through real-world driving. Following standard TTT methods, there is no access to the source data at test time. However, thanks to the auto-labeling nature of trajectory prediction, there is access to delayed GT future trajectory $(\mathbf{x}_{t_{\tau}:t_{\tau} + t_f})$ from a previous time window at time $t_\tau (= t - \tau)$ as depicted in left lower corner of Fig. 2. + +# 3.2. Overall method + +Our method, Test-Time Training of Trajectory Prediction (T4P), enhances the online learning method using supervi + +sion from a delayed GT future trajectory with representation learning from MAE and actor-specific token memory. Following standard TTT [65], the overall framework consists of three phases: offline training, test-time training, and online evaluation. Offline training occurs before test time, and test-time training and online evaluation are executed repeatedly and sequentially during test time. We adopt the ForecastMAE [13] backbone consisting of embedding layers $f$ , a shared encoder $E$ , a reconstruction head $R$ and a motion decoder head $D$ , as depicted in the middle of Fig. 2. The detailed methods during each phase are described below: + +# 3.3. Offline training + +During offline training, the model is trained on source data using both reconstruction loss and regression loss. + +$$ +\min _ {\theta \in \{f, E, R, D \}} \mathbb {E} _ {\mathcal {X}, \mathcal {Y}, \mathcal {M} \in \{\mathbf {X}, \mathbf {Y} \} ^ {S}} \left[ \mathcal {L} _ {\text {r e c o n}} + \mathcal {L} _ {\text {r e g}} \right] \tag {1} +$$ + +In this subsection, subscript $t$ is omitted for simplicity. First, all input elements $(\mathcal{X},\mathcal{Y},\mathcal{M})$ are embedded with their respective embedding layer $(f_x,f_y,f_m)$ . Additionally, we define actor class token $\bar{\alpha}\in \mathbb{R}^{C\times D}$ that learns different motion patterns of each actor class. The actor class token is implemented as a learnable embedding of a transformer structure. Corresponding actor class token $\alpha (c)$ is added to trajectory embedding according to the class of each actor. + +$$ +h _ {x}, h _ {y}, h _ {m} = f _ {x} (\mathcal {X}) + \alpha (c), f _ {y} (\mathcal {Y}) + \alpha (c), f _ {m} (\mathcal {M}) \tag {2} +$$ + +For reconstruction, the history/future trajectory and lane embeddings are fed to the encoder to obtain the encodings $F_{x}, F_{y}, F_{m}$ . Then, segments of the encodings are randomly masked and replaced with masking tokens, $M_{x}, M_{y}, M_{m}$ , + +and the other encodings remain unmasked $(F_x', F_y', F_m')$ . Here, we use random masking for the lane centerline and complementary masking strategy for history/future trajectory following previous works [13]. The masking tokens and the unmasked encodings are fed to the reconstructor to reconstruct the masked elements. + +$$ +F _ {x}, F _ {y}, F _ {m} = E \left(h _ {x}, h _ {y}, h _ {m}\right) \tag {3} +$$ + +$$ +\hat {\mathcal {X}}, \hat {\mathcal {Y}}, \hat {\mathcal {M}} = R \left(M _ {x}, M _ {y}, M _ {m}, F _ {x} ^ {\prime}, F _ {y} ^ {\prime}, F _ {m} ^ {\prime}\right) +$$ + +The encoder and reconstructor both consist of multi-head attention to utilize interaction between history, future and lanes, thus, the reconstruction guides to the model have the capability of interaction reasoning. Finally, a reconstruction loss $\mathcal{L}_{\text{recon}}$ is computed as the MSE loss between the ground truth and the reconstructed outputs. + +$$ +\mathcal {L} _ {\text {r e c o n}} = \frac {1}{N} \sum_ {n} (\mathcal {X} - \hat {\mathcal {X}}) ^ {2} + \frac {1}{N} \sum_ {n} (\mathcal {Y} - \hat {\mathcal {Y}}) ^ {2} + \frac {1}{L} \sum_ {l} (\mathcal {M} - \hat {\mathcal {M}}) ^ {2} \tag {4} +$$ + +For the decoder head, historical trajectory and lanes embeddings are again fed to the same encoder and the output encodings are passed to the motion decoder, composed of MLP layers. The decoder outputs $\mathrm{K}$ candidates for trajectory prediction, and the regression loss $\mathcal{L}_{reg}$ is computed with the widely-used Winner-takes-all (WTA) loss [29, 48]. + +$$ +\hat {\mathcal {Y}} ^ {0: K - 1} = D \left(E \left(h _ {h}, h _ {l}\right)\right) +$$ + +$$ +\mathcal {L} _ {r e g} = \frac {1}{N} \sum_ {n} \operatorname {a r g m i n} _ {k \in K} \left(\mathcal {Y} ^ {n} - \hat {\mathcal {Y}} ^ {n, k}\right) ^ {2} \tag {5} +$$ + +# 3.4. Test-time training + +During test-time, a data sample consisting of trajectories and maps arrives sequentially. Therefore, even though we cannot access the GT future trajectory of current time $(\mathcal{V}_t)$ , we can access both the inputs and GT $(\mathcal{X}_{t_{\tau}},\mathcal{Y}_{t_{\tau}},\mathcal{M}_{t_{\tau}})$ at a previous time $t_\tau$ . With this data, the model is optimized with the same objective as Eq. 1 with target data distribution instead of source data distribution. + +$$ +\min _ {\theta \in \{f, E, R, D \}} \mathbb {E} _ {\chi_ {t _ {\tau}}, \mathcal {Y} _ {t _ {\tau}}, \mathcal {M} _ {t _ {\tau}} \in \{\mathbf {X}, \mathbf {Y} \} ^ {T}} \left[ \mathcal {L} _ {\text {r e c o n}} + \mathcal {L} _ {\text {r e g}} \right] \tag {6} +$$ + +Unlike existing online learning methods that only utilize regression loss, we incorporate an additional reconstruction loss. This enables the model to learn a good representation that considers the complex actor-actor and actor-lane interaction even in the unseen target data distribution. An advantage of representation learning is that the performance stably improves even when the deeper layers are optimized. + +# 3.4.1 Actor-specific token memory + +Unlike during the offline training phase, when the data order is shuffled, data at test-time comes in sequentially. Therefore, at test time, it is possible to keep track of movement + +![](images/e535b160ff33fd972055114707520df3b58efa713657fe5868a37432b5d83e17.jpg) +Figure 3. Actor-specific token memory is colored in gray. It evolves as time passes within a scene. For newborn actors, the corresponding class token is registered. Until the actor disappears, the token is updated through test-time training. At the end of the scene, all tokens are averaged by each class and passed to the next scene as denoted in red arrow and Eq. 7. + +patterns of a specific actor instance. Using this, we propose an actor-specific token memory. + +The overall scheme is described in Fig. 3. During offline training, actor class tokens $\bar{\alpha}_{train} \in \mathbb{R}^{C \times D}$ are trained to reflect the average motion pattern of each of the $C$ classes. At the beginning of the test-time, scene 0, class tokens are initialized from that of training phase ( $\bar{\alpha}_{\text{scene}(0)} \gets \bar{\alpha}_{\text{train}}$ ). When a new $n^{th}$ -actor appears at time $t$ , the class token $\alpha_{n}^{t}(c)$ is cloned from $\bar{\alpha}_{\text{scene}(0)}$ by selecting corresponding class. The newborn tokens are then registered to the actor-specific token memory. The token memory is structured as a dictionary where actor instance ID/corresponding tokens are key/values. At each iteration, the actor-specific tokens are used for both test-time training and online evaluation. As time progresses, the actor-specific tokens evolve and are updated through the reconstruction and regression losses until the actor disappears in the scene. By giving each actor its own specific token that distinguishes it from the sharing of other parts of models with other actors, actor-specific motion patterns can be learned. + +When the scene changes, scene 1, the actors observed during scene 0 are not to be observed anymore, so we need a new averaged actor class token $\bar{\alpha}_{\text{scene}(1)}$ . For that, we average all the tokens in the memory at the final time step $T_{S}$ of scene 0 after gathering by classes as Eq. 7. Here, $N_{c}$ refers to the number of actors of class $c$ . It is because, with a sufficient number of actors class tokens per each class in memory, their average motion can be a representative motion pattern of actor classes. This is more useful than $\bar{\alpha}_{\text{train}}$ because newly averaged tokens are trained on the target dataset while $\bar{\alpha}_{\text{train}}$ contains motion pattern trained on the source dataset. The averaged class tokens are then passed to the next scene and used to initialize tokens for the + +Table 1. Adaptation results in various distribution shifts. The model is trained on source dataset, and test-time trained and evaluated on target dataset (Source $\rightarrow$ Target). All metrics are better in lower value. The best and second-best results are marked in bold and underline. + +
mADE6/mFDE6Short-term exp (1/3/0.1)Long-term exp (2/6/0.5)
INTER → nuSINTER → LyftnuS → WayMeannuS → LyftWay → LyftWay → nuSMean
Source Only1.047 / 2.2471.391 / 2.9450.431 / 1.0310.956 / 2.0741.122 / 2.5770.621 / 1.3471.153 / 2.2200.965 / 2.048
Joint Training1.116 / 2.4451.553 / 3.4580.472 / 1.1251.047 / 2.3431.108 / 2.5970.638 / 1.4041.091 / 2.0310.946 / 2.011
DUA1.118 / 2.4551.516 / 3.3520.516 / 1.2941.050 / 2.3671.365 / 3.2570.790 / 1.8681.270 / 2.6341.142 / 2.586
TENT (w/ sup)1.102 / 2.4231.519 / 3.4050.448 / 1.0711.023 / 2.3001.068 / 2.5140.628 / 1.3811.077 / 2.0120.924 / 1.969
MEK (τ = tf/2)1.012 / 2.4451.283 / 3.4580.445 / 1.1250.913 / 2.3431.079 / 2.5970.629 / 1.4041.079 / 2.0310.929 / 2.011
MEK (τ = tf)0.892 / 1.9520.746 / 1.6540.405 / 1.0610.691 / 1.5561.006 / 2.3690.615 / 1.3511.117 / 2.1400.913 / 1.953
AML (K0)2.093 / 4.6972.695 / 6.6771.624 / 2.1392.137 / 4.5041.787 / 3.0671.322 / 2.5711.618 / 2.9991.576 / 2.879
AML (full)1.149 / 2.5501.042 / 2.6160.764 / 1.7910.985 / 2.3191.462 / 2.5730.977 / 2.1841.495 / 2.9781.311 / 2.578
Ours (T4P)0.537 / 1.1370.391 / 0.8240.336 / 0.8070.421 / 0.9230.776 / 1.8200.549 / 1.1710.996 / 1.7840.774 / 1.592
+ +newborn actors. Please note that while scene 0 is initialized by actor class tokens from the training phase, subsequent scenes obtain as in Eq. 7. More details of memory evolving strategies can be found on the supplementary material. + +$$ +\bar {\alpha} _ {s c e n e (i + 1)} \leftarrow \left\{\frac {1}{N _ {c}} \sum_ {n} ^ {N _ {c}} \alpha_ {n} ^ {T} (c) \right\} _ {s c e n e (i)} \tag {7} +$$ + +# 3.5. Online evaluation + +Using the updated model weight and actor-specific token memory during test-time training, online evaluation is executed. With the input data $(\mathcal{X}_t,\mathcal{M}_t)$ at current time $t$ , the learned encoder and decoder predict multi-modal trajectory $(\mathcal{Y}_t)$ for all actors in the sample. + +# 4. Experiment + +# 4.1. Datasets + +We conducted experiments on well-known datasets, nuScenes [5], Lyft [30], WOMD [21], and INTERACTION [86], to evaluate T4P on various data distribution shifts. These datasets are parsed into the same format using trajdata [37]. Additionally, to verify in various prediction configurations, experiments were conducted with the two most widely used configurations of long-term and short-term prediction. Long-term prediction requires predicting 6 seconds into the future given 2 seconds of the past with a time interval of 0.5s, making the input/output sequence lengths to be 5 and 12, respectively. Short-term prediction requires predicting 3 seconds into the future given 0.9 seconds of the past with a time interval of 0.1s, making the input/output sequence lengths to be 10 and 30, respectively. + +# 4.2. Implementation details + +For actor classes, we use the 5 classes: unknown, vehicle, pedestrian, bicycle and motorcycle. Our method predicts $K = 6$ future candidates for all actors in the sample. We use $\tau$ as $t_f$ to enable the past GT future to contain full prediction horizon. We train and evaluate our model with a single NVIDIA A6000. Learning rates of model weight and actor-specific parameters are set as 0.01 and 0.5, respectively, and + +weight decay is set to 0.001 for all. The gradient is clipped by 15. For metrics, widely used $\mathrm{mADE}_6$ and $\mathrm{mFDE}_6$ are used. Detailed metric definition, model architecture, and training details are included in the supplementary material. + +# 4.3. Baselines + +We compare our $T4P$ with several baselines, including unsupervised/supervised test-time-training methods and online learning trajectory prediction methods. All baseline methods are implemented using the same backbone. + +Source only refers to the backbone model trained on the source dataset only using regression loss. + +Joint training is similar to source only but trained with regression and reconstruction loss jointly. + +DUA [53] is an unsupervised post-hoc regularization method only updates batch normalization statistics in a momentum-updating manner without back-propagation. + +TENT with supervision is a variant of the original TENT [70] in which regression loss is used to optimize the batch normalization layers instead of entropy minimization loss, as entropy minimization is not applicable. + +MEK [73] is an online learning trajectory prediction method utilizing the Modified Extended Kalman filter. It uses only regression loss to optimize the last layer of the decoder. As the prediction horizon is different in our experiment from the original paper, we use both $\frac{1}{2} t_f$ and $t_f$ . + +AML [36] is an Adaptive Meta-learning method. Unlike the other methods that use the same backbone, AML replaces the last decoder layer with a Bayesian linear regression layer for adaptive training. The modified version of backbone without adaptive training is denoted as $K_{0}$ , while the full version with adaptive training is denoted as full. + +# 5. Results + +# 5.1.Quantitative results + +The results of comparing our method with the baselines in various distribution shift scenarios are presented in Tab. 1. We reported three distribution shift scenarios per each time configuration on the table, and other results are included in + +![](images/4fd3c636e118474740e034311324639347cb5a334a62e9040f6797039ff237ae.jpg) +Figure 4. The first row shows prediction before adaptation, and the second row indicates adaptation results by three methods: ours (blue), TENT w/ sup (orange) and MEK (green). Sky blue and orange boxes refer to surrounding actors and actors to be predicted. We depicted only one actor result and one mode among multi-modal predictions closest to the GT for visual simplicity. Please note that our method is multi-modal prediction for all actors method. + +![](images/a52ae0ae537e5b08008adbc411759534ea5d38dd112a24fa1b5c629b8dfbb041.jpg) +Figure 5. Prediction accuracy and execution time on INTER $\rightarrow$ nuS (1/3/0.1) experiment. Adjusting update frequency can balance between accuracy and efficiency. Our method significantly outperforms the baseline methods in both accuracy and efficiency. + +the supplementary material. Notably, our approach consistently surpassed baseline performance across all scenarios. + +DUA consistently exhibits compromised performance across nearly all cases, due to the distinctive characteristics of trajectory data. In contrast to image data, where data is treated as a singular sample, trajectory data involves multiple agents, each exhibiting distinct motion patterns, within a single data. Consequently, holistically updating batch statistics proves to be counterproductive. Similar challenges are encountered by TENT w/ sup. While regression loss prevents performance decline, updating only the batch norm layer has little to no effect on the prediction performance. + +MEK exhibited the most competitive performance among the baselines. While performance improved significantly in short-term settings, it brings limited improvement in long-term cases. As the Kalman filter updates based on the number of prediction steps, short-term configurations with 12 update steps show a better improvement than long-term configurations with only 5 update steps. + +Although AML led to a considerable improvement in the full version, the predictive performance itself was substan + +tially degraded due to the significant performance drop in the modified backbone $(K_0)$ . The limitation of the backbone is due to the Bayesian regression layer being based on probability sampling which is known to be worse than non-probability sampling method of ours [2]. In contrast to all the baseline methods, our method demonstrated state-of-the-art performance in all scenarios featuring various distribution shifts, whether short-term or long-term, showcasing the generalizability of our approach. + +# 5.1.1 Efficiency + +As efficiency is a crucial factor in TTT, we evaluate the frame per second (FPS) along with accuracy $(\mathrm{mADE}_6)$ , shown in Fig. 5. We set the performance of the joint training method w/o adaptation as the benchmark and present MEK and TENT, which demonstrates competitive performance among the baselines. Our approach allows for the adjustment of the update frequency, with a frequency of 1 indicates updating at every opportunity, and 2 means updating every other opportunity. While frequent updates improve prediction performance, they also increase execution time; adjusting the update frequency allows for a balance between efficiency and accuracy. As shown in Fig. 5, our method outperforms in both accuracy and efficiency. Given that the time interval is 0.1 seconds, real-time execution requires a processing speed of at least 10 FPS. Even at the maximum update frequency of 1, our method maintains real-time capability with a superior accuracy of 0.39. When increasing the update frequency to 20, the error increases to 0.81, close to MEK's 0.75. However, the FPS reaches 24.2, demonstrating overwhelmingly faster operation compared to MEK's speed of 3.3 FPS. + +# 5.2. Qualitative results + +Comparison to the baselines: We compare our results with TENT and MEK in Fig. 4. While all methods, includ + +![](images/d085ef7d88ac5057af959f00a6e97261c08931baa7e5625dc918e537e76a8e66.jpg) +Figure 6. The first row indicates masked samples, and the row below shows the reconstructed outputs. The blue/red arrows indicate historical/future trajectories. The black arrows refer to the masked trajectories. The white lines are the lane centerlines, and the gray dashed lines are the masked lane centerlines. + +![](images/c365f1fd0cd1174f5ab1788d0fbd97f4c6916990322d11398392d02b6c30732a.jpg) +Figure 7. Multi modal prediction results (blue arrows) before and after adaptation via our method. Ours generates elaborate samples that consider interaction between lane (above) or actors (below) due to representation learning, which cannot be learned from the GT (red arrow) via regression loss. + +ing ours, perform multi-agent, multi-modal predictions, we only illustrate one actor and the closest mode to the GT for visual simplicity. The first row of the figure represents predictions before adaptation, and below are the results after adaptation using three different methods. MEK and TENT exhibit instances of underfitting or excessive overfitting upon adaptation, whereas our method consistently demonstrates stable and accurate predictions. + +Reconstruction results: Reconstruction examples are depicted in Fig. 6. For all agents and lanes within a data sample, random masking is applied, as shown in the first row. During test-time training, learning for reconstruction is conducted, resulting in successful reconstruction for data with different distributions, as seen in the second row. + +Multi-modal prediction results: As multi-modality is a crucial issue [10, 62, 64], we show that ours can handle multi-modal prediction results in Fig. 7. The adapted predictions showcase diverse yet plausible scenarios, either considering the lane structure (above) or surrounding agents + +Table 2. Effect of type of losses to be optimized. Optimizing all losses shows optimal test-time adaptation performance. + +
Exp.Loss typemADE6 /mFDE6
Actor reconLane reconReg
INTER → Lyft (1/3/0.1)✓1.553 / 3.458
✓✓1.054 / 2.007
✓0.842 / 1.512
0.674 / 1.430
✓✓✓0.391 / 0.824
nuS → Lyft (2/6/0.5)✓1.108 / 2.597
0.987 / 2.304
✓✓0.973 / 2.280
✓0.942 / 2.262
✓✓✓0.776 / 1.820
+ +Table 3. Ablation on the depth of optimizing layer according to the loss types. The right side of the table indicates the optimization of deeper layers. Using only $\mathcal{L}_{reg}$ deteriorates performance when optimizing all layers while ours stably improves as deeper. + +
LossOptimizing layers
DD+ED+E+f h,f,l
Lreg0.864 / 2.0720.840 / 2.0930.942 / 2.262
Lreg + Lrecon0.859 / 2.0600.813 / 1.9230.776 / 1.820
+ +(below). These elaborated samples, although not present in the observed ground truth (GT) future, are learned through the representation learning from reconstruction loss. In addition, it shows that actor-specific tokens do not induce mode collapse to only one motion. + +# 6. Ablation + +# 6.1. Reconstruction objective + +Table. 2 shows ablation studies on optimizing different loss types. Both reconstruction and regression losses individually boost prediction performance, with their joint optimization yielding even greater improvements. Table. 3 compares the effects of using only regression loss versus both losses on prediction performance across different layer depths. Updating just the decoder $(D)$ shows similar results in both scenarios, but extending updates to the encoder $(E)$ significantly enhances performance when using both losses. Furthermore, extending updates to the embedding layers $(f_{h,f,l})$ deteriorates performance when only regression loss is optimized. This highlights the importance of incorporating representation learning through the MAE, as relying solely on regression loss can lead to suboptimal adaptation and damage to learned representations. + +We also conduct ablation studies on the masking ratio for both actors and lane centerlines in Tab. 4. The result is visualized via graphs in Fig. 8 according to lane masking ratio and actor masking ratio, respectively. Around 0.3 of lane and 0.4 of actor masking ratio, tendencies of $\mathrm{mADE}_6$ follow + +Table 4. ${\mathrm{{mADE}}}_{6}$ according to actor and lane masking ratio. + +
INTER → Lyft (1/3/0.1)Lane Masking Ratio
0.10.30.50.70.9
Actor0.10.4180.4070.4640.4810.446
0.30.4230.4430.4430.4450.390
Masking Ratio0.50.4170.4550.5150.3910.435
0.70.5670.4350.4480.4530.491
0.90.4470.4210.4940.4170.398
+ +![](images/4e415c63103f80aa7a0e4edd6a56f39c46758c3de3bcadd677d9ae131bb984c3.jpg) +Figure 8. Tendency of $\mathrm{mADE}_6$ according to actor and lane masking ratio respectively. (INTER $\rightarrow$ Lyft $(1/3/0.1)$ ) + +![](images/78156583d219c4af1bd5b431da4f43dc19101a43a9b0c1fbef7de4d95e1e9685.jpg) + +a U-shape. In case of too small masking ratio, the reconstruction does not learn sufficient representation from the loss, while large masking ratio interrupt interaction learning due to absence of sufficient information. However, in both lane and actor masking, $\mathrm{mADE}_6$ gets improved when the masking ratio increases above 0.8. In that case, the reconstruction network is induced to learn scene-specific information. In addition, unlike regression loss which deteriorates performance, reconstruction loss does not harm performance because it induces learning the semantic relationship than direct regression supervision. + +# 6.2. Actor-specific token + +Table. 5 presents results for the baseline without adaptation, our method without actor-specific tokens, and our full method. The second column reveals that even without actor-specific tokens, the prediction performance is 0.581 and 0.931, surpassing MEK's 0.746 and 1.006. However, incorporating actor-specific tokens for instance-aware adaptation yields notable improvements of $32.7\%$ and $16.7\%$ for short-term and long-term experiments, respectively. The difference in performance between short-term and long-term is influenced by the scene length in the dataset. The average scene length for short-term data with a 0.1 time interval is 200.04, significantly longer than the average of 32.67 for long-term data with a 0.5 time interval. Intuitively, as scene length increases, the time spent observing previously adapted actors also increases, enhancing the effectiveness of actor-specific tokens. To verify this, Fig. 9 adjusts scene length arbitrarily by skipping to the next scene in data loading once a specific scene length is exceeded. The results confirm that as scene length decreases by skipping scenes earlier, the effectiveness of actor-specific tokens diminishes in both short-term and long-term scenarios. + +Table 5. Effect of actor-specific token in $\mathrm{mADE}_6 / \mathrm{mFDE}_6$ . The proposed method enhances adaptation performance by learning actor-wise motion characteristics. + +
Exp.BaselineOurs w/o Actor-specificOurs (Full)
INTER → Lyft (1/3/0.1)1.553 / 3.4580.581 / 1.1510.391 / 0.824
nuS → Lyft (2/6/0.5)1.108 / 2.5970.932 / 2.2200.776 / 1.820
+ +![](images/44509efe139cf590a75aa806d5674052ec86ef6891c09f3d29290fc0c408aca5.jpg) +Figure 9. Effect of scenario length to the effectiveness of actor-specific token. As the scenario length shortens with manual skipping, its effectiveness diminishes because the duration available for the actor-specific token to adapt is reduced. In real-world applications, driving scenarios are continuous, resulting in maximal efficacy of the proposed method. + +# 7. Conclusion + +We propose a test-time training method for trajectory prediction by incorporating the MAE and actor-specific token memory. The introduced MAE objective addresses a limitation of conventional online learning, preventing the loss of representations learned from source data. Consequently, our approach enables learning deeper layers, leading to improved representations and enhanced predictions even for out-of-distribution samples. The integration of actor-specific tokens during test-time allows for instance-wise learning of motion patterns, resulting in substantial performance improvements. This approach, particularly effective in continuous real-world autonomous driving scenarios without scene breaks, demonstrates significant efficacy and holds promise for practical applications. + +Acknowledgements This research was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF2022R1A2B5B03002636) and the Challengeable Future Defense Technology Research and Development Program through the Agency For Defense Development (ADD) funded by the Defense Acquisition Program Administration (DAPA) in 2024 (No.912768601). + +# References + +[1] Görkay Aydemir, Adil Kaan Akan, and Fatma Güney. Adapt: Efficient multi-agent trajectory prediction with adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8295-8305, 2023. 1 +[2] Inhwan Bae, Jin-Hwi Park, and Hae-Gon Jeon. Nonprobability sampling network for stochastic human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6477-6487, 2022. 6 +[3] Inhwan Bae, Jean Oh, and Hae-Gon Jeon. Eigentrajectory: Low-rank descriptors for multi-modal trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10017-10029, 2023. 2 +[4] Collin Burns and Jacob Steinhardt. Limitations of post-hoc feature alignment for robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2525-2533, 2021. 2 +[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5 +[6] Rohan Chandra, Uttaran Bhattacharya, Aniket Bera, and Dinesh Manocha. Traphic: Trajectory prediction in dense and heterogeneous traffic using weighted interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8483-8492, 2019. 1 +[7] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 2 +[8] Hao Chen, Jiaze Wang, Kun Shao, Furui Liu, Jianye Hao, Chenyong Guan, Guangyong Chen, and Pheng-Ann Heng. Traj-mae: Masked autoencoders for trajectory prediction. arXiv preprint arXiv:2303.06697, 2023. 2 +[9] Liang Chen, Yong Zhang, Yibing Song, Ying Shan, and Lingqiao Liu. Improved test-time adaptation for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24172-24182, 2023. 1, 2 +[10] Weihuang Chen, Zhigang Yang, Lingyang Xue, Jinghai Duan, Hongbin Sun, and Nanning Zheng. Multimodal pedestrian trajectory prediction using probabilistic proposal network. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 7 +[11] Xuesong Chen, Shaoshuai Shi, Chao Zhang, Benjamin Zhu, Qiang Wang, Ka Chun Cheung, Simon See, and Hongsheng Li. Trajectoryformer: 3d object tracking transformer with predictive trajectory hypotheses. arXiv preprint arXiv:2306.05888, 2023. 2 +[12] Yuxiao Chen, Boris Ivanovic, and Marco Pavone. Scept: Scene-consistent, policy-based trajectory predictions for planning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17103-17112, 2022. 2 + +[13] Jie Cheng, Xiaodong Mei, and Ming Liu. Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8679-8689, 2023. 2, 3, 4 +[14] Junhyeong Cho, Gilhyun Nam, Sungyeon Kim, Hunmin Yang, and Suha Kwak. Promptstyler: Prompt-driven style generation for source-free domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15702-15712, 2023. 2 +[15] Dooseop Choi and KyoungWook Min. Hierarchical latent structure for multi-modal vehicle trajectory forecasting. In European Conference on Computer Vision, pages 129-145. Springer, 2022. 2 +[16] Sehwan Choi, Jungho Kim, Junyong Yun, and Jun Won Choi. R-pred: Two-stage motion prediction via tube-query attention-based trajectory refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8525-8535, 2023. 1 +[17] Henggang Cui, Vladan Radosavljevic, Fang-Chieh Chou, Tsung-Han Lin, Thi Nguyen, Tzu-Kuo Huang, Jeff Schneider, and Nemanja Djuric. Multimodal trajectory predictions for autonomous driving using deep convolutional networks. In 2019 International Conference on Robotics and Automation (ICRA), pages 2090-2096. IEEE, 2019. 2 +[18] Ning Ding, Yixing Xu, Yehui Tang, Chao Xu, Yunhe Wang, and Dacheng Tao. Source-free domain adaptation via distribution estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7212-7222, 2022. 2 +[19] Nemanja Djuric, Vladan Radosavljevic, Henggang Cui, Thi Nguyen, Fang-Chieh Chou, Tsung-Han Lin, Nitin Singh, and Jeff Schneider. Uncertainty-aware short-term motion prediction of traffic actors for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2095-2104, 2020. 2 +[20] Yonghao Dong, Le Wang, Sanping Zhou, and Gang Hua. Sparse instance conditioned multimodal trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9763-9772, 2023. 2 +[21] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 5 +[22] Francois Fleuret et al. Uncertainty reduction for model adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9613-9623, 2021. 2 +[23] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. Advances in Neural Information Processing Systems, 35:29374-29385, 2022. 2 +[24] Thomas Gilles, Stefano Sabatini, Dzmitry Tsishkou, Bogdan Stanciulescu, and Fabien Moutarde. THOMAS: Trajectory heatmap output with learned multi-agent sampling. In In- + +ternational Conference on Learning Representations, 2022. 2 +[25] Thomas Gilles, Stefano Sabatini, Dzmitry Tsishkou, Bogdan Stanciulescu, and Fabien Moutarde. Uncertainty estimation for cross-dataset performance in trajectory prediction. arXiv preprint arXiv:2205.07310, 2022. 1, 2 +[26] Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. In International Conference on Learning Representations, 2022. 2 +[27] Taesik Gong, Jongheon Jeong, Taewon Kim, Yewon Kim, Jinwoo Shin, and Sung-Ju Lee. NOTE: Robust continual test-time adaptation against temporal correlation. In Advances in Neural Information Processing Systems, 2022. 2 +[28] Junru Gu, Chenxu Hu, Tianyuan Zhang, Xuanyao Chen, Yilun Wang, Yue Wang, and Hang Zhao. Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5496-5506, 2023. 1 +[29] Abner Guzman-Rivera, Dhruv Batra, and Pushmeet Kohli. Multiple choice learning: Learning to produce multiple structured outputs. Advances in neural information processing systems, 25, 2012. 4 +[30] John Houston, Guido Zuidhof, Luca Bergamini, Yawei Ye, Long Chen, Ashesh Jain, Sammy Omari, Vladimir Iglovikov, and Peter Ondruska. One thousand and one hours: Self-driving motion prediction dataset. In Conference on Robot Learning, pages 409-418. PMLR, 2021. 5 +[31] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1 +[32] Pingxuan Huang, Zhenhua Cui, Jing Li, Shenghua Gao, bo Hu, and Yanyan Fang. Cross-domain trajectory prediction with ctp-net, 2022. 1 +[33] Manh Huynh and Gita Alaghband. AOL: Adaptive online learning for human trajectory prediction in dynamic video scenes. arXiv preprint arXiv:2002.06666, 2020. 2 +[34] Manh Huynh and Gita Alaghband. Online adaptive temporal memory with certainty estimation for human trajectory prediction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 940-949, 2023. 2 +[35] David Isele and Akansel Cosgun. Selective experience replay for lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2 +[36] Boris Ivanovic, James Harrison, and Marco Pavone. Expanding the deployment envelope of behavior prediction via adaptive meta-learning. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7786-7793. IEEE, 2023. 1, 2, 5 +[37] Boris Ivanovic, Guanyu Song, Igor Gilitschenski, and Marco Pavone. trajdata: A unified interface to multiple human trajectory datasets. In Proceedings of the Neural Information + +Processing Systems (NeurIPS) Track on Datasets and Benchmarks, New Orleans, USA, 2023. 5 +[38] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9644-9653, 2023. 1 +[39] Ruochen Jiao, Xiangguo Liu, Takami Sato, Qi Alfred Chen, and Qi Zhu. Semi-supervised semantics-guided adversarial training for robust trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8207-8217, 2023. 2 +[40] Junho Kim, Inwoo Hwang, and Young Min Kim. Ev-tta: Test-time adaptation for event-based object recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17745-17754, 2022. 2 +[41] Mihee Lee, Samuel S Sohn, Seonghyeon Moon, Sejong Yoon, Mubbasir Kapadia, and Vladimir Pavlovic. Musevae: multi-scale vae for environment-aware long term trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2221-2230, 2022. 2 +[42] Lihuan Li, Maurice Pagnucco, and Yang Song. Graph-based spatial transformer with memory replay for multifuture pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2231-2241, 2022. 2 +[43] Maosen Li, Siheng Chen, Yanning Shen, Genjia Liu, Ivor W Tsang, and Ya Zhang. Online multi-agent forecasting with interpretable collaborative graph neural networks. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2 +[44] Yingwei Li, Charles R Qi, Yin Zhou, Chenxi Liu, and Dragomir Anguelov. Modar: Using motion forecasting for 3d object detection in point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9329-9339, 2023. 2 +[45] Yushu Li, Xun Xu, Yongyi Su, and Kui Jia. On the robustness of open-world test-time training: Self-training with dynamic prototype expansion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11836-11846, 2023. 2 +[46] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 2 +[47] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. arXiv preprint arXiv:2303.15361, 2023. 2 +[48] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecasting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 541-556. Springer, 2020. 4 +[49] Hyesu Lim, Byeonggeun Kim, Jaegul Choo, and Sungha Choi. Ttn: A domain-shift aware batch normalization in test + +time adaptation. In The Eleventh International Conference on Learning Representations, 2022. 2 +[50] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808-21820, 2021. 2 +[51] Yuejiang Liu, Riccardo Cadei, Jonas Schweizer, Sherwin Bahmani, and Alexandre Alahi. Towards robust and adaptive motion forecasting: A causal representation perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17081-17092, 2022. 1 +[52] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5517-5526, 2023. 1 +[53] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 5 +[54] M Jehanzeb Mirza, Inkyu Shin, Wei Lin, Andreas Schriebl, Kunyang Sun, Jaesung Choe, Mateusz Kozinski, Horst Possegger, In So Kweon, Kuk-Jin Yoon, et al. Mate: Masked autoencoders are online 3d test-time learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16709-16718, 2023. 2 +[55] Jiquan Ngiam, Vijay Vasudevan, Benjamin Caine, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, David J Weiss, Benjamin Sapp, Zhifeng Chen, and Jonathon Shlens. Scene transformer: A unified architecture for predicting future trajectories of multiple agents. In International Conference on Learning Representations, 2022. 1 +[56] Daehee Park, Jaewoo Jeong, and Kuk-Jin Yoon. Improving transferability for cross-domain trajectory prediction via neural stochastic differential equation. arXiv preprint arXiv:2312.15906, 2023. 2 +[57] Daehee Park, Hobin Ryu, Yunseo Yang, Jegyeong Cho, Jiwon Kim, and Kuk-Jin Yoon. Leveraging future relationship reasoning for vehicle trajectory prediction. In The Eleventh International Conference on Learning Representations, 2023. 1 +[58] Mozhgan Pourkeshavarz, Changhe Chen, and Amir Rasouli. Learn tarot with mentor: A meta-learned self-supervised approach for trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8384–8393, 2023. 2 +[59] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. Advances in neural information processing systems, 32, 2019. 2 +[60] Luke Rowe, Martin Ethier, Eli-Henry Dykhne, and Krzysztof Czarnecki. Fjmp: Factorized joint multi-agent motion prediction over learned directed acyclic interaction + +graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13745-13755, 2023. 1 +[61] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 2 +[62] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Wei Tang, Nanning Zheng, and Gang Hua. Representing multimodal behaviors with mean location for pedestrian trajectory prediction. IEEE transactions on pattern analysis and machine intelligence, 2023. 7 +[63] Liushuai Shi, Le Wang, Sanping Zhou, and Gang Hua. Trajectory unified transformer for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9675-9684, 2023. 2 +[64] Jianhua Sun, Yuxuan Li, Hao-Shu Fang, and Cewu Lu. Three steps to multimodal trajectory prediction: Modality clustering, classification and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13250-13259, 2021. 7 +[65] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In Proceedings of the 37th International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 2, 3 +[66] Devavrat Tomar, Guillaume Vray, Behzad Bozortabar, and Jean-Philippe Thiran. Tesla: Test-time self-learning with automatic adversarial augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20341–20350, 2023. 2 +[67] Li-Wu Tsao, Yan-Kai Wang, Hao-Siang Lin, Hong-Han Shuai, Lai-Kuan Wong, and Wen-Huang Cheng. Socialssl: Self-supervised cross-sequence representation learning based on transformers for multi-agent trajectory prediction. In European Conference on Computer Vision, pages 234-250. Springer, 2022. 2 +[68] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[69] Chunnan Wang, Xiang Chen, Junzhe Wang, and Hongzhi Wang. Atpfl: Automatic trajectory prediction model design under federated learning framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6563-6572, 2022. 2 +[70] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 5 +[71] Eason Wang, Henggang Cui, Sai Yalamanchi, Mohana Moorthy, and Nemanja Djuric. Improving movement predictions of traffic actors in bird's-eye view models using gans and differentiable trajectory rasterization. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2340-2348, 2020. 2 + +[72] Jingke Wang, Tengju Ye, Ziqing Gu, and Junbo Chen. Ltp: Lane-based trajectory prediction for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17134-17142, 2022. 2 +[73] Letian Wang, Yeping Hu, and Changliu Liu. Online adaptation of neural network models by modified extended kalman filter for customizable and transferable driving behavior prediction, 2022. 1, 2, 5 +[74] Letian Wang, Yeping Hu, Liting Sun, Wei Zhan, Masayoshi Tomizuka, and Changliu Liu. Transferable and adaptable driving behavior prediction. arXiv preprint arXiv:2202.05140, 2022. 2 +[75] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 2 +[76] Zhibo Wang, Jiayu Guo, Haiqiang Zhang, Ru Wan, Junping Zhang, and Jian Pu. Bridging the gap: Improving domain generalization in trajectory prediction. IEEE Transactions on Intelligent Vehicles, 2023. 1 +[77] Chenfeng Xu, Tian Li, Chen Tang, Lingfeng Sun, Kurt Keutzer, Masayoshi Tomizuka, Alireza Fathi, and Wei Zhan. Pretram: Self-supervised pre-training via connecting trajectory and map. In European Conference on Computer Vision, pages 34-50. Springer, 2022. 2 +[78] Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1410-1420, 2023. 2 +[79] Pei Xu, Jean-Bernard Hayet, and Ioannis Karamouzas. Socialvae: Human trajectory prediction using timewise latents. In European Conference on Computer Vision, pages 511-528. Springer, 2022. 2 +[80] Yi Xu, Lichen Wang, Yizhou Wang, and Yun Fu. Adaptive trajectory prediction via transferable gnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6520-6531, 2022. 1, 2 +[81] Yi Xu, Armin Bazarjani, Hyung-gun Chi, Chiho Choi, and Yun Fu. Uncovering the missing pattern: Unified framework towards trajectory imputation and prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9632-9643, 2023. 1 +[82] Zhijie Yan, Pengfei Li, Zheng Fu, Shaocong Xu, Yongliang Shi, Xiaoxue Chen, Yuhang Zheng, Yang Li, Tianyu Liu, Chuxuan Li, et al. Int2: Interactive trajectory prediction at intersections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8536-8547, 2023. 1 +[83] Tao Yang, Shenglong Zhou, Yuwang Wang, Yan Lu, and Nanning Zheng. Test-time batch normalization. arXiv preprint arXiv:2205.10210, 2022. 2 +[84] Luyao Ye, Zikang Zhou, and Jianping Wang. Improving the generalizability of trajectory prediction models with freenet-based domain normalization. 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11562-11568, 2023. 2 + +[85] Jiangbei Yue, Dinesh Manocha, and He Wang. Human trajectory prediction via neural social physics. In European Conference on Computer Vision, pages 376-394. Springer, 2022. 2 +[86] Wei Zhan, Liting Sun, Di Wang, Haojie Shi, Aubrey Clausse, Maximilian Naumann, Julius Kummerle, Hendrik Konigshof, Christoph Stiller, Arnaud de La Fortelle, et al. Interaction dataset: An international, adversarial and cooperative motion dataset in interactive driving scenarios with semantic maps. arXiv preprint arXiv:1910.03088, 2019. 5 +[87] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems, 35: 38629-38642, 2022. 2 +[88] Xingxuan Zhang, Peng Cui, Renzhe Xu, Linjun Zhou, Yue He, and Zheyan Shen. Deep stable learning for out-of-distribution generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5372-5382, 2021. 1 +[89] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863-17873, 2023. 1 +[90] Dekai Zhu, Guangyao Zhai, Yan Di, Fabian Manhardt, Hendrik Berkemeyer, Tuan Tran, Nassir Navab, Federico Tombari, and Benjamin Busam. Ipcc-tp: Utilizing incremental pearson correlation coefficient for joint multi-agent trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5507-5516, 2023. 1 +[91] Yiyao Zhu, Di Luan, and Shaojie Shen. Biff: Bi-level future fusion with polyline-based coordinate for interactive trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8260-8271, 2023. 2 \ No newline at end of file diff --git a/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/images.zip b/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c17eeb068a525e3eb870b99551d7cf2782ca9745 --- /dev/null +++ b/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:126b644f4273635a0ae0424ba8fa9955f1675502ce0b7d1c85e2be0424cc74e9 +size 622809 diff --git a/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/layout.json b/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a91ca8fe6072e36cc53f8315af044faee7a5b1d7 --- /dev/null +++ b/2024/T4P_ Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory/layout.json @@ -0,0 +1,9833 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 529, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 529, + 140 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 529, + 140 + ], + "type": "text", + "content": "T4P: Test-Time Training of Trajectory Prediction via Masked Autoencoder and Actor-specific Token Memory" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 161, + 493, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 161, + 493, + 188 + ], + "spans": [ + { + "bbox": [ + 100, + 161, + 493, + 188 + ], + "type": "text", + "content": "Daehee Park, Jaeseok Jeong, Sung-Hoon Yoon, Jaewoo Jeong, and Kuk-Jin Yoon KAIST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 191, + 449, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 191, + 449, + 204 + ], + "spans": [ + { + "bbox": [ + 140, + 191, + 449, + 204 + ], + "type": "text", + "content": "{bag2824, jason. jeong, yoon307, jeong207, kjyoon}@kaist.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 258, + 290, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 558 + ], + "type": "text", + "content": "Trajectory prediction is a challenging problem that requires considering interactions among multiple actors and the surrounding environment. While data-driven approaches have been used to address this complex problem, they suffer from unreliable predictions under distribution shifts during test time. Accordingly, several online learning methods have been proposed using regression loss from the ground truth of observed data leveraging the auto-labeling nature of trajectory prediction task. We mainly tackle the following two issues. First, previous works underfit and overfit as they only optimize the last layer of motion decoder. To this end, we employ the masked autoencoder (MAE) for representation learning to encourage complex interaction modeling in shifted test distribution for updating deeper layers. Second, utilizing the sequential nature of driving data, we propose an actor-specific token memory that enables the test-time learning of actor-wise motion characteristics. Our proposed method has been validated across various challenging cross-dataset distribution shift scenarios including nuScenes, Lyft, Waymo, and Interaction. Our method surpasses the performance of existing state-of-the-art online learning methods in terms of both prediction accuracy and computational efficiency. The code is available at https://github.com/daeheepark/T4P." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Trajectory prediction plays a significant role in autonomous systems, enhancing safety and navigation efficiency [28, 31]. Recently, data-driven methods have shown remarkable prediction capabilities [1, 16, 38, 52, 55, 57, 60, 81, 89, 90]; however, they are prone to distribution shift [9, 88]. Trajectory prediction models also produce unreliable output when faced shifts in the data distribution [25], posing significant risks in various real-world applications. This vulnerability stems from the ease with which the trajectory data distri" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 232, + 531, + 420 + ], + "blocks": [ + { + "bbox": [ + 317, + 232, + 531, + 420 + ], + "lines": [ + { + "bbox": [ + 317, + 232, + 531, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 531, + 420 + ], + "type": "image", + "image_path": "33acb881603285214330de90b83340b0d3fc7c37f8a9239ca2e9d5f9f611682c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 426, + 547, + 492 + ], + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 492 + ], + "type": "text", + "content": "Figure 1. Previous methods optimize the last layer of the decoder using regression loss from delayed ground truth. Our method, on the other hand, learns representation via a masked autoencoder, which boosts prediction performance by optimizing deeper layers. In addition, the proposed actor-specific token enables the prediction model to learn actor-wise motion characteristics." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 507, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 547, + 544 + ], + "type": "text", + "content": "bution can be altered by numerous factors, such as scene changes and driving habits; i.e. road layout, interaction between agents, driver demographics [6, 51, 82]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "text", + "content": "To address this challenge, recent methods have proposed domain adaptation and generalization strategies which aim to anticipate the distribution shifts and accordingly train the model [32, 76, 80]. However, due to the wide variety of factors influencing data distribution, the anticipated shifts may differ significantly from those encountered at test time. As a result, several online learning methods have been developed to dynamically adapt models during test time [36, 73]. Since trajectory prediction serves as an auto-labeling task where trajectory data is obtained from object tracking, the observed past and future trajectory provide both input and corresponding ground truth for supervision " + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{reg})" + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "text", + "content": ", as depicted at the top right of Fig. 1. Nevertheless, updating the entire model may ruin the representation learned from the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 1, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 1, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 1, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15065" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "source data, so only part of the network, such as the batch normalization layer, is updated [27, 40, 61, 83]. Particularly, because regression loss is calculated at delayed timestamps and only a few samples are available during test time, this approach risks deteriorating the model's learned representation [73]. Therefore, previous online prediction methods mainly update only the last layer of the decoder." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 290, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 290, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 290, + 419 + ], + "type": "text", + "content": "In this work, we propose a test-time training (TTT) for trajectory prediction with two key aspects. First, we build a masked autoencoder (MAE) framework to adapt deep features, incorporating good representation that captures complex interactions between agents and road structures. Due to the challenge of existing online learning methods in damaging learned representations when updating deeper layers, we employ a MAE to guide representation learning. Second, we introduce an actor-specific token memory that has significant advantages in real-world driving scenario where data arrives continuously and sequentially. As each actor instance has its own driving habits and the past motion pattern of specific actors can be accessed from the arrived observations, we design a token memory in transformer structure [68] and its training strategy to learn actor-wise motion characteristics. The proposed TTT framework is validated on challenging cross-dataset distribution shift cases between nuScenes, Lyft, Waymo, and INTERACTION, and shows state-of-the-art performance surpassing previous online learning methods. Furthermore, we show the practicality of our TTT framework by evaluating its computational efficiency. We summarize our contributions as below:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 421, + 287, + 528 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 421, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 421, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 287, + 468 + ], + "type": "text", + "content": "- We propose a test time training for trajectory prediction (T4P) by utilizing a masked autoencoder to learn deep feature representations that stably improve prediction performance across entire network layers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 469, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 469, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 287, + 491 + ], + "type": "text", + "content": "- We introduce an actor-specific token memory used to learn the different actor characteristics and habits." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 492, + 287, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 492, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 492, + 287, + 528 + ], + "type": "text", + "content": "- Our method is validated across 4 different datasets as well as different temporal configurations. Ours achieves state-of-the-art performance both in accuracy and efficiency." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 542, + 138, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 138, + 554 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 138, + 554 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 563, + 171, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 171, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 171, + 575 + ], + "type": "text", + "content": "2.1. Trajectory Prediction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": "Trajectory prediction garners attention with the emergence of methods that can enhance perception or planning [11, 12, 44]. Its goal is to predict future trajectories of traffic actors based on their historical trajectories and the context of their environment [3, 17, 19, 20, 39, 58, 63, 78, 91]. Historical trajectories, or tracklets of traffic actors, are sequentially acquired via vehicle detection and tracking systems. Some studies employed this temporal property to enhance prediction via memory replay [35, 42, 59]. In the early stages of trajectory prediction, only the historical trajectory of the actors of interest was considered. However, recent stud" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": "ies emphasize the significance of understanding interactions among agents [67, 85] and the rules governed by surrounding environments [72, 77] in improving prediction performance. This has led to the development of models that incorporate multi-head attention or graph-based methods to capture these interactions [24, 26, 42]. Additionally, MAE has been adopted for pretraining to better understand agent interactions [8, 13]. To further refine prediction capabilities, various generative models have been introduced, enabling the generation of future trajectories [15, 41, 71, 79]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 201, + 531, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 201, + 531, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 531, + 213 + ], + "type": "text", + "content": "2.2. Transfer Learning in Trajectory Prediction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 219, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 471 + ], + "type": "text", + "content": "With data-driven approaches offering superior performance in trajectory prediction, their effectiveness diminishes under distribution shifts [25, 56]. In response, several studies have adopted domain adaptation or generalization strategy [69, 74]. Some specifically aimed to reduce the domain gap within unique characteristics of trajectory prediction: differences in road structures [84], actor interaction [80], etc. However, these methods rely on anticipating how to cover domain shifts. Yet, given that trajectory data is subject to influence from numerous factors, predicting and accommodating for shifts may not always be sufficient. Consequently, recent developments have introduced adaptation to unseen test sets using online learning [33, 34, 43]. Among them, some methods [36, 73] showed remarkable prediction performance improvement under severe distribution shifts like cross-dataset cases by utilizing regression loss for online learning. These methods exploit the fact that the input and ground truth (GT) labels are provided at test time as tracking history. However, with the limitations of updating with only a few samples in a delayed time, adaptation becomes restricted to the last layer of the decoder." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 479, + 419, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 479, + 419, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 479, + 419, + 491 + ], + "type": "text", + "content": "2.3. Test Time Training" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Test-time training (TTT) is a method that trains the network on test time data, unseen during training [4, 7, 14, 18, 22, 49, 66, 75]. Unlike domain generalization or adaptation, which are confined to the training phase, TTT extends model adaptation into the test phase by utilizing available test data [47]. TTT methods are categorized into regularization-based approaches for post-hoc regularization of out-of-distribution (OOD) samples [46, 87], and self-supervised approaches that employ pretext tasks on test data for optimal representation learning [7, 9, 45, 50, 54]. Specifically, TTT [65] introduced a Y-shaped network structure consisting of a feature encoder, a pretext branch, and a decoder branch. The decoder branch is fixed, while the encoder and pretext branch are optimized through self-supervision. Adhering to this model, TTT-MAE [23] integrated a MAE in the pretext task. Expanding on this method, we adopt TTT-MAE to the domain of trajectory prediction, leveraging its representation learning capabilities to enhance test-time training." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15066" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 73, + 542, + 264 + ], + "blocks": [ + { + "bbox": [ + 52, + 73, + 542, + 264 + ], + "lines": [ + { + "bbox": [ + 52, + 73, + 542, + 264 + ], + "spans": [ + { + "bbox": [ + 52, + 73, + 542, + 264 + ], + "type": "image", + "image_path": "f3b3e43995adeb1299905a2a1f4552e9eb13aeb779576994b464a229953c164b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 268, + 545, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 268, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 545, + 313 + ], + "type": "text", + "content": "Figure 2. Overall method. During test-time training, the network trained on source dataset is optimized on target data under online setting. The model is optimized both from regression and reconstruction loss. Both losses utilize the data observed at the delayed time stamp " + }, + { + "bbox": [ + 46, + 268, + 545, + 313 + ], + "type": "inline_equation", + "content": "(t_{\\tau})" + }, + { + "bbox": [ + 46, + 268, + 545, + 313 + ], + "type": "text", + "content": ". Actor-specific token is used to learn instance-wise motion pattern during test-time training phase. During online evaluation phase, model and actor-specific token learned from test-time training phase are used." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 325, + 102, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 325, + 102, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 325, + 102, + 337 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 345, + 157, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 345, + 157, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 345, + 157, + 357 + ], + "type": "text", + "content": "3.1. Problem definition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "text", + "content": "Trajectory prediction aims to learn the mapping function between the input, consisting of historical trajectory and map information " + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{X}:\\{\\mathcal{X}_t,\\mathcal{M}_t\\}" + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "text", + "content": ", and the output, consisting of " + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "text", + "content": " possible candidates for future trajectory of " + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "text", + "content": " traffic actors, " + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}:\\{\\mathcal{Y}_t^{0:K - 1}\\}" + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "text", + "content": ", at current time " + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "text", + "content": ". We predict " + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "text", + "content": " different actor classes including vehicle, cyclist, etc. Historical and future trajectories are represented as " + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_t = \\mathbf{x}_{t - t_h:t}^{0:N - 1}" + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_t = \\mathbf{x}_{t:t + t_f}^{0:N - 1}" + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "inline_equation", + "content": "t_h" + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "text", + "content": " represent sequence length of input and output trajectory. Here, " + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^n" + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "text", + "content": " represents the spatial location of actor " + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "text", + "content": ". For map information " + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 363, + 287, + 508 + ], + "type": "text", + "content": " segmented lane centerlines around ego-actors which is widely-used in trajectory prediction methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 509, + 288, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 288, + 666 + ], + "type": "text", + "content": "We deal with the case when the target data distribution during test time " + }, + { + "bbox": [ + 46, + 509, + 288, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X},\\mathbf{Y}\\}^T" + }, + { + "bbox": [ + 46, + 509, + 288, + 666 + ], + "type": "text", + "content": " is different from the source data distribution seen during the training phase " + }, + { + "bbox": [ + 46, + 509, + 288, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X},\\mathbf{Y}\\}^S" + }, + { + "bbox": [ + 46, + 509, + 288, + 666 + ], + "type": "text", + "content": ". We formulate the problem as a online adaptation scenario in which one data sample is given per each time interval as time passes. The test data is consists of multiple distinct scenes. Each scene includes temporally ordered data samples which are captured through real-world driving. Following standard TTT methods, there is no access to the source data at test time. However, thanks to the auto-labeling nature of trajectory prediction, there is access to delayed GT future trajectory " + }, + { + "bbox": [ + 46, + 509, + 288, + 666 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_{t_{\\tau}:t_{\\tau} + t_f})" + }, + { + "bbox": [ + 46, + 509, + 288, + 666 + ], + "type": "text", + "content": " from a previous time window at time " + }, + { + "bbox": [ + 46, + 509, + 288, + 666 + ], + "type": "inline_equation", + "content": "t_\\tau (= t - \\tau)" + }, + { + "bbox": [ + 46, + 509, + 288, + 666 + ], + "type": "text", + "content": " as depicted in left lower corner of Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 671, + 144, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 144, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 144, + 682 + ], + "type": "text", + "content": "3.2. Overall method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Our method, Test-Time Training of Trajectory Prediction (T4P), enhances the online learning method using supervi" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 327, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 459 + ], + "type": "text", + "content": "sion from a delayed GT future trajectory with representation learning from MAE and actor-specific token memory. Following standard TTT [65], the overall framework consists of three phases: offline training, test-time training, and online evaluation. Offline training occurs before test time, and test-time training and online evaluation are executed repeatedly and sequentially during test time. We adopt the ForecastMAE [13] backbone consisting of embedding layers " + }, + { + "bbox": [ + 304, + 327, + 545, + 459 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 327, + 545, + 459 + ], + "type": "text", + "content": ", a shared encoder " + }, + { + "bbox": [ + 304, + 327, + 545, + 459 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 327, + 545, + 459 + ], + "type": "text", + "content": ", a reconstruction head " + }, + { + "bbox": [ + 304, + 327, + 545, + 459 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 327, + 545, + 459 + ], + "type": "text", + "content": " and a motion decoder head " + }, + { + "bbox": [ + 304, + 327, + 545, + 459 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 327, + 545, + 459 + ], + "type": "text", + "content": ", as depicted in the middle of Fig. 2. The detailed methods during each phase are described below:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 465, + 402, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 402, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 402, + 478 + ], + "type": "text", + "content": "3.3. Offline training" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 483, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 508 + ], + "type": "text", + "content": "During offline training, the model is trained on source data using both reconstruction loss and regression loss." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 333, + 516, + 545, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 516, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 333, + 516, + 545, + 536 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta \\in \\{f, E, R, D \\}} \\mathbb {E} _ {\\mathcal {X}, \\mathcal {Y}, \\mathcal {M} \\in \\{\\mathbf {X}, \\mathbf {Y} \\} ^ {S}} \\left[ \\mathcal {L} _ {\\text {r e c o n}} + \\mathcal {L} _ {\\text {r e g}} \\right] \\tag {1}", + "image_path": "acbc23dbc0b6965c3f14b5fc2a31156027262293a00d9995b8e7158f2b07251e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "text", + "content": "In this subsection, subscript " + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "text", + "content": " is omitted for simplicity. First, all input elements " + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "inline_equation", + "content": "(\\mathcal{X},\\mathcal{Y},\\mathcal{M})" + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "text", + "content": " are embedded with their respective embedding layer " + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "inline_equation", + "content": "(f_x,f_y,f_m)" + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "text", + "content": ". Additionally, we define actor class token " + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}\\in \\mathbb{R}^{C\\times D}" + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "text", + "content": " that learns different motion patterns of each actor class. The actor class token is implemented as a learnable embedding of a transformer structure. Corresponding actor class token " + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\alpha (c)" + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "text", + "content": " is added to trajectory embedding according to the class of each actor." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 644, + 545, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 644, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 312, + 644, + 545, + 658 + ], + "type": "interline_equation", + "content": "h _ {x}, h _ {y}, h _ {m} = f _ {x} (\\mathcal {X}) + \\alpha (c), f _ {y} (\\mathcal {Y}) + \\alpha (c), f _ {m} (\\mathcal {M}) \\tag {2}", + "image_path": "8875c0e7509886a66b7cd916a9fa0bca7b8dcd839652cf085daabe66b6194fa4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "text", + "content": "For reconstruction, the history/future trajectory and lane embeddings are fed to the encoder to obtain the encodings " + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "inline_equation", + "content": "F_{x}, F_{y}, F_{m}" + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "text", + "content": ". Then, segments of the encodings are randomly masked and replaced with masking tokens, " + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "inline_equation", + "content": "M_{x}, M_{y}, M_{m}" + }, + { + "bbox": [ + 304, + 666, + 545, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15067" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": "and the other encodings remain unmasked " + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "inline_equation", + "content": "(F_x', F_y', F_m')" + }, + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": ". Here, we use random masking for the lane centerline and complementary masking strategy for history/future trajectory following previous works [13]. The masking tokens and the unmasked encodings are fed to the reconstructor to reconstruct the masked elements." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 80, + 149, + 287, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 149, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 80, + 149, + 287, + 169 + ], + "type": "interline_equation", + "content": "F _ {x}, F _ {y}, F _ {m} = E \\left(h _ {x}, h _ {y}, h _ {m}\\right) \\tag {3}", + "image_path": "a205a4ae9023af3d8e58b932dace36174b523f593aa2ac7d6b524145d38a23ab.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 81, + 167, + 253, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 167, + 253, + 180 + ], + "spans": [ + { + "bbox": [ + 81, + 167, + 253, + 180 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {X}}, \\hat {\\mathcal {Y}}, \\hat {\\mathcal {M}} = R \\left(M _ {x}, M _ {y}, M _ {m}, F _ {x} ^ {\\prime}, F _ {y} ^ {\\prime}, F _ {m} ^ {\\prime}\\right)", + "image_path": "d1e5f538c3434a3e1009632fcfade1f8a4942391ae760fd1ad641a40a7b5f6ee.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "text", + "content": "The encoder and reconstructor both consist of multi-head attention to utilize interaction between history, future and lanes, thus, the reconstruction guides to the model have the capability of interaction reasoning. Finally, a reconstruction loss " + }, + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{recon}}" + }, + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "text", + "content": " is computed as the MSE loss between the ground truth and the reconstructed outputs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 263, + 287, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 263, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 287, + 298 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{N} \\sum_ {n} (\\mathcal {X} - \\hat {\\mathcal {X}}) ^ {2} + \\frac {1}{N} \\sum_ {n} (\\mathcal {Y} - \\hat {\\mathcal {Y}}) ^ {2} + \\frac {1}{L} \\sum_ {l} (\\mathcal {M} - \\hat {\\mathcal {M}}) ^ {2} \\tag {4}", + "image_path": "aef8d5c92bbfd6ee566563a612d51c8418c20b0db4d12571cb0289fcea6c75c5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 299, + 287, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 299, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 287, + 370 + ], + "type": "text", + "content": "For the decoder head, historical trajectory and lanes embeddings are again fed to the same encoder and the output encodings are passed to the motion decoder, composed of MLP layers. The decoder outputs " + }, + { + "bbox": [ + 46, + 299, + 287, + 370 + ], + "type": "inline_equation", + "content": "\\mathrm{K}" + }, + { + "bbox": [ + 46, + 299, + 287, + 370 + ], + "type": "text", + "content": " candidates for trajectory prediction, and the regression loss " + }, + { + "bbox": [ + 46, + 299, + 287, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 46, + 299, + 287, + 370 + ], + "type": "text", + "content": " is computed with the widely-used Winner-takes-all (WTA) loss [29, 48]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 90, + 379, + 195, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 379, + 195, + 393 + ], + "spans": [ + { + "bbox": [ + 90, + 379, + 195, + 393 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {Y}} ^ {0: K - 1} = D \\left(E \\left(h _ {h}, h _ {l}\\right)\\right)", + "image_path": "560246427a4d0fbfcdef9389a380d5e5ebc6d8b0377cb688d5eaaefa22e43590.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 396, + 287, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 396, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 91, + 396, + 287, + 423 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e g} = \\frac {1}{N} \\sum_ {n} \\operatorname {a r g m i n} _ {k \\in K} \\left(\\mathcal {Y} ^ {n} - \\hat {\\mathcal {Y}} ^ {n, k}\\right) ^ {2} \\tag {5}", + "image_path": "5508c24a1f826e17769e651b4a1b9856778ca2148522b70c09839827b8ab65f5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 430, + 154, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 154, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 154, + 442 + ], + "type": "text", + "content": "3.4. Test-time training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 448, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 531 + ], + "type": "text", + "content": "During test-time, a data sample consisting of trajectories and maps arrives sequentially. Therefore, even though we cannot access the GT future trajectory of current time " + }, + { + "bbox": [ + 46, + 448, + 287, + 531 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}_t)" + }, + { + "bbox": [ + 46, + 448, + 287, + 531 + ], + "type": "text", + "content": ", we can access both the inputs and GT " + }, + { + "bbox": [ + 46, + 448, + 287, + 531 + ], + "type": "inline_equation", + "content": "(\\mathcal{X}_{t_{\\tau}},\\mathcal{Y}_{t_{\\tau}},\\mathcal{M}_{t_{\\tau}})" + }, + { + "bbox": [ + 46, + 448, + 287, + 531 + ], + "type": "text", + "content": " at a previous time " + }, + { + "bbox": [ + 46, + 448, + 287, + 531 + ], + "type": "inline_equation", + "content": "t_\\tau" + }, + { + "bbox": [ + 46, + 448, + 287, + 531 + ], + "type": "text", + "content": ". With this data, the model is optimized with the same objective as Eq. 1 with target data distribution instead of source data distribution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 538, + 287, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 538, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 69, + 538, + 287, + 555 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta \\in \\{f, E, R, D \\}} \\mathbb {E} _ {\\chi_ {t _ {\\tau}}, \\mathcal {Y} _ {t _ {\\tau}}, \\mathcal {M} _ {t _ {\\tau}} \\in \\{\\mathbf {X}, \\mathbf {Y} \\} ^ {T}} \\left[ \\mathcal {L} _ {\\text {r e c o n}} + \\mathcal {L} _ {\\text {r e g}} \\right] \\tag {6}", + "image_path": "4b5703986ebb37c0a95be17ecbe43006ac9b89f5e3b76e3d16e546d37af6410f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 560, + 287, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 560, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 287, + 644 + ], + "type": "text", + "content": "Unlike existing online learning methods that only utilize regression loss, we incorporate an additional reconstruction loss. This enables the model to learn a good representation that considers the complex actor-actor and actor-lane interaction even in the unseen target data distribution. An advantage of representation learning is that the performance stably improves even when the deeper layers are optimized." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 658, + 203, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 203, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 203, + 671 + ], + "type": "text", + "content": "3.4.1 Actor-specific token memory" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Unlike during the offline training phase, when the data order is shuffled, data at test-time comes in sequentially. Therefore, at test time, it is possible to keep track of movement" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 541, + 232 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 541, + 232 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 541, + 232 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 541, + 232 + ], + "type": "image", + "image_path": "e535b160ff33fd972055114707520df3b58efa713657fe5868a37432b5d83e17.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 237, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 305, + 237, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 237, + 545, + 304 + ], + "type": "text", + "content": "Figure 3. Actor-specific token memory is colored in gray. It evolves as time passes within a scene. For newborn actors, the corresponding class token is registered. Until the actor disappears, the token is updated through test-time training. At the end of the scene, all tokens are averaged by each class and passed to the next scene as denoted in red arrow and Eq. 7." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 314, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 314, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 314, + 545, + 338 + ], + "type": "text", + "content": "patterns of a specific actor instance. Using this, we propose an actor-specific token memory." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 341, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 341, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 305, + 341, + 545, + 557 + ], + "type": "text", + "content": "The overall scheme is described in Fig. 3. During offline training, actor class tokens " + }, + { + "bbox": [ + 305, + 341, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{train} \\in \\mathbb{R}^{C \\times D}" + }, + { + "bbox": [ + 305, + 341, + 545, + 557 + ], + "type": "text", + "content": " are trained to reflect the average motion pattern of each of the " + }, + { + "bbox": [ + 305, + 341, + 545, + 557 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 341, + 545, + 557 + ], + "type": "text", + "content": " classes. At the beginning of the test-time, scene 0, class tokens are initialized from that of training phase (" + }, + { + "bbox": [ + 305, + 341, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{\\text{scene}(0)} \\gets \\bar{\\alpha}_{\\text{train}}" + }, + { + "bbox": [ + 305, + 341, + 545, + 557 + ], + "type": "text", + "content": "). When a new " + }, + { + "bbox": [ + 305, + 341, + 545, + 557 + ], + "type": "inline_equation", + "content": "n^{th}" + }, + { + "bbox": [ + 305, + 341, + 545, + 557 + ], + "type": "text", + "content": "-actor appears at time " + }, + { + "bbox": [ + 305, + 341, + 545, + 557 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 341, + 545, + 557 + ], + "type": "text", + "content": ", the class token " + }, + { + "bbox": [ + 305, + 341, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\alpha_{n}^{t}(c)" + }, + { + "bbox": [ + 305, + 341, + 545, + 557 + ], + "type": "text", + "content": " is cloned from " + }, + { + "bbox": [ + 305, + 341, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{\\text{scene}(0)}" + }, + { + "bbox": [ + 305, + 341, + 545, + 557 + ], + "type": "text", + "content": " by selecting corresponding class. The newborn tokens are then registered to the actor-specific token memory. The token memory is structured as a dictionary where actor instance ID/corresponding tokens are key/values. At each iteration, the actor-specific tokens are used for both test-time training and online evaluation. As time progresses, the actor-specific tokens evolve and are updated through the reconstruction and regression losses until the actor disappears in the scene. By giving each actor its own specific token that distinguishes it from the sharing of other parts of models with other actors, actor-specific motion patterns can be learned." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "text", + "content": "When the scene changes, scene 1, the actors observed during scene 0 are not to be observed anymore, so we need a new averaged actor class token " + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{\\text{scene}(1)}" + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "text", + "content": ". For that, we average all the tokens in the memory at the final time step " + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "T_{S}" + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "text", + "content": " of scene 0 after gathering by classes as Eq. 7. Here, " + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "text", + "content": " refers to the number of actors of class " + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "text", + "content": ". It is because, with a sufficient number of actors class tokens per each class in memory, their average motion can be a representative motion pattern of actor classes. This is more useful than " + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{\\text{train}}" + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "text", + "content": " because newly averaged tokens are trained on the target dataset while " + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{\\text{train}}" + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "text", + "content": " contains motion pattern trained on the source dataset. The averaged class tokens are then passed to the next scene and used to initialize tokens for the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15068" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 97, + 545, + 210 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "type": "text", + "content": "Table 1. Adaptation results in various distribution shifts. The model is trained on source dataset, and test-time trained and evaluated on target dataset (Source " + }, + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "type": "text", + "content": " Target). All metrics are better in lower value. The best and second-best results are marked in bold and underline." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 97, + 545, + 210 + ], + "lines": [ + { + "bbox": [ + 49, + 97, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 49, + 97, + 545, + 210 + ], + "type": "table", + "html": "
mADE6/mFDE6Short-term exp (1/3/0.1)Long-term exp (2/6/0.5)
INTER → nuSINTER → LyftnuS → WayMeannuS → LyftWay → LyftWay → nuSMean
Source Only1.047 / 2.2471.391 / 2.9450.431 / 1.0310.956 / 2.0741.122 / 2.5770.621 / 1.3471.153 / 2.2200.965 / 2.048
Joint Training1.116 / 2.4451.553 / 3.4580.472 / 1.1251.047 / 2.3431.108 / 2.5970.638 / 1.4041.091 / 2.0310.946 / 2.011
DUA1.118 / 2.4551.516 / 3.3520.516 / 1.2941.050 / 2.3671.365 / 3.2570.790 / 1.8681.270 / 2.6341.142 / 2.586
TENT (w/ sup)1.102 / 2.4231.519 / 3.4050.448 / 1.0711.023 / 2.3001.068 / 2.5140.628 / 1.3811.077 / 2.0120.924 / 1.969
MEK (τ = tf/2)1.012 / 2.4451.283 / 3.4580.445 / 1.1250.913 / 2.3431.079 / 2.5970.629 / 1.4041.079 / 2.0310.929 / 2.011
MEK (τ = tf)0.892 / 1.9520.746 / 1.6540.405 / 1.0610.691 / 1.5561.006 / 2.3690.615 / 1.3511.117 / 2.1400.913 / 1.953
AML (K0)2.093 / 4.6972.695 / 6.6771.624 / 2.1392.137 / 4.5041.787 / 3.0671.322 / 2.5711.618 / 2.9991.576 / 2.879
AML (full)1.149 / 2.5501.042 / 2.6160.764 / 1.7910.985 / 2.3191.462 / 2.5730.977 / 2.1841.495 / 2.9781.311 / 2.578
Ours (T4P)0.537 / 1.1370.391 / 0.8240.336 / 0.8070.421 / 0.9230.776 / 1.8200.549 / 1.1710.996 / 1.7840.774 / 1.592
", + "image_path": "52282f1f8b365f2b19f0c7c4c3f5d0647264d0561bf35ca9874e4f13495e909e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 220, + 288, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 288, + 269 + ], + "type": "text", + "content": "newborn actors. Please note that while scene 0 is initialized by actor class tokens from the training phase, subsequent scenes obtain as in Eq. 7. More details of memory evolving strategies can be found on the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 273, + 287, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 273, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 83, + 273, + 287, + 308 + ], + "type": "interline_equation", + "content": "\\bar {\\alpha} _ {s c e n e (i + 1)} \\leftarrow \\left\\{\\frac {1}{N _ {c}} \\sum_ {n} ^ {N _ {c}} \\alpha_ {n} ^ {T} (c) \\right\\} _ {s c e n e (i)} \\tag {7}", + "image_path": "a864fbc2c567b46aee25e7a7577aa448adc6548d6eae68f7dc3c21bd21247a5d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 312, + 153, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 153, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 153, + 324 + ], + "type": "text", + "content": "3.5. Online evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 331, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 392 + ], + "type": "text", + "content": "Using the updated model weight and actor-specific token memory during test-time training, online evaluation is executed. With the input data " + }, + { + "bbox": [ + 46, + 331, + 287, + 392 + ], + "type": "inline_equation", + "content": "(\\mathcal{X}_t,\\mathcal{M}_t)" + }, + { + "bbox": [ + 46, + 331, + 287, + 392 + ], + "type": "text", + "content": " at current time " + }, + { + "bbox": [ + 46, + 331, + 287, + 392 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 331, + 287, + 392 + ], + "type": "text", + "content": ", the learned encoder and decoder predict multi-modal trajectory " + }, + { + "bbox": [ + 46, + 331, + 287, + 392 + ], + "type": "inline_equation", + "content": "(\\mathcal{Y}_t)" + }, + { + "bbox": [ + 46, + 331, + 287, + 392 + ], + "type": "text", + "content": " for all actors in the sample." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 400, + 124, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 400, + 124, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 400, + 124, + 414 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 419, + 110, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 110, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 110, + 431 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 437, + 287, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 606 + ], + "type": "text", + "content": "We conducted experiments on well-known datasets, nuScenes [5], Lyft [30], WOMD [21], and INTERACTION [86], to evaluate T4P on various data distribution shifts. These datasets are parsed into the same format using trajdata [37]. Additionally, to verify in various prediction configurations, experiments were conducted with the two most widely used configurations of long-term and short-term prediction. Long-term prediction requires predicting 6 seconds into the future given 2 seconds of the past with a time interval of 0.5s, making the input/output sequence lengths to be 5 and 12, respectively. Short-term prediction requires predicting 3 seconds into the future given 0.9 seconds of the past with a time interval of 0.1s, making the input/output sequence lengths to be 10 and 30, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 611, + 178, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 178, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 178, + 624 + ], + "type": "text", + "content": "4.2. Implementation details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "For actor classes, we use the 5 classes: unknown, vehicle, pedestrian, bicycle and motorcycle. Our method predicts " + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "K = 6" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": " future candidates for all actors in the sample. We use " + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": " to enable the past GT future to contain full prediction horizon. We train and evaluate our model with a single NVIDIA A6000. Learning rates of model weight and actor-specific parameters are set as 0.01 and 0.5, respectively, and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "type": "text", + "content": "weight decay is set to 0.001 for all. The gradient is clipped by 15. For metrics, widely used " + }, + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "type": "inline_equation", + "content": "\\mathrm{mADE}_6" + }, + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "type": "inline_equation", + "content": "\\mathrm{mFDE}_6" + }, + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "type": "text", + "content": " are used. Detailed metric definition, model architecture, and training details are included in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 277, + 372, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 277, + 372, + 290 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 372, + 290 + ], + "type": "text", + "content": "4.3. Baselines" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 297, + 545, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 345 + ], + "type": "text", + "content": "We compare our " + }, + { + "bbox": [ + 304, + 297, + 545, + 345 + ], + "type": "inline_equation", + "content": "T4P" + }, + { + "bbox": [ + 304, + 297, + 545, + 345 + ], + "type": "text", + "content": " with several baselines, including unsupervised/supervised test-time-training methods and online learning trajectory prediction methods. All baseline methods are implemented using the same backbone." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 346, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 370 + ], + "type": "text", + "content": "Source only refers to the backbone model trained on the source dataset only using regression loss." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 370, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 545, + 395 + ], + "type": "text", + "content": "Joint training is similar to source only but trained with regression and reconstruction loss jointly." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "text", + "content": "DUA [53] is an unsupervised post-hoc regularization method only updates batch normalization statistics in a momentum-updating manner without back-propagation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 432, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 479 + ], + "type": "text", + "content": "TENT with supervision is a variant of the original TENT [70] in which regression loss is used to optimize the batch normalization layers instead of entropy minimization loss, as entropy minimization is not applicable." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 480, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 545, + 542 + ], + "type": "text", + "content": "MEK [73] is an online learning trajectory prediction method utilizing the Modified Extended Kalman filter. It uses only regression loss to optimize the last layer of the decoder. As the prediction horizon is different in our experiment from the original paper, we use both " + }, + { + "bbox": [ + 304, + 480, + 545, + 542 + ], + "type": "inline_equation", + "content": "\\frac{1}{2} t_f" + }, + { + "bbox": [ + 304, + 480, + 545, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 480, + 545, + 542 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 304, + 480, + 545, + 542 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 541, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 541, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 545, + 613 + ], + "type": "text", + "content": "AML [36] is an Adaptive Meta-learning method. Unlike the other methods that use the same backbone, AML replaces the last decoder layer with a Bayesian linear regression layer for adaptive training. The modified version of backbone without adaptive training is denoted as " + }, + { + "bbox": [ + 304, + 541, + 545, + 613 + ], + "type": "inline_equation", + "content": "K_{0}" + }, + { + "bbox": [ + 304, + 541, + 545, + 613 + ], + "type": "text", + "content": ", while the full version with adaptive training is denoted as full." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 626, + 359, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 626, + 359, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 626, + 359, + 639 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 646, + 421, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 421, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 421, + 658 + ], + "type": "text", + "content": "5.1.Quantitative results" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "The results of comparing our method with the baselines in various distribution shift scenarios are presented in Tab. 1. We reported three distribution shift scenarios per each time configuration on the table, and other results are included in" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15069" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 68, + 544, + 215 + ], + "blocks": [ + { + "bbox": [ + 52, + 68, + 544, + 215 + ], + "lines": [ + { + "bbox": [ + 52, + 68, + 544, + 215 + ], + "spans": [ + { + "bbox": [ + 52, + 68, + 544, + 215 + ], + "type": "image", + "image_path": "4fd3c636e118474740e034311324639347cb5a334a62e9040f6797039ff237ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 218, + 546, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 218, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 546, + 262 + ], + "type": "text", + "content": "Figure 4. The first row shows prediction before adaptation, and the second row indicates adaptation results by three methods: ours (blue), TENT w/ sup (orange) and MEK (green). Sky blue and orange boxes refer to surrounding actors and actors to be predicted. We depicted only one actor result and one mode among multi-modal predictions closest to the GT for visual simplicity. Please note that our method is multi-modal prediction for all actors method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 74, + 271, + 259, + 393 + ], + "blocks": [ + { + "bbox": [ + 74, + 271, + 259, + 393 + ], + "lines": [ + { + "bbox": [ + 74, + 271, + 259, + 393 + ], + "spans": [ + { + "bbox": [ + 74, + 271, + 259, + 393 + ], + "type": "image", + "image_path": "a52ae0ae537e5b08008adbc411759534ea5d38dd112a24fa1b5c629b8dfbb041.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 397, + 287, + 441 + ], + "lines": [ + { + "bbox": [ + 46, + 397, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 287, + 441 + ], + "type": "text", + "content": "Figure 5. Prediction accuracy and execution time on INTER " + }, + { + "bbox": [ + 46, + 397, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 397, + 287, + 441 + ], + "type": "text", + "content": " nuS (1/3/0.1) experiment. Adjusting update frequency can balance between accuracy and efficiency. Our method significantly outperforms the baseline methods in both accuracy and efficiency." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 453, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 477 + ], + "type": "text", + "content": "the supplementary material. Notably, our approach consistently surpassed baseline performance across all scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 482, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 602 + ], + "type": "text", + "content": "DUA consistently exhibits compromised performance across nearly all cases, due to the distinctive characteristics of trajectory data. In contrast to image data, where data is treated as a singular sample, trajectory data involves multiple agents, each exhibiting distinct motion patterns, within a single data. Consequently, holistically updating batch statistics proves to be counterproductive. Similar challenges are encountered by TENT w/ sup. While regression loss prevents performance decline, updating only the batch norm layer has little to no effect on the prediction performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 604, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 287, + 687 + ], + "type": "text", + "content": "MEK exhibited the most competitive performance among the baselines. While performance improved significantly in short-term settings, it brings limited improvement in long-term cases. As the Kalman filter updates based on the number of prediction steps, short-term configurations with 12 update steps show a better improvement than long-term configurations with only 5 update steps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 45, + 689, + 287, + 713 + ], + "type": "text", + "content": "Although AML led to a considerable improvement in the full version, the predictive performance itself was substan" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 270, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 378 + ], + "type": "text", + "content": "tially degraded due to the significant performance drop in the modified backbone " + }, + { + "bbox": [ + 304, + 270, + 545, + 378 + ], + "type": "inline_equation", + "content": "(K_0)" + }, + { + "bbox": [ + 304, + 270, + 545, + 378 + ], + "type": "text", + "content": ". The limitation of the backbone is due to the Bayesian regression layer being based on probability sampling which is known to be worse than non-probability sampling method of ours [2]. In contrast to all the baseline methods, our method demonstrated state-of-the-art performance in all scenarios featuring various distribution shifts, whether short-term or long-term, showcasing the generalizability of our approach." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 392, + 382, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 382, + 405 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 382, + 405 + ], + "type": "text", + "content": "5.1.1 Efficiency" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 411, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 662 + ], + "type": "text", + "content": "As efficiency is a crucial factor in TTT, we evaluate the frame per second (FPS) along with accuracy " + }, + { + "bbox": [ + 304, + 411, + 545, + 662 + ], + "type": "inline_equation", + "content": "(\\mathrm{mADE}_6)" + }, + { + "bbox": [ + 304, + 411, + 545, + 662 + ], + "type": "text", + "content": ", shown in Fig. 5. We set the performance of the joint training method w/o adaptation as the benchmark and present MEK and TENT, which demonstrates competitive performance among the baselines. Our approach allows for the adjustment of the update frequency, with a frequency of 1 indicates updating at every opportunity, and 2 means updating every other opportunity. While frequent updates improve prediction performance, they also increase execution time; adjusting the update frequency allows for a balance between efficiency and accuracy. As shown in Fig. 5, our method outperforms in both accuracy and efficiency. Given that the time interval is 0.1 seconds, real-time execution requires a processing speed of at least 10 FPS. Even at the maximum update frequency of 1, our method maintains real-time capability with a superior accuracy of 0.39. When increasing the update frequency to 20, the error increases to 0.81, close to MEK's 0.75. However, the FPS reaches 24.2, demonstrating overwhelmingly faster operation compared to MEK's speed of 3.3 FPS." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 671, + 415, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 415, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 415, + 684 + ], + "type": "text", + "content": "5.2. Qualitative results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Comparison to the baselines: We compare our results with TENT and MEK in Fig. 4. While all methods, includ" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15070" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 286, + 225 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 286, + 225 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 286, + 225 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 286, + 225 + ], + "type": "image", + "image_path": "d085ef7d88ac5057af959f00a6e97261c08931baa7e5625dc918e537e76a8e66.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 228, + 288, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 288, + 283 + ], + "type": "text", + "content": "Figure 6. The first row indicates masked samples, and the row below shows the reconstructed outputs. The blue/red arrows indicate historical/future trajectories. The black arrows refer to the masked trajectories. The white lines are the lane centerlines, and the gray dashed lines are the masked lane centerlines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 289, + 286, + 415 + ], + "blocks": [ + { + "bbox": [ + 48, + 289, + 286, + 415 + ], + "lines": [ + { + "bbox": [ + 48, + 289, + 286, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 286, + 415 + ], + "type": "image", + "image_path": "c365f1fd0cd1174f5ab1788d0fbd97f4c6916990322d11398392d02b6c30732a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 417, + 288, + 472 + ], + "lines": [ + { + "bbox": [ + 46, + 417, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 417, + 288, + 472 + ], + "type": "text", + "content": "Figure 7. Multi modal prediction results (blue arrows) before and after adaptation via our method. Ours generates elaborate samples that consider interaction between lane (above) or actors (below) due to representation learning, which cannot be learned from the GT (red arrow) via regression loss." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 581 + ], + "type": "text", + "content": "ing ours, perform multi-agent, multi-modal predictions, we only illustrate one actor and the closest mode to the GT for visual simplicity. The first row of the figure represents predictions before adaptation, and below are the results after adaptation using three different methods. MEK and TENT exhibit instances of underfitting or excessive overfitting upon adaptation, whereas our method consistently demonstrates stable and accurate predictions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "type": "text", + "content": "Reconstruction results: Reconstruction examples are depicted in Fig. 6. For all agents and lanes within a data sample, random masking is applied, as shown in the first row. During test-time training, learning for reconstruction is conducted, resulting in successful reconstruction for data with different distributions, as seen in the second row." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Multi-modal prediction results: As multi-modality is a crucial issue [10, 62, 64], we show that ours can handle multi-modal prediction results in Fig. 7. The adapted predictions showcase diverse yet plausible scenarios, either considering the lane structure (above) or surrounding agents" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 101, + 536, + 252 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 2. Effect of type of losses to be optimized. Optimizing all losses shows optimal test-time adaptation performance." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 101, + 536, + 252 + ], + "lines": [ + { + "bbox": [ + 315, + 101, + 536, + 252 + ], + "spans": [ + { + "bbox": [ + 315, + 101, + 536, + 252 + ], + "type": "table", + "html": "
Exp.Loss typemADE6 /mFDE6
Actor reconLane reconReg
INTER → Lyft (1/3/0.1)✓1.553 / 3.458
✓✓1.054 / 2.007
✓0.842 / 1.512
0.674 / 1.430
✓✓✓0.391 / 0.824
nuS → Lyft (2/6/0.5)✓1.108 / 2.597
0.987 / 2.304
✓✓0.973 / 2.280
✓0.942 / 2.262
✓✓✓0.776 / 1.820
", + "image_path": "6af54dae414e3be7f34aaa98918059a3c222e0d1a9c258ad896fa575d5f84300.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 304, + 541, + 352 + ], + "blocks": [ + { + "bbox": [ + 305, + 258, + 547, + 303 + ], + "lines": [ + { + "bbox": [ + 305, + 258, + 547, + 303 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 547, + 303 + ], + "type": "text", + "content": "Table 3. Ablation on the depth of optimizing layer according to the loss types. The right side of the table indicates the optimization of deeper layers. Using only " + }, + { + "bbox": [ + 305, + 258, + 547, + 303 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 305, + 258, + 547, + 303 + ], + "type": "text", + "content": " deteriorates performance when optimizing all layers while ours stably improves as deeper." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 304, + 541, + 352 + ], + "lines": [ + { + "bbox": [ + 309, + 304, + 541, + 352 + ], + "spans": [ + { + "bbox": [ + 309, + 304, + 541, + 352 + ], + "type": "table", + "html": "
LossOptimizing layers
DD+ED+E+f h,f,l
Lreg0.864 / 2.0720.840 / 2.0930.942 / 2.262
Lreg + Lrecon0.859 / 2.0600.813 / 1.9230.776 / 1.820
", + "image_path": "5e4751f96ac34cbaaf31b457c513cea49af647416f9f3835b0ce88ef427a4444.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 364, + 546, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 364, + 546, + 425 + ], + "spans": [ + { + "bbox": [ + 305, + 364, + 546, + 425 + ], + "type": "text", + "content": "(below). These elaborated samples, although not present in the observed ground truth (GT) future, are learned through the representation learning from reconstruction loss. In addition, it shows that actor-specific tokens do not induce mode collapse to only one motion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 436, + 366, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 436, + 366, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 436, + 366, + 448 + ], + "type": "text", + "content": "6. Ablation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 455, + 444, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 444, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 444, + 468 + ], + "type": "text", + "content": "6.1. Reconstruction objective" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 474, + 546, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 546, + 653 + ], + "type": "text", + "content": "Table. 2 shows ablation studies on optimizing different loss types. Both reconstruction and regression losses individually boost prediction performance, with their joint optimization yielding even greater improvements. Table. 3 compares the effects of using only regression loss versus both losses on prediction performance across different layer depths. Updating just the decoder " + }, + { + "bbox": [ + 304, + 474, + 546, + 653 + ], + "type": "inline_equation", + "content": "(D)" + }, + { + "bbox": [ + 304, + 474, + 546, + 653 + ], + "type": "text", + "content": " shows similar results in both scenarios, but extending updates to the encoder " + }, + { + "bbox": [ + 304, + 474, + 546, + 653 + ], + "type": "inline_equation", + "content": "(E)" + }, + { + "bbox": [ + 304, + 474, + 546, + 653 + ], + "type": "text", + "content": " significantly enhances performance when using both losses. Furthermore, extending updates to the embedding layers " + }, + { + "bbox": [ + 304, + 474, + 546, + 653 + ], + "type": "inline_equation", + "content": "(f_{h,f,l})" + }, + { + "bbox": [ + 304, + 474, + 546, + 653 + ], + "type": "text", + "content": " deteriorates performance when only regression loss is optimized. This highlights the importance of incorporating representation learning through the MAE, as relying solely on regression loss can lead to suboptimal adaptation and damage to learned representations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "We also conduct ablation studies on the masking ratio for both actors and lane centerlines in Tab. 4. The result is visualized via graphs in Fig. 8 according to lane masking ratio and actor masking ratio, respectively. Around 0.3 of lane and 0.4 of actor masking ratio, tendencies of " + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{mADE}_6" + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": " follow" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15071" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 85, + 282, + 169 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 275, + 82 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 275, + 82 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 275, + 82 + ], + "type": "text", + "content": "Table 4. " + }, + { + "bbox": [ + 58, + 71, + 275, + 82 + ], + "type": "inline_equation", + "content": "{\\mathrm{{mADE}}}_{6}" + }, + { + "bbox": [ + 58, + 71, + 275, + 82 + ], + "type": "text", + "content": " according to actor and lane masking ratio." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 85, + 282, + 169 + ], + "lines": [ + { + "bbox": [ + 52, + 85, + 282, + 169 + ], + "spans": [ + { + "bbox": [ + 52, + 85, + 282, + 169 + ], + "type": "table", + "html": "
INTER → Lyft (1/3/0.1)Lane Masking Ratio
0.10.30.50.70.9
Actor0.10.4180.4070.4640.4810.446
0.30.4230.4430.4430.4450.390
Masking Ratio0.50.4170.4550.5150.3910.435
0.70.5670.4350.4480.4530.491
0.90.4470.4210.4940.4170.398
", + "image_path": "af9b92b9a8f8ce965deed41d189599df35e0b81e0728634df127a6c8b1e2f046.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 171, + 165, + 269 + ], + "blocks": [ + { + "bbox": [ + 49, + 171, + 165, + 269 + ], + "lines": [ + { + "bbox": [ + 49, + 171, + 165, + 269 + ], + "spans": [ + { + "bbox": [ + 49, + 171, + 165, + 269 + ], + "type": "image", + "image_path": "4e415c63103f80aa7a0e4edd6a56f39c46758c3de3bcadd677d9ae131bb984c3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 272, + 286, + 294 + ], + "lines": [ + { + "bbox": [ + 47, + 272, + 286, + 294 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 286, + 294 + ], + "type": "text", + "content": "Figure 8. Tendency of " + }, + { + "bbox": [ + 47, + 272, + 286, + 294 + ], + "type": "inline_equation", + "content": "\\mathrm{mADE}_6" + }, + { + "bbox": [ + 47, + 272, + 286, + 294 + ], + "type": "text", + "content": " according to actor and lane masking ratio respectively. (INTER " + }, + { + "bbox": [ + 47, + 272, + 286, + 294 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 47, + 272, + 286, + 294 + ], + "type": "text", + "content": " Lyft " + }, + { + "bbox": [ + 47, + 272, + 286, + 294 + ], + "type": "inline_equation", + "content": "(1/3/0.1)" + }, + { + "bbox": [ + 47, + 272, + 286, + 294 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 172, + 282, + 269 + ], + "blocks": [ + { + "bbox": [ + 167, + 172, + 282, + 269 + ], + "lines": [ + { + "bbox": [ + 167, + 172, + 282, + 269 + ], + "spans": [ + { + "bbox": [ + 167, + 172, + 282, + 269 + ], + "type": "image", + "image_path": "78156583d219c4af1bd5b431da4f43dc19101a43a9b0c1fbef7de4d95e1e9685.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 304, + 286, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 286, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 286, + 437 + ], + "type": "text", + "content": "a U-shape. In case of too small masking ratio, the reconstruction does not learn sufficient representation from the loss, while large masking ratio interrupt interaction learning due to absence of sufficient information. However, in both lane and actor masking, " + }, + { + "bbox": [ + 46, + 304, + 286, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{mADE}_6" + }, + { + "bbox": [ + 46, + 304, + 286, + 437 + ], + "type": "text", + "content": " gets improved when the masking ratio increases above 0.8. In that case, the reconstruction network is induced to learn scene-specific information. In addition, unlike regression loss which deteriorates performance, reconstruction loss does not harm performance because it induces learning the semantic relationship than direct regression supervision." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 444, + 162, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 162, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 162, + 456 + ], + "type": "text", + "content": "6.2. Actor-specific token" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 462, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 713 + ], + "type": "text", + "content": "Table. 5 presents results for the baseline without adaptation, our method without actor-specific tokens, and our full method. The second column reveals that even without actor-specific tokens, the prediction performance is 0.581 and 0.931, surpassing MEK's 0.746 and 1.006. However, incorporating actor-specific tokens for instance-aware adaptation yields notable improvements of " + }, + { + "bbox": [ + 46, + 462, + 287, + 713 + ], + "type": "inline_equation", + "content": "32.7\\%" + }, + { + "bbox": [ + 46, + 462, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 462, + 287, + 713 + ], + "type": "inline_equation", + "content": "16.7\\%" + }, + { + "bbox": [ + 46, + 462, + 287, + 713 + ], + "type": "text", + "content": " for short-term and long-term experiments, respectively. The difference in performance between short-term and long-term is influenced by the scene length in the dataset. The average scene length for short-term data with a 0.1 time interval is 200.04, significantly longer than the average of 32.67 for long-term data with a 0.5 time interval. Intuitively, as scene length increases, the time spent observing previously adapted actors also increases, enhancing the effectiveness of actor-specific tokens. To verify this, Fig. 9 adjusts scene length arbitrarily by skipping to the next scene in data loading once a specific scene length is exceeded. The results confirm that as scene length decreases by skipping scenes earlier, the effectiveness of actor-specific tokens diminishes in both short-term and long-term scenarios." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 311, + 108, + 541, + 174 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 102 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 102 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 102 + ], + "type": "text", + "content": "Table 5. Effect of actor-specific token in " + }, + { + "bbox": [ + 305, + 71, + 545, + 102 + ], + "type": "inline_equation", + "content": "\\mathrm{mADE}_6 / \\mathrm{mFDE}_6" + }, + { + "bbox": [ + 305, + 71, + 545, + 102 + ], + "type": "text", + "content": ". The proposed method enhances adaptation performance by learning actor-wise motion characteristics." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 311, + 108, + 541, + 174 + ], + "lines": [ + { + "bbox": [ + 311, + 108, + 541, + 174 + ], + "spans": [ + { + "bbox": [ + 311, + 108, + 541, + 174 + ], + "type": "table", + "html": "
Exp.BaselineOurs w/o Actor-specificOurs (Full)
INTER → Lyft (1/3/0.1)1.553 / 3.4580.581 / 1.1510.391 / 0.824
nuS → Lyft (2/6/0.5)1.108 / 2.5970.932 / 2.2200.776 / 1.820
", + "image_path": "9bf639bafce7eec560c8b5baa67176543b680f12cf0e6779153eb3e814c2ab71.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 311, + 184, + 543, + 332 + ], + "blocks": [ + { + "bbox": [ + 311, + 184, + 543, + 332 + ], + "lines": [ + { + "bbox": [ + 311, + 184, + 543, + 332 + ], + "spans": [ + { + "bbox": [ + 311, + 184, + 543, + 332 + ], + "type": "image", + "image_path": "44509efe139cf590a75aa806d5674052ec86ef6891c09f3d29290fc0c408aca5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 336, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 403 + ], + "type": "text", + "content": "Figure 9. Effect of scenario length to the effectiveness of actor-specific token. As the scenario length shortens with manual skipping, its effectiveness diminishes because the duration available for the actor-specific token to adapt is reduced. In real-world applications, driving scenarios are continuous, resulting in maximal efficacy of the proposed method." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 414, + 378, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 414, + 378, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 414, + 378, + 426 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 434, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 602 + ], + "type": "text", + "content": "We propose a test-time training method for trajectory prediction by incorporating the MAE and actor-specific token memory. The introduced MAE objective addresses a limitation of conventional online learning, preventing the loss of representations learned from source data. Consequently, our approach enables learning deeper layers, leading to improved representations and enhanced predictions even for out-of-distribution samples. The integration of actor-specific tokens during test-time allows for instance-wise learning of motion patterns, resulting in substantial performance improvements. This approach, particularly effective in continuous real-world autonomous driving scenarios without scene breaks, demonstrates significant efficacy and holds promise for practical applications." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements This research was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF2022R1A2B5B03002636) and the Challengeable Future Defense Technology Research and Development Program through the Agency For Defense Development (ADD) funded by the Defense Acquisition Program Administration (DAPA) in 2024 (No.912768601)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15072" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Görkay Aydemir, Adil Kaan Akan, and Fatma Güney. Adapt: Efficient multi-agent trajectory prediction with adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8295-8305, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "type": "text", + "content": "[2] Inhwan Bae, Jin-Hwi Park, and Hae-Gon Jeon. Nonprobability sampling network for stochastic human trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6477-6487, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "text", + "content": "[3] Inhwan Bae, Jean Oh, and Hae-Gon Jeon. Eigentrajectory: Low-rank descriptors for multi-modal trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10017-10029, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "text", + "content": "[4] Collin Burns and Jacob Steinhardt. Limitations of post-hoc feature alignment for robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2525-2533, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 281, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 288, + 346 + ], + "type": "text", + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "type": "text", + "content": "[6] Rohan Chandra, Uttaran Bhattacharya, Aniket Bera, and Dinesh Manocha. Traphic: Trajectory prediction in dense and heterogeneous traffic using weighted interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8483-8492, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "type": "text", + "content": "[7] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "type": "text", + "content": "[8] Hao Chen, Jiaze Wang, Kun Shao, Furui Liu, Jianye Hao, Chenyong Guan, Guangyong Chen, and Pheng-Ann Heng. Traj-mae: Masked autoencoders for trajectory prediction. arXiv preprint arXiv:2303.06697, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 288, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 288, + 545 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 288, + 545 + ], + "type": "text", + "content": "[9] Liang Chen, Yong Zhang, Yibing Song, Ying Shan, and Lingqiao Liu. Improved test-time adaptation for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24172-24182, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 547, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 602 + ], + "type": "text", + "content": "[10] Weihuang Chen, Zhigang Yang, Lingyang Xue, Jinghai Duan, Hongbin Sun, and Nanning Zheng. Multimodal pedestrian trajectory prediction using probabilistic proposal network. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "text", + "content": "[11] Xuesong Chen, Shaoshuai Shi, Chao Zhang, Benjamin Zhu, Qiang Wang, Ka Chun Cheung, Simon See, and Hongsheng Li. Trajectoryformer: 3d object tracking transformer with predictive trajectory hypotheses. arXiv preprint arXiv:2306.05888, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Yuxiao Chen, Boris Ivanovic, and Marco Pavone. Scept: Scene-consistent, policy-based trajectory predictions for planning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17103-17112, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[13] Jie Cheng, Xiaodong Mei, and Ming Liu. Forecast-mae: Self-supervised pre-training for motion forecasting with masked autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8679-8689, 2023. 2, 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "type": "text", + "content": "[14] Junhyeong Cho, Gilhyun Nam, Sungyeon Kim, Hunmin Yang, and Suha Kwak. Promptstyler: Prompt-driven style generation for source-free domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15702-15712, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 185, + 546, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 546, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 546, + 229 + ], + "type": "text", + "content": "[15] Dooseop Choi and KyoungWook Min. Hierarchical latent structure for multi-modal vehicle trajectory forecasting. In European Conference on Computer Vision, pages 129-145. Springer, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 231, + 546, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 546, + 285 + ], + "type": "text", + "content": "[16] Sehwan Choi, Jungho Kim, Junyong Yun, and Jun Won Choi. R-pred: Two-stage motion prediction via tube-query attention-based trajectory refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8525-8535, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 286, + 547, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 547, + 353 + ], + "type": "text", + "content": "[17] Henggang Cui, Vladan Radosavljevic, Fang-Chieh Chou, Tsung-Han Lin, Thi Nguyen, Tzu-Kuo Huang, Jeff Schneider, and Nemanja Djuric. Multimodal trajectory predictions for autonomous driving using deep convolutional networks. In 2019 International Conference on Robotics and Automation (ICRA), pages 2090-2096. IEEE, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 354, + 546, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 546, + 407 + ], + "type": "text", + "content": "[18] Ning Ding, Yixing Xu, Yehui Tang, Chao Xu, Yunhe Wang, and Dacheng Tao. Source-free domain adaptation via distribution estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7212-7222, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 410, + 546, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 410, + 546, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 410, + 546, + 476 + ], + "type": "text", + "content": "[19] Nemanja Djuric, Vladan Radosavljevic, Henggang Cui, Thi Nguyen, Fang-Chieh Chou, Tsung-Han Lin, Nitin Singh, and Jeff Schneider. Uncertainty-aware short-term motion prediction of traffic actors for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2095-2104, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 477, + 546, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 477, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 477, + 546, + 521 + ], + "type": "text", + "content": "[20] Yonghao Dong, Le Wang, Sanping Zhou, and Gang Hua. Sparse instance conditioned multimodal trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9763-9772, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 522, + 547, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 547, + 589 + ], + "type": "text", + "content": "[21] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 590, + 547, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 547, + 634 + ], + "type": "text", + "content": "[22] Francois Fleuret et al. Uncertainty reduction for model adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9613-9623, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 635, + 546, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 546, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 546, + 678 + ], + "type": "text", + "content": "[23] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. Advances in Neural Information Processing Systems, 35:29374-29385, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 679, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 679, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 679, + 546, + 714 + ], + "type": "text", + "content": "[24] Thomas Gilles, Stefano Sabatini, Dzmitry Tsishkou, Bogdan Stanciulescu, and Fabien Moutarde. THOMAS: Trajectory heatmap output with learned multi-agent sampling. In In-" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15073" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "ternational Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[25] Thomas Gilles, Stefano Sabatini, Dzmitry Tsishkou, Bogdan Stanciulescu, and Fabien Moutarde. Uncertainty estimation for cross-dataset performance in trajectory prediction. arXiv preprint arXiv:2205.07310, 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 205 + ], + "type": "text", + "content": "[26] Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "text", + "content": "[27] Taesik Gong, Jongheon Jeong, Taewon Kim, Yewon Kim, Jinwoo Shin, and Sung-Ju Lee. NOTE: Robust continual test-time adaptation against temporal correlation. In Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "text", + "content": "[28] Junru Gu, Chenxu Hu, Tianyuan Zhang, Xuanyao Chen, Yilun Wang, Yue Wang, and Hang Zhao. Vip3d: End-to-end visual trajectory prediction via 3d agent queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5496-5506, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "type": "text", + "content": "[29] Abner Guzman-Rivera, Dhruv Batra, and Pushmeet Kohli. Multiple choice learning: Learning to produce multiple structured outputs. Advances in neural information processing systems, 25, 2012. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "text", + "content": "[30] John Houston, Guido Zuidhof, Luca Bergamini, Yawei Ye, Long Chen, Ashesh Jain, Sammy Omari, Vladimir Iglovikov, and Peter Ondruska. One thousand and one hours: Self-driving motion prediction dataset. In Conference on Robot Learning, pages 409-418. PMLR, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 464 + ], + "type": "text", + "content": "[31] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "type": "text", + "content": "[32] Pingxuan Huang, Zhenhua Cui, Jing Li, Shenghua Gao, bo Hu, and Yanyan Fang. Cross-domain trajectory prediction with ctp-net, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "text", + "content": "[33] Manh Huynh and Gita Alaghband. AOL: Adaptive online learning for human trajectory prediction in dynamic video scenes. arXiv preprint arXiv:2002.06666, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 588 + ], + "type": "text", + "content": "[34] Manh Huynh and Gita Alaghband. Online adaptive temporal memory with certainty estimation for human trajectory prediction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 940-949, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "text", + "content": "[35] David Isele and Akansel Cosgun. Selective experience replay for lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "text", + "content": "[36] Boris Ivanovic, James Harrison, and Marco Pavone. Expanding the deployment envelope of behavior prediction via adaptive meta-learning. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7786-7793. IEEE, 2023. 1, 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[37] Boris Ivanovic, Guanyu Song, Igor Gilitschenski, and Marco Pavone. trajdata: A unified interface to multiple human trajectory datasets. In Proceedings of the Neural Information" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "Processing Systems (NeurIPS) Track on Datasets and Benchmarks, New Orleans, USA, 2023. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 151 + ], + "type": "text", + "content": "[38] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9644-9653, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "text", + "content": "[39] Ruochen Jiao, Xiangguo Liu, Takami Sato, Qi Alfred Chen, and Qi Zhu. Semi-supervised semantics-guided adversarial training for robust trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8207-8217, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "type": "text", + "content": "[40] Junho Kim, Inwoo Hwang, and Young Min Kim. Ev-tta: Test-time adaptation for event-based object recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17745-17754, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 254, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 318 + ], + "type": "text", + "content": "[41] Mihee Lee, Samuel S Sohn, Seonghyeon Moon, Sejong Yoon, Mubbasir Kapadia, and Vladimir Pavlovic. Musevae: multi-scale vae for environment-aware long term trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2221-2230, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 320, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 375 + ], + "type": "text", + "content": "[42] Lihuan Li, Maurice Pagnucco, and Yang Song. Graph-based spatial transformer with memory replay for multifuture pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2231-2241, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 376, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 430 + ], + "type": "text", + "content": "[43] Maosen Li, Siheng Chen, Yanning Shen, Genjia Liu, Ivor W Tsang, and Ya Zhang. Online multi-agent forecasting with interpretable collaborative graph neural networks. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "type": "text", + "content": "[44] Yingwei Li, Charles R Qi, Yin Zhou, Chenxi Liu, and Dragomir Anguelov. Modar: Using motion forecasting for 3d object detection in point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9329-9339, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 489, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 543 + ], + "type": "text", + "content": "[45] Yushu Li, Xun Xu, Yongyi Su, and Kui Jia. On the robustness of open-world test-time training: Self-training with dynamic prototype expansion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11836-11846, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 544, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 588 + ], + "type": "text", + "content": "[46] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "text", + "content": "[47] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. arXiv preprint arXiv:2303.15361, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "text", + "content": "[48] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecasting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 541-556. Springer, 2020. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[49] Hyesu Lim, Byeonggeun Kim, Jaegul Choo, and Sungha Choi. Ttn: A domain-shift aware batch normalization in test" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15074" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "time adaptation. In The Eleventh International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "text", + "content": "[50] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808-21820, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 217 + ], + "type": "text", + "content": "[51] Yuejiang Liu, Riccardo Cadei, Jonas Schweizer, Sherwin Bahmani, and Alexandre Alahi. Towards robust and adaptive motion forecasting: A causal representation perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17081-17092, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "type": "text", + "content": "[52] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5517-5526, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "type": "text", + "content": "[53] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 397 + ], + "type": "text", + "content": "[54] M Jehanzeb Mirza, Inkyu Shin, Wei Lin, Andreas Schriebl, Kunyang Sun, Jaesung Choe, Mateusz Kozinski, Horst Possegger, In So Kweon, Kuk-Jin Yoon, et al. Mate: Masked autoencoders are online 3d test-time learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16709-16718, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 475 + ], + "type": "text", + "content": "[55] Jiquan Ngiam, Vijay Vasudevan, Benjamin Caine, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, David J Weiss, Benjamin Sapp, Zhifeng Chen, and Jonathon Shlens. Scene transformer: A unified architecture for predicting future trajectories of multiple agents. In International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 478, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 520 + ], + "type": "text", + "content": "[56] Daehee Park, Jaewoo Jeong, and Kuk-Jin Yoon. Improving transferability for cross-domain trajectory prediction via neural stochastic differential equation. arXiv preprint arXiv:2312.15906, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 523, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 576 + ], + "type": "text", + "content": "[57] Daehee Park, Hobin Ryu, Yunseo Yang, Jegyeong Cho, Jiwon Kim, and Kuk-Jin Yoon. Leveraging future relationship reasoning for vehicle trajectory prediction. In The Eleventh International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "text", + "content": "[58] Mozhgan Pourkeshavarz, Changhe Chen, and Amir Rasouli. Learn tarot with mentor: A meta-learned self-supervised approach for trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8384–8393, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[59] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[60] Luke Rowe, Martin Ethier, Eli-Henry Dykhne, and Krzysztof Czarnecki. Fjmp: Factorized joint multi-agent motion prediction over learned directed acyclic interaction" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "text", + "content": "graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13745-13755, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 162 + ], + "type": "text", + "content": "[61] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 164, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 217 + ], + "type": "text", + "content": "[62] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Wei Tang, Nanning Zheng, and Gang Hua. Representing multimodal behaviors with mean location for pedestrian trajectory prediction. IEEE transactions on pattern analysis and machine intelligence, 2023. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "text", + "content": "[63] Liushuai Shi, Le Wang, Sanping Zhou, and Gang Hua. Trajectory unified transformer for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9675-9684, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 265, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 545, + 319 + ], + "type": "text", + "content": "[64] Jianhua Sun, Yuxuan Li, Hao-Shu Fang, and Cewu Lu. Three steps to multimodal trajectory prediction: Modality clustering, classification and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13250-13259, 2021. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 322, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 322, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 322, + 545, + 376 + ], + "type": "text", + "content": "[65] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In Proceedings of the 37th International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 378, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 378, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 308, + 378, + 545, + 432 + ], + "type": "text", + "content": "[66] Devavrat Tomar, Guillaume Vray, Behzad Bozortabar, and Jean-Philippe Thiran. Tesla: Test-time self-learning with automatic adversarial augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20341–20350, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 434, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 498 + ], + "type": "text", + "content": "[67] Li-Wu Tsao, Yan-Kai Wang, Hao-Siang Lin, Hong-Han Shuai, Lai-Kuan Wong, and Wen-Huang Cheng. Socialssl: Self-supervised cross-sequence representation learning based on transformers for multi-agent trajectory prediction. In European Conference on Computer Vision, pages 234-250. Springer, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "text", + "content": "[68] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 546, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 545, + 600 + ], + "type": "text", + "content": "[69] Chunnan Wang, Xiang Chen, Junzhe Wang, and Hongzhi Wang. Atpfl: Automatic trajectory prediction model design under federated learning framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6563-6572, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 602, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 645 + ], + "type": "text", + "content": "[70] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[71] Eason Wang, Henggang Cui, Sai Yalamanchi, Mohana Moorthy, and Nemanja Djuric. Improving movement predictions of traffic actors in bird's-eye view models using gans and differentiable trajectory rasterization. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2340-2348, 2020. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15075" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[72] Jingke Wang, Tengju Ye, Ziqing Gu, and Junbo Chen. Ltp: Lane-based trajectory prediction for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17134-17142, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[73] Letian Wang, Yeping Hu, and Changliu Liu. Online adaptation of neural network models by modified extended kalman filter for customizable and transferable driving behavior prediction, 2022. 1, 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "type": "text", + "content": "[74] Letian Wang, Yeping Hu, Liting Sun, Wei Zhan, Masayoshi Tomizuka, and Changliu Liu. Transferable and adaptable driving behavior prediction. arXiv preprint arXiv:2202.05140, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "type": "text", + "content": "[75] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "text", + "content": "[76] Zhibo Wang, Jiayu Guo, Haiqiang Zhang, Ru Wan, Junping Zhang, and Jian Pu. Bridging the gap: Improving domain generalization in trajectory prediction. IEEE Transactions on Intelligent Vehicles, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 354 + ], + "type": "text", + "content": "[77] Chenfeng Xu, Tian Li, Chen Tang, Lingfeng Sun, Kurt Keutzer, Masayoshi Tomizuka, Alireza Fathi, and Wei Zhan. Pretram: Self-supervised pre-training via connecting trajectory and map. In European Conference on Computer Vision, pages 34-50. Springer, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 421 + ], + "type": "text", + "content": "[78] Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1410-1420, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "text", + "content": "[79] Pei Xu, Jean-Bernard Hayet, and Ioannis Karamouzas. Socialvae: Human trajectory prediction using timewise latents. In European Conference on Computer Vision, pages 511-528. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 288, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 288, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 288, + 511 + ], + "type": "text", + "content": "[80] Yi Xu, Lichen Wang, Yizhou Wang, and Yun Fu. Adaptive trajectory prediction via transferable gnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6520-6531, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "text", + "content": "[81] Yi Xu, Armin Bazarjani, Hyung-gun Chi, Chiho Choi, and Yun Fu. Uncovering the missing pattern: Unified framework towards trajectory imputation and prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9632-9643, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 623 + ], + "type": "text", + "content": "[82] Zhijie Yan, Pengfei Li, Zheng Fu, Shaocong Xu, Yongliang Shi, Xiaoxue Chen, Yuhang Zheng, Yang Li, Tianyu Liu, Chuxuan Li, et al. Int2: Interactive trajectory prediction at intersections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8536-8547, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "text", + "content": "[83] Tao Yang, Shenglong Zhou, Yuwang Wang, Yan Lu, and Nanning Zheng. Test-time batch normalization. arXiv preprint arXiv:2205.10210, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[84] Luyao Ye, Zikang Zhou, and Jianping Wang. Improving the generalizability of trajectory prediction models with freenet-based domain normalization. 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11562-11568, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 461 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[85] Jiangbei Yue, Dinesh Manocha, and He Wang. Human trajectory prediction via neural social physics. In European Conference on Computer Vision, pages 376-394. Springer, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 184 + ], + "type": "text", + "content": "[86] Wei Zhan, Liting Sun, Di Wang, Haojie Shi, Aubrey Clausse, Maximilian Naumann, Julius Kummerle, Hendrik Konigshof, Christoph Stiller, Arnaud de La Fortelle, et al. Interaction dataset: An international, adversarial and cooperative motion dataset in interactive driving scenarios with semantic maps. arXiv preprint arXiv:1910.03088, 2019. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "type": "text", + "content": "[87] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems, 35: 38629-38642, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 229, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 229, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 229, + 545, + 284 + ], + "type": "text", + "content": "[88] Xingxuan Zhang, Peng Cui, Renzhe Xu, Linjun Zhou, Yue He, and Zheyan Shen. Deep stable learning for out-of-distribution generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5372-5382, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 285, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 545, + 329 + ], + "type": "text", + "content": "[89] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863-17873, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 331, + 545, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 406 + ], + "type": "text", + "content": "[90] Dekai Zhu, Guangyao Zhai, Yan Di, Fabian Manhardt, Hendrik Berkemeyer, Tuan Tran, Nassir Navab, Federico Tombari, and Benjamin Busam. Ipcc-tp: Utilizing incremental pearson correlation coefficient for joint multi-agent trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5507-5516, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 407, + 545, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 407, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 308, + 407, + 545, + 461 + ], + "type": "text", + "content": "[91] Yiyao Zhu, Di Luan, and Shaojie Shen. Biff: Bi-level future fusion with polyline-based coordinate for interactive trajectory prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8260-8271, 2023. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15076" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/b3378a53-d4f8-49d5-95a4-af3ae029dc3c_content_list.json b/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/b3378a53-d4f8-49d5-95a4-af3ae029dc3c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1797a937478367560f2d6c90860834d989262789 --- /dev/null +++ b/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/b3378a53-d4f8-49d5-95a4-af3ae029dc3c_content_list.json @@ -0,0 +1,1796 @@ +[ + { + "type": "text", + "text": "TACO: Benchmarking Generalizable Bimanual Tool-ACTION-Object Understanding", + "text_level": 1, + "bbox": [ + 243, + 130, + 727, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yun Liu $^{1,2,3}$ , Haolin Yang $^{4}$ , Xu Si $^{1}$ , Ling Liu $^{5}$ , Zipeng Li $^{1}$ , Yuxiang Zhang $^{1}$ , Yebin Liu $^{1}$ , Li Yi $^{\\dagger,1,2,3}$ $^{1}$ Tsinghua University $^{2}$ Shanghai Artificial Intelligence Laboratory $^{3}$ Shanghai Qi Zhi Institute $^{4}$ Beijing University of Posts and Telecommunications $^{5}$ Beijing Institute of Technology", + "bbox": [ + 86, + 200, + 895, + 262 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Third-person View Video", + "bbox": [ + 210, + 295, + 346, + 304 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a7095185afdc72fa1d218f9b36580ec57d97c848f52a363f324c5e0accba8c0b.jpg", + "image_caption": [ + "Egocentric Video" + ], + "image_footnote": [], + "bbox": [ + 80, + 306, + 444, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hand-Object Mesh", + "bbox": [ + 540, + 300, + 635, + 311 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/70734ec12711cbd502686f0cf04fcbe37fe476736f6157b07689d585b57ba2f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 455, + 316, + 702, + 358 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Segmentation", + "bbox": [ + 470, + 364, + 527, + 373 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Realistic Appearance", + "bbox": [ + 535, + 364, + 627, + 373 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Behavior Triplet", + "bbox": [ + 635, + 364, + 705, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/52d6072e2a8a17c8a867f3635b03989e0bfbf803bd218696b52e03321f768a38.jpg", + "image_caption": [ + "Motion Forecasting" + ], + "image_footnote": [], + "bbox": [ + 454, + 375, + 549, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/44766966b467364b968a1578dcc6d7def426b5a760bf69748433cdc3ab198132.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 376, + 627, + 433 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Test Time Generalization", + "bbox": [ + 745, + 295, + 883, + 304 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2148b55abd1a66d0860a6e434ab73586ad739583895c9b239ea9211b5485c30c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 315, + 812, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2cf661d8a44082cdf5294d66548ebb37f4d402dc4ebbd69e428c2091098fec5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 375, + 812, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3c00496f4791b5691aef890c75cd37ebf498446e9bc38296260cf8f20cc34a12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 313, + 883, + 364 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel Triplet", + "bbox": [ + 826, + 367, + 882, + 376 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc06444b2e7b7b9ec0018992f2cb7b04869f3f549acccc2aac86b90e40eae227.jpg", + "image_caption": [ + "Cooperative Grasp Synthesis" + ], + "image_footnote": [], + "bbox": [ + 823, + 381, + 885, + 435 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9e70c1491c30e0d53f4ed2f11665e522b1a0ca9b8b0533c3e42d074a3e92a835.jpg", + "image_caption": [ + "Action Recognition", + "Figure 1. TACO is a large-scale bimanual hand-object manipulation dataset covering extensive tool-action-object combinations in real-world scenarios. It involves videos from both 12 third-person views and one egocentric view, together with precise hand-object meshes, 2D segmentation, realistic hand-object appearances, and behavior triplet annotations. With rich diversities of object shapes and interaction behaviors, TACO supports test-time generalization to unseen object geometries and novel behavior triplets and benchmarks various generalizable research topics, e.g., action recognition, motion forecasting, and cooperative grasp synthesis." + ], + "image_footnote": [], + "bbox": [ + 94, + 470, + 248, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "brush something", + "bbox": [ + 250, + 474, + 321, + 483 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fe7f90247c143cc264e278f8f08604b69edd2e2f50e2aa125d834e63a7ba8cdb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 503, + 318, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1662622d9486b0169786fcbaa9d3fad799678c820f01fa854b9ee8b35ed7d85d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 483, + 606, + 555 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ae181eba02a4fe7cb58fc6f2bf2d3bbba23318160dbf8e1951a348278547f20f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 478, + 748, + 550 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9de51667526a0f7424c1bc7b0f06612d28265765f6604c1298c1ec96b9af9bee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 476, + 852, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 660, + 313, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans commonly work with multiple objects in daily life and can intuitively transfer manipulation skills to novel objects by understanding object functional regularities. However, existing technical approaches for analyzing and synthesizing hand-object manipulation are mostly limited to handling a single hand and object due to the lack of data support. To address this, we construct TACO, an extensive bimanual hand-object-interaction dataset spanning a large variety of tool-action-object compositions for daily human activities. TACO contains 2.5K motion sequences paired with third-person and egocentric views, precise hand-object 3D meshes, and action labels. To rapidly", + "bbox": [ + 75, + 693, + 472, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "expand the data scale, we present a fully automatic data acquisition pipeline combining multi-view sensing with an optical motion capture system. With the vast research fields provided by TACO, we benchmark three generalizable hand-object-interaction tasks: compositional action recognition, generalizable hand-object motion forecasting, and cooperative grasp synthesis. Extensive experiments reveal new insights, challenges, and opportunities for advancing the studies of generalizable hand-object motion analysis and synthesis. Our data and code are available at https://taco2024.github.io.", + "bbox": [ + 496, + 662, + 893, + 829 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21740", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 89, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our everyday lives, humans effortlessly synchronize the movements of both hands to manipulate a pair of objects, such as using a spatula to stir in a pan while cooking. Usually, these objects are not symmetric in their roles, with one acting as a tool to enhance the bimanual action performed on the other object. This allows us to characterize the bimanual behaviors with tool-action-object triplets. Various triplet combinations usually incorporate intricate and differing temporal and spatial coordination among the tool, the object, and the two hands. Understanding versatile bimanual behaviors could immediately benefit numerous applications in VR/AR, human-robot interaction [8, 44], and dexterous manipulation [7, 49, 74, 78], which pose significant challenges for today's computer vision systems.", + "bbox": [ + 76, + 114, + 467, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tackling the above challenges requires the support of large-scale annotation-rich datasets. Existing dataset efforts [2, 4, 15, 18, 26, 39, 72] on hand-object interactions (HOI) primarily focus on unimanual actions. However, simply aggregating two unimanual actions falls short of encompassing bimanual coordination behaviors. Furthermore, some works [13, 19, 34, 66, 75] examine bimanual manipulation of a single object which can hardly extend to interpreting tool-action-object triplets. The most relevant dataset [33] currently available covers only 12 objects and fewer than 20 distinct tool-action-object triplets, severely limiting its potential to facilitate and benchmark the understanding of bimanual tool-action-object interactions that can generalize to novel object geometries or previously unseen triplets. This limitation partially stems from the challenging nature of jointly capturing two dynamic hands interacting with two objects in the real world, but further underscores the urgency of developing methodologies for generalizable bimanual HOI understanding. Therefore, we put generalization as our primary focus while exploring the curation of a large-scale bimanual HOI dataset that can both facilitate and benchmark generalizable understanding.", + "bbox": [ + 76, + 327, + 467, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In particular, we present TACO, a large-scale bimanual manipulation dataset encompassing a diverse array of tool-action-object compositions in real-world settings. Serving as a knowledge base of multi-object cooperation, TACO focuses on daily scenarios involving the use of tools to interact with target objects and collects 131 types of triplets across 20 object categories, 196 object instances, and 15 daily actions. We organize the dataset based on the triplets and make sure triplets have different levels of overlaps, this naturally defines the semantic distance between different motion trajectories and supports studying generalization with different extents. To expedite the expansion of our dataset, we combine the advantages of marker-based and markerless motion capture systems and present a fully automatic data acquisition pipeline that can guarantee motion quality and", + "bbox": [ + 76, + 659, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "visual data quality at the same time. For each time step, the pipeline automatically provides labels including the precise recovery of hand-object mesh and segmentation on markerless vision data. As a result, TACO comprises a total of 2.5K motion sequences and 5.2M video frames incorporating 3rd-person and egocentric views.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Benefiting from the rich and diverse motion sequences within our dataset, we carefully benchmark three tasks aiming at generalizing common human manipulation knowledge to unseen object geometries and novel tool-action-object combinations: 1) compositional action recognition, 2) generalizable hand-object motion forecasting, and 3) cooperative grasp synthesis. Extensive studies are conducted on these benchmarks: First, diverse hand-object interactive motions with action annotations pave the way to learning to recognize hand-object action under novel object combinations in real-world scenarios. Additionally, the availability of 4D dynamic hand-object pose sequences allows for fine-grained hand-object motion tendency forecasting. Finally, we provide contact-rich hand-object meshes to facilitate studying hand grasp synthesis for unseen object geometries and categories under interaction scenarios. Extensive experiments have been conducted on these benchmarks, revealing generalization challenges faced by existing methodologies when dealing with novel object geometries, object categories, and interaction triplets. We anticipate that TACO will offer more opportunities for researching and developing universal generalization strategies in the understanding and creation of hand-object interactions.", + "bbox": [ + 496, + 183, + 890, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are threefold:", + "bbox": [ + 517, + 530, + 849, + 544 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present TACO, the first large-scale real-world 4D bi-manual hand-object-interaction dataset covering diverse tool-action-object compositions and object geometries.", + "- We develop an automatic data acquisition pipeline that provides precise recovery of hand-object mesh and segmentation together with markerless vision data.", + "- We benchmark three tasks toward generalizable hand-object motion analysis and synthesis. We provide a comprehensive discussion and highlight the new challenges posed by TACO." + ], + "bbox": [ + 500, + 547, + 890, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 714, + 640, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Hand-object Interaction Datasets", + "text_level": 1, + "bbox": [ + 500, + 739, + 790, + 756 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Understanding hand-object interaction is an emerging research topic that is supported by a large number of datasets. Many widely-used video datasets [11, 12, 17, 55] capture 2D real-world hand-object interaction videos to facilitate traditional visual perception studies such as action recognition and HOI detection. With the rapid development of 3D computer vision, recent advances have incorporated 3D hand-object mesh annotations into data-capturing processes and presented various 3D [20, 26] or", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21741", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4D [2, 4, 12, 13, 15, 18, 19, 34, 39, 61, 66, 72, 75] datasets with different focuses. Table 1 briefly summarizes the characteristics of these datasets. A majority of existing datasets focus on single-hand grasping, a basic manipulation skill that can benefit existing dexterous robot learning [7, 48, 49, 78]. Beyond grasping, a line of datasets captures diverse and complex manipulation skills requiring humans to perform complicated, multi-step actions like using tools [13, 33, 39], deforming soft bodies [66], and multi-person handovers [75]. From the perspective of object behaviors, KIT [33], HOI4D [39], and AffordPose[26] capture functional manipulations aiming at showing actual functions and usages of objects. KIT further collects multi-object interaction behaviors, however, it lacks object diversities and is hard to support generalizable studies. Our dataset captures diverse bimanual functional manipulation behaviors for multi-object cooperation, serving a wide range of research directions.", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Generalizable Hand-object Motion Analysis", + "text_level": 1, + "bbox": [ + 76, + 375, + 452, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generalizable hand-object-interaction motion analysis aims to understand the attributes of hand-object motion when manipulating novel object instances. To precisely recognize hand actions during the manipulation of unseen objects, a series of compositional action recognition approaches [21, 31, 43, 50, 51, 59, 69, 76] represent objects through their bounding boxes to prevent the network from overfitting to intricate object geometries. To detect and infer action-object pairs under few-shot or zero-shot settings, a prevailing methodology used in recent generalizable HOI detection methods [22-24, 64, 65, 80] decomposes actions and objects into distinct features. This decomposition allows leveraging abundant training data with either similar actions or similar objects to handle rare action-object pairs that emerge at test time. It also provides the potential to discover novel and reasonable action-object pairs [25]. To estimate hand and object poses without object geometries, Zhu et al. [82] propose learning hand-object interaction priors for each object category and transferring them to unseen objects within the same category during test time, enhancing the method adaptability to diverse unseen object instances.", + "bbox": [ + 75, + 397, + 472, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Generalizable Hand-object Motion Synthesis", + "text_level": 1, + "bbox": [ + 76, + 724, + 457, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthesizing realistic hand-object-interaction motion on novel object geometries and categories is an emerging research topic that includes two main challenges: how to model generic, realistic, and diverse hand motion patterns from existing HOI data, and how to apply them to novel object instances. To address the first challenge, a branch of data-driven approaches decomposes complicated hand-object contact into spatial relations between each hand joint [30, 37, 38, 79] or hand surface point [71, 77, 81] and the object's relevant closest local regions, and such", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "relations can be formulated as a combination of whether contact occurs [37, 38, 71, 77, 79, 81], hand-object distances [37, 71, 77, 81], positions [37, 38, 77, 79, 81] and directions [37, 38, 79] of contact points, and whether contact points should be closer [71]. Another line of work [28-30, 35] encodes hand-object interaction integrally to an implicit global feature by a conditional variational autoencoder (CVAE) [57] structure. As solutions to the second challenge, generalizable methods [28, 30, 35, 38, 79] decompose the object's geometry from HOI and encode it as a condition for generative models. CAMS [79] further maps objects into a canonicalized object space to decrease the gap in object geometry. One limitation of existing approaches is that they do not focus on multi-hand and multi-object cooperation due to the lack of data support.", + "bbox": [ + 496, + 90, + 893, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Constructing TACO", + "text_level": 1, + "bbox": [ + 500, + 330, + 694, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we respectively describe the data capturing system (Section 3.1), the data annotating pipeline (Section 3.2), and dataset statistics (Section 3.3).", + "bbox": [ + 496, + 357, + 893, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Capturing", + "text_level": 1, + "bbox": [ + 500, + 412, + 660, + 429 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b71dd1aa85185fdfa4fcd91289d6f36f3fae586aafb9b0f3617a0f282330aa39.jpg", + "image_caption": [ + "Figure 2. Data capturing system and camera views." + ], + "image_footnote": [], + "bbox": [ + 501, + 449, + 890, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We obtain hand motion from multi-view RGB videos while capturing the object motion by attaching four markers on the object surface and tracking them with a mocap system. To this end, as shown in Figure 2, our data capturing system incorporates 12 synchronized industrial FLIR cameras and a NOKOV optical motion capture suite with 6 infrared Mars4H cameras. To capture egocentric RGBD videos, a helmet equipped with a Realsense L515 camera is worn by the actor. Similar to objects, the motion of the L515 camera is tracked by the mocap system. The frequency of our cameras and mocap system is $30\\mathrm{Hz}$ . The resolutions of our allocentric and egocentric images are $4096\\times 3000$ and $1920\\times 1080$ , respectively.", + "bbox": [ + 496, + 628, + 893, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object model acquisition. To capture contacts between hands and objects and support relevant studies, we obtain an accurate 3D mesh for each object with an industrial EinScan 3D scanner. Each object mesh is represented by up to 100K triangular faces for fine-grained geometries.", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21742", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/c5aa334f42663442cb3b45e11d083a231403f310a4cbc4e65b601b7bbbe425ca.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
datasetbimanualmulti-object cooperationfunctional manipulationdata characteristics: category-levelegocentricmulti-viewmarkerlessmocapaction label# number of: sequenceframe
FHPA [15]✓✓✓1.2K105K
ObMan [20]-154K
ContactPose[2]✓✓2.3K3.0M
GRAB [61]✓✓1.3K1.6M
HO3D [18]✓✓2778K
KIT Binamual Manipulation [33]✓✓✓✓✓✓✓5881.6M
H2O [34]✓✓✓✓191571K
DexYCB [4]✓✓1.0K582K
H2O-3D [19]✓✓✓1776K
H2O for handover [75]✓✓✓6.0K5.0M
OakInk [72]✓✓778314K
HOI4D [39]✓✓✓✓✓4.0K1.2M
HMDO [66]✓✓✓1221K
ARCTIC [13]✓✓✓✓3392.1M
AffordPose [26]✓✓✓-27K
TACO (ours)✓✓✓✓✓✓✓✓✓2.5K5.2M
", + "bbox": [ + 96, + 88, + 870, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Comparison of TACO with existing 3D hand-object interaction datasets.", + "bbox": [ + 243, + 297, + 723, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Data Annotating", + "text_level": 1, + "bbox": [ + 76, + 319, + 243, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3 depicts our automatic data acquisition pipeline. Given color images, 3D marker positions, and object models as inputs, the pipeline successively performs object pose optimization, hand keypoint localization, hand pose optimization, hand-object segmentation, and marker removal.", + "bbox": [ + 76, + 342, + 468, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object pose optimization. Since our objects are rigid bodies, their motions are 6-dimensional and can be tracked by capturing marker positions using the mocap system. To accurately associate markers with the object, we draw inspiration from Mosh++ [41] and optimize marker-to-surface correspondence to encourage contact between markers and object surfaces while avoiding penetration. The object poses are obtained by combining marker positions relative to the object mesh and the captured marker motions.", + "bbox": [ + 75, + 417, + 468, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D hand keypoint localization. Leveraging multi-view sensing, we first estimate 2D hand keypoints from each camera view, respectively, and then fuse them into 3D. For 2D hand keypoint estimation, directly applying existing dual-hand approaches [19, 27, 36] could fail when severe occlusion occurs or two hands are far apart. We thus follow [3, 40] and design a method to separately detect two hands using pre-trained YOLOv3 [52] and then utilize MMPose [9] to obtain single-hand 2D keypoints. To fuse multi-view 2D keypoints to 3D under severe occlusion, we use RANSAC [14] to filter out imprecise 2D keypoints and localize 3D keypoints by triangulation.", + "bbox": [ + 75, + 554, + 468, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contact-aware hand pose optimization. We adopt MANO [53] to formulate a 3D hand mesh as $\\Theta_h = \\{\\theta, \\beta, t\\}$ , where $\\theta \\in \\mathbb{R}^{48}$ , $\\beta \\in \\mathbb{R}^{10}$ , and $t \\in \\mathbb{R}^3$ represent hand pose, hand shape, and wrist position, respectively. Given precomputed 3D hand keypoints and object meshes, we optimize MANO hand meshes by minimizing the following loss function:", + "bbox": [ + 76, + 734, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\Theta} _ {h} = \\underset {\\Theta_ {h}} {\\arg \\min } \\left(\\lambda_ {2 D} \\mathcal {L} _ {2 D} + \\lambda_ {3 D} \\mathcal {L} _ {3 D} + \\lambda_ {a n g l e} \\mathcal {L} _ {a n g l e} + \\right. \\\\ \\left. \\lambda_ {t c} \\mathcal {L} _ {t c} + \\lambda_ {p} \\mathcal {L} _ {p} + \\lambda_ {a} \\mathcal {L} _ {a}\\right), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 851, + 468, + 897 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{2D}$ and $\\mathcal{L}_{3D}$ induce hand meshes to get close to the keypoints, $\\mathcal{L}_{angle}$ provides hand pose priors, $\\mathcal{L}_{tc}$ promotes temporal smoothness, $\\mathcal{L}_a$ encourages contacts, and $\\mathcal{L}_p$ prevents hand-object interpenetration. Details of these loss terms are provided in the supplementary material.", + "bbox": [ + 498, + 320, + 890, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hand-object segmentation. We obtain 2D hand-object masks by first rendering hand-object silhouettes from their 3D meshes. Since MANO meshes differ from real hands, we then apply the Segment Anything Model [32] to refine the masks, using sampled hand-object keypoints as prompts. To help detect hands in subsequent frames, we use Track-anything [70] to track the hands in these frames and localize them.", + "bbox": [ + 496, + 396, + 890, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Marker removal. Despite the fact that attaching markers to objects can accurately capture object motion, markers may increase the appearance gap between captured objects and those in the wild. One direct defect is that an object pose tracking network could simply track markers rather than learn to track the target object. To improve the object's appearance and make it more realistic, after acquiring hand and object poses, we automatically remove the markers from RGB images. First, we obtain marker positions in the world coordinate system from the mocap system, and place spheres with a radius two times the marker at those positions in a simulation environment. We then render a 2D mask of the spheres for each real camera view. Benefiting from the accurate marker position estimates, such 2D masks can fully cover the markers in the original RGB images. Finally, we mask the spheres in the original RGB images and use a pre-trained LAMA [60] network to inpaint the sphere regions. Figure 4 shows a marker removal example.", + "bbox": [ + 496, + 517, + 892, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dataset Statistics", + "text_level": 1, + "bbox": [ + 500, + 801, + 668, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TACO contains 2.5K motion sequences describing extensive daily tool-using behaviors covering 20 object categories, 196 fine-grained object 3D models, 14 participants, and 15 actions. To support cross-category generalizable HOI studies, we explore the functional diversity of tools", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21743", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/96ede0e916f3a4a2ee1b2ab0b6cca3059268fc7f3f6abfb3fba29815e032dca1.jpg", + "image_caption": [ + "Figure 3. Automatic data annotating pipeline. The input consists of color frames from allocentric views, pre-scanned object models, and 3D positions of markers attached to object surfaces. We first separately localize 3D hand keypoints and obtain object poses, and then conduct contact-aware optimization to recover MANO [53] hand meshes. We finally segment hands and objects from images and automatically inpaint markers to acquire realistic object appearances." + ], + "image_footnote": [], + "bbox": [ + 76, + 108, + 893, + 305 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2013a814c6edacde0454d319f62960316980a8faf73df327c951e8b82baef3b7.jpg", + "image_caption": [ + "Figure 4. An example of marker removal. Three sub-figures respectively show the captured image patch, the automatically-computed marker mask, and the inpainted image patch." + ], + "image_footnote": [], + "bbox": [ + 78, + 373, + 207, + 467 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/310ba87860ef953b6eb8801108bd2ed82d4e8c8d43af1b11a8aaed368c386f3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 373, + 336, + 467 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ab0a128c12b6087450c47200f423eb4228ba49c866097d4e73397ce1ad86a2cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 373, + 468, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and target objects, with each tool and target object being utilized to perform up to 7 different kinds of actions. We formulate a behavior as using a tool category to perform an action on a target object category and divide our data into $131 <$ tool category, action label, target object category $\\rightharpoondown$ triplets. Some correlated triplets are exemplified in Figure 5. To enable various studies, TACO presents high-resolution color images from 12 allocentric views and RGBD images from one egocentric-view.", + "bbox": [ + 75, + 530, + 468, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Data Quality Evaluation", + "text_level": 1, + "bbox": [ + 76, + 679, + 307, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative contact optimization evaluation. Figure 6 depicts two frames of the TACO dataset. As shown in the figure, the attraction loss $\\mathcal{L}_a$ promotes contact between the hands and the objects but may exacerbate penetration. The penetration loss $\\mathcal{L}_p$ can mitigate penetration but does not address cases where the hands and the objects are not in contact. Our method combines both attraction loss and penetration loss, striking a balance between encouraging contact and preventing penetration.", + "bbox": [ + 75, + 704, + 468, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative hand pose evaluation. To quantitatively examine the quality and value of our hand poses, we conduct a cross-dataset evaluation between TACO and an existing high-quality HOI dataset DexYCB [4]. We select a", + "bbox": [ + 75, + 840, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d1fa6c2bf2b6a49e434c0775b0b8dda0359609cf8149b7cd2b4c5af01a88fe52.jpg", + "image_caption": [ + "Figure 5. Examples of interaction triplets in TACO. The left, middle, and right columns exemplify categories of tool, action, and target object, respectively. Triplets in our dataset are represented by connected paths from tools to target objects." + ], + "image_footnote": [], + "bbox": [ + 547, + 373, + 841, + 561 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2b711509c1a6053a50a41ff9401d072f8b6db4eab9a02b2f859a874593e6c1fa.jpg", + "image_caption": [ + "Figure 6. Qualitative contact optimization evaluation. From left to right: original RGB image, optimization without attraction loss, optimization without penetration loss, our method with attraction loss and penetration loss." + ], + "image_footnote": [], + "bbox": [ + 501, + 623, + 893, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "multi-stage method, CMR [5], and a lightweight network, MobRecon [6], both of which can recover occluded hand", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21744", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fe0a9b453a282e2f3b49d66090cc7947ccf6a653bf5fdb605216ffa5a13ce459.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTrain TestTACODexYCBTACO + DexYCB
CMRTACO10.73 / 30.8713.53 / 39.5410.25 / 28.36
DexYCB19.36 / 76.206.51 / 13.926.44 / 13.62
MobReconTACO9.03 / 20.7612.95 / 34.328.69 / 20.14
DexYCB15.17 / 44.736.51 / 13.616.32 / 12.94
", + "bbox": [ + 81, + 88, + 467, + 164 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Cross-dataset evaluation with DexYCB [4] on 3D hand pose estimation. Results are in PA-MPJPE (mm, lower is better) and MPJPE (mm, lower is better), respectively.", + "bbox": [ + 76, + 166, + 468, + 207 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "meshes from a single-view RGB image captured in HOI scenarios, and then train and test their networks under different data combinations. For DexYCB, we follow its default train/test split (S0). For TACO, we randomly select 60 motion sequences for training and 40 for testing and crop the right hand from each image to align with the single-hand HO3D. Following existing hand mesh recovery methods [5, 6, 45, 67], we select Procrustes-Aligned Mean Per Joint Position Error (PA-MPJPE) and Mean Per Joint Position Error (MPJPE) as evaluation metrics. More details about the metrics can be found in the paper [6]. Table 2 shows the performance of the two methods under different settings. Due to the large gap in hand-object motion between the two datasets, the performances for the two methods would both drop rapidly when the training and the test sets are from different datasets. Nevertheless, combining training data from TACO and DexYCB achieves significant performance gains on both datasets for both the two methods, which demonstrates the accuracy of our hand pose annotations and indicates that TACO is a beneficial data complement for 3D hand pose estimation.", + "bbox": [ + 75, + 215, + 472, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Marker removal quality evaluation. We conduct a novel experiment to examine the reality of our marker-removed images. The key idea is that a network can easily segment markers from the image while being hard to figure out the same regions when markers are replaced by the original object surface appearances. We thus use the originally captured image patches and the marker-removed ones to train a 2D marker segmentation network, respectively, and then apply the two networks to novel image patches with the same image processing setting as training. In practice, spheres larger than markers (as shown in Figure 4(b)) are required to be segmented, since the marker removal process exactly inpaints these regions. We select U-Net [54] as the network backbone and train the two networks with an L2-loss measuring differences between mask predictions and ground truths. When training and testing on raw image patches, U-Net achieves $63.8\\%$ mean Intersection over Union (mIoU), while its performance sharply drops to $11.1\\%$ mIoU after applying marker removal, indicating the reality and naturalness of the inpainted image regions. Figure 7 further compares heatmaps of network predictions under the two settings, exhibiting ambiguities of marker segmentation posed by realistic inpainted images. More details are provided in the supplementary material.", + "bbox": [ + 75, + 537, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c0d87809ba63d4c78231e9ac99541f59ac0bea7fc55e291032bcd1954412c4b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 89, + 594, + 160 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b1a76b24aa839527c3380d75fb4f3807189169d671783cd7c08832ab47332bcc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 90, + 687, + 160 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/270ddb091364333c9892f5f5c7d893d13bfc6d6cf4b9400f9b84c99f984239ac.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 504, + 161, + 594, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a57fa7f153f47d85c7f8be2cfa1e2ee9e9da703d0e44513408e3fa586dc2a2f5.jpg", + "image_caption": [ + "Figure 7. Marker removal quality evaluation. We train and test U-Net [54] on raw captured images (a) and marker-removed images (b), respectively. In each experiment, the left and right columns show the input image patches and the predicted segmentation heatmaps, respectively." + ], + "image_footnote": [], + "bbox": [ + 596, + 161, + 687, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0a966003a09c3e49348500b8bd9308f95f874d96003eef769769d004027f2117.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 90, + 794, + 160 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/db3f0b10190b5f10903a553e4ea8dfdbafbae201e853c38b2c545c8c182c8f2f.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 705, + 161, + 794, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d6d1e87c980527b470b33011076f8cd6ff8b2924cb43d3acd7aa5f85eaa7e6ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 795, + 90, + 887, + 160 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7bf980a1437b8e43178c59bc83a3dc6b1eba837dd5a9d4126079ad9f105bd050.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 161, + 888, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 338, + 633, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first introduce our data split (Section 5.1) supporting different research purposes about generalization. We then present three benchmarks on TACO: compositional action recognition (Section 5.2), generalizable hand-object motion forecasting (Section 5.3), and cooperative grasp synthesis (Section 5.4). Details of evaluation metrics are provided in supplementary materials.", + "bbox": [ + 496, + 364, + 893, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Data Split", + "text_level": 1, + "bbox": [ + 500, + 482, + 614, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With diverse object geometries and interaction triplets, TACO focuses on supporting generalizable studies that apply to unseen object geometries or novel interaction triplets. Hence, we carefully split our test data into four subsets with different generalization purposes:", + "bbox": [ + 496, + 507, + 890, + 583 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Test set 1 (S1): No generalization. Tool geometries and interaction triplets are both involved in the training set.", + "- Test set 2 (S2): Geometry-level generalization. The tool geometry is novel, while the interaction triplet appears in the training set.", + "- Test set 3 (S3): Interaction-level generalization. The interaction triplet is novel, while the tool categories and geometries are included in the training set.", + "- Test set 4 (S4): Compound generalization. The tool category is novel, leading to unseen geometries and triplets." + ], + "bbox": [ + 500, + 587, + 890, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The training set and the four test sets follow a data ratio of 4:1:1:1.5:2.5.", + "bbox": [ + 496, + 742, + 890, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Compositional Action Recognition", + "text_level": 1, + "bbox": [ + 500, + 785, + 802, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Humans can recognize the action disentangled from their context and appearance biases of the objects, which is referred to as compositionality. To test a system's compositional generalization capability, with a diverse array of tool-action-object triplets in TACO, we benchmark compositional action recognition aiming at identifying actions", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21745", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8f7a3976e44021ae54a3b305b93b9659437f500211c4e74ec385ce92247afe3c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Test SetMethodTop-1 (%, ↑)Top-5 (%, ↑)
S1AIM [73]83.0898.85
CACNF [50]86.1599.62
S2AIM [73]82.8198.83
CACNF [50]77.3496.88
S3AIM [73]53.6581.25
CACNF [50]63.0292.97
S4AIM [73]39.3373.67
CACNF [50]44.0080.50
", + "bbox": [ + 109, + 88, + 429, + 218 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Results on compositional action recognition. Methods are examined via Top-1 and Top-5 accuracy.", + "bbox": [ + 75, + 222, + 468, + 251 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in HOI videos. In contrast to traditional action recognition tasks [11, 16, 56, 58], our focus is on scenarios where the tool-action-object triplets in the test set are unseen during training, including novel tool types or geometries. This poses additional challenges for evaluating model generalization and achieving human-like perception.", + "bbox": [ + 75, + 258, + 467, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Problem formulation: In a bimanual HOI scenario, our goal is to identify action labels from first-person RGB video frames, with optional bounding boxes for hands and objects computed based on their poses.", + "bbox": [ + 76, + 351, + 467, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics: Following [50, 73], we use Top-1 Accuracy and Top-5 Accuracy to evaluate the efficacy of action recognition.", + "bbox": [ + 76, + 411, + 467, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines, results, and analysis: We compare two baselines to evaluate the compositional generalization capability: AIM [73], an adaptation of pre-trained image transformer models, representing state-of-the-art traditional action recognition approaches; and CACNF [50], which specializes in compositional and few-shot action recognition. Table 3 shows the results. A significant decrease in accuracy on the most difficult S4 set, compared to high accuracy on the non-generalizable S1 set, underscores compositional generalization challenges. Notably, the S2 set, with novel tool geometries within tool categories, performed similarly to S1, while the S3 set, featuring novel tool-action-object triplets, showed a significant accuracy drop, revealing that while action recognition models inherently possess basic generalizability to geometric changes, they struggle with novel triplets due to the unfamiliarity of interaction contexts. When comparing among baselines, CACNF outperformed AIM in challenging sets S3 and S4, benefiting from its bounding box-based approach that effectively disentangles actions from object and tool geometries, highlighting the advantages of a compositional approach in generalizable cases. All these results underscore the significant impact of compositional generalization on action recognition.", + "bbox": [ + 75, + 458, + 467, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Generalizable Hand-object Motion Forecasting", + "text_level": 1, + "bbox": [ + 76, + 816, + 467, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With abundant 3D hand-object mesh annotations provided by TACO, we benchmark generalizable interaction forecasting aiming to predict the following hand-object motions from seen short motion clips. Unlike the prediction of", + "bbox": [ + 76, + 839, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "low-frequency human motions [42, 46, 47], we observe that hands commonly achieve a complete manipulation behavior (e.g. pouring all the water out of a bowl) in a very short time, making forecasting both interesting and challenging.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Problem formulation: In a bimanual HOI scenario, given object point clouds and poses of two hands, the tool, and the target object in consecutive $N$ frames, our goal is to forecast their poses in subsequent $M$ frames. We select $N = 10$ and $M = 10$ in our dataset.", + "bbox": [ + 498, + 152, + 890, + 226 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics: Following human-object forecasting evaluations [1, 10, 63, 68], we use Mean Per Joint Position Error $J_{e}$ to evaluate predicted left and right-hand skeletons, and leverage translation error $T_{e}$ and rotation error $R_{e}$ to measure object positions and orientations, respectively.", + "bbox": [ + 498, + 228, + 890, + 304 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines, results, and analysis: Due to the lack of hand-object interaction forecasting solutions, we transfer methodologies from human-object motion forecasting as baselines. To cover diverse method designs and possibilities, we select two generative models [62, 68] and two predictive models [10, 68] for comparison. Table 4 compares their performance under different generalization settings. We observed that the tool and the hand holding it play dominant roles during manipulation, making their motion forecasting more challenging compared to others. The generative models, in comparison to predictive ones, yield significantly larger errors for both hands and the tool. One possible reason is that their motion patterns are fast and complex, making the modeling of motion distribution challenging. Enhancing the accuracy of generative methods in hand-object manipulation scenarios is an interesting future direction. When comparing method performances between a non-generalizable set (S1) and sets involving generalization (S2-4), we note that methods consistently exhibit significant performance declines with the right hand and the tool, regardless of whether applied to novel tool geometries (S2, S4) or interaction triplets (S3, S4), meanwhile achieving close results on the others. This trend indicates substantial challenges in generalizing dominant interaction entities.", + "bbox": [ + 498, + 305, + 890, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Cooperative Grasp Synthesis", + "text_level": 1, + "bbox": [ + 500, + 679, + 756, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generating human-like hand grasps benefits various applications in VR/AR and dexterous manipulation. Existing grasp synthesis approaches [28, 30, 38, 81] mainly focus on creating stable grasps on static objects without considering interactive purposes. Taking advantage of extensive interaction behaviors from TACO, we benchmark a novel task that aims to synthesize realistic and physically plausible hand grasps in HOI scenarios. Besides achieving stable grasping for the object, the method must comprehend other interactive objects and human hands to generate cooperative motions and avoid conflicts. We evaluate this task on our four test sets, each targeting different aspects of generalization.", + "bbox": [ + 498, + 703, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Problem formulation: Consider using the right hand to", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21746", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4cf6911ba3f591d41d9817c2c96139ad832b19f2aae0fa78bbe69eb80f4ed3b0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Test SetMethodJe(mm,↓)Te(mm,↓)Re(°,↓)
Right / LeftTool / TargetTool / Target
S1InterVAE [68]54.9 / 48.955.0 / 15.866.94 / 6.63
MDM [62]61.2 / 53.749.4 / 14.952.81 / 5.85
InterRNN [68]29.3 / 20.425.8 / 11.610.08 / 5.22
CAHMP [10]28.8 / 22.123.9 / 12.310.24 / 4.95
S2InterVAE [68]58.8 / 50.854.3 / 12.288.19 / 5.24
MDM [62]65.8 / 55.848.0 / 10.175.78 / 3.59
InterRNN [68]35.2 / 23.131.0 / 8.911.09 / 4.03
CAHMP [10]31.4 / 22.624.7 / 8.610.65 / 3.32
S3InterVAE [68]56.5 / 50.156.7 / 12.768.03 / 5.69
MDM [62]66.4 / 58.046.2 / 11.561.90 / 4.14
InterRNN [68]32.5 / 22.028.2 / 9.711.42 / 4.32
CAHMP [10]29.3 / 21.624.4 / 9.411.18 / 3.80
S4InterVAE [68]63.9 / 54.263.0 / 14.571.38 / 5.01
MDM [62]70.5 / 57.950.2 / 12.270.25 / 3.55
InterRNN [68]36.6 / 22.330.9 / 10.611.92 / 3.86
CAHMP [10]32.9 / 22.226.6 / 10.811.73 / 3.24
", + "bbox": [ + 81, + 88, + 467, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "manipulate a tool and the left hand to operate a target object. During manipulation, given the meshes of the two objects and the left hand, the goal is to generate a right-hand mesh grasping the tool in a manner conducive to the interaction.", + "bbox": [ + 75, + 400, + 468, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation metrics: Following static grasp synthesis methods [28, 29, 38], we use interpenetration volume (Pen. $V$ ) and contact ratio (Con. $R$ ) to measure the physical plausibility of generated hands. To ensure the hand is appropriately interactive, we introduce a new metric named collision ratio (Col. $R$ ), which detects collisions between generated hands and the interaction environment. Nevertheless, to quantitatively examine whether the grasps are realistic, we design the FID score (FID) to measure distances between distributions of human grasps and synthesized ones.", + "bbox": [ + 75, + 462, + 468, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines, results, and analysis: We select two recent CVAE-based approaches ContactGen [38] and HALO-VAE [30], and modify their network structures to incorporate the left hand and the target object as additional VAE conditions. To assess the significance of interaction environments, we also evaluate HALO-VAE $^{-}$ , which omits the use of environment inputs and directly generates from the tool mesh. Due to the strong correlation between hand grasps and tool geometries, we test these approaches with familiar (S1, S3) and unseen tool geometries (S2, S4). Table 5 summarizes quantitative evaluations of the two settings. Notably, HALO-VAE $^{-}$ shows a significant decrease in performance on contact and collision ratios compared to HALO-VAE, underscoring the importance of perceiving and modeling interaction environments. When applied to novel object geometries and categories, all three approaches exhibit higher collision ratios and FID scores, and most of them show a marked decline in their ability to contact tools precisely and prevent penetrations. Figure 8 exemplifies", + "bbox": [ + 75, + 613, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/41a107f40ca36a4d7260ee967eea277aab84b998f2a0a05d7b42e438741084a7.jpg", + "table_caption": [ + "Table 4. Results on generalizable interaction forecasting. We respectively examine predictions of the right hand, the left hand, the tool, and the target object." + ], + "table_footnote": [], + "table_body": "
Test SetMethodPen. V (cm3, ↓)Con. R (%, ↑)Col. R (%, ↓)FID (10-2, ↓)
S1 ∪ S3ContactGen [38]0.2780.1210.4012.18
HALO-VAE-0.3176.274.194.89
HALO-VAE [30]0.2682.862.392.02
S2 ∪ S4ContactGen [38]0.2674.9817.5622.53
HALO-VAE-0.2971.6914.255.09
HALO-VAE [30]0.3283.5211.112.84
", + "bbox": [ + 501, + 88, + 905, + 203 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Results on cooperative grasp synthesis. We measure the physical plausibility of generated grasps by interpenetration volume and contact ratio, meanwhile using collision ratio to examine interaction faithfulness and FID score to assess reality.", + "bbox": [ + 498, + 208, + 890, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "failure cases of synthesized grasps, indicating a large room for improving physical feasibility and naturalness on complex and slender geometries.", + "bbox": [ + 498, + 272, + 890, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b3a1e009f4eed4b73751e651a84d3e8774212406a848a3919311a966c77c4946.jpg", + "image_caption": [ + "Figure 8. Failure cases of cooperative grasp synthesis. (a) and (b) are generated by ContactGen [38], while (c) and (d) are from HALO-VAE [30]." + ], + "image_footnote": [], + "bbox": [ + 501, + 329, + 890, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations and Conclusion", + "text_level": 1, + "bbox": [ + 500, + 542, + 756, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present TACO, the first large-scale, real-world 4D bi-manual hand-object manipulation dataset. It encompasses a wide range of tool-action-object compositions and object geometries. TACO contains a total of $2.5\\mathrm{K}$ motion sequences and $5.2\\mathrm{M}$ video frames, captured from 12 third-person views and one egocentric view. We contribute an automatic data acquisition pipeline that accurately recovers hand-object meshes and segmentation, along with realistic hand-object appearances. Leveraging TACO's diverse data, we benchmark compositional action recognition, generalizable hand-object motion forecasting, and cooperative grasp synthesis. These benchmarks reveal new insights, challenges, and opportunities in the field of generalizable hand-object interaction studies.", + "bbox": [ + 496, + 568, + 890, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "There are three major limitations in TACO. Firstly, TACO currently does not cover articulated objects. Secondly, while TACO offers an extensive exploration of object geometries and HOI behaviors, it lacks scene diversities that are also crucial for understanding human manipulations. Thirdly, our solution to marker removal is an application of generative models, hence the original object appearances cannot be recovered perfectly.", + "bbox": [ + 496, + 779, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21747", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Vida Adeli, Mahsa Ehsanpour, Ian Reid, Juan Carlos Niebles, Silvio Savarese, Ehsan Adeli, and Hamid Rezatofighi. Tripod: Human trajectory and pose dynamics forecasting in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13390-13400, 2021. 7", + "[2] Samarth Brahmbhatt, Chengcheng Tang, Christopher D Twigg, Charles C Kemp, and James Hays. Contactpose: A dataset of grasps with object contact and hand pose. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 361-378. Springer, 2020. 2, 3, 4", + "[3] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 4", + "[4] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 2, 3, 4, 5, 6", + "[5] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5, 6", + "[6] Xingyu Chen, Yufeng Liu, Dong Yajiao, Xiong Zhang, Chongyang Ma, Yanmin Xiong, Yuan Zhang, and Xiaoyan Guo. Mobrecon: Mobile-friendly hand mesh reconstruction from monocular image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5, 6", + "[7] Sammy Christen, Muhammed Kocabas, Emre Aksan, Jemin Hwangbo, Jie Song, and Otmar Hilliges. D-grasp: Physically plausible dynamic grasp synthesis for hand-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20577-20586, 2022. 2, 3", + "[8] Sammy Christen, Wei Yang, Claudia Pérez-D'Arpino, Otmar Hilliges, Dieter Fox, and Yu-Wei Chao. Learning human-to-robot handovers from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9654–9664, 2023. 2", + "[9] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.4", + "[10] Enric Corona, Albert Pumarola, Guillem Alenya, and Francesc Moreno-Noguer. Context-aware human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6992-7001, 2020. 7, 8", + "[11] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European conference on computer vision (ECCV), pages 720-736, 2018. 2, 7", + "[12] Christian RG Dreher, Mirko Wächter, and Tamim Asfour. Learning object-action relations from bimanual human demonstration using graph networks. IEEE Robotics and Automation Letters, 5(1):187-194, 2019. 2, 3", + "[13] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J Black, and Otmar Hilliges. Arctic: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12954, 2023. 2, 3, 4", + "[14] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 4", + "[15] Guillermo Garcia-Hernando, Shanxin Yuan, Seungryul Baek, and Tae-Kyun Kim. First-person hand action benchmark with rgb-d videos and 3d hand pose annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 409-419, 2018. 2, 3, 4", + "[16] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 7", + "[17] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2", + "[18] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3196-3206, 2020. 2, 3, 4", + "[19] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11090-11100, 2022. 2, 3, 4", + "[20] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11807-11816, 2019. 2, 4", + "[21] Roei Herzig, Elad Ben-Avraham, Karttikeya Mangalam, Amir Bar, Gal Chechik, Anna Rohrbach, Trevor Darrell, and Amir Globerson. Object-region video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3148-3159, 2022. 3" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21748", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Zhi Hou, Xiaojiang Peng, Yu Qiao, and Dacheng Tao. Visual compositional learning for human-object interaction detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 584-600. Springer, 2020. 3", + "[23] Zhi Hou, Baosheng Yu, Yu Qiao, Xiaojiang Peng, and Dacheng Tao. Affordance transfer learning for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 495-504, 2021.", + "[24] Zhi Hou, Baosheng Yu, Yu Qiao, Xiaojiang Peng, and Dacheng Tao. Detecting human-object interaction via fabricated compositional learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14646-14655, 2021. 3", + "[25] Zhi Hou, Baosheng Yu, and Dacheng Tao. Discovering human-object interaction concepts via self-compositional learning. In European Conference on Computer Vision, pages 461-478. Springer, 2022. 3", + "[26] Juntao Jian, Xiuping Liu, Manyi Li, Ruizhen Hu, and Jian Liu. Affordpose: A large-scale dataset of hand-object interactions with affordance-driven hand pose. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14713-14724, 2023. 2, 3, 4", + "[27] Changlong Jiang, Yang Xiao, Cunlin Wu, Mingyang Zhang, Jinghong Zheng, Zhiguo Cao, and Joey Tianyi Zhou. A2j-transformer: Anchor-to-joint transformer network for 3d interacting hand pose estimation from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8846-8855, 2023. 4", + "[28] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J Black, Krikamol Muandet, and Siyu Tang. Grasping field: Learning implicit representations for human grasps. In 2020 International Conference on 3D Vision (3DV), pages 333-344. IEEE, 2020. 3, 7, 8", + "[29] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J Black, Krikamol Muandet, and Siyu Tang. Grasping field: Learning implicit representations for human grasps. In 2020 International Conference on 3D Vision (3DV), pages 333-344. IEEE, 2020. 8", + "[30] Korrawe Karunratanakul, Adrian Spurr, Zicong Fan, Otmar Hilliges, and Siyu Tang. A skeleton-driven neural occupancy representation for articulated hands. In 2021 International Conference on 3D Vision (3DV), pages 11-21. IEEE, 2021. 3, 7, 8", + "[31] Tae Soo Kim, Jonathan Jones, and Gregory D Hager. Motion guided attention fusion to recognize interactions from videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13076-13086, 2021. 3", + "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 4", + "[33] Franziska Krebs, Andre Meixner, Isabel Patzer, and Tamim Asfour. The kit bimanual manipulation dataset. In 2020 IEEE-RAS 20th International Conference on Humanoid Robots (Humanoids), pages 499-506. IEEE, 2021. 2, 3, 4" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 2, 3, 4", + "[35] Haoming Li, Xinzhuo Lin, Yang Zhou, Xiang Li, Yuchi Huo, Jiming Chen, and Qi Ye. Contact2grasp: 3d grasp synthesis via hand-object contact constraint. arXiv preprint arXiv:2210.09245, 2022. 3", + "[36] Mengcheng Li, Liang An, Hongwen Zhang, Lianpeng Wu, Feng Chen, Tao Yu, and Yebin Liu. Interacting attention graph for single image two-hand reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2761-2770, 2022. 4", + "[37] Qingtao Liu, Yu Cui, Zhengnan Sun, Haoming Li, Gaofeng Li, Lin Shao, Jiming Chen, and Qi Ye. Dexrepnet: Learning dexterous robotic grasping network with geometric and spatial hand-object representations. arXiv preprint arXiv:2303.09806, 2023. 3", + "[38] Shaowei Liu, Yang Zhou, Jimei Yang, Saurabh Gupta, and Shenlong Wang. Contactgen: Generative contact modeling for grasp generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20609-20620, 2023. 3, 7, 8", + "[39] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 2, 3, 4", + "[40] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 4", + "[41] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 4", + "[42] Wei Mao, Richard I Hartley, Mathieu Salzmann, et al. Contact-aware human motion forecasting. Advances in Neural Information Processing Systems, 35:7356-7367, 2022. 7", + "[43] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-else: Compositional action recognition with spatial-temporal interaction networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1049–1059, 2020. 3", + "[44] Eley Ng, Ziang Liu, and Monroe Kennedy. It takes two: Learning to plan for human-robot cooperative carrying. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7526–7532. IEEE, 2023. 2", + "[45] JoonKyu Park, Yeonguk Oh, Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Handocctnet: Occlusion-robust 3d hand mesh estimation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1496–1505, 2022. 6" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21749", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Xiaogang Peng, Yaodi Shen, Haoran Wang, Binling Nie, Yiqiang Wang, and Zizhao Wu. Somoformer: Social-aware motion transformer for multi-person motion prediction. arXiv preprint arXiv:2208.09224, 2022.7", + "[47] Xiaogang Peng, Siyuan Mao, and Zizhao Wu. Trajectory-aware body interaction transformer for multi-person pose forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17121-17130, 2023. 7", + "[48] Yuzhe Qin, Hao Su, and Xiaolong Wang. From one hand to multiple hands: Imitation learning for dexterous manipulation from single-camera teleoperation. IEEE Robotics and Automation Letters, 7(4):10873-10881, 2022. 3", + "[49] Yuzhe Qin, Yueh-Hua Wu, Shaowei Liu, Hanwen Jiang, Ruihan Yang, Yang Fu, and Xiaolong Wang. Dexamv: Imitation learning for dexterous manipulation from human videos. In European Conference on Computer Vision, pages 570-587. Springer, 2022. 2, 3", + "[50] Gorjan Radevski, Marie-Francine Moens, and Tinne Tuyte-laars. Revisiting spatio-temporal layouts for compositional action recognition. arXiv preprint arXiv:2111.01936, 2021. 3, 7", + "[51] Ramanathan Rajendiran, Debaditya Roy, and Basura Fernando. Modelling spatio-temporal interactions for compositional action recognition. arXiv preprint arXiv:2305.02673, 2023. 3", + "[52] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 4", + "[53] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 4, 5", + "[54] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 6", + "[55] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 2", + "[56] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. 7", + "[57] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 3", + "[58] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 7", + "[59] Pengzhan Sun, Bo Wu, Xunsong Li, Wen Li, Lixin Duan, and Chuang Gan. Counterfactual debiasing inference for compositional action recognition. In Proceedings of the 29th" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ACM International Conference on Multimedia, pages 3220-3228, 2021. 3", + "[60] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021.4", + "[61] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 581-600. Springer, 2020. 3, 4", + "[62] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 7, 8", + "[63] Weilin Wan, Lei Yang, Lingjie Liu, Zhuoying Zhang, Ruixing Jia, Yi-King Choi, Jia Pan, Christian Theobalt, Taku Komura, and Wenping Wang. Learn to predict how humans manipulate large-sized objects from interactive motions. IEEE Robotics and Automation Letters, 7(2):4702-4709, 2022. 7", + "[64] Haozhong Wang, Hua Yu, and Qiang Zhang. Detecting zero-shot human-object interaction with visual-text modeling. In 2023 9th International Conference on Virtual Reality (ICVR), pages 155-162. IEEE, 2023. 3", + "[65] Suchen Wang, Yueqi Duan, Henghui Ding, Yap-Peng Tan, Kim-Hui Yap, and Junsong Yuan. Learning transferable human-object interaction detector with natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 939-948, 2022. 3", + "[66] Wei Xie, Zhipeng Yu, Zimeng Zhao, Binghui Zuo, and Yanggang Wang. Hmdo: Markerless multi-view hand manipulation capture with deformable objects. Graphical Models, 127:101178, 2023. 2, 3, 4", + "[67] Hao Xu, Tianyu Wang, Xiao Tang, and Chi-Wing Fu. H2onet: Hand-occlusion-and-orientation-aware network for real-time 3d hand mesh reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17048–17058, 2023. 6", + "[68] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14928-14940, 2023. 7, 8", + "[69] Rui Yan, Peng Huang, Xiangbo Shu, Junhao Zhang, Yonghua Pan, and Jinhui Tang. Look less think more: Rethinking compositional action recognition. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3666-3675, 2022. 3", + "[70] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 4", + "[71] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21750", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF International Conference on Computer Vision, pages 11097-11106, 2021.3", + "[72] Lixin Yang, Kailin Li, Xinyu Zhan, Fei Wu, Anran Xu, Liu Liu, and Cewu Lu. Oakink: A large-scale knowledge repository for understanding hand-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20953-20962, 2022. 2, 3, 4", + "[73] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. AIM: Adapting image models for efficient video action recognition. In The Eleventh International Conference on Learning Representations, 2023. 7", + "[74] Jianglong Ye, Jiashun Wang, Binghao Huang, Yuzhe Qin, and Xiaolong Wang. Learning continuous grasping function with a dexterous hand from human demonstrations. IEEE Robotics and Automation Letters, 8(5):2882-2889, 2023. 2", + "[75] Ruolin Ye, Wenqiang Xu, Zhendong Xue, Tutian Tang, Yanfeng Wang, and Cewu Lu. H2o: A benchmark for visual human-human object handover analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15762-15771, 2021. 2, 3, 4", + "[76] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Is an object-centric video representation beneficial for transfer? In Proceedings of the Asian Conference on Computer Vision, pages 1976-1994, 2022. 3", + "[77] He Zhang, Yuting Ye, Takaaki Shiratori, and Taku Komura. Manipnet: neural manipulation synthesis with a handobject spatial representation. ACM Transactions on Graphics (ToG), 40(4):1-14, 2021. 3", + "[78] Hui Zhang, Sammy Christen, Zicong Fan, Luocheng Zheng, Jemin Hwangbo, Jie Song, and Otmar Hilliges. Artigrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. arXiv preprint arXiv:2309.03891, 2023. 2, 3", + "[79] Juntian Zheng, Qingyuan Zheng, Lixing Fang, Yun Liu, and Li Yi. Cams: Canonicalized manipulation spaces for category-level functional hand-object manipulation synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 585-594, 2023. 3", + "[80] Desen Zhou, Zhichao Liu, Jian Wang, Leshan Wang, Tao Hu, Errui Ding, and Jingdong Wang. Human-object interaction detection via disentangled transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19568-19577, 2022. 3", + "[81] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Toch: Spatio-temporal object-to-hand correspondence for motion refinement. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3, 7", + "[82] Zehao Zhu, Jiashun Wang, Yuzhe Qin, Deqing Sun, Varun Jampani, and Xiaolong Wang. Contactart: Learning 3d interaction priors for category-level articulated object and hand poses estimation. arXiv preprint arXiv:2305.01618, 2023. 3" + ], + "bbox": [ + 78, + 90, + 468, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "21751", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/b3378a53-d4f8-49d5-95a4-af3ae029dc3c_model.json b/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/b3378a53-d4f8-49d5-95a4-af3ae029dc3c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7573680ab3a02c94b46eb1ed29223e0789a3b935 --- /dev/null +++ b/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/b3378a53-d4f8-49d5-95a4-af3ae029dc3c_model.json @@ -0,0 +1,2765 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.245, + 0.131, + 0.728, + 0.177 + ], + "angle": 0, + "content": "TACO: Benchmarking Generalizable Bimanual Tool-ACTION-Object Understanding" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.202, + 0.897, + 0.263 + ], + "angle": 0, + "content": "Yun Liu\\(^{1,2,3}\\), Haolin Yang\\(^{4}\\), Xu Si\\(^{1}\\), Ling Liu\\(^{5}\\), Zipeng Li\\(^{1}\\), Yuxiang Zhang\\(^{1}\\), Yebin Liu\\(^{1}\\), Li Yi\\(^{\\dagger,1,2,3}\\) \n\\(^{1}\\)Tsinghua University \\(^{2}\\)Shanghai Artificial Intelligence Laboratory \\(^{3}\\)Shanghai Qi Zhi Institute \\(^{4}\\)Beijing University of Posts and Telecommunications \\(^{5}\\)Beijing Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.296, + 0.347, + 0.305 + ], + "angle": 0, + "content": "Third-person View Video" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.307, + 0.445, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.428, + 0.305, + 0.438 + ], + "angle": 0, + "content": "Egocentric Video" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.301, + 0.636, + 0.312 + ], + "angle": 0, + "content": "Hand-Object Mesh" + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.318, + 0.703, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.471, + 0.366, + 0.528, + 0.374 + ], + "angle": 0, + "content": "Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.366, + 0.628, + 0.374 + ], + "angle": 0, + "content": "Realistic Appearance" + }, + { + "type": "text", + "bbox": [ + 0.636, + 0.366, + 0.706, + 0.374 + ], + "angle": 0, + "content": "Behavior Triplet" + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.376, + 0.55, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.377, + 0.629, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.746, + 0.296, + 0.885, + 0.305 + ], + "angle": 0, + "content": "Test Time Generalization" + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.316, + 0.813, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.375, + 0.813, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.825, + 0.314, + 0.885, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.828, + 0.368, + 0.883, + 0.377 + ], + "angle": 0, + "content": "Novel Triplet" + }, + { + "type": "image", + "bbox": [ + 0.824, + 0.382, + 0.887, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.442, + 0.282, + 0.455 + ], + "angle": 0, + "content": "Action Recognition" + }, + { + "type": "image_caption", + "bbox": [ + 0.42, + 0.442, + 0.56, + 0.456 + ], + "angle": 0, + "content": "Motion Forecasting" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.472, + 0.249, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.476, + 0.323, + 0.484 + ], + "angle": 0, + "content": "brush something" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.505, + 0.319, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.484, + 0.607, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.442, + 0.86, + 0.456 + ], + "angle": 0, + "content": "Cooperative Grasp Synthesis" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.479, + 0.749, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.477, + 0.854, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.579, + 0.895, + 0.651 + ], + "angle": 0, + "content": "Figure 1. TACO is a large-scale bimanual hand-object manipulation dataset covering extensive tool-action-object combinations in real-world scenarios. It involves videos from both 12 third-person views and one egocentric view, together with precise hand-object meshes, 2D segmentation, realistic hand-object appearances, and behavior triplet annotations. With rich diversities of object shapes and interaction behaviors, TACO supports test-time generalization to unseen object geometries and novel behavior triplets and benchmarks various generalizable research topics, e.g., action recognition, motion forecasting, and cooperative grasp synthesis." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.661, + 0.314, + 0.677 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.694, + 0.473, + 0.876 + ], + "angle": 0, + "content": "Humans commonly work with multiple objects in daily life and can intuitively transfer manipulation skills to novel objects by understanding object functional regularities. However, existing technical approaches for analyzing and synthesizing hand-object manipulation are mostly limited to handling a single hand and object due to the lack of data support. To address this, we construct TACO, an extensive bimanual hand-object-interaction dataset spanning a large variety of tool-action-object compositions for daily human activities. TACO contains 2.5K motion sequences paired with third-person and egocentric views, precise hand-object 3D meshes, and action labels. To rapidly" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.895, + 0.83 + ], + "angle": 0, + "content": "expand the data scale, we present a fully automatic data acquisition pipeline combining multi-view sensing with an optical motion capture system. With the vast research fields provided by TACO, we benchmark three generalizable hand-object-interaction tasks: compositional action recognition, generalizable hand-object motion forecasting, and cooperative grasp synthesis. Extensive experiments reveal new insights, challenges, and opportunities for advancing the studies of generalizable hand-object motion analysis and synthesis. Our data and code are available at https://taco2024.github.io." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21740" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.208, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.116, + 0.468, + 0.326 + ], + "angle": 0, + "content": "In our everyday lives, humans effortlessly synchronize the movements of both hands to manipulate a pair of objects, such as using a spatula to stir in a pan while cooking. Usually, these objects are not symmetric in their roles, with one acting as a tool to enhance the bimanual action performed on the other object. This allows us to characterize the bimanual behaviors with tool-action-object triplets. Various triplet combinations usually incorporate intricate and differing temporal and spatial coordination among the tool, the object, and the two hands. Understanding versatile bimanual behaviors could immediately benefit numerous applications in VR/AR, human-robot interaction [8, 44], and dexterous manipulation [7, 49, 74, 78], which pose significant challenges for today's computer vision systems." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.328, + 0.468, + 0.658 + ], + "angle": 0, + "content": "Tackling the above challenges requires the support of large-scale annotation-rich datasets. Existing dataset efforts [2, 4, 15, 18, 26, 39, 72] on hand-object interactions (HOI) primarily focus on unimanual actions. However, simply aggregating two unimanual actions falls short of encompassing bimanual coordination behaviors. Furthermore, some works [13, 19, 34, 66, 75] examine bimanual manipulation of a single object which can hardly extend to interpreting tool-action-object triplets. The most relevant dataset [33] currently available covers only 12 objects and fewer than 20 distinct tool-action-object triplets, severely limiting its potential to facilitate and benchmark the understanding of bimanual tool-action-object interactions that can generalize to novel object geometries or previously unseen triplets. This limitation partially stems from the challenging nature of jointly capturing two dynamic hands interacting with two objects in the real world, but further underscores the urgency of developing methodologies for generalizable bimanual HOI understanding. Therefore, we put generalization as our primary focus while exploring the curation of a large-scale bimanual HOI dataset that can both facilitate and benchmark generalizable understanding." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.66, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In particular, we present TACO, a large-scale bimanual manipulation dataset encompassing a diverse array of tool-action-object compositions in real-world settings. Serving as a knowledge base of multi-object cooperation, TACO focuses on daily scenarios involving the use of tools to interact with target objects and collects 131 types of triplets across 20 object categories, 196 object instances, and 15 daily actions. We organize the dataset based on the triplets and make sure triplets have different levels of overlaps, this naturally defines the semantic distance between different motion trajectories and supports studying generalization with different extents. To expedite the expansion of our dataset, we combine the advantages of marker-based and markerless motion capture systems and present a fully automatic data acquisition pipeline that can guarantee motion quality and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "visual data quality at the same time. For each time step, the pipeline automatically provides labels including the precise recovery of hand-object mesh and segmentation on markerless vision data. As a result, TACO comprises a total of 2.5K motion sequences and 5.2M video frames incorporating 3rd-person and egocentric views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.892, + 0.529 + ], + "angle": 0, + "content": "Benefiting from the rich and diverse motion sequences within our dataset, we carefully benchmark three tasks aiming at generalizing common human manipulation knowledge to unseen object geometries and novel tool-action-object combinations: 1) compositional action recognition, 2) generalizable hand-object motion forecasting, and 3) cooperative grasp synthesis. Extensive studies are conducted on these benchmarks: First, diverse hand-object interactive motions with action annotations pave the way to learning to recognize hand-object action under novel object combinations in real-world scenarios. Additionally, the availability of 4D dynamic hand-object pose sequences allows for fine-grained hand-object motion tendency forecasting. Finally, we provide contact-rich hand-object meshes to facilitate studying hand grasp synthesis for unseen object geometries and categories under interaction scenarios. Extensive experiments have been conducted on these benchmarks, revealing generalization challenges faced by existing methodologies when dealing with novel object geometries, object categories, and interaction triplets. We anticipate that TACO will offer more opportunities for researching and developing universal generalization strategies in the understanding and creation of hand-object interactions." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.531, + 0.851, + 0.545 + ], + "angle": 0, + "content": "In summary, our main contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.549, + 0.891, + 0.594 + ], + "angle": 0, + "content": "- We present TACO, the first large-scale real-world 4D bi-manual hand-object-interaction dataset covering diverse tool-action-object compositions and object geometries." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.595, + 0.891, + 0.639 + ], + "angle": 0, + "content": "- We develop an automatic data acquisition pipeline that provides precise recovery of hand-object mesh and segmentation together with markerless vision data." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.64, + 0.891, + 0.701 + ], + "angle": 0, + "content": "- We benchmark three tasks toward generalizable hand-object motion analysis and synthesis. We provide a comprehensive discussion and highlight the new challenges posed by TACO." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.549, + 0.891, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.715, + 0.641, + 0.731 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.791, + 0.757 + ], + "angle": 0, + "content": "2.1. Hand-object Interaction Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Understanding hand-object interaction is an emerging research topic that is supported by a large number of datasets. Many widely-used video datasets [11, 12, 17, 55] capture 2D real-world hand-object interaction videos to facilitate traditional visual perception studies such as action recognition and HOI detection. With the rapid development of 3D computer vision, recent advances have incorporated 3D hand-object mesh annotations into data-capturing processes and presented various 3D [20, 26] or" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21741" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.364 + ], + "angle": 0, + "content": "4D [2, 4, 12, 13, 15, 18, 19, 34, 39, 61, 66, 72, 75] datasets with different focuses. Table 1 briefly summarizes the characteristics of these datasets. A majority of existing datasets focus on single-hand grasping, a basic manipulation skill that can benefit existing dexterous robot learning [7, 48, 49, 78]. Beyond grasping, a line of datasets captures diverse and complex manipulation skills requiring humans to perform complicated, multi-step actions like using tools [13, 33, 39], deforming soft bodies [66], and multi-person handovers [75]. From the perspective of object behaviors, KIT [33], HOI4D [39], and AffordPose[26] capture functional manipulations aiming at showing actual functions and usages of objects. KIT further collects multi-object interaction behaviors, however, it lacks object diversities and is hard to support generalizable studies. Our dataset captures diverse bimanual functional manipulation behaviors for multi-object cooperation, serving a wide range of research directions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.375, + 0.453, + 0.391 + ], + "angle": 0, + "content": "2.2. Generalizable Hand-object Motion Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.473, + 0.716 + ], + "angle": 0, + "content": "Generalizable hand-object-interaction motion analysis aims to understand the attributes of hand-object motion when manipulating novel object instances. To precisely recognize hand actions during the manipulation of unseen objects, a series of compositional action recognition approaches [21, 31, 43, 50, 51, 59, 69, 76] represent objects through their bounding boxes to prevent the network from overfitting to intricate object geometries. To detect and infer action-object pairs under few-shot or zero-shot settings, a prevailing methodology used in recent generalizable HOI detection methods [22-24, 64, 65, 80] decomposes actions and objects into distinct features. This decomposition allows leveraging abundant training data with either similar actions or similar objects to handle rare action-object pairs that emerge at test time. It also provides the potential to discover novel and reasonable action-object pairs [25]. To estimate hand and object poses without object geometries, Zhu et al. [82] propose learning hand-object interaction priors for each object category and transferring them to unseen objects within the same category during test time, enhancing the method adaptability to diverse unseen object instances." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.458, + 0.743 + ], + "angle": 0, + "content": "2.3. Generalizable Hand-object Motion Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Synthesizing realistic hand-object-interaction motion on novel object geometries and categories is an emerging research topic that includes two main challenges: how to model generic, realistic, and diverse hand motion patterns from existing HOI data, and how to apply them to novel object instances. To address the first challenge, a branch of data-driven approaches decomposes complicated hand-object contact into spatial relations between each hand joint [30, 37, 38, 79] or hand surface point [71, 77, 81] and the object's relevant closest local regions, and such" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.32 + ], + "angle": 0, + "content": "relations can be formulated as a combination of whether contact occurs [37, 38, 71, 77, 79, 81], hand-object distances [37, 71, 77, 81], positions [37, 38, 77, 79, 81] and directions [37, 38, 79] of contact points, and whether contact points should be closer [71]. Another line of work [28-30, 35] encodes hand-object interaction integrally to an implicit global feature by a conditional variational autoencoder (CVAE) [57] structure. As solutions to the second challenge, generalizable methods [28, 30, 35, 38, 79] decompose the object's geometry from HOI and encode it as a condition for generative models. CAMS [79] further maps objects into a canonicalized object space to decrease the gap in object geometry. One limitation of existing approaches is that they do not focus on multi-hand and multi-object cooperation due to the lack of data support." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.332, + 0.695, + 0.35 + ], + "angle": 0, + "content": "3. Constructing TACO" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.894, + 0.404 + ], + "angle": 0, + "content": "In this section, we respectively describe the data capturing system (Section 3.1), the data annotating pipeline (Section 3.2), and dataset statistics (Section 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.413, + 0.661, + 0.43 + ], + "angle": 0, + "content": "3.1. Data Capturing" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.45, + 0.892, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.609, + 0.848, + 0.624 + ], + "angle": 0, + "content": "Figure 2. Data capturing system and camera views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.825 + ], + "angle": 0, + "content": "We obtain hand motion from multi-view RGB videos while capturing the object motion by attaching four markers on the object surface and tracking them with a mocap system. To this end, as shown in Figure 2, our data capturing system incorporates 12 synchronized industrial FLIR cameras and a NOKOV optical motion capture suite with 6 infrared Mars4H cameras. To capture egocentric RGBD videos, a helmet equipped with a Realsense L515 camera is worn by the actor. Similar to objects, the motion of the L515 camera is tracked by the mocap system. The frequency of our cameras and mocap system is \\(30\\mathrm{Hz}\\). The resolutions of our allocentric and egocentric images are \\(4096\\times 3000\\) and \\(1920\\times 1080\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Object model acquisition. To capture contacts between hands and objects and support relevant studies, we obtain an accurate 3D mesh for each object with an industrial EinScan 3D scanner. Each object mesh is represented by up to 100K triangular faces for fine-grained geometries." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21742" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.872, + 0.296 + ], + "angle": 0, + "content": "
datasetbimanualmulti-object cooperationfunctional manipulationdata characteristics: category-levelegocentricmulti-viewmarkerlessmocapaction label# number of: sequenceframe
FHPA [15]✓✓✓1.2K105K
ObMan [20]-154K
ContactPose[2]✓✓2.3K3.0M
GRAB [61]✓✓1.3K1.6M
HO3D [18]✓✓2778K
KIT Binamual Manipulation [33]✓✓✓✓✓✓✓5881.6M
H2O [34]✓✓✓✓191571K
DexYCB [4]✓✓1.0K582K
H2O-3D [19]✓✓✓1776K
H2O for handover [75]✓✓✓6.0K5.0M
OakInk [72]✓✓778314K
HOI4D [39]✓✓✓✓✓4.0K1.2M
HMDO [66]✓✓✓1221K
ARCTIC [13]✓✓✓✓3392.1M
AffordPose [26]✓✓✓-27K
TACO (ours)✓✓✓✓✓✓✓✓✓2.5K5.2M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.245, + 0.298, + 0.725, + 0.311 + ], + "angle": 0, + "content": "Table 1. Comparison of TACO with existing 3D hand-object interaction datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.32, + 0.245, + 0.336 + ], + "angle": 0, + "content": "3.2. Data Annotating" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.343, + 0.469, + 0.418 + ], + "angle": 0, + "content": "Figure 3 depicts our automatic data acquisition pipeline. Given color images, 3D marker positions, and object models as inputs, the pipeline successively performs object pose optimization, hand keypoint localization, hand pose optimization, hand-object segmentation, and marker removal." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.469, + 0.554 + ], + "angle": 0, + "content": "Object pose optimization. Since our objects are rigid bodies, their motions are 6-dimensional and can be tracked by capturing marker positions using the mocap system. To accurately associate markers with the object, we draw inspiration from Mosh++ [41] and optimize marker-to-surface correspondence to encourage contact between markers and object surfaces while avoiding penetration. The object poses are obtained by combining marker positions relative to the object mesh and the captured marker motions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.555, + 0.469, + 0.736 + ], + "angle": 0, + "content": "3D hand keypoint localization. Leveraging multi-view sensing, we first estimate 2D hand keypoints from each camera view, respectively, and then fuse them into 3D. For 2D hand keypoint estimation, directly applying existing dual-hand approaches [19, 27, 36] could fail when severe occlusion occurs or two hands are far apart. We thus follow [3, 40] and design a method to separately detect two hands using pre-trained YOLOv3 [52] and then utilize MMPose [9] to obtain single-hand 2D keypoints. To fuse multi-view 2D keypoints to 3D under severe occlusion, we use RANSAC [14] to filter out imprecise 2D keypoints and localize 3D keypoints by triangulation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.736, + 0.469, + 0.842 + ], + "angle": 0, + "content": "Contact-aware hand pose optimization. We adopt MANO [53] to formulate a 3D hand mesh as \\(\\Theta_h = \\{\\theta, \\beta, t\\}\\), where \\(\\theta \\in \\mathbb{R}^{48}\\), \\(\\beta \\in \\mathbb{R}^{10}\\), and \\(t \\in \\mathbb{R}^3\\) represent hand pose, hand shape, and wrist position, respectively. Given precomputed 3D hand keypoints and object meshes, we optimize MANO hand meshes by minimizing the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.852, + 0.469, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\Theta} _ {h} = \\underset {\\Theta_ {h}} {\\arg \\min } \\left(\\lambda_ {2 D} \\mathcal {L} _ {2 D} + \\lambda_ {3 D} \\mathcal {L} _ {3 D} + \\lambda_ {a n g l e} \\mathcal {L} _ {a n g l e} + \\right. \\\\ \\left. \\lambda_ {t c} \\mathcal {L} _ {t c} + \\lambda_ {p} \\mathcal {L} _ {p} + \\lambda_ {a} \\mathcal {L} _ {a}\\right), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.321, + 0.892, + 0.396 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{2D}\\) and \\(\\mathcal{L}_{3D}\\) induce hand meshes to get close to the keypoints, \\(\\mathcal{L}_{angle}\\) provides hand pose priors, \\(\\mathcal{L}_{tc}\\) promotes temporal smoothness, \\(\\mathcal{L}_a\\) encourages contacts, and \\(\\mathcal{L}_p\\) prevents hand-object interpenetration. Details of these loss terms are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.517 + ], + "angle": 0, + "content": "Hand-object segmentation. We obtain 2D hand-object masks by first rendering hand-object silhouettes from their 3D meshes. Since MANO meshes differ from real hands, we then apply the Segment Anything Model [32] to refine the masks, using sampled hand-object keypoints as prompts. To help detect hands in subsequent frames, we use Track-anything [70] to track the hands in these frames and localize them." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.893, + 0.79 + ], + "angle": 0, + "content": "Marker removal. Despite the fact that attaching markers to objects can accurately capture object motion, markers may increase the appearance gap between captured objects and those in the wild. One direct defect is that an object pose tracking network could simply track markers rather than learn to track the target object. To improve the object's appearance and make it more realistic, after acquiring hand and object poses, we automatically remove the markers from RGB images. First, we obtain marker positions in the world coordinate system from the mocap system, and place spheres with a radius two times the marker at those positions in a simulation environment. We then render a 2D mask of the spheres for each real camera view. Benefiting from the accurate marker position estimates, such 2D masks can fully cover the markers in the original RGB images. Finally, we mask the spheres in the original RGB images and use a pre-trained LAMA [60] network to inpaint the sphere regions. Figure 4 shows a marker removal example." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.669, + 0.816 + ], + "angle": 0, + "content": "3.3. Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "TACO contains 2.5K motion sequences describing extensive daily tool-using behaviors covering 20 object categories, 196 fine-grained object 3D models, 14 participants, and 15 actions. To support cross-category generalizable HOI studies, we explore the functional diversity of tools" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21743" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.109, + 0.895, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.312, + 0.895, + 0.368 + ], + "angle": 0, + "content": "Figure 3. Automatic data annotating pipeline. The input consists of color frames from allocentric views, pre-scanned object models, and 3D positions of markers attached to object surfaces. We first separately localize 3D hand keypoints and obtain object poses, and then conduct contact-aware optimization to recover MANO [53] hand meshes. We finally segment hands and objects from images and automatically inpaint markers to acquire realistic object appearances." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.374, + 0.209, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.374, + 0.338, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.374, + 0.47, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.478, + 0.47, + 0.521 + ], + "angle": 0, + "content": "Figure 4. An example of marker removal. Three sub-figures respectively show the captured image patch, the automatically-computed marker mask, and the inpainted image patch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.47, + 0.668 + ], + "angle": 0, + "content": "and target objects, with each tool and target object being utilized to perform up to 7 different kinds of actions. We formulate a behavior as using a tool category to perform an action on a target object category and divide our data into \\(131 <\\) tool category, action label, target object category \\(\\rightharpoondown\\) triplets. Some correlated triplets are exemplified in Figure 5. To enable various studies, TACO presents high-resolution color images from 12 allocentric views and RGBD images from one egocentric-view." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.308, + 0.697 + ], + "angle": 0, + "content": "4. Data Quality Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Qualitative contact optimization evaluation. Figure 6 depicts two frames of the TACO dataset. As shown in the figure, the attraction loss \\(\\mathcal{L}_a\\) promotes contact between the hands and the objects but may exacerbate penetration. The penetration loss \\(\\mathcal{L}_p\\) can mitigate penetration but does not address cases where the hands and the objects are not in contact. Our method combines both attraction loss and penetration loss, striking a balance between encouraging contact and preventing penetration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Quantitative hand pose evaluation. To quantitatively examine the quality and value of our hand poses, we conduct a cross-dataset evaluation between TACO and an existing high-quality HOI dataset DexYCB [4]. We select a" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.374, + 0.843, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.565, + 0.894, + 0.621 + ], + "angle": 0, + "content": "Figure 5. Examples of interaction triplets in TACO. The left, middle, and right columns exemplify categories of tool, action, and target object, respectively. Triplets in our dataset are represented by connected paths from tools to target objects." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.624, + 0.895, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.798, + 0.894, + 0.854 + ], + "angle": 0, + "content": "Figure 6. Qualitative contact optimization evaluation. From left to right: original RGB image, optimization without attraction loss, optimization without penetration loss, our method with attraction loss and penetration loss." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "multi-stage method, CMR [5], and a lightweight network, MobRecon [6], both of which can recover occluded hand" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21744" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.165 + ], + "angle": 0, + "content": "
MethodTrain TestTACODexYCBTACO + DexYCB
CMRTACO10.73 / 30.8713.53 / 39.5410.25 / 28.36
DexYCB19.36 / 76.206.51 / 13.926.44 / 13.62
MobReconTACO9.03 / 20.7612.95 / 34.328.69 / 20.14
DexYCB15.17 / 44.736.51 / 13.616.32 / 12.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.167, + 0.47, + 0.208 + ], + "angle": 0, + "content": "Table 2. Cross-dataset evaluation with DexYCB [4] on 3D hand pose estimation. Results are in PA-MPJPE (mm, lower is better) and MPJPE (mm, lower is better), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.473, + 0.535 + ], + "angle": 0, + "content": "meshes from a single-view RGB image captured in HOI scenarios, and then train and test their networks under different data combinations. For DexYCB, we follow its default train/test split (S0). For TACO, we randomly select 60 motion sequences for training and 40 for testing and crop the right hand from each image to align with the single-hand HO3D. Following existing hand mesh recovery methods [5, 6, 45, 67], we select Procrustes-Aligned Mean Per Joint Position Error (PA-MPJPE) and Mean Per Joint Position Error (MPJPE) as evaluation metrics. More details about the metrics can be found in the paper [6]. Table 2 shows the performance of the two methods under different settings. Due to the large gap in hand-object motion between the two datasets, the performances for the two methods would both drop rapidly when the training and the test sets are from different datasets. Nevertheless, combining training data from TACO and DexYCB achieves significant performance gains on both datasets for both the two methods, which demonstrates the accuracy of our hand pose annotations and indicates that TACO is a beneficial data complement for 3D hand pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Marker removal quality evaluation. We conduct a novel experiment to examine the reality of our marker-removed images. The key idea is that a network can easily segment markers from the image while being hard to figure out the same regions when markers are replaced by the original object surface appearances. We thus use the originally captured image patches and the marker-removed ones to train a 2D marker segmentation network, respectively, and then apply the two networks to novel image patches with the same image processing setting as training. In practice, spheres larger than markers (as shown in Figure 4(b)) are required to be segmented, since the marker removal process exactly inpaints these regions. We select U-Net [54] as the network backbone and train the two networks with an L2-loss measuring differences between mask predictions and ground truths. When training and testing on raw image patches, U-Net achieves \\(63.8\\%\\) mean Intersection over Union (mIoU), while its performance sharply drops to \\(11.1\\%\\) mIoU after applying marker removal, indicating the reality and naturalness of the inpainted image regions. Figure 7 further compares heatmaps of network predictions under the two settings, exhibiting ambiguities of marker segmentation posed by realistic inpainted images. More details are provided in the supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.596, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.091, + 0.688, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.162, + 0.596, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.162, + 0.688, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.232, + 0.603, + 0.24 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.091, + 0.795, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.162, + 0.795, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.796, + 0.091, + 0.888, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.162, + 0.889, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.232, + 0.803, + 0.24 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.259, + 0.895, + 0.33 + ], + "angle": 0, + "content": "Figure 7. Marker removal quality evaluation. We train and test U-Net [54] on raw captured images (a) and marker-removed images (b), respectively. In each experiment, the left and right columns show the input image patches and the predicted segmentation heatmaps, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.339, + 0.634, + 0.357 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.894, + 0.472 + ], + "angle": 0, + "content": "In this section, we first introduce our data split (Section 5.1) supporting different research purposes about generalization. We then present three benchmarks on TACO: compositional action recognition (Section 5.2), generalizable hand-object motion forecasting (Section 5.3), and cooperative grasp synthesis (Section 5.4). Details of evaluation metrics are provided in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.483, + 0.616, + 0.501 + ], + "angle": 0, + "content": "5.1. Data Split" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.584 + ], + "angle": 0, + "content": "With diverse object geometries and interaction triplets, TACO focuses on supporting generalizable studies that apply to unseen object geometries or novel interaction triplets. Hence, we carefully split our test data into four subsets with different generalization purposes:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.588, + 0.892, + 0.617 + ], + "angle": 0, + "content": "- Test set 1 (S1): No generalization. Tool geometries and interaction triplets are both involved in the training set." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.618, + 0.892, + 0.663 + ], + "angle": 0, + "content": "- Test set 2 (S2): Geometry-level generalization. The tool geometry is novel, while the interaction triplet appears in the training set." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.664, + 0.892, + 0.708 + ], + "angle": 0, + "content": "- Test set 3 (S3): Interaction-level generalization. The interaction triplet is novel, while the tool categories and geometries are included in the training set." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.709, + 0.892, + 0.739 + ], + "angle": 0, + "content": "- Test set 4 (S4): Compound generalization. The tool category is novel, leading to unseen geometries and triplets." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.588, + 0.892, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.743, + 0.892, + 0.772 + ], + "angle": 0, + "content": "The training set and the four test sets follow a data ratio of 4:1:1:1.5:2.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.786, + 0.803, + 0.803 + ], + "angle": 0, + "content": "5.2. Compositional Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Humans can recognize the action disentangled from their context and appearance biases of the objects, which is referred to as compositionality. To test a system's compositional generalization capability, with a diverse array of tool-action-object triplets in TACO, we benchmark compositional action recognition aiming at identifying actions" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21745" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.111, + 0.089, + 0.43, + 0.219 + ], + "angle": 0, + "content": "
Test SetMethodTop-1 (%, ↑)Top-5 (%, ↑)
S1AIM [73]83.0898.85
CACNF [50]86.1599.62
S2AIM [73]82.8198.83
CACNF [50]77.3496.88
S3AIM [73]53.6581.25
CACNF [50]63.0292.97
S4AIM [73]39.3373.67
CACNF [50]44.0080.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.223, + 0.47, + 0.252 + ], + "angle": 0, + "content": "Table 3. Results on compositional action recognition. Methods are examined via Top-1 and Top-5 accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.468, + 0.352 + ], + "angle": 0, + "content": "in HOI videos. In contrast to traditional action recognition tasks [11, 16, 56, 58], our focus is on scenarios where the tool-action-object triplets in the test set are unseen during training, including novel tool types or geometries. This poses additional challenges for evaluating model generalization and achieving human-like perception." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.352, + 0.468, + 0.412 + ], + "angle": 0, + "content": "Problem formulation: In a bimanual HOI scenario, our goal is to identify action labels from first-person RGB video frames, with optional bounding boxes for hands and objects computed based on their poses." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.412, + 0.468, + 0.457 + ], + "angle": 0, + "content": "Evaluation metrics: Following [50, 73], we use Top-1 Accuracy and Top-5 Accuracy to evaluate the efficacy of action recognition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.468, + 0.806 + ], + "angle": 0, + "content": "Baselines, results, and analysis: We compare two baselines to evaluate the compositional generalization capability: AIM [73], an adaptation of pre-trained image transformer models, representing state-of-the-art traditional action recognition approaches; and CACNF [50], which specializes in compositional and few-shot action recognition. Table 3 shows the results. A significant decrease in accuracy on the most difficult S4 set, compared to high accuracy on the non-generalizable S1 set, underscores compositional generalization challenges. Notably, the S2 set, with novel tool geometries within tool categories, performed similarly to S1, while the S3 set, featuring novel tool-action-object triplets, showed a significant accuracy drop, revealing that while action recognition models inherently possess basic generalizability to geometric changes, they struggle with novel triplets due to the unfamiliarity of interaction contexts. When comparing among baselines, CACNF outperformed AIM in challenging sets S3 and S4, benefiting from its bounding box-based approach that effectively disentangles actions from object and tool geometries, highlighting the advantages of a compositional approach in generalizable cases. All these results underscore the significant impact of compositional generalization on action recognition." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.468, + 0.833 + ], + "angle": 0, + "content": "5.3. Generalizable Hand-object Motion Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "With abundant 3D hand-object mesh annotations provided by TACO, we benchmark generalizable interaction forecasting aiming to predict the following hand-object motions from seen short motion clips. Unlike the prediction of" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "low-frequency human motions [42, 46, 47], we observe that hands commonly achieve a complete manipulation behavior (e.g. pouring all the water out of a bowl) in a very short time, making forecasting both interesting and challenging." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.153, + 0.892, + 0.227 + ], + "angle": 0, + "content": "Problem formulation: In a bimanual HOI scenario, given object point clouds and poses of two hands, the tool, and the target object in consecutive \\(N\\) frames, our goal is to forecast their poses in subsequent \\(M\\) frames. We select \\(N = 10\\) and \\(M = 10\\) in our dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.229, + 0.892, + 0.305 + ], + "angle": 0, + "content": "Evaluation metrics: Following human-object forecasting evaluations [1, 10, 63, 68], we use Mean Per Joint Position Error \\( J_{e} \\) to evaluate predicted left and right-hand skeletons, and leverage translation error \\( T_{e} \\) and rotation error \\( R_{e} \\) to measure object positions and orientations, respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.306, + 0.892, + 0.668 + ], + "angle": 0, + "content": "Baselines, results, and analysis: Due to the lack of hand-object interaction forecasting solutions, we transfer methodologies from human-object motion forecasting as baselines. To cover diverse method designs and possibilities, we select two generative models [62, 68] and two predictive models [10, 68] for comparison. Table 4 compares their performance under different generalization settings. We observed that the tool and the hand holding it play dominant roles during manipulation, making their motion forecasting more challenging compared to others. The generative models, in comparison to predictive ones, yield significantly larger errors for both hands and the tool. One possible reason is that their motion patterns are fast and complex, making the modeling of motion distribution challenging. Enhancing the accuracy of generative methods in hand-object manipulation scenarios is an interesting future direction. When comparing method performances between a non-generalizable set (S1) and sets involving generalization (S2-4), we note that methods consistently exhibit significant performance declines with the right hand and the tool, regardless of whether applied to novel tool geometries (S2, S4) or interaction triplets (S3, S4), meanwhile achieving close results on the others. This trend indicates substantial challenges in generalizing dominant interaction entities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.757, + 0.696 + ], + "angle": 0, + "content": "5.4. Cooperative Grasp Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.704, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Generating human-like hand grasps benefits various applications in VR/AR and dexterous manipulation. Existing grasp synthesis approaches [28, 30, 38, 81] mainly focus on creating stable grasps on static objects without considering interactive purposes. Taking advantage of extensive interaction behaviors from TACO, we benchmark a novel task that aims to synthesize realistic and physically plausible hand grasps in HOI scenarios. Besides achieving stable grasping for the object, the method must comprehend other interactive objects and human hands to generate cooperative motions and avoid conflicts. We evaluate this task on our four test sets, each targeting different aspects of generalization." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Problem formulation: Consider using the right hand to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21746" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.344 + ], + "angle": 0, + "content": "
Test SetMethodJe(mm,↓)Te(mm,↓)Re(°,↓)
Right / LeftTool / TargetTool / Target
S1InterVAE [68]54.9 / 48.955.0 / 15.866.94 / 6.63
MDM [62]61.2 / 53.749.4 / 14.952.81 / 5.85
InterRNN [68]29.3 / 20.425.8 / 11.610.08 / 5.22
CAHMP [10]28.8 / 22.123.9 / 12.310.24 / 4.95
S2InterVAE [68]58.8 / 50.854.3 / 12.288.19 / 5.24
MDM [62]65.8 / 55.848.0 / 10.175.78 / 3.59
InterRNN [68]35.2 / 23.131.0 / 8.911.09 / 4.03
CAHMP [10]31.4 / 22.624.7 / 8.610.65 / 3.32
S3InterVAE [68]56.5 / 50.156.7 / 12.768.03 / 5.69
MDM [62]66.4 / 58.046.2 / 11.561.90 / 4.14
InterRNN [68]32.5 / 22.028.2 / 9.711.42 / 4.32
CAHMP [10]29.3 / 21.624.4 / 9.411.18 / 3.80
S4InterVAE [68]63.9 / 54.263.0 / 14.571.38 / 5.01
MDM [62]70.5 / 57.950.2 / 12.270.25 / 3.55
InterRNN [68]36.6 / 22.330.9 / 10.611.92 / 3.86
CAHMP [10]32.9 / 22.226.6 / 10.811.73 / 3.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.349, + 0.471, + 0.391 + ], + "angle": 0, + "content": "Table 4. Results on generalizable interaction forecasting. We respectively examine predictions of the right hand, the left hand, the tool, and the target object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.469, + 0.462 + ], + "angle": 0, + "content": "manipulate a tool and the left hand to operate a target object. During manipulation, given the meshes of the two objects and the left hand, the goal is to generate a right-hand mesh grasping the tool in a manner conducive to the interaction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.469, + 0.612 + ], + "angle": 0, + "content": "Evaluation metrics: Following static grasp synthesis methods [28, 29, 38], we use interpenetration volume (Pen. \\(V\\)) and contact ratio (Con. \\(R\\)) to measure the physical plausibility of generated hands. To ensure the hand is appropriately interactive, we introduce a new metric named collision ratio (Col. \\(R\\)), which detects collisions between generated hands and the interaction environment. Nevertheless, to quantitatively examine whether the grasps are realistic, we design the FID score (FID) to measure distances between distributions of human grasps and synthesized ones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Baselines, results, and analysis: We select two recent CVAE-based approaches ContactGen [38] and HALO-VAE [30], and modify their network structures to incorporate the left hand and the target object as additional VAE conditions. To assess the significance of interaction environments, we also evaluate HALO-VAE\\(^{-}\\), which omits the use of environment inputs and directly generates from the tool mesh. Due to the strong correlation between hand grasps and tool geometries, we test these approaches with familiar (S1, S3) and unseen tool geometries (S2, S4). Table 5 summarizes quantitative evaluations of the two settings. Notably, HALO-VAE\\(^{-}\\) shows a significant decrease in performance on contact and collision ratios compared to HALO-VAE, underscoring the importance of perceiving and modeling interaction environments. When applied to novel object geometries and categories, all three approaches exhibit higher collision ratios and FID scores, and most of them show a marked decline in their ability to contact tools precisely and prevent penetrations. Figure 8 exemplifies" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.089, + 0.906, + 0.204 + ], + "angle": 0, + "content": "
Test SetMethodPen. V (cm3, ↓)Con. R (%, ↑)Col. R (%, ↓)FID (10-2, ↓)
S1 ∪ S3ContactGen [38]0.2780.1210.4012.18
HALO-VAE-0.3176.274.194.89
HALO-VAE [30]0.2682.862.392.02
S2 ∪ S4ContactGen [38]0.2674.9817.5622.53
HALO-VAE-0.2971.6914.255.09
HALO-VAE [30]0.3283.5211.112.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.209, + 0.892, + 0.265 + ], + "angle": 0, + "content": "Table 5. Results on cooperative grasp synthesis. We measure the physical plausibility of generated grasps by interpenetration volume and contact ratio, meanwhile using collision ratio to examine interaction faithfulness and FID score to assess reality." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.273, + 0.892, + 0.318 + ], + "angle": 0, + "content": "failure cases of synthesized grasps, indicating a large room for improving physical feasibility and naturalness on complex and slender geometries." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.49, + 0.892, + 0.532 + ], + "angle": 0, + "content": "Figure 8. Failure cases of cooperative grasp synthesis. (a) and (b) are generated by ContactGen [38], while (c) and (d) are from HALO-VAE [30]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.544, + 0.757, + 0.559 + ], + "angle": 0, + "content": "6. Limitations and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.78 + ], + "angle": 0, + "content": "We present TACO, the first large-scale, real-world 4D bi-manual hand-object manipulation dataset. It encompasses a wide range of tool-action-object compositions and object geometries. TACO contains a total of \\(2.5\\mathrm{K}\\) motion sequences and \\(5.2\\mathrm{M}\\) video frames, captured from 12 third-person views and one egocentric view. We contribute an automatic data acquisition pipeline that accurately recovers hand-object meshes and segmentation, along with realistic hand-object appearances. Leveraging TACO's diverse data, we benchmark compositional action recognition, generalizable hand-object motion forecasting, and cooperative grasp synthesis. These benchmarks reveal new insights, challenges, and opportunities in the field of generalizable hand-object interaction studies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.902 + ], + "angle": 0, + "content": "There are three major limitations in TACO. Firstly, TACO currently does not cover articulated objects. Secondly, while TACO offers an extensive exploration of object geometries and HOI behaviors, it lacks scene diversities that are also crucial for understanding human manipulations. Thirdly, our solution to marker removal is an application of generative models, hence the original object appearances cannot be recovered perfectly." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21747" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Vida Adeli, Mahsa Ehsanpour, Ian Reid, Juan Carlos Niebles, Silvio Savarese, Ehsan Adeli, and Hamid Rezatofighi. Tripod: Human trajectory and pose dynamics forecasting in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13390-13400, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.282 + ], + "angle": 0, + "content": "[2] Samarth Brahmbhatt, Chengcheng Tang, Christopher D Twigg, Charles C Kemp, and James Hays. Contactpose: A dataset of grasps with object contact and hand pose. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 361-378. Springer, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.47, + 0.338 + ], + "angle": 0, + "content": "[3] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.47, + 0.436 + ], + "angle": 0, + "content": "[4] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.521 + ], + "angle": 0, + "content": "[5] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.47, + 0.605 + ], + "angle": 0, + "content": "[6] Xingyu Chen, Yufeng Liu, Dong Yajiao, Xiong Zhang, Chongyang Ma, Yanmin Xiong, Yuan Zhang, and Xiaoyan Guo. Mobrecon: Mobile-friendly hand mesh reconstruction from monocular image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.47, + 0.688 + ], + "angle": 0, + "content": "[7] Sammy Christen, Muhammed Kocabas, Emre Aksan, Jemin Hwangbo, Jie Song, and Otmar Hilliges. D-grasp: Physically plausible dynamic grasp synthesis for hand-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20577-20586, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.69, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[8] Sammy Christen, Wei Yang, Claudia Pérez-D'Arpino, Otmar Hilliges, Dieter Fox, and Yu-Wei Chao. Learning human-to-robot handovers from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9654–9664, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.761, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[9] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[10] Enric Corona, Albert Pumarola, Guillem Alenya, and Francesc Moreno-Noguer. Context-aware human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6992-7001, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European conference on computer vision (ECCV), pages 720-736, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[12] Christian RG Dreher, Mirko Wächter, and Tamim Asfour. Learning object-action relations from bimanual human demonstration using graph networks. IEEE Robotics and Automation Letters, 5(1):187-194, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[13] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J Black, and Otmar Hilliges. Arctic: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12954, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[14] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.345, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[15] Guillermo Garcia-Hernando, Shanxin Yuan, Seungryul Baek, and Tae-Kyun Kim. First-person hand action benchmark with rgb-d videos and 3d hand pose annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 409-419, 2018. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.415, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[16] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[17] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[18] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3196-3206, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[19] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11090-11100, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[20] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11807-11816, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[21] Roei Herzig, Elad Ben-Avraham, Karttikeya Mangalam, Amir Bar, Gal Chechik, Anna Rohrbach, Trevor Darrell, and Amir Globerson. Object-region video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3148-3159, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21748" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[22] Zhi Hou, Xiaojiang Peng, Yu Qiao, and Dacheng Tao. Visual compositional learning for human-object interaction detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 584-600. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[23] Zhi Hou, Baosheng Yu, Yu Qiao, Xiaojiang Peng, and Dacheng Tao. Affordance transfer learning for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 495-504, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[24] Zhi Hou, Baosheng Yu, Yu Qiao, Xiaojiang Peng, and Dacheng Tao. Detecting human-object interaction via fabricated compositional learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14646-14655, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.361 + ], + "angle": 0, + "content": "[25] Zhi Hou, Baosheng Yu, and Dacheng Tao. Discovering human-object interaction concepts via self-compositional learning. In European Conference on Computer Vision, pages 461-478. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.47, + 0.432 + ], + "angle": 0, + "content": "[26] Juntao Jian, Xiuping Liu, Manyi Li, Ruizhen Hu, and Jian Liu. Affordpose: A large-scale dataset of hand-object interactions with affordance-driven hand pose. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14713-14724, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.47, + 0.517 + ], + "angle": 0, + "content": "[27] Changlong Jiang, Yang Xiao, Cunlin Wu, Mingyang Zhang, Jinghong Zheng, Zhiguo Cao, and Joey Tianyi Zhou. A2j-transformer: Anchor-to-joint transformer network for 3d interacting hand pose estimation from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8846-8855, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[28] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J Black, Krikamol Muandet, and Siyu Tang. Grasping field: Learning implicit representations for human grasps. In 2020 International Conference on 3D Vision (3DV), pages 333-344. IEEE, 2020. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.47, + 0.658 + ], + "angle": 0, + "content": "[29] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J Black, Krikamol Muandet, and Siyu Tang. Grasping field: Learning implicit representations for human grasps. In 2020 International Conference on 3D Vision (3DV), pages 333-344. IEEE, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[30] Korrawe Karunratanakul, Adrian Spurr, Zicong Fan, Otmar Hilliges, and Siyu Tang. A skeleton-driven neural occupancy representation for articulated hands. In 2021 International Conference on 3D Vision (3DV), pages 11-21. IEEE, 2021. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[31] Tae Soo Kim, Jonathan Jones, and Gregory D Hager. Motion guided attention fusion to recognize interactions from videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13076-13086, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[33] Franziska Krebs, Andre Meixner, Isabel Patzer, and Tamim Asfour. The kit bimanual manipulation dataset. In 2020 IEEE-RAS 20th International Conference on Humanoid Robots (Humanoids), pages 499-506. IEEE, 2021. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[34] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.217 + ], + "angle": 0, + "content": "[35] Haoming Li, Xinzhuo Lin, Yang Zhou, Xiang Li, Yuchi Huo, Jiming Chen, and Qi Ye. Contact2grasp: 3d grasp synthesis via hand-object contact constraint. arXiv preprint arXiv:2210.09245, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.893, + 0.288 + ], + "angle": 0, + "content": "[36] Mengcheng Li, Liang An, Hongwen Zhang, Lianpeng Wu, Feng Chen, Tao Yu, and Yebin Liu. Interacting attention graph for single image two-hand reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2761-2770, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.893, + 0.356 + ], + "angle": 0, + "content": "[37] Qingtao Liu, Yu Cui, Zhengnan Sun, Haoming Li, Gaofeng Li, Lin Shao, Jiming Chen, and Qi Ye. Dexrepnet: Learning dexterous robotic grasping network with geometric and spatial hand-object representations. arXiv preprint arXiv:2303.09806, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.893, + 0.426 + ], + "angle": 0, + "content": "[38] Shaowei Liu, Yang Zhou, Jimei Yang, Saurabh Gupta, and Shenlong Wang. Contactgen: Generative contact modeling for grasp generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20609-20620, 2023. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.893, + 0.509 + ], + "angle": 0, + "content": "[39] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.893, + 0.58 + ], + "angle": 0, + "content": "[40] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.894, + 0.65 + ], + "angle": 0, + "content": "[41] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.893, + 0.692 + ], + "angle": 0, + "content": "[42] Wei Mao, Richard I Hartley, Mathieu Salzmann, et al. Contact-aware human motion forecasting. Advances in Neural Information Processing Systems, 35:7356-7367, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.893, + 0.774 + ], + "angle": 0, + "content": "[43] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-else: Compositional action recognition with spatial-temporal interaction networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1049–1059, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.893, + 0.831 + ], + "angle": 0, + "content": "[44] Eley Ng, Ziang Liu, and Monroe Kennedy. It takes two: Learning to plan for human-robot cooperative carrying. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7526–7532. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[45] JoonKyu Park, Yeonguk Oh, Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Handocctnet: Occlusion-robust 3d hand mesh estimation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1496–1505, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21749" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[46] Xiaogang Peng, Yaodi Shen, Haoran Wang, Binling Nie, Yiqiang Wang, and Zizhao Wu. Somoformer: Social-aware motion transformer for multi-person motion prediction. arXiv preprint arXiv:2208.09224, 2022.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.471, + 0.218 + ], + "angle": 0, + "content": "[47] Xiaogang Peng, Siyuan Mao, and Zizhao Wu. Trajectory-aware body interaction transformer for multi-person pose forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17121-17130, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[48] Yuzhe Qin, Hao Su, and Xiaolong Wang. From one hand to multiple hands: Imitation learning for dexterous manipulation from single-camera teleoperation. IEEE Robotics and Automation Letters, 7(4):10873-10881, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.471, + 0.346 + ], + "angle": 0, + "content": "[49] Yuzhe Qin, Yueh-Hua Wu, Shaowei Liu, Hanwen Jiang, Ruihan Yang, Yang Fu, and Xiaolong Wang. Dexamv: Imitation learning for dexterous manipulation from human videos. In European Conference on Computer Vision, pages 570-587. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.471, + 0.403 + ], + "angle": 0, + "content": "[50] Gorjan Radevski, Marie-Francine Moens, and Tinne Tuyte-laars. Revisiting spatio-temporal layouts for compositional action recognition. arXiv preprint arXiv:2111.01936, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.471, + 0.459 + ], + "angle": 0, + "content": "[51] Ramanathan Rajendiran, Debaditya Roy, and Basura Fernando. Modelling spatio-temporal interactions for compositional action recognition. arXiv preprint arXiv:2305.02673, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.471, + 0.49 + ], + "angle": 0, + "content": "[52] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.471, + 0.533 + ], + "angle": 0, + "content": "[53] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.471, + 0.617 + ], + "angle": 0, + "content": "[54] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[55] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.471, + 0.759 + ], + "angle": 0, + "content": "[56] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.471, + 0.815 + ], + "angle": 0, + "content": "[57] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[58] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[59] Pengzhan Sun, Bo Wu, Xunsong Li, Wen Li, Lixin Duan, and Chuang Gan. Counterfactual debiasing inference for compositional action recognition. In Proceedings of the 29th" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ACM International Conference on Multimedia, pages 3220-3228, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[60] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.894, + 0.277 + ], + "angle": 0, + "content": "[61] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 581-600. Springer, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.894, + 0.331 + ], + "angle": 0, + "content": "[62] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.894, + 0.403 + ], + "angle": 0, + "content": "[63] Weilin Wan, Lei Yang, Lingjie Liu, Zhuoying Zhang, Ruixing Jia, Yi-King Choi, Jia Pan, Christian Theobalt, Taku Komura, and Wenping Wang. Learn to predict how humans manipulate large-sized objects from interactive motions. IEEE Robotics and Automation Letters, 7(2):4702-4709, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.894, + 0.461 + ], + "angle": 0, + "content": "[64] Haozhong Wang, Hua Yu, and Qiang Zhang. Detecting zero-shot human-object interaction with visual-text modeling. In 2023 9th International Conference on Virtual Reality (ICVR), pages 155-162. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.894, + 0.545 + ], + "angle": 0, + "content": "[65] Suchen Wang, Yueqi Duan, Henghui Ding, Yap-Peng Tan, Kim-Hui Yap, and Junsong Yuan. Learning transferable human-object interaction detector with natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 939-948, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.894, + 0.602 + ], + "angle": 0, + "content": "[66] Wei Xie, Zhipeng Yu, Zimeng Zhao, Binghui Zuo, and Yanggang Wang. Hmdo: Markerless multi-view hand manipulation capture with deformable objects. Graphical Models, 127:101178, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.894, + 0.673 + ], + "angle": 0, + "content": "[67] Hao Xu, Tianyu Wang, Xiao Tang, and Chi-Wing Fu. H2onet: Hand-occlusion-and-orientation-aware network for real-time 3d hand mesh reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17048–17058, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.894, + 0.744 + ], + "angle": 0, + "content": "[68] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14928-14940, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.894, + 0.815 + ], + "angle": 0, + "content": "[69] Rui Yan, Peng Huang, Xiangbo Shu, Junhao Zhang, Yonghua Pan, and Jinhui Tang. Look less think more: Rethinking compositional action recognition. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3666-3675, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.894, + 0.858 + ], + "angle": 0, + "content": "[70] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[71] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21750" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision, pages 11097-11106, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[72] Lixin Yang, Kailin Li, Xinyu Zhan, Fei Wu, Anran Xu, Liu Liu, and Cewu Lu. Oakink: A large-scale knowledge repository for understanding hand-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20953-20962, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.247 + ], + "angle": 0, + "content": "[73] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. AIM: Adapting image models for efficient video action recognition. In The Eleventh International Conference on Learning Representations, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[74] Jianglong Ye, Jiashun Wang, Binghao Huang, Yuzhe Qin, and Xiaolong Wang. Learning continuous grasping function with a dexterous hand from human demonstrations. IEEE Robotics and Automation Letters, 8(5):2882-2889, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.374 + ], + "angle": 0, + "content": "[75] Ruolin Ye, Wenqiang Xu, Zhendong Xue, Tutian Tang, Yanfeng Wang, and Cewu Lu. H2o: A benchmark for visual human-human object handover analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15762-15771, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[76] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Is an object-centric video representation beneficial for transfer? In Proceedings of the Asian Conference on Computer Vision, pages 1976-1994, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[77] He Zhang, Yuting Ye, Takaaki Shiratori, and Taku Komura. Manipnet: neural manipulation synthesis with a handobject spatial representation. ACM Transactions on Graphics (ToG), 40(4):1-14, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.469, + 0.557 + ], + "angle": 0, + "content": "[78] Hui Zhang, Sammy Christen, Zicong Fan, Luocheng Zheng, Jemin Hwangbo, Jie Song, and Otmar Hilliges. Artigrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. arXiv preprint arXiv:2309.03891, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.56, + 0.469, + 0.641 + ], + "angle": 0, + "content": "[79] Juntian Zheng, Qingyuan Zheng, Lixing Fang, Yun Liu, and Li Yi. Cams: Canonicalized manipulation spaces for category-level functional hand-object manipulation synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 585-594, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.469, + 0.713 + ], + "angle": 0, + "content": "[80] Desen Zhou, Zhichao Liu, Jian Wang, Leshan Wang, Tao Hu, Errui Ding, and Jingdong Wang. Human-object interaction detection via disentangled transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19568-19577, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.469, + 0.781 + ], + "angle": 0, + "content": "[81] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Toch: Spatio-temporal object-to-hand correspondence for motion refinement. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.785, + 0.469, + 0.84 + ], + "angle": 0, + "content": "[82] Zehao Zhu, Jiashun Wang, Yuzhe Qin, Deqing Sun, Varun Jampani, and Xiaolong Wang. Contactart: Learning 3d interaction priors for category-level articulated object and hand poses estimation. arXiv preprint arXiv:2305.01618, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.84 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21751" + } + ] +] \ No newline at end of file diff --git a/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/b3378a53-d4f8-49d5-95a4-af3ae029dc3c_origin.pdf b/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/b3378a53-d4f8-49d5-95a4-af3ae029dc3c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2b5a50ff630e3655e96629944a37d36bb3f2bee8 --- /dev/null +++ b/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/b3378a53-d4f8-49d5-95a4-af3ae029dc3c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f10bb3646550e607908b6371e5d0068e31be7044db88939041d2d8af6123fd1f +size 7222731 diff --git a/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/full.md b/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..da9ac987abed016a99aec0c79dd6ceb7bcc0dec1 --- /dev/null +++ b/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/full.md @@ -0,0 +1,354 @@ +# TACO: Benchmarking Generalizable Bimanual Tool-ACTION-Object Understanding + +Yun Liu $^{1,2,3}$ , Haolin Yang $^{4}$ , Xu Si $^{1}$ , Ling Liu $^{5}$ , Zipeng Li $^{1}$ , Yuxiang Zhang $^{1}$ , Yebin Liu $^{1}$ , Li Yi $^{\dagger,1,2,3}$ $^{1}$ Tsinghua University $^{2}$ Shanghai Artificial Intelligence Laboratory $^{3}$ Shanghai Qi Zhi Institute $^{4}$ Beijing University of Posts and Telecommunications $^{5}$ Beijing Institute of Technology + +Third-person View Video + +![](images/a7095185afdc72fa1d218f9b36580ec57d97c848f52a363f324c5e0accba8c0b.jpg) +Egocentric Video + +Hand-Object Mesh + +![](images/70734ec12711cbd502686f0cf04fcbe37fe476736f6157b07689d585b57ba2f8.jpg) + +Segmentation + +Realistic Appearance + +Behavior Triplet + +![](images/52d6072e2a8a17c8a867f3635b03989e0bfbf803bd218696b52e03321f768a38.jpg) +Motion Forecasting + +![](images/44766966b467364b968a1578dcc6d7def426b5a760bf69748433cdc3ab198132.jpg) + +Test Time Generalization + +![](images/2148b55abd1a66d0860a6e434ab73586ad739583895c9b239ea9211b5485c30c.jpg) + +![](images/2cf661d8a44082cdf5294d66548ebb37f4d402dc4ebbd69e428c2091098fec5b.jpg) + +![](images/3c00496f4791b5691aef890c75cd37ebf498446e9bc38296260cf8f20cc34a12.jpg) + +Novel Triplet + +![](images/fc06444b2e7b7b9ec0018992f2cb7b04869f3f549acccc2aac86b90e40eae227.jpg) +Cooperative Grasp Synthesis + +![](images/9e70c1491c30e0d53f4ed2f11665e522b1a0ca9b8b0533c3e42d074a3e92a835.jpg) +Action Recognition +Figure 1. TACO is a large-scale bimanual hand-object manipulation dataset covering extensive tool-action-object combinations in real-world scenarios. It involves videos from both 12 third-person views and one egocentric view, together with precise hand-object meshes, 2D segmentation, realistic hand-object appearances, and behavior triplet annotations. With rich diversities of object shapes and interaction behaviors, TACO supports test-time generalization to unseen object geometries and novel behavior triplets and benchmarks various generalizable research topics, e.g., action recognition, motion forecasting, and cooperative grasp synthesis. + +brush something + +![](images/fe7f90247c143cc264e278f8f08604b69edd2e2f50e2aa125d834e63a7ba8cdb.jpg) + +![](images/1662622d9486b0169786fcbaa9d3fad799678c820f01fa854b9ee8b35ed7d85d.jpg) + +![](images/ae181eba02a4fe7cb58fc6f2bf2d3bbba23318160dbf8e1951a348278547f20f.jpg) + +![](images/9de51667526a0f7424c1bc7b0f06612d28265765f6604c1298c1ec96b9af9bee.jpg) + +# Abstract + +Humans commonly work with multiple objects in daily life and can intuitively transfer manipulation skills to novel objects by understanding object functional regularities. However, existing technical approaches for analyzing and synthesizing hand-object manipulation are mostly limited to handling a single hand and object due to the lack of data support. To address this, we construct TACO, an extensive bimanual hand-object-interaction dataset spanning a large variety of tool-action-object compositions for daily human activities. TACO contains 2.5K motion sequences paired with third-person and egocentric views, precise hand-object 3D meshes, and action labels. To rapidly + +expand the data scale, we present a fully automatic data acquisition pipeline combining multi-view sensing with an optical motion capture system. With the vast research fields provided by TACO, we benchmark three generalizable hand-object-interaction tasks: compositional action recognition, generalizable hand-object motion forecasting, and cooperative grasp synthesis. Extensive experiments reveal new insights, challenges, and opportunities for advancing the studies of generalizable hand-object motion analysis and synthesis. Our data and code are available at https://taco2024.github.io. + +# 1. Introduction + +In our everyday lives, humans effortlessly synchronize the movements of both hands to manipulate a pair of objects, such as using a spatula to stir in a pan while cooking. Usually, these objects are not symmetric in their roles, with one acting as a tool to enhance the bimanual action performed on the other object. This allows us to characterize the bimanual behaviors with tool-action-object triplets. Various triplet combinations usually incorporate intricate and differing temporal and spatial coordination among the tool, the object, and the two hands. Understanding versatile bimanual behaviors could immediately benefit numerous applications in VR/AR, human-robot interaction [8, 44], and dexterous manipulation [7, 49, 74, 78], which pose significant challenges for today's computer vision systems. + +Tackling the above challenges requires the support of large-scale annotation-rich datasets. Existing dataset efforts [2, 4, 15, 18, 26, 39, 72] on hand-object interactions (HOI) primarily focus on unimanual actions. However, simply aggregating two unimanual actions falls short of encompassing bimanual coordination behaviors. Furthermore, some works [13, 19, 34, 66, 75] examine bimanual manipulation of a single object which can hardly extend to interpreting tool-action-object triplets. The most relevant dataset [33] currently available covers only 12 objects and fewer than 20 distinct tool-action-object triplets, severely limiting its potential to facilitate and benchmark the understanding of bimanual tool-action-object interactions that can generalize to novel object geometries or previously unseen triplets. This limitation partially stems from the challenging nature of jointly capturing two dynamic hands interacting with two objects in the real world, but further underscores the urgency of developing methodologies for generalizable bimanual HOI understanding. Therefore, we put generalization as our primary focus while exploring the curation of a large-scale bimanual HOI dataset that can both facilitate and benchmark generalizable understanding. + +In particular, we present TACO, a large-scale bimanual manipulation dataset encompassing a diverse array of tool-action-object compositions in real-world settings. Serving as a knowledge base of multi-object cooperation, TACO focuses on daily scenarios involving the use of tools to interact with target objects and collects 131 types of triplets across 20 object categories, 196 object instances, and 15 daily actions. We organize the dataset based on the triplets and make sure triplets have different levels of overlaps, this naturally defines the semantic distance between different motion trajectories and supports studying generalization with different extents. To expedite the expansion of our dataset, we combine the advantages of marker-based and markerless motion capture systems and present a fully automatic data acquisition pipeline that can guarantee motion quality and + +visual data quality at the same time. For each time step, the pipeline automatically provides labels including the precise recovery of hand-object mesh and segmentation on markerless vision data. As a result, TACO comprises a total of 2.5K motion sequences and 5.2M video frames incorporating 3rd-person and egocentric views. + +Benefiting from the rich and diverse motion sequences within our dataset, we carefully benchmark three tasks aiming at generalizing common human manipulation knowledge to unseen object geometries and novel tool-action-object combinations: 1) compositional action recognition, 2) generalizable hand-object motion forecasting, and 3) cooperative grasp synthesis. Extensive studies are conducted on these benchmarks: First, diverse hand-object interactive motions with action annotations pave the way to learning to recognize hand-object action under novel object combinations in real-world scenarios. Additionally, the availability of 4D dynamic hand-object pose sequences allows for fine-grained hand-object motion tendency forecasting. Finally, we provide contact-rich hand-object meshes to facilitate studying hand grasp synthesis for unseen object geometries and categories under interaction scenarios. Extensive experiments have been conducted on these benchmarks, revealing generalization challenges faced by existing methodologies when dealing with novel object geometries, object categories, and interaction triplets. We anticipate that TACO will offer more opportunities for researching and developing universal generalization strategies in the understanding and creation of hand-object interactions. + +In summary, our main contributions are threefold: + +- We present TACO, the first large-scale real-world 4D bi-manual hand-object-interaction dataset covering diverse tool-action-object compositions and object geometries. +- We develop an automatic data acquisition pipeline that provides precise recovery of hand-object mesh and segmentation together with markerless vision data. +- We benchmark three tasks toward generalizable hand-object motion analysis and synthesis. We provide a comprehensive discussion and highlight the new challenges posed by TACO. + +# 2. Related Work + +# 2.1. Hand-object Interaction Datasets + +Understanding hand-object interaction is an emerging research topic that is supported by a large number of datasets. Many widely-used video datasets [11, 12, 17, 55] capture 2D real-world hand-object interaction videos to facilitate traditional visual perception studies such as action recognition and HOI detection. With the rapid development of 3D computer vision, recent advances have incorporated 3D hand-object mesh annotations into data-capturing processes and presented various 3D [20, 26] or + +4D [2, 4, 12, 13, 15, 18, 19, 34, 39, 61, 66, 72, 75] datasets with different focuses. Table 1 briefly summarizes the characteristics of these datasets. A majority of existing datasets focus on single-hand grasping, a basic manipulation skill that can benefit existing dexterous robot learning [7, 48, 49, 78]. Beyond grasping, a line of datasets captures diverse and complex manipulation skills requiring humans to perform complicated, multi-step actions like using tools [13, 33, 39], deforming soft bodies [66], and multi-person handovers [75]. From the perspective of object behaviors, KIT [33], HOI4D [39], and AffordPose[26] capture functional manipulations aiming at showing actual functions and usages of objects. KIT further collects multi-object interaction behaviors, however, it lacks object diversities and is hard to support generalizable studies. Our dataset captures diverse bimanual functional manipulation behaviors for multi-object cooperation, serving a wide range of research directions. + +# 2.2. Generalizable Hand-object Motion Analysis + +Generalizable hand-object-interaction motion analysis aims to understand the attributes of hand-object motion when manipulating novel object instances. To precisely recognize hand actions during the manipulation of unseen objects, a series of compositional action recognition approaches [21, 31, 43, 50, 51, 59, 69, 76] represent objects through their bounding boxes to prevent the network from overfitting to intricate object geometries. To detect and infer action-object pairs under few-shot or zero-shot settings, a prevailing methodology used in recent generalizable HOI detection methods [22-24, 64, 65, 80] decomposes actions and objects into distinct features. This decomposition allows leveraging abundant training data with either similar actions or similar objects to handle rare action-object pairs that emerge at test time. It also provides the potential to discover novel and reasonable action-object pairs [25]. To estimate hand and object poses without object geometries, Zhu et al. [82] propose learning hand-object interaction priors for each object category and transferring them to unseen objects within the same category during test time, enhancing the method adaptability to diverse unseen object instances. + +# 2.3. Generalizable Hand-object Motion Synthesis + +Synthesizing realistic hand-object-interaction motion on novel object geometries and categories is an emerging research topic that includes two main challenges: how to model generic, realistic, and diverse hand motion patterns from existing HOI data, and how to apply them to novel object instances. To address the first challenge, a branch of data-driven approaches decomposes complicated hand-object contact into spatial relations between each hand joint [30, 37, 38, 79] or hand surface point [71, 77, 81] and the object's relevant closest local regions, and such + +relations can be formulated as a combination of whether contact occurs [37, 38, 71, 77, 79, 81], hand-object distances [37, 71, 77, 81], positions [37, 38, 77, 79, 81] and directions [37, 38, 79] of contact points, and whether contact points should be closer [71]. Another line of work [28-30, 35] encodes hand-object interaction integrally to an implicit global feature by a conditional variational autoencoder (CVAE) [57] structure. As solutions to the second challenge, generalizable methods [28, 30, 35, 38, 79] decompose the object's geometry from HOI and encode it as a condition for generative models. CAMS [79] further maps objects into a canonicalized object space to decrease the gap in object geometry. One limitation of existing approaches is that they do not focus on multi-hand and multi-object cooperation due to the lack of data support. + +# 3. Constructing TACO + +In this section, we respectively describe the data capturing system (Section 3.1), the data annotating pipeline (Section 3.2), and dataset statistics (Section 3.3). + +# 3.1. Data Capturing + +![](images/b71dd1aa85185fdfa4fcd91289d6f36f3fae586aafb9b0f3617a0f282330aa39.jpg) +Figure 2. Data capturing system and camera views. + +We obtain hand motion from multi-view RGB videos while capturing the object motion by attaching four markers on the object surface and tracking them with a mocap system. To this end, as shown in Figure 2, our data capturing system incorporates 12 synchronized industrial FLIR cameras and a NOKOV optical motion capture suite with 6 infrared Mars4H cameras. To capture egocentric RGBD videos, a helmet equipped with a Realsense L515 camera is worn by the actor. Similar to objects, the motion of the L515 camera is tracked by the mocap system. The frequency of our cameras and mocap system is $30\mathrm{Hz}$ . The resolutions of our allocentric and egocentric images are $4096\times 3000$ and $1920\times 1080$ , respectively. + +Object model acquisition. To capture contacts between hands and objects and support relevant studies, we obtain an accurate 3D mesh for each object with an industrial EinScan 3D scanner. Each object mesh is represented by up to 100K triangular faces for fine-grained geometries. + +
datasetbimanualmulti-object cooperationfunctional manipulationdata characteristics: category-levelegocentricmulti-viewmarkerlessmocapaction label# number of: sequenceframe
FHPA [15]✓✓✓1.2K105K
ObMan [20]-154K
ContactPose[2]✓✓2.3K3.0M
GRAB [61]✓✓1.3K1.6M
HO3D [18]✓✓2778K
KIT Binamual Manipulation [33]✓✓✓✓✓✓✓5881.6M
H2O [34]✓✓✓✓191571K
DexYCB [4]✓✓1.0K582K
H2O-3D [19]✓✓✓1776K
H2O for handover [75]✓✓✓6.0K5.0M
OakInk [72]✓✓778314K
HOI4D [39]✓✓✓✓✓4.0K1.2M
HMDO [66]✓✓✓1221K
ARCTIC [13]✓✓✓✓3392.1M
AffordPose [26]✓✓✓-27K
TACO (ours)✓✓✓✓✓✓✓✓✓2.5K5.2M
+ +Table 1. Comparison of TACO with existing 3D hand-object interaction datasets. + +# 3.2. Data Annotating + +Figure 3 depicts our automatic data acquisition pipeline. Given color images, 3D marker positions, and object models as inputs, the pipeline successively performs object pose optimization, hand keypoint localization, hand pose optimization, hand-object segmentation, and marker removal. + +Object pose optimization. Since our objects are rigid bodies, their motions are 6-dimensional and can be tracked by capturing marker positions using the mocap system. To accurately associate markers with the object, we draw inspiration from Mosh++ [41] and optimize marker-to-surface correspondence to encourage contact between markers and object surfaces while avoiding penetration. The object poses are obtained by combining marker positions relative to the object mesh and the captured marker motions. + +3D hand keypoint localization. Leveraging multi-view sensing, we first estimate 2D hand keypoints from each camera view, respectively, and then fuse them into 3D. For 2D hand keypoint estimation, directly applying existing dual-hand approaches [19, 27, 36] could fail when severe occlusion occurs or two hands are far apart. We thus follow [3, 40] and design a method to separately detect two hands using pre-trained YOLOv3 [52] and then utilize MMPose [9] to obtain single-hand 2D keypoints. To fuse multi-view 2D keypoints to 3D under severe occlusion, we use RANSAC [14] to filter out imprecise 2D keypoints and localize 3D keypoints by triangulation. + +Contact-aware hand pose optimization. We adopt MANO [53] to formulate a 3D hand mesh as $\Theta_h = \{\theta, \beta, t\}$ , where $\theta \in \mathbb{R}^{48}$ , $\beta \in \mathbb{R}^{10}$ , and $t \in \mathbb{R}^3$ represent hand pose, hand shape, and wrist position, respectively. Given precomputed 3D hand keypoints and object meshes, we optimize MANO hand meshes by minimizing the following loss function: + +$$ +\begin{array}{l} \hat {\Theta} _ {h} = \underset {\Theta_ {h}} {\arg \min } \left(\lambda_ {2 D} \mathcal {L} _ {2 D} + \lambda_ {3 D} \mathcal {L} _ {3 D} + \lambda_ {a n g l e} \mathcal {L} _ {a n g l e} + \right. \\ \left. \lambda_ {t c} \mathcal {L} _ {t c} + \lambda_ {p} \mathcal {L} _ {p} + \lambda_ {a} \mathcal {L} _ {a}\right), \tag {1} \\ \end{array} +$$ + +where $\mathcal{L}_{2D}$ and $\mathcal{L}_{3D}$ induce hand meshes to get close to the keypoints, $\mathcal{L}_{angle}$ provides hand pose priors, $\mathcal{L}_{tc}$ promotes temporal smoothness, $\mathcal{L}_a$ encourages contacts, and $\mathcal{L}_p$ prevents hand-object interpenetration. Details of these loss terms are provided in the supplementary material. + +Hand-object segmentation. We obtain 2D hand-object masks by first rendering hand-object silhouettes from their 3D meshes. Since MANO meshes differ from real hands, we then apply the Segment Anything Model [32] to refine the masks, using sampled hand-object keypoints as prompts. To help detect hands in subsequent frames, we use Track-anything [70] to track the hands in these frames and localize them. + +Marker removal. Despite the fact that attaching markers to objects can accurately capture object motion, markers may increase the appearance gap between captured objects and those in the wild. One direct defect is that an object pose tracking network could simply track markers rather than learn to track the target object. To improve the object's appearance and make it more realistic, after acquiring hand and object poses, we automatically remove the markers from RGB images. First, we obtain marker positions in the world coordinate system from the mocap system, and place spheres with a radius two times the marker at those positions in a simulation environment. We then render a 2D mask of the spheres for each real camera view. Benefiting from the accurate marker position estimates, such 2D masks can fully cover the markers in the original RGB images. Finally, we mask the spheres in the original RGB images and use a pre-trained LAMA [60] network to inpaint the sphere regions. Figure 4 shows a marker removal example. + +# 3.3. Dataset Statistics + +TACO contains 2.5K motion sequences describing extensive daily tool-using behaviors covering 20 object categories, 196 fine-grained object 3D models, 14 participants, and 15 actions. To support cross-category generalizable HOI studies, we explore the functional diversity of tools + +![](images/96ede0e916f3a4a2ee1b2ab0b6cca3059268fc7f3f6abfb3fba29815e032dca1.jpg) +Figure 3. Automatic data annotating pipeline. The input consists of color frames from allocentric views, pre-scanned object models, and 3D positions of markers attached to object surfaces. We first separately localize 3D hand keypoints and obtain object poses, and then conduct contact-aware optimization to recover MANO [53] hand meshes. We finally segment hands and objects from images and automatically inpaint markers to acquire realistic object appearances. + +![](images/2013a814c6edacde0454d319f62960316980a8faf73df327c951e8b82baef3b7.jpg) +Figure 4. An example of marker removal. Three sub-figures respectively show the captured image patch, the automatically-computed marker mask, and the inpainted image patch. + +![](images/310ba87860ef953b6eb8801108bd2ed82d4e8c8d43af1b11a8aaed368c386f3e.jpg) + +![](images/ab0a128c12b6087450c47200f423eb4228ba49c866097d4e73397ce1ad86a2cb.jpg) + +and target objects, with each tool and target object being utilized to perform up to 7 different kinds of actions. We formulate a behavior as using a tool category to perform an action on a target object category and divide our data into $131 <$ tool category, action label, target object category $\rightharpoondown$ triplets. Some correlated triplets are exemplified in Figure 5. To enable various studies, TACO presents high-resolution color images from 12 allocentric views and RGBD images from one egocentric-view. + +# 4. Data Quality Evaluation + +Qualitative contact optimization evaluation. Figure 6 depicts two frames of the TACO dataset. As shown in the figure, the attraction loss $\mathcal{L}_a$ promotes contact between the hands and the objects but may exacerbate penetration. The penetration loss $\mathcal{L}_p$ can mitigate penetration but does not address cases where the hands and the objects are not in contact. Our method combines both attraction loss and penetration loss, striking a balance between encouraging contact and preventing penetration. + +Quantitative hand pose evaluation. To quantitatively examine the quality and value of our hand poses, we conduct a cross-dataset evaluation between TACO and an existing high-quality HOI dataset DexYCB [4]. We select a + +![](images/d1fa6c2bf2b6a49e434c0775b0b8dda0359609cf8149b7cd2b4c5af01a88fe52.jpg) +Figure 5. Examples of interaction triplets in TACO. The left, middle, and right columns exemplify categories of tool, action, and target object, respectively. Triplets in our dataset are represented by connected paths from tools to target objects. + +![](images/2b711509c1a6053a50a41ff9401d072f8b6db4eab9a02b2f859a874593e6c1fa.jpg) +Figure 6. Qualitative contact optimization evaluation. From left to right: original RGB image, optimization without attraction loss, optimization without penetration loss, our method with attraction loss and penetration loss. + +multi-stage method, CMR [5], and a lightweight network, MobRecon [6], both of which can recover occluded hand + +
MethodTrain TestTACODexYCBTACO + DexYCB
CMRTACO10.73 / 30.8713.53 / 39.5410.25 / 28.36
DexYCB19.36 / 76.206.51 / 13.926.44 / 13.62
MobReconTACO9.03 / 20.7612.95 / 34.328.69 / 20.14
DexYCB15.17 / 44.736.51 / 13.616.32 / 12.94
+ +Table 2. Cross-dataset evaluation with DexYCB [4] on 3D hand pose estimation. Results are in PA-MPJPE (mm, lower is better) and MPJPE (mm, lower is better), respectively. + +meshes from a single-view RGB image captured in HOI scenarios, and then train and test their networks under different data combinations. For DexYCB, we follow its default train/test split (S0). For TACO, we randomly select 60 motion sequences for training and 40 for testing and crop the right hand from each image to align with the single-hand HO3D. Following existing hand mesh recovery methods [5, 6, 45, 67], we select Procrustes-Aligned Mean Per Joint Position Error (PA-MPJPE) and Mean Per Joint Position Error (MPJPE) as evaluation metrics. More details about the metrics can be found in the paper [6]. Table 2 shows the performance of the two methods under different settings. Due to the large gap in hand-object motion between the two datasets, the performances for the two methods would both drop rapidly when the training and the test sets are from different datasets. Nevertheless, combining training data from TACO and DexYCB achieves significant performance gains on both datasets for both the two methods, which demonstrates the accuracy of our hand pose annotations and indicates that TACO is a beneficial data complement for 3D hand pose estimation. + +Marker removal quality evaluation. We conduct a novel experiment to examine the reality of our marker-removed images. The key idea is that a network can easily segment markers from the image while being hard to figure out the same regions when markers are replaced by the original object surface appearances. We thus use the originally captured image patches and the marker-removed ones to train a 2D marker segmentation network, respectively, and then apply the two networks to novel image patches with the same image processing setting as training. In practice, spheres larger than markers (as shown in Figure 4(b)) are required to be segmented, since the marker removal process exactly inpaints these regions. We select U-Net [54] as the network backbone and train the two networks with an L2-loss measuring differences between mask predictions and ground truths. When training and testing on raw image patches, U-Net achieves $63.8\%$ mean Intersection over Union (mIoU), while its performance sharply drops to $11.1\%$ mIoU after applying marker removal, indicating the reality and naturalness of the inpainted image regions. Figure 7 further compares heatmaps of network predictions under the two settings, exhibiting ambiguities of marker segmentation posed by realistic inpainted images. More details are provided in the supplementary material. + +![](images/c0d87809ba63d4c78231e9ac99541f59ac0bea7fc55e291032bcd1954412c4b9.jpg) + +![](images/b1a76b24aa839527c3380d75fb4f3807189169d671783cd7c08832ab47332bcc.jpg) + +![](images/270ddb091364333c9892f5f5c7d893d13bfc6d6cf4b9400f9b84c99f984239ac.jpg) +(a) + +![](images/a57fa7f153f47d85c7f8be2cfa1e2ee9e9da703d0e44513408e3fa586dc2a2f5.jpg) +Figure 7. Marker removal quality evaluation. We train and test U-Net [54] on raw captured images (a) and marker-removed images (b), respectively. In each experiment, the left and right columns show the input image patches and the predicted segmentation heatmaps, respectively. + +![](images/0a966003a09c3e49348500b8bd9308f95f874d96003eef769769d004027f2117.jpg) + +![](images/db3f0b10190b5f10903a553e4ea8dfdbafbae201e853c38b2c545c8c182c8f2f.jpg) +(b) + +![](images/d6d1e87c980527b470b33011076f8cd6ff8b2924cb43d3acd7aa5f85eaa7e6ce.jpg) + +![](images/7bf980a1437b8e43178c59bc83a3dc6b1eba837dd5a9d4126079ad9f105bd050.jpg) + +# 5. Experiments + +In this section, we first introduce our data split (Section 5.1) supporting different research purposes about generalization. We then present three benchmarks on TACO: compositional action recognition (Section 5.2), generalizable hand-object motion forecasting (Section 5.3), and cooperative grasp synthesis (Section 5.4). Details of evaluation metrics are provided in supplementary materials. + +# 5.1. Data Split + +With diverse object geometries and interaction triplets, TACO focuses on supporting generalizable studies that apply to unseen object geometries or novel interaction triplets. Hence, we carefully split our test data into four subsets with different generalization purposes: + +- Test set 1 (S1): No generalization. Tool geometries and interaction triplets are both involved in the training set. +- Test set 2 (S2): Geometry-level generalization. The tool geometry is novel, while the interaction triplet appears in the training set. +- Test set 3 (S3): Interaction-level generalization. The interaction triplet is novel, while the tool categories and geometries are included in the training set. +- Test set 4 (S4): Compound generalization. The tool category is novel, leading to unseen geometries and triplets. + +The training set and the four test sets follow a data ratio of 4:1:1:1.5:2.5. + +# 5.2. Compositional Action Recognition + +Humans can recognize the action disentangled from their context and appearance biases of the objects, which is referred to as compositionality. To test a system's compositional generalization capability, with a diverse array of tool-action-object triplets in TACO, we benchmark compositional action recognition aiming at identifying actions + +
Test SetMethodTop-1 (%, ↑)Top-5 (%, ↑)
S1AIM [73]83.0898.85
CACNF [50]86.1599.62
S2AIM [73]82.8198.83
CACNF [50]77.3496.88
S3AIM [73]53.6581.25
CACNF [50]63.0292.97
S4AIM [73]39.3373.67
CACNF [50]44.0080.50
+ +Table 3. Results on compositional action recognition. Methods are examined via Top-1 and Top-5 accuracy. + +in HOI videos. In contrast to traditional action recognition tasks [11, 16, 56, 58], our focus is on scenarios where the tool-action-object triplets in the test set are unseen during training, including novel tool types or geometries. This poses additional challenges for evaluating model generalization and achieving human-like perception. + +Problem formulation: In a bimanual HOI scenario, our goal is to identify action labels from first-person RGB video frames, with optional bounding boxes for hands and objects computed based on their poses. + +Evaluation metrics: Following [50, 73], we use Top-1 Accuracy and Top-5 Accuracy to evaluate the efficacy of action recognition. + +Baselines, results, and analysis: We compare two baselines to evaluate the compositional generalization capability: AIM [73], an adaptation of pre-trained image transformer models, representing state-of-the-art traditional action recognition approaches; and CACNF [50], which specializes in compositional and few-shot action recognition. Table 3 shows the results. A significant decrease in accuracy on the most difficult S4 set, compared to high accuracy on the non-generalizable S1 set, underscores compositional generalization challenges. Notably, the S2 set, with novel tool geometries within tool categories, performed similarly to S1, while the S3 set, featuring novel tool-action-object triplets, showed a significant accuracy drop, revealing that while action recognition models inherently possess basic generalizability to geometric changes, they struggle with novel triplets due to the unfamiliarity of interaction contexts. When comparing among baselines, CACNF outperformed AIM in challenging sets S3 and S4, benefiting from its bounding box-based approach that effectively disentangles actions from object and tool geometries, highlighting the advantages of a compositional approach in generalizable cases. All these results underscore the significant impact of compositional generalization on action recognition. + +# 5.3. Generalizable Hand-object Motion Forecasting + +With abundant 3D hand-object mesh annotations provided by TACO, we benchmark generalizable interaction forecasting aiming to predict the following hand-object motions from seen short motion clips. Unlike the prediction of + +low-frequency human motions [42, 46, 47], we observe that hands commonly achieve a complete manipulation behavior (e.g. pouring all the water out of a bowl) in a very short time, making forecasting both interesting and challenging. + +Problem formulation: In a bimanual HOI scenario, given object point clouds and poses of two hands, the tool, and the target object in consecutive $N$ frames, our goal is to forecast their poses in subsequent $M$ frames. We select $N = 10$ and $M = 10$ in our dataset. + +Evaluation metrics: Following human-object forecasting evaluations [1, 10, 63, 68], we use Mean Per Joint Position Error $J_{e}$ to evaluate predicted left and right-hand skeletons, and leverage translation error $T_{e}$ and rotation error $R_{e}$ to measure object positions and orientations, respectively. + +Baselines, results, and analysis: Due to the lack of hand-object interaction forecasting solutions, we transfer methodologies from human-object motion forecasting as baselines. To cover diverse method designs and possibilities, we select two generative models [62, 68] and two predictive models [10, 68] for comparison. Table 4 compares their performance under different generalization settings. We observed that the tool and the hand holding it play dominant roles during manipulation, making their motion forecasting more challenging compared to others. The generative models, in comparison to predictive ones, yield significantly larger errors for both hands and the tool. One possible reason is that their motion patterns are fast and complex, making the modeling of motion distribution challenging. Enhancing the accuracy of generative methods in hand-object manipulation scenarios is an interesting future direction. When comparing method performances between a non-generalizable set (S1) and sets involving generalization (S2-4), we note that methods consistently exhibit significant performance declines with the right hand and the tool, regardless of whether applied to novel tool geometries (S2, S4) or interaction triplets (S3, S4), meanwhile achieving close results on the others. This trend indicates substantial challenges in generalizing dominant interaction entities. + +# 5.4. Cooperative Grasp Synthesis + +Generating human-like hand grasps benefits various applications in VR/AR and dexterous manipulation. Existing grasp synthesis approaches [28, 30, 38, 81] mainly focus on creating stable grasps on static objects without considering interactive purposes. Taking advantage of extensive interaction behaviors from TACO, we benchmark a novel task that aims to synthesize realistic and physically plausible hand grasps in HOI scenarios. Besides achieving stable grasping for the object, the method must comprehend other interactive objects and human hands to generate cooperative motions and avoid conflicts. We evaluate this task on our four test sets, each targeting different aspects of generalization. + +Problem formulation: Consider using the right hand to + +
Test SetMethodJe(mm,↓)Te(mm,↓)Re(°,↓)
Right / LeftTool / TargetTool / Target
S1InterVAE [68]54.9 / 48.955.0 / 15.866.94 / 6.63
MDM [62]61.2 / 53.749.4 / 14.952.81 / 5.85
InterRNN [68]29.3 / 20.425.8 / 11.610.08 / 5.22
CAHMP [10]28.8 / 22.123.9 / 12.310.24 / 4.95
S2InterVAE [68]58.8 / 50.854.3 / 12.288.19 / 5.24
MDM [62]65.8 / 55.848.0 / 10.175.78 / 3.59
InterRNN [68]35.2 / 23.131.0 / 8.911.09 / 4.03
CAHMP [10]31.4 / 22.624.7 / 8.610.65 / 3.32
S3InterVAE [68]56.5 / 50.156.7 / 12.768.03 / 5.69
MDM [62]66.4 / 58.046.2 / 11.561.90 / 4.14
InterRNN [68]32.5 / 22.028.2 / 9.711.42 / 4.32
CAHMP [10]29.3 / 21.624.4 / 9.411.18 / 3.80
S4InterVAE [68]63.9 / 54.263.0 / 14.571.38 / 5.01
MDM [62]70.5 / 57.950.2 / 12.270.25 / 3.55
InterRNN [68]36.6 / 22.330.9 / 10.611.92 / 3.86
CAHMP [10]32.9 / 22.226.6 / 10.811.73 / 3.24
+ +manipulate a tool and the left hand to operate a target object. During manipulation, given the meshes of the two objects and the left hand, the goal is to generate a right-hand mesh grasping the tool in a manner conducive to the interaction. + +Evaluation metrics: Following static grasp synthesis methods [28, 29, 38], we use interpenetration volume (Pen. $V$ ) and contact ratio (Con. $R$ ) to measure the physical plausibility of generated hands. To ensure the hand is appropriately interactive, we introduce a new metric named collision ratio (Col. $R$ ), which detects collisions between generated hands and the interaction environment. Nevertheless, to quantitatively examine whether the grasps are realistic, we design the FID score (FID) to measure distances between distributions of human grasps and synthesized ones. + +Baselines, results, and analysis: We select two recent CVAE-based approaches ContactGen [38] and HALO-VAE [30], and modify their network structures to incorporate the left hand and the target object as additional VAE conditions. To assess the significance of interaction environments, we also evaluate HALO-VAE $^{-}$ , which omits the use of environment inputs and directly generates from the tool mesh. Due to the strong correlation between hand grasps and tool geometries, we test these approaches with familiar (S1, S3) and unseen tool geometries (S2, S4). Table 5 summarizes quantitative evaluations of the two settings. Notably, HALO-VAE $^{-}$ shows a significant decrease in performance on contact and collision ratios compared to HALO-VAE, underscoring the importance of perceiving and modeling interaction environments. When applied to novel object geometries and categories, all three approaches exhibit higher collision ratios and FID scores, and most of them show a marked decline in their ability to contact tools precisely and prevent penetrations. Figure 8 exemplifies + +Table 4. Results on generalizable interaction forecasting. We respectively examine predictions of the right hand, the left hand, the tool, and the target object. + +
Test SetMethodPen. V (cm3, ↓)Con. R (%, ↑)Col. R (%, ↓)FID (10-2, ↓)
S1 ∪ S3ContactGen [38]0.2780.1210.4012.18
HALO-VAE-0.3176.274.194.89
HALO-VAE [30]0.2682.862.392.02
S2 ∪ S4ContactGen [38]0.2674.9817.5622.53
HALO-VAE-0.2971.6914.255.09
HALO-VAE [30]0.3283.5211.112.84
+ +Table 5. Results on cooperative grasp synthesis. We measure the physical plausibility of generated grasps by interpenetration volume and contact ratio, meanwhile using collision ratio to examine interaction faithfulness and FID score to assess reality. + +failure cases of synthesized grasps, indicating a large room for improving physical feasibility and naturalness on complex and slender geometries. + +![](images/b3a1e009f4eed4b73751e651a84d3e8774212406a848a3919311a966c77c4946.jpg) +Figure 8. Failure cases of cooperative grasp synthesis. (a) and (b) are generated by ContactGen [38], while (c) and (d) are from HALO-VAE [30]. + +# 6. Limitations and Conclusion + +We present TACO, the first large-scale, real-world 4D bi-manual hand-object manipulation dataset. It encompasses a wide range of tool-action-object compositions and object geometries. TACO contains a total of $2.5\mathrm{K}$ motion sequences and $5.2\mathrm{M}$ video frames, captured from 12 third-person views and one egocentric view. We contribute an automatic data acquisition pipeline that accurately recovers hand-object meshes and segmentation, along with realistic hand-object appearances. Leveraging TACO's diverse data, we benchmark compositional action recognition, generalizable hand-object motion forecasting, and cooperative grasp synthesis. These benchmarks reveal new insights, challenges, and opportunities in the field of generalizable hand-object interaction studies. + +There are three major limitations in TACO. Firstly, TACO currently does not cover articulated objects. Secondly, while TACO offers an extensive exploration of object geometries and HOI behaviors, it lacks scene diversities that are also crucial for understanding human manipulations. Thirdly, our solution to marker removal is an application of generative models, hence the original object appearances cannot be recovered perfectly. + +# References + +[1] Vida Adeli, Mahsa Ehsanpour, Ian Reid, Juan Carlos Niebles, Silvio Savarese, Ehsan Adeli, and Hamid Rezatofighi. Tripod: Human trajectory and pose dynamics forecasting in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13390-13400, 2021. 7 +[2] Samarth Brahmbhatt, Chengcheng Tang, Christopher D Twigg, Charles C Kemp, and James Hays. Contactpose: A dataset of grasps with object contact and hand pose. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 361-378. Springer, 2020. 2, 3, 4 +[3] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 4 +[4] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 2, 3, 4, 5, 6 +[5] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5, 6 +[6] Xingyu Chen, Yufeng Liu, Dong Yajiao, Xiong Zhang, Chongyang Ma, Yanmin Xiong, Yuan Zhang, and Xiaoyan Guo. Mobrecon: Mobile-friendly hand mesh reconstruction from monocular image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5, 6 +[7] Sammy Christen, Muhammed Kocabas, Emre Aksan, Jemin Hwangbo, Jie Song, and Otmar Hilliges. D-grasp: Physically plausible dynamic grasp synthesis for hand-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20577-20586, 2022. 2, 3 +[8] Sammy Christen, Wei Yang, Claudia Pérez-D'Arpino, Otmar Hilliges, Dieter Fox, and Yu-Wei Chao. Learning human-to-robot handovers from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9654–9664, 2023. 2 +[9] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.4 +[10] Enric Corona, Albert Pumarola, Guillem Alenya, and Francesc Moreno-Noguer. Context-aware human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6992-7001, 2020. 7, 8 +[11] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide + +Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European conference on computer vision (ECCV), pages 720-736, 2018. 2, 7 +[12] Christian RG Dreher, Mirko Wächter, and Tamim Asfour. Learning object-action relations from bimanual human demonstration using graph networks. IEEE Robotics and Automation Letters, 5(1):187-194, 2019. 2, 3 +[13] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J Black, and Otmar Hilliges. Arctic: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12954, 2023. 2, 3, 4 +[14] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 4 +[15] Guillermo Garcia-Hernando, Shanxin Yuan, Seungryul Baek, and Tae-Kyun Kim. First-person hand action benchmark with rgb-d videos and 3d hand pose annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 409-419, 2018. 2, 3, 4 +[16] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The" something something" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 7 +[17] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2 +[18] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3196-3206, 2020. 2, 3, 4 +[19] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11090-11100, 2022. 2, 3, 4 +[20] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11807-11816, 2019. 2, 4 +[21] Roei Herzig, Elad Ben-Avraham, Karttikeya Mangalam, Amir Bar, Gal Chechik, Anna Rohrbach, Trevor Darrell, and Amir Globerson. Object-region video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3148-3159, 2022. 3 + +[22] Zhi Hou, Xiaojiang Peng, Yu Qiao, and Dacheng Tao. Visual compositional learning for human-object interaction detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 584-600. Springer, 2020. 3 +[23] Zhi Hou, Baosheng Yu, Yu Qiao, Xiaojiang Peng, and Dacheng Tao. Affordance transfer learning for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 495-504, 2021. +[24] Zhi Hou, Baosheng Yu, Yu Qiao, Xiaojiang Peng, and Dacheng Tao. Detecting human-object interaction via fabricated compositional learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14646-14655, 2021. 3 +[25] Zhi Hou, Baosheng Yu, and Dacheng Tao. Discovering human-object interaction concepts via self-compositional learning. In European Conference on Computer Vision, pages 461-478. Springer, 2022. 3 +[26] Juntao Jian, Xiuping Liu, Manyi Li, Ruizhen Hu, and Jian Liu. Affordpose: A large-scale dataset of hand-object interactions with affordance-driven hand pose. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14713-14724, 2023. 2, 3, 4 +[27] Changlong Jiang, Yang Xiao, Cunlin Wu, Mingyang Zhang, Jinghong Zheng, Zhiguo Cao, and Joey Tianyi Zhou. A2j-transformer: Anchor-to-joint transformer network for 3d interacting hand pose estimation from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8846-8855, 2023. 4 +[28] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J Black, Krikamol Muandet, and Siyu Tang. Grasping field: Learning implicit representations for human grasps. In 2020 International Conference on 3D Vision (3DV), pages 333-344. IEEE, 2020. 3, 7, 8 +[29] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J Black, Krikamol Muandet, and Siyu Tang. Grasping field: Learning implicit representations for human grasps. In 2020 International Conference on 3D Vision (3DV), pages 333-344. IEEE, 2020. 8 +[30] Korrawe Karunratanakul, Adrian Spurr, Zicong Fan, Otmar Hilliges, and Siyu Tang. A skeleton-driven neural occupancy representation for articulated hands. In 2021 International Conference on 3D Vision (3DV), pages 11-21. IEEE, 2021. 3, 7, 8 +[31] Tae Soo Kim, Jonathan Jones, and Gregory D Hager. Motion guided attention fusion to recognize interactions from videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13076-13086, 2021. 3 +[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 4 +[33] Franziska Krebs, Andre Meixner, Isabel Patzer, and Tamim Asfour. The kit bimanual manipulation dataset. In 2020 IEEE-RAS 20th International Conference on Humanoid Robots (Humanoids), pages 499-506. IEEE, 2021. 2, 3, 4 + +[34] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 2, 3, 4 +[35] Haoming Li, Xinzhuo Lin, Yang Zhou, Xiang Li, Yuchi Huo, Jiming Chen, and Qi Ye. Contact2grasp: 3d grasp synthesis via hand-object contact constraint. arXiv preprint arXiv:2210.09245, 2022. 3 +[36] Mengcheng Li, Liang An, Hongwen Zhang, Lianpeng Wu, Feng Chen, Tao Yu, and Yebin Liu. Interacting attention graph for single image two-hand reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2761-2770, 2022. 4 +[37] Qingtao Liu, Yu Cui, Zhengnan Sun, Haoming Li, Gaofeng Li, Lin Shao, Jiming Chen, and Qi Ye. Dexrepnet: Learning dexterous robotic grasping network with geometric and spatial hand-object representations. arXiv preprint arXiv:2303.09806, 2023. 3 +[38] Shaowei Liu, Yang Zhou, Jimei Yang, Saurabh Gupta, and Shenlong Wang. Contactgen: Generative contact modeling for grasp generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20609-20620, 2023. 3, 7, 8 +[39] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 2, 3, 4 +[40] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 4 +[41] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 4 +[42] Wei Mao, Richard I Hartley, Mathieu Salzmann, et al. Contact-aware human motion forecasting. Advances in Neural Information Processing Systems, 35:7356-7367, 2022. 7 +[43] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-else: Compositional action recognition with spatial-temporal interaction networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1049–1059, 2020. 3 +[44] Eley Ng, Ziang Liu, and Monroe Kennedy. It takes two: Learning to plan for human-robot cooperative carrying. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7526–7532. IEEE, 2023. 2 +[45] JoonKyu Park, Yeonguk Oh, Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Handocctnet: Occlusion-robust 3d hand mesh estimation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1496–1505, 2022. 6 + +[46] Xiaogang Peng, Yaodi Shen, Haoran Wang, Binling Nie, Yiqiang Wang, and Zizhao Wu. Somoformer: Social-aware motion transformer for multi-person motion prediction. arXiv preprint arXiv:2208.09224, 2022.7 +[47] Xiaogang Peng, Siyuan Mao, and Zizhao Wu. Trajectory-aware body interaction transformer for multi-person pose forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17121-17130, 2023. 7 +[48] Yuzhe Qin, Hao Su, and Xiaolong Wang. From one hand to multiple hands: Imitation learning for dexterous manipulation from single-camera teleoperation. IEEE Robotics and Automation Letters, 7(4):10873-10881, 2022. 3 +[49] Yuzhe Qin, Yueh-Hua Wu, Shaowei Liu, Hanwen Jiang, Ruihan Yang, Yang Fu, and Xiaolong Wang. Dexamv: Imitation learning for dexterous manipulation from human videos. In European Conference on Computer Vision, pages 570-587. Springer, 2022. 2, 3 +[50] Gorjan Radevski, Marie-Francine Moens, and Tinne Tuyte-laars. Revisiting spatio-temporal layouts for compositional action recognition. arXiv preprint arXiv:2111.01936, 2021. 3, 7 +[51] Ramanathan Rajendiran, Debaditya Roy, and Basura Fernando. Modelling spatio-temporal interactions for compositional action recognition. arXiv preprint arXiv:2305.02673, 2023. 3 +[52] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 4 +[53] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 4, 5 +[54] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 6 +[55] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 2 +[56] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. 7 +[57] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 3 +[58] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 7 +[59] Pengzhan Sun, Bo Wu, Xunsong Li, Wen Li, Lixin Duan, and Chuang Gan. Counterfactual debiasing inference for compositional action recognition. In Proceedings of the 29th + +ACM International Conference on Multimedia, pages 3220-3228, 2021. 3 +[60] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021.4 +[61] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 581-600. Springer, 2020. 3, 4 +[62] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 7, 8 +[63] Weilin Wan, Lei Yang, Lingjie Liu, Zhuoying Zhang, Ruixing Jia, Yi-King Choi, Jia Pan, Christian Theobalt, Taku Komura, and Wenping Wang. Learn to predict how humans manipulate large-sized objects from interactive motions. IEEE Robotics and Automation Letters, 7(2):4702-4709, 2022. 7 +[64] Haozhong Wang, Hua Yu, and Qiang Zhang. Detecting zero-shot human-object interaction with visual-text modeling. In 2023 9th International Conference on Virtual Reality (ICVR), pages 155-162. IEEE, 2023. 3 +[65] Suchen Wang, Yueqi Duan, Henghui Ding, Yap-Peng Tan, Kim-Hui Yap, and Junsong Yuan. Learning transferable human-object interaction detector with natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 939-948, 2022. 3 +[66] Wei Xie, Zhipeng Yu, Zimeng Zhao, Binghui Zuo, and Yanggang Wang. Hmdo: Markerless multi-view hand manipulation capture with deformable objects. Graphical Models, 127:101178, 2023. 2, 3, 4 +[67] Hao Xu, Tianyu Wang, Xiao Tang, and Chi-Wing Fu. H2onet: Hand-occlusion-and-orientation-aware network for real-time 3d hand mesh reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17048–17058, 2023. 6 +[68] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14928-14940, 2023. 7, 8 +[69] Rui Yan, Peng Huang, Xiangbo Shu, Junhao Zhang, Yonghua Pan, and Jinhui Tang. Look less think more: Rethinking compositional action recognition. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3666-3675, 2022. 3 +[70] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 4 +[71] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the + +IEEE/CVF International Conference on Computer Vision, pages 11097-11106, 2021.3 +[72] Lixin Yang, Kailin Li, Xinyu Zhan, Fei Wu, Anran Xu, Liu Liu, and Cewu Lu. Oakink: A large-scale knowledge repository for understanding hand-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20953-20962, 2022. 2, 3, 4 +[73] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. AIM: Adapting image models for efficient video action recognition. In The Eleventh International Conference on Learning Representations, 2023. 7 +[74] Jianglong Ye, Jiashun Wang, Binghao Huang, Yuzhe Qin, and Xiaolong Wang. Learning continuous grasping function with a dexterous hand from human demonstrations. IEEE Robotics and Automation Letters, 8(5):2882-2889, 2023. 2 +[75] Ruolin Ye, Wenqiang Xu, Zhendong Xue, Tutian Tang, Yanfeng Wang, and Cewu Lu. H2o: A benchmark for visual human-human object handover analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15762-15771, 2021. 2, 3, 4 +[76] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Is an object-centric video representation beneficial for transfer? In Proceedings of the Asian Conference on Computer Vision, pages 1976-1994, 2022. 3 +[77] He Zhang, Yuting Ye, Takaaki Shiratori, and Taku Komura. Manipnet: neural manipulation synthesis with a handobject spatial representation. ACM Transactions on Graphics (ToG), 40(4):1-14, 2021. 3 +[78] Hui Zhang, Sammy Christen, Zicong Fan, Luocheng Zheng, Jemin Hwangbo, Jie Song, and Otmar Hilliges. Artigrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. arXiv preprint arXiv:2309.03891, 2023. 2, 3 +[79] Juntian Zheng, Qingyuan Zheng, Lixing Fang, Yun Liu, and Li Yi. Cams: Canonicalized manipulation spaces for category-level functional hand-object manipulation synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 585-594, 2023. 3 +[80] Desen Zhou, Zhichao Liu, Jian Wang, Leshan Wang, Tao Hu, Errui Ding, and Jingdong Wang. Human-object interaction detection via disentangled transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19568-19577, 2022. 3 +[81] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Toch: Spatio-temporal object-to-hand correspondence for motion refinement. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3, 7 +[82] Zehao Zhu, Jiashun Wang, Yuzhe Qin, Deqing Sun, Varun Jampani, and Xiaolong Wang. Contactart: Learning 3d interaction priors for category-level articulated object and hand poses estimation. arXiv preprint arXiv:2305.01618, 2023. 3 \ No newline at end of file diff --git a/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/images.zip b/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6b3c972bd551b8d7bc38e971e9fc2065126b6e4b --- /dev/null +++ b/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c086ebfaa86b3967159f8bff8bb537dad31466b997863a699cdb9c14447bbaa +size 573736 diff --git a/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/layout.json b/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a36d35368bc8c9e081f8b1bdaf3b1dcf4361115c --- /dev/null +++ b/2024/TACO_ Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding/layout.json @@ -0,0 +1,9508 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 103, + 445, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 103, + 445, + 140 + ], + "spans": [ + { + "bbox": [ + 149, + 103, + 445, + 140 + ], + "type": "text", + "content": "TACO: Benchmarking Generalizable Bimanual Tool-ACTION-Object Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "text", + "content": "Yun Liu" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "text", + "content": ", Haolin Yang" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "text", + "content": ", Xu Si" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "text", + "content": ", Ling Liu" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "text", + "content": ", Zipeng Li" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "text", + "content": ", Yuxiang Zhang" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "text", + "content": ", Yebin Liu" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "text", + "content": ", Li Yi" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "inline_equation", + "content": "^{\\dagger,1,2,3}" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory " + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "text", + "content": "Shanghai Qi Zhi Institute " + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "text", + "content": "Beijing University of Posts and Telecommunications " + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 53, + 159, + 548, + 208 + ], + "type": "text", + "content": "Beijing Institute of Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 234, + 212, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 234, + 212, + 241 + ], + "spans": [ + { + "bbox": [ + 129, + 234, + 212, + 241 + ], + "type": "text", + "content": "Third-person View Video" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 49, + 243, + 272, + 338 + ], + "blocks": [ + { + "bbox": [ + 49, + 243, + 272, + 338 + ], + "lines": [ + { + "bbox": [ + 49, + 243, + 272, + 338 + ], + "spans": [ + { + "bbox": [ + 49, + 243, + 272, + 338 + ], + "type": "image", + "image_path": "a7095185afdc72fa1d218f9b36580ec57d97c848f52a363f324c5e0accba8c0b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 338, + 186, + 346 + ], + "lines": [ + { + "bbox": [ + 129, + 338, + 186, + 346 + ], + "spans": [ + { + "bbox": [ + 129, + 338, + 186, + 346 + ], + "type": "text", + "content": "Egocentric Video" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 331, + 238, + 389, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 238, + 389, + 247 + ], + "spans": [ + { + "bbox": [ + 331, + 238, + 389, + 247 + ], + "type": "text", + "content": "Hand-Object Mesh" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 279, + 251, + 430, + 284 + ], + "blocks": [ + { + "bbox": [ + 279, + 251, + 430, + 284 + ], + "lines": [ + { + "bbox": [ + 279, + 251, + 430, + 284 + ], + "spans": [ + { + "bbox": [ + 279, + 251, + 430, + 284 + ], + "type": "image", + "image_path": "70734ec12711cbd502686f0cf04fcbe37fe476736f6157b07689d585b57ba2f8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 288, + 289, + 323, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 289, + 323, + 296 + ], + "spans": [ + { + "bbox": [ + 288, + 289, + 323, + 296 + ], + "type": "text", + "content": "Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 328, + 289, + 384, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 289, + 384, + 296 + ], + "spans": [ + { + "bbox": [ + 328, + 289, + 384, + 296 + ], + "type": "text", + "content": "Realistic Appearance" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 389, + 289, + 432, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 289, + 432, + 296 + ], + "spans": [ + { + "bbox": [ + 389, + 289, + 432, + 296 + ], + "type": "text", + "content": "Behavior Triplet" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 278, + 297, + 336, + 343 + ], + "blocks": [ + { + "bbox": [ + 278, + 297, + 336, + 343 + ], + "lines": [ + { + "bbox": [ + 278, + 297, + 336, + 343 + ], + "spans": [ + { + "bbox": [ + 278, + 297, + 336, + 343 + ], + "type": "image", + "image_path": "52d6072e2a8a17c8a867f3635b03989e0bfbf803bd218696b52e03321f768a38.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 350, + 342, + 361 + ], + "lines": [ + { + "bbox": [ + 257, + 350, + 342, + 361 + ], + "spans": [ + { + "bbox": [ + 257, + 350, + 342, + 361 + ], + "type": "text", + "content": "Motion Forecasting" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 341, + 298, + 384, + 343 + ], + "blocks": [ + { + "bbox": [ + 341, + 298, + 384, + 343 + ], + "lines": [ + { + "bbox": [ + 341, + 298, + 384, + 343 + ], + "spans": [ + { + "bbox": [ + 341, + 298, + 384, + 343 + ], + "type": "image", + "image_path": "44766966b467364b968a1578dcc6d7def426b5a760bf69748433cdc3ab198132.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 456, + 234, + 541, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 456, + 234, + 541, + 241 + ], + "spans": [ + { + "bbox": [ + 456, + 234, + 541, + 241 + ], + "type": "text", + "content": "Test Time Generalization" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 441, + 250, + 497, + 291 + ], + "blocks": [ + { + "bbox": [ + 441, + 250, + 497, + 291 + ], + "lines": [ + { + "bbox": [ + 441, + 250, + 497, + 291 + ], + "spans": [ + { + "bbox": [ + 441, + 250, + 497, + 291 + ], + "type": "image", + "image_path": "2148b55abd1a66d0860a6e434ab73586ad739583895c9b239ea9211b5485c30c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 441, + 297, + 497, + 338 + ], + "blocks": [ + { + "bbox": [ + 441, + 297, + 497, + 338 + ], + "lines": [ + { + "bbox": [ + 441, + 297, + 497, + 338 + ], + "spans": [ + { + "bbox": [ + 441, + 297, + 497, + 338 + ], + "type": "image", + "image_path": "2cf661d8a44082cdf5294d66548ebb37f4d402dc4ebbd69e428c2091098fec5b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 504, + 248, + 541, + 289 + ], + "blocks": [ + { + "bbox": [ + 504, + 248, + 541, + 289 + ], + "lines": [ + { + "bbox": [ + 504, + 248, + 541, + 289 + ], + "spans": [ + { + "bbox": [ + 504, + 248, + 541, + 289 + ], + "type": "image", + "image_path": "3c00496f4791b5691aef890c75cd37ebf498446e9bc38296260cf8f20cc34a12.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 506, + 291, + 540, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 506, + 291, + 540, + 298 + ], + "spans": [ + { + "bbox": [ + 506, + 291, + 540, + 298 + ], + "type": "text", + "content": "Novel Triplet" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 504, + 302, + 542, + 345 + ], + "blocks": [ + { + "bbox": [ + 504, + 302, + 542, + 345 + ], + "lines": [ + { + "bbox": [ + 504, + 302, + 542, + 345 + ], + "spans": [ + { + "bbox": [ + 504, + 302, + 542, + 345 + ], + "type": "image", + "image_path": "fc06444b2e7b7b9ec0018992f2cb7b04869f3f549acccc2aac86b90e40eae227.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 350, + 526, + 361 + ], + "lines": [ + { + "bbox": [ + 398, + 350, + 526, + 361 + ], + "spans": [ + { + "bbox": [ + 398, + 350, + 526, + 361 + ], + "type": "text", + "content": "Cooperative Grasp Synthesis" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 58, + 373, + 152, + 447 + ], + "blocks": [ + { + "bbox": [ + 86, + 350, + 172, + 360 + ], + "lines": [ + { + "bbox": [ + 86, + 350, + 172, + 360 + ], + "spans": [ + { + "bbox": [ + 86, + 350, + 172, + 360 + ], + "type": "text", + "content": "Action Recognition" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 373, + 152, + 447 + ], + "lines": [ + { + "bbox": [ + 58, + 373, + 152, + 447 + ], + "spans": [ + { + "bbox": [ + 58, + 373, + 152, + 447 + ], + "type": "image", + "image_path": "9e70c1491c30e0d53f4ed2f11665e522b1a0ca9b8b0533c3e42d074a3e92a835.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 458, + 547, + 515 + ], + "lines": [ + { + "bbox": [ + 46, + 458, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 547, + 515 + ], + "type": "text", + "content": "Figure 1. TACO is a large-scale bimanual hand-object manipulation dataset covering extensive tool-action-object combinations in real-world scenarios. It involves videos from both 12 third-person views and one egocentric view, together with precise hand-object meshes, 2D segmentation, realistic hand-object appearances, and behavior triplet annotations. With rich diversities of object shapes and interaction behaviors, TACO supports test-time generalization to unseen object geometries and novel behavior triplets and benchmarks various generalizable research topics, e.g., action recognition, motion forecasting, and cooperative grasp synthesis." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 153, + 376, + 197, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 376, + 197, + 383 + ], + "spans": [ + { + "bbox": [ + 153, + 376, + 197, + 383 + ], + "type": "text", + "content": "brush something" + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 157, + 399, + 195, + 430 + ], + "blocks": [ + { + "bbox": [ + 157, + 399, + 195, + 430 + ], + "lines": [ + { + "bbox": [ + 157, + 399, + 195, + 430 + ], + "spans": [ + { + "bbox": [ + 157, + 399, + 195, + 430 + ], + "type": "image", + "image_path": "fe7f90247c143cc264e278f8f08604b69edd2e2f50e2aa125d834e63a7ba8cdb.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 220, + 383, + 371, + 440 + ], + "blocks": [ + { + "bbox": [ + 220, + 383, + 371, + 440 + ], + "lines": [ + { + "bbox": [ + 220, + 383, + 371, + 440 + ], + "spans": [ + { + "bbox": [ + 220, + 383, + 371, + 440 + ], + "type": "image", + "image_path": "1662622d9486b0169786fcbaa9d3fad799678c820f01fa854b9ee8b35ed7d85d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 398, + 379, + 458, + 436 + ], + "blocks": [ + { + "bbox": [ + 398, + 379, + 458, + 436 + ], + "lines": [ + { + "bbox": [ + 398, + 379, + 458, + 436 + ], + "spans": [ + { + "bbox": [ + 398, + 379, + 458, + 436 + ], + "type": "image", + "image_path": "ae181eba02a4fe7cb58fc6f2bf2d3bbba23318160dbf8e1951a348278547f20f.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 476, + 377, + 522, + 435 + ], + "blocks": [ + { + "bbox": [ + 476, + 377, + 522, + 435 + ], + "lines": [ + { + "bbox": [ + 476, + 377, + 522, + 435 + ], + "spans": [ + { + "bbox": [ + 476, + 377, + 522, + 435 + ], + "type": "image", + "image_path": "9de51667526a0f7424c1bc7b0f06612d28265765f6604c1298c1ec96b9af9bee.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 143, + 523, + 192, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 523, + 192, + 536 + ], + "spans": [ + { + "bbox": [ + 143, + 523, + 192, + 536 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 549, + 289, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 289, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 289, + 693 + ], + "type": "text", + "content": "Humans commonly work with multiple objects in daily life and can intuitively transfer manipulation skills to novel objects by understanding object functional regularities. However, existing technical approaches for analyzing and synthesizing hand-object manipulation are mostly limited to handling a single hand and object due to the lack of data support. To address this, we construct TACO, an extensive bimanual hand-object-interaction dataset spanning a large variety of tool-action-object compositions for daily human activities. TACO contains 2.5K motion sequences paired with third-person and egocentric views, precise hand-object 3D meshes, and action labels. To rapidly" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 525, + 547, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 547, + 657 + ], + "type": "text", + "content": "expand the data scale, we present a fully automatic data acquisition pipeline combining multi-view sensing with an optical motion capture system. With the vast research fields provided by TACO, we benchmark three generalizable hand-object-interaction tasks: compositional action recognition, generalizable hand-object motion forecasting, and cooperative grasp synthesis. Extensive experiments reveal new insights, challenges, and opportunities for advancing the studies of generalizable hand-object motion analysis and synthesis. Our data and code are available at https://taco2024.github.io." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21740" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 286, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 286, + 258 + ], + "type": "text", + "content": "In our everyday lives, humans effortlessly synchronize the movements of both hands to manipulate a pair of objects, such as using a spatula to stir in a pan while cooking. Usually, these objects are not symmetric in their roles, with one acting as a tool to enhance the bimanual action performed on the other object. This allows us to characterize the bimanual behaviors with tool-action-object triplets. Various triplet combinations usually incorporate intricate and differing temporal and spatial coordination among the tool, the object, and the two hands. Understanding versatile bimanual behaviors could immediately benefit numerous applications in VR/AR, human-robot interaction [8, 44], and dexterous manipulation [7, 49, 74, 78], which pose significant challenges for today's computer vision systems." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 259, + 286, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 259, + 286, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 286, + 521 + ], + "type": "text", + "content": "Tackling the above challenges requires the support of large-scale annotation-rich datasets. Existing dataset efforts [2, 4, 15, 18, 26, 39, 72] on hand-object interactions (HOI) primarily focus on unimanual actions. However, simply aggregating two unimanual actions falls short of encompassing bimanual coordination behaviors. Furthermore, some works [13, 19, 34, 66, 75] examine bimanual manipulation of a single object which can hardly extend to interpreting tool-action-object triplets. The most relevant dataset [33] currently available covers only 12 objects and fewer than 20 distinct tool-action-object triplets, severely limiting its potential to facilitate and benchmark the understanding of bimanual tool-action-object interactions that can generalize to novel object geometries or previously unseen triplets. This limitation partially stems from the challenging nature of jointly capturing two dynamic hands interacting with two objects in the real world, but further underscores the urgency of developing methodologies for generalizable bimanual HOI understanding. Therefore, we put generalization as our primary focus while exploring the curation of a large-scale bimanual HOI dataset that can both facilitate and benchmark generalizable understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 522, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 286, + 713 + ], + "type": "text", + "content": "In particular, we present TACO, a large-scale bimanual manipulation dataset encompassing a diverse array of tool-action-object compositions in real-world settings. Serving as a knowledge base of multi-object cooperation, TACO focuses on daily scenarios involving the use of tools to interact with target objects and collects 131 types of triplets across 20 object categories, 196 object instances, and 15 daily actions. We organize the dataset based on the triplets and make sure triplets have different levels of overlaps, this naturally defines the semantic distance between different motion trajectories and supports studying generalization with different extents. To expedite the expansion of our dataset, we combine the advantages of marker-based and markerless motion capture systems and present a fully automatic data acquisition pipeline that can guarantee motion quality and" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "visual data quality at the same time. For each time step, the pipeline automatically provides labels including the precise recovery of hand-object mesh and segmentation on markerless vision data. As a result, TACO comprises a total of 2.5K motion sequences and 5.2M video frames incorporating 3rd-person and egocentric views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 145, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 545, + 418 + ], + "type": "text", + "content": "Benefiting from the rich and diverse motion sequences within our dataset, we carefully benchmark three tasks aiming at generalizing common human manipulation knowledge to unseen object geometries and novel tool-action-object combinations: 1) compositional action recognition, 2) generalizable hand-object motion forecasting, and 3) cooperative grasp synthesis. Extensive studies are conducted on these benchmarks: First, diverse hand-object interactive motions with action annotations pave the way to learning to recognize hand-object action under novel object combinations in real-world scenarios. Additionally, the availability of 4D dynamic hand-object pose sequences allows for fine-grained hand-object motion tendency forecasting. Finally, we provide contact-rich hand-object meshes to facilitate studying hand grasp synthesis for unseen object geometries and categories under interaction scenarios. Extensive experiments have been conducted on these benchmarks, revealing generalization challenges faced by existing methodologies when dealing with novel object geometries, object categories, and interaction triplets. We anticipate that TACO will offer more opportunities for researching and developing universal generalization strategies in the understanding and creation of hand-object interactions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 420, + 520, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 420, + 520, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 420, + 520, + 431 + ], + "type": "text", + "content": "In summary, our main contributions are threefold:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 434, + 545, + 555 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 434, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 434, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 545, + 470 + ], + "type": "text", + "content": "- We present TACO, the first large-scale real-world 4D bi-manual hand-object-interaction dataset covering diverse tool-action-object compositions and object geometries." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 471, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 471, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 471, + 545, + 506 + ], + "type": "text", + "content": "- We develop an automatic data acquisition pipeline that provides precise recovery of hand-object mesh and segmentation together with markerless vision data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 506, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 545, + 555 + ], + "type": "text", + "content": "- We benchmark three tasks toward generalizable hand-object motion analysis and synthesis. We provide a comprehensive discussion and highlight the new challenges posed by TACO." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 566, + 392, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 566, + 392, + 578 + ], + "spans": [ + { + "bbox": [ + 306, + 566, + 392, + 578 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 586, + 484, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 484, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 484, + 599 + ], + "type": "text", + "content": "2.1. Hand-object Interaction Datasets" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Understanding hand-object interaction is an emerging research topic that is supported by a large number of datasets. Many widely-used video datasets [11, 12, 17, 55] capture 2D real-world hand-object interaction videos to facilitate traditional visual perception studies such as action recognition and HOI detection. With the rapid development of 3D computer vision, recent advances have incorporated 3D hand-object mesh annotations into data-capturing processes and presented various 3D [20, 26] or" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21741" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "4D [2, 4, 12, 13, 15, 18, 19, 34, 39, 61, 66, 72, 75] datasets with different focuses. Table 1 briefly summarizes the characteristics of these datasets. A majority of existing datasets focus on single-hand grasping, a basic manipulation skill that can benefit existing dexterous robot learning [7, 48, 49, 78]. Beyond grasping, a line of datasets captures diverse and complex manipulation skills requiring humans to perform complicated, multi-step actions like using tools [13, 33, 39], deforming soft bodies [66], and multi-person handovers [75]. From the perspective of object behaviors, KIT [33], HOI4D [39], and AffordPose[26] capture functional manipulations aiming at showing actual functions and usages of objects. KIT further collects multi-object interaction behaviors, however, it lacks object diversities and is hard to support generalizable studies. Our dataset captures diverse bimanual functional manipulation behaviors for multi-object cooperation, serving a wide range of research directions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 297, + 277, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 277, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 277, + 309 + ], + "type": "text", + "content": "2.2. Generalizable Hand-object Motion Analysis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 315, + 289, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 289, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 289, + 567 + ], + "type": "text", + "content": "Generalizable hand-object-interaction motion analysis aims to understand the attributes of hand-object motion when manipulating novel object instances. To precisely recognize hand actions during the manipulation of unseen objects, a series of compositional action recognition approaches [21, 31, 43, 50, 51, 59, 69, 76] represent objects through their bounding boxes to prevent the network from overfitting to intricate object geometries. To detect and infer action-object pairs under few-shot or zero-shot settings, a prevailing methodology used in recent generalizable HOI detection methods [22-24, 64, 65, 80] decomposes actions and objects into distinct features. This decomposition allows leveraging abundant training data with either similar actions or similar objects to handle rare action-object pairs that emerge at test time. It also provides the potential to discover novel and reasonable action-object pairs [25]. To estimate hand and object poses without object geometries, Zhu et al. [82] propose learning hand-object interaction priors for each object category and transferring them to unseen objects within the same category during test time, enhancing the method adaptability to diverse unseen object instances." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 574, + 280, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 280, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 280, + 588 + ], + "type": "text", + "content": "2.3. Generalizable Hand-object Motion Synthesis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Synthesizing realistic hand-object-interaction motion on novel object geometries and categories is an emerging research topic that includes two main challenges: how to model generic, realistic, and diverse hand motion patterns from existing HOI data, and how to apply them to novel object instances. To address the first challenge, a branch of data-driven approaches decomposes complicated hand-object contact into spatial relations between each hand joint [30, 37, 38, 79] or hand surface point [71, 77, 81] and the object's relevant closest local regions, and such" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 253 + ], + "type": "text", + "content": "relations can be formulated as a combination of whether contact occurs [37, 38, 71, 77, 79, 81], hand-object distances [37, 71, 77, 81], positions [37, 38, 77, 79, 81] and directions [37, 38, 79] of contact points, and whether contact points should be closer [71]. Another line of work [28-30, 35] encodes hand-object interaction integrally to an implicit global feature by a conditional variational autoencoder (CVAE) [57] structure. As solutions to the second challenge, generalizable methods [28, 30, 35, 38, 79] decompose the object's geometry from HOI and encode it as a condition for generative models. CAMS [79] further maps objects into a canonicalized object space to decrease the gap in object geometry. One limitation of existing approaches is that they do not focus on multi-hand and multi-object cooperation due to the lack of data support." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 262, + 425, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 425, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 425, + 277 + ], + "type": "text", + "content": "3. Constructing TACO" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 283, + 547, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 547, + 319 + ], + "type": "text", + "content": "In this section, we respectively describe the data capturing system (Section 3.1), the data annotating pipeline (Section 3.2), and dataset statistics (Section 3.3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 327, + 404, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 327, + 404, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 327, + 404, + 340 + ], + "type": "text", + "content": "3.1. Data Capturing" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 356, + 545, + 476 + ], + "blocks": [ + { + "bbox": [ + 307, + 356, + 545, + 476 + ], + "lines": [ + { + "bbox": [ + 307, + 356, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 356, + 545, + 476 + ], + "type": "image", + "image_path": "b71dd1aa85185fdfa4fcd91289d6f36f3fae586aafb9b0f3617a0f282330aa39.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 482, + 518, + 494 + ], + "lines": [ + { + "bbox": [ + 331, + 482, + 518, + 494 + ], + "spans": [ + { + "bbox": [ + 331, + 482, + 518, + 494 + ], + "type": "text", + "content": "Figure 2. Data capturing system and camera views." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "text", + "content": "We obtain hand motion from multi-view RGB videos while capturing the object motion by attaching four markers on the object surface and tracking them with a mocap system. To this end, as shown in Figure 2, our data capturing system incorporates 12 synchronized industrial FLIR cameras and a NOKOV optical motion capture suite with 6 infrared Mars4H cameras. To capture egocentric RGBD videos, a helmet equipped with a Realsense L515 camera is worn by the actor. Similar to objects, the motion of the L515 camera is tracked by the mocap system. The frequency of our cameras and mocap system is " + }, + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "inline_equation", + "content": "30\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "text", + "content": ". The resolutions of our allocentric and egocentric images are " + }, + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "inline_equation", + "content": "4096\\times 3000" + }, + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "inline_equation", + "content": "1920\\times 1080" + }, + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Object model acquisition. To capture contacts between hands and objects and support relevant studies, we obtain an accurate 3D mesh for each object with an industrial EinScan 3D scanner. Each object mesh is represented by up to 100K triangular faces for fine-grained geometries." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21742" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 533, + 234 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 533, + 234 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 533, + 234 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 533, + 234 + ], + "type": "table", + "html": "
datasetbimanualmulti-object cooperationfunctional manipulationdata characteristics: category-levelegocentricmulti-viewmarkerlessmocapaction label# number of: sequenceframe
FHPA [15]✓✓✓1.2K105K
ObMan [20]-154K
ContactPose[2]✓✓2.3K3.0M
GRAB [61]✓✓1.3K1.6M
HO3D [18]✓✓2778K
KIT Binamual Manipulation [33]✓✓✓✓✓✓✓5881.6M
H2O [34]✓✓✓✓191571K
DexYCB [4]✓✓1.0K582K
H2O-3D [19]✓✓✓1776K
H2O for handover [75]✓✓✓6.0K5.0M
OakInk [72]✓✓778314K
HOI4D [39]✓✓✓✓✓4.0K1.2M
HMDO [66]✓✓✓1221K
ARCTIC [13]✓✓✓✓3392.1M
AffordPose [26]✓✓✓-27K
TACO (ours)✓✓✓✓✓✓✓✓✓2.5K5.2M
", + "image_path": "c5aa334f42663442cb3b45e11d083a231403f310a4cbc4e65b601b7bbbe425ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 149, + 236, + 443, + 246 + ], + "lines": [ + { + "bbox": [ + 149, + 236, + 443, + 246 + ], + "spans": [ + { + "bbox": [ + 149, + 236, + 443, + 246 + ], + "type": "text", + "content": "Table 1. Comparison of TACO with existing 3D hand-object interaction datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 253, + 149, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 149, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 149, + 266 + ], + "type": "text", + "content": "3.2. Data Annotating" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 271, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 271, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 287, + 331 + ], + "type": "text", + "content": "Figure 3 depicts our automatic data acquisition pipeline. Given color images, 3D marker positions, and object models as inputs, the pipeline successively performs object pose optimization, hand keypoint localization, hand pose optimization, hand-object segmentation, and marker removal." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 331, + 287, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 438 + ], + "type": "text", + "content": "Object pose optimization. Since our objects are rigid bodies, their motions are 6-dimensional and can be tracked by capturing marker positions using the mocap system. To accurately associate markers with the object, we draw inspiration from Mosh++ [41] and optimize marker-to-surface correspondence to encourage contact between markers and object surfaces while avoiding penetration. The object poses are obtained by combining marker positions relative to the object mesh and the captured marker motions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 439, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 439, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 287, + 582 + ], + "type": "text", + "content": "3D hand keypoint localization. Leveraging multi-view sensing, we first estimate 2D hand keypoints from each camera view, respectively, and then fuse them into 3D. For 2D hand keypoint estimation, directly applying existing dual-hand approaches [19, 27, 36] could fail when severe occlusion occurs or two hands are far apart. We thus follow [3, 40] and design a method to separately detect two hands using pre-trained YOLOv3 [52] and then utilize MMPose [9] to obtain single-hand 2D keypoints. To fuse multi-view 2D keypoints to 3D under severe occlusion, we use RANSAC [14] to filter out imprecise 2D keypoints and localize 3D keypoints by triangulation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 582, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 666 + ], + "type": "text", + "content": "Contact-aware hand pose optimization. We adopt MANO [53] to formulate a 3D hand mesh as " + }, + { + "bbox": [ + 47, + 582, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\Theta_h = \\{\\theta, \\beta, t\\}" + }, + { + "bbox": [ + 47, + 582, + 287, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 582, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^{48}" + }, + { + "bbox": [ + 47, + 582, + 287, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 582, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{10}" + }, + { + "bbox": [ + 47, + 582, + 287, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 582, + 287, + 666 + ], + "type": "inline_equation", + "content": "t \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 47, + 582, + 287, + 666 + ], + "type": "text", + "content": " represent hand pose, hand shape, and wrist position, respectively. Given precomputed 3D hand keypoints and object meshes, we optimize MANO hand meshes by minimizing the following loss function:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 674, + 287, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 674, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 56, + 674, + 287, + 711 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\Theta} _ {h} = \\underset {\\Theta_ {h}} {\\arg \\min } \\left(\\lambda_ {2 D} \\mathcal {L} _ {2 D} + \\lambda_ {3 D} \\mathcal {L} _ {3 D} + \\lambda_ {a n g l e} \\mathcal {L} _ {a n g l e} + \\right. \\\\ \\left. \\lambda_ {t c} \\mathcal {L} _ {t c} + \\lambda_ {p} \\mathcal {L} _ {p} + \\lambda_ {a} \\mathcal {L} _ {a}\\right), \\tag {1} \\\\ \\end{array}", + "image_path": "1c9aa3bb76b57e903181df505d6d24795c47cac47848c5146a2881bca6ec54e0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 254, + 545, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 254, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 305, + 254, + 545, + 313 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 254, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2D}" + }, + { + "bbox": [ + 305, + 254, + 545, + 313 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 254, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{3D}" + }, + { + "bbox": [ + 305, + 254, + 545, + 313 + ], + "type": "text", + "content": " induce hand meshes to get close to the keypoints, " + }, + { + "bbox": [ + 305, + 254, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{angle}" + }, + { + "bbox": [ + 305, + 254, + 545, + 313 + ], + "type": "text", + "content": " provides hand pose priors, " + }, + { + "bbox": [ + 305, + 254, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tc}" + }, + { + "bbox": [ + 305, + 254, + 545, + 313 + ], + "type": "text", + "content": " promotes temporal smoothness, " + }, + { + "bbox": [ + 305, + 254, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_a" + }, + { + "bbox": [ + 305, + 254, + 545, + 313 + ], + "type": "text", + "content": " encourages contacts, and " + }, + { + "bbox": [ + 305, + 254, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 305, + 254, + 545, + 313 + ], + "type": "text", + "content": " prevents hand-object interpenetration. Details of these loss terms are provided in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 314, + 545, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 409 + ], + "type": "text", + "content": "Hand-object segmentation. We obtain 2D hand-object masks by first rendering hand-object silhouettes from their 3D meshes. Since MANO meshes differ from real hands, we then apply the Segment Anything Model [32] to refine the masks, using sampled hand-object keypoints as prompts. To help detect hands in subsequent frames, we use Track-anything [70] to track the hands in these frames and localize them." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 410, + 546, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 546, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 546, + 625 + ], + "type": "text", + "content": "Marker removal. Despite the fact that attaching markers to objects can accurately capture object motion, markers may increase the appearance gap between captured objects and those in the wild. One direct defect is that an object pose tracking network could simply track markers rather than learn to track the target object. To improve the object's appearance and make it more realistic, after acquiring hand and object poses, we automatically remove the markers from RGB images. First, we obtain marker positions in the world coordinate system from the mocap system, and place spheres with a radius two times the marker at those positions in a simulation environment. We then render a 2D mask of the spheres for each real camera view. Benefiting from the accurate marker position estimates, such 2D masks can fully cover the markers in the original RGB images. Finally, we mask the spheres in the original RGB images and use a pre-trained LAMA [60] network to inpaint the sphere regions. Figure 4 shows a marker removal example." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 635, + 409, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 409, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 409, + 646 + ], + "type": "text", + "content": "3.3. Dataset Statistics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "TACO contains 2.5K motion sequences describing extensive daily tool-using behaviors covering 20 object categories, 196 fine-grained object 3D models, 14 participants, and 15 actions. To support cross-category generalizable HOI studies, we explore the functional diversity of tools" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21743" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 86, + 547, + 242 + ], + "blocks": [ + { + "bbox": [ + 47, + 86, + 547, + 242 + ], + "lines": [ + { + "bbox": [ + 47, + 86, + 547, + 242 + ], + "spans": [ + { + "bbox": [ + 47, + 86, + 547, + 242 + ], + "type": "image", + "image_path": "96ede0e916f3a4a2ee1b2ab0b6cca3059268fc7f3f6abfb3fba29815e032dca1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 247, + 547, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 547, + 291 + ], + "type": "text", + "content": "Figure 3. Automatic data annotating pipeline. The input consists of color frames from allocentric views, pre-scanned object models, and 3D positions of markers attached to object surfaces. We first separately localize 3D hand keypoints and obtain object poses, and then conduct contact-aware optimization to recover MANO [53] hand meshes. We finally segment hands and objects from images and automatically inpaint markers to acquire realistic object appearances." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 296, + 127, + 370 + ], + "blocks": [ + { + "bbox": [ + 48, + 296, + 127, + 370 + ], + "lines": [ + { + "bbox": [ + 48, + 296, + 127, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 296, + 127, + 370 + ], + "type": "image", + "image_path": "2013a814c6edacde0454d319f62960316980a8faf73df327c951e8b82baef3b7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 378, + 287, + 412 + ], + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 412 + ], + "type": "text", + "content": "Figure 4. An example of marker removal. Three sub-figures respectively show the captured image patch, the automatically-computed marker mask, and the inpainted image patch." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 129, + 296, + 206, + 370 + ], + "blocks": [ + { + "bbox": [ + 129, + 296, + 206, + 370 + ], + "lines": [ + { + "bbox": [ + 129, + 296, + 206, + 370 + ], + "spans": [ + { + "bbox": [ + 129, + 296, + 206, + 370 + ], + "type": "image", + "image_path": "310ba87860ef953b6eb8801108bd2ed82d4e8c8d43af1b11a8aaed368c386f3e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 208, + 296, + 287, + 370 + ], + "blocks": [ + { + "bbox": [ + 208, + 296, + 287, + 370 + ], + "lines": [ + { + "bbox": [ + 208, + 296, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 208, + 296, + 287, + 370 + ], + "type": "image", + "image_path": "ab0a128c12b6087450c47200f423eb4228ba49c866097d4e73397ce1ad86a2cb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 420, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 287, + 529 + ], + "type": "text", + "content": "and target objects, with each tool and target object being utilized to perform up to 7 different kinds of actions. We formulate a behavior as using a tool category to perform an action on a target object category and divide our data into " + }, + { + "bbox": [ + 46, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "131 <" + }, + { + "bbox": [ + 46, + 420, + 287, + 529 + ], + "type": "text", + "content": " tool category, action label, target object category " + }, + { + "bbox": [ + 46, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\rightharpoondown" + }, + { + "bbox": [ + 46, + 420, + 287, + 529 + ], + "type": "text", + "content": " triplets. Some correlated triplets are exemplified in Figure 5. To enable various studies, TACO presents high-resolution color images from 12 allocentric views and RGBD images from one egocentric-view." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 538, + 188, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 188, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 188, + 552 + ], + "type": "text", + "content": "4. Data Quality Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 665 + ], + "type": "text", + "content": "Qualitative contact optimization evaluation. Figure 6 depicts two frames of the TACO dataset. As shown in the figure, the attraction loss " + }, + { + "bbox": [ + 46, + 558, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_a" + }, + { + "bbox": [ + 46, + 558, + 287, + 665 + ], + "type": "text", + "content": " promotes contact between the hands and the objects but may exacerbate penetration. The penetration loss " + }, + { + "bbox": [ + 46, + 558, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 46, + 558, + 287, + 665 + ], + "type": "text", + "content": " can mitigate penetration but does not address cases where the hands and the objects are not in contact. Our method combines both attraction loss and penetration loss, striking a balance between encouraging contact and preventing penetration." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "content": "Quantitative hand pose evaluation. To quantitatively examine the quality and value of our hand poses, we conduct a cross-dataset evaluation between TACO and an existing high-quality HOI dataset DexYCB [4]. We select a" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 335, + 296, + 515, + 445 + ], + "blocks": [ + { + "bbox": [ + 335, + 296, + 515, + 445 + ], + "lines": [ + { + "bbox": [ + 335, + 296, + 515, + 445 + ], + "spans": [ + { + "bbox": [ + 335, + 296, + 515, + 445 + ], + "type": "image", + "image_path": "d1fa6c2bf2b6a49e434c0775b0b8dda0359609cf8149b7cd2b4c5af01a88fe52.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 447, + 547, + 491 + ], + "lines": [ + { + "bbox": [ + 305, + 447, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 447, + 547, + 491 + ], + "type": "text", + "content": "Figure 5. Examples of interaction triplets in TACO. The left, middle, and right columns exemplify categories of tool, action, and target object, respectively. Triplets in our dataset are represented by connected paths from tools to target objects." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 494, + 547, + 620 + ], + "blocks": [ + { + "bbox": [ + 307, + 494, + 547, + 620 + ], + "lines": [ + { + "bbox": [ + 307, + 494, + 547, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 494, + 547, + 620 + ], + "type": "image", + "image_path": "2b711509c1a6053a50a41ff9401d072f8b6db4eab9a02b2f859a874593e6c1fa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 632, + 547, + 676 + ], + "lines": [ + { + "bbox": [ + 305, + 632, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 305, + 632, + 547, + 676 + ], + "type": "text", + "content": "Figure 6. Qualitative contact optimization evaluation. From left to right: original RGB image, optimization without attraction loss, optimization without penetration loss, our method with attraction loss and penetration loss." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "multi-stage method, CMR [5], and a lightweight network, MobRecon [6], both of which can recover occluded hand" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21744" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 130 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 130 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 130 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 130 + ], + "type": "table", + "html": "
MethodTrain TestTACODexYCBTACO + DexYCB
CMRTACO10.73 / 30.8713.53 / 39.5410.25 / 28.36
DexYCB19.36 / 76.206.51 / 13.926.44 / 13.62
MobReconTACO9.03 / 20.7612.95 / 34.328.69 / 20.14
DexYCB15.17 / 44.736.51 / 13.616.32 / 12.94
", + "image_path": "fe0a9b453a282e2f3b49d66090cc7947ccf6a653bf5fdb605216ffa5a13ce459.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 132, + 287, + 164 + ], + "lines": [ + { + "bbox": [ + 47, + 132, + 287, + 164 + ], + "spans": [ + { + "bbox": [ + 47, + 132, + 287, + 164 + ], + "type": "text", + "content": "Table 2. Cross-dataset evaluation with DexYCB [4] on 3D hand pose estimation. Results are in PA-MPJPE (mm, lower is better) and MPJPE (mm, lower is better), respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 171, + 289, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 289, + 423 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 289, + 423 + ], + "type": "text", + "content": "meshes from a single-view RGB image captured in HOI scenarios, and then train and test their networks under different data combinations. For DexYCB, we follow its default train/test split (S0). For TACO, we randomly select 60 motion sequences for training and 40 for testing and crop the right hand from each image to align with the single-hand HO3D. Following existing hand mesh recovery methods [5, 6, 45, 67], we select Procrustes-Aligned Mean Per Joint Position Error (PA-MPJPE) and Mean Per Joint Position Error (MPJPE) as evaluation metrics. More details about the metrics can be found in the paper [6]. Table 2 shows the performance of the two methods under different settings. Due to the large gap in hand-object motion between the two datasets, the performances for the two methods would both drop rapidly when the training and the test sets are from different datasets. Nevertheless, combining training data from TACO and DexYCB achieves significant performance gains on both datasets for both the two methods, which demonstrates the accuracy of our hand pose annotations and indicates that TACO is a beneficial data complement for 3D hand pose estimation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": "Marker removal quality evaluation. We conduct a novel experiment to examine the reality of our marker-removed images. The key idea is that a network can easily segment markers from the image while being hard to figure out the same regions when markers are replaced by the original object surface appearances. We thus use the originally captured image patches and the marker-removed ones to train a 2D marker segmentation network, respectively, and then apply the two networks to novel image patches with the same image processing setting as training. In practice, spheres larger than markers (as shown in Figure 4(b)) are required to be segmented, since the marker removal process exactly inpaints these regions. We select U-Net [54] as the network backbone and train the two networks with an L2-loss measuring differences between mask predictions and ground truths. When training and testing on raw image patches, U-Net achieves " + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "63.8\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": " mean Intersection over Union (mIoU), while its performance sharply drops to " + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "11.1\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": " mIoU after applying marker removal, indicating the reality and naturalness of the inpainted image regions. Figure 7 further compares heatmaps of network predictions under the two settings, exhibiting ambiguities of marker segmentation posed by realistic inpainted images. More details are provided in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 364, + 127 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 364, + 127 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 364, + 127 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 364, + 127 + ], + "type": "image", + "image_path": "c0d87809ba63d4c78231e9ac99541f59ac0bea7fc55e291032bcd1954412c4b9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 365, + 72, + 421, + 127 + ], + "blocks": [ + { + "bbox": [ + 365, + 72, + 421, + 127 + ], + "lines": [ + { + "bbox": [ + 365, + 72, + 421, + 127 + ], + "spans": [ + { + "bbox": [ + 365, + 72, + 421, + 127 + ], + "type": "image", + "image_path": "b1a76b24aa839527c3380d75fb4f3807189169d671783cd7c08832ab47332bcc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 128, + 364, + 182 + ], + "blocks": [ + { + "bbox": [ + 309, + 128, + 364, + 182 + ], + "lines": [ + { + "bbox": [ + 309, + 128, + 364, + 182 + ], + "spans": [ + { + "bbox": [ + 309, + 128, + 364, + 182 + ], + "type": "image", + "image_path": "270ddb091364333c9892f5f5c7d893d13bfc6d6cf4b9400f9b84c99f984239ac.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 183, + 369, + 190 + ], + "lines": [ + { + "bbox": [ + 361, + 183, + 369, + 190 + ], + "spans": [ + { + "bbox": [ + 361, + 183, + 369, + 190 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 365, + 128, + 421, + 182 + ], + "blocks": [ + { + "bbox": [ + 365, + 128, + 421, + 182 + ], + "lines": [ + { + "bbox": [ + 365, + 128, + 421, + 182 + ], + "spans": [ + { + "bbox": [ + 365, + 128, + 421, + 182 + ], + "type": "image", + "image_path": "a57fa7f153f47d85c7f8be2cfa1e2ee9e9da703d0e44513408e3fa586dc2a2f5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 205, + 547, + 261 + ], + "lines": [ + { + "bbox": [ + 304, + 205, + 547, + 261 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 547, + 261 + ], + "type": "text", + "content": "Figure 7. Marker removal quality evaluation. We train and test U-Net [54] on raw captured images (a) and marker-removed images (b), respectively. In each experiment, the left and right columns show the input image patches and the predicted segmentation heatmaps, respectively." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 432, + 72, + 486, + 127 + ], + "blocks": [ + { + "bbox": [ + 432, + 72, + 486, + 127 + ], + "lines": [ + { + "bbox": [ + 432, + 72, + 486, + 127 + ], + "spans": [ + { + "bbox": [ + 432, + 72, + 486, + 127 + ], + "type": "image", + "image_path": "0a966003a09c3e49348500b8bd9308f95f874d96003eef769769d004027f2117.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 432, + 128, + 486, + 182 + ], + "blocks": [ + { + "bbox": [ + 432, + 128, + 486, + 182 + ], + "lines": [ + { + "bbox": [ + 432, + 128, + 486, + 182 + ], + "spans": [ + { + "bbox": [ + 432, + 128, + 486, + 182 + ], + "type": "image", + "image_path": "db3f0b10190b5f10903a553e4ea8dfdbafbae201e853c38b2c545c8c182c8f2f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 183, + 491, + 190 + ], + "lines": [ + { + "bbox": [ + 484, + 183, + 491, + 190 + ], + "spans": [ + { + "bbox": [ + 484, + 183, + 491, + 190 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 487, + 72, + 543, + 127 + ], + "blocks": [ + { + "bbox": [ + 487, + 72, + 543, + 127 + ], + "lines": [ + { + "bbox": [ + 487, + 72, + 543, + 127 + ], + "spans": [ + { + "bbox": [ + 487, + 72, + 543, + 127 + ], + "type": "image", + "image_path": "d6d1e87c980527b470b33011076f8cd6ff8b2924cb43d3acd7aa5f85eaa7e6ce.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 488, + 128, + 544, + 182 + ], + "blocks": [ + { + "bbox": [ + 488, + 128, + 544, + 182 + ], + "lines": [ + { + "bbox": [ + 488, + 128, + 544, + 182 + ], + "spans": [ + { + "bbox": [ + 488, + 128, + 544, + 182 + ], + "type": "image", + "image_path": "7bf980a1437b8e43178c59bc83a3dc6b1eba837dd5a9d4126079ad9f105bd050.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 268, + 388, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 268, + 388, + 282 + ], + "spans": [ + { + "bbox": [ + 306, + 268, + 388, + 282 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 289, + 547, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 547, + 373 + ], + "type": "text", + "content": "In this section, we first introduce our data split (Section 5.1) supporting different research purposes about generalization. We then present three benchmarks on TACO: compositional action recognition (Section 5.2), generalizable hand-object motion forecasting (Section 5.3), and cooperative grasp synthesis (Section 5.4). Details of evaluation metrics are provided in supplementary materials." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 382, + 376, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 382, + 376, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 382, + 376, + 396 + ], + "type": "text", + "content": "5.1. Data Split" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 402, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 462 + ], + "type": "text", + "content": "With diverse object geometries and interaction triplets, TACO focuses on supporting generalizable studies that apply to unseen object geometries or novel interaction triplets. Hence, we carefully split our test data into four subsets with different generalization purposes:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 465, + 545, + 585 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 306, + 465, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 545, + 488 + ], + "type": "text", + "content": "- Test set 1 (S1): No generalization. Tool geometries and interaction triplets are both involved in the training set." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 489, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 489, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 489, + 545, + 525 + ], + "type": "text", + "content": "- Test set 2 (S2): Geometry-level generalization. The tool geometry is novel, while the interaction triplet appears in the training set." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 525, + 545, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 545, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 545, + 560 + ], + "type": "text", + "content": "- Test set 3 (S3): Interaction-level generalization. The interaction triplet is novel, while the tool categories and geometries are included in the training set." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 561, + 545, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 561, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 306, + 561, + 545, + 585 + ], + "type": "text", + "content": "- Test set 4 (S4): Compound generalization. The tool category is novel, leading to unseen geometries and triplets." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 588, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 545, + 611 + ], + "type": "text", + "content": "The training set and the four test sets follow a data ratio of 4:1:1:1.5:2.5." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 622, + 491, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 622, + 491, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 622, + 491, + 635 + ], + "type": "text", + "content": "5.2. Compositional Action Recognition" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Humans can recognize the action disentangled from their context and appearance biases of the objects, which is referred to as compositionality. To test a system's compositional generalization capability, with a diverse array of tool-action-object triplets in TACO, we benchmark compositional action recognition aiming at identifying actions" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21745" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 263, + 173 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 263, + 173 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 263, + 173 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 263, + 173 + ], + "type": "table", + "html": "
Test SetMethodTop-1 (%, ↑)Top-5 (%, ↑)
S1AIM [73]83.0898.85
CACNF [50]86.1599.62
S2AIM [73]82.8198.83
CACNF [50]77.3496.88
S3AIM [73]53.6581.25
CACNF [50]63.0292.97
S4AIM [73]39.3373.67
CACNF [50]44.0080.50
", + "image_path": "8f7a3976e44021ae54a3b305b93b9659437f500211c4e74ec385ce92247afe3c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 176, + 287, + 199 + ], + "lines": [ + { + "bbox": [ + 46, + 176, + 287, + 199 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 287, + 199 + ], + "type": "text", + "content": "Table 3. Results on compositional action recognition. Methods are examined via Top-1 and Top-5 accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 205, + 286, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 286, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 286, + 278 + ], + "type": "text", + "content": "in HOI videos. In contrast to traditional action recognition tasks [11, 16, 56, 58], our focus is on scenarios where the tool-action-object triplets in the test set are unseen during training, including novel tool types or geometries. This poses additional challenges for evaluating model generalization and achieving human-like perception." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 278, + 286, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 278, + 286, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 286, + 326 + ], + "type": "text", + "content": "Problem formulation: In a bimanual HOI scenario, our goal is to identify action labels from first-person RGB video frames, with optional bounding boxes for hands and objects computed based on their poses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 326, + 286, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 326, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 286, + 361 + ], + "type": "text", + "content": "Evaluation metrics: Following [50, 73], we use Top-1 Accuracy and Top-5 Accuracy to evaluate the efficacy of action recognition." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 363, + 286, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 286, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 286, + 638 + ], + "type": "text", + "content": "Baselines, results, and analysis: We compare two baselines to evaluate the compositional generalization capability: AIM [73], an adaptation of pre-trained image transformer models, representing state-of-the-art traditional action recognition approaches; and CACNF [50], which specializes in compositional and few-shot action recognition. Table 3 shows the results. A significant decrease in accuracy on the most difficult S4 set, compared to high accuracy on the non-generalizable S1 set, underscores compositional generalization challenges. Notably, the S2 set, with novel tool geometries within tool categories, performed similarly to S1, while the S3 set, featuring novel tool-action-object triplets, showed a significant accuracy drop, revealing that while action recognition models inherently possess basic generalizability to geometric changes, they struggle with novel triplets due to the unfamiliarity of interaction contexts. When comparing among baselines, CACNF outperformed AIM in challenging sets S3 and S4, benefiting from its bounding box-based approach that effectively disentangles actions from object and tool geometries, highlighting the advantages of a compositional approach in generalizable cases. All these results underscore the significant impact of compositional generalization on action recognition." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 647, + 286, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 286, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 286, + 659 + ], + "type": "text", + "content": "5.3. Generalizable Hand-object Motion Forecasting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": "With abundant 3D hand-object mesh annotations provided by TACO, we benchmark generalizable interaction forecasting aiming to predict the following hand-object motions from seen short motion clips. Unlike the prediction of" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "low-frequency human motions [42, 46, 47], we observe that hands commonly achieve a complete manipulation behavior (e.g. pouring all the water out of a bowl) in a very short time, making forecasting both interesting and challenging." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 121, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 121, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 305, + 121, + 545, + 179 + ], + "type": "text", + "content": "Problem formulation: In a bimanual HOI scenario, given object point clouds and poses of two hands, the tool, and the target object in consecutive " + }, + { + "bbox": [ + 305, + 121, + 545, + 179 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 121, + 545, + 179 + ], + "type": "text", + "content": " frames, our goal is to forecast their poses in subsequent " + }, + { + "bbox": [ + 305, + 121, + 545, + 179 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 121, + 545, + 179 + ], + "type": "text", + "content": " frames. We select " + }, + { + "bbox": [ + 305, + 121, + 545, + 179 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 305, + 121, + 545, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 121, + 545, + 179 + ], + "type": "inline_equation", + "content": "M = 10" + }, + { + "bbox": [ + 305, + 121, + 545, + 179 + ], + "type": "text", + "content": " in our dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 181, + 545, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 181, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 305, + 181, + 545, + 241 + ], + "type": "text", + "content": "Evaluation metrics: Following human-object forecasting evaluations [1, 10, 63, 68], we use Mean Per Joint Position Error " + }, + { + "bbox": [ + 305, + 181, + 545, + 241 + ], + "type": "inline_equation", + "content": "J_{e}" + }, + { + "bbox": [ + 305, + 181, + 545, + 241 + ], + "type": "text", + "content": " to evaluate predicted left and right-hand skeletons, and leverage translation error " + }, + { + "bbox": [ + 305, + 181, + 545, + 241 + ], + "type": "inline_equation", + "content": "T_{e}" + }, + { + "bbox": [ + 305, + 181, + 545, + 241 + ], + "type": "text", + "content": " and rotation error " + }, + { + "bbox": [ + 305, + 181, + 545, + 241 + ], + "type": "inline_equation", + "content": "R_{e}" + }, + { + "bbox": [ + 305, + 181, + 545, + 241 + ], + "type": "text", + "content": " to measure object positions and orientations, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 242, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 242, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 305, + 242, + 545, + 529 + ], + "type": "text", + "content": "Baselines, results, and analysis: Due to the lack of hand-object interaction forecasting solutions, we transfer methodologies from human-object motion forecasting as baselines. To cover diverse method designs and possibilities, we select two generative models [62, 68] and two predictive models [10, 68] for comparison. Table 4 compares their performance under different generalization settings. We observed that the tool and the hand holding it play dominant roles during manipulation, making their motion forecasting more challenging compared to others. The generative models, in comparison to predictive ones, yield significantly larger errors for both hands and the tool. One possible reason is that their motion patterns are fast and complex, making the modeling of motion distribution challenging. Enhancing the accuracy of generative methods in hand-object manipulation scenarios is an interesting future direction. When comparing method performances between a non-generalizable set (S1) and sets involving generalization (S2-4), we note that methods consistently exhibit significant performance declines with the right hand and the tool, regardless of whether applied to novel tool geometries (S2, S4) or interaction triplets (S3, S4), meanwhile achieving close results on the others. This trend indicates substantial challenges in generalizing dominant interaction entities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 538, + 463, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 463, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 463, + 551 + ], + "type": "text", + "content": "5.4. Cooperative Grasp Synthesis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 557, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 557, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 305, + 557, + 545, + 700 + ], + "type": "text", + "content": "Generating human-like hand grasps benefits various applications in VR/AR and dexterous manipulation. Existing grasp synthesis approaches [28, 30, 38, 81] mainly focus on creating stable grasps on static objects without considering interactive purposes. Taking advantage of extensive interaction behaviors from TACO, we benchmark a novel task that aims to synthesize realistic and physically plausible hand grasps in HOI scenarios. Besides achieving stable grasping for the object, the method must comprehend other interactive objects and human hands to generate cooperative motions and avoid conflicts. We evaluate this task on our four test sets, each targeting different aspects of generalization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "Problem formulation: Consider using the right hand to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21746" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 272 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 272 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 272 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 272 + ], + "type": "table", + "html": "
Test SetMethodJe(mm,↓)Te(mm,↓)Re(°,↓)
Right / LeftTool / TargetTool / Target
S1InterVAE [68]54.9 / 48.955.0 / 15.866.94 / 6.63
MDM [62]61.2 / 53.749.4 / 14.952.81 / 5.85
InterRNN [68]29.3 / 20.425.8 / 11.610.08 / 5.22
CAHMP [10]28.8 / 22.123.9 / 12.310.24 / 4.95
S2InterVAE [68]58.8 / 50.854.3 / 12.288.19 / 5.24
MDM [62]65.8 / 55.848.0 / 10.175.78 / 3.59
InterRNN [68]35.2 / 23.131.0 / 8.911.09 / 4.03
CAHMP [10]31.4 / 22.624.7 / 8.610.65 / 3.32
S3InterVAE [68]56.5 / 50.156.7 / 12.768.03 / 5.69
MDM [62]66.4 / 58.046.2 / 11.561.90 / 4.14
InterRNN [68]32.5 / 22.028.2 / 9.711.42 / 4.32
CAHMP [10]29.3 / 21.624.4 / 9.411.18 / 3.80
S4InterVAE [68]63.9 / 54.263.0 / 14.571.38 / 5.01
MDM [62]70.5 / 57.950.2 / 12.270.25 / 3.55
InterRNN [68]36.6 / 22.330.9 / 10.611.92 / 3.86
CAHMP [10]32.9 / 22.226.6 / 10.811.73 / 3.24
", + "image_path": "4cf6911ba3f591d41d9817c2c96139ad832b19f2aae0fa78bbe69eb80f4ed3b0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 317, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 287, + 365 + ], + "type": "text", + "content": "manipulate a tool and the left hand to operate a target object. During manipulation, given the meshes of the two objects and the left hand, the goal is to generate a right-hand mesh grasping the tool in a manner conducive to the interaction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 366, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 484 + ], + "type": "text", + "content": "Evaluation metrics: Following static grasp synthesis methods [28, 29, 38], we use interpenetration volume (Pen. " + }, + { + "bbox": [ + 46, + 366, + 287, + 484 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 366, + 287, + 484 + ], + "type": "text", + "content": ") and contact ratio (Con. " + }, + { + "bbox": [ + 46, + 366, + 287, + 484 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 366, + 287, + 484 + ], + "type": "text", + "content": ") to measure the physical plausibility of generated hands. To ensure the hand is appropriately interactive, we introduce a new metric named collision ratio (Col. " + }, + { + "bbox": [ + 46, + 366, + 287, + 484 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 366, + 287, + 484 + ], + "type": "text", + "content": "), which detects collisions between generated hands and the interaction environment. Nevertheless, to quantitatively examine whether the grasps are realistic, we design the FID score (FID) to measure distances between distributions of human grasps and synthesized ones." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "type": "text", + "content": "Baselines, results, and analysis: We select two recent CVAE-based approaches ContactGen [38] and HALO-VAE [30], and modify their network structures to incorporate the left hand and the target object as additional VAE conditions. To assess the significance of interaction environments, we also evaluate HALO-VAE" + }, + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "type": "inline_equation", + "content": "^{-}" + }, + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "type": "text", + "content": ", which omits the use of environment inputs and directly generates from the tool mesh. Due to the strong correlation between hand grasps and tool geometries, we test these approaches with familiar (S1, S3) and unseen tool geometries (S2, S4). Table 5 summarizes quantitative evaluations of the two settings. Notably, HALO-VAE" + }, + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "type": "inline_equation", + "content": "^{-}" + }, + { + "bbox": [ + 46, + 486, + 287, + 713 + ], + "type": "text", + "content": " shows a significant decrease in performance on contact and collision ratios compared to HALO-VAE, underscoring the importance of perceiving and modeling interaction environments. When applied to novel object geometries and categories, all three approaches exhibit higher collision ratios and FID scores, and most of them show a marked decline in their ability to contact tools precisely and prevent penetrations. Figure 8 exemplifies" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 554, + 161 + ], + "blocks": [ + { + "bbox": [ + 46, + 276, + 288, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 276, + 288, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 288, + 309 + ], + "type": "text", + "content": "Table 4. Results on generalizable interaction forecasting. We respectively examine predictions of the right hand, the left hand, the tool, and the target object." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 70, + 554, + 161 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 554, + 161 + ], + "type": "table", + "html": "
Test SetMethodPen. V (cm3, ↓)Con. R (%, ↑)Col. R (%, ↓)FID (10-2, ↓)
S1 ∪ S3ContactGen [38]0.2780.1210.4012.18
HALO-VAE-0.3176.274.194.89
HALO-VAE [30]0.2682.862.392.02
S2 ∪ S4ContactGen [38]0.2674.9817.5622.53
HALO-VAE-0.2971.6914.255.09
HALO-VAE [30]0.3283.5211.112.84
", + "image_path": "41a107f40ca36a4d7260ee967eea277aab84b998f2a0a05d7b42e438741084a7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 165, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 305, + 165, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 165, + 545, + 209 + ], + "type": "text", + "content": "Table 5. Results on cooperative grasp synthesis. We measure the physical plausibility of generated grasps by interpenetration volume and contact ratio, meanwhile using collision ratio to examine interaction faithfulness and FID score to assess reality." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 216, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 216, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 216, + 545, + 251 + ], + "type": "text", + "content": "failure cases of synthesized grasps, indicating a large room for improving physical feasibility and naturalness on complex and slender geometries." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 261, + 545, + 376 + ], + "blocks": [ + { + "bbox": [ + 307, + 261, + 545, + 376 + ], + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 376 + ], + "type": "image", + "image_path": "b3a1e009f4eed4b73751e651a84d3e8774212406a848a3919311a966c77c4946.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 388, + 545, + 421 + ], + "lines": [ + { + "bbox": [ + 305, + 388, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 545, + 421 + ], + "type": "text", + "content": "Figure 8. Failure cases of cooperative grasp synthesis. (a) and (b) are generated by ContactGen [38], while (c) and (d) are from HALO-VAE [30]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 430, + 463, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 463, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 463, + 442 + ], + "type": "text", + "content": "6. Limitations and Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 450, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 617 + ], + "type": "text", + "content": "We present TACO, the first large-scale, real-world 4D bi-manual hand-object manipulation dataset. It encompasses a wide range of tool-action-object compositions and object geometries. TACO contains a total of " + }, + { + "bbox": [ + 304, + 450, + 545, + 617 + ], + "type": "inline_equation", + "content": "2.5\\mathrm{K}" + }, + { + "bbox": [ + 304, + 450, + 545, + 617 + ], + "type": "text", + "content": " motion sequences and " + }, + { + "bbox": [ + 304, + 450, + 545, + 617 + ], + "type": "inline_equation", + "content": "5.2\\mathrm{M}" + }, + { + "bbox": [ + 304, + 450, + 545, + 617 + ], + "type": "text", + "content": " video frames, captured from 12 third-person views and one egocentric view. We contribute an automatic data acquisition pipeline that accurately recovers hand-object meshes and segmentation, along with realistic hand-object appearances. Leveraging TACO's diverse data, we benchmark compositional action recognition, generalizable hand-object motion forecasting, and cooperative grasp synthesis. These benchmarks reveal new insights, challenges, and opportunities in the field of generalizable hand-object interaction studies." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": "There are three major limitations in TACO. Firstly, TACO currently does not cover articulated objects. Secondly, while TACO offers an extensive exploration of object geometries and HOI behaviors, it lacks scene diversities that are also crucial for understanding human manipulations. Thirdly, our solution to marker removal is an application of generative models, hence the original object appearances cannot be recovered perfectly." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21747" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Vida Adeli, Mahsa Ehsanpour, Ian Reid, Juan Carlos Niebles, Silvio Savarese, Ehsan Adeli, and Hamid Rezatofighi. Tripod: Human trajectory and pose dynamics forecasting in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13390-13400, 2021. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 223 + ], + "type": "text", + "content": "[2] Samarth Brahmbhatt, Chengcheng Tang, Christopher D Twigg, Charles C Kemp, and James Hays. Contactpose: A dataset of grasps with object contact and hand pose. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 361-378. Springer, 2020. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 224, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 267 + ], + "type": "text", + "content": "[3] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 269, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 287, + 345 + ], + "type": "text", + "content": "[4] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 346, + 287, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 412 + ], + "type": "text", + "content": "[5] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 414, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 479 + ], + "type": "text", + "content": "[6] Xingyu Chen, Yufeng Liu, Dong Yajiao, Xiong Zhang, Chongyang Ma, Yanmin Xiong, Yuan Zhang, and Xiaoyan Guo. Mobrecon: Mobile-friendly hand mesh reconstruction from monocular image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 480, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 544 + ], + "type": "text", + "content": "[7] Sammy Christen, Muhammed Kocabas, Emre Aksan, Jemin Hwangbo, Jie Song, and Otmar Hilliges. D-grasp: Physically plausible dynamic grasp synthesis for hand-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20577-20586, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 546, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 546, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 546, + 287, + 601 + ], + "type": "text", + "content": "[8] Sammy Christen, Wei Yang, Claudia Pérez-D'Arpino, Otmar Hilliges, Dieter Fox, and Yu-Wei Chao. Learning human-to-robot handovers from point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9654–9664, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 602, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 602, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 53, + 602, + 287, + 634 + ], + "type": "text", + "content": "[9] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[10] Enric Corona, Albert Pumarola, Guillem Alenya, and Francesc Moreno-Noguer. Context-aware human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6992-7001, 2020. 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[11] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "type": "text", + "content": "Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European conference on computer vision (ECCV), pages 720-736, 2018. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "type": "text", + "content": "[12] Christian RG Dreher, Mirko Wächter, and Tamim Asfour. Learning object-action relations from bimanual human demonstration using graph networks. IEEE Robotics and Automation Letters, 5(1):187-194, 2019. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 162, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 227 + ], + "type": "text", + "content": "[13] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J Black, and Otmar Hilliges. Arctic: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12954, 2023. 2, 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "text", + "content": "[14] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 273, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 273, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 545, + 327 + ], + "type": "text", + "content": "[15] Guillermo Garcia-Hernando, Shanxin Yuan, Seungryul Baek, and Tae-Kyun Kim. First-person hand action benchmark with rgb-d videos and 3d hand pose annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 409-419, 2018. 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 328, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 328, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 328, + 545, + 404 + ], + "type": "text", + "content": "[16] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 405, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 470 + ], + "type": "text", + "content": "[17] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 472, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 525 + ], + "type": "text", + "content": "[18] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3196-3206, 2020. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 526, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 591 + ], + "type": "text", + "content": "[19] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11090-11100, 2022. 2, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 593, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 657 + ], + "type": "text", + "content": "[20] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11807-11816, 2019. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[21] Roei Herzig, Elad Ben-Avraham, Karttikeya Mangalam, Amir Bar, Gal Chechik, Anna Rohrbach, Trevor Darrell, and Amir Globerson. Object-region video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3148-3159, 2022. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21748" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "type": "text", + "content": "[22] Zhi Hou, Xiaojiang Peng, Yu Qiao, and Dacheng Tao. Visual compositional learning for human-object interaction detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 584-600. Springer, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "text", + "content": "[23] Zhi Hou, Baosheng Yu, Yu Qiao, Xiaojiang Peng, and Dacheng Tao. Affordance transfer learning for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 495-504, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "text", + "content": "[24] Zhi Hou, Baosheng Yu, Yu Qiao, Xiaojiang Peng, and Dacheng Tao. Detecting human-object interaction via fabricated compositional learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14646-14655, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "text", + "content": "[25] Zhi Hou, Baosheng Yu, and Dacheng Tao. Discovering human-object interaction concepts via self-compositional learning. In European Conference on Computer Vision, pages 461-478. Springer, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 342 + ], + "type": "text", + "content": "[26] Juntao Jian, Xiuping Liu, Manyi Li, Ruizhen Hu, and Jian Liu. Affordpose: A large-scale dataset of hand-object interactions with affordance-driven hand pose. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14713-14724, 2023. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 343, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 409 + ], + "type": "text", + "content": "[27] Changlong Jiang, Yang Xiao, Cunlin Wu, Mingyang Zhang, Jinghong Zheng, Zhiguo Cao, and Joey Tianyi Zhou. A2j-transformer: Anchor-to-joint transformer network for 3d interacting hand pose estimation from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8846-8855, 2023. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "type": "text", + "content": "[28] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J Black, Krikamol Muandet, and Siyu Tang. Grasping field: Learning implicit representations for human grasps. In 2020 International Conference on 3D Vision (3DV), pages 333-344. IEEE, 2020. 3, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 466, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 521 + ], + "type": "text", + "content": "[29] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J Black, Krikamol Muandet, and Siyu Tang. Grasping field: Learning implicit representations for human grasps. In 2020 International Conference on 3D Vision (3DV), pages 333-344. IEEE, 2020. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 522, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 576 + ], + "type": "text", + "content": "[30] Korrawe Karunratanakul, Adrian Spurr, Zicong Fan, Otmar Hilliges, and Siyu Tang. A skeleton-driven neural occupancy representation for articulated hands. In 2021 International Conference on 3D Vision (3DV), pages 11-21. IEEE, 2021. 3, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "type": "text", + "content": "[31] Tae Soo Kim, Jonathan Jones, and Gregory D Hager. Motion guided attention fusion to recognize interactions from videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13076-13086, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[33] Franziska Krebs, Andre Meixner, Isabel Patzer, and Tamim Asfour. The kit bimanual manipulation dataset. In 2020 IEEE-RAS 20th International Conference on Humanoid Robots (Humanoids), pages 499-506. IEEE, 2021. 2, 3, 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": "[34] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 2, 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 546, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 171 + ], + "type": "text", + "content": "[35] Haoming Li, Xinzhuo Lin, Yang Zhou, Xiang Li, Yuchi Huo, Jiming Chen, and Qi Ye. Contact2grasp: 3d grasp synthesis via hand-object contact constraint. arXiv preprint arXiv:2210.09245, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 172, + 546, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 546, + 228 + ], + "type": "text", + "content": "[36] Mengcheng Li, Liang An, Hongwen Zhang, Lianpeng Wu, Feng Chen, Tao Yu, and Yebin Liu. Interacting attention graph for single image two-hand reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2761-2770, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 228, + 546, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 546, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 546, + 281 + ], + "type": "text", + "content": "[37] Qingtao Liu, Yu Cui, Zhengnan Sun, Haoming Li, Gaofeng Li, Lin Shao, Jiming Chen, and Qi Ye. Dexrepnet: Learning dexterous robotic grasping network with geometric and spatial hand-object representations. arXiv preprint arXiv:2303.09806, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 283, + 546, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 546, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 546, + 337 + ], + "type": "text", + "content": "[38] Shaowei Liu, Yang Zhou, Jimei Yang, Saurabh Gupta, and Shenlong Wang. Contactgen: Generative contact modeling for grasp generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20609-20620, 2023. 3, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 338, + 546, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 546, + 403 + ], + "type": "text", + "content": "[39] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 2, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 404, + 546, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 546, + 459 + ], + "type": "text", + "content": "[40] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 460, + 547, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 547, + 514 + ], + "type": "text", + "content": "[41] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 514, + 546, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 546, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 546, + 548 + ], + "type": "text", + "content": "[42] Wei Mao, Richard I Hartley, Mathieu Salzmann, et al. Contact-aware human motion forecasting. Advances in Neural Information Processing Systems, 35:7356-7367, 2022. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 548, + 546, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 546, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 546, + 613 + ], + "type": "text", + "content": "[43] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-else: Compositional action recognition with spatial-temporal interaction networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1049–1059, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 613, + 546, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 546, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 546, + 658 + ], + "type": "text", + "content": "[44] Eley Ng, Ziang Liu, and Monroe Kennedy. It takes two: Learning to plan for human-robot cooperative carrying. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7526–7532. IEEE, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 658, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 546, + 713 + ], + "type": "text", + "content": "[45] JoonKyu Park, Yeonguk Oh, Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Handocctnet: Occlusion-robust 3d hand mesh estimation network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1496–1505, 2022. 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21749" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[46] Xiaogang Peng, Yaodi Shen, Haoran Wang, Binling Nie, Yiqiang Wang, and Zizhao Wu. Somoformer: Social-aware motion transformer for multi-person motion prediction. arXiv preprint arXiv:2208.09224, 2022.7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "type": "text", + "content": "[47] Xiaogang Peng, Siyuan Mao, and Zizhao Wu. Trajectory-aware body interaction transformer for multi-person pose forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17121-17130, 2023. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 288, + 217 + ], + "type": "text", + "content": "[48] Yuzhe Qin, Hao Su, and Xiaolong Wang. From one hand to multiple hands: Imitation learning for dexterous manipulation from single-camera teleoperation. IEEE Robotics and Automation Letters, 7(4):10873-10881, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 288, + 274 + ], + "type": "text", + "content": "[49] Yuzhe Qin, Yueh-Hua Wu, Shaowei Liu, Hanwen Jiang, Ruihan Yang, Yang Fu, and Xiaolong Wang. Dexamv: Imitation learning for dexterous manipulation from human videos. In European Conference on Computer Vision, pages 570-587. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 288, + 319 + ], + "type": "text", + "content": "[50] Gorjan Radevski, Marie-Francine Moens, and Tinne Tuyte-laars. Revisiting spatio-temporal layouts for compositional action recognition. arXiv preprint arXiv:2111.01936, 2021. 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 288, + 363 + ], + "type": "text", + "content": "[51] Ramanathan Rajendiran, Debaditya Roy, and Basura Fernando. Modelling spatio-temporal interactions for compositional action recognition. arXiv preprint arXiv:2305.02673, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 288, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 388 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 388 + ], + "type": "text", + "content": "[52] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 288, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 422 + ], + "type": "text", + "content": "[53] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 288, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 488 + ], + "type": "text", + "content": "[54] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 489, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 288, + 555 + ], + "type": "text", + "content": "[55] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 556, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 288, + 601 + ], + "type": "text", + "content": "[56] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 288, + 645 + ], + "type": "text", + "content": "[57] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "type": "text", + "content": "[58] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[59] Pengzhan Sun, Bo Wu, Xunsong Li, Wen Li, Lixin Duan, and Chuang Gan. Counterfactual debiasing inference for compositional action recognition. In Proceedings of the 29th" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ACM International Conference on Multimedia, pages 3220-3228, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 162 + ], + "type": "text", + "content": "[60] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021.4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 547, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 219 + ], + "type": "text", + "content": "[61] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 581-600. Springer, 2020. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 220, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 547, + 262 + ], + "type": "text", + "content": "[62] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 264, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 547, + 319 + ], + "type": "text", + "content": "[63] Weilin Wan, Lei Yang, Lingjie Liu, Zhuoying Zhang, Ruixing Jia, Yi-King Choi, Jia Pan, Christian Theobalt, Taku Komura, and Wenping Wang. Learn to predict how humans manipulate large-sized objects from interactive motions. IEEE Robotics and Automation Letters, 7(2):4702-4709, 2022. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 320, + 547, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 547, + 365 + ], + "type": "text", + "content": "[64] Haozhong Wang, Hua Yu, and Qiang Zhang. Detecting zero-shot human-object interaction with visual-text modeling. In 2023 9th International Conference on Virtual Reality (ICVR), pages 155-162. IEEE, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 547, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 547, + 431 + ], + "type": "text", + "content": "[65] Suchen Wang, Yueqi Duan, Henghui Ding, Yap-Peng Tan, Kim-Hui Yap, and Junsong Yuan. Learning transferable human-object interaction detector with natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 939-948, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 433, + 547, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 547, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 547, + 476 + ], + "type": "text", + "content": "[66] Wei Xie, Zhipeng Yu, Zimeng Zhao, Binghui Zuo, and Yanggang Wang. Hmdo: Markerless multi-view hand manipulation capture with deformable objects. Graphical Models, 127:101178, 2023. 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 478, + 547, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 547, + 533 + ], + "type": "text", + "content": "[67] Hao Xu, Tianyu Wang, Xiao Tang, and Chi-Wing Fu. H2onet: Hand-occlusion-and-orientation-aware network for real-time 3d hand mesh reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17048–17058, 2023. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 534, + 547, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 547, + 589 + ], + "type": "text", + "content": "[68] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14928-14940, 2023. 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 590, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 547, + 645 + ], + "type": "text", + "content": "[69] Rui Yan, Peng Huang, Xiangbo Shu, Junhao Zhang, Yonghua Pan, and Jinhui Tang. Look less think more: Rethinking compositional action recognition. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3666-3675, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 547, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 679 + ], + "type": "text", + "content": "[70] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "type": "text", + "content": "[71] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21750" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 665 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision, pages 11097-11106, 2021.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "text", + "content": "[72] Lixin Yang, Kailin Li, Xinyu Zhan, Fei Wu, Anran Xu, Liu Liu, and Cewu Lu. Oakink: A large-scale knowledge repository for understanding hand-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20953-20962, 2022. 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "text", + "content": "[73] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. AIM: Adapting image models for efficient video action recognition. In The Eleventh International Conference on Learning Representations, 2023. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "text", + "content": "[74] Jianglong Ye, Jiashun Wang, Binghao Huang, Yuzhe Qin, and Xiaolong Wang. Learning continuous grasping function with a dexterous hand from human demonstrations. IEEE Robotics and Automation Letters, 8(5):2882-2889, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 296 + ], + "type": "text", + "content": "[75] Ruolin Ye, Wenqiang Xu, Zhendong Xue, Tutian Tang, Yanfeng Wang, and Cewu Lu. H2o: A benchmark for visual human-human object handover analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15762-15771, 2021. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "text", + "content": "[76] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Is an object-centric video representation beneficial for transfer? In Proceedings of the Asian Conference on Computer Vision, pages 1976-1994, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 342, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 385 + ], + "type": "text", + "content": "[77] He Zhang, Yuting Ye, Takaaki Shiratori, and Taku Komura. Manipnet: neural manipulation synthesis with a handobject spatial representation. ACM Transactions on Graphics (ToG), 40(4):1-14, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 387, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 287, + 441 + ], + "type": "text", + "content": "[78] Hui Zhang, Sammy Christen, Zicong Fan, Luocheng Zheng, Jemin Hwangbo, Jie Song, and Otmar Hilliges. Artigrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. arXiv preprint arXiv:2309.03891, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 443, + 287, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 443, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 48, + 443, + 287, + 507 + ], + "type": "text", + "content": "[79] Juntian Zheng, Qingyuan Zheng, Lixing Fang, Yun Liu, and Li Yi. Cams: Canonicalized manipulation spaces for category-level functional hand-object manipulation synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 585-594, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "type": "text", + "content": "[80] Desen Zhou, Zhichao Liu, Jian Wang, Leshan Wang, Tao Hu, Errui Ding, and Jingdong Wang. Human-object interaction detection via disentangled transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19568-19577, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 566, + 287, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 287, + 618 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 287, + 618 + ], + "type": "text", + "content": "[81] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Toch: Spatio-temporal object-to-hand correspondence for motion refinement. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 621, + 287, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 621, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 48, + 621, + 287, + 665 + ], + "type": "text", + "content": "[82] Zehao Zhu, Jiashun Wang, Yuzhe Qin, Deqing Sun, Varun Jampani, and Xiaolong Wang. Contactart: Learning 3d interaction priors for category-level articulated object and hand poses estimation. arXiv preprint arXiv:2305.01618, 2023. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21751" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/f474f9b7-e45c-4a08-b76f-0b118ae61d7f_content_list.json b/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/f474f9b7-e45c-4a08-b76f-0b118ae61d7f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0c865372c8bf7fa0299941287edbaa06e82ebaaa --- /dev/null +++ b/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/f474f9b7-e45c-4a08-b76f-0b118ae61d7f_content_list.json @@ -0,0 +1,1585 @@ +[ + { + "type": "text", + "text": "TAMM: TriAdapter Multi-Modal Learning for 3D Shape Understanding", + "text_level": 1, + "bbox": [ + 117, + 128, + 851, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhihao Zhang $^{1*}$ Shengcao Cao $^{2*}$ Yu-Xiong Wang $^{2}$ $^{1}$ Xi'an Jiaotong University $^{2}$ University of Illinois Urbana-Champaign \n $^{1}$ zh1142@stu.xjtu.edu.cn $^{2}\\{cao44, yxw\\} @illinois.edu$", + "bbox": [ + 187, + 179, + 776, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/06b8d2bbf939a68fcb692a39846a4b7d07d46f557686a0edae21a12585c51933.jpg", + "image_caption": [ + "Figure 1. Our TriAdapter Multi-Modal Learning (TAMM) significantly enhances 3D shape understanding. Left: When aligning features of 3D point clouds (P) with 2D images (I) and/or text (T), prior methods (e.g., ULIP [45] and OpenShape [23]) under-utilize the image modality, due to the overlooked or unsolved image domain gap. TAMM better exploits the image modality and brings more gains when learning from both image and text data. The results are produced by pre-training Point-BERT [47] on ShapeNet [3]. Middle: Our CLIP Image Adapter (CIA) re-aligns the images rendered from 3D shapes with the text descriptions. The rendered images are inaccurately matched with text when the image features are directly extracted by CLIP, and CIA can correct the matching. Right: Our Image Alignment Adapter (IAA) and Text Alignment Adapter (TAA) decouple 3D features with complementary visual and semantic focuses. In the visualized examples, features from one single adapter are matched with classes whose appearance or semantics resemble the true class; using both adapters leads to the correct class." + ], + "image_footnote": [], + "bbox": [ + 84, + 280, + 385, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0db415b090078d7b48e45bd99d3a914a40d606ed9abf803a93a3ce4f554a738a.jpg", + "image_caption": [ + "Point Clouds" + ], + "image_footnote": [], + "bbox": [ + 398, + 287, + 647, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8b8971a677425e6d6c3627af7c8a303d46746dc4691e9e240a62c5c57323c367.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 281, + 764, + 347 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0c1fa06a9826da8db0e3dae11c6bc72bede6338ee24b14bd2a8e49dff4b56f03.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 280, + 870, + 349 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/ad808842328af6a33011e3b0c47c60829626199c1833883211d7a1d4dade7b01.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Matching w/o CIAa gray box with four screws on ita bedside table with a lamp
Matching w/ CIAa desk with drawers on ita bed with white sheet and a lamp
", + "bbox": [ + 398, + 357, + 647, + 410 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "IAA", + "bbox": [ + 658, + 357, + 681, + 368 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dress hat", + "bbox": [ + 714, + 359, + 756, + 369 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mallet", + "bbox": [ + 821, + 359, + 852, + 369 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "TAA", + "bbox": [ + 658, + 378, + 683, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lampshade", + "bbox": [ + 710, + 378, + 761, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "postbox", + "bbox": [ + 818, + 378, + 857, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dual", + "bbox": [ + 658, + 398, + 681, + 409 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mushroom", + "bbox": [ + 710, + 400, + 761, + 409 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "newsletter", + "bbox": [ + 818, + 400, + 857, + 409 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 585, + 313, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The limited scale of current 3D shape datasets hinders the advancements in 3D shape understanding, and motivates multi-modal learning approaches which transfer learned knowledge from data-abundant 2D image and language modalities to 3D shapes. However, even though the image and language representations have been aligned by cross-modal models like CLIP, we find that the image modality fails to contribute as much as the language in existing multi-modal 3D representation learning methods. This is attributed to the domain shift in the 2D images and the distinct focus of each modality. To more effectively leverage both modalities in the pre-training, we introduce TriAdapter Multi-Modal Learning (TAMM) - a novel two-stage learning approach based on three synergistic adapters. First, our CLIP Image Adapter mitigates the domain gap between 3D-rendered images and natural images, by adapting the visual representations of CLIP for synthetic image-text pairs. Subsequently, our Dual Adapters decou", + "bbox": [ + 73, + 618, + 470, + 891 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ple the 3D shape representation space into two complementary sub-spaces: one focusing on visual attributes and the other for semantic understanding, which ensure a more comprehensive and effective multi-modal pre-training. Extensive experiments demonstrate that TAMM consistently enhances 3D representations for a wide range of 3D encoder architectures, pre-training datasets, and downstream tasks. Notably, we boost the zero-shot classification accuracy on Objaverse-LVIS from $46.8\\%$ to $50.7\\%$ , and improve the 5-way 10-shot linear probing classification accuracy on ModelNet40 from $96.1\\%$ to $99.0\\%$ . Project page: https://alanzhangcs.github.io/tamm-page.", + "bbox": [ + 500, + 588, + 890, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 800, + 630, + 815 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the recent success of 3D shape understanding [16, 19, 24, 42, 47, 51, 56], the limited scale of existing 3D shape datasets hinders the learning of more robust and generalizable 3D representations. Due to the significant human labor and expertise required in collecting and annotating 3D", + "bbox": [ + 498, + 825, + 890, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 902, + 205, + 915 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21413", + "bbox": [ + 478, + 950, + 517, + 963 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shape data, building large-scale 3D shape datasets is a prohibitive endeavor [3, 9, 26].", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate the challenge of the limited data scale, recent research in 3D shape representation learning shows a promising direction of multi-modal learning [23, 45]. By establishing connections among data of 3D shapes, 2D images, and text, it becomes possible to pre-train 3D shape representations by distilling learned knowledge from image and language modalities. There exists a comprehensive body of literature of massive datasets [36], high-quality models pre-trained via self-supervised learning for 2D image [2, 4, 15] and language [1, 10, 39] modalities, and furthermore, cross-modal vision-language models [5, 34]. For instance, ULIP [45] creates triplets of 3D point clouds, 2D images, and text by rendering images from 3D shapes and generating text descriptions from their metadata. Then, ULIP adopts contrastive learning to align the 3D shape features with both the image features and text features extracted by CLIP [34].", + "bbox": [ + 75, + 123, + 470, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While the image and language modalities are pre-aligned by CLIP pre-training and share the same feature space, we observe that directly aligning 3D features with image features leads to considerably worse representation quality as compared with aligning 3D features with text features, as shown in Figure 1-left. This phenomenon suggests that existing methods (e.g., ULIP [45] and OpenShape [23]) are not optimally leveraging the 2D image modality in 3D representation pre-training.", + "bbox": [ + 75, + 382, + 468, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are two reasons for this counter-intuitive observation: 1) The 2D images in the generated triplets follow a data distribution different from natural images, on which CLIP is pre-trained. The images are rendered or projected from 3D to 2D, usually lacking a realistic background and texture. Since the image domain is shifted, the image features are no longer well-aligned with the text features, as exemplified in Figure 1-middle. 2) Intuitively, the image features represent more visual attributes including the shape, texture, or color, while the text features have a focus on semantics such as the function of the object. As shown in Figure 1-right, if the 3D features are specifically aligned with one single modality of images or text, they pay attention to different aspects of the 3D shape. Therefore, enforcing the 3D shape to simultaneously align with two modalities that convey subtly distinct information could be challenging. Such an approach may not fully leverage the potential of multi-modal learning signals.", + "bbox": [ + 75, + 520, + 468, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Aiming at addressing these two issues hindering multimodal pre-training for 3D shape understanding, we propose TriAdapter Multi-Modal Learning (TAMM), a two-stage pre-training approach based on three synergistic adapters. In the first stage, to mitigate the domain gap between the 2D images rendered from 3D shapes and the natural images on which CLIP is pre-trained, we adapt the visual representa", + "bbox": [ + 75, + 795, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions of CLIP based on the synthetic image-text pairs. More specifically, we fine-tune a lightweight CLIP Image Adapter on top of CLIP visual encoder through contrastive learning and re-align the adapted image features with the text features in the new domain. This CLIP Image Adapter allows us to establish more accurate relations between 3D shapes, 2D images, and language in an updated feature space, and avoids learning 3D representations from mismatched image features and text features.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the following second stage, to prevent the vision-semantic feature disparity from impairing our 3D representation pre-training, we choose to embrace this disparity and decouple the 3D representations into two sub-spaces. To comprehensively encode a 3D shape, the 3D encoder needs to capture both the visual and semantic aspects of its representation, which are centric to the corresponding image and text features, respectively. Therefore, we decouple the 3D feature space for the two focuses on visual and semantic representations. In particular, we attach two independent feature adapters to the 3D backbone and transform the 3D features into two sub-spaces. One sub-space focuses more on the visual representations, and is aligned with the 2D image feature space; the other sub-space focuses more on the semantic representations, and is aligned with the language feature space. This approach of decoupled feature spaces makes the learned 3D representations more comprehensive and expressive.", + "bbox": [ + 496, + 227, + 892, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our main contribution includes:", + "bbox": [ + 517, + 500, + 830, + 513 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify the under-utilization of the 2D image modality in existing multi-modal methods. The image domain gap and feature disparity in image-text pairs hinder representation learning in 3D shape understanding.", + "- We propose a novel multi-modal learning framework with two learning stages and three unified adapter modules. Our proposed TAMM better exploits both image and language modalities and improves 3D shape representations.", + "- Our TAMM consistently enhances 3D representations for a variety of 3D encoder architectures (e.g., PointBERT [47], SparseConv [6]), pre-training datasets (e.g., ShapeNet [3], an ensembled dataset [23]), and downstream tasks (e.g., zero-shot and linear probing shape classification on Objaverse-LVIS [9], ModelNet40 [43], and ScanObjectNN [40])." + ], + "bbox": [ + 500, + 513, + 890, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 753, + 640, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Shape Understanding. There are two mainstreams for 3D shape representation learning: 1) Projecting 3D shapes into voxel or grid-based formats [37] and then using 2D/3D convolutions [6] for feature extraction. 2) Directly modeling 3D point clouds with point-centric architectures [24, 28, 30, 31, 33, 42, 50, 52]. In this work, to ensure a fair and comprehensive comparison with previous methods [23, 45] on the pre-training scheme, we follow", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21414", + "bbox": [ + 478, + 950, + 519, + 964 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "their selection and utilize two representative 3D encoders from these two mainstreams: SparseConv [6] and Point-BERT [47]. SparseConv is designed for efficiently processing sparse voxels using specialized convolutions that focus computations on non-zero data points. Point-BERT [42] utilizes a Transformer-based architecture [41] and can be self-supervised by masked modeling [10].", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-Modal Representation Learning. Contrastive Language-Image Pre-training (CLIP) [34] has enabled various downstream applications including object detection [48, 53, 55] and language grounding [22]. Recently, CLIP has been extended to 3D-based tasks, such as zero-shot text-to-3D generation [17, 20, 25, 27, 38, 44] and scene-level 3D segmentation [14, 29, 35, 46]. Meanwhile, developing general and robust representations for 3D shape understanding with the foundation of CLIP [19, 23, 45, 51, 54, 56] becomes a major focus. Among these methods, ULIP [45], as a pioneering work, utilizes contrastive learning to distill CLIP features into 3D representations. OpenShape [23] follows this learning paradigm with a focus on building a larger pre-training dataset with enriched and filtered text data. Unlike OpenShape [23], we focus on improving the multi-modal learning paradigm by more effectively leveraging both the image and text modalities via a two-stage pre-training approach based on three synergistic adapters.", + "bbox": [ + 75, + 198, + 470, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 482, + 166, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 shows our proposed two-stage approach, TriAdapter Multi-Modal Learning (TAMM), for pretraining robust and generalizable 3D representations by leveraging 2D images and text. We first revisit the triplet data generation and problem formulation in multi-modal 3D representation learning. Section 3.1 delves into our finetuning of the original CLIP model for better fitting 3D understanding tasks. Section 3.2 details our strategy to enhance alignment across the 3D, 2D, and text feature spaces.", + "bbox": [ + 75, + 507, + 468, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation. Given $n$ triplets $\\{(P_i, I_i, T_i)\\}_{i=1}^n$ , where $P_i$ is a 3D point cloud, $I_i$ represents the corresponding image produced by projecting the 3D point cloud $P_i$ into 2D from an arbitrary perspective, and $T_i$ denotes the associated text generated using advanced vision-language models such as BLIP [21], the objective is to learn high-quality 3D representations from the triplets. The basic framework to achieve this objective is proposed by ULIP [45] (and followed by OpenShape [23]) with the help of CLIP [5, 34].", + "bbox": [ + 75, + 643, + 468, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, the 3D feature $f_{i}^{P} = E_{P}(P_{i})$ is produced by a learnable 3D encoder $E_{P}$ , and the corresponding image feature $f_{i}^{I} = E_{I}(I_{i})$ and text feature $f_{i}^{T} = E_{T}(T_{i})$ are generated by frozen CLIP encoders $E_{I}, E_{T}$ . Then the 3D encoder $E_{P}$ is optimized by aligning the 3D feature space $f^{P}$ simultaneously to the pre-aligned CLIP image space $f^{I}$ and text space $f^{T}$ through contrastive learning. The corresponding contrastive loss $L_{\\mathrm{contrast}}(f^{M_1}, f^{M_2})$ between two", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "modalities $M_{1}, M_{2}$ (3D-2D or 3D-text) is formulated as:", + "bbox": [ + 500, + 90, + 874, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n- \\frac {1}{2 n} \\sum_ {i = 1} ^ {n} \\left(\\log \\frac {\\exp \\left(f _ {i} ^ {M _ {1}} \\cdot f _ {i} ^ {M _ {2}} / \\tau\\right)}{\\sum_ {j = 1} ^ {n} \\exp \\left(f _ {i} ^ {M _ {1}} \\cdot f _ {j} ^ {M _ {2}} / \\tau\\right)} + \\log \\frac {\\exp \\left(f _ {i} ^ {M _ {2}} \\cdot f _ {i} ^ {M _ {1}} / \\tau\\right)}{\\sum_ {j = 1} ^ {n} \\exp \\left(f _ {i} ^ {M _ {2}} \\cdot f _ {j} ^ {M _ {1}} / \\tau\\right)}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 119, + 895, + 178 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ denotes the temperature hyperparameter.", + "bbox": [ + 500, + 180, + 826, + 195 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Image-Text Re-Alignment", + "text_level": 1, + "bbox": [ + 500, + 208, + 735, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike ULIP [45] or OpenShape [23], in which the 3D feature space is aligned with the image-text feature spaces directly produced by CLIP, we introduce an image feature space tuning strategy, aiming to foster better alignment across 3D, 2D, and language modalities (Figure 2-left). We argue that the image feature space produced directly by the CLIP visual encoder does not perfectly align with the text feature space and is thus sub-optimal. The reason is that the 2D images in the triplets, which originate from 3D point cloud projections, lack backgrounds and are visually different from natural images on which CLIP is pre-trained.", + "bbox": [ + 496, + 231, + 890, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, when using such 2D images from a shifted data domain to perform multi-modal pre-training, it becomes necessary to further fine-tune CLIP and re-align its image and language feature spaces. For the first time, we redesign domain adapters [18] for multi-modal contrastive 3D representation learning, and propose a CLIP Image Adapter (CIA), to adapt the image feature space for the rendered 2D images in our data triplets. We append a lightweight, learnable multi-layer perceptron (MLP) to CLIP image encoder, with a residual connection which seamlessly integrates the new knowledge acquired from fine-tuning with the existing knowledge from the pre-trained CLIP backbone. To avoid heavy computation and over-fitting, we only fine-tune the additional parameters in CIA, instead of the whole CLIP backbone. Formally, given the image feature $f_{i}^{I}$ and text feature $f_{i}^{T}$ extracted from the triplet $(P_{i}, T_{i}, I_{i})$ , we use a learnable, two-layer CIA $A_{C}(\\cdot)$ to adapt the image feature, formulated as:", + "bbox": [ + 496, + 398, + 892, + 671 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA _ {C} \\left(f _ {i} ^ {I}\\right) = \\sigma \\left(f _ {i} ^ {I} W _ {1}\\right) \\cdot W _ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 685, + 890, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $W_{1}$ and $W_{2}$ are the parameters associated with the linear transformation layers, and $\\sigma$ is the non-linear activation function. The refined image feature $\\widetilde{f}_i^I$ can be computed with a residual connection:", + "bbox": [ + 496, + 717, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {f} _ {i} ^ {I} = \\alpha A _ {C} \\left(f _ {i} ^ {I}\\right) + (1 - \\alpha) f _ {i} ^ {I}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 792, + 890, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha$ is a hyperparameter. Finally, CIA $A_{C}(\\cdot)$ is optimized by minimizing the contrastive loss function (Equation 1), instantiated as:", + "bbox": [ + 496, + 824, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e a l i g n}} = L _ {\\text {c o n t r a s t}} \\left(\\widetilde {f} ^ {I}, f ^ {T}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 882, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21415", + "bbox": [ + 478, + 950, + 517, + 963 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/60b0520307b5ea6bb9dc2699ea5f4f545b27bc6b2185bd26c63d18b5b0249748.jpg", + "image_caption": [ + "Stage 1: Image-Text Re-Alignment", + "Figure 2. Overview of TAMM. Left: In Stage 1, TAMM fine-tunes a lightweight CLIP Image Adapter (CIA) through contrastive learning and re-aligns the image features with the text features to alleviate the domain shift originated from rendered images. Contrastive learning maximizes inner products between features from corresponding text-image pairs, and reduces similarities of mismatched pairs. Middle: In Stage 2, TAMM introduces Image Alignment Adapter (IAA) and Text Alignment Adapter (TAA) to decouple 3D representations into two sub-spaces: one focusing more on visual attributes and the other for semantic understanding, ensuring a more comprehensive and effective multi-modal pre-training strategy. Right: TAMM adaptively utilizes decoupled 3D features for various downstream tasks including linear probing classification (top) and zero-shot classification (bottom), achieving more robust classification results." + ], + "image_footnote": [], + "bbox": [ + 89, + 99, + 313, + 391 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c46b3fef7b9e42f88f2ce12b763efc58cbfc061cc4ede2b6e2bf236b8597ce8a.jpg", + "image_caption": [ + "Stage 2: Decoupled Tri-Modal Pre-Training" + ], + "image_footnote": [], + "bbox": [ + 321, + 101, + 890, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Decoupled Tri-Modal Pre-Training", + "text_level": 1, + "bbox": [ + 76, + 515, + 382, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the image and language feature spaces have been aligned by CLIP and our CLIP Image Adapter, they still encode subtly different information. For instance, the color of a 3D shape may be inferred from the image feature, but it cannot possibly be included in the text feature if the text description does not contain color information. Similarly, the image feature may not include semantic information such as the function or name of an object. In such cases, enforcing the 3D feature to align with both the image and text features simultaneously is challenging.", + "bbox": [ + 75, + 545, + 468, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to overcome the obstacle introduced by aligning the 3D shape feature space with two distinct modalities, we propose a novel decoupled tri-modal pre-training framework, which aligns two decoupled 3D shape feature spaces with the refined image feature space and the text feature space, respectively (Figure 2-middle). We encourage the 3D encoder to cover both the visual representation and the semantic information inherent to the 3D shape, avoiding the dilemma of aligning with disparate image and text features at the same time. Formally, given the triplets $\\{(P_i,I_i,T_i)\\}_{i = 1}^n$ , we first generate the image feature $\\widetilde{f}_i^I$ using the frozen, adapted 2D image encoder (Equation 3) and the text feature $f_{i}^{T}$ using the original CLIP", + "bbox": [ + 75, + 702, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "text encoder. We introduce a pair of lightweight Dual Adapters: Image Alignment Adapter (IAA) $A_V(\\cdot)$ and Text Alignment Adapter (TAA) $A_S(\\cdot)$ . They split the 3D feature $f_i^P$ originated from the 3D encoder $E_P$ into a vision-focusing feature $f_i^{VP}$ and a semantic-focusing feature $f_i^{SP}$ , respectively. Dual Adapters $A_V(\\cdot)$ and $A_S(\\cdot)$ share the same architecture as the CLIP Image Adapter in Section 3.1 and can be formulated as:", + "bbox": [ + 496, + 516, + 890, + 636 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {V P} = A _ {V} \\left(f _ {i} ^ {P}\\right) = \\sigma \\left(f _ {i} ^ {P} W _ {1} ^ {V}\\right) \\cdot W _ {2} ^ {V}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 640, + 890, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {S P} = A _ {S} (f _ {i} ^ {P}) = \\sigma (f _ {i} ^ {P} W _ {1} ^ {S}) \\cdot W _ {2} ^ {S},\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 661, + 823, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_1^V, W_2^V, W_1^S,$ and $W_2^S$ are the parameters associated with the linear layers, and $\\sigma$ represents the activation function. By decoupling 3D features into these sub-spaces, the 3D encoder $E_P$ interprets the 3D shape with a more comprehensive visual and semantic understanding, improving the expressivity of the learned representations.", + "bbox": [ + 496, + 686, + 890, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, instead of enforcing the 3D shape feature space to directly mimic the pre-aligned image-text feature space, we use the decoupled 3D features $f^{VP}, f^{SP}$ and align them with the adapted image feature space $\\widetilde{f}^T$ and the text feature space $f^T$ , respectively. Moreover, since one single image can only capture the 3D shape from one perspective, we align the 3D feature with the adapted features of multi-view images, to fully exploit the image modality and achieve a", + "bbox": [ + 496, + 777, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21416", + "bbox": [ + 478, + 950, + 519, + 963 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "better alignment between 3D and 2D. The overall loss function is defined as:", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r i m o d a l}} = L _ {\\text {c o n t r a s t}} \\left(f ^ {S P}, f ^ {T}\\right) + \\frac {1}{m} \\sum_ {k} ^ {m} L _ {\\text {c o n t r a s t}} \\left(f ^ {V P}, \\widetilde {f} ^ {I, k}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 125, + 468, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $m$ represents the number of rendered images and $\\widetilde{f}^{I,k}$ is the adapted image feature from the $k$ -th view.", + "bbox": [ + 76, + 165, + 468, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Application in Downstream Tasks. The learned 3D feature sub-spaces, $f^{VP}$ and $f^{SP}$ can be adaptively applied to a variety of downstream tasks (Figure 2-right). Specifically, in zero-shot 3D classification, we leverage both the 3D vision-focusing feature $f^{VP}$ and the 3D semantic-focusing feature $f^{SP}$ . We calculate the similarity between these features and category embeddings generated by the CLIP text encoder, respectively. After summing up the per-category similarity scores, the category with the highest similarity is chosen as the predicted class, yielding more robust and enhanced classification results compared with using a single sub-space alone. In the linear probing classification task, we concatenate the 3D vision-focusing feature $f^{VP}$ and 3D semantic-focusing feature $f^{SP}$ as input to the learnable linear classification layer, providing a more comprehensive and robust representation for 3D understanding.", + "bbox": [ + 76, + 196, + 470, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 452, + 209, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-Training Datasets. Following the prior state-of-the-art method, OpenShape [23], our TAMM is pre-trained on the triplets generated from four datasets: ShapeNetCore [3], 3D-FUTURE [12], ABO [7], and Objaverse [9]. Our training sets are defined as follows: \"ShapeNet\" is a triplet set derived exclusively from the ShapeNetCore dataset, containing 52,470 3D shapes with corresponding images and text; \"Ensembled (no LVIS)\" is a set of 829,460 triplets from the above datasets excluding Objaverse-LVIS; \"Ensembled\" denotes the triplet set comprising data from all four datasets, containing 875,665 3D shapes and their associated images and text.", + "bbox": [ + 76, + 477, + 468, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Datasets. Our TAMM is evaluated on the following datasets: Objaverse-LVIS [9], ModelNet40 [43], ScanObjectNN [40], and ScanNet [8]. Objverse-LVIS encompasses a broad range of categories, featuring 46,832 high-quality shapes distributed across 1,156 LVIS [13] categories. ModelNet40 is a synthetic indoor 3D dataset comprising 40 categories. We use the test split of 2,468 shapes in our experiments. ScanObjectNN consists of scanned objects from 15 common categories. It has three main variants: OBJ-BG, OBJ-ONLY, and PB-T50-RS. ScanNet, characterized by its real-world scans, includes 1,513 indoor scenes containing 36,213 objects. We conduct experiments on four tasks including zero-shot 3D classification, linear probing 3D classification, few-shot linear probing 3D classification, and real-world recognition, to demonstrate the advantages of our TAMM. Other implementation details re", + "bbox": [ + 76, + 659, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "garding pre-training and evaluation are introduced in the supplementary material.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Zero-Shot 3D Classification", + "text_level": 1, + "bbox": [ + 500, + 131, + 748, + 147 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Zero-shot ability is a key metric for reflecting the quality of learned 3D representations, which requires the representations to be directly applicable to datasets where the model has never been explicitly supervised. Without any further tuning, the 3D representations are compared with text embeddings of categories to predict the classes of 3D shapes. To make a fair comparison and keep consistency with prior work, we adopt the same settings as OpenShape [23] on three benchmarks: Objaverse-LVIS [9], ModelNet40 [43], and OBJ-ONLY (ScanObjectNN) [40].", + "bbox": [ + 496, + 155, + 893, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The results are summarized in Table 1. First, we observe that TAMM, benefited from multi-modal pre-training, outperforms PointCLIP [51] and PointCLIP v2 [56] by a large margin. Furthermore, it is evident that our pretrained models consistently outperform those pre-trained by ULIP and OpenShape, irrespective of whether they are pretrained on smaller datasets like ShapeNet or more expansive ones such as the Ensembled dataset. For instance, PointBERT [47] pre-trained by our TAMM on the Ensembled (no LVIS) dataset surpasses ULIP and OpenShape by margins of $+20.6\\%$ and $+2.9\\%$ in Top-1 accuracy on the long-tailed Objaverse-LVIS benchmark, which validates the effectiveness of our multi-modal pre-training scheme.", + "bbox": [ + 496, + 306, + 892, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Linear Probing 3D Classification", + "text_level": 1, + "bbox": [ + 500, + 513, + 787, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further evaluate the 3D representation quality of our pre-trained models, we design and conduct linear probing classification experiments. Here, we first freeze our pretrained Point-BERT model along with the Image Alignment Adapter (IAA) and Text Alignment Adapter (TAA), and then append a single learnable linear layer to the model. Given a batch of point clouds, as illustrated in Figure 2, we generate features from both IAA and TAA and concatenate them, and learn the appended linear classification layer on the concatenated features. We evaluate the accuracy of the linear classifier on three benchmarks: Objverse-LVIS [9], ModelNet40 [43], and ScanObjectNN [40]. Objverse-LVIS is a challenging long-tailed dataset with 1,156 categories, which has not been evaluated by prior methods [23, 45]. To ensure evaluation with statistically meaningful number of samples per class, we exclude categories with fewer than 10 instances, leaving us 1,046 categories. Samples in each class are divided into training and testing sets at an $8:2$ ratio. For ModelNet40 and ScanObjectNN, we use the standard splits, following [32, 42].", + "bbox": [ + 496, + 537, + 890, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The results are shown in Table 2. TAMM consistently outperforms all previous methods by a large margin. For example, pre-trained on the Ensembled dataset, TAMM improves over OpenShape by $+11.2\\%$ and $+2.2\\%$ overall ac", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21417", + "bbox": [ + 478, + 950, + 519, + 964 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/292f3cd4db83329b3f2564ec898f39cb2603b94bfd9fa6141352cbb8a8af60ac.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pre-Training DatasetModelPre-Training MethodObjaverse-LVIS [9]ModelNet40 [43]ScanObjectNN [40]
Top-1Top-3Top-5Top-1Top-3Top-5Top-1Top-3Top-5
-PointCLIP [51]-1.94.15.819.328.634.810.520.830.6
-PointCLIP v2 [56]-4.79.512.963.677.985.042.263.374.5
ShapeNetViT-B/32 [11]CLIP2Point [19]2.75.87.949.571.381.225.544.659.4
Transformer [41]ReCon [32]1.12.73.761.273.978.142.362.575.6
SparseConv [6]OpenShape [23]11.621.827.172.987.293.052.772.783.6
TAMM (Ours)13.624.229.374.688.294.057.975.383.1
Point-BERT [47]ULIP [45]6.213.617.960.479.084.451.571.180.2
OpenShape [23]10.820.225.070.386.991.351.369.478.4
TAMM (Ours)13.724.229.273.188.591.954.874.583.3
Ensembled (no LVIS)SparseConv [6]OpenShape [23]37.058.466.982.695.097.554.976.887.0
TAMM (Ours)39.862.070.485.796.898.357.581.390.0
Point-BERT [47]ULIP [45]21.438.146.071.484.489.246.066.176.4
OpenShape [23]39.160.868.985.396.297.447.272.484.7
TAMM (Ours)42.063.671.786.396.698.156.778.386.1
EnsembledSparseConv [6]OpenShape [23]43.464.872.483.495.697.856.778.988.6
TAMM (Ours)43.866.274.185.496.498.158.581.389.5
Point-BERT [47]ULIP [45]26.844.852.675.188.193.251.672.582.3
OpenShape [23]46.869.177.084.496.598.052.279.788.7
TAMM (Ours)50.773.280.685.096.698.155.780.788.9
", + "bbox": [ + 101, + 74, + 872, + 397 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1367565047a26aa1a5cf5a5c7e70cdfbb51a58040b66d26422b7e8360d4a4da8.jpg", + "table_caption": [ + "Table 1. Zero-shot 3D classification results. TAMM sets new state of the art in zero-shot classification accuracy across Objaverse-LVIS, ModelNet-40, and ScanObjectNN benchmarks, outperforming existing methods in diverse settings of pre-training datasets and 3D model architectures. The performance gain is more significant on the most challenging long-tailed Objverse-LVIS dataset." + ], + "table_footnote": [], + "table_body": "
Pre-Training DatasetMethodO-LVIS [9]M-40 [43]ScanObjectNN [40]
OBJ-BGOBJ-ONLYPB-T50-RS
ShapeNetULIP [45]34.690.675.475.464.8
OpenShape [23]29.388.577.878.564.1
TAMM (Ours)39.191.080.681.168.5
Rel. Improv.+9.8+2.5+2.8+2.6+4.4
EnsembledOpenShape [23]48.391.385.985.478.0
TAMM (Ours)59.593.588.588.080.3
Rel. Improv.+11.2+2.2+2.6+2.6+1.7
", + "bbox": [ + 78, + 465, + 470, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "curacy on Objaverse-LVIS and ModelNet40, respectively. This demonstrates that TAMM effectively exploits knowledge from CLIP and learns generalizable 3D representations that are directly applicable to novel linear classification tasks. Notably, TAMM even surpasses the Point-BERT model pre-trained without multi-modality but allowed to be fully fine-tuned on OBJ-BG (87.4%) [47] by +1.1% accuracy. These strong results indicate that TAMM has learned excellent transferable 3D representations, showing a great potential in real-world applications.", + "bbox": [ + 73, + 665, + 468, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Few-Shot Linear Probing 3D Classification", + "text_level": 1, + "bbox": [ + 75, + 830, + 444, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the previous evaluation, we perform few-shot classification experiments on ModelNet40 [43] to assess TAMM in low-data scenarios. Similar to the linear probing", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5ab40ebaeae016bbba2dbc9e0eb4e9996c94c192d75c9937aa691ba8edd12048.jpg", + "table_caption": [ + "Table 2. Linear probing 3D classification results. TAMM outperforms previous methods by a large margin, e.g., +11.2% accuracy gain on the challenging Objaverse-LVIS dataset." + ], + "table_footnote": [], + "table_body": "
Pre-Training DatasetMethod5-way10-way
10-shot20-shot10-shot20-shot
ShapeNetULIP [45]94.4±3.793.2±4.286.6±5.390.6±5.2
OpenShape [23]95.3±2.697.9±3.989.2±5.192.9±3.9
TAMM (Ours)97.8±1.998.1±1.495.3±3.995.9±2.6
EnsembledOpenShape [23]96.1±2.795.7±2.589.1±4.691.8±3.7
TAMM (Ours)99.0±1.399.4±0.796.8±2.997.4±2.2
", + "bbox": [ + 500, + 465, + 890, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Few-shot linear probing classification results on ModelNet40. We report the average accuracy and standard deviation of 10 independent experiments. Our TAMM consistently achieves both the best average accuracy and the lowest variance in various few-shot settings.", + "bbox": [ + 496, + 574, + 893, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "experiments described in Section 4.2, we extend our model with an additional linear classification layer and only train this linear layer instead of fine-tuning the entire model. Following previous work [47, 50], we adopt the \" $K$ -way $N$ -shot\" configuration, wherein we select $K$ classes at random and sample $(N + 20)$ instances per class. Training is conducted on a support set of $K \\times N$ samples, while evaluation is based on a query set comprised of the remaining 20 instances per class. We assess our model under four distinct scenarios: \"5-way 10-shot,\" \"5-way 20-shot,\" \"10-way 10-shot,\" and \"10-way 20-shot.\" For each scenario, we carry out 10 separate trials, and report both the mean performance and the standard deviation across these trials.", + "bbox": [ + 496, + 670, + 890, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Table 3, our TAMM achieves the best results and sets a new state of the art in the few-shot clas-", + "bbox": [ + 498, + 869, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21418", + "bbox": [ + 478, + 950, + 519, + 964 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8e20e87433acfc8ebe88008e98734b11a3ba3e9040ded5d0d9222a45c1d42c06.jpg", + "table_caption": [], + "table_footnote": [ + "w/ TP. denotes training with the real-world data provided by $\\mathrm{CLIP^2}$ . Results using Point-BERT [47] as 3D encoder, pre-trained on the Ensembled dataset." + ], + "table_body": "
MethodAvg.BedCabChairSofaTablDoorWindBksfPicCntrDeskCurtFridgBathShowrToilSink
CLIP2Point [19]24.920.80.085.143.326.569.90.020.91.731.727.00.01.646.50.022.425.6
PointCLIP w/ TP. [56]26.10.055.772.85.05.11.70.077.20.00.051.70.30.00.040.385.349.2
CLIP2Point w/ TP. [19]35.211.83.045.127.610.561.52.671.90.333.629.94.711.572.292.486.134.0
\\( CLIP^2 \\) [49]38.532.667.269.342.318.319.14.062.61.412.752.840.19.159.741.071.045.5
\\( OpenShape^\\dagger \\) [23]45.666.73.275.883.537.749.247.564.948.21.966.170.21.850.057.145.27.1
TAMM (Ours)†49.466.74.883.684.548.957.948.280.561.31.960.683.67.041.456.148.43.6
", + "bbox": [ + 81, + 74, + 890, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0943b63de7bc751a234e24160e35b69d6b0a97dfef2749d95f85dee59b1af474.jpg", + "table_caption": [ + "Table 4. Zero-shot classification results on the real-world ScanNet dataset. Avg.: the mean average Top-1 accuracy of all classes. TAMM achieves the best results." + ], + "table_footnote": [ + "(a) Adapters in pre-training. Pre-training with CLIP Image Adapter (CIA), Image Alignment Adapter (IAA), and Text Alignment Adapter (TAA) is the most effective." + ], + "table_body": "
CIAIAATAAImg-Txt AccObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
✓✓40.113.523.929.172.888.291.7
✓60.912.522.427.371.486.089.8
✓✓60.912.922.927.773.888.292.5
✓✓60.913.023.028.272.987.890.4
✓✓✓60.913.724.229.273.188.591.9
", + "bbox": [ + 127, + 242, + 455, + 328 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fb4ff0526aaa532992d526cfcdd83524c0072d1b9b4658bca672be83d635444b.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Pre-training stages. Learning the CLIP Image Adapter first and then pre-train the 3D encoder with Dual Adapters is better than training modules all together." + ], + "table_body": "
StageImg-Txt AccObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
One stage55.312.221.427.071.486.490.7
Two stages60.913.724.229.273.188.591.9
", + "bbox": [ + 527, + 258, + 841, + 310 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/81964dee1778efb4ba69942001c0251c3e81f4672b4a743df556a6647d90c65e.jpg", + "table_caption": [], + "table_footnote": [ + "(c) Pre-training modalities. Exploiting both image and text data provides the most performance gain. They contribute almost equally in TAMM." + ], + "table_body": "
ImageTextObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
✓11.920.825.967.687.392.5
✓10.519.223.770.285.088.0
✓✓13.724.229.273.188.591.9
", + "bbox": [ + 76, + 375, + 341, + 438 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ac7d97b09543ac02fa5b609b5a07280f0f0aa3e5c005fdc680c44bbacc25ebf0.jpg", + "table_caption": [], + "table_footnote": [ + "(d) Adapters for inference. Dual Adapters are complementary, and their combination more accurately classifies 3D shapes at test time." + ], + "table_body": "
IAATAAObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
✓13.023.128.368.186.591.4
✓12.922.527.572.486.290.4
✓✓13.724.229.273.188.591.9
", + "bbox": [ + 359, + 375, + 625, + 439 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4edafad04bc06e908738e340fc13fd7a8de76663db70f89feb4f6d1a99d46ace.jpg", + "table_caption": [], + "table_footnote": [ + "(e) Number of images. Aligning a 3D shape with 4 images of it is the best." + ], + "table_body": "
ImagesObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
113.623.628.972.588.092.6
213.523.528.873.187.692.8
413.724.229.273.188.591.9
813.523.128.473.089.092.8
", + "bbox": [ + 643, + 375, + 890, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation study of components in TAMM. We pre-train Point-BERT models on ShapeNet in various settings and evaluate their zero-shot classification performance on Objaverse-LVIS and ModelNet40. The baseline method OpenShape [23] achieves $10.8\\%$ and $70.3\\%$ Top-1 accuracy on these two benchmarks (Table 1), respectively. The setting adopted by TAMM is marked.", + "bbox": [ + 75, + 489, + 892, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sification task. TAMM shows both higher overall accuracy and smaller deviations than other methods, showcasing the generalizability and robustness of TAMM-learned 3D representations. For example, when pre-trained on the Ensembled dataset, our TAMM surpasses OpenShape [23] by $2.9\\%$ , $3.7\\%$ , $7.7\\%$ , $5.6\\%$ , respectively for all four settings. These results indicate that our TAMM is able to learn 3D representations that are more generalizable and can be readily adapted to downstream tasks under low-data regimes.", + "bbox": [ + 75, + 551, + 468, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Real-World Recognition", + "text_level": 1, + "bbox": [ + 76, + 696, + 299, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the capability of TAMM in understanding $3D$ shapes from the real world, we follow the previous work $\\mathrm{CLIP}^2$ [49] and test TAMM on a real-world recognition task, in which the model aims at correctly classifying each instance from a complex scene in a zero-shot manner. Specifically, we select the real-world scene-level dataset ScanNet [8] and adopt same data splits as [49], containing 17 classes. We perform zero-shot classification (Section 4.1) on the point cloud of each object instance extracted from scenes, and report both the Top-1 accuracy of each class and the mean accuracy of all 17 classes. As shown in Table 4, TAMM significantly outperforms all prior meth", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ods, showing improvements of $3.8\\%$ and $10.9\\%$ in average Top-1 accuracy compared with OpenShape and $\\mathrm{CLIP^2}$ , respectively. These results underscore TAMM's ability in recognizing and understanding 3D shapes captured from real-world scenarios. Further results on complex scene recognition and instance segmentation are included in the supplementary material.", + "bbox": [ + 496, + 551, + 890, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 665, + 653, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we provide additional experimental results to further test the performance gain from each design of TAMM. Due to limited computation, we pre-train PointBERT [47] on the small-scale ShapeNet [3], and use two more challenging zero-shot classification benchmarks ModelNet40 [43] and Objaverse-LVIS [9] for evaluation. We alter only one component in each set of experiments.", + "bbox": [ + 496, + 688, + 890, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adapters in Pre-Training. As described in Section 3.1, our CLIP Image Adapter (CIA) mitigates the domain gap between rendered images and natural images. We first measure the contrastive accuracy, computed as the ratio of image-text pairs where the text feature is more similar to the image from the same triplet than any other text. The contrastive accuracy of CLIP without CIA is only $40.1\\%$ ,", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21419", + "bbox": [ + 478, + 950, + 519, + 963 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "indicating that the original image-text feature space is compromised by the domain gap. CIA brings the accuracy up to $60.9\\%$ , showing the effectiveness of our CIA in resolving the shifted data domain. Furthermore, as shown in Table 5a, CIA increases the zero-shot accuracy on Objaverse-LVIS dataset from $12.2\\%$ to $13.7\\%$ , validating its effectiveness on improving the pre-trained 3D representations.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TAMM adopts Image Alignment Adapter (IAA) and Text Alignment Adapter (TAA) to decouple the feature space and enrich the learned 3D representations. To further explore the performance gains introduced by our Dual Adapters (IAA and TAA), we experiment tri-modal pretraining with or without IAA and TAA. As shown in Table 5a, both adapters bring performance gains, and their combination achieves the best result, demonstrating the effectiveness of Dual Adapters.", + "bbox": [ + 75, + 198, + 470, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pre-Training Stages. We investigate the significance of our two-stage pre-training design, which initially learns CLIP Image Adapter (CIA) to re-align image-text pairs and subsequently employs Dual Adapters in decoupled tri-modal pre-training. An alternative could be learning these modules all together in one stage. As shown in Table 5b, remarkably, the two-stage pre-training approach achieves a better performance compared with one-stage pre-training.", + "bbox": [ + 75, + 335, + 470, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pre-Training Modalities. We further explore the effectiveness of image and language modalities in pre-training. Table 5c shows that integrating the 3D modality with both the image and language modalities consistently yields superior performance across various benchmarks, compared with aligning it with either modality alone. Moreover, unlike prior methods which ineffectively learn from images (Figure 1-left), the image and language modalities contribute almost equally in TAMM, which also signifies that TAMM better exploits the image modality and brings more gains by learning from both image and text data.", + "bbox": [ + 75, + 459, + 468, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adapters for Inference. Our IAA and TAA decouple 3D features with visual and semantic focuses, respectively, offering a more accurate and comprehensive understanding for 3D shapes. At inference time, we combine the outputs from Dual Adapters for classification tasks. To investigate this design, we separately utilize features produced solely by either IAA or TAA in zero-shot classification evaluation. As shown in Table 5d, the combined features achieve more accurate classification, demonstrating that the learned features from IAA and TAA are complementary to each other.", + "bbox": [ + 75, + 627, + 468, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of Images. In order to gain more comprehensive knowledge from the image modality, TAMM simultaneously aligns the 3D feature with multi-view 2D images, each projected from a different perspective. We evaluate this design by aligning the 3D feature with a varying number of corresponding 2D images, ranging from 1 to 8 images. As shown in Table 5e, multi-view image features offer a performance gain by fully exploiting the image modality.", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd8efb0de0c4c7191a532931e2cf86427657a9ef603bd873c0deaaf26c026870.jpg", + "image_caption": [ + "Figure 3. Qualitative results. Top: CIA re-aligns the images rendered from 3D shapes with the text descriptions. The rendered images are inaccurately matched with text when the image features are directly extracted by CLIP, and CIA can correct the matching. Bottom: IAA and TAA decouple 3D features with complementary visual and semantic focuses. Features from one single adapter are matched with classes whose appearance or semantics resemble the true class; using both adapters leads to the correct class." + ], + "image_footnote": [], + "bbox": [ + 506, + 79, + 893, + 309 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Aligning a 3D shape with 4 views achieves the best result.", + "bbox": [ + 496, + 452, + 887, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results. Finally, we provide some visualizations to intuitively demonstrate the benefit of our proposed TAMM. As illustrated in Figure 3, CIA successfully mitigates the domain gap introduced by rendered images, and leads to the more accurate matching between images and text. IAA and TAA learn 3D representations with subtly different but complementary focuses on vision and semantics, respectively, and their combination brings more robust and comprehensive 3D representations. More visualized results are included in the supplementary material.", + "bbox": [ + 496, + 467, + 890, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 633, + 619, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we examine the sub-optimal utilization of 2D images in existing multi-modal pre-training methods for 3D shape understanding, and propose TriAdapter MultiModal Learning (TAMM), a novel two-stage representation learning approach built on three synergistic adapter modules. Extensive experiments verify that TAMM consistently learns improved 3D features in various settings.", + "bbox": [ + 496, + 657, + 890, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Due to the limited computation resources, we are not able to perform pre-training on very large-scale 3D backbones with billions of parameters. The quality of the learned 3D representation could be further improved if the 3D backbones are scaled to a larger size.", + "bbox": [ + 496, + 765, + 890, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Jump ARCHES endowment, and the IBM-Illinois Discovery Accelerator Institute. This work used NVIDIA GPUs at NCSA Delta through allocations CIS220014, CIS230012, and CIS230013 from the ACCESS program.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21420", + "bbox": [ + 478, + 950, + 519, + 964 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 2", + "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2", + "[3] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 2, 5, 7", + "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2", + "[5] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 2, 3", + "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D spatio-temporal ConvNets: Minkowski convolutional neural networks. In CVPR, 2019. 2, 3, 6", + "[7] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F. Yago Vicente, Thomas Dideriksien, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. ABO: Dataset and benchmarks for real-world 3D object understanding. In CVPR, 2022. 5", + "[8] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 5, 7", + "[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, 2023. 2, 5, 6, 7", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2, 3", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 6" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3D-FUTURE: 3D furniture shape with texture. IJCV, 129:3313-3337, 2021. 5", + "[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 5", + "[14] Huy Ha and Shuran Song. Semantic abstraction: Openworld 3D scene understanding from 2D vision-language models. In CoRL, 2022. 3", + "[15] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2", + "[16] Deepti Hegde, Jeya Maria Jose Valanarasu, and Vishal Patel. CLIP goes 3D: Leveraging prompt tuning for language grounded 3D recognition. In ICCV, 2023. 1", + "[17] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. AvatarCLIP: Zero-shot text-driven generation and animation of 3D avatars. ACM Transactions on Graphics, 41(4):1-19, 2022. 3", + "[18] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. In ICML, 2019. 3", + "[19] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. CLIP2Point: Transfer CLIP to point cloud classification with image-depth pre-training. In ICCV, 2023. 1, 3, 6, 7", + "[20] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, 2022. 3", + "[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 3", + "[22] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 3", + "[23] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. OpenShape: Scaling up 3D shape representation towards open-world understanding. In NeurIPS, 2023. 1, 2, 3, 5, 6, 7", + "[24] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In ICLR, 2022. 1, 2", + "[25] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2Mesh: Text-driven neural stylization for meshes. In CVPR, 2022. 3", + "[26] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In CVPR, 2019. 2", + "[27] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. CLIP-Mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia, 2022. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21421", + "bbox": [ + 478, + 952, + 517, + 963 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 2", + "[29] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. OpenScene: 3D scene understanding with open vocabularies. In CVPR, 2023. 3", + "[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In CVPR, 2017. 2", + "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. PointNet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2", + "[32] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3D representation learning guided by generative pretraining. In ICML, 2023. 5, 6", + "[33] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. PointNeXt: Revisiting PointNet++ with improved training and scaling strategies. In NeurIPS, 2022. 2", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3", + "[35] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3D semantic segmentation in the wild. In ECCV, 2022. 3", + "[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: An open large-scale dataset for training next generation image-text models. In NeurIPS Datasets and Benchmarks Track, 2022. 2", + "[37] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. PV-RCNN: Point-voxel feature set abstraction for 3D object detection. In CVPR, 2020. 2", + "[38] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. MotionCLIP: Exposing human motion generation to CLIP space. In ECCV, 2022. 3", + "[39] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[40] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 2, 5, 6", + "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3, 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point Transformer V2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 1, 2, 3, 5", + "[43] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In CVPR, 2015. 2, 5, 6, 7", + "[44] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3D: Zero-shot text-to-3D synthesis using 3D shape prior and text-to-image diffusion models. In CVPR, 2023. 3", + "[45] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. ULIP: Learning a unified representation of language, images, and point clouds for 3D understanding. In CVPR, 2023. 1, 2, 3, 5, 6", + "[46] Jihan Yang, Runyu Ding, Zhe Wang, and Xiaojuan Qi. RegionPLC: Regional point-language contrastive learning for open-world 3D scene understanding. arXiv preprint arXiv:2304.00962, 2023. 3", + "[47] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-BERT: Pre-training 3D point cloud transformers with masked point modeling. In CVPR, 2022. 1, 2, 3, 5, 6, 7", + "[48] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In CVPR, 2021. 3", + "[49] Yihan Zeng, Chenhan Jiang, Jiageng Mao, Jianhua Han, Chaoqiang Ye, Qingqiu Huang, Dit-Yan Yeung, Zhen Yang, Xiaodan Liang, and Hang Xu. CLIP $^2$ : Contrastive language-image-point pretraining from real-world point cloud data. In CVPR, 2023. 7", + "[50] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. PointM2AE: Multi-scale masked autoencoders for hierarchical point cloud pre-training. In NeurIPS, 2022. 2, 6", + "[51] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. PointCLIP: Point cloud understanding by CLIP. In CVPR, 2022. 1, 3, 5, 6", + "[52] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3D features on any point-cloud. In ICCV, 2021. 2", + "[53] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. RegionCLIP: Region-based language-image pretraining. In CVPR, 2022. 3", + "[54] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3D: Exploring unified 3D representation at scale. In ICLR, 2024. 3", + "[55] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 3", + "[56] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Point-" + ], + "bbox": [ + 501, + 90, + 890, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21422", + "bbox": [ + 478, + 952, + 519, + 963 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "CLIP V2: Prompting CLIP and GPT for powerful 3D open-world learning. In ICCV, 2023. 1, 3, 5, 6, 7", + "bbox": [ + 109, + 90, + 468, + 119 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21423", + "bbox": [ + 478, + 952, + 517, + 963 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/f474f9b7-e45c-4a08-b76f-0b118ae61d7f_model.json b/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/f474f9b7-e45c-4a08-b76f-0b118ae61d7f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..50e2e334813d46e31dabc6e87fea8649ecce7932 --- /dev/null +++ b/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/f474f9b7-e45c-4a08-b76f-0b118ae61d7f_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.13, + 0.852, + 0.154 + ], + "angle": 0, + "content": "TAMM: TriAdapter Multi-Modal Learning for 3D Shape Understanding" + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.18, + 0.777, + 0.236 + ], + "angle": 0, + "content": "Zhihao Zhang\\(^{1*}\\) Shengcao Cao\\(^{2*}\\) Yu-Xiong Wang\\(^{2}\\) \n\\(^{1}\\)Xi'an Jiaotong University \\(^{2}\\)University of Illinois Urbana-Champaign \n\\(^{1}\\)zh1142@stu.xjtu.edu.cn \\(^{2}\\{cao44, yxw\\} @illinois.edu\\)" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.281, + 0.387, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.288, + 0.648, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.306, + 0.687, + 0.325 + ], + "angle": 0, + "content": "Point Clouds" + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.282, + 0.766, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.812, + 0.281, + 0.871, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.399, + 0.358, + 0.648, + 0.411 + ], + "angle": 0, + "content": "
Matching w/o CIAa gray box with four screws on ita bedside table with a lamp
Matching w/ CIAa desk with drawers on ita bed with white sheet and a lamp
" + }, + { + "type": "text", + "bbox": [ + 0.659, + 0.358, + 0.682, + 0.369 + ], + "angle": 0, + "content": "IAA" + }, + { + "type": "text", + "bbox": [ + 0.715, + 0.361, + 0.758, + 0.371 + ], + "angle": 0, + "content": "dress hat" + }, + { + "type": "text", + "bbox": [ + 0.822, + 0.361, + 0.854, + 0.371 + ], + "angle": 0, + "content": "mallet" + }, + { + "type": "text", + "bbox": [ + 0.66, + 0.38, + 0.684, + 0.39 + ], + "angle": 0, + "content": "TAA" + }, + { + "type": "text", + "bbox": [ + 0.711, + 0.38, + 0.762, + 0.39 + ], + "angle": 0, + "content": "lampshade" + }, + { + "type": "text", + "bbox": [ + 0.819, + 0.38, + 0.858, + 0.39 + ], + "angle": 0, + "content": "postbox" + }, + { + "type": "text", + "bbox": [ + 0.66, + 0.4, + 0.683, + 0.41 + ], + "angle": 0, + "content": "Dual" + }, + { + "type": "text", + "bbox": [ + 0.711, + 0.401, + 0.762, + 0.41 + ], + "angle": 0, + "content": "mushroom" + }, + { + "type": "text", + "bbox": [ + 0.819, + 0.401, + 0.858, + 0.41 + ], + "angle": 0, + "content": "newsletter" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.442, + 0.892, + 0.568 + ], + "angle": 0, + "content": "Figure 1. Our TriAdapter Multi-Modal Learning (TAMM) significantly enhances 3D shape understanding. Left: When aligning features of 3D point clouds (P) with 2D images (I) and/or text (T), prior methods (e.g., ULIP [45] and OpenShape [23]) under-utilize the image modality, due to the overlooked or unsolved image domain gap. TAMM better exploits the image modality and brings more gains when learning from both image and text data. The results are produced by pre-training Point-BERT [47] on ShapeNet [3]. Middle: Our CLIP Image Adapter (CIA) re-aligns the images rendered from 3D shapes with the text descriptions. The rendered images are inaccurately matched with text when the image features are directly extracted by CLIP, and CIA can correct the matching. Right: Our Image Alignment Adapter (IAA) and Text Alignment Adapter (TAA) decouple 3D features with complementary visual and semantic focuses. In the visualized examples, features from one single adapter are matched with classes whose appearance or semantics resemble the true class; using both adapters leads to the correct class." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.587, + 0.314, + 0.603 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.619, + 0.471, + 0.892 + ], + "angle": 0, + "content": "The limited scale of current 3D shape datasets hinders the advancements in 3D shape understanding, and motivates multi-modal learning approaches which transfer learned knowledge from data-abundant 2D image and language modalities to 3D shapes. However, even though the image and language representations have been aligned by cross-modal models like CLIP, we find that the image modality fails to contribute as much as the language in existing multi-modal 3D representation learning methods. This is attributed to the domain shift in the 2D images and the distinct focus of each modality. To more effectively leverage both modalities in the pre-training, we introduce TriAdapter Multi-Modal Learning (TAMM) - a novel two-stage learning approach based on three synergistic adapters. First, our CLIP Image Adapter mitigates the domain gap between 3D-rendered images and natural images, by adapting the visual representations of CLIP for synthetic image-text pairs. Subsequently, our Dual Adapters decou" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.589, + 0.892, + 0.771 + ], + "angle": 0, + "content": "ple the 3D shape representation space into two complementary sub-spaces: one focusing on visual attributes and the other for semantic understanding, which ensure a more comprehensive and effective multi-modal pre-training. Extensive experiments demonstrate that TAMM consistently enhances 3D representations for a wide range of 3D encoder architectures, pre-training datasets, and downstream tasks. Notably, we boost the zero-shot classification accuracy on Objaverse-LVIS from \\(46.8\\%\\) to \\(50.7\\%\\), and improve the 5-way 10-shot linear probing classification accuracy on ModelNet40 from \\(96.1\\%\\) to \\(99.0\\%\\). Project page: https://alanzhangcs.github.io/tamm-page." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.801, + 0.631, + 0.816 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.827, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Despite the recent success of 3D shape understanding [16, 19, 24, 42, 47, 51, 56], the limited scale of existing 3D shape datasets hinders the learning of more robust and generalizable 3D representations. Due to the significant human labor and expertise required in collecting and annotating 3D" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.903, + 0.206, + 0.916 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.952, + 0.519, + 0.964 + ], + "angle": 0, + "content": "21413" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "shape data, building large-scale 3D shape datasets is a prohibitive endeavor [3, 9, 26]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.124, + 0.471, + 0.382 + ], + "angle": 0, + "content": "To mitigate the challenge of the limited data scale, recent research in 3D shape representation learning shows a promising direction of multi-modal learning [23, 45]. By establishing connections among data of 3D shapes, 2D images, and text, it becomes possible to pre-train 3D shape representations by distilling learned knowledge from image and language modalities. There exists a comprehensive body of literature of massive datasets [36], high-quality models pre-trained via self-supervised learning for 2D image [2, 4, 15] and language [1, 10, 39] modalities, and furthermore, cross-modal vision-language models [5, 34]. For instance, ULIP [45] creates triplets of 3D point clouds, 2D images, and text by rendering images from 3D shapes and generating text descriptions from their metadata. Then, ULIP adopts contrastive learning to align the 3D shape features with both the image features and text features extracted by CLIP [34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.47, + 0.519 + ], + "angle": 0, + "content": "While the image and language modalities are pre-aligned by CLIP pre-training and share the same feature space, we observe that directly aligning 3D features with image features leads to considerably worse representation quality as compared with aligning 3D features with text features, as shown in Figure 1-left. This phenomenon suggests that existing methods (e.g., ULIP [45] and OpenShape [23]) are not optimally leveraging the 2D image modality in 3D representation pre-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.793 + ], + "angle": 0, + "content": "There are two reasons for this counter-intuitive observation: 1) The 2D images in the generated triplets follow a data distribution different from natural images, on which CLIP is pre-trained. The images are rendered or projected from 3D to 2D, usually lacking a realistic background and texture. Since the image domain is shifted, the image features are no longer well-aligned with the text features, as exemplified in Figure 1-middle. 2) Intuitively, the image features represent more visual attributes including the shape, texture, or color, while the text features have a focus on semantics such as the function of the object. As shown in Figure 1-right, if the 3D features are specifically aligned with one single modality of images or text, they pay attention to different aspects of the 3D shape. Therefore, enforcing the 3D shape to simultaneously align with two modalities that convey subtly distinct information could be challenging. Such an approach may not fully leverage the potential of multi-modal learning signals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Aiming at addressing these two issues hindering multimodal pre-training for 3D shape understanding, we propose TriAdapter Multi-Modal Learning (TAMM), a two-stage pre-training approach based on three synergistic adapters. In the first stage, to mitigate the domain gap between the 2D images rendered from 3D shapes and the natural images on which CLIP is pre-trained, we adapt the visual representa" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "tions of CLIP based on the synthetic image-text pairs. More specifically, we fine-tune a lightweight CLIP Image Adapter on top of CLIP visual encoder through contrastive learning and re-align the adapted image features with the text features in the new domain. This CLIP Image Adapter allows us to establish more accurate relations between 3D shapes, 2D images, and language in an updated feature space, and avoids learning 3D representations from mismatched image features and text features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.893, + 0.499 + ], + "angle": 0, + "content": "In the following second stage, to prevent the vision-semantic feature disparity from impairing our 3D representation pre-training, we choose to embrace this disparity and decouple the 3D representations into two sub-spaces. To comprehensively encode a 3D shape, the 3D encoder needs to capture both the visual and semantic aspects of its representation, which are centric to the corresponding image and text features, respectively. Therefore, we decouple the 3D feature space for the two focuses on visual and semantic representations. In particular, we attach two independent feature adapters to the 3D backbone and transform the 3D features into two sub-spaces. One sub-space focuses more on the visual representations, and is aligned with the 2D image feature space; the other sub-space focuses more on the semantic representations, and is aligned with the language feature space. This approach of decoupled feature spaces makes the learned 3D representations more comprehensive and expressive." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.5, + 0.831, + 0.514 + ], + "angle": 0, + "content": "To summarize, our main contribution includes:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.515, + 0.891, + 0.574 + ], + "angle": 0, + "content": "- We identify the under-utilization of the 2D image modality in existing multi-modal methods. The image domain gap and feature disparity in image-text pairs hinder representation learning in 3D shape understanding." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.575, + 0.89, + 0.635 + ], + "angle": 0, + "content": "- We propose a novel multi-modal learning framework with two learning stages and three unified adapter modules. Our proposed TAMM better exploits both image and language modalities and improves 3D shape representations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.636, + 0.891, + 0.74 + ], + "angle": 0, + "content": "- Our TAMM consistently enhances 3D representations for a variety of 3D encoder architectures (e.g., PointBERT [47], SparseConv [6]), pre-training datasets (e.g., ShapeNet [3], an ensembled dataset [23]), and downstream tasks (e.g., zero-shot and linear probing shape classification on Objaverse-LVIS [9], ModelNet40 [43], and ScanObjectNN [40])." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.515, + 0.891, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.642, + 0.77 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "3D Shape Understanding. There are two mainstreams for 3D shape representation learning: 1) Projecting 3D shapes into voxel or grid-based formats [37] and then using 2D/3D convolutions [6] for feature extraction. 2) Directly modeling 3D point clouds with point-centric architectures [24, 28, 30, 31, 33, 42, 50, 52]. In this work, to ensure a fair and comprehensive comparison with previous methods [23, 45] on the pre-training scheme, we follow" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.952, + 0.521, + 0.965 + ], + "angle": 0, + "content": "21414" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "their selection and utilize two representative 3D encoders from these two mainstreams: SparseConv [6] and Point-BERT [47]. SparseConv is designed for efficiently processing sparse voxels using specialized convolutions that focus computations on non-zero data points. Point-BERT [42] utilizes a Transformer-based architecture [41] and can be self-supervised by masked modeling [10]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.471, + 0.471 + ], + "angle": 0, + "content": "Multi-Modal Representation Learning. Contrastive Language-Image Pre-training (CLIP) [34] has enabled various downstream applications including object detection [48, 53, 55] and language grounding [22]. Recently, CLIP has been extended to 3D-based tasks, such as zero-shot text-to-3D generation [17, 20, 25, 27, 38, 44] and scene-level 3D segmentation [14, 29, 35, 46]. Meanwhile, developing general and robust representations for 3D shape understanding with the foundation of CLIP [19, 23, 45, 51, 54, 56] becomes a major focus. Among these methods, ULIP [45], as a pioneering work, utilizes contrastive learning to distill CLIP features into 3D representations. OpenShape [23] follows this learning paradigm with a focus on building a larger pre-training dataset with enriched and filtered text data. Unlike OpenShape [23], we focus on improving the multi-modal learning paradigm by more effectively leveraging both the image and text modalities via a two-stage pre-training approach based on three synergistic adapters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.483, + 0.168, + 0.499 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.644 + ], + "angle": 0, + "content": "Figure 2 shows our proposed two-stage approach, TriAdapter Multi-Modal Learning (TAMM), for pretraining robust and generalizable 3D representations by leveraging 2D images and text. We first revisit the triplet data generation and problem formulation in multi-modal 3D representation learning. Section 3.1 delves into our finetuning of the original CLIP model for better fitting 3D understanding tasks. Section 3.2 details our strategy to enhance alignment across the 3D, 2D, and text feature spaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Problem Formulation. Given \\( n \\) triplets \\( \\{(P_i, I_i, T_i)\\}_{i=1}^n \\), where \\( P_i \\) is a 3D point cloud, \\( I_i \\) represents the corresponding image produced by projecting the 3D point cloud \\( P_i \\) into 2D from an arbitrary perspective, and \\( T_i \\) denotes the associated text generated using advanced vision-language models such as BLIP [21], the objective is to learn high-quality 3D representations from the triplets. The basic framework to achieve this objective is proposed by ULIP [45] (and followed by OpenShape [23]) with the help of CLIP [5, 34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Formally, the 3D feature \\( f_{i}^{P} = E_{P}(P_{i}) \\) is produced by a learnable 3D encoder \\( E_{P} \\), and the corresponding image feature \\( f_{i}^{I} = E_{I}(I_{i}) \\) and text feature \\( f_{i}^{T} = E_{T}(T_{i}) \\) are generated by frozen CLIP encoders \\( E_{I}, E_{T} \\). Then the 3D encoder \\( E_{P} \\) is optimized by aligning the 3D feature space \\( f^{P} \\) simultaneously to the pre-aligned CLIP image space \\( f^{I} \\) and text space \\( f^{T} \\) through contrastive learning. The corresponding contrastive loss \\( L_{\\mathrm{contrast}}(f^{M_1}, f^{M_2}) \\) between two" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.875, + 0.107 + ], + "angle": 0, + "content": "modalities \\(M_{1}, M_{2}\\) (3D-2D or 3D-text) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.12, + 0.897, + 0.179 + ], + "angle": 0, + "content": "\\[\n- \\frac {1}{2 n} \\sum_ {i = 1} ^ {n} \\left(\\log \\frac {\\exp \\left(f _ {i} ^ {M _ {1}} \\cdot f _ {i} ^ {M _ {2}} / \\tau\\right)}{\\sum_ {j = 1} ^ {n} \\exp \\left(f _ {i} ^ {M _ {1}} \\cdot f _ {j} ^ {M _ {2}} / \\tau\\right)} + \\log \\frac {\\exp \\left(f _ {i} ^ {M _ {2}} \\cdot f _ {i} ^ {M _ {1}} / \\tau\\right)}{\\sum_ {j = 1} ^ {n} \\exp \\left(f _ {i} ^ {M _ {2}} \\cdot f _ {j} ^ {M _ {1}} / \\tau\\right)}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.181, + 0.828, + 0.196 + ], + "angle": 0, + "content": "where \\(\\tau\\) denotes the temperature hyperparameter." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.209, + 0.736, + 0.225 + ], + "angle": 0, + "content": "3.1. Image-Text Re-Alignment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.232, + 0.892, + 0.399 + ], + "angle": 0, + "content": "Unlike ULIP [45] or OpenShape [23], in which the 3D feature space is aligned with the image-text feature spaces directly produced by CLIP, we introduce an image feature space tuning strategy, aiming to foster better alignment across 3D, 2D, and language modalities (Figure 2-left). We argue that the image feature space produced directly by the CLIP visual encoder does not perfectly align with the text feature space and is thus sub-optimal. The reason is that the 2D images in the triplets, which originate from 3D point cloud projections, lack backgrounds and are visually different from natural images on which CLIP is pre-trained." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.893, + 0.672 + ], + "angle": 0, + "content": "Therefore, when using such 2D images from a shifted data domain to perform multi-modal pre-training, it becomes necessary to further fine-tune CLIP and re-align its image and language feature spaces. For the first time, we redesign domain adapters [18] for multi-modal contrastive 3D representation learning, and propose a CLIP Image Adapter (CIA), to adapt the image feature space for the rendered 2D images in our data triplets. We append a lightweight, learnable multi-layer perceptron (MLP) to CLIP image encoder, with a residual connection which seamlessly integrates the new knowledge acquired from fine-tuning with the existing knowledge from the pre-trained CLIP backbone. To avoid heavy computation and over-fitting, we only fine-tune the additional parameters in CIA, instead of the whole CLIP backbone. Formally, given the image feature \\( f_{i}^{I} \\) and text feature \\( f_{i}^{T} \\) extracted from the triplet \\( (P_{i}, T_{i}, I_{i}) \\), we use a learnable, two-layer CIA \\( A_{C}(\\cdot) \\) to adapt the image feature, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.686, + 0.892, + 0.705 + ], + "angle": 0, + "content": "\\[\nA _ {C} \\left(f _ {i} ^ {I}\\right) = \\sigma \\left(f _ {i} ^ {I} W _ {1}\\right) \\cdot W _ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.78 + ], + "angle": 0, + "content": "where \\(W_{1}\\) and \\(W_{2}\\) are the parameters associated with the linear transformation layers, and \\(\\sigma\\) is the non-linear activation function. The refined image feature \\(\\widetilde{f}_i^I\\) can be computed with a residual connection:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.793, + 0.892, + 0.812 + ], + "angle": 0, + "content": "\\[\n\\tilde {f} _ {i} ^ {I} = \\alpha A _ {C} \\left(f _ {i} ^ {I}\\right) + (1 - \\alpha) f _ {i} ^ {I}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.87 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a hyperparameter. Finally, CIA \\(A_{C}(\\cdot)\\) is optimized by minimizing the contrastive loss function (Equation 1), instantiated as:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.883, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e a l i g n}} = L _ {\\text {c o n t r a s t}} \\left(\\widetilde {f} ^ {I}, f ^ {T}\\right). \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.952, + 0.519, + 0.964 + ], + "angle": 0, + "content": "21415" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.081, + 0.31, + 0.096 + ], + "angle": 0, + "content": "Stage 1: Image-Text Re-Alignment" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.101, + 0.315, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.081, + 0.656, + 0.095 + ], + "angle": 0, + "content": "Stage 2: Decoupled Tri-Modal Pre-Training" + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.102, + 0.891, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.4, + 0.895, + 0.499 + ], + "angle": 0, + "content": "Figure 2. Overview of TAMM. Left: In Stage 1, TAMM fine-tunes a lightweight CLIP Image Adapter (CIA) through contrastive learning and re-aligns the image features with the text features to alleviate the domain shift originated from rendered images. Contrastive learning maximizes inner products between features from corresponding text-image pairs, and reduces similarities of mismatched pairs. Middle: In Stage 2, TAMM introduces Image Alignment Adapter (IAA) and Text Alignment Adapter (TAA) to decouple 3D representations into two sub-spaces: one focusing more on visual attributes and the other for semantic understanding, ensuring a more comprehensive and effective multi-modal pre-training strategy. Right: TAMM adaptively utilizes decoupled 3D features for various downstream tasks including linear probing classification (top) and zero-shot classification (bottom), achieving more robust classification results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.516, + 0.383, + 0.533 + ], + "angle": 0, + "content": "3.2. Decoupled Tri-Modal Pre-Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.698 + ], + "angle": 0, + "content": "Although the image and language feature spaces have been aligned by CLIP and our CLIP Image Adapter, they still encode subtly different information. For instance, the color of a 3D shape may be inferred from the image feature, but it cannot possibly be included in the text feature if the text description does not contain color information. Similarly, the image feature may not include semantic information such as the function or name of an object. In such cases, enforcing the 3D feature to align with both the image and text features simultaneously is challenging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In order to overcome the obstacle introduced by aligning the 3D shape feature space with two distinct modalities, we propose a novel decoupled tri-modal pre-training framework, which aligns two decoupled 3D shape feature spaces with the refined image feature space and the text feature space, respectively (Figure 2-middle). We encourage the 3D encoder to cover both the visual representation and the semantic information inherent to the 3D shape, avoiding the dilemma of aligning with disparate image and text features at the same time. Formally, given the triplets \\(\\{(P_i,I_i,T_i)\\}_{i = 1}^n\\), we first generate the image feature \\(\\widetilde{f}_i^I\\) using the frozen, adapted 2D image encoder (Equation 3) and the text feature \\(f_{i}^{T}\\) using the original CLIP" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.892, + 0.637 + ], + "angle": 0, + "content": "text encoder. We introduce a pair of lightweight Dual Adapters: Image Alignment Adapter (IAA) \\( A_V(\\cdot) \\) and Text Alignment Adapter (TAA) \\( A_S(\\cdot) \\). They split the 3D feature \\( f_i^P \\) originated from the 3D encoder \\( E_P \\) into a vision-focusing feature \\( f_i^{VP} \\) and a semantic-focusing feature \\( f_i^{SP} \\), respectively. Dual Adapters \\( A_V(\\cdot) \\) and \\( A_S(\\cdot) \\) share the same architecture as the CLIP Image Adapter in Section 3.1 and can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.641, + 0.892, + 0.668 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {V P} = A _ {V} \\left(f _ {i} ^ {P}\\right) = \\sigma \\left(f _ {i} ^ {P} W _ {1} ^ {V}\\right) \\cdot W _ {2} ^ {V}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.662, + 0.824, + 0.681 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {S P} = A _ {S} (f _ {i} ^ {P}) = \\sigma (f _ {i} ^ {P} W _ {1} ^ {S}) \\cdot W _ {2} ^ {S},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.779 + ], + "angle": 0, + "content": "where \\( W_1^V, W_2^V, W_1^S, \\) and \\( W_2^S \\) are the parameters associated with the linear layers, and \\( \\sigma \\) represents the activation function. By decoupling 3D features into these sub-spaces, the 3D encoder \\( E_P \\) interprets the 3D shape with a more comprehensive visual and semantic understanding, improving the expressivity of the learned representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Finally, instead of enforcing the 3D shape feature space to directly mimic the pre-aligned image-text feature space, we use the decoupled 3D features \\( f^{VP}, f^{SP} \\) and align them with the adapted image feature space \\( \\widetilde{f}^T \\) and the text feature space \\( f^T \\), respectively. Moreover, since one single image can only capture the 3D shape from one perspective, we align the 3D feature with the adapted features of multi-view images, to fully exploit the image modality and achieve a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.952, + 0.52, + 0.964 + ], + "angle": 0, + "content": "21416" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "better alignment between 3D and 2D. The overall loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.126, + 0.47, + 0.159 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r i m o d a l}} = L _ {\\text {c o n t r a s t}} \\left(f ^ {S P}, f ^ {T}\\right) + \\frac {1}{m} \\sum_ {k} ^ {m} L _ {\\text {c o n t r a s t}} \\left(f ^ {V P}, \\widetilde {f} ^ {I, k}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.166, + 0.469, + 0.198 + ], + "angle": 0, + "content": "where \\(m\\) represents the number of rendered images and \\(\\widetilde{f}^{I,k}\\) is the adapted image feature from the \\(k\\)-th view." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.198, + 0.471, + 0.44 + ], + "angle": 0, + "content": "Application in Downstream Tasks. The learned 3D feature sub-spaces, \\( f^{VP} \\) and \\( f^{SP} \\) can be adaptively applied to a variety of downstream tasks (Figure 2-right). Specifically, in zero-shot 3D classification, we leverage both the 3D vision-focusing feature \\( f^{VP} \\) and the 3D semantic-focusing feature \\( f^{SP} \\). We calculate the similarity between these features and category embeddings generated by the CLIP text encoder, respectively. After summing up the per-category similarity scores, the category with the highest similarity is chosen as the predicted class, yielding more robust and enhanced classification results compared with using a single sub-space alone. In the linear probing classification task, we concatenate the 3D vision-focusing feature \\( f^{VP} \\) and 3D semantic-focusing feature \\( f^{SP} \\) as input to the learnable linear classification layer, providing a more comprehensive and robust representation for 3D understanding." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.453, + 0.21, + 0.47 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.478, + 0.47, + 0.659 + ], + "angle": 0, + "content": "Pre-Training Datasets. Following the prior state-of-the-art method, OpenShape [23], our TAMM is pre-trained on the triplets generated from four datasets: ShapeNetCore [3], 3D-FUTURE [12], ABO [7], and Objaverse [9]. Our training sets are defined as follows: \"ShapeNet\" is a triplet set derived exclusively from the ShapeNetCore dataset, containing 52,470 3D shapes with corresponding images and text; \"Ensembled (no LVIS)\" is a set of 829,460 triplets from the above datasets excluding Objaverse-LVIS; \"Ensembled\" denotes the triplet set comprising data from all four datasets, containing 875,665 3D shapes and their associated images and text." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.66, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Evaluation Datasets. Our TAMM is evaluated on the following datasets: Objaverse-LVIS [9], ModelNet40 [43], ScanObjectNN [40], and ScanNet [8]. Objverse-LVIS encompasses a broad range of categories, featuring 46,832 high-quality shapes distributed across 1,156 LVIS [13] categories. ModelNet40 is a synthetic indoor 3D dataset comprising 40 categories. We use the test split of 2,468 shapes in our experiments. ScanObjectNN consists of scanned objects from 15 common categories. It has three main variants: OBJ-BG, OBJ-ONLY, and PB-T50-RS. ScanNet, characterized by its real-world scans, includes 1,513 indoor scenes containing 36,213 objects. We conduct experiments on four tasks including zero-shot 3D classification, linear probing 3D classification, few-shot linear probing 3D classification, and real-world recognition, to demonstrate the advantages of our TAMM. Other implementation details re" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "garding pre-training and evaluation are introduced in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.132, + 0.749, + 0.148 + ], + "angle": 0, + "content": "4.1. Zero-Shot 3D Classification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.156, + 0.894, + 0.307 + ], + "angle": 0, + "content": "Zero-shot ability is a key metric for reflecting the quality of learned 3D representations, which requires the representations to be directly applicable to datasets where the model has never been explicitly supervised. Without any further tuning, the 3D representations are compared with text embeddings of categories to predict the classes of 3D shapes. To make a fair comparison and keep consistency with prior work, we adopt the same settings as OpenShape [23] on three benchmarks: Objaverse-LVIS [9], ModelNet40 [43], and OBJ-ONLY (ScanObjectNN) [40]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.893, + 0.504 + ], + "angle": 0, + "content": "The results are summarized in Table 1. First, we observe that TAMM, benefited from multi-modal pre-training, outperforms PointCLIP [51] and PointCLIP v2 [56] by a large margin. Furthermore, it is evident that our pretrained models consistently outperform those pre-trained by ULIP and OpenShape, irrespective of whether they are pretrained on smaller datasets like ShapeNet or more expansive ones such as the Ensembled dataset. For instance, PointBERT [47] pre-trained by our TAMM on the Ensembled (no LVIS) dataset surpasses ULIP and OpenShape by margins of \\(+20.6\\%\\) and \\(+2.9\\%\\) in Top-1 accuracy on the long-tailed Objaverse-LVIS benchmark, which validates the effectiveness of our multi-modal pre-training scheme." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.515, + 0.788, + 0.53 + ], + "angle": 0, + "content": "4.2. Linear Probing 3D Classification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.84 + ], + "angle": 0, + "content": "To further evaluate the 3D representation quality of our pre-trained models, we design and conduct linear probing classification experiments. Here, we first freeze our pretrained Point-BERT model along with the Image Alignment Adapter (IAA) and Text Alignment Adapter (TAA), and then append a single learnable linear layer to the model. Given a batch of point clouds, as illustrated in Figure 2, we generate features from both IAA and TAA and concatenate them, and learn the appended linear classification layer on the concatenated features. We evaluate the accuracy of the linear classifier on three benchmarks: Objverse-LVIS [9], ModelNet40 [43], and ScanObjectNN [40]. Objverse-LVIS is a challenging long-tailed dataset with 1,156 categories, which has not been evaluated by prior methods [23, 45]. To ensure evaluation with statistically meaningful number of samples per class, we exclude categories with fewer than 10 instances, leaving us 1,046 categories. Samples in each class are divided into training and testing sets at an \\(8:2\\) ratio. For ModelNet40 and ScanObjectNN, we use the standard splits, following [32, 42]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The results are shown in Table 2. TAMM consistently outperforms all previous methods by a large margin. For example, pre-trained on the Ensembled dataset, TAMM improves over OpenShape by \\(+11.2\\%\\) and \\(+2.2\\%\\) overall ac" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.952, + 0.521, + 0.965 + ], + "angle": 0, + "content": "21417" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.075, + 0.874, + 0.398 + ], + "angle": 0, + "content": "
Pre-Training DatasetModelPre-Training MethodObjaverse-LVIS [9]ModelNet40 [43]ScanObjectNN [40]
Top-1Top-3Top-5Top-1Top-3Top-5Top-1Top-3Top-5
-PointCLIP [51]-1.94.15.819.328.634.810.520.830.6
-PointCLIP v2 [56]-4.79.512.963.677.985.042.263.374.5
ShapeNetViT-B/32 [11]CLIP2Point [19]2.75.87.949.571.381.225.544.659.4
Transformer [41]ReCon [32]1.12.73.761.273.978.142.362.575.6
SparseConv [6]OpenShape [23]11.621.827.172.987.293.052.772.783.6
TAMM (Ours)13.624.229.374.688.294.057.975.383.1
Point-BERT [47]ULIP [45]6.213.617.960.479.084.451.571.180.2
OpenShape [23]10.820.225.070.386.991.351.369.478.4
TAMM (Ours)13.724.229.273.188.591.954.874.583.3
Ensembled (no LVIS)SparseConv [6]OpenShape [23]37.058.466.982.695.097.554.976.887.0
TAMM (Ours)39.862.070.485.796.898.357.581.390.0
Point-BERT [47]ULIP [45]21.438.146.071.484.489.246.066.176.4
OpenShape [23]39.160.868.985.396.297.447.272.484.7
TAMM (Ours)42.063.671.786.396.698.156.778.386.1
EnsembledSparseConv [6]OpenShape [23]43.464.872.483.495.697.856.778.988.6
TAMM (Ours)43.866.274.185.496.498.158.581.389.5
Point-BERT [47]ULIP [45]26.844.852.675.188.193.251.672.582.3
OpenShape [23]46.869.177.084.496.598.052.279.788.7
TAMM (Ours)50.773.280.685.096.698.155.780.788.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.407, + 0.895, + 0.451 + ], + "angle": 0, + "content": "Table 1. Zero-shot 3D classification results. TAMM sets new state of the art in zero-shot classification accuracy across Objaverse-LVIS, ModelNet-40, and ScanObjectNN benchmarks, outperforming existing methods in diverse settings of pre-training datasets and 3D model architectures. The performance gain is more significant on the most challenging long-tailed Objverse-LVIS dataset." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.466, + 0.472, + 0.593 + ], + "angle": 0, + "content": "
Pre-Training DatasetMethodO-LVIS [9]M-40 [43]ScanObjectNN [40]
OBJ-BGOBJ-ONLYPB-T50-RS
ShapeNetULIP [45]34.690.675.475.464.8
OpenShape [23]29.388.577.878.564.1
TAMM (Ours)39.191.080.681.168.5
Rel. Improv.+9.8+2.5+2.8+2.6+4.4
EnsembledOpenShape [23]48.391.385.985.478.0
TAMM (Ours)59.593.588.588.080.3
Rel. Improv.+11.2+2.2+2.6+2.6+1.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.602, + 0.47, + 0.645 + ], + "angle": 0, + "content": "Table 2. Linear probing 3D classification results. TAMM outperforms previous methods by a large margin, e.g., +11.2% accuracy gain on the challenging Objaverse-LVIS dataset." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.666, + 0.47, + 0.819 + ], + "angle": 0, + "content": "curacy on Objaverse-LVIS and ModelNet40, respectively. This demonstrates that TAMM effectively exploits knowledge from CLIP and learns generalizable 3D representations that are directly applicable to novel linear classification tasks. Notably, TAMM even surpasses the Point-BERT model pre-trained without multi-modality but allowed to be fully fine-tuned on OBJ-BG (87.4%) [47] by +1.1% accuracy. These strong results indicate that TAMM has learned excellent transferable 3D representations, showing a great potential in real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.831, + 0.445, + 0.848 + ], + "angle": 0, + "content": "4.3. Few-Shot Linear Probing 3D Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Following the previous evaluation, we perform few-shot classification experiments on ModelNet40 [43] to assess TAMM in low-data scenarios. Similar to the linear probing" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.466, + 0.892, + 0.566 + ], + "angle": 0, + "content": "
Pre-Training DatasetMethod5-way10-way
10-shot20-shot10-shot20-shot
ShapeNetULIP [45]94.4±3.793.2±4.286.6±5.390.6±5.2
OpenShape [23]95.3±2.697.9±3.989.2±5.192.9±3.9
TAMM (Ours)97.8±1.998.1±1.495.3±3.995.9±2.6
EnsembledOpenShape [23]96.1±2.795.7±2.589.1±4.691.8±3.7
TAMM (Ours)99.0±1.399.4±0.796.8±2.997.4±2.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.575, + 0.894, + 0.645 + ], + "angle": 0, + "content": "Table 3. Few-shot linear probing classification results on ModelNet40. We report the average accuracy and standard deviation of 10 independent experiments. Our TAMM consistently achieves both the best average accuracy and the lowest variance in various few-shot settings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.867 + ], + "angle": 0, + "content": "experiments described in Section 4.2, we extend our model with an additional linear classification layer and only train this linear layer instead of fine-tuning the entire model. Following previous work [47, 50], we adopt the \" \\( K \\)-way \\( N \\)-shot\" configuration, wherein we select \\( K \\) classes at random and sample \\( (N + 20) \\) instances per class. Training is conducted on a support set of \\( K \\times N \\) samples, while evaluation is based on a query set comprised of the remaining 20 instances per class. We assess our model under four distinct scenarios: \"5-way 10-shot,\" \"5-way 20-shot,\" \"10-way 10-shot,\" and \"10-way 20-shot.\" For each scenario, we carry out 10 separate trials, and report both the mean performance and the standard deviation across these trials." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Table 3, our TAMM achieves the best results and sets a new state of the art in the few-shot clas-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.952, + 0.521, + 0.965 + ], + "angle": 0, + "content": "21418" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.075, + 0.891, + 0.187 + ], + "angle": 0, + "content": "
MethodAvg.BedCabChairSofaTablDoorWindBksfPicCntrDeskCurtFridgBathShowrToilSink
CLIP2Point [19]24.920.80.085.143.326.569.90.020.91.731.727.00.01.646.50.022.425.6
PointCLIP w/ TP. [56]26.10.055.772.85.05.11.70.077.20.00.051.70.30.00.040.385.349.2
CLIP2Point w/ TP. [19]35.211.83.045.127.610.561.52.671.90.333.629.94.711.572.292.486.134.0
\\( CLIP^2 \\) [49]38.532.667.269.342.318.319.14.062.61.412.752.840.19.159.741.071.045.5
\\( OpenShape^\\dagger \\) [23]45.666.73.275.883.537.749.247.564.948.21.966.170.21.850.057.145.27.1
TAMM (Ours)†49.466.74.883.684.548.957.948.280.561.31.960.683.67.041.456.148.43.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.188, + 0.891, + 0.201 + ], + "angle": 0, + "content": "w/ TP. denotes training with the real-world data provided by \\(\\mathrm{CLIP^2}\\). Results using Point-BERT [47] as 3D encoder, pre-trained on the Ensembled dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.208, + 0.891, + 0.235 + ], + "angle": 0, + "content": "Table 4. Zero-shot classification results on the real-world ScanNet dataset. Avg.: the mean average Top-1 accuracy of all classes. TAMM achieves the best results." + }, + { + "type": "table", + "bbox": [ + 0.128, + 0.243, + 0.457, + 0.329 + ], + "angle": 0, + "content": "
CIAIAATAAImg-Txt AccObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
✓✓40.113.523.929.172.888.291.7
✓60.912.522.427.371.486.089.8
✓✓60.912.922.927.773.888.292.5
✓✓60.913.023.028.272.987.890.4
✓✓✓60.913.724.229.273.188.591.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.128, + 0.331, + 0.456, + 0.367 + ], + "angle": 0, + "content": "(a) Adapters in pre-training. Pre-training with CLIP Image Adapter (CIA), Image Alignment Adapter (IAA), and Text Alignment Adapter (TAA) is the most effective." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.259, + 0.842, + 0.311 + ], + "angle": 0, + "content": "
StageImg-Txt AccObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
One stage55.312.221.427.071.486.490.7
Two stages60.913.724.229.273.188.591.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.529, + 0.314, + 0.842, + 0.351 + ], + "angle": 0, + "content": "(b) Pre-training stages. Learning the CLIP Image Adapter first and then pre-train the 3D encoder with Dual Adapters is better than training modules all together." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.375, + 0.342, + 0.439 + ], + "angle": 0, + "content": "
ImageTextObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
✓11.920.825.967.687.392.5
✓10.519.223.770.285.088.0
✓✓13.724.229.273.188.591.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.078, + 0.442, + 0.341, + 0.478 + ], + "angle": 0, + "content": "(c) Pre-training modalities. Exploiting both image and text data provides the most performance gain. They contribute almost equally in TAMM." + }, + { + "type": "table", + "bbox": [ + 0.361, + 0.375, + 0.626, + 0.44 + ], + "angle": 0, + "content": "
IAATAAObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
✓13.023.128.368.186.591.4
✓12.922.527.572.486.290.4
✓✓13.724.229.273.188.591.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.361, + 0.442, + 0.625, + 0.478 + ], + "angle": 0, + "content": "(d) Adapters for inference. Dual Adapters are complementary, and their combination more accurately classifies 3D shapes at test time." + }, + { + "type": "table", + "bbox": [ + 0.645, + 0.375, + 0.892, + 0.451 + ], + "angle": 0, + "content": "
ImagesObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
113.623.628.972.588.092.6
213.523.528.873.187.692.8
413.724.229.273.188.591.9
813.523.128.473.089.092.8
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.645, + 0.454, + 0.891, + 0.478 + ], + "angle": 0, + "content": "(e) Number of images. Aligning a 3D shape with 4 images of it is the best." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.49, + 0.893, + 0.532 + ], + "angle": 0, + "content": "Table 5. Ablation study of components in TAMM. We pre-train Point-BERT models on ShapeNet in various settings and evaluate their zero-shot classification performance on Objaverse-LVIS and ModelNet40. The baseline method OpenShape [23] achieves \\(10.8\\%\\) and \\(70.3\\%\\) Top-1 accuracy on these two benchmarks (Table 1), respectively. The setting adopted by TAMM is marked." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.688 + ], + "angle": 0, + "content": "sification task. TAMM shows both higher overall accuracy and smaller deviations than other methods, showcasing the generalizability and robustness of TAMM-learned 3D representations. For example, when pre-trained on the Ensembled dataset, our TAMM surpasses OpenShape [23] by \\(2.9\\%\\), \\(3.7\\%\\), \\(7.7\\%\\), \\(5.6\\%\\), respectively for all four settings. These results indicate that our TAMM is able to learn 3D representations that are more generalizable and can be readily adapted to downstream tasks under low-data regimes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.3, + 0.713 + ], + "angle": 0, + "content": "4.4. Real-World Recognition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To evaluate the capability of TAMM in understanding \\(3D\\) shapes from the real world, we follow the previous work \\(\\mathrm{CLIP}^2\\) [49] and test TAMM on a real-world recognition task, in which the model aims at correctly classifying each instance from a complex scene in a zero-shot manner. Specifically, we select the real-world scene-level dataset ScanNet [8] and adopt same data splits as [49], containing 17 classes. We perform zero-shot classification (Section 4.1) on the point cloud of each object instance extracted from scenes, and report both the Top-1 accuracy of each class and the mean accuracy of all 17 classes. As shown in Table 4, TAMM significantly outperforms all prior meth" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.657 + ], + "angle": 0, + "content": "ods, showing improvements of \\(3.8\\%\\) and \\(10.9\\%\\) in average Top-1 accuracy compared with OpenShape and \\(\\mathrm{CLIP^2}\\), respectively. These results underscore TAMM's ability in recognizing and understanding 3D shapes captured from real-world scenarios. Further results on complex scene recognition and instance segmentation are included in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.655, + 0.683 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.795 + ], + "angle": 0, + "content": "In this section, we provide additional experimental results to further test the performance gain from each design of TAMM. Due to limited computation, we pre-train PointBERT [47] on the small-scale ShapeNet [3], and use two more challenging zero-shot classification benchmarks ModelNet40 [43] and Objaverse-LVIS [9] for evaluation. We alter only one component in each set of experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Adapters in Pre-Training. As described in Section 3.1, our CLIP Image Adapter (CIA) mitigates the domain gap between rendered images and natural images. We first measure the contrastive accuracy, computed as the ratio of image-text pairs where the text feature is more similar to the image from the same triplet than any other text. The contrastive accuracy of CLIP without CIA is only \\(40.1\\%\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.952, + 0.52, + 0.964 + ], + "angle": 0, + "content": "21419" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "indicating that the original image-text feature space is compromised by the domain gap. CIA brings the accuracy up to \\(60.9\\%\\), showing the effectiveness of our CIA in resolving the shifted data domain. Furthermore, as shown in Table 5a, CIA increases the zero-shot accuracy on Objaverse-LVIS dataset from \\(12.2\\%\\) to \\(13.7\\%\\), validating its effectiveness on improving the pre-trained 3D representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.471, + 0.336 + ], + "angle": 0, + "content": "TAMM adopts Image Alignment Adapter (IAA) and Text Alignment Adapter (TAA) to decouple the feature space and enrich the learned 3D representations. To further explore the performance gains introduced by our Dual Adapters (IAA and TAA), we experiment tri-modal pretraining with or without IAA and TAA. As shown in Table 5a, both adapters bring performance gains, and their combination achieves the best result, demonstrating the effectiveness of Dual Adapters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.471, + 0.459 + ], + "angle": 0, + "content": "Pre-Training Stages. We investigate the significance of our two-stage pre-training design, which initially learns CLIP Image Adapter (CIA) to re-align image-text pairs and subsequently employs Dual Adapters in decoupled tri-modal pre-training. An alternative could be learning these modules all together in one stage. As shown in Table 5b, remarkably, the two-stage pre-training approach achieves a better performance compared with one-stage pre-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.47, + 0.627 + ], + "angle": 0, + "content": "Pre-Training Modalities. We further explore the effectiveness of image and language modalities in pre-training. Table 5c shows that integrating the 3D modality with both the image and language modalities consistently yields superior performance across various benchmarks, compared with aligning it with either modality alone. Moreover, unlike prior methods which ineffectively learn from images (Figure 1-left), the image and language modalities contribute almost equally in TAMM, which also signifies that TAMM better exploits the image modality and brings more gains by learning from both image and text data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Adapters for Inference. Our IAA and TAA decouple 3D features with visual and semantic focuses, respectively, offering a more accurate and comprehensive understanding for 3D shapes. At inference time, we combine the outputs from Dual Adapters for classification tasks. To investigate this design, we separately utilize features produced solely by either IAA or TAA in zero-shot classification evaluation. As shown in Table 5d, the combined features achieve more accurate classification, demonstrating that the learned features from IAA and TAA are complementary to each other." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Number of Images. In order to gain more comprehensive knowledge from the image modality, TAMM simultaneously aligns the 3D feature with multi-view 2D images, each projected from a different perspective. We evaluate this design by aligning the 3D feature with a varying number of corresponding 2D images, ranging from 1 to 8 images. As shown in Table 5e, multi-view image features offer a performance gain by fully exploiting the image modality." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.08, + 0.895, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.321, + 0.895, + 0.435 + ], + "angle": 0, + "content": "Figure 3. Qualitative results. Top: CIA re-aligns the images rendered from 3D shapes with the text descriptions. The rendered images are inaccurately matched with text when the image features are directly extracted by CLIP, and CIA can correct the matching. Bottom: IAA and TAA decouple 3D features with complementary visual and semantic focuses. Features from one single adapter are matched with classes whose appearance or semantics resemble the true class; using both adapters leads to the correct class." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.888, + 0.468 + ], + "angle": 0, + "content": "Aligning a 3D shape with 4 views achieves the best result." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.892, + 0.621 + ], + "angle": 0, + "content": "Qualitative Results. Finally, we provide some visualizations to intuitively demonstrate the benefit of our proposed TAMM. As illustrated in Figure 3, CIA successfully mitigates the domain gap introduced by rendered images, and leads to the more accurate matching between images and text. IAA and TAA learn 3D representations with subtly different but complementary focuses on vision and semantics, respectively, and their combination brings more robust and comprehensive 3D representations. More visualized results are included in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.634, + 0.62, + 0.649 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.765 + ], + "angle": 0, + "content": "In this work, we examine the sub-optimal utilization of 2D images in existing multi-modal pre-training methods for 3D shape understanding, and propose TriAdapter MultiModal Learning (TAMM), a novel two-stage representation learning approach built on three synergistic adapter modules. Extensive experiments verify that TAMM consistently learns improved 3D features in various settings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.841 + ], + "angle": 0, + "content": "Limitations. Due to the limited computation resources, we are not able to perform pre-training on very large-scale 3D backbones with billions of parameters. The quality of the learned 3D representation could be further improved if the 3D backbones are scaled to a larger size." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Jump ARCHES endowment, and the IBM-Illinois Discovery Accelerator Institute. This work used NVIDIA GPUs at NCSA Delta through allocations CIS220014, CIS230012, and CIS230013 from the ACCESS program." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.952, + 0.521, + 0.965 + ], + "angle": 0, + "content": "21420" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.253 + ], + "angle": 0, + "content": "[1] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.255, + 0.47, + 0.31 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.471, + 0.394 + ], + "angle": 0, + "content": "[3] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.395, + 0.471, + 0.437 + ], + "angle": 0, + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.506 + ], + "angle": 0, + "content": "[5] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D spatio-temporal ConvNets: Minkowski convolutional neural networks. In CVPR, 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.47, + 0.633 + ], + "angle": 0, + "content": "[7] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F. Yago Vicente, Thomas Dideriksien, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. ABO: Dataset and benchmarks for real-world 3D object understanding. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.47, + 0.69 + ], + "angle": 0, + "content": "[8] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.691, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.761, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.818, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[12] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3D-FUTURE: 3D furniture shape with texture. IJCV, 129:3313-3337, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[14] Huy Ha and Shuran Song. Semantic abstraction: Openworld 3D scene understanding from 2D vision-language models. In CoRL, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[15] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[16] Deepti Hegde, Jeya Maria Jose Valanarasu, and Vishal Patel. CLIP goes 3D: Leveraging prompt tuning for language grounded 3D recognition. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[17] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. AvatarCLIP: Zero-shot text-driven generation and animation of 3D avatars. ACM Transactions on Graphics, 41(4):1-19, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.359, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[18] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. In ICML, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[19] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. CLIP2Point: Transfer CLIP to point cloud classification with image-depth pre-training. In ICCV, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[20] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[22] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.892, + 0.706 + ], + "angle": 0, + "content": "[23] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. OpenShape: Scaling up 3D shape representation towards open-world understanding. In NeurIPS, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[24] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In ICLR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[25] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2Mesh: Text-driven neural stylization for meshes. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[26] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. CLIP-Mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.953, + 0.518, + 0.964 + ], + "angle": 0, + "content": "21421" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[28] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[29] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. OpenScene: 3D scene understanding with open vocabularies. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. PointNet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[32] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3D representation learning guided by generative pretraining. In ICML, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[33] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. PointNeXt: Revisiting PointNet++ with improved training and scaling strategies. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[35] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3D semantic segmentation in the wild. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: An open large-scale dataset for training next generation image-text models. In NeurIPS Datasets and Benchmarks Track, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[37] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. PV-RCNN: Point-voxel feature set abstraction for 3D object detection. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[38] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. MotionCLIP: Exposing human motion generation to CLIP space. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[39] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[40] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[42] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point Transformer V2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[43] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In CVPR, 2015. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[44] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3D: Zero-shot text-to-3D synthesis using 3D shape prior and text-to-image diffusion models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[45] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. ULIP: Learning a unified representation of language, images, and point clouds for 3D understanding. In CVPR, 2023. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.332, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[46] Jihan Yang, Runyu Ding, Zhe Wang, and Xiaojuan Qi. RegionPLC: Regional point-language contrastive learning for open-world 3D scene understanding. arXiv preprint arXiv:2304.00962, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[47] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-BERT: Pre-training 3D point cloud transformers with masked point modeling. In CVPR, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[48] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[49] Yihan Zeng, Chenhan Jiang, Jiageng Mao, Jianhua Han, Chaoqiang Ye, Qingqiu Huang, Dit-Yan Yeung, Zhen Yang, Xiaodan Liang, and Hang Xu. CLIP\\(^2\\): Contrastive language-image-point pretraining from real-world point cloud data. In CVPR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.56, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[50] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. PointM2AE: Multi-scale masked autoencoders for hierarchical point cloud pre-training. In NeurIPS, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[51] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. PointCLIP: Point cloud understanding by CLIP. In CVPR, 2022. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.673, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[52] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3D features on any point-cloud. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[53] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. RegionCLIP: Region-based language-image pretraining. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.786, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[54] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3D: Exploring unified 3D representation at scale. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.83, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[55] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[56] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Point-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.953, + 0.52, + 0.964 + ], + "angle": 0, + "content": "21422" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "CLIP V2: Prompting CLIP and GPT for powerful 3D open-world learning. In ICCV, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.953, + 0.519, + 0.964 + ], + "angle": 0, + "content": "21423" + } + ] +] \ No newline at end of file diff --git a/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/f474f9b7-e45c-4a08-b76f-0b118ae61d7f_origin.pdf b/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/f474f9b7-e45c-4a08-b76f-0b118ae61d7f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2c2e8a1e0161968167227d0439f72e3100717f30 --- /dev/null +++ b/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/f474f9b7-e45c-4a08-b76f-0b118ae61d7f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddc6d9c1c135c2a0ba3fadae564ebfd6f8fb20e1b37887a4fdc8327989d10571 +size 1266685 diff --git a/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/full.md b/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f9fb1358f9848260a43a5f2495fad932d9cbc354 --- /dev/null +++ b/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/full.md @@ -0,0 +1,322 @@ +# TAMM: TriAdapter Multi-Modal Learning for 3D Shape Understanding + +Zhihao Zhang $^{1*}$ Shengcao Cao $^{2*}$ Yu-Xiong Wang $^{2}$ $^{1}$ Xi'an Jiaotong University $^{2}$ University of Illinois Urbana-Champaign + $^{1}$ zh1142@stu.xjtu.edu.cn $^{2}\{cao44, yxw\} @illinois.edu$ + +![](images/06b8d2bbf939a68fcb692a39846a4b7d07d46f557686a0edae21a12585c51933.jpg) +Figure 1. Our TriAdapter Multi-Modal Learning (TAMM) significantly enhances 3D shape understanding. Left: When aligning features of 3D point clouds (P) with 2D images (I) and/or text (T), prior methods (e.g., ULIP [45] and OpenShape [23]) under-utilize the image modality, due to the overlooked or unsolved image domain gap. TAMM better exploits the image modality and brings more gains when learning from both image and text data. The results are produced by pre-training Point-BERT [47] on ShapeNet [3]. Middle: Our CLIP Image Adapter (CIA) re-aligns the images rendered from 3D shapes with the text descriptions. The rendered images are inaccurately matched with text when the image features are directly extracted by CLIP, and CIA can correct the matching. Right: Our Image Alignment Adapter (IAA) and Text Alignment Adapter (TAA) decouple 3D features with complementary visual and semantic focuses. In the visualized examples, features from one single adapter are matched with classes whose appearance or semantics resemble the true class; using both adapters leads to the correct class. + +![](images/0db415b090078d7b48e45bd99d3a914a40d606ed9abf803a93a3ce4f554a738a.jpg) +Point Clouds + +![](images/8b8971a677425e6d6c3627af7c8a303d46746dc4691e9e240a62c5c57323c367.jpg) + +![](images/0c1fa06a9826da8db0e3dae11c6bc72bede6338ee24b14bd2a8e49dff4b56f03.jpg) + +
Matching w/o CIAa gray box with four screws on ita bedside table with a lamp
Matching w/ CIAa desk with drawers on ita bed with white sheet and a lamp
+ +IAA + +dress hat + +mallet + +TAA + +lampshade + +postbox + +Dual + +mushroom + +newsletter + +# Abstract + +The limited scale of current 3D shape datasets hinders the advancements in 3D shape understanding, and motivates multi-modal learning approaches which transfer learned knowledge from data-abundant 2D image and language modalities to 3D shapes. However, even though the image and language representations have been aligned by cross-modal models like CLIP, we find that the image modality fails to contribute as much as the language in existing multi-modal 3D representation learning methods. This is attributed to the domain shift in the 2D images and the distinct focus of each modality. To more effectively leverage both modalities in the pre-training, we introduce TriAdapter Multi-Modal Learning (TAMM) - a novel two-stage learning approach based on three synergistic adapters. First, our CLIP Image Adapter mitigates the domain gap between 3D-rendered images and natural images, by adapting the visual representations of CLIP for synthetic image-text pairs. Subsequently, our Dual Adapters decou + +ple the 3D shape representation space into two complementary sub-spaces: one focusing on visual attributes and the other for semantic understanding, which ensure a more comprehensive and effective multi-modal pre-training. Extensive experiments demonstrate that TAMM consistently enhances 3D representations for a wide range of 3D encoder architectures, pre-training datasets, and downstream tasks. Notably, we boost the zero-shot classification accuracy on Objaverse-LVIS from $46.8\%$ to $50.7\%$ , and improve the 5-way 10-shot linear probing classification accuracy on ModelNet40 from $96.1\%$ to $99.0\%$ . Project page: https://alanzhangcs.github.io/tamm-page. + +# 1. Introduction + +Despite the recent success of 3D shape understanding [16, 19, 24, 42, 47, 51, 56], the limited scale of existing 3D shape datasets hinders the learning of more robust and generalizable 3D representations. Due to the significant human labor and expertise required in collecting and annotating 3D + +shape data, building large-scale 3D shape datasets is a prohibitive endeavor [3, 9, 26]. + +To mitigate the challenge of the limited data scale, recent research in 3D shape representation learning shows a promising direction of multi-modal learning [23, 45]. By establishing connections among data of 3D shapes, 2D images, and text, it becomes possible to pre-train 3D shape representations by distilling learned knowledge from image and language modalities. There exists a comprehensive body of literature of massive datasets [36], high-quality models pre-trained via self-supervised learning for 2D image [2, 4, 15] and language [1, 10, 39] modalities, and furthermore, cross-modal vision-language models [5, 34]. For instance, ULIP [45] creates triplets of 3D point clouds, 2D images, and text by rendering images from 3D shapes and generating text descriptions from their metadata. Then, ULIP adopts contrastive learning to align the 3D shape features with both the image features and text features extracted by CLIP [34]. + +While the image and language modalities are pre-aligned by CLIP pre-training and share the same feature space, we observe that directly aligning 3D features with image features leads to considerably worse representation quality as compared with aligning 3D features with text features, as shown in Figure 1-left. This phenomenon suggests that existing methods (e.g., ULIP [45] and OpenShape [23]) are not optimally leveraging the 2D image modality in 3D representation pre-training. + +There are two reasons for this counter-intuitive observation: 1) The 2D images in the generated triplets follow a data distribution different from natural images, on which CLIP is pre-trained. The images are rendered or projected from 3D to 2D, usually lacking a realistic background and texture. Since the image domain is shifted, the image features are no longer well-aligned with the text features, as exemplified in Figure 1-middle. 2) Intuitively, the image features represent more visual attributes including the shape, texture, or color, while the text features have a focus on semantics such as the function of the object. As shown in Figure 1-right, if the 3D features are specifically aligned with one single modality of images or text, they pay attention to different aspects of the 3D shape. Therefore, enforcing the 3D shape to simultaneously align with two modalities that convey subtly distinct information could be challenging. Such an approach may not fully leverage the potential of multi-modal learning signals. + +Aiming at addressing these two issues hindering multimodal pre-training for 3D shape understanding, we propose TriAdapter Multi-Modal Learning (TAMM), a two-stage pre-training approach based on three synergistic adapters. In the first stage, to mitigate the domain gap between the 2D images rendered from 3D shapes and the natural images on which CLIP is pre-trained, we adapt the visual representa + +tions of CLIP based on the synthetic image-text pairs. More specifically, we fine-tune a lightweight CLIP Image Adapter on top of CLIP visual encoder through contrastive learning and re-align the adapted image features with the text features in the new domain. This CLIP Image Adapter allows us to establish more accurate relations between 3D shapes, 2D images, and language in an updated feature space, and avoids learning 3D representations from mismatched image features and text features. + +In the following second stage, to prevent the vision-semantic feature disparity from impairing our 3D representation pre-training, we choose to embrace this disparity and decouple the 3D representations into two sub-spaces. To comprehensively encode a 3D shape, the 3D encoder needs to capture both the visual and semantic aspects of its representation, which are centric to the corresponding image and text features, respectively. Therefore, we decouple the 3D feature space for the two focuses on visual and semantic representations. In particular, we attach two independent feature adapters to the 3D backbone and transform the 3D features into two sub-spaces. One sub-space focuses more on the visual representations, and is aligned with the 2D image feature space; the other sub-space focuses more on the semantic representations, and is aligned with the language feature space. This approach of decoupled feature spaces makes the learned 3D representations more comprehensive and expressive. + +To summarize, our main contribution includes: + +- We identify the under-utilization of the 2D image modality in existing multi-modal methods. The image domain gap and feature disparity in image-text pairs hinder representation learning in 3D shape understanding. +- We propose a novel multi-modal learning framework with two learning stages and three unified adapter modules. Our proposed TAMM better exploits both image and language modalities and improves 3D shape representations. +- Our TAMM consistently enhances 3D representations for a variety of 3D encoder architectures (e.g., PointBERT [47], SparseConv [6]), pre-training datasets (e.g., ShapeNet [3], an ensembled dataset [23]), and downstream tasks (e.g., zero-shot and linear probing shape classification on Objaverse-LVIS [9], ModelNet40 [43], and ScanObjectNN [40]). + +# 2. Related Work + +3D Shape Understanding. There are two mainstreams for 3D shape representation learning: 1) Projecting 3D shapes into voxel or grid-based formats [37] and then using 2D/3D convolutions [6] for feature extraction. 2) Directly modeling 3D point clouds with point-centric architectures [24, 28, 30, 31, 33, 42, 50, 52]. In this work, to ensure a fair and comprehensive comparison with previous methods [23, 45] on the pre-training scheme, we follow + +their selection and utilize two representative 3D encoders from these two mainstreams: SparseConv [6] and Point-BERT [47]. SparseConv is designed for efficiently processing sparse voxels using specialized convolutions that focus computations on non-zero data points. Point-BERT [42] utilizes a Transformer-based architecture [41] and can be self-supervised by masked modeling [10]. + +Multi-Modal Representation Learning. Contrastive Language-Image Pre-training (CLIP) [34] has enabled various downstream applications including object detection [48, 53, 55] and language grounding [22]. Recently, CLIP has been extended to 3D-based tasks, such as zero-shot text-to-3D generation [17, 20, 25, 27, 38, 44] and scene-level 3D segmentation [14, 29, 35, 46]. Meanwhile, developing general and robust representations for 3D shape understanding with the foundation of CLIP [19, 23, 45, 51, 54, 56] becomes a major focus. Among these methods, ULIP [45], as a pioneering work, utilizes contrastive learning to distill CLIP features into 3D representations. OpenShape [23] follows this learning paradigm with a focus on building a larger pre-training dataset with enriched and filtered text data. Unlike OpenShape [23], we focus on improving the multi-modal learning paradigm by more effectively leveraging both the image and text modalities via a two-stage pre-training approach based on three synergistic adapters. + +# 3. Method + +Figure 2 shows our proposed two-stage approach, TriAdapter Multi-Modal Learning (TAMM), for pretraining robust and generalizable 3D representations by leveraging 2D images and text. We first revisit the triplet data generation and problem formulation in multi-modal 3D representation learning. Section 3.1 delves into our finetuning of the original CLIP model for better fitting 3D understanding tasks. Section 3.2 details our strategy to enhance alignment across the 3D, 2D, and text feature spaces. + +Problem Formulation. Given $n$ triplets $\{(P_i, I_i, T_i)\}_{i=1}^n$ , where $P_i$ is a 3D point cloud, $I_i$ represents the corresponding image produced by projecting the 3D point cloud $P_i$ into 2D from an arbitrary perspective, and $T_i$ denotes the associated text generated using advanced vision-language models such as BLIP [21], the objective is to learn high-quality 3D representations from the triplets. The basic framework to achieve this objective is proposed by ULIP [45] (and followed by OpenShape [23]) with the help of CLIP [5, 34]. + +Formally, the 3D feature $f_{i}^{P} = E_{P}(P_{i})$ is produced by a learnable 3D encoder $E_{P}$ , and the corresponding image feature $f_{i}^{I} = E_{I}(I_{i})$ and text feature $f_{i}^{T} = E_{T}(T_{i})$ are generated by frozen CLIP encoders $E_{I}, E_{T}$ . Then the 3D encoder $E_{P}$ is optimized by aligning the 3D feature space $f^{P}$ simultaneously to the pre-aligned CLIP image space $f^{I}$ and text space $f^{T}$ through contrastive learning. The corresponding contrastive loss $L_{\mathrm{contrast}}(f^{M_1}, f^{M_2})$ between two + +modalities $M_{1}, M_{2}$ (3D-2D or 3D-text) is formulated as: + +$$ +- \frac {1}{2 n} \sum_ {i = 1} ^ {n} \left(\log \frac {\exp \left(f _ {i} ^ {M _ {1}} \cdot f _ {i} ^ {M _ {2}} / \tau\right)}{\sum_ {j = 1} ^ {n} \exp \left(f _ {i} ^ {M _ {1}} \cdot f _ {j} ^ {M _ {2}} / \tau\right)} + \log \frac {\exp \left(f _ {i} ^ {M _ {2}} \cdot f _ {i} ^ {M _ {1}} / \tau\right)}{\sum_ {j = 1} ^ {n} \exp \left(f _ {i} ^ {M _ {2}} \cdot f _ {j} ^ {M _ {1}} / \tau\right)}\right), +$$ + +where $\tau$ denotes the temperature hyperparameter. + +# 3.1. Image-Text Re-Alignment + +Unlike ULIP [45] or OpenShape [23], in which the 3D feature space is aligned with the image-text feature spaces directly produced by CLIP, we introduce an image feature space tuning strategy, aiming to foster better alignment across 3D, 2D, and language modalities (Figure 2-left). We argue that the image feature space produced directly by the CLIP visual encoder does not perfectly align with the text feature space and is thus sub-optimal. The reason is that the 2D images in the triplets, which originate from 3D point cloud projections, lack backgrounds and are visually different from natural images on which CLIP is pre-trained. + +Therefore, when using such 2D images from a shifted data domain to perform multi-modal pre-training, it becomes necessary to further fine-tune CLIP and re-align its image and language feature spaces. For the first time, we redesign domain adapters [18] for multi-modal contrastive 3D representation learning, and propose a CLIP Image Adapter (CIA), to adapt the image feature space for the rendered 2D images in our data triplets. We append a lightweight, learnable multi-layer perceptron (MLP) to CLIP image encoder, with a residual connection which seamlessly integrates the new knowledge acquired from fine-tuning with the existing knowledge from the pre-trained CLIP backbone. To avoid heavy computation and over-fitting, we only fine-tune the additional parameters in CIA, instead of the whole CLIP backbone. Formally, given the image feature $f_{i}^{I}$ and text feature $f_{i}^{T}$ extracted from the triplet $(P_{i}, T_{i}, I_{i})$ , we use a learnable, two-layer CIA $A_{C}(\cdot)$ to adapt the image feature, formulated as: + +$$ +A _ {C} \left(f _ {i} ^ {I}\right) = \sigma \left(f _ {i} ^ {I} W _ {1}\right) \cdot W _ {2}, \tag {2} +$$ + +where $W_{1}$ and $W_{2}$ are the parameters associated with the linear transformation layers, and $\sigma$ is the non-linear activation function. The refined image feature $\widetilde{f}_i^I$ can be computed with a residual connection: + +$$ +\tilde {f} _ {i} ^ {I} = \alpha A _ {C} \left(f _ {i} ^ {I}\right) + (1 - \alpha) f _ {i} ^ {I}, \tag {3} +$$ + +where $\alpha$ is a hyperparameter. Finally, CIA $A_{C}(\cdot)$ is optimized by minimizing the contrastive loss function (Equation 1), instantiated as: + +$$ +\mathcal {L} _ {\text {r e a l i g n}} = L _ {\text {c o n t r a s t}} \left(\widetilde {f} ^ {I}, f ^ {T}\right). \tag {4} +$$ + +![](images/60b0520307b5ea6bb9dc2699ea5f4f545b27bc6b2185bd26c63d18b5b0249748.jpg) +Stage 1: Image-Text Re-Alignment +Figure 2. Overview of TAMM. Left: In Stage 1, TAMM fine-tunes a lightweight CLIP Image Adapter (CIA) through contrastive learning and re-aligns the image features with the text features to alleviate the domain shift originated from rendered images. Contrastive learning maximizes inner products between features from corresponding text-image pairs, and reduces similarities of mismatched pairs. Middle: In Stage 2, TAMM introduces Image Alignment Adapter (IAA) and Text Alignment Adapter (TAA) to decouple 3D representations into two sub-spaces: one focusing more on visual attributes and the other for semantic understanding, ensuring a more comprehensive and effective multi-modal pre-training strategy. Right: TAMM adaptively utilizes decoupled 3D features for various downstream tasks including linear probing classification (top) and zero-shot classification (bottom), achieving more robust classification results. + +![](images/c46b3fef7b9e42f88f2ce12b763efc58cbfc061cc4ede2b6e2bf236b8597ce8a.jpg) +Stage 2: Decoupled Tri-Modal Pre-Training + +# 3.2. Decoupled Tri-Modal Pre-Training + +Although the image and language feature spaces have been aligned by CLIP and our CLIP Image Adapter, they still encode subtly different information. For instance, the color of a 3D shape may be inferred from the image feature, but it cannot possibly be included in the text feature if the text description does not contain color information. Similarly, the image feature may not include semantic information such as the function or name of an object. In such cases, enforcing the 3D feature to align with both the image and text features simultaneously is challenging. + +In order to overcome the obstacle introduced by aligning the 3D shape feature space with two distinct modalities, we propose a novel decoupled tri-modal pre-training framework, which aligns two decoupled 3D shape feature spaces with the refined image feature space and the text feature space, respectively (Figure 2-middle). We encourage the 3D encoder to cover both the visual representation and the semantic information inherent to the 3D shape, avoiding the dilemma of aligning with disparate image and text features at the same time. Formally, given the triplets $\{(P_i,I_i,T_i)\}_{i = 1}^n$ , we first generate the image feature $\widetilde{f}_i^I$ using the frozen, adapted 2D image encoder (Equation 3) and the text feature $f_{i}^{T}$ using the original CLIP + +text encoder. We introduce a pair of lightweight Dual Adapters: Image Alignment Adapter (IAA) $A_V(\cdot)$ and Text Alignment Adapter (TAA) $A_S(\cdot)$ . They split the 3D feature $f_i^P$ originated from the 3D encoder $E_P$ into a vision-focusing feature $f_i^{VP}$ and a semantic-focusing feature $f_i^{SP}$ , respectively. Dual Adapters $A_V(\cdot)$ and $A_S(\cdot)$ share the same architecture as the CLIP Image Adapter in Section 3.1 and can be formulated as: + +$$ +f _ {i} ^ {V P} = A _ {V} \left(f _ {i} ^ {P}\right) = \sigma \left(f _ {i} ^ {P} W _ {1} ^ {V}\right) \cdot W _ {2} ^ {V}, \tag {5} +$$ + +$$ +f _ {i} ^ {S P} = A _ {S} (f _ {i} ^ {P}) = \sigma (f _ {i} ^ {P} W _ {1} ^ {S}) \cdot W _ {2} ^ {S}, +$$ + +where $W_1^V, W_2^V, W_1^S,$ and $W_2^S$ are the parameters associated with the linear layers, and $\sigma$ represents the activation function. By decoupling 3D features into these sub-spaces, the 3D encoder $E_P$ interprets the 3D shape with a more comprehensive visual and semantic understanding, improving the expressivity of the learned representations. + +Finally, instead of enforcing the 3D shape feature space to directly mimic the pre-aligned image-text feature space, we use the decoupled 3D features $f^{VP}, f^{SP}$ and align them with the adapted image feature space $\widetilde{f}^T$ and the text feature space $f^T$ , respectively. Moreover, since one single image can only capture the 3D shape from one perspective, we align the 3D feature with the adapted features of multi-view images, to fully exploit the image modality and achieve a + +better alignment between 3D and 2D. The overall loss function is defined as: + +$$ +\mathcal {L} _ {\text {t r i m o d a l}} = L _ {\text {c o n t r a s t}} \left(f ^ {S P}, f ^ {T}\right) + \frac {1}{m} \sum_ {k} ^ {m} L _ {\text {c o n t r a s t}} \left(f ^ {V P}, \widetilde {f} ^ {I, k}\right), \tag {6} +$$ + +where $m$ represents the number of rendered images and $\widetilde{f}^{I,k}$ is the adapted image feature from the $k$ -th view. + +Application in Downstream Tasks. The learned 3D feature sub-spaces, $f^{VP}$ and $f^{SP}$ can be adaptively applied to a variety of downstream tasks (Figure 2-right). Specifically, in zero-shot 3D classification, we leverage both the 3D vision-focusing feature $f^{VP}$ and the 3D semantic-focusing feature $f^{SP}$ . We calculate the similarity between these features and category embeddings generated by the CLIP text encoder, respectively. After summing up the per-category similarity scores, the category with the highest similarity is chosen as the predicted class, yielding more robust and enhanced classification results compared with using a single sub-space alone. In the linear probing classification task, we concatenate the 3D vision-focusing feature $f^{VP}$ and 3D semantic-focusing feature $f^{SP}$ as input to the learnable linear classification layer, providing a more comprehensive and robust representation for 3D understanding. + +# 4. Experiments + +Pre-Training Datasets. Following the prior state-of-the-art method, OpenShape [23], our TAMM is pre-trained on the triplets generated from four datasets: ShapeNetCore [3], 3D-FUTURE [12], ABO [7], and Objaverse [9]. Our training sets are defined as follows: "ShapeNet" is a triplet set derived exclusively from the ShapeNetCore dataset, containing 52,470 3D shapes with corresponding images and text; "Ensembled (no LVIS)" is a set of 829,460 triplets from the above datasets excluding Objaverse-LVIS; "Ensembled" denotes the triplet set comprising data from all four datasets, containing 875,665 3D shapes and their associated images and text. + +Evaluation Datasets. Our TAMM is evaluated on the following datasets: Objaverse-LVIS [9], ModelNet40 [43], ScanObjectNN [40], and ScanNet [8]. Objverse-LVIS encompasses a broad range of categories, featuring 46,832 high-quality shapes distributed across 1,156 LVIS [13] categories. ModelNet40 is a synthetic indoor 3D dataset comprising 40 categories. We use the test split of 2,468 shapes in our experiments. ScanObjectNN consists of scanned objects from 15 common categories. It has three main variants: OBJ-BG, OBJ-ONLY, and PB-T50-RS. ScanNet, characterized by its real-world scans, includes 1,513 indoor scenes containing 36,213 objects. We conduct experiments on four tasks including zero-shot 3D classification, linear probing 3D classification, few-shot linear probing 3D classification, and real-world recognition, to demonstrate the advantages of our TAMM. Other implementation details re + +garding pre-training and evaluation are introduced in the supplementary material. + +# 4.1. Zero-Shot 3D Classification + +Zero-shot ability is a key metric for reflecting the quality of learned 3D representations, which requires the representations to be directly applicable to datasets where the model has never been explicitly supervised. Without any further tuning, the 3D representations are compared with text embeddings of categories to predict the classes of 3D shapes. To make a fair comparison and keep consistency with prior work, we adopt the same settings as OpenShape [23] on three benchmarks: Objaverse-LVIS [9], ModelNet40 [43], and OBJ-ONLY (ScanObjectNN) [40]. + +The results are summarized in Table 1. First, we observe that TAMM, benefited from multi-modal pre-training, outperforms PointCLIP [51] and PointCLIP v2 [56] by a large margin. Furthermore, it is evident that our pretrained models consistently outperform those pre-trained by ULIP and OpenShape, irrespective of whether they are pretrained on smaller datasets like ShapeNet or more expansive ones such as the Ensembled dataset. For instance, PointBERT [47] pre-trained by our TAMM on the Ensembled (no LVIS) dataset surpasses ULIP and OpenShape by margins of $+20.6\%$ and $+2.9\%$ in Top-1 accuracy on the long-tailed Objaverse-LVIS benchmark, which validates the effectiveness of our multi-modal pre-training scheme. + +# 4.2. Linear Probing 3D Classification + +To further evaluate the 3D representation quality of our pre-trained models, we design and conduct linear probing classification experiments. Here, we first freeze our pretrained Point-BERT model along with the Image Alignment Adapter (IAA) and Text Alignment Adapter (TAA), and then append a single learnable linear layer to the model. Given a batch of point clouds, as illustrated in Figure 2, we generate features from both IAA and TAA and concatenate them, and learn the appended linear classification layer on the concatenated features. We evaluate the accuracy of the linear classifier on three benchmarks: Objverse-LVIS [9], ModelNet40 [43], and ScanObjectNN [40]. Objverse-LVIS is a challenging long-tailed dataset with 1,156 categories, which has not been evaluated by prior methods [23, 45]. To ensure evaluation with statistically meaningful number of samples per class, we exclude categories with fewer than 10 instances, leaving us 1,046 categories. Samples in each class are divided into training and testing sets at an $8:2$ ratio. For ModelNet40 and ScanObjectNN, we use the standard splits, following [32, 42]. + +The results are shown in Table 2. TAMM consistently outperforms all previous methods by a large margin. For example, pre-trained on the Ensembled dataset, TAMM improves over OpenShape by $+11.2\%$ and $+2.2\%$ overall ac + +
Pre-Training DatasetModelPre-Training MethodObjaverse-LVIS [9]ModelNet40 [43]ScanObjectNN [40]
Top-1Top-3Top-5Top-1Top-3Top-5Top-1Top-3Top-5
-PointCLIP [51]-1.94.15.819.328.634.810.520.830.6
-PointCLIP v2 [56]-4.79.512.963.677.985.042.263.374.5
ShapeNetViT-B/32 [11]CLIP2Point [19]2.75.87.949.571.381.225.544.659.4
Transformer [41]ReCon [32]1.12.73.761.273.978.142.362.575.6
SparseConv [6]OpenShape [23]11.621.827.172.987.293.052.772.783.6
TAMM (Ours)13.624.229.374.688.294.057.975.383.1
Point-BERT [47]ULIP [45]6.213.617.960.479.084.451.571.180.2
OpenShape [23]10.820.225.070.386.991.351.369.478.4
TAMM (Ours)13.724.229.273.188.591.954.874.583.3
Ensembled (no LVIS)SparseConv [6]OpenShape [23]37.058.466.982.695.097.554.976.887.0
TAMM (Ours)39.862.070.485.796.898.357.581.390.0
Point-BERT [47]ULIP [45]21.438.146.071.484.489.246.066.176.4
OpenShape [23]39.160.868.985.396.297.447.272.484.7
TAMM (Ours)42.063.671.786.396.698.156.778.386.1
EnsembledSparseConv [6]OpenShape [23]43.464.872.483.495.697.856.778.988.6
TAMM (Ours)43.866.274.185.496.498.158.581.389.5
Point-BERT [47]ULIP [45]26.844.852.675.188.193.251.672.582.3
OpenShape [23]46.869.177.084.496.598.052.279.788.7
TAMM (Ours)50.773.280.685.096.698.155.780.788.9
+ +Table 1. Zero-shot 3D classification results. TAMM sets new state of the art in zero-shot classification accuracy across Objaverse-LVIS, ModelNet-40, and ScanObjectNN benchmarks, outperforming existing methods in diverse settings of pre-training datasets and 3D model architectures. The performance gain is more significant on the most challenging long-tailed Objverse-LVIS dataset. + +
Pre-Training DatasetMethodO-LVIS [9]M-40 [43]ScanObjectNN [40]
OBJ-BGOBJ-ONLYPB-T50-RS
ShapeNetULIP [45]34.690.675.475.464.8
OpenShape [23]29.388.577.878.564.1
TAMM (Ours)39.191.080.681.168.5
Rel. Improv.+9.8+2.5+2.8+2.6+4.4
EnsembledOpenShape [23]48.391.385.985.478.0
TAMM (Ours)59.593.588.588.080.3
Rel. Improv.+11.2+2.2+2.6+2.6+1.7
+ +curacy on Objaverse-LVIS and ModelNet40, respectively. This demonstrates that TAMM effectively exploits knowledge from CLIP and learns generalizable 3D representations that are directly applicable to novel linear classification tasks. Notably, TAMM even surpasses the Point-BERT model pre-trained without multi-modality but allowed to be fully fine-tuned on OBJ-BG (87.4%) [47] by +1.1% accuracy. These strong results indicate that TAMM has learned excellent transferable 3D representations, showing a great potential in real-world applications. + +# 4.3. Few-Shot Linear Probing 3D Classification + +Following the previous evaluation, we perform few-shot classification experiments on ModelNet40 [43] to assess TAMM in low-data scenarios. Similar to the linear probing + +Table 2. Linear probing 3D classification results. TAMM outperforms previous methods by a large margin, e.g., +11.2% accuracy gain on the challenging Objaverse-LVIS dataset. + +
Pre-Training DatasetMethod5-way10-way
10-shot20-shot10-shot20-shot
ShapeNetULIP [45]94.4±3.793.2±4.286.6±5.390.6±5.2
OpenShape [23]95.3±2.697.9±3.989.2±5.192.9±3.9
TAMM (Ours)97.8±1.998.1±1.495.3±3.995.9±2.6
EnsembledOpenShape [23]96.1±2.795.7±2.589.1±4.691.8±3.7
TAMM (Ours)99.0±1.399.4±0.796.8±2.997.4±2.2
+ +Table 3. Few-shot linear probing classification results on ModelNet40. We report the average accuracy and standard deviation of 10 independent experiments. Our TAMM consistently achieves both the best average accuracy and the lowest variance in various few-shot settings. + +experiments described in Section 4.2, we extend our model with an additional linear classification layer and only train this linear layer instead of fine-tuning the entire model. Following previous work [47, 50], we adopt the " $K$ -way $N$ -shot" configuration, wherein we select $K$ classes at random and sample $(N + 20)$ instances per class. Training is conducted on a support set of $K \times N$ samples, while evaluation is based on a query set comprised of the remaining 20 instances per class. We assess our model under four distinct scenarios: "5-way 10-shot," "5-way 20-shot," "10-way 10-shot," and "10-way 20-shot." For each scenario, we carry out 10 separate trials, and report both the mean performance and the standard deviation across these trials. + +As illustrated in Table 3, our TAMM achieves the best results and sets a new state of the art in the few-shot clas- + +
MethodAvg.BedCabChairSofaTablDoorWindBksfPicCntrDeskCurtFridgBathShowrToilSink
CLIP2Point [19]24.920.80.085.143.326.569.90.020.91.731.727.00.01.646.50.022.425.6
PointCLIP w/ TP. [56]26.10.055.772.85.05.11.70.077.20.00.051.70.30.00.040.385.349.2
CLIP2Point w/ TP. [19]35.211.83.045.127.610.561.52.671.90.333.629.94.711.572.292.486.134.0
\( CLIP^2 \) [49]38.532.667.269.342.318.319.14.062.61.412.752.840.19.159.741.071.045.5
\( OpenShape^\dagger \) [23]45.666.73.275.883.537.749.247.564.948.21.966.170.21.850.057.145.27.1
TAMM (Ours)†49.466.74.883.684.548.957.948.280.561.31.960.683.67.041.456.148.43.6
+ +w/ TP. denotes training with the real-world data provided by $\mathrm{CLIP^2}$ . Results using Point-BERT [47] as 3D encoder, pre-trained on the Ensembled dataset. + +Table 4. Zero-shot classification results on the real-world ScanNet dataset. Avg.: the mean average Top-1 accuracy of all classes. TAMM achieves the best results. + +
CIAIAATAAImg-Txt AccObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
✓✓40.113.523.929.172.888.291.7
✓60.912.522.427.371.486.089.8
✓✓60.912.922.927.773.888.292.5
✓✓60.913.023.028.272.987.890.4
✓✓✓60.913.724.229.273.188.591.9
+ +(a) Adapters in pre-training. Pre-training with CLIP Image Adapter (CIA), Image Alignment Adapter (IAA), and Text Alignment Adapter (TAA) is the most effective. + +
StageImg-Txt AccObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
One stage55.312.221.427.071.486.490.7
Two stages60.913.724.229.273.188.591.9
+ +(b) Pre-training stages. Learning the CLIP Image Adapter first and then pre-train the 3D encoder with Dual Adapters is better than training modules all together. + +
ImageTextObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
✓11.920.825.967.687.392.5
✓10.519.223.770.285.088.0
✓✓13.724.229.273.188.591.9
+ +(c) Pre-training modalities. Exploiting both image and text data provides the most performance gain. They contribute almost equally in TAMM. + +
IAATAAObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
✓13.023.128.368.186.591.4
✓12.922.527.572.486.290.4
✓✓13.724.229.273.188.591.9
+ +(d) Adapters for inference. Dual Adapters are complementary, and their combination more accurately classifies 3D shapes at test time. + +
ImagesObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
113.623.628.972.588.092.6
213.523.528.873.187.692.8
413.724.229.273.188.591.9
813.523.128.473.089.092.8
+ +(e) Number of images. Aligning a 3D shape with 4 images of it is the best. + +Table 5. Ablation study of components in TAMM. We pre-train Point-BERT models on ShapeNet in various settings and evaluate their zero-shot classification performance on Objaverse-LVIS and ModelNet40. The baseline method OpenShape [23] achieves $10.8\%$ and $70.3\%$ Top-1 accuracy on these two benchmarks (Table 1), respectively. The setting adopted by TAMM is marked. + +sification task. TAMM shows both higher overall accuracy and smaller deviations than other methods, showcasing the generalizability and robustness of TAMM-learned 3D representations. For example, when pre-trained on the Ensembled dataset, our TAMM surpasses OpenShape [23] by $2.9\%$ , $3.7\%$ , $7.7\%$ , $5.6\%$ , respectively for all four settings. These results indicate that our TAMM is able to learn 3D representations that are more generalizable and can be readily adapted to downstream tasks under low-data regimes. + +# 4.4. Real-World Recognition + +To evaluate the capability of TAMM in understanding $3D$ shapes from the real world, we follow the previous work $\mathrm{CLIP}^2$ [49] and test TAMM on a real-world recognition task, in which the model aims at correctly classifying each instance from a complex scene in a zero-shot manner. Specifically, we select the real-world scene-level dataset ScanNet [8] and adopt same data splits as [49], containing 17 classes. We perform zero-shot classification (Section 4.1) on the point cloud of each object instance extracted from scenes, and report both the Top-1 accuracy of each class and the mean accuracy of all 17 classes. As shown in Table 4, TAMM significantly outperforms all prior meth + +ods, showing improvements of $3.8\%$ and $10.9\%$ in average Top-1 accuracy compared with OpenShape and $\mathrm{CLIP^2}$ , respectively. These results underscore TAMM's ability in recognizing and understanding 3D shapes captured from real-world scenarios. Further results on complex scene recognition and instance segmentation are included in the supplementary material. + +# 4.5. Ablation Study + +In this section, we provide additional experimental results to further test the performance gain from each design of TAMM. Due to limited computation, we pre-train PointBERT [47] on the small-scale ShapeNet [3], and use two more challenging zero-shot classification benchmarks ModelNet40 [43] and Objaverse-LVIS [9] for evaluation. We alter only one component in each set of experiments. + +Adapters in Pre-Training. As described in Section 3.1, our CLIP Image Adapter (CIA) mitigates the domain gap between rendered images and natural images. We first measure the contrastive accuracy, computed as the ratio of image-text pairs where the text feature is more similar to the image from the same triplet than any other text. The contrastive accuracy of CLIP without CIA is only $40.1\%$ , + +indicating that the original image-text feature space is compromised by the domain gap. CIA brings the accuracy up to $60.9\%$ , showing the effectiveness of our CIA in resolving the shifted data domain. Furthermore, as shown in Table 5a, CIA increases the zero-shot accuracy on Objaverse-LVIS dataset from $12.2\%$ to $13.7\%$ , validating its effectiveness on improving the pre-trained 3D representations. + +TAMM adopts Image Alignment Adapter (IAA) and Text Alignment Adapter (TAA) to decouple the feature space and enrich the learned 3D representations. To further explore the performance gains introduced by our Dual Adapters (IAA and TAA), we experiment tri-modal pretraining with or without IAA and TAA. As shown in Table 5a, both adapters bring performance gains, and their combination achieves the best result, demonstrating the effectiveness of Dual Adapters. + +Pre-Training Stages. We investigate the significance of our two-stage pre-training design, which initially learns CLIP Image Adapter (CIA) to re-align image-text pairs and subsequently employs Dual Adapters in decoupled tri-modal pre-training. An alternative could be learning these modules all together in one stage. As shown in Table 5b, remarkably, the two-stage pre-training approach achieves a better performance compared with one-stage pre-training. + +Pre-Training Modalities. We further explore the effectiveness of image and language modalities in pre-training. Table 5c shows that integrating the 3D modality with both the image and language modalities consistently yields superior performance across various benchmarks, compared with aligning it with either modality alone. Moreover, unlike prior methods which ineffectively learn from images (Figure 1-left), the image and language modalities contribute almost equally in TAMM, which also signifies that TAMM better exploits the image modality and brings more gains by learning from both image and text data. + +Adapters for Inference. Our IAA and TAA decouple 3D features with visual and semantic focuses, respectively, offering a more accurate and comprehensive understanding for 3D shapes. At inference time, we combine the outputs from Dual Adapters for classification tasks. To investigate this design, we separately utilize features produced solely by either IAA or TAA in zero-shot classification evaluation. As shown in Table 5d, the combined features achieve more accurate classification, demonstrating that the learned features from IAA and TAA are complementary to each other. + +Number of Images. In order to gain more comprehensive knowledge from the image modality, TAMM simultaneously aligns the 3D feature with multi-view 2D images, each projected from a different perspective. We evaluate this design by aligning the 3D feature with a varying number of corresponding 2D images, ranging from 1 to 8 images. As shown in Table 5e, multi-view image features offer a performance gain by fully exploiting the image modality. + +![](images/fd8efb0de0c4c7191a532931e2cf86427657a9ef603bd873c0deaaf26c026870.jpg) +Figure 3. Qualitative results. Top: CIA re-aligns the images rendered from 3D shapes with the text descriptions. The rendered images are inaccurately matched with text when the image features are directly extracted by CLIP, and CIA can correct the matching. Bottom: IAA and TAA decouple 3D features with complementary visual and semantic focuses. Features from one single adapter are matched with classes whose appearance or semantics resemble the true class; using both adapters leads to the correct class. + +Aligning a 3D shape with 4 views achieves the best result. + +Qualitative Results. Finally, we provide some visualizations to intuitively demonstrate the benefit of our proposed TAMM. As illustrated in Figure 3, CIA successfully mitigates the domain gap introduced by rendered images, and leads to the more accurate matching between images and text. IAA and TAA learn 3D representations with subtly different but complementary focuses on vision and semantics, respectively, and their combination brings more robust and comprehensive 3D representations. More visualized results are included in the supplementary material. + +# 5. Conclusion + +In this work, we examine the sub-optimal utilization of 2D images in existing multi-modal pre-training methods for 3D shape understanding, and propose TriAdapter MultiModal Learning (TAMM), a novel two-stage representation learning approach built on three synergistic adapter modules. Extensive experiments verify that TAMM consistently learns improved 3D features in various settings. + +Limitations. Due to the limited computation resources, we are not able to perform pre-training on very large-scale 3D backbones with billions of parameters. The quality of the learned 3D representation could be further improved if the 3D backbones are scaled to a larger size. + +Acknowledgement. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Jump ARCHES endowment, and the IBM-Illinois Discovery Accelerator Institute. This work used NVIDIA GPUs at NCSA Delta through allocations CIS220014, CIS230012, and CIS230013 from the ACCESS program. + +# References + +[1] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 2 +[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2 +[3] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 2, 5, 7 +[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2 +[5] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 2, 3 +[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D spatio-temporal ConvNets: Minkowski convolutional neural networks. In CVPR, 2019. 2, 3, 6 +[7] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F. Yago Vicente, Thomas Dideriksien, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. ABO: Dataset and benchmarks for real-world 3D object understanding. In CVPR, 2022. 5 +[8] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 5, 7 +[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, 2023. 2, 5, 6, 7 +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2, 3 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 6 + +[12] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3D-FUTURE: 3D furniture shape with texture. IJCV, 129:3313-3337, 2021. 5 +[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 5 +[14] Huy Ha and Shuran Song. Semantic abstraction: Openworld 3D scene understanding from 2D vision-language models. In CoRL, 2022. 3 +[15] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2 +[16] Deepti Hegde, Jeya Maria Jose Valanarasu, and Vishal Patel. CLIP goes 3D: Leveraging prompt tuning for language grounded 3D recognition. In ICCV, 2023. 1 +[17] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. AvatarCLIP: Zero-shot text-driven generation and animation of 3D avatars. ACM Transactions on Graphics, 41(4):1-19, 2022. 3 +[18] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. In ICML, 2019. 3 +[19] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. CLIP2Point: Transfer CLIP to point cloud classification with image-depth pre-training. In ICCV, 2023. 1, 3, 6, 7 +[20] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, 2022. 3 +[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 3 +[22] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 3 +[23] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. OpenShape: Scaling up 3D shape representation towards open-world understanding. In NeurIPS, 2023. 1, 2, 3, 5, 6, 7 +[24] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In ICLR, 2022. 1, 2 +[25] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2Mesh: Text-driven neural stylization for meshes. In CVPR, 2022. 3 +[26] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In CVPR, 2019. 2 +[27] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. CLIP-Mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia, 2022. 3 + +[28] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 2 +[29] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. OpenScene: 3D scene understanding with open vocabularies. In CVPR, 2023. 3 +[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In CVPR, 2017. 2 +[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. PointNet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2 +[32] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3D representation learning guided by generative pretraining. In ICML, 2023. 5, 6 +[33] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. PointNeXt: Revisiting PointNet++ with improved training and scaling strategies. In NeurIPS, 2022. 2 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3 +[35] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3D semantic segmentation in the wild. In ECCV, 2022. 3 +[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: An open large-scale dataset for training next generation image-text models. In NeurIPS Datasets and Benchmarks Track, 2022. 2 +[37] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. PV-RCNN: Point-voxel feature set abstraction for 3D object detection. In CVPR, 2020. 2 +[38] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. MotionCLIP: Exposing human motion generation to CLIP space. In ECCV, 2022. 3 +[39] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[40] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 2, 5, 6 +[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3, 6 + +[42] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point Transformer V2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 1, 2, 3, 5 +[43] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In CVPR, 2015. 2, 5, 6, 7 +[44] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3D: Zero-shot text-to-3D synthesis using 3D shape prior and text-to-image diffusion models. In CVPR, 2023. 3 +[45] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. ULIP: Learning a unified representation of language, images, and point clouds for 3D understanding. In CVPR, 2023. 1, 2, 3, 5, 6 +[46] Jihan Yang, Runyu Ding, Zhe Wang, and Xiaojuan Qi. RegionPLC: Regional point-language contrastive learning for open-world 3D scene understanding. arXiv preprint arXiv:2304.00962, 2023. 3 +[47] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-BERT: Pre-training 3D point cloud transformers with masked point modeling. In CVPR, 2022. 1, 2, 3, 5, 6, 7 +[48] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In CVPR, 2021. 3 +[49] Yihan Zeng, Chenhan Jiang, Jiageng Mao, Jianhua Han, Chaoqiang Ye, Qingqiu Huang, Dit-Yan Yeung, Zhen Yang, Xiaodan Liang, and Hang Xu. CLIP $^2$ : Contrastive language-image-point pretraining from real-world point cloud data. In CVPR, 2023. 7 +[50] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. PointM2AE: Multi-scale masked autoencoders for hierarchical point cloud pre-training. In NeurIPS, 2022. 2, 6 +[51] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. PointCLIP: Point cloud understanding by CLIP. In CVPR, 2022. 1, 3, 5, 6 +[52] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3D features on any point-cloud. In ICCV, 2021. 2 +[53] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. RegionCLIP: Region-based language-image pretraining. In CVPR, 2022. 3 +[54] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3D: Exploring unified 3D representation at scale. In ICLR, 2024. 3 +[55] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 3 +[56] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Point- + +CLIP V2: Prompting CLIP and GPT for powerful 3D open-world learning. In ICCV, 2023. 1, 3, 5, 6, 7 \ No newline at end of file diff --git a/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/images.zip b/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e1a0731b50a6b507a1215902bb00e46a460f218b --- /dev/null +++ b/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c44664ca56ef9ec75a6d97dc904ae9e207f714dfa5fe9ea54e2a61295c473c01 +size 584101 diff --git a/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/layout.json b/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..77c3f58a3ab28c59321e7fac002aba1d29dbdcbd --- /dev/null +++ b/2024/TAMM_ TriAdapter Multi-Modal Learning for 3D Shape Understanding/layout.json @@ -0,0 +1,8792 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 521, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 521, + 121 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 521, + 121 + ], + "type": "text", + "content": "TAMM: TriAdapter Multi-Modal Learning for 3D Shape Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 142, + 475, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 142, + 475, + 186 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 475, + 186 + ], + "type": "text", + "content": "Zhihao Zhang" + }, + { + "bbox": [ + 115, + 142, + 475, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 115, + 142, + 475, + 186 + ], + "type": "text", + "content": " Shengcao Cao" + }, + { + "bbox": [ + 115, + 142, + 475, + 186 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 115, + 142, + 475, + 186 + ], + "type": "text", + "content": " Yu-Xiong Wang" + }, + { + "bbox": [ + 115, + 142, + 475, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 142, + 475, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 142, + 475, + 186 + ], + "type": "text", + "content": "Xi'an Jiaotong University " + }, + { + "bbox": [ + 115, + 142, + 475, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 142, + 475, + 186 + ], + "type": "text", + "content": "University of Illinois Urbana-Champaign \n" + }, + { + "bbox": [ + 115, + 142, + 475, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 142, + 475, + 186 + ], + "type": "text", + "content": "zh1142@stu.xjtu.edu.cn " + }, + { + "bbox": [ + 115, + 142, + 475, + 186 + ], + "type": "inline_equation", + "content": "^{2}\\{cao44, yxw\\} @illinois.edu" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 52, + 222, + 236, + 338 + ], + "blocks": [ + { + "bbox": [ + 52, + 222, + 236, + 338 + ], + "lines": [ + { + "bbox": [ + 52, + 222, + 236, + 338 + ], + "spans": [ + { + "bbox": [ + 52, + 222, + 236, + 338 + ], + "type": "image", + "image_path": "06b8d2bbf939a68fcb692a39846a4b7d07d46f557686a0edae21a12585c51933.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 350, + 545, + 449 + ], + "lines": [ + { + "bbox": [ + 46, + 350, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 545, + 449 + ], + "type": "text", + "content": "Figure 1. Our TriAdapter Multi-Modal Learning (TAMM) significantly enhances 3D shape understanding. Left: When aligning features of 3D point clouds (P) with 2D images (I) and/or text (T), prior methods (e.g., ULIP [45] and OpenShape [23]) under-utilize the image modality, due to the overlooked or unsolved image domain gap. TAMM better exploits the image modality and brings more gains when learning from both image and text data. The results are produced by pre-training Point-BERT [47] on ShapeNet [3]. Middle: Our CLIP Image Adapter (CIA) re-aligns the images rendered from 3D shapes with the text descriptions. The rendered images are inaccurately matched with text when the image features are directly extracted by CLIP, and CIA can correct the matching. Right: Our Image Alignment Adapter (IAA) and Text Alignment Adapter (TAA) decouple 3D features with complementary visual and semantic focuses. In the visualized examples, features from one single adapter are matched with classes whose appearance or semantics resemble the true class; using both adapters leads to the correct class." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 244, + 228, + 396, + 277 + ], + "blocks": [ + { + "bbox": [ + 244, + 228, + 396, + 277 + ], + "lines": [ + { + "bbox": [ + 244, + 228, + 396, + 277 + ], + "spans": [ + { + "bbox": [ + 244, + 228, + 396, + 277 + ], + "type": "image", + "image_path": "0db415b090078d7b48e45bd99d3a914a40d606ed9abf803a93a3ce4f554a738a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 242, + 420, + 257 + ], + "lines": [ + { + "bbox": [ + 399, + 242, + 420, + 257 + ], + "spans": [ + { + "bbox": [ + 399, + 242, + 420, + 257 + ], + "type": "text", + "content": "Point Clouds" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 435, + 223, + 468, + 275 + ], + "blocks": [ + { + "bbox": [ + 435, + 223, + 468, + 275 + ], + "lines": [ + { + "bbox": [ + 435, + 223, + 468, + 275 + ], + "spans": [ + { + "bbox": [ + 435, + 223, + 468, + 275 + ], + "type": "image", + "image_path": "8b8971a677425e6d6c3627af7c8a303d46746dc4691e9e240a62c5c57323c367.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 496, + 222, + 533, + 277 + ], + "blocks": [ + { + "bbox": [ + 496, + 222, + 533, + 277 + ], + "lines": [ + { + "bbox": [ + 496, + 222, + 533, + 277 + ], + "spans": [ + { + "bbox": [ + 496, + 222, + 533, + 277 + ], + "type": "image", + "image_path": "0c1fa06a9826da8db0e3dae11c6bc72bede6338ee24b14bd2a8e49dff4b56f03.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 244, + 283, + 396, + 325 + ], + "blocks": [ + { + "bbox": [ + 244, + 283, + 396, + 325 + ], + "lines": [ + { + "bbox": [ + 244, + 283, + 396, + 325 + ], + "spans": [ + { + "bbox": [ + 244, + 283, + 396, + 325 + ], + "type": "table", + "html": "
Matching w/o CIAa gray box with four screws on ita bedside table with a lamp
Matching w/ CIAa desk with drawers on ita bed with white sheet and a lamp
", + "image_path": "ad808842328af6a33011e3b0c47c60829626199c1833883211d7a1d4dade7b01.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 403, + 283, + 417, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 283, + 417, + 292 + ], + "spans": [ + { + "bbox": [ + 403, + 283, + 417, + 292 + ], + "type": "text", + "content": "IAA" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 437, + 285, + 463, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 437, + 285, + 463, + 293 + ], + "spans": [ + { + "bbox": [ + 437, + 285, + 463, + 293 + ], + "type": "text", + "content": "dress hat" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 503, + 285, + 522, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 503, + 285, + 522, + 293 + ], + "spans": [ + { + "bbox": [ + 503, + 285, + 522, + 293 + ], + "type": "text", + "content": "mallet" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 403, + 300, + 418, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 300, + 418, + 308 + ], + "spans": [ + { + "bbox": [ + 403, + 300, + 418, + 308 + ], + "type": "text", + "content": "TAA" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 435, + 300, + 466, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 435, + 300, + 466, + 308 + ], + "spans": [ + { + "bbox": [ + 435, + 300, + 466, + 308 + ], + "type": "text", + "content": "lampshade" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 501, + 300, + 525, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 501, + 300, + 525, + 308 + ], + "spans": [ + { + "bbox": [ + 501, + 300, + 525, + 308 + ], + "type": "text", + "content": "postbox" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 403, + 316, + 417, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 316, + 417, + 324 + ], + "spans": [ + { + "bbox": [ + 403, + 316, + 417, + 324 + ], + "type": "text", + "content": "Dual" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 435, + 317, + 466, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 435, + 317, + 466, + 324 + ], + "spans": [ + { + "bbox": [ + 435, + 317, + 466, + 324 + ], + "type": "text", + "content": "mushroom" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 501, + 317, + 525, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 501, + 317, + 525, + 324 + ], + "spans": [ + { + "bbox": [ + 501, + 317, + 525, + 324 + ], + "type": "text", + "content": "newsletter" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 143, + 464, + 192, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 464, + 192, + 477 + ], + "spans": [ + { + "bbox": [ + 143, + 464, + 192, + 477 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 45, + 490, + 288, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 490, + 288, + 706 + ], + "spans": [ + { + "bbox": [ + 45, + 490, + 288, + 706 + ], + "type": "text", + "content": "The limited scale of current 3D shape datasets hinders the advancements in 3D shape understanding, and motivates multi-modal learning approaches which transfer learned knowledge from data-abundant 2D image and language modalities to 3D shapes. However, even though the image and language representations have been aligned by cross-modal models like CLIP, we find that the image modality fails to contribute as much as the language in existing multi-modal 3D representation learning methods. This is attributed to the domain shift in the 2D images and the distinct focus of each modality. To more effectively leverage both modalities in the pre-training, we introduce TriAdapter Multi-Modal Learning (TAMM) - a novel two-stage learning approach based on three synergistic adapters. First, our CLIP Image Adapter mitigates the domain gap between 3D-rendered images and natural images, by adapting the visual representations of CLIP for synthetic image-text pairs. Subsequently, our Dual Adapters decou" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 466, + 545, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 545, + 610 + ], + "type": "text", + "content": "ple the 3D shape representation space into two complementary sub-spaces: one focusing on visual attributes and the other for semantic understanding, which ensure a more comprehensive and effective multi-modal pre-training. Extensive experiments demonstrate that TAMM consistently enhances 3D representations for a wide range of 3D encoder architectures, pre-training datasets, and downstream tasks. Notably, we boost the zero-shot classification accuracy on Objaverse-LVIS from " + }, + { + "bbox": [ + 306, + 466, + 545, + 610 + ], + "type": "inline_equation", + "content": "46.8\\%" + }, + { + "bbox": [ + 306, + 466, + 545, + 610 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 466, + 545, + 610 + ], + "type": "inline_equation", + "content": "50.7\\%" + }, + { + "bbox": [ + 306, + 466, + 545, + 610 + ], + "type": "text", + "content": ", and improve the 5-way 10-shot linear probing classification accuracy on ModelNet40 from " + }, + { + "bbox": [ + 306, + 466, + 545, + 610 + ], + "type": "inline_equation", + "content": "96.1\\%" + }, + { + "bbox": [ + 306, + 466, + 545, + 610 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 466, + 545, + 610 + ], + "type": "inline_equation", + "content": "99.0\\%" + }, + { + "bbox": [ + 306, + 466, + 545, + 610 + ], + "type": "text", + "content": ". Project page: https://alanzhangcs.github.io/tamm-page." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 634, + 386, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 386, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 386, + 646 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 654, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 715 + ], + "type": "text", + "content": "Despite the recent success of 3D shape understanding [16, 19, 24, 42, 47, 51, 56], the limited scale of existing 3D shape datasets hinders the learning of more robust and generalizable 3D representations. Due to the significant human labor and expertise required in collecting and annotating 3D" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 715, + 126, + 725 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 715, + 126, + 725 + ], + "spans": [ + { + "bbox": [ + 58, + 715, + 126, + 725 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "text", + "content": "21413" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "shape data, building large-scale 3D shape datasets is a prohibitive endeavor [3, 9, 26]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 98, + 288, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 98, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 98, + 288, + 302 + ], + "type": "text", + "content": "To mitigate the challenge of the limited data scale, recent research in 3D shape representation learning shows a promising direction of multi-modal learning [23, 45]. By establishing connections among data of 3D shapes, 2D images, and text, it becomes possible to pre-train 3D shape representations by distilling learned knowledge from image and language modalities. There exists a comprehensive body of literature of massive datasets [36], high-quality models pre-trained via self-supervised learning for 2D image [2, 4, 15] and language [1, 10, 39] modalities, and furthermore, cross-modal vision-language models [5, 34]. For instance, ULIP [45] creates triplets of 3D point clouds, 2D images, and text by rendering images from 3D shapes and generating text descriptions from their metadata. Then, ULIP adopts contrastive learning to align the 3D shape features with both the image features and text features extracted by CLIP [34]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 303, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 287, + 411 + ], + "type": "text", + "content": "While the image and language modalities are pre-aligned by CLIP pre-training and share the same feature space, we observe that directly aligning 3D features with image features leads to considerably worse representation quality as compared with aligning 3D features with text features, as shown in Figure 1-left. This phenomenon suggests that existing methods (e.g., ULIP [45] and OpenShape [23]) are not optimally leveraging the 2D image modality in 3D representation pre-training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 412, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 628 + ], + "type": "text", + "content": "There are two reasons for this counter-intuitive observation: 1) The 2D images in the generated triplets follow a data distribution different from natural images, on which CLIP is pre-trained. The images are rendered or projected from 3D to 2D, usually lacking a realistic background and texture. Since the image domain is shifted, the image features are no longer well-aligned with the text features, as exemplified in Figure 1-middle. 2) Intuitively, the image features represent more visual attributes including the shape, texture, or color, while the text features have a focus on semantics such as the function of the object. As shown in Figure 1-right, if the 3D features are specifically aligned with one single modality of images or text, they pay attention to different aspects of the 3D shape. Therefore, enforcing the 3D shape to simultaneously align with two modalities that convey subtly distinct information could be challenging. Such an approach may not fully leverage the potential of multi-modal learning signals." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": "Aiming at addressing these two issues hindering multimodal pre-training for 3D shape understanding, we propose TriAdapter Multi-Modal Learning (TAMM), a two-stage pre-training approach based on three synergistic adapters. In the first stage, to mitigate the domain gap between the 2D images rendered from 3D shapes and the natural images on which CLIP is pre-trained, we adapt the visual representa" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "tions of CLIP based on the synthetic image-text pairs. More specifically, we fine-tune a lightweight CLIP Image Adapter on top of CLIP visual encoder through contrastive learning and re-align the adapted image features with the text features in the new domain. This CLIP Image Adapter allows us to establish more accurate relations between 3D shapes, 2D images, and language in an updated feature space, and avoids learning 3D representations from mismatched image features and text features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 180, + 546, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 546, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 546, + 395 + ], + "type": "text", + "content": "In the following second stage, to prevent the vision-semantic feature disparity from impairing our 3D representation pre-training, we choose to embrace this disparity and decouple the 3D representations into two sub-spaces. To comprehensively encode a 3D shape, the 3D encoder needs to capture both the visual and semantic aspects of its representation, which are centric to the corresponding image and text features, respectively. Therefore, we decouple the 3D feature space for the two focuses on visual and semantic representations. In particular, we attach two independent feature adapters to the 3D backbone and transform the 3D features into two sub-spaces. One sub-space focuses more on the visual representations, and is aligned with the 2D image feature space; the other sub-space focuses more on the semantic representations, and is aligned with the language feature space. This approach of decoupled feature spaces makes the learned 3D representations more comprehensive and expressive." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 396, + 508, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 396, + 508, + 407 + ], + "spans": [ + { + "bbox": [ + 317, + 396, + 508, + 407 + ], + "type": "text", + "content": "To summarize, our main contribution includes:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 407, + 545, + 586 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 407, + 545, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 545, + 454 + ], + "type": "text", + "content": "- We identify the under-utilization of the 2D image modality in existing multi-modal methods. The image domain gap and feature disparity in image-text pairs hinder representation learning in 3D shape understanding." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 455, + 544, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 544, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 544, + 502 + ], + "type": "text", + "content": "- We propose a novel multi-modal learning framework with two learning stages and three unified adapter modules. Our proposed TAMM better exploits both image and language modalities and improves 3D shape representations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 503, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 545, + 586 + ], + "type": "text", + "content": "- Our TAMM consistently enhances 3D representations for a variety of 3D encoder architectures (e.g., PointBERT [47], SparseConv [6]), pre-training datasets (e.g., ShapeNet [3], an ensembled dataset [23]), and downstream tasks (e.g., zero-shot and linear probing shape classification on Objaverse-LVIS [9], ModelNet40 [43], and ScanObjectNN [40])." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 597, + 392, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 392, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 392, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "3D Shape Understanding. There are two mainstreams for 3D shape representation learning: 1) Projecting 3D shapes into voxel or grid-based formats [37] and then using 2D/3D convolutions [6] for feature extraction. 2) Directly modeling 3D point clouds with point-centric architectures [24, 28, 30, 31, 33, 42, 50, 52]. In this work, to ensure a fair and comprehensive comparison with previous methods [23, 45] on the pre-training scheme, we follow" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 318, + 764 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 318, + 764 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 318, + 764 + ], + "type": "text", + "content": "21414" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "their selection and utilize two representative 3D encoders from these two mainstreams: SparseConv [6] and Point-BERT [47]. SparseConv is designed for efficiently processing sparse voxels using specialized convolutions that focus computations on non-zero data points. Point-BERT [42] utilizes a Transformer-based architecture [41] and can be self-supervised by masked modeling [10]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 288, + 373 + ], + "type": "text", + "content": "Multi-Modal Representation Learning. Contrastive Language-Image Pre-training (CLIP) [34] has enabled various downstream applications including object detection [48, 53, 55] and language grounding [22]. Recently, CLIP has been extended to 3D-based tasks, such as zero-shot text-to-3D generation [17, 20, 25, 27, 38, 44] and scene-level 3D segmentation [14, 29, 35, 46]. Meanwhile, developing general and robust representations for 3D shape understanding with the foundation of CLIP [19, 23, 45, 51, 54, 56] becomes a major focus. Among these methods, ULIP [45], as a pioneering work, utilizes contrastive learning to distill CLIP features into 3D representations. OpenShape [23] follows this learning paradigm with a focus on building a larger pre-training dataset with enriched and filtered text data. Unlike OpenShape [23], we focus on improving the multi-modal learning paradigm by more effectively leveraging both the image and text modalities via a two-stage pre-training approach based on three synergistic adapters." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 382, + 102, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 102, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 102, + 395 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 510 + ], + "type": "text", + "content": "Figure 2 shows our proposed two-stage approach, TriAdapter Multi-Modal Learning (TAMM), for pretraining robust and generalizable 3D representations by leveraging 2D images and text. We first revisit the triplet data generation and problem formulation in multi-modal 3D representation learning. Section 3.1 delves into our finetuning of the original CLIP model for better fitting 3D understanding tasks. Section 3.2 details our strategy to enhance alignment across the 3D, 2D, and text feature spaces." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "content": "Problem Formulation. Given " + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "content": " triplets " + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\{(P_i, I_i, T_i)\\}_{i=1}^n" + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "content": " is a 3D point cloud, " + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "content": " represents the corresponding image produced by projecting the 3D point cloud " + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "content": " into 2D from an arbitrary perspective, and " + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "content": " denotes the associated text generated using advanced vision-language models such as BLIP [21], the objective is to learn high-quality 3D representations from the triplets. The basic framework to achieve this objective is proposed by ULIP [45] (and followed by OpenShape [23]) with the help of CLIP [5, 34]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Formally, the 3D feature " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_{i}^{P} = E_{P}(P_{i})" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " is produced by a learnable 3D encoder " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "E_{P}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ", and the corresponding image feature " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_{i}^{I} = E_{I}(I_{i})" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " and text feature " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_{i}^{T} = E_{T}(T_{i})" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " are generated by frozen CLIP encoders " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "E_{I}, E_{T}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ". Then the 3D encoder " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "E_{P}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " is optimized by aligning the 3D feature space " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "f^{P}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " simultaneously to the pre-aligned CLIP image space " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "f^{I}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " and text space " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "f^{T}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " through contrastive learning. The corresponding contrastive loss " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{contrast}}(f^{M_1}, f^{M_2})" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " between two" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 535, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 535, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 535, + 84 + ], + "type": "text", + "content": "modalities " + }, + { + "bbox": [ + 306, + 72, + 535, + 84 + ], + "type": "inline_equation", + "content": "M_{1}, M_{2}" + }, + { + "bbox": [ + 306, + 72, + 535, + 84 + ], + "type": "text", + "content": " (3D-2D or 3D-text) is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 95, + 548, + 141 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 95, + 548, + 141 + ], + "spans": [ + { + "bbox": [ + 306, + 95, + 548, + 141 + ], + "type": "interline_equation", + "content": "- \\frac {1}{2 n} \\sum_ {i = 1} ^ {n} \\left(\\log \\frac {\\exp \\left(f _ {i} ^ {M _ {1}} \\cdot f _ {i} ^ {M _ {2}} / \\tau\\right)}{\\sum_ {j = 1} ^ {n} \\exp \\left(f _ {i} ^ {M _ {1}} \\cdot f _ {j} ^ {M _ {2}} / \\tau\\right)} + \\log \\frac {\\exp \\left(f _ {i} ^ {M _ {2}} \\cdot f _ {i} ^ {M _ {1}} / \\tau\\right)}{\\sum_ {j = 1} ^ {n} \\exp \\left(f _ {i} ^ {M _ {2}} \\cdot f _ {j} ^ {M _ {1}} / \\tau\\right)}\\right),", + "image_path": "23d0d974d7cef1a093efc4e757be09b2ccd286b91e69e3a10c31d88b335072bb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 143, + 506, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 143, + 506, + 155 + ], + "spans": [ + { + "bbox": [ + 306, + 143, + 506, + 155 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 143, + 506, + 155 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 306, + 143, + 506, + 155 + ], + "type": "text", + "content": " denotes the temperature hyperparameter." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 165, + 450, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 165, + 450, + 178 + ], + "spans": [ + { + "bbox": [ + 306, + 165, + 450, + 178 + ], + "type": "text", + "content": "3.1. Image-Text Re-Alignment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 183, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 183, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 545, + 316 + ], + "type": "text", + "content": "Unlike ULIP [45] or OpenShape [23], in which the 3D feature space is aligned with the image-text feature spaces directly produced by CLIP, we introduce an image feature space tuning strategy, aiming to foster better alignment across 3D, 2D, and language modalities (Figure 2-left). We argue that the image feature space produced directly by the CLIP visual encoder does not perfectly align with the text feature space and is thus sub-optimal. The reason is that the 2D images in the triplets, which originate from 3D point cloud projections, lack backgrounds and are visually different from natural images on which CLIP is pre-trained." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 316, + 546, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 546, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 546, + 532 + ], + "type": "text", + "content": "Therefore, when using such 2D images from a shifted data domain to perform multi-modal pre-training, it becomes necessary to further fine-tune CLIP and re-align its image and language feature spaces. For the first time, we redesign domain adapters [18] for multi-modal contrastive 3D representation learning, and propose a CLIP Image Adapter (CIA), to adapt the image feature space for the rendered 2D images in our data triplets. We append a lightweight, learnable multi-layer perceptron (MLP) to CLIP image encoder, with a residual connection which seamlessly integrates the new knowledge acquired from fine-tuning with the existing knowledge from the pre-trained CLIP backbone. To avoid heavy computation and over-fitting, we only fine-tune the additional parameters in CIA, instead of the whole CLIP backbone. Formally, given the image feature " + }, + { + "bbox": [ + 304, + 316, + 546, + 532 + ], + "type": "inline_equation", + "content": "f_{i}^{I}" + }, + { + "bbox": [ + 304, + 316, + 546, + 532 + ], + "type": "text", + "content": " and text feature " + }, + { + "bbox": [ + 304, + 316, + 546, + 532 + ], + "type": "inline_equation", + "content": "f_{i}^{T}" + }, + { + "bbox": [ + 304, + 316, + 546, + 532 + ], + "type": "text", + "content": " extracted from the triplet " + }, + { + "bbox": [ + 304, + 316, + 546, + 532 + ], + "type": "inline_equation", + "content": "(P_{i}, T_{i}, I_{i})" + }, + { + "bbox": [ + 304, + 316, + 546, + 532 + ], + "type": "text", + "content": ", we use a learnable, two-layer CIA " + }, + { + "bbox": [ + 304, + 316, + 546, + 532 + ], + "type": "inline_equation", + "content": "A_{C}(\\cdot)" + }, + { + "bbox": [ + 304, + 316, + 546, + 532 + ], + "type": "text", + "content": " to adapt the image feature, formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 370, + 543, + 545, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 543, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 370, + 543, + 545, + 558 + ], + "type": "interline_equation", + "content": "A _ {C} \\left(f _ {i} ^ {I}\\right) = \\sigma \\left(f _ {i} ^ {I} W _ {1}\\right) \\cdot W _ {2}, \\tag {2}", + "image_path": "f692a8aa5953495ca804fab250121d8a8cfc94568dcf3d1c365612a0903e53cd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "inline_equation", + "content": "W_{1}" + }, + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "text", + "content": " are the parameters associated with the linear transformation layers, and " + }, + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "text", + "content": " is the non-linear activation function. The refined image feature " + }, + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\widetilde{f}_i^I" + }, + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "text", + "content": " can be computed with a residual connection:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 364, + 628, + 545, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 628, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 364, + 628, + 545, + 643 + ], + "type": "interline_equation", + "content": "\\tilde {f} _ {i} ^ {I} = \\alpha A _ {C} \\left(f _ {i} ^ {I}\\right) + (1 - \\alpha) f _ {i} ^ {I}, \\tag {3}", + "image_path": "9621a4204099c81163635517a6db68e17377fafec65f4954b14049e4f3309a86.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": " is a hyperparameter. Finally, CIA " + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "inline_equation", + "content": "A_{C}(\\cdot)" + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": " is optimized by minimizing the contrastive loss function (Equation 1), instantiated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 370, + 699, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 699, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 370, + 699, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e a l i g n}} = L _ {\\text {c o n t r a s t}} \\left(\\widetilde {f} ^ {I}, f ^ {T}\\right). \\tag {4}", + "image_path": "6a4c33569b25d3af7ef8a27bd1af7713041cf60abb3bb83e9dcb269f7bd29ead.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "text", + "content": "21415" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 79, + 192, + 310 + ], + "blocks": [ + { + "bbox": [ + 53, + 64, + 189, + 76 + ], + "lines": [ + { + "bbox": [ + 53, + 64, + 189, + 76 + ], + "spans": [ + { + "bbox": [ + 53, + 64, + 189, + 76 + ], + "type": "text", + "content": "Stage 1: Image-Text Re-Alignment" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 79, + 192, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 79, + 192, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 79, + 192, + 310 + ], + "type": "image", + "image_path": "60b0520307b5ea6bb9dc2699ea5f4f545b27bc6b2185bd26c63d18b5b0249748.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 316, + 547, + 395 + ], + "lines": [ + { + "bbox": [ + 46, + 316, + 547, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 547, + 395 + ], + "type": "text", + "content": "Figure 2. Overview of TAMM. Left: In Stage 1, TAMM fine-tunes a lightweight CLIP Image Adapter (CIA) through contrastive learning and re-aligns the image features with the text features to alleviate the domain shift originated from rendered images. Contrastive learning maximizes inner products between features from corresponding text-image pairs, and reduces similarities of mismatched pairs. Middle: In Stage 2, TAMM introduces Image Alignment Adapter (IAA) and Text Alignment Adapter (TAA) to decouple 3D representations into two sub-spaces: one focusing more on visual attributes and the other for semantic understanding, ensuring a more comprehensive and effective multi-modal pre-training strategy. Right: TAMM adaptively utilizes decoupled 3D features for various downstream tasks including linear probing classification (top) and zero-shot classification (bottom), achieving more robust classification results." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 197, + 80, + 545, + 305 + ], + "blocks": [ + { + "bbox": [ + 230, + 64, + 401, + 75 + ], + "lines": [ + { + "bbox": [ + 230, + 64, + 401, + 75 + ], + "spans": [ + { + "bbox": [ + 230, + 64, + 401, + 75 + ], + "type": "text", + "content": "Stage 2: Decoupled Tri-Modal Pre-Training" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 197, + 80, + 545, + 305 + ], + "lines": [ + { + "bbox": [ + 197, + 80, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 197, + 80, + 545, + 305 + ], + "type": "image", + "image_path": "c46b3fef7b9e42f88f2ce12b763efc58cbfc061cc4ede2b6e2bf236b8597ce8a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 408, + 234, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 408, + 234, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 408, + 234, + 422 + ], + "type": "text", + "content": "3.2. Decoupled Tri-Modal Pre-Training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 432, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 552 + ], + "type": "text", + "content": "Although the image and language feature spaces have been aligned by CLIP and our CLIP Image Adapter, they still encode subtly different information. For instance, the color of a 3D shape may be inferred from the image feature, but it cannot possibly be included in the text feature if the text description does not contain color information. Similarly, the image feature may not include semantic information such as the function or name of an object. In such cases, enforcing the 3D feature to align with both the image and text features simultaneously is challenging." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 556, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 288, + 715 + ], + "type": "text", + "content": "In order to overcome the obstacle introduced by aligning the 3D shape feature space with two distinct modalities, we propose a novel decoupled tri-modal pre-training framework, which aligns two decoupled 3D shape feature spaces with the refined image feature space and the text feature space, respectively (Figure 2-middle). We encourage the 3D encoder to cover both the visual representation and the semantic information inherent to the 3D shape, avoiding the dilemma of aligning with disparate image and text features at the same time. Formally, given the triplets " + }, + { + "bbox": [ + 46, + 556, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{(P_i,I_i,T_i)\\}_{i = 1}^n" + }, + { + "bbox": [ + 46, + 556, + 288, + 715 + ], + "type": "text", + "content": ", we first generate the image feature " + }, + { + "bbox": [ + 46, + 556, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\widetilde{f}_i^I" + }, + { + "bbox": [ + 46, + 556, + 288, + 715 + ], + "type": "text", + "content": " using the frozen, adapted 2D image encoder (Equation 3) and the text feature " + }, + { + "bbox": [ + 46, + 556, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{i}^{T}" + }, + { + "bbox": [ + 46, + 556, + 288, + 715 + ], + "type": "text", + "content": " using the original CLIP" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "text", + "content": "text encoder. We introduce a pair of lightweight Dual Adapters: Image Alignment Adapter (IAA) " + }, + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "inline_equation", + "content": "A_V(\\cdot)" + }, + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "text", + "content": " and Text Alignment Adapter (TAA) " + }, + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "inline_equation", + "content": "A_S(\\cdot)" + }, + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "text", + "content": ". They split the 3D feature " + }, + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "inline_equation", + "content": "f_i^P" + }, + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "text", + "content": " originated from the 3D encoder " + }, + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "inline_equation", + "content": "E_P" + }, + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "text", + "content": " into a vision-focusing feature " + }, + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "inline_equation", + "content": "f_i^{VP}" + }, + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "text", + "content": " and a semantic-focusing feature " + }, + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "inline_equation", + "content": "f_i^{SP}" + }, + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "text", + "content": ", respectively. Dual Adapters " + }, + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "inline_equation", + "content": "A_V(\\cdot)" + }, + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "inline_equation", + "content": "A_S(\\cdot)" + }, + { + "bbox": [ + 304, + 409, + 545, + 504 + ], + "type": "text", + "content": " share the same architecture as the CLIP Image Adapter in Section 3.1 and can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 347, + 507, + 545, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 507, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 347, + 507, + 545, + 529 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {V P} = A _ {V} \\left(f _ {i} ^ {P}\\right) = \\sigma \\left(f _ {i} ^ {P} W _ {1} ^ {V}\\right) \\cdot W _ {2} ^ {V}, \\tag {5}", + "image_path": "c4f9ff2618d3e9c14dc94c6e6a1e18128690878f29fe62691390d7d87c0a3751.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 354, + 524, + 504, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 524, + 504, + 539 + ], + "spans": [ + { + "bbox": [ + 354, + 524, + 504, + 539 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {S P} = A _ {S} (f _ {i} ^ {P}) = \\sigma (f _ {i} ^ {P} W _ {1} ^ {S}) \\cdot W _ {2} ^ {S},", + "image_path": "26bb1e9c9c92a07ddb61aa39fb5dfc87fcafdb842b80e071032532fbe7db0bb4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "inline_equation", + "content": "W_1^V, W_2^V, W_1^S," + }, + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "inline_equation", + "content": "W_2^S" + }, + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "text", + "content": " are the parameters associated with the linear layers, and " + }, + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "text", + "content": " represents the activation function. By decoupling 3D features into these sub-spaces, the 3D encoder " + }, + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "inline_equation", + "content": "E_P" + }, + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "text", + "content": " interprets the 3D shape with a more comprehensive visual and semantic understanding, improving the expressivity of the learned representations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 616, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 546, + 714 + ], + "type": "text", + "content": "Finally, instead of enforcing the 3D shape feature space to directly mimic the pre-aligned image-text feature space, we use the decoupled 3D features " + }, + { + "bbox": [ + 304, + 616, + 546, + 714 + ], + "type": "inline_equation", + "content": "f^{VP}, f^{SP}" + }, + { + "bbox": [ + 304, + 616, + 546, + 714 + ], + "type": "text", + "content": " and align them with the adapted image feature space " + }, + { + "bbox": [ + 304, + 616, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\widetilde{f}^T" + }, + { + "bbox": [ + 304, + 616, + 546, + 714 + ], + "type": "text", + "content": " and the text feature space " + }, + { + "bbox": [ + 304, + 616, + 546, + 714 + ], + "type": "inline_equation", + "content": "f^T" + }, + { + "bbox": [ + 304, + 616, + 546, + 714 + ], + "type": "text", + "content": ", respectively. Moreover, since one single image can only capture the 3D shape from one perspective, we align the 3D feature with the adapted features of multi-view images, to fully exploit the image modality and achieve a" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "text", + "content": "21416" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "better alignment between 3D and 2D. The overall loss function is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 99, + 287, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 99, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 52, + 99, + 287, + 125 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r i m o d a l}} = L _ {\\text {c o n t r a s t}} \\left(f ^ {S P}, f ^ {T}\\right) + \\frac {1}{m} \\sum_ {k} ^ {m} L _ {\\text {c o n t r a s t}} \\left(f ^ {V P}, \\widetilde {f} ^ {I, k}\\right), \\tag {6}", + "image_path": "2d1bf9a703309a89ea1eac7ba1a4b77c595645218e800ce09ba4a6e85216e00c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 131, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 131, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 131, + 287, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 131, + 287, + 156 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 131, + 287, + 156 + ], + "type": "text", + "content": " represents the number of rendered images and " + }, + { + "bbox": [ + 47, + 131, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\widetilde{f}^{I,k}" + }, + { + "bbox": [ + 47, + 131, + 287, + 156 + ], + "type": "text", + "content": " is the adapted image feature from the " + }, + { + "bbox": [ + 47, + 131, + 287, + 156 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 131, + 287, + 156 + ], + "type": "text", + "content": "-th view." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 156, + 288, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 288, + 348 + ], + "type": "text", + "content": "Application in Downstream Tasks. The learned 3D feature sub-spaces, " + }, + { + "bbox": [ + 47, + 156, + 288, + 348 + ], + "type": "inline_equation", + "content": "f^{VP}" + }, + { + "bbox": [ + 47, + 156, + 288, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 156, + 288, + 348 + ], + "type": "inline_equation", + "content": "f^{SP}" + }, + { + "bbox": [ + 47, + 156, + 288, + 348 + ], + "type": "text", + "content": " can be adaptively applied to a variety of downstream tasks (Figure 2-right). Specifically, in zero-shot 3D classification, we leverage both the 3D vision-focusing feature " + }, + { + "bbox": [ + 47, + 156, + 288, + 348 + ], + "type": "inline_equation", + "content": "f^{VP}" + }, + { + "bbox": [ + 47, + 156, + 288, + 348 + ], + "type": "text", + "content": " and the 3D semantic-focusing feature " + }, + { + "bbox": [ + 47, + 156, + 288, + 348 + ], + "type": "inline_equation", + "content": "f^{SP}" + }, + { + "bbox": [ + 47, + 156, + 288, + 348 + ], + "type": "text", + "content": ". We calculate the similarity between these features and category embeddings generated by the CLIP text encoder, respectively. After summing up the per-category similarity scores, the category with the highest similarity is chosen as the predicted class, yielding more robust and enhanced classification results compared with using a single sub-space alone. In the linear probing classification task, we concatenate the 3D vision-focusing feature " + }, + { + "bbox": [ + 47, + 156, + 288, + 348 + ], + "type": "inline_equation", + "content": "f^{VP}" + }, + { + "bbox": [ + 47, + 156, + 288, + 348 + ], + "type": "text", + "content": " and 3D semantic-focusing feature " + }, + { + "bbox": [ + 47, + 156, + 288, + 348 + ], + "type": "inline_equation", + "content": "f^{SP}" + }, + { + "bbox": [ + 47, + 156, + 288, + 348 + ], + "type": "text", + "content": " as input to the learnable linear classification layer, providing a more comprehensive and robust representation for 3D understanding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 358, + 128, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 128, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 128, + 372 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 378, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 378, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 378, + 287, + 521 + ], + "type": "text", + "content": "Pre-Training Datasets. Following the prior state-of-the-art method, OpenShape [23], our TAMM is pre-trained on the triplets generated from four datasets: ShapeNetCore [3], 3D-FUTURE [12], ABO [7], and Objaverse [9]. Our training sets are defined as follows: \"ShapeNet\" is a triplet set derived exclusively from the ShapeNetCore dataset, containing 52,470 3D shapes with corresponding images and text; \"Ensembled (no LVIS)\" is a set of 829,460 triplets from the above datasets excluding Objaverse-LVIS; \"Ensembled\" denotes the triplet set comprising data from all four datasets, containing 875,665 3D shapes and their associated images and text." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 522, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 288, + 713 + ], + "type": "text", + "content": "Evaluation Datasets. Our TAMM is evaluated on the following datasets: Objaverse-LVIS [9], ModelNet40 [43], ScanObjectNN [40], and ScanNet [8]. Objverse-LVIS encompasses a broad range of categories, featuring 46,832 high-quality shapes distributed across 1,156 LVIS [13] categories. ModelNet40 is a synthetic indoor 3D dataset comprising 40 categories. We use the test split of 2,468 shapes in our experiments. ScanObjectNN consists of scanned objects from 15 common categories. It has three main variants: OBJ-BG, OBJ-ONLY, and PB-T50-RS. ScanNet, characterized by its real-world scans, includes 1,513 indoor scenes containing 36,213 objects. We conduct experiments on four tasks including zero-shot 3D classification, linear probing 3D classification, few-shot linear probing 3D classification, and real-world recognition, to demonstrate the advantages of our TAMM. Other implementation details re" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "garding pre-training and evaluation are introduced in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 104, + 458, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 104, + 458, + 117 + ], + "spans": [ + { + "bbox": [ + 306, + 104, + 458, + 117 + ], + "type": "text", + "content": "4.1. Zero-Shot 3D Classification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 123, + 547, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 123, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 123, + 547, + 243 + ], + "type": "text", + "content": "Zero-shot ability is a key metric for reflecting the quality of learned 3D representations, which requires the representations to be directly applicable to datasets where the model has never been explicitly supervised. Without any further tuning, the 3D representations are compared with text embeddings of categories to predict the classes of 3D shapes. To make a fair comparison and keep consistency with prior work, we adopt the same settings as OpenShape [23] on three benchmarks: Objaverse-LVIS [9], ModelNet40 [43], and OBJ-ONLY (ScanObjectNN) [40]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 243, + 546, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 546, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 546, + 399 + ], + "type": "text", + "content": "The results are summarized in Table 1. First, we observe that TAMM, benefited from multi-modal pre-training, outperforms PointCLIP [51] and PointCLIP v2 [56] by a large margin. Furthermore, it is evident that our pretrained models consistently outperform those pre-trained by ULIP and OpenShape, irrespective of whether they are pretrained on smaller datasets like ShapeNet or more expansive ones such as the Ensembled dataset. For instance, PointBERT [47] pre-trained by our TAMM on the Ensembled (no LVIS) dataset surpasses ULIP and OpenShape by margins of " + }, + { + "bbox": [ + 304, + 243, + 546, + 399 + ], + "type": "inline_equation", + "content": "+20.6\\%" + }, + { + "bbox": [ + 304, + 243, + 546, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 243, + 546, + 399 + ], + "type": "inline_equation", + "content": "+2.9\\%" + }, + { + "bbox": [ + 304, + 243, + 546, + 399 + ], + "type": "text", + "content": " in Top-1 accuracy on the long-tailed Objaverse-LVIS benchmark, which validates the effectiveness of our multi-modal pre-training scheme." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 407, + 482, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 482, + 419 + ], + "type": "text", + "content": "4.2. Linear Probing 3D Classification" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 426, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 665 + ], + "type": "text", + "content": "To further evaluate the 3D representation quality of our pre-trained models, we design and conduct linear probing classification experiments. Here, we first freeze our pretrained Point-BERT model along with the Image Alignment Adapter (IAA) and Text Alignment Adapter (TAA), and then append a single learnable linear layer to the model. Given a batch of point clouds, as illustrated in Figure 2, we generate features from both IAA and TAA and concatenate them, and learn the appended linear classification layer on the concatenated features. We evaluate the accuracy of the linear classifier on three benchmarks: Objverse-LVIS [9], ModelNet40 [43], and ScanObjectNN [40]. Objverse-LVIS is a challenging long-tailed dataset with 1,156 categories, which has not been evaluated by prior methods [23, 45]. To ensure evaluation with statistically meaningful number of samples per class, we exclude categories with fewer than 10 instances, leaving us 1,046 categories. Samples in each class are divided into training and testing sets at an " + }, + { + "bbox": [ + 304, + 426, + 545, + 665 + ], + "type": "inline_equation", + "content": "8:2" + }, + { + "bbox": [ + 304, + 426, + 545, + 665 + ], + "type": "text", + "content": " ratio. For ModelNet40 and ScanObjectNN, we use the standard splits, following [32, 42]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "The results are shown in Table 2. TAMM consistently outperforms all previous methods by a large margin. For example, pre-trained on the Ensembled dataset, TAMM improves over OpenShape by " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "+11.2\\%" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "+2.2\\%" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " overall ac" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 318, + 764 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 318, + 764 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 318, + 764 + ], + "type": "text", + "content": "21417" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 59, + 534, + 315 + ], + "blocks": [ + { + "bbox": [ + 62, + 59, + 534, + 315 + ], + "lines": [ + { + "bbox": [ + 62, + 59, + 534, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 59, + 534, + 315 + ], + "type": "table", + "html": "
Pre-Training DatasetModelPre-Training MethodObjaverse-LVIS [9]ModelNet40 [43]ScanObjectNN [40]
Top-1Top-3Top-5Top-1Top-3Top-5Top-1Top-3Top-5
-PointCLIP [51]-1.94.15.819.328.634.810.520.830.6
-PointCLIP v2 [56]-4.79.512.963.677.985.042.263.374.5
ShapeNetViT-B/32 [11]CLIP2Point [19]2.75.87.949.571.381.225.544.659.4
Transformer [41]ReCon [32]1.12.73.761.273.978.142.362.575.6
SparseConv [6]OpenShape [23]11.621.827.172.987.293.052.772.783.6
TAMM (Ours)13.624.229.374.688.294.057.975.383.1
Point-BERT [47]ULIP [45]6.213.617.960.479.084.451.571.180.2
OpenShape [23]10.820.225.070.386.991.351.369.478.4
TAMM (Ours)13.724.229.273.188.591.954.874.583.3
Ensembled (no LVIS)SparseConv [6]OpenShape [23]37.058.466.982.695.097.554.976.887.0
TAMM (Ours)39.862.070.485.796.898.357.581.390.0
Point-BERT [47]ULIP [45]21.438.146.071.484.489.246.066.176.4
OpenShape [23]39.160.868.985.396.297.447.272.484.7
TAMM (Ours)42.063.671.786.396.698.156.778.386.1
EnsembledSparseConv [6]OpenShape [23]43.464.872.483.495.697.856.778.988.6
TAMM (Ours)43.866.274.185.496.498.158.581.389.5
Point-BERT [47]ULIP [45]26.844.852.675.188.193.251.672.582.3
OpenShape [23]46.869.177.084.496.598.052.279.788.7
TAMM (Ours)50.773.280.685.096.698.155.780.788.9
", + "image_path": "292f3cd4db83329b3f2564ec898f39cb2603b94bfd9fa6141352cbb8a8af60ac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 369, + 288, + 469 + ], + "blocks": [ + { + "bbox": [ + 46, + 322, + 547, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 322, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 547, + 357 + ], + "type": "text", + "content": "Table 1. Zero-shot 3D classification results. TAMM sets new state of the art in zero-shot classification accuracy across Objaverse-LVIS, ModelNet-40, and ScanObjectNN benchmarks, outperforming existing methods in diverse settings of pre-training datasets and 3D model architectures. The performance gain is more significant on the most challenging long-tailed Objverse-LVIS dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 369, + 288, + 469 + ], + "lines": [ + { + "bbox": [ + 48, + 369, + 288, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 369, + 288, + 469 + ], + "type": "table", + "html": "
Pre-Training DatasetMethodO-LVIS [9]M-40 [43]ScanObjectNN [40]
OBJ-BGOBJ-ONLYPB-T50-RS
ShapeNetULIP [45]34.690.675.475.464.8
OpenShape [23]29.388.577.878.564.1
TAMM (Ours)39.191.080.681.168.5
Rel. Improv.+9.8+2.5+2.8+2.6+4.4
EnsembledOpenShape [23]48.391.385.985.478.0
TAMM (Ours)59.593.588.588.080.3
Rel. Improv.+11.2+2.2+2.6+2.6+1.7
", + "image_path": "1367565047a26aa1a5cf5a5c7e70cdfbb51a58040b66d26422b7e8360d4a4da8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 527, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 527, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 45, + 527, + 287, + 648 + ], + "type": "text", + "content": "curacy on Objaverse-LVIS and ModelNet40, respectively. This demonstrates that TAMM effectively exploits knowledge from CLIP and learns generalizable 3D representations that are directly applicable to novel linear classification tasks. Notably, TAMM even surpasses the Point-BERT model pre-trained without multi-modality but allowed to be fully fine-tuned on OBJ-BG (87.4%) [47] by +1.1% accuracy. These strong results indicate that TAMM has learned excellent transferable 3D representations, showing a great potential in real-world applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 658, + 272, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 658, + 272, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 658, + 272, + 671 + ], + "type": "text", + "content": "4.3. Few-Shot Linear Probing 3D Classification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Following the previous evaluation, we perform few-shot classification experiments on ModelNet40 [43] to assess TAMM in low-data scenarios. Similar to the linear probing" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 306, + 369, + 545, + 448 + ], + "blocks": [ + { + "bbox": [ + 46, + 476, + 287, + 510 + ], + "lines": [ + { + "bbox": [ + 46, + 476, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 287, + 510 + ], + "type": "text", + "content": "Table 2. Linear probing 3D classification results. TAMM outperforms previous methods by a large margin, e.g., +11.2% accuracy gain on the challenging Objaverse-LVIS dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 369, + 545, + 448 + ], + "lines": [ + { + "bbox": [ + 306, + 369, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 545, + 448 + ], + "type": "table", + "html": "
Pre-Training DatasetMethod5-way10-way
10-shot20-shot10-shot20-shot
ShapeNetULIP [45]94.4±3.793.2±4.286.6±5.390.6±5.2
OpenShape [23]95.3±2.697.9±3.989.2±5.192.9±3.9
TAMM (Ours)97.8±1.998.1±1.495.3±3.995.9±2.6
EnsembledOpenShape [23]96.1±2.795.7±2.589.1±4.691.8±3.7
TAMM (Ours)99.0±1.399.4±0.796.8±2.997.4±2.2
", + "image_path": "5ab40ebaeae016bbba2dbc9e0eb4e9996c94c192d75c9937aa691ba8edd12048.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 455, + 547, + 510 + ], + "lines": [ + { + "bbox": [ + 304, + 455, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 547, + 510 + ], + "type": "text", + "content": "Table 3. Few-shot linear probing classification results on ModelNet40. We report the average accuracy and standard deviation of 10 independent experiments. Our TAMM consistently achieves both the best average accuracy and the lowest variance in various few-shot settings." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 531, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 686 + ], + "type": "text", + "content": "experiments described in Section 4.2, we extend our model with an additional linear classification layer and only train this linear layer instead of fine-tuning the entire model. Following previous work [47, 50], we adopt the \" " + }, + { + "bbox": [ + 304, + 531, + 545, + 686 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 531, + 545, + 686 + ], + "type": "text", + "content": "-way " + }, + { + "bbox": [ + 304, + 531, + 545, + 686 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 531, + 545, + 686 + ], + "type": "text", + "content": "-shot\" configuration, wherein we select " + }, + { + "bbox": [ + 304, + 531, + 545, + 686 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 531, + 545, + 686 + ], + "type": "text", + "content": " classes at random and sample " + }, + { + "bbox": [ + 304, + 531, + 545, + 686 + ], + "type": "inline_equation", + "content": "(N + 20)" + }, + { + "bbox": [ + 304, + 531, + 545, + 686 + ], + "type": "text", + "content": " instances per class. Training is conducted on a support set of " + }, + { + "bbox": [ + 304, + 531, + 545, + 686 + ], + "type": "inline_equation", + "content": "K \\times N" + }, + { + "bbox": [ + 304, + 531, + 545, + 686 + ], + "type": "text", + "content": " samples, while evaluation is based on a query set comprised of the remaining 20 instances per class. We assess our model under four distinct scenarios: \"5-way 10-shot,\" \"5-way 20-shot,\" \"10-way 10-shot,\" and \"10-way 20-shot.\" For each scenario, we carry out 10 separate trials, and report both the mean performance and the standard deviation across these trials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": "As illustrated in Table 3, our TAMM achieves the best results and sets a new state of the art in the few-shot clas-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 318, + 764 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 318, + 764 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 318, + 764 + ], + "type": "text", + "content": "21418" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 59, + 545, + 148 + ], + "blocks": [ + { + "bbox": [ + 50, + 59, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 50, + 59, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 50, + 59, + 545, + 148 + ], + "type": "table", + "html": "
MethodAvg.BedCabChairSofaTablDoorWindBksfPicCntrDeskCurtFridgBathShowrToilSink
CLIP2Point [19]24.920.80.085.143.326.569.90.020.91.731.727.00.01.646.50.022.425.6
PointCLIP w/ TP. [56]26.10.055.772.85.05.11.70.077.20.00.051.70.30.00.040.385.349.2
CLIP2Point w/ TP. [19]35.211.83.045.127.610.561.52.671.90.333.629.94.711.572.292.486.134.0
\\( CLIP^2 \\) [49]38.532.667.269.342.318.319.14.062.61.412.752.840.19.159.741.071.045.5
\\( OpenShape^\\dagger \\) [23]45.666.73.275.883.537.749.247.564.948.21.966.170.21.850.057.145.27.1
TAMM (Ours)†49.466.74.883.684.548.957.948.280.561.31.960.683.67.041.456.148.43.6
", + "image_path": "8e20e87433acfc8ebe88008e98734b11a3ba3e9040ded5d0d9222a45c1d42c06.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 148, + 545, + 159 + ], + "lines": [ + { + "bbox": [ + 47, + 148, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 47, + 148, + 545, + 159 + ], + "type": "text", + "content": "w/ TP. denotes training with the real-world data provided by " + }, + { + "bbox": [ + 47, + 148, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^2}" + }, + { + "bbox": [ + 47, + 148, + 545, + 159 + ], + "type": "text", + "content": ". Results using Point-BERT [47] as 3D encoder, pre-trained on the Ensembled dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 78, + 192, + 279, + 260 + ], + "blocks": [ + { + "bbox": [ + 47, + 164, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 47, + 164, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 164, + 545, + 186 + ], + "type": "text", + "content": "Table 4. Zero-shot classification results on the real-world ScanNet dataset. Avg.: the mean average Top-1 accuracy of all classes. TAMM achieves the best results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 192, + 279, + 260 + ], + "lines": [ + { + "bbox": [ + 78, + 192, + 279, + 260 + ], + "spans": [ + { + "bbox": [ + 78, + 192, + 279, + 260 + ], + "type": "table", + "html": "
CIAIAATAAImg-Txt AccObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
✓✓40.113.523.929.172.888.291.7
✓60.912.522.427.371.486.089.8
✓✓60.912.922.927.773.888.292.5
✓✓60.913.023.028.272.987.890.4
✓✓✓60.913.724.229.273.188.591.9
", + "image_path": "0943b63de7bc751a234e24160e35b69d6b0a97dfef2749d95f85dee59b1af474.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 78, + 262, + 279, + 290 + ], + "lines": [ + { + "bbox": [ + 78, + 262, + 279, + 290 + ], + "spans": [ + { + "bbox": [ + 78, + 262, + 279, + 290 + ], + "type": "text", + "content": "(a) Adapters in pre-training. Pre-training with CLIP Image Adapter (CIA), Image Alignment Adapter (IAA), and Text Alignment Adapter (TAA) is the most effective." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 323, + 205, + 515, + 246 + ], + "blocks": [ + { + "bbox": [ + 323, + 205, + 515, + 246 + ], + "lines": [ + { + "bbox": [ + 323, + 205, + 515, + 246 + ], + "spans": [ + { + "bbox": [ + 323, + 205, + 515, + 246 + ], + "type": "table", + "html": "
StageImg-Txt AccObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
One stage55.312.221.427.071.486.490.7
Two stages60.913.724.229.273.188.591.9
", + "image_path": "fb4ff0526aaa532992d526cfcdd83524c0072d1b9b4658bca672be83d635444b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 323, + 248, + 515, + 277 + ], + "lines": [ + { + "bbox": [ + 323, + 248, + 515, + 277 + ], + "spans": [ + { + "bbox": [ + 323, + 248, + 515, + 277 + ], + "type": "text", + "content": "(b) Pre-training stages. Learning the CLIP Image Adapter first and then pre-train the 3D encoder with Dual Adapters is better than training modules all together." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 47, + 297, + 209, + 347 + ], + "blocks": [ + { + "bbox": [ + 47, + 297, + 209, + 347 + ], + "lines": [ + { + "bbox": [ + 47, + 297, + 209, + 347 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 209, + 347 + ], + "type": "table", + "html": "
ImageTextObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
✓11.920.825.967.687.392.5
✓10.519.223.770.285.088.0
✓✓13.724.229.273.188.591.9
", + "image_path": "81964dee1778efb4ba69942001c0251c3e81f4672b4a743df556a6647d90c65e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 350, + 208, + 378 + ], + "lines": [ + { + "bbox": [ + 47, + 350, + 208, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 208, + 378 + ], + "type": "text", + "content": "(c) Pre-training modalities. Exploiting both image and text data provides the most performance gain. They contribute almost equally in TAMM." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 220, + 297, + 383, + 348 + ], + "blocks": [ + { + "bbox": [ + 220, + 297, + 383, + 348 + ], + "lines": [ + { + "bbox": [ + 220, + 297, + 383, + 348 + ], + "spans": [ + { + "bbox": [ + 220, + 297, + 383, + 348 + ], + "type": "table", + "html": "
IAATAAObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
✓13.023.128.368.186.591.4
✓12.922.527.572.486.290.4
✓✓13.724.229.273.188.591.9
", + "image_path": "ac7d97b09543ac02fa5b609b5a07280f0f0aa3e5c005fdc680c44bbacc25ebf0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 220, + 350, + 382, + 378 + ], + "lines": [ + { + "bbox": [ + 220, + 350, + 382, + 378 + ], + "spans": [ + { + "bbox": [ + 220, + 350, + 382, + 378 + ], + "type": "text", + "content": "(d) Adapters for inference. Dual Adapters are complementary, and their combination more accurately classifies 3D shapes at test time." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 394, + 297, + 545, + 357 + ], + "blocks": [ + { + "bbox": [ + 394, + 297, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 394, + 297, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 394, + 297, + 545, + 357 + ], + "type": "table", + "html": "
ImagesObjaverse-LVISModelNet-40
Top-1Top-3Top-5Top-1Top-3Top-5
113.623.628.972.588.092.6
213.523.528.873.187.692.8
413.724.229.273.188.591.9
813.523.128.473.089.092.8
", + "image_path": "4edafad04bc06e908738e340fc13fd7a8de76663db70f89feb4f6d1a99d46ace.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 394, + 359, + 545, + 378 + ], + "lines": [ + { + "bbox": [ + 394, + 359, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 394, + 359, + 545, + 378 + ], + "type": "text", + "content": "(e) Number of images. Aligning a 3D shape with 4 images of it is the best." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 388, + 546, + 421 + ], + "lines": [ + { + "bbox": [ + 46, + 388, + 546, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 546, + 421 + ], + "type": "text", + "content": "Table 5. Ablation study of components in TAMM. We pre-train Point-BERT models on ShapeNet in various settings and evaluate their zero-shot classification performance on Objaverse-LVIS and ModelNet40. The baseline method OpenShape [23] achieves " + }, + { + "bbox": [ + 46, + 388, + 546, + 421 + ], + "type": "inline_equation", + "content": "10.8\\%" + }, + { + "bbox": [ + 46, + 388, + 546, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 388, + 546, + 421 + ], + "type": "inline_equation", + "content": "70.3\\%" + }, + { + "bbox": [ + 46, + 388, + 546, + 421 + ], + "type": "text", + "content": " Top-1 accuracy on these two benchmarks (Table 1), respectively. The setting adopted by TAMM is marked." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "content": "sification task. TAMM shows both higher overall accuracy and smaller deviations than other methods, showcasing the generalizability and robustness of TAMM-learned 3D representations. For example, when pre-trained on the Ensembled dataset, our TAMM surpasses OpenShape [23] by " + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "inline_equation", + "content": "7.7\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "content": ", respectively for all four settings. These results indicate that our TAMM is able to learn 3D representations that are more generalizable and can be readily adapted to downstream tasks under low-data regimes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 552, + 183, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 183, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 183, + 564 + ], + "type": "text", + "content": "4.4. Real-World Recognition" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "To evaluate the capability of TAMM in understanding " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " shapes from the real world, we follow the previous work " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " [49] and test TAMM on a real-world recognition task, in which the model aims at correctly classifying each instance from a complex scene in a zero-shot manner. Specifically, we select the real-world scene-level dataset ScanNet [8] and adopt same data splits as [49], containing 17 classes. We perform zero-shot classification (Section 4.1) on the point cloud of each object instance extracted from scenes, and report both the Top-1 accuracy of each class and the mean accuracy of all 17 classes. As shown in Table 4, TAMM significantly outperforms all prior meth" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 437, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 520 + ], + "type": "text", + "content": "ods, showing improvements of " + }, + { + "bbox": [ + 304, + 437, + 545, + 520 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 304, + 437, + 545, + 520 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 437, + 545, + 520 + ], + "type": "inline_equation", + "content": "10.9\\%" + }, + { + "bbox": [ + 304, + 437, + 545, + 520 + ], + "type": "text", + "content": " in average Top-1 accuracy compared with OpenShape and " + }, + { + "bbox": [ + 304, + 437, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP^2}" + }, + { + "bbox": [ + 304, + 437, + 545, + 520 + ], + "type": "text", + "content": ", respectively. These results underscore TAMM's ability in recognizing and understanding 3D shapes captured from real-world scenarios. Further results on complex scene recognition and instance segmentation are included in the supplementary material." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 527, + 400, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 400, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 400, + 540 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "type": "text", + "content": "In this section, we provide additional experimental results to further test the performance gain from each design of TAMM. Due to limited computation, we pre-train PointBERT [47] on the small-scale ShapeNet [3], and use two more challenging zero-shot classification benchmarks ModelNet40 [43] and Objaverse-LVIS [9] for evaluation. We alter only one component in each set of experiments." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Adapters in Pre-Training. As described in Section 3.1, our CLIP Image Adapter (CIA) mitigates the domain gap between rendered images and natural images. We first measure the contrastive accuracy, computed as the ratio of image-text pairs where the text feature is more similar to the image from the same triplet than any other text. The contrastive accuracy of CLIP without CIA is only " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "40.1\\%" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "text", + "content": "21419" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "indicating that the original image-text feature space is compromised by the domain gap. CIA brings the accuracy up to " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "60.9\\%" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": ", showing the effectiveness of our CIA in resolving the shifted data domain. Furthermore, as shown in Table 5a, CIA increases the zero-shot accuracy on Objaverse-LVIS dataset from " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "12.2\\%" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "13.7\\%" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": ", validating its effectiveness on improving the pre-trained 3D representations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 288, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 288, + 266 + ], + "type": "text", + "content": "TAMM adopts Image Alignment Adapter (IAA) and Text Alignment Adapter (TAA) to decouple the feature space and enrich the learned 3D representations. To further explore the performance gains introduced by our Dual Adapters (IAA and TAA), we experiment tri-modal pretraining with or without IAA and TAA. As shown in Table 5a, both adapters bring performance gains, and their combination achieves the best result, demonstrating the effectiveness of Dual Adapters." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 266, + 288, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 288, + 363 + ], + "type": "text", + "content": "Pre-Training Stages. We investigate the significance of our two-stage pre-training design, which initially learns CLIP Image Adapter (CIA) to re-align image-text pairs and subsequently employs Dual Adapters in decoupled tri-modal pre-training. An alternative could be learning these modules all together in one stage. As shown in Table 5b, remarkably, the two-stage pre-training approach achieves a better performance compared with one-stage pre-training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 364, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 287, + 496 + ], + "type": "text", + "content": "Pre-Training Modalities. We further explore the effectiveness of image and language modalities in pre-training. Table 5c shows that integrating the 3D modality with both the image and language modalities consistently yields superior performance across various benchmarks, compared with aligning it with either modality alone. Moreover, unlike prior methods which ineffectively learn from images (Figure 1-left), the image and language modalities contribute almost equally in TAMM, which also signifies that TAMM better exploits the image modality and brings more gains by learning from both image and text data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 497, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 616 + ], + "type": "text", + "content": "Adapters for Inference. Our IAA and TAA decouple 3D features with visual and semantic focuses, respectively, offering a more accurate and comprehensive understanding for 3D shapes. At inference time, we combine the outputs from Dual Adapters for classification tasks. To investigate this design, we separately utilize features produced solely by either IAA or TAA in zero-shot classification evaluation. As shown in Table 5d, the combined features achieve more accurate classification, demonstrating that the learned features from IAA and TAA are complementary to each other." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "Number of Images. In order to gain more comprehensive knowledge from the image modality, TAMM simultaneously aligns the 3D feature with multi-view 2D images, each projected from a different perspective. We evaluate this design by aligning the 3D feature with a varying number of corresponding 2D images, ranging from 1 to 8 images. As shown in Table 5e, multi-view image features offer a performance gain by fully exploiting the image modality." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 63, + 547, + 245 + ], + "blocks": [ + { + "bbox": [ + 310, + 63, + 547, + 245 + ], + "lines": [ + { + "bbox": [ + 310, + 63, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 310, + 63, + 547, + 245 + ], + "type": "image", + "image_path": "fd8efb0de0c4c7191a532931e2cf86427657a9ef603bd873c0deaaf26c026870.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 254, + 547, + 344 + ], + "lines": [ + { + "bbox": [ + 304, + 254, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 547, + 344 + ], + "type": "text", + "content": "Figure 3. Qualitative results. Top: CIA re-aligns the images rendered from 3D shapes with the text descriptions. The rendered images are inaccurately matched with text when the image features are directly extracted by CLIP, and CIA can correct the matching. Bottom: IAA and TAA decouple 3D features with complementary visual and semantic focuses. Features from one single adapter are matched with classes whose appearance or semantics resemble the true class; using both adapters leads to the correct class." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 358, + 543, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 543, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 543, + 370 + ], + "type": "text", + "content": "Aligning a 3D shape with 4 views achieves the best result." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 370, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 545, + 491 + ], + "type": "text", + "content": "Qualitative Results. Finally, we provide some visualizations to intuitively demonstrate the benefit of our proposed TAMM. As illustrated in Figure 3, CIA successfully mitigates the domain gap introduced by rendered images, and leads to the more accurate matching between images and text. IAA and TAA learn 3D representations with subtly different but complementary focuses on vision and semantics, respectively, and their combination brings more robust and comprehensive 3D representations. More visualized results are included in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 502, + 379, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 379, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 379, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "content": "In this work, we examine the sub-optimal utilization of 2D images in existing multi-modal pre-training methods for 3D shape understanding, and propose TriAdapter MultiModal Learning (TAMM), a novel two-stage representation learning approach built on three synergistic adapter modules. Extensive experiments verify that TAMM consistently learns improved 3D features in various settings." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 606, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 666 + ], + "type": "text", + "content": "Limitations. Due to the limited computation resources, we are not able to perform pre-training on very large-scale 3D backbones with billions of parameters. The quality of the learned 3D representation could be further improved if the 3D backbones are scaled to a larger size." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Jump ARCHES endowment, and the IBM-Illinois Discovery Accelerator Institute. This work used NVIDIA GPUs at NCSA Delta through allocations CIS220014, CIS230012, and CIS230013 from the ACCESS program." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 318, + 764 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 318, + 764 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 318, + 764 + ], + "type": "text", + "content": "21420" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 200 + ], + "type": "text", + "content": "[1] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 201, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 201, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 201, + 287, + 245 + ], + "type": "text", + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 246, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 288, + 312 + ], + "type": "text", + "content": "[3] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 2, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 312, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 312, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 312, + 288, + 346 + ], + "type": "text", + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 346, + 287, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 400 + ], + "type": "text", + "content": "[5] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "type": "text", + "content": "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D spatio-temporal ConvNets: Minkowski convolutional neural networks. In CVPR, 2019. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 436, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 501 + ], + "type": "text", + "content": "[7] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F. Yago Vicente, Thomas Dideriksien, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. ABO: Dataset and benchmarks for real-world 3D object understanding. In CVPR, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 502, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 287, + 546 + ], + "type": "text", + "content": "[8] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 547, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 547, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 547, + 287, + 601 + ], + "type": "text", + "content": "[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[12] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3D-FUTURE: 3D furniture shape with texture. IJCV, 129:3313-3337, 2021. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 106, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 139 + ], + "type": "text", + "content": "[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "type": "text", + "content": "[14] Huy Ha and Shuran Song. Semantic abstraction: Openworld 3D scene understanding from 2D vision-language models. In CoRL, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "type": "text", + "content": "[15] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 206, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 239 + ], + "type": "text", + "content": "[16] Deepti Hegde, Jeya Maria Jose Valanarasu, and Vishal Patel. CLIP goes 3D: Leveraging prompt tuning for language grounded 3D recognition. In ICCV, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 282 + ], + "type": "text", + "content": "[17] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. AvatarCLIP: Zero-shot text-driven generation and animation of 3D avatars. ACM Transactions on Graphics, 41(4):1-19, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 284, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 284, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 284, + 545, + 327 + ], + "type": "text", + "content": "[18] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. In ICML, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "type": "text", + "content": "[19] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. CLIP2Point: Transfer CLIP to point cloud classification with image-depth pre-training. In ICCV, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "type": "text", + "content": "[20] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 415, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 459 + ], + "type": "text", + "content": "[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 459, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 545, + 514 + ], + "type": "text", + "content": "[22] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 514, + 545, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 545, + 559 + ], + "type": "text", + "content": "[23] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. OpenShape: Scaling up 3D shape representation towards open-world understanding. In NeurIPS, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 559, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 592 + ], + "type": "text", + "content": "[24] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In ICLR, 2022. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 593, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 624 + ], + "type": "text", + "content": "[25] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2Mesh: Text-driven neural stylization for meshes. In CVPR, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "text", + "content": "[26] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In CVPR, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[27] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. CLIP-Mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia, 2022. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 754, + 317, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 754, + 317, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 754, + 317, + 763 + ], + "type": "text", + "content": "21421" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[28] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "text", + "content": "[29] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. OpenScene: 3D scene understanding with open vocabularies. In CVPR, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "text", + "content": "[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In CVPR, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "type": "text", + "content": "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. PointNet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "text", + "content": "[32] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3D representation learning guided by generative pretraining. In ICML, 2023. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[33] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. PointNeXt: Revisiting PointNet++ with improved training and scaling strategies. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, A. Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 392 + ], + "type": "text", + "content": "[35] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3D semantic segmentation in the wild. In ECCV, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 393, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 480 + ], + "type": "text", + "content": "[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: An open large-scale dataset for training next generation image-text models. In NeurIPS Datasets and Benchmarks Track, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "text", + "content": "[37] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. PV-RCNN: Point-voxel feature set abstraction for 3D object detection. In CVPR, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "type": "text", + "content": "[38] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. MotionCLIP: Exposing human motion generation to CLIP space. In ECCV, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 624 + ], + "type": "text", + "content": "[39] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "text", + "content": "[40] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 115 + ], + "type": "text", + "content": "[42] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point Transformer V2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[43] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In CVPR, 2015. 2, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 162, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 206 + ], + "type": "text", + "content": "[44] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3D: Zero-shot text-to-3D synthesis using 3D shape prior and text-to-image diffusion models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 207, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 261 + ], + "type": "text", + "content": "[45] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. ULIP: Learning a unified representation of language, images, and point clouds for 3D understanding. In CVPR, 2023. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 262, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 545, + 307 + ], + "type": "text", + "content": "[46] Jihan Yang, Runyu Ding, Zhe Wang, and Xiaojuan Qi. RegionPLC: Regional point-language contrastive learning for open-world 3D scene understanding. arXiv preprint arXiv:2304.00962, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "text", + "content": "[47] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-BERT: Pre-training 3D point cloud transformers with masked point modeling. In CVPR, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "type": "text", + "content": "[48] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In CVPR, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "type": "text", + "content": "[49] Yihan Zeng, Chenhan Jiang, Jiageng Mao, Jianhua Han, Chaoqiang Ye, Qingqiu Huang, Dit-Yan Yeung, Zhen Yang, Xiaodan Liang, and Hang Xu. CLIP" + }, + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "type": "text", + "content": ": Contrastive language-image-point pretraining from real-world point cloud data. In CVPR, 2023. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 443, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 443, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 443, + 545, + 487 + ], + "type": "text", + "content": "[50] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. PointM2AE: Multi-scale masked autoencoders for hierarchical point cloud pre-training. In NeurIPS, 2022. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 488, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 532 + ], + "type": "text", + "content": "[51] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. PointCLIP: Point cloud understanding by CLIP. In CVPR, 2022. 1, 3, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "type": "text", + "content": "[52] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3D features on any point-cloud. In ICCV, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 567, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 620 + ], + "type": "text", + "content": "[53] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. RegionCLIP: Region-based language-image pretraining. In CVPR, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 622, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 622, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 622, + 545, + 656 + ], + "type": "text", + "content": "[54] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3D: Exploring unified 3D representation at scale. In ICLR, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 657, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 657, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 657, + 545, + 690 + ], + "type": "text", + "content": "[55] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "text", + "content": "[56] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Point-" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 754, + 318, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 754, + 318, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 754, + 318, + 763 + ], + "type": "text", + "content": "21422" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "CLIP V2: Prompting CLIP and GPT for powerful 3D open-world learning. In ICCV, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 754, + 317, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 754, + 317, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 754, + 317, + 763 + ], + "type": "text", + "content": "21423" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/745b72e4-2dd6-4811-ade0-f52413f14141_content_list.json b/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/745b72e4-2dd6-4811-ade0-f52413f14141_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5134d305bd0b6a510da6f6eb6e3189ae64834f8f --- /dev/null +++ b/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/745b72e4-2dd6-4811-ade0-f52413f14141_content_list.json @@ -0,0 +1,1447 @@ +[ + { + "type": "text", + "text": "TASeg: Temporal Aggregation Network for LiDAR Semantic Segmentation", + "text_level": 1, + "bbox": [ + 106, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaopei Wu $^{1,2}$ , Yuenan Hou $^{2*}$ , Xiaoshui Huang $^{2*}$ , Binbin Lin $^{3,4}$ , Tong He $^{2}$ , Xinge Zhu $^{5}$ , Yuexin Ma $^{6}$ , Boxi Wu $^{4}$ , Haifeng Liu $^{1*}$ , Deng Cai $^{1}$ , Wanli Ouyang $^{2}$", + "bbox": [ + 174, + 179, + 792, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Lab of CAD&CG, Zhejiang University $^{2}$ Shanghai AI Laboratory $^{3}$ Fullong Inc.", + "bbox": [ + 200, + 217, + 764, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ School of Software Technology, Zhejiang University $^{5}$ The Chinese University of Hong Kong $^{6}$ ShanghaiTech University", + "bbox": [ + 109, + 234, + 852, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training deep models for LiDAR semantic segmentation is challenging due to the inherent sparsity of point clouds. Utilizing temporal data is a natural remedy against the sparsity problem as it makes the input signal denser. However, previous multi-frame fusion algorithms fall short in utilizing sufficient temporal information due to the memory constraint, and they also ignore the informative temporal images. To fully exploit rich information hidden in long-term temporal point clouds and images, we present the Temporal Aggregation Network, termed TAseg. Specifically, we propose a Temporal LiDAR Aggregation and Distillation (TLAD) algorithm, which leverages historical priors to assign different aggregation steps for different classes. It can largely reduce memory and time overhead while achieving higher accuracy. Besides, TLAD trains a teacher injected with gt priors to distill the model, further boosting the performance. To make full use of temporal images, we design a Temporal Image Aggregation and Fusion (TIAF) module, which can greatly expand the camera FOV and enhance the present features. Temporal LiDAR points in the camera FOV are used as mediums to transform temporal image features to the present coordinate for temporal multi-modal fusion. Moreover, we develop a Static-Moving Switch Augmentation (SMSA) algorithm, which utilizes sufficient temporal information to enable objects to switch their motion states freely, thus greatly increasing static and moving training samples. Our TAseg ranks $\\mathbf{1}^{st\\dagger}$ on three challenging tracks, i.e., SemanticKITTI single-scan track, multi-scan track and nuScenes LiDAR segmentation track, strongly demonstrating the superiority of our method. Codes are available at https://github.com/LittlePey/TASeg.", + "bbox": [ + 75, + 316, + 473, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 810, + 209, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LiDAR segmentation aims to infer the semantic information of each point in point clouds and plays an indispensable role", + "bbox": [ + 75, + 835, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in the autonomous driving [3, 21, 34, 42]. With the advent of deep learning, a large quantity of LiDAR segmentation algorithms [14, 19, 24, 39, 46] have been put forward. Despite their impressive results, the segmentation performance is still constrained by the inherent sparsity of point clouds. To alleviate this, it is desirable to aggregate temporal data. Previous multi-frame models [4, 9, 22] can only fuse a few LiDAR frames due to the GPU memory constraint, which restricts them from utilizing rich information hidden in the long-term temporal point clouds. Although MSeg3D [20] stacks dozens of LiDAR frames, the memory and time overhead are intolerable. Moreover, previous attempts concentrate on utilizing temporal point clouds, ignoring the valuable information hidden in temporal images.", + "bbox": [ + 500, + 287, + 893, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To better leverage long-term temporal information, we propose a Temporal LiDAR Aggregation and Distillation (TLAD) algorithm. It can perform efficient multi-frame aggregation while achieving higher accuracy. TLAD consists of Flexible Step Aggregation (FSA) and Mask Distillation. FSA is based on the observation that the model needs different amounts of temporal points for different classes. For difficult classes, such as bicyclists, more points are needed to yield accurate predictions. However, for easy classes, such as cars and roads, it is unnecessary to aggregate too many point cloud frames for them. Therefore, we propose to assign different aggregation steps for different classes in point cloud sequences, which can significantly save memory and computation overhead on easy classes while providing sufficient temporal points for difficult classes. To identify classes of temporal point clouds, historical predictions are used, considering the temporal data is processed in chronological order, which is consistent with practical applications.", + "bbox": [ + 496, + 503, + 895, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Interestingly, by assigning different classes with different steps, FSA actually injects historical priors into the aggregated point clouds. For classes with different steps, their patterns can be more discriminative due to their different densities. This makes the point cloud frames accumulated by FSA more conducive for segmentation. Experimental results show that FSA can not only reduce memory and time costs but also enhance overall performance. To further verify our", + "bbox": [ + 496, + 779, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding authors.", + "bbox": [ + 94, + 875, + 232, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† On the date of CVPR deadline, i.e., 2023-11-18 07:59 AM UTC.", + "bbox": [ + 94, + 887, + 446, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15311", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "idea, we directly replace historical predictions of temporal point clouds with their ground truth labels for FSA. We find the performance boosts greatly. This inspires us to use the model injected with gt priors to distill the model injected with historical priors, which we term Mask Distillation.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To fully use the informative temporal images, we devise a Temporal Image Aggregation and Fusion (TIAF) module. Previous multi-modal fusion methods usually suffer from limited image features due to the different FOVs (field of views) between the LiDAR and the camera. We observe that as the ego-vehicle moves forward, cameras can capture different FOVs. By aggregating images of different timestamps, the FOV of the present camera can be enlarged greatly. In addition, temporal images can provide different views for the same region at different timestamps, which can supply richer information. Our TIAF leverages temporal LiDAR points in the FOV of the corresponding temporal camera to transform historical image features to the present coordinate with the pose matrix. After aggregating all temporal image features, we use a series of 3D sparse convolutions to fuse them and convert them to voxel representation. Ultimately, we use temporal LiDAR points to gather temporal image features for temporal multi-modal fusion. The fused features embrace both temporal LiDAR and temporal image information, which contribute to more precise segmentation results. To our knowledge, we are the first to leverage temporal point clouds and temporal images simultaneously for LiDAR semantic segmentation.", + "bbox": [ + 75, + 167, + 473, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, in the multi-scan task, the model needs to distinguish the motion states (moving or static) of a movable object, which is very challenging. In this paper, we make full use of temporal information to remarkably increase the multi-scan perception ability of the model. Specifically, we propose a data augmentation strategy named Static-Moving Switch Augmentation (SMSA). It manipulates temporal point clouds of a movable object to switch its motion state. In this way, we can increase the quantity of static and moving samples significantly, even in the absence of training samples of static or moving classes.", + "bbox": [ + 75, + 515, + 470, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our major contributions are listed as follows:", + "bbox": [ + 94, + 681, + 470, + 696 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a Temporal LiDAR Aggregation and Distillation algorithm, which uses Flexible Step Aggregation and Mask Distillation techniques to largely reduce memory and time costs while achieving higher accuracy.", + "- We devise a Temporal Image Aggregation and Fusion module, which exploits temporal images to enlarge the camera FOV and enhance present features. It also delivers a scheme for temporal multi-modal fusion.", + "- We design a Static-Moving Switch Augmentation algorithm to enable static and moving objects to switch their motion states freely. With this technique, we can greatly increase static and moving training samples." + ], + "bbox": [ + 76, + 705, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Our Temporal Aggregation Network, dubbed TASeg, achieves impressive results on SemanticKITTI and nuScenes benchmarks. Notably, our TASeg ranks $1^{st}$ on three challenging tracks. Thorough ablation studies are provided to demonstrate the efficacy of our approach.", + "bbox": [ + 500, + 90, + 893, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 184, + 640, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LiDAR Semantic Segmentation. LiDAR semantic segmentation aims to assign a unique class label to each point in the input point cloud sequence. Recent years have witnessed an explosion of LiDAR segmentation algorithms [5, 7, 13, 15, 19, 20, 24, 27, 31, 39]. For example, [26] is the pioneering work that approximates the permutation-invariant function with a per-point and shared Multi-Layer Perceptron. [46] changes traditional cubic grids to cylindrical grids and designs a network of asymmetrical 3D convolutions. [19] divides the space with the radial window, which increases the receptive field smoothly and helps improve the performance. Despite their good segmentation performance, these methods still take the single LiDAR frame as input, which does not utilize the rich semantic and geometric information hidden in temporal data.", + "bbox": [ + 496, + 210, + 895, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Frame LiDAR Perception. Compared to a single LiDAR scan, multiple LiDAR scans can provide more sufficient information. Recently, many research efforts have been put on temporal LiDAR segmentation [1, 2, 6, 8, 9, 17, 20, 22, 30, 45]. For example, [8] leverages a Bayes filter to explore the temporal consistency. [4] shunts the historical points into two groups to utilize historical frames efficiently. Despite the success of previous multi-frame methods, they can not leverage the valuable information hidden in long-term temporal point clouds due to the GPU memory constraint. Although [20] stacks dozens of LiDAR scans, the memory and time overhead are intolerable. In this paper, we present an efficient multi-frame aggregation algorithm, which can greatly save memory and computation consumption while achieving higher performance.", + "bbox": [ + 496, + 450, + 895, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Modal Fusion. Since LiDAR and camera are two complementary sensors for 3D semantic segmentation, multimodal fusion has gained increasing attention in recent years [10, 18, 24, 43, 47]. However, these multi-modal fusion methods usually suffer from limited overlapped regions between the LiDAR and camera due to their different FOVs. [39] proposes a cross-modal knowledge distillation method, which is free from images at inference, while it causes much loss of the RGB information. [20] completes the missing camera features using predicted pseudo-camera features, while the used image information is still restricted to the present camera FOV. In this paper, we fully take advantage of temporal images to enlarge the FOV of the present camera and enhance present image features.", + "bbox": [ + 496, + 688, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15312", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge Distillation for LiDAR Perception. Knowledge distillation [12] is widely used in various fields in that it can improve the performance of the student without sacrificing inference efficiency. In LiDAR perception, [13, 40] compress a cumbersome teacher to a lightweight student to reinforce the representation learning of the student as well as maintain high efficiency. [39] distills the prior of 2D images to 3D point clouds with well-designed cross-modal knowledge distillation module. [32, 44] utilizes a multi-frame teacher to help a single-frame student learn dense 3D features. By contrast, our method aims to transfer knowledge from a multi-frame teacher injected with gt priors to a multi-frame student injected with historical priors.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 300, + 212, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Temporal LiDAR Aggregation and Distillation", + "text_level": 1, + "bbox": [ + 76, + 325, + 468, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A simple solution to utilize the temporal information of a consecutive point cloud sequence is to concatenate all points of the sequence, as shown in the following equation:", + "bbox": [ + 75, + 348, + 468, + 393 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nX _ {t} = \\operatorname {c o n c a t} \\left(P _ {t}, T _ {t - 1} P _ {t - 1}, \\dots , T _ {t - \\Delta t} P _ {t - \\Delta t}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 402, + 468, + 426 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nY _ {t} = \\operatorname {c o n c a t} \\left(L _ {t}, L _ {t - 1}, \\dots , L _ {t - \\Delta t}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 422, + 357, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $P_{t - i}$ and $L_{t - i}$ denote the $(t - i)^{th}$ point cloud frame and the corresponding point-wise label. $T_{t - i}$ is the transformation matrix that transforms the coordinate from the $(t - i)^{th}$ frame to the $t^{th}$ frame. $\\Delta t$ is the window size of temporal point clouds and $\\text{concat}(.)$ denotes the concatenation operation. $X_{t}$ and $Y_{t}$ are the aggregated LiDAR frame and point-wise label. Although simple, there are some problems. On the one hand, direct concatenation consumes much GPU memory. On the other hand, the huge memory cost constrains the multi-frame model from utilizing more historical frames, thereby limiting the ultimate performance.", + "bbox": [ + 75, + 450, + 472, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Flexible Step Aggregation", + "text_level": 1, + "bbox": [ + 76, + 635, + 310, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To reduce the GPU memory consumption, we can sample temporal frames with a step, while this method is also suboptimal. A small step introduces huge memory overhead, whilst a large step cannot utilize sufficient temporal information. To this end, we propose Flexible Step Aggregation (FSA). Our method is based on the observation that for different classes, the model needs different amounts of temporal information. For difficult classes, such as bicycles, more points are needed to yield accurate predictions. For easy classes, such as cars and buildings, it is unnecessary to aggregate many point cloud frames for them. Based on the above analysis, we propose to assign different steps for different classes according to their learning difficulty. Specifically, we leverage historical predictions to divide the temporal point clouds into several class groups and assign a specific sampling step for each group. Then, we aggregate temporal", + "bbox": [ + 75, + 659, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "points for each group with the corresponding step. Eventually, temporal points of all groups are concatenated with the current frame, resulting in an aggregated frame.", + "bbox": [ + 496, + 90, + 893, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The group division is not strict as long as it follows the principle that more difficult classes need smaller steps. For example, we can simply divide all classes into three groups according to their segmentation performance, such as [0, 80) mIoU, [80, 90) mIoU and [90, 100] mIoU. Then, we assign the three groups with a step of 2, 4 and $\\infty$ ( $\\infty$ means we do not aggregate temporal points for the group). To save more memory and computation without sacrificing performance, a more fine-grained division can be used. Formally, suppose we divide temporal point clouds into $g$ groups. We aggregate temporal points for the $k^{th}$ group as below:", + "bbox": [ + 496, + 136, + 893, + 303 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nX _ {t} ^ {k} = \\operatorname {c o n c a t} \\left(T _ {o _ {1}} P _ {o _ {1}} M _ {o _ {1}} ^ {k}, \\dots , T _ {o _ {n}} P _ {o _ {n}} M _ {o _ {n}} ^ {k}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 309, + 893, + 335 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nY _ {t} ^ {k} = \\operatorname {c o n c a t} \\left(L _ {o _ {1}} M _ {o _ {1}} ^ {k}, \\dots , L _ {o _ {n}} M _ {o _ {n}} ^ {k}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 330, + 799, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $o = \\{o_i|o_i = t - i\\times s_k,i = 1,2,\\dots,n,n = \\lfloor \\Delta t / s_k\\rfloor \\}$ $s_k$ is the sampling step for the $k^{th}$ class group and $\\lfloor .\\rfloor$ is the floor operation. $M_{o_i}^k$ is the group mask that indicates which point of $P_{o_i}$ belongs to the $k^{th}$ class group. It can be obtained from historical predictions. Finally, we concatenate points and labels of all groups with the current frame:", + "bbox": [ + 496, + 357, + 893, + 448 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nX _ {t} = \\operatorname {c o n c a t} \\left(P _ {t}, X _ {t} ^ {1}, \\dots , X _ {t} ^ {k}, \\dots , X _ {t} ^ {g}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 455, + 893, + 481 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nY _ {t} = \\operatorname {c o n c a t} \\left(L _ {t}, Y _ {t} ^ {- 1}, \\dots , X _ {t} ^ {k}, \\dots , Y _ {t} ^ {g}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 477, + 826, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this way, we discard massive redundant temporal points while maintaining essential temporal information. Since easy classes are usually large and hold a large quantity of points, the GPU memory overhead can be further reduced. Moreover, by assigning different classes with different steps, FSA actually injects historical priors into the aggregated point clouds. For classes with different steps, their patterns can be more discriminative due to their different densities (steps), which makes it easier to segment the multi-frame point clouds aggregated by FSA. Experiments in Table 4 verify that the proposed FSA can not only save memory and time costs but also achieve better performance.", + "bbox": [ + 496, + 502, + 893, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2 Mask Distillation", + "text_level": 1, + "bbox": [ + 500, + 695, + 673, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In FSA, we use historical predictions to generate group masks, which we call pseudo masks. Since ground truth labels are more accurate than historical predictions, a natural question arises: what if we use gt masks (the group masks generated by ground truth labels) for FSA? Our experiment shows that the performance can be improved greatly. Actually, using gt masks for FSA can produce more discriminative patterns for classes with different steps. This motivates us to simulate features of a model trained on temporal point clouds aggregated with gt masks. In this way, our model is guided to learn more discriminative features to distinguish different classes. We call this Mask Distillation. Specifically,", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15313", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9b15e5bcf6fe62695fdcac9bb7297f21a8bc49fd7ab50e8acab5c209ffef3555.jpg", + "image_caption": [ + "Figure 1. Overview of our Temporal Aggregation Network (TASeg). (1) Temporal LiDAR Aggregation and Distillation leverages the proposed Flexible Step Aggregation (FSA) to assign different temporal steps for different classes, and it utilizes a teacher injected with gt priors for knowledge distillation. (2) Temporal Image Aggregation and Fusion takes temporal LiDAR points as mediums to transform historical image features to the present coordinate. 3D sparse convolutions are employed to fuse temporal image features. Finally, we use temporal LiDAR points to gather voxel-wise temporal image features for temporal multi-modal fusion." + ], + "image_footnote": [], + "bbox": [ + 86, + 92, + 883, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we formulate the distillation on the feature map of student $F^{s}\\in \\mathbb{R}^{N^{s}\\times C}$ and the feature map of teacher $F^{t}\\in \\mathbb{R}^{N^{t}\\times C}$ :", + "bbox": [ + 76, + 517, + 473, + 550 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K D}} = \\mathbb {E} \\left[ \\left\\| F ^ {s} m ^ {s} - F ^ {t} m ^ {t} \\right\\| _ {2} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 561, + 470, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since temporal point clouds aggregated with pseudo masks and gt masks are different, we use masks $m^s$ and $m^t$ to select voxels that appeared in both $F^s$ and $F^t$ . Note that our Mask Distillation is distinct from the methods that distill a multi-frame model to a single-frame model [32, 44]. Both the student and teacher in our method are multi-frame models.", + "bbox": [ + 75, + 582, + 470, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Temporal Image Aggregation and Fusion", + "text_level": 1, + "bbox": [ + 76, + 681, + 429, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous multi-modal fusion methods only focus on leveraging present images while ignoring the precious value of temporal images. Temporal images can provide broader camera FOVs and richer information. Besides, they can enable more robust multi-modal fusion under the malfunction condition on some cameras. In this section, we provide an effective solution for aggregating temporal image features and performing temporal multi-modal fusion.", + "bbox": [ + 75, + 704, + 470, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Aggregation and Fusion. Since temporal images are in different feature spaces, it is difficult to establish the relationship between different images for feature aggregation. In our method, we take temporal LiDAR points as mediums to transform temporal image features to the present", + "bbox": [ + 75, + 825, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "coordinate with the pose information. This way, temporal image features are unified to the present 3D space. Specifically, given an image $I_{t - \\Delta t}\\in \\mathbb{R}^{H\\times W\\times 3}$ and point cloud $P_{t - \\Delta t}\\in \\mathbb{R}^{N\\times 3}$ , we use an image network to extract the image feature $Z_{t - \\Delta t}\\in \\mathbb{R}^{H\\times W\\times C}$ . According to the sensor calibration, we can establish the pixel-to-point mapping between 2D pixels and 3D points. Hence, we can project image feature $Z_{t - \\Delta t}$ to 3D space, resulting in point-wise image feature $Q_{t - \\Delta t}\\in \\mathbb{R}^{M\\times C}$ , where $M$ is the number of LiDAR points located on $Z_{t - \\Delta t}$ . By transforming $Q_{t - \\Delta t}$ to the present coordinate with the pose matrix, we realize the aggregation of temporal image features:", + "bbox": [ + 496, + 517, + 893, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {X} _ {t} = \\operatorname {c o n c a t} \\left(Q _ {t}, T _ {o _ {1}} Q _ {o _ {1}}, \\dots , T _ {o _ {n}} Q _ {o _ {n}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 705, + 893, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $o = \\{o_i|o_i = t - i\\times s,i = 1,2,\\dots,\\lfloor \\Delta t / s\\rfloor \\}$ and $s$ is the sampling step for temporal images. With temporal images aggregated, the FOV of the present camera is expanded greatly. Moreover, temporal image feature fusion becomes convenient because they are unified to the same 3D space. Concretely, we can use several 3D sparse convolutions to fuse aggregated temporal image features, which also endows them with geometric information, as shown in Equation 6. In addition, feature map downsampling is also utilized to generate multi-scale voxel features, providing richer information for subsequent temporal multi-modal fusion.", + "bbox": [ + 496, + 734, + 895, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15314", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {V} _ {t} = \\operatorname {S p a r s e C o n v} _ {\\mathrm {3 D}} \\left(\\operatorname {V o x e l i z a t i o n} \\left(\\bar {X} _ {t}\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 104, + 470, + 123 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal Multi-Modal Fusion. Benefiting from the temporal image aggregation, associating temporal image features with temporal LiDAR points also becomes convenient. Specifically, given temporal image features converted to unified voxel representation $\\overline{V}_t$ , we can establish a point-to-voxel association between temporal LiDAR points and $\\overline{V}_t$ . For each temporal LiDAR point, we generate its image feature by pooling its nearby voxel-wise temporal image features with trilinear interpolation instead of hard indexing. To extract richer image features, we perform interpolation on multi-scale feature maps. Finally, we concatenate the point cloud features and aggregated multi-scale image features, resulting in fused features, which convey powerful information of both temporal point clouds and temporal images.", + "bbox": [ + 75, + 128, + 472, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2D and 3D Supervision. To make the extracted image features more informative, we add 2D supervision and 3D supervision on the 2D backbone and 3D convolutions in the image branch, respectively. The 3D supervision is just the label of point clouds. The 2D supervision is obtained by projecting labels of point clouds to the image plane.", + "bbox": [ + 75, + 347, + 470, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Static-Moving Switch Augmentation", + "text_level": 1, + "bbox": [ + 76, + 446, + 392, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the multi-scan task, the model is required to distinguish the motion state of movable objects. To enable the model to explore a large data space, we design an effective data augmentation, Static-Moving Switch Augmentation (SMSA). SMSA enables a movable object to switch its motion state freely, which can remarkably increase the sample quantity of static and moving objects. Concretely, considering that a unique object $b$ has the same instance id in all frames of a sequence, we can use its instance mask to crop its temporal point clouds, which is denoted as $\\mathcal{B} = \\{\\mathcal{B}_{o_i}|i = 0,1,\\dots,\\lfloor \\Delta t / s\\rfloor ,\\mathcal{B}_{o_i}\\in \\mathbb{R}^{N\\times 3}\\}$ . Here $o = \\{o_i|o_i = t - i\\times s,i = 1,2,\\dots,\\lfloor \\Delta t / s\\rfloor \\}$ , $s$ is the sampling step and $\\mathcal{B}_{o_i}$ is the temporal part of $b$ at $t_{o_i}$ . By manipulating $\\mathcal{B}$ , we can change the motion state of $b$ .", + "bbox": [ + 75, + 470, + 470, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Static to Moving. If $b$ is static, all temporal parts of $b$ locate at the same position, as shown in the upper-left of Figure 2. To change $b$ to a moving object, we can shift each temporal part of $b$ with an offset. Considering that objects typically move at a constant speed within a short time, we set the offset between adjacent temporal parts to be the same. For offset itself, it is a random value to increase the diversity of resulting moving samples. As for the direction of the offset, it can be roughly estimated by comparing the width and height of the minimum bounding box of $b$ . Since static objects often park on the side of the road, which is crowded, the shifted temporal parts of static objects may overlap with other objects. To alleviate this, we define a set of anchor points around $b$ and a coverage area for each anchor point.", + "bbox": [ + 75, + 688, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a2410a010fdcb1d5c563c6b1ae14acc2326ee362bd875d1ed10945435cc43774.jpg", + "image_caption": [ + "Figure 2. Visualization of the augmented samples by our Static-Moving Switch Augmentation (SMSA). Our SMSA switches the motion state by manipulating the temporal parts of objects." + ], + "image_footnote": [], + "bbox": [ + 509, + 95, + 893, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we shift all temporal parts of $b$ to the anchor point whose coverage area contains the fewest LiDAR points (refer to the supplementary material for more details).", + "bbox": [ + 496, + 309, + 890, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moving to Static. If $b$ is moving, different temporal parts of $b$ locate at different positions. To switch its motion state to static, we can shift all its temporal parts to the same position. In particular, we calculate the centers of each temporal part of $b$ , which are denoted as $\\mathcal{C} = \\{\\mathcal{C}_{o_i} | \\mathcal{C}_{o_i} \\in \\mathbb{R}^3\\}$ . Considering the trajectory of moving objects in a short time is approximate to a line and the speed is a constant, we simply use $q = \\mathcal{C}_{o_0} - \\mathcal{C}_{o_1}$ as the offset of adjacent temporal parts of $b$ . Eventually, we can obtain a static object by shifting all temporal parts of $b$ to $C_{o_0}$ (see Figure 2).", + "bbox": [ + 496, + 359, + 893, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It should be noted that our SMSA is an online and plug-and-play data augmentation strategy, which consumes negligible storage and computation costs.", + "bbox": [ + 496, + 513, + 893, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Objective", + "text_level": 1, + "bbox": [ + 500, + 566, + 681, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final training objective is comprised of LiDAR segmentation loss, mask distillation loss, fusion segmentation loss, 2D and 3D supervision loss on the image branch:", + "bbox": [ + 496, + 590, + 893, + 636 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {L i D A R}} + \\alpha \\mathcal {L} _ {\\mathrm {K D}} + \\beta \\mathcal {L} _ {\\text {F u s i o n}} + \\gamma (\\mathcal {L} _ {\\mathrm {2 D}} + \\mathcal {L} _ {\\mathrm {3 D}}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 638, + 890, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha, \\beta, \\gamma$ are the coefficients to control the effect of different losses. We set $\\alpha = 1, \\beta = 1, \\gamma = 1$ by default.", + "bbox": [ + 496, + 665, + 893, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 709, + 633, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets & Evaluation Metrics. Following [13, 46], we evaluate the performance on two popular LiDAR segmentation benchmarks, i.e., SemanticKITTI and nuScenes. SemanticKITTI contains 22 point cloud sequences, where sequences 0-10, 8, 11-21 are selected as training, validation and testing, respectively. As for nuScenes, it collects 1, 000 driving scenes, where 850 scenes are chosen for training and validation, and the remaining 150 scenes are used for testing. We conduct experiments on three tracks, i.e., SemanticKITTI single-scan and multi-scan semantic segmentation and nuScenes semantic segmentation.", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15315", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/27f1dd895fec9f632f5cad48f04f57d61a6ae8de8b545a0e7f2c2e2613f91dff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmIoUcarbicyclemotorcycletruckother-veh.personbicyclistmotorcyclistroadparkingsidewalkother-gro.buildingfencevegetationtrunkterrainpoletraffic sign
SqueezeSegV2 [33]39.781.818.517.913.414.020.125.13.988.645.867.617.773.741.171.835.860.220.226.3
RangeNet53++ [25]52.291.425.734.425.723.038.338.84.891.865.075.227.887.458.680.555.164.647.955.9
RandLA-Net [14]55.994.229.832.243.939.148.447.49.490.561.874.024.589.760.483.863.668.651.050.7
SqueezeSegV3 [36]55.992.538.736.529.633.045.646.220.191.763.474.826.489.059.482.058.765.449.658.9
JS3C-Net [38]66.095.859.352.954.346.069.565.439.988.961.972.131.992.570.884.569.867.960.768.7
SPVNAS [31]67.097.250.650.456.658.067.467.150.390.267.675.421.891.666.986.173.471.064.367.3
Cylinder3D [46]68.997.167.663.850.858.573.769.248.092.265.077.032.390.766.585.672.569.862.466.2
RPVNet [37]70.397.668.468.744.261.175.974.443.493.470.380.733.393.572.186.575.171.764.861.4
(AF)2-S3Net [5]70.894.363.081.440.240.076.481.777.792.066.876.245.892.569.678.668.063.164.073.3
PVKD [13]71.297.069.353.567.960.275.173.550.591.877.570.941.092.469.486.573.871.964.965.8
2DPASS [39]72.997.063.663.461.161.577.981.374.189.767.474.740.093.572.986.273.971.065.070.4
SphereFormer [19]74.897.570.170.559.667.779.080.475.391.869.778.241.393.872.886.775.172.466.872.9
UniSeg [24]75.297.971.975.263.674.178.974.860.692.674.079.546.193.472.787.576.373.168.368.5
TASeg (Ours)76.597.771.871.465.278.779.984.678.691.674.078.039.393.573.486.675.071.769.673.8
", + "bbox": [ + 81, + 89, + 890, + 377 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8a4ff21bd597ca0fddc813029ac0406dadc8330c962e35e4934336925dee7e73.jpg", + "table_caption": [ + "Table 1. Comparison with state-of-the-arts on SemanticKITTI test set (single-scan). All results can be found on the online leaderboard." + ], + "table_footnote": [], + "table_body": "
MethodmIoUcar-scar-mtruc.-struc.-mothe.-sothe.-mpers.-spers.-mbicy.-sbicy.-mmotor.-smotor.-m
LatticeNet [29]45.291.154.829.73.523.10.66.849.90.044.60.064.3
TemporalLidarSeg [9]47.092.168.239.22.135.012.414.440.40.042.80.012.9
\\( (AF)^2-S3Net[5] \\)56.991.865.315.75.627.53.916.467.615.166.467.159.6
MarS3D [22]52.795.178.439.75.136.610.016.258.01.267.30.036.3
SVQNet [4]60.596.080.141.05.160.47.128.785.10.072.40.088.1
2DPASS [39]62.496.282.148.216.152.73.835.480.37.971.262.073.1
TASeg (Ours)65.794.872.853.635.271.453.236.780.432.475.212.389.6
", + "bbox": [ + 199, + 407, + 772, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison to the state-of-the-art methods on SemanticKITTI test set (multi-scan). -s indicates static and -m stands for moving.", + "bbox": [ + 78, + 574, + 887, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We use the MinkowskiNet [7] re-implemented by PCSeg [23] as our baseline. Our TASeg is trained with the SGD [28] optimizer on 4 A100 GPUs with batch size 6 for 12 and 36 epochs on SemanticKITTI and nuScenes datasets. The learning rate and weight decay are set to 0.02 and 0.0001. The window size of temporal point clouds is set to 16. Due to a large redundancy of images, we use a step of 12 and a window size of 48 for temporal images. Our data augmentation strategy includes random flipping, rotation, scaling, transformation, LaserMix [16] and PolarMix [35]. During the inference, the LiDAR semantic predictions and image features of the previous timestamp can be saved, so there is no time cost for generating historical predictions and processing temporal image aggregation. More details are provided in the supplementary material.", + "bbox": [ + 75, + 601, + 473, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 76, + 844, + 441, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our TASeg with state-of-the-art LiDAR segmentation methods on SemanticKITTI single-scan and multi", + "bbox": [ + 75, + 869, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "scan track, and nuScenes LiDAR semantic segmentation track, as summarized in Table 1, 2 and 3. On the SemanticKITTI single-scan track, our TASeg is $1.3\\mathrm{mIoU}$ higher than UniSeg [24], the best-performing published algorithm on the SemanticKITTI single-scan leaderboard. On the SemanticKITTI multi-scan track, our approach outperforms the best multi-scan method 2DPASS [39] by $3.3\\mathrm{mIoU}$ . For nuScenes LiDAR semantic segmentation, we also achieve superior performance over previous algorithms. As shown in Table 3, the proposed approach surpasses the UniSeg [24] by $1.1\\mathrm{mIoU}$ . In particular, our approach holds the highest entry on all three tracks. These encouraging results strongly show the effectiveness of our approach.", + "bbox": [ + 496, + 601, + 893, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation studies", + "text_level": 1, + "bbox": [ + 500, + 811, + 661, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We examine the effect of each component through detailed ablations. Unless otherwise specified, the following experiments are trained on the training set of the SemanticKITTI single-scan task and evaluated on the validation set.", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15316", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b121b255f5e74594bd8e7a2b5b141e1dab4d8b771af4528441c9412b7e1f783f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmIoUconstruction
barrierbicyclebuscarconstructionmotorcyclepedestriantraffic-conetrailertruckdriveableother-flatsidewalkterrainmannadevegetation
Cylinder3D [46]77.282.829.884.389.463.079.377.273.484.669.197.770.280.375.590.487.6
SPVCNN [31]77.480.030.091.990.864.779.075.670.981.074.697.469.280.076.189.387.1
2D3DNet [11]80.083.059.488.085.163.784.482.076.084.871.996.967.479.876.092.189.2
2DPASS [39]80.881.755.392.091.873.386.578.572.584.775.597.669.179.975.590.288.0
LidarMultiNet [41]81.480.448.494.390.071.587.285.280.486.974.897.867.380.776.592.189.6
MSeg3D [20]81.183.142.594.992.067.178.685.780.587.577.397.769.881.277.892.490.1
SphereFormer [19]81.983.339.294.792.577.584.284.479.188.478.397.969.081.577.293.490.2
UniSeg [24]83.585.971.292.191.680.588.080.976.086.376.797.771.880.776.791.388.8
TASeg (Ours)84.687.169.490.592.278.790.486.381.988.375.997.870.981.078.293.491.2
", + "bbox": [ + 122, + 88, + 851, + 308 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Flexible Step Aggregation. In Table 4, we compare our FSA with other multi-frame algorithms. MSeg3D [20] utilizes temporal point clouds by directly concatenating all points, which introduces a large memory consumption and achieves limited improvement. SVQNet [4] leverages cross-attention to fuse temporal features while it can only handle fewer LiDAR scans. When increasing LiDAR scans, the memory overhead is also huge. Our FSA assigns different classes with different temporal steps, which can leverage long-term temporal information with less memory and time costs. Moreover, thanks to the utilization of historical priors, our FSA achieves better results (71.3 mIoU) than directly concatenating (69.9 mIoU), given the same temporal window size. To explore the effect of different window sizes on FSA, we present Table 5. Results show that the performance is saturated with a window size of 24.", + "bbox": [ + 75, + 342, + 473, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Group Division. To investigate the effect of different group divisions for FSA, we provide Table 6. For division1, we simply divide all classes into three groups according to their performance on the val set. The group1 consists of the classes in [90, 100) mIoU, such as cars and roads. The group2 consists of classes in [80, 90) mIoU, such as motorcycles. The group3 contains the remaining classes. We assign group1, group2 and group3 with the step of $\\infty$ , 4 and 2. A step of $\\infty$ means we do not aggregate temporal points for classes in the group due to their near-saturate performances. The result shows that division1 can reduce the time and memory costs largely compared to directly stacking (the $2^{st}$ row in Table 4). We can also use performance top1-6, top6-12, and top12-19 for division, resulting in division2. The group division is robust as long as it follows the principle that more difficult classes need smaller steps. To further reduce the memory consumption, we can finetune the group division, such as moving large-size classes (e.g. other-ground and terrain) in group3 to group2 (division3) or moving large-size classes in group3 and group2 to a new", + "bbox": [ + 75, + 598, + 473, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d6d69026c35996167a3e859cd75290059fa9806224e6a484a5be59d6cc8f19ab.jpg", + "table_caption": [ + "Table 3. Performance comparison with start-of-the-art methods on nuScenes test set." + ], + "table_footnote": [], + "table_body": "
MethodLatencyMemorymIoU
Baseline63ms5.0G68.9
+ MSeg3D [20]†284ms69.7G69.9
+ SVQNet [4]†121ms15.4G69.5
+ FSA79ms8.5G71.3
+ FSA w/ gt mask79ms8.5G75.9
+ FSA + M.D.79ms8.5G71.8
+ FSA + M.D. × 0.7572ms6.9G71.6
", + "bbox": [ + 516, + 343, + 877, + 459 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4c5314fd218c1736de730cd7a15343608901615afec72fd25ad17cc4c43e2c73.jpg", + "table_caption": [ + "Table 4. Comparison with different multi-frame algorithms. $\\dagger$ represents our re-implementation. \"M.D.\" represents Mask Distillation. \"×75\" represents that we reduce the model parameters to $75\\%$ ." + ], + "table_footnote": [], + "table_body": "
Window size481216202428
mIoU69.970.470.971.371.471.671.4
", + "bbox": [ + 516, + 513, + 877, + 550 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/66685ffc0e1d35ac50579a0787643e624055db9a2c165568b097682cb22664de.jpg", + "table_caption": [ + "Table 5. Ablation on different window sizes for FSA." + ], + "table_footnote": [], + "table_body": "
MethodLatencyMemorymIoU
FSA w/ division192ms14.1G70.8
FSA w/ division283ms7.2G70.6
FSA w/ division379ms8.5G71.3
FSA w/ division474ms7.7G71.2
FSA w/ division574ms7.4G71.2
", + "bbox": [ + 524, + 577, + 867, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablation on different group divisions for FSA.", + "bbox": [ + 529, + 672, + 859, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "group with step of 8 (division4). Besides, we can divide each class group into a close and a distant group according to the distance (such as $30\\mathrm{m}$ ). Considering close areas need fewer temporal points, we can use twice the original step for temporal aggregation of close groups (division5).", + "bbox": [ + 496, + 694, + 893, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Mask Distillation. At the $5^{th}$ row of Table 4, we use gt masks of historical frames for FSA. The performance is greatly boosted. Considering we cannot get ground truth at inference, we use it as a teacher to distill the model trained with pseudo masks. The result shows that the student is improved by $0.5\\mathrm{mIoU}$ after distillation. With $75\\%$ model complexity, we can further reduce the time and memory cost and still achieve a leading performance.", + "bbox": [ + 496, + 779, + 895, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15317", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Temporal Image Aggregation and Fusion. TIAF delivers a temporal multi-modal fusion scheme to make full use of temporal images. As shown in Table 7, with zero image, only TLAD is used, achieving $71.8\\mathrm{mIoU}$ . With one image, only present images are utilized. Due to the limited FOV of the present camera, only part of the LiDAR points can gather image features, which limits the multi-modal fusion. Our TIAF leverages temporal images to enlarge the FOV of the camera and enhance present features. The results show that with the number of images increasing, the performance rises gradually, which verifies the effectiveness of our TIAF. When using seven temporal images, we achieve $1.0\\mathrm{mIoU}$ improvement on the strong baseline. Note that there is no extra cost to generate historical image features because we can save them at the past moments.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/18ffc2b10ec61de70f86f2abf947c3ae938456c6dc1d58b4cb3a0cff30e48ad7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Num of images01357
mIoU71.872.172.472.772.8
", + "bbox": [ + 91, + 325, + 454, + 362 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/36c7277fdbb3cbb35ab666e290b46a18eda2dc8111b087971c49a19ab59fe6e5.jpg", + "table_caption": [ + "Table 7. Ablation on the number of temporal images for TIAF." + ], + "table_footnote": [], + "table_body": "
Exp2D Sup.3D Sup.Multi-ScalemIoU
(a)✓✓72.4
(b)✓✓72.5
(c)✓✓72.4
(d)✓✓✓72.7
", + "bbox": [ + 142, + 388, + 406, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also provide the ablation on different supervisions and multi-scale features for TIAF in Table 8. In TIAF, we leverage 2D and 3D supervision to guide the extracted image features to be more conducive for segmentation. Multi-scale is also used to provide more discriminative features. The result shows that each of the designs is beneficial for final performance. Note that our temporal multi-modal method is orthogonal to other single-frame multi-modal methods. We use a simple pixel-to-point mapping at the feature projection stage. More complex methods [20, 24] can also be utilized, but it is not the focus of our TIAF.", + "bbox": [ + 75, + 497, + 468, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Static-Moving Switch Augmentation. To verify our SMSA, we present Table 9. From the results, we can find that without SMSA, the accuracy of the model on many classes is lower than $10\\mathrm{IoU}$ . After using SMSA, the performance of moving truck, moving other-vehicle and static bicyclist classes is improved by more than $20\\mathrm{IoU}$ , and the overall performance boosts from $61.3\\mathrm{mIoU}$ to $65.7\\mathrm{mIoU}$ , which strongly demonstrates the efficacy of our SMSA.", + "bbox": [ + 75, + 672, + 468, + 792 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d512a85a8f7b9fb58e8a8f6d7119d217faf8cd4b7185b3e26e198f1e4b0c5bb1.jpg", + "table_caption": [ + "Table 8. Effect of different supervisions and multi-scale for TIAF." + ], + "table_footnote": [], + "table_body": "
SMSAmloUcar-scar-mtrue-strue-mother-sother-mpers-spers-mbicy-sbicy-mmotor-smotor-m
✓61.395.679.349.29.161.05.739.087.00.072.64.376.4
65.794.872.853.635.271.453.236.780.432.475.212.389.6
", + "bbox": [ + 81, + 797, + 467, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization to More Architectures. In addition to MinkowskiNet, we also validate our temporal aggregation method on Cylinder3D [46] and SPVNAS [31]. Detailed results are summarized in Table 10. Our TASeg brings impressive improvements to both baselines, which demonstrates the strong generalization ability of our method.", + "bbox": [ + 496, + 90, + 893, + 181 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/971fd6e83459116b596c16e785355270e930e16737abff4849dddebcc51f3b84.jpg", + "table_caption": [ + "Table 9. Ablation study of our SMSA on the SemanticKITTI test benchmark of the multi-scan track." + ], + "table_footnote": [], + "table_body": "
MethodmIoU
Cylinder3D [46]†66.0
Cylinder3D [46]†+ TASeg69.6
SPVNAS [31]†68.6
SPVNAS [31]†+ TASeg72.7
", + "bbox": [ + 566, + 193, + 826, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison on Complexity, Latency and Accuracy. As shown in Table 11, when not utilizing temporal images, our TASeg achieves superior performance than other methods with comparable or less complexity and latency. With temporal images, TASeg can maintain lower complexity and latency than UniSeg [24] while achieving higher accuracy. Besides, with $75\\%$ parameters, the latency of TASeg can be further reduced, and the accuracy is not affected much.", + "bbox": [ + 496, + 316, + 893, + 436 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/329f087cdd65c3e26136edd04ed16c660e6a7c7f04bf598b53687186be96711e.jpg", + "table_caption": [ + "Table 10. Performance of applying our TASeg to different baselines. $\\dagger$ represents our re-implementation." + ], + "table_footnote": [], + "table_body": "
MethodInputParamsLatencymIoU
MinkowskiNet [7]†L37.9M63ms68.4
Cylinder3D [46]†L55.9M75ms66.0
SPVNAS [31]†L96.7M105ms68.6
RPVNet [37]†L119.0M110ms68.9
UniSeg [24]L+C147.6M145ms71.3
TASeg wo/ TIAFL+T37.9M79ms71.8
TASegL+C+T46.7M116ms72.7
TASeg × 0.75L+C+T27.4M108ms72.5
", + "bbox": [ + 506, + 445, + 885, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 11. Comparison of the model complexity, latency and accuracy on SemanticKITTI val set. $\\dagger$ represents our re-implementation on NVIDIA A100 for a fair comparison. \"L\", \"C\" and \"T\" represent LiDAR, Camera and Temporal data, respectively.", + "bbox": [ + 496, + 573, + 893, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 643, + 617, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present TASeg to fully exploit temporal point clouds and temporal images. Our TASeg is comprised of TLAD, TIAF and SMSA. We perform extensive experiments to validate the efficacy of each component. Moreover, our TASeg set a new state-of-the-art on three challenging tracks, demonstrating the superiority of our method.", + "bbox": [ + 496, + 666, + 893, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 768, + 687, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by The National Nature Science Foundation of China (Grant Nos: 62273301, 62273302, 62036009, 61936006, 62273303), in part by Yongjiang Talent Introduction Programme (Grant No: 2022A-240-G, 2023A-197-G), in part by the National Key R&D Program of China (NO.2022ZD0160100). This work was done during the first author's internship at Shanghai AI Laboratory.", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15318", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mehmet Aygun, Aljosa Osep, Mark Weber, Maxim Maximov, Cyrill Stachniss, Jens Behley, and Laura Leal-Taixe. 4d panoptic lidar segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5527-5537, 2021. 2", + "[2] Hanwen Cao, Yongyi Lu, Cewu Lu, Bo Pang, Gongshen Liu, and Alan Yuille. Asap-net: Attention and structure aware point cloud sequence segmentation. arXiv preprint arXiv:2008.05149, 2020. 2", + "[3] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. arXiv preprint arXiv:2306.16927, 2023. 1", + "[4] Xuechao Chen, Shuangjie Xu, Xiaoyi Zou, Tongyi Cao, Dit-Yan Yeung, and Lu Fang. Svqnet: Sparse voxel-adjacent query network for 4d spatio-temporal lidar semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8569–8578, 2023. 1, 2, 6, 7", + "[5] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. (af)2-S3Net: Attentive Feature Fusion with Adaptive Feature Selection for Sparse Semantic Segmentation Network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12547-12556, 2021. 2, 6", + "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2", + "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 2, 6, 8", + "[8] Ayush Dewan and Wolfram Burgard. Deeptemporalseg: Temporally consistent semantic segmentation of 3d lidar scans. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 2624-2630. IEEE, 2020. 2", + "[9] Fabian Duerr, Mario Pfaller, Hendrik Weigel, and Jürgen Beyerer. Lidar-based recurrent 3d semantic segmentation with temporal memory alignment. In 2020 International Conference on 3D Vision (3DV), pages 781-790. IEEE, 2020. 1, 2, 6", + "[10] Khaled El Madawi, Hazem Rashed, Ahmad El Sallab, Omar Nasr, Hanan Kamel, and Senthil Yogamani. Rgb and lidar fusion based 3d semantic segmentation for autonomous driving. In IEEE Intelligent Transportation Systems Conference, pages 7-12, 2019. 2", + "[11] Kyle Genova, Xiaoqi Yin, Abhijit Kundu, Caroline Pantofaru, Forrester Cole, Avneesh Sud, Brian Brewington, Brian Shucker, and Thomas Funkhouser. Learning 3d semantic segmentation with only 2d image supervision. In 3DV, 2021. 7", + "[12] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3", + "[13] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-Voxel Knowledge Distillation for LiDAR Semantic Segmentation. In IEEE Conference on Computer" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vision and Pattern Recognition, pages 8479-8488, 2022. 2, 3, 5, 6", + "[14] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 6", + "[15] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 228-240, 2023. 2", + "[16] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. Lasermix for semi-supervised lidar semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21705-21715, 2023. 6", + "[17] Lars Kreuzberg, Idil Esen Zulfikar, Sabarinath Mahadevan, Francis Engelmann, and Bastian Leibe. 4d-stop: Panoptic segmentation of 4d lidar using spatio-temporal object proposal generation and aggregation. In European Conference on Computer Vision, pages 537-553. Springer, 2022. 2", + "[18] Georg Krispel, Michael Opitz, Georg Waltner, Horst Possegger, and Horst Bischof. Fuseseg: Lidar point cloud segmentation fusing multi-modal data. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1874-1883, 2020. 2", + "[19] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. arXiv preprint arXiv:2303.12766, 2023. 1, 2, 6, 7", + "[20] Jiale Li, Hang Dai, Hao Han, and Yong Ding. MSeg3D: Multi-modal 3D Semantic Segmentation for Autonomous Driving. arXiv preprint arXiv:2303.08600, 2023. 1, 2, 7, 8", + "[21] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research, 20(6):837-854, 2023. 1", + "[22] Jiahui Liu, Chirui Chang, Jianhui Liu, Xiaoyang Wu, Lan Ma, and Xiaojuan Qi. Mars3d: A plug-and-play motion-aware model for semantic segmentation on multi-scan 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9372-9381, 2023. 1, 2, 6", + "[23] Youquan Liu, Yeqi Bai, Lingdong Kong, Runnan Chen, Yuenan Hou, Botian Shi, and Yikang Li. Pcseg: An open source point cloud segmentation codebase. https://github.com/PJLab-ADG/PCSeg, 2023.6", + "[24] Youquan Liu, Runnan Chen, Xin Li, Lingdong Kong, Yuchen Yang, Zhaoyang Xia, Yeqi Bai, Xinge Zhu, Yuexin Ma, Yikang Li, et al. Uniseg: A unified multi-modal lidar segmentation network and the openpcseg codebase. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21662-21673, 2023. 1, 2, 6, 7, 8", + "[25] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2019. 6", + "[26] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15319", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2", + "[27] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in neural information processing systems, pages 5099-5108, 2017. 2", + "[28] Herbert Robbins and Sutton Monro. A Stochastic Approximation Method. The annals of mathematical statistics, pages 400-407, 1951. 6", + "[29] Radu Alexandru Rosu, Peer Schütt, Jan Quenzel, and Sven Behnke. Latticenet: Fast point cloud segmentation using permutohedral lattices. arXiv preprint arXiv:1912.05905, 2019.6", + "[30] Hanyu Shi, Guosheng Lin, Hao Wang, Tzu-Yi Hung, and Zhenhua Wang. Spsequencenet: Semantic segmentation network on 4d point clouds. In CVPR, 2020. 2", + "[31] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In European Conference on Computer Vision, pages 685-702. Springer, 2020. 2, 6, 7, 8", + "[32] Jiexiong Tang, John Folkesson, and Patric Jensfelt. Sparse2dense: From direct sparse odometry to dense 3-d reconstruction. IEEE Robotics and Automation Letters, 4(2): 530-537, 2019. 3, 4", + "[33] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In 2019 International Conference on Robotics and Automation (ICRA), pages 4376-4382. IEEE, 2019. 6", + "[34] Dong Wu, Man-Wen Liao, Wei-Tian Zhang, Xing-Gang Wang, Xiang Bai, Wen-Qing Cheng, and Wen-Yu Liu. Yolop: You only look once for panoptic driving perception. Machine Intelligence Research, 19(6):550-562, 2022. 1", + "[35] Aoran Xiao, Jiaxing Huang, Dayan Guan, Kaiwen Cui, Shijian Lu, and Ling Shao. PolarMix: A General Data Augmentation Technique for LiDAR Point Clouds. arXiv preprint arXiv:2208.00223, 2022. 6", + "[36] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeezesegv3: Spatially-adaptive convolution for efficient point-cloud segmentation. In European Conference on Computer Vision, pages 1-19. Springer, 2020. 6", + "[37] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. RPVnet: A Deep and Efficient Range-Point-Voxel Fusion Network for Lidar Point Cloud Segmentation. In IEEE International Conference on Computer Vision, pages 16024–16033, 2021. 6, 8", + "[38] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3101-3109, 2021. 6", + "[39] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2DPASS: 2D Priors" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Assisted Semantic Segmentation on LiDAR Point Clouds. In European Conference on Computer Vision, 2022. 1, 2, 3, 6, 7", + "[40] Jihan Yang, Shaoshuai Shi, Runyu Ding, Zhe Wang, and Xiaojuan Qi. Towards efficient 3d object detection with knowledge distillation. Advances in Neural Information Processing Systems, 35:21300-21313, 2022. 3", + "[41] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet: Towards a unified multi-task network for lidar perception. arXiv preprint arXiv:2209.09385, 2022. 7", + "[42] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE access, 8:58443-58469, 2020. 1", + "[43] Zhiwei Zhang, Zhizhong Zhang, Qian Yu, Ran Yi, Yuan Xie, and Lizhuang Ma. Lidar-camera panoptic segmentation via geometry-consistent and semantic-aware alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3662-3671, 2023. 2", + "[44] Wu Zheng, Li Jiang, Fanbin Lu, Yangyang Ye, and Chi-Wing Fu. Boosting single-frame 3d object detection by simulating multi-frame point clouds. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4848-4856, 2022. 3, 4", + "[45] Minghan Zhu, Shizhong Han, Hong Cai, Shubhankar Borse, Maani Ghaffari, and Fatih Porikli. 4d panoptic segmentation as invariant and equivariant field prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22488-22498, 2023. 2", + "[46] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and Asymmetrical 3D Convolution Networks for Lidar Segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 9939-9948, 2021. 1, 2, 5, 6, 7, 8", + "[47] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. In ICCV, 2021. 2" + ], + "bbox": [ + 501, + 92, + 893, + 614 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15320", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/745b72e4-2dd6-4811-ade0-f52413f14141_model.json b/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/745b72e4-2dd6-4811-ade0-f52413f14141_model.json new file mode 100644 index 0000000000000000000000000000000000000000..15819304b1aaa6710fcee369147ec647e1e5ee46 --- /dev/null +++ b/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/745b72e4-2dd6-4811-ade0-f52413f14141_model.json @@ -0,0 +1,2024 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.131, + 0.868, + 0.154 + ], + "angle": 0, + "content": "TASeg: Temporal Aggregation Network for LiDAR Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.18, + 0.794, + 0.217 + ], + "angle": 0, + "content": "Xiaopei Wu\\(^{1,2}\\), Yuenan Hou\\(^{2*}\\), Xiaoshui Huang\\(^{2*}\\), Binbin Lin\\(^{3,4}\\), Tong He\\(^{2}\\), Xinge Zhu\\(^{5}\\), Yuexin Ma\\(^{6}\\), Boxi Wu\\(^{4}\\), Haifeng Liu\\(^{1*}\\), Deng Cai\\(^{1}\\), Wanli Ouyang\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.218, + 0.766, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)State Key Lab of CAD&CG, Zhejiang University \\(^{2}\\)Shanghai AI Laboratory \\(^{3}\\)Fullong Inc." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.236, + 0.853, + 0.252 + ], + "angle": 0, + "content": "\\(^{4}\\)School of Software Technology, Zhejiang University \\(^{5}\\)The Chinese University of Hong Kong \\(^{6}\\)ShanghaiTech University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.474, + 0.789 + ], + "angle": 0, + "content": "Training deep models for LiDAR semantic segmentation is challenging due to the inherent sparsity of point clouds. Utilizing temporal data is a natural remedy against the sparsity problem as it makes the input signal denser. However, previous multi-frame fusion algorithms fall short in utilizing sufficient temporal information due to the memory constraint, and they also ignore the informative temporal images. To fully exploit rich information hidden in long-term temporal point clouds and images, we present the Temporal Aggregation Network, termed TAseg. Specifically, we propose a Temporal LiDAR Aggregation and Distillation (TLAD) algorithm, which leverages historical priors to assign different aggregation steps for different classes. It can largely reduce memory and time overhead while achieving higher accuracy. Besides, TLAD trains a teacher injected with gt priors to distill the model, further boosting the performance. To make full use of temporal images, we design a Temporal Image Aggregation and Fusion (TIAF) module, which can greatly expand the camera FOV and enhance the present features. Temporal LiDAR points in the camera FOV are used as mediums to transform temporal image features to the present coordinate for temporal multi-modal fusion. Moreover, we develop a Static-Moving Switch Augmentation (SMSA) algorithm, which utilizes sufficient temporal information to enable objects to switch their motion states freely, thus greatly increasing static and moving training samples. Our TAseg ranks \\( \\mathbf{1}^{st\\dagger} \\) on three challenging tracks, i.e., SemanticKITTI single-scan track, multi-scan track and nuScenes LiDAR segmentation track, strongly demonstrating the superiority of our method. Codes are available at https://github.com/LittlePey/TASeg." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.811, + 0.21, + 0.827 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.836, + 0.47, + 0.868 + ], + "angle": 0, + "content": "LiDAR segmentation aims to infer the semantic information of each point in point clouds and plays an indispensable role" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.288, + 0.895, + 0.501 + ], + "angle": 0, + "content": "in the autonomous driving [3, 21, 34, 42]. With the advent of deep learning, a large quantity of LiDAR segmentation algorithms [14, 19, 24, 39, 46] have been put forward. Despite their impressive results, the segmentation performance is still constrained by the inherent sparsity of point clouds. To alleviate this, it is desirable to aggregate temporal data. Previous multi-frame models [4, 9, 22] can only fuse a few LiDAR frames due to the GPU memory constraint, which restricts them from utilizing rich information hidden in the long-term temporal point clouds. Although MSeg3D [20] stacks dozens of LiDAR frames, the memory and time overhead are intolerable. Moreover, previous attempts concentrate on utilizing temporal point clouds, ignoring the valuable information hidden in temporal images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.896, + 0.777 + ], + "angle": 0, + "content": "To better leverage long-term temporal information, we propose a Temporal LiDAR Aggregation and Distillation (TLAD) algorithm. It can perform efficient multi-frame aggregation while achieving higher accuracy. TLAD consists of Flexible Step Aggregation (FSA) and Mask Distillation. FSA is based on the observation that the model needs different amounts of temporal points for different classes. For difficult classes, such as bicyclists, more points are needed to yield accurate predictions. However, for easy classes, such as cars and roads, it is unnecessary to aggregate too many point cloud frames for them. Therefore, we propose to assign different aggregation steps for different classes in point cloud sequences, which can significantly save memory and computation overhead on easy classes while providing sufficient temporal points for difficult classes. To identify classes of temporal point clouds, historical predictions are used, considering the temporal data is processed in chronological order, which is consistent with practical applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Interestingly, by assigning different classes with different steps, FSA actually injects historical priors into the aggregated point clouds. For classes with different steps, their patterns can be more discriminative due to their different densities. This makes the point cloud frames accumulated by FSA more conducive for segmentation. Experimental results show that FSA can not only reduce memory and time costs but also enhance overall performance. To further verify our" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.233, + 0.888 + ], + "angle": 0, + "content": "* Corresponding authors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.447, + 0.9 + ], + "angle": 0, + "content": "† On the date of CVPR deadline, i.e., 2023-11-18 07:59 AM UTC." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.447, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15311" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "idea, we directly replace historical predictions of temporal point clouds with their ground truth labels for FSA. We find the performance boosts greatly. This inspires us to use the model injected with gt priors to distill the model injected with historical priors, which we term Mask Distillation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.474, + 0.514 + ], + "angle": 0, + "content": "To fully use the informative temporal images, we devise a Temporal Image Aggregation and Fusion (TIAF) module. Previous multi-modal fusion methods usually suffer from limited image features due to the different FOVs (field of views) between the LiDAR and the camera. We observe that as the ego-vehicle moves forward, cameras can capture different FOVs. By aggregating images of different timestamps, the FOV of the present camera can be enlarged greatly. In addition, temporal images can provide different views for the same region at different timestamps, which can supply richer information. Our TIAF leverages temporal LiDAR points in the FOV of the corresponding temporal camera to transform historical image features to the present coordinate with the pose matrix. After aggregating all temporal image features, we use a series of 3D sparse convolutions to fuse them and convert them to voxel representation. Ultimately, we use temporal LiDAR points to gather temporal image features for temporal multi-modal fusion. The fused features embrace both temporal LiDAR and temporal image information, which contribute to more precise segmentation results. To our knowledge, we are the first to leverage temporal point clouds and temporal images simultaneously for LiDAR semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.471, + 0.682 + ], + "angle": 0, + "content": "Moreover, in the multi-scan task, the model needs to distinguish the motion states (moving or static) of a movable object, which is very challenging. In this paper, we make full use of temporal information to remarkably increase the multi-scan perception ability of the model. Specifically, we propose a data augmentation strategy named Static-Moving Switch Augmentation (SMSA). It manipulates temporal point clouds of a movable object to switch its motion state. In this way, we can increase the quantity of static and moving samples significantly, even in the absence of training samples of static or moving classes." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.682, + 0.471, + 0.697 + ], + "angle": 0, + "content": "In summary, our major contributions are listed as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.706, + 0.472, + 0.768 + ], + "angle": 0, + "content": "- We propose a Temporal LiDAR Aggregation and Distillation algorithm, which uses Flexible Step Aggregation and Mask Distillation techniques to largely reduce memory and time costs while achieving higher accuracy." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.773, + 0.471, + 0.835 + ], + "angle": 0, + "content": "- We devise a Temporal Image Aggregation and Fusion module, which exploits temporal images to enlarge the camera FOV and enhance present features. It also delivers a scheme for temporal multi-modal fusion." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.84, + 0.472, + 0.902 + ], + "angle": 0, + "content": "- We design a Static-Moving Switch Augmentation algorithm to enable static and moving objects to switch their motion states freely. With this technique, we can greatly increase static and moving training samples." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.706, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "- Our Temporal Aggregation Network, dubbed TASeg, achieves impressive results on SemanticKITTI and nuScenes benchmarks. Notably, our TASeg ranks \\(1^{st}\\) on three challenging tracks. Thorough ablation studies are provided to demonstrate the efficacy of our approach." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.185, + 0.642, + 0.202 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.212, + 0.896, + 0.44 + ], + "angle": 0, + "content": "LiDAR Semantic Segmentation. LiDAR semantic segmentation aims to assign a unique class label to each point in the input point cloud sequence. Recent years have witnessed an explosion of LiDAR segmentation algorithms [5, 7, 13, 15, 19, 20, 24, 27, 31, 39]. For example, [26] is the pioneering work that approximates the permutation-invariant function with a per-point and shared Multi-Layer Perceptron. [46] changes traditional cubic grids to cylindrical grids and designs a network of asymmetrical 3D convolutions. [19] divides the space with the radial window, which increases the receptive field smoothly and helps improve the performance. Despite their good segmentation performance, these methods still take the single LiDAR frame as input, which does not utilize the rich semantic and geometric information hidden in temporal data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.451, + 0.897, + 0.679 + ], + "angle": 0, + "content": "Multi-Frame LiDAR Perception. Compared to a single LiDAR scan, multiple LiDAR scans can provide more sufficient information. Recently, many research efforts have been put on temporal LiDAR segmentation [1, 2, 6, 8, 9, 17, 20, 22, 30, 45]. For example, [8] leverages a Bayes filter to explore the temporal consistency. [4] shunts the historical points into two groups to utilize historical frames efficiently. Despite the success of previous multi-frame methods, they can not leverage the valuable information hidden in long-term temporal point clouds due to the GPU memory constraint. Although [20] stacks dozens of LiDAR scans, the memory and time overhead are intolerable. In this paper, we present an efficient multi-frame aggregation algorithm, which can greatly save memory and computation consumption while achieving higher performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Multi-Modal Fusion. Since LiDAR and camera are two complementary sensors for 3D semantic segmentation, multimodal fusion has gained increasing attention in recent years [10, 18, 24, 43, 47]. However, these multi-modal fusion methods usually suffer from limited overlapped regions between the LiDAR and camera due to their different FOVs. [39] proposes a cross-modal knowledge distillation method, which is free from images at inference, while it causes much loss of the RGB information. [20] completes the missing camera features using predicted pseudo-camera features, while the used image information is still restricted to the present camera FOV. In this paper, we fully take advantage of temporal images to enlarge the FOV of the present camera and enhance present image features." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "15312" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.289 + ], + "angle": 0, + "content": "Knowledge Distillation for LiDAR Perception. Knowledge distillation [12] is widely used in various fields in that it can improve the performance of the student without sacrificing inference efficiency. In LiDAR perception, [13, 40] compress a cumbersome teacher to a lightweight student to reinforce the representation learning of the student as well as maintain high efficiency. [39] distills the prior of 2D images to 3D point clouds with well-designed cross-modal knowledge distillation module. [32, 44] utilizes a multi-frame teacher to help a single-frame student learn dense 3D features. By contrast, our method aims to transfer knowledge from a multi-frame teacher injected with gt priors to a multi-frame student injected with historical priors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.301, + 0.213, + 0.318 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.326, + 0.47, + 0.342 + ], + "angle": 0, + "content": "3.1. Temporal LiDAR Aggregation and Distillation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.47, + 0.395 + ], + "angle": 0, + "content": "A simple solution to utilize the temporal information of a consecutive point cloud sequence is to concatenate all points of the sequence, as shown in the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.404, + 0.47, + 0.428 + ], + "angle": 0, + "content": "\\[\nX _ {t} = \\operatorname {c o n c a t} \\left(P _ {t}, T _ {t - 1} P _ {t - 1}, \\dots , T _ {t - \\Delta t} P _ {t - \\Delta t}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.424, + 0.359, + 0.44 + ], + "angle": 0, + "content": "\\[\nY _ {t} = \\operatorname {c o n c a t} \\left(L _ {t}, L _ {t - 1}, \\dots , L _ {t - \\Delta t}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.473, + 0.618 + ], + "angle": 0, + "content": "where \\( P_{t - i} \\) and \\( L_{t - i} \\) denote the \\( (t - i)^{th} \\) point cloud frame and the corresponding point-wise label. \\( T_{t - i} \\) is the transformation matrix that transforms the coordinate from the \\( (t - i)^{th} \\) frame to the \\( t^{th} \\) frame. \\( \\Delta t \\) is the window size of temporal point clouds and \\( \\text{concat}(.) \\) denotes the concatenation operation. \\( X_{t} \\) and \\( Y_{t} \\) are the aggregated LiDAR frame and point-wise label. Although simple, there are some problems. On the one hand, direct concatenation consumes much GPU memory. On the other hand, the huge memory cost constrains the multi-frame model from utilizing more historical frames, thereby limiting the ultimate performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.312, + 0.652 + ], + "angle": 0, + "content": "3.1.1 Flexible Step Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.902 + ], + "angle": 0, + "content": "To reduce the GPU memory consumption, we can sample temporal frames with a step, while this method is also suboptimal. A small step introduces huge memory overhead, whilst a large step cannot utilize sufficient temporal information. To this end, we propose Flexible Step Aggregation (FSA). Our method is based on the observation that for different classes, the model needs different amounts of temporal information. For difficult classes, such as bicycles, more points are needed to yield accurate predictions. For easy classes, such as cars and buildings, it is unnecessary to aggregate many point cloud frames for them. Based on the above analysis, we propose to assign different steps for different classes according to their learning difficulty. Specifically, we leverage historical predictions to divide the temporal point clouds into several class groups and assign a specific sampling step for each group. Then, we aggregate temporal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.137 + ], + "angle": 0, + "content": "points for each group with the corresponding step. Eventually, temporal points of all groups are concatenated with the current frame, resulting in an aggregated frame." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.895, + 0.304 + ], + "angle": 0, + "content": "The group division is not strict as long as it follows the principle that more difficult classes need smaller steps. For example, we can simply divide all classes into three groups according to their segmentation performance, such as [0, 80) mIoU, [80, 90) mIoU and [90, 100] mIoU. Then, we assign the three groups with a step of 2, 4 and \\(\\infty\\) (\\(\\infty\\) means we do not aggregate temporal points for the group). To save more memory and computation without sacrificing performance, a more fine-grained division can be used. Formally, suppose we divide temporal point clouds into \\(g\\) groups. We aggregate temporal points for the \\(k^{th}\\) group as below:" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.31, + 0.894, + 0.337 + ], + "angle": 0, + "content": "\\[\nX _ {t} ^ {k} = \\operatorname {c o n c a t} \\left(T _ {o _ {1}} P _ {o _ {1}} M _ {o _ {1}} ^ {k}, \\dots , T _ {o _ {n}} P _ {o _ {n}} M _ {o _ {n}} ^ {k}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.332, + 0.8, + 0.351 + ], + "angle": 0, + "content": "\\[\nY _ {t} ^ {k} = \\operatorname {c o n c a t} \\left(L _ {o _ {1}} M _ {o _ {1}} ^ {k}, \\dots , L _ {o _ {n}} M _ {o _ {n}} ^ {k}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.895, + 0.449 + ], + "angle": 0, + "content": "Here \\(o = \\{o_i|o_i = t - i\\times s_k,i = 1,2,\\dots,n,n = \\lfloor \\Delta t / s_k\\rfloor \\}\\) \\(s_k\\) is the sampling step for the \\(k^{th}\\) class group and \\(\\lfloor .\\rfloor\\) is the floor operation. \\(M_{o_i}^k\\) is the group mask that indicates which point of \\(P_{o_i}\\) belongs to the \\(k^{th}\\) class group. It can be obtained from historical predictions. Finally, we concatenate points and labels of all groups with the current frame:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.456, + 0.894, + 0.482 + ], + "angle": 0, + "content": "\\[\nX _ {t} = \\operatorname {c o n c a t} \\left(P _ {t}, X _ {t} ^ {1}, \\dots , X _ {t} ^ {k}, \\dots , X _ {t} ^ {g}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.478, + 0.827, + 0.495 + ], + "angle": 0, + "content": "\\[\nY _ {t} = \\operatorname {c o n c a t} \\left(L _ {t}, Y _ {t} ^ {- 1}, \\dots , X _ {t} ^ {k}, \\dots , Y _ {t} ^ {g}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.895, + 0.684 + ], + "angle": 0, + "content": "In this way, we discard massive redundant temporal points while maintaining essential temporal information. Since easy classes are usually large and hold a large quantity of points, the GPU memory overhead can be further reduced. Moreover, by assigning different classes with different steps, FSA actually injects historical priors into the aggregated point clouds. For classes with different steps, their patterns can be more discriminative due to their different densities (steps), which makes it easier to segment the multi-frame point clouds aggregated by FSA. Experiments in Table 4 verify that the proposed FSA can not only save memory and time costs but also achieve better performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.674, + 0.71 + ], + "angle": 0, + "content": "3.1.2 Mask Distillation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In FSA, we use historical predictions to generate group masks, which we call pseudo masks. Since ground truth labels are more accurate than historical predictions, a natural question arises: what if we use gt masks (the group masks generated by ground truth labels) for FSA? Our experiment shows that the performance can be improved greatly. Actually, using gt masks for FSA can produce more discriminative patterns for classes with different steps. This motivates us to simulate features of a model trained on temporal point clouds aggregated with gt masks. In this way, our model is guided to learn more discriminative features to distinguish different classes. We call this Mask Distillation. Specifically," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15313" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.093, + 0.885, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.439, + 0.896, + 0.511 + ], + "angle": 0, + "content": "Figure 1. Overview of our Temporal Aggregation Network (TASeg). (1) Temporal LiDAR Aggregation and Distillation leverages the proposed Flexible Step Aggregation (FSA) to assign different temporal steps for different classes, and it utilizes a teacher injected with gt priors for knowledge distillation. (2) Temporal Image Aggregation and Fusion takes temporal LiDAR points as mediums to transform historical image features to the present coordinate. 3D sparse convolutions are employed to fuse temporal image features. Finally, we use temporal LiDAR points to gather voxel-wise temporal image features for temporal multi-modal fusion." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.518, + 0.475, + 0.551 + ], + "angle": 0, + "content": "we formulate the distillation on the feature map of student \\(F^{s}\\in \\mathbb{R}^{N^{s}\\times C}\\) and the feature map of teacher \\(F^{t}\\in \\mathbb{R}^{N^{t}\\times C}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.562, + 0.471, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {K D}} = \\mathbb {E} \\left[ \\left\\| F ^ {s} m ^ {s} - F ^ {t} m ^ {t} \\right\\| _ {2} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.472, + 0.675 + ], + "angle": 0, + "content": "Since temporal point clouds aggregated with pseudo masks and gt masks are different, we use masks \\( m^s \\) and \\( m^t \\) to select voxels that appeared in both \\( F^s \\) and \\( F^t \\). Note that our Mask Distillation is distinct from the methods that distill a multi-frame model to a single-frame model [32, 44]. Both the student and teacher in our method are multi-frame models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.431, + 0.699 + ], + "angle": 0, + "content": "3.2. Temporal Image Aggregation and Fusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.825 + ], + "angle": 0, + "content": "Previous multi-modal fusion methods only focus on leveraging present images while ignoring the precious value of temporal images. Temporal images can provide broader camera FOVs and richer information. Besides, they can enable more robust multi-modal fusion under the malfunction condition on some cameras. In this section, we provide an effective solution for aggregating temporal image features and performing temporal multi-modal fusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Temporal Aggregation and Fusion. Since temporal images are in different feature spaces, it is difficult to establish the relationship between different images for feature aggregation. In our method, we take temporal LiDAR points as mediums to transform temporal image features to the present" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.895, + 0.701 + ], + "angle": 0, + "content": "coordinate with the pose information. This way, temporal image features are unified to the present 3D space. Specifically, given an image \\(I_{t - \\Delta t}\\in \\mathbb{R}^{H\\times W\\times 3}\\) and point cloud \\(P_{t - \\Delta t}\\in \\mathbb{R}^{N\\times 3}\\), we use an image network to extract the image feature \\(Z_{t - \\Delta t}\\in \\mathbb{R}^{H\\times W\\times C}\\). According to the sensor calibration, we can establish the pixel-to-point mapping between 2D pixels and 3D points. Hence, we can project image feature \\(Z_{t - \\Delta t}\\) to 3D space, resulting in point-wise image feature \\(Q_{t - \\Delta t}\\in \\mathbb{R}^{M\\times C}\\), where \\(M\\) is the number of LiDAR points located on \\(Z_{t - \\Delta t}\\). By transforming \\(Q_{t - \\Delta t}\\) to the present coordinate with the pose matrix, we realize the aggregation of temporal image features:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.706, + 0.895, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\bar {X} _ {t} = \\operatorname {c o n c a t} \\left(Q _ {t}, T _ {o _ {1}} Q _ {o _ {1}}, \\dots , T _ {o _ {n}} Q _ {o _ {n}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.897, + 0.903 + ], + "angle": 0, + "content": "where \\(o = \\{o_i|o_i = t - i\\times s,i = 1,2,\\dots,\\lfloor \\Delta t / s\\rfloor \\}\\) and \\(s\\) is the sampling step for temporal images. With temporal images aggregated, the FOV of the present camera is expanded greatly. Moreover, temporal image feature fusion becomes convenient because they are unified to the same 3D space. Concretely, we can use several 3D sparse convolutions to fuse aggregated temporal image features, which also endows them with geometric information, as shown in Equation 6. In addition, feature map downsampling is also utilized to generate multi-scale voxel features, providing richer information for subsequent temporal multi-modal fusion." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15314" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.138, + 0.105, + 0.471, + 0.124 + ], + "angle": 0, + "content": "\\[\n\\bar {V} _ {t} = \\operatorname {S p a r s e C o n v} _ {\\mathrm {3 D}} \\left(\\operatorname {V o x e l i z a t i o n} \\left(\\bar {X} _ {t}\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.13, + 0.473, + 0.342 + ], + "angle": 0, + "content": "Temporal Multi-Modal Fusion. Benefiting from the temporal image aggregation, associating temporal image features with temporal LiDAR points also becomes convenient. Specifically, given temporal image features converted to unified voxel representation \\(\\overline{V}_t\\), we can establish a point-to-voxel association between temporal LiDAR points and \\(\\overline{V}_t\\). For each temporal LiDAR point, we generate its image feature by pooling its nearby voxel-wise temporal image features with trilinear interpolation instead of hard indexing. To extract richer image features, we perform interpolation on multi-scale feature maps. Finally, we concatenate the point cloud features and aggregated multi-scale image features, resulting in fused features, which convey powerful information of both temporal point clouds and temporal images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.471, + 0.44 + ], + "angle": 0, + "content": "2D and 3D Supervision. To make the extracted image features more informative, we add 2D supervision and 3D supervision on the 2D backbone and 3D convolutions in the image branch, respectively. The 3D supervision is just the label of point clouds. The 2D supervision is obtained by projecting labels of point clouds to the image plane." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.448, + 0.393, + 0.465 + ], + "angle": 0, + "content": "3.3. Static-Moving Switch Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.472, + 0.683 + ], + "angle": 0, + "content": "In the multi-scan task, the model is required to distinguish the motion state of movable objects. To enable the model to explore a large data space, we design an effective data augmentation, Static-Moving Switch Augmentation (SMSA). SMSA enables a movable object to switch its motion state freely, which can remarkably increase the sample quantity of static and moving objects. Concretely, considering that a unique object \\( b \\) has the same instance id in all frames of a sequence, we can use its instance mask to crop its temporal point clouds, which is denoted as \\( \\mathcal{B} = \\{\\mathcal{B}_{o_i}|i = 0,1,\\dots,\\lfloor \\Delta t / s\\rfloor ,\\mathcal{B}_{o_i}\\in \\mathbb{R}^{N\\times 3}\\} \\). Here \\( o = \\{o_i|o_i = t - i\\times s,i = 1,2,\\dots,\\lfloor \\Delta t / s\\rfloor \\} \\), \\( s \\) is the sampling step and \\( \\mathcal{B}_{o_i} \\) is the temporal part of \\( b \\) at \\( t_{o_i} \\). By manipulating \\( \\mathcal{B} \\), we can change the motion state of \\( b \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Static to Moving. If \\( b \\) is static, all temporal parts of \\( b \\) locate at the same position, as shown in the upper-left of Figure 2. To change \\( b \\) to a moving object, we can shift each temporal part of \\( b \\) with an offset. Considering that objects typically move at a constant speed within a short time, we set the offset between adjacent temporal parts to be the same. For offset itself, it is a random value to increase the diversity of resulting moving samples. As for the direction of the offset, it can be roughly estimated by comparing the width and height of the minimum bounding box of \\( b \\). Since static objects often park on the side of the road, which is crowded, the shifted temporal parts of static objects may overlap with other objects. To alleviate this, we define a set of anchor points around \\( b \\) and a coverage area for each anchor point." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.096, + 0.895, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.261, + 0.895, + 0.304 + ], + "angle": 0, + "content": "Figure 2. Visualization of the augmented samples by our Static-Moving Switch Augmentation (SMSA). Our SMSA switches the motion state by manipulating the temporal parts of objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.892, + 0.357 + ], + "angle": 0, + "content": "Then, we shift all temporal parts of \\( b \\) to the anchor point whose coverage area contains the fewest LiDAR points (refer to the supplementary material for more details)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.895, + 0.512 + ], + "angle": 0, + "content": "Moving to Static. If \\( b \\) is moving, different temporal parts of \\( b \\) locate at different positions. To switch its motion state to static, we can shift all its temporal parts to the same position. In particular, we calculate the centers of each temporal part of \\( b \\), which are denoted as \\( \\mathcal{C} = \\{\\mathcal{C}_{o_i} | \\mathcal{C}_{o_i} \\in \\mathbb{R}^3\\} \\). Considering the trajectory of moving objects in a short time is approximate to a line and the speed is a constant, we simply use \\( q = \\mathcal{C}_{o_0} - \\mathcal{C}_{o_1} \\) as the offset of adjacent temporal parts of \\( b \\). Eventually, we can obtain a static object by shifting all temporal parts of \\( b \\) to \\( C_{o_0} \\) (see Figure 2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.895, + 0.56 + ], + "angle": 0, + "content": "It should be noted that our SMSA is an online and plug-and-play data augmentation strategy, which consumes negligible storage and computation costs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.568, + 0.682, + 0.584 + ], + "angle": 0, + "content": "3.4. Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.895, + 0.637 + ], + "angle": 0, + "content": "The final training objective is comprised of LiDAR segmentation loss, mask distillation loss, fusion segmentation loss, 2D and 3D supervision loss on the image branch:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.64, + 0.892, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {L i D A R}} + \\alpha \\mathcal {L} _ {\\mathrm {K D}} + \\beta \\mathcal {L} _ {\\text {F u s i o n}} + \\gamma (\\mathcal {L} _ {\\mathrm {2 D}} + \\mathcal {L} _ {\\mathrm {3 D}}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.666, + 0.894, + 0.697 + ], + "angle": 0, + "content": "where \\(\\alpha, \\beta, \\gamma\\) are the coefficients to control the effect of different losses. We set \\(\\alpha = 1, \\beta = 1, \\gamma = 1\\) by default." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.634, + 0.727 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Datasets & Evaluation Metrics. Following [13, 46], we evaluate the performance on two popular LiDAR segmentation benchmarks, i.e., SemanticKITTI and nuScenes. SemanticKITTI contains 22 point cloud sequences, where sequences 0-10, 8, 11-21 are selected as training, validation and testing, respectively. As for nuScenes, it collects 1, 000 driving scenes, where 850 scenes are chosen for training and validation, and the remaining 150 scenes are used for testing. We conduct experiments on three tracks, i.e., SemanticKITTI single-scan and multi-scan semantic segmentation and nuScenes semantic segmentation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15315" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.891, + 0.378 + ], + "angle": 0, + "content": "
MethodmIoUcarbicyclemotorcycletruckother-veh.personbicyclistmotorcyclistroadparkingsidewalkother-gro.buildingfencevegetationtrunkterrainpoletraffic sign
SqueezeSegV2 [33]39.781.818.517.913.414.020.125.13.988.645.867.617.773.741.171.835.860.220.226.3
RangeNet53++ [25]52.291.425.734.425.723.038.338.84.891.865.075.227.887.458.680.555.164.647.955.9
RandLA-Net [14]55.994.229.832.243.939.148.447.49.490.561.874.024.589.760.483.863.668.651.050.7
SqueezeSegV3 [36]55.992.538.736.529.633.045.646.220.191.763.474.826.489.059.482.058.765.449.658.9
JS3C-Net [38]66.095.859.352.954.346.069.565.439.988.961.972.131.992.570.884.569.867.960.768.7
SPVNAS [31]67.097.250.650.456.658.067.467.150.390.267.675.421.891.666.986.173.471.064.367.3
Cylinder3D [46]68.997.167.663.850.858.573.769.248.092.265.077.032.390.766.585.672.569.862.466.2
RPVNet [37]70.397.668.468.744.261.175.974.443.493.470.380.733.393.572.186.575.171.764.861.4
(AF)2-S3Net [5]70.894.363.081.440.240.076.481.777.792.066.876.245.892.569.678.668.063.164.073.3
PVKD [13]71.297.069.353.567.960.275.173.550.591.877.570.941.092.469.486.573.871.964.965.8
2DPASS [39]72.997.063.663.461.161.577.981.374.189.767.474.740.093.572.986.273.971.065.070.4
SphereFormer [19]74.897.570.170.559.667.779.080.475.391.869.778.241.393.872.886.775.172.466.872.9
UniSeg [24]75.297.971.975.263.674.178.974.860.692.674.079.546.193.472.787.576.373.168.368.5
TASeg (Ours)76.597.771.871.465.278.779.984.678.691.674.078.039.393.573.486.675.071.769.673.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.086, + 0.387, + 0.882, + 0.401 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-arts on SemanticKITTI test set (single-scan). All results can be found on the online leaderboard." + }, + { + "type": "table", + "bbox": [ + 0.2, + 0.408, + 0.773, + 0.567 + ], + "angle": 0, + "content": "
MethodmIoUcar-scar-mtruc.-struc.-mothe.-sothe.-mpers.-spers.-mbicy.-sbicy.-mmotor.-smotor.-m
LatticeNet [29]45.291.154.829.73.523.10.66.849.90.044.60.064.3
TemporalLidarSeg [9]47.092.168.239.22.135.012.414.440.40.042.80.012.9
\\( (AF)^2-S3Net[5] \\)56.991.865.315.75.627.53.916.467.615.166.467.159.6
MarS3D [22]52.795.178.439.75.136.610.016.258.01.267.30.036.3
SVQNet [4]60.596.080.141.05.160.47.128.785.10.072.40.088.1
2DPASS [39]62.496.282.148.216.152.73.835.480.37.971.262.073.1
TASeg (Ours)65.794.872.853.635.271.453.236.780.432.475.212.389.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.575, + 0.888, + 0.59 + ], + "angle": 0, + "content": "Table 2. Comparison to the state-of-the-art methods on SemanticKITTI test set (multi-scan). -s indicates static and -m stands for moving." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.602, + 0.475, + 0.83 + ], + "angle": 0, + "content": "Implementation Details. We use the MinkowskiNet [7] re-implemented by PCSeg [23] as our baseline. Our TASeg is trained with the SGD [28] optimizer on 4 A100 GPUs with batch size 6 for 12 and 36 epochs on SemanticKITTI and nuScenes datasets. The learning rate and weight decay are set to 0.02 and 0.0001. The window size of temporal point clouds is set to 16. Due to a large redundancy of images, we use a step of 12 and a window size of 48 for temporal images. Our data augmentation strategy includes random flipping, rotation, scaling, transformation, LaserMix [16] and PolarMix [35]. During the inference, the LiDAR semantic predictions and image features of the previous timestamp can be saved, so there is no time cost for generating historical predictions and processing temporal image aggregation. More details are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.442, + 0.862 + ], + "angle": 0, + "content": "4.1. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We compare our TASeg with state-of-the-art LiDAR segmentation methods on SemanticKITTI single-scan and multi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.895, + 0.799 + ], + "angle": 0, + "content": "scan track, and nuScenes LiDAR semantic segmentation track, as summarized in Table 1, 2 and 3. On the SemanticKITTI single-scan track, our TASeg is \\(1.3\\mathrm{mIoU}\\) higher than UniSeg [24], the best-performing published algorithm on the SemanticKITTI single-scan leaderboard. On the SemanticKITTI multi-scan track, our approach outperforms the best multi-scan method 2DPASS [39] by \\(3.3\\mathrm{mIoU}\\). For nuScenes LiDAR semantic segmentation, we also achieve superior performance over previous algorithms. As shown in Table 3, the proposed approach surpasses the UniSeg [24] by \\(1.1\\mathrm{mIoU}\\). In particular, our approach holds the highest entry on all three tracks. These encouraging results strongly show the effectiveness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.813, + 0.663, + 0.829 + ], + "angle": 0, + "content": "4.2. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We examine the effect of each component through detailed ablations. Unless otherwise specified, the following experiments are trained on the training set of the SemanticKITTI single-scan task and evaluated on the validation set." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15316" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.089, + 0.852, + 0.309 + ], + "angle": 0, + "content": "
MethodmIoUconstruction
barrierbicyclebuscarconstructionmotorcyclepedestriantraffic-conetrailertruckdriveableother-flatsidewalkterrainmannadevegetation
Cylinder3D [46]77.282.829.884.389.463.079.377.273.484.669.197.770.280.375.590.487.6
SPVCNN [31]77.480.030.091.990.864.779.075.670.981.074.697.469.280.076.189.387.1
2D3DNet [11]80.083.059.488.085.163.784.482.076.084.871.996.967.479.876.092.189.2
2DPASS [39]80.881.755.392.091.873.386.578.572.584.775.597.669.179.975.590.288.0
LidarMultiNet [41]81.480.448.494.390.071.587.285.280.486.974.897.867.380.776.592.189.6
MSeg3D [20]81.183.142.594.992.067.178.685.780.587.577.397.769.881.277.892.490.1
SphereFormer [19]81.983.339.294.792.577.584.284.479.188.478.397.969.081.577.293.490.2
UniSeg [24]83.585.971.292.191.680.588.080.976.086.376.797.771.880.776.791.388.8
TASeg (Ours)84.687.169.490.592.278.790.486.381.988.375.997.870.981.078.293.491.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.235, + 0.317, + 0.735, + 0.332 + ], + "angle": 0, + "content": "Table 3. Performance comparison with start-of-the-art methods on nuScenes test set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.474, + 0.587 + ], + "angle": 0, + "content": "Effect of Flexible Step Aggregation. In Table 4, we compare our FSA with other multi-frame algorithms. MSeg3D [20] utilizes temporal point clouds by directly concatenating all points, which introduces a large memory consumption and achieves limited improvement. SVQNet [4] leverages cross-attention to fuse temporal features while it can only handle fewer LiDAR scans. When increasing LiDAR scans, the memory overhead is also huge. Our FSA assigns different classes with different temporal steps, which can leverage long-term temporal information with less memory and time costs. Moreover, thanks to the utilization of historical priors, our FSA achieves better results (71.3 mIoU) than directly concatenating (69.9 mIoU), given the same temporal window size. To explore the effect of different window sizes on FSA, we present Table 5. Results show that the performance is saturated with a window size of 24." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.474, + 0.903 + ], + "angle": 0, + "content": "Effect of Group Division. To investigate the effect of different group divisions for FSA, we provide Table 6. For division1, we simply divide all classes into three groups according to their performance on the val set. The group1 consists of the classes in [90, 100) mIoU, such as cars and roads. The group2 consists of classes in [80, 90) mIoU, such as motorcycles. The group3 contains the remaining classes. We assign group1, group2 and group3 with the step of \\(\\infty\\), 4 and 2. A step of \\(\\infty\\) means we do not aggregate temporal points for classes in the group due to their near-saturate performances. The result shows that division1 can reduce the time and memory costs largely compared to directly stacking (the \\(2^{st}\\) row in Table 4). We can also use performance top1-6, top6-12, and top12-19 for division, resulting in division2. The group division is robust as long as it follows the principle that more difficult classes need smaller steps. To further reduce the memory consumption, we can finetune the group division, such as moving large-size classes (e.g. other-ground and terrain) in group3 to group2 (division3) or moving large-size classes in group3 and group2 to a new" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.344, + 0.878, + 0.46 + ], + "angle": 0, + "content": "
MethodLatencyMemorymIoU
Baseline63ms5.0G68.9
+ MSeg3D [20]†284ms69.7G69.9
+ SVQNet [4]†121ms15.4G69.5
+ FSA79ms8.5G71.3
+ FSA w/ gt mask79ms8.5G75.9
+ FSA + M.D.79ms8.5G71.8
+ FSA + M.D. × 0.7572ms6.9G71.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.464, + 0.896, + 0.507 + ], + "angle": 0, + "content": "Table 4. Comparison with different multi-frame algorithms. \\(\\dagger\\) represents our re-implementation. \"M.D.\" represents Mask Distillation. \"×75\" represents that we reduce the model parameters to \\(75\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.515, + 0.878, + 0.551 + ], + "angle": 0, + "content": "
Window size481216202428
mIoU69.970.470.971.371.471.671.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.537, + 0.556, + 0.855, + 0.57 + ], + "angle": 0, + "content": "Table 5. Ablation on different window sizes for FSA." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.578, + 0.869, + 0.668 + ], + "angle": 0, + "content": "
MethodLatencyMemorymIoU
FSA w/ division192ms14.1G70.8
FSA w/ division283ms7.2G70.6
FSA w/ division379ms8.5G71.3
FSA w/ division474ms7.7G71.2
FSA w/ division574ms7.4G71.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.53, + 0.674, + 0.86, + 0.688 + ], + "angle": 0, + "content": "Table 6. Ablation on different group divisions for FSA." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.895, + 0.773 + ], + "angle": 0, + "content": "group with step of 8 (division4). Besides, we can divide each class group into a close and a distant group according to the distance (such as \\(30\\mathrm{m}\\)). Considering close areas need fewer temporal points, we can use twice the original step for temporal aggregation of close groups (division5)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Effect of Mask Distillation. At the \\(5^{th}\\) row of Table 4, we use gt masks of historical frames for FSA. The performance is greatly boosted. Considering we cannot get ground truth at inference, we use it as a teacher to distill the model trained with pseudo masks. The result shows that the student is improved by \\(0.5\\mathrm{mIoU}\\) after distillation. With \\(75\\%\\) model complexity, we can further reduce the time and memory cost and still achieve a leading performance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15317" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "Effect of Temporal Image Aggregation and Fusion. TIAF delivers a temporal multi-modal fusion scheme to make full use of temporal images. As shown in Table 7, with zero image, only TLAD is used, achieving \\(71.8\\mathrm{mIoU}\\). With one image, only present images are utilized. Due to the limited FOV of the present camera, only part of the LiDAR points can gather image features, which limits the multi-modal fusion. Our TIAF leverages temporal images to enlarge the FOV of the camera and enhance present features. The results show that with the number of images increasing, the performance rises gradually, which verifies the effectiveness of our TIAF. When using seven temporal images, we achieve \\(1.0\\mathrm{mIoU}\\) improvement on the strong baseline. Note that there is no extra cost to generate historical image features because we can save them at the past moments." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.326, + 0.455, + 0.363 + ], + "angle": 0, + "content": "
Num of images01357
mIoU71.872.172.472.772.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.087, + 0.367, + 0.46, + 0.382 + ], + "angle": 0, + "content": "Table 7. Ablation on the number of temporal images for TIAF." + }, + { + "type": "table", + "bbox": [ + 0.143, + 0.39, + 0.407, + 0.469 + ], + "angle": 0, + "content": "
Exp2D Sup.3D Sup.Multi-ScalemIoU
(a)✓✓72.4
(b)✓✓72.5
(c)✓✓72.4
(d)✓✓✓72.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.478, + 0.468, + 0.492 + ], + "angle": 0, + "content": "Table 8. Effect of different supervisions and multi-scale for TIAF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.47, + 0.663 + ], + "angle": 0, + "content": "We also provide the ablation on different supervisions and multi-scale features for TIAF in Table 8. In TIAF, we leverage 2D and 3D supervision to guide the extracted image features to be more conducive for segmentation. Multi-scale is also used to provide more discriminative features. The result shows that each of the designs is beneficial for final performance. Note that our temporal multi-modal method is orthogonal to other single-frame multi-modal methods. We use a simple pixel-to-point mapping at the feature projection stage. More complex methods [20, 24] can also be utilized, but it is not the focus of our TIAF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Effect of Static-Moving Switch Augmentation. To verify our SMSA, we present Table 9. From the results, we can find that without SMSA, the accuracy of the model on many classes is lower than \\(10\\mathrm{IoU}\\). After using SMSA, the performance of moving truck, moving other-vehicle and static bicyclist classes is improved by more than \\(20\\mathrm{IoU}\\), and the overall performance boosts from \\(61.3\\mathrm{mIoU}\\) to \\(65.7\\mathrm{mIoU}\\), which strongly demonstrates the efficacy of our SMSA." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.799, + 0.468, + 0.869 + ], + "angle": 0, + "content": "
SMSAmloUcar-scar-mtrue-strue-mother-sother-mpers-spers-mbicy-sbicy-mmotor-smotor-m
✓61.395.679.349.29.161.05.739.087.00.072.64.376.4
65.794.872.853.635.271.453.236.780.432.475.212.389.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.873, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Table 9. Ablation study of our SMSA on the SemanticKITTI test benchmark of the multi-scan track." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.183 + ], + "angle": 0, + "content": "Generalization to More Architectures. In addition to MinkowskiNet, we also validate our temporal aggregation method on Cylinder3D [46] and SPVNAS [31]. Detailed results are summarized in Table 10. Our TASeg brings impressive improvements to both baselines, which demonstrates the strong generalization ability of our method." + }, + { + "type": "table", + "bbox": [ + 0.567, + 0.194, + 0.828, + 0.275 + ], + "angle": 0, + "content": "
MethodmIoU
Cylinder3D [46]†66.0
Cylinder3D [46]†+ TASeg69.6
SPVNAS [31]†68.6
SPVNAS [31]†+ TASeg72.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.283, + 0.894, + 0.311 + ], + "angle": 0, + "content": "Table 10. Performance of applying our TASeg to different baselines. \\(\\dagger\\) represents our re-implementation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.895, + 0.438 + ], + "angle": 0, + "content": "Comparison on Complexity, Latency and Accuracy. As shown in Table 11, when not utilizing temporal images, our TASeg achieves superior performance than other methods with comparable or less complexity and latency. With temporal images, TASeg can maintain lower complexity and latency than UniSeg [24] while achieving higher accuracy. Besides, with \\(75\\%\\) parameters, the latency of TASeg can be further reduced, and the accuracy is not affected much." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.446, + 0.887, + 0.566 + ], + "angle": 0, + "content": "
MethodInputParamsLatencymIoU
MinkowskiNet [7]†L37.9M63ms68.4
Cylinder3D [46]†L55.9M75ms66.0
SPVNAS [31]†L96.7M105ms68.6
RPVNet [37]†L119.0M110ms68.9
UniSeg [24]L+C147.6M145ms71.3
TASeg wo/ TIAFL+T37.9M79ms71.8
TASegL+C+T46.7M116ms72.7
TASeg × 0.75L+C+T27.4M108ms72.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.574, + 0.894, + 0.63 + ], + "angle": 0, + "content": "Table 11. Comparison of the model complexity, latency and accuracy on SemanticKITTI val set. \\(\\dagger\\) represents our re-implementation on NVIDIA A100 for a fair comparison. \"L\", \"C\" and \"T\" represent LiDAR, Camera and Temporal data, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.644, + 0.619, + 0.66 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.667, + 0.894, + 0.757 + ], + "angle": 0, + "content": "In this paper, we present TASeg to fully exploit temporal point clouds and temporal images. Our TASeg is comprised of TLAD, TIAF and SMSA. We perform extensive experiments to validate the efficacy of each component. Moreover, our TASeg set a new state-of-the-art on three challenging tracks, demonstrating the superiority of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.688, + 0.788 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by The National Nature Science Foundation of China (Grant Nos: 62273301, 62273302, 62036009, 61936006, 62273303), in part by Yongjiang Talent Introduction Programme (Grant No: 2022A-240-G, 2023A-197-G), in part by the National Key R&D Program of China (NO.2022ZD0160100). This work was done during the first author's internship at Shanghai AI Laboratory." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15318" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.471, + 0.184 + ], + "angle": 0, + "content": "[1] Mehmet Aygun, Aljosa Osep, Mark Weber, Maxim Maximov, Cyrill Stachniss, Jens Behley, and Laura Leal-Taixe. 4d panoptic lidar segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5527-5537, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Hanwen Cao, Yongyi Lu, Cewu Lu, Bo Pang, Gongshen Liu, and Alan Yuille. Asap-net: Attention and structure aware point cloud sequence segmentation. arXiv preprint arXiv:2008.05149, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.472, + 0.297 + ], + "angle": 0, + "content": "[3] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. arXiv preprint arXiv:2306.16927, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.472, + 0.38 + ], + "angle": 0, + "content": "[4] Xuechao Chen, Shuangjie Xu, Xiaoyi Zou, Tongyi Cao, Dit-Yan Yeung, and Lu Fang. Svqnet: Sparse voxel-adjacent query network for 4d spatio-temporal lidar semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8569–8578, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.472, + 0.451 + ], + "angle": 0, + "content": "[5] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. (af)2-S3Net: Attentive Feature Fusion with Adaptive Feature Selection for Sparse Semantic Segmentation Network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12547-12556, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.472, + 0.493 + ], + "angle": 0, + "content": "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.472, + 0.55 + ], + "angle": 0, + "content": "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.472, + 0.606 + ], + "angle": 0, + "content": "[8] Ayush Dewan and Wolfram Burgard. Deeptemporalseg: Temporally consistent semantic segmentation of 3d lidar scans. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 2624-2630. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.472, + 0.675 + ], + "angle": 0, + "content": "[9] Fabian Duerr, Mario Pfaller, Hendrik Weigel, and Jürgen Beyerer. Lidar-based recurrent 3d semantic segmentation with temporal memory alignment. In 2020 International Conference on 3D Vision (3DV), pages 781-790. IEEE, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.472, + 0.746 + ], + "angle": 0, + "content": "[10] Khaled El Madawi, Hazem Rashed, Ahmad El Sallab, Omar Nasr, Hanan Kamel, and Senthil Yogamani. Rgb and lidar fusion based 3d semantic segmentation for autonomous driving. In IEEE Intelligent Transportation Systems Conference, pages 7-12, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.472, + 0.815 + ], + "angle": 0, + "content": "[11] Kyle Genova, Xiaoqi Yin, Abhijit Kundu, Caroline Pantofaru, Forrester Cole, Avneesh Sud, Brian Brewington, Brian Shucker, and Thomas Funkhouser. Learning 3d semantic segmentation with only 2d image supervision. In 3DV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.472, + 0.858 + ], + "angle": 0, + "content": "[12] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[13] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-Voxel Knowledge Distillation for LiDAR Semantic Segmentation. In IEEE Conference on Computer" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.893, + 0.119 + ], + "angle": 0, + "content": "Vision and Pattern Recognition, pages 8479-8488, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.895, + 0.19 + ], + "angle": 0, + "content": "[14] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.895, + 0.259 + ], + "angle": 0, + "content": "[15] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 228-240, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.26, + 0.895, + 0.315 + ], + "angle": 0, + "content": "[16] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. Lasermix for semi-supervised lidar semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21705-21715, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.895, + 0.385 + ], + "angle": 0, + "content": "[17] Lars Kreuzberg, Idil Esen Zulfikar, Sabarinath Mahadevan, Francis Engelmann, and Bastian Leibe. 4d-stop: Panoptic segmentation of 4d lidar using spatio-temporal object proposal generation and aggregation. In European Conference on Computer Vision, pages 537-553. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.895, + 0.453 + ], + "angle": 0, + "content": "[18] Georg Krispel, Michael Opitz, Georg Waltner, Horst Possegger, and Horst Bischof. Fuseseg: Lidar point cloud segmentation fusing multi-modal data. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1874-1883, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.895, + 0.496 + ], + "angle": 0, + "content": "[19] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. arXiv preprint arXiv:2303.12766, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.895, + 0.538 + ], + "angle": 0, + "content": "[20] Jiale Li, Hang Dai, Hao Han, and Yong Ding. MSeg3D: Multi-modal 3D Semantic Segmentation for Autonomous Driving. arXiv preprint arXiv:2303.08600, 2023. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.895, + 0.594 + ], + "angle": 0, + "content": "[21] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research, 20(6):837-854, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.895, + 0.677 + ], + "angle": 0, + "content": "[22] Jiahui Liu, Chirui Chang, Jianhui Liu, Xiaoyang Wu, Lan Ma, and Xiaojuan Qi. Mars3d: A plug-and-play motion-aware model for semantic segmentation on multi-scan 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9372-9381, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.895, + 0.733 + ], + "angle": 0, + "content": "[23] Youquan Liu, Yeqi Bai, Lingdong Kong, Runnan Chen, Yuenan Hou, Botian Shi, and Yikang Li. Pcseg: An open source point cloud segmentation codebase. https://github.com/PJLab-ADG/PCSeg, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.734, + 0.895, + 0.816 + ], + "angle": 0, + "content": "[24] Youquan Liu, Runnan Chen, Xin Li, Lingdong Kong, Yuchen Yang, Zhaoyang Xia, Yeqi Bai, Xinge Zhu, Yuexin Ma, Yikang Li, et al. Uniseg: A unified multi-modal lidar segmentation network and the openpcseg codebase. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21662-21673, 2023. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.895, + 0.872 + ], + "angle": 0, + "content": "[25] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.895, + 0.902 + ], + "angle": 0, + "content": "[26] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15319" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[27] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in neural information processing systems, pages 5099-5108, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.234 + ], + "angle": 0, + "content": "[28] Herbert Robbins and Sutton Monro. A Stochastic Approximation Method. The annals of mathematical statistics, pages 400-407, 1951. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[29] Radu Alexandru Rosu, Peer Schütt, Jan Quenzel, and Sven Behnke. Latticenet: Fast point cloud segmentation using permutohedral lattices. arXiv preprint arXiv:1912.05905, 2019.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.471, + 0.334 + ], + "angle": 0, + "content": "[30] Hanyu Shi, Guosheng Lin, Hao Wang, Tzu-Yi Hung, and Zhenhua Wang. Spsequencenet: Semantic segmentation network on 4d point clouds. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.336, + 0.471, + 0.404 + ], + "angle": 0, + "content": "[31] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In European Conference on Computer Vision, pages 685-702. Springer, 2020. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.471, + 0.462 + ], + "angle": 0, + "content": "[32] Jiexiong Tang, John Folkesson, and Patric Jensfelt. Sparse2dense: From direct sparse odometry to dense 3-d reconstruction. IEEE Robotics and Automation Letters, 4(2): 530-537, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.471, + 0.545 + ], + "angle": 0, + "content": "[33] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In 2019 International Conference on Robotics and Automation (ICRA), pages 4376-4382. IEEE, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.471, + 0.603 + ], + "angle": 0, + "content": "[34] Dong Wu, Man-Wen Liao, Wei-Tian Zhang, Xing-Gang Wang, Xiang Bai, Wen-Qing Cheng, and Wen-Yu Liu. Yolop: You only look once for panoptic driving perception. Machine Intelligence Research, 19(6):550-562, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[35] Aoran Xiao, Jiaxing Huang, Dayan Guan, Kaiwen Cui, Shijian Lu, and Ling Shao. PolarMix: A General Data Augmentation Technique for LiDAR Point Clouds. arXiv preprint arXiv:2208.00223, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.471, + 0.731 + ], + "angle": 0, + "content": "[36] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeezesegv3: Spatially-adaptive convolution for efficient point-cloud segmentation. In European Conference on Computer Vision, pages 1-19. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.471, + 0.8 + ], + "angle": 0, + "content": "[37] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. RPVnet: A Deep and Efficient Range-Point-Voxel Fusion Network for Lidar Point Cloud Segmentation. In IEEE International Conference on Computer Vision, pages 16024–16033, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[38] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3101-3109, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[39] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2DPASS: 2D Priors" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.893, + 0.121 + ], + "angle": 0, + "content": "Assisted Semantic Segmentation on LiDAR Point Clouds. In European Conference on Computer Vision, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[40] Jihan Yang, Shaoshuai Shi, Runyu Ding, Zhe Wang, and Xiaojuan Qi. Towards efficient 3d object detection with knowledge distillation. Advances in Neural Information Processing Systems, 35:21300-21313, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[41] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet: Towards a unified multi-task network for lidar perception. arXiv preprint arXiv:2209.09385, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.894, + 0.29 + ], + "angle": 0, + "content": "[42] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE access, 8:58443-58469, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.893, + 0.361 + ], + "angle": 0, + "content": "[43] Zhiwei Zhang, Zhizhong Zhang, Qian Yu, Ran Yi, Yuan Xie, and Lizhuang Ma. Lidar-camera panoptic segmentation via geometry-consistent and semantic-aware alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3662-3671, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.893, + 0.431 + ], + "angle": 0, + "content": "[44] Wu Zheng, Li Jiang, Fanbin Lu, Yangyang Ye, and Chi-Wing Fu. Boosting single-frame 3d object detection by simulating multi-frame point clouds. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4848-4856, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.894, + 0.502 + ], + "angle": 0, + "content": "[45] Minghan Zhu, Shizhong Han, Hong Cai, Shubhankar Borse, Maani Ghaffari, and Fatih Porikli. 4d panoptic segmentation as invariant and equivariant field prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22488-22498, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.893, + 0.573 + ], + "angle": 0, + "content": "[46] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and Asymmetrical 3D Convolution Networks for Lidar Segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 9939-9948, 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.574, + 0.893, + 0.615 + ], + "angle": 0, + "content": "[47] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. In ICCV, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15320" + } + ] +] \ No newline at end of file diff --git a/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/745b72e4-2dd6-4811-ade0-f52413f14141_origin.pdf b/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/745b72e4-2dd6-4811-ade0-f52413f14141_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..12bd7591d9c1d3f5d61e3c7016c458d74b853033 --- /dev/null +++ b/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/745b72e4-2dd6-4811-ade0-f52413f14141_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10dbf445fece09d12765d051668871afda978d2499b2d97511455316ad440d23 +size 2241572 diff --git a/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/full.md b/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..38dc397a06a1da20b3f34e811d0622bd766b3123 --- /dev/null +++ b/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/full.md @@ -0,0 +1,293 @@ +# TASeg: Temporal Aggregation Network for LiDAR Semantic Segmentation + +Xiaopei Wu $^{1,2}$ , Yuenan Hou $^{2*}$ , Xiaoshui Huang $^{2*}$ , Binbin Lin $^{3,4}$ , Tong He $^{2}$ , Xinge Zhu $^{5}$ , Yuexin Ma $^{6}$ , Boxi Wu $^{4}$ , Haifeng Liu $^{1*}$ , Deng Cai $^{1}$ , Wanli Ouyang $^{2}$ + +$^{1}$ State Key Lab of CAD&CG, Zhejiang University $^{2}$ Shanghai AI Laboratory $^{3}$ Fullong Inc. + +$^{4}$ School of Software Technology, Zhejiang University $^{5}$ The Chinese University of Hong Kong $^{6}$ ShanghaiTech University + +# Abstract + +Training deep models for LiDAR semantic segmentation is challenging due to the inherent sparsity of point clouds. Utilizing temporal data is a natural remedy against the sparsity problem as it makes the input signal denser. However, previous multi-frame fusion algorithms fall short in utilizing sufficient temporal information due to the memory constraint, and they also ignore the informative temporal images. To fully exploit rich information hidden in long-term temporal point clouds and images, we present the Temporal Aggregation Network, termed TAseg. Specifically, we propose a Temporal LiDAR Aggregation and Distillation (TLAD) algorithm, which leverages historical priors to assign different aggregation steps for different classes. It can largely reduce memory and time overhead while achieving higher accuracy. Besides, TLAD trains a teacher injected with gt priors to distill the model, further boosting the performance. To make full use of temporal images, we design a Temporal Image Aggregation and Fusion (TIAF) module, which can greatly expand the camera FOV and enhance the present features. Temporal LiDAR points in the camera FOV are used as mediums to transform temporal image features to the present coordinate for temporal multi-modal fusion. Moreover, we develop a Static-Moving Switch Augmentation (SMSA) algorithm, which utilizes sufficient temporal information to enable objects to switch their motion states freely, thus greatly increasing static and moving training samples. Our TAseg ranks $\mathbf{1}^{st\dagger}$ on three challenging tracks, i.e., SemanticKITTI single-scan track, multi-scan track and nuScenes LiDAR segmentation track, strongly demonstrating the superiority of our method. Codes are available at https://github.com/LittlePey/TASeg. + +# 1. Introduction + +LiDAR segmentation aims to infer the semantic information of each point in point clouds and plays an indispensable role + +in the autonomous driving [3, 21, 34, 42]. With the advent of deep learning, a large quantity of LiDAR segmentation algorithms [14, 19, 24, 39, 46] have been put forward. Despite their impressive results, the segmentation performance is still constrained by the inherent sparsity of point clouds. To alleviate this, it is desirable to aggregate temporal data. Previous multi-frame models [4, 9, 22] can only fuse a few LiDAR frames due to the GPU memory constraint, which restricts them from utilizing rich information hidden in the long-term temporal point clouds. Although MSeg3D [20] stacks dozens of LiDAR frames, the memory and time overhead are intolerable. Moreover, previous attempts concentrate on utilizing temporal point clouds, ignoring the valuable information hidden in temporal images. + +To better leverage long-term temporal information, we propose a Temporal LiDAR Aggregation and Distillation (TLAD) algorithm. It can perform efficient multi-frame aggregation while achieving higher accuracy. TLAD consists of Flexible Step Aggregation (FSA) and Mask Distillation. FSA is based on the observation that the model needs different amounts of temporal points for different classes. For difficult classes, such as bicyclists, more points are needed to yield accurate predictions. However, for easy classes, such as cars and roads, it is unnecessary to aggregate too many point cloud frames for them. Therefore, we propose to assign different aggregation steps for different classes in point cloud sequences, which can significantly save memory and computation overhead on easy classes while providing sufficient temporal points for difficult classes. To identify classes of temporal point clouds, historical predictions are used, considering the temporal data is processed in chronological order, which is consistent with practical applications. + +Interestingly, by assigning different classes with different steps, FSA actually injects historical priors into the aggregated point clouds. For classes with different steps, their patterns can be more discriminative due to their different densities. This makes the point cloud frames accumulated by FSA more conducive for segmentation. Experimental results show that FSA can not only reduce memory and time costs but also enhance overall performance. To further verify our + +idea, we directly replace historical predictions of temporal point clouds with their ground truth labels for FSA. We find the performance boosts greatly. This inspires us to use the model injected with gt priors to distill the model injected with historical priors, which we term Mask Distillation. + +To fully use the informative temporal images, we devise a Temporal Image Aggregation and Fusion (TIAF) module. Previous multi-modal fusion methods usually suffer from limited image features due to the different FOVs (field of views) between the LiDAR and the camera. We observe that as the ego-vehicle moves forward, cameras can capture different FOVs. By aggregating images of different timestamps, the FOV of the present camera can be enlarged greatly. In addition, temporal images can provide different views for the same region at different timestamps, which can supply richer information. Our TIAF leverages temporal LiDAR points in the FOV of the corresponding temporal camera to transform historical image features to the present coordinate with the pose matrix. After aggregating all temporal image features, we use a series of 3D sparse convolutions to fuse them and convert them to voxel representation. Ultimately, we use temporal LiDAR points to gather temporal image features for temporal multi-modal fusion. The fused features embrace both temporal LiDAR and temporal image information, which contribute to more precise segmentation results. To our knowledge, we are the first to leverage temporal point clouds and temporal images simultaneously for LiDAR semantic segmentation. + +Moreover, in the multi-scan task, the model needs to distinguish the motion states (moving or static) of a movable object, which is very challenging. In this paper, we make full use of temporal information to remarkably increase the multi-scan perception ability of the model. Specifically, we propose a data augmentation strategy named Static-Moving Switch Augmentation (SMSA). It manipulates temporal point clouds of a movable object to switch its motion state. In this way, we can increase the quantity of static and moving samples significantly, even in the absence of training samples of static or moving classes. + +In summary, our major contributions are listed as follows: + +- We propose a Temporal LiDAR Aggregation and Distillation algorithm, which uses Flexible Step Aggregation and Mask Distillation techniques to largely reduce memory and time costs while achieving higher accuracy. +- We devise a Temporal Image Aggregation and Fusion module, which exploits temporal images to enlarge the camera FOV and enhance present features. It also delivers a scheme for temporal multi-modal fusion. +- We design a Static-Moving Switch Augmentation algorithm to enable static and moving objects to switch their motion states freely. With this technique, we can greatly increase static and moving training samples. + +- Our Temporal Aggregation Network, dubbed TASeg, achieves impressive results on SemanticKITTI and nuScenes benchmarks. Notably, our TASeg ranks $1^{st}$ on three challenging tracks. Thorough ablation studies are provided to demonstrate the efficacy of our approach. + +# 2. Related Work + +LiDAR Semantic Segmentation. LiDAR semantic segmentation aims to assign a unique class label to each point in the input point cloud sequence. Recent years have witnessed an explosion of LiDAR segmentation algorithms [5, 7, 13, 15, 19, 20, 24, 27, 31, 39]. For example, [26] is the pioneering work that approximates the permutation-invariant function with a per-point and shared Multi-Layer Perceptron. [46] changes traditional cubic grids to cylindrical grids and designs a network of asymmetrical 3D convolutions. [19] divides the space with the radial window, which increases the receptive field smoothly and helps improve the performance. Despite their good segmentation performance, these methods still take the single LiDAR frame as input, which does not utilize the rich semantic and geometric information hidden in temporal data. + +Multi-Frame LiDAR Perception. Compared to a single LiDAR scan, multiple LiDAR scans can provide more sufficient information. Recently, many research efforts have been put on temporal LiDAR segmentation [1, 2, 6, 8, 9, 17, 20, 22, 30, 45]. For example, [8] leverages a Bayes filter to explore the temporal consistency. [4] shunts the historical points into two groups to utilize historical frames efficiently. Despite the success of previous multi-frame methods, they can not leverage the valuable information hidden in long-term temporal point clouds due to the GPU memory constraint. Although [20] stacks dozens of LiDAR scans, the memory and time overhead are intolerable. In this paper, we present an efficient multi-frame aggregation algorithm, which can greatly save memory and computation consumption while achieving higher performance. + +Multi-Modal Fusion. Since LiDAR and camera are two complementary sensors for 3D semantic segmentation, multimodal fusion has gained increasing attention in recent years [10, 18, 24, 43, 47]. However, these multi-modal fusion methods usually suffer from limited overlapped regions between the LiDAR and camera due to their different FOVs. [39] proposes a cross-modal knowledge distillation method, which is free from images at inference, while it causes much loss of the RGB information. [20] completes the missing camera features using predicted pseudo-camera features, while the used image information is still restricted to the present camera FOV. In this paper, we fully take advantage of temporal images to enlarge the FOV of the present camera and enhance present image features. + +Knowledge Distillation for LiDAR Perception. Knowledge distillation [12] is widely used in various fields in that it can improve the performance of the student without sacrificing inference efficiency. In LiDAR perception, [13, 40] compress a cumbersome teacher to a lightweight student to reinforce the representation learning of the student as well as maintain high efficiency. [39] distills the prior of 2D images to 3D point clouds with well-designed cross-modal knowledge distillation module. [32, 44] utilizes a multi-frame teacher to help a single-frame student learn dense 3D features. By contrast, our method aims to transfer knowledge from a multi-frame teacher injected with gt priors to a multi-frame student injected with historical priors. + +# 3. Methodology + +# 3.1. Temporal LiDAR Aggregation and Distillation + +A simple solution to utilize the temporal information of a consecutive point cloud sequence is to concatenate all points of the sequence, as shown in the following equation: + +$$ +X _ {t} = \operatorname {c o n c a t} \left(P _ {t}, T _ {t - 1} P _ {t - 1}, \dots , T _ {t - \Delta t} P _ {t - \Delta t}\right), \tag {1} +$$ + +$$ +Y _ {t} = \operatorname {c o n c a t} \left(L _ {t}, L _ {t - 1}, \dots , L _ {t - \Delta t}\right), +$$ + +where $P_{t - i}$ and $L_{t - i}$ denote the $(t - i)^{th}$ point cloud frame and the corresponding point-wise label. $T_{t - i}$ is the transformation matrix that transforms the coordinate from the $(t - i)^{th}$ frame to the $t^{th}$ frame. $\Delta t$ is the window size of temporal point clouds and $\text{concat}(.)$ denotes the concatenation operation. $X_{t}$ and $Y_{t}$ are the aggregated LiDAR frame and point-wise label. Although simple, there are some problems. On the one hand, direct concatenation consumes much GPU memory. On the other hand, the huge memory cost constrains the multi-frame model from utilizing more historical frames, thereby limiting the ultimate performance. + +# 3.1.1 Flexible Step Aggregation + +To reduce the GPU memory consumption, we can sample temporal frames with a step, while this method is also suboptimal. A small step introduces huge memory overhead, whilst a large step cannot utilize sufficient temporal information. To this end, we propose Flexible Step Aggregation (FSA). Our method is based on the observation that for different classes, the model needs different amounts of temporal information. For difficult classes, such as bicycles, more points are needed to yield accurate predictions. For easy classes, such as cars and buildings, it is unnecessary to aggregate many point cloud frames for them. Based on the above analysis, we propose to assign different steps for different classes according to their learning difficulty. Specifically, we leverage historical predictions to divide the temporal point clouds into several class groups and assign a specific sampling step for each group. Then, we aggregate temporal + +points for each group with the corresponding step. Eventually, temporal points of all groups are concatenated with the current frame, resulting in an aggregated frame. + +The group division is not strict as long as it follows the principle that more difficult classes need smaller steps. For example, we can simply divide all classes into three groups according to their segmentation performance, such as [0, 80) mIoU, [80, 90) mIoU and [90, 100] mIoU. Then, we assign the three groups with a step of 2, 4 and $\infty$ ( $\infty$ means we do not aggregate temporal points for the group). To save more memory and computation without sacrificing performance, a more fine-grained division can be used. Formally, suppose we divide temporal point clouds into $g$ groups. We aggregate temporal points for the $k^{th}$ group as below: + +$$ +X _ {t} ^ {k} = \operatorname {c o n c a t} \left(T _ {o _ {1}} P _ {o _ {1}} M _ {o _ {1}} ^ {k}, \dots , T _ {o _ {n}} P _ {o _ {n}} M _ {o _ {n}} ^ {k}\right), \tag {2} +$$ + +$$ +Y _ {t} ^ {k} = \operatorname {c o n c a t} \left(L _ {o _ {1}} M _ {o _ {1}} ^ {k}, \dots , L _ {o _ {n}} M _ {o _ {n}} ^ {k}\right). +$$ + +Here $o = \{o_i|o_i = t - i\times s_k,i = 1,2,\dots,n,n = \lfloor \Delta t / s_k\rfloor \}$ $s_k$ is the sampling step for the $k^{th}$ class group and $\lfloor .\rfloor$ is the floor operation. $M_{o_i}^k$ is the group mask that indicates which point of $P_{o_i}$ belongs to the $k^{th}$ class group. It can be obtained from historical predictions. Finally, we concatenate points and labels of all groups with the current frame: + +$$ +X _ {t} = \operatorname {c o n c a t} \left(P _ {t}, X _ {t} ^ {1}, \dots , X _ {t} ^ {k}, \dots , X _ {t} ^ {g}\right), \tag {3} +$$ + +$$ +Y _ {t} = \operatorname {c o n c a t} \left(L _ {t}, Y _ {t} ^ {- 1}, \dots , X _ {t} ^ {k}, \dots , Y _ {t} ^ {g}\right). +$$ + +In this way, we discard massive redundant temporal points while maintaining essential temporal information. Since easy classes are usually large and hold a large quantity of points, the GPU memory overhead can be further reduced. Moreover, by assigning different classes with different steps, FSA actually injects historical priors into the aggregated point clouds. For classes with different steps, their patterns can be more discriminative due to their different densities (steps), which makes it easier to segment the multi-frame point clouds aggregated by FSA. Experiments in Table 4 verify that the proposed FSA can not only save memory and time costs but also achieve better performance. + +# 3.1.2 Mask Distillation + +In FSA, we use historical predictions to generate group masks, which we call pseudo masks. Since ground truth labels are more accurate than historical predictions, a natural question arises: what if we use gt masks (the group masks generated by ground truth labels) for FSA? Our experiment shows that the performance can be improved greatly. Actually, using gt masks for FSA can produce more discriminative patterns for classes with different steps. This motivates us to simulate features of a model trained on temporal point clouds aggregated with gt masks. In this way, our model is guided to learn more discriminative features to distinguish different classes. We call this Mask Distillation. Specifically, + +![](images/9b15e5bcf6fe62695fdcac9bb7297f21a8bc49fd7ab50e8acab5c209ffef3555.jpg) +Figure 1. Overview of our Temporal Aggregation Network (TASeg). (1) Temporal LiDAR Aggregation and Distillation leverages the proposed Flexible Step Aggregation (FSA) to assign different temporal steps for different classes, and it utilizes a teacher injected with gt priors for knowledge distillation. (2) Temporal Image Aggregation and Fusion takes temporal LiDAR points as mediums to transform historical image features to the present coordinate. 3D sparse convolutions are employed to fuse temporal image features. Finally, we use temporal LiDAR points to gather voxel-wise temporal image features for temporal multi-modal fusion. + +we formulate the distillation on the feature map of student $F^{s}\in \mathbb{R}^{N^{s}\times C}$ and the feature map of teacher $F^{t}\in \mathbb{R}^{N^{t}\times C}$ : + +$$ +\mathcal {L} _ {\mathrm {K D}} = \mathbb {E} \left[ \left\| F ^ {s} m ^ {s} - F ^ {t} m ^ {t} \right\| _ {2} \right]. \tag {4} +$$ + +Since temporal point clouds aggregated with pseudo masks and gt masks are different, we use masks $m^s$ and $m^t$ to select voxels that appeared in both $F^s$ and $F^t$ . Note that our Mask Distillation is distinct from the methods that distill a multi-frame model to a single-frame model [32, 44]. Both the student and teacher in our method are multi-frame models. + +# 3.2. Temporal Image Aggregation and Fusion + +Previous multi-modal fusion methods only focus on leveraging present images while ignoring the precious value of temporal images. Temporal images can provide broader camera FOVs and richer information. Besides, they can enable more robust multi-modal fusion under the malfunction condition on some cameras. In this section, we provide an effective solution for aggregating temporal image features and performing temporal multi-modal fusion. + +Temporal Aggregation and Fusion. Since temporal images are in different feature spaces, it is difficult to establish the relationship between different images for feature aggregation. In our method, we take temporal LiDAR points as mediums to transform temporal image features to the present + +coordinate with the pose information. This way, temporal image features are unified to the present 3D space. Specifically, given an image $I_{t - \Delta t}\in \mathbb{R}^{H\times W\times 3}$ and point cloud $P_{t - \Delta t}\in \mathbb{R}^{N\times 3}$ , we use an image network to extract the image feature $Z_{t - \Delta t}\in \mathbb{R}^{H\times W\times C}$ . According to the sensor calibration, we can establish the pixel-to-point mapping between 2D pixels and 3D points. Hence, we can project image feature $Z_{t - \Delta t}$ to 3D space, resulting in point-wise image feature $Q_{t - \Delta t}\in \mathbb{R}^{M\times C}$ , where $M$ is the number of LiDAR points located on $Z_{t - \Delta t}$ . By transforming $Q_{t - \Delta t}$ to the present coordinate with the pose matrix, we realize the aggregation of temporal image features: + +$$ +\bar {X} _ {t} = \operatorname {c o n c a t} \left(Q _ {t}, T _ {o _ {1}} Q _ {o _ {1}}, \dots , T _ {o _ {n}} Q _ {o _ {n}}\right), \tag {5} +$$ + +where $o = \{o_i|o_i = t - i\times s,i = 1,2,\dots,\lfloor \Delta t / s\rfloor \}$ and $s$ is the sampling step for temporal images. With temporal images aggregated, the FOV of the present camera is expanded greatly. Moreover, temporal image feature fusion becomes convenient because they are unified to the same 3D space. Concretely, we can use several 3D sparse convolutions to fuse aggregated temporal image features, which also endows them with geometric information, as shown in Equation 6. In addition, feature map downsampling is also utilized to generate multi-scale voxel features, providing richer information for subsequent temporal multi-modal fusion. + +$$ +\bar {V} _ {t} = \operatorname {S p a r s e C o n v} _ {\mathrm {3 D}} \left(\operatorname {V o x e l i z a t i o n} \left(\bar {X} _ {t}\right)\right). \tag {6} +$$ + +Temporal Multi-Modal Fusion. Benefiting from the temporal image aggregation, associating temporal image features with temporal LiDAR points also becomes convenient. Specifically, given temporal image features converted to unified voxel representation $\overline{V}_t$ , we can establish a point-to-voxel association between temporal LiDAR points and $\overline{V}_t$ . For each temporal LiDAR point, we generate its image feature by pooling its nearby voxel-wise temporal image features with trilinear interpolation instead of hard indexing. To extract richer image features, we perform interpolation on multi-scale feature maps. Finally, we concatenate the point cloud features and aggregated multi-scale image features, resulting in fused features, which convey powerful information of both temporal point clouds and temporal images. + +2D and 3D Supervision. To make the extracted image features more informative, we add 2D supervision and 3D supervision on the 2D backbone and 3D convolutions in the image branch, respectively. The 3D supervision is just the label of point clouds. The 2D supervision is obtained by projecting labels of point clouds to the image plane. + +# 3.3. Static-Moving Switch Augmentation + +In the multi-scan task, the model is required to distinguish the motion state of movable objects. To enable the model to explore a large data space, we design an effective data augmentation, Static-Moving Switch Augmentation (SMSA). SMSA enables a movable object to switch its motion state freely, which can remarkably increase the sample quantity of static and moving objects. Concretely, considering that a unique object $b$ has the same instance id in all frames of a sequence, we can use its instance mask to crop its temporal point clouds, which is denoted as $\mathcal{B} = \{\mathcal{B}_{o_i}|i = 0,1,\dots,\lfloor \Delta t / s\rfloor ,\mathcal{B}_{o_i}\in \mathbb{R}^{N\times 3}\}$ . Here $o = \{o_i|o_i = t - i\times s,i = 1,2,\dots,\lfloor \Delta t / s\rfloor \}$ , $s$ is the sampling step and $\mathcal{B}_{o_i}$ is the temporal part of $b$ at $t_{o_i}$ . By manipulating $\mathcal{B}$ , we can change the motion state of $b$ . + +Static to Moving. If $b$ is static, all temporal parts of $b$ locate at the same position, as shown in the upper-left of Figure 2. To change $b$ to a moving object, we can shift each temporal part of $b$ with an offset. Considering that objects typically move at a constant speed within a short time, we set the offset between adjacent temporal parts to be the same. For offset itself, it is a random value to increase the diversity of resulting moving samples. As for the direction of the offset, it can be roughly estimated by comparing the width and height of the minimum bounding box of $b$ . Since static objects often park on the side of the road, which is crowded, the shifted temporal parts of static objects may overlap with other objects. To alleviate this, we define a set of anchor points around $b$ and a coverage area for each anchor point. + +![](images/a2410a010fdcb1d5c563c6b1ae14acc2326ee362bd875d1ed10945435cc43774.jpg) +Figure 2. Visualization of the augmented samples by our Static-Moving Switch Augmentation (SMSA). Our SMSA switches the motion state by manipulating the temporal parts of objects. + +Then, we shift all temporal parts of $b$ to the anchor point whose coverage area contains the fewest LiDAR points (refer to the supplementary material for more details). + +Moving to Static. If $b$ is moving, different temporal parts of $b$ locate at different positions. To switch its motion state to static, we can shift all its temporal parts to the same position. In particular, we calculate the centers of each temporal part of $b$ , which are denoted as $\mathcal{C} = \{\mathcal{C}_{o_i} | \mathcal{C}_{o_i} \in \mathbb{R}^3\}$ . Considering the trajectory of moving objects in a short time is approximate to a line and the speed is a constant, we simply use $q = \mathcal{C}_{o_0} - \mathcal{C}_{o_1}$ as the offset of adjacent temporal parts of $b$ . Eventually, we can obtain a static object by shifting all temporal parts of $b$ to $C_{o_0}$ (see Figure 2). + +It should be noted that our SMSA is an online and plug-and-play data augmentation strategy, which consumes negligible storage and computation costs. + +# 3.4. Training Objective + +The final training objective is comprised of LiDAR segmentation loss, mask distillation loss, fusion segmentation loss, 2D and 3D supervision loss on the image branch: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {L i D A R}} + \alpha \mathcal {L} _ {\mathrm {K D}} + \beta \mathcal {L} _ {\text {F u s i o n}} + \gamma (\mathcal {L} _ {\mathrm {2 D}} + \mathcal {L} _ {\mathrm {3 D}}), \tag {7} +$$ + +where $\alpha, \beta, \gamma$ are the coefficients to control the effect of different losses. We set $\alpha = 1, \beta = 1, \gamma = 1$ by default. + +# 4. Experiments + +Datasets & Evaluation Metrics. Following [13, 46], we evaluate the performance on two popular LiDAR segmentation benchmarks, i.e., SemanticKITTI and nuScenes. SemanticKITTI contains 22 point cloud sequences, where sequences 0-10, 8, 11-21 are selected as training, validation and testing, respectively. As for nuScenes, it collects 1, 000 driving scenes, where 850 scenes are chosen for training and validation, and the remaining 150 scenes are used for testing. We conduct experiments on three tracks, i.e., SemanticKITTI single-scan and multi-scan semantic segmentation and nuScenes semantic segmentation. + +
MethodmIoUcarbicyclemotorcycletruckother-veh.personbicyclistmotorcyclistroadparkingsidewalkother-gro.buildingfencevegetationtrunkterrainpoletraffic sign
SqueezeSegV2 [33]39.781.818.517.913.414.020.125.13.988.645.867.617.773.741.171.835.860.220.226.3
RangeNet53++ [25]52.291.425.734.425.723.038.338.84.891.865.075.227.887.458.680.555.164.647.955.9
RandLA-Net [14]55.994.229.832.243.939.148.447.49.490.561.874.024.589.760.483.863.668.651.050.7
SqueezeSegV3 [36]55.992.538.736.529.633.045.646.220.191.763.474.826.489.059.482.058.765.449.658.9
JS3C-Net [38]66.095.859.352.954.346.069.565.439.988.961.972.131.992.570.884.569.867.960.768.7
SPVNAS [31]67.097.250.650.456.658.067.467.150.390.267.675.421.891.666.986.173.471.064.367.3
Cylinder3D [46]68.997.167.663.850.858.573.769.248.092.265.077.032.390.766.585.672.569.862.466.2
RPVNet [37]70.397.668.468.744.261.175.974.443.493.470.380.733.393.572.186.575.171.764.861.4
(AF)2-S3Net [5]70.894.363.081.440.240.076.481.777.792.066.876.245.892.569.678.668.063.164.073.3
PVKD [13]71.297.069.353.567.960.275.173.550.591.877.570.941.092.469.486.573.871.964.965.8
2DPASS [39]72.997.063.663.461.161.577.981.374.189.767.474.740.093.572.986.273.971.065.070.4
SphereFormer [19]74.897.570.170.559.667.779.080.475.391.869.778.241.393.872.886.775.172.466.872.9
UniSeg [24]75.297.971.975.263.674.178.974.860.692.674.079.546.193.472.787.576.373.168.368.5
TASeg (Ours)76.597.771.871.465.278.779.984.678.691.674.078.039.393.573.486.675.071.769.673.8
+ +Table 1. Comparison with state-of-the-arts on SemanticKITTI test set (single-scan). All results can be found on the online leaderboard. + +
MethodmIoUcar-scar-mtruc.-struc.-mothe.-sothe.-mpers.-spers.-mbicy.-sbicy.-mmotor.-smotor.-m
LatticeNet [29]45.291.154.829.73.523.10.66.849.90.044.60.064.3
TemporalLidarSeg [9]47.092.168.239.22.135.012.414.440.40.042.80.012.9
\( (AF)^2-S3Net[5] \)56.991.865.315.75.627.53.916.467.615.166.467.159.6
MarS3D [22]52.795.178.439.75.136.610.016.258.01.267.30.036.3
SVQNet [4]60.596.080.141.05.160.47.128.785.10.072.40.088.1
2DPASS [39]62.496.282.148.216.152.73.835.480.37.971.262.073.1
TASeg (Ours)65.794.872.853.635.271.453.236.780.432.475.212.389.6
+ +Table 2. Comparison to the state-of-the-art methods on SemanticKITTI test set (multi-scan). -s indicates static and -m stands for moving. + +Implementation Details. We use the MinkowskiNet [7] re-implemented by PCSeg [23] as our baseline. Our TASeg is trained with the SGD [28] optimizer on 4 A100 GPUs with batch size 6 for 12 and 36 epochs on SemanticKITTI and nuScenes datasets. The learning rate and weight decay are set to 0.02 and 0.0001. The window size of temporal point clouds is set to 16. Due to a large redundancy of images, we use a step of 12 and a window size of 48 for temporal images. Our data augmentation strategy includes random flipping, rotation, scaling, transformation, LaserMix [16] and PolarMix [35]. During the inference, the LiDAR semantic predictions and image features of the previous timestamp can be saved, so there is no time cost for generating historical predictions and processing temporal image aggregation. More details are provided in the supplementary material. + +# 4.1. Comparison with State-of-the-art Methods + +We compare our TASeg with state-of-the-art LiDAR segmentation methods on SemanticKITTI single-scan and multi + +scan track, and nuScenes LiDAR semantic segmentation track, as summarized in Table 1, 2 and 3. On the SemanticKITTI single-scan track, our TASeg is $1.3\mathrm{mIoU}$ higher than UniSeg [24], the best-performing published algorithm on the SemanticKITTI single-scan leaderboard. On the SemanticKITTI multi-scan track, our approach outperforms the best multi-scan method 2DPASS [39] by $3.3\mathrm{mIoU}$ . For nuScenes LiDAR semantic segmentation, we also achieve superior performance over previous algorithms. As shown in Table 3, the proposed approach surpasses the UniSeg [24] by $1.1\mathrm{mIoU}$ . In particular, our approach holds the highest entry on all three tracks. These encouraging results strongly show the effectiveness of our approach. + +# 4.2. Ablation studies + +We examine the effect of each component through detailed ablations. Unless otherwise specified, the following experiments are trained on the training set of the SemanticKITTI single-scan task and evaluated on the validation set. + +
MethodmIoUconstruction
barrierbicyclebuscarconstructionmotorcyclepedestriantraffic-conetrailertruckdriveableother-flatsidewalkterrainmannadevegetation
Cylinder3D [46]77.282.829.884.389.463.079.377.273.484.669.197.770.280.375.590.487.6
SPVCNN [31]77.480.030.091.990.864.779.075.670.981.074.697.469.280.076.189.387.1
2D3DNet [11]80.083.059.488.085.163.784.482.076.084.871.996.967.479.876.092.189.2
2DPASS [39]80.881.755.392.091.873.386.578.572.584.775.597.669.179.975.590.288.0
LidarMultiNet [41]81.480.448.494.390.071.587.285.280.486.974.897.867.380.776.592.189.6
MSeg3D [20]81.183.142.594.992.067.178.685.780.587.577.397.769.881.277.892.490.1
SphereFormer [19]81.983.339.294.792.577.584.284.479.188.478.397.969.081.577.293.490.2
UniSeg [24]83.585.971.292.191.680.588.080.976.086.376.797.771.880.776.791.388.8
TASeg (Ours)84.687.169.490.592.278.790.486.381.988.375.997.870.981.078.293.491.2
+ +Effect of Flexible Step Aggregation. In Table 4, we compare our FSA with other multi-frame algorithms. MSeg3D [20] utilizes temporal point clouds by directly concatenating all points, which introduces a large memory consumption and achieves limited improvement. SVQNet [4] leverages cross-attention to fuse temporal features while it can only handle fewer LiDAR scans. When increasing LiDAR scans, the memory overhead is also huge. Our FSA assigns different classes with different temporal steps, which can leverage long-term temporal information with less memory and time costs. Moreover, thanks to the utilization of historical priors, our FSA achieves better results (71.3 mIoU) than directly concatenating (69.9 mIoU), given the same temporal window size. To explore the effect of different window sizes on FSA, we present Table 5. Results show that the performance is saturated with a window size of 24. + +Effect of Group Division. To investigate the effect of different group divisions for FSA, we provide Table 6. For division1, we simply divide all classes into three groups according to their performance on the val set. The group1 consists of the classes in [90, 100) mIoU, such as cars and roads. The group2 consists of classes in [80, 90) mIoU, such as motorcycles. The group3 contains the remaining classes. We assign group1, group2 and group3 with the step of $\infty$ , 4 and 2. A step of $\infty$ means we do not aggregate temporal points for classes in the group due to their near-saturate performances. The result shows that division1 can reduce the time and memory costs largely compared to directly stacking (the $2^{st}$ row in Table 4). We can also use performance top1-6, top6-12, and top12-19 for division, resulting in division2. The group division is robust as long as it follows the principle that more difficult classes need smaller steps. To further reduce the memory consumption, we can finetune the group division, such as moving large-size classes (e.g. other-ground and terrain) in group3 to group2 (division3) or moving large-size classes in group3 and group2 to a new + +Table 3. Performance comparison with start-of-the-art methods on nuScenes test set. + +
MethodLatencyMemorymIoU
Baseline63ms5.0G68.9
+ MSeg3D [20]†284ms69.7G69.9
+ SVQNet [4]†121ms15.4G69.5
+ FSA79ms8.5G71.3
+ FSA w/ gt mask79ms8.5G75.9
+ FSA + M.D.79ms8.5G71.8
+ FSA + M.D. × 0.7572ms6.9G71.6
+ +Table 4. Comparison with different multi-frame algorithms. $\dagger$ represents our re-implementation. "M.D." represents Mask Distillation. "×75" represents that we reduce the model parameters to $75\%$ . + +
Window size481216202428
mIoU69.970.470.971.371.471.671.4
+ +Table 5. Ablation on different window sizes for FSA. + +
MethodLatencyMemorymIoU
FSA w/ division192ms14.1G70.8
FSA w/ division283ms7.2G70.6
FSA w/ division379ms8.5G71.3
FSA w/ division474ms7.7G71.2
FSA w/ division574ms7.4G71.2
+ +Table 6. Ablation on different group divisions for FSA. + +group with step of 8 (division4). Besides, we can divide each class group into a close and a distant group according to the distance (such as $30\mathrm{m}$ ). Considering close areas need fewer temporal points, we can use twice the original step for temporal aggregation of close groups (division5). + +Effect of Mask Distillation. At the $5^{th}$ row of Table 4, we use gt masks of historical frames for FSA. The performance is greatly boosted. Considering we cannot get ground truth at inference, we use it as a teacher to distill the model trained with pseudo masks. The result shows that the student is improved by $0.5\mathrm{mIoU}$ after distillation. With $75\%$ model complexity, we can further reduce the time and memory cost and still achieve a leading performance. + +Effect of Temporal Image Aggregation and Fusion. TIAF delivers a temporal multi-modal fusion scheme to make full use of temporal images. As shown in Table 7, with zero image, only TLAD is used, achieving $71.8\mathrm{mIoU}$ . With one image, only present images are utilized. Due to the limited FOV of the present camera, only part of the LiDAR points can gather image features, which limits the multi-modal fusion. Our TIAF leverages temporal images to enlarge the FOV of the camera and enhance present features. The results show that with the number of images increasing, the performance rises gradually, which verifies the effectiveness of our TIAF. When using seven temporal images, we achieve $1.0\mathrm{mIoU}$ improvement on the strong baseline. Note that there is no extra cost to generate historical image features because we can save them at the past moments. + +
Num of images01357
mIoU71.872.172.472.772.8
+ +Table 7. Ablation on the number of temporal images for TIAF. + +
Exp2D Sup.3D Sup.Multi-ScalemIoU
(a)✓✓72.4
(b)✓✓72.5
(c)✓✓72.4
(d)✓✓✓72.7
+ +We also provide the ablation on different supervisions and multi-scale features for TIAF in Table 8. In TIAF, we leverage 2D and 3D supervision to guide the extracted image features to be more conducive for segmentation. Multi-scale is also used to provide more discriminative features. The result shows that each of the designs is beneficial for final performance. Note that our temporal multi-modal method is orthogonal to other single-frame multi-modal methods. We use a simple pixel-to-point mapping at the feature projection stage. More complex methods [20, 24] can also be utilized, but it is not the focus of our TIAF. + +Effect of Static-Moving Switch Augmentation. To verify our SMSA, we present Table 9. From the results, we can find that without SMSA, the accuracy of the model on many classes is lower than $10\mathrm{IoU}$ . After using SMSA, the performance of moving truck, moving other-vehicle and static bicyclist classes is improved by more than $20\mathrm{IoU}$ , and the overall performance boosts from $61.3\mathrm{mIoU}$ to $65.7\mathrm{mIoU}$ , which strongly demonstrates the efficacy of our SMSA. + +Table 8. Effect of different supervisions and multi-scale for TIAF. + +
SMSAmloUcar-scar-mtrue-strue-mother-sother-mpers-spers-mbicy-sbicy-mmotor-smotor-m
✓61.395.679.349.29.161.05.739.087.00.072.64.376.4
65.794.872.853.635.271.453.236.780.432.475.212.389.6
+ +Generalization to More Architectures. In addition to MinkowskiNet, we also validate our temporal aggregation method on Cylinder3D [46] and SPVNAS [31]. Detailed results are summarized in Table 10. Our TASeg brings impressive improvements to both baselines, which demonstrates the strong generalization ability of our method. + +Table 9. Ablation study of our SMSA on the SemanticKITTI test benchmark of the multi-scan track. + +
MethodmIoU
Cylinder3D [46]†66.0
Cylinder3D [46]†+ TASeg69.6
SPVNAS [31]†68.6
SPVNAS [31]†+ TASeg72.7
+ +Comparison on Complexity, Latency and Accuracy. As shown in Table 11, when not utilizing temporal images, our TASeg achieves superior performance than other methods with comparable or less complexity and latency. With temporal images, TASeg can maintain lower complexity and latency than UniSeg [24] while achieving higher accuracy. Besides, with $75\%$ parameters, the latency of TASeg can be further reduced, and the accuracy is not affected much. + +Table 10. Performance of applying our TASeg to different baselines. $\dagger$ represents our re-implementation. + +
MethodInputParamsLatencymIoU
MinkowskiNet [7]†L37.9M63ms68.4
Cylinder3D [46]†L55.9M75ms66.0
SPVNAS [31]†L96.7M105ms68.6
RPVNet [37]†L119.0M110ms68.9
UniSeg [24]L+C147.6M145ms71.3
TASeg wo/ TIAFL+T37.9M79ms71.8
TASegL+C+T46.7M116ms72.7
TASeg × 0.75L+C+T27.4M108ms72.5
+ +Table 11. Comparison of the model complexity, latency and accuracy on SemanticKITTI val set. $\dagger$ represents our re-implementation on NVIDIA A100 for a fair comparison. "L", "C" and "T" represent LiDAR, Camera and Temporal data, respectively. + +# 5. Conclusion + +In this paper, we present TASeg to fully exploit temporal point clouds and temporal images. Our TASeg is comprised of TLAD, TIAF and SMSA. We perform extensive experiments to validate the efficacy of each component. Moreover, our TASeg set a new state-of-the-art on three challenging tracks, demonstrating the superiority of our method. + +# 6. Acknowledgements + +This work was supported in part by The National Nature Science Foundation of China (Grant Nos: 62273301, 62273302, 62036009, 61936006, 62273303), in part by Yongjiang Talent Introduction Programme (Grant No: 2022A-240-G, 2023A-197-G), in part by the National Key R&D Program of China (NO.2022ZD0160100). This work was done during the first author's internship at Shanghai AI Laboratory. + +# References + +[1] Mehmet Aygun, Aljosa Osep, Mark Weber, Maxim Maximov, Cyrill Stachniss, Jens Behley, and Laura Leal-Taixe. 4d panoptic lidar segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5527-5537, 2021. 2 +[2] Hanwen Cao, Yongyi Lu, Cewu Lu, Bo Pang, Gongshen Liu, and Alan Yuille. Asap-net: Attention and structure aware point cloud sequence segmentation. arXiv preprint arXiv:2008.05149, 2020. 2 +[3] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. arXiv preprint arXiv:2306.16927, 2023. 1 +[4] Xuechao Chen, Shuangjie Xu, Xiaoyi Zou, Tongyi Cao, Dit-Yan Yeung, and Lu Fang. Svqnet: Sparse voxel-adjacent query network for 4d spatio-temporal lidar semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8569–8578, 2023. 1, 2, 6, 7 +[5] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. (af)2-S3Net: Attentive Feature Fusion with Adaptive Feature Selection for Sparse Semantic Segmentation Network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12547-12556, 2021. 2, 6 +[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2 +[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 2, 6, 8 +[8] Ayush Dewan and Wolfram Burgard. Deeptemporalseg: Temporally consistent semantic segmentation of 3d lidar scans. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 2624-2630. IEEE, 2020. 2 +[9] Fabian Duerr, Mario Pfaller, Hendrik Weigel, and Jürgen Beyerer. Lidar-based recurrent 3d semantic segmentation with temporal memory alignment. In 2020 International Conference on 3D Vision (3DV), pages 781-790. IEEE, 2020. 1, 2, 6 +[10] Khaled El Madawi, Hazem Rashed, Ahmad El Sallab, Omar Nasr, Hanan Kamel, and Senthil Yogamani. Rgb and lidar fusion based 3d semantic segmentation for autonomous driving. In IEEE Intelligent Transportation Systems Conference, pages 7-12, 2019. 2 +[11] Kyle Genova, Xiaoqi Yin, Abhijit Kundu, Caroline Pantofaru, Forrester Cole, Avneesh Sud, Brian Brewington, Brian Shucker, and Thomas Funkhouser. Learning 3d semantic segmentation with only 2d image supervision. In 3DV, 2021. 7 +[12] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3 +[13] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-Voxel Knowledge Distillation for LiDAR Semantic Segmentation. In IEEE Conference on Computer + +Vision and Pattern Recognition, pages 8479-8488, 2022. 2, 3, 5, 6 +[14] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 6 +[15] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 228-240, 2023. 2 +[16] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. Lasermix for semi-supervised lidar semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21705-21715, 2023. 6 +[17] Lars Kreuzberg, Idil Esen Zulfikar, Sabarinath Mahadevan, Francis Engelmann, and Bastian Leibe. 4d-stop: Panoptic segmentation of 4d lidar using spatio-temporal object proposal generation and aggregation. In European Conference on Computer Vision, pages 537-553. Springer, 2022. 2 +[18] Georg Krispel, Michael Opitz, Georg Waltner, Horst Possegger, and Horst Bischof. Fuseseg: Lidar point cloud segmentation fusing multi-modal data. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1874-1883, 2020. 2 +[19] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. arXiv preprint arXiv:2303.12766, 2023. 1, 2, 6, 7 +[20] Jiale Li, Hang Dai, Hao Han, and Yong Ding. MSeg3D: Multi-modal 3D Semantic Segmentation for Autonomous Driving. arXiv preprint arXiv:2303.08600, 2023. 1, 2, 7, 8 +[21] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research, 20(6):837-854, 2023. 1 +[22] Jiahui Liu, Chirui Chang, Jianhui Liu, Xiaoyang Wu, Lan Ma, and Xiaojuan Qi. Mars3d: A plug-and-play motion-aware model for semantic segmentation on multi-scan 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9372-9381, 2023. 1, 2, 6 +[23] Youquan Liu, Yeqi Bai, Lingdong Kong, Runnan Chen, Yuenan Hou, Botian Shi, and Yikang Li. Pcseg: An open source point cloud segmentation codebase. https://github.com/PJLab-ADG/PCSeg, 2023.6 +[24] Youquan Liu, Runnan Chen, Xin Li, Lingdong Kong, Yuchen Yang, Zhaoyang Xia, Yeqi Bai, Xinge Zhu, Yuexin Ma, Yikang Li, et al. Uniseg: A unified multi-modal lidar segmentation network and the openpcseg codebase. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21662-21673, 2023. 1, 2, 6, 7, 8 +[25] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2019. 6 +[26] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification + +and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2 +[27] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in neural information processing systems, pages 5099-5108, 2017. 2 +[28] Herbert Robbins and Sutton Monro. A Stochastic Approximation Method. The annals of mathematical statistics, pages 400-407, 1951. 6 +[29] Radu Alexandru Rosu, Peer Schütt, Jan Quenzel, and Sven Behnke. Latticenet: Fast point cloud segmentation using permutohedral lattices. arXiv preprint arXiv:1912.05905, 2019.6 +[30] Hanyu Shi, Guosheng Lin, Hao Wang, Tzu-Yi Hung, and Zhenhua Wang. Spsequencenet: Semantic segmentation network on 4d point clouds. In CVPR, 2020. 2 +[31] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In European Conference on Computer Vision, pages 685-702. Springer, 2020. 2, 6, 7, 8 +[32] Jiexiong Tang, John Folkesson, and Patric Jensfelt. Sparse2dense: From direct sparse odometry to dense 3-d reconstruction. IEEE Robotics and Automation Letters, 4(2): 530-537, 2019. 3, 4 +[33] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In 2019 International Conference on Robotics and Automation (ICRA), pages 4376-4382. IEEE, 2019. 6 +[34] Dong Wu, Man-Wen Liao, Wei-Tian Zhang, Xing-Gang Wang, Xiang Bai, Wen-Qing Cheng, and Wen-Yu Liu. Yolop: You only look once for panoptic driving perception. Machine Intelligence Research, 19(6):550-562, 2022. 1 +[35] Aoran Xiao, Jiaxing Huang, Dayan Guan, Kaiwen Cui, Shijian Lu, and Ling Shao. PolarMix: A General Data Augmentation Technique for LiDAR Point Clouds. arXiv preprint arXiv:2208.00223, 2022. 6 +[36] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeezesegv3: Spatially-adaptive convolution for efficient point-cloud segmentation. In European Conference on Computer Vision, pages 1-19. Springer, 2020. 6 +[37] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. RPVnet: A Deep and Efficient Range-Point-Voxel Fusion Network for Lidar Point Cloud Segmentation. In IEEE International Conference on Computer Vision, pages 16024–16033, 2021. 6, 8 +[38] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3101-3109, 2021. 6 +[39] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2DPASS: 2D Priors + +Assisted Semantic Segmentation on LiDAR Point Clouds. In European Conference on Computer Vision, 2022. 1, 2, 3, 6, 7 +[40] Jihan Yang, Shaoshuai Shi, Runyu Ding, Zhe Wang, and Xiaojuan Qi. Towards efficient 3d object detection with knowledge distillation. Advances in Neural Information Processing Systems, 35:21300-21313, 2022. 3 +[41] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet: Towards a unified multi-task network for lidar perception. arXiv preprint arXiv:2209.09385, 2022. 7 +[42] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE access, 8:58443-58469, 2020. 1 +[43] Zhiwei Zhang, Zhizhong Zhang, Qian Yu, Ran Yi, Yuan Xie, and Lizhuang Ma. Lidar-camera panoptic segmentation via geometry-consistent and semantic-aware alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3662-3671, 2023. 2 +[44] Wu Zheng, Li Jiang, Fanbin Lu, Yangyang Ye, and Chi-Wing Fu. Boosting single-frame 3d object detection by simulating multi-frame point clouds. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4848-4856, 2022. 3, 4 +[45] Minghan Zhu, Shizhong Han, Hong Cai, Shubhankar Borse, Maani Ghaffari, and Fatih Porikli. 4d panoptic segmentation as invariant and equivariant field prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22488-22498, 2023. 2 +[46] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and Asymmetrical 3D Convolution Networks for Lidar Segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 9939-9948, 2021. 1, 2, 5, 6, 7, 8 +[47] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. In ICCV, 2021. 2 \ No newline at end of file diff --git a/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/images.zip b/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..912fb54a56a2d8b0b3fc582a1c784fe6bf3974ec --- /dev/null +++ b/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f39a2922e1752bf915c3c3b407284ed9f1de6f0118416171239a41710188167 +size 767014 diff --git a/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/layout.json b/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6c6d30c295809c97b51c9f45220dafcaf1a99ee7 --- /dev/null +++ b/2024/TASeg_ Temporal Aggregation Network for LiDAR Semantic Segmentation/layout.json @@ -0,0 +1,8231 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 65, + 103, + 531, + 121 + ], + "type": "text", + "content": "TASeg: Temporal Aggregation Network for LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "spans": [ + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": "Xiaopei Wu" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Yuenan Hou" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Xiaoshui Huang" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Binbin Lin" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Tong He" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Xinge Zhu" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Yuexin Ma" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Boxi Wu" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Haifeng Liu" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Deng Cai" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "text", + "content": ", Wanli Ouyang" + }, + { + "bbox": [ + 107, + 142, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 172, + 468, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 172, + 468, + 185 + ], + "spans": [ + { + "bbox": [ + 123, + 172, + 468, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 172, + 468, + 185 + ], + "type": "text", + "content": "State Key Lab of CAD&CG, Zhejiang University " + }, + { + "bbox": [ + 123, + 172, + 468, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 123, + 172, + 468, + 185 + ], + "type": "text", + "content": "Shanghai AI Laboratory " + }, + { + "bbox": [ + 123, + 172, + 468, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 123, + 172, + 468, + 185 + ], + "type": "text", + "content": "Fullong Inc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 186, + 522, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 186, + 522, + 199 + ], + "spans": [ + { + "bbox": [ + 67, + 186, + 522, + 199 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 186, + 522, + 199 + ], + "type": "text", + "content": "School of Software Technology, Zhejiang University " + }, + { + "bbox": [ + 67, + 186, + 522, + 199 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 67, + 186, + 522, + 199 + ], + "type": "text", + "content": "The Chinese University of Hong Kong " + }, + { + "bbox": [ + 67, + 186, + 522, + 199 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 67, + 186, + 522, + 199 + ], + "type": "text", + "content": "ShanghaiTech University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 251, + 290, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 290, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 290, + 624 + ], + "type": "text", + "content": "Training deep models for LiDAR semantic segmentation is challenging due to the inherent sparsity of point clouds. Utilizing temporal data is a natural remedy against the sparsity problem as it makes the input signal denser. However, previous multi-frame fusion algorithms fall short in utilizing sufficient temporal information due to the memory constraint, and they also ignore the informative temporal images. To fully exploit rich information hidden in long-term temporal point clouds and images, we present the Temporal Aggregation Network, termed TAseg. Specifically, we propose a Temporal LiDAR Aggregation and Distillation (TLAD) algorithm, which leverages historical priors to assign different aggregation steps for different classes. It can largely reduce memory and time overhead while achieving higher accuracy. Besides, TLAD trains a teacher injected with gt priors to distill the model, further boosting the performance. To make full use of temporal images, we design a Temporal Image Aggregation and Fusion (TIAF) module, which can greatly expand the camera FOV and enhance the present features. Temporal LiDAR points in the camera FOV are used as mediums to transform temporal image features to the present coordinate for temporal multi-modal fusion. Moreover, we develop a Static-Moving Switch Augmentation (SMSA) algorithm, which utilizes sufficient temporal information to enable objects to switch their motion states freely, thus greatly increasing static and moving training samples. Our TAseg ranks " + }, + { + "bbox": [ + 46, + 251, + 290, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{1}^{st\\dagger}" + }, + { + "bbox": [ + 46, + 251, + 290, + 624 + ], + "type": "text", + "content": " on three challenging tracks, i.e., SemanticKITTI single-scan track, multi-scan track and nuScenes LiDAR segmentation track, strongly demonstrating the superiority of our method. Codes are available at https://github.com/LittlePey/TASeg." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 642, + 128, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 128, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 128, + 654 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 662, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 662, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 662, + 287, + 687 + ], + "type": "text", + "content": "LiDAR segmentation aims to infer the semantic information of each point in point clouds and plays an indispensable role" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 228, + 547, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 547, + 396 + ], + "type": "text", + "content": "in the autonomous driving [3, 21, 34, 42]. With the advent of deep learning, a large quantity of LiDAR segmentation algorithms [14, 19, 24, 39, 46] have been put forward. Despite their impressive results, the segmentation performance is still constrained by the inherent sparsity of point clouds. To alleviate this, it is desirable to aggregate temporal data. Previous multi-frame models [4, 9, 22] can only fuse a few LiDAR frames due to the GPU memory constraint, which restricts them from utilizing rich information hidden in the long-term temporal point clouds. Although MSeg3D [20] stacks dozens of LiDAR frames, the memory and time overhead are intolerable. Moreover, previous attempts concentrate on utilizing temporal point clouds, ignoring the valuable information hidden in temporal images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 399, + 548, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 548, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 548, + 615 + ], + "type": "text", + "content": "To better leverage long-term temporal information, we propose a Temporal LiDAR Aggregation and Distillation (TLAD) algorithm. It can perform efficient multi-frame aggregation while achieving higher accuracy. TLAD consists of Flexible Step Aggregation (FSA) and Mask Distillation. FSA is based on the observation that the model needs different amounts of temporal points for different classes. For difficult classes, such as bicyclists, more points are needed to yield accurate predictions. However, for easy classes, such as cars and roads, it is unnecessary to aggregate too many point cloud frames for them. Therefore, we propose to assign different aggregation steps for different classes in point cloud sequences, which can significantly save memory and computation overhead on easy classes while providing sufficient temporal points for difficult classes. To identify classes of temporal point clouds, historical predictions are used, considering the temporal data is processed in chronological order, which is consistent with practical applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "content": "Interestingly, by assigning different classes with different steps, FSA actually injects historical priors into the aggregated point clouds. For classes with different steps, their patterns can be more discriminative due to their different densities. This makes the point cloud frames accumulated by FSA more conducive for segmentation. Experimental results show that FSA can not only reduce memory and time costs but also enhance overall performance. To further verify our" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 142, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 142, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 142, + 703 + ], + "type": "text", + "content": "* Corresponding authors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 58, + 703, + 273, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 273, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 273, + 712 + ], + "type": "text", + "content": "† On the date of CVPR deadline, i.e., 2023-11-18 07:59 AM UTC." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15311" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "idea, we directly replace historical predictions of temporal point clouds with their ground truth labels for FSA. We find the performance boosts greatly. This inspires us to use the model injected with gt priors to distill the model injected with historical priors, which we term Mask Distillation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 290, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 290, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 290, + 407 + ], + "type": "text", + "content": "To fully use the informative temporal images, we devise a Temporal Image Aggregation and Fusion (TIAF) module. Previous multi-modal fusion methods usually suffer from limited image features due to the different FOVs (field of views) between the LiDAR and the camera. We observe that as the ego-vehicle moves forward, cameras can capture different FOVs. By aggregating images of different timestamps, the FOV of the present camera can be enlarged greatly. In addition, temporal images can provide different views for the same region at different timestamps, which can supply richer information. Our TIAF leverages temporal LiDAR points in the FOV of the corresponding temporal camera to transform historical image features to the present coordinate with the pose matrix. After aggregating all temporal image features, we use a series of 3D sparse convolutions to fuse them and convert them to voxel representation. Ultimately, we use temporal LiDAR points to gather temporal image features for temporal multi-modal fusion. The fused features embrace both temporal LiDAR and temporal image information, which contribute to more precise segmentation results. To our knowledge, we are the first to leverage temporal point clouds and temporal images simultaneously for LiDAR semantic segmentation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 408, + 288, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 288, + 540 + ], + "type": "text", + "content": "Moreover, in the multi-scan task, the model needs to distinguish the motion states (moving or static) of a movable object, which is very challenging. In this paper, we make full use of temporal information to remarkably increase the multi-scan perception ability of the model. Specifically, we propose a data augmentation strategy named Static-Moving Switch Augmentation (SMSA). It manipulates temporal point clouds of a movable object to switch its motion state. In this way, we can increase the quantity of static and moving samples significantly, even in the absence of training samples of static or moving classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 540, + 288, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 540, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 58, + 540, + 288, + 552 + ], + "type": "text", + "content": "In summary, our major contributions are listed as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 559, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 559, + 288, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 288, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 288, + 608 + ], + "type": "text", + "content": "- We propose a Temporal LiDAR Aggregation and Distillation algorithm, which uses Flexible Step Aggregation and Mask Distillation techniques to largely reduce memory and time costs while achieving higher accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 612, + 288, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 612, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 612, + 288, + 661 + ], + "type": "text", + "content": "- We devise a Temporal Image Aggregation and Fusion module, which exploits temporal images to enlarge the camera FOV and enhance present features. It also delivers a scheme for temporal multi-modal fusion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "content": "- We design a Static-Moving Switch Augmentation algorithm to enable static and moving objects to switch their motion states freely. With this technique, we can greatly increase static and moving training samples." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 133 + ], + "type": "text", + "content": "- Our Temporal Aggregation Network, dubbed TASeg, achieves impressive results on SemanticKITTI and nuScenes benchmarks. Notably, our TASeg ranks " + }, + { + "bbox": [ + 306, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 306, + 72, + 547, + 133 + ], + "type": "text", + "content": " on three challenging tracks. Thorough ablation studies are provided to demonstrate the efficacy of our approach." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 146, + 392, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 146, + 392, + 159 + ], + "spans": [ + { + "bbox": [ + 306, + 146, + 392, + 159 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 167, + 548, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 548, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 548, + 348 + ], + "type": "text", + "content": "LiDAR Semantic Segmentation. LiDAR semantic segmentation aims to assign a unique class label to each point in the input point cloud sequence. Recent years have witnessed an explosion of LiDAR segmentation algorithms [5, 7, 13, 15, 19, 20, 24, 27, 31, 39]. For example, [26] is the pioneering work that approximates the permutation-invariant function with a per-point and shared Multi-Layer Perceptron. [46] changes traditional cubic grids to cylindrical grids and designs a network of asymmetrical 3D convolutions. [19] divides the space with the radial window, which increases the receptive field smoothly and helps improve the performance. Despite their good segmentation performance, these methods still take the single LiDAR frame as input, which does not utilize the rich semantic and geometric information hidden in temporal data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 357, + 548, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 548, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 548, + 537 + ], + "type": "text", + "content": "Multi-Frame LiDAR Perception. Compared to a single LiDAR scan, multiple LiDAR scans can provide more sufficient information. Recently, many research efforts have been put on temporal LiDAR segmentation [1, 2, 6, 8, 9, 17, 20, 22, 30, 45]. For example, [8] leverages a Bayes filter to explore the temporal consistency. [4] shunts the historical points into two groups to utilize historical frames efficiently. Despite the success of previous multi-frame methods, they can not leverage the valuable information hidden in long-term temporal point clouds due to the GPU memory constraint. Although [20] stacks dozens of LiDAR scans, the memory and time overhead are intolerable. In this paper, we present an efficient multi-frame aggregation algorithm, which can greatly save memory and computation consumption while achieving higher performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 548, + 715 + ], + "type": "text", + "content": "Multi-Modal Fusion. Since LiDAR and camera are two complementary sensors for 3D semantic segmentation, multimodal fusion has gained increasing attention in recent years [10, 18, 24, 43, 47]. However, these multi-modal fusion methods usually suffer from limited overlapped regions between the LiDAR and camera due to their different FOVs. [39] proposes a cross-modal knowledge distillation method, which is free from images at inference, while it causes much loss of the RGB information. [20] completes the missing camera features using predicted pseudo-camera features, while the used image information is still restricted to the present camera FOV. In this paper, we fully take advantage of temporal images to enlarge the FOV of the present camera and enhance present image features." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15312" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "Knowledge Distillation for LiDAR Perception. Knowledge distillation [12] is widely used in various fields in that it can improve the performance of the student without sacrificing inference efficiency. In LiDAR perception, [13, 40] compress a cumbersome teacher to a lightweight student to reinforce the representation learning of the student as well as maintain high efficiency. [39] distills the prior of 2D images to 3D point clouds with well-designed cross-modal knowledge distillation module. [32, 44] utilizes a multi-frame teacher to help a single-frame student learn dense 3D features. By contrast, our method aims to transfer knowledge from a multi-frame teacher injected with gt priors to a multi-frame student injected with historical priors." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 238, + 130, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 238, + 130, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 130, + 251 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 258, + 287, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 287, + 270 + ], + "type": "text", + "content": "3.1. Temporal LiDAR Aggregation and Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 276, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 287, + 312 + ], + "type": "text", + "content": "A simple solution to utilize the temporal information of a consecutive point cloud sequence is to concatenate all points of the sequence, as shown in the following equation:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 319, + 287, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 319, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 70, + 319, + 287, + 338 + ], + "type": "interline_equation", + "content": "X _ {t} = \\operatorname {c o n c a t} \\left(P _ {t}, T _ {t - 1} P _ {t - 1}, \\dots , T _ {t - \\Delta t} P _ {t - \\Delta t}\\right), \\tag {1}", + "image_path": "72c34f394d48dc7f3adc422287fba4a9b8fab24a486b8afa40d71ac8c976b260.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 75, + 335, + 219, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 335, + 219, + 348 + ], + "spans": [ + { + "bbox": [ + 75, + 335, + 219, + 348 + ], + "type": "interline_equation", + "content": "Y _ {t} = \\operatorname {c o n c a t} \\left(L _ {t}, L _ {t - 1}, \\dots , L _ {t - \\Delta t}\\right),", + "image_path": "8808c0b8b34f048dc1af4419a5de40afdd5fd0985151f5895df4244774b16d1b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "inline_equation", + "content": "P_{t - i}" + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "inline_equation", + "content": "L_{t - i}" + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "inline_equation", + "content": "(t - i)^{th}" + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "text", + "content": " point cloud frame and the corresponding point-wise label. " + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "inline_equation", + "content": "T_{t - i}" + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "text", + "content": " is the transformation matrix that transforms the coordinate from the " + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "inline_equation", + "content": "(t - i)^{th}" + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "text", + "content": " frame to the " + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "inline_equation", + "content": "t^{th}" + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "text", + "content": " frame. " + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "text", + "content": " is the window size of temporal point clouds and " + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "inline_equation", + "content": "\\text{concat}(.)" + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "text", + "content": " denotes the concatenation operation. " + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "inline_equation", + "content": "Y_{t}" + }, + { + "bbox": [ + 46, + 357, + 289, + 489 + ], + "type": "text", + "content": " are the aggregated LiDAR frame and point-wise label. Although simple, there are some problems. On the one hand, direct concatenation consumes much GPU memory. On the other hand, the huge memory cost constrains the multi-frame model from utilizing more historical frames, thereby limiting the ultimate performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 503, + 190, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 190, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 190, + 516 + ], + "type": "text", + "content": "3.1.1 Flexible Step Aggregation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": "To reduce the GPU memory consumption, we can sample temporal frames with a step, while this method is also suboptimal. A small step introduces huge memory overhead, whilst a large step cannot utilize sufficient temporal information. To this end, we propose Flexible Step Aggregation (FSA). Our method is based on the observation that for different classes, the model needs different amounts of temporal information. For difficult classes, such as bicycles, more points are needed to yield accurate predictions. For easy classes, such as cars and buildings, it is unnecessary to aggregate many point cloud frames for them. Based on the above analysis, we propose to assign different steps for different classes according to their learning difficulty. Specifically, we leverage historical predictions to divide the temporal point clouds into several class groups and assign a specific sampling step for each group. Then, we aggregate temporal" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "text", + "content": "points for each group with the corresponding step. Eventually, temporal points of all groups are concatenated with the current frame, resulting in an aggregated frame." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 108, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 547, + 240 + ], + "type": "text", + "content": "The group division is not strict as long as it follows the principle that more difficult classes need smaller steps. For example, we can simply divide all classes into three groups according to their segmentation performance, such as [0, 80) mIoU, [80, 90) mIoU and [90, 100] mIoU. Then, we assign the three groups with a step of 2, 4 and " + }, + { + "bbox": [ + 304, + 108, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 304, + 108, + 547, + 240 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 108, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 304, + 108, + 547, + 240 + ], + "type": "text", + "content": " means we do not aggregate temporal points for the group). To save more memory and computation without sacrificing performance, a more fine-grained division can be used. Formally, suppose we divide temporal point clouds into " + }, + { + "bbox": [ + 304, + 108, + 547, + 240 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 108, + 547, + 240 + ], + "type": "text", + "content": " groups. We aggregate temporal points for the " + }, + { + "bbox": [ + 304, + 108, + 547, + 240 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 304, + 108, + 547, + 240 + ], + "type": "text", + "content": " group as below:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 332, + 245, + 547, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 245, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 332, + 245, + 547, + 266 + ], + "type": "interline_equation", + "content": "X _ {t} ^ {k} = \\operatorname {c o n c a t} \\left(T _ {o _ {1}} P _ {o _ {1}} M _ {o _ {1}} ^ {k}, \\dots , T _ {o _ {n}} P _ {o _ {n}} M _ {o _ {n}} ^ {k}\\right), \\tag {2}", + "image_path": "18d7696a201775a2373e5e61d27c00cc46b01c94e549e8c15cb14417b29d038b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 335, + 262, + 489, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 262, + 489, + 277 + ], + "spans": [ + { + "bbox": [ + 335, + 262, + 489, + 277 + ], + "type": "interline_equation", + "content": "Y _ {t} ^ {k} = \\operatorname {c o n c a t} \\left(L _ {o _ {1}} M _ {o _ {1}} ^ {k}, \\dots , L _ {o _ {n}} M _ {o _ {n}} ^ {k}\\right).", + "image_path": "b1eb20a7935af467b2c877d48a1c19e85ffeadb3e40a39dd6f0aee12b64ab53c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 283, + 547, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 547, + 355 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 283, + 547, + 355 + ], + "type": "inline_equation", + "content": "o = \\{o_i|o_i = t - i\\times s_k,i = 1,2,\\dots,n,n = \\lfloor \\Delta t / s_k\\rfloor \\}" + }, + { + "bbox": [ + 304, + 283, + 547, + 355 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 304, + 283, + 547, + 355 + ], + "type": "text", + "content": " is the sampling step for the " + }, + { + "bbox": [ + 304, + 283, + 547, + 355 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 304, + 283, + 547, + 355 + ], + "type": "text", + "content": " class group and " + }, + { + "bbox": [ + 304, + 283, + 547, + 355 + ], + "type": "inline_equation", + "content": "\\lfloor .\\rfloor" + }, + { + "bbox": [ + 304, + 283, + 547, + 355 + ], + "type": "text", + "content": " is the floor operation. " + }, + { + "bbox": [ + 304, + 283, + 547, + 355 + ], + "type": "inline_equation", + "content": "M_{o_i}^k" + }, + { + "bbox": [ + 304, + 283, + 547, + 355 + ], + "type": "text", + "content": " is the group mask that indicates which point of " + }, + { + "bbox": [ + 304, + 283, + 547, + 355 + ], + "type": "inline_equation", + "content": "P_{o_i}" + }, + { + "bbox": [ + 304, + 283, + 547, + 355 + ], + "type": "text", + "content": " belongs to the " + }, + { + "bbox": [ + 304, + 283, + 547, + 355 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 304, + 283, + 547, + 355 + ], + "type": "text", + "content": " class group. It can be obtained from historical predictions. Finally, we concatenate points and labels of all groups with the current frame:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 343, + 361, + 547, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 361, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 343, + 361, + 547, + 381 + ], + "type": "interline_equation", + "content": "X _ {t} = \\operatorname {c o n c a t} \\left(P _ {t}, X _ {t} ^ {1}, \\dots , X _ {t} ^ {k}, \\dots , X _ {t} ^ {g}\\right), \\tag {3}", + "image_path": "7817d2417af1935865b133c895f29cafa82852ca58da8ca6bd72da6a672970fa.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 347, + 378, + 506, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 378, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 347, + 378, + 506, + 392 + ], + "type": "interline_equation", + "content": "Y _ {t} = \\operatorname {c o n c a t} \\left(L _ {t}, Y _ {t} ^ {- 1}, \\dots , X _ {t} ^ {k}, \\dots , Y _ {t} ^ {g}\\right).", + "image_path": "0cf6cbe01e496668a2c087b4fe53c4cdb24f7360683630141672f5e5b095274a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 398, + 547, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 547, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 547, + 541 + ], + "type": "text", + "content": "In this way, we discard massive redundant temporal points while maintaining essential temporal information. Since easy classes are usually large and hold a large quantity of points, the GPU memory overhead can be further reduced. Moreover, by assigning different classes with different steps, FSA actually injects historical priors into the aggregated point clouds. For classes with different steps, their patterns can be more discriminative due to their different densities (steps), which makes it easier to segment the multi-frame point clouds aggregated by FSA. Experiments in Table 4 verify that the proposed FSA can not only save memory and time costs but also achieve better performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 551, + 412, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 412, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 412, + 562 + ], + "type": "text", + "content": "3.1.2 Mask Distillation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "In FSA, we use historical predictions to generate group masks, which we call pseudo masks. Since ground truth labels are more accurate than historical predictions, a natural question arises: what if we use gt masks (the group masks generated by ground truth labels) for FSA? Our experiment shows that the performance can be improved greatly. Actually, using gt masks for FSA can produce more discriminative patterns for classes with different steps. This motivates us to simulate features of a model trained on temporal point clouds aggregated with gt masks. In this way, our model is guided to learn more discriminative features to distinguish different classes. We call this Mask Distillation. Specifically," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15313" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 73, + 541, + 340 + ], + "blocks": [ + { + "bbox": [ + 53, + 73, + 541, + 340 + ], + "lines": [ + { + "bbox": [ + 53, + 73, + 541, + 340 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 541, + 340 + ], + "type": "image", + "image_path": "9b15e5bcf6fe62695fdcac9bb7297f21a8bc49fd7ab50e8acab5c209ffef3555.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 347, + 548, + 404 + ], + "lines": [ + { + "bbox": [ + 46, + 347, + 548, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 548, + 404 + ], + "type": "text", + "content": "Figure 1. Overview of our Temporal Aggregation Network (TASeg). (1) Temporal LiDAR Aggregation and Distillation leverages the proposed Flexible Step Aggregation (FSA) to assign different temporal steps for different classes, and it utilizes a teacher injected with gt priors for knowledge distillation. (2) Temporal Image Aggregation and Fusion takes temporal LiDAR points as mediums to transform historical image features to the present coordinate. 3D sparse convolutions are employed to fuse temporal image features. Finally, we use temporal LiDAR points to gather voxel-wise temporal image features for temporal multi-modal fusion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 410, + 290, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 410, + 290, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 410, + 290, + 436 + ], + "type": "text", + "content": "we formulate the distillation on the feature map of student " + }, + { + "bbox": [ + 47, + 410, + 290, + 436 + ], + "type": "inline_equation", + "content": "F^{s}\\in \\mathbb{R}^{N^{s}\\times C}" + }, + { + "bbox": [ + 47, + 410, + 290, + 436 + ], + "type": "text", + "content": " and the feature map of teacher " + }, + { + "bbox": [ + 47, + 410, + 290, + 436 + ], + "type": "inline_equation", + "content": "F^{t}\\in \\mathbb{R}^{N^{t}\\times C}" + }, + { + "bbox": [ + 47, + 410, + 290, + 436 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 445, + 288, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 445, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 288, + 459 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {K D}} = \\mathbb {E} \\left[ \\left\\| F ^ {s} m ^ {s} - F ^ {t} m ^ {t} \\right\\| _ {2} \\right]. \\tag {4}", + "image_path": "f0d74a08b66e8088d6e24076269c0f372fa0e089f777171d1d0fc477f26bf548.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 461, + 288, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 288, + 534 + ], + "type": "text", + "content": "Since temporal point clouds aggregated with pseudo masks and gt masks are different, we use masks " + }, + { + "bbox": [ + 46, + 461, + 288, + 534 + ], + "type": "inline_equation", + "content": "m^s" + }, + { + "bbox": [ + 46, + 461, + 288, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 461, + 288, + 534 + ], + "type": "inline_equation", + "content": "m^t" + }, + { + "bbox": [ + 46, + 461, + 288, + 534 + ], + "type": "text", + "content": " to select voxels that appeared in both " + }, + { + "bbox": [ + 46, + 461, + 288, + 534 + ], + "type": "inline_equation", + "content": "F^s" + }, + { + "bbox": [ + 46, + 461, + 288, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 461, + 288, + 534 + ], + "type": "inline_equation", + "content": "F^t" + }, + { + "bbox": [ + 46, + 461, + 288, + 534 + ], + "type": "text", + "content": ". Note that our Mask Distillation is distinct from the methods that distill a multi-frame model to a single-frame model [32, 44]. Both the student and teacher in our method are multi-frame models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 540, + 263, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 263, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 263, + 553 + ], + "type": "text", + "content": "3.2. Temporal Image Aggregation and Fusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 288, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 653 + ], + "type": "text", + "content": "Previous multi-modal fusion methods only focus on leveraging present images while ignoring the precious value of temporal images. Temporal images can provide broader camera FOVs and richer information. Besides, they can enable more robust multi-modal fusion under the malfunction condition on some cameras. In this section, we provide an effective solution for aggregating temporal image features and performing temporal multi-modal fusion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 289, + 715 + ], + "type": "text", + "content": "Temporal Aggregation and Fusion. Since temporal images are in different feature spaces, it is difficult to establish the relationship between different images for feature aggregation. In our method, we take temporal LiDAR points as mediums to transform temporal image features to the present" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "text", + "content": "coordinate with the pose information. This way, temporal image features are unified to the present 3D space. Specifically, given an image " + }, + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "inline_equation", + "content": "I_{t - \\Delta t}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "text", + "content": " and point cloud " + }, + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "inline_equation", + "content": "P_{t - \\Delta t}\\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "text", + "content": ", we use an image network to extract the image feature " + }, + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "inline_equation", + "content": "Z_{t - \\Delta t}\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "text", + "content": ". According to the sensor calibration, we can establish the pixel-to-point mapping between 2D pixels and 3D points. Hence, we can project image feature " + }, + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "inline_equation", + "content": "Z_{t - \\Delta t}" + }, + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "text", + "content": " to 3D space, resulting in point-wise image feature " + }, + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "inline_equation", + "content": "Q_{t - \\Delta t}\\in \\mathbb{R}^{M\\times C}" + }, + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "text", + "content": " is the number of LiDAR points located on " + }, + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "inline_equation", + "content": "Z_{t - \\Delta t}" + }, + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "text", + "content": ". By transforming " + }, + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "inline_equation", + "content": "Q_{t - \\Delta t}" + }, + { + "bbox": [ + 304, + 410, + 547, + 555 + ], + "type": "text", + "content": " to the present coordinate with the pose matrix, we realize the aggregation of temporal image features:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 342, + 559, + 547, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 559, + 547, + 574 + ], + "spans": [ + { + "bbox": [ + 342, + 559, + 547, + 574 + ], + "type": "interline_equation", + "content": "\\bar {X} _ {t} = \\operatorname {c o n c a t} \\left(Q _ {t}, T _ {o _ {1}} Q _ {o _ {1}}, \\dots , T _ {o _ {n}} Q _ {o _ {n}}\\right), \\tag {5}", + "image_path": "b279b90c01e27d51590c83a52ac27a93d9b6dbc290a495ba0d65390129202ee6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "o = \\{o_i|o_i = t - i\\times s,i = 1,2,\\dots,\\lfloor \\Delta t / s\\rfloor \\}" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": " is the sampling step for temporal images. With temporal images aggregated, the FOV of the present camera is expanded greatly. Moreover, temporal image feature fusion becomes convenient because they are unified to the same 3D space. Concretely, we can use several 3D sparse convolutions to fuse aggregated temporal image features, which also endows them with geometric information, as shown in Equation 6. In addition, feature map downsampling is also utilized to generate multi-scale voxel features, providing richer information for subsequent temporal multi-modal fusion." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15314" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 84, + 83, + 288, + 98 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 83, + 288, + 98 + ], + "spans": [ + { + "bbox": [ + 84, + 83, + 288, + 98 + ], + "type": "interline_equation", + "content": "\\bar {V} _ {t} = \\operatorname {S p a r s e C o n v} _ {\\mathrm {3 D}} \\left(\\operatorname {V o x e l i z a t i o n} \\left(\\bar {X} _ {t}\\right)\\right). \\tag {6}", + "image_path": "0b81c188640104ea388b26bf67e2ad6f31950df3fb2b41a3ba6310432ae10537.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 102, + 289, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 102, + 289, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 102, + 289, + 270 + ], + "type": "text", + "content": "Temporal Multi-Modal Fusion. Benefiting from the temporal image aggregation, associating temporal image features with temporal LiDAR points also becomes convenient. Specifically, given temporal image features converted to unified voxel representation " + }, + { + "bbox": [ + 46, + 102, + 289, + 270 + ], + "type": "inline_equation", + "content": "\\overline{V}_t" + }, + { + "bbox": [ + 46, + 102, + 289, + 270 + ], + "type": "text", + "content": ", we can establish a point-to-voxel association between temporal LiDAR points and " + }, + { + "bbox": [ + 46, + 102, + 289, + 270 + ], + "type": "inline_equation", + "content": "\\overline{V}_t" + }, + { + "bbox": [ + 46, + 102, + 289, + 270 + ], + "type": "text", + "content": ". For each temporal LiDAR point, we generate its image feature by pooling its nearby voxel-wise temporal image features with trilinear interpolation instead of hard indexing. To extract richer image features, we perform interpolation on multi-scale feature maps. Finally, we concatenate the point cloud features and aggregated multi-scale image features, resulting in fused features, which convey powerful information of both temporal point clouds and temporal images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 275, + 288, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 288, + 348 + ], + "type": "text", + "content": "2D and 3D Supervision. To make the extracted image features more informative, we add 2D supervision and 3D supervision on the 2D backbone and 3D convolutions in the image branch, respectively. The 3D supervision is just the label of point clouds. The 2D supervision is obtained by projecting labels of point clouds to the image plane." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 354, + 240, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 240, + 368 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 240, + 368 + ], + "type": "text", + "content": "3.3. Static-Moving Switch Augmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "text", + "content": "In the multi-scan task, the model is required to distinguish the motion state of movable objects. To enable the model to explore a large data space, we design an effective data augmentation, Static-Moving Switch Augmentation (SMSA). SMSA enables a movable object to switch its motion state freely, which can remarkably increase the sample quantity of static and moving objects. Concretely, considering that a unique object " + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "text", + "content": " has the same instance id in all frames of a sequence, we can use its instance mask to crop its temporal point clouds, which is denoted as " + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{\\mathcal{B}_{o_i}|i = 0,1,\\dots,\\lfloor \\Delta t / s\\rfloor ,\\mathcal{B}_{o_i}\\in \\mathbb{R}^{N\\times 3}\\}" + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "inline_equation", + "content": "o = \\{o_i|o_i = t - i\\times s,i = 1,2,\\dots,\\lfloor \\Delta t / s\\rfloor \\}" + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "text", + "content": " is the sampling step and " + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{o_i}" + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "text", + "content": " is the temporal part of " + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "inline_equation", + "content": "t_{o_i}" + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "text", + "content": ". By manipulating " + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "text", + "content": ", we can change the motion state of " + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 373, + 288, + 540 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "Static to Moving. If " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " is static, all temporal parts of " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " locate at the same position, as shown in the upper-left of Figure 2. To change " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " to a moving object, we can shift each temporal part of " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " with an offset. Considering that objects typically move at a constant speed within a short time, we set the offset between adjacent temporal parts to be the same. For offset itself, it is a random value to increase the diversity of resulting moving samples. As for the direction of the offset, it can be roughly estimated by comparing the width and height of the minimum bounding box of " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": ". Since static objects often park on the side of the road, which is crowded, the shifted temporal parts of static objects may overlap with other objects. To alleviate this, we define a set of anchor points around " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " and a coverage area for each anchor point." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 312, + 76, + 547, + 200 + ], + "blocks": [ + { + "bbox": [ + 312, + 76, + 547, + 200 + ], + "lines": [ + { + "bbox": [ + 312, + 76, + 547, + 200 + ], + "spans": [ + { + "bbox": [ + 312, + 76, + 547, + 200 + ], + "type": "image", + "image_path": "a2410a010fdcb1d5c563c6b1ae14acc2326ee362bd875d1ed10945435cc43774.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 206, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 304, + 206, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 547, + 240 + ], + "type": "text", + "content": "Figure 2. Visualization of the augmented samples by our Static-Moving Switch Augmentation (SMSA). Our SMSA switches the motion state by manipulating the temporal parts of objects." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 245, + 545, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 545, + 282 + ], + "type": "text", + "content": "Then, we shift all temporal parts of " + }, + { + "bbox": [ + 304, + 245, + 545, + 282 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 245, + 545, + 282 + ], + "type": "text", + "content": " to the anchor point whose coverage area contains the fewest LiDAR points (refer to the supplementary material for more details)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "text", + "content": "Moving to Static. If " + }, + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "text", + "content": " is moving, different temporal parts of " + }, + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "text", + "content": " locate at different positions. To switch its motion state to static, we can shift all its temporal parts to the same position. In particular, we calculate the centers of each temporal part of " + }, + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "text", + "content": ", which are denoted as " + }, + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{\\mathcal{C}_{o_i} | \\mathcal{C}_{o_i} \\in \\mathbb{R}^3\\}" + }, + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "text", + "content": ". Considering the trajectory of moving objects in a short time is approximate to a line and the speed is a constant, we simply use " + }, + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "inline_equation", + "content": "q = \\mathcal{C}_{o_0} - \\mathcal{C}_{o_1}" + }, + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "text", + "content": " as the offset of adjacent temporal parts of " + }, + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "text", + "content": ". Eventually, we can obtain a static object by shifting all temporal parts of " + }, + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "inline_equation", + "content": "C_{o_0}" + }, + { + "bbox": [ + 304, + 285, + 547, + 405 + ], + "type": "text", + "content": " (see Figure 2)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 407, + 547, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 547, + 443 + ], + "type": "text", + "content": "It should be noted that our SMSA is an online and plug-and-play data augmentation strategy, which consumes negligible storage and computation costs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 449, + 417, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 417, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 417, + 462 + ], + "type": "text", + "content": "3.4. Training Objective" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 468, + 547, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 547, + 504 + ], + "type": "text", + "content": "The final training objective is comprised of LiDAR segmentation loss, mask distillation loss, fusion segmentation loss, 2D and 3D supervision loss on the image branch:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 506, + 545, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 506, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 318, + 506, + 545, + 520 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {L i D A R}} + \\alpha \\mathcal {L} _ {\\mathrm {K D}} + \\beta \\mathcal {L} _ {\\text {F u s i o n}} + \\gamma (\\mathcal {L} _ {\\mathrm {2 D}} + \\mathcal {L} _ {\\mathrm {3 D}}), \\tag {7}", + "image_path": "7ca8fcaba51cad82212367ff2430112b6b15fd3aafe81534a617314e7a23af99.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 527, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 527, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 547, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 527, + 547, + 552 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta, \\gamma" + }, + { + "bbox": [ + 304, + 527, + 547, + 552 + ], + "type": "text", + "content": " are the coefficients to control the effect of different losses. We set " + }, + { + "bbox": [ + 304, + 527, + 547, + 552 + ], + "type": "inline_equation", + "content": "\\alpha = 1, \\beta = 1, \\gamma = 1" + }, + { + "bbox": [ + 304, + 527, + 547, + 552 + ], + "type": "text", + "content": " by default." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 562, + 388, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 388, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 388, + 575 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "Datasets & Evaluation Metrics. Following [13, 46], we evaluate the performance on two popular LiDAR segmentation benchmarks, i.e., SemanticKITTI and nuScenes. SemanticKITTI contains 22 point cloud sequences, where sequences 0-10, 8, 11-21 are selected as training, validation and testing, respectively. As for nuScenes, it collects 1, 000 driving scenes, where 850 scenes are chosen for training and validation, and the remaining 150 scenes are used for testing. We conduct experiments on three tracks, i.e., SemanticKITTI single-scan and multi-scan semantic segmentation and nuScenes semantic segmentation." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15315" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 545, + 299 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 299 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 299 + ], + "type": "table", + "html": "
MethodmIoUcarbicyclemotorcycletruckother-veh.personbicyclistmotorcyclistroadparkingsidewalkother-gro.buildingfencevegetationtrunkterrainpoletraffic sign
SqueezeSegV2 [33]39.781.818.517.913.414.020.125.13.988.645.867.617.773.741.171.835.860.220.226.3
RangeNet53++ [25]52.291.425.734.425.723.038.338.84.891.865.075.227.887.458.680.555.164.647.955.9
RandLA-Net [14]55.994.229.832.243.939.148.447.49.490.561.874.024.589.760.483.863.668.651.050.7
SqueezeSegV3 [36]55.992.538.736.529.633.045.646.220.191.763.474.826.489.059.482.058.765.449.658.9
JS3C-Net [38]66.095.859.352.954.346.069.565.439.988.961.972.131.992.570.884.569.867.960.768.7
SPVNAS [31]67.097.250.650.456.658.067.467.150.390.267.675.421.891.666.986.173.471.064.367.3
Cylinder3D [46]68.997.167.663.850.858.573.769.248.092.265.077.032.390.766.585.672.569.862.466.2
RPVNet [37]70.397.668.468.744.261.175.974.443.493.470.380.733.393.572.186.575.171.764.861.4
(AF)2-S3Net [5]70.894.363.081.440.240.076.481.777.792.066.876.245.892.569.678.668.063.164.073.3
PVKD [13]71.297.069.353.567.960.275.173.550.591.877.570.941.092.469.486.573.871.964.965.8
2DPASS [39]72.997.063.663.461.161.577.981.374.189.767.474.740.093.572.986.273.971.065.070.4
SphereFormer [19]74.897.570.170.559.667.779.080.475.391.869.778.241.393.872.886.775.172.466.872.9
UniSeg [24]75.297.971.975.263.674.178.974.860.692.674.079.546.193.472.787.576.373.168.368.5
TASeg (Ours)76.597.771.871.465.278.779.984.678.691.674.078.039.393.573.486.675.071.769.673.8
", + "image_path": "27f1dd895fec9f632f5cad48f04f57d61a6ae8de8b545a0e7f2c2e2613f91dff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 122, + 323, + 473, + 449 + ], + "blocks": [ + { + "bbox": [ + 52, + 306, + 539, + 317 + ], + "lines": [ + { + "bbox": [ + 52, + 306, + 539, + 317 + ], + "spans": [ + { + "bbox": [ + 52, + 306, + 539, + 317 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-arts on SemanticKITTI test set (single-scan). All results can be found on the online leaderboard." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 122, + 323, + 473, + 449 + ], + "lines": [ + { + "bbox": [ + 122, + 323, + 473, + 449 + ], + "spans": [ + { + "bbox": [ + 122, + 323, + 473, + 449 + ], + "type": "table", + "html": "
MethodmIoUcar-scar-mtruc.-struc.-mothe.-sothe.-mpers.-spers.-mbicy.-sbicy.-mmotor.-smotor.-m
LatticeNet [29]45.291.154.829.73.523.10.66.849.90.044.60.064.3
TemporalLidarSeg [9]47.092.168.239.22.135.012.414.440.40.042.80.012.9
\\( (AF)^2-S3Net[5] \\)56.991.865.315.75.627.53.916.467.615.166.467.159.6
MarS3D [22]52.795.178.439.75.136.610.016.258.01.267.30.036.3
SVQNet [4]60.596.080.141.05.160.47.128.785.10.072.40.088.1
2DPASS [39]62.496.282.148.216.152.73.835.480.37.971.262.073.1
TASeg (Ours)65.794.872.853.635.271.453.236.780.432.475.212.389.6
", + "image_path": "8a4ff21bd597ca0fddc813029ac0406dadc8330c962e35e4934336925dee7e73.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 455, + 543, + 467 + ], + "lines": [ + { + "bbox": [ + 48, + 455, + 543, + 467 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 543, + 467 + ], + "type": "text", + "content": "Table 2. Comparison to the state-of-the-art methods on SemanticKITTI test set (multi-scan). -s indicates static and -m stands for moving." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 476, + 290, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 476, + 290, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 290, + 657 + ], + "type": "text", + "content": "Implementation Details. We use the MinkowskiNet [7] re-implemented by PCSeg [23] as our baseline. Our TASeg is trained with the SGD [28] optimizer on 4 A100 GPUs with batch size 6 for 12 and 36 epochs on SemanticKITTI and nuScenes datasets. The learning rate and weight decay are set to 0.02 and 0.0001. The window size of temporal point clouds is set to 16. Due to a large redundancy of images, we use a step of 12 and a window size of 48 for temporal images. Our data augmentation strategy includes random flipping, rotation, scaling, transformation, LaserMix [16] and PolarMix [35]. During the inference, the LiDAR semantic predictions and image features of the previous timestamp can be saved, so there is no time cost for generating historical predictions and processing temporal image aggregation. More details are provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 669, + 270, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 270, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 270, + 682 + ], + "type": "text", + "content": "4.1. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "text", + "content": "We compare our TASeg with state-of-the-art LiDAR segmentation methods on SemanticKITTI single-scan and multi" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 476, + 547, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 547, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 547, + 632 + ], + "type": "text", + "content": "scan track, and nuScenes LiDAR semantic segmentation track, as summarized in Table 1, 2 and 3. On the SemanticKITTI single-scan track, our TASeg is " + }, + { + "bbox": [ + 304, + 476, + 547, + 632 + ], + "type": "inline_equation", + "content": "1.3\\mathrm{mIoU}" + }, + { + "bbox": [ + 304, + 476, + 547, + 632 + ], + "type": "text", + "content": " higher than UniSeg [24], the best-performing published algorithm on the SemanticKITTI single-scan leaderboard. On the SemanticKITTI multi-scan track, our approach outperforms the best multi-scan method 2DPASS [39] by " + }, + { + "bbox": [ + 304, + 476, + 547, + 632 + ], + "type": "inline_equation", + "content": "3.3\\mathrm{mIoU}" + }, + { + "bbox": [ + 304, + 476, + 547, + 632 + ], + "type": "text", + "content": ". For nuScenes LiDAR semantic segmentation, we also achieve superior performance over previous algorithms. As shown in Table 3, the proposed approach surpasses the UniSeg [24] by " + }, + { + "bbox": [ + 304, + 476, + 547, + 632 + ], + "type": "inline_equation", + "content": "1.1\\mathrm{mIoU}" + }, + { + "bbox": [ + 304, + 476, + 547, + 632 + ], + "type": "text", + "content": ". In particular, our approach holds the highest entry on all three tracks. These encouraging results strongly show the effectiveness of our approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 643, + 405, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 643, + 405, + 656 + ], + "spans": [ + { + "bbox": [ + 306, + 643, + 405, + 656 + ], + "type": "text", + "content": "4.2. Ablation studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "We examine the effect of each component through detailed ablations. Unless otherwise specified, the following experiments are trained on the training set of the SemanticKITTI single-scan task and evaluated on the validation set." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15316" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 521, + 244 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 521, + 244 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 521, + 244 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 521, + 244 + ], + "type": "table", + "html": "
MethodmIoUconstruction
barrierbicyclebuscarconstructionmotorcyclepedestriantraffic-conetrailertruckdriveableother-flatsidewalkterrainmannadevegetation
Cylinder3D [46]77.282.829.884.389.463.079.377.273.484.669.197.770.280.375.590.487.6
SPVCNN [31]77.480.030.091.990.864.779.075.670.981.074.697.469.280.076.189.387.1
2D3DNet [11]80.083.059.488.085.163.784.482.076.084.871.996.967.479.876.092.189.2
2DPASS [39]80.881.755.392.091.873.386.578.572.584.775.597.669.179.975.590.288.0
LidarMultiNet [41]81.480.448.494.390.071.587.285.280.486.974.897.867.380.776.592.189.6
MSeg3D [20]81.183.142.594.992.067.178.685.780.587.577.397.769.881.277.892.490.1
SphereFormer [19]81.983.339.294.792.577.584.284.479.188.478.397.969.081.577.293.490.2
UniSeg [24]83.585.971.292.191.680.588.080.976.086.376.797.771.880.776.791.388.8
TASeg (Ours)84.687.169.490.592.278.790.486.381.988.375.997.870.981.078.293.491.2
", + "image_path": "b121b255f5e74594bd8e7a2b5b141e1dab4d8b771af4528441c9412b7e1f783f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 271, + 290, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 290, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 290, + 464 + ], + "type": "text", + "content": "Effect of Flexible Step Aggregation. In Table 4, we compare our FSA with other multi-frame algorithms. MSeg3D [20] utilizes temporal point clouds by directly concatenating all points, which introduces a large memory consumption and achieves limited improvement. SVQNet [4] leverages cross-attention to fuse temporal features while it can only handle fewer LiDAR scans. When increasing LiDAR scans, the memory overhead is also huge. Our FSA assigns different classes with different temporal steps, which can leverage long-term temporal information with less memory and time costs. Moreover, thanks to the utilization of historical priors, our FSA achieves better results (71.3 mIoU) than directly concatenating (69.9 mIoU), given the same temporal window size. To explore the effect of different window sizes on FSA, we present Table 5. Results show that the performance is saturated with a window size of 24." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "text", + "content": "Effect of Group Division. To investigate the effect of different group divisions for FSA, we provide Table 6. For division1, we simply divide all classes into three groups according to their performance on the val set. The group1 consists of the classes in [90, 100) mIoU, such as cars and roads. The group2 consists of classes in [80, 90) mIoU, such as motorcycles. The group3 contains the remaining classes. We assign group1, group2 and group3 with the step of " + }, + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "text", + "content": ", 4 and 2. A step of " + }, + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "text", + "content": " means we do not aggregate temporal points for classes in the group due to their near-saturate performances. The result shows that division1 can reduce the time and memory costs largely compared to directly stacking (the " + }, + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "inline_equation", + "content": "2^{st}" + }, + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "text", + "content": " row in Table 4). We can also use performance top1-6, top6-12, and top12-19 for division, resulting in division2. The group division is robust as long as it follows the principle that more difficult classes need smaller steps. To further reduce the memory consumption, we can finetune the group division, such as moving large-size classes (e.g. other-ground and terrain) in group3 to group2 (division3) or moving large-size classes in group3 and group2 to a new" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 316, + 272, + 537, + 364 + ], + "blocks": [ + { + "bbox": [ + 143, + 251, + 449, + 262 + ], + "lines": [ + { + "bbox": [ + 143, + 251, + 449, + 262 + ], + "spans": [ + { + "bbox": [ + 143, + 251, + 449, + 262 + ], + "type": "text", + "content": "Table 3. Performance comparison with start-of-the-art methods on nuScenes test set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 272, + 537, + 364 + ], + "lines": [ + { + "bbox": [ + 316, + 272, + 537, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 537, + 364 + ], + "type": "table", + "html": "
MethodLatencyMemorymIoU
Baseline63ms5.0G68.9
+ MSeg3D [20]†284ms69.7G69.9
+ SVQNet [4]†121ms15.4G69.5
+ FSA79ms8.5G71.3
+ FSA w/ gt mask79ms8.5G75.9
+ FSA + M.D.79ms8.5G71.8
+ FSA + M.D. × 0.7572ms6.9G71.6
", + "image_path": "d6d69026c35996167a3e859cd75290059fa9806224e6a484a5be59d6cc8f19ab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 316, + 407, + 537, + 436 + ], + "blocks": [ + { + "bbox": [ + 304, + 367, + 548, + 401 + ], + "lines": [ + { + "bbox": [ + 304, + 367, + 548, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 548, + 401 + ], + "type": "text", + "content": "Table 4. Comparison with different multi-frame algorithms. " + }, + { + "bbox": [ + 304, + 367, + 548, + 401 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 367, + 548, + 401 + ], + "type": "text", + "content": " represents our re-implementation. \"M.D.\" represents Mask Distillation. \"×75\" represents that we reduce the model parameters to " + }, + { + "bbox": [ + 304, + 367, + 548, + 401 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 367, + 548, + 401 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 407, + 537, + 436 + ], + "lines": [ + { + "bbox": [ + 316, + 407, + 537, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 407, + 537, + 436 + ], + "type": "table", + "html": "
Window size481216202428
mIoU69.970.470.971.371.471.671.4
", + "image_path": "4c5314fd218c1736de730cd7a15343608901615afec72fd25ad17cc4c43e2c73.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 321, + 457, + 531, + 529 + ], + "blocks": [ + { + "bbox": [ + 328, + 440, + 523, + 451 + ], + "lines": [ + { + "bbox": [ + 328, + 440, + 523, + 451 + ], + "spans": [ + { + "bbox": [ + 328, + 440, + 523, + 451 + ], + "type": "text", + "content": "Table 5. Ablation on different window sizes for FSA." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 457, + 531, + 529 + ], + "lines": [ + { + "bbox": [ + 321, + 457, + 531, + 529 + ], + "spans": [ + { + "bbox": [ + 321, + 457, + 531, + 529 + ], + "type": "table", + "html": "
MethodLatencyMemorymIoU
FSA w/ division192ms14.1G70.8
FSA w/ division283ms7.2G70.6
FSA w/ division379ms8.5G71.3
FSA w/ division474ms7.7G71.2
FSA w/ division574ms7.4G71.2
", + "image_path": "66685ffc0e1d35ac50579a0787643e624055db9a2c165568b097682cb22664de.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 324, + 533, + 526, + 544 + ], + "lines": [ + { + "bbox": [ + 324, + 533, + 526, + 544 + ], + "spans": [ + { + "bbox": [ + 324, + 533, + 526, + 544 + ], + "type": "text", + "content": "Table 6. Ablation on different group divisions for FSA." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 550, + 547, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 547, + 612 + ], + "type": "text", + "content": "group with step of 8 (division4). Besides, we can divide each class group into a close and a distant group according to the distance (such as " + }, + { + "bbox": [ + 304, + 550, + 547, + 612 + ], + "type": "inline_equation", + "content": "30\\mathrm{m}" + }, + { + "bbox": [ + 304, + 550, + 547, + 612 + ], + "type": "text", + "content": "). Considering close areas need fewer temporal points, we can use twice the original step for temporal aggregation of close groups (division5)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "text", + "content": "Effect of Mask Distillation. At the " + }, + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "inline_equation", + "content": "5^{th}" + }, + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "text", + "content": " row of Table 4, we use gt masks of historical frames for FSA. The performance is greatly boosted. Considering we cannot get ground truth at inference, we use it as a teacher to distill the model trained with pseudo masks. The result shows that the student is improved by " + }, + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{mIoU}" + }, + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "text", + "content": " after distillation. With " + }, + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "text", + "content": " model complexity, we can further reduce the time and memory cost and still achieve a leading performance." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15317" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "Effect of Temporal Image Aggregation and Fusion. TIAF delivers a temporal multi-modal fusion scheme to make full use of temporal images. As shown in Table 7, with zero image, only TLAD is used, achieving " + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "71.8\\mathrm{mIoU}" + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": ". With one image, only present images are utilized. Due to the limited FOV of the present camera, only part of the LiDAR points can gather image features, which limits the multi-modal fusion. Our TIAF leverages temporal images to enlarge the FOV of the camera and enhance present features. The results show that with the number of images increasing, the performance rises gradually, which verifies the effectiveness of our TIAF. When using seven temporal images, we achieve " + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "1.0\\mathrm{mIoU}" + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": " improvement on the strong baseline. Note that there is no extra cost to generate historical image features because we can save them at the past moments." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 258, + 278, + 287 + ], + "blocks": [ + { + "bbox": [ + 56, + 258, + 278, + 287 + ], + "lines": [ + { + "bbox": [ + 56, + 258, + 278, + 287 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 278, + 287 + ], + "type": "table", + "html": "
Num of images01357
mIoU71.872.172.472.772.8
", + "image_path": "18ffc2b10ec61de70f86f2abf947c3ae938456c6dc1d58b4cb3a0cff30e48ad7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 87, + 308, + 249, + 371 + ], + "blocks": [ + { + "bbox": [ + 53, + 290, + 281, + 302 + ], + "lines": [ + { + "bbox": [ + 53, + 290, + 281, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 290, + 281, + 302 + ], + "type": "text", + "content": "Table 7. Ablation on the number of temporal images for TIAF." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 87, + 308, + 249, + 371 + ], + "lines": [ + { + "bbox": [ + 87, + 308, + 249, + 371 + ], + "spans": [ + { + "bbox": [ + 87, + 308, + 249, + 371 + ], + "type": "table", + "html": "
Exp2D Sup.3D Sup.Multi-ScalemIoU
(a)✓✓72.4
(b)✓✓72.5
(c)✓✓72.4
(d)✓✓✓72.7
", + "image_path": "36c7277fdbb3cbb35ab666e290b46a18eda2dc8111b087971c49a19ab59fe6e5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 394, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 287, + 525 + ], + "type": "text", + "content": "We also provide the ablation on different supervisions and multi-scale features for TIAF in Table 8. In TIAF, we leverage 2D and 3D supervision to guide the extracted image features to be more conducive for segmentation. Multi-scale is also used to provide more discriminative features. The result shows that each of the designs is beneficial for final performance. Note that our temporal multi-modal method is orthogonal to other single-frame multi-modal methods. We use a simple pixel-to-point mapping at the feature projection stage. More complex methods [20, 24] can also be utilized, but it is not the focus of our TIAF." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 628 + ], + "type": "text", + "content": "Effect of Static-Moving Switch Augmentation. To verify our SMSA, we present Table 9. From the results, we can find that without SMSA, the accuracy of the model on many classes is lower than " + }, + { + "bbox": [ + 46, + 533, + 287, + 628 + ], + "type": "inline_equation", + "content": "10\\mathrm{IoU}" + }, + { + "bbox": [ + 46, + 533, + 287, + 628 + ], + "type": "text", + "content": ". After using SMSA, the performance of moving truck, moving other-vehicle and static bicyclist classes is improved by more than " + }, + { + "bbox": [ + 46, + 533, + 287, + 628 + ], + "type": "inline_equation", + "content": "20\\mathrm{IoU}" + }, + { + "bbox": [ + 46, + 533, + 287, + 628 + ], + "type": "text", + "content": ", and the overall performance boosts from " + }, + { + "bbox": [ + 46, + 533, + 287, + 628 + ], + "type": "inline_equation", + "content": "61.3\\mathrm{mIoU}" + }, + { + "bbox": [ + 46, + 533, + 287, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 533, + 287, + 628 + ], + "type": "inline_equation", + "content": "65.7\\mathrm{mIoU}" + }, + { + "bbox": [ + 46, + 533, + 287, + 628 + ], + "type": "text", + "content": ", which strongly demonstrates the efficacy of our SMSA." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 50, + 632, + 286, + 688 + ], + "blocks": [ + { + "bbox": [ + 47, + 378, + 286, + 389 + ], + "lines": [ + { + "bbox": [ + 47, + 378, + 286, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 378, + 286, + 389 + ], + "type": "text", + "content": "Table 8. Effect of different supervisions and multi-scale for TIAF." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 632, + 286, + 688 + ], + "lines": [ + { + "bbox": [ + 50, + 632, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 50, + 632, + 286, + 688 + ], + "type": "table", + "html": "
SMSAmloUcar-scar-mtrue-strue-mother-sother-mpers-spers-mbicy-sbicy-mmotor-smotor-m
✓61.395.679.349.29.161.05.739.087.00.072.64.376.4
65.794.872.853.635.271.453.236.780.432.475.212.389.6
", + "image_path": "d512a85a8f7b9fb58e8a8f6d7119d217faf8cd4b7185b3e26e198f1e4b0c5bb1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": "Generalization to More Architectures. In addition to MinkowskiNet, we also validate our temporal aggregation method on Cylinder3D [46] and SPVNAS [31]. Detailed results are summarized in Table 10. Our TASeg brings impressive improvements to both baselines, which demonstrates the strong generalization ability of our method." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 347, + 153, + 506, + 217 + ], + "blocks": [ + { + "bbox": [ + 47, + 691, + 286, + 712 + ], + "lines": [ + { + "bbox": [ + 47, + 691, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 691, + 286, + 712 + ], + "type": "text", + "content": "Table 9. Ablation study of our SMSA on the SemanticKITTI test benchmark of the multi-scan track." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 347, + 153, + 506, + 217 + ], + "lines": [ + { + "bbox": [ + 347, + 153, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 347, + 153, + 506, + 217 + ], + "type": "table", + "html": "
MethodmIoU
Cylinder3D [46]†66.0
Cylinder3D [46]†+ TASeg69.6
SPVNAS [31]†68.6
SPVNAS [31]†+ TASeg72.7
", + "image_path": "971fd6e83459116b596c16e785355270e930e16737abff4849dddebcc51f3b84.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 251, + 547, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 547, + 346 + ], + "type": "text", + "content": "Comparison on Complexity, Latency and Accuracy. As shown in Table 11, when not utilizing temporal images, our TASeg achieves superior performance than other methods with comparable or less complexity and latency. With temporal images, TASeg can maintain lower complexity and latency than UniSeg [24] while achieving higher accuracy. Besides, with " + }, + { + "bbox": [ + 304, + 251, + 547, + 346 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 251, + 547, + 346 + ], + "type": "text", + "content": " parameters, the latency of TASeg can be further reduced, and the accuracy is not affected much." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 310, + 353, + 542, + 448 + ], + "blocks": [ + { + "bbox": [ + 304, + 224, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 304, + 224, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 547, + 246 + ], + "type": "text", + "content": "Table 10. Performance of applying our TASeg to different baselines. " + }, + { + "bbox": [ + 304, + 224, + 547, + 246 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 224, + 547, + 246 + ], + "type": "text", + "content": " represents our re-implementation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 353, + 542, + 448 + ], + "lines": [ + { + "bbox": [ + 310, + 353, + 542, + 448 + ], + "spans": [ + { + "bbox": [ + 310, + 353, + 542, + 448 + ], + "type": "table", + "html": "
MethodInputParamsLatencymIoU
MinkowskiNet [7]†L37.9M63ms68.4
Cylinder3D [46]†L55.9M75ms66.0
SPVNAS [31]†L96.7M105ms68.6
RPVNet [37]†L119.0M110ms68.9
UniSeg [24]L+C147.6M145ms71.3
TASeg wo/ TIAFL+T37.9M79ms71.8
TASegL+C+T46.7M116ms72.7
TASeg × 0.75L+C+T27.4M108ms72.5
", + "image_path": "329f087cdd65c3e26136edd04ed16c660e6a7c7f04bf598b53687186be96711e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 454, + 547, + 498 + ], + "lines": [ + { + "bbox": [ + 304, + 454, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 547, + 498 + ], + "type": "text", + "content": "Table 11. Comparison of the model complexity, latency and accuracy on SemanticKITTI val set. " + }, + { + "bbox": [ + 304, + 454, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 454, + 547, + 498 + ], + "type": "text", + "content": " represents our re-implementation on NVIDIA A100 for a fair comparison. \"L\", \"C\" and \"T\" represent LiDAR, Camera and Temporal data, respectively." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 510, + 378, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 510, + 378, + 522 + ], + "spans": [ + { + "bbox": [ + 306, + 510, + 378, + 522 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 528, + 547, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 547, + 599 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 547, + 599 + ], + "type": "text", + "content": "In this paper, we present TASeg to fully exploit temporal point clouds and temporal images. Our TASeg is comprised of TLAD, TIAF and SMSA. We perform extensive experiments to validate the efficacy of each component. Moreover, our TASeg set a new state-of-the-art on three challenging tracks, demonstrating the superiority of our method." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 609, + 421, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 421, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 421, + 624 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "This work was supported in part by The National Nature Science Foundation of China (Grant Nos: 62273301, 62273302, 62036009, 61936006, 62273303), in part by Yongjiang Talent Introduction Programme (Grant No: 2022A-240-G, 2023A-197-G), in part by the National Key R&D Program of China (NO.2022ZD0160100). This work was done during the first author's internship at Shanghai AI Laboratory." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15318" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 145 + ], + "type": "text", + "content": "[1] Mehmet Aygun, Aljosa Osep, Mark Weber, Maxim Maximov, Cyrill Stachniss, Jens Behley, and Laura Leal-Taixe. 4d panoptic lidar segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5527-5537, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 190 + ], + "type": "text", + "content": "[2] Hanwen Cao, Yongyi Lu, Cewu Lu, Bo Pang, Gongshen Liu, and Alan Yuille. Asap-net: Attention and structure aware point cloud sequence segmentation. arXiv preprint arXiv:2008.05149, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "text", + "content": "[3] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. arXiv preprint arXiv:2306.16927, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 300 + ], + "type": "text", + "content": "[4] Xuechao Chen, Shuangjie Xu, Xiaoyi Zou, Tongyi Cao, Dit-Yan Yeung, and Lu Fang. Svqnet: Sparse voxel-adjacent query network for 4d spatio-temporal lidar semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8569–8578, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "type": "text", + "content": "[5] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. (af)2-S3Net: Attentive Feature Fusion with Adaptive Feature Selection for Sparse Semantic Segmentation Network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12547-12556, 2021. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 390 + ], + "type": "text", + "content": "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 435 + ], + "type": "text", + "content": "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 2, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 288, + 479 + ], + "type": "text", + "content": "[8] Ayush Dewan and Wolfram Burgard. Deeptemporalseg: Temporally consistent semantic segmentation of 3d lidar scans. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 2624-2630. IEEE, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 534 + ], + "type": "text", + "content": "[9] Fabian Duerr, Mario Pfaller, Hendrik Weigel, and Jürgen Beyerer. Lidar-based recurrent 3d semantic segmentation with temporal memory alignment. In 2020 International Conference on 3D Vision (3DV), pages 781-790. IEEE, 2020. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "type": "text", + "content": "[10] Khaled El Madawi, Hazem Rashed, Ahmad El Sallab, Omar Nasr, Hanan Kamel, and Senthil Yogamani. Rgb and lidar fusion based 3d semantic segmentation for autonomous driving. In IEEE Intelligent Transportation Systems Conference, pages 7-12, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 645 + ], + "type": "text", + "content": "[11] Kyle Genova, Xiaoqi Yin, Abhijit Kundu, Caroline Pantofaru, Forrester Cole, Avneesh Sud, Brian Brewington, Brian Shucker, and Thomas Funkhouser. Learning 3d semantic segmentation with only 2d image supervision. In 3DV, 2021. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "type": "text", + "content": "[12] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[13] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-Voxel Knowledge Distillation for LiDAR Semantic Segmentation. In IEEE Conference on Computer" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 546, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 546, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 546, + 94 + ], + "type": "text", + "content": "Vision and Pattern Recognition, pages 8479-8488, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 150 + ], + "type": "text", + "content": "[14] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 151, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 547, + 205 + ], + "type": "text", + "content": "[15] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 228-240, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 205, + 547, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 205, + 547, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 205, + 547, + 249 + ], + "type": "text", + "content": "[16] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. Lasermix for semi-supervised lidar semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21705-21715, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 250, + 547, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 547, + 304 + ], + "type": "text", + "content": "[17] Lars Kreuzberg, Idil Esen Zulfikar, Sabarinath Mahadevan, Francis Engelmann, and Bastian Leibe. 4d-stop: Panoptic segmentation of 4d lidar using spatio-temporal object proposal generation and aggregation. In European Conference on Computer Vision, pages 537-553. Springer, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 305, + 547, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 547, + 358 + ], + "type": "text", + "content": "[18] Georg Krispel, Michael Opitz, Georg Waltner, Horst Possegger, and Horst Bischof. Fuseseg: Lidar point cloud segmentation fusing multi-modal data. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1874-1883, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 360, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 547, + 392 + ], + "type": "text", + "content": "[19] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. arXiv preprint arXiv:2303.12766, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 393, + 547, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 547, + 426 + ], + "type": "text", + "content": "[20] Jiale Li, Hang Dai, Hao Han, and Yong Ding. MSeg3D: Multi-modal 3D Semantic Segmentation for Autonomous Driving. arXiv preprint arXiv:2303.08600, 2023. 1, 2, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 426, + 547, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 547, + 470 + ], + "type": "text", + "content": "[21] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research, 20(6):837-854, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 471, + 547, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 547, + 536 + ], + "type": "text", + "content": "[22] Jiahui Liu, Chirui Chang, Jianhui Liu, Xiaoyang Wu, Lan Ma, and Xiaojuan Qi. Mars3d: A plug-and-play motion-aware model for semantic segmentation on multi-scan 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9372-9381, 2023. 1, 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 537, + 547, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 547, + 580 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 547, + 580 + ], + "type": "text", + "content": "[23] Youquan Liu, Yeqi Bai, Lingdong Kong, Runnan Chen, Yuenan Hou, Botian Shi, and Yikang Li. Pcseg: An open source point cloud segmentation codebase. https://github.com/PJLab-ADG/PCSeg, 2023.6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 581, + 547, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 581, + 547, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 581, + 547, + 646 + ], + "type": "text", + "content": "[24] Youquan Liu, Runnan Chen, Xin Li, Lingdong Kong, Yuchen Yang, Zhaoyang Xia, Yeqi Bai, Xinge Zhu, Yuexin Ma, Yikang Li, et al. Uniseg: A unified multi-modal lidar segmentation network and the openpcseg codebase. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21662-21673, 2023. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 647, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 690 + ], + "type": "text", + "content": "[25] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2019. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 547, + 714 + ], + "type": "text", + "content": "[26] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15319" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "text", + "content": "[27] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in neural information processing systems, pages 5099-5108, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 185 + ], + "type": "text", + "content": "[28] Herbert Robbins and Sutton Monro. A Stochastic Approximation Method. The annals of mathematical statistics, pages 400-407, 1951. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "type": "text", + "content": "[29] Radu Alexandru Rosu, Peer Schütt, Jan Quenzel, and Sven Behnke. Latticenet: Fast point cloud segmentation using permutohedral lattices. arXiv preprint arXiv:1912.05905, 2019.6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 288, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 288, + 264 + ], + "type": "text", + "content": "[30] Hanyu Shi, Guosheng Lin, Hao Wang, Tzu-Yi Hung, and Zhenhua Wang. Spsequencenet: Semantic segmentation network on 4d point clouds. In CVPR, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 266, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 266, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 266, + 288, + 319 + ], + "type": "text", + "content": "[31] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In European Conference on Computer Vision, pages 685-702. Springer, 2020. 2, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 321, + 288, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 288, + 365 + ], + "type": "text", + "content": "[32] Jiexiong Tang, John Folkesson, and Patric Jensfelt. Sparse2dense: From direct sparse odometry to dense 3-d reconstruction. IEEE Robotics and Automation Letters, 4(2): 530-537, 2019. 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 366, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 288, + 431 + ], + "type": "text", + "content": "[33] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In 2019 International Conference on Robotics and Automation (ICRA), pages 4376-4382. IEEE, 2019. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 288, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 477 + ], + "type": "text", + "content": "[34] Dong Wu, Man-Wen Liao, Wei-Tian Zhang, Xing-Gang Wang, Xiang Bai, Wen-Qing Cheng, and Wen-Yu Liu. Yolop: You only look once for panoptic driving perception. Machine Intelligence Research, 19(6):550-562, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "type": "text", + "content": "[35] Aoran Xiao, Jiaxing Huang, Dayan Guan, Kaiwen Cui, Shijian Lu, and Ling Shao. PolarMix: A General Data Augmentation Technique for LiDAR Point Clouds. arXiv preprint arXiv:2208.00223, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "type": "text", + "content": "[36] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeezesegv3: Spatially-adaptive convolution for efficient point-cloud segmentation. In European Conference on Computer Vision, pages 1-19. Springer, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "type": "text", + "content": "[37] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. RPVnet: A Deep and Efficient Range-Point-Voxel Fusion Network for Lidar Point Cloud Segmentation. In IEEE International Conference on Computer Vision, pages 16024–16033, 2021. 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 690 + ], + "type": "text", + "content": "[38] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3101-3109, 2021. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[39] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2DPASS: 2D Priors" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 487 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "type": "text", + "content": "Assisted Semantic Segmentation on LiDAR Point Clouds. In European Conference on Computer Vision, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 139 + ], + "type": "text", + "content": "[40] Jihan Yang, Shaoshuai Shi, Runyu Ding, Zhe Wang, and Xiaojuan Qi. Towards efficient 3d object detection with knowledge distillation. Advances in Neural Information Processing Systems, 35:21300-21313, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 185 + ], + "type": "text", + "content": "[41] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet: Towards a unified multi-task network for lidar perception. arXiv preprint arXiv:2209.09385, 2022. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 547, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 229 + ], + "type": "text", + "content": "[42] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE access, 8:58443-58469, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 231, + 546, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 546, + 285 + ], + "type": "text", + "content": "[43] Zhiwei Zhang, Zhizhong Zhang, Qian Yu, Ran Yi, Yuan Xie, and Lizhuang Ma. Lidar-camera panoptic segmentation via geometry-consistent and semantic-aware alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3662-3671, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 286, + 546, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 546, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 546, + 341 + ], + "type": "text", + "content": "[44] Wu Zheng, Li Jiang, Fanbin Lu, Yangyang Ye, and Chi-Wing Fu. Boosting single-frame 3d object detection by simulating multi-frame point clouds. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4848-4856, 2022. 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 342, + 547, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 547, + 397 + ], + "type": "text", + "content": "[45] Minghan Zhu, Shizhong Han, Hong Cai, Shubhankar Borse, Maani Ghaffari, and Fatih Porikli. 4d panoptic segmentation as invariant and equivariant field prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22488-22498, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 398, + 546, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 546, + 453 + ], + "type": "text", + "content": "[46] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and Asymmetrical 3D Convolution Networks for Lidar Segmentation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 9939-9948, 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 454, + 546, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 454, + 546, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 454, + 546, + 487 + ], + "type": "text", + "content": "[47] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. In ICCV, 2021. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15320" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/137896e1-f0a6-4b80-849c-9a14cbb3250c_content_list.json b/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/137896e1-f0a6-4b80-849c-9a14cbb3250c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f3cebf4b957c14d9c1b3370ece50ec2b27eb07e3 --- /dev/null +++ b/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/137896e1-f0a6-4b80-849c-9a14cbb3250c_content_list.json @@ -0,0 +1,1570 @@ +[ + { + "type": "text", + "text": "TCP:Textual-based Class-aware Prompt tuning for Visual-Language Model", + "text_level": 1, + "bbox": [ + 104, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hantao Yao $^{1}$ , Rui Zhang $^{2}$ , Changsheng Xu $^{1,3}$", + "bbox": [ + 346, + 181, + 620, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, China", + "bbox": [ + 91, + 198, + 877, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences, China;", + "bbox": [ + 169, + 215, + 797, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Chinese Academy of Sciences, China", + "bbox": [ + 326, + 233, + 643, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hantao.yao@nlpr.ia.ac.cn; zhangrui@ict.ac.cn; csxu@nlpr.ia.ac.cn", + "bbox": [ + 196, + 253, + 764, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prompt tuning represents a valuable technique for adapting pre-trained visual-language models (VLM) to various downstream tasks. Recent advancements in CoOp-based methods propose a set of learnable domain-shared or image-conditional textual tokens to facilitate the generation of task-specific textual classifiers. However, those textual tokens have a limited generalization ability regarding unseen domains, as they cannot dynamically adjust to the distribution of testing classes. To tackle this issue, we present a novel Textual-based Class-aware Prompt tuning(TCP) that explicitly incorporates prior knowledge about classes to enhance their discriminability. The critical concept of TCP involves leveraging Textual Knowledge Embedding (TKE) to map the high generalizability of class-level textual knowledge into class-aware textual tokens. By seamlessly integrating these class-aware prompts into the Text Encoder, a dynamic class-aware classifier is generated to enhance discriminability for unseen domains. During inference, TKE dynamically generates class-aware prompts related to the unseen classes. Comprehensive evaluations demonstrate that TKE serves as a plug-and-play module effortlessly combinable with existing methods. Furthermore, TCP consistently achieves superior performance while demanding less training time1.", + "bbox": [ + 76, + 335, + 473, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 729, + 209, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale image-text pairs have the ability to train a Visual-language model (VLM) with a powerful generalization capacity for various downstream tasks [2, 33]. However, training these models from scratch requires a huge dataset with labeled images, making it difficult to apply them directly to downstream tasks with fewer images. To address this issue, three common techniques have been pro", + "bbox": [ + 75, + 756, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c2f9ba5a4198bc1eb3d356e143542573c8f5cb36b0644ff32ab0104935c3021b.jpg", + "image_caption": [ + "Figure 1. Comparison with existing frameworks. (a) Domain-shared prompt tuning applies the same learnable prompt between the training and testing domains. (b) Image-conditional prompt tuning combines the image embedding with the learnable prompt; (c) Class-aware Prompt tuning injects the class-level textual embedding into the Text Encoder with the class-aware prompt." + ], + "image_footnote": [], + "bbox": [ + 522, + 301, + 700, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/89bb576f820f247f6119691c614724f1b01f1f263a433bdd0e86aa74c33e7378.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 315, + 870, + 469 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "posed: fine-tuning [30], prompt tuning [50], adapter [12] and LoRA [15]. Among them, prompt tuning is a simple and efficient framework that transfers the essential general knowledge of VLM to the downstream tasks.", + "bbox": [ + 496, + 575, + 890, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prompt tuning $^2$ is a technique that combines learnable textual tokens with class tokens to generate a discriminative textual classifier, known as Context Optimization (CoOp) [50]. Recently, various CoOp-based methods [4, 5, 17, 18, 22, 31, 40, 43, 50, 52] infer the domain-shared prompt tokens between training and testing domains (Figure 1(a)). Nevertheless, as the domain-shared prompt tokens are derived from labeled training images, their performance is suboptimal when confronted with unseen test classes. To enhance the generalization capacity of learnable prompt tokens, image-conditional prompts have been proposed in [45, 49] by fusing image features and learnable tokens (Figure 1(b)). Notably, image-conditional", + "bbox": [ + 496, + 638, + 892, + 835 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "2For the visual-language model, two types of prompt-tuning strategies exist: visual prompt tuning and textual prompt tuning. In this work, we focus on the textual prompt tuning and do not consider the visual prompt tuning.", + "bbox": [ + 500, + 849, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/htyao89/Textual-based_ Class-aware_prompt_tuning", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23438", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "textual tokens encapsulate specific knowledge of each image, particularly for testing images, thereby rendering it more apt for generalization to unseen testing images. However, image-conditional prompt tokens with image-specific knowledge have less ability in elevating the distribution of class embeddings. In summary, classifiers generated by domain-shared and image-conditional textual tokens exhibit suboptimal performance for the unseen classes, primarily due to their inability to explicitly model the class distribution. Consequently, it is imperative to establish a dynamic relationship between the learnable prompt and the textual knowledge of each class to augment its discriminative prowess.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The frozen CLIP, coupled with a hand-crafted prompt, exhibits robust generalization capabilities to novel classes, rendering it a valuable source of prior textual knowledge for each class. By associating the class-level textual knowledge with the learnable prompts, a class-aware prompt can be formulated for hencing the discriminative capacity of the textual classifier. To achieve this, we employ an embedding module to project the class-aware textual knowledge into the class-aware prompt tokens, as shown in Figure 1(c). The resultant class-aware prompt incorporates prior textual knowledge specific to each class, endowing the generated textual classifier with heightened discriminative prowess. Furthermore, the class-aware prompt facilitates the generation of classifiers for both seen and unseen classes by leveraging textual knowledge from both categories. To sum up, the trained embedding module can generate a class-aware prompt for each class based on its description ('class name'), thereby enhancing generalization and discriminative capabilities in class-level textual embeddings.", + "bbox": [ + 75, + 289, + 472, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, we propose a novel Textual-based Class-aware Prompt tuning (TCP) based on the framework of CoOp, shown in Figure 2. In addition to the domain-share textual tokens introduced in CoOp, TCP contributes a novel Textual Knowledge Embedding (TKE) to map the class-level textual knowledge into class-aware prompt tokens. Moreover, a class-aware textual classifier is generated by inserting the class-aware prompt tokens into the middle layer of the Text Encoder. We use standard contrastive loss and knowledge-guided consistency [43] to optimize the TKE and the learnable prompt tokens. During inference, TCP generates a class-aware classifier for unseen classes by feeding the domain-shared prompt tokens and the class-aware prompt tokens generated by the TKE into the frozen Text Encoder.", + "bbox": [ + 75, + 580, + 472, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, the proposed TCP explicitly steers prompts to learn a class-aware knowledge that maximizes the generalization and discriminative of the downstream tasks. Evaluation over 11 image classification datasets on base-to-new generalization, cross-dataset generalization, and few-shot learning verify shows that TCP is an efficient method that", + "bbox": [ + 75, + 809, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "obtain a higher performance with less training time. In summary, the proposed Textual-based Class-aware Prompt tuning(TCP) has the following main contributions:", + "bbox": [ + 496, + 90, + 893, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. An effectively Textual-based Class-aware Prompt tuning is proposed by injecting the textual class-aware prompts generated by Textual Knowledge Embedding(TKE) into the Text Encoder.", + "2. We demonstrate that explicitly incorporating the prior knowledge of each class into the learnable prompt tokens can enhance the discriminative of the class distribution.", + "3. Textual Knowledge Embedding(TKE) is a plug-and-play module that can quickly insert existing prompt tuning methods to improve their performance further." + ], + "bbox": [ + 500, + 137, + 893, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 301, + 650, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Vision-Language Models", + "text_level": 1, + "bbox": [ + 500, + 328, + 728, + 344 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, researchers have shown that the Visual-Language Models(VLM) [2, 33] consisting of visual and textual modalities trained on the large-scale of image-text pairs has a powerful generalization and discriminativeion. To further boost the descriptive ability of VLM, the VLM models are boosted from the following aspect: 1) using a stronger text encoder or visual encoder [25, 41, 46]; 2) deeply fusing the visual and text knowledges [23, 38]; 3) using more images [16, 33, 35, 36]. To boost the diversity of text description, Masked Language Modeling (MLM) [20] [26] randomly erases the words in the text description used for representation learning. Unlike MLM, Masked autoencoder-based methods [13] are proposed to boost the descriptive ability by randomly masking the image patches. Among existing VLM models, CLIP is a representative and straightforward framework for inferring the independent visual and text encoder using the contrastive loss based on 400 million image-text association pairs. As CLIP has a good generalization, most existing CoOp-based methods are proposed based on CLIP for adapting the pre-trained VLM into the downstream task. Similar to existing methods, we conduct the prompt-tuning strategy on the TextEncoder of CLIP to obtain a task-specific textual embedding for prediction.", + "bbox": [ + 496, + 352, + 893, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Prompt Tuning", + "text_level": 1, + "bbox": [ + 500, + 710, + 656, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To adapt the pretrained VLM to the downstream tasks, the prompt tuning [10, 22, 24, 31, 33, 45] always applies task-related textual tokens to infer the task-specific textual knowledge. In CLIP[33], the hand-crafted template \"a photo of a [CLASS]\" is used to embed the textual embedding for zero-shot prediction. However, the hand-crafted prompts have a poor ability to describe the downstream task. Textual prompt tuning is applied to boost the textual embedding by inferring a set of learnable textual tokens combined with the class tokens. For example, Context Optimization(CoOp) [50] replaces the hand-crafted prompts", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23439", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with the learnable soft prompts. To improve the generalization of the learnable textual prompt in CoOp, Conditional Context Optimization(CoCoOp) [49] and VPT [45] generates an image-conditional prompt fusing the image feature and the learnable textual prompt. Moreover, Knowledge-Guided Context Optimization(KgCoOp) [43], ProGrad [51], and Prompt Regularization(ProReg) [52] constrain the proposed learnable prompts contain the essential general knowledge. Unlike the above methods, which consider textual prompts, Ensembling Context Optimization(ECO) [1] employs prompt ensembling to combine multiple prompts. To obtain high-quality task-related tokens, ProDA [27] considers the prompt's prior distribution learning, and Distribution-Aware Prompt Tuning (DAPT) [5] optimizes the learnable prompt by maximizing inter-dispersion. Besides the textual knowledge from \"classname\", Knowledge-Aware Prompt Tuning(KAPT) [17] employs the external knowledge to generate the discriminative knowledge-aware prompt for the novel categories. PLOT [4] applies optimal transport to match the vision and text modalities for generating the discriminative and visual-aligned local textual prompt tokens. Besides the textual prompt tuning, Multi-modal Prompt Learning (MaPLe) [18] and PromptSRC [19] conduct the visual-textual prompt tuning by jointly conducting the prompt tuning on the visual and text encoders. Multitask Vision-Language Prompt Tuning(MVLPT) [37]incorporates crosstalk knowledge into prompt tuning for vision-language models. DenseCLIP [34] uses the context-aware prompt strategy to generate dense prediction tasks, and CLIP-Adapter [11] applies an adapter to adjust the visual or text embeddings.", + "bbox": [ + 76, + 90, + 472, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing methods commonly infer two types of prompt tokens: domain-share and image-conditional. However, the textual classifiers generated with these tokens tend to perform poorly on unseen classes. To mitigate this limitation, we propose a novel Textual-based Class-aware Prompt tuning (TCP) that employs a dynamic class-aware token to enhance the generalization and discriminative capabilities of the learnable textual prompt. Moreover, we introduce a Textual Knowledge Embedding to project class-level textual knowledge into the class-aware prompts, which are then inserted into the Text Encoder to generate a discriminative class-aware classifier. The evaluation results demonstrate that integrating class-level prior knowledge into the prompt tokens significantly enhances the discriminative ability of the prompt tuning process.", + "bbox": [ + 75, + 577, + 472, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 825, + 204, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As the Textual-based Class-aware Prompt tuning(TCP) is proposed based on Context Optimization (CoOp), we first briefly review CoOp and then introduce the proposed TCP.", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 90, + 648, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing CoOp-based methods are proposed based on the powerful Contrastive Language-Image Pre-training(CLIP). Given an image along with its corresponding textual description, CLIP uses visual and text encoders to extract the visual and text embeddings. After that, the constrastive loss between the visual and textual embeddings is calculated to align those two embeddings. To effectively adapt CLIP for the downstream task, CLIP applies the hand-crafted template \"a photo of a $\\{\\}$ \" to extract the general class-level textual embedding, defined as $\\mathbf{W}^{clip} = \\{\\mathbf{w}_i^{clip}\\}_{i=1}^{N_c}$ , where $\\mathbf{w}_i^{clip}$ is the textual embedding of $i$ -th class, and $N_c$ is the number of classes. Given the 'class-name' of $i$ -th class, Word Embedded $e(\\cdot)$ firstly embeds the hand-crafted description into a vectorized textual tokens: $\\mathbf{t}_i^{clip} = e(\"a photo of a \\{class-name\\}\")$ . After that, Text Encoder $\\theta$ maps the vectorized textual tokens $\\mathbf{t}_i^{clip}$ into the class-level embedding: $\\mathbf{w}_i^{clip} = \\theta (\\mathbf{t}_i^{clip})$ .", + "bbox": [ + 496, + 114, + 893, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To improve the discriminative of the class-level embedding, the prompt tuning methods of Context Optimization(CoOp) replace the hand-crafted textual tokens with a set of learnable textual tokens $\\mathbb{T} = \\{\\mathbf{t}_1,\\mathbf{t}_2,\\dots,\\mathbf{t}_M\\}$ , where $M$ is the length of tokens. Similar to CLIP, the corresponding class token $\\mathbf{c}_i$ is concatenated with the learnable tokens $\\mathbb{T}$ for generating the textual token $\\mathbf{t}_i^{coop} = \\{\\mathbf{t}_1,\\mathbf{t}_2,\\dots,\\mathbf{t}_M,\\mathbf{c}_i\\}$ . Next, the textual embedding $\\mathbf{w}_i^{coop}$ is obtained by fedding the textual tokens $\\mathbf{t}_i^{coop}$ into Text Encoder $\\theta$ , i.e., $\\mathbf{w}_i^{coop} = \\theta (\\mathbf{t}_i^{coop})$ . Finally, the textual embeddings of all classes are defined as $\\mathbf{W}^{coop} = \\{\\mathbf{w}_i^{coop}\\}_{i = 1}^{N_c}$ . CoOp infers the learnable textual tokens $\\mathbb{T}$ by minimizing the contrastive loss between the image's embedding $\\mathbf{x}$ and its class embedding $\\mathbf{w}_y^{coop}$ :", + "bbox": [ + 496, + 381, + 893, + 598 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c e} = \\frac {1}{N} \\sum_ {(\\mathbf {x}, y) \\in \\mathcal {D} _ {s}} \\frac {\\exp (d (\\mathbf {x} , \\mathbf {w} _ {y} ^ {c o o p}) / \\tau)}{\\sum_ {i = 1} ^ {N _ {c}} \\exp (d (\\mathbf {x} , \\mathbf {w} _ {i} ^ {c o o p}) / \\tau)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 609, + 890, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_s$ is the seen dataset, and $d(\\cdot)$ is the cosine distance. $\\tau$ is a temperature factor defined in CLIP, and $N$ is the number of training images.", + "bbox": [ + 496, + 666, + 890, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As the generated textual embedding has a good generalization ability for the novel classes, $\\mathrm{KgCoOp}$ further adds an efficient consistency $\\mathcal{L}_{kg}$ between the generated embedding $\\mathbf{W}^{coop}$ and the general embedding $\\mathbf{W}^{clip}$ ,", + "bbox": [ + 496, + 713, + 890, + 776 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {k g} = | | \\mathbf {W} ^ {c l i p} - \\mathbf {W} ^ {c o o p} | | _ {2} ^ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 787, + 890, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, a robust objective for prompt tuning is:", + "bbox": [ + 519, + 823, + 849, + 838 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c e} + \\omega \\mathcal {L} _ {k g}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 854, + 890, + 871 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\omega$ is set as 8.0 same as in KgCoOp [43].", + "bbox": [ + 500, + 885, + 808, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23440", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1b2afd8b794f699c29a1ea474d25a27953c3d54ffced838f82246af0eaaad4f2.jpg", + "image_caption": [ + "Figure 2. The framework of the proposed TCP." + ], + "image_footnote": [], + "bbox": [ + 117, + 87, + 433, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Textual-based Class-aware Prompt tuning", + "text_level": 1, + "bbox": [ + 76, + 301, + 434, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the pre-trained Text Encoder in CLIP, the textual prompt tuning aims to infer a set of the domain-shared or image-conditional textual tokens combined with the general class tokens for generating the specific class embeddings. However, the textual classifier generated by those textual tokens performs a worse generalization on the unseen classes because they can not model the distribution of the testing classes. Research has shown that utilizing the general textual knowledge extracted by frozen CLIP can create a discriminative prior knowledge of novel classes, enhancing the discriminative and generalization of the learnable prompt. Drawing on the benefit of the general textual knowledge of seen and unseen classes, we propose a Textual-based Class-aware Prompt tuning(TCP) to adapt the pretrained CLIP to the downstream task. As shown in Figure 2, TCP employs a Textual Knowledge Embedding(TKE) to transfer the general class-level textual embedding into the class-aware prompt, which is then combined with the learnable textual tokens for generating the class-aware classifiers. TKE is advantageous for unseen classes, as it generates class-specific prompts to obtain an unseen class-aware textual classifier with better discriminative abilities. Moreover, explicitly incorporating seen class-aware prompts can enhance the discriminative power of seen classes.", + "bbox": [ + 75, + 325, + 472, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the general class-level textual embedding $\\mathbf{W}^{clip} \\in \\mathbb{R}^{N_c \\times D_t}$ with $N_c$ training classes, Textual Knowledge Embedding(TKE) $\\mathcal{T}(\\cdot)$ is proposed to project the class-level embedding $\\mathbf{W}^{clip}$ into the class-aware prompt $\\mathbf{T} = \\mathcal{T}(\\mathbf{W}^{clip})$ . As shown in Figure 2, TKE consists of two layers: down-project layer and up-project layer. The down-project layer uses the weight $\\mathbf{W}_{down} \\in \\mathbb{R}^{D_t \\times D_{mid}}$ to project the textual embedding into the low-dimension feature with the dimension of $D_{mid}$ . Next, the weight $\\mathbf{W}_{up} \\in \\mathbb{R}^{D_{mid} \\times D'}$ of the up-project layer maps the low-dimension feature into the high-dimension feature with the dimension of $D'$ . Note that $D'$ is determined by the prompt's length $M$ and dimension $D$ : $D' = M \\times D$ . In", + "bbox": [ + 75, + 704, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "summary, the general textual embedding $\\mathbf{W}^{clip} \\in \\mathbb{R}^{N_c \\times D_t}$ can be projected into the class-aware textual tokens $\\mathbf{T} \\in \\mathbb{R}^{N_c \\times D'}$ , which is further reshaped into the shape of $\\mathbf{T} \\in \\mathbb{R}^{N_c \\times M \\times D}$ for inserting into the middle layer of Text Encoder $\\theta$ .", + "bbox": [ + 496, + 89, + 890, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assuming we insert the class-aware prompt $\\mathbf{T}$ into $l$ -th layer of Text Encoder $\\theta$ . We will give a detailed analysis of the hyperparameter $l$ in the following. Similar to CoOp, by combining the domain-shared learnable textual tokens $\\mathbb{T} = [\\mathbf{t}_1,\\mathbf{t}_2,\\dots,\\mathbf{t}_M]$ and the pre-trained class tokens $\\mathbf{C}$ of all classes, we can obtain the input textual tokens $\\mathbf{F}_0 = \\{\\mathbb{T},\\mathbf{C}\\}$ of the Text Encoder, where $\\mathbf{C} = \\{\\mathbf{c}_i\\}_{i}^{N_c}$ , and $\\mathbf{c}_i$ is the vectorized textual tokens of $i$ -th class. The textual token $\\mathbf{F}_0$ is fed into the first $l$ layers of Text Encoder for obtaining the middle-level textual embedding $\\mathbf{F}_l$ . Formally, the textual token $\\mathbf{F}_i(i\\leq l)$ of $i$ -th layer is defined as:", + "bbox": [ + 496, + 167, + 892, + 333 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {i} = \\theta_ {i} \\left(\\mathbf {F} _ {i - 1}\\right), i \\in [ 1, l ], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 342, + 890, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_{i}$ is the $i$ -th layer of Text Encoder.", + "bbox": [ + 496, + 368, + 777, + 382 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the textual tokens $\\mathbf{F}_l\\in \\mathbb{R}^{N_c\\times N_t\\times D}$ and the class-aware prompt tokens $\\mathbf{T}\\in \\mathbb{R}^{N_c\\times M\\times D}$ , the first dimension is related to the number of classes. Therefore, the learnable prompt tokens are always inserted into the second dimension of $\\mathbf{F}_l$ , the same as CoOp. Formally, the class-aware prompt $\\mathbf{T}$ is inserted into $\\mathbf{F}_l$ to generate the class-aware enhanced tokens $\\mathbf{F}_l'$ .", + "bbox": [ + 496, + 382, + 890, + 489 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {l} ^ {\\prime} = \\left[ \\mathbf {T} _ {1}, \\mathbf {T} _ {2}, \\dots , \\mathbf {T} _ {M}, \\mathbf {F} _ {l, M + 1}, \\mathbf {F} _ {l, M + 2}, \\dots , \\mathbf {F} _ {l, N _ {t}} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 497, + 890, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{T}_i$ denotes the $i$ -th index of $\\mathbf{T}$ in the second dimension, and $\\mathbf{F}_{l,j}$ denote the correspond $j$ -th index of $\\mathbf{F}_l$ in the second dimension, i.e., $\\mathbf{T}_i = \\mathbf{T}[:i,:]$ , and $\\mathbf{F}_{l,j} = \\mathbf{F}_l[:j,:]$ .", + "bbox": [ + 496, + 523, + 890, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After that, the class-enhanced textual tokens $\\mathbf{F}_l^{\\prime}$ are fedded into the rest layers for generating the class-aware textual embedding,", + "bbox": [ + 496, + 568, + 890, + 614 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {i} ^ {\\prime} = \\theta_ {i} (\\mathbf {F} _ {i - 1} ^ {\\prime}), i \\in [ l + 1, L ]. \\qquad (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 623, + 890, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The output of the last layer $\\mathbf{F}_L^{\\prime}$ is treated as the class embedding $\\mathbf{W}^{tcp}$ used for optimization with contrastive loss and knowledge-guided consistency loss in Eq. (3).", + "bbox": [ + 496, + 648, + 890, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 707, + 633, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to CoOp [50], we evaluate the effectiveness of TCP from the three types of tasks: 1) generalization from baseto-new classes within a dataset; 2) few-shot learning with $K$ -shot labeled images; 3) cross-dataset generalization from the imagenet to other datasets. More detailed results will be provided in the Supplementary materials.", + "bbox": [ + 496, + 732, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 498, + 832, + 689, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset. We conduct the base-to-new generalization, few-shot learning, and cross-dataset generalization on 11 benchmarks, i.e., ImageNet [7], Caltech [9], OxfordPets [32],", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23441", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/0ea82750e6c47a62101faf4c779c2448f9196551956cd50b54609164fa5bee1b.jpg", + "table_caption": [ + "Table 1. Comparison on the base-to-new generalization setting with 16-shot samples. 'tp', 'dtp', 'vp', and 'dvp' denote the 'textual prompt', 'deep textual prompt', 'visual prompt', and 'deep visual prompt', respectively. PromptSRC are based on deep visual-textual prompt tuning ('dvp+dtp'). *' denote the performance obtained by our re-implementation." + ], + "table_footnote": [], + "table_body": "
DatasetsSetsCoOp* (JCV22)CoCoOp (CVPR22)DAPT* (ICCV23)ProGrad* (ICCV23)ProDA (CVPR22)KgCoOp (CVPR23)RPO (ICCV23)PLOT* (ICLR23)LFA (ICCV23)MaPLe (CVPR23)DePT (Arxiv23)PromptSRC* (ICCV23)TCP
tptptp+vptptptpdtp+dvptp+vp-dtp+dvptpdtp+dvptp
AverageBase82.3880.4783.1882.4881.5680.7381.1383.9883.6282.2883.6284.1284.13
New67.9671.6969.2770.7572.3073.675.0071.7274.5675.1475.0475.0275.36
H74.4875.8375.5976.1676.6577.077.7877.3778.8378.5579.1079.3179.51
ImageNetBase76.4675.9876.8377.0275.4075.8376.6077.3076.8976.6677.0377.7577.27
New66.3170.4369.2766.6670.2369.9671.5769.8769.3670.5470.1370.7069.87
H71.0273.1072.8571.4672.7272.7874.0073.4072.9373.4773.4274.0673.38
Caltech101Base97.8097.9697.8398.0298.2797.7297.9798.5398.4197.7498.3098.1398.23
New93.2793.8193.0793.8993.2394.3994.3792.8093.9394.3694.6093.9094.67
H95.4895.8495.3995.9195.6896.0396.0395.5896.1396.0296.4195.9796.42
OxfordPetsBase94.4795.2095.0095.0795.4394.6594.6394.5095.1395.4394.3395.5094.67
New96.0097.6995.8397.6397.8397.7697.5096.8396.2397.7697.2397.4097.20
H95.2396.4395.4196.3396.6296.1896.0595.6595.6896.5895.7696.4495.92
CarsBase75.6770.4975.8077.6874.7071.7673.8779.0776.3272.9479.1378.4080.80
New67.5373.5963.9368.6371.2075.0475.5374.8074.8874.0075.4774.7374.13
H71.3772.0169.3672.8872.9173.3674.6976.8875.5973.4777.2675.5277.32
FlowersBase97.2794.8796.9795.5497.7095.0094.1397.9397.3495.9298.0097.9097.73
New67.1371.7560.9071.8768.6874.7376.6773.5375.4472.4676.3776.7775.57
H79.4481.7174.8182.0380.6683.6584.5083.9985.0082.5685.8486.0685.23
Food101Base89.3790.7090.3790.3790.3090.590.3389.8090.5290.7190.5090.6390.57
New88.7791.2991.3089.5988.5791.790.8391.3791.4892.0591.6091.5091.37
H89.0790.9990.8389.9889.4391.0990.5890.5891.0091.3891.0591.0690.97
AircraftBase39.6733.4139.9740.5436.9036.2137.3342.1341.4837.4443.2042.3041.97
New31.2323.7129.8027.5734.1333.5534.2033.7332.2935.6134.8336.9734.43
H34.9527.7434.1432.8235.4634.8335.7037.4636.3136.5038.5739.4637.83
SUN397Base80.8579.7480.9781.2678.6780.2980.6082.2082.1380.8282.3382.8382.63
New68.3476.8676.9774.1776.9376.5377.8073.6377.2078.7077.8079.0078.20
H74.0778.2778.9277.5577.7978.3679.1877.6879.5979.7580.0080.8780.35
DTDBase79.9777.0182.2377.3580.6777.5576.7081.9781.2980.3682.2082.6082.77
New48.6056.0054.2352.3556.4854.9962.1343.8060.6359.1859.1357.5058.07
H60.4664.8565.3662.4566.4464.3568.6157.0969.4668.1668.7867.8068.25
EuroSATBase90.1087.4994.7390.1183.9085.6486.6393.7093.4094.0789.0392.4091.63
New53.0060.0450.3360.8966.0064.3468.9762.6771.2473.2371.0768.4374.73
H66.7471.2165.7472.6773.8873.4876.7975.1180.8382.379.0478.6382.32
UCF101Base84.5382.3384.3084.3385.2382.8983.6786.6086.9783.0085.8086.9387.13
New67.3773.4576.3374.9471.9776.6775.4375.9077.4878.6677.2378.3380.77
H74.9877.6780.1279.3578.0479.6579.3480.9081.9580.7781.2982.4183.83
", + "bbox": [ + 76, + 142, + 944, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "StanfordCars [21], Flowers [29], Food101 [3], FGVCAircraft [28], EuroSAT [14], UCF101 [39], DTD [6], and SUN397 [42]. A detailed introduction of those dataset is given in the Supplementary materials.", + "bbox": [ + 75, + 560, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details. Our implementation is based on CoOp's code [50]. All experiments are conducted based on the CLIP with the backbone of ViT-B/16 [8]. The prompt's length $M$ is set as 4 with random initialization of \"X X X X {}\". The final performance is averaged over three random seeds(1/2/3)for a fair comparison. Adam optimizer is applied for optimization with the learning rate of 2e-3 and the batch size of 32, and the training epochs is 50. All experiments are conducted on RTX 3090.", + "bbox": [ + 75, + 621, + 468, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. Recently CoOp-based methods are used for comparison, e.g., CoOp [50], CoCoOp [49], ProGrad [51], ProDA [27], KgCoOp [43], PromptSRC [19], MaPLe [18], LFA [31], DePT [47], DAPT [5], PLOT [4], TaskRes [44], RPO [22], VPT [45], and TIP-Adapter-F [48].", + "bbox": [ + 75, + 758, + 468, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Base-to-New Class Generalization", + "text_level": 1, + "bbox": [ + 76, + 845, + 375, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To verify the generalization ability of the prompt tuning, the base-to-new class generalization splits each dataset into two", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "disjoint subsets: Base and New classes, where Base classes are used to infer the learnable prompt and TKE, and New classes are used for evaluation. Note that each New class contains 16-shot samples. We summarize the comparison between the proposed TCP and existing methods in Table 1 from the following two aspects: Textual Prompt Tuning and Vivasal-Textual Prompt Tuning.", + "bbox": [ + 496, + 560, + 890, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Textual Prompt Tuning: Based on the visual feature extracted from the frozen visual encoder, the textual prompt tuning (e.g., CoOp, CoCoOp, ProGrad, ProDA, and Kg-CoOp) merely conducts the textual prompt tuning by inserting a set of learnable textual prompts into the Text Encoder for obtaining a discriminative textual classifier, which is same as our proposed TCP. Among all textual prompt tuning methods, KgCoOp [43] is a more substantial baseline, which is also the baseline of our TCP. Compared to KgCoOp, TCP improves the metric term of Base/New/H from $80.73\\% / 73.6\\% / 77.0\\%$ to $84.35\\% / 75.28\\% / 79.56\\%$ . We can observe that a noticeable improvement of $3.62\\%$ and $1.68\\%$ are obtained in the Base class and New class, respectively. The superior performance proves that TCP using the class-aware prompt to inject the class knowledge can generate", + "bbox": [ + 496, + 674, + 892, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23442", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a694851e5404f09d16231a263d24315b169cd2c260ef0809cd73de9a1f91957b.jpg", + "table_caption": [ + "Table 2. Comparison of cross-dataset evaluation. 'tp', 'dtp', 'vp', and 'dvp' denote the 'textual prompt', 'deep textual prompt', 'visual prompt', and 'deep visual prompt', respectively. Note that DAPT and MaPLe are based on visual-textual prompt tuning('vp+tp')." + ], + "table_footnote": [], + "table_body": "
DatasetsCLIPCoOpProGradKgCoOpDePTVPTPLOTPromptSRCMaPLeDAPTTCP
tptptptptp+vptp+vpdtp+tvpdtp+dvptp+vptp
ImageNet66.7071.5172.2470.6672.7769.7371.6071.2770.7271.6071.40
Caltech10193.3093.7091.5293.9294.2393.6792.0793.6093.5393.5093.97
OxfordPets89.1089.1489.6489.8390.0389.2790.1090.2590.4990.6791.25
StandfordCars65.7064.5162.3965.4165.5765.565.7065.7065.5765.9364.69
Flowers70.7068.7167.8770.0170.5770.269.2370.2572.2071.7071.21
Food10185.9085.3085.4086.3686.3786.2786.2386.1586.2086.1086.69
FGVCAircraft24.9018.4720.1622.5123.2722.1325.0023.9024.7423.0323.45
SUN39762.6064.1562.4766.1666.6766.5761.6767.1067.0167.0067.15
DTD44.3041.9239.4246.3545.9746.9338.6046.8746.4944/0044.35
EuroSAT48.3046.3943.4646.0443.5347.4347.8345.5048.0652.4751.45
UCF10167.6066.5564.2968.5069.3067.2067.0068.7568.6968.7368.73
Avg.65.2463.8862.7165.5165.5565.5264.3465.8166.3066.3166.29
", + "bbox": [ + 148, + 128, + 821, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the discriminative textual classifier for both seen and unseen classes. DePT [47] is another work that also treats KgCoOp as the baseline and obtains the best $H$ of $79.10\\%$ among all textual prompt tuning methods. Compared to DePT, the proposed TCP obtains the performance of $84.35\\% / 75.28\\% / 79.56\\%$ in the metric term of Base/New/H, achieving the improvement of $0.51\\% / 0.32\\% / 0.41\\%$ .", + "bbox": [ + 75, + 321, + 470, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual-Textual Prompt Tuning: Different from the textual prompt tuning merely adapts Text Encoder, DAPT [5], PLOT [4], RPO [22], MaPLe [18], and PromptSRC [19] all jointly performs prompt tuning on the text and visual encoders for both improving the discriminative of textual embedding and visual embeddings. Unlike CoOp and TCP, which conduct the prompt tuning on one or two layers, those methods conduct the visual-textual prompt tunings on several layers of visual and text encoders. Compared to them, the proposed TCP with shallow textual prompt tuning obtains a higher performance, e.g., $79.51\\%$ vs $79.31\\%$ for PromptSRC in terms of $H$ . Compared to deep visual-textual prompt tuning, the superior performance demonstrates that class-aware prompt tuning can generate discriminative and generalization textual-level classifiers.", + "bbox": [ + 75, + 429, + 473, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Cross-Dataset Generalization", + "text_level": 1, + "bbox": [ + 76, + 665, + 341, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the Base-to-New generalization setting, New classes always have a similar data distribution as the Base classes. To further verify the generalization of the proposed TCP, in the Cross-Dataset generalization, TCP is trained from the ImageNet and directly evaluated on the unrelated datasets, e.g., the rest ten datasets. The comparison between the proposed TCP and existing methods is summarized in Table 2. From Table 2, we can observe that the proposed TCP obtains the highest average performance among all textual prompt tuning methods (66.29% vs 65.55% of DePT [47]), and obtains a comparable performance with the visual-textual prompt tuning methods (66.29% vs 66.31% of DAPT [5]), demonstrating the effectiveness of TCP in learning the generalization knowledge.", + "bbox": [ + 75, + 688, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fbf152178c98eacaff336c354fe5b3e2534150cd88ea03d6da70c4ab3db45d30.jpg", + "table_caption": [ + "Table 3. Effect of TCP on CoOp, KgCoOp, ProGrad, PromptSRC, and DAPT in the base-to-new generalization setting regarding final average performance." + ], + "table_footnote": [], + "table_body": "
MethodsBaseNewH
CoOp [50]82.6367.9974.6
CoOp+TKE83.10(↑ 0.47)70.17(↑ 2.18)76.09 (↑ 1.49)
KgCoOp [43]80.7373.677.00
KgCoOp +TKE84.13 (↑ 3.40)75.36(↑ 1.76)79.51 (↑ 2.51)
ProGrad [51]82.4870.7576.16
ProGrad+TKE82.61 (↑ 0.13)72.91(↑ 2.16)77.46 (↑ 1.30)
PromptSRC [19]82.0774.8378.03
PromptSRC+TKE83.74 (↑ 1.67)75.85(↑ 1.02)79.60 (↑ 1.57)
DAPT [5]83.1869.2775.59
DAPT+TKE84.15 (↑ 0.97)74.87(↑ 5.60)79.24 (↑ 3.65)
", + "bbox": [ + 535, + 373, + 856, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Few-shot Classification", + "text_level": 1, + "bbox": [ + 498, + 518, + 715, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To verify that the proposed TCP can infer the class-aware knowledge, the few-shot classification is conducted on all 11 datasets using K-shot labeled source images and evaluated on the standard testing domain with the same class space as the training classes. The 4-shot setting comparison between the proposed TCP and existing methods is summarized in Table 4. We can observe that the proposed TCP achieves a higher average performance than existing methods, i.e., obtaining the average performance of $76.32\\%$ .", + "bbox": [ + 496, + 542, + 893, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 498, + 689, + 653, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of TKE's generalization. The critical component of TCP is the class-aware prompt generated by TKE, which embeds the prior class knowledge and can be easily integrated with existing CoOp-based methods. As shown in Table 3, CoOp, KgCoOp, ProGrad, PromptSRC, and DAPT have all demonstrated improved performance when TKE is used. Specifically, improvements of $1.49\\%$ , $2.51\\%$ , $1.30\\%$ , $1.57\\%$ , and $3.65\\%$ were achieved for CoOp, KgCoOp, ProGrad, PromptSRC3, and DAPT, respectively. This superior performance highlights TCP's plug-and-play functionality,", + "bbox": [ + 496, + 712, + 893, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3PromptSRC is re-implemented with the shallow visual-textual prompt tuning.", + "bbox": [ + 500, + 875, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23443", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/28a6341b11fc12a5141aeb2206d65f63bb476ab88923c33de8e01cdce5235b87.jpg", + "table_caption": [ + "Table 4. Comparison of few-shot learning with 4-shot samples." + ], + "table_footnote": [], + "table_body": "
CLIPCoOpCoCoOpProGradKgCoOpMaPLeTIP-Adapter-FDAPTPromptSRCPLOTTaskResTCP
ImageNet66.7069.3770.5570.2170.1970.6770.7870.8070.8070.4062.8770.48
Caltech10193.3094.4494.9894.9394.6594.3094.7794.2394.7795.1394.6795.00
OxfordPets89.1091.3093.0193.2193.2092.0592.2692.1793.2392.5592.0091.90
StandfordCars65.7072.7369.1071.7571.9868.7074.4274.4071.8374.9375.9076.30
Flowers70.7091.1482.5689.9890.6980.8092.9892.3791.3192.9391.5094.40
Food10185.9082.5886.6485.7786.5986.9086.1883.6086.0686.4686.0385.3
FGVCAircraft24.9033.1830.8732.9332.4729.0335.4932.4732.8035.2933.8036.20
SUN39762.6070.1370.5071.1771.7971.4770.6572.2072.8070.4272.7072.11
DTD44.3058.5754.7957.7258.3154.7361.7061.3760.6462.4359.5763.97
EuroSAT48.3068.6263.8370.8471.0654.8778.2772.7375.0280.7072.8777.43
UCF10167.6077.4174.9977.8278.4073.7079.7379.4079.3579.7676.1080.83
Avg.65.3773.5971.9874.2174.4870.6676.1175.0775.3376.4574.3676.72
", + "bbox": [ + 94, + 114, + 874, + 268 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3a4426026c1df7113bc2af0d171a17f81ee2d2f4c4863930abbcf53d206fa2c1.jpg", + "table_caption": [ + "Table 5. Domain-shared prompt vs Class-aware prompt." + ], + "table_footnote": [], + "table_body": "
BaselineDomain-SharedClass-AwareBaseNewH
✓✓80.7373.677
✓✓84.0575.1879.36
✓✓✓84.1375.3679.51
", + "bbox": [ + 88, + 306, + 455, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7fd9172650d9bdbfc8186a426a2601cc631977f3ed0728679d2df6e5b7f0205.jpg", + "image_caption": [ + "Figure 3. Effect of $M$ on the base-to-new generalization setting." + ], + "image_footnote": [ + "$D_{mid}$ in TKE." + ], + "bbox": [ + 122, + 397, + 411, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "allowing easy integration with existing methods.", + "bbox": [ + 76, + 549, + 398, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Domain-shared prompt vs the class-aware prompt. It is common for existing methods to infer the domain-shared prompt tokens. However, TCP differs significantly from them in using a class-aware prompt to capture class-aware knowledge in the Text Encoder. We thus compare the domain-shared and class-aware prompts, and summarize the related results in Table 5. We can observe that simply using the class-aware prompt yields a fantastic performance of $79.36\\%$ in terms of $H$ , higher than existing methods. Additionally, as shown in Table 5, using the class-aware prompt results in a noticeable improvement over the domain-shared prompt. For instance, the Basel/New performance is improved from $80.73\\% / 73.6\\%$ to $84.05\\% / 75.18\\%$ . The superior performance proves the reasonableness and effectiveness of injecting the class-aware knowledge into the prompt tuning. By further considering the domain-shared and class-aware prompts, the final performance is reached at $79.51\\%$ .", + "bbox": [ + 76, + 566, + 468, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Prompt length $M$ . In this study, we examine the impact of prompt length $(M)$ on the Base-to-New generalization setting. We compared prompt lengths of 1, 2, 4, 8, and 16, and the results are presented in Figure 3. The", + "bbox": [ + 76, + 839, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/310d25b9fe2e461642adb82ddb093bef6b64e77316f175c279a345cfd04318a7.jpg", + "table_caption": [ + "Table 6. Comparison of different templates." + ], + "table_footnote": [], + "table_body": "
TemplatesBaseNewH
‘X X X X’84.1375.3679.51
‘a photo of a’83.9474.9479.18
‘this is a picture’84.0074.6679.06
", + "bbox": [ + 575, + 306, + 815, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b18c2a1da65e25595fdbf49500f754ac88e936722af28d90c4f5ff65b93a8f9.jpg", + "image_caption": [ + "Figure 4. Effect of dimension" + ], + "image_footnote": [], + "bbox": [ + 501, + 369, + 691, + 452 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad2c24d0f099c4e9ac56637ca53bc03d6b373c303ae46a847443b36ca6fd1f6a.jpg", + "image_caption": [ + "Figure 5. Effect of insert layer $l$ ." + ], + "image_footnote": [], + "bbox": [ + 694, + 376, + 880, + 457 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "findings indicate that longer prompt lengths produce better performance, with $M = 8$ producing the best results. However, the performance gap between the highest and lowest values of $H$ , which were $79.63\\%$ and $79.26\\%$ , respectively, is only $0.37\\%$ . Therefore, the proposed TCP is relatively insensitive to prompt length $(M)$ .", + "bbox": [ + 496, + 507, + 890, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of different templates. Unlike most CoOp-based methods using the handcrafted template \"a photo of a $\\{\\}$ \", the proposed TCP utilizes the random initialization template (\"X XXX $\\{\\}$ \"). Therefore, we conducted a comparison among different templates. Table 6 shows that using the random template performs better than the handcrafted template. This is because TCP uses the class-ware prompt without considering the domain-share prompt, which has achieved better performance. For instance, in Table 5, TCP with the class-aware prompt achieves comparable performance of $84.05\\% /75.18\\% /79.36\\%$ in terms of Base/New/H compared to existing methods. Moreover, the domain-shared prompt initialized as the random template provides complementary knowledge to the class-aware prompt.", + "bbox": [ + 496, + 598, + 892, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of $D_{mid}$ in TAK. We analyze the effect of the dimension $D_{mid}$ in TKE, and show the comparison in Figure 4. Setting $D_{mid} = 128$ obtains the best performance, and a smaller $D_{mid}$ would get a worse performance.", + "bbox": [ + 496, + 809, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Insert which layer? TCP inserts the class-aware textual prompts into a particular layer of the Text Encoder. The im-", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23444", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/43c447dd9ee37ff8c7435957d196b31c260aa323c171c29401d2c4916fecae84.jpg", + "table_caption": [ + "Table 7. Comparison with single-layer vs multiple layers prompt tuning." + ], + "table_footnote": [], + "table_body": "
ModesLayersBaseNewH
TCP-Shallow{8}84.1375.3679.51
TCP-Deeper{4:8}84.0275.1379.33
TCP-Deeper{8:10}84.2475.2979.51
TCP-Deeper{4:8;10}84.1174.5179.02
TCP-Deeper{4;6;8;10}84.1774.6679.13
", + "bbox": [ + 151, + 114, + 393, + 179 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b606b15fc1b82e7e6bfe74f92517ac8a56f790938ab53d20e244416c7b4698a.jpg", + "image_caption": [ + "Figure 6. Comparison with TCP and Adapter-based methods." + ], + "image_footnote": [], + "bbox": [ + 120, + 186, + 428, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pact of inserting prompts into different layers was analyzed and shown in Figure 5. The best performance was achieved when the class-aware prompt tuning was done on the 8th layer. This is because doing so on earlier layers ( $l < 8$ ) does not transfer class-aware knowledge to the textual embedding, and doing it on later layers can easily affect the textual embedding.", + "bbox": [ + 75, + 354, + 468, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Single-layer vs Multiple-layers? The results of existing MaPLe and PromptSRC show that deep prompt learning applied on multiple layers can perform better than shallow prompt tuning with a single layer. Therefore, we compared the performance of shallow prompt tuning with deep prompt tuning. The results in Table 7 indicate that the shallow class-aware prompt tuning with a single layer provides better performance. For instance, TCP-Shallow with 8-layer achieves the highest $H$ of $79.51\\%$ , which surpasses all variants of TCP-deeper that have two, three, and four layers. Additionally, we observed that TCP-shallow obtains the best performance of $75.36\\%$ for the New class, and using more layers results in lower performance.", + "bbox": [ + 75, + 460, + 468, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compared to Adapter-based methods: In TCP, TKE acts as a special adapter, leading to the proposed TCP is similar to Adapter-based methods. We compared TCP with several Adapter-based methods such as CLIP-Adapter and CoOp-Adapter, which apply the adapter on the general textual-level class embeddings. However, TCP applies TKE to transfer the textual-level class embedding into the class-aware prompt combined with the class tokens in the middle layer of the Text Encoder. As shown in Figure 6, TCP with TKE performs better than the existing CLIP-Adapter and CoOp-Adapter on all 11 datasets.", + "bbox": [ + 75, + 657, + 468, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training time analysis: We further analyze the training time of the proposed TCP. The training time was determined based on the average time to process one image from the ImageNet with a 16-shot setting. According to Table 8, the proposed TCP is an efficient method that performs better", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b85f55d4fe68936adeaf4f208d4be2142ff248a322feb2b5cdb8da32e97c4e5a.jpg", + "table_caption": [ + "Table 8. Training time comparison(ms/image)." + ], + "table_footnote": [], + "table_body": "
CoOpProGradKgCoOpPLOTRPOMaPLePromptSRCTCP
ms/image6ms22ms6ms78.8ms190ms90.7ms43.2ms6.4ms
H74.4876.1677.0077.3777.7878.5579.3179.51
", + "bbox": [ + 501, + 114, + 916, + 154 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea6c2b7fbe926823433b42f97e36931156db764ad71f5036cf0742e84e679fec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 162, + 696, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/293fc5e48e6789d6ae5faf265486c20247ac898afe8395cf8d31dbd6c65db2bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 162, + 852, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c5a2d314927fe6616f08c6243ecd749ce33b972dd85061b13afa5ce1b5df6d08.jpg", + "image_caption": [ + "Figure 7. Visualization of the probability of CoOp and TCP." + ], + "image_footnote": [], + "bbox": [ + 540, + 251, + 700, + 332 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/998afdab57dbe67dda51b7a2ffb65dc2df0565b7908f7a6191b7aedf10f38c35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 251, + 852, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with less running time. This is because TCP adds an efficient TKE component for generating class-aware prompt upon CoOp and KgCoOp, resulting in slightly more time consumption than CoOp and KgCoOp.", + "bbox": [ + 496, + 369, + 890, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization. To prove that the proposed TCP can generate the discriminative textual classifier for prediction, we further visualize the prediction probability among all classes. As shown in Figure 7, the proposed TCP has an obvious inter-class classes than existing methods.", + "bbox": [ + 496, + 434, + 892, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 532, + 617, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To improve the generalization and discriminative abilities of the learnable prompt, we introduce a Textual-based Class-aware Prompt tuning method that takes advantage of the benefits of general class-level textual knowledge. To achieve this, we propose a Textual Knowledge Embedding (TKE) that transfers the class-level textual embedding into the class-aware prompt. This is combined with pre-trained class tokens to generate task-specific textual knowledge. Several benchmarks and tasks have shown that the class-aware prompt is effective for prompt tuning.", + "bbox": [ + 496, + 561, + 892, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, the class-aware prompt in TCP heavily relies on the discriminative ability of the general textual embedding. On the other hand, a weaker textual embedding will produce a weaker textual classifier. For example, TCP does not perform well on the FGVCAircraft dataset. Therefore, in the future, we plan to explore ways to use weaker textual knowledge to obtain a discriminative textual classifier.", + "bbox": [ + 496, + 715, + 890, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: This work was supported by National Science and Technology Major Project(2021ZD0112202), National Natural Science Foundation of China (62376268,U23A20387,62036012,U21B2044,62376196), and Beijing Natural Science Foundation (4222039).", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23445", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lorenzo Agnolucci, Alberto Baldrati, Francesco Todino, Federico Becattini, Marco Bertini, and Alberto Del Bimbo. Eco: Ensembling context optimization for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2811-2815, 2023. 3", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 2", + "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI, pages 446-461. Springer, 2014. 5", + "[4] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. PLOT: prompt learning with optimal transport for vision-language models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 1, 3, 5, 6", + "[5] Eulrang Cho, Jooyeon Kim, and Hyunwoo J Kim. Distribution-aware prompt tuning for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22004-22013, 2023. 1, 3, 5, 6", + "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014, pages 3606-3613. IEEE Computer Society, 2014. 5", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248-255. IEEE Computer Society, 2009. 4", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 5", + "[9] Li Fei-Fei, Robert Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. Comput. Vis. Image Underst., 106(1):59-70, 2007. 4", + "[10] Zhe Gan, Linjie Li, Chunyuan Li, Lijuan Wang, Zicheng Liu, Jianfeng Gao, et al. Vision-language pre-training: Basics, recent advances, and future trends. Foundations and Trends® in Computer Graphics and Vision, 14(3-4):163-352, 2022. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. CoRR, abs/2110.04544, 2021. 3", + "[12] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 1", + "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15979-15988. IEEE, 2022. 2", + "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens., 12(7):2217-2226, 2019. 5", + "[15] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. Open-Review.net, 2022. 1", + "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 4904-4916. PMLR, 2021. 2", + "[17] Baoshuo Kan, Teng Wang, Wenpeng Lu, Xiantong Zhen, Weili Guan, and Feng Zheng. Knowledge-aware prompt tuning for generalizable vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15670-15680, 2023. 1, 3", + "[18] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 1, 3, 5, 6", + "[19] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 3, 5, 6", + "[20] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 5583-5594. PMLR, 2021. 2", + "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 2013 IEEE International Conference on Computer Vision" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23446", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Workshops, ICCV Workshops 2013, Sydney, Australia, December 1-8, 2013, pages 554-561. IEEE Computer Society, 2013. 5", + "[22] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1401-1411, 2023. 1, 2, 5, 6", + "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2", + "[24] Juncheng Li, Minghe Gao, Longhui Wei, Siliang Tang, Wenqiao Zhang, Mengze Li, Wei Ji, Qi Tian, Tat-Seng Chua, and Yueting Zhuang. Gradient-regulated meta-prompt learning for generalizable vision-language models. arXiv preprint arXiv:2303.06571, 2023. 2", + "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2", + "[26] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 13-23, 2019. 2", + "[27] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5196-5205. IEEE, 2022. 3, 5", + "[28] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. CoRR, abs/1306.5151, 2013. 5", + "[29] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Sixth Indian Conference on Computer Vision, Graphics & Image Processing, ICVGIP 2008, Bhubaneswar, India, 16-19 December 2008, pages 722-729. IEEE Computer Society, 2008. 5", + "[30] Maxime Aquab, Léon Bottou, Ivan Laptev, and Josef Sivic. Learning and transferring mid-level image representations using convolutional neural networks. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014, pages 1717-1724. IEEE Computer Society, 2014. 1", + "[31] Yassine Ouali, Adrian Bulat, Brais Matinez, and Georgios Tzimiropoulos. Black box few-shot adaptation for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15534-15546, 2023. 1, 2, 5", + "[32] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI," + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "USA, June 16-21, 2012, pages 3498-3505. IEEE Computer Society, 2012. 4", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 1, 2", + "[34] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 18061-18070. IEEE, 2022. 3", + "[35] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2", + "[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2", + "[37] Sheng Shen, Shijia Yang, Tianjun Zhang, Bohan Zhai, Joseph E. Gonzalez, Kurt Keutzer, and Trevor Darrell. Multitask vision-language prompt tuning. CoRR, abs/2211.11720, 2022. 3", + "[38] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2", + "[39] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR, abs/1212.0402, 2012. 5", + "[40] Jingchen Sun, Jiayu Qin, Zihao Lin, and Changyou Chen. Prompt tuning based adapter for vision-language model adaption. CoRR, abs/2303.15234, 2023. 1", + "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008, 2017. 2", + "[42] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In The Twenty-Third IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2010, San Francisco, CA, USA, 13-18 June 2010, pages 3485-3492. IEEE Computer Society, 2010. 5", + "[43] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23447", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 1, 2, 3, 5, 6", + "[44] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 10899-10909. IEEE, 2023. 5", + "[45] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Unified vision and language prompt learning. CoRR, abs/2210.07225, 2022. 1, 2, 3, 5", + "[46] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12104-12113, 2022. 2", + "[47] Ji Zhang, Shihan Wu, Lianli Gao, Hengtao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. arXiv preprint arXiv:2309.07439, 2023. 5, 6", + "[48] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of CLIP for few-shot classification. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 493-510. Springer, 2022. 5", + "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 16795-16804. IEEE, 2022. 1, 3, 5", + "[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. Int. J. Comput. Vis., 130(9):2337-2348, 2022. 1, 2, 4, 5, 6", + "[51] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 3, 5, 6", + "[52] Beier Zhu, Yulei Niu, Saeil Lee, Minhoe Hur, and Hanwang Zhang. Debiased fine-tuning for vision-language models by prompt regularization. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence, IAAI 2023, Thirteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2023, Washington, DC, USA, February 7-14, 2023, pages 3834-3842. AAAI Press, 2023. 1, 3" + ], + "bbox": [ + 78, + 90, + 470, + 710 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23448", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/137896e1-f0a6-4b80-849c-9a14cbb3250c_model.json b/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/137896e1-f0a6-4b80-849c-9a14cbb3250c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5da901cadf4e864a65791a28b6157b5bcb4e3368 --- /dev/null +++ b/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/137896e1-f0a6-4b80-849c-9a14cbb3250c_model.json @@ -0,0 +1,2235 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.131, + 0.868, + 0.154 + ], + "angle": 0, + "content": "TCP:Textual-based Class-aware Prompt tuning for Visual-Language Model" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.183, + 0.621, + 0.199 + ], + "angle": 0, + "content": "Hantao Yao\\(^{1}\\), Rui Zhang\\(^{2}\\), Changsheng Xu\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.199, + 0.878, + 0.216 + ], + "angle": 0, + "content": "\\(^{1}\\) State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, China" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.217, + 0.798, + 0.233 + ], + "angle": 0, + "content": "\\(^{2}\\) State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences, China;" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.234, + 0.644, + 0.251 + ], + "angle": 0, + "content": "3 University of Chinese Academy of Sciences, China" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.255, + 0.766, + 0.27 + ], + "angle": 0, + "content": "hantao.yao@nlpr.ia.ac.cn; zhangrui@ict.ac.cn; csxu@nlpr.ia.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.337, + 0.474, + 0.701 + ], + "angle": 0, + "content": "Prompt tuning represents a valuable technique for adapting pre-trained visual-language models (VLM) to various downstream tasks. Recent advancements in CoOp-based methods propose a set of learnable domain-shared or image-conditional textual tokens to facilitate the generation of task-specific textual classifiers. However, those textual tokens have a limited generalization ability regarding unseen domains, as they cannot dynamically adjust to the distribution of testing classes. To tackle this issue, we present a novel Textual-based Class-aware Prompt tuning(TCP) that explicitly incorporates prior knowledge about classes to enhance their discriminability. The critical concept of TCP involves leveraging Textual Knowledge Embedding (TKE) to map the high generalizability of class-level textual knowledge into class-aware textual tokens. By seamlessly integrating these class-aware prompts into the Text Encoder, a dynamic class-aware classifier is generated to enhance discriminability for unseen domains. During inference, TKE dynamically generates class-aware prompts related to the unseen classes. Comprehensive evaluations demonstrate that TKE serves as a plug-and-play module effortlessly combinable with existing methods. Furthermore, TCP consistently achieves superior performance while demanding less training time1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.731, + 0.21, + 0.746 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Large-scale image-text pairs have the ability to train a Visual-language model (VLM) with a powerful generalization capacity for various downstream tasks [2, 33]. However, training these models from scratch requires a huge dataset with labeled images, making it difficult to apply them directly to downstream tasks with fewer images. To address this issue, three common techniques have been pro" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.302, + 0.701, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.316, + 0.871, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.481, + 0.895, + 0.566 + ], + "angle": 0, + "content": "Figure 1. Comparison with existing frameworks. (a) Domain-shared prompt tuning applies the same learnable prompt between the training and testing domains. (b) Image-conditional prompt tuning combines the image embedding with the learnable prompt; (c) Class-aware Prompt tuning injects the class-level textual embedding into the Text Encoder with the class-aware prompt." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.892, + 0.636 + ], + "angle": 0, + "content": "posed: fine-tuning [30], prompt tuning [50], adapter [12] and LoRA [15]. Among them, prompt tuning is a simple and efficient framework that transfers the essential general knowledge of VLM to the downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.893, + 0.837 + ], + "angle": 0, + "content": "Prompt tuning\\(^2\\) is a technique that combines learnable textual tokens with class tokens to generate a discriminative textual classifier, known as Context Optimization (CoOp) [50]. Recently, various CoOp-based methods [4, 5, 17, 18, 22, 31, 40, 43, 50, 52] infer the domain-shared prompt tokens between training and testing domains (Figure 1(a)). Nevertheless, as the domain-shared prompt tokens are derived from labeled training images, their performance is suboptimal when confronted with unseen test classes. To enhance the generalization capacity of learnable prompt tokens, image-conditional prompts have been proposed in [45, 49] by fusing image features and learnable tokens (Figure 1(b)). Notably, image-conditional" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.851, + 0.894, + 0.901 + ], + "angle": 0, + "content": "2For the visual-language model, two types of prompt-tuning strategies exist: visual prompt tuning and textual prompt tuning. In this work, we focus on the textual prompt tuning and do not consider the visual prompt tuning." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "1https://github.com/htyao89/Textual-based_ Class-aware_prompt_tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "23438" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "textual tokens encapsulate specific knowledge of each image, particularly for testing images, thereby rendering it more apt for generalization to unseen testing images. However, image-conditional prompt tokens with image-specific knowledge have less ability in elevating the distribution of class embeddings. In summary, classifiers generated by domain-shared and image-conditional textual tokens exhibit suboptimal performance for the unseen classes, primarily due to their inability to explicitly model the class distribution. Consequently, it is imperative to establish a dynamic relationship between the learnable prompt and the textual knowledge of each class to augment its discriminative prowess." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.473, + 0.579 + ], + "angle": 0, + "content": "The frozen CLIP, coupled with a hand-crafted prompt, exhibits robust generalization capabilities to novel classes, rendering it a valuable source of prior textual knowledge for each class. By associating the class-level textual knowledge with the learnable prompts, a class-aware prompt can be formulated for hencing the discriminative capacity of the textual classifier. To achieve this, we employ an embedding module to project the class-aware textual knowledge into the class-aware prompt tokens, as shown in Figure 1(c). The resultant class-aware prompt incorporates prior textual knowledge specific to each class, endowing the generated textual classifier with heightened discriminative prowess. Furthermore, the class-aware prompt facilitates the generation of classifiers for both seen and unseen classes by leveraging textual knowledge from both categories. To sum up, the trained embedding module can generate a class-aware prompt for each class based on its description ('class name'), thereby enhancing generalization and discriminative capabilities in class-level textual embeddings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.473, + 0.807 + ], + "angle": 0, + "content": "Therefore, we propose a novel Textual-based Class-aware Prompt tuning (TCP) based on the framework of CoOp, shown in Figure 2. In addition to the domain-share textual tokens introduced in CoOp, TCP contributes a novel Textual Knowledge Embedding (TKE) to map the class-level textual knowledge into class-aware prompt tokens. Moreover, a class-aware textual classifier is generated by inserting the class-aware prompt tokens into the middle layer of the Text Encoder. We use standard contrastive loss and knowledge-guided consistency [43] to optimize the TKE and the learnable prompt tokens. During inference, TCP generates a class-aware classifier for unseen classes by feeding the domain-shared prompt tokens and the class-aware prompt tokens generated by the TKE into the frozen Text Encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Overall, the proposed TCP explicitly steers prompts to learn a class-aware knowledge that maximizes the generalization and discriminative of the downstream tasks. Evaluation over 11 image classification datasets on base-to-new generalization, cross-dataset generalization, and few-shot learning verify shows that TCP is an efficient method that" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.137 + ], + "angle": 0, + "content": "obtain a higher performance with less training time. In summary, the proposed Textual-based Class-aware Prompt tuning(TCP) has the following main contributions:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.138, + 0.894, + 0.196 + ], + "angle": 0, + "content": "1. An effectively Textual-based Class-aware Prompt tuning is proposed by injecting the textual class-aware prompts generated by Textual Knowledge Embedding(TKE) into the Text Encoder." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.198, + 0.892, + 0.242 + ], + "angle": 0, + "content": "2. We demonstrate that explicitly incorporating the prior knowledge of each class into the learnable prompt tokens can enhance the discriminative of the class distribution." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.243, + 0.894, + 0.288 + ], + "angle": 0, + "content": "3. Textual Knowledge Embedding(TKE) is a plug-and-play module that can quickly insert existing prompt tuning methods to improve their performance further." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.138, + 0.894, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.303, + 0.651, + 0.32 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.329, + 0.729, + 0.345 + ], + "angle": 0, + "content": "2.1. Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.895, + 0.701 + ], + "angle": 0, + "content": "Recently, researchers have shown that the Visual-Language Models(VLM) [2, 33] consisting of visual and textual modalities trained on the large-scale of image-text pairs has a powerful generalization and discriminativeion. To further boost the descriptive ability of VLM, the VLM models are boosted from the following aspect: 1) using a stronger text encoder or visual encoder [25, 41, 46]; 2) deeply fusing the visual and text knowledges [23, 38]; 3) using more images [16, 33, 35, 36]. To boost the diversity of text description, Masked Language Modeling (MLM) [20] [26] randomly erases the words in the text description used for representation learning. Unlike MLM, Masked autoencoder-based methods [13] are proposed to boost the descriptive ability by randomly masking the image patches. Among existing VLM models, CLIP is a representative and straightforward framework for inferring the independent visual and text encoder using the contrastive loss based on 400 million image-text association pairs. As CLIP has a good generalization, most existing CoOp-based methods are proposed based on CLIP for adapting the pre-trained VLM into the downstream task. Similar to existing methods, we conduct the prompt-tuning strategy on the TextEncoder of CLIP to obtain a task-specific textual embedding for prediction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.657, + 0.728 + ], + "angle": 0, + "content": "2.2. Prompt Tuning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To adapt the pretrained VLM to the downstream tasks, the prompt tuning [10, 22, 24, 31, 33, 45] always applies task-related textual tokens to infer the task-specific textual knowledge. In CLIP[33], the hand-crafted template \"a photo of a [CLASS]\" is used to embed the textual embedding for zero-shot prediction. However, the hand-crafted prompts have a poor ability to describe the downstream task. Textual prompt tuning is applied to boost the textual embedding by inferring a set of learnable textual tokens combined with the class tokens. For example, Context Optimization(CoOp) [50] replaces the hand-crafted prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23439" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.575 + ], + "angle": 0, + "content": "with the learnable soft prompts. To improve the generalization of the learnable textual prompt in CoOp, Conditional Context Optimization(CoCoOp) [49] and VPT [45] generates an image-conditional prompt fusing the image feature and the learnable textual prompt. Moreover, Knowledge-Guided Context Optimization(KgCoOp) [43], ProGrad [51], and Prompt Regularization(ProReg) [52] constrain the proposed learnable prompts contain the essential general knowledge. Unlike the above methods, which consider textual prompts, Ensembling Context Optimization(ECO) [1] employs prompt ensembling to combine multiple prompts. To obtain high-quality task-related tokens, ProDA [27] considers the prompt's prior distribution learning, and Distribution-Aware Prompt Tuning (DAPT) [5] optimizes the learnable prompt by maximizing inter-dispersion. Besides the textual knowledge from \"classname\", Knowledge-Aware Prompt Tuning(KAPT) [17] employs the external knowledge to generate the discriminative knowledge-aware prompt for the novel categories. PLOT [4] applies optimal transport to match the vision and text modalities for generating the discriminative and visual-aligned local textual prompt tokens. Besides the textual prompt tuning, Multi-modal Prompt Learning (MaPLe) [18] and PromptSRC [19] conduct the visual-textual prompt tuning by jointly conducting the prompt tuning on the visual and text encoders. Multitask Vision-Language Prompt Tuning(MVLPT) [37]incorporates crosstalk knowledge into prompt tuning for vision-language models. DenseCLIP [34] uses the context-aware prompt strategy to generate dense prediction tasks, and CLIP-Adapter [11] applies an adapter to adjust the visual or text embeddings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.473, + 0.807 + ], + "angle": 0, + "content": "Existing methods commonly infer two types of prompt tokens: domain-share and image-conditional. However, the textual classifiers generated with these tokens tend to perform poorly on unseen classes. To mitigate this limitation, we propose a novel Textual-based Class-aware Prompt tuning (TCP) that employs a dynamic class-aware token to enhance the generalization and discriminative capabilities of the learnable textual prompt. Moreover, we introduce a Textual Knowledge Embedding to project class-level textual knowledge into the class-aware prompts, which are then inserted into the Text Encoder to generate a discriminative class-aware classifier. The evaluation results demonstrate that integrating class-level prior knowledge into the prompt tokens significantly enhances the discriminative ability of the prompt tuning process." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.827, + 0.205, + 0.844 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "As the Textual-based Class-aware Prompt tuning(TCP) is proposed based on Context Optimization (CoOp), we first briefly review CoOp and then introduce the proposed TCP." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.649, + 0.106 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.382 + ], + "angle": 0, + "content": "Existing CoOp-based methods are proposed based on the powerful Contrastive Language-Image Pre-training(CLIP). Given an image along with its corresponding textual description, CLIP uses visual and text encoders to extract the visual and text embeddings. After that, the constrastive loss between the visual and textual embeddings is calculated to align those two embeddings. To effectively adapt CLIP for the downstream task, CLIP applies the hand-crafted template \"a photo of a \\(\\{\\}\\)\" to extract the general class-level textual embedding, defined as \\(\\mathbf{W}^{clip} = \\{\\mathbf{w}_i^{clip}\\}_{i=1}^{N_c}\\), where \\(\\mathbf{w}_i^{clip}\\) is the textual embedding of \\(i\\)-th class, and \\(N_c\\) is the number of classes. Given the 'class-name' of \\(i\\)-th class, Word Embedded \\(e(\\cdot)\\) firstly embeds the hand-crafted description into a vectorized textual tokens: \\(\\mathbf{t}_i^{clip} = e(\"a photo of a \\{class-name\\}\")\\). After that, Text Encoder \\(\\theta\\) maps the vectorized textual tokens \\(\\mathbf{t}_i^{clip}\\) into the class-level embedding: \\(\\mathbf{w}_i^{clip} = \\theta (\\mathbf{t}_i^{clip})\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.895, + 0.599 + ], + "angle": 0, + "content": "To improve the discriminative of the class-level embedding, the prompt tuning methods of Context Optimization(CoOp) replace the hand-crafted textual tokens with a set of learnable textual tokens \\(\\mathbb{T} = \\{\\mathbf{t}_1,\\mathbf{t}_2,\\dots,\\mathbf{t}_M\\}\\), where \\(M\\) is the length of tokens. Similar to CLIP, the corresponding class token \\(\\mathbf{c}_i\\) is concatenated with the learnable tokens \\(\\mathbb{T}\\) for generating the textual token \\(\\mathbf{t}_i^{coop} = \\{\\mathbf{t}_1,\\mathbf{t}_2,\\dots,\\mathbf{t}_M,\\mathbf{c}_i\\}\\). Next, the textual embedding \\(\\mathbf{w}_i^{coop}\\) is obtained by fedding the textual tokens \\(\\mathbf{t}_i^{coop}\\) into Text Encoder \\(\\theta\\), i.e., \\(\\mathbf{w}_i^{coop} = \\theta (\\mathbf{t}_i^{coop})\\). Finally, the textual embeddings of all classes are defined as \\(\\mathbf{W}^{coop} = \\{\\mathbf{w}_i^{coop}\\}_{i = 1}^{N_c}\\). CoOp infers the learnable textual tokens \\(\\mathbb{T}\\) by minimizing the contrastive loss between the image's embedding \\(\\mathbf{x}\\) and its class embedding \\(\\mathbf{w}_y^{coop}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.611, + 0.892, + 0.653 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c e} = \\frac {1}{N} \\sum_ {(\\mathbf {x}, y) \\in \\mathcal {D} _ {s}} \\frac {\\exp (d (\\mathbf {x} , \\mathbf {w} _ {y} ^ {c o o p}) / \\tau)}{\\sum_ {i = 1} ^ {N _ {c}} \\exp (d (\\mathbf {x} , \\mathbf {w} _ {i} ^ {c o o p}) / \\tau)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.667, + 0.892, + 0.713 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_s\\) is the seen dataset, and \\(d(\\cdot)\\) is the cosine distance. \\(\\tau\\) is a temperature factor defined in CLIP, and \\(N\\) is the number of training images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.777 + ], + "angle": 0, + "content": "As the generated textual embedding has a good generalization ability for the novel classes, \\(\\mathrm{KgCoOp}\\) further adds an efficient consistency \\(\\mathcal{L}_{kg}\\) between the generated embedding \\(\\mathbf{W}^{coop}\\) and the general embedding \\(\\mathbf{W}^{clip}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.789, + 0.892, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {k g} = | | \\mathbf {W} ^ {c l i p} - \\mathbf {W} ^ {c o o p} | | _ {2} ^ {2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.824, + 0.851, + 0.839 + ], + "angle": 0, + "content": "Therefore, a robust objective for prompt tuning is:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.855, + 0.892, + 0.872 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c e} + \\omega \\mathcal {L} _ {k g}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.81, + 0.902 + ], + "angle": 0, + "content": "where \\(\\omega\\) is set as 8.0 same as in KgCoOp [43]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23440" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.088, + 0.434, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.279, + 0.414, + 0.293 + ], + "angle": 0, + "content": "Figure 2. The framework of the proposed TCP." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.302, + 0.436, + 0.319 + ], + "angle": 0, + "content": "3.2. Textual-based Class-aware Prompt tuning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.473, + 0.702 + ], + "angle": 0, + "content": "Based on the pre-trained Text Encoder in CLIP, the textual prompt tuning aims to infer a set of the domain-shared or image-conditional textual tokens combined with the general class tokens for generating the specific class embeddings. However, the textual classifier generated by those textual tokens performs a worse generalization on the unseen classes because they can not model the distribution of the testing classes. Research has shown that utilizing the general textual knowledge extracted by frozen CLIP can create a discriminative prior knowledge of novel classes, enhancing the discriminative and generalization of the learnable prompt. Drawing on the benefit of the general textual knowledge of seen and unseen classes, we propose a Textual-based Class-aware Prompt tuning(TCP) to adapt the pretrained CLIP to the downstream task. As shown in Figure 2, TCP employs a Textual Knowledge Embedding(TKE) to transfer the general class-level textual embedding into the class-aware prompt, which is then combined with the learnable textual tokens for generating the class-aware classifiers. TKE is advantageous for unseen classes, as it generates class-specific prompts to obtain an unseen class-aware textual classifier with better discriminative abilities. Moreover, explicitly incorporating seen class-aware prompts can enhance the discriminative power of seen classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Given the general class-level textual embedding \\(\\mathbf{W}^{clip} \\in \\mathbb{R}^{N_c \\times D_t}\\) with \\(N_c\\) training classes, Textual Knowledge Embedding(TKE) \\(\\mathcal{T}(\\cdot)\\) is proposed to project the class-level embedding \\(\\mathbf{W}^{clip}\\) into the class-aware prompt \\(\\mathbf{T} = \\mathcal{T}(\\mathbf{W}^{clip})\\). As shown in Figure 2, TKE consists of two layers: down-project layer and up-project layer. The down-project layer uses the weight \\(\\mathbf{W}_{down} \\in \\mathbb{R}^{D_t \\times D_{mid}}\\) to project the textual embedding into the low-dimension feature with the dimension of \\(D_{mid}\\). Next, the weight \\(\\mathbf{W}_{up} \\in \\mathbb{R}^{D_{mid} \\times D'}\\) of the up-project layer maps the low-dimension feature into the high-dimension feature with the dimension of \\(D'\\). Note that \\(D'\\) is determined by the prompt's length \\(M\\) and dimension \\(D\\): \\(D' = M \\times D\\). In" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.09, + 0.892, + 0.166 + ], + "angle": 0, + "content": "summary, the general textual embedding \\(\\mathbf{W}^{clip} \\in \\mathbb{R}^{N_c \\times D_t}\\) can be projected into the class-aware textual tokens \\(\\mathbf{T} \\in \\mathbb{R}^{N_c \\times D'}\\), which is further reshaped into the shape of \\(\\mathbf{T} \\in \\mathbb{R}^{N_c \\times M \\times D}\\) for inserting into the middle layer of Text Encoder \\(\\theta\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.893, + 0.334 + ], + "angle": 0, + "content": "Assuming we insert the class-aware prompt \\(\\mathbf{T}\\) into \\(l\\)-th layer of Text Encoder \\(\\theta\\). We will give a detailed analysis of the hyperparameter \\(l\\) in the following. Similar to CoOp, by combining the domain-shared learnable textual tokens \\(\\mathbb{T} = [\\mathbf{t}_1,\\mathbf{t}_2,\\dots,\\mathbf{t}_M]\\) and the pre-trained class tokens \\(\\mathbf{C}\\) of all classes, we can obtain the input textual tokens \\(\\mathbf{F}_0 = \\{\\mathbb{T},\\mathbf{C}\\}\\) of the Text Encoder, where \\(\\mathbf{C} = \\{\\mathbf{c}_i\\}_{i}^{N_c}\\), and \\(\\mathbf{c}_i\\) is the vectorized textual tokens of \\(i\\)-th class. The textual token \\(\\mathbf{F}_0\\) is fed into the first \\(l\\) layers of Text Encoder for obtaining the middle-level textual embedding \\(\\mathbf{F}_l\\). Formally, the textual token \\(\\mathbf{F}_i(i\\leq l)\\) of \\(i\\)-th layer is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.343, + 0.892, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {i} = \\theta_ {i} \\left(\\mathbf {F} _ {i - 1}\\right), i \\in [ 1, l ], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.779, + 0.383 + ], + "angle": 0, + "content": "where \\(\\theta_{i}\\) is the \\(i\\)-th layer of Text Encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.49 + ], + "angle": 0, + "content": "For the textual tokens \\(\\mathbf{F}_l\\in \\mathbb{R}^{N_c\\times N_t\\times D}\\) and the class-aware prompt tokens \\(\\mathbf{T}\\in \\mathbb{R}^{N_c\\times M\\times D}\\), the first dimension is related to the number of classes. Therefore, the learnable prompt tokens are always inserted into the second dimension of \\(\\mathbf{F}_l\\), the same as CoOp. Formally, the class-aware prompt \\(\\mathbf{T}\\) is inserted into \\(\\mathbf{F}_l\\) to generate the class-aware enhanced tokens \\(\\mathbf{F}_l'\\)." + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.498, + 0.892, + 0.516 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {l} ^ {\\prime} = \\left[ \\mathbf {T} _ {1}, \\mathbf {T} _ {2}, \\dots , \\mathbf {T} _ {M}, \\mathbf {F} _ {l, M + 1}, \\mathbf {F} _ {l, M + 2}, \\dots , \\mathbf {F} _ {l, N _ {t}} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.569 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}_i\\) denotes the \\(i\\)-th index of \\(\\mathbf{T}\\) in the second dimension, and \\(\\mathbf{F}_{l,j}\\) denote the correspond \\(j\\)-th index of \\(\\mathbf{F}_l\\) in the second dimension, i.e., \\(\\mathbf{T}_i = \\mathbf{T}[:i,:]\\), and \\(\\mathbf{F}_{l,j} = \\mathbf{F}_l[:j,:]\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.615 + ], + "angle": 0, + "content": "After that, the class-enhanced textual tokens \\(\\mathbf{F}_l^{\\prime}\\) are fedded into the rest layers for generating the class-aware textual embedding," + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.624, + 0.892, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {i} ^ {\\prime} = \\theta_ {i} (\\mathbf {F} _ {i - 1} ^ {\\prime}), i \\in [ l + 1, L ]. \\qquad (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.892, + 0.696 + ], + "angle": 0, + "content": "The output of the last layer \\(\\mathbf{F}_L^{\\prime}\\) is treated as the class embedding \\(\\mathbf{W}^{tcp}\\) used for optimization with contrastive loss and knowledge-guided consistency loss in Eq. (3)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.708, + 0.634, + 0.725 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Similar to CoOp [50], we evaluate the effectiveness of TCP from the three types of tasks: 1) generalization from baseto-new classes within a dataset; 2) few-shot learning with \\( K \\)-shot labeled images; 3) cross-dataset generalization from the imagenet to other datasets. More detailed results will be provided in the Supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.833, + 0.691, + 0.849 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Dataset. We conduct the base-to-new generalization, few-shot learning, and cross-dataset generalization on 11 benchmarks, i.e., ImageNet [7], Caltech [9], OxfordPets [32]," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23441" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.133 + ], + "angle": 0, + "content": "Table 1. Comparison on the base-to-new generalization setting with 16-shot samples. 'tp', 'dtp', 'vp', and 'dvp' denote the 'textual prompt', 'deep textual prompt', 'visual prompt', and 'deep visual prompt', respectively. PromptSRC are based on deep visual-textual prompt tuning ('dvp+dtp'). *' denote the performance obtained by our re-implementation." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.143, + 0.946, + 0.546 + ], + "angle": 0, + "content": "
DatasetsSetsCoOp* (JCV22)CoCoOp (CVPR22)DAPT* (ICCV23)ProGrad* (ICCV23)ProDA (CVPR22)KgCoOp (CVPR23)RPO (ICCV23)PLOT* (ICLR23)LFA (ICCV23)MaPLe (CVPR23)DePT (Arxiv23)PromptSRC* (ICCV23)TCP
tptptp+vptptptpdtp+dvptp+vp-dtp+dvptpdtp+dvptp
AverageBase82.3880.4783.1882.4881.5680.7381.1383.9883.6282.2883.6284.1284.13
New67.9671.6969.2770.7572.3073.675.0071.7274.5675.1475.0475.0275.36
H74.4875.8375.5976.1676.6577.077.7877.3778.8378.5579.1079.3179.51
ImageNetBase76.4675.9876.8377.0275.4075.8376.6077.3076.8976.6677.0377.7577.27
New66.3170.4369.2766.6670.2369.9671.5769.8769.3670.5470.1370.7069.87
H71.0273.1072.8571.4672.7272.7874.0073.4072.9373.4773.4274.0673.38
Caltech101Base97.8097.9697.8398.0298.2797.7297.9798.5398.4197.7498.3098.1398.23
New93.2793.8193.0793.8993.2394.3994.3792.8093.9394.3694.6093.9094.67
H95.4895.8495.3995.9195.6896.0396.0395.5896.1396.0296.4195.9796.42
OxfordPetsBase94.4795.2095.0095.0795.4394.6594.6394.5095.1395.4394.3395.5094.67
New96.0097.6995.8397.6397.8397.7697.5096.8396.2397.7697.2397.4097.20
H95.2396.4395.4196.3396.6296.1896.0595.6595.6896.5895.7696.4495.92
CarsBase75.6770.4975.8077.6874.7071.7673.8779.0776.3272.9479.1378.4080.80
New67.5373.5963.9368.6371.2075.0475.5374.8074.8874.0075.4774.7374.13
H71.3772.0169.3672.8872.9173.3674.6976.8875.5973.4777.2675.5277.32
FlowersBase97.2794.8796.9795.5497.7095.0094.1397.9397.3495.9298.0097.9097.73
New67.1371.7560.9071.8768.6874.7376.6773.5375.4472.4676.3776.7775.57
H79.4481.7174.8182.0380.6683.6584.5083.9985.0082.5685.8486.0685.23
Food101Base89.3790.7090.3790.3790.3090.590.3389.8090.5290.7190.5090.6390.57
New88.7791.2991.3089.5988.5791.790.8391.3791.4892.0591.6091.5091.37
H89.0790.9990.8389.9889.4391.0990.5890.5891.0091.3891.0591.0690.97
AircraftBase39.6733.4139.9740.5436.9036.2137.3342.1341.4837.4443.2042.3041.97
New31.2323.7129.8027.5734.1333.5534.2033.7332.2935.6134.8336.9734.43
H34.9527.7434.1432.8235.4634.8335.7037.4636.3136.5038.5739.4637.83
SUN397Base80.8579.7480.9781.2678.6780.2980.6082.2082.1380.8282.3382.8382.63
New68.3476.8676.9774.1776.9376.5377.8073.6377.2078.7077.8079.0078.20
H74.0778.2778.9277.5577.7978.3679.1877.6879.5979.7580.0080.8780.35
DTDBase79.9777.0182.2377.3580.6777.5576.7081.9781.2980.3682.2082.6082.77
New48.6056.0054.2352.3556.4854.9962.1343.8060.6359.1859.1357.5058.07
H60.4664.8565.3662.4566.4464.3568.6157.0969.4668.1668.7867.8068.25
EuroSATBase90.1087.4994.7390.1183.9085.6486.6393.7093.4094.0789.0392.4091.63
New53.0060.0450.3360.8966.0064.3468.9762.6771.2473.2371.0768.4374.73
H66.7471.2165.7472.6773.8873.4876.7975.1180.8382.379.0478.6382.32
UCF101Base84.5382.3384.3084.3385.2382.8983.6786.6086.9783.0085.8086.9387.13
New67.3773.4576.3374.9471.9776.6775.4375.9077.4878.6677.2378.3380.77
H74.9877.6780.1279.3578.0479.6579.3480.9081.9580.7781.2982.4183.83
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.47, + 0.621 + ], + "angle": 0, + "content": "StanfordCars [21], Flowers [29], Food101 [3], FGVCAircraft [28], EuroSAT [14], UCF101 [39], DTD [6], and SUN397 [42]. A detailed introduction of those dataset is given in the Supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.47, + 0.758 + ], + "angle": 0, + "content": "Training Details. Our implementation is based on CoOp's code [50]. All experiments are conducted based on the CLIP with the backbone of ViT-B/16 [8]. The prompt's length \\( M \\) is set as 4 with random initialization of \"X X X X {}\". The final performance is averaged over three random seeds(1/2/3)for a fair comparison. Adam optimizer is applied for optimization with the learning rate of 2e-3 and the batch size of 32, and the training epochs is 50. All experiments are conducted on RTX 3090." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.47, + 0.835 + ], + "angle": 0, + "content": "Baselines. Recently CoOp-based methods are used for comparison, e.g., CoOp [50], CoCoOp [49], ProGrad [51], ProDA [27], KgCoOp [43], PromptSRC [19], MaPLe [18], LFA [31], DePT [47], DAPT [5], PLOT [4], TaskRes [44], RPO [22], VPT [45], and TIP-Adapter-F [48]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.377, + 0.862 + ], + "angle": 0, + "content": "4.2. Base-to-New Class Generalization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To verify the generalization ability of the prompt tuning, the base-to-new class generalization splits each dataset into two" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.892, + 0.669 + ], + "angle": 0, + "content": "disjoint subsets: Base and New classes, where Base classes are used to infer the learnable prompt and TKE, and New classes are used for evaluation. Note that each New class contains 16-shot samples. We summarize the comparison between the proposed TCP and existing methods in Table 1 from the following two aspects: Textual Prompt Tuning and Vivasal-Textual Prompt Tuning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Textual Prompt Tuning: Based on the visual feature extracted from the frozen visual encoder, the textual prompt tuning (e.g., CoOp, CoCoOp, ProGrad, ProDA, and Kg-CoOp) merely conducts the textual prompt tuning by inserting a set of learnable textual prompts into the Text Encoder for obtaining a discriminative textual classifier, which is same as our proposed TCP. Among all textual prompt tuning methods, KgCoOp [43] is a more substantial baseline, which is also the baseline of our TCP. Compared to KgCoOp, TCP improves the metric term of Base/New/H from \\(80.73\\% / 73.6\\% / 77.0\\%\\) to \\(84.35\\% / 75.28\\% / 79.56\\%\\). We can observe that a noticeable improvement of \\(3.62\\%\\) and \\(1.68\\%\\) are obtained in the Base class and New class, respectively. The superior performance proves that TCP using the class-aware prompt to inject the class knowledge can generate" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23442" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.12 + ], + "angle": 0, + "content": "Table 2. Comparison of cross-dataset evaluation. 'tp', 'dtp', 'vp', and 'dvp' denote the 'textual prompt', 'deep textual prompt', 'visual prompt', and 'deep visual prompt', respectively. Note that DAPT and MaPLe are based on visual-textual prompt tuning('vp+tp')." + }, + { + "type": "table", + "bbox": [ + 0.149, + 0.129, + 0.823, + 0.307 + ], + "angle": 0, + "content": "
DatasetsCLIPCoOpProGradKgCoOpDePTVPTPLOTPromptSRCMaPLeDAPTTCP
tptptptptp+vptp+vpdtp+tvpdtp+dvptp+vptp
ImageNet66.7071.5172.2470.6672.7769.7371.6071.2770.7271.6071.40
Caltech10193.3093.7091.5293.9294.2393.6792.0793.6093.5393.5093.97
OxfordPets89.1089.1489.6489.8390.0389.2790.1090.2590.4990.6791.25
StandfordCars65.7064.5162.3965.4165.5765.565.7065.7065.5765.9364.69
Flowers70.7068.7167.8770.0170.5770.269.2370.2572.2071.7071.21
Food10185.9085.3085.4086.3686.3786.2786.2386.1586.2086.1086.69
FGVCAircraft24.9018.4720.1622.5123.2722.1325.0023.9024.7423.0323.45
SUN39762.6064.1562.4766.1666.6766.5761.6767.1067.0167.0067.15
DTD44.3041.9239.4246.3545.9746.9338.6046.8746.4944/0044.35
EuroSAT48.3046.3943.4646.0443.5347.4347.8345.5048.0652.4751.45
UCF10167.6066.5564.2968.5069.3067.2067.0068.7568.6968.7368.73
Avg.65.2463.8862.7165.5165.5565.5264.3465.8166.3066.3166.29
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.471, + 0.429 + ], + "angle": 0, + "content": "the discriminative textual classifier for both seen and unseen classes. DePT [47] is another work that also treats KgCoOp as the baseline and obtains the best \\( H \\) of \\( 79.10\\% \\) among all textual prompt tuning methods. Compared to DePT, the proposed TCP obtains the performance of \\( 84.35\\% / 75.28\\% / 79.56\\% \\) in the metric term of Base/New/H, achieving the improvement of \\( 0.51\\% / 0.32\\% / 0.41\\% \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.474, + 0.656 + ], + "angle": 0, + "content": "Visual-Textual Prompt Tuning: Different from the textual prompt tuning merely adapts Text Encoder, DAPT [5], PLOT [4], RPO [22], MaPLe [18], and PromptSRC [19] all jointly performs prompt tuning on the text and visual encoders for both improving the discriminative of textual embedding and visual embeddings. Unlike CoOp and TCP, which conduct the prompt tuning on one or two layers, those methods conduct the visual-textual prompt tunings on several layers of visual and text encoders. Compared to them, the proposed TCP with shallow textual prompt tuning obtains a higher performance, e.g., \\(79.51\\%\\) vs \\(79.31\\%\\) for PromptSRC in terms of \\(H\\). Compared to deep visual-textual prompt tuning, the superior performance demonstrates that class-aware prompt tuning can generate discriminative and generalization textual-level classifiers." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.342, + 0.683 + ], + "angle": 0, + "content": "4.3. Cross-Dataset Generalization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.472, + 0.903 + ], + "angle": 0, + "content": "In the Base-to-New generalization setting, New classes always have a similar data distribution as the Base classes. To further verify the generalization of the proposed TCP, in the Cross-Dataset generalization, TCP is trained from the ImageNet and directly evaluated on the unrelated datasets, e.g., the rest ten datasets. The comparison between the proposed TCP and existing methods is summarized in Table 2. From Table 2, we can observe that the proposed TCP obtains the highest average performance among all textual prompt tuning methods (66.29% vs 65.55% of DePT [47]), and obtains a comparable performance with the visual-textual prompt tuning methods (66.29% vs 66.31% of DAPT [5]), demonstrating the effectiveness of TCP in learning the generalization knowledge." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.321, + 0.895, + 0.364 + ], + "angle": 0, + "content": "Table 3. Effect of TCP on CoOp, KgCoOp, ProGrad, PromptSRC, and DAPT in the base-to-new generalization setting regarding final average performance." + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.374, + 0.857, + 0.503 + ], + "angle": 0, + "content": "
MethodsBaseNewH
CoOp [50]82.6367.9974.6
CoOp+TKE83.10(↑ 0.47)70.17(↑ 2.18)76.09 (↑ 1.49)
KgCoOp [43]80.7373.677.00
KgCoOp +TKE84.13 (↑ 3.40)75.36(↑ 1.76)79.51 (↑ 2.51)
ProGrad [51]82.4870.7576.16
ProGrad+TKE82.61 (↑ 0.13)72.91(↑ 2.16)77.46 (↑ 1.30)
PromptSRC [19]82.0774.8378.03
PromptSRC+TKE83.74 (↑ 1.67)75.85(↑ 1.02)79.60 (↑ 1.57)
DAPT [5]83.1869.2775.59
DAPT+TKE84.15 (↑ 0.97)74.87(↑ 5.60)79.24 (↑ 3.65)
" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.52, + 0.716, + 0.535 + ], + "angle": 0, + "content": "4.4. Few-shot Classification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.895, + 0.68 + ], + "angle": 0, + "content": "To verify that the proposed TCP can infer the class-aware knowledge, the few-shot classification is conducted on all 11 datasets using K-shot labeled source images and evaluated on the standard testing domain with the same class space as the training classes. The 4-shot setting comparison between the proposed TCP and existing methods is summarized in Table 4. We can observe that the proposed TCP achieves a higher average performance than existing methods, i.e., obtaining the average performance of \\(76.32\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.69, + 0.655, + 0.707 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.895, + 0.867 + ], + "angle": 0, + "content": "Effect of TKE's generalization. The critical component of TCP is the class-aware prompt generated by TKE, which embeds the prior class knowledge and can be easily integrated with existing CoOp-based methods. As shown in Table 3, CoOp, KgCoOp, ProGrad, PromptSRC, and DAPT have all demonstrated improved performance when TKE is used. Specifically, improvements of \\(1.49\\%\\), \\(2.51\\%\\), \\(1.30\\%\\), \\(1.57\\%\\), and \\(3.65\\%\\) were achieved for CoOp, KgCoOp, ProGrad, PromptSRC3, and DAPT, respectively. This superior performance highlights TCP's plug-and-play functionality," + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.895, + 0.903 + ], + "angle": 0, + "content": "3PromptSRC is re-implemented with the shallow visual-textual prompt tuning." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "23443" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.298, + 0.09, + 0.674, + 0.105 + ], + "angle": 0, + "content": "Table 4. Comparison of few-shot learning with 4-shot samples." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.115, + 0.875, + 0.27 + ], + "angle": 0, + "content": "
CLIPCoOpCoCoOpProGradKgCoOpMaPLeTIP-Adapter-FDAPTPromptSRCPLOTTaskResTCP
ImageNet66.7069.3770.5570.2170.1970.6770.7870.8070.8070.4062.8770.48
Caltech10193.3094.4494.9894.9394.6594.3094.7794.2394.7795.1394.6795.00
OxfordPets89.1091.3093.0193.2193.2092.0592.2692.1793.2392.5592.0091.90
StandfordCars65.7072.7369.1071.7571.9868.7074.4274.4071.8374.9375.9076.30
Flowers70.7091.1482.5689.9890.6980.8092.9892.3791.3192.9391.5094.40
Food10185.9082.5886.6485.7786.5986.9086.1883.6086.0686.4686.0385.3
FGVCAircraft24.9033.1830.8732.9332.4729.0335.4932.4732.8035.2933.8036.20
SUN39762.6070.1370.5071.1771.7971.4770.6572.2072.8070.4272.7072.11
DTD44.3058.5754.7957.7258.3154.7361.7061.3760.6462.4359.5763.97
EuroSAT48.3068.6263.8370.8471.0654.8778.2772.7375.0280.7072.8777.43
UCF10167.6077.4174.9977.8278.4073.7079.7379.4079.3579.7676.1080.83
Avg.65.3773.5971.9874.2174.4870.6676.1175.0775.3376.4574.3676.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.106, + 0.281, + 0.441, + 0.296 + ], + "angle": 0, + "content": "Table 5. Domain-shared prompt vs Class-aware prompt." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.307, + 0.457, + 0.368 + ], + "angle": 0, + "content": "
BaselineDomain-SharedClass-AwareBaseNewH
✓✓80.7373.677
✓✓84.0575.1879.36
✓✓✓84.1375.3679.51
" + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.398, + 0.413, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.521, + 0.465, + 0.535 + ], + "angle": 0, + "content": "Figure 3. Effect of \\(M\\) on the base-to-new generalization setting." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.55, + 0.4, + 0.565 + ], + "angle": 0, + "content": "allowing easy integration with existing methods." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.567, + 0.47, + 0.838 + ], + "angle": 0, + "content": "Domain-shared prompt vs the class-aware prompt. It is common for existing methods to infer the domain-shared prompt tokens. However, TCP differs significantly from them in using a class-aware prompt to capture class-aware knowledge in the Text Encoder. We thus compare the domain-shared and class-aware prompts, and summarize the related results in Table 5. We can observe that simply using the class-aware prompt yields a fantastic performance of \\(79.36\\%\\) in terms of \\(H\\), higher than existing methods. Additionally, as shown in Table 5, using the class-aware prompt results in a noticeable improvement over the domain-shared prompt. For instance, the Basel/New performance is improved from \\(80.73\\% / 73.6\\%\\) to \\(84.05\\% / 75.18\\%\\). The superior performance proves the reasonableness and effectiveness of injecting the class-aware knowledge into the prompt tuning. By further considering the domain-shared and class-aware prompts, the final performance is reached at \\(79.51\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Effect of Prompt length \\(M\\). In this study, we examine the impact of prompt length \\((M)\\) on the Base-to-New generalization setting. We compared prompt lengths of 1, 2, 4, 8, and 16, and the results are presented in Figure 3. The" + }, + { + "type": "table_caption", + "bbox": [ + 0.565, + 0.281, + 0.826, + 0.296 + ], + "angle": 0, + "content": "Table 6. Comparison of different templates." + }, + { + "type": "table", + "bbox": [ + 0.576, + 0.307, + 0.816, + 0.366 + ], + "angle": 0, + "content": "
TemplatesBaseNewH
‘X X X X’84.1375.3679.51
‘a photo of a’83.9474.9479.18
‘this is a picture’84.0074.6679.06
" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.371, + 0.692, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.377, + 0.882, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.465, + 0.693, + 0.477 + ], + "angle": 0, + "content": "Figure 4. Effect of dimension" + }, + { + "type": "image_footnote", + "bbox": [ + 0.503, + 0.478, + 0.59, + 0.492 + ], + "angle": 0, + "content": "\\(D_{mid}\\) in TKE." + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.472, + 0.882, + 0.487 + ], + "angle": 0, + "content": "Figure 5. Effect of insert layer \\(l\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.599 + ], + "angle": 0, + "content": "findings indicate that longer prompt lengths produce better performance, with \\( M = 8 \\) producing the best results. However, the performance gap between the highest and lowest values of \\( H \\), which were \\( 79.63\\% \\) and \\( 79.26\\% \\), respectively, is only \\( 0.37\\% \\). Therefore, the proposed TCP is relatively insensitive to prompt length \\( (M) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.893, + 0.81 + ], + "angle": 0, + "content": "Effect of different templates. Unlike most CoOp-based methods using the handcrafted template \"a photo of a \\(\\{\\}\\)\", the proposed TCP utilizes the random initialization template (\"X XXX \\(\\{\\}\\)\"). Therefore, we conducted a comparison among different templates. Table 6 shows that using the random template performs better than the handcrafted template. This is because TCP uses the class-ware prompt without considering the domain-share prompt, which has achieved better performance. For instance, in Table 5, TCP with the class-aware prompt achieves comparable performance of \\(84.05\\% /75.18\\% /79.36\\%\\) in terms of Base/New/H compared to existing methods. Moreover, the domain-shared prompt initialized as the random template provides complementary knowledge to the class-aware prompt." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.871 + ], + "angle": 0, + "content": "Effect of \\(D_{mid}\\) in TAK. We analyze the effect of the dimension \\(D_{mid}\\) in TKE, and show the comparison in Figure 4. Setting \\(D_{mid} = 128\\) obtains the best performance, and a smaller \\(D_{mid}\\) would get a worse performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Insert which layer? TCP inserts the class-aware textual prompts into a particular layer of the Text Encoder. The im-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23444" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.09, + 0.467, + 0.104 + ], + "angle": 0, + "content": "Table 7. Comparison with single-layer vs multiple layers prompt tuning." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.115, + 0.394, + 0.18 + ], + "angle": 0, + "content": "
ModesLayersBaseNewH
TCP-Shallow{8}84.1375.3679.51
TCP-Deeper{4:8}84.0275.1379.33
TCP-Deeper{8:10}84.2475.2979.51
TCP-Deeper{4:8;10}84.1174.5179.02
TCP-Deeper{4;6;8;10}84.1774.6679.13
" + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.187, + 0.429, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.332, + 0.456, + 0.346 + ], + "angle": 0, + "content": "Figure 6. Comparison with TCP and Adapter-based methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.47, + 0.461 + ], + "angle": 0, + "content": "pact of inserting prompts into different layers was analyzed and shown in Figure 5. The best performance was achieved when the class-aware prompt tuning was done on the 8th layer. This is because doing so on earlier layers (\\(l < 8\\)) does not transfer class-aware knowledge to the textual embedding, and doing it on later layers can easily affect the textual embedding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.469, + 0.657 + ], + "angle": 0, + "content": "Single-layer vs Multiple-layers? The results of existing MaPLe and PromptSRC show that deep prompt learning applied on multiple layers can perform better than shallow prompt tuning with a single layer. Therefore, we compared the performance of shallow prompt tuning with deep prompt tuning. The results in Table 7 indicate that the shallow class-aware prompt tuning with a single layer provides better performance. For instance, TCP-Shallow with 8-layer achieves the highest \\( H \\) of \\( 79.51\\% \\), which surpasses all variants of TCP-deeper that have two, three, and four layers. Additionally, we observed that TCP-shallow obtains the best performance of \\( 75.36\\% \\) for the New class, and using more layers results in lower performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Compared to Adapter-based methods: In TCP, TKE acts as a special adapter, leading to the proposed TCP is similar to Adapter-based methods. We compared TCP with several Adapter-based methods such as CLIP-Adapter and CoOp-Adapter, which apply the adapter on the general textual-level class embeddings. However, TCP applies TKE to transfer the textual-level class embedding into the class-aware prompt combined with the class tokens in the middle layer of the Text Encoder. As shown in Figure 6, TCP with TKE performs better than the existing CLIP-Adapter and CoOp-Adapter on all 11 datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Training time analysis: We further analyze the training time of the proposed TCP. The training time was determined based on the average time to process one image from the ImageNet with a 16-shot setting. According to Table 8, the proposed TCP is an efficient method that performs better" + }, + { + "type": "table_caption", + "bbox": [ + 0.557, + 0.09, + 0.835, + 0.104 + ], + "angle": 0, + "content": "Table 8. Training time comparison(ms/image)." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.115, + 0.918, + 0.155 + ], + "angle": 0, + "content": "
CoOpProGradKgCoOpPLOTRPOMaPLePromptSRCTCP
ms/image6ms22ms6ms78.8ms190ms90.7ms43.2ms6.4ms
H74.4876.1677.0077.3777.7878.5579.3179.51
" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.164, + 0.697, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.164, + 0.854, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.252, + 0.702, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.252, + 0.853, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.343, + 0.875, + 0.358 + ], + "angle": 0, + "content": "Figure 7. Visualization of the probability of CoOp and TCP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.433 + ], + "angle": 0, + "content": "with less running time. This is because TCP adds an efficient TKE component for generating class-aware prompt upon CoOp and KgCoOp, resulting in slightly more time consumption than CoOp and KgCoOp." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.893, + 0.511 + ], + "angle": 0, + "content": "Visualization. To prove that the proposed TCP can generate the discriminative textual classifier for prediction, we further visualize the prediction probability among all classes. As shown in Figure 7, the proposed TCP has an obvious inter-class classes than existing methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.534, + 0.619, + 0.55 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.893, + 0.715 + ], + "angle": 0, + "content": "To improve the generalization and discriminative abilities of the learnable prompt, we introduce a Textual-based Class-aware Prompt tuning method that takes advantage of the benefits of general class-level textual knowledge. To achieve this, we propose a Textual Knowledge Embedding (TKE) that transfers the class-level textual embedding into the class-aware prompt. This is combined with pre-trained class tokens to generate task-specific textual knowledge. Several benchmarks and tasks have shown that the class-aware prompt is effective for prompt tuning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.822 + ], + "angle": 0, + "content": "However, the class-aware prompt in TCP heavily relies on the discriminative ability of the general textual embedding. On the other hand, a weaker textual embedding will produce a weaker textual classifier. For example, TCP does not perform well on the FGVCAircraft dataset. Therefore, in the future, we plan to explore ways to use weaker textual knowledge to obtain a discriminative textual classifier." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements: This work was supported by National Science and Technology Major Project(2021ZD0112202), National Natural Science Foundation of China (62376268,U23A20387,62036012,U21B2044,62376196), and Beijing Natural Science Foundation (4222039)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23445" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Lorenzo Agnolucci, Alberto Baldrati, Francesco Todino, Federico Becattini, Marco Bertini, and Alberto Del Bimbo. Eco: Ensembling context optimization for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2811-2815, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.47, + 0.339 + ], + "angle": 0, + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI, pages 446-461. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.47, + 0.422 + ], + "angle": 0, + "content": "[4] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. PLOT: prompt learning with optimal transport for vision-language models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.47, + 0.493 + ], + "angle": 0, + "content": "[5] Eulrang Cho, Jooyeon Kim, and Hyunwoo J Kim. Distribution-aware prompt tuning for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22004-22013, 2023. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.47, + 0.577 + ], + "angle": 0, + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014, pages 3606-3613. IEEE Computer Society, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.58, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248-255. IEEE Computer Society, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.664, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.776, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[9] Li Fei-Fei, Robert Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. Comput. Vis. Image Underst., 106(1):59-70, 2007. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[10] Zhe Gan, Linjie Li, Chunyuan Li, Lijuan Wang, Zicheng Liu, Jianfeng Gao, et al. Vision-language pre-training: Basics, recent advances, and future trends. Foundations and Trends® in Computer Graphics and Vision, 14(3-4):163-352, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[11] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. CoRR, abs/2110.04544, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[12] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15979-15988. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens., 12(7):2217-2226, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[15] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. Open-Review.net, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[17] Baoshuo Kan, Teng Wang, Wenpeng Lu, Xiantong Zhen, Weili Guan, and Feng Zheng. Knowledge-aware prompt tuning for generalizable vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15670-15680, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[18] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[19] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[20] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 5583-5594. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 2013 IEEE International Conference on Computer Vision" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23446" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "Workshops, ICCV Workshops 2013, Sydney, Australia, December 1-8, 2013, pages 554-561. IEEE Computer Society, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.207 + ], + "angle": 0, + "content": "[22] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1401-1411, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.346 + ], + "angle": 0, + "content": "[24] Juncheng Li, Minghe Gao, Longhui Wei, Siliang Tang, Wenqiao Zhang, Mengze Li, Wei Ji, Qi Tian, Tat-Seng Chua, and Yueting Zhuang. Gradient-regulated meta-prompt learning for generalizable vision-language models. arXiv preprint arXiv:2303.06571, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.501 + ], + "angle": 0, + "content": "[26] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 13-23, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[27] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5196-5205. IEEE, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[28] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. CoRR, abs/1306.5151, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[29] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Sixth Indian Conference on Computer Vision, Graphics & Image Processing, ICVGIP 2008, Bhubaneswar, India, 16-19 December 2008, pages 722-729. IEEE Computer Society, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[30] Maxime Aquab, Léon Bottou, Ivan Laptev, and Josef Sivic. Learning and transferring mid-level image representations using convolutional neural networks. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014, pages 1717-1724. IEEE Computer Society, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[31] Yassine Ouali, Adrian Bulat, Brais Matinez, and Georgios Tzimiropoulos. Black box few-shot adaptation for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15534-15546, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[32] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "USA, June 16-21, 2012, pages 3498-3505. IEEE Computer Society, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[34] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 18061-18070. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[35] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.385, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.522 + ], + "angle": 0, + "content": "[37] Sheng Shen, Shijia Yang, Tianjun Zhang, Bohan Zhai, Joseph E. Gonzalez, Kurt Keutzer, and Trevor Darrell. Multitask vision-language prompt tuning. CoRR, abs/2211.11720, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[38] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[39] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR, abs/1212.0402, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[40] Jingchen Sun, Jiayu Qin, Zihao Lin, and Changyou Chen. Prompt tuning based adapter for vision-language model adaption. CoRR, abs/2303.15234, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.777, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[42] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In The Twenty-Third IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2010, San Francisco, CA, USA, 13-18 June 2010, pages 3485-3492. IEEE Computer Society, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[43] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23447" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[44] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 10899-10909. IEEE, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[45] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Unified vision and language prompt learning. CoRR, abs/2210.07225, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[46] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12104-12113, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[47] Ji Zhang, Shihan Wu, Lianli Gao, Hengtao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. arXiv preprint arXiv:2309.07439, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[48] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of CLIP for few-shot classification. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 493-510. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 16795-16804. IEEE, 2022. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. Int. J. Comput. Vis., 130(9):2337-2348, 2022. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[51] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.469, + 0.711 + ], + "angle": 0, + "content": "[52] Beier Zhu, Yulei Niu, Saeil Lee, Minhoe Hur, and Hanwang Zhang. Debiased fine-tuning for vision-language models by prompt regularization. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence, IAAI 2023, Thirteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2023, Washington, DC, USA, February 7-14, 2023, pages 3834-3842. AAAI Press, 2023. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23448" + } + ] +] \ No newline at end of file diff --git a/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/137896e1-f0a6-4b80-849c-9a14cbb3250c_origin.pdf b/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/137896e1-f0a6-4b80-849c-9a14cbb3250c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..36f94b207d8baaad12a83d0bdc0b224a46338002 --- /dev/null +++ b/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/137896e1-f0a6-4b80-849c-9a14cbb3250c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c4fa15e15517a732b9776241b44f5c876b70c0939493a85a46cd6f7d940efe3 +size 1882847 diff --git a/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/full.md b/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cf6ab73e656dd2650fed165f0526659ee65f2dd6 --- /dev/null +++ b/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/full.md @@ -0,0 +1,309 @@ +# TCP:Textual-based Class-aware Prompt tuning for Visual-Language Model + +Hantao Yao $^{1}$ , Rui Zhang $^{2}$ , Changsheng Xu $^{1,3}$ + +$^{1}$ State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, China + +$^{2}$ State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences, China; + +3 University of Chinese Academy of Sciences, China + +hantao.yao@nlpr.ia.ac.cn; zhangrui@ict.ac.cn; csxu@nlpr.ia.ac.cn + +# Abstract + +Prompt tuning represents a valuable technique for adapting pre-trained visual-language models (VLM) to various downstream tasks. Recent advancements in CoOp-based methods propose a set of learnable domain-shared or image-conditional textual tokens to facilitate the generation of task-specific textual classifiers. However, those textual tokens have a limited generalization ability regarding unseen domains, as they cannot dynamically adjust to the distribution of testing classes. To tackle this issue, we present a novel Textual-based Class-aware Prompt tuning(TCP) that explicitly incorporates prior knowledge about classes to enhance their discriminability. The critical concept of TCP involves leveraging Textual Knowledge Embedding (TKE) to map the high generalizability of class-level textual knowledge into class-aware textual tokens. By seamlessly integrating these class-aware prompts into the Text Encoder, a dynamic class-aware classifier is generated to enhance discriminability for unseen domains. During inference, TKE dynamically generates class-aware prompts related to the unseen classes. Comprehensive evaluations demonstrate that TKE serves as a plug-and-play module effortlessly combinable with existing methods. Furthermore, TCP consistently achieves superior performance while demanding less training time1. + +# 1. Introduction + +Large-scale image-text pairs have the ability to train a Visual-language model (VLM) with a powerful generalization capacity for various downstream tasks [2, 33]. However, training these models from scratch requires a huge dataset with labeled images, making it difficult to apply them directly to downstream tasks with fewer images. To address this issue, three common techniques have been pro + +![](images/c2f9ba5a4198bc1eb3d356e143542573c8f5cb36b0644ff32ab0104935c3021b.jpg) +Figure 1. Comparison with existing frameworks. (a) Domain-shared prompt tuning applies the same learnable prompt between the training and testing domains. (b) Image-conditional prompt tuning combines the image embedding with the learnable prompt; (c) Class-aware Prompt tuning injects the class-level textual embedding into the Text Encoder with the class-aware prompt. + +![](images/89bb576f820f247f6119691c614724f1b01f1f263a433bdd0e86aa74c33e7378.jpg) + +posed: fine-tuning [30], prompt tuning [50], adapter [12] and LoRA [15]. Among them, prompt tuning is a simple and efficient framework that transfers the essential general knowledge of VLM to the downstream tasks. + +Prompt tuning $^2$ is a technique that combines learnable textual tokens with class tokens to generate a discriminative textual classifier, known as Context Optimization (CoOp) [50]. Recently, various CoOp-based methods [4, 5, 17, 18, 22, 31, 40, 43, 50, 52] infer the domain-shared prompt tokens between training and testing domains (Figure 1(a)). Nevertheless, as the domain-shared prompt tokens are derived from labeled training images, their performance is suboptimal when confronted with unseen test classes. To enhance the generalization capacity of learnable prompt tokens, image-conditional prompts have been proposed in [45, 49] by fusing image features and learnable tokens (Figure 1(b)). Notably, image-conditional + +textual tokens encapsulate specific knowledge of each image, particularly for testing images, thereby rendering it more apt for generalization to unseen testing images. However, image-conditional prompt tokens with image-specific knowledge have less ability in elevating the distribution of class embeddings. In summary, classifiers generated by domain-shared and image-conditional textual tokens exhibit suboptimal performance for the unseen classes, primarily due to their inability to explicitly model the class distribution. Consequently, it is imperative to establish a dynamic relationship between the learnable prompt and the textual knowledge of each class to augment its discriminative prowess. + +The frozen CLIP, coupled with a hand-crafted prompt, exhibits robust generalization capabilities to novel classes, rendering it a valuable source of prior textual knowledge for each class. By associating the class-level textual knowledge with the learnable prompts, a class-aware prompt can be formulated for hencing the discriminative capacity of the textual classifier. To achieve this, we employ an embedding module to project the class-aware textual knowledge into the class-aware prompt tokens, as shown in Figure 1(c). The resultant class-aware prompt incorporates prior textual knowledge specific to each class, endowing the generated textual classifier with heightened discriminative prowess. Furthermore, the class-aware prompt facilitates the generation of classifiers for both seen and unseen classes by leveraging textual knowledge from both categories. To sum up, the trained embedding module can generate a class-aware prompt for each class based on its description ('class name'), thereby enhancing generalization and discriminative capabilities in class-level textual embeddings. + +Therefore, we propose a novel Textual-based Class-aware Prompt tuning (TCP) based on the framework of CoOp, shown in Figure 2. In addition to the domain-share textual tokens introduced in CoOp, TCP contributes a novel Textual Knowledge Embedding (TKE) to map the class-level textual knowledge into class-aware prompt tokens. Moreover, a class-aware textual classifier is generated by inserting the class-aware prompt tokens into the middle layer of the Text Encoder. We use standard contrastive loss and knowledge-guided consistency [43] to optimize the TKE and the learnable prompt tokens. During inference, TCP generates a class-aware classifier for unseen classes by feeding the domain-shared prompt tokens and the class-aware prompt tokens generated by the TKE into the frozen Text Encoder. + +Overall, the proposed TCP explicitly steers prompts to learn a class-aware knowledge that maximizes the generalization and discriminative of the downstream tasks. Evaluation over 11 image classification datasets on base-to-new generalization, cross-dataset generalization, and few-shot learning verify shows that TCP is an efficient method that + +obtain a higher performance with less training time. In summary, the proposed Textual-based Class-aware Prompt tuning(TCP) has the following main contributions: + +1. An effectively Textual-based Class-aware Prompt tuning is proposed by injecting the textual class-aware prompts generated by Textual Knowledge Embedding(TKE) into the Text Encoder. +2. We demonstrate that explicitly incorporating the prior knowledge of each class into the learnable prompt tokens can enhance the discriminative of the class distribution. +3. Textual Knowledge Embedding(TKE) is a plug-and-play module that can quickly insert existing prompt tuning methods to improve their performance further. + +# 2. Related Works + +# 2.1. Vision-Language Models + +Recently, researchers have shown that the Visual-Language Models(VLM) [2, 33] consisting of visual and textual modalities trained on the large-scale of image-text pairs has a powerful generalization and discriminativeion. To further boost the descriptive ability of VLM, the VLM models are boosted from the following aspect: 1) using a stronger text encoder or visual encoder [25, 41, 46]; 2) deeply fusing the visual and text knowledges [23, 38]; 3) using more images [16, 33, 35, 36]. To boost the diversity of text description, Masked Language Modeling (MLM) [20] [26] randomly erases the words in the text description used for representation learning. Unlike MLM, Masked autoencoder-based methods [13] are proposed to boost the descriptive ability by randomly masking the image patches. Among existing VLM models, CLIP is a representative and straightforward framework for inferring the independent visual and text encoder using the contrastive loss based on 400 million image-text association pairs. As CLIP has a good generalization, most existing CoOp-based methods are proposed based on CLIP for adapting the pre-trained VLM into the downstream task. Similar to existing methods, we conduct the prompt-tuning strategy on the TextEncoder of CLIP to obtain a task-specific textual embedding for prediction. + +# 2.2. Prompt Tuning + +To adapt the pretrained VLM to the downstream tasks, the prompt tuning [10, 22, 24, 31, 33, 45] always applies task-related textual tokens to infer the task-specific textual knowledge. In CLIP[33], the hand-crafted template "a photo of a [CLASS]" is used to embed the textual embedding for zero-shot prediction. However, the hand-crafted prompts have a poor ability to describe the downstream task. Textual prompt tuning is applied to boost the textual embedding by inferring a set of learnable textual tokens combined with the class tokens. For example, Context Optimization(CoOp) [50] replaces the hand-crafted prompts + +with the learnable soft prompts. To improve the generalization of the learnable textual prompt in CoOp, Conditional Context Optimization(CoCoOp) [49] and VPT [45] generates an image-conditional prompt fusing the image feature and the learnable textual prompt. Moreover, Knowledge-Guided Context Optimization(KgCoOp) [43], ProGrad [51], and Prompt Regularization(ProReg) [52] constrain the proposed learnable prompts contain the essential general knowledge. Unlike the above methods, which consider textual prompts, Ensembling Context Optimization(ECO) [1] employs prompt ensembling to combine multiple prompts. To obtain high-quality task-related tokens, ProDA [27] considers the prompt's prior distribution learning, and Distribution-Aware Prompt Tuning (DAPT) [5] optimizes the learnable prompt by maximizing inter-dispersion. Besides the textual knowledge from "classname", Knowledge-Aware Prompt Tuning(KAPT) [17] employs the external knowledge to generate the discriminative knowledge-aware prompt for the novel categories. PLOT [4] applies optimal transport to match the vision and text modalities for generating the discriminative and visual-aligned local textual prompt tokens. Besides the textual prompt tuning, Multi-modal Prompt Learning (MaPLe) [18] and PromptSRC [19] conduct the visual-textual prompt tuning by jointly conducting the prompt tuning on the visual and text encoders. Multitask Vision-Language Prompt Tuning(MVLPT) [37]incorporates crosstalk knowledge into prompt tuning for vision-language models. DenseCLIP [34] uses the context-aware prompt strategy to generate dense prediction tasks, and CLIP-Adapter [11] applies an adapter to adjust the visual or text embeddings. + +Existing methods commonly infer two types of prompt tokens: domain-share and image-conditional. However, the textual classifiers generated with these tokens tend to perform poorly on unseen classes. To mitigate this limitation, we propose a novel Textual-based Class-aware Prompt tuning (TCP) that employs a dynamic class-aware token to enhance the generalization and discriminative capabilities of the learnable textual prompt. Moreover, we introduce a Textual Knowledge Embedding to project class-level textual knowledge into the class-aware prompts, which are then inserted into the Text Encoder to generate a discriminative class-aware classifier. The evaluation results demonstrate that integrating class-level prior knowledge into the prompt tokens significantly enhances the discriminative ability of the prompt tuning process. + +# 3. Methodology + +As the Textual-based Class-aware Prompt tuning(TCP) is proposed based on Context Optimization (CoOp), we first briefly review CoOp and then introduce the proposed TCP. + +# 3.1. Preliminaries + +Existing CoOp-based methods are proposed based on the powerful Contrastive Language-Image Pre-training(CLIP). Given an image along with its corresponding textual description, CLIP uses visual and text encoders to extract the visual and text embeddings. After that, the constrastive loss between the visual and textual embeddings is calculated to align those two embeddings. To effectively adapt CLIP for the downstream task, CLIP applies the hand-crafted template "a photo of a $\{\}$ " to extract the general class-level textual embedding, defined as $\mathbf{W}^{clip} = \{\mathbf{w}_i^{clip}\}_{i=1}^{N_c}$ , where $\mathbf{w}_i^{clip}$ is the textual embedding of $i$ -th class, and $N_c$ is the number of classes. Given the 'class-name' of $i$ -th class, Word Embedded $e(\cdot)$ firstly embeds the hand-crafted description into a vectorized textual tokens: $\mathbf{t}_i^{clip} = e("a photo of a \{class-name\}")$ . After that, Text Encoder $\theta$ maps the vectorized textual tokens $\mathbf{t}_i^{clip}$ into the class-level embedding: $\mathbf{w}_i^{clip} = \theta (\mathbf{t}_i^{clip})$ . + +To improve the discriminative of the class-level embedding, the prompt tuning methods of Context Optimization(CoOp) replace the hand-crafted textual tokens with a set of learnable textual tokens $\mathbb{T} = \{\mathbf{t}_1,\mathbf{t}_2,\dots,\mathbf{t}_M\}$ , where $M$ is the length of tokens. Similar to CLIP, the corresponding class token $\mathbf{c}_i$ is concatenated with the learnable tokens $\mathbb{T}$ for generating the textual token $\mathbf{t}_i^{coop} = \{\mathbf{t}_1,\mathbf{t}_2,\dots,\mathbf{t}_M,\mathbf{c}_i\}$ . Next, the textual embedding $\mathbf{w}_i^{coop}$ is obtained by fedding the textual tokens $\mathbf{t}_i^{coop}$ into Text Encoder $\theta$ , i.e., $\mathbf{w}_i^{coop} = \theta (\mathbf{t}_i^{coop})$ . Finally, the textual embeddings of all classes are defined as $\mathbf{W}^{coop} = \{\mathbf{w}_i^{coop}\}_{i = 1}^{N_c}$ . CoOp infers the learnable textual tokens $\mathbb{T}$ by minimizing the contrastive loss between the image's embedding $\mathbf{x}$ and its class embedding $\mathbf{w}_y^{coop}$ : + +$$ +\mathcal {L} _ {c e} = \frac {1}{N} \sum_ {(\mathbf {x}, y) \in \mathcal {D} _ {s}} \frac {\exp (d (\mathbf {x} , \mathbf {w} _ {y} ^ {c o o p}) / \tau)}{\sum_ {i = 1} ^ {N _ {c}} \exp (d (\mathbf {x} , \mathbf {w} _ {i} ^ {c o o p}) / \tau)}, \tag {1} +$$ + +where $\mathcal{D}_s$ is the seen dataset, and $d(\cdot)$ is the cosine distance. $\tau$ is a temperature factor defined in CLIP, and $N$ is the number of training images. + +As the generated textual embedding has a good generalization ability for the novel classes, $\mathrm{KgCoOp}$ further adds an efficient consistency $\mathcal{L}_{kg}$ between the generated embedding $\mathbf{W}^{coop}$ and the general embedding $\mathbf{W}^{clip}$ , + +$$ +\mathcal {L} _ {k g} = | | \mathbf {W} ^ {c l i p} - \mathbf {W} ^ {c o o p} | | _ {2} ^ {2}. \tag {2} +$$ + +Therefore, a robust objective for prompt tuning is: + +$$ +\mathcal {L} = \mathcal {L} _ {c e} + \omega \mathcal {L} _ {k g}, \tag {3} +$$ + +where $\omega$ is set as 8.0 same as in KgCoOp [43]. + +![](images/1b2afd8b794f699c29a1ea474d25a27953c3d54ffced838f82246af0eaaad4f2.jpg) +Figure 2. The framework of the proposed TCP. + +# 3.2. Textual-based Class-aware Prompt tuning + +Based on the pre-trained Text Encoder in CLIP, the textual prompt tuning aims to infer a set of the domain-shared or image-conditional textual tokens combined with the general class tokens for generating the specific class embeddings. However, the textual classifier generated by those textual tokens performs a worse generalization on the unseen classes because they can not model the distribution of the testing classes. Research has shown that utilizing the general textual knowledge extracted by frozen CLIP can create a discriminative prior knowledge of novel classes, enhancing the discriminative and generalization of the learnable prompt. Drawing on the benefit of the general textual knowledge of seen and unseen classes, we propose a Textual-based Class-aware Prompt tuning(TCP) to adapt the pretrained CLIP to the downstream task. As shown in Figure 2, TCP employs a Textual Knowledge Embedding(TKE) to transfer the general class-level textual embedding into the class-aware prompt, which is then combined with the learnable textual tokens for generating the class-aware classifiers. TKE is advantageous for unseen classes, as it generates class-specific prompts to obtain an unseen class-aware textual classifier with better discriminative abilities. Moreover, explicitly incorporating seen class-aware prompts can enhance the discriminative power of seen classes. + +Given the general class-level textual embedding $\mathbf{W}^{clip} \in \mathbb{R}^{N_c \times D_t}$ with $N_c$ training classes, Textual Knowledge Embedding(TKE) $\mathcal{T}(\cdot)$ is proposed to project the class-level embedding $\mathbf{W}^{clip}$ into the class-aware prompt $\mathbf{T} = \mathcal{T}(\mathbf{W}^{clip})$ . As shown in Figure 2, TKE consists of two layers: down-project layer and up-project layer. The down-project layer uses the weight $\mathbf{W}_{down} \in \mathbb{R}^{D_t \times D_{mid}}$ to project the textual embedding into the low-dimension feature with the dimension of $D_{mid}$ . Next, the weight $\mathbf{W}_{up} \in \mathbb{R}^{D_{mid} \times D'}$ of the up-project layer maps the low-dimension feature into the high-dimension feature with the dimension of $D'$ . Note that $D'$ is determined by the prompt's length $M$ and dimension $D$ : $D' = M \times D$ . In + +summary, the general textual embedding $\mathbf{W}^{clip} \in \mathbb{R}^{N_c \times D_t}$ can be projected into the class-aware textual tokens $\mathbf{T} \in \mathbb{R}^{N_c \times D'}$ , which is further reshaped into the shape of $\mathbf{T} \in \mathbb{R}^{N_c \times M \times D}$ for inserting into the middle layer of Text Encoder $\theta$ . + +Assuming we insert the class-aware prompt $\mathbf{T}$ into $l$ -th layer of Text Encoder $\theta$ . We will give a detailed analysis of the hyperparameter $l$ in the following. Similar to CoOp, by combining the domain-shared learnable textual tokens $\mathbb{T} = [\mathbf{t}_1,\mathbf{t}_2,\dots,\mathbf{t}_M]$ and the pre-trained class tokens $\mathbf{C}$ of all classes, we can obtain the input textual tokens $\mathbf{F}_0 = \{\mathbb{T},\mathbf{C}\}$ of the Text Encoder, where $\mathbf{C} = \{\mathbf{c}_i\}_{i}^{N_c}$ , and $\mathbf{c}_i$ is the vectorized textual tokens of $i$ -th class. The textual token $\mathbf{F}_0$ is fed into the first $l$ layers of Text Encoder for obtaining the middle-level textual embedding $\mathbf{F}_l$ . Formally, the textual token $\mathbf{F}_i(i\leq l)$ of $i$ -th layer is defined as: + +$$ +\mathbf {F} _ {i} = \theta_ {i} \left(\mathbf {F} _ {i - 1}\right), i \in [ 1, l ], \tag {4} +$$ + +where $\theta_{i}$ is the $i$ -th layer of Text Encoder. + +For the textual tokens $\mathbf{F}_l\in \mathbb{R}^{N_c\times N_t\times D}$ and the class-aware prompt tokens $\mathbf{T}\in \mathbb{R}^{N_c\times M\times D}$ , the first dimension is related to the number of classes. Therefore, the learnable prompt tokens are always inserted into the second dimension of $\mathbf{F}_l$ , the same as CoOp. Formally, the class-aware prompt $\mathbf{T}$ is inserted into $\mathbf{F}_l$ to generate the class-aware enhanced tokens $\mathbf{F}_l'$ . + +$$ +\mathbf {F} _ {l} ^ {\prime} = \left[ \mathbf {T} _ {1}, \mathbf {T} _ {2}, \dots , \mathbf {T} _ {M}, \mathbf {F} _ {l, M + 1}, \mathbf {F} _ {l, M + 2}, \dots , \mathbf {F} _ {l, N _ {t}} \right], \tag {5} +$$ + +where $\mathbf{T}_i$ denotes the $i$ -th index of $\mathbf{T}$ in the second dimension, and $\mathbf{F}_{l,j}$ denote the correspond $j$ -th index of $\mathbf{F}_l$ in the second dimension, i.e., $\mathbf{T}_i = \mathbf{T}[:i,:]$ , and $\mathbf{F}_{l,j} = \mathbf{F}_l[:j,:]$ . + +After that, the class-enhanced textual tokens $\mathbf{F}_l^{\prime}$ are fedded into the rest layers for generating the class-aware textual embedding, + +$$ +\mathbf {F} _ {i} ^ {\prime} = \theta_ {i} (\mathbf {F} _ {i - 1} ^ {\prime}), i \in [ l + 1, L ]. \qquad (6) +$$ + +The output of the last layer $\mathbf{F}_L^{\prime}$ is treated as the class embedding $\mathbf{W}^{tcp}$ used for optimization with contrastive loss and knowledge-guided consistency loss in Eq. (3). + +# 4. Experiments + +Similar to CoOp [50], we evaluate the effectiveness of TCP from the three types of tasks: 1) generalization from baseto-new classes within a dataset; 2) few-shot learning with $K$ -shot labeled images; 3) cross-dataset generalization from the imagenet to other datasets. More detailed results will be provided in the Supplementary materials. + +# 4.1. Experimental Setup + +Dataset. We conduct the base-to-new generalization, few-shot learning, and cross-dataset generalization on 11 benchmarks, i.e., ImageNet [7], Caltech [9], OxfordPets [32], + +Table 1. Comparison on the base-to-new generalization setting with 16-shot samples. 'tp', 'dtp', 'vp', and 'dvp' denote the 'textual prompt', 'deep textual prompt', 'visual prompt', and 'deep visual prompt', respectively. PromptSRC are based on deep visual-textual prompt tuning ('dvp+dtp'). *' denote the performance obtained by our re-implementation. + +
DatasetsSetsCoOp* (JCV22)CoCoOp (CVPR22)DAPT* (ICCV23)ProGrad* (ICCV23)ProDA (CVPR22)KgCoOp (CVPR23)RPO (ICCV23)PLOT* (ICLR23)LFA (ICCV23)MaPLe (CVPR23)DePT (Arxiv23)PromptSRC* (ICCV23)TCP
tptptp+vptptptpdtp+dvptp+vp-dtp+dvptpdtp+dvptp
AverageBase82.3880.4783.1882.4881.5680.7381.1383.9883.6282.2883.6284.1284.13
New67.9671.6969.2770.7572.3073.675.0071.7274.5675.1475.0475.0275.36
H74.4875.8375.5976.1676.6577.077.7877.3778.8378.5579.1079.3179.51
ImageNetBase76.4675.9876.8377.0275.4075.8376.6077.3076.8976.6677.0377.7577.27
New66.3170.4369.2766.6670.2369.9671.5769.8769.3670.5470.1370.7069.87
H71.0273.1072.8571.4672.7272.7874.0073.4072.9373.4773.4274.0673.38
Caltech101Base97.8097.9697.8398.0298.2797.7297.9798.5398.4197.7498.3098.1398.23
New93.2793.8193.0793.8993.2394.3994.3792.8093.9394.3694.6093.9094.67
H95.4895.8495.3995.9195.6896.0396.0395.5896.1396.0296.4195.9796.42
OxfordPetsBase94.4795.2095.0095.0795.4394.6594.6394.5095.1395.4394.3395.5094.67
New96.0097.6995.8397.6397.8397.7697.5096.8396.2397.7697.2397.4097.20
H95.2396.4395.4196.3396.6296.1896.0595.6595.6896.5895.7696.4495.92
CarsBase75.6770.4975.8077.6874.7071.7673.8779.0776.3272.9479.1378.4080.80
New67.5373.5963.9368.6371.2075.0475.5374.8074.8874.0075.4774.7374.13
H71.3772.0169.3672.8872.9173.3674.6976.8875.5973.4777.2675.5277.32
FlowersBase97.2794.8796.9795.5497.7095.0094.1397.9397.3495.9298.0097.9097.73
New67.1371.7560.9071.8768.6874.7376.6773.5375.4472.4676.3776.7775.57
H79.4481.7174.8182.0380.6683.6584.5083.9985.0082.5685.8486.0685.23
Food101Base89.3790.7090.3790.3790.3090.590.3389.8090.5290.7190.5090.6390.57
New88.7791.2991.3089.5988.5791.790.8391.3791.4892.0591.6091.5091.37
H89.0790.9990.8389.9889.4391.0990.5890.5891.0091.3891.0591.0690.97
AircraftBase39.6733.4139.9740.5436.9036.2137.3342.1341.4837.4443.2042.3041.97
New31.2323.7129.8027.5734.1333.5534.2033.7332.2935.6134.8336.9734.43
H34.9527.7434.1432.8235.4634.8335.7037.4636.3136.5038.5739.4637.83
SUN397Base80.8579.7480.9781.2678.6780.2980.6082.2082.1380.8282.3382.8382.63
New68.3476.8676.9774.1776.9376.5377.8073.6377.2078.7077.8079.0078.20
H74.0778.2778.9277.5577.7978.3679.1877.6879.5979.7580.0080.8780.35
DTDBase79.9777.0182.2377.3580.6777.5576.7081.9781.2980.3682.2082.6082.77
New48.6056.0054.2352.3556.4854.9962.1343.8060.6359.1859.1357.5058.07
H60.4664.8565.3662.4566.4464.3568.6157.0969.4668.1668.7867.8068.25
EuroSATBase90.1087.4994.7390.1183.9085.6486.6393.7093.4094.0789.0392.4091.63
New53.0060.0450.3360.8966.0064.3468.9762.6771.2473.2371.0768.4374.73
H66.7471.2165.7472.6773.8873.4876.7975.1180.8382.379.0478.6382.32
UCF101Base84.5382.3384.3084.3385.2382.8983.6786.6086.9783.0085.8086.9387.13
New67.3773.4576.3374.9471.9776.6775.4375.9077.4878.6677.2378.3380.77
H74.9877.6780.1279.3578.0479.6579.3480.9081.9580.7781.2982.4183.83
+ +StanfordCars [21], Flowers [29], Food101 [3], FGVCAircraft [28], EuroSAT [14], UCF101 [39], DTD [6], and SUN397 [42]. A detailed introduction of those dataset is given in the Supplementary materials. + +Training Details. Our implementation is based on CoOp's code [50]. All experiments are conducted based on the CLIP with the backbone of ViT-B/16 [8]. The prompt's length $M$ is set as 4 with random initialization of "X X X X {}". The final performance is averaged over three random seeds(1/2/3)for a fair comparison. Adam optimizer is applied for optimization with the learning rate of 2e-3 and the batch size of 32, and the training epochs is 50. All experiments are conducted on RTX 3090. + +Baselines. Recently CoOp-based methods are used for comparison, e.g., CoOp [50], CoCoOp [49], ProGrad [51], ProDA [27], KgCoOp [43], PromptSRC [19], MaPLe [18], LFA [31], DePT [47], DAPT [5], PLOT [4], TaskRes [44], RPO [22], VPT [45], and TIP-Adapter-F [48]. + +# 4.2. Base-to-New Class Generalization + +To verify the generalization ability of the prompt tuning, the base-to-new class generalization splits each dataset into two + +disjoint subsets: Base and New classes, where Base classes are used to infer the learnable prompt and TKE, and New classes are used for evaluation. Note that each New class contains 16-shot samples. We summarize the comparison between the proposed TCP and existing methods in Table 1 from the following two aspects: Textual Prompt Tuning and Vivasal-Textual Prompt Tuning. + +Textual Prompt Tuning: Based on the visual feature extracted from the frozen visual encoder, the textual prompt tuning (e.g., CoOp, CoCoOp, ProGrad, ProDA, and Kg-CoOp) merely conducts the textual prompt tuning by inserting a set of learnable textual prompts into the Text Encoder for obtaining a discriminative textual classifier, which is same as our proposed TCP. Among all textual prompt tuning methods, KgCoOp [43] is a more substantial baseline, which is also the baseline of our TCP. Compared to KgCoOp, TCP improves the metric term of Base/New/H from $80.73\% / 73.6\% / 77.0\%$ to $84.35\% / 75.28\% / 79.56\%$ . We can observe that a noticeable improvement of $3.62\%$ and $1.68\%$ are obtained in the Base class and New class, respectively. The superior performance proves that TCP using the class-aware prompt to inject the class knowledge can generate + +Table 2. Comparison of cross-dataset evaluation. 'tp', 'dtp', 'vp', and 'dvp' denote the 'textual prompt', 'deep textual prompt', 'visual prompt', and 'deep visual prompt', respectively. Note that DAPT and MaPLe are based on visual-textual prompt tuning('vp+tp'). + +
DatasetsCLIPCoOpProGradKgCoOpDePTVPTPLOTPromptSRCMaPLeDAPTTCP
tptptptptp+vptp+vpdtp+tvpdtp+dvptp+vptp
ImageNet66.7071.5172.2470.6672.7769.7371.6071.2770.7271.6071.40
Caltech10193.3093.7091.5293.9294.2393.6792.0793.6093.5393.5093.97
OxfordPets89.1089.1489.6489.8390.0389.2790.1090.2590.4990.6791.25
StandfordCars65.7064.5162.3965.4165.5765.565.7065.7065.5765.9364.69
Flowers70.7068.7167.8770.0170.5770.269.2370.2572.2071.7071.21
Food10185.9085.3085.4086.3686.3786.2786.2386.1586.2086.1086.69
FGVCAircraft24.9018.4720.1622.5123.2722.1325.0023.9024.7423.0323.45
SUN39762.6064.1562.4766.1666.6766.5761.6767.1067.0167.0067.15
DTD44.3041.9239.4246.3545.9746.9338.6046.8746.4944/0044.35
EuroSAT48.3046.3943.4646.0443.5347.4347.8345.5048.0652.4751.45
UCF10167.6066.5564.2968.5069.3067.2067.0068.7568.6968.7368.73
Avg.65.2463.8862.7165.5165.5565.5264.3465.8166.3066.3166.29
+ +the discriminative textual classifier for both seen and unseen classes. DePT [47] is another work that also treats KgCoOp as the baseline and obtains the best $H$ of $79.10\%$ among all textual prompt tuning methods. Compared to DePT, the proposed TCP obtains the performance of $84.35\% / 75.28\% / 79.56\%$ in the metric term of Base/New/H, achieving the improvement of $0.51\% / 0.32\% / 0.41\%$ . + +Visual-Textual Prompt Tuning: Different from the textual prompt tuning merely adapts Text Encoder, DAPT [5], PLOT [4], RPO [22], MaPLe [18], and PromptSRC [19] all jointly performs prompt tuning on the text and visual encoders for both improving the discriminative of textual embedding and visual embeddings. Unlike CoOp and TCP, which conduct the prompt tuning on one or two layers, those methods conduct the visual-textual prompt tunings on several layers of visual and text encoders. Compared to them, the proposed TCP with shallow textual prompt tuning obtains a higher performance, e.g., $79.51\%$ vs $79.31\%$ for PromptSRC in terms of $H$ . Compared to deep visual-textual prompt tuning, the superior performance demonstrates that class-aware prompt tuning can generate discriminative and generalization textual-level classifiers. + +# 4.3. Cross-Dataset Generalization + +In the Base-to-New generalization setting, New classes always have a similar data distribution as the Base classes. To further verify the generalization of the proposed TCP, in the Cross-Dataset generalization, TCP is trained from the ImageNet and directly evaluated on the unrelated datasets, e.g., the rest ten datasets. The comparison between the proposed TCP and existing methods is summarized in Table 2. From Table 2, we can observe that the proposed TCP obtains the highest average performance among all textual prompt tuning methods (66.29% vs 65.55% of DePT [47]), and obtains a comparable performance with the visual-textual prompt tuning methods (66.29% vs 66.31% of DAPT [5]), demonstrating the effectiveness of TCP in learning the generalization knowledge. + +Table 3. Effect of TCP on CoOp, KgCoOp, ProGrad, PromptSRC, and DAPT in the base-to-new generalization setting regarding final average performance. + +
MethodsBaseNewH
CoOp [50]82.6367.9974.6
CoOp+TKE83.10(↑ 0.47)70.17(↑ 2.18)76.09 (↑ 1.49)
KgCoOp [43]80.7373.677.00
KgCoOp +TKE84.13 (↑ 3.40)75.36(↑ 1.76)79.51 (↑ 2.51)
ProGrad [51]82.4870.7576.16
ProGrad+TKE82.61 (↑ 0.13)72.91(↑ 2.16)77.46 (↑ 1.30)
PromptSRC [19]82.0774.8378.03
PromptSRC+TKE83.74 (↑ 1.67)75.85(↑ 1.02)79.60 (↑ 1.57)
DAPT [5]83.1869.2775.59
DAPT+TKE84.15 (↑ 0.97)74.87(↑ 5.60)79.24 (↑ 3.65)
+ +# 4.4. Few-shot Classification + +To verify that the proposed TCP can infer the class-aware knowledge, the few-shot classification is conducted on all 11 datasets using K-shot labeled source images and evaluated on the standard testing domain with the same class space as the training classes. The 4-shot setting comparison between the proposed TCP and existing methods is summarized in Table 4. We can observe that the proposed TCP achieves a higher average performance than existing methods, i.e., obtaining the average performance of $76.32\%$ . + +# 4.5. Ablation Study + +Effect of TKE's generalization. The critical component of TCP is the class-aware prompt generated by TKE, which embeds the prior class knowledge and can be easily integrated with existing CoOp-based methods. As shown in Table 3, CoOp, KgCoOp, ProGrad, PromptSRC, and DAPT have all demonstrated improved performance when TKE is used. Specifically, improvements of $1.49\%$ , $2.51\%$ , $1.30\%$ , $1.57\%$ , and $3.65\%$ were achieved for CoOp, KgCoOp, ProGrad, PromptSRC3, and DAPT, respectively. This superior performance highlights TCP's plug-and-play functionality, + +Table 4. Comparison of few-shot learning with 4-shot samples. + +
CLIPCoOpCoCoOpProGradKgCoOpMaPLeTIP-Adapter-FDAPTPromptSRCPLOTTaskResTCP
ImageNet66.7069.3770.5570.2170.1970.6770.7870.8070.8070.4062.8770.48
Caltech10193.3094.4494.9894.9394.6594.3094.7794.2394.7795.1394.6795.00
OxfordPets89.1091.3093.0193.2193.2092.0592.2692.1793.2392.5592.0091.90
StandfordCars65.7072.7369.1071.7571.9868.7074.4274.4071.8374.9375.9076.30
Flowers70.7091.1482.5689.9890.6980.8092.9892.3791.3192.9391.5094.40
Food10185.9082.5886.6485.7786.5986.9086.1883.6086.0686.4686.0385.3
FGVCAircraft24.9033.1830.8732.9332.4729.0335.4932.4732.8035.2933.8036.20
SUN39762.6070.1370.5071.1771.7971.4770.6572.2072.8070.4272.7072.11
DTD44.3058.5754.7957.7258.3154.7361.7061.3760.6462.4359.5763.97
EuroSAT48.3068.6263.8370.8471.0654.8778.2772.7375.0280.7072.8777.43
UCF10167.6077.4174.9977.8278.4073.7079.7379.4079.3579.7676.1080.83
Avg.65.3773.5971.9874.2174.4870.6676.1175.0775.3376.4574.3676.72
+ +Table 5. Domain-shared prompt vs Class-aware prompt. + +
BaselineDomain-SharedClass-AwareBaseNewH
✓✓80.7373.677
✓✓84.0575.1879.36
✓✓✓84.1375.3679.51
+ +Figure 3. Effect of $M$ on the base-to-new generalization setting. +![](images/a7fd9172650d9bdbfc8186a426a2601cc631977f3ed0728679d2df6e5b7f0205.jpg) +$D_{mid}$ in TKE. + +allowing easy integration with existing methods. + +Domain-shared prompt vs the class-aware prompt. It is common for existing methods to infer the domain-shared prompt tokens. However, TCP differs significantly from them in using a class-aware prompt to capture class-aware knowledge in the Text Encoder. We thus compare the domain-shared and class-aware prompts, and summarize the related results in Table 5. We can observe that simply using the class-aware prompt yields a fantastic performance of $79.36\%$ in terms of $H$ , higher than existing methods. Additionally, as shown in Table 5, using the class-aware prompt results in a noticeable improvement over the domain-shared prompt. For instance, the Basel/New performance is improved from $80.73\% / 73.6\%$ to $84.05\% / 75.18\%$ . The superior performance proves the reasonableness and effectiveness of injecting the class-aware knowledge into the prompt tuning. By further considering the domain-shared and class-aware prompts, the final performance is reached at $79.51\%$ . + +Effect of Prompt length $M$ . In this study, we examine the impact of prompt length $(M)$ on the Base-to-New generalization setting. We compared prompt lengths of 1, 2, 4, 8, and 16, and the results are presented in Figure 3. The + +Table 6. Comparison of different templates. + +
TemplatesBaseNewH
‘X X X X’84.1375.3679.51
‘a photo of a’83.9474.9479.18
‘this is a picture’84.0074.6679.06
+ +![](images/5b18c2a1da65e25595fdbf49500f754ac88e936722af28d90c4f5ff65b93a8f9.jpg) +Figure 4. Effect of dimension + +![](images/ad2c24d0f099c4e9ac56637ca53bc03d6b373c303ae46a847443b36ca6fd1f6a.jpg) +Figure 5. Effect of insert layer $l$ . + +findings indicate that longer prompt lengths produce better performance, with $M = 8$ producing the best results. However, the performance gap between the highest and lowest values of $H$ , which were $79.63\%$ and $79.26\%$ , respectively, is only $0.37\%$ . Therefore, the proposed TCP is relatively insensitive to prompt length $(M)$ . + +Effect of different templates. Unlike most CoOp-based methods using the handcrafted template "a photo of a $\{\}$ ", the proposed TCP utilizes the random initialization template ("X XXX $\{\}$ "). Therefore, we conducted a comparison among different templates. Table 6 shows that using the random template performs better than the handcrafted template. This is because TCP uses the class-ware prompt without considering the domain-share prompt, which has achieved better performance. For instance, in Table 5, TCP with the class-aware prompt achieves comparable performance of $84.05\% /75.18\% /79.36\%$ in terms of Base/New/H compared to existing methods. Moreover, the domain-shared prompt initialized as the random template provides complementary knowledge to the class-aware prompt. + +Effect of $D_{mid}$ in TAK. We analyze the effect of the dimension $D_{mid}$ in TKE, and show the comparison in Figure 4. Setting $D_{mid} = 128$ obtains the best performance, and a smaller $D_{mid}$ would get a worse performance. + +Insert which layer? TCP inserts the class-aware textual prompts into a particular layer of the Text Encoder. The im- + +Table 7. Comparison with single-layer vs multiple layers prompt tuning. + +
ModesLayersBaseNewH
TCP-Shallow{8}84.1375.3679.51
TCP-Deeper{4:8}84.0275.1379.33
TCP-Deeper{8:10}84.2475.2979.51
TCP-Deeper{4:8;10}84.1174.5179.02
TCP-Deeper{4;6;8;10}84.1774.6679.13
+ +![](images/5b606b15fc1b82e7e6bfe74f92517ac8a56f790938ab53d20e244416c7b4698a.jpg) +Figure 6. Comparison with TCP and Adapter-based methods. + +pact of inserting prompts into different layers was analyzed and shown in Figure 5. The best performance was achieved when the class-aware prompt tuning was done on the 8th layer. This is because doing so on earlier layers ( $l < 8$ ) does not transfer class-aware knowledge to the textual embedding, and doing it on later layers can easily affect the textual embedding. + +Single-layer vs Multiple-layers? The results of existing MaPLe and PromptSRC show that deep prompt learning applied on multiple layers can perform better than shallow prompt tuning with a single layer. Therefore, we compared the performance of shallow prompt tuning with deep prompt tuning. The results in Table 7 indicate that the shallow class-aware prompt tuning with a single layer provides better performance. For instance, TCP-Shallow with 8-layer achieves the highest $H$ of $79.51\%$ , which surpasses all variants of TCP-deeper that have two, three, and four layers. Additionally, we observed that TCP-shallow obtains the best performance of $75.36\%$ for the New class, and using more layers results in lower performance. + +Compared to Adapter-based methods: In TCP, TKE acts as a special adapter, leading to the proposed TCP is similar to Adapter-based methods. We compared TCP with several Adapter-based methods such as CLIP-Adapter and CoOp-Adapter, which apply the adapter on the general textual-level class embeddings. However, TCP applies TKE to transfer the textual-level class embedding into the class-aware prompt combined with the class tokens in the middle layer of the Text Encoder. As shown in Figure 6, TCP with TKE performs better than the existing CLIP-Adapter and CoOp-Adapter on all 11 datasets. + +Training time analysis: We further analyze the training time of the proposed TCP. The training time was determined based on the average time to process one image from the ImageNet with a 16-shot setting. According to Table 8, the proposed TCP is an efficient method that performs better + +Table 8. Training time comparison(ms/image). + +
CoOpProGradKgCoOpPLOTRPOMaPLePromptSRCTCP
ms/image6ms22ms6ms78.8ms190ms90.7ms43.2ms6.4ms
H74.4876.1677.0077.3777.7878.5579.3179.51
+ +![](images/ea6c2b7fbe926823433b42f97e36931156db764ad71f5036cf0742e84e679fec.jpg) + +![](images/293fc5e48e6789d6ae5faf265486c20247ac898afe8395cf8d31dbd6c65db2bb.jpg) + +![](images/c5a2d314927fe6616f08c6243ecd749ce33b972dd85061b13afa5ce1b5df6d08.jpg) +Figure 7. Visualization of the probability of CoOp and TCP. + +![](images/998afdab57dbe67dda51b7a2ffb65dc2df0565b7908f7a6191b7aedf10f38c35.jpg) + +with less running time. This is because TCP adds an efficient TKE component for generating class-aware prompt upon CoOp and KgCoOp, resulting in slightly more time consumption than CoOp and KgCoOp. + +Visualization. To prove that the proposed TCP can generate the discriminative textual classifier for prediction, we further visualize the prediction probability among all classes. As shown in Figure 7, the proposed TCP has an obvious inter-class classes than existing methods. + +# 5. Conclusion + +To improve the generalization and discriminative abilities of the learnable prompt, we introduce a Textual-based Class-aware Prompt tuning method that takes advantage of the benefits of general class-level textual knowledge. To achieve this, we propose a Textual Knowledge Embedding (TKE) that transfers the class-level textual embedding into the class-aware prompt. This is combined with pre-trained class tokens to generate task-specific textual knowledge. Several benchmarks and tasks have shown that the class-aware prompt is effective for prompt tuning. + +However, the class-aware prompt in TCP heavily relies on the discriminative ability of the general textual embedding. On the other hand, a weaker textual embedding will produce a weaker textual classifier. For example, TCP does not perform well on the FGVCAircraft dataset. Therefore, in the future, we plan to explore ways to use weaker textual knowledge to obtain a discriminative textual classifier. + +Acknowledgements: This work was supported by National Science and Technology Major Project(2021ZD0112202), National Natural Science Foundation of China (62376268,U23A20387,62036012,U21B2044,62376196), and Beijing Natural Science Foundation (4222039). + +# References + +[1] Lorenzo Agnolucci, Alberto Baldrati, Francesco Todino, Federico Becattini, Marco Bertini, and Alberto Del Bimbo. Eco: Ensembling context optimization for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2811-2815, 2023. 3 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 2 +[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI, pages 446-461. Springer, 2014. 5 +[4] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. PLOT: prompt learning with optimal transport for vision-language models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 1, 3, 5, 6 +[5] Eulrang Cho, Jooyeon Kim, and Hyunwoo J Kim. Distribution-aware prompt tuning for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22004-22013, 2023. 1, 3, 5, 6 +[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014, pages 3606-3613. IEEE Computer Society, 2014. 5 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248-255. IEEE Computer Society, 2009. 4 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 5 +[9] Li Fei-Fei, Robert Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. Comput. Vis. Image Underst., 106(1):59-70, 2007. 4 +[10] Zhe Gan, Linjie Li, Chunyuan Li, Lijuan Wang, Zicheng Liu, Jianfeng Gao, et al. Vision-language pre-training: Basics, recent advances, and future trends. Foundations and Trends® in Computer Graphics and Vision, 14(3-4):163-352, 2022. 2 + +[11] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. CoRR, abs/2110.04544, 2021. 3 +[12] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 1 +[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15979-15988. IEEE, 2022. 2 +[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens., 12(7):2217-2226, 2019. 5 +[15] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. Open-Review.net, 2022. 1 +[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 4904-4916. PMLR, 2021. 2 +[17] Baoshuo Kan, Teng Wang, Wenpeng Lu, Xiantong Zhen, Weili Guan, and Feng Zheng. Knowledge-aware prompt tuning for generalizable vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15670-15680, 2023. 1, 3 +[18] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 1, 3, 5, 6 +[19] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 3, 5, 6 +[20] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 5583-5594. PMLR, 2021. 2 +[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 2013 IEEE International Conference on Computer Vision + +Workshops, ICCV Workshops 2013, Sydney, Australia, December 1-8, 2013, pages 554-561. IEEE Computer Society, 2013. 5 +[22] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1401-1411, 2023. 1, 2, 5, 6 +[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2 +[24] Juncheng Li, Minghe Gao, Longhui Wei, Siliang Tang, Wenqiao Zhang, Mengze Li, Wei Ji, Qi Tian, Tat-Seng Chua, and Yueting Zhuang. Gradient-regulated meta-prompt learning for generalizable vision-language models. arXiv preprint arXiv:2303.06571, 2023. 2 +[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2 +[26] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 13-23, 2019. 2 +[27] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5196-5205. IEEE, 2022. 3, 5 +[28] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. CoRR, abs/1306.5151, 2013. 5 +[29] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Sixth Indian Conference on Computer Vision, Graphics & Image Processing, ICVGIP 2008, Bhubaneswar, India, 16-19 December 2008, pages 722-729. IEEE Computer Society, 2008. 5 +[30] Maxime Aquab, Léon Bottou, Ivan Laptev, and Josef Sivic. Learning and transferring mid-level image representations using convolutional neural networks. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014, pages 1717-1724. IEEE Computer Society, 2014. 1 +[31] Yassine Ouali, Adrian Bulat, Brais Matinez, and Georgios Tzimiropoulos. Black box few-shot adaptation for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15534-15546, 2023. 1, 2, 5 +[32] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI, + +USA, June 16-21, 2012, pages 3498-3505. IEEE Computer Society, 2012. 4 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 1, 2 +[34] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 18061-18070. IEEE, 2022. 3 +[35] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2 +[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2 +[37] Sheng Shen, Shijia Yang, Tianjun Zhang, Bohan Zhai, Joseph E. Gonzalez, Kurt Keutzer, and Trevor Darrell. Multitask vision-language prompt tuning. CoRR, abs/2211.11720, 2022. 3 +[38] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2 +[39] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR, abs/1212.0402, 2012. 5 +[40] Jingchen Sun, Jiayu Qin, Zihao Lin, and Changyou Chen. Prompt tuning based adapter for vision-language model adaption. CoRR, abs/2303.15234, 2023. 1 +[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008, 2017. 2 +[42] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In The Twenty-Third IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2010, San Francisco, CA, USA, 13-18 June 2010, pages 3485-3492. IEEE Computer Society, 2010. 5 +[43] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 1, 2, 3, 5, 6 +[44] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 10899-10909. IEEE, 2023. 5 +[45] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Unified vision and language prompt learning. CoRR, abs/2210.07225, 2022. 1, 2, 3, 5 +[46] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12104-12113, 2022. 2 +[47] Ji Zhang, Shihan Wu, Lianli Gao, Hengtao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. arXiv preprint arXiv:2309.07439, 2023. 5, 6 +[48] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of CLIP for few-shot classification. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 493-510. Springer, 2022. 5 +[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 16795-16804. IEEE, 2022. 1, 3, 5 +[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. Int. J. Comput. Vis., 130(9):2337-2348, 2022. 1, 2, 4, 5, 6 +[51] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 3, 5, 6 +[52] Beier Zhu, Yulei Niu, Saeil Lee, Minhoe Hur, and Hanwang Zhang. Debiased fine-tuning for vision-language models by prompt regularization. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence, IAAI 2023, Thirteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2023, Washington, DC, USA, February 7-14, 2023, pages 3834-3842. AAAI Press, 2023. 1, 3 \ No newline at end of file diff --git a/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/images.zip b/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fbef9af495ce11372810539ca3e92ab6b763d25a --- /dev/null +++ b/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d027c198e74cc75e3a576f2264fba3cf68669b670678a3e4c91a0798b564644b +size 672771 diff --git a/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/layout.json b/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cb2964c0672754f93fd6f6234a5301e8110a134d --- /dev/null +++ b/2024/TCP_Textual-based Class-aware Prompt tuning for Visual-Language Model/layout.json @@ -0,0 +1,9568 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 64, + 103, + 531, + 121 + ], + "type": "text", + "content": "TCP:Textual-based Class-aware Prompt tuning for Visual-Language Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 212, + 144, + 380, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 144, + 380, + 157 + ], + "spans": [ + { + "bbox": [ + 212, + 144, + 380, + 157 + ], + "type": "text", + "content": "Hantao Yao" + }, + { + "bbox": [ + 212, + 144, + 380, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 212, + 144, + 380, + 157 + ], + "type": "text", + "content": ", Rui Zhang" + }, + { + "bbox": [ + 212, + 144, + 380, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 212, + 144, + 380, + 157 + ], + "type": "text", + "content": ", Changsheng Xu" + }, + { + "bbox": [ + 212, + 144, + 380, + 157 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 157, + 537, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 157, + 537, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 157, + 537, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 157, + 537, + 171 + ], + "type": "text", + "content": " State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 171, + 488, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 171, + 488, + 184 + ], + "spans": [ + { + "bbox": [ + 104, + 171, + 488, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 171, + 488, + 184 + ], + "type": "text", + "content": " State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences, China;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 200, + 185, + 394, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 185, + 394, + 198 + ], + "spans": [ + { + "bbox": [ + 200, + 185, + 394, + 198 + ], + "type": "text", + "content": "3 University of Chinese Academy of Sciences, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 120, + 201, + 468, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 201, + 468, + 213 + ], + "spans": [ + { + "bbox": [ + 120, + 201, + 468, + 213 + ], + "type": "text", + "content": "hantao.yao@nlpr.ia.ac.cn; zhangrui@ict.ac.cn; csxu@nlpr.ia.ac.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 266, + 290, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 290, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 290, + 555 + ], + "type": "text", + "content": "Prompt tuning represents a valuable technique for adapting pre-trained visual-language models (VLM) to various downstream tasks. Recent advancements in CoOp-based methods propose a set of learnable domain-shared or image-conditional textual tokens to facilitate the generation of task-specific textual classifiers. However, those textual tokens have a limited generalization ability regarding unseen domains, as they cannot dynamically adjust to the distribution of testing classes. To tackle this issue, we present a novel Textual-based Class-aware Prompt tuning(TCP) that explicitly incorporates prior knowledge about classes to enhance their discriminability. The critical concept of TCP involves leveraging Textual Knowledge Embedding (TKE) to map the high generalizability of class-level textual knowledge into class-aware textual tokens. By seamlessly integrating these class-aware prompts into the Text Encoder, a dynamic class-aware classifier is generated to enhance discriminability for unseen domains. During inference, TKE dynamically generates class-aware prompts related to the unseen classes. Comprehensive evaluations demonstrate that TKE serves as a plug-and-play module effortlessly combinable with existing methods. Furthermore, TCP consistently achieves superior performance while demanding less training time1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 599, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 287, + 684 + ], + "type": "text", + "content": "Large-scale image-text pairs have the ability to train a Visual-language model (VLM) with a powerful generalization capacity for various downstream tasks [2, 33]. However, training these models from scratch requires a huge dataset with labeled images, making it difficult to apply them directly to downstream tasks with fewer images. To address this issue, three common techniques have been pro" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 320, + 239, + 429, + 373 + ], + "blocks": [ + { + "bbox": [ + 320, + 239, + 429, + 373 + ], + "lines": [ + { + "bbox": [ + 320, + 239, + 429, + 373 + ], + "spans": [ + { + "bbox": [ + 320, + 239, + 429, + 373 + ], + "type": "image", + "image_path": "c2f9ba5a4198bc1eb3d356e143542573c8f5cb36b0644ff32ab0104935c3021b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 380, + 547, + 448 + ], + "lines": [ + { + "bbox": [ + 304, + 380, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 547, + 448 + ], + "type": "text", + "content": "Figure 1. Comparison with existing frameworks. (a) Domain-shared prompt tuning applies the same learnable prompt between the training and testing domains. (b) Image-conditional prompt tuning combines the image embedding with the learnable prompt; (c) Class-aware Prompt tuning injects the class-level textual embedding into the Text Encoder with the class-aware prompt." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 430, + 250, + 533, + 372 + ], + "blocks": [ + { + "bbox": [ + 430, + 250, + 533, + 372 + ], + "lines": [ + { + "bbox": [ + 430, + 250, + 533, + 372 + ], + "spans": [ + { + "bbox": [ + 430, + 250, + 533, + 372 + ], + "type": "image", + "image_path": "89bb576f820f247f6119691c614724f1b01f1f263a433bdd0e86aa74c33e7378.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 456, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 503 + ], + "type": "text", + "content": "posed: fine-tuning [30], prompt tuning [50], adapter [12] and LoRA [15]. Among them, prompt tuning is a simple and efficient framework that transfers the essential general knowledge of VLM to the downstream tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 506, + 546, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 546, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 546, + 662 + ], + "type": "text", + "content": "Prompt tuning" + }, + { + "bbox": [ + 304, + 506, + 546, + 662 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 506, + 546, + 662 + ], + "type": "text", + "content": " is a technique that combines learnable textual tokens with class tokens to generate a discriminative textual classifier, known as Context Optimization (CoOp) [50]. Recently, various CoOp-based methods [4, 5, 17, 18, 22, 31, 40, 43, 50, 52] infer the domain-shared prompt tokens between training and testing domains (Figure 1(a)). Nevertheless, as the domain-shared prompt tokens are derived from labeled training images, their performance is suboptimal when confronted with unseen test classes. To enhance the generalization capacity of learnable prompt tokens, image-conditional prompts have been proposed in [45, 49] by fusing image features and learnable tokens (Figure 1(b)). Notably, image-conditional" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 306, + 673, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 673, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 673, + 547, + 713 + ], + "type": "text", + "content": "2For the visual-language model, two types of prompt-tuning strategies exist: visual prompt tuning and textual prompt tuning. In this work, we focus on the textual prompt tuning and do not consider the visual prompt tuning." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "1https://github.com/htyao89/Textual-based_ Class-aware_prompt_tuning" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23438" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "textual tokens encapsulate specific knowledge of each image, particularly for testing images, thereby rendering it more apt for generalization to unseen testing images. However, image-conditional prompt tokens with image-specific knowledge have less ability in elevating the distribution of class embeddings. In summary, classifiers generated by domain-shared and image-conditional textual tokens exhibit suboptimal performance for the unseen classes, primarily due to their inability to explicitly model the class distribution. Consequently, it is imperative to establish a dynamic relationship between the learnable prompt and the textual knowledge of each class to augment its discriminative prowess." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 229, + 289, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 289, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 289, + 458 + ], + "type": "text", + "content": "The frozen CLIP, coupled with a hand-crafted prompt, exhibits robust generalization capabilities to novel classes, rendering it a valuable source of prior textual knowledge for each class. By associating the class-level textual knowledge with the learnable prompts, a class-aware prompt can be formulated for hencing the discriminative capacity of the textual classifier. To achieve this, we employ an embedding module to project the class-aware textual knowledge into the class-aware prompt tokens, as shown in Figure 1(c). The resultant class-aware prompt incorporates prior textual knowledge specific to each class, endowing the generated textual classifier with heightened discriminative prowess. Furthermore, the class-aware prompt facilitates the generation of classifiers for both seen and unseen classes by leveraging textual knowledge from both categories. To sum up, the trained embedding module can generate a class-aware prompt for each class based on its description ('class name'), thereby enhancing generalization and discriminative capabilities in class-level textual embeddings." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 460, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 289, + 639 + ], + "type": "text", + "content": "Therefore, we propose a novel Textual-based Class-aware Prompt tuning (TCP) based on the framework of CoOp, shown in Figure 2. In addition to the domain-share textual tokens introduced in CoOp, TCP contributes a novel Textual Knowledge Embedding (TKE) to map the class-level textual knowledge into class-aware prompt tokens. Moreover, a class-aware textual classifier is generated by inserting the class-aware prompt tokens into the middle layer of the Text Encoder. We use standard contrastive loss and knowledge-guided consistency [43] to optimize the TKE and the learnable prompt tokens. During inference, TCP generates a class-aware classifier for unseen classes by feeding the domain-shared prompt tokens and the class-aware prompt tokens generated by the TKE into the frozen Text Encoder." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "content": "Overall, the proposed TCP explicitly steers prompts to learn a class-aware knowledge that maximizes the generalization and discriminative of the downstream tasks. Evaluation over 11 image classification datasets on base-to-new generalization, cross-dataset generalization, and few-shot learning verify shows that TCP is an efficient method that" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "text", + "content": "obtain a higher performance with less training time. In summary, the proposed Textual-based Class-aware Prompt tuning(TCP) has the following main contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 109, + 547, + 228 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 109, + 547, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 109, + 547, + 155 + ], + "spans": [ + { + "bbox": [ + 306, + 109, + 547, + 155 + ], + "type": "text", + "content": "1. An effectively Textual-based Class-aware Prompt tuning is proposed by injecting the textual class-aware prompts generated by Textual Knowledge Embedding(TKE) into the Text Encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 156, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 545, + 191 + ], + "type": "text", + "content": "2. We demonstrate that explicitly incorporating the prior knowledge of each class into the learnable prompt tokens can enhance the discriminative of the class distribution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 192, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 192, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 306, + 192, + 547, + 228 + ], + "type": "text", + "content": "3. Textual Knowledge Embedding(TKE) is a plug-and-play module that can quickly insert existing prompt tuning methods to improve their performance further." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 239, + 398, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 398, + 253 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 398, + 253 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 260, + 446, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 260, + 446, + 273 + ], + "spans": [ + { + "bbox": [ + 306, + 260, + 446, + 273 + ], + "type": "text", + "content": "2.1. Vision-Language Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 279, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 547, + 555 + ], + "type": "text", + "content": "Recently, researchers have shown that the Visual-Language Models(VLM) [2, 33] consisting of visual and textual modalities trained on the large-scale of image-text pairs has a powerful generalization and discriminativeion. To further boost the descriptive ability of VLM, the VLM models are boosted from the following aspect: 1) using a stronger text encoder or visual encoder [25, 41, 46]; 2) deeply fusing the visual and text knowledges [23, 38]; 3) using more images [16, 33, 35, 36]. To boost the diversity of text description, Masked Language Modeling (MLM) [20] [26] randomly erases the words in the text description used for representation learning. Unlike MLM, Masked autoencoder-based methods [13] are proposed to boost the descriptive ability by randomly masking the image patches. Among existing VLM models, CLIP is a representative and straightforward framework for inferring the independent visual and text encoder using the contrastive loss based on 400 million image-text association pairs. As CLIP has a good generalization, most existing CoOp-based methods are proposed based on CLIP for adapting the pre-trained VLM into the downstream task. Similar to existing methods, we conduct the prompt-tuning strategy on the TextEncoder of CLIP to obtain a task-specific textual embedding for prediction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 563, + 402, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 402, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 402, + 576 + ], + "type": "text", + "content": "2.2. Prompt Tuning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "To adapt the pretrained VLM to the downstream tasks, the prompt tuning [10, 22, 24, 31, 33, 45] always applies task-related textual tokens to infer the task-specific textual knowledge. In CLIP[33], the hand-crafted template \"a photo of a [CLASS]\" is used to embed the textual embedding for zero-shot prediction. However, the hand-crafted prompts have a poor ability to describe the downstream task. Textual prompt tuning is applied to boost the textual embedding by inferring a set of learnable textual tokens combined with the class tokens. For example, Context Optimization(CoOp) [50] replaces the hand-crafted prompts" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23439" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 455 + ], + "type": "text", + "content": "with the learnable soft prompts. To improve the generalization of the learnable textual prompt in CoOp, Conditional Context Optimization(CoCoOp) [49] and VPT [45] generates an image-conditional prompt fusing the image feature and the learnable textual prompt. Moreover, Knowledge-Guided Context Optimization(KgCoOp) [43], ProGrad [51], and Prompt Regularization(ProReg) [52] constrain the proposed learnable prompts contain the essential general knowledge. Unlike the above methods, which consider textual prompts, Ensembling Context Optimization(ECO) [1] employs prompt ensembling to combine multiple prompts. To obtain high-quality task-related tokens, ProDA [27] considers the prompt's prior distribution learning, and Distribution-Aware Prompt Tuning (DAPT) [5] optimizes the learnable prompt by maximizing inter-dispersion. Besides the textual knowledge from \"classname\", Knowledge-Aware Prompt Tuning(KAPT) [17] employs the external knowledge to generate the discriminative knowledge-aware prompt for the novel categories. PLOT [4] applies optimal transport to match the vision and text modalities for generating the discriminative and visual-aligned local textual prompt tokens. Besides the textual prompt tuning, Multi-modal Prompt Learning (MaPLe) [18] and PromptSRC [19] conduct the visual-textual prompt tuning by jointly conducting the prompt tuning on the visual and text encoders. Multitask Vision-Language Prompt Tuning(MVLPT) [37]incorporates crosstalk knowledge into prompt tuning for vision-language models. DenseCLIP [34] uses the context-aware prompt strategy to generate dense prediction tasks, and CLIP-Adapter [11] applies an adapter to adjust the visual or text embeddings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 457, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 289, + 639 + ], + "type": "text", + "content": "Existing methods commonly infer two types of prompt tokens: domain-share and image-conditional. However, the textual classifiers generated with these tokens tend to perform poorly on unseen classes. To mitigate this limitation, we propose a novel Textual-based Class-aware Prompt tuning (TCP) that employs a dynamic class-aware token to enhance the generalization and discriminative capabilities of the learnable textual prompt. Moreover, we introduce a Textual Knowledge Embedding to project class-level textual knowledge into the class-aware prompts, which are then inserted into the Text Encoder to generate a discriminative class-aware classifier. The evaluation results demonstrate that integrating class-level prior knowledge into the prompt tokens significantly enhances the discriminative ability of the prompt tuning process." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 654, + 125, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 125, + 668 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 125, + 668 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "As the Textual-based Class-aware Prompt tuning(TCP) is proposed based on Context Optimization (CoOp), we first briefly review CoOp and then introduce the proposed TCP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 72, + 397, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 397, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 397, + 83 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "text", + "content": "Existing CoOp-based methods are proposed based on the powerful Contrastive Language-Image Pre-training(CLIP). Given an image along with its corresponding textual description, CLIP uses visual and text encoders to extract the visual and text embeddings. After that, the constrastive loss between the visual and textual embeddings is calculated to align those two embeddings. To effectively adapt CLIP for the downstream task, CLIP applies the hand-crafted template \"a photo of a " + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "inline_equation", + "content": "\\{\\}" + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "text", + "content": "\" to extract the general class-level textual embedding, defined as " + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{clip} = \\{\\mathbf{w}_i^{clip}\\}_{i=1}^{N_c}" + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i^{clip}" + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "text", + "content": " is the textual embedding of " + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "text", + "content": "-th class, and " + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "inline_equation", + "content": "N_c" + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "text", + "content": " is the number of classes. Given the 'class-name' of " + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "text", + "content": "-th class, Word Embedded " + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "inline_equation", + "content": "e(\\cdot)" + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "text", + "content": " firstly embeds the hand-crafted description into a vectorized textual tokens: " + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i^{clip} = e(\"a photo of a \\{class-name\\}\")" + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "text", + "content": ". After that, Text Encoder " + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "text", + "content": " maps the vectorized textual tokens " + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i^{clip}" + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "text", + "content": " into the class-level embedding: " + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i^{clip} = \\theta (\\mathbf{t}_i^{clip})" + }, + { + "bbox": [ + 304, + 91, + 547, + 302 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "text", + "content": "To improve the discriminative of the class-level embedding, the prompt tuning methods of Context Optimization(CoOp) replace the hand-crafted textual tokens with a set of learnable textual tokens " + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathbb{T} = \\{\\mathbf{t}_1,\\mathbf{t}_2,\\dots,\\mathbf{t}_M\\}" + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "text", + "content": " is the length of tokens. Similar to CLIP, the corresponding class token " + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "text", + "content": " is concatenated with the learnable tokens " + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "text", + "content": " for generating the textual token " + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i^{coop} = \\{\\mathbf{t}_1,\\mathbf{t}_2,\\dots,\\mathbf{t}_M,\\mathbf{c}_i\\}" + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "text", + "content": ". Next, the textual embedding " + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i^{coop}" + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "text", + "content": " is obtained by fedding the textual tokens " + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i^{coop}" + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "text", + "content": " into Text Encoder " + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i^{coop} = \\theta (\\mathbf{t}_i^{coop})" + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "text", + "content": ". Finally, the textual embeddings of all classes are defined as " + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{coop} = \\{\\mathbf{w}_i^{coop}\\}_{i = 1}^{N_c}" + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "text", + "content": ". CoOp infers the learnable textual tokens " + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "text", + "content": " by minimizing the contrastive loss between the image's embedding " + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "text", + "content": " and its class embedding " + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_y^{coop}" + }, + { + "bbox": [ + 304, + 302, + 547, + 474 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 329, + 483, + 545, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 483, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 329, + 483, + 545, + 517 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c e} = \\frac {1}{N} \\sum_ {(\\mathbf {x}, y) \\in \\mathcal {D} _ {s}} \\frac {\\exp (d (\\mathbf {x} , \\mathbf {w} _ {y} ^ {c o o p}) / \\tau)}{\\sum_ {i = 1} ^ {N _ {c}} \\exp (d (\\mathbf {x} , \\mathbf {w} _ {i} ^ {c o o p}) / \\tau)}, \\tag {1}", + "image_path": "e7221603aea598fda71b11a0fafd857877fe27a75f2c20df31323b44cabac015.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 528, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 545, + 564 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 528, + 545, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 304, + 528, + 545, + 564 + ], + "type": "text", + "content": " is the seen dataset, and " + }, + { + "bbox": [ + 304, + 528, + 545, + 564 + ], + "type": "inline_equation", + "content": "d(\\cdot)" + }, + { + "bbox": [ + 304, + 528, + 545, + 564 + ], + "type": "text", + "content": " is the cosine distance. " + }, + { + "bbox": [ + 304, + 528, + 545, + 564 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 528, + 545, + 564 + ], + "type": "text", + "content": " is a temperature factor defined in CLIP, and " + }, + { + "bbox": [ + 304, + 528, + 545, + 564 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 528, + 545, + 564 + ], + "type": "text", + "content": " is the number of training images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 565, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 615 + ], + "type": "text", + "content": "As the generated textual embedding has a good generalization ability for the novel classes, " + }, + { + "bbox": [ + 304, + 565, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\mathrm{KgCoOp}" + }, + { + "bbox": [ + 304, + 565, + 545, + 615 + ], + "type": "text", + "content": " further adds an efficient consistency " + }, + { + "bbox": [ + 304, + 565, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{kg}" + }, + { + "bbox": [ + 304, + 565, + 545, + 615 + ], + "type": "text", + "content": " between the generated embedding " + }, + { + "bbox": [ + 304, + 565, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{coop}" + }, + { + "bbox": [ + 304, + 565, + 545, + 615 + ], + "type": "text", + "content": " and the general embedding " + }, + { + "bbox": [ + 304, + 565, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{clip}" + }, + { + "bbox": [ + 304, + 565, + 545, + 615 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 367, + 624, + 545, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 624, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 367, + 624, + 545, + 640 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {k g} = | | \\mathbf {W} ^ {c l i p} - \\mathbf {W} ^ {c o o p} | | _ {2} ^ {2}. \\tag {2}", + "image_path": "eda7c1834796add64c7e743fbb3c17fbf26809ba4763c03abb91cd3104c3daaa.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 652, + 520, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 652, + 520, + 664 + ], + "spans": [ + { + "bbox": [ + 318, + 652, + 520, + 664 + ], + "type": "text", + "content": "Therefore, a robust objective for prompt tuning is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 388, + 677, + 545, + 690 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 677, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 388, + 677, + 545, + 690 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c e} + \\omega \\mathcal {L} _ {k g}, \\tag {3}", + "image_path": "ea60c9bd60eba6f4d9bea5665d4d596647b29c835890408697e72daaf43ea1f2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 701, + 495, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 495, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 495, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 701, + 495, + 714 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 306, + 701, + 495, + 714 + ], + "type": "text", + "content": " is set as 8.0 same as in KgCoOp [43]." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23440" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 69, + 265, + 215 + ], + "blocks": [ + { + "bbox": [ + 72, + 69, + 265, + 215 + ], + "lines": [ + { + "bbox": [ + 72, + 69, + 265, + 215 + ], + "spans": [ + { + "bbox": [ + 72, + 69, + 265, + 215 + ], + "type": "image", + "image_path": "1b2afd8b794f699c29a1ea474d25a27953c3d54ffced838f82246af0eaaad4f2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 220, + 253, + 232 + ], + "lines": [ + { + "bbox": [ + 80, + 220, + 253, + 232 + ], + "spans": [ + { + "bbox": [ + 80, + 220, + 253, + 232 + ], + "type": "text", + "content": "Figure 2. The framework of the proposed TCP." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 239, + 266, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 266, + 252 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 266, + 252 + ], + "type": "text", + "content": "3.2. Textual-based Class-aware Prompt tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 258, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 289, + 555 + ], + "type": "text", + "content": "Based on the pre-trained Text Encoder in CLIP, the textual prompt tuning aims to infer a set of the domain-shared or image-conditional textual tokens combined with the general class tokens for generating the specific class embeddings. However, the textual classifier generated by those textual tokens performs a worse generalization on the unseen classes because they can not model the distribution of the testing classes. Research has shown that utilizing the general textual knowledge extracted by frozen CLIP can create a discriminative prior knowledge of novel classes, enhancing the discriminative and generalization of the learnable prompt. Drawing on the benefit of the general textual knowledge of seen and unseen classes, we propose a Textual-based Class-aware Prompt tuning(TCP) to adapt the pretrained CLIP to the downstream task. As shown in Figure 2, TCP employs a Textual Knowledge Embedding(TKE) to transfer the general class-level textual embedding into the class-aware prompt, which is then combined with the learnable textual tokens for generating the class-aware classifiers. TKE is advantageous for unseen classes, as it generates class-specific prompts to obtain an unseen class-aware textual classifier with better discriminative abilities. Moreover, explicitly incorporating seen class-aware prompts can enhance the discriminative power of seen classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": "Given the general class-level textual embedding " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{clip} \\in \\mathbb{R}^{N_c \\times D_t}" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "N_c" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": " training classes, Textual Knowledge Embedding(TKE) " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": " is proposed to project the class-level embedding " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{clip}" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": " into the class-aware prompt " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = \\mathcal{T}(\\mathbf{W}^{clip})" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": ". As shown in Figure 2, TKE consists of two layers: down-project layer and up-project layer. The down-project layer uses the weight " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{down} \\in \\mathbb{R}^{D_t \\times D_{mid}}" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": " to project the textual embedding into the low-dimension feature with the dimension of " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "D_{mid}" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": ". Next, the weight " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{up} \\in \\mathbb{R}^{D_{mid} \\times D'}" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": " of the up-project layer maps the low-dimension feature into the high-dimension feature with the dimension of " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "D'" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "D'" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": " is determined by the prompt's length " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": " and dimension " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "D' = M \\times D" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": ". In" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 71, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 71, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 71, + 545, + 131 + ], + "type": "text", + "content": "summary, the general textual embedding " + }, + { + "bbox": [ + 304, + 71, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{clip} \\in \\mathbb{R}^{N_c \\times D_t}" + }, + { + "bbox": [ + 304, + 71, + 545, + 131 + ], + "type": "text", + "content": " can be projected into the class-aware textual tokens " + }, + { + "bbox": [ + 304, + 71, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in \\mathbb{R}^{N_c \\times D'}" + }, + { + "bbox": [ + 304, + 71, + 545, + 131 + ], + "type": "text", + "content": ", which is further reshaped into the shape of " + }, + { + "bbox": [ + 304, + 71, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in \\mathbb{R}^{N_c \\times M \\times D}" + }, + { + "bbox": [ + 304, + 71, + 545, + 131 + ], + "type": "text", + "content": " for inserting into the middle layer of Text Encoder " + }, + { + "bbox": [ + 304, + 71, + 545, + 131 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 71, + 545, + 131 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "text", + "content": "Assuming we insert the class-aware prompt " + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "text", + "content": "-th layer of Text Encoder " + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "text", + "content": ". We will give a detailed analysis of the hyperparameter " + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "text", + "content": " in the following. Similar to CoOp, by combining the domain-shared learnable textual tokens " + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "inline_equation", + "content": "\\mathbb{T} = [\\mathbf{t}_1,\\mathbf{t}_2,\\dots,\\mathbf{t}_M]" + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "text", + "content": " and the pre-trained class tokens " + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "text", + "content": " of all classes, we can obtain the input textual tokens " + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_0 = \\{\\mathbb{T},\\mathbf{C}\\}" + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "text", + "content": " of the Text Encoder, where " + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = \\{\\mathbf{c}_i\\}_{i}^{N_c}" + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "text", + "content": " is the vectorized textual tokens of " + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "text", + "content": "-th class. The textual token " + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_0" + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "text", + "content": " is fed into the first " + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "text", + "content": " layers of Text Encoder for obtaining the middle-level textual embedding " + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l" + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "text", + "content": ". Formally, the textual token " + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i(i\\leq l)" + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 133, + 546, + 264 + ], + "type": "text", + "content": "-th layer is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 373, + 271, + 545, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 271, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 373, + 271, + 545, + 285 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {i} = \\theta_ {i} \\left(\\mathbf {F} _ {i - 1}\\right), i \\in [ 1, l ], \\tag {4}", + "image_path": "3917f5f367c45753290541c486ab7418a9cf2821747ee72608bb480ed4d2d837.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 292, + 476, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 476, + 303 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 476, + 303 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 292, + 476, + 303 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 304, + 292, + 476, + 303 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 292, + 476, + 303 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 292, + 476, + 303 + ], + "type": "text", + "content": "-th layer of Text Encoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 303, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 388 + ], + "type": "text", + "content": "For the textual tokens " + }, + { + "bbox": [ + 304, + 303, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l\\in \\mathbb{R}^{N_c\\times N_t\\times D}" + }, + { + "bbox": [ + 304, + 303, + 545, + 388 + ], + "type": "text", + "content": " and the class-aware prompt tokens " + }, + { + "bbox": [ + 304, + 303, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{T}\\in \\mathbb{R}^{N_c\\times M\\times D}" + }, + { + "bbox": [ + 304, + 303, + 545, + 388 + ], + "type": "text", + "content": ", the first dimension is related to the number of classes. Therefore, the learnable prompt tokens are always inserted into the second dimension of " + }, + { + "bbox": [ + 304, + 303, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l" + }, + { + "bbox": [ + 304, + 303, + 545, + 388 + ], + "type": "text", + "content": ", the same as CoOp. Formally, the class-aware prompt " + }, + { + "bbox": [ + 304, + 303, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 303, + 545, + 388 + ], + "type": "text", + "content": " is inserted into " + }, + { + "bbox": [ + 304, + 303, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l" + }, + { + "bbox": [ + 304, + 303, + 545, + 388 + ], + "type": "text", + "content": " to generate the class-aware enhanced tokens " + }, + { + "bbox": [ + 304, + 303, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l'" + }, + { + "bbox": [ + 304, + 303, + 545, + 388 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 394, + 545, + 408 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 394, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 315, + 394, + 545, + 408 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {l} ^ {\\prime} = \\left[ \\mathbf {T} _ {1}, \\mathbf {T} _ {2}, \\dots , \\mathbf {T} _ {M}, \\mathbf {F} _ {l, M + 1}, \\mathbf {F} _ {l, M + 2}, \\dots , \\mathbf {F} _ {l, N _ {t}} \\right], \\tag {5}", + "image_path": "437bce7bed2346a835a493c759392e23b4e6d1f5ae842885436c8e00d56eeb9a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i" + }, + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "text", + "content": "-th index of " + }, + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "text", + "content": " in the second dimension, and " + }, + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{l,j}" + }, + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "text", + "content": " denote the correspond " + }, + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "text", + "content": "-th index of " + }, + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l" + }, + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "text", + "content": " in the second dimension, i.e., " + }, + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i = \\mathbf{T}[:i,:]" + }, + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{l,j} = \\mathbf{F}_l[:j,:]" + }, + { + "bbox": [ + 304, + 415, + 545, + 450 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 450, + 545, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 487 + ], + "type": "text", + "content": "After that, the class-enhanced textual tokens " + }, + { + "bbox": [ + 304, + 450, + 545, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l^{\\prime}" + }, + { + "bbox": [ + 304, + 450, + 545, + 487 + ], + "type": "text", + "content": " are fedded into the rest layers for generating the class-aware textual embedding," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 364, + 494, + 545, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 494, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 364, + 494, + 545, + 509 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {i} ^ {\\prime} = \\theta_ {i} (\\mathbf {F} _ {i - 1} ^ {\\prime}), i \\in [ l + 1, L ]. \\qquad (6)", + "image_path": "93edaabf21aaf08eac58ced418c090ee981c9c023d2806a4a8816d4f41117cd0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 514, + 545, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 551 + ], + "type": "text", + "content": "The output of the last layer " + }, + { + "bbox": [ + 304, + 514, + 545, + 551 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_L^{\\prime}" + }, + { + "bbox": [ + 304, + 514, + 545, + 551 + ], + "type": "text", + "content": " is treated as the class embedding " + }, + { + "bbox": [ + 304, + 514, + 545, + 551 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{tcp}" + }, + { + "bbox": [ + 304, + 514, + 545, + 551 + ], + "type": "text", + "content": " used for optimization with contrastive loss and knowledge-guided consistency loss in Eq. (3)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 560, + 388, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 560, + 388, + 574 + ], + "spans": [ + { + "bbox": [ + 305, + 560, + 388, + 574 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 580, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 653 + ], + "type": "text", + "content": "Similar to CoOp [50], we evaluate the effectiveness of TCP from the three types of tasks: 1) generalization from baseto-new classes within a dataset; 2) few-shot learning with " + }, + { + "bbox": [ + 304, + 580, + 545, + 653 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 580, + 545, + 653 + ], + "type": "text", + "content": "-shot labeled images; 3) cross-dataset generalization from the imagenet to other datasets. More detailed results will be provided in the Supplementary materials." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 659, + 422, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 659, + 422, + 672 + ], + "spans": [ + { + "bbox": [ + 305, + 659, + 422, + 672 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "Dataset. We conduct the base-to-new generalization, few-shot learning, and cross-dataset generalization on 11 benchmarks, i.e., ImageNet [7], Caltech [9], OxfordPets [32]," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23441" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 113, + 578, + 432 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 105 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 105 + ], + "type": "text", + "content": "Table 1. Comparison on the base-to-new generalization setting with 16-shot samples. 'tp', 'dtp', 'vp', and 'dvp' denote the 'textual prompt', 'deep textual prompt', 'visual prompt', and 'deep visual prompt', respectively. PromptSRC are based on deep visual-textual prompt tuning ('dvp+dtp'). *' denote the performance obtained by our re-implementation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 113, + 578, + 432 + ], + "lines": [ + { + "bbox": [ + 47, + 113, + 578, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 113, + 578, + 432 + ], + "type": "table", + "html": "
DatasetsSetsCoOp* (JCV22)CoCoOp (CVPR22)DAPT* (ICCV23)ProGrad* (ICCV23)ProDA (CVPR22)KgCoOp (CVPR23)RPO (ICCV23)PLOT* (ICLR23)LFA (ICCV23)MaPLe (CVPR23)DePT (Arxiv23)PromptSRC* (ICCV23)TCP
tptptp+vptptptpdtp+dvptp+vp-dtp+dvptpdtp+dvptp
AverageBase82.3880.4783.1882.4881.5680.7381.1383.9883.6282.2883.6284.1284.13
New67.9671.6969.2770.7572.3073.675.0071.7274.5675.1475.0475.0275.36
H74.4875.8375.5976.1676.6577.077.7877.3778.8378.5579.1079.3179.51
ImageNetBase76.4675.9876.8377.0275.4075.8376.6077.3076.8976.6677.0377.7577.27
New66.3170.4369.2766.6670.2369.9671.5769.8769.3670.5470.1370.7069.87
H71.0273.1072.8571.4672.7272.7874.0073.4072.9373.4773.4274.0673.38
Caltech101Base97.8097.9697.8398.0298.2797.7297.9798.5398.4197.7498.3098.1398.23
New93.2793.8193.0793.8993.2394.3994.3792.8093.9394.3694.6093.9094.67
H95.4895.8495.3995.9195.6896.0396.0395.5896.1396.0296.4195.9796.42
OxfordPetsBase94.4795.2095.0095.0795.4394.6594.6394.5095.1395.4394.3395.5094.67
New96.0097.6995.8397.6397.8397.7697.5096.8396.2397.7697.2397.4097.20
H95.2396.4395.4196.3396.6296.1896.0595.6595.6896.5895.7696.4495.92
CarsBase75.6770.4975.8077.6874.7071.7673.8779.0776.3272.9479.1378.4080.80
New67.5373.5963.9368.6371.2075.0475.5374.8074.8874.0075.4774.7374.13
H71.3772.0169.3672.8872.9173.3674.6976.8875.5973.4777.2675.5277.32
FlowersBase97.2794.8796.9795.5497.7095.0094.1397.9397.3495.9298.0097.9097.73
New67.1371.7560.9071.8768.6874.7376.6773.5375.4472.4676.3776.7775.57
H79.4481.7174.8182.0380.6683.6584.5083.9985.0082.5685.8486.0685.23
Food101Base89.3790.7090.3790.3790.3090.590.3389.8090.5290.7190.5090.6390.57
New88.7791.2991.3089.5988.5791.790.8391.3791.4892.0591.6091.5091.37
H89.0790.9990.8389.9889.4391.0990.5890.5891.0091.3891.0591.0690.97
AircraftBase39.6733.4139.9740.5436.9036.2137.3342.1341.4837.4443.2042.3041.97
New31.2323.7129.8027.5734.1333.5534.2033.7332.2935.6134.8336.9734.43
H34.9527.7434.1432.8235.4634.8335.7037.4636.3136.5038.5739.4637.83
SUN397Base80.8579.7480.9781.2678.6780.2980.6082.2082.1380.8282.3382.8382.63
New68.3476.8676.9774.1776.9376.5377.8073.6377.2078.7077.8079.0078.20
H74.0778.2778.9277.5577.7978.3679.1877.6879.5979.7580.0080.8780.35
DTDBase79.9777.0182.2377.3580.6777.5576.7081.9781.2980.3682.2082.6082.77
New48.6056.0054.2352.3556.4854.9962.1343.8060.6359.1859.1357.5058.07
H60.4664.8565.3662.4566.4464.3568.6157.0969.4668.1668.7867.8068.25
EuroSATBase90.1087.4994.7390.1183.9085.6486.6393.7093.4094.0789.0392.4091.63
New53.0060.0450.3360.8966.0064.3468.9762.6771.2473.2371.0768.4374.73
H66.7471.2165.7472.6773.8873.4876.7975.1180.8382.379.0478.6382.32
UCF101Base84.5382.3384.3084.3385.2382.8983.6786.6086.9783.0085.8086.9387.13
New67.3773.4576.3374.9471.9776.6775.4375.9077.4878.6677.2378.3380.77
H74.9877.6780.1279.3578.0479.6579.3480.9081.9580.7781.2982.4183.83
", + "image_path": "0ea82750e6c47a62101faf4c779c2448f9196551956cd50b54609164fa5bee1b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 444, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 287, + 491 + ], + "type": "text", + "content": "StanfordCars [21], Flowers [29], Food101 [3], FGVCAircraft [28], EuroSAT [14], UCF101 [39], DTD [6], and SUN397 [42]. A detailed introduction of those dataset is given in the Supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 492, + 287, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 287, + 600 + ], + "type": "text", + "content": "Training Details. Our implementation is based on CoOp's code [50]. All experiments are conducted based on the CLIP with the backbone of ViT-B/16 [8]. The prompt's length " + }, + { + "bbox": [ + 46, + 492, + 287, + 600 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 492, + 287, + 600 + ], + "type": "text", + "content": " is set as 4 with random initialization of \"X X X X {}\". The final performance is averaged over three random seeds(1/2/3)for a fair comparison. Adam optimizer is applied for optimization with the learning rate of 2e-3 and the batch size of 32, and the training epochs is 50. All experiments are conducted on RTX 3090." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 601, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 661 + ], + "type": "text", + "content": "Baselines. Recently CoOp-based methods are used for comparison, e.g., CoOp [50], CoCoOp [49], ProGrad [51], ProDA [27], KgCoOp [43], PromptSRC [19], MaPLe [18], LFA [31], DePT [47], DAPT [5], PLOT [4], TaskRes [44], RPO [22], VPT [45], and TIP-Adapter-F [48]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 670, + 230, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 230, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 230, + 682 + ], + "type": "text", + "content": "4.2. Base-to-New Class Generalization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "To verify the generalization ability of the prompt tuning, the base-to-new class generalization splits each dataset into two" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 444, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 545, + 529 + ], + "type": "text", + "content": "disjoint subsets: Base and New classes, where Base classes are used to infer the learnable prompt and TKE, and New classes are used for evaluation. Note that each New class contains 16-shot samples. We summarize the comparison between the proposed TCP and existing methods in Table 1 from the following two aspects: Textual Prompt Tuning and Vivasal-Textual Prompt Tuning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "type": "text", + "content": "Textual Prompt Tuning: Based on the visual feature extracted from the frozen visual encoder, the textual prompt tuning (e.g., CoOp, CoCoOp, ProGrad, ProDA, and Kg-CoOp) merely conducts the textual prompt tuning by inserting a set of learnable textual prompts into the Text Encoder for obtaining a discriminative textual classifier, which is same as our proposed TCP. Among all textual prompt tuning methods, KgCoOp [43] is a more substantial baseline, which is also the baseline of our TCP. Compared to KgCoOp, TCP improves the metric term of Base/New/H from " + }, + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "type": "inline_equation", + "content": "80.73\\% / 73.6\\% / 77.0\\%" + }, + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "type": "inline_equation", + "content": "84.35\\% / 75.28\\% / 79.56\\%" + }, + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "type": "text", + "content": ". We can observe that a noticeable improvement of " + }, + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "type": "inline_equation", + "content": "3.62\\%" + }, + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "type": "inline_equation", + "content": "1.68\\%" + }, + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "type": "text", + "content": " are obtained in the Base class and New class, respectively. The superior performance proves that TCP using the class-aware prompt to inject the class knowledge can generate" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23442" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 91, + 102, + 503, + 243 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": "Table 2. Comparison of cross-dataset evaluation. 'tp', 'dtp', 'vp', and 'dvp' denote the 'textual prompt', 'deep textual prompt', 'visual prompt', and 'deep visual prompt', respectively. Note that DAPT and MaPLe are based on visual-textual prompt tuning('vp+tp')." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 91, + 102, + 503, + 243 + ], + "lines": [ + { + "bbox": [ + 91, + 102, + 503, + 243 + ], + "spans": [ + { + "bbox": [ + 91, + 102, + 503, + 243 + ], + "type": "table", + "html": "
DatasetsCLIPCoOpProGradKgCoOpDePTVPTPLOTPromptSRCMaPLeDAPTTCP
tptptptptp+vptp+vpdtp+tvpdtp+dvptp+vptp
ImageNet66.7071.5172.2470.6672.7769.7371.6071.2770.7271.6071.40
Caltech10193.3093.7091.5293.9294.2393.6792.0793.6093.5393.5093.97
OxfordPets89.1089.1489.6489.8390.0389.2790.1090.2590.4990.6791.25
StandfordCars65.7064.5162.3965.4165.5765.565.7065.7065.5765.9364.69
Flowers70.7068.7167.8770.0170.5770.269.2370.2572.2071.7071.21
Food10185.9085.3085.4086.3686.3786.2786.2386.1586.2086.1086.69
FGVCAircraft24.9018.4720.1622.5123.2722.1325.0023.9024.7423.0323.45
SUN39762.6064.1562.4766.1666.6766.5761.6767.1067.0167.0067.15
DTD44.3041.9239.4246.3545.9746.9338.6046.8746.4944/0044.35
EuroSAT48.3046.3943.4646.0443.5347.4347.8345.5048.0652.4751.45
UCF10167.6066.5564.2968.5069.3067.2067.0068.7568.6968.7368.73
Avg.65.2463.8862.7165.5165.5565.5264.3465.8166.3066.3166.29
", + "image_path": "a694851e5404f09d16231a263d24315b169cd2c260ef0809cd73de9a1f91957b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 255, + 288, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 288, + 339 + ], + "type": "text", + "content": "the discriminative textual classifier for both seen and unseen classes. DePT [47] is another work that also treats KgCoOp as the baseline and obtains the best " + }, + { + "bbox": [ + 46, + 255, + 288, + 339 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 255, + 288, + 339 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 255, + 288, + 339 + ], + "type": "inline_equation", + "content": "79.10\\%" + }, + { + "bbox": [ + 46, + 255, + 288, + 339 + ], + "type": "text", + "content": " among all textual prompt tuning methods. Compared to DePT, the proposed TCP obtains the performance of " + }, + { + "bbox": [ + 46, + 255, + 288, + 339 + ], + "type": "inline_equation", + "content": "84.35\\% / 75.28\\% / 79.56\\%" + }, + { + "bbox": [ + 46, + 255, + 288, + 339 + ], + "type": "text", + "content": " in the metric term of Base/New/H, achieving the improvement of " + }, + { + "bbox": [ + 46, + 255, + 288, + 339 + ], + "type": "inline_equation", + "content": "0.51\\% / 0.32\\% / 0.41\\%" + }, + { + "bbox": [ + 46, + 255, + 288, + 339 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 340, + 290, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 290, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 290, + 519 + ], + "type": "text", + "content": "Visual-Textual Prompt Tuning: Different from the textual prompt tuning merely adapts Text Encoder, DAPT [5], PLOT [4], RPO [22], MaPLe [18], and PromptSRC [19] all jointly performs prompt tuning on the text and visual encoders for both improving the discriminative of textual embedding and visual embeddings. Unlike CoOp and TCP, which conduct the prompt tuning on one or two layers, those methods conduct the visual-textual prompt tunings on several layers of visual and text encoders. Compared to them, the proposed TCP with shallow textual prompt tuning obtains a higher performance, e.g., " + }, + { + "bbox": [ + 46, + 340, + 290, + 519 + ], + "type": "inline_equation", + "content": "79.51\\%" + }, + { + "bbox": [ + 46, + 340, + 290, + 519 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 46, + 340, + 290, + 519 + ], + "type": "inline_equation", + "content": "79.31\\%" + }, + { + "bbox": [ + 46, + 340, + 290, + 519 + ], + "type": "text", + "content": " for PromptSRC in terms of " + }, + { + "bbox": [ + 46, + 340, + 290, + 519 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 340, + 290, + 519 + ], + "type": "text", + "content": ". Compared to deep visual-textual prompt tuning, the superior performance demonstrates that class-aware prompt tuning can generate discriminative and generalization textual-level classifiers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 527, + 209, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 209, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 209, + 540 + ], + "type": "text", + "content": "4.3. Cross-Dataset Generalization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "content": "In the Base-to-New generalization setting, New classes always have a similar data distribution as the Base classes. To further verify the generalization of the proposed TCP, in the Cross-Dataset generalization, TCP is trained from the ImageNet and directly evaluated on the unrelated datasets, e.g., the rest ten datasets. The comparison between the proposed TCP and existing methods is summarized in Table 2. From Table 2, we can observe that the proposed TCP obtains the highest average performance among all textual prompt tuning methods (66.29% vs 65.55% of DePT [47]), and obtains a comparable performance with the visual-textual prompt tuning methods (66.29% vs 66.31% of DAPT [5]), demonstrating the effectiveness of TCP in learning the generalization knowledge." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 328, + 296, + 524, + 398 + ], + "blocks": [ + { + "bbox": [ + 305, + 254, + 547, + 288 + ], + "lines": [ + { + "bbox": [ + 305, + 254, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 305, + 254, + 547, + 288 + ], + "type": "text", + "content": "Table 3. Effect of TCP on CoOp, KgCoOp, ProGrad, PromptSRC, and DAPT in the base-to-new generalization setting regarding final average performance." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 296, + 524, + 398 + ], + "lines": [ + { + "bbox": [ + 328, + 296, + 524, + 398 + ], + "spans": [ + { + "bbox": [ + 328, + 296, + 524, + 398 + ], + "type": "table", + "html": "
MethodsBaseNewH
CoOp [50]82.6367.9974.6
CoOp+TKE83.10(↑ 0.47)70.17(↑ 2.18)76.09 (↑ 1.49)
KgCoOp [43]80.7373.677.00
KgCoOp +TKE84.13 (↑ 3.40)75.36(↑ 1.76)79.51 (↑ 2.51)
ProGrad [51]82.4870.7576.16
ProGrad+TKE82.61 (↑ 0.13)72.91(↑ 2.16)77.46 (↑ 1.30)
PromptSRC [19]82.0774.8378.03
PromptSRC+TKE83.74 (↑ 1.67)75.85(↑ 1.02)79.60 (↑ 1.57)
DAPT [5]83.1869.2775.59
DAPT+TKE84.15 (↑ 0.97)74.87(↑ 5.60)79.24 (↑ 3.65)
", + "image_path": "fbf152178c98eacaff336c354fe5b3e2534150cd88ea03d6da70c4ab3db45d30.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 411, + 438, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 438, + 423 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 438, + 423 + ], + "type": "text", + "content": "4.4. Few-shot Classification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 430, + 547, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 547, + 538 + ], + "type": "text", + "content": "To verify that the proposed TCP can infer the class-aware knowledge, the few-shot classification is conducted on all 11 datasets using K-shot labeled source images and evaluated on the standard testing domain with the same class space as the training classes. The 4-shot setting comparison between the proposed TCP and existing methods is summarized in Table 4. We can observe that the proposed TCP achieves a higher average performance than existing methods, i.e., obtaining the average performance of " + }, + { + "bbox": [ + 304, + 430, + 547, + 538 + ], + "type": "inline_equation", + "content": "76.32\\%" + }, + { + "bbox": [ + 304, + 430, + 547, + 538 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 546, + 400, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 546, + 400, + 559 + ], + "spans": [ + { + "bbox": [ + 305, + 546, + 400, + 559 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 564, + 547, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 547, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 547, + 686 + ], + "type": "text", + "content": "Effect of TKE's generalization. The critical component of TCP is the class-aware prompt generated by TKE, which embeds the prior class knowledge and can be easily integrated with existing CoOp-based methods. As shown in Table 3, CoOp, KgCoOp, ProGrad, PromptSRC, and DAPT have all demonstrated improved performance when TKE is used. Specifically, improvements of " + }, + { + "bbox": [ + 304, + 564, + 547, + 686 + ], + "type": "inline_equation", + "content": "1.49\\%" + }, + { + "bbox": [ + 304, + 564, + 547, + 686 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 564, + 547, + 686 + ], + "type": "inline_equation", + "content": "2.51\\%" + }, + { + "bbox": [ + 304, + 564, + 547, + 686 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 564, + 547, + 686 + ], + "type": "inline_equation", + "content": "1.30\\%" + }, + { + "bbox": [ + 304, + 564, + 547, + 686 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 564, + 547, + 686 + ], + "type": "inline_equation", + "content": "1.57\\%" + }, + { + "bbox": [ + 304, + 564, + 547, + 686 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 564, + 547, + 686 + ], + "type": "inline_equation", + "content": "3.65\\%" + }, + { + "bbox": [ + 304, + 564, + 547, + 686 + ], + "type": "text", + "content": " were achieved for CoOp, KgCoOp, ProGrad, PromptSRC3, and DAPT, respectively. This superior performance highlights TCP's plug-and-play functionality," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 547, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 547, + 715 + ], + "type": "text", + "content": "3PromptSRC is re-implemented with the shallow visual-textual prompt tuning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23443" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 91, + 535, + 213 + ], + "blocks": [ + { + "bbox": [ + 182, + 71, + 412, + 83 + ], + "lines": [ + { + "bbox": [ + 182, + 71, + 412, + 83 + ], + "spans": [ + { + "bbox": [ + 182, + 71, + 412, + 83 + ], + "type": "text", + "content": "Table 4. Comparison of few-shot learning with 4-shot samples." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 91, + 535, + 213 + ], + "lines": [ + { + "bbox": [ + 58, + 91, + 535, + 213 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 535, + 213 + ], + "type": "table", + "html": "
CLIPCoOpCoCoOpProGradKgCoOpMaPLeTIP-Adapter-FDAPTPromptSRCPLOTTaskResTCP
ImageNet66.7069.3770.5570.2170.1970.6770.7870.8070.8070.4062.8770.48
Caltech10193.3094.4494.9894.9394.6594.3094.7794.2394.7795.1394.6795.00
OxfordPets89.1091.3093.0193.2193.2092.0592.2692.1793.2392.5592.0091.90
StandfordCars65.7072.7369.1071.7571.9868.7074.4274.4071.8374.9375.9076.30
Flowers70.7091.1482.5689.9890.6980.8092.9892.3791.3192.9391.5094.40
Food10185.9082.5886.6485.7786.5986.9086.1883.6086.0686.4686.0385.3
FGVCAircraft24.9033.1830.8732.9332.4729.0335.4932.4732.8035.2933.8036.20
SUN39762.6070.1370.5071.1771.7971.4770.6572.2072.8070.4272.7072.11
DTD44.3058.5754.7957.7258.3154.7361.7061.3760.6462.4359.5763.97
EuroSAT48.3068.6263.8370.8471.0654.8778.2772.7375.0280.7072.8777.43
UCF10167.6077.4174.9977.8278.4073.7079.7379.4079.3579.7676.1080.83
Avg.65.3773.5971.9874.2174.4870.6676.1175.0775.3376.4574.3676.72
", + "image_path": "28a6341b11fc12a5141aeb2206d65f63bb476ab88923c33de8e01cdce5235b87.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 54, + 243, + 279, + 291 + ], + "blocks": [ + { + "bbox": [ + 64, + 222, + 269, + 234 + ], + "lines": [ + { + "bbox": [ + 64, + 222, + 269, + 234 + ], + "spans": [ + { + "bbox": [ + 64, + 222, + 269, + 234 + ], + "type": "text", + "content": "Table 5. Domain-shared prompt vs Class-aware prompt." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 243, + 279, + 291 + ], + "lines": [ + { + "bbox": [ + 54, + 243, + 279, + 291 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 279, + 291 + ], + "type": "table", + "html": "
BaselineDomain-SharedClass-AwareBaseNewH
✓✓80.7373.677
✓✓84.0575.1879.36
✓✓✓84.1375.3679.51
", + "image_path": "3a4426026c1df7113bc2af0d171a17f81ee2d2f4c4863930abbcf53d206fa2c1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 75, + 315, + 252, + 402 + ], + "blocks": [ + { + "bbox": [ + 75, + 315, + 252, + 402 + ], + "lines": [ + { + "bbox": [ + 75, + 315, + 252, + 402 + ], + "spans": [ + { + "bbox": [ + 75, + 315, + 252, + 402 + ], + "type": "image", + "image_path": "a7fd9172650d9bdbfc8186a426a2601cc631977f3ed0728679d2df6e5b7f0205.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 412, + 284, + 423 + ], + "lines": [ + { + "bbox": [ + 50, + 412, + 284, + 423 + ], + "spans": [ + { + "bbox": [ + 50, + 412, + 284, + 423 + ], + "type": "text", + "content": "Figure 3. Effect of " + }, + { + "bbox": [ + 50, + 412, + 284, + 423 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 50, + 412, + 284, + 423 + ], + "type": "text", + "content": " on the base-to-new generalization setting." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 378, + 361, + 389 + ], + "lines": [ + { + "bbox": [ + 307, + 378, + 361, + 389 + ], + "spans": [ + { + "bbox": [ + 307, + 378, + 361, + 389 + ], + "type": "inline_equation", + "content": "D_{mid}" + }, + { + "bbox": [ + 307, + 378, + 361, + 389 + ], + "type": "text", + "content": " in TKE." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 435, + 244, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 435, + 244, + 447 + ], + "spans": [ + { + "bbox": [ + 47, + 435, + 244, + 447 + ], + "type": "text", + "content": "allowing easy integration with existing methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 449, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 287, + 663 + ], + "type": "text", + "content": "Domain-shared prompt vs the class-aware prompt. It is common for existing methods to infer the domain-shared prompt tokens. However, TCP differs significantly from them in using a class-aware prompt to capture class-aware knowledge in the Text Encoder. We thus compare the domain-shared and class-aware prompts, and summarize the related results in Table 5. We can observe that simply using the class-aware prompt yields a fantastic performance of " + }, + { + "bbox": [ + 47, + 449, + 287, + 663 + ], + "type": "inline_equation", + "content": "79.36\\%" + }, + { + "bbox": [ + 47, + 449, + 287, + 663 + ], + "type": "text", + "content": " in terms of " + }, + { + "bbox": [ + 47, + 449, + 287, + 663 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 47, + 449, + 287, + 663 + ], + "type": "text", + "content": ", higher than existing methods. Additionally, as shown in Table 5, using the class-aware prompt results in a noticeable improvement over the domain-shared prompt. For instance, the Basel/New performance is improved from " + }, + { + "bbox": [ + 47, + 449, + 287, + 663 + ], + "type": "inline_equation", + "content": "80.73\\% / 73.6\\%" + }, + { + "bbox": [ + 47, + 449, + 287, + 663 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 449, + 287, + 663 + ], + "type": "inline_equation", + "content": "84.05\\% / 75.18\\%" + }, + { + "bbox": [ + 47, + 449, + 287, + 663 + ], + "type": "text", + "content": ". The superior performance proves the reasonableness and effectiveness of injecting the class-aware knowledge into the prompt tuning. By further considering the domain-shared and class-aware prompts, the final performance is reached at " + }, + { + "bbox": [ + 47, + 449, + 287, + 663 + ], + "type": "inline_equation", + "content": "79.51\\%" + }, + { + "bbox": [ + 47, + 449, + 287, + 663 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "content": "Effect of Prompt length " + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "content": ". In this study, we examine the impact of prompt length " + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "(M)" + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "content": " on the Base-to-New generalization setting. We compared prompt lengths of 1, 2, 4, 8, and 16, and the results are presented in Figure 3. The" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 352, + 243, + 499, + 289 + ], + "blocks": [ + { + "bbox": [ + 345, + 222, + 505, + 234 + ], + "lines": [ + { + "bbox": [ + 345, + 222, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 345, + 222, + 505, + 234 + ], + "type": "text", + "content": "Table 6. Comparison of different templates." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 352, + 243, + 499, + 289 + ], + "lines": [ + { + "bbox": [ + 352, + 243, + 499, + 289 + ], + "spans": [ + { + "bbox": [ + 352, + 243, + 499, + 289 + ], + "type": "table", + "html": "
TemplatesBaseNewH
‘X X X X’84.1375.3679.51
‘a photo of a’83.9474.9479.18
‘this is a picture’84.0074.6679.06
", + "image_path": "310d25b9fe2e461642adb82ddb093bef6b64e77316f175c279a345cfd04318a7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 293, + 423, + 358 + ], + "blocks": [ + { + "bbox": [ + 307, + 293, + 423, + 358 + ], + "lines": [ + { + "bbox": [ + 307, + 293, + 423, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 423, + 358 + ], + "type": "image", + "image_path": "5b18c2a1da65e25595fdbf49500f754ac88e936722af28d90c4f5ff65b93a8f9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 368, + 424, + 377 + ], + "lines": [ + { + "bbox": [ + 306, + 368, + 424, + 377 + ], + "spans": [ + { + "bbox": [ + 306, + 368, + 424, + 377 + ], + "type": "text", + "content": "Figure 4. Effect of dimension" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 425, + 298, + 539, + 362 + ], + "blocks": [ + { + "bbox": [ + 425, + 298, + 539, + 362 + ], + "lines": [ + { + "bbox": [ + 425, + 298, + 539, + 362 + ], + "spans": [ + { + "bbox": [ + 425, + 298, + 539, + 362 + ], + "type": "image", + "image_path": "ad2c24d0f099c4e9ac56637ca53bc03d6b373c303ae46a847443b36ca6fd1f6a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 373, + 539, + 385 + ], + "lines": [ + { + "bbox": [ + 424, + 373, + 539, + 385 + ], + "spans": [ + { + "bbox": [ + 424, + 373, + 539, + 385 + ], + "type": "text", + "content": "Figure 5. Effect of insert layer " + }, + { + "bbox": [ + 424, + 373, + 539, + 385 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 424, + 373, + 539, + 385 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 402, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 474 + ], + "type": "text", + "content": "findings indicate that longer prompt lengths produce better performance, with " + }, + { + "bbox": [ + 304, + 402, + 545, + 474 + ], + "type": "inline_equation", + "content": "M = 8" + }, + { + "bbox": [ + 304, + 402, + 545, + 474 + ], + "type": "text", + "content": " producing the best results. However, the performance gap between the highest and lowest values of " + }, + { + "bbox": [ + 304, + 402, + 545, + 474 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 402, + 545, + 474 + ], + "type": "text", + "content": ", which were " + }, + { + "bbox": [ + 304, + 402, + 545, + 474 + ], + "type": "inline_equation", + "content": "79.63\\%" + }, + { + "bbox": [ + 304, + 402, + 545, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 402, + 545, + 474 + ], + "type": "inline_equation", + "content": "79.26\\%" + }, + { + "bbox": [ + 304, + 402, + 545, + 474 + ], + "type": "text", + "content": ", respectively, is only " + }, + { + "bbox": [ + 304, + 402, + 545, + 474 + ], + "type": "inline_equation", + "content": "0.37\\%" + }, + { + "bbox": [ + 304, + 402, + 545, + 474 + ], + "type": "text", + "content": ". Therefore, the proposed TCP is relatively insensitive to prompt length " + }, + { + "bbox": [ + 304, + 402, + 545, + 474 + ], + "type": "inline_equation", + "content": "(M)" + }, + { + "bbox": [ + 304, + 402, + 545, + 474 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 474, + 546, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 546, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 546, + 641 + ], + "type": "text", + "content": "Effect of different templates. Unlike most CoOp-based methods using the handcrafted template \"a photo of a " + }, + { + "bbox": [ + 304, + 474, + 546, + 641 + ], + "type": "inline_equation", + "content": "\\{\\}" + }, + { + "bbox": [ + 304, + 474, + 546, + 641 + ], + "type": "text", + "content": "\", the proposed TCP utilizes the random initialization template (\"X XXX " + }, + { + "bbox": [ + 304, + 474, + 546, + 641 + ], + "type": "inline_equation", + "content": "\\{\\}" + }, + { + "bbox": [ + 304, + 474, + 546, + 641 + ], + "type": "text", + "content": "\"). Therefore, we conducted a comparison among different templates. Table 6 shows that using the random template performs better than the handcrafted template. This is because TCP uses the class-ware prompt without considering the domain-share prompt, which has achieved better performance. For instance, in Table 5, TCP with the class-aware prompt achieves comparable performance of " + }, + { + "bbox": [ + 304, + 474, + 546, + 641 + ], + "type": "inline_equation", + "content": "84.05\\% /75.18\\% /79.36\\%" + }, + { + "bbox": [ + 304, + 474, + 546, + 641 + ], + "type": "text", + "content": " in terms of Base/New/H compared to existing methods. Moreover, the domain-shared prompt initialized as the random template provides complementary knowledge to the class-aware prompt." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": "Effect of " + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "D_{mid}" + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": " in TAK. We analyze the effect of the dimension " + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "D_{mid}" + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": " in TKE, and show the comparison in Figure 4. Setting " + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "D_{mid} = 128" + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": " obtains the best performance, and a smaller " + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "D_{mid}" + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": " would get a worse performance." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "Insert which layer? TCP inserts the class-aware textual prompts into a particular layer of the Text Encoder. The im-" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23444" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 91, + 241, + 142 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 285, + 82 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 285, + 82 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 285, + 82 + ], + "type": "text", + "content": "Table 7. Comparison with single-layer vs multiple layers prompt tuning." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 91, + 241, + 142 + ], + "lines": [ + { + "bbox": [ + 93, + 91, + 241, + 142 + ], + "spans": [ + { + "bbox": [ + 93, + 91, + 241, + 142 + ], + "type": "table", + "html": "
ModesLayersBaseNewH
TCP-Shallow{8}84.1375.3679.51
TCP-Deeper{4:8}84.0275.1379.33
TCP-Deeper{8:10}84.2475.2979.51
TCP-Deeper{4:8;10}84.1174.5179.02
TCP-Deeper{4;6;8;10}84.1774.6679.13
", + "image_path": "43c447dd9ee37ff8c7435957d196b31c260aa323c171c29401d2c4916fecae84.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 74, + 148, + 262, + 253 + ], + "blocks": [ + { + "bbox": [ + 74, + 148, + 262, + 253 + ], + "lines": [ + { + "bbox": [ + 74, + 148, + 262, + 253 + ], + "spans": [ + { + "bbox": [ + 74, + 148, + 262, + 253 + ], + "type": "image", + "image_path": "5b606b15fc1b82e7e6bfe74f92517ac8a56f790938ab53d20e244416c7b4698a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 262, + 279, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 279, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 279, + 274 + ], + "type": "text", + "content": "Figure 6. Comparison with TCP and Adapter-based methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "text", + "content": "pact of inserting prompts into different layers was analyzed and shown in Figure 5. The best performance was achieved when the class-aware prompt tuning was done on the 8th layer. This is because doing so on earlier layers (" + }, + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "inline_equation", + "content": "l < 8" + }, + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "text", + "content": ") does not transfer class-aware knowledge to the textual embedding, and doing it on later layers can easily affect the textual embedding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "type": "text", + "content": "Single-layer vs Multiple-layers? The results of existing MaPLe and PromptSRC show that deep prompt learning applied on multiple layers can perform better than shallow prompt tuning with a single layer. Therefore, we compared the performance of shallow prompt tuning with deep prompt tuning. The results in Table 7 indicate that the shallow class-aware prompt tuning with a single layer provides better performance. For instance, TCP-Shallow with 8-layer achieves the highest " + }, + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "type": "inline_equation", + "content": "79.51\\%" + }, + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "type": "text", + "content": ", which surpasses all variants of TCP-deeper that have two, three, and four layers. Additionally, we observed that TCP-shallow obtains the best performance of " + }, + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "type": "inline_equation", + "content": "75.36\\%" + }, + { + "bbox": [ + 46, + 365, + 287, + 520 + ], + "type": "text", + "content": " for the New class, and using more layers results in lower performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": "Compared to Adapter-based methods: In TCP, TKE acts as a special adapter, leading to the proposed TCP is similar to Adapter-based methods. We compared TCP with several Adapter-based methods such as CLIP-Adapter and CoOp-Adapter, which apply the adapter on the general textual-level class embeddings. However, TCP applies TKE to transfer the textual-level class embedding into the class-aware prompt combined with the class tokens in the middle layer of the Text Encoder. As shown in Figure 6, TCP with TKE performs better than the existing CLIP-Adapter and CoOp-Adapter on all 11 datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Training time analysis: We further analyze the training time of the proposed TCP. The training time was determined based on the average time to process one image from the ImageNet with a 16-shot setting. According to Table 8, the proposed TCP is an efficient method that performs better" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 91, + 561, + 122 + ], + "blocks": [ + { + "bbox": [ + 340, + 71, + 511, + 82 + ], + "lines": [ + { + "bbox": [ + 340, + 71, + 511, + 82 + ], + "spans": [ + { + "bbox": [ + 340, + 71, + 511, + 82 + ], + "type": "text", + "content": "Table 8. Training time comparison(ms/image)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 91, + 561, + 122 + ], + "lines": [ + { + "bbox": [ + 307, + 91, + 561, + 122 + ], + "spans": [ + { + "bbox": [ + 307, + 91, + 561, + 122 + ], + "type": "table", + "html": "
CoOpProGradKgCoOpPLOTRPOMaPLePromptSRCTCP
ms/image6ms22ms6ms78.8ms190ms90.7ms43.2ms6.4ms
H74.4876.1677.0077.3777.7878.5579.3179.51
", + "image_path": "b85f55d4fe68936adeaf4f208d4be2142ff248a322feb2b5cdb8da32e97c4e5a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 332, + 129, + 426, + 190 + ], + "blocks": [ + { + "bbox": [ + 332, + 129, + 426, + 190 + ], + "lines": [ + { + "bbox": [ + 332, + 129, + 426, + 190 + ], + "spans": [ + { + "bbox": [ + 332, + 129, + 426, + 190 + ], + "type": "image", + "image_path": "ea6c2b7fbe926823433b42f97e36931156db764ad71f5036cf0742e84e679fec.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 434, + 129, + 522, + 190 + ], + "blocks": [ + { + "bbox": [ + 434, + 129, + 522, + 190 + ], + "lines": [ + { + "bbox": [ + 434, + 129, + 522, + 190 + ], + "spans": [ + { + "bbox": [ + 434, + 129, + 522, + 190 + ], + "type": "image", + "image_path": "293fc5e48e6789d6ae5faf265486c20247ac898afe8395cf8d31dbd6c65db2bb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 331, + 199, + 429, + 263 + ], + "blocks": [ + { + "bbox": [ + 331, + 199, + 429, + 263 + ], + "lines": [ + { + "bbox": [ + 331, + 199, + 429, + 263 + ], + "spans": [ + { + "bbox": [ + 331, + 199, + 429, + 263 + ], + "type": "image", + "image_path": "c5a2d314927fe6616f08c6243ecd749ce33b972dd85061b13afa5ce1b5df6d08.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 271, + 535, + 283 + ], + "lines": [ + { + "bbox": [ + 315, + 271, + 535, + 283 + ], + "spans": [ + { + "bbox": [ + 315, + 271, + 535, + 283 + ], + "type": "text", + "content": "Figure 7. Visualization of the probability of CoOp and TCP." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 436, + 199, + 522, + 262 + ], + "blocks": [ + { + "bbox": [ + 436, + 199, + 522, + 262 + ], + "lines": [ + { + "bbox": [ + 436, + 199, + 522, + 262 + ], + "spans": [ + { + "bbox": [ + 436, + 199, + 522, + 262 + ], + "type": "image", + "image_path": "998afdab57dbe67dda51b7a2ffb65dc2df0565b7908f7a6191b7aedf10f38c35.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 293, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 342 + ], + "type": "text", + "content": "with less running time. This is because TCP adds an efficient TKE component for generating class-aware prompt upon CoOp and KgCoOp, resulting in slightly more time consumption than CoOp and KgCoOp." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 344, + 546, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 546, + 404 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 546, + 404 + ], + "type": "text", + "content": "Visualization. To prove that the proposed TCP can generate the discriminative textual classifier for prediction, we further visualize the prediction probability among all classes. As shown in Figure 7, the proposed TCP has an obvious inter-class classes than existing methods." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 422, + 378, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 378, + 435 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 378, + 435 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 445, + 546, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 546, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 546, + 566 + ], + "type": "text", + "content": "To improve the generalization and discriminative abilities of the learnable prompt, we introduce a Textual-based Class-aware Prompt tuning method that takes advantage of the benefits of general class-level textual knowledge. To achieve this, we propose a Textual Knowledge Embedding (TKE) that transfers the class-level textual embedding into the class-aware prompt. This is combined with pre-trained class tokens to generate task-specific textual knowledge. Several benchmarks and tasks have shown that the class-aware prompt is effective for prompt tuning." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "content": "However, the class-aware prompt in TCP heavily relies on the discriminative ability of the general textual embedding. On the other hand, a weaker textual embedding will produce a weaker textual classifier. For example, TCP does not perform well on the FGVCAircraft dataset. Therefore, in the future, we plan to explore ways to use weaker textual knowledge to obtain a discriminative textual classifier." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements: This work was supported by National Science and Technology Major Project(2021ZD0112202), National Natural Science Foundation of China (62376268,U23A20387,62036012,U21B2044,62376196), and Beijing Natural Science Foundation (4222039)." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23445" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Lorenzo Agnolucci, Alberto Baldrati, Francesco Todino, Federico Becattini, Marco Bertini, and Alberto Del Bimbo. Eco: Ensembling context optimization for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2811-2815, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 214, + 287, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 268 + ], + "type": "text", + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI, pages 446-461. Springer, 2014. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 270, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 287, + 334 + ], + "type": "text", + "content": "[4] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. PLOT: prompt learning with optimal transport for vision-language models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 1, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 335, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 390 + ], + "type": "text", + "content": "[5] Eulrang Cho, Jooyeon Kim, and Hyunwoo J Kim. Distribution-aware prompt tuning for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22004-22013, 2023. 1, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 392, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 287, + 456 + ], + "type": "text", + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014, pages 3606-3613. IEEE Computer Society, 2014. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 459, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 459, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 459, + 287, + 524 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248-255. IEEE Computer Society, 2009. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 525, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 287, + 612 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 614, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 614, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 53, + 614, + 287, + 657 + ], + "type": "text", + "content": "[9] Li Fei-Fei, Robert Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. Comput. Vis. Image Underst., 106(1):59-70, 2007. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[10] Zhe Gan, Linjie Li, Chunyuan Li, Lijuan Wang, Zicheng Liu, Jianfeng Gao, et al. Vision-language pre-training: Basics, recent advances, and future trends. Foundations and Trends® in Computer Graphics and Vision, 14(3-4):163-352, 2022. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[11] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. CoRR, abs/2110.04544, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 119, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 547, + 173 + ], + "type": "text", + "content": "[12] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "type": "text", + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15979-15988. IEEE, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 243, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 296 + ], + "type": "text", + "content": "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens., 12(7):2217-2226, 2019. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 299, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 363 + ], + "type": "text", + "content": "[15] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. Open-Review.net, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 365, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 441 + ], + "type": "text", + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "text", + "content": "[17] Baoshuo Kan, Teng Wang, Wenpeng Lu, Xiantong Zhen, Weili Guan, and Feng Zheng. Knowledge-aware prompt tuning for generalizable vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15670-15680, 2023. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "text", + "content": "[18] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 1, 3, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 557, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 622 + ], + "type": "text", + "content": "[19] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 3, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "text", + "content": "[20] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 5583-5594. PMLR, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 2013 IEEE International Conference on Computer Vision" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23446" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "text", + "content": "Workshops, ICCV Workshops 2013, Sydney, Australia, December 1-8, 2013, pages 554-561. IEEE Computer Society, 2013. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 163 + ], + "type": "text", + "content": "[22] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1401-1411, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 217 + ], + "type": "text", + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "type": "text", + "content": "[24] Juncheng Li, Minghe Gao, Longhui Wei, Siliang Tang, Wenqiao Zhang, Mengze Li, Wei Ji, Qi Tian, Tat-Seng Chua, and Yueting Zhuang. Gradient-regulated meta-prompt learning for generalizable vision-language models. arXiv preprint arXiv:2303.06571, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "text", + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 321, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 396 + ], + "type": "text", + "content": "[26] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 13-23, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "text", + "content": "[27] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5196-5205. IEEE, 2022. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 456, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 487 + ], + "type": "text", + "content": "[28] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. CoRR, abs/1306.5151, 2013. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 490, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 554 + ], + "type": "text", + "content": "[29] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Sixth Indian Conference on Computer Vision, Graphics & Image Processing, ICVGIP 2008, Bhubaneswar, India, 16-19 December 2008, pages 722-729. IEEE Computer Society, 2008. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 557, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 623 + ], + "type": "text", + "content": "[30] Maxime Aquab, Léon Bottou, Ivan Laptev, and Josef Sivic. Learning and transferring mid-level image representations using convolutional neural networks. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014, pages 1717-1724. IEEE Computer Society, 2014. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "text", + "content": "[31] Yassine Ouali, Adrian Bulat, Brais Matinez, and Georgios Tzimiropoulos. Black box few-shot adaptation for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15534-15546, 2023. 1, 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[32] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "USA, June 16-21, 2012, pages 3498-3505. IEEE Computer Society, 2012. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 182 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 184, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 249 + ], + "type": "text", + "content": "[34] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 18061-18070. IEEE, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 251, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 304 + ], + "type": "text", + "content": "[35] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 304, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 370 + ], + "type": "text", + "content": "[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 372, + 545, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 413 + ], + "type": "text", + "content": "[37] Sheng Shen, Shijia Yang, Tianjun Zhang, Bohan Zhai, Joseph E. Gonzalez, Kurt Keutzer, and Trevor Darrell. Multitask vision-language prompt tuning. CoRR, abs/2211.11720, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 415, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 480 + ], + "type": "text", + "content": "[38] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "text", + "content": "[39] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR, abs/1212.0402, 2012. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 515, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 547 + ], + "type": "text", + "content": "[40] Jingchen Sun, Jiayu Qin, Zihao Lin, and Changyou Chen. Prompt tuning based adapter for vision-language model adaption. CoRR, abs/2303.15234, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "text", + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 615, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 615, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 615, + 545, + 680 + ], + "type": "text", + "content": "[42] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In The Twenty-Third IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2010, San Francisco, CA, USA, 13-18 June 2010, pages 3485-3492. IEEE Computer Society, 2010. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[43] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23447" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 563 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "text", + "content": "[44] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 10899-10909. IEEE, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "text", + "content": "[45] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Unified vision and language prompt learning. CoRR, abs/2210.07225, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "type": "text", + "content": "[46] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12104-12113, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 263 + ], + "type": "text", + "content": "[47] Ji Zhang, Shihan Wu, Lianli Gao, Hengtao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. arXiv preprint arXiv:2309.07439, 2023. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 265, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 330 + ], + "type": "text", + "content": "[48] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of CLIP for few-shot classification. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 493-510. Springer, 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 385 + ], + "type": "text", + "content": "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 16795-16804. IEEE, 2022. 1, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 387, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 287, + 419 + ], + "type": "text", + "content": "[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. Int. J. Comput. Vis., 130(9):2337-2348, 2022. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 421, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 464 + ], + "type": "text", + "content": "[51] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 466, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 563 + ], + "type": "text", + "content": "[52] Beier Zhu, Yulei Niu, Saeil Lee, Minhoe Hur, and Hanwang Zhang. Debiased fine-tuning for vision-language models by prompt regularization. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence, IAAI 2023, Thirteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2023, Washington, DC, USA, February 7-14, 2023, pages 3834-3842. AAAI Press, 2023. 1, 3" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23448" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/63fde73a-d557-47b6-aee9-074fa8fa7370_content_list.json b/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/63fde73a-d557-47b6-aee9-074fa8fa7370_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c08ea49b38a3c749884df9d75ce52958317628f6 --- /dev/null +++ b/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/63fde73a-d557-47b6-aee9-074fa8fa7370_content_list.json @@ -0,0 +1,1444 @@ +[ + { + "type": "text", + "text": "TE-TAD: Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression", + "text_level": 1, + "bbox": [ + 148, + 127, + 821, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ho-Joong Kim", + "bbox": [ + 169, + 191, + 300, + 210 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jung-Ho Hong", + "bbox": [ + 336, + 191, + 462, + 210 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Heejo Kong²", + "bbox": [ + 500, + 191, + 606, + 210 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seong-Whan Lee $^{1*}$", + "bbox": [ + 640, + 191, + 795, + 210 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Dept. of Artificial Intelligence, Korea University, Seoul, Korea", + "bbox": [ + 272, + 214, + 696, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Dept. of Brain and Cognitive Engineering, Korea University, Seoul, Korea", + "bbox": [ + 236, + 232, + 732, + 248 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hojoong_kim, jungho-hong, hj_kong, sw.lee}@korea.ac.kr", + "bbox": [ + 241, + 250, + 720, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 290, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we investigate that the normalized coordinate expression is a key factor as reliance on handcrafted components in query-based detectors for temporal action detection (TAD). Despite significant advancements towards an end-to-end framework in object detection, query-based detectors have been limited in achieving full end-to-end modeling in TAD. To address this issue, we propose TE-TAD, a full end-to-end temporal action detection transformer that integrates time-aligned coordinate expression. We reformulate coordinate expression utilizing actual timeline values, ensuring length-invariant representations from the extremely diverse video duration environment. Furthermore, our proposed adaptive query selection dynamically adjusts the number of queries based on video length, providing a suitable solution for varying video durations compared to a fixed query set. Our approach not only simplifies the TAD process by eliminating the need for hand-crafted components but also significantly improves the performance of query-based detectors. Our TE-TAD outperforms the previous query-based detectors and achieves competitive performance compared to state-of-the-art methods on popular benchmark datasets. Code is available at: https://github.com/Dotori-HJ/TE-TAD", + "bbox": [ + 76, + 316, + 473, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 672, + 209, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Temporal action detection (TAD) plays an essential role in video understanding and its numerous real-world applications, such as video surveillance, video summarization, and video retrieval. TAD aims to recognize and localize actions within untrimmed video sequences by identifying the class labels with precise start and end times of action instances. Recently, TAD methods can be mainly divided by three approaches: anchor-based [3, 15, 16, 21, 23, 31, 34], anchor-free [6, 14, 27, 32], and query-based [18, 26, 28] detector.", + "bbox": [ + 76, + 698, + 468, + 833 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Query-based detectors, inspired by DETR [4], have attracted interest because of their potential to eliminate reliance on hand-crafted components, such as the sliding win", + "bbox": [ + 76, + 834, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5b1cb195368c23efc6302d8e872d53cc15ab5bfb5ba26733d817014f906f129e.jpg", + "image_caption": [ + "Figure 1. Performance comparison of query-based detectors across various feature coverages on THUMOS14, demonstrating how extending the feature coverage impacts detection performance, as measured by mean Average Precision (mAP@AVG). The full end-to-end setting that without coverage constraints is denoted by $\\infty$ ." + ], + "image_footnote": [], + "bbox": [ + 516, + 290, + 877, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dow and non-maximum suppression (NMS). This potential derives from adopting a set-prediction mechanism, which aims to provide an end-to-end detection process by utilizing a one-to-one matching paradigm. Despite these advantages, query-based detectors encounter significant challenges in two aspects: (1) they show decreased performance when dealing with extended temporal coverage, often making them a less favorable option compared to anchor-free detectors, and (2) due to limited extended temporal coverage, the reliance on the sliding window approach leads to redundant proposals and necessitates the use of NMS.", + "bbox": [ + 496, + 537, + 893, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To demonstrate these issues, we conduct experiments by increasing the feature coverages on existing query-based detectors. Fig. 1 demonstrates the change in performance across various feature coverage. Except for unrestricted feature coverage denoted by $\\infty$ , feature coverage is calculated from the window size of the sliding window method. As shown in the graph, even though wide feature coverage is beneficial to capturing longer context, the performance of existing query-based detectors diminishes in mean Average Precision (mAP) as feature coverage increments. This result indicates that there are significant limitations to query-based detectors in scalability to temporal length, despite efforts to address long durations [32] in TAD.", + "bbox": [ + 496, + 704, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 101, + 887, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18837", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Furthermore, Fig. 2 shows the second issue, illustrating the limitations of the sliding window approach. As shown in Fig. 2(a), the sliding window is limited to address long-range duration due to its limited window size. Moreover, the sliding window approach contains overlapping areas to prevent predictions from being truncated in the middle. These overlapping areas generate duplicate predictions, necessitating the use of NMS to filter the false positive cases. This reliance on NMS contradicts the set-prediction goal of minimizing hand-crafted components, thus hindering the achievement of a fully end-to-end TAD. To address these issues, we investigate why extended temporal coverage adversely affects the performance of query-based detectors (Sec. 3.3). Our investigation reveals that the conventional use of normalized coordinate expressions is a significant factor, disturbing the achievement of a full end-to-end TAD.", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a full end-to-end temporal action detection transformer that integrates time-aligned coordinate expression (TE-TAD), which reformulates normalized coordinate expression to actual timeline video values. Our reformulation enables the query-based detector to effectively address length-invariant modeling by avoiding the distortion of the normalizing process, which not only enhances the detection performance but also simplifies the detection process. Our TE-TAD stabilizes the training process of the query-based detector when dealing with extended videos; our approach shows significant improvements and completely removes the reliance on hand-crafted components such as the sliding window and NMS. Furthermore, we introduce an adaptive query selection that effectively addresses various video lengths, dynamically adjusting the number of queries in response to the temporal length of each video. In contrast to relying on a fixed set of queries, our TE-TAD provides a suitable approach to process diverse video lengths. Our approach shows significant improvements compared to previous query-based detectors and achieves competitive performance with state-of-the-art methods on popular benchmark datasets: THUMOS14 [10], ActivityNet v1.3 [8], and EpicKitchens [7].", + "bbox": [ + 75, + 335, + 472, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as three-fold:", + "bbox": [ + 76, + 686, + 397, + 700 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a full end-to-end temporal action detection transformer that integrates time-aligned coordinate expression (TE-TAD), which preserves the set-prediction mechanism and enables a full end-to-end modeling for TAD by eliminating the hand-crafted components.", + "- Our approach introduces a length-invariant mechanism to query-based detectors, significantly improving scalability in handling varying lengths of videos.", + "- Our TE-TAD significantly outperforms the previous query-based detectors and achieves competitive performance compared to state-of-the-art methods, even without hand-crafted components such as the sliding window and NMS." + ], + "bbox": [ + 76, + 704, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/233dc2ece837e5460e3600ca008616558e77436becdf465f0d9554c68689fc7a.jpg", + "image_caption": [ + "(a) Sliding window" + ], + "image_footnote": [], + "bbox": [ + 526, + 87, + 867, + 200 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6eeb3c774ffc851cf4718d15d445845300eea92438c41572e32ebca0560966e2.jpg", + "image_caption": [ + "(b) End-to-end setting", + "Figure 2. Comparison between sliding window and end-to-end settings. The sliding window generates redundant proposals in the duplicated area." + ], + "image_footnote": [], + "bbox": [ + 526, + 215, + 864, + 314 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 388, + 640, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Action Recognition Action Recognition is a foundational task in video understanding, categorizing video sequences into distinct action classes. Notable models include I3D [5], which enhances the inception network with 3D convolutions, and $\\mathrm{R}(2 + 1)\\mathrm{D}$ [29], separating 3D convolutions into 2D spatial and 1D temporal parts for efficient processing. TSP [2] introduces temporal channel shifting for effective temporal modeling without extra computation. VideoSwin [20], employing the SwinTransformer [19] architecture, excels in complex video data recognition. These models serve as backbones for extracting video features, directly impacting performance in subtasks like TAD.", + "bbox": [ + 496, + 414, + 890, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Anchor-based Detector Anchor-based detectors [3, 15, 16, 21, 23, 31, 34] leverage predefined anchor boxes to generate action proposals. These hand-designed anchors hinder the diverse range of action instances because of a lack of flexibility in the localization of the action instances. This approach inherently limits and necessitates additional post-processing steps to discard redundant proposals because they model the one-to-many assignment training strategy.", + "bbox": [ + 496, + 597, + 890, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Anchor-free Detector Anchor-free detectors [6, 14, 25, 27, 32, 35] offer more flexibility in action instance localization compared to anchor-based detectors by adopting an asymmetric modeling approach. For instance, ActionFormer [32] significantly enhances TAD performance by employing a transformer-based architecture that captures long-range video dependencies. TriDet [27] demonstrates superior performance in TAD by employing a trident prediction scheme and their proposed convolution-based architecture. Despite their improvements, anchor-free detectors rely on handcrafted components to remove redundant proposals using NMS because they adopt a one-to-many assignment train-", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18838", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing manner. In contrast, our approach directly addresses a one-to-one matching scheme, eliminating the use of NMS. Query-based Detector Query-based detectors, inspired by DETR [4], introduce a set-prediction mechanism, thereby reducing the reliance on hand-crafted components, which ideally prevents the need for NMS. However, existing query-based detectors still require NMS because their model design inherently breaks the one-to-one matching paradigm. RTD-Net [28] utilizes a one-to-many matching to mitigate the slow convergence issue associated with the detection transformer. This approach inherently breaks the one-to-one assignment. ReAct [26] modify the decoder's self-attention, called relational attention, only adopt self-attention between their defined relations. This partial adoption of the decoder's self-attention disturbs the set-prediction mechanism because they cannot capture the whole context of queries. Furthermore, previous query-based detectors [18, 26, 28] adopt the sliding window method that contains overlapping areas, causing redundant proposals. Furthermore, TadTR [18] deals with one-to-many matching at training loss. TadTR employs cross-window fusion (CWF), which applies NMS to overlapping areas to remove redundant proposals. In contrast, our approach entirely preserves the one-to-one matching paradigm, which enables a full end-to-end modeling.", + "bbox": [ + 76, + 90, + 472, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Approach", + "text_level": 1, + "bbox": [ + 76, + 484, + 225, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 508, + 187, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first discuss about the limitations of existing query-based detectors in TAD, focusing on the normalized coordinate expression. The normalized coordinate expression, used in existing models, causes matching instability and sensitivity, especially in extended video scenarios. Subsequently, to introduce our TE-TAD, we describe the reformulation of normalized coordinate expression to timeline coordinate expression and adaptive query proposals to ensure the length-invariant modeling. The overall architecture of TE-TAD is illustrated in Fig. 5.", + "bbox": [ + 75, + 532, + 468, + 685 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 695, + 207, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $X \\in \\mathbb{R}^{T_0 \\times C}$ denote the video feature sequence extracted by the backbone network, where $T_0$ is the temporal length of the features, and $C$ is the dimension of the video feature. Each element in the video feature sequence represented as $X = \\{x_t\\}_{t=1}^{T_0}$ , corresponds to a snippet at timestep $t$ , with each snippet comprising a few consecutive frames. These snippets are processed using a pre-trained backbone network such as I3D [5] or SlowFast [9]. Each video contains numerous action instances, and each action instance contains start and end timestamps $s$ and $e$ , along with its action class $c$ . Formally, the set of action instances in a video is represented as $A = \\{(s_n, e_n, c_n)\\}_{n=1}^N$ , where", + "bbox": [ + 75, + 718, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ee5b6860d177876d4266404ba5716c5c6b3ac6bc433f20dbdc5db9af9244223a.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 540, + 90, + 851, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d770c65a35e14d8450b047700d5a9d1044ec2d9f2c59fbd347ea0448a5b77756.jpg", + "image_caption": [ + "(b)", + "Figure 3. Comparative analysis of instability and detection performance on THUMOS14: (a) variance in instance matching quantified by IS; (b) the performance change in mAP. Feature coverage lengths are denoted by the values following the dash (-). The symbol $\\infty$ denotes an unrestricted end-to-end setting." + ], + "image_footnote": [], + "bbox": [ + 537, + 232, + 846, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$N$ is the number of action instances, and $s_n$ and $e_n$ are the start and end timestamps of an action instance, respectively, and $c_n$ is its action class. The main goal of TAD is to accurately predict the set of action instances $\\mathcal{A}$ for any given video. Previous query-based detectors [18, 26, 28] typically compute predicted values of center $\\hat{c}$ and width $\\hat{d}$ using a sigmoid function $(\\sigma)$ . Consequently, existing models decode predicted start $\\hat{s}$ and end $\\hat{e}$ timestamps by $\\sigma(\\hat{c}) - \\sigma(\\hat{d})$ and $\\sigma(\\hat{c}) + \\sigma(\\hat{d})$ to start and end timestamps, respectively.", + "bbox": [ + 496, + 457, + 890, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Exploring Key Issues", + "text_level": 1, + "bbox": [ + 500, + 604, + 697, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Matching Instability As discussed in Sec. 1, extended video lengths significantly influence the performance of existing query-based methods. To investigate this issue, we conduct a comparative study on the instability (IS) [13] and the detection performance across different feature coverage scenarios on THUMOS14. Here, IS is a quantitative measurement of the inconsistency of matching during the training process. For query-based detectors, fluctuations in matched targets compel the model to be learned from different values for the same input, leading to performance degradation. In the TadTR setting, smaller window sizes have more steps per epoch due to generating more sliding windows, whereas larger window sizes yield fewer steps per epoch. This mismatch in the number of steps results that with fewer updates per epoch, there is inherently less change to the model, which shows lower instability. For a fair comparison, we match the number of iteration steps per epoch, aligning with TadTR-34.1s (original TadTR).", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18839", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f6fe4afb96fd47c0096ae186d2b67e30bb005ac7721244494c928eae0ecb58a5.jpg", + "image_caption": [ + "Fig. 3 illustrates the instability and detection performance across diverse feature coverage. As feature coverage increases, we observe a rise in IS across the training epoch, indicating less stable instance matching, which leads to a decline in detection performance. This analysis underscores the challenge of maintaining consistent learning when with extended temporal lengths. Furthermore, a direct comparison between TadTR models and our method reveals that Ours- $\\infty$ maintains a level of stability comparable to TadTR-34.1s. This indicates that our approach significantly stabilizes the training process relative to TadTR- $\\infty$ . Moreover, our method demonstrates similar levels of matching instability yet shows significant improvements under more challenging conditions for matching problems, even when compared to models with shorter feature coverage like TadTR-34.1s and TadTR-273.1s.", + "(a) $\\sigma (\\hat{c} +\\epsilon)$", + "Figure 4. Analysis of noise tolerance on predicted value. The noise level $\\epsilon$ sampled from uniform distribution and injected before the sigmoid function." + ], + "image_footnote": [], + "bbox": [ + 81, + 348, + 272, + 468 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2de3b7dc477067254ee38879ce9dcc6e070b50402a7425e1b0ba3bdb96a06a31.jpg", + "image_caption": [ + "(b) $\\sigma (\\hat{d} +\\epsilon)$" + ], + "image_footnote": [], + "bbox": [ + 274, + 348, + 467, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sensitivity to Localize Action Instances To further explore this issue, we investigate the sensitivity of model prediction measured by adopting minor perturbations in predictions of localization. We inject a small scale of noise sampled from a uniform distribution before adopting the sigmoid function that normalizes coordinates within a [0, 1] range. Fig. 4 shows that even minimal noise injection significantly affects the prediction when applying the center prediction value. The noise $\\epsilon \\sim \\mathrm{Uniform}(-\\alpha, \\alpha)$ is sampled from the uniform distribution. As illustrated in Fig. 4(a), the model's sensitivity to small shifts is evident from the significant decline in $\\Delta \\mathrm{mAP}@\\mathrm{AVG}$ upon introducing noise to the center predicted value. Notably, with noise levels of $\\pm 0.01$ and $\\pm 0.1$ , after processing through the sigmoid function $(\\sigma)$ , the maximal shifts are confined within $\\pm 0.0025$ and $\\pm 0.025$ in the normalized coordinate space, respectively. These findings highlight that even minor output variations amplify in extended videos, leading to considerable drops in performance due to the heightened sensitivity in the normalized coordinate framework. To address this issue, we introduce a time-aligned coordinate expression that is not normalized, thereby ensuring independence from video length and reducing sensitivity.", + "bbox": [ + 76, + 551, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. TE-TAD", + "text_level": 1, + "bbox": [ + 500, + 90, + 602, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This part describes our TE-TAD for a full end-to-end TAD. We adopt the TadTR [18] architecture as a baseline method, including encoder, decoder, and temporal deformable attention architecture. Starting with the baseline, we mainly address three aspects: (1) adopting multi-scale and two-stage methods from the previous methods in object detection to bridge the performance gap between query-based and anchor-free detectors, (2) reformulating coordinate expression utilizing the actual time values to address extended video length in an end-to-end setting, and (3) proposing an adaptive query selection that dynamically adjusts the number of queries based on the diverse length of videos.", + "bbox": [ + 496, + 113, + 890, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Embedding & Multi-Scale Features We project input features $X$ using a single convolutional neural network to align them with the dimension of the transformer architecture. The projection maps the input feature $X$ to the embedded feature $Z_{1} \\in \\mathbb{R}^{D \\times T_{1}}$ , where $D$ denotes the channel dimension of the encoder and decoder transformer architecture. The temporal length $T_{1}$ of the embedded features $Z_{1}$ remains the same as the original $T_{0}$ . Subsequently, following previous approaches [27, 32], we incorporate multi-scale generation to effectively address varying lengths of actions. Unlike utilizing a transformer [32], we employ a single convolution layer with a stride of 2 to produce features at each scale level as follows:", + "bbox": [ + 496, + 295, + 892, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {l} = \\operatorname {L a y e r N o r m} _ {l} \\left(\\operatorname {C o n v} _ {l} \\left(Z _ {l - 1}\\right)\\right), l \\in \\{2, \\dots , L \\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 503, + 890, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Z_{l} \\in \\mathbb{R}^{C \\times T_{l}}$ represents the embedded features at each level $l$ , and $L$ denotes the total number of feature levels. Each subsequent level $l$ has a temporal length $T_{l}$ that is half of the temporal level of the previous level $T_{l-1}$ . The LayerNorm $_{l}$ and Conv $_{l}$ denote $l$ -th layer normalization and convolutional neural networks, respectively. We do not apply any activation function in this process to deliver the raw feature representations to the transformer detector.", + "bbox": [ + 496, + 527, + 890, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Time-Aligned Query Generation Our method follows a two-stage approach [33, 37] that generates initial action proposals using the transformer encoder. The previous two-stage approach [37] provides the reference to the encoder's outputs. The transformer encoder predicts a binary foreground score, $p^{(0)}$ , and segment offsets $\\Delta c^{(0)}$ and $\\Delta d^{(0)}$ to refine segments based on reference, for each time $t$ and level $l$ . We define reference for center $c^{\\mathrm{ref}}$ and width $d^{\\mathrm{ref}}$ predictions, aligning with the real timeline of the video. For each scale level $l$ , the reference for the center is computed as follows:", + "bbox": [ + 496, + 651, + 890, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc ^ {\\text {r e f}} = \\left\\{t \\times \\frac {f}{w \\times 2 ^ {l - 1}} + \\frac {w \\times 2 ^ {l - 1}}{2} \\right\\} _ {t = 1} ^ {T _ {l}}, l \\in \\{1, 2, \\dots , L \\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 824, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f$ denotes the frame-per-second rate of the video, and $w$ represents step size for feature extracting that indicates", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18840", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/70316b20b3e11eafe01f58796fcb637d3d5f1f21d2e5d6895a23e5164ec9bd8b.jpg", + "image_caption": [ + "Figure 5. Overview of the TE-TAD. Starting with video input, the architecture processes through a backbone for feature extraction, generating multi-scale features $Z$ . These are encoded and subsequently passed through an adaptive query selection, aligning with the video timeline for initial query generation. The decoder refines these queries layer-by-layer, culminating in the refinement of action proposals." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 901, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "how many frames to step when extracting features. The factor $2^{l - 1}$ fits the temporal lengths to the embedded features from the multi-scale generation, corresponding with the actual timeline. Subsequently, the reference for the width is computed as follows:", + "bbox": [ + 75, + 409, + 470, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd ^ {\\text {r e f}} = \\alpha \\cdot f \\times 2 ^ {l - 1}, l \\in \\{1, 2, \\dots , L \\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 492, + 468, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is the base scale for encoder proposals, which adjusts the length of the reference width length. Consequently, the time-aligned queries are decoded as follows:", + "bbox": [ + 75, + 518, + 468, + 565 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n(\\hat {c} ^ {(0)}, \\hat {d} ^ {(0)}) = (c ^ {\\mathrm {r e f}} + \\Delta c ^ {(0)} \\cdot d ^ {\\mathrm {r e f}}, \\exp (\\ln (d ^ {\\mathrm {r e f}}) + \\Delta d ^ {(0)})) (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 571, + 468, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{c}^{(0)}$ and $\\hat{d}^{(0)}$ denote the center and width of the proposals, respectively. Our approach utilizes the scaling of the center offsets with width $a^{\\mathrm{ref}}$ and exp at the proposals to address the scale-invariant approach. The decoder then refines these initial locations of queries after the proposed adaptive query selection.", + "bbox": [ + 75, + 597, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adaptive Query Selection In a conventional two-stage approach in object detection, a fixed top- $k$ selection method based on binary class predictions $p^{(0)}$ is typically employed. However, this static method may not be optimal for TAD, where the number and duration of action instances within videos vary significantly. In TAD, there's often a direct correlation between the length of a video and the number of action instances. Typically, longer videos contain more action instances, while shorter videos have fewer. This variation presents a challenge for the fixed selection method, which fails to adapt to video content characteristics.", + "bbox": [ + 75, + 688, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method divides videos into sectors of a base length $T_{\\mathrm{sector}}$ . We then perform top- $k$ selection for each sector. This individual selection allows the detector to adapt to the", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "video length, preventing the selection of only some parts when dealing with long videos. In practice, we redistribute any remaining timesteps to ensure that no sector at the end part of the video is smaller than $T_{\\mathrm{sector}}$ . The total number of sectors, $S$ , is calculated by dividing the number of the first layer's feature $T_{1}$ by $T_{\\mathrm{sector}}$ , applying a floor function. For each sector $s$ , where $s \\in \\{1, 2, \\dots, S\\}$ , we select the top- $k$ proposals from all levels based on the encoder's binary class scores within that sector. We define a subset of encoder output scores, $P_{s}$ , for each sector, and then select the top- $k$ proposals from this subset. The adaptive query selection is represented as follows:", + "bbox": [ + 496, + 409, + 893, + 589 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Q} = \\bigcup_ {s = 1} ^ {S} \\left\\{\\left(\\hat {c} _ {t, l} ^ {(0)}, \\hat {d} _ {t, l} ^ {(0)}\\right) \\mid (t, l) \\in \\text {i n d i c e s o f t o p - K i n} P _ {s} \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 595, + 890, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{Q} = \\{(\\hat{c}_q^{(0)},\\hat{d}_q^{(0)})\\}_{q = 1}^{N_q}$ is the aggregated set of selected queries. Here, $K$ is the number of queries selected from each sector, and the total number of queries $N_{q}$ equals the sum of the top- $k$ proposals across all sectors, denoted as $N_{q} = \\sum_{s = 1}^{S}K$", + "bbox": [ + 496, + 638, + 893, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Time-Aligned Segment Refinement Existing query-based models [18, 26, 28] employ the sigmoid function to express normalized coordinates from the range 0 to 1. Moreover, TadTR [18] utilizes the refining step in the decoder of the transformer using the predicted center and width of each layer. The layer-wise segment refinement step in the normalized coordinate expression is defined as $\\sigma (\\sigma^{-1}(\\hat{c}_q^{(n - 1)}) + \\Delta c_q^{(n)})$ and $\\sigma (\\sigma^{-1}(\\hat{d}_q^{(n - 1)}) + \\Delta d_q^{(n)})$ for center prediction and width prediction, respectively. This refinement step restricts the values within a [0, 1] but enables layer-by-layer updates. We reorganize the previous segment refinement step without the normalized expression.", + "bbox": [ + 496, + 718, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18841", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our approach, given selected queries $\\mathcal{Q}$ are utilized for segment refinement. Formally, segment refinement of each layer is as follows:", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {c} _ {q} ^ {(n)} = \\hat {c} _ {q} ^ {(n - 1)} + \\Delta c _ {q} ^ {(n)} \\cdot \\hat {d} _ {q} ^ {(n - 1)}, n \\in \\{1, 2, \\dots , L _ {D} \\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 143, + 468, + 162 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {d} _ {q} ^ {(n)} = \\exp \\left(\\ln \\left(\\hat {d} _ {q} ^ {(n - 1)}\\right) + \\Delta \\hat {d} _ {q} ^ {(n)}\\right), n \\in \\{1, 2, \\dots , L _ {D} \\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 193, + 468, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{c}_q^{(n)}$ and $\\hat{d}_q^{(n)}$ represent the predicted outputs of the center point and width from the decoder, respectively, of $n$ -th layer for each query. Similar to the encoder case, we utilize the scaling of the center offsets with width $d^{(n)}$ and exp function at the proposals to address the scale-invariant approach. The start $\\hat{s}^{(L_D)}$ and end $\\hat{e}^{(L_D)}$ timestamps for predictions are decoded as $\\hat{c}^{(L_D)} - \\hat{d}^{(L_D)}$ and $\\hat{c}^{(L_D)} + \\hat{d}^{(L_D)}$ , respectively. Consequently, the final predicted proposals is defined as $\\hat{\\mathcal{A}} = (\\hat{s}^{(L_D)}, \\hat{e}^{(L_D)}, \\hat{p}^{(L_D)})$ , where $L_D$ is the number of decoder layers, and $\\hat{p}^{(L_D)}$ is the prediction of confidence score of last layer of decoder action class.", + "bbox": [ + 75, + 227, + 468, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Training and Inference", + "text_level": 1, + "bbox": [ + 76, + 405, + 290, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training We follow the standard bipartite matching loss [4]. The total loss $\\mathcal{L}_{\\text{total}}$ is defined as follows:", + "bbox": [ + 76, + 428, + 468, + 459 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} (\\mathcal {A}, \\hat {\\mathcal {A}}) = \\sum_ {i = 1} ^ {N _ {q}} \\mathcal {L} _ {\\text {m a t c h}} \\left(\\mathcal {A} _ {i}, \\hat {\\mathcal {A}} _ {\\pi (i)}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 469, + 468, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{match}$ is the bipartite matching loss that incorporates both the classification probabilities and the distance between the ground truth and the predicted segments, and $\\pi$ represent permutation indices obtained by bipartite matching [12]. This cost function, $\\mathcal{L}_{match}$ , is a composite of the classification loss, and the regression loss. For the classification loss, we use the focal loss [17], which effectively addresses the class imbalance issue. For the regression loss, our model utilizes DIoU [36] and log-ratio distance for the width. DIoU evaluates the relative center distance and GIoU [24] once, while the log-ratio compares the widths relatively. Following the previous query-based approaches [4, 18, 37], we utilize the auxiliary decoding loss at every decoder layer for a more effective learning process. Furthermore, we do not apply any other losses without bipartite matching loss, such as actionness loss [18] or action classification enhancement loss [26]. More detailed descriptions of each loss are described in the supplementary materials.", + "bbox": [ + 75, + 523, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference TE-TAD introduces a significant innovation by completely removing the need for common post-processing steps, such as NMS and temporal scaling. This is a direct consequence of our model's unique ability to work with an end-to-end approach and actual timeline values of the video, streamlining the inference process. The predictions from the final layer of the decoder $\\hat{\\mathcal{A}}$ are directly used.", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 89, + 632, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 500, + 114, + 581, + 130 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets We conduct experiments on three datasets: THUMOS14 [10], ActivityNet v1.3 [8], and EpicKitchens [7]. THUMOS14 comprises 20 action classes with 200 and 213 untrimmed videos in the validation and test sets, containing 3,007 and 3,358 action instances, respectively. ActivityNet v1.3 are large-scale datasets with 200 action classes. They consist of 10,024 videos for training and 4,926 videos for validation, respectively. EpicKitchens, a first-person vision dataset, includes two sub-tasks: noun and verb. It contains 495 and 138 videos with 67,217 and 9,668 action instances for training and test, with 300 and 97 action classes for nouns and verbs, respectively. These datasets contain diverse actions and scenes, providing a rigorous evaluation setup for our method.", + "bbox": [ + 496, + 137, + 890, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric We follow the standard evaluation protocol for all datasets, utilizing mAP at different intersections over union (IoU) thresholds to evaluate TAD performance. The IoU thresholds for THUMOS14 and EpicKitchens are set at [0.3:0.7:0.1] and [0.1:0.5:0.1] respectively, while for ActivityNet v1.3, the results are reported at IoU threshold [0.5, 0.75, 0.95] with the average mAP computed at [0.5:0.95:0.05].", + "bbox": [ + 496, + 349, + 890, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details To ensure the clarity and focus of our main manuscript, detailed descriptions of the hyperparameters and experimental environments are provided in the supplementary materials.", + "bbox": [ + 496, + 469, + 890, + 530 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e42b2a9d845e7cdf36525a6b048648972c1d2979285918148efe68ccf0f40fda.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Performance comparison with state-of-the-art methods on THUMOS14. * and † denote TSN [30] and Swin-B [20] backbones, respectively. ‡ represents R(2+1)D [29]. Others employ I3D [5] backbone. The symbol △ indicates partial adoption of NMS." + ], + "table_body": "
TypeMethodNMSmAP
0.30.40.50.60.7Avg.
Anchor -basedBSN [15]*✓53.545.036.928.420.036.8
BMN [16]*✓56.047.438.829.720.538.5
BC-GNN [3]*✓57.149.140.431.223.140.2
G-TAD [31]*✓54.547.640.330.823.439.3
VSGN [34]*✓66.760.452.441.030.450.2
TCANet [23]*✓60.653.244.636.826.744.3
Anchor -freeAFSD [14]✓67.362.455.543.731.152.0
\\( MENet\\ \\left[35\\right]^‡ \\)✓70.765.358.849.134.055.6
TALLFormer [6]†✓76.0-63.2-34.559.2
ActionFormer [32]✓82.177.871.059.443.966.8
TriDet [27]✓83.680.172.962.447.469.3
Query -basedRTD-Net [28]✓68.362.351.938.823.749.0
ReAct [26]✓69.265.057.147.835.655.0
Self-DETR [11]Δ74.669.560.047.631.856.7
TadTR [18]Δ74.869.160.146.632.856.7
TE-TAD (Ours)×81.776.669.559.344.866.4
Ours w/ NMS✓83.378.471.360.745.667.9
", + "bbox": [ + 511, + 542, + 877, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 500, + 845, + 638, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "THUMOS14 Table 1 provides a comparison with the state-of-the-art methods on THUMOS14. Our TE-TAD shows a", + "bbox": [ + 496, + 869, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18842", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9cb8a551a2e907d77ca8c0681854673372c343be46426c5f4899e19ee297ef32.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodFeaturemAP
0.50.750.95Avg.
Anchor -basedBSN [15]TSN [30]46.530.08.030.0
BMN [16]TSN [30]50.134.88.333.9
BC-GNN [3]TSN [30]50.634.89.434.3
G-TAD [31]TSN [30]50.434.69.034.1
VSGN [34]TSN [30]52.436.08.435.1
TCANet [23]TSN [30]52.336.76.935.5
Anchor -freeAFSD [14]I3D [5]52.435.36.534.4
TALLFormer [6]Swin-B [20]54.136.27.935.6
ActionFormer [32]R(2+1)D [29]54.737.88.436.6
TriDet [27]R(2+1)D [29]54.738.08.436.8
MENet [35]R(2+1)D [29]54.738.410.537.7
Query -basedRTD-Net [28]TSN [30]47.230.78.630.8
ReAct [26]TSN [30]49.633.08.632.6
TadTR [18]R(2+1)D [29]53.637.510.636.8
TE-TAD (Ours)R(2+1)D [29]54.038.210.637.0
Ours w/ NMSR(2+1)D [29]54.238.110.637.1
", + "bbox": [ + 99, + 88, + 444, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "significant margin of improvement over other query-based detectors, even without NMS. While TadTR achieves an average mAP of $56.7\\%$ with partial NMS utilization through their proposed cross window fusion (CWF) denoted by $\\pmb{\\Delta}$ , our method without NMS achieves a superior performance average mAP of $66.4\\%$ . This significant improvement indicates the length-invariant capability of our TE-TAD model because THUMOS14 contains extremely diverse lengths of videos. Furthermore, even compared to anchor-free detectors, our method demonstrates competitive performance.", + "bbox": [ + 75, + 347, + 470, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ActivityNet v1.3 Following the conventional approach [18, 27, 32], the external classification score is used to evaluate ActivityNet v1.3. The pre-extracted classification scores are combined with predictions from binary detectors to obtain class labels. Table 2 presents a performance comparison of our TE-TAD with state-of-the-art approaches. While our TE-TAD method exhibits a slight improvement in query-based detectors on ActivityNet v1.3 compared to the significant gains shown in Table 1, this is reflective of the intrinsic characteristics of the dataset. ActivityNet v1.3 does not contain the diverse length of the video relative to THUMOS14. Despite this different condition, our approach still demonstrates an improvement on ActivityNet v1.3, showing performance improvement even though it does not align with the primary issues our TE-TAD aims to resolve.", + "bbox": [ + 75, + 503, + 468, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "EpicKitchens Table 3 shows the comparison with state-of-the-art methods on EpicKitchens. Our model achieves competitive performance without relying on NMS, indicating TE-TAD robustness in diverse and complex action detection scenarios. EpicKitchen contains an extremely diverse length of action instances, like THUMOS14. The results indicate the robustness of TE-TAD in handling a wide range of action lengths and complexities. The comparable performance of TE-TAD is meaningful in query-based approaches, a relatively less explored field than the more established anchor-free methods.", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7e456a09681ec079eb22afe1fa365bb4d897a00edb03635932743416ddb415c8.jpg", + "table_caption": [ + "Table 2. Comparison with state-of-the-art methods on ActivityNet v1.3." + ], + "table_footnote": [], + "table_body": "
TaskMethodNMSmAP
0.10.20.30.40.5Avg.
VerbBMN [16]✓10.88.88.47.15.68.4
G-TAD [31]✓12.111.09.48.16.59.4
ActionFormer [32]✓26.625.424.222.319.123.5
ASL [25]✓27.9-25.5-19.824.6
TriDet [27]✓28.627.426.124.220.825.4
TE-TAD (Ours)✗27.025.924.622.920.024.1
Ours w/ NMS✓27.926.825.423.420.024.7
NounBMN [16]✓10.38.36.24.53.46.5
G-TAD [31]✓11.010.08.67.05.48.4
ActionFormer [32]✓25.224.122.720.517.021.9
ASL [25]✓26.0-23.4-17.722.6
TriDet [27]✓27.426.324.622.218.323.8
TE-TAD (Ours)✗26.024.823.220.818.322.6
Ours w/ NMS✓26.325.223.221.018.222.8
", + "bbox": [ + 521, + 88, + 867, + 263 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1c41f18f8b0ee52dc200b2a1be07b89ae0b67dc6c50bf6a4af718a8fc15d95c4.jpg", + "table_caption": [ + "Table 3. Comparison with state-of-the-art methods on EpicK-itchens. All methods employ SlowFast [9] as a backbone." + ], + "table_footnote": [], + "table_body": "
TypeMethodNMSmAP@AVG
Anchor-freeActionFormer [32]X43.2 (-23.6)
✓66.8
TriDet [27]X44.9 (-24.4)
✓69.3
Query-basedReAct [26]X19.8 (-35.7)
✓55.0
TadTR [18]X53.1 (-3.6)
Δ56.7
TE-TAD (Ours)X66.4 (-1.5)
✓67.9
", + "bbox": [ + 550, + 299, + 838, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Effect of NMS on the mAP across various anchor-free and query-based methods on THUMOS14. The value in parentheses represents the decrease in mAP when NMS is not applied. The symbol $\\Delta$ indicates partial adoption of NMS.", + "bbox": [ + 498, + 457, + 890, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Further Analysis", + "text_level": 1, + "bbox": [ + 500, + 518, + 666, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of NMS In Table 4, we evaluate how NMS influences the performance of various TAD methods on THUMOS14. NMS is particularly crucial for anchor-free detectors, which employ a one-to-many assignment strategy that leads to duplicated predictions for the same instance. Furthermore, even though ReAct [26] is a query-based approach, removing NMS at the ReAct significantly affects the performance. This is why ReAct adopts partial self-attention in the decoder called relational queries, which does not address whole queries. As shown in Table 6, our approach also drops the performance without NMS when removing the decoder's self-attention layer. This result indicates that addressing whole queries by the decoder's self-attention is crucial to preserving the set-prediction mechanism and full end-to-end modeling. Furthermore, our proposed method exhibits a minimal performance decrement of only -1.5 when NMS is excluded. This indicates that our approach effectively achieves full end-to-end modeling.", + "bbox": [ + 496, + 536, + 890, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Component Contribution Analysis Table 5 shows the incremental impact of each key component in our TE-TAD on THUMOS14. The performance is measured without NMS. We conduct experiments based on full end-to-end TadTR, referred to as TadTR-∞ in Sec. 3.3. Starting with the TadTR-∞ baseline, incorporating multi-scale features and", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18843", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/05c83ac846bb691def57dea19545c55f7d2d89278f09229a7ded77674bfb713f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BaselineMulti-ScaleTwo-stage [35]TEAQSmAP@AVG
TadTR-34.1s w/ NMS✓56.7
✓56.5
✓✓57.3
57.0
TadTR-∞ w/o NMS✓40.2
✓42.6
✓✓43.3
✓✓59.5
✓✓46.1
✓✓✓63.6
✓✓✓✓66.4
", + "bbox": [ + 133, + 88, + 410, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6826bbb8fbd4a4083c9ed1f9a74cce62780a9aa5451f6ebb152ef813dce94d66.jpg", + "table_caption": [ + "Table 5. Analysis of contributions of each component on THUMOS14. The all-empty check is the denoted baseline." + ], + "table_footnote": [], + "table_body": "
EncoderDecodermAP@AVG
Self-attn.Self-attn.Cross-attn.w/o NMSw/ NMS
#1✓✓61.263.8
#2✓✓53.063.4
#3✓✓0.10.2
#4✓✓✓66.167.7
", + "bbox": [ + 138, + 232, + 403, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Analysis of the specific roles of self-attention and cross-attention layers in the encoder and decoder, and their impact with and without utilizing NMS.", + "bbox": [ + 76, + 311, + 468, + 353 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03f94a0b93f3c9b722df3ce22c0d198715641f942fa5a6576098ebe4c33eecfa.jpg", + "image_caption": [ + "Figure 6. Comparison of false negative rates by instance lengths (XS, S, M, L, XL) on THUMOS14." + ], + "image_footnote": [], + "bbox": [ + 148, + 354, + 397, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a two-stage approach shows slight improvements. Nonetheless, these adoptions of improved methods for query-based detectors do not reach the performance level of the original TadTR. However, including our time-aligned expression (TE) significantly improves performance, demonstrating the value of our time-aligned representations. Finally, incorporating the adaptive query selection (AQS) mechanism shows the highest performance.", + "bbox": [ + 75, + 477, + 467, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Role of Each Attention To clarify the role of encoder and decoder architecture, we conduct experiments for removing the attention layers of the encoder and decoder. Table 6 shows the individual contributions of the encoder and decoder layers' attention. As shown in Row #1, removing self-attention in the encoder slightly decreases the detection performance, indicating the encoder's self-attention affects representational ability. Row #2 reveals that the decoder's self-attention is the core role of the set-prediction mechanism by showing the performance degradation without using NMS. Finally, Row #3 shows that removing the decoder's cross-attention does not work because only location information is provided to the decoder, which does not capture the content information without cross-attention.", + "bbox": [ + 75, + 598, + 467, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "False Negative Analysis To further compare with the baseline method, we evaluate the false negative rate on THUMOS14. Fig. 6 shows the false negative rates across varying action instance sizes: extra small (XS), small (S), medium (M), large (L), and extra large (XL) based on DETAD [1]. These results show that even though TadTR- $\\infty$ shows the", + "bbox": [ + 75, + 810, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fc111cc3333e8fe69c840e52cf9a74c9293f2227b45374c5b96707734a7141af.jpg", + "image_caption": [ + "Figure 7. Comparison between query selection method. Ground truth (GT) is illustrated in the first row. The second row shows encoder prediction scores $p^{(0)}$ . The third and fourth row visualizes the generated queries at the top 50 points and 10 points for each sector by AQS, respectively. Different colors in the fourth row represent proposals selected by different sectors." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 890, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "worse overall performance in mAP, TadTR- $\\infty$ more captures the XL cases, indicating the shorter feature coverage cannot capture the long duration of instances. These results show that our method significantly reduces false negatives, particularly in XS and XL cases.", + "bbox": [ + 496, + 316, + 890, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effetiveness of Adaptive Query Selection Fig. 7 compares the query selection method by denoting the center point $c^{\\mathrm{ref}}$ of selected queries. Selected points both w/o AQS and w/AQS are generated by the same encoder output scores. As illustrated in Fig. 7, the fixed top- $k$ proposal method misses the fourth ground location as a foreground candidate. When a foreground proposal is too distant from actual action instances, it necessitates extensive refinement from the decoder. Refining the segments from missed candidates leads to an additional workload for the decoder layers. In contrast, our AQS method successfully captures the part missed by w/o AQS. By dividing the video into sectors and selecting queries within these local units, AQS dynamically adjusts the number of queries and more accurately detects foreground candidates.", + "bbox": [ + 496, + 392, + 892, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 625, + 617, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a full end-to-end temporal action detection transformer that integrates time-aligned coordinate expression, called TE-TAD, which eliminates reliance on hand-crafted components such as the sliding window and NMS. By aligning coordinate expression with the actual video timeline, our model not only simplifies the detection process but also significantly enhances the performance of query-based detectors. Furthermore, our TE-TAD has a length-invariant property by combining the proposed time-aligned coordinate expression and adaptive query selection, showing the potential of query-based detectors.", + "bbox": [ + 496, + 645, + 890, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2019-0-00079, Artificial Intelligence Graduate School Program (Korea University), No. 2021-0-02068, Artificial Intelligence Innovation Hub, and No. 2022-0-00984, Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation).", + "bbox": [ + 496, + 815, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18844", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Humam Alwassel, Fabian Caba Heilbron, Victor Escorchia, and Bernard Ghanem. Diagnosing error in temporal action detectors. In Proceedings of the European Conference on Computer Vision, pages 256-272, 2018. 8", + "[2] Humam Alwassel, Silvio Giancola, and Bernard Ghanem. Tsp: Temporally-sensitive pretraining of video encoders for localization tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 3173-3183, 2021. 2", + "[3] Yueran Bai, Yingying Wang, Yunhai Tong, Yang Yang, Qiyue Liu, and Junhui Liu. Boundary content graph neural network for temporal action proposal generation. In Proceedings of the European Conference on Computer Vision, pages 121-137, 2020. 1, 2, 6, 7", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision, pages 213-229, 2020. 1, 3, 6", + "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 2, 3, 6, 7", + "[6] Feng Cheng and Gedas Bertasius. Tallformer: Temporal action localization with long-memory transformer. In Proceedings of the European Conference on Computer Vision, pages 503-521, 2022. 1, 2, 6, 7", + "[7] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, and Will Price. The epic-kitchens dataset: Collection, challenges and baselines. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):4125-4141, 2020. 2, 6", + "[8] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 961-970, 2015. 2, 6", + "[9] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6202-6211, 2019. 3, 7", + "[10] Haroon Idrees, Amir R Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The thumos challenge on action recognition for videos “in the wild”. Computer Vision and Image Understanding, 155:1-23, 2017. 2, 6", + "[11] Jihwan Kim, Miso Lee, and Jae-Pil Heo. Self-feedback detr for temporal action detection. In ICCV, pages 10286-10296, 2023. 6", + "[12] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6, 1", + "[13] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 3", + "[14] Chuming Lin, Chengming Xu, Donghao Luo, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Learning salient boundary feature for anchor-free temporal action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3320-3329, 2021. 1, 2, 6, 7", + "[15] Tianwei Lin, Xu Zhao, Haisheng Su, Chongjing Wang, and Ming Yang. Bsn: Boundary sensitive network for temporal action proposal generation. In Proceedings of the European Conference on Computer Vision, pages 3-19, 2018. 1, 2, 6, 7", + "[16] Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, and Shilei Wen. Bmn: Boundary-matching network for temporal action proposal generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3889-3898, 2019. 1, 2, 6, 7", + "[17] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 2980-2988, 2017. 6, 1", + "[18] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 2022. 1, 3, 4, 5, 6, 7", + "[19] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2", + "[20] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 2, 6, 7", + "[21] Fuchen Long, Ting Yao, Zhaofan Qiu, Xinmei Tian, Jiebo Luo, and Tao Mei. Gaussian temporal awareness networks for action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 344-353, 2019. 1, 2", + "[22] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 1", + "[23] Zhiwu Qing, Haisheng Su, Weihao Gan, Dongliang Wang, Wei Wu, Xiang Wang, Yu Qiao, Junjie Yan, Changxin Gao, and Nong Sang. Temporal context aggregation network for temporal action proposal refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 485-494, 2021. 1, 2, 6, 7", + "[24] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 658-666, 2019. 6, 1", + "[25] Jiayi Shao, Xiaohan Wang, Ruijie Quan, Junjun Zheng, Jiang Yang, and Yi Yang. Action sensitivity learning for temporal" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18845", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "action localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 7", + "[26] Dingfeng Shi, Yujie Zhong, Qiong Cao, Jing Zhang, Lin Ma, Jia Li, and Dacheng Tao. React: Temporal action detection with relational queries. In Proceedings of the European Conference on Computer Vision, pages 105–121, 2022. 1, 3, 5, 6, 7", + "[27] Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, and Dacheng Tao. Tridet: Temporal action detection with relative boundary modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18857-18866, 2023. 1, 2, 4, 6, 7", + "[28] Jing Tan, Jiaqi Tang, Limin Wang, and Gangshan Wu. Relaxed transformer decoders for direct action proposal generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13526-13535, 2021. 1, 3, 5, 6, 7", + "[29] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 2, 6, 7", + "[30] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In Proceedings of the European Conference on Computer Vision, pages 20-36, 2016. 6, 7", + "[31] Mengmeng Xu, Chen Zhao, David S Rojas, Ali Thabet, and Bernard Ghanem. G-tad: Sub-graph localization for temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10156-10165, 2020. 1, 2, 6, 7", + "[32] Chenlin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In Proceedings of the European Conference on Computer Vision, pages 492-510, 2022. 1, 2, 4, 6, 7", + "[33] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In Proceedings of the European Conference on Computer Vision, 2022. 4, 1", + "[34] Chen Zhao, Ali K Thabet, and Bernard Ghanem. Video self-stitching graph network for temporal action localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13658-13667, 2021. 1, 2, 6, 7", + "[35] Zixuan Zhao, Dongqi Wang, and Xu Zhao. Movement enhancement toward multi-scale video feature representation for temporal action detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13555-13564, 2023. 2, 6, 7", + "[36] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12993-13000, 2020. 6, 1", + "[37] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "for end-to-end object detection. In International Conference on Learning Representations, 2020. 4, 6, 1", + "bbox": [ + 532, + 90, + 890, + 119 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18846", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/63fde73a-d557-47b6-aee9-074fa8fa7370_model.json b/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/63fde73a-d557-47b6-aee9-074fa8fa7370_model.json new file mode 100644 index 0000000000000000000000000000000000000000..49ad2fac1feb58f8400a1fa90f8ca5646a4535be --- /dev/null +++ b/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/63fde73a-d557-47b6-aee9-074fa8fa7370_model.json @@ -0,0 +1,1980 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.128, + 0.823, + 0.175 + ], + "angle": 0, + "content": "TE-TAD: Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.192, + 0.301, + 0.211 + ], + "angle": 0, + "content": "Ho-Joong Kim" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.193, + 0.463, + 0.211 + ], + "angle": 0, + "content": "Jung-Ho Hong" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.192, + 0.607, + 0.211 + ], + "angle": 0, + "content": "Heejo Kong²" + }, + { + "type": "text", + "bbox": [ + 0.642, + 0.192, + 0.796, + 0.211 + ], + "angle": 0, + "content": "Seong-Whan Lee\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.215, + 0.697, + 0.231 + ], + "angle": 0, + "content": "\\(^{1}\\)Dept. of Artificial Intelligence, Korea University, Seoul, Korea" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.233, + 0.733, + 0.249 + ], + "angle": 0, + "content": "\\(^{2}\\)Dept. of Brain and Cognitive Engineering, Korea University, Seoul, Korea" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.251, + 0.721, + 0.266 + ], + "angle": 0, + "content": "{hojoong_kim, jungho-hong, hj_kong, sw.lee}@korea.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.291, + 0.314, + 0.307 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.317, + 0.474, + 0.664 + ], + "angle": 0, + "content": "In this paper, we investigate that the normalized coordinate expression is a key factor as reliance on handcrafted components in query-based detectors for temporal action detection (TAD). Despite significant advancements towards an end-to-end framework in object detection, query-based detectors have been limited in achieving full end-to-end modeling in TAD. To address this issue, we propose TE-TAD, a full end-to-end temporal action detection transformer that integrates time-aligned coordinate expression. We reformulate coordinate expression utilizing actual timeline values, ensuring length-invariant representations from the extremely diverse video duration environment. Furthermore, our proposed adaptive query selection dynamically adjusts the number of queries based on video length, providing a suitable solution for varying video durations compared to a fixed query set. Our approach not only simplifies the TAD process by eliminating the need for hand-crafted components but also significantly improves the performance of query-based detectors. Our TE-TAD outperforms the previous query-based detectors and achieves competitive performance compared to state-of-the-art methods on popular benchmark datasets. Code is available at: https://github.com/Dotori-HJ/TE-TAD" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.673, + 0.21, + 0.689 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.699, + 0.47, + 0.834 + ], + "angle": 0, + "content": "Temporal action detection (TAD) plays an essential role in video understanding and its numerous real-world applications, such as video surveillance, video summarization, and video retrieval. TAD aims to recognize and localize actions within untrimmed video sequences by identifying the class labels with precise start and end times of action instances. Recently, TAD methods can be mainly divided by three approaches: anchor-based [3, 15, 16, 21, 23, 31, 34], anchor-free [6, 14, 27, 32], and query-based [18, 26, 28] detector." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.835, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Query-based detectors, inspired by DETR [4], have attracted interest because of their potential to eliminate reliance on hand-crafted components, such as the sliding win" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.291, + 0.878, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.45, + 0.895, + 0.534 + ], + "angle": 0, + "content": "Figure 1. Performance comparison of query-based detectors across various feature coverages on THUMOS14, demonstrating how extending the feature coverage impacts detection performance, as measured by mean Average Precision (mAP@AVG). The full end-to-end setting that without coverage constraints is denoted by \\(\\infty\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.704 + ], + "angle": 0, + "content": "dow and non-maximum suppression (NMS). This potential derives from adopting a set-prediction mechanism, which aims to provide an end-to-end detection process by utilizing a one-to-one matching paradigm. Despite these advantages, query-based detectors encounter significant challenges in two aspects: (1) they show decreased performance when dealing with extended temporal coverage, often making them a less favorable option compared to anchor-free detectors, and (2) due to limited extended temporal coverage, the reliance on the sliding window approach leads to redundant proposals and necessitates the use of NMS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.896, + 0.902 + ], + "angle": 0, + "content": "To demonstrate these issues, we conduct experiments by increasing the feature coverages on existing query-based detectors. Fig. 1 demonstrates the change in performance across various feature coverage. Except for unrestricted feature coverage denoted by \\(\\infty\\), feature coverage is calculated from the window size of the sliding window method. As shown in the graph, even though wide feature coverage is beneficial to capturing longer context, the performance of existing query-based detectors diminishes in mean Average Precision (mAP) as feature coverage increments. This result indicates that there are significant limitations to query-based detectors in scalability to temporal length, despite efforts to address long durations [32] in TAD." + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.227, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18837" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "Furthermore, Fig. 2 shows the second issue, illustrating the limitations of the sliding window approach. As shown in Fig. 2(a), the sliding window is limited to address long-range duration due to its limited window size. Moreover, the sliding window approach contains overlapping areas to prevent predictions from being truncated in the middle. These overlapping areas generate duplicate predictions, necessitating the use of NMS to filter the false positive cases. This reliance on NMS contradicts the set-prediction goal of minimizing hand-crafted components, thus hindering the achievement of a fully end-to-end TAD. To address these issues, we investigate why extended temporal coverage adversely affects the performance of query-based detectors (Sec. 3.3). Our investigation reveals that the conventional use of normalized coordinate expressions is a significant factor, disturbing the achievement of a full end-to-end TAD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.473, + 0.684 + ], + "angle": 0, + "content": "In this paper, we propose a full end-to-end temporal action detection transformer that integrates time-aligned coordinate expression (TE-TAD), which reformulates normalized coordinate expression to actual timeline video values. Our reformulation enables the query-based detector to effectively address length-invariant modeling by avoiding the distortion of the normalizing process, which not only enhances the detection performance but also simplifies the detection process. Our TE-TAD stabilizes the training process of the query-based detector when dealing with extended videos; our approach shows significant improvements and completely removes the reliance on hand-crafted components such as the sliding window and NMS. Furthermore, we introduce an adaptive query selection that effectively addresses various video lengths, dynamically adjusting the number of queries in response to the temporal length of each video. In contrast to relying on a fixed set of queries, our TE-TAD provides a suitable approach to process diverse video lengths. Our approach shows significant improvements compared to previous query-based detectors and achieves competitive performance with state-of-the-art methods on popular benchmark datasets: THUMOS14 [10], ActivityNet v1.3 [8], and EpicKitchens [7]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.687, + 0.398, + 0.701 + ], + "angle": 0, + "content": "Our contributions are summarized as three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.705, + 0.47, + 0.78 + ], + "angle": 0, + "content": "- We propose a full end-to-end temporal action detection transformer that integrates time-aligned coordinate expression (TE-TAD), which preserves the set-prediction mechanism and enables a full end-to-end modeling for TAD by eliminating the hand-crafted components." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.781, + 0.47, + 0.825 + ], + "angle": 0, + "content": "- Our approach introduces a length-invariant mechanism to query-based detectors, significantly improving scalability in handling varying lengths of videos." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.826, + 0.47, + 0.9 + ], + "angle": 0, + "content": "- Our TE-TAD significantly outperforms the previous query-based detectors and achieves competitive performance compared to state-of-the-art methods, even without hand-crafted components such as the sliding window and NMS." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.705, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.088, + 0.868, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.203, + 0.748, + 0.215 + ], + "angle": 0, + "content": "(a) Sliding window" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.217, + 0.866, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.318, + 0.756, + 0.33 + ], + "angle": 0, + "content": "(b) End-to-end setting" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.333, + 0.894, + 0.375 + ], + "angle": 0, + "content": "Figure 2. Comparison between sliding window and end-to-end settings. The sliding window generates redundant proposals in the duplicated area." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.389, + 0.642, + 0.404 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.596 + ], + "angle": 0, + "content": "Action Recognition Action Recognition is a foundational task in video understanding, categorizing video sequences into distinct action classes. Notable models include I3D [5], which enhances the inception network with 3D convolutions, and \\(\\mathrm{R}(2 + 1)\\mathrm{D}\\) [29], separating 3D convolutions into 2D spatial and 1D temporal parts for efficient processing. TSP [2] introduces temporal channel shifting for effective temporal modeling without extra computation. VideoSwin [20], employing the SwinTransformer [19] architecture, excels in complex video data recognition. These models serve as backbones for extracting video features, directly impacting performance in subtasks like TAD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Anchor-based Detector Anchor-based detectors [3, 15, 16, 21, 23, 31, 34] leverage predefined anchor boxes to generate action proposals. These hand-designed anchors hinder the diverse range of action instances because of a lack of flexibility in the localization of the action instances. This approach inherently limits and necessitates additional post-processing steps to discard redundant proposals because they model the one-to-many assignment training strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Anchor-free Detector Anchor-free detectors [6, 14, 25, 27, 32, 35] offer more flexibility in action instance localization compared to anchor-based detectors by adopting an asymmetric modeling approach. For instance, ActionFormer [32] significantly enhances TAD performance by employing a transformer-based architecture that captures long-range video dependencies. TriDet [27] demonstrates superior performance in TAD by employing a trident prediction scheme and their proposed convolution-based architecture. Despite their improvements, anchor-free detectors rely on handcrafted components to remove redundant proposals using NMS because they adopt a one-to-many assignment train-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18838" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.471 + ], + "angle": 0, + "content": "ing manner. In contrast, our approach directly addresses a one-to-one matching scheme, eliminating the use of NMS. Query-based Detector Query-based detectors, inspired by DETR [4], introduce a set-prediction mechanism, thereby reducing the reliance on hand-crafted components, which ideally prevents the need for NMS. However, existing query-based detectors still require NMS because their model design inherently breaks the one-to-one matching paradigm. RTD-Net [28] utilizes a one-to-many matching to mitigate the slow convergence issue associated with the detection transformer. This approach inherently breaks the one-to-one assignment. ReAct [26] modify the decoder's self-attention, called relational attention, only adopt self-attention between their defined relations. This partial adoption of the decoder's self-attention disturbs the set-prediction mechanism because they cannot capture the whole context of queries. Furthermore, previous query-based detectors [18, 26, 28] adopt the sliding window method that contains overlapping areas, causing redundant proposals. Furthermore, TadTR [18] deals with one-to-many matching at training loss. TadTR employs cross-window fusion (CWF), which applies NMS to overlapping areas to remove redundant proposals. In contrast, our approach entirely preserves the one-to-one matching paradigm, which enables a full end-to-end modeling." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.485, + 0.226, + 0.502 + ], + "angle": 0, + "content": "3. Our Approach" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.51, + 0.188, + 0.525 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.686 + ], + "angle": 0, + "content": "In this section, we first discuss about the limitations of existing query-based detectors in TAD, focusing on the normalized coordinate expression. The normalized coordinate expression, used in existing models, causes matching instability and sensitivity, especially in extended video scenarios. Subsequently, to introduce our TE-TAD, we describe the reformulation of normalized coordinate expression to timeline coordinate expression and adaptive query proposals to ensure the length-invariant modeling. The overall architecture of TE-TAD is illustrated in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.208, + 0.712 + ], + "angle": 0, + "content": "3.2. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Let \\( X \\in \\mathbb{R}^{T_0 \\times C} \\) denote the video feature sequence extracted by the backbone network, where \\( T_0 \\) is the temporal length of the features, and \\( C \\) is the dimension of the video feature. Each element in the video feature sequence represented as \\( X = \\{x_t\\}_{t=1}^{T_0} \\), corresponds to a snippet at timestep \\( t \\), with each snippet comprising a few consecutive frames. These snippets are processed using a pre-trained backbone network such as I3D [5] or SlowFast [9]. Each video contains numerous action instances, and each action instance contains start and end timestamps \\( s \\) and \\( e \\), along with its action class \\( c \\). Formally, the set of action instances in a video is represented as \\( A = \\{(s_n, e_n, c_n)\\}_{n=1}^N \\), where" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.091, + 0.852, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.219, + 0.706, + 0.23 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.233, + 0.848, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.362, + 0.704, + 0.373 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.376, + 0.892, + 0.446 + ], + "angle": 0, + "content": "Figure 3. Comparative analysis of instability and detection performance on THUMOS14: (a) variance in instance matching quantified by IS; (b) the performance change in mAP. Feature coverage lengths are denoted by the values following the dash (-). The symbol \\(\\infty\\) denotes an unrestricted end-to-end setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.892, + 0.596 + ], + "angle": 0, + "content": "\\(N\\) is the number of action instances, and \\(s_n\\) and \\(e_n\\) are the start and end timestamps of an action instance, respectively, and \\(c_n\\) is its action class. The main goal of TAD is to accurately predict the set of action instances \\(\\mathcal{A}\\) for any given video. Previous query-based detectors [18, 26, 28] typically compute predicted values of center \\(\\hat{c}\\) and width \\(\\hat{d}\\) using a sigmoid function \\((\\sigma)\\). Consequently, existing models decode predicted start \\(\\hat{s}\\) and end \\(\\hat{e}\\) timestamps by \\(\\sigma(\\hat{c}) - \\sigma(\\hat{d})\\) and \\(\\sigma(\\hat{c}) + \\sigma(\\hat{d})\\) to start and end timestamps, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.699, + 0.622 + ], + "angle": 0, + "content": "3.3. Exploring Key Issues" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Matching Instability As discussed in Sec. 1, extended video lengths significantly influence the performance of existing query-based methods. To investigate this issue, we conduct a comparative study on the instability (IS) [13] and the detection performance across different feature coverage scenarios on THUMOS14. Here, IS is a quantitative measurement of the inconsistency of matching during the training process. For query-based detectors, fluctuations in matched targets compel the model to be learned from different values for the same input, leading to performance degradation. In the TadTR setting, smaller window sizes have more steps per epoch due to generating more sliding windows, whereas larger window sizes yield fewer steps per epoch. This mismatch in the number of steps results that with fewer updates per epoch, there is inherently less change to the model, which shows lower instability. For a fair comparison, we match the number of iteration steps per epoch, aligning with TadTR-34.1s (original TadTR)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18839" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.332 + ], + "angle": 0, + "content": "Fig. 3 illustrates the instability and detection performance across diverse feature coverage. As feature coverage increases, we observe a rise in IS across the training epoch, indicating less stable instance matching, which leads to a decline in detection performance. This analysis underscores the challenge of maintaining consistent learning when with extended temporal lengths. Furthermore, a direct comparison between TadTR models and our method reveals that Ours- \\(\\infty\\) maintains a level of stability comparable to TadTR-34.1s. This indicates that our approach significantly stabilizes the training process relative to TadTR- \\(\\infty\\). Moreover, our method demonstrates similar levels of matching instability yet shows significant improvements under more challenging conditions for matching problems, even when compared to models with shorter feature coverage like TadTR-34.1s and TadTR-273.1s." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.349, + 0.273, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.474, + 0.21, + 0.487 + ], + "angle": 0, + "content": "(a) \\(\\sigma (\\hat{c} +\\epsilon)\\)" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.349, + 0.468, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.473, + 0.405, + 0.487 + ], + "angle": 0, + "content": "(b) \\(\\sigma (\\hat{d} +\\epsilon)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.533 + ], + "angle": 0, + "content": "Figure 4. Analysis of noise tolerance on predicted value. The noise level \\(\\epsilon\\) sampled from uniform distribution and injected before the sigmoid function." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.553, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Sensitivity to Localize Action Instances To further explore this issue, we investigate the sensitivity of model prediction measured by adopting minor perturbations in predictions of localization. We inject a small scale of noise sampled from a uniform distribution before adopting the sigmoid function that normalizes coordinates within a [0, 1] range. Fig. 4 shows that even minimal noise injection significantly affects the prediction when applying the center prediction value. The noise \\(\\epsilon \\sim \\mathrm{Uniform}(-\\alpha, \\alpha)\\) is sampled from the uniform distribution. As illustrated in Fig. 4(a), the model's sensitivity to small shifts is evident from the significant decline in \\(\\Delta \\mathrm{mAP}@\\mathrm{AVG}\\) upon introducing noise to the center predicted value. Notably, with noise levels of \\(\\pm 0.01\\) and \\(\\pm 0.1\\), after processing through the sigmoid function \\((\\sigma)\\), the maximal shifts are confined within \\(\\pm 0.0025\\) and \\(\\pm 0.025\\) in the normalized coordinate space, respectively. These findings highlight that even minor output variations amplify in extended videos, leading to considerable drops in performance due to the heightened sensitivity in the normalized coordinate framework. To address this issue, we introduce a time-aligned coordinate expression that is not normalized, thereby ensuring independence from video length and reducing sensitivity." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.091, + 0.603, + 0.105 + ], + "angle": 0, + "content": "3.4. TE-TAD" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.295 + ], + "angle": 0, + "content": "This part describes our TE-TAD for a full end-to-end TAD. We adopt the TadTR [18] architecture as a baseline method, including encoder, decoder, and temporal deformable attention architecture. Starting with the baseline, we mainly address three aspects: (1) adopting multi-scale and two-stage methods from the previous methods in object detection to bridge the performance gap between query-based and anchor-free detectors, (2) reformulating coordinate expression utilizing the actual time values to address extended video length in an end-to-end setting, and (3) proposing an adaptive query selection that dynamically adjusts the number of queries based on the diverse length of videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.893, + 0.492 + ], + "angle": 0, + "content": "Embedding & Multi-Scale Features We project input features \\(X\\) using a single convolutional neural network to align them with the dimension of the transformer architecture. The projection maps the input feature \\(X\\) to the embedded feature \\(Z_{1} \\in \\mathbb{R}^{D \\times T_{1}}\\), where \\(D\\) denotes the channel dimension of the encoder and decoder transformer architecture. The temporal length \\(T_{1}\\) of the embedded features \\(Z_{1}\\) remains the same as the original \\(T_{0}\\). Subsequently, following previous approaches [27, 32], we incorporate multi-scale generation to effectively address varying lengths of actions. Unlike utilizing a transformer [32], we employ a single convolution layer with a stride of 2 to produce features at each scale level as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.504, + 0.891, + 0.52 + ], + "angle": 0, + "content": "\\[\nZ _ {l} = \\operatorname {L a y e r N o r m} _ {l} \\left(\\operatorname {C o n v} _ {l} \\left(Z _ {l - 1}\\right)\\right), l \\in \\{2, \\dots , L \\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.892, + 0.65 + ], + "angle": 0, + "content": "where \\( Z_{l} \\in \\mathbb{R}^{C \\times T_{l}} \\) represents the embedded features at each level \\( l \\), and \\( L \\) denotes the total number of feature levels. Each subsequent level \\( l \\) has a temporal length \\( T_{l} \\) that is half of the temporal level of the previous level \\( T_{l-1} \\). The LayerNorm \\( _{l} \\) and Conv \\( _{l} \\) denote \\( l \\)-th layer normalization and convolutional neural networks, respectively. We do not apply any activation function in this process to deliver the raw feature representations to the transformer detector." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.892, + 0.817 + ], + "angle": 0, + "content": "Time-Aligned Query Generation Our method follows a two-stage approach [33, 37] that generates initial action proposals using the transformer encoder. The previous two-stage approach [37] provides the reference to the encoder's outputs. The transformer encoder predicts a binary foreground score, \\( p^{(0)} \\), and segment offsets \\( \\Delta c^{(0)} \\) and \\( \\Delta d^{(0)} \\) to refine segments based on reference, for each time \\( t \\) and level \\( l \\). We define reference for center \\( c^{\\mathrm{ref}} \\) and width \\( d^{\\mathrm{ref}} \\) predictions, aligning with the real timeline of the video. For each scale level \\( l \\), the reference for the center is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.825, + 0.891, + 0.87 + ], + "angle": 0, + "content": "\\[\nc ^ {\\text {r e f}} = \\left\\{t \\times \\frac {f}{w \\times 2 ^ {l - 1}} + \\frac {w \\times 2 ^ {l - 1}}{2} \\right\\} _ {t = 1} ^ {T _ {l}}, l \\in \\{1, 2, \\dots , L \\}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( f \\) denotes the frame-per-second rate of the video, and \\( w \\) represents step size for feature extracting that indicates" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18840" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.087, + 0.902, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.354, + 0.895, + 0.398 + ], + "angle": 0, + "content": "Figure 5. Overview of the TE-TAD. Starting with video input, the architecture processes through a backbone for feature extraction, generating multi-scale features \\( Z \\). These are encoded and subsequently passed through an adaptive query selection, aligning with the video timeline for initial query generation. The decoder refines these queries layer-by-layer, culminating in the refinement of action proposals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.471, + 0.485 + ], + "angle": 0, + "content": "how many frames to step when extracting features. The factor \\(2^{l - 1}\\) fits the temporal lengths to the embedded features from the multi-scale generation, corresponding with the actual timeline. Subsequently, the reference for the width is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.493, + 0.47, + 0.512 + ], + "angle": 0, + "content": "\\[\nd ^ {\\text {r e f}} = \\alpha \\cdot f \\times 2 ^ {l - 1}, l \\in \\{1, 2, \\dots , L \\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.566 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the base scale for encoder proposals, which adjusts the length of the reference width length. Consequently, the time-aligned queries are decoded as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.572, + 0.47, + 0.593 + ], + "angle": 0, + "content": "\\[\n(\\hat {c} ^ {(0)}, \\hat {d} ^ {(0)}) = (c ^ {\\mathrm {r e f}} + \\Delta c ^ {(0)} \\cdot d ^ {\\mathrm {r e f}}, \\exp (\\ln (d ^ {\\mathrm {r e f}}) + \\Delta d ^ {(0)})) (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.689 + ], + "angle": 0, + "content": "where \\(\\hat{c}^{(0)}\\) and \\(\\hat{d}^{(0)}\\) denote the center and width of the proposals, respectively. Our approach utilizes the scaling of the center offsets with width \\(a^{\\mathrm{ref}}\\) and exp at the proposals to address the scale-invariant approach. The decoder then refines these initial locations of queries after the proposed adaptive query selection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Adaptive Query Selection In a conventional two-stage approach in object detection, a fixed top-\\(k\\) selection method based on binary class predictions \\(p^{(0)}\\) is typically employed. However, this static method may not be optimal for TAD, where the number and duration of action instances within videos vary significantly. In TAD, there's often a direct correlation between the length of a video and the number of action instances. Typically, longer videos contain more action instances, while shorter videos have fewer. This variation presents a challenge for the fixed selection method, which fails to adapt to video content characteristics." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Our method divides videos into sectors of a base length \\( T_{\\mathrm{sector}} \\). We then perform top-\\( k \\) selection for each sector. This individual selection allows the detector to adapt to the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.895, + 0.59 + ], + "angle": 0, + "content": "video length, preventing the selection of only some parts when dealing with long videos. In practice, we redistribute any remaining timesteps to ensure that no sector at the end part of the video is smaller than \\( T_{\\mathrm{sector}} \\). The total number of sectors, \\( S \\), is calculated by dividing the number of the first layer's feature \\( T_{1} \\) by \\( T_{\\mathrm{sector}} \\), applying a floor function. For each sector \\( s \\), where \\( s \\in \\{1, 2, \\dots, S\\} \\), we select the top- \\( k \\) proposals from all levels based on the encoder's binary class scores within that sector. We define a subset of encoder output scores, \\( P_{s} \\), for each sector, and then select the top- \\( k \\) proposals from this subset. The adaptive query selection is represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.596, + 0.892, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Q} = \\bigcup_ {s = 1} ^ {S} \\left\\{\\left(\\hat {c} _ {t, l} ^ {(0)}, \\hat {d} _ {t, l} ^ {(0)}\\right) \\mid (t, l) \\in \\text {i n d i c e s o f t o p - K i n} P _ {s} \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.894, + 0.719 + ], + "angle": 0, + "content": "where \\(\\mathcal{Q} = \\{(\\hat{c}_q^{(0)},\\hat{d}_q^{(0)})\\}_{q = 1}^{N_q}\\) is the aggregated set of selected queries. Here, \\(K\\) is the number of queries selected from each sector, and the total number of queries \\(N_{q}\\) equals the sum of the top- \\(k\\) proposals across all sectors, denoted as \\(N_{q} = \\sum_{s = 1}^{S}K\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Time-Aligned Segment Refinement Existing query-based models [18, 26, 28] employ the sigmoid function to express normalized coordinates from the range 0 to 1. Moreover, TadTR [18] utilizes the refining step in the decoder of the transformer using the predicted center and width of each layer. The layer-wise segment refinement step in the normalized coordinate expression is defined as \\(\\sigma (\\sigma^{-1}(\\hat{c}_q^{(n - 1)}) + \\Delta c_q^{(n)})\\) and \\(\\sigma (\\sigma^{-1}(\\hat{d}_q^{(n - 1)}) + \\Delta d_q^{(n)})\\) for center prediction and width prediction, respectively. This refinement step restricts the values within a [0, 1] but enables layer-by-layer updates. We reorganize the previous segment refinement step without the normalized expression." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18841" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "In our approach, given selected queries \\(\\mathcal{Q}\\) are utilized for segment refinement. Formally, segment refinement of each layer is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.144, + 0.47, + 0.163 + ], + "angle": 0, + "content": "\\[\n\\hat {c} _ {q} ^ {(n)} = \\hat {c} _ {q} ^ {(n - 1)} + \\Delta c _ {q} ^ {(n)} \\cdot \\hat {d} _ {q} ^ {(n - 1)}, n \\in \\{1, 2, \\dots , L _ {D} \\}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.194, + 0.469, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\hat {d} _ {q} ^ {(n)} = \\exp \\left(\\ln \\left(\\hat {d} _ {q} ^ {(n - 1)}\\right) + \\Delta \\hat {d} _ {q} ^ {(n)}\\right), n \\in \\{1, 2, \\dots , L _ {D} \\}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.47, + 0.397 + ], + "angle": 0, + "content": "where \\(\\hat{c}_q^{(n)}\\) and \\(\\hat{d}_q^{(n)}\\) represent the predicted outputs of the center point and width from the decoder, respectively, of \\(n\\)-th layer for each query. Similar to the encoder case, we utilize the scaling of the center offsets with width \\(d^{(n)}\\) and exp function at the proposals to address the scale-invariant approach. The start \\(\\hat{s}^{(L_D)}\\) and end \\(\\hat{e}^{(L_D)}\\) timestamps for predictions are decoded as \\(\\hat{c}^{(L_D)} - \\hat{d}^{(L_D)}\\) and \\(\\hat{c}^{(L_D)} + \\hat{d}^{(L_D)}\\), respectively. Consequently, the final predicted proposals is defined as \\(\\hat{\\mathcal{A}} = (\\hat{s}^{(L_D)}, \\hat{e}^{(L_D)}, \\hat{p}^{(L_D)})\\), where \\(L_D\\) is the number of decoder layers, and \\(\\hat{p}^{(L_D)}\\) is the prediction of confidence score of last layer of decoder action class." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.406, + 0.292, + 0.422 + ], + "angle": 0, + "content": "3.5. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.429, + 0.469, + 0.46 + ], + "angle": 0, + "content": "Training We follow the standard bipartite matching loss [4]. The total loss \\(\\mathcal{L}_{\\text{total}}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.47, + 0.469, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} (\\mathcal {A}, \\hat {\\mathcal {A}}) = \\sum_ {i = 1} ^ {N _ {q}} \\mathcal {L} _ {\\text {m a t c h}} \\left(\\mathcal {A} _ {i}, \\hat {\\mathcal {A}} _ {\\pi (i)}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.469, + 0.794 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{match}\\) is the bipartite matching loss that incorporates both the classification probabilities and the distance between the ground truth and the predicted segments, and \\(\\pi\\) represent permutation indices obtained by bipartite matching [12]. This cost function, \\(\\mathcal{L}_{match}\\), is a composite of the classification loss, and the regression loss. For the classification loss, we use the focal loss [17], which effectively addresses the class imbalance issue. For the regression loss, our model utilizes DIoU [36] and log-ratio distance for the width. DIoU evaluates the relative center distance and GIoU [24] once, while the log-ratio compares the widths relatively. Following the previous query-based approaches [4, 18, 37], we utilize the auxiliary decoding loss at every decoder layer for a more effective learning process. Furthermore, we do not apply any other losses without bipartite matching loss, such as actionness loss [18] or action classification enhancement loss [26]. More detailed descriptions of each loss are described in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Inference TE-TAD introduces a significant innovation by completely removing the need for common post-processing steps, such as NMS and temporal scaling. This is a direct consequence of our model's unique ability to work with an end-to-end approach and actual timeline values of the video, streamlining the inference process. The predictions from the final layer of the decoder \\(\\hat{\\mathcal{A}}\\) are directly used." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.633, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.582, + 0.131 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Datasets We conduct experiments on three datasets: THUMOS14 [10], ActivityNet v1.3 [8], and EpicKitchens [7]. THUMOS14 comprises 20 action classes with 200 and 213 untrimmed videos in the validation and test sets, containing 3,007 and 3,358 action instances, respectively. ActivityNet v1.3 are large-scale datasets with 200 action classes. They consist of 10,024 videos for training and 4,926 videos for validation, respectively. EpicKitchens, a first-person vision dataset, includes two sub-tasks: noun and verb. It contains 495 and 138 videos with 67,217 and 9,668 action instances for training and test, with 300 and 97 action classes for nouns and verbs, respectively. These datasets contain diverse actions and scenes, providing a rigorous evaluation setup for our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.47 + ], + "angle": 0, + "content": "Evaluation Metric We follow the standard evaluation protocol for all datasets, utilizing mAP at different intersections over union (IoU) thresholds to evaluate TAD performance. The IoU thresholds for THUMOS14 and EpicKitchens are set at [0.3:0.7:0.1] and [0.1:0.5:0.1] respectively, while for ActivityNet v1.3, the results are reported at IoU threshold [0.5, 0.75, 0.95] with the average mAP computed at [0.5:0.95:0.05]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.531 + ], + "angle": 0, + "content": "Implementation Details To ensure the clarity and focus of our main manuscript, detailed descriptions of the hyperparameters and experimental environments are provided in the supplementary materials." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.543, + 0.878, + 0.759 + ], + "angle": 0, + "content": "
TypeMethodNMSmAP
0.30.40.50.60.7Avg.
Anchor -basedBSN [15]*✓53.545.036.928.420.036.8
BMN [16]*✓56.047.438.829.720.538.5
BC-GNN [3]*✓57.149.140.431.223.140.2
G-TAD [31]*✓54.547.640.330.823.439.3
VSGN [34]*✓66.760.452.441.030.450.2
TCANet [23]*✓60.653.244.636.826.744.3
Anchor -freeAFSD [14]✓67.362.455.543.731.152.0
\\( MENet\\ \\left[35\\right]^‡ \\)✓70.765.358.849.134.055.6
TALLFormer [6]†✓76.0-63.2-34.559.2
ActionFormer [32]✓82.177.871.059.443.966.8
TriDet [27]✓83.680.172.962.447.469.3
Query -basedRTD-Net [28]✓68.362.351.938.823.749.0
ReAct [26]✓69.265.057.147.835.655.0
Self-DETR [11]Δ74.669.560.047.631.856.7
TadTR [18]Δ74.869.160.146.632.856.7
TE-TAD (Ours)×81.776.669.559.344.866.4
Ours w/ NMS✓83.378.471.360.745.667.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.831 + ], + "angle": 0, + "content": "Table 1. Performance comparison with state-of-the-art methods on THUMOS14. * and † denote TSN [30] and Swin-B [20] backbones, respectively. ‡ represents R(2+1)D [29]. Others employ I3D [5] backbone. The symbol △ indicates partial adoption of NMS." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.639, + 0.862 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "THUMOS14 Table 1 provides a comparison with the state-of-the-art methods on THUMOS14. Our TE-TAD shows a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18842" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.445, + 0.299 + ], + "angle": 0, + "content": "
TypeMethodFeaturemAP
0.50.750.95Avg.
Anchor -basedBSN [15]TSN [30]46.530.08.030.0
BMN [16]TSN [30]50.134.88.333.9
BC-GNN [3]TSN [30]50.634.89.434.3
G-TAD [31]TSN [30]50.434.69.034.1
VSGN [34]TSN [30]52.436.08.435.1
TCANet [23]TSN [30]52.336.76.935.5
Anchor -freeAFSD [14]I3D [5]52.435.36.534.4
TALLFormer [6]Swin-B [20]54.136.27.935.6
ActionFormer [32]R(2+1)D [29]54.737.88.436.6
TriDet [27]R(2+1)D [29]54.738.08.436.8
MENet [35]R(2+1)D [29]54.738.410.537.7
Query -basedRTD-Net [28]TSN [30]47.230.78.630.8
ReAct [26]TSN [30]49.633.08.632.6
TadTR [18]R(2+1)D [29]53.637.510.636.8
TE-TAD (Ours)R(2+1)D [29]54.038.210.637.0
Ours w/ NMSR(2+1)D [29]54.238.110.637.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.306, + 0.47, + 0.334 + ], + "angle": 0, + "content": "Table 2. Comparison with state-of-the-art methods on ActivityNet v1.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.471, + 0.501 + ], + "angle": 0, + "content": "significant margin of improvement over other query-based detectors, even without NMS. While TadTR achieves an average mAP of \\(56.7\\%\\) with partial NMS utilization through their proposed cross window fusion (CWF) denoted by \\(\\pmb{\\Delta}\\), our method without NMS achieves a superior performance average mAP of \\(66.4\\%\\). This significant improvement indicates the length-invariant capability of our TE-TAD model because THUMOS14 contains extremely diverse lengths of videos. Furthermore, even compared to anchor-free detectors, our method demonstrates competitive performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.732 + ], + "angle": 0, + "content": "ActivityNet v1.3 Following the conventional approach [18, 27, 32], the external classification score is used to evaluate ActivityNet v1.3. The pre-extracted classification scores are combined with predictions from binary detectors to obtain class labels. Table 2 presents a performance comparison of our TE-TAD with state-of-the-art approaches. While our TE-TAD method exhibits a slight improvement in query-based detectors on ActivityNet v1.3 compared to the significant gains shown in Table 1, this is reflective of the intrinsic characteristics of the dataset. ActivityNet v1.3 does not contain the diverse length of the video relative to THUMOS14. Despite this different condition, our approach still demonstrates an improvement on ActivityNet v1.3, showing performance improvement even though it does not align with the primary issues our TE-TAD aims to resolve." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": "EpicKitchens Table 3 shows the comparison with state-of-the-art methods on EpicKitchens. Our model achieves competitive performance without relying on NMS, indicating TE-TAD robustness in diverse and complex action detection scenarios. EpicKitchen contains an extremely diverse length of action instances, like THUMOS14. The results indicate the robustness of TE-TAD in handling a wide range of action lengths and complexities. The comparable performance of TE-TAD is meaningful in query-based approaches, a relatively less explored field than the more established anchor-free methods." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.089, + 0.868, + 0.265 + ], + "angle": 0, + "content": "
TaskMethodNMSmAP
0.10.20.30.40.5Avg.
VerbBMN [16]✓10.88.88.47.15.68.4
G-TAD [31]✓12.111.09.48.16.59.4
ActionFormer [32]✓26.625.424.222.319.123.5
ASL [25]✓27.9-25.5-19.824.6
TriDet [27]✓28.627.426.124.220.825.4
TE-TAD (Ours)✗27.025.924.622.920.024.1
Ours w/ NMS✓27.926.825.423.420.024.7
NounBMN [16]✓10.38.36.24.53.46.5
G-TAD [31]✓11.010.08.67.05.48.4
ActionFormer [32]✓25.224.122.720.517.021.9
ASL [25]✓26.0-23.4-17.722.6
TriDet [27]✓27.426.324.622.218.323.8
TE-TAD (Ours)✗26.024.823.220.818.322.6
Ours w/ NMS✓26.325.223.221.018.222.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.268, + 0.89, + 0.295 + ], + "angle": 0, + "content": "Table 3. Comparison with state-of-the-art methods on EpicK-itchens. All methods employ SlowFast [9] as a backbone." + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.3, + 0.839, + 0.453 + ], + "angle": 0, + "content": "
TypeMethodNMSmAP@AVG
Anchor-freeActionFormer [32]X43.2 (-23.6)
✓66.8
TriDet [27]X44.9 (-24.4)
✓69.3
Query-basedReAct [26]X19.8 (-35.7)
✓55.0
TadTR [18]X53.1 (-3.6)
Δ56.7
TE-TAD (Ours)X66.4 (-1.5)
✓67.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.458, + 0.892, + 0.514 + ], + "angle": 0, + "content": "Table 4. Effect of NMS on the mAP across various anchor-free and query-based methods on THUMOS14. The value in parentheses represents the decrease in mAP when NMS is not applied. The symbol \\(\\Delta\\) indicates partial adoption of NMS." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.519, + 0.667, + 0.535 + ], + "angle": 0, + "content": "4.3. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Impact of NMS In Table 4, we evaluate how NMS influences the performance of various TAD methods on THUMOS14. NMS is particularly crucial for anchor-free detectors, which employ a one-to-many assignment strategy that leads to duplicated predictions for the same instance. Furthermore, even though ReAct [26] is a query-based approach, removing NMS at the ReAct significantly affects the performance. This is why ReAct adopts partial self-attention in the decoder called relational queries, which does not address whole queries. As shown in Table 6, our approach also drops the performance without NMS when removing the decoder's self-attention layer. This result indicates that addressing whole queries by the decoder's self-attention is crucial to preserving the set-prediction mechanism and full end-to-end modeling. Furthermore, our proposed method exhibits a minimal performance decrement of only -1.5 when NMS is excluded. This indicates that our approach effectively achieves full end-to-end modeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Component Contribution Analysis Table 5 shows the incremental impact of each key component in our TE-TAD on THUMOS14. The performance is measured without NMS. We conduct experiments based on full end-to-end TadTR, referred to as TadTR-∞ in Sec. 3.3. Starting with the TadTR-∞ baseline, incorporating multi-scale features and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18843" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.134, + 0.089, + 0.411, + 0.201 + ], + "angle": 0, + "content": "
BaselineMulti-ScaleTwo-stage [35]TEAQSmAP@AVG
TadTR-34.1s w/ NMS✓56.7
✓56.5
✓✓57.3
57.0
TadTR-∞ w/o NMS✓40.2
✓42.6
✓✓43.3
✓✓59.5
✓✓46.1
✓✓✓63.6
✓✓✓✓66.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.204, + 0.47, + 0.232 + ], + "angle": 0, + "content": "Table 5. Analysis of contributions of each component on THUMOS14. The all-empty check is the denoted baseline." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.233, + 0.404, + 0.308 + ], + "angle": 0, + "content": "
EncoderDecodermAP@AVG
Self-attn.Self-attn.Cross-attn.w/o NMSw/ NMS
#1✓✓61.263.8
#2✓✓53.063.4
#3✓✓0.10.2
#4✓✓✓66.167.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.312, + 0.47, + 0.354 + ], + "angle": 0, + "content": "Table 6. Analysis of the specific roles of self-attention and cross-attention layers in the encoder and decoder, and their impact with and without utilizing NMS." + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.356, + 0.398, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.45, + 0.468, + 0.476 + ], + "angle": 0, + "content": "Figure 6. Comparison of false negative rates by instance lengths (XS, S, M, L, XL) on THUMOS14." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.468, + 0.598 + ], + "angle": 0, + "content": "a two-stage approach shows slight improvements. Nonetheless, these adoptions of improved methods for query-based detectors do not reach the performance level of the original TadTR. However, including our time-aligned expression (TE) significantly improves performance, demonstrating the value of our time-aligned representations. Finally, incorporating the adaptive query selection (AQS) mechanism shows the highest performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.468, + 0.809 + ], + "angle": 0, + "content": "Role of Each Attention To clarify the role of encoder and decoder architecture, we conduct experiments for removing the attention layers of the encoder and decoder. Table 6 shows the individual contributions of the encoder and decoder layers' attention. As shown in Row #1, removing self-attention in the encoder slightly decreases the detection performance, indicating the encoder's self-attention affects representational ability. Row #2 reveals that the decoder's self-attention is the core role of the set-prediction mechanism by showing the performance degradation without using NMS. Finally, Row #3 shows that removing the decoder's cross-attention does not work because only location information is provided to the decoder, which does not capture the content information without cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.901 + ], + "angle": 0, + "content": "False Negative Analysis To further compare with the baseline method, we evaluate the false negative rate on THUMOS14. Fig. 6 shows the false negative rates across varying action instance sizes: extra small (XS), small (S), medium (M), large (L), and extra large (XL) based on DETAD [1]. These results show that even though TadTR- \\(\\infty\\) shows the" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.891, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.229, + 0.892, + 0.311 + ], + "angle": 0, + "content": "Figure 7. Comparison between query selection method. Ground truth (GT) is illustrated in the first row. The second row shows encoder prediction scores \\( p^{(0)} \\). The third and fourth row visualizes the generated queries at the top 50 points and 10 points for each sector by AQS, respectively. Different colors in the fourth row represent proposals selected by different sectors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.892, + 0.391 + ], + "angle": 0, + "content": "worse overall performance in mAP, TadTR-\\(\\infty\\) more captures the XL cases, indicating the shorter feature coverage cannot capture the long duration of instances. These results show that our method significantly reduces false negatives, particularly in XS and XL cases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.893, + 0.619 + ], + "angle": 0, + "content": "Effetiveness of Adaptive Query Selection Fig. 7 compares the query selection method by denoting the center point \\( c^{\\mathrm{ref}} \\) of selected queries. Selected points both w/o AQS and w/AQS are generated by the same encoder output scores. As illustrated in Fig. 7, the fixed top- \\( k \\) proposal method misses the fourth ground location as a foreground candidate. When a foreground proposal is too distant from actual action instances, it necessitates extensive refinement from the decoder. Refining the segments from missed candidates leads to an additional workload for the decoder layers. In contrast, our AQS method successfully captures the part missed by w/o AQS. By dividing the video into sectors and selecting queries within these local units, AQS dynamically adjusts the number of queries and more accurately detects foreground candidates." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.626, + 0.619, + 0.64 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.892, + 0.812 + ], + "angle": 0, + "content": "In this paper, we propose a full end-to-end temporal action detection transformer that integrates time-aligned coordinate expression, called TE-TAD, which eliminates reliance on hand-crafted components such as the sliding window and NMS. By aligning coordinate expression with the actual video timeline, our model not only simplifies the detection process but also significantly enhances the performance of query-based detectors. Furthermore, our TE-TAD has a length-invariant property by combining the proposed time-aligned coordinate expression and adaptive query selection, showing the potential of query-based detectors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.816, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2019-0-00079, Artificial Intelligence Graduate School Program (Korea University), No. 2021-0-02068, Artificial Intelligence Innovation Hub, and No. 2022-0-00984, Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18844" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Humam Alwassel, Fabian Caba Heilbron, Victor Escorchia, and Bernard Ghanem. Diagnosing error in temporal action detectors. In Proceedings of the European Conference on Computer Vision, pages 256-272, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Humam Alwassel, Silvio Giancola, and Bernard Ghanem. Tsp: Temporally-sensitive pretraining of video encoders for localization tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 3173-3183, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[3] Yueran Bai, Yingying Wang, Yunhai Tong, Yang Yang, Qiyue Liu, and Junhui Liu. Boundary content graph neural network for temporal action proposal generation. In Proceedings of the European Conference on Computer Vision, pages 121-137, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision, pages 213-229, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.437 + ], + "angle": 0, + "content": "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.492 + ], + "angle": 0, + "content": "[6] Feng Cheng and Gedas Bertasius. Tallformer: Temporal action localization with long-memory transformer. In Proceedings of the European Conference on Computer Vision, pages 503-521, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.471, + 0.577 + ], + "angle": 0, + "content": "[7] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, and Will Price. The epic-kitchens dataset: Collection, challenges and baselines. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):4125-4141, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.472, + 0.647 + ], + "angle": 0, + "content": "[8] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 961-970, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.471, + 0.703 + ], + "angle": 0, + "content": "[9] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6202-6211, 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.704, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[10] Haroon Idrees, Amir R Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The thumos challenge on action recognition for videos “in the wild”. Computer Vision and Image Understanding, 155:1-23, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.775, + 0.471, + 0.814 + ], + "angle": 0, + "content": "[11] Jihwan Kim, Miso Lee, and Jae-Pil Heo. Self-feedback detr for temporal action detection. In ICCV, pages 10286-10296, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.817, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[12] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.859, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[13] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[14] Chuming Lin, Chengming Xu, Donghao Luo, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Learning salient boundary feature for anchor-free temporal action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3320-3329, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.894, + 0.274 + ], + "angle": 0, + "content": "[15] Tianwei Lin, Xu Zhao, Haisheng Su, Chongjing Wang, and Ming Yang. Bsn: Boundary sensitive network for temporal action proposal generation. In Proceedings of the European Conference on Computer Vision, pages 3-19, 2018. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[16] Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, and Shilei Wen. Bmn: Boundary-matching network for temporal action proposal generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3889-3898, 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[17] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 2980-2988, 2017. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[18] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 2022. 1, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[19] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[20] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[21] Fuchen Long, Ting Yao, Zhaofan Qiu, Xinmei Tian, Jiebo Luo, and Tao Mei. Gaussian temporal awareness networks for action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 344-353, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[22] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.894, + 0.787 + ], + "angle": 0, + "content": "[23] Zhiwu Qing, Haisheng Su, Weihao Gan, Dongliang Wang, Wei Wu, Xiang Wang, Yu Qiao, Junjie Yan, Changxin Gao, and Nong Sang. Temporal context aggregation network for temporal action proposal refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 485-494, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[24] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 658-666, 2019. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Jiayi Shao, Xiaohan Wang, Ruijie Quan, Junjun Zheng, Jiang Yang, and Yi Yang. Action sensitivity learning for temporal" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18845" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "action localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[26] Dingfeng Shi, Yujie Zhong, Qiong Cao, Jing Zhang, Lin Ma, Jia Li, and Dacheng Tao. React: Temporal action detection with relational queries. In Proceedings of the European Conference on Computer Vision, pages 105–121, 2022. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[27] Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, and Dacheng Tao. Tridet: Temporal action detection with relative boundary modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18857-18866, 2023. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[28] Jing Tan, Jiaqi Tang, Limin Wang, and Gangshan Wu. Relaxed transformer decoders for direct action proposal generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13526-13535, 2021. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.404 + ], + "angle": 0, + "content": "[29] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.475 + ], + "angle": 0, + "content": "[30] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In Proceedings of the European Conference on Computer Vision, pages 20-36, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.469, + 0.545 + ], + "angle": 0, + "content": "[31] Mengmeng Xu, Chen Zhao, David S Rojas, Ali Thabet, and Bernard Ghanem. G-tad: Sub-graph localization for temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10156-10165, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[32] Chenlin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In Proceedings of the European Conference on Computer Vision, pages 492-510, 2022. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[33] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In Proceedings of the European Conference on Computer Vision, 2022. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[34] Chen Zhao, Ali K Thabet, and Bernard Ghanem. Video self-stitching graph network for temporal action localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13658-13667, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[35] Zixuan Zhao, Dongqi Wang, and Xu Zhao. Movement enhancement toward multi-scale video feature representation for temporal action detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13555-13564, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[36] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12993-13000, 2020. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.092, + 0.892, + 0.12 + ], + "angle": 0, + "content": "for end-to-end object detection. In International Conference on Learning Representations, 2020. 4, 6, 1" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18846" + } + ] +] \ No newline at end of file diff --git a/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/63fde73a-d557-47b6-aee9-074fa8fa7370_origin.pdf b/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/63fde73a-d557-47b6-aee9-074fa8fa7370_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..78e25ad155e08380440b3ddf33900d55871aa20f --- /dev/null +++ b/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/63fde73a-d557-47b6-aee9-074fa8fa7370_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:baa2f51ed5ffe079f234f780a98e5f625ae0d614a519e92433382e9b412f31dd +size 1985797 diff --git a/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/full.md b/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e739d5a5993de1ab0a8832b1bd78436d3e8022d7 --- /dev/null +++ b/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/full.md @@ -0,0 +1,285 @@ +# TE-TAD: Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression + +Ho-Joong Kim + +Jung-Ho Hong + +Heejo Kong² + +Seong-Whan Lee $^{1*}$ + +$^{1}$ Dept. of Artificial Intelligence, Korea University, Seoul, Korea + +$^{2}$ Dept. of Brain and Cognitive Engineering, Korea University, Seoul, Korea + +{hojoong_kim, jungho-hong, hj_kong, sw.lee}@korea.ac.kr + +# Abstract + +In this paper, we investigate that the normalized coordinate expression is a key factor as reliance on handcrafted components in query-based detectors for temporal action detection (TAD). Despite significant advancements towards an end-to-end framework in object detection, query-based detectors have been limited in achieving full end-to-end modeling in TAD. To address this issue, we propose TE-TAD, a full end-to-end temporal action detection transformer that integrates time-aligned coordinate expression. We reformulate coordinate expression utilizing actual timeline values, ensuring length-invariant representations from the extremely diverse video duration environment. Furthermore, our proposed adaptive query selection dynamically adjusts the number of queries based on video length, providing a suitable solution for varying video durations compared to a fixed query set. Our approach not only simplifies the TAD process by eliminating the need for hand-crafted components but also significantly improves the performance of query-based detectors. Our TE-TAD outperforms the previous query-based detectors and achieves competitive performance compared to state-of-the-art methods on popular benchmark datasets. Code is available at: https://github.com/Dotori-HJ/TE-TAD + +# 1. Introduction + +Temporal action detection (TAD) plays an essential role in video understanding and its numerous real-world applications, such as video surveillance, video summarization, and video retrieval. TAD aims to recognize and localize actions within untrimmed video sequences by identifying the class labels with precise start and end times of action instances. Recently, TAD methods can be mainly divided by three approaches: anchor-based [3, 15, 16, 21, 23, 31, 34], anchor-free [6, 14, 27, 32], and query-based [18, 26, 28] detector. + +Query-based detectors, inspired by DETR [4], have attracted interest because of their potential to eliminate reliance on hand-crafted components, such as the sliding win + +![](images/5b1cb195368c23efc6302d8e872d53cc15ab5bfb5ba26733d817014f906f129e.jpg) +Figure 1. Performance comparison of query-based detectors across various feature coverages on THUMOS14, demonstrating how extending the feature coverage impacts detection performance, as measured by mean Average Precision (mAP@AVG). The full end-to-end setting that without coverage constraints is denoted by $\infty$ . + +dow and non-maximum suppression (NMS). This potential derives from adopting a set-prediction mechanism, which aims to provide an end-to-end detection process by utilizing a one-to-one matching paradigm. Despite these advantages, query-based detectors encounter significant challenges in two aspects: (1) they show decreased performance when dealing with extended temporal coverage, often making them a less favorable option compared to anchor-free detectors, and (2) due to limited extended temporal coverage, the reliance on the sliding window approach leads to redundant proposals and necessitates the use of NMS. + +To demonstrate these issues, we conduct experiments by increasing the feature coverages on existing query-based detectors. Fig. 1 demonstrates the change in performance across various feature coverage. Except for unrestricted feature coverage denoted by $\infty$ , feature coverage is calculated from the window size of the sliding window method. As shown in the graph, even though wide feature coverage is beneficial to capturing longer context, the performance of existing query-based detectors diminishes in mean Average Precision (mAP) as feature coverage increments. This result indicates that there are significant limitations to query-based detectors in scalability to temporal length, despite efforts to address long durations [32] in TAD. + +Furthermore, Fig. 2 shows the second issue, illustrating the limitations of the sliding window approach. As shown in Fig. 2(a), the sliding window is limited to address long-range duration due to its limited window size. Moreover, the sliding window approach contains overlapping areas to prevent predictions from being truncated in the middle. These overlapping areas generate duplicate predictions, necessitating the use of NMS to filter the false positive cases. This reliance on NMS contradicts the set-prediction goal of minimizing hand-crafted components, thus hindering the achievement of a fully end-to-end TAD. To address these issues, we investigate why extended temporal coverage adversely affects the performance of query-based detectors (Sec. 3.3). Our investigation reveals that the conventional use of normalized coordinate expressions is a significant factor, disturbing the achievement of a full end-to-end TAD. + +In this paper, we propose a full end-to-end temporal action detection transformer that integrates time-aligned coordinate expression (TE-TAD), which reformulates normalized coordinate expression to actual timeline video values. Our reformulation enables the query-based detector to effectively address length-invariant modeling by avoiding the distortion of the normalizing process, which not only enhances the detection performance but also simplifies the detection process. Our TE-TAD stabilizes the training process of the query-based detector when dealing with extended videos; our approach shows significant improvements and completely removes the reliance on hand-crafted components such as the sliding window and NMS. Furthermore, we introduce an adaptive query selection that effectively addresses various video lengths, dynamically adjusting the number of queries in response to the temporal length of each video. In contrast to relying on a fixed set of queries, our TE-TAD provides a suitable approach to process diverse video lengths. Our approach shows significant improvements compared to previous query-based detectors and achieves competitive performance with state-of-the-art methods on popular benchmark datasets: THUMOS14 [10], ActivityNet v1.3 [8], and EpicKitchens [7]. + +Our contributions are summarized as three-fold: + +- We propose a full end-to-end temporal action detection transformer that integrates time-aligned coordinate expression (TE-TAD), which preserves the set-prediction mechanism and enables a full end-to-end modeling for TAD by eliminating the hand-crafted components. +- Our approach introduces a length-invariant mechanism to query-based detectors, significantly improving scalability in handling varying lengths of videos. +- Our TE-TAD significantly outperforms the previous query-based detectors and achieves competitive performance compared to state-of-the-art methods, even without hand-crafted components such as the sliding window and NMS. + +![](images/233dc2ece837e5460e3600ca008616558e77436becdf465f0d9554c68689fc7a.jpg) +(a) Sliding window + +![](images/6eeb3c774ffc851cf4718d15d445845300eea92438c41572e32ebca0560966e2.jpg) +(b) End-to-end setting +Figure 2. Comparison between sliding window and end-to-end settings. The sliding window generates redundant proposals in the duplicated area. + +# 2. Related Work + +Action Recognition Action Recognition is a foundational task in video understanding, categorizing video sequences into distinct action classes. Notable models include I3D [5], which enhances the inception network with 3D convolutions, and $\mathrm{R}(2 + 1)\mathrm{D}$ [29], separating 3D convolutions into 2D spatial and 1D temporal parts for efficient processing. TSP [2] introduces temporal channel shifting for effective temporal modeling without extra computation. VideoSwin [20], employing the SwinTransformer [19] architecture, excels in complex video data recognition. These models serve as backbones for extracting video features, directly impacting performance in subtasks like TAD. + +Anchor-based Detector Anchor-based detectors [3, 15, 16, 21, 23, 31, 34] leverage predefined anchor boxes to generate action proposals. These hand-designed anchors hinder the diverse range of action instances because of a lack of flexibility in the localization of the action instances. This approach inherently limits and necessitates additional post-processing steps to discard redundant proposals because they model the one-to-many assignment training strategy. + +Anchor-free Detector Anchor-free detectors [6, 14, 25, 27, 32, 35] offer more flexibility in action instance localization compared to anchor-based detectors by adopting an asymmetric modeling approach. For instance, ActionFormer [32] significantly enhances TAD performance by employing a transformer-based architecture that captures long-range video dependencies. TriDet [27] demonstrates superior performance in TAD by employing a trident prediction scheme and their proposed convolution-based architecture. Despite their improvements, anchor-free detectors rely on handcrafted components to remove redundant proposals using NMS because they adopt a one-to-many assignment train- + +ing manner. In contrast, our approach directly addresses a one-to-one matching scheme, eliminating the use of NMS. Query-based Detector Query-based detectors, inspired by DETR [4], introduce a set-prediction mechanism, thereby reducing the reliance on hand-crafted components, which ideally prevents the need for NMS. However, existing query-based detectors still require NMS because their model design inherently breaks the one-to-one matching paradigm. RTD-Net [28] utilizes a one-to-many matching to mitigate the slow convergence issue associated with the detection transformer. This approach inherently breaks the one-to-one assignment. ReAct [26] modify the decoder's self-attention, called relational attention, only adopt self-attention between their defined relations. This partial adoption of the decoder's self-attention disturbs the set-prediction mechanism because they cannot capture the whole context of queries. Furthermore, previous query-based detectors [18, 26, 28] adopt the sliding window method that contains overlapping areas, causing redundant proposals. Furthermore, TadTR [18] deals with one-to-many matching at training loss. TadTR employs cross-window fusion (CWF), which applies NMS to overlapping areas to remove redundant proposals. In contrast, our approach entirely preserves the one-to-one matching paradigm, which enables a full end-to-end modeling. + +# 3. Our Approach + +# 3.1. Overview + +In this section, we first discuss about the limitations of existing query-based detectors in TAD, focusing on the normalized coordinate expression. The normalized coordinate expression, used in existing models, causes matching instability and sensitivity, especially in extended video scenarios. Subsequently, to introduce our TE-TAD, we describe the reformulation of normalized coordinate expression to timeline coordinate expression and adaptive query proposals to ensure the length-invariant modeling. The overall architecture of TE-TAD is illustrated in Fig. 5. + +# 3.2. Preliminary + +Let $X \in \mathbb{R}^{T_0 \times C}$ denote the video feature sequence extracted by the backbone network, where $T_0$ is the temporal length of the features, and $C$ is the dimension of the video feature. Each element in the video feature sequence represented as $X = \{x_t\}_{t=1}^{T_0}$ , corresponds to a snippet at timestep $t$ , with each snippet comprising a few consecutive frames. These snippets are processed using a pre-trained backbone network such as I3D [5] or SlowFast [9]. Each video contains numerous action instances, and each action instance contains start and end timestamps $s$ and $e$ , along with its action class $c$ . Formally, the set of action instances in a video is represented as $A = \{(s_n, e_n, c_n)\}_{n=1}^N$ , where + +![](images/ee5b6860d177876d4266404ba5716c5c6b3ac6bc433f20dbdc5db9af9244223a.jpg) +(a) + +![](images/d770c65a35e14d8450b047700d5a9d1044ec2d9f2c59fbd347ea0448a5b77756.jpg) +(b) +Figure 3. Comparative analysis of instability and detection performance on THUMOS14: (a) variance in instance matching quantified by IS; (b) the performance change in mAP. Feature coverage lengths are denoted by the values following the dash (-). The symbol $\infty$ denotes an unrestricted end-to-end setting. + +$N$ is the number of action instances, and $s_n$ and $e_n$ are the start and end timestamps of an action instance, respectively, and $c_n$ is its action class. The main goal of TAD is to accurately predict the set of action instances $\mathcal{A}$ for any given video. Previous query-based detectors [18, 26, 28] typically compute predicted values of center $\hat{c}$ and width $\hat{d}$ using a sigmoid function $(\sigma)$ . Consequently, existing models decode predicted start $\hat{s}$ and end $\hat{e}$ timestamps by $\sigma(\hat{c}) - \sigma(\hat{d})$ and $\sigma(\hat{c}) + \sigma(\hat{d})$ to start and end timestamps, respectively. + +# 3.3. Exploring Key Issues + +Matching Instability As discussed in Sec. 1, extended video lengths significantly influence the performance of existing query-based methods. To investigate this issue, we conduct a comparative study on the instability (IS) [13] and the detection performance across different feature coverage scenarios on THUMOS14. Here, IS is a quantitative measurement of the inconsistency of matching during the training process. For query-based detectors, fluctuations in matched targets compel the model to be learned from different values for the same input, leading to performance degradation. In the TadTR setting, smaller window sizes have more steps per epoch due to generating more sliding windows, whereas larger window sizes yield fewer steps per epoch. This mismatch in the number of steps results that with fewer updates per epoch, there is inherently less change to the model, which shows lower instability. For a fair comparison, we match the number of iteration steps per epoch, aligning with TadTR-34.1s (original TadTR). + +![](images/f6fe4afb96fd47c0096ae186d2b67e30bb005ac7721244494c928eae0ecb58a5.jpg) +Fig. 3 illustrates the instability and detection performance across diverse feature coverage. As feature coverage increases, we observe a rise in IS across the training epoch, indicating less stable instance matching, which leads to a decline in detection performance. This analysis underscores the challenge of maintaining consistent learning when with extended temporal lengths. Furthermore, a direct comparison between TadTR models and our method reveals that Ours- $\infty$ maintains a level of stability comparable to TadTR-34.1s. This indicates that our approach significantly stabilizes the training process relative to TadTR- $\infty$ . Moreover, our method demonstrates similar levels of matching instability yet shows significant improvements under more challenging conditions for matching problems, even when compared to models with shorter feature coverage like TadTR-34.1s and TadTR-273.1s. +(a) $\sigma (\hat{c} +\epsilon)$ +Figure 4. Analysis of noise tolerance on predicted value. The noise level $\epsilon$ sampled from uniform distribution and injected before the sigmoid function. + +![](images/2de3b7dc477067254ee38879ce9dcc6e070b50402a7425e1b0ba3bdb96a06a31.jpg) +(b) $\sigma (\hat{d} +\epsilon)$ + +Sensitivity to Localize Action Instances To further explore this issue, we investigate the sensitivity of model prediction measured by adopting minor perturbations in predictions of localization. We inject a small scale of noise sampled from a uniform distribution before adopting the sigmoid function that normalizes coordinates within a [0, 1] range. Fig. 4 shows that even minimal noise injection significantly affects the prediction when applying the center prediction value. The noise $\epsilon \sim \mathrm{Uniform}(-\alpha, \alpha)$ is sampled from the uniform distribution. As illustrated in Fig. 4(a), the model's sensitivity to small shifts is evident from the significant decline in $\Delta \mathrm{mAP}@\mathrm{AVG}$ upon introducing noise to the center predicted value. Notably, with noise levels of $\pm 0.01$ and $\pm 0.1$ , after processing through the sigmoid function $(\sigma)$ , the maximal shifts are confined within $\pm 0.0025$ and $\pm 0.025$ in the normalized coordinate space, respectively. These findings highlight that even minor output variations amplify in extended videos, leading to considerable drops in performance due to the heightened sensitivity in the normalized coordinate framework. To address this issue, we introduce a time-aligned coordinate expression that is not normalized, thereby ensuring independence from video length and reducing sensitivity. + +# 3.4. TE-TAD + +This part describes our TE-TAD for a full end-to-end TAD. We adopt the TadTR [18] architecture as a baseline method, including encoder, decoder, and temporal deformable attention architecture. Starting with the baseline, we mainly address three aspects: (1) adopting multi-scale and two-stage methods from the previous methods in object detection to bridge the performance gap between query-based and anchor-free detectors, (2) reformulating coordinate expression utilizing the actual time values to address extended video length in an end-to-end setting, and (3) proposing an adaptive query selection that dynamically adjusts the number of queries based on the diverse length of videos. + +Embedding & Multi-Scale Features We project input features $X$ using a single convolutional neural network to align them with the dimension of the transformer architecture. The projection maps the input feature $X$ to the embedded feature $Z_{1} \in \mathbb{R}^{D \times T_{1}}$ , where $D$ denotes the channel dimension of the encoder and decoder transformer architecture. The temporal length $T_{1}$ of the embedded features $Z_{1}$ remains the same as the original $T_{0}$ . Subsequently, following previous approaches [27, 32], we incorporate multi-scale generation to effectively address varying lengths of actions. Unlike utilizing a transformer [32], we employ a single convolution layer with a stride of 2 to produce features at each scale level as follows: + +$$ +Z _ {l} = \operatorname {L a y e r N o r m} _ {l} \left(\operatorname {C o n v} _ {l} \left(Z _ {l - 1}\right)\right), l \in \{2, \dots , L \}, \tag {1} +$$ + +where $Z_{l} \in \mathbb{R}^{C \times T_{l}}$ represents the embedded features at each level $l$ , and $L$ denotes the total number of feature levels. Each subsequent level $l$ has a temporal length $T_{l}$ that is half of the temporal level of the previous level $T_{l-1}$ . The LayerNorm $_{l}$ and Conv $_{l}$ denote $l$ -th layer normalization and convolutional neural networks, respectively. We do not apply any activation function in this process to deliver the raw feature representations to the transformer detector. + +Time-Aligned Query Generation Our method follows a two-stage approach [33, 37] that generates initial action proposals using the transformer encoder. The previous two-stage approach [37] provides the reference to the encoder's outputs. The transformer encoder predicts a binary foreground score, $p^{(0)}$ , and segment offsets $\Delta c^{(0)}$ and $\Delta d^{(0)}$ to refine segments based on reference, for each time $t$ and level $l$ . We define reference for center $c^{\mathrm{ref}}$ and width $d^{\mathrm{ref}}$ predictions, aligning with the real timeline of the video. For each scale level $l$ , the reference for the center is computed as follows: + +$$ +c ^ {\text {r e f}} = \left\{t \times \frac {f}{w \times 2 ^ {l - 1}} + \frac {w \times 2 ^ {l - 1}}{2} \right\} _ {t = 1} ^ {T _ {l}}, l \in \{1, 2, \dots , L \}, \tag {2} +$$ + +where $f$ denotes the frame-per-second rate of the video, and $w$ represents step size for feature extracting that indicates + +![](images/70316b20b3e11eafe01f58796fcb637d3d5f1f21d2e5d6895a23e5164ec9bd8b.jpg) +Figure 5. Overview of the TE-TAD. Starting with video input, the architecture processes through a backbone for feature extraction, generating multi-scale features $Z$ . These are encoded and subsequently passed through an adaptive query selection, aligning with the video timeline for initial query generation. The decoder refines these queries layer-by-layer, culminating in the refinement of action proposals. + +how many frames to step when extracting features. The factor $2^{l - 1}$ fits the temporal lengths to the embedded features from the multi-scale generation, corresponding with the actual timeline. Subsequently, the reference for the width is computed as follows: + +$$ +d ^ {\text {r e f}} = \alpha \cdot f \times 2 ^ {l - 1}, l \in \{1, 2, \dots , L \}, \tag {3} +$$ + +where $\alpha$ is the base scale for encoder proposals, which adjusts the length of the reference width length. Consequently, the time-aligned queries are decoded as follows: + +$$ +(\hat {c} ^ {(0)}, \hat {d} ^ {(0)}) = (c ^ {\mathrm {r e f}} + \Delta c ^ {(0)} \cdot d ^ {\mathrm {r e f}}, \exp (\ln (d ^ {\mathrm {r e f}}) + \Delta d ^ {(0)})) (4) +$$ + +where $\hat{c}^{(0)}$ and $\hat{d}^{(0)}$ denote the center and width of the proposals, respectively. Our approach utilizes the scaling of the center offsets with width $a^{\mathrm{ref}}$ and exp at the proposals to address the scale-invariant approach. The decoder then refines these initial locations of queries after the proposed adaptive query selection. + +Adaptive Query Selection In a conventional two-stage approach in object detection, a fixed top- $k$ selection method based on binary class predictions $p^{(0)}$ is typically employed. However, this static method may not be optimal for TAD, where the number and duration of action instances within videos vary significantly. In TAD, there's often a direct correlation between the length of a video and the number of action instances. Typically, longer videos contain more action instances, while shorter videos have fewer. This variation presents a challenge for the fixed selection method, which fails to adapt to video content characteristics. + +Our method divides videos into sectors of a base length $T_{\mathrm{sector}}$ . We then perform top- $k$ selection for each sector. This individual selection allows the detector to adapt to the + +video length, preventing the selection of only some parts when dealing with long videos. In practice, we redistribute any remaining timesteps to ensure that no sector at the end part of the video is smaller than $T_{\mathrm{sector}}$ . The total number of sectors, $S$ , is calculated by dividing the number of the first layer's feature $T_{1}$ by $T_{\mathrm{sector}}$ , applying a floor function. For each sector $s$ , where $s \in \{1, 2, \dots, S\}$ , we select the top- $k$ proposals from all levels based on the encoder's binary class scores within that sector. We define a subset of encoder output scores, $P_{s}$ , for each sector, and then select the top- $k$ proposals from this subset. The adaptive query selection is represented as follows: + +$$ +\mathcal {Q} = \bigcup_ {s = 1} ^ {S} \left\{\left(\hat {c} _ {t, l} ^ {(0)}, \hat {d} _ {t, l} ^ {(0)}\right) \mid (t, l) \in \text {i n d i c e s o f t o p - K i n} P _ {s} \right\}, \tag {5} +$$ + +where $\mathcal{Q} = \{(\hat{c}_q^{(0)},\hat{d}_q^{(0)})\}_{q = 1}^{N_q}$ is the aggregated set of selected queries. Here, $K$ is the number of queries selected from each sector, and the total number of queries $N_{q}$ equals the sum of the top- $k$ proposals across all sectors, denoted as $N_{q} = \sum_{s = 1}^{S}K$ + +Time-Aligned Segment Refinement Existing query-based models [18, 26, 28] employ the sigmoid function to express normalized coordinates from the range 0 to 1. Moreover, TadTR [18] utilizes the refining step in the decoder of the transformer using the predicted center and width of each layer. The layer-wise segment refinement step in the normalized coordinate expression is defined as $\sigma (\sigma^{-1}(\hat{c}_q^{(n - 1)}) + \Delta c_q^{(n)})$ and $\sigma (\sigma^{-1}(\hat{d}_q^{(n - 1)}) + \Delta d_q^{(n)})$ for center prediction and width prediction, respectively. This refinement step restricts the values within a [0, 1] but enables layer-by-layer updates. We reorganize the previous segment refinement step without the normalized expression. + +In our approach, given selected queries $\mathcal{Q}$ are utilized for segment refinement. Formally, segment refinement of each layer is as follows: + +$$ +\hat {c} _ {q} ^ {(n)} = \hat {c} _ {q} ^ {(n - 1)} + \Delta c _ {q} ^ {(n)} \cdot \hat {d} _ {q} ^ {(n - 1)}, n \in \{1, 2, \dots , L _ {D} \}, \tag {6} +$$ + +$$ +\hat {d} _ {q} ^ {(n)} = \exp \left(\ln \left(\hat {d} _ {q} ^ {(n - 1)}\right) + \Delta \hat {d} _ {q} ^ {(n)}\right), n \in \{1, 2, \dots , L _ {D} \}, \tag {7} +$$ + +where $\hat{c}_q^{(n)}$ and $\hat{d}_q^{(n)}$ represent the predicted outputs of the center point and width from the decoder, respectively, of $n$ -th layer for each query. Similar to the encoder case, we utilize the scaling of the center offsets with width $d^{(n)}$ and exp function at the proposals to address the scale-invariant approach. The start $\hat{s}^{(L_D)}$ and end $\hat{e}^{(L_D)}$ timestamps for predictions are decoded as $\hat{c}^{(L_D)} - \hat{d}^{(L_D)}$ and $\hat{c}^{(L_D)} + \hat{d}^{(L_D)}$ , respectively. Consequently, the final predicted proposals is defined as $\hat{\mathcal{A}} = (\hat{s}^{(L_D)}, \hat{e}^{(L_D)}, \hat{p}^{(L_D)})$ , where $L_D$ is the number of decoder layers, and $\hat{p}^{(L_D)}$ is the prediction of confidence score of last layer of decoder action class. + +# 3.5. Training and Inference + +Training We follow the standard bipartite matching loss [4]. The total loss $\mathcal{L}_{\text{total}}$ is defined as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} (\mathcal {A}, \hat {\mathcal {A}}) = \sum_ {i = 1} ^ {N _ {q}} \mathcal {L} _ {\text {m a t c h}} \left(\mathcal {A} _ {i}, \hat {\mathcal {A}} _ {\pi (i)}\right), \tag {8} +$$ + +where $\mathcal{L}_{match}$ is the bipartite matching loss that incorporates both the classification probabilities and the distance between the ground truth and the predicted segments, and $\pi$ represent permutation indices obtained by bipartite matching [12]. This cost function, $\mathcal{L}_{match}$ , is a composite of the classification loss, and the regression loss. For the classification loss, we use the focal loss [17], which effectively addresses the class imbalance issue. For the regression loss, our model utilizes DIoU [36] and log-ratio distance for the width. DIoU evaluates the relative center distance and GIoU [24] once, while the log-ratio compares the widths relatively. Following the previous query-based approaches [4, 18, 37], we utilize the auxiliary decoding loss at every decoder layer for a more effective learning process. Furthermore, we do not apply any other losses without bipartite matching loss, such as actionness loss [18] or action classification enhancement loss [26]. More detailed descriptions of each loss are described in the supplementary materials. + +Inference TE-TAD introduces a significant innovation by completely removing the need for common post-processing steps, such as NMS and temporal scaling. This is a direct consequence of our model's unique ability to work with an end-to-end approach and actual timeline values of the video, streamlining the inference process. The predictions from the final layer of the decoder $\hat{\mathcal{A}}$ are directly used. + +# 4. Experiments + +# 4.1. Setup + +Datasets We conduct experiments on three datasets: THUMOS14 [10], ActivityNet v1.3 [8], and EpicKitchens [7]. THUMOS14 comprises 20 action classes with 200 and 213 untrimmed videos in the validation and test sets, containing 3,007 and 3,358 action instances, respectively. ActivityNet v1.3 are large-scale datasets with 200 action classes. They consist of 10,024 videos for training and 4,926 videos for validation, respectively. EpicKitchens, a first-person vision dataset, includes two sub-tasks: noun and verb. It contains 495 and 138 videos with 67,217 and 9,668 action instances for training and test, with 300 and 97 action classes for nouns and verbs, respectively. These datasets contain diverse actions and scenes, providing a rigorous evaluation setup for our method. + +Evaluation Metric We follow the standard evaluation protocol for all datasets, utilizing mAP at different intersections over union (IoU) thresholds to evaluate TAD performance. The IoU thresholds for THUMOS14 and EpicKitchens are set at [0.3:0.7:0.1] and [0.1:0.5:0.1] respectively, while for ActivityNet v1.3, the results are reported at IoU threshold [0.5, 0.75, 0.95] with the average mAP computed at [0.5:0.95:0.05]. + +Implementation Details To ensure the clarity and focus of our main manuscript, detailed descriptions of the hyperparameters and experimental environments are provided in the supplementary materials. + +
TypeMethodNMSmAP
0.30.40.50.60.7Avg.
Anchor -basedBSN [15]*✓53.545.036.928.420.036.8
BMN [16]*✓56.047.438.829.720.538.5
BC-GNN [3]*✓57.149.140.431.223.140.2
G-TAD [31]*✓54.547.640.330.823.439.3
VSGN [34]*✓66.760.452.441.030.450.2
TCANet [23]*✓60.653.244.636.826.744.3
Anchor -freeAFSD [14]✓67.362.455.543.731.152.0
\( MENet\ \left[35\right]^‡ \)✓70.765.358.849.134.055.6
TALLFormer [6]†✓76.0-63.2-34.559.2
ActionFormer [32]✓82.177.871.059.443.966.8
TriDet [27]✓83.680.172.962.447.469.3
Query -basedRTD-Net [28]✓68.362.351.938.823.749.0
ReAct [26]✓69.265.057.147.835.655.0
Self-DETR [11]Δ74.669.560.047.631.856.7
TadTR [18]Δ74.869.160.146.632.856.7
TE-TAD (Ours)×81.776.669.559.344.866.4
Ours w/ NMS✓83.378.471.360.745.667.9
+ +Table 1. Performance comparison with state-of-the-art methods on THUMOS14. * and † denote TSN [30] and Swin-B [20] backbones, respectively. ‡ represents R(2+1)D [29]. Others employ I3D [5] backbone. The symbol △ indicates partial adoption of NMS. + +# 4.2. Main Results + +THUMOS14 Table 1 provides a comparison with the state-of-the-art methods on THUMOS14. Our TE-TAD shows a + +
TypeMethodFeaturemAP
0.50.750.95Avg.
Anchor -basedBSN [15]TSN [30]46.530.08.030.0
BMN [16]TSN [30]50.134.88.333.9
BC-GNN [3]TSN [30]50.634.89.434.3
G-TAD [31]TSN [30]50.434.69.034.1
VSGN [34]TSN [30]52.436.08.435.1
TCANet [23]TSN [30]52.336.76.935.5
Anchor -freeAFSD [14]I3D [5]52.435.36.534.4
TALLFormer [6]Swin-B [20]54.136.27.935.6
ActionFormer [32]R(2+1)D [29]54.737.88.436.6
TriDet [27]R(2+1)D [29]54.738.08.436.8
MENet [35]R(2+1)D [29]54.738.410.537.7
Query -basedRTD-Net [28]TSN [30]47.230.78.630.8
ReAct [26]TSN [30]49.633.08.632.6
TadTR [18]R(2+1)D [29]53.637.510.636.8
TE-TAD (Ours)R(2+1)D [29]54.038.210.637.0
Ours w/ NMSR(2+1)D [29]54.238.110.637.1
+ +significant margin of improvement over other query-based detectors, even without NMS. While TadTR achieves an average mAP of $56.7\%$ with partial NMS utilization through their proposed cross window fusion (CWF) denoted by $\pmb{\Delta}$ , our method without NMS achieves a superior performance average mAP of $66.4\%$ . This significant improvement indicates the length-invariant capability of our TE-TAD model because THUMOS14 contains extremely diverse lengths of videos. Furthermore, even compared to anchor-free detectors, our method demonstrates competitive performance. + +ActivityNet v1.3 Following the conventional approach [18, 27, 32], the external classification score is used to evaluate ActivityNet v1.3. The pre-extracted classification scores are combined with predictions from binary detectors to obtain class labels. Table 2 presents a performance comparison of our TE-TAD with state-of-the-art approaches. While our TE-TAD method exhibits a slight improvement in query-based detectors on ActivityNet v1.3 compared to the significant gains shown in Table 1, this is reflective of the intrinsic characteristics of the dataset. ActivityNet v1.3 does not contain the diverse length of the video relative to THUMOS14. Despite this different condition, our approach still demonstrates an improvement on ActivityNet v1.3, showing performance improvement even though it does not align with the primary issues our TE-TAD aims to resolve. + +EpicKitchens Table 3 shows the comparison with state-of-the-art methods on EpicKitchens. Our model achieves competitive performance without relying on NMS, indicating TE-TAD robustness in diverse and complex action detection scenarios. EpicKitchen contains an extremely diverse length of action instances, like THUMOS14. The results indicate the robustness of TE-TAD in handling a wide range of action lengths and complexities. The comparable performance of TE-TAD is meaningful in query-based approaches, a relatively less explored field than the more established anchor-free methods. + +Table 2. Comparison with state-of-the-art methods on ActivityNet v1.3. + +
TaskMethodNMSmAP
0.10.20.30.40.5Avg.
VerbBMN [16]✓10.88.88.47.15.68.4
G-TAD [31]✓12.111.09.48.16.59.4
ActionFormer [32]✓26.625.424.222.319.123.5
ASL [25]✓27.9-25.5-19.824.6
TriDet [27]✓28.627.426.124.220.825.4
TE-TAD (Ours)✗27.025.924.622.920.024.1
Ours w/ NMS✓27.926.825.423.420.024.7
NounBMN [16]✓10.38.36.24.53.46.5
G-TAD [31]✓11.010.08.67.05.48.4
ActionFormer [32]✓25.224.122.720.517.021.9
ASL [25]✓26.0-23.4-17.722.6
TriDet [27]✓27.426.324.622.218.323.8
TE-TAD (Ours)✗26.024.823.220.818.322.6
Ours w/ NMS✓26.325.223.221.018.222.8
+ +Table 3. Comparison with state-of-the-art methods on EpicK-itchens. All methods employ SlowFast [9] as a backbone. + +
TypeMethodNMSmAP@AVG
Anchor-freeActionFormer [32]X43.2 (-23.6)
✓66.8
TriDet [27]X44.9 (-24.4)
✓69.3
Query-basedReAct [26]X19.8 (-35.7)
✓55.0
TadTR [18]X53.1 (-3.6)
Δ56.7
TE-TAD (Ours)X66.4 (-1.5)
✓67.9
+ +Table 4. Effect of NMS on the mAP across various anchor-free and query-based methods on THUMOS14. The value in parentheses represents the decrease in mAP when NMS is not applied. The symbol $\Delta$ indicates partial adoption of NMS. + +# 4.3. Further Analysis + +Impact of NMS In Table 4, we evaluate how NMS influences the performance of various TAD methods on THUMOS14. NMS is particularly crucial for anchor-free detectors, which employ a one-to-many assignment strategy that leads to duplicated predictions for the same instance. Furthermore, even though ReAct [26] is a query-based approach, removing NMS at the ReAct significantly affects the performance. This is why ReAct adopts partial self-attention in the decoder called relational queries, which does not address whole queries. As shown in Table 6, our approach also drops the performance without NMS when removing the decoder's self-attention layer. This result indicates that addressing whole queries by the decoder's self-attention is crucial to preserving the set-prediction mechanism and full end-to-end modeling. Furthermore, our proposed method exhibits a minimal performance decrement of only -1.5 when NMS is excluded. This indicates that our approach effectively achieves full end-to-end modeling. + +Component Contribution Analysis Table 5 shows the incremental impact of each key component in our TE-TAD on THUMOS14. The performance is measured without NMS. We conduct experiments based on full end-to-end TadTR, referred to as TadTR-∞ in Sec. 3.3. Starting with the TadTR-∞ baseline, incorporating multi-scale features and + +
BaselineMulti-ScaleTwo-stage [35]TEAQSmAP@AVG
TadTR-34.1s w/ NMS✓56.7
✓56.5
✓✓57.3
57.0
TadTR-∞ w/o NMS✓40.2
✓42.6
✓✓43.3
✓✓59.5
✓✓46.1
✓✓✓63.6
✓✓✓✓66.4
+ +Table 5. Analysis of contributions of each component on THUMOS14. The all-empty check is the denoted baseline. + +
EncoderDecodermAP@AVG
Self-attn.Self-attn.Cross-attn.w/o NMSw/ NMS
#1✓✓61.263.8
#2✓✓53.063.4
#3✓✓0.10.2
#4✓✓✓66.167.7
+ +Table 6. Analysis of the specific roles of self-attention and cross-attention layers in the encoder and decoder, and their impact with and without utilizing NMS. + +![](images/03f94a0b93f3c9b722df3ce22c0d198715641f942fa5a6576098ebe4c33eecfa.jpg) +Figure 6. Comparison of false negative rates by instance lengths (XS, S, M, L, XL) on THUMOS14. + +a two-stage approach shows slight improvements. Nonetheless, these adoptions of improved methods for query-based detectors do not reach the performance level of the original TadTR. However, including our time-aligned expression (TE) significantly improves performance, demonstrating the value of our time-aligned representations. Finally, incorporating the adaptive query selection (AQS) mechanism shows the highest performance. + +Role of Each Attention To clarify the role of encoder and decoder architecture, we conduct experiments for removing the attention layers of the encoder and decoder. Table 6 shows the individual contributions of the encoder and decoder layers' attention. As shown in Row #1, removing self-attention in the encoder slightly decreases the detection performance, indicating the encoder's self-attention affects representational ability. Row #2 reveals that the decoder's self-attention is the core role of the set-prediction mechanism by showing the performance degradation without using NMS. Finally, Row #3 shows that removing the decoder's cross-attention does not work because only location information is provided to the decoder, which does not capture the content information without cross-attention. + +False Negative Analysis To further compare with the baseline method, we evaluate the false negative rate on THUMOS14. Fig. 6 shows the false negative rates across varying action instance sizes: extra small (XS), small (S), medium (M), large (L), and extra large (XL) based on DETAD [1]. These results show that even though TadTR- $\infty$ shows the + +![](images/fc111cc3333e8fe69c840e52cf9a74c9293f2227b45374c5b96707734a7141af.jpg) +Figure 7. Comparison between query selection method. Ground truth (GT) is illustrated in the first row. The second row shows encoder prediction scores $p^{(0)}$ . The third and fourth row visualizes the generated queries at the top 50 points and 10 points for each sector by AQS, respectively. Different colors in the fourth row represent proposals selected by different sectors. + +worse overall performance in mAP, TadTR- $\infty$ more captures the XL cases, indicating the shorter feature coverage cannot capture the long duration of instances. These results show that our method significantly reduces false negatives, particularly in XS and XL cases. + +Effetiveness of Adaptive Query Selection Fig. 7 compares the query selection method by denoting the center point $c^{\mathrm{ref}}$ of selected queries. Selected points both w/o AQS and w/AQS are generated by the same encoder output scores. As illustrated in Fig. 7, the fixed top- $k$ proposal method misses the fourth ground location as a foreground candidate. When a foreground proposal is too distant from actual action instances, it necessitates extensive refinement from the decoder. Refining the segments from missed candidates leads to an additional workload for the decoder layers. In contrast, our AQS method successfully captures the part missed by w/o AQS. By dividing the video into sectors and selecting queries within these local units, AQS dynamically adjusts the number of queries and more accurately detects foreground candidates. + +# 5. Conclusion + +In this paper, we propose a full end-to-end temporal action detection transformer that integrates time-aligned coordinate expression, called TE-TAD, which eliminates reliance on hand-crafted components such as the sliding window and NMS. By aligning coordinate expression with the actual video timeline, our model not only simplifies the detection process but also significantly enhances the performance of query-based detectors. Furthermore, our TE-TAD has a length-invariant property by combining the proposed time-aligned coordinate expression and adaptive query selection, showing the potential of query-based detectors. + +Acknowledgement This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2019-0-00079, Artificial Intelligence Graduate School Program (Korea University), No. 2021-0-02068, Artificial Intelligence Innovation Hub, and No. 2022-0-00984, Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation). + +# References + +[1] Humam Alwassel, Fabian Caba Heilbron, Victor Escorchia, and Bernard Ghanem. Diagnosing error in temporal action detectors. In Proceedings of the European Conference on Computer Vision, pages 256-272, 2018. 8 +[2] Humam Alwassel, Silvio Giancola, and Bernard Ghanem. Tsp: Temporally-sensitive pretraining of video encoders for localization tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 3173-3183, 2021. 2 +[3] Yueran Bai, Yingying Wang, Yunhai Tong, Yang Yang, Qiyue Liu, and Junhui Liu. Boundary content graph neural network for temporal action proposal generation. In Proceedings of the European Conference on Computer Vision, pages 121-137, 2020. 1, 2, 6, 7 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision, pages 213-229, 2020. 1, 3, 6 +[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 2, 3, 6, 7 +[6] Feng Cheng and Gedas Bertasius. Tallformer: Temporal action localization with long-memory transformer. In Proceedings of the European Conference on Computer Vision, pages 503-521, 2022. 1, 2, 6, 7 +[7] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, and Will Price. The epic-kitchens dataset: Collection, challenges and baselines. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):4125-4141, 2020. 2, 6 +[8] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 961-970, 2015. 2, 6 +[9] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6202-6211, 2019. 3, 7 +[10] Haroon Idrees, Amir R Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The thumos challenge on action recognition for videos “in the wild”. Computer Vision and Image Understanding, 155:1-23, 2017. 2, 6 +[11] Jihwan Kim, Miso Lee, and Jae-Pil Heo. Self-feedback detr for temporal action detection. In ICCV, pages 10286-10296, 2023. 6 +[12] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6, 1 +[13] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 3 +[14] Chuming Lin, Chengming Xu, Donghao Luo, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Learning salient boundary feature for anchor-free temporal action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3320-3329, 2021. 1, 2, 6, 7 +[15] Tianwei Lin, Xu Zhao, Haisheng Su, Chongjing Wang, and Ming Yang. Bsn: Boundary sensitive network for temporal action proposal generation. In Proceedings of the European Conference on Computer Vision, pages 3-19, 2018. 1, 2, 6, 7 +[16] Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, and Shilei Wen. Bmn: Boundary-matching network for temporal action proposal generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3889-3898, 2019. 1, 2, 6, 7 +[17] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 2980-2988, 2017. 6, 1 +[18] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 2022. 1, 3, 4, 5, 6, 7 +[19] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2 +[20] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 2, 6, 7 +[21] Fuchen Long, Ting Yao, Zhaofan Qiu, Xinmei Tian, Jiebo Luo, and Tao Mei. Gaussian temporal awareness networks for action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 344-353, 2019. 1, 2 +[22] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 1 +[23] Zhiwu Qing, Haisheng Su, Weihao Gan, Dongliang Wang, Wei Wu, Xiang Wang, Yu Qiao, Junjie Yan, Changxin Gao, and Nong Sang. Temporal context aggregation network for temporal action proposal refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 485-494, 2021. 1, 2, 6, 7 +[24] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 658-666, 2019. 6, 1 +[25] Jiayi Shao, Xiaohan Wang, Ruijie Quan, Junjun Zheng, Jiang Yang, and Yi Yang. Action sensitivity learning for temporal + +action localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 7 +[26] Dingfeng Shi, Yujie Zhong, Qiong Cao, Jing Zhang, Lin Ma, Jia Li, and Dacheng Tao. React: Temporal action detection with relational queries. In Proceedings of the European Conference on Computer Vision, pages 105–121, 2022. 1, 3, 5, 6, 7 +[27] Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, and Dacheng Tao. Tridet: Temporal action detection with relative boundary modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18857-18866, 2023. 1, 2, 4, 6, 7 +[28] Jing Tan, Jiaqi Tang, Limin Wang, and Gangshan Wu. Relaxed transformer decoders for direct action proposal generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13526-13535, 2021. 1, 3, 5, 6, 7 +[29] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 2, 6, 7 +[30] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In Proceedings of the European Conference on Computer Vision, pages 20-36, 2016. 6, 7 +[31] Mengmeng Xu, Chen Zhao, David S Rojas, Ali Thabet, and Bernard Ghanem. G-tad: Sub-graph localization for temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10156-10165, 2020. 1, 2, 6, 7 +[32] Chenlin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In Proceedings of the European Conference on Computer Vision, pages 492-510, 2022. 1, 2, 4, 6, 7 +[33] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In Proceedings of the European Conference on Computer Vision, 2022. 4, 1 +[34] Chen Zhao, Ali K Thabet, and Bernard Ghanem. Video self-stitching graph network for temporal action localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13658-13667, 2021. 1, 2, 6, 7 +[35] Zixuan Zhao, Dongqi Wang, and Xu Zhao. Movement enhancement toward multi-scale video feature representation for temporal action detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13555-13564, 2023. 2, 6, 7 +[36] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12993-13000, 2020. 6, 1 +[37] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers + +for end-to-end object detection. In International Conference on Learning Representations, 2020. 4, 6, 1 \ No newline at end of file diff --git a/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/images.zip b/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ed38c002ee7f311bd7e1c5bf28d5ac3ef0fa2c1b --- /dev/null +++ b/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3785f25402ee75ad4f9179ee891e90097389ffbf49b350b83aa007c167e82fb9 +size 508665 diff --git a/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/layout.json b/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2b4ed901560d178f2d056f33438cdf9376ef58d2 --- /dev/null +++ b/2024/TE-TAD_ Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression/layout.json @@ -0,0 +1,8410 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 101, + 503, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 101, + 503, + 138 + ], + "spans": [ + { + "bbox": [ + 91, + 101, + 503, + 138 + ], + "type": "text", + "content": "TE-TAD: Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 152, + 184, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 152, + 184, + 167 + ], + "spans": [ + { + "bbox": [ + 104, + 152, + 184, + 167 + ], + "type": "text", + "content": "Ho-Joong Kim" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 206, + 152, + 283, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 152, + 283, + 167 + ], + "spans": [ + { + "bbox": [ + 206, + 152, + 283, + 167 + ], + "type": "text", + "content": "Jung-Ho Hong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 152, + 371, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 152, + 371, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 152, + 371, + 167 + ], + "type": "text", + "content": "Heejo Kong²" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 392, + 152, + 487, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 152, + 487, + 167 + ], + "spans": [ + { + "bbox": [ + 392, + 152, + 487, + 167 + ], + "type": "text", + "content": "Seong-Whan Lee" + }, + { + "bbox": [ + 392, + 152, + 487, + 167 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 167, + 170, + 426, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 170, + 426, + 182 + ], + "spans": [ + { + "bbox": [ + 167, + 170, + 426, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 167, + 170, + 426, + 182 + ], + "type": "text", + "content": "Dept. of Artificial Intelligence, Korea University, Seoul, Korea" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 145, + 184, + 448, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 184, + 448, + 197 + ], + "spans": [ + { + "bbox": [ + 145, + 184, + 448, + 197 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 184, + 448, + 197 + ], + "type": "text", + "content": "Dept. of Brain and Cognitive Engineering, Korea University, Seoul, Korea" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 148, + 198, + 441, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 198, + 441, + 210 + ], + "spans": [ + { + "bbox": [ + 148, + 198, + 441, + 210 + ], + "type": "text", + "content": "{hojoong_kim, jungho-hong, hj_kong, sw.lee}@korea.ac.kr" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 230, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 230, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 230, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 251, + 290, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 290, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 290, + 525 + ], + "type": "text", + "content": "In this paper, we investigate that the normalized coordinate expression is a key factor as reliance on handcrafted components in query-based detectors for temporal action detection (TAD). Despite significant advancements towards an end-to-end framework in object detection, query-based detectors have been limited in achieving full end-to-end modeling in TAD. To address this issue, we propose TE-TAD, a full end-to-end temporal action detection transformer that integrates time-aligned coordinate expression. We reformulate coordinate expression utilizing actual timeline values, ensuring length-invariant representations from the extremely diverse video duration environment. Furthermore, our proposed adaptive query selection dynamically adjusts the number of queries based on video length, providing a suitable solution for varying video durations compared to a fixed query set. Our approach not only simplifies the TAD process by eliminating the need for hand-crafted components but also significantly improves the performance of query-based detectors. Our TE-TAD outperforms the previous query-based detectors and achieves competitive performance compared to state-of-the-art methods on popular benchmark datasets. Code is available at: https://github.com/Dotori-HJ/TE-TAD" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 533, + 128, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 128, + 545 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 128, + 545 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 553, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 553, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 47, + 553, + 287, + 660 + ], + "type": "text", + "content": "Temporal action detection (TAD) plays an essential role in video understanding and its numerous real-world applications, such as video surveillance, video summarization, and video retrieval. TAD aims to recognize and localize actions within untrimmed video sequences by identifying the class labels with precise start and end times of action instances. Recently, TAD methods can be mainly divided by three approaches: anchor-based [3, 15, 16, 21, 23, 31, 34], anchor-free [6, 14, 27, 32], and query-based [18, 26, 28] detector." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 661, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 661, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 661, + 287, + 696 + ], + "type": "text", + "content": "Query-based detectors, inspired by DETR [4], have attracted interest because of their potential to eliminate reliance on hand-crafted components, such as the sliding win" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 316, + 230, + 537, + 353 + ], + "blocks": [ + { + "bbox": [ + 316, + 230, + 537, + 353 + ], + "lines": [ + { + "bbox": [ + 316, + 230, + 537, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 537, + 353 + ], + "type": "image", + "image_path": "5b1cb195368c23efc6302d8e872d53cc15ab5bfb5ba26733d817014f906f129e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 356, + 547, + 422 + ], + "lines": [ + { + "bbox": [ + 304, + 356, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 547, + 422 + ], + "type": "text", + "content": "Figure 1. Performance comparison of query-based detectors across various feature coverages on THUMOS14, demonstrating how extending the feature coverage impacts detection performance, as measured by mean Average Precision (mAP@AVG). The full end-to-end setting that without coverage constraints is denoted by " + }, + { + "bbox": [ + 304, + 356, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 304, + 356, + 547, + 422 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 426, + 547, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 557 + ], + "type": "text", + "content": "dow and non-maximum suppression (NMS). This potential derives from adopting a set-prediction mechanism, which aims to provide an end-to-end detection process by utilizing a one-to-one matching paradigm. Despite these advantages, query-based detectors encounter significant challenges in two aspects: (1) they show decreased performance when dealing with extended temporal coverage, often making them a less favorable option compared to anchor-free detectors, and (2) due to limited extended temporal coverage, the reliance on the sliding window approach leads to redundant proposals and necessitates the use of NMS." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 558, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 548, + 714 + ], + "type": "text", + "content": "To demonstrate these issues, we conduct experiments by increasing the feature coverages on existing query-based detectors. Fig. 1 demonstrates the change in performance across various feature coverage. Except for unrestricted feature coverage denoted by " + }, + { + "bbox": [ + 304, + 558, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 304, + 558, + 548, + 714 + ], + "type": "text", + "content": ", feature coverage is calculated from the window size of the sliding window method. As shown in the graph, even though wide feature coverage is beneficial to capturing longer context, the performance of existing query-based detectors diminishes in mean Average Precision (mAP) as feature coverage increments. This result indicates that there are significant limitations to query-based detectors in scalability to temporal length, despite efforts to address long durations [32] in TAD." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 138, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18837" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "Furthermore, Fig. 2 shows the second issue, illustrating the limitations of the sliding window approach. As shown in Fig. 2(a), the sliding window is limited to address long-range duration due to its limited window size. Moreover, the sliding window approach contains overlapping areas to prevent predictions from being truncated in the middle. These overlapping areas generate duplicate predictions, necessitating the use of NMS to filter the false positive cases. This reliance on NMS contradicts the set-prediction goal of minimizing hand-crafted components, thus hindering the achievement of a fully end-to-end TAD. To address these issues, we investigate why extended temporal coverage adversely affects the performance of query-based detectors (Sec. 3.3). Our investigation reveals that the conventional use of normalized coordinate expressions is a significant factor, disturbing the achievement of a full end-to-end TAD." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 289, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 541 + ], + "type": "text", + "content": "In this paper, we propose a full end-to-end temporal action detection transformer that integrates time-aligned coordinate expression (TE-TAD), which reformulates normalized coordinate expression to actual timeline video values. Our reformulation enables the query-based detector to effectively address length-invariant modeling by avoiding the distortion of the normalizing process, which not only enhances the detection performance but also simplifies the detection process. Our TE-TAD stabilizes the training process of the query-based detector when dealing with extended videos; our approach shows significant improvements and completely removes the reliance on hand-crafted components such as the sliding window and NMS. Furthermore, we introduce an adaptive query selection that effectively addresses various video lengths, dynamically adjusting the number of queries in response to the temporal length of each video. In contrast to relying on a fixed set of queries, our TE-TAD provides a suitable approach to process diverse video lengths. Our approach shows significant improvements compared to previous query-based detectors and achieves competitive performance with state-of-the-art methods on popular benchmark datasets: THUMOS14 [10], ActivityNet v1.3 [8], and EpicKitchens [7]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 544, + 243, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 243, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 243, + 555 + ], + "type": "text", + "content": "Our contributions are summarized as three-fold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 558, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 558, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 287, + 617 + ], + "type": "text", + "content": "- We propose a full end-to-end temporal action detection transformer that integrates time-aligned coordinate expression (TE-TAD), which preserves the set-prediction mechanism and enables a full end-to-end modeling for TAD by eliminating the hand-crafted components." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 618, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 287, + 653 + ], + "type": "text", + "content": "- Our approach introduces a length-invariant mechanism to query-based detectors, significantly improving scalability in handling varying lengths of videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "text", + "content": "- Our TE-TAD significantly outperforms the previous query-based detectors and achieves competitive performance compared to state-of-the-art methods, even without hand-crafted components such as the sliding window and NMS." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 322, + 69, + 531, + 159 + ], + "blocks": [ + { + "bbox": [ + 322, + 69, + 531, + 159 + ], + "lines": [ + { + "bbox": [ + 322, + 69, + 531, + 159 + ], + "spans": [ + { + "bbox": [ + 322, + 69, + 531, + 159 + ], + "type": "image", + "image_path": "233dc2ece837e5460e3600ca008616558e77436becdf465f0d9554c68689fc7a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 160, + 457, + 170 + ], + "lines": [ + { + "bbox": [ + 394, + 160, + 457, + 170 + ], + "spans": [ + { + "bbox": [ + 394, + 160, + 457, + 170 + ], + "type": "text", + "content": "(a) Sliding window" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 171, + 529, + 249 + ], + "blocks": [ + { + "bbox": [ + 322, + 171, + 529, + 249 + ], + "lines": [ + { + "bbox": [ + 322, + 171, + 529, + 249 + ], + "spans": [ + { + "bbox": [ + 322, + 171, + 529, + 249 + ], + "type": "image", + "image_path": "6eeb3c774ffc851cf4718d15d445845300eea92438c41572e32ebca0560966e2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 251, + 462, + 261 + ], + "lines": [ + { + "bbox": [ + 390, + 251, + 462, + 261 + ], + "spans": [ + { + "bbox": [ + 390, + 251, + 462, + 261 + ], + "type": "text", + "content": "(b) End-to-end setting" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 263, + 547, + 297 + ], + "lines": [ + { + "bbox": [ + 305, + 263, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 263, + 547, + 297 + ], + "type": "text", + "content": "Figure 2. Comparison between sliding window and end-to-end settings. The sliding window generates redundant proposals in the duplicated area." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 308, + 392, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 392, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 392, + 319 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 328, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 472 + ], + "type": "text", + "content": "Action Recognition Action Recognition is a foundational task in video understanding, categorizing video sequences into distinct action classes. Notable models include I3D [5], which enhances the inception network with 3D convolutions, and " + }, + { + "bbox": [ + 304, + 328, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{R}(2 + 1)\\mathrm{D}" + }, + { + "bbox": [ + 304, + 328, + 545, + 472 + ], + "type": "text", + "content": " [29], separating 3D convolutions into 2D spatial and 1D temporal parts for efficient processing. TSP [2] introduces temporal channel shifting for effective temporal modeling without extra computation. VideoSwin [20], employing the SwinTransformer [19] architecture, excels in complex video data recognition. These models serve as backbones for extracting video features, directly impacting performance in subtasks like TAD." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 473, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 569 + ], + "type": "text", + "content": "Anchor-based Detector Anchor-based detectors [3, 15, 16, 21, 23, 31, 34] leverage predefined anchor boxes to generate action proposals. These hand-designed anchors hinder the diverse range of action instances because of a lack of flexibility in the localization of the action instances. This approach inherently limits and necessitates additional post-processing steps to discard redundant proposals because they model the one-to-many assignment training strategy." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Anchor-free Detector Anchor-free detectors [6, 14, 25, 27, 32, 35] offer more flexibility in action instance localization compared to anchor-based detectors by adopting an asymmetric modeling approach. For instance, ActionFormer [32] significantly enhances TAD performance by employing a transformer-based architecture that captures long-range video dependencies. TriDet [27] demonstrates superior performance in TAD by employing a trident prediction scheme and their proposed convolution-based architecture. Despite their improvements, anchor-free detectors rely on handcrafted components to remove redundant proposals using NMS because they adopt a one-to-many assignment train-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18838" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 373 + ], + "type": "text", + "content": "ing manner. In contrast, our approach directly addresses a one-to-one matching scheme, eliminating the use of NMS. Query-based Detector Query-based detectors, inspired by DETR [4], introduce a set-prediction mechanism, thereby reducing the reliance on hand-crafted components, which ideally prevents the need for NMS. However, existing query-based detectors still require NMS because their model design inherently breaks the one-to-one matching paradigm. RTD-Net [28] utilizes a one-to-many matching to mitigate the slow convergence issue associated with the detection transformer. This approach inherently breaks the one-to-one assignment. ReAct [26] modify the decoder's self-attention, called relational attention, only adopt self-attention between their defined relations. This partial adoption of the decoder's self-attention disturbs the set-prediction mechanism because they cannot capture the whole context of queries. Furthermore, previous query-based detectors [18, 26, 28] adopt the sliding window method that contains overlapping areas, causing redundant proposals. Furthermore, TadTR [18] deals with one-to-many matching at training loss. TadTR employs cross-window fusion (CWF), which applies NMS to overlapping areas to remove redundant proposals. In contrast, our approach entirely preserves the one-to-one matching paradigm, which enables a full end-to-end modeling." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 384, + 138, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 138, + 397 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 138, + 397 + ], + "type": "text", + "content": "3. Our Approach" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 403, + 115, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 115, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 115, + 415 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 422, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 543 + ], + "type": "text", + "content": "In this section, we first discuss about the limitations of existing query-based detectors in TAD, focusing on the normalized coordinate expression. The normalized coordinate expression, used in existing models, causes matching instability and sensitivity, especially in extended video scenarios. Subsequently, to introduce our TE-TAD, we describe the reformulation of normalized coordinate expression to timeline coordinate expression and adaptive query proposals to ensure the length-invariant modeling. The overall architecture of TE-TAD is illustrated in Fig. 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 551, + 127, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 127, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 127, + 563 + ], + "type": "text", + "content": "3.2. Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{T_0 \\times C}" + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "text", + "content": " denote the video feature sequence extracted by the backbone network, where " + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "inline_equation", + "content": "T_0" + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "text", + "content": " is the temporal length of the features, and " + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "text", + "content": " is the dimension of the video feature. Each element in the video feature sequence represented as " + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "inline_equation", + "content": "X = \\{x_t\\}_{t=1}^{T_0}" + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "text", + "content": ", corresponds to a snippet at timestep " + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "text", + "content": ", with each snippet comprising a few consecutive frames. These snippets are processed using a pre-trained backbone network such as I3D [5] or SlowFast [9]. Each video contains numerous action instances, and each action instance contains start and end timestamps " + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "text", + "content": ", along with its action class " + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "text", + "content": ". Formally, the set of action instances in a video is represented as " + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "inline_equation", + "content": "A = \\{(s_n, e_n, c_n)\\}_{n=1}^N" + }, + { + "bbox": [ + 46, + 569, + 288, + 715 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 331, + 72, + 521, + 171 + ], + "blocks": [ + { + "bbox": [ + 331, + 72, + 521, + 171 + ], + "lines": [ + { + "bbox": [ + 331, + 72, + 521, + 171 + ], + "spans": [ + { + "bbox": [ + 331, + 72, + 521, + 171 + ], + "type": "image", + "image_path": "ee5b6860d177876d4266404ba5716c5c6b3ac6bc433f20dbdc5db9af9244223a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 173, + 432, + 182 + ], + "lines": [ + { + "bbox": [ + 421, + 173, + 432, + 182 + ], + "spans": [ + { + "bbox": [ + 421, + 173, + 432, + 182 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 329, + 184, + 518, + 284 + ], + "blocks": [ + { + "bbox": [ + 329, + 184, + 518, + 284 + ], + "lines": [ + { + "bbox": [ + 329, + 184, + 518, + 284 + ], + "spans": [ + { + "bbox": [ + 329, + 184, + 518, + 284 + ], + "type": "image", + "image_path": "d770c65a35e14d8450b047700d5a9d1044ec2d9f2c59fbd347ea0448a5b77756.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 286, + 430, + 295 + ], + "lines": [ + { + "bbox": [ + 419, + 286, + 430, + 295 + ], + "spans": [ + { + "bbox": [ + 419, + 286, + 430, + 295 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 297, + 545, + 353 + ], + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 353 + ], + "type": "text", + "content": "Figure 3. Comparative analysis of instability and detection performance on THUMOS14: (a) variance in instance matching quantified by IS; (b) the performance change in mAP. Feature coverage lengths are denoted by the values following the dash (-). The symbol " + }, + { + "bbox": [ + 304, + 297, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 304, + 297, + 545, + 353 + ], + "type": "text", + "content": " denotes an unrestricted end-to-end setting." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "text", + "content": " is the number of action instances, and " + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "inline_equation", + "content": "s_n" + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "inline_equation", + "content": "e_n" + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "text", + "content": " are the start and end timestamps of an action instance, respectively, and " + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "inline_equation", + "content": "c_n" + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "text", + "content": " is its action class. The main goal of TAD is to accurately predict the set of action instances " + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "text", + "content": " for any given video. Previous query-based detectors [18, 26, 28] typically compute predicted values of center " + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\hat{d}" + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "text", + "content": " using a sigmoid function " + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "inline_equation", + "content": "(\\sigma)" + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "text", + "content": ". Consequently, existing models decode predicted start " + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\hat{s}" + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "text", + "content": " and end " + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\hat{e}" + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "text", + "content": " timestamps by " + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\sigma(\\hat{c}) - \\sigma(\\hat{d})" + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\sigma(\\hat{c}) + \\sigma(\\hat{d})" + }, + { + "bbox": [ + 304, + 362, + 545, + 472 + ], + "type": "text", + "content": " to start and end timestamps, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 479, + 427, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 427, + 492 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 427, + 492 + ], + "type": "text", + "content": "3.3. Exploring Key Issues" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Matching Instability As discussed in Sec. 1, extended video lengths significantly influence the performance of existing query-based methods. To investigate this issue, we conduct a comparative study on the instability (IS) [13] and the detection performance across different feature coverage scenarios on THUMOS14. Here, IS is a quantitative measurement of the inconsistency of matching during the training process. For query-based detectors, fluctuations in matched targets compel the model to be learned from different values for the same input, leading to performance degradation. In the TadTR setting, smaller window sizes have more steps per epoch due to generating more sliding windows, whereas larger window sizes yield fewer steps per epoch. This mismatch in the number of steps results that with fewer updates per epoch, there is inherently less change to the model, which shows lower instability. For a fair comparison, we match the number of iteration steps per epoch, aligning with TadTR-34.1s (original TadTR)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18839" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 276, + 167, + 371 + ], + "blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 262 + ], + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 262 + ], + "type": "text", + "content": "Fig. 3 illustrates the instability and detection performance across diverse feature coverage. As feature coverage increases, we observe a rise in IS across the training epoch, indicating less stable instance matching, which leads to a decline in detection performance. This analysis underscores the challenge of maintaining consistent learning when with extended temporal lengths. Furthermore, a direct comparison between TadTR models and our method reveals that Ours- " + }, + { + "bbox": [ + 47, + 72, + 289, + 262 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 47, + 72, + 289, + 262 + ], + "type": "text", + "content": " maintains a level of stability comparable to TadTR-34.1s. This indicates that our approach significantly stabilizes the training process relative to TadTR- " + }, + { + "bbox": [ + 47, + 72, + 289, + 262 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 47, + 72, + 289, + 262 + ], + "type": "text", + "content": ". Moreover, our method demonstrates similar levels of matching instability yet shows significant improvements under more challenging conditions for matching problems, even when compared to models with shorter feature coverage like TadTR-34.1s and TadTR-273.1s." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 276, + 167, + 371 + ], + "lines": [ + { + "bbox": [ + 50, + 276, + 167, + 371 + ], + "spans": [ + { + "bbox": [ + 50, + 276, + 167, + 371 + ], + "type": "image", + "image_path": "f6fe4afb96fd47c0096ae186d2b67e30bb005ac7721244494c928eae0ecb58a5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 375, + 128, + 385 + ], + "lines": [ + { + "bbox": [ + 87, + 375, + 128, + 385 + ], + "spans": [ + { + "bbox": [ + 87, + 375, + 128, + 385 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 87, + 375, + 128, + 385 + ], + "type": "inline_equation", + "content": "\\sigma (\\hat{c} +\\epsilon)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 388, + 287, + 422 + ], + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 422 + ], + "type": "text", + "content": "Figure 4. Analysis of noise tolerance on predicted value. The noise level " + }, + { + "bbox": [ + 46, + 388, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 388, + 287, + 422 + ], + "type": "text", + "content": " sampled from uniform distribution and injected before the sigmoid function." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 168, + 276, + 286, + 370 + ], + "blocks": [ + { + "bbox": [ + 168, + 276, + 286, + 370 + ], + "lines": [ + { + "bbox": [ + 168, + 276, + 286, + 370 + ], + "spans": [ + { + "bbox": [ + 168, + 276, + 286, + 370 + ], + "type": "image", + "image_path": "2de3b7dc477067254ee38879ce9dcc6e070b50402a7425e1b0ba3bdb96a06a31.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 374, + 247, + 385 + ], + "lines": [ + { + "bbox": [ + 205, + 374, + 247, + 385 + ], + "spans": [ + { + "bbox": [ + 205, + 374, + 247, + 385 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 205, + 374, + 247, + 385 + ], + "type": "inline_equation", + "content": "\\sigma (\\hat{d} +\\epsilon)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 437, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 289, + 715 + ], + "type": "text", + "content": "Sensitivity to Localize Action Instances To further explore this issue, we investigate the sensitivity of model prediction measured by adopting minor perturbations in predictions of localization. We inject a small scale of noise sampled from a uniform distribution before adopting the sigmoid function that normalizes coordinates within a [0, 1] range. Fig. 4 shows that even minimal noise injection significantly affects the prediction when applying the center prediction value. The noise " + }, + { + "bbox": [ + 47, + 437, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathrm{Uniform}(-\\alpha, \\alpha)" + }, + { + "bbox": [ + 47, + 437, + 289, + 715 + ], + "type": "text", + "content": " is sampled from the uniform distribution. As illustrated in Fig. 4(a), the model's sensitivity to small shifts is evident from the significant decline in " + }, + { + "bbox": [ + 47, + 437, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{mAP}@\\mathrm{AVG}" + }, + { + "bbox": [ + 47, + 437, + 289, + 715 + ], + "type": "text", + "content": " upon introducing noise to the center predicted value. Notably, with noise levels of " + }, + { + "bbox": [ + 47, + 437, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\pm 0.01" + }, + { + "bbox": [ + 47, + 437, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 437, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\pm 0.1" + }, + { + "bbox": [ + 47, + 437, + 289, + 715 + ], + "type": "text", + "content": ", after processing through the sigmoid function " + }, + { + "bbox": [ + 47, + 437, + 289, + 715 + ], + "type": "inline_equation", + "content": "(\\sigma)" + }, + { + "bbox": [ + 47, + 437, + 289, + 715 + ], + "type": "text", + "content": ", the maximal shifts are confined within " + }, + { + "bbox": [ + 47, + 437, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\pm 0.0025" + }, + { + "bbox": [ + 47, + 437, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 437, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\pm 0.025" + }, + { + "bbox": [ + 47, + 437, + 289, + 715 + ], + "type": "text", + "content": " in the normalized coordinate space, respectively. These findings highlight that even minor output variations amplify in extended videos, leading to considerable drops in performance due to the heightened sensitivity in the normalized coordinate framework. To address this issue, we introduce a time-aligned coordinate expression that is not normalized, thereby ensuring independence from video length and reducing sensitivity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 369, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 369, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 369, + 83 + ], + "type": "text", + "content": "3.4. TE-TAD" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "content": "This part describes our TE-TAD for a full end-to-end TAD. We adopt the TadTR [18] architecture as a baseline method, including encoder, decoder, and temporal deformable attention architecture. Starting with the baseline, we mainly address three aspects: (1) adopting multi-scale and two-stage methods from the previous methods in object detection to bridge the performance gap between query-based and anchor-free detectors, (2) reformulating coordinate expression utilizing the actual time values to address extended video length in an end-to-end setting, and (3) proposing an adaptive query selection that dynamically adjusts the number of queries based on the diverse length of videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 234, + 546, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 546, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 546, + 389 + ], + "type": "text", + "content": "Embedding & Multi-Scale Features We project input features " + }, + { + "bbox": [ + 304, + 234, + 546, + 389 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 234, + 546, + 389 + ], + "type": "text", + "content": " using a single convolutional neural network to align them with the dimension of the transformer architecture. The projection maps the input feature " + }, + { + "bbox": [ + 304, + 234, + 546, + 389 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 234, + 546, + 389 + ], + "type": "text", + "content": " to the embedded feature " + }, + { + "bbox": [ + 304, + 234, + 546, + 389 + ], + "type": "inline_equation", + "content": "Z_{1} \\in \\mathbb{R}^{D \\times T_{1}}" + }, + { + "bbox": [ + 304, + 234, + 546, + 389 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 234, + 546, + 389 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 234, + 546, + 389 + ], + "type": "text", + "content": " denotes the channel dimension of the encoder and decoder transformer architecture. The temporal length " + }, + { + "bbox": [ + 304, + 234, + 546, + 389 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 304, + 234, + 546, + 389 + ], + "type": "text", + "content": " of the embedded features " + }, + { + "bbox": [ + 304, + 234, + 546, + 389 + ], + "type": "inline_equation", + "content": "Z_{1}" + }, + { + "bbox": [ + 304, + 234, + 546, + 389 + ], + "type": "text", + "content": " remains the same as the original " + }, + { + "bbox": [ + 304, + 234, + 546, + 389 + ], + "type": "inline_equation", + "content": "T_{0}" + }, + { + "bbox": [ + 304, + 234, + 546, + 389 + ], + "type": "text", + "content": ". Subsequently, following previous approaches [27, 32], we incorporate multi-scale generation to effectively address varying lengths of actions. Unlike utilizing a transformer [32], we employ a single convolution layer with a stride of 2 to produce features at each scale level as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 334, + 399, + 545, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 399, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 334, + 399, + 545, + 411 + ], + "type": "interline_equation", + "content": "Z _ {l} = \\operatorname {L a y e r N o r m} _ {l} \\left(\\operatorname {C o n v} _ {l} \\left(Z _ {l - 1}\\right)\\right), l \\in \\{2, \\dots , L \\}, \\tag {1}", + "image_path": "8ef01edf206dd0c2dbffcad907afb953d7ff31d2406b6e36d242b22b83b38259.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "inline_equation", + "content": "Z_{l} \\in \\mathbb{R}^{C \\times T_{l}}" + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "text", + "content": " represents the embedded features at each level " + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "text", + "content": " denotes the total number of feature levels. Each subsequent level " + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "text", + "content": " has a temporal length " + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "inline_equation", + "content": "T_{l}" + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "text", + "content": " that is half of the temporal level of the previous level " + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "inline_equation", + "content": "T_{l-1}" + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "text", + "content": ". The LayerNorm " + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "inline_equation", + "content": "_{l}" + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "text", + "content": " and Conv " + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "inline_equation", + "content": "_{l}" + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "text", + "content": " denote " + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "text", + "content": "-th layer normalization and convolutional neural networks, respectively. We do not apply any activation function in this process to deliver the raw feature representations to the transformer detector." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "text", + "content": "Time-Aligned Query Generation Our method follows a two-stage approach [33, 37] that generates initial action proposals using the transformer encoder. The previous two-stage approach [37] provides the reference to the encoder's outputs. The transformer encoder predicts a binary foreground score, " + }, + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "inline_equation", + "content": "p^{(0)}" + }, + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "text", + "content": ", and segment offsets " + }, + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\Delta c^{(0)}" + }, + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\Delta d^{(0)}" + }, + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "text", + "content": " to refine segments based on reference, for each time " + }, + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "text", + "content": " and level " + }, + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "text", + "content": ". We define reference for center " + }, + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "inline_equation", + "content": "c^{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "inline_equation", + "content": "d^{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "text", + "content": " predictions, aligning with the real timeline of the video. For each scale level " + }, + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 516, + 545, + 647 + ], + "type": "text", + "content": ", the reference for the center is computed as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 310, + 653, + 545, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 653, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 310, + 653, + 545, + 689 + ], + "type": "interline_equation", + "content": "c ^ {\\text {r e f}} = \\left\\{t \\times \\frac {f}{w \\times 2 ^ {l - 1}} + \\frac {w \\times 2 ^ {l - 1}}{2} \\right\\} _ {t = 1} ^ {T _ {l}}, l \\in \\{1, 2, \\dots , L \\}, \\tag {2}", + "image_path": "3566b394c7271209344d6abc7322ac8db4f09a4fb945be48d2947f20b9f76982.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " denotes the frame-per-second rate of the video, and " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " represents step size for feature extracting that indicates" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18840" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 552, + 277 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 552, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 552, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 552, + 277 + ], + "type": "image", + "image_path": "70316b20b3e11eafe01f58796fcb637d3d5f1f21d2e5d6895a23e5164ec9bd8b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 280, + 547, + 315 + ], + "lines": [ + { + "bbox": [ + 46, + 280, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 547, + 315 + ], + "type": "text", + "content": "Figure 5. Overview of the TE-TAD. Starting with video input, the architecture processes through a backbone for feature extraction, generating multi-scale features " + }, + { + "bbox": [ + 46, + 280, + 547, + 315 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 280, + 547, + 315 + ], + "type": "text", + "content": ". These are encoded and subsequently passed through an adaptive query selection, aligning with the video timeline for initial query generation. The decoder refines these queries layer-by-layer, culminating in the refinement of action proposals." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 324, + 288, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 288, + 384 + ], + "type": "text", + "content": "how many frames to step when extracting features. The factor " + }, + { + "bbox": [ + 46, + 324, + 288, + 384 + ], + "type": "inline_equation", + "content": "2^{l - 1}" + }, + { + "bbox": [ + 46, + 324, + 288, + 384 + ], + "type": "text", + "content": " fits the temporal lengths to the embedded features from the multi-scale generation, corresponding with the actual timeline. Subsequently, the reference for the width is computed as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 390, + 287, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 390, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 93, + 390, + 287, + 405 + ], + "type": "interline_equation", + "content": "d ^ {\\text {r e f}} = \\alpha \\cdot f \\times 2 ^ {l - 1}, l \\in \\{1, 2, \\dots , L \\}, \\tag {3}", + "image_path": "5561ca0c44edf76b6cb7b2e648cda1ff5a1759e433afa6207e4839970aca40fb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 411, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 411, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 411, + 287, + 448 + ], + "type": "text", + "content": " is the base scale for encoder proposals, which adjusts the length of the reference width length. Consequently, the time-aligned queries are decoded as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 453, + 287, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 287, + 469 + ], + "type": "interline_equation", + "content": "(\\hat {c} ^ {(0)}, \\hat {d} ^ {(0)}) = (c ^ {\\mathrm {r e f}} + \\Delta c ^ {(0)} \\cdot d ^ {\\mathrm {r e f}}, \\exp (\\ln (d ^ {\\mathrm {r e f}}) + \\Delta d ^ {(0)})) (4)", + "image_path": "44068560d411ea4a7aa48831b873581b6dedacff475c069d8431bd89bf9af7b7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 473, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 473, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\hat{c}^{(0)}" + }, + { + "bbox": [ + 46, + 473, + 287, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 473, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\hat{d}^{(0)}" + }, + { + "bbox": [ + 46, + 473, + 287, + 545 + ], + "type": "text", + "content": " denote the center and width of the proposals, respectively. Our approach utilizes the scaling of the center offsets with width " + }, + { + "bbox": [ + 46, + 473, + 287, + 545 + ], + "type": "inline_equation", + "content": "a^{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 473, + 287, + 545 + ], + "type": "text", + "content": " and exp at the proposals to address the scale-invariant approach. The decoder then refines these initial locations of queries after the proposed adaptive query selection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "content": "Adaptive Query Selection In a conventional two-stage approach in object detection, a fixed top-" + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "content": " selection method based on binary class predictions " + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "inline_equation", + "content": "p^{(0)}" + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "content": " is typically employed. However, this static method may not be optimal for TAD, where the number and duration of action instances within videos vary significantly. In TAD, there's often a direct correlation between the length of a video and the number of action instances. Typically, longer videos contain more action instances, while shorter videos have fewer. This variation presents a challenge for the fixed selection method, which fails to adapt to video content characteristics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "Our method divides videos into sectors of a base length " + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{sector}}" + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": ". We then perform top-" + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": " selection for each sector. This individual selection allows the detector to adapt to the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "text", + "content": "video length, preventing the selection of only some parts when dealing with long videos. In practice, we redistribute any remaining timesteps to ensure that no sector at the end part of the video is smaller than " + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{sector}}" + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "text", + "content": ". The total number of sectors, " + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "text", + "content": ", is calculated by dividing the number of the first layer's feature " + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{sector}}" + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "text", + "content": ", applying a floor function. For each sector " + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "inline_equation", + "content": "s \\in \\{1, 2, \\dots, S\\}" + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "text", + "content": ", we select the top- " + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "text", + "content": " proposals from all levels based on the encoder's binary class scores within that sector. We define a subset of encoder output scores, " + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "inline_equation", + "content": "P_{s}" + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "text", + "content": ", for each sector, and then select the top- " + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 324, + 547, + 467 + ], + "type": "text", + "content": " proposals from this subset. The adaptive query selection is represented as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 472, + 545, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 472, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 315, + 472, + 545, + 502 + ], + "type": "interline_equation", + "content": "\\mathcal {Q} = \\bigcup_ {s = 1} ^ {S} \\left\\{\\left(\\hat {c} _ {t, l} ^ {(0)}, \\hat {d} _ {t, l} ^ {(0)}\\right) \\mid (t, l) \\in \\text {i n d i c e s o f t o p - K i n} P _ {s} \\right\\}, \\tag {5}", + "image_path": "234be9688e1f0fd5a9debf7f0c0bb25005a0604dfd94a9dae7c9964254bf9b47.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 506, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 506, + 547, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\{(\\hat{c}_q^{(0)},\\hat{d}_q^{(0)})\\}_{q = 1}^{N_q}" + }, + { + "bbox": [ + 304, + 506, + 547, + 569 + ], + "type": "text", + "content": " is the aggregated set of selected queries. Here, " + }, + { + "bbox": [ + 304, + 506, + 547, + 569 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 506, + 547, + 569 + ], + "type": "text", + "content": " is the number of queries selected from each sector, and the total number of queries " + }, + { + "bbox": [ + 304, + 506, + 547, + 569 + ], + "type": "inline_equation", + "content": "N_{q}" + }, + { + "bbox": [ + 304, + 506, + 547, + 569 + ], + "type": "text", + "content": " equals the sum of the top- " + }, + { + "bbox": [ + 304, + 506, + 547, + 569 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 506, + 547, + 569 + ], + "type": "text", + "content": " proposals across all sectors, denoted as " + }, + { + "bbox": [ + 304, + 506, + 547, + 569 + ], + "type": "inline_equation", + "content": "N_{q} = \\sum_{s = 1}^{S}K" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 569, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 547, + 713 + ], + "type": "text", + "content": "Time-Aligned Segment Refinement Existing query-based models [18, 26, 28] employ the sigmoid function to express normalized coordinates from the range 0 to 1. Moreover, TadTR [18] utilizes the refining step in the decoder of the transformer using the predicted center and width of each layer. The layer-wise segment refinement step in the normalized coordinate expression is defined as " + }, + { + "bbox": [ + 304, + 569, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\sigma (\\sigma^{-1}(\\hat{c}_q^{(n - 1)}) + \\Delta c_q^{(n)})" + }, + { + "bbox": [ + 304, + 569, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 569, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\sigma (\\sigma^{-1}(\\hat{d}_q^{(n - 1)}) + \\Delta d_q^{(n)})" + }, + { + "bbox": [ + 304, + 569, + 547, + 713 + ], + "type": "text", + "content": " for center prediction and width prediction, respectively. This refinement step restricts the values within a [0, 1] but enables layer-by-layer updates. We reorganize the previous segment refinement step without the normalized expression." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18841" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "In our approach, given selected queries " + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": " are utilized for segment refinement. Formally, segment refinement of each layer is as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 114, + 287, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 287, + 129 + ], + "type": "interline_equation", + "content": "\\hat {c} _ {q} ^ {(n)} = \\hat {c} _ {q} ^ {(n - 1)} + \\Delta c _ {q} ^ {(n)} \\cdot \\hat {d} _ {q} ^ {(n - 1)}, n \\in \\{1, 2, \\dots , L _ {D} \\}, \\tag {6}", + "image_path": "601cf2dac2665c17b3dc0b50bdba8e9b2c4caae95c336f6c6afb8a6810553ddb.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 153, + 287, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 153, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 53, + 153, + 287, + 172 + ], + "type": "interline_equation", + "content": "\\hat {d} _ {q} ^ {(n)} = \\exp \\left(\\ln \\left(\\hat {d} _ {q} ^ {(n - 1)}\\right) + \\Delta \\hat {d} _ {q} ^ {(n)}\\right), n \\in \\{1, 2, \\dots , L _ {D} \\}, \\tag {7}", + "image_path": "77a083e3e68e005f1e5e2b54bbe058381bf84843229c886d579427f29f4eed2c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "inline_equation", + "content": "\\hat{c}_q^{(n)}" + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "inline_equation", + "content": "\\hat{d}_q^{(n)}" + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "text", + "content": " represent the predicted outputs of the center point and width from the decoder, respectively, of " + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "text", + "content": "-th layer for each query. Similar to the encoder case, we utilize the scaling of the center offsets with width " + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "inline_equation", + "content": "d^{(n)}" + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "text", + "content": " and exp function at the proposals to address the scale-invariant approach. The start " + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "inline_equation", + "content": "\\hat{s}^{(L_D)}" + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "text", + "content": " and end " + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "inline_equation", + "content": "\\hat{e}^{(L_D)}" + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "text", + "content": " timestamps for predictions are decoded as " + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "inline_equation", + "content": "\\hat{c}^{(L_D)} - \\hat{d}^{(L_D)}" + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "inline_equation", + "content": "\\hat{c}^{(L_D)} + \\hat{d}^{(L_D)}" + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "text", + "content": ", respectively. Consequently, the final predicted proposals is defined as " + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{A}} = (\\hat{s}^{(L_D)}, \\hat{e}^{(L_D)}, \\hat{p}^{(L_D)})" + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "inline_equation", + "content": "L_D" + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "text", + "content": " is the number of decoder layers, and " + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "inline_equation", + "content": "\\hat{p}^{(L_D)}" + }, + { + "bbox": [ + 46, + 180, + 287, + 314 + ], + "type": "text", + "content": " is the prediction of confidence score of last layer of decoder action class." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 321, + 178, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 321, + 178, + 334 + ], + "spans": [ + { + "bbox": [ + 47, + 321, + 178, + 334 + ], + "type": "text", + "content": "3.5. Training and Inference" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 339, + 287, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 339, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 339, + 287, + 364 + ], + "type": "text", + "content": "Training We follow the standard bipartite matching loss [4]. The total loss " + }, + { + "bbox": [ + 47, + 339, + 287, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{total}}" + }, + { + "bbox": [ + 47, + 339, + 287, + 364 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 372, + 287, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 372, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 85, + 372, + 287, + 405 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} (\\mathcal {A}, \\hat {\\mathcal {A}}) = \\sum_ {i = 1} ^ {N _ {q}} \\mathcal {L} _ {\\text {m a t c h}} \\left(\\mathcal {A} _ {i}, \\hat {\\mathcal {A}} _ {\\pi (i)}\\right), \\tag {8}", + "image_path": "26ad1ed36b4f429eedd628cab696a465ec900c53047bcbd52127f8f7e11f275c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 415, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 415, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{match}" + }, + { + "bbox": [ + 46, + 415, + 287, + 628 + ], + "type": "text", + "content": " is the bipartite matching loss that incorporates both the classification probabilities and the distance between the ground truth and the predicted segments, and " + }, + { + "bbox": [ + 46, + 415, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 415, + 287, + 628 + ], + "type": "text", + "content": " represent permutation indices obtained by bipartite matching [12]. This cost function, " + }, + { + "bbox": [ + 46, + 415, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{match}" + }, + { + "bbox": [ + 46, + 415, + 287, + 628 + ], + "type": "text", + "content": ", is a composite of the classification loss, and the regression loss. For the classification loss, we use the focal loss [17], which effectively addresses the class imbalance issue. For the regression loss, our model utilizes DIoU [36] and log-ratio distance for the width. DIoU evaluates the relative center distance and GIoU [24] once, while the log-ratio compares the widths relatively. Following the previous query-based approaches [4, 18, 37], we utilize the auxiliary decoding loss at every decoder layer for a more effective learning process. Furthermore, we do not apply any other losses without bipartite matching loss, such as actionness loss [18] or action classification enhancement loss [26]. More detailed descriptions of each loss are described in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Inference TE-TAD introduces a significant innovation by completely removing the need for common post-processing steps, such as NMS and temporal scaling. This is a direct consequence of our model's unique ability to work with an end-to-end approach and actual timeline values of the video, streamlining the inference process. The predictions from the final layer of the decoder " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{A}}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " are directly used." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 71, + 387, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 387, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 387, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 91, + 356, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 356, + 103 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 356, + 103 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 109, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 545, + 276 + ], + "type": "text", + "content": "Datasets We conduct experiments on three datasets: THUMOS14 [10], ActivityNet v1.3 [8], and EpicKitchens [7]. THUMOS14 comprises 20 action classes with 200 and 213 untrimmed videos in the validation and test sets, containing 3,007 and 3,358 action instances, respectively. ActivityNet v1.3 are large-scale datasets with 200 action classes. They consist of 10,024 videos for training and 4,926 videos for validation, respectively. EpicKitchens, a first-person vision dataset, includes two sub-tasks: noun and verb. It contains 495 and 138 videos with 67,217 and 9,668 action instances for training and test, with 300 and 97 action classes for nouns and verbs, respectively. These datasets contain diverse actions and scenes, providing a rigorous evaluation setup for our method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 277, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 372 + ], + "type": "text", + "content": "Evaluation Metric We follow the standard evaluation protocol for all datasets, utilizing mAP at different intersections over union (IoU) thresholds to evaluate TAD performance. The IoU thresholds for THUMOS14 and EpicKitchens are set at [0.3:0.7:0.1] and [0.1:0.5:0.1] respectively, while for ActivityNet v1.3, the results are reported at IoU threshold [0.5, 0.75, 0.95] with the average mAP computed at [0.5:0.95:0.05]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 372, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 420 + ], + "type": "text", + "content": "Implementation Details To ensure the clarity and focus of our main manuscript, detailed descriptions of the hyperparameters and experimental environments are provided in the supplementary materials." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 313, + 430, + 537, + 601 + ], + "blocks": [ + { + "bbox": [ + 313, + 430, + 537, + 601 + ], + "lines": [ + { + "bbox": [ + 313, + 430, + 537, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 537, + 601 + ], + "type": "table", + "html": "
TypeMethodNMSmAP
0.30.40.50.60.7Avg.
Anchor -basedBSN [15]*✓53.545.036.928.420.036.8
BMN [16]*✓56.047.438.829.720.538.5
BC-GNN [3]*✓57.149.140.431.223.140.2
G-TAD [31]*✓54.547.640.330.823.439.3
VSGN [34]*✓66.760.452.441.030.450.2
TCANet [23]*✓60.653.244.636.826.744.3
Anchor -freeAFSD [14]✓67.362.455.543.731.152.0
\\( MENet\\ \\left[35\\right]^‡ \\)✓70.765.358.849.134.055.6
TALLFormer [6]†✓76.0-63.2-34.559.2
ActionFormer [32]✓82.177.871.059.443.966.8
TriDet [27]✓83.680.172.962.447.469.3
Query -basedRTD-Net [28]✓68.362.351.938.823.749.0
ReAct [26]✓69.265.057.147.835.655.0
Self-DETR [11]Δ74.669.560.047.631.856.7
TadTR [18]Δ74.869.160.146.632.856.7
TE-TAD (Ours)×81.776.669.559.344.866.4
Ours w/ NMS✓83.378.471.360.745.667.9
", + "image_path": "e42b2a9d845e7cdf36525a6b048648972c1d2979285918148efe68ccf0f40fda.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 304, + 604, + 545, + 658 + ], + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 658 + ], + "type": "text", + "content": "Table 1. Performance comparison with state-of-the-art methods on THUMOS14. * and † denote TSN [30] and Swin-B [20] backbones, respectively. ‡ represents R(2+1)D [29]. Others employ I3D [5] backbone. The symbol △ indicates partial adoption of NMS." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 670, + 391, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 391, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 391, + 682 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "text", + "content": "THUMOS14 Table 1 provides a comparison with the state-of-the-art methods on THUMOS14. Our TE-TAD shows a" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18842" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 272, + 236 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 272, + 236 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 272, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 272, + 236 + ], + "type": "table", + "html": "
TypeMethodFeaturemAP
0.50.750.95Avg.
Anchor -basedBSN [15]TSN [30]46.530.08.030.0
BMN [16]TSN [30]50.134.88.333.9
BC-GNN [3]TSN [30]50.634.89.434.3
G-TAD [31]TSN [30]50.434.69.034.1
VSGN [34]TSN [30]52.436.08.435.1
TCANet [23]TSN [30]52.336.76.935.5
Anchor -freeAFSD [14]I3D [5]52.435.36.534.4
TALLFormer [6]Swin-B [20]54.136.27.935.6
ActionFormer [32]R(2+1)D [29]54.737.88.436.6
TriDet [27]R(2+1)D [29]54.738.08.436.8
MENet [35]R(2+1)D [29]54.738.410.537.7
Query -basedRTD-Net [28]TSN [30]47.230.78.630.8
ReAct [26]TSN [30]49.633.08.632.6
TadTR [18]R(2+1)D [29]53.637.510.636.8
TE-TAD (Ours)R(2+1)D [29]54.038.210.637.0
Ours w/ NMSR(2+1)D [29]54.238.110.637.1
", + "image_path": "9cb8a551a2e907d77ca8c0681854673372c343be46426c5f4899e19ee297ef32.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 275, + 288, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 288, + 396 + ], + "type": "text", + "content": "significant margin of improvement over other query-based detectors, even without NMS. While TadTR achieves an average mAP of " + }, + { + "bbox": [ + 46, + 275, + 288, + 396 + ], + "type": "inline_equation", + "content": "56.7\\%" + }, + { + "bbox": [ + 46, + 275, + 288, + 396 + ], + "type": "text", + "content": " with partial NMS utilization through their proposed cross window fusion (CWF) denoted by " + }, + { + "bbox": [ + 46, + 275, + 288, + 396 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}" + }, + { + "bbox": [ + 46, + 275, + 288, + 396 + ], + "type": "text", + "content": ", our method without NMS achieves a superior performance average mAP of " + }, + { + "bbox": [ + 46, + 275, + 288, + 396 + ], + "type": "inline_equation", + "content": "66.4\\%" + }, + { + "bbox": [ + 46, + 275, + 288, + 396 + ], + "type": "text", + "content": ". This significant improvement indicates the length-invariant capability of our TE-TAD model because THUMOS14 contains extremely diverse lengths of videos. Furthermore, even compared to anchor-free detectors, our method demonstrates competitive performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 579 + ], + "type": "text", + "content": "ActivityNet v1.3 Following the conventional approach [18, 27, 32], the external classification score is used to evaluate ActivityNet v1.3. The pre-extracted classification scores are combined with predictions from binary detectors to obtain class labels. Table 2 presents a performance comparison of our TE-TAD with state-of-the-art approaches. While our TE-TAD method exhibits a slight improvement in query-based detectors on ActivityNet v1.3 compared to the significant gains shown in Table 1, this is reflective of the intrinsic characteristics of the dataset. ActivityNet v1.3 does not contain the diverse length of the video relative to THUMOS14. Despite this different condition, our approach still demonstrates an improvement on ActivityNet v1.3, showing performance improvement even though it does not align with the primary issues our TE-TAD aims to resolve." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "EpicKitchens Table 3 shows the comparison with state-of-the-art methods on EpicKitchens. Our model achieves competitive performance without relying on NMS, indicating TE-TAD robustness in diverse and complex action detection scenarios. EpicKitchen contains an extremely diverse length of action instances, like THUMOS14. The results indicate the robustness of TE-TAD in handling a wide range of action lengths and complexities. The comparable performance of TE-TAD is meaningful in query-based approaches, a relatively less explored field than the more established anchor-free methods." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 319, + 70, + 531, + 209 + ], + "blocks": [ + { + "bbox": [ + 46, + 242, + 287, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 242, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 287, + 264 + ], + "type": "text", + "content": "Table 2. Comparison with state-of-the-art methods on ActivityNet v1.3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 70, + 531, + 209 + ], + "lines": [ + { + "bbox": [ + 319, + 70, + 531, + 209 + ], + "spans": [ + { + "bbox": [ + 319, + 70, + 531, + 209 + ], + "type": "table", + "html": "
TaskMethodNMSmAP
0.10.20.30.40.5Avg.
VerbBMN [16]✓10.88.88.47.15.68.4
G-TAD [31]✓12.111.09.48.16.59.4
ActionFormer [32]✓26.625.424.222.319.123.5
ASL [25]✓27.9-25.5-19.824.6
TriDet [27]✓28.627.426.124.220.825.4
TE-TAD (Ours)✗27.025.924.622.920.024.1
Ours w/ NMS✓27.926.825.423.420.024.7
NounBMN [16]✓10.38.36.24.53.46.5
G-TAD [31]✓11.010.08.67.05.48.4
ActionFormer [32]✓25.224.122.720.517.021.9
ASL [25]✓26.0-23.4-17.722.6
TriDet [27]✓27.426.324.622.218.323.8
TE-TAD (Ours)✗26.024.823.220.818.322.6
Ours w/ NMS✓26.325.223.221.018.222.8
", + "image_path": "7e456a09681ec079eb22afe1fa365bb4d897a00edb03635932743416ddb415c8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 337, + 237, + 513, + 358 + ], + "blocks": [ + { + "bbox": [ + 306, + 212, + 544, + 233 + ], + "lines": [ + { + "bbox": [ + 306, + 212, + 544, + 233 + ], + "spans": [ + { + "bbox": [ + 306, + 212, + 544, + 233 + ], + "type": "text", + "content": "Table 3. Comparison with state-of-the-art methods on EpicK-itchens. All methods employ SlowFast [9] as a backbone." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 237, + 513, + 358 + ], + "lines": [ + { + "bbox": [ + 337, + 237, + 513, + 358 + ], + "spans": [ + { + "bbox": [ + 337, + 237, + 513, + 358 + ], + "type": "table", + "html": "
TypeMethodNMSmAP@AVG
Anchor-freeActionFormer [32]X43.2 (-23.6)
✓66.8
TriDet [27]X44.9 (-24.4)
✓69.3
Query-basedReAct [26]X19.8 (-35.7)
✓55.0
TadTR [18]X53.1 (-3.6)
Δ56.7
TE-TAD (Ours)X66.4 (-1.5)
✓67.9
", + "image_path": "1c41f18f8b0ee52dc200b2a1be07b89ae0b67dc6c50bf6a4af718a8fc15d95c4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 362, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 305, + 362, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 362, + 545, + 407 + ], + "type": "text", + "content": "Table 4. Effect of NMS on the mAP across various anchor-free and query-based methods on THUMOS14. The value in parentheses represents the decrease in mAP when NMS is not applied. The symbol " + }, + { + "bbox": [ + 305, + 362, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 305, + 362, + 545, + 407 + ], + "type": "text", + "content": " indicates partial adoption of NMS." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 411, + 408, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 411, + 408, + 423 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 408, + 423 + ], + "type": "text", + "content": "4.3. Further Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 425, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 640 + ], + "type": "text", + "content": "Impact of NMS In Table 4, we evaluate how NMS influences the performance of various TAD methods on THUMOS14. NMS is particularly crucial for anchor-free detectors, which employ a one-to-many assignment strategy that leads to duplicated predictions for the same instance. Furthermore, even though ReAct [26] is a query-based approach, removing NMS at the ReAct significantly affects the performance. This is why ReAct adopts partial self-attention in the decoder called relational queries, which does not address whole queries. As shown in Table 6, our approach also drops the performance without NMS when removing the decoder's self-attention layer. This result indicates that addressing whole queries by the decoder's self-attention is crucial to preserving the set-prediction mechanism and full end-to-end modeling. Furthermore, our proposed method exhibits a minimal performance decrement of only -1.5 when NMS is excluded. This indicates that our approach effectively achieves full end-to-end modeling." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Component Contribution Analysis Table 5 shows the incremental impact of each key component in our TE-TAD on THUMOS14. The performance is measured without NMS. We conduct experiments based on full end-to-end TadTR, referred to as TadTR-∞ in Sec. 3.3. Starting with the TadTR-∞ baseline, incorporating multi-scale features and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18843" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 70, + 251, + 159 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 251, + 159 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 251, + 159 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 251, + 159 + ], + "type": "table", + "html": "
BaselineMulti-ScaleTwo-stage [35]TEAQSmAP@AVG
TadTR-34.1s w/ NMS✓56.7
✓56.5
✓✓57.3
57.0
TadTR-∞ w/o NMS✓40.2
✓42.6
✓✓43.3
✓✓59.5
✓✓46.1
✓✓✓63.6
✓✓✓✓66.4
", + "image_path": "05c83ac846bb691def57dea19545c55f7d2d89278f09229a7ded77674bfb713f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 85, + 184, + 247, + 243 + ], + "blocks": [ + { + "bbox": [ + 47, + 161, + 287, + 183 + ], + "lines": [ + { + "bbox": [ + 47, + 161, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 47, + 161, + 287, + 183 + ], + "type": "text", + "content": "Table 5. Analysis of contributions of each component on THUMOS14. The all-empty check is the denoted baseline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 184, + 247, + 243 + ], + "lines": [ + { + "bbox": [ + 85, + 184, + 247, + 243 + ], + "spans": [ + { + "bbox": [ + 85, + 184, + 247, + 243 + ], + "type": "table", + "html": "
EncoderDecodermAP@AVG
Self-attn.Self-attn.Cross-attn.w/o NMSw/ NMS
#1✓✓61.263.8
#2✓✓53.063.4
#3✓✓0.10.2
#4✓✓✓66.167.7
", + "image_path": "6826bbb8fbd4a4083c9ed1f9a74cce62780a9aa5451f6ebb152ef813dce94d66.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 247, + 287, + 280 + ], + "lines": [ + { + "bbox": [ + 47, + 247, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 287, + 280 + ], + "type": "text", + "content": "Table 6. Analysis of the specific roles of self-attention and cross-attention layers in the encoder and decoder, and their impact with and without utilizing NMS." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 91, + 281, + 243, + 355 + ], + "blocks": [ + { + "bbox": [ + 91, + 281, + 243, + 355 + ], + "lines": [ + { + "bbox": [ + 91, + 281, + 243, + 355 + ], + "spans": [ + { + "bbox": [ + 91, + 281, + 243, + 355 + ], + "type": "image", + "image_path": "03f94a0b93f3c9b722df3ce22c0d198715641f942fa5a6576098ebe4c33eecfa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 356, + 286, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 356, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 286, + 376 + ], + "type": "text", + "content": "Figure 6. Comparison of false negative rates by instance lengths (XS, S, M, L, XL) on THUMOS14." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 378, + 286, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 286, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 286, + 473 + ], + "type": "text", + "content": "a two-stage approach shows slight improvements. Nonetheless, these adoptions of improved methods for query-based detectors do not reach the performance level of the original TadTR. However, including our time-aligned expression (TE) significantly improves performance, demonstrating the value of our time-aligned representations. Finally, incorporating the adaptive query selection (AQS) mechanism shows the highest performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 474, + 286, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 286, + 640 + ], + "type": "text", + "content": "Role of Each Attention To clarify the role of encoder and decoder architecture, we conduct experiments for removing the attention layers of the encoder and decoder. Table 6 shows the individual contributions of the encoder and decoder layers' attention. As shown in Row #1, removing self-attention in the encoder slightly decreases the detection performance, indicating the encoder's self-attention affects representational ability. Row #2 reveals that the decoder's self-attention is the core role of the set-prediction mechanism by showing the performance degradation without using NMS. Finally, Row #3 shows that removing the decoder's cross-attention does not work because only location information is provided to the decoder, which does not capture the content information without cross-attention." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "content": "False Negative Analysis To further compare with the baseline method, we evaluate the false negative rate on THUMOS14. Fig. 6 shows the false negative rates across varying action instance sizes: extra small (XS), small (S), medium (M), large (L), and extra large (XL) based on DETAD [1]. These results show that even though TadTR- " + }, + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "content": " shows the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 545, + 180 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 545, + 180 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 545, + 180 + ], + "type": "image", + "image_path": "fc111cc3333e8fe69c840e52cf9a74c9293f2227b45374c5b96707734a7141af.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 181, + 545, + 246 + ], + "lines": [ + { + "bbox": [ + 304, + 181, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 545, + 246 + ], + "type": "text", + "content": "Figure 7. Comparison between query selection method. Ground truth (GT) is illustrated in the first row. The second row shows encoder prediction scores " + }, + { + "bbox": [ + 304, + 181, + 545, + 246 + ], + "type": "inline_equation", + "content": "p^{(0)}" + }, + { + "bbox": [ + 304, + 181, + 545, + 246 + ], + "type": "text", + "content": ". The third and fourth row visualizes the generated queries at the top 50 points and 10 points for each sector by AQS, respectively. Different colors in the fourth row represent proposals selected by different sectors." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 251, + 545, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 309 + ], + "type": "text", + "content": "worse overall performance in mAP, TadTR-" + }, + { + "bbox": [ + 304, + 251, + 545, + 309 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 304, + 251, + 545, + 309 + ], + "type": "text", + "content": " more captures the XL cases, indicating the shorter feature coverage cannot capture the long duration of instances. These results show that our method significantly reduces false negatives, particularly in XS and XL cases." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 311, + 546, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 546, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 546, + 490 + ], + "type": "text", + "content": "Effetiveness of Adaptive Query Selection Fig. 7 compares the query selection method by denoting the center point " + }, + { + "bbox": [ + 304, + 311, + 546, + 490 + ], + "type": "inline_equation", + "content": "c^{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 311, + 546, + 490 + ], + "type": "text", + "content": " of selected queries. Selected points both w/o AQS and w/AQS are generated by the same encoder output scores. As illustrated in Fig. 7, the fixed top- " + }, + { + "bbox": [ + 304, + 311, + 546, + 490 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 311, + 546, + 490 + ], + "type": "text", + "content": " proposal method misses the fourth ground location as a foreground candidate. When a foreground proposal is too distant from actual action instances, it necessitates extensive refinement from the decoder. Refining the segments from missed candidates leads to an additional workload for the decoder layers. In contrast, our AQS method successfully captures the part missed by w/o AQS. By dividing the video into sectors and selecting queries within these local units, AQS dynamically adjusts the number of queries and more accurately detects foreground candidates." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 495, + 378, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 378, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 378, + 506 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 511, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 511, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 545, + 643 + ], + "type": "text", + "content": "In this paper, we propose a full end-to-end temporal action detection transformer that integrates time-aligned coordinate expression, called TE-TAD, which eliminates reliance on hand-crafted components such as the sliding window and NMS. By aligning coordinate expression with the actual video timeline, our model not only simplifies the detection process but also significantly enhances the performance of query-based detectors. Furthermore, our TE-TAD has a length-invariant property by combining the proposed time-aligned coordinate expression and adaptive query selection, showing the potential of query-based detectors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 646, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 646, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 646, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2019-0-00079, Artificial Intelligence Graduate School Program (Korea University), No. 2021-0-02068, Artificial Intelligence Innovation Hub, and No. 2022-0-00984, Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18844" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Humam Alwassel, Fabian Caba Heilbron, Victor Escorchia, and Bernard Ghanem. Diagnosing error in temporal action detectors. In Proceedings of the European Conference on Computer Vision, pages 256-272, 2018. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "text", + "content": "[2] Humam Alwassel, Silvio Giancola, and Bernard Ghanem. Tsp: Temporally-sensitive pretraining of video encoders for localization tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 3173-3183, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "text", + "content": "[3] Yueran Bai, Yingying Wang, Yunhai Tong, Yang Yang, Qiyue Liu, and Junhui Liu. Boundary content graph neural network for temporal action proposal generation. In Proceedings of the European Conference on Computer Vision, pages 121-137, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision, pages 213-229, 2020. 1, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "type": "text", + "content": "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 288, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 389 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 389 + ], + "type": "text", + "content": "[6] Feng Cheng and Gedas Bertasius. Tallformer: Temporal action localization with long-memory transformer. In Proceedings of the European Conference on Computer Vision, pages 503-521, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 456 + ], + "type": "text", + "content": "[7] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, and Will Price. The epic-kitchens dataset: Collection, challenges and baselines. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):4125-4141, 2020. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 288, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 288, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 288, + 512 + ], + "type": "text", + "content": "[8] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 961-970, 2015. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 288, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 288, + 556 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 288, + 556 + ], + "type": "text", + "content": "[9] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6202-6211, 2019. 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "type": "text", + "content": "[10] Haroon Idrees, Amir R Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The thumos challenge on action recognition for videos “in the wild”. Computer Vision and Image Understanding, 155:1-23, 2017. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 644 + ], + "type": "text", + "content": "[11] Jihwan Kim, Miso Lee, and Jae-Pil Heo. Self-feedback detr for temporal action detection. In ICCV, pages 10286-10296, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 678 + ], + "type": "text", + "content": "[12] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6, 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[13] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 162 + ], + "type": "text", + "content": "[14] Chuming Lin, Chengming Xu, Donghao Luo, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Learning salient boundary feature for anchor-free temporal action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3320-3329, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "text", + "content": "[15] Tianwei Lin, Xu Zhao, Haisheng Su, Chongjing Wang, and Ming Yang. Bsn: Boundary sensitive network for temporal action proposal generation. In Proceedings of the European Conference on Computer Vision, pages 3-19, 2018. 1, 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 219, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 274 + ], + "type": "text", + "content": "[16] Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, and Shilei Wen. Bmn: Boundary-matching network for temporal action proposal generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3889-3898, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "text", + "content": "[17] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 2980-2988, 2017. 6, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 320, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 365 + ], + "type": "text", + "content": "[18] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 2022. 1, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "text", + "content": "[19] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "type": "text", + "content": "[20] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 467, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 521 + ], + "type": "text", + "content": "[21] Fuchen Long, Ting Yao, Zhaofan Qiu, Xinmei Tian, Jiebo Luo, and Tao Mei. Gaussian temporal awareness networks for action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 344-353, 2019. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "text", + "content": "[22] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 557, + 547, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 547, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 547, + 623 + ], + "type": "text", + "content": "[23] Zhiwu Qing, Haisheng Su, Weihao Gan, Dongliang Wang, Wei Wu, Xiang Wang, Yu Qiao, Junjie Yan, Changxin Gao, and Nong Sang. Temporal context aggregation network for temporal action proposal refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 485-494, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "text", + "content": "[24] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 658-666, 2019. 6, 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[25] Jiayi Shao, Xiaohan Wang, Ruijie Quan, Junjun Zheng, Jiang Yang, and Yi Yang. Action sensitivity learning for temporal" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18845" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "action localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "text", + "content": "[26] Dingfeng Shi, Yujie Zhong, Qiong Cao, Jing Zhang, Lin Ma, Jia Li, and Dacheng Tao. React: Temporal action detection with relational queries. In Proceedings of the European Conference on Computer Vision, pages 105–121, 2022. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "text", + "content": "[27] Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, and Dacheng Tao. Tridet: Temporal action detection with relative boundary modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18857-18866, 2023. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "text", + "content": "[28] Jing Tan, Jiaqi Tang, Limin Wang, and Gangshan Wu. Relaxed transformer decoders for direct action proposal generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13526-13535, 2021. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "text", + "content": "[29] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 376 + ], + "type": "text", + "content": "[30] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In Proceedings of the European Conference on Computer Vision, pages 20-36, 2016. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 431 + ], + "type": "text", + "content": "[31] Mengmeng Xu, Chen Zhao, David S Rojas, Ali Thabet, and Bernard Ghanem. G-tad: Sub-graph localization for temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10156-10165, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 433, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 287, + 476 + ], + "type": "text", + "content": "[32] Chenlin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In Proceedings of the European Conference on Computer Vision, pages 492-510, 2022. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 478, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 533 + ], + "type": "text", + "content": "[33] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In Proceedings of the European Conference on Computer Vision, 2022. 4, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "text", + "content": "[34] Chen Zhao, Ali K Thabet, and Bernard Ghanem. Video self-stitching graph network for temporal action localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13658-13667, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "text", + "content": "[35] Zixuan Zhao, Dongqi Wang, and Xu Zhao. Movement enhancement toward multi-scale video feature representation for temporal action detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13555-13564, 2023. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[36] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12993-13000, 2020. 6, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[37] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "text", + "content": "for end-to-end object detection. In International Conference on Learning Representations, 2020. 4, 6, 1" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18846" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/TEA_ Test-time Energy Adaptation/c8c44200-6095-44b9-b188-9639c747d9ec_content_list.json b/2024/TEA_ Test-time Energy Adaptation/c8c44200-6095-44b9-b188-9639c747d9ec_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5eff25871a7288a58597a89f7f66f72a1367c7b2 --- /dev/null +++ b/2024/TEA_ Test-time Energy Adaptation/c8c44200-6095-44b9-b188-9639c747d9ec_content_list.json @@ -0,0 +1,1555 @@ +[ + { + "type": "text", + "text": "TEA: Test-time Energy Adaptation", + "text_level": 1, + "bbox": [ + 305, + 130, + 665, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yige Yuan $^{1,2}$ , Bingbing Xu $^{1*}$ , Liang Hou $^{3}$ , Fei Sun $^{1}$ , Huawei Shen $^{1,2}$ , Xueqi Cheng $^{1,2}$ $^{1}$ CAS Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences \n $^{2}$ University of Chinese Academy of Sciences Kuaishou Technology", + "bbox": [ + 86, + 179, + 897, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yuanyige20z, xubingbing, sunfei, shenhuawei, cxq}@ict.ac.cn, lianghou96@gmail.com", + "bbox": [ + 127, + 234, + 851, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 285, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Test Time Adaptation (TTA) aims to improve model generalizability when test data diverges from training distribution, with the distinct advantage of not requiring access to training data and processes, especially valuable in the context of pre-trained models. However, current TTA methods fail to address the fundamental issue: covariate shift, i.e., the decreased generalizability can be attributed to the model's reliance on the marginal distribution of the training data, which may impair model calibration and introduce confirmation bias. To address this, we propose a novel energy-based perspective, enhancing the model's perception of target data distributions without requiring access to training data or processes. Building on this perspective, we introduce Test-time Energy Adaptation (TEA), which transforms the trained classifier into an energy-based model and aligns the model's distribution with the test data's, enhancing its ability to perceive test distributions and thus improving overall generalizability. Extensive experiments across multiple tasks, benchmarks and architectures demonstrate TEA's superior generalization performance against state-of-the-art methods. Further in-depth analyses reveal that TEA can equip the model with a comprehensive perception of test distribution, ultimately paving the way toward improved generalization and calibration1.", + "bbox": [ + 76, + 318, + 473, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 713, + 209, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks, despite their remarkable performance under the assumption of independent and identically distributed (i.i.d.) training and test data [26, 33, 59], significantly degrade in real-world scenarios where unseen test data diverges from the training distribution. This limitation, known as distribution shift or domain shift, emphasizes the pressing need for generalizability across shifted test distributions [49, 57, 61, 67]. To tackle this issue, recent studies", + "bbox": [ + 76, + 739, + 468, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa6c2f82dc88978e89460dcdfffbb24da0097b6b96bbc49a6d228c5660aae940.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 287, + 870, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dda722254ab59e1395cee0a0e4b27cd5e972c8e1c2a0cc2110341aa13e671888.jpg", + "image_caption": [ + "Figure 1. Performance vs. energy on model trained with original distribution, tested across various shifted distributions. Upper: error rate change within energy score groups. Lower: loss variation with energy scores, each point denoting a distribution. Marker styles and opacity reflect distribution types and divergence." + ], + "image_footnote": [], + "bbox": [ + 522, + 417, + 870, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of Test Time Adaptation (TTA) [37, 60, 70] began considering the incorporation of unlabeled test data and leveraging it to fine-tune the source model in an unsupervised manner. This paradigm has garnered significant attention due to its ability to operate without access to training data or involvement with the training process. In the era of large open-source models [9, 58, 63], where models are publicly available but the training data and training process remain inaccessible due to privacy and resource restrictions [6, 23], TTA emerges as especially beneficial and practical.", + "bbox": [ + 496, + 641, + 892, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing TTA methods can be broadly categorized into three classes [37]. Normalization-based methods [41, 52] adjust the BatchNorm statistics of the model with test data statistics. Entropy-based methods [45, 46, 60] fine-tune the model by minimizing the prediction entropy. Pseudolabeling-based methods [34, 36] utilize test-time generated labels for updates. While these methods have been empi", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 875, + 222, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code is available at https://github.com/yuanyige/tea", + "bbox": [ + 94, + 887, + 452, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23901", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cally effective, these methods all fail to address a fundamental issue: covariate shift. Specifically, the decrease in generalization ability on test data with distribution shift can be attributed to the model's reliance on the marginal distribution of the training data. However, previous TTA methods do not address this shift due to their lack of connection with marginal distributions, impairing model calibration [14, 24] and introducing confirmation bias [1].", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To combat the above challenges, we propose a novel way rooted in an energy-based perspective. Within this way, energy is defined as an unnormalized probability assigned to a sample, where a lower score corresponds to a higher likelihood of that sample within a distribution [32, 54]. Proposing such a way to improve test time adaptation is twofold.", + "bbox": [ + 75, + 215, + 468, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Firstly, test samples that correspond to lower energy within the model's distribution tend to exhibit higher performance. This is demonstrated by examining the energy scores of various test datasets in relation to a model trained on a specific training distribution. As depicted in Fig. 1, an increase in the divergence between the test and training distributions is accompanied by a drastic escalation in energy scores, leading to a significant degradation in performance. Secondly, the energy-based way can address covariate shift under TTA via directly injecting the model with a comprehensive perception of test distribution. Addressing covariate shift in TTA is particularly challenging, as it is neither feasible to access the training dataset to align the marginal distribution between training and testing [53], nor possible to modify the training process to mitigate the influence of marginal training distribution [48]. Under such circumstances, the energy-based way can directly manipulate the trained model's likelihood landscape [39] via an implicit distribution modeling process without requiring training process and training data, becoming a promising way. This stands in contrast to other models such as GANs [12, 19], Flows [50], and VAEs [28] which are advantageous only when the training data are accessible.", + "bbox": [ + 75, + 309, + 470, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on the above energy-based way, we propose Test-time Energy Adaptation, abbreviated as TEA, which constructs an energy-based model from the trained (classifier) model by reinterpreting the negative log-sum-exp of logits as an energy function, and employs Contrastive Divergence [18] as the adaptation objective to decrease the energy of test samples while increase that of samples generated by Stochastic Gradient Langevin Dynamics [62]. This approach prevents a trivial solution that indiscriminately reduces the energy across the entire data space to ensure an increased likelihood for target test samples within the model's distribution. TEA enables a gradual alignment between the distributions of the trained model and the test data, bolstering the trained model's perception of the test distribution and paving the way for superior adaptability and performance when confronted with the corresponding test data.", + "bbox": [ + 75, + 659, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We investigate the effectiveness of TEA under image corruption and domain generalization on four popular benchmarks CIFAR-10, CIFAR-100, TinyImageNet and PACS, across three architectures WRN-28-10, ResNet50 and ResNet-18. Experimental results underscore that TEA significantly outperforms current best-performing TTA methods in terms of generalizability, with an average increment of $4.7\\%$ . We further reveal that TEA can equip the model with a comprehensive perception of the test distribution. This, in turn, significantly improves the generalization and calibration of the trained model.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions include:", + "bbox": [ + 517, + 260, + 732, + 273 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Promising Way: We propose a new energy-based way for test time adaptation, which marks a departure from traditional methods and sheds light on potential avenues for mitigating the impact of distribution shifts.", + "- Innovative Method: We propose TEA to decrease the energy of the test data within the model's distribution, thereby equipping the trained model with a perception of the test distribution and enhancing generalizability.", + "- Extensive Experiments: Experiments across extensive settings validate TEA's superiority over current leading methods. Further in-depth analyses extend the understanding of how energy reduction enhances the model's perception of test distribution, ultimately paving the way toward improved generalization and calibration." + ], + "bbox": [ + 500, + 277, + 890, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 510, + 640, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Test Time Adaptation Test Time Adaptation (TTA) [37] is a paradigm aiming to enhance a model's generalizability on specific test data through unsupervised fine-tuning with these data. Note that the model is originally trained on a distinct training dataset, which is not available during the adaptation phase. Approaches like TTT [55] adapt models through self-supervised proxy task during testing but require the training of the same proxy task during training procedure. DDA [11, 65] explores adapting the test data, yet faces limitations due to model structure and training constraints. Recent research [60] highlights a scenario where the training process and training data is entirely agnostic, leading to three main categories of approaches: For normalization-based, BN [52] adapts the BatchNorm [22] statistics with test data. DUA [41] uses a tiny fraction of test data and its augmentation for BatchNorm statistics adaptation. For entropy-based, TENT [60] fine-tunes BatchNorm layers using entropy minimization during the test phase. EATA [45] employs a Fisher regularizer to limit excessive model parameter changes. SAR [46] removes high-gradient samples and promotes flat minimum weights. For pseudolabeling-based, PL [34] fine-tunes parameters using confident pseudo labels. SHOT [36] combines entropy minimization methods with pseudo labeling.", + "bbox": [ + 496, + 537, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23902", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/198c4a0b0b4d7b4f169e3b9371d315286bb65caca129d77218ecc0970e42fc8d.jpg", + "image_caption": [ + "Figure 2. Overview of Test-time Energy Adaptation (TEA). Given a trained model (classifier) and in-coming test data, TEA directly integrates test data distribution into the trained classifier by fine-tuning its normalization layers through energy-based training: TEA constructs an Energy-Based Model from the classifier by reinterpreting the negative log-sum-exp of logits as an energy function, and employs Contrastive Divergence as the adaptation objective which decreases the energy of test samples while increases the energy of negative samples generated by Langevin Dynamics. This adaptation increases the likelihood of test samples under the classifier's distribution, enabling a gradual alignment between the distributions of the trained classifier and the test data, thereby enhancing generalizability." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 877, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Energy Based Model Energy-Based Models (EBMs) are a type of non-normalized probabilistic models. Unlike most other probabilistic models, EBMs do not necessitate the normalizing constant to be tractable [32, 54] and do not require an explicit neural network for sample generation, implying the generation process is implicit [7]. These lead to increased flexibility in parameterization and allow for modeling a wider range of probability distributions. Due to their flexibility, EBMs can construct hybrid models with both discriminative and generative capabilities, integrating the generative competencies into discriminative models without sacrificing their discriminative capabilities [7, 13, 15]. Among these, JEM [13] is particularly representative, reinterpreting classifiers as an EBM and achieving impressive results in both classification and generation.", + "bbox": [ + 75, + 429, + 472, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 671, + 168, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we detail our method, Test-time Energy Adaptation (TEA). Initially, we present a thorough description of the notation and overall architecture in Sec. 3.1, after which we proceed to explain energy adaption (Sec. 3.2) and modulation parameters (Sec. 3.3), respectively. Furthermore, we engage in a discussion about the difference between our method and entropy-based adaptation in Sec. 3.4.", + "bbox": [ + 75, + 698, + 468, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Notation and Overall Architecture", + "text_level": 1, + "bbox": [ + 76, + 816, + 377, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The labeled training dataset is denoted as $\\{(\\mathbf{x}_{\\mathrm{train}},y_{\\mathrm{train}})\\} \\subset \\mathcal{X}\\times \\mathcal{Y}$ and the unlabeled test data are represented by $\\mathbf{x}_{\\mathrm{test}}\\in$ $\\mathcal{X}$ , where $\\mathcal{X}$ and $\\mathcal{Y}$ are data and label spaces. The respective marginal distributions of the training and test data are given", + "bbox": [ + 76, + 839, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "by $\\mathbf{x}_{\\mathrm{train}}\\sim p_{\\mathrm{train}}(\\mathbf{x})$ and $\\mathbf{x}_{\\mathrm{test}}\\sim p_{\\mathrm{test}}(\\mathbf{x})$ . A classifier model trained on the training dataset and parameterized by $\\theta$ , is denoted as $f_{\\theta}:\\mathcal{X}\\to \\mathcal{Y}$ . The data distribution learned by this trained classifier is denoted by $p_{\\theta}(\\mathbf{x})$ , which will be referred to as the model distribution henceforth.", + "bbox": [ + 496, + 429, + 890, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework of TEA is depicted in Fig. 2. The motivation behind TEA is rooted in the issue of covariate shifts [25], where the degradation of model generalization on test data $\\mathbf{x}_{\\mathrm{test}}$ is attributed to the model's reliance on the training distribution $p_{\\mathrm{train}}(\\mathbf{x})$ . To overcome this issue without accessing the training data or training process, TEA directly integrates the test data distribution into the trained model. TEA constructs an energy-based model from the trained classifier by reinterpreting the negative log-sum-exp of logits as an energy function. Through this lens, TEA employs contrastive divergence [18] as the adaptation objective, which serves to decrease the energy (increase the likelihood) of the test samples under the model distribution $p_{\\theta}(\\mathbf{x})$ . This adaptation enables the gradual alignment of distribution between the trained model and test data, thereby bolstering the model's perception of the test distribution and enhancing generalizability. Following previous TTA methods, TEA freezes the majority of the model's parameters, permitting only minor adjustments for efficient adaptation.", + "bbox": [ + 496, + 503, + 892, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Energy Adaptation for Test Distribution", + "text_level": 1, + "bbox": [ + 500, + 801, + 844, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Enhancing the model's perception of test distribution from an energy-based perspective involves two key steps: constructing the energy-based model and optimizing it.", + "bbox": [ + 496, + 825, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Constructing the energy-based model. Energy-based models (EBMs) [13, 32, 54] represent a class of proba", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23903", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "bilistic models that are characterized by an energy function. Consider a sample $\\mathbf{x} \\in \\mathbb{R}^D$ , the energy function $E: \\mathbb{R}^D \\to \\mathbb{R}$ maps each sample into an energy that can be considered as an unnormalized probability, with lower scores indicating higher likelihoods [7]. Consequently, the probability density $p(\\mathbf{x})$ , as defined by EBM, can be expressed using the Boltzmann distribution [38], as shown in Eq. (1), where the partition function $Z = \\int \\exp(-E(\\mathbf{x})) \\, \\mathrm{d}\\mathbf{x}$ serves to normalize the probability density.", + "bbox": [ + 75, + 90, + 468, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (\\mathbf {x}) = \\exp (- E (\\mathbf {x})) / Z. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 231, + 468, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Constructing an energy-based model from a trained classifier $f_{\\theta}$ is founded on the fundamental analysis that an energy-based framework inherently underlies any discriminative model [13]. In this framework, the energy of one sample for a corresponding class can be represented as its logit produced by the trained classifier, denoted by $E_{\\theta}(\\mathbf{x},y) = -f_{\\theta}(\\mathbf{x})[y]$ . Therefore, the joint probability distribution of $\\mathbf{x}$ and $y$ can be defined as,", + "bbox": [ + 75, + 251, + 468, + 371 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (\\mathbf {x}, y) = \\exp \\left(f _ {\\theta} (\\mathbf {x}) [ y ]\\right) / Z _ {\\theta}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 376, + 468, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "then the distribution of $\\mathbf{x}$ can be obtained by marginalizing over $y$ , as shown below,", + "bbox": [ + 75, + 396, + 468, + 425 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (\\mathbf {x}) = \\sum_ {y} p _ {\\theta} (\\mathbf {x}, y) = \\sum_ {y} \\exp \\left(f _ {\\theta} (\\mathbf {x}) [ y ]\\right) / Z _ {\\theta}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 429, + 468, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By substituting Eq. (3) into Eq. (1), we can obtain the form of the energy function as follows:", + "bbox": [ + 75, + 465, + 468, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\theta} (\\mathbf {x}) = - \\log \\sum_ {y} \\exp (f _ {\\theta} (\\mathbf {x}) [ y ]). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 498, + 468, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the aforementioned steps, we repurpose and reinterpret the logits produced by the trained classifier to establish an energy-based model and define the energy function as the negative log-sum-exp of the logits.", + "bbox": [ + 75, + 534, + 468, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the logits reinterpretation, we can construct a energy-based model for the test data $\\mathbf{x}_{\\mathrm{test}}$ using the trained classifier, where $Z_{\\theta} = \\int \\sum_{y}\\exp \\left(f_{\\theta}(\\mathbf{x})[y]\\right)\\mathrm{d}\\mathbf{x}$", + "bbox": [ + 75, + 595, + 468, + 642 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {x} _ {\\text {t e s t}}\\right) = \\frac {\\exp \\left(- E _ {\\theta} \\left(\\mathbf {x} _ {\\text {t e s t}}\\right)\\right)}{Z _ {\\theta}} = \\frac {\\sum_ {y} \\exp \\left(f _ {\\theta} \\left(\\mathbf {x} _ {\\text {t e s t}}\\right) [ y ]\\right)}{Z _ {\\theta}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 645, + 468, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimizing the energy-based model. Optimizing Eq. (5) is challenging as the partition function $Z_{\\theta}$ necessitates integration across the whole input space of $\\mathbf{x}$ , typically making it computationally intractable. Thus, direct maximizing the log-likelihood of test data $\\log p_{\\theta}(\\mathbf{x}_{\\mathrm{test}})$ presents significant difficulties when training the parameter $\\theta$ of our energy-based model. To overcome this difficulty, we propose to use contrastive divergence [3, 18] via estimating the gradient of the log-likelihood,", + "bbox": [ + 75, + 681, + 468, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\log p _ {\\theta} (\\mathbf {x} _ {\\text {t e s t}})}{\\partial \\theta} = \\mathbb {E} _ {\\tilde {\\mathbf {x}} \\sim p _ {\\theta}} \\left[ \\frac {\\partial E _ {\\theta} (\\tilde {\\mathbf {x}})}{\\partial \\theta} \\right] - \\frac {\\partial E _ {\\theta} (\\mathbf {x} _ {\\text {t e s t}})}{\\partial \\theta}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 819, + 468, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Eq. (6), the notation $\\tilde{\\mathbf{x}}\\sim p_{\\theta}$ denotes a random sample drawn from the distribution over $\\mathbf{x}$ , which is defined by the model's distribution. The sampling procedure can be", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Test-time Energy Adaptation" + ], + "code_body": "Input: $f_{\\theta},N,T,p_0,\\alpha ,\\beta ,\\mathbf{x}_{\\mathrm{test}}$ Output: $f_{\\theta}(\\mathbf{x}_{\\mathrm{test}})$ \n1 $E_{\\theta}(\\cdot)\\gets -\\log \\sum_{y}\\exp (f_{\\theta}(\\cdot)[y])$ \n2 for $i\\gets 0,1,\\dots ,N - 1$ do \n3 $\\tilde{\\mathbf{x}}_0\\gets \\mathrm{sample}(p_0)$ \n4 for $t\\gets 0,1,\\ldots ,T - 1$ do \n5 $\\begin{array}{r}\\epsilon \\leftarrow \\mathrm{sample}(\\mathcal{N}(\\mathbf{0},\\mathbf{I})) \\end{array}$ \n6 $\\tilde{\\mathbf{x}}_{t + 1}\\gets \\tilde{\\mathbf{x}}_t - \\frac{\\alpha}{2}\\frac{\\partial E_\\theta(\\tilde{\\mathbf{x}}_t)}{\\partial\\tilde{\\mathbf{x}}_t} +\\alpha \\epsilon$ \n7 end \n8 $\\tilde{\\mathbf{x}}\\gets \\tilde{\\mathbf{x}}_{T - 1}$ \n9 $\\theta \\gets \\theta -\\beta \\nabla_{\\theta}[E_{\\theta}(\\mathbf{x}_{\\mathrm{test}}) - E_{\\theta}(\\tilde{\\mathbf{x}})]$ \n10 end \n11 return $f_{\\theta}(\\underline{\\mathbf{x}}_{\\mathrm{test}})$", + "bbox": [ + 500, + 111, + 781, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "performed through Stochastic Gradient Langevin Dynamics (SGLD) [62], which iteratively generates samples by using the gradient information [13, 44]. In this context, $p_0$ represents an initial noise distribution, $\\alpha$ denotes the step-size, and $t = 0, 1, \\dots, T - 1$ is the time step. After $T$ steps of updating, a fictitious sample, generated by the energy-based model governed by the classifier, can be obtained.", + "bbox": [ + 496, + 340, + 890, + 445 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {x}} _ {t + 1} = \\mathbf {x} _ {t} - \\frac {\\alpha}{2} \\frac {\\partial E _ {\\theta} \\left(\\tilde {\\mathbf {x}} _ {t}\\right)}{\\partial \\tilde {\\mathbf {x}} _ {t}} + \\sqrt {\\alpha} \\boldsymbol {\\epsilon}, \\quad \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 454, + 890, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "this sampling process in Eq. (7) essentially optimizes the sample by moving in the direction of energy reduction. Consequently, the objective of Eqs. (6) and (7) can be fundamentally understood as a min-max game (Eq. (8)), which minimizes the energy derived from the incoming test samples while concurrently amplifying the energy of fictitious samples obtained via SGLD from the classifier's distribution. Significantly, the latter plays a pivotal role in preventing a collapse towards a trivial solution where energy is indiscriminately minimized throughout the entire space, rather than focusing on the target test data.", + "bbox": [ + 496, + 492, + 890, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\theta} \\left[ \\min _ {\\tilde {\\mathbf {x}}} E _ {\\theta} (\\tilde {\\mathbf {x}}) - E _ {\\theta} \\left(\\mathbf {x} _ {\\text {t e s t}}\\right) \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 667, + 890, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By adapting through this objective, the classifier's distribution continually converges towards the test distribution. The convergence point of the min-max game is reached when the energy of samples drawn from the classifier's distribution matches the energy of the test samples. At this stage, the classifier reaches a low-energy steady state with respect to the test data. The likelihood landscape of the classifier then exhibits a higher probability for samples that are consistent with the test data distribution, and conversely, a lower probability for samples that deviate from it, leading to a comprehensive perception of the classifier towards this test distribution and ultimately enhancing its generalizability. The pseudocode for TEA can be found in Algorithm 1.", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23904", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Modulation Parameters", + "text_level": 1, + "bbox": [ + 76, + 90, + 297, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As outlined in Eq. (8), TEA requires updating the parameters of the trained model using the aforementioned energy adaptation to adjust to test data. In line with previous methods, we opt to update the parameters of the normalization layer due to the following two factors: (1) Practicality and Efficiency: In the era of large-scale models, the practice of fine-tuning a selected group of parameters has gained prominence [20]. For both practicality and efficiency, it's crucial to avoid updating all parameters as this process would be excessively time-consuming. Note that the parameters of the normalization layer account for a modest $1\\%$ of total model parameters [60], making their update become more manageable. (2) Direct impact on data distribution: The parameters within normalization layers possess the potential to capture intrinsic features of data, thereby exerting direct influence upon the corresponding data distribution. As evidenced by [21], simple modulation to the mean and variance in a generator's normalization layers can modify image style, underlining the normalization parameters' profound impact on data distribution. This aligns well with our goal of manipulating the energy of test data to make it compatible with the model distribution.", + "bbox": [ + 76, + 114, + 472, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Discussion", + "text_level": 1, + "bbox": [ + 76, + 458, + 194, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As entropy-based adaptation has been the representative adaptation method, we further discuss the difference between TEA and entropy-based adaptation. Intriguingly, TEA may have a connection with entropy-based adaptation, given that the negative entropy $\\mathrm{NegEntropy}(\\mathbf{x}) = \\sum_{i}x_{i}\\log x_{i}$ is the convex conjugate of the free energy $\\mathrm{LogSumExp}(\\mathbf{x}) = \\log \\sum_{i}\\exp (x_{i})$ , as established in the literature [2, 43]. Despite this connection, entropy-based methods, which apply softmax normalization in the label space $y$ and strive to minimize entropy, can result in diminished uncertainty within classification probabilities, leading to compromised model calibration. In contrast, TEA, utilizing the log-sum-exp function within the data space $\\mathbf{x}$ can not only effectively avoid the pitfalls associated with entropy-based methods, but also improve calibration by introduces uncertainty to each class. This conjecture has been substantiated through experiments (refer to Sec. 4.3.3).", + "bbox": [ + 76, + 481, + 472, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 753, + 202, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the following sections, we compare TEA with state-of-the-art methods across various tasks, benchmarks and architectures. Then, we delve into deeper understanding of our method by exploring its desirable properties and identifying significant components that contribute to its improvements. Due to the space limitations, more comprehensive experiments including full results on corruptions and other analyses, are provided in Appendix Sec. 7.", + "bbox": [ + 76, + 780, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 90, + 689, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Metrics We focus on two tasks to verify the performance of TEA: generalization on image corruption and domain generalization. Image corruption includes clean and corrupted datasets from CIFAR-10(C), CIFAR-100(C) and TinyImageNet-200(C) [17, 29, 31], incorporating 15 types of corruption at 5 severity levels. Domain generalization considers PACS [35], encompassing 4 domains (Photo, Art Painting, Cartoon, Sketch) across 7 categories. We follow [68] to use Accuracy and mean Corruption Error (mCE) [17] as evaluation metrics. The evaluations are conducted at both the most severe level and the average of all severity levels to ensure thorough analysis.", + "bbox": [ + 496, + 113, + 893, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Backbones and Baselines We use two architectures for image corruption: WideResNet-28-10 [69] with BatchNorm [22], and ResNet-50 [16] with GroupNorm [64], consistent with the implementations of TENT [60] and SAR [46]. We use ResNet-18 [16] for domain generalization. We evaluate our method against eight leading TTA methods across three categories: (1) Normalization-based methods: BN [52] and DUA [41]. (2) Entropy-based methods: TENT [60], ETA, EATA [45], and SAR [46]; (3) Pseudo-labeling-based methods: PL [34] and SHOT [36]. Source denotes the original model without any adaptation.", + "bbox": [ + 496, + 314, + 893, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation We implement methods based on PyTorch [47]. Consistency in model weight is ensured by the RobustBench protocol [5], which provides pretrained weights for the WideResNet-28-10 (BatchNorm) on CIFAR-10. In the case where RobustBench weights are unavailable, we train models in accordance with the guidelines specified in [69]. All adaptation employ Adam [27], except for SAR, which originally uses SAM [10] with SGD [51]. Baselines are replicated using their original hyper-parameters, except when these were unspecified. More details and setups are deferred to Appendix Sec. 8.", + "bbox": [ + 496, + 500, + 893, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Adaptation Results", + "text_level": 1, + "bbox": [ + 500, + 676, + 684, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section compares the generalizability of TEA with baselines in image corruption and domain generalization.", + "bbox": [ + 496, + 700, + 890, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Corruption As reported in Tab. 1, we conducted experiments on three benchmarks against eight baselines for corruption scenarios. TEA markedly surpasses all baselines on the vast majority of datasets and severity levels. Specifically, TEA outshines the best-performing baseline by an average of $4.7\\%$ at the most severe level. The only exception is on TinyImageNet across all levels, where TEA ranks second, merely trailing by a minimal margin $0.1\\%$ . We further incorporate the ResNet50 with GroupNorm for a broader validation. Tab. 2 indicates that TEA still delivers the best", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23905", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fcf32817d0e9e14fcb279c830c4af3fb683fe12ef780ae14de312d8ff47012ce.jpg", + "table_caption": [ + "Table 1. Comparisons of TEA and baselines for image corruption on CIFAR-10(C), CIFAR-100(C), and TinyImageNet(C) using WRN28-10 with BatchNorm. Accuracy and mCE are evaluated at the most severe level and across all levels with asterisk (*) indicating the results are taken from the original paper [56]. The best adaptation results are highlighted in boldface." + ], + "table_footnote": [], + "table_body": "
WRN-28-10 BatchNormCIFAR-10(C)CIFAR-100(C)Tiny-ImageNet(C)
Clean Acc (↑)Corr Severity 5 Acc (↑)mCE (↓)Corr Severity 1-5 Acc (↑)mCE (↓)Clean Acc (↑)Corr Severity 5 Acc (↑)mCE (↓)Corr Severity 1-5 Acc (↑)mCE (↓)Clean Acc (↑)Corr Severity 5 Acc (↑)mCE (↓)Corr Severity 1-5 Acc (↑)mCE (↓)
Source94.7756.47100.0073.45100.0081.7935.39100.0052.12100.0063.1921.21100.0034.13100.00
NormBN [52]93.9779.5652.6585.6360.0080.8360.0663.5468.1169.4245.0427.7493.4234.27100.96
DUA* [41]-80.1050.78------------
PseudoPL [34]93.7551.42106.9872.6299.3780.5253.4072.1264.5375.2947.8428.2691.2239.8391.67
SHOT [36]93.2574.7763.1982.3572.6180.5256.5368.0166.0073.2847.9529.1490.1640.0191.41
EntropyTENT [60]93.6681.4148.1386.7556.1780.1463.0959.4269.4767.8039.5426.3195.5232.03104.49
ETA [45]93.9679.5852.6485.6359.9980.6559.8264.5267.1772.4043.2027.2894.1233.46102.25
EATA [45]93.9679.5952.6285.6459.9880.6860.2463.7567.4871.6643.4227.2894.0933.47102.24
SAR [46]93.9779.7751.9485.8358.9780.8462.9559.3770.0165.9941.5828.2192.8234.60100.47
EnergyTEA94.0983.3443.6987.8852.0080.8865.1056.0771.2263.7251.6531.6787.9939.9692.12
", + "bbox": [ + 80, + 142, + 890, + 319 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8e762e1a917b032d640c8c6542395627b75b05462b3d317553a4534ad75fe229.jpg", + "table_caption": [ + "Table 2. Comparisons for image corruption on CIFAR-10(C), CIFAR-100(C), and TinyImageNet(C) using ResNet-50 with GroupNorm across all severity levels. Best results are in boldface." + ], + "table_footnote": [], + "table_body": "
ResNet50 GroupNormCIFAR-10(C)CIFAR-100(C)Tiny-ImageNet(C)
Acc (↑)mCE (↓)Acc (↑)mCE (↓)Acc (↑)mCE (↓)
Source78.71100.0054.98100.0026.64100.00
PseudoPL79.4394.7656.6896.0226.6099.92
SHOT81.9886.6558.3193.4529.1196.73
EntropyTENT77.29102.8856.3496.8826.6599.94
ETA78.68100.0956.7296.3729.2596.42
EATA78.70100.0256.7696.2829.2596.42
SAR78.7899.6555.2899.3327.0599.41
EnergyTEA83.0579.0959.6789.3230.4194.81
", + "bbox": [ + 81, + 376, + 468, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "performance in both average accuracy and mCE. These results underscore the efficacy of TEA in handling image corruption scenarios, ensuring its universality for model architecture or normalization techniques.", + "bbox": [ + 76, + 527, + 468, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Domain Generalization Tab. 3 provides a comparison between TEA and state-of-the-art TTA methods on the PACS dataset. It is evident that when trained on the photo and art domains, TEA exhibits substantial improvements over the best-performing baseline, achieving increases of $7.1\\%$ and $8.5\\%$ , respectively. Compared to photo and art, cartoon and sketch domains may exhibit greater domain discrepancies, posing significant challenges to model generalization. Despite these challenging conditions, in the adaptation from cartoon to art and the sketch domain, TEA achieved improvements when all baselines experienced significant declines. TENT shows decline of $14.73\\%$ compared to the source in sketch, whereas TEA bucks the trend and enhances performance by $4.5\\%$ , highlighting its stability in the face of severe domain shifts.", + "bbox": [ + 75, + 612, + 468, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results from both tasks highlight TEA's superior generalizability. These effectiveness may originate from the reduced energy, the enhanced distribution perception, and the improved calibration. In next section, we will delve into", + "bbox": [ + 76, + 839, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1f4f1e8b1c0f8ca455ad4e262a4c191dc65bec1c24cdf28045e1f65a997e8f4e.jpg", + "table_caption": [ + "Table 3. Single source domain generalization comparisons on PACS datasets using ResNet-18 with BatchNorm in terms of Accuracy. The best adaptation results are highlighted in boldface." + ], + "table_footnote": [], + "table_body": "
Source DomainMethodTarget DomainAvg
PhotoArtCartoonSketch
PhotoSource-26.7622.4016.6221.93
BN-26.6627.9415.9623.52
TENT-26.9529.8617.5424.78
EATA-26.6628.1115.9823.59
SAR-26.7128.4115.9823.70
SHOT-26.6129.8620.9225.80
TEA-28.8133.6220.4927.64
ArtSource49.04-36.4324.4836.65
BN46.65-28.2822.7332.55
TENT50.78-30.1224.6135.17
EATA46.83-29.3123.4233.19
SAR47.90-33.0226.2735.73
SHOT50.24-34.3029.3737.97
TEA56.29-38.5728.7141.19
CartoonSource42.6929.79-29.4733.98
BN28.6825.15-20.8724.90
TENT30.9623.34-22.6525.65
EATA28.8025.10-25.0426.31
SAR29.7025.78-21.5125.66
SHOT37.7222.66-23.1427.84
TEA36.0531.44-22.8830.12
SketchSource19.9418.7032.21-23.62
BN13.4717.1429.86-20.16
TENT13.5317.3829.52-20.14
EATA13.1717.3330.08-20.19
SAR13.2918.8029.95-20.68
SHOT19.7618.7530.46-22.99
TEA19.6421.2433.19-24.69
", + "bbox": [ + 504, + 376, + 890, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "these aspects for further analysis and discussion.", + "bbox": [ + 500, + 747, + 821, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Analysis and Discussion", + "text_level": 1, + "bbox": [ + 500, + 782, + 718, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we delve into the mechanisms driving TEA's effectiveness and explore its desirable properties. Specifically, we studied three key aspects of TEA: (1) The correlation between energy reduction and generalizability enhancement; (2) The distribution perception and generation capabilities; (3) The confidence calibration improvements.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23906", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4840f1e565db7c005955e6e04e2c736f2afd5422126a50a80b55b082c00d9e42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 88, + 352, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a841b33e495cc40209b4b86ca44091ef0f3eaba949279b03f6d2e0b81cbc7d0.jpg", + "image_caption": [ + "Figure 3. This illustration captures the energy reduction and generalizability enhancement achieved by TEA across CIFAR-10-C, CIFAR-100-C, and TinyImageNet-200-C, displayed from left to right. The upper set of graphs trace the evolution of energy score, corresponding loss and accuracy in response to incrementally increasing TEA adaptation steps. The lower set uncovers the extent of energy reduction and the consequent performance improvement before and after executing TEA adaptation, under different levels of distribution shift." + ], + "image_footnote": [], + "bbox": [ + 78, + 188, + 344, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/151379108d2937a8d16b7cdca4910fb1524568e9392b9898ed5beb5b19bad047.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 88, + 620, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7fa31240c1e245b963432bbcdaf3699bddb2ba78ffe8b16a01f2072dea6ce905.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 188, + 609, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d458ea1f2acae09b7abaf870e1c7d3dc107ef5a171d1dd624f17025b904badb3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 88, + 890, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/353a1543a5ce03d4aed30294a098ca05c48b571caad9451008f29809157e9442.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 188, + 880, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1 Relation between TEA's Energy Reduction and Generalizability Enhancement", + "text_level": 1, + "bbox": [ + 76, + 354, + 470, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This experiment validate TEA's energy reduction capability, and its impact on generalizability, spanning two scenarios: (1) In the first scenario, we focus on the trends of energy scores, loss, and accuracy on the same test data over increasing adaptation steps. (2) In the second scenario, we explore how the extent of energy reduction correlates with performance improvements, before and after adaptation, using varied test data that exhibit different distribution shifts.", + "bbox": [ + 75, + 395, + 468, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results from both scenarios are depicted in Fig. 3. It is observable that $(i)$ As the iteration step of TEA increases, there is a consistent reduction in energy and corresponding decrease in loss, coupled with an ongoing enhancement in accuracy. $(ii)$ As the distribution shift increases, TEA's energy reduction becomes more pronounced. Concurrently, the enhancement in performance over the baseline also increases. $(iii)$ As the distribution shift increases, there is a sharp degradation observed in the baseline performance. However, the model adapted via TEA maintains its stability and robustness, demonstrating resilience against strong distribution shifts. In summary, these aforementioned trends are consistently observed across three datasets, demonstrating TEA's significant effectiveness in reducing energy. Notably, a greater reduction in energy correlates with an increased improvement in the model's generalizability", + "bbox": [ + 75, + 517, + 470, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2 TEA's Distribution Perception and Generation", + "text_level": 1, + "bbox": [ + 76, + 784, + 455, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This experiment aims to validate the capability of TEA in modeling test data distribution. We framed the experiment within two scenarios: identical training and testing distributions, and shifted training and testing distributions. (1) The first scenario include four settings: (i) Source model on MNIST training set, TEA adaptation on MNIST test set;", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a8f47726808e381e9bf96ac9165d32f1b8489a9690feaaf5ddb0a046a99f8aa.jpg", + "image_caption": [ + "Figure 4. Test distribution perception visualization for identical training and testing distributions on MNIST and CIFAR-10." + ], + "image_footnote": [], + "bbox": [ + 501, + 352, + 890, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c7e891cbc808f79805eac9d43182006756dc400b5b576c1338eee615640a4e11.jpg", + "image_caption": [ + "Figure 5. Test distribution perception visualization (upper) and real samples (lower) on shifted distribution: A model trained on PACS-A dataset then individually tested with TEA adaptation across PACS-P, PACS-A, PACS-C, PACS-S datasets." + ], + "image_footnote": [], + "bbox": [ + 501, + 467, + 890, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(ii) Source model on MNIST training set, TEA adaptation on 90-degree rotated MNIST test set; (iii) Source model on CIFAR-10 training set, TEA adaptation on CIFAR-10 test set; (iv) Source models on MNIST or CIFAR-10 training sets, tested without TEA adaptation. (2) In the second scenario, source model was trained on the PACS-A dataset and individually tested with TEA adaptation on PACS-P, PACS-A, PACS-C and PACS-S datasets.", + "bbox": [ + 498, + 700, + 890, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The outcomes from both scenarios are respectively illustrated in Fig. 4 and Fig. 5. Fig. 4 (1,3) indicates that, in scenarios where the distributions are closely identical, TEA has the potential to reconstruct samples that maintain discernible patterns. Fig. 4 (2) reveals that the sample distri", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23907", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "bution indeed originates from the test data rather than provoking recollections from the training datasets. Fig. 4 (4) confirms that models without adaptation, or those adapted using other methods, fail to effectively characterize the testing distributions on both the simple MNIST dataset (upper) and the more complex CIFAR-10 dataset (lower). Drawing conclusions from Fig. 5, we can infer that under significant distribution/domain shift, our method can still characterize the key features of the shifted test distribution, such as style, texture and color schemes.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, TEA endows the model with generative ability for test data via energy-based fine-tuning of the normalization layers. This ability may incorporates generative self-supervised information from the test data into the model and improves the model's thorough understanding of the test distribution, which in turn strengthens its generalization performance on that distribution.", + "bbox": [ + 75, + 243, + 472, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.3 TEA's Improvements in Confidence Calibration", + "text_level": 1, + "bbox": [ + 76, + 375, + 465, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This experiment compares model calibration across the source model, the entropy-based method TENT, the pseudolabeling-based method SHOT, and our energy-based TEA, on CIFAR-10. In accordance with the protocol in [4, 14], we illustrate reliability histogram and compute two scalar summary statistics: Expected Calibration Error (ECE) and Maximum Calibration Error (MCE) [14], to evaluate calibration implemented by torchmetrics. The procedures are implemented as follows: For the reliability histogram, we divided the model's predictions into ten bins based on the confidence score of the highest probability class and calculated the average accuracy for each bin.", + "bbox": [ + 75, + 398, + 468, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results are depicted in Fig. 6. From the perspective of the histogram, an optimally calibrated model should have its bar graph in a diagonal shape to achieve the smallest gap area. However, the bars for TENT and SHOT are observed to fall significantly below this line, manifesting an even inferior performance compared to the source model without any adaptation. These phenomena provide evidence that both the entropy-based methods and the pseudo-label-based method could potentially harm the confidence calibration by inducing overconfidence in their predictions. In stark contrast, our TEA has significantly narrowed the gap area and improved alignment with the diagonal line. For quantitative metrics, ECE and MCE, TEA has improved by $2.43\\%$ and $18.31\\%$ respectively compared to the source model.", + "bbox": [ + 75, + 582, + 468, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The improvement in calibration of TEA over competitors may come from that neural networks inherently tend to be overconfident [14]. The softmax function enforces exponential normalization among classes, which tends to amplify the probability of dominant classes, thus inherently not advantageous for calibration. TENT and SHOT exacerbate this dominance of certain classes by reducing the un", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/541afa8af5637d6d201a9f4c500afd276df4c7936f75e0c1bd7bc09ce2dca5e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 689, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8a8f66e1f1fa7729f3176fb5038b85b264417a8c639a08293184e284b22fbdd7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 88, + 890, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b5a682aa1125634be43cacd5bf5e40321b592bcb79a8543d317b6d510b9b28e2.jpg", + "image_caption": [ + "Figure 6. Calibration comparison between TEA and baselines on CIFAR-10 dataset. In an ideal scenario for optimal calibration, blue bars should align with the diagonal line, and a smaller grey gap area is preferred. Quantitative measures are provided via ECE and MCE metrics, where lower values indicate better calibration." + ], + "image_footnote": [], + "bbox": [ + 501, + 229, + 689, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/77b50894f51975d703b78ce5ac1ec58aa9f32d217e78a851331486070e049d07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 229, + 890, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "certainty of class probabilities, further amplifying the overconfidence of the classifier. On the contrary, TEA dose not perform normalization in the label space, but maximizes the log-sum-exp of classifier logits, which essentially introduces a certain level of uncertainty to each class and empowers TEA with the ability to enhance calibration.", + "bbox": [ + 496, + 445, + 892, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 549, + 619, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To achieve test time adaptation, we introduce an innovative energy-based perspective to mitigate the impact derived from distribution shifts. The proposed TEA aims to decrease the energy of the test data within the pretrained model's distribution. TEA guides the model towards achieving a harmonious low-energy equilibrium state for the test data, which mitigates the model's distribution discrepancy and boosts its generalizability towards test distributions. Comprehensive experiments across multiple tasks, benchmarks, and architectures confirm TEA's superiority over current leading methods. Further in-depth analyses of TEA's underlying mechanisms deepen our understanding of how energy reduction can enhance the model's perception of the test distribution, ultimately paving the way for improved generalization and calibration.", + "bbox": [ + 496, + 574, + 893, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 815, + 658, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the Strategic Priority Research Program of the CAS under Grants No. XDB0680302, and the National Key R&D Program of China (No. 2022YFB3103700, No. 2022YFB3103704).", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23908", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E. O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning, 2020. 2", + "[2] Stephen P Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. 5", + "[3] Miguel A Carreira-Perpinan and Geoffrey Hinton. On contrastive divergence learning. In International workshop on artificial intelligence and statistics, pages 33–40. PMLR, 2005. 4", + "[4] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 8", + "[5] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robust-bench: a standardized adversarial robustness benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 5", + "[6] Sabrina De Capitani Di Vimercati, Sara Foresti, Giovanni Livraga, and Pierangela Samarati. Data privacy: Definitions and techniques. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 20(06):793-817, 2012. 1", + "[7] Yilun Du and Igor Mordatch. Implicit generation and generalization in energy-based models. arXiv preprint arXiv:1903.08689, 2019. 3, 4", + "[8] Yilun Du, Shuang Li, Joshua Tenenbaum, and Igor Mordatch. Improved contrastive divergence training of energy-based models. In Proceedings of the 38th International Conference on Machine Learning, pages 2837-2848. PMLR, 2021. 2", + "[9] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. arXiv preprint arXiv:2103.10360, 2021. 1", + "[10] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, 2021. 5", + "[11] Jin Gao, Jialing Zhang, Xihui Liu, Trevor Darrell, Evan Shelhamer, and Dequan Wang. Back to the source: Diffusion-driven adaptation to test-time corruption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11786-11796, 2023. 2", + "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2", + "[13] Will Grathwohl, Kuan-Chieh Wang, Joern-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. In International Conference on Learning Representations, 2020. 3, 4, 2", + "[14] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "conference on machine learning, pages 1321-1330. PMLR, 2017. 2, 8", + "[15] Tian Han, Erik Nijkamp, Xiaolin Fang, Mitch Hill, Song-Chun Zhu, and Ying Nian Wu. Divergence triangle for joint training of generator model, energy-based model, and inferential model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8670-8679, 2019. 3", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[17] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 5, 1, 2", + "[18] Geoffrey E Hinton. Training products of experts by minimizing contrastive divergence. Neural computation, 14(8): 1771-1800, 2002. 2, 3, 4", + "[19] Liang Hou, Qi Cao, Yige Yuan, Songtao Zhao, Chongyang Ma, Siyuan Pan, Pengfei Wan, Zhongyuan Wang, Huawei Shen, and Xueqi Cheng. Augmentation-aware self-supervision for data-efficient gan training. arXiv preprint arXiv:2205.15677, 2022. 2", + "[20] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 5", + "[21] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 5", + "[22] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 2, 5", + "[23] Priyank Jain, Manasi Gyanchandani, and Nilay Khare. Big data privacy: a technological perspective and review. Journal of Big Data, 3:1-25, 2016. 1", + "[24] Minguk Jang, Sae-Young Chung, and Hye Won Chung. Test-time adaptation via self-training with nearest neighbor information. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[25] Jing Jiang. A literature survey on domain adaptation of statistical classifiers. URL: http://sifaka.cs.iuic.edu/jiang4/domainadaptation/survey, 3(1-12):3, 2008. 3", + "[26] Michael I Jordan and Tom M Mitchell. Machine learning: Trends, perspectives, and prospects. Science, 349(6245): 255-260, 2015. 1", + "[27] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5", + "[28] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23909", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, 2014. 2", + "[29] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5, 1", + "[30] Jogendra Nath Kundu, Akshay R Kulkarni, Suvaansh Bhambri, Deepesh Mehta, Shreyas Anand Kulkarni, Varun Jampani, and Venkatesh Babu Radhakrishnan. Balancing discriminability and transferability for source-free domain adaptation. In International Conference on Machine Learning, pages 11710-11728. PMLR, 2022. 2", + "[31] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 5, 1", + "[32] Yann LeCun, Sumit Chopra, Raia Hadsell, M Ranzato, and Fujie Huang. A tutorial on energy-based learning. Predicting structured data, 1(0), 2006. 2, 3", + "[33] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436-444, 2015. 1", + "[34] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 1, 2, 5, 6", + "[35] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pages 5542-5550, 2017. 5, 2", + "[36] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 1, 2, 5, 6", + "[37] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. arXiv preprint arXiv:2303.15361, 2023. 1, 2", + "[38] EM Lifshitz and Lev Davidovich Landau. Statistical physics, course of theoretical physics. In Part 2: Theory of the Condensed State. Butterworth-Heinemann Pergamon, London, 1980. 4", + "[39] Fu Lin, Rohit Mittapalli, Prithvijit Chattopadhyay, Daniel Bolya, and Judy Hoffman. Likelihood landscapes: A unifying principle behind many adversarial defenses. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 39-54. Springer, 2020. 2", + "[40] Weijian Luo, Hao Jiang, Tianyang Hu, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Training energy-based models with diffusion contrastive divergences. arXiv preprint arXiv:2307.01668, 2023. 2", + "[41] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 1, 2, 5, 6", + "[42] Weili Nie, Arash Vahdat, and Anima Anandkumar. Controllable and compositional generation with latent-space energy-based models. Advances in Neural Information Processing Systems, 34:13497-13510, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Frank Nielsen and Ke Sun. Guaranteed bounds on information-theoretic measures of univariate mixtures using piecewise log-sum-exp inequalities. Entropy, 18(12):442, 2016. 5", + "[44] Erik Nijkamp, Mitch Hill, Song-Chun Zhu, and Ying Nian Wu. Learning non-convergent non-persistent short-run mcmc toward energy-based model. Advances in Neural Information Processing Systems, 32, 2019. 4", + "[45] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yafo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 1, 2, 5, 6", + "[46] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yafo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 5, 6", + "[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5", + "[48] Francesco Pinto, Harry Yang, Ser Nam Lim, Philip Torr, and Puneet Dokania. Using mixup as a regularizer can surprisingly improve accuracy & out-of-distribution robustness. In Advances in Neural Information Processing Systems, pages 14608-14622. Curran Associates, Inc., 2022. 2", + "[49] Joaquin Quinonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset shift in machine learning. Mit Press, 2008. 1", + "[50] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530-1538. PMLR, 2015. 2", + "[51] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 5", + "[52] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 1, 2, 5, 6", + "[53] Hidetoshi Shimodaira. Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of statistical planning and inference, 90(2):227-244, 2000. 2", + "[54] Yang Song and Diederik P Kingma. How to train your energy-based models. arXiv preprint arXiv:2101.03288, 2021. 2, 3", + "[55] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23910", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the 37th International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 2", + "[56] Yushun Tang, Ce Zhang, Heng Xu, Shuoshuo Chen, Jie Cheng, Luziwei Leng, Qinghai Guo, and Zhihai He. Neuromodulated hebbian learning for fully test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3728-3738, 2023. 6", + "[57] Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring robustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems, 33: 18583-18599, 2020. 1", + "[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[59] Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 1999. 1", + "[60] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 1, 2, 5, 6", + "[61] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. Generalizing to unseen domains: A survey on domain generalization. IEEE Transactions on Knowledge and Data Engineering, 2022. 1", + "[62] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688, 2011. 2, 4", + "[63] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-mond, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019. 1", + "[64] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 5", + "[65] Zehao Xiao, Xiantong Zhen, Shengcai Liao, and Cees G. M. Snoek. Energy-based test sample adaptation for domain generalization. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[66] Xiulong Yang and Shihao Ji. Jem++: Improved techniques for training jem. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6494-6503, 2021. 2", + "[67] Yige Yuan, Bingbing Xu, Huawei Shen, Qi Cao, Keting Cen, Wen Zheng, and Xueqi Cheng. Towards generalizable graph contrastive learning: An information theory perspective. arXiv preprint arXiv:2211.10929, 2022. 1", + "[68] Yige Yuan, Bingbing Xu, Bo Lin, Liang Hou, Fei Sun, Huawei Shen, and Xueqi Cheng. Pde+: Enhancing generalization via pde with adaptive distributional diffusion. arXiv preprint arXiv:2305.15835, 2023. 5" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[69] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In British Machine Vision Conference 2016. British Machine Vision Association, 2016. 5", + "[70] Hao Zhao, Yuejiang Liu, Alexandre Alahi, and Tao Lin. On pitfalls of test-time adaptation. In ICLR 2023 Workshop on Pitfalls of limited data and computation for Trustworthy ML, 2023. 1" + ], + "bbox": [ + 501, + 90, + 893, + 189 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23911", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/TEA_ Test-time Energy Adaptation/c8c44200-6095-44b9-b188-9639c747d9ec_model.json b/2024/TEA_ Test-time Energy Adaptation/c8c44200-6095-44b9-b188-9639c747d9ec_model.json new file mode 100644 index 0000000000000000000000000000000000000000..64728ce80c2bd68e74db7a0cfc77062587d04137 --- /dev/null +++ b/2024/TEA_ Test-time Energy Adaptation/c8c44200-6095-44b9-b188-9639c747d9ec_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.306, + 0.131, + 0.666, + 0.154 + ], + "angle": 0, + "content": "TEA: Test-time Energy Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.18, + 0.898, + 0.234 + ], + "angle": 0, + "content": "Yige Yuan\\(^{1,2}\\), Bingbing Xu\\(^{1*}\\), Liang Hou\\(^{3}\\), Fei Sun\\(^{1}\\), Huawei Shen\\(^{1,2}\\), Xueqi Cheng\\(^{1,2}\\) \n\\(^{1}\\)CAS Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences \n\\(^{2}\\)University of Chinese Academy of Sciences Kuaishou Technology" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.236, + 0.852, + 0.253 + ], + "angle": 0, + "content": "{yuanyige20z, xubingbing, sunfei, shenhuawei, cxq}@ict.ac.cn, lianghou96@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.286, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.319, + 0.474, + 0.683 + ], + "angle": 0, + "content": "Test Time Adaptation (TTA) aims to improve model generalizability when test data diverges from training distribution, with the distinct advantage of not requiring access to training data and processes, especially valuable in the context of pre-trained models. However, current TTA methods fail to address the fundamental issue: covariate shift, i.e., the decreased generalizability can be attributed to the model's reliance on the marginal distribution of the training data, which may impair model calibration and introduce confirmation bias. To address this, we propose a novel energy-based perspective, enhancing the model's perception of target data distributions without requiring access to training data or processes. Building on this perspective, we introduce Test-time Energy Adaptation (TEA), which transforms the trained classifier into an energy-based model and aligns the model's distribution with the test data's, enhancing its ability to perceive test distributions and thus improving overall generalizability. Extensive experiments across multiple tasks, benchmarks and architectures demonstrate TEA's superior generalization performance against state-of-the-art methods. Further in-depth analyses reveal that TEA can equip the model with a comprehensive perception of test distribution, ultimately paving the way toward improved generalization and calibration1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.714, + 0.21, + 0.73 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.741, + 0.47, + 0.862 + ], + "angle": 0, + "content": "Deep neural networks, despite their remarkable performance under the assumption of independent and identically distributed (i.i.d.) training and test data [26, 33, 59], significantly degrade in real-world scenarios where unseen test data diverges from the training distribution. This limitation, known as distribution shift or domain shift, emphasizes the pressing need for generalizability across shifted test distributions [49, 57, 61, 67]. To tackle this issue, recent studies" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.288, + 0.871, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.419, + 0.871, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.553, + 0.894, + 0.624 + ], + "angle": 0, + "content": "Figure 1. Performance vs. energy on model trained with original distribution, tested across various shifted distributions. Upper: error rate change within energy score groups. Lower: loss variation with energy scores, each point denoting a distribution. Marker styles and opacity reflect distribution types and divergence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.893, + 0.794 + ], + "angle": 0, + "content": "of Test Time Adaptation (TTA) [37, 60, 70] began considering the incorporation of unlabeled test data and leveraging it to fine-tune the source model in an unsupervised manner. This paradigm has garnered significant attention due to its ability to operate without access to training data or involvement with the training process. In the era of large open-source models [9, 58, 63], where models are publicly available but the training data and training process remain inaccessible due to privacy and resource restrictions [6, 23], TTA emerges as especially beneficial and practical." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Existing TTA methods can be broadly categorized into three classes [37]. Normalization-based methods [41, 52] adjust the BatchNorm statistics of the model with test data statistics. Entropy-based methods [45, 46, 60] fine-tune the model by minimizing the prediction entropy. Pseudolabeling-based methods [34, 36] utilize test-time generated labels for updates. While these methods have been empi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.223, + 0.888 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.454, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Code is available at https://github.com/yuanyige/tea" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.454, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "23901" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "cally effective, these methods all fail to address a fundamental issue: covariate shift. Specifically, the decrease in generalization ability on test data with distribution shift can be attributed to the model's reliance on the marginal distribution of the training data. However, previous TTA methods do not address this shift due to their lack of connection with marginal distributions, impairing model calibration [14, 24] and introducing confirmation bias [1]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.216, + 0.47, + 0.306 + ], + "angle": 0, + "content": "To combat the above challenges, we propose a novel way rooted in an energy-based perspective. Within this way, energy is defined as an unnormalized probability assigned to a sample, where a lower score corresponds to a higher likelihood of that sample within a distribution [32, 54]. Proposing such a way to improve test time adaptation is twofold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.471, + 0.656 + ], + "angle": 0, + "content": "Firstly, test samples that correspond to lower energy within the model's distribution tend to exhibit higher performance. This is demonstrated by examining the energy scores of various test datasets in relation to a model trained on a specific training distribution. As depicted in Fig. 1, an increase in the divergence between the test and training distributions is accompanied by a drastic escalation in energy scores, leading to a significant degradation in performance. Secondly, the energy-based way can address covariate shift under TTA via directly injecting the model with a comprehensive perception of test distribution. Addressing covariate shift in TTA is particularly challenging, as it is neither feasible to access the training dataset to align the marginal distribution between training and testing [53], nor possible to modify the training process to mitigate the influence of marginal training distribution [48]. Under such circumstances, the energy-based way can directly manipulate the trained model's likelihood landscape [39] via an implicit distribution modeling process without requiring training process and training data, becoming a promising way. This stands in contrast to other models such as GANs [12, 19], Flows [50], and VAEs [28] which are advantageous only when the training data are accessible." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Building on the above energy-based way, we propose Test-time Energy Adaptation, abbreviated as TEA, which constructs an energy-based model from the trained (classifier) model by reinterpreting the negative log-sum-exp of logits as an energy function, and employs Contrastive Divergence [18] as the adaptation objective to decrease the energy of test samples while increase that of samples generated by Stochastic Gradient Langevin Dynamics [62]. This approach prevents a trivial solution that indiscriminately reduces the energy across the entire data space to ensure an increased likelihood for target test samples within the model's distribution. TEA enables a gradual alignment between the distributions of the trained model and the test data, bolstering the trained model's perception of the test distribution and paving the way for superior adaptability and performance when confronted with the corresponding test data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "We investigate the effectiveness of TEA under image corruption and domain generalization on four popular benchmarks CIFAR-10, CIFAR-100, TinyImageNet and PACS, across three architectures WRN-28-10, ResNet50 and ResNet-18. Experimental results underscore that TEA significantly outperforms current best-performing TTA methods in terms of generalizability, with an average increment of \\(4.7\\%\\). We further reveal that TEA can equip the model with a comprehensive perception of the test distribution. This, in turn, significantly improves the generalization and calibration of the trained model." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.261, + 0.733, + 0.274 + ], + "angle": 0, + "content": "Our main contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.278, + 0.891, + 0.338 + ], + "angle": 0, + "content": "- Promising Way: We propose a new energy-based way for test time adaptation, which marks a departure from traditional methods and sheds light on potential avenues for mitigating the impact of distribution shifts." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.338, + 0.892, + 0.399 + ], + "angle": 0, + "content": "- Innovative Method: We propose TEA to decrease the energy of the test data within the model's distribution, thereby equipping the trained model with a perception of the test distribution and enhancing generalizability." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.399, + 0.892, + 0.489 + ], + "angle": 0, + "content": "- Extensive Experiments: Experiments across extensive settings validate TEA's superiority over current leading methods. Further in-depth analyses extend the understanding of how energy reduction enhances the model's perception of test distribution, ultimately paving the way toward improved generalization and calibration." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.278, + 0.892, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.511, + 0.642, + 0.527 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Test Time Adaptation Test Time Adaptation (TTA) [37] is a paradigm aiming to enhance a model's generalizability on specific test data through unsupervised fine-tuning with these data. Note that the model is originally trained on a distinct training dataset, which is not available during the adaptation phase. Approaches like TTT [55] adapt models through self-supervised proxy task during testing but require the training of the same proxy task during training procedure. DDA [11, 65] explores adapting the test data, yet faces limitations due to model structure and training constraints. Recent research [60] highlights a scenario where the training process and training data is entirely agnostic, leading to three main categories of approaches: For normalization-based, BN [52] adapts the BatchNorm [22] statistics with test data. DUA [41] uses a tiny fraction of test data and its augmentation for BatchNorm statistics adaptation. For entropy-based, TENT [60] fine-tunes BatchNorm layers using entropy minimization during the test phase. EATA [45] employs a Fisher regularizer to limit excessive model parameter changes. SAR [46] removes high-gradient samples and promotes flat minimum weights. For pseudolabeling-based, PL [34] fine-tunes parameters using confident pseudo labels. SHOT [36] combines entropy minimization methods with pseudo labeling." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23902" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.088, + 0.878, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.332, + 0.892, + 0.416 + ], + "angle": 0, + "content": "Figure 2. Overview of Test-time Energy Adaptation (TEA). Given a trained model (classifier) and in-coming test data, TEA directly integrates test data distribution into the trained classifier by fine-tuning its normalization layers through energy-based training: TEA constructs an Energy-Based Model from the classifier by reinterpreting the negative log-sum-exp of logits as an energy function, and employs Contrastive Divergence as the adaptation objective which decreases the energy of test samples while increases the energy of negative samples generated by Langevin Dynamics. This adaptation increases the likelihood of test samples under the classifier's distribution, enabling a gradual alignment between the distributions of the trained classifier and the test data, thereby enhancing generalizability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.473, + 0.657 + ], + "angle": 0, + "content": "Energy Based Model Energy-Based Models (EBMs) are a type of non-normalized probabilistic models. Unlike most other probabilistic models, EBMs do not necessitate the normalizing constant to be tractable [32, 54] and do not require an explicit neural network for sample generation, implying the generation process is implicit [7]. These lead to increased flexibility in parameterization and allow for modeling a wider range of probability distributions. Due to their flexibility, EBMs can construct hybrid models with both discriminative and generative capabilities, integrating the generative competencies into discriminative models without sacrificing their discriminative capabilities [7, 13, 15]. Among these, JEM [13] is particularly representative, reinterpreting classifiers as an EBM and achieving impressive results in both classification and generation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.672, + 0.169, + 0.687 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.47, + 0.804 + ], + "angle": 0, + "content": "In this section, we detail our method, Test-time Energy Adaptation (TEA). Initially, we present a thorough description of the notation and overall architecture in Sec. 3.1, after which we proceed to explain energy adaption (Sec. 3.2) and modulation parameters (Sec. 3.3), respectively. Furthermore, we engage in a discussion about the difference between our method and entropy-based adaptation in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.379, + 0.832 + ], + "angle": 0, + "content": "3.1. Notation and Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The labeled training dataset is denoted as \\(\\{(\\mathbf{x}_{\\mathrm{train}},y_{\\mathrm{train}})\\} \\subset \\mathcal{X}\\times \\mathcal{Y}\\) and the unlabeled test data are represented by \\(\\mathbf{x}_{\\mathrm{test}}\\in\\) \\(\\mathcal{X}\\), where \\(\\mathcal{X}\\) and \\(\\mathcal{Y}\\) are data and label spaces. The respective marginal distributions of the training and test data are given" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.503 + ], + "angle": 0, + "content": "by \\(\\mathbf{x}_{\\mathrm{train}}\\sim p_{\\mathrm{train}}(\\mathbf{x})\\) and \\(\\mathbf{x}_{\\mathrm{test}}\\sim p_{\\mathrm{test}}(\\mathbf{x})\\). A classifier model trained on the training dataset and parameterized by \\(\\theta\\), is denoted as \\(f_{\\theta}:\\mathcal{X}\\to \\mathcal{Y}\\). The data distribution learned by this trained classifier is denoted by \\(p_{\\theta}(\\mathbf{x})\\), which will be referred to as the model distribution henceforth." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.893, + 0.793 + ], + "angle": 0, + "content": "The overall framework of TEA is depicted in Fig. 2. The motivation behind TEA is rooted in the issue of covariate shifts [25], where the degradation of model generalization on test data \\(\\mathbf{x}_{\\mathrm{test}}\\) is attributed to the model's reliance on the training distribution \\(p_{\\mathrm{train}}(\\mathbf{x})\\). To overcome this issue without accessing the training data or training process, TEA directly integrates the test data distribution into the trained model. TEA constructs an energy-based model from the trained classifier by reinterpreting the negative log-sum-exp of logits as an energy function. Through this lens, TEA employs contrastive divergence [18] as the adaptation objective, which serves to decrease the energy (increase the likelihood) of the test samples under the model distribution \\(p_{\\theta}(\\mathbf{x})\\). This adaptation enables the gradual alignment of distribution between the trained model and test data, thereby bolstering the model's perception of the test distribution and enhancing generalizability. Following previous TTA methods, TEA freezes the majority of the model's parameters, permitting only minor adjustments for efficient adaptation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.845, + 0.817 + ], + "angle": 0, + "content": "3.2. Energy Adaptation for Test Distribution" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Enhancing the model's perception of test distribution from an energy-based perspective involves two key steps: constructing the energy-based model and optimizing it." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Constructing the energy-based model. Energy-based models (EBMs) [13, 32, 54] represent a class of proba" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23903" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.228 + ], + "angle": 0, + "content": "bilistic models that are characterized by an energy function. Consider a sample \\(\\mathbf{x} \\in \\mathbb{R}^D\\), the energy function \\(E: \\mathbb{R}^D \\to \\mathbb{R}\\) maps each sample into an energy that can be considered as an unnormalized probability, with lower scores indicating higher likelihoods [7]. Consequently, the probability density \\(p(\\mathbf{x})\\), as defined by EBM, can be expressed using the Boltzmann distribution [38], as shown in Eq. (1), where the partition function \\(Z = \\int \\exp(-E(\\mathbf{x})) \\, \\mathrm{d}\\mathbf{x}\\) serves to normalize the probability density." + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.232, + 0.469, + 0.248 + ], + "angle": 0, + "content": "\\[\np (\\mathbf {x}) = \\exp (- E (\\mathbf {x})) / Z. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.252, + 0.47, + 0.372 + ], + "angle": 0, + "content": "Constructing an energy-based model from a trained classifier \\( f_{\\theta} \\) is founded on the fundamental analysis that an energy-based framework inherently underlies any discriminative model [13]. In this framework, the energy of one sample for a corresponding class can be represented as its logit produced by the trained classifier, denoted by \\( E_{\\theta}(\\mathbf{x},y) = -f_{\\theta}(\\mathbf{x})[y] \\). Therefore, the joint probability distribution of \\( \\mathbf{x} \\) and \\( y \\) can be defined as," + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.377, + 0.469, + 0.394 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (\\mathbf {x}, y) = \\exp \\left(f _ {\\theta} (\\mathbf {x}) [ y ]\\right) / Z _ {\\theta}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.469, + 0.426 + ], + "angle": 0, + "content": "then the distribution of \\( \\mathbf{x} \\) can be obtained by marginalizing over \\( y \\), as shown below," + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.43, + 0.469, + 0.462 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (\\mathbf {x}) = \\sum_ {y} p _ {\\theta} (\\mathbf {x}, y) = \\sum_ {y} \\exp \\left(f _ {\\theta} (\\mathbf {x}) [ y ]\\right) / Z _ {\\theta}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.469, + 0.496 + ], + "angle": 0, + "content": "By substituting Eq. (3) into Eq. (1), we can obtain the form of the energy function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.499, + 0.469, + 0.531 + ], + "angle": 0, + "content": "\\[\nE _ {\\theta} (\\mathbf {x}) = - \\log \\sum_ {y} \\exp (f _ {\\theta} (\\mathbf {x}) [ y ]). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.469, + 0.595 + ], + "angle": 0, + "content": "Following the aforementioned steps, we repurpose and reinterpret the logits produced by the trained classifier to establish an energy-based model and define the energy function as the negative log-sum-exp of the logits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.469, + 0.643 + ], + "angle": 0, + "content": "After the logits reinterpretation, we can construct a energy-based model for the test data \\(\\mathbf{x}_{\\mathrm{test}}\\) using the trained classifier, where \\(Z_{\\theta} = \\int \\sum_{y}\\exp \\left(f_{\\theta}(\\mathbf{x})[y]\\right)\\mathrm{d}\\mathbf{x}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.646, + 0.469, + 0.679 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {x} _ {\\text {t e s t}}\\right) = \\frac {\\exp \\left(- E _ {\\theta} \\left(\\mathbf {x} _ {\\text {t e s t}}\\right)\\right)}{Z _ {\\theta}} = \\frac {\\sum_ {y} \\exp \\left(f _ {\\theta} \\left(\\mathbf {x} _ {\\text {t e s t}}\\right) [ y ]\\right)}{Z _ {\\theta}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.469, + 0.817 + ], + "angle": 0, + "content": "Optimizing the energy-based model. Optimizing Eq. (5) is challenging as the partition function \\( Z_{\\theta} \\) necessitates integration across the whole input space of \\( \\mathbf{x} \\), typically making it computationally intractable. Thus, direct maximizing the log-likelihood of test data \\( \\log p_{\\theta}(\\mathbf{x}_{\\mathrm{test}}) \\) presents significant difficulties when training the parameter \\( \\theta \\) of our energy-based model. To overcome this difficulty, we propose to use contrastive divergence [3, 18] via estimating the gradient of the log-likelihood," + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.82, + 0.469, + 0.854 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\log p _ {\\theta} (\\mathbf {x} _ {\\text {t e s t}})}{\\partial \\theta} = \\mathbb {E} _ {\\tilde {\\mathbf {x}} \\sim p _ {\\theta}} \\left[ \\frac {\\partial E _ {\\theta} (\\tilde {\\mathbf {x}})}{\\partial \\theta} \\right] - \\frac {\\partial E _ {\\theta} (\\mathbf {x} _ {\\text {t e s t}})}{\\partial \\theta}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In Eq. (6), the notation \\(\\tilde{\\mathbf{x}}\\sim p_{\\theta}\\) denotes a random sample drawn from the distribution over \\(\\mathbf{x}\\), which is defined by the model's distribution. The sampling procedure can be" + }, + { + "type": "code_caption", + "bbox": [ + 0.51, + 0.095, + 0.798, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: Test-time Energy Adaptation" + }, + { + "type": "algorithm", + "bbox": [ + 0.501, + 0.112, + 0.782, + 0.321 + ], + "angle": 0, + "content": "Input: \\(f_{\\theta},N,T,p_0,\\alpha ,\\beta ,\\mathbf{x}_{\\mathrm{test}}\\) Output: \\(f_{\\theta}(\\mathbf{x}_{\\mathrm{test}})\\) \n1 \\(E_{\\theta}(\\cdot)\\gets -\\log \\sum_{y}\\exp (f_{\\theta}(\\cdot)[y])\\) \n2 for \\(i\\gets 0,1,\\dots ,N - 1\\) do \n3 \\(\\tilde{\\mathbf{x}}_0\\gets \\mathrm{sample}(p_0)\\) \n4 for \\(t\\gets 0,1,\\ldots ,T - 1\\) do \n5 \\(\\begin{array}{r}\\epsilon \\leftarrow \\mathrm{sample}(\\mathcal{N}(\\mathbf{0},\\mathbf{I})) \\end{array}\\) \n6 \\(\\tilde{\\mathbf{x}}_{t + 1}\\gets \\tilde{\\mathbf{x}}_t - \\frac{\\alpha}{2}\\frac{\\partial E_\\theta(\\tilde{\\mathbf{x}}_t)}{\\partial\\tilde{\\mathbf{x}}_t} +\\alpha \\epsilon\\) \n7 end \n8 \\(\\tilde{\\mathbf{x}}\\gets \\tilde{\\mathbf{x}}_{T - 1}\\) \n9 \\(\\theta \\gets \\theta -\\beta \\nabla_{\\theta}[E_{\\theta}(\\mathbf{x}_{\\mathrm{test}}) - E_{\\theta}(\\tilde{\\mathbf{x}})]\\) \n10 end \n11 return \\(f_{\\theta}(\\underline{\\mathbf{x}}_{\\mathrm{test}})\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.341, + 0.892, + 0.446 + ], + "angle": 0, + "content": "performed through Stochastic Gradient Langevin Dynamics (SGLD) [62], which iteratively generates samples by using the gradient information [13, 44]. In this context, \\( p_0 \\) represents an initial noise distribution, \\( \\alpha \\) denotes the step-size, and \\( t = 0, 1, \\dots, T - 1 \\) is the time step. After \\( T \\) steps of updating, a fictitious sample, generated by the energy-based model governed by the classifier, can be obtained." + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.455, + 0.891, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {x}} _ {t + 1} = \\mathbf {x} _ {t} - \\frac {\\alpha}{2} \\frac {\\partial E _ {\\theta} \\left(\\tilde {\\mathbf {x}} _ {t}\\right)}{\\partial \\tilde {\\mathbf {x}} _ {t}} + \\sqrt {\\alpha} \\boldsymbol {\\epsilon}, \\quad \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.659 + ], + "angle": 0, + "content": "this sampling process in Eq. (7) essentially optimizes the sample by moving in the direction of energy reduction. Consequently, the objective of Eqs. (6) and (7) can be fundamentally understood as a min-max game (Eq. (8)), which minimizes the energy derived from the incoming test samples while concurrently amplifying the energy of fictitious samples obtained via SGLD from the classifier's distribution. Significantly, the latter plays a pivotal role in preventing a collapse towards a trivial solution where energy is indiscriminately minimized throughout the entire space, rather than focusing on the target test data." + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.668, + 0.891, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\theta} \\left[ \\min _ {\\tilde {\\mathbf {x}}} E _ {\\theta} (\\tilde {\\mathbf {x}}) - E _ {\\theta} \\left(\\mathbf {x} _ {\\text {t e s t}}\\right) \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "By adapting through this objective, the classifier's distribution continually converges towards the test distribution. The convergence point of the min-max game is reached when the energy of samples drawn from the classifier's distribution matches the energy of the test samples. At this stage, the classifier reaches a low-energy steady state with respect to the test data. The likelihood landscape of the classifier then exhibits a higher probability for samples that are consistent with the test data distribution, and conversely, a lower probability for samples that deviate from it, leading to a comprehensive perception of the classifier towards this test distribution and ultimately enhancing its generalizability. The pseudocode for TEA can be found in Algorithm 1." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23904" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.298, + 0.106 + ], + "angle": 0, + "content": "3.3. Modulation Parameters" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.473, + 0.449 + ], + "angle": 0, + "content": "As outlined in Eq. (8), TEA requires updating the parameters of the trained model using the aforementioned energy adaptation to adjust to test data. In line with previous methods, we opt to update the parameters of the normalization layer due to the following two factors: (1) Practicality and Efficiency: In the era of large-scale models, the practice of fine-tuning a selected group of parameters has gained prominence [20]. For both practicality and efficiency, it's crucial to avoid updating all parameters as this process would be excessively time-consuming. Note that the parameters of the normalization layer account for a modest \\(1\\%\\) of total model parameters [60], making their update become more manageable. (2) Direct impact on data distribution: The parameters within normalization layers possess the potential to capture intrinsic features of data, thereby exerting direct influence upon the corresponding data distribution. As evidenced by [21], simple modulation to the mean and variance in a generator's normalization layers can modify image style, underlining the normalization parameters' profound impact on data distribution. This aligns well with our goal of manipulating the energy of test data to make it compatible with the model distribution." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.459, + 0.196, + 0.473 + ], + "angle": 0, + "content": "3.4. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.482, + 0.473, + 0.741 + ], + "angle": 0, + "content": "As entropy-based adaptation has been the representative adaptation method, we further discuss the difference between TEA and entropy-based adaptation. Intriguingly, TEA may have a connection with entropy-based adaptation, given that the negative entropy \\(\\mathrm{NegEntropy}(\\mathbf{x}) = \\sum_{i}x_{i}\\log x_{i}\\) is the convex conjugate of the free energy \\(\\mathrm{LogSumExp}(\\mathbf{x}) = \\log \\sum_{i}\\exp (x_{i})\\), as established in the literature [2, 43]. Despite this connection, entropy-based methods, which apply softmax normalization in the label space \\(y\\) and strive to minimize entropy, can result in diminished uncertainty within classification probabilities, leading to compromised model calibration. In contrast, TEA, utilizing the log-sum-exp function within the data space \\(\\mathbf{x}\\) can not only effectively avoid the pitfalls associated with entropy-based methods, but also improve calibration by introduces uncertainty to each class. This conjecture has been substantiated through experiments (refer to Sec. 4.3.3)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.203, + 0.772 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.781, + 0.472, + 0.903 + ], + "angle": 0, + "content": "In the following sections, we compare TEA with state-of-the-art methods across various tasks, benchmarks and architectures. Then, we delve into deeper understanding of our method by exploring its desirable properties and identifying significant components that contribute to its improvements. Due to the space limitations, more comprehensive experiments including full results on corruptions and other analyses, are provided in Appendix Sec. 7." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.691, + 0.108 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.895, + 0.298 + ], + "angle": 0, + "content": "Datasets and Metrics We focus on two tasks to verify the performance of TEA: generalization on image corruption and domain generalization. Image corruption includes clean and corrupted datasets from CIFAR-10(C), CIFAR-100(C) and TinyImageNet-200(C) [17, 29, 31], incorporating 15 types of corruption at 5 severity levels. Domain generalization considers PACS [35], encompassing 4 domains (Photo, Art Painting, Cartoon, Sketch) across 7 categories. We follow [68] to use Accuracy and mean Corruption Error (mCE) [17] as evaluation metrics. The evaluations are conducted at both the most severe level and the average of all severity levels to ensure thorough analysis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.315, + 0.895, + 0.483 + ], + "angle": 0, + "content": "Backbones and Baselines We use two architectures for image corruption: WideResNet-28-10 [69] with BatchNorm [22], and ResNet-50 [16] with GroupNorm [64], consistent with the implementations of TENT [60] and SAR [46]. We use ResNet-18 [16] for domain generalization. We evaluate our method against eight leading TTA methods across three categories: (1) Normalization-based methods: BN [52] and DUA [41]. (2) Entropy-based methods: TENT [60], ETA, EATA [45], and SAR [46]; (3) Pseudo-labeling-based methods: PL [34] and SHOT [36]. Source denotes the original model without any adaptation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.895, + 0.669 + ], + "angle": 0, + "content": "Implementation We implement methods based on PyTorch [47]. Consistency in model weight is ensured by the RobustBench protocol [5], which provides pretrained weights for the WideResNet-28-10 (BatchNorm) on CIFAR-10. In the case where RobustBench weights are unavailable, we train models in accordance with the guidelines specified in [69]. All adaptation employ Adam [27], except for SAR, which originally uses SAM [10] with SGD [51]. Baselines are replicated using their original hyper-parameters, except when these were unspecified. More details and setups are deferred to Appendix Sec. 8." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.677, + 0.685, + 0.693 + ], + "angle": 0, + "content": "4.2. Adaptation Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.732 + ], + "angle": 0, + "content": "This section compares the generalizability of TEA with baselines in image corruption and domain generalization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Image Corruption As reported in Tab. 1, we conducted experiments on three benchmarks against eight baselines for corruption scenarios. TEA markedly surpasses all baselines on the vast majority of datasets and severity levels. Specifically, TEA outshines the best-performing baseline by an average of \\(4.7\\%\\) at the most severe level. The only exception is on TinyImageNet across all levels, where TEA ranks second, merely trailing by a minimal margin \\(0.1\\%\\). We further incorporate the ResNet50 with GroupNorm for a broader validation. Tab. 2 indicates that TEA still delivers the best" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23905" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.096, + 0.892, + 0.138 + ], + "angle": 0, + "content": "Table 1. Comparisons of TEA and baselines for image corruption on CIFAR-10(C), CIFAR-100(C), and TinyImageNet(C) using WRN28-10 with BatchNorm. Accuracy and mCE are evaluated at the most severe level and across all levels with asterisk (*) indicating the results are taken from the original paper [56]. The best adaptation results are highlighted in boldface." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.143, + 0.892, + 0.32 + ], + "angle": 0, + "content": "
WRN-28-10 BatchNormCIFAR-10(C)CIFAR-100(C)Tiny-ImageNet(C)
Clean Acc (↑)Corr Severity 5 Acc (↑)mCE (↓)Corr Severity 1-5 Acc (↑)mCE (↓)Clean Acc (↑)Corr Severity 5 Acc (↑)mCE (↓)Corr Severity 1-5 Acc (↑)mCE (↓)Clean Acc (↑)Corr Severity 5 Acc (↑)mCE (↓)Corr Severity 1-5 Acc (↑)mCE (↓)
Source94.7756.47100.0073.45100.0081.7935.39100.0052.12100.0063.1921.21100.0034.13100.00
NormBN [52]93.9779.5652.6585.6360.0080.8360.0663.5468.1169.4245.0427.7493.4234.27100.96
DUA* [41]-80.1050.78------------
PseudoPL [34]93.7551.42106.9872.6299.3780.5253.4072.1264.5375.2947.8428.2691.2239.8391.67
SHOT [36]93.2574.7763.1982.3572.6180.5256.5368.0166.0073.2847.9529.1490.1640.0191.41
EntropyTENT [60]93.6681.4148.1386.7556.1780.1463.0959.4269.4767.8039.5426.3195.5232.03104.49
ETA [45]93.9679.5852.6485.6359.9980.6559.8264.5267.1772.4043.2027.2894.1233.46102.25
EATA [45]93.9679.5952.6285.6459.9880.6860.2463.7567.4871.6643.4227.2894.0933.47102.24
SAR [46]93.9779.7751.9485.8358.9780.8462.9559.3770.0165.9941.5828.2192.8234.60100.47
EnergyTEA94.0983.3443.6987.8852.0080.8865.1056.0771.2263.7251.6531.6787.9939.9692.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.33, + 0.47, + 0.372 + ], + "angle": 0, + "content": "Table 2. Comparisons for image corruption on CIFAR-10(C), CIFAR-100(C), and TinyImageNet(C) using ResNet-50 with GroupNorm across all severity levels. Best results are in boldface." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.377, + 0.469, + 0.519 + ], + "angle": 0, + "content": "
ResNet50 GroupNormCIFAR-10(C)CIFAR-100(C)Tiny-ImageNet(C)
Acc (↑)mCE (↓)Acc (↑)mCE (↓)Acc (↑)mCE (↓)
Source78.71100.0054.98100.0026.64100.00
PseudoPL79.4394.7656.6896.0226.6099.92
SHOT81.9886.6558.3193.4529.1196.73
EntropyTENT77.29102.8856.3496.8826.6599.94
ETA78.68100.0956.7296.3729.2596.42
EATA78.70100.0256.7696.2829.2596.42
SAR78.7899.6555.2899.3327.0599.41
EnergyTEA83.0579.0959.6789.3230.4194.81
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.529, + 0.47, + 0.59 + ], + "angle": 0, + "content": "performance in both average accuracy and mCE. These results underscore the efficacy of TEA in handling image corruption scenarios, ensuring its universality for model architecture or normalization techniques." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Domain Generalization Tab. 3 provides a comparison between TEA and state-of-the-art TTA methods on the PACS dataset. It is evident that when trained on the photo and art domains, TEA exhibits substantial improvements over the best-performing baseline, achieving increases of \\(7.1\\%\\) and \\(8.5\\%\\), respectively. Compared to photo and art, cartoon and sketch domains may exhibit greater domain discrepancies, posing significant challenges to model generalization. Despite these challenging conditions, in the adaptation from cartoon to art and the sketch domain, TEA achieved improvements when all baselines experienced significant declines. TENT shows decline of \\(14.73\\%\\) compared to the source in sketch, whereas TEA bucks the trend and enhances performance by \\(4.5\\%\\), highlighting its stability in the face of severe domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The results from both tasks highlight TEA's superior generalizability. These effectiveness may originate from the reduced energy, the enhanced distribution perception, and the improved calibration. In next section, we will delve into" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.33, + 0.892, + 0.373 + ], + "angle": 0, + "content": "Table 3. Single source domain generalization comparisons on PACS datasets using ResNet-18 with BatchNorm in terms of Accuracy. The best adaptation results are highlighted in boldface." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.377, + 0.892, + 0.73 + ], + "angle": 0, + "content": "
Source DomainMethodTarget DomainAvg
PhotoArtCartoonSketch
PhotoSource-26.7622.4016.6221.93
BN-26.6627.9415.9623.52
TENT-26.9529.8617.5424.78
EATA-26.6628.1115.9823.59
SAR-26.7128.4115.9823.70
SHOT-26.6129.8620.9225.80
TEA-28.8133.6220.4927.64
ArtSource49.04-36.4324.4836.65
BN46.65-28.2822.7332.55
TENT50.78-30.1224.6135.17
EATA46.83-29.3123.4233.19
SAR47.90-33.0226.2735.73
SHOT50.24-34.3029.3737.97
TEA56.29-38.5728.7141.19
CartoonSource42.6929.79-29.4733.98
BN28.6825.15-20.8724.90
TENT30.9623.34-22.6525.65
EATA28.8025.10-25.0426.31
SAR29.7025.78-21.5125.66
SHOT37.7222.66-23.1427.84
TEA36.0531.44-22.8830.12
SketchSource19.9418.7032.21-23.62
BN13.4717.1429.86-20.16
TENT13.5317.3829.52-20.14
EATA13.1717.3330.08-20.19
SAR13.2918.8029.95-20.68
SHOT19.7618.7530.46-22.99
TEA19.6421.2433.19-24.69
" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.748, + 0.822, + 0.763 + ], + "angle": 0, + "content": "these aspects for further analysis and discussion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.783, + 0.72, + 0.8 + ], + "angle": 0, + "content": "4.3. Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we delve into the mechanisms driving TEA's effectiveness and explore its desirable properties. Specifically, we studied three key aspects of TEA: (1) The correlation between energy reduction and generalizability enhancement; (2) The distribution perception and generation capabilities; (3) The confidence calibration improvements." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23906" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.353, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.189, + 0.346, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.089, + 0.622, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.189, + 0.61, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.089, + 0.892, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.189, + 0.882, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.285, + 0.894, + 0.342 + ], + "angle": 0, + "content": "Figure 3. This illustration captures the energy reduction and generalizability enhancement achieved by TEA across CIFAR-10-C, CIFAR-100-C, and TinyImageNet-200-C, displayed from left to right. The upper set of graphs trace the evolution of energy score, corresponding loss and accuracy in response to incrementally increasing TEA adaptation steps. The lower set uncovers the extent of energy reduction and the consequent performance improvement before and after executing TEA adaptation, under different levels of distribution shift." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.355, + 0.471, + 0.387 + ], + "angle": 0, + "content": "4.3.1 Relation between TEA's Energy Reduction and Generalizability Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.47, + 0.517 + ], + "angle": 0, + "content": "This experiment validate TEA's energy reduction capability, and its impact on generalizability, spanning two scenarios: (1) In the first scenario, we focus on the trends of energy scores, loss, and accuracy on the same test data over increasing adaptation steps. (2) In the second scenario, we explore how the extent of energy reduction correlates with performance improvements, before and after adaptation, using varied test data that exhibit different distribution shifts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.471, + 0.761 + ], + "angle": 0, + "content": "The results from both scenarios are depicted in Fig. 3. It is observable that \\((i)\\) As the iteration step of TEA increases, there is a consistent reduction in energy and corresponding decrease in loss, coupled with an ongoing enhancement in accuracy. \\((ii)\\) As the distribution shift increases, TEA's energy reduction becomes more pronounced. Concurrently, the enhancement in performance over the baseline also increases. \\((iii)\\) As the distribution shift increases, there is a sharp degradation observed in the baseline performance. However, the model adapted via TEA maintains its stability and robustness, demonstrating resilience against strong distribution shifts. In summary, these aforementioned trends are consistently observed across three datasets, demonstrating TEA's significant effectiveness in reducing energy. Notably, a greater reduction in energy correlates with an increased improvement in the model's generalizability" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.456, + 0.8 + ], + "angle": 0, + "content": "4.3.2 TEA's Distribution Perception and Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "This experiment aims to validate the capability of TEA in modeling test data distribution. We framed the experiment within two scenarios: identical training and testing distributions, and shifted training and testing distributions. (1) The first scenario include four settings: (i) Source model on MNIST training set, TEA adaptation on MNIST test set;" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.353, + 0.892, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.433, + 0.892, + 0.461 + ], + "angle": 0, + "content": "Figure 4. Test distribution perception visualization for identical training and testing distributions on MNIST and CIFAR-10." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.468, + 0.892, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.623, + 0.894, + 0.679 + ], + "angle": 0, + "content": "Figure 5. Test distribution perception visualization (upper) and real samples (lower) on shifted distribution: A model trained on PACS-A dataset then individually tested with TEA adaptation across PACS-P, PACS-A, PACS-C, PACS-S datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.701, + 0.892, + 0.822 + ], + "angle": 0, + "content": "(ii) Source model on MNIST training set, TEA adaptation on 90-degree rotated MNIST test set; (iii) Source model on CIFAR-10 training set, TEA adaptation on CIFAR-10 test set; (iv) Source models on MNIST or CIFAR-10 training sets, tested without TEA adaptation. (2) In the second scenario, source model was trained on the PACS-A dataset and individually tested with TEA adaptation on PACS-P, PACS-A, PACS-C and PACS-S datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The outcomes from both scenarios are respectively illustrated in Fig. 4 and Fig. 5. Fig. 4 (1,3) indicates that, in scenarios where the distributions are closely identical, TEA has the potential to reconstruct samples that maintain discernible patterns. Fig. 4 (2) reveals that the sample distri" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23907" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "bution indeed originates from the test data rather than provoking recollections from the training datasets. Fig. 4 (4) confirms that models without adaptation, or those adapted using other methods, fail to effectively characterize the testing distributions on both the simple MNIST dataset (upper) and the more complex CIFAR-10 dataset (lower). Drawing conclusions from Fig. 5, we can infer that under significant distribution/domain shift, our method can still characterize the key features of the shifted test distribution, such as style, texture and color schemes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.473, + 0.351 + ], + "angle": 0, + "content": "In summary, TEA endows the model with generative ability for test data via energy-based fine-tuning of the normalization layers. This ability may incorporates generative self-supervised information from the test data into the model and improves the model's thorough understanding of the test distribution, which in turn strengthens its generalization performance on that distribution." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.375, + 0.466, + 0.39 + ], + "angle": 0, + "content": "4.3.3 TEA's Improvements in Confidence Calibration" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.47, + 0.581 + ], + "angle": 0, + "content": "This experiment compares model calibration across the source model, the entropy-based method TENT, the pseudolabeling-based method SHOT, and our energy-based TEA, on CIFAR-10. In accordance with the protocol in [4, 14], we illustrate reliability histogram and compute two scalar summary statistics: Expected Calibration Error (ECE) and Maximum Calibration Error (MCE) [14], to evaluate calibration implemented by torchmetrics. The procedures are implemented as follows: For the reliability histogram, we divided the model's predictions into ten bins based on the confidence score of the highest probability class and calculated the average accuracy for each bin." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.794 + ], + "angle": 0, + "content": "The results are depicted in Fig. 6. From the perspective of the histogram, an optimally calibrated model should have its bar graph in a diagonal shape to achieve the smallest gap area. However, the bars for TENT and SHOT are observed to fall significantly below this line, manifesting an even inferior performance compared to the source model without any adaptation. These phenomena provide evidence that both the entropy-based methods and the pseudo-label-based method could potentially harm the confidence calibration by inducing overconfidence in their predictions. In stark contrast, our TEA has significantly narrowed the gap area and improved alignment with the diagonal line. For quantitative metrics, ECE and MCE, TEA has improved by \\(2.43\\%\\) and \\(18.31\\%\\) respectively compared to the source model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The improvement in calibration of TEA over competitors may come from that neural networks inherently tend to be overconfident [14]. The softmax function enforces exponential normalization among classes, which tends to amplify the probability of dominant classes, thus inherently not advantageous for calibration. TENT and SHOT exacerbate this dominance of certain classes by reducing the un" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.691, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.089, + 0.892, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.23, + 0.69, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.23, + 0.892, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.363, + 0.894, + 0.432 + ], + "angle": 0, + "content": "Figure 6. Calibration comparison between TEA and baselines on CIFAR-10 dataset. In an ideal scenario for optimal calibration, blue bars should align with the diagonal line, and a smaller grey gap area is preferred. Quantitative measures are provided via ECE and MCE metrics, where lower values indicate better calibration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.893, + 0.537 + ], + "angle": 0, + "content": "certainty of class probabilities, further amplifying the overconfidence of the classifier. On the contrary, TEA dose not perform normalization in the label space, but maximizes the log-sum-exp of classifier logits, which essentially introduces a certain level of uncertainty to each class and empowers TEA with the ability to enhance calibration." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.55, + 0.62, + 0.566 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.895, + 0.803 + ], + "angle": 0, + "content": "To achieve test time adaptation, we introduce an innovative energy-based perspective to mitigate the impact derived from distribution shifts. The proposed TEA aims to decrease the energy of the test data within the pretrained model's distribution. TEA guides the model towards achieving a harmonious low-energy equilibrium state for the test data, which mitigates the model's distribution discrepancy and boosts its generalizability towards test distributions. Comprehensive experiments across multiple tasks, benchmarks, and architectures confirm TEA's superiority over current leading methods. Further in-depth analyses of TEA's underlying mechanisms deepen our understanding of how energy reduction can enhance the model's perception of the test distribution, ultimately paving the way for improved generalization and calibration." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.66, + 0.833 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the Strategic Priority Research Program of the CAS under Grants No. XDB0680302, and the National Key R&D Program of China (No. 2022YFB3103700, No. 2022YFB3103704)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23908" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.156 + ], + "angle": 0, + "content": "[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E. O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.469, + 0.186 + ], + "angle": 0, + "content": "[2] Stephen P Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.469, + 0.24 + ], + "angle": 0, + "content": "[3] Miguel A Carreira-Perpinan and Geoffrey Hinton. On contrastive divergence learning. In International workshop on artificial intelligence and statistics, pages 33–40. PMLR, 2005. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.469, + 0.297 + ], + "angle": 0, + "content": "[4] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.299, + 0.469, + 0.382 + ], + "angle": 0, + "content": "[5] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robust-bench: a standardized adversarial robustness benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.469, + 0.45 + ], + "angle": 0, + "content": "[6] Sabrina De Capitani Di Vimercati, Sara Foresti, Giovanni Livraga, and Pierangela Samarati. Data privacy: Definitions and techniques. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 20(06):793-817, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.469, + 0.493 + ], + "angle": 0, + "content": "[7] Yilun Du and Igor Mordatch. Implicit generation and generalization in energy-based models. arXiv preprint arXiv:1903.08689, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.469, + 0.562 + ], + "angle": 0, + "content": "[8] Yilun Du, Shuang Li, Joshua Tenenbaum, and Igor Mordatch. Improved contrastive divergence training of energy-based models. In Proceedings of the 38th International Conference on Machine Learning, pages 2837-2848. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[9] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. arXiv preprint arXiv:2103.10360, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[10] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[11] Jin Gao, Jialing Zhang, Xihui Liu, Trevor Darrell, Evan Shelhamer, and Dequan Wang. Back to the source: Diffusion-driven adaptation to test-time corruption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11786-11796, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[13] Will Grathwohl, Kuan-Chieh Wang, Joern-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. In International Conference on Learning Representations, 2020. 3, 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "conference on machine learning, pages 1321-1330. PMLR, 2017. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[15] Tian Han, Erik Nijkamp, Xiaolin Fang, Mitch Hill, Song-Chun Zhu, and Ying Nian Wu. Divergence triangle for joint training of generator model, energy-based model, and inferential model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8670-8679, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[17] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 5, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[18] Geoffrey E Hinton. Training products of experts by minimizing contrastive divergence. Neural computation, 14(8): 1771-1800, 2002. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[19] Liang Hou, Qi Cao, Yige Yuan, Songtao Zhao, Chongyang Ma, Siyuan Pan, Pengfei Wan, Zhongyuan Wang, Huawei Shen, and Xueqi Cheng. Augmentation-aware self-supervision for data-efficient gan training. arXiv preprint arXiv:2205.15677, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[20] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.503, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[21] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[22] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[23] Priyank Jain, Manasi Gyanchandani, and Nilay Khare. Big data privacy: a technological perspective and review. Journal of Big Data, 3:1-25, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[24] Minguk Jang, Sae-Young Chung, and Hye Won Chung. Test-time adaptation via self-training with nearest neighbor information. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.717, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[25] Jing Jiang. A literature survey on domain adaptation of statistical classifiers. URL: http://sifaka.cs.iuic.edu/jiang4/domainadaptation/survey, 3(1-12):3, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[26] Michael I Jordan and Tom M Mitchell. Machine learning: Trends, perspectives, and prospects. Science, 349(6245): 255-260, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.869 + ], + "angle": 0, + "content": "[27] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23909" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.47, + 0.151 + ], + "angle": 0, + "content": "[29] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.153, + 0.471, + 0.236 + ], + "angle": 0, + "content": "[30] Jogendra Nath Kundu, Akshay R Kulkarni, Suvaansh Bhambri, Deepesh Mehta, Shreyas Anand Kulkarni, Varun Jampani, and Venkatesh Babu Radhakrishnan. Balancing discriminability and transferability for source-free domain adaptation. In International Conference on Machine Learning, pages 11710-11728. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.238, + 0.47, + 0.266 + ], + "angle": 0, + "content": "[31] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.268, + 0.47, + 0.31 + ], + "angle": 0, + "content": "[32] Yann LeCun, Sumit Chopra, Raia Hadsell, M Ranzato, and Fujie Huang. A tutorial on energy-based learning. Predicting structured data, 1(0), 2006. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.312, + 0.469, + 0.34 + ], + "angle": 0, + "content": "[33] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436-444, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.342, + 0.47, + 0.398 + ], + "angle": 0, + "content": "[34] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.47, + 0.456 + ], + "angle": 0, + "content": "[35] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pages 5542-5550, 2017. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.458, + 0.47, + 0.525 + ], + "angle": 0, + "content": "[36] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.529, + 0.47, + 0.571 + ], + "angle": 0, + "content": "[37] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. arXiv preprint arXiv:2303.15361, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.573, + 0.47, + 0.627 + ], + "angle": 0, + "content": "[38] EM Lifshitz and Lev Davidovich Landau. Statistical physics, course of theoretical physics. In Part 2: Theory of the Condensed State. Butterworth-Heinemann Pergamon, London, 1980. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.47, + 0.714 + ], + "angle": 0, + "content": "[39] Fu Lin, Rohit Mittapalli, Prithvijit Chattopadhyay, Daniel Bolya, and Judy Hoffman. Likelihood landscapes: A unifying principle behind many adversarial defenses. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 39-54. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[40] Weijian Luo, Hao Jiang, Tianyang Hu, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Training energy-based models with diffusion contrastive divergences. arXiv preprint arXiv:2307.01668, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[41] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[42] Weili Nie, Arash Vahdat, and Anima Anandkumar. Controllable and compositional generation with latent-space energy-based models. Advances in Neural Information Processing Systems, 34:13497-13510, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[43] Frank Nielsen and Ke Sun. Guaranteed bounds on information-theoretic measures of univariate mixtures using piecewise log-sum-exp inequalities. Entropy, 18(12):442, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[44] Erik Nijkamp, Mitch Hill, Song-Chun Zhu, and Ying Nian Wu. Learning non-convergent non-persistent short-run mcmc toward energy-based model. Advances in Neural Information Processing Systems, 32, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[45] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yafo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[46] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yafo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.475, + 0.892, + 0.556 + ], + "angle": 0, + "content": "[48] Francesco Pinto, Harry Yang, Ser Nam Lim, Philip Torr, and Puneet Dokania. Using mixup as a regularizer can surprisingly improve accuracy & out-of-distribution robustness. In Advances in Neural Information Processing Systems, pages 14608-14622. Curran Associates, Inc., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.559, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[49] Joaquin Quinonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset shift in machine learning. Mit Press, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[50] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530-1538. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[51] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.688, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[52] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.759, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[53] Hidetoshi Shimodaira. Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of statistical planning and inference, 90(2):227-244, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[54] Yang Song and Diederik P Kingma. How to train your energy-based models. arXiv preprint arXiv:2101.03288, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[55] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23910" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "In Proceedings of the 37th International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.472, + 0.193 + ], + "angle": 0, + "content": "[56] Yushun Tang, Ce Zhang, Heng Xu, Shuoshuo Chen, Jie Cheng, Luziwei Leng, Qinghai Guo, and Zhihai He. Neuromodulated hebbian learning for fully test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3728-3738, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.471, + 0.263 + ], + "angle": 0, + "content": "[57] Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring robustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems, 33: 18583-18599, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.471, + 0.336 + ], + "angle": 0, + "content": "[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.338, + 0.47, + 0.366 + ], + "angle": 0, + "content": "[59] Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.368, + 0.471, + 0.424 + ], + "angle": 0, + "content": "[60] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.471, + 0.496 + ], + "angle": 0, + "content": "[61] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. Generalizing to unseen domains: A survey on domain generalization. IEEE Transactions on Knowledge and Data Engineering, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.471, + 0.554 + ], + "angle": 0, + "content": "[62] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688, 2011. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.556, + 0.471, + 0.626 + ], + "angle": 0, + "content": "[63] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-mond, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.628, + 0.471, + 0.669 + ], + "angle": 0, + "content": "[64] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.471, + 0.728 + ], + "angle": 0, + "content": "[65] Zehao Xiao, Xiantong Zhen, Shengcai Liao, and Cees G. M. Snoek. Energy-based test sample adaptation for domain generalization. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.471, + 0.784 + ], + "angle": 0, + "content": "[66] Xiulong Yang and Shihao Ji. Jem++: Improved techniques for training jem. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6494-6503, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[67] Yige Yuan, Bingbing Xu, Huawei Shen, Qi Cao, Keting Cen, Wen Zheng, and Xueqi Cheng. Towards generalizable graph contrastive learning: An information theory perspective. arXiv preprint arXiv:2211.10929, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[68] Yige Yuan, Bingbing Xu, Bo Lin, Liang Hou, Fei Sun, Huawei Shen, and Xueqi Cheng. Pde+: Enhancing generalization via pde with adaptive distributional diffusion. arXiv preprint arXiv:2305.15835, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.134 + ], + "angle": 0, + "content": "[69] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In British Machine Vision Conference 2016. British Machine Vision Association, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[70] Hao Zhao, Yuejiang Liu, Alexandre Alahi, and Tao Lin. On pitfalls of test-time adaptation. In ICLR 2023 Workshop on Pitfalls of limited data and computation for Trustworthy ML, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23911" + } + ] +] \ No newline at end of file diff --git a/2024/TEA_ Test-time Energy Adaptation/c8c44200-6095-44b9-b188-9639c747d9ec_origin.pdf b/2024/TEA_ Test-time Energy Adaptation/c8c44200-6095-44b9-b188-9639c747d9ec_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7348c9c2d2da0f5f29426f5cd74ca3264c138d57 --- /dev/null +++ b/2024/TEA_ Test-time Energy Adaptation/c8c44200-6095-44b9-b188-9639c747d9ec_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31c26f0b2cff8b8dea8cc51ad533c5c296b5c2080cd0271228b403e050535dd8 +size 5219683 diff --git a/2024/TEA_ Test-time Energy Adaptation/full.md b/2024/TEA_ Test-time Energy Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..65bbc78b5b0d3f810c958a4948c33ef8deff1061 --- /dev/null +++ b/2024/TEA_ Test-time Energy Adaptation/full.md @@ -0,0 +1,329 @@ +# TEA: Test-time Energy Adaptation + +Yige Yuan $^{1,2}$ , Bingbing Xu $^{1*}$ , Liang Hou $^{3}$ , Fei Sun $^{1}$ , Huawei Shen $^{1,2}$ , Xueqi Cheng $^{1,2}$ $^{1}$ CAS Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences + $^{2}$ University of Chinese Academy of Sciences Kuaishou Technology + +{yuanyige20z, xubingbing, sunfei, shenhuawei, cxq}@ict.ac.cn, lianghou96@gmail.com + +# Abstract + +Test Time Adaptation (TTA) aims to improve model generalizability when test data diverges from training distribution, with the distinct advantage of not requiring access to training data and processes, especially valuable in the context of pre-trained models. However, current TTA methods fail to address the fundamental issue: covariate shift, i.e., the decreased generalizability can be attributed to the model's reliance on the marginal distribution of the training data, which may impair model calibration and introduce confirmation bias. To address this, we propose a novel energy-based perspective, enhancing the model's perception of target data distributions without requiring access to training data or processes. Building on this perspective, we introduce Test-time Energy Adaptation (TEA), which transforms the trained classifier into an energy-based model and aligns the model's distribution with the test data's, enhancing its ability to perceive test distributions and thus improving overall generalizability. Extensive experiments across multiple tasks, benchmarks and architectures demonstrate TEA's superior generalization performance against state-of-the-art methods. Further in-depth analyses reveal that TEA can equip the model with a comprehensive perception of test distribution, ultimately paving the way toward improved generalization and calibration1. + +# 1. Introduction + +Deep neural networks, despite their remarkable performance under the assumption of independent and identically distributed (i.i.d.) training and test data [26, 33, 59], significantly degrade in real-world scenarios where unseen test data diverges from the training distribution. This limitation, known as distribution shift or domain shift, emphasizes the pressing need for generalizability across shifted test distributions [49, 57, 61, 67]. To tackle this issue, recent studies + +![](images/fa6c2f82dc88978e89460dcdfffbb24da0097b6b96bbc49a6d228c5660aae940.jpg) + +![](images/dda722254ab59e1395cee0a0e4b27cd5e972c8e1c2a0cc2110341aa13e671888.jpg) +Figure 1. Performance vs. energy on model trained with original distribution, tested across various shifted distributions. Upper: error rate change within energy score groups. Lower: loss variation with energy scores, each point denoting a distribution. Marker styles and opacity reflect distribution types and divergence. + +of Test Time Adaptation (TTA) [37, 60, 70] began considering the incorporation of unlabeled test data and leveraging it to fine-tune the source model in an unsupervised manner. This paradigm has garnered significant attention due to its ability to operate without access to training data or involvement with the training process. In the era of large open-source models [9, 58, 63], where models are publicly available but the training data and training process remain inaccessible due to privacy and resource restrictions [6, 23], TTA emerges as especially beneficial and practical. + +Existing TTA methods can be broadly categorized into three classes [37]. Normalization-based methods [41, 52] adjust the BatchNorm statistics of the model with test data statistics. Entropy-based methods [45, 46, 60] fine-tune the model by minimizing the prediction entropy. Pseudolabeling-based methods [34, 36] utilize test-time generated labels for updates. While these methods have been empi + +cally effective, these methods all fail to address a fundamental issue: covariate shift. Specifically, the decrease in generalization ability on test data with distribution shift can be attributed to the model's reliance on the marginal distribution of the training data. However, previous TTA methods do not address this shift due to their lack of connection with marginal distributions, impairing model calibration [14, 24] and introducing confirmation bias [1]. + +To combat the above challenges, we propose a novel way rooted in an energy-based perspective. Within this way, energy is defined as an unnormalized probability assigned to a sample, where a lower score corresponds to a higher likelihood of that sample within a distribution [32, 54]. Proposing such a way to improve test time adaptation is twofold. + +Firstly, test samples that correspond to lower energy within the model's distribution tend to exhibit higher performance. This is demonstrated by examining the energy scores of various test datasets in relation to a model trained on a specific training distribution. As depicted in Fig. 1, an increase in the divergence between the test and training distributions is accompanied by a drastic escalation in energy scores, leading to a significant degradation in performance. Secondly, the energy-based way can address covariate shift under TTA via directly injecting the model with a comprehensive perception of test distribution. Addressing covariate shift in TTA is particularly challenging, as it is neither feasible to access the training dataset to align the marginal distribution between training and testing [53], nor possible to modify the training process to mitigate the influence of marginal training distribution [48]. Under such circumstances, the energy-based way can directly manipulate the trained model's likelihood landscape [39] via an implicit distribution modeling process without requiring training process and training data, becoming a promising way. This stands in contrast to other models such as GANs [12, 19], Flows [50], and VAEs [28] which are advantageous only when the training data are accessible. + +Building on the above energy-based way, we propose Test-time Energy Adaptation, abbreviated as TEA, which constructs an energy-based model from the trained (classifier) model by reinterpreting the negative log-sum-exp of logits as an energy function, and employs Contrastive Divergence [18] as the adaptation objective to decrease the energy of test samples while increase that of samples generated by Stochastic Gradient Langevin Dynamics [62]. This approach prevents a trivial solution that indiscriminately reduces the energy across the entire data space to ensure an increased likelihood for target test samples within the model's distribution. TEA enables a gradual alignment between the distributions of the trained model and the test data, bolstering the trained model's perception of the test distribution and paving the way for superior adaptability and performance when confronted with the corresponding test data. + +We investigate the effectiveness of TEA under image corruption and domain generalization on four popular benchmarks CIFAR-10, CIFAR-100, TinyImageNet and PACS, across three architectures WRN-28-10, ResNet50 and ResNet-18. Experimental results underscore that TEA significantly outperforms current best-performing TTA methods in terms of generalizability, with an average increment of $4.7\%$ . We further reveal that TEA can equip the model with a comprehensive perception of the test distribution. This, in turn, significantly improves the generalization and calibration of the trained model. + +Our main contributions include: + +- Promising Way: We propose a new energy-based way for test time adaptation, which marks a departure from traditional methods and sheds light on potential avenues for mitigating the impact of distribution shifts. +- Innovative Method: We propose TEA to decrease the energy of the test data within the model's distribution, thereby equipping the trained model with a perception of the test distribution and enhancing generalizability. +- Extensive Experiments: Experiments across extensive settings validate TEA's superiority over current leading methods. Further in-depth analyses extend the understanding of how energy reduction enhances the model's perception of test distribution, ultimately paving the way toward improved generalization and calibration. + +# 2. Related Work + +Test Time Adaptation Test Time Adaptation (TTA) [37] is a paradigm aiming to enhance a model's generalizability on specific test data through unsupervised fine-tuning with these data. Note that the model is originally trained on a distinct training dataset, which is not available during the adaptation phase. Approaches like TTT [55] adapt models through self-supervised proxy task during testing but require the training of the same proxy task during training procedure. DDA [11, 65] explores adapting the test data, yet faces limitations due to model structure and training constraints. Recent research [60] highlights a scenario where the training process and training data is entirely agnostic, leading to three main categories of approaches: For normalization-based, BN [52] adapts the BatchNorm [22] statistics with test data. DUA [41] uses a tiny fraction of test data and its augmentation for BatchNorm statistics adaptation. For entropy-based, TENT [60] fine-tunes BatchNorm layers using entropy minimization during the test phase. EATA [45] employs a Fisher regularizer to limit excessive model parameter changes. SAR [46] removes high-gradient samples and promotes flat minimum weights. For pseudolabeling-based, PL [34] fine-tunes parameters using confident pseudo labels. SHOT [36] combines entropy minimization methods with pseudo labeling. + +![](images/198c4a0b0b4d7b4f169e3b9371d315286bb65caca129d77218ecc0970e42fc8d.jpg) +Figure 2. Overview of Test-time Energy Adaptation (TEA). Given a trained model (classifier) and in-coming test data, TEA directly integrates test data distribution into the trained classifier by fine-tuning its normalization layers through energy-based training: TEA constructs an Energy-Based Model from the classifier by reinterpreting the negative log-sum-exp of logits as an energy function, and employs Contrastive Divergence as the adaptation objective which decreases the energy of test samples while increases the energy of negative samples generated by Langevin Dynamics. This adaptation increases the likelihood of test samples under the classifier's distribution, enabling a gradual alignment between the distributions of the trained classifier and the test data, thereby enhancing generalizability. + +Energy Based Model Energy-Based Models (EBMs) are a type of non-normalized probabilistic models. Unlike most other probabilistic models, EBMs do not necessitate the normalizing constant to be tractable [32, 54] and do not require an explicit neural network for sample generation, implying the generation process is implicit [7]. These lead to increased flexibility in parameterization and allow for modeling a wider range of probability distributions. Due to their flexibility, EBMs can construct hybrid models with both discriminative and generative capabilities, integrating the generative competencies into discriminative models without sacrificing their discriminative capabilities [7, 13, 15]. Among these, JEM [13] is particularly representative, reinterpreting classifiers as an EBM and achieving impressive results in both classification and generation. + +# 3. Method + +In this section, we detail our method, Test-time Energy Adaptation (TEA). Initially, we present a thorough description of the notation and overall architecture in Sec. 3.1, after which we proceed to explain energy adaption (Sec. 3.2) and modulation parameters (Sec. 3.3), respectively. Furthermore, we engage in a discussion about the difference between our method and entropy-based adaptation in Sec. 3.4. + +# 3.1. Notation and Overall Architecture + +The labeled training dataset is denoted as $\{(\mathbf{x}_{\mathrm{train}},y_{\mathrm{train}})\} \subset \mathcal{X}\times \mathcal{Y}$ and the unlabeled test data are represented by $\mathbf{x}_{\mathrm{test}}\in$ $\mathcal{X}$ , where $\mathcal{X}$ and $\mathcal{Y}$ are data and label spaces. The respective marginal distributions of the training and test data are given + +by $\mathbf{x}_{\mathrm{train}}\sim p_{\mathrm{train}}(\mathbf{x})$ and $\mathbf{x}_{\mathrm{test}}\sim p_{\mathrm{test}}(\mathbf{x})$ . A classifier model trained on the training dataset and parameterized by $\theta$ , is denoted as $f_{\theta}:\mathcal{X}\to \mathcal{Y}$ . The data distribution learned by this trained classifier is denoted by $p_{\theta}(\mathbf{x})$ , which will be referred to as the model distribution henceforth. + +The overall framework of TEA is depicted in Fig. 2. The motivation behind TEA is rooted in the issue of covariate shifts [25], where the degradation of model generalization on test data $\mathbf{x}_{\mathrm{test}}$ is attributed to the model's reliance on the training distribution $p_{\mathrm{train}}(\mathbf{x})$ . To overcome this issue without accessing the training data or training process, TEA directly integrates the test data distribution into the trained model. TEA constructs an energy-based model from the trained classifier by reinterpreting the negative log-sum-exp of logits as an energy function. Through this lens, TEA employs contrastive divergence [18] as the adaptation objective, which serves to decrease the energy (increase the likelihood) of the test samples under the model distribution $p_{\theta}(\mathbf{x})$ . This adaptation enables the gradual alignment of distribution between the trained model and test data, thereby bolstering the model's perception of the test distribution and enhancing generalizability. Following previous TTA methods, TEA freezes the majority of the model's parameters, permitting only minor adjustments for efficient adaptation. + +# 3.2. Energy Adaptation for Test Distribution + +Enhancing the model's perception of test distribution from an energy-based perspective involves two key steps: constructing the energy-based model and optimizing it. + +Constructing the energy-based model. Energy-based models (EBMs) [13, 32, 54] represent a class of proba + +bilistic models that are characterized by an energy function. Consider a sample $\mathbf{x} \in \mathbb{R}^D$ , the energy function $E: \mathbb{R}^D \to \mathbb{R}$ maps each sample into an energy that can be considered as an unnormalized probability, with lower scores indicating higher likelihoods [7]. Consequently, the probability density $p(\mathbf{x})$ , as defined by EBM, can be expressed using the Boltzmann distribution [38], as shown in Eq. (1), where the partition function $Z = \int \exp(-E(\mathbf{x})) \, \mathrm{d}\mathbf{x}$ serves to normalize the probability density. + +$$ +p (\mathbf {x}) = \exp (- E (\mathbf {x})) / Z. \tag {1} +$$ + +Constructing an energy-based model from a trained classifier $f_{\theta}$ is founded on the fundamental analysis that an energy-based framework inherently underlies any discriminative model [13]. In this framework, the energy of one sample for a corresponding class can be represented as its logit produced by the trained classifier, denoted by $E_{\theta}(\mathbf{x},y) = -f_{\theta}(\mathbf{x})[y]$ . Therefore, the joint probability distribution of $\mathbf{x}$ and $y$ can be defined as, + +$$ +p _ {\theta} (\mathbf {x}, y) = \exp \left(f _ {\theta} (\mathbf {x}) [ y ]\right) / Z _ {\theta}, \tag {2} +$$ + +then the distribution of $\mathbf{x}$ can be obtained by marginalizing over $y$ , as shown below, + +$$ +p _ {\theta} (\mathbf {x}) = \sum_ {y} p _ {\theta} (\mathbf {x}, y) = \sum_ {y} \exp \left(f _ {\theta} (\mathbf {x}) [ y ]\right) / Z _ {\theta}. \tag {3} +$$ + +By substituting Eq. (3) into Eq. (1), we can obtain the form of the energy function as follows: + +$$ +E _ {\theta} (\mathbf {x}) = - \log \sum_ {y} \exp (f _ {\theta} (\mathbf {x}) [ y ]). \tag {4} +$$ + +Following the aforementioned steps, we repurpose and reinterpret the logits produced by the trained classifier to establish an energy-based model and define the energy function as the negative log-sum-exp of the logits. + +After the logits reinterpretation, we can construct a energy-based model for the test data $\mathbf{x}_{\mathrm{test}}$ using the trained classifier, where $Z_{\theta} = \int \sum_{y}\exp \left(f_{\theta}(\mathbf{x})[y]\right)\mathrm{d}\mathbf{x}$ + +$$ +p _ {\theta} \left(\mathbf {x} _ {\text {t e s t}}\right) = \frac {\exp \left(- E _ {\theta} \left(\mathbf {x} _ {\text {t e s t}}\right)\right)}{Z _ {\theta}} = \frac {\sum_ {y} \exp \left(f _ {\theta} \left(\mathbf {x} _ {\text {t e s t}}\right) [ y ]\right)}{Z _ {\theta}}. \tag {5} +$$ + +Optimizing the energy-based model. Optimizing Eq. (5) is challenging as the partition function $Z_{\theta}$ necessitates integration across the whole input space of $\mathbf{x}$ , typically making it computationally intractable. Thus, direct maximizing the log-likelihood of test data $\log p_{\theta}(\mathbf{x}_{\mathrm{test}})$ presents significant difficulties when training the parameter $\theta$ of our energy-based model. To overcome this difficulty, we propose to use contrastive divergence [3, 18] via estimating the gradient of the log-likelihood, + +$$ +\frac {\partial \log p _ {\theta} (\mathbf {x} _ {\text {t e s t}})}{\partial \theta} = \mathbb {E} _ {\tilde {\mathbf {x}} \sim p _ {\theta}} \left[ \frac {\partial E _ {\theta} (\tilde {\mathbf {x}})}{\partial \theta} \right] - \frac {\partial E _ {\theta} (\mathbf {x} _ {\text {t e s t}})}{\partial \theta}. \tag {6} +$$ + +In Eq. (6), the notation $\tilde{\mathbf{x}}\sim p_{\theta}$ denotes a random sample drawn from the distribution over $\mathbf{x}$ , which is defined by the model's distribution. The sampling procedure can be + +Algorithm 1: Test-time Energy Adaptation +Input: $f_{\theta},N,T,p_0,\alpha ,\beta ,\mathbf{x}_{\mathrm{test}}$ Output: $f_{\theta}(\mathbf{x}_{\mathrm{test}})$ +1 $E_{\theta}(\cdot)\gets -\log \sum_{y}\exp (f_{\theta}(\cdot)[y])$ +2 for $i\gets 0,1,\dots ,N - 1$ do +3 $\tilde{\mathbf{x}}_0\gets \mathrm{sample}(p_0)$ +4 for $t\gets 0,1,\ldots ,T - 1$ do +5 $\begin{array}{r}\epsilon \leftarrow \mathrm{sample}(\mathcal{N}(\mathbf{0},\mathbf{I})) \end{array}$ +6 $\tilde{\mathbf{x}}_{t + 1}\gets \tilde{\mathbf{x}}_t - \frac{\alpha}{2}\frac{\partial E_\theta(\tilde{\mathbf{x}}_t)}{\partial\tilde{\mathbf{x}}_t} +\alpha \epsilon$ +7 end +8 $\tilde{\mathbf{x}}\gets \tilde{\mathbf{x}}_{T - 1}$ +9 $\theta \gets \theta -\beta \nabla_{\theta}[E_{\theta}(\mathbf{x}_{\mathrm{test}}) - E_{\theta}(\tilde{\mathbf{x}})]$ +10 end +11 return $f_{\theta}(\underline{\mathbf{x}}_{\mathrm{test}})$ + +performed through Stochastic Gradient Langevin Dynamics (SGLD) [62], which iteratively generates samples by using the gradient information [13, 44]. In this context, $p_0$ represents an initial noise distribution, $\alpha$ denotes the step-size, and $t = 0, 1, \dots, T - 1$ is the time step. After $T$ steps of updating, a fictitious sample, generated by the energy-based model governed by the classifier, can be obtained. + +$$ +\tilde {\mathbf {x}} _ {t + 1} = \mathbf {x} _ {t} - \frac {\alpha}{2} \frac {\partial E _ {\theta} \left(\tilde {\mathbf {x}} _ {t}\right)}{\partial \tilde {\mathbf {x}} _ {t}} + \sqrt {\alpha} \boldsymbol {\epsilon}, \quad \boldsymbol {\epsilon} \sim \mathcal {N} (\mathbf {0}, \mathbf {I}), \tag {7} +$$ + +this sampling process in Eq. (7) essentially optimizes the sample by moving in the direction of energy reduction. Consequently, the objective of Eqs. (6) and (7) can be fundamentally understood as a min-max game (Eq. (8)), which minimizes the energy derived from the incoming test samples while concurrently amplifying the energy of fictitious samples obtained via SGLD from the classifier's distribution. Significantly, the latter plays a pivotal role in preventing a collapse towards a trivial solution where energy is indiscriminately minimized throughout the entire space, rather than focusing on the target test data. + +$$ +\max _ {\theta} \left[ \min _ {\tilde {\mathbf {x}}} E _ {\theta} (\tilde {\mathbf {x}}) - E _ {\theta} \left(\mathbf {x} _ {\text {t e s t}}\right) \right]. \tag {8} +$$ + +By adapting through this objective, the classifier's distribution continually converges towards the test distribution. The convergence point of the min-max game is reached when the energy of samples drawn from the classifier's distribution matches the energy of the test samples. At this stage, the classifier reaches a low-energy steady state with respect to the test data. The likelihood landscape of the classifier then exhibits a higher probability for samples that are consistent with the test data distribution, and conversely, a lower probability for samples that deviate from it, leading to a comprehensive perception of the classifier towards this test distribution and ultimately enhancing its generalizability. The pseudocode for TEA can be found in Algorithm 1. + +# 3.3. Modulation Parameters + +As outlined in Eq. (8), TEA requires updating the parameters of the trained model using the aforementioned energy adaptation to adjust to test data. In line with previous methods, we opt to update the parameters of the normalization layer due to the following two factors: (1) Practicality and Efficiency: In the era of large-scale models, the practice of fine-tuning a selected group of parameters has gained prominence [20]. For both practicality and efficiency, it's crucial to avoid updating all parameters as this process would be excessively time-consuming. Note that the parameters of the normalization layer account for a modest $1\%$ of total model parameters [60], making their update become more manageable. (2) Direct impact on data distribution: The parameters within normalization layers possess the potential to capture intrinsic features of data, thereby exerting direct influence upon the corresponding data distribution. As evidenced by [21], simple modulation to the mean and variance in a generator's normalization layers can modify image style, underlining the normalization parameters' profound impact on data distribution. This aligns well with our goal of manipulating the energy of test data to make it compatible with the model distribution. + +# 3.4. Discussion + +As entropy-based adaptation has been the representative adaptation method, we further discuss the difference between TEA and entropy-based adaptation. Intriguingly, TEA may have a connection with entropy-based adaptation, given that the negative entropy $\mathrm{NegEntropy}(\mathbf{x}) = \sum_{i}x_{i}\log x_{i}$ is the convex conjugate of the free energy $\mathrm{LogSumExp}(\mathbf{x}) = \log \sum_{i}\exp (x_{i})$ , as established in the literature [2, 43]. Despite this connection, entropy-based methods, which apply softmax normalization in the label space $y$ and strive to minimize entropy, can result in diminished uncertainty within classification probabilities, leading to compromised model calibration. In contrast, TEA, utilizing the log-sum-exp function within the data space $\mathbf{x}$ can not only effectively avoid the pitfalls associated with entropy-based methods, but also improve calibration by introduces uncertainty to each class. This conjecture has been substantiated through experiments (refer to Sec. 4.3.3). + +# 4. Experiment + +In the following sections, we compare TEA with state-of-the-art methods across various tasks, benchmarks and architectures. Then, we delve into deeper understanding of our method by exploring its desirable properties and identifying significant components that contribute to its improvements. Due to the space limitations, more comprehensive experiments including full results on corruptions and other analyses, are provided in Appendix Sec. 7. + +# 4.1. Experimental Setup + +Datasets and Metrics We focus on two tasks to verify the performance of TEA: generalization on image corruption and domain generalization. Image corruption includes clean and corrupted datasets from CIFAR-10(C), CIFAR-100(C) and TinyImageNet-200(C) [17, 29, 31], incorporating 15 types of corruption at 5 severity levels. Domain generalization considers PACS [35], encompassing 4 domains (Photo, Art Painting, Cartoon, Sketch) across 7 categories. We follow [68] to use Accuracy and mean Corruption Error (mCE) [17] as evaluation metrics. The evaluations are conducted at both the most severe level and the average of all severity levels to ensure thorough analysis. + +Backbones and Baselines We use two architectures for image corruption: WideResNet-28-10 [69] with BatchNorm [22], and ResNet-50 [16] with GroupNorm [64], consistent with the implementations of TENT [60] and SAR [46]. We use ResNet-18 [16] for domain generalization. We evaluate our method against eight leading TTA methods across three categories: (1) Normalization-based methods: BN [52] and DUA [41]. (2) Entropy-based methods: TENT [60], ETA, EATA [45], and SAR [46]; (3) Pseudo-labeling-based methods: PL [34] and SHOT [36]. Source denotes the original model without any adaptation. + +Implementation We implement methods based on PyTorch [47]. Consistency in model weight is ensured by the RobustBench protocol [5], which provides pretrained weights for the WideResNet-28-10 (BatchNorm) on CIFAR-10. In the case where RobustBench weights are unavailable, we train models in accordance with the guidelines specified in [69]. All adaptation employ Adam [27], except for SAR, which originally uses SAM [10] with SGD [51]. Baselines are replicated using their original hyper-parameters, except when these were unspecified. More details and setups are deferred to Appendix Sec. 8. + +# 4.2. Adaptation Results + +This section compares the generalizability of TEA with baselines in image corruption and domain generalization. + +Image Corruption As reported in Tab. 1, we conducted experiments on three benchmarks against eight baselines for corruption scenarios. TEA markedly surpasses all baselines on the vast majority of datasets and severity levels. Specifically, TEA outshines the best-performing baseline by an average of $4.7\%$ at the most severe level. The only exception is on TinyImageNet across all levels, where TEA ranks second, merely trailing by a minimal margin $0.1\%$ . We further incorporate the ResNet50 with GroupNorm for a broader validation. Tab. 2 indicates that TEA still delivers the best + +Table 1. Comparisons of TEA and baselines for image corruption on CIFAR-10(C), CIFAR-100(C), and TinyImageNet(C) using WRN28-10 with BatchNorm. Accuracy and mCE are evaluated at the most severe level and across all levels with asterisk (*) indicating the results are taken from the original paper [56]. The best adaptation results are highlighted in boldface. + +
WRN-28-10 BatchNormCIFAR-10(C)CIFAR-100(C)Tiny-ImageNet(C)
Clean Acc (↑)Corr Severity 5 Acc (↑)mCE (↓)Corr Severity 1-5 Acc (↑)mCE (↓)Clean Acc (↑)Corr Severity 5 Acc (↑)mCE (↓)Corr Severity 1-5 Acc (↑)mCE (↓)Clean Acc (↑)Corr Severity 5 Acc (↑)mCE (↓)Corr Severity 1-5 Acc (↑)mCE (↓)
Source94.7756.47100.0073.45100.0081.7935.39100.0052.12100.0063.1921.21100.0034.13100.00
NormBN [52]93.9779.5652.6585.6360.0080.8360.0663.5468.1169.4245.0427.7493.4234.27100.96
DUA* [41]-80.1050.78------------
PseudoPL [34]93.7551.42106.9872.6299.3780.5253.4072.1264.5375.2947.8428.2691.2239.8391.67
SHOT [36]93.2574.7763.1982.3572.6180.5256.5368.0166.0073.2847.9529.1490.1640.0191.41
EntropyTENT [60]93.6681.4148.1386.7556.1780.1463.0959.4269.4767.8039.5426.3195.5232.03104.49
ETA [45]93.9679.5852.6485.6359.9980.6559.8264.5267.1772.4043.2027.2894.1233.46102.25
EATA [45]93.9679.5952.6285.6459.9880.6860.2463.7567.4871.6643.4227.2894.0933.47102.24
SAR [46]93.9779.7751.9485.8358.9780.8462.9559.3770.0165.9941.5828.2192.8234.60100.47
EnergyTEA94.0983.3443.6987.8852.0080.8865.1056.0771.2263.7251.6531.6787.9939.9692.12
+ +Table 2. Comparisons for image corruption on CIFAR-10(C), CIFAR-100(C), and TinyImageNet(C) using ResNet-50 with GroupNorm across all severity levels. Best results are in boldface. + +
ResNet50 GroupNormCIFAR-10(C)CIFAR-100(C)Tiny-ImageNet(C)
Acc (↑)mCE (↓)Acc (↑)mCE (↓)Acc (↑)mCE (↓)
Source78.71100.0054.98100.0026.64100.00
PseudoPL79.4394.7656.6896.0226.6099.92
SHOT81.9886.6558.3193.4529.1196.73
EntropyTENT77.29102.8856.3496.8826.6599.94
ETA78.68100.0956.7296.3729.2596.42
EATA78.70100.0256.7696.2829.2596.42
SAR78.7899.6555.2899.3327.0599.41
EnergyTEA83.0579.0959.6789.3230.4194.81
+ +performance in both average accuracy and mCE. These results underscore the efficacy of TEA in handling image corruption scenarios, ensuring its universality for model architecture or normalization techniques. + +Domain Generalization Tab. 3 provides a comparison between TEA and state-of-the-art TTA methods on the PACS dataset. It is evident that when trained on the photo and art domains, TEA exhibits substantial improvements over the best-performing baseline, achieving increases of $7.1\%$ and $8.5\%$ , respectively. Compared to photo and art, cartoon and sketch domains may exhibit greater domain discrepancies, posing significant challenges to model generalization. Despite these challenging conditions, in the adaptation from cartoon to art and the sketch domain, TEA achieved improvements when all baselines experienced significant declines. TENT shows decline of $14.73\%$ compared to the source in sketch, whereas TEA bucks the trend and enhances performance by $4.5\%$ , highlighting its stability in the face of severe domain shifts. + +The results from both tasks highlight TEA's superior generalizability. These effectiveness may originate from the reduced energy, the enhanced distribution perception, and the improved calibration. In next section, we will delve into + +Table 3. Single source domain generalization comparisons on PACS datasets using ResNet-18 with BatchNorm in terms of Accuracy. The best adaptation results are highlighted in boldface. + +
Source DomainMethodTarget DomainAvg
PhotoArtCartoonSketch
PhotoSource-26.7622.4016.6221.93
BN-26.6627.9415.9623.52
TENT-26.9529.8617.5424.78
EATA-26.6628.1115.9823.59
SAR-26.7128.4115.9823.70
SHOT-26.6129.8620.9225.80
TEA-28.8133.6220.4927.64
ArtSource49.04-36.4324.4836.65
BN46.65-28.2822.7332.55
TENT50.78-30.1224.6135.17
EATA46.83-29.3123.4233.19
SAR47.90-33.0226.2735.73
SHOT50.24-34.3029.3737.97
TEA56.29-38.5728.7141.19
CartoonSource42.6929.79-29.4733.98
BN28.6825.15-20.8724.90
TENT30.9623.34-22.6525.65
EATA28.8025.10-25.0426.31
SAR29.7025.78-21.5125.66
SHOT37.7222.66-23.1427.84
TEA36.0531.44-22.8830.12
SketchSource19.9418.7032.21-23.62
BN13.4717.1429.86-20.16
TENT13.5317.3829.52-20.14
EATA13.1717.3330.08-20.19
SAR13.2918.8029.95-20.68
SHOT19.7618.7530.46-22.99
TEA19.6421.2433.19-24.69
+ +these aspects for further analysis and discussion. + +# 4.3. Analysis and Discussion + +In this section, we delve into the mechanisms driving TEA's effectiveness and explore its desirable properties. Specifically, we studied three key aspects of TEA: (1) The correlation between energy reduction and generalizability enhancement; (2) The distribution perception and generation capabilities; (3) The confidence calibration improvements. + +![](images/4840f1e565db7c005955e6e04e2c736f2afd5422126a50a80b55b082c00d9e42.jpg) + +![](images/4a841b33e495cc40209b4b86ca44091ef0f3eaba949279b03f6d2e0b81cbc7d0.jpg) +Figure 3. This illustration captures the energy reduction and generalizability enhancement achieved by TEA across CIFAR-10-C, CIFAR-100-C, and TinyImageNet-200-C, displayed from left to right. The upper set of graphs trace the evolution of energy score, corresponding loss and accuracy in response to incrementally increasing TEA adaptation steps. The lower set uncovers the extent of energy reduction and the consequent performance improvement before and after executing TEA adaptation, under different levels of distribution shift. + +![](images/151379108d2937a8d16b7cdca4910fb1524568e9392b9898ed5beb5b19bad047.jpg) + +![](images/7fa31240c1e245b963432bbcdaf3699bddb2ba78ffe8b16a01f2072dea6ce905.jpg) + +![](images/d458ea1f2acae09b7abaf870e1c7d3dc107ef5a171d1dd624f17025b904badb3.jpg) + +![](images/353a1543a5ce03d4aed30294a098ca05c48b571caad9451008f29809157e9442.jpg) + +# 4.3.1 Relation between TEA's Energy Reduction and Generalizability Enhancement + +This experiment validate TEA's energy reduction capability, and its impact on generalizability, spanning two scenarios: (1) In the first scenario, we focus on the trends of energy scores, loss, and accuracy on the same test data over increasing adaptation steps. (2) In the second scenario, we explore how the extent of energy reduction correlates with performance improvements, before and after adaptation, using varied test data that exhibit different distribution shifts. + +The results from both scenarios are depicted in Fig. 3. It is observable that $(i)$ As the iteration step of TEA increases, there is a consistent reduction in energy and corresponding decrease in loss, coupled with an ongoing enhancement in accuracy. $(ii)$ As the distribution shift increases, TEA's energy reduction becomes more pronounced. Concurrently, the enhancement in performance over the baseline also increases. $(iii)$ As the distribution shift increases, there is a sharp degradation observed in the baseline performance. However, the model adapted via TEA maintains its stability and robustness, demonstrating resilience against strong distribution shifts. In summary, these aforementioned trends are consistently observed across three datasets, demonstrating TEA's significant effectiveness in reducing energy. Notably, a greater reduction in energy correlates with an increased improvement in the model's generalizability + +# 4.3.2 TEA's Distribution Perception and Generation + +This experiment aims to validate the capability of TEA in modeling test data distribution. We framed the experiment within two scenarios: identical training and testing distributions, and shifted training and testing distributions. (1) The first scenario include four settings: (i) Source model on MNIST training set, TEA adaptation on MNIST test set; + +![](images/9a8f47726808e381e9bf96ac9165d32f1b8489a9690feaaf5ddb0a046a99f8aa.jpg) +Figure 4. Test distribution perception visualization for identical training and testing distributions on MNIST and CIFAR-10. + +![](images/c7e891cbc808f79805eac9d43182006756dc400b5b576c1338eee615640a4e11.jpg) +Figure 5. Test distribution perception visualization (upper) and real samples (lower) on shifted distribution: A model trained on PACS-A dataset then individually tested with TEA adaptation across PACS-P, PACS-A, PACS-C, PACS-S datasets. + +(ii) Source model on MNIST training set, TEA adaptation on 90-degree rotated MNIST test set; (iii) Source model on CIFAR-10 training set, TEA adaptation on CIFAR-10 test set; (iv) Source models on MNIST or CIFAR-10 training sets, tested without TEA adaptation. (2) In the second scenario, source model was trained on the PACS-A dataset and individually tested with TEA adaptation on PACS-P, PACS-A, PACS-C and PACS-S datasets. + +The outcomes from both scenarios are respectively illustrated in Fig. 4 and Fig. 5. Fig. 4 (1,3) indicates that, in scenarios where the distributions are closely identical, TEA has the potential to reconstruct samples that maintain discernible patterns. Fig. 4 (2) reveals that the sample distri + +bution indeed originates from the test data rather than provoking recollections from the training datasets. Fig. 4 (4) confirms that models without adaptation, or those adapted using other methods, fail to effectively characterize the testing distributions on both the simple MNIST dataset (upper) and the more complex CIFAR-10 dataset (lower). Drawing conclusions from Fig. 5, we can infer that under significant distribution/domain shift, our method can still characterize the key features of the shifted test distribution, such as style, texture and color schemes. + +In summary, TEA endows the model with generative ability for test data via energy-based fine-tuning of the normalization layers. This ability may incorporates generative self-supervised information from the test data into the model and improves the model's thorough understanding of the test distribution, which in turn strengthens its generalization performance on that distribution. + +# 4.3.3 TEA's Improvements in Confidence Calibration + +This experiment compares model calibration across the source model, the entropy-based method TENT, the pseudolabeling-based method SHOT, and our energy-based TEA, on CIFAR-10. In accordance with the protocol in [4, 14], we illustrate reliability histogram and compute two scalar summary statistics: Expected Calibration Error (ECE) and Maximum Calibration Error (MCE) [14], to evaluate calibration implemented by torchmetrics. The procedures are implemented as follows: For the reliability histogram, we divided the model's predictions into ten bins based on the confidence score of the highest probability class and calculated the average accuracy for each bin. + +The results are depicted in Fig. 6. From the perspective of the histogram, an optimally calibrated model should have its bar graph in a diagonal shape to achieve the smallest gap area. However, the bars for TENT and SHOT are observed to fall significantly below this line, manifesting an even inferior performance compared to the source model without any adaptation. These phenomena provide evidence that both the entropy-based methods and the pseudo-label-based method could potentially harm the confidence calibration by inducing overconfidence in their predictions. In stark contrast, our TEA has significantly narrowed the gap area and improved alignment with the diagonal line. For quantitative metrics, ECE and MCE, TEA has improved by $2.43\%$ and $18.31\%$ respectively compared to the source model. + +The improvement in calibration of TEA over competitors may come from that neural networks inherently tend to be overconfident [14]. The softmax function enforces exponential normalization among classes, which tends to amplify the probability of dominant classes, thus inherently not advantageous for calibration. TENT and SHOT exacerbate this dominance of certain classes by reducing the un + +![](images/541afa8af5637d6d201a9f4c500afd276df4c7936f75e0c1bd7bc09ce2dca5e0.jpg) + +![](images/8a8f66e1f1fa7729f3176fb5038b85b264417a8c639a08293184e284b22fbdd7.jpg) + +![](images/b5a682aa1125634be43cacd5bf5e40321b592bcb79a8543d317b6d510b9b28e2.jpg) +Figure 6. Calibration comparison between TEA and baselines on CIFAR-10 dataset. In an ideal scenario for optimal calibration, blue bars should align with the diagonal line, and a smaller grey gap area is preferred. Quantitative measures are provided via ECE and MCE metrics, where lower values indicate better calibration. + +![](images/77b50894f51975d703b78ce5ac1ec58aa9f32d217e78a851331486070e049d07.jpg) + +certainty of class probabilities, further amplifying the overconfidence of the classifier. On the contrary, TEA dose not perform normalization in the label space, but maximizes the log-sum-exp of classifier logits, which essentially introduces a certain level of uncertainty to each class and empowers TEA with the ability to enhance calibration. + +# 5. Conclusion + +To achieve test time adaptation, we introduce an innovative energy-based perspective to mitigate the impact derived from distribution shifts. The proposed TEA aims to decrease the energy of the test data within the pretrained model's distribution. TEA guides the model towards achieving a harmonious low-energy equilibrium state for the test data, which mitigates the model's distribution discrepancy and boosts its generalizability towards test distributions. Comprehensive experiments across multiple tasks, benchmarks, and architectures confirm TEA's superiority over current leading methods. Further in-depth analyses of TEA's underlying mechanisms deepen our understanding of how energy reduction can enhance the model's perception of the test distribution, ultimately paving the way for improved generalization and calibration. + +# Acknowledgments + +This work was supported by the Strategic Priority Research Program of the CAS under Grants No. XDB0680302, and the National Key R&D Program of China (No. 2022YFB3103700, No. 2022YFB3103704). + +# References + +[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E. O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning, 2020. 2 +[2] Stephen P Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. 5 +[3] Miguel A Carreira-Perpinan and Geoffrey Hinton. On contrastive divergence learning. In International workshop on artificial intelligence and statistics, pages 33–40. PMLR, 2005. 4 +[4] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 8 +[5] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robust-bench: a standardized adversarial robustness benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 5 +[6] Sabrina De Capitani Di Vimercati, Sara Foresti, Giovanni Livraga, and Pierangela Samarati. Data privacy: Definitions and techniques. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 20(06):793-817, 2012. 1 +[7] Yilun Du and Igor Mordatch. Implicit generation and generalization in energy-based models. arXiv preprint arXiv:1903.08689, 2019. 3, 4 +[8] Yilun Du, Shuang Li, Joshua Tenenbaum, and Igor Mordatch. Improved contrastive divergence training of energy-based models. In Proceedings of the 38th International Conference on Machine Learning, pages 2837-2848. PMLR, 2021. 2 +[9] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. arXiv preprint arXiv:2103.10360, 2021. 1 +[10] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, 2021. 5 +[11] Jin Gao, Jialing Zhang, Xihui Liu, Trevor Darrell, Evan Shelhamer, and Dequan Wang. Back to the source: Diffusion-driven adaptation to test-time corruption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11786-11796, 2023. 2 +[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2 +[13] Will Grathwohl, Kuan-Chieh Wang, Joern-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. In International Conference on Learning Representations, 2020. 3, 4, 2 +[14] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International + +conference on machine learning, pages 1321-1330. PMLR, 2017. 2, 8 +[15] Tian Han, Erik Nijkamp, Xiaolin Fang, Mitch Hill, Song-Chun Zhu, and Ying Nian Wu. Divergence triangle for joint training of generator model, energy-based model, and inferential model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8670-8679, 2019. 3 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[17] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 5, 1, 2 +[18] Geoffrey E Hinton. Training products of experts by minimizing contrastive divergence. Neural computation, 14(8): 1771-1800, 2002. 2, 3, 4 +[19] Liang Hou, Qi Cao, Yige Yuan, Songtao Zhao, Chongyang Ma, Siyuan Pan, Pengfei Wan, Zhongyuan Wang, Huawei Shen, and Xueqi Cheng. Augmentation-aware self-supervision for data-efficient gan training. arXiv preprint arXiv:2205.15677, 2022. 2 +[20] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 5 +[21] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 5 +[22] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 2, 5 +[23] Priyank Jain, Manasi Gyanchandani, and Nilay Khare. Big data privacy: a technological perspective and review. Journal of Big Data, 3:1-25, 2016. 1 +[24] Minguk Jang, Sae-Young Chung, and Hye Won Chung. Test-time adaptation via self-training with nearest neighbor information. In The Eleventh International Conference on Learning Representations, 2023. 2 +[25] Jing Jiang. A literature survey on domain adaptation of statistical classifiers. URL: http://sifaka.cs.iuic.edu/jiang4/domainadaptation/survey, 3(1-12):3, 2008. 3 +[26] Michael I Jordan and Tom M Mitchell. Machine learning: Trends, perspectives, and prospects. Science, 349(6245): 255-260, 2015. 1 +[27] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5 +[28] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning + +Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, 2014. 2 +[29] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5, 1 +[30] Jogendra Nath Kundu, Akshay R Kulkarni, Suvaansh Bhambri, Deepesh Mehta, Shreyas Anand Kulkarni, Varun Jampani, and Venkatesh Babu Radhakrishnan. Balancing discriminability and transferability for source-free domain adaptation. In International Conference on Machine Learning, pages 11710-11728. PMLR, 2022. 2 +[31] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 5, 1 +[32] Yann LeCun, Sumit Chopra, Raia Hadsell, M Ranzato, and Fujie Huang. A tutorial on energy-based learning. Predicting structured data, 1(0), 2006. 2, 3 +[33] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436-444, 2015. 1 +[34] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 1, 2, 5, 6 +[35] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pages 5542-5550, 2017. 5, 2 +[36] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 1, 2, 5, 6 +[37] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. arXiv preprint arXiv:2303.15361, 2023. 1, 2 +[38] EM Lifshitz and Lev Davidovich Landau. Statistical physics, course of theoretical physics. In Part 2: Theory of the Condensed State. Butterworth-Heinemann Pergamon, London, 1980. 4 +[39] Fu Lin, Rohit Mittapalli, Prithvijit Chattopadhyay, Daniel Bolya, and Judy Hoffman. Likelihood landscapes: A unifying principle behind many adversarial defenses. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 39-54. Springer, 2020. 2 +[40] Weijian Luo, Hao Jiang, Tianyang Hu, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Training energy-based models with diffusion contrastive divergences. arXiv preprint arXiv:2307.01668, 2023. 2 +[41] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 1, 2, 5, 6 +[42] Weili Nie, Arash Vahdat, and Anima Anandkumar. Controllable and compositional generation with latent-space energy-based models. Advances in Neural Information Processing Systems, 34:13497-13510, 2021. 2 + +[43] Frank Nielsen and Ke Sun. Guaranteed bounds on information-theoretic measures of univariate mixtures using piecewise log-sum-exp inequalities. Entropy, 18(12):442, 2016. 5 +[44] Erik Nijkamp, Mitch Hill, Song-Chun Zhu, and Ying Nian Wu. Learning non-convergent non-persistent short-run mcmc toward energy-based model. Advances in Neural Information Processing Systems, 32, 2019. 4 +[45] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yafo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 1, 2, 5, 6 +[46] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yafo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 5, 6 +[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5 +[48] Francesco Pinto, Harry Yang, Ser Nam Lim, Philip Torr, and Puneet Dokania. Using mixup as a regularizer can surprisingly improve accuracy & out-of-distribution robustness. In Advances in Neural Information Processing Systems, pages 14608-14622. Curran Associates, Inc., 2022. 2 +[49] Joaquin Quinonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset shift in machine learning. Mit Press, 2008. 1 +[50] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530-1538. PMLR, 2015. 2 +[51] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 5 +[52] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 1, 2, 5, 6 +[53] Hidetoshi Shimodaira. Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of statistical planning and inference, 90(2):227-244, 2000. 2 +[54] Yang Song and Diederik P Kingma. How to train your energy-based models. arXiv preprint arXiv:2101.03288, 2021. 2, 3 +[55] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. + +In Proceedings of the 37th International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 2 +[56] Yushun Tang, Ce Zhang, Heng Xu, Shuoshuo Chen, Jie Cheng, Luziwei Leng, Qinghai Guo, and Zhihai He. Neuromodulated hebbian learning for fully test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3728-3738, 2023. 6 +[57] Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring robustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems, 33: 18583-18599, 2020. 1 +[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[59] Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 1999. 1 +[60] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 1, 2, 5, 6 +[61] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. Generalizing to unseen domains: A survey on domain generalization. IEEE Transactions on Knowledge and Data Engineering, 2022. 1 +[62] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688, 2011. 2, 4 +[63] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-mond, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019. 1 +[64] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 5 +[65] Zehao Xiao, Xiantong Zhen, Shengcai Liao, and Cees G. M. Snoek. Energy-based test sample adaptation for domain generalization. In The Eleventh International Conference on Learning Representations, 2023. 2 +[66] Xiulong Yang and Shihao Ji. Jem++: Improved techniques for training jem. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6494-6503, 2021. 2 +[67] Yige Yuan, Bingbing Xu, Huawei Shen, Qi Cao, Keting Cen, Wen Zheng, and Xueqi Cheng. Towards generalizable graph contrastive learning: An information theory perspective. arXiv preprint arXiv:2211.10929, 2022. 1 +[68] Yige Yuan, Bingbing Xu, Bo Lin, Liang Hou, Fei Sun, Huawei Shen, and Xueqi Cheng. Pde+: Enhancing generalization via pde with adaptive distributional diffusion. arXiv preprint arXiv:2305.15835, 2023. 5 + +[69] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In British Machine Vision Conference 2016. British Machine Vision Association, 2016. 5 +[70] Hao Zhao, Yuejiang Liu, Alexandre Alahi, and Tao Lin. On pitfalls of test-time adaptation. In ICLR 2023 Workshop on Pitfalls of limited data and computation for Trustworthy ML, 2023. 1 \ No newline at end of file diff --git a/2024/TEA_ Test-time Energy Adaptation/images.zip b/2024/TEA_ Test-time Energy Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fb77f5a6ca3db42ef641a5608e30e9b1d6fac115 --- /dev/null +++ b/2024/TEA_ Test-time Energy Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0a64fdf74755659921e24fc26855556eacd610d731b99b2352180bde182e09b +size 722367 diff --git a/2024/TEA_ Test-time Energy Adaptation/layout.json b/2024/TEA_ Test-time Energy Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f69bfd074e0057c3616ee43730c231998096d25b --- /dev/null +++ b/2024/TEA_ Test-time Energy Adaptation/layout.json @@ -0,0 +1,8587 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 187, + 103, + 407, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 103, + 407, + 121 + ], + "spans": [ + { + "bbox": [ + 187, + 103, + 407, + 121 + ], + "type": "text", + "content": "TEA: Test-time Energy Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "spans": [ + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "type": "text", + "content": "Yige Yuan" + }, + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "type": "text", + "content": ", Bingbing Xu" + }, + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "type": "text", + "content": ", Liang Hou" + }, + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "type": "text", + "content": ", Fei Sun" + }, + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "type": "text", + "content": ", Huawei Shen" + }, + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "type": "text", + "content": ", Xueqi Cheng" + }, + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "type": "text", + "content": "CAS Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences \n" + }, + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 549, + 185 + ], + "type": "text", + "content": "University of Chinese Academy of Sciences Kuaishou Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 186, + 521, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 186, + 521, + 200 + ], + "spans": [ + { + "bbox": [ + 78, + 186, + 521, + 200 + ], + "type": "text", + "content": "{yuanyige20z, xubingbing, sunfei, shenhuawei, cxq}@ict.ac.cn, lianghou96@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 252, + 290, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 290, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 290, + 540 + ], + "type": "text", + "content": "Test Time Adaptation (TTA) aims to improve model generalizability when test data diverges from training distribution, with the distinct advantage of not requiring access to training data and processes, especially valuable in the context of pre-trained models. However, current TTA methods fail to address the fundamental issue: covariate shift, i.e., the decreased generalizability can be attributed to the model's reliance on the marginal distribution of the training data, which may impair model calibration and introduce confirmation bias. To address this, we propose a novel energy-based perspective, enhancing the model's perception of target data distributions without requiring access to training data or processes. Building on this perspective, we introduce Test-time Energy Adaptation (TEA), which transforms the trained classifier into an energy-based model and aligns the model's distribution with the test data's, enhancing its ability to perceive test distributions and thus improving overall generalizability. Extensive experiments across multiple tasks, benchmarks and architectures demonstrate TEA's superior generalization performance against state-of-the-art methods. Further in-depth analyses reveal that TEA can equip the model with a comprehensive perception of test distribution, ultimately paving the way toward improved generalization and calibration1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 565, + 128, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 565, + 128, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 565, + 128, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 586, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 287, + 682 + ], + "type": "text", + "content": "Deep neural networks, despite their remarkable performance under the assumption of independent and identically distributed (i.i.d.) training and test data [26, 33, 59], significantly degrade in real-world scenarios where unseen test data diverges from the training distribution. This limitation, known as distribution shift or domain shift, emphasizes the pressing need for generalizability across shifted test distributions [49, 57, 61, 67]. To tackle this issue, recent studies" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 228, + 533, + 330 + ], + "blocks": [ + { + "bbox": [ + 319, + 228, + 533, + 330 + ], + "lines": [ + { + "bbox": [ + 319, + 228, + 533, + 330 + ], + "spans": [ + { + "bbox": [ + 319, + 228, + 533, + 330 + ], + "type": "image", + "image_path": "fa6c2f82dc88978e89460dcdfffbb24da0097b6b96bbc49a6d228c5660aae940.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 331, + 533, + 434 + ], + "blocks": [ + { + "bbox": [ + 320, + 331, + 533, + 434 + ], + "lines": [ + { + "bbox": [ + 320, + 331, + 533, + 434 + ], + "spans": [ + { + "bbox": [ + 320, + 331, + 533, + 434 + ], + "type": "image", + "image_path": "dda722254ab59e1395cee0a0e4b27cd5e972c8e1c2a0cc2110341aa13e671888.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 437, + 547, + 494 + ], + "lines": [ + { + "bbox": [ + 305, + 437, + 547, + 494 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 547, + 494 + ], + "type": "text", + "content": "Figure 1. Performance vs. energy on model trained with original distribution, tested across various shifted distributions. Upper: error rate change within energy score groups. Lower: loss variation with energy scores, each point denoting a distribution. Marker styles and opacity reflect distribution types and divergence." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 508, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 546, + 628 + ], + "type": "text", + "content": "of Test Time Adaptation (TTA) [37, 60, 70] began considering the incorporation of unlabeled test data and leveraging it to fine-tune the source model in an unsupervised manner. This paradigm has garnered significant attention due to its ability to operate without access to training data or involvement with the training process. In the era of large open-source models [9, 58, 63], where models are publicly available but the training data and training process remain inaccessible due to privacy and resource restrictions [6, 23], TTA emerges as especially beneficial and practical." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Existing TTA methods can be broadly categorized into three classes [37]. Normalization-based methods [41, 52] adjust the BatchNorm statistics of the model with test data statistics. Entropy-based methods [45, 46, 60] fine-tune the model by minimizing the prediction entropy. Pseudolabeling-based methods [34, 36] utilize test-time generated labels for updates. While these methods have been empi" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 136, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 136, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 136, + 703 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 58, + 703, + 277, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 277, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 277, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 703, + 277, + 713 + ], + "type": "text", + "content": "Code is available at https://github.com/yuanyige/tea" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "23901" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "cally effective, these methods all fail to address a fundamental issue: covariate shift. Specifically, the decrease in generalization ability on test data with distribution shift can be attributed to the model's reliance on the marginal distribution of the training data. However, previous TTA methods do not address this shift due to their lack of connection with marginal distributions, impairing model calibration [14, 24] and introducing confirmation bias [1]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 171, + 287, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 242 + ], + "type": "text", + "content": "To combat the above challenges, we propose a novel way rooted in an energy-based perspective. Within this way, energy is defined as an unnormalized probability assigned to a sample, where a lower score corresponds to a higher likelihood of that sample within a distribution [32, 54]. Proposing such a way to improve test time adaptation is twofold." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 245, + 288, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 288, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 288, + 519 + ], + "type": "text", + "content": "Firstly, test samples that correspond to lower energy within the model's distribution tend to exhibit higher performance. This is demonstrated by examining the energy scores of various test datasets in relation to a model trained on a specific training distribution. As depicted in Fig. 1, an increase in the divergence between the test and training distributions is accompanied by a drastic escalation in energy scores, leading to a significant degradation in performance. Secondly, the energy-based way can address covariate shift under TTA via directly injecting the model with a comprehensive perception of test distribution. Addressing covariate shift in TTA is particularly challenging, as it is neither feasible to access the training dataset to align the marginal distribution between training and testing [53], nor possible to modify the training process to mitigate the influence of marginal training distribution [48]. Under such circumstances, the energy-based way can directly manipulate the trained model's likelihood landscape [39] via an implicit distribution modeling process without requiring training process and training data, becoming a promising way. This stands in contrast to other models such as GANs [12, 19], Flows [50], and VAEs [28] which are advantageous only when the training data are accessible." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": "Building on the above energy-based way, we propose Test-time Energy Adaptation, abbreviated as TEA, which constructs an energy-based model from the trained (classifier) model by reinterpreting the negative log-sum-exp of logits as an energy function, and employs Contrastive Divergence [18] as the adaptation objective to decrease the energy of test samples while increase that of samples generated by Stochastic Gradient Langevin Dynamics [62]. This approach prevents a trivial solution that indiscriminately reduces the energy across the entire data space to ensure an increased likelihood for target test samples within the model's distribution. TEA enables a gradual alignment between the distributions of the trained model and the test data, bolstering the trained model's perception of the test distribution and paving the way for superior adaptability and performance when confronted with the corresponding test data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "We investigate the effectiveness of TEA under image corruption and domain generalization on four popular benchmarks CIFAR-10, CIFAR-100, TinyImageNet and PACS, across three architectures WRN-28-10, ResNet50 and ResNet-18. Experimental results underscore that TEA significantly outperforms current best-performing TTA methods in terms of generalizability, with an average increment of " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": ". We further reveal that TEA can equip the model with a comprehensive perception of the test distribution. This, in turn, significantly improves the generalization and calibration of the trained model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 206, + 448, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 448, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 448, + 217 + ], + "type": "text", + "content": "Our main contributions include:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 220, + 545, + 387 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 220, + 545, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 220, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 545, + 267 + ], + "type": "text", + "content": "- Promising Way: We propose a new energy-based way for test time adaptation, which marks a departure from traditional methods and sheds light on potential avenues for mitigating the impact of distribution shifts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 267, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 267, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 267, + 545, + 316 + ], + "type": "text", + "content": "- Innovative Method: We propose TEA to decrease the energy of the test data within the model's distribution, thereby equipping the trained model with a perception of the test distribution and enhancing generalizability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 316, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 316, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 545, + 387 + ], + "type": "text", + "content": "- Extensive Experiments: Experiments across extensive settings validate TEA's superiority over current leading methods. Further in-depth analyses extend the understanding of how energy reduction enhances the model's perception of test distribution, ultimately paving the way toward improved generalization and calibration." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 404, + 392, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 404, + 392, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 404, + 392, + 417 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 426, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 713 + ], + "type": "text", + "content": "Test Time Adaptation Test Time Adaptation (TTA) [37] is a paradigm aiming to enhance a model's generalizability on specific test data through unsupervised fine-tuning with these data. Note that the model is originally trained on a distinct training dataset, which is not available during the adaptation phase. Approaches like TTT [55] adapt models through self-supervised proxy task during testing but require the training of the same proxy task during training procedure. DDA [11, 65] explores adapting the test data, yet faces limitations due to model structure and training constraints. Recent research [60] highlights a scenario where the training process and training data is entirely agnostic, leading to three main categories of approaches: For normalization-based, BN [52] adapts the BatchNorm [22] statistics with test data. DUA [41] uses a tiny fraction of test data and its augmentation for BatchNorm statistics adaptation. For entropy-based, TENT [60] fine-tunes BatchNorm layers using entropy minimization during the test phase. EATA [45] employs a Fisher regularizer to limit excessive model parameter changes. SAR [46] removes high-gradient samples and promotes flat minimum weights. For pseudolabeling-based, PL [34] fine-tunes parameters using confident pseudo labels. SHOT [36] combines entropy minimization methods with pseudo labeling." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23902" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 537, + 259 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 537, + 259 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 537, + 259 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 537, + 259 + ], + "type": "image", + "image_path": "198c4a0b0b4d7b4f169e3b9371d315286bb65caca129d77218ecc0970e42fc8d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 545, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 545, + 329 + ], + "type": "text", + "content": "Figure 2. Overview of Test-time Energy Adaptation (TEA). Given a trained model (classifier) and in-coming test data, TEA directly integrates test data distribution into the trained classifier by fine-tuning its normalization layers through energy-based training: TEA constructs an Energy-Based Model from the classifier by reinterpreting the negative log-sum-exp of logits as an energy function, and employs Contrastive Divergence as the adaptation objective which decreases the energy of test samples while increases the energy of negative samples generated by Langevin Dynamics. This adaptation increases the likelihood of test samples under the classifier's distribution, enabling a gradual alignment between the distributions of the trained classifier and the test data, thereby enhancing generalizability." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 340, + 289, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 289, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 289, + 520 + ], + "type": "text", + "content": "Energy Based Model Energy-Based Models (EBMs) are a type of non-normalized probabilistic models. Unlike most other probabilistic models, EBMs do not necessitate the normalizing constant to be tractable [32, 54] and do not require an explicit neural network for sample generation, implying the generation process is implicit [7]. These lead to increased flexibility in parameterization and allow for modeling a wider range of probability distributions. Due to their flexibility, EBMs can construct hybrid models with both discriminative and generative capabilities, integrating the generative competencies into discriminative models without sacrificing their discriminative capabilities [7, 13, 15]. Among these, JEM [13] is particularly representative, reinterpreting classifiers as an EBM and achieving impressive results in both classification and generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 532, + 103, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 103, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 103, + 544 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 553, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 287, + 636 + ], + "type": "text", + "content": "In this section, we detail our method, Test-time Energy Adaptation (TEA). Initially, we present a thorough description of the notation and overall architecture in Sec. 3.1, after which we proceed to explain energy adaption (Sec. 3.2) and modulation parameters (Sec. 3.3), respectively. Furthermore, we engage in a discussion about the difference between our method and entropy-based adaptation in Sec. 3.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 647, + 231, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 231, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 231, + 658 + ], + "type": "text", + "content": "3.1. Notation and Overall Architecture" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": "The labeled training dataset is denoted as " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{x}_{\\mathrm{train}},y_{\\mathrm{train}})\\} \\subset \\mathcal{X}\\times \\mathcal{Y}" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": " and the unlabeled test data are represented by " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{test}}\\in" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": " are data and label spaces. The respective marginal distributions of the training and test data are given" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 340, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 398 + ], + "type": "text", + "content": "by " + }, + { + "bbox": [ + 304, + 340, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{train}}\\sim p_{\\mathrm{train}}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 340, + 545, + 398 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 340, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{test}}\\sim p_{\\mathrm{test}}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 340, + 545, + 398 + ], + "type": "text", + "content": ". A classifier model trained on the training dataset and parameterized by " + }, + { + "bbox": [ + 304, + 340, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 340, + 545, + 398 + ], + "type": "text", + "content": ", is denoted as " + }, + { + "bbox": [ + 304, + 340, + 545, + 398 + ], + "type": "inline_equation", + "content": "f_{\\theta}:\\mathcal{X}\\to \\mathcal{Y}" + }, + { + "bbox": [ + 304, + 340, + 545, + 398 + ], + "type": "text", + "content": ". The data distribution learned by this trained classifier is denoted by " + }, + { + "bbox": [ + 304, + 340, + 545, + 398 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 340, + 545, + 398 + ], + "type": "text", + "content": ", which will be referred to as the model distribution henceforth." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 399, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 546, + 628 + ], + "type": "text", + "content": "The overall framework of TEA is depicted in Fig. 2. The motivation behind TEA is rooted in the issue of covariate shifts [25], where the degradation of model generalization on test data " + }, + { + "bbox": [ + 304, + 399, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{test}}" + }, + { + "bbox": [ + 304, + 399, + 546, + 628 + ], + "type": "text", + "content": " is attributed to the model's reliance on the training distribution " + }, + { + "bbox": [ + 304, + 399, + 546, + 628 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{train}}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 399, + 546, + 628 + ], + "type": "text", + "content": ". To overcome this issue without accessing the training data or training process, TEA directly integrates the test data distribution into the trained model. TEA constructs an energy-based model from the trained classifier by reinterpreting the negative log-sum-exp of logits as an energy function. Through this lens, TEA employs contrastive divergence [18] as the adaptation objective, which serves to decrease the energy (increase the likelihood) of the test samples under the model distribution " + }, + { + "bbox": [ + 304, + 399, + 546, + 628 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 399, + 546, + 628 + ], + "type": "text", + "content": ". This adaptation enables the gradual alignment of distribution between the trained model and test data, thereby bolstering the model's perception of the test distribution and enhancing generalizability. Following previous TTA methods, TEA freezes the majority of the model's parameters, permitting only minor adjustments for efficient adaptation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 635, + 517, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 517, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 517, + 647 + ], + "type": "text", + "content": "3.2. Energy Adaptation for Test Distribution" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": "Enhancing the model's perception of test distribution from an energy-based perspective involves two key steps: constructing the energy-based model and optimizing it." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Constructing the energy-based model. Energy-based models (EBMs) [13, 32, 54] represent a class of proba" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23903" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "text", + "content": "bilistic models that are characterized by an energy function. Consider a sample " + }, + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "text", + "content": ", the energy function " + }, + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "inline_equation", + "content": "E: \\mathbb{R}^D \\to \\mathbb{R}" + }, + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "text", + "content": " maps each sample into an energy that can be considered as an unnormalized probability, with lower scores indicating higher likelihoods [7]. Consequently, the probability density " + }, + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "text", + "content": ", as defined by EBM, can be expressed using the Boltzmann distribution [38], as shown in Eq. (1), where the partition function " + }, + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "inline_equation", + "content": "Z = \\int \\exp(-E(\\mathbf{x})) \\, \\mathrm{d}\\mathbf{x}" + }, + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "text", + "content": " serves to normalize the probability density." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 183, + 287, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 183, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 116, + 183, + 287, + 196 + ], + "type": "interline_equation", + "content": "p (\\mathbf {x}) = \\exp (- E (\\mathbf {x})) / Z. \\tag {1}", + "image_path": "1098ef053de095e7341397b026632241ee8f30ac0f24cdf9a76bbbec1fa6df95.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 199, + 287, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 199, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 287, + 294 + ], + "type": "text", + "content": "Constructing an energy-based model from a trained classifier " + }, + { + "bbox": [ + 46, + 199, + 287, + 294 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 199, + 287, + 294 + ], + "type": "text", + "content": " is founded on the fundamental analysis that an energy-based framework inherently underlies any discriminative model [13]. In this framework, the energy of one sample for a corresponding class can be represented as its logit produced by the trained classifier, denoted by " + }, + { + "bbox": [ + 46, + 199, + 287, + 294 + ], + "type": "inline_equation", + "content": "E_{\\theta}(\\mathbf{x},y) = -f_{\\theta}(\\mathbf{x})[y]" + }, + { + "bbox": [ + 46, + 199, + 287, + 294 + ], + "type": "text", + "content": ". Therefore, the joint probability distribution of " + }, + { + "bbox": [ + 46, + 199, + 287, + 294 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 199, + 287, + 294 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 199, + 287, + 294 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 199, + 287, + 294 + ], + "type": "text", + "content": " can be defined as," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 298, + 287, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 298, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 104, + 298, + 287, + 312 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (\\mathbf {x}, y) = \\exp \\left(f _ {\\theta} (\\mathbf {x}) [ y ]\\right) / Z _ {\\theta}, \\tag {2}", + "image_path": "2f0b8bbe5f500ca8a5156a09d90271e7cbc50f6209e52059fa944977955b0158.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 314, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 337 + ], + "type": "text", + "content": "then the distribution of " + }, + { + "bbox": [ + 46, + 314, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 314, + 287, + 337 + ], + "type": "text", + "content": " can be obtained by marginalizing over " + }, + { + "bbox": [ + 46, + 314, + 287, + 337 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 314, + 287, + 337 + ], + "type": "text", + "content": ", as shown below," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 340, + 287, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 340, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 63, + 340, + 287, + 365 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (\\mathbf {x}) = \\sum_ {y} p _ {\\theta} (\\mathbf {x}, y) = \\sum_ {y} \\exp \\left(f _ {\\theta} (\\mathbf {x}) [ y ]\\right) / Z _ {\\theta}. \\tag {3}", + "image_path": "836a35789ecf4090f35c4e889ef64eec46fe85d869e2c1eceb1f77ea1cc4af3f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 369, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 392 + ], + "type": "text", + "content": "By substituting Eq. (3) into Eq. (1), we can obtain the form of the energy function as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 395, + 287, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 395, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 94, + 395, + 287, + 420 + ], + "type": "interline_equation", + "content": "E _ {\\theta} (\\mathbf {x}) = - \\log \\sum_ {y} \\exp (f _ {\\theta} (\\mathbf {x}) [ y ]). \\tag {4}", + "image_path": "cb5ce0fb6c5b340d2b502681f408b21bde043547bd550fd1edbc8b03a0fdb63b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 423, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 471 + ], + "type": "text", + "content": "Following the aforementioned steps, we repurpose and reinterpret the logits produced by the trained classifier to establish an energy-based model and define the energy function as the negative log-sum-exp of the logits." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 472, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 509 + ], + "type": "text", + "content": "After the logits reinterpretation, we can construct a energy-based model for the test data " + }, + { + "bbox": [ + 46, + 472, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{test}}" + }, + { + "bbox": [ + 46, + 472, + 287, + 509 + ], + "type": "text", + "content": " using the trained classifier, where " + }, + { + "bbox": [ + 46, + 472, + 287, + 509 + ], + "type": "inline_equation", + "content": "Z_{\\theta} = \\int \\sum_{y}\\exp \\left(f_{\\theta}(\\mathbf{x})[y]\\right)\\mathrm{d}\\mathbf{x}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 511, + 287, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 287, + 537 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {x} _ {\\text {t e s t}}\\right) = \\frac {\\exp \\left(- E _ {\\theta} \\left(\\mathbf {x} _ {\\text {t e s t}}\\right)\\right)}{Z _ {\\theta}} = \\frac {\\sum_ {y} \\exp \\left(f _ {\\theta} \\left(\\mathbf {x} _ {\\text {t e s t}}\\right) [ y ]\\right)}{Z _ {\\theta}}. \\tag {5}", + "image_path": "faa258fafc274115909e4c92927acfcec8849f1224187342c34b7b57c7b997a7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "text", + "content": "Optimizing the energy-based model. Optimizing Eq. (5) is challenging as the partition function " + }, + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "inline_equation", + "content": "Z_{\\theta}" + }, + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "text", + "content": " necessitates integration across the whole input space of " + }, + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "text", + "content": ", typically making it computationally intractable. Thus, direct maximizing the log-likelihood of test data " + }, + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\log p_{\\theta}(\\mathbf{x}_{\\mathrm{test}})" + }, + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "text", + "content": " presents significant difficulties when training the parameter " + }, + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "text", + "content": " of our energy-based model. To overcome this difficulty, we propose to use contrastive divergence [3, 18] via estimating the gradient of the log-likelihood," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 649, + 287, + 676 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 649, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 649, + 287, + 676 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\log p _ {\\theta} (\\mathbf {x} _ {\\text {t e s t}})}{\\partial \\theta} = \\mathbb {E} _ {\\tilde {\\mathbf {x}} \\sim p _ {\\theta}} \\left[ \\frac {\\partial E _ {\\theta} (\\tilde {\\mathbf {x}})}{\\partial \\theta} \\right] - \\frac {\\partial E _ {\\theta} (\\mathbf {x} _ {\\text {t e s t}})}{\\partial \\theta}. \\tag {6}", + "image_path": "e5af6e2bb8032992f64ef8d875ceeb05711f3f1636fa807df4467f2241197a61.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "In Eq. (6), the notation " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}\\sim p_{\\theta}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " denotes a random sample drawn from the distribution over " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": ", which is defined by the model's distribution. The sampling procedure can be" + } + ] + } + ], + "index": 13 + }, + { + "type": "code", + "bbox": [ + 306, + 88, + 478, + 254 + ], + "blocks": [ + { + "bbox": [ + 312, + 75, + 488, + 87 + ], + "lines": [ + { + "bbox": [ + 312, + 75, + 488, + 87 + ], + "spans": [ + { + "bbox": [ + 312, + 75, + 488, + 87 + ], + "type": "text", + "content": "Algorithm 1: Test-time Energy Adaptation" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "lines": [ + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "spans": [ + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "inline_equation", + "content": "f_{\\theta},N,T,p_0,\\alpha ,\\beta ,\\mathbf{x}_{\\mathrm{test}}" + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "text", + "content": " Output: " + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathbf{x}_{\\mathrm{test}})" + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "text", + "content": " \n1 " + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "inline_equation", + "content": "E_{\\theta}(\\cdot)\\gets -\\log \\sum_{y}\\exp (f_{\\theta}(\\cdot)[y])" + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "text", + "content": " \n2 for " + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "inline_equation", + "content": "i\\gets 0,1,\\dots ,N - 1" + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "text", + "content": " do \n3 " + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_0\\gets \\mathrm{sample}(p_0)" + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "text", + "content": " \n4 for " + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "inline_equation", + "content": "t\\gets 0,1,\\ldots ,T - 1" + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "text", + "content": " do \n5 " + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "inline_equation", + "content": "\\begin{array}{r}\\epsilon \\leftarrow \\mathrm{sample}(\\mathcal{N}(\\mathbf{0},\\mathbf{I})) \\end{array}" + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "text", + "content": " \n6 " + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_{t + 1}\\gets \\tilde{\\mathbf{x}}_t - \\frac{\\alpha}{2}\\frac{\\partial E_\\theta(\\tilde{\\mathbf{x}}_t)}{\\partial\\tilde{\\mathbf{x}}_t} +\\alpha \\epsilon" + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "text", + "content": " \n7 end \n8 " + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}\\gets \\tilde{\\mathbf{x}}_{T - 1}" + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "text", + "content": " \n9 " + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "inline_equation", + "content": "\\theta \\gets \\theta -\\beta \\nabla_{\\theta}[E_{\\theta}(\\mathbf{x}_{\\mathrm{test}}) - E_{\\theta}(\\tilde{\\mathbf{x}})]" + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "text", + "content": " \n10 end \n11 return " + }, + { + "bbox": [ + 306, + 88, + 478, + 254 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\underline{\\mathbf{x}}_{\\mathrm{test}})" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "code_body" + } + ], + "index": 15, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 270, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 353 + ], + "type": "text", + "content": "performed through Stochastic Gradient Langevin Dynamics (SGLD) [62], which iteratively generates samples by using the gradient information [13, 44]. In this context, " + }, + { + "bbox": [ + 304, + 270, + 545, + 353 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 304, + 270, + 545, + 353 + ], + "type": "text", + "content": " represents an initial noise distribution, " + }, + { + "bbox": [ + 304, + 270, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 270, + 545, + 353 + ], + "type": "text", + "content": " denotes the step-size, and " + }, + { + "bbox": [ + 304, + 270, + 545, + 353 + ], + "type": "inline_equation", + "content": "t = 0, 1, \\dots, T - 1" + }, + { + "bbox": [ + 304, + 270, + 545, + 353 + ], + "type": "text", + "content": " is the time step. After " + }, + { + "bbox": [ + 304, + 270, + 545, + 353 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 270, + 545, + 353 + ], + "type": "text", + "content": " steps of updating, a fictitious sample, generated by the energy-based model governed by the classifier, can be obtained." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 360, + 545, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 360, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 318, + 360, + 545, + 385 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {x}} _ {t + 1} = \\mathbf {x} _ {t} - \\frac {\\alpha}{2} \\frac {\\partial E _ {\\theta} \\left(\\tilde {\\mathbf {x}} _ {t}\\right)}{\\partial \\tilde {\\mathbf {x}} _ {t}} + \\sqrt {\\alpha} \\boldsymbol {\\epsilon}, \\quad \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), \\tag {7}", + "image_path": "cf5357d3138e3a8ec3a456db07e634dcf872506e42c18e9f57f168cb216de529.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 390, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 521 + ], + "type": "text", + "content": "this sampling process in Eq. (7) essentially optimizes the sample by moving in the direction of energy reduction. Consequently, the objective of Eqs. (6) and (7) can be fundamentally understood as a min-max game (Eq. (8)), which minimizes the energy derived from the incoming test samples while concurrently amplifying the energy of fictitious samples obtained via SGLD from the classifier's distribution. Significantly, the latter plays a pivotal role in preventing a collapse towards a trivial solution where energy is indiscriminately minimized throughout the entire space, rather than focusing on the target test data." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 358, + 529, + 545, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 529, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 358, + 529, + 545, + 550 + ], + "type": "interline_equation", + "content": "\\max _ {\\theta} \\left[ \\min _ {\\tilde {\\mathbf {x}}} E _ {\\theta} (\\tilde {\\mathbf {x}}) - E _ {\\theta} \\left(\\mathbf {x} _ {\\text {t e s t}}\\right) \\right]. \\tag {8}", + "image_path": "33d18588e84b44226824f6e99505cdddd6cee1f2793567d8559dec0d52771c70.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "By adapting through this objective, the classifier's distribution continually converges towards the test distribution. The convergence point of the min-max game is reached when the energy of samples drawn from the classifier's distribution matches the energy of the test samples. At this stage, the classifier reaches a low-energy steady state with respect to the test data. The likelihood landscape of the classifier then exhibits a higher probability for samples that are consistent with the test data distribution, and conversely, a lower probability for samples that deviate from it, leading to a comprehensive perception of the classifier towards this test distribution and ultimately enhancing its generalizability. The pseudocode for TEA can be found in Algorithm 1." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23904" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 182, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 182, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 182, + 83 + ], + "type": "text", + "content": "3.3. Modulation Parameters" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 289, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 289, + 355 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 289, + 355 + ], + "type": "text", + "content": "As outlined in Eq. (8), TEA requires updating the parameters of the trained model using the aforementioned energy adaptation to adjust to test data. In line with previous methods, we opt to update the parameters of the normalization layer due to the following two factors: (1) Practicality and Efficiency: In the era of large-scale models, the practice of fine-tuning a selected group of parameters has gained prominence [20]. For both practicality and efficiency, it's crucial to avoid updating all parameters as this process would be excessively time-consuming. Note that the parameters of the normalization layer account for a modest " + }, + { + "bbox": [ + 47, + 91, + 289, + 355 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 47, + 91, + 289, + 355 + ], + "type": "text", + "content": " of total model parameters [60], making their update become more manageable. (2) Direct impact on data distribution: The parameters within normalization layers possess the potential to capture intrinsic features of data, thereby exerting direct influence upon the corresponding data distribution. As evidenced by [21], simple modulation to the mean and variance in a generator's normalization layers can modify image style, underlining the normalization parameters' profound impact on data distribution. This aligns well with our goal of manipulating the energy of test data to make it compatible with the model distribution." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 363, + 119, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 119, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 119, + 374 + ], + "type": "text", + "content": "3.4. Discussion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 381, + 289, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 289, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 289, + 586 + ], + "type": "text", + "content": "As entropy-based adaptation has been the representative adaptation method, we further discuss the difference between TEA and entropy-based adaptation. Intriguingly, TEA may have a connection with entropy-based adaptation, given that the negative entropy " + }, + { + "bbox": [ + 47, + 381, + 289, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{NegEntropy}(\\mathbf{x}) = \\sum_{i}x_{i}\\log x_{i}" + }, + { + "bbox": [ + 47, + 381, + 289, + 586 + ], + "type": "text", + "content": " is the convex conjugate of the free energy " + }, + { + "bbox": [ + 47, + 381, + 289, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{LogSumExp}(\\mathbf{x}) = \\log \\sum_{i}\\exp (x_{i})" + }, + { + "bbox": [ + 47, + 381, + 289, + 586 + ], + "type": "text", + "content": ", as established in the literature [2, 43]. Despite this connection, entropy-based methods, which apply softmax normalization in the label space " + }, + { + "bbox": [ + 47, + 381, + 289, + 586 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 381, + 289, + 586 + ], + "type": "text", + "content": " and strive to minimize entropy, can result in diminished uncertainty within classification probabilities, leading to compromised model calibration. In contrast, TEA, utilizing the log-sum-exp function within the data space " + }, + { + "bbox": [ + 47, + 381, + 289, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 381, + 289, + 586 + ], + "type": "text", + "content": " can not only effectively avoid the pitfalls associated with entropy-based methods, but also improve calibration by introduces uncertainty to each class. This conjecture has been substantiated through experiments (refer to Sec. 4.3.3)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 597, + 124, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 124, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 124, + 611 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "text", + "content": "In the following sections, we compare TEA with state-of-the-art methods across various tasks, benchmarks and architectures. Then, we delve into deeper understanding of our method by exploring its desirable properties and identifying significant components that contribute to its improvements. Due to the space limitations, more comprehensive experiments including full results on corruptions and other analyses, are provided in Appendix Sec. 7." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 422, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 422, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 422, + 85 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 90, + 547, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 547, + 236 + ], + "type": "text", + "content": "Datasets and Metrics We focus on two tasks to verify the performance of TEA: generalization on image corruption and domain generalization. Image corruption includes clean and corrupted datasets from CIFAR-10(C), CIFAR-100(C) and TinyImageNet-200(C) [17, 29, 31], incorporating 15 types of corruption at 5 severity levels. Domain generalization considers PACS [35], encompassing 4 domains (Photo, Art Painting, Cartoon, Sketch) across 7 categories. We follow [68] to use Accuracy and mean Corruption Error (mCE) [17] as evaluation metrics. The evaluations are conducted at both the most severe level and the average of all severity levels to ensure thorough analysis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 249, + 547, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 249, + 547, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 249, + 547, + 382 + ], + "type": "text", + "content": "Backbones and Baselines We use two architectures for image corruption: WideResNet-28-10 [69] with BatchNorm [22], and ResNet-50 [16] with GroupNorm [64], consistent with the implementations of TENT [60] and SAR [46]. We use ResNet-18 [16] for domain generalization. We evaluate our method against eight leading TTA methods across three categories: (1) Normalization-based methods: BN [52] and DUA [41]. (2) Entropy-based methods: TENT [60], ETA, EATA [45], and SAR [46]; (3) Pseudo-labeling-based methods: PL [34] and SHOT [36]. Source denotes the original model without any adaptation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "text", + "content": "Implementation We implement methods based on PyTorch [47]. Consistency in model weight is ensured by the RobustBench protocol [5], which provides pretrained weights for the WideResNet-28-10 (BatchNorm) on CIFAR-10. In the case where RobustBench weights are unavailable, we train models in accordance with the guidelines specified in [69]. All adaptation employ Adam [27], except for SAR, which originally uses SAM [10] with SGD [51]. Baselines are replicated using their original hyper-parameters, except when these were unspecified. More details and setups are deferred to Appendix Sec. 8." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 536, + 419, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 419, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 419, + 548 + ], + "type": "text", + "content": "4.2. Adaptation Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 555, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 579 + ], + "type": "text", + "content": "This section compares the generalizability of TEA with baselines in image corruption and domain generalization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Image Corruption As reported in Tab. 1, we conducted experiments on three benchmarks against eight baselines for corruption scenarios. TEA markedly surpasses all baselines on the vast majority of datasets and severity levels. Specifically, TEA outshines the best-performing baseline by an average of " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " at the most severe level. The only exception is on TinyImageNet across all levels, where TEA ranks second, merely trailing by a minimal margin " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ". We further incorporate the ResNet50 with GroupNorm for a broader validation. Tab. 2 indicates that TEA still delivers the best" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23905" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 113, + 545, + 253 + ], + "blocks": [ + { + "bbox": [ + 47, + 76, + 545, + 109 + ], + "lines": [ + { + "bbox": [ + 47, + 76, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 76, + 545, + 109 + ], + "type": "text", + "content": "Table 1. Comparisons of TEA and baselines for image corruption on CIFAR-10(C), CIFAR-100(C), and TinyImageNet(C) using WRN28-10 with BatchNorm. Accuracy and mCE are evaluated at the most severe level and across all levels with asterisk (*) indicating the results are taken from the original paper [56]. The best adaptation results are highlighted in boldface." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 113, + 545, + 253 + ], + "lines": [ + { + "bbox": [ + 49, + 113, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 49, + 113, + 545, + 253 + ], + "type": "table", + "html": "
WRN-28-10 BatchNormCIFAR-10(C)CIFAR-100(C)Tiny-ImageNet(C)
Clean Acc (↑)Corr Severity 5 Acc (↑)mCE (↓)Corr Severity 1-5 Acc (↑)mCE (↓)Clean Acc (↑)Corr Severity 5 Acc (↑)mCE (↓)Corr Severity 1-5 Acc (↑)mCE (↓)Clean Acc (↑)Corr Severity 5 Acc (↑)mCE (↓)Corr Severity 1-5 Acc (↑)mCE (↓)
Source94.7756.47100.0073.45100.0081.7935.39100.0052.12100.0063.1921.21100.0034.13100.00
NormBN [52]93.9779.5652.6585.6360.0080.8360.0663.5468.1169.4245.0427.7493.4234.27100.96
DUA* [41]-80.1050.78------------
PseudoPL [34]93.7551.42106.9872.6299.3780.5253.4072.1264.5375.2947.8428.2691.2239.8391.67
SHOT [36]93.2574.7763.1982.3572.6180.5256.5368.0166.0073.2847.9529.1490.1640.0191.41
EntropyTENT [60]93.6681.4148.1386.7556.1780.1463.0959.4269.4767.8039.5426.3195.5232.03104.49
ETA [45]93.9679.5852.6485.6359.9980.6559.8264.5267.1772.4043.2027.2894.1233.46102.25
EATA [45]93.9679.5952.6285.6459.9880.6860.2463.7567.4871.6643.4227.2894.0933.47102.24
SAR [46]93.9779.7751.9485.8358.9780.8462.9559.3770.0165.9941.5828.2192.8234.60100.47
EnergyTEA94.0983.3443.6987.8852.0080.8865.1056.0771.2263.7251.6531.6787.9939.9692.12
", + "image_path": "fcf32817d0e9e14fcb279c830c4af3fb683fe12ef780ae14de312d8ff47012ce.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 298, + 287, + 411 + ], + "blocks": [ + { + "bbox": [ + 47, + 261, + 287, + 294 + ], + "lines": [ + { + "bbox": [ + 47, + 261, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 47, + 261, + 287, + 294 + ], + "type": "text", + "content": "Table 2. Comparisons for image corruption on CIFAR-10(C), CIFAR-100(C), and TinyImageNet(C) using ResNet-50 with GroupNorm across all severity levels. Best results are in boldface." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 298, + 287, + 411 + ], + "lines": [ + { + "bbox": [ + 50, + 298, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 50, + 298, + 287, + 411 + ], + "type": "table", + "html": "
ResNet50 GroupNormCIFAR-10(C)CIFAR-100(C)Tiny-ImageNet(C)
Acc (↑)mCE (↓)Acc (↑)mCE (↓)Acc (↑)mCE (↓)
Source78.71100.0054.98100.0026.64100.00
PseudoPL79.4394.7656.6896.0226.6099.92
SHOT81.9886.6558.3193.4529.1196.73
EntropyTENT77.29102.8856.3496.8826.6599.94
ETA78.68100.0956.7296.3729.2596.42
EATA78.70100.0256.7696.2829.2596.42
SAR78.7899.6555.2899.3327.0599.41
EnergyTEA83.0579.0959.6789.3230.4194.81
", + "image_path": "8e762e1a917b032d640c8c6542395627b75b05462b3d317553a4534ad75fe229.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 418, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 287, + 467 + ], + "type": "text", + "content": "performance in both average accuracy and mCE. These results underscore the efficacy of TEA in handling image corruption scenarios, ensuring its universality for model architecture or normalization techniques." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "text", + "content": "Domain Generalization Tab. 3 provides a comparison between TEA and state-of-the-art TTA methods on the PACS dataset. It is evident that when trained on the photo and art domains, TEA exhibits substantial improvements over the best-performing baseline, achieving increases of " + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "inline_equation", + "content": "7.1\\%" + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "inline_equation", + "content": "8.5\\%" + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "text", + "content": ", respectively. Compared to photo and art, cartoon and sketch domains may exhibit greater domain discrepancies, posing significant challenges to model generalization. Despite these challenging conditions, in the adaptation from cartoon to art and the sketch domain, TEA achieved improvements when all baselines experienced significant declines. TENT shows decline of " + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "inline_equation", + "content": "14.73\\%" + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "text", + "content": " compared to the source in sketch, whereas TEA bucks the trend and enhances performance by " + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 46, + 485, + 287, + 664 + ], + "type": "text", + "content": ", highlighting its stability in the face of severe domain shifts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "content": "The results from both tasks highlight TEA's superior generalizability. These effectiveness may originate from the reduced energy, the enhanced distribution perception, and the improved calibration. In next section, we will delve into" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 298, + 545, + 578 + ], + "blocks": [ + { + "bbox": [ + 305, + 261, + 545, + 295 + ], + "lines": [ + { + "bbox": [ + 305, + 261, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 305, + 261, + 545, + 295 + ], + "type": "text", + "content": "Table 3. Single source domain generalization comparisons on PACS datasets using ResNet-18 with BatchNorm in terms of Accuracy. The best adaptation results are highlighted in boldface." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 298, + 545, + 578 + ], + "lines": [ + { + "bbox": [ + 309, + 298, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 309, + 298, + 545, + 578 + ], + "type": "table", + "html": "
Source DomainMethodTarget DomainAvg
PhotoArtCartoonSketch
PhotoSource-26.7622.4016.6221.93
BN-26.6627.9415.9623.52
TENT-26.9529.8617.5424.78
EATA-26.6628.1115.9823.59
SAR-26.7128.4115.9823.70
SHOT-26.6129.8620.9225.80
TEA-28.8133.6220.4927.64
ArtSource49.04-36.4324.4836.65
BN46.65-28.2822.7332.55
TENT50.78-30.1224.6135.17
EATA46.83-29.3123.4233.19
SAR47.90-33.0226.2735.73
SHOT50.24-34.3029.3737.97
TEA56.29-38.5728.7141.19
CartoonSource42.6929.79-29.4733.98
BN28.6825.15-20.8724.90
TENT30.9623.34-22.6525.65
EATA28.8025.10-25.0426.31
SAR29.7025.78-21.5125.66
SHOT37.7222.66-23.1427.84
TEA36.0531.44-22.8830.12
SketchSource19.9418.7032.21-23.62
BN13.4717.1429.86-20.16
TENT13.5317.3829.52-20.14
EATA13.1717.3330.08-20.19
SAR13.2918.8029.95-20.68
SHOT19.7618.7530.46-22.99
TEA19.6421.2433.19-24.69
", + "image_path": "1f4f1e8b1c0f8ca455ad4e262a4c191dc65bec1c24cdf28045e1f65a997e8f4e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 592, + 503, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 592, + 503, + 604 + ], + "spans": [ + { + "bbox": [ + 306, + 592, + 503, + 604 + ], + "type": "text", + "content": "these aspects for further analysis and discussion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 620, + 440, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 440, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 440, + 633 + ], + "type": "text", + "content": "4.3. Analysis and Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "In this section, we delve into the mechanisms driving TEA's effectiveness and explore its desirable properties. Specifically, we studied three key aspects of TEA: (1) The correlation between energy reduction and generalizability enhancement; (2) The distribution perception and generation capabilities; (3) The confidence calibration improvements." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23906" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 216, + 148 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 216, + 148 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 216, + 148 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 216, + 148 + ], + "type": "image", + "image_path": "4840f1e565db7c005955e6e04e2c736f2afd5422126a50a80b55b082c00d9e42.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 149, + 211, + 222 + ], + "blocks": [ + { + "bbox": [ + 48, + 149, + 211, + 222 + ], + "lines": [ + { + "bbox": [ + 48, + 149, + 211, + 222 + ], + "spans": [ + { + "bbox": [ + 48, + 149, + 211, + 222 + ], + "type": "image", + "image_path": "4a841b33e495cc40209b4b86ca44091ef0f3eaba949279b03f6d2e0b81cbc7d0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 225, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 225, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 547, + 270 + ], + "type": "text", + "content": "Figure 3. This illustration captures the energy reduction and generalizability enhancement achieved by TEA across CIFAR-10-C, CIFAR-100-C, and TinyImageNet-200-C, displayed from left to right. The upper set of graphs trace the evolution of energy score, corresponding loss and accuracy in response to incrementally increasing TEA adaptation steps. The lower set uncovers the extent of energy reduction and the consequent performance improvement before and after executing TEA adaptation, under different levels of distribution shift." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 217, + 70, + 380, + 148 + ], + "blocks": [ + { + "bbox": [ + 217, + 70, + 380, + 148 + ], + "lines": [ + { + "bbox": [ + 217, + 70, + 380, + 148 + ], + "spans": [ + { + "bbox": [ + 217, + 70, + 380, + 148 + ], + "type": "image", + "image_path": "151379108d2937a8d16b7cdca4910fb1524568e9392b9898ed5beb5b19bad047.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 215, + 149, + 373, + 222 + ], + "blocks": [ + { + "bbox": [ + 215, + 149, + 373, + 222 + ], + "lines": [ + { + "bbox": [ + 215, + 149, + 373, + 222 + ], + "spans": [ + { + "bbox": [ + 215, + 149, + 373, + 222 + ], + "type": "image", + "image_path": "7fa31240c1e245b963432bbcdaf3699bddb2ba78ffe8b16a01f2072dea6ce905.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 382, + 70, + 545, + 148 + ], + "blocks": [ + { + "bbox": [ + 382, + 70, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 382, + 70, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 382, + 70, + 545, + 148 + ], + "type": "image", + "image_path": "d458ea1f2acae09b7abaf870e1c7d3dc107ef5a171d1dd624f17025b904badb3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 380, + 149, + 539, + 222 + ], + "blocks": [ + { + "bbox": [ + 380, + 149, + 539, + 222 + ], + "lines": [ + { + "bbox": [ + 380, + 149, + 539, + 222 + ], + "spans": [ + { + "bbox": [ + 380, + 149, + 539, + 222 + ], + "type": "image", + "image_path": "353a1543a5ce03d4aed30294a098ca05c48b571caad9451008f29809157e9442.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 281, + 288, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 288, + 306 + ], + "type": "text", + "content": "4.3.1 Relation between TEA's Energy Reduction and Generalizability Enhancement" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 313, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 287, + 409 + ], + "type": "text", + "content": "This experiment validate TEA's energy reduction capability, and its impact on generalizability, spanning two scenarios: (1) In the first scenario, we focus on the trends of energy scores, loss, and accuracy on the same test data over increasing adaptation steps. (2) In the second scenario, we explore how the extent of energy reduction correlates with performance improvements, before and after adaptation, using varied test data that exhibit different distribution shifts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 410, + 288, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 288, + 602 + ], + "type": "text", + "content": "The results from both scenarios are depicted in Fig. 3. It is observable that " + }, + { + "bbox": [ + 46, + 410, + 288, + 602 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 46, + 410, + 288, + 602 + ], + "type": "text", + "content": " As the iteration step of TEA increases, there is a consistent reduction in energy and corresponding decrease in loss, coupled with an ongoing enhancement in accuracy. " + }, + { + "bbox": [ + 46, + 410, + 288, + 602 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 46, + 410, + 288, + 602 + ], + "type": "text", + "content": " As the distribution shift increases, TEA's energy reduction becomes more pronounced. Concurrently, the enhancement in performance over the baseline also increases. " + }, + { + "bbox": [ + 46, + 410, + 288, + 602 + ], + "type": "inline_equation", + "content": "(iii)" + }, + { + "bbox": [ + 46, + 410, + 288, + 602 + ], + "type": "text", + "content": " As the distribution shift increases, there is a sharp degradation observed in the baseline performance. However, the model adapted via TEA maintains its stability and robustness, demonstrating resilience against strong distribution shifts. In summary, these aforementioned trends are consistently observed across three datasets, demonstrating TEA's significant effectiveness in reducing energy. Notably, a greater reduction in energy correlates with an increased improvement in the model's generalizability" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 621, + 279, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 279, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 279, + 633 + ], + "type": "text", + "content": "4.3.2 TEA's Distribution Perception and Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "This experiment aims to validate the capability of TEA in modeling test data distribution. We framed the experiment within two scenarios: identical training and testing distributions, and shifted training and testing distributions. (1) The first scenario include four settings: (i) Source model on MNIST training set, TEA adaptation on MNIST test set;" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 279, + 545, + 339 + ], + "blocks": [ + { + "bbox": [ + 307, + 279, + 545, + 339 + ], + "lines": [ + { + "bbox": [ + 307, + 279, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 307, + 279, + 545, + 339 + ], + "type": "image", + "image_path": "9a8f47726808e381e9bf96ac9165d32f1b8489a9690feaaf5ddb0a046a99f8aa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 342, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 306, + 342, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 545, + 365 + ], + "type": "text", + "content": "Figure 4. Test distribution perception visualization for identical training and testing distributions on MNIST and CIFAR-10." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 370, + 545, + 490 + ], + "blocks": [ + { + "bbox": [ + 307, + 370, + 545, + 490 + ], + "lines": [ + { + "bbox": [ + 307, + 370, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 307, + 370, + 545, + 490 + ], + "type": "image", + "image_path": "c7e891cbc808f79805eac9d43182006756dc400b5b576c1338eee615640a4e11.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 493, + 547, + 537 + ], + "lines": [ + { + "bbox": [ + 305, + 493, + 547, + 537 + ], + "spans": [ + { + "bbox": [ + 305, + 493, + 547, + 537 + ], + "type": "text", + "content": "Figure 5. Test distribution perception visualization (upper) and real samples (lower) on shifted distribution: A model trained on PACS-A dataset then individually tested with TEA adaptation across PACS-P, PACS-A, PACS-C, PACS-S datasets." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 555, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 555, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 305, + 555, + 545, + 651 + ], + "type": "text", + "content": "(ii) Source model on MNIST training set, TEA adaptation on 90-degree rotated MNIST test set; (iii) Source model on CIFAR-10 training set, TEA adaptation on CIFAR-10 test set; (iv) Source models on MNIST or CIFAR-10 training sets, tested without TEA adaptation. (2) In the second scenario, source model was trained on the PACS-A dataset and individually tested with TEA adaptation on PACS-P, PACS-A, PACS-C and PACS-S datasets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "The outcomes from both scenarios are respectively illustrated in Fig. 4 and Fig. 5. Fig. 4 (1,3) indicates that, in scenarios where the distributions are closely identical, TEA has the potential to reconstruct samples that maintain discernible patterns. Fig. 4 (2) reveals that the sample distri" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23907" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "bution indeed originates from the test data rather than provoking recollections from the training datasets. Fig. 4 (4) confirms that models without adaptation, or those adapted using other methods, fail to effectively characterize the testing distributions on both the simple MNIST dataset (upper) and the more complex CIFAR-10 dataset (lower). Drawing conclusions from Fig. 5, we can infer that under significant distribution/domain shift, our method can still characterize the key features of the shifted test distribution, such as style, texture and color schemes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 289, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 289, + 277 + ], + "type": "text", + "content": "In summary, TEA endows the model with generative ability for test data via energy-based fine-tuning of the normalization layers. This ability may incorporates generative self-supervised information from the test data into the model and improves the model's thorough understanding of the test distribution, which in turn strengthens its generalization performance on that distribution." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 297, + 285, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 285, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 285, + 308 + ], + "type": "text", + "content": "4.3.3 TEA's Improvements in Confidence Calibration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 460 + ], + "type": "text", + "content": "This experiment compares model calibration across the source model, the entropy-based method TENT, the pseudolabeling-based method SHOT, and our energy-based TEA, on CIFAR-10. In accordance with the protocol in [4, 14], we illustrate reliability histogram and compute two scalar summary statistics: Expected Calibration Error (ECE) and Maximum Calibration Error (MCE) [14], to evaluate calibration implemented by torchmetrics. The procedures are implemented as follows: For the reliability histogram, we divided the model's predictions into ten bins based on the confidence score of the highest probability class and calculated the average accuracy for each bin." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 461, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 628 + ], + "type": "text", + "content": "The results are depicted in Fig. 6. From the perspective of the histogram, an optimally calibrated model should have its bar graph in a diagonal shape to achieve the smallest gap area. However, the bars for TENT and SHOT are observed to fall significantly below this line, manifesting an even inferior performance compared to the source model without any adaptation. These phenomena provide evidence that both the entropy-based methods and the pseudo-label-based method could potentially harm the confidence calibration by inducing overconfidence in their predictions. In stark contrast, our TEA has significantly narrowed the gap area and improved alignment with the diagonal line. For quantitative metrics, ECE and MCE, TEA has improved by " + }, + { + "bbox": [ + 46, + 461, + 287, + 628 + ], + "type": "inline_equation", + "content": "2.43\\%" + }, + { + "bbox": [ + 46, + 461, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 461, + 287, + 628 + ], + "type": "inline_equation", + "content": "18.31\\%" + }, + { + "bbox": [ + 46, + 461, + 287, + 628 + ], + "type": "text", + "content": " respectively compared to the source model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "The improvement in calibration of TEA over competitors may come from that neural networks inherently tend to be overconfident [14]. The softmax function enforces exponential normalization among classes, which tends to amplify the probability of dominant classes, thus inherently not advantageous for calibration. TENT and SHOT exacerbate this dominance of certain classes by reducing the un" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 422, + 182 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 422, + 182 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 422, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 422, + 182 + ], + "type": "image", + "image_path": "541afa8af5637d6d201a9f4c500afd276df4c7936f75e0c1bd7bc09ce2dca5e0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 422, + 70, + 545, + 182 + ], + "blocks": [ + { + "bbox": [ + 422, + 70, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 422, + 70, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 422, + 70, + 545, + 182 + ], + "type": "image", + "image_path": "8a8f66e1f1fa7729f3176fb5038b85b264417a8c639a08293184e284b22fbdd7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 182, + 422, + 283 + ], + "blocks": [ + { + "bbox": [ + 307, + 182, + 422, + 283 + ], + "lines": [ + { + "bbox": [ + 307, + 182, + 422, + 283 + ], + "spans": [ + { + "bbox": [ + 307, + 182, + 422, + 283 + ], + "type": "image", + "image_path": "b5a682aa1125634be43cacd5bf5e40321b592bcb79a8543d317b6d510b9b28e2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 287, + 547, + 342 + ], + "lines": [ + { + "bbox": [ + 304, + 287, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 547, + 342 + ], + "type": "text", + "content": "Figure 6. Calibration comparison between TEA and baselines on CIFAR-10 dataset. In an ideal scenario for optimal calibration, blue bars should align with the diagonal line, and a smaller grey gap area is preferred. Quantitative measures are provided via ECE and MCE metrics, where lower values indicate better calibration." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 422, + 182, + 545, + 283 + ], + "blocks": [ + { + "bbox": [ + 422, + 182, + 545, + 283 + ], + "lines": [ + { + "bbox": [ + 422, + 182, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 422, + 182, + 545, + 283 + ], + "type": "image", + "image_path": "77b50894f51975d703b78ce5ac1ec58aa9f32d217e78a851331486070e049d07.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 353, + 546, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 546, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 546, + 425 + ], + "type": "text", + "content": "certainty of class probabilities, further amplifying the overconfidence of the classifier. On the contrary, TEA dose not perform normalization in the label space, but maximizes the log-sum-exp of classifier logits, which essentially introduces a certain level of uncertainty to each class and empowers TEA with the ability to enhance calibration." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 435, + 379, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 435, + 379, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 435, + 379, + 448 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 455, + 547, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 547, + 635 + ], + "type": "text", + "content": "To achieve test time adaptation, we introduce an innovative energy-based perspective to mitigate the impact derived from distribution shifts. The proposed TEA aims to decrease the energy of the test data within the pretrained model's distribution. TEA guides the model towards achieving a harmonious low-energy equilibrium state for the test data, which mitigates the model's distribution discrepancy and boosts its generalizability towards test distributions. Comprehensive experiments across multiple tasks, benchmarks, and architectures confirm TEA's superiority over current leading methods. Further in-depth analyses of TEA's underlying mechanisms deepen our understanding of how energy reduction can enhance the model's perception of the test distribution, ultimately paving the way for improved generalization and calibration." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 646, + 403, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 403, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 403, + 659 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "This work was supported by the Strategic Priority Research Program of the CAS under Grants No. XDB0680302, and the National Key R&D Program of China (No. 2022YFB3103700, No. 2022YFB3103704)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23908" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Eric Arazo, Diego Ortego, Paul Albert, Noel E. O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 287, + 147 + ], + "type": "text", + "content": "[2] Stephen P Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 287, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 287, + 190 + ], + "type": "text", + "content": "[3] Miguel A Carreira-Perpinan and Geoffrey Hinton. On contrastive divergence learning. In International workshop on artificial intelligence and statistics, pages 33–40. PMLR, 2005. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 192, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 287, + 235 + ], + "type": "text", + "content": "[4] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 287, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 302 + ], + "type": "text", + "content": "[5] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robust-bench: a standardized adversarial robustness benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 303, + 287, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 287, + 356 + ], + "type": "text", + "content": "[6] Sabrina De Capitani Di Vimercati, Sara Foresti, Giovanni Livraga, and Pierangela Samarati. Data privacy: Definitions and techniques. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 20(06):793-817, 2012. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 358, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 390 + ], + "type": "text", + "content": "[7] Yilun Du and Igor Mordatch. Implicit generation and generalization in energy-based models. arXiv preprint arXiv:1903.08689, 2019. 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 392, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 287, + 445 + ], + "type": "text", + "content": "[8] Yilun Du, Shuang Li, Joshua Tenenbaum, and Igor Mordatch. Improved contrastive divergence training of energy-based models. In Proceedings of the 38th International Conference on Machine Learning, pages 2837-2848. PMLR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "type": "text", + "content": "[9] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. arXiv preprint arXiv:2103.10360, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "text", + "content": "[10] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, 2021. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 591 + ], + "type": "text", + "content": "[11] Jin Gao, Jialing Zhang, Xihui Liu, Trevor Darrell, Evan Shelhamer, and Dequan Wang. Back to the source: Diffusion-driven adaptation to test-time corruption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11786-11796, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "text", + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 690 + ], + "type": "text", + "content": "[13] Will Grathwohl, Kuan-Chieh Wang, Joern-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. In International Conference on Learning Representations, 2020. 3, 4, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "conference on machine learning, pages 1321-1330. PMLR, 2017. 2, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "type": "text", + "content": "[15] Tian Han, Erik Nijkamp, Xiaolin Fang, Mitch Hill, Song-Chun Zhu, and Ying Nian Wu. Divergence triangle for joint training of generator model, energy-based model, and inferential model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8670-8679, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "type": "text", + "content": "[17] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 5, 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "text", + "content": "[18] Geoffrey E Hinton. Training products of experts by minimizing contrastive divergence. Neural computation, 14(8): 1771-1800, 2002. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "text", + "content": "[19] Liang Hou, Qi Cao, Yige Yuan, Songtao Zhao, Chongyang Ma, Siyuan Pan, Pengfei Wan, Zhongyuan Wang, Huawei Shen, and Xueqi Cheng. Augmentation-aware self-supervision for data-efficient gan training. arXiv preprint arXiv:2205.15677, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "type": "text", + "content": "[20] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 398, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 398, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 398, + 545, + 442 + ], + "type": "text", + "content": "[21] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 444, + 545, + 487 + ], + "type": "text", + "content": "[22] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 488, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 545, + 521 + ], + "type": "text", + "content": "[23] Priyank Jain, Manasi Gyanchandani, and Nilay Khare. Big data privacy: a technological perspective and review. Journal of Big Data, 3:1-25, 2016. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 522, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 545, + 566 + ], + "type": "text", + "content": "[24] Minguk Jang, Sae-Young Chung, and Hye Won Chung. Test-time adaptation via self-training with nearest neighbor information. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 567, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 601 + ], + "type": "text", + "content": "[25] Jing Jiang. A literature survey on domain adaptation of statistical classifiers. URL: http://sifaka.cs.iuic.edu/jiang4/domainadaptation/survey, 3(1-12):3, 2008. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 601, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 633 + ], + "type": "text", + "content": "[26] Michael I Jordan and Tom M Mitchell. Machine learning: Trends, perspectives, and prospects. Science, 349(6245): 255-260, 2015. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 635, + 545, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 688 + ], + "type": "text", + "content": "[27] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[28] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23909" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, 2014. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 287, + 119 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 287, + 119 + ], + "type": "text", + "content": "[29] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 121, + 288, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 121, + 288, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 121, + 288, + 186 + ], + "type": "text", + "content": "[30] Jogendra Nath Kundu, Akshay R Kulkarni, Suvaansh Bhambri, Deepesh Mehta, Shreyas Anand Kulkarni, Varun Jampani, and Venkatesh Babu Radhakrishnan. Balancing discriminability and transferability for source-free domain adaptation. In International Conference on Machine Learning, pages 11710-11728. PMLR, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 188, + 287, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 188, + 287, + 210 + ], + "spans": [ + { + "bbox": [ + 48, + 188, + 287, + 210 + ], + "type": "text", + "content": "[31] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 5, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 212, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 212, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 48, + 212, + 287, + 245 + ], + "type": "text", + "content": "[32] Yann LeCun, Sumit Chopra, Raia Hadsell, M Ranzato, and Fujie Huang. A tutorial on energy-based learning. Predicting structured data, 1(0), 2006. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 247, + 287, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 247, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 48, + 247, + 287, + 269 + ], + "type": "text", + "content": "[33] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436-444, 2015. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 270, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 270, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 270, + 287, + 315 + ], + "type": "text", + "content": "[34] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 1, 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 316, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 361 + ], + "type": "text", + "content": "[35] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pages 5542-5550, 2017. 5, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 362, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 362, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 362, + 287, + 415 + ], + "type": "text", + "content": "[36] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 418, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 418, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 48, + 418, + 287, + 452 + ], + "type": "text", + "content": "[37] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. arXiv preprint arXiv:2303.15361, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 453, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 496 + ], + "type": "text", + "content": "[38] EM Lifshitz and Lev Davidovich Landau. Statistical physics, course of theoretical physics. In Part 2: Theory of the Condensed State. Butterworth-Heinemann Pergamon, London, 1980. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 499, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 565 + ], + "type": "text", + "content": "[39] Fu Lin, Rohit Mittapalli, Prithvijit Chattopadhyay, Daniel Bolya, and Judy Hoffman. Likelihood landscapes: A unifying principle behind many adversarial defenses. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 39-54. Springer, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "text", + "content": "[40] Weijian Luo, Hao Jiang, Tianyang Hu, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Training energy-based models with diffusion contrastive divergences. arXiv preprint arXiv:2307.01668, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "text", + "content": "[41] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[42] Weili Nie, Arash Vahdat, and Anima Anandkumar. Controllable and compositional generation with latent-space energy-based models. Advances in Neural Information Processing Systems, 34:13497-13510, 2021. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[43] Frank Nielsen and Ke Sun. Guaranteed bounds on information-theoretic measures of univariate mixtures using piecewise log-sum-exp inequalities. Entropy, 18(12):442, 2016. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[44] Erik Nijkamp, Mitch Hill, Song-Chun Zhu, and Ying Nian Wu. Learning non-convergent non-persistent short-run mcmc toward energy-based model. Advances in Neural Information Processing Systems, 32, 2019. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[45] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yafo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "text", + "content": "[46] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yafo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 276, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 374 + ], + "type": "text", + "content": "[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 376, + 545, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 440 + ], + "type": "text", + "content": "[48] Francesco Pinto, Harry Yang, Ser Nam Lim, Philip Torr, and Puneet Dokania. Using mixup as a regularizer can surprisingly improve accuracy & out-of-distribution robustness. In Advances in Neural Information Processing Systems, pages 14608-14622. Curran Associates, Inc., 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 442, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 442, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 442, + 545, + 475 + ], + "type": "text", + "content": "[49] Joaquin Quinonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset shift in machine learning. Mit Press, 2008. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "type": "text", + "content": "[50] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530-1538. PMLR, 2015. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "type": "text", + "content": "[51] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 544, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 599 + ], + "type": "text", + "content": "[52] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 601, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 643 + ], + "type": "text", + "content": "[53] Hidetoshi Shimodaira. Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of statistical planning and inference, 90(2):227-244, 2000. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 646, + 545, + 678 + ], + "type": "text", + "content": "[54] Yang Song and Diederik P Kingma. How to train your energy-based models. arXiv preprint arXiv:2101.03288, 2021. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[55] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23910" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "In Proceedings of the 37th International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 288, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 288, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 288, + 152 + ], + "type": "text", + "content": "[56] Yushun Tang, Ce Zhang, Heng Xu, Shuoshuo Chen, Jie Cheng, Luziwei Leng, Qinghai Guo, and Zhihai He. Neuromodulated hebbian learning for fully test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3728-3738, 2023. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 288, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 288, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 288, + 208 + ], + "type": "text", + "content": "[57] Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring robustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems, 33: 18583-18599, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 210, + 288, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 288, + 266 + ], + "type": "text", + "content": "[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 267, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 267, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 267, + 287, + 289 + ], + "type": "text", + "content": "[59] Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 1999. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 291, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 291, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 291, + 288, + 335 + ], + "type": "text", + "content": "[60] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 337, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 288, + 392 + ], + "type": "text", + "content": "[61] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. Generalizing to unseen domains: A survey on domain generalization. IEEE Transactions on Knowledge and Data Engineering, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 394, + 288, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 288, + 438 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 288, + 438 + ], + "type": "text", + "content": "[62] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688, 2011. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 440, + 288, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 440, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 288, + 495 + ], + "type": "text", + "content": "[63] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-mond, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 497, + 288, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 497, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 497, + 288, + 529 + ], + "type": "text", + "content": "[64] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 532, + 288, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 288, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 288, + 576 + ], + "type": "text", + "content": "[65] Zehao Xiao, Xiantong Zhen, Shengcai Liao, and Cees G. M. Snoek. Energy-based test sample adaptation for domain generalization. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 288, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 620 + ], + "type": "text", + "content": "[66] Xiulong Yang and Shihao Ji. Jem++: Improved techniques for training jem. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6494-6503, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 623, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 288, + 667 + ], + "type": "text", + "content": "[67] Yige Yuan, Bingbing Xu, Huawei Shen, Qi Cao, Keting Cen, Wen Zheng, and Xueqi Cheng. Towards generalizable graph contrastive learning: An information theory perspective. arXiv preprint arXiv:2211.10929, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 714 + ], + "type": "text", + "content": "[68] Yige Yuan, Bingbing Xu, Bo Lin, Liang Hou, Fei Sun, Huawei Shen, and Xueqi Cheng. Pde+: Enhancing generalization via pde with adaptive distributional diffusion. arXiv preprint arXiv:2305.15835, 2023. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 150 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "type": "text", + "content": "[69] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In British Machine Vision Conference 2016. British Machine Vision Association, 2016. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "text", + "content": "[70] Hao Zhao, Yuejiang Liu, Alexandre Alahi, and Tao Lin. On pitfalls of test-time adaptation. In ICLR 2023 Workshop on Pitfalls of limited data and computation for Trustworthy ML, 2023. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23911" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/a91d56f7-042a-42e5-9c77-ba0dbaac2109_content_list.json b/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/a91d56f7-042a-42e5-9c77-ba0dbaac2109_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8f7598179af78a4aa881185903ccaa3c516c8730 --- /dev/null +++ b/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/a91d56f7-042a-42e5-9c77-ba0dbaac2109_content_list.json @@ -0,0 +1,1508 @@ +[ + { + "type": "text", + "text": "TFMQ-DM: Temporal Feature Maintenance Quantization for Diffusion Models", + "text_level": 1, + "bbox": [ + 84, + 130, + 887, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yushi Huang $^{1,2*}$ , Ruihao Gong $^{1,2*}$ , Jing Liu $^{2,3}$ , Tianlong Chen $^{4}$ , Xianglong Liu $^{1*}$ $^{1}$ Beihang University $^{2}$ SenseTime Research $^{3}$ Monash University $^{4}$ UT Austin", + "bbox": [ + 176, + 179, + 802, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{huangyushi, gongruihao, liujing5} @sentetime.com, tianlong.chen@utexas.edu, xlliu@buaa.edu.cn", + "bbox": [ + 132, + 220, + 831, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Diffusion model, a prevalent framework for image generation, encounters significant challenges in terms of broad applicability due to its extended inference times and substantial memory requirements. Efficient Post-training Quantization (PTQ) is pivotal for addressing these issues in traditional models. Different from traditional models, diffusion models heavily depend on the time-step $t$ to achieve satisfactory multi-round denoising. Usually, $t$ from the finite set $\\{1,\\dots,T\\}$ is encoded to a temporal feature by a few modules totally irrespective of the sampling data. However, existing PTQ methods do not optimize these modules separately. They adopt inappropriate reconstruction targets and complex calibration methods, resulting in a severe disturbance of the temporal feature and denoising trajectory, as well as a low compression efficiency. To solve these, we propose a Temporal Feature Maintenance Quantization (TFMQ) framework building upon a Temporal Information Block which is just related to the time-step $t$ and unrelated to the sampling data. Powered by the pioneering block design, we devise temporal information aware reconstruction (TIAR) and finite set calibration (FSC) to align the full-precision temporal features in a limited time. Equipped with the framework, we can maintain the most temporal information and ensure the end-to-end generation quality. Extensive experiments on various datasets and diffusion models prove our state-of-the-art results. Remarkably, our quantization approach, for the first time, achieves model performance nearly on par with the full-precision model under 4-bit weight quantization. Additionally, our method incurs almost no extra computational cost and accelerates quantization time by $2.0\\times$ on LSUN-Bedrooms $256\\times 256$ compared to previous works. Our code is publicly available at https://github.com/ModelTC/TFMQ-DM.", + "bbox": [ + 75, + 308, + 472, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 268, + 630, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative modeling plays a crucial role in machine learning, particularly in applications like image [13, 14, 17, 38, 44], voice [37, 42], and text synthesis [2, 51]. Diffusion models have showcased impressive capabilities in producing high-quality samples across diverse domains. In comparison to generative adversarial networks (GANs) [8] and variational autoencoders (VAEs) [20], diffusion models successfully sidestep issues such as model collapse and posterior collapse, resulting in a more stable training process. However, the substantial computational cost poses a critical bottleneck hampering the widespread adoption of diffusion models. Furthermore, the computational cost for diffusion models can be attributed to two primary factors. First, these models typically require hundreds of denoising steps to generate images, rendering the procedure considerably slower than that of GANs. Prior efforts [21, 27, 29, 44] have addressed this challenge by seeking shorter and more efficient sampling trajectories, thereby reducing the number of denoising steps. Second, the substantial network architecture of diffusion models demands considerable time and memory resources, particularly for foundational models pre-trained on large-scale datasets, e.g., LDM [38] and Stable Diffusion. Our work aims to tackle the latter challenge, focusing on the compression of diffusion models.", + "bbox": [ + 496, + 295, + 893, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Quantization is currently the most widely used method [1, 5, 7, 31, 32] for compressing models by mapping high-precision floating-point numbers to low-precision numbers. Among different quantization methods, Posttraining quantization (PTQ) [15, 31, 47] incurs lower overhead and is more user-friendly without the need for retraining or fine-tuning. While PTQ on conventional models has undergone extensive study [7, 24, 31, 47], its application to diffusion models has shown huge performance degradation, especially under low-bit settings. For instance, Q-Diffusion [23] exhibits severe accuracy drop on some datasets [48] under 4-bit quantization. We believe the reason they fail to achieve better results is that they all overlook the sampling data independence and uniqueness of temporal features, which are generated from time-step $t$ through a few modules, used to control the denoising trajectory in dif", + "bbox": [ + 496, + 659, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution.", + "bbox": [ + 93, + 862, + 209, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Work done during an internship at SenseTime Research.", + "bbox": [ + 98, + 875, + 397, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Corresponding Author.", + "bbox": [ + 98, + 887, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7362", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fusion models. Furthermore, we observe that temporal feature disturbance significantly impacts model performance in the aforementioned methods.", + "bbox": [ + 76, + 90, + 468, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle temporal feature disturbance, we first find that the modules generating temporal features are independent of the sampling data and define the whole modules as the Temporal Information Block. All existing methods do not separately optimize this block during the quantization process, causing temporal features to overfit to limited calibration data. On the other hand, since the maximum time-step for denoising is a finite positive integer, the temporal feature and the activations during its generation form a finite set. The optimal approach is to optimize each element in this set individually. Based on these observations and analyses, we propose a novel quantization reconstruction approach, temporal information aware reconstruction (TIAR), specifically optimizing for temporal features. It aims to reduce temporal feature loss as the optimization objective while isolating the network's components related to sampled data from the generation of temporal features during calibration. Furthermore, we also introduce a calibration strategy, finite set calibration (FSC), for the finite set of temporal features and activations during its generation. This strategy employs different quantization parameters for activations corresponding to different time-steps. Moreover, the calibration speed of this method is faster than existing mainstream methods [1, 5], for example, we speedup quantization time by $2.0 \\times$ on LSUN-Bedrooms $256 \\times 256$ dataset, yet the strategy incurs negligible inference and storage overhead. The overview of our framework can be seen in Fig. 1. In summary, our contributions are as follows:", + "bbox": [ + 76, + 147, + 472, + 570 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We discover that existing quantization methods suffer from temporal feature disturbance, disrupting the denoising trajectory of diffusion models and significantly affecting the quality of generated images.", + "- We reveal that the disturbance comes from two aspects: inappropriate reconstruction target and unaware of finite activations. Both inducements ignore the special characteristics of time information-related modules.", + "- An advanced framework (TFMQ-DM) is proposed, consisting of temporal information aware reconstruction (TIAR) for weight quantization and finite set calibration (FSC) for activation quantization. Both are based on a Temporal Information Block specially devised for diffusion models.", + "- Extensive experiments on various datasets show that our novel framework achieves a new state-of-the-art result in PTQ of diffusion models, especially under 4-bit weight quantization, and significantly accelerates quantization time. For some hard tasks, e.g., CelebA-HQ $256 \\times 256$ , our method reduces the FID score by 6.71 (images in appendix)." + ], + "bbox": [ + 76, + 580, + 470, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 642, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Efficient Diffusion Models", + "text_level": 1, + "bbox": [ + 500, + 114, + 736, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are diverse perspectives to accelerate the inference of diffusion model, e.g., retraining-based model design [3, 6, 30, 52] and retraining-free sampler strategy [21, 44, 50]. However, the retraining-based method proves resource-intensive and time-consuming. The efficient samplers can reduce sampling iterations but still suffer from diffusion models' extensive parameters and computational complexity. In this paper, we focus on diminishing the time and memory overhead of the single-step denoising process using low-bit quantization in a training-free manner, a method orthogonal to previous speedup techniques.", + "bbox": [ + 496, + 137, + 893, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Model Quantization", + "text_level": 1, + "bbox": [ + 500, + 311, + 691, + 327 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Quantization is a predominant technique for minimizing storage and computational costs. It can be categorized into quantization-aware training (QAT) [7, 16, 28, 49, 53] and post-training quantization (PTQ) [15, 24, 26, 31, 47]. QAT requires intensive model training with substantial data and computational demands. Correspondingly, PTQ compresses models without re-training, making it a preferred method due to its minimal data requirements and easy deployment on real hardware. In PTQ, high-precision values are mapped into discrete levels using uniform quantization expressed as:", + "bbox": [ + 496, + 334, + 893, + 501 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} = \\Phi \\left(\\left\\lfloor \\frac {x}{s} \\right\\rceil + z, 0, 2 ^ {b} - 1\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 506, + 890, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $x$ represents floating-point data, $\\hat{x}$ is the quantized value, $s$ is the quantization step size, $z$ is the zero offset, and $b$ is the target bit-width. The clamp function $\\Phi(\\cdot)$ clips the rounded value $\\left\\lfloor \\frac{x}{s} \\right\\rfloor + z$ within the range of $[0, 2^b - 1]$ . However, naive quantization may lead to accuracy degradation, especially for low-bit quantization. Recent studies [15, 24, 31, 46, 47] have explored innovative strategies based on reconstruction to preserve model performance after low-bit quantization.", + "bbox": [ + 496, + 537, + 893, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, the iterative denoising process in diffusion models presents new challenges for PTQ in comparison to traditional models. PTQ4DM [41] represents the initial attempt to quantize diffusion models to 8-bit, albeit with limited experiments and lower resolutions. Conversely, Q-Diffusion [23] achieves enhanced performance and is evaluated on a broader dataset range. Moreover, PTQD [10] eliminates quantization noise through correlated and residual noise correction. Notably, traditional single-time-step PTQ calibration methods are unsuitable for diffusion models due to significant activation distribution changes with each time-step [23, 41, 43, 45]. ADP-DM [45] proposes group-wise quantization across time-steps for diffusion models, and TDQ [43] introduces distinct quantization parameters for different time-steps. However, all of the", + "bbox": [ + 496, + 674, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7363", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "above works overlook the specificity of temporal features. To address temporal feature disturbance in the aforementioned works, our study delves into the inducements of the phenomenon and introduces a novel reconstruction and calibration framework, significantly enhancing quantized diffusion model performance.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 194, + 215, + 209 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models. Diffusion models [14, 44] iteratively add Gaussian noise with a variance schedule $\\beta_{1},\\ldots ,\\beta_{T}\\in$ $(0,1)$ to data $\\mathbf{x}_0\\sim q(\\mathbf{x})$ for $T$ times as sampling process, resulting in a sequence of noisy samples $\\mathbf{x}_1,\\dots ,\\mathbf{x}_T$ In DDPMs [14], the former sampling process is a Markov chain, taking the form:", + "bbox": [ + 75, + 220, + 468, + 310 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 320, + 468, + 361 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1}) = \\mathcal {N} (\\mathbf {x} _ {t}; \\sqrt {\\alpha_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}),\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 364, + 395, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t} = 1 - \\beta_{t}$ . Conversely, the denoising process removes noise from a sample from Gaussian noise $\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ to gradually generate high-fidelity images. Nevertheless, due to the unavailability of the true reverse conditional distribution $q(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)$ , diffusion models approximate it via variational inference by learning a Gaussian distribution $p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t) = \\mathcal{N}(\\mathbf{x}_{t - 1};\\boldsymbol {\\mu}_\\theta (\\mathbf{x}_t,t),\\boldsymbol {\\Sigma}_\\theta (\\mathbf{x}_t,t))$ , the $\\boldsymbol{\\mu}_{\\theta}$ can be derived by reparameterization trick as follows:", + "bbox": [ + 75, + 391, + 468, + 525 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 534, + 468, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_{t} = \\prod_{i=1}^{t} \\alpha_{i}$ and $\\epsilon_{\\theta}(\\cdot)$ is a noise estimation model. The variance $\\Sigma_{\\theta}(\\mathbf{x}_{t}, t)$ can be either learned [33] or fixed to a constant schedule [14] $\\sigma_{t}$ . When employing the latter method, $\\mathbf{x}_{t-1}$ can be expressed as:", + "bbox": [ + 75, + 579, + 468, + 640 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {x} _ {t}, t)\\right) + \\sigma_ {t} \\mathbf {z}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 650, + 468, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{z}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$", + "bbox": [ + 76, + 693, + 210, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reconstruction on diffusion models. UNet [39], the predominant model employed as $\\epsilon_{\\theta}(\\cdot)$ in Eq. 4 to predict Gaussian noise, can be divided into blocks that incorporate residual connections (such as Residual Bottleneck Blocks or Transformer Blocks [35]) and the remaining layers. Numerous PTQ approaches for diffusion models are grounded in layer/block-wise reconstruction [10, 23, 41, 43] to obtain optimal quantization parameters. For example, in the Residual Bottleneck Block, this approach typically minimizes the following loss function as its optimization objective:", + "bbox": [ + 75, + 720, + 468, + 872 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i} = \\left\\| f _ {i} (\\cdot) - \\widehat {f} _ {i} (\\cdot) \\right\\| _ {F} ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 882, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|_F^2$ denotes the Frobenius norm. The function $f_i(\\cdot)$ represents the $i^{\\text{th}}$ Residual Bottleneck Block, and $\\widehat{f}_i(\\cdot)$ is its quantized counterpart. Furthermore, in the ensuing sections, we use $n$ to denote the total number of Residual Bottleneck Blocks in a single diffusion model.", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal feature in diffusion models. Seeing from Fig. 1 (a), time-step $t$ is encoded with time embed $^{1}$ and then passes through the embedding layer $^{2}$ in each Residual Bottleneck Block, resulting in a series of unique activations. In this paper, we denote these activations as temporal features. Notably, temporal features are independent of $\\mathbf{x}_t$ and unrelated to other temporal features from different time-steps. To enhance clarity, we simplify our notation as follows: we represent time embed as $h(\\cdot)$ , the embedding layer in the $i^{\\text{th}}$ Residual Bottleneck Block as $g_i(\\cdot)$ , and denote the $i^{\\text{th}}$ temporal feature at time-step $t$ as $\\mathbf{emb}_{t,i}$ . Moreover, as illustrated in Fig. 1 (a), the relationship is explicitly expressed by the equation:", + "bbox": [ + 498, + 181, + 890, + 377 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {T e m p o r a l} \\quad \\mathbf {e m b} _ {t, i} = g _ {i} (h (t)) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 388, + 890, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, we have found that temporal features play a crucial role in the context of the diffusion model, holding unique and substantial physical implications. These features encompass temporal information that signifies the current image's temporal position along the denoising trajectory. Within the UNet structure, each time-step transforms into these temporal features, thereby controlling the denoising trajectory by applying them to the features of images generated at each iteration.", + "bbox": [ + 498, + 417, + 890, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. TFMQ for Diffusion Models", + "text_level": 1, + "bbox": [ + 500, + 566, + 761, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present our novel PTQ framework in this section. We first observe temporal disturbance in previous methods in Sec. 4.1 and then analyze the inducements in Sec. 4.2. Finally, we propose our solutions in Sec. 4.3.", + "bbox": [ + 498, + 592, + 890, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Temporal Feature Disturbance", + "text_level": 1, + "bbox": [ + 500, + 662, + 772, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on Sec. 3, we investigate the impact of previous PTQ works on temporal features, and we identify the phenomenon of temporal feature disturbance, which significantly deteriorates the quality of generated images.", + "bbox": [ + 498, + 685, + 890, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal feature error. We thoroughly analyze temporal feature variations before and after the quantization of embedding layers and time embed in the Stable Diffusion model ( $T = 50, i = 11$ ). Prior to this analysis, we introduce the temporal feature error as defined by:", + "bbox": [ + 498, + 758, + 890, + 834 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\cos \\left(\\widehat {\\mathbf {e m b}} _ {t, i}, \\widehat {\\mathbf {e m b}} _ {t, i}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 844, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "PyTorch time embed implementation in diffusion models.", + "bbox": [ + 517, + 875, + 828, + 887 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "PyTorch embedding layer implementation in diffusion models.", + "bbox": [ + 517, + 887, + 852, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7364", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/87b6455d1bf6e3904f1661eca93aa8229ab12114fb2c24f4ad02d0f7cadc3d67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 95, + 465, + 204 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/01437bc9a586d1dad1f66ae44a29260ebd857443fb7eb429f28491d61385d568.jpg", + "image_caption": [ + "Figure 1. Overview of the proposed Temporal Feature Maintenance Quantization. (a) Temporal Feature $\\mathbf{emb}_{t,i}$ , belonging to a finite set representing temporal information, has been overlooked in previous works due to inappropriate reconstruction targets (box with a solid line). (b) This oversight leads to a severe disturbance for $\\mathbf{emb}_{t,i}$ and results in the mismatch of crucial temporal information for the diffusion model's generation, causing a deviation in the denoising trajectory and a significant drop in accuracy. (c) Based on these analyses, we introduce a Temporal Information Block that exclusively correlates with the time-step $t$ . Leveraging this $\\mathbf{x}_t$ -unrelated block, we enable Temporal Information Aware Reconstruction and Finite Set Calibration (utilizing the finite number of $t$ ). This approach achieves the maintenance of temporal features and yields state-of-the-art results." + ], + "image_footnote": [], + "bbox": [ + 88, + 207, + 465, + 262 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/afc432d837c9c40687168e5dcb02f97600585a64ffcb5e6266585921821bfed2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 89, + 883, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\cos (\\cdot)$ denotes cosine similarity, and $\\widehat{\\mathbf{emb}_{t,i}}$ signifies the temporal feature corresponding to $\\mathbf{emb}_{t,i}$ in the quantized model. As illustrated in Fig. 2 (Left), quantization induces notable temporal feature errors. We term this phenomenon, characterized by substantial temporal feature errors within diffusion models, as temporal feature disturbance.", + "bbox": [ + 75, + 385, + 468, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal information mismatch. Temporal feature disturbance alters the original embedded temporal information. Specifically, $\\mathbf{emb}_{t,i}$ is intended to correspond to timestep $t$ . However, due to significant errors, the quantized model's $\\widehat{\\mathbf{emb}}_{t,i}$ is no longer accurately associated with $t$ , resulting in what we term as temporal information mismatch:", + "bbox": [ + 75, + 506, + 468, + 611 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt \\leftarrow \\mathbf {e m b} _ {t, i}, \\quad t \\nleftarrow \\widehat {\\mathbf {e m b}} _ {t, i}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 611, + 468, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, as depicted in Fig. 2 (Right), we even observe a pronounced temporal information mismatch. Specifically, the temporal feature generated by the quantized model at time-step $t$ exhibits a divergence from that of the full-precision model at the corresponding time-step. Instead, it tends to align more closely with the temporal feature corresponding to $t + \\delta_t$ , importing wrong temporal information from $t + \\delta_t$ .", + "bbox": [ + 75, + 633, + 468, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Trajectory deviation. Temporal information mismatch delivers wrong temporal information, therefore, causing a deviation in the corresponding temporal position of the image within the denoising trajectory, ultimately leading to:", + "bbox": [ + 76, + 766, + 468, + 827 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t} \\neq \\mathbf {x} _ {t - 1}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 834, + 468, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we apply disrupted temporal features to the model. Evidently, as the deviation in the denoising trajectory accumulates with the increase in the number of denoising", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9a683213a2eeebc8b675f563ced683121f11e6445e523c26ff4265a00d6f6dd2.jpg", + "image_caption": [ + "Figure 2. (Left) Temporal feature disturbance. The inflection points serve as indicators of temporal feature errors at different time-steps, and they highlight the significant phenomenon of temporal feature disturbance. (Right) Temporal information mismatch. The coordinates of the inflection points on the blue curve can be denoted as $(t,t + \\delta_{t,i})$ . It indicates $\\mathbf{emb}_{t + \\delta_{t,i},i}$ exhibits the highest similarity with $\\mathbf{emb}_{t,i}$ ." + ], + "image_footnote": [], + "bbox": [ + 501, + 385, + 694, + 526 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/677aa6975168f144167cfeb2bc4f955dd11a8e1777b0b6c487ff1cd255bc00ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 385, + 888, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "iterations, the final generated image struggles to align with $\\mathbf{x}_0$ . This evolution is illustrated in Fig. 3, where we maintain UNet excluding embedding layers and time embed in full precision.", + "bbox": [ + 496, + 654, + 890, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Inducement Analyses", + "text_level": 1, + "bbox": [ + 500, + 727, + 699, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we explore the two inducements of temporal feature disturbance. For the purpose of clarity, in the subsequent sections, \"reconstruction\" specifically points to slight weight adjustment for minimal quantization error, while \"calibration\" specifically refers to activation calibration.", + "bbox": [ + 496, + 751, + 890, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inappropriate reconstruction target. Previous PTQ works [10, 23, 43] have achieved remarkable progress on diffusion models. However, these existing methods overlook the temporal feature's independence and its dis", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7365", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1304d60b10183617411e1fa61cc9de06863ae7c71a1307ba4a64fe68af16fa04.jpg", + "image_caption": [ + "Figure 3. Denoising process of full-precision (Upper) and w4a8 quantized (Lower) Stable-Diffusion $(T = 50)$ under the same experiment settings and prompt: A man in the snow on a snow board. We represent $\\{\\mathbf{emb}_{t,i}\\}_{i=0,\\dots,n}$ and $\\{\\widehat{\\mathbf{emb}}_{t,i}\\}_{i=0,\\dots,n}$ as $\\mathbf{EMB}_t$ and $\\widehat{\\mathbf{EMB}_t}$ , respectively. Additionally, we denote $\\widehat{\\mathbf{x}}_t$ as $\\mathbf{x}_t$ in the context of the quantized diffusion model. It is noteworthy that, in the quantized model employed here, to showcase the impact of temporal features, only the layers in Temporal Information Block are quantized and the components unrelated to the generation of temporal features are maintained in full precision." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 890, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tive physical significance. During their reconstruction processes, there was a lack of optimization for the embedding layers; instead, a Residual Bottleneck Block of coarser granularity was selected as the reconstruction target. This method involves two potential factors causing temporal feature disturbance:", + "bbox": [ + 75, + 353, + 468, + 443 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Optimize the objective as expressed in Eq. 5 to decrease the reconstruction loss of Residual Bottleneck Block, as opposed to direct reduction of temporal feature loss.", + "- During backpropagation of the reconstruction process, embedding layers independent from $\\mathbf{x}_t$ are affected by $\\mathbf{x}_t$ , resulting in an overfitting scenario on limited calibration data." + ], + "bbox": [ + 76, + 446, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further validate our analyses, we respectively evaluate the FID [12] and sFID [40] for this reconstruction method, e.g., BRECQ [24] and the approach where we freeze the parameters of the embedding layers during the reconstruction of the Residual Bottleneck Block, initializing the embedding layers solely through Min-max [32] for comparison. As shown in Tab. 1, the Freeze strategy exhibits better results, which verify that embedding layers serve as their own optimization objective and maintain their independence of $\\mathbf{x}_t$ can significance mitigate temporal feature disturbance, especially at low-bit.", + "bbox": [ + 75, + 553, + 470, + 733 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/aac2fe3feffdc825fa50a35d1aa4dd34b59ee4b19f0f8213fffd2bd50bae272b.jpg", + "table_caption": [ + "Table 1. FID and sFID on LSUN-Bedrooms $256 \\times 256$ [48] for LDM-4 with 50000 sampling images. Prev represents BRECQ. Freeze denotes our trial." + ], + "table_footnote": [], + "table_body": "
MethodsBits (W/A)FID↓sFID↓
Full Prec.32/322.987.09
Prev8/87.5112.54
Freeze8/85.76 (-1.75)8.42 (-4.12)
Prev4/89.3622.73
Freeze4/87.08 (-2.28)16.82 (-5.91)
", + "bbox": [ + 98, + 801, + 444, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unaware of finite activations within $h(\\cdot)$ and $g_{i}(\\cdot)$ . We observe that, given $T$ as a finite positive integer, the set of all possible activation values for embedding layers and time embed is finite and strictly dependent on timesteps. Within this set, activations corresponding to the same layer display notable range variations across different timesteps (refer to the appendix). Previous methods [43, 45] mainly focus on finding the optimal calibration method for $\\widehat{\\mathbf{x}}_t$ -related network components. Moreover, akin to the first inducement, their calibration is directly towards the Residual Bottleneck Block, which proves suboptimal (refer to the appendix). However, based on the finite activations, we can employ calibration methods, especially for these time information-related activations, to better adapt to their range variations.", + "bbox": [ + 496, + 353, + 890, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Temporal Feature Maintenance", + "text_level": 1, + "bbox": [ + 498, + 590, + 779, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the problem of temporal feature disturbance, we design a novel Temporal Information Block to maintain the temporal features. Based on the block, Temporal Information Aware Reconstruction and Finite Set Calibration are proposed to solve the two inducements analyzed above.", + "bbox": [ + 496, + 616, + 890, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal Information Block. Based on the inducements, it is crucial to meticulously separate the reconstruction and calibration process for each embedding layer and Residual Bottleneck Block to enhance quantized model performance. Considering the unique structure of the UNet, we consolidate all embedding layers and time embed into a unified Temporal Information Block, which can be denoted as $\\{g_{i}(h(\\cdot))\\}_{i = 0,\\dots,n}$ (see Fig. 1 (c)).", + "bbox": [ + 496, + 705, + 890, + 828 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal information aware reconstruction. Based on the Temporal Information Block, we propose Temporal information aware reconstruction (TIAR) to tackle the first inducement. The optimization objective for the block during", + "bbox": [ + 496, + 839, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7366", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the reconstruction process is as follows:", + "bbox": [ + 76, + 90, + 341, + 104 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T I A R} = \\sum_ {i = 0} ^ {n} \\| g _ {i} (h (t)) - \\widehat {g} _ {i} (\\widehat {h} (t)) \\| _ {F} ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 119, + 468, + 159 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\widehat{h} (\\cdot)$ and $\\widehat{g}_i(\\cdot)$ are quantized versions of $h(\\cdot)$ and $g(\\cdot)$ , respectively. With this reconstruction, weights are adjusted to pursue a minimal disturbance for temporal features.", + "bbox": [ + 76, + 172, + 468, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finite set calibration. To address the challenge posed by the wide span of activations within a finite set for the second inducement, we propose Finite Set Calibration (FSC) for activation quantization. This strategy employs $T$ sets of quantization parameters for every activation within all the embedding layers and time embed, such as $\\{(s_{T},z_{T}),\\ldots ,(s_{1},z_{1})\\}$ for activation $\\mathbf{x}$ . In time-step $t$ , the quantization function for the $\\mathbf{x}$ can be expressed as:", + "bbox": [ + 76, + 233, + 468, + 354 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} = \\Phi \\left(\\left\\lfloor \\frac {\\mathbf {x}}{s _ {t}} \\right\\rceil + z _ {t}, 0, 2 ^ {b} - 1\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 367, + 468, + 401 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To be noted, the calibration target is also aligned with the output of the Temporal Information Block. For a more specific estimation method of activation ranges, since the solution space within the finite set is limited, we find the Minmax [32] method can achieve satisfactory results with high efficiency (more evidence in Sec. 5.3).", + "bbox": [ + 75, + 412, + 468, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 518, + 209, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 545, + 290, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models and datasets. In this section, we conduct image generation experiments to evaluate the proposed TFMQ-DM framework on various diffusion models: pixel-space diffusion model DDPM [14] for unconditional image generation, latent-space diffusion model LDM [38] for unconditional image generation and class-conditional image generation. We also apply our work to Stable Diffusion-v1-4 for text-guided image generation. In our experiments, We use seven standard benchmarks: CIFAR-10 $32 \\times 32$ [22], LSUN-Bedrooms $256 \\times 256$ [48], LSUN-Churches $256 \\times 256$ [48], CelebA-HQ $256 \\times 256$ [18], ImageNet $256 \\times 256$ [4], FFHQ $256 \\times 256$ [19] and MS-COCO [25].", + "bbox": [ + 75, + 569, + 468, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantization settings. We use channel-wise quantization for weights and layer-wise quantization for activations, as it is a common practice. In our experimental setup, we employ BRECQ [24] and AdaRound [31]. Drawing from empirical insights derived from conventional model quantization practices [9, 36], we maintain the input and output layers of the model in full precision. Calibration sets, integral to our methodology, are generated through full-precision", + "bbox": [ + 75, + 779, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "diffusion models, mirroring the approach outlined in Q-Diffusion [23]. Moreover, in weight quantization, we reconstruct quantized weights for 20k iterations with a minibatch size of 32 for DDPM and LDM, and 8 for Stable Diffusion. In activation quantization, we utilize EMA [16] to estimate the ranges of activations with a mini-batch size of 16 on all models. More details can be found in the appendix.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. For each experiment, we evaluate the performance of diffusion models with Fréchet Inception Distance (FID) [12]. In the case of LDM and Stable Diffusion experiments, we also include sFID [40], which better captures spatial relationships than FID. For ImageNet and CIFAR-10 experiments, we additionally provide Inception Score (IS) [40] as a reference metric. Further, in the context of Stable Diffusion experiments, we extend our evaluation to include the compatibility of image-caption pairs, employing the CLIP score [11]. The ViT-B/32 is used as the backbone when computing the CLIP score. To ensure consistency in the reported outcomes, all results are derived from our implementation or from other papers, where experiments are conducted under conditions consistent with ours. More specifically, in the evaluation process of each experiment, we sample 50k images from DDPM or LDM, or 30k images from Stable-Diffusion. All experiments are conducted utilizing one H800 GPU and implemented with the PyTorch framework [34].", + "bbox": [ + 496, + 210, + 892, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Main Results", + "text_level": 1, + "bbox": [ + 500, + 512, + 638, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unconditional image generation. In the experiments conducted on the LDM, we maintain the same experimental settings as presented in [38], including the number of steps, variance schedule, and classifier-free guidance scale (denoted by eta and CFG in the following, respectively). As shown in Tab. 2, the FID performance differences relative to the full precision (FP) model are all within 0.7 for all settings. Specifically, on the CelebA-HQ $256 \\times 256$ dataset, our method exhibits a FID reduction of 6.71 and a sFID reduction of 6.60 in the w4a8 setting compared to the current state-of-the-art (SOTA). It is noticeable that existing methods, whether in 4-bit or 8-bit, show significant performance degradation when compared to the FP model on face datasets like CelebA-HQ $256 \\times 256$ and FFHQ $256 \\times 256$ , whereas our TFMQ-DM shows almost no performance degradation compared to the FP model. Importantly, our method achieves significant performance improvement on the LSUN-Bedrooms $256 \\times 256$ in the w4a8 setting, with FID and sFID reductions of 2.26 and 7.51 compared to PTQD [10], respectively. Regarding LDM-8 on LSUN-Churches $256 \\times 256$ , we attribute the moderate improvement, compared to other datasets. We believe that the use of the LDM-8 model with a downsampling factor of 8 may be more quantization-friendly. Existing methods have al", + "bbox": [ + 496, + 537, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7367", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e404313a48862ba617f4abcd4d79dfee743f1f84c7fe069c77185e3050759eef.jpg", + "table_caption": [ + "Table 2. Quantization results for unconditional image generation with LDM-4 on LSUN-Bedrooms 256, FFHQ 256 and CelebA-HQ ${256} \\times {256}$ ,LDM-8 on LSUN-Churches ${256} \\times {256}$ . * represents our implementation according to open-source codes and †means directly rerunning open-source codes." + ], + "table_footnote": [], + "table_body": "
MethodsBits (W/A)LSUN-Bedrooms 256 × 256LSUN-Churches 256 × 256CelebA-HQ 256 × 256FFHQ 256 × 256
FID↓sFID↓FID↓sFID↓FID↓sFID↓FID↓sFID↓
Full Prec.32/322.987.094.1210.898.7410.169.368.67
PTQ4DM* [41]4/324.837.944.9213.9413.6714.7211.7412.18
Q-Diffusion† [23]4/324.207.664.5511.9011.0912.0011.6010.30
PTQD* [10]4/324.427.884.6713.6811.0612.2112.0111.12
TFMQ-DM (Ours)4/323.60 (-0.60)7.61 (-0.05)4.07 (-0.48)11.41 (-0.49)8.74 (-2.32)10.18 (-1.82)9.89 (-1.71)9.06 (-1.24)
PTQ4DM* [41]8/84.759.594.8013.4814.4215.0610.7311.65
Q-Diffusion† [23]8/84.518.174.4112.2312.8514.1610.8710.01
PTQD [10]8/83.759.894.89*14.89*12.76*13.54*10.69*10.97*
TFMQ-DM (Ours)8/83.14 (-0.61)7.26 (-0.91)4.01 (-0.40)10.98 (-1.25)8.71 (-4.05)10.20 (-3.34)9.46 (-1.23)8.73 (-1.28)
PTQ4DM [41]4/820.7254.304.97*14.87*17.08*17.48*11.83*12.91*
Q-Diffusion† [23]4/86.4017.934.6613.9415.5516.8611.4511.15
PTQD [10]4/85.9415.165.10*13.23*15.47*17.38*11.42*11.43*
TFMQ-DM (Ours)4/83.68 (-2.26)7.65 (-7.51)4.14 (-0.52)11.46 (-1.77)8.76 (-6.71)10.26 (-6.60)9.97 (-1.45)9.14 (-2.01)
", + "bbox": [ + 78, + 142, + 890, + 353 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c57ebc81250bdca08b82d6ac6894296d51a4c9eeb1a97bf98a74f9ea54d7474b.jpg", + "table_caption": [ + "Table 3. Quantization results for unconditional image generation with class-conditional image generation with LDM-8 on ImageNet $256 \\times 256$ ." + ], + "table_footnote": [], + "table_body": "
MethodsBits (W/A)ImageNet 256 × 256
IS↑FID↓sFID↓
Full Prec.32/32235.6410.917.67
PTQ4DM [41]4/32---
Q-Diffusion* [23]4/32213.5611.878.76
PTQD† [10]4/32201.7811.659.06
TFMQ-DM (Ours)4/32223.81 (+10.25)10.50 (-1.15)7.98 (-0.78)
PTQ4DM [41]8/8161.7512.59-
Q-Diffusion* [23]8/8187.6512.809.87
PTQD [10]8/8153.9211.948.03
TFMQ-DM (Ours)8/8198.86 (+11.21)10.79 (-1.15)7.65 (-0.38)
PTQ4DM [41]4/8---
Q-Diffusion* [23]4/8212.5110.6814.85
PTQD [10]4/8214.7310.4012.63
TFMQ-DM (Ours)4/8221.82 (+7.09)10.29 (-0.11)7.35 (-5.28)
", + "bbox": [ + 80, + 412, + 468, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ready achieved satisfactory results on this dataset. Nonetheless, our method still approaches the performance of the FP model more closely compared to existing methods. Besides the experiments on LDM, We have also conducted experiments with DDPM on CIFAR-10 $32 \\times 32$ , which can be found in the appendix.", + "bbox": [ + 75, + 643, + 468, + 737 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Class-conditional image generation. On the ImageNet $256 \\times 256$ dataset, we employed a denoising process with 20 iterations, setting eta and CFG to 0.0 and 3.0, respectively. Compared to PTQD, our method achieved a FID reduction of 1.15 on both w4a32 and w8a8. Simultaneously, in the w4a8 setting, sFID decreased by 5.28. Under the same conditions, we observed an improvement of over 7 in IS. Particularly noteworthy is that, across various quantization settings, our method consistently achieved lower FID compared to the FP model.", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d691e5eb0c5e7ea9dc86e69f553368f352f186b757ce7d634d0fec17ede660c4.jpg", + "table_caption": [ + "Table 4. Quantization results for text-guided image generation with Stable-Diffusion on MS-COCO captions." + ], + "table_footnote": [], + "table_body": "
MethodsBits (W/A)MS-COCO
FID↓sFID↓CLIP↑
Full Prec.32/3213.1519.310.3146
Q-Diffusion† [23]4/3213.5819.500.3143
TFMQ-DM (Ours)4/3213.21 (-0.37)19.03 (-0.47)0.3144 (+0.0001)
Q-Diffusion† [23]8/813.3120.540.3134
TFMQ-DM (Ours)8/813.09 (-0.22)19.91 (-0.63)0.3134 (+0.0000)
Q-Diffusion† [23]4/814.4920.430.3121
TFMQ-DM (Ours)4/813.36 (-1.13)20.14 (-0.29)0.3128 (+0.0007)
", + "bbox": [ + 503, + 398, + 890, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text-guided image generation. In this experiment, we sample high-resolution images of $512 \\times 512$ pixels with 50 denoising steps and fix CFG to the default 7.5 in Stable Diffusion as the trade-off between sample quality and diversity. In Tab. 4, compared to the current SOTA Q-Diffusion, our approach achieves an FID reduction of 1.13 on w4a8. Simultaneously, our FID on w8a8 and sFID on w4a32 are even lower than those of the full precision model. However, existing metrics fail to adequately assess the semantic consistency of generated images. Nevertheless, based on the images generated in the appendix, our method produces higher-quality images with more realistic details, better demonstrating semantic information. Furthermore, our generated images closely approximate the effects of FP model.", + "bbox": [ + 496, + 544, + 893, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 498, + 785, + 653, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of each proposed method, we perform a thorough ablation study on the LSUN-Bedrooms $256 \\times 256$ dataset with w4a8 quantization, utilizing the LDM-4 model with a DDIM sampler, as outlined in Tab. 5. We begin the assessment with a baseline BRECQ [24], which shows ineffective in denoising images when oper", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7368", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/65db997857da4b9f02d806c74c5a952b3fb4dae69f4a9be70997ed4663c0fd09.jpg", + "table_caption": [ + "Table 5. The effect of different methods proposed in the paper on LSUN-Bedrooms $256 \\times 256$ ." + ], + "table_footnote": [], + "table_body": "
MethodsBits (W/A)FID↓sFID↓
Full Prec.32/322.987.09
BRECQ [24] (Baseline)4/89.3622.73
+TIAR4/84.849.29
+FSC4/86.0711.31
+TFMQ-DM (TIAR + FSC)4/83.68 (-5.68)7.65 (-15.08)
", + "bbox": [ + 78, + 127, + 467, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ating with 4-bit quantization. Additional ablation experiments can be found in the appendix.", + "bbox": [ + 76, + 224, + 467, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of TIAR. It can be observed that, compared to the baseline, our TIAR method reduces FID and sFID by 4.52 and 13.44, respectively. Furthermore, As shown in Fig. 4, our method's temporal feature disturbance is significantly far weaker than existing SOTA PTQD. This indicates the effectiveness of our method in maintaining temporal information contained in temporal features. Further details of the effects are presented in the appendix.", + "bbox": [ + 75, + 268, + 470, + 390 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/079e6a510385420edf81ad2fe64a74293be6be63b07f66558ee8bf525313dc3c.jpg", + "image_caption": [ + "Figure 4. Temporal feature errors across different PTQ methods." + ], + "image_footnote": [], + "bbox": [ + 132, + 402, + 415, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of FSC. Our FSC method has also achieved remarkably positive results. Compared to the baseline, it reduces FID and sFID by 3.29 and 11.42, respectively. Compared to PTQD, from Fig. 4, our method's temporal feature error is significantly smaller than that of PTQD. Since we employ layer-wise quantization for activations, this introduces less than one percent of additional parameters. The inference time overhead incurred by switching different step sizes and zero points during multiple time-steps inference is negligible, as detailed in the appendix.", + "bbox": [ + 75, + 645, + 468, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Efficiency of FSC. For FSC, there are various methods to evaluate the range of activations to determine the optimal step size. We try several methods and assess the GPU time consumed during calibration, as detailed in Tab. 6. Since the improvements achieved by these methods in model performance are similar, we opt for the simplest and most efficient", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1f2a2e0fd6aab3ef99f36ba8eec0daca28a7b9e315267036a542f623994012ff.jpg", + "table_caption": [ + "Table 6. Different calibration methods for FTFC. We note $T$ as calibration GPU hours in the table." + ], + "table_footnote": [], + "table_body": "
TSC MethodsBits (W/A)FID↓sFID↓T (hours)
FIAR8/322.847.082.18
+LSQ [5]8/83.177.182.48
+KL-divergence8/83.277.3219.67
+Percentile8/83.347.4112.00
+MSE8/83.127.128.89
+Min-max [32]8/83.14 (+0.02)7.26 (+0.14)0.12 (-2.36)
FIAR4/323.047.162.20
+LSQ [5]4/83.697.482.57
+KL-divergence4/83.947.4219.65
+Percentile4/83.748.0212.04
+MSE4/83.627.488.89
+Min-max [32]4/83.68 (+0.06)7.65 (+0.23)0.12 (-2.45)
", + "bbox": [ + 503, + 128, + 888, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Min-max [32] method as our specific calibration strategy, striking a balance between calibration time and effectiveness. Notably, we have found that PTQD and Q-Diffusion cost 4.68 and 5.29 GPU hours in their PTQ methods under w4a8 quantization on LSUN-Bedrooms $256 \\times 256$ , respectively. However, our framework only spends 2.32 GPU hours (as shown in Tab. 6).", + "bbox": [ + 496, + 328, + 890, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, for the only learning-based method in the table, LSQ [5], which is one of the most commonly used methods in previous works [10, 23, 41], we observe that it did not outperform other methods and, in some cases, performs even worse. We speculate that the main reason might be that, for a fixed time-step, the calibration data used to determine the quantization parameters is relatively limited compared to all calibration data. In such cases, LSQ may struggle to learn an optimal set of parameters.", + "bbox": [ + 496, + 434, + 892, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion & Discussions of Limitations", + "text_level": 1, + "bbox": [ + 498, + 584, + 862, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research explores the application of quantization for accelerating diffusion models. In this work, we identify a novel and significant problem, namely temporal feature disturbance, in the quantization of diffusion models. We conducted a detailed analysis of the root causes of this problem and introduced our TFMQ-DM quantization framework. In 4-bit quantization on extensive datasets and different diffusion models, this framework exhibits minimal performance degradation compared to the FP model and speedup quantization time.", + "bbox": [ + 496, + 609, + 890, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, besides temporal features, we have also found the textual features introduced in Stable Diffusion have a physical meaning and influence the generation effect. Nevertheless, these textual features are not taken into consideration in current methods. Secondly, temporal feature maintenance is suitable for both PTQ and QAT scenarios. But in this study, our main focus is on the PTQ setting and achieving 4-bit quantization. The extension to the QAT setting should be conducted in the future to achieve lower-bit quantization and further performance enhancements.", + "bbox": [ + 496, + 761, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7369", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 118, + 174, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yash Bhalgat, Jinwon Lee, Markus Nagel, Tijmen Blankevoort, and Nojun Kwak. $\\mathrm{Lsq}+$ : Improving low-bit quantization through learnable offsets and better initialization, 2020. 1, 2", + "[2] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 1", + "[3] Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In CVPR, pages 12413-12422, 2022. 2", + "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6", + "[5] Steven K. Esser, Jeffrey L. McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S. Modha. Learned step size quantization. In ICLR, 2020. 1, 2, 8", + "[6] Giulio Franzese, Simone Rossi, Lixuan Yang, Alessandro Finamore, Dario Rossi, Maurizio Filippone, and Pietro Michiardi. How much is enough? a study on diffusion times in score-based generative models. arXiv preprint arXiv:2206.05173, 2022. 2", + "[7] Ruihao Gong, Xianglong Liu, Shenghu Jiang, Tianxiang Li, Peng Hu, Jiazhen Lin, Fengwei Yu, and Junjie Yan. Differentiable soft quantization: Bridging full-precision and low-bit neural networks. In ICCV, pages 4852–4861, 2019. 1, 2", + "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 1", + "[9] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2015. 6", + "[10] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. In NeurIPS, 2023. 2, 3, 4, 6, 7, 8", + "[11] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning, 2022. 6", + "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018. 5, 6", + "[13] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 1" + ], + "bbox": [ + 78, + 143, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3, 6", + "[15] Itay Hubara, Yury Nahshan, Yair Hanani, Ron Banner, and Daniel Soudry. Improving post training neural quantization: Layer-wise calibration and integer programming. arXiv preprint arXiv:2006.10518, 2020. 1, 2", + "[16] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In CVPR, pages 2704-2713, 2018. 2, 6", + "[17] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis, 2023. 1", + "[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation, 2018. 6", + "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks, 2019. 6", + "[20] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2022. 1", + "[21] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. arXiv preprint arXiv:2106.00132, 2021. 1, 2", + "[22] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[23] Xiuyu Li, Long Lian, Yijiang Liu, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In ICCV, 2023. 1, 2, 3, 4, 6, 7, 8", + "[24] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. BRECQ: pushing the limit of post-training quantization by block reconstruction. In ICLR, 2021. 1, 2, 5, 6, 7, 8", + "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 6", + "[26] Yang Lin, Tianyu Zhang, Peiqin Sun, Zheng Li, and Shuchang Zhou. Fq-vit: Post-training quantization for fully quantized vision transformer. In *IJCAI*, pages 1173–1179, 2022. 2", + "[27] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In ICLR, 2022. 1", + "[28] Christos Louizos, Matthias Reisser, Tijmen Blankevoort, Efstratos Gavves, and Max Welling. Relaxed quantization for discretized neural networks. In ICLR, 2019. 2", + "[29] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, 2022. 1", + "[30] Zhaoyang Lyu, Xudong Xu, Ceyuan Yang, Dahua Lin, and Bo Dai. Accelerating diffusion models via early stop of the diffusion process. arXiv preprint arXiv:2205.12524, 2022. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7370", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Markus Nagel, Rana Ali Amjad, Mart van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. In ICML, pages 7197-7206, 2020. 1, 2, 6", + "[32] Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bondarenko, Mart van Baalen, and Tijmen Blankevoort. A white paper on neural network quantization, 2021. 1, 5, 6, 8", + "[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171, 2021. 3", + "[34] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 6", + "[35] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 3", + "[36] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. In ECCV, pages 525-542, 2016. 6", + "[37] Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. Fastspeech 2: Fast and high-quality end-to-end text to speech, 2022. 1", + "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 6", + "[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation, 2015. 3", + "[40] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans, 2016. 5, 6", + "[41] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In CVPR, 2023. 2, 3, 7, 8", + "[42] Jonathan Shen, Ruoming Pang, Ron J. Weiss, Mike Schuster, Navdeep Jaitly, Zongheng Yang, Zhifeng Chen, Yu Zhang, Yuxuan Wang, RJ Skerry-Ryan, Rif A. Saurous, Yannis Agiomyrgiannakis, and Yonghui Wu. Natural tts synthesis by conditioning wavenet on mel spectrogram predictions, 2018. 1", + "[43] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for diffusion models. In NeurIPS, 2023. 2, 3, 4, 5", + "[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1, 2, 3", + "[45] Changyuan Wang, Ziwei Wang, Xiuwei Xu, Yansong Tang, Jie Zhou, and Jiwen Lu. Towards accurate data-free quantization for diffusion models. arXiv preprint arXiv:2305.18723, 2023. 2, 5", + "[46] Peisong Wang, Qiang Chen, Xiangyu He, and Jian Cheng. Towards accurate post-training network quantization via bitsplit and stitching. In ICML, pages 9847-9856, 2020. 2", + "[47] Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. Qdrop: Randomly dropping quantization for" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "extremely low-bit post-training quantization. In ICLR, 2022. 1, 2", + "[48] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop, 2016. 1, 5, 6", + "[49] Luoming Zhang, Yefei He, Zhenyu Lou, Xin Ye, Yuxing Wang, and Hong Zhou. Root quantization: a self-adaptive supplement ste. Applied Intelligence, 53(6):6266-6275, 2023. 2", + "[50] Qinsheng Zhang, Molei Tao, and Yongxin Chen. gddim: Generalized denoising diffusion implicit models. arXiv preprint arXiv:2206.05564, 2022. 2", + "[51] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. Opt: Open pre-trained transformer language models, 2022. 1", + "[52] Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Truncated diffusion probabilistic models. stat, 1050:7, 2022. 2", + "[53] Bohan Zhuang, Chunhua Shen, Mingkui Tan, Lingqiao Liu, and Ian Reid. Towards effective low-bitwidth convolutional neural networks. In CVPR, pages 7920-7928, 2018. 2" + ], + "bbox": [ + 501, + 92, + 890, + 445 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7371", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/a91d56f7-042a-42e5-9c77-ba0dbaac2109_model.json b/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/a91d56f7-042a-42e5-9c77-ba0dbaac2109_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f7ece7ec1609dfc01f681ce8d7f76fd67ee83440 --- /dev/null +++ b/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/a91d56f7-042a-42e5-9c77-ba0dbaac2109_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.085, + 0.131, + 0.888, + 0.153 + ], + "angle": 0, + "content": "TFMQ-DM: Temporal Feature Maintenance Quantization for Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.177, + 0.18, + 0.803, + 0.22 + ], + "angle": 0, + "content": "Yushi Huang\\(^{1,2*}\\), Ruihao Gong\\(^{1,2*}\\), Jing Liu\\(^{2,3}\\), Tianlong Chen\\(^{4}\\), Xianglong Liu\\(^{1*}\\) \n\\(^{1}\\)Beihang University \\(^{2}\\)SenseTime Research \\(^{3}\\)Monash University \\(^{4}\\)UT Austin" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.221, + 0.832, + 0.235 + ], + "angle": 0, + "content": "{huangyushi, gongruihao, liujing5} @sentetime.com, tianlong.chen@utexas.edu, xlliu@buaa.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.473, + 0.807 + ], + "angle": 0, + "content": "The Diffusion model, a prevalent framework for image generation, encounters significant challenges in terms of broad applicability due to its extended inference times and substantial memory requirements. Efficient Post-training Quantization (PTQ) is pivotal for addressing these issues in traditional models. Different from traditional models, diffusion models heavily depend on the time-step \\( t \\) to achieve satisfactory multi-round denoising. Usually, \\( t \\) from the finite set \\( \\{1,\\dots,T\\} \\) is encoded to a temporal feature by a few modules totally irrespective of the sampling data. However, existing PTQ methods do not optimize these modules separately. They adopt inappropriate reconstruction targets and complex calibration methods, resulting in a severe disturbance of the temporal feature and denoising trajectory, as well as a low compression efficiency. To solve these, we propose a Temporal Feature Maintenance Quantization (TFMQ) framework building upon a Temporal Information Block which is just related to the time-step \\( t \\) and unrelated to the sampling data. Powered by the pioneering block design, we devise temporal information aware reconstruction (TIAR) and finite set calibration (FSC) to align the full-precision temporal features in a limited time. Equipped with the framework, we can maintain the most temporal information and ensure the end-to-end generation quality. Extensive experiments on various datasets and diffusion models prove our state-of-the-art results. Remarkably, our quantization approach, for the first time, achieves model performance nearly on par with the full-precision model under 4-bit weight quantization. Additionally, our method incurs almost no extra computational cost and accelerates quantization time by \\( 2.0\\times \\) on LSUN-Bedrooms \\( 256\\times 256 \\) compared to previous works. Our code is publicly available at https://github.com/ModelTC/TFMQ-DM." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.269, + 0.632, + 0.285 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.895, + 0.658 + ], + "angle": 0, + "content": "Generative modeling plays a crucial role in machine learning, particularly in applications like image [13, 14, 17, 38, 44], voice [37, 42], and text synthesis [2, 51]. Diffusion models have showcased impressive capabilities in producing high-quality samples across diverse domains. In comparison to generative adversarial networks (GANs) [8] and variational autoencoders (VAEs) [20], diffusion models successfully sidestep issues such as model collapse and posterior collapse, resulting in a more stable training process. However, the substantial computational cost poses a critical bottleneck hampering the widespread adoption of diffusion models. Furthermore, the computational cost for diffusion models can be attributed to two primary factors. First, these models typically require hundreds of denoising steps to generate images, rendering the procedure considerably slower than that of GANs. Prior efforts [21, 27, 29, 44] have addressed this challenge by seeking shorter and more efficient sampling trajectories, thereby reducing the number of denoising steps. Second, the substantial network architecture of diffusion models demands considerable time and memory resources, particularly for foundational models pre-trained on large-scale datasets, e.g., LDM [38] and Stable Diffusion. Our work aims to tackle the latter challenge, focusing on the compression of diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Quantization is currently the most widely used method [1, 5, 7, 31, 32] for compressing models by mapping high-precision floating-point numbers to low-precision numbers. Among different quantization methods, Posttraining quantization (PTQ) [15, 31, 47] incurs lower overhead and is more user-friendly without the need for retraining or fine-tuning. While PTQ on conventional models has undergone extensive study [7, 24, 31, 47], its application to diffusion models has shown huge performance degradation, especially under low-bit settings. For instance, Q-Diffusion [23] exhibits severe accuracy drop on some datasets [48] under 4-bit quantization. We believe the reason they fail to achieve better results is that they all overlook the sampling data independence and uniqueness of temporal features, which are generated from time-step \\( t \\) through a few modules, used to control the denoising trajectory in dif" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.863, + 0.21, + 0.875 + ], + "angle": 0, + "content": "*Equal Contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.876, + 0.398, + 0.888 + ], + "angle": 0, + "content": "\\(\\dagger\\) Work done during an internship at SenseTime Research." + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.888, + 0.226, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Corresponding Author." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.863, + 0.398, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7362" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "fusion models. Furthermore, we observe that temporal feature disturbance significantly impacts model performance in the aforementioned methods." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.148, + 0.473, + 0.571 + ], + "angle": 0, + "content": "To tackle temporal feature disturbance, we first find that the modules generating temporal features are independent of the sampling data and define the whole modules as the Temporal Information Block. All existing methods do not separately optimize this block during the quantization process, causing temporal features to overfit to limited calibration data. On the other hand, since the maximum time-step for denoising is a finite positive integer, the temporal feature and the activations during its generation form a finite set. The optimal approach is to optimize each element in this set individually. Based on these observations and analyses, we propose a novel quantization reconstruction approach, temporal information aware reconstruction (TIAR), specifically optimizing for temporal features. It aims to reduce temporal feature loss as the optimization objective while isolating the network's components related to sampled data from the generation of temporal features during calibration. Furthermore, we also introduce a calibration strategy, finite set calibration (FSC), for the finite set of temporal features and activations during its generation. This strategy employs different quantization parameters for activations corresponding to different time-steps. Moreover, the calibration speed of this method is faster than existing mainstream methods [1, 5], for example, we speedup quantization time by \\(2.0 \\times\\) on LSUN-Bedrooms \\(256 \\times 256\\) dataset, yet the strategy incurs negligible inference and storage overhead. The overview of our framework can be seen in Fig. 1. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.581, + 0.47, + 0.642 + ], + "angle": 0, + "content": "- We discover that existing quantization methods suffer from temporal feature disturbance, disrupting the denoising trajectory of diffusion models and significantly affecting the quality of generated images." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.643, + 0.47, + 0.701 + ], + "angle": 0, + "content": "- We reveal that the disturbance comes from two aspects: inappropriate reconstruction target and unaware of finite activations. Both inducements ignore the special characteristics of time information-related modules." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.702, + 0.47, + 0.792 + ], + "angle": 0, + "content": "- An advanced framework (TFMQ-DM) is proposed, consisting of temporal information aware reconstruction (TIAR) for weight quantization and finite set calibration (FSC) for activation quantization. Both are based on a Temporal Information Block specially devised for diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.793, + 0.471, + 0.899 + ], + "angle": 0, + "content": "- Extensive experiments on various datasets show that our novel framework achieves a new state-of-the-art result in PTQ of diffusion models, especially under 4-bit weight quantization, and significantly accelerates quantization time. For some hard tasks, e.g., CelebA-HQ \\(256 \\times 256\\), our method reduces the FID score by 6.71 (images in appendix)." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.581, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.643, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.738, + 0.13 + ], + "angle": 0, + "content": "2.1. Efficient Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.894, + 0.306 + ], + "angle": 0, + "content": "There are diverse perspectives to accelerate the inference of diffusion model, e.g., retraining-based model design [3, 6, 30, 52] and retraining-free sampler strategy [21, 44, 50]. However, the retraining-based method proves resource-intensive and time-consuming. The efficient samplers can reduce sampling iterations but still suffer from diffusion models' extensive parameters and computational complexity. In this paper, we focus on diminishing the time and memory overhead of the single-step denoising process using low-bit quantization in a training-free manner, a method orthogonal to previous speedup techniques." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.312, + 0.692, + 0.328 + ], + "angle": 0, + "content": "2.2. Model Quantization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.895, + 0.502 + ], + "angle": 0, + "content": "Quantization is a predominant technique for minimizing storage and computational costs. It can be categorized into quantization-aware training (QAT) [7, 16, 28, 49, 53] and post-training quantization (PTQ) [15, 24, 26, 31, 47]. QAT requires intensive model training with substantial data and computational demands. Correspondingly, PTQ compresses models without re-training, making it a preferred method due to its minimal data requirements and easy deployment on real hardware. In PTQ, high-precision values are mapped into discrete levels using uniform quantization expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.507, + 0.892, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\hat {x} = \\Phi \\left(\\left\\lfloor \\frac {x}{s} \\right\\rceil + z, 0, 2 ^ {b} - 1\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.894, + 0.674 + ], + "angle": 0, + "content": "where \\( x \\) represents floating-point data, \\( \\hat{x} \\) is the quantized value, \\( s \\) is the quantization step size, \\( z \\) is the zero offset, and \\( b \\) is the target bit-width. The clamp function \\( \\Phi(\\cdot) \\) clips the rounded value \\( \\left\\lfloor \\frac{x}{s} \\right\\rfloor + z \\) within the range of \\( [0, 2^b - 1] \\). However, naive quantization may lead to accuracy degradation, especially for low-bit quantization. Recent studies [15, 24, 31, 46, 47] have explored innovative strategies based on reconstruction to preserve model performance after low-bit quantization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In contrast, the iterative denoising process in diffusion models presents new challenges for PTQ in comparison to traditional models. PTQ4DM [41] represents the initial attempt to quantize diffusion models to 8-bit, albeit with limited experiments and lower resolutions. Conversely, Q-Diffusion [23] achieves enhanced performance and is evaluated on a broader dataset range. Moreover, PTQD [10] eliminates quantization noise through correlated and residual noise correction. Notably, traditional single-time-step PTQ calibration methods are unsuitable for diffusion models due to significant activation distribution changes with each time-step [23, 41, 43, 45]. ADP-DM [45] proposes group-wise quantization across time-steps for diffusion models, and TDQ [43] introduces distinct quantization parameters for different time-steps. However, all of the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7363" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "above works overlook the specificity of temporal features. To address temporal feature disturbance in the aforementioned works, our study delves into the inducements of the phenomenon and introduces a novel reconstruction and calibration framework, significantly enhancing quantized diffusion model performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.195, + 0.216, + 0.21 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.221, + 0.469, + 0.311 + ], + "angle": 0, + "content": "Diffusion models. Diffusion models [14, 44] iteratively add Gaussian noise with a variance schedule \\(\\beta_{1},\\ldots ,\\beta_{T}\\in\\) \\((0,1)\\) to data \\(\\mathbf{x}_0\\sim q(\\mathbf{x})\\) for \\(T\\) times as sampling process, resulting in a sequence of noisy samples \\(\\mathbf{x}_1,\\dots ,\\mathbf{x}_T\\) In DDPMs [14], the former sampling process is a Markov chain, taking the form:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.321, + 0.469, + 0.362 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.365, + 0.397, + 0.383 + ], + "angle": 0, + "content": "\\[\nq (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1}) = \\mathcal {N} (\\mathbf {x} _ {t}; \\sqrt {\\alpha_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.469, + 0.526 + ], + "angle": 0, + "content": "where \\(\\alpha_{t} = 1 - \\beta_{t}\\). Conversely, the denoising process removes noise from a sample from Gaussian noise \\(\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) to gradually generate high-fidelity images. Nevertheless, due to the unavailability of the true reverse conditional distribution \\(q(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)\\), diffusion models approximate it via variational inference by learning a Gaussian distribution \\(p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t) = \\mathcal{N}(\\mathbf{x}_{t - 1};\\boldsymbol {\\mu}_\\theta (\\mathbf{x}_t,t),\\boldsymbol {\\Sigma}_\\theta (\\mathbf{x}_t,t))\\), the \\(\\boldsymbol{\\mu}_{\\theta}\\) can be derived by reparameterization trick as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.535, + 0.469, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.469, + 0.641 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_{t} = \\prod_{i=1}^{t} \\alpha_{i}\\) and \\(\\epsilon_{\\theta}(\\cdot)\\) is a noise estimation model. The variance \\(\\Sigma_{\\theta}(\\mathbf{x}_{t}, t)\\) can be either learned [33] or fixed to a constant schedule [14] \\(\\sigma_{t}\\). When employing the latter method, \\(\\mathbf{x}_{t-1}\\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.651, + 0.469, + 0.685 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {x} _ {t}, t)\\right) + \\sigma_ {t} \\mathbf {z}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.694, + 0.212, + 0.709 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.722, + 0.469, + 0.873 + ], + "angle": 0, + "content": "Reconstruction on diffusion models. UNet [39], the predominant model employed as \\(\\epsilon_{\\theta}(\\cdot)\\) in Eq. 4 to predict Gaussian noise, can be divided into blocks that incorporate residual connections (such as Residual Bottleneck Blocks or Transformer Blocks [35]) and the remaining layers. Numerous PTQ approaches for diffusion models are grounded in layer/block-wise reconstruction [10, 23, 41, 43] to obtain optimal quantization parameters. For example, in the Residual Bottleneck Block, this approach typically minimizes the following loss function as its optimization objective:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.883, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i} = \\left\\| f _ {i} (\\cdot) - \\widehat {f} _ {i} (\\cdot) \\right\\| _ {F} ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.169 + ], + "angle": 0, + "content": "where \\( \\| \\cdot \\|_F^2 \\) denotes the Frobenius norm. The function \\( f_i(\\cdot) \\) represents the \\( i^{\\text{th}} \\) Residual Bottleneck Block, and \\( \\widehat{f}_i(\\cdot) \\) is its quantized counterpart. Furthermore, in the ensuing sections, we use \\( n \\) to denote the total number of Residual Bottleneck Blocks in a single diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.182, + 0.892, + 0.378 + ], + "angle": 0, + "content": "Temporal feature in diffusion models. Seeing from Fig. 1 (a), time-step \\( t \\) is encoded with time embed\\(^{1}\\) and then passes through the embedding layer\\(^{2}\\) in each Residual Bottleneck Block, resulting in a series of unique activations. In this paper, we denote these activations as temporal features. Notably, temporal features are independent of \\( \\mathbf{x}_t \\) and unrelated to other temporal features from different time-steps. To enhance clarity, we simplify our notation as follows: we represent time embed as \\( h(\\cdot) \\), the embedding layer in the \\( i^{\\text{th}} \\) Residual Bottleneck Block as \\( g_i(\\cdot) \\), and denote the \\( i^{\\text{th}} \\) temporal feature at time-step \\( t \\) as \\( \\mathbf{emb}_{t,i} \\). Moreover, as illustrated in Fig. 1 (a), the relationship is explicitly expressed by the equation:" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.39, + 0.892, + 0.407 + ], + "angle": 0, + "content": "\\[\n\\text {T e m p o r a l} \\quad \\mathbf {e m b} _ {t, i} = g _ {i} (h (t)) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.418, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Additionally, we have found that temporal features play a crucial role in the context of the diffusion model, holding unique and substantial physical implications. These features encompass temporal information that signifies the current image's temporal position along the denoising trajectory. Within the UNet structure, each time-step transforms into these temporal features, thereby controlling the denoising trajectory by applying them to the features of images generated at each iteration." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.567, + 0.762, + 0.584 + ], + "angle": 0, + "content": "4. TFMQ for Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.593, + 0.892, + 0.655 + ], + "angle": 0, + "content": "We present our novel PTQ framework in this section. We first observe temporal disturbance in previous methods in Sec. 4.1 and then analyze the inducements in Sec. 4.2. Finally, we propose our solutions in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.663, + 0.774, + 0.679 + ], + "angle": 0, + "content": "4.1. Temporal Feature Disturbance" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.686, + 0.892, + 0.747 + ], + "angle": 0, + "content": "Based on Sec. 3, we investigate the impact of previous PTQ works on temporal features, and we identify the phenomenon of temporal feature disturbance, which significantly deteriorates the quality of generated images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.759, + 0.892, + 0.835 + ], + "angle": 0, + "content": "Temporal feature error. We thoroughly analyze temporal feature variations before and after the quantization of embedding layers and time embed in the Stable Diffusion model (\\(T = 50, i = 11\\)). Prior to this analysis, we introduce the temporal feature error as defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.845, + 0.892, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\cos \\left(\\widehat {\\mathbf {e m b}} _ {t, i}, \\widehat {\\mathbf {e m b}} _ {t, i}\\right), \\tag {7}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.875, + 0.829, + 0.888 + ], + "angle": 0, + "content": "PyTorch time embed implementation in diffusion models." + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.888, + 0.854, + 0.9 + ], + "angle": 0, + "content": "PyTorch embedding layer implementation in diffusion models." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.875, + 0.854, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7364" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.096, + 0.466, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.208, + 0.466, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.09, + 0.885, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.275, + 0.892, + 0.371 + ], + "angle": 0, + "content": "Figure 1. Overview of the proposed Temporal Feature Maintenance Quantization. (a) Temporal Feature \\(\\mathbf{emb}_{t,i}\\), belonging to a finite set representing temporal information, has been overlooked in previous works due to inappropriate reconstruction targets (box with a solid line). (b) This oversight leads to a severe disturbance for \\(\\mathbf{emb}_{t,i}\\) and results in the mismatch of crucial temporal information for the diffusion model's generation, causing a deviation in the denoising trajectory and a significant drop in accuracy. (c) Based on these analyses, we introduce a Temporal Information Block that exclusively correlates with the time-step \\(t\\). Leveraging this \\(\\mathbf{x}_t\\)-unrelated block, we enable Temporal Information Aware Reconstruction and Finite Set Calibration (utilizing the finite number of \\(t\\)). This approach achieves the maintenance of temporal features and yields state-of-the-art results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.47, + 0.493 + ], + "angle": 0, + "content": "where \\(\\cos (\\cdot)\\) denotes cosine similarity, and \\(\\widehat{\\mathbf{emb}_{t,i}}\\) signifies the temporal feature corresponding to \\(\\mathbf{emb}_{t,i}\\) in the quantized model. As illustrated in Fig. 2 (Left), quantization induces notable temporal feature errors. We term this phenomenon, characterized by substantial temporal feature errors within diffusion models, as temporal feature disturbance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.469, + 0.612 + ], + "angle": 0, + "content": "Temporal information mismatch. Temporal feature disturbance alters the original embedded temporal information. Specifically, \\(\\mathbf{emb}_{t,i}\\) is intended to correspond to timestep \\(t\\). However, due to significant errors, the quantized model's \\(\\widehat{\\mathbf{emb}}_{t,i}\\) is no longer accurately associated with \\(t\\), resulting in what we term as temporal information mismatch:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.612, + 0.469, + 0.631 + ], + "angle": 0, + "content": "\\[\nt \\leftarrow \\mathbf {e m b} _ {t, i}, \\quad t \\nleftarrow \\widehat {\\mathbf {e m b}} _ {t, i}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.634, + 0.469, + 0.754 + ], + "angle": 0, + "content": "Furthermore, as depicted in Fig. 2 (Right), we even observe a pronounced temporal information mismatch. Specifically, the temporal feature generated by the quantized model at time-step \\( t \\) exhibits a divergence from that of the full-precision model at the corresponding time-step. Instead, it tends to align more closely with the temporal feature corresponding to \\( t + \\delta_t \\), importing wrong temporal information from \\( t + \\delta_t \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.767, + 0.469, + 0.828 + ], + "angle": 0, + "content": "Trajectory deviation. Temporal information mismatch delivers wrong temporal information, therefore, causing a deviation in the corresponding temporal position of the image within the denoising trajectory, ultimately leading to:" + }, + { + "type": "equation", + "bbox": [ + 0.231, + 0.835, + 0.469, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t} \\neq \\mathbf {x} _ {t - 1}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where we apply disrupted temporal features to the model. Evidently, as the deviation in the denoising trajectory accumulates with the increase in the number of denoising" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.386, + 0.696, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.386, + 0.89, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.637 + ], + "angle": 0, + "content": "Figure 2. (Left) Temporal feature disturbance. The inflection points serve as indicators of temporal feature errors at different time-steps, and they highlight the significant phenomenon of temporal feature disturbance. (Right) Temporal information mismatch. The coordinates of the inflection points on the blue curve can be denoted as \\((t,t + \\delta_{t,i})\\). It indicates \\(\\mathbf{emb}_{t + \\delta_{t,i},i}\\) exhibits the highest similarity with \\(\\mathbf{emb}_{t,i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.715 + ], + "angle": 0, + "content": "iterations, the final generated image struggles to align with \\(\\mathbf{x}_0\\). This evolution is illustrated in Fig. 3, where we maintain UNet excluding embedding layers and time embed in full precision." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.728, + 0.7, + 0.743 + ], + "angle": 0, + "content": "4.2. Inducement Analyses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.752, + 0.892, + 0.827 + ], + "angle": 0, + "content": "In this section, we explore the two inducements of temporal feature disturbance. For the purpose of clarity, in the subsequent sections, \"reconstruction\" specifically points to slight weight adjustment for minimal quantization error, while \"calibration\" specifically refers to activation calibration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Inappropriate reconstruction target. Previous PTQ works [10, 23, 43] have achieved remarkable progress on diffusion models. However, these existing methods overlook the temporal feature's independence and its dis" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "7365" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.892, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.266, + 0.894, + 0.34 + ], + "angle": 0, + "content": "Figure 3. Denoising process of full-precision (Upper) and w4a8 quantized (Lower) Stable-Diffusion \\((T = 50)\\) under the same experiment settings and prompt: A man in the snow on a snow board. We represent \\(\\{\\mathbf{emb}_{t,i}\\}_{i=0,\\dots,n}\\) and \\(\\{\\widehat{\\mathbf{emb}}_{t,i}\\}_{i=0,\\dots,n}\\) as \\(\\mathbf{EMB}_t\\) and \\(\\widehat{\\mathbf{EMB}_t}\\), respectively. Additionally, we denote \\(\\widehat{\\mathbf{x}}_t\\) as \\(\\mathbf{x}_t\\) in the context of the quantized diffusion model. It is noteworthy that, in the quantized model employed here, to showcase the impact of temporal features, only the layers in Temporal Information Block are quantized and the components unrelated to the generation of temporal features are maintained in full precision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.47, + 0.444 + ], + "angle": 0, + "content": "tive physical significance. During their reconstruction processes, there was a lack of optimization for the embedding layers; instead, a Residual Bottleneck Block of coarser granularity was selected as the reconstruction target. This method involves two potential factors causing temporal feature disturbance:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.447, + 0.47, + 0.492 + ], + "angle": 0, + "content": "- Optimize the objective as expressed in Eq. 5 to decrease the reconstruction loss of Residual Bottleneck Block, as opposed to direct reduction of temporal feature loss." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.492, + 0.47, + 0.552 + ], + "angle": 0, + "content": "- During backpropagation of the reconstruction process, embedding layers independent from \\(\\mathbf{x}_t\\) are affected by \\(\\mathbf{x}_t\\), resulting in an overfitting scenario on limited calibration data." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.447, + 0.47, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.471, + 0.734 + ], + "angle": 0, + "content": "To further validate our analyses, we respectively evaluate the FID [12] and sFID [40] for this reconstruction method, e.g., BRECQ [24] and the approach where we freeze the parameters of the embedding layers during the reconstruction of the Residual Bottleneck Block, initializing the embedding layers solely through Min-max [32] for comparison. As shown in Tab. 1, the Freeze strategy exhibits better results, which verify that embedding layers serve as their own optimization objective and maintain their independence of \\(\\mathbf{x}_t\\) can significance mitigate temporal feature disturbance, especially at low-bit." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.749, + 0.471, + 0.791 + ], + "angle": 0, + "content": "Table 1. FID and sFID on LSUN-Bedrooms \\(256 \\times 256\\) [48] for LDM-4 with 50000 sampling images. Prev represents BRECQ. Freeze denotes our trial." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.803, + 0.446, + 0.899 + ], + "angle": 0, + "content": "
MethodsBits (W/A)FID↓sFID↓
Full Prec.32/322.987.09
Prev8/87.5112.54
Freeze8/85.76 (-1.75)8.42 (-4.12)
Prev4/89.3622.73
Freeze4/87.08 (-2.28)16.82 (-5.91)
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.892, + 0.58 + ], + "angle": 0, + "content": "Unaware of finite activations within \\( h(\\cdot) \\) and \\( g_{i}(\\cdot) \\). We observe that, given \\( T \\) as a finite positive integer, the set of all possible activation values for embedding layers and time embed is finite and strictly dependent on timesteps. Within this set, activations corresponding to the same layer display notable range variations across different timesteps (refer to the appendix). Previous methods [43, 45] mainly focus on finding the optimal calibration method for \\( \\widehat{\\mathbf{x}}_t \\)-related network components. Moreover, akin to the first inducement, their calibration is directly towards the Residual Bottleneck Block, which proves suboptimal (refer to the appendix). However, based on the finite activations, we can employ calibration methods, especially for these time information-related activations, to better adapt to their range variations." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.592, + 0.78, + 0.608 + ], + "angle": 0, + "content": "4.3. Temporal Feature Maintenance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.892, + 0.694 + ], + "angle": 0, + "content": "To address the problem of temporal feature disturbance, we design a novel Temporal Information Block to maintain the temporal features. Based on the block, Temporal Information Aware Reconstruction and Finite Set Calibration are proposed to solve the two inducements analyzed above." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.706, + 0.892, + 0.829 + ], + "angle": 0, + "content": "Temporal Information Block. Based on the inducements, it is crucial to meticulously separate the reconstruction and calibration process for each embedding layer and Residual Bottleneck Block to enhance quantized model performance. Considering the unique structure of the UNet, we consolidate all embedding layers and time embed into a unified Temporal Information Block, which can be denoted as \\(\\{g_{i}(h(\\cdot))\\}_{i = 0,\\dots,n}\\) (see Fig. 1 (c))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Temporal information aware reconstruction. Based on the Temporal Information Block, we propose Temporal information aware reconstruction (TIAR) to tackle the first inducement. The optimization objective for the block during" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7366" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.343, + 0.106 + ], + "angle": 0, + "content": "the reconstruction process is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.12, + 0.47, + 0.16 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T I A R} = \\sum_ {i = 0} ^ {n} \\| g _ {i} (h (t)) - \\widehat {g} _ {i} (\\widehat {h} (t)) \\| _ {F} ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.174, + 0.47, + 0.222 + ], + "angle": 0, + "content": "where \\(\\widehat{h} (\\cdot)\\) and \\(\\widehat{g}_i(\\cdot)\\) are quantized versions of \\(h(\\cdot)\\) and \\(g(\\cdot)\\), respectively. With this reconstruction, weights are adjusted to pursue a minimal disturbance for temporal features." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.234, + 0.469, + 0.356 + ], + "angle": 0, + "content": "Finite set calibration. To address the challenge posed by the wide span of activations within a finite set for the second inducement, we propose Finite Set Calibration (FSC) for activation quantization. This strategy employs \\( T \\) sets of quantization parameters for every activation within all the embedding layers and time embed, such as \\( \\{(s_{T},z_{T}),\\ldots ,(s_{1},z_{1})\\} \\) for activation \\( \\mathbf{x} \\). In time-step \\( t \\), the quantization function for the \\( \\mathbf{x} \\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.368, + 0.47, + 0.402 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} = \\Phi \\left(\\left\\lfloor \\frac {\\mathbf {x}}{s _ {t}} \\right\\rceil + z _ {t}, 0, 2 ^ {b} - 1\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.469, + 0.504 + ], + "angle": 0, + "content": "To be noted, the calibration target is also aligned with the output of the Temporal Information Block. For a more specific estimation method of activation ranges, since the solution space within the finite set is limited, we find the Minmax [32] method can achieve satisfactory results with high efficiency (more evidence in Sec. 5.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.52, + 0.21, + 0.538 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.546, + 0.292, + 0.563 + ], + "angle": 0, + "content": "5.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.768 + ], + "angle": 0, + "content": "Models and datasets. In this section, we conduct image generation experiments to evaluate the proposed TFMQ-DM framework on various diffusion models: pixel-space diffusion model DDPM [14] for unconditional image generation, latent-space diffusion model LDM [38] for unconditional image generation and class-conditional image generation. We also apply our work to Stable Diffusion-v1-4 for text-guided image generation. In our experiments, We use seven standard benchmarks: CIFAR-10 \\(32 \\times 32\\) [22], LSUN-Bedrooms \\(256 \\times 256\\) [48], LSUN-Churches \\(256 \\times 256\\) [48], CelebA-HQ \\(256 \\times 256\\) [18], ImageNet \\(256 \\times 256\\) [4], FFHQ \\(256 \\times 256\\) [19] and MS-COCO [25]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Quantization settings. We use channel-wise quantization for weights and layer-wise quantization for activations, as it is a common practice. In our experimental setup, we employ BRECQ [24] and AdaRound [31]. Drawing from empirical insights derived from conventional model quantization practices [9, 36], we maintain the input and output layers of the model in full precision. Calibration sets, integral to our methodology, are generated through full-precision" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "diffusion models, mirroring the approach outlined in Q-Diffusion [23]. Moreover, in weight quantization, we reconstruct quantized weights for 20k iterations with a minibatch size of 32 for DDPM and LDM, and 8 for Stable Diffusion. In activation quantization, we utilize EMA [16] to estimate the ranges of activations with a mini-batch size of 16 on all models. More details can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.212, + 0.893, + 0.5 + ], + "angle": 0, + "content": "Evaluation metrics. For each experiment, we evaluate the performance of diffusion models with Fréchet Inception Distance (FID) [12]. In the case of LDM and Stable Diffusion experiments, we also include sFID [40], which better captures spatial relationships than FID. For ImageNet and CIFAR-10 experiments, we additionally provide Inception Score (IS) [40] as a reference metric. Further, in the context of Stable Diffusion experiments, we extend our evaluation to include the compatibility of image-caption pairs, employing the CLIP score [11]. The ViT-B/32 is used as the backbone when computing the CLIP score. To ensure consistency in the reported outcomes, all results are derived from our implementation or from other papers, where experiments are conducted under conditions consistent with ours. More specifically, in the evaluation process of each experiment, we sample 50k images from DDPM or LDM, or 30k images from Stable-Diffusion. All experiments are conducted utilizing one H800 GPU and implemented with the PyTorch framework [34]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.64, + 0.529 + ], + "angle": 0, + "content": "5.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Unconditional image generation. In the experiments conducted on the LDM, we maintain the same experimental settings as presented in [38], including the number of steps, variance schedule, and classifier-free guidance scale (denoted by eta and CFG in the following, respectively). As shown in Tab. 2, the FID performance differences relative to the full precision (FP) model are all within 0.7 for all settings. Specifically, on the CelebA-HQ \\(256 \\times 256\\) dataset, our method exhibits a FID reduction of 6.71 and a sFID reduction of 6.60 in the w4a8 setting compared to the current state-of-the-art (SOTA). It is noticeable that existing methods, whether in 4-bit or 8-bit, show significant performance degradation when compared to the FP model on face datasets like CelebA-HQ \\(256 \\times 256\\) and FFHQ \\(256 \\times 256\\), whereas our TFMQ-DM shows almost no performance degradation compared to the FP model. Importantly, our method achieves significant performance improvement on the LSUN-Bedrooms \\(256 \\times 256\\) in the w4a8 setting, with FID and sFID reductions of 2.26 and 7.51 compared to PTQD [10], respectively. Regarding LDM-8 on LSUN-Churches \\(256 \\times 256\\), we attribute the moderate improvement, compared to other datasets. We believe that the use of the LDM-8 model with a downsampling factor of 8 may be more quantization-friendly. Existing methods have al" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7367" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 2. Quantization results for unconditional image generation with LDM-4 on LSUN-Bedrooms 256, FFHQ 256 and CelebA-HQ \\( {256} \\times {256} \\) ,LDM-8 on LSUN-Churches \\( {256} \\times {256} \\) . * represents our implementation according to open-source codes and †means directly rerunning open-source codes." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.143, + 0.891, + 0.354 + ], + "angle": 0, + "content": "
MethodsBits (W/A)LSUN-Bedrooms 256 × 256LSUN-Churches 256 × 256CelebA-HQ 256 × 256FFHQ 256 × 256
FID↓sFID↓FID↓sFID↓FID↓sFID↓FID↓sFID↓
Full Prec.32/322.987.094.1210.898.7410.169.368.67
PTQ4DM* [41]4/324.837.944.9213.9413.6714.7211.7412.18
Q-Diffusion† [23]4/324.207.664.5511.9011.0912.0011.6010.30
PTQD* [10]4/324.427.884.6713.6811.0612.2112.0111.12
TFMQ-DM (Ours)4/323.60 (-0.60)7.61 (-0.05)4.07 (-0.48)11.41 (-0.49)8.74 (-2.32)10.18 (-1.82)9.89 (-1.71)9.06 (-1.24)
PTQ4DM* [41]8/84.759.594.8013.4814.4215.0610.7311.65
Q-Diffusion† [23]8/84.518.174.4112.2312.8514.1610.8710.01
PTQD [10]8/83.759.894.89*14.89*12.76*13.54*10.69*10.97*
TFMQ-DM (Ours)8/83.14 (-0.61)7.26 (-0.91)4.01 (-0.40)10.98 (-1.25)8.71 (-4.05)10.20 (-3.34)9.46 (-1.23)8.73 (-1.28)
PTQ4DM [41]4/820.7254.304.97*14.87*17.08*17.48*11.83*12.91*
Q-Diffusion† [23]4/86.4017.934.6613.9415.5516.8611.4511.15
PTQD [10]4/85.9415.165.10*13.23*15.47*17.38*11.42*11.43*
TFMQ-DM (Ours)4/83.68 (-2.26)7.65 (-7.51)4.14 (-0.52)11.46 (-1.77)8.76 (-6.71)10.26 (-6.60)9.97 (-1.45)9.14 (-2.01)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.36, + 0.472, + 0.402 + ], + "angle": 0, + "content": "Table 3. Quantization results for unconditional image generation with class-conditional image generation with LDM-8 on ImageNet \\(256 \\times 256\\)." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.413, + 0.469, + 0.626 + ], + "angle": 0, + "content": "
MethodsBits (W/A)ImageNet 256 × 256
IS↑FID↓sFID↓
Full Prec.32/32235.6410.917.67
PTQ4DM [41]4/32---
Q-Diffusion* [23]4/32213.5611.878.76
PTQD† [10]4/32201.7811.659.06
TFMQ-DM (Ours)4/32223.81 (+10.25)10.50 (-1.15)7.98 (-0.78)
PTQ4DM [41]8/8161.7512.59-
Q-Diffusion* [23]8/8187.6512.809.87
PTQD [10]8/8153.9211.948.03
TFMQ-DM (Ours)8/8198.86 (+11.21)10.79 (-1.15)7.65 (-0.38)
PTQ4DM [41]4/8---
Q-Diffusion* [23]4/8212.5110.6814.85
PTQD [10]4/8214.7310.4012.63
TFMQ-DM (Ours)4/8221.82 (+7.09)10.29 (-0.11)7.35 (-5.28)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.738 + ], + "angle": 0, + "content": "ready achieved satisfactory results on this dataset. Nonetheless, our method still approaches the performance of the FP model more closely compared to existing methods. Besides the experiments on LDM, We have also conducted experiments with DDPM on CIFAR-10 \\(32 \\times 32\\), which can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Class-conditional image generation. On the ImageNet \\(256 \\times 256\\) dataset, we employed a denoising process with 20 iterations, setting eta and CFG to 0.0 and 3.0, respectively. Compared to PTQD, our method achieved a FID reduction of 1.15 on both w4a32 and w8a8. Simultaneously, in the w4a8 setting, sFID decreased by 5.28. Under the same conditions, we observed an improvement of over 7 in IS. Particularly noteworthy is that, across various quantization settings, our method consistently achieved lower FID compared to the FP model." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.36, + 0.892, + 0.389 + ], + "angle": 0, + "content": "Table 4. Quantization results for text-guided image generation with Stable-Diffusion on MS-COCO captions." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.4, + 0.891, + 0.535 + ], + "angle": 0, + "content": "
MethodsBits (W/A)MS-COCO
FID↓sFID↓CLIP↑
Full Prec.32/3213.1519.310.3146
Q-Diffusion† [23]4/3213.5819.500.3143
TFMQ-DM (Ours)4/3213.21 (-0.37)19.03 (-0.47)0.3144 (+0.0001)
Q-Diffusion† [23]8/813.3120.540.3134
TFMQ-DM (Ours)8/813.09 (-0.22)19.91 (-0.63)0.3134 (+0.0000)
Q-Diffusion† [23]4/814.4920.430.3121
TFMQ-DM (Ours)4/813.36 (-1.13)20.14 (-0.29)0.3128 (+0.0007)
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.895, + 0.773 + ], + "angle": 0, + "content": "Text-guided image generation. In this experiment, we sample high-resolution images of \\(512 \\times 512\\) pixels with 50 denoising steps and fix CFG to the default 7.5 in Stable Diffusion as the trade-off between sample quality and diversity. In Tab. 4, compared to the current SOTA Q-Diffusion, our approach achieves an FID reduction of 1.13 on w4a8. Simultaneously, our FID on w8a8 and sFID on w4a32 are even lower than those of the full precision model. However, existing metrics fail to adequately assess the semantic consistency of generated images. Nevertheless, based on the images generated in the appendix, our method produces higher-quality images with more realistic details, better demonstrating semantic information. Furthermore, our generated images closely approximate the effects of FP model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.786, + 0.655, + 0.803 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To evaluate the effectiveness of each proposed method, we perform a thorough ablation study on the LSUN-Bedrooms \\(256 \\times 256\\) dataset with w4a8 quantization, utilizing the LDM-4 model with a DDIM sampler, as outlined in Tab. 5. We begin the assessment with a baseline BRECQ [24], which shows ineffective in denoising images when oper" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7368" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.117 + ], + "angle": 0, + "content": "Table 5. The effect of different methods proposed in the paper on LSUN-Bedrooms \\(256 \\times 256\\)." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.128, + 0.468, + 0.219 + ], + "angle": 0, + "content": "
MethodsBits (W/A)FID↓sFID↓
Full Prec.32/322.987.09
BRECQ [24] (Baseline)4/89.3622.73
+TIAR4/84.849.29
+FSC4/86.0711.31
+TFMQ-DM (TIAR + FSC)4/83.68 (-5.68)7.65 (-15.08)
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.226, + 0.468, + 0.259 + ], + "angle": 0, + "content": "ating with 4-bit quantization. Additional ablation experiments can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.269, + 0.471, + 0.391 + ], + "angle": 0, + "content": "Effect of TIAR. It can be observed that, compared to the baseline, our TIAR method reduces FID and sFID by 4.52 and 13.44, respectively. Furthermore, As shown in Fig. 4, our method's temporal feature disturbance is significantly far weaker than existing SOTA PTQD. This indicates the effectiveness of our method in maintaining temporal information contained in temporal features. Further details of the effects are presented in the appendix." + }, + { + "type": "image", + "bbox": [ + 0.133, + 0.404, + 0.416, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.614, + 0.465, + 0.63 + ], + "angle": 0, + "content": "Figure 4. Temporal feature errors across different PTQ methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.47, + 0.799 + ], + "angle": 0, + "content": "Effect of FSC. Our FSC method has also achieved remarkably positive results. Compared to the baseline, it reduces FID and sFID by 3.29 and 11.42, respectively. Compared to PTQD, from Fig. 4, our method's temporal feature error is significantly smaller than that of PTQD. Since we employ layer-wise quantization for activations, this introduces less than one percent of additional parameters. The inference time overhead incurred by switching different step sizes and zero points during multiple time-steps inference is negligible, as detailed in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Efficiency of FSC. For FSC, there are various methods to evaluate the range of activations to determine the optimal step size. We try several methods and assess the GPU time consumed during calibration, as detailed in Tab. 6. Since the improvements achieved by these methods in model performance are similar, we opt for the simplest and most efficient" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 6. Different calibration methods for FTFC. We note \\( T \\) as calibration GPU hours in the table." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.129, + 0.89, + 0.321 + ], + "angle": 0, + "content": "
TSC MethodsBits (W/A)FID↓sFID↓T (hours)
FIAR8/322.847.082.18
+LSQ [5]8/83.177.182.48
+KL-divergence8/83.277.3219.67
+Percentile8/83.347.4112.00
+MSE8/83.127.128.89
+Min-max [32]8/83.14 (+0.02)7.26 (+0.14)0.12 (-2.36)
FIAR4/323.047.162.20
+LSQ [5]4/83.697.482.57
+KL-divergence4/83.947.4219.65
+Percentile4/83.748.0212.04
+MSE4/83.627.488.89
+Min-max [32]4/83.68 (+0.06)7.65 (+0.23)0.12 (-2.45)
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.329, + 0.892, + 0.435 + ], + "angle": 0, + "content": "Min-max [32] method as our specific calibration strategy, striking a balance between calibration time and effectiveness. Notably, we have found that PTQD and Q-Diffusion cost 4.68 and 5.29 GPU hours in their PTQ methods under w4a8 quantization on LSUN-Bedrooms \\(256 \\times 256\\), respectively. However, our framework only spends 2.32 GPU hours (as shown in Tab. 6)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.893, + 0.572 + ], + "angle": 0, + "content": "Furthermore, for the only learning-based method in the table, LSQ [5], which is one of the most commonly used methods in previous works [10, 23, 41], we observe that it did not outperform other methods and, in some cases, performs even worse. We speculate that the main reason might be that, for a fixed time-step, the calibration data used to determine the quantization parameters is relatively limited compared to all calibration data. In such cases, LSQ may struggle to learn an optimal set of parameters." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.585, + 0.864, + 0.603 + ], + "angle": 0, + "content": "6. Conclusion & Discussions of Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.892, + 0.761 + ], + "angle": 0, + "content": "This research explores the application of quantization for accelerating diffusion models. In this work, we identify a novel and significant problem, namely temporal feature disturbance, in the quantization of diffusion models. We conducted a detailed analysis of the root causes of this problem and introduced our TFMQ-DM quantization framework. In 4-bit quantization on extensive datasets and different diffusion models, this framework exhibits minimal performance degradation compared to the FP model and speedup quantization time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.892, + 0.902 + ], + "angle": 0, + "content": "However, besides temporal features, we have also found the textual features introduced in Stable Diffusion have a physical meaning and influence the generation effect. Nevertheless, these textual features are not taken into consideration in current methods. Secondly, temporal feature maintenance is suitable for both PTQ and QAT scenarios. But in this study, our main focus is on the PTQ setting and achieving 4-bit quantization. The extension to the QAT setting should be conducted in the future to achieve lower-bit quantization and further performance enhancements." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7369" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.119, + 0.176, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.47, + 0.199 + ], + "angle": 0, + "content": "[1] Yash Bhalgat, Jinwon Lee, Markus Nagel, Tijmen Blankevoort, and Nojun Kwak. \\(\\mathrm{Lsq}+\\): Improving low-bit quantization through learnable offsets and better initialization, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[2] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[3] Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In CVPR, pages 12413-12422, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.472, + 0.452 + ], + "angle": 0, + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.47, + 0.493 + ], + "angle": 0, + "content": "[5] Steven K. Esser, Jeffrey L. McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S. Modha. Learned step size quantization. In ICLR, 2020. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.471, + 0.563 + ], + "angle": 0, + "content": "[6] Giulio Franzese, Simone Rossi, Lixuan Yang, Alessandro Finamore, Dario Rossi, Maurizio Filippone, and Pietro Michiardi. How much is enough? a study on diffusion times in score-based generative models. arXiv preprint arXiv:2206.05173, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[7] Ruihao Gong, Xianglong Liu, Shenghu Jiang, Tianxiang Li, Peng Hu, Jiazhen Lin, Fengwei Yu, and Junjie Yan. Differentiable soft quantization: Bridging full-precision and low-bit neural networks. In ICCV, pages 4852–4861, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.677, + 0.47, + 0.731 + ], + "angle": 0, + "content": "[9] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.733, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[10] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. In NeurIPS, 2023. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[11] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.145, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[15] Itay Hubara, Yury Nahshan, Yair Hanani, Ron Banner, and Daniel Soudry. Improving post training neural quantization: Layer-wise calibration and integer programming. arXiv preprint arXiv:2006.10518, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.18, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[16] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In CVPR, pages 2704-2713, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.251, + 0.892, + 0.293 + ], + "angle": 0, + "content": "[17] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.296, + 0.892, + 0.335 + ], + "angle": 0, + "content": "[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.338, + 0.892, + 0.378 + ], + "angle": 0, + "content": "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.382, + 0.892, + 0.409 + ], + "angle": 0, + "content": "[20] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.412, + 0.892, + 0.452 + ], + "angle": 0, + "content": "[21] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. arXiv preprint arXiv:2106.00132, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[22] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.486, + 0.892, + 0.54 + ], + "angle": 0, + "content": "[23] Xiuyu Li, Long Lian, Yijiang Liu, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In ICCV, 2023. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.543, + 0.892, + 0.597 + ], + "angle": 0, + "content": "[24] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. BRECQ: pushing the limit of post-training quantization by block reconstruction. In ICLR, 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.6, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.711 + ], + "angle": 0, + "content": "[26] Yang Lin, Tianyu Zhang, Peiqin Sun, Zheng Li, and Shuchang Zhou. Fq-vit: Post-training quantization for fully quantized vision transformer. In *IJCAI*, pages 1173–1179, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.715, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[27] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In ICLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.758, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[28] Christos Louizos, Matthias Reisser, Tijmen Blankevoort, Efstratos Gavves, and Max Welling. Relaxed quantization for discretized neural networks. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[29] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Zhaoyang Lyu, Xudong Xu, Ceyuan Yang, Dahua Lin, and Bo Dai. Accelerating diffusion models via early stop of the diffusion process. arXiv preprint arXiv:2205.12524, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7370" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[31] Markus Nagel, Rana Ali Amjad, Mart van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. In ICML, pages 7197-7206, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[32] Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bondarenko, Mart van Baalen, and Tijmen Blankevoort. A white paper on neural network quantization, 2021. 1, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[34] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.319, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[35] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.402 + ], + "angle": 0, + "content": "[36] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. In ECCV, pages 525-542, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[37] Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. Fastspeech 2: Fast and high-quality end-to-end text to speech, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[40] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[41] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In CVPR, 2023. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[42] Jonathan Shen, Ruoming Pang, Ron J. Weiss, Mike Schuster, Navdeep Jaitly, Zongheng Yang, Zhifeng Chen, Yu Zhang, Yuxuan Wang, RJ Skerry-Ryan, Rif A. Saurous, Yannis Agiomyrgiannakis, and Yonghui Wu. Natural tts synthesis by conditioning wavenet on mel spectrogram predictions, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[43] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for diffusion models. In NeurIPS, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[45] Changyuan Wang, Ziwei Wang, Xiuwei Xu, Yansong Tang, Jie Zhou, and Jiwen Lu. Towards accurate data-free quantization for diffusion models. arXiv preprint arXiv:2305.18723, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[46] Peisong Wang, Qiang Chen, Xiangyu He, and Jian Cheng. Towards accurate post-training network quantization via bitsplit and stitching. In ICML, pages 9847-9856, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[47] Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. Qdrop: Randomly dropping quantization for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "extremely low-bit post-training quantization. In ICLR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[48] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop, 2016. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[49] Luoming Zhang, Yefei He, Zhenyu Lou, Xin Ye, Yuxing Wang, and Hong Zhou. Root quantization: a self-adaptive supplement ste. Applied Intelligence, 53(6):6266-6275, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[50] Qinsheng Zhang, Molei Tao, and Yongxin Chen. gddim: Generalized denoising diffusion implicit models. arXiv preprint arXiv:2206.05564, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[51] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. Opt: Open pre-trained transformer language models, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[52] Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Truncated diffusion probabilistic models. stat, 1050:7, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[53] Bohan Zhuang, Chunhua Shen, Mingkui Tan, Lingqiao Liu, and Ian Reid. Towards effective low-bitwidth convolutional neural networks. In CVPR, pages 7920-7928, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "7371" + } + ] +] \ No newline at end of file diff --git a/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/a91d56f7-042a-42e5-9c77-ba0dbaac2109_origin.pdf b/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/a91d56f7-042a-42e5-9c77-ba0dbaac2109_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..961ddf54c197955b19414ecce6d38b9b87933c18 --- /dev/null +++ b/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/a91d56f7-042a-42e5-9c77-ba0dbaac2109_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4eba4bde740e658fbebadf8c772619cd6f5beb4fdfd28fec626442213bd2f098 +size 8028919 diff --git a/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/full.md b/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d3748134c39a9c09dc2cfac0f23127148012994b --- /dev/null +++ b/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/full.md @@ -0,0 +1,305 @@ +# TFMQ-DM: Temporal Feature Maintenance Quantization for Diffusion Models + +Yushi Huang $^{1,2*}$ , Ruihao Gong $^{1,2*}$ , Jing Liu $^{2,3}$ , Tianlong Chen $^{4}$ , Xianglong Liu $^{1*}$ $^{1}$ Beihang University $^{2}$ SenseTime Research $^{3}$ Monash University $^{4}$ UT Austin + +{huangyushi, gongruihao, liujing5} @sentetime.com, tianlong.chen@utexas.edu, xlliu@buaa.edu.cn + +# Abstract + +The Diffusion model, a prevalent framework for image generation, encounters significant challenges in terms of broad applicability due to its extended inference times and substantial memory requirements. Efficient Post-training Quantization (PTQ) is pivotal for addressing these issues in traditional models. Different from traditional models, diffusion models heavily depend on the time-step $t$ to achieve satisfactory multi-round denoising. Usually, $t$ from the finite set $\{1,\dots,T\}$ is encoded to a temporal feature by a few modules totally irrespective of the sampling data. However, existing PTQ methods do not optimize these modules separately. They adopt inappropriate reconstruction targets and complex calibration methods, resulting in a severe disturbance of the temporal feature and denoising trajectory, as well as a low compression efficiency. To solve these, we propose a Temporal Feature Maintenance Quantization (TFMQ) framework building upon a Temporal Information Block which is just related to the time-step $t$ and unrelated to the sampling data. Powered by the pioneering block design, we devise temporal information aware reconstruction (TIAR) and finite set calibration (FSC) to align the full-precision temporal features in a limited time. Equipped with the framework, we can maintain the most temporal information and ensure the end-to-end generation quality. Extensive experiments on various datasets and diffusion models prove our state-of-the-art results. Remarkably, our quantization approach, for the first time, achieves model performance nearly on par with the full-precision model under 4-bit weight quantization. Additionally, our method incurs almost no extra computational cost and accelerates quantization time by $2.0\times$ on LSUN-Bedrooms $256\times 256$ compared to previous works. Our code is publicly available at https://github.com/ModelTC/TFMQ-DM. + +# 1. Introduction + +Generative modeling plays a crucial role in machine learning, particularly in applications like image [13, 14, 17, 38, 44], voice [37, 42], and text synthesis [2, 51]. Diffusion models have showcased impressive capabilities in producing high-quality samples across diverse domains. In comparison to generative adversarial networks (GANs) [8] and variational autoencoders (VAEs) [20], diffusion models successfully sidestep issues such as model collapse and posterior collapse, resulting in a more stable training process. However, the substantial computational cost poses a critical bottleneck hampering the widespread adoption of diffusion models. Furthermore, the computational cost for diffusion models can be attributed to two primary factors. First, these models typically require hundreds of denoising steps to generate images, rendering the procedure considerably slower than that of GANs. Prior efforts [21, 27, 29, 44] have addressed this challenge by seeking shorter and more efficient sampling trajectories, thereby reducing the number of denoising steps. Second, the substantial network architecture of diffusion models demands considerable time and memory resources, particularly for foundational models pre-trained on large-scale datasets, e.g., LDM [38] and Stable Diffusion. Our work aims to tackle the latter challenge, focusing on the compression of diffusion models. + +Quantization is currently the most widely used method [1, 5, 7, 31, 32] for compressing models by mapping high-precision floating-point numbers to low-precision numbers. Among different quantization methods, Posttraining quantization (PTQ) [15, 31, 47] incurs lower overhead and is more user-friendly without the need for retraining or fine-tuning. While PTQ on conventional models has undergone extensive study [7, 24, 31, 47], its application to diffusion models has shown huge performance degradation, especially under low-bit settings. For instance, Q-Diffusion [23] exhibits severe accuracy drop on some datasets [48] under 4-bit quantization. We believe the reason they fail to achieve better results is that they all overlook the sampling data independence and uniqueness of temporal features, which are generated from time-step $t$ through a few modules, used to control the denoising trajectory in dif + +fusion models. Furthermore, we observe that temporal feature disturbance significantly impacts model performance in the aforementioned methods. + +To tackle temporal feature disturbance, we first find that the modules generating temporal features are independent of the sampling data and define the whole modules as the Temporal Information Block. All existing methods do not separately optimize this block during the quantization process, causing temporal features to overfit to limited calibration data. On the other hand, since the maximum time-step for denoising is a finite positive integer, the temporal feature and the activations during its generation form a finite set. The optimal approach is to optimize each element in this set individually. Based on these observations and analyses, we propose a novel quantization reconstruction approach, temporal information aware reconstruction (TIAR), specifically optimizing for temporal features. It aims to reduce temporal feature loss as the optimization objective while isolating the network's components related to sampled data from the generation of temporal features during calibration. Furthermore, we also introduce a calibration strategy, finite set calibration (FSC), for the finite set of temporal features and activations during its generation. This strategy employs different quantization parameters for activations corresponding to different time-steps. Moreover, the calibration speed of this method is faster than existing mainstream methods [1, 5], for example, we speedup quantization time by $2.0 \times$ on LSUN-Bedrooms $256 \times 256$ dataset, yet the strategy incurs negligible inference and storage overhead. The overview of our framework can be seen in Fig. 1. In summary, our contributions are as follows: + +- We discover that existing quantization methods suffer from temporal feature disturbance, disrupting the denoising trajectory of diffusion models and significantly affecting the quality of generated images. +- We reveal that the disturbance comes from two aspects: inappropriate reconstruction target and unaware of finite activations. Both inducements ignore the special characteristics of time information-related modules. +- An advanced framework (TFMQ-DM) is proposed, consisting of temporal information aware reconstruction (TIAR) for weight quantization and finite set calibration (FSC) for activation quantization. Both are based on a Temporal Information Block specially devised for diffusion models. +- Extensive experiments on various datasets show that our novel framework achieves a new state-of-the-art result in PTQ of diffusion models, especially under 4-bit weight quantization, and significantly accelerates quantization time. For some hard tasks, e.g., CelebA-HQ $256 \times 256$ , our method reduces the FID score by 6.71 (images in appendix). + +# 2. Related Work + +# 2.1. Efficient Diffusion Models + +There are diverse perspectives to accelerate the inference of diffusion model, e.g., retraining-based model design [3, 6, 30, 52] and retraining-free sampler strategy [21, 44, 50]. However, the retraining-based method proves resource-intensive and time-consuming. The efficient samplers can reduce sampling iterations but still suffer from diffusion models' extensive parameters and computational complexity. In this paper, we focus on diminishing the time and memory overhead of the single-step denoising process using low-bit quantization in a training-free manner, a method orthogonal to previous speedup techniques. + +# 2.2. Model Quantization + +Quantization is a predominant technique for minimizing storage and computational costs. It can be categorized into quantization-aware training (QAT) [7, 16, 28, 49, 53] and post-training quantization (PTQ) [15, 24, 26, 31, 47]. QAT requires intensive model training with substantial data and computational demands. Correspondingly, PTQ compresses models without re-training, making it a preferred method due to its minimal data requirements and easy deployment on real hardware. In PTQ, high-precision values are mapped into discrete levels using uniform quantization expressed as: + +$$ +\hat {x} = \Phi \left(\left\lfloor \frac {x}{s} \right\rceil + z, 0, 2 ^ {b} - 1\right), \tag {1} +$$ + +where $x$ represents floating-point data, $\hat{x}$ is the quantized value, $s$ is the quantization step size, $z$ is the zero offset, and $b$ is the target bit-width. The clamp function $\Phi(\cdot)$ clips the rounded value $\left\lfloor \frac{x}{s} \right\rfloor + z$ within the range of $[0, 2^b - 1]$ . However, naive quantization may lead to accuracy degradation, especially for low-bit quantization. Recent studies [15, 24, 31, 46, 47] have explored innovative strategies based on reconstruction to preserve model performance after low-bit quantization. + +In contrast, the iterative denoising process in diffusion models presents new challenges for PTQ in comparison to traditional models. PTQ4DM [41] represents the initial attempt to quantize diffusion models to 8-bit, albeit with limited experiments and lower resolutions. Conversely, Q-Diffusion [23] achieves enhanced performance and is evaluated on a broader dataset range. Moreover, PTQD [10] eliminates quantization noise through correlated and residual noise correction. Notably, traditional single-time-step PTQ calibration methods are unsuitable for diffusion models due to significant activation distribution changes with each time-step [23, 41, 43, 45]. ADP-DM [45] proposes group-wise quantization across time-steps for diffusion models, and TDQ [43] introduces distinct quantization parameters for different time-steps. However, all of the + +above works overlook the specificity of temporal features. To address temporal feature disturbance in the aforementioned works, our study delves into the inducements of the phenomenon and introduces a novel reconstruction and calibration framework, significantly enhancing quantized diffusion model performance. + +# 3. Preliminaries + +Diffusion models. Diffusion models [14, 44] iteratively add Gaussian noise with a variance schedule $\beta_{1},\ldots ,\beta_{T}\in$ $(0,1)$ to data $\mathbf{x}_0\sim q(\mathbf{x})$ for $T$ times as sampling process, resulting in a sequence of noisy samples $\mathbf{x}_1,\dots ,\mathbf{x}_T$ In DDPMs [14], the former sampling process is a Markov chain, taking the form: + +$$ +q \left(\mathbf {x} _ {1: T} \mid \mathbf {x} _ {0}\right) = \prod_ {t = 1} ^ {T} q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {t - 1}\right), \tag {2} +$$ + +$$ +q (\mathbf {x} _ {t} | \mathbf {x} _ {t - 1}) = \mathcal {N} (\mathbf {x} _ {t}; \sqrt {\alpha_ {t}} \mathbf {x} _ {t - 1}, \beta_ {t} \mathbf {I}), +$$ + +where $\alpha_{t} = 1 - \beta_{t}$ . Conversely, the denoising process removes noise from a sample from Gaussian noise $\mathbf{x}_T\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ to gradually generate high-fidelity images. Nevertheless, due to the unavailability of the true reverse conditional distribution $q(\mathbf{x}_{t - 1}|\mathbf{x}_t)$ , diffusion models approximate it via variational inference by learning a Gaussian distribution $p_{\theta}(\mathbf{x}_{t - 1}|\mathbf{x}_t) = \mathcal{N}(\mathbf{x}_{t - 1};\boldsymbol {\mu}_\theta (\mathbf{x}_t,t),\boldsymbol {\Sigma}_\theta (\mathbf{x}_t,t))$ , the $\boldsymbol{\mu}_{\theta}$ can be derived by reparameterization trick as follows: + +$$ +\boldsymbol {\mu} _ {\theta} \left(\mathbf {x} _ {t}, t\right) = \frac {1}{\sqrt {\alpha_ {t}}} \left(\mathbf {x} _ {t} - \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \boldsymbol {\epsilon} _ {\theta} \left(\mathbf {x} _ {t}, t\right)\right), \tag {3} +$$ + +where $\bar{\alpha}_{t} = \prod_{i=1}^{t} \alpha_{i}$ and $\epsilon_{\theta}(\cdot)$ is a noise estimation model. The variance $\Sigma_{\theta}(\mathbf{x}_{t}, t)$ can be either learned [33] or fixed to a constant schedule [14] $\sigma_{t}$ . When employing the latter method, $\mathbf{x}_{t-1}$ can be expressed as: + +$$ +\mathbf {x} _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\mathbf {x} _ {t} - \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \boldsymbol {\epsilon} _ {\theta} (\mathbf {x} _ {t}, t)\right) + \sigma_ {t} \mathbf {z}, \tag {4} +$$ + +where $\mathbf{z}\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ + +Reconstruction on diffusion models. UNet [39], the predominant model employed as $\epsilon_{\theta}(\cdot)$ in Eq. 4 to predict Gaussian noise, can be divided into blocks that incorporate residual connections (such as Residual Bottleneck Blocks or Transformer Blocks [35]) and the remaining layers. Numerous PTQ approaches for diffusion models are grounded in layer/block-wise reconstruction [10, 23, 41, 43] to obtain optimal quantization parameters. For example, in the Residual Bottleneck Block, this approach typically minimizes the following loss function as its optimization objective: + +$$ +\mathcal {L} _ {i} = \left\| f _ {i} (\cdot) - \widehat {f} _ {i} (\cdot) \right\| _ {F} ^ {2}, \tag {5} +$$ + +where $\| \cdot \|_F^2$ denotes the Frobenius norm. The function $f_i(\cdot)$ represents the $i^{\text{th}}$ Residual Bottleneck Block, and $\widehat{f}_i(\cdot)$ is its quantized counterpart. Furthermore, in the ensuing sections, we use $n$ to denote the total number of Residual Bottleneck Blocks in a single diffusion model. + +Temporal feature in diffusion models. Seeing from Fig. 1 (a), time-step $t$ is encoded with time embed $^{1}$ and then passes through the embedding layer $^{2}$ in each Residual Bottleneck Block, resulting in a series of unique activations. In this paper, we denote these activations as temporal features. Notably, temporal features are independent of $\mathbf{x}_t$ and unrelated to other temporal features from different time-steps. To enhance clarity, we simplify our notation as follows: we represent time embed as $h(\cdot)$ , the embedding layer in the $i^{\text{th}}$ Residual Bottleneck Block as $g_i(\cdot)$ , and denote the $i^{\text{th}}$ temporal feature at time-step $t$ as $\mathbf{emb}_{t,i}$ . Moreover, as illustrated in Fig. 1 (a), the relationship is explicitly expressed by the equation: + +$$ +\text {T e m p o r a l} \quad \mathbf {e m b} _ {t, i} = g _ {i} (h (t)) \tag {6} +$$ + +Additionally, we have found that temporal features play a crucial role in the context of the diffusion model, holding unique and substantial physical implications. These features encompass temporal information that signifies the current image's temporal position along the denoising trajectory. Within the UNet structure, each time-step transforms into these temporal features, thereby controlling the denoising trajectory by applying them to the features of images generated at each iteration. + +# 4. TFMQ for Diffusion Models + +We present our novel PTQ framework in this section. We first observe temporal disturbance in previous methods in Sec. 4.1 and then analyze the inducements in Sec. 4.2. Finally, we propose our solutions in Sec. 4.3. + +# 4.1. Temporal Feature Disturbance + +Based on Sec. 3, we investigate the impact of previous PTQ works on temporal features, and we identify the phenomenon of temporal feature disturbance, which significantly deteriorates the quality of generated images. + +Temporal feature error. We thoroughly analyze temporal feature variations before and after the quantization of embedding layers and time embed in the Stable Diffusion model ( $T = 50, i = 11$ ). Prior to this analysis, we introduce the temporal feature error as defined by: + +$$ +\cos \left(\widehat {\mathbf {e m b}} _ {t, i}, \widehat {\mathbf {e m b}} _ {t, i}\right), \tag {7} +$$ + +![](images/87b6455d1bf6e3904f1661eca93aa8229ab12114fb2c24f4ad02d0f7cadc3d67.jpg) + +![](images/01437bc9a586d1dad1f66ae44a29260ebd857443fb7eb429f28491d61385d568.jpg) +Figure 1. Overview of the proposed Temporal Feature Maintenance Quantization. (a) Temporal Feature $\mathbf{emb}_{t,i}$ , belonging to a finite set representing temporal information, has been overlooked in previous works due to inappropriate reconstruction targets (box with a solid line). (b) This oversight leads to a severe disturbance for $\mathbf{emb}_{t,i}$ and results in the mismatch of crucial temporal information for the diffusion model's generation, causing a deviation in the denoising trajectory and a significant drop in accuracy. (c) Based on these analyses, we introduce a Temporal Information Block that exclusively correlates with the time-step $t$ . Leveraging this $\mathbf{x}_t$ -unrelated block, we enable Temporal Information Aware Reconstruction and Finite Set Calibration (utilizing the finite number of $t$ ). This approach achieves the maintenance of temporal features and yields state-of-the-art results. + +![](images/afc432d837c9c40687168e5dcb02f97600585a64ffcb5e6266585921821bfed2.jpg) + +where $\cos (\cdot)$ denotes cosine similarity, and $\widehat{\mathbf{emb}_{t,i}}$ signifies the temporal feature corresponding to $\mathbf{emb}_{t,i}$ in the quantized model. As illustrated in Fig. 2 (Left), quantization induces notable temporal feature errors. We term this phenomenon, characterized by substantial temporal feature errors within diffusion models, as temporal feature disturbance. + +Temporal information mismatch. Temporal feature disturbance alters the original embedded temporal information. Specifically, $\mathbf{emb}_{t,i}$ is intended to correspond to timestep $t$ . However, due to significant errors, the quantized model's $\widehat{\mathbf{emb}}_{t,i}$ is no longer accurately associated with $t$ , resulting in what we term as temporal information mismatch: + +$$ +t \leftarrow \mathbf {e m b} _ {t, i}, \quad t \nleftarrow \widehat {\mathbf {e m b}} _ {t, i}. \tag {8} +$$ + +Furthermore, as depicted in Fig. 2 (Right), we even observe a pronounced temporal information mismatch. Specifically, the temporal feature generated by the quantized model at time-step $t$ exhibits a divergence from that of the full-precision model at the corresponding time-step. Instead, it tends to align more closely with the temporal feature corresponding to $t + \delta_t$ , importing wrong temporal information from $t + \delta_t$ . + +Trajectory deviation. Temporal information mismatch delivers wrong temporal information, therefore, causing a deviation in the corresponding temporal position of the image within the denoising trajectory, ultimately leading to: + +$$ +\mathbf {x} _ {t} \neq \mathbf {x} _ {t - 1}, \tag {9} +$$ + +where we apply disrupted temporal features to the model. Evidently, as the deviation in the denoising trajectory accumulates with the increase in the number of denoising + +![](images/9a683213a2eeebc8b675f563ced683121f11e6445e523c26ff4265a00d6f6dd2.jpg) +Figure 2. (Left) Temporal feature disturbance. The inflection points serve as indicators of temporal feature errors at different time-steps, and they highlight the significant phenomenon of temporal feature disturbance. (Right) Temporal information mismatch. The coordinates of the inflection points on the blue curve can be denoted as $(t,t + \delta_{t,i})$ . It indicates $\mathbf{emb}_{t + \delta_{t,i},i}$ exhibits the highest similarity with $\mathbf{emb}_{t,i}$ . + +![](images/677aa6975168f144167cfeb2bc4f955dd11a8e1777b0b6c487ff1cd255bc00ce.jpg) + +iterations, the final generated image struggles to align with $\mathbf{x}_0$ . This evolution is illustrated in Fig. 3, where we maintain UNet excluding embedding layers and time embed in full precision. + +# 4.2. Inducement Analyses + +In this section, we explore the two inducements of temporal feature disturbance. For the purpose of clarity, in the subsequent sections, "reconstruction" specifically points to slight weight adjustment for minimal quantization error, while "calibration" specifically refers to activation calibration. + +Inappropriate reconstruction target. Previous PTQ works [10, 23, 43] have achieved remarkable progress on diffusion models. However, these existing methods overlook the temporal feature's independence and its dis + +![](images/1304d60b10183617411e1fa61cc9de06863ae7c71a1307ba4a64fe68af16fa04.jpg) +Figure 3. Denoising process of full-precision (Upper) and w4a8 quantized (Lower) Stable-Diffusion $(T = 50)$ under the same experiment settings and prompt: A man in the snow on a snow board. We represent $\{\mathbf{emb}_{t,i}\}_{i=0,\dots,n}$ and $\{\widehat{\mathbf{emb}}_{t,i}\}_{i=0,\dots,n}$ as $\mathbf{EMB}_t$ and $\widehat{\mathbf{EMB}_t}$ , respectively. Additionally, we denote $\widehat{\mathbf{x}}_t$ as $\mathbf{x}_t$ in the context of the quantized diffusion model. It is noteworthy that, in the quantized model employed here, to showcase the impact of temporal features, only the layers in Temporal Information Block are quantized and the components unrelated to the generation of temporal features are maintained in full precision. + +tive physical significance. During their reconstruction processes, there was a lack of optimization for the embedding layers; instead, a Residual Bottleneck Block of coarser granularity was selected as the reconstruction target. This method involves two potential factors causing temporal feature disturbance: + +- Optimize the objective as expressed in Eq. 5 to decrease the reconstruction loss of Residual Bottleneck Block, as opposed to direct reduction of temporal feature loss. +- During backpropagation of the reconstruction process, embedding layers independent from $\mathbf{x}_t$ are affected by $\mathbf{x}_t$ , resulting in an overfitting scenario on limited calibration data. + +To further validate our analyses, we respectively evaluate the FID [12] and sFID [40] for this reconstruction method, e.g., BRECQ [24] and the approach where we freeze the parameters of the embedding layers during the reconstruction of the Residual Bottleneck Block, initializing the embedding layers solely through Min-max [32] for comparison. As shown in Tab. 1, the Freeze strategy exhibits better results, which verify that embedding layers serve as their own optimization objective and maintain their independence of $\mathbf{x}_t$ can significance mitigate temporal feature disturbance, especially at low-bit. + +Table 1. FID and sFID on LSUN-Bedrooms $256 \times 256$ [48] for LDM-4 with 50000 sampling images. Prev represents BRECQ. Freeze denotes our trial. + +
MethodsBits (W/A)FID↓sFID↓
Full Prec.32/322.987.09
Prev8/87.5112.54
Freeze8/85.76 (-1.75)8.42 (-4.12)
Prev4/89.3622.73
Freeze4/87.08 (-2.28)16.82 (-5.91)
+ +Unaware of finite activations within $h(\cdot)$ and $g_{i}(\cdot)$ . We observe that, given $T$ as a finite positive integer, the set of all possible activation values for embedding layers and time embed is finite and strictly dependent on timesteps. Within this set, activations corresponding to the same layer display notable range variations across different timesteps (refer to the appendix). Previous methods [43, 45] mainly focus on finding the optimal calibration method for $\widehat{\mathbf{x}}_t$ -related network components. Moreover, akin to the first inducement, their calibration is directly towards the Residual Bottleneck Block, which proves suboptimal (refer to the appendix). However, based on the finite activations, we can employ calibration methods, especially for these time information-related activations, to better adapt to their range variations. + +# 4.3. Temporal Feature Maintenance + +To address the problem of temporal feature disturbance, we design a novel Temporal Information Block to maintain the temporal features. Based on the block, Temporal Information Aware Reconstruction and Finite Set Calibration are proposed to solve the two inducements analyzed above. + +Temporal Information Block. Based on the inducements, it is crucial to meticulously separate the reconstruction and calibration process for each embedding layer and Residual Bottleneck Block to enhance quantized model performance. Considering the unique structure of the UNet, we consolidate all embedding layers and time embed into a unified Temporal Information Block, which can be denoted as $\{g_{i}(h(\cdot))\}_{i = 0,\dots,n}$ (see Fig. 1 (c)). + +Temporal information aware reconstruction. Based on the Temporal Information Block, we propose Temporal information aware reconstruction (TIAR) to tackle the first inducement. The optimization objective for the block during + +the reconstruction process is as follows: + +$$ +\mathcal {L} _ {T I A R} = \sum_ {i = 0} ^ {n} \| g _ {i} (h (t)) - \widehat {g} _ {i} (\widehat {h} (t)) \| _ {F} ^ {2}, \tag {10} +$$ + +where $\widehat{h} (\cdot)$ and $\widehat{g}_i(\cdot)$ are quantized versions of $h(\cdot)$ and $g(\cdot)$ , respectively. With this reconstruction, weights are adjusted to pursue a minimal disturbance for temporal features. + +Finite set calibration. To address the challenge posed by the wide span of activations within a finite set for the second inducement, we propose Finite Set Calibration (FSC) for activation quantization. This strategy employs $T$ sets of quantization parameters for every activation within all the embedding layers and time embed, such as $\{(s_{T},z_{T}),\ldots ,(s_{1},z_{1})\}$ for activation $\mathbf{x}$ . In time-step $t$ , the quantization function for the $\mathbf{x}$ can be expressed as: + +$$ +\hat {\mathbf {x}} = \Phi \left(\left\lfloor \frac {\mathbf {x}}{s _ {t}} \right\rceil + z _ {t}, 0, 2 ^ {b} - 1\right). \tag {11} +$$ + +To be noted, the calibration target is also aligned with the output of the Temporal Information Block. For a more specific estimation method of activation ranges, since the solution space within the finite set is limited, we find the Minmax [32] method can achieve satisfactory results with high efficiency (more evidence in Sec. 5.3). + +# 5. Experiments + +# 5.1. Implementation details + +Models and datasets. In this section, we conduct image generation experiments to evaluate the proposed TFMQ-DM framework on various diffusion models: pixel-space diffusion model DDPM [14] for unconditional image generation, latent-space diffusion model LDM [38] for unconditional image generation and class-conditional image generation. We also apply our work to Stable Diffusion-v1-4 for text-guided image generation. In our experiments, We use seven standard benchmarks: CIFAR-10 $32 \times 32$ [22], LSUN-Bedrooms $256 \times 256$ [48], LSUN-Churches $256 \times 256$ [48], CelebA-HQ $256 \times 256$ [18], ImageNet $256 \times 256$ [4], FFHQ $256 \times 256$ [19] and MS-COCO [25]. + +Quantization settings. We use channel-wise quantization for weights and layer-wise quantization for activations, as it is a common practice. In our experimental setup, we employ BRECQ [24] and AdaRound [31]. Drawing from empirical insights derived from conventional model quantization practices [9, 36], we maintain the input and output layers of the model in full precision. Calibration sets, integral to our methodology, are generated through full-precision + +diffusion models, mirroring the approach outlined in Q-Diffusion [23]. Moreover, in weight quantization, we reconstruct quantized weights for 20k iterations with a minibatch size of 32 for DDPM and LDM, and 8 for Stable Diffusion. In activation quantization, we utilize EMA [16] to estimate the ranges of activations with a mini-batch size of 16 on all models. More details can be found in the appendix. + +Evaluation metrics. For each experiment, we evaluate the performance of diffusion models with Fréchet Inception Distance (FID) [12]. In the case of LDM and Stable Diffusion experiments, we also include sFID [40], which better captures spatial relationships than FID. For ImageNet and CIFAR-10 experiments, we additionally provide Inception Score (IS) [40] as a reference metric. Further, in the context of Stable Diffusion experiments, we extend our evaluation to include the compatibility of image-caption pairs, employing the CLIP score [11]. The ViT-B/32 is used as the backbone when computing the CLIP score. To ensure consistency in the reported outcomes, all results are derived from our implementation or from other papers, where experiments are conducted under conditions consistent with ours. More specifically, in the evaluation process of each experiment, we sample 50k images from DDPM or LDM, or 30k images from Stable-Diffusion. All experiments are conducted utilizing one H800 GPU and implemented with the PyTorch framework [34]. + +# 5.2. Main Results + +Unconditional image generation. In the experiments conducted on the LDM, we maintain the same experimental settings as presented in [38], including the number of steps, variance schedule, and classifier-free guidance scale (denoted by eta and CFG in the following, respectively). As shown in Tab. 2, the FID performance differences relative to the full precision (FP) model are all within 0.7 for all settings. Specifically, on the CelebA-HQ $256 \times 256$ dataset, our method exhibits a FID reduction of 6.71 and a sFID reduction of 6.60 in the w4a8 setting compared to the current state-of-the-art (SOTA). It is noticeable that existing methods, whether in 4-bit or 8-bit, show significant performance degradation when compared to the FP model on face datasets like CelebA-HQ $256 \times 256$ and FFHQ $256 \times 256$ , whereas our TFMQ-DM shows almost no performance degradation compared to the FP model. Importantly, our method achieves significant performance improvement on the LSUN-Bedrooms $256 \times 256$ in the w4a8 setting, with FID and sFID reductions of 2.26 and 7.51 compared to PTQD [10], respectively. Regarding LDM-8 on LSUN-Churches $256 \times 256$ , we attribute the moderate improvement, compared to other datasets. We believe that the use of the LDM-8 model with a downsampling factor of 8 may be more quantization-friendly. Existing methods have al + +Table 2. Quantization results for unconditional image generation with LDM-4 on LSUN-Bedrooms 256, FFHQ 256 and CelebA-HQ ${256} \times {256}$ ,LDM-8 on LSUN-Churches ${256} \times {256}$ . * represents our implementation according to open-source codes and †means directly rerunning open-source codes. + +
MethodsBits (W/A)LSUN-Bedrooms 256 × 256LSUN-Churches 256 × 256CelebA-HQ 256 × 256FFHQ 256 × 256
FID↓sFID↓FID↓sFID↓FID↓sFID↓FID↓sFID↓
Full Prec.32/322.987.094.1210.898.7410.169.368.67
PTQ4DM* [41]4/324.837.944.9213.9413.6714.7211.7412.18
Q-Diffusion† [23]4/324.207.664.5511.9011.0912.0011.6010.30
PTQD* [10]4/324.427.884.6713.6811.0612.2112.0111.12
TFMQ-DM (Ours)4/323.60 (-0.60)7.61 (-0.05)4.07 (-0.48)11.41 (-0.49)8.74 (-2.32)10.18 (-1.82)9.89 (-1.71)9.06 (-1.24)
PTQ4DM* [41]8/84.759.594.8013.4814.4215.0610.7311.65
Q-Diffusion† [23]8/84.518.174.4112.2312.8514.1610.8710.01
PTQD [10]8/83.759.894.89*14.89*12.76*13.54*10.69*10.97*
TFMQ-DM (Ours)8/83.14 (-0.61)7.26 (-0.91)4.01 (-0.40)10.98 (-1.25)8.71 (-4.05)10.20 (-3.34)9.46 (-1.23)8.73 (-1.28)
PTQ4DM [41]4/820.7254.304.97*14.87*17.08*17.48*11.83*12.91*
Q-Diffusion† [23]4/86.4017.934.6613.9415.5516.8611.4511.15
PTQD [10]4/85.9415.165.10*13.23*15.47*17.38*11.42*11.43*
TFMQ-DM (Ours)4/83.68 (-2.26)7.65 (-7.51)4.14 (-0.52)11.46 (-1.77)8.76 (-6.71)10.26 (-6.60)9.97 (-1.45)9.14 (-2.01)
+ +Table 3. Quantization results for unconditional image generation with class-conditional image generation with LDM-8 on ImageNet $256 \times 256$ . + +
MethodsBits (W/A)ImageNet 256 × 256
IS↑FID↓sFID↓
Full Prec.32/32235.6410.917.67
PTQ4DM [41]4/32---
Q-Diffusion* [23]4/32213.5611.878.76
PTQD† [10]4/32201.7811.659.06
TFMQ-DM (Ours)4/32223.81 (+10.25)10.50 (-1.15)7.98 (-0.78)
PTQ4DM [41]8/8161.7512.59-
Q-Diffusion* [23]8/8187.6512.809.87
PTQD [10]8/8153.9211.948.03
TFMQ-DM (Ours)8/8198.86 (+11.21)10.79 (-1.15)7.65 (-0.38)
PTQ4DM [41]4/8---
Q-Diffusion* [23]4/8212.5110.6814.85
PTQD [10]4/8214.7310.4012.63
TFMQ-DM (Ours)4/8221.82 (+7.09)10.29 (-0.11)7.35 (-5.28)
+ +ready achieved satisfactory results on this dataset. Nonetheless, our method still approaches the performance of the FP model more closely compared to existing methods. Besides the experiments on LDM, We have also conducted experiments with DDPM on CIFAR-10 $32 \times 32$ , which can be found in the appendix. + +Class-conditional image generation. On the ImageNet $256 \times 256$ dataset, we employed a denoising process with 20 iterations, setting eta and CFG to 0.0 and 3.0, respectively. Compared to PTQD, our method achieved a FID reduction of 1.15 on both w4a32 and w8a8. Simultaneously, in the w4a8 setting, sFID decreased by 5.28. Under the same conditions, we observed an improvement of over 7 in IS. Particularly noteworthy is that, across various quantization settings, our method consistently achieved lower FID compared to the FP model. + +Table 4. Quantization results for text-guided image generation with Stable-Diffusion on MS-COCO captions. + +
MethodsBits (W/A)MS-COCO
FID↓sFID↓CLIP↑
Full Prec.32/3213.1519.310.3146
Q-Diffusion† [23]4/3213.5819.500.3143
TFMQ-DM (Ours)4/3213.21 (-0.37)19.03 (-0.47)0.3144 (+0.0001)
Q-Diffusion† [23]8/813.3120.540.3134
TFMQ-DM (Ours)8/813.09 (-0.22)19.91 (-0.63)0.3134 (+0.0000)
Q-Diffusion† [23]4/814.4920.430.3121
TFMQ-DM (Ours)4/813.36 (-1.13)20.14 (-0.29)0.3128 (+0.0007)
+ +Text-guided image generation. In this experiment, we sample high-resolution images of $512 \times 512$ pixels with 50 denoising steps and fix CFG to the default 7.5 in Stable Diffusion as the trade-off between sample quality and diversity. In Tab. 4, compared to the current SOTA Q-Diffusion, our approach achieves an FID reduction of 1.13 on w4a8. Simultaneously, our FID on w8a8 and sFID on w4a32 are even lower than those of the full precision model. However, existing metrics fail to adequately assess the semantic consistency of generated images. Nevertheless, based on the images generated in the appendix, our method produces higher-quality images with more realistic details, better demonstrating semantic information. Furthermore, our generated images closely approximate the effects of FP model. + +# 5.3. Ablation Study + +To evaluate the effectiveness of each proposed method, we perform a thorough ablation study on the LSUN-Bedrooms $256 \times 256$ dataset with w4a8 quantization, utilizing the LDM-4 model with a DDIM sampler, as outlined in Tab. 5. We begin the assessment with a baseline BRECQ [24], which shows ineffective in denoising images when oper + +Table 5. The effect of different methods proposed in the paper on LSUN-Bedrooms $256 \times 256$ . + +
MethodsBits (W/A)FID↓sFID↓
Full Prec.32/322.987.09
BRECQ [24] (Baseline)4/89.3622.73
+TIAR4/84.849.29
+FSC4/86.0711.31
+TFMQ-DM (TIAR + FSC)4/83.68 (-5.68)7.65 (-15.08)
+ +ating with 4-bit quantization. Additional ablation experiments can be found in the appendix. + +Effect of TIAR. It can be observed that, compared to the baseline, our TIAR method reduces FID and sFID by 4.52 and 13.44, respectively. Furthermore, As shown in Fig. 4, our method's temporal feature disturbance is significantly far weaker than existing SOTA PTQD. This indicates the effectiveness of our method in maintaining temporal information contained in temporal features. Further details of the effects are presented in the appendix. + +![](images/079e6a510385420edf81ad2fe64a74293be6be63b07f66558ee8bf525313dc3c.jpg) +Figure 4. Temporal feature errors across different PTQ methods. + +Effect of FSC. Our FSC method has also achieved remarkably positive results. Compared to the baseline, it reduces FID and sFID by 3.29 and 11.42, respectively. Compared to PTQD, from Fig. 4, our method's temporal feature error is significantly smaller than that of PTQD. Since we employ layer-wise quantization for activations, this introduces less than one percent of additional parameters. The inference time overhead incurred by switching different step sizes and zero points during multiple time-steps inference is negligible, as detailed in the appendix. + +Efficiency of FSC. For FSC, there are various methods to evaluate the range of activations to determine the optimal step size. We try several methods and assess the GPU time consumed during calibration, as detailed in Tab. 6. Since the improvements achieved by these methods in model performance are similar, we opt for the simplest and most efficient + +Table 6. Different calibration methods for FTFC. We note $T$ as calibration GPU hours in the table. + +
TSC MethodsBits (W/A)FID↓sFID↓T (hours)
FIAR8/322.847.082.18
+LSQ [5]8/83.177.182.48
+KL-divergence8/83.277.3219.67
+Percentile8/83.347.4112.00
+MSE8/83.127.128.89
+Min-max [32]8/83.14 (+0.02)7.26 (+0.14)0.12 (-2.36)
FIAR4/323.047.162.20
+LSQ [5]4/83.697.482.57
+KL-divergence4/83.947.4219.65
+Percentile4/83.748.0212.04
+MSE4/83.627.488.89
+Min-max [32]4/83.68 (+0.06)7.65 (+0.23)0.12 (-2.45)
+ +Min-max [32] method as our specific calibration strategy, striking a balance between calibration time and effectiveness. Notably, we have found that PTQD and Q-Diffusion cost 4.68 and 5.29 GPU hours in their PTQ methods under w4a8 quantization on LSUN-Bedrooms $256 \times 256$ , respectively. However, our framework only spends 2.32 GPU hours (as shown in Tab. 6). + +Furthermore, for the only learning-based method in the table, LSQ [5], which is one of the most commonly used methods in previous works [10, 23, 41], we observe that it did not outperform other methods and, in some cases, performs even worse. We speculate that the main reason might be that, for a fixed time-step, the calibration data used to determine the quantization parameters is relatively limited compared to all calibration data. In such cases, LSQ may struggle to learn an optimal set of parameters. + +# 6. Conclusion & Discussions of Limitations + +This research explores the application of quantization for accelerating diffusion models. In this work, we identify a novel and significant problem, namely temporal feature disturbance, in the quantization of diffusion models. We conducted a detailed analysis of the root causes of this problem and introduced our TFMQ-DM quantization framework. In 4-bit quantization on extensive datasets and different diffusion models, this framework exhibits minimal performance degradation compared to the FP model and speedup quantization time. + +However, besides temporal features, we have also found the textual features introduced in Stable Diffusion have a physical meaning and influence the generation effect. Nevertheless, these textual features are not taken into consideration in current methods. Secondly, temporal feature maintenance is suitable for both PTQ and QAT scenarios. But in this study, our main focus is on the PTQ setting and achieving 4-bit quantization. The extension to the QAT setting should be conducted in the future to achieve lower-bit quantization and further performance enhancements. + +# References + +[1] Yash Bhalgat, Jinwon Lee, Markus Nagel, Tijmen Blankevoort, and Nojun Kwak. $\mathrm{Lsq}+$ : Improving low-bit quantization through learnable offsets and better initialization, 2020. 1, 2 +[2] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 1 +[3] Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In CVPR, pages 12413-12422, 2022. 2 +[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6 +[5] Steven K. Esser, Jeffrey L. McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S. Modha. Learned step size quantization. In ICLR, 2020. 1, 2, 8 +[6] Giulio Franzese, Simone Rossi, Lixuan Yang, Alessandro Finamore, Dario Rossi, Maurizio Filippone, and Pietro Michiardi. How much is enough? a study on diffusion times in score-based generative models. arXiv preprint arXiv:2206.05173, 2022. 2 +[7] Ruihao Gong, Xianglong Liu, Shenghu Jiang, Tianxiang Li, Peng Hu, Jiazhen Lin, Fengwei Yu, and Junjie Yan. Differentiable soft quantization: Bridging full-precision and low-bit neural networks. In ICCV, pages 4852–4861, 2019. 1, 2 +[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 1 +[9] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2015. 6 +[10] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. In NeurIPS, 2023. 2, 3, 4, 6, 7, 8 +[11] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning, 2022. 6 +[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018. 5, 6 +[13] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 1 + +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3, 6 +[15] Itay Hubara, Yury Nahshan, Yair Hanani, Ron Banner, and Daniel Soudry. Improving post training neural quantization: Layer-wise calibration and integer programming. arXiv preprint arXiv:2006.10518, 2020. 1, 2 +[16] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In CVPR, pages 2704-2713, 2018. 2, 6 +[17] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis, 2023. 1 +[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation, 2018. 6 +[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks, 2019. 6 +[20] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2022. 1 +[21] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. arXiv preprint arXiv:2106.00132, 2021. 1, 2 +[22] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[23] Xiuyu Li, Long Lian, Yijiang Liu, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In ICCV, 2023. 1, 2, 3, 4, 6, 7, 8 +[24] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. BRECQ: pushing the limit of post-training quantization by block reconstruction. In ICLR, 2021. 1, 2, 5, 6, 7, 8 +[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 6 +[26] Yang Lin, Tianyu Zhang, Peiqin Sun, Zheng Li, and Shuchang Zhou. Fq-vit: Post-training quantization for fully quantized vision transformer. In *IJCAI*, pages 1173–1179, 2022. 2 +[27] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In ICLR, 2022. 1 +[28] Christos Louizos, Matthias Reisser, Tijmen Blankevoort, Efstratos Gavves, and Max Welling. Relaxed quantization for discretized neural networks. In ICLR, 2019. 2 +[29] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, 2022. 1 +[30] Zhaoyang Lyu, Xudong Xu, Ceyuan Yang, Dahua Lin, and Bo Dai. Accelerating diffusion models via early stop of the diffusion process. arXiv preprint arXiv:2205.12524, 2022. 2 + +[31] Markus Nagel, Rana Ali Amjad, Mart van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. In ICML, pages 7197-7206, 2020. 1, 2, 6 +[32] Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bondarenko, Mart van Baalen, and Tijmen Blankevoort. A white paper on neural network quantization, 2021. 1, 5, 6, 8 +[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171, 2021. 3 +[34] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 6 +[35] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 3 +[36] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. In ECCV, pages 525-542, 2016. 6 +[37] Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. Fastspeech 2: Fast and high-quality end-to-end text to speech, 2022. 1 +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 6 +[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation, 2015. 3 +[40] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans, 2016. 5, 6 +[41] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In CVPR, 2023. 2, 3, 7, 8 +[42] Jonathan Shen, Ruoming Pang, Ron J. Weiss, Mike Schuster, Navdeep Jaitly, Zongheng Yang, Zhifeng Chen, Yu Zhang, Yuxuan Wang, RJ Skerry-Ryan, Rif A. Saurous, Yannis Agiomyrgiannakis, and Yonghui Wu. Natural tts synthesis by conditioning wavenet on mel spectrogram predictions, 2018. 1 +[43] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for diffusion models. In NeurIPS, 2023. 2, 3, 4, 5 +[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1, 2, 3 +[45] Changyuan Wang, Ziwei Wang, Xiuwei Xu, Yansong Tang, Jie Zhou, and Jiwen Lu. Towards accurate data-free quantization for diffusion models. arXiv preprint arXiv:2305.18723, 2023. 2, 5 +[46] Peisong Wang, Qiang Chen, Xiangyu He, and Jian Cheng. Towards accurate post-training network quantization via bitsplit and stitching. In ICML, pages 9847-9856, 2020. 2 +[47] Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. Qdrop: Randomly dropping quantization for + +extremely low-bit post-training quantization. In ICLR, 2022. 1, 2 +[48] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop, 2016. 1, 5, 6 +[49] Luoming Zhang, Yefei He, Zhenyu Lou, Xin Ye, Yuxing Wang, and Hong Zhou. Root quantization: a self-adaptive supplement ste. Applied Intelligence, 53(6):6266-6275, 2023. 2 +[50] Qinsheng Zhang, Molei Tao, and Yongxin Chen. gddim: Generalized denoising diffusion implicit models. arXiv preprint arXiv:2206.05564, 2022. 2 +[51] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. Opt: Open pre-trained transformer language models, 2022. 1 +[52] Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Truncated diffusion probabilistic models. stat, 1050:7, 2022. 2 +[53] Bohan Zhuang, Chunhua Shen, Mingkui Tan, Lingqiao Liu, and Ian Reid. Towards effective low-bitwidth convolutional neural networks. In CVPR, pages 7920-7928, 2018. 2 \ No newline at end of file diff --git a/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/images.zip b/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7fb46866e130637e0b8ff2756300b85382fda365 --- /dev/null +++ b/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd8e7fed7d6a3a4c069cb3ec77d29bf463dfbd61ae7f842e27aafd70b2abce27 +size 619910 diff --git a/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/layout.json b/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d6540b6c5ea8c49352b0d604ca0e513c84f718d0 --- /dev/null +++ b/2024/TFMQ-DM_ Temporal Feature Maintenance Quantization for Diffusion Models/layout.json @@ -0,0 +1,9055 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 103, + 543, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 103, + 543, + 121 + ], + "spans": [ + { + "bbox": [ + 52, + 103, + 543, + 121 + ], + "type": "text", + "content": "TFMQ-DM: Temporal Feature Maintenance Quantization for Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "spans": [ + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "text", + "content": "Yushi Huang" + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "text", + "content": ", Ruihao Gong" + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "text", + "content": ", Jing Liu" + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "text", + "content": ", Tianlong Chen" + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "text", + "content": ", Xianglong Liu" + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "text", + "content": "Beihang University " + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "text", + "content": "SenseTime Research " + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "text", + "content": "Monash University " + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 108, + 142, + 491, + 174 + ], + "type": "text", + "content": "UT Austin" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 81, + 175, + 509, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 175, + 509, + 186 + ], + "spans": [ + { + "bbox": [ + 81, + 175, + 509, + 186 + ], + "type": "text", + "content": "{huangyushi, gongruihao, liujing5} @sentetime.com, tianlong.chen@utexas.edu, xlliu@buaa.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 244, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 244, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 289, + 639 + ], + "type": "text", + "content": "The Diffusion model, a prevalent framework for image generation, encounters significant challenges in terms of broad applicability due to its extended inference times and substantial memory requirements. Efficient Post-training Quantization (PTQ) is pivotal for addressing these issues in traditional models. Different from traditional models, diffusion models heavily depend on the time-step " + }, + { + "bbox": [ + 46, + 244, + 289, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 244, + 289, + 639 + ], + "type": "text", + "content": " to achieve satisfactory multi-round denoising. Usually, " + }, + { + "bbox": [ + 46, + 244, + 289, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 244, + 289, + 639 + ], + "type": "text", + "content": " from the finite set " + }, + { + "bbox": [ + 46, + 244, + 289, + 639 + ], + "type": "inline_equation", + "content": "\\{1,\\dots,T\\}" + }, + { + "bbox": [ + 46, + 244, + 289, + 639 + ], + "type": "text", + "content": " is encoded to a temporal feature by a few modules totally irrespective of the sampling data. However, existing PTQ methods do not optimize these modules separately. They adopt inappropriate reconstruction targets and complex calibration methods, resulting in a severe disturbance of the temporal feature and denoising trajectory, as well as a low compression efficiency. To solve these, we propose a Temporal Feature Maintenance Quantization (TFMQ) framework building upon a Temporal Information Block which is just related to the time-step " + }, + { + "bbox": [ + 46, + 244, + 289, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 244, + 289, + 639 + ], + "type": "text", + "content": " and unrelated to the sampling data. Powered by the pioneering block design, we devise temporal information aware reconstruction (TIAR) and finite set calibration (FSC) to align the full-precision temporal features in a limited time. Equipped with the framework, we can maintain the most temporal information and ensure the end-to-end generation quality. Extensive experiments on various datasets and diffusion models prove our state-of-the-art results. Remarkably, our quantization approach, for the first time, achieves model performance nearly on par with the full-precision model under 4-bit weight quantization. Additionally, our method incurs almost no extra computational cost and accelerates quantization time by " + }, + { + "bbox": [ + 46, + 244, + 289, + 639 + ], + "type": "inline_equation", + "content": "2.0\\times" + }, + { + "bbox": [ + 46, + 244, + 289, + 639 + ], + "type": "text", + "content": " on LSUN-Bedrooms " + }, + { + "bbox": [ + 46, + 244, + 289, + 639 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 46, + 244, + 289, + 639 + ], + "type": "text", + "content": " compared to previous works. Our code is publicly available at https://github.com/ModelTC/TFMQ-DM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 213, + 386, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 213, + 386, + 225 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 386, + 225 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 234, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 547, + 521 + ], + "type": "text", + "content": "Generative modeling plays a crucial role in machine learning, particularly in applications like image [13, 14, 17, 38, 44], voice [37, 42], and text synthesis [2, 51]. Diffusion models have showcased impressive capabilities in producing high-quality samples across diverse domains. In comparison to generative adversarial networks (GANs) [8] and variational autoencoders (VAEs) [20], diffusion models successfully sidestep issues such as model collapse and posterior collapse, resulting in a more stable training process. However, the substantial computational cost poses a critical bottleneck hampering the widespread adoption of diffusion models. Furthermore, the computational cost for diffusion models can be attributed to two primary factors. First, these models typically require hundreds of denoising steps to generate images, rendering the procedure considerably slower than that of GANs. Prior efforts [21, 27, 29, 44] have addressed this challenge by seeking shorter and more efficient sampling trajectories, thereby reducing the number of denoising steps. Second, the substantial network architecture of diffusion models demands considerable time and memory resources, particularly for foundational models pre-trained on large-scale datasets, e.g., LDM [38] and Stable Diffusion. Our work aims to tackle the latter challenge, focusing on the compression of diffusion models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "content": "Quantization is currently the most widely used method [1, 5, 7, 31, 32] for compressing models by mapping high-precision floating-point numbers to low-precision numbers. Among different quantization methods, Posttraining quantization (PTQ) [15, 31, 47] incurs lower overhead and is more user-friendly without the need for retraining or fine-tuning. While PTQ on conventional models has undergone extensive study [7, 24, 31, 47], its application to diffusion models has shown huge performance degradation, especially under low-bit settings. For instance, Q-Diffusion [23] exhibits severe accuracy drop on some datasets [48] under 4-bit quantization. We believe the reason they fail to achieve better results is that they all overlook the sampling data independence and uniqueness of temporal features, which are generated from time-step " + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "content": " through a few modules, used to control the denoising trajectory in dif" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 683, + 128, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 683, + 128, + 693 + ], + "spans": [ + { + "bbox": [ + 57, + 683, + 128, + 693 + ], + "type": "text", + "content": "*Equal Contribution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 60, + 693, + 243, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 693, + 243, + 703 + ], + "spans": [ + { + "bbox": [ + 60, + 693, + 243, + 703 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 60, + 693, + 243, + 703 + ], + "type": "text", + "content": " Work done during an internship at SenseTime Research." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 60, + 703, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 703, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 60, + 703, + 138, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 60, + 703, + 138, + 712 + ], + "type": "text", + "content": " Corresponding Author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7362" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": "fusion models. Furthermore, we observe that temporal feature disturbance significantly impacts model performance in the aforementioned methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 117, + 289, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 117, + 289, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 117, + 289, + 452 + ], + "type": "text", + "content": "To tackle temporal feature disturbance, we first find that the modules generating temporal features are independent of the sampling data and define the whole modules as the Temporal Information Block. All existing methods do not separately optimize this block during the quantization process, causing temporal features to overfit to limited calibration data. On the other hand, since the maximum time-step for denoising is a finite positive integer, the temporal feature and the activations during its generation form a finite set. The optimal approach is to optimize each element in this set individually. Based on these observations and analyses, we propose a novel quantization reconstruction approach, temporal information aware reconstruction (TIAR), specifically optimizing for temporal features. It aims to reduce temporal feature loss as the optimization objective while isolating the network's components related to sampled data from the generation of temporal features during calibration. Furthermore, we also introduce a calibration strategy, finite set calibration (FSC), for the finite set of temporal features and activations during its generation. This strategy employs different quantization parameters for activations corresponding to different time-steps. Moreover, the calibration speed of this method is faster than existing mainstream methods [1, 5], for example, we speedup quantization time by " + }, + { + "bbox": [ + 47, + 117, + 289, + 452 + ], + "type": "inline_equation", + "content": "2.0 \\times" + }, + { + "bbox": [ + 47, + 117, + 289, + 452 + ], + "type": "text", + "content": " on LSUN-Bedrooms " + }, + { + "bbox": [ + 47, + 117, + 289, + 452 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 47, + 117, + 289, + 452 + ], + "type": "text", + "content": " dataset, yet the strategy incurs negligible inference and storage overhead. The overview of our framework can be seen in Fig. 1. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 460, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 460, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 287, + 508 + ], + "type": "text", + "content": "- We discover that existing quantization methods suffer from temporal feature disturbance, disrupting the denoising trajectory of diffusion models and significantly affecting the quality of generated images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 509, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 509, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 509, + 287, + 555 + ], + "type": "text", + "content": "- We reveal that the disturbance comes from two aspects: inappropriate reconstruction target and unaware of finite activations. Both inducements ignore the special characteristics of time information-related modules." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 555, + 287, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 287, + 627 + ], + "type": "text", + "content": "- An advanced framework (TFMQ-DM) is proposed, consisting of temporal information aware reconstruction (TIAR) for weight quantization and finite set calibration (FSC) for activation quantization. Both are based on a Temporal Information Block specially devised for diffusion models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 628, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 288, + 712 + ], + "type": "text", + "content": "- Extensive experiments on various datasets show that our novel framework achieves a new state-of-the-art result in PTQ of diffusion models, especially under 4-bit weight quantization, and significantly accelerates quantization time. For some hard tasks, e.g., CelebA-HQ " + }, + { + "bbox": [ + 47, + 628, + 288, + 712 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 47, + 628, + 288, + 712 + ], + "type": "text", + "content": ", our method reduces the FID score by 6.71 (images in appendix)." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 393, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 393, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 393, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 91, + 451, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 451, + 102 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 451, + 102 + ], + "type": "text", + "content": "2.1. Efficient Diffusion Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 109, + 547, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 547, + 242 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 547, + 242 + ], + "type": "text", + "content": "There are diverse perspectives to accelerate the inference of diffusion model, e.g., retraining-based model design [3, 6, 30, 52] and retraining-free sampler strategy [21, 44, 50]. However, the retraining-based method proves resource-intensive and time-consuming. The efficient samplers can reduce sampling iterations but still suffer from diffusion models' extensive parameters and computational complexity. In this paper, we focus on diminishing the time and memory overhead of the single-step denoising process using low-bit quantization in a training-free manner, a method orthogonal to previous speedup techniques." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 247, + 423, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 247, + 423, + 259 + ], + "spans": [ + { + "bbox": [ + 306, + 247, + 423, + 259 + ], + "type": "text", + "content": "2.2. Model Quantization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 265, + 547, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 547, + 397 + ], + "type": "text", + "content": "Quantization is a predominant technique for minimizing storage and computational costs. It can be categorized into quantization-aware training (QAT) [7, 16, 28, 49, 53] and post-training quantization (PTQ) [15, 24, 26, 31, 47]. QAT requires intensive model training with substantial data and computational demands. Correspondingly, PTQ compresses models without re-training, making it a preferred method due to its minimal data requirements and easy deployment on real hardware. In PTQ, high-precision values are mapped into discrete levels using uniform quantization expressed as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 368, + 401, + 545, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 401, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 368, + 401, + 545, + 422 + ], + "type": "interline_equation", + "content": "\\hat {x} = \\Phi \\left(\\left\\lfloor \\frac {x}{s} \\right\\rceil + z, 0, 2 ^ {b} - 1\\right), \\tag {1}", + "image_path": "d9bbb2ae423b01403db5c266c25d6c4d24b3bf83d49e925c3c6224872c0d7040.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "text", + "content": " represents floating-point data, " + }, + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "text", + "content": " is the quantized value, " + }, + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "text", + "content": " is the quantization step size, " + }, + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "text", + "content": " is the zero offset, and " + }, + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "text", + "content": " is the target bit-width. The clamp function " + }, + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "inline_equation", + "content": "\\Phi(\\cdot)" + }, + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "text", + "content": " clips the rounded value " + }, + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "inline_equation", + "content": "\\left\\lfloor \\frac{x}{s} \\right\\rfloor + z" + }, + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "text", + "content": " within the range of " + }, + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "inline_equation", + "content": "[0, 2^b - 1]" + }, + { + "bbox": [ + 304, + 426, + 547, + 533 + ], + "type": "text", + "content": ". However, naive quantization may lead to accuracy degradation, especially for low-bit quantization. Recent studies [15, 24, 31, 46, 47] have explored innovative strategies based on reconstruction to preserve model performance after low-bit quantization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "type": "text", + "content": "In contrast, the iterative denoising process in diffusion models presents new challenges for PTQ in comparison to traditional models. PTQ4DM [41] represents the initial attempt to quantize diffusion models to 8-bit, albeit with limited experiments and lower resolutions. Conversely, Q-Diffusion [23] achieves enhanced performance and is evaluated on a broader dataset range. Moreover, PTQD [10] eliminates quantization noise through correlated and residual noise correction. Notably, traditional single-time-step PTQ calibration methods are unsuitable for diffusion models due to significant activation distribution changes with each time-step [23, 41, 43, 45]. ADP-DM [45] proposes group-wise quantization across time-steps for diffusion models, and TDQ [43] introduces distinct quantization parameters for different time-steps. However, all of the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7363" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "above works overlook the specificity of temporal features. To address temporal feature disturbance in the aforementioned works, our study delves into the inducements of the phenomenon and introduces a novel reconstruction and calibration framework, significantly enhancing quantized diffusion model performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 154, + 132, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 154, + 132, + 166 + ], + "spans": [ + { + "bbox": [ + 47, + 154, + 132, + 166 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 175, + 287, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 287, + 246 + ], + "type": "text", + "content": "Diffusion models. Diffusion models [14, 44] iteratively add Gaussian noise with a variance schedule " + }, + { + "bbox": [ + 46, + 175, + 287, + 246 + ], + "type": "inline_equation", + "content": "\\beta_{1},\\ldots ,\\beta_{T}\\in" + }, + { + "bbox": [ + 46, + 175, + 287, + 246 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 46, + 175, + 287, + 246 + ], + "type": "text", + "content": " to data " + }, + { + "bbox": [ + 46, + 175, + 287, + 246 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0\\sim q(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 175, + 287, + 246 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 175, + 287, + 246 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 175, + 287, + 246 + ], + "type": "text", + "content": " times as sampling process, resulting in a sequence of noisy samples " + }, + { + "bbox": [ + 46, + 175, + 287, + 246 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1,\\dots ,\\mathbf{x}_T" + }, + { + "bbox": [ + 46, + 175, + 287, + 246 + ], + "type": "text", + "content": " In DDPMs [14], the former sampling process is a Markov chain, taking the form:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 254, + 287, + 286 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 254, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 90, + 254, + 287, + 286 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right), \\tag {2}", + "image_path": "46deac7f886bb5648c5513f3d975914089d49e5bcaa72cc3a30f0603875b0340.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 289, + 242, + 303 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 289, + 242, + 303 + ], + "spans": [ + { + "bbox": [ + 91, + 289, + 242, + 303 + ], + "type": "interline_equation", + "content": "q (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1}) = \\mathcal {N} (\\mathbf {x} _ {t}; \\sqrt {\\alpha_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}),", + "image_path": "b2f0671058ababfecb4efc55f7889b4d849e188010037e0aec602d69b8fa42ce.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 310, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 287, + 416 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 310, + 287, + 416 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 46, + 310, + 287, + 416 + ], + "type": "text", + "content": ". Conversely, the denoising process removes noise from a sample from Gaussian noise " + }, + { + "bbox": [ + 46, + 310, + 287, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 310, + 287, + 416 + ], + "type": "text", + "content": " to gradually generate high-fidelity images. Nevertheless, due to the unavailability of the true reverse conditional distribution " + }, + { + "bbox": [ + 46, + 310, + 287, + 416 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)" + }, + { + "bbox": [ + 46, + 310, + 287, + 416 + ], + "type": "text", + "content": ", diffusion models approximate it via variational inference by learning a Gaussian distribution " + }, + { + "bbox": [ + 46, + 310, + 287, + 416 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t) = \\mathcal{N}(\\mathbf{x}_{t - 1};\\boldsymbol {\\mu}_\\theta (\\mathbf{x}_t,t),\\boldsymbol {\\Sigma}_\\theta (\\mathbf{x}_t,t))" + }, + { + "bbox": [ + 46, + 310, + 287, + 416 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 46, + 310, + 287, + 416 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\mu}_{\\theta}" + }, + { + "bbox": [ + 46, + 310, + 287, + 416 + ], + "type": "text", + "content": " can be derived by reparameterization trick as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 72, + 423, + 287, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 423, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 72, + 423, + 287, + 450 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right)\\right), \\tag {3}", + "image_path": "12ce7f487b5848b3b96dfc154bef40d3342b5847cb8fb55444ffb90f530dadf9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 459, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 507 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 459, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t} = \\prod_{i=1}^{t} \\alpha_{i}" + }, + { + "bbox": [ + 46, + 459, + 287, + 507 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 459, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 46, + 459, + 287, + 507 + ], + "type": "text", + "content": " is a noise estimation model. The variance " + }, + { + "bbox": [ + 46, + 459, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\theta}(\\mathbf{x}_{t}, t)" + }, + { + "bbox": [ + 46, + 459, + 287, + 507 + ], + "type": "text", + "content": " can be either learned [33] or fixed to a constant schedule [14] " + }, + { + "bbox": [ + 46, + 459, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 46, + 459, + 287, + 507 + ], + "type": "text", + "content": ". When employing the latter method, " + }, + { + "bbox": [ + 46, + 459, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t-1}" + }, + { + "bbox": [ + 46, + 459, + 287, + 507 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 515, + 287, + 542 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 515, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 515, + 287, + 542 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {x} _ {t}, t)\\right) + \\sigma_ {t} \\mathbf {z}, \\tag {4}", + "image_path": "71867454a23c8108e895eb80e3355b451730a333387f039ed8fe07bc3cd57ccf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 549, + 129, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 129, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 129, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 549, + 129, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{z}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 571, + 287, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 571, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 287, + 691 + ], + "type": "text", + "content": "Reconstruction on diffusion models. UNet [39], the predominant model employed as " + }, + { + "bbox": [ + 46, + 571, + 287, + 691 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 46, + 571, + 287, + 691 + ], + "type": "text", + "content": " in Eq. 4 to predict Gaussian noise, can be divided into blocks that incorporate residual connections (such as Residual Bottleneck Blocks or Transformer Blocks [35]) and the remaining layers. Numerous PTQ approaches for diffusion models are grounded in layer/block-wise reconstruction [10, 23, 41, 43] to obtain optimal quantization parameters. For example, in the Residual Bottleneck Block, this approach typically minimizes the following loss function as its optimization objective:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 119, + 699, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 699, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 119, + 699, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i} = \\left\\| f _ {i} (\\cdot) - \\widehat {f} _ {i} (\\cdot) \\right\\| _ {F} ^ {2}, \\tag {5}", + "image_path": "d9ec9c96ae2fd131a4388ef723bdcf9e179357f5dcc59da1fd757a1b5b1e8de7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_F^2" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " denotes the Frobenius norm. The function " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "f_i(\\cdot)" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " Residual Bottleneck Block, and " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\widehat{f}_i(\\cdot)" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " is its quantized counterpart. Furthermore, in the ensuing sections, we use " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " to denote the total number of Residual Bottleneck Blocks in a single diffusion model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "text", + "content": "Temporal feature in diffusion models. Seeing from Fig. 1 (a), time-step " + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "text", + "content": " is encoded with time embed" + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "text", + "content": " and then passes through the embedding layer" + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "text", + "content": " in each Residual Bottleneck Block, resulting in a series of unique activations. In this paper, we denote these activations as temporal features. Notably, temporal features are independent of " + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "text", + "content": " and unrelated to other temporal features from different time-steps. To enhance clarity, we simplify our notation as follows: we represent time embed as " + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "text", + "content": ", the embedding layer in the " + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "text", + "content": " Residual Bottleneck Block as " + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "inline_equation", + "content": "g_i(\\cdot)" + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "text", + "content": ", and denote the " + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "text", + "content": " temporal feature at time-step " + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{emb}_{t,i}" + }, + { + "bbox": [ + 305, + 144, + 545, + 299 + ], + "type": "text", + "content": ". Moreover, as illustrated in Fig. 1 (a), the relationship is explicitly expressed by the equation:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 348, + 308, + 545, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 308, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 348, + 308, + 545, + 322 + ], + "type": "interline_equation", + "content": "\\text {T e m p o r a l} \\quad \\mathbf {e m b} _ {t, i} = g _ {i} (h (t)) \\tag {6}", + "image_path": "d10e0299a08982efff1f2b0bdb702ca77c51a0359e9fb4d47c01f5eaef6f02fd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 331, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 545, + 437 + ], + "type": "text", + "content": "Additionally, we have found that temporal features play a crucial role in the context of the diffusion model, holding unique and substantial physical implications. These features encompass temporal information that signifies the current image's temporal position along the denoising trajectory. Within the UNet structure, each time-step transforms into these temporal features, thereby controlling the denoising trajectory by applying them to the features of images generated at each iteration." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 449, + 466, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 466, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 466, + 462 + ], + "type": "text", + "content": "4. TFMQ for Diffusion Models" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 469, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 469, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 305, + 469, + 545, + 518 + ], + "type": "text", + "content": "We present our novel PTQ framework in this section. We first observe temporal disturbance in previous methods in Sec. 4.1 and then analyze the inducements in Sec. 4.2. Finally, we propose our solutions in Sec. 4.3." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 525, + 473, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 473, + 537 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 473, + 537 + ], + "type": "text", + "content": "4.1. Temporal Feature Disturbance" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 543, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 543, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 305, + 543, + 545, + 591 + ], + "type": "text", + "content": "Based on Sec. 3, we investigate the impact of previous PTQ works on temporal features, and we identify the phenomenon of temporal feature disturbance, which significantly deteriorates the quality of generated images." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 601, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 601, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 305, + 601, + 545, + 661 + ], + "type": "text", + "content": "Temporal feature error. We thoroughly analyze temporal feature variations before and after the quantization of embedding layers and time embed in the Stable Diffusion model (" + }, + { + "bbox": [ + 305, + 601, + 545, + 661 + ], + "type": "inline_equation", + "content": "T = 50, i = 11" + }, + { + "bbox": [ + 305, + 601, + 545, + 661 + ], + "type": "text", + "content": "). Prior to this analysis, we introduce the temporal feature error as defined by:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 380, + 669, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 669, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 380, + 669, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\cos \\left(\\widehat {\\mathbf {e m b}} _ {t, i}, \\widehat {\\mathbf {e m b}} _ {t, i}\\right), \\tag {7}", + "image_path": "fd1ac6cb0158e34d7956fe6f6fab9f6018e1f0c27d0aecfcb78edc0f9f13c702.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 693, + 507, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 693, + 507, + 703 + ], + "spans": [ + { + "bbox": [ + 317, + 693, + 507, + 703 + ], + "type": "text", + "content": "PyTorch time embed implementation in diffusion models." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 703, + 522, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 522, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 522, + 712 + ], + "type": "text", + "content": "PyTorch embedding layer implementation in diffusion models." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7364" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 76, + 285, + 162 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 285, + 162 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 285, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 285, + 162 + ], + "type": "image", + "image_path": "87b6455d1bf6e3904f1661eca93aa8229ab12114fb2c24f4ad02d0f7cadc3d67.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 54, + 164, + 285, + 208 + ], + "blocks": [ + { + "bbox": [ + 54, + 164, + 285, + 208 + ], + "lines": [ + { + "bbox": [ + 54, + 164, + 285, + 208 + ], + "spans": [ + { + "bbox": [ + 54, + 164, + 285, + 208 + ], + "type": "image", + "image_path": "01437bc9a586d1dad1f66ae44a29260ebd857443fb7eb429f28491d61385d568.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 545, + 293 + ], + "type": "text", + "content": "Figure 1. Overview of the proposed Temporal Feature Maintenance Quantization. (a) Temporal Feature " + }, + { + "bbox": [ + 46, + 217, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{emb}_{t,i}" + }, + { + "bbox": [ + 46, + 217, + 545, + 293 + ], + "type": "text", + "content": ", belonging to a finite set representing temporal information, has been overlooked in previous works due to inappropriate reconstruction targets (box with a solid line). (b) This oversight leads to a severe disturbance for " + }, + { + "bbox": [ + 46, + 217, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{emb}_{t,i}" + }, + { + "bbox": [ + 46, + 217, + 545, + 293 + ], + "type": "text", + "content": " and results in the mismatch of crucial temporal information for the diffusion model's generation, causing a deviation in the denoising trajectory and a significant drop in accuracy. (c) Based on these analyses, we introduce a Temporal Information Block that exclusively correlates with the time-step " + }, + { + "bbox": [ + 46, + 217, + 545, + 293 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 217, + 545, + 293 + ], + "type": "text", + "content": ". Leveraging this " + }, + { + "bbox": [ + 46, + 217, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 217, + 545, + 293 + ], + "type": "text", + "content": "-unrelated block, we enable Temporal Information Aware Reconstruction and Finite Set Calibration (utilizing the finite number of " + }, + { + "bbox": [ + 46, + 217, + 545, + 293 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 217, + 545, + 293 + ], + "type": "text", + "content": "). This approach achieves the maintenance of temporal features and yields state-of-the-art results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 286, + 71, + 541, + 209 + ], + "blocks": [ + { + "bbox": [ + 286, + 71, + 541, + 209 + ], + "lines": [ + { + "bbox": [ + 286, + 71, + 541, + 209 + ], + "spans": [ + { + "bbox": [ + 286, + 71, + 541, + 209 + ], + "type": "image", + "image_path": "afc432d837c9c40687168e5dcb02f97600585a64ffcb5e6266585921821bfed2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot)" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": " denotes cosine similarity, and " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{emb}_{t,i}}" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": " signifies the temporal feature corresponding to " + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{emb}_{t,i}" + }, + { + "bbox": [ + 46, + 305, + 287, + 390 + ], + "type": "text", + "content": " in the quantized model. As illustrated in Fig. 2 (Left), quantization induces notable temporal feature errors. We term this phenomenon, characterized by substantial temporal feature errors within diffusion models, as temporal feature disturbance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 401, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 484 + ], + "type": "text", + "content": "Temporal information mismatch. Temporal feature disturbance alters the original embedded temporal information. Specifically, " + }, + { + "bbox": [ + 46, + 401, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{emb}_{t,i}" + }, + { + "bbox": [ + 46, + 401, + 287, + 484 + ], + "type": "text", + "content": " is intended to correspond to timestep " + }, + { + "bbox": [ + 46, + 401, + 287, + 484 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 401, + 287, + 484 + ], + "type": "text", + "content": ". However, due to significant errors, the quantized model's " + }, + { + "bbox": [ + 46, + 401, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{emb}}_{t,i}" + }, + { + "bbox": [ + 46, + 401, + 287, + 484 + ], + "type": "text", + "content": " is no longer accurately associated with " + }, + { + "bbox": [ + 46, + 401, + 287, + 484 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 401, + 287, + 484 + ], + "type": "text", + "content": ", resulting in what we term as temporal information mismatch:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 484, + 287, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 484, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 107, + 484, + 287, + 499 + ], + "type": "interline_equation", + "content": "t \\leftarrow \\mathbf {e m b} _ {t, i}, \\quad t \\nleftarrow \\widehat {\\mathbf {e m b}} _ {t, i}. \\tag {8}", + "image_path": "658945fb37efedf62123425b188cc5a91536c0a63c4bf106ff6130fc28a5943b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 502, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 597 + ], + "type": "text", + "content": "Furthermore, as depicted in Fig. 2 (Right), we even observe a pronounced temporal information mismatch. Specifically, the temporal feature generated by the quantized model at time-step " + }, + { + "bbox": [ + 46, + 502, + 287, + 597 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 502, + 287, + 597 + ], + "type": "text", + "content": " exhibits a divergence from that of the full-precision model at the corresponding time-step. Instead, it tends to align more closely with the temporal feature corresponding to " + }, + { + "bbox": [ + 46, + 502, + 287, + 597 + ], + "type": "inline_equation", + "content": "t + \\delta_t" + }, + { + "bbox": [ + 46, + 502, + 287, + 597 + ], + "type": "text", + "content": ", importing wrong temporal information from " + }, + { + "bbox": [ + 46, + 502, + 287, + 597 + ], + "type": "inline_equation", + "content": "t + \\delta_t" + }, + { + "bbox": [ + 46, + 502, + 287, + 597 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 607, + 287, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 607, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 47, + 607, + 287, + 655 + ], + "type": "text", + "content": "Trajectory deviation. Temporal information mismatch delivers wrong temporal information, therefore, causing a deviation in the corresponding temporal position of the image within the denoising trajectory, ultimately leading to:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 661, + 287, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 661, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 141, + 661, + 287, + 673 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t} \\neq \\mathbf {x} _ {t - 1}, \\tag {9}", + "image_path": "e986d7f70fc564b210f8a6b185000566807f3f2ed8d6e62a994f491c54e7c56d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "where we apply disrupted temporal features to the model. Evidently, as the deviation in the denoising trajectory accumulates with the increase in the number of denoising" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 305, + 425, + 417 + ], + "blocks": [ + { + "bbox": [ + 307, + 305, + 425, + 417 + ], + "lines": [ + { + "bbox": [ + 307, + 305, + 425, + 417 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 425, + 417 + ], + "type": "image", + "image_path": "9a683213a2eeebc8b675f563ced683121f11e6445e523c26ff4265a00d6f6dd2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 424, + 545, + 504 + ], + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 504 + ], + "type": "text", + "content": "Figure 2. (Left) Temporal feature disturbance. The inflection points serve as indicators of temporal feature errors at different time-steps, and they highlight the significant phenomenon of temporal feature disturbance. (Right) Temporal information mismatch. The coordinates of the inflection points on the blue curve can be denoted as " + }, + { + "bbox": [ + 304, + 424, + 545, + 504 + ], + "type": "inline_equation", + "content": "(t,t + \\delta_{t,i})" + }, + { + "bbox": [ + 304, + 424, + 545, + 504 + ], + "type": "text", + "content": ". It indicates " + }, + { + "bbox": [ + 304, + 424, + 545, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{emb}_{t + \\delta_{t,i},i}" + }, + { + "bbox": [ + 304, + 424, + 545, + 504 + ], + "type": "text", + "content": " exhibits the highest similarity with " + }, + { + "bbox": [ + 304, + 424, + 545, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{emb}_{t,i}" + }, + { + "bbox": [ + 304, + 424, + 545, + 504 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 427, + 305, + 544, + 418 + ], + "blocks": [ + { + "bbox": [ + 427, + 305, + 544, + 418 + ], + "lines": [ + { + "bbox": [ + 427, + 305, + 544, + 418 + ], + "spans": [ + { + "bbox": [ + 427, + 305, + 544, + 418 + ], + "type": "image", + "image_path": "677aa6975168f144167cfeb2bc4f955dd11a8e1777b0b6c487ff1cd255bc00ce.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 518, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 566 + ], + "type": "text", + "content": "iterations, the final generated image struggles to align with " + }, + { + "bbox": [ + 304, + 518, + 545, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 304, + 518, + 545, + 566 + ], + "type": "text", + "content": ". This evolution is illustrated in Fig. 3, where we maintain UNet excluding embedding layers and time embed in full precision." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 576, + 428, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 576, + 428, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 576, + 428, + 588 + ], + "type": "text", + "content": "4.2. Inducement Analyses" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 595, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 595, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 595, + 545, + 654 + ], + "type": "text", + "content": "In this section, we explore the two inducements of temporal feature disturbance. For the purpose of clarity, in the subsequent sections, \"reconstruction\" specifically points to slight weight adjustment for minimal quantization error, while \"calibration\" specifically refers to activation calibration." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "Inappropriate reconstruction target. Previous PTQ works [10, 23, 43] have achieved remarkable progress on diffusion models. However, these existing methods overlook the temporal feature's independence and its dis" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7365" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 205 + ], + "type": "image", + "image_path": "1304d60b10183617411e1fa61cc9de06863ae7c71a1307ba4a64fe68af16fa04.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 210, + 547, + 269 + ], + "lines": [ + { + "bbox": [ + 46, + 210, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 547, + 269 + ], + "type": "text", + "content": "Figure 3. Denoising process of full-precision (Upper) and w4a8 quantized (Lower) Stable-Diffusion " + }, + { + "bbox": [ + 46, + 210, + 547, + 269 + ], + "type": "inline_equation", + "content": "(T = 50)" + }, + { + "bbox": [ + 46, + 210, + 547, + 269 + ], + "type": "text", + "content": " under the same experiment settings and prompt: A man in the snow on a snow board. We represent " + }, + { + "bbox": [ + 46, + 210, + 547, + 269 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{emb}_{t,i}\\}_{i=0,\\dots,n}" + }, + { + "bbox": [ + 46, + 210, + 547, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 210, + 547, + 269 + ], + "type": "inline_equation", + "content": "\\{\\widehat{\\mathbf{emb}}_{t,i}\\}_{i=0,\\dots,n}" + }, + { + "bbox": [ + 46, + 210, + 547, + 269 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 210, + 547, + 269 + ], + "type": "inline_equation", + "content": "\\mathbf{EMB}_t" + }, + { + "bbox": [ + 46, + 210, + 547, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 210, + 547, + 269 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{EMB}_t}" + }, + { + "bbox": [ + 46, + 210, + 547, + 269 + ], + "type": "text", + "content": ", respectively. Additionally, we denote " + }, + { + "bbox": [ + 46, + 210, + 547, + 269 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{x}}_t" + }, + { + "bbox": [ + 46, + 210, + 547, + 269 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 210, + 547, + 269 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 210, + 547, + 269 + ], + "type": "text", + "content": " in the context of the quantized diffusion model. It is noteworthy that, in the quantized model employed here, to showcase the impact of temporal features, only the layers in Temporal Information Block are quantized and the components unrelated to the generation of temporal features are maintained in full precision." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 280, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 287, + 351 + ], + "type": "text", + "content": "tive physical significance. During their reconstruction processes, there was a lack of optimization for the embedding layers; instead, a Residual Bottleneck Block of coarser granularity was selected as the reconstruction target. This method involves two potential factors causing temporal feature disturbance:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 354, + 287, + 437 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 354, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 287, + 389 + ], + "type": "text", + "content": "- Optimize the objective as expressed in Eq. 5 to decrease the reconstruction loss of Residual Bottleneck Block, as opposed to direct reduction of temporal feature loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 389, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 389, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 287, + 437 + ], + "type": "text", + "content": "- During backpropagation of the reconstruction process, embedding layers independent from " + }, + { + "bbox": [ + 47, + 389, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 47, + 389, + 287, + 437 + ], + "type": "text", + "content": " are affected by " + }, + { + "bbox": [ + 47, + 389, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 47, + 389, + 287, + 437 + ], + "type": "text", + "content": ", resulting in an overfitting scenario on limited calibration data." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 438, + 288, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 288, + 581 + ], + "type": "text", + "content": "To further validate our analyses, we respectively evaluate the FID [12] and sFID [40] for this reconstruction method, e.g., BRECQ [24] and the approach where we freeze the parameters of the embedding layers during the reconstruction of the Residual Bottleneck Block, initializing the embedding layers solely through Min-max [32] for comparison. As shown in Tab. 1, the Freeze strategy exhibits better results, which verify that embedding layers serve as their own optimization objective and maintain their independence of " + }, + { + "bbox": [ + 46, + 438, + 288, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 438, + 288, + 581 + ], + "type": "text", + "content": " can significance mitigate temporal feature disturbance, especially at low-bit." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 60, + 635, + 272, + 712 + ], + "blocks": [ + { + "bbox": [ + 46, + 593, + 288, + 626 + ], + "lines": [ + { + "bbox": [ + 46, + 593, + 288, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 288, + 626 + ], + "type": "text", + "content": "Table 1. FID and sFID on LSUN-Bedrooms " + }, + { + "bbox": [ + 46, + 593, + 288, + 626 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 593, + 288, + 626 + ], + "type": "text", + "content": " [48] for LDM-4 with 50000 sampling images. Prev represents BRECQ. Freeze denotes our trial." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 635, + 272, + 712 + ], + "lines": [ + { + "bbox": [ + 60, + 635, + 272, + 712 + ], + "spans": [ + { + "bbox": [ + 60, + 635, + 272, + 712 + ], + "type": "table", + "html": "
MethodsBits (W/A)FID↓sFID↓
Full Prec.32/322.987.09
Prev8/87.5112.54
Freeze8/85.76 (-1.75)8.42 (-4.12)
Prev4/89.3622.73
Freeze4/87.08 (-2.28)16.82 (-5.91)
", + "image_path": "aac2fe3feffdc825fa50a35d1aa4dd34b59ee4b19f0f8213fffd2bd50bae272b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 280, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 545, + 459 + ], + "type": "text", + "content": "Unaware of finite activations within " + }, + { + "bbox": [ + 304, + 280, + 545, + 459 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 304, + 280, + 545, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 280, + 545, + 459 + ], + "type": "inline_equation", + "content": "g_{i}(\\cdot)" + }, + { + "bbox": [ + 304, + 280, + 545, + 459 + ], + "type": "text", + "content": ". We observe that, given " + }, + { + "bbox": [ + 304, + 280, + 545, + 459 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 280, + 545, + 459 + ], + "type": "text", + "content": " as a finite positive integer, the set of all possible activation values for embedding layers and time embed is finite and strictly dependent on timesteps. Within this set, activations corresponding to the same layer display notable range variations across different timesteps (refer to the appendix). Previous methods [43, 45] mainly focus on finding the optimal calibration method for " + }, + { + "bbox": [ + 304, + 280, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{x}}_t" + }, + { + "bbox": [ + 304, + 280, + 545, + 459 + ], + "type": "text", + "content": "-related network components. Moreover, akin to the first inducement, their calibration is directly towards the Residual Bottleneck Block, which proves suboptimal (refer to the appendix). However, based on the finite activations, we can employ calibration methods, especially for these time information-related activations, to better adapt to their range variations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 468, + 477, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 468, + 477, + 481 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 477, + 481 + ], + "type": "text", + "content": "4.3. Temporal Feature Maintenance" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 488, + 545, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 545, + 549 + ], + "type": "text", + "content": "To address the problem of temporal feature disturbance, we design a novel Temporal Information Block to maintain the temporal features. Based on the block, Temporal Information Aware Reconstruction and Finite Set Calibration are proposed to solve the two inducements analyzed above." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 559, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 545, + 656 + ], + "type": "text", + "content": "Temporal Information Block. Based on the inducements, it is crucial to meticulously separate the reconstruction and calibration process for each embedding layer and Residual Bottleneck Block to enhance quantized model performance. Considering the unique structure of the UNet, we consolidate all embedding layers and time embed into a unified Temporal Information Block, which can be denoted as " + }, + { + "bbox": [ + 304, + 559, + 545, + 656 + ], + "type": "inline_equation", + "content": "\\{g_{i}(h(\\cdot))\\}_{i = 0,\\dots,n}" + }, + { + "bbox": [ + 304, + 559, + 545, + 656 + ], + "type": "text", + "content": " (see Fig. 1 (c))." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": "Temporal information aware reconstruction. Based on the Temporal Information Block, we propose Temporal information aware reconstruction (TIAR) to tackle the first inducement. The optimization objective for the block during" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7366" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 209, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 209, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 209, + 83 + ], + "type": "text", + "content": "the reconstruction process is as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 87, + 95, + 287, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 95, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 87, + 95, + 287, + 126 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T I A R} = \\sum_ {i = 0} ^ {n} \\| g _ {i} (h (t)) - \\widehat {g} _ {i} (\\widehat {h} (t)) \\| _ {F} ^ {2}, \\tag {10}", + "image_path": "a270b78f8dd017fddf054dd41a25101b55dbee809d87d5703e8d4ac4a351c503.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 137, + 287, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 137, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 47, + 137, + 287, + 175 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 137, + 287, + 175 + ], + "type": "inline_equation", + "content": "\\widehat{h} (\\cdot)" + }, + { + "bbox": [ + 47, + 137, + 287, + 175 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 137, + 287, + 175 + ], + "type": "inline_equation", + "content": "\\widehat{g}_i(\\cdot)" + }, + { + "bbox": [ + 47, + 137, + 287, + 175 + ], + "type": "text", + "content": " are quantized versions of " + }, + { + "bbox": [ + 47, + 137, + 287, + 175 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 47, + 137, + 287, + 175 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 137, + 287, + 175 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 47, + 137, + 287, + 175 + ], + "type": "text", + "content": ", respectively. With this reconstruction, weights are adjusted to pursue a minimal disturbance for temporal features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 185, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 185, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 185, + 287, + 281 + ], + "type": "text", + "content": "Finite set calibration. To address the challenge posed by the wide span of activations within a finite set for the second inducement, we propose Finite Set Calibration (FSC) for activation quantization. This strategy employs " + }, + { + "bbox": [ + 47, + 185, + 287, + 281 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 185, + 287, + 281 + ], + "type": "text", + "content": " sets of quantization parameters for every activation within all the embedding layers and time embed, such as " + }, + { + "bbox": [ + 47, + 185, + 287, + 281 + ], + "type": "inline_equation", + "content": "\\{(s_{T},z_{T}),\\ldots ,(s_{1},z_{1})\\}" + }, + { + "bbox": [ + 47, + 185, + 287, + 281 + ], + "type": "text", + "content": " for activation " + }, + { + "bbox": [ + 47, + 185, + 287, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 185, + 287, + 281 + ], + "type": "text", + "content": ". In time-step " + }, + { + "bbox": [ + 47, + 185, + 287, + 281 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 185, + 287, + 281 + ], + "type": "text", + "content": ", the quantization function for the " + }, + { + "bbox": [ + 47, + 185, + 287, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 185, + 287, + 281 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 291, + 287, + 318 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 291, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 104, + 291, + 287, + 318 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} = \\Phi \\left(\\left\\lfloor \\frac {\\mathbf {x}}{s _ {t}} \\right\\rceil + z _ {t}, 0, 2 ^ {b} - 1\\right). \\tag {11}", + "image_path": "af7357500e32ddd06418e6252cb5245622be15de9fba8bd4a8355c2035c1f024.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 327, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 399 + ], + "type": "text", + "content": "To be noted, the calibration target is also aligned with the output of the Temporal Information Block. For a more specific estimation method of activation ranges, since the solution space within the finite set is limited, we find the Minmax [32] method can achieve satisfactory results with high efficiency (more evidence in Sec. 5.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 411, + 128, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 128, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 128, + 426 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 432, + 178, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 178, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 178, + 445 + ], + "type": "text", + "content": "5.1. Implementation details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 451, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 608 + ], + "type": "text", + "content": "Models and datasets. In this section, we conduct image generation experiments to evaluate the proposed TFMQ-DM framework on various diffusion models: pixel-space diffusion model DDPM [14] for unconditional image generation, latent-space diffusion model LDM [38] for unconditional image generation and class-conditional image generation. We also apply our work to Stable Diffusion-v1-4 for text-guided image generation. In our experiments, We use seven standard benchmarks: CIFAR-10 " + }, + { + "bbox": [ + 46, + 451, + 287, + 608 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 46, + 451, + 287, + 608 + ], + "type": "text", + "content": " [22], LSUN-Bedrooms " + }, + { + "bbox": [ + 46, + 451, + 287, + 608 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 451, + 287, + 608 + ], + "type": "text", + "content": " [48], LSUN-Churches " + }, + { + "bbox": [ + 46, + 451, + 287, + 608 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 451, + 287, + 608 + ], + "type": "text", + "content": " [48], CelebA-HQ " + }, + { + "bbox": [ + 46, + 451, + 287, + 608 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 451, + 287, + 608 + ], + "type": "text", + "content": " [18], ImageNet " + }, + { + "bbox": [ + 46, + 451, + 287, + 608 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 451, + 287, + 608 + ], + "type": "text", + "content": " [4], FFHQ " + }, + { + "bbox": [ + 46, + 451, + 287, + 608 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 451, + 287, + 608 + ], + "type": "text", + "content": " [19] and MS-COCO [25]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": "Quantization settings. We use channel-wise quantization for weights and layer-wise quantization for activations, as it is a common practice. In our experimental setup, we employ BRECQ [24] and AdaRound [31]. Drawing from empirical insights derived from conventional model quantization practices [9, 36], we maintain the input and output layers of the model in full precision. Calibration sets, integral to our methodology, are generated through full-precision" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "diffusion models, mirroring the approach outlined in Q-Diffusion [23]. Moreover, in weight quantization, we reconstruct quantized weights for 20k iterations with a minibatch size of 32 for DDPM and LDM, and 8 for Stable Diffusion. In activation quantization, we utilize EMA [16] to estimate the ranges of activations with a mini-batch size of 16 on all models. More details can be found in the appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 167, + 546, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 546, + 396 + ], + "type": "text", + "content": "Evaluation metrics. For each experiment, we evaluate the performance of diffusion models with Fréchet Inception Distance (FID) [12]. In the case of LDM and Stable Diffusion experiments, we also include sFID [40], which better captures spatial relationships than FID. For ImageNet and CIFAR-10 experiments, we additionally provide Inception Score (IS) [40] as a reference metric. Further, in the context of Stable Diffusion experiments, we extend our evaluation to include the compatibility of image-caption pairs, employing the CLIP score [11]. The ViT-B/32 is used as the backbone when computing the CLIP score. To ensure consistency in the reported outcomes, all results are derived from our implementation or from other papers, where experiments are conducted under conditions consistent with ours. More specifically, in the evaluation process of each experiment, we sample 50k images from DDPM or LDM, or 30k images from Stable-Diffusion. All experiments are conducted utilizing one H800 GPU and implemented with the PyTorch framework [34]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 406, + 391, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 391, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 391, + 418 + ], + "type": "text", + "content": "5.2. Main Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 426, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 546, + 715 + ], + "type": "text", + "content": "Unconditional image generation. In the experiments conducted on the LDM, we maintain the same experimental settings as presented in [38], including the number of steps, variance schedule, and classifier-free guidance scale (denoted by eta and CFG in the following, respectively). As shown in Tab. 2, the FID performance differences relative to the full precision (FP) model are all within 0.7 for all settings. Specifically, on the CelebA-HQ " + }, + { + "bbox": [ + 304, + 426, + 546, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 426, + 546, + 715 + ], + "type": "text", + "content": " dataset, our method exhibits a FID reduction of 6.71 and a sFID reduction of 6.60 in the w4a8 setting compared to the current state-of-the-art (SOTA). It is noticeable that existing methods, whether in 4-bit or 8-bit, show significant performance degradation when compared to the FP model on face datasets like CelebA-HQ " + }, + { + "bbox": [ + 304, + 426, + 546, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 426, + 546, + 715 + ], + "type": "text", + "content": " and FFHQ " + }, + { + "bbox": [ + 304, + 426, + 546, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 426, + 546, + 715 + ], + "type": "text", + "content": ", whereas our TFMQ-DM shows almost no performance degradation compared to the FP model. Importantly, our method achieves significant performance improvement on the LSUN-Bedrooms " + }, + { + "bbox": [ + 304, + 426, + 546, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 426, + 546, + 715 + ], + "type": "text", + "content": " in the w4a8 setting, with FID and sFID reductions of 2.26 and 7.51 compared to PTQD [10], respectively. Regarding LDM-8 on LSUN-Churches " + }, + { + "bbox": [ + 304, + 426, + 546, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 426, + 546, + 715 + ], + "type": "text", + "content": ", we attribute the moderate improvement, compared to other datasets. We believe that the use of the LDM-8 model with a downsampling factor of 8 may be more quantization-friendly. Existing methods have al" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7367" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 113, + 545, + 280 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 2. Quantization results for unconditional image generation with LDM-4 on LSUN-Bedrooms 256, FFHQ 256 and CelebA-HQ " + }, + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "{256} \\times {256}" + }, + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "type": "text", + "content": " ,LDM-8 on LSUN-Churches " + }, + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "{256} \\times {256}" + }, + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "type": "text", + "content": " . * represents our implementation according to open-source codes and †means directly rerunning open-source codes." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 113, + 545, + 280 + ], + "lines": [ + { + "bbox": [ + 48, + 113, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 48, + 113, + 545, + 280 + ], + "type": "table", + "html": "
MethodsBits (W/A)LSUN-Bedrooms 256 × 256LSUN-Churches 256 × 256CelebA-HQ 256 × 256FFHQ 256 × 256
FID↓sFID↓FID↓sFID↓FID↓sFID↓FID↓sFID↓
Full Prec.32/322.987.094.1210.898.7410.169.368.67
PTQ4DM* [41]4/324.837.944.9213.9413.6714.7211.7412.18
Q-Diffusion† [23]4/324.207.664.5511.9011.0912.0011.6010.30
PTQD* [10]4/324.427.884.6713.6811.0612.2112.0111.12
TFMQ-DM (Ours)4/323.60 (-0.60)7.61 (-0.05)4.07 (-0.48)11.41 (-0.49)8.74 (-2.32)10.18 (-1.82)9.89 (-1.71)9.06 (-1.24)
PTQ4DM* [41]8/84.759.594.8013.4814.4215.0610.7311.65
Q-Diffusion† [23]8/84.518.174.4112.2312.8514.1610.8710.01
PTQD [10]8/83.759.894.89*14.89*12.76*13.54*10.69*10.97*
TFMQ-DM (Ours)8/83.14 (-0.61)7.26 (-0.91)4.01 (-0.40)10.98 (-1.25)8.71 (-4.05)10.20 (-3.34)9.46 (-1.23)8.73 (-1.28)
PTQ4DM [41]4/820.7254.304.97*14.87*17.08*17.48*11.83*12.91*
Q-Diffusion† [23]4/86.4017.934.6613.9415.5516.8611.4511.15
PTQD [10]4/85.9415.165.10*13.23*15.47*17.38*11.42*11.43*
TFMQ-DM (Ours)4/83.68 (-2.26)7.65 (-7.51)4.14 (-0.52)11.46 (-1.77)8.76 (-6.71)10.26 (-6.60)9.97 (-1.45)9.14 (-2.01)
", + "image_path": "e404313a48862ba617f4abcd4d79dfee743f1f84c7fe069c77185e3050759eef.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 327, + 287, + 495 + ], + "blocks": [ + { + "bbox": [ + 47, + 285, + 288, + 318 + ], + "lines": [ + { + "bbox": [ + 47, + 285, + 288, + 318 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 288, + 318 + ], + "type": "text", + "content": "Table 3. Quantization results for unconditional image generation with class-conditional image generation with LDM-8 on ImageNet " + }, + { + "bbox": [ + 47, + 285, + 288, + 318 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 47, + 285, + 288, + 318 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 327, + 287, + 495 + ], + "lines": [ + { + "bbox": [ + 49, + 327, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 49, + 327, + 287, + 495 + ], + "type": "table", + "html": "
MethodsBits (W/A)ImageNet 256 × 256
IS↑FID↓sFID↓
Full Prec.32/32235.6410.917.67
PTQ4DM [41]4/32---
Q-Diffusion* [23]4/32213.5611.878.76
PTQD† [10]4/32201.7811.659.06
TFMQ-DM (Ours)4/32223.81 (+10.25)10.50 (-1.15)7.98 (-0.78)
PTQ4DM [41]8/8161.7512.59-
Q-Diffusion* [23]8/8187.6512.809.87
PTQD [10]8/8153.9211.948.03
TFMQ-DM (Ours)8/8198.86 (+11.21)10.79 (-1.15)7.65 (-0.38)
PTQ4DM [41]4/8---
Q-Diffusion* [23]4/8212.5110.6814.85
PTQD [10]4/8214.7310.4012.63
TFMQ-DM (Ours)4/8221.82 (+7.09)10.29 (-0.11)7.35 (-5.28)
", + "image_path": "c57ebc81250bdca08b82d6ac6894296d51a4c9eeb1a97bf98a74f9ea54d7474b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 287, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 584 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 584 + ], + "type": "text", + "content": "ready achieved satisfactory results on this dataset. Nonetheless, our method still approaches the performance of the FP model more closely compared to existing methods. Besides the experiments on LDM, We have also conducted experiments with DDPM on CIFAR-10 " + }, + { + "bbox": [ + 46, + 510, + 287, + 584 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 46, + 510, + 287, + 584 + ], + "type": "text", + "content": ", which can be found in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Class-conditional image generation. On the ImageNet " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " dataset, we employed a denoising process with 20 iterations, setting eta and CFG to 0.0 and 3.0, respectively. Compared to PTQD, our method achieved a FID reduction of 1.15 on both w4a32 and w8a8. Simultaneously, in the w4a8 setting, sFID decreased by 5.28. Under the same conditions, we observed an improvement of over 7 in IS. Particularly noteworthy is that, across various quantization settings, our method consistently achieved lower FID compared to the FP model." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 308, + 316, + 545, + 423 + ], + "blocks": [ + { + "bbox": [ + 305, + 285, + 545, + 308 + ], + "lines": [ + { + "bbox": [ + 305, + 285, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 545, + 308 + ], + "type": "text", + "content": "Table 4. Quantization results for text-guided image generation with Stable-Diffusion on MS-COCO captions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 316, + 545, + 423 + ], + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 423 + ], + "type": "table", + "html": "
MethodsBits (W/A)MS-COCO
FID↓sFID↓CLIP↑
Full Prec.32/3213.1519.310.3146
Q-Diffusion† [23]4/3213.5819.500.3143
TFMQ-DM (Ours)4/3213.21 (-0.37)19.03 (-0.47)0.3144 (+0.0001)
Q-Diffusion† [23]8/813.3120.540.3134
TFMQ-DM (Ours)8/813.09 (-0.22)19.91 (-0.63)0.3134 (+0.0000)
Q-Diffusion† [23]4/814.4920.430.3121
TFMQ-DM (Ours)4/813.36 (-1.13)20.14 (-0.29)0.3128 (+0.0007)
", + "image_path": "d691e5eb0c5e7ea9dc86e69f553368f352f186b757ce7d634d0fec17ede660c4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 431, + 547, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 547, + 612 + ], + "type": "text", + "content": "Text-guided image generation. In this experiment, we sample high-resolution images of " + }, + { + "bbox": [ + 304, + 431, + 547, + 612 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 431, + 547, + 612 + ], + "type": "text", + "content": " pixels with 50 denoising steps and fix CFG to the default 7.5 in Stable Diffusion as the trade-off between sample quality and diversity. In Tab. 4, compared to the current SOTA Q-Diffusion, our approach achieves an FID reduction of 1.13 on w4a8. Simultaneously, our FID on w8a8 and sFID on w4a32 are even lower than those of the full precision model. However, existing metrics fail to adequately assess the semantic consistency of generated images. Nevertheless, based on the images generated in the appendix, our method produces higher-quality images with more realistic details, better demonstrating semantic information. Furthermore, our generated images closely approximate the effects of FP model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 622, + 400, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 622, + 400, + 635 + ], + "spans": [ + { + "bbox": [ + 305, + 622, + 400, + 635 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "To evaluate the effectiveness of each proposed method, we perform a thorough ablation study on the LSUN-Bedrooms " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": " dataset with w4a8 quantization, utilizing the LDM-4 model with a DDIM sampler, as outlined in Tab. 5. We begin the assessment with a baseline BRECQ [24], which shows ineffective in denoising images when oper" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7368" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 101, + 286, + 173 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "type": "text", + "content": "Table 5. The effect of different methods proposed in the paper on LSUN-Bedrooms " + }, + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 101, + 286, + 173 + ], + "lines": [ + { + "bbox": [ + 48, + 101, + 286, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 101, + 286, + 173 + ], + "type": "table", + "html": "
MethodsBits (W/A)FID↓sFID↓
Full Prec.32/322.987.09
BRECQ [24] (Baseline)4/89.3622.73
+TIAR4/84.849.29
+FSC4/86.0711.31
+TFMQ-DM (TIAR + FSC)4/83.68 (-5.68)7.65 (-15.08)
", + "image_path": "65db997857da4b9f02d806c74c5a952b3fb4dae69f4a9be70997ed4663c0fd09.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 178, + 286, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 178, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 286, + 205 + ], + "type": "text", + "content": "ating with 4-bit quantization. Additional ablation experiments can be found in the appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 213, + 288, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 288, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 288, + 309 + ], + "type": "text", + "content": "Effect of TIAR. It can be observed that, compared to the baseline, our TIAR method reduces FID and sFID by 4.52 and 13.44, respectively. Furthermore, As shown in Fig. 4, our method's temporal feature disturbance is significantly far weaker than existing SOTA PTQD. This indicates the effectiveness of our method in maintaining temporal information contained in temporal features. Further details of the effects are presented in the appendix." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 81, + 319, + 254, + 480 + ], + "blocks": [ + { + "bbox": [ + 81, + 319, + 254, + 480 + ], + "lines": [ + { + "bbox": [ + 81, + 319, + 254, + 480 + ], + "spans": [ + { + "bbox": [ + 81, + 319, + 254, + 480 + ], + "type": "image", + "image_path": "079e6a510385420edf81ad2fe64a74293be6be63b07f66558ee8bf525313dc3c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 486, + 284, + 498 + ], + "lines": [ + { + "bbox": [ + 49, + 486, + 284, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 486, + 284, + 498 + ], + "type": "text", + "content": "Figure 4. Temporal feature errors across different PTQ methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 511, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 511, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 287, + 632 + ], + "type": "text", + "content": "Effect of FSC. Our FSC method has also achieved remarkably positive results. Compared to the baseline, it reduces FID and sFID by 3.29 and 11.42, respectively. Compared to PTQD, from Fig. 4, our method's temporal feature error is significantly smaller than that of PTQD. Since we employ layer-wise quantization for activations, this introduces less than one percent of additional parameters. The inference time overhead incurred by switching different step sizes and zero points during multiple time-steps inference is negligible, as detailed in the appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Efficiency of FSC. For FSC, there are various methods to evaluate the range of activations to determine the optimal step size. We try several methods and assess the GPU time consumed during calibration, as detailed in Tab. 6. Since the improvements achieved by these methods in model performance are similar, we opt for the simplest and most efficient" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 102, + 544, + 254 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 6. Different calibration methods for FTFC. We note " + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": " as calibration GPU hours in the table." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 102, + 544, + 254 + ], + "lines": [ + { + "bbox": [ + 308, + 102, + 544, + 254 + ], + "spans": [ + { + "bbox": [ + 308, + 102, + 544, + 254 + ], + "type": "table", + "html": "
TSC MethodsBits (W/A)FID↓sFID↓T (hours)
FIAR8/322.847.082.18
+LSQ [5]8/83.177.182.48
+KL-divergence8/83.277.3219.67
+Percentile8/83.347.4112.00
+MSE8/83.127.128.89
+Min-max [32]8/83.14 (+0.02)7.26 (+0.14)0.12 (-2.36)
FIAR4/323.047.162.20
+LSQ [5]4/83.697.482.57
+KL-divergence4/83.947.4219.65
+Percentile4/83.748.0212.04
+MSE4/83.627.488.89
+Min-max [32]4/83.68 (+0.06)7.65 (+0.23)0.12 (-2.45)
", + "image_path": "1f2a2e0fd6aab3ef99f36ba8eec0daca28a7b9e315267036a542f623994012ff.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 260, + 545, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 260, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 260, + 545, + 344 + ], + "type": "text", + "content": "Min-max [32] method as our specific calibration strategy, striking a balance between calibration time and effectiveness. Notably, we have found that PTQD and Q-Diffusion cost 4.68 and 5.29 GPU hours in their PTQ methods under w4a8 quantization on LSUN-Bedrooms " + }, + { + "bbox": [ + 304, + 260, + 545, + 344 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 260, + 545, + 344 + ], + "type": "text", + "content": ", respectively. However, our framework only spends 2.32 GPU hours (as shown in Tab. 6)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 344, + 546, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 546, + 453 + ], + "type": "text", + "content": "Furthermore, for the only learning-based method in the table, LSQ [5], which is one of the most commonly used methods in previous works [10, 23, 41], we observe that it did not outperform other methods and, in some cases, performs even worse. We speculate that the main reason might be that, for a fixed time-step, the calibration data used to determine the quantization parameters is relatively limited compared to all calibration data. In such cases, LSQ may struggle to learn an optimal set of parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 463, + 528, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 463, + 528, + 477 + ], + "spans": [ + { + "bbox": [ + 305, + 463, + 528, + 477 + ], + "type": "text", + "content": "6. Conclusion & Discussions of Limitations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 483, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 602 + ], + "type": "text", + "content": "This research explores the application of quantization for accelerating diffusion models. In this work, we identify a novel and significant problem, namely temporal feature disturbance, in the quantization of diffusion models. We conducted a detailed analysis of the root causes of this problem and introduced our TFMQ-DM quantization framework. In 4-bit quantization on extensive datasets and different diffusion models, this framework exhibits minimal performance degradation compared to the FP model and speedup quantization time." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 603, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 603, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 545, + 714 + ], + "type": "text", + "content": "However, besides temporal features, we have also found the textual features introduced in Stable Diffusion have a physical meaning and influence the generation effect. Nevertheless, these textual features are not taken into consideration in current methods. Secondly, temporal feature maintenance is suitable for both PTQ and QAT scenarios. But in this study, our main focus is on the PTQ setting and achieving 4-bit quantization. The extension to the QAT setting should be conducted in the future to achieve lower-bit quantization and further performance enhancements." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7369" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 94, + 107, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 94, + 107, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 94, + 107, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 114, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 114, + 287, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 287, + 157 + ], + "type": "text", + "content": "[1] Yash Bhalgat, Jinwon Lee, Markus Nagel, Tijmen Blankevoort, and Nojun Kwak. " + }, + { + "bbox": [ + 53, + 114, + 287, + 157 + ], + "type": "inline_equation", + "content": "\\mathrm{Lsq}+" + }, + { + "bbox": [ + 53, + 114, + 287, + 157 + ], + "type": "text", + "content": ": Improving low-bit quantization through learnable offsets and better initialization, 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 267 + ], + "type": "text", + "content": "[2] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "text", + "content": "[3] Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In CVPR, pages 12413-12422, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 314, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 288, + 357 + ], + "type": "text", + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 358, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 390 + ], + "type": "text", + "content": "[5] Steven K. Esser, Jeffrey L. McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S. Modha. Learned step size quantization. In ICLR, 2020. 1, 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 392, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 445 + ], + "type": "text", + "content": "[6] Giulio Franzese, Simone Rossi, Lixuan Yang, Alessandro Finamore, Dario Rossi, Maurizio Filippone, and Pietro Michiardi. How much is enough? a study on diffusion times in score-based generative models. arXiv preprint arXiv:2206.05173, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "type": "text", + "content": "[7] Ruihao Gong, Xianglong Liu, Shenghu Jiang, Tianxiang Li, Peng Hu, Jiazhen Lin, Fengwei Yu, and Junjie Yan. Differentiable soft quantization: Bridging full-precision and low-bit neural networks. In ICCV, pages 4852–4861, 2019. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 491, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 287, + 535 + ], + "type": "text", + "content": "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 536, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 536, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 536, + 287, + 578 + ], + "type": "text", + "content": "[9] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2015. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 580, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 613 + ], + "type": "text", + "content": "[10] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. In NeurIPS, 2023. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "text", + "content": "[11] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning, 2022. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[13] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 97, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 547, + 140 + ], + "type": "text", + "content": "[15] Itay Hubara, Yury Nahshan, Yair Hanani, Ron Banner, and Daniel Soudry. Improving post training neural quantization: Layer-wise calibration and integer programming. arXiv preprint arXiv:2006.10518, 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 142, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 142, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 142, + 547, + 196 + ], + "type": "text", + "content": "[16] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In CVPR, pages 2704-2713, 2018. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 198, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 232 + ], + "type": "text", + "content": "[17] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 234, + 545, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 234, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 545, + 265 + ], + "type": "text", + "content": "[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation, 2018. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 267, + 545, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 267, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 307, + 267, + 545, + 299 + ], + "type": "text", + "content": "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks, 2019. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 302, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 302, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 302, + 545, + 323 + ], + "type": "text", + "content": "[20] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 326, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 326, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 326, + 545, + 357 + ], + "type": "text", + "content": "[21] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. arXiv preprint arXiv:2106.00132, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 360, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 382 + ], + "type": "text", + "content": "[22] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 384, + 545, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 384, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 307, + 384, + 545, + 427 + ], + "type": "text", + "content": "[23] Xiuyu Li, Long Lian, Yijiang Liu, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In ICCV, 2023. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 430, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 430, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 307, + 430, + 545, + 472 + ], + "type": "text", + "content": "[24] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. BRECQ: pushing the limit of post-training quantization by block reconstruction. In ICLR, 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "type": "text", + "content": "[26] Yang Lin, Tianyu Zhang, Peiqin Sun, Zheng Li, and Shuchang Zhou. Fq-vit: Post-training quantization for fully quantized vision transformer. In *IJCAI*, pages 1173–1179, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 566, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 566, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 307, + 566, + 545, + 597 + ], + "type": "text", + "content": "[27] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In ICLR, 2022. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 600, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 600, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 600, + 545, + 632 + ], + "type": "text", + "content": "[28] Christos Louizos, Matthias Reisser, Tijmen Blankevoort, Efstratos Gavves, and Max Welling. Relaxed quantization for discretized neural networks. In ICLR, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[29] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[30] Zhaoyang Lyu, Xudong Xu, Ceyuan Yang, Dahua Lin, and Bo Dai. Accelerating diffusion models via early stop of the diffusion process. arXiv preprint arXiv:2205.12524, 2022. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "7370" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[31] Markus Nagel, Rana Ali Amjad, Mart van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. In ICML, pages 7197-7206, 2020. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "type": "text", + "content": "[32] Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bondarenko, Mart van Baalen, and Tijmen Blankevoort. A white paper on neural network quantization, 2021. 1, 5, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 195 + ], + "type": "text", + "content": "[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 250 + ], + "type": "text", + "content": "[34] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 252, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 252, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 252, + 287, + 274 + ], + "type": "text", + "content": "[35] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 318 + ], + "type": "text", + "content": "[36] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. In ECCV, pages 525-542, 2016. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 319, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 352 + ], + "type": "text", + "content": "[37] Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. Fastspeech 2: Fast and high-quality end-to-end text to speech, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 354, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 386 + ], + "type": "text", + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 388, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 419 + ], + "type": "text", + "content": "[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 422, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 454 + ], + "type": "text", + "content": "[40] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans, 2016. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 456, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 487 + ], + "type": "text", + "content": "[41] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In CVPR, 2023. 2, 3, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 489, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 553 + ], + "type": "text", + "content": "[42] Jonathan Shen, Ruoming Pang, Ron J. Weiss, Mike Schuster, Navdeep Jaitly, Zongheng Yang, Zhifeng Chen, Yu Zhang, Yuxuan Wang, RJ Skerry-Ryan, Rif A. Saurous, Yannis Agiomyrgiannakis, and Yonghui Wu. Natural tts synthesis by conditioning wavenet on mel spectrogram predictions, 2018. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 556, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 588 + ], + "type": "text", + "content": "[43] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for diffusion models. In NeurIPS, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 590, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 611 + ], + "type": "text", + "content": "[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1, 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "type": "text", + "content": "[45] Changyuan Wang, Ziwei Wang, Xiuwei Xu, Yansong Tang, Jie Zhou, and Jiwen Lu. Towards accurate data-free quantization for diffusion models. arXiv preprint arXiv:2305.18723, 2023. 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "text", + "content": "[46] Peisong Wang, Qiang Chen, Xiangyu He, and Jian Cheng. Towards accurate post-training network quantization via bitsplit and stitching. In ICML, pages 9847-9856, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[47] Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. Qdrop: Randomly dropping quantization for" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 353 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "text", + "content": "extremely low-bit post-training quantization. In ICLR, 2022. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[48] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop, 2016. 1, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 141, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 183 + ], + "type": "text", + "content": "[49] Luoming Zhang, Yefei He, Zhenyu Lou, Xin Ye, Yuxing Wang, and Hong Zhou. Root quantization: a self-adaptive supplement ste. Applied Intelligence, 53(6):6266-6275, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "text", + "content": "[50] Qinsheng Zhang, Molei Tao, and Yongxin Chen. gddim: Generalized denoising diffusion implicit models. arXiv preprint arXiv:2206.05564, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 220, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 285 + ], + "type": "text", + "content": "[51] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. Opt: Open pre-trained transformer language models, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "text", + "content": "[52] Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Truncated diffusion probabilistic models. stat, 1050:7, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "type": "text", + "content": "[53] Bohan Zhuang, Chunhua Shen, Mingkui Tan, Lingqiao Liu, and Ian Reid. Towards effective low-bitwidth convolutional neural networks. In CVPR, pages 7920-7928, 2018. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7371" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/da9cf0d5-8d8b-4df8-b396-af5ff1ab4d0d_content_list.json b/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/da9cf0d5-8d8b-4df8-b396-af5ff1ab4d0d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..06e25e62df6066fa4912bbbaeeb51c88c70731bc --- /dev/null +++ b/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/da9cf0d5-8d8b-4df8-b396-af5ff1ab4d0d_content_list.json @@ -0,0 +1,1588 @@ +[ + { + "type": "text", + "text": "THRONE: An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models", + "text_level": 1, + "bbox": [ + 173, + 130, + 792, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prannay Kaul $^{1*}$ Zhizhong Li $^{2\\dagger}$ Hao Yang $^{2}$ Yonatan Dukler $^{2}$ Ashwin Swaminathan $^{2}$ C. J. Taylor $^{2}$ Stefano Soatto $^{2}$ VGG, University of Oxford $^{1}$ AWS AI Labs $^{2}$", + "bbox": [ + 236, + 184, + 732, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "prannay@robots.ox.ac.uk {lzhizhon,haoyng,dukler,swashwin,taylorcj,soattos}@amazon.com", + "bbox": [ + 101, + 241, + 857, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mitigating hallucinations in large vision-language models (LVLMs) remains an open problem. Recent benchmarks do not address hallucinations in open-ended free-form responses, which we term \"Type I hallucinations\". Instead, they focus on hallucinations responding to very specific question formats—typically a multiple-choice response regarding a particular object or attribute—which we term \"Type II hallucinations\". Additionally, such benchmarks often require external API calls to models which are subject to change. In practice, we observe that a reduction in Type II hallucinations does not lead to a reduction in Type I hallucinations but rather that the two forms of hallucinations are often anti-correlated. To address this, we propose THRONE, a novel object-based automatic framework for quantitatively evaluating Type I hallucinations in LVLM free-form outputs. We use public language models (LMs) to identify hallucinations in LVLM responses and compute informative metrics. By evaluating a large selection of recent LVLMs using public datasets, we show that an improvement in existing metrics do not lead to a reduction in Type I hallucinations, and that established benchmarks for measuring Type I hallucinations are incomplete. Finally, we provide a simple and effective data augmentation method to reduce Type I and Type II hallucinations as a strong baseline.", + "bbox": [ + 75, + 323, + 472, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 76, + 703, + 218, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper proposes a benchmark to evaluate hallucinations by large vision-language models (LVLMs) when generating free-form responses, specifically detailed descriptions, based on a given image.", + "bbox": [ + 75, + 726, + 468, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid advancement in large language models (LLMs) [52] has pushed the development of large vision-language models (LVLMs). [1, 6, 7, 17, 24, 25, 30, 36, 44, 48, 54] LVLMs take input text and images and generate text responses to enable multi-modal perception and com", + "bbox": [ + 75, + 787, + 470, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "prehension.", + "bbox": [ + 500, + 292, + 578, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LVLMs are largely built on LLMs and therefore inherit both their advantages and their disadvantages. LLMs have been shown to produce hallucinations [47, 51], generated text responses that are coherent and plausible but factually incorrect. LVLMs echo this behavior with generated text contradicting with the visual or text input [53]. Hallucinations prevent the use of LVLMs in safety-critical situations and therefore evaluating and mitigating hallucinations in LVLMs is crucial for their deployment in such settings. Determining the presence and cause of hallucinations in LVLMs remains an open question [45, 53].", + "bbox": [ + 496, + 308, + 893, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We divide LVLM hallucinations into two types. Type I hallucinations occur in response to open-ended questions with a very large set of possible responses — e.g. What is happening in this image?. Type II hallucinations are incorrect responses to a factual question regarding a specific concept about the image with a fixed set of options such as yes/no—e.g. Is there a traffic light in this image? Fig. 3 illustrates the difference between these two types of hallucination. Reducing hallucinations in both cases is required for useful, multi-purpose LVLMs. However, later in Fig. 4 we observe that the same LVLM can give contradicting answers when being evaluated for Type I vs. Type II hallucinations. This implies that measuring and reducing one type does not necessarily reduce the other.", + "bbox": [ + 496, + 476, + 893, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing works to evaluate LVLMs often avoid direct quantification of hallucinations and instead develop comprehensive benchmarks that judge various other desirable abilities such as: optical character recognition, fine-grained recognition and attribute detection [14, 22, 32]. The extent of hallucinations in these benchmarks is obfuscated, since it is only one of many factors influencing other metrics. It requires human effort to inspect individual predictions. There are two major established works which specifically develop a benchmark for evaluating hallucinations in vision-language models: POPE [26] and CHAIR [40], which we discuss in detail in Sec. 2. However, we observe they both have shortcomings in effectively evaluating hallucinations:", + "bbox": [ + 496, + 688, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work conducted during an internship at Amazon", + "bbox": [ + 94, + 875, + 359, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27228", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/086844dd337d4200da7236eb37d1d27179d6b14b2daac80a4009bc170907dc43.jpg", + "image_caption": [ + "Figure 1. THRONE (Ours): LVLMs are prompted with a concept neutral instruction. An external LM performs abstractive QA on the response to establish the existence of Type I hallucinations." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 477, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "POPE [26] is a recent work addressing Type II hallucinations with respect to object classes. However, we find Type I and Type II hallucinations are disconnected, and that POPE gives an incomplete picture on LVLM hallucinations. Moreover, POPE systematically under-samples negative object categories leading to a large underestimation of Type II hallucinations (see Sec. 5.4).", + "bbox": [ + 75, + 284, + 468, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CHAIR [40] does address Type I hallucinations—establishing object category hallucination in short image captions using simple text matching. However, CHAIR is not suited to current LVLMs because the simple text matching it employs cannot comprehend abstract or hypothetical concepts present in today's LVLMs (see Fig. 4). Further, hand-crafted rules for each set of classes are required for usable text matching, and trivial model answers can attain a perfect CHAIR score.", + "bbox": [ + 75, + 392, + 468, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the issues of current LVLM hallucination benchmarks, we propose THRONE (Text-from-image Hallucination Recognition with Object-probes for open-ended Evaluation). THRONE leverages language models (LMs) to evaluate Type I hallucinations in free-form, open-ended image descriptions with respect to a pre-defined object vocabulary of interest. By utilizing LMs, THRONE is able to accurately judge whether an object mentioned in an LVLM response is implied to exist in the image or is abstractly mentioned with no implication about its existence (see Fig. 4).", + "bbox": [ + 75, + 532, + 468, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, in THRONE, we provide easy access to our benchmark, by leveraging open-source LMs that can run on common GPUs, instead of relying on closed-source commercial models [35] that are subject to arbitrary change, as done in other works [4, 22, 32]. Through combining multiple open-source LMs, we mitigate any single-model biases in judging hallucinations when calculating Type-I hallucination scores with THRONE.", + "bbox": [ + 75, + 700, + 468, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We make four contributions: first, we establish an accurate and accessible benchmark to quantitatively evaluate object hallucinations in free-form responses, leveraging LMs to judge the existence of Type I hallucinations—quantitatively showing half the judgement errors of CHAIR;", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d0ffd45c22381ac0c708d6e5c516b7ac37149eda31e5a621be3c195e88db01d0.jpg", + "image_caption": [ + "Figure 2. POPE: Questions with specific concepts prompt an LVLM directly to evaluate Type II hallucinations [26]. Handcrafted rules parse LVLM responses to give yes/no labels." + ], + "image_footnote": [], + "bbox": [ + 490, + 88, + 890, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "second, we evaluate a number of current LVLMs on THRONE and demonstrate that observed progress in reducing Type II hallucinations do not translate to a corresponding reduction in Type I hallucinations; third, we show a recent method, POPE, is inaccurately capturing the extent of Type II hallucination discovery due to its sampling strategy; we mitigate this issue in our implementation of THRONE while presenting results for a complete version of POPE; and fourth, we provide a simple augmentation of visual instruction tuning data which significantly improves performance on Type I hallucinations while maintaining or improving Type II hallucination performance.", + "bbox": [ + 496, + 284, + 893, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 500, + 481, + 651, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hallucination Benchmarks for LVLMs. In response to the development of LVLMs, few evaluation benchmarks focusing on hallucinations have been introduced. CHAIR [40], one of the first works to assess hallucinations, is designed for short image captions. CHAIR uses a fixed set of object classes (extended with their synonyms) as a set of text strings to find predicted object classes in image captions via exact text matching. Subsequently, each class matched in the caption is compared to the COCO [27] captions and bounding box annotations to establish object hallucinations. CHAIR was developed prior to current instruction-tuned LVLMs which produce long free-form responses ( $10 \\times$ longer than image captions) with a diverse vocabulary, limiting its applicability to modern LVLMs. As shown in Fig. 3, exact text matching in CHAIR is prone to incorrectly matching the vocabulary classes with abstract concepts in the free-form response and the and synonyms of the class list must be manually selected to prevent confusion during evaluation. For example, in CHAIR the word \"chair\" will match all responses for the phrase \"toilet seat\", because the exact text matching classifies \"seat\" as the COCO class \"chair\". Finally, CHAIR metrics compare the overall number of predicted objects to the overall number of predicted objects judged to be hallucinations—ignoring the recall of ground-truth objects and the distribution of object classes. This means that a single correct", + "bbox": [ + 496, + 507, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27229", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "prediction across the entire evaluation dataset along with a generic response otherwise e.g. A natural scene, achieves a perfect score $(\\frac{0 \\text{ hallucinated objects}}{1 \\text{ predicted objects}} = 0.0)$ . See the Supplementary Material for a full overview of CHAIR. In contrast, THRONE uses pre-trained LMs that go beyond direct synonym matching, to automatically judge the existence of concepts and hallucinations in free-form responses. In addition, our method considers both recall and precision to yield a holistic benchmark and does not require any manual curation of synonyms. THRONE and CHAIR both evaluate Type I hallucinations—hallucinations in response to concept-neutral prompts e.g. Describe this image in detail.", + "bbox": [ + 75, + 90, + 472, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "POPE [26] is a recently proposed benchmark to evaluate object hallucinations in LVLMs—specifically Type II hallucinations, in which an LVLM is directly queried with a yes-no question regarding the existence of a particular object of the form: Is there $\\{\\mathrm{a / an}\\}$ {object_class_name} in the image?. The LVLM response is parsed using simple rules to determine whether a Type II hallucination has occurred. Precision and recall metrics are compiled using the parsed LVLM responses and the ground-truth annotation data. Despite focusing on measuring hallucinations, POPE only queries an LVLM with 3 positive and 3 sampled negative questions per COCO image i.e. the evaluation is artificially balanced. This means many potential hallucinations with respect to the COCO categories are not captured by their method. In Sec. 5.4, we show POPE dramatically underestimates Type II hallucinations and present the results of a complete version. Our method, THRONE, evaluates the prevalence of Type I hallucinations, which we observe are disjoint to Type II hallucinations.", + "bbox": [ + 75, + 289, + 472, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Comprehensive Benchmarks for LVLMs have recently grown in number. MMBench [32] and MM-Vet [49] assess various aspects of LVLM performance such as: color perception, celebrity recognition, and numerical calculation. However many of these works integrate evolving APIs which are modified often (or even discontinued) and are inherently stochastic. Over time this greatly reduces the consistency of these benchmarks. Exceptions are MME [14] and SEED-Bench [22], but the impact of Type II hallucinations on final metrics is conflated with a number of other aspects of model performance. Our method, THRONE, directly addresses Type I hallucinations, only making use of open-source language models and datasets.", + "bbox": [ + 75, + 580, + 472, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Vision Language Models (LVLMs) have rapidly developed by harnessing advancements in large language models (LLM) [5, 39, 43] and by directly integrating pretrained LLMs into their architectures. In contrast to earlier vision-language models such as CLIP [18, 38], LVLMs are generally comprised of a pretrained LLM and image encoder, aligned with a connector module of varying complexity. Some works are highlighted here. Frozen [44] is", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/dbc5b80114e6f8da66eecc2b07e061e2a54e4fdc1a63e8677464de6974aafa7b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ImagePromptResponse
What is happening in this image?In the image, a man is hanging clothes on a clothesline that is attached to the back of a moving car.
Is there a traffic light in this image?Yes, there is a traffic light in the image, and the man is standing on the back of a car near it.
", + "bbox": [ + 501, + 89, + 893, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 3. Type I vs. Type II Hallucinations: (Top) LVLMs prompted with concept-neutral instructions produce Type I hallucinations. (Bottom) Instructions specifying a concept produce Type II hallucinations. Examples from LLaVA-v1.5 [29].", + "bbox": [ + 496, + 250, + 893, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "an early work fine-tuning the vision encoder to dynamically prefix the prompt to a frozen LLM. Flamingo [1] combines visual and language features using cross-attention layers in an otherwise frozen LLM. BLIP-2 [25] uses a frozen image encoder to learn a Querying-Transformer (Q-Former) on image-text pairs that is used as the connector. This architecture is adapted in [7] for dialogue via training with visual instruction tuning. LLaVA [30] uses COCO [27] annotations and GPT-4 [35] to generate visual instruction tuning data in a plain-text pipeline. Combining this generated data with standard VQA datasets (VQAv2 etc.) further boosts performance [29]. Different works modify the training approach by using efficient adaptation [16, 48] multiple training stages [3] or introduce the use of discrete tokens for localization [36]. We note however, that most of these works evaluate performance on traditional vision-language datasets like VQAv2 [2], which do not consider the extent of hallucinations, a known problem with LLMs [47, 51].", + "bbox": [ + 496, + 316, + 893, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 THRONE", + "text_level": 1, + "bbox": [ + 500, + 602, + 625, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recap of existing methods. POPE [26] and other benchmarks (MME [14], MMBench [32]) directly query LVLMs with a restricted desired answer space, e.g. yes-no (MME, POPE) and A-B-C-D multiple choice (MMBench), as shown in Fig. 2. These benchmarks only consider such short answer formats, whereas a key quality of LVLMs is in their ability to generate free-form coherent text. Moreover, POPE, which addresses Type II hallucinations, undersamples negative classes meaning hallucinations are dramatically underestimated (see Sec. 5.4 and Fig. 7). In contrast, we skip class subsampling and enumerate all classes for every image, ensuring a full evaluation of Type I hallucinations of the ground-truth classes.", + "bbox": [ + 496, + 627, + 890, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CHAIR [40] also evaluates Type I hallucinations, but was developed when typical vision-language models could only generate short and simple captions similar to those in COCO Captions [12]. Moreover, it lacks accurate comprehension of responses (see the right side of Fig. 4) and ig-", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27230", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/634cddb0c8f4c8f79f8782ffb0b14a0cc478117e417498c07dc7450b376f3331.jpg", + "image_caption": [ + "Figure 4. A Comparison of POPE, CHAIR and THRONE: Directly querying LVLMs for object existence (person, banana etc.) using concept-specific instructions, as in POPE (bottom left), does not produce the same hallucinations as using concept-neutral instructions (right). We highlight the Type I hallucinations in orange. CHAIR relies on exact text matching to a fixed set of objects and synonyms, thus incorrectly labels \"customers\" and \"shoppers\" as hallucinations, highlighted in red. THRONE is designed for the rich vocabulary and the free-form generations of modern LVLMs by harnessing LMs to establish object existence. By using an LM to pass judgement, our evaluation correctly captures \"customers\" and \"shoppers\" as hypothetical content in the free-form generation." + ], + "image_footnote": [], + "bbox": [ + 168, + 88, + 380, + 207 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/67c6795603d6391106d156263b0b685600a7eadf06269ce452f093f6b4d82376.jpg", + "table_caption": [ + "Type II Hallucination Evaluation (POPE)" + ], + "table_footnote": [], + "table_body": "
Instruction [GT]Response [Prediction]
Is there... \na person [No] \na banana [Yes] \nan apple [No] \nan orange [No] \n...in the image?No, there is no person in the image. [No] \nYes, there is a banana in the image. [Yes] \nNo, there is not an apple in the image. [No] \nNo, there is no orange in the image. [No] \nNo Type II Hallucinations Found
", + "bbox": [ + 80, + 219, + 468, + 314 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/fc94d9cdcb8fd795548283e02c2c5aef72cbd5016af2e605647e3486326c3133.jpg", + "table_caption": [ + "Type I Hallucination Evaluation" + ], + "table_footnote": [], + "table_body": "
InstructionResponse
Describe this image in detail.The image showcases a fruit stand at a grocery store, featuring a variety of fruits on display. There are several bunches of bananas, with some placed in the foreground and others in the background. The bananas are arranged in different sections, creating an appealing presentation for customers. In addition to the bananas, there are also apples and oranges on display. The apples are located towards the left side of the image, while the oranges are placed in the middle and right side of the stand. The fruits are well-organized and presented in an attractive manner, making it an inviting sight for shoppers. Type I Hallucinations Present and Found
MSCOCO Object Prediction from Description
HumanCHAIRTHRONE (Ours)
banana apple orangebanana person apple orangebanana apple orange
", + "bbox": [ + 475, + 99, + 888, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "nores the recall of ground-truth objects. In Sec. 5.5, we describe quantitative evaluations, using a human oracle, which demonstrate THRONE halves the rate of hallucination misjudgement in CHAIR. See Sec. 2 and the Supplementary Material for more details on CHAIR and its shortcomings. Fig. 4 shows an overview of the three aforementioned methods: POPE, CHAIR and our method, THRONE.", + "bbox": [ + 75, + 417, + 468, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Evaluating Hallucinations with THRONE", + "text_level": 1, + "bbox": [ + 76, + 532, + 434, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address these limitations, we propose a framework, THRONE, shown in Fig. 1, to evaluate the prevalence of Type I hallucinations in LVLM responses conditioned on an image and a neutral text prompt.", + "bbox": [ + 75, + 551, + 468, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each image in a labeled dataset, $\\mathcal{I}$ , addressing a set of classes, $\\mathcal{C}$ , the LVLM is queried with the same instruction: Describe this image in detail., regardless of image content. The LVLM response, which is expected to be long free-form text containing an image description, is generated and stored. Next, a publicly available, open-source, external language model (LM) performs abstractive question answering (AQA) using the LVLM response as context and a question of the form: Please answer yes or no. Is there $\\{\\mathrm{a / an}\\}$ {object class name} in this image? or similar, for every class in $\\mathcal{C}$ (right side of Fig. 1). By selecting an appropriate LM and using a simple prompt template (see Sec. 4 for specific details), we ensure the AQA response is either yes or no—our method does not require any additional parsing. This is in contrast to other works which require added parsing or interpretation by a closed-source model.", + "bbox": [ + 75, + 613, + 468, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After performing AQA on each response generated by the LVLM for every class in $\\mathcal{C}$ , we obtain an array of pre", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dicted labels:", + "bbox": [ + 500, + 419, + 589, + 431 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Y}} \\in \\{0, 1 \\} ^ {| \\mathcal {I} | \\times | \\mathcal {C} |} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 431, + 890, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $0/1$ indicates a negative/positive existence judgement by the LM with respect to the relevant LVLM response.", + "bbox": [ + 496, + 460, + 890, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly using the ground-truth data for $\\mathcal{I}$ , an array of ground-truth labels can be constructed:", + "bbox": [ + 498, + 491, + 890, + 520 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} \\in \\{0, 1 \\} ^ {| \\mathcal {I} | \\times | \\mathcal {C} |} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 522, + 890, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using these two arrays, we calculate four metrics: (1) Overall Precision, $P_{\\mathrm{ALL}}$ ; (2) Overall Recall, $R_{\\mathrm{ALL}}$ ; (3) Class-wise Precision, $P_{\\mathrm{CLS}}$ ; and (4) Class-wise Recall, $R_{\\mathrm{CLS}}$ . Overall metrics are calculated in a class-agnostic manner. Class-wise metrics are calculated in a class-conscious manner by computing precision and recall for each category separately and then averaging. This follows common practice in object detection and instance segmentation [10, 27].", + "bbox": [ + 496, + 550, + 890, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "False positives in LVLMs reduce precision and are dominated by hallucinations—precision indicates the extent of Type I hallucinations in LVLM responses. The recall metrics inform the level of class coverage by an LVLM when producing image descriptions. The class-wise metrics give a general measure of performance as the overall metrics are skewed towards the most common categories. A common way to combine precision, $P$ , and recall, $R$ , metrics is through the generalized F score, $F_{\\beta}$ :", + "bbox": [ + 496, + 686, + 890, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {\\beta} = \\left(1 + \\beta^ {2}\\right) \\cdot \\frac {P \\cdot R}{\\left(\\beta^ {2} \\cdot P\\right) + R} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 830, + 890, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prior work such as POPE [26], use the common balanced F-score (or $F^1$ -score) which equally weights precision and", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27231", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "recall. However, given that THRONE is concerned with measuring hallucinations and is therefore particularly interested in precision, we choose $\\beta = 0.5$ or the $F^{0.5}$ -score, thus weighing precision twice as important as recall. The $F^{0.5}$ -score is commonly used in pandemic misinformation filters [37], recommender systems [34], active stock selection [41] and other areas where false positives are costlier than false negatives. From this we can calculate the overall and class-wise $F^{0.5}$ -scores, $F_{\\mathrm{ALL}}^{0.5}$ and $F_{\\mathrm{CLS}}^{0.5}$ , respectively. To mitigate against class imbalance issues, we use $F_{\\mathrm{CLS}}^{0.5}$ as the principle metric of comparison between LVLMs in THRONE.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Ensuring Robustness via Ensembling", + "text_level": 1, + "bbox": [ + 76, + 290, + 408, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Any LM used for AQA in THRONE may misjudge Type II hallucinations—no LM is perfect. Fig. 5 (top) shows two different variants from the same model family (FLANT5 [33]), yielding opposite responses when prompted with the same response and question. Moreover, an LM may yield different answers to semantically identical questions, despite conditioning on the same response—shown in Fig. 5 (bottom). To ensure THRONE is robust to spurious performance by any one LM, we ensemble various LMs and semantically equivalent question formats. The use of $N$ distinct LMs and $M$ distinct question formats yields $NM$ answers for each (LVLM response, class) combination. This set of $NM$ answers is combined based on equal voting by each (LM, question) pair to \"elect\" the predicted answer. Continuing the notation from Eq. (1), stacking predictions from each (LM, question) pair yields a 3D array of predicted labels: $\\bar{\\mathbf{Y}} \\in \\{0,1\\}^{|\\mathcal{I}_{\\mathrm{OURS}}| \\times |\\mathcal{C}_{\\mathrm{OURS}}| \\times NM}$ . To combine the answers from each (LM, question) pair via voting, we require agreement between at least $k$ answers. Where sufficient agreement does not exist we introduce an \"ignore\" label. Mathematically, the elements in $\\hat{\\mathbf{Y}}$ (from Eq. (1)) are calculated as:", + "bbox": [ + 75, + 313, + 472, + 648 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Y} _ {i, j} = \\left\\{ \\begin{array}{l l} 0, & \\sum_ {k = 1} ^ {N M} \\bar {Y} _ {i, j, k} \\leq (N M - k) \\\\ 1, & \\sum_ {k = 1} ^ {N M} \\bar {Y} _ {i, j, k} \\geq k \\\\ - 1, & \\text {o t h e r w i s e} \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 659, + 415, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where an \"ignore\" label exists in $\\bar{\\mathbf{Y}}$ , it is removed from the calculations of $P_{\\mathrm{ALL}}$ , $R_{\\mathrm{ALL}}$ , $P_{\\mathrm{CLS}}$ , and $R_{\\mathrm{CLS}}$ , as we cannot be confident in the AQA process for that particular (LVLM response, class) combination. The choice of $k$ reflects the desired level of confidence. We make use of a unanimous voting mechanism ( $k = NM$ )—use the prediction only if all (LM, question) pairs agree, otherwise ignore. Human evaluation of our benchmark and choice of voting mechanism is found in the Supplementary Material. Once we have applied this voting mechanism, we can calculate the metrics described at the end of Sec. 3.1.", + "bbox": [ + 75, + 733, + 472, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9839b4821e70a308c64e94e2823eca3305e1faa403fb063339a77b55ea633175.jpg", + "image_caption": [ + "Figure 5. AQA Ensembling in Evaluation: Using different LMs or different prompts when running AQA on LVLM generated responses can produce opposing answers to identical prompts or identical LMs. To ensure THRONE is robust to this, we ensemble multiple LMs and multiple prompts in our evaluation pipeline." + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 880, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Implementation", + "text_level": 1, + "bbox": [ + 498, + 295, + 669, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide details of our framework regarding the selection of public datasets, public LMs and LM prompts.", + "bbox": [ + 496, + 318, + 890, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Dataset", + "text_level": 1, + "bbox": [ + 498, + 354, + 604, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Any proper evaluation of object hallucinations (a type of false positive error) requires knowing, with certainty, which classes are absent in an image. In our benchmark, we use COCO [27] for a number of reasons: (1) its annotations of 80 categories are exhaustive at an image level (image-level recall $\\approx 99\\%$ [27])—if there are many book instances in a COCO image, at least one is annotated with bounding boxes; (2) many LVLMs are partly trained on COCO data and so should be familiar with the set of categories; (3) its images generally contain complex scenes suitable for generating long free-form descriptions unlike image recognition datasets like ImageNet [8].", + "bbox": [ + 496, + 375, + 890, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize the validation set of COCO 2017, which contains $|\\mathcal{I}| = 5000$ images and $|\\mathcal{C}| = 80$ categories. Using the single LVLM text prompt Describe this image in detail., we generate 5000 responses. As we query each LVLM response for each category in $\\mathcal{C}$ , a single LM performs AQA $|\\mathcal{I}| \\times |\\mathcal{C}| = 400k$ times across the LVLM responses—one instance of AQA per (image response, class) pair. In Sec. 5.3, we also present results using Objects365 [42] which is rarely used in LVLM training.", + "bbox": [ + 496, + 556, + 890, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2 Language Models", + "text_level": 1, + "bbox": [ + 498, + 699, + 683, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To assess Type I hallucinations in an LVLM response using THRONE, we require a language model (LM) which can answer questions on the existence of object categories based on the LVLM response. MMBench [32] makes use of a ChatGPT model to identify multiple choice answer selections and still reports mistakes. In our experience, some LMs give rather incoherent judgements when used to assess hallucinations when the prompt is changed (see supplemental material). Therefore for THRONE, we choose FLAN-T5 models [33, 39]. We make this choice because FLAN-T5 model family: (1) have undergone instruction tuning with thousands of tasks [33]; (2) are open-source and", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27232", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/70dd6100dc7ef3fcbf7658d68e8ef1293b48be008f8020e0991b96e9512a8cc8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelL\\( P_{ALL} \\)\\( R_{ALL} \\)\\( F_{ALL}^{1} \\)\\( F_{ALL}^{0.5} \\)\\( P_{CLS} \\)\\( R_{CLS} \\)\\( F_{CLS}^{1} \\)\\( F_{CLS}^{0.5} \\)
Adapter-v2 [15]51463.673.368.165.368.270.669.468.7
Adapter-v2.1 [15]51263.873.768.465.567.471.269.368.1
InstructBLIP [7]52570.874.372.571.577.271.974.576.1
Otter-Image [23]25733.031.232.132.725.216.920.222.9
MiniGPT4 [54]47381.759.869.076.179.961.869.775.5
MiniGPT-v2 [6]38179.066.672.376.277.667.071.975.2
mPLUG-Owl [48]55555.571.962.658.166.368.367.366.7
LRV-Instruction-v2 [28]10382.056.767.075.378.458.867.273.5
LLaVA-v1.3* [30]53280.565.272.176.979.965.371.976.5
LLaVA-v1.5 [29]50968.161.064.466.669.956.462.566.8
LLaVA-Mistral [19, 31]52486.871.878.383.684.464.270.877.5
", + "bbox": [ + 78, + 89, + 468, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. THRONE Results with COCO for a selection of instruction-tuned LVLMs. We select $F_{\\mathrm{CLS}}^{0,5}$ as the principal metric for evaluation in our benchmark to balance across classes and to prioritize precision (which reflects the extent of hallucination) over recall. Best and second-best performance are denoted by blue and red, respectively. *Our implementation using official code to enable fair comparison. $L$ corresponds to the median response length (measured in # of characters).", + "bbox": [ + 75, + 215, + 472, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "therefore accessible to the community; (3) can fit locally on a single GPU for the models we consider; (4) follow user's instruction to only respond yes or no; and (5) are optimized for use in the free and public Text-Generation-Inference API [11] for acceleration. As described in Sec. 3.2, we utilize $N$ LMs to ensure our method is robust. Specifically, we use $N = 3$ variants of FLAN-T5, namely: FLAN-T5-Base (250M parameters), FLAN-T5-Large (780M parameters), and FLAN-T5-XL (3B parameters).", + "bbox": [ + 75, + 349, + 468, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3 Prompt Ensembling", + "text_level": 1, + "bbox": [ + 76, + 497, + 277, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To guarantee each of these FLAN-T5 variants faithfully produce responses of either yes or no only during AQA, we use the following input template to each LM, reflecting the format used when training FLAN-T51:", + "bbox": [ + 75, + 517, + 468, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text: {LVLM Response} Read the text about an image and answer the question. \nQuestion: Please answer yes or no. \n{Question}", + "bbox": [ + 107, + 583, + 416, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use $M = 3$ semantically identical questions:", + "bbox": [ + 94, + 633, + 416, + 647 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Is there a/an {class_name} in this image?", + "- Does the text imply a/an {class_name} is in the image?", + "- Does the text explicitly mention a/an {class_name} is in the image?" + ], + "bbox": [ + 80, + 648, + 457, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As outlined in Sec. 3.2, we use a unanimous voting mechanism to combine the answers from each (LM, question) pair and so $k = 3 \\times 3 = 9$ .", + "bbox": [ + 75, + 700, + 468, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 Evaluation Results", + "text_level": 1, + "bbox": [ + 76, + 750, + 269, + 766 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we: (1) outline our LVLM selection and reasoning; (2) present THRONE on COCO for evaluating Type I hallucinations; (3) extend THRONE to Objects365 (containing a larger vocabulary); (4) analyze and extend POPE to enable improved evaluation of Type II hallucinations; and (5) highlight results from our ablation studies found in the Supplemental Material.", + "bbox": [ + 75, + 773, + 468, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1 Models", + "text_level": 1, + "bbox": [ + 500, + 90, + 602, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For fair comparison between existing LVLMs, each publicly available model we evaluate uses an LLM with $\\sim 7\\mathrm{B}$ parameters. The LVLMs generally have different sized image encoders—e.g. LLaVA [30] uses a CLIP ViT-L/14 [9, 38] with an input resolution of $336\\times 336$ , while InstructBLIP [7] uses a ViT-g/14 [50] trained with EVA [13] and an input resolution of $224\\times 224$ . Note that image encoder size and resolution is not something we can easily control in a pre-trained model. Each model we consider contains instruction tuning in the final training phase—instruction tuned models provide free-form descriptions; THRONE focuses on models that can generate free-form descriptions. E.g. we leave out BLIP-2 [25] (response median length 31 characters) in favor of InstructBLIP (median response length 525).", + "bbox": [ + 496, + 109, + 893, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2 THRONE Results on COCO", + "text_level": 1, + "bbox": [ + 498, + 342, + 758, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results are shown in Tab. 1. The principal metric that we use to judge model performance is the classwise $F^{0.5}$ -score (highlighted gray). We also report all the metrics outlined in Sec. 3.1, $(P, R, F^1, F^{0.5})$ for overall (left) and class-wise averaging (right), utilizing the unanimous voting presented in Sec. 3.2. See the Supplementary Material for results and an analysis of different voting mechanisms. These results demonstrate that improvements on other benchmarks (POPE, MME, MMBench etc.) may be orthogonal and potentially at odds with improved performance on THRONE. Using the results of the 11 LVLMs that we evaluate, THRONE and POPE, which measure Type I and Type II hallucinations, respectively, have a Spearman's rank correlation coefficient of just 0.2, and THRONE vs POPE-C has just 0.4—the relationship between performance on POPE and THRONE on the same dataset is far from monotonic. For class-wise precision— $P_{\\mathrm{CLS}}$ , the best performing models hallucinate $\\sim 20\\%$ of the objects. We show in the Supplementary Material that the vast majority of false positive objects in the free-form image descriptions evaluated are direct hallucinations rather than misclassifications of visually similar objects (e.g. mistaking a squash racket for a tennis racket). These results demonstrate that much work still remains to adequately suppress Type I hallucinations in LVLMs.", + "bbox": [ + 496, + 361, + 892, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3 THRONE Results on Objects365", + "text_level": 1, + "bbox": [ + 498, + 744, + 790, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Many LVLMs train on COCO directly or indirectly, thus to demonstrate generality we apply THRONE to the Objects365 dataset [42]. Like COCO, Objects365 aims to be exhaustive in its image-level class labeling (it aims to label at least one instance for each class present), but it has a larger object vocabulary and is not used as training data for the LVLMs that we evaluate. To gather a manageable subset of the Objects365 validation set (80k images), we use the natural sampling algorithm from [21], resulting in 5110 im", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https://tinyurl.com/5n6nexze", + "bbox": [ + 94, + 886, + 325, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27233", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/84d946f9861ade220df8aec711d5c09edd046c825cd5d4d4c8f4f4290fa44c21.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelPALLRALLF1ALLF0.5ALLPCLSRCLSF1CLSF0.5CLS
Adapter-v2 [15]46.733.939.343.448.928.536.042.8
Adapter-v2.1 [15]46.834.039.443.548.828.836.242.8
InstructBLIP [7]54.537.244.249.853.733.641.348.0
Otter-Image [23]21.412.716.018.89.54.46.07.7
MiniGPT4 [54]53.032.940.647.249.931.939.044.9
MiniGPT-v2 [6]54.536.043.449.451.334.641.346.8
mPLUG-Owl [48]43.733.437.841.248.229.036.242.6
LRV-Instruction-v2 [28]57.526.736.546.751.426.635.143.3
LLaVA-v1.3* [30]57.632.941.950.152.630.538.645.9
LLaVA-v1.5 [29]54.039.545.650.353.934.341.948.4
LLaVA-Mistral [19, 31]58.339.146.953.157.835.944.351.5
", + "bbox": [ + 80, + 89, + 467, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. THRONE Evaluation with Objects365. Evaluation results for a selection of instruction-tuned LVLMs, we use a subset of Objects365 for the THRONE evaluation. Best and second-best performance are denoted by blue and red, respectively. *Our implementation using official code to enable fair comparison.", + "bbox": [ + 75, + 239, + 468, + 309 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8bccb59c82dd1aac12a91be6f40a11c47060dee4bf85013f0e2030a36ed26b4.jpg", + "image_caption": [ + "Figure 6. Comparing THRONE on COCO and Objects365. We observe that despite variations in the data distribution, THRONE metrics, which measure Type I hallucinations generalize and have a strong Spearman's rank correlation of $r = 0.900$ . One model (red circle) designated as an outlier and ignored when calculating ranking correlation." + ], + "image_footnote": [], + "bbox": [ + 83, + 311, + 460, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ages (the COCO validation set has 5k images). We present the results for THRONE on Objects365 in Tab. 2. Figure 6 shows the strong correlation in THRONE performance between evaluating on COCO and Objects365. This demonstrates that measuring Type I hallucinations in an LVLM using THRONE with a relatively small dataset like COCO, is indeed indicative of the intrinsic level of Type I hallucination in a given LVLM.", + "bbox": [ + 75, + 601, + 467, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4 Completing POPE for Type II Hallucinations", + "text_level": 1, + "bbox": [ + 75, + 729, + 467, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our experiments have used THRONE to evaluate the prevalence of Type I hallucinations in LVLM responses on COCO. POPE evaluates Type II hallucinations on COCO, but we find POPE is largely incomplete. First, POPE only evaluates on 500 COCO validation set images. Second, for each image only a subset of classes (at most 12) are evaluated—each image is only queried with $15\\%$ of possible questions. Finally, POPE artificially balances evaluation questions between positives and negatives, despite object class existence in images being inherently imbalanced.", + "bbox": [ + 73, + 750, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ebc4c81c9640a9eb8b01b8e4c79436addb25706a877a5e3844e203e900861bc5.jpg", + "image_caption": [ + "Figure 7. Instability of POPE to complete evaluation of Type II Hallucinations. Extending POPE to an exhaustive analysis on all COCO images and classes (POPE-C) leads to a dramatic reduction in performance across all 11 models. POPE is sensitive to the sampling mechanism, and by undersampling negative classes, POPE severely underestimates Type II hallucinations in LVLMs." + ], + "image_footnote": [], + "bbox": [ + 504, + 92, + 883, + 189 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The above reasons make the evaluation of Type II hallucinations using POPE insufficient (see Sec. 2 for more details on POPE). To correct this, we complete POPE by using all images to exhaustively query each LVLM for every class in the COCO vocabulary, as done in THRONE. We name this POPE-Complete (POPE-C). Fig. 7 shows the extreme difference in evaluation between POPE and our exhaustive version, POPE-C—reporting $F^{1}$ -score (POPE does not utilize $F^{0.5}$ -score). As POPE only evaluates at most 9 negative classes, only a small subset of potential hallucinations of COCO classes are evaluated, thereby heavily underestimating the extent of Type II hallucination. For each LVLM analyzed, we observe a large—in many cases an extreme—reduction in precision and therefore in F1-score.", + "bbox": [ + 496, + 294, + 890, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our evaluation contains three pairs of LVLMs in which one is the follow-up work to the other, where each follow-up work generally trains on more data for more tasks with a more advanced language model. Comparing the right hand side of Fig. 7 to Tab. 1, we observe that these follow-up works generally show an improvement in POPE (and POPE-C), but surprisingly indicate a small reduction in performance on THRONE with COCO. This observation suggests that progress in reducing Type II hallucinations can be orthogonal to reducing Type I hallucinations.", + "bbox": [ + 496, + 506, + 890, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5 Ablations", + "text_level": 1, + "bbox": [ + 500, + 666, + 619, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the Supplementary Material we present three key ablation experiments and give an executive summary of results here.", + "bbox": [ + 496, + 686, + 890, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, after subsampling COCO images and LVLM responses, we replace the LMs in THRONE with human judgement as an oracle for Type I hallucination occurrence. When comparing THRONE and CHAIR with human judgements, we estimate using THRONE improves the precision of judging Type I hallucinations to $96\\%$ versus $91\\%$ when using CHAIR—this reduces the false discovery rate by more than $50\\%$ . Note that we find most estimated errors in THRONE arise from the particular class definitions in COCO, e.g. the COCO class tv includes computer monitors, which the oracle judgement is aware of.", + "bbox": [ + 496, + 718, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Second, we apply the same class (and image) sampling", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27234", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "strategy as in POPE to THRONE and show this sampling overestimates $F_{\\mathrm{CLS}}^{0.5}$ by an average of 12.3 points compared to the complete use of classes in THRONE (Tab. 1).", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we vary the choice of $k$ i.e. the voting mechanism used to combine answers from multiple (LM, question) pairs. We use the unanimous voting mechanism ( $k = 9$ ) in THRONE to minimize the false discovery rate and find the valid alternatives of simple majority ( $k = 5$ ) or all-but-one ( $k = 8$ ) voting mechanisms have strong correlations and rank correlations across all compute metrics, in THRONE, of $> 0.99$ and $> 0.94$ , respectively.", + "bbox": [ + 75, + 136, + 468, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6 Improved Baselines", + "text_level": 1, + "bbox": [ + 76, + 263, + 276, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Much needs to be done to study Type I hallucinations. As a first step to their mitigation, we demonstrate a baseline method to augment the visual instruction tuning data for LLaVA models [29, 30], yielding improvements on THRONE while maintaining similar performance regarding Type II hallucinations on POPE.", + "bbox": [ + 75, + 282, + 468, + 373 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.1 Visual Instruction Tuning Data Augmentation", + "text_level": 1, + "bbox": [ + 75, + 378, + 468, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similar to chain-of-thought learning [46], during instruction tuning, we augment all visual instruction tuning samples constructed by LLaVA [30] by preponding the task of enumerating a list of objects (present and absent) and indicating approximate locations, if applicable. Other than this, the LLaVA data and training pipeline remains unchanged. To generate the new data for this object enumeration task, we use the same COCO bounding box annotations used to generate the vision instruction tuning data in LLaVA. The simple text-only format (no special tokens) we use is:", + "bbox": [ + 75, + 397, + 468, + 547 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "Instruction: Give a list of objects and locations in the image. \nResponse: {class_name_1} {{location_1}}/absent] {class_name_N} {{location_N}}/absent]", + "guess_lang": "txt", + "bbox": [ + 76, + 556, + 470, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where location_i is a plain text indicator representing the location of the center point of the relevant object in the image on a $3 \\times 3$ grid (e.g. bottom left). To provide negatives in the training data, if class_name_i is not present, we use the plain text indicator absent. Prior work [53] shows that classes that frequently co-occur in the training data are the most common hallucinations, therefore we bias our negative sampling towards class pairs that frequently co-occur using a correlation matrix. We detail and ablate this choice in the Supplementary Material.", + "bbox": [ + 75, + 619, + 468, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2 Improved Baseline Results", + "text_level": 1, + "bbox": [ + 76, + 775, + 326, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 3 shows the result of evaluating our improved baseline method on THRONE, POPE, and POPE-C. During inference, we approximate our training data augmentation by first prompting the LVLM to perform the object enumeration tasks and then generating a response to the prompt from the relevant benchmark. We additionally show results of utilizing VisualGenome bounding box annotations [20] on the", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b52ce66fed017e3d5717e1f5a1965db4d4a0c80c348923d6f3d5f429853c1f64.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelObject \nEnumeration DataTHRONEPOPEPOPE-C
PCLSRCLSF0.5CLSPRF1PRF1
LLaVA-v1.3X79.965.376.558.098.473.07.799.214.3
COCO83.268.879.973.288.280.09.869.417.2
COCO + VG86.267.081.583.082.582.813.850.421.7
LLaVA-v1.5X69.956.466.881.990.886.158.785.769.7
COCO87.276.684.988.685.387.058.987.570.4
COCO + VG86.177.084.189.883.786.764.586.173.7
", + "bbox": [ + 503, + 89, + 890, + 185 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Improved Baseline via Object Enumeration: Adding our object enumeration task to LLaVA training and inference leads to large improvements on THRONE particularly in terms of class-wise precision, $P_{\\mathrm{CLS}}$ , over standard LLaVA models, demonstrating a reduction in Type I hallucinations, as well as small reductions in Type II hallucinations judged by POPE and POPE-C.", + "bbox": [ + 496, + 196, + 890, + 279 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "COCO images where available. On THRONE, we observe a large increase in classwise precision and therefore $F_{\\mathrm{CLS}}^{0.5}$ , particularly for LLaVA-v1.5, demonstrating the ability of our method to reduce Type I hallucinations. Moreover, on POPE and POPE-C, using our object enumeration yields small improvements in precision, indicating reduced Type II hallucinations as well. In the Supplementary Material, we ablate the sampling of negatives during object enumeration training and the effect of removing the object enumeration task during inference.", + "bbox": [ + 496, + 287, + 890, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 498, + 445, + 627, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We establish a novel benchmark, THRONE, for evaluating hallucinations generated by LVLMs in free-form image descriptions i.e. Type I hallucinations. Our benchmark utilizes multiple LMs and prompt formats with a simple voting mechanism to yield an accurate evaluation of Type I hallucinations in LVLM responses. We ensure that THRONE is broadly accessible by utilizing open-source LMs capable of running on a single commercial GPU. Using THRONE, we benchmark 11 publicly available LVLMs on two datasets, COCO and Objects365, and demonstrate that limited progress has been made in addressing Type I hallucinations. Moreover, we show how the established benchmark, POPE, underestimates Type II hallucinations, which occur in response to specific questions e.g. yes-no questions. We present results for a completed version (POPE-C) to enable a comparison of Type I hallucinations through THRONE and Type II hallucinations using POPE-C. Finally, we propose a simple data augmentation for LVLM training that can result in a large reduction in Type I hallucinations whilst maintaining or improving Type II hallucination performance.", + "bbox": [ + 496, + 467, + 890, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Ethical Considerations are discussed in the Supplementary Material.", + "bbox": [ + 498, + 785, + 890, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 821, + 596, + 837 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in", + "bbox": [ + 506, + 844, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27235", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Neural Information Processing Systems, 35:23716-23736, 2022. 1, 3", + "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the International Conference on Computer Vision, pages 2425-2433, 2015. 3", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 3", + "[4] Yonatan Bitton, Hritik Bansal, Jack Hessel, Rulin Shao, Wanrong Zhu, Anas Awadalla, Josh Gardner, Rohan Taori, and Ludwig Schmidt. Visit-bench: A benchmark for vision-language instruction following inspired by real-world use, 2023. 2", + "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems, 33:1877-1901, 2020. 3", + "[6] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 1, 6, 7", + "[7] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1, 3, 6, 7", + "[8] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2009. 5", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. Proceedings of the International Conference on Learning Representations, 2021. 6", + "[10] Mark Everingham, Luc Van Gool, Chris K. I. Williams, John Winn, and Andrew Zisserman. The PASCAL Visual Object Classes (VOC) challenge. International Journal of Computer Vision, 2010. 4", + "[11] Hugging Face. Text generation inference. https://github.com/huggingface/text-generation-inference, 2023. Accessed: November 10, 2023. 6", + "[12] Hao Fang, Saurabh Gupta, Forrest Iandola, Rupesh K Srivastava, Li Deng, Piotr Dollar, Jianfeng Gao, Xiaodong He, Margaret Mitchell, John C Platt, et al. From captions to visual concepts and back. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015. 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 6", + "[14] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 1, 3", + "[15] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 6, 7", + "[16] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations, 2022. 3", + "[17] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, Qiang Liu, Kriti Aggarwal, Zewen Chi, Johan Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, and Furu Wei. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 1", + "[18] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the International Conference on Machine Learning, pages 4904-4916, 2021. 3", + "[19] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6, 7", + "[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123:32-73, 2017. 8", + "[21] Kibok Lee, Hao Yang, Satyaki Chakraborty, Zhaowei Cai, Gurumurthy Swaminathan, Avinash Ravichandran, and Onkar Dabeer. Rethinking few-shot object detection on a multi-domain benchmark. In Proceedings of the European Conference on Computer Vision, 2022. 6", + "[22] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 1, 2, 3", + "[23] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27236", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 1", + "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning, 2023. 1, 3, 6", + "[26] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In Proceedings of the Conference on Empirical Methods in Natural Language, 2023. 1, 2, 3, 4", + "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, 2014. 2, 3, 4, 5", + "[28] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 6, 7", + "[29] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3, 6, 7, 8", + "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1, 3, 6, 7, 8", + "[31] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.6,7", + "[32] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 1, 2, 3, 5", + "[33] Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V Le, Barret Zoph, Jason Wei, and Adam Roberts. The flan collection: Designing data and methods for effective instruction tuning. In Proceedings of the International Conference on Machine Learning, 2023. 5", + "[34] Hwee Tou Ng, Siew Mei Wu, Ted Briscoe, Christian Hadiwinoto, Raymond Hendy Susanto, and Christopher Bryant. The CoNLL-2014 shared task on grammatical error correction. In Proceedings of the Eighteenth Conference on Computational Natural Language Learning: Shared Task, 2014. 5", + "[35] OpenAI. Gpt-4 technical report, 2023. 2, 3", + "[36] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 1, 3", + "[37] Tina D Purnat, Paolo Vacca, Christine Czerniak, Sarah Ball, Stefano Burzo, Tim Zecchin, Amy Wright, Supriya" + ], + "bbox": [ + 78, + 90, + 468, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bezbaruah, Faizza Tanggol, Éve Dubé, Fabienne Labbe, Maude Dionne, Jaya Lamichhane, Avichal Mahajan, Sylvie Briand, and Tim Nguyen. Infodemic signal detection during the COVID-19 pandemic: Development of a methodology for identifying potential information voids in online conversations. *JMIR Infodemiology*, 1(1):e30971, 2021. 5", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 6", + "[39] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 2020. 3, 5", + "[40] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. In Proceedings of the Conference on Empirical Methods in Natural Language, pages 4035-4045, 2018. 1, 2, 3", + "[41] Giuliano Rossi, Jakub Kolodziej, and Gurvinder Brar. A recommender system for active stock selection. Computational Management Science, 17, 2020. 5", + "[42] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the International Conference on Computer Vision, 2019. 5, 6", + "[43] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3", + "[44] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. In Advances in Neural Information Processing Systems, pages 200–212, 2021. 1, 3", + "[45] Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, Jitao Sang, and Haoyu Tang. Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126, 2023. 1", + "[46] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 2022. 8", + "[47] Hongbin Ye, Tong Liu, Aijia Zhang, Wei Hua, and Weiqiang Jia. Cognitive mirage: A review of hallucinations in large language models. arXiv preprint arXiv:2309.06794, 2023. 1, 3", + "[48] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1, 3, 6, 7" + ], + "bbox": [ + 501, + 90, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27237", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 3", + "[50] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 6", + "[51] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the ai ocean: A survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 1, 3", + "[52] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 1", + "[53] Yiyang Zhou, Chenhang Cui, Jaehong Yoon, Linjun Zhang, Zhun Deng, Chelsea Finn, Mohit Bansal, and Huaxiu Yao. Analyzing and mitigating object hallucination in large vision-language models, 2023. 1, 8", + "[54] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 444 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27238", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/da9cf0d5-8d8b-4df8-b396-af5ff1ab4d0d_model.json b/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/da9cf0d5-8d8b-4df8-b396-af5ff1ab4d0d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ff14b13ee95cf2e1f527d26e30a7ce76f7c35a22 --- /dev/null +++ b/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/da9cf0d5-8d8b-4df8-b396-af5ff1ab4d0d_model.json @@ -0,0 +1,2202 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.174, + 0.131, + 0.793, + 0.177 + ], + "angle": 0, + "content": "THRONE: An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.185, + 0.733, + 0.24 + ], + "angle": 0, + "content": "Prannay Kaul\\(^{1*}\\) Zhizhong Li\\(^{2\\dagger}\\) Hao Yang\\(^{2}\\) Yonatan Dukler\\(^{2}\\) Ashwin Swaminathan\\(^{2}\\) C. J. Taylor\\(^{2}\\) Stefano Soatto\\(^{2}\\) VGG, University of Oxford\\(^{1}\\) AWS AI Labs\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.242, + 0.859, + 0.258 + ], + "angle": 0, + "content": "prannay@robots.ox.ac.uk {lzhizhon,haoyng,dukler,swashwin,taylorcj,soattos}@amazon.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.473, + 0.687 + ], + "angle": 0, + "content": "Mitigating hallucinations in large vision-language models (LVLMs) remains an open problem. Recent benchmarks do not address hallucinations in open-ended free-form responses, which we term \"Type I hallucinations\". Instead, they focus on hallucinations responding to very specific question formats—typically a multiple-choice response regarding a particular object or attribute—which we term \"Type II hallucinations\". Additionally, such benchmarks often require external API calls to models which are subject to change. In practice, we observe that a reduction in Type II hallucinations does not lead to a reduction in Type I hallucinations but rather that the two forms of hallucinations are often anti-correlated. To address this, we propose THRONE, a novel object-based automatic framework for quantitatively evaluating Type I hallucinations in LVLM free-form outputs. We use public language models (LMs) to identify hallucinations in LVLM responses and compute informative metrics. By evaluating a large selection of recent LVLMs using public datasets, we show that an improvement in existing metrics do not lead to a reduction in Type I hallucinations, and that established benchmarks for measuring Type I hallucinations are incomplete. Finally, we provide a simple and effective data augmentation method to reduce Type I and Type II hallucinations as a strong baseline." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.704, + 0.219, + 0.719 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.47, + 0.788 + ], + "angle": 0, + "content": "This paper proposes a benchmark to evaluate hallucinations by large vision-language models (LVLMs) when generating free-form responses, specifically detailed descriptions, based on a given image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.788, + 0.471, + 0.864 + ], + "angle": 0, + "content": "The rapid advancement in large language models (LLMs) [52] has pushed the development of large vision-language models (LVLMs). [1, 6, 7, 17, 24, 25, 30, 36, 44, 48, 54] LVLMs take input text and images and generate text responses to enable multi-modal perception and com" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.294, + 0.58, + 0.308 + ], + "angle": 0, + "content": "prehension." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.309, + 0.895, + 0.476 + ], + "angle": 0, + "content": "LVLMs are largely built on LLMs and therefore inherit both their advantages and their disadvantages. LLMs have been shown to produce hallucinations [47, 51], generated text responses that are coherent and plausible but factually incorrect. LVLMs echo this behavior with generated text contradicting with the visual or text input [53]. Hallucinations prevent the use of LVLMs in safety-critical situations and therefore evaluating and mitigating hallucinations in LVLMs is crucial for their deployment in such settings. Determining the presence and cause of hallucinations in LVLMs remains an open question [45, 53]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.895, + 0.689 + ], + "angle": 0, + "content": "We divide LVLM hallucinations into two types. Type I hallucinations occur in response to open-ended questions with a very large set of possible responses — e.g. What is happening in this image?. Type II hallucinations are incorrect responses to a factual question regarding a specific concept about the image with a fixed set of options such as yes/no—e.g. Is there a traffic light in this image? Fig. 3 illustrates the difference between these two types of hallucination. Reducing hallucinations in both cases is required for useful, multi-purpose LVLMs. However, later in Fig. 4 we observe that the same LVLM can give contradicting answers when being evaluated for Type I vs. Type II hallucinations. This implies that measuring and reducing one type does not necessarily reduce the other." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.896, + 0.901 + ], + "angle": 0, + "content": "Existing works to evaluate LVLMs often avoid direct quantification of hallucinations and instead develop comprehensive benchmarks that judge various other desirable abilities such as: optical character recognition, fine-grained recognition and attribute detection [14, 22, 32]. The extent of hallucinations in these benchmarks is obfuscated, since it is only one of many factors influencing other metrics. It requires human effort to inspect individual predictions. There are two major established works which specifically develop a benchmark for evaluating hallucinations in vision-language models: POPE [26] and CHAIR [40], which we discuss in detail in Sec. 2. However, we observe they both have shortcomings in effectively evaluating hallucinations:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.361, + 0.888 + ], + "angle": 0, + "content": "*Work conducted during an internship at Amazon" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.361, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27228" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.478, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.234, + 0.48, + 0.276 + ], + "angle": 0, + "content": "Figure 1. THRONE (Ours): LVLMs are prompted with a concept neutral instruction. An external LM performs abstractive QA on the response to establish the existence of Type I hallucinations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.285, + 0.47, + 0.39 + ], + "angle": 0, + "content": "POPE [26] is a recent work addressing Type II hallucinations with respect to object classes. However, we find Type I and Type II hallucinations are disconnected, and that POPE gives an incomplete picture on LVLM hallucinations. Moreover, POPE systematically under-samples negative object categories leading to a large underestimation of Type II hallucinations (see Sec. 5.4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.47, + 0.528 + ], + "angle": 0, + "content": "CHAIR [40] does address Type I hallucinations—establishing object category hallucination in short image captions using simple text matching. However, CHAIR is not suited to current LVLMs because the simple text matching it employs cannot comprehend abstract or hypothetical concepts present in today's LVLMs (see Fig. 4). Further, hand-crafted rules for each set of classes are required for usable text matching, and trivial model answers can attain a perfect CHAIR score." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.47, + 0.699 + ], + "angle": 0, + "content": "To address the issues of current LVLM hallucination benchmarks, we propose THRONE (Text-from-image Hallucination Recognition with Object-probes for open-ended Evaluation). THRONE leverages language models (LMs) to evaluate Type I hallucinations in free-form, open-ended image descriptions with respect to a pre-defined object vocabulary of interest. By utilizing LMs, THRONE is able to accurately judge whether an object mentioned in an LVLM response is implied to exist in the image or is abstractly mentioned with no implication about its existence (see Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.47, + 0.821 + ], + "angle": 0, + "content": "Moreover, in THRONE, we provide easy access to our benchmark, by leveraging open-source LMs that can run on common GPUs, instead of relying on closed-source commercial models [35] that are subject to arbitrary change, as done in other works [4, 22, 32]. Through combining multiple open-source LMs, we mitigate any single-model biases in judging hallucinations when calculating Type-I hallucination scores with THRONE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We make four contributions: first, we establish an accurate and accessible benchmark to quantitatively evaluate object hallucinations in free-form responses, leveraging LMs to judge the existence of Type I hallucinations—quantitatively showing half the judgement errors of CHAIR;" + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.089, + 0.892, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.234, + 0.892, + 0.276 + ], + "angle": 0, + "content": "Figure 2. POPE: Questions with specific concepts prompt an LVLM directly to evaluate Type II hallucinations [26]. Handcrafted rules parse LVLM responses to give yes/no labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.285, + 0.895, + 0.468 + ], + "angle": 0, + "content": "second, we evaluate a number of current LVLMs on THRONE and demonstrate that observed progress in reducing Type II hallucinations do not translate to a corresponding reduction in Type I hallucinations; third, we show a recent method, POPE, is inaccurately capturing the extent of Type II hallucination discovery due to its sampling strategy; we mitigate this issue in our implementation of THRONE while presenting results for a complete version of POPE; and fourth, we provide a simple augmentation of visual instruction tuning data which significantly improves performance on Type I hallucinations while maintaining or improving Type II hallucination performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.482, + 0.652, + 0.499 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Hallucination Benchmarks for LVLMs. In response to the development of LVLMs, few evaluation benchmarks focusing on hallucinations have been introduced. CHAIR [40], one of the first works to assess hallucinations, is designed for short image captions. CHAIR uses a fixed set of object classes (extended with their synonyms) as a set of text strings to find predicted object classes in image captions via exact text matching. Subsequently, each class matched in the caption is compared to the COCO [27] captions and bounding box annotations to establish object hallucinations. CHAIR was developed prior to current instruction-tuned LVLMs which produce long free-form responses (\\(10 \\times\\) longer than image captions) with a diverse vocabulary, limiting its applicability to modern LVLMs. As shown in Fig. 3, exact text matching in CHAIR is prone to incorrectly matching the vocabulary classes with abstract concepts in the free-form response and the and synonyms of the class list must be manually selected to prevent confusion during evaluation. For example, in CHAIR the word \"chair\" will match all responses for the phrase \"toilet seat\", because the exact text matching classifies \"seat\" as the COCO class \"chair\". Finally, CHAIR metrics compare the overall number of predicted objects to the overall number of predicted objects judged to be hallucinations—ignoring the recall of ground-truth objects and the distribution of object classes. This means that a single correct" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27229" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.287 + ], + "angle": 0, + "content": "prediction across the entire evaluation dataset along with a generic response otherwise e.g. A natural scene, achieves a perfect score \\((\\frac{0 \\text{ hallucinated objects}}{1 \\text{ predicted objects}} = 0.0)\\). See the Supplementary Material for a full overview of CHAIR. In contrast, THRONE uses pre-trained LMs that go beyond direct synonym matching, to automatically judge the existence of concepts and hallucinations in free-form responses. In addition, our method considers both recall and precision to yield a holistic benchmark and does not require any manual curation of synonyms. THRONE and CHAIR both evaluate Type I hallucinations—hallucinations in response to concept-neutral prompts e.g. Describe this image in detail." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.473, + 0.578 + ], + "angle": 0, + "content": "POPE [26] is a recently proposed benchmark to evaluate object hallucinations in LVLMs—specifically Type II hallucinations, in which an LVLM is directly queried with a yes-no question regarding the existence of a particular object of the form: Is there \\(\\{\\mathrm{a / an}\\}\\) {object_class_name} in the image?. The LVLM response is parsed using simple rules to determine whether a Type II hallucination has occurred. Precision and recall metrics are compiled using the parsed LVLM responses and the ground-truth annotation data. Despite focusing on measuring hallucinations, POPE only queries an LVLM with 3 positive and 3 sampled negative questions per COCO image i.e. the evaluation is artificially balanced. This means many potential hallucinations with respect to the COCO categories are not captured by their method. In Sec. 5.4, we show POPE dramatically underestimates Type II hallucinations and present the results of a complete version. Our method, THRONE, evaluates the prevalence of Type I hallucinations, which we observe are disjoint to Type II hallucinations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.473, + 0.779 + ], + "angle": 0, + "content": "Comprehensive Benchmarks for LVLMs have recently grown in number. MMBench [32] and MM-Vet [49] assess various aspects of LVLM performance such as: color perception, celebrity recognition, and numerical calculation. However many of these works integrate evolving APIs which are modified often (or even discontinued) and are inherently stochastic. Over time this greatly reduces the consistency of these benchmarks. Exceptions are MME [14] and SEED-Bench [22], but the impact of Type II hallucinations on final metrics is conflated with a number of other aspects of model performance. Our method, THRONE, directly addresses Type I hallucinations, only making use of open-source language models and datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Large Vision Language Models (LVLMs) have rapidly developed by harnessing advancements in large language models (LLM) [5, 39, 43] and by directly integrating pretrained LLMs into their architectures. In contrast to earlier vision-language models such as CLIP [18, 38], LVLMs are generally comprised of a pretrained LLM and image encoder, aligned with a connector module of varying complexity. Some works are highlighted here. Frozen [44] is" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.09, + 0.895, + 0.242 + ], + "angle": 0, + "content": "
ImagePromptResponse
What is happening in this image?In the image, a man is hanging clothes on a clothesline that is attached to the back of a moving car.
Is there a traffic light in this image?Yes, there is a traffic light in the image, and the man is standing on the back of a car near it.
" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.251, + 0.895, + 0.308 + ], + "angle": 0, + "content": "Figure 3. Type I vs. Type II Hallucinations: (Top) LVLMs prompted with concept-neutral instructions produce Type I hallucinations. (Bottom) Instructions specifying a concept produce Type II hallucinations. Examples from LLaVA-v1.5 [29]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.895, + 0.593 + ], + "angle": 0, + "content": "an early work fine-tuning the vision encoder to dynamically prefix the prompt to a frozen LLM. Flamingo [1] combines visual and language features using cross-attention layers in an otherwise frozen LLM. BLIP-2 [25] uses a frozen image encoder to learn a Querying-Transformer (Q-Former) on image-text pairs that is used as the connector. This architecture is adapted in [7] for dialogue via training with visual instruction tuning. LLaVA [30] uses COCO [27] annotations and GPT-4 [35] to generate visual instruction tuning data in a plain-text pipeline. Combining this generated data with standard VQA datasets (VQAv2 etc.) further boosts performance [29]. Different works modify the training approach by using efficient adaptation [16, 48] multiple training stages [3] or introduce the use of discrete tokens for localization [36]. We note however, that most of these works evaluate performance on traditional vision-language datasets like VQAv2 [2], which do not consider the extent of hallucinations, a known problem with LLMs [47, 51]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.603, + 0.627, + 0.619 + ], + "angle": 0, + "content": "3 THRONE" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Recap of existing methods. POPE [26] and other benchmarks (MME [14], MMBench [32]) directly query LVLMs with a restricted desired answer space, e.g. yes-no (MME, POPE) and A-B-C-D multiple choice (MMBench), as shown in Fig. 2. These benchmarks only consider such short answer formats, whereas a key quality of LVLMs is in their ability to generate free-form coherent text. Moreover, POPE, which addresses Type II hallucinations, undersamples negative classes meaning hallucinations are dramatically underestimated (see Sec. 5.4 and Fig. 7). In contrast, we skip class subsampling and enumerate all classes for every image, ensuring a full evaluation of Type I hallucinations of the ground-truth classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "CHAIR [40] also evaluates Type I hallucinations, but was developed when typical vision-language models could only generate short and simple captions similar to those in COCO Captions [12]. Moreover, it lacks accurate comprehension of responses (see the right side of Fig. 4) and ig-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27230" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.169, + 0.089, + 0.382, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.168, + 0.209, + 0.383, + 0.219 + ], + "angle": 0, + "content": "Type II Hallucination Evaluation (POPE)" + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.22, + 0.47, + 0.315 + ], + "angle": 0, + "content": "
Instruction [GT]Response [Prediction]
Is there... \na person [No] \na banana [Yes] \nan apple [No] \nan orange [No] \n...in the image?No, there is no person in the image. [No] \nYes, there is a banana in the image. [Yes] \nNo, there is not an apple in the image. [No] \nNo, there is no orange in the image. [No] \nNo Type II Hallucinations Found
" + }, + { + "type": "table_caption", + "bbox": [ + 0.594, + 0.09, + 0.768, + 0.1 + ], + "angle": 0, + "content": "Type I Hallucination Evaluation" + }, + { + "type": "table", + "bbox": [ + 0.476, + 0.101, + 0.889, + 0.315 + ], + "angle": 0, + "content": "
InstructionResponse
Describe this image in detail.The image showcases a fruit stand at a grocery store, featuring a variety of fruits on display. There are several bunches of bananas, with some placed in the foreground and others in the background. The bananas are arranged in different sections, creating an appealing presentation for customers. In addition to the bananas, there are also apples and oranges on display. The apples are located towards the left side of the image, while the oranges are placed in the middle and right side of the stand. The fruits are well-organized and presented in an attractive manner, making it an inviting sight for shoppers. Type I Hallucinations Present and Found
MSCOCO Object Prediction from Description
HumanCHAIRTHRONE (Ours)
banana apple orangebanana person apple orangebanana apple orange
" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.327, + 0.892, + 0.412 + ], + "angle": 0, + "content": "Figure 4. A Comparison of POPE, CHAIR and THRONE: Directly querying LVLMs for object existence (person, banana etc.) using concept-specific instructions, as in POPE (bottom left), does not produce the same hallucinations as using concept-neutral instructions (right). We highlight the Type I hallucinations in orange. CHAIR relies on exact text matching to a fixed set of objects and synonyms, thus incorrectly labels \"customers\" and \"shoppers\" as hallucinations, highlighted in red. THRONE is designed for the rich vocabulary and the free-form generations of modern LVLMs by harnessing LMs to establish object existence. By using an LM to pass judgement, our evaluation correctly captures \"customers\" and \"shoppers\" as hypothetical content in the free-form generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.469, + 0.525 + ], + "angle": 0, + "content": "nores the recall of ground-truth objects. In Sec. 5.5, we describe quantitative evaluations, using a human oracle, which demonstrate THRONE halves the rate of hallucination misjudgement in CHAIR. See Sec. 2 and the Supplementary Material for more details on CHAIR and its shortcomings. Fig. 4 shows an overview of the three aforementioned methods: POPE, CHAIR and our method, THRONE." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.533, + 0.436, + 0.548 + ], + "angle": 0, + "content": "3.1 Evaluating Hallucinations with THRONE" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.613 + ], + "angle": 0, + "content": "To address these limitations, we propose a framework, THRONE, shown in Fig. 1, to evaluate the prevalence of Type I hallucinations in LVLM responses conditioned on an image and a neutral text prompt." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.869 + ], + "angle": 0, + "content": "For each image in a labeled dataset, \\(\\mathcal{I}\\), addressing a set of classes, \\(\\mathcal{C}\\), the LVLM is queried with the same instruction: Describe this image in detail., regardless of image content. The LVLM response, which is expected to be long free-form text containing an image description, is generated and stored. Next, a publicly available, open-source, external language model (LM) performs abstractive question answering (AQA) using the LVLM response as context and a question of the form: Please answer yes or no. Is there \\(\\{\\mathrm{a / an}\\}\\) {object class name} in this image? or similar, for every class in \\(\\mathcal{C}\\) (right side of Fig. 1). By selecting an appropriate LM and using a simple prompt template (see Sec. 4 for specific details), we ensure the AQA response is either yes or no—our method does not require any additional parsing. This is in contrast to other works which require added parsing or interpretation by a closed-source model." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "After performing AQA on each response generated by the LVLM for every class in \\(\\mathcal{C}\\), we obtain an array of pre" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.42, + 0.591, + 0.432 + ], + "angle": 0, + "content": "dicted labels:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.433, + 0.891, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Y}} \\in \\{0, 1 \\} ^ {| \\mathcal {I} | \\times | \\mathcal {C} |} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.492 + ], + "angle": 0, + "content": "where \\(0/1\\) indicates a negative/positive existence judgement by the LM with respect to the relevant LVLM response." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.492, + 0.892, + 0.521 + ], + "angle": 0, + "content": "Similarly using the ground-truth data for \\(\\mathcal{I}\\), an array of ground-truth labels can be constructed:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.523, + 0.891, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} \\in \\{0, 1 \\} ^ {| \\mathcal {I} | \\times | \\mathcal {C} |} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.686 + ], + "angle": 0, + "content": "Using these two arrays, we calculate four metrics: (1) Overall Precision, \\( P_{\\mathrm{ALL}} \\); (2) Overall Recall, \\( R_{\\mathrm{ALL}} \\); (3) Class-wise Precision, \\( P_{\\mathrm{CLS}} \\); and (4) Class-wise Recall, \\( R_{\\mathrm{CLS}} \\). Overall metrics are calculated in a class-agnostic manner. Class-wise metrics are calculated in a class-conscious manner by computing precision and recall for each category separately and then averaging. This follows common practice in object detection and instance segmentation [10, 27]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.823 + ], + "angle": 0, + "content": "False positives in LVLMs reduce precision and are dominated by hallucinations—precision indicates the extent of Type I hallucinations in LVLM responses. The recall metrics inform the level of class coverage by an LVLM when producing image descriptions. The class-wise metrics give a general measure of performance as the overall metrics are skewed towards the most common categories. A common way to combine precision, \\( P \\), and recall, \\( R \\), metrics is through the generalized F score, \\( F_{\\beta} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.831, + 0.891, + 0.864 + ], + "angle": 0, + "content": "\\[\nF ^ {\\beta} = \\left(1 + \\beta^ {2}\\right) \\cdot \\frac {P \\cdot R}{\\left(\\beta^ {2} \\cdot P\\right) + R} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Prior work such as POPE [26], use the common balanced F-score (or \\(F^1\\)-score) which equally weights precision and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27231" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "recall. However, given that THRONE is concerned with measuring hallucinations and is therefore particularly interested in precision, we choose \\(\\beta = 0.5\\) or the \\(F^{0.5}\\)-score, thus weighing precision twice as important as recall. The \\(F^{0.5}\\)-score is commonly used in pandemic misinformation filters [37], recommender systems [34], active stock selection [41] and other areas where false positives are costlier than false negatives. From this we can calculate the overall and class-wise \\(F^{0.5}\\)-scores, \\(F_{\\mathrm{ALL}}^{0.5}\\) and \\(F_{\\mathrm{CLS}}^{0.5}\\), respectively. To mitigate against class imbalance issues, we use \\(F_{\\mathrm{CLS}}^{0.5}\\) as the principle metric of comparison between LVLMs in THRONE." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.291, + 0.409, + 0.307 + ], + "angle": 0, + "content": "3.2 Ensuring Robustness via Ensembling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.473, + 0.649 + ], + "angle": 0, + "content": "Any LM used for AQA in THRONE may misjudge Type II hallucinations—no LM is perfect. Fig. 5 (top) shows two different variants from the same model family (FLANT5 [33]), yielding opposite responses when prompted with the same response and question. Moreover, an LM may yield different answers to semantically identical questions, despite conditioning on the same response—shown in Fig. 5 (bottom). To ensure THRONE is robust to spurious performance by any one LM, we ensemble various LMs and semantically equivalent question formats. The use of \\( N \\) distinct LMs and \\( M \\) distinct question formats yields \\( NM \\) answers for each (LVLM response, class) combination. This set of \\( NM \\) answers is combined based on equal voting by each (LM, question) pair to \"elect\" the predicted answer. Continuing the notation from Eq. (1), stacking predictions from each (LM, question) pair yields a 3D array of predicted labels: \\( \\bar{\\mathbf{Y}} \\in \\{0,1\\}^{|\\mathcal{I}_{\\mathrm{OURS}}| \\times |\\mathcal{C}_{\\mathrm{OURS}}| \\times NM} \\). To combine the answers from each (LM, question) pair via voting, we require agreement between at least \\( k \\) answers. Where sufficient agreement does not exist we introduce an \"ignore\" label. Mathematically, the elements in \\( \\hat{\\mathbf{Y}} \\) (from Eq. (1)) are calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.66, + 0.416, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\hat {Y} _ {i, j} = \\left\\{ \\begin{array}{l l} 0, & \\sum_ {k = 1} ^ {N M} \\bar {Y} _ {i, j, k} \\leq (N M - k) \\\\ 1, & \\sum_ {k = 1} ^ {N M} \\bar {Y} _ {i, j, k} \\geq k \\\\ - 1, & \\text {o t h e r w i s e} \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.473, + 0.901 + ], + "angle": 0, + "content": "where an \"ignore\" label exists in \\(\\bar{\\mathbf{Y}}\\), it is removed from the calculations of \\(P_{\\mathrm{ALL}}\\), \\(R_{\\mathrm{ALL}}\\), \\(P_{\\mathrm{CLS}}\\), and \\(R_{\\mathrm{CLS}}\\), as we cannot be confident in the AQA process for that particular (LVLM response, class) combination. The choice of \\(k\\) reflects the desired level of confidence. We make use of a unanimous voting mechanism (\\(k = NM\\))—use the prediction only if all (LM, question) pairs agree, otherwise ignore. Human evaluation of our benchmark and choice of voting mechanism is found in the Supplementary Material. Once we have applied this voting mechanism, we can calculate the metrics described at the end of Sec. 3.1." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.092, + 0.882, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.219, + 0.895, + 0.29 + ], + "angle": 0, + "content": "Figure 5. AQA Ensembling in Evaluation: Using different LMs or different prompts when running AQA on LVLM generated responses can produce opposing answers to identical prompts or identical LMs. To ensure THRONE is robust to this, we ensemble multiple LMs and multiple prompts in our evaluation pipeline." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.296, + 0.67, + 0.314 + ], + "angle": 0, + "content": "4 Implementation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.892, + 0.351 + ], + "angle": 0, + "content": "We provide details of our framework regarding the selection of public datasets, public LMs and LM prompts." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.356, + 0.605, + 0.371 + ], + "angle": 0, + "content": "4.1 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.376, + 0.892, + 0.557 + ], + "angle": 0, + "content": "Any proper evaluation of object hallucinations (a type of false positive error) requires knowing, with certainty, which classes are absent in an image. In our benchmark, we use COCO [27] for a number of reasons: (1) its annotations of 80 categories are exhaustive at an image level (image-level recall \\(\\approx 99\\%\\) [27])—if there are many book instances in a COCO image, at least one is annotated with bounding boxes; (2) many LVLMs are partly trained on COCO data and so should be familiar with the set of categories; (3) its images generally contain complex scenes suitable for generating long free-form descriptions unlike image recognition datasets like ImageNet [8]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.892, + 0.694 + ], + "angle": 0, + "content": "We utilize the validation set of COCO 2017, which contains \\( |\\mathcal{I}| = 5000 \\) images and \\( |\\mathcal{C}| = 80 \\) categories. Using the single LVLM text prompt Describe this image in detail., we generate 5000 responses. As we query each LVLM response for each category in \\( \\mathcal{C} \\), a single LM performs AQA \\( |\\mathcal{I}| \\times |\\mathcal{C}| = 400k \\) times across the LVLM responses—one instance of AQA per (image response, class) pair. In Sec. 5.3, we also present results using Objects365 [42] which is rarely used in LVLM training." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.7, + 0.684, + 0.716 + ], + "angle": 0, + "content": "4.2 Language Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To assess Type I hallucinations in an LVLM response using THRONE, we require a language model (LM) which can answer questions on the existence of object categories based on the LVLM response. MMBench [32] makes use of a ChatGPT model to identify multiple choice answer selections and still reports mistakes. In our experience, some LMs give rather incoherent judgements when used to assess hallucinations when the prompt is changed (see supplemental material). Therefore for THRONE, we choose FLAN-T5 models [33, 39]. We make this choice because FLAN-T5 model family: (1) have undergone instruction tuning with thousands of tasks [33]; (2) are open-source and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27232" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.09, + 0.47, + 0.206 + ], + "angle": 0, + "content": "
ModelL\\( P_{ALL} \\)\\( R_{ALL} \\)\\( F_{ALL}^{1} \\)\\( F_{ALL}^{0.5} \\)\\( P_{CLS} \\)\\( R_{CLS} \\)\\( F_{CLS}^{1} \\)\\( F_{CLS}^{0.5} \\)
Adapter-v2 [15]51463.673.368.165.368.270.669.468.7
Adapter-v2.1 [15]51263.873.768.465.567.471.269.368.1
InstructBLIP [7]52570.874.372.571.577.271.974.576.1
Otter-Image [23]25733.031.232.132.725.216.920.222.9
MiniGPT4 [54]47381.759.869.076.179.961.869.775.5
MiniGPT-v2 [6]38179.066.672.376.277.667.071.975.2
mPLUG-Owl [48]55555.571.962.658.166.368.367.366.7
LRV-Instruction-v2 [28]10382.056.767.075.378.458.867.273.5
LLaVA-v1.3* [30]53280.565.272.176.979.965.371.976.5
LLaVA-v1.5 [29]50968.161.064.466.669.956.462.566.8
LLaVA-Mistral [19, 31]52486.871.878.383.684.464.270.877.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.216, + 0.473, + 0.328 + ], + "angle": 0, + "content": "Table 1. THRONE Results with COCO for a selection of instruction-tuned LVLMs. We select \\( F_{\\mathrm{CLS}}^{0,5} \\) as the principal metric for evaluation in our benchmark to balance across classes and to prioritize precision (which reflects the extent of hallucination) over recall. Best and second-best performance are denoted by blue and red, respectively. *Our implementation using official code to enable fair comparison. \\( L \\) corresponds to the median response length (measured in # of characters)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.47, + 0.488 + ], + "angle": 0, + "content": "therefore accessible to the community; (3) can fit locally on a single GPU for the models we consider; (4) follow user's instruction to only respond yes or no; and (5) are optimized for use in the free and public Text-Generation-Inference API [11] for acceleration. As described in Sec. 3.2, we utilize \\( N \\) LMs to ensure our method is robust. Specifically, we use \\( N = 3 \\) variants of FLAN-T5, namely: FLAN-T5-Base (250M parameters), FLAN-T5-Large (780M parameters), and FLAN-T5-XL (3B parameters)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.279, + 0.515 + ], + "angle": 0, + "content": "4.3 Prompt Ensembling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.47, + 0.578 + ], + "angle": 0, + "content": "To guarantee each of these FLAN-T5 variants faithfully produce responses of either yes or no only during AQA, we use the following input template to each LM, reflecting the format used when training FLAN-T51:" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.584, + 0.418, + 0.628 + ], + "angle": 0, + "content": "Text: {LVLM Response} Read the text about an image and answer the question. \nQuestion: Please answer yes or no. \n{Question}" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.634, + 0.418, + 0.648 + ], + "angle": 0, + "content": "We use \\( M = 3 \\) semantically identical questions:" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.649, + 0.414, + 0.659 + ], + "angle": 0, + "content": "- Is there a/an {class_name} in this image?" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.659, + 0.458, + 0.679 + ], + "angle": 0, + "content": "- Does the text imply a/an {class_name} is in the image?" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.679, + 0.385, + 0.699 + ], + "angle": 0, + "content": "- Does the text explicitly mention a/an {class_name} is in the image?" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.649, + 0.458, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.47, + 0.746 + ], + "angle": 0, + "content": "As outlined in Sec. 3.2, we use a unanimous voting mechanism to combine the answers from each (LM, question) pair and so \\( k = 3 \\times 3 = 9 \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.751, + 0.27, + 0.767 + ], + "angle": 0, + "content": "5 Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.774, + 0.47, + 0.88 + ], + "angle": 0, + "content": "In this section, we: (1) outline our LVLM selection and reasoning; (2) present THRONE on COCO for evaluating Type I hallucinations; (3) extend THRONE to Objects365 (containing a larger vocabulary); (4) analyze and extend POPE to enable improved evaluation of Type II hallucinations; and (5) highlight results from our ablation studies found in the Supplemental Material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.603, + 0.105 + ], + "angle": 0, + "content": "5.1 Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.11, + 0.894, + 0.337 + ], + "angle": 0, + "content": "For fair comparison between existing LVLMs, each publicly available model we evaluate uses an LLM with \\(\\sim 7\\mathrm{B}\\) parameters. The LVLMs generally have different sized image encoders—e.g. LLaVA [30] uses a CLIP ViT-L/14 [9, 38] with an input resolution of \\(336\\times 336\\), while InstructBLIP [7] uses a ViT-g/14 [50] trained with EVA [13] and an input resolution of \\(224\\times 224\\). Note that image encoder size and resolution is not something we can easily control in a pre-trained model. Each model we consider contains instruction tuning in the final training phase—instruction tuned models provide free-form descriptions; THRONE focuses on models that can generate free-form descriptions. E.g. we leave out BLIP-2 [25] (response median length 31 characters) in favor of InstructBLIP (median response length 525)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.343, + 0.759, + 0.358 + ], + "angle": 0, + "content": "5.2 THRONE Results on COCO" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.893, + 0.739 + ], + "angle": 0, + "content": "Results are shown in Tab. 1. The principal metric that we use to judge model performance is the classwise \\( F^{0.5} \\)-score (highlighted gray). We also report all the metrics outlined in Sec. 3.1, \\( (P, R, F^1, F^{0.5}) \\) for overall (left) and class-wise averaging (right), utilizing the unanimous voting presented in Sec. 3.2. See the Supplementary Material for results and an analysis of different voting mechanisms. These results demonstrate that improvements on other benchmarks (POPE, MME, MMBench etc.) may be orthogonal and potentially at odds with improved performance on THRONE. Using the results of the 11 LVLMs that we evaluate, THRONE and POPE, which measure Type I and Type II hallucinations, respectively, have a Spearman's rank correlation coefficient of just 0.2, and THRONE vs POPE-C has just 0.4—the relationship between performance on POPE and THRONE on the same dataset is far from monotonic. For class-wise precision—\\( P_{\\mathrm{CLS}} \\), the best performing models hallucinate \\( \\sim 20\\% \\) of the objects. We show in the Supplementary Material that the vast majority of false positive objects in the free-form image descriptions evaluated are direct hallucinations rather than misclassifications of visually similar objects (e.g. mistaking a squash racket for a tennis racket). These results demonstrate that much work still remains to adequately suppress Type I hallucinations in LVLMs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.746, + 0.791, + 0.762 + ], + "angle": 0, + "content": "5.3 THRONE Results on Objects365" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Many LVLMs train on COCO directly or indirectly, thus to demonstrate generality we apply THRONE to the Objects365 dataset [42]. Like COCO, Objects365 aims to be exhaustive in its image-level class labeling (it aims to label at least one instance for each class present), but it has a larger object vocabulary and is not used as training data for the LVLMs that we evaluate. To gather a manageable subset of the Objects365 validation set (80k images), we use the natural sampling algorithm from [21], resulting in 5110 im" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.326, + 0.9 + ], + "angle": 0, + "content": "1https://tinyurl.com/5n6nexze" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27233" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.09, + 0.468, + 0.228 + ], + "angle": 0, + "content": "
ModelPALLRALLF1ALLF0.5ALLPCLSRCLSF1CLSF0.5CLS
Adapter-v2 [15]46.733.939.343.448.928.536.042.8
Adapter-v2.1 [15]46.834.039.443.548.828.836.242.8
InstructBLIP [7]54.537.244.249.853.733.641.348.0
Otter-Image [23]21.412.716.018.89.54.46.07.7
MiniGPT4 [54]53.032.940.647.249.931.939.044.9
MiniGPT-v2 [6]54.536.043.449.451.334.641.346.8
mPLUG-Owl [48]43.733.437.841.248.229.036.242.6
LRV-Instruction-v2 [28]57.526.736.546.751.426.635.143.3
LLaVA-v1.3* [30]57.632.941.950.152.630.538.645.9
LLaVA-v1.5 [29]54.039.545.650.353.934.341.948.4
LLaVA-Mistral [19, 31]58.339.146.953.157.835.944.351.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.24, + 0.47, + 0.31 + ], + "angle": 0, + "content": "Table 2. THRONE Evaluation with Objects365. Evaluation results for a selection of instruction-tuned LVLMs, we use a subset of Objects365 for the THRONE evaluation. Best and second-best performance are denoted by blue and red, respectively. *Our implementation using official code to enable fair comparison." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.313, + 0.462, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.509, + 0.47, + 0.593 + ], + "angle": 0, + "content": "Figure 6. Comparing THRONE on COCO and Objects365. We observe that despite variations in the data distribution, THRONE metrics, which measure Type I hallucinations generalize and have a strong Spearman's rank correlation of \\( r = 0.900 \\). One model (red circle) designated as an outlier and ignored when calculating ranking correlation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.602, + 0.468, + 0.722 + ], + "angle": 0, + "content": "ages (the COCO validation set has 5k images). We present the results for THRONE on Objects365 in Tab. 2. Figure 6 shows the strong correlation in THRONE performance between evaluating on COCO and Objects365. This demonstrates that measuring Type I hallucinations in an LVLM using THRONE with a relatively small dataset like COCO, is indeed indicative of the intrinsic level of Type I hallucination in a given LVLM." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.73, + 0.468, + 0.746 + ], + "angle": 0, + "content": "5.4 Completing POPE for Type II Hallucinations" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Our experiments have used THRONE to evaluate the prevalence of Type I hallucinations in LVLM responses on COCO. POPE evaluates Type II hallucinations on COCO, but we find POPE is largely incomplete. First, POPE only evaluates on 500 COCO validation set images. Second, for each image only a subset of classes (at most 12) are evaluated—each image is only queried with \\(15\\%\\) of possible questions. Finally, POPE artificially balances evaluation questions between positives and negatives, despite object class existence in images being inherently imbalanced." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.093, + 0.885, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.2, + 0.892, + 0.284 + ], + "angle": 0, + "content": "Figure 7. Instability of POPE to complete evaluation of Type II Hallucinations. Extending POPE to an exhaustive analysis on all COCO images and classes (POPE-C) leads to a dramatic reduction in performance across all 11 models. POPE is sensitive to the sampling mechanism, and by undersampling negative classes, POPE severely underestimates Type II hallucinations in LVLMs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.892, + 0.505 + ], + "angle": 0, + "content": "The above reasons make the evaluation of Type II hallucinations using POPE insufficient (see Sec. 2 for more details on POPE). To correct this, we complete POPE by using all images to exhaustively query each LVLM for every class in the COCO vocabulary, as done in THRONE. We name this POPE-Complete (POPE-C). Fig. 7 shows the extreme difference in evaluation between POPE and our exhaustive version, POPE-C—reporting \\( F^{1} \\)-score (POPE does not utilize \\( F^{0.5} \\)-score). As POPE only evaluates at most 9 negative classes, only a small subset of potential hallucinations of COCO classes are evaluated, thereby heavily underestimating the extent of Type II hallucination. For each LVLM analyzed, we observe a large—in many cases an extreme—reduction in precision and therefore in F1-score." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Our evaluation contains three pairs of LVLMs in which one is the follow-up work to the other, where each follow-up work generally trains on more data for more tasks with a more advanced language model. Comparing the right hand side of Fig. 7 to Tab. 1, we observe that these follow-up works generally show an improvement in POPE (and POPE-C), but surprisingly indicate a small reduction in performance on THRONE with COCO. This observation suggests that progress in reducing Type II hallucinations can be orthogonal to reducing Type I hallucinations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.667, + 0.62, + 0.682 + ], + "angle": 0, + "content": "5.5 Ablations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.891, + 0.718 + ], + "angle": 0, + "content": "In the Supplementary Material we present three key ablation experiments and give an executive summary of results here." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.884 + ], + "angle": 0, + "content": "First, after subsampling COCO images and LVLM responses, we replace the LMs in THRONE with human judgement as an oracle for Type I hallucination occurrence. When comparing THRONE and CHAIR with human judgements, we estimate using THRONE improves the precision of judging Type I hallucinations to \\(96\\%\\) versus \\(91\\%\\) when using CHAIR—this reduces the false discovery rate by more than \\(50\\%\\). Note that we find most estimated errors in THRONE arise from the particular class definitions in COCO, e.g. the COCO class tv includes computer monitors, which the oracle judgement is aware of." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Second, we apply the same class (and image) sampling" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27234" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.137 + ], + "angle": 0, + "content": "strategy as in POPE to THRONE and show this sampling overestimates \\( F_{\\mathrm{CLS}}^{0.5} \\) by an average of 12.3 points compared to the complete use of classes in THRONE (Tab. 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.47, + 0.258 + ], + "angle": 0, + "content": "Finally, we vary the choice of \\( k \\) i.e. the voting mechanism used to combine answers from multiple (LM, question) pairs. We use the unanimous voting mechanism (\\( k = 9 \\)) in THRONE to minimize the false discovery rate and find the valid alternatives of simple majority (\\( k = 5 \\)) or all-but-one (\\( k = 8 \\)) voting mechanisms have strong correlations and rank correlations across all compute metrics, in THRONE, of \\( > 0.99 \\) and \\( > 0.94 \\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.264, + 0.277, + 0.281 + ], + "angle": 0, + "content": "6 Improved Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.469, + 0.374 + ], + "angle": 0, + "content": "Much needs to be done to study Type I hallucinations. As a first step to their mitigation, we demonstrate a baseline method to augment the visual instruction tuning data for LLaVA models [29, 30], yielding improvements on THRONE while maintaining similar performance regarding Type II hallucinations on POPE." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.379, + 0.469, + 0.395 + ], + "angle": 0, + "content": "6.1 Visual Instruction Tuning Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.47, + 0.549 + ], + "angle": 0, + "content": "Similar to chain-of-thought learning [46], during instruction tuning, we augment all visual instruction tuning samples constructed by LLaVA [30] by preponding the task of enumerating a list of objects (present and absent) and indicating approximate locations, if applicable. Other than this, the LLaVA data and training pipeline remains unchanged. To generate the new data for this object enumeration task, we use the same COCO bounding box annotations used to generate the vision instruction tuning data in LLaVA. The simple text-only format (no special tokens) we use is:" + }, + { + "type": "code", + "bbox": [ + 0.077, + 0.557, + 0.472, + 0.611 + ], + "angle": 0, + "content": "Instruction: Give a list of objects and locations in the image. \nResponse: {class_name_1} {{location_1}}/absent] {class_name_N} {{location_N}}/absent]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.469, + 0.771 + ], + "angle": 0, + "content": "where location_i is a plain text indicator representing the location of the center point of the relevant object in the image on a \\(3 \\times 3\\) grid (e.g. bottom left). To provide negatives in the training data, if class_name_i is not present, we use the plain text indicator absent. Prior work [53] shows that classes that frequently co-occur in the training data are the most common hallucinations, therefore we bias our negative sampling towards class pairs that frequently co-occur using a correlation matrix. We detail and ablate this choice in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.776, + 0.327, + 0.791 + ], + "angle": 0, + "content": "6.2 Improved Baseline Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Tab. 3 shows the result of evaluating our improved baseline method on THRONE, POPE, and POPE-C. During inference, we approximate our training data augmentation by first prompting the LVLM to perform the object enumeration tasks and then generating a response to the prompt from the relevant benchmark. We additionally show results of utilizing VisualGenome bounding box annotations [20] on the" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.09, + 0.892, + 0.186 + ], + "angle": 0, + "content": "
ModelObject \nEnumeration DataTHRONEPOPEPOPE-C
PCLSRCLSF0.5CLSPRF1PRF1
LLaVA-v1.3X79.965.376.558.098.473.07.799.214.3
COCO83.268.879.973.288.280.09.869.417.2
COCO + VG86.267.081.583.082.582.813.850.421.7
LLaVA-v1.5X69.956.466.881.990.886.158.785.769.7
COCO87.276.684.988.685.387.058.987.570.4
COCO + VG86.177.084.189.883.786.764.586.173.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.197, + 0.892, + 0.28 + ], + "angle": 0, + "content": "Table 3. Improved Baseline via Object Enumeration: Adding our object enumeration task to LLaVA training and inference leads to large improvements on THRONE particularly in terms of class-wise precision, \\( P_{\\mathrm{CLS}} \\), over standard LLaVA models, demonstrating a reduction in Type I hallucinations, as well as small reductions in Type II hallucinations judged by POPE and POPE-C." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.439 + ], + "angle": 0, + "content": "COCO images where available. On THRONE, we observe a large increase in classwise precision and therefore \\( F_{\\mathrm{CLS}}^{0.5} \\), particularly for LLaVA-v1.5, demonstrating the ability of our method to reduce Type I hallucinations. Moreover, on POPE and POPE-C, using our object enumeration yields small improvements in precision, indicating reduced Type II hallucinations as well. In the Supplementary Material, we ablate the sampling of negatives during object enumeration training and the effect of removing the object enumeration task during inference." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.446, + 0.628, + 0.461 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.892, + 0.785 + ], + "angle": 0, + "content": "We establish a novel benchmark, THRONE, for evaluating hallucinations generated by LVLMs in free-form image descriptions i.e. Type I hallucinations. Our benchmark utilizes multiple LMs and prompt formats with a simple voting mechanism to yield an accurate evaluation of Type I hallucinations in LVLM responses. We ensure that THRONE is broadly accessible by utilizing open-source LMs capable of running on a single commercial GPU. Using THRONE, we benchmark 11 publicly available LVLMs on two datasets, COCO and Objects365, and demonstrate that limited progress has been made in addressing Type I hallucinations. Moreover, we show how the established benchmark, POPE, underestimates Type II hallucinations, which occur in response to specific questions e.g. yes-no questions. We present results for a completed version (POPE-C) to enable a comparison of Type I hallucinations through THRONE and Type II hallucinations using POPE-C. Finally, we propose a simple data augmentation for LVLM training that can result in a large reduction in Type I hallucinations whilst maintaining or improving Type II hallucination performance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.786, + 0.892, + 0.816 + ], + "angle": 0, + "content": "Limitations and Ethical Considerations are discussed in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.822, + 0.597, + 0.838 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27235" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Neural Information Processing Systems, 35:23716-23736, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.47, + 0.188 + ], + "angle": 0, + "content": "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the International Conference on Computer Vision, pages 2425-2433, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.19, + 0.469, + 0.259 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.261, + 0.469, + 0.327 + ], + "angle": 0, + "content": "[4] Yonatan Bitton, Hritik Bansal, Jack Hessel, Rulin Shao, Wanrong Zhu, Anas Awadalla, Josh Gardner, Rohan Taori, and Ludwig Schmidt. Visit-bench: A benchmark for vision-language instruction following inspired by real-world use, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.33, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems, 33:1877-1901, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.401, + 0.469, + 0.481 + ], + "angle": 0, + "content": "[6] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.484, + 0.469, + 0.564 + ], + "angle": 0, + "content": "[7] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.567, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[8] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.623, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. Proceedings of the International Conference on Learning Representations, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[10] Mark Everingham, Luc Van Gool, Chris K. I. Williams, John Winn, and Andrew Zisserman. The PASCAL Visual Object Classes (VOC) challenge. International Journal of Computer Vision, 2010. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[11] Hugging Face. Text generation inference. https://github.com/huggingface/text-generation-inference, 2023. Accessed: November 10, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[12] Hao Fang, Saurabh Gupta, Forrest Iandola, Rupesh K Srivastava, Li Deng, Piotr Dollar, Jianfeng Gao, Xiaodong He, Margaret Mitchell, John C Platt, et al. From captions to visual concepts and back. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[13] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.243 + ], + "angle": 0, + "content": "[14] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[15] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[16] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[17] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, Qiang Liu, Kriti Aggarwal, Zewen Chi, Johan Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, and Furu Wei. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[18] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the International Conference on Machine Learning, pages 4904-4916, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.634 + ], + "angle": 0, + "content": "[19] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123:32-73, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[21] Kibok Lee, Hao Yang, Satyaki Chakraborty, Zhaowei Cai, Gurumurthy Swaminathan, Avinash Ravichandran, and Onkar Dabeer. Rethinking few-shot object detection on a multi-domain benchmark. In Proceedings of the European Conference on Computer Vision, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[22] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[23] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27236" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[26] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In Proceedings of the Conference on Empirical Methods in Natural Language, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.47, + 0.374 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, 2014. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.47, + 0.431 + ], + "angle": 0, + "content": "[28] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.432, + 0.47, + 0.474 + ], + "angle": 0, + "content": "[29] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.47, + 0.517 + ], + "angle": 0, + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.47, + 0.56 + ], + "angle": 0, + "content": "[31] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.47, + 0.63 + ], + "angle": 0, + "content": "[32] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.47, + 0.714 + ], + "angle": 0, + "content": "[33] Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V Le, Barret Zoph, Jason Wei, and Adam Roberts. The flan collection: Designing data and methods for effective instruction tuning. In Proceedings of the International Conference on Machine Learning, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.47, + 0.798 + ], + "angle": 0, + "content": "[34] Hwee Tou Ng, Siew Mei Wu, Ted Briscoe, Christian Hadiwinoto, Raymond Hendy Susanto, and Christopher Bryant. The CoNLL-2014 shared task on grammatical error correction. In Proceedings of the Eighteenth Conference on Computational Natural Language Learning: Shared Task, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.801, + 0.37, + 0.815 + ], + "angle": 0, + "content": "[35] OpenAI. Gpt-4 technical report, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[36] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[37] Tina D Purnat, Paolo Vacca, Christine Czerniak, Sarah Ball, Stefano Burzo, Tim Zecchin, Amy Wright, Supriya" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.092, + 0.892, + 0.176 + ], + "angle": 0, + "content": "Bezbaruah, Faizza Tanggol, Éve Dubé, Fabienne Labbe, Maude Dionne, Jaya Lamichhane, Avichal Mahajan, Sylvie Briand, and Tim Nguyen. Infodemic signal detection during the COVID-19 pandemic: Development of a methodology for identifying potential information voids in online conversations. *JMIR Infodemiology*, 1(1):e30971, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.26, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[39] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.329, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[40] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. In Proceedings of the Conference on Empirical Methods in Natural Language, pages 4035-4045, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.399, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[41] Giuliano Rossi, Jakub Kolodziej, and Gurvinder Brar. A recommender system for active stock selection. Computational Management Science, 17, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[42] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the International Conference on Computer Vision, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[43] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[44] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. In Advances in Neural Information Processing Systems, pages 200–212, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[45] Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, Jitao Sang, and Haoyu Tang. Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[46] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[47] Hongbin Ye, Tong Liu, Aijia Zhang, Wei Hua, and Weiqiang Jia. Cognitive mirage: A review of hallucinations in large language models. arXiv preprint arXiv:2309.06794, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[48] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27237" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[49] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[50] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.47, + 0.274 + ], + "angle": 0, + "content": "[51] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the ai ocean: A survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[52] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.333, + 0.469, + 0.388 + ], + "angle": 0, + "content": "[53] Yiyang Zhou, Chenhang Cui, Jaehong Yoon, Linjun Zhang, Zhun Deng, Chelsea Finn, Mohit Bansal, and Huaxiu Yao. Analyzing and mitigating object hallucination in large vision-language models, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.39, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[54] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "27238" + } + ] +] \ No newline at end of file diff --git a/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/da9cf0d5-8d8b-4df8-b396-af5ff1ab4d0d_origin.pdf b/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/da9cf0d5-8d8b-4df8-b396-af5ff1ab4d0d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..333f97cb6277d35962a210998e77eacd906056ad --- /dev/null +++ b/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/da9cf0d5-8d8b-4df8-b396-af5ff1ab4d0d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:795b13bd99c06d6026c2c6cfe6bd854ce7871dbf41a2f3bc3fddccee17401e7b +size 1523147 diff --git a/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/full.md b/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6ff83fce0e370c051fbf3c82dc8561148e700210 --- /dev/null +++ b/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/full.md @@ -0,0 +1,314 @@ +# THRONE: An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models + +Prannay Kaul $^{1*}$ Zhizhong Li $^{2\dagger}$ Hao Yang $^{2}$ Yonatan Dukler $^{2}$ Ashwin Swaminathan $^{2}$ C. J. Taylor $^{2}$ Stefano Soatto $^{2}$ VGG, University of Oxford $^{1}$ AWS AI Labs $^{2}$ + +prannay@robots.ox.ac.uk {lzhizhon,haoyng,dukler,swashwin,taylorcj,soattos}@amazon.com + +# Abstract + +Mitigating hallucinations in large vision-language models (LVLMs) remains an open problem. Recent benchmarks do not address hallucinations in open-ended free-form responses, which we term "Type I hallucinations". Instead, they focus on hallucinations responding to very specific question formats—typically a multiple-choice response regarding a particular object or attribute—which we term "Type II hallucinations". Additionally, such benchmarks often require external API calls to models which are subject to change. In practice, we observe that a reduction in Type II hallucinations does not lead to a reduction in Type I hallucinations but rather that the two forms of hallucinations are often anti-correlated. To address this, we propose THRONE, a novel object-based automatic framework for quantitatively evaluating Type I hallucinations in LVLM free-form outputs. We use public language models (LMs) to identify hallucinations in LVLM responses and compute informative metrics. By evaluating a large selection of recent LVLMs using public datasets, we show that an improvement in existing metrics do not lead to a reduction in Type I hallucinations, and that established benchmarks for measuring Type I hallucinations are incomplete. Finally, we provide a simple and effective data augmentation method to reduce Type I and Type II hallucinations as a strong baseline. + +# 1 Introduction + +This paper proposes a benchmark to evaluate hallucinations by large vision-language models (LVLMs) when generating free-form responses, specifically detailed descriptions, based on a given image. + +The rapid advancement in large language models (LLMs) [52] has pushed the development of large vision-language models (LVLMs). [1, 6, 7, 17, 24, 25, 30, 36, 44, 48, 54] LVLMs take input text and images and generate text responses to enable multi-modal perception and com + +prehension. + +LVLMs are largely built on LLMs and therefore inherit both their advantages and their disadvantages. LLMs have been shown to produce hallucinations [47, 51], generated text responses that are coherent and plausible but factually incorrect. LVLMs echo this behavior with generated text contradicting with the visual or text input [53]. Hallucinations prevent the use of LVLMs in safety-critical situations and therefore evaluating and mitigating hallucinations in LVLMs is crucial for their deployment in such settings. Determining the presence and cause of hallucinations in LVLMs remains an open question [45, 53]. + +We divide LVLM hallucinations into two types. Type I hallucinations occur in response to open-ended questions with a very large set of possible responses — e.g. What is happening in this image?. Type II hallucinations are incorrect responses to a factual question regarding a specific concept about the image with a fixed set of options such as yes/no—e.g. Is there a traffic light in this image? Fig. 3 illustrates the difference between these two types of hallucination. Reducing hallucinations in both cases is required for useful, multi-purpose LVLMs. However, later in Fig. 4 we observe that the same LVLM can give contradicting answers when being evaluated for Type I vs. Type II hallucinations. This implies that measuring and reducing one type does not necessarily reduce the other. + +Existing works to evaluate LVLMs often avoid direct quantification of hallucinations and instead develop comprehensive benchmarks that judge various other desirable abilities such as: optical character recognition, fine-grained recognition and attribute detection [14, 22, 32]. The extent of hallucinations in these benchmarks is obfuscated, since it is only one of many factors influencing other metrics. It requires human effort to inspect individual predictions. There are two major established works which specifically develop a benchmark for evaluating hallucinations in vision-language models: POPE [26] and CHAIR [40], which we discuss in detail in Sec. 2. However, we observe they both have shortcomings in effectively evaluating hallucinations: + +![](images/086844dd337d4200da7236eb37d1d27179d6b14b2daac80a4009bc170907dc43.jpg) +Figure 1. THRONE (Ours): LVLMs are prompted with a concept neutral instruction. An external LM performs abstractive QA on the response to establish the existence of Type I hallucinations. + +POPE [26] is a recent work addressing Type II hallucinations with respect to object classes. However, we find Type I and Type II hallucinations are disconnected, and that POPE gives an incomplete picture on LVLM hallucinations. Moreover, POPE systematically under-samples negative object categories leading to a large underestimation of Type II hallucinations (see Sec. 5.4). + +CHAIR [40] does address Type I hallucinations—establishing object category hallucination in short image captions using simple text matching. However, CHAIR is not suited to current LVLMs because the simple text matching it employs cannot comprehend abstract or hypothetical concepts present in today's LVLMs (see Fig. 4). Further, hand-crafted rules for each set of classes are required for usable text matching, and trivial model answers can attain a perfect CHAIR score. + +To address the issues of current LVLM hallucination benchmarks, we propose THRONE (Text-from-image Hallucination Recognition with Object-probes for open-ended Evaluation). THRONE leverages language models (LMs) to evaluate Type I hallucinations in free-form, open-ended image descriptions with respect to a pre-defined object vocabulary of interest. By utilizing LMs, THRONE is able to accurately judge whether an object mentioned in an LVLM response is implied to exist in the image or is abstractly mentioned with no implication about its existence (see Fig. 4). + +Moreover, in THRONE, we provide easy access to our benchmark, by leveraging open-source LMs that can run on common GPUs, instead of relying on closed-source commercial models [35] that are subject to arbitrary change, as done in other works [4, 22, 32]. Through combining multiple open-source LMs, we mitigate any single-model biases in judging hallucinations when calculating Type-I hallucination scores with THRONE. + +We make four contributions: first, we establish an accurate and accessible benchmark to quantitatively evaluate object hallucinations in free-form responses, leveraging LMs to judge the existence of Type I hallucinations—quantitatively showing half the judgement errors of CHAIR; + +![](images/d0ffd45c22381ac0c708d6e5c516b7ac37149eda31e5a621be3c195e88db01d0.jpg) +Figure 2. POPE: Questions with specific concepts prompt an LVLM directly to evaluate Type II hallucinations [26]. Handcrafted rules parse LVLM responses to give yes/no labels. + +second, we evaluate a number of current LVLMs on THRONE and demonstrate that observed progress in reducing Type II hallucinations do not translate to a corresponding reduction in Type I hallucinations; third, we show a recent method, POPE, is inaccurately capturing the extent of Type II hallucination discovery due to its sampling strategy; we mitigate this issue in our implementation of THRONE while presenting results for a complete version of POPE; and fourth, we provide a simple augmentation of visual instruction tuning data which significantly improves performance on Type I hallucinations while maintaining or improving Type II hallucination performance. + +# 2 Related Work + +Hallucination Benchmarks for LVLMs. In response to the development of LVLMs, few evaluation benchmarks focusing on hallucinations have been introduced. CHAIR [40], one of the first works to assess hallucinations, is designed for short image captions. CHAIR uses a fixed set of object classes (extended with their synonyms) as a set of text strings to find predicted object classes in image captions via exact text matching. Subsequently, each class matched in the caption is compared to the COCO [27] captions and bounding box annotations to establish object hallucinations. CHAIR was developed prior to current instruction-tuned LVLMs which produce long free-form responses ( $10 \times$ longer than image captions) with a diverse vocabulary, limiting its applicability to modern LVLMs. As shown in Fig. 3, exact text matching in CHAIR is prone to incorrectly matching the vocabulary classes with abstract concepts in the free-form response and the and synonyms of the class list must be manually selected to prevent confusion during evaluation. For example, in CHAIR the word "chair" will match all responses for the phrase "toilet seat", because the exact text matching classifies "seat" as the COCO class "chair". Finally, CHAIR metrics compare the overall number of predicted objects to the overall number of predicted objects judged to be hallucinations—ignoring the recall of ground-truth objects and the distribution of object classes. This means that a single correct + +prediction across the entire evaluation dataset along with a generic response otherwise e.g. A natural scene, achieves a perfect score $(\frac{0 \text{ hallucinated objects}}{1 \text{ predicted objects}} = 0.0)$ . See the Supplementary Material for a full overview of CHAIR. In contrast, THRONE uses pre-trained LMs that go beyond direct synonym matching, to automatically judge the existence of concepts and hallucinations in free-form responses. In addition, our method considers both recall and precision to yield a holistic benchmark and does not require any manual curation of synonyms. THRONE and CHAIR both evaluate Type I hallucinations—hallucinations in response to concept-neutral prompts e.g. Describe this image in detail. + +POPE [26] is a recently proposed benchmark to evaluate object hallucinations in LVLMs—specifically Type II hallucinations, in which an LVLM is directly queried with a yes-no question regarding the existence of a particular object of the form: Is there $\{\mathrm{a / an}\}$ {object_class_name} in the image?. The LVLM response is parsed using simple rules to determine whether a Type II hallucination has occurred. Precision and recall metrics are compiled using the parsed LVLM responses and the ground-truth annotation data. Despite focusing on measuring hallucinations, POPE only queries an LVLM with 3 positive and 3 sampled negative questions per COCO image i.e. the evaluation is artificially balanced. This means many potential hallucinations with respect to the COCO categories are not captured by their method. In Sec. 5.4, we show POPE dramatically underestimates Type II hallucinations and present the results of a complete version. Our method, THRONE, evaluates the prevalence of Type I hallucinations, which we observe are disjoint to Type II hallucinations. + +Comprehensive Benchmarks for LVLMs have recently grown in number. MMBench [32] and MM-Vet [49] assess various aspects of LVLM performance such as: color perception, celebrity recognition, and numerical calculation. However many of these works integrate evolving APIs which are modified often (or even discontinued) and are inherently stochastic. Over time this greatly reduces the consistency of these benchmarks. Exceptions are MME [14] and SEED-Bench [22], but the impact of Type II hallucinations on final metrics is conflated with a number of other aspects of model performance. Our method, THRONE, directly addresses Type I hallucinations, only making use of open-source language models and datasets. + +Large Vision Language Models (LVLMs) have rapidly developed by harnessing advancements in large language models (LLM) [5, 39, 43] and by directly integrating pretrained LLMs into their architectures. In contrast to earlier vision-language models such as CLIP [18, 38], LVLMs are generally comprised of a pretrained LLM and image encoder, aligned with a connector module of varying complexity. Some works are highlighted here. Frozen [44] is + +
ImagePromptResponse
What is happening in this image?In the image, a man is hanging clothes on a clothesline that is attached to the back of a moving car.
Is there a traffic light in this image?Yes, there is a traffic light in the image, and the man is standing on the back of a car near it.
+ +Figure 3. Type I vs. Type II Hallucinations: (Top) LVLMs prompted with concept-neutral instructions produce Type I hallucinations. (Bottom) Instructions specifying a concept produce Type II hallucinations. Examples from LLaVA-v1.5 [29]. + +an early work fine-tuning the vision encoder to dynamically prefix the prompt to a frozen LLM. Flamingo [1] combines visual and language features using cross-attention layers in an otherwise frozen LLM. BLIP-2 [25] uses a frozen image encoder to learn a Querying-Transformer (Q-Former) on image-text pairs that is used as the connector. This architecture is adapted in [7] for dialogue via training with visual instruction tuning. LLaVA [30] uses COCO [27] annotations and GPT-4 [35] to generate visual instruction tuning data in a plain-text pipeline. Combining this generated data with standard VQA datasets (VQAv2 etc.) further boosts performance [29]. Different works modify the training approach by using efficient adaptation [16, 48] multiple training stages [3] or introduce the use of discrete tokens for localization [36]. We note however, that most of these works evaluate performance on traditional vision-language datasets like VQAv2 [2], which do not consider the extent of hallucinations, a known problem with LLMs [47, 51]. + +# 3 THRONE + +Recap of existing methods. POPE [26] and other benchmarks (MME [14], MMBench [32]) directly query LVLMs with a restricted desired answer space, e.g. yes-no (MME, POPE) and A-B-C-D multiple choice (MMBench), as shown in Fig. 2. These benchmarks only consider such short answer formats, whereas a key quality of LVLMs is in their ability to generate free-form coherent text. Moreover, POPE, which addresses Type II hallucinations, undersamples negative classes meaning hallucinations are dramatically underestimated (see Sec. 5.4 and Fig. 7). In contrast, we skip class subsampling and enumerate all classes for every image, ensuring a full evaluation of Type I hallucinations of the ground-truth classes. + +CHAIR [40] also evaluates Type I hallucinations, but was developed when typical vision-language models could only generate short and simple captions similar to those in COCO Captions [12]. Moreover, it lacks accurate comprehension of responses (see the right side of Fig. 4) and ig- + +![](images/634cddb0c8f4c8f79f8782ffb0b14a0cc478117e417498c07dc7450b376f3331.jpg) +Figure 4. A Comparison of POPE, CHAIR and THRONE: Directly querying LVLMs for object existence (person, banana etc.) using concept-specific instructions, as in POPE (bottom left), does not produce the same hallucinations as using concept-neutral instructions (right). We highlight the Type I hallucinations in orange. CHAIR relies on exact text matching to a fixed set of objects and synonyms, thus incorrectly labels "customers" and "shoppers" as hallucinations, highlighted in red. THRONE is designed for the rich vocabulary and the free-form generations of modern LVLMs by harnessing LMs to establish object existence. By using an LM to pass judgement, our evaluation correctly captures "customers" and "shoppers" as hypothetical content in the free-form generation. + +Type II Hallucination Evaluation (POPE) + +
Instruction [GT]Response [Prediction]
Is there... +a person [No] +a banana [Yes] +an apple [No] +an orange [No] +...in the image?No, there is no person in the image. [No] +Yes, there is a banana in the image. [Yes] +No, there is not an apple in the image. [No] +No, there is no orange in the image. [No] +No Type II Hallucinations Found
+ +Type I Hallucination Evaluation + +
InstructionResponse
Describe this image in detail.The image showcases a fruit stand at a grocery store, featuring a variety of fruits on display. There are several bunches of bananas, with some placed in the foreground and others in the background. The bananas are arranged in different sections, creating an appealing presentation for customers. In addition to the bananas, there are also apples and oranges on display. The apples are located towards the left side of the image, while the oranges are placed in the middle and right side of the stand. The fruits are well-organized and presented in an attractive manner, making it an inviting sight for shoppers. Type I Hallucinations Present and Found
MSCOCO Object Prediction from Description
HumanCHAIRTHRONE (Ours)
banana apple orangebanana person apple orangebanana apple orange
+ +nores the recall of ground-truth objects. In Sec. 5.5, we describe quantitative evaluations, using a human oracle, which demonstrate THRONE halves the rate of hallucination misjudgement in CHAIR. See Sec. 2 and the Supplementary Material for more details on CHAIR and its shortcomings. Fig. 4 shows an overview of the three aforementioned methods: POPE, CHAIR and our method, THRONE. + +# 3.1 Evaluating Hallucinations with THRONE + +To address these limitations, we propose a framework, THRONE, shown in Fig. 1, to evaluate the prevalence of Type I hallucinations in LVLM responses conditioned on an image and a neutral text prompt. + +For each image in a labeled dataset, $\mathcal{I}$ , addressing a set of classes, $\mathcal{C}$ , the LVLM is queried with the same instruction: Describe this image in detail., regardless of image content. The LVLM response, which is expected to be long free-form text containing an image description, is generated and stored. Next, a publicly available, open-source, external language model (LM) performs abstractive question answering (AQA) using the LVLM response as context and a question of the form: Please answer yes or no. Is there $\{\mathrm{a / an}\}$ {object class name} in this image? or similar, for every class in $\mathcal{C}$ (right side of Fig. 1). By selecting an appropriate LM and using a simple prompt template (see Sec. 4 for specific details), we ensure the AQA response is either yes or no—our method does not require any additional parsing. This is in contrast to other works which require added parsing or interpretation by a closed-source model. + +After performing AQA on each response generated by the LVLM for every class in $\mathcal{C}$ , we obtain an array of pre + +dicted labels: + +$$ +\hat {\mathbf {Y}} \in \{0, 1 \} ^ {| \mathcal {I} | \times | \mathcal {C} |} \tag {1} +$$ + +where $0/1$ indicates a negative/positive existence judgement by the LM with respect to the relevant LVLM response. + +Similarly using the ground-truth data for $\mathcal{I}$ , an array of ground-truth labels can be constructed: + +$$ +\mathbf {Y} \in \{0, 1 \} ^ {| \mathcal {I} | \times | \mathcal {C} |} \tag {2} +$$ + +Using these two arrays, we calculate four metrics: (1) Overall Precision, $P_{\mathrm{ALL}}$ ; (2) Overall Recall, $R_{\mathrm{ALL}}$ ; (3) Class-wise Precision, $P_{\mathrm{CLS}}$ ; and (4) Class-wise Recall, $R_{\mathrm{CLS}}$ . Overall metrics are calculated in a class-agnostic manner. Class-wise metrics are calculated in a class-conscious manner by computing precision and recall for each category separately and then averaging. This follows common practice in object detection and instance segmentation [10, 27]. + +False positives in LVLMs reduce precision and are dominated by hallucinations—precision indicates the extent of Type I hallucinations in LVLM responses. The recall metrics inform the level of class coverage by an LVLM when producing image descriptions. The class-wise metrics give a general measure of performance as the overall metrics are skewed towards the most common categories. A common way to combine precision, $P$ , and recall, $R$ , metrics is through the generalized F score, $F_{\beta}$ : + +$$ +F ^ {\beta} = \left(1 + \beta^ {2}\right) \cdot \frac {P \cdot R}{\left(\beta^ {2} \cdot P\right) + R} \tag {3} +$$ + +Prior work such as POPE [26], use the common balanced F-score (or $F^1$ -score) which equally weights precision and + +recall. However, given that THRONE is concerned with measuring hallucinations and is therefore particularly interested in precision, we choose $\beta = 0.5$ or the $F^{0.5}$ -score, thus weighing precision twice as important as recall. The $F^{0.5}$ -score is commonly used in pandemic misinformation filters [37], recommender systems [34], active stock selection [41] and other areas where false positives are costlier than false negatives. From this we can calculate the overall and class-wise $F^{0.5}$ -scores, $F_{\mathrm{ALL}}^{0.5}$ and $F_{\mathrm{CLS}}^{0.5}$ , respectively. To mitigate against class imbalance issues, we use $F_{\mathrm{CLS}}^{0.5}$ as the principle metric of comparison between LVLMs in THRONE. + +# 3.2 Ensuring Robustness via Ensembling + +Any LM used for AQA in THRONE may misjudge Type II hallucinations—no LM is perfect. Fig. 5 (top) shows two different variants from the same model family (FLANT5 [33]), yielding opposite responses when prompted with the same response and question. Moreover, an LM may yield different answers to semantically identical questions, despite conditioning on the same response—shown in Fig. 5 (bottom). To ensure THRONE is robust to spurious performance by any one LM, we ensemble various LMs and semantically equivalent question formats. The use of $N$ distinct LMs and $M$ distinct question formats yields $NM$ answers for each (LVLM response, class) combination. This set of $NM$ answers is combined based on equal voting by each (LM, question) pair to "elect" the predicted answer. Continuing the notation from Eq. (1), stacking predictions from each (LM, question) pair yields a 3D array of predicted labels: $\bar{\mathbf{Y}} \in \{0,1\}^{|\mathcal{I}_{\mathrm{OURS}}| \times |\mathcal{C}_{\mathrm{OURS}}| \times NM}$ . To combine the answers from each (LM, question) pair via voting, we require agreement between at least $k$ answers. Where sufficient agreement does not exist we introduce an "ignore" label. Mathematically, the elements in $\hat{\mathbf{Y}}$ (from Eq. (1)) are calculated as: + +$$ +\hat {Y} _ {i, j} = \left\{ \begin{array}{l l} 0, & \sum_ {k = 1} ^ {N M} \bar {Y} _ {i, j, k} \leq (N M - k) \\ 1, & \sum_ {k = 1} ^ {N M} \bar {Y} _ {i, j, k} \geq k \\ - 1, & \text {o t h e r w i s e} \end{array} \right. +$$ + +where an "ignore" label exists in $\bar{\mathbf{Y}}$ , it is removed from the calculations of $P_{\mathrm{ALL}}$ , $R_{\mathrm{ALL}}$ , $P_{\mathrm{CLS}}$ , and $R_{\mathrm{CLS}}$ , as we cannot be confident in the AQA process for that particular (LVLM response, class) combination. The choice of $k$ reflects the desired level of confidence. We make use of a unanimous voting mechanism ( $k = NM$ )—use the prediction only if all (LM, question) pairs agree, otherwise ignore. Human evaluation of our benchmark and choice of voting mechanism is found in the Supplementary Material. Once we have applied this voting mechanism, we can calculate the metrics described at the end of Sec. 3.1. + +![](images/9839b4821e70a308c64e94e2823eca3305e1faa403fb063339a77b55ea633175.jpg) +Figure 5. AQA Ensembling in Evaluation: Using different LMs or different prompts when running AQA on LVLM generated responses can produce opposing answers to identical prompts or identical LMs. To ensure THRONE is robust to this, we ensemble multiple LMs and multiple prompts in our evaluation pipeline. + +# 4 Implementation + +We provide details of our framework regarding the selection of public datasets, public LMs and LM prompts. + +# 4.1 Dataset + +Any proper evaluation of object hallucinations (a type of false positive error) requires knowing, with certainty, which classes are absent in an image. In our benchmark, we use COCO [27] for a number of reasons: (1) its annotations of 80 categories are exhaustive at an image level (image-level recall $\approx 99\%$ [27])—if there are many book instances in a COCO image, at least one is annotated with bounding boxes; (2) many LVLMs are partly trained on COCO data and so should be familiar with the set of categories; (3) its images generally contain complex scenes suitable for generating long free-form descriptions unlike image recognition datasets like ImageNet [8]. + +We utilize the validation set of COCO 2017, which contains $|\mathcal{I}| = 5000$ images and $|\mathcal{C}| = 80$ categories. Using the single LVLM text prompt Describe this image in detail., we generate 5000 responses. As we query each LVLM response for each category in $\mathcal{C}$ , a single LM performs AQA $|\mathcal{I}| \times |\mathcal{C}| = 400k$ times across the LVLM responses—one instance of AQA per (image response, class) pair. In Sec. 5.3, we also present results using Objects365 [42] which is rarely used in LVLM training. + +# 4.2 Language Models + +To assess Type I hallucinations in an LVLM response using THRONE, we require a language model (LM) which can answer questions on the existence of object categories based on the LVLM response. MMBench [32] makes use of a ChatGPT model to identify multiple choice answer selections and still reports mistakes. In our experience, some LMs give rather incoherent judgements when used to assess hallucinations when the prompt is changed (see supplemental material). Therefore for THRONE, we choose FLAN-T5 models [33, 39]. We make this choice because FLAN-T5 model family: (1) have undergone instruction tuning with thousands of tasks [33]; (2) are open-source and + +
ModelL\( P_{ALL} \)\( R_{ALL} \)\( F_{ALL}^{1} \)\( F_{ALL}^{0.5} \)\( P_{CLS} \)\( R_{CLS} \)\( F_{CLS}^{1} \)\( F_{CLS}^{0.5} \)
Adapter-v2 [15]51463.673.368.165.368.270.669.468.7
Adapter-v2.1 [15]51263.873.768.465.567.471.269.368.1
InstructBLIP [7]52570.874.372.571.577.271.974.576.1
Otter-Image [23]25733.031.232.132.725.216.920.222.9
MiniGPT4 [54]47381.759.869.076.179.961.869.775.5
MiniGPT-v2 [6]38179.066.672.376.277.667.071.975.2
mPLUG-Owl [48]55555.571.962.658.166.368.367.366.7
LRV-Instruction-v2 [28]10382.056.767.075.378.458.867.273.5
LLaVA-v1.3* [30]53280.565.272.176.979.965.371.976.5
LLaVA-v1.5 [29]50968.161.064.466.669.956.462.566.8
LLaVA-Mistral [19, 31]52486.871.878.383.684.464.270.877.5
+ +Table 1. THRONE Results with COCO for a selection of instruction-tuned LVLMs. We select $F_{\mathrm{CLS}}^{0,5}$ as the principal metric for evaluation in our benchmark to balance across classes and to prioritize precision (which reflects the extent of hallucination) over recall. Best and second-best performance are denoted by blue and red, respectively. *Our implementation using official code to enable fair comparison. $L$ corresponds to the median response length (measured in # of characters). + +therefore accessible to the community; (3) can fit locally on a single GPU for the models we consider; (4) follow user's instruction to only respond yes or no; and (5) are optimized for use in the free and public Text-Generation-Inference API [11] for acceleration. As described in Sec. 3.2, we utilize $N$ LMs to ensure our method is robust. Specifically, we use $N = 3$ variants of FLAN-T5, namely: FLAN-T5-Base (250M parameters), FLAN-T5-Large (780M parameters), and FLAN-T5-XL (3B parameters). + +# 4.3 Prompt Ensembling + +To guarantee each of these FLAN-T5 variants faithfully produce responses of either yes or no only during AQA, we use the following input template to each LM, reflecting the format used when training FLAN-T51: + +Text: {LVLM Response} Read the text about an image and answer the question. +Question: Please answer yes or no. +{Question} + +We use $M = 3$ semantically identical questions: + +- Is there a/an {class_name} in this image? +- Does the text imply a/an {class_name} is in the image? +- Does the text explicitly mention a/an {class_name} is in the image? + +As outlined in Sec. 3.2, we use a unanimous voting mechanism to combine the answers from each (LM, question) pair and so $k = 3 \times 3 = 9$ . + +# 5 Evaluation Results + +In this section, we: (1) outline our LVLM selection and reasoning; (2) present THRONE on COCO for evaluating Type I hallucinations; (3) extend THRONE to Objects365 (containing a larger vocabulary); (4) analyze and extend POPE to enable improved evaluation of Type II hallucinations; and (5) highlight results from our ablation studies found in the Supplemental Material. + +# 5.1 Models + +For fair comparison between existing LVLMs, each publicly available model we evaluate uses an LLM with $\sim 7\mathrm{B}$ parameters. The LVLMs generally have different sized image encoders—e.g. LLaVA [30] uses a CLIP ViT-L/14 [9, 38] with an input resolution of $336\times 336$ , while InstructBLIP [7] uses a ViT-g/14 [50] trained with EVA [13] and an input resolution of $224\times 224$ . Note that image encoder size and resolution is not something we can easily control in a pre-trained model. Each model we consider contains instruction tuning in the final training phase—instruction tuned models provide free-form descriptions; THRONE focuses on models that can generate free-form descriptions. E.g. we leave out BLIP-2 [25] (response median length 31 characters) in favor of InstructBLIP (median response length 525). + +# 5.2 THRONE Results on COCO + +Results are shown in Tab. 1. The principal metric that we use to judge model performance is the classwise $F^{0.5}$ -score (highlighted gray). We also report all the metrics outlined in Sec. 3.1, $(P, R, F^1, F^{0.5})$ for overall (left) and class-wise averaging (right), utilizing the unanimous voting presented in Sec. 3.2. See the Supplementary Material for results and an analysis of different voting mechanisms. These results demonstrate that improvements on other benchmarks (POPE, MME, MMBench etc.) may be orthogonal and potentially at odds with improved performance on THRONE. Using the results of the 11 LVLMs that we evaluate, THRONE and POPE, which measure Type I and Type II hallucinations, respectively, have a Spearman's rank correlation coefficient of just 0.2, and THRONE vs POPE-C has just 0.4—the relationship between performance on POPE and THRONE on the same dataset is far from monotonic. For class-wise precision— $P_{\mathrm{CLS}}$ , the best performing models hallucinate $\sim 20\%$ of the objects. We show in the Supplementary Material that the vast majority of false positive objects in the free-form image descriptions evaluated are direct hallucinations rather than misclassifications of visually similar objects (e.g. mistaking a squash racket for a tennis racket). These results demonstrate that much work still remains to adequately suppress Type I hallucinations in LVLMs. + +# 5.3 THRONE Results on Objects365 + +Many LVLMs train on COCO directly or indirectly, thus to demonstrate generality we apply THRONE to the Objects365 dataset [42]. Like COCO, Objects365 aims to be exhaustive in its image-level class labeling (it aims to label at least one instance for each class present), but it has a larger object vocabulary and is not used as training data for the LVLMs that we evaluate. To gather a manageable subset of the Objects365 validation set (80k images), we use the natural sampling algorithm from [21], resulting in 5110 im + +
ModelPALLRALLF1ALLF0.5ALLPCLSRCLSF1CLSF0.5CLS
Adapter-v2 [15]46.733.939.343.448.928.536.042.8
Adapter-v2.1 [15]46.834.039.443.548.828.836.242.8
InstructBLIP [7]54.537.244.249.853.733.641.348.0
Otter-Image [23]21.412.716.018.89.54.46.07.7
MiniGPT4 [54]53.032.940.647.249.931.939.044.9
MiniGPT-v2 [6]54.536.043.449.451.334.641.346.8
mPLUG-Owl [48]43.733.437.841.248.229.036.242.6
LRV-Instruction-v2 [28]57.526.736.546.751.426.635.143.3
LLaVA-v1.3* [30]57.632.941.950.152.630.538.645.9
LLaVA-v1.5 [29]54.039.545.650.353.934.341.948.4
LLaVA-Mistral [19, 31]58.339.146.953.157.835.944.351.5
+ +Table 2. THRONE Evaluation with Objects365. Evaluation results for a selection of instruction-tuned LVLMs, we use a subset of Objects365 for the THRONE evaluation. Best and second-best performance are denoted by blue and red, respectively. *Our implementation using official code to enable fair comparison. + +![](images/f8bccb59c82dd1aac12a91be6f40a11c47060dee4bf85013f0e2030a36ed26b4.jpg) +Figure 6. Comparing THRONE on COCO and Objects365. We observe that despite variations in the data distribution, THRONE metrics, which measure Type I hallucinations generalize and have a strong Spearman's rank correlation of $r = 0.900$ . One model (red circle) designated as an outlier and ignored when calculating ranking correlation. + +ages (the COCO validation set has 5k images). We present the results for THRONE on Objects365 in Tab. 2. Figure 6 shows the strong correlation in THRONE performance between evaluating on COCO and Objects365. This demonstrates that measuring Type I hallucinations in an LVLM using THRONE with a relatively small dataset like COCO, is indeed indicative of the intrinsic level of Type I hallucination in a given LVLM. + +# 5.4 Completing POPE for Type II Hallucinations + +Our experiments have used THRONE to evaluate the prevalence of Type I hallucinations in LVLM responses on COCO. POPE evaluates Type II hallucinations on COCO, but we find POPE is largely incomplete. First, POPE only evaluates on 500 COCO validation set images. Second, for each image only a subset of classes (at most 12) are evaluated—each image is only queried with $15\%$ of possible questions. Finally, POPE artificially balances evaluation questions between positives and negatives, despite object class existence in images being inherently imbalanced. + +![](images/ebc4c81c9640a9eb8b01b8e4c79436addb25706a877a5e3844e203e900861bc5.jpg) +Figure 7. Instability of POPE to complete evaluation of Type II Hallucinations. Extending POPE to an exhaustive analysis on all COCO images and classes (POPE-C) leads to a dramatic reduction in performance across all 11 models. POPE is sensitive to the sampling mechanism, and by undersampling negative classes, POPE severely underestimates Type II hallucinations in LVLMs. + +The above reasons make the evaluation of Type II hallucinations using POPE insufficient (see Sec. 2 for more details on POPE). To correct this, we complete POPE by using all images to exhaustively query each LVLM for every class in the COCO vocabulary, as done in THRONE. We name this POPE-Complete (POPE-C). Fig. 7 shows the extreme difference in evaluation between POPE and our exhaustive version, POPE-C—reporting $F^{1}$ -score (POPE does not utilize $F^{0.5}$ -score). As POPE only evaluates at most 9 negative classes, only a small subset of potential hallucinations of COCO classes are evaluated, thereby heavily underestimating the extent of Type II hallucination. For each LVLM analyzed, we observe a large—in many cases an extreme—reduction in precision and therefore in F1-score. + +Our evaluation contains three pairs of LVLMs in which one is the follow-up work to the other, where each follow-up work generally trains on more data for more tasks with a more advanced language model. Comparing the right hand side of Fig. 7 to Tab. 1, we observe that these follow-up works generally show an improvement in POPE (and POPE-C), but surprisingly indicate a small reduction in performance on THRONE with COCO. This observation suggests that progress in reducing Type II hallucinations can be orthogonal to reducing Type I hallucinations. + +# 5.5 Ablations + +In the Supplementary Material we present three key ablation experiments and give an executive summary of results here. + +First, after subsampling COCO images and LVLM responses, we replace the LMs in THRONE with human judgement as an oracle for Type I hallucination occurrence. When comparing THRONE and CHAIR with human judgements, we estimate using THRONE improves the precision of judging Type I hallucinations to $96\%$ versus $91\%$ when using CHAIR—this reduces the false discovery rate by more than $50\%$ . Note that we find most estimated errors in THRONE arise from the particular class definitions in COCO, e.g. the COCO class tv includes computer monitors, which the oracle judgement is aware of. + +Second, we apply the same class (and image) sampling + +strategy as in POPE to THRONE and show this sampling overestimates $F_{\mathrm{CLS}}^{0.5}$ by an average of 12.3 points compared to the complete use of classes in THRONE (Tab. 1). + +Finally, we vary the choice of $k$ i.e. the voting mechanism used to combine answers from multiple (LM, question) pairs. We use the unanimous voting mechanism ( $k = 9$ ) in THRONE to minimize the false discovery rate and find the valid alternatives of simple majority ( $k = 5$ ) or all-but-one ( $k = 8$ ) voting mechanisms have strong correlations and rank correlations across all compute metrics, in THRONE, of $> 0.99$ and $> 0.94$ , respectively. + +# 6 Improved Baselines + +Much needs to be done to study Type I hallucinations. As a first step to their mitigation, we demonstrate a baseline method to augment the visual instruction tuning data for LLaVA models [29, 30], yielding improvements on THRONE while maintaining similar performance regarding Type II hallucinations on POPE. + +# 6.1 Visual Instruction Tuning Data Augmentation + +Similar to chain-of-thought learning [46], during instruction tuning, we augment all visual instruction tuning samples constructed by LLaVA [30] by preponding the task of enumerating a list of objects (present and absent) and indicating approximate locations, if applicable. Other than this, the LLaVA data and training pipeline remains unchanged. To generate the new data for this object enumeration task, we use the same COCO bounding box annotations used to generate the vision instruction tuning data in LLaVA. The simple text-only format (no special tokens) we use is: + +```txt +Instruction: Give a list of objects and locations in the image. +Response: {class_name_1} {{location_1}}/absent] {class_name_N} {{location_N}}/absent] +``` + +where location_i is a plain text indicator representing the location of the center point of the relevant object in the image on a $3 \times 3$ grid (e.g. bottom left). To provide negatives in the training data, if class_name_i is not present, we use the plain text indicator absent. Prior work [53] shows that classes that frequently co-occur in the training data are the most common hallucinations, therefore we bias our negative sampling towards class pairs that frequently co-occur using a correlation matrix. We detail and ablate this choice in the Supplementary Material. + +# 6.2 Improved Baseline Results + +Tab. 3 shows the result of evaluating our improved baseline method on THRONE, POPE, and POPE-C. During inference, we approximate our training data augmentation by first prompting the LVLM to perform the object enumeration tasks and then generating a response to the prompt from the relevant benchmark. We additionally show results of utilizing VisualGenome bounding box annotations [20] on the + +
ModelObject +Enumeration DataTHRONEPOPEPOPE-C
PCLSRCLSF0.5CLSPRF1PRF1
LLaVA-v1.3X79.965.376.558.098.473.07.799.214.3
COCO83.268.879.973.288.280.09.869.417.2
COCO + VG86.267.081.583.082.582.813.850.421.7
LLaVA-v1.5X69.956.466.881.990.886.158.785.769.7
COCO87.276.684.988.685.387.058.987.570.4
COCO + VG86.177.084.189.883.786.764.586.173.7
+ +Table 3. Improved Baseline via Object Enumeration: Adding our object enumeration task to LLaVA training and inference leads to large improvements on THRONE particularly in terms of class-wise precision, $P_{\mathrm{CLS}}$ , over standard LLaVA models, demonstrating a reduction in Type I hallucinations, as well as small reductions in Type II hallucinations judged by POPE and POPE-C. + +COCO images where available. On THRONE, we observe a large increase in classwise precision and therefore $F_{\mathrm{CLS}}^{0.5}$ , particularly for LLaVA-v1.5, demonstrating the ability of our method to reduce Type I hallucinations. Moreover, on POPE and POPE-C, using our object enumeration yields small improvements in precision, indicating reduced Type II hallucinations as well. In the Supplementary Material, we ablate the sampling of negatives during object enumeration training and the effect of removing the object enumeration task during inference. + +# 7 Conclusion + +We establish a novel benchmark, THRONE, for evaluating hallucinations generated by LVLMs in free-form image descriptions i.e. Type I hallucinations. Our benchmark utilizes multiple LMs and prompt formats with a simple voting mechanism to yield an accurate evaluation of Type I hallucinations in LVLM responses. We ensure that THRONE is broadly accessible by utilizing open-source LMs capable of running on a single commercial GPU. Using THRONE, we benchmark 11 publicly available LVLMs on two datasets, COCO and Objects365, and demonstrate that limited progress has been made in addressing Type I hallucinations. Moreover, we show how the established benchmark, POPE, underestimates Type II hallucinations, which occur in response to specific questions e.g. yes-no questions. We present results for a completed version (POPE-C) to enable a comparison of Type I hallucinations through THRONE and Type II hallucinations using POPE-C. Finally, we propose a simple data augmentation for LVLM training that can result in a large reduction in Type I hallucinations whilst maintaining or improving Type II hallucination performance. + +Limitations and Ethical Considerations are discussed in the Supplementary Material. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in + +Neural Information Processing Systems, 35:23716-23736, 2022. 1, 3 +[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the International Conference on Computer Vision, pages 2425-2433, 2015. 3 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 3 +[4] Yonatan Bitton, Hritik Bansal, Jack Hessel, Rulin Shao, Wanrong Zhu, Anas Awadalla, Josh Gardner, Rohan Taori, and Ludwig Schmidt. Visit-bench: A benchmark for vision-language instruction following inspired by real-world use, 2023. 2 +[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems, 33:1877-1901, 2020. 3 +[6] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 1, 6, 7 +[7] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1, 3, 6, 7 +[8] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2009. 5 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. Proceedings of the International Conference on Learning Representations, 2021. 6 +[10] Mark Everingham, Luc Van Gool, Chris K. I. Williams, John Winn, and Andrew Zisserman. The PASCAL Visual Object Classes (VOC) challenge. International Journal of Computer Vision, 2010. 4 +[11] Hugging Face. Text generation inference. https://github.com/huggingface/text-generation-inference, 2023. Accessed: November 10, 2023. 6 +[12] Hao Fang, Saurabh Gupta, Forrest Iandola, Rupesh K Srivastava, Li Deng, Piotr Dollar, Jianfeng Gao, Xiaodong He, Margaret Mitchell, John C Platt, et al. From captions to visual concepts and back. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015. 3 + +[13] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 6 +[14] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 1, 3 +[15] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 6, 7 +[16] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations, 2022. 3 +[17] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, Qiang Liu, Kriti Aggarwal, Zewen Chi, Johan Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, and Furu Wei. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 1 +[18] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the International Conference on Machine Learning, pages 4904-4916, 2021. 3 +[19] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6, 7 +[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123:32-73, 2017. 8 +[21] Kibok Lee, Hao Yang, Satyaki Chakraborty, Zhaowei Cai, Gurumurthy Swaminathan, Avinash Ravichandran, and Onkar Dabeer. Rethinking few-shot object detection on a multi-domain benchmark. In Proceedings of the European Conference on Computer Vision, 2022. 6 +[22] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 1, 2, 3 +[23] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 6, 7 + +[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 1 +[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning, 2023. 1, 3, 6 +[26] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In Proceedings of the Conference on Empirical Methods in Natural Language, 2023. 1, 2, 3, 4 +[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, 2014. 2, 3, 4, 5 +[28] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 6, 7 +[29] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3, 6, 7, 8 +[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1, 3, 6, 7, 8 +[31] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.6,7 +[32] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 1, 2, 3, 5 +[33] Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V Le, Barret Zoph, Jason Wei, and Adam Roberts. The flan collection: Designing data and methods for effective instruction tuning. In Proceedings of the International Conference on Machine Learning, 2023. 5 +[34] Hwee Tou Ng, Siew Mei Wu, Ted Briscoe, Christian Hadiwinoto, Raymond Hendy Susanto, and Christopher Bryant. The CoNLL-2014 shared task on grammatical error correction. In Proceedings of the Eighteenth Conference on Computational Natural Language Learning: Shared Task, 2014. 5 +[35] OpenAI. Gpt-4 technical report, 2023. 2, 3 +[36] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 1, 3 +[37] Tina D Purnat, Paolo Vacca, Christine Czerniak, Sarah Ball, Stefano Burzo, Tim Zecchin, Amy Wright, Supriya + +Bezbaruah, Faizza Tanggol, Éve Dubé, Fabienne Labbe, Maude Dionne, Jaya Lamichhane, Avichal Mahajan, Sylvie Briand, and Tim Nguyen. Infodemic signal detection during the COVID-19 pandemic: Development of a methodology for identifying potential information voids in online conversations. *JMIR Infodemiology*, 1(1):e30971, 2021. 5 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 6 +[39] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 2020. 3, 5 +[40] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. In Proceedings of the Conference on Empirical Methods in Natural Language, pages 4035-4045, 2018. 1, 2, 3 +[41] Giuliano Rossi, Jakub Kolodziej, and Gurvinder Brar. A recommender system for active stock selection. Computational Management Science, 17, 2020. 5 +[42] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the International Conference on Computer Vision, 2019. 5, 6 +[43] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3 +[44] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. In Advances in Neural Information Processing Systems, pages 200–212, 2021. 1, 3 +[45] Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, Jitao Sang, and Haoyu Tang. Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126, 2023. 1 +[46] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 2022. 8 +[47] Hongbin Ye, Tong Liu, Aijia Zhang, Wei Hua, and Weiqiang Jia. Cognitive mirage: A review of hallucinations in large language models. arXiv preprint arXiv:2309.06794, 2023. 1, 3 +[48] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1, 3, 6, 7 + +[49] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 3 +[50] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 6 +[51] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the ai ocean: A survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 1, 3 +[52] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 1 +[53] Yiyang Zhou, Chenhang Cui, Jaehong Yoon, Linjun Zhang, Zhun Deng, Chelsea Finn, Mohit Bansal, and Huaxiu Yao. Analyzing and mitigating object hallucination in large vision-language models, 2023. 1, 8 +[54] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 6, 7 \ No newline at end of file diff --git a/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/images.zip b/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b1c8b87a408c2212510b19037f3068cd1bc8d17c --- /dev/null +++ b/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:612caa72fc54da8063d08e20620e7eb9a9b073da377061987224e016a0984af0 +size 459332 diff --git a/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/layout.json b/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b04fcfa4c6abb7fbecde8a8757ca5afc8a80d31a --- /dev/null +++ b/2024/THRONE_ An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models/layout.json @@ -0,0 +1,8332 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 106, + 103, + 485, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 103, + 485, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 485, + 140 + ], + "type": "text", + "content": "THRONE: An Object-based Hallucination Benchmark for the Free-form Generations of Large Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "spans": [ + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "text", + "content": "Prannay Kaul" + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "text", + "content": " Zhizhong Li" + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "text", + "content": " Hao Yang" + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "text", + "content": " Yonatan Dukler" + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "text", + "content": " Ashwin Swaminathan" + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "text", + "content": " C. J. Taylor" + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "text", + "content": " Stefano Soatto" + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "text", + "content": " VGG, University of Oxford" + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "text", + "content": " AWS AI Labs" + }, + { + "bbox": [ + 145, + 146, + 448, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 191, + 525, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 525, + 204 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 525, + 204 + ], + "type": "text", + "content": "prannay@robots.ox.ac.uk {lzhizhon,haoyng,dukler,swashwin,taylorcj,soattos}@amazon.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 256, + 289, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 289, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 289, + 544 + ], + "type": "text", + "content": "Mitigating hallucinations in large vision-language models (LVLMs) remains an open problem. Recent benchmarks do not address hallucinations in open-ended free-form responses, which we term \"Type I hallucinations\". Instead, they focus on hallucinations responding to very specific question formats—typically a multiple-choice response regarding a particular object or attribute—which we term \"Type II hallucinations\". Additionally, such benchmarks often require external API calls to models which are subject to change. In practice, we observe that a reduction in Type II hallucinations does not lead to a reduction in Type I hallucinations but rather that the two forms of hallucinations are often anti-correlated. To address this, we propose THRONE, a novel object-based automatic framework for quantitatively evaluating Type I hallucinations in LVLM free-form outputs. We use public language models (LMs) to identify hallucinations in LVLM responses and compute informative metrics. By evaluating a large selection of recent LVLMs using public datasets, we show that an improvement in existing metrics do not lead to a reduction in Type I hallucinations, and that established benchmarks for measuring Type I hallucinations are incomplete. Finally, we provide a simple and effective data augmentation method to reduce Type I and Type II hallucinations as a strong baseline." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 557, + 134, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 557, + 134, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 557, + 134, + 569 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 575, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 287, + 624 + ], + "type": "text", + "content": "This paper proposes a benchmark to evaluate hallucinations by large vision-language models (LVLMs) when generating free-form responses, specifically detailed descriptions, based on a given image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 624, + 288, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 288, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 288, + 684 + ], + "type": "text", + "content": "The rapid advancement in large language models (LLMs) [52] has pushed the development of large vision-language models (LVLMs). [1, 6, 7, 17, 24, 25, 30, 36, 44, 48, 54] LVLMs take input text and images and generate text responses to enable multi-modal perception and com" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 232, + 354, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 354, + 243 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 354, + 243 + ], + "type": "text", + "content": "prehension." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 244, + 547, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 547, + 376 + ], + "type": "text", + "content": "LVLMs are largely built on LLMs and therefore inherit both their advantages and their disadvantages. LLMs have been shown to produce hallucinations [47, 51], generated text responses that are coherent and plausible but factually incorrect. LVLMs echo this behavior with generated text contradicting with the visual or text input [53]. Hallucinations prevent the use of LVLMs in safety-critical situations and therefore evaluating and mitigating hallucinations in LVLMs is crucial for their deployment in such settings. Determining the presence and cause of hallucinations in LVLMs remains an open question [45, 53]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 377, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 547, + 545 + ], + "type": "text", + "content": "We divide LVLM hallucinations into two types. Type I hallucinations occur in response to open-ended questions with a very large set of possible responses — e.g. What is happening in this image?. Type II hallucinations are incorrect responses to a factual question regarding a specific concept about the image with a fixed set of options such as yes/no—e.g. Is there a traffic light in this image? Fig. 3 illustrates the difference between these two types of hallucination. Reducing hallucinations in both cases is required for useful, multi-purpose LVLMs. However, later in Fig. 4 we observe that the same LVLM can give contradicting answers when being evaluated for Type I vs. Type II hallucinations. This implies that measuring and reducing one type does not necessarily reduce the other." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "text", + "content": "Existing works to evaluate LVLMs often avoid direct quantification of hallucinations and instead develop comprehensive benchmarks that judge various other desirable abilities such as: optical character recognition, fine-grained recognition and attribute detection [14, 22, 32]. The extent of hallucinations in these benchmarks is obfuscated, since it is only one of many factors influencing other metrics. It requires human effort to inspect individual predictions. There are two major established works which specifically develop a benchmark for evaluating hallucinations in vision-language models: POPE [26] and CHAIR [40], which we discuss in detail in Sec. 2. However, we observe they both have shortcomings in effectively evaluating hallucinations:" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 220, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 220, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 220, + 703 + ], + "type": "text", + "content": "*Work conducted during an internship at Amazon" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27228" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 292, + 178 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 292, + 178 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 292, + 178 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 292, + 178 + ], + "type": "image", + "image_path": "086844dd337d4200da7236eb37d1d27179d6b14b2daac80a4009bc170907dc43.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 185, + 293, + 218 + ], + "lines": [ + { + "bbox": [ + 46, + 185, + 293, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 293, + 218 + ], + "type": "text", + "content": "Figure 1. THRONE (Ours): LVLMs are prompted with a concept neutral instruction. An external LM performs abstractive QA on the response to establish the existence of Type I hallucinations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 225, + 287, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 225, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 287, + 308 + ], + "type": "text", + "content": "POPE [26] is a recent work addressing Type II hallucinations with respect to object classes. However, we find Type I and Type II hallucinations are disconnected, and that POPE gives an incomplete picture on LVLM hallucinations. Moreover, POPE systematically under-samples negative object categories leading to a large underestimation of Type II hallucinations (see Sec. 5.4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 311, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 287, + 418 + ], + "type": "text", + "content": "CHAIR [40] does address Type I hallucinations—establishing object category hallucination in short image captions using simple text matching. However, CHAIR is not suited to current LVLMs because the simple text matching it employs cannot comprehend abstract or hypothetical concepts present in today's LVLMs (see Fig. 4). Further, hand-crafted rules for each set of classes are required for usable text matching, and trivial model answers can attain a perfect CHAIR score." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "content": "To address the issues of current LVLM hallucination benchmarks, we propose THRONE (Text-from-image Hallucination Recognition with Object-probes for open-ended Evaluation). THRONE leverages language models (LMs) to evaluate Type I hallucinations in free-form, open-ended image descriptions with respect to a pre-defined object vocabulary of interest. By utilizing LMs, THRONE is able to accurately judge whether an object mentioned in an LVLM response is implied to exist in the image or is abstractly mentioned with no implication about its existence (see Fig. 4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 555, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 650 + ], + "type": "text", + "content": "Moreover, in THRONE, we provide easy access to our benchmark, by leveraging open-source LMs that can run on common GPUs, instead of relying on closed-source commercial models [35] that are subject to arbitrary change, as done in other works [4, 22, 32]. Through combining multiple open-source LMs, we mitigate any single-model biases in judging hallucinations when calculating Type-I hallucination scores with THRONE." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "We make four contributions: first, we establish an accurate and accessible benchmark to quantitatively evaluate object hallucinations in free-form responses, leveraging LMs to judge the existence of Type I hallucinations—quantitatively showing half the judgement errors of CHAIR;" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 300, + 70, + 545, + 178 + ], + "blocks": [ + { + "bbox": [ + 300, + 70, + 545, + 178 + ], + "lines": [ + { + "bbox": [ + 300, + 70, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 300, + 70, + 545, + 178 + ], + "type": "image", + "image_path": "d0ffd45c22381ac0c708d6e5c516b7ac37149eda31e5a621be3c195e88db01d0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 185, + 545, + 218 + ], + "lines": [ + { + "bbox": [ + 299, + 185, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 299, + 185, + 545, + 218 + ], + "type": "text", + "content": "Figure 2. POPE: Questions with specific concepts prompt an LVLM directly to evaluate Type II hallucinations [26]. Handcrafted rules parse LVLM responses to give yes/no labels." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 225, + 547, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 225, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 225, + 547, + 370 + ], + "type": "text", + "content": "second, we evaluate a number of current LVLMs on THRONE and demonstrate that observed progress in reducing Type II hallucinations do not translate to a corresponding reduction in Type I hallucinations; third, we show a recent method, POPE, is inaccurately capturing the extent of Type II hallucination discovery due to its sampling strategy; we mitigate this issue in our implementation of THRONE while presenting results for a complete version of POPE; and fourth, we provide a simple augmentation of visual instruction tuning data which significantly improves performance on Type I hallucinations while maintaining or improving Type II hallucination performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 381, + 399, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 381, + 399, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 381, + 399, + 395 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "content": "Hallucination Benchmarks for LVLMs. In response to the development of LVLMs, few evaluation benchmarks focusing on hallucinations have been introduced. CHAIR [40], one of the first works to assess hallucinations, is designed for short image captions. CHAIR uses a fixed set of object classes (extended with their synonyms) as a set of text strings to find predicted object classes in image captions via exact text matching. Subsequently, each class matched in the caption is compared to the COCO [27] captions and bounding box annotations to establish object hallucinations. CHAIR was developed prior to current instruction-tuned LVLMs which produce long free-form responses (" + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "content": " longer than image captions) with a diverse vocabulary, limiting its applicability to modern LVLMs. As shown in Fig. 3, exact text matching in CHAIR is prone to incorrectly matching the vocabulary classes with abstract concepts in the free-form response and the and synonyms of the class list must be manually selected to prevent confusion during evaluation. For example, in CHAIR the word \"chair\" will match all responses for the phrase \"toilet seat\", because the exact text matching classifies \"seat\" as the COCO class \"chair\". Finally, CHAIR metrics compare the overall number of predicted objects to the overall number of predicted objects judged to be hallucinations—ignoring the recall of ground-truth objects and the distribution of object classes. This means that a single correct" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27229" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": "prediction across the entire evaluation dataset along with a generic response otherwise e.g. A natural scene, achieves a perfect score " + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "inline_equation", + "content": "(\\frac{0 \\text{ hallucinated objects}}{1 \\text{ predicted objects}} = 0.0)" + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": ". See the Supplementary Material for a full overview of CHAIR. In contrast, THRONE uses pre-trained LMs that go beyond direct synonym matching, to automatically judge the existence of concepts and hallucinations in free-form responses. In addition, our method considers both recall and precision to yield a holistic benchmark and does not require any manual curation of synonyms. THRONE and CHAIR both evaluate Type I hallucinations—hallucinations in response to concept-neutral prompts e.g. Describe this image in detail." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 229, + 289, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 289, + 457 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 289, + 457 + ], + "type": "text", + "content": "POPE [26] is a recently proposed benchmark to evaluate object hallucinations in LVLMs—specifically Type II hallucinations, in which an LVLM is directly queried with a yes-no question regarding the existence of a particular object of the form: Is there " + }, + { + "bbox": [ + 46, + 229, + 289, + 457 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{a / an}\\}" + }, + { + "bbox": [ + 46, + 229, + 289, + 457 + ], + "type": "text", + "content": " {object_class_name} in the image?. The LVLM response is parsed using simple rules to determine whether a Type II hallucination has occurred. Precision and recall metrics are compiled using the parsed LVLM responses and the ground-truth annotation data. Despite focusing on measuring hallucinations, POPE only queries an LVLM with 3 positive and 3 sampled negative questions per COCO image i.e. the evaluation is artificially balanced. This means many potential hallucinations with respect to the COCO categories are not captured by their method. In Sec. 5.4, we show POPE dramatically underestimates Type II hallucinations and present the results of a complete version. Our method, THRONE, evaluates the prevalence of Type I hallucinations, which we observe are disjoint to Type II hallucinations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 460, + 289, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 289, + 616 + ], + "type": "text", + "content": "Comprehensive Benchmarks for LVLMs have recently grown in number. MMBench [32] and MM-Vet [49] assess various aspects of LVLM performance such as: color perception, celebrity recognition, and numerical calculation. However many of these works integrate evolving APIs which are modified often (or even discontinued) and are inherently stochastic. Over time this greatly reduces the consistency of these benchmarks. Exceptions are MME [14] and SEED-Bench [22], but the impact of Type II hallucinations on final metrics is conflated with a number of other aspects of model performance. Our method, THRONE, directly addresses Type I hallucinations, only making use of open-source language models and datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "Large Vision Language Models (LVLMs) have rapidly developed by harnessing advancements in large language models (LLM) [5, 39, 43] and by directly integrating pretrained LLMs into their architectures. In contrast to earlier vision-language models such as CLIP [18, 38], LVLMs are generally comprised of a pretrained LLM and image encoder, aligned with a connector module of varying complexity. Some works are highlighted here. Frozen [44] is" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 547, + 191 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 547, + 191 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 547, + 191 + ], + "type": "table", + "html": "
ImagePromptResponse
What is happening in this image?In the image, a man is hanging clothes on a clothesline that is attached to the back of a moving car.
Is there a traffic light in this image?Yes, there is a traffic light in the image, and the man is standing on the back of a car near it.
", + "image_path": "dbc5b80114e6f8da66eecc2b07e061e2a54e4fdc1a63e8677464de6974aafa7b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 198, + 547, + 243 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 243 + ], + "type": "text", + "content": "Figure 3. Type I vs. Type II Hallucinations: (Top) LVLMs prompted with concept-neutral instructions produce Type I hallucinations. (Bottom) Instructions specifying a concept produce Type II hallucinations. Examples from LLaVA-v1.5 [29]." + } + ] + } + ], + "index": 5, + "type": "text" + }, + { + "bbox": [ + 304, + 251, + 547, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 547, + 469 + ], + "type": "text", + "content": "an early work fine-tuning the vision encoder to dynamically prefix the prompt to a frozen LLM. Flamingo [1] combines visual and language features using cross-attention layers in an otherwise frozen LLM. BLIP-2 [25] uses a frozen image encoder to learn a Querying-Transformer (Q-Former) on image-text pairs that is used as the connector. This architecture is adapted in [7] for dialogue via training with visual instruction tuning. LLaVA [30] uses COCO [27] annotations and GPT-4 [35] to generate visual instruction tuning data in a plain-text pipeline. Combining this generated data with standard VQA datasets (VQAv2 etc.) further boosts performance [29]. Different works modify the training approach by using efficient adaptation [16, 48] multiple training stages [3] or introduce the use of discrete tokens for localization [36]. We note however, that most of these works evaluate performance on traditional vision-language datasets like VQAv2 [2], which do not consider the extent of hallucinations, a known problem with LLMs [47, 51]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 477, + 383, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 477, + 383, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 477, + 383, + 490 + ], + "type": "text", + "content": "3 THRONE" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 497, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 652 + ], + "type": "text", + "content": "Recap of existing methods. POPE [26] and other benchmarks (MME [14], MMBench [32]) directly query LVLMs with a restricted desired answer space, e.g. yes-no (MME, POPE) and A-B-C-D multiple choice (MMBench), as shown in Fig. 2. These benchmarks only consider such short answer formats, whereas a key quality of LVLMs is in their ability to generate free-form coherent text. Moreover, POPE, which addresses Type II hallucinations, undersamples negative classes meaning hallucinations are dramatically underestimated (see Sec. 5.4 and Fig. 7). In contrast, we skip class subsampling and enumerate all classes for every image, ensuring a full evaluation of Type I hallucinations of the ground-truth classes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "CHAIR [40] also evaluates Type I hallucinations, but was developed when typical vision-language models could only generate short and simple captions similar to those in COCO Captions [12]. Moreover, it lacks accurate comprehension of responses (see the right side of Fig. 4) and ig-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27230" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 103, + 70, + 233, + 164 + ], + "blocks": [ + { + "bbox": [ + 103, + 70, + 233, + 164 + ], + "lines": [ + { + "bbox": [ + 103, + 70, + 233, + 164 + ], + "spans": [ + { + "bbox": [ + 103, + 70, + 233, + 164 + ], + "type": "image", + "image_path": "634cddb0c8f4c8f79f8782ffb0b14a0cc478117e417498c07dc7450b376f3331.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 258, + 545, + 326 + ], + "lines": [ + { + "bbox": [ + 45, + 258, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 45, + 258, + 545, + 326 + ], + "type": "text", + "content": "Figure 4. A Comparison of POPE, CHAIR and THRONE: Directly querying LVLMs for object existence (person, banana etc.) using concept-specific instructions, as in POPE (bottom left), does not produce the same hallucinations as using concept-neutral instructions (right). We highlight the Type I hallucinations in orange. CHAIR relies on exact text matching to a fixed set of objects and synonyms, thus incorrectly labels \"customers\" and \"shoppers\" as hallucinations, highlighted in red. THRONE is designed for the rich vocabulary and the free-form generations of modern LVLMs by harnessing LMs to establish object existence. By using an LM to pass judgement, our evaluation correctly captures \"customers\" and \"shoppers\" as hypothetical content in the free-form generation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 174, + 287, + 249 + ], + "blocks": [ + { + "bbox": [ + 102, + 165, + 234, + 173 + ], + "lines": [ + { + "bbox": [ + 102, + 165, + 234, + 173 + ], + "spans": [ + { + "bbox": [ + 102, + 165, + 234, + 173 + ], + "type": "text", + "content": "Type II Hallucination Evaluation (POPE)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 174, + 287, + 249 + ], + "lines": [ + { + "bbox": [ + 49, + 174, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 49, + 174, + 287, + 249 + ], + "type": "table", + "html": "
Instruction [GT]Response [Prediction]
Is there... \na person [No] \na banana [Yes] \nan apple [No] \nan orange [No] \n...in the image?No, there is no person in the image. [No] \nYes, there is a banana in the image. [Yes] \nNo, there is not an apple in the image. [No] \nNo, there is no orange in the image. [No] \nNo Type II Hallucinations Found
", + "image_path": "67c6795603d6391106d156263b0b685600a7eadf06269ce452f093f6b4d82376.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 291, + 79, + 544, + 249 + ], + "blocks": [ + { + "bbox": [ + 363, + 71, + 470, + 79 + ], + "lines": [ + { + "bbox": [ + 363, + 71, + 470, + 79 + ], + "spans": [ + { + "bbox": [ + 363, + 71, + 470, + 79 + ], + "type": "text", + "content": "Type I Hallucination Evaluation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 291, + 79, + 544, + 249 + ], + "lines": [ + { + "bbox": [ + 291, + 79, + 544, + 249 + ], + "spans": [ + { + "bbox": [ + 291, + 79, + 544, + 249 + ], + "type": "table", + "html": "
InstructionResponse
Describe this image in detail.The image showcases a fruit stand at a grocery store, featuring a variety of fruits on display. There are several bunches of bananas, with some placed in the foreground and others in the background. The bananas are arranged in different sections, creating an appealing presentation for customers. In addition to the bananas, there are also apples and oranges on display. The apples are located towards the left side of the image, while the oranges are placed in the middle and right side of the stand. The fruits are well-organized and presented in an attractive manner, making it an inviting sight for shoppers. Type I Hallucinations Present and Found
MSCOCO Object Prediction from Description
HumanCHAIRTHRONE (Ours)
banana apple orangebanana person apple orangebanana apple orange
", + "image_path": "fc94d9cdcb8fd795548283e02c2c5aef72cbd5016af2e605647e3486326c3133.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 331, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 415 + ], + "type": "text", + "content": "nores the recall of ground-truth objects. In Sec. 5.5, we describe quantitative evaluations, using a human oracle, which demonstrate THRONE halves the rate of hallucination misjudgement in CHAIR. See Sec. 2 and the Supplementary Material for more details on CHAIR and its shortcomings. Fig. 4 shows an overview of the three aforementioned methods: POPE, CHAIR and our method, THRONE." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 422, + 266, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 266, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 266, + 434 + ], + "type": "text", + "content": "3.1 Evaluating Hallucinations with THRONE" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 437, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 485 + ], + "type": "text", + "content": "To address these limitations, we propose a framework, THRONE, shown in Fig. 1, to evaluate the prevalence of Type I hallucinations in LVLM responses conditioned on an image and a neutral text prompt." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "text", + "content": "For each image in a labeled dataset, " + }, + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "text", + "content": ", addressing a set of classes, " + }, + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "text", + "content": ", the LVLM is queried with the same instruction: Describe this image in detail., regardless of image content. The LVLM response, which is expected to be long free-form text containing an image description, is generated and stored. Next, a publicly available, open-source, external language model (LM) performs abstractive question answering (AQA) using the LVLM response as context and a question of the form: Please answer yes or no. Is there " + }, + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{a / an}\\}" + }, + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "text", + "content": " {object class name} in this image? or similar, for every class in " + }, + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "text", + "content": " (right side of Fig. 1). By selecting an appropriate LM and using a simple prompt template (see Sec. 4 for specific details), we ensure the AQA response is either yes or no—our method does not require any additional parsing. This is in contrast to other works which require added parsing or interpretation by a closed-source model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "After performing AQA on each response generated by the LVLM for every class in " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": ", we obtain an array of pre" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 332, + 361, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 332, + 361, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 332, + 361, + 342 + ], + "type": "text", + "content": "dicted labels:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 389, + 342, + 545, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 342, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 389, + 342, + 545, + 358 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Y}} \\in \\{0, 1 \\} ^ {| \\mathcal {I} | \\times | \\mathcal {C} |} \\tag {1}", + "image_path": "c1affb4cef06a86702badba0a205baf973dcf0218624966c65f78bb12537928f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 365, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 365, + 545, + 389 + ], + "type": "inline_equation", + "content": "0/1" + }, + { + "bbox": [ + 304, + 365, + 545, + 389 + ], + "type": "text", + "content": " indicates a negative/positive existence judgement by the LM with respect to the relevant LVLM response." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 389, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 389, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 305, + 389, + 545, + 412 + ], + "type": "text", + "content": "Similarly using the ground-truth data for " + }, + { + "bbox": [ + 305, + 389, + 545, + 412 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 305, + 389, + 545, + 412 + ], + "type": "text", + "content": ", an array of ground-truth labels can be constructed:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 389, + 414, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 414, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 389, + 414, + 545, + 430 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} \\in \\{0, 1 \\} ^ {| \\mathcal {I} | \\times | \\mathcal {C} |} \\tag {2}", + "image_path": "98a2414706a68d2175957f829fb787add782acbbe7ee55669fa61c7cb0fc70a7.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 436, + 545, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 543 + ], + "type": "text", + "content": "Using these two arrays, we calculate four metrics: (1) Overall Precision, " + }, + { + "bbox": [ + 304, + 436, + 545, + 543 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{ALL}}" + }, + { + "bbox": [ + 304, + 436, + 545, + 543 + ], + "type": "text", + "content": "; (2) Overall Recall, " + }, + { + "bbox": [ + 304, + 436, + 545, + 543 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{ALL}}" + }, + { + "bbox": [ + 304, + 436, + 545, + 543 + ], + "type": "text", + "content": "; (3) Class-wise Precision, " + }, + { + "bbox": [ + 304, + 436, + 545, + 543 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{CLS}}" + }, + { + "bbox": [ + 304, + 436, + 545, + 543 + ], + "type": "text", + "content": "; and (4) Class-wise Recall, " + }, + { + "bbox": [ + 304, + 436, + 545, + 543 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{CLS}}" + }, + { + "bbox": [ + 304, + 436, + 545, + 543 + ], + "type": "text", + "content": ". Overall metrics are calculated in a class-agnostic manner. Class-wise metrics are calculated in a class-conscious manner by computing precision and recall for each category separately and then averaging. This follows common practice in object detection and instance segmentation [10, 27]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "text", + "content": "False positives in LVLMs reduce precision and are dominated by hallucinations—precision indicates the extent of Type I hallucinations in LVLM responses. The recall metrics inform the level of class coverage by an LVLM when producing image descriptions. The class-wise metrics give a general measure of performance as the overall metrics are skewed towards the most common categories. A common way to combine precision, " + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "text", + "content": ", and recall, " + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "text", + "content": ", metrics is through the generalized F score, " + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 304, + 544, + 545, + 651 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 361, + 658, + 545, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 658, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 361, + 658, + 545, + 684 + ], + "type": "interline_equation", + "content": "F ^ {\\beta} = \\left(1 + \\beta^ {2}\\right) \\cdot \\frac {P \\cdot R}{\\left(\\beta^ {2} \\cdot P\\right) + R} \\tag {3}", + "image_path": "66fc9d845115f0a34c52f302929aec3f98bb46b7f38e393d26c99a4d909677ed.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Prior work such as POPE [26], use the common balanced F-score (or " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "F^1" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "-score) which equally weights precision and" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27231" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "recall. However, given that THRONE is concerned with measuring hallucinations and is therefore particularly interested in precision, we choose " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\beta = 0.5" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " or the " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "F^{0.5}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "-score, thus weighing precision twice as important as recall. The " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "F^{0.5}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "-score is commonly used in pandemic misinformation filters [37], recommender systems [34], active stock selection [41] and other areas where false positives are costlier than false negatives. From this we can calculate the overall and class-wise " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "F^{0.5}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "-scores, " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{ALL}}^{0.5}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{CLS}}^{0.5}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ", respectively. To mitigate against class imbalance issues, we use " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{CLS}}^{0.5}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " as the principle metric of comparison between LVLMs in THRONE." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 230, + 250, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 230, + 250, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 230, + 250, + 243 + ], + "type": "text", + "content": "3.2 Ensuring Robustness via Ensembling" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 248, + 289, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 289, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 289, + 514 + ], + "type": "text", + "content": "Any LM used for AQA in THRONE may misjudge Type II hallucinations—no LM is perfect. Fig. 5 (top) shows two different variants from the same model family (FLANT5 [33]), yielding opposite responses when prompted with the same response and question. Moreover, an LM may yield different answers to semantically identical questions, despite conditioning on the same response—shown in Fig. 5 (bottom). To ensure THRONE is robust to spurious performance by any one LM, we ensemble various LMs and semantically equivalent question formats. The use of " + }, + { + "bbox": [ + 46, + 248, + 289, + 514 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 248, + 289, + 514 + ], + "type": "text", + "content": " distinct LMs and " + }, + { + "bbox": [ + 46, + 248, + 289, + 514 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 248, + 289, + 514 + ], + "type": "text", + "content": " distinct question formats yields " + }, + { + "bbox": [ + 46, + 248, + 289, + 514 + ], + "type": "inline_equation", + "content": "NM" + }, + { + "bbox": [ + 46, + 248, + 289, + 514 + ], + "type": "text", + "content": " answers for each (LVLM response, class) combination. This set of " + }, + { + "bbox": [ + 46, + 248, + 289, + 514 + ], + "type": "inline_equation", + "content": "NM" + }, + { + "bbox": [ + 46, + 248, + 289, + 514 + ], + "type": "text", + "content": " answers is combined based on equal voting by each (LM, question) pair to \"elect\" the predicted answer. Continuing the notation from Eq. (1), stacking predictions from each (LM, question) pair yields a 3D array of predicted labels: " + }, + { + "bbox": [ + 46, + 248, + 289, + 514 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{Y}} \\in \\{0,1\\}^{|\\mathcal{I}_{\\mathrm{OURS}}| \\times |\\mathcal{C}_{\\mathrm{OURS}}| \\times NM}" + }, + { + "bbox": [ + 46, + 248, + 289, + 514 + ], + "type": "text", + "content": ". To combine the answers from each (LM, question) pair via voting, we require agreement between at least " + }, + { + "bbox": [ + 46, + 248, + 289, + 514 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 248, + 289, + 514 + ], + "type": "text", + "content": " answers. Where sufficient agreement does not exist we introduce an \"ignore\" label. Mathematically, the elements in " + }, + { + "bbox": [ + 46, + 248, + 289, + 514 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 46, + 248, + 289, + 514 + ], + "type": "text", + "content": " (from Eq. (1)) are calculated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 522, + 254, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 522, + 254, + 569 + ], + "spans": [ + { + "bbox": [ + 79, + 522, + 254, + 569 + ], + "type": "interline_equation", + "content": "\\hat {Y} _ {i, j} = \\left\\{ \\begin{array}{l l} 0, & \\sum_ {k = 1} ^ {N M} \\bar {Y} _ {i, j, k} \\leq (N M - k) \\\\ 1, & \\sum_ {k = 1} ^ {N M} \\bar {Y} _ {i, j, k} \\geq k \\\\ - 1, & \\text {o t h e r w i s e} \\end{array} \\right.", + "image_path": "fae83dde654ec8af03a5ff27cc09acc21c106dab5ad870e567fe0398ce69f35c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 581, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 289, + 713 + ], + "type": "text", + "content": "where an \"ignore\" label exists in " + }, + { + "bbox": [ + 46, + 581, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{Y}}" + }, + { + "bbox": [ + 46, + 581, + 289, + 713 + ], + "type": "text", + "content": ", it is removed from the calculations of " + }, + { + "bbox": [ + 46, + 581, + 289, + 713 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{ALL}}" + }, + { + "bbox": [ + 46, + 581, + 289, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 581, + 289, + 713 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{ALL}}" + }, + { + "bbox": [ + 46, + 581, + 289, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 581, + 289, + 713 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{CLS}}" + }, + { + "bbox": [ + 46, + 581, + 289, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 581, + 289, + 713 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{CLS}}" + }, + { + "bbox": [ + 46, + 581, + 289, + 713 + ], + "type": "text", + "content": ", as we cannot be confident in the AQA process for that particular (LVLM response, class) combination. The choice of " + }, + { + "bbox": [ + 46, + 581, + 289, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 581, + 289, + 713 + ], + "type": "text", + "content": " reflects the desired level of confidence. We make use of a unanimous voting mechanism (" + }, + { + "bbox": [ + 46, + 581, + 289, + 713 + ], + "type": "inline_equation", + "content": "k = NM" + }, + { + "bbox": [ + 46, + 581, + 289, + 713 + ], + "type": "text", + "content": ")—use the prediction only if all (LM, question) pairs agree, otherwise ignore. Human evaluation of our benchmark and choice of voting mechanism is found in the Supplementary Material. Once we have applied this voting mechanism, we can calculate the metrics described at the end of Sec. 3.1." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 539, + 170 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 539, + 170 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 539, + 170 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 539, + 170 + ], + "type": "image", + "image_path": "9839b4821e70a308c64e94e2823eca3305e1faa403fb063339a77b55ea633175.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 173, + 547, + 229 + ], + "lines": [ + { + "bbox": [ + 304, + 173, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 547, + 229 + ], + "type": "text", + "content": "Figure 5. AQA Ensembling in Evaluation: Using different LMs or different prompts when running AQA on LVLM generated responses can produce opposing answers to identical prompts or identical LMs. To ensure THRONE is robust to this, we ensemble multiple LMs and multiple prompts in our evaluation pipeline." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 234, + 410, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 234, + 410, + 248 + ], + "spans": [ + { + "bbox": [ + 305, + 234, + 410, + 248 + ], + "type": "text", + "content": "4 Implementation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 252, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 545, + 277 + ], + "type": "text", + "content": "We provide details of our framework regarding the selection of public datasets, public LMs and LM prompts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 281, + 370, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 370, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 370, + 293 + ], + "type": "text", + "content": "4.1 Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 297, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 441 + ], + "type": "text", + "content": "Any proper evaluation of object hallucinations (a type of false positive error) requires knowing, with certainty, which classes are absent in an image. In our benchmark, we use COCO [27] for a number of reasons: (1) its annotations of 80 categories are exhaustive at an image level (image-level recall " + }, + { + "bbox": [ + 304, + 297, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\approx 99\\%" + }, + { + "bbox": [ + 304, + 297, + 545, + 441 + ], + "type": "text", + "content": " [27])—if there are many book instances in a COCO image, at least one is annotated with bounding boxes; (2) many LVLMs are partly trained on COCO data and so should be familiar with the set of categories; (3) its images generally contain complex scenes suitable for generating long free-form descriptions unlike image recognition datasets like ImageNet [8]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 441, + 545, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 549 + ], + "type": "text", + "content": "We utilize the validation set of COCO 2017, which contains " + }, + { + "bbox": [ + 304, + 441, + 545, + 549 + ], + "type": "inline_equation", + "content": "|\\mathcal{I}| = 5000" + }, + { + "bbox": [ + 304, + 441, + 545, + 549 + ], + "type": "text", + "content": " images and " + }, + { + "bbox": [ + 304, + 441, + 545, + 549 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}| = 80" + }, + { + "bbox": [ + 304, + 441, + 545, + 549 + ], + "type": "text", + "content": " categories. Using the single LVLM text prompt Describe this image in detail., we generate 5000 responses. As we query each LVLM response for each category in " + }, + { + "bbox": [ + 304, + 441, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 441, + 545, + 549 + ], + "type": "text", + "content": ", a single LM performs AQA " + }, + { + "bbox": [ + 304, + 441, + 545, + 549 + ], + "type": "inline_equation", + "content": "|\\mathcal{I}| \\times |\\mathcal{C}| = 400k" + }, + { + "bbox": [ + 304, + 441, + 545, + 549 + ], + "type": "text", + "content": " times across the LVLM responses—one instance of AQA per (image response, class) pair. In Sec. 5.3, we also present results using Objects365 [42] which is rarely used in LVLM training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 554, + 418, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 554, + 418, + 567 + ], + "spans": [ + { + "bbox": [ + 305, + 554, + 418, + 567 + ], + "type": "text", + "content": "4.2 Language Models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "To assess Type I hallucinations in an LVLM response using THRONE, we require a language model (LM) which can answer questions on the existence of object categories based on the LVLM response. MMBench [32] makes use of a ChatGPT model to identify multiple choice answer selections and still reports mistakes. In our experience, some LMs give rather incoherent judgements when used to assess hallucinations when the prompt is changed (see supplemental material). Therefore for THRONE, we choose FLAN-T5 models [33, 39]. We make this choice because FLAN-T5 model family: (1) have undergone instruction tuning with thousands of tasks [33]; (2) are open-source and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27232" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 287, + 163 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 287, + 163 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 287, + 163 + ], + "type": "table", + "html": "
ModelL\\( P_{ALL} \\)\\( R_{ALL} \\)\\( F_{ALL}^{1} \\)\\( F_{ALL}^{0.5} \\)\\( P_{CLS} \\)\\( R_{CLS} \\)\\( F_{CLS}^{1} \\)\\( F_{CLS}^{0.5} \\)
Adapter-v2 [15]51463.673.368.165.368.270.669.468.7
Adapter-v2.1 [15]51263.873.768.465.567.471.269.368.1
InstructBLIP [7]52570.874.372.571.577.271.974.576.1
Otter-Image [23]25733.031.232.132.725.216.920.222.9
MiniGPT4 [54]47381.759.869.076.179.961.869.775.5
MiniGPT-v2 [6]38179.066.672.376.277.667.071.975.2
mPLUG-Owl [48]55555.571.962.658.166.368.367.366.7
LRV-Instruction-v2 [28]10382.056.767.075.378.458.867.273.5
LLaVA-v1.3* [30]53280.565.272.176.979.965.371.976.5
LLaVA-v1.5 [29]50968.161.064.466.669.956.462.566.8
LLaVA-Mistral [19, 31]52486.871.878.383.684.464.270.877.5
", + "image_path": "70dd6100dc7ef3fcbf7658d68e8ef1293b48be008f8020e0991b96e9512a8cc8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 171, + 289, + 259 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 289, + 259 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 289, + 259 + ], + "type": "text", + "content": "Table 1. THRONE Results with COCO for a selection of instruction-tuned LVLMs. We select " + }, + { + "bbox": [ + 46, + 171, + 289, + 259 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{CLS}}^{0,5}" + }, + { + "bbox": [ + 46, + 171, + 289, + 259 + ], + "type": "text", + "content": " as the principal metric for evaluation in our benchmark to balance across classes and to prioritize precision (which reflects the extent of hallucination) over recall. Best and second-best performance are denoted by blue and red, respectively. *Our implementation using official code to enable fair comparison. " + }, + { + "bbox": [ + 46, + 171, + 289, + 259 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 171, + 289, + 259 + ], + "type": "text", + "content": " corresponds to the median response length (measured in # of characters)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 277, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 386 + ], + "type": "text", + "content": "therefore accessible to the community; (3) can fit locally on a single GPU for the models we consider; (4) follow user's instruction to only respond yes or no; and (5) are optimized for use in the free and public Text-Generation-Inference API [11] for acceleration. As described in Sec. 3.2, we utilize " + }, + { + "bbox": [ + 46, + 277, + 287, + 386 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 277, + 287, + 386 + ], + "type": "text", + "content": " LMs to ensure our method is robust. Specifically, we use " + }, + { + "bbox": [ + 46, + 277, + 287, + 386 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 46, + 277, + 287, + 386 + ], + "type": "text", + "content": " variants of FLAN-T5, namely: FLAN-T5-Base (250M parameters), FLAN-T5-Large (780M parameters), and FLAN-T5-XL (3B parameters)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 394, + 170, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 170, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 170, + 407 + ], + "type": "text", + "content": "4.3 Prompt Ensembling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 410, + 287, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 287, + 457 + ], + "type": "text", + "content": "To guarantee each of these FLAN-T5 variants faithfully produce responses of either yes or no only during AQA, we use the following input template to each LM, reflecting the format used when training FLAN-T51:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 462, + 255, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 462, + 255, + 497 + ], + "spans": [ + { + "bbox": [ + 66, + 462, + 255, + 497 + ], + "type": "text", + "content": "Text: {LVLM Response} Read the text about an image and answer the question. \nQuestion: Please answer yes or no. \n{Question}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 502, + 255, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 502, + 255, + 513 + ], + "spans": [ + { + "bbox": [ + 58, + 502, + 255, + 513 + ], + "type": "text", + "content": "We use " + }, + { + "bbox": [ + 58, + 502, + 255, + 513 + ], + "type": "inline_equation", + "content": "M = 3" + }, + { + "bbox": [ + 58, + 502, + 255, + 513 + ], + "type": "text", + "content": " semantically identical questions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 514, + 280, + 553 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 49, + 514, + 253, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 514, + 253, + 521 + ], + "spans": [ + { + "bbox": [ + 49, + 514, + 253, + 521 + ], + "type": "text", + "content": "- Is there a/an {class_name} in this image?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 521, + 280, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 521, + 280, + 537 + ], + "spans": [ + { + "bbox": [ + 49, + 521, + 280, + 537 + ], + "type": "text", + "content": "- Does the text imply a/an {class_name} is in the image?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 537, + 235, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 537, + 235, + 553 + ], + "spans": [ + { + "bbox": [ + 49, + 537, + 235, + 553 + ], + "type": "text", + "content": "- Does the text explicitly mention a/an {class_name} is in the image?" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "text", + "content": "As outlined in Sec. 3.2, we use a unanimous voting mechanism to combine the answers from each (LM, question) pair and so " + }, + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "inline_equation", + "content": "k = 3 \\times 3 = 9" + }, + { + "bbox": [ + 46, + 555, + 287, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 594, + 165, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 165, + 607 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 165, + 607 + ], + "type": "text", + "content": "5 Evaluation Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 613, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 287, + 696 + ], + "type": "text", + "content": "In this section, we: (1) outline our LVLM selection and reasoning; (2) present THRONE on COCO for evaluating Type I hallucinations; (3) extend THRONE to Objects365 (containing a larger vocabulary); (4) analyze and extend POPE to enable improved evaluation of Type II hallucinations; and (5) highlight results from our ablation studies found in the Supplemental Material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 369, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 369, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 369, + 83 + ], + "type": "text", + "content": "5.1 Models" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 87, + 547, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 87, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 87, + 547, + 266 + ], + "type": "text", + "content": "For fair comparison between existing LVLMs, each publicly available model we evaluate uses an LLM with " + }, + { + "bbox": [ + 304, + 87, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\sim 7\\mathrm{B}" + }, + { + "bbox": [ + 304, + 87, + 547, + 266 + ], + "type": "text", + "content": " parameters. The LVLMs generally have different sized image encoders—e.g. LLaVA [30] uses a CLIP ViT-L/14 [9, 38] with an input resolution of " + }, + { + "bbox": [ + 304, + 87, + 547, + 266 + ], + "type": "inline_equation", + "content": "336\\times 336" + }, + { + "bbox": [ + 304, + 87, + 547, + 266 + ], + "type": "text", + "content": ", while InstructBLIP [7] uses a ViT-g/14 [50] trained with EVA [13] and an input resolution of " + }, + { + "bbox": [ + 304, + 87, + 547, + 266 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 304, + 87, + 547, + 266 + ], + "type": "text", + "content": ". Note that image encoder size and resolution is not something we can easily control in a pre-trained model. Each model we consider contains instruction tuning in the final training phase—instruction tuned models provide free-form descriptions; THRONE focuses on models that can generate free-form descriptions. E.g. we leave out BLIP-2 [25] (response median length 31 characters) in favor of InstructBLIP (median response length 525)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 271, + 464, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 271, + 464, + 283 + ], + "spans": [ + { + "bbox": [ + 305, + 271, + 464, + 283 + ], + "type": "text", + "content": "5.2 THRONE Results on COCO" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 286, + 546, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 546, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 546, + 585 + ], + "type": "text", + "content": "Results are shown in Tab. 1. The principal metric that we use to judge model performance is the classwise " + }, + { + "bbox": [ + 304, + 286, + 546, + 585 + ], + "type": "inline_equation", + "content": "F^{0.5}" + }, + { + "bbox": [ + 304, + 286, + 546, + 585 + ], + "type": "text", + "content": "-score (highlighted gray). We also report all the metrics outlined in Sec. 3.1, " + }, + { + "bbox": [ + 304, + 286, + 546, + 585 + ], + "type": "inline_equation", + "content": "(P, R, F^1, F^{0.5})" + }, + { + "bbox": [ + 304, + 286, + 546, + 585 + ], + "type": "text", + "content": " for overall (left) and class-wise averaging (right), utilizing the unanimous voting presented in Sec. 3.2. See the Supplementary Material for results and an analysis of different voting mechanisms. These results demonstrate that improvements on other benchmarks (POPE, MME, MMBench etc.) may be orthogonal and potentially at odds with improved performance on THRONE. Using the results of the 11 LVLMs that we evaluate, THRONE and POPE, which measure Type I and Type II hallucinations, respectively, have a Spearman's rank correlation coefficient of just 0.2, and THRONE vs POPE-C has just 0.4—the relationship between performance on POPE and THRONE on the same dataset is far from monotonic. For class-wise precision—" + }, + { + "bbox": [ + 304, + 286, + 546, + 585 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{CLS}}" + }, + { + "bbox": [ + 304, + 286, + 546, + 585 + ], + "type": "text", + "content": ", the best performing models hallucinate " + }, + { + "bbox": [ + 304, + 286, + 546, + 585 + ], + "type": "inline_equation", + "content": "\\sim 20\\%" + }, + { + "bbox": [ + 304, + 286, + 546, + 585 + ], + "type": "text", + "content": " of the objects. We show in the Supplementary Material that the vast majority of false positive objects in the free-form image descriptions evaluated are direct hallucinations rather than misclassifications of visually similar objects (e.g. mistaking a squash racket for a tennis racket). These results demonstrate that much work still remains to adequately suppress Type I hallucinations in LVLMs." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 590, + 484, + 603 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 590, + 484, + 603 + ], + "spans": [ + { + "bbox": [ + 305, + 590, + 484, + 603 + ], + "type": "text", + "content": "5.3 THRONE Results on Objects365" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Many LVLMs train on COCO directly or indirectly, thus to demonstrate generality we apply THRONE to the Objects365 dataset [42]. Like COCO, Objects365 aims to be exhaustive in its image-level class labeling (it aims to label at least one instance for each class present), but it has a larger object vocabulary and is not used as training data for the LVLMs that we evaluate. To gather a manageable subset of the Objects365 validation set (80k images), we use the natural sampling algorithm from [21], resulting in 5110 im" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 199, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 199, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 199, + 712 + ], + "type": "text", + "content": "1https://tinyurl.com/5n6nexze" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27233" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 71, + 286, + 180 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 286, + 180 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 286, + 180 + ], + "type": "table", + "html": "
ModelPALLRALLF1ALLF0.5ALLPCLSRCLSF1CLSF0.5CLS
Adapter-v2 [15]46.733.939.343.448.928.536.042.8
Adapter-v2.1 [15]46.834.039.443.548.828.836.242.8
InstructBLIP [7]54.537.244.249.853.733.641.348.0
Otter-Image [23]21.412.716.018.89.54.46.07.7
MiniGPT4 [54]53.032.940.647.249.931.939.044.9
MiniGPT-v2 [6]54.536.043.449.451.334.641.346.8
mPLUG-Owl [48]43.733.437.841.248.229.036.242.6
LRV-Instruction-v2 [28]57.526.736.546.751.426.635.143.3
LLaVA-v1.3* [30]57.632.941.950.152.630.538.645.9
LLaVA-v1.5 [29]54.039.545.650.353.934.341.948.4
LLaVA-Mistral [19, 31]58.339.146.953.157.835.944.351.5
", + "image_path": "84d946f9861ade220df8aec711d5c09edd046c825cd5d4d4c8f4f4290fa44c21.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 190, + 287, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 287, + 245 + ], + "type": "text", + "content": "Table 2. THRONE Evaluation with Objects365. Evaluation results for a selection of instruction-tuned LVLMs, we use a subset of Objects365 for the THRONE evaluation. Best and second-best performance are denoted by blue and red, respectively. *Our implementation using official code to enable fair comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 51, + 247, + 282, + 402 + ], + "blocks": [ + { + "bbox": [ + 51, + 247, + 282, + 402 + ], + "lines": [ + { + "bbox": [ + 51, + 247, + 282, + 402 + ], + "spans": [ + { + "bbox": [ + 51, + 247, + 282, + 402 + ], + "type": "image", + "image_path": "f8bccb59c82dd1aac12a91be6f40a11c47060dee4bf85013f0e2030a36ed26b4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 403, + 287, + 469 + ], + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 469 + ], + "type": "text", + "content": "Figure 6. Comparing THRONE on COCO and Objects365. We observe that despite variations in the data distribution, THRONE metrics, which measure Type I hallucinations generalize and have a strong Spearman's rank correlation of " + }, + { + "bbox": [ + 46, + 403, + 287, + 469 + ], + "type": "inline_equation", + "content": "r = 0.900" + }, + { + "bbox": [ + 46, + 403, + 287, + 469 + ], + "type": "text", + "content": ". One model (red circle) designated as an outlier and ignored when calculating ranking correlation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 476, + 286, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 476, + 286, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 286, + 571 + ], + "type": "text", + "content": "ages (the COCO validation set has 5k images). We present the results for THRONE on Objects365 in Tab. 2. Figure 6 shows the strong correlation in THRONE performance between evaluating on COCO and Objects365. This demonstrates that measuring Type I hallucinations in an LVLM using THRONE with a relatively small dataset like COCO, is indeed indicative of the intrinsic level of Type I hallucination in a given LVLM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 578, + 286, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 286, + 590 + ], + "type": "text", + "content": "5.4 Completing POPE for Type II Hallucinations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 45, + 594, + 286, + 713 + ], + "type": "text", + "content": "Our experiments have used THRONE to evaluate the prevalence of Type I hallucinations in LVLM responses on COCO. POPE evaluates Type II hallucinations on COCO, but we find POPE is largely incomplete. First, POPE only evaluates on 500 COCO validation set images. Second, for each image only a subset of classes (at most 12) are evaluated—each image is only queried with " + }, + { + "bbox": [ + 45, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 45, + 594, + 286, + 713 + ], + "type": "text", + "content": " of possible questions. Finally, POPE artificially balances evaluation questions between positives and negatives, despite object class existence in images being inherently imbalanced." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 73, + 541, + 150 + ], + "blocks": [ + { + "bbox": [ + 309, + 73, + 541, + 150 + ], + "lines": [ + { + "bbox": [ + 309, + 73, + 541, + 150 + ], + "spans": [ + { + "bbox": [ + 309, + 73, + 541, + 150 + ], + "type": "image", + "image_path": "ebc4c81c9640a9eb8b01b8e4c79436addb25706a877a5e3844e203e900861bc5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 158, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 304, + 158, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 545, + 224 + ], + "type": "text", + "content": "Figure 7. Instability of POPE to complete evaluation of Type II Hallucinations. Extending POPE to an exhaustive analysis on all COCO images and classes (POPE-C) leads to a dramatic reduction in performance across all 11 models. POPE is sensitive to the sampling mechanism, and by undersampling negative classes, POPE severely underestimates Type II hallucinations in LVLMs." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 233, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 545, + 399 + ], + "type": "text", + "content": "The above reasons make the evaluation of Type II hallucinations using POPE insufficient (see Sec. 2 for more details on POPE). To correct this, we complete POPE by using all images to exhaustively query each LVLM for every class in the COCO vocabulary, as done in THRONE. We name this POPE-Complete (POPE-C). Fig. 7 shows the extreme difference in evaluation between POPE and our exhaustive version, POPE-C—reporting " + }, + { + "bbox": [ + 304, + 233, + 545, + 399 + ], + "type": "inline_equation", + "content": "F^{1}" + }, + { + "bbox": [ + 304, + 233, + 545, + 399 + ], + "type": "text", + "content": "-score (POPE does not utilize " + }, + { + "bbox": [ + 304, + 233, + 545, + 399 + ], + "type": "inline_equation", + "content": "F^{0.5}" + }, + { + "bbox": [ + 304, + 233, + 545, + 399 + ], + "type": "text", + "content": "-score). As POPE only evaluates at most 9 negative classes, only a small subset of potential hallucinations of COCO classes are evaluated, thereby heavily underestimating the extent of Type II hallucination. For each LVLM analyzed, we observe a large—in many cases an extreme—reduction in precision and therefore in F1-score." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 401, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 521 + ], + "type": "text", + "content": "Our evaluation contains three pairs of LVLMs in which one is the follow-up work to the other, where each follow-up work generally trains on more data for more tasks with a more advanced language model. Comparing the right hand side of Fig. 7 to Tab. 1, we observe that these follow-up works generally show an improvement in POPE (and POPE-C), but surprisingly indicate a small reduction in performance on THRONE with COCO. This observation suggests that progress in reducing Type II hallucinations can be orthogonal to reducing Type I hallucinations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 528, + 379, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 528, + 379, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 379, + 540 + ], + "type": "text", + "content": "5.5 Ablations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 544, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 568 + ], + "type": "text", + "content": "In the Supplementary Material we present three key ablation experiments and give an executive summary of results here." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "text", + "content": "First, after subsampling COCO images and LVLM responses, we replace the LMs in THRONE with human judgement as an oracle for Type I hallucination occurrence. When comparing THRONE and CHAIR with human judgements, we estimate using THRONE improves the precision of judging Type I hallucinations to " + }, + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "text", + "content": " versus " + }, + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "inline_equation", + "content": "91\\%" + }, + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "text", + "content": " when using CHAIR—this reduces the false discovery rate by more than " + }, + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 569, + 545, + 700 + ], + "type": "text", + "content": ". Note that we find most estimated errors in THRONE arise from the particular class definitions in COCO, e.g. the COCO class tv includes computer monitors, which the oracle judgement is aware of." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Second, we apply the same class (and image) sampling" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27234" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "strategy as in POPE to THRONE and show this sampling overestimates " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{CLS}}^{0.5}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " by an average of 12.3 points compared to the complete use of classes in THRONE (Tab. 1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "type": "text", + "content": "Finally, we vary the choice of " + }, + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "type": "text", + "content": " i.e. the voting mechanism used to combine answers from multiple (LM, question) pairs. We use the unanimous voting mechanism (" + }, + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "type": "inline_equation", + "content": "k = 9" + }, + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "type": "text", + "content": ") in THRONE to minimize the false discovery rate and find the valid alternatives of simple majority (" + }, + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "type": "text", + "content": ") or all-but-one (" + }, + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "type": "inline_equation", + "content": "k = 8" + }, + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "type": "text", + "content": ") voting mechanisms have strong correlations and rank correlations across all compute metrics, in THRONE, of " + }, + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "type": "inline_equation", + "content": "> 0.99" + }, + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "type": "inline_equation", + "content": "> 0.94" + }, + { + "bbox": [ + 46, + 108, + 287, + 204 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 209, + 169, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 209, + 169, + 222 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 169, + 222 + ], + "type": "text", + "content": "6 Improved Baselines" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 224, + 287, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 287, + 296 + ], + "type": "text", + "content": "Much needs to be done to study Type I hallucinations. As a first step to their mitigation, we demonstrate a baseline method to augment the visual instruction tuning data for LLaVA models [29, 30], yielding improvements on THRONE while maintaining similar performance regarding Type II hallucinations on POPE." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 300, + 287, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 312 + ], + "type": "text", + "content": "6.1 Visual Instruction Tuning Data Augmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 315, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 287, + 434 + ], + "type": "text", + "content": "Similar to chain-of-thought learning [46], during instruction tuning, we augment all visual instruction tuning samples constructed by LLaVA [30] by preponding the task of enumerating a list of objects (present and absent) and indicating approximate locations, if applicable. Other than this, the LLaVA data and training pipeline remains unchanged. To generate the new data for this object enumeration task, we use the same COCO bounding box annotations used to generate the vision instruction tuning data in LLaVA. The simple text-only format (no special tokens) we use is:" + } + ] + } + ], + "index": 5 + }, + { + "type": "code", + "bbox": [ + 47, + 441, + 288, + 483 + ], + "blocks": [ + { + "bbox": [ + 47, + 441, + 288, + 483 + ], + "lines": [ + { + "bbox": [ + 47, + 441, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 288, + 483 + ], + "type": "text", + "content": "Instruction: Give a list of objects and locations in the image. \nResponse: {class_name_1} {{location_1}}/absent] {class_name_N} {{location_N}}/absent]" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "code_body" + } + ], + "index": 6, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 46, + 491, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 610 + ], + "type": "text", + "content": "where location_i is a plain text indicator representing the location of the center point of the relevant object in the image on a " + }, + { + "bbox": [ + 46, + 491, + 287, + 610 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 491, + 287, + 610 + ], + "type": "text", + "content": " grid (e.g. bottom left). To provide negatives in the training data, if class_name_i is not present, we use the plain text indicator absent. Prior work [53] shows that classes that frequently co-occur in the training data are the most common hallucinations, therefore we bias our negative sampling towards class pairs that frequently co-occur using a correlation matrix. We detail and ablate this choice in the Supplementary Material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 614, + 200, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 614, + 200, + 626 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 200, + 626 + ], + "type": "text", + "content": "6.2 Improved Baseline Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Tab. 3 shows the result of evaluating our improved baseline method on THRONE, POPE, and POPE-C. During inference, we approximate our training data augmentation by first prompting the LVLM to perform the object enumeration tasks and then generating a response to the prompt from the relevant benchmark. We additionally show results of utilizing VisualGenome bounding box annotations [20] on the" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 308, + 71, + 545, + 147 + ], + "blocks": [ + { + "bbox": [ + 308, + 71, + 545, + 147 + ], + "lines": [ + { + "bbox": [ + 308, + 71, + 545, + 147 + ], + "spans": [ + { + "bbox": [ + 308, + 71, + 545, + 147 + ], + "type": "table", + "html": "
ModelObject \nEnumeration DataTHRONEPOPEPOPE-C
PCLSRCLSF0.5CLSPRF1PRF1
LLaVA-v1.3X79.965.376.558.098.473.07.799.214.3
COCO83.268.879.973.288.280.09.869.417.2
COCO + VG86.267.081.583.082.582.813.850.421.7
LLaVA-v1.5X69.956.466.881.990.886.158.785.769.7
COCO87.276.684.988.685.387.058.987.570.4
COCO + VG86.177.084.189.883.786.764.586.173.7
", + "image_path": "b52ce66fed017e3d5717e1f5a1965db4d4a0c80c348923d6f3d5f429853c1f64.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 156, + 545, + 221 + ], + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 221 + ], + "type": "text", + "content": "Table 3. Improved Baseline via Object Enumeration: Adding our object enumeration task to LLaVA training and inference leads to large improvements on THRONE particularly in terms of class-wise precision, " + }, + { + "bbox": [ + 304, + 156, + 545, + 221 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{CLS}}" + }, + { + "bbox": [ + 304, + 156, + 545, + 221 + ], + "type": "text", + "content": ", over standard LLaVA models, demonstrating a reduction in Type I hallucinations, as well as small reductions in Type II hallucinations judged by POPE and POPE-C." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 228, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 347 + ], + "type": "text", + "content": "COCO images where available. On THRONE, we observe a large increase in classwise precision and therefore " + }, + { + "bbox": [ + 304, + 228, + 545, + 347 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{CLS}}^{0.5}" + }, + { + "bbox": [ + 304, + 228, + 545, + 347 + ], + "type": "text", + "content": ", particularly for LLaVA-v1.5, demonstrating the ability of our method to reduce Type I hallucinations. Moreover, on POPE and POPE-C, using our object enumeration yields small improvements in precision, indicating reduced Type II hallucinations as well. In the Supplementary Material, we ablate the sampling of negatives during object enumeration training and the effect of removing the object enumeration task during inference." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 353, + 384, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 353, + 384, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 353, + 384, + 365 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 370, + 545, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 545, + 621 + ], + "type": "text", + "content": "We establish a novel benchmark, THRONE, for evaluating hallucinations generated by LVLMs in free-form image descriptions i.e. Type I hallucinations. Our benchmark utilizes multiple LMs and prompt formats with a simple voting mechanism to yield an accurate evaluation of Type I hallucinations in LVLM responses. We ensure that THRONE is broadly accessible by utilizing open-source LMs capable of running on a single commercial GPU. Using THRONE, we benchmark 11 publicly available LVLMs on two datasets, COCO and Objects365, and demonstrate that limited progress has been made in addressing Type I hallucinations. Moreover, we show how the established benchmark, POPE, underestimates Type II hallucinations, which occur in response to specific questions e.g. yes-no questions. We present results for a completed version (POPE-C) to enable a comparison of Type I hallucinations through THRONE and Type II hallucinations using POPE-C. Finally, we propose a simple data augmentation for LVLM training that can result in a large reduction in Type I hallucinations whilst maintaining or improving Type II hallucination performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 622, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 622, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 305, + 622, + 545, + 646 + ], + "type": "text", + "content": "Limitations and Ethical Considerations are discussed in the Supplementary Material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 651, + 365, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 365, + 663 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 365, + 663 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 310, + 669, + 545, + 713 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27235" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "Neural Information Processing Systems, 35:23716-23736, 2022. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 287, + 148 + ], + "type": "text", + "content": "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the International Conference on Computer Vision, pages 2425-2433, 2015. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 150, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 287, + 205 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 206, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 206, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 206, + 287, + 258 + ], + "type": "text", + "content": "[4] Yonatan Bitton, Hritik Bansal, Jack Hessel, Rulin Shao, Wanrong Zhu, Anas Awadalla, Josh Gardner, Rohan Taori, and Ludwig Schmidt. Visit-bench: A benchmark for vision-language instruction following inspired by real-world use, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 261, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 287, + 316 + ], + "type": "text", + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems, 33:1877-1901, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 317, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 287, + 380 + ], + "type": "text", + "content": "[6] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 1, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 383, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 383, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 383, + 287, + 446 + ], + "type": "text", + "content": "[7] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 449, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 449, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 449, + 287, + 491 + ], + "type": "text", + "content": "[8] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2009. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 493, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 493, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 493, + 287, + 568 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. Proceedings of the International Conference on Learning Representations, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "text", + "content": "[10] Mark Everingham, Luc Van Gool, Chris K. I. Williams, John Winn, and Andrew Zisserman. The PASCAL Visual Object Classes (VOC) challenge. International Journal of Computer Vision, 2010. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "text", + "content": "[11] Hugging Face. Text generation inference. https://github.com/huggingface/text-generation-inference, 2023. Accessed: November 10, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[12] Hao Fang, Saurabh Gupta, Forrest Iandola, Rupesh K Srivastava, Li Deng, Piotr Dollar, Jianfeng Gao, Xiaodong He, Margaret Mitchell, John C Platt, et al. From captions to visual concepts and back. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[13] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 545, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 192 + ], + "type": "text", + "content": "[14] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 194, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 248 + ], + "type": "text", + "content": "[15] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 250, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 304 + ], + "type": "text", + "content": "[16] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 305, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 545, + 380 + ], + "type": "text", + "content": "[17] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, Qiang Liu, Kriti Aggarwal, Zewen Chi, Johan Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, and Furu Wei. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 383, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 446 + ], + "type": "text", + "content": "[18] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the International Conference on Machine Learning, pages 4904-4916, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "type": "text", + "content": "[19] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 504, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 568 + ], + "type": "text", + "content": "[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123:32-73, 2017. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "type": "text", + "content": "[21] Kibok Lee, Hao Yang, Satyaki Chakraborty, Zhaowei Cai, Gurumurthy Swaminathan, Avinash Ravichandran, and Onkar Dabeer. Rethinking few-shot object detection on a multi-domain benchmark. In Proceedings of the European Conference on Computer Vision, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 625, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 667 + ], + "type": "text", + "content": "[22] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 1, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[23] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 6, 7" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "27236" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "type": "text", + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning, 2023. 1, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "type": "text", + "content": "[26] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In Proceedings of the Conference on Empirical Methods in Natural Language, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 287, + 296 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, 2014. 2, 3, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "text", + "content": "[28] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 287, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 375 + ], + "type": "text", + "content": "[29] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 409 + ], + "type": "text", + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 410, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 287, + 443 + ], + "type": "text", + "content": "[31] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.6,7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "type": "text", + "content": "[32] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 500, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 565 + ], + "type": "text", + "content": "[33] Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V Le, Barret Zoph, Jason Wei, and Adam Roberts. The flan collection: Designing data and methods for effective instruction tuning. In Proceedings of the International Conference on Machine Learning, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 632 + ], + "type": "text", + "content": "[34] Hwee Tou Ng, Siew Mei Wu, Ted Briscoe, Christian Hadiwinoto, Raymond Hendy Susanto, and Christopher Bryant. The CoNLL-2014 shared task on grammatical error correction. In Proceedings of the Eighteenth Conference on Computational Natural Language Learning: Shared Task, 2014. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 634, + 226, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 634, + 226, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 634, + 226, + 645 + ], + "type": "text", + "content": "[35] OpenAI. Gpt-4 technical report, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[36] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[37] Tina D Purnat, Paolo Vacca, Christine Czerniak, Sarah Ball, Stefano Burzo, Tim Zecchin, Amy Wright, Supriya" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 72, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 72, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 327, + 72, + 545, + 139 + ], + "type": "text", + "content": "Bezbaruah, Faizza Tanggol, Éve Dubé, Fabienne Labbe, Maude Dionne, Jaya Lamichhane, Avichal Mahajan, Sylvie Briand, and Tim Nguyen. Infodemic signal detection during the COVID-19 pandemic: Development of a methodology for identifying potential information voids in online conversations. *JMIR Infodemiology*, 1(1):e30971, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 205 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 205, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 205, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 205, + 545, + 259 + ], + "type": "text", + "content": "[39] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 2020. 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 260, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 260, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 307, + 260, + 545, + 314 + ], + "type": "text", + "content": "[40] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. In Proceedings of the Conference on Empirical Methods in Natural Language, pages 4035-4045, 2018. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "type": "text", + "content": "[41] Giuliano Rossi, Jakub Kolodziej, and Gurvinder Brar. A recommender system for active stock selection. Computational Management Science, 17, 2020. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 349, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 545, + 403 + ], + "type": "text", + "content": "[42] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the International Conference on Computer Vision, 2019. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 404, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 458 + ], + "type": "text", + "content": "[43] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 460, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 503 + ], + "type": "text", + "content": "[44] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. In Advances in Neural Information Processing Systems, pages 200–212, 2021. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 504, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 558 + ], + "type": "text", + "content": "[45] Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, Jitao Sang, and Haoyu Tang. Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 613 + ], + "type": "text", + "content": "[46] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 2022. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "type": "text", + "content": "[47] Hongbin Ye, Tong Liu, Aijia Zhang, Wei Hua, and Weiqiang Jia. Cognitive mirage: A review of hallucinations in large language models. arXiv preprint arXiv:2309.06794, 2023. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[48] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27237" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 352 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[49] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "text", + "content": "[50] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "type": "text", + "content": "[51] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the ai ocean: A survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 287, + 262 + ], + "type": "text", + "content": "[52] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 263, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 263, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 49, + 263, + 287, + 307 + ], + "type": "text", + "content": "[53] Yiyang Zhou, Chenhang Cui, Jaehong Yoon, Linjun Zhang, Zhun Deng, Chelsea Finn, Mohit Bansal, and Huaxiu Yao. Analyzing and mitigating object hallucination in large vision-language models, 2023. 1, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 308, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 308, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 49, + 308, + 287, + 352 + ], + "type": "text", + "content": "[54] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 6, 7" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "27238" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/17566a41-32ff-45d4-9eb1-0099d95d87b1_content_list.json b/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/17566a41-32ff-45d4-9eb1-0099d95d87b1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5450d913dc32372b1b484b52b7eec81313c29fb4 --- /dev/null +++ b/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/17566a41-32ff-45d4-9eb1-0099d95d87b1_content_list.json @@ -0,0 +1,1596 @@ +[ + { + "type": "text", + "text": "TI2V-Zero: Zero-Shot Image Conditioning for Text-to-Video Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 130, + 879, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haomiao Ni $^{1*}$ Bernhard Egger $^{2}$ Toshiaki Koike-Akino $^{3}$", + "bbox": [ + 122, + 179, + 406, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Pennsylvania State University, USA", + "bbox": [ + 156, + 217, + 405, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Suhas Lohit3 \nSharon X. Huang1", + "bbox": [ + 444, + 180, + 589, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anoop Cherian3 Ye Wang3 Tim K. Marks3", + "bbox": [ + 591, + 180, + 846, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Friedrich-Alexander-Universität Erlangen-Nürnberg, Germany", + "bbox": [ + 434, + 218, + 812, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Mitsubishi Electric Research Laboratories (MERL), USA", + "bbox": [ + 308, + 234, + 658, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}\\{\\mathrm{hfn5052, suh972}\\} @ \\mathrm{psu.edu}$ $^{2}$ bernhard.egger@fau.de $^{3}\\{$ slohit,cherian,yewang,koike,tmarks\\}@merl.com https://merl.com/demos/TI2V-Zero", + "bbox": [ + 147, + 253, + 823, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 337 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-conditioned image-to-video generation (TI2V) aims to synthesize a realistic video starting from a given image (e.g., a woman's photo) and a text description (e.g., \"a woman is drinking water\"). Existing TI2V frameworks often require costly training on video-text datasets and specific model designs for text and image conditioning. In this paper, we propose TI2V-Zero, a zero-shot, tuning-free method that empowers a pretrained text-to-video (T2V) diffusion model to be conditioned on a provided image, enabling TI2V generation without any optimization, finetuning, or introducing external modules. Our approach leverages a pretrained T2V diffusion foundation model as the generative prior. To guide video generation with the additional image input, we propose a \"repeat-and-slide\" strategy that modulates the reverse denoising process, allowing the frozen diffusion model to synthesize a video frame-by-frame starting from the provided image. To ensure temporal continuity, we employ a DDPM inversion strategy to initialize Gaussian noise for each newly synthesized frame and a resampling technique to help preserve visual details. We conduct comprehensive experiments on both domain-specific and open-domain datasets, where TI2V-Zero consistently outperforms a recent open-domain TI2V model. Furthermore, we show that TI2V-Zero can seamlessly extend to other tasks such as video infilling and prediction when provided with more images. Its autoregressive design also supports long video generation.", + "bbox": [ + 75, + 353, + 473, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 777, + 209, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image-to-video (I2V) generation is an appealing topic with various applications, including artistic creation, entertainment, and data augmentation for machine learning [39]. Given a single image $x^0$ and a text prompt $y$ , text-conditioned image-to-video (TI2V) generation aims to syn", + "bbox": [ + 75, + 801, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a552a36a79232e32b7df66896081cb440cc36b38a41e9cba63a86b6e6101bc7.jpg", + "image_caption": [ + "\"A man with the expression of slight happiness on his face.\"" + ], + "image_footnote": [], + "bbox": [ + 522, + 321, + 612, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dd64b4d666b3b23a4249fb9b9314ff6ab5fb3f25a1ec5eda2d6d7e7a5757c230.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 321, + 782, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2b7a08a4ccf8d6ac829afff081b3aa5308b445a4bd3bcdd6a561a5cb5b7d757.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 321, + 870, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab8124efa72cd21ec71fda21bf1d7b17a2db63a69309a7dedb957f5d1868bcd4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 402, + 611, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3341998c58263fcb3ee286e3c475bdd78408c220aba8b4560a43d8834c2682a7.jpg", + "image_caption": [ + "\"A person is drumming.\"" + ], + "image_footnote": [], + "bbox": [ + 612, + 402, + 782, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f82fd823a376b841b9226db745d0fa81ebc9f9ba64f7ae1eb7a189b629d0d551.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 402, + 870, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7a9448aedbf0e9a0f0ed0937494a78221e422e242cf482ee702896a063f88270.jpg", + "image_caption": [ + "\"A serene mountain cabin covered in a fresh blanket of snow.\"", + "Figure 1. Examples of generated video frames using our proposed T12V-Zero. The given first image $x^0$ is highlighted with the red box, and the text condition $y$ is shown under each row of the video. The remaining columns show the 6th, 11th, and 16th frames of the generated output videos. Each generated video has 16 frames with a resolution of $256 \\times 256$ ." + ], + "image_footnote": [], + "bbox": [ + 522, + 484, + 612, + 553 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/de3ef1e82e3d3559bd02ac5efec8fb5ef96ff3502308f0884ba35e9ec68ac2c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 484, + 784, + 551 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd730428c2e921fad09caaf0a6596fb7577fa91d821f48af29ef50fc427d7fad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 484, + 870, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "thesize $M$ new frames to yield a realistic video, $\\hat{\\mathbf{x}} = \\langle x^0,\\hat{x}^1,\\dots ,\\hat{x}^M\\rangle$ , starting from the given frame $x^0$ and satisfying the text description $y$ . Current TI2V generation methods [59, 63, 70] typically rely on computationally-heavy training on video-text datasets and specific architecture designs to enable text and image conditioning. Some [12, 25] are constrained to specific domains due to the lack of training with large-scale open-domain datasets. Other approaches, such as [14, 67], utilize pretrained foundation models to reduce training costs, but they still need to train additional modules using video data.", + "bbox": [ + 496, + 672, + 892, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose TI2V-Zero, which achieves zero-shot TI2V generation using only an open-domain pretrained text-to-video (T2V) latent diffusion model [60]. Here \"zero-shot\" means that when using the diffusion", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during an internship at MERL.", + "bbox": [ + 94, + 886, + 326, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9015", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "model (DM) that was trained only for text conditioning, our framework enables image conditioning without any optimization, fine-tuning, or introduction of additional modules. Specifically, we guide the generation process by incorporating the provided image $x^0$ into the output latent code at each reverse denoising step. To ensure that the temporal attention layers of the pretrained DM focus on information from the given image, we propose a \"repeat-and-slide\" strategy to synthesize the video in a frame-by-frame manner, rather than directly generating the entire video volume. Notably, TI2V-Zero is not trained for the specific domain of the provided image, thus allowing the model to generalize to any image during inference. Additionally, its autoregressive generation makes the synthesis of long videos possible.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While the standard denoising sampling process starting with randomly initialized Gaussian noise can produce matching semantics, it often results in temporally inconsistent videos. Therefore, we introduce an inversion strategy based on the DDPM [20] forward process, to provide a more suitable initial noise for generating each new frame. We also apply a resampling technique [33] in the video DM to help preserve the generated visual details. Our approach ensures that the network maintains temporal consistency, generating visually convincing videos conditioned on the given starting image (see Fig. 1).", + "bbox": [ + 75, + 303, + 472, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments on MUG [1], UCF-101 [56], and a new open-domain dataset. In these experiments, TI2V-Zero consistently performs well, outperforming a state-of-the-art model [67] that was based on a video diffusion foundation model [8] and was specifically trained to enable open-domain TI2V generation.", + "bbox": [ + 76, + 470, + 472, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 578, + 218, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Conditional Image-to-Video Generation", + "text_level": 1, + "bbox": [ + 76, + 604, + 419, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conditional video generation aims to synthesize videos guided by user-provided signals. It can be classified according to which type(s) of conditions are given, such as text-to-video (T2V) generation [5, 16, 21, 23, 31, 65], video-to-video (V2V) generation [7, 38, 40, 45, 61, 64], and image-to-video (I2V) generation [4, 10, 25, 34, 39, 69]. Here we discuss previous text-conditioned image-to-video (TI2V) generation methods [12, 14, 22, 44, 63, 70]. Hu et al. [25] introduced MAGE, a TI2V generator that integrates a motion anchor structure to store appearance-motion-aligned representations through three-dimensional axial transformers. Yin et al. [70] proposed DragNUWA, a diffusion-based model capable of generating videos controlled by text, image, and trajectory information with three modules including a trajectory sampler, a multi-scale fusion, and an adaptive training strategy. However, these TI2V frameworks require computationally expensive training on videotext datasets and a particular model design to support text", + "bbox": [ + 75, + 628, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and-image-conditioned training. In contrast, our proposed TI2V-Zero leverages a pretrained T2V diffusion model to achieve zero-shot TI2V generation without additional optimization or fine-tuning, making it suitable for a wide range of applications.", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Adaptation of Diffusion Foundation Models", + "text_level": 1, + "bbox": [ + 498, + 175, + 870, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the recent successful application of diffusion models (DM) [20, 42, 47, 54, 55] to both image and video generation, visual diffusion foundation models have gained prominence. These include text-to-image (T2I) models such as Imagen [50] and Stable Diffusion [47], as well as text-to-video (T2V) models such as ModelScopeT2V [60] and VideoCrafter1 [8]. These models are trained with large-scale open-domain datasets, often including LAION-400M [52] and WebVid-10M [2]. They have shown immense potential for adapting their acquired knowledge base to address a wide range of downstream tasks, thereby reducing or eliminating the need for extensive labeled data. For example, previous works have explored the application of large T2I models to personalized image generation [13, 49], image editing [17, 33, 35-37], image segmentation [3, 68], video editing [45, 62], and video generation [14, 27, 53, 66]. In contrast to T2I models, there are fewer works on the adaptation of large-scale T2V models. Xing et al. [67] proposed DynamicCrafter for open-domain TI2V generation by adapting a T2V foundation model [8]. To control the generative process, they first employed a learnable image encoding network to project the given image into a text-aligned image embedding space. Subsequently, they utilized dual cross-attention layers to fuse text and image information and also concatenated the image with the initial noise to provide the video DM with more precise image details. In contrast, in this paper we explore how to inject the provided image to guide the DM sampling process based solely on the pretrained T2V model itself, with no additional training for the new TI2V task.", + "bbox": [ + 496, + 196, + 893, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 498, + 664, + 635, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given one starting image $x^0$ and text $y$ , let $\\mathbf{x} = \\langle x^0, x^1, \\ldots, x^M \\rangle$ represent a real video corresponding to text $y$ . The objective of text-conditioned image-to-video (TI2V) generation is to synthesize a video $\\hat{\\mathbf{x}} = \\langle x^0, \\hat{x}^1, \\ldots, \\hat{x}^M \\rangle$ , such that the conditional distribution of $\\hat{\\mathbf{x}}$ given $x^0$ and $y$ is identical to the conditional distribution of $\\mathbf{x}$ given $x^0$ and $y$ , i.e., $p(\\hat{\\mathbf{x}} | x^0, y) = p(\\mathbf{x} | x^0, y)$ . Our proposed TI2V-Zero can be built on a pretrained T2V diffusion model with a 3D-UNet-based denoising network. Here we choose ModelScopeT2V [60] as backbone due to its promising open-domain T2V generation ability. Below, we first introduce preliminaries about diffusion models, then introduce the architecture of the pretrained T2V model, and finally present the details of our TI2V-Zero.", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9016", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cc501315653f24ec3a0c89a54855a81cf4fd0a353a3f95841365890bf17fc7e0.jpg", + "image_caption": [ + "Figure 2. Illustration of the process of applying TI2V-Zero to generate the new frame $\\hat{x}^{i+1}$ , given the starting image $x^0$ and text $y$ . TI2V-Zero is built upon a frozen pretrained T2V diffusion model, including frame encoder $\\mathcal{E}$ , frame decoder $\\mathcal{D}$ , and the denoising U-Net $\\epsilon_{\\theta}$ . At the beginning of generation ( $i=0$ ), we encode $x^0$ as $z^0$ and repeat it $K$ times to form the queue $\\mathbf{s}_0$ . We then apply DDPM-based inversion to $\\mathbf{s}_0$ to produce the initial Gaussian noise $\\hat{\\mathbf{z}}_T$ . Subsequently, in each reverse denoising step using U-Net $\\epsilon_{\\theta}$ , we keep replacing the first $K$ frames of $\\hat{\\mathbf{z}}_t$ with the noisy latent code $\\mathbf{s}_t$ derived from $\\mathbf{s}_0$ . Resampling is also applied within each step to improve motion coherence. We finally decode the final frame of the clean latent code $\\hat{\\mathbf{z}}_0$ as the new synthesized frame $\\hat{x}^{i+1}$ . To compute the new $\\mathbf{s}_0$ for the next iteration of generation ( $i>0$ ), we perform a sliding operation by dequeuing $s_0^0$ and enqueuing $\\hat{z}_0^K$ within $\\mathbf{s}_0$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries: Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 419, + 361, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models (DM) [20, 54, 55] are probabilistic models designed to learn a data distribution. Here we introduce the fundamental concepts of Denoising Diffusion Probabilistic Models (DDPM). Given a sample from the data distribution $\\mathbf{z}_0 \\sim q(\\mathbf{z}_0)$ , the forward diffusion process of a DM produces a Markov chain $\\mathbf{z}_1, \\ldots, \\mathbf{z}_T$ by iteratively adding Gaussian noise to $\\mathbf{z}_0$ according to a variance schedule $\\beta_1, \\ldots, \\beta_T$ , that is:", + "bbox": [ + 75, + 441, + 468, + 553 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {z} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 556, + 468, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where variances $\\beta_{t}$ are constant. When the $\\beta_{t}$ are small, the posterior $q(\\mathbf{z}_{t - 1}|\\mathbf{z}_t)$ can be well approximated by a diagonal Gaussian [41, 54]. Furthermore, if the length of the chain, denoted by $T$ , is sufficiently large, $\\mathbf{z}_T$ can be well approximated by a standard Gaussian distribution $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . These suggest that the true posterior $q(\\mathbf{z}_{t - 1}|\\mathbf{z}_t)$ can be estimated by $p_{\\theta}(\\mathbf{z}_{t - 1}|\\mathbf{z}_t)$ defined as:", + "bbox": [ + 75, + 574, + 468, + 672 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {z} _ {t - 1} \\mid \\mathbf {z} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {z} _ {t}\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 674, + 468, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where variances $\\sigma_t$ are also constants. The reverse denoising process in the DM (also termed sampling) then generates samples $\\mathbf{z}_0 \\sim p_{\\theta}(\\mathbf{z}_0)$ by starting with Gaussian noise $\\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ and gradually reducing noise in a Markov chain $\\mathbf{z}_{T-1}, \\mathbf{z}_{T-2}, \\ldots, \\mathbf{z}_0$ using a learned $p_{\\theta}(\\mathbf{z}_{t-1}|\\mathbf{z}_t)$ . To learn $p_{\\theta}(\\mathbf{z}_{t-1}|\\mathbf{z}_t)$ , Gaussian noise $\\epsilon$ is first added to $\\mathbf{z}_0$ to generate samples $\\mathbf{z}_t$ . Utilizing the independence property of the noise added at each forward step in Eq. (1), we can calculate the total noise variance as $\\bar{\\alpha}_t = \\prod_{i=0}^t (1 - \\beta_i)$ and transform $\\mathbf{z}_0$ to $\\mathbf{z}_t$ in a single step:", + "bbox": [ + 75, + 691, + 468, + 834 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {z} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 835, + 468, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then a model $\\epsilon_{\\theta}$ is trained to predict $\\epsilon$ using the following mean-squared error loss:", + "bbox": [ + 76, + 854, + 468, + 882 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL = \\mathbb {E} _ {t \\sim \\mathcal {U} (1, T), \\mathbf {z} _ {0} \\sim q (\\mathbf {z} _ {0}), \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t) \\| ^ {2} \\right] ， \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 883, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where diffusion step $t$ is uniformly sampled from $\\{1, \\dots, T\\}$ . Then $\\mu_{\\theta}(\\mathbf{z}_t)$ in Eq. (2) can be derived from $\\epsilon_{\\theta}(\\mathbf{z}_t, t)$ to model $p_{\\theta}(\\mathbf{z}_{t-1}|\\mathbf{z}_t)$ [20]. The denoising model $\\epsilon_{\\theta}$ is implemented using a time-conditioned U-Net [48] with residual blocks [15] and self-attention layers [58]. Diffusion step $t$ is specified to $\\epsilon_{\\theta}$ by the sinusoidal position embedding [58]. Conditional generation that samples $\\mathbf{z}_0 \\sim p_{\\theta}(\\mathbf{z}_0|y)$ can be achieved by learning a $y$ -conditioned model $\\epsilon_{\\theta}(\\mathbf{z}_t, t, y)$ [41, 47] with classifier-free guidance [19]. During training, the condition $y$ in $\\epsilon_{\\theta}(\\mathbf{z}_t, t, y)$ is replaced by a null label $\\emptyset$ with a fixed probability. When sampling, the output is generated as follows:", + "bbox": [ + 496, + 419, + 890, + 587 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} _ {\\theta} \\left(\\mathbf {z} _ {t}, t, y\\right) = \\epsilon_ {\\theta} \\left(\\mathbf {z} _ {t}, t, \\emptyset\\right) + g \\cdot \\left(\\epsilon_ {\\theta} \\left(\\mathbf {z} _ {t}, t, y\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {z} _ {t}, t, \\emptyset\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 589, + 890, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $g$ is the guidance scale.", + "bbox": [ + 504, + 608, + 702, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Architecture of Pretrained T2V Model", + "text_level": 1, + "bbox": [ + 500, + 633, + 831, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TI2V-Zero can be built upon a pretrained T2V diffusion model with a 3D-UNet-based denoising network. Here we choose ModelScopeT2V [60] as the pretrained model (denoted $\\mathcal{M}$ ). We now describe this T2V model in detail.", + "bbox": [ + 496, + 657, + 890, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Structure Overview. Given a text prompt $y$ , the T2V model $\\mathcal{M}$ synthesizes a video $\\hat{\\mathbf{x}} = \\langle \\hat{x}^0, \\hat{x}^1, \\dots, \\hat{x}^K \\rangle$ with a pre-defined video of length $(K + 1)$ using a latent video diffusion model. Similar to Latent Diffusion Models (LDM) [47], $\\mathcal{M}$ incorporates a frame auto-encoder [11, 28] for the conversion of data between pixel space $\\mathcal{X}$ and latent space $\\mathcal{Z}$ through its encoder $\\mathcal{E}$ and decoder $\\mathcal{D}$ . Given the real video $\\mathbf{x} = \\langle x^0, x^1, \\dots, x^K \\rangle$ , $\\mathcal{M}$ first utilizes the frame encoder $\\mathcal{E}$ to encode the video $\\mathbf{x}$ as $\\mathbf{z} = \\langle z^0, z^1, \\dots, z^K \\rangle$ . Here the sizes of pixel frame $x$ and latent frame $z$ are $H_x \\times W_x \\times 3$ and $H_z \\times W_z \\times C_z$ , respectively. To be consistent with the notation used for the DM, we denote the", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9017", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Generation using our TI2V-Zero approach." + ], + "code_body": "Input: The starting frame $x^0$ ; The text prompt $y$ ; The pretrained T2V Model $\\mathcal{M}$ for generating $(K + 1)$ -frame videos, including frame encoder $\\mathcal{E}$ and frame decoder $\\mathcal{D}$ , and the DM de-noising networks $\\epsilon_{\\theta}$ ; The iteration number $U$ for resampling; The parameter $M$ to control the length of the output video. \nOutput: A synthesized video $\\hat{\\mathbf{x}}$ with $(M + 1)$ frames. \n1: $z^0 \\gets \\mathcal{E}(x^0)$ // Encode $x^0$ \n2: $\\mathbf{s}_0 \\gets \\langle z^0, z^0, \\dots, z^0 \\rangle$ // Repeat $z^0$ for $K$ times \n3: $\\hat{\\mathbf{x}} \\gets \\langle x^0 \\rangle$ \n4: for $i = 1, 2, \\dots, M$ do \n// Generate one new frame $\\hat{x}^i$ \n5: $\\mathbf{s}_T \\sim \\mathcal{N}(\\sqrt{\\bar{\\alpha}_T} \\mathbf{s}_0, (1 - \\bar{\\alpha}_T) \\mathbf{I})$ // DDPM Inversion \n6: $\\hat{z}_T^K \\sim \\mathcal{N}(\\sqrt{\\bar{\\alpha}_T} s_0^{K-1}, (1 - \\bar{\\alpha}_T) \\mathbf{I})$ \n7: $\\hat{\\mathbf{z}}_T \\gets \\mathbf{s}_T \\cup \\hat{z}_T^K$ // Initialize $\\hat{\\mathbf{z}}_T$ \n8: for $t = T-1, \\dots, 2, 1$ do \n $\\mathbf{s}_t \\sim \\mathcal{N}(\\sqrt{\\bar{\\alpha}_t} \\mathbf{s}_0, (1 - \\bar{\\alpha}_t) \\mathbf{I})$ \nfor $u = 1, 2, \\dots, U$ do \n $\\langle \\hat{z}_t^0, \\hat{z}_t^1, \\dots, \\hat{z}_t^{K-1} \\rangle \\gets \\mathbf{s}_t$ // Replace \n $\\hat{\\mathbf{z}}_{t-1} \\sim \\mathcal{N}(\\mu_\\theta (\\hat{\\mathbf{z}}_t, y), \\sigma_t^2 \\mathbf{I})$ \nif $u < U$ and $t > 1$ then \n $\\hat{\\mathbf{z}}_t \\sim \\mathcal{N}(\\sqrt{1 - \\beta_t} \\hat{\\mathbf{z}}_{t-1}, \\beta_t \\mathbf{I})$ // Resample \nend if \nend for \nend for \n $\\mathbf{s}_0 \\gets \\langle s_0^1, s_0^2, \\dots, s_0^{K-1} \\rangle \\cup \\hat{z}_0^K$ // Slide \n $\\hat{x}^i \\gets D(\\hat{z}_0^K)$ // Decode $\\hat{z}_0^K$ $\\hat{\\mathbf{x}} \\gets \\hat{\\mathbf{x}} \\cup \\hat{x}^i$ \nend for \nreturn $\\hat{\\mathbf{x}}$", + "bbox": [ + 76, + 109, + 472, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "clean video latent $\\mathbf{z} = \\mathbf{z}_0 = \\langle z_0^0, z_0^1, \\ldots, z_0^K \\rangle$ . $\\mathcal{M}$ then learns a DM on the latent space $\\mathcal{Z}$ through a 3D denoising U-Net $\\epsilon_{\\theta}$ [9]. Let $\\mathbf{z}_t = \\langle z_t^0, z_t^1, \\ldots, z_t^K \\rangle$ represent the latent sequence that results from adding noise over $t$ steps to the original latent sequence $\\mathbf{z}_0$ . When training, the forward diffusion process of a DM transforms the initial latent sequence $\\mathbf{z}_0$ into $\\mathbf{z}_T$ by iteratively adding Gaussian noise $\\epsilon$ for $T$ steps. During inference, denoising U-Net $\\epsilon_{\\theta}$ predicts the added noise at each step, enabling the generation of the clean latent sequence $\\hat{\\mathbf{z}}_0 = \\langle \\hat{z}_0^0, \\hat{z}_0^1, \\ldots, \\hat{z}_0^K \\rangle$ starting from randomly sampled Gaussian noise $\\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ .", + "bbox": [ + 75, + 546, + 468, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text Conditioning Mechanism. $\\mathcal{M}$ employs a cross-attention mechanism [47] to incorporate text information into the generative process as guidance. Specifically, $\\mathcal{M}$ uses a pretrained CLIP model [46] to encode the prompt $y$ as the text embedding $e$ . The embedding $e$ is later used as the key and value in the multi-head attention layer within the spatial attention blocks, thus enabling the integration of text features with the intermediate U-Net features in $\\epsilon_{\\theta}$ .", + "bbox": [ + 75, + 715, + 468, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Denoising U-Net. The denoising U-Net $\\epsilon_{\\theta}$ includes four key building blocks: the initial block, the downsampling block, the spatio-temporal block, and the upsampling block. The initial block transfers the input into the embedding", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4be811a2a4ed11d79312b906b57c48e810421e0172805ae5595114d846b4ecc4.jpg", + "image_caption": [ + "Video Time", + "Figure 3. Illustration of the motivation behind our framework. We explore the application of a replacing-based baseline approach (rows 2-4, labeled \"Replacing\") and our TI2V-Zero (rows 5-6, labeled \"TI2V-Zero\") in various video generation tasks. The given real frames for each task are highlighted by red boxes and the text input is shown under the block. The replacing-based approach is only effective at predicting a single frame when all the other frames in the video are provided, while TI2V-Zero generates temporally coherent videos for both the TI2V and video infilling tasks." + ], + "image_footnote": [], + "bbox": [ + 516, + 98, + 875, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "space, while the downsampling and upsampling blocks are responsible for spatially downsampling and upsampling the feature maps. The spatio-temporal block is designed to capture spatial and temporal dependencies in the latent space, which comprises 2D spatial convolution, 1D temporal convolution, 2D spatial attention, and 1D temporal attention.", + "bbox": [ + 496, + 554, + 890, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Our Framework", + "text_level": 1, + "bbox": [ + 500, + 657, + 663, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Leveraging the pretrained T2V foundation model $\\mathcal{M}$ , we first propose a straightforward replacing-based baseline for adapting $\\mathcal{M}$ to TI2V generation. We then analyze the possible reasons why it fails and introduce our TI2V-Zero framework, which includes a repeat-and-slide strategy, DDPM-based inversion, and resampling. Figure 2 and Algorithm 1 demonstrate the inference process of TI2V-Zero.", + "bbox": [ + 496, + 681, + 890, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Replacing-based Baseline. We assume that the pretrained model $\\mathcal{M}$ is designed to generate the video with a fixed length of $(K + 1)$ . So we first consider synthesizing videos with that same length $(K + 1)$ , i.e., $M = K$ . Since the DM process operates within the latent space $\\mathcal{Z}$ , we use the encoder $\\mathcal{E}$ to map the given starting frame $x^0$ into the latent representation $z^0$ . Additionally, we denote $z^0 = z_0^0$ to specify that the latent is clean and corresponds", + "bbox": [ + 496, + 789, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9018", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to diffusion step 0 of the DM. Note that each reverse denoising step in Eq. (2) from $\\hat{\\mathbf{z}}_t$ to $\\hat{\\mathbf{z}}_{t - 1}$ depends solely on $\\hat{\\mathbf{z}}_t = \\langle \\hat{z}_t^0,\\hat{z}_t^1,\\dots ,\\hat{z}_t^K\\rangle$ . To ensure that the first frame of the final synthesized clean video latent $\\hat{\\mathbf{z}}_0 = \\langle \\hat{z}_0^0,\\hat{z}_0^1,\\dots ,\\hat{z}_0^K\\rangle$ at step 0 matches the provided image latent, i.e., $\\hat{z}_0^0 = z_0^0$ we can modify the first generated latent $\\hat{z}_t^0$ of $\\hat{\\mathbf{z}}_t$ at each reverse step, as long as the signal-to-noise ratio of each frame latent in $\\hat{\\mathbf{z}}_t$ remains consistent. Using Eq. (3), we can add $t$ steps of noise to the provided image latent $z_0^0$ , allowing us to sample $z_{t}^{0}$ through a single-step calculation. By replacing the first generated latent $\\hat{z}_t^0$ with the noisy image latent $z_{t}^{0}$ at each reverse denoising step, we might expect that the video generation process can be guided by $z_0^0$ with the following expressions defined for each reverse step:", + "bbox": [ + 76, + 90, + 472, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} ^ {0} \\sim \\mathcal {N} \\left(\\sqrt {\\bar {\\alpha}} _ {t} z _ {0} ^ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {6a}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 292, + 468, + 309 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {t} ^ {0} \\leftarrow z _ {t} ^ {0}, \\tag {6b}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 311, + 468, + 328 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {z}} _ {t - 1} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} \\left(\\hat {\\mathbf {z}} _ {t}, y\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right). \\tag {6c}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 332, + 468, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, in each reverse step from $\\hat{\\mathbf{z}}_t$ to $\\hat{\\mathbf{z}}_{t - 1}$ , as shown in Eq. (6a), we first compute the noisy latent $z_{t}^{0}$ by adding Gaussian noise to the given image latent $z_0^0$ over $t$ steps. Then, we replace the first latent $\\hat{z}_t^0$ of $\\hat{\\mathbf{z}}_t$ with $z_{t}^{0}$ in Eq. (6b) to incorporate the provided image into the generation process. Finally, in Eq. (6c), we pass $\\hat{\\mathbf{z}}_t$ through the denoising network to generate $\\hat{\\mathbf{z}}_{t - 1}$ , where the text $y$ is integrated by classifier-free guidance (Eq. (5)). After $T$ iterations, the final clean latent $\\hat{\\mathbf{z}}_0$ at diffusion step 0 can be mapped back into the image space $\\mathcal{X}$ using the decoder $\\mathcal{D}$ .", + "bbox": [ + 75, + 357, + 468, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using this replacing-based baseline, we might expect that the temporal attention layers in $\\epsilon_{\\theta}$ can utilize the context provided by the first frame latent $\\hat{z}_t^0$ to generate the subsequent frame latents in a manner that harmonizes with $\\hat{z}_t^0$ . However, as shown in Fig. 3, row 2, this replacing-based approach fails to produce a video that is temporally consistent with the first image. The generated frames are consistent with each other, but not with the provided first frame.", + "bbox": [ + 75, + 508, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To analyze possible reasons for failure, we apply this baseline to a simpler video infilling task, where every other frame is provided and the model needs to predict the interspersed frames. In this case, the baseline replaces the generated frame latents at positions corresponding to real frames with noisy provided-frame latents in each reverse step. The resulting video, in Fig. 3, row 3, looks like a combination of two independent videos: the generated (even) frames are consistent with each other but not with the provided (odd) frames. We speculate that this may result from the intrinsic dissimilarity between frame latents derived from the given real images and those sampled from $\\epsilon_{\\theta}$ . Thus, the temporal attention values between frame latents sampled in the same way (both from the given images or both from $\\epsilon_{\\theta}$ ) will be higher, while the attention values between frame latents sampled in different ways (one from the given image and the other from $\\epsilon_{\\theta}$ ) will be lower. Therefore, the temporal attention layers of $\\mathcal{M}$ tend to utilize the information from latents", + "bbox": [ + 75, + 628, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/56a3fa54e63647ab3a9f3f42593a71457e609a3ccf459ff78b944420e667deec.jpg", + "image_caption": [ + "Figure 4. Qualitative ablation study comparing different sampling strategies for our TI2V-Zero on MUG. The first image $\\hat{x}^0$ is highlighted with the red box and text $y$ is shown under the block. The 1st, 6th, 11th, and 16th frames of the videos are shown in each column. The terms Inversion, DDIM, and Resample denote the application of DDPM inversion, the steps using DDIM sampling, and the iteration number using resampling, respectively." + ], + "image_footnote": [], + "bbox": [ + 521, + 90, + 870, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "produced by $\\epsilon_{\\theta}$ to synthesize new frames at each reverse step, ignoring the provided frames. We further simplify the task to single-frame prediction, where the model only needs to predict a single frame when all the other frames in the video are given. In this setting, all the frame latents except for the final frame are replaced by noisy provided-frame latents in each reverse step. Thus, temporal attention layers can only use information from the real frames. In this case, Fig. 3, row 4, shows that the baseline can now generate a final frame that is consistent with the previous frames.", + "bbox": [ + 496, + 554, + 890, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Repeat-and-Slide Strategy. Inspired by the observation in Fig. 3, to guarantee that the temporal attention layers of $\\mathcal{M}$ depend solely on the given image, we make two major changes to the proposed replacing-based baseline: (1) instead of using $\\mathcal{M}$ to directly synthesize the entire $(K + 1)$ -frame video, we switch to a frame-by-frame generation approach, i.e., we generate only one new frame latent in each complete DM sampling process; (2) for each sampling process generating the new frame latent, we ensure that only one frame latent is produced from $\\epsilon_{\\theta}$ , while the other $K$ frame latents are derived from the given real image and previously synthesized frames, thereby forcing temporal attention layers to only use the information from these frame latents. Specifically, we construct a queue of", + "bbox": [ + 496, + 705, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9019", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$K$ frame latents, denoted as $\\mathbf{s}_0 = \\langle s_0^0,s_0^1,\\dots ,s_0^{K - 1}\\rangle$ . We also define $\\mathbf{s}_t = \\langle s_t^0,s_t^1,\\dots ,s_t^{K - 1}\\rangle$ , which is obtained by adding $t$ steps of Gaussian noise to the clean $\\mathbf{s}_0$ . Similar to our replacing-based baseline in the single-frame prediction task, in each reverse step from $\\hat{\\mathbf{z}}_t$ to $\\hat{\\mathbf{z}}_{t - 1}$ , we replace the first $K$ frame latents in $\\hat{\\mathbf{z}}_t$ by $\\mathbf{s}_t$ . Consequently, the temporal attention layers have to utilize information from $\\mathbf{s}_0$ to synthesize the new frame's latent, $\\hat{z}_0^K$ . Considering that only one starting image latent $z^0$ is provided, we propose a \"repeat-and-slide\" strategy to construct $\\mathbf{s}_0$ . At the beginning of video generation, we repeat $z^0$ for $K$ frames to form $\\mathbf{s}_0$ , and gradually perform a sliding operation within the queue $\\mathbf{s}_0$ by dequeuing the first frame latent $s_0^0$ and enqueuing the newly generated latent $\\hat{z}_0^K$ after each complete DM sampling process. Note that though the initial $\\mathbf{s}_0$ is created by repeating $z^0$ , the noise added to get $\\mathbf{s}_t$ is different for each frame's latent in $\\mathbf{s}_t$ , thus ensuring diversity. The following expressions define one reverse step in the DM sampling process:", + "bbox": [ + 76, + 89, + 472, + 357 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {t} \\sim \\mathcal {N} \\left(\\sqrt {\\bar {\\alpha}} _ {t} \\mathbf {s} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {7a}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 366, + 468, + 383 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\langle \\hat {z} _ {t} ^ {0}, \\hat {z} _ {t} ^ {1}, \\dots , \\hat {z} _ {t} ^ {K - 1} \\right\\rangle \\leftarrow \\mathbf {s} _ {t}, \\tag {7b}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 385, + 468, + 402 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {z}} _ {t - 1} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} \\left(\\hat {\\mathbf {z}} _ {t}, y\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right). \\tag {7c}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 405, + 468, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, in each reverse denoising step from $\\hat{\\mathbf{z}}_t$ to $\\hat{\\mathbf{z}}_{t - 1}$ , we first add $t$ steps of Gaussian noise to the queue $\\mathbf{s}_0$ to yield $\\mathbf{s}_t$ in Eq. (7a). Subsequently, we replace the previous $K$ frames of $\\hat{\\mathbf{z}}_t$ with $\\mathbf{s}_t$ in Eq. (7b) and input $\\hat{\\mathbf{z}}_t$ to the denoising network to produce the less noisy latent $\\hat{\\mathbf{z}}_{t - 1}$ (Eq. (7c)).", + "bbox": [ + 76, + 431, + 468, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the repeat-and-slide strategy, model $\\mathcal{M}$ is tasked with predicting only one new frame, while the preceding $K$ frames are incorporated into the reverse process to ensure that the temporal attention layers depend solely on information derived from the provided image.", + "bbox": [ + 76, + 508, + 468, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DDPM-based Inversion. Though the DM sampling process starting with randomly sampled Gaussian noise produces matching semantics, the generated video is often temporally inconsistent (Fig. 4, row 2). To provide initial noise that can produce more temporally consistent results, we introduce an inversion strategy based on the DDPM [20] forward process when generating the new frame latent. Specifically, at the beginning of each DM sampling process to synthesize the new frame latent $\\hat{z}_0^K$ , instead of starting with the $\\hat{\\mathbf{z}}_T$ randomly sampled from $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , we add $T$ full steps of Gaussian noise to $\\mathbf{s}_0$ to obtain $\\mathbf{s}_T$ using Eq. (3). Note that $\\hat{\\mathbf{z}}$ has $K + 1$ frames, while $\\mathbf{s}$ has $K$ frames. We then use $\\mathbf{s}_T$ to initialize the first $K$ frames of $\\hat{\\mathbf{z}}_T$ . We copy the last frame $s_T^{K - 1}$ of $\\mathbf{s}_T$ to initialize the final frame $\\hat{z}_T^K$ , as the $(K - 1)\\mathrm{th}$ frame is the closest to the $K$ th frame.", + "bbox": [ + 76, + 583, + 470, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Resampling. Similar to [24, 33], we further apply a resampling technique, which was initially designed for the image inpainting task, to the video DM to enhance motion coherence. Particularly, after performing a one-step denoising operation in the reversed process, we add one-step noise again to revert the latent. This procedure is repeated mul", + "bbox": [ + 76, + 810, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5b1d8355d50eefbe5f759ef960183fdece5e7e2f6602a266b18fe9a94a97c1b8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
InversionDDIMResampleFVD↓sFVD↓tFVD↓
X1001656.372074.77±411.741798.05±235.34
✓100339.89443.97±139.10405.22±61.58
✓500463.55581.32±234.09535.06±85.27
✓102207.62299.14±87.24278.73±47.84
✓104180.09267.17±74.72252.77±39.02
", + "bbox": [ + 500, + 88, + 893, + 157 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6024d8e42df2a9a00dc2ab17fa2698aa3bb0491442a95921910bfb8da50488ce.jpg", + "table_caption": [ + "Table 1. Quantitative ablation study comparing different sampling strategies for proposed TI2V-Zero on the MUG dataset. Inversion, DDIM, and Resample denote the application of DDPM-based inversion, the steps using DDIM sampling, and the iteration number using resampling, respectively." + ], + "table_footnote": [], + "table_body": "
Distributions for ComparisonFVD↓tFVD↓
TI2V-Zero-Fake vs. ModelScopeT2V366.41921.31±251.85
TI2V-Zero-Real vs. Real Videos477.191306.75±271.82
ModelScopeT2V vs. Real Videos985.822264.08±501.28
TI2V-Zero-Fake vs. Real Videos937.112177.70±436.71
", + "bbox": [ + 519, + 250, + 874, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Result analysis of TI2V-Zero starting from the real (i.e., TI2V-Zero-Real) or synthesized frames (i.e., TI2V-Zero-Fake) on the UCF101 dataset.", + "bbox": [ + 498, + 323, + 892, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tiple times for each diffusion step, ensuring harmonization between the predicted and conditioning frame latents (see Algorithm 1 for details).", + "bbox": [ + 498, + 380, + 890, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 441, + 633, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 500, + 467, + 697, + 482 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct comprehensive experiments on three datasets. More details about datasets, such as selected subjects and text prompts, can be found in our Supplementary Materials.", + "bbox": [ + 498, + 491, + 890, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MUG facial expression dataset [1] contains 1,009 videos of 52 subjects performing 7 different expressions. We include this dataset to evaluate the performance of models in scenarios with small motion and a simple, unchanged background. To simplify the experiments, we randomly select 5 male and 5 female subjects, and 4 expressions. We use the text prompt templates like \"A woman with the expression of slight {label} on her face.\" to change the expression class label to be text input. Since the expressions shown in the videos of MUG are often not obvious, we add \"slight\" in the text input to avoid large motion.", + "bbox": [ + 496, + 537, + 892, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "UCF101 action recognition dataset [56] contains 13,320 videos from 101 human action classes. We include this dataset to measure performance under complicated motion and complex, changing backgrounds. To simplify the experiments, we select 10 action classes and the first 10 subjects within each class. We use text prompt templates such as \"A person is performing {label}.\" to change the class label to text input.", + "bbox": [ + 496, + 718, + 893, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to the above two datasets, we create an OPEN dataset to assess the model's performance in open-domain TI2V generation. We first utilize ChatGPT [43] to generate 10 text prompts. Subsequently, we employ Stable", + "bbox": [ + 498, + 840, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9020", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/60831c16be81240768b095d1a822dfd337e1bba6492c2354dd01a4e53c9eb508.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison among different methods on multiple datasets for TI2V generation. Columns in each block display the 1st, 6th, 11th, and 16th frames of the output videos, respectively. There are 16 frames with a resolution of $256 \\times 256$ for each video. The given image $x^0$ is highlighted with the red box and the text prompt $y$ is shown under each block." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 305 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0b4b9f03fc32a6c3c1d3bd0718643eb1f7b3d4124f27a4274b3445b93fcc2ed6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMUGUCF101
FVD↓sFVD↓tFVD↓FVD↓tFVD↓
DynamiCrafter [67]1094.721359.86±257.731223.89±105.94589.591540.02±199.59
TI2V-Zero w/o Resample (Ours)339.89443.97±139.10405.22±61.58493.191319.77±283.87
TI2V-Zero w/ Resample (Ours)180.09267.17±74.72252.77±39.02477.191306.75±271.82
", + "bbox": [ + 158, + 369, + 815, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative comparison among different methods on multiple datasets for TI2V generation.", + "bbox": [ + 184, + 450, + 782, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Diffusion 1.5 [47] to synthesize 100 images from each text prompt, generating a total of 1,000 starting images and 10 text prompts for evaluating TI2V models.", + "bbox": [ + 75, + 491, + 468, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data Preprocessing. We resize all the videos/images to $256 \\times 256$ resolution. For UCF101, since most of the video frames are not square, we crop the central part of the frames. To obtain ground truth videos for computing metrics, we uniformly sample 16 frames from each video in the datasets to generate the video clips with a fixed length.", + "bbox": [ + 75, + 537, + 468, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. Following prior work [21, 22, 25], we assess the visual quality, temporal coherence, and sample diversity of generated videos using Fréchet Video Distance (FVD) [57]. Similar to Fréchet Inception Distance (FID) [18], which is used for image quality evaluation, FVD utilizes a video classification network I3D [6] pretrained on Kinetics-400 dataset [26] to extract feature representation of real and synthesized videos. Then it calculates the Fréchet distance between the distributions of the real and synthesized video features. To measure how well a generated video aligns with the text prompt $y$ (condition accuracy) and the given image $x_0$ (subject relevance), following [39], we design two variants of FVD, namely text-conditioned FVD (tFVD) and subject-conditioned FVD (sFVD). tFVD and sFVD compare the distance between real and synthesized video feature distributions under the same text $y$ or the same subject image $x_0$ , respectively. We first compute tFVD and sFVD for each condition $y$ and image $x_0$ , then report their mean", + "bbox": [ + 75, + 628, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and variance as final results. In our experiments, we generate 1,000 videos for all the models to estimate the feature distributions. We compute both tFVD and sFVD on the MUG dataset, but for UCF101, we only consider tFVD since it doesn't contain videos of different actions for the same subject. For the OPEN dataset, we only present qualitative results due to the lack of ground truth videos. Unless otherwise specified, all the generated videos are 16 frames (i.e., $M = 15$ ) with resolution $256 \\times 256$ .", + "bbox": [ + 496, + 491, + 890, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 635, + 718, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model Implementation. We take the ModelScopeT2V 1.4.2 [60] as basis and implement our modifications. For text-conditioned generation, we employ classifier-free guidance with $g = 9.0$ in Eq. (5). Determined by our preliminary experiments, we choose 10-step DDIM and 4-step resampling as the default setting for MUG and OPEN, and 50-step DDIM and 2-step resampling for UCF101.", + "bbox": [ + 496, + 657, + 890, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation of SOTA Model. We compare our TI2V-Zero with a state-of-the-art (SOTA) model DynamamiCrafter, a recent open-domain TI2V framework [67]. DynamamiCrafter is based on a large-scale pretrained T2V foundation model VideoCrafter1 [16]. It introduces a learnable projection network to enable image-conditioned generation and then fine-tunes the entire framework. We implement DynamamiCrafter using their provided code with their default settings. For a fair comparison, all the generated videos are", + "bbox": [ + 496, + 765, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9021", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1fe46997d8682b6916475eed46a7cce2607959dffa8f63729ccc7d5ffd8b019a.jpg", + "image_caption": [ + "\"A mesmerizing display of the northern lights in the Arctic.\"", + "Figure 6. Example of long video generation using our TI2V-Zero on the OPEN dataset. The given image $x^0$ is highlighted with a red box, and the text prompt $y$ is shown under the set of frames. There are a total of 128 video frames ( $M = 127$ ), and the synthesized results for every 14 frames are presented." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 444, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "centrally-cropped and resized to $256 \\times 256$ .", + "bbox": [ + 76, + 332, + 367, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Result Analysis", + "text_level": 1, + "bbox": [ + 76, + 359, + 232, + 376 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study. We conduct ablation study of different sampling strategies on MUG. As shown in Tab. 1 and Fig. 4, compared with generating using randomly sampled Gaussian noise, initializing the input noise with DDPM inversion is important for generating temporally continuous videos, improving all of the metrics dramatically. For MUG, increasing the DDIM sampling steps from 10 to 50 does not enhance the video quality but requires more inference time. Thus, we choose 10-step DDIM as the default setting on MUG. As shown in Fig. 4 and Tab. 1, adding resampling helps preserve identity details (e.g., hairstyle and facial appearance), resulting in lower FVD scores. Increasing resampling steps from 2 to 4 further improves FVD scores.", + "bbox": [ + 75, + 383, + 468, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Real/Synthesized Starting Frames. We also explore the effect of video generation starting with real or synthesized frames on UCF101. We initially use the first frame of the real videos to generate videos with our TI2V-Zero, termed TI2V-Zero-Real. Additionally, we utilize the backbone model ModelScopeT2V [60] to generate synthetic videos using the text inputs of UCF101. We then employ TI2V-Zero to create videos from the first frame of the generated fake videos, denoted as TI2V-Zero-Fake. As shown in Tab. 2, [TI2V-Zero-Fake vs. ModelScopeT2V] can achieve better FVD scores than [TI2V-Zero-Real vs. Real Videos]. The reason may be that frames generated by ModelScopeT2V can be considered as in-distribution data since TI2V-Zero is built upon it. We also compare the output video distribution of TI2V-Zero-Fake and ModelScopeT2V with real videos in Tab. 2. Though starting from the same synthesized frames, TI2V-Zero-Fake can generate more realistic videos than the backbone model.", + "bbox": [ + 75, + 582, + 467, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with SOTA Model. We compare our proposed TI2V-Zero with DynamiCrafter [67] in Tab. 3 and Fig. 5. From Fig. 5, one can find that DynamiCrafter", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "struggles to preserve details from the given image, and the motion of its generated videos is also less diverse. Note that DynamiCrafter requires additional fine-tuning to enable TI2V generation. In contrast, without using any fine-tuning or introducing external modules, our proposed TI2V-Zero can precisely start with the given image and output more visually-pleasing results, thus achieving much better FVD scores on both MUG and UCF101 datasets in Tab. 3. The comparison between our TI2V-Zero models with and without using resampling in Fig. 5 and Tab. 3 also demonstrates the effectiveness of using resampling, which can help maintain identity and background details.", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extension to Other Applications. TI2V-Zero can also be extended to other tasks as long as we can construct $\\mathbf{s}_0$ with $K$ images at the beginning. These images can be obtained either from ground truth videos or by applying the repeating operation. Then we can slide $\\mathbf{s}_0$ when generating the subsequent frames. We have applied TI2V-Zero in video infilling (see the last row in Fig. 3), video prediction (see Supplementary Materials), and long video generation (see Fig. 6). As shown in Fig. 6, when generating a 128-frame video on the OPEN dataset, our method can preserve the mountain shape in the background, even at the 71st frame (frame $\\hat{x}^{70}$ ). The generated video examples and additional experimental results are in our Supplementary Materials.", + "bbox": [ + 496, + 279, + 890, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 512, + 617, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a zero-shot text-conditioned image-to-video framework, TI2V-Zero, to generate videos by modulating the sampling process of a pretrained video diffusion model without any optimization or fine-tuning. Comprehensive experiments show that TI2V-Zero can achieve promising performance on multiple datasets.", + "bbox": [ + 496, + 546, + 890, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While showing impressive potential, our proposed TI2V-Zero still has some limitations. First, as TI2V-Zero relies on a pretrained T2V diffusion model, the generation quality of TI2V-Zero is constrained by the capabilities and limitations of the pretrained T2V model. We plan to extend our method to more powerful video diffusion foundation models in the future. Second, our method sometimes generates videos that are blurry or contain flickering artifacts. One possible solution is to apply post-processing methods such as blind video deflickering [30] or image/video deblurring [51] to enhance the quality of final output videos or the newly synthesized frame in each generation. Finally, compared with GAN and standard video diffusion models, our approach is considerably slower because it requires running the entire diffusion process for each frame generation. We will investigate some faster sampling methods [29, 32] to reduce generation time.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9022", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Niki Aifanti, Christos Papachristou, and Anastasios Delopoulos. The mug facial expression database. In 11th International Workshop on Image Analysis for Multimedia Interactive Services WIAMIS 10, pages 1-4. IEEE, 2010. 2, 6", + "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 2", + "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 2", + "[4] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Bjorn Ommer. Understanding object dynamics for interactive image-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5171-5181, 2021. 2", + "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2", + "[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7", + "[7] Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A Efros. Everybody dance now. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5933-5942, 2019. 2", + "[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videographer1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2", + "[9] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3d u-net: learning dense volumetric segmentation from sparse annotation. In International conference on medical image computing and computer-assisted intervention, pages 424-432. Springer, 2016. 4", + "[10] Michael Dorkenwald, Timo Milbich, Andreas Blattmann, Robin Rombach, Konstantinos G Derpanis, and Bjorn Ommer. Stochastic image-to-video synthesis using cinns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3742-3753, 2021. 2", + "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3", + "[12] Tsu-Jui Fu, Licheng Yu, Ning Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. Tell me what happened: Unifying text-guided video completion via multimodal masked video generation. In Proceedings of" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10681-10692, 2023. 1, 2", + "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2", + "[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 1, 2", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3", + "[16] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 2, 7", + "[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2", + "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7", + "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 6", + "[21] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 7", + "[22] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 7", + "[23] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2", + "[24] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. arXiv preprint arXiv:2206.07696, 2022. 6", + "[25] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18219-18228, 2022. 1, 2, 7", + "[26] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola," + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9023", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tim Green, Trevor Back, Paul Natev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 7", + "[27] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 2", + "[28] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[29] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. arXiv preprint arXiv:2106.00132, 2021.8", + "[30] Chenyang Lei, Xuanchi Ren, Zhaoxiang Zhang, and Qifeng Chen. Blind video deflickering by neural filtering with a flawed atlas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10439-10448, 2023. 8", + "[31] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2", + "[32] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927, 2022. 8", + "[33] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2, 6", + "[34] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3667-3676, 2022. 2", + "[35] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2", + "[36] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022.", + "[37] Nithin Gopalakrishnan Nair, Anoop Cherian, Suhas Lohit, Ye Wang, Toshiaki Koike-Akino, Vishal M Patel, and Tim K Marks. Steered diffusion: A generalized framework for plug-and-play conditional image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20850-20860, 2023. 2", + "[38] Haomiao Ni, Yihao Liu, Sharon X Huang, and Yuan Xue. Cross-identity video motion retargeting with joint transformation and synthesis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 412–422, 2023. 2", + "[39] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18444-18455, 2023. 1, 2, 7" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Haomiao Ni, Jiachen Liu, Yuan Xue, and Sharon X Huang. 3d-aware talking-head video motion transfer. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4954-4964, 2024. 2", + "[41] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3", + "[42] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2", + "[43] OpenAI. Openai: Introducing chatgpt. URL https://openai.com/blog/chatgpt, 2022.6", + "[44] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 2", + "[45] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 2", + "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4", + "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 4, 7", + "[48] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3", + "[49] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2", + "[50] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2", + "[51] Siddhant Sahu, Manoj Kumar Lenka, and Pankaj Kumar Sa. Blind deblurring using deep learning: A survey. arXiv preprint arXiv:1907.10128, 2019. 8", + "[52] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m:" + ], + "bbox": [ + 501, + 92, + 892, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9024", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2", + "[53] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2", + "[54] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2, 3", + "[55] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 2, 3", + "[56] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2, 6", + "[57] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7", + "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[59] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 1", + "[60] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2, 3, 7, 8", + "[61] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. arXiv preprint arXiv:1808.06601, 2018. 2", + "[62] Wen Wang, kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023. 2", + "[63] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 1, 2", + "[64] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022. 2", + "[65] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2", + "[66] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2", + "[67] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 1, 2, 7, 8", + "[68] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 2", + "[69] Ceyuan Yang, Zhe Wang, Xinge Zhu, Chen Huang, Jianping Shi, and Dahua Lin. Pose guided human video generation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 201-216, 2018. 2", + "[70] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023.1, 2" + ], + "bbox": [ + 503, + 92, + 890, + 387 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9025", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/17566a41-32ff-45d4-9eb1-0099d95d87b1_model.json b/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/17566a41-32ff-45d4-9eb1-0099d95d87b1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3cb853f893ca361d280e6e0ee663e9d7e41dfe72 --- /dev/null +++ b/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/17566a41-32ff-45d4-9eb1-0099d95d87b1_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.131, + 0.88, + 0.152 + ], + "angle": 0, + "content": "TI2V-Zero: Zero-Shot Image Conditioning for Text-to-Video Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.18, + 0.408, + 0.216 + ], + "angle": 0, + "content": "Haomiao Ni\\(^{1*}\\) Bernhard Egger\\(^{2}\\) Toshiaki Koike-Akino\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.218, + 0.406, + 0.234 + ], + "angle": 0, + "content": "1The Pennsylvania State University, USA" + }, + { + "type": "text", + "bbox": [ + 0.445, + 0.181, + 0.59, + 0.218 + ], + "angle": 0, + "content": "Suhas Lohit3 \nSharon X. Huang1" + }, + { + "type": "text", + "bbox": [ + 0.593, + 0.181, + 0.847, + 0.216 + ], + "angle": 0, + "content": "Anoop Cherian3 Ye Wang3 Tim K. Marks3" + }, + { + "type": "text", + "bbox": [ + 0.435, + 0.219, + 0.813, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\)Friedrich-Alexander-Universität Erlangen-Nürnberg, Germany" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.236, + 0.66, + 0.251 + ], + "angle": 0, + "content": "\\(^{3}\\)Mitsubishi Electric Research Laboratories (MERL), USA" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.255, + 0.825, + 0.287 + ], + "angle": 0, + "content": "\\(^{1}\\{\\mathrm{hfn5052, suh972}\\} @ \\mathrm{psu.edu}\\) \\(^{2}\\)bernhard.egger@fau.de \\(^{3}\\{\\)slohit,cherian,yewang,koike,tmarks\\}@merl.com https://merl.com/demos/TI2V-Zero" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.338 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.474, + 0.763 + ], + "angle": 0, + "content": "Text-conditioned image-to-video generation (TI2V) aims to synthesize a realistic video starting from a given image (e.g., a woman's photo) and a text description (e.g., \"a woman is drinking water\"). Existing TI2V frameworks often require costly training on video-text datasets and specific model designs for text and image conditioning. In this paper, we propose TI2V-Zero, a zero-shot, tuning-free method that empowers a pretrained text-to-video (T2V) diffusion model to be conditioned on a provided image, enabling TI2V generation without any optimization, finetuning, or introducing external modules. Our approach leverages a pretrained T2V diffusion foundation model as the generative prior. To guide video generation with the additional image input, we propose a \"repeat-and-slide\" strategy that modulates the reverse denoising process, allowing the frozen diffusion model to synthesize a video frame-by-frame starting from the provided image. To ensure temporal continuity, we employ a DDPM inversion strategy to initialize Gaussian noise for each newly synthesized frame and a resampling technique to help preserve visual details. We conduct comprehensive experiments on both domain-specific and open-domain datasets, where TI2V-Zero consistently outperforms a recent open-domain TI2V model. Furthermore, we show that TI2V-Zero can seamlessly extend to other tasks such as video infilling and prediction when provided with more images. Its autoregressive design also supports long video generation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.778, + 0.21, + 0.793 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.803, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Image-to-video (I2V) generation is an appealing topic with various applications, including artistic creation, entertainment, and data augmentation for machine learning [39]. Given a single image \\( x^0 \\) and a text prompt \\( y \\), text-conditioned image-to-video (TI2V) generation aims to syn" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.322, + 0.614, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.39, + 0.865, + 0.401 + ], + "angle": 0, + "content": "\"A man with the expression of slight happiness on his face.\"" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.322, + 0.784, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.322, + 0.872, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.404, + 0.612, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.404, + 0.784, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.404, + 0.872, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.471, + 0.766, + 0.482 + ], + "angle": 0, + "content": "\"A person is drumming.\"" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.485, + 0.613, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.485, + 0.785, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.485, + 0.872, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.554, + 0.871, + 0.566 + ], + "angle": 0, + "content": "\"A serene mountain cabin covered in a fresh blanket of snow.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.576, + 0.895, + 0.659 + ], + "angle": 0, + "content": "Figure 1. Examples of generated video frames using our proposed T12V-Zero. The given first image \\( x^0 \\) is highlighted with the red box, and the text condition \\( y \\) is shown under each row of the video. The remaining columns show the 6th, 11th, and 16th frames of the generated output videos. Each generated video has 16 frames with a resolution of \\( 256 \\times 256 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.893, + 0.84 + ], + "angle": 0, + "content": "thesize \\(M\\) new frames to yield a realistic video, \\(\\hat{\\mathbf{x}} = \\langle x^0,\\hat{x}^1,\\dots ,\\hat{x}^M\\rangle\\), starting from the given frame \\(x^0\\) and satisfying the text description \\(y\\). Current TI2V generation methods [59, 63, 70] typically rely on computationally-heavy training on video-text datasets and specific architecture designs to enable text and image conditioning. Some [12, 25] are constrained to specific domains due to the lack of training with large-scale open-domain datasets. Other approaches, such as [14, 67], utilize pretrained foundation models to reduce training costs, but they still need to train additional modules using video data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose TI2V-Zero, which achieves zero-shot TI2V generation using only an open-domain pretrained text-to-video (T2V) latent diffusion model [60]. Here \"zero-shot\" means that when using the diffusion" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.327, + 0.9 + ], + "angle": 0, + "content": "*Work done during an internship at MERL." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9015" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "model (DM) that was trained only for text conditioning, our framework enables image conditioning without any optimization, fine-tuning, or introduction of additional modules. Specifically, we guide the generation process by incorporating the provided image \\( x^0 \\) into the output latent code at each reverse denoising step. To ensure that the temporal attention layers of the pretrained DM focus on information from the given image, we propose a \"repeat-and-slide\" strategy to synthesize the video in a frame-by-frame manner, rather than directly generating the entire video volume. Notably, TI2V-Zero is not trained for the specific domain of the provided image, thus allowing the model to generalize to any image during inference. Additionally, its autoregressive generation makes the synthesis of long videos possible." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.473, + 0.472 + ], + "angle": 0, + "content": "While the standard denoising sampling process starting with randomly initialized Gaussian noise can produce matching semantics, it often results in temporally inconsistent videos. Therefore, we introduce an inversion strategy based on the DDPM [20] forward process, to provide a more suitable initial noise for generating each new frame. We also apply a resampling technique [33] in the video DM to help preserve the generated visual details. Our approach ensures that the network maintains temporal consistency, generating visually convincing videos conditioned on the given starting image (see Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.471, + 0.473, + 0.563 + ], + "angle": 0, + "content": "We conduct extensive experiments on MUG [1], UCF-101 [56], and a new open-domain dataset. In these experiments, TI2V-Zero consistently performs well, outperforming a state-of-the-art model [67] that was based on a video diffusion foundation model [8] and was specifically trained to enable open-domain TI2V generation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.579, + 0.22, + 0.595 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.605, + 0.421, + 0.621 + ], + "angle": 0, + "content": "2.1. Conditional Image-to-Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Conditional video generation aims to synthesize videos guided by user-provided signals. It can be classified according to which type(s) of conditions are given, such as text-to-video (T2V) generation [5, 16, 21, 23, 31, 65], video-to-video (V2V) generation [7, 38, 40, 45, 61, 64], and image-to-video (I2V) generation [4, 10, 25, 34, 39, 69]. Here we discuss previous text-conditioned image-to-video (TI2V) generation methods [12, 14, 22, 44, 63, 70]. Hu et al. [25] introduced MAGE, a TI2V generator that integrates a motion anchor structure to store appearance-motion-aligned representations through three-dimensional axial transformers. Yin et al. [70] proposed DragNUWA, a diffusion-based model capable of generating videos controlled by text, image, and trajectory information with three modules including a trajectory sampler, a multi-scale fusion, and an adaptive training strategy. However, these TI2V frameworks require computationally expensive training on videotext datasets and a particular model design to support text" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "and-image-conditioned training. In contrast, our proposed TI2V-Zero leverages a pretrained T2V diffusion model to achieve zero-shot TI2V generation without additional optimization or fine-tuning, making it suitable for a wide range of applications." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.176, + 0.872, + 0.192 + ], + "angle": 0, + "content": "2.2. Adaptation of Diffusion Foundation Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.895, + 0.654 + ], + "angle": 0, + "content": "Due to the recent successful application of diffusion models (DM) [20, 42, 47, 54, 55] to both image and video generation, visual diffusion foundation models have gained prominence. These include text-to-image (T2I) models such as Imagen [50] and Stable Diffusion [47], as well as text-to-video (T2V) models such as ModelScopeT2V [60] and VideoCrafter1 [8]. These models are trained with large-scale open-domain datasets, often including LAION-400M [52] and WebVid-10M [2]. They have shown immense potential for adapting their acquired knowledge base to address a wide range of downstream tasks, thereby reducing or eliminating the need for extensive labeled data. For example, previous works have explored the application of large T2I models to personalized image generation [13, 49], image editing [17, 33, 35-37], image segmentation [3, 68], video editing [45, 62], and video generation [14, 27, 53, 66]. In contrast to T2I models, there are fewer works on the adaptation of large-scale T2V models. Xing et al. [67] proposed DynamicCrafter for open-domain TI2V generation by adapting a T2V foundation model [8]. To control the generative process, they first employed a learnable image encoding network to project the given image into a text-aligned image embedding space. Subsequently, they utilized dual cross-attention layers to fuse text and image information and also concatenated the image with the initial noise to provide the video DM with more precise image details. In contrast, in this paper we explore how to inject the provided image to guide the DM sampling process based solely on the pretrained T2V model itself, with no additional training for the new TI2V task." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.665, + 0.637, + 0.682 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Given one starting image \\( x^0 \\) and text \\( y \\), let \\( \\mathbf{x} = \\langle x^0, x^1, \\ldots, x^M \\rangle \\) represent a real video corresponding to text \\( y \\). The objective of text-conditioned image-to-video (TI2V) generation is to synthesize a video \\( \\hat{\\mathbf{x}} = \\langle x^0, \\hat{x}^1, \\ldots, \\hat{x}^M \\rangle \\), such that the conditional distribution of \\( \\hat{\\mathbf{x}} \\) given \\( x^0 \\) and \\( y \\) is identical to the conditional distribution of \\( \\mathbf{x} \\) given \\( x^0 \\) and \\( y \\), i.e., \\( p(\\hat{\\mathbf{x}} | x^0, y) = p(\\mathbf{x} | x^0, y) \\). Our proposed TI2V-Zero can be built on a pretrained T2V diffusion model with a 3D-UNet-based denoising network. Here we choose ModelScopeT2V [60] as backbone due to its promising open-domain T2V generation ability. Below, we first introduce preliminaries about diffusion models, then introduce the architecture of the pretrained T2V model, and finally present the details of our TI2V-Zero." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "9016" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.296, + 0.895, + 0.396 + ], + "angle": 0, + "content": "Figure 2. Illustration of the process of applying TI2V-Zero to generate the new frame \\(\\hat{x}^{i+1}\\), given the starting image \\(x^0\\) and text \\(y\\). TI2V-Zero is built upon a frozen pretrained T2V diffusion model, including frame encoder \\(\\mathcal{E}\\), frame decoder \\(\\mathcal{D}\\), and the denoising U-Net \\(\\epsilon_{\\theta}\\). At the beginning of generation (\\(i=0\\)), we encode \\(x^0\\) as \\(z^0\\) and repeat it \\(K\\) times to form the queue \\(\\mathbf{s}_0\\). We then apply DDPM-based inversion to \\(\\mathbf{s}_0\\) to produce the initial Gaussian noise \\(\\hat{\\mathbf{z}}_T\\). Subsequently, in each reverse denoising step using U-Net \\(\\epsilon_{\\theta}\\), we keep replacing the first \\(K\\) frames of \\(\\hat{\\mathbf{z}}_t\\) with the noisy latent code \\(\\mathbf{s}_t\\) derived from \\(\\mathbf{s}_0\\). Resampling is also applied within each step to improve motion coherence. We finally decode the final frame of the clean latent code \\(\\hat{\\mathbf{z}}_0\\) as the new synthesized frame \\(\\hat{x}^{i+1}\\). To compute the new \\(\\mathbf{s}_0\\) for the next iteration of generation (\\(i>0\\)), we perform a sliding operation by dequeuing \\(s_0^0\\) and enqueuing \\(\\hat{z}_0^K\\) within \\(\\mathbf{s}_0\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.42, + 0.362, + 0.435 + ], + "angle": 0, + "content": "3.1. Preliminaries: Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.47, + 0.554 + ], + "angle": 0, + "content": "Diffusion Models (DM) [20, 54, 55] are probabilistic models designed to learn a data distribution. Here we introduce the fundamental concepts of Denoising Diffusion Probabilistic Models (DDPM). Given a sample from the data distribution \\(\\mathbf{z}_0 \\sim q(\\mathbf{z}_0)\\), the forward diffusion process of a DM produces a Markov chain \\(\\mathbf{z}_1, \\ldots, \\mathbf{z}_T\\) by iteratively adding Gaussian noise to \\(\\mathbf{z}_0\\) according to a variance schedule \\(\\beta_1, \\ldots, \\beta_T\\), that is:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.557, + 0.469, + 0.573 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {z} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.47, + 0.673 + ], + "angle": 0, + "content": "where variances \\(\\beta_{t}\\) are constant. When the \\(\\beta_{t}\\) are small, the posterior \\(q(\\mathbf{z}_{t - 1}|\\mathbf{z}_t)\\) can be well approximated by a diagonal Gaussian [41, 54]. Furthermore, if the length of the chain, denoted by \\(T\\), is sufficiently large, \\(\\mathbf{z}_T\\) can be well approximated by a standard Gaussian distribution \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). These suggest that the true posterior \\(q(\\mathbf{z}_{t - 1}|\\mathbf{z}_t)\\) can be estimated by \\(p_{\\theta}(\\mathbf{z}_{t - 1}|\\mathbf{z}_t)\\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.675, + 0.469, + 0.691 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {z} _ {t - 1} \\mid \\mathbf {z} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {z} _ {t}\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.693, + 0.47, + 0.835 + ], + "angle": 0, + "content": "where variances \\(\\sigma_t\\) are also constants. The reverse denoising process in the DM (also termed sampling) then generates samples \\(\\mathbf{z}_0 \\sim p_{\\theta}(\\mathbf{z}_0)\\) by starting with Gaussian noise \\(\\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) and gradually reducing noise in a Markov chain \\(\\mathbf{z}_{T-1}, \\mathbf{z}_{T-2}, \\ldots, \\mathbf{z}_0\\) using a learned \\(p_{\\theta}(\\mathbf{z}_{t-1}|\\mathbf{z}_t)\\). To learn \\(p_{\\theta}(\\mathbf{z}_{t-1}|\\mathbf{z}_t)\\), Gaussian noise \\(\\epsilon\\) is first added to \\(\\mathbf{z}_0\\) to generate samples \\(\\mathbf{z}_t\\). Utilizing the independence property of the noise added at each forward step in Eq. (1), we can calculate the total noise variance as \\(\\bar{\\alpha}_t = \\prod_{i=0}^t (1 - \\beta_i)\\) and transform \\(\\mathbf{z}_0\\) to \\(\\mathbf{z}_t\\) in a single step:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.837, + 0.469, + 0.853 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {z} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.855, + 0.469, + 0.883 + ], + "angle": 0, + "content": "Then a model \\(\\epsilon_{\\theta}\\) is trained to predict \\(\\epsilon\\) using the following mean-squared error loss:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.885, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\nL = \\mathbb {E} _ {t \\sim \\mathcal {U} (1, T), \\mathbf {z} _ {0} \\sim q (\\mathbf {z} _ {0}), \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t) \\| ^ {2} \\right] ， \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.892, + 0.588 + ], + "angle": 0, + "content": "where diffusion step \\( t \\) is uniformly sampled from \\( \\{1, \\dots, T\\} \\). Then \\( \\mu_{\\theta}(\\mathbf{z}_t) \\) in Eq. (2) can be derived from \\( \\epsilon_{\\theta}(\\mathbf{z}_t, t) \\) to model \\( p_{\\theta}(\\mathbf{z}_{t-1}|\\mathbf{z}_t) \\) [20]. The denoising model \\( \\epsilon_{\\theta} \\) is implemented using a time-conditioned U-Net [48] with residual blocks [15] and self-attention layers [58]. Diffusion step \\( t \\) is specified to \\( \\epsilon_{\\theta} \\) by the sinusoidal position embedding [58]. Conditional generation that samples \\( \\mathbf{z}_0 \\sim p_{\\theta}(\\mathbf{z}_0|y) \\) can be achieved by learning a \\( y \\)-conditioned model \\( \\epsilon_{\\theta}(\\mathbf{z}_t, t, y) \\) [41, 47] with classifier-free guidance [19]. During training, the condition \\( y \\) in \\( \\epsilon_{\\theta}(\\mathbf{z}_t, t, y) \\) is replaced by a null label \\( \\emptyset \\) with a fixed probability. When sampling, the output is generated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.505, + 0.59, + 0.892, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} _ {\\theta} \\left(\\mathbf {z} _ {t}, t, y\\right) = \\epsilon_ {\\theta} \\left(\\mathbf {z} _ {t}, t, \\emptyset\\right) + g \\cdot \\left(\\epsilon_ {\\theta} \\left(\\mathbf {z} _ {t}, t, y\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {z} _ {t}, t, \\emptyset\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.609, + 0.704, + 0.624 + ], + "angle": 0, + "content": "where \\(g\\) is the guidance scale." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.635, + 0.833, + 0.65 + ], + "angle": 0, + "content": "3.2. Architecture of Pretrained T2V Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.718 + ], + "angle": 0, + "content": "TI2V-Zero can be built upon a pretrained T2V diffusion model with a 3D-UNet-based denoising network. Here we choose ModelScopeT2V [60] as the pretrained model (denoted \\(\\mathcal{M}\\)). We now describe this T2V model in detail." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Structure Overview. Given a text prompt \\(y\\), the T2V model \\(\\mathcal{M}\\) synthesizes a video \\(\\hat{\\mathbf{x}} = \\langle \\hat{x}^0, \\hat{x}^1, \\dots, \\hat{x}^K \\rangle\\) with a pre-defined video of length \\((K + 1)\\) using a latent video diffusion model. Similar to Latent Diffusion Models (LDM) [47], \\(\\mathcal{M}\\) incorporates a frame auto-encoder [11, 28] for the conversion of data between pixel space \\(\\mathcal{X}\\) and latent space \\(\\mathcal{Z}\\) through its encoder \\(\\mathcal{E}\\) and decoder \\(\\mathcal{D}\\). Given the real video \\(\\mathbf{x} = \\langle x^0, x^1, \\dots, x^K \\rangle\\), \\(\\mathcal{M}\\) first utilizes the frame encoder \\(\\mathcal{E}\\) to encode the video \\(\\mathbf{x}\\) as \\(\\mathbf{z} = \\langle z^0, z^1, \\dots, z^K \\rangle\\). Here the sizes of pixel frame \\(x\\) and latent frame \\(z\\) are \\(H_x \\times W_x \\times 3\\) and \\(H_z \\times W_z \\times C_z\\), respectively. To be consistent with the notation used for the DM, we denote the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9017" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.079, + 0.092, + 0.449, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Generation using our TI2V-Zero approach." + }, + { + "type": "algorithm", + "bbox": [ + 0.078, + 0.111, + 0.473, + 0.514 + ], + "angle": 0, + "content": "Input: The starting frame \\(x^0\\); The text prompt \\(y\\); The pretrained T2V Model \\(\\mathcal{M}\\) for generating \\((K + 1)\\)-frame videos, including frame encoder \\(\\mathcal{E}\\) and frame decoder \\(\\mathcal{D}\\), and the DM de-noising networks \\(\\epsilon_{\\theta}\\); The iteration number \\(U\\) for resampling; The parameter \\(M\\) to control the length of the output video. \nOutput: A synthesized video \\(\\hat{\\mathbf{x}}\\) with \\((M + 1)\\) frames. \n1: \\(z^0 \\gets \\mathcal{E}(x^0)\\) // Encode \\(x^0\\) \n2: \\(\\mathbf{s}_0 \\gets \\langle z^0, z^0, \\dots, z^0 \\rangle\\) // Repeat \\(z^0\\) for \\(K\\) times \n3: \\(\\hat{\\mathbf{x}} \\gets \\langle x^0 \\rangle\\) \n4: for \\(i = 1, 2, \\dots, M\\) do \n// Generate one new frame \\(\\hat{x}^i\\) \n5: \\(\\mathbf{s}_T \\sim \\mathcal{N}(\\sqrt{\\bar{\\alpha}_T} \\mathbf{s}_0, (1 - \\bar{\\alpha}_T) \\mathbf{I})\\) // DDPM Inversion \n6: \\(\\hat{z}_T^K \\sim \\mathcal{N}(\\sqrt{\\bar{\\alpha}_T} s_0^{K-1}, (1 - \\bar{\\alpha}_T) \\mathbf{I})\\) \n7: \\(\\hat{\\mathbf{z}}_T \\gets \\mathbf{s}_T \\cup \\hat{z}_T^K\\) // Initialize \\(\\hat{\\mathbf{z}}_T\\) \n8: for \\(t = T-1, \\dots, 2, 1\\) do \n\\(\\mathbf{s}_t \\sim \\mathcal{N}(\\sqrt{\\bar{\\alpha}_t} \\mathbf{s}_0, (1 - \\bar{\\alpha}_t) \\mathbf{I})\\) \nfor \\(u = 1, 2, \\dots, U\\) do \n\\(\\langle \\hat{z}_t^0, \\hat{z}_t^1, \\dots, \\hat{z}_t^{K-1} \\rangle \\gets \\mathbf{s}_t\\) // Replace \n\\(\\hat{\\mathbf{z}}_{t-1} \\sim \\mathcal{N}(\\mu_\\theta (\\hat{\\mathbf{z}}_t, y), \\sigma_t^2 \\mathbf{I})\\) \nif \\(u < U\\) and \\(t > 1\\) then \n\\(\\hat{\\mathbf{z}}_t \\sim \\mathcal{N}(\\sqrt{1 - \\beta_t} \\hat{\\mathbf{z}}_{t-1}, \\beta_t \\mathbf{I})\\) // Resample \nend if \nend for \nend for \n\\(\\mathbf{s}_0 \\gets \\langle s_0^1, s_0^2, \\dots, s_0^{K-1} \\rangle \\cup \\hat{z}_0^K\\) // Slide \n\\(\\hat{x}^i \\gets D(\\hat{z}_0^K)\\) // Decode \\(\\hat{z}_0^K\\) \n\\(\\hat{\\mathbf{x}} \\gets \\hat{\\mathbf{x}} \\cup \\hat{x}^i\\) \nend for \nreturn \\(\\hat{\\mathbf{x}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.715 + ], + "angle": 0, + "content": "clean video latent \\(\\mathbf{z} = \\mathbf{z}_0 = \\langle z_0^0, z_0^1, \\ldots, z_0^K \\rangle\\). \\(\\mathcal{M}\\) then learns a DM on the latent space \\(\\mathcal{Z}\\) through a 3D denoising U-Net \\(\\epsilon_{\\theta}\\) [9]. Let \\(\\mathbf{z}_t = \\langle z_t^0, z_t^1, \\ldots, z_t^K \\rangle\\) represent the latent sequence that results from adding noise over \\(t\\) steps to the original latent sequence \\(\\mathbf{z}_0\\). When training, the forward diffusion process of a DM transforms the initial latent sequence \\(\\mathbf{z}_0\\) into \\(\\mathbf{z}_T\\) by iteratively adding Gaussian noise \\(\\epsilon\\) for \\(T\\) steps. During inference, denoising U-Net \\(\\epsilon_{\\theta}\\) predicts the added noise at each step, enabling the generation of the clean latent sequence \\(\\hat{\\mathbf{z}}_0 = \\langle \\hat{z}_0^0, \\hat{z}_0^1, \\ldots, \\hat{z}_0^K \\rangle\\) starting from randomly sampled Gaussian noise \\(\\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.47, + 0.838 + ], + "angle": 0, + "content": "Text Conditioning Mechanism. \\(\\mathcal{M}\\) employs a cross-attention mechanism [47] to incorporate text information into the generative process as guidance. Specifically, \\(\\mathcal{M}\\) uses a pretrained CLIP model [46] to encode the prompt \\(y\\) as the text embedding \\(e\\). The embedding \\(e\\) is later used as the key and value in the multi-head attention layer within the spatial attention blocks, thus enabling the integration of text features with the intermediate U-Net features in \\(\\epsilon_{\\theta}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Denoising U-Net. The denoising U-Net \\(\\epsilon_{\\theta}\\) includes four key building blocks: the initial block, the downsampling block, the spatio-temporal block, and the upsampling block. The initial block transfers the input into the embedding" + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.089, + 0.846, + 0.099 + ], + "angle": 0, + "content": "Video Time" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.099, + 0.877, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.414, + 0.895, + 0.54 + ], + "angle": 0, + "content": "Figure 3. Illustration of the motivation behind our framework. We explore the application of a replacing-based baseline approach (rows 2-4, labeled \"Replacing\") and our TI2V-Zero (rows 5-6, labeled \"TI2V-Zero\") in various video generation tasks. The given real frames for each task are highlighted by red boxes and the text input is shown under the block. The replacing-based approach is only effective at predicting a single frame when all the other frames in the video are provided, while TI2V-Zero generates temporally coherent videos for both the TI2V and video infilling tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.555, + 0.892, + 0.648 + ], + "angle": 0, + "content": "space, while the downsampling and upsampling blocks are responsible for spatially downsampling and upsampling the feature maps. The spatio-temporal block is designed to capture spatial and temporal dependencies in the latent space, which comprises 2D spatial convolution, 1D temporal convolution, 2D spatial attention, and 1D temporal attention." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.659, + 0.664, + 0.673 + ], + "angle": 0, + "content": "3.3. Our Framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.892, + 0.789 + ], + "angle": 0, + "content": "Leveraging the pretrained T2V foundation model \\(\\mathcal{M}\\), we first propose a straightforward replacing-based baseline for adapting \\(\\mathcal{M}\\) to TI2V generation. We then analyze the possible reasons why it fails and introduce our TI2V-Zero framework, which includes a repeat-and-slide strategy, DDPM-based inversion, and resampling. Figure 2 and Algorithm 1 demonstrate the inference process of TI2V-Zero." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.79, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Replacing-based Baseline. We assume that the pretrained model \\(\\mathcal{M}\\) is designed to generate the video with a fixed length of \\((K + 1)\\). So we first consider synthesizing videos with that same length \\((K + 1)\\), i.e., \\(M = K\\). Since the DM process operates within the latent space \\(\\mathcal{Z}\\), we use the encoder \\(\\mathcal{E}\\) to map the given starting frame \\(x^0\\) into the latent representation \\(z^0\\). Additionally, we denote \\(z^0 = z_0^0\\) to specify that the latent is clean and corresponds" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9018" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "to diffusion step 0 of the DM. Note that each reverse denoising step in Eq. (2) from \\(\\hat{\\mathbf{z}}_t\\) to \\(\\hat{\\mathbf{z}}_{t - 1}\\) depends solely on \\(\\hat{\\mathbf{z}}_t = \\langle \\hat{z}_t^0,\\hat{z}_t^1,\\dots ,\\hat{z}_t^K\\rangle\\) . To ensure that the first frame of the final synthesized clean video latent \\(\\hat{\\mathbf{z}}_0 = \\langle \\hat{z}_0^0,\\hat{z}_0^1,\\dots ,\\hat{z}_0^K\\rangle\\) at step 0 matches the provided image latent, i.e., \\(\\hat{z}_0^0 = z_0^0\\) we can modify the first generated latent \\(\\hat{z}_t^0\\) of \\(\\hat{\\mathbf{z}}_t\\) at each reverse step, as long as the signal-to-noise ratio of each frame latent in \\(\\hat{\\mathbf{z}}_t\\) remains consistent. Using Eq. (3), we can add \\(t\\) steps of noise to the provided image latent \\(z_0^0\\) , allowing us to sample \\(z_{t}^{0}\\) through a single-step calculation. By replacing the first generated latent \\(\\hat{z}_t^0\\) with the noisy image latent \\(z_{t}^{0}\\) at each reverse denoising step, we might expect that the video generation process can be guided by \\(z_0^0\\) with the following expressions defined for each reverse step:" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.293, + 0.47, + 0.31 + ], + "angle": 0, + "content": "\\[\nz _ {t} ^ {0} \\sim \\mathcal {N} \\left(\\sqrt {\\bar {\\alpha}} _ {t} z _ {0} ^ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {6a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.313, + 0.469, + 0.329 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {t} ^ {0} \\leftarrow z _ {t} ^ {0}, \\tag {6b}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.333, + 0.469, + 0.35 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {z}} _ {t - 1} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} \\left(\\hat {\\mathbf {z}} _ {t}, y\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right). \\tag {6c}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.508 + ], + "angle": 0, + "content": "Specifically, in each reverse step from \\(\\hat{\\mathbf{z}}_t\\) to \\(\\hat{\\mathbf{z}}_{t - 1}\\), as shown in Eq. (6a), we first compute the noisy latent \\(z_{t}^{0}\\) by adding Gaussian noise to the given image latent \\(z_0^0\\) over \\(t\\) steps. Then, we replace the first latent \\(\\hat{z}_t^0\\) of \\(\\hat{\\mathbf{z}}_t\\) with \\(z_{t}^{0}\\) in Eq. (6b) to incorporate the provided image into the generation process. Finally, in Eq. (6c), we pass \\(\\hat{\\mathbf{z}}_t\\) through the denoising network to generate \\(\\hat{\\mathbf{z}}_{t - 1}\\), where the text \\(y\\) is integrated by classifier-free guidance (Eq. (5)). After \\(T\\) iterations, the final clean latent \\(\\hat{\\mathbf{z}}_0\\) at diffusion step 0 can be mapped back into the image space \\(\\mathcal{X}\\) using the decoder \\(\\mathcal{D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.47, + 0.629 + ], + "angle": 0, + "content": "Using this replacing-based baseline, we might expect that the temporal attention layers in \\(\\epsilon_{\\theta}\\) can utilize the context provided by the first frame latent \\(\\hat{z}_t^0\\) to generate the subsequent frame latents in a manner that harmonizes with \\(\\hat{z}_t^0\\). However, as shown in Fig. 3, row 2, this replacing-based approach fails to produce a video that is temporally consistent with the first image. The generated frames are consistent with each other, but not with the provided first frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.471, + 0.901 + ], + "angle": 0, + "content": "To analyze possible reasons for failure, we apply this baseline to a simpler video infilling task, where every other frame is provided and the model needs to predict the interspersed frames. In this case, the baseline replaces the generated frame latents at positions corresponding to real frames with noisy provided-frame latents in each reverse step. The resulting video, in Fig. 3, row 3, looks like a combination of two independent videos: the generated (even) frames are consistent with each other but not with the provided (odd) frames. We speculate that this may result from the intrinsic dissimilarity between frame latents derived from the given real images and those sampled from \\(\\epsilon_{\\theta}\\). Thus, the temporal attention values between frame latents sampled in the same way (both from the given images or both from \\(\\epsilon_{\\theta}\\)) will be higher, while the attention values between frame latents sampled in different ways (one from the given image and the other from \\(\\epsilon_{\\theta}\\)) will be lower. Therefore, the temporal attention layers of \\(\\mathcal{M}\\) tend to utilize the information from latents" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.091, + 0.871, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.44, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Figure 4. Qualitative ablation study comparing different sampling strategies for our TI2V-Zero on MUG. The first image \\(\\hat{x}^0\\) is highlighted with the red box and text \\(y\\) is shown under the block. The 1st, 6th, 11th, and 16th frames of the videos are shown in each column. The terms Inversion, DDIM, and Resample denote the application of DDPM inversion, the steps using DDIM sampling, and the iteration number using resampling, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.555, + 0.892, + 0.706 + ], + "angle": 0, + "content": "produced by \\(\\epsilon_{\\theta}\\) to synthesize new frames at each reverse step, ignoring the provided frames. We further simplify the task to single-frame prediction, where the model only needs to predict a single frame when all the other frames in the video are given. In this setting, all the frame latents except for the final frame are replaced by noisy provided-frame latents in each reverse step. Thus, temporal attention layers can only use information from the real frames. In this case, Fig. 3, row 4, shows that the baseline can now generate a final frame that is consistent with the previous frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.706, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Repeat-and-Slide Strategy. Inspired by the observation in Fig. 3, to guarantee that the temporal attention layers of \\(\\mathcal{M}\\) depend solely on the given image, we make two major changes to the proposed replacing-based baseline: (1) instead of using \\(\\mathcal{M}\\) to directly synthesize the entire \\((K + 1)\\)-frame video, we switch to a frame-by-frame generation approach, i.e., we generate only one new frame latent in each complete DM sampling process; (2) for each sampling process generating the new frame latent, we ensure that only one frame latent is produced from \\(\\epsilon_{\\theta}\\), while the other \\(K\\) frame latents are derived from the given real image and previously synthesized frames, thereby forcing temporal attention layers to only use the information from these frame latents. Specifically, we construct a queue of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9019" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.09, + 0.473, + 0.358 + ], + "angle": 0, + "content": "\\(K\\) frame latents, denoted as \\(\\mathbf{s}_0 = \\langle s_0^0,s_0^1,\\dots ,s_0^{K - 1}\\rangle\\). We also define \\(\\mathbf{s}_t = \\langle s_t^0,s_t^1,\\dots ,s_t^{K - 1}\\rangle\\), which is obtained by adding \\(t\\) steps of Gaussian noise to the clean \\(\\mathbf{s}_0\\). Similar to our replacing-based baseline in the single-frame prediction task, in each reverse step from \\(\\hat{\\mathbf{z}}_t\\) to \\(\\hat{\\mathbf{z}}_{t - 1}\\), we replace the first \\(K\\) frame latents in \\(\\hat{\\mathbf{z}}_t\\) by \\(\\mathbf{s}_t\\). Consequently, the temporal attention layers have to utilize information from \\(\\mathbf{s}_0\\) to synthesize the new frame's latent, \\(\\hat{z}_0^K\\). Considering that only one starting image latent \\(z^0\\) is provided, we propose a \"repeat-and-slide\" strategy to construct \\(\\mathbf{s}_0\\). At the beginning of video generation, we repeat \\(z^0\\) for \\(K\\) frames to form \\(\\mathbf{s}_0\\), and gradually perform a sliding operation within the queue \\(\\mathbf{s}_0\\) by dequeuing the first frame latent \\(s_0^0\\) and enqueuing the newly generated latent \\(\\hat{z}_0^K\\) after each complete DM sampling process. Note that though the initial \\(\\mathbf{s}_0\\) is created by repeating \\(z^0\\), the noise added to get \\(\\mathbf{s}_t\\) is different for each frame's latent in \\(\\mathbf{s}_t\\), thus ensuring diversity. The following expressions define one reverse step in the DM sampling process:" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.367, + 0.47, + 0.384 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {t} \\sim \\mathcal {N} \\left(\\sqrt {\\bar {\\alpha}} _ {t} \\mathbf {s} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {7a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.386, + 0.469, + 0.403 + ], + "angle": 0, + "content": "\\[\n\\left\\langle \\hat {z} _ {t} ^ {0}, \\hat {z} _ {t} ^ {1}, \\dots , \\hat {z} _ {t} ^ {K - 1} \\right\\rangle \\leftarrow \\mathbf {s} _ {t}, \\tag {7b}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.406, + 0.469, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {z}} _ {t - 1} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} \\left(\\hat {\\mathbf {z}} _ {t}, y\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right). \\tag {7c}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.433, + 0.47, + 0.509 + ], + "angle": 0, + "content": "Specifically, in each reverse denoising step from \\(\\hat{\\mathbf{z}}_t\\) to \\(\\hat{\\mathbf{z}}_{t - 1}\\), we first add \\(t\\) steps of Gaussian noise to the queue \\(\\mathbf{s}_0\\) to yield \\(\\mathbf{s}_t\\) in Eq. (7a). Subsequently, we replace the previous \\(K\\) frames of \\(\\hat{\\mathbf{z}}_t\\) with \\(\\mathbf{s}_t\\) in Eq. (7b) and input \\(\\hat{\\mathbf{z}}_t\\) to the denoising network to produce the less noisy latent \\(\\hat{\\mathbf{z}}_{t - 1}\\) (Eq. (7c))." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.509, + 0.47, + 0.583 + ], + "angle": 0, + "content": "With the repeat-and-slide strategy, model \\(\\mathcal{M}\\) is tasked with predicting only one new frame, while the preceding \\(K\\) frames are incorporated into the reverse process to ensure that the temporal attention layers depend solely on information derived from the provided image." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.584, + 0.471, + 0.81 + ], + "angle": 0, + "content": "DDPM-based Inversion. Though the DM sampling process starting with randomly sampled Gaussian noise produces matching semantics, the generated video is often temporally inconsistent (Fig. 4, row 2). To provide initial noise that can produce more temporally consistent results, we introduce an inversion strategy based on the DDPM [20] forward process when generating the new frame latent. Specifically, at the beginning of each DM sampling process to synthesize the new frame latent \\(\\hat{z}_0^K\\), instead of starting with the \\(\\hat{\\mathbf{z}}_T\\) randomly sampled from \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\), we add \\(T\\) full steps of Gaussian noise to \\(\\mathbf{s}_0\\) to obtain \\(\\mathbf{s}_T\\) using Eq. (3). Note that \\(\\hat{\\mathbf{z}}\\) has \\(K + 1\\) frames, while \\(\\mathbf{s}\\) has \\(K\\) frames. We then use \\(\\mathbf{s}_T\\) to initialize the first \\(K\\) frames of \\(\\hat{\\mathbf{z}}_T\\). We copy the last frame \\(s_T^{K - 1}\\) of \\(\\mathbf{s}_T\\) to initialize the final frame \\(\\hat{z}_T^K\\), as the \\((K - 1)\\mathrm{th}\\) frame is the closest to the \\(K\\)th frame." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Resampling. Similar to [24, 33], we further apply a resampling technique, which was initially designed for the image inpainting task, to the video DM to enhance motion coherence. Particularly, after performing a one-step denoising operation in the reversed process, we add one-step noise again to revert the latent. This procedure is repeated mul" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.894, + 0.158 + ], + "angle": 0, + "content": "
InversionDDIMResampleFVD↓sFVD↓tFVD↓
X1001656.372074.77±411.741798.05±235.34
✓100339.89443.97±139.10405.22±61.58
✓500463.55581.32±234.09535.06±85.27
✓102207.62299.14±87.24278.73±47.84
✓104180.09267.17±74.72252.77±39.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.167, + 0.894, + 0.238 + ], + "angle": 0, + "content": "Table 1. Quantitative ablation study comparing different sampling strategies for proposed TI2V-Zero on the MUG dataset. Inversion, DDIM, and Resample denote the application of DDPM-based inversion, the steps using DDIM sampling, and the iteration number using resampling, respectively." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.251, + 0.875, + 0.315 + ], + "angle": 0, + "content": "
Distributions for ComparisonFVD↓tFVD↓
TI2V-Zero-Fake vs. ModelScopeT2V366.41921.31±251.85
TI2V-Zero-Real vs. Real Videos477.191306.75±271.82
ModelScopeT2V vs. Real Videos985.822264.08±501.28
TI2V-Zero-Fake vs. Real Videos937.112177.70±436.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.324, + 0.893, + 0.366 + ], + "angle": 0, + "content": "Table 2. Result analysis of TI2V-Zero starting from the real (i.e., TI2V-Zero-Real) or synthesized frames (i.e., TI2V-Zero-Fake) on the UCF101 dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.381, + 0.892, + 0.427 + ], + "angle": 0, + "content": "tiple times for each diffusion step, ensuring harmonization between the predicted and conditioning frame latents (see Algorithm 1 for details)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.443, + 0.634, + 0.46 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.468, + 0.699, + 0.483 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.492, + 0.892, + 0.537 + ], + "angle": 0, + "content": "We conduct comprehensive experiments on three datasets. More details about datasets, such as selected subjects and text prompts, can be found in our Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.893, + 0.718 + ], + "angle": 0, + "content": "MUG facial expression dataset [1] contains 1,009 videos of 52 subjects performing 7 different expressions. We include this dataset to evaluate the performance of models in scenarios with small motion and a simple, unchanged background. To simplify the experiments, we randomly select 5 male and 5 female subjects, and 4 expressions. We use the text prompt templates like \"A woman with the expression of slight {label} on her face.\" to change the expression class label to be text input. Since the expressions shown in the videos of MUG are often not obvious, we add \"slight\" in the text input to avoid large motion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.894, + 0.84 + ], + "angle": 0, + "content": "UCF101 action recognition dataset [56] contains 13,320 videos from 101 human action classes. We include this dataset to measure performance under complicated motion and complex, changing backgrounds. To simplify the experiments, we select 10 action classes and the first 10 subjects within each class. We use text prompt templates such as \"A person is performing {label}.\" to change the class label to text input." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.841, + 0.894, + 0.902 + ], + "angle": 0, + "content": "In addition to the above two datasets, we create an OPEN dataset to assess the model's performance in open-domain TI2V generation. We first utilize ChatGPT [43] to generate 10 text prompts. Subsequently, we employ Stable" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9020" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.315, + 0.895, + 0.358 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison among different methods on multiple datasets for TI2V generation. Columns in each block display the 1st, 6th, 11th, and 16th frames of the output videos, respectively. There are 16 frames with a resolution of \\(256 \\times 256\\) for each video. The given image \\(x^0\\) is highlighted with the red box and the text prompt \\(y\\) is shown under each block." + }, + { + "type": "table", + "bbox": [ + 0.159, + 0.371, + 0.816, + 0.442 + ], + "angle": 0, + "content": "
ModelMUGUCF101
FVD↓sFVD↓tFVD↓FVD↓tFVD↓
DynamiCrafter [67]1094.721359.86±257.731223.89±105.94589.591540.02±199.59
TI2V-Zero w/o Resample (Ours)339.89443.97±139.10405.22±61.58493.191319.77±283.87
TI2V-Zero w/ Resample (Ours)180.09267.17±74.72252.77±39.02477.191306.75±271.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.185, + 0.451, + 0.784, + 0.466 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison among different methods on multiple datasets for TI2V generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.538 + ], + "angle": 0, + "content": "Diffusion 1.5 [47] to synthesize 100 images from each text prompt, generating a total of 1,000 starting images and 10 text prompts for evaluating TI2V models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.629 + ], + "angle": 0, + "content": "Data Preprocessing. We resize all the videos/images to \\(256 \\times 256\\) resolution. For UCF101, since most of the video frames are not square, we crop the central part of the frames. To obtain ground truth videos for computing metrics, we uniformly sample 16 frames from each video in the datasets to generate the video clips with a fixed length." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Metrics. Following prior work [21, 22, 25], we assess the visual quality, temporal coherence, and sample diversity of generated videos using Fréchet Video Distance (FVD) [57]. Similar to Fréchet Inception Distance (FID) [18], which is used for image quality evaluation, FVD utilizes a video classification network I3D [6] pretrained on Kinetics-400 dataset [26] to extract feature representation of real and synthesized videos. Then it calculates the Fréchet distance between the distributions of the real and synthesized video features. To measure how well a generated video aligns with the text prompt \\( y \\) (condition accuracy) and the given image \\( x_0 \\) (subject relevance), following [39], we design two variants of FVD, namely text-conditioned FVD (tFVD) and subject-conditioned FVD (sFVD). tFVD and sFVD compare the distance between real and synthesized video feature distributions under the same text \\( y \\) or the same subject image \\( x_0 \\), respectively. We first compute tFVD and sFVD for each condition \\( y \\) and image \\( x_0 \\), then report their mean" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.628 + ], + "angle": 0, + "content": "and variance as final results. In our experiments, we generate 1,000 videos for all the models to estimate the feature distributions. We compute both tFVD and sFVD on the MUG dataset, but for UCF101, we only consider tFVD since it doesn't contain videos of different actions for the same subject. For the OPEN dataset, we only present qualitative results due to the lack of ground truth videos. Unless otherwise specified, all the generated videos are 16 frames (i.e., \\(M = 15\\)) with resolution \\(256 \\times 256\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.636, + 0.719, + 0.652 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Model Implementation. We take the ModelScopeT2V 1.4.2 [60] as basis and implement our modifications. For text-conditioned generation, we employ classifier-free guidance with \\( g = 9.0 \\) in Eq. (5). Determined by our preliminary experiments, we choose 10-step DDIM and 4-step resampling as the default setting for MUG and OPEN, and 50-step DDIM and 2-step resampling for UCF101." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Implementation of SOTA Model. We compare our TI2V-Zero with a state-of-the-art (SOTA) model DynamamiCrafter, a recent open-domain TI2V framework [67]. DynamamiCrafter is based on a large-scale pretrained T2V foundation model VideoCrafter1 [16]. It introduces a learnable projection network to enable image-conditioned generation and then fine-tunes the entire framework. We implement DynamamiCrafter using their provided code with their default settings. For a fair comparison, all the generated videos are" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "9021" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.446, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.221, + 0.441, + 0.233 + ], + "angle": 0, + "content": "\"A mesmerizing display of the northern lights in the Arctic.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.246, + 0.47, + 0.318 + ], + "angle": 0, + "content": "Figure 6. Example of long video generation using our TI2V-Zero on the OPEN dataset. The given image \\( x^0 \\) is highlighted with a red box, and the text prompt \\( y \\) is shown under the set of frames. There are a total of 128 video frames (\\( M = 127 \\)), and the synthesized results for every 14 frames are presented." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.333, + 0.368, + 0.348 + ], + "angle": 0, + "content": "centrally-cropped and resized to \\(256 \\times 256\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.361, + 0.233, + 0.377 + ], + "angle": 0, + "content": "4.3. Result Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.47, + 0.581 + ], + "angle": 0, + "content": "Ablation Study. We conduct ablation study of different sampling strategies on MUG. As shown in Tab. 1 and Fig. 4, compared with generating using randomly sampled Gaussian noise, initializing the input noise with DDPM inversion is important for generating temporally continuous videos, improving all of the metrics dramatically. For MUG, increasing the DDIM sampling steps from 10 to 50 does not enhance the video quality but requires more inference time. Thus, we choose 10-step DDIM as the default setting on MUG. As shown in Fig. 4 and Tab. 1, adding resampling helps preserve identity details (e.g., hairstyle and facial appearance), resulting in lower FVD scores. Increasing resampling steps from 2 to 4 further improves FVD scores." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.468, + 0.853 + ], + "angle": 0, + "content": "Effect of Real/Synthesized Starting Frames. We also explore the effect of video generation starting with real or synthesized frames on UCF101. We initially use the first frame of the real videos to generate videos with our TI2V-Zero, termed TI2V-Zero-Real. Additionally, we utilize the backbone model ModelScopeT2V [60] to generate synthetic videos using the text inputs of UCF101. We then employ TI2V-Zero to create videos from the first frame of the generated fake videos, denoted as TI2V-Zero-Fake. As shown in Tab. 2, [TI2V-Zero-Fake vs. ModelScopeT2V] can achieve better FVD scores than [TI2V-Zero-Real vs. Real Videos]. The reason may be that frames generated by ModelScopeT2V can be considered as in-distribution data since TI2V-Zero is built upon it. We also compare the output video distribution of TI2V-Zero-Fake and ModelScopeT2V with real videos in Tab. 2. Though starting from the same synthesized frames, TI2V-Zero-Fake can generate more realistic videos than the backbone model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Comparison with SOTA Model. We compare our proposed TI2V-Zero with DynamiCrafter [67] in Tab. 3 and Fig. 5. From Fig. 5, one can find that DynamiCrafter" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "struggles to preserve details from the given image, and the motion of its generated videos is also less diverse. Note that DynamiCrafter requires additional fine-tuning to enable TI2V generation. In contrast, without using any fine-tuning or introducing external modules, our proposed TI2V-Zero can precisely start with the given image and output more visually-pleasing results, thus achieving much better FVD scores on both MUG and UCF101 datasets in Tab. 3. The comparison between our TI2V-Zero models with and without using resampling in Fig. 5 and Tab. 3 also demonstrates the effectiveness of using resampling, which can help maintain identity and background details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.892, + 0.477 + ], + "angle": 0, + "content": "Extension to Other Applications. TI2V-Zero can also be extended to other tasks as long as we can construct \\(\\mathbf{s}_0\\) with \\(K\\) images at the beginning. These images can be obtained either from ground truth videos or by applying the repeating operation. Then we can slide \\(\\mathbf{s}_0\\) when generating the subsequent frames. We have applied TI2V-Zero in video infilling (see the last row in Fig. 3), video prediction (see Supplementary Materials), and long video generation (see Fig. 6). As shown in Fig. 6, when generating a 128-frame video on the OPEN dataset, our method can preserve the mountain shape in the background, even at the 71st frame (frame \\(\\hat{x}^{70}\\)). The generated video examples and additional experimental results are in our Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.618, + 0.529 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.637 + ], + "angle": 0, + "content": "In this paper, we propose a zero-shot text-conditioned image-to-video framework, TI2V-Zero, to generate videos by modulating the sampling process of a pretrained video diffusion model without any optimization or fine-tuning. Comprehensive experiments show that TI2V-Zero can achieve promising performance on multiple datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "While showing impressive potential, our proposed TI2V-Zero still has some limitations. First, as TI2V-Zero relies on a pretrained T2V diffusion model, the generation quality of TI2V-Zero is constrained by the capabilities and limitations of the pretrained T2V model. We plan to extend our method to more powerful video diffusion foundation models in the future. Second, our method sometimes generates videos that are blurry or contain flickering artifacts. One possible solution is to apply post-processing methods such as blind video deflickering [30] or image/video deblurring [51] to enhance the quality of final output videos or the newly synthesized frame in each generation. Finally, compared with GAN and standard video diffusion models, our approach is considerably slower because it requires running the entire diffusion process for each frame generation. We will investigate some faster sampling methods [29, 32] to reduce generation time." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9022" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.183 + ], + "angle": 0, + "content": "[1] Niki Aifanti, Christos Papachristou, and Anastasios Delopoulos. The mug facial expression database. In 11th International Workshop on Image Analysis for Multimedia Interactive Services WIAMIS 10, pages 1-4. IEEE, 2010. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Bjorn Ommer. Understanding object dynamics for interactive image-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5171-5181, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.45 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.471, + 0.508 + ], + "angle": 0, + "content": "[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[7] Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A Efros. Everybody dance now. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5933-5942, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.634 + ], + "angle": 0, + "content": "[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videographer1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.471, + 0.717 + ], + "angle": 0, + "content": "[9] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3d u-net: learning dense volumetric segmentation from sparse annotation. In International conference on medical image computing and computer-assisted intervention, pages 424-432. Springer, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[10] Michael Dorkenwald, Timo Milbich, Andreas Blattmann, Robin Rombach, Konstantinos G Derpanis, and Bjorn Ommer. Stochastic image-to-video synthesis using cinns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3742-3753, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[12] Tsu-Jui Fu, Licheng Yu, Ning Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. Tell me what happened: Unifying text-guided video completion via multimodal masked video generation. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10681-10692, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[16] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.532, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[21] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.645, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[22] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.702, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[23] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.759, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[24] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. arXiv preprint arXiv:2206.07696, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[25] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18219-18228, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.872, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola," + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9023" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "Tim Green, Trevor Back, Paul Natev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[27] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[28] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[29] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. arXiv preprint arXiv:2106.00132, 2021.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[30] Chenyang Lei, Xuanchi Ren, Zhaoxiang Zhang, and Qifeng Chen. Blind video deflickering by neural filtering with a flawed atlas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10439-10448, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[31] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.442 + ], + "angle": 0, + "content": "[32] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[33] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.47, + 0.567 + ], + "angle": 0, + "content": "[34] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3667-3676, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.622 + ], + "angle": 0, + "content": "[35] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.469, + 0.678 + ], + "angle": 0, + "content": "[36] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.762 + ], + "angle": 0, + "content": "[37] Nithin Gopalakrishnan Nair, Anoop Cherian, Suhas Lohit, Ye Wang, Toshiaki Koike-Akino, Vishal M Patel, and Tim K Marks. Steered diffusion: A generalized framework for plug-and-play conditional image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20850-20860, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[38] Haomiao Ni, Yihao Liu, Sharon X Huang, and Yuan Xue. Cross-identity video motion retargeting with joint transformation and synthesis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 412–422, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[39] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18444-18455, 2023. 1, 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.148 + ], + "angle": 0, + "content": "[40] Haomiao Ni, Jiachen Liu, Yuan Xue, and Sharon X Huang. 3d-aware talking-head video motion transfer. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4954-4964, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.219 + ], + "angle": 0, + "content": "[41] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.893, + 0.275 + ], + "angle": 0, + "content": "[42] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.276, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[43] OpenAI. Openai: Introducing chatgpt. URL https://openai.com/blog/chatgpt, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.893, + 0.389 + ], + "angle": 0, + "content": "[44] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[45] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.893, + 0.531 + ], + "angle": 0, + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.893, + 0.603 + ], + "angle": 0, + "content": "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.603, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[48] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[49] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[50] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[51] Siddhant Sahu, Manoj Kumar Lenka, and Pankaj Kumar Sa. Blind deblurring using deep learning: A survey. arXiv preprint arXiv:1907.10128, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m:" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9024" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.188 + ], + "angle": 0, + "content": "[53] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.19, + 0.469, + 0.258 + ], + "angle": 0, + "content": "[54] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.261, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[55] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.302, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[56] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.344, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[57] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.401, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.457, + 0.469, + 0.536 + ], + "angle": 0, + "content": "[59] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.539, + 0.469, + 0.592 + ], + "angle": 0, + "content": "[60] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.595, + 0.469, + 0.636 + ], + "angle": 0, + "content": "[61] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. arXiv preprint arXiv:1808.06601, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.638, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[62] Wen Wang, kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.692, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[63] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.762, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[64] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.818, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[65] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[66] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.15, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[67] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[68] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[69] Ceyuan Yang, Zhe Wang, Xinge Zhu, Chen Huang, Jianping Shi, and Dahua Lin. Pose guided human video generation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 201-216, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[70] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023.1, 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "9025" + } + ] +] \ No newline at end of file diff --git a/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/17566a41-32ff-45d4-9eb1-0099d95d87b1_origin.pdf b/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/17566a41-32ff-45d4-9eb1-0099d95d87b1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9dd8ecbd33f67bbf34fa51bd0ec531093ffa2797 --- /dev/null +++ b/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/17566a41-32ff-45d4-9eb1-0099d95d87b1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0559c44000475aae83190f0508ee27f0f592cc634274aadbde858ba3f1e0aecb +size 2046495 diff --git a/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/full.md b/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3ebfce1c8d63a8c28dab0c49a797a4aae85df894 --- /dev/null +++ b/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/full.md @@ -0,0 +1,357 @@ +# TI2V-Zero: Zero-Shot Image Conditioning for Text-to-Video Diffusion Models + +Haomiao Ni $^{1*}$ Bernhard Egger $^{2}$ Toshiaki Koike-Akino $^{3}$ + +1The Pennsylvania State University, USA + +Suhas Lohit3 +Sharon X. Huang1 + +Anoop Cherian3 Ye Wang3 Tim K. Marks3 + +$^{2}$ Friedrich-Alexander-Universität Erlangen-Nürnberg, Germany + +$^{3}$ Mitsubishi Electric Research Laboratories (MERL), USA + +$^{1}\{\mathrm{hfn5052, suh972}\} @ \mathrm{psu.edu}$ $^{2}$ bernhard.egger@fau.de $^{3}\{$ slohit,cherian,yewang,koike,tmarks\}@merl.com https://merl.com/demos/TI2V-Zero + +# Abstract + +Text-conditioned image-to-video generation (TI2V) aims to synthesize a realistic video starting from a given image (e.g., a woman's photo) and a text description (e.g., "a woman is drinking water"). Existing TI2V frameworks often require costly training on video-text datasets and specific model designs for text and image conditioning. In this paper, we propose TI2V-Zero, a zero-shot, tuning-free method that empowers a pretrained text-to-video (T2V) diffusion model to be conditioned on a provided image, enabling TI2V generation without any optimization, finetuning, or introducing external modules. Our approach leverages a pretrained T2V diffusion foundation model as the generative prior. To guide video generation with the additional image input, we propose a "repeat-and-slide" strategy that modulates the reverse denoising process, allowing the frozen diffusion model to synthesize a video frame-by-frame starting from the provided image. To ensure temporal continuity, we employ a DDPM inversion strategy to initialize Gaussian noise for each newly synthesized frame and a resampling technique to help preserve visual details. We conduct comprehensive experiments on both domain-specific and open-domain datasets, where TI2V-Zero consistently outperforms a recent open-domain TI2V model. Furthermore, we show that TI2V-Zero can seamlessly extend to other tasks such as video infilling and prediction when provided with more images. Its autoregressive design also supports long video generation. + +# 1. Introduction + +Image-to-video (I2V) generation is an appealing topic with various applications, including artistic creation, entertainment, and data augmentation for machine learning [39]. Given a single image $x^0$ and a text prompt $y$ , text-conditioned image-to-video (TI2V) generation aims to syn + +![](images/2a552a36a79232e32b7df66896081cb440cc36b38a41e9cba63a86b6e6101bc7.jpg) +"A man with the expression of slight happiness on his face." + +![](images/dd64b4d666b3b23a4249fb9b9314ff6ab5fb3f25a1ec5eda2d6d7e7a5757c230.jpg) + +![](images/d2b7a08a4ccf8d6ac829afff081b3aa5308b445a4bd3bcdd6a561a5cb5b7d757.jpg) + +![](images/ab8124efa72cd21ec71fda21bf1d7b17a2db63a69309a7dedb957f5d1868bcd4.jpg) + +![](images/3341998c58263fcb3ee286e3c475bdd78408c220aba8b4560a43d8834c2682a7.jpg) +"A person is drumming." + +![](images/f82fd823a376b841b9226db745d0fa81ebc9f9ba64f7ae1eb7a189b629d0d551.jpg) + +![](images/7a9448aedbf0e9a0f0ed0937494a78221e422e242cf482ee702896a063f88270.jpg) +"A serene mountain cabin covered in a fresh blanket of snow." +Figure 1. Examples of generated video frames using our proposed T12V-Zero. The given first image $x^0$ is highlighted with the red box, and the text condition $y$ is shown under each row of the video. The remaining columns show the 6th, 11th, and 16th frames of the generated output videos. Each generated video has 16 frames with a resolution of $256 \times 256$ . + +![](images/de3ef1e82e3d3559bd02ac5efec8fb5ef96ff3502308f0884ba35e9ec68ac2c8.jpg) + +![](images/bd730428c2e921fad09caaf0a6596fb7577fa91d821f48af29ef50fc427d7fad.jpg) + +thesize $M$ new frames to yield a realistic video, $\hat{\mathbf{x}} = \langle x^0,\hat{x}^1,\dots ,\hat{x}^M\rangle$ , starting from the given frame $x^0$ and satisfying the text description $y$ . Current TI2V generation methods [59, 63, 70] typically rely on computationally-heavy training on video-text datasets and specific architecture designs to enable text and image conditioning. Some [12, 25] are constrained to specific domains due to the lack of training with large-scale open-domain datasets. Other approaches, such as [14, 67], utilize pretrained foundation models to reduce training costs, but they still need to train additional modules using video data. + +In this paper, we propose TI2V-Zero, which achieves zero-shot TI2V generation using only an open-domain pretrained text-to-video (T2V) latent diffusion model [60]. Here "zero-shot" means that when using the diffusion + +model (DM) that was trained only for text conditioning, our framework enables image conditioning without any optimization, fine-tuning, or introduction of additional modules. Specifically, we guide the generation process by incorporating the provided image $x^0$ into the output latent code at each reverse denoising step. To ensure that the temporal attention layers of the pretrained DM focus on information from the given image, we propose a "repeat-and-slide" strategy to synthesize the video in a frame-by-frame manner, rather than directly generating the entire video volume. Notably, TI2V-Zero is not trained for the specific domain of the provided image, thus allowing the model to generalize to any image during inference. Additionally, its autoregressive generation makes the synthesis of long videos possible. + +While the standard denoising sampling process starting with randomly initialized Gaussian noise can produce matching semantics, it often results in temporally inconsistent videos. Therefore, we introduce an inversion strategy based on the DDPM [20] forward process, to provide a more suitable initial noise for generating each new frame. We also apply a resampling technique [33] in the video DM to help preserve the generated visual details. Our approach ensures that the network maintains temporal consistency, generating visually convincing videos conditioned on the given starting image (see Fig. 1). + +We conduct extensive experiments on MUG [1], UCF-101 [56], and a new open-domain dataset. In these experiments, TI2V-Zero consistently performs well, outperforming a state-of-the-art model [67] that was based on a video diffusion foundation model [8] and was specifically trained to enable open-domain TI2V generation. + +# 2. Related Work + +# 2.1. Conditional Image-to-Video Generation + +Conditional video generation aims to synthesize videos guided by user-provided signals. It can be classified according to which type(s) of conditions are given, such as text-to-video (T2V) generation [5, 16, 21, 23, 31, 65], video-to-video (V2V) generation [7, 38, 40, 45, 61, 64], and image-to-video (I2V) generation [4, 10, 25, 34, 39, 69]. Here we discuss previous text-conditioned image-to-video (TI2V) generation methods [12, 14, 22, 44, 63, 70]. Hu et al. [25] introduced MAGE, a TI2V generator that integrates a motion anchor structure to store appearance-motion-aligned representations through three-dimensional axial transformers. Yin et al. [70] proposed DragNUWA, a diffusion-based model capable of generating videos controlled by text, image, and trajectory information with three modules including a trajectory sampler, a multi-scale fusion, and an adaptive training strategy. However, these TI2V frameworks require computationally expensive training on videotext datasets and a particular model design to support text + +and-image-conditioned training. In contrast, our proposed TI2V-Zero leverages a pretrained T2V diffusion model to achieve zero-shot TI2V generation without additional optimization or fine-tuning, making it suitable for a wide range of applications. + +# 2.2. Adaptation of Diffusion Foundation Models + +Due to the recent successful application of diffusion models (DM) [20, 42, 47, 54, 55] to both image and video generation, visual diffusion foundation models have gained prominence. These include text-to-image (T2I) models such as Imagen [50] and Stable Diffusion [47], as well as text-to-video (T2V) models such as ModelScopeT2V [60] and VideoCrafter1 [8]. These models are trained with large-scale open-domain datasets, often including LAION-400M [52] and WebVid-10M [2]. They have shown immense potential for adapting their acquired knowledge base to address a wide range of downstream tasks, thereby reducing or eliminating the need for extensive labeled data. For example, previous works have explored the application of large T2I models to personalized image generation [13, 49], image editing [17, 33, 35-37], image segmentation [3, 68], video editing [45, 62], and video generation [14, 27, 53, 66]. In contrast to T2I models, there are fewer works on the adaptation of large-scale T2V models. Xing et al. [67] proposed DynamicCrafter for open-domain TI2V generation by adapting a T2V foundation model [8]. To control the generative process, they first employed a learnable image encoding network to project the given image into a text-aligned image embedding space. Subsequently, they utilized dual cross-attention layers to fuse text and image information and also concatenated the image with the initial noise to provide the video DM with more precise image details. In contrast, in this paper we explore how to inject the provided image to guide the DM sampling process based solely on the pretrained T2V model itself, with no additional training for the new TI2V task. + +# 3. Methodology + +Given one starting image $x^0$ and text $y$ , let $\mathbf{x} = \langle x^0, x^1, \ldots, x^M \rangle$ represent a real video corresponding to text $y$ . The objective of text-conditioned image-to-video (TI2V) generation is to synthesize a video $\hat{\mathbf{x}} = \langle x^0, \hat{x}^1, \ldots, \hat{x}^M \rangle$ , such that the conditional distribution of $\hat{\mathbf{x}}$ given $x^0$ and $y$ is identical to the conditional distribution of $\mathbf{x}$ given $x^0$ and $y$ , i.e., $p(\hat{\mathbf{x}} | x^0, y) = p(\mathbf{x} | x^0, y)$ . Our proposed TI2V-Zero can be built on a pretrained T2V diffusion model with a 3D-UNet-based denoising network. Here we choose ModelScopeT2V [60] as backbone due to its promising open-domain T2V generation ability. Below, we first introduce preliminaries about diffusion models, then introduce the architecture of the pretrained T2V model, and finally present the details of our TI2V-Zero. + +![](images/cc501315653f24ec3a0c89a54855a81cf4fd0a353a3f95841365890bf17fc7e0.jpg) +Figure 2. Illustration of the process of applying TI2V-Zero to generate the new frame $\hat{x}^{i+1}$ , given the starting image $x^0$ and text $y$ . TI2V-Zero is built upon a frozen pretrained T2V diffusion model, including frame encoder $\mathcal{E}$ , frame decoder $\mathcal{D}$ , and the denoising U-Net $\epsilon_{\theta}$ . At the beginning of generation ( $i=0$ ), we encode $x^0$ as $z^0$ and repeat it $K$ times to form the queue $\mathbf{s}_0$ . We then apply DDPM-based inversion to $\mathbf{s}_0$ to produce the initial Gaussian noise $\hat{\mathbf{z}}_T$ . Subsequently, in each reverse denoising step using U-Net $\epsilon_{\theta}$ , we keep replacing the first $K$ frames of $\hat{\mathbf{z}}_t$ with the noisy latent code $\mathbf{s}_t$ derived from $\mathbf{s}_0$ . Resampling is also applied within each step to improve motion coherence. We finally decode the final frame of the clean latent code $\hat{\mathbf{z}}_0$ as the new synthesized frame $\hat{x}^{i+1}$ . To compute the new $\mathbf{s}_0$ for the next iteration of generation ( $i>0$ ), we perform a sliding operation by dequeuing $s_0^0$ and enqueuing $\hat{z}_0^K$ within $\mathbf{s}_0$ . + +# 3.1. Preliminaries: Diffusion Models + +Diffusion Models (DM) [20, 54, 55] are probabilistic models designed to learn a data distribution. Here we introduce the fundamental concepts of Denoising Diffusion Probabilistic Models (DDPM). Given a sample from the data distribution $\mathbf{z}_0 \sim q(\mathbf{z}_0)$ , the forward diffusion process of a DM produces a Markov chain $\mathbf{z}_1, \ldots, \mathbf{z}_T$ by iteratively adding Gaussian noise to $\mathbf{z}_0$ according to a variance schedule $\beta_1, \ldots, \beta_T$ , that is: + +$$ +q \left(\mathbf {z} _ {t} \mid \mathbf {z} _ {t - 1}\right) = \mathcal {N} \left(\mathbf {z} _ {t}; \sqrt {1 - \beta_ {t}} \mathbf {z} _ {t - 1}, \beta_ {t} \mathbf {I}\right), \tag {1} +$$ + +where variances $\beta_{t}$ are constant. When the $\beta_{t}$ are small, the posterior $q(\mathbf{z}_{t - 1}|\mathbf{z}_t)$ can be well approximated by a diagonal Gaussian [41, 54]. Furthermore, if the length of the chain, denoted by $T$ , is sufficiently large, $\mathbf{z}_T$ can be well approximated by a standard Gaussian distribution $\mathcal{N}(\mathbf{0},\mathbf{I})$ . These suggest that the true posterior $q(\mathbf{z}_{t - 1}|\mathbf{z}_t)$ can be estimated by $p_{\theta}(\mathbf{z}_{t - 1}|\mathbf{z}_t)$ defined as: + +$$ +p _ {\theta} \left(\mathbf {z} _ {t - 1} \mid \mathbf {z} _ {t}\right) = \mathcal {N} \left(\mathbf {z} _ {t - 1}; \mu_ {\theta} \left(\mathbf {z} _ {t}\right), \sigma_ {t} ^ {2} \mathbf {I}\right), \tag {2} +$$ + +where variances $\sigma_t$ are also constants. The reverse denoising process in the DM (also termed sampling) then generates samples $\mathbf{z}_0 \sim p_{\theta}(\mathbf{z}_0)$ by starting with Gaussian noise $\mathbf{z}_T \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ and gradually reducing noise in a Markov chain $\mathbf{z}_{T-1}, \mathbf{z}_{T-2}, \ldots, \mathbf{z}_0$ using a learned $p_{\theta}(\mathbf{z}_{t-1}|\mathbf{z}_t)$ . To learn $p_{\theta}(\mathbf{z}_{t-1}|\mathbf{z}_t)$ , Gaussian noise $\epsilon$ is first added to $\mathbf{z}_0$ to generate samples $\mathbf{z}_t$ . Utilizing the independence property of the noise added at each forward step in Eq. (1), we can calculate the total noise variance as $\bar{\alpha}_t = \prod_{i=0}^t (1 - \beta_i)$ and transform $\mathbf{z}_0$ to $\mathbf{z}_t$ in a single step: + +$$ +q \left(\mathbf {z} _ {t} \mid \mathbf {z} _ {0}\right) = \mathcal {N} \left(\mathbf {z} _ {t}; \sqrt {\bar {\alpha} _ {t}} \mathbf {z} _ {0}, (1 - \bar {\alpha} _ {t}) \mathbf {I}\right). \tag {3} +$$ + +Then a model $\epsilon_{\theta}$ is trained to predict $\epsilon$ using the following mean-squared error loss: + +$$ +L = \mathbb {E} _ {t \sim \mathcal {U} (1, T), \mathbf {z} _ {0} \sim q (\mathbf {z} _ {0}), \epsilon \sim \mathcal {N} (\mathbf {0}, \mathbf {I})} \left[ \| \epsilon - \epsilon_ {\theta} (\mathbf {z} _ {t}, t) \| ^ {2} \right] ， \tag {4} +$$ + +where diffusion step $t$ is uniformly sampled from $\{1, \dots, T\}$ . Then $\mu_{\theta}(\mathbf{z}_t)$ in Eq. (2) can be derived from $\epsilon_{\theta}(\mathbf{z}_t, t)$ to model $p_{\theta}(\mathbf{z}_{t-1}|\mathbf{z}_t)$ [20]. The denoising model $\epsilon_{\theta}$ is implemented using a time-conditioned U-Net [48] with residual blocks [15] and self-attention layers [58]. Diffusion step $t$ is specified to $\epsilon_{\theta}$ by the sinusoidal position embedding [58]. Conditional generation that samples $\mathbf{z}_0 \sim p_{\theta}(\mathbf{z}_0|y)$ can be achieved by learning a $y$ -conditioned model $\epsilon_{\theta}(\mathbf{z}_t, t, y)$ [41, 47] with classifier-free guidance [19]. During training, the condition $y$ in $\epsilon_{\theta}(\mathbf{z}_t, t, y)$ is replaced by a null label $\emptyset$ with a fixed probability. When sampling, the output is generated as follows: + +$$ +\hat {\epsilon} _ {\theta} \left(\mathbf {z} _ {t}, t, y\right) = \epsilon_ {\theta} \left(\mathbf {z} _ {t}, t, \emptyset\right) + g \cdot \left(\epsilon_ {\theta} \left(\mathbf {z} _ {t}, t, y\right) - \epsilon_ {\theta} \left(\mathbf {z} _ {t}, t, \emptyset\right)\right), \tag {5} +$$ + +where $g$ is the guidance scale. + +# 3.2. Architecture of Pretrained T2V Model + +TI2V-Zero can be built upon a pretrained T2V diffusion model with a 3D-UNet-based denoising network. Here we choose ModelScopeT2V [60] as the pretrained model (denoted $\mathcal{M}$ ). We now describe this T2V model in detail. + +Structure Overview. Given a text prompt $y$ , the T2V model $\mathcal{M}$ synthesizes a video $\hat{\mathbf{x}} = \langle \hat{x}^0, \hat{x}^1, \dots, \hat{x}^K \rangle$ with a pre-defined video of length $(K + 1)$ using a latent video diffusion model. Similar to Latent Diffusion Models (LDM) [47], $\mathcal{M}$ incorporates a frame auto-encoder [11, 28] for the conversion of data between pixel space $\mathcal{X}$ and latent space $\mathcal{Z}$ through its encoder $\mathcal{E}$ and decoder $\mathcal{D}$ . Given the real video $\mathbf{x} = \langle x^0, x^1, \dots, x^K \rangle$ , $\mathcal{M}$ first utilizes the frame encoder $\mathcal{E}$ to encode the video $\mathbf{x}$ as $\mathbf{z} = \langle z^0, z^1, \dots, z^K \rangle$ . Here the sizes of pixel frame $x$ and latent frame $z$ are $H_x \times W_x \times 3$ and $H_z \times W_z \times C_z$ , respectively. To be consistent with the notation used for the DM, we denote the + +Algorithm 1 Generation using our TI2V-Zero approach. +Input: The starting frame $x^0$ ; The text prompt $y$ ; The pretrained T2V Model $\mathcal{M}$ for generating $(K + 1)$ -frame videos, including frame encoder $\mathcal{E}$ and frame decoder $\mathcal{D}$ , and the DM de-noising networks $\epsilon_{\theta}$ ; The iteration number $U$ for resampling; The parameter $M$ to control the length of the output video. +Output: A synthesized video $\hat{\mathbf{x}}$ with $(M + 1)$ frames. +1: $z^0 \gets \mathcal{E}(x^0)$ // Encode $x^0$ +2: $\mathbf{s}_0 \gets \langle z^0, z^0, \dots, z^0 \rangle$ // Repeat $z^0$ for $K$ times +3: $\hat{\mathbf{x}} \gets \langle x^0 \rangle$ +4: for $i = 1, 2, \dots, M$ do +// Generate one new frame $\hat{x}^i$ +5: $\mathbf{s}_T \sim \mathcal{N}(\sqrt{\bar{\alpha}_T} \mathbf{s}_0, (1 - \bar{\alpha}_T) \mathbf{I})$ // DDPM Inversion +6: $\hat{z}_T^K \sim \mathcal{N}(\sqrt{\bar{\alpha}_T} s_0^{K-1}, (1 - \bar{\alpha}_T) \mathbf{I})$ +7: $\hat{\mathbf{z}}_T \gets \mathbf{s}_T \cup \hat{z}_T^K$ // Initialize $\hat{\mathbf{z}}_T$ +8: for $t = T-1, \dots, 2, 1$ do + $\mathbf{s}_t \sim \mathcal{N}(\sqrt{\bar{\alpha}_t} \mathbf{s}_0, (1 - \bar{\alpha}_t) \mathbf{I})$ +for $u = 1, 2, \dots, U$ do + $\langle \hat{z}_t^0, \hat{z}_t^1, \dots, \hat{z}_t^{K-1} \rangle \gets \mathbf{s}_t$ // Replace + $\hat{\mathbf{z}}_{t-1} \sim \mathcal{N}(\mu_\theta (\hat{\mathbf{z}}_t, y), \sigma_t^2 \mathbf{I})$ +if $u < U$ and $t > 1$ then + $\hat{\mathbf{z}}_t \sim \mathcal{N}(\sqrt{1 - \beta_t} \hat{\mathbf{z}}_{t-1}, \beta_t \mathbf{I})$ // Resample +end if +end for +end for + $\mathbf{s}_0 \gets \langle s_0^1, s_0^2, \dots, s_0^{K-1} \rangle \cup \hat{z}_0^K$ // Slide + $\hat{x}^i \gets D(\hat{z}_0^K)$ // Decode $\hat{z}_0^K$ $\hat{\mathbf{x}} \gets \hat{\mathbf{x}} \cup \hat{x}^i$ +end for +return $\hat{\mathbf{x}}$ + +clean video latent $\mathbf{z} = \mathbf{z}_0 = \langle z_0^0, z_0^1, \ldots, z_0^K \rangle$ . $\mathcal{M}$ then learns a DM on the latent space $\mathcal{Z}$ through a 3D denoising U-Net $\epsilon_{\theta}$ [9]. Let $\mathbf{z}_t = \langle z_t^0, z_t^1, \ldots, z_t^K \rangle$ represent the latent sequence that results from adding noise over $t$ steps to the original latent sequence $\mathbf{z}_0$ . When training, the forward diffusion process of a DM transforms the initial latent sequence $\mathbf{z}_0$ into $\mathbf{z}_T$ by iteratively adding Gaussian noise $\epsilon$ for $T$ steps. During inference, denoising U-Net $\epsilon_{\theta}$ predicts the added noise at each step, enabling the generation of the clean latent sequence $\hat{\mathbf{z}}_0 = \langle \hat{z}_0^0, \hat{z}_0^1, \ldots, \hat{z}_0^K \rangle$ starting from randomly sampled Gaussian noise $\mathbf{z}_T \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ . + +Text Conditioning Mechanism. $\mathcal{M}$ employs a cross-attention mechanism [47] to incorporate text information into the generative process as guidance. Specifically, $\mathcal{M}$ uses a pretrained CLIP model [46] to encode the prompt $y$ as the text embedding $e$ . The embedding $e$ is later used as the key and value in the multi-head attention layer within the spatial attention blocks, thus enabling the integration of text features with the intermediate U-Net features in $\epsilon_{\theta}$ . + +Denoising U-Net. The denoising U-Net $\epsilon_{\theta}$ includes four key building blocks: the initial block, the downsampling block, the spatio-temporal block, and the upsampling block. The initial block transfers the input into the embedding + +![](images/4be811a2a4ed11d79312b906b57c48e810421e0172805ae5595114d846b4ecc4.jpg) +Video Time +Figure 3. Illustration of the motivation behind our framework. We explore the application of a replacing-based baseline approach (rows 2-4, labeled "Replacing") and our TI2V-Zero (rows 5-6, labeled "TI2V-Zero") in various video generation tasks. The given real frames for each task are highlighted by red boxes and the text input is shown under the block. The replacing-based approach is only effective at predicting a single frame when all the other frames in the video are provided, while TI2V-Zero generates temporally coherent videos for both the TI2V and video infilling tasks. + +space, while the downsampling and upsampling blocks are responsible for spatially downsampling and upsampling the feature maps. The spatio-temporal block is designed to capture spatial and temporal dependencies in the latent space, which comprises 2D spatial convolution, 1D temporal convolution, 2D spatial attention, and 1D temporal attention. + +# 3.3. Our Framework + +Leveraging the pretrained T2V foundation model $\mathcal{M}$ , we first propose a straightforward replacing-based baseline for adapting $\mathcal{M}$ to TI2V generation. We then analyze the possible reasons why it fails and introduce our TI2V-Zero framework, which includes a repeat-and-slide strategy, DDPM-based inversion, and resampling. Figure 2 and Algorithm 1 demonstrate the inference process of TI2V-Zero. + +Replacing-based Baseline. We assume that the pretrained model $\mathcal{M}$ is designed to generate the video with a fixed length of $(K + 1)$ . So we first consider synthesizing videos with that same length $(K + 1)$ , i.e., $M = K$ . Since the DM process operates within the latent space $\mathcal{Z}$ , we use the encoder $\mathcal{E}$ to map the given starting frame $x^0$ into the latent representation $z^0$ . Additionally, we denote $z^0 = z_0^0$ to specify that the latent is clean and corresponds + +to diffusion step 0 of the DM. Note that each reverse denoising step in Eq. (2) from $\hat{\mathbf{z}}_t$ to $\hat{\mathbf{z}}_{t - 1}$ depends solely on $\hat{\mathbf{z}}_t = \langle \hat{z}_t^0,\hat{z}_t^1,\dots ,\hat{z}_t^K\rangle$ . To ensure that the first frame of the final synthesized clean video latent $\hat{\mathbf{z}}_0 = \langle \hat{z}_0^0,\hat{z}_0^1,\dots ,\hat{z}_0^K\rangle$ at step 0 matches the provided image latent, i.e., $\hat{z}_0^0 = z_0^0$ we can modify the first generated latent $\hat{z}_t^0$ of $\hat{\mathbf{z}}_t$ at each reverse step, as long as the signal-to-noise ratio of each frame latent in $\hat{\mathbf{z}}_t$ remains consistent. Using Eq. (3), we can add $t$ steps of noise to the provided image latent $z_0^0$ , allowing us to sample $z_{t}^{0}$ through a single-step calculation. By replacing the first generated latent $\hat{z}_t^0$ with the noisy image latent $z_{t}^{0}$ at each reverse denoising step, we might expect that the video generation process can be guided by $z_0^0$ with the following expressions defined for each reverse step: + +$$ +z _ {t} ^ {0} \sim \mathcal {N} \left(\sqrt {\bar {\alpha}} _ {t} z _ {0} ^ {0}, (1 - \bar {\alpha} _ {t}) \mathbf {I}\right), \tag {6a} +$$ + +$$ +\hat {z} _ {t} ^ {0} \leftarrow z _ {t} ^ {0}, \tag {6b} +$$ + +$$ +\hat {\mathbf {z}} _ {t - 1} \sim \mathcal {N} \left(\mu_ {\theta} \left(\hat {\mathbf {z}} _ {t}, y\right), \sigma_ {t} ^ {2} \mathbf {I}\right). \tag {6c} +$$ + +Specifically, in each reverse step from $\hat{\mathbf{z}}_t$ to $\hat{\mathbf{z}}_{t - 1}$ , as shown in Eq. (6a), we first compute the noisy latent $z_{t}^{0}$ by adding Gaussian noise to the given image latent $z_0^0$ over $t$ steps. Then, we replace the first latent $\hat{z}_t^0$ of $\hat{\mathbf{z}}_t$ with $z_{t}^{0}$ in Eq. (6b) to incorporate the provided image into the generation process. Finally, in Eq. (6c), we pass $\hat{\mathbf{z}}_t$ through the denoising network to generate $\hat{\mathbf{z}}_{t - 1}$ , where the text $y$ is integrated by classifier-free guidance (Eq. (5)). After $T$ iterations, the final clean latent $\hat{\mathbf{z}}_0$ at diffusion step 0 can be mapped back into the image space $\mathcal{X}$ using the decoder $\mathcal{D}$ . + +Using this replacing-based baseline, we might expect that the temporal attention layers in $\epsilon_{\theta}$ can utilize the context provided by the first frame latent $\hat{z}_t^0$ to generate the subsequent frame latents in a manner that harmonizes with $\hat{z}_t^0$ . However, as shown in Fig. 3, row 2, this replacing-based approach fails to produce a video that is temporally consistent with the first image. The generated frames are consistent with each other, but not with the provided first frame. + +To analyze possible reasons for failure, we apply this baseline to a simpler video infilling task, where every other frame is provided and the model needs to predict the interspersed frames. In this case, the baseline replaces the generated frame latents at positions corresponding to real frames with noisy provided-frame latents in each reverse step. The resulting video, in Fig. 3, row 3, looks like a combination of two independent videos: the generated (even) frames are consistent with each other but not with the provided (odd) frames. We speculate that this may result from the intrinsic dissimilarity between frame latents derived from the given real images and those sampled from $\epsilon_{\theta}$ . Thus, the temporal attention values between frame latents sampled in the same way (both from the given images or both from $\epsilon_{\theta}$ ) will be higher, while the attention values between frame latents sampled in different ways (one from the given image and the other from $\epsilon_{\theta}$ ) will be lower. Therefore, the temporal attention layers of $\mathcal{M}$ tend to utilize the information from latents + +![](images/56a3fa54e63647ab3a9f3f42593a71457e609a3ccf459ff78b944420e667deec.jpg) +Figure 4. Qualitative ablation study comparing different sampling strategies for our TI2V-Zero on MUG. The first image $\hat{x}^0$ is highlighted with the red box and text $y$ is shown under the block. The 1st, 6th, 11th, and 16th frames of the videos are shown in each column. The terms Inversion, DDIM, and Resample denote the application of DDPM inversion, the steps using DDIM sampling, and the iteration number using resampling, respectively. + +produced by $\epsilon_{\theta}$ to synthesize new frames at each reverse step, ignoring the provided frames. We further simplify the task to single-frame prediction, where the model only needs to predict a single frame when all the other frames in the video are given. In this setting, all the frame latents except for the final frame are replaced by noisy provided-frame latents in each reverse step. Thus, temporal attention layers can only use information from the real frames. In this case, Fig. 3, row 4, shows that the baseline can now generate a final frame that is consistent with the previous frames. + +Repeat-and-Slide Strategy. Inspired by the observation in Fig. 3, to guarantee that the temporal attention layers of $\mathcal{M}$ depend solely on the given image, we make two major changes to the proposed replacing-based baseline: (1) instead of using $\mathcal{M}$ to directly synthesize the entire $(K + 1)$ -frame video, we switch to a frame-by-frame generation approach, i.e., we generate only one new frame latent in each complete DM sampling process; (2) for each sampling process generating the new frame latent, we ensure that only one frame latent is produced from $\epsilon_{\theta}$ , while the other $K$ frame latents are derived from the given real image and previously synthesized frames, thereby forcing temporal attention layers to only use the information from these frame latents. Specifically, we construct a queue of + +$K$ frame latents, denoted as $\mathbf{s}_0 = \langle s_0^0,s_0^1,\dots ,s_0^{K - 1}\rangle$ . We also define $\mathbf{s}_t = \langle s_t^0,s_t^1,\dots ,s_t^{K - 1}\rangle$ , which is obtained by adding $t$ steps of Gaussian noise to the clean $\mathbf{s}_0$ . Similar to our replacing-based baseline in the single-frame prediction task, in each reverse step from $\hat{\mathbf{z}}_t$ to $\hat{\mathbf{z}}_{t - 1}$ , we replace the first $K$ frame latents in $\hat{\mathbf{z}}_t$ by $\mathbf{s}_t$ . Consequently, the temporal attention layers have to utilize information from $\mathbf{s}_0$ to synthesize the new frame's latent, $\hat{z}_0^K$ . Considering that only one starting image latent $z^0$ is provided, we propose a "repeat-and-slide" strategy to construct $\mathbf{s}_0$ . At the beginning of video generation, we repeat $z^0$ for $K$ frames to form $\mathbf{s}_0$ , and gradually perform a sliding operation within the queue $\mathbf{s}_0$ by dequeuing the first frame latent $s_0^0$ and enqueuing the newly generated latent $\hat{z}_0^K$ after each complete DM sampling process. Note that though the initial $\mathbf{s}_0$ is created by repeating $z^0$ , the noise added to get $\mathbf{s}_t$ is different for each frame's latent in $\mathbf{s}_t$ , thus ensuring diversity. The following expressions define one reverse step in the DM sampling process: + +$$ +\mathbf {s} _ {t} \sim \mathcal {N} \left(\sqrt {\bar {\alpha}} _ {t} \mathbf {s} _ {0}, (1 - \bar {\alpha} _ {t}) \mathbf {I}\right), \tag {7a} +$$ + +$$ +\left\langle \hat {z} _ {t} ^ {0}, \hat {z} _ {t} ^ {1}, \dots , \hat {z} _ {t} ^ {K - 1} \right\rangle \leftarrow \mathbf {s} _ {t}, \tag {7b} +$$ + +$$ +\hat {\mathbf {z}} _ {t - 1} \sim \mathcal {N} \left(\mu_ {\theta} \left(\hat {\mathbf {z}} _ {t}, y\right), \sigma_ {t} ^ {2} \mathbf {I}\right). \tag {7c} +$$ + +Specifically, in each reverse denoising step from $\hat{\mathbf{z}}_t$ to $\hat{\mathbf{z}}_{t - 1}$ , we first add $t$ steps of Gaussian noise to the queue $\mathbf{s}_0$ to yield $\mathbf{s}_t$ in Eq. (7a). Subsequently, we replace the previous $K$ frames of $\hat{\mathbf{z}}_t$ with $\mathbf{s}_t$ in Eq. (7b) and input $\hat{\mathbf{z}}_t$ to the denoising network to produce the less noisy latent $\hat{\mathbf{z}}_{t - 1}$ (Eq. (7c)). + +With the repeat-and-slide strategy, model $\mathcal{M}$ is tasked with predicting only one new frame, while the preceding $K$ frames are incorporated into the reverse process to ensure that the temporal attention layers depend solely on information derived from the provided image. + +DDPM-based Inversion. Though the DM sampling process starting with randomly sampled Gaussian noise produces matching semantics, the generated video is often temporally inconsistent (Fig. 4, row 2). To provide initial noise that can produce more temporally consistent results, we introduce an inversion strategy based on the DDPM [20] forward process when generating the new frame latent. Specifically, at the beginning of each DM sampling process to synthesize the new frame latent $\hat{z}_0^K$ , instead of starting with the $\hat{\mathbf{z}}_T$ randomly sampled from $\mathcal{N}(\mathbf{0},\mathbf{I})$ , we add $T$ full steps of Gaussian noise to $\mathbf{s}_0$ to obtain $\mathbf{s}_T$ using Eq. (3). Note that $\hat{\mathbf{z}}$ has $K + 1$ frames, while $\mathbf{s}$ has $K$ frames. We then use $\mathbf{s}_T$ to initialize the first $K$ frames of $\hat{\mathbf{z}}_T$ . We copy the last frame $s_T^{K - 1}$ of $\mathbf{s}_T$ to initialize the final frame $\hat{z}_T^K$ , as the $(K - 1)\mathrm{th}$ frame is the closest to the $K$ th frame. + +Resampling. Similar to [24, 33], we further apply a resampling technique, which was initially designed for the image inpainting task, to the video DM to enhance motion coherence. Particularly, after performing a one-step denoising operation in the reversed process, we add one-step noise again to revert the latent. This procedure is repeated mul + +
InversionDDIMResampleFVD↓sFVD↓tFVD↓
X1001656.372074.77±411.741798.05±235.34
✓100339.89443.97±139.10405.22±61.58
✓500463.55581.32±234.09535.06±85.27
✓102207.62299.14±87.24278.73±47.84
✓104180.09267.17±74.72252.77±39.02
+ +Table 1. Quantitative ablation study comparing different sampling strategies for proposed TI2V-Zero on the MUG dataset. Inversion, DDIM, and Resample denote the application of DDPM-based inversion, the steps using DDIM sampling, and the iteration number using resampling, respectively. + +
Distributions for ComparisonFVD↓tFVD↓
TI2V-Zero-Fake vs. ModelScopeT2V366.41921.31±251.85
TI2V-Zero-Real vs. Real Videos477.191306.75±271.82
ModelScopeT2V vs. Real Videos985.822264.08±501.28
TI2V-Zero-Fake vs. Real Videos937.112177.70±436.71
+ +Table 2. Result analysis of TI2V-Zero starting from the real (i.e., TI2V-Zero-Real) or synthesized frames (i.e., TI2V-Zero-Fake) on the UCF101 dataset. + +tiple times for each diffusion step, ensuring harmonization between the predicted and conditioning frame latents (see Algorithm 1 for details). + +# 4. Experiments + +# 4.1. Datasets and Metrics + +We conduct comprehensive experiments on three datasets. More details about datasets, such as selected subjects and text prompts, can be found in our Supplementary Materials. + +MUG facial expression dataset [1] contains 1,009 videos of 52 subjects performing 7 different expressions. We include this dataset to evaluate the performance of models in scenarios with small motion and a simple, unchanged background. To simplify the experiments, we randomly select 5 male and 5 female subjects, and 4 expressions. We use the text prompt templates like "A woman with the expression of slight {label} on her face." to change the expression class label to be text input. Since the expressions shown in the videos of MUG are often not obvious, we add "slight" in the text input to avoid large motion. + +UCF101 action recognition dataset [56] contains 13,320 videos from 101 human action classes. We include this dataset to measure performance under complicated motion and complex, changing backgrounds. To simplify the experiments, we select 10 action classes and the first 10 subjects within each class. We use text prompt templates such as "A person is performing {label}." to change the class label to text input. + +In addition to the above two datasets, we create an OPEN dataset to assess the model's performance in open-domain TI2V generation. We first utilize ChatGPT [43] to generate 10 text prompts. Subsequently, we employ Stable + +![](images/60831c16be81240768b095d1a822dfd337e1bba6492c2354dd01a4e53c9eb508.jpg) +Figure 5. Qualitative comparison among different methods on multiple datasets for TI2V generation. Columns in each block display the 1st, 6th, 11th, and 16th frames of the output videos, respectively. There are 16 frames with a resolution of $256 \times 256$ for each video. The given image $x^0$ is highlighted with the red box and the text prompt $y$ is shown under each block. + +
ModelMUGUCF101
FVD↓sFVD↓tFVD↓FVD↓tFVD↓
DynamiCrafter [67]1094.721359.86±257.731223.89±105.94589.591540.02±199.59
TI2V-Zero w/o Resample (Ours)339.89443.97±139.10405.22±61.58493.191319.77±283.87
TI2V-Zero w/ Resample (Ours)180.09267.17±74.72252.77±39.02477.191306.75±271.82
+ +Table 3. Quantitative comparison among different methods on multiple datasets for TI2V generation. + +Diffusion 1.5 [47] to synthesize 100 images from each text prompt, generating a total of 1,000 starting images and 10 text prompts for evaluating TI2V models. + +Data Preprocessing. We resize all the videos/images to $256 \times 256$ resolution. For UCF101, since most of the video frames are not square, we crop the central part of the frames. To obtain ground truth videos for computing metrics, we uniformly sample 16 frames from each video in the datasets to generate the video clips with a fixed length. + +Metrics. Following prior work [21, 22, 25], we assess the visual quality, temporal coherence, and sample diversity of generated videos using Fréchet Video Distance (FVD) [57]. Similar to Fréchet Inception Distance (FID) [18], which is used for image quality evaluation, FVD utilizes a video classification network I3D [6] pretrained on Kinetics-400 dataset [26] to extract feature representation of real and synthesized videos. Then it calculates the Fréchet distance between the distributions of the real and synthesized video features. To measure how well a generated video aligns with the text prompt $y$ (condition accuracy) and the given image $x_0$ (subject relevance), following [39], we design two variants of FVD, namely text-conditioned FVD (tFVD) and subject-conditioned FVD (sFVD). tFVD and sFVD compare the distance between real and synthesized video feature distributions under the same text $y$ or the same subject image $x_0$ , respectively. We first compute tFVD and sFVD for each condition $y$ and image $x_0$ , then report their mean + +and variance as final results. In our experiments, we generate 1,000 videos for all the models to estimate the feature distributions. We compute both tFVD and sFVD on the MUG dataset, but for UCF101, we only consider tFVD since it doesn't contain videos of different actions for the same subject. For the OPEN dataset, we only present qualitative results due to the lack of ground truth videos. Unless otherwise specified, all the generated videos are 16 frames (i.e., $M = 15$ ) with resolution $256 \times 256$ . + +# 4.2. Implementation Details + +Model Implementation. We take the ModelScopeT2V 1.4.2 [60] as basis and implement our modifications. For text-conditioned generation, we employ classifier-free guidance with $g = 9.0$ in Eq. (5). Determined by our preliminary experiments, we choose 10-step DDIM and 4-step resampling as the default setting for MUG and OPEN, and 50-step DDIM and 2-step resampling for UCF101. + +Implementation of SOTA Model. We compare our TI2V-Zero with a state-of-the-art (SOTA) model DynamamiCrafter, a recent open-domain TI2V framework [67]. DynamamiCrafter is based on a large-scale pretrained T2V foundation model VideoCrafter1 [16]. It introduces a learnable projection network to enable image-conditioned generation and then fine-tunes the entire framework. We implement DynamamiCrafter using their provided code with their default settings. For a fair comparison, all the generated videos are + +![](images/1fe46997d8682b6916475eed46a7cce2607959dffa8f63729ccc7d5ffd8b019a.jpg) +"A mesmerizing display of the northern lights in the Arctic." +Figure 6. Example of long video generation using our TI2V-Zero on the OPEN dataset. The given image $x^0$ is highlighted with a red box, and the text prompt $y$ is shown under the set of frames. There are a total of 128 video frames ( $M = 127$ ), and the synthesized results for every 14 frames are presented. + +centrally-cropped and resized to $256 \times 256$ . + +# 4.3. Result Analysis + +Ablation Study. We conduct ablation study of different sampling strategies on MUG. As shown in Tab. 1 and Fig. 4, compared with generating using randomly sampled Gaussian noise, initializing the input noise with DDPM inversion is important for generating temporally continuous videos, improving all of the metrics dramatically. For MUG, increasing the DDIM sampling steps from 10 to 50 does not enhance the video quality but requires more inference time. Thus, we choose 10-step DDIM as the default setting on MUG. As shown in Fig. 4 and Tab. 1, adding resampling helps preserve identity details (e.g., hairstyle and facial appearance), resulting in lower FVD scores. Increasing resampling steps from 2 to 4 further improves FVD scores. + +Effect of Real/Synthesized Starting Frames. We also explore the effect of video generation starting with real or synthesized frames on UCF101. We initially use the first frame of the real videos to generate videos with our TI2V-Zero, termed TI2V-Zero-Real. Additionally, we utilize the backbone model ModelScopeT2V [60] to generate synthetic videos using the text inputs of UCF101. We then employ TI2V-Zero to create videos from the first frame of the generated fake videos, denoted as TI2V-Zero-Fake. As shown in Tab. 2, [TI2V-Zero-Fake vs. ModelScopeT2V] can achieve better FVD scores than [TI2V-Zero-Real vs. Real Videos]. The reason may be that frames generated by ModelScopeT2V can be considered as in-distribution data since TI2V-Zero is built upon it. We also compare the output video distribution of TI2V-Zero-Fake and ModelScopeT2V with real videos in Tab. 2. Though starting from the same synthesized frames, TI2V-Zero-Fake can generate more realistic videos than the backbone model. + +Comparison with SOTA Model. We compare our proposed TI2V-Zero with DynamiCrafter [67] in Tab. 3 and Fig. 5. From Fig. 5, one can find that DynamiCrafter + +struggles to preserve details from the given image, and the motion of its generated videos is also less diverse. Note that DynamiCrafter requires additional fine-tuning to enable TI2V generation. In contrast, without using any fine-tuning or introducing external modules, our proposed TI2V-Zero can precisely start with the given image and output more visually-pleasing results, thus achieving much better FVD scores on both MUG and UCF101 datasets in Tab. 3. The comparison between our TI2V-Zero models with and without using resampling in Fig. 5 and Tab. 3 also demonstrates the effectiveness of using resampling, which can help maintain identity and background details. + +Extension to Other Applications. TI2V-Zero can also be extended to other tasks as long as we can construct $\mathbf{s}_0$ with $K$ images at the beginning. These images can be obtained either from ground truth videos or by applying the repeating operation. Then we can slide $\mathbf{s}_0$ when generating the subsequent frames. We have applied TI2V-Zero in video infilling (see the last row in Fig. 3), video prediction (see Supplementary Materials), and long video generation (see Fig. 6). As shown in Fig. 6, when generating a 128-frame video on the OPEN dataset, our method can preserve the mountain shape in the background, even at the 71st frame (frame $\hat{x}^{70}$ ). The generated video examples and additional experimental results are in our Supplementary Materials. + +# 5. Conclusion + +In this paper, we propose a zero-shot text-conditioned image-to-video framework, TI2V-Zero, to generate videos by modulating the sampling process of a pretrained video diffusion model without any optimization or fine-tuning. Comprehensive experiments show that TI2V-Zero can achieve promising performance on multiple datasets. + +While showing impressive potential, our proposed TI2V-Zero still has some limitations. First, as TI2V-Zero relies on a pretrained T2V diffusion model, the generation quality of TI2V-Zero is constrained by the capabilities and limitations of the pretrained T2V model. We plan to extend our method to more powerful video diffusion foundation models in the future. Second, our method sometimes generates videos that are blurry or contain flickering artifacts. One possible solution is to apply post-processing methods such as blind video deflickering [30] or image/video deblurring [51] to enhance the quality of final output videos or the newly synthesized frame in each generation. Finally, compared with GAN and standard video diffusion models, our approach is considerably slower because it requires running the entire diffusion process for each frame generation. We will investigate some faster sampling methods [29, 32] to reduce generation time. + +# References + +[1] Niki Aifanti, Christos Papachristou, and Anastasios Delopoulos. The mug facial expression database. In 11th International Workshop on Image Analysis for Multimedia Interactive Services WIAMIS 10, pages 1-4. IEEE, 2010. 2, 6 +[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 2 +[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 2 +[4] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Bjorn Ommer. Understanding object dynamics for interactive image-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5171-5181, 2021. 2 +[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2 +[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7 +[7] Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A Efros. Everybody dance now. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5933-5942, 2019. 2 +[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videographer1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2 +[9] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3d u-net: learning dense volumetric segmentation from sparse annotation. In International conference on medical image computing and computer-assisted intervention, pages 424-432. Springer, 2016. 4 +[10] Michael Dorkenwald, Timo Milbich, Andreas Blattmann, Robin Rombach, Konstantinos G Derpanis, and Bjorn Ommer. Stochastic image-to-video synthesis using cinns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3742-3753, 2021. 2 +[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3 +[12] Tsu-Jui Fu, Licheng Yu, Ning Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. Tell me what happened: Unifying text-guided video completion via multimodal masked video generation. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10681-10692, 2023. 1, 2 +[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2 +[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 1, 2 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3 +[16] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 2, 7 +[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2 +[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7 +[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 6 +[21] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 7 +[22] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 7 +[23] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2 +[24] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. arXiv preprint arXiv:2206.07696, 2022. 6 +[25] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18219-18228, 2022. 1, 2, 7 +[26] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, + +Tim Green, Trevor Back, Paul Natev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 7 +[27] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 2 +[28] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[29] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. arXiv preprint arXiv:2106.00132, 2021.8 +[30] Chenyang Lei, Xuanchi Ren, Zhaoxiang Zhang, and Qifeng Chen. Blind video deflickering by neural filtering with a flawed atlas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10439-10448, 2023. 8 +[31] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2 +[32] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927, 2022. 8 +[33] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2, 6 +[34] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3667-3676, 2022. 2 +[35] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2 +[36] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. +[37] Nithin Gopalakrishnan Nair, Anoop Cherian, Suhas Lohit, Ye Wang, Toshiaki Koike-Akino, Vishal M Patel, and Tim K Marks. Steered diffusion: A generalized framework for plug-and-play conditional image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20850-20860, 2023. 2 +[38] Haomiao Ni, Yihao Liu, Sharon X Huang, and Yuan Xue. Cross-identity video motion retargeting with joint transformation and synthesis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 412–422, 2023. 2 +[39] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18444-18455, 2023. 1, 2, 7 + +[40] Haomiao Ni, Jiachen Liu, Yuan Xue, and Sharon X Huang. 3d-aware talking-head video motion transfer. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4954-4964, 2024. 2 +[41] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3 +[42] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2 +[43] OpenAI. Openai: Introducing chatgpt. URL https://openai.com/blog/chatgpt, 2022.6 +[44] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 2 +[45] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 2 +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4 +[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 4, 7 +[48] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3 +[49] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2 +[50] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2 +[51] Siddhant Sahu, Manoj Kumar Lenka, and Pankaj Kumar Sa. Blind deblurring using deep learning: A survey. arXiv preprint arXiv:1907.10128, 2019. 8 +[52] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: + +Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2 +[53] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2 +[54] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2, 3 +[55] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 2, 3 +[56] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2, 6 +[57] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7 +[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[59] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 1 +[60] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2, 3, 7, 8 +[61] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. arXiv preprint arXiv:1808.06601, 2018. 2 +[62] Wen Wang, kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023. 2 +[63] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 1, 2 +[64] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022. 2 +[65] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2 +[66] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu + +Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2 +[67] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 1, 2, 7, 8 +[68] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 2 +[69] Ceyuan Yang, Zhe Wang, Xinge Zhu, Chen Huang, Jianping Shi, and Dahua Lin. Pose guided human video generation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 201-216, 2018. 2 +[70] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023.1, 2 \ No newline at end of file diff --git a/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/images.zip b/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..985e9cf55bb64c15e93197dfcc7f09822e710211 --- /dev/null +++ b/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b91d52805a915bfc8f2da321c9eecb832166c23dd3f160ddc5e67894b78bfdb2 +size 538894 diff --git a/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/layout.json b/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d40300351c28ed79a5b4c2bf55ab9d3bb8af0052 --- /dev/null +++ b/2024/TI2V-Zero_ Zero-Shot Image Conditioning for Text-to-Video Diffusion Models/layout.json @@ -0,0 +1,13332 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 55, + 103, + 538, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 538, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 538, + 120 + ], + "type": "text", + "content": "TI2V-Zero: Zero-Shot Image Conditioning for Text-to-Video Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 142, + 249, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 142, + 249, + 171 + ], + "spans": [ + { + "bbox": [ + 75, + 142, + 249, + 171 + ], + "type": "text", + "content": "Haomiao Ni" + }, + { + "bbox": [ + 75, + 142, + 249, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 75, + 142, + 249, + 171 + ], + "type": "text", + "content": " Bernhard Egger" + }, + { + "bbox": [ + 75, + 142, + 249, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 142, + 249, + 171 + ], + "type": "text", + "content": " Toshiaki Koike-Akino" + }, + { + "bbox": [ + 75, + 142, + 249, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 96, + 172, + 248, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 172, + 248, + 185 + ], + "spans": [ + { + "bbox": [ + 96, + 172, + 248, + 185 + ], + "type": "text", + "content": "1The Pennsylvania State University, USA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 272, + 143, + 361, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 143, + 361, + 172 + ], + "spans": [ + { + "bbox": [ + 272, + 143, + 361, + 172 + ], + "type": "text", + "content": "Suhas Lohit3 \nSharon X. Huang1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 362, + 143, + 518, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 143, + 518, + 171 + ], + "spans": [ + { + "bbox": [ + 362, + 143, + 518, + 171 + ], + "type": "text", + "content": "Anoop Cherian3 Ye Wang3 Tim K. Marks3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 266, + 173, + 497, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 173, + 497, + 185 + ], + "spans": [ + { + "bbox": [ + 266, + 173, + 497, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 266, + 173, + 497, + 185 + ], + "type": "text", + "content": "Friedrich-Alexander-Universität Erlangen-Nürnberg, Germany" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 189, + 186, + 403, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 186, + 403, + 198 + ], + "spans": [ + { + "bbox": [ + 189, + 186, + 403, + 198 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 189, + 186, + 403, + 198 + ], + "type": "text", + "content": "Mitsubishi Electric Research Laboratories (MERL), USA" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 90, + 201, + 504, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 201, + 504, + 227 + ], + "spans": [ + { + "bbox": [ + 90, + 201, + 504, + 227 + ], + "type": "inline_equation", + "content": "^{1}\\{\\mathrm{hfn5052, suh972}\\} @ \\mathrm{psu.edu}" + }, + { + "bbox": [ + 90, + 201, + 504, + 227 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 201, + 504, + 227 + ], + "type": "text", + "content": "bernhard.egger@fau.de " + }, + { + "bbox": [ + 90, + 201, + 504, + 227 + ], + "type": "inline_equation", + "content": "^{3}\\{" + }, + { + "bbox": [ + 90, + 201, + 504, + 227 + ], + "type": "text", + "content": "slohit,cherian,yewang,koike,tmarks\\}@merl.com https://merl.com/demos/TI2V-Zero" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 255, + 192, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 267 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 267 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 280, + 290, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 290, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 290, + 604 + ], + "type": "text", + "content": "Text-conditioned image-to-video generation (TI2V) aims to synthesize a realistic video starting from a given image (e.g., a woman's photo) and a text description (e.g., \"a woman is drinking water\"). Existing TI2V frameworks often require costly training on video-text datasets and specific model designs for text and image conditioning. In this paper, we propose TI2V-Zero, a zero-shot, tuning-free method that empowers a pretrained text-to-video (T2V) diffusion model to be conditioned on a provided image, enabling TI2V generation without any optimization, finetuning, or introducing external modules. Our approach leverages a pretrained T2V diffusion foundation model as the generative prior. To guide video generation with the additional image input, we propose a \"repeat-and-slide\" strategy that modulates the reverse denoising process, allowing the frozen diffusion model to synthesize a video frame-by-frame starting from the provided image. To ensure temporal continuity, we employ a DDPM inversion strategy to initialize Gaussian noise for each newly synthesized frame and a resampling technique to help preserve visual details. We conduct comprehensive experiments on both domain-specific and open-domain datasets, where TI2V-Zero consistently outperforms a recent open-domain TI2V model. Furthermore, we show that TI2V-Zero can seamlessly extend to other tasks such as video infilling and prediction when provided with more images. Its autoregressive design also supports long video generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "content": "Image-to-video (I2V) generation is an appealing topic with various applications, including artistic creation, entertainment, and data augmentation for machine learning [39]. Given a single image " + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "inline_equation", + "content": "x^0" + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "content": " and a text prompt " + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "content": ", text-conditioned image-to-video (TI2V) generation aims to syn" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 320, + 255, + 375, + 308 + ], + "blocks": [ + { + "bbox": [ + 320, + 255, + 375, + 308 + ], + "lines": [ + { + "bbox": [ + 320, + 255, + 375, + 308 + ], + "spans": [ + { + "bbox": [ + 320, + 255, + 375, + 308 + ], + "type": "image", + "image_path": "2a552a36a79232e32b7df66896081cb440cc36b38a41e9cba63a86b6e6101bc7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 308, + 529, + 317 + ], + "lines": [ + { + "bbox": [ + 326, + 308, + 529, + 317 + ], + "spans": [ + { + "bbox": [ + 326, + 308, + 529, + 317 + ], + "type": "text", + "content": "\"A man with the expression of slight happiness on his face.\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 376, + 255, + 479, + 308 + ], + "blocks": [ + { + "bbox": [ + 376, + 255, + 479, + 308 + ], + "lines": [ + { + "bbox": [ + 376, + 255, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 376, + 255, + 479, + 308 + ], + "type": "image", + "image_path": "dd64b4d666b3b23a4249fb9b9314ff6ab5fb3f25a1ec5eda2d6d7e7a5757c230.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 480, + 255, + 533, + 308 + ], + "blocks": [ + { + "bbox": [ + 480, + 255, + 533, + 308 + ], + "lines": [ + { + "bbox": [ + 480, + 255, + 533, + 308 + ], + "spans": [ + { + "bbox": [ + 480, + 255, + 533, + 308 + ], + "type": "image", + "image_path": "d2b7a08a4ccf8d6ac829afff081b3aa5308b445a4bd3bcdd6a561a5cb5b7d757.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 320, + 319, + 374, + 373 + ], + "blocks": [ + { + "bbox": [ + 320, + 319, + 374, + 373 + ], + "lines": [ + { + "bbox": [ + 320, + 319, + 374, + 373 + ], + "spans": [ + { + "bbox": [ + 320, + 319, + 374, + 373 + ], + "type": "image", + "image_path": "ab8124efa72cd21ec71fda21bf1d7b17a2db63a69309a7dedb957f5d1868bcd4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 375, + 319, + 479, + 373 + ], + "blocks": [ + { + "bbox": [ + 375, + 319, + 479, + 373 + ], + "lines": [ + { + "bbox": [ + 375, + 319, + 479, + 373 + ], + "spans": [ + { + "bbox": [ + 375, + 319, + 479, + 373 + ], + "type": "image", + "image_path": "3341998c58263fcb3ee286e3c475bdd78408c220aba8b4560a43d8834c2682a7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 373, + 468, + 381 + ], + "lines": [ + { + "bbox": [ + 386, + 373, + 468, + 381 + ], + "spans": [ + { + "bbox": [ + 386, + 373, + 468, + 381 + ], + "type": "text", + "content": "\"A person is drumming.\"" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 480, + 319, + 533, + 373 + ], + "blocks": [ + { + "bbox": [ + 480, + 319, + 533, + 373 + ], + "lines": [ + { + "bbox": [ + 480, + 319, + 533, + 373 + ], + "spans": [ + { + "bbox": [ + 480, + 319, + 533, + 373 + ], + "type": "image", + "image_path": "f82fd823a376b841b9226db745d0fa81ebc9f9ba64f7ae1eb7a189b629d0d551.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 320, + 384, + 375, + 438 + ], + "blocks": [ + { + "bbox": [ + 320, + 384, + 375, + 438 + ], + "lines": [ + { + "bbox": [ + 320, + 384, + 375, + 438 + ], + "spans": [ + { + "bbox": [ + 320, + 384, + 375, + 438 + ], + "type": "image", + "image_path": "7a9448aedbf0e9a0f0ed0937494a78221e422e242cf482ee702896a063f88270.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 438, + 533, + 448 + ], + "lines": [ + { + "bbox": [ + 322, + 438, + 533, + 448 + ], + "spans": [ + { + "bbox": [ + 322, + 438, + 533, + 448 + ], + "type": "text", + "content": "\"A serene mountain cabin covered in a fresh blanket of snow.\"" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 456, + 547, + 521 + ], + "lines": [ + { + "bbox": [ + 304, + 456, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 547, + 521 + ], + "type": "text", + "content": "Figure 1. Examples of generated video frames using our proposed T12V-Zero. The given first image " + }, + { + "bbox": [ + 304, + 456, + 547, + 521 + ], + "type": "inline_equation", + "content": "x^0" + }, + { + "bbox": [ + 304, + 456, + 547, + 521 + ], + "type": "text", + "content": " is highlighted with the red box, and the text condition " + }, + { + "bbox": [ + 304, + 456, + 547, + 521 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 456, + 547, + 521 + ], + "type": "text", + "content": " is shown under each row of the video. The remaining columns show the 6th, 11th, and 16th frames of the generated output videos. Each generated video has 16 frames with a resolution of " + }, + { + "bbox": [ + 304, + 456, + 547, + 521 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 456, + 547, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 376, + 384, + 480, + 437 + ], + "blocks": [ + { + "bbox": [ + 376, + 384, + 480, + 437 + ], + "lines": [ + { + "bbox": [ + 376, + 384, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 376, + 384, + 480, + 437 + ], + "type": "image", + "image_path": "de3ef1e82e3d3559bd02ac5efec8fb5ef96ff3502308f0884ba35e9ec68ac2c8.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 480, + 384, + 533, + 437 + ], + "blocks": [ + { + "bbox": [ + 480, + 384, + 533, + 437 + ], + "lines": [ + { + "bbox": [ + 480, + 384, + 533, + 437 + ], + "spans": [ + { + "bbox": [ + 480, + 384, + 533, + 437 + ], + "type": "image", + "image_path": "bd730428c2e921fad09caaf0a6596fb7577fa91d821f48af29ef50fc427d7fad.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 533, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 665 + ], + "type": "text", + "content": "thesize " + }, + { + "bbox": [ + 304, + 533, + 546, + 665 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 533, + 546, + 665 + ], + "type": "text", + "content": " new frames to yield a realistic video, " + }, + { + "bbox": [ + 304, + 533, + 546, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}} = \\langle x^0,\\hat{x}^1,\\dots ,\\hat{x}^M\\rangle" + }, + { + "bbox": [ + 304, + 533, + 546, + 665 + ], + "type": "text", + "content": ", starting from the given frame " + }, + { + "bbox": [ + 304, + 533, + 546, + 665 + ], + "type": "inline_equation", + "content": "x^0" + }, + { + "bbox": [ + 304, + 533, + 546, + 665 + ], + "type": "text", + "content": " and satisfying the text description " + }, + { + "bbox": [ + 304, + 533, + 546, + 665 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 533, + 546, + 665 + ], + "type": "text", + "content": ". Current TI2V generation methods [59, 63, 70] typically rely on computationally-heavy training on video-text datasets and specific architecture designs to enable text and image conditioning. Some [12, 25] are constrained to specific domains due to the lack of training with large-scale open-domain datasets. Other approaches, such as [14, 67], utilize pretrained foundation models to reduce training costs, but they still need to train additional modules using video data." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we propose TI2V-Zero, which achieves zero-shot TI2V generation using only an open-domain pretrained text-to-video (T2V) latent diffusion model [60]. Here \"zero-shot\" means that when using the diffusion" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 200, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 200, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 200, + 712 + ], + "type": "text", + "content": "*Work done during an internship at MERL." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9015" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "model (DM) that was trained only for text conditioning, our framework enables image conditioning without any optimization, fine-tuning, or introduction of additional modules. Specifically, we guide the generation process by incorporating the provided image " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "x^0" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " into the output latent code at each reverse denoising step. To ensure that the temporal attention layers of the pretrained DM focus on information from the given image, we propose a \"repeat-and-slide\" strategy to synthesize the video in a frame-by-frame manner, rather than directly generating the entire video volume. Notably, TI2V-Zero is not trained for the specific domain of the provided image, thus allowing the model to generalize to any image during inference. Additionally, its autoregressive generation makes the synthesis of long videos possible." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 240, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 289, + 373 + ], + "type": "text", + "content": "While the standard denoising sampling process starting with randomly initialized Gaussian noise can produce matching semantics, it often results in temporally inconsistent videos. Therefore, we introduce an inversion strategy based on the DDPM [20] forward process, to provide a more suitable initial noise for generating each new frame. We also apply a resampling technique [33] in the video DM to help preserve the generated visual details. Our approach ensures that the network maintains temporal consistency, generating visually convincing videos conditioned on the given starting image (see Fig. 1)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 373, + 289, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 289, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 289, + 445 + ], + "type": "text", + "content": "We conduct extensive experiments on MUG [1], UCF-101 [56], and a new open-domain dataset. In these experiments, TI2V-Zero consistently performs well, outperforming a state-of-the-art model [67] that was based on a video diffusion foundation model [8] and was specifically trained to enable open-domain TI2V generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 458, + 134, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 458, + 134, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 458, + 134, + 471 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 479, + 257, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 257, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 257, + 491 + ], + "type": "text", + "content": "2.1. Conditional Image-to-Video Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": "Conditional video generation aims to synthesize videos guided by user-provided signals. It can be classified according to which type(s) of conditions are given, such as text-to-video (T2V) generation [5, 16, 21, 23, 31, 65], video-to-video (V2V) generation [7, 38, 40, 45, 61, 64], and image-to-video (I2V) generation [4, 10, 25, 34, 39, 69]. Here we discuss previous text-conditioned image-to-video (TI2V) generation methods [12, 14, 22, 44, 63, 70]. Hu et al. [25] introduced MAGE, a TI2V generator that integrates a motion anchor structure to store appearance-motion-aligned representations through three-dimensional axial transformers. Yin et al. [70] proposed DragNUWA, a diffusion-based model capable of generating videos controlled by text, image, and trajectory information with three modules including a trajectory sampler, a multi-scale fusion, and an adaptive training strategy. However, these TI2V frameworks require computationally expensive training on videotext datasets and a particular model design to support text" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "and-image-conditioned training. In contrast, our proposed TI2V-Zero leverages a pretrained T2V diffusion model to achieve zero-shot TI2V generation without additional optimization or fine-tuning, making it suitable for a wide range of applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 139, + 533, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 139, + 533, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 139, + 533, + 152 + ], + "type": "text", + "content": "2.2. Adaptation of Diffusion Foundation Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 156, + 547, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 547, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 547, + 517 + ], + "type": "text", + "content": "Due to the recent successful application of diffusion models (DM) [20, 42, 47, 54, 55] to both image and video generation, visual diffusion foundation models have gained prominence. These include text-to-image (T2I) models such as Imagen [50] and Stable Diffusion [47], as well as text-to-video (T2V) models such as ModelScopeT2V [60] and VideoCrafter1 [8]. These models are trained with large-scale open-domain datasets, often including LAION-400M [52] and WebVid-10M [2]. They have shown immense potential for adapting their acquired knowledge base to address a wide range of downstream tasks, thereby reducing or eliminating the need for extensive labeled data. For example, previous works have explored the application of large T2I models to personalized image generation [13, 49], image editing [17, 33, 35-37], image segmentation [3, 68], video editing [45, 62], and video generation [14, 27, 53, 66]. In contrast to T2I models, there are fewer works on the adaptation of large-scale T2V models. Xing et al. [67] proposed DynamicCrafter for open-domain TI2V generation by adapting a T2V foundation model [8]. To control the generative process, they first employed a learnable image encoding network to project the given image into a text-aligned image embedding space. Subsequently, they utilized dual cross-attention layers to fuse text and image information and also concatenated the image with the initial noise to provide the video DM with more precise image details. In contrast, in this paper we explore how to inject the provided image to guide the DM sampling process based solely on the pretrained T2V model itself, with no additional training for the new TI2V task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 526, + 389, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 526, + 389, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 526, + 389, + 540 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "Given one starting image " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "x^0" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " and text " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\langle x^0, x^1, \\ldots, x^M \\rangle" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " represent a real video corresponding to text " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": ". The objective of text-conditioned image-to-video (TI2V) generation is to synthesize a video " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}} = \\langle x^0, \\hat{x}^1, \\ldots, \\hat{x}^M \\rangle" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": ", such that the conditional distribution of " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "x^0" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " is identical to the conditional distribution of " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "x^0" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "p(\\hat{\\mathbf{x}} | x^0, y) = p(\\mathbf{x} | x^0, y)" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": ". Our proposed TI2V-Zero can be built on a pretrained T2V diffusion model with a 3D-UNet-based denoising network. Here we choose ModelScopeT2V [60] as backbone due to its promising open-domain T2V generation ability. Below, we first introduce preliminaries about diffusion models, then introduce the architecture of the pretrained T2V model, and finally present the details of our TI2V-Zero." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "9016" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 225 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 225 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 225 + ], + "type": "image", + "image_path": "cc501315653f24ec3a0c89a54855a81cf4fd0a353a3f95841365890bf17fc7e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": "Figure 2. Illustration of the process of applying TI2V-Zero to generate the new frame " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "\\hat{x}^{i+1}" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": ", given the starting image " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "x^0" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": " and text " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": ". TI2V-Zero is built upon a frozen pretrained T2V diffusion model, including frame encoder " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": ", frame decoder " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": ", and the denoising U-Net " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": ". At the beginning of generation (" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "i=0" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": "), we encode " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "x^0" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "z^0" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": " and repeat it " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": " times to form the queue " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": ". We then apply DDPM-based inversion to " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": " to produce the initial Gaussian noise " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_T" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": ". Subsequently, in each reverse denoising step using U-Net " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": ", we keep replacing the first " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": " frames of " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": " with the noisy latent code " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": " derived from " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": ". Resampling is also applied within each step to improve motion coherence. We finally decode the final frame of the clean latent code " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_0" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": " as the new synthesized frame " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "\\hat{x}^{i+1}" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": ". To compute the new " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": " for the next iteration of generation (" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "i>0" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": "), we perform a sliding operation by dequeuing " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "s_0^0" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": " and enqueuing " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "\\hat{z}_0^K" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 46, + 234, + 547, + 313 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 332, + 221, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 332, + 221, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 332, + 221, + 344 + ], + "type": "text", + "content": "3.1. Preliminaries: Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 287, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 438 + ], + "type": "text", + "content": "Diffusion Models (DM) [20, 54, 55] are probabilistic models designed to learn a data distribution. Here we introduce the fundamental concepts of Denoising Diffusion Probabilistic Models (DDPM). Given a sample from the data distribution " + }, + { + "bbox": [ + 46, + 350, + 287, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0 \\sim q(\\mathbf{z}_0)" + }, + { + "bbox": [ + 46, + 350, + 287, + 438 + ], + "type": "text", + "content": ", the forward diffusion process of a DM produces a Markov chain " + }, + { + "bbox": [ + 46, + 350, + 287, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_1, \\ldots, \\mathbf{z}_T" + }, + { + "bbox": [ + 46, + 350, + 287, + 438 + ], + "type": "text", + "content": " by iteratively adding Gaussian noise to " + }, + { + "bbox": [ + 46, + 350, + 287, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 46, + 350, + 287, + 438 + ], + "type": "text", + "content": " according to a variance schedule " + }, + { + "bbox": [ + 46, + 350, + 287, + 438 + ], + "type": "inline_equation", + "content": "\\beta_1, \\ldots, \\beta_T" + }, + { + "bbox": [ + 46, + 350, + 287, + 438 + ], + "type": "text", + "content": ", that is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 87, + 441, + 287, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 441, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 87, + 441, + 287, + 453 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {z} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}", + "image_path": "4d7a5dbafe743c94900213e138c7b1dfa96fc1308f089d9e1fa7bf81b078c521.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "text", + "content": "where variances " + }, + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "text", + "content": " are constant. When the " + }, + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "text", + "content": " are small, the posterior " + }, + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "inline_equation", + "content": "q(\\mathbf{z}_{t - 1}|\\mathbf{z}_t)" + }, + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "text", + "content": " can be well approximated by a diagonal Gaussian [41, 54]. Furthermore, if the length of the chain, denoted by " + }, + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "text", + "content": ", is sufficiently large, " + }, + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T" + }, + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "text", + "content": " can be well approximated by a standard Gaussian distribution " + }, + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "text", + "content": ". These suggest that the true posterior " + }, + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "inline_equation", + "content": "q(\\mathbf{z}_{t - 1}|\\mathbf{z}_t)" + }, + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "text", + "content": " can be estimated by " + }, + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{z}_{t - 1}|\\mathbf{z}_t)" + }, + { + "bbox": [ + 46, + 455, + 287, + 533 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 534, + 287, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 534, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 93, + 534, + 287, + 547 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {z} _ {t - 1} \\mid \\mathbf {z} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {z} _ {t}\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right), \\tag {2}", + "image_path": "4fd7e4e06f499c95467afbb057f08598b6f468f5efd53fea1a997ce637540e40.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "text", + "content": "where variances " + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "text", + "content": " are also constants. The reverse denoising process in the DM (also termed sampling) then generates samples " + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0 \\sim p_{\\theta}(\\mathbf{z}_0)" + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "text", + "content": " by starting with Gaussian noise " + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "text", + "content": " and gradually reducing noise in a Markov chain " + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{T-1}, \\mathbf{z}_{T-2}, \\ldots, \\mathbf{z}_0" + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "text", + "content": " using a learned " + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{z}_{t-1}|\\mathbf{z}_t)" + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "text", + "content": ". To learn " + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{z}_{t-1}|\\mathbf{z}_t)" + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "text", + "content": ", Gaussian noise " + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "text", + "content": " is first added to " + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "text", + "content": " to generate samples " + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "text", + "content": ". Utilizing the independence property of the noise added at each forward step in Eq. (1), we can calculate the total noise variance as " + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{i=0}^t (1 - \\beta_i)" + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "text", + "content": " and transform " + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 46, + 548, + 287, + 661 + ], + "type": "text", + "content": " in a single step:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 93, + 662, + 287, + 675 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 662, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 93, + 662, + 287, + 675 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {z} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right). \\tag {3}", + "image_path": "c97ac807d6898da316bbf4bee7997438a1a876d918f1667ff9941281af19c67e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 287, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 699 + ], + "type": "text", + "content": "Then a model " + }, + { + "bbox": [ + 47, + 677, + 287, + 699 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 47, + 677, + 287, + 699 + ], + "type": "text", + "content": " is trained to predict " + }, + { + "bbox": [ + 47, + 677, + 287, + 699 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 47, + 677, + 287, + 699 + ], + "type": "text", + "content": " using the following mean-squared error loss:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 61, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "L = \\mathbb {E} _ {t \\sim \\mathcal {U} (1, T), \\mathbf {z} _ {0} \\sim q (\\mathbf {z} _ {0}), \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t) \\| ^ {2} \\right] ， \\tag {4}", + "image_path": "811908554395d1229ffc8f7d31f58b61fb0f90d6fd5261a6a6ca88f9fa085f74.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "text", + "content": "where diffusion step " + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "text", + "content": " is uniformly sampled from " + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\{1, \\dots, T\\}" + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(\\mathbf{z}_t)" + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "text", + "content": " in Eq. (2) can be derived from " + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{z}_t, t)" + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "text", + "content": " to model " + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{z}_{t-1}|\\mathbf{z}_t)" + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "text", + "content": " [20]. The denoising model " + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "text", + "content": " is implemented using a time-conditioned U-Net [48] with residual blocks [15] and self-attention layers [58]. Diffusion step " + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "text", + "content": " is specified to " + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "text", + "content": " by the sinusoidal position embedding [58]. Conditional generation that samples " + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0 \\sim p_{\\theta}(\\mathbf{z}_0|y)" + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "text", + "content": " can be achieved by learning a " + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "text", + "content": "-conditioned model " + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{z}_t, t, y)" + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "text", + "content": " [41, 47] with classifier-free guidance [19]. During training, the condition " + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{z}_t, t, y)" + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "text", + "content": " is replaced by a null label " + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "inline_equation", + "content": "\\emptyset" + }, + { + "bbox": [ + 304, + 332, + 545, + 465 + ], + "type": "text", + "content": " with a fixed probability. When sampling, the output is generated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 309, + 467, + 545, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 467, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 309, + 467, + 545, + 479 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} _ {\\theta} \\left(\\mathbf {z} _ {t}, t, y\\right) = \\epsilon_ {\\theta} \\left(\\mathbf {z} _ {t}, t, \\emptyset\\right) + g \\cdot \\left(\\epsilon_ {\\theta} \\left(\\mathbf {z} _ {t}, t, y\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {z} _ {t}, t, \\emptyset\\right)\\right), \\tag {5}", + "image_path": "675741078b5d4df5970c13ac1ae17cd416fa4727f47e01dca9feb34c58dc9d12.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 309, + 482, + 430, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 482, + 430, + 494 + ], + "spans": [ + { + "bbox": [ + 309, + 482, + 430, + 494 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 309, + 482, + 430, + 494 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 309, + 482, + 430, + 494 + ], + "type": "text", + "content": " is the guidance scale." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 502, + 509, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 509, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 509, + 514 + ], + "type": "text", + "content": "3.2. Architecture of Pretrained T2V Model" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "text", + "content": "TI2V-Zero can be built upon a pretrained T2V diffusion model with a 3D-UNet-based denoising network. Here we choose ModelScopeT2V [60] as the pretrained model (denoted " + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 521, + 545, + 568 + ], + "type": "text", + "content": "). We now describe this T2V model in detail." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Structure Overview. Given a text prompt " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ", the T2V model " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " synthesizes a video " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}} = \\langle \\hat{x}^0, \\hat{x}^1, \\dots, \\hat{x}^K \\rangle" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " with a pre-defined video of length " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "(K + 1)" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " using a latent video diffusion model. Similar to Latent Diffusion Models (LDM) [47], " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " incorporates a frame auto-encoder [11, 28] for the conversion of data between pixel space " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " and latent space " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " through its encoder " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ". Given the real video " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\langle x^0, x^1, \\dots, x^K \\rangle" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " first utilizes the frame encoder " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " to encode the video " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\langle z^0, z^1, \\dots, z^K \\rangle" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ". Here the sizes of pixel frame " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " and latent frame " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "H_x \\times W_x \\times 3" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "H_z \\times W_z \\times C_z" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ", respectively. To be consistent with the notation used for the DM, we denote the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9017" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 47, + 87, + 289, + 407 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 274, + 84 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 274, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 274, + 84 + ], + "type": "text", + "content": "Algorithm 1 Generation using our TI2V-Zero approach." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "lines": [ + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": "Input: The starting frame " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "x^0" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": "; The text prompt " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": "; The pretrained T2V Model " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " for generating " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "(K + 1)" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": "-frame videos, including frame encoder " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " and frame decoder " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": ", and the DM de-noising networks " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": "; The iteration number " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " for resampling; The parameter " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " to control the length of the output video. \nOutput: A synthesized video " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "(M + 1)" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " frames. \n1: " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "z^0 \\gets \\mathcal{E}(x^0)" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " // Encode " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "x^0" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " \n2: " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0 \\gets \\langle z^0, z^0, \\dots, z^0 \\rangle" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " // Repeat " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "z^0" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " times \n3: " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}} \\gets \\langle x^0 \\rangle" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " \n4: for " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "i = 1, 2, \\dots, M" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " do \n// Generate one new frame " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\hat{x}^i" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_T \\sim \\mathcal{N}(\\sqrt{\\bar{\\alpha}_T} \\mathbf{s}_0, (1 - \\bar{\\alpha}_T) \\mathbf{I})" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " // DDPM Inversion \n6: " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\hat{z}_T^K \\sim \\mathcal{N}(\\sqrt{\\bar{\\alpha}_T} s_0^{K-1}, (1 - \\bar{\\alpha}_T) \\mathbf{I})" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_T \\gets \\mathbf{s}_T \\cup \\hat{z}_T^K" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " // Initialize " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_T" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " \n8: for " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "t = T-1, \\dots, 2, 1" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " do \n" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t \\sim \\mathcal{N}(\\sqrt{\\bar{\\alpha}_t} \\mathbf{s}_0, (1 - \\bar{\\alpha}_t) \\mathbf{I})" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "u = 1, 2, \\dots, U" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " do \n" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\langle \\hat{z}_t^0, \\hat{z}_t^1, \\dots, \\hat{z}_t^{K-1} \\rangle \\gets \\mathbf{s}_t" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " // Replace \n" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{t-1} \\sim \\mathcal{N}(\\mu_\\theta (\\hat{\\mathbf{z}}_t, y), \\sigma_t^2 \\mathbf{I})" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " \nif " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "u < U" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "t > 1" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " then \n" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t \\sim \\mathcal{N}(\\sqrt{1 - \\beta_t} \\hat{\\mathbf{z}}_{t-1}, \\beta_t \\mathbf{I})" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " // Resample \nend if \nend for \nend for \n" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0 \\gets \\langle s_0^1, s_0^2, \\dots, s_0^{K-1} \\rangle \\cup \\hat{z}_0^K" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " // Slide \n" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\hat{x}^i \\gets D(\\hat{z}_0^K)" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " // Decode " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\hat{z}_0^K" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}} \\gets \\hat{\\mathbf{x}} \\cup \\hat{x}^i" + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "text", + "content": " \nend for \nreturn " + }, + { + "bbox": [ + 47, + 87, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "text", + "content": "clean video latent " + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\mathbf{z}_0 = \\langle z_0^0, z_0^1, \\ldots, z_0^K \\rangle" + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "text", + "content": " then learns a DM on the latent space " + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "text", + "content": " through a 3D denoising U-Net " + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "text", + "content": " [9]. Let " + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t = \\langle z_t^0, z_t^1, \\ldots, z_t^K \\rangle" + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "text", + "content": " represent the latent sequence that results from adding noise over " + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "text", + "content": " steps to the original latent sequence " + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "text", + "content": ". When training, the forward diffusion process of a DM transforms the initial latent sequence " + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T" + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "text", + "content": " by iteratively adding Gaussian noise " + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "text", + "content": " steps. During inference, denoising U-Net " + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "text", + "content": " predicts the added noise at each step, enabling the generation of the clean latent sequence " + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_0 = \\langle \\hat{z}_0^0, \\hat{z}_0^1, \\ldots, \\hat{z}_0^K \\rangle" + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "text", + "content": " starting from randomly sampled Gaussian noise " + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 46, + 433, + 287, + 566 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "text", + "content": "Text Conditioning Mechanism. " + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "text", + "content": " employs a cross-attention mechanism [47] to incorporate text information into the generative process as guidance. Specifically, " + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "text", + "content": " uses a pretrained CLIP model [46] to encode the prompt " + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "text", + "content": " as the text embedding " + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "text", + "content": ". The embedding " + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "text", + "content": " is later used as the key and value in the multi-head attention layer within the spatial attention blocks, thus enabling the integration of text features with the intermediate U-Net features in " + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 567, + 287, + 663 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Denoising U-Net. The denoising U-Net " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " includes four key building blocks: the initial block, the downsampling block, the spatio-temporal block, and the upsampling block. The initial block transfers the input into the embedding" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 78, + 536, + 318 + ], + "blocks": [ + { + "bbox": [ + 481, + 70, + 517, + 78 + ], + "lines": [ + { + "bbox": [ + 481, + 70, + 517, + 78 + ], + "spans": [ + { + "bbox": [ + 481, + 70, + 517, + 78 + ], + "type": "text", + "content": "Video Time" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 78, + 536, + 318 + ], + "lines": [ + { + "bbox": [ + 316, + 78, + 536, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 78, + 536, + 318 + ], + "type": "image", + "image_path": "4be811a2a4ed11d79312b906b57c48e810421e0172805ae5595114d846b4ecc4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 327, + 547, + 427 + ], + "lines": [ + { + "bbox": [ + 304, + 327, + 547, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 547, + 427 + ], + "type": "text", + "content": "Figure 3. Illustration of the motivation behind our framework. We explore the application of a replacing-based baseline approach (rows 2-4, labeled \"Replacing\") and our TI2V-Zero (rows 5-6, labeled \"TI2V-Zero\") in various video generation tasks. The given real frames for each task are highlighted by red boxes and the text input is shown under the block. The replacing-based approach is only effective at predicting a single frame when all the other frames in the video are provided, while TI2V-Zero generates temporally coherent videos for both the TI2V and video infilling tasks." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 439, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 439, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 439, + 545, + 513 + ], + "type": "text", + "content": "space, while the downsampling and upsampling blocks are responsible for spatially downsampling and upsampling the feature maps. The spatio-temporal block is designed to capture spatial and temporal dependencies in the latent space, which comprises 2D spatial convolution, 1D temporal convolution, 2D spatial attention, and 1D temporal attention." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 521, + 406, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 521, + 406, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 521, + 406, + 533 + ], + "type": "text", + "content": "3.3. Our Framework" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "content": "Leveraging the pretrained T2V foundation model " + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "content": ", we first propose a straightforward replacing-based baseline for adapting " + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "content": " to TI2V generation. We then analyze the possible reasons why it fails and introduce our TI2V-Zero framework, which includes a repeat-and-slide strategy, DDPM-based inversion, and resampling. Figure 2 and Algorithm 1 demonstrate the inference process of TI2V-Zero." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "text", + "content": "Replacing-based Baseline. We assume that the pretrained model " + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "text", + "content": " is designed to generate the video with a fixed length of " + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "inline_equation", + "content": "(K + 1)" + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "text", + "content": ". So we first consider synthesizing videos with that same length " + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "inline_equation", + "content": "(K + 1)" + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "inline_equation", + "content": "M = K" + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "text", + "content": ". Since the DM process operates within the latent space " + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "text", + "content": ", we use the encoder " + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "text", + "content": " to map the given starting frame " + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "inline_equation", + "content": "x^0" + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "text", + "content": " into the latent representation " + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "inline_equation", + "content": "z^0" + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "text", + "content": ". Additionally, we denote " + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "inline_equation", + "content": "z^0 = z_0^0" + }, + { + "bbox": [ + 304, + 625, + 546, + 714 + ], + "type": "text", + "content": " to specify that the latent is clean and corresponds" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9018" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "content": "to diffusion step 0 of the DM. Note that each reverse denoising step in Eq. (2) from " + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{t - 1}" + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "content": " depends solely on " + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t = \\langle \\hat{z}_t^0,\\hat{z}_t^1,\\dots ,\\hat{z}_t^K\\rangle" + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "content": " . To ensure that the first frame of the final synthesized clean video latent " + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_0 = \\langle \\hat{z}_0^0,\\hat{z}_0^1,\\dots ,\\hat{z}_0^K\\rangle" + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "content": " at step 0 matches the provided image latent, i.e., " + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\hat{z}_0^0 = z_0^0" + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "content": " we can modify the first generated latent " + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^0" + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "content": " at each reverse step, as long as the signal-to-noise ratio of each frame latent in " + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "content": " remains consistent. Using Eq. (3), we can add " + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "content": " steps of noise to the provided image latent " + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "z_0^0" + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "content": " , allowing us to sample " + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "z_{t}^{0}" + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "content": " through a single-step calculation. By replacing the first generated latent " + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^0" + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "content": " with the noisy image latent " + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "z_{t}^{0}" + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "content": " at each reverse denoising step, we might expect that the video generation process can be guided by " + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "z_0^0" + }, + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "content": " with the following expressions defined for each reverse step:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 110, + 232, + 287, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 232, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 110, + 232, + 287, + 245 + ], + "type": "interline_equation", + "content": "z _ {t} ^ {0} \\sim \\mathcal {N} \\left(\\sqrt {\\bar {\\alpha}} _ {t} z _ {0} ^ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {6a}", + "image_path": "2c44548737003b2e8a4885af28c7168a52627ee8fe627abe89e4ccfe267c8cda.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 111, + 247, + 287, + 260 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 247, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 111, + 247, + 287, + 260 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {t} ^ {0} \\leftarrow z _ {t} ^ {0}, \\tag {6b}", + "image_path": "9b1f3a3f62ed4db6a3e8d4fe85389c2744e6178ca057a4fc15d4fdce44c7e3c1.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 263, + 287, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 263, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 110, + 263, + 287, + 277 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {z}} _ {t - 1} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} \\left(\\hat {\\mathbf {z}} _ {t}, y\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right). \\tag {6c}", + "image_path": "7eaed2e97e88a125ca98ac716baa82692b1531eaab41385bfa6330a1d184fcb5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "text", + "content": "Specifically, in each reverse step from " + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{t - 1}" + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "text", + "content": ", as shown in Eq. (6a), we first compute the noisy latent " + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "inline_equation", + "content": "z_{t}^{0}" + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "text", + "content": " by adding Gaussian noise to the given image latent " + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "inline_equation", + "content": "z_0^0" + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "text", + "content": " steps. Then, we replace the first latent " + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^0" + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "inline_equation", + "content": "z_{t}^{0}" + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "text", + "content": " in Eq. (6b) to incorporate the provided image into the generation process. Finally, in Eq. (6c), we pass " + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "text", + "content": " through the denoising network to generate " + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{t - 1}" + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "text", + "content": ", where the text " + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "text", + "content": " is integrated by classifier-free guidance (Eq. (5)). After " + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "text", + "content": " iterations, the final clean latent " + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_0" + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "text", + "content": " at diffusion step 0 can be mapped back into the image space " + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "text", + "content": " using the decoder " + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 283, + 287, + 402 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 403, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 498 + ], + "type": "text", + "content": "Using this replacing-based baseline, we might expect that the temporal attention layers in " + }, + { + "bbox": [ + 46, + 403, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 403, + 287, + 498 + ], + "type": "text", + "content": " can utilize the context provided by the first frame latent " + }, + { + "bbox": [ + 46, + 403, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^0" + }, + { + "bbox": [ + 46, + 403, + 287, + 498 + ], + "type": "text", + "content": " to generate the subsequent frame latents in a manner that harmonizes with " + }, + { + "bbox": [ + 46, + 403, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^0" + }, + { + "bbox": [ + 46, + 403, + 287, + 498 + ], + "type": "text", + "content": ". However, as shown in Fig. 3, row 2, this replacing-based approach fails to produce a video that is temporally consistent with the first image. The generated frames are consistent with each other, but not with the provided first frame." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "text", + "content": "To analyze possible reasons for failure, we apply this baseline to a simpler video infilling task, where every other frame is provided and the model needs to predict the interspersed frames. In this case, the baseline replaces the generated frame latents at positions corresponding to real frames with noisy provided-frame latents in each reverse step. The resulting video, in Fig. 3, row 3, looks like a combination of two independent videos: the generated (even) frames are consistent with each other but not with the provided (odd) frames. We speculate that this may result from the intrinsic dissimilarity between frame latents derived from the given real images and those sampled from " + }, + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "text", + "content": ". Thus, the temporal attention values between frame latents sampled in the same way (both from the given images or both from " + }, + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "text", + "content": ") will be higher, while the attention values between frame latents sampled in different ways (one from the given image and the other from " + }, + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "text", + "content": ") will be lower. Therefore, the temporal attention layers of " + }, + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "text", + "content": " tend to utilize the information from latents" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 319, + 72, + 533, + 339 + ], + "blocks": [ + { + "bbox": [ + 319, + 72, + 533, + 339 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 533, + 339 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 533, + 339 + ], + "type": "image", + "image_path": "56a3fa54e63647ab3a9f3f42593a71457e609a3ccf459ff78b944420e667deec.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 348, + 545, + 426 + ], + "lines": [ + { + "bbox": [ + 305, + 348, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 545, + 426 + ], + "type": "text", + "content": "Figure 4. Qualitative ablation study comparing different sampling strategies for our TI2V-Zero on MUG. The first image " + }, + { + "bbox": [ + 305, + 348, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\hat{x}^0" + }, + { + "bbox": [ + 305, + 348, + 545, + 426 + ], + "type": "text", + "content": " is highlighted with the red box and text " + }, + { + "bbox": [ + 305, + 348, + 545, + 426 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 348, + 545, + 426 + ], + "type": "text", + "content": " is shown under the block. The 1st, 6th, 11th, and 16th frames of the videos are shown in each column. The terms Inversion, DDIM, and Resample denote the application of DDPM inversion, the steps using DDIM sampling, and the iteration number using resampling, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "type": "text", + "content": "produced by " + }, + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "type": "text", + "content": " to synthesize new frames at each reverse step, ignoring the provided frames. We further simplify the task to single-frame prediction, where the model only needs to predict a single frame when all the other frames in the video are given. In this setting, all the frame latents except for the final frame are replaced by noisy provided-frame latents in each reverse step. Thus, temporal attention layers can only use information from the real frames. In this case, Fig. 3, row 4, shows that the baseline can now generate a final frame that is consistent with the previous frames." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "text", + "content": "Repeat-and-Slide Strategy. Inspired by the observation in Fig. 3, to guarantee that the temporal attention layers of " + }, + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "text", + "content": " depend solely on the given image, we make two major changes to the proposed replacing-based baseline: (1) instead of using " + }, + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "text", + "content": " to directly synthesize the entire " + }, + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "inline_equation", + "content": "(K + 1)" + }, + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "text", + "content": "-frame video, we switch to a frame-by-frame generation approach, i.e., we generate only one new frame latent in each complete DM sampling process; (2) for each sampling process generating the new frame latent, we ensure that only one frame latent is produced from " + }, + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "text", + "content": ", while the other " + }, + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "text", + "content": " frame latents are derived from the given real image and previously synthesized frames, thereby forcing temporal attention layers to only use the information from these frame latents. Specifically, we construct a queue of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9019" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": " frame latents, denoted as " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0 = \\langle s_0^0,s_0^1,\\dots ,s_0^{K - 1}\\rangle" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": ". We also define " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t = \\langle s_t^0,s_t^1,\\dots ,s_t^{K - 1}\\rangle" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": ", which is obtained by adding " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": " steps of Gaussian noise to the clean " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": ". Similar to our replacing-based baseline in the single-frame prediction task, in each reverse step from " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{t - 1}" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": ", we replace the first " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": " frame latents in " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": ". Consequently, the temporal attention layers have to utilize information from " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": " to synthesize the new frame's latent, " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "\\hat{z}_0^K" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": ". Considering that only one starting image latent " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "z^0" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": " is provided, we propose a \"repeat-and-slide\" strategy to construct " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": ". At the beginning of video generation, we repeat " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "z^0" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": " frames to form " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": ", and gradually perform a sliding operation within the queue " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": " by dequeuing the first frame latent " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "s_0^0" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": " and enqueuing the newly generated latent " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "\\hat{z}_0^K" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": " after each complete DM sampling process. Note that though the initial " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": " is created by repeating " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "z^0" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": ", the noise added to get " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": " is different for each frame's latent in " + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t" + }, + { + "bbox": [ + 47, + 71, + 289, + 283 + ], + "type": "text", + "content": ", thus ensuring diversity. The following expressions define one reverse step in the DM sampling process:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 110, + 290, + 287, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 290, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 110, + 290, + 287, + 304 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {t} \\sim \\mathcal {N} \\left(\\sqrt {\\bar {\\alpha}} _ {t} \\mathbf {s} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {7a}", + "image_path": "4f7c5742b56b633428a7220f713927976f64bf8a3be2d436b5250f995521ddac.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 112, + 305, + 287, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 305, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 112, + 305, + 287, + 319 + ], + "type": "interline_equation", + "content": "\\left\\langle \\hat {z} _ {t} ^ {0}, \\hat {z} _ {t} ^ {1}, \\dots , \\hat {z} _ {t} ^ {K - 1} \\right\\rangle \\leftarrow \\mathbf {s} _ {t}, \\tag {7b}", + "image_path": "3831aa3dcae13e09ef4b360a0536fbaaec34ce33af4f4cdb72dc38df5f9896e4.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 321, + 287, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 321, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 111, + 321, + 287, + 335 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {z}} _ {t - 1} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} \\left(\\hat {\\mathbf {z}} _ {t}, y\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right). \\tag {7c}", + "image_path": "e66731cceef32f6bc74d6e919cfdc3410ed90354d9d547075d77c32aa57771a0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "text", + "content": "Specifically, in each reverse denoising step from " + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{t - 1}" + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "text", + "content": ", we first add " + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "text", + "content": " steps of Gaussian noise to the queue " + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "text", + "content": " to yield " + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t" + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "text", + "content": " in Eq. (7a). Subsequently, we replace the previous " + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "text", + "content": " frames of " + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t" + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "text", + "content": " in Eq. (7b) and input " + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_t" + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "text", + "content": " to the denoising network to produce the less noisy latent " + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{t - 1}" + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "text", + "content": " (Eq. (7c))." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 403, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 287, + 461 + ], + "type": "text", + "content": "With the repeat-and-slide strategy, model " + }, + { + "bbox": [ + 47, + 403, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 403, + 287, + 461 + ], + "type": "text", + "content": " is tasked with predicting only one new frame, while the preceding " + }, + { + "bbox": [ + 47, + 403, + 287, + 461 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 403, + 287, + 461 + ], + "type": "text", + "content": " frames are incorporated into the reverse process to ensure that the temporal attention layers depend solely on information derived from the provided image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": "DDPM-based Inversion. Though the DM sampling process starting with randomly sampled Gaussian noise produces matching semantics, the generated video is often temporally inconsistent (Fig. 4, row 2). To provide initial noise that can produce more temporally consistent results, we introduce an inversion strategy based on the DDPM [20] forward process when generating the new frame latent. Specifically, at the beginning of each DM sampling process to synthesize the new frame latent " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\hat{z}_0^K" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": ", instead of starting with the " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_T" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": " randomly sampled from " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": ", we add " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": " full steps of Gaussian noise to " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_T" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": " using Eq. (3). Note that " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": " frames, while " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": " frames. We then use " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_T" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": " to initialize the first " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": " frames of " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_T" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": ". We copy the last frame " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "s_T^{K - 1}" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_T" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": " to initialize the final frame " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\hat{z}_T^K" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": ", as the " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "(K - 1)\\mathrm{th}" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": " frame is the closest to the " + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 462, + 288, + 641 + ], + "type": "text", + "content": "th frame." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "text", + "content": "Resampling. Similar to [24, 33], we further apply a resampling technique, which was initially designed for the image inpainting task, to the video DM to enhance motion coherence. Particularly, after performing a one-step denoising operation in the reversed process, we add one-step noise again to revert the latent. This procedure is repeated mul" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 547, + 125 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 547, + 125 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 547, + 125 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 547, + 125 + ], + "type": "table", + "html": "
InversionDDIMResampleFVD↓sFVD↓tFVD↓
X1001656.372074.77±411.741798.05±235.34
✓100339.89443.97±139.10405.22±61.58
✓500463.55581.32±234.09535.06±85.27
✓102207.62299.14±87.24278.73±47.84
✓104180.09267.17±74.72252.77±39.02
", + "image_path": "5b1d8355d50eefbe5f759ef960183fdece5e7e2f6602a266b18fe9a94a97c1b8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 198, + 535, + 249 + ], + "blocks": [ + { + "bbox": [ + 305, + 132, + 547, + 188 + ], + "lines": [ + { + "bbox": [ + 305, + 132, + 547, + 188 + ], + "spans": [ + { + "bbox": [ + 305, + 132, + 547, + 188 + ], + "type": "text", + "content": "Table 1. Quantitative ablation study comparing different sampling strategies for proposed TI2V-Zero on the MUG dataset. Inversion, DDIM, and Resample denote the application of DDPM-based inversion, the steps using DDIM sampling, and the iteration number using resampling, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 198, + 535, + 249 + ], + "lines": [ + { + "bbox": [ + 318, + 198, + 535, + 249 + ], + "spans": [ + { + "bbox": [ + 318, + 198, + 535, + 249 + ], + "type": "table", + "html": "
Distributions for ComparisonFVD↓tFVD↓
TI2V-Zero-Fake vs. ModelScopeT2V366.41921.31±251.85
TI2V-Zero-Real vs. Real Videos477.191306.75±271.82
ModelScopeT2V vs. Real Videos985.822264.08±501.28
TI2V-Zero-Fake vs. Real Videos937.112177.70±436.71
", + "image_path": "6024d8e42df2a9a00dc2ab17fa2698aa3bb0491442a95921910bfb8da50488ce.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 256, + 546, + 289 + ], + "lines": [ + { + "bbox": [ + 305, + 256, + 546, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 256, + 546, + 289 + ], + "type": "text", + "content": "Table 2. Result analysis of TI2V-Zero starting from the real (i.e., TI2V-Zero-Real) or synthesized frames (i.e., TI2V-Zero-Fake) on the UCF101 dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 301, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 301, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 301, + 545, + 338 + ], + "type": "text", + "content": "tiple times for each diffusion step, ensuring harmonization between the predicted and conditioning frame latents (see Algorithm 1 for details)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 350, + 388, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 350, + 388, + 364 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 388, + 364 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 370, + 427, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 427, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 427, + 382 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 389, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 389, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 305, + 389, + 545, + 425 + ], + "type": "text", + "content": "We conduct comprehensive experiments on three datasets. More details about datasets, such as selected subjects and text prompts, can be found in our Supplementary Materials." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 426, + 546, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 546, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 546, + 568 + ], + "type": "text", + "content": "MUG facial expression dataset [1] contains 1,009 videos of 52 subjects performing 7 different expressions. We include this dataset to evaluate the performance of models in scenarios with small motion and a simple, unchanged background. To simplify the experiments, we randomly select 5 male and 5 female subjects, and 4 expressions. We use the text prompt templates like \"A woman with the expression of slight {label} on her face.\" to change the expression class label to be text input. Since the expressions shown in the videos of MUG are often not obvious, we add \"slight\" in the text input to avoid large motion." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "type": "text", + "content": "UCF101 action recognition dataset [56] contains 13,320 videos from 101 human action classes. We include this dataset to measure performance under complicated motion and complex, changing backgrounds. To simplify the experiments, we select 10 action classes and the first 10 subjects within each class. We use text prompt templates such as \"A person is performing {label}.\" to change the class label to text input." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 666, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 666, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 666, + 547, + 714 + ], + "type": "text", + "content": "In addition to the above two datasets, we create an OPEN dataset to assess the model's performance in open-domain TI2V generation. We first utilize ChatGPT [43] to generate 10 text prompts. Subsequently, we employ Stable" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9020" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 242 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 242 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 242 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 242 + ], + "type": "image", + "image_path": "60831c16be81240768b095d1a822dfd337e1bba6492c2354dd01a4e53c9eb508.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 249, + 547, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 249, + 547, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 547, + 283 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison among different methods on multiple datasets for TI2V generation. Columns in each block display the 1st, 6th, 11th, and 16th frames of the output videos, respectively. There are 16 frames with a resolution of " + }, + { + "bbox": [ + 46, + 249, + 547, + 283 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 249, + 547, + 283 + ], + "type": "text", + "content": " for each video. The given image " + }, + { + "bbox": [ + 46, + 249, + 547, + 283 + ], + "type": "inline_equation", + "content": "x^0" + }, + { + "bbox": [ + 46, + 249, + 547, + 283 + ], + "type": "text", + "content": " is highlighted with the red box and the text prompt " + }, + { + "bbox": [ + 46, + 249, + 547, + 283 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 249, + 547, + 283 + ], + "type": "text", + "content": " is shown under each block." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 97, + 293, + 499, + 350 + ], + "blocks": [ + { + "bbox": [ + 97, + 293, + 499, + 350 + ], + "lines": [ + { + "bbox": [ + 97, + 293, + 499, + 350 + ], + "spans": [ + { + "bbox": [ + 97, + 293, + 499, + 350 + ], + "type": "table", + "html": "
ModelMUGUCF101
FVD↓sFVD↓tFVD↓FVD↓tFVD↓
DynamiCrafter [67]1094.721359.86±257.731223.89±105.94589.591540.02±199.59
TI2V-Zero w/o Resample (Ours)339.89443.97±139.10405.22±61.58493.191319.77±283.87
TI2V-Zero w/ Resample (Ours)180.09267.17±74.72252.77±39.02477.191306.75±271.82
", + "image_path": "0b4b9f03fc32a6c3c1d3bd0718643eb1f7b3d4124f27a4274b3445b93fcc2ed6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 357, + 479, + 369 + ], + "lines": [ + { + "bbox": [ + 113, + 357, + 479, + 369 + ], + "spans": [ + { + "bbox": [ + 113, + 357, + 479, + 369 + ], + "type": "text", + "content": "Table 3. Quantitative comparison among different methods on multiple datasets for TI2V generation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 389, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 426 + ], + "type": "text", + "content": "Diffusion 1.5 [47] to synthesize 100 images from each text prompt, generating a total of 1,000 starting images and 10 text prompts for evaluating TI2V models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 498 + ], + "type": "text", + "content": "Data Preprocessing. We resize all the videos/images to " + }, + { + "bbox": [ + 46, + 426, + 287, + 498 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 426, + 287, + 498 + ], + "type": "text", + "content": " resolution. For UCF101, since most of the video frames are not square, we crop the central part of the frames. To obtain ground truth videos for computing metrics, we uniformly sample 16 frames from each video in the datasets to generate the video clips with a fixed length." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": "Metrics. Following prior work [21, 22, 25], we assess the visual quality, temporal coherence, and sample diversity of generated videos using Fréchet Video Distance (FVD) [57]. Similar to Fréchet Inception Distance (FID) [18], which is used for image quality evaluation, FVD utilizes a video classification network I3D [6] pretrained on Kinetics-400 dataset [26] to extract feature representation of real and synthesized videos. Then it calculates the Fréchet distance between the distributions of the real and synthesized video features. To measure how well a generated video aligns with the text prompt " + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": " (condition accuracy) and the given image " + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": " (subject relevance), following [39], we design two variants of FVD, namely text-conditioned FVD (tFVD) and subject-conditioned FVD (sFVD). tFVD and sFVD compare the distance between real and synthesized video feature distributions under the same text " + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": " or the same subject image " + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": ", respectively. We first compute tFVD and sFVD for each condition " + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": " and image " + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": ", then report their mean" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 389, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 497 + ], + "type": "text", + "content": "and variance as final results. In our experiments, we generate 1,000 videos for all the models to estimate the feature distributions. We compute both tFVD and sFVD on the MUG dataset, but for UCF101, we only consider tFVD since it doesn't contain videos of different actions for the same subject. For the OPEN dataset, we only present qualitative results due to the lack of ground truth videos. Unless otherwise specified, all the generated videos are 16 frames (i.e., " + }, + { + "bbox": [ + 304, + 389, + 545, + 497 + ], + "type": "inline_equation", + "content": "M = 15" + }, + { + "bbox": [ + 304, + 389, + 545, + 497 + ], + "type": "text", + "content": ") with resolution " + }, + { + "bbox": [ + 304, + 389, + 545, + 497 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 389, + 545, + 497 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 503, + 440, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 503, + 440, + 516 + ], + "spans": [ + { + "bbox": [ + 305, + 503, + 440, + 516 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "content": "Model Implementation. We take the ModelScopeT2V 1.4.2 [60] as basis and implement our modifications. For text-conditioned generation, we employ classifier-free guidance with " + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "inline_equation", + "content": "g = 9.0" + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "content": " in Eq. (5). Determined by our preliminary experiments, we choose 10-step DDIM and 4-step resampling as the default setting for MUG and OPEN, and 50-step DDIM and 2-step resampling for UCF101." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 606, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 546, + 713 + ], + "type": "text", + "content": "Implementation of SOTA Model. We compare our TI2V-Zero with a state-of-the-art (SOTA) model DynamamiCrafter, a recent open-domain TI2V framework [67]. DynamamiCrafter is based on a large-scale pretrained T2V foundation model VideoCrafter1 [16]. It introduces a learnable projection network to enable image-conditioned generation and then fine-tunes the entire framework. We implement DynamamiCrafter using their provided code with their default settings. For a fair comparison, all the generated videos are" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9021" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 272, + 173 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 272, + 173 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 272, + 173 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 272, + 173 + ], + "type": "image", + "image_path": "1fe46997d8682b6916475eed46a7cce2607959dffa8f63729ccc7d5ffd8b019a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 175, + 269, + 184 + ], + "lines": [ + { + "bbox": [ + 66, + 175, + 269, + 184 + ], + "spans": [ + { + "bbox": [ + 66, + 175, + 269, + 184 + ], + "type": "text", + "content": "\"A mesmerizing display of the northern lights in the Arctic.\"" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 194, + 287, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 287, + 251 + ], + "type": "text", + "content": "Figure 6. Example of long video generation using our TI2V-Zero on the OPEN dataset. The given image " + }, + { + "bbox": [ + 46, + 194, + 287, + 251 + ], + "type": "inline_equation", + "content": "x^0" + }, + { + "bbox": [ + 46, + 194, + 287, + 251 + ], + "type": "text", + "content": " is highlighted with a red box, and the text prompt " + }, + { + "bbox": [ + 46, + 194, + 287, + 251 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 194, + 287, + 251 + ], + "type": "text", + "content": " is shown under the set of frames. There are a total of 128 video frames (" + }, + { + "bbox": [ + 46, + 194, + 287, + 251 + ], + "type": "inline_equation", + "content": "M = 127" + }, + { + "bbox": [ + 46, + 194, + 287, + 251 + ], + "type": "text", + "content": "), and the synthesized results for every 14 frames are presented." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 263, + 225, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 263, + 225, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 225, + 275 + ], + "type": "text", + "content": "centrally-cropped and resized to " + }, + { + "bbox": [ + 47, + 263, + 225, + 275 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 47, + 263, + 225, + 275 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 285, + 142, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 142, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 142, + 298 + ], + "type": "text", + "content": "4.3. Result Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 304, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 460 + ], + "type": "text", + "content": "Ablation Study. We conduct ablation study of different sampling strategies on MUG. As shown in Tab. 1 and Fig. 4, compared with generating using randomly sampled Gaussian noise, initializing the input noise with DDPM inversion is important for generating temporally continuous videos, improving all of the metrics dramatically. For MUG, increasing the DDIM sampling steps from 10 to 50 does not enhance the video quality but requires more inference time. Thus, we choose 10-step DDIM as the default setting on MUG. As shown in Fig. 4 and Tab. 1, adding resampling helps preserve identity details (e.g., hairstyle and facial appearance), resulting in lower FVD scores. Increasing resampling steps from 2 to 4 further improves FVD scores." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 461, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 286, + 675 + ], + "type": "text", + "content": "Effect of Real/Synthesized Starting Frames. We also explore the effect of video generation starting with real or synthesized frames on UCF101. We initially use the first frame of the real videos to generate videos with our TI2V-Zero, termed TI2V-Zero-Real. Additionally, we utilize the backbone model ModelScopeT2V [60] to generate synthetic videos using the text inputs of UCF101. We then employ TI2V-Zero to create videos from the first frame of the generated fake videos, denoted as TI2V-Zero-Fake. As shown in Tab. 2, [TI2V-Zero-Fake vs. ModelScopeT2V] can achieve better FVD scores than [TI2V-Zero-Real vs. Real Videos]. The reason may be that frames generated by ModelScopeT2V can be considered as in-distribution data since TI2V-Zero is built upon it. We also compare the output video distribution of TI2V-Zero-Fake and ModelScopeT2V with real videos in Tab. 2. Though starting from the same synthesized frames, TI2V-Zero-Fake can generate more realistic videos than the backbone model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "Comparison with SOTA Model. We compare our proposed TI2V-Zero with DynamiCrafter [67] in Tab. 3 and Fig. 5. From Fig. 5, one can find that DynamiCrafter" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "struggles to preserve details from the given image, and the motion of its generated videos is also less diverse. Note that DynamiCrafter requires additional fine-tuning to enable TI2V generation. In contrast, without using any fine-tuning or introducing external modules, our proposed TI2V-Zero can precisely start with the given image and output more visually-pleasing results, thus achieving much better FVD scores on both MUG and UCF101 datasets in Tab. 3. The comparison between our TI2V-Zero models with and without using resampling in Fig. 5 and Tab. 3 also demonstrates the effectiveness of using resampling, which can help maintain identity and background details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 221, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 545, + 377 + ], + "type": "text", + "content": "Extension to Other Applications. TI2V-Zero can also be extended to other tasks as long as we can construct " + }, + { + "bbox": [ + 304, + 221, + 545, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 304, + 221, + 545, + 377 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 221, + 545, + 377 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 221, + 545, + 377 + ], + "type": "text", + "content": " images at the beginning. These images can be obtained either from ground truth videos or by applying the repeating operation. Then we can slide " + }, + { + "bbox": [ + 304, + 221, + 545, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 304, + 221, + 545, + 377 + ], + "type": "text", + "content": " when generating the subsequent frames. We have applied TI2V-Zero in video infilling (see the last row in Fig. 3), video prediction (see Supplementary Materials), and long video generation (see Fig. 6). As shown in Fig. 6, when generating a 128-frame video on the OPEN dataset, our method can preserve the mountain shape in the background, even at the 71st frame (frame " + }, + { + "bbox": [ + 304, + 221, + 545, + 377 + ], + "type": "inline_equation", + "content": "\\hat{x}^{70}" + }, + { + "bbox": [ + 304, + 221, + 545, + 377 + ], + "type": "text", + "content": "). The generated video examples and additional experimental results are in our Supplementary Materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 406, + 378, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 378, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 378, + 418 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 433, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 504 + ], + "type": "text", + "content": "In this paper, we propose a zero-shot text-conditioned image-to-video framework, TI2V-Zero, to generate videos by modulating the sampling process of a pretrained video diffusion model without any optimization or fine-tuning. Comprehensive experiments show that TI2V-Zero can achieve promising performance on multiple datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "While showing impressive potential, our proposed TI2V-Zero still has some limitations. First, as TI2V-Zero relies on a pretrained T2V diffusion model, the generation quality of TI2V-Zero is constrained by the capabilities and limitations of the pretrained T2V model. We plan to extend our method to more powerful video diffusion foundation models in the future. Second, our method sometimes generates videos that are blurry or contain flickering artifacts. One possible solution is to apply post-processing methods such as blind video deflickering [30] or image/video deblurring [51] to enhance the quality of final output videos or the newly synthesized frame in each generation. Finally, compared with GAN and standard video diffusion models, our approach is considerably slower because it requires running the entire diffusion process for each frame generation. We will investigate some faster sampling methods [29, 32] to reduce generation time." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9022" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "text", + "content": "[1] Niki Aifanti, Christos Papachristou, and Anastasios Delopoulos. The mug facial expression database. In 11th International Workshop on Image Analysis for Multimedia Interactive Services WIAMIS 10, pages 1-4. IEEE, 2010. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 190 + ], + "type": "text", + "content": "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 192, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 288, + 235 + ], + "type": "text", + "content": "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Bjorn Ommer. Understanding object dynamics for interactive image-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5171-5181, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 356 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 288, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 402 + ], + "type": "text", + "content": "[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "type": "text", + "content": "[7] Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A Efros. Everybody dance now. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5933-5942, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 502 + ], + "type": "text", + "content": "[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videographer1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 288, + 567 + ], + "type": "text", + "content": "[9] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3d u-net: learning dense volumetric segmentation from sparse annotation. In International conference on medical image computing and computer-assisted intervention, pages 424-432. Springer, 2016. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 624 + ], + "type": "text", + "content": "[10] Michael Dorkenwald, Timo Milbich, Andreas Blattmann, Robin Rombach, Konstantinos G Derpanis, and Bjorn Ommer. Stochastic image-to-video synthesis using cinns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3742-3753, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "type": "text", + "content": "[12] Tsu-Jui Fu, Licheng Yu, Ning Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. Tell me what happened: Unifying text-guided video completion via multimodal masked video generation. In Proceedings of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10681-10692, 2023. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "type": "text", + "content": "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "text", + "content": "[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 197, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 240 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "type": "text", + "content": "[16] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "text", + "content": "[17] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "type": "text", + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 387, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 387, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 545, + 420 + ], + "type": "text", + "content": "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 421, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 421, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 421, + 545, + 454 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "text", + "content": "[21] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 510, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 510, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 308, + 510, + 545, + 553 + ], + "type": "text", + "content": "[22] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 555, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 555, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 555, + 545, + 599 + ], + "type": "text", + "content": "[23] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "type": "text", + "content": "[24] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion models for video prediction and infilling. arXiv preprint arXiv:2206.07696, 2022. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "text", + "content": "[25] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make it move: controllable image-to-video generation with text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18219-18228, 2022. 1, 2, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 690, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 690, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 690, + 545, + 713 + ], + "type": "text", + "content": "[26] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9023" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "Tim Green, Trevor Back, Paul Natev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 162 + ], + "type": "text", + "content": "[27] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 183 + ], + "type": "text", + "content": "[28] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 216 + ], + "type": "text", + "content": "[29] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. arXiv preprint arXiv:2106.00132, 2021.8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "type": "text", + "content": "[30] Chenyang Lei, Xuanchi Ren, Zhaoxiang Zhang, and Qifeng Chen. Blind video deflickering by neural filtering with a flawed atlas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10439-10448, 2023. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "text", + "content": "[31] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 350 + ], + "type": "text", + "content": "[32] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927, 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 350, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 404 + ], + "type": "text", + "content": "[33] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 404, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 449 + ], + "type": "text", + "content": "[34] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3667-3676, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 492 + ], + "type": "text", + "content": "[35] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "text", + "content": "[36] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 537, + 287, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 603 + ], + "type": "text", + "content": "[37] Nithin Gopalakrishnan Nair, Anoop Cherian, Suhas Lohit, Ye Wang, Toshiaki Koike-Akino, Vishal M Patel, and Tim K Marks. Steered diffusion: A generalized framework for plug-and-play conditional image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20850-20860, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "text", + "content": "[38] Haomiao Ni, Yihao Liu, Sharon X Huang, and Yuan Xue. Cross-identity video motion retargeting with joint transformation and synthesis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 412–422, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[39] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18444-18455, 2023. 1, 2, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 117 + ], + "type": "text", + "content": "[40] Haomiao Ni, Jiachen Liu, Yuan Xue, and Sharon X Huang. 3d-aware talking-head video motion transfer. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4954-4964, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 546, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 173 + ], + "type": "text", + "content": "[41] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 174, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 546, + 217 + ], + "type": "text", + "content": "[42] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 218, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 218, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 307, + 218, + 545, + 242 + ], + "type": "text", + "content": "[43] OpenAI. Openai: Introducing chatgpt. URL https://openai.com/blog/chatgpt, 2022.6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 242, + 546, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 546, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 546, + 308 + ], + "type": "text", + "content": "[44] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 354 + ], + "type": "text", + "content": "[45] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 354, + 546, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 546, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 546, + 420 + ], + "type": "text", + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 421, + 546, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 546, + 477 + ], + "type": "text", + "content": "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 4, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 477, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 477, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 477, + 545, + 533 + ], + "type": "text", + "content": "[48] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "type": "text", + "content": "[49] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 578, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 645 + ], + "type": "text", + "content": "[50] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 646, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 679 + ], + "type": "text", + "content": "[51] Siddhant Sahu, Manoj Kumar Lenka, and Pankaj Kumar Sa. Blind deblurring using deep learning: A survey. arXiv preprint arXiv:1907.10128, 2019. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[52] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m:" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9024" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "text", + "content": "Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "type": "text", + "content": "[53] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 150, + 287, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 150, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 49, + 150, + 287, + 204 + ], + "type": "text", + "content": "[54] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 206, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 206, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 49, + 206, + 287, + 237 + ], + "type": "text", + "content": "[55] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 239, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 239, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 49, + 239, + 287, + 271 + ], + "type": "text", + "content": "[56] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 272, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 272, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 49, + 272, + 287, + 316 + ], + "type": "text", + "content": "[57] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 317, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 317, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 49, + 317, + 287, + 360 + ], + "type": "text", + "content": "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 361, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 361, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 49, + 361, + 287, + 424 + ], + "type": "text", + "content": "[59] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 426, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 426, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 49, + 426, + 287, + 468 + ], + "type": "text", + "content": "[60] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 471, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 471, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 49, + 471, + 287, + 503 + ], + "type": "text", + "content": "[61] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. arXiv preprint arXiv:1808.06601, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 505, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 505, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 49, + 505, + 287, + 546 + ], + "type": "text", + "content": "[62] Wen Wang, kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 548, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 548, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 49, + 548, + 287, + 601 + ], + "type": "text", + "content": "[63] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 603, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 603, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 49, + 603, + 287, + 645 + ], + "type": "text", + "content": "[64] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 690 + ], + "type": "text", + "content": "[65] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "text", + "content": "[66] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 307 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "type": "text", + "content": "Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 162 + ], + "type": "text", + "content": "[67] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 1, 2, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "text", + "content": "[68] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "type": "text", + "content": "[69] Ceyuan Yang, Zhe Wang, Xinge Zhu, Chen Huang, Jianping Shi, and Dahua Lin. Pose guided human video generation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 201-216, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 263, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 545, + 307 + ], + "type": "text", + "content": "[70] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023.1, 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9025" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/8fe5d2c1-4f78-435e-8bde-f7a4afed8604_content_list.json b/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/8fe5d2c1-4f78-435e-8bde-f7a4afed8604_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1e94635737662735fa1e7b00804f3aab46d5eb4f --- /dev/null +++ b/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/8fe5d2c1-4f78-435e-8bde-f7a4afed8604_content_list.json @@ -0,0 +1,1696 @@ +[ + { + "type": "text", + "text": "TIGER: Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process", + "text_level": 1, + "bbox": [ + 212, + 130, + 756, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhiyuan Ren, Minchul Kim, Feng Liu, Xiaoming Liu \nMichigan State University", + "bbox": [ + 272, + 203, + 696, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{renzhiy1, kimminc2, liufeng6, liuxm}@msu.edu", + "bbox": [ + 287, + 241, + 684, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, diffusion models have emerged as a new powerful generative method for 3D point cloud generation tasks. However, few works study the effect of the architecture of the diffusion model in the 3D point cloud, resorting to the typical UNet model developed for 2D images. Inspired by the wide adoption of Transformers, we study the complementary role of convolution (from UNet) and attention (from Transformers). We discover that their respective importance change according to the timestep in the diffusion process. At early stage, attention has an outsized influence because Transformers are found to generate the overall shape more quickly, and at later stages when adding fine detail, convolution starts having a larger impact on the generated point cloud's local surface quality. In light of this observation, we propose a time-varying two-stream denoising model combined with convolution layers and transformer blocks. We generate an estimizable mask from each timestep to reweigh global and local features, obtaining time-varying fused features. Experimentally, we demonstrate that our proposed method quantitatively outperforms other state-of-the-art methods regarding visual quality and diversity. Code is available https://github.com/Zhiyuan-R/Tiger-Diffusion.", + "bbox": [ + 76, + 325, + 473, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 707, + 209, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The point cloud is an essential 3D shape representation that is easy to obtain (sensed directly from laser scanning), versatile (serving as a building block of mesh), and easy to manipulate (geometrically transformed in a straightforward manner by affine matrices). 3D point cloud generative models benefit extensive applications across robotics [35], medicine [15] and content creation [36], and lay the groundwork for other vision tasks like point cloud upsampling [52], point cloud completion [21, 31] and mesh generation [9]. However, compared to 2D images, the cost and complexity of acquiring 3D point clouds make it crucial to ex", + "bbox": [ + 76, + 734, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/20777058ec642375a5e5a540ccee9faeee2ff02eec539403aa98250cb074a8ea.jpg", + "image_caption": [ + "Figure 1. An illustration of different roles played by convolution and attention operations in the denoising model. Convolution is good for learning local relationships, and attention is optimal for modeling global relationships. We propose to merge these two properties across different timesteps in the diffusion process." + ], + "image_footnote": [], + "bbox": [ + 506, + 290, + 888, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plore and develop efficient and effective model architectures for 3D point cloud generation.", + "bbox": [ + 498, + 546, + 890, + 577 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing point cloud generative models are built on a range of frameworks, including generative adversarial networks (GANs) [1, 5], variational autoencoders (VAEs) [24], normalizing flows [23, 27, 51]. However, these methods neither are stable in training nor produce high-fidelity results. Recently, the denoising diffusion model [18], a novel generative method, has demonstrated exceptional results in 3D point cloud generation. Typically, diffusion models perturb a point cloud using a forward process and denoising algorithm. Once trained, the model can be used to generate new point clouds by iterating the reverse process over a sequence of time steps, with each step adding more noise to the input point cloud.", + "bbox": [ + 496, + 580, + 892, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several recent works, including DPM [34] and PVD [55], have successfully applied diffusion models to 3D point cloud generation, achieving high levels of naturalness and diversity. Very recently, LION [53] further improves by encoding point clouds into latent space. However, these methods commonly utilize UNet-like convolutional networks that are originally designed for image processing. There has been limited research on developing suitable de", + "bbox": [ + 496, + 779, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9462", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "noising model architectures that are specifically tailored for 3D point cloud generation. For example, PVD and LION both utilize Point-Voxel CNN (PVCNN) [30] as their denoising model. However, we observed that these PVCNN-based denoising models require a considerable number of timesteps to establish a rough shape since the limited receptive field cannot capture the global distribution of noise.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to convolution operations, we observe that the attention mechanism used in Transformers [10, 48] is more effective at capturing long-range dependencies and thus the overall shape of 3D points. This difference is illustrated in the first two rows of Fig. 1. Compared to the convolution-based diffusion model (first row), the attention-based diffusion model (second row) correctly conforms to the overall object shape at earlier denoising time steps (middle column). However, the convolution operation excels in modeling local relationships. Therefore, the details of the final output are better modeled with convolutions (last column) which is also discussed in the ablation analysis of Sec. 4.3. The finding suggests that the two operations, the convolution and the attention are complementary in their roles at different denoising timesteps (the last row of Fig. 1) and we ask the following question.", + "bbox": [ + 75, + 198, + 470, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "How to leverage the strengths of both CNNs and Transformers and enable them to dynamically capture complementary information in denoising diffusion for 3D point cloud generation?", + "bbox": [ + 102, + 457, + 444, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To answer this question, we propose a Time-varying denoising model for 3D point cloud generation (TIGER), a two-stream architecture combining a shallow CNN branch and Transformer branch with a timestep-dependent mask that adaptively reassigns weights to the global Transformer feature and the local CNN feature. To enable this feature-level fusion, we design an encoder-decoder to downsample the point cloud for reducing the number of points to a manageable size for Transformer. And we learn a selection mask that finds the optimal configuration at each timestep.", + "bbox": [ + 75, + 537, + 468, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, in this two-stream architecture where global and local information is modeled separately, we discover that supplying 3D points position information to the global branch is critical to the performance. To supply the 3D position information more effectively, we design two novel position encoding methods: phase-shifted position encoding and Base- $\\lambda$ position encoding. They create 3D position information in a continuous space and retain the property of linear expression of relative position. The effect of the time-dependent dual branch mask and the proposed position embedding is ablated in Sec. 4.3.", + "bbox": [ + 75, + 688, + 468, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A unique property of TIGER is that the weights of the two branches are dependent on the diffusion timesteps, and they are learned and optimized for the 3D point cloud gen", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "eration. Interestingly, experiments demonstrate that the weight of the convolution branch is monotonically increasing with the timestep, which is in accordance with our initial observation that convolutions are good for modeling fine details. Empirically we show that TIGER achieves state-of-the-art (SoTA) performance in 3D point cloud generation.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions include:", + "bbox": [ + 517, + 183, + 732, + 196 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel two-stream denoising model, which uses timestep to optimally reweigh the global feature from Transformer and the local feature from shallow CNN.", + "- We design two novel 3D continuous position encoding methods and position-aware self-attention to enable our Transformer branch to aggregate global features effectively from the latent point cloud input.", + "- Experiments show that our method outperforms prior works in 3D point cloud generation tasks qualitatively and quantitatively on the ShapeNetv2 dataset." + ], + "bbox": [ + 500, + 198, + 890, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 381, + 640, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Point Cloud Generation. The advances in 3D point cloud generation can be categorized by how the input is represented. Early methods [1, 12, 14, 38, 39, 44] represent the point cloud as a $N \\times 3$ matrix, where $N$ is the predefined number of 3D points. The backbone comprises 1D convolutions and max-pooling to allow the points' permutation invariance. So, the shape is represented by a global feature vector.", + "bbox": [ + 496, + 407, + 890, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recent works [5, 7, 13, 27, 34, 51] have improved upon the input representation by introducing a probabilistic distribution assumption, where each point cloud is considered as one data point sampled from the distribution. For instance, PointFlow [51] models the distribution of points by continuous normalizing flows, allowing the sampling of an arbitrary number of points to represent a point cloud. However, the backbone architecture is still based on 1D convolution-based encoder with the max-pooling operation [38].", + "bbox": [ + 496, + 527, + 890, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, PVD [55] and LION [53] adopt a hybrid point-voxel representation [30] which voxelize the point clouds for 3D convolution. This representation with rich spatial information combines well with the diffusion objective and produces promising results [53]. However, the backbone is still confined to convolution-based networks, lacking the capacity to effectively model the long range-dependencies. We, on the other hand, find the complementary roles of attention and convolution in the point cloud generator and propose a network that combines the benefits of both.", + "bbox": [ + 496, + 680, + 890, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models. Diffusion [18] and score-based methods [22, 46] have enjoyed remarkable success in images synthesis. Diffusion has been extended to many generation", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9463", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0030f138726516636a8cd7a1379cda31749f6cdb1ba7a50177aa783fc8154afb.jpg", + "image_caption": [ + "Figure 2. Illustration of our time-varying two-stream architecture (TIGER). The network's input is a noisy point cloud $\\mathbf{X}_t$ at timestep $t$ , and the goal is to predict the noise in $\\mathbf{X}_t$ to obtain the denoised point cloud $\\mathbf{X}_{t-1}$ . We first encode $\\mathbf{X}_t$ into latent point cloud $\\hat{\\mathbf{X}}_t$ and export coordinates information for position encoding and decoder. We feed the latent point cloud into both the Latent Point Cloud Transformer and a shallow CNN branch to get global feature $\\mathbf{X}_{tr}$ and local feature $\\mathbf{X}_c$ . Then, we combine two features using timestep dependent mask $\\mathbf{M}_c$ to create a fused feature $\\hat{\\mathbf{X}}_F$ . Last, we apply our latent point cloud decoder to predict the noise from $\\hat{\\mathbf{X}}_F$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 104, + 880, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tasks [2, 8, 17, 20, 25, 28, 41-43, 47], and largely improved by innovative techniques, such as fast sampling [33, 45], classifier-free guidance [19, 37] and latent diffusion [4, 42]. Recently 3D point cloud generation with diffusion [34, 53, 55] is also proposed with remarkable diversity and fidelity.", + "bbox": [ + 75, + 375, + 468, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, previous works do not investigate the behavior of the diffusion backbone with respect to the diffusion timestep. We find that the network equipped with both convolution and attention operations learns to utilize the local and global shape representation with different magnitudes based on the timestep. This time adaptive architecture allows the 3D point cloud generator to be trained faster than LION [53] while producing higher-quality outputs.", + "bbox": [ + 75, + 450, + 472, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer for Point Clouds. Transformer architecture [48] is intrinsically a set operator, which is suitable for permutation-invariant point clouds. PCT [16] applies global attention to the point cloud. Point Transformer V1 [54] and V2 [49] convert global attention to local attention to reduce memory consumption and computational complexity.", + "bbox": [ + 75, + 589, + 468, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrary to the previous methods, our method decouples the global and local information by using Transformer to aggregate global information effectively while using CNN to model the local information. Also, we find that position embedding is crucial to performance and propose a new position embedding to further improve the performance.", + "bbox": [ + 75, + 680, + 468, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 784, + 169, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first briefly formulate the probabilistic model of forward and reverse diffusion processes for 3D point cloud generation. Then, we dive into the details of our time-varying two-stream architecture, including the encoder part, latent point Transformer, time mask generator, and decoder part. The entire architecture is shown in Fig. 2.", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 498, + 375, + 702, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a point cloud $\\mathbf{X}_0\\in \\mathbb{R}^{N\\times 3}$ , where $N$ is the number of points, the diffusion forward process gradually diffuses the original point cloud into Gaussian noise $\\mathbf{X}_T$ ; the forward process can be formulated with a transition probability $q(\\mathbf{X}_t|\\mathbf{X}_{t - 1})$ . By giving a sequence of pre-defined noising scales $\\beta_{1},\\beta_{2},\\ldots ,\\beta_{T}$ , the transition probability can be expressed as:", + "bbox": [ + 496, + 396, + 890, + 503 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {X} _ {t} \\mid \\mathbf {X} _ {t - 1}\\right) = \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} \\mathbf {X} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 508, + 890, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the reverse process, we use learn the $p_{\\theta}(\\mathbf{X}_{t - 1}|\\mathbf{X}_t)$ a Gaussian distribution which approximates the intractable real distribution $q(\\mathbf{X}_{t - 1}|\\mathbf{X}_t)$ . Specifically,", + "bbox": [ + 496, + 534, + 890, + 580 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} = \\mathcal {N} \\left(\\mu_ {\\theta} \\left(\\mathbf {X} _ {t}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 585, + 890, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where we learn the mean $\\mu_{\\theta}(\\mathbf{X}_t,t)$ and set the variance $\\sigma_t^2$ to a fixed schedule as in [18]. Conventionally, $\\mu_{\\theta}(\\mathbf{X}_t,t)$ is modeled using a UNet backbone and it depends on the timestep $t$ to scale and shift the feature. We design a two-stream backbone where the timestep modulates the weight between the two streams. To sample a 3D point cloud, we iteratively denoise $\\mathbf{X}_T,T$ times to finally generate the highfidelity samples. In training, we optimize the MSE loss:", + "bbox": [ + 496, + 608, + 890, + 731 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {s i m p l e}} = \\mathbb {E} _ {t \\sim [ 1, T ]} \\| \\mu - \\mu_ {\\theta} (\\mathbf {X} _ {t}, t) \\| _ {2} ^ {2} \\tag {3} \\\\ = \\mathbb {E} _ {t \\sim [ 1, T ]} \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {X} _ {t}, t) \\| _ {2} ^ {2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 734, + 890, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon$ is the ground truth noise and $\\epsilon_{\\theta}$ is the noise prediction by our diffusion model.", + "bbox": [ + 496, + 779, + 890, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Noisy Point Cloud Encoder", + "text_level": 1, + "bbox": [ + 498, + 816, + 746, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To effectively represent the $N\\times 3$ points, we adopt the voxelization encoding scheme of PVCNN [30]. This representation works well with both CNN and Transformer architecture as we can downsample the voxel-grid following [11] to", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9464", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5d0d12675c8a255d5c173a700feee286dce8d645432f1cc4d6a57748aa521066.jpg", + "image_caption": [ + "Figure 3. The overview of our encoder. It extracts features in voxel space and downsamples the point cloud by applying trilinear interpolation of voxel with grouped points." + ], + "image_footnote": [], + "bbox": [ + 86, + 95, + 467, + 218 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "increase computation efficiency. Specifically, a noisy point cloud encoder $\\mathcal{E}:\\mathbb{R}^{N\\times 3}\\to \\mathbb{R}^{M\\times d}$ transforms the noisy point cloud $\\mathbf{X}_t\\in \\mathbb{R}^{N\\times 3}$ at timestep $t$ into a latent point cloud representation $\\hat{\\mathbf{X}}_t\\in \\mathbb{R}^{M\\times d}$ , where $N$ is the number of original points, $M$ is the number of downsampled points and $d$ is the dimension of each latent point's feature. The overview of our encoder is illustrated in Fig. 3.", + "bbox": [ + 75, + 299, + 468, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To aggregate features in a rigid field, we first voxelize the noisy point cloud $\\{x_{n},y_{n},z_{n}\\} = \\mathbf{X}_{t}\\in \\mathbb{R}^{N\\times 3}$ into the voxel grids $\\{\\mathbf{V}_{u,v,w}\\} = \\mathbf{V}\\in \\mathbb{R}^{L\\times L\\times L}$ , where $L$ is the resolution and $n$ is the point index in $N$ . Specifically,", + "bbox": [ + 75, + 404, + 468, + 465 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} _ {u, v, w} = \\sum_ {n = 1} ^ {N} \\mathbb {I} \\left(x _ {n} \\in \\mathbb {N} (u, \\delta), y _ {n} \\in \\mathbb {N} (v, \\delta), z _ {n} \\in \\mathbb {N} (w, \\delta)\\right) * \\mathbf {f} _ {n}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 474, + 468, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{N}(z,\\delta) = \\{x|x\\in [z - \\delta ,z + \\delta)\\}$ , $\\mathbf{f}_n$ denotes the $n^{th}$ extracted latent feature corresponding to the $n^{th}$ point $(\\mathbf{X}_t)_n$ and $\\mathbb{I}(\\cdot)$ is an indicator of whether the $n^{th}$ point falls into the voxel grid $(u,v,w)$ . After voxelization, we apply a series of 3D convolution, Swish [40] activation and Group-Norm [50]. After applying a series of these operations, this results in a latent volume $\\hat{\\mathbf{V}}\\in \\mathbb{R}^{L\\times L\\times L\\times d}$ , where $d$ is the channel size for each voxel.", + "bbox": [ + 75, + 505, + 468, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use furthest point sampling algorithm [11] to downsample the input noisy point cloud $\\mathbf{X}_t\\in \\mathbb{R}^{N\\times 3}$ into a sparser point cloud $\\mathbf{X}_t^s\\in \\mathbb{R}^{M\\times 3}$ ( $M < N$ ). Then, as the position in volume is discrete, we query the latent volume $\\hat{\\mathbf{V}}$ by the sparse point cloud $\\mathbf{X}_t^s$ via trilinear interpolation, eventually transformed into latent point cloud $\\hat{\\mathbf{X}}_t\\in \\mathbb{R}^{M\\times d}$ . It is worth noting we preserve the coordinates of $\\mathbf{X}_t$ and $\\mathbf{X}_t^s$ to do upsampling with the decoder and provide position embedding for our latent point cloud Transformer.", + "bbox": [ + 75, + 627, + 468, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Latent Point Cloud Transformer", + "text_level": 1, + "bbox": [ + 76, + 771, + 362, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous denoising models such as PVCNN [30] simply apply stacked convolutional layers to create feature representations. Contrarily, we aim to better model the global relationship by introducing a novel Transformer branch named latent point cloud Transformer.", + "bbox": [ + 75, + 794, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tokenization. After encoding, we get a latent point cloud representation $\\hat{\\mathbf{X}}_t\\in \\mathbb{R}^{M\\times d}$ and its coordinates $\\mathbf{X}_t^s\\in \\mathbb{R}^{M\\times 3}$ .", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2348bc51c4bc105e548c4f3b0f78a9e8932dc7feaf9fea47027652e82f247cec.jpg", + "image_caption": [ + "(a) PSPE" + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 694, + 199 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c1db3196b9deaa2816c4d79e4ad6b7cbee8d8f1fbf38de23e567196690137c13.jpg", + "image_caption": [ + "(b) BλPE", + "Figure 4. Illustration of two examples of the position embedding from PSPE and BλPE respectively. Both methods show distinguished representation for each position." + ], + "image_footnote": [], + "bbox": [ + 697, + 90, + 887, + 199 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following [29], we use dual PatchNorm to project the latent point cloud into tokens, which place LayerNorm before and after an MLP layer for more stable training:", + "bbox": [ + 498, + 292, + 890, + 337 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} = \\operatorname {L N} \\left(\\operatorname {M L P} \\left(\\operatorname {L N} \\left(\\hat {\\mathbf {X}} _ {\\mathbf {t}}\\right)\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 345, + 890, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{T} \\in \\mathbb{R}^{M \\times D}$ and $D$ is the dimension in the token space (in our paper, $D = 2d$ ).", + "bbox": [ + 498, + 371, + 890, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Position Encoding. Since the order of the latent points is arbitrary in a point cloud, it's necessary to create a 3D position embedding and fuse it into the feature, so that the model knows the position of each point. Contrary to the position embedding in image domains [10, 26], the design of 3D position embedding plays a vital role in the performance and our position embedding produces the best result. We propose two novel 3D space continuous position encoding methods: Phase Shift Position Encoding (PSPE) and Base- $\\lambda$ Position Encoding (B $\\lambda$ PE). For PSPE, we use Sine and Cosine functions of different frequencies and phases:", + "bbox": [ + 498, + 402, + 890, + 568 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {e m b} \\left(p o s _ {j}, 6 i + 2 * (j - 1)\\right) = \\sin \\left(\\frac {p o s _ {j}}{1 0 0 0 0 ^ {\\frac {2 i}{D}}} + (j - 1) \\frac {2 \\pi}{3}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 575, + 890, + 618 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {e m b} (p o s _ {j}, 6 i + 1 + 2 * (j - 1)) = \\cos \\left(\\frac {p o s _ {j}}{1 0 0 0 0 ^ {\\frac {2 i}{D}}} + (j - 1) \\frac {2 \\pi}{3}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 618, + 895, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $pos_{j}$ is $j$ -th element in each point's position, specifically $pos_{1} = x$ , $pos_{2} = y$ , $pos_{3} = z$ , and $i$ is the index of the dimension. We maximize the phase shift difference $(\\frac{4\\pi}{3})$ of different axes to distinguish them and combine Sine and Cosine to guarantee the linear representation of relative position, which will be proved mathematically in Supp. An example is shown in Fig. 4a and $\\mathbf{P}_{emb}$ is shaped $\\mathbb{R}^{M\\times D}$ .", + "bbox": [ + 498, + 660, + 890, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, PSPE requires 6 channels to represent one frequency feature, which is quite inefficient. To improve channel utilization, we apply BλPE to compress 3D position into one scalar:", + "bbox": [ + 498, + 766, + 890, + 823 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np o s = \\lambda^ {2} * z + \\lambda * y + x \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 825, + 890, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {e m b} (p o s, 2 i) = \\sin \\left(\\frac {p o s}{1 0 0 0 0 ^ {\\frac {2 i}{D}}}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 844, + 890, + 873 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {e m b} (p o s, 2 i + 1) = \\cos \\left(\\frac {p o s}{1 0 0 0 0 ^ {\\frac {2 i}{D}}}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 876, + 890, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9465", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/49b1d9a93ac5046301acf56d70da9c791842c0f7556fa73c8c610a5cf832f53b.jpg", + "image_caption": [ + "Figure 5. The overview of the position-aware self-attention. We compute position relationship map $\\mathbf{H}$ by previous position embedding to reweigh the $\\mathbf{Q}\\mathbf{K}^T$ matrix." + ], + "image_footnote": [], + "bbox": [ + 81, + 92, + 459, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $pos$ is a polynomial expression of $\\lambda$ . Specifically, we set $\\lambda$ to 1000, which means this preserves three decimal places of precision and one example is shown in Fig. 4b. The choice between these two position encoding methods is essentially a trade-off between granularity and expressiveness. We ablate them and other non-position encoding and learnable position encoding in Sec. 4.3.", + "bbox": [ + 75, + 393, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transformer and Position-Aware Attention We stack $L$ layers of Transformer block [48] which comprises of LayerNorm(LN) [3] and Multihead Self Attention (MHA) and MultiLayer Perceptrons. We let the token $\\mathbf{T}_0 = \\mathbf{T}$ and let", + "bbox": [ + 75, + 500, + 468, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {l *} = \\operatorname {M H A} \\left(\\ln \\left(\\mathbf {T} _ {l - 1}\\right)\\right) + \\mathbf {T} _ {l - 1}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 575, + 468, + 592 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {l} = \\operatorname {M L P} \\left(\\operatorname {L N} \\left(\\mathbf {T} _ {l *}\\right)\\right) + \\mathbf {T} _ {l *}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 594, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {t r} := \\mathbf {T} _ {L}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 613, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To add position embedding into the model, one can simply add it to $\\mathbf{T}$ . But as the Transformer blocks are gradually stacked, the continuous 3D position encoding is difficult to retain in the token representation. Inspired by [54], we explicitly add the position encoding to each MHA in order to make our attention map position-aware, as shown in Fig. 5. Specifically, our position-aware MHA modified as", + "bbox": [ + 75, + 642, + 468, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} = \\operatorname {S o f t m a x} \\left(\\left(\\mathbf {P} _ {e m b} \\mathbf {W} _ {p}\\right) \\left(\\mathbf {P} _ {e m b} \\mathbf {W} _ {p}\\right) ^ {T}\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 762, + 468, + 781 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {l *} = \\operatorname {S o f t m a x} \\left(\\frac {\\left(\\mathbf {T} _ {l} \\mathbf {W} _ {q}\\right) \\left(\\mathbf {T} _ {l} \\mathbf {W} _ {k}\\right) ^ {T} \\odot \\mathbf {H}}{\\sqrt {D}}\\right) \\left(\\mathbf {T} _ {l} \\mathbf {W} _ {v}\\right) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 795, + 468, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_q,\\mathbf{W}_k,\\mathbf{W}_v,\\mathbf{W}_p\\in \\mathbb{R}^{D\\times Z}$ are projection matrices and $\\mathbf{H}\\in \\mathbb{R}^{M\\times M}$ is the position relationship map. Modified MHA could incorporate position information explicitly and sufficiently during training. $\\mathbf{X}_{tr}$ is the final output.", + "bbox": [ + 76, + 838, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ad7b3332409ec3d6e4b12c4f9b36848faeea9c190734e000f0bcb3e3d116d1ca.jpg", + "image_caption": [ + "Figure 6. The overview of our decoder. By querying the latent volume $\\hat{\\mathbf{X}}_F$ with previous coordinates information $\\mathbf{X}_t$ , we can upsample the latent point cloud into 3D space." + ], + "image_footnote": [], + "bbox": [ + 506, + 95, + 890, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Time Mask Generator", + "text_level": 1, + "bbox": [ + 500, + 265, + 707, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We apply a depth-2 PVCNN [30] to get local feature $\\mathbf{X}_c\\in$ $\\mathbb{R}^{M\\times D}$ and our latent point cloud Transformer to get global feature $\\mathbf{X}_{tr}\\in \\mathbb{R}^{M\\times D}$ . Time mask generator module scales the local and global features by the time variable. We force the sum of two learnable time masks to be 1. With this constraint, the trade-off of local and global features can be learned by the network.", + "bbox": [ + 498, + 287, + 890, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first align the dimension and scale of $\\mathbf{X}_c$ and $\\mathbf{X}_{tr}$ by one layer MLP and Sigmoid function $\\sigma$ .", + "bbox": [ + 498, + 393, + 890, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {c} ^ {*} = \\sigma (\\operatorname {M L P} (\\mathbf {X} _ {c})), \\quad \\mathbf {X} _ {t r} ^ {*} = \\sigma (\\operatorname {M L P} (\\mathbf {X} _ {t r})). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 433, + 890, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we encode a timestep $t$ to a sinusoidal position embedding [18, 48] to $\\mathbf{t}_{emb} \\in \\mathbb{R}^c$ . We use two stacked MLP to map $\\mathbf{t}_{emb}$ to a higher dimension space and apply Sigmoid to keep one of the masks between 0 to 1. Specifically,", + "bbox": [ + 498, + 460, + 890, + 521 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {c} = \\sigma (\\mathrm {M L P} (\\text {L e a k y R e L U} (\\mathrm {M L P} (\\mathbf {t} _ {\\text {e m b}}))), \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 531, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {t r} = \\mathbf {1} - \\mathbf {M} _ {c}. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 550, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{M}_c$ , $\\mathbf{M}_{tr} \\in \\mathbb{R}^D$ are the time masks for CNN and Transformer feature respectively. We ablate the dimension of masks in Sec. 4.3 to be $\\mathbb{R}^D$ or $\\mathbb{R}^1$ . We multiply these two masks by the respective features and adjust the fused feature by MLP.", + "bbox": [ + 498, + 577, + 890, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {X}} _ {F} = \\operatorname {M L P} \\left(\\mathbf {X} _ {c} ^ {*} \\odot \\mathbf {M} _ {c} + \\mathbf {X} _ {t r} ^ {*} \\odot \\mathbf {M} _ {c}\\right), \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 662, + 890, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{X}}_F\\in \\mathbb{R}^{M\\times D}$ is the final feature adaptively aggregating local and global features based on the timestep.", + "bbox": [ + 498, + 690, + 890, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Latent Point Cloud Decoder", + "text_level": 1, + "bbox": [ + 500, + 729, + 751, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to upsample the latent point cloud and predict the corresponding noise, we design a latent point cloud decoder $\\mathcal{D}$ , which converts the latent point cloud $\\hat{\\mathbf{X}}_F \\in \\mathbb{R}^{M \\times D}$ to $\\mathbf{X}_t \\in \\mathbb{R}^{N \\times 3}$ . Similar to Sec. 3.2, we can voxelize $\\hat{\\mathbf{X}}_F$ into a latent volume $\\tilde{\\mathbf{V}}_F \\in \\mathbb{R}^{L \\times L \\times L \\times D}$ , and take $\\mathbf{X}_t$ as the query for trilinear interpolation, eventually resulting in the final noise prediction $\\epsilon_{\\theta} \\in \\mathbb{R}^{N \\times 3}$ . This formulation can be expressed as:", + "bbox": [ + 498, + 753, + 890, + 873 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\theta} = \\operatorname {T r i l i n e a r} \\left(\\operatorname {V o x e l i z e} \\left(\\hat {\\mathbf {X}} _ {F}\\right), \\mathbf {X} _ {t}\\right) \\mathbf {W} ^ {D \\times 3}, \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 883, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9466", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b0f2c3c6cdadb2cbb64979d8cbe9ff08666450cf659e482c94372239e3e5c254.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGenerative ModelAirplaneChairCar
CD→50%EMD→50%CD→50%EMD→50%CD→50%EMD→50%
r-GAN [1]GAN98.4096.7983.6999.7094.4699.01
l-GAN(CD) [1]GAN87.3093.9568.5883.8466.4988.78
l-GAN(EMD) [1]GAN89.4976.9171.9064.6571.1666.19
PointFlow [51]Normalizing Flow75.6870.7462.8460.5758.1056.52
DPF-Net [27]Normalizing Flow75.1865.5562.0058.5362.3554.48
ShapeGF [5]GAN80.0076.1768.9665.4863.2056.53
SoftFlow [23]Normalizing Flow76.0565.8059.2160.0564.7760.09
SetVAE [24]VAE76.5467.6558.8460.5759.9559.94
DPM [34]Diffusion76.4286.9160.0574.7768.8979.97
PVD [55]Diffusion73.8264.8156.2653.3254.5553.83
TIGERDiffusion71.8555.8254.6152.7154.3152.24
LION [53]Diffusion67.4161.2353.7052.3453.4151.14
TIGERDiffusion67.2156.2654.3251.7154.1250.24
", + "bbox": [ + 119, + 88, + 849, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison with baselines using 1-NN. Both CD and EMD are considered, where CD is multiplied by ${10}^{3}$ and EMD is multiplied by ${10}^{2}$ . The closer the score is to 50%,the better the quality and diversity of generated samples. [Key: Best, Second Best]", + "bbox": [ + 75, + 311, + 893, + 340 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/18846db4820861bec226d324c3415b846fd7c495f424a059089b22abe3ad0c49.jpg", + "image_caption": [ + "Figure 7. Our generation results (right) compared to baseline models (left). TIGER generates high-quality and diverse 3D point clouds." + ], + "image_footnote": [], + "bbox": [ + 101, + 353, + 867, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{W}^{D\\times 3}$ is the projection matrix to map the noise into 3D space and $\\mathbf{X}_t$ is the original point cloud coordinates. The overview of our decoder is shown in Fig. 6.", + "bbox": [ + 75, + 609, + 468, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 670, + 209, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 695, + 266, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We follow previous works and choose ShapeNetv2 [6] as our main dataset, which is pre-processed by PointFlow [51]. To fairly compare with other different baseline methods, we train and evaluate on three categories: airplane, chair and car. For each shape, we sample 2,048 points and normalize them globally across the entire dataset. The training set consists of 2,832,4,612, and 2,458 shapes, while the evaluation set is composed of 405,662, and 352 shapes for airplanes, chairs, and cars, respectively.", + "bbox": [ + 75, + 719, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Following the baselines PVD [55] and LION [53], we use 1-NN (1-nearest neighbor) accuracy [32] as our evaluation metric. This metric has been shown to effectively", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "measure both the quality and diversity of generated point clouds and a score closer to $50\\%$ indicates superior performance [51]. To calculate the 1-NN distance matrix, we consider two commonly used metrics for measuring the distance between point clouds: CD (Chamfer Distance) and EMD (Earth Movers' Distance).", + "bbox": [ + 496, + 609, + 890, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with SoTA methods", + "text_level": 1, + "bbox": [ + 498, + 710, + 785, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. To compare our time-varying denoising model with SoTA methods, we evaluate its performance against competitive models such as SetVAE [24], DPM [34], PVD [55] and LION [53]. It is noteworthy that in order to compare with LION, which uses a different dataset splitting strategy (sampling from the first 10,000 points instead of the latter 5,000 points), we evaluate our model under both LION's strategy and the standard strategy used by other methods. As shown in Tab. 1, we outperform LION in four out of six metrics. Compared to other methods, our performance is significantly better. Notably, our EMD-based", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9467", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/56fde96708d3b6d5720031c3100d036c0e299dbf21347878db742b80b7a58dc7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTraining Time (GPU hours)Inference Time (s)
PVD [55]1428.46
LION [53]55027.12
Tiger1649.73
", + "bbox": [ + 81, + 88, + 467, + 151 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2616913c8571b50801082ca6647f18dc08129298d005e968bfd16c62d3d77c28.jpg", + "table_caption": [ + "Table 2. Training and inference time comparison. Training time is counted on average for three categories: airplane, chair and car." + ], + "table_footnote": [], + "table_body": "
MethodW1W2Mean CDMean EMD
Convolution1061.3456.31
Transformer0162.4155.85
Conv+Tran0.50.564.4557.92
Ourslearnable (scalar)60.7954.12
Ourslearnable (channel-wise)59.4353.51
", + "bbox": [ + 81, + 191, + 467, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "1-NN accuracy shows a significant improvement of $13.45\\%$ relative to PVD and $8.12\\%$ relative to LION for airplanes. This is attributed to the efficient aggregation of global features by our Transformer branch, which is especially beneficial for EMD as it is primarily influenced by global distribution. Fig. 7 shows that our generated samples are of high quality and diversity compared with other methods.", + "bbox": [ + 75, + 339, + 467, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficiency. To compare the training and inference speed, we report the GPU hours with NVidia V100 GPUs as in LION [53]. As shown in Tab. 2, our model trains much faster than LION, with only a quarter of its training time and a third of its inference time. Our training and inference speed is comparable to that of PVD while surpassing LION and PVD in the generation quality as shown in Tab. 1. More computation analysis can be found in Supp.", + "bbox": [ + 75, + 445, + 467, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation and Analysis", + "text_level": 1, + "bbox": [ + 76, + 575, + 282, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Time-varying two-stream architecture. We ablate our time-varying two-stream architecture by comparing it with single ConvNet, single Transformer, and two-stream networks with equal weight. In this ablation, we also compare the performance of our time masking with scalar value setting and channel-wise value setting. The results, presented in the first two rows of Tab. 3, show that Transformer performs better on the EMD metric, while ConvNet performs better on the CD metric. Using a fixed equal weight for the two-stream network leads to a significant performance drop on both metrics. However, the learnable time mask improves the performance on both metrics, indicating the importance of the time-varying design. Further, channel-wise time masking improves performance slightly more than scale value setting due to its greater expressivity. Transformer backbones, position encoding, and self-attention strategies. Here we carried out three ablations: (1) We compare the performance of our Transformer backbone with two classic Transformer backbones, PCT [16] and Point Transformer (PointTran) [54]. (2) To assess", + "bbox": [ + 75, + 598, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/01e482d857c59e94089b584f2c8e01aa5b7a27c5e0992ad59941197e883f93b6.jpg", + "table_caption": [ + "Table 3. Ablation of different architecture designs. $W_{1}$ and $W_{2}$ are the weights for the ConvNet branch and Transformer branch. [Key: Best, Second Best]" + ], + "table_footnote": [], + "table_body": "
TBPEAPMean CDMean EMD
PCT [16]Learnable-65.8260.21
PointTran [54]LearnableLearnable66.4262.39
Ours--63.5157.98
OursLearnableLearnable62.1756.13
OursPSPE-60.1155.74
OursPSPELearnable61.4855.97
OursPSPEPSPE59.0254.49
OursBλPE-61.3554.27
OursBλPELearnable61.8956.02
OursBλPEBλPE59.4353.51
", + "bbox": [ + 506, + 88, + 885, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation of Transformer backbones, position encoding, and self-attention strategies. [Key: Best, Second Best; TB=Transformer backbone; PE=position encoding; AP=additional position information]", + "bbox": [ + 498, + 244, + 890, + 299 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7487b1c2e2fa1ec6dd7186593fd646c1f0b70302f4d35ce36c14cbcbedd6530.jpg", + "image_caption": [ + "Figure 8. Visualization of attention maps in 3D point clouds. Column (c) is the result of column (b) multiplied by position relationship map H. Taking the first row as an example, queried by the airplane head's position, our method can aggregate the position information from the airplane's right wing and left elevator compared with standard self-attention. This comparison highlights the effectiveness of our position-aware self-attention module in capturing long-range relationships in 3D point clouds. [SA: Self-attention, PASA: Position-aware Self-attention]" + ], + "image_footnote": [], + "bbox": [ + 506, + 313, + 718, + 459 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23fe7c856fb0923408d7a66d7d8252ac8419dd825c62a27118ab97168b3fa04a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 758, + 311, + 883, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our proposed position encoding methods, PSPE and BλPE, we compare them with no position encoding and learnable position encoding. (3) We investigate our position-aware self-attention module by comparing it with standard self-attention and self-attention with learnable position encoding as proposed in [54]. As shown in Tab. 4, our latent point cloud Transformer outperforms both PCT and Point Transformer under various settings. Furthermore, our proposed position encoding methods, PSPE and BλPE, significantly improve performance compared to no position encoding or learnable position encoding. Specifically, PSPE improves the CD metric, while BλPE improves the EMD metric. Our position-aware self-attention module, which aggregates corresponding position encoding, achieves better results than standard self-attention and self-attention with learnable position encoding. To illustrate this, Fig. 8 visualizes the attention map in 3D point space for both standard self-attention and our position-aware self-attention. This visualization demonstrates the ability of our method to capture", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9468", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8009a5ef648b8eb0755451f9d170aed35e6f8460236d5baebd74ddf9fe163948.jpg", + "image_caption": [ + "(a) Airplane" + ], + "image_footnote": [], + "bbox": [ + 86, + 85, + 348, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/82263e7e086d08818f9f6e6e8d6709972c4a468e9cb82c856ff8a9dc4b62c7e0.jpg", + "image_caption": [ + "(b) Chair" + ], + "image_footnote": [], + "bbox": [ + 349, + 87, + 612, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0b9120aa39eab04542649825872d9840f49d658450f75d08f536bfa3f5e388fb.jpg", + "image_caption": [ + "(c) Car" + ], + "image_footnote": [], + "bbox": [ + 614, + 87, + 880, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8ad4e30cfcb2378ba7eac682e165aa3026c9cf196214f26bb2d995d07b2eb7db.jpg", + "image_caption": [ + "Figure 9. The changes of ConvNet importance for each category over timesteps with the channel-wise time mask setting. The thick curve is the result of Gaussian smoothing.", + "Figure 10. High quality and diverse 3D point clouds generated from our TIGER model trained on 55 ShapeNetv2 categories." + ], + "image_footnote": [], + "bbox": [ + 81, + 263, + 890, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "long-range relationships in the data.", + "bbox": [ + 75, + 486, + 315, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Weight changing over timesteps. To validate our motivation of reweighing the local and global features from ConvNet and Transformer at different time steps, we visualize the changes in weights. For channel-wise time mask setting, we use the ratio of the weight of the local feature over the weight of the global feature to measure the importance of ConvNet and Transformer at each timestep. Fig. 9 indicates that the importance of ConvNet for each category increases nearly monotonically, implying that the diffusion process of 3D point cloud generation benefits from global features at the early stages of diffusion and requires the aggregation of more local features at later stages. Moreover, we observe that the curves of airplanes and chairs change more drastically than that of cars, which we hypothesize is due to the complicated shapes of airplanes and chairs, resulting in a more complex diffusion process.", + "bbox": [ + 75, + 501, + 468, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unconditional generation of multiple classes. To further demonstrate the generalizability of our model, we train a universal TIGER model on all 55 categories of ShapeNetv2 [6]. Fig. 10 shows that our universal model can produce shapes with high quality and diversity.", + "bbox": [ + 75, + 743, + 468, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 829, + 204, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel time-varying denoising model with a latent point transformer backbone. Experimentally, we show how the time mask optimally aggre", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gates local and global features at different timesteps, which quantitatively reaches SoTA performance. In accordance with our hypothesis, Transformer is more important at early diffusing stages and ConvNet has a larger weight at later timesteps in the 3D point cloud generation task. Future work will be focused on exploring how our time-varying denoising model would behave in other diffusion-based tasks.", + "bbox": [ + 496, + 486, + 890, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Although we generate high-quality and natural samples, we cannot control the category of the generated shape. One potential solution is to encode the category into latent space and train the diffusion model conditioned on the latent code. We also acknowledge that our model needs to get the local feature and global feature by two subnetworks, which is not quite efficient. In this work, we gain efficiency with the model converging with fewer epochs. But future works can increase the backbone efficiency by proposing time-varying properties with only one network.", + "bbox": [ + 496, + 609, + 892, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Potential Negative Impacts. While it has the potential to speed up progress in both areas, it also has the risk of eliminating the jobs of game and furniture designers due to automation. However, our model should have a mostly positive impact, because the most likely use of our method would be to assist designers not completely replace their jobs. It is still necessary to select suitable shapes among many generated samples.", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9469", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, 2018. 1, 2, 6", + "[2] Vishal Asnani, John Collomosse, Tu Bui, Xiaoming Liu, and Shruti Agarwal. ProMark: Proactive diffusion watermarking for causal attribution. In CVPR, 2024. 3", + "[3] Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450, 2016. 5", + "[4] Fan Bao, Kun Xu, Chongxuan Li, Lanqing Hong, Jun Zhu, and Bo Zhang. Variational (gradient) estimate of the score function in energy-based latent variable models. In ICML, 2021. 3", + "[5] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snavely, and Bharath Hariharan. Learning gradient fields for shape generation. In ECCV, 2020. 1, 2, 6", + "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 8", + "[7] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. NeurIPS, 31, 2018. 2", + "[8] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusiondet: Diffusion model for object detection. arXiv preprint arXiv:2211.09788, 2022. 3", + "[9] Tamal K Dey and James Hudson. Pmr: Point to mesh rendering, a feature-based approach. In VIS, 2002. 1", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2, 4", + "[11] Yuval Eldar, Michael Lindenbaum, Moshe Porat, and Yehoshua Y Zeevi. The farthest point strategy for progressive image sampling. In ICPR, 1994. 3, 4", + "[12] Matheus Gadelha, Rui Wang, and Subhransu Maji. Multiresolution tree networks for 3d point cloud processing. In ECCV, 2018. 2", + "[13] Will Grathwohl, Ricky TQ Chen, Jesse Bettencourt, Ilya Sutskever, and David Duvenaud. Ffjord: Free-form continuous dynamics for scalable reversible generative models. arXiv preprint arXiv:1810.01367, 2018. 2", + "[14] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3D surface generation. In CVPR, 2018. 2", + "[15] Jia Guo, Xuanxia Yao, Mengyu Shen, Jiafei Wang, and Wanyou Liao. A deep learning network for point cloud of medicine structure. In ITME, 2018. 1", + "[16] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 3, 7", + "[17] Kilian Konstantin Haefeli, Karolis Martinkus, Nathanael Perraudin, and Roger Wattenhofer. Diffusion models for" + ], + "bbox": [ + 78, + 114, + 468, + 896 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "graphs benefit from discrete state spaces. arXiv preprint arXiv:2210.01549, 2022.3", + "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2, 3, 5", + "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3", + "[20] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 3", + "[21] Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. In CVPR, 2020. 1", + "[22] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 2005. 2", + "[23] Hyeongju Kim, Hyeonseung Lee, Woo Hyun Kang, Joun Yeop Lee, and Nam Soo Kim. Softflow: Probabilistic framework for normalizing flow on manifolds. In NeurIPS, 2020. 1, 6", + "[24] Jinwoo Kim, Jaehoon Yoo, Juho Lee, and Seunghoon Hong. Setvae: Learning hierarchical composition for generative modeling of set-structured data. In CVPR, 2021. 1, 6", + "[25] Minchul Kim, Feng Liu, Anil Jain, and Xiaoming Liu. DCFace: Synthetic face generation with dual condition diffusion model. In CVPR, 2023. 3", + "[26] Minchul Kim, Yiyang Su, Feng Liu, Anil Jain, and Xiaoming Liu. Keypoint relative position encoding for face recognition. In CVPR, 2024. 4", + "[27] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. In ECCV, 2020. 1, 2, 6", + "[28] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 3", + "[29] Manoj Kumar, Mostafa Dehghani, and Neil Houlsby. Dual patchnorm. arXiv preprint arXiv:2302.01327, 2023. 4", + "[30] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. Pointvoxel CNN for efficient 3d deep learning. In NeurIPS, 2019. 2, 3, 4, 5", + "[31] Yunfei Long, Daniel Morris, Xiaoming Liu, Marcos Paul Gerardo Castro, Punarjay Chakravarty, and Praveen Narayanan. Radar-camera pixel depth association for depth completion. In CVPR, 2021. 1", + "[32] David Lopez-Paz and Maxime Oquab. Revisiting classifier two-sample tests. arXiv preprint arXiv:1610.06545, 2016. 6", + "[33] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927, 2022. 3", + "[34] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, 2021. 1, 2, 3, 6", + "[35] Peiyuan Ni, Wenguang Zhang, Xiaoxiao Zhu, and Qixin Cao. Pointnet++ grasping: Learning an end-to-end spatial grasp generation algorithm from sparse point clouds. In ICRA, 2020. 1", + "[36] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 1", + "[37] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh," + ], + "bbox": [ + 503, + 93, + 890, + 897 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9470", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 3", + "[38] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 2", + "[39] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2", + "[40] Prajit Ramachandran, Barret Zoph, and Quoc V. Le. Searching for activation functions. In ICLRW, 2018. 4", + "[41] Zhiyuan Ren, Zhihong Pan, Xin Zhou, and Le Kang. Diffusion motion: Generate text-guided 3d human motion by diffusion model. arXiv preprint arXiv:2210.12315, 2022. 3", + "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3", + "[43] Nikolay Savinov, Junyoung Chung, Mikolaj Binkowski, Erich Elsen, and Aaron van den Oord. Step-unrolled denoising autoencoders for text generation. arXiv preprint arXiv:2112.06749, 2021. 3", + "[44] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3D object proposal generation and detection from point cloud. In CVPR, 2019. 2", + "[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3", + "[46] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 2019. 2", + "[47] Haoru Tan, Sitong Wu, and Jimin Pi. Semantic diffusion network for semantic segmentation. arXiv preprint arXiv:2302.02057, 2023. 3", + "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2, 3, 5", + "[49] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 3", + "[50] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 4", + "[51] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In ICCV, 2019. 1, 2, 6", + "[52] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-net: Point cloud upsampling network. In CVPR, 2018. 1", + "[53] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. LION: Latent point diffusion models for 3d shape generation. In NeurIPS, 2022. 1, 2, 3, 6, 7", + "[54] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 3, 5, 7", + "[55] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In ICCV, 2021. 1, 2, 3, 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 871 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9471", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/8fe5d2c1-4f78-435e-8bde-f7a4afed8604_model.json b/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/8fe5d2c1-4f78-435e-8bde-f7a4afed8604_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8270a1a052c9180e04b2211a5046bffa6825be39 --- /dev/null +++ b/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/8fe5d2c1-4f78-435e-8bde-f7a4afed8604_model.json @@ -0,0 +1,2387 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.213, + 0.131, + 0.758, + 0.175 + ], + "angle": 0, + "content": "TIGER: Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.204, + 0.697, + 0.24 + ], + "angle": 0, + "content": "Zhiyuan Ren, Minchul Kim, Feng Liu, Xiaoming Liu \nMichigan State University" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.242, + 0.686, + 0.256 + ], + "angle": 0, + "content": "{renzhiy1, kimminc2, liufeng6, liuxm}@msu.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.326, + 0.474, + 0.675 + ], + "angle": 0, + "content": "Recently, diffusion models have emerged as a new powerful generative method for 3D point cloud generation tasks. However, few works study the effect of the architecture of the diffusion model in the 3D point cloud, resorting to the typical UNet model developed for 2D images. Inspired by the wide adoption of Transformers, we study the complementary role of convolution (from UNet) and attention (from Transformers). We discover that their respective importance change according to the timestep in the diffusion process. At early stage, attention has an outsized influence because Transformers are found to generate the overall shape more quickly, and at later stages when adding fine detail, convolution starts having a larger impact on the generated point cloud's local surface quality. In light of this observation, we propose a time-varying two-stream denoising model combined with convolution layers and transformer blocks. We generate an estimizable mask from each timestep to reweigh global and local features, obtaining time-varying fused features. Experimentally, we demonstrate that our proposed method quantitatively outperforms other state-of-the-art methods regarding visual quality and diversity. Code is available https://github.com/Zhiyuan-R/Tiger-Diffusion." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.708, + 0.21, + 0.723 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The point cloud is an essential 3D shape representation that is easy to obtain (sensed directly from laser scanning), versatile (serving as a building block of mesh), and easy to manipulate (geometrically transformed in a straightforward manner by affine matrices). 3D point cloud generative models benefit extensive applications across robotics [35], medicine [15] and content creation [36], and lay the groundwork for other vision tasks like point cloud upsampling [52], point cloud completion [21, 31] and mesh generation [9]. However, compared to 2D images, the cost and complexity of acquiring 3D point clouds make it crucial to ex" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.291, + 0.889, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.456, + 0.895, + 0.527 + ], + "angle": 0, + "content": "Figure 1. An illustration of different roles played by convolution and attention operations in the denoising model. Convolution is good for learning local relationships, and attention is optimal for modeling global relationships. We propose to merge these two properties across different timesteps in the diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.547, + 0.892, + 0.578 + ], + "angle": 0, + "content": "plore and develop efficient and effective model architectures for 3D point cloud generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.893, + 0.777 + ], + "angle": 0, + "content": "Existing point cloud generative models are built on a range of frameworks, including generative adversarial networks (GANs) [1, 5], variational autoencoders (VAEs) [24], normalizing flows [23, 27, 51]. However, these methods neither are stable in training nor produce high-fidelity results. Recently, the denoising diffusion model [18], a novel generative method, has demonstrated exceptional results in 3D point cloud generation. Typically, diffusion models perturb a point cloud using a forward process and denoising algorithm. Once trained, the model can be used to generate new point clouds by iterating the reverse process over a sequence of time steps, with each step adding more noise to the input point cloud." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Several recent works, including DPM [34] and PVD [55], have successfully applied diffusion models to 3D point cloud generation, achieving high levels of naturalness and diversity. Very recently, LION [53] further improves by encoding point clouds into latent space. However, these methods commonly utilize UNet-like convolutional networks that are originally designed for image processing. There has been limited research on developing suitable de" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9462" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "noising model architectures that are specifically tailored for 3D point cloud generation. For example, PVD and LION both utilize Point-Voxel CNN (PVCNN) [30] as their denoising model. However, we observed that these PVCNN-based denoising models require a considerable number of timesteps to establish a rough shape since the limited receptive field cannot capture the global distribution of noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.471, + 0.44 + ], + "angle": 0, + "content": "In contrast to convolution operations, we observe that the attention mechanism used in Transformers [10, 48] is more effective at capturing long-range dependencies and thus the overall shape of 3D points. This difference is illustrated in the first two rows of Fig. 1. Compared to the convolution-based diffusion model (first row), the attention-based diffusion model (second row) correctly conforms to the overall object shape at earlier denoising time steps (middle column). However, the convolution operation excels in modeling local relationships. Therefore, the details of the final output are better modeled with convolutions (last column) which is also discussed in the ablation analysis of Sec. 4.3. The finding suggests that the two operations, the convolution and the attention are complementary in their roles at different denoising timesteps (the last row of Fig. 1) and we ask the following question." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.458, + 0.445, + 0.52 + ], + "angle": 0, + "content": "How to leverage the strengths of both CNNs and Transformers and enable them to dynamically capture complementary information in denoising diffusion for 3D point cloud generation?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.689 + ], + "angle": 0, + "content": "To answer this question, we propose a Time-varying denoising model for 3D point cloud generation (TIGER), a two-stream architecture combining a shallow CNN branch and Transformer branch with a timestep-dependent mask that adaptively reassigns weights to the global Transformer feature and the local CNN feature. To enable this feature-level fusion, we design an encoder-decoder to downsample the point cloud for reducing the number of points to a manageable size for Transformer. And we learn a selection mask that finds the optimal configuration at each timestep." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Furthermore, in this two-stream architecture where global and local information is modeled separately, we discover that supplying 3D points position information to the global branch is critical to the performance. To supply the 3D position information more effectively, we design two novel position encoding methods: phase-shifted position encoding and Base- \\(\\lambda\\) position encoding. They create 3D position information in a continuous space and retain the property of linear expression of relative position. The effect of the time-dependent dual branch mask and the proposed position embedding is ablated in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "A unique property of TIGER is that the weights of the two branches are dependent on the diffusion timesteps, and they are learned and optimized for the 3D point cloud gen" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "eration. Interestingly, experiments demonstrate that the weight of the convolution branch is monotonically increasing with the timestep, which is in accordance with our initial observation that convolutions are good for modeling fine details. Empirically we show that TIGER achieves state-of-the-art (SoTA) performance in 3D point cloud generation." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.184, + 0.733, + 0.197 + ], + "angle": 0, + "content": "Our main contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.199, + 0.892, + 0.258 + ], + "angle": 0, + "content": "- We propose a novel two-stream denoising model, which uses timestep to optimally reweigh the global feature from Transformer and the local feature from shallow CNN." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.26, + 0.892, + 0.319 + ], + "angle": 0, + "content": "- We design two novel 3D continuous position encoding methods and position-aware self-attention to enable our Transformer branch to aggregate global features effectively from the latent point cloud input." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.32, + 0.892, + 0.365 + ], + "angle": 0, + "content": "- Experiments show that our method outperforms prior works in 3D point cloud generation tasks qualitatively and quantitatively on the ShapeNetv2 dataset." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.199, + 0.892, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.382, + 0.642, + 0.397 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.527 + ], + "angle": 0, + "content": "3D Point Cloud Generation. The advances in 3D point cloud generation can be categorized by how the input is represented. Early methods [1, 12, 14, 38, 39, 44] represent the point cloud as a \\(N \\times 3\\) matrix, where \\(N\\) is the predefined number of 3D points. The backbone comprises 1D convolutions and max-pooling to allow the points' permutation invariance. So, the shape is represented by a global feature vector." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.892, + 0.68 + ], + "angle": 0, + "content": "More recent works [5, 7, 13, 27, 34, 51] have improved upon the input representation by introducing a probabilistic distribution assumption, where each point cloud is considered as one data point sampled from the distribution. For instance, PointFlow [51] models the distribution of points by continuous normalizing flows, allowing the sampling of an arbitrary number of points to represent a point cloud. However, the backbone architecture is still based on 1D convolution-based encoder with the max-pooling operation [38]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.681, + 0.892, + 0.833 + ], + "angle": 0, + "content": "Recently, PVD [55] and LION [53] adopt a hybrid point-voxel representation [30] which voxelize the point clouds for 3D convolution. This representation with rich spatial information combines well with the diffusion objective and produces promising results [53]. However, the backbone is still confined to convolution-based networks, lacking the capacity to effectively model the long range-dependencies. We, on the other hand, find the complementary roles of attention and convolution in the point cloud generator and propose a network that combines the benefits of both." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Diffusion Models. Diffusion [18] and score-based methods [22, 46] have enjoyed remarkable success in images synthesis. Diffusion has been extended to many generation" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9463" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.105, + 0.881, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.283, + 0.895, + 0.354 + ], + "angle": 0, + "content": "Figure 2. Illustration of our time-varying two-stream architecture (TIGER). The network's input is a noisy point cloud \\(\\mathbf{X}_t\\) at timestep \\(t\\), and the goal is to predict the noise in \\(\\mathbf{X}_t\\) to obtain the denoised point cloud \\(\\mathbf{X}_{t-1}\\). We first encode \\(\\mathbf{X}_t\\) into latent point cloud \\(\\hat{\\mathbf{X}}_t\\) and export coordinates information for position encoding and decoder. We feed the latent point cloud into both the Latent Point Cloud Transformer and a shallow CNN branch to get global feature \\(\\mathbf{X}_{tr}\\) and local feature \\(\\mathbf{X}_c\\). Then, we combine two features using timestep dependent mask \\(\\mathbf{M}_c\\) to create a fused feature \\(\\hat{\\mathbf{X}}_F\\). Last, we apply our latent point cloud decoder to predict the noise from \\(\\hat{\\mathbf{X}}_F\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.376, + 0.47, + 0.451 + ], + "angle": 0, + "content": "tasks [2, 8, 17, 20, 25, 28, 41-43, 47], and largely improved by innovative techniques, such as fast sampling [33, 45], classifier-free guidance [19, 37] and latent diffusion [4, 42]. Recently 3D point cloud generation with diffusion [34, 53, 55] is also proposed with remarkable diversity and fidelity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.473, + 0.573 + ], + "angle": 0, + "content": "However, previous works do not investigate the behavior of the diffusion backbone with respect to the diffusion timestep. We find that the network equipped with both convolution and attention operations learns to utilize the local and global shape representation with different magnitudes based on the timestep. This time adaptive architecture allows the 3D point cloud generator to be trained faster than LION [53] while producing higher-quality outputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.47, + 0.68 + ], + "angle": 0, + "content": "Transformer for Point Clouds. Transformer architecture [48] is intrinsically a set operator, which is suitable for permutation-invariant point clouds. PCT [16] applies global attention to the point cloud. Point Transformer V1 [54] and V2 [49] convert global attention to local attention to reduce memory consumption and computational complexity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.773 + ], + "angle": 0, + "content": "Contrary to the previous methods, our method decouples the global and local information by using Transformer to aggregate global information effectively while using CNN to model the local information. Also, we find that position embedding is crucial to performance and propose a new position embedding to further improve the performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.17, + 0.8 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this section, we first briefly formulate the probabilistic model of forward and reverse diffusion processes for 3D point cloud generation. Then, we dive into the details of our time-varying two-stream architecture, including the encoder part, latent point Transformer, time mask generator, and decoder part. The entire architecture is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.375, + 0.704, + 0.39 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.504 + ], + "angle": 0, + "content": "Given a point cloud \\(\\mathbf{X}_0\\in \\mathbb{R}^{N\\times 3}\\), where \\(N\\) is the number of points, the diffusion forward process gradually diffuses the original point cloud into Gaussian noise \\(\\mathbf{X}_T\\); the forward process can be formulated with a transition probability \\(q(\\mathbf{X}_t|\\mathbf{X}_{t - 1})\\). By giving a sequence of pre-defined noising scales \\(\\beta_{1},\\beta_{2},\\ldots ,\\beta_{T}\\), the transition probability can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.51, + 0.892, + 0.53 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {X} _ {t} \\mid \\mathbf {X} _ {t - 1}\\right) = \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} \\mathbf {X} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.581 + ], + "angle": 0, + "content": "For the reverse process, we use learn the \\(p_{\\theta}(\\mathbf{X}_{t - 1}|\\mathbf{X}_t)\\) a Gaussian distribution which approximates the intractable real distribution \\(q(\\mathbf{X}_{t - 1}|\\mathbf{X}_t)\\). Specifically," + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.586, + 0.892, + 0.605 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} = \\mathcal {N} \\left(\\mu_ {\\theta} \\left(\\mathbf {X} _ {t}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.892, + 0.732 + ], + "angle": 0, + "content": "where we learn the mean \\(\\mu_{\\theta}(\\mathbf{X}_t,t)\\) and set the variance \\(\\sigma_t^2\\) to a fixed schedule as in [18]. Conventionally, \\(\\mu_{\\theta}(\\mathbf{X}_t,t)\\) is modeled using a UNet backbone and it depends on the timestep \\(t\\) to scale and shift the feature. We design a two-stream backbone where the timestep modulates the weight between the two streams. To sample a 3D point cloud, we iteratively denoise \\(\\mathbf{X}_T,T\\) times to finally generate the highfidelity samples. In training, we optimize the MSE loss:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.736, + 0.892, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {s i m p l e}} = \\mathbb {E} _ {t \\sim [ 1, T ]} \\| \\mu - \\mu_ {\\theta} (\\mathbf {X} _ {t}, t) \\| _ {2} ^ {2} \\tag {3} \\\\ = \\mathbb {E} _ {t \\sim [ 1, T ]} \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {X} _ {t}, t) \\| _ {2} ^ {2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.81 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is the ground truth noise and \\(\\epsilon_{\\theta}\\) is the noise prediction by our diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.818, + 0.747, + 0.833 + ], + "angle": 0, + "content": "3.2. Noisy Point Cloud Encoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To effectively represent the \\(N\\times 3\\) points, we adopt the voxelization encoding scheme of PVCNN [30]. This representation works well with both CNN and Transformer architecture as we can downsample the voxel-grid following [11] to" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9464" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.096, + 0.468, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.232, + 0.471, + 0.275 + ], + "angle": 0, + "content": "Figure 3. The overview of our encoder. It extracts features in voxel space and downsamples the point cloud by applying trilinear interpolation of voxel with grouped points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.469, + 0.405 + ], + "angle": 0, + "content": "increase computation efficiency. Specifically, a noisy point cloud encoder \\(\\mathcal{E}:\\mathbb{R}^{N\\times 3}\\to \\mathbb{R}^{M\\times d}\\) transforms the noisy point cloud \\(\\mathbf{X}_t\\in \\mathbb{R}^{N\\times 3}\\) at timestep \\(t\\) into a latent point cloud representation \\(\\hat{\\mathbf{X}}_t\\in \\mathbb{R}^{M\\times d}\\), where \\(N\\) is the number of original points, \\(M\\) is the number of downsampled points and \\(d\\) is the dimension of each latent point's feature. The overview of our encoder is illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.405, + 0.469, + 0.466 + ], + "angle": 0, + "content": "To aggregate features in a rigid field, we first voxelize the noisy point cloud \\(\\{x_{n},y_{n},z_{n}\\} = \\mathbf{X}_{t}\\in \\mathbb{R}^{N\\times 3}\\) into the voxel grids \\(\\{\\mathbf{V}_{u,v,w}\\} = \\mathbf{V}\\in \\mathbb{R}^{L\\times L\\times L}\\), where \\(L\\) is the resolution and \\(n\\) is the point index in \\(N\\). Specifically," + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.476, + 0.469, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} _ {u, v, w} = \\sum_ {n = 1} ^ {N} \\mathbb {I} \\left(x _ {n} \\in \\mathbb {N} (u, \\delta), y _ {n} \\in \\mathbb {N} (v, \\delta), z _ {n} \\in \\mathbb {N} (w, \\delta)\\right) * \\mathbf {f} _ {n}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.469, + 0.626 + ], + "angle": 0, + "content": "where \\(\\mathbb{N}(z,\\delta) = \\{x|x\\in [z - \\delta ,z + \\delta)\\}\\), \\(\\mathbf{f}_n\\) denotes the \\(n^{th}\\) extracted latent feature corresponding to the \\(n^{th}\\) point \\((\\mathbf{X}_t)_n\\) and \\(\\mathbb{I}(\\cdot)\\) is an indicator of whether the \\(n^{th}\\) point falls into the voxel grid \\((u,v,w)\\). After voxelization, we apply a series of 3D convolution, Swish [40] activation and Group-Norm [50]. After applying a series of these operations, this results in a latent volume \\(\\hat{\\mathbf{V}}\\in \\mathbb{R}^{L\\times L\\times L\\times d}\\), where \\(d\\) is the channel size for each voxel." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.469, + 0.763 + ], + "angle": 0, + "content": "We use furthest point sampling algorithm [11] to downsample the input noisy point cloud \\(\\mathbf{X}_t\\in \\mathbb{R}^{N\\times 3}\\) into a sparser point cloud \\(\\mathbf{X}_t^s\\in \\mathbb{R}^{M\\times 3}\\) (\\(M < N\\)). Then, as the position in volume is discrete, we query the latent volume \\(\\hat{\\mathbf{V}}\\) by the sparse point cloud \\(\\mathbf{X}_t^s\\) via trilinear interpolation, eventually transformed into latent point cloud \\(\\hat{\\mathbf{X}}_t\\in \\mathbb{R}^{M\\times d}\\). It is worth noting we preserve the coordinates of \\(\\mathbf{X}_t\\) and \\(\\mathbf{X}_t^s\\) to do upsampling with the decoder and provide position embedding for our latent point cloud Transformer." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.364, + 0.787 + ], + "angle": 0, + "content": "3.3. Latent Point Cloud Transformer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Previous denoising models such as PVCNN [30] simply apply stacked convolutional layers to create feature representations. Contrarily, we aim to better model the global relationship by introducing a novel Transformer branch named latent point cloud Transformer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Tokenization. After encoding, we get a latent point cloud representation \\(\\hat{\\mathbf{X}}_t\\in \\mathbb{R}^{M\\times d}\\) and its coordinates \\(\\mathbf{X}_t^s\\in \\mathbb{R}^{M\\times 3}\\)." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.695, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.204, + 0.626, + 0.215 + ], + "angle": 0, + "content": "(a) PSPE" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.091, + 0.888, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.203, + 0.822, + 0.215 + ], + "angle": 0, + "content": "(b) BλPE" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.226, + 0.892, + 0.268 + ], + "angle": 0, + "content": "Figure 4. Illustration of two examples of the position embedding from PSPE and BλPE respectively. Both methods show distinguished representation for each position." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.293, + 0.892, + 0.338 + ], + "angle": 0, + "content": "Following [29], we use dual PatchNorm to project the latent point cloud into tokens, which place LayerNorm before and after an MLP layer for more stable training:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.346, + 0.892, + 0.365 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} = \\operatorname {L N} \\left(\\operatorname {M L P} \\left(\\operatorname {L N} \\left(\\hat {\\mathbf {X}} _ {\\mathbf {t}}\\right)\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.372, + 0.891, + 0.403 + ], + "angle": 0, + "content": "where \\(\\mathbf{T} \\in \\mathbb{R}^{M \\times D}\\) and \\(D\\) is the dimension in the token space (in our paper, \\(D = 2d\\))." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.403, + 0.892, + 0.569 + ], + "angle": 0, + "content": "Position Encoding. Since the order of the latent points is arbitrary in a point cloud, it's necessary to create a 3D position embedding and fuse it into the feature, so that the model knows the position of each point. Contrary to the position embedding in image domains [10, 26], the design of 3D position embedding plays a vital role in the performance and our position embedding produces the best result. We propose two novel 3D space continuous position encoding methods: Phase Shift Position Encoding (PSPE) and Base- \\(\\lambda\\) Position Encoding (B\\(\\lambda\\)PE). For PSPE, we use Sine and Cosine functions of different frequencies and phases:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.577, + 0.891, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {e m b} \\left(p o s _ {j}, 6 i + 2 * (j - 1)\\right) = \\sin \\left(\\frac {p o s _ {j}}{1 0 0 0 0 ^ {\\frac {2 i}{D}}} + (j - 1) \\frac {2 \\pi}{3}\\right) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.619, + 0.897, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {e m b} (p o s _ {j}, 6 i + 1 + 2 * (j - 1)) = \\cos \\left(\\frac {p o s _ {j}}{1 0 0 0 0 ^ {\\frac {2 i}{D}}} + (j - 1) \\frac {2 \\pi}{3}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.661, + 0.892, + 0.766 + ], + "angle": 0, + "content": "where \\(pos_{j}\\) is \\(j\\)-th element in each point's position, specifically \\(pos_{1} = x\\), \\(pos_{2} = y\\), \\(pos_{3} = z\\), and \\(i\\) is the index of the dimension. We maximize the phase shift difference \\((\\frac{4\\pi}{3})\\) of different axes to distinguish them and combine Sine and Cosine to guarantee the linear representation of relative position, which will be proved mathematically in Supp. An example is shown in Fig. 4a and \\(\\mathbf{P}_{emb}\\) is shaped \\(\\mathbb{R}^{M\\times D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.767, + 0.892, + 0.824 + ], + "angle": 0, + "content": "However, PSPE requires 6 channels to represent one frequency feature, which is quite inefficient. To improve channel utilization, we apply BλPE to compress 3D position into one scalar:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.826, + 0.891, + 0.842 + ], + "angle": 0, + "content": "\\[\np o s = \\lambda^ {2} * z + \\lambda * y + x \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.845, + 0.891, + 0.874 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {e m b} (p o s, 2 i) = \\sin \\left(\\frac {p o s}{1 0 0 0 0 ^ {\\frac {2 i}{D}}}\\right) \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.877, + 0.891, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {e m b} (p o s, 2 i + 1) = \\cos \\left(\\frac {p o s}{1 0 0 0 0 ^ {\\frac {2 i}{D}}}\\right), \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9465" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.093, + 0.46, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.324, + 0.471, + 0.368 + ], + "angle": 0, + "content": "Figure 5. The overview of the position-aware self-attention. We compute position relationship map \\(\\mathbf{H}\\) by previous position embedding to reweigh the \\(\\mathbf{Q}\\mathbf{K}^T\\) matrix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.47, + 0.5 + ], + "angle": 0, + "content": "where \\(pos\\) is a polynomial expression of \\(\\lambda\\). Specifically, we set \\(\\lambda\\) to 1000, which means this preserves three decimal places of precision and one example is shown in Fig. 4b. The choice between these two position encoding methods is essentially a trade-off between granularity and expressiveness. We ablate them and other non-position encoding and learnable position encoding in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.47, + 0.562 + ], + "angle": 0, + "content": "Transformer and Position-Aware Attention We stack \\(L\\) layers of Transformer block [48] which comprises of LayerNorm(LN) [3] and Multihead Self Attention (MHA) and MultiLayer Perceptrons. We let the token \\(\\mathbf{T}_0 = \\mathbf{T}\\) and let" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.576, + 0.469, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {l *} = \\operatorname {M H A} \\left(\\ln \\left(\\mathbf {T} _ {l - 1}\\right)\\right) + \\mathbf {T} _ {l - 1}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.595, + 0.469, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {l} = \\operatorname {M L P} \\left(\\operatorname {L N} \\left(\\mathbf {T} _ {l *}\\right)\\right) + \\mathbf {T} _ {l *}, \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.614, + 0.469, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {t r} := \\mathbf {T} _ {L}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.751 + ], + "angle": 0, + "content": "To add position embedding into the model, one can simply add it to \\(\\mathbf{T}\\). But as the Transformer blocks are gradually stacked, the continuous 3D position encoding is difficult to retain in the token representation. Inspired by [54], we explicitly add the position encoding to each MHA in order to make our attention map position-aware, as shown in Fig. 5. Specifically, our position-aware MHA modified as" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.763, + 0.469, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} = \\operatorname {S o f t m a x} \\left(\\left(\\mathbf {P} _ {e m b} \\mathbf {W} _ {p}\\right) \\left(\\mathbf {P} _ {e m b} \\mathbf {W} _ {p}\\right) ^ {T}\\right) \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.796, + 0.469, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {l *} = \\operatorname {S o f t m a x} \\left(\\frac {\\left(\\mathbf {T} _ {l} \\mathbf {W} _ {q}\\right) \\left(\\mathbf {T} _ {l} \\mathbf {W} _ {k}\\right) ^ {T} \\odot \\mathbf {H}}{\\sqrt {D}}\\right) \\left(\\mathbf {T} _ {l} \\mathbf {W} _ {v}\\right) \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.47, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_q,\\mathbf{W}_k,\\mathbf{W}_v,\\mathbf{W}_p\\in \\mathbb{R}^{D\\times Z}\\) are projection matrices and \\(\\mathbf{H}\\in \\mathbb{R}^{M\\times M}\\) is the position relationship map. Modified MHA could incorporate position information explicitly and sufficiently during training. \\(\\mathbf{X}_{tr}\\) is the final output." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.097, + 0.891, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.202, + 0.893, + 0.245 + ], + "angle": 0, + "content": "Figure 6. The overview of our decoder. By querying the latent volume \\(\\hat{\\mathbf{X}}_F\\) with previous coordinates information \\(\\mathbf{X}_t\\), we can upsample the latent point cloud into 3D space." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.266, + 0.708, + 0.28 + ], + "angle": 0, + "content": "3.4. Time Mask Generator" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.288, + 0.892, + 0.393 + ], + "angle": 0, + "content": "We apply a depth-2 PVCNN [30] to get local feature \\(\\mathbf{X}_c\\in\\) \\(\\mathbb{R}^{M\\times D}\\) and our latent point cloud Transformer to get global feature \\(\\mathbf{X}_{tr}\\in \\mathbb{R}^{M\\times D}\\). Time mask generator module scales the local and global features by the time variable. We force the sum of two learnable time masks to be 1. With this constraint, the trade-off of local and global features can be learned by the network." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.394, + 0.892, + 0.424 + ], + "angle": 0, + "content": "We first align the dimension and scale of \\(\\mathbf{X}_c\\) and \\(\\mathbf{X}_{tr}\\) by one layer MLP and Sigmoid function \\(\\sigma\\)." + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.434, + 0.891, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {c} ^ {*} = \\sigma (\\operatorname {M L P} (\\mathbf {X} _ {c})), \\quad \\mathbf {X} _ {t r} ^ {*} = \\sigma (\\operatorname {M L P} (\\mathbf {X} _ {t r})). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.461, + 0.892, + 0.522 + ], + "angle": 0, + "content": "Then we encode a timestep \\(t\\) to a sinusoidal position embedding [18, 48] to \\(\\mathbf{t}_{emb} \\in \\mathbb{R}^c\\). We use two stacked MLP to map \\(\\mathbf{t}_{emb}\\) to a higher dimension space and apply Sigmoid to keep one of the masks between 0 to 1. Specifically," + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.532, + 0.891, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {c} = \\sigma (\\mathrm {M L P} (\\text {L e a k y R e L U} (\\mathrm {M L P} (\\mathbf {t} _ {\\text {e m b}}))), \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.551, + 0.891, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {t r} = \\mathbf {1} - \\mathbf {M} _ {c}. \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.578, + 0.892, + 0.655 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}_c\\), \\(\\mathbf{M}_{tr} \\in \\mathbb{R}^D\\) are the time masks for CNN and Transformer feature respectively. We ablate the dimension of masks in Sec. 4.3 to be \\(\\mathbb{R}^D\\) or \\(\\mathbb{R}^1\\). We multiply these two masks by the respective features and adjust the fused feature by MLP." + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.664, + 0.891, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {X}} _ {F} = \\operatorname {M L P} \\left(\\mathbf {X} _ {c} ^ {*} \\odot \\mathbf {M} _ {c} + \\mathbf {X} _ {t r} ^ {*} \\odot \\mathbf {M} _ {c}\\right), \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.691, + 0.891, + 0.723 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{X}}_F\\in \\mathbb{R}^{M\\times D}\\) is the final feature adaptively aggregating local and global features based on the timestep." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.731, + 0.753, + 0.746 + ], + "angle": 0, + "content": "3.5. Latent Point Cloud Decoder" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.754, + 0.892, + 0.874 + ], + "angle": 0, + "content": "In order to upsample the latent point cloud and predict the corresponding noise, we design a latent point cloud decoder \\(\\mathcal{D}\\), which converts the latent point cloud \\(\\hat{\\mathbf{X}}_F \\in \\mathbb{R}^{M \\times D}\\) to \\(\\mathbf{X}_t \\in \\mathbb{R}^{N \\times 3}\\). Similar to Sec. 3.2, we can voxelize \\(\\hat{\\mathbf{X}}_F\\) into a latent volume \\(\\tilde{\\mathbf{V}}_F \\in \\mathbb{R}^{L \\times L \\times L \\times D}\\), and take \\(\\mathbf{X}_t\\) as the query for trilinear interpolation, eventually resulting in the final noise prediction \\(\\epsilon_{\\theta} \\in \\mathbb{R}^{N \\times 3}\\). This formulation can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.884, + 0.891, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\theta} = \\operatorname {T r i l i n e a r} \\left(\\operatorname {V o x e l i z e} \\left(\\hat {\\mathbf {X}} _ {F}\\right), \\mathbf {X} _ {t}\\right) \\mathbf {W} ^ {D \\times 3}, \\tag {20}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9466" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.851, + 0.306 + ], + "angle": 0, + "content": "
MethodGenerative ModelAirplaneChairCar
CD→50%EMD→50%CD→50%EMD→50%CD→50%EMD→50%
r-GAN [1]GAN98.4096.7983.6999.7094.4699.01
l-GAN(CD) [1]GAN87.3093.9568.5883.8466.4988.78
l-GAN(EMD) [1]GAN89.4976.9171.9064.6571.1666.19
PointFlow [51]Normalizing Flow75.6870.7462.8460.5758.1056.52
DPF-Net [27]Normalizing Flow75.1865.5562.0058.5362.3554.48
ShapeGF [5]GAN80.0076.1768.9665.4863.2056.53
SoftFlow [23]Normalizing Flow76.0565.8059.2160.0564.7760.09
SetVAE [24]VAE76.5467.6558.8460.5759.9559.94
DPM [34]Diffusion76.4286.9160.0574.7768.8979.97
PVD [55]Diffusion73.8264.8156.2653.3254.5553.83
TIGERDiffusion71.8555.8254.6152.7154.3152.24
LION [53]Diffusion67.4161.2353.7052.3453.4151.14
TIGERDiffusion67.2156.2654.3251.7154.1250.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.312, + 0.894, + 0.342 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with baselines using 1-NN. Both CD and EMD are considered, where CD is multiplied by \\( {10}^{3} \\) and EMD is multiplied by \\( {10}^{2} \\) . The closer the score is to 50%,the better the quality and diversity of generated samples. [Key: Best, Second Best]" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.354, + 0.868, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.083, + 0.57, + 0.885, + 0.586 + ], + "angle": 0, + "content": "Figure 7. Our generation results (right) compared to baseline models (left). TIGER generates high-quality and diverse 3D point clouds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.47, + 0.658 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}^{D\\times 3}\\) is the projection matrix to map the noise into 3D space and \\(\\mathbf{X}_t\\) is the original point cloud coordinates. The overview of our decoder is shown in Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.671, + 0.21, + 0.688 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.267, + 0.713 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Datasets. We follow previous works and choose ShapeNetv2 [6] as our main dataset, which is pre-processed by PointFlow [51]. To fairly compare with other different baseline methods, we train and evaluate on three categories: airplane, chair and car. For each shape, we sample 2,048 points and normalize them globally across the entire dataset. The training set consists of 2,832,4,612, and 2,458 shapes, while the evaluation set is composed of 405,662, and 352 shapes for airplanes, chairs, and cars, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Metrics. Following the baselines PVD [55] and LION [53], we use 1-NN (1-nearest neighbor) accuracy [32] as our evaluation metric. This metric has been shown to effectively" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.892, + 0.702 + ], + "angle": 0, + "content": "measure both the quality and diversity of generated point clouds and a score closer to \\(50\\%\\) indicates superior performance [51]. To calculate the 1-NN distance matrix, we consider two commonly used metrics for measuring the distance between point clouds: CD (Chamfer Distance) and EMD (Earth Movers' Distance)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.712, + 0.787, + 0.728 + ], + "angle": 0, + "content": "4.2. Comparison with SoTA methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Results. To compare our time-varying denoising model with SoTA methods, we evaluate its performance against competitive models such as SetVAE [24], DPM [34], PVD [55] and LION [53]. It is noteworthy that in order to compare with LION, which uses a different dataset splitting strategy (sampling from the first 10,000 points instead of the latter 5,000 points), we evaluate our model under both LION's strategy and the standard strategy used by other methods. As shown in Tab. 1, we outperform LION in four out of six metrics. Compared to other methods, our performance is significantly better. Notably, our EMD-based" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9467" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.152 + ], + "angle": 0, + "content": "
MethodTraining Time (GPU hours)Inference Time (s)
PVD [55]1428.46
LION [53]55027.12
Tiger1649.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.157, + 0.468, + 0.185 + ], + "angle": 0, + "content": "Table 2. Training and inference time comparison. Training time is counted on average for three categories: airplane, chair and car." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.192, + 0.468, + 0.274 + ], + "angle": 0, + "content": "
MethodW1W2Mean CDMean EMD
Convolution1061.3456.31
Transformer0162.4155.85
Conv+Tran0.50.564.4557.92
Ourslearnable (scalar)60.7954.12
Ourslearnable (channel-wise)59.4353.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.28, + 0.468, + 0.322 + ], + "angle": 0, + "content": "Table 3. Ablation of different architecture designs. \\( W_{1} \\) and \\( W_{2} \\) are the weights for the ConvNet branch and Transformer branch. [Key: Best, Second Best]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.468, + 0.445 + ], + "angle": 0, + "content": "1-NN accuracy shows a significant improvement of \\(13.45\\%\\) relative to PVD and \\(8.12\\%\\) relative to LION for airplanes. This is attributed to the efficient aggregation of global features by our Transformer branch, which is especially beneficial for EMD as it is primarily influenced by global distribution. Fig. 7 shows that our generated samples are of high quality and diversity compared with other methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.468, + 0.566 + ], + "angle": 0, + "content": "Efficiency. To compare the training and inference speed, we report the GPU hours with NVidia V100 GPUs as in LION [53]. As shown in Tab. 2, our model trains much faster than LION, with only a quarter of its training time and a third of its inference time. Our training and inference speed is comparable to that of PVD while surpassing LION and PVD in the generation quality as shown in Tab. 1. More computation analysis can be found in Supp." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.576, + 0.283, + 0.59 + ], + "angle": 0, + "content": "4.3. Ablation and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Time-varying two-stream architecture. We ablate our time-varying two-stream architecture by comparing it with single ConvNet, single Transformer, and two-stream networks with equal weight. In this ablation, we also compare the performance of our time masking with scalar value setting and channel-wise value setting. The results, presented in the first two rows of Tab. 3, show that Transformer performs better on the EMD metric, while ConvNet performs better on the CD metric. Using a fixed equal weight for the two-stream network leads to a significant performance drop on both metrics. However, the learnable time mask improves the performance on both metrics, indicating the importance of the time-varying design. Further, channel-wise time masking improves performance slightly more than scale value setting due to its greater expressivity. Transformer backbones, position encoding, and self-attention strategies. Here we carried out three ablations: (1) We compare the performance of our Transformer backbone with two classic Transformer backbones, PCT [16] and Point Transformer (PointTran) [54]. (2) To assess" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.089, + 0.886, + 0.24 + ], + "angle": 0, + "content": "
TBPEAPMean CDMean EMD
PCT [16]Learnable-65.8260.21
PointTran [54]LearnableLearnable66.4262.39
Ours--63.5157.98
OursLearnableLearnable62.1756.13
OursPSPE-60.1155.74
OursPSPELearnable61.4855.97
OursPSPEPSPE59.0254.49
OursBλPE-61.3554.27
OursBλPELearnable61.8956.02
OursBλPEBλPE59.4353.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.245, + 0.892, + 0.3 + ], + "angle": 0, + "content": "Table 4. Ablation of Transformer backbones, position encoding, and self-attention strategies. [Key: Best, Second Best; TB=Transformer backbone; PE=position encoding; AP=additional position information]" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.314, + 0.72, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.312, + 0.884, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.47, + 0.892, + 0.594 + ], + "angle": 0, + "content": "Figure 8. Visualization of attention maps in 3D point clouds. Column (c) is the result of column (b) multiplied by position relationship map H. Taking the first row as an example, queried by the airplane head's position, our method can aggregate the position information from the airplane's right wing and left elevator compared with standard self-attention. This comparison highlights the effectiveness of our position-aware self-attention module in capturing long-range relationships in 3D point clouds. [SA: Self-attention, PASA: Position-aware Self-attention]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "our proposed position encoding methods, PSPE and BλPE, we compare them with no position encoding and learnable position encoding. (3) We investigate our position-aware self-attention module by comparing it with standard self-attention and self-attention with learnable position encoding as proposed in [54]. As shown in Tab. 4, our latent point cloud Transformer outperforms both PCT and Point Transformer under various settings. Furthermore, our proposed position encoding methods, PSPE and BλPE, significantly improve performance compared to no position encoding or learnable position encoding. Specifically, PSPE improves the CD metric, while BλPE improves the EMD metric. Our position-aware self-attention module, which aggregates corresponding position encoding, achieves better results than standard self-attention and self-attention with learnable position encoding. To illustrate this, Fig. 8 visualizes the attention map in 3D point space for both standard self-attention and our position-aware self-attention. This visualization demonstrates the ability of our method to capture" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9468" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.087, + 0.349, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.206, + 0.26, + 0.219 + ], + "angle": 0, + "content": "(a) Airplane" + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.088, + 0.614, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.458, + 0.206, + 0.514, + 0.218 + ], + "angle": 0, + "content": "(b) Chair" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.088, + 0.882, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.206, + 0.77, + 0.218 + ], + "angle": 0, + "content": "(c) Car" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.228, + 0.895, + 0.259 + ], + "angle": 0, + "content": "Figure 9. The changes of ConvNet importance for each category over timesteps with the channel-wise time mask setting. The thick curve is the result of Gaussian smoothing." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.265, + 0.891, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.45, + 0.852, + 0.466 + ], + "angle": 0, + "content": "Figure 10. High quality and diverse 3D point clouds generated from our TIGER model trained on 55 ShapeNetv2 categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.316, + 0.502 + ], + "angle": 0, + "content": "long-range relationships in the data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.47, + 0.744 + ], + "angle": 0, + "content": "Weight changing over timesteps. To validate our motivation of reweighing the local and global features from ConvNet and Transformer at different time steps, we visualize the changes in weights. For channel-wise time mask setting, we use the ratio of the weight of the local feature over the weight of the global feature to measure the importance of ConvNet and Transformer at each timestep. Fig. 9 indicates that the importance of ConvNet for each category increases nearly monotonically, implying that the diffusion process of 3D point cloud generation benefits from global features at the early stages of diffusion and requires the aggregation of more local features at later stages. Moreover, we observe that the curves of airplanes and chairs change more drastically than that of cars, which we hypothesize is due to the complicated shapes of airplanes and chairs, resulting in a more complex diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.47, + 0.82 + ], + "angle": 0, + "content": "Unconditional generation of multiple classes. To further demonstrate the generalizability of our model, we train a universal TIGER model on all 55 categories of ShapeNetv2 [6]. Fig. 10 shows that our universal model can produce shapes with high quality and diversity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.205, + 0.847 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose a novel time-varying denoising model with a latent point transformer backbone. Experimentally, we show how the time mask optimally aggre" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.593 + ], + "angle": 0, + "content": "gates local and global features at different timesteps, which quantitatively reaches SoTA performance. In accordance with our hypothesis, Transformer is more important at early diffusing stages and ConvNet has a larger weight at later timesteps in the 3D point cloud generation task. Future work will be focused on exploring how our time-varying denoising model would behave in other diffusion-based tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.893, + 0.763 + ], + "angle": 0, + "content": "Limitations. Although we generate high-quality and natural samples, we cannot control the category of the generated shape. One potential solution is to encode the category into latent space and train the diffusion model conditioned on the latent code. We also acknowledge that our model needs to get the local feature and global feature by two subnetworks, which is not quite efficient. In this work, we gain efficiency with the model converging with fewer epochs. But future works can increase the backbone efficiency by proposing time-varying properties with only one network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Potential Negative Impacts. While it has the potential to speed up progress in both areas, it also has the risk of eliminating the jobs of game and furniture designers due to automation. However, our model should have a mostly positive impact, because the most likely use of our method would be to assist designers not completely replace their jobs. It is still necessary to select suitable shapes among many generated samples." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9469" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.196 + ], + "angle": 0, + "content": "[2] Vishal Asnani, John Collomosse, Tu Bui, Xiaoming Liu, and Shruti Agarwal. ProMark: Proactive diffusion watermarking for causal attribution. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.197, + 0.468, + 0.223 + ], + "angle": 0, + "content": "[3] Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.223, + 0.468, + 0.276 + ], + "angle": 0, + "content": "[4] Fan Bao, Kun Xu, Chongxuan Li, Lanqing Hong, Jun Zhu, and Bo Zhang. Variational (gradient) estimate of the score function in energy-based latent variable models. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.277, + 0.468, + 0.33 + ], + "angle": 0, + "content": "[5] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snavely, and Bharath Hariharan. Learning gradient fields for shape generation. In ECCV, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.468, + 0.399 + ], + "angle": 0, + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.4, + 0.468, + 0.439 + ], + "angle": 0, + "content": "[7] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. NeurIPS, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.44, + 0.468, + 0.479 + ], + "angle": 0, + "content": "[8] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusiondet: Diffusion model for object detection. arXiv preprint arXiv:2211.09788, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.468, + 0.506 + ], + "angle": 0, + "content": "[9] Tamal K Dey and James Hudson. Pmr: Point to mesh rendering, a feature-based approach. In VIS, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.506, + 0.468, + 0.588 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.468, + 0.628 + ], + "angle": 0, + "content": "[11] Yuval Eldar, Michael Lindenbaum, Moshe Porat, and Yehoshua Y Zeevi. The farthest point strategy for progressive image sampling. In ICPR, 1994. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.468, + 0.667 + ], + "angle": 0, + "content": "[12] Matheus Gadelha, Rui Wang, and Subhransu Maji. Multiresolution tree networks for 3d point cloud processing. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.668, + 0.468, + 0.722 + ], + "angle": 0, + "content": "[13] Will Grathwohl, Ricky TQ Chen, Jesse Bettencourt, Ilya Sutskever, and David Duvenaud. Ffjord: Free-form continuous dynamics for scalable reversible generative models. arXiv preprint arXiv:1810.01367, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.722, + 0.468, + 0.775 + ], + "angle": 0, + "content": "[14] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3D surface generation. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.468, + 0.816 + ], + "angle": 0, + "content": "[15] Jia Guo, Xuanxia Yao, Mengyu Shen, Jiafei Wang, and Wanyou Liao. A deep learning network for point cloud of medicine structure. In ITME, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.468, + 0.869 + ], + "angle": 0, + "content": "[16] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.87, + 0.468, + 0.897 + ], + "angle": 0, + "content": "[17] Kilian Konstantin Haefeli, Karolis Martinkus, Nathanael Perraudin, and Roger Wattenhofer. Diffusion models for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.094, + 0.892, + 0.119 + ], + "angle": 0, + "content": "graphs benefit from discrete state spaces. arXiv preprint arXiv:2210.01549, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.12, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.147, + 0.892, + 0.173 + ], + "angle": 0, + "content": "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.173, + 0.892, + 0.213 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.213, + 0.892, + 0.252 + ], + "angle": 0, + "content": "[21] Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.253, + 0.892, + 0.293 + ], + "angle": 0, + "content": "[22] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[23] Hyeongju Kim, Hyeonseung Lee, Woo Hyun Kang, Joun Yeop Lee, and Nam Soo Kim. Softflow: Probabilistic framework for normalizing flow on manifolds. In NeurIPS, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.347, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[24] Jinwoo Kim, Jaehoon Yoo, Juho Lee, and Seunghoon Hong. Setvae: Learning hierarchical composition for generative modeling of set-structured data. In CVPR, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.388, + 0.892, + 0.428 + ], + "angle": 0, + "content": "[25] Minchul Kim, Feng Liu, Anil Jain, and Xiaoming Liu. DCFace: Synthetic face generation with dual condition diffusion model. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.429, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[26] Minchul Kim, Yiyang Su, Feng Liu, Anil Jain, and Xiaoming Liu. Keypoint relative position encoding for face recognition. In CVPR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.468, + 0.892, + 0.508 + ], + "angle": 0, + "content": "[27] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. In ECCV, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.509, + 0.892, + 0.549 + ], + "angle": 0, + "content": "[28] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.549, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[29] Manoj Kumar, Mostafa Dehghani, and Neil Houlsby. Dual patchnorm. arXiv preprint arXiv:2302.01327, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.614 + ], + "angle": 0, + "content": "[30] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. Pointvoxel CNN for efficient 3d deep learning. In NeurIPS, 2019. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.615, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[31] Yunfei Long, Daniel Morris, Xiaoming Liu, Marcos Paul Gerardo Castro, Punarjay Chakravarty, and Praveen Narayanan. Radar-camera pixel depth association for depth completion. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.669, + 0.892, + 0.696 + ], + "angle": 0, + "content": "[32] David Lopez-Paz and Maxime Oquab. Revisiting classifier two-sample tests. arXiv preprint arXiv:1610.06545, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.696, + 0.892, + 0.75 + ], + "angle": 0, + "content": "[33] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.75, + 0.892, + 0.776 + ], + "angle": 0, + "content": "[34] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[35] Peiyuan Ni, Wenguang Zhang, Xiaoxiao Zhu, and Qixin Cao. Pointnet++ grasping: Learning an end-to-end spatial grasp generation algorithm from sparse point clouds. In ICRA, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.83, + 0.892, + 0.884 + ], + "angle": 0, + "content": "[36] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.884, + 0.892, + 0.898 + ], + "angle": 0, + "content": "[37] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh," + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9470" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.147 + ], + "angle": 0, + "content": "Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.469, + 0.187 + ], + "angle": 0, + "content": "[38] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.188, + 0.469, + 0.227 + ], + "angle": 0, + "content": "[39] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.228, + 0.468, + 0.254 + ], + "angle": 0, + "content": "[40] Prajit Ramachandran, Barret Zoph, and Quoc V. Le. Searching for activation functions. In ICLRW, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.255, + 0.468, + 0.295 + ], + "angle": 0, + "content": "[41] Zhiyuan Ren, Zhihong Pan, Xin Zhou, and Le Kang. Diffusion motion: Generate text-guided 3d human motion by diffusion model. arXiv preprint arXiv:2210.12315, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.296, + 0.468, + 0.334 + ], + "angle": 0, + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.469, + 0.388 + ], + "angle": 0, + "content": "[43] Nikolay Savinov, Junyoung Chung, Mikolaj Binkowski, Erich Elsen, and Aaron van den Oord. Step-unrolled denoising autoencoders for text generation. arXiv preprint arXiv:2112.06749, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.389, + 0.469, + 0.429 + ], + "angle": 0, + "content": "[44] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3D object proposal generation and detection from point cloud. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.43, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.457, + 0.469, + 0.494 + ], + "angle": 0, + "content": "[46] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.496, + 0.469, + 0.535 + ], + "angle": 0, + "content": "[47] Haoru Tan, Sitong Wu, and Jimin Pi. Semantic diffusion network for semantic segmentation. arXiv preprint arXiv:2302.02057, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.537, + 0.469, + 0.589 + ], + "angle": 0, + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[49] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.469, + 0.655 + ], + "angle": 0, + "content": "[50] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.469, + 0.71 + ], + "angle": 0, + "content": "[51] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In ICCV, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.711, + 0.469, + 0.75 + ], + "angle": 0, + "content": "[52] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-net: Point cloud upsampling network. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.751, + 0.469, + 0.804 + ], + "angle": 0, + "content": "[53] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. LION: Latent point diffusion models for 3d shape generation. In NeurIPS, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.806, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[54] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[55] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In ICCV, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.872 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9471" + } + ] +] \ No newline at end of file diff --git a/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/8fe5d2c1-4f78-435e-8bde-f7a4afed8604_origin.pdf b/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/8fe5d2c1-4f78-435e-8bde-f7a4afed8604_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c817b0bc8e08a52d655c6dab6943b3ed11f477a2 --- /dev/null +++ b/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/8fe5d2c1-4f78-435e-8bde-f7a4afed8604_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5ee4a56fb5e4570faf923028fd7b8d35b87c88c8bc93bdddafec204b7199586 +size 7086245 diff --git a/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/full.md b/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eb40644b193d6b053d95a18eb2946da39147bf96 --- /dev/null +++ b/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/full.md @@ -0,0 +1,367 @@ +# TIGER: Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process + +Zhiyuan Ren, Minchul Kim, Feng Liu, Xiaoming Liu +Michigan State University + +{renzhiy1, kimminc2, liufeng6, liuxm}@msu.edu + +# Abstract + +Recently, diffusion models have emerged as a new powerful generative method for 3D point cloud generation tasks. However, few works study the effect of the architecture of the diffusion model in the 3D point cloud, resorting to the typical UNet model developed for 2D images. Inspired by the wide adoption of Transformers, we study the complementary role of convolution (from UNet) and attention (from Transformers). We discover that their respective importance change according to the timestep in the diffusion process. At early stage, attention has an outsized influence because Transformers are found to generate the overall shape more quickly, and at later stages when adding fine detail, convolution starts having a larger impact on the generated point cloud's local surface quality. In light of this observation, we propose a time-varying two-stream denoising model combined with convolution layers and transformer blocks. We generate an estimizable mask from each timestep to reweigh global and local features, obtaining time-varying fused features. Experimentally, we demonstrate that our proposed method quantitatively outperforms other state-of-the-art methods regarding visual quality and diversity. Code is available https://github.com/Zhiyuan-R/Tiger-Diffusion. + +# 1. Introduction + +The point cloud is an essential 3D shape representation that is easy to obtain (sensed directly from laser scanning), versatile (serving as a building block of mesh), and easy to manipulate (geometrically transformed in a straightforward manner by affine matrices). 3D point cloud generative models benefit extensive applications across robotics [35], medicine [15] and content creation [36], and lay the groundwork for other vision tasks like point cloud upsampling [52], point cloud completion [21, 31] and mesh generation [9]. However, compared to 2D images, the cost and complexity of acquiring 3D point clouds make it crucial to ex + +![](images/20777058ec642375a5e5a540ccee9faeee2ff02eec539403aa98250cb074a8ea.jpg) +Figure 1. An illustration of different roles played by convolution and attention operations in the denoising model. Convolution is good for learning local relationships, and attention is optimal for modeling global relationships. We propose to merge these two properties across different timesteps in the diffusion process. + +plore and develop efficient and effective model architectures for 3D point cloud generation. + +Existing point cloud generative models are built on a range of frameworks, including generative adversarial networks (GANs) [1, 5], variational autoencoders (VAEs) [24], normalizing flows [23, 27, 51]. However, these methods neither are stable in training nor produce high-fidelity results. Recently, the denoising diffusion model [18], a novel generative method, has demonstrated exceptional results in 3D point cloud generation. Typically, diffusion models perturb a point cloud using a forward process and denoising algorithm. Once trained, the model can be used to generate new point clouds by iterating the reverse process over a sequence of time steps, with each step adding more noise to the input point cloud. + +Several recent works, including DPM [34] and PVD [55], have successfully applied diffusion models to 3D point cloud generation, achieving high levels of naturalness and diversity. Very recently, LION [53] further improves by encoding point clouds into latent space. However, these methods commonly utilize UNet-like convolutional networks that are originally designed for image processing. There has been limited research on developing suitable de + +noising model architectures that are specifically tailored for 3D point cloud generation. For example, PVD and LION both utilize Point-Voxel CNN (PVCNN) [30] as their denoising model. However, we observed that these PVCNN-based denoising models require a considerable number of timesteps to establish a rough shape since the limited receptive field cannot capture the global distribution of noise. + +In contrast to convolution operations, we observe that the attention mechanism used in Transformers [10, 48] is more effective at capturing long-range dependencies and thus the overall shape of 3D points. This difference is illustrated in the first two rows of Fig. 1. Compared to the convolution-based diffusion model (first row), the attention-based diffusion model (second row) correctly conforms to the overall object shape at earlier denoising time steps (middle column). However, the convolution operation excels in modeling local relationships. Therefore, the details of the final output are better modeled with convolutions (last column) which is also discussed in the ablation analysis of Sec. 4.3. The finding suggests that the two operations, the convolution and the attention are complementary in their roles at different denoising timesteps (the last row of Fig. 1) and we ask the following question. + +How to leverage the strengths of both CNNs and Transformers and enable them to dynamically capture complementary information in denoising diffusion for 3D point cloud generation? + +To answer this question, we propose a Time-varying denoising model for 3D point cloud generation (TIGER), a two-stream architecture combining a shallow CNN branch and Transformer branch with a timestep-dependent mask that adaptively reassigns weights to the global Transformer feature and the local CNN feature. To enable this feature-level fusion, we design an encoder-decoder to downsample the point cloud for reducing the number of points to a manageable size for Transformer. And we learn a selection mask that finds the optimal configuration at each timestep. + +Furthermore, in this two-stream architecture where global and local information is modeled separately, we discover that supplying 3D points position information to the global branch is critical to the performance. To supply the 3D position information more effectively, we design two novel position encoding methods: phase-shifted position encoding and Base- $\lambda$ position encoding. They create 3D position information in a continuous space and retain the property of linear expression of relative position. The effect of the time-dependent dual branch mask and the proposed position embedding is ablated in Sec. 4.3. + +A unique property of TIGER is that the weights of the two branches are dependent on the diffusion timesteps, and they are learned and optimized for the 3D point cloud gen + +eration. Interestingly, experiments demonstrate that the weight of the convolution branch is monotonically increasing with the timestep, which is in accordance with our initial observation that convolutions are good for modeling fine details. Empirically we show that TIGER achieves state-of-the-art (SoTA) performance in 3D point cloud generation. + +Our main contributions include: + +- We propose a novel two-stream denoising model, which uses timestep to optimally reweigh the global feature from Transformer and the local feature from shallow CNN. +- We design two novel 3D continuous position encoding methods and position-aware self-attention to enable our Transformer branch to aggregate global features effectively from the latent point cloud input. +- Experiments show that our method outperforms prior works in 3D point cloud generation tasks qualitatively and quantitatively on the ShapeNetv2 dataset. + +# 2. Related Work + +3D Point Cloud Generation. The advances in 3D point cloud generation can be categorized by how the input is represented. Early methods [1, 12, 14, 38, 39, 44] represent the point cloud as a $N \times 3$ matrix, where $N$ is the predefined number of 3D points. The backbone comprises 1D convolutions and max-pooling to allow the points' permutation invariance. So, the shape is represented by a global feature vector. + +More recent works [5, 7, 13, 27, 34, 51] have improved upon the input representation by introducing a probabilistic distribution assumption, where each point cloud is considered as one data point sampled from the distribution. For instance, PointFlow [51] models the distribution of points by continuous normalizing flows, allowing the sampling of an arbitrary number of points to represent a point cloud. However, the backbone architecture is still based on 1D convolution-based encoder with the max-pooling operation [38]. + +Recently, PVD [55] and LION [53] adopt a hybrid point-voxel representation [30] which voxelize the point clouds for 3D convolution. This representation with rich spatial information combines well with the diffusion objective and produces promising results [53]. However, the backbone is still confined to convolution-based networks, lacking the capacity to effectively model the long range-dependencies. We, on the other hand, find the complementary roles of attention and convolution in the point cloud generator and propose a network that combines the benefits of both. + +Diffusion Models. Diffusion [18] and score-based methods [22, 46] have enjoyed remarkable success in images synthesis. Diffusion has been extended to many generation + +![](images/0030f138726516636a8cd7a1379cda31749f6cdb1ba7a50177aa783fc8154afb.jpg) +Figure 2. Illustration of our time-varying two-stream architecture (TIGER). The network's input is a noisy point cloud $\mathbf{X}_t$ at timestep $t$ , and the goal is to predict the noise in $\mathbf{X}_t$ to obtain the denoised point cloud $\mathbf{X}_{t-1}$ . We first encode $\mathbf{X}_t$ into latent point cloud $\hat{\mathbf{X}}_t$ and export coordinates information for position encoding and decoder. We feed the latent point cloud into both the Latent Point Cloud Transformer and a shallow CNN branch to get global feature $\mathbf{X}_{tr}$ and local feature $\mathbf{X}_c$ . Then, we combine two features using timestep dependent mask $\mathbf{M}_c$ to create a fused feature $\hat{\mathbf{X}}_F$ . Last, we apply our latent point cloud decoder to predict the noise from $\hat{\mathbf{X}}_F$ . + +tasks [2, 8, 17, 20, 25, 28, 41-43, 47], and largely improved by innovative techniques, such as fast sampling [33, 45], classifier-free guidance [19, 37] and latent diffusion [4, 42]. Recently 3D point cloud generation with diffusion [34, 53, 55] is also proposed with remarkable diversity and fidelity. + +However, previous works do not investigate the behavior of the diffusion backbone with respect to the diffusion timestep. We find that the network equipped with both convolution and attention operations learns to utilize the local and global shape representation with different magnitudes based on the timestep. This time adaptive architecture allows the 3D point cloud generator to be trained faster than LION [53] while producing higher-quality outputs. + +Transformer for Point Clouds. Transformer architecture [48] is intrinsically a set operator, which is suitable for permutation-invariant point clouds. PCT [16] applies global attention to the point cloud. Point Transformer V1 [54] and V2 [49] convert global attention to local attention to reduce memory consumption and computational complexity. + +Contrary to the previous methods, our method decouples the global and local information by using Transformer to aggregate global information effectively while using CNN to model the local information. Also, we find that position embedding is crucial to performance and propose a new position embedding to further improve the performance. + +# 3. Method + +In this section, we first briefly formulate the probabilistic model of forward and reverse diffusion processes for 3D point cloud generation. Then, we dive into the details of our time-varying two-stream architecture, including the encoder part, latent point Transformer, time mask generator, and decoder part. The entire architecture is shown in Fig. 2. + +# 3.1. Problem Formulation + +Given a point cloud $\mathbf{X}_0\in \mathbb{R}^{N\times 3}$ , where $N$ is the number of points, the diffusion forward process gradually diffuses the original point cloud into Gaussian noise $\mathbf{X}_T$ ; the forward process can be formulated with a transition probability $q(\mathbf{X}_t|\mathbf{X}_{t - 1})$ . By giving a sequence of pre-defined noising scales $\beta_{1},\beta_{2},\ldots ,\beta_{T}$ , the transition probability can be expressed as: + +$$ +q \left(\mathbf {X} _ {t} \mid \mathbf {X} _ {t - 1}\right) = \mathcal {N} \left(\sqrt {1 - \beta_ {t}} \mathbf {X} _ {t - 1}, \beta_ {t} \mathbf {I}\right). \tag {1} +$$ + +For the reverse process, we use learn the $p_{\theta}(\mathbf{X}_{t - 1}|\mathbf{X}_t)$ a Gaussian distribution which approximates the intractable real distribution $q(\mathbf{X}_{t - 1}|\mathbf{X}_t)$ . Specifically, + +$$ +p _ {\theta} = \mathcal {N} \left(\mu_ {\theta} \left(\mathbf {X} _ {t}, t\right), \sigma_ {t} ^ {2} \mathbf {I}\right), \tag {2} +$$ + +where we learn the mean $\mu_{\theta}(\mathbf{X}_t,t)$ and set the variance $\sigma_t^2$ to a fixed schedule as in [18]. Conventionally, $\mu_{\theta}(\mathbf{X}_t,t)$ is modeled using a UNet backbone and it depends on the timestep $t$ to scale and shift the feature. We design a two-stream backbone where the timestep modulates the weight between the two streams. To sample a 3D point cloud, we iteratively denoise $\mathbf{X}_T,T$ times to finally generate the highfidelity samples. In training, we optimize the MSE loss: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {s i m p l e}} = \mathbb {E} _ {t \sim [ 1, T ]} \| \mu - \mu_ {\theta} (\mathbf {X} _ {t}, t) \| _ {2} ^ {2} \tag {3} \\ = \mathbb {E} _ {t \sim [ 1, T ]} \| \epsilon - \epsilon_ {\theta} (\mathbf {X} _ {t}, t) \| _ {2} ^ {2}, \\ \end{array} +$$ + +where $\epsilon$ is the ground truth noise and $\epsilon_{\theta}$ is the noise prediction by our diffusion model. + +# 3.2. Noisy Point Cloud Encoder + +To effectively represent the $N\times 3$ points, we adopt the voxelization encoding scheme of PVCNN [30]. This representation works well with both CNN and Transformer architecture as we can downsample the voxel-grid following [11] to + +![](images/5d0d12675c8a255d5c173a700feee286dce8d645432f1cc4d6a57748aa521066.jpg) +Figure 3. The overview of our encoder. It extracts features in voxel space and downsamples the point cloud by applying trilinear interpolation of voxel with grouped points. + +increase computation efficiency. Specifically, a noisy point cloud encoder $\mathcal{E}:\mathbb{R}^{N\times 3}\to \mathbb{R}^{M\times d}$ transforms the noisy point cloud $\mathbf{X}_t\in \mathbb{R}^{N\times 3}$ at timestep $t$ into a latent point cloud representation $\hat{\mathbf{X}}_t\in \mathbb{R}^{M\times d}$ , where $N$ is the number of original points, $M$ is the number of downsampled points and $d$ is the dimension of each latent point's feature. The overview of our encoder is illustrated in Fig. 3. + +To aggregate features in a rigid field, we first voxelize the noisy point cloud $\{x_{n},y_{n},z_{n}\} = \mathbf{X}_{t}\in \mathbb{R}^{N\times 3}$ into the voxel grids $\{\mathbf{V}_{u,v,w}\} = \mathbf{V}\in \mathbb{R}^{L\times L\times L}$ , where $L$ is the resolution and $n$ is the point index in $N$ . Specifically, + +$$ +\mathbf {V} _ {u, v, w} = \sum_ {n = 1} ^ {N} \mathbb {I} \left(x _ {n} \in \mathbb {N} (u, \delta), y _ {n} \in \mathbb {N} (v, \delta), z _ {n} \in \mathbb {N} (w, \delta)\right) * \mathbf {f} _ {n}, \tag {4} +$$ + +where $\mathbb{N}(z,\delta) = \{x|x\in [z - \delta ,z + \delta)\}$ , $\mathbf{f}_n$ denotes the $n^{th}$ extracted latent feature corresponding to the $n^{th}$ point $(\mathbf{X}_t)_n$ and $\mathbb{I}(\cdot)$ is an indicator of whether the $n^{th}$ point falls into the voxel grid $(u,v,w)$ . After voxelization, we apply a series of 3D convolution, Swish [40] activation and Group-Norm [50]. After applying a series of these operations, this results in a latent volume $\hat{\mathbf{V}}\in \mathbb{R}^{L\times L\times L\times d}$ , where $d$ is the channel size for each voxel. + +We use furthest point sampling algorithm [11] to downsample the input noisy point cloud $\mathbf{X}_t\in \mathbb{R}^{N\times 3}$ into a sparser point cloud $\mathbf{X}_t^s\in \mathbb{R}^{M\times 3}$ ( $M < N$ ). Then, as the position in volume is discrete, we query the latent volume $\hat{\mathbf{V}}$ by the sparse point cloud $\mathbf{X}_t^s$ via trilinear interpolation, eventually transformed into latent point cloud $\hat{\mathbf{X}}_t\in \mathbb{R}^{M\times d}$ . It is worth noting we preserve the coordinates of $\mathbf{X}_t$ and $\mathbf{X}_t^s$ to do upsampling with the decoder and provide position embedding for our latent point cloud Transformer. + +# 3.3. Latent Point Cloud Transformer + +Previous denoising models such as PVCNN [30] simply apply stacked convolutional layers to create feature representations. Contrarily, we aim to better model the global relationship by introducing a novel Transformer branch named latent point cloud Transformer. + +Tokenization. After encoding, we get a latent point cloud representation $\hat{\mathbf{X}}_t\in \mathbb{R}^{M\times d}$ and its coordinates $\mathbf{X}_t^s\in \mathbb{R}^{M\times 3}$ . + +![](images/2348bc51c4bc105e548c4f3b0f78a9e8932dc7feaf9fea47027652e82f247cec.jpg) +(a) PSPE + +![](images/c1db3196b9deaa2816c4d79e4ad6b7cbee8d8f1fbf38de23e567196690137c13.jpg) +(b) BλPE +Figure 4. Illustration of two examples of the position embedding from PSPE and BλPE respectively. Both methods show distinguished representation for each position. + +Following [29], we use dual PatchNorm to project the latent point cloud into tokens, which place LayerNorm before and after an MLP layer for more stable training: + +$$ +\mathbf {T} = \operatorname {L N} \left(\operatorname {M L P} \left(\operatorname {L N} \left(\hat {\mathbf {X}} _ {\mathbf {t}}\right)\right)\right), \tag {5} +$$ + +where $\mathbf{T} \in \mathbb{R}^{M \times D}$ and $D$ is the dimension in the token space (in our paper, $D = 2d$ ). + +Position Encoding. Since the order of the latent points is arbitrary in a point cloud, it's necessary to create a 3D position embedding and fuse it into the feature, so that the model knows the position of each point. Contrary to the position embedding in image domains [10, 26], the design of 3D position embedding plays a vital role in the performance and our position embedding produces the best result. We propose two novel 3D space continuous position encoding methods: Phase Shift Position Encoding (PSPE) and Base- $\lambda$ Position Encoding (B $\lambda$ PE). For PSPE, we use Sine and Cosine functions of different frequencies and phases: + +$$ +\mathbf {P} _ {e m b} \left(p o s _ {j}, 6 i + 2 * (j - 1)\right) = \sin \left(\frac {p o s _ {j}}{1 0 0 0 0 ^ {\frac {2 i}{D}}} + (j - 1) \frac {2 \pi}{3}\right) \tag {6} +$$ + +$$ +\mathbf {P} _ {e m b} (p o s _ {j}, 6 i + 1 + 2 * (j - 1)) = \cos \left(\frac {p o s _ {j}}{1 0 0 0 0 ^ {\frac {2 i}{D}}} + (j - 1) \frac {2 \pi}{3}\right), \tag {7} +$$ + +where $pos_{j}$ is $j$ -th element in each point's position, specifically $pos_{1} = x$ , $pos_{2} = y$ , $pos_{3} = z$ , and $i$ is the index of the dimension. We maximize the phase shift difference $(\frac{4\pi}{3})$ of different axes to distinguish them and combine Sine and Cosine to guarantee the linear representation of relative position, which will be proved mathematically in Supp. An example is shown in Fig. 4a and $\mathbf{P}_{emb}$ is shaped $\mathbb{R}^{M\times D}$ . + +However, PSPE requires 6 channels to represent one frequency feature, which is quite inefficient. To improve channel utilization, we apply BλPE to compress 3D position into one scalar: + +$$ +p o s = \lambda^ {2} * z + \lambda * y + x \tag {8} +$$ + +$$ +\mathbf {P} _ {e m b} (p o s, 2 i) = \sin \left(\frac {p o s}{1 0 0 0 0 ^ {\frac {2 i}{D}}}\right) \tag {9} +$$ + +$$ +\mathbf {P} _ {e m b} (p o s, 2 i + 1) = \cos \left(\frac {p o s}{1 0 0 0 0 ^ {\frac {2 i}{D}}}\right), \tag {10} +$$ + +![](images/49b1d9a93ac5046301acf56d70da9c791842c0f7556fa73c8c610a5cf832f53b.jpg) +Figure 5. The overview of the position-aware self-attention. We compute position relationship map $\mathbf{H}$ by previous position embedding to reweigh the $\mathbf{Q}\mathbf{K}^T$ matrix. + +where $pos$ is a polynomial expression of $\lambda$ . Specifically, we set $\lambda$ to 1000, which means this preserves three decimal places of precision and one example is shown in Fig. 4b. The choice between these two position encoding methods is essentially a trade-off between granularity and expressiveness. We ablate them and other non-position encoding and learnable position encoding in Sec. 4.3. + +Transformer and Position-Aware Attention We stack $L$ layers of Transformer block [48] which comprises of LayerNorm(LN) [3] and Multihead Self Attention (MHA) and MultiLayer Perceptrons. We let the token $\mathbf{T}_0 = \mathbf{T}$ and let + +$$ +\mathbf {T} _ {l *} = \operatorname {M H A} \left(\ln \left(\mathbf {T} _ {l - 1}\right)\right) + \mathbf {T} _ {l - 1}, \tag {11} +$$ + +$$ +\mathbf {T} _ {l} = \operatorname {M L P} \left(\operatorname {L N} \left(\mathbf {T} _ {l *}\right)\right) + \mathbf {T} _ {l *}, \tag {12} +$$ + +$$ +\mathbf {X} _ {t r} := \mathbf {T} _ {L}. \tag {13} +$$ + +To add position embedding into the model, one can simply add it to $\mathbf{T}$ . But as the Transformer blocks are gradually stacked, the continuous 3D position encoding is difficult to retain in the token representation. Inspired by [54], we explicitly add the position encoding to each MHA in order to make our attention map position-aware, as shown in Fig. 5. Specifically, our position-aware MHA modified as + +$$ +\mathbf {H} = \operatorname {S o f t m a x} \left(\left(\mathbf {P} _ {e m b} \mathbf {W} _ {p}\right) \left(\mathbf {P} _ {e m b} \mathbf {W} _ {p}\right) ^ {T}\right) \tag {14} +$$ + +$$ +\mathbf {T} _ {l *} = \operatorname {S o f t m a x} \left(\frac {\left(\mathbf {T} _ {l} \mathbf {W} _ {q}\right) \left(\mathbf {T} _ {l} \mathbf {W} _ {k}\right) ^ {T} \odot \mathbf {H}}{\sqrt {D}}\right) \left(\mathbf {T} _ {l} \mathbf {W} _ {v}\right) \tag {15} +$$ + +where $\mathbf{W}_q,\mathbf{W}_k,\mathbf{W}_v,\mathbf{W}_p\in \mathbb{R}^{D\times Z}$ are projection matrices and $\mathbf{H}\in \mathbb{R}^{M\times M}$ is the position relationship map. Modified MHA could incorporate position information explicitly and sufficiently during training. $\mathbf{X}_{tr}$ is the final output. + +![](images/ad7b3332409ec3d6e4b12c4f9b36848faeea9c190734e000f0bcb3e3d116d1ca.jpg) +Figure 6. The overview of our decoder. By querying the latent volume $\hat{\mathbf{X}}_F$ with previous coordinates information $\mathbf{X}_t$ , we can upsample the latent point cloud into 3D space. + +# 3.4. Time Mask Generator + +We apply a depth-2 PVCNN [30] to get local feature $\mathbf{X}_c\in$ $\mathbb{R}^{M\times D}$ and our latent point cloud Transformer to get global feature $\mathbf{X}_{tr}\in \mathbb{R}^{M\times D}$ . Time mask generator module scales the local and global features by the time variable. We force the sum of two learnable time masks to be 1. With this constraint, the trade-off of local and global features can be learned by the network. + +We first align the dimension and scale of $\mathbf{X}_c$ and $\mathbf{X}_{tr}$ by one layer MLP and Sigmoid function $\sigma$ . + +$$ +\mathbf {X} _ {c} ^ {*} = \sigma (\operatorname {M L P} (\mathbf {X} _ {c})), \quad \mathbf {X} _ {t r} ^ {*} = \sigma (\operatorname {M L P} (\mathbf {X} _ {t r})). \tag {16} +$$ + +Then we encode a timestep $t$ to a sinusoidal position embedding [18, 48] to $\mathbf{t}_{emb} \in \mathbb{R}^c$ . We use two stacked MLP to map $\mathbf{t}_{emb}$ to a higher dimension space and apply Sigmoid to keep one of the masks between 0 to 1. Specifically, + +$$ +\mathbf {M} _ {c} = \sigma (\mathrm {M L P} (\text {L e a k y R e L U} (\mathrm {M L P} (\mathbf {t} _ {\text {e m b}}))), \tag {17} +$$ + +$$ +\mathbf {M} _ {t r} = \mathbf {1} - \mathbf {M} _ {c}. \tag {18} +$$ + +where $\mathbf{M}_c$ , $\mathbf{M}_{tr} \in \mathbb{R}^D$ are the time masks for CNN and Transformer feature respectively. We ablate the dimension of masks in Sec. 4.3 to be $\mathbb{R}^D$ or $\mathbb{R}^1$ . We multiply these two masks by the respective features and adjust the fused feature by MLP. + +$$ +\hat {\mathbf {X}} _ {F} = \operatorname {M L P} \left(\mathbf {X} _ {c} ^ {*} \odot \mathbf {M} _ {c} + \mathbf {X} _ {t r} ^ {*} \odot \mathbf {M} _ {c}\right), \tag {19} +$$ + +where $\hat{\mathbf{X}}_F\in \mathbb{R}^{M\times D}$ is the final feature adaptively aggregating local and global features based on the timestep. + +# 3.5. Latent Point Cloud Decoder + +In order to upsample the latent point cloud and predict the corresponding noise, we design a latent point cloud decoder $\mathcal{D}$ , which converts the latent point cloud $\hat{\mathbf{X}}_F \in \mathbb{R}^{M \times D}$ to $\mathbf{X}_t \in \mathbb{R}^{N \times 3}$ . Similar to Sec. 3.2, we can voxelize $\hat{\mathbf{X}}_F$ into a latent volume $\tilde{\mathbf{V}}_F \in \mathbb{R}^{L \times L \times L \times D}$ , and take $\mathbf{X}_t$ as the query for trilinear interpolation, eventually resulting in the final noise prediction $\epsilon_{\theta} \in \mathbb{R}^{N \times 3}$ . This formulation can be expressed as: + +$$ +\epsilon_ {\theta} = \operatorname {T r i l i n e a r} \left(\operatorname {V o x e l i z e} \left(\hat {\mathbf {X}} _ {F}\right), \mathbf {X} _ {t}\right) \mathbf {W} ^ {D \times 3}, \tag {20} +$$ + +
MethodGenerative ModelAirplaneChairCar
CD→50%EMD→50%CD→50%EMD→50%CD→50%EMD→50%
r-GAN [1]GAN98.4096.7983.6999.7094.4699.01
l-GAN(CD) [1]GAN87.3093.9568.5883.8466.4988.78
l-GAN(EMD) [1]GAN89.4976.9171.9064.6571.1666.19
PointFlow [51]Normalizing Flow75.6870.7462.8460.5758.1056.52
DPF-Net [27]Normalizing Flow75.1865.5562.0058.5362.3554.48
ShapeGF [5]GAN80.0076.1768.9665.4863.2056.53
SoftFlow [23]Normalizing Flow76.0565.8059.2160.0564.7760.09
SetVAE [24]VAE76.5467.6558.8460.5759.9559.94
DPM [34]Diffusion76.4286.9160.0574.7768.8979.97
PVD [55]Diffusion73.8264.8156.2653.3254.5553.83
TIGERDiffusion71.8555.8254.6152.7154.3152.24
LION [53]Diffusion67.4161.2353.7052.3453.4151.14
TIGERDiffusion67.2156.2654.3251.7154.1250.24
+ +Table 1. Quantitative comparison with baselines using 1-NN. Both CD and EMD are considered, where CD is multiplied by ${10}^{3}$ and EMD is multiplied by ${10}^{2}$ . The closer the score is to 50%,the better the quality and diversity of generated samples. [Key: Best, Second Best] + +![](images/18846db4820861bec226d324c3415b846fd7c495f424a059089b22abe3ad0c49.jpg) +Figure 7. Our generation results (right) compared to baseline models (left). TIGER generates high-quality and diverse 3D point clouds. + +where $\mathbf{W}^{D\times 3}$ is the projection matrix to map the noise into 3D space and $\mathbf{X}_t$ is the original point cloud coordinates. The overview of our decoder is shown in Fig. 6. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We follow previous works and choose ShapeNetv2 [6] as our main dataset, which is pre-processed by PointFlow [51]. To fairly compare with other different baseline methods, we train and evaluate on three categories: airplane, chair and car. For each shape, we sample 2,048 points and normalize them globally across the entire dataset. The training set consists of 2,832,4,612, and 2,458 shapes, while the evaluation set is composed of 405,662, and 352 shapes for airplanes, chairs, and cars, respectively. + +Metrics. Following the baselines PVD [55] and LION [53], we use 1-NN (1-nearest neighbor) accuracy [32] as our evaluation metric. This metric has been shown to effectively + +measure both the quality and diversity of generated point clouds and a score closer to $50\%$ indicates superior performance [51]. To calculate the 1-NN distance matrix, we consider two commonly used metrics for measuring the distance between point clouds: CD (Chamfer Distance) and EMD (Earth Movers' Distance). + +# 4.2. Comparison with SoTA methods + +Results. To compare our time-varying denoising model with SoTA methods, we evaluate its performance against competitive models such as SetVAE [24], DPM [34], PVD [55] and LION [53]. It is noteworthy that in order to compare with LION, which uses a different dataset splitting strategy (sampling from the first 10,000 points instead of the latter 5,000 points), we evaluate our model under both LION's strategy and the standard strategy used by other methods. As shown in Tab. 1, we outperform LION in four out of six metrics. Compared to other methods, our performance is significantly better. Notably, our EMD-based + +
MethodTraining Time (GPU hours)Inference Time (s)
PVD [55]1428.46
LION [53]55027.12
Tiger1649.73
+ +Table 2. Training and inference time comparison. Training time is counted on average for three categories: airplane, chair and car. + +
MethodW1W2Mean CDMean EMD
Convolution1061.3456.31
Transformer0162.4155.85
Conv+Tran0.50.564.4557.92
Ourslearnable (scalar)60.7954.12
Ourslearnable (channel-wise)59.4353.51
+ +1-NN accuracy shows a significant improvement of $13.45\%$ relative to PVD and $8.12\%$ relative to LION for airplanes. This is attributed to the efficient aggregation of global features by our Transformer branch, which is especially beneficial for EMD as it is primarily influenced by global distribution. Fig. 7 shows that our generated samples are of high quality and diversity compared with other methods. + +Efficiency. To compare the training and inference speed, we report the GPU hours with NVidia V100 GPUs as in LION [53]. As shown in Tab. 2, our model trains much faster than LION, with only a quarter of its training time and a third of its inference time. Our training and inference speed is comparable to that of PVD while surpassing LION and PVD in the generation quality as shown in Tab. 1. More computation analysis can be found in Supp. + +# 4.3. Ablation and Analysis + +Time-varying two-stream architecture. We ablate our time-varying two-stream architecture by comparing it with single ConvNet, single Transformer, and two-stream networks with equal weight. In this ablation, we also compare the performance of our time masking with scalar value setting and channel-wise value setting. The results, presented in the first two rows of Tab. 3, show that Transformer performs better on the EMD metric, while ConvNet performs better on the CD metric. Using a fixed equal weight for the two-stream network leads to a significant performance drop on both metrics. However, the learnable time mask improves the performance on both metrics, indicating the importance of the time-varying design. Further, channel-wise time masking improves performance slightly more than scale value setting due to its greater expressivity. Transformer backbones, position encoding, and self-attention strategies. Here we carried out three ablations: (1) We compare the performance of our Transformer backbone with two classic Transformer backbones, PCT [16] and Point Transformer (PointTran) [54]. (2) To assess + +Table 3. Ablation of different architecture designs. $W_{1}$ and $W_{2}$ are the weights for the ConvNet branch and Transformer branch. [Key: Best, Second Best] + +
TBPEAPMean CDMean EMD
PCT [16]Learnable-65.8260.21
PointTran [54]LearnableLearnable66.4262.39
Ours--63.5157.98
OursLearnableLearnable62.1756.13
OursPSPE-60.1155.74
OursPSPELearnable61.4855.97
OursPSPEPSPE59.0254.49
OursBλPE-61.3554.27
OursBλPELearnable61.8956.02
OursBλPEBλPE59.4353.51
+ +Table 4. Ablation of Transformer backbones, position encoding, and self-attention strategies. [Key: Best, Second Best; TB=Transformer backbone; PE=position encoding; AP=additional position information] + +![](images/a7487b1c2e2fa1ec6dd7186593fd646c1f0b70302f4d35ce36c14cbcbedd6530.jpg) +Figure 8. Visualization of attention maps in 3D point clouds. Column (c) is the result of column (b) multiplied by position relationship map H. Taking the first row as an example, queried by the airplane head's position, our method can aggregate the position information from the airplane's right wing and left elevator compared with standard self-attention. This comparison highlights the effectiveness of our position-aware self-attention module in capturing long-range relationships in 3D point clouds. [SA: Self-attention, PASA: Position-aware Self-attention] + +![](images/23fe7c856fb0923408d7a66d7d8252ac8419dd825c62a27118ab97168b3fa04a.jpg) + +our proposed position encoding methods, PSPE and BλPE, we compare them with no position encoding and learnable position encoding. (3) We investigate our position-aware self-attention module by comparing it with standard self-attention and self-attention with learnable position encoding as proposed in [54]. As shown in Tab. 4, our latent point cloud Transformer outperforms both PCT and Point Transformer under various settings. Furthermore, our proposed position encoding methods, PSPE and BλPE, significantly improve performance compared to no position encoding or learnable position encoding. Specifically, PSPE improves the CD metric, while BλPE improves the EMD metric. Our position-aware self-attention module, which aggregates corresponding position encoding, achieves better results than standard self-attention and self-attention with learnable position encoding. To illustrate this, Fig. 8 visualizes the attention map in 3D point space for both standard self-attention and our position-aware self-attention. This visualization demonstrates the ability of our method to capture + +![](images/8009a5ef648b8eb0755451f9d170aed35e6f8460236d5baebd74ddf9fe163948.jpg) +(a) Airplane + +![](images/82263e7e086d08818f9f6e6e8d6709972c4a468e9cb82c856ff8a9dc4b62c7e0.jpg) +(b) Chair + +![](images/0b9120aa39eab04542649825872d9840f49d658450f75d08f536bfa3f5e388fb.jpg) +(c) Car + +![](images/8ad4e30cfcb2378ba7eac682e165aa3026c9cf196214f26bb2d995d07b2eb7db.jpg) +Figure 9. The changes of ConvNet importance for each category over timesteps with the channel-wise time mask setting. The thick curve is the result of Gaussian smoothing. +Figure 10. High quality and diverse 3D point clouds generated from our TIGER model trained on 55 ShapeNetv2 categories. + +long-range relationships in the data. + +Weight changing over timesteps. To validate our motivation of reweighing the local and global features from ConvNet and Transformer at different time steps, we visualize the changes in weights. For channel-wise time mask setting, we use the ratio of the weight of the local feature over the weight of the global feature to measure the importance of ConvNet and Transformer at each timestep. Fig. 9 indicates that the importance of ConvNet for each category increases nearly monotonically, implying that the diffusion process of 3D point cloud generation benefits from global features at the early stages of diffusion and requires the aggregation of more local features at later stages. Moreover, we observe that the curves of airplanes and chairs change more drastically than that of cars, which we hypothesize is due to the complicated shapes of airplanes and chairs, resulting in a more complex diffusion process. + +Unconditional generation of multiple classes. To further demonstrate the generalizability of our model, we train a universal TIGER model on all 55 categories of ShapeNetv2 [6]. Fig. 10 shows that our universal model can produce shapes with high quality and diversity. + +# 5. Conclusions + +In this work, we propose a novel time-varying denoising model with a latent point transformer backbone. Experimentally, we show how the time mask optimally aggre + +gates local and global features at different timesteps, which quantitatively reaches SoTA performance. In accordance with our hypothesis, Transformer is more important at early diffusing stages and ConvNet has a larger weight at later timesteps in the 3D point cloud generation task. Future work will be focused on exploring how our time-varying denoising model would behave in other diffusion-based tasks. + +Limitations. Although we generate high-quality and natural samples, we cannot control the category of the generated shape. One potential solution is to encode the category into latent space and train the diffusion model conditioned on the latent code. We also acknowledge that our model needs to get the local feature and global feature by two subnetworks, which is not quite efficient. In this work, we gain efficiency with the model converging with fewer epochs. But future works can increase the backbone efficiency by proposing time-varying properties with only one network. + +Potential Negative Impacts. While it has the potential to speed up progress in both areas, it also has the risk of eliminating the jobs of game and furniture designers due to automation. However, our model should have a mostly positive impact, because the most likely use of our method would be to assist designers not completely replace their jobs. It is still necessary to select suitable shapes among many generated samples. + +# References + +[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, 2018. 1, 2, 6 +[2] Vishal Asnani, John Collomosse, Tu Bui, Xiaoming Liu, and Shruti Agarwal. ProMark: Proactive diffusion watermarking for causal attribution. In CVPR, 2024. 3 +[3] Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450, 2016. 5 +[4] Fan Bao, Kun Xu, Chongxuan Li, Lanqing Hong, Jun Zhu, and Bo Zhang. Variational (gradient) estimate of the score function in energy-based latent variable models. In ICML, 2021. 3 +[5] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snavely, and Bharath Hariharan. Learning gradient fields for shape generation. In ECCV, 2020. 1, 2, 6 +[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 8 +[7] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. NeurIPS, 31, 2018. 2 +[8] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusiondet: Diffusion model for object detection. arXiv preprint arXiv:2211.09788, 2022. 3 +[9] Tamal K Dey and James Hudson. Pmr: Point to mesh rendering, a feature-based approach. In VIS, 2002. 1 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2, 4 +[11] Yuval Eldar, Michael Lindenbaum, Moshe Porat, and Yehoshua Y Zeevi. The farthest point strategy for progressive image sampling. In ICPR, 1994. 3, 4 +[12] Matheus Gadelha, Rui Wang, and Subhransu Maji. Multiresolution tree networks for 3d point cloud processing. In ECCV, 2018. 2 +[13] Will Grathwohl, Ricky TQ Chen, Jesse Bettencourt, Ilya Sutskever, and David Duvenaud. Ffjord: Free-form continuous dynamics for scalable reversible generative models. arXiv preprint arXiv:1810.01367, 2018. 2 +[14] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3D surface generation. In CVPR, 2018. 2 +[15] Jia Guo, Xuanxia Yao, Mengyu Shen, Jiafei Wang, and Wanyou Liao. A deep learning network for point cloud of medicine structure. In ITME, 2018. 1 +[16] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 3, 7 +[17] Kilian Konstantin Haefeli, Karolis Martinkus, Nathanael Perraudin, and Roger Wattenhofer. Diffusion models for + +graphs benefit from discrete state spaces. arXiv preprint arXiv:2210.01549, 2022.3 +[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2, 3, 5 +[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3 +[20] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 3 +[21] Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. In CVPR, 2020. 1 +[22] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 2005. 2 +[23] Hyeongju Kim, Hyeonseung Lee, Woo Hyun Kang, Joun Yeop Lee, and Nam Soo Kim. Softflow: Probabilistic framework for normalizing flow on manifolds. In NeurIPS, 2020. 1, 6 +[24] Jinwoo Kim, Jaehoon Yoo, Juho Lee, and Seunghoon Hong. Setvae: Learning hierarchical composition for generative modeling of set-structured data. In CVPR, 2021. 1, 6 +[25] Minchul Kim, Feng Liu, Anil Jain, and Xiaoming Liu. DCFace: Synthetic face generation with dual condition diffusion model. In CVPR, 2023. 3 +[26] Minchul Kim, Yiyang Su, Feng Liu, Anil Jain, and Xiaoming Liu. Keypoint relative position encoding for face recognition. In CVPR, 2024. 4 +[27] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. In ECCV, 2020. 1, 2, 6 +[28] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 3 +[29] Manoj Kumar, Mostafa Dehghani, and Neil Houlsby. Dual patchnorm. arXiv preprint arXiv:2302.01327, 2023. 4 +[30] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. Pointvoxel CNN for efficient 3d deep learning. In NeurIPS, 2019. 2, 3, 4, 5 +[31] Yunfei Long, Daniel Morris, Xiaoming Liu, Marcos Paul Gerardo Castro, Punarjay Chakravarty, and Praveen Narayanan. Radar-camera pixel depth association for depth completion. In CVPR, 2021. 1 +[32] David Lopez-Paz and Maxime Oquab. Revisiting classifier two-sample tests. arXiv preprint arXiv:1610.06545, 2016. 6 +[33] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927, 2022. 3 +[34] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, 2021. 1, 2, 3, 6 +[35] Peiyuan Ni, Wenguang Zhang, Xiaoxiao Zhu, and Qixin Cao. Pointnet++ grasping: Learning an end-to-end spatial grasp generation algorithm from sparse point clouds. In ICRA, 2020. 1 +[36] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 1 +[37] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, + +Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 3 +[38] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 2 +[39] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2 +[40] Prajit Ramachandran, Barret Zoph, and Quoc V. Le. Searching for activation functions. In ICLRW, 2018. 4 +[41] Zhiyuan Ren, Zhihong Pan, Xin Zhou, and Le Kang. Diffusion motion: Generate text-guided 3d human motion by diffusion model. arXiv preprint arXiv:2210.12315, 2022. 3 +[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3 +[43] Nikolay Savinov, Junyoung Chung, Mikolaj Binkowski, Erich Elsen, and Aaron van den Oord. Step-unrolled denoising autoencoders for text generation. arXiv preprint arXiv:2112.06749, 2021. 3 +[44] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3D object proposal generation and detection from point cloud. In CVPR, 2019. 2 +[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3 +[46] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 2019. 2 +[47] Haoru Tan, Sitong Wu, and Jimin Pi. Semantic diffusion network for semantic segmentation. arXiv preprint arXiv:2302.02057, 2023. 3 +[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2, 3, 5 +[49] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 3 +[50] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 4 +[51] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In ICCV, 2019. 1, 2, 6 +[52] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-net: Point cloud upsampling network. In CVPR, 2018. 1 +[53] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. LION: Latent point diffusion models for 3d shape generation. In NeurIPS, 2022. 1, 2, 3, 6, 7 +[54] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 3, 5, 7 +[55] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In ICCV, 2021. 1, 2, 3, 6, 7 \ No newline at end of file diff --git a/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/images.zip b/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..324b71815703a9d3b2a66c1da8d061718aa0a920 --- /dev/null +++ b/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e42dc7198d854b5fa28c69c7d64c0ba0fc31dbadbbcfff8dbc512e12a438d143 +size 693730 diff --git a/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/layout.json b/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4c50705d23bf25ba4cebab32f140b6be0d46311c --- /dev/null +++ b/2024/TIGER_ Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process/layout.json @@ -0,0 +1,9725 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 130, + 103, + 463, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 103, + 463, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 103, + 463, + 138 + ], + "type": "text", + "content": "TIGER: Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 167, + 161, + 426, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 161, + 426, + 190 + ], + "spans": [ + { + "bbox": [ + 167, + 161, + 426, + 190 + ], + "type": "text", + "content": "Zhiyuan Ren, Minchul Kim, Feng Liu, Xiaoming Liu \nMichigan State University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 176, + 191, + 419, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 191, + 419, + 202 + ], + "spans": [ + { + "bbox": [ + 176, + 191, + 419, + 202 + ], + "type": "text", + "content": "{renzhiy1, kimminc2, liufeng6, liuxm}@msu.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 258, + 290, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 290, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 290, + 534 + ], + "type": "text", + "content": "Recently, diffusion models have emerged as a new powerful generative method for 3D point cloud generation tasks. However, few works study the effect of the architecture of the diffusion model in the 3D point cloud, resorting to the typical UNet model developed for 2D images. Inspired by the wide adoption of Transformers, we study the complementary role of convolution (from UNet) and attention (from Transformers). We discover that their respective importance change according to the timestep in the diffusion process. At early stage, attention has an outsized influence because Transformers are found to generate the overall shape more quickly, and at later stages when adding fine detail, convolution starts having a larger impact on the generated point cloud's local surface quality. In light of this observation, we propose a time-varying two-stream denoising model combined with convolution layers and transformer blocks. We generate an estimizable mask from each timestep to reweigh global and local features, obtaining time-varying fused features. Experimentally, we demonstrate that our proposed method quantitatively outperforms other state-of-the-art methods regarding visual quality and diversity. Code is available https://github.com/Zhiyuan-R/Tiger-Diffusion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 560, + 128, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 128, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 128, + 572 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 715 + ], + "type": "text", + "content": "The point cloud is an essential 3D shape representation that is easy to obtain (sensed directly from laser scanning), versatile (serving as a building block of mesh), and easy to manipulate (geometrically transformed in a straightforward manner by affine matrices). 3D point cloud generative models benefit extensive applications across robotics [35], medicine [15] and content creation [36], and lay the groundwork for other vision tasks like point cloud upsampling [52], point cloud completion [21, 31] and mesh generation [9]. However, compared to 2D images, the cost and complexity of acquiring 3D point clouds make it crucial to ex" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 230, + 544, + 353 + ], + "blocks": [ + { + "bbox": [ + 310, + 230, + 544, + 353 + ], + "lines": [ + { + "bbox": [ + 310, + 230, + 544, + 353 + ], + "spans": [ + { + "bbox": [ + 310, + 230, + 544, + 353 + ], + "type": "image", + "image_path": "20777058ec642375a5e5a540ccee9faeee2ff02eec539403aa98250cb074a8ea.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 361, + 547, + 417 + ], + "lines": [ + { + "bbox": [ + 305, + 361, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 547, + 417 + ], + "type": "text", + "content": "Figure 1. An illustration of different roles played by convolution and attention operations in the denoising model. Convolution is good for learning local relationships, and attention is optimal for modeling global relationships. We propose to merge these two properties across different timesteps in the diffusion process." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 433, + 545, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 433, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 305, + 433, + 545, + 457 + ], + "type": "text", + "content": "plore and develop efficient and effective model architectures for 3D point cloud generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 460, + 546, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 546, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 546, + 615 + ], + "type": "text", + "content": "Existing point cloud generative models are built on a range of frameworks, including generative adversarial networks (GANs) [1, 5], variational autoencoders (VAEs) [24], normalizing flows [23, 27, 51]. However, these methods neither are stable in training nor produce high-fidelity results. Recently, the denoising diffusion model [18], a novel generative method, has demonstrated exceptional results in 3D point cloud generation. Typically, diffusion models perturb a point cloud using a forward process and denoising algorithm. Once trained, the model can be used to generate new point clouds by iterating the reverse process over a sequence of time steps, with each step adding more noise to the input point cloud." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": "Several recent works, including DPM [34] and PVD [55], have successfully applied diffusion models to 3D point cloud generation, achieving high levels of naturalness and diversity. Very recently, LION [53] further improves by encoding point clouds into latent space. However, these methods commonly utilize UNet-like convolutional networks that are originally designed for image processing. There has been limited research on developing suitable de" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9462" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "noising model architectures that are specifically tailored for 3D point cloud generation. For example, PVD and LION both utilize Point-Voxel CNN (PVCNN) [30] as their denoising model. However, we observed that these PVCNN-based denoising models require a considerable number of timesteps to establish a rough shape since the limited receptive field cannot capture the global distribution of noise." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 288, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 288, + 348 + ], + "type": "text", + "content": "In contrast to convolution operations, we observe that the attention mechanism used in Transformers [10, 48] is more effective at capturing long-range dependencies and thus the overall shape of 3D points. This difference is illustrated in the first two rows of Fig. 1. Compared to the convolution-based diffusion model (first row), the attention-based diffusion model (second row) correctly conforms to the overall object shape at earlier denoising time steps (middle column). However, the convolution operation excels in modeling local relationships. Therefore, the details of the final output are better modeled with convolutions (last column) which is also discussed in the ablation analysis of Sec. 4.3. The finding suggests that the two operations, the convolution and the attention are complementary in their roles at different denoising timesteps (the last row of Fig. 1) and we ask the following question." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 362, + 272, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 362, + 272, + 411 + ], + "spans": [ + { + "bbox": [ + 63, + 362, + 272, + 411 + ], + "type": "text", + "content": "How to leverage the strengths of both CNNs and Transformers and enable them to dynamically capture complementary information in denoising diffusion for 3D point cloud generation?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "type": "text", + "content": "To answer this question, we propose a Time-varying denoising model for 3D point cloud generation (TIGER), a two-stream architecture combining a shallow CNN branch and Transformer branch with a timestep-dependent mask that adaptively reassigns weights to the global Transformer feature and the local CNN feature. To enable this feature-level fusion, we design an encoder-decoder to downsample the point cloud for reducing the number of points to a manageable size for Transformer. And we learn a selection mask that finds the optimal configuration at each timestep." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "content": "Furthermore, in this two-stream architecture where global and local information is modeled separately, we discover that supplying 3D points position information to the global branch is critical to the performance. To supply the 3D position information more effectively, we design two novel position encoding methods: phase-shifted position encoding and Base- " + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "content": " position encoding. They create 3D position information in a continuous space and retain the property of linear expression of relative position. The effect of the time-dependent dual branch mask and the proposed position embedding is ablated in Sec. 4.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "A unique property of TIGER is that the weights of the two branches are dependent on the diffusion timesteps, and they are learned and optimized for the 3D point cloud gen" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "eration. Interestingly, experiments demonstrate that the weight of the convolution branch is monotonically increasing with the timestep, which is in accordance with our initial observation that convolutions are good for modeling fine details. Empirically we show that TIGER achieves state-of-the-art (SoTA) performance in 3D point cloud generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 145, + 448, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 145, + 448, + 156 + ], + "spans": [ + { + "bbox": [ + 317, + 145, + 448, + 156 + ], + "type": "text", + "content": "Our main contributions include:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 157, + 545, + 289 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 157, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 157, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 157, + 545, + 204 + ], + "type": "text", + "content": "- We propose a novel two-stream denoising model, which uses timestep to optimally reweigh the global feature from Transformer and the local feature from shallow CNN." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 205, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 205, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 545, + 252 + ], + "type": "text", + "content": "- We design two novel 3D continuous position encoding methods and position-aware self-attention to enable our Transformer branch to aggregate global features effectively from the latent point cloud input." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 253, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 253, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 545, + 289 + ], + "type": "text", + "content": "- Experiments show that our method outperforms prior works in 3D point cloud generation tasks qualitatively and quantitatively on the ShapeNetv2 dataset." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 302, + 392, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 302, + 392, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 302, + 392, + 314 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 323, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 417 + ], + "type": "text", + "content": "3D Point Cloud Generation. The advances in 3D point cloud generation can be categorized by how the input is represented. Early methods [1, 12, 14, 38, 39, 44] represent the point cloud as a " + }, + { + "bbox": [ + 304, + 323, + 545, + 417 + ], + "type": "inline_equation", + "content": "N \\times 3" + }, + { + "bbox": [ + 304, + 323, + 545, + 417 + ], + "type": "text", + "content": " matrix, where " + }, + { + "bbox": [ + 304, + 323, + 545, + 417 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 323, + 545, + 417 + ], + "type": "text", + "content": " is the predefined number of 3D points. The backbone comprises 1D convolutions and max-pooling to allow the points' permutation invariance. So, the shape is represented by a global feature vector." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 418, + 545, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 538 + ], + "type": "text", + "content": "More recent works [5, 7, 13, 27, 34, 51] have improved upon the input representation by introducing a probabilistic distribution assumption, where each point cloud is considered as one data point sampled from the distribution. For instance, PointFlow [51] models the distribution of points by continuous normalizing flows, allowing the sampling of an arbitrary number of points to represent a point cloud. However, the backbone architecture is still based on 1D convolution-based encoder with the max-pooling operation [38]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 539, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 545, + 659 + ], + "type": "text", + "content": "Recently, PVD [55] and LION [53] adopt a hybrid point-voxel representation [30] which voxelize the point clouds for 3D convolution. This representation with rich spatial information combines well with the diffusion objective and produces promising results [53]. However, the backbone is still confined to convolution-based networks, lacking the capacity to effectively model the long range-dependencies. We, on the other hand, find the complementary roles of attention and convolution in the point cloud generator and propose a network that combines the benefits of both." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "Diffusion Models. Diffusion [18] and score-based methods [22, 46] have enjoyed remarkable success in images synthesis. Diffusion has been extended to many generation" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9463" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 83, + 539, + 208 + ], + "blocks": [ + { + "bbox": [ + 58, + 83, + 539, + 208 + ], + "lines": [ + { + "bbox": [ + 58, + 83, + 539, + 208 + ], + "spans": [ + { + "bbox": [ + 58, + 83, + 539, + 208 + ], + "type": "image", + "image_path": "0030f138726516636a8cd7a1379cda31749f6cdb1ba7a50177aa783fc8154afb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "text", + "content": "Figure 2. Illustration of our time-varying two-stream architecture (TIGER). The network's input is a noisy point cloud " + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "text", + "content": ", and the goal is to predict the noise in " + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "text", + "content": " to obtain the denoised point cloud " + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{t-1}" + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "text", + "content": ". We first encode " + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "text", + "content": " into latent point cloud " + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_t" + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "text", + "content": " and export coordinates information for position encoding and decoder. We feed the latent point cloud into both the Latent Point Cloud Transformer and a shallow CNN branch to get global feature " + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{tr}" + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "text", + "content": " and local feature " + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_c" + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "text", + "content": ". Then, we combine two features using timestep dependent mask " + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_c" + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "text", + "content": " to create a fused feature " + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_F" + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "text", + "content": ". Last, we apply our latent point cloud decoder to predict the noise from " + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_F" + }, + { + "bbox": [ + 46, + 224, + 547, + 280 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 297, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 287, + 357 + ], + "type": "text", + "content": "tasks [2, 8, 17, 20, 25, 28, 41-43, 47], and largely improved by innovative techniques, such as fast sampling [33, 45], classifier-free guidance [19, 37] and latent diffusion [4, 42]. Recently 3D point cloud generation with diffusion [34, 53, 55] is also proposed with remarkable diversity and fidelity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 357, + 289, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 289, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 289, + 453 + ], + "type": "text", + "content": "However, previous works do not investigate the behavior of the diffusion backbone with respect to the diffusion timestep. We find that the network equipped with both convolution and attention operations learns to utilize the local and global shape representation with different magnitudes based on the timestep. This time adaptive architecture allows the 3D point cloud generator to be trained faster than LION [53] while producing higher-quality outputs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 467, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 287, + 538 + ], + "type": "text", + "content": "Transformer for Point Clouds. Transformer architecture [48] is intrinsically a set operator, which is suitable for permutation-invariant point clouds. PCT [16] applies global attention to the point cloud. Point Transformer V1 [54] and V2 [49] convert global attention to local attention to reduce memory consumption and computational complexity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 539, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 612 + ], + "type": "text", + "content": "Contrary to the previous methods, our method decouples the global and local information by using Transformer to aggregate global information effectively while using CNN to model the local information. Also, we find that position embedding is crucial to performance and propose a new position embedding to further improve the performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 621, + 104, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 104, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 104, + 633 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "In this section, we first briefly formulate the probabilistic model of forward and reverse diffusion processes for 3D point cloud generation. Then, we dive into the details of our time-varying two-stream architecture, including the encoder part, latent point Transformer, time mask generator, and decoder part. The entire architecture is shown in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 297, + 430, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 430, + 308 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 430, + 308 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 314, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 399 + ], + "type": "text", + "content": "Given a point cloud " + }, + { + "bbox": [ + 304, + 314, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0\\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 304, + 314, + 545, + 399 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 314, + 545, + 399 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 314, + 545, + 399 + ], + "type": "text", + "content": " is the number of points, the diffusion forward process gradually diffuses the original point cloud into Gaussian noise " + }, + { + "bbox": [ + 304, + 314, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_T" + }, + { + "bbox": [ + 304, + 314, + 545, + 399 + ], + "type": "text", + "content": "; the forward process can be formulated with a transition probability " + }, + { + "bbox": [ + 304, + 314, + 545, + 399 + ], + "type": "inline_equation", + "content": "q(\\mathbf{X}_t|\\mathbf{X}_{t - 1})" + }, + { + "bbox": [ + 304, + 314, + 545, + 399 + ], + "type": "text", + "content": ". By giving a sequence of pre-defined noising scales " + }, + { + "bbox": [ + 304, + 314, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\beta_{1},\\beta_{2},\\ldots ,\\beta_{T}" + }, + { + "bbox": [ + 304, + 314, + 545, + 399 + ], + "type": "text", + "content": ", the transition probability can be expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 343, + 403, + 545, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 403, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 343, + 403, + 545, + 419 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {X} _ {t} \\mid \\mathbf {X} _ {t - 1}\\right) = \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} \\mathbf {X} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right). \\tag {1}", + "image_path": "7f28da8a1a12cce97455b3d913c29ef553b7e5aabd8b5972ea15fa055ee3905a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "type": "text", + "content": "For the reverse process, we use learn the " + }, + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{X}_{t - 1}|\\mathbf{X}_t)" + }, + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "type": "text", + "content": " a Gaussian distribution which approximates the intractable real distribution " + }, + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "type": "inline_equation", + "content": "q(\\mathbf{X}_{t - 1}|\\mathbf{X}_t)" + }, + { + "bbox": [ + 304, + 423, + 545, + 460 + ], + "type": "text", + "content": ". Specifically," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 372, + 464, + 545, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 464, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 372, + 464, + 545, + 479 + ], + "type": "interline_equation", + "content": "p _ {\\theta} = \\mathcal {N} \\left(\\mu_ {\\theta} \\left(\\mathbf {X} _ {t}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right), \\tag {2}", + "image_path": "e57574d592eeebeaf7a2deb3b9347df5c5b00ff9df0e00637738156404ac6098.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 482, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 545, + 579 + ], + "type": "text", + "content": "where we learn the mean " + }, + { + "bbox": [ + 304, + 482, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(\\mathbf{X}_t,t)" + }, + { + "bbox": [ + 304, + 482, + 545, + 579 + ], + "type": "text", + "content": " and set the variance " + }, + { + "bbox": [ + 304, + 482, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\sigma_t^2" + }, + { + "bbox": [ + 304, + 482, + 545, + 579 + ], + "type": "text", + "content": " to a fixed schedule as in [18]. Conventionally, " + }, + { + "bbox": [ + 304, + 482, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(\\mathbf{X}_t,t)" + }, + { + "bbox": [ + 304, + 482, + 545, + 579 + ], + "type": "text", + "content": " is modeled using a UNet backbone and it depends on the timestep " + }, + { + "bbox": [ + 304, + 482, + 545, + 579 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 482, + 545, + 579 + ], + "type": "text", + "content": " to scale and shift the feature. We design a two-stream backbone where the timestep modulates the weight between the two streams. To sample a 3D point cloud, we iteratively denoise " + }, + { + "bbox": [ + 304, + 482, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_T,T" + }, + { + "bbox": [ + 304, + 482, + 545, + 579 + ], + "type": "text", + "content": " times to finally generate the highfidelity samples. In training, we optimize the MSE loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 349, + 582, + 545, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 582, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 349, + 582, + 545, + 613 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {s i m p l e}} = \\mathbb {E} _ {t \\sim [ 1, T ]} \\| \\mu - \\mu_ {\\theta} (\\mathbf {X} _ {t}, t) \\| _ {2} ^ {2} \\tag {3} \\\\ = \\mathbb {E} _ {t \\sim [ 1, T ]} \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {X} _ {t}, t) \\| _ {2} ^ {2}, \\\\ \\end{array}", + "image_path": "0185b5ccf2ec5f75051a249c91decffe9b91d1fa165904c873313fa7ea6de5d6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 617, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 617, + 545, + 641 + ], + "type": "text", + "content": " is the ground truth noise and " + }, + { + "bbox": [ + 304, + 617, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 617, + 545, + 641 + ], + "type": "text", + "content": " is the noise prediction by our diffusion model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 647, + 457, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 647, + 457, + 659 + ], + "spans": [ + { + "bbox": [ + 305, + 647, + 457, + 659 + ], + "type": "text", + "content": "3.2. Noisy Point Cloud Encoder" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "To effectively represent the " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "N\\times 3" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": " points, we adopt the voxelization encoding scheme of PVCNN [30]. This representation works well with both CNN and Transformer architecture as we can downsample the voxel-grid following [11] to" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9464" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 76, + 286, + 173 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 286, + 173 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 286, + 173 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 286, + 173 + ], + "type": "image", + "image_path": "5d0d12675c8a255d5c173a700feee286dce8d645432f1cc4d6a57748aa521066.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 183, + 288, + 217 + ], + "lines": [ + { + "bbox": [ + 46, + 183, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 288, + 217 + ], + "type": "text", + "content": "Figure 3. The overview of our encoder. It extracts features in voxel space and downsamples the point cloud by applying trilinear interpolation of voxel with grouped points." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 237, + 287, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 287, + 320 + ], + "type": "text", + "content": "increase computation efficiency. Specifically, a noisy point cloud encoder " + }, + { + "bbox": [ + 46, + 237, + 287, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{E}:\\mathbb{R}^{N\\times 3}\\to \\mathbb{R}^{M\\times d}" + }, + { + "bbox": [ + 46, + 237, + 287, + 320 + ], + "type": "text", + "content": " transforms the noisy point cloud " + }, + { + "bbox": [ + 46, + 237, + 287, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t\\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 46, + 237, + 287, + 320 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 46, + 237, + 287, + 320 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 237, + 287, + 320 + ], + "type": "text", + "content": " into a latent point cloud representation " + }, + { + "bbox": [ + 46, + 237, + 287, + 320 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_t\\in \\mathbb{R}^{M\\times d}" + }, + { + "bbox": [ + 46, + 237, + 287, + 320 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 237, + 287, + 320 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 237, + 287, + 320 + ], + "type": "text", + "content": " is the number of original points, " + }, + { + "bbox": [ + 46, + 237, + 287, + 320 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 237, + 287, + 320 + ], + "type": "text", + "content": " is the number of downsampled points and " + }, + { + "bbox": [ + 46, + 237, + 287, + 320 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 237, + 287, + 320 + ], + "type": "text", + "content": " is the dimension of each latent point's feature. The overview of our encoder is illustrated in Fig. 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "text", + "content": "To aggregate features in a rigid field, we first voxelize the noisy point cloud " + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\{x_{n},y_{n},z_{n}\\} = \\mathbf{X}_{t}\\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "text", + "content": " into the voxel grids " + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{V}_{u,v,w}\\} = \\mathbf{V}\\in \\mathbb{R}^{L\\times L\\times L}" + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "text", + "content": " is the resolution and " + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "text", + "content": " is the point index in " + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "text", + "content": ". Specifically," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 376, + 287, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 287, + 399 + ], + "type": "interline_equation", + "content": "\\mathbf {V} _ {u, v, w} = \\sum_ {n = 1} ^ {N} \\mathbb {I} \\left(x _ {n} \\in \\mathbb {N} (u, \\delta), y _ {n} \\in \\mathbb {N} (v, \\delta), z _ {n} \\in \\mathbb {N} (w, \\delta)\\right) * \\mathbf {f} _ {n}, \\tag {4}", + "image_path": "2ecd39b1260ea6a2f6a51a6dbd6093168f5faa11c98233630082bd08ef3628cd.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathbb{N}(z,\\delta) = \\{x|x\\in [z - \\delta ,z + \\delta)\\}" + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_n" + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "inline_equation", + "content": "n^{th}" + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "text", + "content": " extracted latent feature corresponding to the " + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "inline_equation", + "content": "n^{th}" + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "text", + "content": " point " + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}_t)_n" + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(\\cdot)" + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "text", + "content": " is an indicator of whether the " + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "inline_equation", + "content": "n^{th}" + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "text", + "content": " point falls into the voxel grid " + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "inline_equation", + "content": "(u,v,w)" + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "text", + "content": ". After voxelization, we apply a series of 3D convolution, Swish [40] activation and Group-Norm [50]. After applying a series of these operations, this results in a latent volume " + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}\\in \\mathbb{R}^{L\\times L\\times L\\times d}" + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 400, + 287, + 495 + ], + "type": "text", + "content": " is the channel size for each voxel." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "text", + "content": "We use furthest point sampling algorithm [11] to downsample the input noisy point cloud " + }, + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t\\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "text", + "content": " into a sparser point cloud " + }, + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^s\\in \\mathbb{R}^{M\\times 3}" + }, + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "inline_equation", + "content": "M < N" + }, + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "text", + "content": "). Then, as the position in volume is discrete, we query the latent volume " + }, + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}" + }, + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "text", + "content": " by the sparse point cloud " + }, + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^s" + }, + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "text", + "content": " via trilinear interpolation, eventually transformed into latent point cloud " + }, + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_t\\in \\mathbb{R}^{M\\times d}" + }, + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "text", + "content": ". It is worth noting we preserve the coordinates of " + }, + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^s" + }, + { + "bbox": [ + 46, + 497, + 287, + 604 + ], + "type": "text", + "content": " to do upsampling with the decoder and provide position embedding for our latent point cloud Transformer." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 611, + 222, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 222, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 222, + 623 + ], + "type": "text", + "content": "3.3. Latent Point Cloud Transformer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": "Previous denoising models such as PVCNN [30] simply apply stacked convolutional layers to create feature representations. Contrarily, we aim to better model the global relationship by introducing a novel Transformer branch named latent point cloud Transformer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Tokenization. After encoding, we get a latent point cloud representation " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_t\\in \\mathbb{R}^{M\\times d}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " and its coordinates " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^s\\in \\mathbb{R}^{M\\times 3}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 425, + 158 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 425, + 158 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 425, + 158 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 425, + 158 + ], + "type": "image", + "image_path": "2348bc51c4bc105e548c4f3b0f78a9e8932dc7feaf9fea47027652e82f247cec.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 161, + 383, + 170 + ], + "lines": [ + { + "bbox": [ + 351, + 161, + 383, + 170 + ], + "spans": [ + { + "bbox": [ + 351, + 161, + 383, + 170 + ], + "type": "text", + "content": "(a) PSPE" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 543, + 158 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 543, + 158 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 543, + 158 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 543, + 158 + ], + "type": "image", + "image_path": "c1db3196b9deaa2816c4d79e4ad6b7cbee8d8f1fbf38de23e567196690137c13.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 160, + 503, + 170 + ], + "lines": [ + { + "bbox": [ + 468, + 160, + 503, + 170 + ], + "spans": [ + { + "bbox": [ + 468, + 160, + 503, + 170 + ], + "type": "text", + "content": "(b) BλPE" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 178, + 545, + 212 + ], + "lines": [ + { + "bbox": [ + 305, + 178, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 545, + 212 + ], + "type": "text", + "content": "Figure 4. Illustration of two examples of the position embedding from PSPE and BλPE respectively. Both methods show distinguished representation for each position." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 232, + 545, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 232, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 305, + 232, + 545, + 267 + ], + "type": "text", + "content": "Following [29], we use dual PatchNorm to project the latent point cloud into tokens, which place LayerNorm before and after an MLP layer for more stable training:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 369, + 274, + 545, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 274, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 369, + 274, + 545, + 289 + ], + "type": "interline_equation", + "content": "\\mathbf {T} = \\operatorname {L N} \\left(\\operatorname {M L P} \\left(\\operatorname {L N} \\left(\\hat {\\mathbf {X}} _ {\\mathbf {t}}\\right)\\right)\\right), \\tag {5}", + "image_path": "ef1887d14b5421d527dcecfcdaae93726022106ef47102d157d8422839a8cda2.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 294, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 294, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 294, + 545, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 294, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 305, + 294, + 545, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 294, + 545, + 319 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 294, + 545, + 319 + ], + "type": "text", + "content": " is the dimension in the token space (in our paper, " + }, + { + "bbox": [ + 305, + 294, + 545, + 319 + ], + "type": "inline_equation", + "content": "D = 2d" + }, + { + "bbox": [ + 305, + 294, + 545, + 319 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 319, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 545, + 450 + ], + "type": "text", + "content": "Position Encoding. Since the order of the latent points is arbitrary in a point cloud, it's necessary to create a 3D position embedding and fuse it into the feature, so that the model knows the position of each point. Contrary to the position embedding in image domains [10, 26], the design of 3D position embedding plays a vital role in the performance and our position embedding produces the best result. We propose two novel 3D space continuous position encoding methods: Phase Shift Position Encoding (PSPE) and Base- " + }, + { + "bbox": [ + 305, + 319, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 319, + 545, + 450 + ], + "type": "text", + "content": " Position Encoding (B" + }, + { + "bbox": [ + 305, + 319, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 319, + 545, + 450 + ], + "type": "text", + "content": "PE). For PSPE, we use Sine and Cosine functions of different frequencies and phases:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 456, + 545, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 545, + 490 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {e m b} \\left(p o s _ {j}, 6 i + 2 * (j - 1)\\right) = \\sin \\left(\\frac {p o s _ {j}}{1 0 0 0 0 ^ {\\frac {2 i}{D}}} + (j - 1) \\frac {2 \\pi}{3}\\right) \\tag {6}", + "image_path": "7057abf5c33482bad19a8280f65c6f519760d705b80c5d1b82f39c18443ecfce.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 490, + 548, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 548, + 522 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 548, + 522 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {e m b} (p o s _ {j}, 6 i + 1 + 2 * (j - 1)) = \\cos \\left(\\frac {p o s _ {j}}{1 0 0 0 0 ^ {\\frac {2 i}{D}}} + (j - 1) \\frac {2 \\pi}{3}\\right), \\tag {7}", + "image_path": "6531ca0415f74e9f7a7bc5ac015c93d46ee9fbc0f43a2e50d1be781ba75fb40d.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "inline_equation", + "content": "pos_{j}" + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "text", + "content": "-th element in each point's position, specifically " + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "inline_equation", + "content": "pos_{1} = x" + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "inline_equation", + "content": "pos_{2} = y" + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "inline_equation", + "content": "pos_{3} = z" + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "text", + "content": " is the index of the dimension. We maximize the phase shift difference " + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "inline_equation", + "content": "(\\frac{4\\pi}{3})" + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "text", + "content": " of different axes to distinguish them and combine Sine and Cosine to guarantee the linear representation of relative position, which will be proved mathematically in Supp. An example is shown in Fig. 4a and " + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{emb}" + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "text", + "content": " is shaped " + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{M\\times D}" + }, + { + "bbox": [ + 305, + 523, + 545, + 606 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 607, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 607, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 305, + 607, + 545, + 652 + ], + "type": "text", + "content": "However, PSPE requires 6 channels to represent one frequency feature, which is quite inefficient. To improve channel utilization, we apply BλPE to compress 3D position into one scalar:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 372, + 654, + 545, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 654, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 372, + 654, + 545, + 666 + ], + "type": "interline_equation", + "content": "p o s = \\lambda^ {2} * z + \\lambda * y + x \\tag {8}", + "image_path": "80deae241a9dba80921dfaeba58d5ac36756ad3c60f3266f884ad3cb9199d91f.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 361, + 669, + 545, + 692 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 669, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 361, + 669, + 545, + 692 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {e m b} (p o s, 2 i) = \\sin \\left(\\frac {p o s}{1 0 0 0 0 ^ {\\frac {2 i}{D}}}\\right) \\tag {9}", + "image_path": "8cec2d537da2197a81234f5be1384b6496e489a32661ba7885aa452c6475ff31.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 350, + 694, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 694, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 350, + 694, + 545, + 717 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {e m b} (p o s, 2 i + 1) = \\cos \\left(\\frac {p o s}{1 0 0 0 0 ^ {\\frac {2 i}{D}}}\\right), \\tag {10}", + "image_path": "5d34648d804817503d9dab6e80cdaef7bc69b58f2d077173620829624da9684d.jpg" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9465" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 73, + 281, + 249 + ], + "blocks": [ + { + "bbox": [ + 50, + 73, + 281, + 249 + ], + "lines": [ + { + "bbox": [ + 50, + 73, + 281, + 249 + ], + "spans": [ + { + "bbox": [ + 50, + 73, + 281, + 249 + ], + "type": "image", + "image_path": "49b1d9a93ac5046301acf56d70da9c791842c0f7556fa73c8c610a5cf832f53b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 256, + 288, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 256, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 288, + 291 + ], + "type": "text", + "content": "Figure 5. The overview of the position-aware self-attention. We compute position relationship map " + }, + { + "bbox": [ + 46, + 256, + 288, + 291 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 46, + 256, + 288, + 291 + ], + "type": "text", + "content": " by previous position embedding to reweigh the " + }, + { + "bbox": [ + 46, + 256, + 288, + 291 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}\\mathbf{K}^T" + }, + { + "bbox": [ + 46, + 256, + 288, + 291 + ], + "type": "text", + "content": " matrix." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 312, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 312, + 287, + 396 + ], + "type": "inline_equation", + "content": "pos" + }, + { + "bbox": [ + 46, + 312, + 287, + 396 + ], + "type": "text", + "content": " is a polynomial expression of " + }, + { + "bbox": [ + 46, + 312, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 312, + 287, + 396 + ], + "type": "text", + "content": ". Specifically, we set " + }, + { + "bbox": [ + 46, + 312, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 312, + 287, + 396 + ], + "type": "text", + "content": " to 1000, which means this preserves three decimal places of precision and one example is shown in Fig. 4b. The choice between these two position encoding methods is essentially a trade-off between granularity and expressiveness. We ablate them and other non-position encoding and learnable position encoding in Sec. 4.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "text", + "content": "Transformer and Position-Aware Attention We stack " + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "text", + "content": " layers of Transformer block [48] which comprises of LayerNorm(LN) [3] and Multihead Self Attention (MHA) and MultiLayer Perceptrons. We let the token " + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_0 = \\mathbf{T}" + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "text", + "content": " and let" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 456, + 287, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 456, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 94, + 456, + 287, + 469 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {l *} = \\operatorname {M H A} \\left(\\ln \\left(\\mathbf {T} _ {l - 1}\\right)\\right) + \\mathbf {T} _ {l - 1}, \\tag {11}", + "image_path": "75ee3dc46cce22e394150c6b910c3c481d989ea1de105d1a85ac1db1d9a723ff.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 471, + 287, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 471, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 96, + 471, + 287, + 483 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {l} = \\operatorname {M L P} \\left(\\operatorname {L N} \\left(\\mathbf {T} _ {l *}\\right)\\right) + \\mathbf {T} _ {l *}, \\tag {12}", + "image_path": "0646bfce44b727845f3e618864f5a0000eb188278213110ad983e24ab3cefb57.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 486, + 287, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 486, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 96, + 486, + 287, + 498 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {t r} := \\mathbf {T} _ {L}. \\tag {13}", + "image_path": "36f3f75140ad664534f37d92b6fe889ec7faa5d7ac55e28d61a7be68fca6e0ff.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 509, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 594 + ], + "type": "text", + "content": "To add position embedding into the model, one can simply add it to " + }, + { + "bbox": [ + 46, + 509, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 46, + 509, + 287, + 594 + ], + "type": "text", + "content": ". But as the Transformer blocks are gradually stacked, the continuous 3D position encoding is difficult to retain in the token representation. Inspired by [54], we explicitly add the position encoding to each MHA in order to make our attention map position-aware, as shown in Fig. 5. Specifically, our position-aware MHA modified as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 82, + 604, + 287, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 604, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 82, + 604, + 287, + 619 + ], + "type": "interline_equation", + "content": "\\mathbf {H} = \\operatorname {S o f t m a x} \\left(\\left(\\mathbf {P} _ {e m b} \\mathbf {W} _ {p}\\right) \\left(\\mathbf {P} _ {e m b} \\mathbf {W} _ {p}\\right) ^ {T}\\right) \\tag {14}", + "image_path": "d963d01feee836e5187f9fb0338aa8a74d39f8f7f717b67db82455c8ad6b3be0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 630, + 287, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 630, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 53, + 630, + 287, + 658 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {l *} = \\operatorname {S o f t m a x} \\left(\\frac {\\left(\\mathbf {T} _ {l} \\mathbf {W} _ {q}\\right) \\left(\\mathbf {T} _ {l} \\mathbf {W} _ {k}\\right) ^ {T} \\odot \\mathbf {H}}{\\sqrt {D}}\\right) \\left(\\mathbf {T} _ {l} \\mathbf {W} _ {v}\\right) \\tag {15}", + "image_path": "42a0f7c73a820d24021291dd02d2e51d1eca630d104f18c922a83da287b79767.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q,\\mathbf{W}_k,\\mathbf{W}_v,\\mathbf{W}_p\\in \\mathbb{R}^{D\\times Z}" + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "text", + "content": " are projection matrices and " + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{H}\\in \\mathbb{R}^{M\\times M}" + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "text", + "content": " is the position relationship map. Modified MHA could incorporate position information explicitly and sufficiently during training. " + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{tr}" + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "text", + "content": " is the final output." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 76, + 545, + 150 + ], + "blocks": [ + { + "bbox": [ + 310, + 76, + 545, + 150 + ], + "lines": [ + { + "bbox": [ + 310, + 76, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 310, + 76, + 545, + 150 + ], + "type": "image", + "image_path": "ad7b3332409ec3d6e4b12c4f9b36848faeea9c190734e000f0bcb3e3d116d1ca.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 159, + 546, + 194 + ], + "lines": [ + { + "bbox": [ + 305, + 159, + 546, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 159, + 546, + 194 + ], + "type": "text", + "content": "Figure 6. The overview of our decoder. By querying the latent volume " + }, + { + "bbox": [ + 305, + 159, + 546, + 194 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_F" + }, + { + "bbox": [ + 305, + 159, + 546, + 194 + ], + "type": "text", + "content": " with previous coordinates information " + }, + { + "bbox": [ + 305, + 159, + 546, + 194 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 305, + 159, + 546, + 194 + ], + "type": "text", + "content": ", we can upsample the latent point cloud into 3D space." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 210, + 433, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 210, + 433, + 221 + ], + "spans": [ + { + "bbox": [ + 306, + 210, + 433, + 221 + ], + "type": "text", + "content": "3.4. Time Mask Generator" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 228, + 545, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 228, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 545, + 311 + ], + "type": "text", + "content": "We apply a depth-2 PVCNN [30] to get local feature " + }, + { + "bbox": [ + 305, + 228, + 545, + 311 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_c\\in" + }, + { + "bbox": [ + 305, + 228, + 545, + 311 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{M\\times D}" + }, + { + "bbox": [ + 305, + 228, + 545, + 311 + ], + "type": "text", + "content": " and our latent point cloud Transformer to get global feature " + }, + { + "bbox": [ + 305, + 228, + 545, + 311 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{tr}\\in \\mathbb{R}^{M\\times D}" + }, + { + "bbox": [ + 305, + 228, + 545, + 311 + ], + "type": "text", + "content": ". Time mask generator module scales the local and global features by the time variable. We force the sum of two learnable time masks to be 1. With this constraint, the trade-off of local and global features can be learned by the network." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 312, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 545, + 335 + ], + "type": "text", + "content": "We first align the dimension and scale of " + }, + { + "bbox": [ + 305, + 312, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_c" + }, + { + "bbox": [ + 305, + 312, + 545, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 312, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{tr}" + }, + { + "bbox": [ + 305, + 312, + 545, + 335 + ], + "type": "text", + "content": " by one layer MLP and Sigmoid function " + }, + { + "bbox": [ + 305, + 312, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 305, + 312, + 545, + 335 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 343, + 545, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 343, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 321, + 343, + 545, + 357 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {c} ^ {*} = \\sigma (\\operatorname {M L P} (\\mathbf {X} _ {c})), \\quad \\mathbf {X} _ {t r} ^ {*} = \\sigma (\\operatorname {M L P} (\\mathbf {X} _ {t r})). \\tag {16}", + "image_path": "341cc9a80d326c577d568455ddf1d8f676dc684187e7a605980e6542c7c63207.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 365, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 413 + ], + "type": "text", + "content": "Then we encode a timestep " + }, + { + "bbox": [ + 305, + 365, + 545, + 413 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 365, + 545, + 413 + ], + "type": "text", + "content": " to a sinusoidal position embedding [18, 48] to " + }, + { + "bbox": [ + 305, + 365, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{emb} \\in \\mathbb{R}^c" + }, + { + "bbox": [ + 305, + 365, + 545, + 413 + ], + "type": "text", + "content": ". We use two stacked MLP to map " + }, + { + "bbox": [ + 305, + 365, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{emb}" + }, + { + "bbox": [ + 305, + 365, + 545, + 413 + ], + "type": "text", + "content": " to a higher dimension space and apply Sigmoid to keep one of the masks between 0 to 1. Specifically," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 327, + 421, + 545, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 421, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 327, + 421, + 545, + 434 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {c} = \\sigma (\\mathrm {M L P} (\\text {L e a k y R e L U} (\\mathrm {M L P} (\\mathbf {t} _ {\\text {e m b}}))), \\tag {17}", + "image_path": "1929b81bc18d1e6c8433971207a75fb347799d3144fd9c83ad125071ac2bc0dc.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 325, + 436, + 545, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 436, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 325, + 436, + 545, + 449 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {t r} = \\mathbf {1} - \\mathbf {M} _ {c}. \\tag {18}", + "image_path": "d263518d04ba4d03914a45441f21ba76c5a10b946ea9015d7eccaaab7146ae4a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 457, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 457, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 305, + 457, + 545, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 457, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_c" + }, + { + "bbox": [ + 305, + 457, + 545, + 518 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 457, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{tr} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 305, + 457, + 545, + 518 + ], + "type": "text", + "content": " are the time masks for CNN and Transformer feature respectively. We ablate the dimension of masks in Sec. 4.3 to be " + }, + { + "bbox": [ + 305, + 457, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^D" + }, + { + "bbox": [ + 305, + 457, + 545, + 518 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 305, + 457, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^1" + }, + { + "bbox": [ + 305, + 457, + 545, + 518 + ], + "type": "text", + "content": ". We multiply these two masks by the respective features and adjust the fused feature by MLP." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 345, + 525, + 545, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 525, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 345, + 525, + 545, + 539 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {X}} _ {F} = \\operatorname {M L P} \\left(\\mathbf {X} _ {c} ^ {*} \\odot \\mathbf {M} _ {c} + \\mathbf {X} _ {t r} ^ {*} \\odot \\mathbf {M} _ {c}\\right), \\tag {19}", + "image_path": "7b9acc6734e80d70947c681ad43d0e0ae8752ce1a8c8337193100b5ae3eefd3a.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 547, + 545, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 547, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 305, + 547, + 545, + 572 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 547, + 545, + 572 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_F\\in \\mathbb{R}^{M\\times D}" + }, + { + "bbox": [ + 305, + 547, + 545, + 572 + ], + "type": "text", + "content": " is the final feature adaptively aggregating local and global features based on the timestep." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 578, + 460, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 578, + 460, + 590 + ], + "spans": [ + { + "bbox": [ + 306, + 578, + 460, + 590 + ], + "type": "text", + "content": "3.5. Latent Point Cloud Decoder" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 597, + 545, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 597, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 305, + 597, + 545, + 692 + ], + "type": "text", + "content": "In order to upsample the latent point cloud and predict the corresponding noise, we design a latent point cloud decoder " + }, + { + "bbox": [ + 305, + 597, + 545, + 692 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 597, + 545, + 692 + ], + "type": "text", + "content": ", which converts the latent point cloud " + }, + { + "bbox": [ + 305, + 597, + 545, + 692 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_F \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 305, + 597, + 545, + 692 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 597, + 545, + 692 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 305, + 597, + 545, + 692 + ], + "type": "text", + "content": ". Similar to Sec. 3.2, we can voxelize " + }, + { + "bbox": [ + 305, + 597, + 545, + 692 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_F" + }, + { + "bbox": [ + 305, + 597, + 545, + 692 + ], + "type": "text", + "content": " into a latent volume " + }, + { + "bbox": [ + 305, + 597, + 545, + 692 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}}_F \\in \\mathbb{R}^{L \\times L \\times L \\times D}" + }, + { + "bbox": [ + 305, + 597, + 545, + 692 + ], + "type": "text", + "content": ", and take " + }, + { + "bbox": [ + 305, + 597, + 545, + 692 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 305, + 597, + 545, + 692 + ], + "type": "text", + "content": " as the query for trilinear interpolation, eventually resulting in the final noise prediction " + }, + { + "bbox": [ + 305, + 597, + 545, + 692 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 305, + 597, + 545, + 692 + ], + "type": "text", + "content": ". This formulation can be expressed as:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 328, + 700, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 700, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 328, + 700, + 545, + 714 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\theta} = \\operatorname {T r i l i n e a r} \\left(\\operatorname {V o x e l i z e} \\left(\\hat {\\mathbf {X}} _ {F}\\right), \\mathbf {X} _ {t}\\right) \\mathbf {W} ^ {D \\times 3}, \\tag {20}", + "image_path": "d0d1633bcde405193825a9f31b30e3814295973577ae8d844a564cfb5c63b88f.jpg" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9466" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 520, + 242 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 520, + 242 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 520, + 242 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 520, + 242 + ], + "type": "table", + "html": "
MethodGenerative ModelAirplaneChairCar
CD→50%EMD→50%CD→50%EMD→50%CD→50%EMD→50%
r-GAN [1]GAN98.4096.7983.6999.7094.4699.01
l-GAN(CD) [1]GAN87.3093.9568.5883.8466.4988.78
l-GAN(EMD) [1]GAN89.4976.9171.9064.6571.1666.19
PointFlow [51]Normalizing Flow75.6870.7462.8460.5758.1056.52
DPF-Net [27]Normalizing Flow75.1865.5562.0058.5362.3554.48
ShapeGF [5]GAN80.0076.1768.9665.4863.2056.53
SoftFlow [23]Normalizing Flow76.0565.8059.2160.0564.7760.09
SetVAE [24]VAE76.5467.6558.8460.5759.9559.94
DPM [34]Diffusion76.4286.9160.0574.7768.8979.97
PVD [55]Diffusion73.8264.8156.2653.3254.5553.83
TIGERDiffusion71.8555.8254.6152.7154.3152.24
LION [53]Diffusion67.4161.2353.7052.3453.4151.14
TIGERDiffusion67.2156.2654.3251.7154.1250.24
", + "image_path": "b0f2c3c6cdadb2cbb64979d8cbe9ff08666450cf659e482c94372239e3e5c254.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 247, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 547, + 270 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with baselines using 1-NN. Both CD and EMD are considered, where CD is multiplied by " + }, + { + "bbox": [ + 46, + 247, + 547, + 270 + ], + "type": "inline_equation", + "content": "{10}^{3}" + }, + { + "bbox": [ + 46, + 247, + 547, + 270 + ], + "type": "text", + "content": " and EMD is multiplied by " + }, + { + "bbox": [ + 46, + 247, + 547, + 270 + ], + "type": "inline_equation", + "content": "{10}^{2}" + }, + { + "bbox": [ + 46, + 247, + 547, + 270 + ], + "type": "text", + "content": " . The closer the score is to 50%,the better the quality and diversity of generated samples. [Key: Best, Second Best]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 62, + 280, + 531, + 445 + ], + "blocks": [ + { + "bbox": [ + 62, + 280, + 531, + 445 + ], + "lines": [ + { + "bbox": [ + 62, + 280, + 531, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 531, + 445 + ], + "type": "image", + "image_path": "18846db4820861bec226d324c3415b846fd7c495f424a059089b22abe3ad0c49.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 451, + 541, + 464 + ], + "lines": [ + { + "bbox": [ + 50, + 451, + 541, + 464 + ], + "spans": [ + { + "bbox": [ + 50, + 451, + 541, + 464 + ], + "type": "text", + "content": "Figure 7. Our generation results (right) compared to baseline models (left). TIGER generates high-quality and diverse 3D point clouds." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 483, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 483, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{D\\times 3}" + }, + { + "bbox": [ + 46, + 483, + 287, + 521 + ], + "type": "text", + "content": " is the projection matrix to map the noise into 3D space and " + }, + { + "bbox": [ + 46, + 483, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 46, + 483, + 287, + 521 + ], + "type": "text", + "content": " is the original point cloud coordinates. The overview of our decoder is shown in Fig. 6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 531, + 128, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 128, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 128, + 544 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 551, + 163, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 163, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 163, + 564 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 677 + ], + "type": "text", + "content": "Datasets. We follow previous works and choose ShapeNetv2 [6] as our main dataset, which is pre-processed by PointFlow [51]. To fairly compare with other different baseline methods, we train and evaluate on three categories: airplane, chair and car. For each shape, we sample 2,048 points and normalize them globally across the entire dataset. The training set consists of 2,832,4,612, and 2,458 shapes, while the evaluation set is composed of 405,662, and 352 shapes for airplanes, chairs, and cars, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Metrics. Following the baselines PVD [55] and LION [53], we use 1-NN (1-nearest neighbor) accuracy [32] as our evaluation metric. This metric has been shown to effectively" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "type": "text", + "content": "measure both the quality and diversity of generated point clouds and a score closer to " + }, + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "type": "text", + "content": " indicates superior performance [51]. To calculate the 1-NN distance matrix, we consider two commonly used metrics for measuring the distance between point clouds: CD (Chamfer Distance) and EMD (Earth Movers' Distance)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 563, + 481, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 481, + 576 + ], + "type": "text", + "content": "4.2. Comparison with SoTA methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Results. To compare our time-varying denoising model with SoTA methods, we evaluate its performance against competitive models such as SetVAE [24], DPM [34], PVD [55] and LION [53]. It is noteworthy that in order to compare with LION, which uses a different dataset splitting strategy (sampling from the first 10,000 points instead of the latter 5,000 points), we evaluate our model under both LION's strategy and the standard strategy used by other methods. As shown in Tab. 1, we outperform LION in four out of six metrics. Compared to other methods, our performance is significantly better. Notably, our EMD-based" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9467" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 120 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 120 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 120 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 120 + ], + "type": "table", + "html": "
MethodTraining Time (GPU hours)Inference Time (s)
PVD [55]1428.46
LION [53]55027.12
Tiger1649.73
", + "image_path": "56fde96708d3b6d5720031c3100d036c0e299dbf21347878db742b80b7a58dc7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 152, + 286, + 217 + ], + "blocks": [ + { + "bbox": [ + 47, + 124, + 286, + 146 + ], + "lines": [ + { + "bbox": [ + 47, + 124, + 286, + 146 + ], + "spans": [ + { + "bbox": [ + 47, + 124, + 286, + 146 + ], + "type": "text", + "content": "Table 2. Training and inference time comparison. Training time is counted on average for three categories: airplane, chair and car." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 152, + 286, + 217 + ], + "lines": [ + { + "bbox": [ + 50, + 152, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 50, + 152, + 286, + 217 + ], + "type": "table", + "html": "
MethodW1W2Mean CDMean EMD
Convolution1061.3456.31
Transformer0162.4155.85
Conv+Tran0.50.564.4557.92
Ourslearnable (scalar)60.7954.12
Ourslearnable (channel-wise)59.4353.51
", + "image_path": "2616913c8571b50801082ca6647f18dc08129298d005e968bfd16c62d3d77c28.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 269, + 286, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 286, + 352 + ], + "type": "text", + "content": "1-NN accuracy shows a significant improvement of " + }, + { + "bbox": [ + 46, + 269, + 286, + 352 + ], + "type": "inline_equation", + "content": "13.45\\%" + }, + { + "bbox": [ + 46, + 269, + 286, + 352 + ], + "type": "text", + "content": " relative to PVD and " + }, + { + "bbox": [ + 46, + 269, + 286, + 352 + ], + "type": "inline_equation", + "content": "8.12\\%" + }, + { + "bbox": [ + 46, + 269, + 286, + 352 + ], + "type": "text", + "content": " relative to LION for airplanes. This is attributed to the efficient aggregation of global features by our Transformer branch, which is especially beneficial for EMD as it is primarily influenced by global distribution. Fig. 7 shows that our generated samples are of high quality and diversity compared with other methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 353, + 286, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 286, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 286, + 448 + ], + "type": "text", + "content": "Efficiency. To compare the training and inference speed, we report the GPU hours with NVidia V100 GPUs as in LION [53]. As shown in Tab. 2, our model trains much faster than LION, with only a quarter of its training time and a third of its inference time. Our training and inference speed is comparable to that of PVD while surpassing LION and PVD in the generation quality as shown in Tab. 1. More computation analysis can be found in Supp." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 456, + 173, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 173, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 173, + 467 + ], + "type": "text", + "content": "4.3. Ablation and Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 474, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 286, + 713 + ], + "type": "text", + "content": "Time-varying two-stream architecture. We ablate our time-varying two-stream architecture by comparing it with single ConvNet, single Transformer, and two-stream networks with equal weight. In this ablation, we also compare the performance of our time masking with scalar value setting and channel-wise value setting. The results, presented in the first two rows of Tab. 3, show that Transformer performs better on the EMD metric, while ConvNet performs better on the CD metric. Using a fixed equal weight for the two-stream network leads to a significant performance drop on both metrics. However, the learnable time mask improves the performance on both metrics, indicating the importance of the time-varying design. Further, channel-wise time masking improves performance slightly more than scale value setting due to its greater expressivity. Transformer backbones, position encoding, and self-attention strategies. Here we carried out three ablations: (1) We compare the performance of our Transformer backbone with two classic Transformer backbones, PCT [16] and Point Transformer (PointTran) [54]. (2) To assess" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 542, + 190 + ], + "blocks": [ + { + "bbox": [ + 47, + 221, + 286, + 255 + ], + "lines": [ + { + "bbox": [ + 47, + 221, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 221, + 286, + 255 + ], + "type": "text", + "content": "Table 3. Ablation of different architecture designs. " + }, + { + "bbox": [ + 47, + 221, + 286, + 255 + ], + "type": "inline_equation", + "content": "W_{1}" + }, + { + "bbox": [ + 47, + 221, + 286, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 221, + 286, + 255 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 47, + 221, + 286, + 255 + ], + "type": "text", + "content": " are the weights for the ConvNet branch and Transformer branch. [Key: Best, Second Best]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 70, + 542, + 190 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 542, + 190 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 542, + 190 + ], + "type": "table", + "html": "
TBPEAPMean CDMean EMD
PCT [16]Learnable-65.8260.21
PointTran [54]LearnableLearnable66.4262.39
Ours--63.5157.98
OursLearnableLearnable62.1756.13
OursPSPE-60.1155.74
OursPSPELearnable61.4855.97
OursPSPEPSPE59.0254.49
OursBλPE-61.3554.27
OursBλPELearnable61.8956.02
OursBλPEBλPE59.4353.51
", + "image_path": "01e482d857c59e94089b584f2c8e01aa5b7a27c5e0992ad59941197e883f93b6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 194, + 545, + 237 + ], + "lines": [ + { + "bbox": [ + 305, + 194, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 305, + 194, + 545, + 237 + ], + "type": "text", + "content": "Table 4. Ablation of Transformer backbones, position encoding, and self-attention strategies. [Key: Best, Second Best; TB=Transformer backbone; PE=position encoding; AP=additional position information]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 310, + 248, + 440, + 364 + ], + "blocks": [ + { + "bbox": [ + 310, + 248, + 440, + 364 + ], + "lines": [ + { + "bbox": [ + 310, + 248, + 440, + 364 + ], + "spans": [ + { + "bbox": [ + 310, + 248, + 440, + 364 + ], + "type": "image", + "image_path": "a7487b1c2e2fa1ec6dd7186593fd646c1f0b70302f4d35ce36c14cbcbedd6530.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 372, + 545, + 470 + ], + "lines": [ + { + "bbox": [ + 305, + 372, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 545, + 470 + ], + "type": "text", + "content": "Figure 8. Visualization of attention maps in 3D point clouds. Column (c) is the result of column (b) multiplied by position relationship map H. Taking the first row as an example, queried by the airplane head's position, our method can aggregate the position information from the airplane's right wing and left elevator compared with standard self-attention. This comparison highlights the effectiveness of our position-aware self-attention module in capturing long-range relationships in 3D point clouds. [SA: Self-attention, PASA: Position-aware Self-attention]" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 464, + 247, + 541, + 364 + ], + "blocks": [ + { + "bbox": [ + 464, + 247, + 541, + 364 + ], + "lines": [ + { + "bbox": [ + 464, + 247, + 541, + 364 + ], + "spans": [ + { + "bbox": [ + 464, + 247, + 541, + 364 + ], + "type": "image", + "image_path": "23fe7c856fb0923408d7a66d7d8252ac8419dd825c62a27118ab97168b3fa04a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "our proposed position encoding methods, PSPE and BλPE, we compare them with no position encoding and learnable position encoding. (3) We investigate our position-aware self-attention module by comparing it with standard self-attention and self-attention with learnable position encoding as proposed in [54]. As shown in Tab. 4, our latent point cloud Transformer outperforms both PCT and Point Transformer under various settings. Furthermore, our proposed position encoding methods, PSPE and BλPE, significantly improve performance compared to no position encoding or learnable position encoding. Specifically, PSPE improves the CD metric, while BλPE improves the EMD metric. Our position-aware self-attention module, which aggregates corresponding position encoding, achieves better results than standard self-attention and self-attention with learnable position encoding. To illustrate this, Fig. 8 visualizes the attention map in 3D point space for both standard self-attention and our position-aware self-attention. This visualization demonstrates the ability of our method to capture" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9468" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 68, + 213, + 159 + ], + "blocks": [ + { + "bbox": [ + 53, + 68, + 213, + 159 + ], + "lines": [ + { + "bbox": [ + 53, + 68, + 213, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 68, + 213, + 159 + ], + "type": "image", + "image_path": "8009a5ef648b8eb0755451f9d170aed35e6f8460236d5baebd74ddf9fe163948.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 163, + 159, + 173 + ], + "lines": [ + { + "bbox": [ + 113, + 163, + 159, + 173 + ], + "spans": [ + { + "bbox": [ + 113, + 163, + 159, + 173 + ], + "type": "text", + "content": "(a) Airplane" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 214, + 69, + 375, + 159 + ], + "blocks": [ + { + "bbox": [ + 214, + 69, + 375, + 159 + ], + "lines": [ + { + "bbox": [ + 214, + 69, + 375, + 159 + ], + "spans": [ + { + "bbox": [ + 214, + 69, + 375, + 159 + ], + "type": "image", + "image_path": "82263e7e086d08818f9f6e6e8d6709972c4a468e9cb82c856ff8a9dc4b62c7e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 163, + 314, + 172 + ], + "lines": [ + { + "bbox": [ + 280, + 163, + 314, + 172 + ], + "spans": [ + { + "bbox": [ + 280, + 163, + 314, + 172 + ], + "type": "text", + "content": "(b) Chair" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 376, + 69, + 539, + 160 + ], + "blocks": [ + { + "bbox": [ + 376, + 69, + 539, + 160 + ], + "lines": [ + { + "bbox": [ + 376, + 69, + 539, + 160 + ], + "spans": [ + { + "bbox": [ + 376, + 69, + 539, + 160 + ], + "type": "image", + "image_path": "0b9120aa39eab04542649825872d9840f49d658450f75d08f536bfa3f5e388fb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 163, + 471, + 172 + ], + "lines": [ + { + "bbox": [ + 444, + 163, + 471, + 172 + ], + "spans": [ + { + "bbox": [ + 444, + 163, + 471, + 172 + ], + "type": "text", + "content": "(c) Car" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 209, + 545, + 352 + ], + "blocks": [ + { + "bbox": [ + 46, + 180, + 547, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 180, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 547, + 205 + ], + "type": "text", + "content": "Figure 9. The changes of ConvNet importance for each category over timesteps with the channel-wise time mask setting. The thick curve is the result of Gaussian smoothing." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 209, + 545, + 352 + ], + "lines": [ + { + "bbox": [ + 50, + 209, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 50, + 209, + 545, + 352 + ], + "type": "image", + "image_path": "8ad4e30cfcb2378ba7eac682e165aa3026c9cf196214f26bb2d995d07b2eb7db.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 356, + 521, + 369 + ], + "lines": [ + { + "bbox": [ + 67, + 356, + 521, + 369 + ], + "spans": [ + { + "bbox": [ + 67, + 356, + 521, + 369 + ], + "type": "text", + "content": "Figure 10. High quality and diverse 3D point clouds generated from our TIGER model trained on 55 ShapeNetv2 categories." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 385, + 193, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 193, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 193, + 397 + ], + "type": "text", + "content": "long-range relationships in the data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 397, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 287, + 589 + ], + "type": "text", + "content": "Weight changing over timesteps. To validate our motivation of reweighing the local and global features from ConvNet and Transformer at different time steps, we visualize the changes in weights. For channel-wise time mask setting, we use the ratio of the weight of the local feature over the weight of the global feature to measure the importance of ConvNet and Transformer at each timestep. Fig. 9 indicates that the importance of ConvNet for each category increases nearly monotonically, implying that the diffusion process of 3D point cloud generation benefits from global features at the early stages of diffusion and requires the aggregation of more local features at later stages. Moreover, we observe that the curves of airplanes and chairs change more drastically than that of cars, which we hypothesize is due to the complicated shapes of airplanes and chairs, resulting in a more complex diffusion process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 589, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 287, + 649 + ], + "type": "text", + "content": "Unconditional generation of multiple classes. To further demonstrate the generalizability of our model, we train a universal TIGER model on all 55 categories of ShapeNetv2 [6]. Fig. 10 shows that our universal model can produce shapes with high quality and diversity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 657, + 125, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 125, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 125, + 670 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "In this work, we propose a novel time-varying denoising model with a latent point transformer backbone. Experimentally, we show how the time mask optimally aggre" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 385, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 469 + ], + "type": "text", + "content": "gates local and global features at different timesteps, which quantitatively reaches SoTA performance. In accordance with our hypothesis, Transformer is more important at early diffusing stages and ConvNet has a larger weight at later timesteps in the 3D point cloud generation task. Future work will be focused on exploring how our time-varying denoising model would behave in other diffusion-based tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 483, + 546, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 546, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 546, + 604 + ], + "type": "text", + "content": "Limitations. Although we generate high-quality and natural samples, we cannot control the category of the generated shape. One potential solution is to encode the category into latent space and train the diffusion model conditioned on the latent code. We also acknowledge that our model needs to get the local feature and global feature by two subnetworks, which is not quite efficient. In this work, we gain efficiency with the model converging with fewer epochs. But future works can increase the backbone efficiency by proposing time-varying properties with only one network." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Potential Negative Impacts. While it has the potential to speed up progress in both areas, it also has the risk of eliminating the jobs of game and furniture designers due to automation. However, our model should have a mostly positive impact, because the most likely use of our method would be to assist designers not completely replace their jobs. It is still necessary to select suitable shapes among many generated samples." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9469" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 710 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, 2018. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 155 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 155 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 155 + ], + "type": "text", + "content": "[2] Vishal Asnani, John Collomosse, Tu Bui, Xiaoming Liu, and Shruti Agarwal. ProMark: Proactive diffusion watermarking for causal attribution. In CVPR, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 156, + 286, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 156, + 286, + 176 + ], + "spans": [ + { + "bbox": [ + 53, + 156, + 286, + 176 + ], + "type": "text", + "content": "[3] Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450, 2016. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 176, + 286, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 176, + 286, + 218 + ], + "spans": [ + { + "bbox": [ + 53, + 176, + 286, + 218 + ], + "type": "text", + "content": "[4] Fan Bao, Kun Xu, Chongxuan Li, Lanqing Hong, Jun Zhu, and Bo Zhang. Variational (gradient) estimate of the score function in energy-based latent variable models. In ICML, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 219, + 286, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 219, + 286, + 261 + ], + "spans": [ + { + "bbox": [ + 53, + 219, + 286, + 261 + ], + "type": "text", + "content": "[5] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snavely, and Bharath Hariharan. Learning gradient fields for shape generation. In ECCV, 2020. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 261, + 286, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 286, + 316 + ], + "type": "text", + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 316, + 286, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 316, + 286, + 347 + ], + "spans": [ + { + "bbox": [ + 53, + 316, + 286, + 347 + ], + "type": "text", + "content": "[7] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. NeurIPS, 31, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 348, + 286, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 348, + 286, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 348, + 286, + 379 + ], + "type": "text", + "content": "[8] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusiondet: Diffusion model for object detection. arXiv preprint arXiv:2211.09788, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 286, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 286, + 400 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 286, + 400 + ], + "type": "text", + "content": "[9] Tamal K Dey and James Hudson. Pmr: Point to mesh rendering, a feature-based approach. In VIS, 2002. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 400, + 286, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 400, + 286, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 400, + 286, + 465 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 465, + 286, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 286, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 286, + 497 + ], + "type": "text", + "content": "[11] Yuval Eldar, Michael Lindenbaum, Moshe Porat, and Yehoshua Y Zeevi. The farthest point strategy for progressive image sampling. In ICPR, 1994. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 498, + 286, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 286, + 528 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 286, + 528 + ], + "type": "text", + "content": "[12] Matheus Gadelha, Rui Wang, and Subhransu Maji. Multiresolution tree networks for 3d point cloud processing. In ECCV, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 529, + 286, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 529, + 286, + 571 + ], + "spans": [ + { + "bbox": [ + 48, + 529, + 286, + 571 + ], + "type": "text", + "content": "[13] Will Grathwohl, Ricky TQ Chen, Jesse Bettencourt, Ilya Sutskever, and David Duvenaud. Ffjord: Free-form continuous dynamics for scalable reversible generative models. arXiv preprint arXiv:1810.01367, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 571, + 286, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 286, + 613 + ], + "type": "text", + "content": "[14] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3D surface generation. In CVPR, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 614, + 286, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 286, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 286, + 646 + ], + "type": "text", + "content": "[15] Jia Guo, Xuanxia Yao, Mengyu Shen, Jiafei Wang, and Wanyou Liao. A deep learning network for point cloud of medicine structure. In ITME, 2018. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 647, + 286, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 688 + ], + "type": "text", + "content": "[16] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 689, + 286, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 689, + 286, + 710 + ], + "spans": [ + { + "bbox": [ + 48, + 689, + 286, + 710 + ], + "type": "text", + "content": "[17] Kilian Konstantin Haefeli, Karolis Martinkus, Nathanael Perraudin, and Roger Wattenhofer. Diffusion models for" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 711 + ], + "type": "list", + "angle": 0, + "index": 40, + "blocks": [ + { + "bbox": [ + 327, + 74, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 74, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 74, + 545, + 94 + ], + "type": "text", + "content": "graphs benefit from discrete state spaces. arXiv preprint arXiv:2210.01549, 2022.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 95, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 115 + ], + "type": "text", + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 116, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 116, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 308, + 116, + 545, + 137 + ], + "type": "text", + "content": "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 137, + 545, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 137, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 308, + 137, + 545, + 168 + ], + "type": "text", + "content": "[20] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 168, + 545, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 168, + 545, + 199 + ], + "spans": [ + { + "bbox": [ + 308, + 168, + 545, + 199 + ], + "type": "text", + "content": "[21] Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. In CVPR, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 200, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 200, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 308, + 200, + 545, + 232 + ], + "type": "text", + "content": "[22] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 2005. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "text", + "content": "[23] Hyeongju Kim, Hyeonseung Lee, Woo Hyun Kang, Joun Yeop Lee, and Nam Soo Kim. Softflow: Probabilistic framework for normalizing flow on manifolds. In NeurIPS, 2020. 1, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 274, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 545, + 307 + ], + "type": "text", + "content": "[24] Jinwoo Kim, Jaehoon Yoo, Juho Lee, and Seunghoon Hong. Setvae: Learning hierarchical composition for generative modeling of set-structured data. In CVPR, 2021. 1, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 307, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 307, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 307, + 545, + 338 + ], + "type": "text", + "content": "[25] Minchul Kim, Feng Liu, Anil Jain, and Xiaoming Liu. DCFace: Synthetic face generation with dual condition diffusion model. In CVPR, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 339, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 339, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 339, + 545, + 370 + ], + "type": "text", + "content": "[26] Minchul Kim, Yiyang Su, Feng Liu, Anil Jain, and Xiaoming Liu. Keypoint relative position encoding for face recognition. In CVPR, 2024. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 370, + 545, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 370, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 308, + 370, + 545, + 402 + ], + "type": "text", + "content": "[27] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. In ECCV, 2020. 1, 2, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 403, + 545, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 403, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 308, + 403, + 545, + 434 + ], + "type": "text", + "content": "[28] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 434, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 455 + ], + "type": "text", + "content": "[29] Manoj Kumar, Mostafa Dehghani, and Neil Houlsby. Dual patchnorm. arXiv preprint arXiv:2302.01327, 2023. 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 456, + 545, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 486 + ], + "type": "text", + "content": "[30] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. Pointvoxel CNN for efficient 3d deep learning. In NeurIPS, 2019. 2, 3, 4, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 487, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 487, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 487, + 545, + 529 + ], + "type": "text", + "content": "[31] Yunfei Long, Daniel Morris, Xiaoming Liu, Marcos Paul Gerardo Castro, Punarjay Chakravarty, and Praveen Narayanan. Radar-camera pixel depth association for depth completion. In CVPR, 2021. 1" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 529, + 545, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 529, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 308, + 529, + 545, + 551 + ], + "type": "text", + "content": "[32] David Lopez-Paz and Maxime Oquab. Revisiting classifier two-sample tests. arXiv preprint arXiv:1610.06545, 2016. 6" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 551, + 545, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 551, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 308, + 551, + 545, + 594 + ], + "type": "text", + "content": "[33] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927, 2022. 3" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 308, + 594, + 545, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 594, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 308, + 594, + 545, + 614 + ], + "type": "text", + "content": "[34] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "type": "text", + "content": "[35] Peiyuan Ni, Wenguang Zhang, Xiaoxiao Zhu, and Qixin Cao. Pointnet++ grasping: Learning an end-to-end spatial grasp generation algorithm from sparse point clouds. In ICRA, 2020. 1" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 308, + 657, + 545, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 657, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 308, + 657, + 545, + 700 + ], + "type": "text", + "content": "[36] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 1" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 308, + 700, + 545, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 700, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 308, + 700, + 545, + 711 + ], + "type": "text", + "content": "[37] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh," + } + ] + } + ], + "index": 39 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9470" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 690 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "text", + "content": "Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 148 + ], + "type": "text", + "content": "[38] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 148, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 148, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 48, + 148, + 287, + 179 + ], + "type": "text", + "content": "[39] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 180, + 286, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 180, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 48, + 180, + 286, + 201 + ], + "type": "text", + "content": "[40] Prajit Ramachandran, Barret Zoph, and Quoc V. Le. Searching for activation functions. In ICLRW, 2018. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 201, + 286, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 201, + 286, + 233 + ], + "spans": [ + { + "bbox": [ + 48, + 201, + 286, + 233 + ], + "type": "text", + "content": "[41] Zhiyuan Ren, Zhihong Pan, Xin Zhou, and Le Kang. Diffusion motion: Generate text-guided 3d human motion by diffusion model. arXiv preprint arXiv:2210.12315, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 234, + 286, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 234, + 286, + 264 + ], + "spans": [ + { + "bbox": [ + 48, + 234, + 286, + 264 + ], + "type": "text", + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 265, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 307 + ], + "type": "text", + "content": "[43] Nikolay Savinov, Junyoung Chung, Mikolaj Binkowski, Erich Elsen, and Aaron van den Oord. Step-unrolled denoising autoencoders for text generation. arXiv preprint arXiv:2112.06749, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 308, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 339 + ], + "type": "text", + "content": "[44] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3D object proposal generation and detection from point cloud. In CVPR, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 340, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 340, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 340, + 287, + 360 + ], + "type": "text", + "content": "[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 361, + 287, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 391 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 391 + ], + "type": "text", + "content": "[46] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 392, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 392, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 48, + 392, + 287, + 423 + ], + "type": "text", + "content": "[47] Haoru Tan, Sitong Wu, and Jimin Pi. Semantic diffusion network for semantic segmentation. arXiv preprint arXiv:2302.02057, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 425, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 425, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 425, + 287, + 466 + ], + "type": "text", + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 468, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 498 + ], + "type": "text", + "content": "[49] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 499, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 518 + ], + "type": "text", + "content": "[50] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 520, + 287, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 562 + ], + "type": "text", + "content": "[51] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In ICCV, 2019. 1, 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 563, + 287, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 563, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 48, + 563, + 287, + 594 + ], + "type": "text", + "content": "[52] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-net: Point cloud upsampling network. In CVPR, 2018. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 594, + 287, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 594, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 48, + 594, + 287, + 636 + ], + "type": "text", + "content": "[53] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. LION: Latent point diffusion models for 3d shape generation. In NeurIPS, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 638, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 638, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 638, + 287, + 658 + ], + "type": "text", + "content": "[54] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 3, 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "text", + "content": "[55] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In ICCV, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9471" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/03af8456-92c0-4a22-a6d9-94775506b39a_content_list.json b/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/03af8456-92c0-4a22-a6d9-94775506b39a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6c79a4c81a419680c7e21978513ce40f3dc2e256 --- /dev/null +++ b/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/03af8456-92c0-4a22-a6d9-94775506b39a_content_list.json @@ -0,0 +1,1779 @@ +[ + { + "type": "text", + "text": "TIM: A Time Interval Machine for Audio-Visual Action Recognition", + "text_level": 1, + "bbox": [ + 140, + 130, + 828, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jacob Chalk $^{1*}$", + "bbox": [ + 225, + 179, + 339, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaesung Huh $^{2*}$", + "bbox": [ + 398, + 180, + 517, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Andrew Zisserman", + "bbox": [ + 316, + 199, + 475, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Evangelos Kazakos3", + "bbox": [ + 578, + 180, + 743, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dima Damen", + "bbox": [ + 537, + 199, + 651, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ University of Bristol", + "bbox": [ + 107, + 215, + 281, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ VGG, University of Oxford", + "bbox": [ + 300, + 215, + 529, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Czech Technical University in Prague", + "bbox": [ + 547, + 215, + 861, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diverse actions give rise to rich audio-visual signals in long videos. Recent works showcase that the two modalities of audio and video exhibit different temporal extents of events and distinct labels. We address the interplay between the two modalities in long videos by explicitly modelling the temporal extents of audio and visual events. We propose the Time Interval Machine (TIM) where a modality-specific time interval poses as a query to a transformer encoder that ingests a long video input. The encoder then attends to the specified interval, as well as the surrounding context in both modalities, in order to recognise the ongoing action.", + "bbox": [ + 75, + 301, + 470, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We test TIM on three long audio-visual video datasets: EPIC-KITCHENS, Perception Test, and AVE, reporting state-of-the-art (SOTA) for recognition. On EPIC-KITCHENS, we beat previous SOTA that utilises LLMs and significantly larger pre-training by $2.9\\%$ top-1 action recognition accuracy. Additionally, we show that TIM can be adapted for action detection, using dense multi-scale interval queries, outperforming SOTA on EPIC-KITCHENS-100 for most metrics, and showing strong performance on the Perception Test. Our ablations show the critical role of integrating the two modalities and modelling their time intervals in achieving this performance. Code and models at: https://github.com/JacobChalk/TIM.", + "bbox": [ + 73, + 468, + 468, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 696, + 207, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Long videos exhibit a quick succession of auditory and visual events. The latest attempts to annotate events in these modalities separately [15, 30], showcase that both the temporal extents and class labels differ between the two. However, these events still remain correlated - identifying temporally close events in both modalities can improve recognition of actions in both visual and audio.", + "bbox": [ + 75, + 720, + 468, + 827 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Furthermore, most methods to date typically only utilise the exact temporal extent of an action; a precise, trimmed clip of the action is fed into a convolutional [5, 8, 39] or", + "bbox": [ + 75, + 829, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/38991d2abc51082a7ec61d91e08a9075d5fcefb1634a1e45e00d8fdb8a0d655f.jpg", + "image_caption": [ + "Figure 1. Time Interval Machine (TIM): Top: Given a visual and auditory stream input, the ongoing action in a particular time interval is determined by a query specifying the start and end time of the interval, along with the modality of interest. Bottom: TIM can query for visual (e.g. 'Rinse Sponge') and auditory (e.g. 'Water') action classes, as well as distinguish between overlapping actions within the same modality ('Glass Collision' and 'Scrub / Scrape')." + ], + "image_footnote": [], + "bbox": [ + 500, + 266, + 893, + 419 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "transformer-based [1, 11, 22] backbone, which predicts the action taking place. Even when the surrounding context is utilised to improve action recognition [18, 31, 42], again, this context is supplied in the form of exact clips of neighbouring actions, rather than the untrimmed long input video.", + "bbox": [ + 496, + 551, + 890, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose an approach that encodes multiple events that occur in both visual and auditory streams of a long video input. We achieve this by elevating time intervals to first-class citizens, utilising them to specify a query within an accompanying modality. We term this mechanism a Time Interval Machine (TIM). It is able to receive a long video input, and output the actions that occur within the queried intervals of the queried modalities.", + "bbox": [ + 496, + 628, + 890, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Consider the example in Figure 1. The input contains the sound of water running while a sponge is being rinsed, which is then used to wipe a surface. These distinct events may vary significantly in duration and may be more prominent in the audio or visual modality. Despite differences between these events, there are likely many correlations between them and the surrounding context, which may be beneficial to recognising a given event (e.g. the sound of water is relevant to rinsing a sponge, providing useful information for recognising the visual action). TIM is able", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal technical contribution.", + "bbox": [ + 101, + 887, + 263, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18153", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to exploit this by accessing the context within both modalities, including the background when no events occur. It can then distinguish between different, potentially overlapping, events within the same input by querying the time interval of a particular event within a given modality.", + "bbox": [ + 75, + 90, + 467, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We test TIM on three challenging audio-visual recognition datasets consisting of long videos: EPIC-KITCHENS [7], which recently offered distinct audio annotations through EPIC-SOUNDS [15], the Perception Test [30], and AVE [36]. We show that TIM can effectively learn both visual and auditory classes within a long input, outperforming the current SOTA top-1 accuracy on EPIC-KITCHENS by $2.9\\%$ and $1.4\\%$ on EPIC-SOUNDS, despite competing methods for the former using far larger pre-training datasets, large language models or higher resolution input. We also outperform models pre-trained with public datasets on AVE by $0.6\\%$ and improve over a strong baseline on the Perception Test in visual and audio action recognition by $9.9\\%$ and $3.2\\%$ respectively.", + "bbox": [ + 75, + 167, + 467, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, we adapt TIM to action detection, through fixed multi-scale dense querying with an added interval regression loss. We report strong detection results on EPIC-KITCHENS and the Perception Test, outperforming Action Former [49] by 1.6 avg mAP and 4.3 respectively.", + "bbox": [ + 76, + 378, + 467, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarised as: (i) we propose the TIM query mechanism for attending to modality-specific intervals in long videos. (ii) we efficiently train TIM to encode/query multiple audio-visual actions using time intervals. (iii) we showcase the value of TIM for both visual and auditory action recognition, and adapt it for detection with an added interval regression loss. (iv) we achieve new SOTA in both video and multi-modal recognition on multiple datasets.", + "bbox": [ + 75, + 454, + 467, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 603, + 225, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Audio-visual action recognition. A number of works have employed audio and visual modalities for action recognition [10, 17, 18, 26, 41, 44]. Some introduce new architectures to effectively fuse modalities [17, 18, 26, 44]; others propose unique training techniques to solve problems occurring while training multi-modal models, such as Gradient Blending [41], to tackle overfitting at different speeds for each modality, or contrastive learning for cross-modal discrimination [25]. However, these works use the same set of semantic and temporal labels for both modalities. Recent works have shown that both the temporal intervals and semantics of events differ between modalities [15, 30]. [37] temporally annotates visual and auditory events independently, although they share the same set of labels. In this work, we train with distinct labels for each modality to leverage discriminative audio and visual actions.", + "bbox": [ + 75, + 628, + 467, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Leveraging temporal context. Several works have considered incorporating temporal context [18, 27, 42, 43, 48], a", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "direction orthogonal to employing multiple modalities and particularly useful in untrimmed videos. An auto-regressive LSTM-based encoder-decoder is proposed in [27] for action sequence classification, effectively leveraging past action context to predict the current action. The Temporal Query Network [48] uses learnable query vectors that correspond to specific attributes of a long video, allowing the model to attend to the aspects of the video and its surrounding context to produce a response for each attribute. [42] proposes to enhance the representation of the action by aggregating temporal context from neighbouring action clips using a Long-Term Feature Bank along with an attention mechanism. [43] crafts a more sophisticated memory bank by storing keys and values of all the intermediate layers of a transformer to aggregate the past context. Lastly, [18] exploits multi-modal temporal context from surrounding actions using vision, audio, and language.", + "bbox": [ + 496, + 90, + 890, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[18, 42, 43] are the closest to our approach, in that the common goal is to enrich the representation of the action of interest using surrounding context from the untrimmed video, rather than neighbouring clips. Nevertheless, [42, 43] are single modality models, recognising visual actions solely. [18] assumes the temporal extents of all actions are known, including for the test set, which is restrictive.", + "bbox": [ + 496, + 347, + 890, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Queries in visual models. Learning visual queries with Transformer architectures has gained recent attention [4, 14, 16, 23, 48]. Commonly, approaches employ a set of learnable vectors that are used to inquire about the presence of a concept in the input. For example, in [4, 23] the learnable queries correspond to different objects, whereas in [14] they are used for multi-task learning and each learnable query corresponds to a different task. [16] has incorporated learnable queries for adapting a pre-trained model while keeping the rest of its parameters frozen. Closest to our motivation is [48], where the queries correspond to events and their attributes for fine-grained action recognition in videos. The authors note that the queries also have the role of temporally localising the events in untrimmed videos.", + "bbox": [ + 496, + 454, + 890, + 662 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different from [48] and other works, our queries are primarily temporal with no semantic interpretation and are applied to multiple modalities. Importantly, since time is continuous, we cannot use a predefined set of queries. Instead, we employ an MLP architecture to encode time, in a form akin to a universal clock. We present our approach next.", + "bbox": [ + 496, + 665, + 890, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Time Interval Machine", + "text_level": 1, + "bbox": [ + 500, + 768, + 718, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we describe the Time Interval Machine (TIM), a multi-modal transformer encoder architecture where all inputs, both features and queries, are encoded with their associated time intervals. A time interval incorporates the duration and position of each audio and visual feature and is also used to query the network for any action occurring within the given time interval.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18154", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9c05bfa74f8ae34037b09debdaf43548f11dbe523908bfd63294c4cd34f20c93.jpg", + "image_caption": [ + "Figure 2. Overview of the Time Interval Machine (TIM). The model ingests a sequence of audio and visual features from a video, with each feature time-stamped by the temporal interval it spans, and encoded with its modality. To infer the action occurring over a temporal interval (a visual or audio event) a query is formed specifying the interval and modality of interest." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 893, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The architecture of TIM is illustrated in Figure 2. It ingests a large video input, represented as a sequence of audio and visual features, and outputs the ongoing auditory or visual action label for the provided query time intervals.", + "bbox": [ + 75, + 388, + 470, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model architecture", + "text_level": 1, + "bbox": [ + 76, + 459, + 261, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input. The input to TIM is a long crop of the untrimmed video, represented by extracted features. When considering two modality inputs, such as video and audio, each modality is embedded separately as follows: for each modality $m$ , let $\\mathbf{X}^m = [X_1^m, \\dots, X_{N^m}^m]$ be $N^m$ temporally-ordered feature representations of the input video, obtained from a pre-trained feature extractor $S^m(\\cdot)$ . We feed the features through modality-specific embedding layers $g^m(\\cdot)$ , projecting them to a lower, common dimension $D$ across all modalities. The embedded features are then tagged with modality encodings and time interval encodings, forming the input to the transformer encoder. We now detail how we encode the time intervals.", + "bbox": [ + 75, + 482, + 468, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Encoding Time Intervals. In this work, we introduce a new type of learnt query network, the Time Interval MLP, which produces a single $D$ -dimensional vector representing a given time interval. This network is used within TIM to encode the time intervals of the input features and the time interval we wish to query, and later classify. Figure 3 illustrates the concept of this network.", + "bbox": [ + 75, + 681, + 468, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Time Interval MLP $I(\\cdot): \\mathbb{R}^2 \\to \\mathbb{R}^D$ receives a time interval, represented by start and end time, and produces a single $D$ -dimensional encoding. Note that this is distinct from encoding the start and end times separately. Specifically, let $t_s$ and $t_e$ be the start and end time of an interval of interest, normalised by the length of the long video input.", + "bbox": [ + 75, + 786, + 470, + 878 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0bce42347eaca559eb4fe3ccc488ad575a1b99c5d3e336b01c07d63c6037b55a.jpg", + "image_caption": [ + "Figure 3. Illustration of the Time Interval MLP $I(\\cdot)$ . It inputs the two-dimensional vector, start and end times of an interval, and produces a single vector, which can be concatenated along the channel dimension to either input features or [CLS] tokens. The figure shows three time interval inputs and three corresponding outputs. Note that in practice, time intervals are ingested simultaneously." + ], + "image_footnote": [], + "bbox": [ + 526, + 387, + 870, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$I(\\cdot)$ receives the interval $\\tilde{t} = [t_s, t_e]$ as input, and outputs a $D$ -dimensional vector encoding of that interval. This vector encodes both the relative position of the time interval within the input, as well as its duration. This vector then acts as a query for the model concerning the action taking place within the interval. Furthermore, each feature $\\{X_i^m\\}$ spans a certain time interval within the input. Thus, it is important to also encode the time intervals of the features.", + "bbox": [ + 496, + 604, + 890, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, the Time Interval MLP acts as an universal clock, which encodes the temporal extent of features, from any modality, within the input. Note that it is critical that the same Time Interval MLP is used for encoding all time intervals of the input features and queries across both modalities to accurately encode universal time. It is also important to note that Time interval MLP can cover continuous time intervals, whereas traditional positional encoding only covers a fixed set of positions of the input features. The Time Interval MLP is trained end-to-end along with the transformer.", + "bbox": [ + 496, + 729, + 890, + 880 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer Feature Inputs. Let $\\tilde{\\mathbf{t}}^m = [\\tilde{t}_1^m, \\dots, \\tilde{t}_{N^m}^m]$", + "bbox": [ + 500, + 883, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1Note that the number of features can differ between modalities", + "bbox": [ + 94, + 886, + 433, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18155", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "be the corresponding time intervals of the video's features $\\mathbf{X}^m$ from the modality $m$ . We inject the encoded time interval $I(\\tilde{\\mathbf{t}}^m)$ into the embedded features via channelwise concatenation. A learnable modality-specific encoding $e^m \\in \\mathbb{R}^{2D}$ is then summed to the temporally-encoded features to discriminate between each modality. In summary, the feature inputs $\\mathbf{E}^m$ for TIM are computed by,", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {i} ^ {m} = \\left[ g ^ {m} \\left(X _ {i} ^ {m}\\right), I \\left(\\tilde {t} _ {i} ^ {m}\\right) \\right] + e ^ {m} \\quad \\forall i \\in [ 1, \\dots , N ^ {m} ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 207, + 468, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[\\cdot ,\\cdot ]$ indicates concatenation.", + "bbox": [ + 76, + 234, + 313, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer Query Inputs. To query for an action within an interval of interest, we adopt a standard approach of appending a learnable classification token to the input sequence, $\\mathsf{CLS}^m$ . If $\\tilde{t}_Q$ is an interval of interest, we concatenate the time interval representation $I(\\tilde{t}_Q)$ to this classification token along the channel dimension, which acts as a query for the network in order to predict the corresponding action happening within $\\tilde{t}_Q$ . We also add the modality-specific encoding $e^m$ to each classification token, as a flag to distinguish between which modality we are querying. The encoded $[\\mathsf{CLS}]^m$ tokens can be more formally defined as:", + "bbox": [ + 75, + 252, + 468, + 419 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n[ \\mathrm {C L S} ] ^ {m} = [ \\mathrm {C L S} ^ {m}, I (\\tilde {t} _ {Q}) ] + e ^ {m} (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 429, + 468, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, we add a classification token for each action within the input video, resulting in multiple [CLS] tokens across both modalities.", + "bbox": [ + 75, + 458, + 468, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer Encoder. We use a transformer encoder to perform self-attention on the input sequence to aggregate relevant temporal context and cross-modal relations.", + "bbox": [ + 75, + 506, + 468, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We form the transformer input sequence with the encoded feature inputs $\\mathbf{E}^m$ and one or more classification tokens $[\\mathrm{CLS}]^m$ , representing each time interval query, and feed these into the encoder. Note that we recognise all actions from any modality simultaneously by appending multiple $\\mathrm{CLS}^m$ tokens to the input. The transformer output representation of $[\\mathrm{CLS}]^m$ , namely $Z_{\\mathrm{CLS}}^m$ , is then passed to the corresponding linear classifier to predict the action labels.", + "bbox": [ + 75, + 551, + 468, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Importantly, we use an attention mask to prevent queries from attending to one another, and we similarly prevent input features from attending to queries. This ensures each query is recognised without the privileged knowledge of any other query, or action boundary, during inference.", + "bbox": [ + 75, + 672, + 468, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Training and Testing in TIM", + "text_level": 1, + "bbox": [ + 76, + 756, + 333, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train TIM, we consider all long segments of $W$ seconds and stride $H_{w}$ across the entirety of the untrimmed videos. We randomly select batches from these. For each window, we query all annotated audio and visual actions that overlap with the window by more than $\\delta = 0.2$ seconds.", + "bbox": [ + 75, + 779, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All queries in the window are encoded and concatenated to separate CLS tokens. To classify queries, let $h_{\\mathrm{CLS}}^m(\\cdot)$ be a linear classifier for modality $m$ , and let $\\hat{y}_{\\mathrm{CLS}}^m = h^m(Z_{\\mathrm{CLS}}^m)$", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "be the predicted action of the output representation $Z_{\\mathrm{CLS}}^{m}$ . We train TIM by using a cross-entropy classification loss $CE(\\cdot)$ on the ground truth $y_{\\mathrm{CLS}}^{m}$ by:", + "bbox": [ + 498, + 90, + 890, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL ^ {m} = \\frac {1}{N _ {Q}} \\sum^ {N _ {Q}} C E \\left(\\hat {y} _ {\\mathrm {C L S}} ^ {m}, y _ {\\mathrm {C L S}} ^ {m}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 148, + 890, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N_{Q}$ is the number of queries within the batch.", + "bbox": [ + 498, + 196, + 846, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Distance Loss. In addition to the standard classification loss, we introduce a Temporal Distance (TD) loss as an auxiliary loss for training TIM. Inspired by [21], where relative patch positions in token embeddings are learnt using self-supervision, we similarly train the network to take two transformer outputs and predict the elapsed time between their corresponding time intervals.", + "bbox": [ + 496, + 214, + 890, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let the $\\mathbf{Z}_{1:\\sum_mN^m}$ be the transformer outputs of the features from all modalities. We randomly sample a set of feature pairs $\\mathbb{B} \\subset \\mathbf{Z}_{1:\\sum_mN^m}$ from these outputs, concatenate along the channel dimension and feed them to the temporal distance regression head $h_{\\tilde{t}}(\\cdot):\\mathbb{R}^{4D} \\to \\mathbb{R}^1$ to predict the time interval difference between each pair. Note that feature pairs can be sampled both within and across modalities. In our case, we sample across modalities by pairing one visual feature with another audio feature. This helps the model to learn the temporal relations between modalities.", + "bbox": [ + 496, + 319, + 892, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, a TD loss $L^{td}$ is computed as:", + "bbox": [ + 517, + 469, + 787, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL ^ {t d} = \\sum_ {\\left\\{Z _ {i}, Z _ {j} \\right\\} \\in \\mathbb {B}} \\left| h _ {\\tilde {i}} \\left(Z _ {i}, Z _ {j}\\right) - d _ {i j} \\right| \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 497, + 890, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_{ij}$ is the temporal distance between intervals $\\tilde{t}_i,\\tilde{t}_j$", + "bbox": [ + 498, + 542, + 885, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training objective and regime. For our final training loss, we sum the losses across modalities along with the TD loss:", + "bbox": [ + 498, + 561, + 890, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL ^ {t o t a l} = \\left(\\sum_ {m \\in \\mathbb {M}} \\lambda^ {m} L ^ {m}\\right) + \\lambda^ {t d} L ^ {t d} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 602, + 890, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{M}$ is a set of modalities, $\\lambda^m$ controls the strength of each modality's loss and $\\lambda^{td}$ is a hyperparameter that controls the strength of the TD loss.", + "bbox": [ + 496, + 654, + 890, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Test-Time Augmentation. We use test-time augmentation, as this generally increases prediction robustness and performance [29, 32]. In TIM, we use a sliding window across the untrimmed video, thus feeding the same interval query with varying contexts. We then aggregate predictions of the same interval query across windows to make the final prediction.", + "bbox": [ + 496, + 700, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Adapting for Detection", + "text_level": 1, + "bbox": [ + 500, + 801, + 712, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While primarily designed for recognition, we can adapt TIM for detection. The backbone remains largely unchanged from recognition, but there are two main differences. First, we construct dense multi-scale interval queries spanning the entirety of the video input at each scale. These", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18156", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are used as interval queries in both training and detection inference. The multi-scale intervals allow for detecting both long and short actions. Second, we introduce an additional interval regression head, which regresses the query interval to the action's exact temporal duration.", + "bbox": [ + 75, + 90, + 467, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we deem any query in the multi-scale pyramid that overlaps with a ground truth action by more than some IoU threshold as a positive query. In addition to classifying the query, we train a DIOU regression loss [52] to predict the exact interval of the action. Both classification and interval regression losses are trained jointly. We provide full details in the ArXiv appendix.", + "bbox": [ + 75, + 167, + 467, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 285, + 207, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section describes the datasets used to evaluate our model, implementation details and results along with a comparison to the state-of-the-art methods.", + "bbox": [ + 75, + 310, + 467, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 76, + 364, + 171, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "EPIC-KITCHENS-100 [7] is a large-scale video dataset, including 700 egocentric videos recording actions in kitchens. It consists of 89,977 segments of fine-grained actions. Inspired by prior works [11, 34, 35], we directly predict the action out of 3806 classes present in the train and validation set to avoid predicting invalid actions.", + "bbox": [ + 75, + 388, + 467, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "EPIC-SOUNDS [15] offers audio annotations capturing temporal extents and class labels within the audio stream of EPIC-KITCHENS-100. The annotations contain 78,366 labelled audio events. We combine the visual annotations from EPIC-KITCHENS with the audio annotations from EPIC-SOUNDS to train our audio-visual model. TIM can recognise actions from both datasets using a single model.", + "bbox": [ + 75, + 483, + 467, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AVE [36] contains 4,143 videos covering a range of real-life scenes and labelled with 27 categories, such as church bell, male speaking and dog barking. Each video is equally divided into 10 segments, each 1 second in length. We evaluate TIM on the supervised audio-visual event localisation task. Given a 1 second segment, we recognise the ongoing action out of the 27 categories plus a background class.", + "bbox": [ + 75, + 590, + 467, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Perception Test [30] is a recent multimodal video benchmark of 11,620 videos with an average length of 23 seconds, and provides both temporal action and sound annotations. There are 73,503 visual annotations spanning 63 classes, versus 137,128 sound annotations over 16 classes.", + "bbox": [ + 75, + 699, + 467, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 784, + 292, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architectural Details. Visual and audio embedding layers $g_{m}$ consist of a single 512-D feed-forward layer, followed by a GELU [13] activation and layer normalisation [2] are used to project the features to a common space. The Time Interval MLP $I$ consists of three linear layers with 512-D hidden dimension, followed by ReLU activations, with", + "bbox": [ + 75, + 809, + 467, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "layer normalisation after the output of the last linear layer. We include 512-D learnable [CLS] tokens: $[\\mathrm{CLS}]_{\\text {action }}^{m}$ for each query in each modality which become 1024-D after concatenation with the encoded time interval. They are then summed with 1024-D modality encodings; $e^{m}$ .", + "bbox": [ + 496, + 90, + 890, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The audio-visual transformer contains four encoder layers, each with 8 attention heads, GELU activations, and 1024-D keys, queries and values. A dropout rate of $p = 0.1$ is applied within the encoder layers. We also apply channelwise dropout with $p = 0.5$ directly to the raw input features, as well as to the encoded transformer input. The temporal distance head consists of two linear layers with a hidden dimension of 1024 and a third which outputs a single number corresponding to the elapsed time between each time interval. We include the architectural ablations on encoder layers and the temporal distance head in the ArXiv appendix.", + "bbox": [ + 496, + 167, + 890, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training / Validation Details. We train each model for 100 epochs, using AdamW [24] with a batch size of 64 and a weight decay of 1e-4. A linear learning rate warm-up is applied for the first two epochs, starting from 1e-6 to a target learning rate, and we use a cosine learning rate scheduler. We set the TD loss weight $\\lambda^{td}$ to 0.3. We pad the queries for each window in the batch to the maximum number of queries in a single window in each dataset. We provide implementation details per dataset in the ArXiv appendix.", + "bbox": [ + 496, + 335, + 890, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Results", + "text_level": 1, + "bbox": [ + 500, + 481, + 591, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare TIM with SOTA models for each dataset.", + "bbox": [ + 500, + 505, + 857, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "EPIC-KITCHENS / EPIC-SOUNDS Results. We train a single model on both visual and audio labels of the EPIC-KITCHENS videos, reporting results on both datasets.", + "bbox": [ + 500, + 522, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the visual features, we concatenate Omnivore [11] and VideoMAE-L [38] features along the channel dimension, forming 2048-D features. For the audio features, we use Auditory SlowFast [19], which generalises well across diverse audio domains [40]. For both modalities, we extract 1 second features every 0.2s. For training, we extract additional augmented feature sets - with RandAugment [6] for visual and SpecAugment [28] for audio features.", + "bbox": [ + 496, + 568, + 890, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 compares TIM with the SOTA models on EPIC-KITCHCHENS-100. We outperform M&M Mix [45] by $5.1\\%$ on verb, $0.9\\%$ on noun and $3.9\\%$ on action. Compared to our model, both MTV and M&M Mix are trained with an additional private dataset [33] which contains 194K hours of 70 million videos while we only use the open-source visual backbone pre-trained with public datasets. We also outperform LaViLa [51] and AVION [50] which leverage pre-trained LLMs to learn video representations.", + "bbox": [ + 496, + 688, + 890, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We note that we outperform all prior works, often without additional techniques that boost performance. For example, we use short-sided cropped $224 \\times 224$ images while [45] uses $420 \\times 420$ , which enlarges the spatial resolution of objects in the egocentric video, enabling better noun recog", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18157", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8710e2663bfa710d1f8fd539d1d0b9869fb1f3aa6b85b0031f3c04896ec4ab1e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelxpLLMVerbNounAction
Visual-only models
MFormer-HR [29]336p×67.058.544.5
MoViNet-A6 [20]320p×72.257.347.7
MeMViT [43]224p×71.460.348.4
Omnivore [11]224p×69.561.749.9
MTV [47]280p×69.963.950.5
LaViLa (TSF-L) [51]224p✓72.062.951.0
AVION (ViT-L) [50]224p✓73.065.454.4
TIM (ours)224p×76.266.456.4
Audio-visual models
TBN [17]224p×66.047.236.7
MBT [26]224p×64.858.043.4
MTCN [18]336p×70.762.149.6
M&M [45]420p×72.066.353.6
TIM (ours)224p×77.167.257.5
", + "bbox": [ + 81, + 87, + 467, + 303 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ceacd609df448fb4e644139e7cc5af867940a5b84a200e13de4032cf6b5e67cc.jpg", + "table_caption": [ + "Table 1. Comparisons to state-of-the-art recognition models on the EPIC-KITCHENS validation set. We report the top-1 accuracy for verb, noun and action (\\%). LLM: large language model is used during pre-training. xp: input resolution of $x \\times x$ ." + ], + "table_footnote": [], + "table_body": "
ModelSSAST [12]ASF [19]DiffSED [3]TIM (A)TIM (A+V)
Top-1 acc53.553.856.955.758.3
", + "bbox": [ + 81, + 386, + 467, + 411 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b400b1a94434cc4badaa7efb2775c433284816d39200d91bfeb1a2d740211671.jpg", + "table_caption": [ + "Table 2. Comparisons to state-of-the-art sound recognition models on EPIC-SOUNDS. We report the top-1 accuracy $(\\%)$ on Val. The performance of SSAST and ASF are from [15]." + ], + "table_footnote": [], + "table_body": "
ModelVAAverage Precision (AP)
Task@0.1@0.2@0.3@0.4@0.5Avg.
G-TAD [46]✓✗Verb12.111.09.48.16.59.4
Noun11.010.08.67.05.48.4
ActionFormer [49]✓✗Verb26.625.424.222.319.123.5
Noun25.224.122.720.517.021.9
ActionFormer - Our Features✓✗Verb29.628.826.924.421.626.3
Noun34.332.630.227.422.629.4
TIM✓✓Verb32.931.629.627.022.228.6
Noun36.434.832.128.722.731.0
", + "bbox": [ + 81, + 479, + 465, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "nition. We expect a further performance boost when implementing any of: higher resolution feature extractors, additional large-scale pre-training and the introduction of a LLM. We leave this as an avenue for future work.", + "bbox": [ + 76, + 652, + 468, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 compares TIM against prior results on EPIC-SOUNDS, where TIM outperforms SOTA by $1.4\\%$ .", + "bbox": [ + 76, + 714, + 467, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For detection, we show that TIM can produce competitive results when compared to models primarily designed for this task in Table 3. TIM adapted for detection outperforms ActionFormer [49] by $2.3\\mathrm{mAP}$ on verb and $1.6\\mathrm{mAP}$ on noun using the same set of features.", + "bbox": [ + 76, + 746, + 468, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "AVE Results. As this dataset contains joint audio-visual labels, we train TIM by duplicating the query, i.e. using a [CLS] for each modality, and combine their logits during training and inference. We use the pre-trained publicly available models from [36] for a fair comparison with other", + "bbox": [ + 76, + 824, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/802fdf91ec8fa1805c2cc1e3ed3ab82b4022e89c884b4de7b4d55cb26cf38835.jpg", + "table_caption": [ + "Table 3. Comparisons to state-of-the-art detection models on the EPIC-KITCHENS validation set. We report the average precision at IOU thresholds [0.1, 0.2, 0.3, 0.4, 0.5] as well as their average across all thresholds on verb, noun." + ], + "table_footnote": [], + "table_body": "
ModelPSP [53]CPSP [54]CSSNet [9]†TIM
VAAVAV*
Top-1 acc77.878.680.562.865.579.279.8
", + "bbox": [ + 504, + 87, + 890, + 128 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0c262f06788d41fb431998f018d27b94f6329b43fc3d9eec6a67e71176c98bdb.jpg", + "table_caption": [ + "Table 4. Top-1 event classification accuracy $(\\%)$ on AVE Test set. $\\dagger$ : no official code or public model provided to replicate results. We show the models trained only with publicly available datasets. $\\star$ : results with Omni+ASF features." + ], + "table_footnote": [], + "table_body": "
Perception Test Action
ModelMLP (V)MTCN [18](A+V)TIM (V)TIM (A+V)
Top-1 acc43.751.256.161.1
Perception Test Sound
ModelMLP (A)MTCN [18](A+V)TIM (A)TIM (A+V)
Top-1 acc50.652.954.856.1
", + "bbox": [ + 503, + 213, + 888, + 287 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d708fbdcf6c48bd7d7f7e529f21effa11823a5d4a218db77a09d93cb6f4e6c16.jpg", + "table_caption": [ + "Table 5. Comparisons to trained recognition baselines on the Perception Test validation split. We show both action and sound recognition and the benefit of including audio-visual in TIM for both challenges. V : visual and A : audio input features. MLP is the result by training an MLP classifier with the features directly." + ], + "table_footnote": [], + "table_body": "
ModelAverage Precision (AP)
@0.1@0.2@0.3@0.4@0.5Avg.
Perception Test Action
ActionFormer [49]27.827.625.223.020.024.5
TIM33.532.229.826.422.028.8
Perception Test Sound
ActionFormer [49]34.731.327.522.717.726.8
TIM37.533.127.922.817.227.7
", + "bbox": [ + 504, + 385, + 888, + 482 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 6. Comparisons to strong detection models on the Perception Test validation set for action and sound localisation. We report the average precision at IOU thresholds [0.1, 0.2, 0.3, 0.4, 0.5] as well as the average across all thresholds.", + "bbox": [ + 498, + 493, + 890, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "works. We also apply AVGA [36] to spatial visual features from VGG-19 before feeding them to the transformer.", + "bbox": [ + 498, + 563, + 890, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4 shows our results on the AVE dataset. Combining audio and video significantly improves the performance on TIM. The results from [9] perform best but could not be replicated. We also report TIM using the Omnivore visual features and Auditory Slowfast features used for EPIC-KITCHENS, which achieves a $0.6\\%$ boost in performance.", + "bbox": [ + 498, + 595, + 890, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Perception Test Results. We use the same backbone for the Omnivore features and Auditory Slowfast features and train a single model using both the visual and audio labels. Table 5 compares the results on the newly introduced Perception Test. We train an MLP classifier, with two linear layers and a ReLU activation, on the features directly as a baseline. We also evaluate on an audio-visual model that does use context with MTCN. Compared to these methods TIM clearly shows significant improvements. Results are improved over MTCN by $9.9\\%$ and $3.2\\%$ on visual and audio recognition tasks respectively. We also provide detection results in Table 6. TIM improves over ActionFormer [49] by 3.3 average mAP on visual actions and by 0.9 average mAP on sound, when using the same features.", + "bbox": [ + 496, + 688, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18158", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/99ebabf585cd33faad664ece1324409a3e3f93644d45fcc2925113ac8b8ac581.jpg", + "image_caption": [ + "Figure 4. Qualitative results for all datasets. PRED: Prediction by TIM, TIQ: Time Interval Queries, V/AGT: Visual/Audio Ground Truth." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 470, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd10edd49098b623f5b0f85b4effc0bf5bd1ec03e71ced203c1475089241fb67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 85, + 890, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-Modality in TIM. When referring to our previous results, we see that including the additional modality provides a performance boost in all cases, highlighting TIM's ability to utilise and distinguish between different modalities. For example, on EPIC-KITCHENS-100, including audio improves visual action accuracy by $0.9\\%$ . For EPICSOUNDS, the visual modality further improves accuracy by $2.6\\%$ . In the Perception Test, including the audio modality improves visual recognition by $5.0\\%$ , and visual increases sound recognition by $1.3\\%$ . Finally, for AVE, we see a significant improvement, where an audio-visual model increases accuracy by $13.7\\%$ from audio-only.", + "bbox": [ + 75, + 412, + 468, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. We present qualitative results in Figure 4. We see that in EPIC-KITCHENS, TIM can competently recognise actions across the two modalities including overlapping queries. Furthermore, we see consecutive actions are correctly recognised with varying interval lengths, such as the 'open / close' audio actions between 0.2s and 4.5s. For AVE, TIM is able to distinguish between the background and a 'barking' audio-visual event based on the time interval query. For the Perception Test, we see that TIM can distinguish between heavily overlapping actions across both modalities, such as between 'break', 'human speech', 'hit' and 'put something'. However, there are also failure cases, such as in EPIC-KITCHENS when the action 'take washing up liquid' is recognised as 'wash hand', as the model is likely confused by the context predominantly associated with the highly overlapping 'wash hand' actions.", + "bbox": [ + 75, + 595, + 470, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Analysing Time Intervals", + "text_level": 1, + "bbox": [ + 76, + 847, + 307, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We showcase the importance of effectively encoding time intervals, as well as how they differ from alternative strate-", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fcb9398663fa72e6ca36f8b903cb62b76c2adc38c0a8483e790fdd315583d383.jpg", + "image_caption": [ + "Figure 5. TSNE plot for time encodings $I(\\cdot)$ on all datasets. In each plot, we use colour maps to indicate encodings of the time interval's duration (left), start time (middle) and end time (right)." + ], + "image_footnote": [], + "bbox": [ + 498, + 409, + 890, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "gies. We perform this analysis on the EPIC-KITCHENS-100 and EPIC-SOUNDS recognition tasks.", + "bbox": [ + 498, + 699, + 890, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Time Encoding Representation. To show TIM encodes time intervals across all datasets in Figure 5. We use three colour maps on the same TSNE projection to show three properties of the encoded interval: duration, start time and end time. Interestingly, the 1D time encoding perfectly captures all three attributes and across the datasets. While the encodings differ per dataset, as these differ in the positions and durations of actions, we see clear similarities in the learnt time encoding projections. For example, the duration is perfectly captured across the x-axis of the TSNE plot with lower values indicating longer time intervals.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18159", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79bc2ccf94924ad8be6bfba4acdc35e7bc0366f47ffe04c2bbfe24b4e649eddd.jpg", + "image_caption": [ + "Figure 6. Attention heatmaps of the second encoder layer for two random 30s clips in EPIC-KITCHENS. x-axis: input feature time intervals; y-axis query time intervals of varying position and duration. The attention magnitude relates to the query CLS token." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 467, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9796b37d53ca4aae70507d2799aa1a95b390944ed8618727740fb19ecdaf61c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 369, + 428, + 460 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e962224ee5858897afd4669f2fb4f8108ffa74f8a1070e3beedcaa0e32feba15.jpg", + "image_caption": [ + "Figure 7. The impact of shifting the time interval query on visual performance (top) and audio performance (bottom), both on short actions $(< 2\\mathrm{sec})$ and long actions $(>2\\mathrm{sec})$ and overall validation set (all) for EPIC-KITCHENS-100 and EPIC-SOUNDS." + ], + "image_footnote": [], + "bbox": [ + 116, + 463, + 428, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Interval Query Attention. We plot two attention heatmaps in Figure 6 for 5 separate queries with varying positions and scales in EPIC-KITCHENS-100. We extract the attention weights from the second transformer encoder layer, as this appears to be the most relevant to the interval query. The learnt attention clearly applies to the feature time intervals contained within the query. We note the similarity between the attention in the two randomly selected windows.", + "bbox": [ + 75, + 625, + 467, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Shifting Intervals. To show how TIM effectively encodes the time interval of actions, we shift the time interval queries from their correct action interval by -1.5s to 1.5s, assessing the impact of these adjustments on the performance.", + "bbox": [ + 75, + 750, + 467, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 7 shows the result. We see the performance gradually drops, both in visual and audio, as the query interval moves away from the correct action interval. The drop is also symmetric showcasing no bias. Unsurprisingly, the performance drops significantly when shifting short actions both in video $(-57.9\\%)$ and audio $(-35.2\\%)$ , while it is less", + "bbox": [ + 75, + 810, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/822a16d57cacc349cf0f7c6b61d9a619f9fc1c6934021d96297fcac44ff9b72c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EPIC-KITCHENSEPIC-SOUNDS
EncodingVerbNounActionAudio Actions
Learned43.844.329.623.7
Sinusoidal43.844.630.013.4
Centre74.365.855.656.4
Separate-add76.066.256.457.7
Interval-add76.366.556.958.8
Separate-cat76.867.457.158.4
Interval-cat (proposed)77.167.257.558.3
", + "bbox": [ + 504, + 88, + 890, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablating the choice of encoding time intervals.", + "bbox": [ + 529, + 215, + 861, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "extreme in long actions $(-14.5\\%$ and $-11.2\\%)$ . We assess the impact of scaling the time interval in the ArXiv appendix.", + "bbox": [ + 498, + 247, + 890, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Time Interval Encodings. The Time Interval MLP encodes the interval of the query. Here, we compare this to traditional positional encodings, both sinusoidal and learned. We also experiment on five different variations of the Time Interval MLP, namely: (i) Centre - we only encode the centre timestamp of the interval; (ii) Separate-Add/Cat - we encode the intervals' start and end time separately and add the encoded output vectors together, or concatenate along the channel dimension; and (iii) Interval-Add/Cat - we encode the start and end time within the same vector and add, or concatenate, the encoded output to the input sequence.", + "bbox": [ + 496, + 279, + 890, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show the results in Table 7. In all cases, the final encoding is of the same dimension for comparable results. Performance is significantly worse with sinusoidal or learned positional encoding, as they are unable to capture the complexity of overlapping actions. There is also a drop when encoding only the centre of the time interval.", + "bbox": [ + 496, + 446, + 890, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Separate-Add/Cat are alternative ways to encode the intervals (and hence include duration information) resulting in comparable performance to the interval counterparts. Our proposed approach to encoding the interval into the MLP shows the best performance for visual and while maintaining strong auditory performance.", + "bbox": [ + 496, + 537, + 890, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 641, + 625, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose to utilise the action's time interval as a query to an audio-visual transformer which learns to recognise the action from its interval and the unaltered surrounding context. We jointly train the model on modality-specific time intervals and label sets, allowing the Time Interval Machine (TIM) to recognise multiple events across both visual and auditory modalities.", + "bbox": [ + 496, + 666, + 890, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TIM is sensitive to the interval's position and duration. This allows the model, as is, to produce competitive results on action detection through multi-scale dense querying.", + "bbox": [ + 496, + 772, + 890, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work uses public datasets. It is supported by EPSRC Doctoral Training Program, EPSRC UMPIRE EP/T004991/1 and EPSRC Programme Grant VisualAI EP/T028572/1; and by the use of the EPSRC funded Tier 2 facility JADE-II.", + "bbox": [ + 496, + 816, + 890, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18160", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 1", + "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5", + "[3] Swapnil Bhosale, Sauradip Nag, Diptesh Kanojia, Jiankang Deng, and Xiatian Zhu. Diffsed: Sound event detection with denoising diffusion. arXiv preprint arXiv:2308.07293, 2023. 6", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. ECCV, 2020. 2", + "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1", + "[6] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 5", + "[7] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 2021. 2, 5", + "[8] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast Networks for Video Recognition. In Proceedings of International Conference on Computer Vision (ICCV), 2019. 1", + "[9] Fan Feng, Yue Ming, Nannan Hu, Hui Yu, and Yuanan Liu. CSS-net: A consistent segment selection network for audiovisual event localization. IEEE Transactions on Multimedia, 2023. 6", + "[10] Ruohan Gao, Tae-Hyun Oh, Kristen Grauman, and Lorenzo Torresani. Listen to look: Action recognition by previewing audio. In Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[11] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A Single Model for Many Visual Modalities. In CVPR, 2022. 1, 5, 6", + "[12] Yuan Gong, Cheng-I Lai, Yu-An Chung, and James Glass. Ssast: Self-supervised audio spectrogram transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10699-10709, 2022. 6", + "[13] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5", + "[14] Roei Herzig, Ofir Abramovich, Elad Ben-Avraham, Assaf Arbelle, Leonid Karlinsky, Ariel Shamir, Trevor Darrell, and" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Amir Globerson. Promptonomyvit: Multi-task prompt learning improves video transformers using synthetic scene data. arXiv preprint arXiv:2212.04821, 2022. 2", + "[15] Jaesung Huh, Jacob Chalk, Evangelos Kazakos, Dima Damen, and Andrew Zisserman. EPIC-SOUNDS: A LargeScale Dataset of Actions that Sound. In IEEE International Conference on Acoustics, Speech, & Signal Processing (ICASSP), 2023. 1, 2, 5, 6", + "[16] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022. 2", + "[17] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. Epic-fusion: Audio-visual temporal binding for egocentric action recognition. In Proceedings of International Conference on Computer Vision (ICCV), 2019. 2, 6", + "[18] Evangelos Kazakos, Jaesung Huh, Arsha Nagrani, Andrew Zisserman, and Dima Damen. With a little help from my temporal context: Multimodal egocentric action recognition. In Proc. BMVC, 2021. 1, 2, 6", + "[19] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. Slow-fast auditory streams for audio recognition. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021. 5, 6", + "[20] Dan Kondratyuk, Liangzhe Yuan, Yandong Li, Li Zhang, Mingxing Tan, Matthew Brown, and Boqing Gong. Movinets: Mobile video networks for efficient video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16020-16030, 2021. 6", + "[21] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco Nadai. Efficient training of visual transformers with small datasets. Advances in Neural Information Processing Systems, 34, 2021. 4", + "[22] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3202-3211, 2022. 1", + "[23] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2", + "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[25] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12475-12486, 2021. 2", + "[26] Arsha Nagrani, Shan Yang, Anurag Arnab, Cordelia Schmid, and Chen Sun. Attention bottlenecks for multimodal fusion. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 6", + "[27] Yan Bin Ng and Basura Fernando. Human action sequence classification. CoRR, abs/1910.02602, 2019. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18161", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Daniel S Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D Cubuk, and Quoc V Le. Specaugment: A simple data augmentation method for automatic speech recognition. arXiv preprint arXiv:1904.08779, 2019. 5", + "[29] Mandela Patrick, Dylan Campbell, Yuki M. Asano, Ishaan Misra Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and João F. Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 4, 6", + "[30] Viorica Pătrăucean, Lucas Smaira, Ankush Gupta, Adrià Recasens Continente, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Joseph Heyward, Mateusz Malinowski, Yi Yang, Carl Doersch, Tatiana Matejovicova, Yury Sulsky, Antoine Miech, Alex Frechette, Hanna Klimczak, Raphael Koster, Junlin Zhang, Stephanie Winkler, Yusuf Aytar, Simon Osindero, Dima Damen, Andrew Zisserman, and João Carreira. Perception test: A diagnostic benchmark for multimodal video models, 2023. 1, 2, 5", + "[31] Fadime Sener, Dipika Singhania, and Angela Yao. Temporal aggregate representations for long-range video understanding. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 1", + "[32] Divya Shanmugam, Davis Blalock, Guha Balakrishnan, and John Guttag. Better aggregation in test-time augmentation. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 4", + "[33] Jonathan C Stroud, Zhichao Lu, Chen Sun, Jia Deng, Rahul Sukthankar, Cordelia Schmid, and David A Ross. Learning video representations from textual web supervision. arXiv preprint arXiv:2007.14937, 2020. 5", + "[34] Swathikiran Sudhakaran, Adrian Bulat, Juan-Manuel Perez-Rua, Alex Falcon, Sergio Escalera, Oswald Lanz, Brais Martinez, and Georgios Tzimiropoulos. Saic.cambridge-hupba-fbk submission to the epic-kitchens-100 action recognition challenge 2021. arXiv preprint arXiv:2110.02902, 2021. 5", + "[35] Tsung-Ming Tai, Oswald Lanz, Giuseppe Fiameni, Yi-Kwan Wong, Sze-Sen Poon, Cheng-Kuang Lee, Ka-Chun Cheung, and Simon See. Nvidia-unibz submission for epic-kitchens-100 action anticipation challenge 2022. arXiv preprint arXiv:2206.10869, 2022. 5", + "[36] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In ECCV, 2018. 2, 5, 6", + "[37] Yapeng Tian, Dingzeyu Li, and Chenliang Xu. Unified multisensory perception: Weakly-supervised audio-visual video parsing. In Proc. ECCV, 2020. 2", + "[38] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In Advances in Neural Information Processing Systems, 2022. 5", + "[39] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 1" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Luyu Wang, Pauline Luc, Yan Wu, Adria Recasens, Lucas Smaira, Andrew Brock, Andrew Jaegle, Jean-Baptiste Alayrac, Sander Dieleman, Joao Carreira, et al. Towards learning universal audio representations. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4593-4597. IEEE, 2022. 5", + "[41] Weiyao Wang, Du Tran, and Matt Feiszli. What makes training multi-modal classification networks hard? In Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[42] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 284-293, 2019. 1, 2", + "[43] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 2, 6", + "[44] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740, 2020. 2", + "[45] Xuehan Xiong, Anurag Arnab, Arsha Nagrani, and Cordelia Schmid. M&m mix: A multimodal multiview transformer ensemble. arXiv preprint arXiv:2206.09852, 2022. 5, 6", + "[46] Mengmeng Xu, Chen Zhao, David S. Rojas, Ali Thabet, and Bernard Ghanem. G-tad: Sub-graph localization for temporal action detection, 2020. 6", + "[47] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3333-3343, 2022. 6", + "[48] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Temporal query networks for fine-grained video understanding. In Proc. CVPR, 2021. 2", + "[49] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, pages 492-510, 2022. 2, 6", + "[50] Yue Zhao and Philipp Kähenbuhl. Training a large video model on a single machine in a day. arXiv preprint arXiv:2309.16669, 2023. 5, 6", + "[51] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, 2023. 5, 6", + "[52] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In The AAAI Conference on Artificial Intelligence (AAAI), pages 12993-13000, 2020. 5", + "[53] Jinxing Zhou, Liang Zheng, Yiran Zhong, Shijie Hao, and Meng Wang. Positive sample propagation along the audiovisual event line. In Proceedings of the IEEE/CVF Con-" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18162", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer Vision and Pattern Recognition, pages 8436-8444, 2021. 6", + "[54] Jinxing Zhou, Dan Guo, and Meng Wang. Contrastive positive sample propagation along the audio-visual event line. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 6" + ], + "bbox": [ + 78, + 90, + 468, + 176 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18163", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/03af8456-92c0-4a22-a6d9-94775506b39a_model.json b/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/03af8456-92c0-4a22-a6d9-94775506b39a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..53aa805217a564c706ee2172d4affe69559d908d --- /dev/null +++ b/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/03af8456-92c0-4a22-a6d9-94775506b39a_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.141, + 0.131, + 0.83, + 0.154 + ], + "angle": 0, + "content": "TIM: A Time Interval Machine for Audio-Visual Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.18, + 0.341, + 0.199 + ], + "angle": 0, + "content": "Jacob Chalk\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.399, + 0.181, + 0.519, + 0.199 + ], + "angle": 0, + "content": "Jaesung Huh\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.2, + 0.477, + 0.216 + ], + "angle": 0, + "content": "Andrew Zisserman" + }, + { + "type": "text", + "bbox": [ + 0.579, + 0.181, + 0.744, + 0.199 + ], + "angle": 0, + "content": "Evangelos Kazakos3" + }, + { + "type": "text", + "bbox": [ + 0.538, + 0.2, + 0.653, + 0.216 + ], + "angle": 0, + "content": "Dima Damen" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.216, + 0.282, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)University of Bristol" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.217, + 0.53, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\)VGG, University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.549, + 0.217, + 0.862, + 0.235 + ], + "angle": 0, + "content": "3 Czech Technical University in Prague" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.471, + 0.468 + ], + "angle": 0, + "content": "Diverse actions give rise to rich audio-visual signals in long videos. Recent works showcase that the two modalities of audio and video exhibit different temporal extents of events and distinct labels. We address the interplay between the two modalities in long videos by explicitly modelling the temporal extents of audio and visual events. We propose the Time Interval Machine (TIM) where a modality-specific time interval poses as a query to a transformer encoder that ingests a long video input. The encoder then attends to the specified interval, as well as the surrounding context in both modalities, in order to recognise the ongoing action." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.469, + 0.47, + 0.665 + ], + "angle": 0, + "content": "We test TIM on three long audio-visual video datasets: EPIC-KITCHENS, Perception Test, and AVE, reporting state-of-the-art (SOTA) for recognition. On EPIC-KITCHENS, we beat previous SOTA that utilises LLMs and significantly larger pre-training by \\(2.9\\%\\) top-1 action recognition accuracy. Additionally, we show that TIM can be adapted for action detection, using dense multi-scale interval queries, outperforming SOTA on EPIC-KITCHENS-100 for most metrics, and showing strong performance on the Perception Test. Our ablations show the critical role of integrating the two modalities and modelling their time intervals in achieving this performance. Code and models at: https://github.com/JacobChalk/TIM." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.208, + 0.712 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.722, + 0.469, + 0.828 + ], + "angle": 0, + "content": "Long videos exhibit a quick succession of auditory and visual events. The latest attempts to annotate events in these modalities separately [15, 30], showcase that both the temporal extents and class labels differ between the two. However, these events still remain correlated - identifying temporally close events in both modalities can improve recognition of actions in both visual and audio." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.83, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Furthermore, most methods to date typically only utilise the exact temporal extent of an action; a precise, trimmed clip of the action is fed into a convolutional [5, 8, 39] or" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.267, + 0.895, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.428, + 0.895, + 0.526 + ], + "angle": 0, + "content": "Figure 1. Time Interval Machine (TIM): Top: Given a visual and auditory stream input, the ongoing action in a particular time interval is determined by a query specifying the start and end time of the interval, along with the modality of interest. Bottom: TIM can query for visual (e.g. 'Rinse Sponge') and auditory (e.g. 'Water') action classes, as well as distinguish between overlapping actions within the same modality ('Glass Collision' and 'Scrub / Scrape')." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.629 + ], + "angle": 0, + "content": "transformer-based [1, 11, 22] backbone, which predicts the action taking place. Even when the surrounding context is utilised to improve action recognition [18, 31, 42], again, this context is supplied in the form of exact clips of neighbouring actions, rather than the untrimmed long input video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.75 + ], + "angle": 0, + "content": "In this paper, we propose an approach that encodes multiple events that occur in both visual and auditory streams of a long video input. We achieve this by elevating time intervals to first-class citizens, utilising them to specify a query within an accompanying modality. We term this mechanism a Time Interval Machine (TIM). It is able to receive a long video input, and output the actions that occur within the queried intervals of the queried modalities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Consider the example in Figure 1. The input contains the sound of water running while a sponge is being rinsed, which is then used to wipe a surface. These distinct events may vary significantly in duration and may be more prominent in the audio or visual modality. Despite differences between these events, there are likely many correlations between them and the surrounding context, which may be beneficial to recognising a given event (e.g. the sound of water is relevant to rinsing a sponge, providing useful information for recognising the visual action). TIM is able" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.264, + 0.9 + ], + "angle": 0, + "content": "*Equal technical contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18153" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.166 + ], + "angle": 0, + "content": "to exploit this by accessing the context within both modalities, including the background when no events occur. It can then distinguish between different, potentially overlapping, events within the same input by querying the time interval of a particular event within a given modality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.378 + ], + "angle": 0, + "content": "We test TIM on three challenging audio-visual recognition datasets consisting of long videos: EPIC-KITCHENS [7], which recently offered distinct audio annotations through EPIC-SOUNDS [15], the Perception Test [30], and AVE [36]. We show that TIM can effectively learn both visual and auditory classes within a long input, outperforming the current SOTA top-1 accuracy on EPIC-KITCHENS by \\(2.9\\%\\) and \\(1.4\\%\\) on EPIC-SOUNDS, despite competing methods for the former using far larger pre-training datasets, large language models or higher resolution input. We also outperform models pre-trained with public datasets on AVE by \\(0.6\\%\\) and improve over a strong baseline on the Perception Test in visual and audio action recognition by \\(9.9\\%\\) and \\(3.2\\%\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.379, + 0.468, + 0.453 + ], + "angle": 0, + "content": "Additionally, we adapt TIM to action detection, through fixed multi-scale dense querying with an added interval regression loss. We report strong detection results on EPIC-KITCHENS and the Perception Test, outperforming Action Former [49] by 1.6 avg mAP and 4.3 respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.468, + 0.589 + ], + "angle": 0, + "content": "Our contributions are summarised as: (i) we propose the TIM query mechanism for attending to modality-specific intervals in long videos. (ii) we efficiently train TIM to encode/query multiple audio-visual actions using time intervals. (iii) we showcase the value of TIM for both visual and auditory action recognition, and adapt it for detection with an added interval regression loss. (iv) we achieve new SOTA in both video and multi-modal recognition on multiple datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.604, + 0.227, + 0.619 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.468, + 0.869 + ], + "angle": 0, + "content": "Audio-visual action recognition. A number of works have employed audio and visual modalities for action recognition [10, 17, 18, 26, 41, 44]. Some introduce new architectures to effectively fuse modalities [17, 18, 26, 44]; others propose unique training techniques to solve problems occurring while training multi-modal models, such as Gradient Blending [41], to tackle overfitting at different speeds for each modality, or contrastive learning for cross-modal discrimination [25]. However, these works use the same set of semantic and temporal labels for both modalities. Recent works have shown that both the temporal intervals and semantics of events differ between modalities [15, 30]. [37] temporally annotates visual and auditory events independently, although they share the same set of labels. In this work, we train with distinct labels for each modality to leverage discriminative audio and visual actions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Leveraging temporal context. Several works have considered incorporating temporal context [18, 27, 42, 43, 48], a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.348 + ], + "angle": 0, + "content": "direction orthogonal to employing multiple modalities and particularly useful in untrimmed videos. An auto-regressive LSTM-based encoder-decoder is proposed in [27] for action sequence classification, effectively leveraging past action context to predict the current action. The Temporal Query Network [48] uses learnable query vectors that correspond to specific attributes of a long video, allowing the model to attend to the aspects of the video and its surrounding context to produce a response for each attribute. [42] proposes to enhance the representation of the action by aggregating temporal context from neighbouring action clips using a Long-Term Feature Bank along with an attention mechanism. [43] crafts a more sophisticated memory bank by storing keys and values of all the intermediate layers of a transformer to aggregate the past context. Lastly, [18] exploits multi-modal temporal context from surrounding actions using vision, audio, and language." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[18, 42, 43] are the closest to our approach, in that the common goal is to enrich the representation of the action of interest using surrounding context from the untrimmed video, rather than neighbouring clips. Nevertheless, [42, 43] are single modality models, recognising visual actions solely. [18] assumes the temporal extents of all actions are known, including for the test set, which is restrictive." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.664 + ], + "angle": 0, + "content": "Queries in visual models. Learning visual queries with Transformer architectures has gained recent attention [4, 14, 16, 23, 48]. Commonly, approaches employ a set of learnable vectors that are used to inquire about the presence of a concept in the input. For example, in [4, 23] the learnable queries correspond to different objects, whereas in [14] they are used for multi-task learning and each learnable query corresponds to a different task. [16] has incorporated learnable queries for adapting a pre-trained model while keeping the rest of its parameters frozen. Closest to our motivation is [48], where the queries correspond to events and their attributes for fine-grained action recognition in videos. The authors note that the queries also have the role of temporally localising the events in untrimmed videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.666, + 0.892, + 0.756 + ], + "angle": 0, + "content": "Different from [48] and other works, our queries are primarily temporal with no semantic interpretation and are applied to multiple modalities. Importantly, since time is continuous, we cannot use a predefined set of queries. Instead, we employ an MLP architecture to encode time, in a form akin to a universal clock. We present our approach next." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.719, + 0.785 + ], + "angle": 0, + "content": "3. Time Interval Machine" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we describe the Time Interval Machine (TIM), a multi-modal transformer encoder architecture where all inputs, both features and queries, are encoded with their associated time intervals. A time interval incorporates the duration and position of each audio and visual feature and is also used to query the network for any action occurring within the given time interval." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18154" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.895, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.336, + 0.895, + 0.38 + ], + "angle": 0, + "content": "Figure 2. Overview of the Time Interval Machine (TIM). The model ingests a sequence of audio and visual features from a video, with each feature time-stamped by the temporal interval it spans, and encoded with its modality. To infer the action occurring over a temporal interval (a visual or audio event) a query is formed specifying the interval and modality of interest." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.471, + 0.451 + ], + "angle": 0, + "content": "The architecture of TIM is illustrated in Figure 2. It ingests a large video input, represented as a sequence of audio and visual features, and outputs the ongoing auditory or visual action label for the provided query time intervals." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.46, + 0.262, + 0.475 + ], + "angle": 0, + "content": "3.1. Model architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.47, + 0.679 + ], + "angle": 0, + "content": "Input. The input to TIM is a long crop of the untrimmed video, represented by extracted features. When considering two modality inputs, such as video and audio, each modality is embedded separately as follows: for each modality \\( m \\), let \\( \\mathbf{X}^m = [X_1^m, \\dots, X_{N^m}^m] \\) be \\( N^m \\) temporally-ordered feature representations of the input video, obtained from a pre-trained feature extractor \\( S^m(\\cdot) \\). We feed the features through modality-specific embedding layers \\( g^m(\\cdot) \\), projecting them to a lower, common dimension \\( D \\) across all modalities. The embedded features are then tagged with modality encodings and time interval encodings, forming the input to the transformer encoder. We now detail how we encode the time intervals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.788 + ], + "angle": 0, + "content": "Encoding Time Intervals. In this work, we introduce a new type of learnt query network, the Time Interval MLP, which produces a single \\(D\\)-dimensional vector representing a given time interval. This network is used within TIM to encode the time intervals of the input features and the time interval we wish to query, and later classify. Figure 3 illustrates the concept of this network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.787, + 0.471, + 0.88 + ], + "angle": 0, + "content": "The Time Interval MLP \\( I(\\cdot): \\mathbb{R}^2 \\to \\mathbb{R}^D \\) receives a time interval, represented by start and end time, and produces a single \\( D \\)-dimensional encoding. Note that this is distinct from encoding the start and end times separately. Specifically, let \\( t_s \\) and \\( t_e \\) be the start and end time of an interval of interest, normalised by the length of the long video input." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.388, + 0.872, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.59 + ], + "angle": 0, + "content": "Figure 3. Illustration of the Time Interval MLP \\( I(\\cdot) \\). It inputs the two-dimensional vector, start and end times of an interval, and produces a single vector, which can be concatenated along the channel dimension to either input features or [CLS] tokens. The figure shows three time interval inputs and three corresponding outputs. Note that in practice, time intervals are ingested simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.892, + 0.728 + ], + "angle": 0, + "content": "\\(I(\\cdot)\\) receives the interval \\(\\tilde{t} = [t_s, t_e]\\) as input, and outputs a \\(D\\)-dimensional vector encoding of that interval. This vector encodes both the relative position of the time interval within the input, as well as its duration. This vector then acts as a query for the model concerning the action taking place within the interval. Furthermore, each feature \\(\\{X_i^m\\}\\) spans a certain time interval within the input. Thus, it is important to also encode the time intervals of the features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.73, + 0.892, + 0.881 + ], + "angle": 0, + "content": "In summary, the Time Interval MLP acts as an universal clock, which encodes the temporal extent of features, from any modality, within the input. Note that it is critical that the same Time Interval MLP is used for encoding all time intervals of the input features and queries across both modalities to accurately encode universal time. It is also important to note that Time interval MLP can cover continuous time intervals, whereas traditional positional encoding only covers a fixed set of positions of the input features. The Time Interval MLP is trained end-to-end along with the transformer." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.885, + 0.891, + 0.904 + ], + "angle": 0, + "content": "Transformer Feature Inputs. Let \\(\\tilde{\\mathbf{t}}^m = [\\tilde{t}_1^m, \\dots, \\tilde{t}_{N^m}^m]\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.434, + 0.9 + ], + "angle": 0, + "content": "1Note that the number of features can differ between modalities" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18155" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "be the corresponding time intervals of the video's features \\(\\mathbf{X}^m\\) from the modality \\(m\\). We inject the encoded time interval \\(I(\\tilde{\\mathbf{t}}^m)\\) into the embedded features via channelwise concatenation. A learnable modality-specific encoding \\(e^m \\in \\mathbb{R}^{2D}\\) is then summed to the temporally-encoded features to discriminate between each modality. In summary, the feature inputs \\(\\mathbf{E}^m\\) for TIM are computed by," + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.208, + 0.47, + 0.227 + ], + "angle": 0, + "content": "\\[\nE _ {i} ^ {m} = \\left[ g ^ {m} \\left(X _ {i} ^ {m}\\right), I \\left(\\tilde {t} _ {i} ^ {m}\\right) \\right] + e ^ {m} \\quad \\forall i \\in [ 1, \\dots , N ^ {m} ] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.235, + 0.314, + 0.251 + ], + "angle": 0, + "content": "where \\([\\cdot ,\\cdot ]\\) indicates concatenation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.253, + 0.47, + 0.42 + ], + "angle": 0, + "content": "Transformer Query Inputs. To query for an action within an interval of interest, we adopt a standard approach of appending a learnable classification token to the input sequence, \\(\\mathsf{CLS}^m\\). If \\(\\tilde{t}_Q\\) is an interval of interest, we concatenate the time interval representation \\(I(\\tilde{t}_Q)\\) to this classification token along the channel dimension, which acts as a query for the network in order to predict the corresponding action happening within \\(\\tilde{t}_Q\\). We also add the modality-specific encoding \\(e^m\\) to each classification token, as a flag to distinguish between which modality we are querying. The encoded \\([\\mathsf{CLS}]^m\\) tokens can be more formally defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.43, + 0.47, + 0.45 + ], + "angle": 0, + "content": "\\[\n[ \\mathrm {C L S} ] ^ {m} = [ \\mathrm {C L S} ^ {m}, I (\\tilde {t} _ {Q}) ] + e ^ {m} (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.503 + ], + "angle": 0, + "content": "During training, we add a classification token for each action within the input video, resulting in multiple [CLS] tokens across both modalities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.55 + ], + "angle": 0, + "content": "Transformer Encoder. We use a transformer encoder to perform self-attention on the input sequence to aggregate relevant temporal context and cross-modal relations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.469, + 0.672 + ], + "angle": 0, + "content": "We form the transformer input sequence with the encoded feature inputs \\(\\mathbf{E}^m\\) and one or more classification tokens \\([\\mathrm{CLS}]^m\\), representing each time interval query, and feed these into the encoder. Note that we recognise all actions from any modality simultaneously by appending multiple \\(\\mathrm{CLS}^m\\) tokens to the input. The transformer output representation of \\([\\mathrm{CLS}]^m\\), namely \\(Z_{\\mathrm{CLS}}^m\\), is then passed to the corresponding linear classifier to predict the action labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Importantly, we use an attention mask to prevent queries from attending to one another, and we similarly prevent input features from attending to queries. This ensures each query is recognised without the privileged knowledge of any other query, or action boundary, during inference." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.334, + 0.773 + ], + "angle": 0, + "content": "3.2. Training and Testing in TIM" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.855 + ], + "angle": 0, + "content": "To train TIM, we consider all long segments of \\( W \\) seconds and stride \\( H_{w} \\) across the entirety of the untrimmed videos. We randomly select batches from these. For each window, we query all annotated audio and visual actions that overlap with the window by more than \\( \\delta = 0.2 \\) seconds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "All queries in the window are encoded and concatenated to separate CLS tokens. To classify queries, let \\( h_{\\mathrm{CLS}}^m(\\cdot) \\) be a linear classifier for modality \\( m \\), and let \\( \\hat{y}_{\\mathrm{CLS}}^m = h^m(Z_{\\mathrm{CLS}}^m) \\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "be the predicted action of the output representation \\( Z_{\\mathrm{CLS}}^{m} \\). We train TIM by using a cross-entropy classification loss \\( CE(\\cdot) \\) on the ground truth \\( y_{\\mathrm{CLS}}^{m} \\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.149, + 0.892, + 0.189 + ], + "angle": 0, + "content": "\\[\nL ^ {m} = \\frac {1}{N _ {Q}} \\sum^ {N _ {Q}} C E \\left(\\hat {y} _ {\\mathrm {C L S}} ^ {m}, y _ {\\mathrm {C L S}} ^ {m}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.197, + 0.848, + 0.213 + ], + "angle": 0, + "content": "where \\(N_{Q}\\) is the number of queries within the batch." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.215, + 0.892, + 0.32 + ], + "angle": 0, + "content": "Temporal Distance Loss. In addition to the standard classification loss, we introduce a Temporal Distance (TD) loss as an auxiliary loss for training TIM. Inspired by [21], where relative patch positions in token embeddings are learnt using self-supervision, we similarly train the network to take two transformer outputs and predict the elapsed time between their corresponding time intervals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.893, + 0.47 + ], + "angle": 0, + "content": "Let the \\(\\mathbf{Z}_{1:\\sum_mN^m}\\) be the transformer outputs of the features from all modalities. We randomly sample a set of feature pairs \\(\\mathbb{B} \\subset \\mathbf{Z}_{1:\\sum_mN^m}\\) from these outputs, concatenate along the channel dimension and feed them to the temporal distance regression head \\(h_{\\tilde{t}}(\\cdot):\\mathbb{R}^{4D} \\to \\mathbb{R}^1\\) to predict the time interval difference between each pair. Note that feature pairs can be sampled both within and across modalities. In our case, we sample across modalities by pairing one visual feature with another audio feature. This helps the model to learn the temporal relations between modalities." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.47, + 0.788, + 0.486 + ], + "angle": 0, + "content": "Formally, a TD loss \\(L^{td}\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.498, + 0.892, + 0.533 + ], + "angle": 0, + "content": "\\[\nL ^ {t d} = \\sum_ {\\left\\{Z _ {i}, Z _ {j} \\right\\} \\in \\mathbb {B}} \\left| h _ {\\tilde {i}} \\left(Z _ {i}, Z _ {j}\\right) - d _ {i j} \\right| \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.544, + 0.887, + 0.561 + ], + "angle": 0, + "content": "where \\(d_{ij}\\) is the temporal distance between intervals \\(\\tilde{t}_i,\\tilde{t}_j\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.562, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Training objective and regime. For our final training loss, we sum the losses across modalities along with the TD loss:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.603, + 0.892, + 0.645 + ], + "angle": 0, + "content": "\\[\nL ^ {t o t a l} = \\left(\\sum_ {m \\in \\mathbb {M}} \\lambda^ {m} L ^ {m}\\right) + \\lambda^ {t d} L ^ {t d} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.7 + ], + "angle": 0, + "content": "where \\(\\mathbb{M}\\) is a set of modalities, \\(\\lambda^m\\) controls the strength of each modality's loss and \\(\\lambda^{td}\\) is a hyperparameter that controls the strength of the TD loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Test-Time Augmentation. We use test-time augmentation, as this generally increases prediction robustness and performance [29, 32]. In TIM, we use a sliding window across the untrimmed video, thus feeding the same interval query with varying contexts. We then aggregate predictions of the same interval query across windows to make the final prediction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.714, + 0.818 + ], + "angle": 0, + "content": "3.3. Adapting for Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "While primarily designed for recognition, we can adapt TIM for detection. The backbone remains largely unchanged from recognition, but there are two main differences. First, we construct dense multi-scale interval queries spanning the entirety of the video input at each scale. These" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18156" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.166 + ], + "angle": 0, + "content": "are used as interval queries in both training and detection inference. The multi-scale intervals allow for detecting both long and short actions. Second, we introduce an additional interval regression head, which regresses the query interval to the action's exact temporal duration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.273 + ], + "angle": 0, + "content": "During training, we deem any query in the multi-scale pyramid that overlaps with a ground truth action by more than some IoU threshold as a positive query. In addition to classifying the query, we train a DIOU regression loss [52] to predict the exact interval of the action. Both classification and interval regression losses are trained jointly. We provide full details in the ArXiv appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.286, + 0.208, + 0.303 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.468, + 0.356 + ], + "angle": 0, + "content": "This section describes the datasets used to evaluate our model, implementation details and results along with a comparison to the state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.365, + 0.172, + 0.38 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.468, + 0.48 + ], + "angle": 0, + "content": "EPIC-KITCHENS-100 [7] is a large-scale video dataset, including 700 egocentric videos recording actions in kitchens. It consists of 89,977 segments of fine-grained actions. Inspired by prior works [11, 34, 35], we directly predict the action out of 3806 classes present in the train and validation set to avoid predicting invalid actions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.484, + 0.468, + 0.589 + ], + "angle": 0, + "content": "EPIC-SOUNDS [15] offers audio annotations capturing temporal extents and class labels within the audio stream of EPIC-KITCHENS-100. The annotations contain 78,366 labelled audio events. We combine the visual annotations from EPIC-KITCHENS with the audio annotations from EPIC-SOUNDS to train our audio-visual model. TIM can recognise actions from both datasets using a single model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.468, + 0.697 + ], + "angle": 0, + "content": "AVE [36] contains 4,143 videos covering a range of real-life scenes and labelled with 27 categories, such as church bell, male speaking and dog barking. Each video is equally divided into 10 segments, each 1 second in length. We evaluate TIM on the supervised audio-visual event localisation task. Given a 1 second segment, we recognise the ongoing action out of the 27 categories plus a background class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.468, + 0.775 + ], + "angle": 0, + "content": "Perception Test [30] is a recent multimodal video benchmark of 11,620 videos with an average length of 23 seconds, and provides both temporal action and sound annotations. There are 73,503 visual annotations spanning 63 classes, versus 137,128 sound annotations over 16 classes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.294, + 0.801 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Architectural Details. Visual and audio embedding layers \\( g_{m} \\) consist of a single 512-D feed-forward layer, followed by a GELU [13] activation and layer normalisation [2] are used to project the features to a common space. The Time Interval MLP \\( I \\) consists of three linear layers with 512-D hidden dimension, followed by ReLU activations, with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.891, + 0.166 + ], + "angle": 0, + "content": "layer normalisation after the output of the last linear layer. We include 512-D learnable [CLS] tokens: \\([\\mathrm{CLS}]_{\\text {action }}^{m}\\) for each query in each modality which become 1024-D after concatenation with the encoded time interval. They are then summed with 1024-D modality encodings; \\(e^{m}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.333 + ], + "angle": 0, + "content": "The audio-visual transformer contains four encoder layers, each with 8 attention heads, GELU activations, and 1024-D keys, queries and values. A dropout rate of \\( p = 0.1 \\) is applied within the encoder layers. We also apply channelwise dropout with \\( p = 0.5 \\) directly to the raw input features, as well as to the encoded transformer input. The temporal distance head consists of two linear layers with a hidden dimension of 1024 and a third which outputs a single number corresponding to the elapsed time between each time interval. We include the architectural ablations on encoder layers and the temporal distance head in the ArXiv appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Training / Validation Details. We train each model for 100 epochs, using AdamW [24] with a batch size of 64 and a weight decay of 1e-4. A linear learning rate warm-up is applied for the first two epochs, starting from 1e-6 to a target learning rate, and we use a cosine learning rate scheduler. We set the TD loss weight \\(\\lambda^{td}\\) to 0.3. We pad the queries for each window in the batch to the maximum number of queries in a single window in each dataset. We provide implementation details per dataset in the ArXiv appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.482, + 0.593, + 0.497 + ], + "angle": 0, + "content": "4.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.506, + 0.859, + 0.52 + ], + "angle": 0, + "content": "We compare TIM with SOTA models for each dataset." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.523, + 0.891, + 0.567 + ], + "angle": 0, + "content": "EPIC-KITCHENS / EPIC-SOUNDS Results. We train a single model on both visual and audio labels of the EPIC-KITCHENS videos, reporting results on both datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.688 + ], + "angle": 0, + "content": "For the visual features, we concatenate Omnivore [11] and VideoMAE-L [38] features along the channel dimension, forming 2048-D features. For the audio features, we use Auditory SlowFast [19], which generalises well across diverse audio domains [40]. For both modalities, we extract 1 second features every 0.2s. For training, we extract additional augmented feature sets - with RandAugment [6] for visual and SpecAugment [28] for audio features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Table 1 compares TIM with the SOTA models on EPIC-KITCHCHENS-100. We outperform M&M Mix [45] by \\(5.1\\%\\) on verb, \\(0.9\\%\\) on noun and \\(3.9\\%\\) on action. Compared to our model, both MTV and M&M Mix are trained with an additional private dataset [33] which contains 194K hours of 70 million videos while we only use the open-source visual backbone pre-trained with public datasets. We also outperform LaViLa [51] and AVION [50] which leverage pre-trained LLMs to learn video representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We note that we outperform all prior works, often without additional techniques that boost performance. For example, we use short-sided cropped \\(224 \\times 224\\) images while [45] uses \\(420 \\times 420\\), which enlarges the spatial resolution of objects in the egocentric video, enabling better noun recog" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18157" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.088, + 0.468, + 0.304 + ], + "angle": 0, + "content": "
ModelxpLLMVerbNounAction
Visual-only models
MFormer-HR [29]336p×67.058.544.5
MoViNet-A6 [20]320p×72.257.347.7
MeMViT [43]224p×71.460.348.4
Omnivore [11]224p×69.561.749.9
MTV [47]280p×69.963.950.5
LaViLa (TSF-L) [51]224p✓72.062.951.0
AVION (ViT-L) [50]224p✓73.065.454.4
TIM (ours)224p×76.266.456.4
Audio-visual models
TBN [17]224p×66.047.236.7
MBT [26]224p×64.858.043.4
MTCN [18]336p×70.762.149.6
M&M [45]420p×72.066.353.6
TIM (ours)224p×77.167.257.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.315, + 0.47, + 0.371 + ], + "angle": 0, + "content": "Table 1. Comparisons to state-of-the-art recognition models on the EPIC-KITCHENS validation set. We report the top-1 accuracy for verb, noun and action (\\%). LLM: large language model is used during pre-training. xp: input resolution of \\( x \\times x \\)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.387, + 0.468, + 0.412 + ], + "angle": 0, + "content": "
ModelSSAST [12]ASF [19]DiffSED [3]TIM (A)TIM (A+V)
Top-1 acc53.553.856.955.758.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.422, + 0.469, + 0.464 + ], + "angle": 0, + "content": "Table 2. Comparisons to state-of-the-art sound recognition models on EPIC-SOUNDS. We report the top-1 accuracy \\((\\%)\\) on Val. The performance of SSAST and ASF are from [15]." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.481, + 0.466, + 0.572 + ], + "angle": 0, + "content": "
ModelVAAverage Precision (AP)
Task@0.1@0.2@0.3@0.4@0.5Avg.
G-TAD [46]✓✗Verb12.111.09.48.16.59.4
Noun11.010.08.67.05.48.4
ActionFormer [49]✓✗Verb26.625.424.222.319.123.5
Noun25.224.122.720.517.021.9
ActionFormer - Our Features✓✗Verb29.628.826.924.421.626.3
Noun34.332.630.227.422.629.4
TIM✓✓Verb32.931.629.627.022.228.6
Noun36.434.832.128.722.731.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.583, + 0.469, + 0.638 + ], + "angle": 0, + "content": "Table 3. Comparisons to state-of-the-art detection models on the EPIC-KITCHENS validation set. We report the average precision at IOU thresholds [0.1, 0.2, 0.3, 0.4, 0.5] as well as their average across all thresholds on verb, noun." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.653, + 0.469, + 0.713 + ], + "angle": 0, + "content": "nition. We expect a further performance boost when implementing any of: higher resolution feature extractors, additional large-scale pre-training and the introduction of a LLM. We leave this as an avenue for future work." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.715, + 0.468, + 0.745 + ], + "angle": 0, + "content": "Table 2 compares TIM against prior results on EPIC-SOUNDS, where TIM outperforms SOTA by \\(1.4\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.747, + 0.469, + 0.822 + ], + "angle": 0, + "content": "For detection, we show that TIM can produce competitive results when compared to models primarily designed for this task in Table 3. TIM adapted for detection outperforms ActionFormer [49] by \\(2.3\\mathrm{mAP}\\) on verb and \\(1.6\\mathrm{mAP}\\) on noun using the same set of features." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "AVE Results. As this dataset contains joint audio-visual labels, we train TIM by duplicating the query, i.e. using a [CLS] for each modality, and combine their logits during training and inference. We use the pre-trained publicly available models from [36] for a fair comparison with other" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.088, + 0.892, + 0.13 + ], + "angle": 0, + "content": "
ModelPSP [53]CPSP [54]CSSNet [9]†TIM
VAAVAV*
Top-1 acc77.878.680.562.865.579.279.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.141, + 0.892, + 0.197 + ], + "angle": 0, + "content": "Table 4. Top-1 event classification accuracy \\((\\%)\\) on AVE Test set. \\(\\dagger\\) : no official code or public model provided to replicate results. We show the models trained only with publicly available datasets. \\(\\star\\) : results with Omni+ASF features." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.214, + 0.89, + 0.289 + ], + "angle": 0, + "content": "
Perception Test Action
ModelMLP (V)MTCN [18](A+V)TIM (V)TIM (A+V)
Top-1 acc43.751.256.161.1
Perception Test Sound
ModelMLP (A)MTCN [18](A+V)TIM (A)TIM (A+V)
Top-1 acc50.652.954.856.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.3, + 0.892, + 0.369 + ], + "angle": 0, + "content": "Table 5. Comparisons to trained recognition baselines on the Perception Test validation split. We show both action and sound recognition and the benefit of including audio-visual in TIM for both challenges. V : visual and A : audio input features. MLP is the result by training an MLP classifier with the features directly." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.386, + 0.889, + 0.483 + ], + "angle": 0, + "content": "
ModelAverage Precision (AP)
@0.1@0.2@0.3@0.4@0.5Avg.
Perception Test Action
ActionFormer [49]27.827.625.223.020.024.5
TIM33.532.229.826.422.028.8
Perception Test Sound
ActionFormer [49]34.731.327.522.717.726.8
TIM37.533.127.922.817.227.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.494, + 0.892, + 0.55 + ], + "angle": 0, + "content": "Table 6. Comparisons to strong detection models on the Perception Test validation set for action and sound localisation. We report the average precision at IOU thresholds [0.1, 0.2, 0.3, 0.4, 0.5] as well as the average across all thresholds." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.564, + 0.891, + 0.594 + ], + "angle": 0, + "content": "works. We also apply AVGA [36] to spatial visual features from VGG-19 before feeding them to the transformer." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.596, + 0.892, + 0.686 + ], + "angle": 0, + "content": "Table 4 shows our results on the AVE dataset. Combining audio and video significantly improves the performance on TIM. The results from [9] perform best but could not be replicated. We also report TIM using the Omnivore visual features and Auditory Slowfast features used for EPIC-KITCHENS, which achieves a \\(0.6\\%\\) boost in performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Perception Test Results. We use the same backbone for the Omnivore features and Auditory Slowfast features and train a single model using both the visual and audio labels. Table 5 compares the results on the newly introduced Perception Test. We train an MLP classifier, with two linear layers and a ReLU activation, on the features directly as a baseline. We also evaluate on an audio-visual model that does use context with MTCN. Compared to these methods TIM clearly shows significant improvements. Results are improved over MTCN by \\(9.9\\%\\) and \\(3.2\\%\\) on visual and audio recognition tasks respectively. We also provide detection results in Table 6. TIM improves over ActionFormer [49] by 3.3 average mAP on visual actions and by 0.9 average mAP on sound, when using the same features." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18158" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.472, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.087, + 0.891, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.385, + 0.891, + 0.399 + ], + "angle": 0, + "content": "Figure 4. Qualitative results for all datasets. PRED: Prediction by TIM, TIQ: Time Interval Queries, V/AGT: Visual/Audio Ground Truth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.47, + 0.594 + ], + "angle": 0, + "content": "Cross-Modality in TIM. When referring to our previous results, we see that including the additional modality provides a performance boost in all cases, highlighting TIM's ability to utilise and distinguish between different modalities. For example, on EPIC-KITCHENS-100, including audio improves visual action accuracy by \\(0.9\\%\\). For EPICSOUNDS, the visual modality further improves accuracy by \\(2.6\\%\\). In the Perception Test, including the audio modality improves visual recognition by \\(5.0\\%\\), and visual increases sound recognition by \\(1.3\\%\\). Finally, for AVE, we see a significant improvement, where an audio-visual model increases accuracy by \\(13.7\\%\\) from audio-only." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.471, + 0.839 + ], + "angle": 0, + "content": "Qualitative Results. We present qualitative results in Figure 4. We see that in EPIC-KITCHENS, TIM can competently recognise actions across the two modalities including overlapping queries. Furthermore, we see consecutive actions are correctly recognised with varying interval lengths, such as the 'open / close' audio actions between 0.2s and 4.5s. For AVE, TIM is able to distinguish between the background and a 'barking' audio-visual event based on the time interval query. For the Perception Test, we see that TIM can distinguish between heavily overlapping actions across both modalities, such as between 'break', 'human speech', 'hit' and 'put something'. However, there are also failure cases, such as in EPIC-KITCHENS when the action 'take washing up liquid' is recognised as 'wash hand', as the model is likely confused by the context predominantly associated with the highly overlapping 'wash hand' actions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.308, + 0.864 + ], + "angle": 0, + "content": "4.4. Analysing Time Intervals" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We showcase the importance of effectively encoding time intervals, as well as how they differ from alternative strate-" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.642, + 0.892, + 0.684 + ], + "angle": 0, + "content": "Figure 5. TSNE plot for time encodings \\( I(\\cdot) \\) on all datasets. In each plot, we use colour maps to indicate encodings of the time interval's duration (left), start time (middle) and end time (right)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.7, + 0.891, + 0.731 + ], + "angle": 0, + "content": "gies. We perform this analysis on the EPIC-KITCHENS-100 and EPIC-SOUNDS recognition tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Time Encoding Representation. To show TIM encodes time intervals across all datasets in Figure 5. We use three colour maps on the same TSNE projection to show three properties of the encoded interval: duration, start time and end time. Interestingly, the 1D time encoding perfectly captures all three attributes and across the datasets. While the encodings differ per dataset, as these differ in the positions and durations of actions, we see clear similarities in the learnt time encoding projections. For example, the duration is perfectly captured across the x-axis of the TSNE plot with lower values indicating longer time intervals." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18159" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.301, + 0.47, + 0.358 + ], + "angle": 0, + "content": "Figure 6. Attention heatmaps of the second encoder layer for two random 30s clips in EPIC-KITCHENS. x-axis: input feature time intervals; y-axis query time intervals of varying position and duration. The attention magnitude relates to the query CLS token." + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.371, + 0.429, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.464, + 0.429, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.561, + 0.47, + 0.618 + ], + "angle": 0, + "content": "Figure 7. The impact of shifting the time interval query on visual performance (top) and audio performance (bottom), both on short actions \\((< 2\\mathrm{sec})\\) and long actions \\((>2\\mathrm{sec})\\) and overall validation set (all) for EPIC-KITCHENS-100 and EPIC-SOUNDS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.468, + 0.746 + ], + "angle": 0, + "content": "Interval Query Attention. We plot two attention heatmaps in Figure 6 for 5 separate queries with varying positions and scales in EPIC-KITCHENS-100. We extract the attention weights from the second transformer encoder layer, as this appears to be the most relevant to the interval query. The learnt attention clearly applies to the feature time intervals contained within the query. We note the similarity between the attention in the two randomly selected windows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.81 + ], + "angle": 0, + "content": "Shifting Intervals. To show how TIM effectively encodes the time interval of actions, we shift the time interval queries from their correct action interval by -1.5s to 1.5s, assessing the impact of these adjustments on the performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Figure 7 shows the result. We see the performance gradually drops, both in visual and audio, as the query interval moves away from the correct action interval. The drop is also symmetric showcasing no bias. Unsurprisingly, the performance drops significantly when shifting short actions both in video \\((-57.9\\%)\\) and audio \\((-35.2\\%)\\), while it is less" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.892, + 0.205 + ], + "angle": 0, + "content": "
EPIC-KITCHENSEPIC-SOUNDS
EncodingVerbNounActionAudio Actions
Learned43.844.329.623.7
Sinusoidal43.844.630.013.4
Centre74.365.855.656.4
Separate-add76.066.256.457.7
Interval-add76.366.556.958.8
Separate-cat76.867.457.158.4
Interval-cat (proposed)77.167.257.558.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.53, + 0.216, + 0.862, + 0.23 + ], + "angle": 0, + "content": "Table 7. Ablating the choice of encoding time intervals." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.248, + 0.892, + 0.278 + ], + "angle": 0, + "content": "extreme in long actions \\((-14.5\\%\\) and \\(-11.2\\%)\\). We assess the impact of scaling the time interval in the ArXiv appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.892, + 0.446 + ], + "angle": 0, + "content": "Time Interval Encodings. The Time Interval MLP encodes the interval of the query. Here, we compare this to traditional positional encodings, both sinusoidal and learned. We also experiment on five different variations of the Time Interval MLP, namely: (i) Centre - we only encode the centre timestamp of the interval; (ii) Separate-Add/Cat - we encode the intervals' start and end time separately and add the encoded output vectors together, or concatenate along the channel dimension; and (iii) Interval-Add/Cat - we encode the start and end time within the same vector and add, or concatenate, the encoded output to the input sequence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.537 + ], + "angle": 0, + "content": "We show the results in Table 7. In all cases, the final encoding is of the same dimension for comparable results. Performance is significantly worse with sinusoidal or learned positional encoding, as they are unable to capture the complexity of overlapping actions. There is also a drop when encoding only the centre of the time interval." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.628 + ], + "angle": 0, + "content": "Separate-Add/Cat are alternative ways to encode the intervals (and hence include duration information) resulting in comparable performance to the interval counterparts. Our proposed approach to encoding the interval into the MLP shows the best performance for visual and while maintaining strong auditory performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.642, + 0.627, + 0.658 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.667, + 0.892, + 0.772 + ], + "angle": 0, + "content": "In this paper, we propose to utilise the action's time interval as a query to an audio-visual transformer which learns to recognise the action from its interval and the unaltered surrounding context. We jointly train the model on modality-specific time intervals and label sets, allowing the Time Interval Machine (TIM) to recognise multiple events across both visual and auditory modalities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.773, + 0.892, + 0.817 + ], + "angle": 0, + "content": "TIM is sensitive to the interval's position and duration. This allows the model, as is, to produce competitive results on action detection through multi-scale dense querying." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.818, + 0.892, + 0.887 + ], + "angle": 0, + "content": "Acknowledgements. This work uses public datasets. It is supported by EPSRC Doctoral Training Program, EPSRC UMPIRE EP/T004991/1 and EPSRC Programme Grant VisualAI EP/T028572/1; and by the use of the EPSRC funded Tier 2 facility JADE-II." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18160" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.229, + 0.47, + 0.284 + ], + "angle": 0, + "content": "[3] Swapnil Bhosale, Sauradip Nag, Diptesh Kanojia, Jiankang Deng, and Xiatian Zhu. Diffsed: Sound event detection with denoising diffusion. arXiv preprint arXiv:2308.07293, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.47, + 0.4 + ], + "angle": 0, + "content": "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.471, + 0.47 + ], + "angle": 0, + "content": "[6] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.47, + 0.556 + ], + "angle": 0, + "content": "[7] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.558, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[8] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast Networks for Video Recognition. In Proceedings of International Conference on Computer Vision (ICCV), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.615, + 0.47, + 0.669 + ], + "angle": 0, + "content": "[9] Fan Feng, Yue Ming, Nannan Hu, Hui Yu, and Yuanan Liu. CSS-net: A consistent segment selection network for audiovisual event localization. IEEE Transactions on Multimedia, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.47, + 0.727 + ], + "angle": 0, + "content": "[10] Ruohan Gao, Tae-Hyun Oh, Kristen Grauman, and Lorenzo Torresani. Listen to look: Action recognition by previewing audio. In Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.47, + 0.784 + ], + "angle": 0, + "content": "[11] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A Single Model for Many Visual Modalities. In CVPR, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.786, + 0.47, + 0.842 + ], + "angle": 0, + "content": "[12] Yuan Gong, Cheng-I Lai, Yu-An Chung, and James Glass. Ssast: Self-supervised audio spectrogram transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10699-10709, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.844, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[13] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Roei Herzig, Ofir Abramovich, Elad Ben-Avraham, Assaf Arbelle, Leonid Karlinsky, Ariel Shamir, Trevor Darrell, and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Amir Globerson. Promptonomyvit: Multi-task prompt learning improves video transformers using synthetic scene data. arXiv preprint arXiv:2212.04821, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[15] Jaesung Huh, Jacob Chalk, Evangelos Kazakos, Dima Damen, and Andrew Zisserman. EPIC-SOUNDS: A LargeScale Dataset of Actions that Sound. In IEEE International Conference on Acoustics, Speech, & Signal Processing (ICASSP), 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[16] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[17] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. Epic-fusion: Audio-visual temporal binding for egocentric action recognition. In Proceedings of International Conference on Computer Vision (ICCV), 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[18] Evangelos Kazakos, Jaesung Huh, Arsha Nagrani, Andrew Zisserman, and Dima Damen. With a little help from my temporal context: Multimodal egocentric action recognition. In Proc. BMVC, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[19] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. Slow-fast auditory streams for audio recognition. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[20] Dan Kondratyuk, Liangzhe Yuan, Yandong Li, Li Zhang, Mingxing Tan, Matthew Brown, and Boqing Gong. Movinets: Mobile video networks for efficient video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16020-16030, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[21] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco Nadai. Efficient training of visual transformers with small datasets. Advances in Neural Information Processing Systems, 34, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[22] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3202-3211, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[23] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[25] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12475-12486, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[26] Arsha Nagrani, Shan Yang, Anurag Arnab, Cordelia Schmid, and Chen Sun. Attention bottlenecks for multimodal fusion. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Yan Bin Ng and Basura Fernando. Human action sequence classification. CoRR, abs/1910.02602, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18161" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[28] Daniel S Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D Cubuk, and Quoc V Le. Specaugment: A simple data augmentation method for automatic speech recognition. arXiv preprint arXiv:1904.08779, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.246 + ], + "angle": 0, + "content": "[29] Mandela Patrick, Dylan Campbell, Yuki M. Asano, Ishaan Misra Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and João F. Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.248, + 0.472, + 0.373 + ], + "angle": 0, + "content": "[30] Viorica Pătrăucean, Lucas Smaira, Ankush Gupta, Adrià Recasens Continente, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Joseph Heyward, Mateusz Malinowski, Yi Yang, Carl Doersch, Tatiana Matejovicova, Yury Sulsky, Antoine Miech, Alex Frechette, Hanna Klimczak, Raphael Koster, Junlin Zhang, Stephanie Winkler, Yusuf Aytar, Simon Osindero, Dima Damen, Andrew Zisserman, and João Carreira. Perception test: A diagnostic benchmark for multimodal video models, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.472, + 0.43 + ], + "angle": 0, + "content": "[31] Fadime Sener, Dipika Singhania, and Angela Yao. Temporal aggregate representations for long-range video understanding. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.432, + 0.472, + 0.487 + ], + "angle": 0, + "content": "[32] Divya Shanmugam, Davis Blalock, Guha Balakrishnan, and John Guttag. Better aggregation in test-time augmentation. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.472, + 0.545 + ], + "angle": 0, + "content": "[33] Jonathan C Stroud, Zhichao Lu, Chen Sun, Jia Deng, Rahul Sukthankar, Cordelia Schmid, and David A Ross. Learning video representations from textual web supervision. arXiv preprint arXiv:2007.14937, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.472, + 0.616 + ], + "angle": 0, + "content": "[34] Swathikiran Sudhakaran, Adrian Bulat, Juan-Manuel Perez-Rua, Alex Falcon, Sergio Escalera, Oswald Lanz, Brais Martinez, and Georgios Tzimiropoulos. Saic.cambridge-hupba-fbk submission to the epic-kitchens-100 action recognition challenge 2021. arXiv preprint arXiv:2110.02902, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.472, + 0.686 + ], + "angle": 0, + "content": "[35] Tsung-Ming Tai, Oswald Lanz, Giuseppe Fiameni, Yi-Kwan Wong, Sze-Sen Poon, Cheng-Kuang Lee, Ka-Chun Cheung, and Simon See. Nvidia-unibz submission for epic-kitchens-100 action anticipation challenge 2022. arXiv preprint arXiv:2206.10869, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.472, + 0.729 + ], + "angle": 0, + "content": "[36] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In ECCV, 2018. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[37] Yapeng Tian, Dingzeyu Li, and Chenliang Xu. Unified multisensory perception: Weakly-supervised audio-visual video parsing. In Proc. ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.472, + 0.83 + ], + "angle": 0, + "content": "[38] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In Advances in Neural Information Processing Systems, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[39] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.188 + ], + "angle": 0, + "content": "[40] Luyu Wang, Pauline Luc, Yan Wu, Adria Recasens, Lucas Smaira, Andrew Brock, Andrew Jaegle, Jean-Baptiste Alayrac, Sander Dieleman, Joao Carreira, et al. Towards learning universal audio representations. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4593-4597. IEEE, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.893, + 0.246 + ], + "angle": 0, + "content": "[41] Weiyao Wang, Du Tran, and Matt Feiszli. What makes training multi-modal classification networks hard? In Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.894, + 0.317 + ], + "angle": 0, + "content": "[42] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 284-293, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.319, + 0.894, + 0.401 + ], + "angle": 0, + "content": "[43] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.403, + 0.893, + 0.457 + ], + "angle": 0, + "content": "[44] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.46, + 0.893, + 0.502 + ], + "angle": 0, + "content": "[45] Xuehan Xiong, Anurag Arnab, Arsha Nagrani, and Cordelia Schmid. M&m mix: A multimodal multiview transformer ensemble. arXiv preprint arXiv:2206.09852, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.893, + 0.544 + ], + "angle": 0, + "content": "[46] Mengmeng Xu, Chen Zhao, David S. Rojas, Ali Thabet, and Bernard Ghanem. G-tad: Sub-graph localization for temporal action detection, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.546, + 0.893, + 0.615 + ], + "angle": 0, + "content": "[47] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3333-3343, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.893, + 0.657 + ], + "angle": 0, + "content": "[48] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Temporal query networks for fine-grained video understanding. In Proc. CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.893, + 0.702 + ], + "angle": 0, + "content": "[49] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, pages 492-510, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.893, + 0.744 + ], + "angle": 0, + "content": "[50] Yue Zhao and Philipp Kähenbuhl. Training a large video model on a single machine in a day. arXiv preprint arXiv:2309.16669, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.893, + 0.787 + ], + "angle": 0, + "content": "[51] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.893, + 0.856 + ], + "angle": 0, + "content": "[52] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In The AAAI Conference on Artificial Intelligence (AAAI), pages 12993-13000, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[53] Jinxing Zhou, Liang Zheng, Yiran Zhong, Shijie Hao, and Meng Wang. Positive sample propagation along the audiovisual event line. In Proceedings of the IEEE/CVF Con-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18162" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.119 + ], + "angle": 0, + "content": "ference on Computer Vision and Pattern Recognition, pages 8436-8444, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[54] Jinxing Zhou, Dan Guo, and Meng Wang. Contrastive positive sample propagation along the audio-visual event line. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18163" + } + ] +] \ No newline at end of file diff --git a/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/03af8456-92c0-4a22-a6d9-94775506b39a_origin.pdf b/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/03af8456-92c0-4a22-a6d9-94775506b39a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..985e917ac72cfe1bf23c169053eb2cf2bd08d63f --- /dev/null +++ b/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/03af8456-92c0-4a22-a6d9-94775506b39a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3b21959acc21e8f8d9ee66d3b34fd10e8d3e365068e05cd3fcdfcb918b6858f +size 8367638 diff --git a/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/full.md b/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9783fc23aee20e33058af4f68da09d4a97105c6d --- /dev/null +++ b/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/full.md @@ -0,0 +1,343 @@ +# TIM: A Time Interval Machine for Audio-Visual Action Recognition + +Jacob Chalk $^{1*}$ + +Jaesung Huh $^{2*}$ + +Andrew Zisserman + +Evangelos Kazakos3 + +Dima Damen + +$^{1}$ University of Bristol + +$^{2}$ VGG, University of Oxford + +3 Czech Technical University in Prague + +# Abstract + +Diverse actions give rise to rich audio-visual signals in long videos. Recent works showcase that the two modalities of audio and video exhibit different temporal extents of events and distinct labels. We address the interplay between the two modalities in long videos by explicitly modelling the temporal extents of audio and visual events. We propose the Time Interval Machine (TIM) where a modality-specific time interval poses as a query to a transformer encoder that ingests a long video input. The encoder then attends to the specified interval, as well as the surrounding context in both modalities, in order to recognise the ongoing action. + +We test TIM on three long audio-visual video datasets: EPIC-KITCHENS, Perception Test, and AVE, reporting state-of-the-art (SOTA) for recognition. On EPIC-KITCHENS, we beat previous SOTA that utilises LLMs and significantly larger pre-training by $2.9\%$ top-1 action recognition accuracy. Additionally, we show that TIM can be adapted for action detection, using dense multi-scale interval queries, outperforming SOTA on EPIC-KITCHENS-100 for most metrics, and showing strong performance on the Perception Test. Our ablations show the critical role of integrating the two modalities and modelling their time intervals in achieving this performance. Code and models at: https://github.com/JacobChalk/TIM. + +# 1. Introduction + +Long videos exhibit a quick succession of auditory and visual events. The latest attempts to annotate events in these modalities separately [15, 30], showcase that both the temporal extents and class labels differ between the two. However, these events still remain correlated - identifying temporally close events in both modalities can improve recognition of actions in both visual and audio. + +Furthermore, most methods to date typically only utilise the exact temporal extent of an action; a precise, trimmed clip of the action is fed into a convolutional [5, 8, 39] or + +![](images/38991d2abc51082a7ec61d91e08a9075d5fcefb1634a1e45e00d8fdb8a0d655f.jpg) +Figure 1. Time Interval Machine (TIM): Top: Given a visual and auditory stream input, the ongoing action in a particular time interval is determined by a query specifying the start and end time of the interval, along with the modality of interest. Bottom: TIM can query for visual (e.g. 'Rinse Sponge') and auditory (e.g. 'Water') action classes, as well as distinguish between overlapping actions within the same modality ('Glass Collision' and 'Scrub / Scrape'). + +transformer-based [1, 11, 22] backbone, which predicts the action taking place. Even when the surrounding context is utilised to improve action recognition [18, 31, 42], again, this context is supplied in the form of exact clips of neighbouring actions, rather than the untrimmed long input video. + +In this paper, we propose an approach that encodes multiple events that occur in both visual and auditory streams of a long video input. We achieve this by elevating time intervals to first-class citizens, utilising them to specify a query within an accompanying modality. We term this mechanism a Time Interval Machine (TIM). It is able to receive a long video input, and output the actions that occur within the queried intervals of the queried modalities. + +Consider the example in Figure 1. The input contains the sound of water running while a sponge is being rinsed, which is then used to wipe a surface. These distinct events may vary significantly in duration and may be more prominent in the audio or visual modality. Despite differences between these events, there are likely many correlations between them and the surrounding context, which may be beneficial to recognising a given event (e.g. the sound of water is relevant to rinsing a sponge, providing useful information for recognising the visual action). TIM is able + +to exploit this by accessing the context within both modalities, including the background when no events occur. It can then distinguish between different, potentially overlapping, events within the same input by querying the time interval of a particular event within a given modality. + +We test TIM on three challenging audio-visual recognition datasets consisting of long videos: EPIC-KITCHENS [7], which recently offered distinct audio annotations through EPIC-SOUNDS [15], the Perception Test [30], and AVE [36]. We show that TIM can effectively learn both visual and auditory classes within a long input, outperforming the current SOTA top-1 accuracy on EPIC-KITCHENS by $2.9\%$ and $1.4\%$ on EPIC-SOUNDS, despite competing methods for the former using far larger pre-training datasets, large language models or higher resolution input. We also outperform models pre-trained with public datasets on AVE by $0.6\%$ and improve over a strong baseline on the Perception Test in visual and audio action recognition by $9.9\%$ and $3.2\%$ respectively. + +Additionally, we adapt TIM to action detection, through fixed multi-scale dense querying with an added interval regression loss. We report strong detection results on EPIC-KITCHENS and the Perception Test, outperforming Action Former [49] by 1.6 avg mAP and 4.3 respectively. + +Our contributions are summarised as: (i) we propose the TIM query mechanism for attending to modality-specific intervals in long videos. (ii) we efficiently train TIM to encode/query multiple audio-visual actions using time intervals. (iii) we showcase the value of TIM for both visual and auditory action recognition, and adapt it for detection with an added interval regression loss. (iv) we achieve new SOTA in both video and multi-modal recognition on multiple datasets. + +# 2. Related Works + +Audio-visual action recognition. A number of works have employed audio and visual modalities for action recognition [10, 17, 18, 26, 41, 44]. Some introduce new architectures to effectively fuse modalities [17, 18, 26, 44]; others propose unique training techniques to solve problems occurring while training multi-modal models, such as Gradient Blending [41], to tackle overfitting at different speeds for each modality, or contrastive learning for cross-modal discrimination [25]. However, these works use the same set of semantic and temporal labels for both modalities. Recent works have shown that both the temporal intervals and semantics of events differ between modalities [15, 30]. [37] temporally annotates visual and auditory events independently, although they share the same set of labels. In this work, we train with distinct labels for each modality to leverage discriminative audio and visual actions. + +Leveraging temporal context. Several works have considered incorporating temporal context [18, 27, 42, 43, 48], a + +direction orthogonal to employing multiple modalities and particularly useful in untrimmed videos. An auto-regressive LSTM-based encoder-decoder is proposed in [27] for action sequence classification, effectively leveraging past action context to predict the current action. The Temporal Query Network [48] uses learnable query vectors that correspond to specific attributes of a long video, allowing the model to attend to the aspects of the video and its surrounding context to produce a response for each attribute. [42] proposes to enhance the representation of the action by aggregating temporal context from neighbouring action clips using a Long-Term Feature Bank along with an attention mechanism. [43] crafts a more sophisticated memory bank by storing keys and values of all the intermediate layers of a transformer to aggregate the past context. Lastly, [18] exploits multi-modal temporal context from surrounding actions using vision, audio, and language. + +[18, 42, 43] are the closest to our approach, in that the common goal is to enrich the representation of the action of interest using surrounding context from the untrimmed video, rather than neighbouring clips. Nevertheless, [42, 43] are single modality models, recognising visual actions solely. [18] assumes the temporal extents of all actions are known, including for the test set, which is restrictive. + +Queries in visual models. Learning visual queries with Transformer architectures has gained recent attention [4, 14, 16, 23, 48]. Commonly, approaches employ a set of learnable vectors that are used to inquire about the presence of a concept in the input. For example, in [4, 23] the learnable queries correspond to different objects, whereas in [14] they are used for multi-task learning and each learnable query corresponds to a different task. [16] has incorporated learnable queries for adapting a pre-trained model while keeping the rest of its parameters frozen. Closest to our motivation is [48], where the queries correspond to events and their attributes for fine-grained action recognition in videos. The authors note that the queries also have the role of temporally localising the events in untrimmed videos. + +Different from [48] and other works, our queries are primarily temporal with no semantic interpretation and are applied to multiple modalities. Importantly, since time is continuous, we cannot use a predefined set of queries. Instead, we employ an MLP architecture to encode time, in a form akin to a universal clock. We present our approach next. + +# 3. Time Interval Machine + +In this section, we describe the Time Interval Machine (TIM), a multi-modal transformer encoder architecture where all inputs, both features and queries, are encoded with their associated time intervals. A time interval incorporates the duration and position of each audio and visual feature and is also used to query the network for any action occurring within the given time interval. + +![](images/9c05bfa74f8ae34037b09debdaf43548f11dbe523908bfd63294c4cd34f20c93.jpg) +Figure 2. Overview of the Time Interval Machine (TIM). The model ingests a sequence of audio and visual features from a video, with each feature time-stamped by the temporal interval it spans, and encoded with its modality. To infer the action occurring over a temporal interval (a visual or audio event) a query is formed specifying the interval and modality of interest. + +The architecture of TIM is illustrated in Figure 2. It ingests a large video input, represented as a sequence of audio and visual features, and outputs the ongoing auditory or visual action label for the provided query time intervals. + +# 3.1. Model architecture + +Input. The input to TIM is a long crop of the untrimmed video, represented by extracted features. When considering two modality inputs, such as video and audio, each modality is embedded separately as follows: for each modality $m$ , let $\mathbf{X}^m = [X_1^m, \dots, X_{N^m}^m]$ be $N^m$ temporally-ordered feature representations of the input video, obtained from a pre-trained feature extractor $S^m(\cdot)$ . We feed the features through modality-specific embedding layers $g^m(\cdot)$ , projecting them to a lower, common dimension $D$ across all modalities. The embedded features are then tagged with modality encodings and time interval encodings, forming the input to the transformer encoder. We now detail how we encode the time intervals. + +Encoding Time Intervals. In this work, we introduce a new type of learnt query network, the Time Interval MLP, which produces a single $D$ -dimensional vector representing a given time interval. This network is used within TIM to encode the time intervals of the input features and the time interval we wish to query, and later classify. Figure 3 illustrates the concept of this network. + +The Time Interval MLP $I(\cdot): \mathbb{R}^2 \to \mathbb{R}^D$ receives a time interval, represented by start and end time, and produces a single $D$ -dimensional encoding. Note that this is distinct from encoding the start and end times separately. Specifically, let $t_s$ and $t_e$ be the start and end time of an interval of interest, normalised by the length of the long video input. + +![](images/0bce42347eaca559eb4fe3ccc488ad575a1b99c5d3e336b01c07d63c6037b55a.jpg) +Figure 3. Illustration of the Time Interval MLP $I(\cdot)$ . It inputs the two-dimensional vector, start and end times of an interval, and produces a single vector, which can be concatenated along the channel dimension to either input features or [CLS] tokens. The figure shows three time interval inputs and three corresponding outputs. Note that in practice, time intervals are ingested simultaneously. + +$I(\cdot)$ receives the interval $\tilde{t} = [t_s, t_e]$ as input, and outputs a $D$ -dimensional vector encoding of that interval. This vector encodes both the relative position of the time interval within the input, as well as its duration. This vector then acts as a query for the model concerning the action taking place within the interval. Furthermore, each feature $\{X_i^m\}$ spans a certain time interval within the input. Thus, it is important to also encode the time intervals of the features. + +In summary, the Time Interval MLP acts as an universal clock, which encodes the temporal extent of features, from any modality, within the input. Note that it is critical that the same Time Interval MLP is used for encoding all time intervals of the input features and queries across both modalities to accurately encode universal time. It is also important to note that Time interval MLP can cover continuous time intervals, whereas traditional positional encoding only covers a fixed set of positions of the input features. The Time Interval MLP is trained end-to-end along with the transformer. + +Transformer Feature Inputs. Let $\tilde{\mathbf{t}}^m = [\tilde{t}_1^m, \dots, \tilde{t}_{N^m}^m]$ + +be the corresponding time intervals of the video's features $\mathbf{X}^m$ from the modality $m$ . We inject the encoded time interval $I(\tilde{\mathbf{t}}^m)$ into the embedded features via channelwise concatenation. A learnable modality-specific encoding $e^m \in \mathbb{R}^{2D}$ is then summed to the temporally-encoded features to discriminate between each modality. In summary, the feature inputs $\mathbf{E}^m$ for TIM are computed by, + +$$ +E _ {i} ^ {m} = \left[ g ^ {m} \left(X _ {i} ^ {m}\right), I \left(\tilde {t} _ {i} ^ {m}\right) \right] + e ^ {m} \quad \forall i \in [ 1, \dots , N ^ {m} ] \tag {1} +$$ + +where $[\cdot ,\cdot ]$ indicates concatenation. + +Transformer Query Inputs. To query for an action within an interval of interest, we adopt a standard approach of appending a learnable classification token to the input sequence, $\mathsf{CLS}^m$ . If $\tilde{t}_Q$ is an interval of interest, we concatenate the time interval representation $I(\tilde{t}_Q)$ to this classification token along the channel dimension, which acts as a query for the network in order to predict the corresponding action happening within $\tilde{t}_Q$ . We also add the modality-specific encoding $e^m$ to each classification token, as a flag to distinguish between which modality we are querying. The encoded $[\mathsf{CLS}]^m$ tokens can be more formally defined as: + +$$ +[ \mathrm {C L S} ] ^ {m} = [ \mathrm {C L S} ^ {m}, I (\tilde {t} _ {Q}) ] + e ^ {m} (2) +$$ + +During training, we add a classification token for each action within the input video, resulting in multiple [CLS] tokens across both modalities. + +Transformer Encoder. We use a transformer encoder to perform self-attention on the input sequence to aggregate relevant temporal context and cross-modal relations. + +We form the transformer input sequence with the encoded feature inputs $\mathbf{E}^m$ and one or more classification tokens $[\mathrm{CLS}]^m$ , representing each time interval query, and feed these into the encoder. Note that we recognise all actions from any modality simultaneously by appending multiple $\mathrm{CLS}^m$ tokens to the input. The transformer output representation of $[\mathrm{CLS}]^m$ , namely $Z_{\mathrm{CLS}}^m$ , is then passed to the corresponding linear classifier to predict the action labels. + +Importantly, we use an attention mask to prevent queries from attending to one another, and we similarly prevent input features from attending to queries. This ensures each query is recognised without the privileged knowledge of any other query, or action boundary, during inference. + +# 3.2. Training and Testing in TIM + +To train TIM, we consider all long segments of $W$ seconds and stride $H_{w}$ across the entirety of the untrimmed videos. We randomly select batches from these. For each window, we query all annotated audio and visual actions that overlap with the window by more than $\delta = 0.2$ seconds. + +All queries in the window are encoded and concatenated to separate CLS tokens. To classify queries, let $h_{\mathrm{CLS}}^m(\cdot)$ be a linear classifier for modality $m$ , and let $\hat{y}_{\mathrm{CLS}}^m = h^m(Z_{\mathrm{CLS}}^m)$ + +be the predicted action of the output representation $Z_{\mathrm{CLS}}^{m}$ . We train TIM by using a cross-entropy classification loss $CE(\cdot)$ on the ground truth $y_{\mathrm{CLS}}^{m}$ by: + +$$ +L ^ {m} = \frac {1}{N _ {Q}} \sum^ {N _ {Q}} C E \left(\hat {y} _ {\mathrm {C L S}} ^ {m}, y _ {\mathrm {C L S}} ^ {m}\right) \tag {3} +$$ + +where $N_{Q}$ is the number of queries within the batch. + +Temporal Distance Loss. In addition to the standard classification loss, we introduce a Temporal Distance (TD) loss as an auxiliary loss for training TIM. Inspired by [21], where relative patch positions in token embeddings are learnt using self-supervision, we similarly train the network to take two transformer outputs and predict the elapsed time between their corresponding time intervals. + +Let the $\mathbf{Z}_{1:\sum_mN^m}$ be the transformer outputs of the features from all modalities. We randomly sample a set of feature pairs $\mathbb{B} \subset \mathbf{Z}_{1:\sum_mN^m}$ from these outputs, concatenate along the channel dimension and feed them to the temporal distance regression head $h_{\tilde{t}}(\cdot):\mathbb{R}^{4D} \to \mathbb{R}^1$ to predict the time interval difference between each pair. Note that feature pairs can be sampled both within and across modalities. In our case, we sample across modalities by pairing one visual feature with another audio feature. This helps the model to learn the temporal relations between modalities. + +Formally, a TD loss $L^{td}$ is computed as: + +$$ +L ^ {t d} = \sum_ {\left\{Z _ {i}, Z _ {j} \right\} \in \mathbb {B}} \left| h _ {\tilde {i}} \left(Z _ {i}, Z _ {j}\right) - d _ {i j} \right| \tag {4} +$$ + +where $d_{ij}$ is the temporal distance between intervals $\tilde{t}_i,\tilde{t}_j$ + +Training objective and regime. For our final training loss, we sum the losses across modalities along with the TD loss: + +$$ +L ^ {t o t a l} = \left(\sum_ {m \in \mathbb {M}} \lambda^ {m} L ^ {m}\right) + \lambda^ {t d} L ^ {t d} \tag {5} +$$ + +where $\mathbb{M}$ is a set of modalities, $\lambda^m$ controls the strength of each modality's loss and $\lambda^{td}$ is a hyperparameter that controls the strength of the TD loss. + +Test-Time Augmentation. We use test-time augmentation, as this generally increases prediction robustness and performance [29, 32]. In TIM, we use a sliding window across the untrimmed video, thus feeding the same interval query with varying contexts. We then aggregate predictions of the same interval query across windows to make the final prediction. + +# 3.3. Adapting for Detection + +While primarily designed for recognition, we can adapt TIM for detection. The backbone remains largely unchanged from recognition, but there are two main differences. First, we construct dense multi-scale interval queries spanning the entirety of the video input at each scale. These + +are used as interval queries in both training and detection inference. The multi-scale intervals allow for detecting both long and short actions. Second, we introduce an additional interval regression head, which regresses the query interval to the action's exact temporal duration. + +During training, we deem any query in the multi-scale pyramid that overlaps with a ground truth action by more than some IoU threshold as a positive query. In addition to classifying the query, we train a DIOU regression loss [52] to predict the exact interval of the action. Both classification and interval regression losses are trained jointly. We provide full details in the ArXiv appendix. + +# 4. Experiments + +This section describes the datasets used to evaluate our model, implementation details and results along with a comparison to the state-of-the-art methods. + +# 4.1. Dataset + +EPIC-KITCHENS-100 [7] is a large-scale video dataset, including 700 egocentric videos recording actions in kitchens. It consists of 89,977 segments of fine-grained actions. Inspired by prior works [11, 34, 35], we directly predict the action out of 3806 classes present in the train and validation set to avoid predicting invalid actions. + +EPIC-SOUNDS [15] offers audio annotations capturing temporal extents and class labels within the audio stream of EPIC-KITCHENS-100. The annotations contain 78,366 labelled audio events. We combine the visual annotations from EPIC-KITCHENS with the audio annotations from EPIC-SOUNDS to train our audio-visual model. TIM can recognise actions from both datasets using a single model. + +AVE [36] contains 4,143 videos covering a range of real-life scenes and labelled with 27 categories, such as church bell, male speaking and dog barking. Each video is equally divided into 10 segments, each 1 second in length. We evaluate TIM on the supervised audio-visual event localisation task. Given a 1 second segment, we recognise the ongoing action out of the 27 categories plus a background class. + +Perception Test [30] is a recent multimodal video benchmark of 11,620 videos with an average length of 23 seconds, and provides both temporal action and sound annotations. There are 73,503 visual annotations spanning 63 classes, versus 137,128 sound annotations over 16 classes. + +# 4.2. Implementation Details + +Architectural Details. Visual and audio embedding layers $g_{m}$ consist of a single 512-D feed-forward layer, followed by a GELU [13] activation and layer normalisation [2] are used to project the features to a common space. The Time Interval MLP $I$ consists of three linear layers with 512-D hidden dimension, followed by ReLU activations, with + +layer normalisation after the output of the last linear layer. We include 512-D learnable [CLS] tokens: $[\mathrm{CLS}]_{\text {action }}^{m}$ for each query in each modality which become 1024-D after concatenation with the encoded time interval. They are then summed with 1024-D modality encodings; $e^{m}$ . + +The audio-visual transformer contains four encoder layers, each with 8 attention heads, GELU activations, and 1024-D keys, queries and values. A dropout rate of $p = 0.1$ is applied within the encoder layers. We also apply channelwise dropout with $p = 0.5$ directly to the raw input features, as well as to the encoded transformer input. The temporal distance head consists of two linear layers with a hidden dimension of 1024 and a third which outputs a single number corresponding to the elapsed time between each time interval. We include the architectural ablations on encoder layers and the temporal distance head in the ArXiv appendix. + +Training / Validation Details. We train each model for 100 epochs, using AdamW [24] with a batch size of 64 and a weight decay of 1e-4. A linear learning rate warm-up is applied for the first two epochs, starting from 1e-6 to a target learning rate, and we use a cosine learning rate scheduler. We set the TD loss weight $\lambda^{td}$ to 0.3. We pad the queries for each window in the batch to the maximum number of queries in a single window in each dataset. We provide implementation details per dataset in the ArXiv appendix. + +# 4.3. Results + +We compare TIM with SOTA models for each dataset. + +EPIC-KITCHENS / EPIC-SOUNDS Results. We train a single model on both visual and audio labels of the EPIC-KITCHENS videos, reporting results on both datasets. + +For the visual features, we concatenate Omnivore [11] and VideoMAE-L [38] features along the channel dimension, forming 2048-D features. For the audio features, we use Auditory SlowFast [19], which generalises well across diverse audio domains [40]. For both modalities, we extract 1 second features every 0.2s. For training, we extract additional augmented feature sets - with RandAugment [6] for visual and SpecAugment [28] for audio features. + +Table 1 compares TIM with the SOTA models on EPIC-KITCHCHENS-100. We outperform M&M Mix [45] by $5.1\%$ on verb, $0.9\%$ on noun and $3.9\%$ on action. Compared to our model, both MTV and M&M Mix are trained with an additional private dataset [33] which contains 194K hours of 70 million videos while we only use the open-source visual backbone pre-trained with public datasets. We also outperform LaViLa [51] and AVION [50] which leverage pre-trained LLMs to learn video representations. + +We note that we outperform all prior works, often without additional techniques that boost performance. For example, we use short-sided cropped $224 \times 224$ images while [45] uses $420 \times 420$ , which enlarges the spatial resolution of objects in the egocentric video, enabling better noun recog + +
ModelxpLLMVerbNounAction
Visual-only models
MFormer-HR [29]336p×67.058.544.5
MoViNet-A6 [20]320p×72.257.347.7
MeMViT [43]224p×71.460.348.4
Omnivore [11]224p×69.561.749.9
MTV [47]280p×69.963.950.5
LaViLa (TSF-L) [51]224p✓72.062.951.0
AVION (ViT-L) [50]224p✓73.065.454.4
TIM (ours)224p×76.266.456.4
Audio-visual models
TBN [17]224p×66.047.236.7
MBT [26]224p×64.858.043.4
MTCN [18]336p×70.762.149.6
M&M [45]420p×72.066.353.6
TIM (ours)224p×77.167.257.5
+ +Table 1. Comparisons to state-of-the-art recognition models on the EPIC-KITCHENS validation set. We report the top-1 accuracy for verb, noun and action (\%). LLM: large language model is used during pre-training. xp: input resolution of $x \times x$ . + +
ModelSSAST [12]ASF [19]DiffSED [3]TIM (A)TIM (A+V)
Top-1 acc53.553.856.955.758.3
+ +Table 2. Comparisons to state-of-the-art sound recognition models on EPIC-SOUNDS. We report the top-1 accuracy $(\%)$ on Val. The performance of SSAST and ASF are from [15]. + +
ModelVAAverage Precision (AP)
Task@0.1@0.2@0.3@0.4@0.5Avg.
G-TAD [46]✓✗Verb12.111.09.48.16.59.4
Noun11.010.08.67.05.48.4
ActionFormer [49]✓✗Verb26.625.424.222.319.123.5
Noun25.224.122.720.517.021.9
ActionFormer - Our Features✓✗Verb29.628.826.924.421.626.3
Noun34.332.630.227.422.629.4
TIM✓✓Verb32.931.629.627.022.228.6
Noun36.434.832.128.722.731.0
+ +nition. We expect a further performance boost when implementing any of: higher resolution feature extractors, additional large-scale pre-training and the introduction of a LLM. We leave this as an avenue for future work. + +Table 2 compares TIM against prior results on EPIC-SOUNDS, where TIM outperforms SOTA by $1.4\%$ . + +For detection, we show that TIM can produce competitive results when compared to models primarily designed for this task in Table 3. TIM adapted for detection outperforms ActionFormer [49] by $2.3\mathrm{mAP}$ on verb and $1.6\mathrm{mAP}$ on noun using the same set of features. + +AVE Results. As this dataset contains joint audio-visual labels, we train TIM by duplicating the query, i.e. using a [CLS] for each modality, and combine their logits during training and inference. We use the pre-trained publicly available models from [36] for a fair comparison with other + +Table 3. Comparisons to state-of-the-art detection models on the EPIC-KITCHENS validation set. We report the average precision at IOU thresholds [0.1, 0.2, 0.3, 0.4, 0.5] as well as their average across all thresholds on verb, noun. + +
ModelPSP [53]CPSP [54]CSSNet [9]†TIM
VAAVAV*
Top-1 acc77.878.680.562.865.579.279.8
+ +Table 4. Top-1 event classification accuracy $(\%)$ on AVE Test set. $\dagger$ : no official code or public model provided to replicate results. We show the models trained only with publicly available datasets. $\star$ : results with Omni+ASF features. + +
Perception Test Action
ModelMLP (V)MTCN [18](A+V)TIM (V)TIM (A+V)
Top-1 acc43.751.256.161.1
Perception Test Sound
ModelMLP (A)MTCN [18](A+V)TIM (A)TIM (A+V)
Top-1 acc50.652.954.856.1
+ +Table 5. Comparisons to trained recognition baselines on the Perception Test validation split. We show both action and sound recognition and the benefit of including audio-visual in TIM for both challenges. V : visual and A : audio input features. MLP is the result by training an MLP classifier with the features directly. + +
ModelAverage Precision (AP)
@0.1@0.2@0.3@0.4@0.5Avg.
Perception Test Action
ActionFormer [49]27.827.625.223.020.024.5
TIM33.532.229.826.422.028.8
Perception Test Sound
ActionFormer [49]34.731.327.522.717.726.8
TIM37.533.127.922.817.227.7
+ +Table 6. Comparisons to strong detection models on the Perception Test validation set for action and sound localisation. We report the average precision at IOU thresholds [0.1, 0.2, 0.3, 0.4, 0.5] as well as the average across all thresholds. + +works. We also apply AVGA [36] to spatial visual features from VGG-19 before feeding them to the transformer. + +Table 4 shows our results on the AVE dataset. Combining audio and video significantly improves the performance on TIM. The results from [9] perform best but could not be replicated. We also report TIM using the Omnivore visual features and Auditory Slowfast features used for EPIC-KITCHENS, which achieves a $0.6\%$ boost in performance. + +Perception Test Results. We use the same backbone for the Omnivore features and Auditory Slowfast features and train a single model using both the visual and audio labels. Table 5 compares the results on the newly introduced Perception Test. We train an MLP classifier, with two linear layers and a ReLU activation, on the features directly as a baseline. We also evaluate on an audio-visual model that does use context with MTCN. Compared to these methods TIM clearly shows significant improvements. Results are improved over MTCN by $9.9\%$ and $3.2\%$ on visual and audio recognition tasks respectively. We also provide detection results in Table 6. TIM improves over ActionFormer [49] by 3.3 average mAP on visual actions and by 0.9 average mAP on sound, when using the same features. + +![](images/99ebabf585cd33faad664ece1324409a3e3f93644d45fcc2925113ac8b8ac581.jpg) +Figure 4. Qualitative results for all datasets. PRED: Prediction by TIM, TIQ: Time Interval Queries, V/AGT: Visual/Audio Ground Truth. + +![](images/bd10edd49098b623f5b0f85b4effc0bf5bd1ec03e71ced203c1475089241fb67.jpg) + +Cross-Modality in TIM. When referring to our previous results, we see that including the additional modality provides a performance boost in all cases, highlighting TIM's ability to utilise and distinguish between different modalities. For example, on EPIC-KITCHENS-100, including audio improves visual action accuracy by $0.9\%$ . For EPICSOUNDS, the visual modality further improves accuracy by $2.6\%$ . In the Perception Test, including the audio modality improves visual recognition by $5.0\%$ , and visual increases sound recognition by $1.3\%$ . Finally, for AVE, we see a significant improvement, where an audio-visual model increases accuracy by $13.7\%$ from audio-only. + +Qualitative Results. We present qualitative results in Figure 4. We see that in EPIC-KITCHENS, TIM can competently recognise actions across the two modalities including overlapping queries. Furthermore, we see consecutive actions are correctly recognised with varying interval lengths, such as the 'open / close' audio actions between 0.2s and 4.5s. For AVE, TIM is able to distinguish between the background and a 'barking' audio-visual event based on the time interval query. For the Perception Test, we see that TIM can distinguish between heavily overlapping actions across both modalities, such as between 'break', 'human speech', 'hit' and 'put something'. However, there are also failure cases, such as in EPIC-KITCHENS when the action 'take washing up liquid' is recognised as 'wash hand', as the model is likely confused by the context predominantly associated with the highly overlapping 'wash hand' actions. + +# 4.4. Analysing Time Intervals + +We showcase the importance of effectively encoding time intervals, as well as how they differ from alternative strate- + +![](images/fcb9398663fa72e6ca36f8b903cb62b76c2adc38c0a8483e790fdd315583d383.jpg) +Figure 5. TSNE plot for time encodings $I(\cdot)$ on all datasets. In each plot, we use colour maps to indicate encodings of the time interval's duration (left), start time (middle) and end time (right). + +gies. We perform this analysis on the EPIC-KITCHENS-100 and EPIC-SOUNDS recognition tasks. + +Time Encoding Representation. To show TIM encodes time intervals across all datasets in Figure 5. We use three colour maps on the same TSNE projection to show three properties of the encoded interval: duration, start time and end time. Interestingly, the 1D time encoding perfectly captures all three attributes and across the datasets. While the encodings differ per dataset, as these differ in the positions and durations of actions, we see clear similarities in the learnt time encoding projections. For example, the duration is perfectly captured across the x-axis of the TSNE plot with lower values indicating longer time intervals. + +![](images/79bc2ccf94924ad8be6bfba4acdc35e7bc0366f47ffe04c2bbfe24b4e649eddd.jpg) +Figure 6. Attention heatmaps of the second encoder layer for two random 30s clips in EPIC-KITCHENS. x-axis: input feature time intervals; y-axis query time intervals of varying position and duration. The attention magnitude relates to the query CLS token. + +![](images/9796b37d53ca4aae70507d2799aa1a95b390944ed8618727740fb19ecdaf61c5.jpg) + +![](images/e962224ee5858897afd4669f2fb4f8108ffa74f8a1070e3beedcaa0e32feba15.jpg) +Figure 7. The impact of shifting the time interval query on visual performance (top) and audio performance (bottom), both on short actions $(< 2\mathrm{sec})$ and long actions $(>2\mathrm{sec})$ and overall validation set (all) for EPIC-KITCHENS-100 and EPIC-SOUNDS. + +Interval Query Attention. We plot two attention heatmaps in Figure 6 for 5 separate queries with varying positions and scales in EPIC-KITCHENS-100. We extract the attention weights from the second transformer encoder layer, as this appears to be the most relevant to the interval query. The learnt attention clearly applies to the feature time intervals contained within the query. We note the similarity between the attention in the two randomly selected windows. + +Shifting Intervals. To show how TIM effectively encodes the time interval of actions, we shift the time interval queries from their correct action interval by -1.5s to 1.5s, assessing the impact of these adjustments on the performance. + +Figure 7 shows the result. We see the performance gradually drops, both in visual and audio, as the query interval moves away from the correct action interval. The drop is also symmetric showcasing no bias. Unsurprisingly, the performance drops significantly when shifting short actions both in video $(-57.9\%)$ and audio $(-35.2\%)$ , while it is less + +
EPIC-KITCHENSEPIC-SOUNDS
EncodingVerbNounActionAudio Actions
Learned43.844.329.623.7
Sinusoidal43.844.630.013.4
Centre74.365.855.656.4
Separate-add76.066.256.457.7
Interval-add76.366.556.958.8
Separate-cat76.867.457.158.4
Interval-cat (proposed)77.167.257.558.3
+ +Table 7. Ablating the choice of encoding time intervals. + +extreme in long actions $(-14.5\%$ and $-11.2\%)$ . We assess the impact of scaling the time interval in the ArXiv appendix. + +Time Interval Encodings. The Time Interval MLP encodes the interval of the query. Here, we compare this to traditional positional encodings, both sinusoidal and learned. We also experiment on five different variations of the Time Interval MLP, namely: (i) Centre - we only encode the centre timestamp of the interval; (ii) Separate-Add/Cat - we encode the intervals' start and end time separately and add the encoded output vectors together, or concatenate along the channel dimension; and (iii) Interval-Add/Cat - we encode the start and end time within the same vector and add, or concatenate, the encoded output to the input sequence. + +We show the results in Table 7. In all cases, the final encoding is of the same dimension for comparable results. Performance is significantly worse with sinusoidal or learned positional encoding, as they are unable to capture the complexity of overlapping actions. There is also a drop when encoding only the centre of the time interval. + +Separate-Add/Cat are alternative ways to encode the intervals (and hence include duration information) resulting in comparable performance to the interval counterparts. Our proposed approach to encoding the interval into the MLP shows the best performance for visual and while maintaining strong auditory performance. + +# 5. Conclusions + +In this paper, we propose to utilise the action's time interval as a query to an audio-visual transformer which learns to recognise the action from its interval and the unaltered surrounding context. We jointly train the model on modality-specific time intervals and label sets, allowing the Time Interval Machine (TIM) to recognise multiple events across both visual and auditory modalities. + +TIM is sensitive to the interval's position and duration. This allows the model, as is, to produce competitive results on action detection through multi-scale dense querying. + +Acknowledgements. This work uses public datasets. It is supported by EPSRC Doctoral Training Program, EPSRC UMPIRE EP/T004991/1 and EPSRC Programme Grant VisualAI EP/T028572/1; and by the use of the EPSRC funded Tier 2 facility JADE-II. + +# References + +[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 1 +[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5 +[3] Swapnil Bhosale, Sauradip Nag, Diptesh Kanojia, Jiankang Deng, and Xiatian Zhu. Diffsed: Sound event detection with denoising diffusion. arXiv preprint arXiv:2308.07293, 2023. 6 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. ECCV, 2020. 2 +[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1 +[6] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 5 +[7] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 2021. 2, 5 +[8] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast Networks for Video Recognition. In Proceedings of International Conference on Computer Vision (ICCV), 2019. 1 +[9] Fan Feng, Yue Ming, Nannan Hu, Hui Yu, and Yuanan Liu. CSS-net: A consistent segment selection network for audiovisual event localization. IEEE Transactions on Multimedia, 2023. 6 +[10] Ruohan Gao, Tae-Hyun Oh, Kristen Grauman, and Lorenzo Torresani. Listen to look: Action recognition by previewing audio. In Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[11] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A Single Model for Many Visual Modalities. In CVPR, 2022. 1, 5, 6 +[12] Yuan Gong, Cheng-I Lai, Yu-An Chung, and James Glass. Ssast: Self-supervised audio spectrogram transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10699-10709, 2022. 6 +[13] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5 +[14] Roei Herzig, Ofir Abramovich, Elad Ben-Avraham, Assaf Arbelle, Leonid Karlinsky, Ariel Shamir, Trevor Darrell, and + +Amir Globerson. Promptonomyvit: Multi-task prompt learning improves video transformers using synthetic scene data. arXiv preprint arXiv:2212.04821, 2022. 2 +[15] Jaesung Huh, Jacob Chalk, Evangelos Kazakos, Dima Damen, and Andrew Zisserman. EPIC-SOUNDS: A LargeScale Dataset of Actions that Sound. In IEEE International Conference on Acoustics, Speech, & Signal Processing (ICASSP), 2023. 1, 2, 5, 6 +[16] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022. 2 +[17] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. Epic-fusion: Audio-visual temporal binding for egocentric action recognition. In Proceedings of International Conference on Computer Vision (ICCV), 2019. 2, 6 +[18] Evangelos Kazakos, Jaesung Huh, Arsha Nagrani, Andrew Zisserman, and Dima Damen. With a little help from my temporal context: Multimodal egocentric action recognition. In Proc. BMVC, 2021. 1, 2, 6 +[19] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. Slow-fast auditory streams for audio recognition. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021. 5, 6 +[20] Dan Kondratyuk, Liangzhe Yuan, Yandong Li, Li Zhang, Mingxing Tan, Matthew Brown, and Boqing Gong. Movinets: Mobile video networks for efficient video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16020-16030, 2021. 6 +[21] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco Nadai. Efficient training of visual transformers with small datasets. Advances in Neural Information Processing Systems, 34, 2021. 4 +[22] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3202-3211, 2022. 1 +[23] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2 +[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[25] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12475-12486, 2021. 2 +[26] Arsha Nagrani, Shan Yang, Anurag Arnab, Cordelia Schmid, and Chen Sun. Attention bottlenecks for multimodal fusion. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 6 +[27] Yan Bin Ng and Basura Fernando. Human action sequence classification. CoRR, abs/1910.02602, 2019. 2 + +[28] Daniel S Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D Cubuk, and Quoc V Le. Specaugment: A simple data augmentation method for automatic speech recognition. arXiv preprint arXiv:1904.08779, 2019. 5 +[29] Mandela Patrick, Dylan Campbell, Yuki M. Asano, Ishaan Misra Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and João F. Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 4, 6 +[30] Viorica Pătrăucean, Lucas Smaira, Ankush Gupta, Adrià Recasens Continente, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Joseph Heyward, Mateusz Malinowski, Yi Yang, Carl Doersch, Tatiana Matejovicova, Yury Sulsky, Antoine Miech, Alex Frechette, Hanna Klimczak, Raphael Koster, Junlin Zhang, Stephanie Winkler, Yusuf Aytar, Simon Osindero, Dima Damen, Andrew Zisserman, and João Carreira. Perception test: A diagnostic benchmark for multimodal video models, 2023. 1, 2, 5 +[31] Fadime Sener, Dipika Singhania, and Angela Yao. Temporal aggregate representations for long-range video understanding. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 1 +[32] Divya Shanmugam, Davis Blalock, Guha Balakrishnan, and John Guttag. Better aggregation in test-time augmentation. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 4 +[33] Jonathan C Stroud, Zhichao Lu, Chen Sun, Jia Deng, Rahul Sukthankar, Cordelia Schmid, and David A Ross. Learning video representations from textual web supervision. arXiv preprint arXiv:2007.14937, 2020. 5 +[34] Swathikiran Sudhakaran, Adrian Bulat, Juan-Manuel Perez-Rua, Alex Falcon, Sergio Escalera, Oswald Lanz, Brais Martinez, and Georgios Tzimiropoulos. Saic.cambridge-hupba-fbk submission to the epic-kitchens-100 action recognition challenge 2021. arXiv preprint arXiv:2110.02902, 2021. 5 +[35] Tsung-Ming Tai, Oswald Lanz, Giuseppe Fiameni, Yi-Kwan Wong, Sze-Sen Poon, Cheng-Kuang Lee, Ka-Chun Cheung, and Simon See. Nvidia-unibz submission for epic-kitchens-100 action anticipation challenge 2022. arXiv preprint arXiv:2206.10869, 2022. 5 +[36] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In ECCV, 2018. 2, 5, 6 +[37] Yapeng Tian, Dingzeyu Li, and Chenliang Xu. Unified multisensory perception: Weakly-supervised audio-visual video parsing. In Proc. ECCV, 2020. 2 +[38] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In Advances in Neural Information Processing Systems, 2022. 5 +[39] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 1 + +[40] Luyu Wang, Pauline Luc, Yan Wu, Adria Recasens, Lucas Smaira, Andrew Brock, Andrew Jaegle, Jean-Baptiste Alayrac, Sander Dieleman, Joao Carreira, et al. Towards learning universal audio representations. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4593-4597. IEEE, 2022. 5 +[41] Weiyao Wang, Du Tran, and Matt Feiszli. What makes training multi-modal classification networks hard? In Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[42] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 284-293, 2019. 1, 2 +[43] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 2, 6 +[44] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740, 2020. 2 +[45] Xuehan Xiong, Anurag Arnab, Arsha Nagrani, and Cordelia Schmid. M&m mix: A multimodal multiview transformer ensemble. arXiv preprint arXiv:2206.09852, 2022. 5, 6 +[46] Mengmeng Xu, Chen Zhao, David S. Rojas, Ali Thabet, and Bernard Ghanem. G-tad: Sub-graph localization for temporal action detection, 2020. 6 +[47] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3333-3343, 2022. 6 +[48] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Temporal query networks for fine-grained video understanding. In Proc. CVPR, 2021. 2 +[49] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, pages 492-510, 2022. 2, 6 +[50] Yue Zhao and Philipp Kähenbuhl. Training a large video model on a single machine in a day. arXiv preprint arXiv:2309.16669, 2023. 5, 6 +[51] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, 2023. 5, 6 +[52] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In The AAAI Conference on Artificial Intelligence (AAAI), pages 12993-13000, 2020. 5 +[53] Jinxing Zhou, Liang Zheng, Yiran Zhong, Shijie Hao, and Meng Wang. Positive sample propagation along the audiovisual event line. In Proceedings of the IEEE/CVF Con- + +ference on Computer Vision and Pattern Recognition, pages 8436-8444, 2021. 6 +[54] Jinxing Zhou, Dan Guo, and Meng Wang. Contrastive positive sample propagation along the audio-visual event line. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 6 \ No newline at end of file diff --git a/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/images.zip b/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d0ff97d65f613d5f79e55f88b8cf2397c7874fab --- /dev/null +++ b/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2e40677e8359bbe052343d9cae44a66e0d00842768cac0b5b108dd43c75e75e +size 696513 diff --git a/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/layout.json b/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e0bc0235c719817caf90b755b0f69c8d5f206330 --- /dev/null +++ b/2024/TIM_ A Time Interval Machine for Audio-Visual Action Recognition/layout.json @@ -0,0 +1,9244 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 507, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 507, + 121 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 507, + 121 + ], + "type": "text", + "content": "TIM: A Time Interval Machine for Audio-Visual Action Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 208, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 208, + 157 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 208, + 157 + ], + "type": "text", + "content": "Jacob Chalk" + }, + { + "bbox": [ + 138, + 142, + 208, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 244, + 143, + 317, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 143, + 317, + 157 + ], + "spans": [ + { + "bbox": [ + 244, + 143, + 317, + 157 + ], + "type": "text", + "content": "Jaesung Huh" + }, + { + "bbox": [ + 244, + 143, + 317, + 157 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 194, + 158, + 291, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 158, + 291, + 171 + ], + "spans": [ + { + "bbox": [ + 194, + 158, + 291, + 171 + ], + "type": "text", + "content": "Andrew Zisserman" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 354, + 143, + 455, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 143, + 455, + 157 + ], + "spans": [ + { + "bbox": [ + 354, + 143, + 455, + 157 + ], + "type": "text", + "content": "Evangelos Kazakos3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 329, + 158, + 399, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 158, + 399, + 171 + ], + "spans": [ + { + "bbox": [ + 329, + 158, + 399, + 171 + ], + "type": "text", + "content": "Dima Damen" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 171, + 172, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 171, + 172, + 186 + ], + "spans": [ + { + "bbox": [ + 66, + 171, + 172, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 171, + 172, + 186 + ], + "type": "text", + "content": "University of Bristol" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 184, + 171, + 324, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 171, + 324, + 185 + ], + "spans": [ + { + "bbox": [ + 184, + 171, + 324, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 184, + 171, + 324, + 185 + ], + "type": "text", + "content": "VGG, University of Oxford" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 335, + 171, + 527, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 171, + 527, + 186 + ], + "spans": [ + { + "bbox": [ + 335, + 171, + 527, + 186 + ], + "type": "text", + "content": "3 Czech Technical University in Prague" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 239, + 288, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 288, + 370 + ], + "type": "text", + "content": "Diverse actions give rise to rich audio-visual signals in long videos. Recent works showcase that the two modalities of audio and video exhibit different temporal extents of events and distinct labels. We address the interplay between the two modalities in long videos by explicitly modelling the temporal extents of audio and visual events. We propose the Time Interval Machine (TIM) where a modality-specific time interval poses as a query to a transformer encoder that ingests a long video input. The encoder then attends to the specified interval, as well as the surrounding context in both modalities, in order to recognise the ongoing action." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 45, + 371, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 371, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 45, + 371, + 287, + 526 + ], + "type": "text", + "content": "We test TIM on three long audio-visual video datasets: EPIC-KITCHENS, Perception Test, and AVE, reporting state-of-the-art (SOTA) for recognition. On EPIC-KITCHENS, we beat previous SOTA that utilises LLMs and significantly larger pre-training by " + }, + { + "bbox": [ + 45, + 371, + 287, + 526 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 45, + 371, + 287, + 526 + ], + "type": "text", + "content": " top-1 action recognition accuracy. Additionally, we show that TIM can be adapted for action detection, using dense multi-scale interval queries, outperforming SOTA on EPIC-KITCHENS-100 for most metrics, and showing strong performance on the Perception Test. Our ablations show the critical role of integrating the two modalities and modelling their time intervals in achieving this performance. Code and models at: https://github.com/JacobChalk/TIM." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 552, + 127, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 127, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 127, + 563 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 571, + 287, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 571, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 287, + 655 + ], + "type": "text", + "content": "Long videos exhibit a quick succession of auditory and visual events. The latest attempts to annotate events in these modalities separately [15, 30], showcase that both the temporal extents and class labels differ between the two. However, these events still remain correlated - identifying temporally close events in both modalities can improve recognition of actions in both visual and audio." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 657, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 657, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 657, + 287, + 693 + ], + "type": "text", + "content": "Furthermore, most methods to date typically only utilise the exact temporal extent of an action; a precise, trimmed clip of the action is fed into a convolutional [5, 8, 39] or" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 306, + 211, + 547, + 332 + ], + "blocks": [ + { + "bbox": [ + 306, + 211, + 547, + 332 + ], + "lines": [ + { + "bbox": [ + 306, + 211, + 547, + 332 + ], + "spans": [ + { + "bbox": [ + 306, + 211, + 547, + 332 + ], + "type": "image", + "image_path": "38991d2abc51082a7ec61d91e08a9075d5fcefb1634a1e45e00d8fdb8a0d655f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 338, + 547, + 416 + ], + "lines": [ + { + "bbox": [ + 305, + 338, + 547, + 416 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 547, + 416 + ], + "type": "text", + "content": "Figure 1. Time Interval Machine (TIM): Top: Given a visual and auditory stream input, the ongoing action in a particular time interval is determined by a query specifying the start and end time of the interval, along with the modality of interest. Bottom: TIM can query for visual (e.g. 'Rinse Sponge') and auditory (e.g. 'Water') action classes, as well as distinguish between overlapping actions within the same modality ('Glass Collision' and 'Scrub / Scrape')." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": "transformer-based [1, 11, 22] backbone, which predicts the action taking place. Even when the surrounding context is utilised to improve action recognition [18, 31, 42], again, this context is supplied in the form of exact clips of neighbouring actions, rather than the untrimmed long input video." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": "In this paper, we propose an approach that encodes multiple events that occur in both visual and auditory streams of a long video input. We achieve this by elevating time intervals to first-class citizens, utilising them to specify a query within an accompanying modality. We term this mechanism a Time Interval Machine (TIM). It is able to receive a long video input, and output the actions that occur within the queried intervals of the queried modalities." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Consider the example in Figure 1. The input contains the sound of water running while a sponge is being rinsed, which is then used to wipe a surface. These distinct events may vary significantly in duration and may be more prominent in the audio or visual modality. Despite differences between these events, there are likely many correlations between them and the surrounding context, which may be beneficial to recognising a given event (e.g. the sound of water is relevant to rinsing a sponge, providing useful information for recognising the visual action). TIM is able" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 161, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 161, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 161, + 712 + ], + "type": "text", + "content": "*Equal technical contribution." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18153" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": "to exploit this by accessing the context within both modalities, including the background when no events occur. It can then distinguish between different, potentially overlapping, events within the same input by querying the time interval of a particular event within a given modality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 299 + ], + "type": "text", + "content": "We test TIM on three challenging audio-visual recognition datasets consisting of long videos: EPIC-KITCHENS [7], which recently offered distinct audio annotations through EPIC-SOUNDS [15], the Perception Test [30], and AVE [36]. We show that TIM can effectively learn both visual and auditory classes within a long input, outperforming the current SOTA top-1 accuracy on EPIC-KITCHENS by " + }, + { + "bbox": [ + 46, + 133, + 286, + 299 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 46, + 133, + 286, + 299 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 133, + 286, + 299 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 46, + 133, + 286, + 299 + ], + "type": "text", + "content": " on EPIC-SOUNDS, despite competing methods for the former using far larger pre-training datasets, large language models or higher resolution input. We also outperform models pre-trained with public datasets on AVE by " + }, + { + "bbox": [ + 46, + 133, + 286, + 299 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 46, + 133, + 286, + 299 + ], + "type": "text", + "content": " and improve over a strong baseline on the Perception Test in visual and audio action recognition by " + }, + { + "bbox": [ + 46, + 133, + 286, + 299 + ], + "type": "inline_equation", + "content": "9.9\\%" + }, + { + "bbox": [ + 46, + 133, + 286, + 299 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 133, + 286, + 299 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 46, + 133, + 286, + 299 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 300, + 286, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 286, + 358 + ], + "type": "text", + "content": "Additionally, we adapt TIM to action detection, through fixed multi-scale dense querying with an added interval regression loss. We report strong detection results on EPIC-KITCHENS and the Perception Test, outperforming Action Former [49] by 1.6 avg mAP and 4.3 respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 360, + 286, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 286, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 286, + 466 + ], + "type": "text", + "content": "Our contributions are summarised as: (i) we propose the TIM query mechanism for attending to modality-specific intervals in long videos. (ii) we efficiently train TIM to encode/query multiple audio-visual actions using time intervals. (iii) we showcase the value of TIM for both visual and auditory action recognition, and adapt it for detection with an added interval regression loss. (iv) we achieve new SOTA in both video and multi-modal recognition on multiple datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 478, + 138, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 138, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 138, + 490 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 286, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 286, + 688 + ], + "type": "text", + "content": "Audio-visual action recognition. A number of works have employed audio and visual modalities for action recognition [10, 17, 18, 26, 41, 44]. Some introduce new architectures to effectively fuse modalities [17, 18, 26, 44]; others propose unique training techniques to solve problems occurring while training multi-modal models, such as Gradient Blending [41], to tackle overfitting at different speeds for each modality, or contrastive learning for cross-modal discrimination [25]. However, these works use the same set of semantic and temporal labels for both modalities. Recent works have shown that both the temporal intervals and semantics of events differ between modalities [15, 30]. [37] temporally annotates visual and auditory events independently, although they share the same set of labels. In this work, we train with distinct labels for each modality to leverage discriminative audio and visual actions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "Leveraging temporal context. Several works have considered incorporating temporal context [18, 27, 42, 43, 48], a" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "type": "text", + "content": "direction orthogonal to employing multiple modalities and particularly useful in untrimmed videos. An auto-regressive LSTM-based encoder-decoder is proposed in [27] for action sequence classification, effectively leveraging past action context to predict the current action. The Temporal Query Network [48] uses learnable query vectors that correspond to specific attributes of a long video, allowing the model to attend to the aspects of the video and its surrounding context to produce a response for each attribute. [42] proposes to enhance the representation of the action by aggregating temporal context from neighbouring action clips using a Long-Term Feature Bank along with an attention mechanism. [43] crafts a more sophisticated memory bank by storing keys and values of all the intermediate layers of a transformer to aggregate the past context. Lastly, [18] exploits multi-modal temporal context from surrounding actions using vision, audio, and language." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 275, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 545, + 358 + ], + "type": "text", + "content": "[18, 42, 43] are the closest to our approach, in that the common goal is to enrich the representation of the action of interest using surrounding context from the untrimmed video, rather than neighbouring clips. Nevertheless, [42, 43] are single modality models, recognising visual actions solely. [18] assumes the temporal extents of all actions are known, including for the test set, which is restrictive." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 360, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 525 + ], + "type": "text", + "content": "Queries in visual models. Learning visual queries with Transformer architectures has gained recent attention [4, 14, 16, 23, 48]. Commonly, approaches employ a set of learnable vectors that are used to inquire about the presence of a concept in the input. For example, in [4, 23] the learnable queries correspond to different objects, whereas in [14] they are used for multi-task learning and each learnable query corresponds to a different task. [16] has incorporated learnable queries for adapting a pre-trained model while keeping the rest of its parameters frozen. Closest to our motivation is [48], where the queries correspond to events and their attributes for fine-grained action recognition in videos. The authors note that the queries also have the role of temporally localising the events in untrimmed videos." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 527, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 527, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 545, + 598 + ], + "type": "text", + "content": "Different from [48] and other works, our queries are primarily temporal with no semantic interpretation and are applied to multiple modalities. Importantly, since time is continuous, we cannot use a predefined set of queries. Instead, we employ an MLP architecture to encode time, in a form akin to a universal clock. We present our approach next." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 609, + 440, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 440, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 440, + 621 + ], + "type": "text", + "content": "3. Time Interval Machine" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "In this section, we describe the Time Interval Machine (TIM), a multi-modal transformer encoder architecture where all inputs, both features and queries, are encoded with their associated time intervals. A time interval incorporates the duration and position of each audio and visual feature and is also used to query the network for any action occurring within the given time interval." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18154" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 547, + 259 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 259 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 259 + ], + "type": "image", + "image_path": "9c05bfa74f8ae34037b09debdaf43548f11dbe523908bfd63294c4cd34f20c93.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 266, + 547, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 547, + 300 + ], + "type": "text", + "content": "Figure 2. Overview of the Time Interval Machine (TIM). The model ingests a sequence of audio and visual features from a video, with each feature time-stamped by the temporal interval it spans, and encoded with its modality. To infer the action occurring over a temporal interval (a visual or audio event) a query is formed specifying the interval and modality of interest." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 308, + 288, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 288, + 357 + ], + "type": "text", + "content": "The architecture of TIM is illustrated in Figure 2. It ingests a large video input, represented as a sequence of audio and visual features, and outputs the ongoing auditory or visual action label for the provided query time intervals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 364, + 160, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 364, + 160, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 364, + 160, + 376 + ], + "type": "text", + "content": "3.1. Model architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 382, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 287, + 537 + ], + "type": "text", + "content": "Input. The input to TIM is a long crop of the untrimmed video, represented by extracted features. When considering two modality inputs, such as video and audio, each modality is embedded separately as follows: for each modality " + }, + { + "bbox": [ + 46, + 382, + 287, + 537 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 382, + 287, + 537 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 46, + 382, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^m = [X_1^m, \\dots, X_{N^m}^m]" + }, + { + "bbox": [ + 46, + 382, + 287, + 537 + ], + "type": "text", + "content": " be " + }, + { + "bbox": [ + 46, + 382, + 287, + 537 + ], + "type": "inline_equation", + "content": "N^m" + }, + { + "bbox": [ + 46, + 382, + 287, + 537 + ], + "type": "text", + "content": " temporally-ordered feature representations of the input video, obtained from a pre-trained feature extractor " + }, + { + "bbox": [ + 46, + 382, + 287, + 537 + ], + "type": "inline_equation", + "content": "S^m(\\cdot)" + }, + { + "bbox": [ + 46, + 382, + 287, + 537 + ], + "type": "text", + "content": ". We feed the features through modality-specific embedding layers " + }, + { + "bbox": [ + 46, + 382, + 287, + 537 + ], + "type": "inline_equation", + "content": "g^m(\\cdot)" + }, + { + "bbox": [ + 46, + 382, + 287, + 537 + ], + "type": "text", + "content": ", projecting them to a lower, common dimension " + }, + { + "bbox": [ + 46, + 382, + 287, + 537 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 382, + 287, + 537 + ], + "type": "text", + "content": " across all modalities. The embedded features are then tagged with modality encodings and time interval encodings, forming the input to the transformer encoder. We now detail how we encode the time intervals." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "content": "Encoding Time Intervals. In this work, we introduce a new type of learnt query network, the Time Interval MLP, which produces a single " + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 540, + 287, + 624 + ], + "type": "text", + "content": "-dimensional vector representing a given time interval. This network is used within TIM to encode the time intervals of the input features and the time interval we wish to query, and later classify. Figure 3 illustrates the concept of this network." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 623, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 623, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 623, + 288, + 696 + ], + "type": "text", + "content": "The Time Interval MLP " + }, + { + "bbox": [ + 46, + 623, + 288, + 696 + ], + "type": "inline_equation", + "content": "I(\\cdot): \\mathbb{R}^2 \\to \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 623, + 288, + 696 + ], + "type": "text", + "content": " receives a time interval, represented by start and end time, and produces a single " + }, + { + "bbox": [ + 46, + 623, + 288, + 696 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 623, + 288, + 696 + ], + "type": "text", + "content": "-dimensional encoding. Note that this is distinct from encoding the start and end times separately. Specifically, let " + }, + { + "bbox": [ + 46, + 623, + 288, + 696 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 46, + 623, + 288, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 623, + 288, + 696 + ], + "type": "inline_equation", + "content": "t_e" + }, + { + "bbox": [ + 46, + 623, + 288, + 696 + ], + "type": "text", + "content": " be the start and end time of an interval of interest, normalised by the length of the long video input." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 322, + 307, + 533, + 392 + ], + "blocks": [ + { + "bbox": [ + 322, + 307, + 533, + 392 + ], + "lines": [ + { + "bbox": [ + 322, + 307, + 533, + 392 + ], + "spans": [ + { + "bbox": [ + 322, + 307, + 533, + 392 + ], + "type": "image", + "image_path": "0bce42347eaca559eb4fe3ccc488ad575a1b99c5d3e336b01c07d63c6037b55a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 399, + 545, + 467 + ], + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 467 + ], + "type": "text", + "content": "Figure 3. Illustration of the Time Interval MLP " + }, + { + "bbox": [ + 304, + 399, + 545, + 467 + ], + "type": "inline_equation", + "content": "I(\\cdot)" + }, + { + "bbox": [ + 304, + 399, + 545, + 467 + ], + "type": "text", + "content": ". It inputs the two-dimensional vector, start and end times of an interval, and produces a single vector, which can be concatenated along the channel dimension to either input features or [CLS] tokens. The figure shows three time interval inputs and three corresponding outputs. Note that in practice, time intervals are ingested simultaneously." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 479, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 576 + ], + "type": "inline_equation", + "content": "I(\\cdot)" + }, + { + "bbox": [ + 304, + 479, + 545, + 576 + ], + "type": "text", + "content": " receives the interval " + }, + { + "bbox": [ + 304, + 479, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\tilde{t} = [t_s, t_e]" + }, + { + "bbox": [ + 304, + 479, + 545, + 576 + ], + "type": "text", + "content": " as input, and outputs a " + }, + { + "bbox": [ + 304, + 479, + 545, + 576 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 479, + 545, + 576 + ], + "type": "text", + "content": "-dimensional vector encoding of that interval. This vector encodes both the relative position of the time interval within the input, as well as its duration. This vector then acts as a query for the model concerning the action taking place within the interval. Furthermore, each feature " + }, + { + "bbox": [ + 304, + 479, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\{X_i^m\\}" + }, + { + "bbox": [ + 304, + 479, + 545, + 576 + ], + "type": "text", + "content": " spans a certain time interval within the input. Thus, it is important to also encode the time intervals of the features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 578, + 545, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 697 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 697 + ], + "type": "text", + "content": "In summary, the Time Interval MLP acts as an universal clock, which encodes the temporal extent of features, from any modality, within the input. Note that it is critical that the same Time Interval MLP is used for encoding all time intervals of the input features and queries across both modalities to accurately encode universal time. It is also important to note that Time interval MLP can cover continuous time intervals, whereas traditional positional encoding only covers a fixed set of positions of the input features. The Time Interval MLP is trained end-to-end along with the transformer." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 700, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 700, + 545, + 715 + ], + "type": "text", + "content": "Transformer Feature Inputs. Let " + }, + { + "bbox": [ + 306, + 700, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{t}}^m = [\\tilde{t}_1^m, \\dots, \\tilde{t}_{N^m}^m]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 265, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 265, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 265, + 712 + ], + "type": "text", + "content": "1Note that the number of features can differ between modalities" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18155" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "be the corresponding time intervals of the video's features " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^m" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " from the modality " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ". We inject the encoded time interval " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "I(\\tilde{\\mathbf{t}}^m)" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " into the embedded features via channelwise concatenation. A learnable modality-specific encoding " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "e^m \\in \\mathbb{R}^{2D}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " is then summed to the temporally-encoded features to discriminate between each modality. In summary, the feature inputs " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^m" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " for TIM are computed by," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 164, + 287, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 164, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 59, + 164, + 287, + 179 + ], + "type": "interline_equation", + "content": "E _ {i} ^ {m} = \\left[ g ^ {m} \\left(X _ {i} ^ {m}\\right), I \\left(\\tilde {t} _ {i} ^ {m}\\right) \\right] + e ^ {m} \\quad \\forall i \\in [ 1, \\dots , N ^ {m} ] \\tag {1}", + "image_path": "ac4712093ea32912c5082786ac6ddc531cfe67ec069764f5e2c22f21bed6fb3b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 186, + 192, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 186, + 192, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 192, + 198 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 186, + 192, + 198 + ], + "type": "inline_equation", + "content": "[\\cdot ,\\cdot ]" + }, + { + "bbox": [ + 47, + 186, + 192, + 198 + ], + "type": "text", + "content": " indicates concatenation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "text", + "content": "Transformer Query Inputs. To query for an action within an interval of interest, we adopt a standard approach of appending a learnable classification token to the input sequence, " + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "inline_equation", + "content": "\\mathsf{CLS}^m" + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "inline_equation", + "content": "\\tilde{t}_Q" + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "text", + "content": " is an interval of interest, we concatenate the time interval representation " + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "inline_equation", + "content": "I(\\tilde{t}_Q)" + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "text", + "content": " to this classification token along the channel dimension, which acts as a query for the network in order to predict the corresponding action happening within " + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "inline_equation", + "content": "\\tilde{t}_Q" + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "text", + "content": ". We also add the modality-specific encoding " + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "inline_equation", + "content": "e^m" + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "text", + "content": " to each classification token, as a flag to distinguish between which modality we are querying. The encoded " + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "inline_equation", + "content": "[\\mathsf{CLS}]^m" + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "text", + "content": " tokens can be more formally defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 340, + 287, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 340, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 100, + 340, + 287, + 356 + ], + "type": "interline_equation", + "content": "[ \\mathrm {C L S} ] ^ {m} = [ \\mathrm {C L S} ^ {m}, I (\\tilde {t} _ {Q}) ] + e ^ {m} (2)", + "image_path": "83785bb6c8c0ca7a57ec7ad745f4660f00a12851b338c77d565bedee419aa675.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 363, + 287, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 398 + ], + "type": "text", + "content": "During training, we add a classification token for each action within the input video, resulting in multiple [CLS] tokens across both modalities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 401, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 435 + ], + "type": "text", + "content": "Transformer Encoder. We use a transformer encoder to perform self-attention on the input sequence to aggregate relevant temporal context and cross-modal relations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 437, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 532 + ], + "type": "text", + "content": "We form the transformer input sequence with the encoded feature inputs " + }, + { + "bbox": [ + 46, + 437, + 287, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^m" + }, + { + "bbox": [ + 46, + 437, + 287, + 532 + ], + "type": "text", + "content": " and one or more classification tokens " + }, + { + "bbox": [ + 46, + 437, + 287, + 532 + ], + "type": "inline_equation", + "content": "[\\mathrm{CLS}]^m" + }, + { + "bbox": [ + 46, + 437, + 287, + 532 + ], + "type": "text", + "content": ", representing each time interval query, and feed these into the encoder. Note that we recognise all actions from any modality simultaneously by appending multiple " + }, + { + "bbox": [ + 46, + 437, + 287, + 532 + ], + "type": "inline_equation", + "content": "\\mathrm{CLS}^m" + }, + { + "bbox": [ + 46, + 437, + 287, + 532 + ], + "type": "text", + "content": " tokens to the input. The transformer output representation of " + }, + { + "bbox": [ + 46, + 437, + 287, + 532 + ], + "type": "inline_equation", + "content": "[\\mathrm{CLS}]^m" + }, + { + "bbox": [ + 46, + 437, + 287, + 532 + ], + "type": "text", + "content": ", namely " + }, + { + "bbox": [ + 46, + 437, + 287, + 532 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{CLS}}^m" + }, + { + "bbox": [ + 46, + 437, + 287, + 532 + ], + "type": "text", + "content": ", is then passed to the corresponding linear classifier to predict the action labels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 593 + ], + "type": "text", + "content": "Importantly, we use an attention mask to prevent queries from attending to one another, and we similarly prevent input features from attending to queries. This ensures each query is recognised without the privileged knowledge of any other query, or action boundary, during inference." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 599, + 204, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 204, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 204, + 612 + ], + "type": "text", + "content": "3.2. Training and Testing in TIM" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "content": "To train TIM, we consider all long segments of " + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "content": " seconds and stride " + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "inline_equation", + "content": "H_{w}" + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "content": " across the entirety of the untrimmed videos. We randomly select batches from these. For each window, we query all annotated audio and visual actions that overlap with the window by more than " + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\delta = 0.2" + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "content": " seconds." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "All queries in the window are encoded and concatenated to separate CLS tokens. To classify queries, let " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{CLS}}^m(\\cdot)" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " be a linear classifier for modality " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": ", and let " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{y}_{\\mathrm{CLS}}^m = h^m(Z_{\\mathrm{CLS}}^m)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": "be the predicted action of the output representation " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{CLS}}^{m}" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": ". We train TIM by using a cross-entropy classification loss " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "CE(\\cdot)" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " on the ground truth " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{CLS}}^{m}" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 118, + 545, + 149 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 118, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 361, + 118, + 545, + 149 + ], + "type": "interline_equation", + "content": "L ^ {m} = \\frac {1}{N _ {Q}} \\sum^ {N _ {Q}} C E \\left(\\hat {y} _ {\\mathrm {C L S}} ^ {m}, y _ {\\mathrm {C L S}} ^ {m}\\right) \\tag {3}", + "image_path": "d238c211334eecf3bf0de606abbd809c3f21dbc273c5e28904c282824a1adbb6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 156, + 518, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 156, + 518, + 168 + ], + "spans": [ + { + "bbox": [ + 305, + 156, + 518, + 168 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 156, + 518, + 168 + ], + "type": "inline_equation", + "content": "N_{Q}" + }, + { + "bbox": [ + 305, + 156, + 518, + 168 + ], + "type": "text", + "content": " is the number of queries within the batch." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 170, + 545, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 545, + 253 + ], + "type": "text", + "content": "Temporal Distance Loss. In addition to the standard classification loss, we introduce a Temporal Distance (TD) loss as an auxiliary loss for training TIM. Inspired by [21], where relative patch positions in token embeddings are learnt using self-supervision, we similarly train the network to take two transformer outputs and predict the elapsed time between their corresponding time intervals." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 253, + 546, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 546, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 546, + 372 + ], + "type": "text", + "content": "Let the " + }, + { + "bbox": [ + 304, + 253, + 546, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{1:\\sum_mN^m}" + }, + { + "bbox": [ + 304, + 253, + 546, + 372 + ], + "type": "text", + "content": " be the transformer outputs of the features from all modalities. We randomly sample a set of feature pairs " + }, + { + "bbox": [ + 304, + 253, + 546, + 372 + ], + "type": "inline_equation", + "content": "\\mathbb{B} \\subset \\mathbf{Z}_{1:\\sum_mN^m}" + }, + { + "bbox": [ + 304, + 253, + 546, + 372 + ], + "type": "text", + "content": " from these outputs, concatenate along the channel dimension and feed them to the temporal distance regression head " + }, + { + "bbox": [ + 304, + 253, + 546, + 372 + ], + "type": "inline_equation", + "content": "h_{\\tilde{t}}(\\cdot):\\mathbb{R}^{4D} \\to \\mathbb{R}^1" + }, + { + "bbox": [ + 304, + 253, + 546, + 372 + ], + "type": "text", + "content": " to predict the time interval difference between each pair. Note that feature pairs can be sampled both within and across modalities. In our case, we sample across modalities by pairing one visual feature with another audio feature. This helps the model to learn the temporal relations between modalities." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 372, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 482, + 384 + ], + "type": "text", + "content": "Formally, a TD loss " + }, + { + "bbox": [ + 317, + 372, + 482, + 384 + ], + "type": "inline_equation", + "content": "L^{td}" + }, + { + "bbox": [ + 317, + 372, + 482, + 384 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 353, + 394, + 545, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 394, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 353, + 394, + 545, + 422 + ], + "type": "interline_equation", + "content": "L ^ {t d} = \\sum_ {\\left\\{Z _ {i}, Z _ {j} \\right\\} \\in \\mathbb {B}} \\left| h _ {\\tilde {i}} \\left(Z _ {i}, Z _ {j}\\right) - d _ {i j} \\right| \\tag {4}", + "image_path": "4a024bc1c366433a88d22dcedc498db7828b8ad4a2543c6507b9b3095c7b4387.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 430, + 542, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 430, + 542, + 444 + ], + "spans": [ + { + "bbox": [ + 305, + 430, + 542, + 444 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 430, + 542, + 444 + ], + "type": "inline_equation", + "content": "d_{ij}" + }, + { + "bbox": [ + 305, + 430, + 542, + 444 + ], + "type": "text", + "content": " is the temporal distance between intervals " + }, + { + "bbox": [ + 305, + 430, + 542, + 444 + ], + "type": "inline_equation", + "content": "\\tilde{t}_i,\\tilde{t}_j" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 445, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 445, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 305, + 445, + 545, + 469 + ], + "type": "text", + "content": "Training objective and regime. For our final training loss, we sum the losses across modalities along with the TD loss:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 353, + 477, + 545, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 477, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 353, + 477, + 545, + 510 + ], + "type": "interline_equation", + "content": "L ^ {t o t a l} = \\left(\\sum_ {m \\in \\mathbb {M}} \\lambda^ {m} L ^ {m}\\right) + \\lambda^ {t d} L ^ {t d} \\tag {5}", + "image_path": "ae8881cf487a4add4de09c716647442de0c32207d998532d53970f3e5a56353b.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 518, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 554 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 518, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathbb{M}" + }, + { + "bbox": [ + 304, + 518, + 545, + 554 + ], + "type": "text", + "content": " is a set of modalities, " + }, + { + "bbox": [ + 304, + 518, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\lambda^m" + }, + { + "bbox": [ + 304, + 518, + 545, + 554 + ], + "type": "text", + "content": " controls the strength of each modality's loss and " + }, + { + "bbox": [ + 304, + 518, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\lambda^{td}" + }, + { + "bbox": [ + 304, + 518, + 545, + 554 + ], + "type": "text", + "content": " is a hyperparameter that controls the strength of the TD loss." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "text", + "content": "Test-Time Augmentation. We use test-time augmentation, as this generally increases prediction robustness and performance [29, 32]. In TIM, we use a sliding window across the untrimmed video, thus feeding the same interval query with varying contexts. We then aggregate predictions of the same interval query across windows to make the final prediction." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 635, + 436, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 436, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 436, + 647 + ], + "type": "text", + "content": "3.3. Adapting for Detection" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "While primarily designed for recognition, we can adapt TIM for detection. The backbone remains largely unchanged from recognition, but there are two main differences. First, we construct dense multi-scale interval queries spanning the entirety of the video input at each scale. These" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18156" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": "are used as interval queries in both training and detection inference. The multi-scale intervals allow for detecting both long and short actions. Second, we introduce an additional interval regression head, which regresses the query interval to the action's exact temporal duration." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 216 + ], + "type": "text", + "content": "During training, we deem any query in the multi-scale pyramid that overlaps with a ground truth action by more than some IoU threshold as a positive query. In addition to classifying the query, we train a DIOU regression loss [52] to predict the exact interval of the action. Both classification and interval regression losses are trained jointly. We provide full details in the ArXiv appendix." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 226, + 127, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 226, + 127, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 226, + 127, + 239 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 246, + 286, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 286, + 281 + ], + "type": "text", + "content": "This section describes the datasets used to evaluate our model, implementation details and results along with a comparison to the state-of-the-art methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 289, + 105, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 105, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 105, + 300 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 308, + 286, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 286, + 380 + ], + "type": "text", + "content": "EPIC-KITCHENS-100 [7] is a large-scale video dataset, including 700 egocentric videos recording actions in kitchens. It consists of 89,977 segments of fine-grained actions. Inspired by prior works [11, 34, 35], we directly predict the action out of 3806 classes present in the train and validation set to avoid predicting invalid actions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 383, + 286, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 383, + 286, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 383, + 286, + 466 + ], + "type": "text", + "content": "EPIC-SOUNDS [15] offers audio annotations capturing temporal extents and class labels within the audio stream of EPIC-KITCHENS-100. The annotations contain 78,366 labelled audio events. We combine the visual annotations from EPIC-KITCHENS with the audio annotations from EPIC-SOUNDS to train our audio-visual model. TIM can recognise actions from both datasets using a single model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 468, + 286, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 286, + 552 + ], + "type": "text", + "content": "AVE [36] contains 4,143 videos covering a range of real-life scenes and labelled with 27 categories, such as church bell, male speaking and dog barking. Each video is equally divided into 10 segments, each 1 second in length. We evaluate TIM on the supervised audio-visual event localisation task. Given a 1 second segment, we recognise the ongoing action out of the 27 categories plus a background class." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 554, + 286, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 286, + 613 + ], + "type": "text", + "content": "Perception Test [30] is a recent multimodal video benchmark of 11,620 videos with an average length of 23 seconds, and provides both temporal action and sound annotations. There are 73,503 visual annotations spanning 63 classes, versus 137,128 sound annotations over 16 classes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 621, + 179, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 179, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 179, + 634 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "text", + "content": "Architectural Details. Visual and audio embedding layers " + }, + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "inline_equation", + "content": "g_{m}" + }, + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "text", + "content": " consist of a single 512-D feed-forward layer, followed by a GELU [13] activation and layer normalisation [2] are used to project the features to a common space. The Time Interval MLP " + }, + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 641, + 286, + 712 + ], + "type": "text", + "content": " consists of three linear layers with 512-D hidden dimension, followed by ReLU activations, with" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "layer normalisation after the output of the last linear layer. We include 512-D learnable [CLS] tokens: " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "[\\mathrm{CLS}]_{\\text {action }}^{m}" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " for each query in each modality which become 1024-D after concatenation with the encoded time interval. They are then summed with 1024-D modality encodings; " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "e^{m}" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "type": "text", + "content": "The audio-visual transformer contains four encoder layers, each with 8 attention heads, GELU activations, and 1024-D keys, queries and values. A dropout rate of " + }, + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "type": "inline_equation", + "content": "p = 0.1" + }, + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "type": "text", + "content": " is applied within the encoder layers. We also apply channelwise dropout with " + }, + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "type": "inline_equation", + "content": "p = 0.5" + }, + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "type": "text", + "content": " directly to the raw input features, as well as to the encoded transformer input. The temporal distance head consists of two linear layers with a hidden dimension of 1024 and a third which outputs a single number corresponding to the elapsed time between each time interval. We include the architectural ablations on encoder layers and the temporal distance head in the ArXiv appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 266, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 373 + ], + "type": "text", + "content": "Training / Validation Details. We train each model for 100 epochs, using AdamW [24] with a batch size of 64 and a weight decay of 1e-4. A linear learning rate warm-up is applied for the first two epochs, starting from 1e-6 to a target learning rate, and we use a cosine learning rate scheduler. We set the TD loss weight " + }, + { + "bbox": [ + 304, + 266, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\lambda^{td}" + }, + { + "bbox": [ + 304, + 266, + 545, + 373 + ], + "type": "text", + "content": " to 0.3. We pad the queries for each window in the batch to the maximum number of queries in a single window in each dataset. We provide implementation details per dataset in the ArXiv appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 381, + 362, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 381, + 362, + 393 + ], + "spans": [ + { + "bbox": [ + 306, + 381, + 362, + 393 + ], + "type": "text", + "content": "4.3. Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 400, + 525, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 400, + 525, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 400, + 525, + 411 + ], + "type": "text", + "content": "We compare TIM with SOTA models for each dataset." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 414, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 414, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 414, + 545, + 449 + ], + "type": "text", + "content": "EPIC-KITCHENS / EPIC-SOUNDS Results. We train a single model on both visual and audio labels of the EPIC-KITCHENS videos, reporting results on both datasets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 450, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 544 + ], + "type": "text", + "content": "For the visual features, we concatenate Omnivore [11] and VideoMAE-L [38] features along the channel dimension, forming 2048-D features. For the audio features, we use Auditory SlowFast [19], which generalises well across diverse audio domains [40]. For both modalities, we extract 1 second features every 0.2s. For training, we extract additional augmented feature sets - with RandAugment [6] for visual and SpecAugment [28] for audio features." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "text", + "content": "Table 1 compares TIM with the SOTA models on EPIC-KITCHCHENS-100. We outperform M&M Mix [45] by " + }, + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "inline_equation", + "content": "5.1\\%" + }, + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "text", + "content": " on verb, " + }, + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "text", + "content": " on noun and " + }, + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 304, + 545, + 545, + 653 + ], + "type": "text", + "content": " on action. Compared to our model, both MTV and M&M Mix are trained with an additional private dataset [33] which contains 194K hours of 70 million videos while we only use the open-source visual backbone pre-trained with public datasets. We also outperform LaViLa [51] and AVION [50] which leverage pre-trained LLMs to learn video representations." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "We note that we outperform all prior works, often without additional techniques that boost performance. For example, we use short-sided cropped " + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": " images while [45] uses " + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "inline_equation", + "content": "420 \\times 420" + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": ", which enlarges the spatial resolution of objects in the egocentric video, enabling better noun recog" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18157" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 69, + 286, + 240 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 286, + 240 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 286, + 240 + ], + "type": "table", + "html": "
ModelxpLLMVerbNounAction
Visual-only models
MFormer-HR [29]336p×67.058.544.5
MoViNet-A6 [20]320p×72.257.347.7
MeMViT [43]224p×71.460.348.4
Omnivore [11]224p×69.561.749.9
MTV [47]280p×69.963.950.5
LaViLa (TSF-L) [51]224p✓72.062.951.0
AVION (ViT-L) [50]224p✓73.065.454.4
TIM (ours)224p×76.266.456.4
Audio-visual models
TBN [17]224p×66.047.236.7
MBT [26]224p×64.858.043.4
MTCN [18]336p×70.762.149.6
M&M [45]420p×72.066.353.6
TIM (ours)224p×77.167.257.5
", + "image_path": "8710e2663bfa710d1f8fd539d1d0b9869fb1f3aa6b85b0031f3c04896ec4ab1e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 306, + 286, + 326 + ], + "blocks": [ + { + "bbox": [ + 47, + 249, + 287, + 293 + ], + "lines": [ + { + "bbox": [ + 47, + 249, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 249, + 287, + 293 + ], + "type": "text", + "content": "Table 1. Comparisons to state-of-the-art recognition models on the EPIC-KITCHENS validation set. We report the top-1 accuracy for verb, noun and action (\\%). LLM: large language model is used during pre-training. xp: input resolution of " + }, + { + "bbox": [ + 47, + 249, + 287, + 293 + ], + "type": "inline_equation", + "content": "x \\times x" + }, + { + "bbox": [ + 47, + 249, + 287, + 293 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 306, + 286, + 326 + ], + "lines": [ + { + "bbox": [ + 50, + 306, + 286, + 326 + ], + "spans": [ + { + "bbox": [ + 50, + 306, + 286, + 326 + ], + "type": "table", + "html": "
ModelSSAST [12]ASF [19]DiffSED [3]TIM (A)TIM (A+V)
Top-1 acc53.553.856.955.758.3
", + "image_path": "ceacd609df448fb4e644139e7cc5af867940a5b84a200e13de4032cf6b5e67cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 380, + 285, + 453 + ], + "blocks": [ + { + "bbox": [ + 47, + 334, + 287, + 367 + ], + "lines": [ + { + "bbox": [ + 47, + 334, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 287, + 367 + ], + "type": "text", + "content": "Table 2. Comparisons to state-of-the-art sound recognition models on EPIC-SOUNDS. We report the top-1 accuracy " + }, + { + "bbox": [ + 47, + 334, + 287, + 367 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 47, + 334, + 287, + 367 + ], + "type": "text", + "content": " on Val. The performance of SSAST and ASF are from [15]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 380, + 285, + 453 + ], + "lines": [ + { + "bbox": [ + 50, + 380, + 285, + 453 + ], + "spans": [ + { + "bbox": [ + 50, + 380, + 285, + 453 + ], + "type": "table", + "html": "
ModelVAAverage Precision (AP)
Task@0.1@0.2@0.3@0.4@0.5Avg.
G-TAD [46]✓✗Verb12.111.09.48.16.59.4
Noun11.010.08.67.05.48.4
ActionFormer [49]✓✗Verb26.625.424.222.319.123.5
Noun25.224.122.720.517.021.9
ActionFormer - Our Features✓✗Verb29.628.826.924.421.626.3
Noun34.332.630.227.422.629.4
TIM✓✓Verb32.931.629.627.022.228.6
Noun36.434.832.128.722.731.0
", + "image_path": "b400b1a94434cc4badaa7efb2775c433284816d39200d91bfeb1a2d740211671.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 517, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 287, + 564 + ], + "type": "text", + "content": "nition. We expect a further performance boost when implementing any of: higher resolution feature extractors, additional large-scale pre-training and the introduction of a LLM. We leave this as an avenue for future work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 566, + 286, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 566, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 286, + 590 + ], + "type": "text", + "content": "Table 2 compares TIM against prior results on EPIC-SOUNDS, where TIM outperforms SOTA by " + }, + { + "bbox": [ + 47, + 566, + 286, + 590 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 47, + 566, + 286, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 591, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 287, + 651 + ], + "type": "text", + "content": "For detection, we show that TIM can produce competitive results when compared to models primarily designed for this task in Table 3. TIM adapted for detection outperforms ActionFormer [49] by " + }, + { + "bbox": [ + 47, + 591, + 287, + 651 + ], + "type": "inline_equation", + "content": "2.3\\mathrm{mAP}" + }, + { + "bbox": [ + 47, + 591, + 287, + 651 + ], + "type": "text", + "content": " on verb and " + }, + { + "bbox": [ + 47, + 591, + 287, + 651 + ], + "type": "inline_equation", + "content": "1.6\\mathrm{mAP}" + }, + { + "bbox": [ + 47, + 591, + 287, + 651 + ], + "type": "text", + "content": " on noun using the same set of features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": "AVE Results. As this dataset contains joint audio-visual labels, we train TIM by duplicating the query, i.e. using a [CLS] for each modality, and combine their logits during training and inference. We use the pre-trained publicly available models from [36] for a fair comparison with other" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 69, + 545, + 102 + ], + "blocks": [ + { + "bbox": [ + 47, + 461, + 287, + 505 + ], + "lines": [ + { + "bbox": [ + 47, + 461, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 47, + 461, + 287, + 505 + ], + "type": "text", + "content": "Table 3. Comparisons to state-of-the-art detection models on the EPIC-KITCHENS validation set. We report the average precision at IOU thresholds [0.1, 0.2, 0.3, 0.4, 0.5] as well as their average across all thresholds on verb, noun." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 69, + 545, + 102 + ], + "lines": [ + { + "bbox": [ + 309, + 69, + 545, + 102 + ], + "spans": [ + { + "bbox": [ + 309, + 69, + 545, + 102 + ], + "type": "table", + "html": "
ModelPSP [53]CPSP [54]CSSNet [9]†TIM
VAAVAV*
Top-1 acc77.878.680.562.865.579.279.8
", + "image_path": "802fdf91ec8fa1805c2cc1e3ed3ab82b4022e89c884b4de7b4d55cb26cf38835.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 169, + 544, + 228 + ], + "blocks": [ + { + "bbox": [ + 306, + 111, + 545, + 156 + ], + "lines": [ + { + "bbox": [ + 306, + 111, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 111, + 545, + 156 + ], + "type": "text", + "content": "Table 4. Top-1 event classification accuracy " + }, + { + "bbox": [ + 306, + 111, + 545, + 156 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 306, + 111, + 545, + 156 + ], + "type": "text", + "content": " on AVE Test set. " + }, + { + "bbox": [ + 306, + 111, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 306, + 111, + 545, + 156 + ], + "type": "text", + "content": " : no official code or public model provided to replicate results. We show the models trained only with publicly available datasets. " + }, + { + "bbox": [ + 306, + 111, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 306, + 111, + 545, + 156 + ], + "type": "text", + "content": " : results with Omni+ASF features." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 169, + 544, + 228 + ], + "lines": [ + { + "bbox": [ + 308, + 169, + 544, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 169, + 544, + 228 + ], + "type": "table", + "html": "
Perception Test Action
ModelMLP (V)MTCN [18](A+V)TIM (V)TIM (A+V)
Top-1 acc43.751.256.161.1
Perception Test Sound
ModelMLP (A)MTCN [18](A+V)TIM (A)TIM (A+V)
Top-1 acc50.652.954.856.1
", + "image_path": "0c262f06788d41fb431998f018d27b94f6329b43fc3d9eec6a67e71176c98bdb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 309, + 305, + 544, + 382 + ], + "blocks": [ + { + "bbox": [ + 305, + 237, + 545, + 292 + ], + "lines": [ + { + "bbox": [ + 305, + 237, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 305, + 237, + 545, + 292 + ], + "type": "text", + "content": "Table 5. Comparisons to trained recognition baselines on the Perception Test validation split. We show both action and sound recognition and the benefit of including audio-visual in TIM for both challenges. V : visual and A : audio input features. MLP is the result by training an MLP classifier with the features directly." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 305, + 544, + 382 + ], + "lines": [ + { + "bbox": [ + 309, + 305, + 544, + 382 + ], + "spans": [ + { + "bbox": [ + 309, + 305, + 544, + 382 + ], + "type": "table", + "html": "
ModelAverage Precision (AP)
@0.1@0.2@0.3@0.4@0.5Avg.
Perception Test Action
ActionFormer [49]27.827.625.223.020.024.5
TIM33.532.229.826.422.028.8
Perception Test Sound
ActionFormer [49]34.731.327.522.717.726.8
TIM37.533.127.922.817.227.7
", + "image_path": "d708fbdcf6c48bd7d7f7e529f21effa11823a5d4a218db77a09d93cb6f4e6c16.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 391, + 545, + 435 + ], + "lines": [ + { + "bbox": [ + 305, + 391, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 305, + 391, + 545, + 435 + ], + "type": "text", + "content": "Table 6. Comparisons to strong detection models on the Perception Test validation set for action and sound localisation. We report the average precision at IOU thresholds [0.1, 0.2, 0.3, 0.4, 0.5] as well as the average across all thresholds." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 446, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 446, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 305, + 446, + 545, + 470 + ], + "type": "text", + "content": "works. We also apply AVGA [36] to spatial visual features from VGG-19 before feeding them to the transformer." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 472, + 545, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 472, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 545, + 543 + ], + "type": "text", + "content": "Table 4 shows our results on the AVE dataset. Combining audio and video significantly improves the performance on TIM. The results from [9] perform best but could not be replicated. We also report TIM using the Omnivore visual features and Auditory Slowfast features used for EPIC-KITCHENS, which achieves a " + }, + { + "bbox": [ + 305, + 472, + 545, + 543 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 305, + 472, + 545, + 543 + ], + "type": "text", + "content": " boost in performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": "Perception Test Results. We use the same backbone for the Omnivore features and Auditory Slowfast features and train a single model using both the visual and audio labels. Table 5 compares the results on the newly introduced Perception Test. We train an MLP classifier, with two linear layers and a ReLU activation, on the features directly as a baseline. We also evaluate on an audio-visual model that does use context with MTCN. Compared to these methods TIM clearly shows significant improvements. Results are improved over MTCN by " + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "inline_equation", + "content": "9.9\\%" + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": " on visual and audio recognition tasks respectively. We also provide detection results in Table 6. TIM improves over ActionFormer [49] by 3.3 average mAP on visual actions and by 0.9 average mAP on sound, when using the same features." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18158" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 288, + 297 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 288, + 297 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 288, + 297 + ], + "type": "image", + "image_path": "99ebabf585cd33faad664ece1324409a3e3f93644d45fcc2925113ac8b8ac581.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 304, + 545, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 545, + 316 + ], + "type": "text", + "content": "Figure 4. Qualitative results for all datasets. PRED: Prediction by TIM, TIQ: Time Interval Queries, V/AGT: Visual/Audio Ground Truth." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 293, + 68, + 545, + 295 + ], + "blocks": [ + { + "bbox": [ + 293, + 68, + 545, + 295 + ], + "lines": [ + { + "bbox": [ + 293, + 68, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 293, + 68, + 545, + 295 + ], + "type": "image", + "image_path": "bd10edd49098b623f5b0f85b4effc0bf5bd1ec03e71ced203c1475089241fb67.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 327, + 287, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 470 + ], + "type": "text", + "content": "Cross-Modality in TIM. When referring to our previous results, we see that including the additional modality provides a performance boost in all cases, highlighting TIM's ability to utilise and distinguish between different modalities. For example, on EPIC-KITCHENS-100, including audio improves visual action accuracy by " + }, + { + "bbox": [ + 46, + 327, + 287, + 470 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 46, + 327, + 287, + 470 + ], + "type": "text", + "content": ". For EPICSOUNDS, the visual modality further improves accuracy by " + }, + { + "bbox": [ + 46, + 327, + 287, + 470 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 46, + 327, + 287, + 470 + ], + "type": "text", + "content": ". In the Perception Test, including the audio modality improves visual recognition by " + }, + { + "bbox": [ + 46, + 327, + 287, + 470 + ], + "type": "inline_equation", + "content": "5.0\\%" + }, + { + "bbox": [ + 46, + 327, + 287, + 470 + ], + "type": "text", + "content": ", and visual increases sound recognition by " + }, + { + "bbox": [ + 46, + 327, + 287, + 470 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 46, + 327, + 287, + 470 + ], + "type": "text", + "content": ". Finally, for AVE, we see a significant improvement, where an audio-visual model increases accuracy by " + }, + { + "bbox": [ + 46, + 327, + 287, + 470 + ], + "type": "inline_equation", + "content": "13.7\\%" + }, + { + "bbox": [ + 46, + 327, + 287, + 470 + ], + "type": "text", + "content": " from audio-only." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 472, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 288, + 664 + ], + "type": "text", + "content": "Qualitative Results. We present qualitative results in Figure 4. We see that in EPIC-KITCHENS, TIM can competently recognise actions across the two modalities including overlapping queries. Furthermore, we see consecutive actions are correctly recognised with varying interval lengths, such as the 'open / close' audio actions between 0.2s and 4.5s. For AVE, TIM is able to distinguish between the background and a 'barking' audio-visual event based on the time interval query. For the Perception Test, we see that TIM can distinguish between heavily overlapping actions across both modalities, such as between 'break', 'human speech', 'hit' and 'put something'. However, there are also failure cases, such as in EPIC-KITCHENS when the action 'take washing up liquid' is recognised as 'wash hand', as the model is likely confused by the context predominantly associated with the highly overlapping 'wash hand' actions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 671, + 188, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 188, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 188, + 684 + ], + "type": "text", + "content": "4.4. Analysing Time Intervals" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We showcase the importance of effectively encoding time intervals, as well as how they differ from alternative strate-" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 305, + 324, + 545, + 499 + ], + "blocks": [ + { + "bbox": [ + 305, + 324, + 545, + 499 + ], + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 499 + ], + "type": "image", + "image_path": "fcb9398663fa72e6ca36f8b903cb62b76c2adc38c0a8483e790fdd315583d383.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 508, + 545, + 541 + ], + "lines": [ + { + "bbox": [ + 305, + 508, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 305, + 508, + 545, + 541 + ], + "type": "text", + "content": "Figure 5. TSNE plot for time encodings " + }, + { + "bbox": [ + 305, + 508, + 545, + 541 + ], + "type": "inline_equation", + "content": "I(\\cdot)" + }, + { + "bbox": [ + 305, + 508, + 545, + 541 + ], + "type": "text", + "content": " on all datasets. In each plot, we use colour maps to indicate encodings of the time interval's duration (left), start time (middle) and end time (right)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 554, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 554, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 305, + 554, + 545, + 578 + ], + "type": "text", + "content": "gies. We perform this analysis on the EPIC-KITCHENS-100 and EPIC-SOUNDS recognition tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Time Encoding Representation. To show TIM encodes time intervals across all datasets in Figure 5. We use three colour maps on the same TSNE projection to show three properties of the encoded interval: duration, start time and end time. Interestingly, the 1D time encoding perfectly captures all three attributes and across the datasets. While the encodings differ per dataset, as these differ in the positions and durations of actions, we see clear similarities in the learnt time encoding projections. For example, the duration is perfectly captured across the x-axis of the TSNE plot with lower values indicating longer time intervals." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18159" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 286, + 228 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 228 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 228 + ], + "type": "image", + "image_path": "79bc2ccf94924ad8be6bfba4acdc35e7bc0366f47ffe04c2bbfe24b4e649eddd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 238, + 287, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 287, + 283 + ], + "type": "text", + "content": "Figure 6. Attention heatmaps of the second encoder layer for two random 30s clips in EPIC-KITCHENS. x-axis: input feature time intervals; y-axis query time intervals of varying position and duration. The attention magnitude relates to the query CLS token." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 71, + 293, + 262, + 365 + ], + "blocks": [ + { + "bbox": [ + 71, + 293, + 262, + 365 + ], + "lines": [ + { + "bbox": [ + 71, + 293, + 262, + 365 + ], + "spans": [ + { + "bbox": [ + 71, + 293, + 262, + 365 + ], + "type": "image", + "image_path": "9796b37d53ca4aae70507d2799aa1a95b390944ed8618727740fb19ecdaf61c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 71, + 367, + 262, + 435 + ], + "blocks": [ + { + "bbox": [ + 71, + 367, + 262, + 435 + ], + "lines": [ + { + "bbox": [ + 71, + 367, + 262, + 435 + ], + "spans": [ + { + "bbox": [ + 71, + 367, + 262, + 435 + ], + "type": "image", + "image_path": "e962224ee5858897afd4669f2fb4f8108ffa74f8a1070e3beedcaa0e32feba15.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 444, + 287, + 489 + ], + "lines": [ + { + "bbox": [ + 46, + 444, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 287, + 489 + ], + "type": "text", + "content": "Figure 7. The impact of shifting the time interval query on visual performance (top) and audio performance (bottom), both on short actions " + }, + { + "bbox": [ + 46, + 444, + 287, + 489 + ], + "type": "inline_equation", + "content": "(< 2\\mathrm{sec})" + }, + { + "bbox": [ + 46, + 444, + 287, + 489 + ], + "type": "text", + "content": " and long actions " + }, + { + "bbox": [ + 46, + 444, + 287, + 489 + ], + "type": "inline_equation", + "content": "(>2\\mathrm{sec})" + }, + { + "bbox": [ + 46, + 444, + 287, + 489 + ], + "type": "text", + "content": " and overall validation set (all) for EPIC-KITCHENS-100 and EPIC-SOUNDS." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 495, + 286, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 286, + 590 + ], + "type": "text", + "content": "Interval Query Attention. We plot two attention heatmaps in Figure 6 for 5 separate queries with varying positions and scales in EPIC-KITCHENS-100. We extract the attention weights from the second transformer encoder layer, as this appears to be the most relevant to the interval query. The learnt attention clearly applies to the feature time intervals contained within the query. We note the similarity between the attention in the two randomly selected windows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 286, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 641 + ], + "type": "text", + "content": "Shifting Intervals. To show how TIM effectively encodes the time interval of actions, we shift the time interval queries from their correct action interval by -1.5s to 1.5s, assessing the impact of these adjustments on the performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "content": "Figure 7 shows the result. We see the performance gradually drops, both in visual and audio, as the query interval moves away from the correct action interval. The drop is also symmetric showcasing no bias. Unsurprisingly, the performance drops significantly when shifting short actions both in video " + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "inline_equation", + "content": "(-57.9\\%)" + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "content": " and audio " + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "inline_equation", + "content": "(-35.2\\%)" + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "content": ", while it is less" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 162 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 545, + 162 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 162 + ], + "type": "table", + "html": "
EPIC-KITCHENSEPIC-SOUNDS
EncodingVerbNounActionAudio Actions
Learned43.844.329.623.7
Sinusoidal43.844.630.013.4
Centre74.365.855.656.4
Separate-add76.066.256.457.7
Interval-add76.366.556.958.8
Separate-cat76.867.457.158.4
Interval-cat (proposed)77.167.257.558.3
", + "image_path": "822a16d57cacc349cf0f7c6b61d9a619f9fc1c6934021d96297fcac44ff9b72c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 324, + 171, + 527, + 182 + ], + "lines": [ + { + "bbox": [ + 324, + 171, + 527, + 182 + ], + "spans": [ + { + "bbox": [ + 324, + 171, + 527, + 182 + ], + "type": "text", + "content": "Table 7. Ablating the choice of encoding time intervals." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 196, + 545, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 196, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 305, + 196, + 545, + 220 + ], + "type": "text", + "content": "extreme in long actions " + }, + { + "bbox": [ + 305, + 196, + 545, + 220 + ], + "type": "inline_equation", + "content": "(-14.5\\%" + }, + { + "bbox": [ + 305, + 196, + 545, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 196, + 545, + 220 + ], + "type": "inline_equation", + "content": "-11.2\\%)" + }, + { + "bbox": [ + 305, + 196, + 545, + 220 + ], + "type": "text", + "content": ". We assess the impact of scaling the time interval in the ArXiv appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 221, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 545, + 353 + ], + "type": "text", + "content": "Time Interval Encodings. The Time Interval MLP encodes the interval of the query. Here, we compare this to traditional positional encodings, both sinusoidal and learned. We also experiment on five different variations of the Time Interval MLP, namely: (i) Centre - we only encode the centre timestamp of the interval; (ii) Separate-Add/Cat - we encode the intervals' start and end time separately and add the encoded output vectors together, or concatenate along the channel dimension; and (iii) Interval-Add/Cat - we encode the start and end time within the same vector and add, or concatenate, the encoded output to the input sequence." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 354, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 425 + ], + "type": "text", + "content": "We show the results in Table 7. In all cases, the final encoding is of the same dimension for comparable results. Performance is significantly worse with sinusoidal or learned positional encoding, as they are unable to capture the complexity of overlapping actions. There is also a drop when encoding only the centre of the time interval." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 426, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 497 + ], + "type": "text", + "content": "Separate-Add/Cat are alternative ways to encode the intervals (and hence include duration information) resulting in comparable performance to the interval counterparts. Our proposed approach to encoding the interval into the MLP shows the best performance for visual and while maintaining strong auditory performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 508, + 383, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 508, + 383, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 508, + 383, + 521 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 528, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 545, + 611 + ], + "type": "text", + "content": "In this paper, we propose to utilise the action's time interval as a query to an audio-visual transformer which learns to recognise the action from its interval and the unaltered surrounding context. We jointly train the model on modality-specific time intervals and label sets, allowing the Time Interval Machine (TIM) to recognise multiple events across both visual and auditory modalities." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 612, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 612, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 545, + 647 + ], + "type": "text", + "content": "TIM is sensitive to the interval's position and duration. This allows the model, as is, to produce competitive results on action detection through multi-scale dense querying." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 647, + 545, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 702 + ], + "type": "text", + "content": "Acknowledgements. This work uses public datasets. It is supported by EPSRC Doctoral Training Program, EPSRC UMPIRE EP/T004991/1 and EPSRC Programme Grant VisualAI EP/T028572/1; and by the use of the EPSRC funded Tier 2 facility JADE-II." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18160" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 179 + ], + "type": "text", + "content": "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 181, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 181, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 181, + 287, + 224 + ], + "type": "text", + "content": "[3] Swapnil Bhosale, Sauradip Nag, Diptesh Kanojia, Jiankang Deng, and Xiatian Zhu. Diffsed: Sound event detection with denoising diffusion. arXiv preprint arXiv:2308.07293, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 227, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 287, + 270 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. ECCV, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "type": "text", + "content": "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 288, + 372 + ], + "type": "text", + "content": "[6] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 373, + 287, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 287, + 440 + ], + "type": "text", + "content": "[7] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 2021. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 441, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 287, + 485 + ], + "type": "text", + "content": "[8] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. SlowFast Networks for Video Recognition. In Proceedings of International Conference on Computer Vision (ICCV), 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 529 + ], + "type": "text", + "content": "[9] Fan Feng, Yue Ming, Nannan Hu, Hui Yu, and Yuanan Liu. CSS-net: A consistent segment selection network for audiovisual event localization. IEEE Transactions on Multimedia, 2023. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 532, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 287, + 575 + ], + "type": "text", + "content": "[10] Ruohan Gao, Tae-Hyun Oh, Kristen Grauman, and Lorenzo Torresani. Listen to look: Action recognition by previewing audio. In Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "type": "text", + "content": "[11] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A Single Model for Many Visual Modalities. In CVPR, 2022. 1, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 622, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 287, + 666 + ], + "type": "text", + "content": "[12] Yuan Gong, Cheng-I Lai, Yu-An Chung, and James Glass. Ssast: Self-supervised audio spectrogram transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10699-10709, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 668, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 668, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 668, + 287, + 690 + ], + "type": "text", + "content": "[13] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Roei Herzig, Ofir Abramovich, Elad Ben-Avraham, Assaf Arbelle, Leonid Karlinsky, Ariel Shamir, Trevor Darrell, and" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Amir Globerson. Promptonomyvit: Multi-task prompt learning improves video transformers using synthetic scene data. arXiv preprint arXiv:2212.04821, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 161 + ], + "type": "text", + "content": "[15] Jaesung Huh, Jacob Chalk, Evangelos Kazakos, Dima Damen, and Andrew Zisserman. EPIC-SOUNDS: A LargeScale Dataset of Actions that Sound. In IEEE International Conference on Acoustics, Speech, & Signal Processing (ICASSP), 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "text", + "content": "[16] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "type": "text", + "content": "[17] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. Epic-fusion: Audio-visual temporal binding for egocentric action recognition. In Proceedings of International Conference on Computer Vision (ICCV), 2019. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "text", + "content": "[18] Evangelos Kazakos, Jaesung Huh, Arsha Nagrani, Andrew Zisserman, and Dima Damen. With a little help from my temporal context: Multimodal egocentric action recognition. In Proc. BMVC, 2021. 1, 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 305, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 358 + ], + "type": "text", + "content": "[19] Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. Slow-fast auditory streams for audio recognition. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 360, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 425 + ], + "type": "text", + "content": "[20] Dan Kondratyuk, Liangzhe Yuan, Yandong Li, Li Zhang, Mingxing Tan, Matthew Brown, and Boqing Gong. Movinets: Mobile video networks for efficient video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16020-16030, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "type": "text", + "content": "[21] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco Nadai. Efficient training of visual transformers with small datasets. Advances in Neural Information Processing Systems, 34, 2021. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "text", + "content": "[22] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3202-3211, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 569 + ], + "type": "text", + "content": "[23] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 570, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 592 + ], + "type": "text", + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "type": "text", + "content": "[25] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12475-12486, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "text", + "content": "[26] Arsha Nagrani, Shan Yang, Anurag Arnab, Cordelia Schmid, and Chen Sun. Attention bottlenecks for multimodal fusion. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[27] Yan Bin Ng and Basura Fernando. Human action sequence classification. CoRR, abs/1910.02602, 2019. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18161" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[28] Daniel S Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D Cubuk, and Quoc V Le. Specaugment: A simple data augmentation method for automatic speech recognition. arXiv preprint arXiv:1904.08779, 2019. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 194 + ], + "type": "text", + "content": "[29] Mandela Patrick, Dylan Campbell, Yuki M. Asano, Ishaan Misra Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and João F. Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 196, + 288, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 196, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 48, + 196, + 288, + 295 + ], + "type": "text", + "content": "[30] Viorica Pătrăucean, Lucas Smaira, Ankush Gupta, Adrià Recasens Continente, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Joseph Heyward, Mateusz Malinowski, Yi Yang, Carl Doersch, Tatiana Matejovicova, Yury Sulsky, Antoine Miech, Alex Frechette, Hanna Klimczak, Raphael Koster, Junlin Zhang, Stephanie Winkler, Yusuf Aytar, Simon Osindero, Dima Damen, Andrew Zisserman, and João Carreira. Perception test: A diagnostic benchmark for multimodal video models, 2023. 1, 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 297, + 288, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 288, + 340 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 288, + 340 + ], + "type": "text", + "content": "[31] Fadime Sener, Dipika Singhania, and Angela Yao. Temporal aggregate representations for long-range video understanding. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 342, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 385 + ], + "type": "text", + "content": "[32] Divya Shanmugam, Davis Blalock, Guha Balakrishnan, and John Guttag. Better aggregation in test-time augmentation. In Proceedings of the IEEE/CVF international conference on computer vision, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 387, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 288, + 431 + ], + "type": "text", + "content": "[33] Jonathan C Stroud, Zhichao Lu, Chen Sun, Jia Deng, Rahul Sukthankar, Cordelia Schmid, and David A Ross. Learning video representations from textual web supervision. arXiv preprint arXiv:2007.14937, 2020. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 433, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 288, + 487 + ], + "type": "text", + "content": "[34] Swathikiran Sudhakaran, Adrian Bulat, Juan-Manuel Perez-Rua, Alex Falcon, Sergio Escalera, Oswald Lanz, Brais Martinez, and Georgios Tzimiropoulos. Saic.cambridge-hupba-fbk submission to the epic-kitchens-100 action recognition challenge 2021. arXiv preprint arXiv:2110.02902, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 488, + 288, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 288, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 288, + 543 + ], + "type": "text", + "content": "[35] Tsung-Ming Tai, Oswald Lanz, Giuseppe Fiameni, Yi-Kwan Wong, Sze-Sen Poon, Cheng-Kuang Lee, Ka-Chun Cheung, and Simon See. Nvidia-unibz submission for epic-kitchens-100 action anticipation challenge 2022. arXiv preprint arXiv:2206.10869, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 544, + 288, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 577 + ], + "type": "text", + "content": "[36] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In ECCV, 2018. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 578, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 612 + ], + "type": "text", + "content": "[37] Yapeng Tian, Dingzeyu Li, and Chenliang Xu. Unified multisensory perception: Weakly-supervised audio-visual video parsing. In Proc. ECCV, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 657 + ], + "type": "text", + "content": "[38] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In Advances in Neural Information Processing Systems, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 714 + ], + "type": "text", + "content": "[39] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 148 + ], + "type": "text", + "content": "[40] Luyu Wang, Pauline Luc, Yan Wu, Adria Recasens, Lucas Smaira, Andrew Brock, Andrew Jaegle, Jean-Baptiste Alayrac, Sander Dieleman, Joao Carreira, et al. Towards learning universal audio representations. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4593-4597. IEEE, 2022. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 150, + 546, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 546, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 546, + 194 + ], + "type": "text", + "content": "[41] Weiyao Wang, Du Tran, and Matt Feiszli. What makes training multi-modal classification networks hard? In Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 196, + 547, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 547, + 251 + ], + "type": "text", + "content": "[42] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 284-293, 2019. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 252, + 547, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 252, + 547, + 317 + ], + "spans": [ + { + "bbox": [ + 308, + 252, + 547, + 317 + ], + "type": "text", + "content": "[43] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 319, + 546, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 546, + 361 + ], + "type": "text", + "content": "[44] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 364, + 546, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 364, + 546, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 364, + 546, + 397 + ], + "type": "text", + "content": "[45] Xuehan Xiong, Anurag Arnab, Arsha Nagrani, and Cordelia Schmid. M&m mix: A multimodal multiview transformer ensemble. arXiv preprint arXiv:2206.09852, 2022. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 399, + 546, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 546, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 546, + 430 + ], + "type": "text", + "content": "[46] Mengmeng Xu, Chen Zhao, David S. Rojas, Ali Thabet, and Bernard Ghanem. G-tad: Sub-graph localization for temporal action detection, 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 432, + 546, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 432, + 546, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 432, + 546, + 487 + ], + "type": "text", + "content": "[47] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3333-3343, 2022. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 488, + 546, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 546, + 520 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 546, + 520 + ], + "type": "text", + "content": "[48] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Temporal query networks for fine-grained video understanding. In Proc. CVPR, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 522, + 546, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 546, + 555 + ], + "type": "text", + "content": "[49] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, pages 492-510, 2022. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 557, + 546, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 546, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 546, + 589 + ], + "type": "text", + "content": "[50] Yue Zhao and Philipp Kähenbuhl. Training a large video model on a single machine in a day. arXiv preprint arXiv:2309.16669, 2023. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 590, + 546, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 546, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 546, + 623 + ], + "type": "text", + "content": "[51] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, 2023. 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 624, + 546, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 546, + 677 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 546, + 677 + ], + "type": "text", + "content": "[52] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In The AAAI Conference on Artificial Intelligence (AAAI), pages 12993-13000, 2020. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 680, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 546, + 713 + ], + "type": "text", + "content": "[53] Jinxing Zhou, Liang Zheng, Yiran Zhong, Shijie Hao, and Meng Wang. Positive sample propagation along the audiovisual event line. In Proceedings of the IEEE/CVF Con-" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18162" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 140 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "type": "text", + "content": "ference on Computer Vision and Pattern Recognition, pages 8436-8444, 2021. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[54] Jinxing Zhou, Dan Guo, and Meng Wang. Contrastive positive sample propagation along the audio-visual event line. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 6" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18163" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/TRINS_ Towards Multimodal Language Models that Can Read/527c7573-3fb6-42ad-a14f-b7b957e97243_content_list.json b/2024/TRINS_ Towards Multimodal Language Models that Can Read/527c7573-3fb6-42ad-a14f-b7b957e97243_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1b2bf889df4838345f78cc5b41ecfaef4313cfec --- /dev/null +++ b/2024/TRINS_ Towards Multimodal Language Models that Can Read/527c7573-3fb6-42ad-a14f-b7b957e97243_content_list.json @@ -0,0 +1,1484 @@ +[ + { + "type": "text", + "text": "TRINS: Towards Multimodal Language Models that Can Read", + "text_level": 1, + "bbox": [ + 166, + 130, + 803, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruiyi Zhang $^{1}$ , Yanzhe Zhang $^{2}$ , Jian Chen $^{3}$ , Yufan Zhou $^{1}$ , Jiumiang Gu $^{1}$ , Changyou Chen $^{3}$ , Tong Sun $^{1}$", + "bbox": [ + 174, + 179, + 803, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Adobe Research $^{2}$ Georgia Institute of Technology $^{3}$ State University of New York at Buffalo", + "bbox": [ + 98, + 215, + 880, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ruizhang}@adobe.com", + "bbox": [ + 398, + 236, + 576, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large multimodal language models have shown remarkable proficiency in understanding and editing images. However, a majority of these visually-tuned models struggle to comprehend the textual content embedded in images, primarily due to the limitation of training data. In this work, we introduce TRINS: a Text-Rich image1 INstruction dataset, with the objective of enhancing the reading ability of the multimodal large language model. TRINS is built upon LAION2 using hybrid data annotation strategies that include machine-assisted and human-assisted annotation process. It contains 39,153 text-rich images, captions, and 102,437 questions. Specifically, we show that the number of words per annotation in TRINS is significantly longer than that of related datasets, providing new challenges. Furthermore, we introduce a simple and effective architecture, called a Language-Vision Reading Assistant (LaRA), which is good at understanding textual content within images. LaRA outperforms existing state-of-the-art multimodal large language models on the TRINS dataset as well as other classical benchmarks. Lastly, we conducted a comprehensive evaluation with TRINS on various text-rich image understanding and generation tasks, demonstrating its effectiveness.", + "bbox": [ + 76, + 316, + 473, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 676, + 209, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Instruction tuning [9, 38] has shown a great generalization ability on unseen tasks and has contributed to the growing popularity of large language models (LLMs), such as ChatGPT [37]. Recently, multimodal language models benefit from visual instruction finetuning [1, 18, 19, 28, 63], and has shown great success in real-world applications. These models leverage visual encoders such as CLIP-ViT [12, 39] to empower LLMs with image comprehension ability. However, challenges arise in comprehension of textual information within images, which may stem from the prevalence of natu", + "bbox": [ + 75, + 703, + 470, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ral images in training datasets, such as Conceptual Captions [5] and COCO [26]), as highlighted by Liu et al. [30]. Recognizing the importance of visual textual understanding for effective collaboration between agents and humans, Zhang et al. [62] proposed enhancing end-to-end visual instruction-tuned models by introducing noisy Optical Character Recognition (OCR) annotations to improve vision-language alignment. In this work, we surpass existing achievements and collect a new Text-Rich image INStruction dataset named TRINS, which contains 39,153 text-rich images, captions and 102,437 questions.", + "bbox": [ + 500, + 287, + 893, + 454 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "TRINS is created in a semi-automatic manner for a more controllable and faithful collection. Specifically, we exploited large-scale pre-trained models such as CLIP [40] and GPT-4 in the annotation process. This semi-automatic process significantly reduces the time and resources required for manual annotation and surprisingly improves the overall quality of annotations. TRINS dataset is composed of three datasets for captioning, visual question answering (VQA) and image generation, respectively. Specifically, human-annotated captions for text-rich images are first collected because they can best translate text-rich images into texts. During this process, extracted OCR words and recognize-anything model tags are provided to the annotators for better and efficient annotations. With detailed image descriptions, VAQ data is built and fulfilled by large language models, such as GPT-4 [37] and LLaMA-2 [51]. In detailed statistics and analysis, we found that both annotated captions and collected question-answer pairs are more comprehensive and contain significantly more details than the existing dataset. Therefore, we show the superior advantage of TRINS compared to existing instruction fine-tuning datasets. As a by-product, high-quality image-caption pairs can serve as a good benchmark for text-rich image generation, which is still a very challenging task [6]. At the same time, we propose a new simple and effective multimodal language model architecture that includes OCR as a component. We call it Language-vision Reading Assistant (LaRA) and show that LaRA fine-tuned on TRINS brings the best text-rich image understanding ability. Our contributions are as follows:", + "bbox": [ + 496, + 462, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1In this work, we use the phrase \"text-rich images\" to describe images with rich textual information, such as posters and book covers.", + "bbox": [ + 75, + 862, + 468, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "2Work done during Q3 2023.", + "bbox": [ + 96, + 887, + 250, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22584", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel dataset (TRINS) containing 39,153 captions and 102,437 high-quality text-rich image instruction pairs. TRINS is annotated with a novel semi-automatic annotation framework that is scalable and reliable.", + "- We develop several evaluation benchmarks for text-rich image understanding and generation tasks. Various methods are evaluated on TRINS, demonstrating the effectiveness of the dataset.", + "- The TRINS datasets are high-quality and comprehensive, which is reflected not only in the dataset statistics but also from the results of multiple baseline models. Especially, our proposed LaRA finetuned on TRINS outperforms existing state-of-the-art methods on text-rich image understanding tasks." + ], + "bbox": [ + 76, + 90, + 468, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 329, + 217, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal Instruction Tuning Multi-modal instruction tuning, including image, video [32, 59], and audio [16, 58] settings, has been an active research topic. MiniGPT-4 [63] uses ChatGPT to generate high-quality instruction-following data, while LLaVA [28] generates such data by prompting GPT-4 with captions and bounding boxes. LLaMA-Adapter [14, 60] aligns text-image features using COCO data, and mPLUG-owl [56] combines extensive image-text pairs for pretraining and a mixture of data for finetuning. Despite this, many models, according to Liu et al. [30], struggle with OCR tasks. InstructBLIP [10] addresses this by transforming 13 vision language tasks into an instruction-following format. mPLUGOwl [55, 56] apply multitask instruction funetuing using existing document datasets. A comprehensive survey is available in Li et al. [20]. LLaVAR [62] exploits GPT-4 to collect fine-tuning data without human annotations using OCR and captioning tools. It discovered that resolution plays a significant role in recognizing textual information and explored several options. Monkey [24] performed a surgery between simple text labels and high input resolution, enabling remarkable performance in visually-rich document images with dense text. TRINS exploits human-machine collaboration for data annotation and can provide more accurate information, reducing the problem of hallucination.", + "bbox": [ + 76, + 354, + 470, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-Rich Image Datasets Visual question answering or captioning datasets are widely used in task-specific finetuning and large multimodal model evaluation. TextCap [46] is the first text-rich image captioning dataset. Compared to TextCap, TRINS-Cap provides more detailed annotations that can fulfill the requirement of instruction finetuning. TextOCR [48] aims comprehend text in the context of an image, which is similar to our motivation but focuses more on text recognition in images instead of understanding. STVQA [13] uses spatial and textual information to answer visually grounded questions, effectively integrating visual", + "bbox": [ + 76, + 734, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/00c482fa61e018f2cb1af6e9ef212201c62f6f9fdedd4a1ee8e7e9a3ddb89008.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetYearSizeAnnotation Type
OCR-VQA2019200KQA (▲)
TextVQA201945KQA (▲)
TextCap2020140kCaption (▲)
TextOCR2021145kText Bbox (▲)
DocVQA202050kQA (▲)
40kCaption (▲)
TRINS (Ours)2023100kQA (▲ +▲)
40kText Bbox (▲)
", + "bbox": [ + 514, + 88, + 879, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison between TRINS and other related datasets.", + "bbox": [ + 503, + 234, + 885, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and textual cues. OCR-VQA [36] focuses on incorporating optical character recognition (OCR) into visual question answering (VQA), which operates primarily on text within images. TextVQA [47] also takes advantage of the textual information present in the images to answer questions, but with an emphasis on open questions. DocVQA takes this one step further by applying VQA to document images, handling a variety of layouts and formats. InfoVQA [35] and ChartQA [33] focus on specific subdomains and aim to answer questions about information graphics and chart images, respectively. In summary, these related works provide datasets for leveraging spatial and textual cues. TRINS-VQA is a dataset that exploits the semi-automatic annotation process. It can be used for general domain instruction finetuning and model evaluations.", + "bbox": [ + 496, + 267, + 893, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Text-Rich Image Instruction Dataset", + "text_level": 1, + "bbox": [ + 500, + 510, + 830, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To equip multimodal language models with the ability to recognize text and relate it to its visual context, we have curated a new dataset named Text-Rich Instruction (TRINS). The ultimate goal is to enable these models to have spatial, semantic, and visual reasoning between multiple text tokens and visual entities. In this section, we present TRINS, a dataset crafted through a semi-automatic process. Specifically, we leverage large-scale pre-trained models like CLIP [40] and GPT-4 in the annotation process, offering potential advantages: (i) Significant reduction in annotation time and resources: using these models significantly reduces the time and resources required for manual annotation. (ii) Enhancement of annotation data quality through post-processing: the involvement of large models contributes to improving the overall quality of annotation data through subsequent post-processing. (iii) Functionality of large models as knowledge bases: large models can serve as effective knowledge bases, aiding in the annotation process by virtue of their extensive training in diverse datasets.", + "bbox": [ + 496, + 537, + 893, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first outline the document image collection process for TRINS, utilizing CLIP models. Then, we present data statistics to facilitate a comprehensive understanding. We delve into three distinct tasks derived from the TRINS dataset in detail: i) TRINS-Cap: Visual Captioning, ii) TRINS-VQA:", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22585", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2e07ccf16200e8bd467810bedd6b190e2e9d962a7d624f7fa2783da4a2892e6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 90, + 890, + 210 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9ff8f48869fbb1da497f7a7c8a13946925c11b54325674a532753e38d772c26c.jpg", + "image_caption": [ + "Figure 1. Overview of TRINS data collection process, which consists of three datasets. Text-rich images are first selected from web images and then ask annotators to describe the image in details. i) TRINS-Cap is extracted from human annotations with heuristic data processing for text-rich image summarization tasks. ii) TRINS-VQA is built upon human annotations and generate question-answer pairs for training by prompting text-only large language models. iii) TRINS-Gen combined human annotations and text boxes for text-rich image generation." + ], + "image_footnote": [], + "bbox": [ + 76, + 282, + 875, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Question Answering and iii) TRINS-Gen: Text-to-Image Generation. The overview of the TRINS data collection process is illustrated in Figure 1. To provide a succinct overview, TRINS-Cap undergoes full annotation by human annotators, while TRINS-Gen and TRINS-VQA are constructed based on TRINS-Cap with the help of pre-trained models. A summary of the various TRINS datasets is presented in Table 7 of Appendix A.", + "bbox": [ + 75, + 469, + 472, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Machine-Assisted Text-rich Image Selections", + "text_level": 1, + "bbox": [ + 76, + 599, + 455, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Beginning with the LAION-5B dataset [44], our objective is to selectively retain images that exhibit a significant presence of text. Recognizing that document images typically contain substantial textual content, we initially formed a binary classification dataset by combining natural images with document data. Subsequently, we trained an image classifier using a DiT [21] base backbone, fine-tuned on the RVL-CDIP dataset [15]. The purpose of this classifier was to predict whether an image contains text. Then a subset was constructed by selecting images with a predicted probability greater than 0.8, while also adhering to the criteria $p(\\text{watermark}) < 0.8$ and $p(\\text{unsafe}) < 0.5$ , where both probabilities are derived from the metadata of the LAION dataset. Acknowledging the noise introduced due to the classifier's limitations, we further refined the dataset by incorporating human judgment. A random sample of 20,000 images", + "bbox": [ + 75, + 622, + 470, + 866 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/63d498ffbbc00fa4d8ed66ff2ea00ceb34d9bb86fb247519e51be4241bbd4848.jpg", + "image_caption": [ + "Figure 2. CLIP-based categorization of our collected images and selected representative data samples from each category.", + "(a)" + ], + "image_footnote": [], + "bbox": [ + 501, + 467, + 694, + 570 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2d4b79776fd32e46a986dacbe18dd014d9d88e6071323881fd99ef94c145ec2c.jpg", + "image_caption": [ + "(b)", + "Figure 3. Word clouds of (a) predicted tags and (b) detected words from the text-rich images of TRINS." + ], + "image_footnote": [], + "bbox": [ + 697, + 468, + 890, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "from the filtered LAION-5B was clustered into 50 groups based on CLIP-ViT-B/32 visual features. After inspecting the clustering results, one cluster was meticulously chosen, encompassing diverse text-rich images such as posters, covers, advertisements, and educational documents. This cluster model then served as the filtering mechanism for collecting images that comprise the TRINS dataset. For reference, we present a CLIP-based categorization [39] in Figure 2 to depict the distribution of images in the collected data. The major class is book cover images, further categorized on the basis of book themes and contents. To enhance our understanding of text-rich images, we employed a Recognize Anything Model (RAM) [17, 61] to extract tags from TRINS images. Figure 3a displays word clouds of RAM tags, where \"book\" and \"poster\" emerge as major keywords. Additionally, we utilize the Azure Read API $^{4}$ and PaddleOCR to extract", + "bbox": [ + 496, + 619, + 893, + 861 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3https://huggingface.co/datasets/laion/laion-high-resolution", + "bbox": [ + 76, + 875, + 472, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "4https://azure.microsoft.com/en-us/updates/computer-vision-v3-preview-6/", + "bbox": [ + 500, + 875, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22586", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e04720c003cdc236deadbd4276ac8aeb06b8f88947330391a6d3af5f56fcafb2.jpg", + "image_caption": [ + "(a)", + "Figure 4. OCR word (a), Caption (b), Question (c) and Answer (d) statistics for TRINS." + ], + "image_footnote": [], + "bbox": [ + 91, + 90, + 277, + 196 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1b45b02cdc6115d23f128d22e7825dd04090b80d81de7594d3a234ec0cff66de.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 290, + 90, + 478, + 196 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3749cb3f0b249d37623a18c5cb51b1d1e78b2c644405fdbeb87b481be6831862.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 488, + 90, + 674, + 196 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/88282c48c156dd73e6fcc27346188ec1b9a5b044249e6cb45219a32213fc3b80.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 687, + 90, + 870, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "text within TRINS images. The word cloud of the extracted texts is presented in Figure 3b. Figure 4a illustrates the distribution of OCR words per image, indicating that most of the images in TextCap have fewer than 10 words, while the TRINS images average 31.4 OCR words. Recognizing texts within TRINS images is more challenging because of the presence of numerous small words. In summary, TRINS images encompass rich visual content, seamlessly integrating textual information into the image context.", + "bbox": [ + 75, + 231, + 472, + 367 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b68098fd62743113499630aa43e14fa47bcc3c2eeb206808982629c26467f68c.jpg", + "image_caption": [ + "Figure 5. Question type statistics based on key words." + ], + "image_footnote": [], + "bbox": [ + 84, + 378, + 450, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Annotation Details", + "text_level": 1, + "bbox": [ + 76, + 590, + 259, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Annotator Selections All annotators (with the tag $100\\%$ Job Success and Top Rated Plus) are native English speakers and have experience with document annotations. We first asked all annotators to annotate a 200-image set and provide them with detailed annotation guidelines with multiple examples. In addition, we use Labelbox as an annotation tool and set quality control questions.", + "bbox": [ + 75, + 613, + 470, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Heuristic Filters We first use EasyOCR to extract texts from images and retain text phrases with more than three characters, a height greater than $5\\%$ of the canvas height, and a confidence score greater than 0.1. For each phrase retrieved, we employ an edit distance-based string matching algorithm (due to potentially erroneous OCR results) to search for its optimal matching substrings within the human-generated caption. The average score for all extracted phrases serves as a metric.", + "bbox": [ + 75, + 719, + 468, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Manual Reviews We accept annotations with high metric scores and reject the lowest for rework. We manually review other annotations.", + "bbox": [ + 75, + 854, + 470, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sensitive Images We combined neural models with human efforts to filter the images. The first step involves an initial data filtering by 2-3 individuals to filter out sensitive images for training. The second step involves hiring additional people to perform a further screening on the data. We engaged annotators from various countries to check the selected images.", + "bbox": [ + 496, + 231, + 893, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. TRINS-Cap: Text-Rich Image Captioning", + "text_level": 1, + "bbox": [ + 498, + 359, + 859, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Annotation Process TRINS-Cap is a dataset fully annotated by human annotators. We hired 20 native English speakers with experience in document annotation through Upwork. The annotation process, conducted in LabelBox, involved a total of 2,079 hours to annotate 40,576 text-rich images, with an additional 159 hours allocated for result review. After filtering low-quality annotations and addressing missing images, we obtained a final set of 39,153 image-annotation pairs. The dataset is partitioned into train, validation, and test splits with sizes of 29,153, 5,000, and 5,000, respectively. All annotations undergo an initial automated review that involves matching the OCR words with the annotations. Subsequently, human evaluators conduct a thorough review, rejecting annotations with errors, and prompting annotators to rework them for enhancement. We provide comprehensive annotation instructions to all annotators to ensure that each annotation includes: (i) detailed descriptions of visual components. (ii) describe texts' location, attributes, and put texts into annotations. (iii) optional insights or abstract descriptions.", + "bbox": [ + 496, + 387, + 893, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Statistics and Analysis The primary objective of the annotation process is to facilitate a human or machine's full comprehension of the information conveyed in the image without direct viewing. Consequently, the average annotation length for TRINS is 65.1 words, significantly exceeding that of COCO (10.6 words) and TextCaps (12.4 words). Figure 4b shows the caption length distributions for TRINS-Cap, COCO, and TextCap, demonstrating the comprehensive nature of the TRINS-Cap annotations. TRINS with more contexts can generally provide a better description of complex images, where short captions are insufficient. Hence, LLMs finetund on TRINS can better understand images with complex texts and layouts, which has been further verified in Section 5.", + "bbox": [ + 496, + 688, + 895, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22587", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. TRINS-VQA: Multimodal Question Answering", + "text_level": 1, + "bbox": [ + 76, + 90, + 470, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The annotation process for question answering is inherently complex, primarily due to the necessity for annotators to generate high-quality questions. Creating an effective question is more challenging than providing an answer. As a result, annotators frequently gravitate toward formulating concrete and extractive questions (e.g., \"Who is the author of this book?\") rather than abstract ones (e.g., \"How does the design of the book cover reflect the content of the book?\"). We introduce semi-automatic annotation methods to generate high-quality visual question-answering data for TRINS-VQA. This dataset is designed to train general vision language assistants through instruction fine-tuning, and its benefits on model performance are evaluated in Section 5.1. TRINS-Cap, on the other hand, serves as human-assisted annotations, offering a comprehensive but non-instructive dataset for fine-tuning. To utilize the wealth of high-quality annotations available, we incorporated semi-automatic annotation by using large language models (LLMs) such as OpenAI's GPT-4 [37] and Llama-70B [50, 51] to enhance our data annotation pipeline. OCR results and detailed descriptions of each image are provided to LLMs. Furthermore, high-quality human-crafted demonstrations and detailed annotation rules are provided to LLMs. One demonstration focused on extract questions, while the other emphasized abstract questions, creating a more balanced dataset.", + "bbox": [ + 76, + 114, + 473, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Human Annotations To facilitate a robust evaluation of model performance, we hired 10 Upwork annotators, whose native language is English, to annotate the test dataset, following a methodology similar to previous work [13, 34, 47]. The test dataset comprises 5,000 images with 18,764 question-answer pairs. These data collected are used exclusively for evaluation purposes.", + "bbox": [ + 76, + 500, + 470, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Statistics and Analysis Building upon prior research [28, 49, 53], we provide visualizations of instructions in Figure 5 based on question keywords. The inner cycle illustrates the distribution of the first word in the questions, while the outer cycle presents extracted keywords determined by carefully designed heuristics. Types of questions are categorized according to keywords found in questions.", + "bbox": [ + 76, + 613, + 470, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Figure 4c and 4d, we present statistics on the number of words per question-answer pair, comparing them with previous work. Generally, the average length of questions for TRINS-VQA is 10.5, surpassing that of DocVQA (8.3), OCR-VQA (6.5), and TextVQA (7.1). Surprisingly, the average answer length for TRINS is 23.9, significantly longer than related datasets (all less than 4). This discrepancy arises from TRINS containing more abstract questions that typically have longer answers. Similarly, the dataset is divided into train, validation, and test splits. For extract questions, the accuracy of the answers is calculated similarly to Liu et al. [28], while for abstract questions, generation metrics", + "bbox": [ + 76, + 719, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/38df59e7c1c89b20c52f2a641f14a48022598067c0247d375ea5fd9356ab05e8.jpg", + "image_caption": [ + "Figure 6. Model overview of the LaRA. the CLIP model processes the input image to generate patch-wise features. These features then serve as input to a projection layer, yielding visual tokens. Concurrently, an OCR tool extracts textual data from the image, which is then merged with the user instruction." + ], + "image_footnote": [], + "bbox": [ + 527, + 90, + 851, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "such as BLEU scores are used to assess the quality of the answers.", + "bbox": [ + 498, + 359, + 890, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The question-answering and instruction data we obtain are extensive, encompassing a balanced mix of extract and abstract questions. This includes detailed descriptions, summaries, question-answer pairs, tasks that promote creativity and the generation of novel thoughts, and conversational tasks. The dataset spans a diverse range of concepts, ranging from visual presentation and visual language relations to intricate reasoning tasks. Compared to previous methods [28], captions generated by models such as BLIP-2 [22] exhibit less informativeness than detailed human annotations, as demonstrated in Section 5.2. Additionally, captioning models may be susceptible to hallucinations [42], a concern mitigated in TRINS, which provides more comprehensive and reliable visual descriptions. Although OCR tools are robust, they can introduce noisy information. To address this, we utilize both the Azure Read API and PaddleOCR to extract text information. We added the potential unreliability in our system prompts to LLMs, instructing them to generate questions with assured answers. Ultimately, we directly leverage the responses from large language models (LLMs) to construct TRINS-VQA. The quality of the instruction data can be further enhanced through self-alignment [23] or by seeking verification from human annotators. Although this has not been explored in this work, we leave it as a potential direction for future research.", + "bbox": [ + 496, + 391, + 893, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. The Language-vision Reading Assistant (LaRA) Model", + "text_level": 1, + "bbox": [ + 498, + 781, + 890, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To verify our data, we adopt a simple architecture based on LLaVA [28] for language-vision reading assistant, as illustrated in Figure 6. A low-resolution visual encoder is a significant bottleneck for extracting textual information from images [62]. On the contrary, a high-resolution", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22588", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5d4118388ee25eab4de75386866960a5381643c4ebfd47d717cb4f4a740214f0.jpg", + "image_caption": [ + "Figure 7. Examples generated by different multimodal language models on the abstract TRINS-VQA benchmarkk." + ], + "image_footnote": [], + "bbox": [ + 89, + 93, + 305, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Question: Can you describe the design of the book cover?", + "bbox": [ + 320, + 93, + 552, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GT: The book cover is quite vibrant and engaging. The title, \"Explore Your World: Weird, Wild, Amazing!\" is written in large, white fonts that span across the top and bottom of the cover, making it immediately eye-catching. The author's name, Tim Flannery, is prominently displayed, and the illustrator, Sam Caldwell, is credited on the right side of the cover. The background of the cover is filled with various animals, adding a dynamic and exciting element that would likely appeal to children.", + "bbox": [ + 318, + 108, + 846, + 146 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qwen-VL-Chat: The book cover is a black background with various colorful illustrations of animals such as a crocodile, alligator, penguin, and snake in white and green. The title \"Weird Wild, Amazing!\" is written in white letters with a yellow border at the top. The author's name \"Tim Flannery\" is written in white letters above the title, and the publisher's name \"Explore Your World\" is written in green letters below the title. The cover also has a winner sticker for the Environment Children's Literature.", + "bbox": [ + 320, + 150, + 844, + 185 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mPLUG-Owl2: The book cover features a colorful design with a blue background and a yellow bird. The text on the cover is written in a fun and playful font, emphasizing the title \"Tim Flannery: Explore Your World Weird Wild Amazing.\"", + "bbox": [ + 320, + 191, + 844, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GPT-4V: The book cover features a collage of various whimsical and colorful creatures, reminiscent of a style one might find in a children's book. The title \"Weird, Wild, Amazing!\" is prominently displayed at the center top in large, playful, and colorful lettering, which suggests the book may be about the variety and eccentricity of life on Earth. Below the title, the text \"Explore Your World\" is in a smaller font, inviting the reader to discover the contents. The name \"Tim Flannery\" appears above the title, indicating the author, with the descriptor \"Explore Your World\" just below the title. The art by Sam Caldwell is acknowledged at the bottom right, and there's a badge on the bottom left corner stating \"Winner Environment Award for Children's Literature,\" which suggests the book has been recognized for its contribution to environmental education. The background is dark, which makes the bright colors of the animals and other elements stand out. The animals appear to be drawn in a stylized and engaging manner, with expressions and poses that give them character and life. The overall design is vibrant and seems to be targeting a young audience, promising education through fun and engagement with the natural world.", + "bbox": [ + 318, + 215, + 844, + 279 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LaRA: The book cover features the title \"Weird, Wild, Amazing!\" in bold letters, with the author's name, Tim Flannery, mentioned at the top. The backdrop of the cover includes illustrations of various animals, suggesting that the book may be about wildlife or exploration.", + "bbox": [ + 318, + 282, + 844, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "visual encoder generates an excessive number of patches, drastically impacting model efficiency. Our conviction is that a low-resolution visual encoder is designed to capture visual information more effectively, including layouts. In contrast, a considerably smaller OCR tool is employed to extract text from high-resolution images. Instruction-tuning on TRINS could serve as an effective method of training LLMs to better align OCR texts and extracted visual features. In addition, LLMs can autonomously rectify errors generated by OCR tools. For the visual encoder $V$ , we employ CLIP-ViT-L/14-336 at a resolution of $336 \\times 336$ . The grid features before the last transformer layer are then mapped into the word embedding space of the language decoder using a trainable projection matrix $W$ . Regarding the language decoder $D$ , we utilize Vicuna-1.5-13B [8], a language model tuned for instructions through LLaMA 2 [50].", + "bbox": [ + 75, + 349, + 472, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow the two-stage training design of LLaVA but adopt the pre-trained projection layer of LLaVAR [62]. Training targets remain similar: generating output responses $(<\\text{res}>)$ for the input instructions $(<\\text{ins}>)$ , alongside OCR results $(<\\text{ocr}>)$ . The transformed image tokens $(<\\text{img}>)$ are introduced before or after the first input instruction randomly when building the instruction finetuning data. During the finetuning stage, both the projection matrix $W$ and the language decoder $D$ are trained. We consolidate our nearly 90K visual question-answering data with the 158K instruction-following data from LLaVA to form the training set. It should be noted that the visual encoder remains frozen throughout the training period. Compared with previous approaches, LaRA incorporates OCR words as part of the input, a simple way to enhance visual text understanding.", + "bbox": [ + 75, + 592, + 472, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 829, + 209, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we present three downstream tasks based on the TRINS dataset and outline their evaluation metrics. The proposed method LaRA is used in both the text-rich", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0e7aca71eed90e32fab63f8e0c2d48ce103dae80d9c7511ff756d9849337b15e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRecog.VQASVQADKIEFinal Score
Gemini215174128134651
GPT-4v167163146160636
Monkey1741619188514
mPLUG-Owl21531534119366
LLaVAR1861222513346
LLaVA1.5-13B176129197331
mPLUG-Owl172104183297
MiniGPT-V21242940157
LaRA21114785105548
", + "bbox": [ + 506, + 347, + 888, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Results of LMMs on OCRBench. Recog. represents text recognition, $\\mathrm{VQA}^S$ represents Scene Text-Centric VQA, $\\mathrm{VQA}^D$ represents Document-Oriented VQA.", + "bbox": [ + 498, + 502, + 890, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "image summarization and visual question-answer tasks. All experiments were conducted on NVIDIA A100 80GB GPUs. In fine-tuning, we use a cosine annealing schedule with an initial learning rate of $2e^{-5}$ and a batch size of 32.", + "bbox": [ + 498, + 558, + 893, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. TRINS-VQA: Text-Rich Image Visual Question Answering", + "text_level": 1, + "bbox": [ + 498, + 633, + 890, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first performed experiments to evaluate the zero-shot performance of LaRA on classical benchmarks [28, 54]. The results are reported in Table 2 and Table 10. The proposed LaRA model exhibits significant performance improvement across all the datasets compared to other models. Even without OCR, LaRA outperforms other models in most cases, highlighting its robustness and effectiveness in handling visual question-answering tasks. The inclusion of OCR in LaRA further enhances performance, suggesting that Large Language Models (LLMs) can proficiently utilize textual information. However, the ability to directly extract text from images remains limited. LaRA, with its simple structure, significantly boosts model performance on text-rich images, offering an alternative solution to overcome the limitations of pre-trained image encoders. We further evaluate different", + "bbox": [ + 496, + 674, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22589", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/919b4673629bbb617b4650b8d3fd7faae0359c6957ccbc901f358e59573eba09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 89, + 401, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GT: There are big white words \"Urban Sketching\" at the bottom left of the cover. In the upper left corner of the cover, it reads \"Draw What You See.\" In the background of the picture, there is a drawing book and pen. It shows a signature in the bottom right corner of the book.", + "bbox": [ + 416, + 93, + 849, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LLaVA: The image is an urban sketching project displayed on a tabletop, featuring a variety of hand-drawn city scenes. Several people can be seen in the drawing, walking around and engaging in different activities. Among the drawings, there is an image of an alleyway with a prominent Umbrella on the right side.", + "bbox": [ + 416, + 128, + 849, + 165 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LLaVAR: The main subjects in this image are a notebook with a pen placed on top of it, and a photograph of a city street. The notebook and pen appear to be the focal point, with the urban scene in the background.", + "bbox": [ + 416, + 170, + 849, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instruct-BLIP: The main subjects in the image are a series of sketches and drawings depicting urban scenes.", + "bbox": [ + 416, + 205, + 849, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LaRA: This is a blog image. Black text in the upper left corner reads \"Draw What You See\". Large white text in the bottom left reads \"Urban Sketching\". The image background is a blue spiral notebook with a pen in the upper right and a picture of a colorful urban scene in the middle.", + "bbox": [ + 98, + 232, + 849, + 253 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/513a715c3fc31547b95481cdb1111873cd113bd1725e590c7923e5a899cb3519.jpg", + "table_caption": [ + "Figure 8. Examples generated by different multimodal language models on the TRINS-Cap benchmarkk." + ], + "table_footnote": [], + "table_body": "
MethodResolutionExtract AccuracyAbstract
B@1B@2B@3B@4METEORROUGECIDEr
Instruct-BLIP [11]\\( {224}^{2} \\)43.913.59.26.75.28.916.723.5
Mini-GPTv215.329.419.413.810.417.228.744.3
mPLUG-Owl2 [57]\\( {448}^{2} \\)61.035.925.519.315.217.738.097.2
Qwen-VL [3]63.640.227.319.915.220.436.479.4
LLaVA [28]23.733.120.914.410.518.031.848.2
LLaVAR [62]\\( {336}^{2} \\)51.738.025.919.114.717.435.684.9
LLaVAR w/ OCR58.140.328.020.916.318.737.397.3
LLaVA 1.5 [27]38.829.820.915.812.620.834.940.7
LLaVAR (finetuned)\\( {336}^{2} \\)61.245.331.923.918.720.439.6104.7
LaRA62.845.134.227.322.621.946.5186.6
", + "bbox": [ + 98, + 284, + 869, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Results of different models on TRINS-VQA for text-rich image question-answering tasks.", + "bbox": [ + 189, + 464, + 774, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods on the TRINS-VQA dataset, as shown in Table 3. For extraction questions, we use the same metric as Wu et al. [54]. For abstract questions, where the answer is typically a longer sentence, we evaluate them based on text similarity metrics such as BLEU [25], ROUGE [25], and CIDEr [52]. In zero-shot inference, LLaVAR with OCR exhibits the best performance, reinforcing the importance of extracting textual information. Furthermore, mPLUG-Owl2 and Qwen-VL perform well and represent the best methods in extract question evaluations, showing that a high-resolution encoder can significantly improve model performance. Instruct-BLIP demonstrates good performance on extract questions, but did not fare as well on abstract questions, given that the answers provided are usually short and concise. Figure 7 shows an example of the responses of different methods on the abstract TRINS-VQA dataset, and more examples can be found in the Appendix D. Qwen-VL includes all details but does not provide high-level insights, such as the ground-truth annotation. Both mPLUG-Owl2 and GPT-4V suffer from hallucination issues.", + "bbox": [ + 75, + 493, + 472, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. TRINS-Cap: Text-rich image Captioning", + "text_level": 1, + "bbox": [ + 76, + 801, + 429, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our experiments on TRINS-Cap, we ask large multimodal models to generate summaries based on text-rich images. The data set was divided into train, validation, and test sets. We compared LaRA with popular baselines, including InstructBLIP [11], Mini-GPT4 [63], LLaVA [28],", + "bbox": [ + 75, + 824, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LLaVAR [62], mPLUG-Owl2 [56] and Qwen-VL [3]. Given that BLIP-2 faced challenges in generating comprehensive and meaningful results for text-rich images, we considered InstructBLIP as an alternative. For all methods, we randomly selected three prompts from ten as instructions for the model (details provided in the Appendix C.1).", + "bbox": [ + 496, + 493, + 890, + 584 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 presents the results of different methods in terms of classical captioning metrics. Models with enhanced visual text understanding generally outperform general multimodal models, such as LLaVA, Mini-GPT4, and Instruct-BLIP. LaRA (zero-shot) refers to the LaRA model fine-tuned on the TRINS-QA dataset and demonstrates improved performance. Comparison of fine-tuned LaRA variants indicates that text recognition ability is still limited for OCR-free methods, suggesting that the CLIP encoder or feature projection process may cause visual information loss. Addressing this limitation may involve employing a better trained encoder on text-rich images or designing a more carefully crafted architecture, a direction we leave for future exploration. When fine-tuned with TRINS-Cap, LaRA exhibits much better performance, underscoring the importance of high-quality human-annotated data. Figure 8 shows examples of different models on the TRINS-Cap benchmark. It shows the great capability of LaRA in recognizing text and relating it to its visual contexts, demonstrating the effectiveness of the TRINS dataset.", + "bbox": [ + 496, + 598, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22590", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/88e8b677e6b78b56fb6d842481ec7ea6a44ffb984ab468765ddb39c5c30505da.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneB@1B@2B@3B@4METEORROUGECIDEr
mPLUG-Owl2 [57]Llama-2-7B4.93.22.31.78.719.64.5
Instruct-BLIP [11]Vicuna-7B13.48.55.74.19.217.95.7
Qwen-VL [3]Qwen-7B28.818.412.28.614.124.316.6
Instruct-BLIP [11]15.99.96.44.49.618.68.0
Mini-GPT4 [63]31.116.18.65.011.420.86.3
Mini-GPT-v2 [7]Vicuna-13B27.914.78.04.810.820.97.3
LLaVA [28]35.118.29.55.413.222.28.8
LLaVAR [62]18.911.27.25.010.820.111.4
LLaVA 1.5 [27]31.116.49.56.111.921.815.4
LLaVAR w/ OCRVicuna-13B21.413.18.86.412.022.213.0
LaRA (zero-shot)29.319.112.99.214.826.321.3
LLaVAR (fine-tuned)Vicuna-13B36.525.418.013.417.832.435.7
LaRA37.726.418.914.218.433.246.7
", + "bbox": [ + 89, + 88, + 877, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Additional Experiments", + "text_level": 1, + "bbox": [ + 76, + 328, + 299, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance on general visual tasks after TRINS fine-tuning. We adopted the evaluation protocols of MiniGPT-v2 [7] and compared LaRA with LLaVA [28] on traditional visual question answering benchmarks in table 5. LaRA shows a comparable performance on knowledgeability and better performance on reasoning and spatial awareness. This further verifies the effectiveness of the TRINS dataset, demonstrating that fine-tuning on text-rich images does not degrade performance on natural images, but instead enhances the results.", + "bbox": [ + 75, + 351, + 473, + 502 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0821fc75c608feb5722e7392fee5ceea3eeb058ed2ccd4dcfcf65bd3de1ebef7.jpg", + "table_caption": [ + "Table 4. Results of different models on text-rich image captioning tasks." + ], + "table_footnote": [], + "table_body": "
OKVQAGQAVSRVizWiz
LLaVA [28]57.841.351.245.0
LaRA58.142.453.053.1
", + "bbox": [ + 102, + 507, + 442, + 566 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4f5ba0ae4e88177a07c6039e5454fb396b0bde33a9cf78c89599950b21c1400f.jpg", + "table_caption": [ + "Table 5. Quantitative Results on the public visual benchmarks." + ], + "table_footnote": [], + "table_body": "
MetricsControlNetDeepFloydTextDiffuser
FID (↓)51.5949.9651.26
CLIP Score (↑)0.37170.39170.3707
OCR Acc. (↑)0.42410.21920.5027
", + "bbox": [ + 83, + 601, + 460, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Empirical Results on TRINS-Gen (easy) benchmark.", + "bbox": [ + 88, + 678, + 455, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text-to-document generation Diffusion-based text-to-image generation has shown great success, while precise textual renderings remain a big challenge. TextDiffuser introduced the MARIO-Eval benchmark, drawing from works such as DrawBench [43] and DrawTextCreative [29]. However, most text prompts in MARIO-Eval are short and cannot serve as a good evaluation dataset to handle complex real-world human instructions. We take advantage of human annotations from TRINS-Cap and build the TRINS-Gen benchmark. It is still difficult to render too many words in a single image [6]. In response to this, we filter out images with more than 20 OCR words, resulting in a curated set of 2,104 images. We divide these images into two sets (easy", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and difficult) based on the number of OCR words and the length of the longest OCR string per annotation, where all text prompts in the easy set have less than 9 OCR words. We evaluated existing methods using their public checkpoints and reported the results in Table 6 and detailed results in Table 9 (Appendix C.3).", + "bbox": [ + 496, + 329, + 893, + 421 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d14b7ec39e5dcd0c0af0d8040545cc064affff63bec2753311ee8ee8c6723be9.jpg", + "image_caption": [ + "Text Prompts: The image is a book cover for 'Abandoned San Diego' by Jessica D. Johnson, featuring a background image of an abandoned house.", + "Ground Truth", + "Figure 9. Examples generated by different text-to-image models on the TRINS-Gen benchmarkk." + ], + "image_footnote": [], + "bbox": [ + 506, + 455, + 586, + 529 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/77c326b75a08897e5707afb1967f5638896f5fe815fc5020a09a5dcf670f10ba.jpg", + "image_caption": [ + "ControlNet" + ], + "image_footnote": [], + "bbox": [ + 589, + 455, + 686, + 527 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6efe6c91bee5bc0a943c4fe7ef9b58e15adc8e74234a93385aefed8b51987df7.jpg", + "image_caption": [ + "Deepfloyd" + ], + "image_footnote": [], + "bbox": [ + 692, + 455, + 787, + 527 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cdf991f751a98211d980d47cc20151ecab9dad8d5c8545126d8e660952705b7a.jpg", + "image_caption": [ + "TextDiffuser" + ], + "image_footnote": [], + "bbox": [ + 792, + 455, + 887, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 602, + 627, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite the challenges posed by the prevalence of natural images in training data, the significance of visual textual understanding cannot be understated. In this paper, we introduce TRINS, a Text-Rich Image INStruction dataset, comprising a diverse collection of text-rich images, captions, and questions. This dataset, created through a semi-automatic process leveraging large-scale pre-trained models, not only significantly reduces annotation time but also elevates annotation quality. Furthermore, we propose a novel multimodal language model architecture, LaRA, which incorporates OCR as a pivotal enhancement for text-rich image understanding. We anticipate that continued progress in multimodal language model architectures, fine-tuning techniques, and the expansion of diverse, text-rich datasets like TRINS will push the boundaries of visual textual understanding. This, in turn, will facilitate more efficient collaboration between humans and agents, potentially revolutionizing numerous real-world applications.", + "bbox": [ + 496, + 628, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22591", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1", + "[2] Anas Awadalla, Irena Gao, Joshua Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortzman, and Ludwig Schmidt. Openflamingo, 2023. 22", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.7, 8", + "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 12", + "[5] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts, 2021. 1", + "[6] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. arXiv preprint arXiv:2305.10855, 2023. 1, 8, 12", + "[7] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 8", + "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality, 2023. 6", + "[9] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling instruction-finetuned language models, 2022. 1", + "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2", + "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructblip: Towards" + ], + "bbox": [ + 78, + 114, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "general-purpose vision-language models with instruction tuning. ArXiv, abs/2305.06500, 2023. 7, 8", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2020. 1", + "[13] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marcal Rusinol, Minesh Mathew, C.V. Jawahar, Ernest Valveny, and Dimosthenis Karatzas. Icdar 2019 competition on scene text visual question answering. 2019 International Conference on Document Analysis and Recognition (ICDAR), 2019. 2, 5", + "[14] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient visual instruction model, 2023. 2", + "[15] Adam W. Harley, Alex Ufkes, and Konstantinos G. Derpanis. Evaluation of deep convolutional nets for document image classification and retrieval, 2015. 3", + "[16] Rongjie Huang, Mingze Li, Dongchao Yang, Jiatong Shi, Xuankai Chang, Zhenhui Ye, Yuning Wu, Zhiqing Hong, Jia-Bin Huang, Jinglin Liu, Yixiang Ren, Zhou Zhao, and Shinji Watanabe. Audiogpt: Understanding and generating speech, music, sound, and talking head. ArXiv, abs/2304.12995, 2023. 2", + "[17] Xinyu Huang, Youcai Zhang, Jinyu Ma, Weiwei Tian, Rui Feng, Yuejie Zhang, Yaqian Li, Yandong Guo, and Lei Zhang. Tag2text: Guiding vision-language model via image tagging. arXiv preprint arXiv:2303.05657, 2023. 3", + "[18] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 1", + "[19] Chunyuan Li. Large multimodal models: Notes on cvpr 2023 tutorial. ArXiv, abs/2306.14895, 2023. 1", + "[20] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 1, 2023. 2", + "[21] Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, and Furu Wei. Dit: Self-supervised pre-training for document image transformer. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 3", + "[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023. 5, 22", + "[23] Xian Li, Ping Yu, Chunting Zhou, Timo Schick, Luke Zettlemoyer, Omer Levy, Jason Weston, and Mike Lewis. Self-alignment with instruction backtranslation. arXiv preprint arXiv:2308.06259, 2023. 5", + "[24] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. arXiv preprint arXiv:2311.06607, 2023. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22592", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 7", + "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 1", + "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 7, 8", + "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 2, 5, 6, 7, 8, 22", + "[29] Rosanne Liu, Dan Garrette, Chitwan Sahara, William Chan, Adam Roberts, Sharan Narang, Irina Blok, RJ Mical, Mohammad Norouzi, and Noah Constant. Character-aware models improve visual text rendering. arXiv preprint arXiv:2212.10562, 2022.8, 12", + "[30] Yuliang Liu, Zhang Li, Hongliang Li, Wenwen Yu, Mingxin Huang, Dezhi Peng, Mingyu Liu, Mingrui Chen, Chunyuan Li, Cheng lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models, 2023. 1, 2, 22", + "[31] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Learning to draw Chinese characters in image synthesis models coherently, 2023. 12", + "[32] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models, 2023. 2", + "[33] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 2", + "[34] Minesh Mathew, Dimosthenis Karatzas, and C. V. Jawahar. Docvqa: A dataset for vqa on document images, 2020. 5", + "[35] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 2", + "[36] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947–952. IEEE, 2019. 2", + "[37] OpenAI. Gpt-4 technical report, 2023. 1, 5", + "[38] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback, 2022. 1", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1, 3" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2", + "[41] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 12", + "[42] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. arXiv preprint arXiv:1809.02156, 2018. 5", + "[43] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 8, 12", + "[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022.3", + "[45] Alex Shonenkov, Misha Konstantinov, Daria Bakshandaeva, Christoph Schuhmann, Ksenia Ivanova, and Nadija Klokova. Deepfloyd. https://github.com/deep-floyd/if, 2023.12", + "[46] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 742-758. Springer, 2020. 2", + "[47] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 5", + "[48] Amanpreet Singh, Guan Pang, Mandy Toh, Jing Huang, Wojciech Galuba, and Tal Hassner. Textocr: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8802-8812, 2021. 2", + "[49] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023.5", + "[50] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 5, 6" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22593", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 5", + "[52] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 7", + "[53] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions, 2022. 5", + "[54] Zizhang Wu, Xinyuan Chen, Jizheng Wang, Xiaoquan Wang, Yanzhu Gan, Muqing Fang, and Tianhao Xu. Ocr-rtps: anOCR-based real-time positioning system for the valet parking. Applied Intelligence, pages 1-15, 2023. 6, 7", + "[55] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 2", + "[56] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 2, 7, 22", + "[57] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. arXiv preprint arXiv:2311.04257, 2023. 7, 8, 22", + "[58] Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities, 2023. 2", + "[59] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2", + "[60] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention, 2023. 2", + "[61] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023. 3", + "[62] Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 1, 2, 5, 6, 7, 8, 22", + "[63] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 1, 2, 7, 8, 22" + ], + "bbox": [ + 78, + 90, + 470, + 881 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22594", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/TRINS_ Towards Multimodal Language Models that Can Read/527c7573-3fb6-42ad-a14f-b7b957e97243_model.json b/2024/TRINS_ Towards Multimodal Language Models that Can Read/527c7573-3fb6-42ad-a14f-b7b957e97243_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6cfe324c9cefdb46d348466fb2d40e571f7d5f87 --- /dev/null +++ b/2024/TRINS_ Towards Multimodal Language Models that Can Read/527c7573-3fb6-42ad-a14f-b7b957e97243_model.json @@ -0,0 +1,2312 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.168, + 0.131, + 0.804, + 0.152 + ], + "angle": 0, + "content": "TRINS: Towards Multimodal Language Models that Can Read" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.18, + 0.804, + 0.217 + ], + "angle": 0, + "content": "Ruiyi Zhang\\(^{1}\\), Yanzhe Zhang\\(^{2}\\), Jian Chen\\(^{3}\\), Yufan Zhou\\(^{1}\\), Jiumiang Gu\\(^{1}\\), Changyou Chen\\(^{3}\\), Tong Sun\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.217, + 0.882, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\) Adobe Research \\(^{2}\\) Georgia Institute of Technology \\(^{3}\\) State University of New York at Buffalo" + }, + { + "type": "text", + "bbox": [ + 0.4, + 0.237, + 0.578, + 0.252 + ], + "angle": 0, + "content": "{ruizhang}@adobe.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.318, + 0.474, + 0.651 + ], + "angle": 0, + "content": "Large multimodal language models have shown remarkable proficiency in understanding and editing images. However, a majority of these visually-tuned models struggle to comprehend the textual content embedded in images, primarily due to the limitation of training data. In this work, we introduce TRINS: a Text-Rich image1 INstruction dataset, with the objective of enhancing the reading ability of the multimodal large language model. TRINS is built upon LAION2 using hybrid data annotation strategies that include machine-assisted and human-assisted annotation process. It contains 39,153 text-rich images, captions, and 102,437 questions. Specifically, we show that the number of words per annotation in TRINS is significantly longer than that of related datasets, providing new challenges. Furthermore, we introduce a simple and effective architecture, called a Language-Vision Reading Assistant (LaRA), which is good at understanding textual content within images. LaRA outperforms existing state-of-the-art multimodal large language models on the TRINS dataset as well as other classical benchmarks. Lastly, we conducted a comprehensive evaluation with TRINS on various text-rich image understanding and generation tasks, demonstrating its effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.678, + 0.21, + 0.695 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.856 + ], + "angle": 0, + "content": "Instruction tuning [9, 38] has shown a great generalization ability on unseen tasks and has contributed to the growing popularity of large language models (LLMs), such as ChatGPT [37]. Recently, multimodal language models benefit from visual instruction finetuning [1, 18, 19, 28, 63], and has shown great success in real-world applications. These models leverage visual encoders such as CLIP-ViT [12, 39] to empower LLMs with image comprehension ability. However, challenges arise in comprehension of textual information within images, which may stem from the prevalence of natu" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.288, + 0.895, + 0.455 + ], + "angle": 0, + "content": "ral images in training datasets, such as Conceptual Captions [5] and COCO [26]), as highlighted by Liu et al. [30]. Recognizing the importance of visual textual understanding for effective collaboration between agents and humans, Zhang et al. [62] proposed enhancing end-to-end visual instruction-tuned models by introducing noisy Optical Character Recognition (OCR) annotations to improve vision-language alignment. In this work, we surpass existing achievements and collect a new Text-Rich image INStruction dataset named TRINS, which contains 39,153 text-rich images, captions and 102,437 questions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.897, + 0.903 + ], + "angle": 0, + "content": "TRINS is created in a semi-automatic manner for a more controllable and faithful collection. Specifically, we exploited large-scale pre-trained models such as CLIP [40] and GPT-4 in the annotation process. This semi-automatic process significantly reduces the time and resources required for manual annotation and surprisingly improves the overall quality of annotations. TRINS dataset is composed of three datasets for captioning, visual question answering (VQA) and image generation, respectively. Specifically, human-annotated captions for text-rich images are first collected because they can best translate text-rich images into texts. During this process, extracted OCR words and recognize-anything model tags are provided to the annotators for better and efficient annotations. With detailed image descriptions, VAQ data is built and fulfilled by large language models, such as GPT-4 [37] and LLaMA-2 [51]. In detailed statistics and analysis, we found that both annotated captions and collected question-answer pairs are more comprehensive and contain significantly more details than the existing dataset. Therefore, we show the superior advantage of TRINS compared to existing instruction fine-tuning datasets. As a by-product, high-quality image-caption pairs can serve as a good benchmark for text-rich image generation, which is still a very challenging task [6]. At the same time, we propose a new simple and effective multimodal language model architecture that includes OCR as a component. We call it Language-vision Reading Assistant (LaRA) and show that LaRA fine-tuned on TRINS brings the best text-rich image understanding ability. Our contributions are as follows:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.469, + 0.888 + ], + "angle": 0, + "content": "1In this work, we use the phrase \"text-rich images\" to describe images with rich textual information, such as posters and book covers." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.251, + 0.9 + ], + "angle": 0, + "content": "2Work done during Q3 2023." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.863, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22584" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "- We introduce a novel dataset (TRINS) containing 39,153 captions and 102,437 high-quality text-rich image instruction pairs. TRINS is annotated with a novel semi-automatic annotation framework that is scalable and reliable." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.167, + 0.47, + 0.226 + ], + "angle": 0, + "content": "- We develop several evaluation benchmarks for text-rich image understanding and generation tasks. Various methods are evaluated on TRINS, demonstrating the effectiveness of the dataset." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.228, + 0.47, + 0.318 + ], + "angle": 0, + "content": "- The TRINS datasets are high-quality and comprehensive, which is reflected not only in the dataset statistics but also from the results of multiple baseline models. Especially, our proposed LaRA finetuned on TRINS outperforms existing state-of-the-art methods on text-rich image understanding tasks." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.092, + 0.47, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.33, + 0.218, + 0.346 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.355, + 0.471, + 0.718 + ], + "angle": 0, + "content": "Multimodal Instruction Tuning Multi-modal instruction tuning, including image, video [32, 59], and audio [16, 58] settings, has been an active research topic. MiniGPT-4 [63] uses ChatGPT to generate high-quality instruction-following data, while LLaVA [28] generates such data by prompting GPT-4 with captions and bounding boxes. LLaMA-Adapter [14, 60] aligns text-image features using COCO data, and mPLUG-owl [56] combines extensive image-text pairs for pretraining and a mixture of data for finetuning. Despite this, many models, according to Liu et al. [30], struggle with OCR tasks. InstructBLIP [10] addresses this by transforming 13 vision language tasks into an instruction-following format. mPLUGOwl [55, 56] apply multitask instruction funetuing using existing document datasets. A comprehensive survey is available in Li et al. [20]. LLaVAR [62] exploits GPT-4 to collect fine-tuning data without human annotations using OCR and captioning tools. It discovered that resolution plays a significant role in recognizing textual information and explored several options. Monkey [24] performed a surgery between simple text labels and high input resolution, enabling remarkable performance in visually-rich document images with dense text. TRINS exploits human-machine collaboration for data annotation and can provide more accurate information, reducing the problem of hallucination." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.735, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Text-Rich Image Datasets Visual question answering or captioning datasets are widely used in task-specific finetuning and large multimodal model evaluation. TextCap [46] is the first text-rich image captioning dataset. Compared to TextCap, TRINS-Cap provides more detailed annotations that can fulfill the requirement of instruction finetuning. TextOCR [48] aims comprehend text in the context of an image, which is similar to our motivation but focuses more on text recognition in images instead of understanding. STVQA [13] uses spatial and textual information to answer visually grounded questions, effectively integrating visual" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.88, + 0.226 + ], + "angle": 0, + "content": "
DatasetYearSizeAnnotation Type
OCR-VQA2019200KQA (▲)
TextVQA201945KQA (▲)
TextCap2020140kCaption (▲)
TextOCR2021145kText Bbox (▲)
DocVQA202050kQA (▲)
40kCaption (▲)
TRINS (Ours)2023100kQA (▲ +▲)
40kText Bbox (▲)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.504, + 0.236, + 0.886, + 0.25 + ], + "angle": 0, + "content": "Table 1. Comparison between TRINS and other related datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.268, + 0.895, + 0.493 + ], + "angle": 0, + "content": "and textual cues. OCR-VQA [36] focuses on incorporating optical character recognition (OCR) into visual question answering (VQA), which operates primarily on text within images. TextVQA [47] also takes advantage of the textual information present in the images to answer questions, but with an emphasis on open questions. DocVQA takes this one step further by applying VQA to document images, handling a variety of layouts and formats. InfoVQA [35] and ChartQA [33] focus on specific subdomains and aim to answer questions about information graphics and chart images, respectively. In summary, these related works provide datasets for leveraging spatial and textual cues. TRINS-VQA is a dataset that exploits the semi-automatic annotation process. It can be used for general domain instruction finetuning and model evaluations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.511, + 0.831, + 0.529 + ], + "angle": 0, + "content": "3. Text-Rich Image Instruction Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.895, + 0.825 + ], + "angle": 0, + "content": "To equip multimodal language models with the ability to recognize text and relate it to its visual context, we have curated a new dataset named Text-Rich Instruction (TRINS). The ultimate goal is to enable these models to have spatial, semantic, and visual reasoning between multiple text tokens and visual entities. In this section, we present TRINS, a dataset crafted through a semi-automatic process. Specifically, we leverage large-scale pre-trained models like CLIP [40] and GPT-4 in the annotation process, offering potential advantages: (i) Significant reduction in annotation time and resources: using these models significantly reduces the time and resources required for manual annotation. (ii) Enhancement of annotation data quality through post-processing: the involvement of large models contributes to improving the overall quality of annotation data through subsequent post-processing. (iii) Functionality of large models as knowledge bases: large models can serve as effective knowledge bases, aiding in the annotation process by virtue of their extensive training in diverse datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We first outline the document image collection process for TRINS, utilizing CLIP models. Then, we present data statistics to facilitate a comprehensive understanding. We delve into three distinct tasks derived from the TRINS dataset in detail: i) TRINS-Cap: Visual Captioning, ii) TRINS-VQA:" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22585" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.891, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.224, + 0.893, + 0.281 + ], + "angle": 0, + "content": "Figure 1. Overview of TRINS data collection process, which consists of three datasets. Text-rich images are first selected from web images and then ask annotators to describe the image in details. i) TRINS-Cap is extracted from human annotations with heuristic data processing for text-rich image summarization tasks. ii) TRINS-VQA is built upon human annotations and generate question-answer pairs for training by prompting text-only large language models. iii) TRINS-Gen combined human annotations and text boxes for text-rich image generation." + }, + { + "type": "image", + "bbox": [ + 0.077, + 0.283, + 0.877, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.442, + 0.843, + 0.457 + ], + "angle": 0, + "content": "Figure 2. CLIP-based categorization of our collected images and selected representative data samples from each category." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.473, + 0.592 + ], + "angle": 0, + "content": "Visual Question Answering and iii) TRINS-Gen: Text-to-Image Generation. The overview of the TRINS data collection process is illustrated in Figure 1. To provide a succinct overview, TRINS-Cap undergoes full annotation by human annotators, while TRINS-Gen and TRINS-VQA are constructed based on TRINS-Cap with the help of pre-trained models. A summary of the various TRINS datasets is presented in Table 7 of Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.601, + 0.456, + 0.617 + ], + "angle": 0, + "content": "3.1. Machine-Assisted Text-rich Image Selections" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.472, + 0.867 + ], + "angle": 0, + "content": "Beginning with the LAION-5B dataset [44], our objective is to selectively retain images that exhibit a significant presence of text. Recognizing that document images typically contain substantial textual content, we initially formed a binary classification dataset by combining natural images with document data. Subsequently, we trained an image classifier using a DiT [21] base backbone, fine-tuned on the RVL-CDIP dataset [15]. The purpose of this classifier was to predict whether an image contains text. Then a subset was constructed by selecting images with a predicted probability greater than 0.8, while also adhering to the criteria \\( p(\\text{watermark}) < 0.8 \\) and \\( p(\\text{unsafe}) < 0.5 \\), where both probabilities are derived from the metadata of the LAION dataset. Acknowledging the noise introduced due to the classifier's limitations, we further refined the dataset by incorporating human judgment. A random sample of 20,000 images" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.468, + 0.695, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.573, + 0.608, + 0.584 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.469, + 0.892, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.573, + 0.804, + 0.584 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.587, + 0.892, + 0.614 + ], + "angle": 0, + "content": "Figure 3. Word clouds of (a) predicted tags and (b) detected words from the text-rich images of TRINS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.895, + 0.862 + ], + "angle": 0, + "content": "from the filtered LAION-5B was clustered into 50 groups based on CLIP-ViT-B/32 visual features. After inspecting the clustering results, one cluster was meticulously chosen, encompassing diverse text-rich images such as posters, covers, advertisements, and educational documents. This cluster model then served as the filtering mechanism for collecting images that comprise the TRINS dataset. For reference, we present a CLIP-based categorization [39] in Figure 2 to depict the distribution of images in the collected data. The major class is book cover images, further categorized on the basis of book themes and contents. To enhance our understanding of text-rich images, we employed a Recognize Anything Model (RAM) [17, 61] to extract tags from TRINS images. Figure 3a displays word clouds of RAM tags, where \"book\" and \"poster\" emerge as major keywords. Additionally, we utilize the Azure Read API \\(^{4}\\) and PaddleOCR to extract" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.473, + 0.9 + ], + "angle": 0, + "content": "3https://huggingface.co/datasets/laion/laion-high-resolution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.501, + 0.875, + 0.893, + 0.901 + ], + "angle": 0, + "content": "4https://azure.microsoft.com/en-us/updates/computer-vision-v3-preview-6/" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22586" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.092, + 0.278, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.198, + 0.194, + 0.206 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.092, + 0.48, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.198, + 0.392, + 0.207 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.092, + 0.675, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.198, + 0.59, + 0.207 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.092, + 0.872, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.198, + 0.789, + 0.207 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.208, + 0.744, + 0.221 + ], + "angle": 0, + "content": "Figure 4. OCR word (a), Caption (b), Question (c) and Answer (d) statistics for TRINS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.473, + 0.368 + ], + "angle": 0, + "content": "text within TRINS images. The word cloud of the extracted texts is presented in Figure 3b. Figure 4a illustrates the distribution of OCR words per image, indicating that most of the images in TextCap have fewer than 10 words, while the TRINS images average 31.4 OCR words. Recognizing texts within TRINS images is more challenging because of the presence of numerous small words. In summary, TRINS images encompass rich visual content, seamlessly integrating textual information into the image context." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.379, + 0.451, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.565, + 0.434, + 0.58 + ], + "angle": 0, + "content": "Figure 5. Question type statistics based on key words." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.26, + 0.606 + ], + "angle": 0, + "content": "3.2. Annotation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.72 + ], + "angle": 0, + "content": "Annotator Selections All annotators (with the tag \\(100\\%\\) Job Success and Top Rated Plus) are native English speakers and have experience with document annotations. We first asked all annotators to annotate a 200-image set and provide them with detailed annotation guidelines with multiple examples. In addition, we use Labelbox as an annotation tool and set quality control questions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Heuristic Filters We first use EasyOCR to extract texts from images and retain text phrases with more than three characters, a height greater than \\(5\\%\\) of the canvas height, and a confidence score greater than 0.1. For each phrase retrieved, we employ an edit distance-based string matching algorithm (due to potentially erroneous OCR results) to search for its optimal matching substrings within the human-generated caption. The average score for all extracted phrases serves as a metric." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.9 + ], + "angle": 0, + "content": "Manual Reviews We accept annotations with high metric scores and reject the lowest for rework. We manually review other annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.232, + 0.895, + 0.339 + ], + "angle": 0, + "content": "Sensitive Images We combined neural models with human efforts to filter the images. The first step involves an initial data filtering by 2-3 individuals to filter out sensitive images for training. The second step involves hiring additional people to perform a further screening on the data. We engaged annotators from various countries to check the selected images." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.36, + 0.861, + 0.377 + ], + "angle": 0, + "content": "3.3. TRINS-Cap: Text-Rich Image Captioning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.895, + 0.689 + ], + "angle": 0, + "content": "Annotation Process TRINS-Cap is a dataset fully annotated by human annotators. We hired 20 native English speakers with experience in document annotation through Upwork. The annotation process, conducted in LabelBox, involved a total of 2,079 hours to annotate 40,576 text-rich images, with an additional 159 hours allocated for result review. After filtering low-quality annotations and addressing missing images, we obtained a final set of 39,153 image-annotation pairs. The dataset is partitioned into train, validation, and test splits with sizes of 29,153, 5,000, and 5,000, respectively. All annotations undergo an initial automated review that involves matching the OCR words with the annotations. Subsequently, human evaluators conduct a thorough review, rejecting annotations with errors, and prompting annotators to rework them for enhancement. We provide comprehensive annotation instructions to all annotators to ensure that each annotation includes: (i) detailed descriptions of visual components. (ii) describe texts' location, attributes, and put texts into annotations. (iii) optional insights or abstract descriptions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.896, + 0.9 + ], + "angle": 0, + "content": "Statistics and Analysis The primary objective of the annotation process is to facilitate a human or machine's full comprehension of the information conveyed in the image without direct viewing. Consequently, the average annotation length for TRINS is 65.1 words, significantly exceeding that of COCO (10.6 words) and TextCaps (12.4 words). Figure 4b shows the caption length distributions for TRINS-Cap, COCO, and TextCap, demonstrating the comprehensive nature of the TRINS-Cap annotations. TRINS with more contexts can generally provide a better description of complex images, where short captions are insufficient. Hence, LLMs finetund on TRINS can better understand images with complex texts and layouts, which has been further verified in Section 5." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22587" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.471, + 0.108 + ], + "angle": 0, + "content": "3.4. TRINS-VQA: Multimodal Question Answering" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.474, + 0.493 + ], + "angle": 0, + "content": "The annotation process for question answering is inherently complex, primarily due to the necessity for annotators to generate high-quality questions. Creating an effective question is more challenging than providing an answer. As a result, annotators frequently gravitate toward formulating concrete and extractive questions (e.g., \"Who is the author of this book?\") rather than abstract ones (e.g., \"How does the design of the book cover reflect the content of the book?\"). We introduce semi-automatic annotation methods to generate high-quality visual question-answering data for TRINS-VQA. This dataset is designed to train general vision language assistants through instruction fine-tuning, and its benefits on model performance are evaluated in Section 5.1. TRINS-Cap, on the other hand, serves as human-assisted annotations, offering a comprehensive but non-instructive dataset for fine-tuning. To utilize the wealth of high-quality annotations available, we incorporated semi-automatic annotation by using large language models (LLMs) such as OpenAI's GPT-4 [37] and Llama-70B [50, 51] to enhance our data annotation pipeline. OCR results and detailed descriptions of each image are provided to LLMs. Furthermore, high-quality human-crafted demonstrations and detailed annotation rules are provided to LLMs. One demonstration focused on extract questions, while the other emphasized abstract questions, creating a more balanced dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.5, + 0.472, + 0.606 + ], + "angle": 0, + "content": "Human Annotations To facilitate a robust evaluation of model performance, we hired 10 Upwork annotators, whose native language is English, to annotate the test dataset, following a methodology similar to previous work [13, 34, 47]. The test dataset comprises 5,000 images with 18,764 question-answer pairs. These data collected are used exclusively for evaluation purposes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.614, + 0.472, + 0.72 + ], + "angle": 0, + "content": "Statistics and Analysis Building upon prior research [28, 49, 53], we provide visualizations of instructions in Figure 5 based on question keywords. The inner cycle illustrates the distribution of the first word in the questions, while the outer cycle presents extracted keywords determined by carefully designed heuristics. Types of questions are categorized according to keywords found in questions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.472, + 0.902 + ], + "angle": 0, + "content": "In Figure 4c and 4d, we present statistics on the number of words per question-answer pair, comparing them with previous work. Generally, the average length of questions for TRINS-VQA is 10.5, surpassing that of DocVQA (8.3), OCR-VQA (6.5), and TextVQA (7.1). Surprisingly, the average answer length for TRINS is 23.9, significantly longer than related datasets (all less than 4). This discrepancy arises from TRINS containing more abstract questions that typically have longer answers. Similarly, the dataset is divided into train, validation, and test splits. For extract questions, the accuracy of the answers is calculated similarly to Liu et al. [28], while for abstract questions, generation metrics" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.091, + 0.852, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.277, + 0.895, + 0.348 + ], + "angle": 0, + "content": "Figure 6. Model overview of the LaRA. the CLIP model processes the input image to generate patch-wise features. These features then serve as input to a projection layer, yielding visual tokens. Concurrently, an OCR tool extracts textual data from the image, which is then merged with the user instruction." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.361, + 0.892, + 0.389 + ], + "angle": 0, + "content": "such as BLEU scores are used to assess the quality of the answers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.895, + 0.768 + ], + "angle": 0, + "content": "The question-answering and instruction data we obtain are extensive, encompassing a balanced mix of extract and abstract questions. This includes detailed descriptions, summaries, question-answer pairs, tasks that promote creativity and the generation of novel thoughts, and conversational tasks. The dataset spans a diverse range of concepts, ranging from visual presentation and visual language relations to intricate reasoning tasks. Compared to previous methods [28], captions generated by models such as BLIP-2 [22] exhibit less informativeness than detailed human annotations, as demonstrated in Section 5.2. Additionally, captioning models may be susceptible to hallucinations [42], a concern mitigated in TRINS, which provides more comprehensive and reliable visual descriptions. Although OCR tools are robust, they can introduce noisy information. To address this, we utilize both the Azure Read API and PaddleOCR to extract text information. We added the potential unreliability in our system prompts to LLMs, instructing them to generate questions with assured answers. Ultimately, we directly leverage the responses from large language models (LLMs) to construct TRINS-VQA. The quality of the instruction data can be further enhanced through self-alignment [23] or by seeking verification from human annotators. Although this has not been explored in this work, we leave it as a potential direction for future research." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.782, + 0.892, + 0.816 + ], + "angle": 0, + "content": "4. The Language-vision Reading Assistant (LaRA) Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To verify our data, we adopt a simple architecture based on LLaVA [28] for language-vision reading assistant, as illustrated in Figure 6. A low-resolution visual encoder is a significant bottleneck for extracting textual information from images [62]. On the contrary, a high-resolution" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22588" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.094, + 0.306, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.094, + 0.553, + 0.105 + ], + "angle": 0, + "content": "Question: Can you describe the design of the book cover?" + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.109, + 0.847, + 0.147 + ], + "angle": 0, + "content": "GT: The book cover is quite vibrant and engaging. The title, \"Explore Your World: Weird, Wild, Amazing!\" is written in large, white fonts that span across the top and bottom of the cover, making it immediately eye-catching. The author's name, Tim Flannery, is prominently displayed, and the illustrator, Sam Caldwell, is credited on the right side of the cover. The background of the cover is filled with various animals, adding a dynamic and exciting element that would likely appeal to children." + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.151, + 0.846, + 0.186 + ], + "angle": 0, + "content": "Qwen-VL-Chat: The book cover is a black background with various colorful illustrations of animals such as a crocodile, alligator, penguin, and snake in white and green. The title \"Weird Wild, Amazing!\" is written in white letters with a yellow border at the top. The author's name \"Tim Flannery\" is written in white letters above the title, and the publisher's name \"Explore Your World\" is written in green letters below the title. The cover also has a winner sticker for the Environment Children's Literature." + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.193, + 0.846, + 0.213 + ], + "angle": 0, + "content": "mPLUG-Owl2: The book cover features a colorful design with a blue background and a yellow bird. The text on the cover is written in a fun and playful font, emphasizing the title \"Tim Flannery: Explore Your World Weird Wild Amazing.\"" + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.217, + 0.846, + 0.28 + ], + "angle": 0, + "content": "GPT-4V: The book cover features a collage of various whimsical and colorful creatures, reminiscent of a style one might find in a children's book. The title \"Weird, Wild, Amazing!\" is prominently displayed at the center top in large, playful, and colorful lettering, which suggests the book may be about the variety and eccentricity of life on Earth. Below the title, the text \"Explore Your World\" is in a smaller font, inviting the reader to discover the contents. The name \"Tim Flannery\" appears above the title, indicating the author, with the descriptor \"Explore Your World\" just below the title. The art by Sam Caldwell is acknowledged at the bottom right, and there's a badge on the bottom left corner stating \"Winner Environment Award for Children's Literature,\" which suggests the book has been recognized for its contribution to environmental education. The background is dark, which makes the bright colors of the animals and other elements stand out. The animals appear to be drawn in a stylized and engaging manner, with expressions and poses that give them character and life. The overall design is vibrant and seems to be targeting a young audience, promising education through fun and engagement with the natural world." + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.283, + 0.846, + 0.313 + ], + "angle": 0, + "content": "LaRA: The book cover features the title \"Weird, Wild, Amazing!\" in bold letters, with the author's name, Tim Flannery, mentioned at the top. The backdrop of the cover includes illustrations of various animals, suggesting that the book may be about wildlife or exploration." + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.321, + 0.82, + 0.336 + ], + "angle": 0, + "content": "Figure 7. Examples generated by different multimodal language models on the abstract TRINS-VQA benchmarkk." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.473, + 0.593 + ], + "angle": 0, + "content": "visual encoder generates an excessive number of patches, drastically impacting model efficiency. Our conviction is that a low-resolution visual encoder is designed to capture visual information more effectively, including layouts. In contrast, a considerably smaller OCR tool is employed to extract text from high-resolution images. Instruction-tuning on TRINS could serve as an effective method of training LLMs to better align OCR texts and extracted visual features. In addition, LLMs can autonomously rectify errors generated by OCR tools. For the visual encoder \\( V \\), we employ CLIP-ViT-L/14-336 at a resolution of \\( 336 \\times 336 \\). The grid features before the last transformer layer are then mapped into the word embedding space of the language decoder using a trainable projection matrix \\( W \\). Regarding the language decoder \\( D \\), we utilize Vicuna-1.5-13B [8], a language model tuned for instructions through LLaMA 2 [50]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.473, + 0.821 + ], + "angle": 0, + "content": "We follow the two-stage training design of LLaVA but adopt the pre-trained projection layer of LLaVAR [62]. Training targets remain similar: generating output responses \\((<\\text{res}>)\\) for the input instructions \\((<\\text{ins}>)\\), alongside OCR results \\((<\\text{ocr}>)\\). The transformed image tokens \\((<\\text{img}>)\\) are introduced before or after the first input instruction randomly when building the instruction finetuning data. During the finetuning stage, both the projection matrix \\(W\\) and the language decoder \\(D\\) are trained. We consolidate our nearly 90K visual question-answering data with the 158K instruction-following data from LLaVA to form the training set. It should be noted that the visual encoder remains frozen throughout the training period. Compared with previous approaches, LaRA incorporates OCR words as part of the input, a simple way to enhance visual text understanding." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.21, + 0.848 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "In this section, we present three downstream tasks based on the TRINS dataset and outline their evaluation metrics. The proposed method LaRA is used in both the text-rich" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.348, + 0.889, + 0.492 + ], + "angle": 0, + "content": "
MethodRecog.VQASVQADKIEFinal Score
Gemini215174128134651
GPT-4v167163146160636
Monkey1741619188514
mPLUG-Owl21531534119366
LLaVAR1861222513346
LLaVA1.5-13B176129197331
mPLUG-Owl172104183297
MiniGPT-V21242940157
LaRA21114785105548
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.503, + 0.892, + 0.545 + ], + "angle": 0, + "content": "Table 2. Results of LMMs on OCRBench. Recog. represents text recognition, \\(\\mathrm{VQA}^S\\) represents Scene Text-Centric VQA, \\(\\mathrm{VQA}^D\\) represents Document-Oriented VQA." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.559, + 0.894, + 0.62 + ], + "angle": 0, + "content": "image summarization and visual question-answer tasks. All experiments were conducted on NVIDIA A100 80GB GPUs. In fine-tuning, we use a cosine annealing schedule with an initial learning rate of \\(2e^{-5}\\) and a batch size of 32." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.634, + 0.892, + 0.666 + ], + "angle": 0, + "content": "5.1. TRINS-VQA: Text-Rich Image Visual Question Answering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We first performed experiments to evaluate the zero-shot performance of LaRA on classical benchmarks [28, 54]. The results are reported in Table 2 and Table 10. The proposed LaRA model exhibits significant performance improvement across all the datasets compared to other models. Even without OCR, LaRA outperforms other models in most cases, highlighting its robustness and effectiveness in handling visual question-answering tasks. The inclusion of OCR in LaRA further enhances performance, suggesting that Large Language Models (LLMs) can proficiently utilize textual information. However, the ability to directly extract text from images remains limited. LaRA, with its simple structure, significantly boosts model performance on text-rich images, offering an alternative solution to overcome the limitations of pre-trained image encoders. We further evaluate different" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22589" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.403, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.094, + 0.851, + 0.125 + ], + "angle": 0, + "content": "GT: There are big white words \"Urban Sketching\" at the bottom left of the cover. In the upper left corner of the cover, it reads \"Draw What You See.\" In the background of the picture, there is a drawing book and pen. It shows a signature in the bottom right corner of the book." + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.13, + 0.851, + 0.166 + ], + "angle": 0, + "content": "LLaVA: The image is an urban sketching project displayed on a tabletop, featuring a variety of hand-drawn city scenes. Several people can be seen in the drawing, walking around and engaging in different activities. Among the drawings, there is an image of an alleyway with a prominent Umbrella on the right side." + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.171, + 0.851, + 0.201 + ], + "angle": 0, + "content": "LLaVAR: The main subjects in this image are a notebook with a pen placed on top of it, and a photograph of a city street. The notebook and pen appear to be the focal point, with the urban scene in the background." + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.206, + 0.851, + 0.226 + ], + "angle": 0, + "content": "Instruct-BLIP: The main subjects in the image are a series of sketches and drawings depicting urban scenes." + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.233, + 0.851, + 0.255 + ], + "angle": 0, + "content": "LaRA: This is a blog image. Black text in the upper left corner reads \"Draw What You See\". Large white text in the bottom left reads \"Urban Sketching\". The image background is a blue spiral notebook with a pen in the upper right and a picture of a colorful urban scene in the middle." + }, + { + "type": "table_caption", + "bbox": [ + 0.177, + 0.264, + 0.792, + 0.278 + ], + "angle": 0, + "content": "Figure 8. Examples generated by different multimodal language models on the TRINS-Cap benchmarkk." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.285, + 0.87, + 0.462 + ], + "angle": 0, + "content": "
MethodResolutionExtract AccuracyAbstract
B@1B@2B@3B@4METEORROUGECIDEr
Instruct-BLIP [11]\\( {224}^{2} \\)43.913.59.26.75.28.916.723.5
Mini-GPTv215.329.419.413.810.417.228.744.3
mPLUG-Owl2 [57]\\( {448}^{2} \\)61.035.925.519.315.217.738.097.2
Qwen-VL [3]63.640.227.319.915.220.436.479.4
LLaVA [28]23.733.120.914.410.518.031.848.2
LLaVAR [62]\\( {336}^{2} \\)51.738.025.919.114.717.435.684.9
LLaVAR w/ OCR58.140.328.020.916.318.737.397.3
LLaVA 1.5 [27]38.829.820.915.812.620.834.940.7
LLaVAR (finetuned)\\( {336}^{2} \\)61.245.331.923.918.720.439.6104.7
LaRA62.845.134.227.322.621.946.5186.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.191, + 0.465, + 0.776, + 0.48 + ], + "angle": 0, + "content": "Table 3. Results of different models on TRINS-VQA for text-rich image question-answering tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.473, + 0.795 + ], + "angle": 0, + "content": "methods on the TRINS-VQA dataset, as shown in Table 3. For extraction questions, we use the same metric as Wu et al. [54]. For abstract questions, where the answer is typically a longer sentence, we evaluate them based on text similarity metrics such as BLEU [25], ROUGE [25], and CIDEr [52]. In zero-shot inference, LLaVAR with OCR exhibits the best performance, reinforcing the importance of extracting textual information. Furthermore, mPLUG-Owl2 and Qwen-VL perform well and represent the best methods in extract question evaluations, showing that a high-resolution encoder can significantly improve model performance. Instruct-BLIP demonstrates good performance on extract questions, but did not fare as well on abstract questions, given that the answers provided are usually short and concise. Figure 7 shows an example of the responses of different methods on the abstract TRINS-VQA dataset, and more examples can be found in the Appendix D. Qwen-VL includes all details but does not provide high-level insights, such as the ground-truth annotation. Both mPLUG-Owl2 and GPT-4V suffer from hallucination issues." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.803, + 0.431, + 0.82 + ], + "angle": 0, + "content": "5.2. TRINS-Cap: Text-rich image Captioning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.472, + 0.902 + ], + "angle": 0, + "content": "In our experiments on TRINS-Cap, we ask large multimodal models to generate summaries based on text-rich images. The data set was divided into train, validation, and test sets. We compared LaRA with popular baselines, including InstructBLIP [11], Mini-GPT4 [63], LLaVA [28]," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.585 + ], + "angle": 0, + "content": "LLaVAR [62], mPLUG-Owl2 [56] and Qwen-VL [3]. Given that BLIP-2 faced challenges in generating comprehensive and meaningful results for text-rich images, we considered InstructBLIP as an alternative. For all methods, we randomly selected three prompts from ten as instructions for the model (details provided in the Appendix C.1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Table 4 presents the results of different methods in terms of classical captioning metrics. Models with enhanced visual text understanding generally outperform general multimodal models, such as LLaVA, Mini-GPT4, and Instruct-BLIP. LaRA (zero-shot) refers to the LaRA model fine-tuned on the TRINS-QA dataset and demonstrates improved performance. Comparison of fine-tuned LaRA variants indicates that text recognition ability is still limited for OCR-free methods, suggesting that the CLIP encoder or feature projection process may cause visual information loss. Addressing this limitation may involve employing a better trained encoder on text-rich images or designing a more carefully crafted architecture, a direction we leave for future exploration. When fine-tuned with TRINS-Cap, LaRA exhibits much better performance, underscoring the importance of high-quality human-annotated data. Figure 8 shows examples of different models on the TRINS-Cap benchmark. It shows the great capability of LaRA in recognizing text and relating it to its visual contexts, demonstrating the effectiveness of the TRINS dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22590" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.878, + 0.296 + ], + "angle": 0, + "content": "
MethodBackboneB@1B@2B@3B@4METEORROUGECIDEr
mPLUG-Owl2 [57]Llama-2-7B4.93.22.31.78.719.64.5
Instruct-BLIP [11]Vicuna-7B13.48.55.74.19.217.95.7
Qwen-VL [3]Qwen-7B28.818.412.28.614.124.316.6
Instruct-BLIP [11]15.99.96.44.49.618.68.0
Mini-GPT4 [63]31.116.18.65.011.420.86.3
Mini-GPT-v2 [7]Vicuna-13B27.914.78.04.810.820.97.3
LLaVA [28]35.118.29.55.413.222.28.8
LLaVAR [62]18.911.27.25.010.820.111.4
LLaVA 1.5 [27]31.116.49.56.111.921.815.4
LLaVAR w/ OCRVicuna-13B21.413.18.86.412.022.213.0
LaRA (zero-shot)29.319.112.99.214.826.321.3
LLaVAR (fine-tuned)Vicuna-13B36.525.418.013.417.832.435.7
LaRA37.726.418.914.218.433.246.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.27, + 0.3, + 0.698, + 0.314 + ], + "angle": 0, + "content": "Table 4. Results of different models on text-rich image captioning tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.329, + 0.3, + 0.345 + ], + "angle": 0, + "content": "5.3. Additional Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.474, + 0.503 + ], + "angle": 0, + "content": "Performance on general visual tasks after TRINS fine-tuning. We adopted the evaluation protocols of MiniGPT-v2 [7] and compared LaRA with LLaVA [28] on traditional visual question answering benchmarks in table 5. LaRA shows a comparable performance on knowledgeability and better performance on reasoning and spatial awareness. This further verifies the effectiveness of the TRINS dataset, demonstrating that fine-tuning on text-rich images does not degrade performance on natural images, but instead enhances the results." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.508, + 0.443, + 0.567 + ], + "angle": 0, + "content": "
OKVQAGQAVSRVizWiz
LLaVA [28]57.841.351.245.0
LaRA58.142.453.053.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.087, + 0.572, + 0.458, + 0.586 + ], + "angle": 0, + "content": "Table 5. Quantitative Results on the public visual benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.602, + 0.462, + 0.675 + ], + "angle": 0, + "content": "
MetricsControlNetDeepFloydTextDiffuser
FID (↓)51.5949.9651.26
CLIP Score (↑)0.37170.39170.3707
OCR Acc. (↑)0.42410.21920.5027
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.679, + 0.456, + 0.693 + ], + "angle": 0, + "content": "Table 6. Empirical Results on TRINS-Gen (easy) benchmark." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Text-to-document generation Diffusion-based text-to-image generation has shown great success, while precise textual renderings remain a big challenge. TextDiffuser introduced the MARIO-Eval benchmark, drawing from works such as DrawBench [43] and DrawTextCreative [29]. However, most text prompts in MARIO-Eval are short and cannot serve as a good evaluation dataset to handle complex real-world human instructions. We take advantage of human annotations from TRINS-Cap and build the TRINS-Gen benchmark. It is still difficult to render too many words in a single image [6]. In response to this, we filter out images with more than 20 OCR words, resulting in a curated set of 2,104 images. We divide these images into two sets (easy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.33, + 0.895, + 0.422 + ], + "angle": 0, + "content": "and difficult) based on the number of OCR words and the length of the longest OCR string per annotation, where all text prompts in the easy set have less than 9 OCR words. We evaluated existing methods using their public checkpoints and reported the results in Table 6 and detailed results in Table 9 (Appendix C.3)." + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.436, + 0.889, + 0.455 + ], + "angle": 0, + "content": "Text Prompts: The image is a book cover for 'Abandoned San Diego' by Jessica D. Johnson, featuring a background image of an abandoned house." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.456, + 0.588, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.53, + 0.574, + 0.538 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.456, + 0.687, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.53, + 0.665, + 0.539 + ], + "angle": 0, + "content": "ControlNet" + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.456, + 0.788, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.53, + 0.763, + 0.538 + ], + "angle": 0, + "content": "Deepfloyd" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.456, + 0.888, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.818, + 0.53, + 0.865, + 0.538 + ], + "angle": 0, + "content": "TextDiffuser" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.554, + 0.892, + 0.582 + ], + "angle": 0, + "content": "Figure 9. Examples generated by different text-to-image models on the TRINS-Gen benchmarkk." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.603, + 0.628, + 0.619 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Despite the challenges posed by the prevalence of natural images in training data, the significance of visual textual understanding cannot be understated. In this paper, we introduce TRINS, a Text-Rich Image INStruction dataset, comprising a diverse collection of text-rich images, captions, and questions. This dataset, created through a semi-automatic process leveraging large-scale pre-trained models, not only significantly reduces annotation time but also elevates annotation quality. Furthermore, we propose a novel multimodal language model architecture, LaRA, which incorporates OCR as a pivotal enhancement for text-rich image understanding. We anticipate that continued progress in multimodal language model architectures, fine-tuning techniques, and the expansion of diverse, text-rich datasets like TRINS will push the boundaries of visual textual understanding. This, in turn, will facilitate more efficient collaboration between humans and agents, potentially revolutionizing numerous real-world applications." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22591" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.198 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.27 + ], + "angle": 0, + "content": "[2] Anas Awadalla, Irena Gao, Joshua Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortzman, and Ludwig Schmidt. Openflamingo, 2023. 22" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.272, + 0.472, + 0.342 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.472, + 0.412 + ], + "angle": 0, + "content": "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.472, + 0.458 + ], + "angle": 0, + "content": "[5] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.472, + 0.502 + ], + "angle": 0, + "content": "[6] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. arXiv preprint arXiv:2305.10855, 2023. 1, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.472, + 0.586 + ], + "angle": 0, + "content": "[7] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.472, + 0.658 + ], + "angle": 0, + "content": "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.661, + 0.472, + 0.799 + ], + "angle": 0, + "content": "[9] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling instruction-finetuned language models, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.802, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.859, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructblip: Towards" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.895, + 0.12 + ], + "angle": 0, + "content": "general-purpose vision-language models with instruction tuning. ArXiv, abs/2305.06500, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.895, + 0.204 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.895, + 0.288 + ], + "angle": 0, + "content": "[13] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marcal Rusinol, Minesh Mathew, C.V. Jawahar, Ernest Valveny, and Dimosthenis Karatzas. Icdar 2019 competition on scene text visual question answering. 2019 International Conference on Document Analysis and Recognition (ICDAR), 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.895, + 0.345 + ], + "angle": 0, + "content": "[14] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient visual instruction model, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.895, + 0.388 + ], + "angle": 0, + "content": "[15] Adam W. Harley, Alex Ufkes, and Konstantinos G. Derpanis. Evaluation of deep convolutional nets for document image classification and retrieval, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.895, + 0.472 + ], + "angle": 0, + "content": "[16] Rongjie Huang, Mingze Li, Dongchao Yang, Jiatong Shi, Xuankai Chang, Zhenhui Ye, Yuning Wu, Zhiqing Hong, Jia-Bin Huang, Jinglin Liu, Yixiang Ren, Zhou Zhao, and Shinji Watanabe. Audiogpt: Understanding and generating speech, music, sound, and talking head. ArXiv, abs/2304.12995, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.475, + 0.895, + 0.531 + ], + "angle": 0, + "content": "[17] Xinyu Huang, Youcai Zhang, Jinyu Ma, Weiwei Tian, Rui Feng, Yuejie Zhang, Yaqian Li, Yandong Guo, and Lei Zhang. Tag2text: Guiding vision-language model via image tagging. arXiv preprint arXiv:2303.05657, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.895, + 0.587 + ], + "angle": 0, + "content": "[18] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[19] Chunyuan Li. Large multimodal models: Notes on cvpr 2023 tutorial. ArXiv, abs/2306.14895, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.895, + 0.673 + ], + "angle": 0, + "content": "[20] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 1, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.895, + 0.731 + ], + "angle": 0, + "content": "[21] Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, and Furu Wei. Dit: Self-supervised pre-training for document image transformer. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.895, + 0.773 + ], + "angle": 0, + "content": "[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023. 5, 22" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.895, + 0.829 + ], + "angle": 0, + "content": "[23] Xian Li, Ping Yu, Chunting Zhou, Timo Schick, Luke Zettlemoyer, Omer Levy, Jason Weston, and Mike Lewis. Self-alignment with instruction backtranslation. arXiv preprint arXiv:2308.06259, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.895, + 0.899 + ], + "angle": 0, + "content": "[24] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. arXiv preprint arXiv:2311.06607, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22592" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[25] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.472, + 0.192 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.194, + 0.471, + 0.236 + ], + "angle": 0, + "content": "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.239, + 0.471, + 0.266 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 2, 5, 6, 7, 8, 22" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.268, + 0.471, + 0.337 + ], + "angle": 0, + "content": "[29] Rosanne Liu, Dan Garrette, Chitwan Sahara, William Chan, Adam Roberts, Sharan Narang, Irina Blok, RJ Mical, Mohammad Norouzi, and Noah Constant. Character-aware models improve visual text rendering. arXiv preprint arXiv:2212.10562, 2022.8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.34, + 0.47, + 0.396 + ], + "angle": 0, + "content": "[30] Yuliang Liu, Zhang Li, Hongliang Li, Wenwen Yu, Mingxin Huang, Dezhi Peng, Mingyu Liu, Mingrui Chen, Chunyuan Li, Cheng lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models, 2023. 1, 2, 22" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.398, + 0.471, + 0.452 + ], + "angle": 0, + "content": "[31] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Learning to draw Chinese characters in image synthesis models coherently, 2023. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.455, + 0.471, + 0.497 + ], + "angle": 0, + "content": "[32] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.499, + 0.471, + 0.554 + ], + "angle": 0, + "content": "[33] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.557, + 0.471, + 0.585 + ], + "angle": 0, + "content": "[34] Minesh Mathew, Dimosthenis Karatzas, and C. V. Jawahar. Docvqa: A dataset for vqa on document images, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.587, + 0.471, + 0.642 + ], + "angle": 0, + "content": "[35] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.645, + 0.471, + 0.713 + ], + "angle": 0, + "content": "[36] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947–952. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.716, + 0.369, + 0.73 + ], + "angle": 0, + "content": "[37] OpenAI. Gpt-4 technical report, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.733, + 0.471, + 0.829 + ], + "angle": 0, + "content": "[38] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.832, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.175 + ], + "angle": 0, + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.894, + 0.246 + ], + "angle": 0, + "content": "[41] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.893, + 0.29 + ], + "angle": 0, + "content": "[42] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. arXiv preprint arXiv:1809.02156, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.894, + 0.375 + ], + "angle": 0, + "content": "[43] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.894, + 0.459 + ], + "angle": 0, + "content": "[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.894, + 0.516 + ], + "angle": 0, + "content": "[45] Alex Shonenkov, Misha Konstantinov, Daria Bakshandaeva, Christoph Schuhmann, Ksenia Ivanova, and Nadija Klokova. Deepfloyd. https://github.com/deep-floyd/if, 2023.12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.894, + 0.601 + ], + "angle": 0, + "content": "[46] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 742-758. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.894, + 0.672 + ], + "angle": 0, + "content": "[47] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.894, + 0.745 + ], + "angle": 0, + "content": "[48] Amanpreet Singh, Guan Pang, Mandy Toh, Jing Huang, Wojciech Galuba, and Tal Hassner. Textocr: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8802-8812, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.894, + 0.815 + ], + "angle": 0, + "content": "[49] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[50] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22593" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[51] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[52] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.472, + 0.275 + ], + "angle": 0, + "content": "[53] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[54] Zizhang Wu, Xinyuan Chen, Jizheng Wang, Xiaoquan Wang, Yanzhu Gan, Muqing Fang, and Tianhao Xu. Ocr-rtps: anOCR-based real-time positioning system for the valet parking. Applied Intelligence, pages 1-15, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.472, + 0.403 + ], + "angle": 0, + "content": "[55] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.472, + 0.487 + ], + "angle": 0, + "content": "[56] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 2, 7, 22" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.488, + 0.472, + 0.556 + ], + "angle": 0, + "content": "[57] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. arXiv preprint arXiv:2311.04257, 2023. 7, 8, 22" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.558, + 0.472, + 0.613 + ], + "angle": 0, + "content": "[58] Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.615, + 0.472, + 0.657 + ], + "angle": 0, + "content": "[59] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.472, + 0.713 + ], + "angle": 0, + "content": "[60] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.472, + 0.77 + ], + "angle": 0, + "content": "[61] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.771, + 0.472, + 0.827 + ], + "angle": 0, + "content": "[62] Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 1, 2, 5, 6, 7, 8, 22" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.828, + 0.472, + 0.882 + ], + "angle": 0, + "content": "[63] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 1, 2, 7, 8, 22" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.882 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22594" + } + ] +] \ No newline at end of file diff --git a/2024/TRINS_ Towards Multimodal Language Models that Can Read/527c7573-3fb6-42ad-a14f-b7b957e97243_origin.pdf b/2024/TRINS_ Towards Multimodal Language Models that Can Read/527c7573-3fb6-42ad-a14f-b7b957e97243_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..02a5ef201a4d8aef21b70ab2cada30fa763b0683 --- /dev/null +++ b/2024/TRINS_ Towards Multimodal Language Models that Can Read/527c7573-3fb6-42ad-a14f-b7b957e97243_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77ef3f898165dc0c9b474c5cbf1a9944939e2f70462ed1856840183c2a432598 +size 10040346 diff --git a/2024/TRINS_ Towards Multimodal Language Models that Can Read/full.md b/2024/TRINS_ Towards Multimodal Language Models that Can Read/full.md new file mode 100644 index 0000000000000000000000000000000000000000..87ee94dcc173dc2528c374c2b07cc88059198904 --- /dev/null +++ b/2024/TRINS_ Towards Multimodal Language Models that Can Read/full.md @@ -0,0 +1,287 @@ +# TRINS: Towards Multimodal Language Models that Can Read + +Ruiyi Zhang $^{1}$ , Yanzhe Zhang $^{2}$ , Jian Chen $^{3}$ , Yufan Zhou $^{1}$ , Jiumiang Gu $^{1}$ , Changyou Chen $^{3}$ , Tong Sun $^{1}$ + +$^{1}$ Adobe Research $^{2}$ Georgia Institute of Technology $^{3}$ State University of New York at Buffalo + +{ruizhang}@adobe.com + +# Abstract + +Large multimodal language models have shown remarkable proficiency in understanding and editing images. However, a majority of these visually-tuned models struggle to comprehend the textual content embedded in images, primarily due to the limitation of training data. In this work, we introduce TRINS: a Text-Rich image1 INstruction dataset, with the objective of enhancing the reading ability of the multimodal large language model. TRINS is built upon LAION2 using hybrid data annotation strategies that include machine-assisted and human-assisted annotation process. It contains 39,153 text-rich images, captions, and 102,437 questions. Specifically, we show that the number of words per annotation in TRINS is significantly longer than that of related datasets, providing new challenges. Furthermore, we introduce a simple and effective architecture, called a Language-Vision Reading Assistant (LaRA), which is good at understanding textual content within images. LaRA outperforms existing state-of-the-art multimodal large language models on the TRINS dataset as well as other classical benchmarks. Lastly, we conducted a comprehensive evaluation with TRINS on various text-rich image understanding and generation tasks, demonstrating its effectiveness. + +# 1. Introduction + +Instruction tuning [9, 38] has shown a great generalization ability on unseen tasks and has contributed to the growing popularity of large language models (LLMs), such as ChatGPT [37]. Recently, multimodal language models benefit from visual instruction finetuning [1, 18, 19, 28, 63], and has shown great success in real-world applications. These models leverage visual encoders such as CLIP-ViT [12, 39] to empower LLMs with image comprehension ability. However, challenges arise in comprehension of textual information within images, which may stem from the prevalence of natu + +ral images in training datasets, such as Conceptual Captions [5] and COCO [26]), as highlighted by Liu et al. [30]. Recognizing the importance of visual textual understanding for effective collaboration between agents and humans, Zhang et al. [62] proposed enhancing end-to-end visual instruction-tuned models by introducing noisy Optical Character Recognition (OCR) annotations to improve vision-language alignment. In this work, we surpass existing achievements and collect a new Text-Rich image INStruction dataset named TRINS, which contains 39,153 text-rich images, captions and 102,437 questions. + +TRINS is created in a semi-automatic manner for a more controllable and faithful collection. Specifically, we exploited large-scale pre-trained models such as CLIP [40] and GPT-4 in the annotation process. This semi-automatic process significantly reduces the time and resources required for manual annotation and surprisingly improves the overall quality of annotations. TRINS dataset is composed of three datasets for captioning, visual question answering (VQA) and image generation, respectively. Specifically, human-annotated captions for text-rich images are first collected because they can best translate text-rich images into texts. During this process, extracted OCR words and recognize-anything model tags are provided to the annotators for better and efficient annotations. With detailed image descriptions, VAQ data is built and fulfilled by large language models, such as GPT-4 [37] and LLaMA-2 [51]. In detailed statistics and analysis, we found that both annotated captions and collected question-answer pairs are more comprehensive and contain significantly more details than the existing dataset. Therefore, we show the superior advantage of TRINS compared to existing instruction fine-tuning datasets. As a by-product, high-quality image-caption pairs can serve as a good benchmark for text-rich image generation, which is still a very challenging task [6]. At the same time, we propose a new simple and effective multimodal language model architecture that includes OCR as a component. We call it Language-vision Reading Assistant (LaRA) and show that LaRA fine-tuned on TRINS brings the best text-rich image understanding ability. Our contributions are as follows: + +- We introduce a novel dataset (TRINS) containing 39,153 captions and 102,437 high-quality text-rich image instruction pairs. TRINS is annotated with a novel semi-automatic annotation framework that is scalable and reliable. +- We develop several evaluation benchmarks for text-rich image understanding and generation tasks. Various methods are evaluated on TRINS, demonstrating the effectiveness of the dataset. +- The TRINS datasets are high-quality and comprehensive, which is reflected not only in the dataset statistics but also from the results of multiple baseline models. Especially, our proposed LaRA finetuned on TRINS outperforms existing state-of-the-art methods on text-rich image understanding tasks. + +# 2. Related Work + +Multimodal Instruction Tuning Multi-modal instruction tuning, including image, video [32, 59], and audio [16, 58] settings, has been an active research topic. MiniGPT-4 [63] uses ChatGPT to generate high-quality instruction-following data, while LLaVA [28] generates such data by prompting GPT-4 with captions and bounding boxes. LLaMA-Adapter [14, 60] aligns text-image features using COCO data, and mPLUG-owl [56] combines extensive image-text pairs for pretraining and a mixture of data for finetuning. Despite this, many models, according to Liu et al. [30], struggle with OCR tasks. InstructBLIP [10] addresses this by transforming 13 vision language tasks into an instruction-following format. mPLUGOwl [55, 56] apply multitask instruction funetuing using existing document datasets. A comprehensive survey is available in Li et al. [20]. LLaVAR [62] exploits GPT-4 to collect fine-tuning data without human annotations using OCR and captioning tools. It discovered that resolution plays a significant role in recognizing textual information and explored several options. Monkey [24] performed a surgery between simple text labels and high input resolution, enabling remarkable performance in visually-rich document images with dense text. TRINS exploits human-machine collaboration for data annotation and can provide more accurate information, reducing the problem of hallucination. + +Text-Rich Image Datasets Visual question answering or captioning datasets are widely used in task-specific finetuning and large multimodal model evaluation. TextCap [46] is the first text-rich image captioning dataset. Compared to TextCap, TRINS-Cap provides more detailed annotations that can fulfill the requirement of instruction finetuning. TextOCR [48] aims comprehend text in the context of an image, which is similar to our motivation but focuses more on text recognition in images instead of understanding. STVQA [13] uses spatial and textual information to answer visually grounded questions, effectively integrating visual + +
DatasetYearSizeAnnotation Type
OCR-VQA2019200KQA (▲)
TextVQA201945KQA (▲)
TextCap2020140kCaption (▲)
TextOCR2021145kText Bbox (▲)
DocVQA202050kQA (▲)
40kCaption (▲)
TRINS (Ours)2023100kQA (▲ +▲)
40kText Bbox (▲)
+ +Table 1. Comparison between TRINS and other related datasets. + +and textual cues. OCR-VQA [36] focuses on incorporating optical character recognition (OCR) into visual question answering (VQA), which operates primarily on text within images. TextVQA [47] also takes advantage of the textual information present in the images to answer questions, but with an emphasis on open questions. DocVQA takes this one step further by applying VQA to document images, handling a variety of layouts and formats. InfoVQA [35] and ChartQA [33] focus on specific subdomains and aim to answer questions about information graphics and chart images, respectively. In summary, these related works provide datasets for leveraging spatial and textual cues. TRINS-VQA is a dataset that exploits the semi-automatic annotation process. It can be used for general domain instruction finetuning and model evaluations. + +# 3. Text-Rich Image Instruction Dataset + +To equip multimodal language models with the ability to recognize text and relate it to its visual context, we have curated a new dataset named Text-Rich Instruction (TRINS). The ultimate goal is to enable these models to have spatial, semantic, and visual reasoning between multiple text tokens and visual entities. In this section, we present TRINS, a dataset crafted through a semi-automatic process. Specifically, we leverage large-scale pre-trained models like CLIP [40] and GPT-4 in the annotation process, offering potential advantages: (i) Significant reduction in annotation time and resources: using these models significantly reduces the time and resources required for manual annotation. (ii) Enhancement of annotation data quality through post-processing: the involvement of large models contributes to improving the overall quality of annotation data through subsequent post-processing. (iii) Functionality of large models as knowledge bases: large models can serve as effective knowledge bases, aiding in the annotation process by virtue of their extensive training in diverse datasets. + +We first outline the document image collection process for TRINS, utilizing CLIP models. Then, we present data statistics to facilitate a comprehensive understanding. We delve into three distinct tasks derived from the TRINS dataset in detail: i) TRINS-Cap: Visual Captioning, ii) TRINS-VQA: + +![](images/2e07ccf16200e8bd467810bedd6b190e2e9d962a7d624f7fa2783da4a2892e6e.jpg) + +![](images/9ff8f48869fbb1da497f7a7c8a13946925c11b54325674a532753e38d772c26c.jpg) +Figure 1. Overview of TRINS data collection process, which consists of three datasets. Text-rich images are first selected from web images and then ask annotators to describe the image in details. i) TRINS-Cap is extracted from human annotations with heuristic data processing for text-rich image summarization tasks. ii) TRINS-VQA is built upon human annotations and generate question-answer pairs for training by prompting text-only large language models. iii) TRINS-Gen combined human annotations and text boxes for text-rich image generation. + +Visual Question Answering and iii) TRINS-Gen: Text-to-Image Generation. The overview of the TRINS data collection process is illustrated in Figure 1. To provide a succinct overview, TRINS-Cap undergoes full annotation by human annotators, while TRINS-Gen and TRINS-VQA are constructed based on TRINS-Cap with the help of pre-trained models. A summary of the various TRINS datasets is presented in Table 7 of Appendix A. + +# 3.1. Machine-Assisted Text-rich Image Selections + +Beginning with the LAION-5B dataset [44], our objective is to selectively retain images that exhibit a significant presence of text. Recognizing that document images typically contain substantial textual content, we initially formed a binary classification dataset by combining natural images with document data. Subsequently, we trained an image classifier using a DiT [21] base backbone, fine-tuned on the RVL-CDIP dataset [15]. The purpose of this classifier was to predict whether an image contains text. Then a subset was constructed by selecting images with a predicted probability greater than 0.8, while also adhering to the criteria $p(\text{watermark}) < 0.8$ and $p(\text{unsafe}) < 0.5$ , where both probabilities are derived from the metadata of the LAION dataset. Acknowledging the noise introduced due to the classifier's limitations, we further refined the dataset by incorporating human judgment. A random sample of 20,000 images + +![](images/63d498ffbbc00fa4d8ed66ff2ea00ceb34d9bb86fb247519e51be4241bbd4848.jpg) +Figure 2. CLIP-based categorization of our collected images and selected representative data samples from each category. +(a) + +![](images/2d4b79776fd32e46a986dacbe18dd014d9d88e6071323881fd99ef94c145ec2c.jpg) +(b) +Figure 3. Word clouds of (a) predicted tags and (b) detected words from the text-rich images of TRINS. + +from the filtered LAION-5B was clustered into 50 groups based on CLIP-ViT-B/32 visual features. After inspecting the clustering results, one cluster was meticulously chosen, encompassing diverse text-rich images such as posters, covers, advertisements, and educational documents. This cluster model then served as the filtering mechanism for collecting images that comprise the TRINS dataset. For reference, we present a CLIP-based categorization [39] in Figure 2 to depict the distribution of images in the collected data. The major class is book cover images, further categorized on the basis of book themes and contents. To enhance our understanding of text-rich images, we employed a Recognize Anything Model (RAM) [17, 61] to extract tags from TRINS images. Figure 3a displays word clouds of RAM tags, where "book" and "poster" emerge as major keywords. Additionally, we utilize the Azure Read API $^{4}$ and PaddleOCR to extract + +![](images/e04720c003cdc236deadbd4276ac8aeb06b8f88947330391a6d3af5f56fcafb2.jpg) +(a) +Figure 4. OCR word (a), Caption (b), Question (c) and Answer (d) statistics for TRINS. + +![](images/1b45b02cdc6115d23f128d22e7825dd04090b80d81de7594d3a234ec0cff66de.jpg) +(b) + +![](images/3749cb3f0b249d37623a18c5cb51b1d1e78b2c644405fdbeb87b481be6831862.jpg) +(c) + +![](images/88282c48c156dd73e6fcc27346188ec1b9a5b044249e6cb45219a32213fc3b80.jpg) +(d) + +text within TRINS images. The word cloud of the extracted texts is presented in Figure 3b. Figure 4a illustrates the distribution of OCR words per image, indicating that most of the images in TextCap have fewer than 10 words, while the TRINS images average 31.4 OCR words. Recognizing texts within TRINS images is more challenging because of the presence of numerous small words. In summary, TRINS images encompass rich visual content, seamlessly integrating textual information into the image context. + +![](images/b68098fd62743113499630aa43e14fa47bcc3c2eeb206808982629c26467f68c.jpg) +Figure 5. Question type statistics based on key words. + +# 3.2. Annotation Details + +Annotator Selections All annotators (with the tag $100\%$ Job Success and Top Rated Plus) are native English speakers and have experience with document annotations. We first asked all annotators to annotate a 200-image set and provide them with detailed annotation guidelines with multiple examples. In addition, we use Labelbox as an annotation tool and set quality control questions. + +Heuristic Filters We first use EasyOCR to extract texts from images and retain text phrases with more than three characters, a height greater than $5\%$ of the canvas height, and a confidence score greater than 0.1. For each phrase retrieved, we employ an edit distance-based string matching algorithm (due to potentially erroneous OCR results) to search for its optimal matching substrings within the human-generated caption. The average score for all extracted phrases serves as a metric. + +Manual Reviews We accept annotations with high metric scores and reject the lowest for rework. We manually review other annotations. + +Sensitive Images We combined neural models with human efforts to filter the images. The first step involves an initial data filtering by 2-3 individuals to filter out sensitive images for training. The second step involves hiring additional people to perform a further screening on the data. We engaged annotators from various countries to check the selected images. + +# 3.3. TRINS-Cap: Text-Rich Image Captioning + +Annotation Process TRINS-Cap is a dataset fully annotated by human annotators. We hired 20 native English speakers with experience in document annotation through Upwork. The annotation process, conducted in LabelBox, involved a total of 2,079 hours to annotate 40,576 text-rich images, with an additional 159 hours allocated for result review. After filtering low-quality annotations and addressing missing images, we obtained a final set of 39,153 image-annotation pairs. The dataset is partitioned into train, validation, and test splits with sizes of 29,153, 5,000, and 5,000, respectively. All annotations undergo an initial automated review that involves matching the OCR words with the annotations. Subsequently, human evaluators conduct a thorough review, rejecting annotations with errors, and prompting annotators to rework them for enhancement. We provide comprehensive annotation instructions to all annotators to ensure that each annotation includes: (i) detailed descriptions of visual components. (ii) describe texts' location, attributes, and put texts into annotations. (iii) optional insights or abstract descriptions. + +Statistics and Analysis The primary objective of the annotation process is to facilitate a human or machine's full comprehension of the information conveyed in the image without direct viewing. Consequently, the average annotation length for TRINS is 65.1 words, significantly exceeding that of COCO (10.6 words) and TextCaps (12.4 words). Figure 4b shows the caption length distributions for TRINS-Cap, COCO, and TextCap, demonstrating the comprehensive nature of the TRINS-Cap annotations. TRINS with more contexts can generally provide a better description of complex images, where short captions are insufficient. Hence, LLMs finetund on TRINS can better understand images with complex texts and layouts, which has been further verified in Section 5. + +# 3.4. TRINS-VQA: Multimodal Question Answering + +The annotation process for question answering is inherently complex, primarily due to the necessity for annotators to generate high-quality questions. Creating an effective question is more challenging than providing an answer. As a result, annotators frequently gravitate toward formulating concrete and extractive questions (e.g., "Who is the author of this book?") rather than abstract ones (e.g., "How does the design of the book cover reflect the content of the book?"). We introduce semi-automatic annotation methods to generate high-quality visual question-answering data for TRINS-VQA. This dataset is designed to train general vision language assistants through instruction fine-tuning, and its benefits on model performance are evaluated in Section 5.1. TRINS-Cap, on the other hand, serves as human-assisted annotations, offering a comprehensive but non-instructive dataset for fine-tuning. To utilize the wealth of high-quality annotations available, we incorporated semi-automatic annotation by using large language models (LLMs) such as OpenAI's GPT-4 [37] and Llama-70B [50, 51] to enhance our data annotation pipeline. OCR results and detailed descriptions of each image are provided to LLMs. Furthermore, high-quality human-crafted demonstrations and detailed annotation rules are provided to LLMs. One demonstration focused on extract questions, while the other emphasized abstract questions, creating a more balanced dataset. + +Human Annotations To facilitate a robust evaluation of model performance, we hired 10 Upwork annotators, whose native language is English, to annotate the test dataset, following a methodology similar to previous work [13, 34, 47]. The test dataset comprises 5,000 images with 18,764 question-answer pairs. These data collected are used exclusively for evaluation purposes. + +Statistics and Analysis Building upon prior research [28, 49, 53], we provide visualizations of instructions in Figure 5 based on question keywords. The inner cycle illustrates the distribution of the first word in the questions, while the outer cycle presents extracted keywords determined by carefully designed heuristics. Types of questions are categorized according to keywords found in questions. + +In Figure 4c and 4d, we present statistics on the number of words per question-answer pair, comparing them with previous work. Generally, the average length of questions for TRINS-VQA is 10.5, surpassing that of DocVQA (8.3), OCR-VQA (6.5), and TextVQA (7.1). Surprisingly, the average answer length for TRINS is 23.9, significantly longer than related datasets (all less than 4). This discrepancy arises from TRINS containing more abstract questions that typically have longer answers. Similarly, the dataset is divided into train, validation, and test splits. For extract questions, the accuracy of the answers is calculated similarly to Liu et al. [28], while for abstract questions, generation metrics + +![](images/38df59e7c1c89b20c52f2a641f14a48022598067c0247d375ea5fd9356ab05e8.jpg) +Figure 6. Model overview of the LaRA. the CLIP model processes the input image to generate patch-wise features. These features then serve as input to a projection layer, yielding visual tokens. Concurrently, an OCR tool extracts textual data from the image, which is then merged with the user instruction. + +such as BLEU scores are used to assess the quality of the answers. + +The question-answering and instruction data we obtain are extensive, encompassing a balanced mix of extract and abstract questions. This includes detailed descriptions, summaries, question-answer pairs, tasks that promote creativity and the generation of novel thoughts, and conversational tasks. The dataset spans a diverse range of concepts, ranging from visual presentation and visual language relations to intricate reasoning tasks. Compared to previous methods [28], captions generated by models such as BLIP-2 [22] exhibit less informativeness than detailed human annotations, as demonstrated in Section 5.2. Additionally, captioning models may be susceptible to hallucinations [42], a concern mitigated in TRINS, which provides more comprehensive and reliable visual descriptions. Although OCR tools are robust, they can introduce noisy information. To address this, we utilize both the Azure Read API and PaddleOCR to extract text information. We added the potential unreliability in our system prompts to LLMs, instructing them to generate questions with assured answers. Ultimately, we directly leverage the responses from large language models (LLMs) to construct TRINS-VQA. The quality of the instruction data can be further enhanced through self-alignment [23] or by seeking verification from human annotators. Although this has not been explored in this work, we leave it as a potential direction for future research. + +# 4. The Language-vision Reading Assistant (LaRA) Model + +To verify our data, we adopt a simple architecture based on LLaVA [28] for language-vision reading assistant, as illustrated in Figure 6. A low-resolution visual encoder is a significant bottleneck for extracting textual information from images [62]. On the contrary, a high-resolution + +![](images/5d4118388ee25eab4de75386866960a5381643c4ebfd47d717cb4f4a740214f0.jpg) +Figure 7. Examples generated by different multimodal language models on the abstract TRINS-VQA benchmarkk. + +Question: Can you describe the design of the book cover? + +GT: The book cover is quite vibrant and engaging. The title, "Explore Your World: Weird, Wild, Amazing!" is written in large, white fonts that span across the top and bottom of the cover, making it immediately eye-catching. The author's name, Tim Flannery, is prominently displayed, and the illustrator, Sam Caldwell, is credited on the right side of the cover. The background of the cover is filled with various animals, adding a dynamic and exciting element that would likely appeal to children. + +Qwen-VL-Chat: The book cover is a black background with various colorful illustrations of animals such as a crocodile, alligator, penguin, and snake in white and green. The title "Weird Wild, Amazing!" is written in white letters with a yellow border at the top. The author's name "Tim Flannery" is written in white letters above the title, and the publisher's name "Explore Your World" is written in green letters below the title. The cover also has a winner sticker for the Environment Children's Literature. + +mPLUG-Owl2: The book cover features a colorful design with a blue background and a yellow bird. The text on the cover is written in a fun and playful font, emphasizing the title "Tim Flannery: Explore Your World Weird Wild Amazing." + +GPT-4V: The book cover features a collage of various whimsical and colorful creatures, reminiscent of a style one might find in a children's book. The title "Weird, Wild, Amazing!" is prominently displayed at the center top in large, playful, and colorful lettering, which suggests the book may be about the variety and eccentricity of life on Earth. Below the title, the text "Explore Your World" is in a smaller font, inviting the reader to discover the contents. The name "Tim Flannery" appears above the title, indicating the author, with the descriptor "Explore Your World" just below the title. The art by Sam Caldwell is acknowledged at the bottom right, and there's a badge on the bottom left corner stating "Winner Environment Award for Children's Literature," which suggests the book has been recognized for its contribution to environmental education. The background is dark, which makes the bright colors of the animals and other elements stand out. The animals appear to be drawn in a stylized and engaging manner, with expressions and poses that give them character and life. The overall design is vibrant and seems to be targeting a young audience, promising education through fun and engagement with the natural world. + +LaRA: The book cover features the title "Weird, Wild, Amazing!" in bold letters, with the author's name, Tim Flannery, mentioned at the top. The backdrop of the cover includes illustrations of various animals, suggesting that the book may be about wildlife or exploration. + +visual encoder generates an excessive number of patches, drastically impacting model efficiency. Our conviction is that a low-resolution visual encoder is designed to capture visual information more effectively, including layouts. In contrast, a considerably smaller OCR tool is employed to extract text from high-resolution images. Instruction-tuning on TRINS could serve as an effective method of training LLMs to better align OCR texts and extracted visual features. In addition, LLMs can autonomously rectify errors generated by OCR tools. For the visual encoder $V$ , we employ CLIP-ViT-L/14-336 at a resolution of $336 \times 336$ . The grid features before the last transformer layer are then mapped into the word embedding space of the language decoder using a trainable projection matrix $W$ . Regarding the language decoder $D$ , we utilize Vicuna-1.5-13B [8], a language model tuned for instructions through LLaMA 2 [50]. + +We follow the two-stage training design of LLaVA but adopt the pre-trained projection layer of LLaVAR [62]. Training targets remain similar: generating output responses $(<\text{res}>)$ for the input instructions $(<\text{ins}>)$ , alongside OCR results $(<\text{ocr}>)$ . The transformed image tokens $(<\text{img}>)$ are introduced before or after the first input instruction randomly when building the instruction finetuning data. During the finetuning stage, both the projection matrix $W$ and the language decoder $D$ are trained. We consolidate our nearly 90K visual question-answering data with the 158K instruction-following data from LLaVA to form the training set. It should be noted that the visual encoder remains frozen throughout the training period. Compared with previous approaches, LaRA incorporates OCR words as part of the input, a simple way to enhance visual text understanding. + +# 5. Experiments + +In this section, we present three downstream tasks based on the TRINS dataset and outline their evaluation metrics. The proposed method LaRA is used in both the text-rich + +
MethodRecog.VQASVQADKIEFinal Score
Gemini215174128134651
GPT-4v167163146160636
Monkey1741619188514
mPLUG-Owl21531534119366
LLaVAR1861222513346
LLaVA1.5-13B176129197331
mPLUG-Owl172104183297
MiniGPT-V21242940157
LaRA21114785105548
+ +Table 2. Results of LMMs on OCRBench. Recog. represents text recognition, $\mathrm{VQA}^S$ represents Scene Text-Centric VQA, $\mathrm{VQA}^D$ represents Document-Oriented VQA. + +image summarization and visual question-answer tasks. All experiments were conducted on NVIDIA A100 80GB GPUs. In fine-tuning, we use a cosine annealing schedule with an initial learning rate of $2e^{-5}$ and a batch size of 32. + +# 5.1. TRINS-VQA: Text-Rich Image Visual Question Answering + +We first performed experiments to evaluate the zero-shot performance of LaRA on classical benchmarks [28, 54]. The results are reported in Table 2 and Table 10. The proposed LaRA model exhibits significant performance improvement across all the datasets compared to other models. Even without OCR, LaRA outperforms other models in most cases, highlighting its robustness and effectiveness in handling visual question-answering tasks. The inclusion of OCR in LaRA further enhances performance, suggesting that Large Language Models (LLMs) can proficiently utilize textual information. However, the ability to directly extract text from images remains limited. LaRA, with its simple structure, significantly boosts model performance on text-rich images, offering an alternative solution to overcome the limitations of pre-trained image encoders. We further evaluate different + +![](images/919b4673629bbb617b4650b8d3fd7faae0359c6957ccbc901f358e59573eba09.jpg) + +GT: There are big white words "Urban Sketching" at the bottom left of the cover. In the upper left corner of the cover, it reads "Draw What You See." In the background of the picture, there is a drawing book and pen. It shows a signature in the bottom right corner of the book. + +LLaVA: The image is an urban sketching project displayed on a tabletop, featuring a variety of hand-drawn city scenes. Several people can be seen in the drawing, walking around and engaging in different activities. Among the drawings, there is an image of an alleyway with a prominent Umbrella on the right side. + +LLaVAR: The main subjects in this image are a notebook with a pen placed on top of it, and a photograph of a city street. The notebook and pen appear to be the focal point, with the urban scene in the background. + +Instruct-BLIP: The main subjects in the image are a series of sketches and drawings depicting urban scenes. + +LaRA: This is a blog image. Black text in the upper left corner reads "Draw What You See". Large white text in the bottom left reads "Urban Sketching". The image background is a blue spiral notebook with a pen in the upper right and a picture of a colorful urban scene in the middle. + +Figure 8. Examples generated by different multimodal language models on the TRINS-Cap benchmarkk. + +
MethodResolutionExtract AccuracyAbstract
B@1B@2B@3B@4METEORROUGECIDEr
Instruct-BLIP [11]\( {224}^{2} \)43.913.59.26.75.28.916.723.5
Mini-GPTv215.329.419.413.810.417.228.744.3
mPLUG-Owl2 [57]\( {448}^{2} \)61.035.925.519.315.217.738.097.2
Qwen-VL [3]63.640.227.319.915.220.436.479.4
LLaVA [28]23.733.120.914.410.518.031.848.2
LLaVAR [62]\( {336}^{2} \)51.738.025.919.114.717.435.684.9
LLaVAR w/ OCR58.140.328.020.916.318.737.397.3
LLaVA 1.5 [27]38.829.820.915.812.620.834.940.7
LLaVAR (finetuned)\( {336}^{2} \)61.245.331.923.918.720.439.6104.7
LaRA62.845.134.227.322.621.946.5186.6
+ +Table 3. Results of different models on TRINS-VQA for text-rich image question-answering tasks. + +methods on the TRINS-VQA dataset, as shown in Table 3. For extraction questions, we use the same metric as Wu et al. [54]. For abstract questions, where the answer is typically a longer sentence, we evaluate them based on text similarity metrics such as BLEU [25], ROUGE [25], and CIDEr [52]. In zero-shot inference, LLaVAR with OCR exhibits the best performance, reinforcing the importance of extracting textual information. Furthermore, mPLUG-Owl2 and Qwen-VL perform well and represent the best methods in extract question evaluations, showing that a high-resolution encoder can significantly improve model performance. Instruct-BLIP demonstrates good performance on extract questions, but did not fare as well on abstract questions, given that the answers provided are usually short and concise. Figure 7 shows an example of the responses of different methods on the abstract TRINS-VQA dataset, and more examples can be found in the Appendix D. Qwen-VL includes all details but does not provide high-level insights, such as the ground-truth annotation. Both mPLUG-Owl2 and GPT-4V suffer from hallucination issues. + +# 5.2. TRINS-Cap: Text-rich image Captioning + +In our experiments on TRINS-Cap, we ask large multimodal models to generate summaries based on text-rich images. The data set was divided into train, validation, and test sets. We compared LaRA with popular baselines, including InstructBLIP [11], Mini-GPT4 [63], LLaVA [28], + +LLaVAR [62], mPLUG-Owl2 [56] and Qwen-VL [3]. Given that BLIP-2 faced challenges in generating comprehensive and meaningful results for text-rich images, we considered InstructBLIP as an alternative. For all methods, we randomly selected three prompts from ten as instructions for the model (details provided in the Appendix C.1). + +Table 4 presents the results of different methods in terms of classical captioning metrics. Models with enhanced visual text understanding generally outperform general multimodal models, such as LLaVA, Mini-GPT4, and Instruct-BLIP. LaRA (zero-shot) refers to the LaRA model fine-tuned on the TRINS-QA dataset and demonstrates improved performance. Comparison of fine-tuned LaRA variants indicates that text recognition ability is still limited for OCR-free methods, suggesting that the CLIP encoder or feature projection process may cause visual information loss. Addressing this limitation may involve employing a better trained encoder on text-rich images or designing a more carefully crafted architecture, a direction we leave for future exploration. When fine-tuned with TRINS-Cap, LaRA exhibits much better performance, underscoring the importance of high-quality human-annotated data. Figure 8 shows examples of different models on the TRINS-Cap benchmark. It shows the great capability of LaRA in recognizing text and relating it to its visual contexts, demonstrating the effectiveness of the TRINS dataset. + +
MethodBackboneB@1B@2B@3B@4METEORROUGECIDEr
mPLUG-Owl2 [57]Llama-2-7B4.93.22.31.78.719.64.5
Instruct-BLIP [11]Vicuna-7B13.48.55.74.19.217.95.7
Qwen-VL [3]Qwen-7B28.818.412.28.614.124.316.6
Instruct-BLIP [11]15.99.96.44.49.618.68.0
Mini-GPT4 [63]31.116.18.65.011.420.86.3
Mini-GPT-v2 [7]Vicuna-13B27.914.78.04.810.820.97.3
LLaVA [28]35.118.29.55.413.222.28.8
LLaVAR [62]18.911.27.25.010.820.111.4
LLaVA 1.5 [27]31.116.49.56.111.921.815.4
LLaVAR w/ OCRVicuna-13B21.413.18.86.412.022.213.0
LaRA (zero-shot)29.319.112.99.214.826.321.3
LLaVAR (fine-tuned)Vicuna-13B36.525.418.013.417.832.435.7
LaRA37.726.418.914.218.433.246.7
+ +# 5.3. Additional Experiments + +Performance on general visual tasks after TRINS fine-tuning. We adopted the evaluation protocols of MiniGPT-v2 [7] and compared LaRA with LLaVA [28] on traditional visual question answering benchmarks in table 5. LaRA shows a comparable performance on knowledgeability and better performance on reasoning and spatial awareness. This further verifies the effectiveness of the TRINS dataset, demonstrating that fine-tuning on text-rich images does not degrade performance on natural images, but instead enhances the results. + +Table 4. Results of different models on text-rich image captioning tasks. + +
OKVQAGQAVSRVizWiz
LLaVA [28]57.841.351.245.0
LaRA58.142.453.053.1
+ +Table 5. Quantitative Results on the public visual benchmarks. + +
MetricsControlNetDeepFloydTextDiffuser
FID (↓)51.5949.9651.26
CLIP Score (↑)0.37170.39170.3707
OCR Acc. (↑)0.42410.21920.5027
+ +Table 6. Empirical Results on TRINS-Gen (easy) benchmark. + +Text-to-document generation Diffusion-based text-to-image generation has shown great success, while precise textual renderings remain a big challenge. TextDiffuser introduced the MARIO-Eval benchmark, drawing from works such as DrawBench [43] and DrawTextCreative [29]. However, most text prompts in MARIO-Eval are short and cannot serve as a good evaluation dataset to handle complex real-world human instructions. We take advantage of human annotations from TRINS-Cap and build the TRINS-Gen benchmark. It is still difficult to render too many words in a single image [6]. In response to this, we filter out images with more than 20 OCR words, resulting in a curated set of 2,104 images. We divide these images into two sets (easy + +and difficult) based on the number of OCR words and the length of the longest OCR string per annotation, where all text prompts in the easy set have less than 9 OCR words. We evaluated existing methods using their public checkpoints and reported the results in Table 6 and detailed results in Table 9 (Appendix C.3). + +![](images/d14b7ec39e5dcd0c0af0d8040545cc064affff63bec2753311ee8ee8c6723be9.jpg) +Text Prompts: The image is a book cover for 'Abandoned San Diego' by Jessica D. Johnson, featuring a background image of an abandoned house. +Ground Truth +Figure 9. Examples generated by different text-to-image models on the TRINS-Gen benchmarkk. + +![](images/77c326b75a08897e5707afb1967f5638896f5fe815fc5020a09a5dcf670f10ba.jpg) +ControlNet + +![](images/6efe6c91bee5bc0a943c4fe7ef9b58e15adc8e74234a93385aefed8b51987df7.jpg) +Deepfloyd + +![](images/cdf991f751a98211d980d47cc20151ecab9dad8d5c8545126d8e660952705b7a.jpg) +TextDiffuser + +# 6. Conclusions + +Despite the challenges posed by the prevalence of natural images in training data, the significance of visual textual understanding cannot be understated. In this paper, we introduce TRINS, a Text-Rich Image INStruction dataset, comprising a diverse collection of text-rich images, captions, and questions. This dataset, created through a semi-automatic process leveraging large-scale pre-trained models, not only significantly reduces annotation time but also elevates annotation quality. Furthermore, we propose a novel multimodal language model architecture, LaRA, which incorporates OCR as a pivotal enhancement for text-rich image understanding. We anticipate that continued progress in multimodal language model architectures, fine-tuning techniques, and the expansion of diverse, text-rich datasets like TRINS will push the boundaries of visual textual understanding. This, in turn, will facilitate more efficient collaboration between humans and agents, potentially revolutionizing numerous real-world applications. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1 +[2] Anas Awadalla, Irena Gao, Joshua Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortzman, and Ludwig Schmidt. Openflamingo, 2023. 22 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.7, 8 +[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 12 +[5] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts, 2021. 1 +[6] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. arXiv preprint arXiv:2305.10855, 2023. 1, 8, 12 +[7] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 8 +[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 6 +[9] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling instruction-finetuned language models, 2022. 1 +[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2 +[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructblip: Towards + +general-purpose vision-language models with instruction tuning. ArXiv, abs/2305.06500, 2023. 7, 8 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2020. 1 +[13] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marcal Rusinol, Minesh Mathew, C.V. Jawahar, Ernest Valveny, and Dimosthenis Karatzas. Icdar 2019 competition on scene text visual question answering. 2019 International Conference on Document Analysis and Recognition (ICDAR), 2019. 2, 5 +[14] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient visual instruction model, 2023. 2 +[15] Adam W. Harley, Alex Ufkes, and Konstantinos G. Derpanis. Evaluation of deep convolutional nets for document image classification and retrieval, 2015. 3 +[16] Rongjie Huang, Mingze Li, Dongchao Yang, Jiatong Shi, Xuankai Chang, Zhenhui Ye, Yuning Wu, Zhiqing Hong, Jia-Bin Huang, Jinglin Liu, Yixiang Ren, Zhou Zhao, and Shinji Watanabe. Audiogpt: Understanding and generating speech, music, sound, and talking head. ArXiv, abs/2304.12995, 2023. 2 +[17] Xinyu Huang, Youcai Zhang, Jinyu Ma, Weiwei Tian, Rui Feng, Yuejie Zhang, Yaqian Li, Yandong Guo, and Lei Zhang. Tag2text: Guiding vision-language model via image tagging. arXiv preprint arXiv:2303.05657, 2023. 3 +[18] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 1 +[19] Chunyuan Li. Large multimodal models: Notes on cvpr 2023 tutorial. ArXiv, abs/2306.14895, 2023. 1 +[20] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 1, 2023. 2 +[21] Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, and Furu Wei. Dit: Self-supervised pre-training for document image transformer. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 3 +[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023. 5, 22 +[23] Xian Li, Ping Yu, Chunting Zhou, Timo Schick, Luke Zettlemoyer, Omer Levy, Jason Weston, and Mike Lewis. Self-alignment with instruction backtranslation. arXiv preprint arXiv:2308.06259, 2023. 5 +[24] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. arXiv preprint arXiv:2311.06607, 2023. 2 + +[25] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 7 +[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 1 +[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 7, 8 +[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 2, 5, 6, 7, 8, 22 +[29] Rosanne Liu, Dan Garrette, Chitwan Sahara, William Chan, Adam Roberts, Sharan Narang, Irina Blok, RJ Mical, Mohammad Norouzi, and Noah Constant. Character-aware models improve visual text rendering. arXiv preprint arXiv:2212.10562, 2022.8, 12 +[30] Yuliang Liu, Zhang Li, Hongliang Li, Wenwen Yu, Mingxin Huang, Dezhi Peng, Mingyu Liu, Mingrui Chen, Chunyuan Li, Cheng lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models, 2023. 1, 2, 22 +[31] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Learning to draw Chinese characters in image synthesis models coherently, 2023. 12 +[32] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models, 2023. 2 +[33] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 2 +[34] Minesh Mathew, Dimosthenis Karatzas, and C. V. Jawahar. Docvqa: A dataset for vqa on document images, 2020. 5 +[35] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 2 +[36] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947–952. IEEE, 2019. 2 +[37] OpenAI. Gpt-4 technical report, 2023. 1, 5 +[38] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback, 2022. 1 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1, 3 + +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2 +[41] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 12 +[42] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. arXiv preprint arXiv:1809.02156, 2018. 5 +[43] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 8, 12 +[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022.3 +[45] Alex Shonenkov, Misha Konstantinov, Daria Bakshandaeva, Christoph Schuhmann, Ksenia Ivanova, and Nadija Klokova. Deepfloyd. https://github.com/deep-floyd/if, 2023.12 +[46] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 742-758. Springer, 2020. 2 +[47] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 5 +[48] Amanpreet Singh, Guan Pang, Mandy Toh, Jing Huang, Wojciech Galuba, and Tal Hassner. Textocr: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8802-8812, 2021. 2 +[49] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023.5 +[50] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 5, 6 + +[51] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 5 +[52] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 7 +[53] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions, 2022. 5 +[54] Zizhang Wu, Xinyuan Chen, Jizheng Wang, Xiaoquan Wang, Yanzhu Gan, Muqing Fang, and Tianhao Xu. Ocr-rtps: anOCR-based real-time positioning system for the valet parking. Applied Intelligence, pages 1-15, 2023. 6, 7 +[55] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 2 +[56] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 2, 7, 22 +[57] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. arXiv preprint arXiv:2311.04257, 2023. 7, 8, 22 +[58] Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities, 2023. 2 +[59] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2 +[60] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention, 2023. 2 +[61] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023. 3 +[62] Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 1, 2, 5, 6, 7, 8, 22 +[63] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 1, 2, 7, 8, 22 \ No newline at end of file diff --git a/2024/TRINS_ Towards Multimodal Language Models that Can Read/images.zip b/2024/TRINS_ Towards Multimodal Language Models that Can Read/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6ea16120f25c52f221f1b0a8aca09e36dba466ea --- /dev/null +++ b/2024/TRINS_ Towards Multimodal Language Models that Can Read/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b67c0b6c7c3c1b51c3f450c89dec3f81b667235868f12a8e0ce51185ff21cfe1 +size 665158 diff --git a/2024/TRINS_ Towards Multimodal Language Models that Can Read/layout.json b/2024/TRINS_ Towards Multimodal Language Models that Can Read/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..35b837ab38ebca67e30a22593ba65c36ec487671 --- /dev/null +++ b/2024/TRINS_ Towards Multimodal Language Models that Can Read/layout.json @@ -0,0 +1,7674 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 492, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 492, + 120 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 492, + 120 + ], + "type": "text", + "content": "TRINS: Towards Multimodal Language Models that Can Read" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 142, + 492, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 142, + 492, + 171 + ], + "spans": [ + { + "bbox": [ + 107, + 142, + 492, + 171 + ], + "type": "text", + "content": "Ruiyi Zhang" + }, + { + "bbox": [ + 107, + 142, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 142, + 492, + 171 + ], + "type": "text", + "content": ", Yanzhe Zhang" + }, + { + "bbox": [ + 107, + 142, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 142, + 492, + 171 + ], + "type": "text", + "content": ", Jian Chen" + }, + { + "bbox": [ + 107, + 142, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 142, + 492, + 171 + ], + "type": "text", + "content": ", Yufan Zhou" + }, + { + "bbox": [ + 107, + 142, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 142, + 492, + 171 + ], + "type": "text", + "content": ", Jiumiang Gu" + }, + { + "bbox": [ + 107, + 142, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 142, + 492, + 171 + ], + "type": "text", + "content": ", Changyou Chen" + }, + { + "bbox": [ + 107, + 142, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 142, + 492, + 171 + ], + "type": "text", + "content": ", Tong Sun" + }, + { + "bbox": [ + 107, + 142, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 171, + 539, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 171, + 539, + 186 + ], + "spans": [ + { + "bbox": [ + 60, + 171, + 539, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 60, + 171, + 539, + 186 + ], + "type": "text", + "content": " Adobe Research " + }, + { + "bbox": [ + 60, + 171, + 539, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 60, + 171, + 539, + 186 + ], + "type": "text", + "content": " Georgia Institute of Technology " + }, + { + "bbox": [ + 60, + 171, + 539, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 60, + 171, + 539, + 186 + ], + "type": "text", + "content": " State University of New York at Buffalo" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 244, + 187, + 353, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 187, + 353, + 199 + ], + "spans": [ + { + "bbox": [ + 244, + 187, + 353, + 199 + ], + "type": "text", + "content": "{ruizhang}@adobe.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 251, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 290, + 515 + ], + "type": "text", + "content": "Large multimodal language models have shown remarkable proficiency in understanding and editing images. However, a majority of these visually-tuned models struggle to comprehend the textual content embedded in images, primarily due to the limitation of training data. In this work, we introduce TRINS: a Text-Rich image1 INstruction dataset, with the objective of enhancing the reading ability of the multimodal large language model. TRINS is built upon LAION2 using hybrid data annotation strategies that include machine-assisted and human-assisted annotation process. It contains 39,153 text-rich images, captions, and 102,437 questions. Specifically, we show that the number of words per annotation in TRINS is significantly longer than that of related datasets, providing new challenges. Furthermore, we introduce a simple and effective architecture, called a Language-Vision Reading Assistant (LaRA), which is good at understanding textual content within images. LaRA outperforms existing state-of-the-art multimodal large language models on the TRINS dataset as well as other classical benchmarks. Lastly, we conducted a comprehensive evaluation with TRINS on various text-rich image understanding and generation tasks, demonstrating its effectiveness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 557, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 677 + ], + "type": "text", + "content": "Instruction tuning [9, 38] has shown a great generalization ability on unseen tasks and has contributed to the growing popularity of large language models (LLMs), such as ChatGPT [37]. Recently, multimodal language models benefit from visual instruction finetuning [1, 18, 19, 28, 63], and has shown great success in real-world applications. These models leverage visual encoders such as CLIP-ViT [12, 39] to empower LLMs with image comprehension ability. However, challenges arise in comprehension of textual information within images, which may stem from the prevalence of natu" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 228, + 547, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 547, + 360 + ], + "type": "text", + "content": "ral images in training datasets, such as Conceptual Captions [5] and COCO [26]), as highlighted by Liu et al. [30]. Recognizing the importance of visual textual understanding for effective collaboration between agents and humans, Zhang et al. [62] proposed enhancing end-to-end visual instruction-tuned models by introducing noisy Optical Character Recognition (OCR) annotations to improve vision-language alignment. In this work, we surpass existing achievements and collect a new Text-Rich image INStruction dataset named TRINS, which contains 39,153 text-rich images, captions and 102,437 questions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 366, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 548, + 715 + ], + "type": "text", + "content": "TRINS is created in a semi-automatic manner for a more controllable and faithful collection. Specifically, we exploited large-scale pre-trained models such as CLIP [40] and GPT-4 in the annotation process. This semi-automatic process significantly reduces the time and resources required for manual annotation and surprisingly improves the overall quality of annotations. TRINS dataset is composed of three datasets for captioning, visual question answering (VQA) and image generation, respectively. Specifically, human-annotated captions for text-rich images are first collected because they can best translate text-rich images into texts. During this process, extracted OCR words and recognize-anything model tags are provided to the annotators for better and efficient annotations. With detailed image descriptions, VAQ data is built and fulfilled by large language models, such as GPT-4 [37] and LLaMA-2 [51]. In detailed statistics and analysis, we found that both annotated captions and collected question-answer pairs are more comprehensive and contain significantly more details than the existing dataset. Therefore, we show the superior advantage of TRINS compared to existing instruction fine-tuning datasets. As a by-product, high-quality image-caption pairs can serve as a good benchmark for text-rich image generation, which is still a very challenging task [6]. At the same time, we propose a new simple and effective multimodal language model architecture that includes OCR as a component. We call it Language-vision Reading Assistant (LaRA) and show that LaRA fine-tuned on TRINS brings the best text-rich image understanding ability. Our contributions are as follows:" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 683, + 287, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 703 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 703 + ], + "type": "text", + "content": "1In this work, we use the phrase \"text-rich images\" to describe images with rich textual information, such as posters and book covers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 703, + 153, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 153, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 153, + 712 + ], + "type": "text", + "content": "2Work done during Q3 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22584" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 251 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "content": "- We introduce a novel dataset (TRINS) containing 39,153 captions and 102,437 high-quality text-rich image instruction pairs. TRINS is annotated with a novel semi-automatic annotation framework that is scalable and reliable." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 132, + 287, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 132, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 47, + 132, + 287, + 178 + ], + "type": "text", + "content": "- We develop several evaluation benchmarks for text-rich image understanding and generation tasks. Various methods are evaluated on TRINS, demonstrating the effectiveness of the dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 180, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 180, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 180, + 287, + 251 + ], + "type": "text", + "content": "- The TRINS datasets are high-quality and comprehensive, which is reflected not only in the dataset statistics but also from the results of multiple baseline models. Especially, our proposed LaRA finetuned on TRINS outperforms existing state-of-the-art methods on text-rich image understanding tasks." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 261, + 133, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 261, + 133, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 261, + 133, + 274 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 281, + 288, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 288, + 568 + ], + "type": "text", + "content": "Multimodal Instruction Tuning Multi-modal instruction tuning, including image, video [32, 59], and audio [16, 58] settings, has been an active research topic. MiniGPT-4 [63] uses ChatGPT to generate high-quality instruction-following data, while LLaVA [28] generates such data by prompting GPT-4 with captions and bounding boxes. LLaMA-Adapter [14, 60] aligns text-image features using COCO data, and mPLUG-owl [56] combines extensive image-text pairs for pretraining and a mixture of data for finetuning. Despite this, many models, according to Liu et al. [30], struggle with OCR tasks. InstructBLIP [10] addresses this by transforming 13 vision language tasks into an instruction-following format. mPLUGOwl [55, 56] apply multitask instruction funetuing using existing document datasets. A comprehensive survey is available in Li et al. [20]. LLaVAR [62] exploits GPT-4 to collect fine-tuning data without human annotations using OCR and captioning tools. It discovered that resolution plays a significant role in recognizing textual information and explored several options. Monkey [24] performed a surgery between simple text labels and high input resolution, enabling remarkable performance in visually-rich document images with dense text. TRINS exploits human-machine collaboration for data annotation and can provide more accurate information, reducing the problem of hallucination." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 582, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 714 + ], + "type": "text", + "content": "Text-Rich Image Datasets Visual question answering or captioning datasets are widely used in task-specific finetuning and large multimodal model evaluation. TextCap [46] is the first text-rich image captioning dataset. Compared to TextCap, TRINS-Cap provides more detailed annotations that can fulfill the requirement of instruction finetuning. TextOCR [48] aims comprehend text in the context of an image, which is similar to our motivation but focuses more on text recognition in images instead of understanding. STVQA [13] uses spatial and textual information to answer visually grounded questions, effectively integrating visual" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 538, + 178 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 538, + 178 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 538, + 178 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 538, + 178 + ], + "type": "table", + "html": "
DatasetYearSizeAnnotation Type
OCR-VQA2019200KQA (▲)
TextVQA201945KQA (▲)
TextCap2020140kCaption (▲)
TextOCR2021145kText Bbox (▲)
DocVQA202050kQA (▲)
40kCaption (▲)
TRINS (Ours)2023100kQA (▲ +▲)
40kText Bbox (▲)
", + "image_path": "00c482fa61e018f2cb1af6e9ef212201c62f6f9fdedd4a1ee8e7e9a3ddb89008.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 186, + 542, + 198 + ], + "lines": [ + { + "bbox": [ + 308, + 186, + 542, + 198 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 542, + 198 + ], + "type": "text", + "content": "Table 1. Comparison between TRINS and other related datasets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 212, + 547, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 212, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 212, + 547, + 390 + ], + "type": "text", + "content": "and textual cues. OCR-VQA [36] focuses on incorporating optical character recognition (OCR) into visual question answering (VQA), which operates primarily on text within images. TextVQA [47] also takes advantage of the textual information present in the images to answer questions, but with an emphasis on open questions. DocVQA takes this one step further by applying VQA to document images, handling a variety of layouts and formats. InfoVQA [35] and ChartQA [33] focus on specific subdomains and aim to answer questions about information graphics and chart images, respectively. In summary, these related works provide datasets for leveraging spatial and textual cues. TRINS-VQA is a dataset that exploits the semi-automatic annotation process. It can be used for general domain instruction finetuning and model evaluations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 404, + 508, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 404, + 508, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 404, + 508, + 418 + ], + "type": "text", + "content": "3. Text-Rich Image Instruction Dataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 426, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 653 + ], + "type": "text", + "content": "To equip multimodal language models with the ability to recognize text and relate it to its visual context, we have curated a new dataset named Text-Rich Instruction (TRINS). The ultimate goal is to enable these models to have spatial, semantic, and visual reasoning between multiple text tokens and visual entities. In this section, we present TRINS, a dataset crafted through a semi-automatic process. Specifically, we leverage large-scale pre-trained models like CLIP [40] and GPT-4 in the annotation process, offering potential advantages: (i) Significant reduction in annotation time and resources: using these models significantly reduces the time and resources required for manual annotation. (ii) Enhancement of annotation data quality through post-processing: the involvement of large models contributes to improving the overall quality of annotation data through subsequent post-processing. (iii) Functionality of large models as knowledge bases: large models can serve as effective knowledge bases, aiding in the annotation process by virtue of their extensive training in diverse datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "We first outline the document image collection process for TRINS, utilizing CLIP models. Then, we present data statistics to facilitate a comprehensive understanding. We delve into three distinct tasks derived from the TRINS dataset in detail: i) TRINS-Cap: Visual Captioning, ii) TRINS-VQA:" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22585" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 545, + 167 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 545, + 167 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 545, + 167 + ], + "type": "image", + "image_path": "2e07ccf16200e8bd467810bedd6b190e2e9d962a7d624f7fa2783da4a2892e6e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 224, + 536, + 342 + ], + "blocks": [ + { + "bbox": [ + 46, + 177, + 546, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 177, + 546, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 546, + 222 + ], + "type": "text", + "content": "Figure 1. Overview of TRINS data collection process, which consists of three datasets. Text-rich images are first selected from web images and then ask annotators to describe the image in details. i) TRINS-Cap is extracted from human annotations with heuristic data processing for text-rich image summarization tasks. ii) TRINS-VQA is built upon human annotations and generate question-answer pairs for training by prompting text-only large language models. iii) TRINS-Gen combined human annotations and text boxes for text-rich image generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 224, + 536, + 342 + ], + "lines": [ + { + "bbox": [ + 47, + 224, + 536, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 536, + 342 + ], + "type": "image", + "image_path": "9ff8f48869fbb1da497f7a7c8a13946925c11b54325674a532753e38d772c26c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 372, + 289, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 289, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 289, + 468 + ], + "type": "text", + "content": "Visual Question Answering and iii) TRINS-Gen: Text-to-Image Generation. The overview of the TRINS data collection process is illustrated in Figure 1. To provide a succinct overview, TRINS-Cap undergoes full annotation by human annotators, while TRINS-Gen and TRINS-VQA are constructed based on TRINS-Cap with the help of pre-trained models. A summary of the various TRINS datasets is presented in Table 7 of Appendix A." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 475, + 279, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 279, + 488 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 279, + 488 + ], + "type": "text", + "content": "3.1. Machine-Assisted Text-rich Image Selections" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 493, + 288, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 288, + 686 + ], + "type": "text", + "content": "Beginning with the LAION-5B dataset [44], our objective is to selectively retain images that exhibit a significant presence of text. Recognizing that document images typically contain substantial textual content, we initially formed a binary classification dataset by combining natural images with document data. Subsequently, we trained an image classifier using a DiT [21] base backbone, fine-tuned on the RVL-CDIP dataset [15]. The purpose of this classifier was to predict whether an image contains text. Then a subset was constructed by selecting images with a predicted probability greater than 0.8, while also adhering to the criteria " + }, + { + "bbox": [ + 46, + 493, + 288, + 686 + ], + "type": "inline_equation", + "content": "p(\\text{watermark}) < 0.8" + }, + { + "bbox": [ + 46, + 493, + 288, + 686 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 493, + 288, + 686 + ], + "type": "inline_equation", + "content": "p(\\text{unsafe}) < 0.5" + }, + { + "bbox": [ + 46, + 493, + 288, + 686 + ], + "type": "text", + "content": ", where both probabilities are derived from the metadata of the LAION dataset. Acknowledging the noise introduced due to the classifier's limitations, we further refined the dataset by incorporating human judgment. A random sample of 20,000 images" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 370, + 425, + 452 + ], + "blocks": [ + { + "bbox": [ + 75, + 350, + 515, + 361 + ], + "lines": [ + { + "bbox": [ + 75, + 350, + 515, + 361 + ], + "spans": [ + { + "bbox": [ + 75, + 350, + 515, + 361 + ], + "type": "text", + "content": "Figure 2. CLIP-based categorization of our collected images and selected representative data samples from each category." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 370, + 425, + 452 + ], + "lines": [ + { + "bbox": [ + 307, + 370, + 425, + 452 + ], + "spans": [ + { + "bbox": [ + 307, + 370, + 425, + 452 + ], + "type": "image", + "image_path": "63d498ffbbc00fa4d8ed66ff2ea00ceb34d9bb86fb247519e51be4241bbd4848.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 453, + 372, + 462 + ], + "lines": [ + { + "bbox": [ + 361, + 453, + 372, + 462 + ], + "spans": [ + { + "bbox": [ + 361, + 453, + 372, + 462 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 427, + 371, + 545, + 452 + ], + "blocks": [ + { + "bbox": [ + 427, + 371, + 545, + 452 + ], + "lines": [ + { + "bbox": [ + 427, + 371, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 427, + 371, + 545, + 452 + ], + "type": "image", + "image_path": "2d4b79776fd32e46a986dacbe18dd014d9d88e6071323881fd99ef94c145ec2c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 453, + 492, + 462 + ], + "lines": [ + { + "bbox": [ + 481, + 453, + 492, + 462 + ], + "spans": [ + { + "bbox": [ + 481, + 453, + 492, + 462 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 464, + 545, + 486 + ], + "lines": [ + { + "bbox": [ + 306, + 464, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 545, + 486 + ], + "type": "text", + "content": "Figure 3. Word clouds of (a) predicted tags and (b) detected words from the text-rich images of TRINS." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 491, + 547, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 547, + 682 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 547, + 682 + ], + "type": "text", + "content": "from the filtered LAION-5B was clustered into 50 groups based on CLIP-ViT-B/32 visual features. After inspecting the clustering results, one cluster was meticulously chosen, encompassing diverse text-rich images such as posters, covers, advertisements, and educational documents. This cluster model then served as the filtering mechanism for collecting images that comprise the TRINS dataset. For reference, we present a CLIP-based categorization [39] in Figure 2 to depict the distribution of images in the collected data. The major class is book cover images, further categorized on the basis of book themes and contents. To enhance our understanding of text-rich images, we employed a Recognize Anything Model (RAM) [17, 61] to extract tags from TRINS images. Figure 3a displays word clouds of RAM tags, where \"book\" and \"poster\" emerge as major keywords. Additionally, we utilize the Azure Read API " + }, + { + "bbox": [ + 304, + 491, + 547, + 682 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 304, + 491, + 547, + 682 + ], + "type": "text", + "content": " and PaddleOCR to extract" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 289, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 289, + 712 + ], + "type": "text", + "content": "3https://huggingface.co/datasets/laion/laion-high-resolution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 693, + 546, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 546, + 713 + ], + "type": "text", + "content": "4https://azure.microsoft.com/en-us/updates/computer-vision-v3-preview-6/" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22586" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 170, + 156 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 170, + 156 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 170, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 170, + 156 + ], + "type": "image", + "image_path": "e04720c003cdc236deadbd4276ac8aeb06b8f88947330391a6d3af5f56fcafb2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 156, + 118, + 163 + ], + "lines": [ + { + "bbox": [ + 109, + 156, + 118, + 163 + ], + "spans": [ + { + "bbox": [ + 109, + 156, + 118, + 163 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 136, + 164, + 455, + 175 + ], + "lines": [ + { + "bbox": [ + 136, + 164, + 455, + 175 + ], + "spans": [ + { + "bbox": [ + 136, + 164, + 455, + 175 + ], + "type": "text", + "content": "Figure 4. OCR word (a), Caption (b), Question (c) and Answer (d) statistics for TRINS." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 178, + 72, + 293, + 156 + ], + "blocks": [ + { + "bbox": [ + 178, + 72, + 293, + 156 + ], + "lines": [ + { + "bbox": [ + 178, + 72, + 293, + 156 + ], + "spans": [ + { + "bbox": [ + 178, + 72, + 293, + 156 + ], + "type": "image", + "image_path": "1b45b02cdc6115d23f128d22e7825dd04090b80d81de7594d3a234ec0cff66de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 156, + 239, + 163 + ], + "lines": [ + { + "bbox": [ + 230, + 156, + 239, + 163 + ], + "spans": [ + { + "bbox": [ + 230, + 156, + 239, + 163 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 299, + 72, + 413, + 156 + ], + "blocks": [ + { + "bbox": [ + 299, + 72, + 413, + 156 + ], + "lines": [ + { + "bbox": [ + 299, + 72, + 413, + 156 + ], + "spans": [ + { + "bbox": [ + 299, + 72, + 413, + 156 + ], + "type": "image", + "image_path": "3749cb3f0b249d37623a18c5cb51b1d1e78b2c644405fdbeb87b481be6831862.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 156, + 361, + 163 + ], + "lines": [ + { + "bbox": [ + 351, + 156, + 361, + 163 + ], + "spans": [ + { + "bbox": [ + 351, + 156, + 361, + 163 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 421, + 72, + 533, + 156 + ], + "blocks": [ + { + "bbox": [ + 421, + 72, + 533, + 156 + ], + "lines": [ + { + "bbox": [ + 421, + 72, + 533, + 156 + ], + "spans": [ + { + "bbox": [ + 421, + 72, + 533, + 156 + ], + "type": "image", + "image_path": "88282c48c156dd73e6fcc27346188ec1b9a5b044249e6cb45219a32213fc3b80.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 156, + 482, + 163 + ], + "lines": [ + { + "bbox": [ + 473, + 156, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 473, + 156, + 482, + 163 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 183, + 289, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 289, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 289, + 291 + ], + "type": "text", + "content": "text within TRINS images. The word cloud of the extracted texts is presented in Figure 3b. Figure 4a illustrates the distribution of OCR words per image, indicating that most of the images in TextCap have fewer than 10 words, while the TRINS images average 31.4 OCR words. Recognizing texts within TRINS images is more challenging because of the presence of numerous small words. In summary, TRINS images encompass rich visual content, seamlessly integrating textual information into the image context." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 52, + 300, + 276, + 439 + ], + "blocks": [ + { + "bbox": [ + 52, + 300, + 276, + 439 + ], + "lines": [ + { + "bbox": [ + 52, + 300, + 276, + 439 + ], + "spans": [ + { + "bbox": [ + 52, + 300, + 276, + 439 + ], + "type": "image", + "image_path": "b68098fd62743113499630aa43e14fa47bcc3c2eeb206808982629c26467f68c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 447, + 265, + 459 + ], + "lines": [ + { + "bbox": [ + 67, + 447, + 265, + 459 + ], + "spans": [ + { + "bbox": [ + 67, + 447, + 265, + 459 + ], + "type": "text", + "content": "Figure 5. Question type statistics based on key words." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 468, + 159, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 159, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 159, + 479 + ], + "type": "text", + "content": "3.2. Annotation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 486, + 288, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 570 + ], + "type": "text", + "content": "Annotator Selections All annotators (with the tag " + }, + { + "bbox": [ + 46, + 486, + 288, + 570 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 486, + 288, + 570 + ], + "type": "text", + "content": " Job Success and Top Rated Plus) are native English speakers and have experience with document annotations. We first asked all annotators to annotate a 200-image set and provide them with detailed annotation guidelines with multiple examples. In addition, we use Labelbox as an annotation tool and set quality control questions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 570, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 676 + ], + "type": "text", + "content": "Heuristic Filters We first use EasyOCR to extract texts from images and retain text phrases with more than three characters, a height greater than " + }, + { + "bbox": [ + 46, + 570, + 287, + 676 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 570, + 287, + 676 + ], + "type": "text", + "content": " of the canvas height, and a confidence score greater than 0.1. For each phrase retrieved, we employ an edit distance-based string matching algorithm (due to potentially erroneous OCR results) to search for its optimal matching substrings within the human-generated caption. The average score for all extracted phrases serves as a metric." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 677, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 712 + ], + "type": "text", + "content": "Manual Reviews We accept annotations with high metric scores and reject the lowest for rework. We manually review other annotations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 183, + 547, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 183, + 547, + 268 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 547, + 268 + ], + "type": "text", + "content": "Sensitive Images We combined neural models with human efforts to filter the images. The first step involves an initial data filtering by 2-3 individuals to filter out sensitive images for training. The second step involves hiring additional people to perform a further screening on the data. We engaged annotators from various countries to check the selected images." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 285, + 526, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 526, + 298 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 526, + 298 + ], + "type": "text", + "content": "3.3. TRINS-Cap: Text-Rich Image Captioning" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 307, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 547, + 545 + ], + "type": "text", + "content": "Annotation Process TRINS-Cap is a dataset fully annotated by human annotators. We hired 20 native English speakers with experience in document annotation through Upwork. The annotation process, conducted in LabelBox, involved a total of 2,079 hours to annotate 40,576 text-rich images, with an additional 159 hours allocated for result review. After filtering low-quality annotations and addressing missing images, we obtained a final set of 39,153 image-annotation pairs. The dataset is partitioned into train, validation, and test splits with sizes of 29,153, 5,000, and 5,000, respectively. All annotations undergo an initial automated review that involves matching the OCR words with the annotations. Subsequently, human evaluators conduct a thorough review, rejecting annotations with errors, and prompting annotators to rework them for enhancement. We provide comprehensive annotation instructions to all annotators to ensure that each annotation includes: (i) detailed descriptions of visual components. (ii) describe texts' location, attributes, and put texts into annotations. (iii) optional insights or abstract descriptions." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 545, + 548, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 548, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 548, + 712 + ], + "type": "text", + "content": "Statistics and Analysis The primary objective of the annotation process is to facilitate a human or machine's full comprehension of the information conveyed in the image without direct viewing. Consequently, the average annotation length for TRINS is 65.1 words, significantly exceeding that of COCO (10.6 words) and TextCaps (12.4 words). Figure 4b shows the caption length distributions for TRINS-Cap, COCO, and TextCap, demonstrating the comprehensive nature of the TRINS-Cap annotations. TRINS with more contexts can generally provide a better description of complex images, where short captions are insufficient. Hence, LLMs finetund on TRINS can better understand images with complex texts and layouts, which has been further verified in Section 5." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22587" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 85 + ], + "type": "text", + "content": "3.4. TRINS-VQA: Multimodal Question Answering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 290, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 290, + 390 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 290, + 390 + ], + "type": "text", + "content": "The annotation process for question answering is inherently complex, primarily due to the necessity for annotators to generate high-quality questions. Creating an effective question is more challenging than providing an answer. As a result, annotators frequently gravitate toward formulating concrete and extractive questions (e.g., \"Who is the author of this book?\") rather than abstract ones (e.g., \"How does the design of the book cover reflect the content of the book?\"). We introduce semi-automatic annotation methods to generate high-quality visual question-answering data for TRINS-VQA. This dataset is designed to train general vision language assistants through instruction fine-tuning, and its benefits on model performance are evaluated in Section 5.1. TRINS-Cap, on the other hand, serves as human-assisted annotations, offering a comprehensive but non-instructive dataset for fine-tuning. To utilize the wealth of high-quality annotations available, we incorporated semi-automatic annotation by using large language models (LLMs) such as OpenAI's GPT-4 [37] and Llama-70B [50, 51] to enhance our data annotation pipeline. OCR results and detailed descriptions of each image are provided to LLMs. Furthermore, high-quality human-crafted demonstrations and detailed annotation rules are provided to LLMs. One demonstration focused on extract questions, while the other emphasized abstract questions, creating a more balanced dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 396, + 288, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 288, + 479 + ], + "type": "text", + "content": "Human Annotations To facilitate a robust evaluation of model performance, we hired 10 Upwork annotators, whose native language is English, to annotate the test dataset, following a methodology similar to previous work [13, 34, 47]. The test dataset comprises 5,000 images with 18,764 question-answer pairs. These data collected are used exclusively for evaluation purposes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 486, + 288, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 486, + 288, + 570 + ], + "spans": [ + { + "bbox": [ + 47, + 486, + 288, + 570 + ], + "type": "text", + "content": "Statistics and Analysis Building upon prior research [28, 49, 53], we provide visualizations of instructions in Figure 5 based on question keywords. The inner cycle illustrates the distribution of the first word in the questions, while the outer cycle presents extracted keywords determined by carefully designed heuristics. Types of questions are categorized according to keywords found in questions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 288, + 714 + ], + "type": "text", + "content": "In Figure 4c and 4d, we present statistics on the number of words per question-answer pair, comparing them with previous work. Generally, the average length of questions for TRINS-VQA is 10.5, surpassing that of DocVQA (8.3), OCR-VQA (6.5), and TextVQA (7.1). Surprisingly, the average answer length for TRINS is 23.9, significantly longer than related datasets (all less than 4). This discrepancy arises from TRINS containing more abstract questions that typically have longer answers. Similarly, the dataset is divided into train, validation, and test splits. For extract questions, the accuracy of the answers is calculated similarly to Liu et al. [28], while for abstract questions, generation metrics" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 323, + 72, + 521, + 208 + ], + "blocks": [ + { + "bbox": [ + 323, + 72, + 521, + 208 + ], + "lines": [ + { + "bbox": [ + 323, + 72, + 521, + 208 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 521, + 208 + ], + "type": "image", + "image_path": "38df59e7c1c89b20c52f2a641f14a48022598067c0247d375ea5fd9356ab05e8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 219, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 305, + 219, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 305, + 219, + 547, + 275 + ], + "type": "text", + "content": "Figure 6. Model overview of the LaRA. the CLIP model processes the input image to generate patch-wise features. These features then serve as input to a projection layer, yielding visual tokens. Concurrently, an OCR tool extracts textual data from the image, which is then merged with the user instruction." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 285, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 545, + 308 + ], + "type": "text", + "content": "such as BLEU scores are used to assess the quality of the answers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 310, + 547, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 547, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 547, + 608 + ], + "type": "text", + "content": "The question-answering and instruction data we obtain are extensive, encompassing a balanced mix of extract and abstract questions. This includes detailed descriptions, summaries, question-answer pairs, tasks that promote creativity and the generation of novel thoughts, and conversational tasks. The dataset spans a diverse range of concepts, ranging from visual presentation and visual language relations to intricate reasoning tasks. Compared to previous methods [28], captions generated by models such as BLIP-2 [22] exhibit less informativeness than detailed human annotations, as demonstrated in Section 5.2. Additionally, captioning models may be susceptible to hallucinations [42], a concern mitigated in TRINS, which provides more comprehensive and reliable visual descriptions. Although OCR tools are robust, they can introduce noisy information. To address this, we utilize both the Azure Read API and PaddleOCR to extract text information. We added the potential unreliability in our system prompts to LLMs, instructing them to generate questions with assured answers. Ultimately, we directly leverage the responses from large language models (LLMs) to construct TRINS-VQA. The quality of the instruction data can be further enhanced through self-alignment [23] or by seeking verification from human annotators. Although this has not been explored in this work, we leave it as a potential direction for future research." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 619, + 545, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 619, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 305, + 619, + 545, + 646 + ], + "type": "text", + "content": "4. The Language-vision Reading Assistant (LaRA) Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "To verify our data, we adopt a simple architecture based on LLaVA [28] for language-vision reading assistant, as illustrated in Figure 6. A low-resolution visual encoder is a significant bottleneck for extracting textual information from images [62]. On the contrary, a high-resolution" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22588" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 74, + 187, + 249 + ], + "blocks": [ + { + "bbox": [ + 55, + 74, + 187, + 249 + ], + "lines": [ + { + "bbox": [ + 55, + 74, + 187, + 249 + ], + "spans": [ + { + "bbox": [ + 55, + 74, + 187, + 249 + ], + "type": "image", + "image_path": "5d4118388ee25eab4de75386866960a5381643c4ebfd47d717cb4f4a740214f0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 254, + 501, + 266 + ], + "lines": [ + { + "bbox": [ + 90, + 254, + 501, + 266 + ], + "spans": [ + { + "bbox": [ + 90, + 254, + 501, + 266 + ], + "type": "text", + "content": "Figure 7. Examples generated by different multimodal language models on the abstract TRINS-VQA benchmarkk." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 196, + 74, + 338, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 74, + 338, + 83 + ], + "spans": [ + { + "bbox": [ + 196, + 74, + 338, + 83 + ], + "type": "text", + "content": "Question: Can you describe the design of the book cover?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 195, + 86, + 518, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 86, + 518, + 116 + ], + "spans": [ + { + "bbox": [ + 195, + 86, + 518, + 116 + ], + "type": "text", + "content": "GT: The book cover is quite vibrant and engaging. The title, \"Explore Your World: Weird, Wild, Amazing!\" is written in large, white fonts that span across the top and bottom of the cover, making it immediately eye-catching. The author's name, Tim Flannery, is prominently displayed, and the illustrator, Sam Caldwell, is credited on the right side of the cover. The background of the cover is filled with various animals, adding a dynamic and exciting element that would likely appeal to children." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 196, + 119, + 517, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 119, + 517, + 147 + ], + "spans": [ + { + "bbox": [ + 196, + 119, + 517, + 147 + ], + "type": "text", + "content": "Qwen-VL-Chat: The book cover is a black background with various colorful illustrations of animals such as a crocodile, alligator, penguin, and snake in white and green. The title \"Weird Wild, Amazing!\" is written in white letters with a yellow border at the top. The author's name \"Tim Flannery\" is written in white letters above the title, and the publisher's name \"Explore Your World\" is written in green letters below the title. The cover also has a winner sticker for the Environment Children's Literature." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 196, + 152, + 517, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 152, + 517, + 168 + ], + "spans": [ + { + "bbox": [ + 196, + 152, + 517, + 168 + ], + "type": "text", + "content": "mPLUG-Owl2: The book cover features a colorful design with a blue background and a yellow bird. The text on the cover is written in a fun and playful font, emphasizing the title \"Tim Flannery: Explore Your World Weird Wild Amazing.\"" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 195, + 171, + 517, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 171, + 517, + 221 + ], + "spans": [ + { + "bbox": [ + 195, + 171, + 517, + 221 + ], + "type": "text", + "content": "GPT-4V: The book cover features a collage of various whimsical and colorful creatures, reminiscent of a style one might find in a children's book. The title \"Weird, Wild, Amazing!\" is prominently displayed at the center top in large, playful, and colorful lettering, which suggests the book may be about the variety and eccentricity of life on Earth. Below the title, the text \"Explore Your World\" is in a smaller font, inviting the reader to discover the contents. The name \"Tim Flannery\" appears above the title, indicating the author, with the descriptor \"Explore Your World\" just below the title. The art by Sam Caldwell is acknowledged at the bottom right, and there's a badge on the bottom left corner stating \"Winner Environment Award for Children's Literature,\" which suggests the book has been recognized for its contribution to environmental education. The background is dark, which makes the bright colors of the animals and other elements stand out. The animals appear to be drawn in a stylized and engaging manner, with expressions and poses that give them character and life. The overall design is vibrant and seems to be targeting a young audience, promising education through fun and engagement with the natural world." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 195, + 224, + 517, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 224, + 517, + 247 + ], + "spans": [ + { + "bbox": [ + 195, + 224, + 517, + 247 + ], + "type": "text", + "content": "LaRA: The book cover features the title \"Weird, Wild, Amazing!\" in bold letters, with the author's name, Tim Flannery, mentioned at the top. The backdrop of the cover includes illustrations of various animals, suggesting that the book may be about wildlife or exploration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 277, + 289, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 469 + ], + "type": "text", + "content": "visual encoder generates an excessive number of patches, drastically impacting model efficiency. Our conviction is that a low-resolution visual encoder is designed to capture visual information more effectively, including layouts. In contrast, a considerably smaller OCR tool is employed to extract text from high-resolution images. Instruction-tuning on TRINS could serve as an effective method of training LLMs to better align OCR texts and extracted visual features. In addition, LLMs can autonomously rectify errors generated by OCR tools. For the visual encoder " + }, + { + "bbox": [ + 46, + 277, + 289, + 469 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 277, + 289, + 469 + ], + "type": "text", + "content": ", we employ CLIP-ViT-L/14-336 at a resolution of " + }, + { + "bbox": [ + 46, + 277, + 289, + 469 + ], + "type": "inline_equation", + "content": "336 \\times 336" + }, + { + "bbox": [ + 46, + 277, + 289, + 469 + ], + "type": "text", + "content": ". The grid features before the last transformer layer are then mapped into the word embedding space of the language decoder using a trainable projection matrix " + }, + { + "bbox": [ + 46, + 277, + 289, + 469 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 277, + 289, + 469 + ], + "type": "text", + "content": ". Regarding the language decoder " + }, + { + "bbox": [ + 46, + 277, + 289, + 469 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 277, + 289, + 469 + ], + "type": "text", + "content": ", we utilize Vicuna-1.5-13B [8], a language model tuned for instructions through LLaMA 2 [50]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 469, + 289, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 289, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 289, + 650 + ], + "type": "text", + "content": "We follow the two-stage training design of LLaVA but adopt the pre-trained projection layer of LLaVAR [62]. Training targets remain similar: generating output responses " + }, + { + "bbox": [ + 46, + 469, + 289, + 650 + ], + "type": "inline_equation", + "content": "(<\\text{res}>)" + }, + { + "bbox": [ + 46, + 469, + 289, + 650 + ], + "type": "text", + "content": " for the input instructions " + }, + { + "bbox": [ + 46, + 469, + 289, + 650 + ], + "type": "inline_equation", + "content": "(<\\text{ins}>)" + }, + { + "bbox": [ + 46, + 469, + 289, + 650 + ], + "type": "text", + "content": ", alongside OCR results " + }, + { + "bbox": [ + 46, + 469, + 289, + 650 + ], + "type": "inline_equation", + "content": "(<\\text{ocr}>)" + }, + { + "bbox": [ + 46, + 469, + 289, + 650 + ], + "type": "text", + "content": ". The transformed image tokens " + }, + { + "bbox": [ + 46, + 469, + 289, + 650 + ], + "type": "inline_equation", + "content": "(<\\text{img}>)" + }, + { + "bbox": [ + 46, + 469, + 289, + 650 + ], + "type": "text", + "content": " are introduced before or after the first input instruction randomly when building the instruction finetuning data. During the finetuning stage, both the projection matrix " + }, + { + "bbox": [ + 46, + 469, + 289, + 650 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 469, + 289, + 650 + ], + "type": "text", + "content": " and the language decoder " + }, + { + "bbox": [ + 46, + 469, + 289, + 650 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 469, + 289, + 650 + ], + "type": "text", + "content": " are trained. We consolidate our nearly 90K visual question-answering data with the 158K instruction-following data from LLaVA to form the training set. It should be noted that the visual encoder remains frozen throughout the training period. Compared with previous approaches, LaRA incorporates OCR words as part of the input, a simple way to enhance visual text understanding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 657, + 128, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 128, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 128, + 671 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "In this section, we present three downstream tasks based on the TRINS dataset and outline their evaluation metrics. The proposed method LaRA is used in both the text-rich" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 310, + 275, + 544, + 389 + ], + "blocks": [ + { + "bbox": [ + 310, + 275, + 544, + 389 + ], + "lines": [ + { + "bbox": [ + 310, + 275, + 544, + 389 + ], + "spans": [ + { + "bbox": [ + 310, + 275, + 544, + 389 + ], + "type": "table", + "html": "
MethodRecog.VQASVQADKIEFinal Score
Gemini215174128134651
GPT-4v167163146160636
Monkey1741619188514
mPLUG-Owl21531534119366
LLaVAR1861222513346
LLaVA1.5-13B176129197331
mPLUG-Owl172104183297
MiniGPT-V21242940157
LaRA21114785105548
", + "image_path": "0e7aca71eed90e32fab63f8e0c2d48ce103dae80d9c7511ff756d9849337b15e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 398, + 545, + 431 + ], + "lines": [ + { + "bbox": [ + 305, + 398, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 398, + 545, + 431 + ], + "type": "text", + "content": "Table 2. Results of LMMs on OCRBench. Recog. represents text recognition, " + }, + { + "bbox": [ + 305, + 398, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\mathrm{VQA}^S" + }, + { + "bbox": [ + 305, + 398, + 545, + 431 + ], + "type": "text", + "content": " represents Scene Text-Centric VQA, " + }, + { + "bbox": [ + 305, + 398, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\mathrm{VQA}^D" + }, + { + "bbox": [ + 305, + 398, + 545, + 431 + ], + "type": "text", + "content": " represents Document-Oriented VQA." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 442, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 442, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 442, + 547, + 491 + ], + "type": "text", + "content": "image summarization and visual question-answer tasks. All experiments were conducted on NVIDIA A100 80GB GPUs. In fine-tuning, we use a cosine annealing schedule with an initial learning rate of " + }, + { + "bbox": [ + 305, + 442, + 547, + 491 + ], + "type": "inline_equation", + "content": "2e^{-5}" + }, + { + "bbox": [ + 305, + 442, + 547, + 491 + ], + "type": "text", + "content": " and a batch size of 32." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 502, + 545, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 545, + 527 + ], + "type": "text", + "content": "5.1. TRINS-VQA: Text-Rich Image Visual Question Answering" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": "We first performed experiments to evaluate the zero-shot performance of LaRA on classical benchmarks [28, 54]. The results are reported in Table 2 and Table 10. The proposed LaRA model exhibits significant performance improvement across all the datasets compared to other models. Even without OCR, LaRA outperforms other models in most cases, highlighting its robustness and effectiveness in handling visual question-answering tasks. The inclusion of OCR in LaRA further enhances performance, suggesting that Large Language Models (LLMs) can proficiently utilize textual information. However, the ability to directly extract text from images remains limited. LaRA, with its simple structure, significantly boosts model performance on text-rich images, offering an alternative solution to overcome the limitations of pre-trained image encoders. We further evaluate different" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22589" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 246, + 180 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 246, + 180 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 246, + 180 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 246, + 180 + ], + "type": "image", + "image_path": "919b4673629bbb617b4650b8d3fd7faae0359c6957ccbc901f358e59573eba09.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 255, + 74, + 520, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 74, + 520, + 99 + ], + "spans": [ + { + "bbox": [ + 255, + 74, + 520, + 99 + ], + "type": "text", + "content": "GT: There are big white words \"Urban Sketching\" at the bottom left of the cover. In the upper left corner of the cover, it reads \"Draw What You See.\" In the background of the picture, there is a drawing book and pen. It shows a signature in the bottom right corner of the book." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 255, + 102, + 520, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 102, + 520, + 131 + ], + "spans": [ + { + "bbox": [ + 255, + 102, + 520, + 131 + ], + "type": "text", + "content": "LLaVA: The image is an urban sketching project displayed on a tabletop, featuring a variety of hand-drawn city scenes. Several people can be seen in the drawing, walking around and engaging in different activities. Among the drawings, there is an image of an alleyway with a prominent Umbrella on the right side." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 255, + 135, + 520, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 135, + 520, + 159 + ], + "spans": [ + { + "bbox": [ + 255, + 135, + 520, + 159 + ], + "type": "text", + "content": "LLaVAR: The main subjects in this image are a notebook with a pen placed on top of it, and a photograph of a city street. The notebook and pen appear to be the focal point, with the urban scene in the background." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 255, + 163, + 520, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 163, + 520, + 178 + ], + "spans": [ + { + "bbox": [ + 255, + 163, + 520, + 178 + ], + "type": "text", + "content": "Instruct-BLIP: The main subjects in the image are a series of sketches and drawings depicting urban scenes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 60, + 184, + 520, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 184, + 520, + 201 + ], + "spans": [ + { + "bbox": [ + 60, + 184, + 520, + 201 + ], + "type": "text", + "content": "LaRA: This is a blog image. Black text in the upper left corner reads \"Draw What You See\". Large white text in the bottom left reads \"Urban Sketching\". The image background is a blue spiral notebook with a pen in the upper right and a picture of a colorful urban scene in the middle." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 60, + 225, + 532, + 365 + ], + "blocks": [ + { + "bbox": [ + 108, + 209, + 484, + 220 + ], + "lines": [ + { + "bbox": [ + 108, + 209, + 484, + 220 + ], + "spans": [ + { + "bbox": [ + 108, + 209, + 484, + 220 + ], + "type": "text", + "content": "Figure 8. Examples generated by different multimodal language models on the TRINS-Cap benchmarkk." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 225, + 532, + 365 + ], + "lines": [ + { + "bbox": [ + 60, + 225, + 532, + 365 + ], + "spans": [ + { + "bbox": [ + 60, + 225, + 532, + 365 + ], + "type": "table", + "html": "
MethodResolutionExtract AccuracyAbstract
B@1B@2B@3B@4METEORROUGECIDEr
Instruct-BLIP [11]\\( {224}^{2} \\)43.913.59.26.75.28.916.723.5
Mini-GPTv215.329.419.413.810.417.228.744.3
mPLUG-Owl2 [57]\\( {448}^{2} \\)61.035.925.519.315.217.738.097.2
Qwen-VL [3]63.640.227.319.915.220.436.479.4
LLaVA [28]23.733.120.914.410.518.031.848.2
LLaVAR [62]\\( {336}^{2} \\)51.738.025.919.114.717.435.684.9
LLaVAR w/ OCR58.140.328.020.916.318.737.397.3
LLaVA 1.5 [27]38.829.820.915.812.620.834.940.7
LLaVAR (finetuned)\\( {336}^{2} \\)61.245.331.923.918.720.439.6104.7
LaRA62.845.134.227.322.621.946.5186.6
", + "image_path": "513a715c3fc31547b95481cdb1111873cd113bd1725e590c7923e5a899cb3519.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 368, + 474, + 380 + ], + "lines": [ + { + "bbox": [ + 116, + 368, + 474, + 380 + ], + "spans": [ + { + "bbox": [ + 116, + 368, + 474, + 380 + ], + "type": "text", + "content": "Table 3. Results of different models on TRINS-VQA for text-rich image question-answering tasks." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 391, + 289, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 289, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 289, + 629 + ], + "type": "text", + "content": "methods on the TRINS-VQA dataset, as shown in Table 3. For extraction questions, we use the same metric as Wu et al. [54]. For abstract questions, where the answer is typically a longer sentence, we evaluate them based on text similarity metrics such as BLEU [25], ROUGE [25], and CIDEr [52]. In zero-shot inference, LLaVAR with OCR exhibits the best performance, reinforcing the importance of extracting textual information. Furthermore, mPLUG-Owl2 and Qwen-VL perform well and represent the best methods in extract question evaluations, showing that a high-resolution encoder can significantly improve model performance. Instruct-BLIP demonstrates good performance on extract questions, but did not fare as well on abstract questions, given that the answers provided are usually short and concise. Figure 7 shows an example of the responses of different methods on the abstract TRINS-VQA dataset, and more examples can be found in the Appendix D. Qwen-VL includes all details but does not provide high-level insights, such as the ground-truth annotation. Both mPLUG-Owl2 and GPT-4V suffer from hallucination issues." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 635, + 263, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 263, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 263, + 649 + ], + "type": "text", + "content": "5.2. TRINS-Cap: Text-rich image Captioning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": "In our experiments on TRINS-Cap, we ask large multimodal models to generate summaries based on text-rich images. The data set was divided into train, validation, and test sets. We compared LaRA with popular baselines, including InstructBLIP [11], Mini-GPT4 [63], LLaVA [28]," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 391, + 545, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 463 + ], + "type": "text", + "content": "LLaVAR [62], mPLUG-Owl2 [56] and Qwen-VL [3]. Given that BLIP-2 faced challenges in generating comprehensive and meaningful results for text-rich images, we considered InstructBLIP as an alternative. For all methods, we randomly selected three prompts from ten as instructions for the model (details provided in the Appendix C.1)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "content": "Table 4 presents the results of different methods in terms of classical captioning metrics. Models with enhanced visual text understanding generally outperform general multimodal models, such as LLaVA, Mini-GPT4, and Instruct-BLIP. LaRA (zero-shot) refers to the LaRA model fine-tuned on the TRINS-QA dataset and demonstrates improved performance. Comparison of fine-tuned LaRA variants indicates that text recognition ability is still limited for OCR-free methods, suggesting that the CLIP encoder or feature projection process may cause visual information loss. Addressing this limitation may involve employing a better trained encoder on text-rich images or designing a more carefully crafted architecture, a direction we leave for future exploration. When fine-tuned with TRINS-Cap, LaRA exhibits much better performance, underscoring the importance of high-quality human-annotated data. Figure 8 shows examples of different models on the TRINS-Cap benchmark. It shows the great capability of LaRA in recognizing text and relating it to its visual contexts, demonstrating the effectiveness of the TRINS dataset." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22590" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 537, + 234 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 537, + 234 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 537, + 234 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 537, + 234 + ], + "type": "table", + "html": "
MethodBackboneB@1B@2B@3B@4METEORROUGECIDEr
mPLUG-Owl2 [57]Llama-2-7B4.93.22.31.78.719.64.5
Instruct-BLIP [11]Vicuna-7B13.48.55.74.19.217.95.7
Qwen-VL [3]Qwen-7B28.818.412.28.614.124.316.6
Instruct-BLIP [11]15.99.96.44.49.618.68.0
Mini-GPT4 [63]31.116.18.65.011.420.86.3
Mini-GPT-v2 [7]Vicuna-13B27.914.78.04.810.820.97.3
LLaVA [28]35.118.29.55.413.222.28.8
LLaVAR [62]18.911.27.25.010.820.111.4
LLaVA 1.5 [27]31.116.49.56.111.921.815.4
LLaVAR w/ OCRVicuna-13B21.413.18.86.412.022.213.0
LaRA (zero-shot)29.319.112.99.214.826.321.3
LLaVAR (fine-tuned)Vicuna-13B36.525.418.013.417.832.435.7
LaRA37.726.418.914.218.433.246.7
", + "image_path": "88e8b677e6b78b56fb6d842481ec7ea6a44ffb984ab468765ddb39c5c30505da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 260, + 183, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 260, + 183, + 273 + ], + "spans": [ + { + "bbox": [ + 47, + 260, + 183, + 273 + ], + "type": "text", + "content": "5.3. Additional Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 278, + 290, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 290, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 290, + 398 + ], + "type": "text", + "content": "Performance on general visual tasks after TRINS fine-tuning. We adopted the evaluation protocols of MiniGPT-v2 [7] and compared LaRA with LLaVA [28] on traditional visual question answering benchmarks in table 5. LaRA shows a comparable performance on knowledgeability and better performance on reasoning and spatial awareness. This further verifies the effectiveness of the TRINS dataset, demonstrating that fine-tuning on text-rich images does not degrade performance on natural images, but instead enhances the results." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 63, + 402, + 271, + 449 + ], + "blocks": [ + { + "bbox": [ + 165, + 237, + 427, + 248 + ], + "lines": [ + { + "bbox": [ + 165, + 237, + 427, + 248 + ], + "spans": [ + { + "bbox": [ + 165, + 237, + 427, + 248 + ], + "type": "text", + "content": "Table 4. Results of different models on text-rich image captioning tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 402, + 271, + 449 + ], + "lines": [ + { + "bbox": [ + 63, + 402, + 271, + 449 + ], + "spans": [ + { + "bbox": [ + 63, + 402, + 271, + 449 + ], + "type": "table", + "html": "
OKVQAGQAVSRVizWiz
LLaVA [28]57.841.351.245.0
LaRA58.142.453.053.1
", + "image_path": "0821fc75c608feb5722e7392fee5ceea3eeb058ed2ccd4dcfcf65bd3de1ebef7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 51, + 476, + 282, + 534 + ], + "blocks": [ + { + "bbox": [ + 53, + 453, + 280, + 464 + ], + "lines": [ + { + "bbox": [ + 53, + 453, + 280, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 280, + 464 + ], + "type": "text", + "content": "Table 5. Quantitative Results on the public visual benchmarks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 476, + 282, + 534 + ], + "lines": [ + { + "bbox": [ + 51, + 476, + 282, + 534 + ], + "spans": [ + { + "bbox": [ + 51, + 476, + 282, + 534 + ], + "type": "table", + "html": "
MetricsControlNetDeepFloydTextDiffuser
FID (↓)51.5949.9651.26
CLIP Score (↑)0.37170.39170.3707
OCR Acc. (↑)0.42410.21920.5027
", + "image_path": "4f5ba0ae4e88177a07c6039e5454fb396b0bde33a9cf78c89599950b21c1400f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 537, + 279, + 548 + ], + "lines": [ + { + "bbox": [ + 54, + 537, + 279, + 548 + ], + "spans": [ + { + "bbox": [ + 54, + 537, + 279, + 548 + ], + "type": "text", + "content": "Table 6. Empirical Results on TRINS-Gen (easy) benchmark." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "Text-to-document generation Diffusion-based text-to-image generation has shown great success, while precise textual renderings remain a big challenge. TextDiffuser introduced the MARIO-Eval benchmark, drawing from works such as DrawBench [43] and DrawTextCreative [29]. However, most text prompts in MARIO-Eval are short and cannot serve as a good evaluation dataset to handle complex real-world human instructions. We take advantage of human annotations from TRINS-Cap and build the TRINS-Gen benchmark. It is still difficult to render too many words in a single image [6]. In response to this, we filter out images with more than 20 OCR words, resulting in a curated set of 2,104 images. We divide these images into two sets (easy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 261, + 547, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 547, + 334 + ], + "type": "text", + "content": "and difficult) based on the number of OCR words and the length of the longest OCR string per annotation, where all text prompts in the easy set have less than 9 OCR words. We evaluated existing methods using their public checkpoints and reported the results in Table 6 and detailed results in Table 9 (Appendix C.3)." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 361, + 359, + 419 + ], + "blocks": [ + { + "bbox": [ + 309, + 345, + 544, + 360 + ], + "lines": [ + { + "bbox": [ + 309, + 345, + 544, + 360 + ], + "spans": [ + { + "bbox": [ + 309, + 345, + 544, + 360 + ], + "type": "text", + "content": "Text Prompts: The image is a book cover for 'Abandoned San Diego' by Jessica D. Johnson, featuring a background image of an abandoned house." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 361, + 359, + 419 + ], + "lines": [ + { + "bbox": [ + 310, + 361, + 359, + 419 + ], + "spans": [ + { + "bbox": [ + 310, + 361, + 359, + 419 + ], + "type": "image", + "image_path": "d14b7ec39e5dcd0c0af0d8040545cc064affff63bec2753311ee8ee8c6723be9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 419, + 351, + 426 + ], + "lines": [ + { + "bbox": [ + 318, + 419, + 351, + 426 + ], + "spans": [ + { + "bbox": [ + 318, + 419, + 351, + 426 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 438, + 545, + 460 + ], + "lines": [ + { + "bbox": [ + 305, + 438, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 438, + 545, + 460 + ], + "type": "text", + "content": "Figure 9. Examples generated by different text-to-image models on the TRINS-Gen benchmarkk." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 361, + 361, + 420, + 418 + ], + "blocks": [ + { + "bbox": [ + 361, + 361, + 420, + 418 + ], + "lines": [ + { + "bbox": [ + 361, + 361, + 420, + 418 + ], + "spans": [ + { + "bbox": [ + 361, + 361, + 420, + 418 + ], + "type": "image", + "image_path": "77c326b75a08897e5707afb1967f5638896f5fe815fc5020a09a5dcf670f10ba.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 419, + 406, + 426 + ], + "lines": [ + { + "bbox": [ + 379, + 419, + 406, + 426 + ], + "spans": [ + { + "bbox": [ + 379, + 419, + 406, + 426 + ], + "type": "text", + "content": "ControlNet" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 424, + 361, + 482, + 418 + ], + "blocks": [ + { + "bbox": [ + 424, + 361, + 482, + 418 + ], + "lines": [ + { + "bbox": [ + 424, + 361, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 424, + 361, + 482, + 418 + ], + "type": "image", + "image_path": "6efe6c91bee5bc0a943c4fe7ef9b58e15adc8e74234a93385aefed8b51987df7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 419, + 466, + 426 + ], + "lines": [ + { + "bbox": [ + 441, + 419, + 466, + 426 + ], + "spans": [ + { + "bbox": [ + 441, + 419, + 466, + 426 + ], + "type": "text", + "content": "Deepfloyd" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 485, + 361, + 543, + 418 + ], + "blocks": [ + { + "bbox": [ + 485, + 361, + 543, + 418 + ], + "lines": [ + { + "bbox": [ + 485, + 361, + 543, + 418 + ], + "spans": [ + { + "bbox": [ + 485, + 361, + 543, + 418 + ], + "type": "image", + "image_path": "cdf991f751a98211d980d47cc20151ecab9dad8d5c8545126d8e660952705b7a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 419, + 529, + 426 + ], + "lines": [ + { + "bbox": [ + 500, + 419, + 529, + 426 + ], + "spans": [ + { + "bbox": [ + 500, + 419, + 529, + 426 + ], + "type": "text", + "content": "TextDiffuser" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 477, + 384, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 477, + 384, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 477, + 384, + 490 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "content": "Despite the challenges posed by the prevalence of natural images in training data, the significance of visual textual understanding cannot be understated. In this paper, we introduce TRINS, a Text-Rich Image INStruction dataset, comprising a diverse collection of text-rich images, captions, and questions. This dataset, created through a semi-automatic process leveraging large-scale pre-trained models, not only significantly reduces annotation time but also elevates annotation quality. Furthermore, we propose a novel multimodal language model architecture, LaRA, which incorporates OCR as a pivotal enhancement for text-rich image understanding. We anticipate that continued progress in multimodal language model architectures, fine-tuning techniques, and the expansion of diverse, text-rich datasets like TRINS will push the boundaries of visual textual understanding. This, in turn, will facilitate more efficient collaboration between humans and agents, potentially revolutionizing numerous real-world applications." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22591" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 156 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 213 + ], + "type": "text", + "content": "[2] Anas Awadalla, Irena Gao, Joshua Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortzman, and Ludwig Schmidt. Openflamingo, 2023. 22" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 215, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 215, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 288, + 270 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 272, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 326 + ], + "type": "text", + "content": "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 328, + 288, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 288, + 362 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 288, + 362 + ], + "type": "text", + "content": "[5] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 364, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 288, + 397 + ], + "type": "text", + "content": "[6] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. arXiv preprint arXiv:2305.10855, 2023. 1, 8, 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 399, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 288, + 464 + ], + "type": "text", + "content": "[7] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 466, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 288, + 521 + ], + "type": "text", + "content": "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 53, + 466, + 288, + 521 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 53, + 466, + 288, + 521 + ], + "type": "text", + "content": " * chatgpt quality, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 523, + 288, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 523, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 53, + 523, + 288, + 632 + ], + "type": "text", + "content": "[9] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling instruction-finetuned language models, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "text", + "content": "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructblip: Towards" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 95 + ], + "type": "text", + "content": "general-purpose vision-language models with instruction tuning. ArXiv, abs/2305.06500, 2023. 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 161 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 163, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 228 + ], + "type": "text", + "content": "[13] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marcal Rusinol, Minesh Mathew, C.V. Jawahar, Ernest Valveny, and Dimosthenis Karatzas. Icdar 2019 competition on scene text visual question answering. 2019 International Conference on Document Analysis and Recognition (ICDAR), 2019. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 230, + 547, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 547, + 273 + ], + "type": "text", + "content": "[14] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient visual instruction model, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 275, + 547, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 547, + 307 + ], + "type": "text", + "content": "[15] Adam W. Harley, Alex Ufkes, and Konstantinos G. Derpanis. Evaluation of deep convolutional nets for document image classification and retrieval, 2015. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 309, + 547, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 547, + 373 + ], + "type": "text", + "content": "[16] Rongjie Huang, Mingze Li, Dongchao Yang, Jiatong Shi, Xuankai Chang, Zhenhui Ye, Yuning Wu, Zhiqing Hong, Jia-Bin Huang, Jinglin Liu, Yixiang Ren, Zhou Zhao, and Shinji Watanabe. Audiogpt: Understanding and generating speech, music, sound, and talking head. ArXiv, abs/2304.12995, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 376, + 547, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 547, + 420 + ], + "type": "text", + "content": "[17] Xinyu Huang, Youcai Zhang, Jinyu Ma, Weiwei Tian, Rui Feng, Yuejie Zhang, Yaqian Li, Yandong Guo, and Lei Zhang. Tag2text: Guiding vision-language model via image tagging. arXiv preprint arXiv:2303.05657, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 421, + 547, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 547, + 464 + ], + "type": "text", + "content": "[18] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 466, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 487 + ], + "type": "text", + "content": "[19] Chunyuan Li. Large multimodal models: Notes on cvpr 2023 tutorial. ArXiv, abs/2306.14895, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 489, + 547, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 547, + 533 + ], + "type": "text", + "content": "[20] Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao. Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint arXiv:2309.10020, 1, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 534, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 547, + 578 + ], + "type": "text", + "content": "[21] Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, and Furu Wei. Dit: Self-supervised pre-training for document image transformer. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 579, + 547, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 547, + 612 + ], + "type": "text", + "content": "[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023. 5, 22" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 613, + 547, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 547, + 656 + ], + "type": "text", + "content": "[23] Xian Li, Ping Yu, Chunting Zhou, Timo Schick, Luke Zettlemoyer, Omer Levy, Jason Weston, and Mike Lewis. Self-alignment with instruction backtranslation. arXiv preprint arXiv:2308.06259, 2023. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "type": "text", + "content": "[24] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. arXiv preprint arXiv:2311.06607, 2023. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22592" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[25] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 288, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 288, + 152 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 288, + 152 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 153, + 288, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 153, + 288, + 186 + ], + "spans": [ + { + "bbox": [ + 49, + 153, + 288, + 186 + ], + "type": "text", + "content": "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 189, + 288, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 189, + 288, + 210 + ], + "spans": [ + { + "bbox": [ + 49, + 189, + 288, + 210 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 2, 5, 6, 7, 8, 22" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 212, + 288, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 212, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 49, + 212, + 288, + 266 + ], + "type": "text", + "content": "[29] Rosanne Liu, Dan Garrette, Chitwan Sahara, William Chan, Adam Roberts, Sharan Narang, Irina Blok, RJ Mical, Mohammad Norouzi, and Noah Constant. Character-aware models improve visual text rendering. arXiv preprint arXiv:2212.10562, 2022.8, 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 269, + 287, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 269, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 49, + 269, + 287, + 313 + ], + "type": "text", + "content": "[30] Yuliang Liu, Zhang Li, Hongliang Li, Wenwen Yu, Mingxin Huang, Dezhi Peng, Mingyu Liu, Mingrui Chen, Chunyuan Li, Cheng lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models, 2023. 1, 2, 22" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 315, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 315, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 49, + 315, + 288, + 357 + ], + "type": "text", + "content": "[31] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Learning to draw Chinese characters in image synthesis models coherently, 2023. 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 360, + 288, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 360, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 49, + 360, + 288, + 393 + ], + "type": "text", + "content": "[32] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 395, + 288, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 395, + 288, + 438 + ], + "spans": [ + { + "bbox": [ + 49, + 395, + 288, + 438 + ], + "type": "text", + "content": "[33] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 441, + 288, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 441, + 288, + 463 + ], + "spans": [ + { + "bbox": [ + 49, + 441, + 288, + 463 + ], + "type": "text", + "content": "[34] Minesh Mathew, Dimosthenis Karatzas, and C. V. Jawahar. Docvqa: A dataset for vqa on document images, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 464, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 464, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 49, + 464, + 288, + 508 + ], + "type": "text", + "content": "[35] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 510, + 288, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 510, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 49, + 510, + 288, + 564 + ], + "type": "text", + "content": "[36] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947–952. IEEE, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 567, + 225, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 567, + 225, + 578 + ], + "spans": [ + { + "bbox": [ + 49, + 567, + 225, + 578 + ], + "type": "text", + "content": "[37] OpenAI. Gpt-4 technical report, 2023. 1, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 580, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 580, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 49, + 580, + 288, + 656 + ], + "type": "text", + "content": "[38] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 288, + 713 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "type": "text", + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 194 + ], + "type": "text", + "content": "[41] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 12" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 546, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 546, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 546, + 229 + ], + "type": "text", + "content": "[42] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. arXiv preprint arXiv:1809.02156, 2018. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 232, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 547, + 297 + ], + "type": "text", + "content": "[43] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 8, 12" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 299, + 547, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 547, + 363 + ], + "type": "text", + "content": "[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022.3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 365, + 547, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 547, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 547, + 408 + ], + "type": "text", + "content": "[45] Alex Shonenkov, Misha Konstantinov, Daria Bakshandaeva, Christoph Schuhmann, Ksenia Ivanova, and Nadija Klokova. Deepfloyd. https://github.com/deep-floyd/if, 2023.12" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 411, + 547, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 547, + 475 + ], + "type": "text", + "content": "[46] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 742-758. Springer, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 478, + 547, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 547, + 532 + ], + "type": "text", + "content": "[47] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 535, + 547, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 547, + 590 + ], + "type": "text", + "content": "[48] Amanpreet Singh, Guan Pang, Mandy Toh, Jing Huang, Wojciech Galuba, and Tal Hassner. Textocr: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8802-8812, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 591, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 547, + 645 + ], + "type": "text", + "content": "[49] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023.5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 547, + 712 + ], + "type": "text", + "content": "[50] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 5, 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22593" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 698 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[51] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "text", + "content": "[52] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 288, + 217 + ], + "type": "text", + "content": "[53] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 288, + 262 + ], + "type": "text", + "content": "[54] Zizhang Wu, Xinyuan Chen, Jizheng Wang, Xiaoquan Wang, Yanzhu Gan, Muqing Fang, and Tianhao Xu. Ocr-rtps: anOCR-based real-time positioning system for the valet parking. Applied Intelligence, pages 1-15, 2023. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "type": "text", + "content": "[55] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 288, + 385 + ], + "type": "text", + "content": "[56] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 2, 7, 22" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 386, + 288, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 386, + 288, + 440 + ], + "spans": [ + { + "bbox": [ + 48, + 386, + 288, + 440 + ], + "type": "text", + "content": "[57] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. arXiv preprint arXiv:2311.04257, 2023. 7, 8, 22" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 441, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 441, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 441, + 288, + 485 + ], + "type": "text", + "content": "[58] Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 487, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 288, + 520 + ], + "type": "text", + "content": "[59] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 520, + 288, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 288, + 564 + ], + "type": "text", + "content": "[60] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 565, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 565, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 565, + 288, + 609 + ], + "type": "text", + "content": "[61] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 610, + 288, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 610, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 610, + 288, + 654 + ], + "type": "text", + "content": "[62] Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 1, 2, 5, 6, 7, 8, 22" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 655, + 288, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 655, + 288, + 698 + ], + "spans": [ + { + "bbox": [ + 48, + 655, + 288, + 698 + ], + "type": "text", + "content": "[63] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 1, 2, 7, 8, 22" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22594" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/51572829-1348-4002-a40a-f8702f404317_content_list.json b/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/51572829-1348-4002-a40a-f8702f404317_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e9405b83dd7e283bb03549eadd2f173144c8abf0 --- /dev/null +++ b/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/51572829-1348-4002-a40a-f8702f404317_content_list.json @@ -0,0 +1,1626 @@ +[ + { + "type": "text", + "text": "TRIP: Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models*", + "text_level": 1, + "bbox": [ + 184, + 130, + 785, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhongwei Zhang†, Fuchen Long§, Yingwei Pan§, Zhaofan Qiu§, Ting Yao§, Yang Cao† and Tao Mei§ \n†University of Science and Technology of China, Hefei, China \n§HiDream.ai Inc.", + "bbox": [ + 150, + 204, + 815, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhwzhang@mail.ustc.edu.cn,{longfuchen, pandy, qiuzhaofan, tiyao}@hidream.ai forrest@ustc.edu.cn, tmei@hidream.ai", + "bbox": [ + 156, + 258, + 818, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 312, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in text-to-video generation have demonstrated the utility of powerful diffusion models. Nevertheless, the problem is not trivial when shaping diffusion models to animate static image (i.e., image-to-video generation). The difficulty originates from the aspect that the diffusion process of subsequent animated frames should not only preserve the faithful alignment with the given image but also pursue temporal coherence among adjacent frames. To alleviate this, we present TRIP, a new recipe of image-to-video diffusion paradigm that pivots on image noise prior derived from static image to jointly trigger inter-frame relational reasoning and ease the coherent temporal modeling via temporal residual learning. Technically, the image noise prior is first attained through one-step backward diffusion process based on both static image and noised video latent codes. Next, TRIP executes a residual-like dual-path scheme for noise prediction: 1) a shortcut path that directly takes image noise prior as the reference noise of each frame to amplify the alignment between the first frame and subsequent frames; 2) a residual path that employs 3D-UNet over noised video and static image latent codes to enable inter-frame relational reasoning, thereby easing the learning of the residual noise for each frame. Furthermore, both reference and residual noise of each frame are dynamically merged via attention mechanism for final video generation. Extensive experiments on WebVid-10M, DTDB and MSR-VTT datasets demonstrate the effectiveness of our TRIP for image-to-video generation. Please see our project page at https://trip-i2v.github.io/TRIP/.", + "bbox": [ + 75, + 358, + 473, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 823, + 209, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, deep generative models have demonstrated impressive capabilities to create high-quality visual con", + "bbox": [ + 75, + 848, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/983675b4f8017e056172a5a2c80320ab12aa4c9f5a2e3d5ae88402bca7bde2e9.jpg", + "image_caption": [ + "(a) Independent noise prediction", + "Figure 1. Comparison between (a) independent noise prediction and (b) our residual-like noise prediction with image noise prior in I2V diffusion models." + ], + "image_footnote": [], + "bbox": [ + 511, + 325, + 686, + 488 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dfad6fbaebb001f86268039a5a504716ef7e2de6923644bf2a57e1eae95db71e.jpg", + "image_caption": [ + "(b) Residual-like noise prediction with image noise prior" + ], + "image_footnote": [], + "bbox": [ + 691, + 325, + 875, + 486 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tent. In between, Generative Adversarial Networks (GANs) [16, 75] and diffusion models [7, 24] have brought forward milestone improvement for a series of generative tasks in computer vision field, e.g., text-to-image generation [44, 50], image editing [3, 22], and text-to-video generation [2, 25, 26]. In this work, we are interested in imaged-to-video generation (I2V), i.e., animating a static image by endowing it with motion dynamics, which has great potential real-world applications (e.g., entertainment and creative media). Conventional I2V techniques [9, 12, 40] commonly target for generating video from a natural scene image with random or very coarse motion dynamics (such as clouds moving or fluid flowing). In contrast, our focus is a more challenging text-driven I2V scenario: given the input static image and text prompt, the animated frames should not only faithfully align with the given first frame, but also be temporally coherent and semantically matched with text prompt.", + "bbox": [ + 496, + 551, + 892, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several recent pioneering practices [17, 60] for I2V is to remould the typical latent diffusion model of text-to-video generation by directly leveraging the given static image as additional condition during diffusion process. Figure 1 (a) illustrates such conventional noise prediction strategy in latent diffusion model. Concretely, by taking the given", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was performed at HiDream.ai.", + "bbox": [ + 94, + 886, + 346, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8671", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "static image as the first frame, 2D Variational Auto-encoder (VAE) is first employed to encode the first frame as image latent code. This encoded image latent code is further concatenated with noised video latent code (i.e., a sequence of Gaussian noise) to predict the backward diffusion noise of each subsequent frame via a learnable 3D-UNet. Nevertheless, such independent noise prediction strategy of each frame leaves the inherent relation between the given image and each subsequent frame under-exploited, and is lacking in efficacy of modeling temporal coherence among adjacent frames. As observed in I2V results of Figure 1 (a), both the foreground content (a glass) and the background content (e.g., a table) in the synthesized second frame are completely different from the ones in the given first frame. To alleviate this issue, our work shapes a new way to formulate the noise prediction in I2V diffusion model as temporal residual learning on the basis of amplified guidance of given image (i.e., image noise prior). As illustrated in Figure 1 (b), our unique design is to integrate the typical noise prediction with an additional shortcut path that directly estimates the backward diffusion noise of each frame by solely referring image noise prior. Note that this image noise prior is calculated based on both input image and noised video latent codes, which explicitly models the inherent correlation between the given first frame and each subsequent frame. Meanwhile, such residual-like scheme re-shapes the typical noise prediction via 3D-UNet as residual noise prediction to ease temporal modeling among adjacent frames, thereby leading to more temporally coherent I2V results.", + "bbox": [ + 76, + 90, + 472, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By materializing the idea of executing image-conditioned noise prediction in a residual manner, we present a novel diffusion model, namely Temporal Residual learning with Image noise Prior (TRIP), to facilitate I2V. Specifically, given the input noised video latent code and the corresponding static image latent code, TRIP performs residual-like noise prediction along two pathways. One is the shortcut path that first achieves image noise prior based on static image and noised video latent codes through one-step backward diffusion process, which is directly regarded as the reference noise of each frame. The other is the residual path which concatenates static image and noised video latent codes along temporal dimension and feeds them into 3D-UNet to learn the residual noise of each frame. Eventually, a Transformer-based temporal noise fusion module is leveraged to dynamically fuse the reference and residual noises of each frame, yielding high-fidelity video that faithfully aligns with given image.", + "bbox": [ + 76, + 529, + 472, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 815, + 218, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-Video Diffusion Models. The great success of diffusion models [44, 50] to generate fancy images based on text prompts has been witnessed in recent years. The foundation of the basic generative diffusion models fur", + "bbox": [ + 76, + 839, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ther encourages the development of customized image synthesis, such as image editing [3, 22, 46] and personalized image generation [8, 13, 33, 42, 51, 74]. Inspired by the impressive results of diffusion models on image generation, a series of text-to-video (T2V) diffusion models [2, 19, 25, 26, 32, 39, 52, 59, 64, 65, 70] start to emerge. VDM [26] is one of the earlier works that remoulds text-to-image (T2I) 2D-UNet architecture with temporal attention for text-to-video synthesis. Later in [2] and [18], the temporal modules (e.g., temporal convolution and self-attention) in diffusion models are solely optimized to emphasize motion learning. To further alleviate the burden on temporal modeling, Ge et al. [14] present a mixed video noise model under the assumption that each frame should correspond to a shared common noise and its own independent noise. Similar recipe is also adopted by VideoFusion [39] which estimates the basic noise and residual noise via two image diffusion models for video denoising. Other works go one step further by applying diffusion models for customized video synthesis, such as video editing [4, 6, 15, 45, 69] and personalized video generation [18, 21, 73]. In this work, we choose the basic T2V diffusion model with 3D-UNet [59] as our backbone for image-to-video generation task.", + "bbox": [ + 496, + 90, + 893, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-to-Video Generation. Image-to-Video (I2V) generation is one kind of conditional video synthesis, where the core condition is a given static image. According to the availability of motion cues, I2V approaches can be grouped into two categories: stochastic generation (solely using input image as condition) [27, 34, 41, 66] and conditional generation (using image and other types of conditions) [31, 38, 40, 43, 61]. Early attempts of I2V generation mostly belong to the stochastic models which usually focus on the short motion of fluid elements (e.g., water flow) [12, 27] or human poses [62] for landscape or human video generation. To further enhance the controllability of motion modeling, the conditional models exploit more types of conditions like text prompts [9, 17] in the procedure of video generation. Recent advances start to explore diffusion models [60, 72] for conditional I2V task. One pioneering work (VideoComposer [60]) flexibly composes a video with additional conditions of texts, optical flow or depth mapping sequence for I2V generation. Nevertheless, VideoComposer independently predicts the noise of each frame and leaves the inherent relation between input image and other frames under-exploited, thereby resulting in temporal inconsistency among frames.", + "bbox": [ + 496, + 443, + 893, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep Residual Learning. The effectiveness of learning residual components with additional shortcut connections in deep neural networks [20, 29, 54] has been verified in various vision tasks. Besides typical convolution-based structure, recent advances [10, 56] also integrate the shortcut connection into the emerging transformer-based structure [5, 35-37, 71]. Intuitively, learning residual part with refer", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8672", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/067e4d0946162a94f0a173b77564d0ad26b672674905f7e58a4e7b1e6cdfd349.jpg", + "image_caption": [ + "Figure 2. An overview of (a) our Temporal Residual learning with Image noise Prior (TRIP), and (b) the conceptual comparison between typical frame-wise denoising and our temporal noise fusion. During training, the input video clip is encoded into a sequence of image latent codes via a pre-trained 2D VAE. The latent code of first frame is then concatenated with noised video latent code as the input of the 3D-UNet. Then, TRIP performs a residual-like noise prediction along two pathways, i.e., shortcut path and residual path. The shortcut path computes the image noise prior as the reference noise based on the first frame latent codes and the noised video latent codes through a one-step backward diffusion. The residual path leverages 3D-UNet to predict the residual noise with reference to the image noise prior. By dynamically merging the reference and residual noise via a Transformer-based Temporal Noise Fusion (TNF) module, the target noise vector for video denoising is attained. Both 3D-UNet and TNF module are jointly optimized with the typical MSE loss for I2V generation." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 624, + 251 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/22e148f9973d9fd46f9a651a29b2ea155ec091d6223a2912e1c7d6cb68444910.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 89, + 875, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ence to the optimal identity mapping will ease the network optimization. In this work, we capitalize on such principle and formulate noise prediction in I2V diffusion as temporal residual learning on the basis of the image noise prior derived from static image to enhance temporal coherence.", + "bbox": [ + 75, + 383, + 468, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our work designs a novel image-to-video diffusion paradigm with residual-like noise prediction. The proposed TRIP contributes by studying not only how to excavate the prior knowledge of given image as noise prior to guide temporal modeling, and how to strengthen temporal consistency conditioned on such image noise prior.", + "bbox": [ + 75, + 460, + 468, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Approach", + "text_level": 1, + "bbox": [ + 76, + 571, + 223, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present our Temporal Residual learning with Image noise Prior (TRIP) for I2V generation. Figure 2 illustrates an overview of our TRIP. Given a video clip at training, TRIP first extracts the per-frame latent code via a pre-trained VAE and groups them as a video latent code. The image latent code of the first frame (i.e., the given static image) is regarded as additional condition, which is further concatenated with the noised video latent code along temporal dimension as the input of the 3D-UNet. Next, TRIP executes a residual-like noise prediction along two pathways (i.e., shortcut path and residual path), aiming to trigger inter-frame relational reasoning. In shortcut path, the image noise prior of each frame is calculated based on the correlation between the first frame and subsequent frames through one-step backward diffusion process. In residual path, the residual noise is estimated by employing 3D-UNet over the concatenated noised video latent code plus text prompt. Finally, a transformer-based Temporal Noise Fusion (TNF) module dynamically merges the image noise prior and the residual noise as the predicted noise for video generation.", + "bbox": [ + 75, + 598, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries: Video Diffusion Model", + "text_level": 1, + "bbox": [ + 500, + 383, + 823, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "T2V diffusion models [2, 26] are usually constructed by remoulding T2I diffusion models [44, 49]. A 3D-UNet with text prompt condition is commonly learnt to denoise a randomly sampled noise sequence in Gaussian distribution for video synthesis. Specifically, given an input video clip with $N$ frames, a pre-trained 2D VAE encoder $\\mathcal{E}(\\cdot)$ first extracts latent code of each frame to constitute a latent code sequence $\\{z_0^i\\}_{i=1}^N$ . Then, this latent sequence is grouped via concatenation along temporal dimension as a video latent code $z_0$ . Next, the Gaussian noise is gradually added to $z_0$ through forward diffusion procedure (length: $T$ , noise scheduler: $\\{\\beta_t\\}_{t=1}^T$ ). The noised latent code $z_t$ at each time step $t$ is thus formulated as:", + "bbox": [ + 498, + 407, + 890, + 603 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\epsilon \\sim \\mathcal {N} (0, I), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 612, + 890, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_t$ , $\\alpha_t = 1 - \\beta_t$ and $\\epsilon$ is the adding noise vector sampled from standard normal distribution $\\mathcal{N}(0, I)$ . 3D-UNet with parameter $\\theta$ aims to estimate the noise $\\epsilon$ based on the multiple inputs (i.e., noised video latent code $z_t$ , time step $t$ , and text prompt feature $c$ extracted by CLIP [48]). The Mean Square Error (MSE) loss is leveraged as the final objective $L$ :", + "bbox": [ + 498, + 637, + 890, + 743 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {\\epsilon \\sim \\mathcal {N} (0, I), t, c} [ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, c) \\| ^ {2} ]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 752, + 890, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the inference stage, the video latent code $\\hat{z}_0$ is iteratively denoised by estimating noise $\\epsilon_{\\theta}$ in each time step via 3DUNet, and the following 2D VAE decoder $\\mathcal{D}(\\cdot)$ reconstructs each frame based on the video latent code $\\hat{z}_0$ .", + "bbox": [ + 498, + 777, + 890, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Image Noise Prior", + "text_level": 1, + "bbox": [ + 500, + 847, + 676, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from typical T2V diffusion model, I2V generation model goes one step further by additionally emphasizing the", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8673", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "faithful alignment between the given first frame and the subsequent frames. Most existing I2V techniques [17, 60] control the video synthesis by concatenating the image latent code of first frame with the noised video latent code along channel/ temporal dimension. In this way, the visual information of first frame would be propagated across all frames via the temporal modules (e.g., temporal convolution and self-attention). Nevertheless, these solutions only rely on the temporal modules to calibrate video synthesis and leave the inherent relation between given image and each subsequent frame under-exploited, which is lacking in efficacy of temporal coherence modeling. To alleviate this issue, we formulate the typical noise prediction in I2V generation as temporal residual learning with reference to an image noise prior, targeting for amplifying the alignment between synthesized frames and first static image.", + "bbox": [ + 75, + 90, + 472, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, given the image latent code $z_0^1$ of first frame, we concatenate it with the noised video latent code sequence $\\{z_t^i\\}_{i=1}^N$ ( $t$ denotes time steps in forward diffusion) along temporal dimension, yielding the conditional input $\\{z_0^1, z_t^1, z_t^2, \\dots, z_t^N\\}$ for 3D-UNet. Meanwhile, we attempt to excavate the image noise prior to reflect the correlation between the first frame latent code $z_0^1$ and $i$ -th noised frame latent code $z_t^i$ . According to Eq. (1), we can reconstruct the $i$ -th frame latent code $z_0^i$ in one-step backward diffusion process as follows:", + "bbox": [ + 76, + 332, + 472, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {0} ^ {i} = \\frac {z _ {t} ^ {i} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} ^ {i}}{\\sqrt {\\bar {\\alpha} _ {t}}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 489, + 468, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon_t^i$ denotes Gaussian noise which is added to the $i$ -th frame latent code. Considering the basic assumption in I2V that all frames in a short video clip are inherently correlated to the first frame, the $i$ -th frame latent code $z_0^i$ can be naturally formulated as the combination of the first frame latent code $z_0^1$ and a residual item $\\Delta z^i$ as:", + "bbox": [ + 75, + 532, + 472, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {0} ^ {i} = z _ {0} ^ {1} + \\Delta z ^ {i}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 628, + 468, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we construct a variable $C_t^i$ by adding a scale ratio to the residual item $\\Delta z^i$ as follows:", + "bbox": [ + 75, + 655, + 468, + 683 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {t} ^ {i} = \\frac {\\sqrt {\\bar {\\alpha} _ {t}} \\Delta z ^ {i}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 689, + 468, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus Eq. (5) can be re-written as follows:", + "bbox": [ + 76, + 729, + 354, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta z ^ {i} = \\frac {\\sqrt {1 - \\bar {\\alpha} _ {t}} C _ {t} ^ {i}}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 751, + 468, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After integrating Eq. (3) and Eq. (6) into Eq. (4), the first frame latent code $z_0^1$ is denoted as:", + "bbox": [ + 75, + 790, + 468, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} z _ {0} ^ {1} = z _ {0} ^ {i} - \\Delta z ^ {i} = \\frac {z _ {t} ^ {i} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} ^ {i}}{\\sqrt {\\bar {\\alpha} _ {t}}} - \\frac {\\sqrt {1 - \\bar {\\alpha} _ {t}} C _ {t} ^ {i}}{\\sqrt {\\bar {\\alpha} _ {t}}} \\\\ = \\frac {z _ {t} ^ {i} - \\sqrt {1 - \\bar {\\alpha} _ {t}} (\\epsilon_ {t} ^ {i} + C _ {t} ^ {i})}{\\sqrt {\\bar {\\alpha} _ {t}}} = \\frac {z _ {t} ^ {i} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} ^ {i \\rightarrow 1}}{\\sqrt {\\bar {\\alpha} _ {t}}}, (7) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 828, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon_t^{i\\to 1}$ is defined as the image noise prior which can be interpreted as the relationship between the first frame latent code $z_0^1$ and $i$ -th noised frame latent code $z_{t}^{i}$ . The image noise prior $\\epsilon_t^{i\\rightarrow 1}$ is thus measured as:", + "bbox": [ + 496, + 90, + 893, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {t} ^ {i \\rightarrow 1} = \\frac {z _ {t} ^ {i} - \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} ^ {1}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 162, + 890, + 198 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to Eq. (8), the latent code $z_0^1$ of first frame can be directly reconstructed through the one-step backward diffusion by using the image noise prior $\\epsilon_t^{i\\to 1}$ and the noised latent code $z_{t}^{i}$ of $i$ -th frame. In this way, such image noise prior $\\epsilon_t^{i\\to 1}$ acts as a reference noise for Gaussian noise $\\epsilon_t^i$ which is added in $i$ -th frame latent code, when $z_0^i$ is close to $z_0^1$ (i.e., $i$ -th frame is similar to the given first frame).", + "bbox": [ + 496, + 208, + 893, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Temporal Residual Learning", + "text_level": 1, + "bbox": [ + 498, + 324, + 756, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recall that the image noise prior learnt in the shortcut pathway facilitates temporal modeling to amplify the alignment between the first frame and subsequent frames. To further strengthen temporal coherence among all adjacent frames, a residual pathway is designed to estimate the residual noise of each frame through inter-frame relational reasoning. Inspired by exploring knowledge prior in pre-trained 3DUNet for video editing [58], we propose to fine-tune 3DUNet to estimate the residual noise of each frame. Technically, the estimated noise $\\tilde{\\epsilon}_t^i$ of $i$ -th frame is formulated as the combination of the reference noise $\\epsilon_t^{i\\rightarrow 1}$ (i.e., image noise prior) and the estimated residual noise $\\Delta \\tilde{\\epsilon}_t^i$ :", + "bbox": [ + 496, + 347, + 890, + 529 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} _ {t} ^ {i} = \\lambda^ {i} \\epsilon_ {t} ^ {i \\rightarrow 1} + (1 - \\lambda^ {i}) \\Delta \\tilde {\\epsilon} _ {t} ^ {i}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 539, + 890, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda^i$ is a trade-off parameter. $\\Delta \\tilde{\\epsilon}_t^i$ is learnt by 3D-UNet (parameter: $\\theta$ ), and $\\epsilon_t^{i\\rightarrow 1}$ can be computed by Eq. (8) on the fly. Finally, the I2V denoising diffusion objective $\\tilde{L}$ is calculated as follows:", + "bbox": [ + 496, + 569, + 893, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathcal {L}} = \\mathbb {E} _ {\\epsilon \\sim \\mathcal {N} (0, I), t, c, i} [ \\| \\epsilon_ {t} ^ {i} - \\tilde {\\epsilon} _ {t} ^ {i} \\| ^ {2} ]. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 640, + 890, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that the computation of Eq. (9) is operated through the Temporal Noise Fusion (TNF) module that estimates the backward diffusion noise of each frame. Moreover, since the temporal correlation between $i$ -th frame and the first frame will decrease when increasing frame index $i$ with enlarged timespan, we shape the trade-off parameter $\\lambda^i$ as a linear decay parameter with respect to frame index.", + "bbox": [ + 496, + 670, + 890, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Temporal Noise Fusion Module", + "text_level": 1, + "bbox": [ + 500, + 786, + 777, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our residual-like dual-path noise prediction explores the image noise prior as a reference to facilitate temporal modeling among adjacent frames. However, the handcrafted design of TNF module with simple linear fusion (i.e., Eq. (9)) requires a careful tuning of hyper-parameter $\\lambda^i$ , leading to a sub-optimal solution. Instead, we devise a new", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8674", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6375901c1e2ffd877e7c7ae2e0b8b0ec637cacd7929c76bb081efc47edf21450.jpg", + "image_caption": [ + "Figure 3. Detailed structure of Temporal Noise Fusion module." + ], + "image_footnote": [], + "bbox": [ + 117, + 80, + 433, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transformer-based temporal noise fusion module to dynamically fuse reference and residual noises, aiming to further exploit the relations in between and boost noise fusion.", + "bbox": [ + 75, + 260, + 468, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 3 depicts the structure of Temporal Noise Fusion module. Specifically, given the estimated residual noise $\\Delta \\tilde{\\epsilon}_t^i$ , we first leverage the adaptive layer norm operations [68] which is modulated by the time step $t$ for spatial feature normalization. After that, one self-attention layer is employed for feature enhancement, followed by the same adaptive layer normalization operation. Next, by taking the in-process intermedia feature as query and the concatenated feature of $[\\epsilon_t^{i\\rightarrow 1},\\Delta \\tilde{\\epsilon}_t^i ]$ as key/value, TNF module achieves the final backward diffusion noise $\\tilde{\\epsilon}_{\\theta}(z_{t},t,c,i)$ through a cross-attention layer:", + "bbox": [ + 75, + 306, + 472, + 473 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} _ {\\theta} \\left(z _ {t}, t, c, i\\right) = \\tilde {\\epsilon} _ {t} ^ {i} = \\varphi \\left(\\Delta \\tilde {\\epsilon} _ {t} ^ {i}, \\epsilon^ {i \\rightarrow 1}, t\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 479, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varphi(\\cdot)$ denotes the operation of TNF module. Compared to the simple linear fusion strategy for noise estimation, our Transformer-based TNF module sidesteps the hyper-parameter tuning and provides an elegant alternative to dynamically merge the reference and residual noises of each frame to yield high-fidelity videos.", + "bbox": [ + 75, + 507, + 472, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 609, + 209, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 635, + 284, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We empirically verify the merit of our TRIP model on WebVid-10M [1], DTDB [30] and MSR-VTT [67] datasets. Our TRIP is trained over the training set of WebVid-10M, and the validation sets of all three datasets are used for evaluation. The WebVid-10M dataset consists of about $10.7M$ video-caption pairs with a total of $57K$ video hours. $5K$ videos are used for validation and the remaining videos are utilized for training. We further sample 2,048 videos from the validation set of WebVid-10M for evaluation. DTDB contains more than $10K$ dynamic texture videos of natural scenes. We follow the standard protocols in [9] to evaluate models on a video subset derived from 4 categories of scenes (i.e., fires, clouds, vegetation, and waterfall). The texts of each category are used as the text prompt since there is no text description. For the MSR-VTT dataset, there are $10K$ clips in training set and 2,990", + "bbox": [ + 75, + 657, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3e59f825e397aba8bbbda3337304ad8a8897f0077c5a24607aa6004a52529a7b.jpg", + "table_caption": [ + "Table 1. Performances of F-Consistency (F-Consistency4: consistency among the first four frames, F-Consistencyall: consistency among all frames) and FVD on WebVid-10M." + ], + "table_footnote": [], + "table_body": "
ApproachF-Consistency4 (↑)F-Consistencyall (↑)FVD (↓)
T2V-Zero [32]91.5992.15279
VideoComposer [60]88.7892.52231
TRIP95.3696.4138.9
", + "bbox": [ + 503, + 127, + 893, + 185 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "clips in validation set. Each video clip is annotated with 20 text descriptions. We adopt the standard setting in [59] to pre-process the video clips in validation set for evaluation.", + "bbox": [ + 496, + 194, + 890, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We implement our TRIP on PyTorch platform with Diffusers [57] library. The 3D-UNet [59] derived from Stable-Diffusion v2.1 [50] is exploited as the video backbone. Each training sample is a 16-frames clip and the sampling rate is 4 fps. We fix the resolution of each frame as $256 \\times 256$ , which is centrally cropped from the resized video. The noise scheduler $\\{\\beta_t\\}_{t=1}^T$ is set as linear scheduler ( $\\beta_1 = 1 \\times 10^{-4}$ and $\\beta_T = 2 \\times 10^{-2}$ ). We set the number of time steps $T$ as 1,000 in model training. The sampling strategy for video generation is DDIM [53] with 50 steps. The 3D-UNet and TNF module are trained with AdamW optimizer (learning rate: $2 \\times 10^{-6}$ for 3D-UNet and $2 \\times 10^{-4}$ for TNF module). All experiments are conducted on 8 NVIDIA A800 GPUs.", + "bbox": [ + 496, + 239, + 892, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. Since there is no standard evaluation metric for I2V task, we choose the metrics of frame consistency (F-Consistency) which is commonly adopted in video editing [58, 63] and Frechet Video Distance (FVD) [55] widely used in T2V generation [2, 39] for the evaluation on WebVid-10M. Note that the vision model used to measure F-Consistency is the CLIP ViT-L/14 [48]. Specifically, we report the frame consistency of the first 4 frames (F-Consistency $_4$ ) and all the 16 frames (F-Consistency $_{all}$ ). For the evaluation on DTDB and MSR-VTT, we follow [9, 59] and report the frame-wise Frechet Inception Distance (FID) [23] and FVD performances.", + "bbox": [ + 496, + 452, + 893, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Human Evaluation. The user study is also designed to assess the temporal coherence, motion fidelity, and visual quality of the generated videos. In particular, we randomly sample 256 testing videos from the validation set in WebVid-10M for evaluation. Through the Amazon MTurk platform, we invite 32 evaluators, and each evaluator is asked to choose the better one from two synthetic videos generated by two different methods given the same inputs.", + "bbox": [ + 496, + 635, + 893, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparisons with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 498, + 768, + 875, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our TRIP with several state-of-the-art I2V generation methods, e.g., VideoComposer [60] and T2V-Zero [32], on WebVid-10M. For the evaluation on DTDB, we set AL [11] and cINN [9] as baselines. Text-to-Video generation advances, e.g., CogVideo [28] and Make-A-Video [52] are included for comparison on MSR-VTT.", + "bbox": [ + 496, + 792, + 890, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation on WebVid-10M. Table 1 summarizes per", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8675", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6f16fa540ecea4fdd6408090ee56ff24ee5602f70a1030184a9bd37a583e06d7.jpg", + "image_caption": [ + "Figure 4. Performance comparisons of F-Consistency by using different number of frames on WebVid-10M." + ], + "image_footnote": [], + "bbox": [ + 117, + 73, + 423, + 248 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f7762f8a0a4a2d09546410498ff11397101226a735adc39fc788e3ae7599dc73.jpg", + "table_caption": [ + "Table 2. Performances of averaged FID and FVD over four scene categories on DTDB dataset." + ], + "table_footnote": [], + "table_body": "
ApproachZero-shotFID (↓)FVD (↓)
AL [11]No65.1934.2
cINN [9]No31.9451.6
TRIPYes24.8433.9
", + "bbox": [ + 81, + 313, + 468, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "formance comparison on WebVid-10M. Overall, our TRIP consistently leads to performance boosts against existing diffusion-based baselines across different metrics. Specifically, TRIP achieves the F-Consistency4 of 95.36%, which outperforms the best competitor T2V-Zero by 3.77%. The highest frame consistency among the first four frames of our TRIP generally validates the best temporal coherence within the beginning of the whole video. Such results basically demonstrate the advantage of exploring image noise prior as the guidance to strengthen the faithful visual alignment between given image and subsequent frames. To further support our claim, Figure 4 visualizes the frame consistency of each run which is measured among different numbers of frames. The temporal consistencies of TRIP across various frames constantly surpass other baselines. It is worthy to note that TRIP even shows large superiority when only calculating the consistency between the first two frames. This observation again verifies the effectiveness of our residual-like noise prediction paradigm that nicely preserves the alignment with given image. Considering the metric of FVD, TRIP also achieves the best performance, which indicates that the holistic motion dynamics learnt by our TRIP are well-aligned with the ground-truth video data distribution in WebVid-10M.", + "bbox": [ + 75, + 386, + 468, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 5 showcases four I2V generation results across three methods (VideoComposer, T2V-Zero, and our TRIP) based on the same given first frame and text prompt. Generally, compared to the two baselines, our TRIP synthesizes videos with higher quality by nicely preserving the faithful alignment with the given image & text prompt and meanwhile reflecting temporal coherence. For example, T2V-Zero produces frames that are inconsistent with the second prompt \"the ship passes close to the camera\" and VideoComposer generates temporally inconsistent frames for the", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/39249d7376660e6f8c4dc06404c0ab333bea4633dddf58e436df8d6ca099d43d.jpg", + "table_caption": [ + "Table 3. Performances of FID and FVD on MSR-VTT dataset." + ], + "table_footnote": [], + "table_body": "
ApproachModel TypeFID (↓)FVD (↓)
CogVideo [28]T2V23.591294
Make-A-Video [52]T2V13.17-
VideoComposer [60]T2V-580
ModelScopeT2V [59]T2V11.09550
VideoComposer [60]I2V31.29208
TRIPI2V9.6891.3
", + "bbox": [ + 504, + 89, + 890, + 171 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a9a91ee204a4f50ba2db70881273d3d131a8646ac872ff26b815df5e09b20b8d.jpg", + "table_caption": [ + "Table 4. Human evaluation of the preference ratios between TRIP v.s. other approaches on WebVid-10M dataset." + ], + "table_footnote": [], + "table_body": "
Evaluation Itemsv.s. T2V-Zerov.s. VideoComposer
Temporal Coherence96.9 v.s. 3.184.4 v.s. 15.6
Motion Fidelity93.8 v.s. 6.281.3 v.s. 18.7
Visual Quality90.6 v.s. 9.487.5 v.s. 12.5
", + "bbox": [ + 503, + 219, + 890, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "last prompt \"the portrait of the girl in glasses\". In contrast, the synthesized videos of our TRIP are more temporally coherent and better aligned with inputs. This again confirms the merit of temporal residual learning with amplified guidance of image noise prior for I2V generation.", + "bbox": [ + 496, + 291, + 890, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on DTDB. Next, we evaluate the generalization ability of our TRIP on DTDB dataset in a zero-shot manner. Table 2 lists the performances of averaged FID and FVD over four scene-related categories. In general, our TRIP outperforms two strong stochastic models (AL and cINN) over both two metrics. Note that cINN also exploits the residual information for I2V generation, but it combines image and residual features as clip feature by learning a bijective mapping. Our TRIP is fundamentally different in that we integrate residual-like noise estimation into the video denoising procedure to ease temporal modeling in the diffusion model. Figure 6 shows the I2V generation results of four different scenes produced by our TRIP. As shown in this figure, TRIP manages to be nicely generalized to scene-related I2V generation task and synthesizes realistic videos conditioned on the given first frame plus class label.", + "bbox": [ + 496, + 367, + 892, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on MSR-VTT. We further compare our TRIP with recent video generation advances on MSR-VTT. Note that here we also evaluate TRIP in a zero-shot manner. Table 3 details the frame-wise FID and FVD of different runs. Similar to the observation on DTDB, our TRIP again surpasses all baselines over both metrics. The results basically validate the strong generalization ability of our proposal to formulate the open-domain motion dynamics.", + "bbox": [ + 496, + 609, + 890, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Human Evaluation", + "text_level": 1, + "bbox": [ + 500, + 741, + 681, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to the evaluation over automatic metrics, we also perform human evaluation to investigate user preference with regard to three perspectives (i.e., temporal coherence, motion fidelity, and visual quality) across different I2V approaches. Table 4 shows the comparisons of user preference on the generated videos over WebVid-10M dataset. Overall, our TRIP is clearly the winner in terms of all three criteria compared to both baselines (VideoComposer and T2V-Zero). This demonstrates the powerful temporal modeling", + "bbox": [ + 496, + 763, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8676", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2735994c1c4f7aa3ba0a31c6b543e61c369857720f6e3bcc3d6e9e77dddba9d0.jpg", + "image_caption": [ + "Figure 5. Examples of I2V generation results by three methods (VideoComposer, T2V-Zero, and our TRIP) on WebVid-10M dataset. We uniformly sample four frames of each generated video for visualization." + ], + "image_footnote": [], + "bbox": [ + 138, + 70, + 477, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c16510228d637cecab59ab75ff35e778f49b803ebd9b7a4438e73f76ab8e15f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 70, + 828, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b7eb760c442529b6482b1c25a0fb23d5512bf99e40c428205b1aba56888fe6e2.jpg", + "image_caption": [ + "Figure 6. Image-to-video generation results of four different natural scenes by our TRIP on DTDB dataset." + ], + "image_footnote": [], + "bbox": [ + 99, + 405, + 444, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of motion dynamics through our residual-like noise prediction with image noise prior.", + "bbox": [ + 76, + 617, + 467, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study on TRIP", + "text_level": 1, + "bbox": [ + 76, + 662, + 302, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we perform ablation study to delve into the design of TRIP for I2V generation. Here all experiments are conducted on WebVid-10M for performance comparison.", + "bbox": [ + 76, + 686, + 467, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First Frame Condition. We first investigate different condition strategies to exploit the given first frame as additional condition for I2V generation. Table 5 summarizes the performance comparisons among different variants of our TRIP. $\\mathbf{TRIP}_C$ follows the common recipe [60] to concatenate the clean image latent code of first frame with the noise video latent code along channel dimension. $\\mathbf{TRIP}_{TE}$ is implemented by concatenating image latent code at the end of the noise video latent code along temporal dimension, while our proposal (TRIP) temporally concatenates image latent code at the beginning of video latent code. In par", + "bbox": [ + 75, + 734, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bc19f61505e6fea799b69ea2e6a21beba6b0cd072bee8206661b88d795bc2a5f.jpg", + "table_caption": [ + "Table 5. Performance comparisons among different condition approaches on WebVid-10M dataset." + ], + "table_footnote": [], + "table_body": "
ModelF-Consistency4 (↑)F-Consistencyall (↑)FVD (↓)
TRIPC94.7796.1341.3
TRIPTE95.1796.2039.8
TRIP95.3696.4138.9
", + "bbox": [ + 503, + 436, + 890, + 503 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5f06d0f3ffa4cb1722897c11c0c315dc4de492911296b316fe03de8a7a2fb601.jpg", + "table_caption": [ + "Table 6. Evaluation of temporal residual learning in terms of F-Consistency and FVD on WebVid-10M dataset." + ], + "table_footnote": [], + "table_body": "
ModelF-Consistency4 (↑)F-Consistencyall (↑)FVD (↓)
TRIP-94.6695.9239.9
TRIPW95.2295.9643.0
TRIP95.3696.4138.9
", + "bbox": [ + 503, + 542, + 890, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ticular, $\\mathrm{TRIP}_{TE}$ exhibits better performances than $\\mathrm{TRIP}_C$ , which indicates that temporal concatenation is a practical choice for augmenting additional condition in I2V generation. TRIP further outperforms $\\mathrm{TRIP}_{TE}$ across both F-Consistency and FVD metrics. Compared to concatenating image latent code at the end of video, augmenting video latent code with input first frame latent code at the start position basically matches with the normal temporal direction, thereby facilitating temporal modeling.", + "bbox": [ + 496, + 613, + 890, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Temporal Residual Learning. Next, we study how each design of our TRIP influences the performances of both temporal consistency and visual quality for I2V generation. Here we include two additional variants of TRIP: (1) $\\mathbf{TRIP}^{-}$ removes the shortcut path of temporal residual learning and only capitalizes on 3D-UNet for noise prediction; (2) $\\mathbf{TRIP}^{W}$ is a degraded variant of temporal residual learning with image noise prior by simply fusing reference and residual noise via linear fusion strategy. As shown in Table 6, $\\mathbf{TRIP}^{W}$ exhibits better performances than $\\mathbf{TRIP}^{-}$", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8677", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f9067e6bd7cca475750b152828b1dd38a34356cda17a544b0762e20b4a949640.jpg", + "image_caption": [ + "A castle in the middle of a forest with a river running through it" + ], + "image_footnote": [], + "bbox": [ + 138, + 88, + 480, + 142 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b87c206034da2155825c89c1103dc9138285fa045afa37101baadb0eb4a35cc9.jpg", + "image_caption": [ + "Crabs with bubbles crawls on the beach with sunset in the background" + ], + "image_footnote": [], + "bbox": [ + 483, + 88, + 826, + 142 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8e4618d020e042182948f7ada9891254b3774754940ae03467dd3467bd555c8a.jpg", + "image_caption": [ + "A ship sails in the sea." + ], + "image_footnote": [], + "bbox": [ + 138, + 150, + 480, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/76d9677ef005c7c27bef7306a98c182645d2fede00bc546df2e82894ebc36d3c.jpg", + "image_caption": [ + "WALL-E on the beach with a plant in his hand" + ], + "image_footnote": [], + "bbox": [ + 483, + 150, + 826, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e459fa9626d9d84dda509f10dae4efe49028429e38615bbebe7e775d5c11d67.jpg", + "image_caption": [ + "Halloweep pumpkin with glowing eyes and smoke coming out of it" + ], + "image_footnote": [], + "bbox": [ + 138, + 212, + 480, + 266 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8527d2a2d525e666fb73bd1ed5995137a017fab990a4cadd3a63f2a0d6e075da.jpg", + "image_caption": [ + "A bear is playing a guitar in front of a tank in the snow." + ], + "image_footnote": [], + "bbox": [ + 483, + 212, + 826, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/04b4d4e3592b1d6ac4e85f3cebddfc169199d84b6d1fa4b327c3a1faf9158980.jpg", + "image_caption": [ + "Figure 7. Customized image animation with Stable-Diffusion XL [47] for text-to-video generation. We first employ Stable-Diffusion XL to synthesize fancy images and then exploit TRIP for image animation. Six generated videos with corresponding text prompts are presented.", + "Figure 8. Visualization of I2V example with text prompt \"Young male tourist taking a photograph of old city on sunset\" by using variants of our TRIP." + ], + "image_footnote": [], + "bbox": [ + 86, + 311, + 454, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in terms of F-consistency. The results highlight the advantage of leveraging image noise prior as the reference to amplify the alignment between the first frame and subsequent frames. By further upgrading temporal residual learning $(\\mathbf{TRIP}^W)$ with Transformer-based temporal noise fusion, our TRIP achieves the best performances across all metrics. Figure 8 further showcases the results of different variants, which clearly confirm the effectiveness of our designs.", + "bbox": [ + 75, + 512, + 467, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Application: Customized Image Animation", + "text_level": 1, + "bbox": [ + 76, + 648, + 444, + 665 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we provide two interesting applications of customized image animation by directly extending TRIP in a tuning-free manner. The first application is text-to-video pipeline that first leverages text-to-image diffusion models (e.g., Stable-Diffusion XL [47]) to synthesize fancy images and then employs our TRIP to animate them as videos. Figure 7 shows six text-to-video generation examples by using Stable-Diffusion XL and TRIP. The results generally demonstrate the generalization ability of our TRIP when taking text-to-image results as additional reference for text-to-video generation. To further improve controllability of text-to-video generation in real-world scenarios, we can additionally utilize image editing models (e.g., InstructPix2Pix [3] and ControlNet [74]) to modify the visual content of input images and then feed them into TRIP", + "bbox": [ + 75, + 674, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0c9fd19058f83c28b6cf0841be39307611a30aa28e7e9541f1c688cff9f994ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 314, + 869, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c091633088a02362fb475a8f07df76d740e33e2524af34e8bb6b4af3e0bace66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 369, + 869, + 431 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f9144f2263ecb43d5143464ef8c9aaad5aa40da93e6c436284ab0a64fc5e17b.jpg", + "image_caption": [ + "A beautiful duck swims on the pond", + "Figure 9. Customized image animation with image editing models (e.g., InstructPix2Pix [3] and ControlNet [74]). The first row shows image animation results given the primary real-world first image. The second and third rows show I2V results given the edited images via InstructPix2Pix and ControlNet, respectively." + ], + "image_footnote": [], + "bbox": [ + 519, + 431, + 869, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for video generation. As shown in Figure 9, our TRIP generates vivid videos based on the edited images with promising temporal coherence and visual quality.", + "bbox": [ + 498, + 585, + 890, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 647, + 625, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper explores inherent inter-frame correlation in diffusion model for image-to-video generation. Particularly, we study the problem from a novel viewpoint of leveraging temporal residual learning with reference to the image noise prior of given image for enabling coherent temporal modeling. To materialize our idea, we have devised TRIP, which executes image-conditioned noise prediction in a residual fashion. One shortcut path is novelly designed to learn image noise prior based on given image and noised video latent code, leading to reference noise for enhancing alignment between synthesized frames and given image. Meanwhile, another residual path is exploited to estimate the residual noise of each frame through eased temporal modeling. Experiments conducted on three datasets validate the superiority of our proposal over state-of-the-art approaches.", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8678", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. In ICCV, 2021. 5", + "[2] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. In CVPR, 2023. 1, 2, 3, 5", + "[3] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning to Follow Image Editing Instructions. In CVPR, 2023. 1, 2, 8", + "[4] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. StableVideo: Text-driven Consistency-aware Diffusion Video Editing. In ICCV, 2023. 2", + "[5] Zhikai Chen, Fuchen Long, Zhaofan Qiu, Ting Yao, Wengang Zhou, Jiebo Luo, and Tao Mei. AnchorFormer: Point Cloud Completion from Discriminative Nodes. In CVPR, 2023. 2", + "[6] Paul Couairon, Clément Rambour, Jean-Emmanuel Haugeard, and Nicolas Thome. VidEdit: Zero-Shot and Spatially Aware Text-Driven Video Editing. arXiv preprint arXiv:2306.08707, 2023. 2", + "[7] Prafulla Dhariwal and Alex Nichol. Diffusion Models Beat GANs on Image Synthesis. In NeurIPS, 2022. 1", + "[8] Ziyi Dong, Pengxu Wei, and Liang Lin. DreamArtist: Towards Controllable One-Shot Text-to-Image Generation via Positive-Negative Prompt-Tuning. arXiv preprint arXiv:2211.11337, 2022. 2", + "[9] Michael Dorkenwald, Timo Milbich, Andreas Blattmann, Robin Rombach, Konstantinos G. Derpanis, and Björn Ommer. Stochastic Image-to-Video Synthesis using cINNs. In CVPR, 2021. 1, 2, 5, 6", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 2", + "[11] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating Landscape: Self-Supervised Learning of Decoupled Motion and Appearance for Single-Image Video Synthesis. TOG, 2019. 5, 6", + "[12] Siming Fan, Jingtan Piao, Chen Qian, Kwan-Yee Lin, and Hongsheng Li. Simulating Fluids in Real-World Still Images. In ICCV, 2023. 1, 2", + "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashni, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion. In ICLR, 2023. 2", + "[14] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve Your Own Correlation: A Noise Prior for Video Diffusion Models. In ICCV, 2023. 2", + "[15] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. TokenFlow: Consistent Diffusion Features for Consistent Video Editing. arXiv preprint arXiv:2307.10373, 2023. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative Adversarial Networks. In NeurIPS, 2014. 1", + "[17] Xianfan Gu, Chuan Wen, Jiaming Song, and Yang Gao. Seer: Language Instructed Video Prediction with Latent Diffusion Models. arXiv preprint arXiv:2303.14897, 2023. 1, 2, 4", + "[18] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning. arXiv preprint arXiv:2307.04725, 2023. 2", + "[19] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible Diffusion Modeling of Long Videos. In NeurIPS, 2022. 2", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 2", + "[21] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, and Qifeng Chen. animate-A-Story: Storytelling with Retrieval-Augmented Video Generation. arXiv preprint arXiv:2307.06940, 2023. 2", + "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-Prompt Image Editing with Cross-Attention Control. In ICLR, 2023. 1, 2", + "[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. In NeuIPS, 2017. 5", + "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In NeurIPS, 2020. 1", + "[25] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen Video: High Definition Video Generation with Diffusion Models. In CVPR, 2022. 1, 2", + "[26] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video Diffusion Models. In NeurIPS, 2022. 1, 2, 3", + "[27] Aleksander Holynski, Brian Curless, Steven M. Seitz, and Richard Szeliski. Animating Pictures with Eulerian Motion Fields. In CVPR, 2021. 2", + "[28] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. CogVideo: Large-Scale Pretraining for Text-to-Video Generation via Transformers. In ICLR, 2023. 5, 6", + "[29] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely Connected Convolutional Networks. In CVPR, 2017. 2", + "[30] Hadji Isma and Wildes Richard P. A New Large Scale Dynamic Texture Dataset with Application to Convnet Understanding. In ECCV, 2018. 5", + "[31] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. DreamPose: Fashion Image-to-Video Synthesis via Stable Diffusion. In ICCV, 2023. 2", + "[32] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8679", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Navasardyan, and Humphrey Shi. Text2Video-Zero: Text-to-Image Diffusion Models are Zero-Shot Video Generators. In ICCV, 2023. 2, 5", + "[33] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-Concept Customization of Text-to-Image Diffusion. In CVPR, 2023. 2", + "[34] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Flow-Grounded Spatial-Temporal Video Prediction from Still Images. In ECCV, 2018. 2", + "[35] Yehao Li, Ting Yao, Yingwei Pan, and Tao Mei. Contextual Transformer Networks for Visual Recognition. IEEE Trans. on PAMI, 2022. 2", + "[36] Fuchen Long, Zhaofan Qiu, Yingwei Pan, Ting Yao, Jiebo Luo, and Tao Mei. Stand-Alone Inter-Frame Attention in Video Models. In CVPR, 2022.", + "[37] Fuchen Long, Zhaofan Qiu, Yingwei Pan, Ting Yao, Chong-Wah Ngo, and Tao Mei. Dynamic Temporal Filtering in Video Models. In ECCV, 2022. 2", + "[38] Fuchen Long, Zhaofan Qiu, Ting Yao, and Tao Mei. Video-Drafter: Content-Consistent Multi-Scene Video Generation with LLM. arXiv:2401.01256, 2024. 2", + "[39] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. VideoFusion: Decomposed Diffusion Models for High-Quality Video Generation. In CVPR, 2023. 2, 5", + "[40] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable Animation of Fluid Elements in Still Images. In CVPR, 2022. 1, 2", + "[41] Babaeizadeh Mohammad, Finn Chelsea, Erhan Dumitru, Campbell Roy H, and Levine Sergey. Stochastic Variational Video Prediction. In ICLR, 2018. 2", + "[42] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2I-Adapter: Learning Adapters to Dig out More Controllable Ability for Text-to-Image Diffusion Models. arXiv preprint arXiv:2302.08453, 2023. 2", + "[43] Haomiao Ni, Changhao Shi, Kai Li, Sharon X. Huang, and Martin Renqiang Min. Conditional Image-to-Video Generation with Latent Flow Diffusion Models. In CVPR, 2023. 2", + "[44] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. In ICML, 2022. 1, 2, 3", + "[45] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. CoDeF: Content Deformation Fields for Temporally Consistent Video Processing. arXiv preprint arXiv:2308.07926, 2023. 2", + "[46] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot Image-to-Image Translation. In ACM SIGGRAPH, 2023. 2", + "[47] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. arXiv preprint arXiv:2307.01952, 2023. 8" + ], + "bbox": [ + 78, + 90, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 3, 5", + "[49] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv preprint arXiv:2204.06125, 2022.3", + "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022. 1, 2, 5", + "[51] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. In CVPR, 2023. 2", + "[52] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-Video: Text-to-Video Generation without Text-Video Data. In ICLR, 2023. 2, 5, 6", + "[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models. In ICLR, 2021. 5", + "[54] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Palur. A Closer Look at Spatiotemporal Convolutions for Action Recognition. In CVPR, 2018. 2", + "[55] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new Metric for Video Generation. In ICLR Deep-GenStruct Workshop, 2019. 5", + "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. In NeurIPS, 2017. 2", + "[57] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-Of-The-Art Diffusion Models. 5", + "[58] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-L-Video: Multi-Text to Long Video Generation via Temporal Co-Denoising. arXiv preprint arXiv:2305.18264, 2023. 4, 5", + "[59] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. ModelScope Text-to-Video Technical Report. arXiv preprint arXiv:2308.06571, 2023. 2, 5, 6", + "[60] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional Video Synthesis with Motion Controllability. In NeurIPS, 2023. 1, 2, 4, 5, 6, 7", + "[61] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent Image Animator: Learning to Animate Images via Latent Space Navigation. In ICLR, 2022. 2", + "[62] Yaohui Wang, Xin Ma, Xinyuan Chen, Antitza Dantcheva, Bo Dai, and Yu Qiao. LEO: Generative Latent Image" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8680", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Animator for Human Video Synthesis. arXiv preprint arXiv:2305.03989, 2023. 2", + "[63] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-A-Video: One-Shot Tuning of Image Diffusion Models for Text-to-Video Generation. In ICCV, 2023. 5", + "[64] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A Survey on Video Diffusion Models. arXiv preprint arXiv:2310.10647, 2023. 2", + "[65] Zhen Xing, Qi Dai, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. SimDA: Simple Diffusion Adapter for Efficient Video Generation. In CVPR, 2024. 2", + "[66] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to Generate Time-Lapse Videos Using Multi-Stage Dynamic Generative Adversarial Networks. In CVPR, 2018.", + "[67] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A Large Video Description Dataset for Bridging Video and Language. In CVPR, 2016. 5", + "[68] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and Improving Layer Normalization. In NeurIPS, 2019. 5", + "[69] Hanshu Yan, Jun Hao Liew, Long Mai, Shanchuan Lin, and Jiashi Feng. MagicProp: Diffusion-based Video Editing via Motion-aware Appearance Propagation. arXiv preprint arXiv:2309.00908, 2023. 2", + "[70] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion Probabilistic Modeling for Video Generation. arXiv preprint arXiv:2203.09481, 2022. 2", + "[71] Ting Yao, Yingwei Pan, Yehao Li, Chong-Wah Ngo, and Tao Mei. Wave-ViT: Unifying Wavelet and Transformers for Visual Representation Learning. In ECCV, 2022. 2", + "[72] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. DragNUWA: Fine-grained Control in Video Generation by Integrating Text, Image, and Trajectory. arXiv preprint arXiv:2308.08089, 2023. 2", + "[73] Jianfeng Zhang, Hanshu Yan, Zhongcong Xu, Jiashi Feng, and Jun Hao Liew. MagicAvatar: Multimodal Avatar Generation and Animation. arXiv preprint arXiv:2308.14748, 2023. 2", + "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In ICCV, 2023. 2, 8", + "[75] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks. In ICCV, 2017. 1" + ], + "bbox": [ + 78, + 90, + 468, + 771 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8681", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/51572829-1348-4002-a40a-f8702f404317_model.json b/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/51572829-1348-4002-a40a-f8702f404317_model.json new file mode 100644 index 0000000000000000000000000000000000000000..519a57a09dd94d0e4c65387e506d921c5b8cbd30 --- /dev/null +++ b/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/51572829-1348-4002-a40a-f8702f404317_model.json @@ -0,0 +1,2543 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.185, + 0.131, + 0.787, + 0.175 + ], + "angle": 0, + "content": "TRIP: Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models*" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.205, + 0.816, + 0.255 + ], + "angle": 0, + "content": "Zhongwei Zhang†, Fuchen Long§, Yingwei Pan§, Zhaofan Qiu§, Ting Yao§, Yang Cao† and Tao Mei§ \n†University of Science and Technology of China, Hefei, China \n§HiDream.ai Inc." + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.26, + 0.819, + 0.291 + ], + "angle": 0, + "content": "zhwzhang@mail.ustc.edu.cn,{longfuchen, pandy, qiuzhaofan, tiyao}@hidream.ai forrest@ustc.edu.cn, tmei@hidream.ai" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.313, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.474, + 0.798 + ], + "angle": 0, + "content": "Recent advances in text-to-video generation have demonstrated the utility of powerful diffusion models. Nevertheless, the problem is not trivial when shaping diffusion models to animate static image (i.e., image-to-video generation). The difficulty originates from the aspect that the diffusion process of subsequent animated frames should not only preserve the faithful alignment with the given image but also pursue temporal coherence among adjacent frames. To alleviate this, we present TRIP, a new recipe of image-to-video diffusion paradigm that pivots on image noise prior derived from static image to jointly trigger inter-frame relational reasoning and ease the coherent temporal modeling via temporal residual learning. Technically, the image noise prior is first attained through one-step backward diffusion process based on both static image and noised video latent codes. Next, TRIP executes a residual-like dual-path scheme for noise prediction: 1) a shortcut path that directly takes image noise prior as the reference noise of each frame to amplify the alignment between the first frame and subsequent frames; 2) a residual path that employs 3D-UNet over noised video and static image latent codes to enable inter-frame relational reasoning, thereby easing the learning of the residual noise for each frame. Furthermore, both reference and residual noise of each frame are dynamically merged via attention mechanism for final video generation. Extensive experiments on WebVid-10M, DTDB and MSR-VTT datasets demonstrate the effectiveness of our TRIP for image-to-video generation. Please see our project page at https://trip-i2v.github.io/TRIP/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.824, + 0.21, + 0.84 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.849, + 0.47, + 0.88 + ], + "angle": 0, + "content": "In recent years, deep generative models have demonstrated impressive capabilities to create high-quality visual con" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.348, + 0.9 + ], + "angle": 0, + "content": "*This work was performed at HiDream.ai." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.327, + 0.687, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.49, + 0.672, + 0.499 + ], + "angle": 0, + "content": "(a) Independent noise prediction" + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.327, + 0.876, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.487, + 0.852, + 0.505 + ], + "angle": 0, + "content": "(b) Residual-like noise prediction with image noise prior" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.505, + 0.892, + 0.545 + ], + "angle": 0, + "content": "Figure 1. Comparison between (a) independent noise prediction and (b) our residual-like noise prediction with image noise prior in I2V diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.893, + 0.81 + ], + "angle": 0, + "content": "tent. In between, Generative Adversarial Networks (GANs) [16, 75] and diffusion models [7, 24] have brought forward milestone improvement for a series of generative tasks in computer vision field, e.g., text-to-image generation [44, 50], image editing [3, 22], and text-to-video generation [2, 25, 26]. In this work, we are interested in imaged-to-video generation (I2V), i.e., animating a static image by endowing it with motion dynamics, which has great potential real-world applications (e.g., entertainment and creative media). Conventional I2V techniques [9, 12, 40] commonly target for generating video from a natural scene image with random or very coarse motion dynamics (such as clouds moving or fluid flowing). In contrast, our focus is a more challenging text-driven I2V scenario: given the input static image and text prompt, the animated frames should not only faithfully align with the given first frame, but also be temporally coherent and semantically matched with text prompt." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Several recent pioneering practices [17, 60] for I2V is to remould the typical latent diffusion model of text-to-video generation by directly leveraging the given static image as additional condition during diffusion process. Figure 1 (a) illustrates such conventional noise prediction strategy in latent diffusion model. Concretely, by taking the given" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8671" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.529 + ], + "angle": 0, + "content": "static image as the first frame, 2D Variational Auto-encoder (VAE) is first employed to encode the first frame as image latent code. This encoded image latent code is further concatenated with noised video latent code (i.e., a sequence of Gaussian noise) to predict the backward diffusion noise of each subsequent frame via a learnable 3D-UNet. Nevertheless, such independent noise prediction strategy of each frame leaves the inherent relation between the given image and each subsequent frame under-exploited, and is lacking in efficacy of modeling temporal coherence among adjacent frames. As observed in I2V results of Figure 1 (a), both the foreground content (a glass) and the background content (e.g., a table) in the synthesized second frame are completely different from the ones in the given first frame. To alleviate this issue, our work shapes a new way to formulate the noise prediction in I2V diffusion model as temporal residual learning on the basis of amplified guidance of given image (i.e., image noise prior). As illustrated in Figure 1 (b), our unique design is to integrate the typical noise prediction with an additional shortcut path that directly estimates the backward diffusion noise of each frame by solely referring image noise prior. Note that this image noise prior is calculated based on both input image and noised video latent codes, which explicitly models the inherent correlation between the given first frame and each subsequent frame. Meanwhile, such residual-like scheme re-shapes the typical noise prediction via 3D-UNet as residual noise prediction to ease temporal modeling among adjacent frames, thereby leading to more temporally coherent I2V results." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.53, + 0.473, + 0.804 + ], + "angle": 0, + "content": "By materializing the idea of executing image-conditioned noise prediction in a residual manner, we present a novel diffusion model, namely Temporal Residual learning with Image noise Prior (TRIP), to facilitate I2V. Specifically, given the input noised video latent code and the corresponding static image latent code, TRIP performs residual-like noise prediction along two pathways. One is the shortcut path that first achieves image noise prior based on static image and noised video latent codes through one-step backward diffusion process, which is directly regarded as the reference noise of each frame. The other is the residual path which concatenates static image and noised video latent codes along temporal dimension and feeds them into 3D-UNet to learn the residual noise of each frame. Eventually, a Transformer-based temporal noise fusion module is leveraged to dynamically fuse the reference and residual noises of each frame, yielding high-fidelity video that faithfully aligns with given image." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.816, + 0.22, + 0.831 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Text-to-Video Diffusion Models. The great success of diffusion models [44, 50] to generate fancy images based on text prompts has been witnessed in recent years. The foundation of the basic generative diffusion models fur" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.44 + ], + "angle": 0, + "content": "ther encourages the development of customized image synthesis, such as image editing [3, 22, 46] and personalized image generation [8, 13, 33, 42, 51, 74]. Inspired by the impressive results of diffusion models on image generation, a series of text-to-video (T2V) diffusion models [2, 19, 25, 26, 32, 39, 52, 59, 64, 65, 70] start to emerge. VDM [26] is one of the earlier works that remoulds text-to-image (T2I) 2D-UNet architecture with temporal attention for text-to-video synthesis. Later in [2] and [18], the temporal modules (e.g., temporal convolution and self-attention) in diffusion models are solely optimized to emphasize motion learning. To further alleviate the burden on temporal modeling, Ge et al. [14] present a mixed video noise model under the assumption that each frame should correspond to a shared common noise and its own independent noise. Similar recipe is also adopted by VideoFusion [39] which estimates the basic noise and residual noise via two image diffusion models for video denoising. Other works go one step further by applying diffusion models for customized video synthesis, such as video editing [4, 6, 15, 45, 69] and personalized video generation [18, 21, 73]. In this work, we choose the basic T2V diffusion model with 3D-UNet [59] as our backbone for image-to-video generation task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.895, + 0.791 + ], + "angle": 0, + "content": "Image-to-Video Generation. Image-to-Video (I2V) generation is one kind of conditional video synthesis, where the core condition is a given static image. According to the availability of motion cues, I2V approaches can be grouped into two categories: stochastic generation (solely using input image as condition) [27, 34, 41, 66] and conditional generation (using image and other types of conditions) [31, 38, 40, 43, 61]. Early attempts of I2V generation mostly belong to the stochastic models which usually focus on the short motion of fluid elements (e.g., water flow) [12, 27] or human poses [62] for landscape or human video generation. To further enhance the controllability of motion modeling, the conditional models exploit more types of conditions like text prompts [9, 17] in the procedure of video generation. Recent advances start to explore diffusion models [60, 72] for conditional I2V task. One pioneering work (VideoComposer [60]) flexibly composes a video with additional conditions of texts, optical flow or depth mapping sequence for I2V generation. Nevertheless, VideoComposer independently predicts the noise of each frame and leaves the inherent relation between input image and other frames under-exploited, thereby resulting in temporal inconsistency among frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Deep Residual Learning. The effectiveness of learning residual components with additional shortcut connections in deep neural networks [20, 29, 54] has been verified in various vision tasks. Besides typical convolution-based structure, recent advances [10, 56] also integrate the shortcut connection into the emerging transformer-based structure [5, 35-37, 71]. Intuitively, learning residual part with refer" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8672" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.092, + 0.625, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.09, + 0.877, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.256, + 0.892, + 0.368 + ], + "angle": 0, + "content": "Figure 2. An overview of (a) our Temporal Residual learning with Image noise Prior (TRIP), and (b) the conceptual comparison between typical frame-wise denoising and our temporal noise fusion. During training, the input video clip is encoded into a sequence of image latent codes via a pre-trained 2D VAE. The latent code of first frame is then concatenated with noised video latent code as the input of the 3D-UNet. Then, TRIP performs a residual-like noise prediction along two pathways, i.e., shortcut path and residual path. The shortcut path computes the image noise prior as the reference noise based on the first frame latent codes and the noised video latent codes through a one-step backward diffusion. The residual path leverages 3D-UNet to predict the residual noise with reference to the image noise prior. By dynamically merging the reference and residual noise via a Transformer-based Temporal Noise Fusion (TNF) module, the target noise vector for video denoising is attained. Both 3D-UNet and TNF module are jointly optimized with the typical MSE loss for I2V generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.469, + 0.461 + ], + "angle": 0, + "content": "ence to the optimal identity mapping will ease the network optimization. In this work, we capitalize on such principle and formulate noise prediction in I2V diffusion as temporal residual learning on the basis of the image noise prior derived from static image to enhance temporal coherence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.469, + 0.553 + ], + "angle": 0, + "content": "In summary, our work designs a novel image-to-video diffusion paradigm with residual-like noise prediction. The proposed TRIP contributes by studying not only how to excavate the prior knowledge of given image as noise prior to guide temporal modeling, and how to strengthen temporal consistency conditioned on such image noise prior." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.572, + 0.225, + 0.589 + ], + "angle": 0, + "content": "3. Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In this section, we present our Temporal Residual learning with Image noise Prior (TRIP) for I2V generation. Figure 2 illustrates an overview of our TRIP. Given a video clip at training, TRIP first extracts the per-frame latent code via a pre-trained VAE and groups them as a video latent code. The image latent code of the first frame (i.e., the given static image) is regarded as additional condition, which is further concatenated with the noised video latent code along temporal dimension as the input of the 3D-UNet. Next, TRIP executes a residual-like noise prediction along two pathways (i.e., shortcut path and residual path), aiming to trigger inter-frame relational reasoning. In shortcut path, the image noise prior of each frame is calculated based on the correlation between the first frame and subsequent frames through one-step backward diffusion process. In residual path, the residual noise is estimated by employing 3D-UNet over the concatenated noised video latent code plus text prompt. Finally, a transformer-based Temporal Noise Fusion (TNF) module dynamically merges the image noise prior and the residual noise as the predicted noise for video generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.384, + 0.825, + 0.399 + ], + "angle": 0, + "content": "3.1. Preliminaries: Video Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.408, + 0.892, + 0.604 + ], + "angle": 0, + "content": "T2V diffusion models [2, 26] are usually constructed by remoulding T2I diffusion models [44, 49]. A 3D-UNet with text prompt condition is commonly learnt to denoise a randomly sampled noise sequence in Gaussian distribution for video synthesis. Specifically, given an input video clip with \\(N\\) frames, a pre-trained 2D VAE encoder \\(\\mathcal{E}(\\cdot)\\) first extracts latent code of each frame to constitute a latent code sequence \\(\\{z_0^i\\}_{i=1}^N\\). Then, this latent sequence is grouped via concatenation along temporal dimension as a video latent code \\(z_0\\). Next, the Gaussian noise is gradually added to \\(z_0\\) through forward diffusion procedure (length: \\(T\\), noise scheduler: \\(\\{\\beta_t\\}_{t=1}^T\\)). The noised latent code \\(z_t\\) at each time step \\(t\\) is thus formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.613, + 0.891, + 0.63 + ], + "angle": 0, + "content": "\\[\nz _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\epsilon \\sim \\mathcal {N} (0, I), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.638, + 0.892, + 0.744 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_t\\), \\(\\alpha_t = 1 - \\beta_t\\) and \\(\\epsilon\\) is the adding noise vector sampled from standard normal distribution \\(\\mathcal{N}(0, I)\\). 3D-UNet with parameter \\(\\theta\\) aims to estimate the noise \\(\\epsilon\\) based on the multiple inputs (i.e., noised video latent code \\(z_t\\), time step \\(t\\), and text prompt feature \\(c\\) extracted by CLIP [48]). The Mean Square Error (MSE) loss is leveraged as the final objective \\(L\\):" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.753, + 0.891, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {\\epsilon \\sim \\mathcal {N} (0, I), t, c} [ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, c) \\| ^ {2} ]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.779, + 0.891, + 0.839 + ], + "angle": 0, + "content": "In the inference stage, the video latent code \\(\\hat{z}_0\\) is iteratively denoised by estimating noise \\(\\epsilon_{\\theta}\\) in each time step via 3DUNet, and the following 2D VAE decoder \\(\\mathcal{D}(\\cdot)\\) reconstructs each frame based on the video latent code \\(\\hat{z}_0\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.677, + 0.864 + ], + "angle": 0, + "content": "3.2. Image Noise Prior" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Different from typical T2V diffusion model, I2V generation model goes one step further by additionally emphasizing the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8673" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.332 + ], + "angle": 0, + "content": "faithful alignment between the given first frame and the subsequent frames. Most existing I2V techniques [17, 60] control the video synthesis by concatenating the image latent code of first frame with the noised video latent code along channel/ temporal dimension. In this way, the visual information of first frame would be propagated across all frames via the temporal modules (e.g., temporal convolution and self-attention). Nevertheless, these solutions only rely on the temporal modules to calibrate video synthesis and leave the inherent relation between given image and each subsequent frame under-exploited, which is lacking in efficacy of temporal coherence modeling. To alleviate this issue, we formulate the typical noise prediction in I2V generation as temporal residual learning with reference to an image noise prior, targeting for amplifying the alignment between synthesized frames and first static image." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.333, + 0.473, + 0.485 + ], + "angle": 0, + "content": "Formally, given the image latent code \\( z_0^1 \\) of first frame, we concatenate it with the noised video latent code sequence \\( \\{z_t^i\\}_{i=1}^N \\) (\\( t \\) denotes time steps in forward diffusion) along temporal dimension, yielding the conditional input \\( \\{z_0^1, z_t^1, z_t^2, \\dots, z_t^N\\} \\) for 3D-UNet. Meanwhile, we attempt to excavate the image noise prior to reflect the correlation between the first frame latent code \\( z_0^1 \\) and \\( i \\)-th noised frame latent code \\( z_t^i \\). According to Eq. (1), we can reconstruct the \\( i \\)-th frame latent code \\( z_0^i \\) in one-step backward diffusion process as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.491, + 0.47, + 0.526 + ], + "angle": 0, + "content": "\\[\nz _ {0} ^ {i} = \\frac {z _ {t} ^ {i} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} ^ {i}}{\\sqrt {\\bar {\\alpha} _ {t}}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.473, + 0.625 + ], + "angle": 0, + "content": "where \\(\\epsilon_t^i\\) denotes Gaussian noise which is added to the \\(i\\)-th frame latent code. Considering the basic assumption in I2V that all frames in a short video clip are inherently correlated to the first frame, the \\(i\\)-th frame latent code \\(z_0^i\\) can be naturally formulated as the combination of the first frame latent code \\(z_0^1\\) and a residual item \\(\\Delta z^i\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.63, + 0.47, + 0.649 + ], + "angle": 0, + "content": "\\[\nz _ {0} ^ {i} = z _ {0} ^ {1} + \\Delta z ^ {i}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.684 + ], + "angle": 0, + "content": "Next, we construct a variable \\( C_t^i \\) by adding a scale ratio to the residual item \\( \\Delta z^i \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.69, + 0.47, + 0.725 + ], + "angle": 0, + "content": "\\[\nC _ {t} ^ {i} = \\frac {\\sqrt {\\bar {\\alpha} _ {t}} \\Delta z ^ {i}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.73, + 0.355, + 0.745 + ], + "angle": 0, + "content": "Thus Eq. (5) can be re-written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.752, + 0.47, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\Delta z ^ {i} = \\frac {\\sqrt {1 - \\bar {\\alpha} _ {t}} C _ {t} ^ {i}}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.791, + 0.47, + 0.822 + ], + "angle": 0, + "content": "After integrating Eq. (3) and Eq. (6) into Eq. (4), the first frame latent code \\( z_0^1 \\) is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.829, + 0.47, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} z _ {0} ^ {1} = z _ {0} ^ {i} - \\Delta z ^ {i} = \\frac {z _ {t} ^ {i} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} ^ {i}}{\\sqrt {\\bar {\\alpha} _ {t}}} - \\frac {\\sqrt {1 - \\bar {\\alpha} _ {t}} C _ {t} ^ {i}}{\\sqrt {\\bar {\\alpha} _ {t}}} \\\\ = \\frac {z _ {t} ^ {i} - \\sqrt {1 - \\bar {\\alpha} _ {t}} (\\epsilon_ {t} ^ {i} + C _ {t} ^ {i})}{\\sqrt {\\bar {\\alpha} _ {t}}} = \\frac {z _ {t} ^ {i} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} ^ {i \\rightarrow 1}}{\\sqrt {\\bar {\\alpha} _ {t}}}, (7) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.152 + ], + "angle": 0, + "content": "where \\(\\epsilon_t^{i\\to 1}\\) is defined as the image noise prior which can be interpreted as the relationship between the first frame latent code \\(z_0^1\\) and \\(i\\)-th noised frame latent code \\(z_{t}^{i}\\). The image noise prior \\(\\epsilon_t^{i\\rightarrow 1}\\) is thus measured as:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.163, + 0.892, + 0.199 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {t} ^ {i \\rightarrow 1} = \\frac {z _ {t} ^ {i} - \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} ^ {1}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.209, + 0.895, + 0.316 + ], + "angle": 0, + "content": "According to Eq. (8), the latent code \\( z_0^1 \\) of first frame can be directly reconstructed through the one-step backward diffusion by using the image noise prior \\( \\epsilon_t^{i\\to 1} \\) and the noised latent code \\( z_{t}^{i} \\) of \\( i \\)-th frame. In this way, such image noise prior \\( \\epsilon_t^{i\\to 1} \\) acts as a reference noise for Gaussian noise \\( \\epsilon_t^i \\) which is added in \\( i \\)-th frame latent code, when \\( z_0^i \\) is close to \\( z_0^1 \\) (i.e., \\( i \\)-th frame is similar to the given first frame)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.325, + 0.758, + 0.342 + ], + "angle": 0, + "content": "3.3. Temporal Residual Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.892, + 0.53 + ], + "angle": 0, + "content": "Recall that the image noise prior learnt in the shortcut pathway facilitates temporal modeling to amplify the alignment between the first frame and subsequent frames. To further strengthen temporal coherence among all adjacent frames, a residual pathway is designed to estimate the residual noise of each frame through inter-frame relational reasoning. Inspired by exploring knowledge prior in pre-trained 3DUNet for video editing [58], we propose to fine-tune 3DUNet to estimate the residual noise of each frame. Technically, the estimated noise \\(\\tilde{\\epsilon}_t^i\\) of \\(i\\)-th frame is formulated as the combination of the reference noise \\(\\epsilon_t^{i\\rightarrow 1}\\) (i.e., image noise prior) and the estimated residual noise \\(\\Delta \\tilde{\\epsilon}_t^i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.54, + 0.892, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} _ {t} ^ {i} = \\lambda^ {i} \\epsilon_ {t} ^ {i \\rightarrow 1} + (1 - \\lambda^ {i}) \\Delta \\tilde {\\epsilon} _ {t} ^ {i}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.894, + 0.63 + ], + "angle": 0, + "content": "where \\(\\lambda^i\\) is a trade-off parameter. \\(\\Delta \\tilde{\\epsilon}_t^i\\) is learnt by 3D-UNet (parameter: \\(\\theta\\)), and \\(\\epsilon_t^{i\\rightarrow 1}\\) can be computed by Eq. (8) on the fly. Finally, the I2V denoising diffusion objective \\(\\tilde{L}\\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.641, + 0.892, + 0.661 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathcal {L}} = \\mathbb {E} _ {\\epsilon \\sim \\mathcal {N} (0, I), t, c, i} [ \\| \\epsilon_ {t} ^ {i} - \\tilde {\\epsilon} _ {t} ^ {i} \\| ^ {2} ]. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Note that the computation of Eq. (9) is operated through the Temporal Noise Fusion (TNF) module that estimates the backward diffusion noise of each frame. Moreover, since the temporal correlation between \\(i\\)-th frame and the first frame will decrease when increasing frame index \\(i\\) with enlarged timespan, we shape the trade-off parameter \\(\\lambda^i\\) as a linear decay parameter with respect to frame index." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.779, + 0.803 + ], + "angle": 0, + "content": "3.4. Temporal Noise Fusion Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Our residual-like dual-path noise prediction explores the image noise prior as a reference to facilitate temporal modeling among adjacent frames. However, the handcrafted design of TNF module with simple linear fusion (i.e., Eq. (9)) requires a careful tuning of hyper-parameter \\(\\lambda^i\\), leading to a sub-optimal solution. Instead, we devise a new" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8674" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.081, + 0.434, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.241, + 0.462, + 0.256 + ], + "angle": 0, + "content": "Figure 3. Detailed structure of Temporal Noise Fusion module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.47, + 0.307 + ], + "angle": 0, + "content": "Transformer-based temporal noise fusion module to dynamically fuse reference and residual noises, aiming to further exploit the relations in between and boost noise fusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.473, + 0.474 + ], + "angle": 0, + "content": "Figure 3 depicts the structure of Temporal Noise Fusion module. Specifically, given the estimated residual noise \\(\\Delta \\tilde{\\epsilon}_t^i\\), we first leverage the adaptive layer norm operations [68] which is modulated by the time step \\(t\\) for spatial feature normalization. After that, one self-attention layer is employed for feature enhancement, followed by the same adaptive layer normalization operation. Next, by taking the in-process intermedia feature as query and the concatenated feature of \\([\\epsilon_t^{i\\rightarrow 1},\\Delta \\tilde{\\epsilon}_t^i ]\\) as key/value, TNF module achieves the final backward diffusion noise \\(\\tilde{\\epsilon}_{\\theta}(z_{t},t,c,i)\\) through a cross-attention layer:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.481, + 0.47, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} _ {\\theta} \\left(z _ {t}, t, c, i\\right) = \\tilde {\\epsilon} _ {t} ^ {i} = \\varphi \\left(\\Delta \\tilde {\\epsilon} _ {t} ^ {i}, \\epsilon^ {i \\rightarrow 1}, t\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.473, + 0.6 + ], + "angle": 0, + "content": "where \\(\\varphi(\\cdot)\\) denotes the operation of TNF module. Compared to the simple linear fusion strategy for noise estimation, our Transformer-based TNF module sidesteps the hyper-parameter tuning and provides an elegant alternative to dynamically merge the reference and residual noises of each frame to yield high-fidelity videos." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.611, + 0.21, + 0.629 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.285, + 0.654 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Datasets. We empirically verify the merit of our TRIP model on WebVid-10M [1], DTDB [30] and MSR-VTT [67] datasets. Our TRIP is trained over the training set of WebVid-10M, and the validation sets of all three datasets are used for evaluation. The WebVid-10M dataset consists of about \\(10.7M\\) video-caption pairs with a total of \\(57K\\) video hours. \\(5K\\) videos are used for validation and the remaining videos are utilized for training. We further sample 2,048 videos from the validation set of WebVid-10M for evaluation. DTDB contains more than \\(10K\\) dynamic texture videos of natural scenes. We follow the standard protocols in [9] to evaluate models on a video subset derived from 4 categories of scenes (i.e., fires, clouds, vegetation, and waterfall). The texts of each category are used as the text prompt since there is no text description. For the MSR-VTT dataset, there are \\(10K\\) clips in training set and 2,990" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.081, + 0.894, + 0.123 + ], + "angle": 0, + "content": "Table 1. Performances of F-Consistency (F-Consistency4: consistency among the first four frames, F-Consistencyall: consistency among all frames) and FVD on WebVid-10M." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.128, + 0.894, + 0.186 + ], + "angle": 0, + "content": "
ApproachF-Consistency4 (↑)F-Consistencyall (↑)FVD (↓)
T2V-Zero [32]91.5992.15279
VideoComposer [60]88.7892.52231
TRIP95.3696.4138.9
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.195, + 0.892, + 0.24 + ], + "angle": 0, + "content": "clips in validation set. Each video clip is annotated with 20 text descriptions. We adopt the standard setting in [59] to pre-process the video clips in validation set for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.241, + 0.893, + 0.452 + ], + "angle": 0, + "content": "Implementation Details. We implement our TRIP on PyTorch platform with Diffusers [57] library. The 3D-UNet [59] derived from Stable-Diffusion v2.1 [50] is exploited as the video backbone. Each training sample is a 16-frames clip and the sampling rate is 4 fps. We fix the resolution of each frame as \\(256 \\times 256\\), which is centrally cropped from the resized video. The noise scheduler \\(\\{\\beta_t\\}_{t=1}^T\\) is set as linear scheduler (\\(\\beta_1 = 1 \\times 10^{-4}\\) and \\(\\beta_T = 2 \\times 10^{-2}\\)). We set the number of time steps \\(T\\) as 1,000 in model training. The sampling strategy for video generation is DDIM [53] with 50 steps. The 3D-UNet and TNF module are trained with AdamW optimizer (learning rate: \\(2 \\times 10^{-6}\\) for 3D-UNet and \\(2 \\times 10^{-4}\\) for TNF module). All experiments are conducted on 8 NVIDIA A800 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.894, + 0.636 + ], + "angle": 0, + "content": "Evaluation Metrics. Since there is no standard evaluation metric for I2V task, we choose the metrics of frame consistency (F-Consistency) which is commonly adopted in video editing [58, 63] and Frechet Video Distance (FVD) [55] widely used in T2V generation [2, 39] for the evaluation on WebVid-10M. Note that the vision model used to measure F-Consistency is the CLIP ViT-L/14 [48]. Specifically, we report the frame consistency of the first 4 frames (F-Consistency\\(_4\\)) and all the 16 frames (F-Consistency\\(_{all}\\)). For the evaluation on DTDB and MSR-VTT, we follow [9, 59] and report the frame-wise Frechet Inception Distance (FID) [23] and FVD performances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.894, + 0.758 + ], + "angle": 0, + "content": "Human Evaluation. The user study is also designed to assess the temporal coherence, motion fidelity, and visual quality of the generated videos. In particular, we randomly sample 256 testing videos from the validation set in WebVid-10M for evaluation. Through the Amazon MTurk platform, we invite 32 evaluators, and each evaluator is asked to choose the better one from two synthetic videos generated by two different methods given the same inputs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.77, + 0.877, + 0.788 + ], + "angle": 0, + "content": "4.2. Comparisons with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.794, + 0.892, + 0.884 + ], + "angle": 0, + "content": "We compare our TRIP with several state-of-the-art I2V generation methods, e.g., VideoComposer [60] and T2V-Zero [32], on WebVid-10M. For the evaluation on DTDB, we set AL [11] and cINN [9] as baselines. Text-to-Video generation advances, e.g., CogVideo [28] and Make-A-Video [52] are included for comparison on MSR-VTT." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Evaluation on WebVid-10M. Table 1 summarizes per" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8675" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.074, + 0.424, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.251, + 0.47, + 0.278 + ], + "angle": 0, + "content": "Figure 4. Performance comparisons of F-Consistency by using different number of frames on WebVid-10M." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.281, + 0.469, + 0.309 + ], + "angle": 0, + "content": "Table 2. Performances of averaged FID and FVD over four scene categories on DTDB dataset." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.314, + 0.469, + 0.378 + ], + "angle": 0, + "content": "
ApproachZero-shotFID (↓)FVD (↓)
AL [11]No65.1934.2
cINN [9]No31.9451.6
TRIPYes24.8433.9
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.47, + 0.748 + ], + "angle": 0, + "content": "formance comparison on WebVid-10M. Overall, our TRIP consistently leads to performance boosts against existing diffusion-based baselines across different metrics. Specifically, TRIP achieves the F-Consistency4 of 95.36%, which outperforms the best competitor T2V-Zero by 3.77%. The highest frame consistency among the first four frames of our TRIP generally validates the best temporal coherence within the beginning of the whole video. Such results basically demonstrate the advantage of exploring image noise prior as the guidance to strengthen the faithful visual alignment between given image and subsequent frames. To further support our claim, Figure 4 visualizes the frame consistency of each run which is measured among different numbers of frames. The temporal consistencies of TRIP across various frames constantly surpass other baselines. It is worthy to note that TRIP even shows large superiority when only calculating the consistency between the first two frames. This observation again verifies the effectiveness of our residual-like noise prediction paradigm that nicely preserves the alignment with given image. Considering the metric of FVD, TRIP also achieves the best performance, which indicates that the holistic motion dynamics learnt by our TRIP are well-aligned with the ground-truth video data distribution in WebVid-10M." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Figure 5 showcases four I2V generation results across three methods (VideoComposer, T2V-Zero, and our TRIP) based on the same given first frame and text prompt. Generally, compared to the two baselines, our TRIP synthesizes videos with higher quality by nicely preserving the faithful alignment with the given image & text prompt and meanwhile reflecting temporal coherence. For example, T2V-Zero produces frames that are inconsistent with the second prompt \"the ship passes close to the camera\" and VideoComposer generates temporally inconsistent frames for the" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.072, + 0.882, + 0.084 + ], + "angle": 0, + "content": "Table 3. Performances of FID and FVD on MSR-VTT dataset." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.892, + 0.172 + ], + "angle": 0, + "content": "
ApproachModel TypeFID (↓)FVD (↓)
CogVideo [28]T2V23.591294
Make-A-Video [52]T2V13.17-
VideoComposer [60]T2V-580
ModelScopeT2V [59]T2V11.09550
VideoComposer [60]I2V31.29208
TRIPI2V9.6891.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.188, + 0.892, + 0.215 + ], + "angle": 0, + "content": "Table 4. Human evaluation of the preference ratios between TRIP v.s. other approaches on WebVid-10M dataset." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.22, + 0.892, + 0.282 + ], + "angle": 0, + "content": "
Evaluation Itemsv.s. T2V-Zerov.s. VideoComposer
Temporal Coherence96.9 v.s. 3.184.4 v.s. 15.6
Motion Fidelity93.8 v.s. 6.281.3 v.s. 18.7
Visual Quality90.6 v.s. 9.487.5 v.s. 12.5
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.367 + ], + "angle": 0, + "content": "last prompt \"the portrait of the girl in glasses\". In contrast, the synthesized videos of our TRIP are more temporally coherent and better aligned with inputs. This again confirms the merit of temporal residual learning with amplified guidance of image noise prior for I2V generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.893, + 0.609 + ], + "angle": 0, + "content": "Evaluation on DTDB. Next, we evaluate the generalization ability of our TRIP on DTDB dataset in a zero-shot manner. Table 2 lists the performances of averaged FID and FVD over four scene-related categories. In general, our TRIP outperforms two strong stochastic models (AL and cINN) over both two metrics. Note that cINN also exploits the residual information for I2V generation, but it combines image and residual features as clip feature by learning a bijective mapping. Our TRIP is fundamentally different in that we integrate residual-like noise estimation into the video denoising procedure to ease temporal modeling in the diffusion model. Figure 6 shows the I2V generation results of four different scenes produced by our TRIP. As shown in this figure, TRIP manages to be nicely generalized to scene-related I2V generation task and synthesizes realistic videos conditioned on the given first frame plus class label." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Evaluation on MSR-VTT. We further compare our TRIP with recent video generation advances on MSR-VTT. Note that here we also evaluate TRIP in a zero-shot manner. Table 3 details the frame-wise FID and FVD of different runs. Similar to the observation on DTDB, our TRIP again surpasses all baselines over both metrics. The results basically validate the strong generalization ability of our proposal to formulate the open-domain motion dynamics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.683, + 0.756 + ], + "angle": 0, + "content": "4.3. Human Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In addition to the evaluation over automatic metrics, we also perform human evaluation to investigate user preference with regard to three perspectives (i.e., temporal coherence, motion fidelity, and visual quality) across different I2V approaches. Table 4 shows the comparisons of user preference on the generated videos over WebVid-10M dataset. Overall, our TRIP is clearly the winner in terms of all three criteria compared to both baselines (VideoComposer and T2V-Zero). This demonstrates the powerful temporal modeling" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8676" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.14, + 0.071, + 0.478, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.071, + 0.829, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.37, + 0.892, + 0.399 + ], + "angle": 0, + "content": "Figure 5. Examples of I2V generation results by three methods (VideoComposer, T2V-Zero, and our TRIP) on WebVid-10M dataset. We uniformly sample four frames of each generated video for visualization." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.406, + 0.446, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.579, + 0.468, + 0.606 + ], + "angle": 0, + "content": "Figure 6. Image-to-video generation results of four different natural scenes by our TRIP on DTDB dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.618, + 0.468, + 0.648 + ], + "angle": 0, + "content": "of motion dynamics through our residual-like noise prediction with image noise prior." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.663, + 0.303, + 0.679 + ], + "angle": 0, + "content": "4.4. Ablation Study on TRIP" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.688, + 0.468, + 0.734 + ], + "angle": 0, + "content": "In this section, we perform ablation study to delve into the design of TRIP for I2V generation. Here all experiments are conducted on WebVid-10M for performance comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.901 + ], + "angle": 0, + "content": "First Frame Condition. We first investigate different condition strategies to exploit the given first frame as additional condition for I2V generation. Table 5 summarizes the performance comparisons among different variants of our TRIP. \\(\\mathbf{TRIP}_C\\) follows the common recipe [60] to concatenate the clean image latent code of first frame with the noise video latent code along channel dimension. \\(\\mathbf{TRIP}_{TE}\\) is implemented by concatenating image latent code at the end of the noise video latent code along temporal dimension, while our proposal (TRIP) temporally concatenates image latent code at the beginning of video latent code. In par" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.406, + 0.891, + 0.433 + ], + "angle": 0, + "content": "Table 5. Performance comparisons among different condition approaches on WebVid-10M dataset." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.438, + 0.892, + 0.504 + ], + "angle": 0, + "content": "
ModelF-Consistency4 (↑)F-Consistencyall (↑)FVD (↓)
TRIPC94.7796.1341.3
TRIPTE95.1796.2039.8
TRIP95.3696.4138.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.51, + 0.891, + 0.537 + ], + "angle": 0, + "content": "Table 6. Evaluation of temporal residual learning in terms of F-Consistency and FVD on WebVid-10M dataset." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.543, + 0.892, + 0.61 + ], + "angle": 0, + "content": "
ModelF-Consistency4 (↑)F-Consistencyall (↑)FVD (↓)
TRIP-94.6695.9239.9
TRIPW95.2295.9643.0
TRIP95.3696.4138.9
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.75 + ], + "angle": 0, + "content": "ticular, \\(\\mathrm{TRIP}_{TE}\\) exhibits better performances than \\(\\mathrm{TRIP}_C\\), which indicates that temporal concatenation is a practical choice for augmenting additional condition in I2V generation. TRIP further outperforms \\(\\mathrm{TRIP}_{TE}\\) across both F-Consistency and FVD metrics. Compared to concatenating image latent code at the end of video, augmenting video latent code with input first frame latent code at the start position basically matches with the normal temporal direction, thereby facilitating temporal modeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Temporal Residual Learning. Next, we study how each design of our TRIP influences the performances of both temporal consistency and visual quality for I2V generation. Here we include two additional variants of TRIP: (1) \\(\\mathbf{TRIP}^{-}\\) removes the shortcut path of temporal residual learning and only capitalizes on 3D-UNet for noise prediction; (2) \\(\\mathbf{TRIP}^{W}\\) is a degraded variant of temporal residual learning with image noise prior by simply fusing reference and residual noise via linear fusion strategy. As shown in Table 6, \\(\\mathbf{TRIP}^{W}\\) exhibits better performances than \\(\\mathbf{TRIP}^{-}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8677" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.089, + 0.481, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.204, + 0.143, + 0.416, + 0.15 + ], + "angle": 0, + "content": "A castle in the middle of a forest with a river running through it" + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.089, + 0.828, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.143, + 0.774, + 0.15 + ], + "angle": 0, + "content": "Crabs with bubbles crawls on the beach with sunset in the background" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.151, + 0.481, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.205, + 0.345, + 0.212 + ], + "angle": 0, + "content": "A ship sails in the sea." + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.151, + 0.828, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.205, + 0.735, + 0.212 + ], + "angle": 0, + "content": "WALL-E on the beach with a plant in his hand" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.213, + 0.481, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.267, + 0.414, + 0.275 + ], + "angle": 0, + "content": "Halloweep pumpkin with glowing eyes and smoke coming out of it" + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.213, + 0.828, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.267, + 0.746, + 0.275 + ], + "angle": 0, + "content": "A bear is playing a guitar in front of a tank in the snow." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.279, + 0.892, + 0.307 + ], + "angle": 0, + "content": "Figure 7. Customized image animation with Stable-Diffusion XL [47] for text-to-video generation. We first employ Stable-Diffusion XL to synthesize fancy images and then exploit TRIP for image animation. Six generated videos with corresponding text prompts are presented." + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.313, + 0.455, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.459, + 0.468, + 0.499 + ], + "angle": 0, + "content": "Figure 8. Visualization of I2V example with text prompt \"Young male tourist taking a photograph of old city on sunset\" by using variants of our TRIP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.468, + 0.634 + ], + "angle": 0, + "content": "in terms of F-consistency. The results highlight the advantage of leveraging image noise prior as the reference to amplify the alignment between the first frame and subsequent frames. By further upgrading temporal residual learning \\((\\mathbf{TRIP}^W)\\) with Transformer-based temporal noise fusion, our TRIP achieves the best performances across all metrics. Figure 8 further showcases the results of different variants, which clearly confirm the effectiveness of our designs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.649, + 0.445, + 0.666 + ], + "angle": 0, + "content": "4.5. Application: Customized Image Animation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In this section, we provide two interesting applications of customized image animation by directly extending TRIP in a tuning-free manner. The first application is text-to-video pipeline that first leverages text-to-image diffusion models (e.g., Stable-Diffusion XL [47]) to synthesize fancy images and then employs our TRIP to animate them as videos. Figure 7 shows six text-to-video generation examples by using Stable-Diffusion XL and TRIP. The results generally demonstrate the generalization ability of our TRIP when taking text-to-image results as additional reference for text-to-video generation. To further improve controllability of text-to-video generation in real-world scenarios, we can additionally utilize image editing models (e.g., InstructPix2Pix [3] and ControlNet [74]) to modify the visual content of input images and then feed them into TRIP" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.315, + 0.87, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.371, + 0.87, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.432, + 0.87, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.493, + 0.771, + 0.502 + ], + "angle": 0, + "content": "A beautiful duck swims on the pond" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.506, + 0.892, + 0.576 + ], + "angle": 0, + "content": "Figure 9. Customized image animation with image editing models (e.g., InstructPix2Pix [3] and ControlNet [74]). The first row shows image animation results given the primary real-world first image. The second and third rows show I2V results given the edited images via InstructPix2Pix and ControlNet, respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.587, + 0.891, + 0.632 + ], + "angle": 0, + "content": "for video generation. As shown in Figure 9, our TRIP generates vivid videos based on the edited images with promising temporal coherence and visual quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.648, + 0.627, + 0.664 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This paper explores inherent inter-frame correlation in diffusion model for image-to-video generation. Particularly, we study the problem from a novel viewpoint of leveraging temporal residual learning with reference to the image noise prior of given image for enabling coherent temporal modeling. To materialize our idea, we have devised TRIP, which executes image-conditioned noise prediction in a residual fashion. One shortcut path is novelly designed to learn image noise prior based on given image and noised video latent code, leading to reference noise for enhancing alignment between synthesized frames and given image. Meanwhile, another residual path is exploited to estimate the residual noise of each frame through eased temporal modeling. Experiments conducted on three datasets validate the superiority of our proposal over state-of-the-art approaches." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8678" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.471, + 0.212 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. In CVPR, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.255 + ], + "angle": 0, + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning to Follow Image Editing Instructions. In CVPR, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.298 + ], + "angle": 0, + "content": "[4] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. StableVideo: Text-driven Consistency-aware Diffusion Video Editing. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.299, + 0.471, + 0.352 + ], + "angle": 0, + "content": "[5] Zhikai Chen, Fuchen Long, Zhaofan Qiu, Ting Yao, Wengang Zhou, Jiebo Luo, and Tao Mei. AnchorFormer: Point Cloud Completion from Discriminative Nodes. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.471, + 0.409 + ], + "angle": 0, + "content": "[6] Paul Couairon, Clément Rambour, Jean-Emmanuel Haugeard, and Nicolas Thome. VidEdit: Zero-Shot and Spatially Aware Text-Driven Video Editing. arXiv preprint arXiv:2306.08707, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.471, + 0.438 + ], + "angle": 0, + "content": "[7] Prafulla Dhariwal and Alex Nichol. Diffusion Models Beat GANs on Image Synthesis. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[8] Ziyi Dong, Pengxu Wei, and Liang Lin. DreamArtist: Towards Controllable One-Shot Text-to-Image Generation via Positive-Negative Prompt-Tuning. arXiv preprint arXiv:2211.11337, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[9] Michael Dorkenwald, Timo Milbich, Andreas Blattmann, Robin Rombach, Konstantinos G. Derpanis, and Björn Ommer. Stochastic Image-to-Video Synthesis using cINNs. In CVPR, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.551, + 0.47, + 0.633 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.47, + 0.689 + ], + "angle": 0, + "content": "[11] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating Landscape: Self-Supervised Learning of Decoupled Motion and Appearance for Single-Image Video Synthesis. TOG, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.47, + 0.733 + ], + "angle": 0, + "content": "[12] Siming Fan, Jingtan Piao, Chen Qian, Kwan-Yee Lin, and Hongsheng Li. Simulating Fluids in Real-World Still Images. In ICCV, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashni, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.856 + ], + "angle": 0, + "content": "[14] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve Your Own Correlation: A Noise Prior for Video Diffusion Models. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. TokenFlow: Consistent Diffusion Features for Consistent Video Editing. arXiv preprint arXiv:2307.10373, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[16] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative Adversarial Networks. In NeurIPS, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[17] Xianfan Gu, Chuan Wen, Jiaming Song, and Yang Gao. Seer: Language Instructed Video Prediction with Latent Diffusion Models. arXiv preprint arXiv:2303.14897, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[18] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning. arXiv preprint arXiv:2307.04725, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[19] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible Diffusion Modeling of Long Videos. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[21] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, and Qifeng Chen. animate-A-Story: Storytelling with Retrieval-Augmented Video Generation. arXiv preprint arXiv:2307.06940, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-Prompt Image Editing with Cross-Attention Control. In ICLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. In NeuIPS, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In NeurIPS, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[25] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen Video: High Definition Video Generation with Diffusion Models. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[26] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video Diffusion Models. In NeurIPS, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[27] Aleksander Holynski, Brian Curless, Steven M. Seitz, and Richard Szeliski. Animating Pictures with Eulerian Motion Fields. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[28] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. CogVideo: Large-Scale Pretraining for Text-to-Video Generation via Transformers. In ICLR, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[29] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely Connected Convolutional Networks. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[30] Hadji Isma and Wildes Richard P. A New Large Scale Dynamic Texture Dataset with Application to Convnet Understanding. In ECCV, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[31] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. DreamPose: Fashion Image-to-Video Synthesis via Stable Diffusion. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8679" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.133 + ], + "angle": 0, + "content": "Navasardyan, and Humphrey Shi. Text2Video-Zero: Text-to-Image Diffusion Models are Zero-Shot Video Generators. In ICCV, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.468, + 0.175 + ], + "angle": 0, + "content": "[33] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-Concept Customization of Text-to-Image Diffusion. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.468, + 0.218 + ], + "angle": 0, + "content": "[34] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Flow-Grounded Spatial-Temporal Video Prediction from Still Images. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.468, + 0.259 + ], + "angle": 0, + "content": "[35] Yehao Li, Ting Yao, Yingwei Pan, and Tao Mei. Contextual Transformer Networks for Visual Recognition. IEEE Trans. on PAMI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.468, + 0.3 + ], + "angle": 0, + "content": "[36] Fuchen Long, Zhaofan Qiu, Yingwei Pan, Ting Yao, Jiebo Luo, and Tao Mei. Stand-Alone Inter-Frame Attention in Video Models. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.468, + 0.342 + ], + "angle": 0, + "content": "[37] Fuchen Long, Zhaofan Qiu, Yingwei Pan, Ting Yao, Chong-Wah Ngo, and Tao Mei. Dynamic Temporal Filtering in Video Models. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.468, + 0.384 + ], + "angle": 0, + "content": "[38] Fuchen Long, Zhaofan Qiu, Ting Yao, and Tao Mei. Video-Drafter: Content-Consistent Multi-Scene Video Generation with LLM. arXiv:2401.01256, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.468, + 0.441 + ], + "angle": 0, + "content": "[39] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. VideoFusion: Decomposed Diffusion Models for High-Quality Video Generation. In CVPR, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.468, + 0.481 + ], + "angle": 0, + "content": "[40] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable Animation of Fluid Elements in Still Images. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.468, + 0.523 + ], + "angle": 0, + "content": "[41] Babaeizadeh Mohammad, Finn Chelsea, Erhan Dumitru, Campbell Roy H, and Levine Sergey. Stochastic Variational Video Prediction. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.468, + 0.593 + ], + "angle": 0, + "content": "[42] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2I-Adapter: Learning Adapters to Dig out More Controllable Ability for Text-to-Image Diffusion Models. arXiv preprint arXiv:2302.08453, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.468, + 0.648 + ], + "angle": 0, + "content": "[43] Haomiao Ni, Changhao Shi, Kai Li, Sharon X. Huang, and Martin Renqiang Min. Conditional Image-to-Video Generation with Latent Flow Diffusion Models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.468, + 0.718 + ], + "angle": 0, + "content": "[44] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. In ICML, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.468, + 0.788 + ], + "angle": 0, + "content": "[45] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. CoDeF: Content Deformation Fields for Temporally Consistent Video Processing. arXiv preprint arXiv:2308.07926, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.468, + 0.83 + ], + "angle": 0, + "content": "[46] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot Image-to-Image Translation. In ACM SIGGRAPH, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[47] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. arXiv preprint arXiv:2307.01952, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.176, + 0.892, + 0.23 + ], + "angle": 0, + "content": "[49] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv preprint arXiv:2204.06125, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.892, + 0.285 + ], + "angle": 0, + "content": "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.287, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[51] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.411 + ], + "angle": 0, + "content": "[52] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-Video: Text-to-Video Generation without Text-Video Data. In ICLR, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models. In ICLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.493 + ], + "angle": 0, + "content": "[54] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Palur. A Closer Look at Spatiotemporal Convolutions for Action Recognition. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.496, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[55] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new Metric for Video Generation. In ICLR Deep-GenStruct Workshop, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[57] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-Of-The-Art Diffusion Models. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[58] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-L-Video: Multi-Text to Long Video Generation via Temporal Co-Denoising. arXiv preprint arXiv:2305.18264, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[59] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. ModelScope Text-to-Video Technical Report. arXiv preprint arXiv:2308.06571, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[60] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional Video Synthesis with Motion Controllability. In NeurIPS, 2023. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[61] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent Image Animator: Learning to Animate Images via Latent Space Navigation. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[62] Yaohui Wang, Xin Ma, Xinyuan Chen, Antitza Dantcheva, Bo Dai, and Yu Qiao. LEO: Generative Latent Image" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8680" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Animator for Human Video Synthesis. arXiv preprint arXiv:2305.03989, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[63] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-A-Video: One-Shot Tuning of Image Diffusion Models for Text-to-Video Generation. In ICCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.246 + ], + "angle": 0, + "content": "[64] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A Survey on Video Diffusion Models. arXiv preprint arXiv:2310.10647, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[65] Zhen Xing, Qi Dai, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. SimDA: Simple Diffusion Adapter for Efficient Video Generation. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[66] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to Generate Time-Lapse Videos Using Multi-Stage Dynamic Generative Adversarial Networks. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[67] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A Large Video Description Dataset for Bridging Video and Language. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[68] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and Improving Layer Normalization. In NeurIPS, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[69] Hanshu Yan, Jun Hao Liew, Long Mai, Shanchuan Lin, and Jiashi Feng. MagicProp: Diffusion-based Video Editing via Motion-aware Appearance Propagation. arXiv preprint arXiv:2309.00908, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[70] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion Probabilistic Modeling for Video Generation. arXiv preprint arXiv:2203.09481, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[71] Ting Yao, Yingwei Pan, Yehao Li, Chong-Wah Ngo, and Tao Mei. Wave-ViT: Unifying Wavelet and Transformers for Visual Representation Learning. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[72] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. DragNUWA: Fine-grained Control in Video Generation by Integrating Text, Image, and Trajectory. arXiv preprint arXiv:2308.08089, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[73] Jianfeng Zhang, Hanshu Yan, Zhongcong Xu, Jiashi Feng, and Jun Hao Liew. MagicAvatar: Multimodal Avatar Generation and Animation. arXiv preprint arXiv:2308.14748, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In ICCV, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[75] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks. In ICCV, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.772 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8681" + } + ] +] \ No newline at end of file diff --git a/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/51572829-1348-4002-a40a-f8702f404317_origin.pdf b/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/51572829-1348-4002-a40a-f8702f404317_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..55700d346c58bbf5cd6236e172d731b0996d955c --- /dev/null +++ b/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/51572829-1348-4002-a40a-f8702f404317_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adfccd378321b9ddda8e306a5879c0052038235f8c927097f6e10cf2b0aac137 +size 3656538 diff --git a/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/full.md b/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..34264aee667c6fc619b5d7b774c4307372464007 --- /dev/null +++ b/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/full.md @@ -0,0 +1,353 @@ +# TRIP: Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models* + +Zhongwei Zhang†, Fuchen Long§, Yingwei Pan§, Zhaofan Qiu§, Ting Yao§, Yang Cao† and Tao Mei§ +†University of Science and Technology of China, Hefei, China +§HiDream.ai Inc. + +zhwzhang@mail.ustc.edu.cn,{longfuchen, pandy, qiuzhaofan, tiyao}@hidream.ai forrest@ustc.edu.cn, tmei@hidream.ai + +# Abstract + +Recent advances in text-to-video generation have demonstrated the utility of powerful diffusion models. Nevertheless, the problem is not trivial when shaping diffusion models to animate static image (i.e., image-to-video generation). The difficulty originates from the aspect that the diffusion process of subsequent animated frames should not only preserve the faithful alignment with the given image but also pursue temporal coherence among adjacent frames. To alleviate this, we present TRIP, a new recipe of image-to-video diffusion paradigm that pivots on image noise prior derived from static image to jointly trigger inter-frame relational reasoning and ease the coherent temporal modeling via temporal residual learning. Technically, the image noise prior is first attained through one-step backward diffusion process based on both static image and noised video latent codes. Next, TRIP executes a residual-like dual-path scheme for noise prediction: 1) a shortcut path that directly takes image noise prior as the reference noise of each frame to amplify the alignment between the first frame and subsequent frames; 2) a residual path that employs 3D-UNet over noised video and static image latent codes to enable inter-frame relational reasoning, thereby easing the learning of the residual noise for each frame. Furthermore, both reference and residual noise of each frame are dynamically merged via attention mechanism for final video generation. Extensive experiments on WebVid-10M, DTDB and MSR-VTT datasets demonstrate the effectiveness of our TRIP for image-to-video generation. Please see our project page at https://trip-i2v.github.io/TRIP/. + +# 1. Introduction + +In recent years, deep generative models have demonstrated impressive capabilities to create high-quality visual con + +![](images/983675b4f8017e056172a5a2c80320ab12aa4c9f5a2e3d5ae88402bca7bde2e9.jpg) +(a) Independent noise prediction +Figure 1. Comparison between (a) independent noise prediction and (b) our residual-like noise prediction with image noise prior in I2V diffusion models. + +![](images/dfad6fbaebb001f86268039a5a504716ef7e2de6923644bf2a57e1eae95db71e.jpg) +(b) Residual-like noise prediction with image noise prior + +tent. In between, Generative Adversarial Networks (GANs) [16, 75] and diffusion models [7, 24] have brought forward milestone improvement for a series of generative tasks in computer vision field, e.g., text-to-image generation [44, 50], image editing [3, 22], and text-to-video generation [2, 25, 26]. In this work, we are interested in imaged-to-video generation (I2V), i.e., animating a static image by endowing it with motion dynamics, which has great potential real-world applications (e.g., entertainment and creative media). Conventional I2V techniques [9, 12, 40] commonly target for generating video from a natural scene image with random or very coarse motion dynamics (such as clouds moving or fluid flowing). In contrast, our focus is a more challenging text-driven I2V scenario: given the input static image and text prompt, the animated frames should not only faithfully align with the given first frame, but also be temporally coherent and semantically matched with text prompt. + +Several recent pioneering practices [17, 60] for I2V is to remould the typical latent diffusion model of text-to-video generation by directly leveraging the given static image as additional condition during diffusion process. Figure 1 (a) illustrates such conventional noise prediction strategy in latent diffusion model. Concretely, by taking the given + +static image as the first frame, 2D Variational Auto-encoder (VAE) is first employed to encode the first frame as image latent code. This encoded image latent code is further concatenated with noised video latent code (i.e., a sequence of Gaussian noise) to predict the backward diffusion noise of each subsequent frame via a learnable 3D-UNet. Nevertheless, such independent noise prediction strategy of each frame leaves the inherent relation between the given image and each subsequent frame under-exploited, and is lacking in efficacy of modeling temporal coherence among adjacent frames. As observed in I2V results of Figure 1 (a), both the foreground content (a glass) and the background content (e.g., a table) in the synthesized second frame are completely different from the ones in the given first frame. To alleviate this issue, our work shapes a new way to formulate the noise prediction in I2V diffusion model as temporal residual learning on the basis of amplified guidance of given image (i.e., image noise prior). As illustrated in Figure 1 (b), our unique design is to integrate the typical noise prediction with an additional shortcut path that directly estimates the backward diffusion noise of each frame by solely referring image noise prior. Note that this image noise prior is calculated based on both input image and noised video latent codes, which explicitly models the inherent correlation between the given first frame and each subsequent frame. Meanwhile, such residual-like scheme re-shapes the typical noise prediction via 3D-UNet as residual noise prediction to ease temporal modeling among adjacent frames, thereby leading to more temporally coherent I2V results. + +By materializing the idea of executing image-conditioned noise prediction in a residual manner, we present a novel diffusion model, namely Temporal Residual learning with Image noise Prior (TRIP), to facilitate I2V. Specifically, given the input noised video latent code and the corresponding static image latent code, TRIP performs residual-like noise prediction along two pathways. One is the shortcut path that first achieves image noise prior based on static image and noised video latent codes through one-step backward diffusion process, which is directly regarded as the reference noise of each frame. The other is the residual path which concatenates static image and noised video latent codes along temporal dimension and feeds them into 3D-UNet to learn the residual noise of each frame. Eventually, a Transformer-based temporal noise fusion module is leveraged to dynamically fuse the reference and residual noises of each frame, yielding high-fidelity video that faithfully aligns with given image. + +# 2. Related Work + +Text-to-Video Diffusion Models. The great success of diffusion models [44, 50] to generate fancy images based on text prompts has been witnessed in recent years. The foundation of the basic generative diffusion models fur + +ther encourages the development of customized image synthesis, such as image editing [3, 22, 46] and personalized image generation [8, 13, 33, 42, 51, 74]. Inspired by the impressive results of diffusion models on image generation, a series of text-to-video (T2V) diffusion models [2, 19, 25, 26, 32, 39, 52, 59, 64, 65, 70] start to emerge. VDM [26] is one of the earlier works that remoulds text-to-image (T2I) 2D-UNet architecture with temporal attention for text-to-video synthesis. Later in [2] and [18], the temporal modules (e.g., temporal convolution and self-attention) in diffusion models are solely optimized to emphasize motion learning. To further alleviate the burden on temporal modeling, Ge et al. [14] present a mixed video noise model under the assumption that each frame should correspond to a shared common noise and its own independent noise. Similar recipe is also adopted by VideoFusion [39] which estimates the basic noise and residual noise via two image diffusion models for video denoising. Other works go one step further by applying diffusion models for customized video synthesis, such as video editing [4, 6, 15, 45, 69] and personalized video generation [18, 21, 73]. In this work, we choose the basic T2V diffusion model with 3D-UNet [59] as our backbone for image-to-video generation task. + +Image-to-Video Generation. Image-to-Video (I2V) generation is one kind of conditional video synthesis, where the core condition is a given static image. According to the availability of motion cues, I2V approaches can be grouped into two categories: stochastic generation (solely using input image as condition) [27, 34, 41, 66] and conditional generation (using image and other types of conditions) [31, 38, 40, 43, 61]. Early attempts of I2V generation mostly belong to the stochastic models which usually focus on the short motion of fluid elements (e.g., water flow) [12, 27] or human poses [62] for landscape or human video generation. To further enhance the controllability of motion modeling, the conditional models exploit more types of conditions like text prompts [9, 17] in the procedure of video generation. Recent advances start to explore diffusion models [60, 72] for conditional I2V task. One pioneering work (VideoComposer [60]) flexibly composes a video with additional conditions of texts, optical flow or depth mapping sequence for I2V generation. Nevertheless, VideoComposer independently predicts the noise of each frame and leaves the inherent relation between input image and other frames under-exploited, thereby resulting in temporal inconsistency among frames. + +Deep Residual Learning. The effectiveness of learning residual components with additional shortcut connections in deep neural networks [20, 29, 54] has been verified in various vision tasks. Besides typical convolution-based structure, recent advances [10, 56] also integrate the shortcut connection into the emerging transformer-based structure [5, 35-37, 71]. Intuitively, learning residual part with refer + +![](images/067e4d0946162a94f0a173b77564d0ad26b672674905f7e58a4e7b1e6cdfd349.jpg) +Figure 2. An overview of (a) our Temporal Residual learning with Image noise Prior (TRIP), and (b) the conceptual comparison between typical frame-wise denoising and our temporal noise fusion. During training, the input video clip is encoded into a sequence of image latent codes via a pre-trained 2D VAE. The latent code of first frame is then concatenated with noised video latent code as the input of the 3D-UNet. Then, TRIP performs a residual-like noise prediction along two pathways, i.e., shortcut path and residual path. The shortcut path computes the image noise prior as the reference noise based on the first frame latent codes and the noised video latent codes through a one-step backward diffusion. The residual path leverages 3D-UNet to predict the residual noise with reference to the image noise prior. By dynamically merging the reference and residual noise via a Transformer-based Temporal Noise Fusion (TNF) module, the target noise vector for video denoising is attained. Both 3D-UNet and TNF module are jointly optimized with the typical MSE loss for I2V generation. + +![](images/22e148f9973d9fd46f9a651a29b2ea155ec091d6223a2912e1c7d6cb68444910.jpg) + +ence to the optimal identity mapping will ease the network optimization. In this work, we capitalize on such principle and formulate noise prediction in I2V diffusion as temporal residual learning on the basis of the image noise prior derived from static image to enhance temporal coherence. + +In summary, our work designs a novel image-to-video diffusion paradigm with residual-like noise prediction. The proposed TRIP contributes by studying not only how to excavate the prior knowledge of given image as noise prior to guide temporal modeling, and how to strengthen temporal consistency conditioned on such image noise prior. + +# 3. Our Approach + +In this section, we present our Temporal Residual learning with Image noise Prior (TRIP) for I2V generation. Figure 2 illustrates an overview of our TRIP. Given a video clip at training, TRIP first extracts the per-frame latent code via a pre-trained VAE and groups them as a video latent code. The image latent code of the first frame (i.e., the given static image) is regarded as additional condition, which is further concatenated with the noised video latent code along temporal dimension as the input of the 3D-UNet. Next, TRIP executes a residual-like noise prediction along two pathways (i.e., shortcut path and residual path), aiming to trigger inter-frame relational reasoning. In shortcut path, the image noise prior of each frame is calculated based on the correlation between the first frame and subsequent frames through one-step backward diffusion process. In residual path, the residual noise is estimated by employing 3D-UNet over the concatenated noised video latent code plus text prompt. Finally, a transformer-based Temporal Noise Fusion (TNF) module dynamically merges the image noise prior and the residual noise as the predicted noise for video generation. + +# 3.1. Preliminaries: Video Diffusion Model + +T2V diffusion models [2, 26] are usually constructed by remoulding T2I diffusion models [44, 49]. A 3D-UNet with text prompt condition is commonly learnt to denoise a randomly sampled noise sequence in Gaussian distribution for video synthesis. Specifically, given an input video clip with $N$ frames, a pre-trained 2D VAE encoder $\mathcal{E}(\cdot)$ first extracts latent code of each frame to constitute a latent code sequence $\{z_0^i\}_{i=1}^N$ . Then, this latent sequence is grouped via concatenation along temporal dimension as a video latent code $z_0$ . Next, the Gaussian noise is gradually added to $z_0$ through forward diffusion procedure (length: $T$ , noise scheduler: $\{\beta_t\}_{t=1}^T$ ). The noised latent code $z_t$ at each time step $t$ is thus formulated as: + +$$ +z _ {t} = \sqrt {\bar {\alpha} _ {t}} z _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \epsilon \sim \mathcal {N} (0, I), \tag {1} +$$ + +where $\bar{\alpha}_t = \prod_{i=1}^t \alpha_t$ , $\alpha_t = 1 - \beta_t$ and $\epsilon$ is the adding noise vector sampled from standard normal distribution $\mathcal{N}(0, I)$ . 3D-UNet with parameter $\theta$ aims to estimate the noise $\epsilon$ based on the multiple inputs (i.e., noised video latent code $z_t$ , time step $t$ , and text prompt feature $c$ extracted by CLIP [48]). The Mean Square Error (MSE) loss is leveraged as the final objective $L$ : + +$$ +\mathcal {L} = \mathbb {E} _ {\epsilon \sim \mathcal {N} (0, I), t, c} [ \| \epsilon - \epsilon_ {\theta} (z _ {t}, t, c) \| ^ {2} ]. \tag {2} +$$ + +In the inference stage, the video latent code $\hat{z}_0$ is iteratively denoised by estimating noise $\epsilon_{\theta}$ in each time step via 3DUNet, and the following 2D VAE decoder $\mathcal{D}(\cdot)$ reconstructs each frame based on the video latent code $\hat{z}_0$ . + +# 3.2. Image Noise Prior + +Different from typical T2V diffusion model, I2V generation model goes one step further by additionally emphasizing the + +faithful alignment between the given first frame and the subsequent frames. Most existing I2V techniques [17, 60] control the video synthesis by concatenating the image latent code of first frame with the noised video latent code along channel/ temporal dimension. In this way, the visual information of first frame would be propagated across all frames via the temporal modules (e.g., temporal convolution and self-attention). Nevertheless, these solutions only rely on the temporal modules to calibrate video synthesis and leave the inherent relation between given image and each subsequent frame under-exploited, which is lacking in efficacy of temporal coherence modeling. To alleviate this issue, we formulate the typical noise prediction in I2V generation as temporal residual learning with reference to an image noise prior, targeting for amplifying the alignment between synthesized frames and first static image. + +Formally, given the image latent code $z_0^1$ of first frame, we concatenate it with the noised video latent code sequence $\{z_t^i\}_{i=1}^N$ ( $t$ denotes time steps in forward diffusion) along temporal dimension, yielding the conditional input $\{z_0^1, z_t^1, z_t^2, \dots, z_t^N\}$ for 3D-UNet. Meanwhile, we attempt to excavate the image noise prior to reflect the correlation between the first frame latent code $z_0^1$ and $i$ -th noised frame latent code $z_t^i$ . According to Eq. (1), we can reconstruct the $i$ -th frame latent code $z_0^i$ in one-step backward diffusion process as follows: + +$$ +z _ {0} ^ {i} = \frac {z _ {t} ^ {i} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t} ^ {i}}{\sqrt {\bar {\alpha} _ {t}}}, \tag {3} +$$ + +where $\epsilon_t^i$ denotes Gaussian noise which is added to the $i$ -th frame latent code. Considering the basic assumption in I2V that all frames in a short video clip are inherently correlated to the first frame, the $i$ -th frame latent code $z_0^i$ can be naturally formulated as the combination of the first frame latent code $z_0^1$ and a residual item $\Delta z^i$ as: + +$$ +z _ {0} ^ {i} = z _ {0} ^ {1} + \Delta z ^ {i}. \tag {4} +$$ + +Next, we construct a variable $C_t^i$ by adding a scale ratio to the residual item $\Delta z^i$ as follows: + +$$ +C _ {t} ^ {i} = \frac {\sqrt {\bar {\alpha} _ {t}} \Delta z ^ {i}}{\sqrt {1 - \bar {\alpha} _ {t}}}. \tag {5} +$$ + +Thus Eq. (5) can be re-written as follows: + +$$ +\Delta z ^ {i} = \frac {\sqrt {1 - \bar {\alpha} _ {t}} C _ {t} ^ {i}}{\sqrt {\bar {\alpha} _ {t}}}. \tag {6} +$$ + +After integrating Eq. (3) and Eq. (6) into Eq. (4), the first frame latent code $z_0^1$ is denoted as: + +$$ +\begin{array}{l} z _ {0} ^ {1} = z _ {0} ^ {i} - \Delta z ^ {i} = \frac {z _ {t} ^ {i} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t} ^ {i}}{\sqrt {\bar {\alpha} _ {t}}} - \frac {\sqrt {1 - \bar {\alpha} _ {t}} C _ {t} ^ {i}}{\sqrt {\bar {\alpha} _ {t}}} \\ = \frac {z _ {t} ^ {i} - \sqrt {1 - \bar {\alpha} _ {t}} (\epsilon_ {t} ^ {i} + C _ {t} ^ {i})}{\sqrt {\bar {\alpha} _ {t}}} = \frac {z _ {t} ^ {i} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t} ^ {i \rightarrow 1}}{\sqrt {\bar {\alpha} _ {t}}}, (7) \\ \end{array} +$$ + +where $\epsilon_t^{i\to 1}$ is defined as the image noise prior which can be interpreted as the relationship between the first frame latent code $z_0^1$ and $i$ -th noised frame latent code $z_{t}^{i}$ . The image noise prior $\epsilon_t^{i\rightarrow 1}$ is thus measured as: + +$$ +\epsilon_ {t} ^ {i \rightarrow 1} = \frac {z _ {t} ^ {i} - \sqrt {\bar {\alpha} _ {t}} z _ {0} ^ {1}}{\sqrt {1 - \bar {\alpha} _ {t}}}. \tag {8} +$$ + +According to Eq. (8), the latent code $z_0^1$ of first frame can be directly reconstructed through the one-step backward diffusion by using the image noise prior $\epsilon_t^{i\to 1}$ and the noised latent code $z_{t}^{i}$ of $i$ -th frame. In this way, such image noise prior $\epsilon_t^{i\to 1}$ acts as a reference noise for Gaussian noise $\epsilon_t^i$ which is added in $i$ -th frame latent code, when $z_0^i$ is close to $z_0^1$ (i.e., $i$ -th frame is similar to the given first frame). + +# 3.3. Temporal Residual Learning + +Recall that the image noise prior learnt in the shortcut pathway facilitates temporal modeling to amplify the alignment between the first frame and subsequent frames. To further strengthen temporal coherence among all adjacent frames, a residual pathway is designed to estimate the residual noise of each frame through inter-frame relational reasoning. Inspired by exploring knowledge prior in pre-trained 3DUNet for video editing [58], we propose to fine-tune 3DUNet to estimate the residual noise of each frame. Technically, the estimated noise $\tilde{\epsilon}_t^i$ of $i$ -th frame is formulated as the combination of the reference noise $\epsilon_t^{i\rightarrow 1}$ (i.e., image noise prior) and the estimated residual noise $\Delta \tilde{\epsilon}_t^i$ : + +$$ +\tilde {\epsilon} _ {t} ^ {i} = \lambda^ {i} \epsilon_ {t} ^ {i \rightarrow 1} + (1 - \lambda^ {i}) \Delta \tilde {\epsilon} _ {t} ^ {i}, \tag {9} +$$ + +where $\lambda^i$ is a trade-off parameter. $\Delta \tilde{\epsilon}_t^i$ is learnt by 3D-UNet (parameter: $\theta$ ), and $\epsilon_t^{i\rightarrow 1}$ can be computed by Eq. (8) on the fly. Finally, the I2V denoising diffusion objective $\tilde{L}$ is calculated as follows: + +$$ +\tilde {\mathcal {L}} = \mathbb {E} _ {\epsilon \sim \mathcal {N} (0, I), t, c, i} [ \| \epsilon_ {t} ^ {i} - \tilde {\epsilon} _ {t} ^ {i} \| ^ {2} ]. \tag {10} +$$ + +Note that the computation of Eq. (9) is operated through the Temporal Noise Fusion (TNF) module that estimates the backward diffusion noise of each frame. Moreover, since the temporal correlation between $i$ -th frame and the first frame will decrease when increasing frame index $i$ with enlarged timespan, we shape the trade-off parameter $\lambda^i$ as a linear decay parameter with respect to frame index. + +# 3.4. Temporal Noise Fusion Module + +Our residual-like dual-path noise prediction explores the image noise prior as a reference to facilitate temporal modeling among adjacent frames. However, the handcrafted design of TNF module with simple linear fusion (i.e., Eq. (9)) requires a careful tuning of hyper-parameter $\lambda^i$ , leading to a sub-optimal solution. Instead, we devise a new + +![](images/6375901c1e2ffd877e7c7ae2e0b8b0ec637cacd7929c76bb081efc47edf21450.jpg) +Figure 3. Detailed structure of Temporal Noise Fusion module. + +Transformer-based temporal noise fusion module to dynamically fuse reference and residual noises, aiming to further exploit the relations in between and boost noise fusion. + +Figure 3 depicts the structure of Temporal Noise Fusion module. Specifically, given the estimated residual noise $\Delta \tilde{\epsilon}_t^i$ , we first leverage the adaptive layer norm operations [68] which is modulated by the time step $t$ for spatial feature normalization. After that, one self-attention layer is employed for feature enhancement, followed by the same adaptive layer normalization operation. Next, by taking the in-process intermedia feature as query and the concatenated feature of $[\epsilon_t^{i\rightarrow 1},\Delta \tilde{\epsilon}_t^i ]$ as key/value, TNF module achieves the final backward diffusion noise $\tilde{\epsilon}_{\theta}(z_{t},t,c,i)$ through a cross-attention layer: + +$$ +\tilde {\epsilon} _ {\theta} \left(z _ {t}, t, c, i\right) = \tilde {\epsilon} _ {t} ^ {i} = \varphi \left(\Delta \tilde {\epsilon} _ {t} ^ {i}, \epsilon^ {i \rightarrow 1}, t\right), \tag {11} +$$ + +where $\varphi(\cdot)$ denotes the operation of TNF module. Compared to the simple linear fusion strategy for noise estimation, our Transformer-based TNF module sidesteps the hyper-parameter tuning and provides an elegant alternative to dynamically merge the reference and residual noises of each frame to yield high-fidelity videos. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. We empirically verify the merit of our TRIP model on WebVid-10M [1], DTDB [30] and MSR-VTT [67] datasets. Our TRIP is trained over the training set of WebVid-10M, and the validation sets of all three datasets are used for evaluation. The WebVid-10M dataset consists of about $10.7M$ video-caption pairs with a total of $57K$ video hours. $5K$ videos are used for validation and the remaining videos are utilized for training. We further sample 2,048 videos from the validation set of WebVid-10M for evaluation. DTDB contains more than $10K$ dynamic texture videos of natural scenes. We follow the standard protocols in [9] to evaluate models on a video subset derived from 4 categories of scenes (i.e., fires, clouds, vegetation, and waterfall). The texts of each category are used as the text prompt since there is no text description. For the MSR-VTT dataset, there are $10K$ clips in training set and 2,990 + +Table 1. Performances of F-Consistency (F-Consistency4: consistency among the first four frames, F-Consistencyall: consistency among all frames) and FVD on WebVid-10M. + +
ApproachF-Consistency4 (↑)F-Consistencyall (↑)FVD (↓)
T2V-Zero [32]91.5992.15279
VideoComposer [60]88.7892.52231
TRIP95.3696.4138.9
+ +clips in validation set. Each video clip is annotated with 20 text descriptions. We adopt the standard setting in [59] to pre-process the video clips in validation set for evaluation. + +Implementation Details. We implement our TRIP on PyTorch platform with Diffusers [57] library. The 3D-UNet [59] derived from Stable-Diffusion v2.1 [50] is exploited as the video backbone. Each training sample is a 16-frames clip and the sampling rate is 4 fps. We fix the resolution of each frame as $256 \times 256$ , which is centrally cropped from the resized video. The noise scheduler $\{\beta_t\}_{t=1}^T$ is set as linear scheduler ( $\beta_1 = 1 \times 10^{-4}$ and $\beta_T = 2 \times 10^{-2}$ ). We set the number of time steps $T$ as 1,000 in model training. The sampling strategy for video generation is DDIM [53] with 50 steps. The 3D-UNet and TNF module are trained with AdamW optimizer (learning rate: $2 \times 10^{-6}$ for 3D-UNet and $2 \times 10^{-4}$ for TNF module). All experiments are conducted on 8 NVIDIA A800 GPUs. + +Evaluation Metrics. Since there is no standard evaluation metric for I2V task, we choose the metrics of frame consistency (F-Consistency) which is commonly adopted in video editing [58, 63] and Frechet Video Distance (FVD) [55] widely used in T2V generation [2, 39] for the evaluation on WebVid-10M. Note that the vision model used to measure F-Consistency is the CLIP ViT-L/14 [48]. Specifically, we report the frame consistency of the first 4 frames (F-Consistency $_4$ ) and all the 16 frames (F-Consistency $_{all}$ ). For the evaluation on DTDB and MSR-VTT, we follow [9, 59] and report the frame-wise Frechet Inception Distance (FID) [23] and FVD performances. + +Human Evaluation. The user study is also designed to assess the temporal coherence, motion fidelity, and visual quality of the generated videos. In particular, we randomly sample 256 testing videos from the validation set in WebVid-10M for evaluation. Through the Amazon MTurk platform, we invite 32 evaluators, and each evaluator is asked to choose the better one from two synthetic videos generated by two different methods given the same inputs. + +# 4.2. Comparisons with State-of-the-Art Methods + +We compare our TRIP with several state-of-the-art I2V generation methods, e.g., VideoComposer [60] and T2V-Zero [32], on WebVid-10M. For the evaluation on DTDB, we set AL [11] and cINN [9] as baselines. Text-to-Video generation advances, e.g., CogVideo [28] and Make-A-Video [52] are included for comparison on MSR-VTT. + +Evaluation on WebVid-10M. Table 1 summarizes per + +![](images/6f16fa540ecea4fdd6408090ee56ff24ee5602f70a1030184a9bd37a583e06d7.jpg) +Figure 4. Performance comparisons of F-Consistency by using different number of frames on WebVid-10M. + +Table 2. Performances of averaged FID and FVD over four scene categories on DTDB dataset. + +
ApproachZero-shotFID (↓)FVD (↓)
AL [11]No65.1934.2
cINN [9]No31.9451.6
TRIPYes24.8433.9
+ +formance comparison on WebVid-10M. Overall, our TRIP consistently leads to performance boosts against existing diffusion-based baselines across different metrics. Specifically, TRIP achieves the F-Consistency4 of 95.36%, which outperforms the best competitor T2V-Zero by 3.77%. The highest frame consistency among the first four frames of our TRIP generally validates the best temporal coherence within the beginning of the whole video. Such results basically demonstrate the advantage of exploring image noise prior as the guidance to strengthen the faithful visual alignment between given image and subsequent frames. To further support our claim, Figure 4 visualizes the frame consistency of each run which is measured among different numbers of frames. The temporal consistencies of TRIP across various frames constantly surpass other baselines. It is worthy to note that TRIP even shows large superiority when only calculating the consistency between the first two frames. This observation again verifies the effectiveness of our residual-like noise prediction paradigm that nicely preserves the alignment with given image. Considering the metric of FVD, TRIP also achieves the best performance, which indicates that the holistic motion dynamics learnt by our TRIP are well-aligned with the ground-truth video data distribution in WebVid-10M. + +Figure 5 showcases four I2V generation results across three methods (VideoComposer, T2V-Zero, and our TRIP) based on the same given first frame and text prompt. Generally, compared to the two baselines, our TRIP synthesizes videos with higher quality by nicely preserving the faithful alignment with the given image & text prompt and meanwhile reflecting temporal coherence. For example, T2V-Zero produces frames that are inconsistent with the second prompt "the ship passes close to the camera" and VideoComposer generates temporally inconsistent frames for the + +Table 3. Performances of FID and FVD on MSR-VTT dataset. + +
ApproachModel TypeFID (↓)FVD (↓)
CogVideo [28]T2V23.591294
Make-A-Video [52]T2V13.17-
VideoComposer [60]T2V-580
ModelScopeT2V [59]T2V11.09550
VideoComposer [60]I2V31.29208
TRIPI2V9.6891.3
+ +Table 4. Human evaluation of the preference ratios between TRIP v.s. other approaches on WebVid-10M dataset. + +
Evaluation Itemsv.s. T2V-Zerov.s. VideoComposer
Temporal Coherence96.9 v.s. 3.184.4 v.s. 15.6
Motion Fidelity93.8 v.s. 6.281.3 v.s. 18.7
Visual Quality90.6 v.s. 9.487.5 v.s. 12.5
+ +last prompt "the portrait of the girl in glasses". In contrast, the synthesized videos of our TRIP are more temporally coherent and better aligned with inputs. This again confirms the merit of temporal residual learning with amplified guidance of image noise prior for I2V generation. + +Evaluation on DTDB. Next, we evaluate the generalization ability of our TRIP on DTDB dataset in a zero-shot manner. Table 2 lists the performances of averaged FID and FVD over four scene-related categories. In general, our TRIP outperforms two strong stochastic models (AL and cINN) over both two metrics. Note that cINN also exploits the residual information for I2V generation, but it combines image and residual features as clip feature by learning a bijective mapping. Our TRIP is fundamentally different in that we integrate residual-like noise estimation into the video denoising procedure to ease temporal modeling in the diffusion model. Figure 6 shows the I2V generation results of four different scenes produced by our TRIP. As shown in this figure, TRIP manages to be nicely generalized to scene-related I2V generation task and synthesizes realistic videos conditioned on the given first frame plus class label. + +Evaluation on MSR-VTT. We further compare our TRIP with recent video generation advances on MSR-VTT. Note that here we also evaluate TRIP in a zero-shot manner. Table 3 details the frame-wise FID and FVD of different runs. Similar to the observation on DTDB, our TRIP again surpasses all baselines over both metrics. The results basically validate the strong generalization ability of our proposal to formulate the open-domain motion dynamics. + +# 4.3. Human Evaluation + +In addition to the evaluation over automatic metrics, we also perform human evaluation to investigate user preference with regard to three perspectives (i.e., temporal coherence, motion fidelity, and visual quality) across different I2V approaches. Table 4 shows the comparisons of user preference on the generated videos over WebVid-10M dataset. Overall, our TRIP is clearly the winner in terms of all three criteria compared to both baselines (VideoComposer and T2V-Zero). This demonstrates the powerful temporal modeling + +![](images/2735994c1c4f7aa3ba0a31c6b543e61c369857720f6e3bcc3d6e9e77dddba9d0.jpg) +Figure 5. Examples of I2V generation results by three methods (VideoComposer, T2V-Zero, and our TRIP) on WebVid-10M dataset. We uniformly sample four frames of each generated video for visualization. + +![](images/c16510228d637cecab59ab75ff35e778f49b803ebd9b7a4438e73f76ab8e15f6.jpg) + +![](images/b7eb760c442529b6482b1c25a0fb23d5512bf99e40c428205b1aba56888fe6e2.jpg) +Figure 6. Image-to-video generation results of four different natural scenes by our TRIP on DTDB dataset. + +of motion dynamics through our residual-like noise prediction with image noise prior. + +# 4.4. Ablation Study on TRIP + +In this section, we perform ablation study to delve into the design of TRIP for I2V generation. Here all experiments are conducted on WebVid-10M for performance comparison. + +First Frame Condition. We first investigate different condition strategies to exploit the given first frame as additional condition for I2V generation. Table 5 summarizes the performance comparisons among different variants of our TRIP. $\mathbf{TRIP}_C$ follows the common recipe [60] to concatenate the clean image latent code of first frame with the noise video latent code along channel dimension. $\mathbf{TRIP}_{TE}$ is implemented by concatenating image latent code at the end of the noise video latent code along temporal dimension, while our proposal (TRIP) temporally concatenates image latent code at the beginning of video latent code. In par + +Table 5. Performance comparisons among different condition approaches on WebVid-10M dataset. + +
ModelF-Consistency4 (↑)F-Consistencyall (↑)FVD (↓)
TRIPC94.7796.1341.3
TRIPTE95.1796.2039.8
TRIP95.3696.4138.9
+ +Table 6. Evaluation of temporal residual learning in terms of F-Consistency and FVD on WebVid-10M dataset. + +
ModelF-Consistency4 (↑)F-Consistencyall (↑)FVD (↓)
TRIP-94.6695.9239.9
TRIPW95.2295.9643.0
TRIP95.3696.4138.9
+ +ticular, $\mathrm{TRIP}_{TE}$ exhibits better performances than $\mathrm{TRIP}_C$ , which indicates that temporal concatenation is a practical choice for augmenting additional condition in I2V generation. TRIP further outperforms $\mathrm{TRIP}_{TE}$ across both F-Consistency and FVD metrics. Compared to concatenating image latent code at the end of video, augmenting video latent code with input first frame latent code at the start position basically matches with the normal temporal direction, thereby facilitating temporal modeling. + +Temporal Residual Learning. Next, we study how each design of our TRIP influences the performances of both temporal consistency and visual quality for I2V generation. Here we include two additional variants of TRIP: (1) $\mathbf{TRIP}^{-}$ removes the shortcut path of temporal residual learning and only capitalizes on 3D-UNet for noise prediction; (2) $\mathbf{TRIP}^{W}$ is a degraded variant of temporal residual learning with image noise prior by simply fusing reference and residual noise via linear fusion strategy. As shown in Table 6, $\mathbf{TRIP}^{W}$ exhibits better performances than $\mathbf{TRIP}^{-}$ + +![](images/f9067e6bd7cca475750b152828b1dd38a34356cda17a544b0762e20b4a949640.jpg) +A castle in the middle of a forest with a river running through it + +![](images/b87c206034da2155825c89c1103dc9138285fa045afa37101baadb0eb4a35cc9.jpg) +Crabs with bubbles crawls on the beach with sunset in the background + +![](images/8e4618d020e042182948f7ada9891254b3774754940ae03467dd3467bd555c8a.jpg) +A ship sails in the sea. + +![](images/76d9677ef005c7c27bef7306a98c182645d2fede00bc546df2e82894ebc36d3c.jpg) +WALL-E on the beach with a plant in his hand + +![](images/5e459fa9626d9d84dda509f10dae4efe49028429e38615bbebe7e775d5c11d67.jpg) +Halloweep pumpkin with glowing eyes and smoke coming out of it + +![](images/8527d2a2d525e666fb73bd1ed5995137a017fab990a4cadd3a63f2a0d6e075da.jpg) +A bear is playing a guitar in front of a tank in the snow. + +![](images/04b4d4e3592b1d6ac4e85f3cebddfc169199d84b6d1fa4b327c3a1faf9158980.jpg) +Figure 7. Customized image animation with Stable-Diffusion XL [47] for text-to-video generation. We first employ Stable-Diffusion XL to synthesize fancy images and then exploit TRIP for image animation. Six generated videos with corresponding text prompts are presented. +Figure 8. Visualization of I2V example with text prompt "Young male tourist taking a photograph of old city on sunset" by using variants of our TRIP. + +in terms of F-consistency. The results highlight the advantage of leveraging image noise prior as the reference to amplify the alignment between the first frame and subsequent frames. By further upgrading temporal residual learning $(\mathbf{TRIP}^W)$ with Transformer-based temporal noise fusion, our TRIP achieves the best performances across all metrics. Figure 8 further showcases the results of different variants, which clearly confirm the effectiveness of our designs. + +# 4.5. Application: Customized Image Animation + +In this section, we provide two interesting applications of customized image animation by directly extending TRIP in a tuning-free manner. The first application is text-to-video pipeline that first leverages text-to-image diffusion models (e.g., Stable-Diffusion XL [47]) to synthesize fancy images and then employs our TRIP to animate them as videos. Figure 7 shows six text-to-video generation examples by using Stable-Diffusion XL and TRIP. The results generally demonstrate the generalization ability of our TRIP when taking text-to-image results as additional reference for text-to-video generation. To further improve controllability of text-to-video generation in real-world scenarios, we can additionally utilize image editing models (e.g., InstructPix2Pix [3] and ControlNet [74]) to modify the visual content of input images and then feed them into TRIP + +![](images/0c9fd19058f83c28b6cf0841be39307611a30aa28e7e9541f1c688cff9f994ab.jpg) + +![](images/c091633088a02362fb475a8f07df76d740e33e2524af34e8bb6b4af3e0bace66.jpg) + +![](images/7f9144f2263ecb43d5143464ef8c9aaad5aa40da93e6c436284ab0a64fc5e17b.jpg) +A beautiful duck swims on the pond +Figure 9. Customized image animation with image editing models (e.g., InstructPix2Pix [3] and ControlNet [74]). The first row shows image animation results given the primary real-world first image. The second and third rows show I2V results given the edited images via InstructPix2Pix and ControlNet, respectively. + +for video generation. As shown in Figure 9, our TRIP generates vivid videos based on the edited images with promising temporal coherence and visual quality. + +# 5. Conclusions + +This paper explores inherent inter-frame correlation in diffusion model for image-to-video generation. Particularly, we study the problem from a novel viewpoint of leveraging temporal residual learning with reference to the image noise prior of given image for enabling coherent temporal modeling. To materialize our idea, we have devised TRIP, which executes image-conditioned noise prediction in a residual fashion. One shortcut path is novelly designed to learn image noise prior based on given image and noised video latent code, leading to reference noise for enhancing alignment between synthesized frames and given image. Meanwhile, another residual path is exploited to estimate the residual noise of each frame through eased temporal modeling. Experiments conducted on three datasets validate the superiority of our proposal over state-of-the-art approaches. + +# References + +[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. In ICCV, 2021. 5 +[2] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. In CVPR, 2023. 1, 2, 3, 5 +[3] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning to Follow Image Editing Instructions. In CVPR, 2023. 1, 2, 8 +[4] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. StableVideo: Text-driven Consistency-aware Diffusion Video Editing. In ICCV, 2023. 2 +[5] Zhikai Chen, Fuchen Long, Zhaofan Qiu, Ting Yao, Wengang Zhou, Jiebo Luo, and Tao Mei. AnchorFormer: Point Cloud Completion from Discriminative Nodes. In CVPR, 2023. 2 +[6] Paul Couairon, Clément Rambour, Jean-Emmanuel Haugeard, and Nicolas Thome. VidEdit: Zero-Shot and Spatially Aware Text-Driven Video Editing. arXiv preprint arXiv:2306.08707, 2023. 2 +[7] Prafulla Dhariwal and Alex Nichol. Diffusion Models Beat GANs on Image Synthesis. In NeurIPS, 2022. 1 +[8] Ziyi Dong, Pengxu Wei, and Liang Lin. DreamArtist: Towards Controllable One-Shot Text-to-Image Generation via Positive-Negative Prompt-Tuning. arXiv preprint arXiv:2211.11337, 2022. 2 +[9] Michael Dorkenwald, Timo Milbich, Andreas Blattmann, Robin Rombach, Konstantinos G. Derpanis, and Björn Ommer. Stochastic Image-to-Video Synthesis using cINNs. In CVPR, 2021. 1, 2, 5, 6 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 2 +[11] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating Landscape: Self-Supervised Learning of Decoupled Motion and Appearance for Single-Image Video Synthesis. TOG, 2019. 5, 6 +[12] Siming Fan, Jingtan Piao, Chen Qian, Kwan-Yee Lin, and Hongsheng Li. Simulating Fluids in Real-World Still Images. In ICCV, 2023. 1, 2 +[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashni, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion. In ICLR, 2023. 2 +[14] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve Your Own Correlation: A Noise Prior for Video Diffusion Models. In ICCV, 2023. 2 +[15] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. TokenFlow: Consistent Diffusion Features for Consistent Video Editing. arXiv preprint arXiv:2307.10373, 2023. 2 + +[16] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative Adversarial Networks. In NeurIPS, 2014. 1 +[17] Xianfan Gu, Chuan Wen, Jiaming Song, and Yang Gao. Seer: Language Instructed Video Prediction with Latent Diffusion Models. arXiv preprint arXiv:2303.14897, 2023. 1, 2, 4 +[18] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning. arXiv preprint arXiv:2307.04725, 2023. 2 +[19] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible Diffusion Modeling of Long Videos. In NeurIPS, 2022. 2 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 2 +[21] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, and Qifeng Chen. animate-A-Story: Storytelling with Retrieval-Augmented Video Generation. arXiv preprint arXiv:2307.06940, 2023. 2 +[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-Prompt Image Editing with Cross-Attention Control. In ICLR, 2023. 1, 2 +[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. In NeuIPS, 2017. 5 +[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In NeurIPS, 2020. 1 +[25] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen Video: High Definition Video Generation with Diffusion Models. In CVPR, 2022. 1, 2 +[26] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video Diffusion Models. In NeurIPS, 2022. 1, 2, 3 +[27] Aleksander Holynski, Brian Curless, Steven M. Seitz, and Richard Szeliski. Animating Pictures with Eulerian Motion Fields. In CVPR, 2021. 2 +[28] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. CogVideo: Large-Scale Pretraining for Text-to-Video Generation via Transformers. In ICLR, 2023. 5, 6 +[29] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely Connected Convolutional Networks. In CVPR, 2017. 2 +[30] Hadji Isma and Wildes Richard P. A New Large Scale Dynamic Texture Dataset with Application to Convnet Understanding. In ECCV, 2018. 5 +[31] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. DreamPose: Fashion Image-to-Video Synthesis via Stable Diffusion. In ICCV, 2023. 2 +[32] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant + +Navasardyan, and Humphrey Shi. Text2Video-Zero: Text-to-Image Diffusion Models are Zero-Shot Video Generators. In ICCV, 2023. 2, 5 +[33] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-Concept Customization of Text-to-Image Diffusion. In CVPR, 2023. 2 +[34] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Flow-Grounded Spatial-Temporal Video Prediction from Still Images. In ECCV, 2018. 2 +[35] Yehao Li, Ting Yao, Yingwei Pan, and Tao Mei. Contextual Transformer Networks for Visual Recognition. IEEE Trans. on PAMI, 2022. 2 +[36] Fuchen Long, Zhaofan Qiu, Yingwei Pan, Ting Yao, Jiebo Luo, and Tao Mei. Stand-Alone Inter-Frame Attention in Video Models. In CVPR, 2022. +[37] Fuchen Long, Zhaofan Qiu, Yingwei Pan, Ting Yao, Chong-Wah Ngo, and Tao Mei. Dynamic Temporal Filtering in Video Models. In ECCV, 2022. 2 +[38] Fuchen Long, Zhaofan Qiu, Ting Yao, and Tao Mei. Video-Drafter: Content-Consistent Multi-Scene Video Generation with LLM. arXiv:2401.01256, 2024. 2 +[39] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. VideoFusion: Decomposed Diffusion Models for High-Quality Video Generation. In CVPR, 2023. 2, 5 +[40] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable Animation of Fluid Elements in Still Images. In CVPR, 2022. 1, 2 +[41] Babaeizadeh Mohammad, Finn Chelsea, Erhan Dumitru, Campbell Roy H, and Levine Sergey. Stochastic Variational Video Prediction. In ICLR, 2018. 2 +[42] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2I-Adapter: Learning Adapters to Dig out More Controllable Ability for Text-to-Image Diffusion Models. arXiv preprint arXiv:2302.08453, 2023. 2 +[43] Haomiao Ni, Changhao Shi, Kai Li, Sharon X. Huang, and Martin Renqiang Min. Conditional Image-to-Video Generation with Latent Flow Diffusion Models. In CVPR, 2023. 2 +[44] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. In ICML, 2022. 1, 2, 3 +[45] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. CoDeF: Content Deformation Fields for Temporally Consistent Video Processing. arXiv preprint arXiv:2308.07926, 2023. 2 +[46] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot Image-to-Image Translation. In ACM SIGGRAPH, 2023. 2 +[47] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. arXiv preprint arXiv:2307.01952, 2023. 8 + +[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 3, 5 +[49] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv preprint arXiv:2204.06125, 2022.3 +[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022. 1, 2, 5 +[51] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. In CVPR, 2023. 2 +[52] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-Video: Text-to-Video Generation without Text-Video Data. In ICLR, 2023. 2, 5, 6 +[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models. In ICLR, 2021. 5 +[54] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Palur. A Closer Look at Spatiotemporal Convolutions for Action Recognition. In CVPR, 2018. 2 +[55] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new Metric for Video Generation. In ICLR Deep-GenStruct Workshop, 2019. 5 +[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. In NeurIPS, 2017. 2 +[57] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-Of-The-Art Diffusion Models. 5 +[58] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-L-Video: Multi-Text to Long Video Generation via Temporal Co-Denoising. arXiv preprint arXiv:2305.18264, 2023. 4, 5 +[59] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. ModelScope Text-to-Video Technical Report. arXiv preprint arXiv:2308.06571, 2023. 2, 5, 6 +[60] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional Video Synthesis with Motion Controllability. In NeurIPS, 2023. 1, 2, 4, 5, 6, 7 +[61] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent Image Animator: Learning to Animate Images via Latent Space Navigation. In ICLR, 2022. 2 +[62] Yaohui Wang, Xin Ma, Xinyuan Chen, Antitza Dantcheva, Bo Dai, and Yu Qiao. LEO: Generative Latent Image + +Animator for Human Video Synthesis. arXiv preprint arXiv:2305.03989, 2023. 2 +[63] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-A-Video: One-Shot Tuning of Image Diffusion Models for Text-to-Video Generation. In ICCV, 2023. 5 +[64] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A Survey on Video Diffusion Models. arXiv preprint arXiv:2310.10647, 2023. 2 +[65] Zhen Xing, Qi Dai, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. SimDA: Simple Diffusion Adapter for Efficient Video Generation. In CVPR, 2024. 2 +[66] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to Generate Time-Lapse Videos Using Multi-Stage Dynamic Generative Adversarial Networks. In CVPR, 2018. +[67] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A Large Video Description Dataset for Bridging Video and Language. In CVPR, 2016. 5 +[68] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and Improving Layer Normalization. In NeurIPS, 2019. 5 +[69] Hanshu Yan, Jun Hao Liew, Long Mai, Shanchuan Lin, and Jiashi Feng. MagicProp: Diffusion-based Video Editing via Motion-aware Appearance Propagation. arXiv preprint arXiv:2309.00908, 2023. 2 +[70] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion Probabilistic Modeling for Video Generation. arXiv preprint arXiv:2203.09481, 2022. 2 +[71] Ting Yao, Yingwei Pan, Yehao Li, Chong-Wah Ngo, and Tao Mei. Wave-ViT: Unifying Wavelet and Transformers for Visual Representation Learning. In ECCV, 2022. 2 +[72] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. DragNUWA: Fine-grained Control in Video Generation by Integrating Text, Image, and Trajectory. arXiv preprint arXiv:2308.08089, 2023. 2 +[73] Jianfeng Zhang, Hanshu Yan, Zhongcong Xu, Jiashi Feng, and Jun Hao Liew. MagicAvatar: Multimodal Avatar Generation and Animation. arXiv preprint arXiv:2308.14748, 2023. 2 +[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In ICCV, 2023. 2, 8 +[75] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks. In ICCV, 2017. 1 \ No newline at end of file diff --git a/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/images.zip b/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bf885404cbd7dd04e99db4b3782d163c9a8f2cca --- /dev/null +++ b/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c5def167b77482820e4da64f6d5aed4225428a020b418df7e94e5c4add8a7bc +size 696989 diff --git a/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/layout.json b/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b3d153e3b172651d5d8f9bc7a008b967dcb90d05 --- /dev/null +++ b/2024/TRIP_ Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models/layout.json @@ -0,0 +1,9856 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 113, + 103, + 481, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 103, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 113, + 103, + 481, + 138 + ], + "type": "text", + "content": "TRIP: Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models*" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 162, + 499, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 162, + 499, + 201 + ], + "spans": [ + { + "bbox": [ + 92, + 162, + 499, + 201 + ], + "type": "text", + "content": "Zhongwei Zhang†, Fuchen Long§, Yingwei Pan§, Zhaofan Qiu§, Ting Yao§, Yang Cao† and Tao Mei§ \n†University of Science and Technology of China, Hefei, China \n§HiDream.ai Inc." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 96, + 205, + 501, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 205, + 501, + 230 + ], + "spans": [ + { + "bbox": [ + 96, + 205, + 501, + 230 + ], + "type": "text", + "content": "zhwzhang@mail.ustc.edu.cn,{longfuchen, pandy, qiuzhaofan, tiyao}@hidream.ai forrest@ustc.edu.cn, tmei@hidream.ai" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 258, + 191, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 191, + 271 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 191, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 284, + 290, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 290, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 290, + 632 + ], + "type": "text", + "content": "Recent advances in text-to-video generation have demonstrated the utility of powerful diffusion models. Nevertheless, the problem is not trivial when shaping diffusion models to animate static image (i.e., image-to-video generation). The difficulty originates from the aspect that the diffusion process of subsequent animated frames should not only preserve the faithful alignment with the given image but also pursue temporal coherence among adjacent frames. To alleviate this, we present TRIP, a new recipe of image-to-video diffusion paradigm that pivots on image noise prior derived from static image to jointly trigger inter-frame relational reasoning and ease the coherent temporal modeling via temporal residual learning. Technically, the image noise prior is first attained through one-step backward diffusion process based on both static image and noised video latent codes. Next, TRIP executes a residual-like dual-path scheme for noise prediction: 1) a shortcut path that directly takes image noise prior as the reference noise of each frame to amplify the alignment between the first frame and subsequent frames; 2) a residual path that employs 3D-UNet over noised video and static image latent codes to enable inter-frame relational reasoning, thereby easing the learning of the residual noise for each frame. Furthermore, both reference and residual noise of each frame are dynamically merged via attention mechanism for final video generation. Extensive experiments on WebVid-10M, DTDB and MSR-VTT datasets demonstrate the effectiveness of our TRIP for image-to-video generation. Please see our project page at https://trip-i2v.github.io/TRIP/." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 652, + 128, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 652, + 128, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 652, + 128, + 665 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 672, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 672, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 672, + 287, + 696 + ], + "type": "text", + "content": "In recent years, deep generative models have demonstrated impressive capabilities to create high-quality visual con" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 258, + 420, + 387 + ], + "blocks": [ + { + "bbox": [ + 313, + 258, + 420, + 387 + ], + "lines": [ + { + "bbox": [ + 313, + 258, + 420, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 420, + 387 + ], + "type": "image", + "image_path": "983675b4f8017e056172a5a2c80320ab12aa4c9f5a2e3d5ae88402bca7bde2e9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 388, + 411, + 395 + ], + "lines": [ + { + "bbox": [ + 325, + 388, + 411, + 395 + ], + "spans": [ + { + "bbox": [ + 325, + 388, + 411, + 395 + ], + "type": "text", + "content": "(a) Independent noise prediction" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 399, + 545, + 431 + ], + "lines": [ + { + "bbox": [ + 305, + 399, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 545, + 431 + ], + "type": "text", + "content": "Figure 1. Comparison between (a) independent noise prediction and (b) our residual-like noise prediction with image noise prior in I2V diffusion models." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 423, + 258, + 536, + 385 + ], + "blocks": [ + { + "bbox": [ + 423, + 258, + 536, + 385 + ], + "lines": [ + { + "bbox": [ + 423, + 258, + 536, + 385 + ], + "spans": [ + { + "bbox": [ + 423, + 258, + 536, + 385 + ], + "type": "image", + "image_path": "dfad6fbaebb001f86268039a5a504716ef7e2de6923644bf2a57e1eae95db71e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 385, + 521, + 399 + ], + "lines": [ + { + "bbox": [ + 436, + 385, + 521, + 399 + ], + "spans": [ + { + "bbox": [ + 436, + 385, + 521, + 399 + ], + "type": "text", + "content": "(b) Residual-like noise prediction with image noise prior" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 437, + 546, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 641 + ], + "type": "text", + "content": "tent. In between, Generative Adversarial Networks (GANs) [16, 75] and diffusion models [7, 24] have brought forward milestone improvement for a series of generative tasks in computer vision field, e.g., text-to-image generation [44, 50], image editing [3, 22], and text-to-video generation [2, 25, 26]. In this work, we are interested in imaged-to-video generation (I2V), i.e., animating a static image by endowing it with motion dynamics, which has great potential real-world applications (e.g., entertainment and creative media). Conventional I2V techniques [9, 12, 40] commonly target for generating video from a natural scene image with random or very coarse motion dynamics (such as clouds moving or fluid flowing). In contrast, our focus is a more challenging text-driven I2V scenario: given the input static image and text prompt, the animated frames should not only faithfully align with the given first frame, but also be temporally coherent and semantically matched with text prompt." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "Several recent pioneering practices [17, 60] for I2V is to remould the typical latent diffusion model of text-to-video generation by directly leveraging the given static image as additional condition during diffusion process. Figure 1 (a) illustrates such conventional noise prediction strategy in latent diffusion model. Concretely, by taking the given" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 212, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 212, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 212, + 712 + ], + "type": "text", + "content": "*This work was performed at HiDream.ai." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8671" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 418 + ], + "type": "text", + "content": "static image as the first frame, 2D Variational Auto-encoder (VAE) is first employed to encode the first frame as image latent code. This encoded image latent code is further concatenated with noised video latent code (i.e., a sequence of Gaussian noise) to predict the backward diffusion noise of each subsequent frame via a learnable 3D-UNet. Nevertheless, such independent noise prediction strategy of each frame leaves the inherent relation between the given image and each subsequent frame under-exploited, and is lacking in efficacy of modeling temporal coherence among adjacent frames. As observed in I2V results of Figure 1 (a), both the foreground content (a glass) and the background content (e.g., a table) in the synthesized second frame are completely different from the ones in the given first frame. To alleviate this issue, our work shapes a new way to formulate the noise prediction in I2V diffusion model as temporal residual learning on the basis of amplified guidance of given image (i.e., image noise prior). As illustrated in Figure 1 (b), our unique design is to integrate the typical noise prediction with an additional shortcut path that directly estimates the backward diffusion noise of each frame by solely referring image noise prior. Note that this image noise prior is calculated based on both input image and noised video latent codes, which explicitly models the inherent correlation between the given first frame and each subsequent frame. Meanwhile, such residual-like scheme re-shapes the typical noise prediction via 3D-UNet as residual noise prediction to ease temporal modeling among adjacent frames, thereby leading to more temporally coherent I2V results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 419, + 289, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 289, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 289, + 636 + ], + "type": "text", + "content": "By materializing the idea of executing image-conditioned noise prediction in a residual manner, we present a novel diffusion model, namely Temporal Residual learning with Image noise Prior (TRIP), to facilitate I2V. Specifically, given the input noised video latent code and the corresponding static image latent code, TRIP performs residual-like noise prediction along two pathways. One is the shortcut path that first achieves image noise prior based on static image and noised video latent codes through one-step backward diffusion process, which is directly regarded as the reference noise of each frame. The other is the residual path which concatenates static image and noised video latent codes along temporal dimension and feeds them into 3D-UNet to learn the residual noise of each frame. Eventually, a Transformer-based temporal noise fusion module is leveraged to dynamically fuse the reference and residual noises of each frame, yielding high-fidelity video that faithfully aligns with given image." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 646, + 134, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 134, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 134, + 658 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "type": "text", + "content": "Text-to-Video Diffusion Models. The great success of diffusion models [44, 50] to generate fancy images based on text prompts has been witnessed in recent years. The foundation of the basic generative diffusion models fur" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 348 + ], + "type": "text", + "content": "ther encourages the development of customized image synthesis, such as image editing [3, 22, 46] and personalized image generation [8, 13, 33, 42, 51, 74]. Inspired by the impressive results of diffusion models on image generation, a series of text-to-video (T2V) diffusion models [2, 19, 25, 26, 32, 39, 52, 59, 64, 65, 70] start to emerge. VDM [26] is one of the earlier works that remoulds text-to-image (T2I) 2D-UNet architecture with temporal attention for text-to-video synthesis. Later in [2] and [18], the temporal modules (e.g., temporal convolution and self-attention) in diffusion models are solely optimized to emphasize motion learning. To further alleviate the burden on temporal modeling, Ge et al. [14] present a mixed video noise model under the assumption that each frame should correspond to a shared common noise and its own independent noise. Similar recipe is also adopted by VideoFusion [39] which estimates the basic noise and residual noise via two image diffusion models for video denoising. Other works go one step further by applying diffusion models for customized video synthesis, such as video editing [4, 6, 15, 45, 69] and personalized video generation [18, 21, 73]. In this work, we choose the basic T2V diffusion model with 3D-UNet [59] as our backbone for image-to-video generation task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 351, + 547, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 547, + 626 + ], + "type": "text", + "content": "Image-to-Video Generation. Image-to-Video (I2V) generation is one kind of conditional video synthesis, where the core condition is a given static image. According to the availability of motion cues, I2V approaches can be grouped into two categories: stochastic generation (solely using input image as condition) [27, 34, 41, 66] and conditional generation (using image and other types of conditions) [31, 38, 40, 43, 61]. Early attempts of I2V generation mostly belong to the stochastic models which usually focus on the short motion of fluid elements (e.g., water flow) [12, 27] or human poses [62] for landscape or human video generation. To further enhance the controllability of motion modeling, the conditional models exploit more types of conditions like text prompts [9, 17] in the procedure of video generation. Recent advances start to explore diffusion models [60, 72] for conditional I2V task. One pioneering work (VideoComposer [60]) flexibly composes a video with additional conditions of texts, optical flow or depth mapping sequence for I2V generation. Nevertheless, VideoComposer independently predicts the noise of each frame and leaves the inherent relation between input image and other frames under-exploited, thereby resulting in temporal inconsistency among frames." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Deep Residual Learning. The effectiveness of learning residual components with additional shortcut connections in deep neural networks [20, 29, 54] has been verified in various vision tasks. Besides typical convolution-based structure, recent advances [10, 56] also integrate the shortcut connection into the emerging transformer-based structure [5, 35-37, 71]. Intuitively, learning residual part with refer" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8672" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 382, + 199 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 382, + 199 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 382, + 199 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 382, + 199 + ], + "type": "image", + "image_path": "067e4d0946162a94f0a173b77564d0ad26b672674905f7e58a4e7b1e6cdfd349.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 202, + 545, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 545, + 291 + ], + "type": "text", + "content": "Figure 2. An overview of (a) our Temporal Residual learning with Image noise Prior (TRIP), and (b) the conceptual comparison between typical frame-wise denoising and our temporal noise fusion. During training, the input video clip is encoded into a sequence of image latent codes via a pre-trained 2D VAE. The latent code of first frame is then concatenated with noised video latent code as the input of the 3D-UNet. Then, TRIP performs a residual-like noise prediction along two pathways, i.e., shortcut path and residual path. The shortcut path computes the image noise prior as the reference noise based on the first frame latent codes and the noised video latent codes through a one-step backward diffusion. The residual path leverages 3D-UNet to predict the residual noise with reference to the image noise prior. By dynamically merging the reference and residual noise via a Transformer-based Temporal Noise Fusion (TNF) module, the target noise vector for video denoising is attained. Both 3D-UNet and TNF module are jointly optimized with the typical MSE loss for I2V generation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 389, + 71, + 536, + 199 + ], + "blocks": [ + { + "bbox": [ + 389, + 71, + 536, + 199 + ], + "lines": [ + { + "bbox": [ + 389, + 71, + 536, + 199 + ], + "spans": [ + { + "bbox": [ + 389, + 71, + 536, + 199 + ], + "type": "image", + "image_path": "22e148f9973d9fd46f9a651a29b2ea155ec091d6223a2912e1c7d6cb68444910.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 304, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 365 + ], + "type": "text", + "content": "ence to the optimal identity mapping will ease the network optimization. In this work, we capitalize on such principle and formulate noise prediction in I2V diffusion as temporal residual learning on the basis of the image noise prior derived from static image to enhance temporal coherence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 365, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 437 + ], + "type": "text", + "content": "In summary, our work designs a novel image-to-video diffusion paradigm with residual-like noise prediction. The proposed TRIP contributes by studying not only how to excavate the prior knowledge of given image as noise prior to guide temporal modeling, and how to strengthen temporal consistency conditioned on such image noise prior." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 453, + 137, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 137, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 137, + 466 + ], + "type": "text", + "content": "3. Our Approach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "type": "text", + "content": "In this section, we present our Temporal Residual learning with Image noise Prior (TRIP) for I2V generation. Figure 2 illustrates an overview of our TRIP. Given a video clip at training, TRIP first extracts the per-frame latent code via a pre-trained VAE and groups them as a video latent code. The image latent code of the first frame (i.e., the given static image) is regarded as additional condition, which is further concatenated with the noised video latent code along temporal dimension as the input of the 3D-UNet. Next, TRIP executes a residual-like noise prediction along two pathways (i.e., shortcut path and residual path), aiming to trigger inter-frame relational reasoning. In shortcut path, the image noise prior of each frame is calculated based on the correlation between the first frame and subsequent frames through one-step backward diffusion process. In residual path, the residual noise is estimated by employing 3D-UNet over the concatenated noised video latent code plus text prompt. Finally, a transformer-based Temporal Noise Fusion (TNF) module dynamically merges the image noise prior and the residual noise as the predicted noise for video generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 304, + 504, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 304, + 504, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 504, + 316 + ], + "type": "text", + "content": "3.1. Preliminaries: Video Diffusion Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "text", + "content": "T2V diffusion models [2, 26] are usually constructed by remoulding T2I diffusion models [44, 49]. A 3D-UNet with text prompt condition is commonly learnt to denoise a randomly sampled noise sequence in Gaussian distribution for video synthesis. Specifically, given an input video clip with " + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "text", + "content": " frames, a pre-trained 2D VAE encoder " + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\cdot)" + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "text", + "content": " first extracts latent code of each frame to constitute a latent code sequence " + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "inline_equation", + "content": "\\{z_0^i\\}_{i=1}^N" + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "text", + "content": ". Then, this latent sequence is grouped via concatenation along temporal dimension as a video latent code " + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "text", + "content": ". Next, the Gaussian noise is gradually added to " + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "text", + "content": " through forward diffusion procedure (length: " + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "text", + "content": ", noise scheduler: " + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "inline_equation", + "content": "\\{\\beta_t\\}_{t=1}^T" + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "text", + "content": "). The noised latent code " + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "text", + "content": " at each time step " + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 323, + 545, + 478 + ], + "type": "text", + "content": " is thus formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 344, + 485, + 545, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 485, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 344, + 485, + 545, + 498 + ], + "type": "interline_equation", + "content": "z _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\epsilon \\sim \\mathcal {N} (0, I), \\tag {1}", + "image_path": "de0269a6f97b1e8ae72557ebbc62a96728e418044eaf874c42b8145fc4bef427.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_t" + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "inline_equation", + "content": "\\alpha_t = 1 - \\beta_t" + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "text", + "content": " is the adding noise vector sampled from standard normal distribution " + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, I)" + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "text", + "content": ". 3D-UNet with parameter " + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "text", + "content": " aims to estimate the noise " + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "text", + "content": " based on the multiple inputs (i.e., noised video latent code " + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "text", + "content": ", time step " + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "text", + "content": ", and text prompt feature " + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "text", + "content": " extracted by CLIP [48]). The Mean Square Error (MSE) loss is leveraged as the final objective " + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 505, + 545, + 589 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 347, + 596, + 545, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 596, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 347, + 596, + 545, + 610 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {\\epsilon \\sim \\mathcal {N} (0, I), t, c} [ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, c) \\| ^ {2} ]. \\tag {2}", + "image_path": "d0cb51c691d742c65b7bbafbc42109a2ff6ca2866515c45a6bb10690c72d3251.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 616, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 616, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 305, + 616, + 545, + 664 + ], + "type": "text", + "content": "In the inference stage, the video latent code " + }, + { + "bbox": [ + 305, + 616, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 305, + 616, + 545, + 664 + ], + "type": "text", + "content": " is iteratively denoised by estimating noise " + }, + { + "bbox": [ + 305, + 616, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 305, + 616, + 545, + 664 + ], + "type": "text", + "content": " in each time step via 3DUNet, and the following 2D VAE decoder " + }, + { + "bbox": [ + 305, + 616, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot)" + }, + { + "bbox": [ + 305, + 616, + 545, + 664 + ], + "type": "text", + "content": " reconstructs each frame based on the video latent code " + }, + { + "bbox": [ + 305, + 616, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 305, + 616, + 545, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 671, + 414, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 414, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 414, + 684 + ], + "type": "text", + "content": "3.2. Image Noise Prior" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Different from typical T2V diffusion model, I2V generation model goes one step further by additionally emphasizing the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8673" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "content": "faithful alignment between the given first frame and the subsequent frames. Most existing I2V techniques [17, 60] control the video synthesis by concatenating the image latent code of first frame with the noised video latent code along channel/ temporal dimension. In this way, the visual information of first frame would be propagated across all frames via the temporal modules (e.g., temporal convolution and self-attention). Nevertheless, these solutions only rely on the temporal modules to calibrate video synthesis and leave the inherent relation between given image and each subsequent frame under-exploited, which is lacking in efficacy of temporal coherence modeling. To alleviate this issue, we formulate the typical noise prediction in I2V generation as temporal residual learning with reference to an image noise prior, targeting for amplifying the alignment between synthesized frames and first static image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "text", + "content": "Formally, given the image latent code " + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "inline_equation", + "content": "z_0^1" + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "text", + "content": " of first frame, we concatenate it with the noised video latent code sequence " + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\{z_t^i\\}_{i=1}^N" + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "text", + "content": " denotes time steps in forward diffusion) along temporal dimension, yielding the conditional input " + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\{z_0^1, z_t^1, z_t^2, \\dots, z_t^N\\}" + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "text", + "content": " for 3D-UNet. Meanwhile, we attempt to excavate the image noise prior to reflect the correlation between the first frame latent code " + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "inline_equation", + "content": "z_0^1" + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "text", + "content": "-th noised frame latent code " + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "inline_equation", + "content": "z_t^i" + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "text", + "content": ". According to Eq. (1), we can reconstruct the " + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "text", + "content": "-th frame latent code " + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "inline_equation", + "content": "z_0^i" + }, + { + "bbox": [ + 47, + 263, + 289, + 384 + ], + "type": "text", + "content": " in one-step backward diffusion process as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 119, + 388, + 287, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 388, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 119, + 388, + 287, + 416 + ], + "type": "interline_equation", + "content": "z _ {0} ^ {i} = \\frac {z _ {t} ^ {i} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} ^ {i}}{\\sqrt {\\bar {\\alpha} _ {t}}}, \\tag {3}", + "image_path": "27b87de5c9f978c2e60cb0f0c4a8b4e6a50e395d304767a874b911b90bdb5a69.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 422, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 289, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 422, + 289, + 495 + ], + "type": "inline_equation", + "content": "\\epsilon_t^i" + }, + { + "bbox": [ + 46, + 422, + 289, + 495 + ], + "type": "text", + "content": " denotes Gaussian noise which is added to the " + }, + { + "bbox": [ + 46, + 422, + 289, + 495 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 422, + 289, + 495 + ], + "type": "text", + "content": "-th frame latent code. Considering the basic assumption in I2V that all frames in a short video clip are inherently correlated to the first frame, the " + }, + { + "bbox": [ + 46, + 422, + 289, + 495 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 422, + 289, + 495 + ], + "type": "text", + "content": "-th frame latent code " + }, + { + "bbox": [ + 46, + 422, + 289, + 495 + ], + "type": "inline_equation", + "content": "z_0^i" + }, + { + "bbox": [ + 46, + 422, + 289, + 495 + ], + "type": "text", + "content": " can be naturally formulated as the combination of the first frame latent code " + }, + { + "bbox": [ + 46, + 422, + 289, + 495 + ], + "type": "inline_equation", + "content": "z_0^1" + }, + { + "bbox": [ + 46, + 422, + 289, + 495 + ], + "type": "text", + "content": " and a residual item " + }, + { + "bbox": [ + 46, + 422, + 289, + 495 + ], + "type": "inline_equation", + "content": "\\Delta z^i" + }, + { + "bbox": [ + 46, + 422, + 289, + 495 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 498, + 287, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 287, + 514 + ], + "type": "interline_equation", + "content": "z _ {0} ^ {i} = z _ {0} ^ {1} + \\Delta z ^ {i}. \\tag {4}", + "image_path": "eb99fcefd098eb19f0fecfef6b90cc564d93cd99b6a8e0b767cada9f8335097f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 519, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 541 + ], + "type": "text", + "content": "Next, we construct a variable " + }, + { + "bbox": [ + 46, + 519, + 287, + 541 + ], + "type": "inline_equation", + "content": "C_t^i" + }, + { + "bbox": [ + 46, + 519, + 287, + 541 + ], + "type": "text", + "content": " by adding a scale ratio to the residual item " + }, + { + "bbox": [ + 46, + 519, + 287, + 541 + ], + "type": "inline_equation", + "content": "\\Delta z^i" + }, + { + "bbox": [ + 46, + 519, + 287, + 541 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 546, + 287, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 287, + 574 + ], + "type": "interline_equation", + "content": "C _ {t} ^ {i} = \\frac {\\sqrt {\\bar {\\alpha} _ {t}} \\Delta z ^ {i}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}. \\tag {5}", + "image_path": "4a1e101796b42237486784adb4c234885e593a1c08de2a9eaf5f51040ed0c54a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 578, + 217, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 217, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 217, + 590 + ], + "type": "text", + "content": "Thus Eq. (5) can be re-written as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 124, + 595, + 287, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 595, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 124, + 595, + 287, + 622 + ], + "type": "interline_equation", + "content": "\\Delta z ^ {i} = \\frac {\\sqrt {1 - \\bar {\\alpha} _ {t}} C _ {t} ^ {i}}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {6}", + "image_path": "c691a1f0a28f602528efa667f0d015e26f921056ddac51ed7be9c32a9fd4ce8f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 626, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 626, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 626, + 287, + 651 + ], + "type": "text", + "content": "After integrating Eq. (3) and Eq. (6) into Eq. (4), the first frame latent code " + }, + { + "bbox": [ + 46, + 626, + 287, + 651 + ], + "type": "inline_equation", + "content": "z_0^1" + }, + { + "bbox": [ + 46, + 626, + 287, + 651 + ], + "type": "text", + "content": " is denoted as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 52, + 656, + 287, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 656, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 52, + 656, + 287, + 712 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} z _ {0} ^ {1} = z _ {0} ^ {i} - \\Delta z ^ {i} = \\frac {z _ {t} ^ {i} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} ^ {i}}{\\sqrt {\\bar {\\alpha} _ {t}}} - \\frac {\\sqrt {1 - \\bar {\\alpha} _ {t}} C _ {t} ^ {i}}{\\sqrt {\\bar {\\alpha} _ {t}}} \\\\ = \\frac {z _ {t} ^ {i} - \\sqrt {1 - \\bar {\\alpha} _ {t}} (\\epsilon_ {t} ^ {i} + C _ {t} ^ {i})}{\\sqrt {\\bar {\\alpha} _ {t}}} = \\frac {z _ {t} ^ {i} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} ^ {i \\rightarrow 1}}{\\sqrt {\\bar {\\alpha} _ {t}}}, (7) \\\\ \\end{array}", + "image_path": "5ca8ba78919452be91c47edc479b52e5ce39d255551cc4654267f4bd382c9e9c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "\\epsilon_t^{i\\to 1}" + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": " is defined as the image noise prior which can be interpreted as the relationship between the first frame latent code " + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "z_0^1" + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": "-th noised frame latent code " + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "z_{t}^{i}" + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": ". The image noise prior " + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "\\epsilon_t^{i\\rightarrow 1}" + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": " is thus measured as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 380, + 129, + 545, + 157 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 129, + 545, + 157 + ], + "spans": [ + { + "bbox": [ + 380, + 129, + 545, + 157 + ], + "type": "interline_equation", + "content": "\\epsilon_ {t} ^ {i \\rightarrow 1} = \\frac {z _ {t} ^ {i} - \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} ^ {1}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}. \\tag {8}", + "image_path": "cd2e1ed968315af266c36c388730c145cbbc8b371054af6b63457b9a43f5bb81.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "spans": [ + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "text", + "content": "According to Eq. (8), the latent code " + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "inline_equation", + "content": "z_0^1" + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "text", + "content": " of first frame can be directly reconstructed through the one-step backward diffusion by using the image noise prior " + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "inline_equation", + "content": "\\epsilon_t^{i\\to 1}" + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "text", + "content": " and the noised latent code " + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "inline_equation", + "content": "z_{t}^{i}" + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "text", + "content": "-th frame. In this way, such image noise prior " + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "inline_equation", + "content": "\\epsilon_t^{i\\to 1}" + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "text", + "content": " acts as a reference noise for Gaussian noise " + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "inline_equation", + "content": "\\epsilon_t^i" + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "text", + "content": " which is added in " + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "text", + "content": "-th frame latent code, when " + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "inline_equation", + "content": "z_0^i" + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "text", + "content": " is close to " + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "inline_equation", + "content": "z_0^1" + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 165, + 547, + 250 + ], + "type": "text", + "content": "-th frame is similar to the given first frame)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 257, + 463, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 257, + 463, + 270 + ], + "spans": [ + { + "bbox": [ + 305, + 257, + 463, + 270 + ], + "type": "text", + "content": "3.3. Temporal Residual Learning" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 275, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 545, + 419 + ], + "type": "text", + "content": "Recall that the image noise prior learnt in the shortcut pathway facilitates temporal modeling to amplify the alignment between the first frame and subsequent frames. To further strengthen temporal coherence among all adjacent frames, a residual pathway is designed to estimate the residual noise of each frame through inter-frame relational reasoning. Inspired by exploring knowledge prior in pre-trained 3DUNet for video editing [58], we propose to fine-tune 3DUNet to estimate the residual noise of each frame. Technically, the estimated noise " + }, + { + "bbox": [ + 304, + 275, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}_t^i" + }, + { + "bbox": [ + 304, + 275, + 545, + 419 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 275, + 545, + 419 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 275, + 545, + 419 + ], + "type": "text", + "content": "-th frame is formulated as the combination of the reference noise " + }, + { + "bbox": [ + 304, + 275, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\epsilon_t^{i\\rightarrow 1}" + }, + { + "bbox": [ + 304, + 275, + 545, + 419 + ], + "type": "text", + "content": " (i.e., image noise prior) and the estimated residual noise " + }, + { + "bbox": [ + 304, + 275, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\Delta \\tilde{\\epsilon}_t^i" + }, + { + "bbox": [ + 304, + 275, + 545, + 419 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 366, + 427, + 545, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 427, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 366, + 427, + 545, + 443 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} _ {t} ^ {i} = \\lambda^ {i} \\epsilon_ {t} ^ {i \\rightarrow 1} + (1 - \\lambda^ {i}) \\Delta \\tilde {\\epsilon} _ {t} ^ {i}, \\tag {9}", + "image_path": "bf1ff7d7dd74dea4cd52013c052f6ba74d76171e090203319d6f87f8b767556e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 451, + 547, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 547, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 451, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\lambda^i" + }, + { + "bbox": [ + 304, + 451, + 547, + 498 + ], + "type": "text", + "content": " is a trade-off parameter. " + }, + { + "bbox": [ + 304, + 451, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\Delta \\tilde{\\epsilon}_t^i" + }, + { + "bbox": [ + 304, + 451, + 547, + 498 + ], + "type": "text", + "content": " is learnt by 3D-UNet (parameter: " + }, + { + "bbox": [ + 304, + 451, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 451, + 547, + 498 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 304, + 451, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\epsilon_t^{i\\rightarrow 1}" + }, + { + "bbox": [ + 304, + 451, + 547, + 498 + ], + "type": "text", + "content": " can be computed by Eq. (8) on the fly. Finally, the I2V denoising diffusion objective " + }, + { + "bbox": [ + 304, + 451, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\tilde{L}" + }, + { + "bbox": [ + 304, + 451, + 547, + 498 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 360, + 507, + 545, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 507, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 360, + 507, + 545, + 523 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathcal {L}} = \\mathbb {E} _ {\\epsilon \\sim \\mathcal {N} (0, I), t, c, i} [ \\| \\epsilon_ {t} ^ {i} - \\tilde {\\epsilon} _ {t} ^ {i} \\| ^ {2} ]. \\tag {10}", + "image_path": "40d3fe078fbed0a8a07eaf0dde71d51ba35dbc9965d6ebb212c4d3bcd103a51e.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 531, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 616 + ], + "type": "text", + "content": "Note that the computation of Eq. (9) is operated through the Temporal Noise Fusion (TNF) module that estimates the backward diffusion noise of each frame. Moreover, since the temporal correlation between " + }, + { + "bbox": [ + 304, + 531, + 545, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 531, + 545, + 616 + ], + "type": "text", + "content": "-th frame and the first frame will decrease when increasing frame index " + }, + { + "bbox": [ + 304, + 531, + 545, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 531, + 545, + 616 + ], + "type": "text", + "content": " with enlarged timespan, we shape the trade-off parameter " + }, + { + "bbox": [ + 304, + 531, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\lambda^i" + }, + { + "bbox": [ + 304, + 531, + 545, + 616 + ], + "type": "text", + "content": " as a linear decay parameter with respect to frame index." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 623, + 476, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 476, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 476, + 635 + ], + "type": "text", + "content": "3.4. Temporal Noise Fusion Module" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "Our residual-like dual-path noise prediction explores the image noise prior as a reference to facilitate temporal modeling among adjacent frames. However, the handcrafted design of TNF module with simple linear fusion (i.e., Eq. (9)) requires a careful tuning of hyper-parameter " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\lambda^i" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": ", leading to a sub-optimal solution. Instead, we devise a new" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8674" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 64, + 265, + 187 + ], + "blocks": [ + { + "bbox": [ + 72, + 64, + 265, + 187 + ], + "lines": [ + { + "bbox": [ + 72, + 64, + 265, + 187 + ], + "spans": [ + { + "bbox": [ + 72, + 64, + 265, + 187 + ], + "type": "image", + "image_path": "6375901c1e2ffd877e7c7ae2e0b8b0ec637cacd7929c76bb081efc47edf21450.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 51, + 190, + 282, + 202 + ], + "lines": [ + { + "bbox": [ + 51, + 190, + 282, + 202 + ], + "spans": [ + { + "bbox": [ + 51, + 190, + 282, + 202 + ], + "type": "text", + "content": "Figure 3. Detailed structure of Temporal Noise Fusion module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 206, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 287, + 243 + ], + "type": "text", + "content": "Transformer-based temporal noise fusion module to dynamically fuse reference and residual noises, aiming to further exploit the relations in between and boost noise fusion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 243, + 289, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 289, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 289, + 375 + ], + "type": "text", + "content": "Figure 3 depicts the structure of Temporal Noise Fusion module. Specifically, given the estimated residual noise " + }, + { + "bbox": [ + 46, + 243, + 289, + 375 + ], + "type": "inline_equation", + "content": "\\Delta \\tilde{\\epsilon}_t^i" + }, + { + "bbox": [ + 46, + 243, + 289, + 375 + ], + "type": "text", + "content": ", we first leverage the adaptive layer norm operations [68] which is modulated by the time step " + }, + { + "bbox": [ + 46, + 243, + 289, + 375 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 243, + 289, + 375 + ], + "type": "text", + "content": " for spatial feature normalization. After that, one self-attention layer is employed for feature enhancement, followed by the same adaptive layer normalization operation. Next, by taking the in-process intermedia feature as query and the concatenated feature of " + }, + { + "bbox": [ + 46, + 243, + 289, + 375 + ], + "type": "inline_equation", + "content": "[\\epsilon_t^{i\\rightarrow 1},\\Delta \\tilde{\\epsilon}_t^i ]" + }, + { + "bbox": [ + 46, + 243, + 289, + 375 + ], + "type": "text", + "content": " as key/value, TNF module achieves the final backward diffusion noise " + }, + { + "bbox": [ + 46, + 243, + 289, + 375 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}_{\\theta}(z_{t},t,c,i)" + }, + { + "bbox": [ + 46, + 243, + 289, + 375 + ], + "type": "text", + "content": " through a cross-attention layer:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 380, + 287, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 380, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 91, + 380, + 287, + 396 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} _ {\\theta} \\left(z _ {t}, t, c, i\\right) = \\tilde {\\epsilon} _ {t} ^ {i} = \\varphi \\left(\\Delta \\tilde {\\epsilon} _ {t} ^ {i}, \\epsilon^ {i \\rightarrow 1}, t\\right), \\tag {11}", + "image_path": "1aed18a8d1b999fa0caa3039b7914542f8374e5543ab1accb423aac133cb5d64.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 402, + 289, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 289, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 289, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 402, + 289, + 475 + ], + "type": "inline_equation", + "content": "\\varphi(\\cdot)" + }, + { + "bbox": [ + 46, + 402, + 289, + 475 + ], + "type": "text", + "content": " denotes the operation of TNF module. Compared to the simple linear fusion strategy for noise estimation, our Transformer-based TNF module sidesteps the hyper-parameter tuning and provides an elegant alternative to dynamically merge the reference and residual noises of each frame to yield high-fidelity videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 483, + 128, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 128, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 128, + 498 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 503, + 174, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 174, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 174, + 517 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": "Datasets. We empirically verify the merit of our TRIP model on WebVid-10M [1], DTDB [30] and MSR-VTT [67] datasets. Our TRIP is trained over the training set of WebVid-10M, and the validation sets of all three datasets are used for evaluation. The WebVid-10M dataset consists of about " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "10.7M" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": " video-caption pairs with a total of " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "57K" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": " video hours. " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "5K" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": " videos are used for validation and the remaining videos are utilized for training. We further sample 2,048 videos from the validation set of WebVid-10M for evaluation. DTDB contains more than " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "10K" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": " dynamic texture videos of natural scenes. We follow the standard protocols in [9] to evaluate models on a video subset derived from 4 categories of scenes (i.e., fires, clouds, vegetation, and waterfall). The texts of each category are used as the text prompt since there is no text description. For the MSR-VTT dataset, there are " + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "10K" + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": " clips in training set and 2,990" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 101, + 547, + 147 + ], + "blocks": [ + { + "bbox": [ + 305, + 64, + 547, + 97 + ], + "lines": [ + { + "bbox": [ + 305, + 64, + 547, + 97 + ], + "spans": [ + { + "bbox": [ + 305, + 64, + 547, + 97 + ], + "type": "text", + "content": "Table 1. Performances of F-Consistency (F-Consistency4: consistency among the first four frames, F-Consistencyall: consistency among all frames) and FVD on WebVid-10M." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 101, + 547, + 147 + ], + "lines": [ + { + "bbox": [ + 308, + 101, + 547, + 147 + ], + "spans": [ + { + "bbox": [ + 308, + 101, + 547, + 147 + ], + "type": "table", + "html": "
ApproachF-Consistency4 (↑)F-Consistencyall (↑)FVD (↓)
T2V-Zero [32]91.5992.15279
VideoComposer [60]88.7892.52231
TRIP95.3696.4138.9
", + "image_path": "3e59f825e397aba8bbbda3337304ad8a8897f0077c5a24607aa6004a52529a7b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 154, + 545, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 154, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 304, + 154, + 545, + 190 + ], + "type": "text", + "content": "clips in validation set. Each video clip is annotated with 20 text descriptions. We adopt the standard setting in [59] to pre-process the video clips in validation set for evaluation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 190, + 546, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 546, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 546, + 357 + ], + "type": "text", + "content": "Implementation Details. We implement our TRIP on PyTorch platform with Diffusers [57] library. The 3D-UNet [59] derived from Stable-Diffusion v2.1 [50] is exploited as the video backbone. Each training sample is a 16-frames clip and the sampling rate is 4 fps. We fix the resolution of each frame as " + }, + { + "bbox": [ + 304, + 190, + 546, + 357 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 190, + 546, + 357 + ], + "type": "text", + "content": ", which is centrally cropped from the resized video. The noise scheduler " + }, + { + "bbox": [ + 304, + 190, + 546, + 357 + ], + "type": "inline_equation", + "content": "\\{\\beta_t\\}_{t=1}^T" + }, + { + "bbox": [ + 304, + 190, + 546, + 357 + ], + "type": "text", + "content": " is set as linear scheduler (" + }, + { + "bbox": [ + 304, + 190, + 546, + 357 + ], + "type": "inline_equation", + "content": "\\beta_1 = 1 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 190, + 546, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 190, + 546, + 357 + ], + "type": "inline_equation", + "content": "\\beta_T = 2 \\times 10^{-2}" + }, + { + "bbox": [ + 304, + 190, + 546, + 357 + ], + "type": "text", + "content": "). We set the number of time steps " + }, + { + "bbox": [ + 304, + 190, + 546, + 357 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 190, + 546, + 357 + ], + "type": "text", + "content": " as 1,000 in model training. The sampling strategy for video generation is DDIM [53] with 50 steps. The 3D-UNet and TNF module are trained with AdamW optimizer (learning rate: " + }, + { + "bbox": [ + 304, + 190, + 546, + 357 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-6}" + }, + { + "bbox": [ + 304, + 190, + 546, + 357 + ], + "type": "text", + "content": " for 3D-UNet and " + }, + { + "bbox": [ + 304, + 190, + 546, + 357 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 190, + 546, + 357 + ], + "type": "text", + "content": " for TNF module). All experiments are conducted on 8 NVIDIA A800 GPUs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 358, + 547, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 547, + 503 + ], + "type": "text", + "content": "Evaluation Metrics. Since there is no standard evaluation metric for I2V task, we choose the metrics of frame consistency (F-Consistency) which is commonly adopted in video editing [58, 63] and Frechet Video Distance (FVD) [55] widely used in T2V generation [2, 39] for the evaluation on WebVid-10M. Note that the vision model used to measure F-Consistency is the CLIP ViT-L/14 [48]. Specifically, we report the frame consistency of the first 4 frames (F-Consistency" + }, + { + "bbox": [ + 304, + 358, + 547, + 503 + ], + "type": "inline_equation", + "content": "_4" + }, + { + "bbox": [ + 304, + 358, + 547, + 503 + ], + "type": "text", + "content": ") and all the 16 frames (F-Consistency" + }, + { + "bbox": [ + 304, + 358, + 547, + 503 + ], + "type": "inline_equation", + "content": "_{all}" + }, + { + "bbox": [ + 304, + 358, + 547, + 503 + ], + "type": "text", + "content": "). For the evaluation on DTDB and MSR-VTT, we follow [9, 59] and report the frame-wise Frechet Inception Distance (FID) [23] and FVD performances." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 503, + 547, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 547, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 547, + 600 + ], + "type": "text", + "content": "Human Evaluation. The user study is also designed to assess the temporal coherence, motion fidelity, and visual quality of the generated videos. In particular, we randomly sample 256 testing videos from the validation set in WebVid-10M for evaluation. Through the Amazon MTurk platform, we invite 32 evaluators, and each evaluator is asked to choose the better one from two synthetic videos generated by two different methods given the same inputs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 609, + 536, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 609, + 536, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 609, + 536, + 624 + ], + "type": "text", + "content": "4.2. Comparisons with State-of-the-Art Methods" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "type": "text", + "content": "We compare our TRIP with several state-of-the-art I2V generation methods, e.g., VideoComposer [60] and T2V-Zero [32], on WebVid-10M. For the evaluation on DTDB, we set AL [11] and cINN [9] as baselines. Text-to-Video generation advances, e.g., CogVideo [28] and Make-A-Video [52] are included for comparison on MSR-VTT." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Evaluation on WebVid-10M. Table 1 summarizes per" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8675" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 58, + 259, + 197 + ], + "blocks": [ + { + "bbox": [ + 72, + 58, + 259, + 197 + ], + "lines": [ + { + "bbox": [ + 72, + 58, + 259, + 197 + ], + "spans": [ + { + "bbox": [ + 72, + 58, + 259, + 197 + ], + "type": "image", + "image_path": "6f16fa540ecea4fdd6408090ee56ff24ee5602f70a1030184a9bd37a583e06d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 287, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 220 + ], + "type": "text", + "content": "Figure 4. Performance comparisons of F-Consistency by using different number of frames on WebVid-10M." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 248, + 287, + 299 + ], + "blocks": [ + { + "bbox": [ + 47, + 222, + 287, + 244 + ], + "lines": [ + { + "bbox": [ + 47, + 222, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 47, + 222, + 287, + 244 + ], + "type": "text", + "content": "Table 2. Performances of averaged FID and FVD over four scene categories on DTDB dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 248, + 287, + 299 + ], + "lines": [ + { + "bbox": [ + 50, + 248, + 287, + 299 + ], + "spans": [ + { + "bbox": [ + 50, + 248, + 287, + 299 + ], + "type": "table", + "html": "
ApproachZero-shotFID (↓)FVD (↓)
AL [11]No65.1934.2
cINN [9]No31.9451.6
TRIPYes24.8433.9
", + "image_path": "f7762f8a0a4a2d09546410498ff11397101226a735adc39fc788e3ae7599dc73.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 306, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 287, + 592 + ], + "type": "text", + "content": "formance comparison on WebVid-10M. Overall, our TRIP consistently leads to performance boosts against existing diffusion-based baselines across different metrics. Specifically, TRIP achieves the F-Consistency4 of 95.36%, which outperforms the best competitor T2V-Zero by 3.77%. The highest frame consistency among the first four frames of our TRIP generally validates the best temporal coherence within the beginning of the whole video. Such results basically demonstrate the advantage of exploring image noise prior as the guidance to strengthen the faithful visual alignment between given image and subsequent frames. To further support our claim, Figure 4 visualizes the frame consistency of each run which is measured among different numbers of frames. The temporal consistencies of TRIP across various frames constantly surpass other baselines. It is worthy to note that TRIP even shows large superiority when only calculating the consistency between the first two frames. This observation again verifies the effectiveness of our residual-like noise prediction paradigm that nicely preserves the alignment with given image. Considering the metric of FVD, TRIP also achieves the best performance, which indicates that the holistic motion dynamics learnt by our TRIP are well-aligned with the ground-truth video data distribution in WebVid-10M." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Figure 5 showcases four I2V generation results across three methods (VideoComposer, T2V-Zero, and our TRIP) based on the same given first frame and text prompt. Generally, compared to the two baselines, our TRIP synthesizes videos with higher quality by nicely preserving the faithful alignment with the given image & text prompt and meanwhile reflecting temporal coherence. For example, T2V-Zero produces frames that are inconsistent with the second prompt \"the ship passes close to the camera\" and VideoComposer generates temporally inconsistent frames for the" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 545, + 136 + ], + "blocks": [ + { + "bbox": [ + 312, + 57, + 539, + 66 + ], + "lines": [ + { + "bbox": [ + 312, + 57, + 539, + 66 + ], + "spans": [ + { + "bbox": [ + 312, + 57, + 539, + 66 + ], + "type": "text", + "content": "Table 3. Performances of FID and FVD on MSR-VTT dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 71, + 545, + 136 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 136 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 136 + ], + "type": "table", + "html": "
ApproachModel TypeFID (↓)FVD (↓)
CogVideo [28]T2V23.591294
Make-A-Video [52]T2V13.17-
VideoComposer [60]T2V-580
ModelScopeT2V [59]T2V11.09550
VideoComposer [60]I2V31.29208
TRIPI2V9.6891.3
", + "image_path": "39249d7376660e6f8c4dc06404c0ab333bea4633dddf58e436df8d6ca099d43d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 174, + 545, + 223 + ], + "blocks": [ + { + "bbox": [ + 306, + 148, + 545, + 170 + ], + "lines": [ + { + "bbox": [ + 306, + 148, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 306, + 148, + 545, + 170 + ], + "type": "text", + "content": "Table 4. Human evaluation of the preference ratios between TRIP v.s. other approaches on WebVid-10M dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 174, + 545, + 223 + ], + "lines": [ + { + "bbox": [ + 308, + 174, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 308, + 174, + 545, + 223 + ], + "type": "table", + "html": "
Evaluation Itemsv.s. T2V-Zerov.s. VideoComposer
Temporal Coherence96.9 v.s. 3.184.4 v.s. 15.6
Motion Fidelity93.8 v.s. 6.281.3 v.s. 18.7
Visual Quality90.6 v.s. 9.487.5 v.s. 12.5
", + "image_path": "a9a91ee204a4f50ba2db70881273d3d131a8646ac872ff26b815df5e09b20b8d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 231, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 290 + ], + "type": "text", + "content": "last prompt \"the portrait of the girl in glasses\". In contrast, the synthesized videos of our TRIP are more temporally coherent and better aligned with inputs. This again confirms the merit of temporal residual learning with amplified guidance of image noise prior for I2V generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 291, + 546, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 546, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 546, + 482 + ], + "type": "text", + "content": "Evaluation on DTDB. Next, we evaluate the generalization ability of our TRIP on DTDB dataset in a zero-shot manner. Table 2 lists the performances of averaged FID and FVD over four scene-related categories. In general, our TRIP outperforms two strong stochastic models (AL and cINN) over both two metrics. Note that cINN also exploits the residual information for I2V generation, but it combines image and residual features as clip feature by learning a bijective mapping. Our TRIP is fundamentally different in that we integrate residual-like noise estimation into the video denoising procedure to ease temporal modeling in the diffusion model. Figure 6 shows the I2V generation results of four different scenes produced by our TRIP. As shown in this figure, TRIP manages to be nicely generalized to scene-related I2V generation task and synthesizes realistic videos conditioned on the given first frame plus class label." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 483, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 579 + ], + "type": "text", + "content": "Evaluation on MSR-VTT. We further compare our TRIP with recent video generation advances on MSR-VTT. Note that here we also evaluate TRIP in a zero-shot manner. Table 3 details the frame-wise FID and FVD of different runs. Similar to the observation on DTDB, our TRIP again surpasses all baselines over both metrics. The results basically validate the strong generalization ability of our proposal to formulate the open-domain motion dynamics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 587, + 417, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 417, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 417, + 598 + ], + "type": "text", + "content": "4.3. Human Evaluation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "In addition to the evaluation over automatic metrics, we also perform human evaluation to investigate user preference with regard to three perspectives (i.e., temporal coherence, motion fidelity, and visual quality) across different I2V approaches. Table 4 shows the comparisons of user preference on the generated videos over WebVid-10M dataset. Overall, our TRIP is clearly the winner in terms of all three criteria compared to both baselines (VideoComposer and T2V-Zero). This demonstrates the powerful temporal modeling" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8676" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 56, + 292, + 289 + ], + "blocks": [ + { + "bbox": [ + 85, + 56, + 292, + 289 + ], + "lines": [ + { + "bbox": [ + 85, + 56, + 292, + 289 + ], + "spans": [ + { + "bbox": [ + 85, + 56, + 292, + 289 + ], + "type": "image", + "image_path": "2735994c1c4f7aa3ba0a31c6b543e61c369857720f6e3bcc3d6e9e77dddba9d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 293, + 545, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 545, + 316 + ], + "type": "text", + "content": "Figure 5. Examples of I2V generation results by three methods (VideoComposer, T2V-Zero, and our TRIP) on WebVid-10M dataset. We uniformly sample four frames of each generated video for visualization." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 299, + 56, + 507, + 289 + ], + "blocks": [ + { + "bbox": [ + 299, + 56, + 507, + 289 + ], + "lines": [ + { + "bbox": [ + 299, + 56, + 507, + 289 + ], + "spans": [ + { + "bbox": [ + 299, + 56, + 507, + 289 + ], + "type": "image", + "image_path": "c16510228d637cecab59ab75ff35e778f49b803ebd9b7a4438e73f76ab8e15f6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 61, + 321, + 272, + 456 + ], + "blocks": [ + { + "bbox": [ + 61, + 321, + 272, + 456 + ], + "lines": [ + { + "bbox": [ + 61, + 321, + 272, + 456 + ], + "spans": [ + { + "bbox": [ + 61, + 321, + 272, + 456 + ], + "type": "image", + "image_path": "b7eb760c442529b6482b1c25a0fb23d5512bf99e40c428205b1aba56888fe6e2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 458, + 286, + 479 + ], + "lines": [ + { + "bbox": [ + 47, + 458, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 458, + 286, + 479 + ], + "type": "text", + "content": "Figure 6. Image-to-video generation results of four different natural scenes by our TRIP on DTDB dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 489, + 286, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 286, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 286, + 513 + ], + "type": "text", + "content": "of motion dynamics through our residual-like noise prediction with image noise prior." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 525, + 185, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 185, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 185, + 537 + ], + "type": "text", + "content": "4.4. Ablation Study on TRIP" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 544, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 286, + 581 + ], + "type": "text", + "content": "In this section, we perform ablation study to delve into the design of TRIP for I2V generation. Here all experiments are conducted on WebVid-10M for performance comparison." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": "First Frame Condition. We first investigate different condition strategies to exploit the given first frame as additional condition for I2V generation. Table 5 summarizes the performance comparisons among different variants of our TRIP. " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{TRIP}_C" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": " follows the common recipe [60] to concatenate the clean image latent code of first frame with the noise video latent code along channel dimension. " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{TRIP}_{TE}" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": " is implemented by concatenating image latent code at the end of the noise video latent code along temporal dimension, while our proposal (TRIP) temporally concatenates image latent code at the beginning of video latent code. In par" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 346, + 545, + 399 + ], + "blocks": [ + { + "bbox": [ + 306, + 321, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 306, + 321, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 321, + 545, + 342 + ], + "type": "text", + "content": "Table 5. Performance comparisons among different condition approaches on WebVid-10M dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 346, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 308, + 346, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 346, + 545, + 399 + ], + "type": "table", + "html": "
ModelF-Consistency4 (↑)F-Consistencyall (↑)FVD (↓)
TRIPC94.7796.1341.3
TRIPTE95.1796.2039.8
TRIP95.3696.4138.9
", + "image_path": "bc19f61505e6fea799b69ea2e6a21beba6b0cd072bee8206661b88d795bc2a5f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 430, + 545, + 483 + ], + "blocks": [ + { + "bbox": [ + 306, + 403, + 545, + 425 + ], + "lines": [ + { + "bbox": [ + 306, + 403, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 545, + 425 + ], + "type": "text", + "content": "Table 6. Evaluation of temporal residual learning in terms of F-Consistency and FVD on WebVid-10M dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 430, + 545, + 483 + ], + "lines": [ + { + "bbox": [ + 308, + 430, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 308, + 430, + 545, + 483 + ], + "type": "table", + "html": "
ModelF-Consistency4 (↑)F-Consistencyall (↑)FVD (↓)
TRIP-94.6695.9239.9
TRIPW95.2295.9643.0
TRIP95.3696.4138.9
", + "image_path": "5f06d0f3ffa4cb1722897c11c0c315dc4de492911296b316fe03de8a7a2fb601.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "content": "ticular, " + }, + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{TRIP}_{TE}" + }, + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "content": " exhibits better performances than " + }, + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{TRIP}_C" + }, + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "content": ", which indicates that temporal concatenation is a practical choice for augmenting additional condition in I2V generation. TRIP further outperforms " + }, + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{TRIP}_{TE}" + }, + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "content": " across both F-Consistency and FVD metrics. Compared to concatenating image latent code at the end of video, augmenting video latent code with input first frame latent code at the start position basically matches with the normal temporal direction, thereby facilitating temporal modeling." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Temporal Residual Learning. Next, we study how each design of our TRIP influences the performances of both temporal consistency and visual quality for I2V generation. Here we include two additional variants of TRIP: (1) " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{TRIP}^{-}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " removes the shortcut path of temporal residual learning and only capitalizes on 3D-UNet for noise prediction; (2) " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{TRIP}^{W}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " is a degraded variant of temporal residual learning with image noise prior by simply fusing reference and residual noise via linear fusion strategy. As shown in Table 6, " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{TRIP}^{W}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " exhibits better performances than " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{TRIP}^{-}" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8677" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 70, + 294, + 113 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 294, + 113 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 294, + 113 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 294, + 113 + ], + "type": "image", + "image_path": "f9067e6bd7cca475750b152828b1dd38a34356cda17a544b0762e20b4a949640.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 113, + 254, + 118 + ], + "lines": [ + { + "bbox": [ + 124, + 113, + 254, + 118 + ], + "spans": [ + { + "bbox": [ + 124, + 113, + 254, + 118 + ], + "type": "text", + "content": "A castle in the middle of a forest with a river running through it" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 296, + 70, + 506, + 113 + ], + "blocks": [ + { + "bbox": [ + 296, + 70, + 506, + 113 + ], + "lines": [ + { + "bbox": [ + 296, + 70, + 506, + 113 + ], + "spans": [ + { + "bbox": [ + 296, + 70, + 506, + 113 + ], + "type": "image", + "image_path": "b87c206034da2155825c89c1103dc9138285fa045afa37101baadb0eb4a35cc9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 113, + 473, + 118 + ], + "lines": [ + { + "bbox": [ + 331, + 113, + 473, + 118 + ], + "spans": [ + { + "bbox": [ + 331, + 113, + 473, + 118 + ], + "type": "text", + "content": "Crabs with bubbles crawls on the beach with sunset in the background" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 85, + 119, + 294, + 162 + ], + "blocks": [ + { + "bbox": [ + 85, + 119, + 294, + 162 + ], + "lines": [ + { + "bbox": [ + 85, + 119, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 85, + 119, + 294, + 162 + ], + "type": "image", + "image_path": "8e4618d020e042182948f7ada9891254b3774754940ae03467dd3467bd555c8a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 162, + 211, + 167 + ], + "lines": [ + { + "bbox": [ + 166, + 162, + 211, + 167 + ], + "spans": [ + { + "bbox": [ + 166, + 162, + 211, + 167 + ], + "type": "text", + "content": "A ship sails in the sea." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 296, + 119, + 506, + 162 + ], + "blocks": [ + { + "bbox": [ + 296, + 119, + 506, + 162 + ], + "lines": [ + { + "bbox": [ + 296, + 119, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 296, + 119, + 506, + 162 + ], + "type": "image", + "image_path": "76d9677ef005c7c27bef7306a98c182645d2fede00bc546df2e82894ebc36d3c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 162, + 449, + 167 + ], + "lines": [ + { + "bbox": [ + 357, + 162, + 449, + 167 + ], + "spans": [ + { + "bbox": [ + 357, + 162, + 449, + 167 + ], + "type": "text", + "content": "WALL-E on the beach with a plant in his hand" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 85, + 168, + 294, + 211 + ], + "blocks": [ + { + "bbox": [ + 85, + 168, + 294, + 211 + ], + "lines": [ + { + "bbox": [ + 85, + 168, + 294, + 211 + ], + "spans": [ + { + "bbox": [ + 85, + 168, + 294, + 211 + ], + "type": "image", + "image_path": "5e459fa9626d9d84dda509f10dae4efe49028429e38615bbebe7e775d5c11d67.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 211, + 253, + 217 + ], + "lines": [ + { + "bbox": [ + 123, + 211, + 253, + 217 + ], + "spans": [ + { + "bbox": [ + 123, + 211, + 253, + 217 + ], + "type": "text", + "content": "Halloweep pumpkin with glowing eyes and smoke coming out of it" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 296, + 168, + 506, + 210 + ], + "blocks": [ + { + "bbox": [ + 296, + 168, + 506, + 210 + ], + "lines": [ + { + "bbox": [ + 296, + 168, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 296, + 168, + 506, + 210 + ], + "type": "image", + "image_path": "8527d2a2d525e666fb73bd1ed5995137a017fab990a4cadd3a63f2a0d6e075da.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 211, + 456, + 217 + ], + "lines": [ + { + "bbox": [ + 345, + 211, + 456, + 217 + ], + "spans": [ + { + "bbox": [ + 345, + 211, + 456, + 217 + ], + "type": "text", + "content": "A bear is playing a guitar in front of a tank in the snow." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 53, + 247, + 278, + 361 + ], + "blocks": [ + { + "bbox": [ + 46, + 220, + 545, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 545, + 243 + ], + "type": "text", + "content": "Figure 7. Customized image animation with Stable-Diffusion XL [47] for text-to-video generation. We first employ Stable-Diffusion XL to synthesize fancy images and then exploit TRIP for image animation. Six generated videos with corresponding text prompts are presented." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 247, + 278, + 361 + ], + "lines": [ + { + "bbox": [ + 53, + 247, + 278, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 278, + 361 + ], + "type": "image", + "image_path": "04b4d4e3592b1d6ac4e85f3cebddfc169199d84b6d1fa4b327c3a1faf9158980.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 363, + 286, + 395 + ], + "lines": [ + { + "bbox": [ + 47, + 363, + 286, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 286, + 395 + ], + "type": "text", + "content": "Figure 8. Visualization of I2V example with text prompt \"Young male tourist taking a photograph of old city on sunset\" by using variants of our TRIP." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 406, + 286, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 286, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 286, + 502 + ], + "type": "text", + "content": "in terms of F-consistency. The results highlight the advantage of leveraging image noise prior as the reference to amplify the alignment between the first frame and subsequent frames. By further upgrading temporal residual learning " + }, + { + "bbox": [ + 46, + 406, + 286, + 502 + ], + "type": "inline_equation", + "content": "(\\mathbf{TRIP}^W)" + }, + { + "bbox": [ + 46, + 406, + 286, + 502 + ], + "type": "text", + "content": " with Transformer-based temporal noise fusion, our TRIP achieves the best performances across all metrics. Figure 8 further showcases the results of different variants, which clearly confirm the effectiveness of our designs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 514, + 272, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 272, + 527 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 272, + 527 + ], + "type": "text", + "content": "4.5. Application: Customized Image Animation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "In this section, we provide two interesting applications of customized image animation by directly extending TRIP in a tuning-free manner. The first application is text-to-video pipeline that first leverages text-to-image diffusion models (e.g., Stable-Diffusion XL [47]) to synthesize fancy images and then employs our TRIP to animate them as videos. Figure 7 shows six text-to-video generation examples by using Stable-Diffusion XL and TRIP. The results generally demonstrate the generalization ability of our TRIP when taking text-to-image results as additional reference for text-to-video generation. To further improve controllability of text-to-video generation in real-world scenarios, we can additionally utilize image editing models (e.g., InstructPix2Pix [3] and ControlNet [74]) to modify the visual content of input images and then feed them into TRIP" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 318, + 249, + 532, + 293 + ], + "blocks": [ + { + "bbox": [ + 318, + 249, + 532, + 293 + ], + "lines": [ + { + "bbox": [ + 318, + 249, + 532, + 293 + ], + "spans": [ + { + "bbox": [ + 318, + 249, + 532, + 293 + ], + "type": "image", + "image_path": "0c9fd19058f83c28b6cf0841be39307611a30aa28e7e9541f1c688cff9f994ab.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 318, + 293, + 532, + 342 + ], + "blocks": [ + { + "bbox": [ + 318, + 293, + 532, + 342 + ], + "lines": [ + { + "bbox": [ + 318, + 293, + 532, + 342 + ], + "spans": [ + { + "bbox": [ + 318, + 293, + 532, + 342 + ], + "type": "image", + "image_path": "c091633088a02362fb475a8f07df76d740e33e2524af34e8bb6b4af3e0bace66.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 318, + 342, + 532, + 390 + ], + "blocks": [ + { + "bbox": [ + 318, + 342, + 532, + 390 + ], + "lines": [ + { + "bbox": [ + 318, + 342, + 532, + 390 + ], + "spans": [ + { + "bbox": [ + 318, + 342, + 532, + 390 + ], + "type": "image", + "image_path": "7f9144f2263ecb43d5143464ef8c9aaad5aa40da93e6c436284ab0a64fc5e17b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 390, + 471, + 397 + ], + "lines": [ + { + "bbox": [ + 381, + 390, + 471, + 397 + ], + "spans": [ + { + "bbox": [ + 381, + 390, + 471, + 397 + ], + "type": "text", + "content": "A beautiful duck swims on the pond" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 400, + 545, + 456 + ], + "lines": [ + { + "bbox": [ + 305, + 400, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 400, + 545, + 456 + ], + "type": "text", + "content": "Figure 9. Customized image animation with image editing models (e.g., InstructPix2Pix [3] and ControlNet [74]). The first row shows image animation results given the primary real-world first image. The second and third rows show I2V results given the edited images via InstructPix2Pix and ControlNet, respectively." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 464, + 545, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 464, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 305, + 464, + 545, + 500 + ], + "type": "text", + "content": "for video generation. As shown in Figure 9, our TRIP generates vivid videos based on the edited images with promising temporal coherence and visual quality." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 513, + 383, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 513, + 383, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 513, + 383, + 525 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "This paper explores inherent inter-frame correlation in diffusion model for image-to-video generation. Particularly, we study the problem from a novel viewpoint of leveraging temporal residual learning with reference to the image noise prior of given image for enabling coherent temporal modeling. To materialize our idea, we have devised TRIP, which executes image-conditioned noise prediction in a residual fashion. One shortcut path is novelly designed to learn image noise prior based on given image and noised video latent code, leading to reference noise for enhancing alignment between synthesized frames and given image. Meanwhile, another residual path is exploited to estimate the residual noise of each frame through eased temporal modeling. Experiments conducted on three datasets validate the superiority of our proposal over state-of-the-art approaches." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8678" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. In ICCV, 2021. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. In CVPR, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 201 + ], + "type": "text", + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning to Follow Image Editing Instructions. In CVPR, 2023. 1, 2, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 236 + ], + "type": "text", + "content": "[4] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. StableVideo: Text-driven Consistency-aware Diffusion Video Editing. In ICCV, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 288, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 278 + ], + "type": "text", + "content": "[5] Zhikai Chen, Fuchen Long, Zhaofan Qiu, Ting Yao, Wengang Zhou, Jiebo Luo, and Tao Mei. AnchorFormer: Point Cloud Completion from Discriminative Nodes. In CVPR, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 280, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 323 + ], + "type": "text", + "content": "[6] Paul Couairon, Clément Rambour, Jean-Emmanuel Haugeard, and Nicolas Thome. VidEdit: Zero-Shot and Spatially Aware Text-Driven Video Editing. arXiv preprint arXiv:2306.08707, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 324, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 346 + ], + "type": "text", + "content": "[7] Prafulla Dhariwal and Alex Nichol. Diffusion Models Beat GANs on Image Synthesis. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 347, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 288, + 390 + ], + "type": "text", + "content": "[8] Ziyi Dong, Pengxu Wei, and Liang Lin. DreamArtist: Towards Controllable One-Shot Text-to-Image Generation via Positive-Negative Prompt-Tuning. arXiv preprint arXiv:2211.11337, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 392, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 435 + ], + "type": "text", + "content": "[9] Michael Dorkenwald, Timo Milbich, Andreas Blattmann, Robin Rombach, Konstantinos G. Derpanis, and Björn Ommer. Stochastic Image-to-Video Synthesis using cINNs. In CVPR, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 436, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 436, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 436, + 287, + 501 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "type": "text", + "content": "[11] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating Landscape: Self-Supervised Learning of Decoupled Motion and Appearance for Single-Image Video Synthesis. TOG, 2019. 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 547, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 580 + ], + "type": "text", + "content": "[12] Siming Fan, Jingtan Piao, Chen Qian, Kwan-Yee Lin, and Hongsheng Li. Simulating Fluids in Real-World Still Images. In ICCV, 2023. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "text", + "content": "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashni, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion. In ICLR, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "text", + "content": "[14] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve Your Own Correlation: A Noise Prior for Video Diffusion Models. In ICCV, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[15] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. TokenFlow: Consistent Diffusion Features for Consistent Video Editing. arXiv preprint arXiv:2307.10373, 2023. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[16] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative Adversarial Networks. In NeurIPS, 2014. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[17] Xianfan Gu, Chuan Wen, Jiaming Song, and Yang Gao. Seer: Language Instructed Video Prediction with Latent Diffusion Models. arXiv preprint arXiv:2303.14897, 2023. 1, 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[18] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning. arXiv preprint arXiv:2307.04725, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "text", + "content": "[19] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible Diffusion Modeling of Long Videos. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "type": "text", + "content": "[21] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, and Qifeng Chen. animate-A-Story: Storytelling with Retrieval-Augmented Video Generation. arXiv preprint arXiv:2307.06940, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 320, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 352 + ], + "type": "text", + "content": "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-Prompt Image Editing with Cross-Attention Control. In ICLR, 2023. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "text", + "content": "[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. In NeuIPS, 2017. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 399, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 420 + ], + "type": "text", + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In NeurIPS, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "type": "text", + "content": "[25] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen Video: High Definition Video Generation with Diffusion Models. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "type": "text", + "content": "[26] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video Diffusion Models. In NeurIPS, 2022. 1, 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "type": "text", + "content": "[27] Aleksander Holynski, Brian Curless, Steven M. Seitz, and Richard Szeliski. Animating Pictures with Eulerian Motion Fields. In CVPR, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 544, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 578 + ], + "type": "text", + "content": "[28] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. CogVideo: Large-Scale Pretraining for Text-to-Video Generation via Transformers. In ICLR, 2023. 5, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "type": "text", + "content": "[29] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely Connected Convolutional Networks. In CVPR, 2017. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 612, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 644 + ], + "type": "text", + "content": "[30] Hadji Isma and Wildes Richard P. A New Large Scale Dynamic Texture Dataset with Application to Convnet Understanding. In ECCV, 2018. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 646, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 689 + ], + "type": "text", + "content": "[31] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. DreamPose: Fashion Image-to-Video Synthesis via Stable Diffusion. In ICCV, 2023. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 690, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 713 + ], + "type": "text", + "content": "[32] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8679" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 105 + ], + "type": "text", + "content": "Navasardyan, and Humphrey Shi. Text2Video-Zero: Text-to-Image Diffusion Models are Zero-Shot Video Generators. In ICCV, 2023. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 286, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 286, + 138 + ], + "type": "text", + "content": "[33] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-Concept Customization of Text-to-Image Diffusion. In CVPR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 286, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 286, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 286, + 172 + ], + "type": "text", + "content": "[34] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Flow-Grounded Spatial-Temporal Video Prediction from Still Images. In ECCV, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 173, + 286, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 286, + 205 + ], + "type": "text", + "content": "[35] Yehao Li, Ting Yao, Yingwei Pan, and Tao Mei. Contextual Transformer Networks for Visual Recognition. IEEE Trans. on PAMI, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 205, + 286, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 286, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 286, + 237 + ], + "type": "text", + "content": "[36] Fuchen Long, Zhaofan Qiu, Yingwei Pan, Ting Yao, Jiebo Luo, and Tao Mei. Stand-Alone Inter-Frame Attention in Video Models. In CVPR, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 239, + 286, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 286, + 270 + ], + "type": "text", + "content": "[37] Fuchen Long, Zhaofan Qiu, Yingwei Pan, Ting Yao, Chong-Wah Ngo, and Tao Mei. Dynamic Temporal Filtering in Video Models. In ECCV, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 272, + 286, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 286, + 304 + ], + "type": "text", + "content": "[38] Fuchen Long, Zhaofan Qiu, Ting Yao, and Tao Mei. Video-Drafter: Content-Consistent Multi-Scene Video Generation with LLM. arXiv:2401.01256, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 305, + 286, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 286, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 286, + 349 + ], + "type": "text", + "content": "[39] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. VideoFusion: Decomposed Diffusion Models for High-Quality Video Generation. In CVPR, 2023. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 350, + 286, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 286, + 380 + ], + "type": "text", + "content": "[40] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable Animation of Fluid Elements in Still Images. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 382, + 286, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 286, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 286, + 414 + ], + "type": "text", + "content": "[41] Babaeizadeh Mohammad, Finn Chelsea, Erhan Dumitru, Campbell Roy H, and Levine Sergey. Stochastic Variational Video Prediction. In ICLR, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 415, + 286, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 286, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 286, + 469 + ], + "type": "text", + "content": "[42] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2I-Adapter: Learning Adapters to Dig out More Controllable Ability for Text-to-Image Diffusion Models. arXiv preprint arXiv:2302.08453, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 471, + 286, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 286, + 513 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 286, + 513 + ], + "type": "text", + "content": "[43] Haomiao Ni, Changhao Shi, Kai Li, Sharon X. Huang, and Martin Renqiang Min. Conditional Image-to-Video Generation with Latent Flow Diffusion Models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 515, + 286, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 286, + 568 + ], + "type": "text", + "content": "[44] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. In ICML, 2022. 1, 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 570, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 286, + 624 + ], + "type": "text", + "content": "[45] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. CoDeF: Content Deformation Fields for Temporally Consistent Video Processing. arXiv preprint arXiv:2308.07926, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 625, + 286, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 286, + 657 + ], + "type": "text", + "content": "[46] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot Image-to-Image Translation. In ACM SIGGRAPH, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "text", + "content": "[47] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. arXiv preprint arXiv:2307.01952, 2023. 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "text", + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 139, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 139, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 139, + 545, + 182 + ], + "type": "text", + "content": "[49] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv preprint arXiv:2204.06125, 2022.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 183, + 545, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 183, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 545, + 225 + ], + "type": "text", + "content": "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022. 1, 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 227, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 227, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 227, + 545, + 270 + ], + "type": "text", + "content": "[51] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 272, + 545, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 325 + ], + "type": "text", + "content": "[52] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-Video: Text-to-Video Generation without Text-Video Data. In ICLR, 2023. 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 327, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 348 + ], + "type": "text", + "content": "[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models. In ICLR, 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 350, + 545, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 390 + ], + "type": "text", + "content": "[54] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Palur. A Closer Look at Spatiotemporal Convolutions for Action Recognition. In CVPR, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 392, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 392, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 392, + 545, + 436 + ], + "type": "text", + "content": "[55] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new Metric for Video Generation. In ICLR Deep-GenStruct Workshop, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "type": "text", + "content": "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 471, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 513 + ], + "type": "text", + "content": "[57] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-Of-The-Art Diffusion Models. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 514, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 545, + 558 + ], + "type": "text", + "content": "[58] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-L-Video: Multi-Text to Long Video Generation via Temporal Co-Denoising. arXiv preprint arXiv:2305.18264, 2023. 4, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 559, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 601 + ], + "type": "text", + "content": "[59] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. ModelScope Text-to-Video Technical Report. arXiv preprint arXiv:2308.06571, 2023. 2, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 603, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 655 + ], + "type": "text", + "content": "[60] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional Video Synthesis with Motion Controllability. In NeurIPS, 2023. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "type": "text", + "content": "[61] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent Image Animator: Learning to Animate Images via Latent Space Navigation. In ICLR, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[62] Yaohui Wang, Xin Ma, Xinyuan Chen, Antitza Dantcheva, Bo Dai, and Yu Qiao. LEO: Generative Latent Image" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8680" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 611 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "Animator for Human Video Synthesis. arXiv preprint arXiv:2305.03989, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[63] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-A-Video: One-Shot Tuning of Image Diffusion Models for Text-to-Video Generation. In ICCV, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 194 + ], + "type": "text", + "content": "[64] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A Survey on Video Diffusion Models. arXiv preprint arXiv:2310.10647, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 228 + ], + "type": "text", + "content": "[65] Zhen Xing, Qi Dai, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. SimDA: Simple Diffusion Adapter for Efficient Video Generation. In CVPR, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 230, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 272 + ], + "type": "text", + "content": "[66] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to Generate Time-Lapse Videos Using Multi-Stage Dynamic Generative Adversarial Networks. In CVPR, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 274, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 308 + ], + "type": "text", + "content": "[67] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A Large Video Description Dataset for Bridging Video and Language. In CVPR, 2016. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "text", + "content": "[68] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and Improving Layer Normalization. In NeurIPS, 2019. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 343, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 386 + ], + "type": "text", + "content": "[69] Hanshu Yan, Jun Hao Liew, Long Mai, Shanchuan Lin, and Jiashi Feng. MagicProp: Diffusion-based Video Editing via Motion-aware Appearance Propagation. arXiv preprint arXiv:2309.00908, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "type": "text", + "content": "[70] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion Probabilistic Modeling for Video Generation. arXiv preprint arXiv:2203.09481, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 422, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 454 + ], + "type": "text", + "content": "[71] Ting Yao, Yingwei Pan, Yehao Li, Chong-Wah Ngo, and Tao Mei. Wave-ViT: Unifying Wavelet and Transformers for Visual Representation Learning. In ECCV, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "type": "text", + "content": "[72] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. DragNUWA: Fine-grained Control in Video Generation by Integrating Text, Image, and Trajectory. arXiv preprint arXiv:2308.08089, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 501, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 543 + ], + "type": "text", + "content": "[73] Jianfeng Zhang, Hanshu Yan, Zhongcong Xu, Jiashi Feng, and Jun Hao Liew. MagicAvatar: Multimodal Avatar Generation and Animation. arXiv preprint arXiv:2308.14748, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "text", + "content": "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In ICCV, 2023. 2, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 579, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 611 + ], + "type": "text", + "content": "[75] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks. In ICCV, 2017. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "8681" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/fa811509-bdf9-4c56-97db-551fcc2b0a3d_content_list.json b/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/fa811509-bdf9-4c56-97db-551fcc2b0a3d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4efa3813aae41ec9039ea93c211ff3bd3813e610 --- /dev/null +++ b/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/fa811509-bdf9-4c56-97db-551fcc2b0a3d_content_list.json @@ -0,0 +1,1585 @@ +[ + { + "type": "text", + "text": "TTA-EVF: Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation", + "text_level": 1, + "bbox": [ + 78, + 130, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hoonhee Cho, Taewoo Kim, Yuhwan Jeong, and Kuk-Jin Yoon KAIST", + "bbox": [ + 233, + 203, + 736, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{gnsgnsgml, intelpro, jeongyh98, kjyoon}@kaist.ac.kr", + "bbox": [ + 264, + 241, + 699, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Frame Interpolation (VFI), which aims at generating high-frame-rate videos from low-frame-rate inputs, is a highly challenging task. The emergence of bio-inspired sensors known as event cameras, which boast microsecond-level temporal resolution, has ushered in a transformative era for VFI. Nonetheless, the application of event-based VFI techniques in domains with distinct environments from the training data can be problematic. This is mainly because event camera data distribution can undergo substantial variations based on camera settings and scene conditions, presenting challenges for effective adaptation. In this paper, we propose a test-time adaptation method for event-based VFI to address the gap between the source and target domains. Our approach enables sequential learning in an online manner on the target domain, which only provides low-frame-rate videos. We present an approach that leverages confident pixels as pseudo ground-truths, enabling stable and accurate online learning from low-frame-rate videos. Furthermore, to prevent overfitting during the continuous online process where the same scene is encountered repeatedly, we propose a method of blending historical samples with current scenes. Extensive experiments validate the effectiveness of our method, both in cross-domain and continuous domain shifting setups. The code is available at https://github.com/Chohoonhee/TTA-EVF.", + "bbox": [ + 76, + 325, + 473, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 737, + 209, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Frame Interpolation (VFI) is a well-established problem in the field of computer vision, aiming to enhance the temporal resolution of videos. In recent times, deep learning-based VFI approaches [2, 3, 15, 20-22, 33, 35, 37, 38, 50, 60] have achieved remarkable performances across various benchmark datasets. However, accurate motion estimation becomes challenging in scenes with complex and non-linear motions due to the lack of information, often resulting in the generation of erroneous inter frames.", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3d32e0b745fa88d50c391b586107d24461690ab16bdd21c71a45d404dc85ceb5.jpg", + "image_caption": [ + "Figure 1. TTA-EVF efficiently produces high-quality results by adapting the network to the target domain in an online manner, alleviating the need for offline data supply and addressing the performance degradation observed when applying a well-trained event-based VFI network to a domain with a different distribution." + ], + "image_footnote": [], + "bbox": [ + 516, + 290, + 875, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the inherent lack of intermediate information in frame cameras, recent researches [16, 24, 51, 52, 56] have explored the potential of event cameras as a promising solution for VFI, especially in scenarios involving complex motion. Event cameras sense dynamic changes in pixel intensity, triggering an event when the change surpasses a predefined threshold. The incorporation of event cameras as additional devices in VFI mitigates the challenges associated with modeling complex motion.", + "bbox": [ + 496, + 638, + 893, + 775 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Typically, frame-based VFI methods [17, 33, 38] are evaluated on test datasets (e.g., SNU-FILM [12], X4K1000FPS [43]) with distributions different from their training dataset (e.g., Vimeo90K [61]). Frame-based VFI demonstrates a certain level of generalization ability even without additional modules, providing evidence that it can be applied to real scenario applications, such as dealing with changes in camera devices. These device and environment", + "bbox": [ + 496, + 779, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 1, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25701", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "variations are inherent in VFI, which must learn from High-Frame-Rate (HFR) cameras and perform inference on Low-Frame-Rate (LFR) cameras, making them unavoidable.", + "bbox": [ + 76, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, when training event-based VFI using a source dataset and testing on a different domain, there is a significant performance drop compared to frame-based approaches. This is attributed not only to the inherent noise from external environment (e.g., illumination) but also to the distribution variations due to different camera settings (e.g., event trigger threshold). To address this, one can utilize self-supervised learning [16, 39] with low-frame-rate and event data from the target domain, but there is a drawback in applying them online. In practical application, event cameras exhibit continuous changes in distribution with variations in lighting conditions. It is not feasible to acquire new data for novel scenes each time they arise and wait for the network to learn offline. In this regard, an online adaptation of the network to the environment is essential from a practical perspective and is well-known as Test-Time Adaptation (TTA) [47, 53, 63]. TTA assumes a setting that does not allow access to source data and prefers online learning for adaptation. In our case, the target domain lacks ground-truth data, which means no high-frame-rate video is available in the target domain.", + "bbox": [ + 76, + 138, + 470, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose Test-time Adaptation for Event-based Video Frame Interpolation (TTA-EVF) to alleviate performance degradation in the face of domain shifts. As shown in Fig. 1, when encountering a new domain without performing adaptation, even well-trained networks exhibit performance drops. In contrast, TTA-EVF reliably and rapidly fine-tunes in the target domain through self-training, even without high-frame-rate videos. For TTA in event-based VFI, there are two main challenges: Firstly, even without exposure to HFR videos in the target domain, the network should adapt to the scene while simultaneously learning knowledge for HFR generation from LFR data. To address this challenge, we propose Reliable Pixel Sampling (RPS), a self-training scheme for selecting pixels from generated HFR videos, utilizing a confidence estimation based on a teacher-student framework [49]. The second challenge is overfitting when continuously provided with data that has a similar distribution within a successive sequence within online learning. To mitigate this, we employ a memory bank that stores only reliable samples from the generated images. We then propose Patch-Mixed Sampling (PMS) by blending reliable samples with current images, effectively preventing overfitting through an augmentation approach.", + "bbox": [ + 76, + 459, + 468, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of our works are as follows: (I) We propose a novel Test-Time Adaptation (TTA) framework for event-based VFI. (II) We introduce a novel EventRGB Distribution Shift (ERDS) dataset for event-based video frame interpolation, where the distribution of the dataset is continually changing. (III) We compare our", + "bbox": [ + 76, + 810, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "method with existing works and demonstrate its superiority by outperforming supervised methods.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 135, + 650, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event-based Video Frame Interpolation. Frame-based VFI methods [2, 3, 15, 20-22, 33, 35, 37, 38, 50, 60] have continuously achieved performance improvements; however, it still has weaknesses in handling complex motion and requires motion approximation when synthesizing intermediate frames at arbitrary time. Event-based VFI methods [30, 41, 44, 45, 55, 59, 64, 65], which utilize event cameras as an additional sensor [6, 7, 9, 28, 34, 45], enable not only handling complex motion but also interpolating intermediate frames for irregular objects. However, there is still room for improvement compared to frame-based VFI methods. Typically, frame-based VFI methods perform cross-domain performance evaluations, considering real scenario applications. Well-trained frame-based VFI methods exhibit domain generalization ability, maintaining considerable performance even on different domains without the need for additional training process for adaptation. On the other hand, event-based VFI experiences a significant drop in performance when moving to different domains [16]. Therefore, resolving cross-domain challenges in event-based VFI is vital for widespread applications.", + "bbox": [ + 496, + 160, + 890, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Test-Time Adaptation. Unlike Unsupervised Domain Adaptation [8, 48, 58], which necessitates offline learning for each domain change and relies on labeled data from the source domain, Test-Time Adaptation (TTA) is more challenging setup as it adapts to the target domain without accessing source data. In recent times, the significance of data privacy has escalated, leading to increased interest in TTA, a source-free domain adaptation approach. This is primarily due to the inefficiency of accessing source domain data during inference. TTA is actively being researched in the image domain [5, 14, 19, 27, 31, 32, 36, 40, 42, 47, 53, 63, 66], especially in areas like object recognition and segmentation. While there is an approach of its application in event-based visual recognition, Ev-TTA [23], there is still a lack of research in the domain of video frame interpolation, particularly in relation to events. The existing work, MetaVFI [11], proposed for the frame-based VFI task, learns to estimate scene-adaptive motion with minimal parameter updates through meta-learning [13]. However, this method is well-suited when there is not a significant domain gap between training and testing. We demonstrate that this approach is suboptimal when applied to event-based VFI with a substantial domain gap between the training and test sets.", + "bbox": [ + 496, + 478, + 890, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose the first-ever Test-Time Adaptation (TTA) approach for event-based Video Frame Interpolation (VFI). This is essential because event distributions can vary significantly due to external environmental factors and camera parameter settings even when using the same camera.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25702", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1910b9000bc1557cb3ec1f47454fba3dd0162dbe3dc4b0155ad053fc59d0192e.jpg", + "image_caption": [ + "Figure 2. Overview of our Test-Time Adaptation for Event-based Video Frame Interpolation (TTA-EVF) framework. Our approach involves four steps for test-time training in Ev-VFI: 1) We calculate the adaptive threshold, $\\rho$ , for RPS using the low frame rate image triples, $I_0, I_1$ , and $I_2$ with interval events $E_{0\\rightarrow 1}$ and $E_{2\\rightarrow 1}$ . 2) Reliable Pixel Sampling (RPS) is applied for arbitrary $t_0$ between 0 and 1, and arbitrary $t_1$ between 1 and 2, and in this process, even without intermediate images for training, reliable pixels are estimated through the student-teacher framework and self-training is performed. 3) Using the estimated intermediate frames $\\hat{I}_{t_0}$ and $\\hat{I}_{t_1}$ along with $I_1$ , we form a new triplet, $\\mathcal{V}$ , check if the entire sampled generated image is reliable, and in that case, update the memory bank. 4) Patch-Mixed Sampling (PMS) regularizes the network by mixing the generated videos, $\\mathcal{V}$ with videos from the memory bank, $\\tilde{\\mathcal{V}}$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 84, + 870, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, our method follows the recent trend of TTA by addressing the adaptation to unknown distribution shifts encountered during test-time in an online manner.", + "bbox": [ + 75, + 458, + 470, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 517, + 250, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 542, + 187, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given key frames $I_0, I_1$ and the inter-frame event stream $\\mathcal{E}_{0\\rightarrow 1}$ , Event-based Video Frame Interpolation (Ev-VFI) networks aim to generate the intermediate frame $I_{t}$ on an arbitrary time $t \\in [0,1]$ . Following the previous works, we convert a event stream as a voxel grid [67]. We denote $E_{t_a \\to t_b}$ as a voxel grid form between times $t_a$ and $t_b$ .", + "bbox": [ + 75, + 565, + 468, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TTA-EVF adapts a pre-trained Ev-VFI network trained on the source domain to a target domain with domain shift in the device and setting. In the pre-training phase, we train the network using source data with high-frame rate videos in a supervised manner (Sec. 3.2). We then utilize this pretrained Ev-VFI network for test-time adaptation using target data without accessing the source data (Sec. 3.3).", + "bbox": [ + 75, + 656, + 470, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Pre-training on Source Dataset", + "text_level": 1, + "bbox": [ + 76, + 771, + 352, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In general, TTA methods perform test-time training stably by updating only a subset of network parameters, such as Batch Normalization (BN) [18]. As mentioned in [53], updating the full parameters at test-time, especially in an online manner, leads to instability and makes it challenging to achieve optimal performance. However, in the case of low-level vision tasks like VFI, normalization layer (e.g. BN)", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1059df73b9d437f508be764524a1da6673eeb85b7f3135d0264724301565e7b4.jpg", + "image_caption": [ + "Figure 3. The architecture of Norm-Residual (NR) Block." + ], + "image_footnote": [], + "bbox": [ + 527, + 455, + 859, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is often avoided by most networks [20, 24, 51] since it removes range flexibility and can lead to a performance decrease [29]. Some alternatives for handling this issue include using slightly more flexible normalization techniques (e.g. instance normalization) or selectively updating specific convolutional layers. Updating only the parameters of the normalization layer in VFI can lead to sub-optimal performance when there is significant variance between image patches in the training and test environments. Conversely, updating specific convolution layers with more parameters can remove all statistical modulation, causing instability during the test-time training process. To address this issue, we propose a solution by designing a Norm-Residual (NR) Block that includes both convolution and normalization layers. This approach aims to strike a balance between updating parameters effectively and maintaining statistical modulation during the learning process.", + "bbox": [ + 496, + 597, + 890, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 3, given a feature $\\mathbf{Z}$ , NR Block first split a feature into two feature maps $\\mathbf{Z}^{(1)}$ and $\\mathbf{Z}^{(2)}$ along the channel dimension. The first map $\\mathbf{Z}^{(1)}$ is skipped to keep", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25703", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the context information, and rest feature $\\mathbf{Z}^{(2)}$ is normalized by normalization layer. Then, we fuse the two features by concatenate and convolutional layer. Finally, we apply a element-wise summation with original feature $\\mathbf{Z}$ to obtain the output $\\hat{\\mathbf{Z}}$ . We use the group normalization [57] in normalization layer, which shows the best performance in our experiments. For efficiency, we exclusively deploy the NR Block in the encoders of the networks, replacing the conventional residual blocks of existing networks. Therefore, TTA-EVF is a general and model-agnostic method specifically crafted to be adaptable to any network architecture used in Ev-VFI. Further details can be found in the supple.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Network Optimization. The modified Ev-VFI network with the NR Block is trained on the source dataset using ground truth with high-frame-rate videos. Specifically, we optimize the Ev-VFI network, $\\mathcal{F}$ , by the supervision as:", + "bbox": [ + 75, + 272, + 470, + 334 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\theta (\\mathcal {F})} {\\arg \\min } \\left(\\mathcal {L} _ {\\mathcal {F}} \\left(\\hat {I} _ {t}, I _ {t}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 345, + 468, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathcal{F}}$ is the loss function proposed by each Ev-VFI network and the generated intermediate frame $\\hat{I}_t$ is obtained by $\\hat{I}_t = \\mathcal{F}(I_0, I_1, E_{0 \\to t}, E_{1 \\to t})$ .", + "bbox": [ + 76, + 383, + 470, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Test-time Adaptation on Target Dataset", + "text_level": 1, + "bbox": [ + 76, + 441, + 416, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our TTA setup, the model generates intermediate frames at each iteration and simultaneously updates the parameters of the NR Block. Since the target domain only contains LFR frames, not HFR frames, self-supervised optimization using only LFR is required.", + "bbox": [ + 75, + 464, + 468, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Reliable Pixel Sampling (RPS)", + "text_level": 1, + "bbox": [ + 76, + 561, + 344, + 577 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the target domain, even without High-Frame-Rate (HFR) ground truth, when given Low-Frame-Rate (LFR) videos $I_0$ , $I_1$ , $I_2$ as triplets, and the events between them $E_{0\\to 1}$ and $E_{2\\to 1}$ , the network can be trained using the following self-supervised loss:", + "bbox": [ + 75, + 585, + 468, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S} = \\mathcal {L} _ {\\mathcal {F}} (\\hat {I} _ {1}, I _ {1}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 672, + 468, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{I}_1 = \\mathcal{F}(I_0, I_2, E_{0\\rightarrow 1}, E_{2\\rightarrow 1})$ . However, this is problematic as the network cannot learn the knowledge of generating HFR because it is generating LFR again from temporally subsampled LFR images, leading to the loss of the ability to generate HFR trained from the source domain. As a way to address this, applying the cyclic consistency loss [16, 39] from the existing unsupervised VFI is one option. However, directly applying these methods to TTA of Ev-VFI is intractable, since there is no guarantee that the generated intermediate frames will be clean and artifact-free on target domains. Therefore, we need a new paradigm.", + "bbox": [ + 75, + 702, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, given the $I_0$ , $I_1$ , $I_2$ triplet, we design Reliable Pixel Sampling (RPS) for the left intermediate time", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$t_0$ and the right intermediate time $t_1$ . Here, $t_0$ and $t_1$ are selected at random from the time intervals 0 to 1 and 1 to 2, respectively. Since the RPS process is identical for both $t_0$ and $t_1$ , we introduce it here for $t_0$ only. Reliable data sampling without the intermediate ground truth frames, $I_{t_0}$ , is highly challenging and can lead to unstable learning if estimated incorrectly. To address this, we adopt the teacher-student framework [49] commonly used for pseudo labeling [10, 58] in other tasks. In this framework, the student learns from pseudo labels through self-training, while the teacher is updated solely from the momentum using the parameters of the student. The motivation behind proposed RPS stems from the observation that pixels estimated similarly by two networks with slightly different parameters, yet achieving similar performance in VFI, are likely to be reliable, whereas parts that are difficult to estimate and prone to errors would result in differing predictions between the two networks. Hence, we classify pixels consistently estimated by both teacher-student networks as reliable pixels, while considering pixels with inconsistent predictions as unreliable. Given intermediate frames generated from the student network $\\mathcal{F}_{\\mathcal{S}}$ and the teacher network $\\mathcal{F}_{\\mathcal{T}}$ denoted as $\\hat{I}_{t_0}^{\\mathcal{S}}$ and $\\hat{I}_{t_0}^{\\mathcal{T}}$ respectively, the reliable map, $R_{t_0}$ , with spatial dimensions $H\\times W\\times 1$ (where $H$ and $W$ represent the height and width of the image $I_0$ ) is defined as follows:", + "bbox": [ + 496, + 90, + 893, + 470 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {t _ {0}} (i, j) = \\left\\{ \\begin{array}{l} 1, \\text {i f} \\| \\hat {I} _ {t _ {0}} ^ {\\mathcal {S}} (i, j) - \\hat {I} _ {t _ {0}} ^ {\\mathcal {T}} (i, j) \\| < \\rho , \\\\ 0, \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 491, + 890, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i = 1,\\dots ,H$ $j = 1,\\ldots ,W$ and $\\rho$ denote the threshold. When estimating such reliability, it is crucial to determine a specific threshold. However, assigning the same threshold to all scenes is inappropriate since the performance of VFI significantly varies across scenes. To achieve adaptive thresholding for each scene, we utilize the LFR frames $I_{0}$ and $I_{2}$ , along with the interceding events $E_{0\\rightarrow 1}$ and $E_{2\\rightarrow 1}$ , to estimate $\\hat{I}_1$ . Subsequently, we compute a pixel-wise error map, $e = \\| \\hat{I}_1 - I_1\\|$ , between $\\hat{I}_1$ and $I_{1}$ , allowing us to calculate $\\rho$ , using the average value of $e$ .", + "bbox": [ + 496, + 551, + 893, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, this approach encounters issues when dealing with scenes that contain motionless, static regions (e.g., backgrounds). In such cases, these areas exhibit minimal pixel-wise errors, which leads to an inadequate adaptive threshold calculation for the scene if these pixels are included in the average calculation. To address this, we propose an approach that excludes motionless regions based on optical flow to refine the threshold calculation. Since most Ev-VFI methods compute optical flow towards the intermediate frame from key frames, we can directly calculate a motion magnitude map [54], $M$ , from the intermediate output of optical flow, denoted as $F_{10}$ and $F_{12}$ , without requiring", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25704", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 PMS Pseudo Code", + "text_level": 1, + "bbox": [ + 78, + 90, + 292, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Generated HFR set $\\mathcal{V} = \\{\\hat{I}_{t_0}, I_1, \\hat{I}_{t_1}, E_{t_0 \\to 1}, E_{t_1 \\to 1}\\}$", + "bbox": [ + 80, + 112, + 467, + 128 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initialize $\\mathcal{L}_M\\gets 0$", + "2: Get reliability of $\\nu$ based on Eq.(9)", + "3: Sample $\\mathcal{V} = \\{\\bar{I}_{\\tilde{t}_0},\\bar{I}_1,\\bar{I}_{\\tilde{t}_1},\\bar{E}_{\\tilde{t}_0\\to 1},E_{\\tilde{t}_1\\to 1}\\}$ from $\\mathcal{Q}$", + "4: Sample mixing ratio $\\lambda \\in (0.7, 1.0)$", + "5: if $\\nu$ is reliable then", + "6: $\\mathcal{M}_{t_0} = \\lambda \\hat{I}_{t_0} + (1 - \\lambda)\\tilde{I}_1, \\mathcal{M}_{t_1} = \\lambda \\hat{I}_{t_1} + (1 - \\lambda)\\tilde{I}_1$", + "7: $\\mathcal{M}_1 = \\lambda I_1 + (1 - \\lambda)\\tilde{I}_1$", + "8: $\\mathcal{D}_{t_0\\to 1} = E_{t_0\\to 1},\\mathcal{D}_{t_1\\to 1} = E_{t_1\\to 1}$", + "9: Push $\\mathcal{V}$ into memory bank $\\mathcal{Q}$" + ], + "bbox": [ + 86, + 128, + 467, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "10: else", + "bbox": [ + 81, + 265, + 135, + 276 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "11: $\\mathcal{M}_{t_0} = (1 - \\lambda)I_1 + \\lambda \\tilde{I}_{\\tilde{t}_0}$ , $\\mathcal{M}_{t_1} = (1 - \\lambda)I_1 + \\lambda \\tilde{I}_{\\tilde{t}_1}$ .", + "12: $\\mathcal{M}_1 = (1 - \\lambda)I_1 + \\lambda \\tilde{I}_1$", + "13: $\\mathcal{D}_{t_0\\to 1} = \\bar{E}_{\\tilde{t}_0\\to 1},\\mathcal{D}_{t_1\\to 1} = \\bar{E}_{\\tilde{t}_{1}\\to 1}$", + "14: if $\\mathrm{len}(\\mathcal{Q}) > \\mathrm{max}$ queue size then", + "15: Pop oldest ones out of $\\mathcal{Q}$", + "16: $\\mathcal{L}_M\\gets \\mathcal{L}_{\\mathcal{F}}(\\mathcal{M}_1,\\mathcal{F}_S(\\mathcal{M}_{t_0},\\mathcal{M}_{t_1},\\mathcal{D}_{t_0\\to 1},\\mathcal{D}_{t_1\\to 1}))$" + ], + "bbox": [ + 81, + 277, + 465, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: $\\mathcal{L}_M$", + "bbox": [ + 78, + 381, + 171, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing additional computations as:", + "bbox": [ + 76, + 402, + 285, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM (i, j) = \\left(\\| F _ {1 0} (i, j) \\| + \\| F _ {1 2} (i, j) \\|\\right) / 2, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 426, + 468, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\|\\cdot\\|$ denote the magnitude of the motion vector. The motion mask $\\widehat{M}$ , which excludes motionless areas based on the motion magnitude map for each pixel, can be created as follows:", + "bbox": [ + 76, + 450, + 468, + 511 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {M} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} M (i, j) \\geq m, \\forall i, \\forall j \\\\ 0 & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 518, + 468, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, $m$ represents a predefined minimal amount of motion, and in our experiment, we use a value of 1. Then, we calculate the scene adaptive threshold as:", + "bbox": [ + 76, + 565, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\rho = \\frac {1}{| \\widehat {M} |} \\sum \\widehat {M} \\odot e. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 617, + 468, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define the self-training loss between the teacher and student networks using a pixel-wise reliable map in Eq.(3) and Eq.(6):", + "bbox": [ + 76, + 657, + 468, + 703 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R} ^ {t _ {0}} = \\left\\| \\hat {I} _ {t _ {0}} ^ {\\mathcal {S}} (i, j) - \\hat {I} _ {t _ {0}} ^ {\\mathcal {T}} (i, j) \\right\\| \\odot R _ {t _ {0}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 709, + 468, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the reliable pixel-wise loss, taking into account both intermediate times $t_0$ and $t_1$ , is defined as follows:", + "bbox": [ + 76, + 734, + 468, + 765 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R} = \\mathcal {L} _ {R} ^ {t _ {0}} + \\mathcal {L} _ {R} ^ {t _ {1}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 773, + 468, + 791 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Patch-Mixed Sampling (PMS)", + "text_level": 1, + "bbox": [ + 76, + 816, + 339, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The issues that can arise in TTA include overfitting to specific scenes due to the sequential data for the same scene being continuously fed, as well as the problem of catastrophic forgetting [4, 26] where existing knowledge is lost.", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/967d2dd49df540049f3700bd81fe0dcdef83d7bfb79f55ddc063184ae4673f38.jpg", + "image_caption": [ + "Figure 4. The examples of patch-mixed sampling (PMS) process." + ], + "image_footnote": [], + "bbox": [ + 514, + 90, + 696, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b7df87b8dadc8a0ab456d0427afcce87026021b561a2b727ddaf4b089776bd16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 90, + 874, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this, we propose the Patch-Mixed Sampling (PMS) method. In order to avoid overfitting to specific scenes, PMS involves sampling from the memory bank, which accumulates reliable samples from the generated videos of each iteration. The detailed process of PMS is provided in Algorithm 1.", + "bbox": [ + 498, + 276, + 890, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reliability Check. Denoting the HFR video and event set, including $\\hat{I}_{t_0}$ and $\\hat{I}_{t_1}$ , from the RPS process as $\\mathcal{V}$ , we assess the reliability for the generated frames. If the pixel-wise reliability map $R$ obtained from the RPS module exceeds $\\eta$ fraction of the resolution of the entire image, we deem the generated images as reliable. This is outlined as follows:", + "bbox": [ + 498, + 366, + 890, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {R e l i a b l i t y} = \\left\\{ \\begin{array}{l} \\text {T r u e , i f} \\frac {\\| R _ {t _ {0}} \\|}{H \\times W} > \\eta \\text {a n d} \\frac {\\| R _ {t _ {1}} \\|}{H \\times W} > \\eta \\\\ \\text {F a l s e , o t h e r w i s e ,} \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 464, + 890, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Patch Mixing. Figure 4 shows an example of a patch mixing process. If the currently generated HFR video set $\\mathcal{V}$ is reliable (left of Fig. 4), we fetch a single intermediate frame $\\tilde{I}_1$ from the memory bank and blend it. Since the mixed frames, $M_{t_0}, M_1$ , and $M_{t_1}$ , all incorporate the same $\\tilde{I}_1$ , any events that require pixel changes can't arise from $\\tilde{I}_1$ , thus we directly retain the events $E_{t_0 \\to 1}$ and $E_{t_1 \\to 1}$ from $\\mathcal{V}$ . On the other hand, if $\\mathcal{V}$ is unreliable (right of Fig. 4), we retrieve both images and event sets from the memory bank. From the currently generated $\\mathcal{V}$ , we only fetch $I_1$ and blend it. Similar to before, since the same $I_1$ is blended into all three images, events originating from $I_1$ do not occur. Therefore, we directly retain the memory bank events $\\tilde{E}_{\\tilde{t}_0 \\to 1}$ and $\\tilde{E}_{\\tilde{t}_1 \\to 1}$ . For the mixing ratio of blending images, we adopt a beta distribution [62] that generates values between 0.7 and 1.0. We assign a larger ratio to the reliable set among the two sets being blended. Finally, using the patch-mixed sets, we train the student network of the RPS using the loss of Ev-VFI as follows:", + "bbox": [ + 498, + 512, + 892, + 800 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M} = \\mathcal {L} _ {\\mathcal {F}} \\left(\\mathcal {M} _ {1}, \\mathcal {F} _ {\\mathcal {S}} \\left(\\mathcal {M} _ {t _ {0}}, \\mathcal {M} _ {t _ {1}}, \\mathcal {D} _ {t _ {0} \\rightarrow 1}, \\mathcal {D} _ {t _ {1} \\rightarrow 1}\\right)\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 809, + 890, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, the total objective $\\mathcal{L}$ of TTA-EVF is as follows:", + "bbox": [ + 500, + 834, + 888, + 849 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {S} + \\lambda_ {1} \\mathcal {L} _ {R} + \\lambda_ {2} \\mathcal {L} _ {M} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 864, + 890, + 880 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_1, \\lambda_2$ are the hyper-parameters.", + "bbox": [ + 500, + 886, + 758, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25705", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a1de13773749c1f8bae56c306ed1257e8241fd74af58b5e492a41d076cb5a74e.jpg", + "image_caption": [ + "Figure 5. Visual comparisons of interpolated frames on multiple cross-domain settings using various models." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 872, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 351, + 209, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 376, + 294, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All networks are trained using an Adam optimizer [25] with a batch size of 2 and a learning rate of $1 \\times 10^{-4}$ , for both the source and target domains. For efficiency, during online training in the target domain, we crop images to a resolution of $320 \\times 320$ . In our approach, we perform network updates only once for each triplet provided. For instance, when interpolating 7 skipped frames, we generate a total of 14 frames from $I_0$ , $I_1$ , and $I_2$ , and update the network parameters only once using the loss from Eq.(11), making it highly efficient. All experiments are conducted on a single NVIDIA Titan RTX. We set the $\\eta$ as 0.9 in Eq.(9), bin size of voxel grid as 5, and $\\lambda_1, \\lambda_2$ as 1, 0.5 in Eq.(11).", + "bbox": [ + 75, + 398, + 470, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Test-Time Adaptation Results", + "text_level": 1, + "bbox": [ + 76, + 590, + 341, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the pre-training phase, we utilize the train sequences from the source domain for learning, and in the target domain, we follow the same sequential online approach as in the traditional TTA setup [53], using only the test sequences without utilizing the train set.", + "bbox": [ + 75, + 613, + 468, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our framework can be applied to existing Ev-VFI methods regardless of the model. We conduct experiments by integrating it with TimeLens [51] and CBMNet [24], using publicly available code provided by the authors. For TimeReplayer [16], as public code is unavailable, we re-implement it based on the foundational code [20] as described in the paper. To the best of our knowledge, we are the first to propose test-time adaptation for event-based video frame interpolation. Therefore, for comparison, we adopt the previously proposed meta-learning approach [11] for frame-based VFI and the self-supervised optimization [16] for Ev-VFI. When training at test time, the performance tends to be low if we simply apply the existing methods [16] as is. Therefore, we keep our NR Block", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a2203e4705380265836f9d2ddd8fac22d0b11a930ec6ede90d07d9715dfb9837.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons with existing method in cross-domain datasets. S and T denote the source and target datasets." + ], + "table_footnote": [], + "table_body": "
S TMethodsTimeReplayer [16]TimeLens [51]CBMNet [24]
PSNRSSIMPSNRSSIMPSNRSSIM
B RNo Adapt32.240.915432.040.907331.980.9081
Meta [11]32.750.918032.430.911232.210.9109
Cycle [16]32.870.920833.010.916533.090.9214
TTA-EVF34.070.928634.240.929934.420.9312
B HNo Adapt29.780.837829.960.841329.870.8395
Meta [11]30.130.839330.140.834130.220.8407
Cycle [16]30.320.810430.660.844130.750.8466
TTA-EVF31.450.855132.130.859932.070.8584
R BNo Adapt21.940.643422.320.674521.890.6427
Meta [11]22.830.681823.140.702022.750.6822
Cycle [16]23.980.715424.110.722324.430.7305
TTA-EVF25.850.763826.910.775227.160.7768
R HNo Adapt27.830.815827.610.812127.920.8222
Meta [11]28.110.820628.320.823428.540.8299
Cycle [16]28.430.830629.120.838929.430.8410
TTA-EVF30.330.846431.830.858731.860.8590
", + "bbox": [ + 516, + 378, + 877, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "update and TTA setup intact while only incorporating the cyclic self-supervised mechanism [16] into Eq.(11).", + "bbox": [ + 496, + 604, + 890, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TTA in Cross-domain Datasets. We conduct cross-domain experiments using three existing high-quality real event datasets: BS-ERGB [52], HS-ERGB [51], and HighREV [46]. We refer to each datasets as B, H, and R, respectively. In each experiment, we alternate between the source and target datasets, but HS-ERGB exclusively provides with test sequences and is used solely as the target dataset.", + "bbox": [ + 496, + 638, + 890, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 presents the quantitative results in cross-domain datasets. We can confirm that TTA-EVF has significant advantages over other methods. First, Meta [11] using the meta-learning shows an overall but subtle improvement in performance. It indicates that it is not particularly effective in adapting to entirely new domains, especially when the distribution of event data changes, operating independently for each scene. In contrast, we continuously adapt online to scenes, learning in an optimally directional manner for the entire domain rather than for each individual scene. Further-", + "bbox": [ + 496, + 750, + 892, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25706", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0a5b743447736f793be1034f825dd10ed7311d0027359a249e9edb11106b710e.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons with existing method in continuous domain shifting datasets." + ], + "table_footnote": [], + "table_body": "
NetworksMethodsTarget Sequences
123456Total
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
TimeReplayer [16]No Adapt34.670.909028.520.670733.200.748030.940.702434.870.906936.500.905432.810.7960
Meta [11]33.930.906328.650.673232.340.730331.140.704534.910.908336.950.910632.910.7996
Cycle [16]34.740.909827.430.667432.540.734531.220.705135.100.909137.050.909732.990.7968
TTA-EVF35.050.910129.440.676533.470.752931.900.717536.790.921337.270.914433.710.8040
TimeLens [51]No Adapt33.910.895528.170.666432.710.742830.580.702433.990.900234.290.892432.070.7896
Meta [11]34.050.900127.840.664232.800.741230.150.702233.690.890734.490.900432.090.7898
Cycle [16]34.010.900827.960.668032.410.743530.220.704134.070.900834.980.901332.140.7906
TTA-EVF34.410.905528.630.673733.090.749331.290.710034.940.905936.050.904032.800.7974
CBMNet [24]No Adapt32.700.901427.190.659732.050.732630.140.701532.240.896236.820.911731.420.7879
Meta [11]32.840.902627.320.660832.140.733430.160.701232.310.897436.650.910131.480.7885
Cycle [16]33.620.909727.820.669532.610.737230.340.705832.850.901337.190.914532.410.7988
TTA-EVF35.820.913729.250.680933.260.754031.650.720335.120.910437.540.928833.530.8023
", + "bbox": [ + 122, + 103, + 848, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/adda0ba7ec56e92ef9049e96fef547b5c23b936b0bbf9529055b774277fa1125.jpg", + "image_caption": [ + "Figure 6. Visual comparisons of interpolated frames on continuous domain shifting settings. Please zoom for better view." + ], + "image_footnote": [], + "bbox": [ + 124, + 287, + 841, + 508 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6ec6464f54577aa25dc87f644520b2a0b5699021133bfa17636c69b6df24d619.jpg", + "table_caption": [ + "Table 3. Comparison results between our TTA method and existing supervised/unsupervised methods in the target domain. * denotes that the values are taken from the original paper." + ], + "table_footnote": [], + "table_body": "
STMethodsOnlineTimeReplayer [16]TimeLens [51]
PSNRSSIMPSNRSSIM
RRSupervised Unsupervised--32.81*0.901*
32.420.898--
BRTTA-EVF✓34.070.92934.240.930
HH-farSupervised Unsupervised--32.31*0.869*
30.07*0.834*--
BH-farTTA-EVF✓31.720.85932.420.870
HH-closeSupervised Unsupervised--31.68*0.835*
29.83*0.816*--
BH-closeTTA-EVF✓30.950.82932.110.860
", + "bbox": [ + 86, + 571, + 459, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "more, when we examine the results of Cycle [16], which incorporate a self-supervised loss into our TTA approach, it becomes evident that simply maintaining continuous online updates does not yield significant performance gains. In the case of Cycle [16], learning cyclically with generated videos during limited online updates is unable to amplify adaptation ability in situations where unavoidable errors occur as domains change. In contrast, our method identifies significant errors in pixel and samples for each scene, producing substantial improvements even within the same setting. Figure 5 also demonstrates the effectiveness of our method across various datasets and networks.", + "bbox": [ + 75, + 717, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, we present the comparison of our TTA approaches with supervised/unsupervised methods in Table 3. Despite the fact that our method involves online updates on the target dataset after pre-trained from other source domain datasets, it outperforms methods trained on the target dataset using supervised or unsupervised approaches. For instance, on the HighREV (R) dataset using the TimeLens [51] model, our approach achieved a $1.43\\mathrm{dB}$ higher PSNR than supervised learning, even when trained in a different domain, highlighting its ability to bridge domain gaps and improve the versatility of event-based VFI.", + "bbox": [ + 496, + 531, + 890, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TTA in Continuous Domain Shift. We experiments in a highly challenging setting where the distribution of the dataset continually undergoes significant changes. Since there were no existing datasets with consistently changing domains in high-quality RGB-Event data, we acquire the novel Event-RGB Distribution Shift (ERDS) dataset. In the ERDS dataset, the training dataset includes HFR videos with event data in typical lighting conditions. The test dataset, on the other hand, includes scenarios where lighting conditions are consistently altered, and the event trigger threshold is adjusted accordingly to obtain sufficient information from the events. We present the results in Table 2. In the highly challenging setting of continuous domain shift", + "bbox": [ + 496, + 703, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25707", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d68ccf7a6d672a0cd3e87510e4de4039e26d51ac8a663abfb7abe3e686d8fd20.jpg", + "table_caption": [ + "Table 4. Ablation study of the proposed components." + ], + "table_footnote": [], + "table_body": "
LSNR BlockLRLMCrossContinuous
No Adaptation29.9631.42
✓30.4231.98
✓✓31.5332.76
✓✓✓31.8433.01
✓✓✓31.9233.15
✓✓✓✓32.1333.53
", + "bbox": [ + 155, + 103, + 393, + 189 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0ec67d5d57595024c158b916854d50d3e542c327b7d93262cb75f03aa793ffdd.jpg", + "table_caption": [ + "Table 5. Results based on various parameter update method." + ], + "table_footnote": [], + "table_body": "
UpdatePSNR ↑SSIM ↑
Full Parameters31.040.8489
Encoder Parameters30.890.8455
Batch Normalization (BN) [18]30.440.8423
Layer Normalization (LN) [1]30.920.8477
NR Block w/. BN [18]31.570.8502
NR Block w/. LN [1]31.940.8524
NR Block w/. GN [57]32.130.8599
", + "bbox": [ + 138, + 214, + 406, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing, other methods fail to deliver significant improvements. In contrast, our method can achieve substantial performance gains compared to competing methods. Figure 6 clearly demonstrates the strengths of our method.", + "bbox": [ + 75, + 318, + 467, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 392, + 228, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we conduct experiments to access the efficacy of the proposed approaches. In the cross-dataset setting, we use the TimeLens [51] model, training it with BS-ERGB and evaluating it on HS-ERGB. In the continuous setting, we utilize the CBMNet [24] model.", + "bbox": [ + 75, + 417, + 467, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of each components. We conduct experiments in Tab. 4 while adding each component of TTA-EVF one by one. When optimizing the network using only the loss $\\mathcal{L}_S$ in Eq.(2) under the TTA setup, it can be observed that there is not a significant improvement in performance. This is because directly self-training with only LFR videos prevents the network from learning the knowledge of generating HFR, and as training progresses, it loses the ability to generate HFR learned from the source domain. In contrast, it can be observed that there is an impressive improvement in performance with the addition of the proposed modules. Ablation study on NR Block. We perform ablation by keeping all our other modules intact while changing only the update methods. As shown in Tab. 5, batch normalization (BN) [18] and layer normalization (LN) [1] lead to a decrease in performance. The proposed NR Block update method significantly improves performance, with group normalization (GN) [57] being particularly effective. The benefits of the adaptive threshold in RPS. We calculate the threshold for the reliable map between the student and teacher networks adaptively for each scene in the RPS module. Table 7 compares the adaptive thresholding method for the reliable map in the RPS module with a fixed-value thresholding approach.", + "bbox": [ + 75, + 493, + 467, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ablation study of mixing ratio $\\lambda$ in PMS. Table 6 presents an analysis of the mixing ratio between the samples in the memory bank and the generated videos in PMS.", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4c28ce03cdf19a28634f762ee7dee6e021e645ca4171e462b9f91efafe1d823c.jpg", + "table_caption": [ + "Table 6. Performance based on minimum value of mixing ratio $\\lambda$ ." + ], + "table_footnote": [], + "table_body": "
min(λ)0.50.60.70.80.9
Cross31.9232.0632.1332.0431.93
Continuous33.3033.4933.5333.4233.36
", + "bbox": [ + 563, + 104, + 828, + 148 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a3cde57dfa83b200599c4eec0109754c217fa243ec865fa11ce33f394a7022a1.jpg", + "table_caption": [ + "Table 7. Analysis on the reliable threshold." + ], + "table_footnote": [], + "table_body": "
ρ∞0.20.10.050.01Adaptive (Ours)
Cross30.8730.9331.4431.8531.4232.13
Continuous32.2132.6433.0232.9732.7833.53
", + "bbox": [ + 519, + 175, + 872, + 215 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5f9549a23d13213e00b36707e3c014ef5a39a36709c2d25043b243eec84b30e0.jpg", + "table_caption": [ + "Table 8. Inference time on different updating intervals. In the case of TTA-EVF, the inference time includes the training time as well." + ], + "table_footnote": [], + "table_body": "
S TMethods (TimeLens [51])TimePSNRSSIM
B H-closeNo adapt80 min29.960.8413
TTA-EVF (Interval=5)81 min31.440.8541
TTA-EVF (Interval=3)82 min31.620.8571
TTA-EVF (Interval=1)88 min32.110.8598
", + "bbox": [ + 516, + 253, + 875, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We select the mixing ratio $\\lambda$ uniformly, but it can be set with a minimum value. The best performance is observed when the minimum value, $\\min (\\lambda)$ , is set to 0.7, demonstrating robustness to parameter variations.", + "bbox": [ + 498, + 320, + 890, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Analysis of Updating Interval", + "text_level": 1, + "bbox": [ + 500, + 396, + 769, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present the performance based on the updating interval in Table 8. Our TTA method is highly cost-effective despite performing training during the inference process. TTA-EVF, which updates only once during the multiple inferences when an video triplet is input, exhibits minimal time difference compared to simple inference. When the interval is increased, the time required is almost the same as simple inference, yet the quality of the generated videos is significantly higher. For example, with an interval of 5, there is only a 1 minute difference in total inference time, but the PSNR is improved by $1.48\\mathrm{dB}$ .", + "bbox": [ + 496, + 421, + 890, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 602, + 617, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a Test-Time Adaptation (TTA) for event-based video frame interpolation. We effectively perform online TTA using the newly proposed modules in both cross-domain and continuous domain-shifting settings. Additionally, we acquire the ERDS dataset, containing high-quality RGB and event datasets with continuous distribution changes. Our TTA-EVF, despite online learning, shows minimal speed differences compared to simple inference and outperforms both supervised and unsupervised methods in cross-domain experiments, showcasing its effectiveness.", + "bbox": [ + 496, + 628, + 890, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This research was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF2022R1A2B5B03002636) and the Challengeable Future Defense Technology Research and Development Program through the Agency For Defense Development (ADD) funded by the Defense Acquisition Program Administration (DAPA) in 2024 (No.912768601).", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25708", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.8", + "[2] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3703-3712, 2019. 1, 2", + "[3] Wenbo Bao, Wei-Sheng Lai, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Memc-net: Motion estimation and motion compensation driven neural network for video interpolation and enhancement. IEEE transactions on pattern analysis and machine intelligence, 43(3):933-948, 2019. 1, 2", + "[4] Magdalena Biesialska, Katarzyna Biesialska, and Marta R Costa-Jussa. Continual lifelong learning in natural language processing: A survey. arXiv preprint arXiv:2012.09823, 2020.5", + "[5] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 2", + "[6] Hoonhee Cho and Kuk-Jin Yoon. Event-image fusion stereo using cross-modality feature propagation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 454-462, 2022. 2", + "[7] Hoonhee Cho and Kuk-Jin Yoon. Selection and cross similarity for event-image deep stereo. In European Conference on Computer Vision, pages 470-486. Springer, 2022. 2", + "[8] Hoonhee Cho, Jegyeong Cho, and Kuk-Jin Yoon. Learning adaptive dense event stereo from the image domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17797-17807, 2023. 2", + "[9] Hoonhee Cho, Yuhwan Jeong, Taewoo Kim, and Kuk-Jin Yoon. Non-coaxial event-guided motion deblurring with spatial alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12492-12503, 2023. 2", + "[10] Hoonhee Cho, Hyeonseong Kim, Yujeong Chae, and Kuk-Jin Yoon. Label-free event-based object recognition via joint learning with image reconstruction from events. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19866-19877, 2023. 4", + "[11] Myungsub Choi, Janghoon Choi, Sungyong Baik, Tae Hyun Kim, and Kyoung Mu Lee. Scene-adaptive video frame interpolation via meta-learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9444-9453, 2020. 2, 6, 7", + "[12] Myungsub Choi, Heewon Kim, Bohyung Han, Ning Xu, and Young Mu Lee. Channel attention is all you need for video frame interpolation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10663-10671, 2020. 1", + "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. Advances in Neural Information Processing Systems, 35:29374-29385, 2022. 2", + "[15] Shurui Gui, Chaoyue Wang, Qihua Chen, and Dacheng Tao. Featureflow: Robust video interpolation via structure-to-texture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14004-14013, 2020. 1, 2", + "[16] Weihua He, Kaichao You, Zhendong Qiao, Xu Jia, Ziyang Zhang, Wenhui Wang, Huchuan Lu, Yaoyuan Wang, and Jianxing Liao. Timereplayer: Unlocking the potential of event cameras for video interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17804-17813, 2022. 1, 2, 4, 6, 7", + "[17] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 1", + "[18] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 3, 8", + "[19] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 2", + "[20] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9000-9008, 2018. 1, 2, 3, 6", + "[21] Shihao Jiang, Yao Lu, Hongdong Li, and Richard Hartley. Learning optical flow from a few matches. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16592-16600, 2021.", + "[22] Rico Jonschkowski, Austin Stone, Jonathan T Barron, Ariel Gordon, Kurt Konolige, and Anelia Angelova. What matters in unsupervised optical flow. In European Conference on Computer Vision, pages 557-572. Springer, 2020. 1, 2", + "[23] Junho Kim, Inwoo Hwang, and Young Min Kim. Ev-tta: Test-time adaptation for event-based object recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17745-17754, 2022. 2", + "[24] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 1, 3, 6, 7, 8", + "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neu" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25709", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ral networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 5", + "[27] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 2", + "[28] Jinxiu Liang, Yixin Yang, Boyu Li, Peiqi Duan, Yong Xu, and Boxin Shi. Coherent event guided low-light video enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10615-10625, 2023. 2", + "[29] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 3", + "[30] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. Learning event-driven video deblurring and interpolation. ECCV, 2020. 2", + "[31] Wei Lin, Muhammad Jehanzeb Mirza, Mateusz Kozinski, Horst Possegger, Hilde Kuehne, and Horst Bischof. Video test-time adaptation for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22952-22961, 2023. 2", + "[32] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808-21820, 2021. 2", + "[33] Huaijia Lin Jiangbo Lu Liying Lu, Ruizheng Wu and Jiaya Jia. Video frame interpolation with transformer. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2", + "[34] Weng Fei Low and Gim Hee Lee. Robust e-nerf: Nerf from sparse & noisy events under non-uniform motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 18335–18346, 2023. 2", + "[35] Kunming Luo, Chuan Wang, Shuaicheng Liu, Haoqiang Fan, Jue Wang, and Jian Sun. Upflow: Upsampling pyramid for unsupervised optical flow learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1045-1054, 2021. 1, 2", + "[36] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 2", + "[37] Junheum Park, Keunsoo Ko, Chul Lee, and Chang-Su Kim. Bmbc: Bilateral motion estimation with bilateral cost volume for video interpolation. In European Conference on Computer Vision, 2020. 1, 2", + "[38] Junheum Park, Chul Lee, and Chang-Su Kim. Asymmetric bilateral motion estimation for video frame interpolation. In International Conference on Computer Vision, 2021. 1, 2", + "[39] Fitssum A Reda, Deqing Sun, Aysegul Dundar, Mohammad Shoeybi, Guilin Liu, Kevin J Shih, Andrew Tao, Jan Kautz," + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Bryan Catanzaro. Unsupervised video interpolation using cycle consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 892-900, 2019. 2, 4", + "[40] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 2", + "[41] Wei Shang, Dongwei Ren, Dongqing Zou, Jimmy S. Ren, Ping Luo, and Wangmeng Zuo. Bringing events into video deblurring with non-consecutively blurry frames. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4531-4540, 2021. 2", + "[42] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. Advances in Neural Information Processing Systems, 35:14274-14289, 2022. 2", + "[43] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. Xvfi: extreme video frame interpolation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1", + "[44] Chen Song, Qixing Huang, and Chandrajit Bajaj. E-cir: Event-enhanced continuous intensity recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7803-7812, 2022. 2", + "[45] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European Conference on Computer Vision (ECCV), 2022. 2", + "[46] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18043-18052, 2023. 6", + "[47] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 2", + "[48] Zhaoning Sun, Nico Messikommer, Daniel Gehrig, and Davide Scaramuzza. Ess: Learning event-based semantic segmentation from still images. ArXiv, abs/2203.10016, 2022. 2", + "[49] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2, 4", + "[50] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 1, 2", + "[51] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25710", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16155-16164, 2021. 1, 3, 6, 7, 8", + "[52] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. arXiv preprint arXiv:2203.17191, 2022. 1, 6", + "[53] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 2, 3, 6", + "[54] Yusheng Wang, Yunfan Lu, Ye Gao, Lin Wang, Zhihang Zhong, Yinqiang Zheng, and Atsushi Yamashita. Efficient video deblurring guided by motion magnitude. In European Conference on Computer Vision, pages 413-429. Springer, 2022. 4", + "[55] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based blurry frame interpolation under blind exposure. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1588-1598, 2023. 2", + "[56] Song Wu, Kaichao You, Weihua He, Chen Yang, Yang Tian, Yaoyuan Wang, Ziyang Zhang, and Jianxing Liao. Video interpolation by event-driven anisotropic adjustment of optical flow. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1", + "[57] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 4, 8", + "[58] Ruihao Xia, Chaoqiang Zhao, Meng Zheng, Ziyan Wu, Qiyu Sun, and Yang Tang. Cmda: Cross-modality domain adaptation for nighttime semantic segmentation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 21515-21524, 2023. 2, 4", + "[59] Fang Xu, Lei Yu, Bishan Wang, Wen Yang, Gui-Song Xia, Xu Jia, Zhendong Qiao, and Jianzhuang Liu. Motion deblurring with real events. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2583-2592, 2021. 2", + "[60] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. In NeurIPS, 2019. 1, 2", + "[61] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127(8): 1106-1125, 2019. 1", + "[62] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 5", + "[63] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems, 35: 38629-38642, 2022. 2", + "[64] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proceedings of" + ], + "bbox": [ + 78, + 90, + 468, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17765-17774, 2022. 2", + "[65] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17765-17774, 2022. 2", + "[66] Yizhe Zhang, Shubhankar Borse, Hong Cai, and Fatih Porikli. Auxadapt: Stable and efficient test-time adaptation for temporally consistent video semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2339-2348, 2022. 2", + "[67] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based optical flow using motion compensation. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 3" + ], + "bbox": [ + 501, + 92, + 893, + 316 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25711", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/fa811509-bdf9-4c56-97db-551fcc2b0a3d_model.json b/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/fa811509-bdf9-4c56-97db-551fcc2b0a3d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e79b77af2943b3d420a7d88cda017d54521945c2 --- /dev/null +++ b/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/fa811509-bdf9-4c56-97db-551fcc2b0a3d_model.json @@ -0,0 +1,2510 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.002, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.131, + 0.895, + 0.177 + ], + "angle": 0, + "content": "TTA-EVF: Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.204, + 0.737, + 0.238 + ], + "angle": 0, + "content": "Hoonhee Cho, Taewoo Kim, Yuhwan Jeong, and Kuk-Jin Yoon KAIST" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.242, + 0.7, + 0.258 + ], + "angle": 0, + "content": "{gnsgnsgml, intelpro, jeongyh98, kjyoon}@kaist.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.326, + 0.474, + 0.705 + ], + "angle": 0, + "content": "Video Frame Interpolation (VFI), which aims at generating high-frame-rate videos from low-frame-rate inputs, is a highly challenging task. The emergence of bio-inspired sensors known as event cameras, which boast microsecond-level temporal resolution, has ushered in a transformative era for VFI. Nonetheless, the application of event-based VFI techniques in domains with distinct environments from the training data can be problematic. This is mainly because event camera data distribution can undergo substantial variations based on camera settings and scene conditions, presenting challenges for effective adaptation. In this paper, we propose a test-time adaptation method for event-based VFI to address the gap between the source and target domains. Our approach enables sequential learning in an online manner on the target domain, which only provides low-frame-rate videos. We present an approach that leverages confident pixels as pseudo ground-truths, enabling stable and accurate online learning from low-frame-rate videos. Furthermore, to prevent overfitting during the continuous online process where the same scene is encountered repeatedly, we propose a method of blending historical samples with current scenes. Extensive experiments validate the effectiveness of our method, both in cross-domain and continuous domain shifting setups. The code is available at https://github.com/Chohoonhee/TTA-EVF." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.738, + 0.21, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Video Frame Interpolation (VFI) is a well-established problem in the field of computer vision, aiming to enhance the temporal resolution of videos. In recent times, deep learning-based VFI approaches [2, 3, 15, 20-22, 33, 35, 37, 38, 50, 60] have achieved remarkable performances across various benchmark datasets. However, accurate motion estimation becomes challenging in scenes with complex and non-linear motions due to the lack of information, often resulting in the generation of erroneous inter frames." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.291, + 0.877, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.553, + 0.895, + 0.622 + ], + "angle": 0, + "content": "Figure 1. TTA-EVF efficiently produces high-quality results by adapting the network to the target domain in an online manner, alleviating the need for offline data supply and addressing the performance degradation observed when applying a well-trained event-based VFI network to a domain with a different distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.895, + 0.776 + ], + "angle": 0, + "content": "To address the inherent lack of intermediate information in frame cameras, recent researches [16, 24, 51, 52, 56] have explored the potential of event cameras as a promising solution for VFI, especially in scenarios involving complex motion. Event cameras sense dynamic changes in pixel intensity, triggering an event when the change surpasses a predefined threshold. The incorporation of event cameras as additional devices in VFI mitigates the challenges associated with modeling complex motion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.896, + 0.901 + ], + "angle": 0, + "content": "Typically, frame-based VFI methods [17, 33, 38] are evaluated on test datasets (e.g., SNU-FILM [12], X4K1000FPS [43]) with distributions different from their training dataset (e.g., Vimeo90K [61]). Frame-based VFI demonstrates a certain level of generalization ability even without additional modules, providing evidence that it can be applied to real scenario applications, such as dealing with changes in camera devices. These device and environment" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25701" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "variations are inherent in VFI, which must learn from High-Frame-Rate (HFR) cameras and perform inference on Low-Frame-Rate (LFR) cameras, making them unavoidable." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.14, + 0.471, + 0.456 + ], + "angle": 0, + "content": "However, when training event-based VFI using a source dataset and testing on a different domain, there is a significant performance drop compared to frame-based approaches. This is attributed not only to the inherent noise from external environment (e.g., illumination) but also to the distribution variations due to different camera settings (e.g., event trigger threshold). To address this, one can utilize self-supervised learning [16, 39] with low-frame-rate and event data from the target domain, but there is a drawback in applying them online. In practical application, event cameras exhibit continuous changes in distribution with variations in lighting conditions. It is not feasible to acquire new data for novel scenes each time they arise and wait for the network to learn offline. In this regard, an online adaptation of the network to the environment is essential from a practical perspective and is well-known as Test-Time Adaptation (TTA) [47, 53, 63]. TTA assumes a setting that does not allow access to source data and prefers online learning for adaptation. In our case, the target domain lacks ground-truth data, which means no high-frame-rate video is available in the target domain." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.46, + 0.47, + 0.808 + ], + "angle": 0, + "content": "In this paper, we propose Test-time Adaptation for Event-based Video Frame Interpolation (TTA-EVF) to alleviate performance degradation in the face of domain shifts. As shown in Fig. 1, when encountering a new domain without performing adaptation, even well-trained networks exhibit performance drops. In contrast, TTA-EVF reliably and rapidly fine-tunes in the target domain through self-training, even without high-frame-rate videos. For TTA in event-based VFI, there are two main challenges: Firstly, even without exposure to HFR videos in the target domain, the network should adapt to the scene while simultaneously learning knowledge for HFR generation from LFR data. To address this challenge, we propose Reliable Pixel Sampling (RPS), a self-training scheme for selecting pixels from generated HFR videos, utilizing a confidence estimation based on a teacher-student framework [49]. The second challenge is overfitting when continuously provided with data that has a similar distribution within a successive sequence within online learning. To mitigate this, we employ a memory bank that stores only reliable samples from the generated images. We then propose Patch-Mixed Sampling (PMS) by blending reliable samples with current images, effectively preventing overfitting through an augmentation approach." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The main contributions of our works are as follows: (I) We propose a novel Test-Time Adaptation (TTA) framework for event-based VFI. (II) We introduce a novel EventRGB Distribution Shift (ERDS) dataset for event-based video frame interpolation, where the distribution of the dataset is continually changing. (III) We compare our" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "method with existing works and demonstrate its superiority by outperforming supervised methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.136, + 0.651, + 0.151 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.161, + 0.892, + 0.478 + ], + "angle": 0, + "content": "Event-based Video Frame Interpolation. Frame-based VFI methods [2, 3, 15, 20-22, 33, 35, 37, 38, 50, 60] have continuously achieved performance improvements; however, it still has weaknesses in handling complex motion and requires motion approximation when synthesizing intermediate frames at arbitrary time. Event-based VFI methods [30, 41, 44, 45, 55, 59, 64, 65], which utilize event cameras as an additional sensor [6, 7, 9, 28, 34, 45], enable not only handling complex motion but also interpolating intermediate frames for irregular objects. However, there is still room for improvement compared to frame-based VFI methods. Typically, frame-based VFI methods perform cross-domain performance evaluations, considering real scenario applications. Well-trained frame-based VFI methods exhibit domain generalization ability, maintaining considerable performance even on different domains without the need for additional training process for adaptation. On the other hand, event-based VFI experiences a significant drop in performance when moving to different domains [16]. Therefore, resolving cross-domain challenges in event-based VFI is vital for widespread applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Test-Time Adaptation. Unlike Unsupervised Domain Adaptation [8, 48, 58], which necessitates offline learning for each domain change and relies on labeled data from the source domain, Test-Time Adaptation (TTA) is more challenging setup as it adapts to the target domain without accessing source data. In recent times, the significance of data privacy has escalated, leading to increased interest in TTA, a source-free domain adaptation approach. This is primarily due to the inefficiency of accessing source domain data during inference. TTA is actively being researched in the image domain [5, 14, 19, 27, 31, 32, 36, 40, 42, 47, 53, 63, 66], especially in areas like object recognition and segmentation. While there is an approach of its application in event-based visual recognition, Ev-TTA [23], there is still a lack of research in the domain of video frame interpolation, particularly in relation to events. The existing work, MetaVFI [11], proposed for the frame-based VFI task, learns to estimate scene-adaptive motion with minimal parameter updates through meta-learning [13]. However, this method is well-suited when there is not a significant domain gap between training and testing. We demonstrate that this approach is suboptimal when applied to event-based VFI with a substantial domain gap between the training and test sets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We propose the first-ever Test-Time Adaptation (TTA) approach for event-based Video Frame Interpolation (VFI). This is essential because event distributions can vary significantly due to external environmental factors and camera parameter settings even when using the same camera." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25702" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.085, + 0.872, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.347, + 0.896, + 0.446 + ], + "angle": 0, + "content": "Figure 2. Overview of our Test-Time Adaptation for Event-based Video Frame Interpolation (TTA-EVF) framework. Our approach involves four steps for test-time training in Ev-VFI: 1) We calculate the adaptive threshold, \\(\\rho\\), for RPS using the low frame rate image triples, \\(I_0, I_1\\), and \\(I_2\\) with interval events \\(E_{0\\rightarrow 1}\\) and \\(E_{2\\rightarrow 1}\\). 2) Reliable Pixel Sampling (RPS) is applied for arbitrary \\(t_0\\) between 0 and 1, and arbitrary \\(t_1\\) between 1 and 2, and in this process, even without intermediate images for training, reliable pixels are estimated through the student-teacher framework and self-training is performed. 3) Using the estimated intermediate frames \\(\\hat{I}_{t_0}\\) and \\(\\hat{I}_{t_1}\\) along with \\(I_1\\), we form a new triplet, \\(\\mathcal{V}\\), check if the entire sampled generated image is reliable, and in that case, update the memory bank. 4) Patch-Mixed Sampling (PMS) regularizes the network by mixing the generated videos, \\(\\mathcal{V}\\) with videos from the memory bank, \\(\\tilde{\\mathcal{V}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.471, + 0.505 + ], + "angle": 0, + "content": "Furthermore, our method follows the recent trend of TTA by addressing the adaptation to unknown distribution shifts encountered during test-time in an online manner." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.518, + 0.251, + 0.535 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.543, + 0.188, + 0.557 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.47, + 0.656 + ], + "angle": 0, + "content": "Given key frames \\( I_0, I_1 \\) and the inter-frame event stream \\( \\mathcal{E}_{0\\rightarrow 1} \\), Event-based Video Frame Interpolation (Ev-VFI) networks aim to generate the intermediate frame \\( I_{t} \\) on an arbitrary time \\( t \\in [0,1] \\). Following the previous works, we convert a event stream as a voxel grid [67]. We denote \\( E_{t_a \\to t_b} \\) as a voxel grid form between times \\( t_a \\) and \\( t_b \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.471, + 0.763 + ], + "angle": 0, + "content": "TTA-EVF adapts a pre-trained Ev-VFI network trained on the source domain to a target domain with domain shift in the device and setting. In the pre-training phase, we train the network using source data with high-frame rate videos in a supervised manner (Sec. 3.2). We then utilize this pretrained Ev-VFI network for test-time adaptation using target data without accessing the source data (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.353, + 0.789 + ], + "angle": 0, + "content": "3.2. Pre-training on Source Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In general, TTA methods perform test-time training stably by updating only a subset of network parameters, such as Batch Normalization (BN) [18]. As mentioned in [53], updating the full parameters at test-time, especially in an online manner, leads to instability and makes it challenging to achieve optimal performance. However, in the case of low-level vision tasks like VFI, normalization layer (e.g. BN)" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.457, + 0.861, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.573, + 0.868, + 0.587 + ], + "angle": 0, + "content": "Figure 3. The architecture of Norm-Residual (NR) Block." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.854 + ], + "angle": 0, + "content": "is often avoided by most networks [20, 24, 51] since it removes range flexibility and can lead to a performance decrease [29]. Some alternatives for handling this issue include using slightly more flexible normalization techniques (e.g. instance normalization) or selectively updating specific convolutional layers. Updating only the parameters of the normalization layer in VFI can lead to sub-optimal performance when there is significant variance between image patches in the training and test environments. Conversely, updating specific convolution layers with more parameters can remove all statistical modulation, causing instability during the test-time training process. To address this issue, we propose a solution by designing a Norm-Residual (NR) Block that includes both convolution and normalization layers. This approach aims to strike a balance between updating parameters effectively and maintaining statistical modulation during the learning process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig. 3, given a feature \\(\\mathbf{Z}\\), NR Block first split a feature into two feature maps \\(\\mathbf{Z}^{(1)}\\) and \\(\\mathbf{Z}^{(2)}\\) along the channel dimension. The first map \\(\\mathbf{Z}^{(1)}\\) is skipped to keep" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25703" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.273 + ], + "angle": 0, + "content": "the context information, and rest feature \\(\\mathbf{Z}^{(2)}\\) is normalized by normalization layer. Then, we fuse the two features by concatenate and convolutional layer. Finally, we apply a element-wise summation with original feature \\(\\mathbf{Z}\\) to obtain the output \\(\\hat{\\mathbf{Z}}\\). We use the group normalization [57] in normalization layer, which shows the best performance in our experiments. For efficiency, we exclusively deploy the NR Block in the encoders of the networks, replacing the conventional residual blocks of existing networks. Therefore, TTA-EVF is a general and model-agnostic method specifically crafted to be adaptable to any network architecture used in Ev-VFI. Further details can be found in the supple." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.472, + 0.335 + ], + "angle": 0, + "content": "Network Optimization. The modified Ev-VFI network with the NR Block is trained on the source dataset using ground truth with high-frame-rate videos. Specifically, we optimize the Ev-VFI network, \\(\\mathcal{F}\\), by the supervision as:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.346, + 0.47, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\underset {\\theta (\\mathcal {F})} {\\arg \\min } \\left(\\mathcal {L} _ {\\mathcal {F}} \\left(\\hat {I} _ {t}, I _ {t}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.385, + 0.471, + 0.432 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathcal{F}}\\) is the loss function proposed by each Ev-VFI network and the generated intermediate frame \\(\\hat{I}_t\\) is obtained by \\(\\hat{I}_t = \\mathcal{F}(I_0, I_1, E_{0 \\to t}, E_{1 \\to t})\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.442, + 0.418, + 0.459 + ], + "angle": 0, + "content": "3.3. Test-time Adaptation on Target Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.465, + 0.47, + 0.542 + ], + "angle": 0, + "content": "In our TTA setup, the model generates intermediate frames at each iteration and simultaneously updates the parameters of the NR Block. Since the target domain only contains LFR frames, not HFR frames, self-supervised optimization using only LFR is required." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.562, + 0.345, + 0.578 + ], + "angle": 0, + "content": "3.3.1 Reliable Pixel Sampling (RPS)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.586, + 0.47, + 0.663 + ], + "angle": 0, + "content": "In the target domain, even without High-Frame-Rate (HFR) ground truth, when given Low-Frame-Rate (LFR) videos \\( I_0 \\), \\( I_1 \\), \\( I_2 \\) as triplets, and the events between them \\( E_{0\\to 1} \\) and \\( E_{2\\to 1} \\), the network can be trained using the following self-supervised loss:" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.674, + 0.47, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S} = \\mathcal {L} _ {\\mathcal {F}} (\\hat {I} _ {1}, I _ {1}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.87 + ], + "angle": 0, + "content": "where \\(\\hat{I}_1 = \\mathcal{F}(I_0, I_2, E_{0\\rightarrow 1}, E_{2\\rightarrow 1})\\). However, this is problematic as the network cannot learn the knowledge of generating HFR because it is generating LFR again from temporally subsampled LFR images, leading to the loss of the ability to generate HFR trained from the source domain. As a way to address this, applying the cyclic consistency loss [16, 39] from the existing unsupervised VFI is one option. However, directly applying these methods to TTA of Ev-VFI is intractable, since there is no guarantee that the generated intermediate frames will be clean and artifact-free on target domains. Therefore, we need a new paradigm." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "As shown in Fig. 2, given the \\( I_0 \\), \\( I_1 \\), \\( I_2 \\) triplet, we design Reliable Pixel Sampling (RPS) for the left intermediate time" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.472 + ], + "angle": 0, + "content": "\\(t_0\\) and the right intermediate time \\(t_1\\). Here, \\(t_0\\) and \\(t_1\\) are selected at random from the time intervals 0 to 1 and 1 to 2, respectively. Since the RPS process is identical for both \\(t_0\\) and \\(t_1\\), we introduce it here for \\(t_0\\) only. Reliable data sampling without the intermediate ground truth frames, \\(I_{t_0}\\), is highly challenging and can lead to unstable learning if estimated incorrectly. To address this, we adopt the teacher-student framework [49] commonly used for pseudo labeling [10, 58] in other tasks. In this framework, the student learns from pseudo labels through self-training, while the teacher is updated solely from the momentum using the parameters of the student. The motivation behind proposed RPS stems from the observation that pixels estimated similarly by two networks with slightly different parameters, yet achieving similar performance in VFI, are likely to be reliable, whereas parts that are difficult to estimate and prone to errors would result in differing predictions between the two networks. Hence, we classify pixels consistently estimated by both teacher-student networks as reliable pixels, while considering pixels with inconsistent predictions as unreliable. Given intermediate frames generated from the student network \\(\\mathcal{F}_{\\mathcal{S}}\\) and the teacher network \\(\\mathcal{F}_{\\mathcal{T}}\\) denoted as \\(\\hat{I}_{t_0}^{\\mathcal{S}}\\) and \\(\\hat{I}_{t_0}^{\\mathcal{T}}\\) respectively, the reliable map, \\(R_{t_0}\\), with spatial dimensions \\(H\\times W\\times 1\\) (where \\(H\\) and \\(W\\) represent the height and width of the image \\(I_0\\)) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.492, + 0.892, + 0.533 + ], + "angle": 0, + "content": "\\[\nR _ {t _ {0}} (i, j) = \\left\\{ \\begin{array}{l} 1, \\text {i f} \\| \\hat {I} _ {t _ {0}} ^ {\\mathcal {S}} (i, j) - \\hat {I} _ {t _ {0}} ^ {\\mathcal {T}} (i, j) \\| < \\rho , \\\\ 0, \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.894, + 0.705 + ], + "angle": 0, + "content": "where \\(i = 1,\\dots ,H\\) \\(j = 1,\\ldots ,W\\) and \\(\\rho\\) denote the threshold. When estimating such reliability, it is crucial to determine a specific threshold. However, assigning the same threshold to all scenes is inappropriate since the performance of VFI significantly varies across scenes. To achieve adaptive thresholding for each scene, we utilize the LFR frames \\(I_{0}\\) and \\(I_{2}\\), along with the interceding events \\(E_{0\\rightarrow 1}\\) and \\(E_{2\\rightarrow 1}\\), to estimate \\(\\hat{I}_1\\). Subsequently, we compute a pixel-wise error map, \\(e = \\| \\hat{I}_1 - I_1\\|\\), between \\(\\hat{I}_1\\) and \\(I_{1}\\), allowing us to calculate \\(\\rho\\), using the average value of \\(e\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "However, this approach encounters issues when dealing with scenes that contain motionless, static regions (e.g., backgrounds). In such cases, these areas exhibit minimal pixel-wise errors, which leads to an inadequate adaptive threshold calculation for the scene if these pixels are included in the average calculation. To address this, we propose an approach that excludes motionless regions based on optical flow to refine the threshold calculation. Since most Ev-VFI methods compute optical flow towards the intermediate frame from key frames, we can directly calculate a motion magnitude map [54], \\(M\\), from the intermediate output of optical flow, denoted as \\(F_{10}\\) and \\(F_{12}\\), without requiring" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25704" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.092, + 0.294, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 PMS Pseudo Code" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.113, + 0.468, + 0.129 + ], + "angle": 0, + "content": "Input: Generated HFR set \\(\\mathcal{V} = \\{\\hat{I}_{t_0}, I_1, \\hat{I}_{t_1}, E_{t_0 \\to 1}, E_{t_1 \\to 1}\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.13, + 0.232, + 0.143 + ], + "angle": 0, + "content": "1: Initialize \\(\\mathcal{L}_M\\gets 0\\)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.145, + 0.34, + 0.159 + ], + "angle": 0, + "content": "2: Get reliability of \\(\\nu\\) based on Eq.(9)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.16, + 0.423, + 0.175 + ], + "angle": 0, + "content": "3: Sample \\(\\mathcal{V} = \\{\\bar{I}_{\\tilde{t}_0},\\bar{I}_1,\\bar{I}_{\\tilde{t}_1},\\bar{E}_{\\tilde{t}_0\\to 1},E_{\\tilde{t}_1\\to 1}\\}\\) from \\(\\mathcal{Q}\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.176, + 0.337, + 0.189 + ], + "angle": 0, + "content": "4: Sample mixing ratio \\(\\lambda \\in (0.7, 1.0)\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.19, + 0.238, + 0.202 + ], + "angle": 0, + "content": "5: if \\(\\nu\\) is reliable then" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.203, + 0.468, + 0.22 + ], + "angle": 0, + "content": "6: \\(\\mathcal{M}_{t_0} = \\lambda \\hat{I}_{t_0} + (1 - \\lambda)\\tilde{I}_1, \\mathcal{M}_{t_1} = \\lambda \\hat{I}_{t_1} + (1 - \\lambda)\\tilde{I}_1\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.22, + 0.283, + 0.235 + ], + "angle": 0, + "content": "7: \\(\\mathcal{M}_1 = \\lambda I_1 + (1 - \\lambda)\\tilde{I}_1\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.235, + 0.33, + 0.25 + ], + "angle": 0, + "content": "8: \\(\\mathcal{D}_{t_0\\to 1} = E_{t_0\\to 1},\\mathcal{D}_{t_1\\to 1} = E_{t_1\\to 1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.251, + 0.315, + 0.264 + ], + "angle": 0, + "content": "9: Push \\(\\mathcal{V}\\) into memory bank \\(\\mathcal{Q}\\)" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.13, + 0.468, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.083, + 0.266, + 0.136, + 0.277 + ], + "angle": 0, + "content": "10: else" + }, + { + "type": "text", + "bbox": [ + 0.083, + 0.278, + 0.466, + 0.297 + ], + "angle": 0, + "content": "11: \\(\\mathcal{M}_{t_0} = (1 - \\lambda)I_1 + \\lambda \\tilde{I}_{\\tilde{t}_0}\\), \\(\\mathcal{M}_{t_1} = (1 - \\lambda)I_1 + \\lambda \\tilde{I}_{\\tilde{t}_1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.083, + 0.298, + 0.282, + 0.312 + ], + "angle": 0, + "content": "12: \\(\\mathcal{M}_1 = (1 - \\lambda)I_1 + \\lambda \\tilde{I}_1\\)" + }, + { + "type": "text", + "bbox": [ + 0.083, + 0.313, + 0.33, + 0.33 + ], + "angle": 0, + "content": "13: \\(\\mathcal{D}_{t_0\\to 1} = \\bar{E}_{\\tilde{t}_0\\to 1},\\mathcal{D}_{t_1\\to 1} = \\bar{E}_{\\tilde{t}_{1}\\to 1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.083, + 0.332, + 0.325, + 0.346 + ], + "angle": 0, + "content": "14: if \\(\\mathrm{len}(\\mathcal{Q}) > \\mathrm{max}\\) queue size then" + }, + { + "type": "text", + "bbox": [ + 0.083, + 0.348, + 0.289, + 0.362 + ], + "angle": 0, + "content": "15: Pop oldest ones out of \\(\\mathcal{Q}\\)" + }, + { + "type": "text", + "bbox": [ + 0.083, + 0.365, + 0.428, + 0.381 + ], + "angle": 0, + "content": "16: \\(\\mathcal{L}_M\\gets \\mathcal{L}_{\\mathcal{F}}(\\mathcal{M}_1,\\mathcal{F}_S(\\mathcal{M}_{t_0},\\mathcal{M}_{t_1},\\mathcal{D}_{t_0\\to 1},\\mathcal{D}_{t_1\\to 1}))\\)" + }, + { + "type": "list", + "bbox": [ + 0.083, + 0.278, + 0.466, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.382, + 0.172, + 0.395 + ], + "angle": 0, + "content": "Output: \\(\\mathcal{L}_M\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.404, + 0.286, + 0.419 + ], + "angle": 0, + "content": "ing additional computations as:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.428, + 0.469, + 0.445 + ], + "angle": 0, + "content": "\\[\nM (i, j) = \\left(\\| F _ {1 0} (i, j) \\| + \\| F _ {1 2} (i, j) \\|\\right) / 2, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.452, + 0.469, + 0.512 + ], + "angle": 0, + "content": "where \\( \\|\\cdot\\| \\) denote the magnitude of the motion vector. The motion mask \\( \\widehat{M} \\), which excludes motionless areas based on the motion magnitude map for each pixel, can be created as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.519, + 0.469, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\widehat {M} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} M (i, j) \\geq m, \\forall i, \\forall j \\\\ 0 & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.566, + 0.469, + 0.611 + ], + "angle": 0, + "content": "where, \\( m \\) represents a predefined minimal amount of motion, and in our experiment, we use a value of 1. Then, we calculate the scene adaptive threshold as:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.618, + 0.469, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\rho = \\frac {1}{| \\widehat {M} |} \\sum \\widehat {M} \\odot e. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.659, + 0.469, + 0.704 + ], + "angle": 0, + "content": "We define the self-training loss between the teacher and student networks using a pixel-wise reliable map in Eq.(3) and Eq.(6):" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.71, + 0.469, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R} ^ {t _ {0}} = \\left\\| \\hat {I} _ {t _ {0}} ^ {\\mathcal {S}} (i, j) - \\hat {I} _ {t _ {0}} ^ {\\mathcal {T}} (i, j) \\right\\| \\odot R _ {t _ {0}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.736, + 0.469, + 0.766 + ], + "angle": 0, + "content": "Finally, the reliable pixel-wise loss, taking into account both intermediate times \\( t_0 \\) and \\( t_1 \\), is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.774, + 0.469, + 0.792 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R} = \\mathcal {L} _ {R} ^ {t _ {0}} + \\mathcal {L} _ {R} ^ {t _ {1}}, \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.34, + 0.832 + ], + "angle": 0, + "content": "3.3.2 Patch-Mixed Sampling (PMS)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The issues that can arise in TTA include overfitting to specific scenes due to the sequential data for the same scene being continuously fed, as well as the problem of catastrophic forgetting [4, 26] where existing knowledge is lost." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.091, + 0.697, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.092, + 0.875, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.258, + 0.889, + 0.272 + ], + "angle": 0, + "content": "Figure 4. The examples of patch-mixed sampling (PMS) process." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.277, + 0.892, + 0.367 + ], + "angle": 0, + "content": "To address this, we propose the Patch-Mixed Sampling (PMS) method. In order to avoid overfitting to specific scenes, PMS involves sampling from the memory bank, which accumulates reliable samples from the generated videos of each iteration. The detailed process of PMS is provided in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.367, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Reliability Check. Denoting the HFR video and event set, including \\(\\hat{I}_{t_0}\\) and \\(\\hat{I}_{t_1}\\), from the RPS process as \\(\\mathcal{V}\\), we assess the reliability for the generated frames. If the pixel-wise reliability map \\(R\\) obtained from the RPS module exceeds \\(\\eta\\) fraction of the resolution of the entire image, we deem the generated images as reliable. This is outlined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.465, + 0.891, + 0.506 + ], + "angle": 0, + "content": "\\[\n\\text {R e l i a b l i t y} = \\left\\{ \\begin{array}{l} \\text {T r u e , i f} \\frac {\\| R _ {t _ {0}} \\|}{H \\times W} > \\eta \\text {a n d} \\frac {\\| R _ {t _ {1}} \\|}{H \\times W} > \\eta \\\\ \\text {F a l s e , o t h e r w i s e ,} \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.513, + 0.893, + 0.801 + ], + "angle": 0, + "content": "Patch Mixing. Figure 4 shows an example of a patch mixing process. If the currently generated HFR video set \\(\\mathcal{V}\\) is reliable (left of Fig. 4), we fetch a single intermediate frame \\(\\tilde{I}_1\\) from the memory bank and blend it. Since the mixed frames, \\(M_{t_0}, M_1\\), and \\(M_{t_1}\\), all incorporate the same \\(\\tilde{I}_1\\), any events that require pixel changes can't arise from \\(\\tilde{I}_1\\), thus we directly retain the events \\(E_{t_0 \\to 1}\\) and \\(E_{t_1 \\to 1}\\) from \\(\\mathcal{V}\\). On the other hand, if \\(\\mathcal{V}\\) is unreliable (right of Fig. 4), we retrieve both images and event sets from the memory bank. From the currently generated \\(\\mathcal{V}\\), we only fetch \\(I_1\\) and blend it. Similar to before, since the same \\(I_1\\) is blended into all three images, events originating from \\(I_1\\) do not occur. Therefore, we directly retain the memory bank events \\(\\tilde{E}_{\\tilde{t}_0 \\to 1}\\) and \\(\\tilde{E}_{\\tilde{t}_1 \\to 1}\\). For the mixing ratio of blending images, we adopt a beta distribution [62] that generates values between 0.7 and 1.0. We assign a larger ratio to the reliable set among the two sets being blended. Finally, using the patch-mixed sets, we train the student network of the RPS using the loss of Ev-VFI as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.81, + 0.891, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M} = \\mathcal {L} _ {\\mathcal {F}} \\left(\\mathcal {M} _ {1}, \\mathcal {F} _ {\\mathcal {S}} \\left(\\mathcal {M} _ {t _ {0}}, \\mathcal {M} _ {t _ {1}}, \\mathcal {D} _ {t _ {0} \\rightarrow 1}, \\mathcal {D} _ {t _ {1} \\rightarrow 1}\\right)\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.835, + 0.89, + 0.85 + ], + "angle": 0, + "content": "In summary, the total objective \\(\\mathcal{L}\\) of TTA-EVF is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.865, + 0.891, + 0.881 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {S} + \\lambda_ {1} \\mathcal {L} _ {R} + \\lambda_ {2} \\mathcal {L} _ {M} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.887, + 0.759, + 0.901 + ], + "angle": 0, + "content": "where \\(\\lambda_1, \\lambda_2\\) are the hyper-parameters." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25705" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.873, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.33, + 0.807, + 0.344 + ], + "angle": 0, + "content": "Figure 5. Visual comparisons of interpolated frames on multiple cross-domain settings using various models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.352, + 0.21, + 0.37 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.377, + 0.295, + 0.394 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.471, + 0.581 + ], + "angle": 0, + "content": "All networks are trained using an Adam optimizer [25] with a batch size of 2 and a learning rate of \\(1 \\times 10^{-4}\\), for both the source and target domains. For efficiency, during online training in the target domain, we crop images to a resolution of \\(320 \\times 320\\). In our approach, we perform network updates only once for each triplet provided. For instance, when interpolating 7 skipped frames, we generate a total of 14 frames from \\(I_0\\), \\(I_1\\), and \\(I_2\\), and update the network parameters only once using the loss from Eq.(11), making it highly efficient. All experiments are conducted on a single NVIDIA Titan RTX. We set the \\(\\eta\\) as 0.9 in Eq.(9), bin size of voxel grid as 5, and \\(\\lambda_1, \\lambda_2\\) as 1, 0.5 in Eq.(11)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.342, + 0.607 + ], + "angle": 0, + "content": "4.2. Test-Time Adaptation Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.689 + ], + "angle": 0, + "content": "During the pre-training phase, we utilize the train sequences from the source domain for learning, and in the target domain, we follow the same sequential online approach as in the traditional TTA setup [53], using only the test sequences without utilizing the train set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Our framework can be applied to existing Ev-VFI methods regardless of the model. We conduct experiments by integrating it with TimeLens [51] and CBMNet [24], using publicly available code provided by the authors. For TimeReplayer [16], as public code is unavailable, we re-implement it based on the foundational code [20] as described in the paper. To the best of our knowledge, we are the first to propose test-time adaptation for event-based video frame interpolation. Therefore, for comparison, we adopt the previously proposed meta-learning approach [11] for frame-based VFI and the self-supervised optimization [16] for Ev-VFI. When training at test time, the performance tends to be low if we simply apply the existing methods [16] as is. Therefore, we keep our NR Block" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.352, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons with existing method in cross-domain datasets. S and T denote the source and target datasets." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.38, + 0.878, + 0.596 + ], + "angle": 0, + "content": "
S TMethodsTimeReplayer [16]TimeLens [51]CBMNet [24]
PSNRSSIMPSNRSSIMPSNRSSIM
B RNo Adapt32.240.915432.040.907331.980.9081
Meta [11]32.750.918032.430.911232.210.9109
Cycle [16]32.870.920833.010.916533.090.9214
TTA-EVF34.070.928634.240.929934.420.9312
B HNo Adapt29.780.837829.960.841329.870.8395
Meta [11]30.130.839330.140.834130.220.8407
Cycle [16]30.320.810430.660.844130.750.8466
TTA-EVF31.450.855132.130.859932.070.8584
R BNo Adapt21.940.643422.320.674521.890.6427
Meta [11]22.830.681823.140.702022.750.6822
Cycle [16]23.980.715424.110.722324.430.7305
TTA-EVF25.850.763826.910.775227.160.7768
R HNo Adapt27.830.815827.610.812127.920.8222
Meta [11]28.110.820628.320.823428.540.8299
Cycle [16]28.430.830629.120.838929.430.8410
TTA-EVF30.330.846431.830.858731.860.8590
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.891, + 0.636 + ], + "angle": 0, + "content": "update and TTA setup intact while only incorporating the cyclic self-supervised mechanism [16] into Eq.(11)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.746 + ], + "angle": 0, + "content": "TTA in Cross-domain Datasets. We conduct cross-domain experiments using three existing high-quality real event datasets: BS-ERGB [52], HS-ERGB [51], and HighREV [46]. We refer to each datasets as B, H, and R, respectively. In each experiment, we alternate between the source and target datasets, but HS-ERGB exclusively provides with test sequences and is used solely as the target dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Table 1 presents the quantitative results in cross-domain datasets. We can confirm that TTA-EVF has significant advantages over other methods. First, Meta [11] using the meta-learning shows an overall but subtle improvement in performance. It indicates that it is not particularly effective in adapting to entirely new domains, especially when the distribution of event data changes, operating independently for each scene. In contrast, we continuously adapt online to scenes, learning in an optimally directional manner for the entire domain rather than for each individual scene. Further-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25706" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.201, + 0.09, + 0.768, + 0.103 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons with existing method in continuous domain shifting datasets." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.104, + 0.849, + 0.284 + ], + "angle": 0, + "content": "
NetworksMethodsTarget Sequences
123456Total
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
TimeReplayer [16]No Adapt34.670.909028.520.670733.200.748030.940.702434.870.906936.500.905432.810.7960
Meta [11]33.930.906328.650.673232.340.730331.140.704534.910.908336.950.910632.910.7996
Cycle [16]34.740.909827.430.667432.540.734531.220.705135.100.909137.050.909732.990.7968
TTA-EVF35.050.910129.440.676533.470.752931.900.717536.790.921337.270.914433.710.8040
TimeLens [51]No Adapt33.910.895528.170.666432.710.742830.580.702433.990.900234.290.892432.070.7896
Meta [11]34.050.900127.840.664232.800.741230.150.702233.690.890734.490.900432.090.7898
Cycle [16]34.010.900827.960.668032.410.743530.220.704134.070.900834.980.901332.140.7906
TTA-EVF34.410.905528.630.673733.090.749331.290.710034.940.905936.050.904032.800.7974
CBMNet [24]No Adapt32.700.901427.190.659732.050.732630.140.701532.240.896236.820.911731.420.7879
Meta [11]32.840.902627.320.660832.140.733430.160.701232.310.897436.650.910131.480.7885
Cycle [16]33.620.909727.820.669532.610.737230.340.705832.850.901337.190.914532.410.7988
TTA-EVF35.820.913729.250.680933.260.754031.650.720335.120.910437.540.928833.530.8023
" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.288, + 0.842, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.511, + 0.842, + 0.524 + ], + "angle": 0, + "content": "Figure 6. Visual comparisons of interpolated frames on continuous domain shifting settings. Please zoom for better view." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.53, + 0.47, + 0.572 + ], + "angle": 0, + "content": "Table 3. Comparison results between our TTA method and existing supervised/unsupervised methods in the target domain. * denotes that the values are taken from the original paper." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.573, + 0.46, + 0.713 + ], + "angle": 0, + "content": "
STMethodsOnlineTimeReplayer [16]TimeLens [51]
PSNRSSIMPSNRSSIM
RRSupervised Unsupervised--32.81*0.901*
32.420.898--
BRTTA-EVF✓34.070.92934.240.930
HH-farSupervised Unsupervised--32.31*0.869*
30.07*0.834*--
BH-farTTA-EVF✓31.720.85932.420.870
HH-closeSupervised Unsupervised--31.68*0.835*
29.83*0.816*--
BH-closeTTA-EVF✓30.950.82932.110.860
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.899 + ], + "angle": 0, + "content": "more, when we examine the results of Cycle [16], which incorporate a self-supervised loss into our TTA approach, it becomes evident that simply maintaining continuous online updates does not yield significant performance gains. In the case of Cycle [16], learning cyclically with generated videos during limited online updates is unable to amplify adaptation ability in situations where unavoidable errors occur as domains change. In contrast, our method identifies significant errors in pixel and samples for each scene, producing substantial improvements even within the same setting. Figure 5 also demonstrates the effectiveness of our method across various datasets and networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.892, + 0.699 + ], + "angle": 0, + "content": "In addition, we present the comparison of our TTA approaches with supervised/unsupervised methods in Table 3. Despite the fact that our method involves online updates on the target dataset after pre-trained from other source domain datasets, it outperforms methods trained on the target dataset using supervised or unsupervised approaches. For instance, on the HighREV (R) dataset using the TimeLens [51] model, our approach achieved a \\(1.43\\mathrm{dB}\\) higher PSNR than supervised learning, even when trained in a different domain, highlighting its ability to bridge domain gaps and improve the versatility of event-based VFI." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.901 + ], + "angle": 0, + "content": "TTA in Continuous Domain Shift. We experiments in a highly challenging setting where the distribution of the dataset continually undergoes significant changes. Since there were no existing datasets with consistently changing domains in high-quality RGB-Event data, we acquire the novel Event-RGB Distribution Shift (ERDS) dataset. In the ERDS dataset, the training dataset includes HFR videos with event data in typical lighting conditions. The test dataset, on the other hand, includes scenarios where lighting conditions are consistently altered, and the event trigger threshold is adjusted accordingly to obtain sufficient information from the events. We present the results in Table 2. In the highly challenging setting of continuous domain shift" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25707" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.09, + 0.431, + 0.103 + ], + "angle": 0, + "content": "Table 4. Ablation study of the proposed components." + }, + { + "type": "table", + "bbox": [ + 0.156, + 0.104, + 0.394, + 0.19 + ], + "angle": 0, + "content": "
LSNR BlockLRLMCrossContinuous
No Adaptation29.9631.42
✓30.4231.98
✓✓31.5332.76
✓✓✓31.8433.01
✓✓✓31.9233.15
✓✓✓✓32.1333.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.2, + 0.451, + 0.213 + ], + "angle": 0, + "content": "Table 5. Results based on various parameter update method." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.215, + 0.408, + 0.315 + ], + "angle": 0, + "content": "
UpdatePSNR ↑SSIM ↑
Full Parameters31.040.8489
Encoder Parameters30.890.8455
Batch Normalization (BN) [18]30.440.8423
Layer Normalization (LN) [1]30.920.8477
NR Block w/. BN [18]31.570.8502
NR Block w/. LN [1]31.940.8524
NR Block w/. GN [57]32.130.8599
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.468, + 0.38 + ], + "angle": 0, + "content": "ing, other methods fail to deliver significant improvements. In contrast, our method can achieve substantial performance gains compared to competing methods. Figure 6 clearly demonstrates the strengths of our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.393, + 0.229, + 0.41 + ], + "angle": 0, + "content": "5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.468, + 0.493 + ], + "angle": 0, + "content": "In this section, we conduct experiments to access the efficacy of the proposed approaches. In the cross-dataset setting, we use the TimeLens [51] model, training it with BS-ERGB and evaluating it on HS-ERGB. In the continuous setting, we utilize the CBMNet [24] model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.468, + 0.856 + ], + "angle": 0, + "content": "Effectiveness of each components. We conduct experiments in Tab. 4 while adding each component of TTA-EVF one by one. When optimizing the network using only the loss \\(\\mathcal{L}_S\\) in Eq.(2) under the TTA setup, it can be observed that there is not a significant improvement in performance. This is because directly self-training with only LFR videos prevents the network from learning the knowledge of generating HFR, and as training progresses, it loses the ability to generate HFR learned from the source domain. In contrast, it can be observed that there is an impressive improvement in performance with the addition of the proposed modules. Ablation study on NR Block. We perform ablation by keeping all our other modules intact while changing only the update methods. As shown in Tab. 5, batch normalization (BN) [18] and layer normalization (LN) [1] lead to a decrease in performance. The proposed NR Block update method significantly improves performance, with group normalization (GN) [57] being particularly effective. The benefits of the adaptive threshold in RPS. We calculate the threshold for the reliable map between the student and teacher networks adaptively for each scene in the RPS module. Table 7 compares the adaptive thresholding method for the reliable map in the RPS module with a fixed-value thresholding approach." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The ablation study of mixing ratio \\(\\lambda\\) in PMS. Table 6 presents an analysis of the mixing ratio between the samples in the memory bank and the generated videos in PMS." + }, + { + "type": "table_caption", + "bbox": [ + 0.501, + 0.09, + 0.892, + 0.104 + ], + "angle": 0, + "content": "Table 6. Performance based on minimum value of mixing ratio \\( \\lambda \\) ." + }, + { + "type": "table", + "bbox": [ + 0.564, + 0.106, + 0.83, + 0.149 + ], + "angle": 0, + "content": "
min(λ)0.50.60.70.80.9
Cross31.9232.0632.1332.0431.93
Continuous33.3033.4933.5333.4233.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.568, + 0.161, + 0.824, + 0.174 + ], + "angle": 0, + "content": "Table 7. Analysis on the reliable threshold." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.176, + 0.873, + 0.217 + ], + "angle": 0, + "content": "
ρ∞0.20.10.050.01Adaptive (Ours)
Cross30.8730.9331.4431.8531.4232.13
Continuous32.2132.6433.0232.9732.7833.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.226, + 0.892, + 0.253 + ], + "angle": 0, + "content": "Table 8. Inference time on different updating intervals. In the case of TTA-EVF, the inference time includes the training time as well." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.254, + 0.876, + 0.315 + ], + "angle": 0, + "content": "
S TMethods (TimeLens [51])TimePSNRSSIM
B H-closeNo adapt80 min29.960.8413
TTA-EVF (Interval=5)81 min31.440.8541
TTA-EVF (Interval=3)82 min31.620.8571
TTA-EVF (Interval=1)88 min32.110.8598
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.321, + 0.892, + 0.381 + ], + "angle": 0, + "content": "We select the mixing ratio \\(\\lambda\\) uniformly, but it can be set with a minimum value. The best performance is observed when the minimum value, \\(\\min (\\lambda)\\), is set to 0.7, demonstrating robustness to parameter variations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.397, + 0.771, + 0.414 + ], + "angle": 0, + "content": "6. Analysis of Updating Interval" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.892, + 0.589 + ], + "angle": 0, + "content": "We present the performance based on the updating interval in Table 8. Our TTA method is highly cost-effective despite performing training during the inference process. TTA-EVF, which updates only once during the multiple inferences when an video triplet is input, exhibits minimal time difference compared to simple inference. When the interval is increased, the time required is almost the same as simple inference, yet the quality of the generated videos is significantly higher. For example, with an interval of 5, there is only a 1 minute difference in total inference time, but the PSNR is improved by \\(1.48\\mathrm{dB}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.603, + 0.619, + 0.619 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.779 + ], + "angle": 0, + "content": "In this paper, we present a Test-Time Adaptation (TTA) for event-based video frame interpolation. We effectively perform online TTA using the newly proposed modules in both cross-domain and continuous domain-shifting settings. Additionally, we acquire the ERDS dataset, containing high-quality RGB and event datasets with continuous distribution changes. Our TTA-EVF, despite online learning, shows minimal speed differences compared to simple inference and outperforms both supervised and unsupervised methods in cross-domain experiments, showcasing its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This research was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF2022R1A2B5B03002636) and the Challengeable Future Defense Technology Research and Development Program through the Agency For Defense Development (ADD) funded by the Defense Acquisition Program Administration (DAPA) in 2024 (No.912768601)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25708" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3703-3712, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.31 + ], + "angle": 0, + "content": "[3] Wenbo Bao, Wei-Sheng Lai, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Memc-net: Motion estimation and motion compensation driven neural network for video interpolation and enhancement. IEEE transactions on pattern analysis and machine intelligence, 43(3):933-948, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[4] Magdalena Biesialska, Katarzyna Biesialska, and Marta R Costa-Jussa. Continual lifelong learning in natural language processing: A survey. arXiv preprint arXiv:2012.09823, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.424 + ], + "angle": 0, + "content": "[5] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.472, + 0.479 + ], + "angle": 0, + "content": "[6] Hoonhee Cho and Kuk-Jin Yoon. Event-image fusion stereo using cross-modality feature propagation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 454-462, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.471, + 0.522 + ], + "angle": 0, + "content": "[7] Hoonhee Cho and Kuk-Jin Yoon. Selection and cross similarity for event-image deep stereo. In European Conference on Computer Vision, pages 470-486. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[8] Hoonhee Cho, Jegyeong Cho, and Kuk-Jin Yoon. Learning adaptive dense event stereo from the image domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17797-17807, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.58, + 0.471, + 0.647 + ], + "angle": 0, + "content": "[9] Hoonhee Cho, Yuhwan Jeong, Taewoo Kim, and Kuk-Jin Yoon. Non-coaxial event-guided motion deblurring with spatial alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12492-12503, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[10] Hoonhee Cho, Hyeonseong Kim, Yujeong Chae, and Kuk-Jin Yoon. Label-free event-based object recognition via joint learning with image reconstruction from events. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19866-19877, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[11] Myungsub Choi, Janghoon Choi, Sungyong Baik, Tae Hyun Kim, and Kyoung Mu Lee. Scene-adaptive video frame interpolation via meta-learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9444-9453, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[12] Myungsub Choi, Heewon Kim, Bohyung Han, Ning Xu, and Young Mu Lee. Channel attention is all you need for video frame interpolation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10663-10671, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. Advances in Neural Information Processing Systems, 35:29374-29385, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[15] Shurui Gui, Chaoyue Wang, Qihua Chen, and Dacheng Tao. Featureflow: Robust video interpolation via structure-to-texture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14004-14013, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.894, + 0.304 + ], + "angle": 0, + "content": "[16] Weihua He, Kaichao You, Zhendong Qiao, Xu Jia, Ziyang Zhang, Wenhui Wang, Huchuan Lu, Yaoyuan Wang, and Jianxing Liao. Timereplayer: Unlocking the potential of event cameras for video interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17804-17813, 2022. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[17] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[18] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[19] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[20] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9000-9008, 2018. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.56, + 0.894, + 0.616 + ], + "angle": 0, + "content": "[21] Shihao Jiang, Yao Lu, Hongdong Li, and Richard Hartley. Learning optical flow from a few matches. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16592-16600, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[22] Rico Jonschkowski, Austin Stone, Jonathan T Barron, Ariel Gordon, Kurt Konolige, and Anelia Angelova. What matters in unsupervised optical flow. In European Conference on Computer Vision, pages 557-572. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[23] Junho Kim, Inwoo Hwang, and Young Min Kim. Ev-tta: Test-time adaptation for event-based object recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17745-17754, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[24] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neu" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25709" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "ral networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[27] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[28] Jinxiu Liang, Yixin Yang, Boyu Li, Peiqi Duan, Yong Xu, and Boxin Shi. Coherent event guided low-light video enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10615-10625, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.471, + 0.319 + ], + "angle": 0, + "content": "[29] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[30] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. Learning event-driven video deblurring and interpolation. ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[31] Wei Lin, Muhammad Jehanzeb Mirza, Mateusz Kozinski, Horst Possegger, Hilde Kuehne, and Horst Bischof. Video test-time adaptation for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22952-22961, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[32] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808-21820, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[33] Huaijia Lin Jiangbo Lu Liying Lu, Ruizheng Wu and Jiaya Jia. Video frame interpolation with transformer. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[34] Weng Fei Low and Gim Hee Lee. Robust e-nerf: Nerf from sparse & noisy events under non-uniform motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 18335–18346, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.47, + 0.701 + ], + "angle": 0, + "content": "[35] Kunming Luo, Chuan Wang, Shuaicheng Liu, Haoqiang Fan, Jue Wang, and Jian Sun. Upflow: Upsampling pyramid for unsupervised optical flow learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1045-1054, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[36] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[37] Junheum Park, Keunsoo Ko, Chul Lee, and Chang-Su Kim. Bmbc: Bilateral motion estimation with bilateral cost volume for video interpolation. In European Conference on Computer Vision, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[38] Junheum Park, Chul Lee, and Chang-Su Kim. Asymmetric bilateral motion estimation for video frame interpolation. In International Conference on Computer Vision, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Fitssum A Reda, Deqing Sun, Aysegul Dundar, Mohammad Shoeybi, Guilin Liu, Kevin J Shih, Andrew Tao, Jan Kautz," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "and Bryan Catanzaro. Unsupervised video interpolation using cycle consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 892-900, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[40] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[41] Wei Shang, Dongwei Ren, Dongqing Zou, Jimmy S. Ren, Ping Luo, and Wangmeng Zuo. Bringing events into video deblurring with non-consecutively blurry frames. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4531-4540, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[42] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. Advances in Neural Information Processing Systems, 35:14274-14289, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[43] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. Xvfi: extreme video frame interpolation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[44] Chen Song, Qixing Huang, and Chandrajit Bajaj. E-cir: Event-enhanced continuous intensity recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7803-7812, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[45] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[46] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18043-18052, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[47] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[48] Zhaoning Sun, Nico Messikommer, Daniel Gehrig, and Davide Scaramuzza. Ess: Learning event-based semantic segmentation from still images. ArXiv, abs/2203.10016, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[49] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[50] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[51] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25710" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16155-16164, 2021. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[52] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. arXiv preprint arXiv:2203.17191, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.47, + 0.261 + ], + "angle": 0, + "content": "[53] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.331 + ], + "angle": 0, + "content": "[54] Yusheng Wang, Yunfan Lu, Ye Gao, Lin Wang, Zhihang Zhong, Yinqiang Zheng, and Atsushi Yamashita. Efficient video deblurring guided by motion magnitude. In European Conference on Computer Vision, pages 413-429. Springer, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.39 + ], + "angle": 0, + "content": "[55] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based blurry frame interpolation under blind exposure. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1588-1598, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.47, + 0.46 + ], + "angle": 0, + "content": "[56] Song Wu, Kaichao You, Weihua He, Chen Yang, Yang Tian, Yaoyuan Wang, Ziyang Zhang, and Jianxing Liao. Video interpolation by event-driven anisotropic adjustment of optical flow. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.47, + 0.504 + ], + "angle": 0, + "content": "[57] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.47, + 0.573 + ], + "angle": 0, + "content": "[58] Ruihao Xia, Chaoqiang Zhao, Meng Zheng, Ziyan Wu, Qiyu Sun, and Yang Tang. Cmda: Cross-modality domain adaptation for nighttime semantic segmentation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 21515-21524, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.47, + 0.644 + ], + "angle": 0, + "content": "[59] Fang Xu, Lei Yu, Bishan Wang, Wen Yang, Gui-Song Xia, Xu Jia, Zhendong Qiao, and Jianzhuang Liu. Motion deblurring with real events. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2583-2592, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[60] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. In NeurIPS, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[61] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127(8): 1106-1125, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[62] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[63] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems, 35: 38629-38642, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[64] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17765-17774, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[65] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17765-17774, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[66] Yizhe Zhang, Shubhankar Borse, Hong Cai, and Fatih Porikli. Auxadapt: Stable and efficient test-time adaptation for temporally consistent video semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2339-2348, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.894, + 0.318 + ], + "angle": 0, + "content": "[67] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based optical flow using motion compensation. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25711" + } + ] +] \ No newline at end of file diff --git a/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/fa811509-bdf9-4c56-97db-551fcc2b0a3d_origin.pdf b/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/fa811509-bdf9-4c56-97db-551fcc2b0a3d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e23de42af2f7031dac6eabb722b606d3a0e18608 --- /dev/null +++ b/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/fa811509-bdf9-4c56-97db-551fcc2b0a3d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db3e6c30bfd33d6996a523bee660bdb338f177295ef92cde4b6cc37730cfd2f0 +size 3319639 diff --git a/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/full.md b/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5f36d50e8e2740739aa3236104736316ebe94496 --- /dev/null +++ b/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/full.md @@ -0,0 +1,347 @@ +# TTA-EVF: Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation + +Hoonhee Cho, Taewoo Kim, Yuhwan Jeong, and Kuk-Jin Yoon KAIST + +{gnsgnsgml, intelpro, jeongyh98, kjyoon}@kaist.ac.kr + +# Abstract + +Video Frame Interpolation (VFI), which aims at generating high-frame-rate videos from low-frame-rate inputs, is a highly challenging task. The emergence of bio-inspired sensors known as event cameras, which boast microsecond-level temporal resolution, has ushered in a transformative era for VFI. Nonetheless, the application of event-based VFI techniques in domains with distinct environments from the training data can be problematic. This is mainly because event camera data distribution can undergo substantial variations based on camera settings and scene conditions, presenting challenges for effective adaptation. In this paper, we propose a test-time adaptation method for event-based VFI to address the gap between the source and target domains. Our approach enables sequential learning in an online manner on the target domain, which only provides low-frame-rate videos. We present an approach that leverages confident pixels as pseudo ground-truths, enabling stable and accurate online learning from low-frame-rate videos. Furthermore, to prevent overfitting during the continuous online process where the same scene is encountered repeatedly, we propose a method of blending historical samples with current scenes. Extensive experiments validate the effectiveness of our method, both in cross-domain and continuous domain shifting setups. The code is available at https://github.com/Chohoonhee/TTA-EVF. + +# 1. Introduction + +Video Frame Interpolation (VFI) is a well-established problem in the field of computer vision, aiming to enhance the temporal resolution of videos. In recent times, deep learning-based VFI approaches [2, 3, 15, 20-22, 33, 35, 37, 38, 50, 60] have achieved remarkable performances across various benchmark datasets. However, accurate motion estimation becomes challenging in scenes with complex and non-linear motions due to the lack of information, often resulting in the generation of erroneous inter frames. + +![](images/3d32e0b745fa88d50c391b586107d24461690ab16bdd21c71a45d404dc85ceb5.jpg) +Figure 1. TTA-EVF efficiently produces high-quality results by adapting the network to the target domain in an online manner, alleviating the need for offline data supply and addressing the performance degradation observed when applying a well-trained event-based VFI network to a domain with a different distribution. + +To address the inherent lack of intermediate information in frame cameras, recent researches [16, 24, 51, 52, 56] have explored the potential of event cameras as a promising solution for VFI, especially in scenarios involving complex motion. Event cameras sense dynamic changes in pixel intensity, triggering an event when the change surpasses a predefined threshold. The incorporation of event cameras as additional devices in VFI mitigates the challenges associated with modeling complex motion. + +Typically, frame-based VFI methods [17, 33, 38] are evaluated on test datasets (e.g., SNU-FILM [12], X4K1000FPS [43]) with distributions different from their training dataset (e.g., Vimeo90K [61]). Frame-based VFI demonstrates a certain level of generalization ability even without additional modules, providing evidence that it can be applied to real scenario applications, such as dealing with changes in camera devices. These device and environment + +variations are inherent in VFI, which must learn from High-Frame-Rate (HFR) cameras and perform inference on Low-Frame-Rate (LFR) cameras, making them unavoidable. + +However, when training event-based VFI using a source dataset and testing on a different domain, there is a significant performance drop compared to frame-based approaches. This is attributed not only to the inherent noise from external environment (e.g., illumination) but also to the distribution variations due to different camera settings (e.g., event trigger threshold). To address this, one can utilize self-supervised learning [16, 39] with low-frame-rate and event data from the target domain, but there is a drawback in applying them online. In practical application, event cameras exhibit continuous changes in distribution with variations in lighting conditions. It is not feasible to acquire new data for novel scenes each time they arise and wait for the network to learn offline. In this regard, an online adaptation of the network to the environment is essential from a practical perspective and is well-known as Test-Time Adaptation (TTA) [47, 53, 63]. TTA assumes a setting that does not allow access to source data and prefers online learning for adaptation. In our case, the target domain lacks ground-truth data, which means no high-frame-rate video is available in the target domain. + +In this paper, we propose Test-time Adaptation for Event-based Video Frame Interpolation (TTA-EVF) to alleviate performance degradation in the face of domain shifts. As shown in Fig. 1, when encountering a new domain without performing adaptation, even well-trained networks exhibit performance drops. In contrast, TTA-EVF reliably and rapidly fine-tunes in the target domain through self-training, even without high-frame-rate videos. For TTA in event-based VFI, there are two main challenges: Firstly, even without exposure to HFR videos in the target domain, the network should adapt to the scene while simultaneously learning knowledge for HFR generation from LFR data. To address this challenge, we propose Reliable Pixel Sampling (RPS), a self-training scheme for selecting pixels from generated HFR videos, utilizing a confidence estimation based on a teacher-student framework [49]. The second challenge is overfitting when continuously provided with data that has a similar distribution within a successive sequence within online learning. To mitigate this, we employ a memory bank that stores only reliable samples from the generated images. We then propose Patch-Mixed Sampling (PMS) by blending reliable samples with current images, effectively preventing overfitting through an augmentation approach. + +The main contributions of our works are as follows: (I) We propose a novel Test-Time Adaptation (TTA) framework for event-based VFI. (II) We introduce a novel EventRGB Distribution Shift (ERDS) dataset for event-based video frame interpolation, where the distribution of the dataset is continually changing. (III) We compare our + +method with existing works and demonstrate its superiority by outperforming supervised methods. + +# 2. Related Works + +Event-based Video Frame Interpolation. Frame-based VFI methods [2, 3, 15, 20-22, 33, 35, 37, 38, 50, 60] have continuously achieved performance improvements; however, it still has weaknesses in handling complex motion and requires motion approximation when synthesizing intermediate frames at arbitrary time. Event-based VFI methods [30, 41, 44, 45, 55, 59, 64, 65], which utilize event cameras as an additional sensor [6, 7, 9, 28, 34, 45], enable not only handling complex motion but also interpolating intermediate frames for irregular objects. However, there is still room for improvement compared to frame-based VFI methods. Typically, frame-based VFI methods perform cross-domain performance evaluations, considering real scenario applications. Well-trained frame-based VFI methods exhibit domain generalization ability, maintaining considerable performance even on different domains without the need for additional training process for adaptation. On the other hand, event-based VFI experiences a significant drop in performance when moving to different domains [16]. Therefore, resolving cross-domain challenges in event-based VFI is vital for widespread applications. + +Test-Time Adaptation. Unlike Unsupervised Domain Adaptation [8, 48, 58], which necessitates offline learning for each domain change and relies on labeled data from the source domain, Test-Time Adaptation (TTA) is more challenging setup as it adapts to the target domain without accessing source data. In recent times, the significance of data privacy has escalated, leading to increased interest in TTA, a source-free domain adaptation approach. This is primarily due to the inefficiency of accessing source domain data during inference. TTA is actively being researched in the image domain [5, 14, 19, 27, 31, 32, 36, 40, 42, 47, 53, 63, 66], especially in areas like object recognition and segmentation. While there is an approach of its application in event-based visual recognition, Ev-TTA [23], there is still a lack of research in the domain of video frame interpolation, particularly in relation to events. The existing work, MetaVFI [11], proposed for the frame-based VFI task, learns to estimate scene-adaptive motion with minimal parameter updates through meta-learning [13]. However, this method is well-suited when there is not a significant domain gap between training and testing. We demonstrate that this approach is suboptimal when applied to event-based VFI with a substantial domain gap between the training and test sets. + +We propose the first-ever Test-Time Adaptation (TTA) approach for event-based Video Frame Interpolation (VFI). This is essential because event distributions can vary significantly due to external environmental factors and camera parameter settings even when using the same camera. + +![](images/1910b9000bc1557cb3ec1f47454fba3dd0162dbe3dc4b0155ad053fc59d0192e.jpg) +Figure 2. Overview of our Test-Time Adaptation for Event-based Video Frame Interpolation (TTA-EVF) framework. Our approach involves four steps for test-time training in Ev-VFI: 1) We calculate the adaptive threshold, $\rho$ , for RPS using the low frame rate image triples, $I_0, I_1$ , and $I_2$ with interval events $E_{0\rightarrow 1}$ and $E_{2\rightarrow 1}$ . 2) Reliable Pixel Sampling (RPS) is applied for arbitrary $t_0$ between 0 and 1, and arbitrary $t_1$ between 1 and 2, and in this process, even without intermediate images for training, reliable pixels are estimated through the student-teacher framework and self-training is performed. 3) Using the estimated intermediate frames $\hat{I}_{t_0}$ and $\hat{I}_{t_1}$ along with $I_1$ , we form a new triplet, $\mathcal{V}$ , check if the entire sampled generated image is reliable, and in that case, update the memory bank. 4) Patch-Mixed Sampling (PMS) regularizes the network by mixing the generated videos, $\mathcal{V}$ with videos from the memory bank, $\tilde{\mathcal{V}}$ . + +Furthermore, our method follows the recent trend of TTA by addressing the adaptation to unknown distribution shifts encountered during test-time in an online manner. + +# 3. Proposed Method + +# 3.1. Overview + +Given key frames $I_0, I_1$ and the inter-frame event stream $\mathcal{E}_{0\rightarrow 1}$ , Event-based Video Frame Interpolation (Ev-VFI) networks aim to generate the intermediate frame $I_{t}$ on an arbitrary time $t \in [0,1]$ . Following the previous works, we convert a event stream as a voxel grid [67]. We denote $E_{t_a \to t_b}$ as a voxel grid form between times $t_a$ and $t_b$ . + +TTA-EVF adapts a pre-trained Ev-VFI network trained on the source domain to a target domain with domain shift in the device and setting. In the pre-training phase, we train the network using source data with high-frame rate videos in a supervised manner (Sec. 3.2). We then utilize this pretrained Ev-VFI network for test-time adaptation using target data without accessing the source data (Sec. 3.3). + +# 3.2. Pre-training on Source Dataset + +In general, TTA methods perform test-time training stably by updating only a subset of network parameters, such as Batch Normalization (BN) [18]. As mentioned in [53], updating the full parameters at test-time, especially in an online manner, leads to instability and makes it challenging to achieve optimal performance. However, in the case of low-level vision tasks like VFI, normalization layer (e.g. BN) + +![](images/1059df73b9d437f508be764524a1da6673eeb85b7f3135d0264724301565e7b4.jpg) +Figure 3. The architecture of Norm-Residual (NR) Block. + +is often avoided by most networks [20, 24, 51] since it removes range flexibility and can lead to a performance decrease [29]. Some alternatives for handling this issue include using slightly more flexible normalization techniques (e.g. instance normalization) or selectively updating specific convolutional layers. Updating only the parameters of the normalization layer in VFI can lead to sub-optimal performance when there is significant variance between image patches in the training and test environments. Conversely, updating specific convolution layers with more parameters can remove all statistical modulation, causing instability during the test-time training process. To address this issue, we propose a solution by designing a Norm-Residual (NR) Block that includes both convolution and normalization layers. This approach aims to strike a balance between updating parameters effectively and maintaining statistical modulation during the learning process. + +As shown in Fig. 3, given a feature $\mathbf{Z}$ , NR Block first split a feature into two feature maps $\mathbf{Z}^{(1)}$ and $\mathbf{Z}^{(2)}$ along the channel dimension. The first map $\mathbf{Z}^{(1)}$ is skipped to keep + +the context information, and rest feature $\mathbf{Z}^{(2)}$ is normalized by normalization layer. Then, we fuse the two features by concatenate and convolutional layer. Finally, we apply a element-wise summation with original feature $\mathbf{Z}$ to obtain the output $\hat{\mathbf{Z}}$ . We use the group normalization [57] in normalization layer, which shows the best performance in our experiments. For efficiency, we exclusively deploy the NR Block in the encoders of the networks, replacing the conventional residual blocks of existing networks. Therefore, TTA-EVF is a general and model-agnostic method specifically crafted to be adaptable to any network architecture used in Ev-VFI. Further details can be found in the supple. + +Network Optimization. The modified Ev-VFI network with the NR Block is trained on the source dataset using ground truth with high-frame-rate videos. Specifically, we optimize the Ev-VFI network, $\mathcal{F}$ , by the supervision as: + +$$ +\underset {\theta (\mathcal {F})} {\arg \min } \left(\mathcal {L} _ {\mathcal {F}} \left(\hat {I} _ {t}, I _ {t}\right)\right), \tag {1} +$$ + +where $\mathcal{L}_{\mathcal{F}}$ is the loss function proposed by each Ev-VFI network and the generated intermediate frame $\hat{I}_t$ is obtained by $\hat{I}_t = \mathcal{F}(I_0, I_1, E_{0 \to t}, E_{1 \to t})$ . + +# 3.3. Test-time Adaptation on Target Dataset + +In our TTA setup, the model generates intermediate frames at each iteration and simultaneously updates the parameters of the NR Block. Since the target domain only contains LFR frames, not HFR frames, self-supervised optimization using only LFR is required. + +# 3.3.1 Reliable Pixel Sampling (RPS) + +In the target domain, even without High-Frame-Rate (HFR) ground truth, when given Low-Frame-Rate (LFR) videos $I_0$ , $I_1$ , $I_2$ as triplets, and the events between them $E_{0\to 1}$ and $E_{2\to 1}$ , the network can be trained using the following self-supervised loss: + +$$ +\mathcal {L} _ {S} = \mathcal {L} _ {\mathcal {F}} (\hat {I} _ {1}, I _ {1}), \tag {2} +$$ + +where $\hat{I}_1 = \mathcal{F}(I_0, I_2, E_{0\rightarrow 1}, E_{2\rightarrow 1})$ . However, this is problematic as the network cannot learn the knowledge of generating HFR because it is generating LFR again from temporally subsampled LFR images, leading to the loss of the ability to generate HFR trained from the source domain. As a way to address this, applying the cyclic consistency loss [16, 39] from the existing unsupervised VFI is one option. However, directly applying these methods to TTA of Ev-VFI is intractable, since there is no guarantee that the generated intermediate frames will be clean and artifact-free on target domains. Therefore, we need a new paradigm. + +As shown in Fig. 2, given the $I_0$ , $I_1$ , $I_2$ triplet, we design Reliable Pixel Sampling (RPS) for the left intermediate time + +$t_0$ and the right intermediate time $t_1$ . Here, $t_0$ and $t_1$ are selected at random from the time intervals 0 to 1 and 1 to 2, respectively. Since the RPS process is identical for both $t_0$ and $t_1$ , we introduce it here for $t_0$ only. Reliable data sampling without the intermediate ground truth frames, $I_{t_0}$ , is highly challenging and can lead to unstable learning if estimated incorrectly. To address this, we adopt the teacher-student framework [49] commonly used for pseudo labeling [10, 58] in other tasks. In this framework, the student learns from pseudo labels through self-training, while the teacher is updated solely from the momentum using the parameters of the student. The motivation behind proposed RPS stems from the observation that pixels estimated similarly by two networks with slightly different parameters, yet achieving similar performance in VFI, are likely to be reliable, whereas parts that are difficult to estimate and prone to errors would result in differing predictions between the two networks. Hence, we classify pixels consistently estimated by both teacher-student networks as reliable pixels, while considering pixels with inconsistent predictions as unreliable. Given intermediate frames generated from the student network $\mathcal{F}_{\mathcal{S}}$ and the teacher network $\mathcal{F}_{\mathcal{T}}$ denoted as $\hat{I}_{t_0}^{\mathcal{S}}$ and $\hat{I}_{t_0}^{\mathcal{T}}$ respectively, the reliable map, $R_{t_0}$ , with spatial dimensions $H\times W\times 1$ (where $H$ and $W$ represent the height and width of the image $I_0$ ) is defined as follows: + +$$ +R _ {t _ {0}} (i, j) = \left\{ \begin{array}{l} 1, \text {i f} \| \hat {I} _ {t _ {0}} ^ {\mathcal {S}} (i, j) - \hat {I} _ {t _ {0}} ^ {\mathcal {T}} (i, j) \| < \rho , \\ 0, \text {o t h e r w i s e}, \end{array} \right. \tag {3} +$$ + +where $i = 1,\dots ,H$ $j = 1,\ldots ,W$ and $\rho$ denote the threshold. When estimating such reliability, it is crucial to determine a specific threshold. However, assigning the same threshold to all scenes is inappropriate since the performance of VFI significantly varies across scenes. To achieve adaptive thresholding for each scene, we utilize the LFR frames $I_{0}$ and $I_{2}$ , along with the interceding events $E_{0\rightarrow 1}$ and $E_{2\rightarrow 1}$ , to estimate $\hat{I}_1$ . Subsequently, we compute a pixel-wise error map, $e = \| \hat{I}_1 - I_1\|$ , between $\hat{I}_1$ and $I_{1}$ , allowing us to calculate $\rho$ , using the average value of $e$ . + +However, this approach encounters issues when dealing with scenes that contain motionless, static regions (e.g., backgrounds). In such cases, these areas exhibit minimal pixel-wise errors, which leads to an inadequate adaptive threshold calculation for the scene if these pixels are included in the average calculation. To address this, we propose an approach that excludes motionless regions based on optical flow to refine the threshold calculation. Since most Ev-VFI methods compute optical flow towards the intermediate frame from key frames, we can directly calculate a motion magnitude map [54], $M$ , from the intermediate output of optical flow, denoted as $F_{10}$ and $F_{12}$ , without requiring + +# Algorithm 1 PMS Pseudo Code + +Input: Generated HFR set $\mathcal{V} = \{\hat{I}_{t_0}, I_1, \hat{I}_{t_1}, E_{t_0 \to 1}, E_{t_1 \to 1}\}$ + +1: Initialize $\mathcal{L}_M\gets 0$ +2: Get reliability of $\nu$ based on Eq.(9) +3: Sample $\mathcal{V} = \{\bar{I}_{\tilde{t}_0},\bar{I}_1,\bar{I}_{\tilde{t}_1},\bar{E}_{\tilde{t}_0\to 1},E_{\tilde{t}_1\to 1}\}$ from $\mathcal{Q}$ +4: Sample mixing ratio $\lambda \in (0.7, 1.0)$ +5: if $\nu$ is reliable then +6: $\mathcal{M}_{t_0} = \lambda \hat{I}_{t_0} + (1 - \lambda)\tilde{I}_1, \mathcal{M}_{t_1} = \lambda \hat{I}_{t_1} + (1 - \lambda)\tilde{I}_1$ +7: $\mathcal{M}_1 = \lambda I_1 + (1 - \lambda)\tilde{I}_1$ +8: $\mathcal{D}_{t_0\to 1} = E_{t_0\to 1},\mathcal{D}_{t_1\to 1} = E_{t_1\to 1}$ +9: Push $\mathcal{V}$ into memory bank $\mathcal{Q}$ + +10: else + +11: $\mathcal{M}_{t_0} = (1 - \lambda)I_1 + \lambda \tilde{I}_{\tilde{t}_0}$ , $\mathcal{M}_{t_1} = (1 - \lambda)I_1 + \lambda \tilde{I}_{\tilde{t}_1}$ . +12: $\mathcal{M}_1 = (1 - \lambda)I_1 + \lambda \tilde{I}_1$ +13: $\mathcal{D}_{t_0\to 1} = \bar{E}_{\tilde{t}_0\to 1},\mathcal{D}_{t_1\to 1} = \bar{E}_{\tilde{t}_{1}\to 1}$ +14: if $\mathrm{len}(\mathcal{Q}) > \mathrm{max}$ queue size then +15: Pop oldest ones out of $\mathcal{Q}$ +16: $\mathcal{L}_M\gets \mathcal{L}_{\mathcal{F}}(\mathcal{M}_1,\mathcal{F}_S(\mathcal{M}_{t_0},\mathcal{M}_{t_1},\mathcal{D}_{t_0\to 1},\mathcal{D}_{t_1\to 1}))$ + +Output: $\mathcal{L}_M$ + +ing additional computations as: + +$$ +M (i, j) = \left(\| F _ {1 0} (i, j) \| + \| F _ {1 2} (i, j) \|\right) / 2, \tag {4} +$$ + +where $\|\cdot\|$ denote the magnitude of the motion vector. The motion mask $\widehat{M}$ , which excludes motionless areas based on the motion magnitude map for each pixel, can be created as follows: + +$$ +\widehat {M} (i, j) = \left\{ \begin{array}{l l} 1 & \text {i f} M (i, j) \geq m, \forall i, \forall j \\ 0 & \text {o t h e r w i s e ,} \end{array} \right. \tag {5} +$$ + +where, $m$ represents a predefined minimal amount of motion, and in our experiment, we use a value of 1. Then, we calculate the scene adaptive threshold as: + +$$ +\rho = \frac {1}{| \widehat {M} |} \sum \widehat {M} \odot e. \tag {6} +$$ + +We define the self-training loss between the teacher and student networks using a pixel-wise reliable map in Eq.(3) and Eq.(6): + +$$ +\mathcal {L} _ {R} ^ {t _ {0}} = \left\| \hat {I} _ {t _ {0}} ^ {\mathcal {S}} (i, j) - \hat {I} _ {t _ {0}} ^ {\mathcal {T}} (i, j) \right\| \odot R _ {t _ {0}}, \tag {7} +$$ + +Finally, the reliable pixel-wise loss, taking into account both intermediate times $t_0$ and $t_1$ , is defined as follows: + +$$ +\mathcal {L} _ {R} = \mathcal {L} _ {R} ^ {t _ {0}} + \mathcal {L} _ {R} ^ {t _ {1}}, \tag {8} +$$ + +# 3.3.2 Patch-Mixed Sampling (PMS) + +The issues that can arise in TTA include overfitting to specific scenes due to the sequential data for the same scene being continuously fed, as well as the problem of catastrophic forgetting [4, 26] where existing knowledge is lost. + +![](images/967d2dd49df540049f3700bd81fe0dcdef83d7bfb79f55ddc063184ae4673f38.jpg) +Figure 4. The examples of patch-mixed sampling (PMS) process. + +![](images/b7df87b8dadc8a0ab456d0427afcce87026021b561a2b727ddaf4b089776bd16.jpg) + +To address this, we propose the Patch-Mixed Sampling (PMS) method. In order to avoid overfitting to specific scenes, PMS involves sampling from the memory bank, which accumulates reliable samples from the generated videos of each iteration. The detailed process of PMS is provided in Algorithm 1. + +Reliability Check. Denoting the HFR video and event set, including $\hat{I}_{t_0}$ and $\hat{I}_{t_1}$ , from the RPS process as $\mathcal{V}$ , we assess the reliability for the generated frames. If the pixel-wise reliability map $R$ obtained from the RPS module exceeds $\eta$ fraction of the resolution of the entire image, we deem the generated images as reliable. This is outlined as follows: + +$$ +\text {R e l i a b l i t y} = \left\{ \begin{array}{l} \text {T r u e , i f} \frac {\| R _ {t _ {0}} \|}{H \times W} > \eta \text {a n d} \frac {\| R _ {t _ {1}} \|}{H \times W} > \eta \\ \text {F a l s e , o t h e r w i s e ,} \end{array} \right. \tag {9} +$$ + +Patch Mixing. Figure 4 shows an example of a patch mixing process. If the currently generated HFR video set $\mathcal{V}$ is reliable (left of Fig. 4), we fetch a single intermediate frame $\tilde{I}_1$ from the memory bank and blend it. Since the mixed frames, $M_{t_0}, M_1$ , and $M_{t_1}$ , all incorporate the same $\tilde{I}_1$ , any events that require pixel changes can't arise from $\tilde{I}_1$ , thus we directly retain the events $E_{t_0 \to 1}$ and $E_{t_1 \to 1}$ from $\mathcal{V}$ . On the other hand, if $\mathcal{V}$ is unreliable (right of Fig. 4), we retrieve both images and event sets from the memory bank. From the currently generated $\mathcal{V}$ , we only fetch $I_1$ and blend it. Similar to before, since the same $I_1$ is blended into all three images, events originating from $I_1$ do not occur. Therefore, we directly retain the memory bank events $\tilde{E}_{\tilde{t}_0 \to 1}$ and $\tilde{E}_{\tilde{t}_1 \to 1}$ . For the mixing ratio of blending images, we adopt a beta distribution [62] that generates values between 0.7 and 1.0. We assign a larger ratio to the reliable set among the two sets being blended. Finally, using the patch-mixed sets, we train the student network of the RPS using the loss of Ev-VFI as follows: + +$$ +\mathcal {L} _ {M} = \mathcal {L} _ {\mathcal {F}} \left(\mathcal {M} _ {1}, \mathcal {F} _ {\mathcal {S}} \left(\mathcal {M} _ {t _ {0}}, \mathcal {M} _ {t _ {1}}, \mathcal {D} _ {t _ {0} \rightarrow 1}, \mathcal {D} _ {t _ {1} \rightarrow 1}\right)\right). \tag {10} +$$ + +In summary, the total objective $\mathcal{L}$ of TTA-EVF is as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {S} + \lambda_ {1} \mathcal {L} _ {R} + \lambda_ {2} \mathcal {L} _ {M} \tag {11} +$$ + +where $\lambda_1, \lambda_2$ are the hyper-parameters. + +![](images/a1de13773749c1f8bae56c306ed1257e8241fd74af58b5e492a41d076cb5a74e.jpg) +Figure 5. Visual comparisons of interpolated frames on multiple cross-domain settings using various models. + +# 4. Experiments + +# 4.1. Implementation Details + +All networks are trained using an Adam optimizer [25] with a batch size of 2 and a learning rate of $1 \times 10^{-4}$ , for both the source and target domains. For efficiency, during online training in the target domain, we crop images to a resolution of $320 \times 320$ . In our approach, we perform network updates only once for each triplet provided. For instance, when interpolating 7 skipped frames, we generate a total of 14 frames from $I_0$ , $I_1$ , and $I_2$ , and update the network parameters only once using the loss from Eq.(11), making it highly efficient. All experiments are conducted on a single NVIDIA Titan RTX. We set the $\eta$ as 0.9 in Eq.(9), bin size of voxel grid as 5, and $\lambda_1, \lambda_2$ as 1, 0.5 in Eq.(11). + +# 4.2. Test-Time Adaptation Results + +During the pre-training phase, we utilize the train sequences from the source domain for learning, and in the target domain, we follow the same sequential online approach as in the traditional TTA setup [53], using only the test sequences without utilizing the train set. + +Our framework can be applied to existing Ev-VFI methods regardless of the model. We conduct experiments by integrating it with TimeLens [51] and CBMNet [24], using publicly available code provided by the authors. For TimeReplayer [16], as public code is unavailable, we re-implement it based on the foundational code [20] as described in the paper. To the best of our knowledge, we are the first to propose test-time adaptation for event-based video frame interpolation. Therefore, for comparison, we adopt the previously proposed meta-learning approach [11] for frame-based VFI and the self-supervised optimization [16] for Ev-VFI. When training at test time, the performance tends to be low if we simply apply the existing methods [16] as is. Therefore, we keep our NR Block + +Table 1. Quantitative comparisons with existing method in cross-domain datasets. S and T denote the source and target datasets. + +
S TMethodsTimeReplayer [16]TimeLens [51]CBMNet [24]
PSNRSSIMPSNRSSIMPSNRSSIM
B RNo Adapt32.240.915432.040.907331.980.9081
Meta [11]32.750.918032.430.911232.210.9109
Cycle [16]32.870.920833.010.916533.090.9214
TTA-EVF34.070.928634.240.929934.420.9312
B HNo Adapt29.780.837829.960.841329.870.8395
Meta [11]30.130.839330.140.834130.220.8407
Cycle [16]30.320.810430.660.844130.750.8466
TTA-EVF31.450.855132.130.859932.070.8584
R BNo Adapt21.940.643422.320.674521.890.6427
Meta [11]22.830.681823.140.702022.750.6822
Cycle [16]23.980.715424.110.722324.430.7305
TTA-EVF25.850.763826.910.775227.160.7768
R HNo Adapt27.830.815827.610.812127.920.8222
Meta [11]28.110.820628.320.823428.540.8299
Cycle [16]28.430.830629.120.838929.430.8410
TTA-EVF30.330.846431.830.858731.860.8590
+ +update and TTA setup intact while only incorporating the cyclic self-supervised mechanism [16] into Eq.(11). + +TTA in Cross-domain Datasets. We conduct cross-domain experiments using three existing high-quality real event datasets: BS-ERGB [52], HS-ERGB [51], and HighREV [46]. We refer to each datasets as B, H, and R, respectively. In each experiment, we alternate between the source and target datasets, but HS-ERGB exclusively provides with test sequences and is used solely as the target dataset. + +Table 1 presents the quantitative results in cross-domain datasets. We can confirm that TTA-EVF has significant advantages over other methods. First, Meta [11] using the meta-learning shows an overall but subtle improvement in performance. It indicates that it is not particularly effective in adapting to entirely new domains, especially when the distribution of event data changes, operating independently for each scene. In contrast, we continuously adapt online to scenes, learning in an optimally directional manner for the entire domain rather than for each individual scene. Further- + +Table 2. Quantitative comparisons with existing method in continuous domain shifting datasets. + +
NetworksMethodsTarget Sequences
123456Total
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
TimeReplayer [16]No Adapt34.670.909028.520.670733.200.748030.940.702434.870.906936.500.905432.810.7960
Meta [11]33.930.906328.650.673232.340.730331.140.704534.910.908336.950.910632.910.7996
Cycle [16]34.740.909827.430.667432.540.734531.220.705135.100.909137.050.909732.990.7968
TTA-EVF35.050.910129.440.676533.470.752931.900.717536.790.921337.270.914433.710.8040
TimeLens [51]No Adapt33.910.895528.170.666432.710.742830.580.702433.990.900234.290.892432.070.7896
Meta [11]34.050.900127.840.664232.800.741230.150.702233.690.890734.490.900432.090.7898
Cycle [16]34.010.900827.960.668032.410.743530.220.704134.070.900834.980.901332.140.7906
TTA-EVF34.410.905528.630.673733.090.749331.290.710034.940.905936.050.904032.800.7974
CBMNet [24]No Adapt32.700.901427.190.659732.050.732630.140.701532.240.896236.820.911731.420.7879
Meta [11]32.840.902627.320.660832.140.733430.160.701232.310.897436.650.910131.480.7885
Cycle [16]33.620.909727.820.669532.610.737230.340.705832.850.901337.190.914532.410.7988
TTA-EVF35.820.913729.250.680933.260.754031.650.720335.120.910437.540.928833.530.8023
+ +![](images/adda0ba7ec56e92ef9049e96fef547b5c23b936b0bbf9529055b774277fa1125.jpg) +Figure 6. Visual comparisons of interpolated frames on continuous domain shifting settings. Please zoom for better view. + +Table 3. Comparison results between our TTA method and existing supervised/unsupervised methods in the target domain. * denotes that the values are taken from the original paper. + +
STMethodsOnlineTimeReplayer [16]TimeLens [51]
PSNRSSIMPSNRSSIM
RRSupervised Unsupervised--32.81*0.901*
32.420.898--
BRTTA-EVF✓34.070.92934.240.930
HH-farSupervised Unsupervised--32.31*0.869*
30.07*0.834*--
BH-farTTA-EVF✓31.720.85932.420.870
HH-closeSupervised Unsupervised--31.68*0.835*
29.83*0.816*--
BH-closeTTA-EVF✓30.950.82932.110.860
+ +more, when we examine the results of Cycle [16], which incorporate a self-supervised loss into our TTA approach, it becomes evident that simply maintaining continuous online updates does not yield significant performance gains. In the case of Cycle [16], learning cyclically with generated videos during limited online updates is unable to amplify adaptation ability in situations where unavoidable errors occur as domains change. In contrast, our method identifies significant errors in pixel and samples for each scene, producing substantial improvements even within the same setting. Figure 5 also demonstrates the effectiveness of our method across various datasets and networks. + +In addition, we present the comparison of our TTA approaches with supervised/unsupervised methods in Table 3. Despite the fact that our method involves online updates on the target dataset after pre-trained from other source domain datasets, it outperforms methods trained on the target dataset using supervised or unsupervised approaches. For instance, on the HighREV (R) dataset using the TimeLens [51] model, our approach achieved a $1.43\mathrm{dB}$ higher PSNR than supervised learning, even when trained in a different domain, highlighting its ability to bridge domain gaps and improve the versatility of event-based VFI. + +TTA in Continuous Domain Shift. We experiments in a highly challenging setting where the distribution of the dataset continually undergoes significant changes. Since there were no existing datasets with consistently changing domains in high-quality RGB-Event data, we acquire the novel Event-RGB Distribution Shift (ERDS) dataset. In the ERDS dataset, the training dataset includes HFR videos with event data in typical lighting conditions. The test dataset, on the other hand, includes scenarios where lighting conditions are consistently altered, and the event trigger threshold is adjusted accordingly to obtain sufficient information from the events. We present the results in Table 2. In the highly challenging setting of continuous domain shift + +Table 4. Ablation study of the proposed components. + +
LSNR BlockLRLMCrossContinuous
No Adaptation29.9631.42
✓30.4231.98
✓✓31.5332.76
✓✓✓31.8433.01
✓✓✓31.9233.15
✓✓✓✓32.1333.53
+ +Table 5. Results based on various parameter update method. + +
UpdatePSNR ↑SSIM ↑
Full Parameters31.040.8489
Encoder Parameters30.890.8455
Batch Normalization (BN) [18]30.440.8423
Layer Normalization (LN) [1]30.920.8477
NR Block w/. BN [18]31.570.8502
NR Block w/. LN [1]31.940.8524
NR Block w/. GN [57]32.130.8599
+ +ing, other methods fail to deliver significant improvements. In contrast, our method can achieve substantial performance gains compared to competing methods. Figure 6 clearly demonstrates the strengths of our method. + +# 5. Ablation Study + +In this section, we conduct experiments to access the efficacy of the proposed approaches. In the cross-dataset setting, we use the TimeLens [51] model, training it with BS-ERGB and evaluating it on HS-ERGB. In the continuous setting, we utilize the CBMNet [24] model. + +Effectiveness of each components. We conduct experiments in Tab. 4 while adding each component of TTA-EVF one by one. When optimizing the network using only the loss $\mathcal{L}_S$ in Eq.(2) under the TTA setup, it can be observed that there is not a significant improvement in performance. This is because directly self-training with only LFR videos prevents the network from learning the knowledge of generating HFR, and as training progresses, it loses the ability to generate HFR learned from the source domain. In contrast, it can be observed that there is an impressive improvement in performance with the addition of the proposed modules. Ablation study on NR Block. We perform ablation by keeping all our other modules intact while changing only the update methods. As shown in Tab. 5, batch normalization (BN) [18] and layer normalization (LN) [1] lead to a decrease in performance. The proposed NR Block update method significantly improves performance, with group normalization (GN) [57] being particularly effective. The benefits of the adaptive threshold in RPS. We calculate the threshold for the reliable map between the student and teacher networks adaptively for each scene in the RPS module. Table 7 compares the adaptive thresholding method for the reliable map in the RPS module with a fixed-value thresholding approach. + +The ablation study of mixing ratio $\lambda$ in PMS. Table 6 presents an analysis of the mixing ratio between the samples in the memory bank and the generated videos in PMS. + +Table 6. Performance based on minimum value of mixing ratio $\lambda$ . + +
min(λ)0.50.60.70.80.9
Cross31.9232.0632.1332.0431.93
Continuous33.3033.4933.5333.4233.36
+ +Table 7. Analysis on the reliable threshold. + +
ρ∞0.20.10.050.01Adaptive (Ours)
Cross30.8730.9331.4431.8531.4232.13
Continuous32.2132.6433.0232.9732.7833.53
+ +Table 8. Inference time on different updating intervals. In the case of TTA-EVF, the inference time includes the training time as well. + +
S TMethods (TimeLens [51])TimePSNRSSIM
B H-closeNo adapt80 min29.960.8413
TTA-EVF (Interval=5)81 min31.440.8541
TTA-EVF (Interval=3)82 min31.620.8571
TTA-EVF (Interval=1)88 min32.110.8598
+ +We select the mixing ratio $\lambda$ uniformly, but it can be set with a minimum value. The best performance is observed when the minimum value, $\min (\lambda)$ , is set to 0.7, demonstrating robustness to parameter variations. + +# 6. Analysis of Updating Interval + +We present the performance based on the updating interval in Table 8. Our TTA method is highly cost-effective despite performing training during the inference process. TTA-EVF, which updates only once during the multiple inferences when an video triplet is input, exhibits minimal time difference compared to simple inference. When the interval is increased, the time required is almost the same as simple inference, yet the quality of the generated videos is significantly higher. For example, with an interval of 5, there is only a 1 minute difference in total inference time, but the PSNR is improved by $1.48\mathrm{dB}$ . + +# 7. Conclusion + +In this paper, we present a Test-Time Adaptation (TTA) for event-based video frame interpolation. We effectively perform online TTA using the newly proposed modules in both cross-domain and continuous domain-shifting settings. Additionally, we acquire the ERDS dataset, containing high-quality RGB and event datasets with continuous distribution changes. Our TTA-EVF, despite online learning, shows minimal speed differences compared to simple inference and outperforms both supervised and unsupervised methods in cross-domain experiments, showcasing its effectiveness. + +Acknowledgements. This research was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF2022R1A2B5B03002636) and the Challengeable Future Defense Technology Research and Development Program through the Agency For Defense Development (ADD) funded by the Defense Acquisition Program Administration (DAPA) in 2024 (No.912768601). + +# References + +[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.8 +[2] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3703-3712, 2019. 1, 2 +[3] Wenbo Bao, Wei-Sheng Lai, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Memc-net: Motion estimation and motion compensation driven neural network for video interpolation and enhancement. IEEE transactions on pattern analysis and machine intelligence, 43(3):933-948, 2019. 1, 2 +[4] Magdalena Biesialska, Katarzyna Biesialska, and Marta R Costa-Jussa. Continual lifelong learning in natural language processing: A survey. arXiv preprint arXiv:2012.09823, 2020.5 +[5] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 2 +[6] Hoonhee Cho and Kuk-Jin Yoon. Event-image fusion stereo using cross-modality feature propagation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 454-462, 2022. 2 +[7] Hoonhee Cho and Kuk-Jin Yoon. Selection and cross similarity for event-image deep stereo. In European Conference on Computer Vision, pages 470-486. Springer, 2022. 2 +[8] Hoonhee Cho, Jegyeong Cho, and Kuk-Jin Yoon. Learning adaptive dense event stereo from the image domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17797-17807, 2023. 2 +[9] Hoonhee Cho, Yuhwan Jeong, Taewoo Kim, and Kuk-Jin Yoon. Non-coaxial event-guided motion deblurring with spatial alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12492-12503, 2023. 2 +[10] Hoonhee Cho, Hyeonseong Kim, Yujeong Chae, and Kuk-Jin Yoon. Label-free event-based object recognition via joint learning with image reconstruction from events. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19866-19877, 2023. 4 +[11] Myungsub Choi, Janghoon Choi, Sungyong Baik, Tae Hyun Kim, and Kyoung Mu Lee. Scene-adaptive video frame interpolation via meta-learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9444-9453, 2020. 2, 6, 7 +[12] Myungsub Choi, Heewon Kim, Bohyung Han, Ning Xu, and Young Mu Lee. Channel attention is all you need for video frame interpolation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10663-10671, 2020. 1 +[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2 + +[14] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. Advances in Neural Information Processing Systems, 35:29374-29385, 2022. 2 +[15] Shurui Gui, Chaoyue Wang, Qihua Chen, and Dacheng Tao. Featureflow: Robust video interpolation via structure-to-texture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14004-14013, 2020. 1, 2 +[16] Weihua He, Kaichao You, Zhendong Qiao, Xu Jia, Ziyang Zhang, Wenhui Wang, Huchuan Lu, Yaoyuan Wang, and Jianxing Liao. Timereplayer: Unlocking the potential of event cameras for video interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17804-17813, 2022. 1, 2, 4, 6, 7 +[17] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 1 +[18] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 3, 8 +[19] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 2 +[20] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9000-9008, 2018. 1, 2, 3, 6 +[21] Shihao Jiang, Yao Lu, Hongdong Li, and Richard Hartley. Learning optical flow from a few matches. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16592-16600, 2021. +[22] Rico Jonschkowski, Austin Stone, Jonathan T Barron, Ariel Gordon, Kurt Konolige, and Anelia Angelova. What matters in unsupervised optical flow. In European Conference on Computer Vision, pages 557-572. Springer, 2020. 1, 2 +[23] Junho Kim, Inwoo Hwang, and Young Min Kim. Ev-tta: Test-time adaptation for event-based object recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17745-17754, 2022. 2 +[24] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 1, 3, 6, 7, 8 +[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neu + +ral networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 5 +[27] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 2 +[28] Jinxiu Liang, Yixin Yang, Boyu Li, Peiqi Duan, Yong Xu, and Boxin Shi. Coherent event guided low-light video enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10615-10625, 2023. 2 +[29] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 3 +[30] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. Learning event-driven video deblurring and interpolation. ECCV, 2020. 2 +[31] Wei Lin, Muhammad Jehanzeb Mirza, Mateusz Kozinski, Horst Possegger, Hilde Kuehne, and Horst Bischof. Video test-time adaptation for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22952-22961, 2023. 2 +[32] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808-21820, 2021. 2 +[33] Huaijia Lin Jiangbo Lu Liying Lu, Ruizheng Wu and Jiaya Jia. Video frame interpolation with transformer. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2 +[34] Weng Fei Low and Gim Hee Lee. Robust e-nerf: Nerf from sparse & noisy events under non-uniform motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 18335–18346, 2023. 2 +[35] Kunming Luo, Chuan Wang, Shuaicheng Liu, Haoqiang Fan, Jue Wang, and Jian Sun. Upflow: Upsampling pyramid for unsupervised optical flow learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1045-1054, 2021. 1, 2 +[36] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 2 +[37] Junheum Park, Keunsoo Ko, Chul Lee, and Chang-Su Kim. Bmbc: Bilateral motion estimation with bilateral cost volume for video interpolation. In European Conference on Computer Vision, 2020. 1, 2 +[38] Junheum Park, Chul Lee, and Chang-Su Kim. Asymmetric bilateral motion estimation for video frame interpolation. In International Conference on Computer Vision, 2021. 1, 2 +[39] Fitssum A Reda, Deqing Sun, Aysegul Dundar, Mohammad Shoeybi, Guilin Liu, Kevin J Shih, Andrew Tao, Jan Kautz, + +and Bryan Catanzaro. Unsupervised video interpolation using cycle consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 892-900, 2019. 2, 4 +[40] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 2 +[41] Wei Shang, Dongwei Ren, Dongqing Zou, Jimmy S. Ren, Ping Luo, and Wangmeng Zuo. Bringing events into video deblurring with non-consecutively blurry frames. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4531-4540, 2021. 2 +[42] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. Advances in Neural Information Processing Systems, 35:14274-14289, 2022. 2 +[43] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. Xvfi: extreme video frame interpolation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1 +[44] Chen Song, Qixing Huang, and Chandrajit Bajaj. E-cir: Event-enhanced continuous intensity recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7803-7812, 2022. 2 +[45] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European Conference on Computer Vision (ECCV), 2022. 2 +[46] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18043-18052, 2023. 6 +[47] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 2 +[48] Zhaoning Sun, Nico Messikommer, Daniel Gehrig, and Davide Scaramuzza. Ess: Learning event-based semantic segmentation from still images. ArXiv, abs/2203.10016, 2022. 2 +[49] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2, 4 +[50] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 1, 2 +[51] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation + +tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16155-16164, 2021. 1, 3, 6, 7, 8 +[52] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. arXiv preprint arXiv:2203.17191, 2022. 1, 6 +[53] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 2, 3, 6 +[54] Yusheng Wang, Yunfan Lu, Ye Gao, Lin Wang, Zhihang Zhong, Yinqiang Zheng, and Atsushi Yamashita. Efficient video deblurring guided by motion magnitude. In European Conference on Computer Vision, pages 413-429. Springer, 2022. 4 +[55] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based blurry frame interpolation under blind exposure. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1588-1598, 2023. 2 +[56] Song Wu, Kaichao You, Weihua He, Chen Yang, Yang Tian, Yaoyuan Wang, Ziyang Zhang, and Jianxing Liao. Video interpolation by event-driven anisotropic adjustment of optical flow. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1 +[57] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 4, 8 +[58] Ruihao Xia, Chaoqiang Zhao, Meng Zheng, Ziyan Wu, Qiyu Sun, and Yang Tang. Cmda: Cross-modality domain adaptation for nighttime semantic segmentation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 21515-21524, 2023. 2, 4 +[59] Fang Xu, Lei Yu, Bishan Wang, Wen Yang, Gui-Song Xia, Xu Jia, Zhendong Qiao, and Jianzhuang Liu. Motion deblurring with real events. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2583-2592, 2021. 2 +[60] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. In NeurIPS, 2019. 1, 2 +[61] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127(8): 1106-1125, 2019. 1 +[62] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 5 +[63] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems, 35: 38629-38642, 2022. 2 +[64] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17765-17774, 2022. 2 +[65] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17765-17774, 2022. 2 +[66] Yizhe Zhang, Shubhankar Borse, Hong Cai, and Fatih Porikli. Auxadapt: Stable and efficient test-time adaptation for temporally consistent video semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2339-2348, 2022. 2 +[67] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based optical flow using motion compensation. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 3 \ No newline at end of file diff --git a/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/images.zip b/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3b947db0718a3f613a45a46648f6df2ac65bdc70 --- /dev/null +++ b/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81fa52b3f32c4a858133e424b4e9b9f64d9cdf18bc008bd14d7f3cb10158fa57 +size 863899 diff --git a/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/layout.json b/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d5ecb0ebcc867260b3e71a843a97931ccc800caa --- /dev/null +++ b/2024/TTA-EVF_ Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation/layout.json @@ -0,0 +1,10182 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 547, + 140 + ], + "type": "text", + "content": "TTA-EVF: Test-Time Adaptation for Event-based Video Frame Interpolation via Reliable Pixel and Sample Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 161, + 451, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 161, + 451, + 188 + ], + "spans": [ + { + "bbox": [ + 143, + 161, + 451, + 188 + ], + "type": "text", + "content": "Hoonhee Cho, Taewoo Kim, Yuhwan Jeong, and Kuk-Jin Yoon KAIST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 162, + 191, + 428, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 191, + 428, + 204 + ], + "spans": [ + { + "bbox": [ + 162, + 191, + 428, + 204 + ], + "type": "text", + "content": "{gnsgnsgml, intelpro, jeongyh98, kjyoon}@kaist.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 258, + 290, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 290, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 290, + 558 + ], + "type": "text", + "content": "Video Frame Interpolation (VFI), which aims at generating high-frame-rate videos from low-frame-rate inputs, is a highly challenging task. The emergence of bio-inspired sensors known as event cameras, which boast microsecond-level temporal resolution, has ushered in a transformative era for VFI. Nonetheless, the application of event-based VFI techniques in domains with distinct environments from the training data can be problematic. This is mainly because event camera data distribution can undergo substantial variations based on camera settings and scene conditions, presenting challenges for effective adaptation. In this paper, we propose a test-time adaptation method for event-based VFI to address the gap between the source and target domains. Our approach enables sequential learning in an online manner on the target domain, which only provides low-frame-rate videos. We present an approach that leverages confident pixels as pseudo ground-truths, enabling stable and accurate online learning from low-frame-rate videos. Furthermore, to prevent overfitting during the continuous online process where the same scene is encountered repeatedly, we propose a method of blending historical samples with current scenes. Extensive experiments validate the effectiveness of our method, both in cross-domain and continuous domain shifting setups. The code is available at https://github.com/Chohoonhee/TTA-EVF." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Video Frame Interpolation (VFI) is a well-established problem in the field of computer vision, aiming to enhance the temporal resolution of videos. In recent times, deep learning-based VFI approaches [2, 3, 15, 20-22, 33, 35, 37, 38, 50, 60] have achieved remarkable performances across various benchmark datasets. However, accurate motion estimation becomes challenging in scenes with complex and non-linear motions due to the lack of information, often resulting in the generation of erroneous inter frames." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 230, + 536, + 435 + ], + "blocks": [ + { + "bbox": [ + 316, + 230, + 536, + 435 + ], + "lines": [ + { + "bbox": [ + 316, + 230, + 536, + 435 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 536, + 435 + ], + "type": "image", + "image_path": "3d32e0b745fa88d50c391b586107d24461690ab16bdd21c71a45d404dc85ceb5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 437, + 547, + 492 + ], + "lines": [ + { + "bbox": [ + 305, + 437, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 547, + 492 + ], + "type": "text", + "content": "Figure 1. TTA-EVF efficiently produces high-quality results by adapting the network to the target domain in an online manner, alleviating the need for offline data supply and addressing the performance degradation observed when applying a well-trained event-based VFI network to a domain with a different distribution." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 506, + 547, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 614 + ], + "type": "text", + "content": "To address the inherent lack of intermediate information in frame cameras, recent researches [16, 24, 51, 52, 56] have explored the potential of event cameras as a promising solution for VFI, especially in scenarios involving complex motion. Event cameras sense dynamic changes in pixel intensity, triggering an event when the change surpasses a predefined threshold. The incorporation of event cameras as additional devices in VFI mitigates the challenges associated with modeling complex motion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 548, + 713 + ], + "type": "text", + "content": "Typically, frame-based VFI methods [17, 33, 38] are evaluated on test datasets (e.g., SNU-FILM [12], X4K1000FPS [43]) with distributions different from their training dataset (e.g., Vimeo90K [61]). Frame-based VFI demonstrates a certain level of generalization ability even without additional modules, providing evidence that it can be applied to real scenario applications, such as dealing with changes in camera devices. These device and environment" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 1, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 1, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 1, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25701" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": "variations are inherent in VFI, which must learn from High-Frame-Rate (HFR) cameras and perform inference on Low-Frame-Rate (LFR) cameras, making them unavoidable." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 110, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 110, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 110, + 288, + 361 + ], + "type": "text", + "content": "However, when training event-based VFI using a source dataset and testing on a different domain, there is a significant performance drop compared to frame-based approaches. This is attributed not only to the inherent noise from external environment (e.g., illumination) but also to the distribution variations due to different camera settings (e.g., event trigger threshold). To address this, one can utilize self-supervised learning [16, 39] with low-frame-rate and event data from the target domain, but there is a drawback in applying them online. In practical application, event cameras exhibit continuous changes in distribution with variations in lighting conditions. It is not feasible to acquire new data for novel scenes each time they arise and wait for the network to learn offline. In this regard, an online adaptation of the network to the environment is essential from a practical perspective and is well-known as Test-Time Adaptation (TTA) [47, 53, 63]. TTA assumes a setting that does not allow access to source data and prefers online learning for adaptation. In our case, the target domain lacks ground-truth data, which means no high-frame-rate video is available in the target domain." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 364, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 364, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 364, + 287, + 639 + ], + "type": "text", + "content": "In this paper, we propose Test-time Adaptation for Event-based Video Frame Interpolation (TTA-EVF) to alleviate performance degradation in the face of domain shifts. As shown in Fig. 1, when encountering a new domain without performing adaptation, even well-trained networks exhibit performance drops. In contrast, TTA-EVF reliably and rapidly fine-tunes in the target domain through self-training, even without high-frame-rate videos. For TTA in event-based VFI, there are two main challenges: Firstly, even without exposure to HFR videos in the target domain, the network should adapt to the scene while simultaneously learning knowledge for HFR generation from LFR data. To address this challenge, we propose Reliable Pixel Sampling (RPS), a self-training scheme for selecting pixels from generated HFR videos, utilizing a confidence estimation based on a teacher-student framework [49]. The second challenge is overfitting when continuously provided with data that has a similar distribution within a successive sequence within online learning. To mitigate this, we employ a memory bank that stores only reliable samples from the generated images. We then propose Patch-Mixed Sampling (PMS) by blending reliable samples with current images, effectively preventing overfitting through an augmentation approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 714 + ], + "type": "text", + "content": "The main contributions of our works are as follows: (I) We propose a novel Test-Time Adaptation (TTA) framework for event-based VFI. (II) We introduce a novel EventRGB Distribution Shift (ERDS) dataset for event-based video frame interpolation, where the distribution of the dataset is continually changing. (III) We compare our" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "method with existing works and demonstrate its superiority by outperforming supervised methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 107, + 398, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 107, + 398, + 119 + ], + "spans": [ + { + "bbox": [ + 306, + 107, + 398, + 119 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 127, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 127, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 127, + 545, + 378 + ], + "type": "text", + "content": "Event-based Video Frame Interpolation. Frame-based VFI methods [2, 3, 15, 20-22, 33, 35, 37, 38, 50, 60] have continuously achieved performance improvements; however, it still has weaknesses in handling complex motion and requires motion approximation when synthesizing intermediate frames at arbitrary time. Event-based VFI methods [30, 41, 44, 45, 55, 59, 64, 65], which utilize event cameras as an additional sensor [6, 7, 9, 28, 34, 45], enable not only handling complex motion but also interpolating intermediate frames for irregular objects. However, there is still room for improvement compared to frame-based VFI methods. Typically, frame-based VFI methods perform cross-domain performance evaluations, considering real scenario applications. Well-trained frame-based VFI methods exhibit domain generalization ability, maintaining considerable performance even on different domains without the need for additional training process for adaptation. On the other hand, event-based VFI experiences a significant drop in performance when moving to different domains [16]. Therefore, resolving cross-domain challenges in event-based VFI is vital for widespread applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 379, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 545, + 653 + ], + "type": "text", + "content": "Test-Time Adaptation. Unlike Unsupervised Domain Adaptation [8, 48, 58], which necessitates offline learning for each domain change and relies on labeled data from the source domain, Test-Time Adaptation (TTA) is more challenging setup as it adapts to the target domain without accessing source data. In recent times, the significance of data privacy has escalated, leading to increased interest in TTA, a source-free domain adaptation approach. This is primarily due to the inefficiency of accessing source domain data during inference. TTA is actively being researched in the image domain [5, 14, 19, 27, 31, 32, 36, 40, 42, 47, 53, 63, 66], especially in areas like object recognition and segmentation. While there is an approach of its application in event-based visual recognition, Ev-TTA [23], there is still a lack of research in the domain of video frame interpolation, particularly in relation to events. The existing work, MetaVFI [11], proposed for the frame-based VFI task, learns to estimate scene-adaptive motion with minimal parameter updates through meta-learning [13]. However, this method is well-suited when there is not a significant domain gap between training and testing. We demonstrate that this approach is suboptimal when applied to event-based VFI with a substantial domain gap between the training and test sets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "We propose the first-ever Test-Time Adaptation (TTA) approach for event-based Video Frame Interpolation (VFI). This is essential because event distributions can vary significantly due to external environmental factors and camera parameter settings even when using the same camera." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25702" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 67, + 533, + 270 + ], + "blocks": [ + { + "bbox": [ + 58, + 67, + 533, + 270 + ], + "lines": [ + { + "bbox": [ + 58, + 67, + 533, + 270 + ], + "spans": [ + { + "bbox": [ + 58, + 67, + 533, + 270 + ], + "type": "image", + "image_path": "1910b9000bc1557cb3ec1f47454fba3dd0162dbe3dc4b0155ad053fc59d0192e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "text", + "content": "Figure 2. Overview of our Test-Time Adaptation for Event-based Video Frame Interpolation (TTA-EVF) framework. Our approach involves four steps for test-time training in Ev-VFI: 1) We calculate the adaptive threshold, " + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "text", + "content": ", for RPS using the low frame rate image triples, " + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "inline_equation", + "content": "I_0, I_1" + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "inline_equation", + "content": "I_2" + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "text", + "content": " with interval events " + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "inline_equation", + "content": "E_{0\\rightarrow 1}" + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "inline_equation", + "content": "E_{2\\rightarrow 1}" + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "text", + "content": ". 2) Reliable Pixel Sampling (RPS) is applied for arbitrary " + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "text", + "content": " between 0 and 1, and arbitrary " + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "text", + "content": " between 1 and 2, and in this process, even without intermediate images for training, reliable pixels are estimated through the student-teacher framework and self-training is performed. 3) Using the estimated intermediate frames " + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "inline_equation", + "content": "\\hat{I}_{t_0}" + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "inline_equation", + "content": "\\hat{I}_{t_1}" + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "text", + "content": " along with " + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "text", + "content": ", we form a new triplet, " + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "text", + "content": ", check if the entire sampled generated image is reliable, and in that case, update the memory bank. 4) Patch-Mixed Sampling (PMS) regularizes the network by mixing the generated videos, " + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "text", + "content": " with videos from the memory bank, " + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{V}}" + }, + { + "bbox": [ + 46, + 274, + 548, + 353 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 363, + 288, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 288, + 399 + ], + "type": "text", + "content": "Furthermore, our method follows the recent trend of TTA by addressing the adaptation to unknown distribution shifts encountered during test-time in an online manner." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 410, + 153, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 410, + 153, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 410, + 153, + 423 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 430, + 115, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 115, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 115, + 441 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 448, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 519 + ], + "type": "text", + "content": "Given key frames " + }, + { + "bbox": [ + 46, + 448, + 287, + 519 + ], + "type": "inline_equation", + "content": "I_0, I_1" + }, + { + "bbox": [ + 46, + 448, + 287, + 519 + ], + "type": "text", + "content": " and the inter-frame event stream " + }, + { + "bbox": [ + 46, + 448, + 287, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{0\\rightarrow 1}" + }, + { + "bbox": [ + 46, + 448, + 287, + 519 + ], + "type": "text", + "content": ", Event-based Video Frame Interpolation (Ev-VFI) networks aim to generate the intermediate frame " + }, + { + "bbox": [ + 46, + 448, + 287, + 519 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 448, + 287, + 519 + ], + "type": "text", + "content": " on an arbitrary time " + }, + { + "bbox": [ + 46, + 448, + 287, + 519 + ], + "type": "inline_equation", + "content": "t \\in [0,1]" + }, + { + "bbox": [ + 46, + 448, + 287, + 519 + ], + "type": "text", + "content": ". Following the previous works, we convert a event stream as a voxel grid [67]. We denote " + }, + { + "bbox": [ + 46, + 448, + 287, + 519 + ], + "type": "inline_equation", + "content": "E_{t_a \\to t_b}" + }, + { + "bbox": [ + 46, + 448, + 287, + 519 + ], + "type": "text", + "content": " as a voxel grid form between times " + }, + { + "bbox": [ + 46, + 448, + 287, + 519 + ], + "type": "inline_equation", + "content": "t_a" + }, + { + "bbox": [ + 46, + 448, + 287, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 448, + 287, + 519 + ], + "type": "inline_equation", + "content": "t_b" + }, + { + "bbox": [ + 46, + 448, + 287, + 519 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 520, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 288, + 604 + ], + "type": "text", + "content": "TTA-EVF adapts a pre-trained Ev-VFI network trained on the source domain to a target domain with domain shift in the device and setting. In the pre-training phase, we train the network using source data with high-frame rate videos in a supervised manner (Sec. 3.2). We then utilize this pretrained Ev-VFI network for test-time adaptation using target data without accessing the source data (Sec. 3.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 611, + 216, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 216, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 216, + 624 + ], + "type": "text", + "content": "3.2. Pre-training on Source Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "In general, TTA methods perform test-time training stably by updating only a subset of network parameters, such as Batch Normalization (BN) [18]. As mentioned in [53], updating the full parameters at test-time, especially in an online manner, leads to instability and makes it challenging to achieve optimal performance. However, in the case of low-level vision tasks like VFI, normalization layer (e.g. BN)" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 323, + 361, + 526, + 449 + ], + "blocks": [ + { + "bbox": [ + 323, + 361, + 526, + 449 + ], + "lines": [ + { + "bbox": [ + 323, + 361, + 526, + 449 + ], + "spans": [ + { + "bbox": [ + 323, + 361, + 526, + 449 + ], + "type": "image", + "image_path": "1059df73b9d437f508be764524a1da6673eeb85b7f3135d0264724301565e7b4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 453, + 531, + 464 + ], + "lines": [ + { + "bbox": [ + 320, + 453, + 531, + 464 + ], + "spans": [ + { + "bbox": [ + 320, + 453, + 531, + 464 + ], + "type": "text", + "content": "Figure 3. The architecture of Norm-Residual (NR) Block." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 473, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 676 + ], + "type": "text", + "content": "is often avoided by most networks [20, 24, 51] since it removes range flexibility and can lead to a performance decrease [29]. Some alternatives for handling this issue include using slightly more flexible normalization techniques (e.g. instance normalization) or selectively updating specific convolutional layers. Updating only the parameters of the normalization layer in VFI can lead to sub-optimal performance when there is significant variance between image patches in the training and test environments. Conversely, updating specific convolution layers with more parameters can remove all statistical modulation, causing instability during the test-time training process. To address this issue, we propose a solution by designing a Norm-Residual (NR) Block that includes both convolution and normalization layers. This approach aims to strike a balance between updating parameters effectively and maintaining statistical modulation during the learning process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "As shown in Fig. 3, given a feature " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": ", NR Block first split a feature into two feature maps " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(1)}" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(2)}" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " along the channel dimension. The first map " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(1)}" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " is skipped to keep" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25703" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "the context information, and rest feature " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{(2)}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " is normalized by normalization layer. Then, we fuse the two features by concatenate and convolutional layer. Finally, we apply a element-wise summation with original feature " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " to obtain the output " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ". We use the group normalization [57] in normalization layer, which shows the best performance in our experiments. For efficiency, we exclusively deploy the NR Block in the encoders of the networks, replacing the conventional residual blocks of existing networks. Therefore, TTA-EVF is a general and model-agnostic method specifically crafted to be adaptable to any network architecture used in Ev-VFI. Further details can be found in the supple." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 288, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 288, + 265 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 288, + 265 + ], + "type": "text", + "content": "Network Optimization. The modified Ev-VFI network with the NR Block is trained on the source dataset using ground truth with high-frame-rate videos. Specifically, we optimize the Ev-VFI network, " + }, + { + "bbox": [ + 46, + 216, + 288, + 265 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 216, + 288, + 265 + ], + "type": "text", + "content": ", by the supervision as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 123, + 274, + 287, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 274, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 123, + 274, + 287, + 297 + ], + "type": "interline_equation", + "content": "\\underset {\\theta (\\mathcal {F})} {\\arg \\min } \\left(\\mathcal {L} _ {\\mathcal {F}} \\left(\\hat {I} _ {t}, I _ {t}\\right)\\right), \\tag {1}", + "image_path": "d6f643ac66f324a0179c4e191bceba591ef928847eba75b662e0f95665d29aa7.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 304, + 288, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 288, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 304, + 288, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{F}}" + }, + { + "bbox": [ + 47, + 304, + 288, + 342 + ], + "type": "text", + "content": " is the loss function proposed by each Ev-VFI network and the generated intermediate frame " + }, + { + "bbox": [ + 47, + 304, + 288, + 342 + ], + "type": "inline_equation", + "content": "\\hat{I}_t" + }, + { + "bbox": [ + 47, + 304, + 288, + 342 + ], + "type": "text", + "content": " is obtained by " + }, + { + "bbox": [ + 47, + 304, + 288, + 342 + ], + "type": "inline_equation", + "content": "\\hat{I}_t = \\mathcal{F}(I_0, I_1, E_{0 \\to t}, E_{1 \\to t})" + }, + { + "bbox": [ + 47, + 304, + 288, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 350, + 255, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 255, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 255, + 363 + ], + "type": "text", + "content": "3.3. Test-time Adaptation on Target Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 368, + 287, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 368, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 287, + 429 + ], + "type": "text", + "content": "In our TTA setup, the model generates intermediate frames at each iteration and simultaneously updates the parameters of the NR Block. Since the target domain only contains LFR frames, not HFR frames, self-supervised optimization using only LFR is required." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 445, + 211, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 211, + 457 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 211, + 457 + ], + "type": "text", + "content": "3.3.1 Reliable Pixel Sampling (RPS)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 464, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 525 + ], + "type": "text", + "content": "In the target domain, even without High-Frame-Rate (HFR) ground truth, when given Low-Frame-Rate (LFR) videos " + }, + { + "bbox": [ + 46, + 464, + 287, + 525 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 46, + 464, + 287, + 525 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 464, + 287, + 525 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 46, + 464, + 287, + 525 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 464, + 287, + 525 + ], + "type": "inline_equation", + "content": "I_2" + }, + { + "bbox": [ + 46, + 464, + 287, + 525 + ], + "type": "text", + "content": " as triplets, and the events between them " + }, + { + "bbox": [ + 46, + 464, + 287, + 525 + ], + "type": "inline_equation", + "content": "E_{0\\to 1}" + }, + { + "bbox": [ + 46, + 464, + 287, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 464, + 287, + 525 + ], + "type": "inline_equation", + "content": "E_{2\\to 1}" + }, + { + "bbox": [ + 46, + 464, + 287, + 525 + ], + "type": "text", + "content": ", the network can be trained using the following self-supervised loss:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 129, + 533, + 287, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 533, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 129, + 533, + 287, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S} = \\mathcal {L} _ {\\mathcal {F}} (\\hat {I} _ {1}, I _ {1}), \\tag {2}", + "image_path": "0afeae231435ffe41ad56ff1b08d4a10578a55fb96c31c3a858cacea9cbd7de7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 556, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 556, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\hat{I}_1 = \\mathcal{F}(I_0, I_2, E_{0\\rightarrow 1}, E_{2\\rightarrow 1})" + }, + { + "bbox": [ + 46, + 556, + 287, + 689 + ], + "type": "text", + "content": ". However, this is problematic as the network cannot learn the knowledge of generating HFR because it is generating LFR again from temporally subsampled LFR images, leading to the loss of the ability to generate HFR trained from the source domain. As a way to address this, applying the cyclic consistency loss [16, 39] from the existing unsupervised VFI is one option. However, directly applying these methods to TTA of Ev-VFI is intractable, since there is no guarantee that the generated intermediate frames will be clean and artifact-free on target domains. Therefore, we need a new paradigm." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "As shown in Fig. 2, given the " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "I_2" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": " triplet, we design Reliable Pixel Sampling (RPS) for the left intermediate time" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": " and the right intermediate time " + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": " are selected at random from the time intervals 0 to 1 and 1 to 2, respectively. Since the RPS process is identical for both " + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": ", we introduce it here for " + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": " only. Reliable data sampling without the intermediate ground truth frames, " + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "inline_equation", + "content": "I_{t_0}" + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": ", is highly challenging and can lead to unstable learning if estimated incorrectly. To address this, we adopt the teacher-student framework [49] commonly used for pseudo labeling [10, 58] in other tasks. In this framework, the student learns from pseudo labels through self-training, while the teacher is updated solely from the momentum using the parameters of the student. The motivation behind proposed RPS stems from the observation that pixels estimated similarly by two networks with slightly different parameters, yet achieving similar performance in VFI, are likely to be reliable, whereas parts that are difficult to estimate and prone to errors would result in differing predictions between the two networks. Hence, we classify pixels consistently estimated by both teacher-student networks as reliable pixels, while considering pixels with inconsistent predictions as unreliable. Given intermediate frames generated from the student network " + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{S}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": " and the teacher network " + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{T}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": " denoted as " + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "inline_equation", + "content": "\\hat{I}_{t_0}^{\\mathcal{S}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "inline_equation", + "content": "\\hat{I}_{t_0}^{\\mathcal{T}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": " respectively, the reliable map, " + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "inline_equation", + "content": "R_{t_0}" + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": ", with spatial dimensions " + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "inline_equation", + "content": "H\\times W\\times 1" + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": " represent the height and width of the image " + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 304, + 72, + 547, + 373 + ], + "type": "text", + "content": ") is defined as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 389, + 545, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 389, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 318, + 389, + 545, + 422 + ], + "type": "interline_equation", + "content": "R _ {t _ {0}} (i, j) = \\left\\{ \\begin{array}{l} 1, \\text {i f} \\| \\hat {I} _ {t _ {0}} ^ {\\mathcal {S}} (i, j) - \\hat {I} _ {t _ {0}} ^ {\\mathcal {T}} (i, j) \\| < \\rho , \\\\ 0, \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}", + "image_path": "f57d2cc0cd746dd51e1411c0694f5a71ffe256727ae1dafcdea4c39c5d1344fd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "inline_equation", + "content": "i = 1,\\dots ,H" + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "inline_equation", + "content": "j = 1,\\ldots ,W" + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "text", + "content": " denote the threshold. When estimating such reliability, it is crucial to determine a specific threshold. However, assigning the same threshold to all scenes is inappropriate since the performance of VFI significantly varies across scenes. To achieve adaptive thresholding for each scene, we utilize the LFR frames " + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "inline_equation", + "content": "I_{0}" + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "text", + "content": ", along with the interceding events " + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "inline_equation", + "content": "E_{0\\rightarrow 1}" + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "inline_equation", + "content": "E_{2\\rightarrow 1}" + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "text", + "content": ", to estimate " + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "inline_equation", + "content": "\\hat{I}_1" + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "text", + "content": ". Subsequently, we compute a pixel-wise error map, " + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "inline_equation", + "content": "e = \\| \\hat{I}_1 - I_1\\|" + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "text", + "content": ", between " + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "inline_equation", + "content": "\\hat{I}_1" + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "text", + "content": ", allowing us to calculate " + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "text", + "content": ", using the average value of " + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 437, + 547, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "However, this approach encounters issues when dealing with scenes that contain motionless, static regions (e.g., backgrounds). In such cases, these areas exhibit minimal pixel-wise errors, which leads to an inadequate adaptive threshold calculation for the scene if these pixels are included in the average calculation. To address this, we propose an approach that excludes motionless regions based on optical flow to refine the threshold calculation. Since most Ev-VFI methods compute optical flow towards the intermediate frame from key frames, we can directly calculate a motion magnitude map [54], " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": ", from the intermediate output of optical flow, denoted as " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "F_{10}" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "F_{12}" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": ", without requiring" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25704" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 179, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 179, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 179, + 83 + ], + "type": "text", + "content": "Algorithm 1 PMS Pseudo Code" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 89, + 286, + 102 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 89, + 286, + 102 + ], + "spans": [ + { + "bbox": [ + 49, + 89, + 286, + 102 + ], + "type": "text", + "content": "Input: Generated HFR set " + }, + { + "bbox": [ + 49, + 89, + 286, + 102 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{\\hat{I}_{t_0}, I_1, \\hat{I}_{t_1}, E_{t_0 \\to 1}, E_{t_1 \\to 1}\\}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 102, + 286, + 209 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 53, + 102, + 141, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 102, + 141, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 102, + 141, + 113 + ], + "type": "text", + "content": "1: Initialize " + }, + { + "bbox": [ + 53, + 102, + 141, + 113 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_M\\gets 0" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 114, + 208, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 208, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 208, + 125 + ], + "type": "text", + "content": "2: Get reliability of " + }, + { + "bbox": [ + 53, + 114, + 208, + 125 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 53, + 114, + 208, + 125 + ], + "type": "text", + "content": " based on Eq.(9)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 126, + 258, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 258, + 138 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 258, + 138 + ], + "type": "text", + "content": "3: Sample " + }, + { + "bbox": [ + 53, + 126, + 258, + 138 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{\\bar{I}_{\\tilde{t}_0},\\bar{I}_1,\\bar{I}_{\\tilde{t}_1},\\bar{E}_{\\tilde{t}_0\\to 1},E_{\\tilde{t}_1\\to 1}\\}" + }, + { + "bbox": [ + 53, + 126, + 258, + 138 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 53, + 126, + 258, + 138 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 139, + 206, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 139, + 206, + 149 + ], + "spans": [ + { + "bbox": [ + 53, + 139, + 206, + 149 + ], + "type": "text", + "content": "4: Sample mixing ratio " + }, + { + "bbox": [ + 53, + 139, + 206, + 149 + ], + "type": "inline_equation", + "content": "\\lambda \\in (0.7, 1.0)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 150, + 145, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 145, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 145, + 159 + ], + "type": "text", + "content": "5: if " + }, + { + "bbox": [ + 53, + 150, + 145, + 159 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 53, + 150, + 145, + 159 + ], + "type": "text", + "content": " is reliable then" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 160, + 286, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 286, + 174 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 286, + 174 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 53, + 160, + 286, + 174 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t_0} = \\lambda \\hat{I}_{t_0} + (1 - \\lambda)\\tilde{I}_1, \\mathcal{M}_{t_1} = \\lambda \\hat{I}_{t_1} + (1 - \\lambda)\\tilde{I}_1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 174, + 173, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 174, + 173, + 186 + ], + "spans": [ + { + "bbox": [ + 53, + 174, + 173, + 186 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 53, + 174, + 173, + 186 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_1 = \\lambda I_1 + (1 - \\lambda)\\tilde{I}_1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 186, + 201, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 201, + 198 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 201, + 198 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 53, + 186, + 201, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{t_0\\to 1} = E_{t_0\\to 1},\\mathcal{D}_{t_1\\to 1} = E_{t_1\\to 1}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 198, + 192, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 192, + 209 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 192, + 209 + ], + "type": "text", + "content": "9: Push " + }, + { + "bbox": [ + 53, + 198, + 192, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 53, + 198, + 192, + 209 + ], + "type": "text", + "content": " into memory bank " + }, + { + "bbox": [ + 53, + 198, + 192, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 50, + 210, + 83, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 210, + 83, + 219 + ], + "spans": [ + { + "bbox": [ + 50, + 210, + 83, + 219 + ], + "type": "text", + "content": "10: else" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 50, + 220, + 285, + 301 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 50, + 220, + 285, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 220, + 285, + 235 + ], + "spans": [ + { + "bbox": [ + 50, + 220, + 285, + 235 + ], + "type": "text", + "content": "11: " + }, + { + "bbox": [ + 50, + 220, + 285, + 235 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t_0} = (1 - \\lambda)I_1 + \\lambda \\tilde{I}_{\\tilde{t}_0}" + }, + { + "bbox": [ + 50, + 220, + 285, + 235 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 50, + 220, + 285, + 235 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t_1} = (1 - \\lambda)I_1 + \\lambda \\tilde{I}_{\\tilde{t}_1}" + }, + { + "bbox": [ + 50, + 220, + 285, + 235 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 50, + 236, + 172, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 236, + 172, + 247 + ], + "spans": [ + { + "bbox": [ + 50, + 236, + 172, + 247 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 50, + 236, + 172, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_1 = (1 - \\lambda)I_1 + \\lambda \\tilde{I}_1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 50, + 247, + 201, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 247, + 201, + 261 + ], + "spans": [ + { + "bbox": [ + 50, + 247, + 201, + 261 + ], + "type": "text", + "content": "13: " + }, + { + "bbox": [ + 50, + 247, + 201, + 261 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{t_0\\to 1} = \\bar{E}_{\\tilde{t}_0\\to 1},\\mathcal{D}_{t_1\\to 1} = \\bar{E}_{\\tilde{t}_{1}\\to 1}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 50, + 262, + 198, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 262, + 198, + 274 + ], + "spans": [ + { + "bbox": [ + 50, + 262, + 198, + 274 + ], + "type": "text", + "content": "14: if " + }, + { + "bbox": [ + 50, + 262, + 198, + 274 + ], + "type": "inline_equation", + "content": "\\mathrm{len}(\\mathcal{Q}) > \\mathrm{max}" + }, + { + "bbox": [ + 50, + 262, + 198, + 274 + ], + "type": "text", + "content": " queue size then" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 50, + 275, + 176, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 275, + 176, + 286 + ], + "spans": [ + { + "bbox": [ + 50, + 275, + 176, + 286 + ], + "type": "text", + "content": "15: Pop oldest ones out of " + }, + { + "bbox": [ + 50, + 275, + 176, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 50, + 289, + 261, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 289, + 261, + 301 + ], + "spans": [ + { + "bbox": [ + 50, + 289, + 261, + 301 + ], + "type": "text", + "content": "16: " + }, + { + "bbox": [ + 50, + 289, + 261, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_M\\gets \\mathcal{L}_{\\mathcal{F}}(\\mathcal{M}_1,\\mathcal{F}_S(\\mathcal{M}_{t_0},\\mathcal{M}_{t_1},\\mathcal{D}_{t_0\\to 1},\\mathcal{D}_{t_1\\to 1}))" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 48, + 302, + 105, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 302, + 105, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 302, + 105, + 312 + ], + "type": "text", + "content": "Output: " + }, + { + "bbox": [ + 48, + 302, + 105, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_M" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 319, + 175, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 175, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 175, + 331 + ], + "type": "text", + "content": "ing additional computations as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 82, + 338, + 287, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 338, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 82, + 338, + 287, + 352 + ], + "type": "interline_equation", + "content": "M (i, j) = \\left(\\| F _ {1 0} (i, j) \\| + \\| F _ {1 2} (i, j) \\|\\right) / 2, \\tag {4}", + "image_path": "1e65676c29c909baea6e18635f98e3993c7ea767f36fd535d677a26d3e6572d3.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 357, + 287, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 287, + 405 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 357, + 287, + 405 + ], + "type": "inline_equation", + "content": "\\|\\cdot\\|" + }, + { + "bbox": [ + 47, + 357, + 287, + 405 + ], + "type": "text", + "content": " denote the magnitude of the motion vector. The motion mask " + }, + { + "bbox": [ + 47, + 357, + 287, + 405 + ], + "type": "inline_equation", + "content": "\\widehat{M}" + }, + { + "bbox": [ + 47, + 357, + 287, + 405 + ], + "type": "text", + "content": ", which excludes motionless areas based on the motion magnitude map for each pixel, can be created as follows:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 82, + 411, + 287, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 411, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 82, + 411, + 287, + 442 + ], + "type": "interline_equation", + "content": "\\widehat {M} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} M (i, j) \\geq m, \\forall i, \\forall j \\\\ 0 & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {5}", + "image_path": "5e8f3708c4632c9fb75ce3a1d3d83491f4c832c07ae004522e7e0185c84ad7fc.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 47, + 448, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 448, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 448, + 287, + 483 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 47, + 448, + 287, + 483 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 448, + 287, + 483 + ], + "type": "text", + "content": " represents a predefined minimal amount of motion, and in our experiment, we use a value of 1. Then, we calculate the scene adaptive threshold as:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 123, + 489, + 287, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 489, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 123, + 489, + 287, + 516 + ], + "type": "interline_equation", + "content": "\\rho = \\frac {1}{| \\widehat {M} |} \\sum \\widehat {M} \\odot e. \\tag {6}", + "image_path": "88ef31281e84c8b8cafc9ae43a3d013f671397815f72e8b8a86763a12ea82519.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 47, + 521, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 287, + 557 + ], + "type": "text", + "content": "We define the self-training loss between the teacher and student networks using a pixel-wise reliable map in Eq.(3) and Eq.(6):" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 93, + 562, + 287, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 562, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 93, + 562, + 287, + 578 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R} ^ {t _ {0}} = \\left\\| \\hat {I} _ {t _ {0}} ^ {\\mathcal {S}} (i, j) - \\hat {I} _ {t _ {0}} ^ {\\mathcal {T}} (i, j) \\right\\| \\odot R _ {t _ {0}}, \\tag {7}", + "image_path": "03d1e9becb8352030e15f01d6ad91cff6def7e98a16f22189b91faad50b56bf1.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 47, + 582, + 287, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 606 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 606 + ], + "type": "text", + "content": "Finally, the reliable pixel-wise loss, taking into account both intermediate times " + }, + { + "bbox": [ + 47, + 582, + 287, + 606 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 47, + 582, + 287, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 582, + 287, + 606 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 47, + 582, + 287, + 606 + ], + "type": "text", + "content": ", is defined as follows:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 130, + 613, + 287, + 627 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 287, + 627 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R} = \\mathcal {L} _ {R} ^ {t _ {0}} + \\mathcal {L} _ {R} ^ {t _ {1}}, \\tag {8}", + "image_path": "f9a67c171477fea8b953f4b1a479d96c9438449dd74a44ff5f976bd50de8360d.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 47, + 647, + 208, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 208, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 208, + 658 + ], + "type": "text", + "content": "3.3.2 Patch-Mixed Sampling (PMS)" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "The issues that can arise in TTA include overfitting to specific scenes due to the sequential data for the same scene being continuously fed, as well as the problem of catastrophic forgetting [4, 26] where existing knowledge is lost." + } + ] + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 315, + 72, + 426, + 201 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 426, + 201 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 426, + 201 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 426, + 201 + ], + "type": "image", + "image_path": "967d2dd49df540049f3700bd81fe0dcdef83d7bfb79f55ddc063184ae4673f38.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 204, + 544, + 215 + ], + "lines": [ + { + "bbox": [ + 306, + 204, + 544, + 215 + ], + "spans": [ + { + "bbox": [ + 306, + 204, + 544, + 215 + ], + "type": "text", + "content": "Figure 4. The examples of patch-mixed sampling (PMS) process." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 535, + 201 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 535, + 201 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 535, + 201 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 535, + 201 + ], + "type": "image", + "image_path": "b7df87b8dadc8a0ab456d0427afcce87026021b561a2b727ddaf4b089776bd16.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "bbox": [ + 305, + 219, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 219, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 305, + 219, + 545, + 290 + ], + "type": "text", + "content": "To address this, we propose the Patch-Mixed Sampling (PMS) method. In order to avoid overfitting to specific scenes, PMS involves sampling from the memory bank, which accumulates reliable samples from the generated videos of each iteration. The detailed process of PMS is provided in Algorithm 1." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 305, + 290, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 290, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 290, + 545, + 361 + ], + "type": "text", + "content": "Reliability Check. Denoting the HFR video and event set, including " + }, + { + "bbox": [ + 305, + 290, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\hat{I}_{t_0}" + }, + { + "bbox": [ + 305, + 290, + 545, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 290, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\hat{I}_{t_1}" + }, + { + "bbox": [ + 305, + 290, + 545, + 361 + ], + "type": "text", + "content": ", from the RPS process as " + }, + { + "bbox": [ + 305, + 290, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 305, + 290, + 545, + 361 + ], + "type": "text", + "content": ", we assess the reliability for the generated frames. If the pixel-wise reliability map " + }, + { + "bbox": [ + 305, + 290, + 545, + 361 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 290, + 545, + 361 + ], + "type": "text", + "content": " obtained from the RPS module exceeds " + }, + { + "bbox": [ + 305, + 290, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 305, + 290, + 545, + 361 + ], + "type": "text", + "content": " fraction of the resolution of the entire image, we deem the generated images as reliable. This is outlined as follows:" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 312, + 368, + 545, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 368, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 312, + 368, + 545, + 400 + ], + "type": "interline_equation", + "content": "\\text {R e l i a b l i t y} = \\left\\{ \\begin{array}{l} \\text {T r u e , i f} \\frac {\\| R _ {t _ {0}} \\|}{H \\times W} > \\eta \\text {a n d} \\frac {\\| R _ {t _ {1}} \\|}{H \\times W} > \\eta \\\\ \\text {F a l s e , o t h e r w i s e ,} \\end{array} \\right. \\tag {9}", + "image_path": "9e458ead02d71738a9623a24ac5a0e70d9b18fd2f83da37b87a33a2af2f4a8d1.jpg" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "spans": [ + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "content": "Patch Mixing. Figure 4 shows an example of a patch mixing process. If the currently generated HFR video set " + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "content": " is reliable (left of Fig. 4), we fetch a single intermediate frame " + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "inline_equation", + "content": "\\tilde{I}_1" + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "content": " from the memory bank and blend it. Since the mixed frames, " + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "inline_equation", + "content": "M_{t_0}, M_1" + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "inline_equation", + "content": "M_{t_1}" + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "content": ", all incorporate the same " + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "inline_equation", + "content": "\\tilde{I}_1" + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "content": ", any events that require pixel changes can't arise from " + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "inline_equation", + "content": "\\tilde{I}_1" + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "content": ", thus we directly retain the events " + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "inline_equation", + "content": "E_{t_0 \\to 1}" + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "inline_equation", + "content": "E_{t_1 \\to 1}" + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "content": ". On the other hand, if " + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "content": " is unreliable (right of Fig. 4), we retrieve both images and event sets from the memory bank. From the currently generated " + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "content": ", we only fetch " + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "content": " and blend it. Similar to before, since the same " + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "content": " is blended into all three images, events originating from " + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "content": " do not occur. Therefore, we directly retain the memory bank events " + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "inline_equation", + "content": "\\tilde{E}_{\\tilde{t}_0 \\to 1}" + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "inline_equation", + "content": "\\tilde{E}_{\\tilde{t}_1 \\to 1}" + }, + { + "bbox": [ + 305, + 406, + 546, + 634 + ], + "type": "text", + "content": ". For the mixing ratio of blending images, we adopt a beta distribution [62] that generates values between 0.7 and 1.0. We assign a larger ratio to the reliable set among the two sets being blended. Finally, using the patch-mixed sets, we train the student network of the RPS using the loss of Ev-VFI as follows:" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 318, + 641, + 545, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 641, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 318, + 641, + 545, + 654 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M} = \\mathcal {L} _ {\\mathcal {F}} \\left(\\mathcal {M} _ {1}, \\mathcal {F} _ {\\mathcal {S}} \\left(\\mathcal {M} _ {t _ {0}}, \\mathcal {M} _ {t _ {1}}, \\mathcal {D} _ {t _ {0} \\rightarrow 1}, \\mathcal {D} _ {t _ {1} \\rightarrow 1}\\right)\\right). \\tag {10}", + "image_path": "a3dd96e517cc5992c989b54924633f9bd09bd039e20b9605ad447f99173a859a.jpg" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 306, + 661, + 544, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 661, + 544, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 661, + 544, + 673 + ], + "type": "text", + "content": "In summary, the total objective " + }, + { + "bbox": [ + 306, + 661, + 544, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 306, + 661, + 544, + 673 + ], + "type": "text", + "content": " of TTA-EVF is as follows:" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 370, + 685, + 545, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 685, + 545, + 697 + ], + "spans": [ + { + "bbox": [ + 370, + 685, + 545, + 697 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {S} + \\lambda_ {1} \\mathcal {L} _ {R} + \\lambda_ {2} \\mathcal {L} _ {M} \\tag {11}", + "image_path": "17e2cdc21b18279efd610416fa940132ef9e7c518e8c8ba85a3405e22821c03f.jpg" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 306, + 702, + 464, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 702, + 464, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 702, + 464, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 702, + 464, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2" + }, + { + "bbox": [ + 306, + 702, + 464, + 713 + ], + "type": "text", + "content": " are the hyper-parameters." + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25705" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 534, + 260 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 534, + 260 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 534, + 260 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 534, + 260 + ], + "type": "image", + "image_path": "a1de13773749c1f8bae56c306ed1257e8241fd74af58b5e492a41d076cb5a74e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 261, + 493, + 272 + ], + "lines": [ + { + "bbox": [ + 99, + 261, + 493, + 272 + ], + "spans": [ + { + "bbox": [ + 99, + 261, + 493, + 272 + ], + "type": "text", + "content": "Figure 5. Visual comparisons of interpolated frames on multiple cross-domain settings using various models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 278, + 128, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 278, + 128, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 128, + 293 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 298, + 180, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 298, + 180, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 180, + 312 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 316, + 288, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 460 + ], + "type": "text", + "content": "All networks are trained using an Adam optimizer [25] with a batch size of 2 and a learning rate of " + }, + { + "bbox": [ + 46, + 316, + 288, + 460 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 316, + 288, + 460 + ], + "type": "text", + "content": ", for both the source and target domains. For efficiency, during online training in the target domain, we crop images to a resolution of " + }, + { + "bbox": [ + 46, + 316, + 288, + 460 + ], + "type": "inline_equation", + "content": "320 \\times 320" + }, + { + "bbox": [ + 46, + 316, + 288, + 460 + ], + "type": "text", + "content": ". In our approach, we perform network updates only once for each triplet provided. For instance, when interpolating 7 skipped frames, we generate a total of 14 frames from " + }, + { + "bbox": [ + 46, + 316, + 288, + 460 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 46, + 316, + 288, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 316, + 288, + 460 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 46, + 316, + 288, + 460 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 316, + 288, + 460 + ], + "type": "inline_equation", + "content": "I_2" + }, + { + "bbox": [ + 46, + 316, + 288, + 460 + ], + "type": "text", + "content": ", and update the network parameters only once using the loss from Eq.(11), making it highly efficient. All experiments are conducted on a single NVIDIA Titan RTX. We set the " + }, + { + "bbox": [ + 46, + 316, + 288, + 460 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 316, + 288, + 460 + ], + "type": "text", + "content": " as 0.9 in Eq.(9), bin size of voxel grid as 5, and " + }, + { + "bbox": [ + 46, + 316, + 288, + 460 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2" + }, + { + "bbox": [ + 46, + 316, + 288, + 460 + ], + "type": "text", + "content": " as 1, 0.5 in Eq.(11)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 468, + 209, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 209, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 209, + 480 + ], + "type": "text", + "content": "4.2. Test-Time Adaptation Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "text", + "content": "During the pre-training phase, we utilize the train sequences from the source domain for learning, and in the target domain, we follow the same sequential online approach as in the traditional TTA setup [53], using only the test sequences without utilizing the train set." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Our framework can be applied to existing Ev-VFI methods regardless of the model. We conduct experiments by integrating it with TimeLens [51] and CBMNet [24], using publicly available code provided by the authors. For TimeReplayer [16], as public code is unavailable, we re-implement it based on the foundational code [20] as described in the paper. To the best of our knowledge, we are the first to propose test-time adaptation for event-based video frame interpolation. Therefore, for comparison, we adopt the previously proposed meta-learning approach [11] for frame-based VFI and the self-supervised optimization [16] for Ev-VFI. When training at test time, the performance tends to be low if we simply apply the existing methods [16] as is. Therefore, we keep our NR Block" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 300, + 537, + 472 + ], + "blocks": [ + { + "bbox": [ + 305, + 278, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 305, + 278, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 278, + 545, + 300 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons with existing method in cross-domain datasets. S and T denote the source and target datasets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 300, + 537, + 472 + ], + "lines": [ + { + "bbox": [ + 316, + 300, + 537, + 472 + ], + "spans": [ + { + "bbox": [ + 316, + 300, + 537, + 472 + ], + "type": "table", + "html": "
S TMethodsTimeReplayer [16]TimeLens [51]CBMNet [24]
PSNRSSIMPSNRSSIMPSNRSSIM
B RNo Adapt32.240.915432.040.907331.980.9081
Meta [11]32.750.918032.430.911232.210.9109
Cycle [16]32.870.920833.010.916533.090.9214
TTA-EVF34.070.928634.240.929934.420.9312
B HNo Adapt29.780.837829.960.841329.870.8395
Meta [11]30.130.839330.140.834130.220.8407
Cycle [16]30.320.810430.660.844130.750.8466
TTA-EVF31.450.855132.130.859932.070.8584
R BNo Adapt21.940.643422.320.674521.890.6427
Meta [11]22.830.681823.140.702022.750.6822
Cycle [16]23.980.715424.110.722324.430.7305
TTA-EVF25.850.763826.910.775227.160.7768
R HNo Adapt27.830.815827.610.812127.920.8222
Meta [11]28.110.820628.320.823428.540.8299
Cycle [16]28.430.830629.120.838929.430.8410
TTA-EVF30.330.846431.830.858731.860.8590
", + "image_path": "a2203e4705380265836f9d2ddd8fac22d0b11a930ec6ede90d07d9715dfb9837.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 479, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 503 + ], + "type": "text", + "content": "update and TTA setup intact while only incorporating the cyclic self-supervised mechanism [16] into Eq.(11)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 506, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 590 + ], + "type": "text", + "content": "TTA in Cross-domain Datasets. We conduct cross-domain experiments using three existing high-quality real event datasets: BS-ERGB [52], HS-ERGB [51], and HighREV [46]. We refer to each datasets as B, H, and R, respectively. In each experiment, we alternate between the source and target datasets, but HS-ERGB exclusively provides with test sequences and is used solely as the target dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 712 + ], + "type": "text", + "content": "Table 1 presents the quantitative results in cross-domain datasets. We can confirm that TTA-EVF has significant advantages over other methods. First, Meta [11] using the meta-learning shows an overall but subtle improvement in performance. It indicates that it is not particularly effective in adapting to entirely new domains, especially when the distribution of event data changes, operating independently for each scene. In contrast, we continuously adapt online to scenes, learning in an optimally directional manner for the entire domain rather than for each individual scene. Further-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25706" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 82, + 519, + 224 + ], + "blocks": [ + { + "bbox": [ + 123, + 71, + 470, + 81 + ], + "lines": [ + { + "bbox": [ + 123, + 71, + 470, + 81 + ], + "spans": [ + { + "bbox": [ + 123, + 71, + 470, + 81 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons with existing method in continuous domain shifting datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 82, + 519, + 224 + ], + "lines": [ + { + "bbox": [ + 75, + 82, + 519, + 224 + ], + "spans": [ + { + "bbox": [ + 75, + 82, + 519, + 224 + ], + "type": "table", + "html": "
NetworksMethodsTarget Sequences
123456Total
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
TimeReplayer [16]No Adapt34.670.909028.520.670733.200.748030.940.702434.870.906936.500.905432.810.7960
Meta [11]33.930.906328.650.673232.340.730331.140.704534.910.908336.950.910632.910.7996
Cycle [16]34.740.909827.430.667432.540.734531.220.705135.100.909137.050.909732.990.7968
TTA-EVF35.050.910129.440.676533.470.752931.900.717536.790.921337.270.914433.710.8040
TimeLens [51]No Adapt33.910.895528.170.666432.710.742830.580.702433.990.900234.290.892432.070.7896
Meta [11]34.050.900127.840.664232.800.741230.150.702233.690.890734.490.900432.090.7898
Cycle [16]34.010.900827.960.668032.410.743530.220.704134.070.900834.980.901332.140.7906
TTA-EVF34.410.905528.630.673733.090.749331.290.710034.940.905936.050.904032.800.7974
CBMNet [24]No Adapt32.700.901427.190.659732.050.732630.140.701532.240.896236.820.911731.420.7879
Meta [11]32.840.902627.320.660832.140.733430.160.701232.310.897436.650.910131.480.7885
Cycle [16]33.620.909727.820.669532.610.737230.340.705832.850.901337.190.914532.410.7988
TTA-EVF35.820.913729.250.680933.260.754031.650.720335.120.910437.540.928833.530.8023
", + "image_path": "0a5b743447736f793be1034f825dd10ed7311d0027359a249e9edb11106b710e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 76, + 228, + 515, + 403 + ], + "blocks": [ + { + "bbox": [ + 76, + 228, + 515, + 403 + ], + "lines": [ + { + "bbox": [ + 76, + 228, + 515, + 403 + ], + "spans": [ + { + "bbox": [ + 76, + 228, + 515, + 403 + ], + "type": "image", + "image_path": "adda0ba7ec56e92ef9049e96fef547b5c23b936b0bbf9529055b774277fa1125.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 404, + 515, + 415 + ], + "lines": [ + { + "bbox": [ + 77, + 404, + 515, + 415 + ], + "spans": [ + { + "bbox": [ + 77, + 404, + 515, + 415 + ], + "type": "text", + "content": "Figure 6. Visual comparisons of interpolated frames on continuous domain shifting settings. Please zoom for better view." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 53, + 453, + 281, + 564 + ], + "blocks": [ + { + "bbox": [ + 47, + 419, + 287, + 453 + ], + "lines": [ + { + "bbox": [ + 47, + 419, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 287, + 453 + ], + "type": "text", + "content": "Table 3. Comparison results between our TTA method and existing supervised/unsupervised methods in the target domain. * denotes that the values are taken from the original paper." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 453, + 281, + 564 + ], + "lines": [ + { + "bbox": [ + 53, + 453, + 281, + 564 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 281, + 564 + ], + "type": "table", + "html": "
STMethodsOnlineTimeReplayer [16]TimeLens [51]
PSNRSSIMPSNRSSIM
RRSupervised Unsupervised--32.81*0.901*
32.420.898--
BRTTA-EVF✓34.070.92934.240.930
HH-farSupervised Unsupervised--32.31*0.869*
30.07*0.834*--
BH-farTTA-EVF✓31.720.85932.420.870
HH-closeSupervised Unsupervised--31.68*0.835*
29.83*0.816*--
BH-closeTTA-EVF✓30.950.82932.110.860
", + "image_path": "6ec6464f54577aa25dc87f644520b2a0b5699021133bfa17636c69b6df24d619.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 568, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 712 + ], + "type": "text", + "content": "more, when we examine the results of Cycle [16], which incorporate a self-supervised loss into our TTA approach, it becomes evident that simply maintaining continuous online updates does not yield significant performance gains. In the case of Cycle [16], learning cyclically with generated videos during limited online updates is unable to amplify adaptation ability in situations where unavoidable errors occur as domains change. In contrast, our method identifies significant errors in pixel and samples for each scene, producing substantial improvements even within the same setting. Figure 5 also demonstrates the effectiveness of our method across various datasets and networks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 421, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 545, + 553 + ], + "type": "text", + "content": "In addition, we present the comparison of our TTA approaches with supervised/unsupervised methods in Table 3. Despite the fact that our method involves online updates on the target dataset after pre-trained from other source domain datasets, it outperforms methods trained on the target dataset using supervised or unsupervised approaches. For instance, on the HighREV (R) dataset using the TimeLens [51] model, our approach achieved a " + }, + { + "bbox": [ + 304, + 421, + 545, + 553 + ], + "type": "inline_equation", + "content": "1.43\\mathrm{dB}" + }, + { + "bbox": [ + 304, + 421, + 545, + 553 + ], + "type": "text", + "content": " higher PSNR than supervised learning, even when trained in a different domain, highlighting its ability to bridge domain gaps and improve the versatility of event-based VFI." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": "TTA in Continuous Domain Shift. We experiments in a highly challenging setting where the distribution of the dataset continually undergoes significant changes. Since there were no existing datasets with consistently changing domains in high-quality RGB-Event data, we acquire the novel Event-RGB Distribution Shift (ERDS) dataset. In the ERDS dataset, the training dataset includes HFR videos with event data in typical lighting conditions. The test dataset, on the other hand, includes scenarios where lighting conditions are consistently altered, and the event trigger threshold is adjusted accordingly to obtain sufficient information from the events. We present the results in Table 2. In the highly challenging setting of continuous domain shift" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25707" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 95, + 82, + 241, + 150 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 263, + 81 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 263, + 81 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 263, + 81 + ], + "type": "text", + "content": "Table 4. Ablation study of the proposed components." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 95, + 82, + 241, + 150 + ], + "lines": [ + { + "bbox": [ + 95, + 82, + 241, + 150 + ], + "spans": [ + { + "bbox": [ + 95, + 82, + 241, + 150 + ], + "type": "table", + "html": "
LSNR BlockLRLMCrossContinuous
No Adaptation29.9631.42
✓30.4231.98
✓✓31.5332.76
✓✓✓31.8433.01
✓✓✓31.9233.15
✓✓✓✓32.1333.53
", + "image_path": "d68ccf7a6d672a0cd3e87510e4de4039e26d51ac8a663abfb7abe3e686d8fd20.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 85, + 170, + 249, + 249 + ], + "blocks": [ + { + "bbox": [ + 58, + 158, + 276, + 168 + ], + "lines": [ + { + "bbox": [ + 58, + 158, + 276, + 168 + ], + "spans": [ + { + "bbox": [ + 58, + 158, + 276, + 168 + ], + "type": "text", + "content": "Table 5. Results based on various parameter update method." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 170, + 249, + 249 + ], + "lines": [ + { + "bbox": [ + 85, + 170, + 249, + 249 + ], + "spans": [ + { + "bbox": [ + 85, + 170, + 249, + 249 + ], + "type": "table", + "html": "
UpdatePSNR ↑SSIM ↑
Full Parameters31.040.8489
Encoder Parameters30.890.8455
Batch Normalization (BN) [18]30.440.8423
Layer Normalization (LN) [1]30.920.8477
NR Block w/. BN [18]31.570.8502
NR Block w/. LN [1]31.940.8524
NR Block w/. GN [57]32.130.8599
", + "image_path": "0ec67d5d57595024c158b916854d50d3e542c327b7d93262cb75f03aa793ffdd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 252, + 286, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 286, + 300 + ], + "type": "text", + "content": "ing, other methods fail to deliver significant improvements. In contrast, our method can achieve substantial performance gains compared to competing methods. Figure 6 clearly demonstrates the strengths of our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 311, + 140, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 140, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 140, + 324 + ], + "type": "text", + "content": "5. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 331, + 286, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 286, + 390 + ], + "type": "text", + "content": "In this section, we conduct experiments to access the efficacy of the proposed approaches. In the cross-dataset setting, we use the TimeLens [51] model, training it with BS-ERGB and evaluating it on HS-ERGB. In the continuous setting, we utilize the CBMNet [24] model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 391, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 286, + 677 + ], + "type": "text", + "content": "Effectiveness of each components. We conduct experiments in Tab. 4 while adding each component of TTA-EVF one by one. When optimizing the network using only the loss " + }, + { + "bbox": [ + 46, + 391, + 286, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_S" + }, + { + "bbox": [ + 46, + 391, + 286, + 677 + ], + "type": "text", + "content": " in Eq.(2) under the TTA setup, it can be observed that there is not a significant improvement in performance. This is because directly self-training with only LFR videos prevents the network from learning the knowledge of generating HFR, and as training progresses, it loses the ability to generate HFR learned from the source domain. In contrast, it can be observed that there is an impressive improvement in performance with the addition of the proposed modules. Ablation study on NR Block. We perform ablation by keeping all our other modules intact while changing only the update methods. As shown in Tab. 5, batch normalization (BN) [18] and layer normalization (LN) [1] lead to a decrease in performance. The proposed NR Block update method significantly improves performance, with group normalization (GN) [57] being particularly effective. The benefits of the adaptive threshold in RPS. We calculate the threshold for the reliable map between the student and teacher networks adaptively for each scene in the RPS module. Table 7 compares the adaptive thresholding method for the reliable map in the RPS module with a fixed-value thresholding approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "The ablation study of mixing ratio " + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": " in PMS. Table 6 presents an analysis of the mixing ratio between the samples in the memory bank and the generated videos in PMS." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 345, + 83, + 507, + 118 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 545, + 82 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 545, + 82 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 545, + 82 + ], + "type": "text", + "content": "Table 6. Performance based on minimum value of mixing ratio " + }, + { + "bbox": [ + 306, + 71, + 545, + 82 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 306, + 71, + 545, + 82 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 83, + 507, + 118 + ], + "lines": [ + { + "bbox": [ + 345, + 83, + 507, + 118 + ], + "spans": [ + { + "bbox": [ + 345, + 83, + 507, + 118 + ], + "type": "table", + "html": "
min(λ)0.50.60.70.80.9
Cross31.9232.0632.1332.0431.93
Continuous33.3033.4933.5333.4233.36
", + "image_path": "4c28ce03cdf19a28634f762ee7dee6e021e645ca4171e462b9f91efafe1d823c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 318, + 139, + 534, + 171 + ], + "blocks": [ + { + "bbox": [ + 347, + 127, + 504, + 137 + ], + "lines": [ + { + "bbox": [ + 347, + 127, + 504, + 137 + ], + "spans": [ + { + "bbox": [ + 347, + 127, + 504, + 137 + ], + "type": "text", + "content": "Table 7. Analysis on the reliable threshold." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 139, + 534, + 171 + ], + "lines": [ + { + "bbox": [ + 318, + 139, + 534, + 171 + ], + "spans": [ + { + "bbox": [ + 318, + 139, + 534, + 171 + ], + "type": "table", + "html": "
ρ∞0.20.10.050.01Adaptive (Ours)
Cross30.8730.9331.4431.8531.4232.13
Continuous32.2132.6433.0232.9732.7833.53
", + "image_path": "a3cde57dfa83b200599c4eec0109754c217fa243ec865fa11ce33f394a7022a1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 316, + 201, + 536, + 249 + ], + "blocks": [ + { + "bbox": [ + 306, + 178, + 545, + 200 + ], + "lines": [ + { + "bbox": [ + 306, + 178, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 545, + 200 + ], + "type": "text", + "content": "Table 8. Inference time on different updating intervals. In the case of TTA-EVF, the inference time includes the training time as well." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 201, + 536, + 249 + ], + "lines": [ + { + "bbox": [ + 316, + 201, + 536, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 201, + 536, + 249 + ], + "type": "table", + "html": "
S TMethods (TimeLens [51])TimePSNRSSIM
B H-closeNo adapt80 min29.960.8413
TTA-EVF (Interval=5)81 min31.440.8541
TTA-EVF (Interval=3)82 min31.620.8571
TTA-EVF (Interval=1)88 min32.110.8598
", + "image_path": "5f9549a23d13213e00b36707e3c014ef5a39a36709c2d25043b243eec84b30e0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 254, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 254, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 305, + 254, + 545, + 301 + ], + "type": "text", + "content": "We select the mixing ratio " + }, + { + "bbox": [ + 305, + 254, + 545, + 301 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 254, + 545, + 301 + ], + "type": "text", + "content": " uniformly, but it can be set with a minimum value. The best performance is observed when the minimum value, " + }, + { + "bbox": [ + 305, + 254, + 545, + 301 + ], + "type": "inline_equation", + "content": "\\min (\\lambda)" + }, + { + "bbox": [ + 305, + 254, + 545, + 301 + ], + "type": "text", + "content": ", is set to 0.7, demonstrating robustness to parameter variations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 314, + 471, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 314, + 471, + 327 + ], + "spans": [ + { + "bbox": [ + 306, + 314, + 471, + 327 + ], + "type": "text", + "content": "6. Analysis of Updating Interval" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 334, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 545, + 466 + ], + "type": "text", + "content": "We present the performance based on the updating interval in Table 8. Our TTA method is highly cost-effective despite performing training during the inference process. TTA-EVF, which updates only once during the multiple inferences when an video triplet is input, exhibits minimal time difference compared to simple inference. When the interval is increased, the time required is almost the same as simple inference, yet the quality of the generated videos is significantly higher. For example, with an interval of 5, there is only a 1 minute difference in total inference time, but the PSNR is improved by " + }, + { + "bbox": [ + 304, + 334, + 545, + 466 + ], + "type": "inline_equation", + "content": "1.48\\mathrm{dB}" + }, + { + "bbox": [ + 304, + 334, + 545, + 466 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 477, + 378, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 477, + 378, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 477, + 378, + 490 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 498, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 616 + ], + "type": "text", + "content": "In this paper, we present a Test-Time Adaptation (TTA) for event-based video frame interpolation. We effectively perform online TTA using the newly proposed modules in both cross-domain and continuous domain-shifting settings. Additionally, we acquire the ERDS dataset, containing high-quality RGB and event datasets with continuous distribution changes. Our TTA-EVF, despite online learning, shows minimal speed differences compared to simple inference and outperforms both supervised and unsupervised methods in cross-domain experiments, showcasing its effectiveness." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This research was supported by National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (NRF2022R1A2B5B03002636) and the Challengeable Future Defense Technology Research and Development Program through the Agency For Defense Development (ADD) funded by the Defense Acquisition Program Administration (DAPA) in 2024 (No.912768601)." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25708" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 179 + ], + "type": "text", + "content": "[2] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3703-3712, 2019. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 245 + ], + "type": "text", + "content": "[3] Wenbo Bao, Wei-Sheng Lai, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Memc-net: Motion estimation and motion compensation driven neural network for video interpolation and enhancement. IEEE transactions on pattern analysis and machine intelligence, 43(3):933-948, 2019. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 246, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 288, + 290 + ], + "type": "text", + "content": "[4] Magdalena Biesialska, Katarzyna Biesialska, and Marta R Costa-Jussa. Continual lifelong learning in natural language processing: A survey. arXiv preprint arXiv:2012.09823, 2020.5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "text", + "content": "[5] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 336, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 288, + 379 + ], + "type": "text", + "content": "[6] Hoonhee Cho and Kuk-Jin Yoon. Event-image fusion stereo using cross-modality feature propagation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 454-462, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "type": "text", + "content": "[7] Hoonhee Cho and Kuk-Jin Yoon. Selection and cross similarity for event-image deep stereo. In European Conference on Computer Vision, pages 470-486. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 458 + ], + "type": "text", + "content": "[8] Hoonhee Cho, Jegyeong Cho, and Kuk-Jin Yoon. Learning adaptive dense event stereo from the image domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17797-17807, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 459, + 288, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 459, + 288, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 459, + 288, + 512 + ], + "type": "text", + "content": "[9] Hoonhee Cho, Yuhwan Jeong, Taewoo Kim, and Kuk-Jin Yoon. Non-coaxial event-guided motion deblurring with spatial alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12492-12503, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "text", + "content": "[10] Hoonhee Cho, Hyeonseong Kim, Yujeong Chae, and Kuk-Jin Yoon. Label-free event-based object recognition via joint learning with image reconstruction from events. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19866-19877, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "type": "text", + "content": "[11] Myungsub Choi, Janghoon Choi, Sungyong Baik, Tae Hyun Kim, and Kyoung Mu Lee. Scene-adaptive video frame interpolation via meta-learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9444-9453, 2020. 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "text", + "content": "[12] Myungsub Choi, Heewon Kim, Bohyung Han, Ning Xu, and Young Mu Lee. Channel attention is all you need for video frame interpolation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10663-10671, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. Advances in Neural Information Processing Systems, 35:29374-29385, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "text", + "content": "[15] Shurui Gui, Chaoyue Wang, Qihua Chen, and Dacheng Tao. Featureflow: Robust video interpolation via structure-to-texture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14004-14013, 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 547, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 240 + ], + "type": "text", + "content": "[16] Weihua He, Kaichao You, Zhendong Qiao, Xu Jia, Ziyang Zhang, Wenhui Wang, Huchuan Lu, Yaoyuan Wang, and Jianxing Liao. Timereplayer: Unlocking the potential of event cameras for video interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17804-17813, 2022. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[17] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Computer Vision, pages 624-642. Springer, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 286, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 331 + ], + "type": "text", + "content": "[18] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 3, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 332, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 374 + ], + "type": "text", + "content": "[19] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 376, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 441 + ], + "type": "text", + "content": "[20] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9000-9008, 2018. 1, 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 443, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 443, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 443, + 547, + 487 + ], + "type": "text", + "content": "[21] Shihao Jiang, Yao Lu, Hongdong Li, and Richard Hartley. Learning optical flow from a few matches. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16592-16600, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 488, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 545, + 533 + ], + "type": "text", + "content": "[22] Rico Jonschkowski, Austin Stone, Jonathan T Barron, Ariel Gordon, Kurt Konolige, and Anelia Angelova. What matters in unsupervised optical flow. In European Conference on Computer Vision, pages 557-572. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "text", + "content": "[23] Junho Kim, Inwoo Hwang, and Young Min Kim. Ev-tta: Test-time adaptation for event-based object recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17745-17754, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 579, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 634 + ], + "type": "text", + "content": "[24] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 635, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 667 + ], + "type": "text", + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neu" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25709" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "ral networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[27] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 288, + 195 + ], + "type": "text", + "content": "[28] Jinxiu Liang, Yixin Yang, Boyu Li, Peiqi Duan, Yong Xu, and Boxin Shi. Coherent event guided low-light video enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10615-10625, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 252 + ], + "type": "text", + "content": "[29] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "text", + "content": "[30] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. Learning event-driven video deblurring and interpolation. ECCV, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "type": "text", + "content": "[31] Wei Lin, Muhammad Jehanzeb Mirza, Mateusz Kozinski, Horst Possegger, Hilde Kuehne, and Horst Bischof. Video test-time adaptation for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22952-22961, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "text", + "content": "[32] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808-21820, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 410, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 287, + 453 + ], + "type": "text", + "content": "[33] Huaijia Lin Jiangbo Lu Liying Lu, Ruizheng Wu and Jiaya Jia. Video frame interpolation with transformer. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 455, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 499 + ], + "type": "text", + "content": "[34] Weng Fei Low and Gim Hee Lee. Robust e-nerf: Nerf from sparse & noisy events under non-uniform motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 18335–18346, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "text", + "content": "[35] Kunming Luo, Chuan Wang, Shuaicheng Liu, Haoqiang Fan, Jue Wang, and Jian Sun. Upflow: Upsampling pyramid for unsupervised optical flow learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1045-1054, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 556, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 610 + ], + "type": "text", + "content": "[36] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 612, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 656 + ], + "type": "text", + "content": "[37] Junheum Park, Keunsoo Ko, Chul Lee, and Chang-Su Kim. Bmbc: Bilateral motion estimation with bilateral cost volume for video interpolation. In European Conference on Computer Vision, 2020. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "text", + "content": "[38] Junheum Park, Chul Lee, and Chang-Su Kim. Asymmetric bilateral motion estimation for video frame interpolation. In International Conference on Computer Vision, 2021. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[39] Fitssum A Reda, Deqing Sun, Aysegul Dundar, Mohammad Shoeybi, Guilin Liu, Kevin J Shih, Andrew Tao, Jan Kautz," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "and Bryan Catanzaro. Unsupervised video interpolation using cycle consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 892-900, 2019. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "text", + "content": "[40] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems, 33:11539-11551, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "text", + "content": "[41] Wei Shang, Dongwei Ren, Dongqing Zou, Jimmy S. Ren, Ping Luo, and Wangmeng Zuo. Bringing events into video deblurring with non-consecutively blurry frames. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4531-4540, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 231, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 285 + ], + "type": "text", + "content": "[42] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. Advances in Neural Information Processing Systems, 35:14274-14289, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "text", + "content": "[43] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. Xvfi: extreme video frame interpolation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 331, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 376 + ], + "type": "text", + "content": "[44] Chen Song, Qixing Huang, and Chandrajit Bajaj. E-cir: Event-enhanced continuous intensity recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7803-7812, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 376, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 431 + ], + "type": "text", + "content": "[45] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 433, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 498 + ], + "type": "text", + "content": "[46] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18043-18052, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 500, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 554 + ], + "type": "text", + "content": "[47] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 555, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 599 + ], + "type": "text", + "content": "[48] Zhaoning Sun, Nico Messikommer, Daniel Gehrig, and Davide Scaramuzza. Ess: Learning event-based semantic segmentation from still images. ArXiv, abs/2203.10016, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "text", + "content": "[49] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "text", + "content": "[50] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[51] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25710" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16155-16164, 2021. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[52] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. arXiv preprint arXiv:2203.17191, 2022. 1, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "text", + "content": "[53] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "text", + "content": "[54] Yusheng Wang, Yunfan Lu, Ye Gao, Lin Wang, Zhihang Zhong, Yinqiang Zheng, and Atsushi Yamashita. Efficient video deblurring guided by motion magnitude. In European Conference on Computer Vision, pages 413-429. Springer, 2022. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "text", + "content": "[55] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based blurry frame interpolation under blind exposure. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1588-1598, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 364 + ], + "type": "text", + "content": "[56] Song Wu, Kaichao You, Weihua He, Chen Yang, Yang Tian, Yaoyuan Wang, Ziyang Zhang, and Jianxing Liao. Video interpolation by event-driven anisotropic adjustment of optical flow. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 399 + ], + "type": "text", + "content": "[57] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 4, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "text", + "content": "[58] Ruihao Xia, Chaoqiang Zhao, Meng Zheng, Ziyan Wu, Qiyu Sun, and Yang Tang. Cmda: Cross-modality domain adaptation for nighttime semantic segmentation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 21515-21524, 2023. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 455, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 510 + ], + "type": "text", + "content": "[59] Fang Xu, Lei Yu, Bishan Wang, Wen Yang, Gui-Song Xia, Xu Jia, Zhendong Qiao, and Jianzhuang Liu. Motion deblurring with real events. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2583-2592, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 510, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 544 + ], + "type": "text", + "content": "[60] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. In NeurIPS, 2019. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "text", + "content": "[61] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127(8): 1106-1125, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "text", + "content": "[62] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[63] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems, 35: 38629-38642, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[64] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proceedings of" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 251 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17765-17774, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "text", + "content": "[65] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17765-17774, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 196 + ], + "type": "text", + "content": "[66] Yizhe Zhang, Shubhankar Borse, Hong Cai, and Fatih Porikli. Auxadapt: Stable and efficient test-time adaptation for temporally consistent video semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2339-2348, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 198, + 547, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 547, + 251 + ], + "type": "text", + "content": "[67] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based optical flow using motion compensation. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25711" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/532b2907-27ca-48aa-bd4d-3e8291e221ab_content_list.json b/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/532b2907-27ca-48aa-bd4d-3e8291e221ab_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..228dd43fd9e35a1fe4c5cdd7f38eca63575e7fc3 --- /dev/null +++ b/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/532b2907-27ca-48aa-bd4d-3e8291e221ab_content_list.json @@ -0,0 +1,1420 @@ +[ + { + "type": "text", + "text": "TULIP: Multi-camera 3D Precision Assessment of Parkinson's Disease", + "text_level": 1, + "bbox": [ + 127, + 130, + 841, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kyungdo Kim $^{1}$ , Sihan Lyu $^{1}$ , Sneha Mantri $^{2}$ , Timothy W. Dunn $^{1}$ \nDepartment of ${}^{1}$ Biomedical Engineering, $^{2}$ Neurology $\\}$ , Duke University, Durham, NC, USA \n{kyungdo.kim, sihan.lyu, sneha.mantri, timothy.dunn}@duke.edu", + "bbox": [ + 114, + 179, + 854, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Parkinson's disease (PD) is a devastating movement disorder accelerating in global prevalence, but a lack of precision symptom measurement has made the development of effective therapies challenging. The Unified Parkinson's Disease Rating Scale (UPDRS) is the gold standard for assessing motor symptom severity, yet its manual scoring criteria are vague and subjective, resulting in coarse and noisy clinical assessments. Machine learning approaches have the potential to modernize PD symptom assessments by making them more quantitative, objective, and scalable. However, the lack of benchmark video datasets for PD motor exams hinders model development. Here, we introduce the TULIP dataset to bridge this gap. TULIP emphasizes precision and comprehensiveness, comprising multi-view video recordings (6 cameras) of 25 UPDRS motor exam activities, together with ratings by 3 clinical experts, in a cohort of Parkinson's patients and healthy controls. The multi-view recordings enable 3D reconstructions of body movement that better capture disease signatures than more conventional 2D methods. Using the dataset, we establish a baseline model for predicting UPDRS scores from 3D poses, illustrating how existing diagnostics could be automated. Looking ahead, TULIP could aid the development of new precision diagnostics that transcend UPDRS scores, providing a deeper understanding of PD and its potential treatments.", + "bbox": [ + 75, + 300, + 473, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 723, + 209, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Parkinson's disease (PD) is a progressive neurodegenerative disorder leading to a spectrum of motor impairments that differ between individuals and worsen over time. In the last quarter-century, the global prevalence of PD has increased two-fold and continues to rise, now impacting over 8.5 million individuals [1]. Despite decades of research, no cure has emerged, and the focus has shifted towards managing symptoms to improve quality of life.", + "bbox": [ + 75, + 748, + 468, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The current standard for PD clinical assessment, the Movement Disorder Society-Sponsored Revision of the", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/49c569d8bc840333529ab118ef18738df658f4fffb958be4025afcc046429620.jpg", + "image_caption": [ + "Figure 1. Landscape of Parkinson's disease datasets. The black icons indicate which activities were recorded. Previous studies recorded a small number of activities, typically either walking, finger tapping, or sitting/standing. TULIP includes 25 different activities, capturing a wider set of movements and postures." + ], + "image_footnote": [], + "bbox": [ + 503, + 267, + 890, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unified Parkinson's Disease Rating Scale (UPDRS), relies on a manual scoring system that can lead to imprecise diagnoses. While the scale is validated [2-4], examiners score subjects coarsely based on vague guidelines. For instance, 'global spontaneity' is rated on an ordinal scale from '1 (Slight)' to '4 (Severe)' [5]. This manual scoring is labor-intensive and a source of variability, and it precludes identification of fine-scale motor pathologies. Numerous studies have identified significant inter-rater differences in UPDRS scoring, with disagreements as large as $53\\%$ [6].", + "bbox": [ + 496, + 628, + 890, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in machine learning have greatly enhanced healthcare workflows and patient outcomes across many clinical domains, but applying machine learning approaches to PD remains challenging due to a shortage of training and benchmark datasets. Some approaches have explored modeling PD using electronic health record [7] or neuroimaging data [8], but ultimately such datasets rely on clinical diagnosis of motor symptoms [9]. While small", + "bbox": [ + 496, + 780, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22551", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "subsets of symptoms, such as tremor, can be measured precisely with wearable devices [10], video recordings can capture the wide diversity of PD symptoms, which occur on multiple spatiotemporal scales. However, while there are some video-based datasets for other neurological disorders [11, 12], there are currently no benchmark video datasets for PD.", + "bbox": [ + 75, + 90, + 467, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous studies have reported PD video analyses, but these studies have been restricted to single-camera recordings, which inherently limit kinematic measurement precision. Precision can be enhanced by analyzing 3D data extracted from multi-camera recordings, but thus far 3D data has only been collected with the aid of markers (i.e., motion capture) [13, 14]. Previous video analyses have also been restricted to a very small subset of UPDRS activities as shown in Figure 1. Past approaches have thus been unable to capture the full spectrum of PD motor deficits, limiting their sensitivity and utility.", + "bbox": [ + 75, + 196, + 467, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, we introduce TULIP (Three-dimensional (3D) Understanding and Learning of Impairments in Parkinson's disease) to address this unmet need. To our knowledge, TULIP is the world's first multi-camera video dataset of PD motor exams to: 1) be available for benchmarking, 2) include all of the body movement tests of UPDRS (TULIP excludes the speech test), and 3) enable markerless 3D kinematic analysis. TULIP also nearly triples the acquisition frame rate of past studies (Figure 1).", + "bbox": [ + 75, + 362, + 467, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Using TULIP, we show that UPDRS classifiers built with 3D features significantly outperform those with 2D features. We also identify interpretable distinctions in movement variables between PD patients and healthy controls. We chose the name TULIP, a nod to the floral emblem of PD research and advocacy, to symbolize our goal for this dataset, to foster transformative new machine learning approaches for PD understanding and treatment.", + "bbox": [ + 75, + 498, + 467, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Previous work", + "text_level": 1, + "bbox": [ + 76, + 638, + 220, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Machine learning for neurological disorders", + "text_level": 1, + "bbox": [ + 76, + 664, + 449, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Machine learning is widely used for analyzing motor symptoms in neurological disorders like stroke [12]. Zhang et al. used cameras to detect ADHD by tracking child facial expressions and limb movements [15]. Seifallahi et al. employed Kinect cameras for Alzheimer's gait analysis [16]. In PD, existing computational methods focus on sit-to-stand [17], gait [13, 18] and finger tapping activities [19, 20], which have been the easiest to measure with computer vision. Machine learning has also been pivotal in predicting PD progression and symptom severity. Nilashi et al. proposed a system for predicting UPDRS scores and making holistic decisions only on biomedical voice measurements [21]. Exley et al. used machine learning to build more objective UPDRS motor symptom scores by way of force", + "bbox": [ + 75, + 688, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "places [22]. These efforts highlight the evolving landscape of PD research, combining traditional clinical assessments with advanced analytical techniques for better understanding and managing the disease.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. PD datasets", + "text_level": 1, + "bbox": [ + 500, + 162, + 625, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing datasets for PD detection and analysis fall into three primary categories: speech datasets, handwriting datasets, and body movement datasets. Speech datasets [23, 24] encompass sustained phonation tests, and voice signal analysis can be employed for PD diagnosis. Handwriting datasets [25-27], gathered in image form, primarily focus on subjects' writing of letters or shapes to assess the impact of PD on fine motor control. Body movement datasets are particularly valuable for PD analysis, as they characterize the motor symptoms and motor dysfunction commonly associated with PD progression and can be used to formulate treatment plans. Body movement data are commonly collected using expensive wearable sensors that only capture a small number of body parts [13, 28-32]. Alternatively, video-based data collected by RGB cameras [20, 33-37] provide a contactless, cost-effective, and comprehensive approach to quantify motor disorders. However, while the UPDRS comprises several dozen motor activities, most video-based data collection is limited to one or two activities, usually gait or index finger tapping. Furthermore, these video datasets use only a single camera and thus are fundamentally limited in their precision when capturing fine-scale features of body movement, an inherently 3D process. Another significant limitation of these datasets is that they are not shared with the community. This makes replication difficult and precludes further algorithmic advances by other resesearchers.", + "bbox": [ + 496, + 186, + 890, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. TULIP Dataset", + "text_level": 1, + "bbox": [ + 500, + 609, + 651, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Clinical motivation", + "text_level": 1, + "bbox": [ + 500, + 635, + 683, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The complexity of PD progression and symptomatology necessitates a high-resolution, multi-dimensional dataset of PD body kinematics. Current clinical assessment tools, such as the UPDRS, are coarse, subjective, and noisy. Studies have shown that the UPDRS score for a single exam can vary significantly between raters, with inter-rater Kappa values as low as 0.37, indicating minimal agreement between raters [38, 39]. This underscores a need for more objective, quantitative, and sensitive measures.", + "bbox": [ + 496, + 659, + 890, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this need, we require a rich data streams that accurately capture the complexities and nuances of PD motor symptoms. Research has shown that wearables [40] and force plates [22] can track some PD symptoms, but they struggle to align with UPDRS guidelines and face hospital implementation challenges [41, 42]. Video recordings, however, can efficiently capture a broad range", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22552", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a2e3bbfb6c3f9ddf0254a5dd1f55da5e1cb321922ba322849aa559de62884b58.jpg", + "image_caption": [ + "Figure 2. Overview of the TULIP Dataset. In this figure, subject faces are occluded for privacy, and * denotes there are separate left and right activity recordings. In addition to multi-view video recordings, TULIP contains UPDRS scores for each subject. We tracked 2D and 3D poses to build UPDRS classifiers." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 890, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of PD movement symptoms and can be more flexibly integrated into clinical settings. For TULIP, we recorded activities using a multi-camera setup to more accurately capture complex 3D movement features. For single-camera recordings, 2D features are measured as projections onto a plane. The resulting 2D features thus vary with perspective and depth, reducing the resolution with which articulated 3D bodies moving freely through 3D space can be measured. We designed TULIP to address this issue.", + "bbox": [ + 75, + 608, + 468, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Marker-based multi-camera motion capture systems are traditionally used to track 3D movement quantities, and recent studies have used such systems for movement disorder analyses [43-45]. But requiring that patients wear markers, together with the size of the motion capture camera array, is not feasible for routine PD assessments. A markerless diagnostic system could easily be deployed in the clinic, or even at home, where it could be used to support decision making or monitor disease progression longitudinally. Furthermore, it is easier to capture facial", + "bbox": [ + 75, + 744, + 470, + 893 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "expressions, a key element of PD assessment, with markerless systems. Markerless 3D analysis could also provide the necessary granularity to identify subtle but clinically significant changes in motor function, thereby enhancing the sensitivity of clinical trials.", + "bbox": [ + 496, + 608, + 890, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Dataset structure", + "text_level": 1, + "bbox": [ + 500, + 710, + 669, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the TULIP dataset, we enrolled 15 subjects, comprising PD patients and healthy individuals. Ten were clinically diagnosed with PD, while five had no prior PD diagnosis. Subjects demographics are detailed in Supplementary Table 2. The study was approved and conducted in accordance with the ethical guidelines set by the Duke Institutional Review Board.", + "bbox": [ + 496, + 734, + 890, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TULIP focuses on observational Part III: Motor Examination UPDRS components (Figure 2). With clinician guidance, we recorded 25 total activities from which UPDRS scores could be derived, including unilateral tasks, such", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22553", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as guided hand movement, index finger tapping, and stepping, and bilateral tasks, such as arm straightening and gait. Twenty-one of these activities are described in the UPDRS guidelines, and 4, including standing on one leg and finger tapping using all fingers, were added to encompass a wider range of body movements, on the advice of clinicians. See Supplementary Section 3 for full activity list and other paradigm details.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Video collection", + "text_level": 1, + "bbox": [ + 76, + 220, + 235, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To comprehensively capture the activities enumerated in the UPDRS, we set up the system to record a $6.3 \\times 3$ meter hexagonal space fully capturing subjects in all 6 cameras (1920 × 1200 pixels; 80 fps; Basler acA1920-155uc). Cameras were synchronized via GPIO cables controlled by campy [46]. Intrinsic and extrinsic camera parameters were fit before and after each recording session.", + "bbox": [ + 75, + 244, + 468, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. UPDRS Labeling", + "text_level": 1, + "bbox": [ + 76, + 359, + 246, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For our study, we recruited three clinical expert neurologists (2 professors, 1 MD fellow) in PD diagnosis and management with over 30 years of PD experience combined. These professionals reviewed the video recordings and scored subjects according to UPDRS guidelines, and further made an overall judgment about whether the subject had PD or was healthy. The evaluations were conducted without access to the subjects' personal or medical history information to ensure unbiased scoring. From the videos, we obtained 29 distinct UPDRS Part III scores summarizing motor function, including bilateral Hand movements, Kinetic body tremor, Resting tremor amplitude arms and elbows, Facial expressions, Gait, Posture, Dyskinesias, and more. The mapping between the 25 recorded activities and the 29 UPDRS scores is delineated in Supplementary Section 3. We derived 'gold-standard' UPDRS labels from clinician scores using majority voting, as per prior research [35].", + "bbox": [ + 75, + 383, + 472, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Analysis Methods", + "text_level": 1, + "bbox": [ + 76, + 656, + 251, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Pose Estimation", + "text_level": 1, + "bbox": [ + 76, + 680, + 236, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We used TULIP to test automated UPDRS scoring models for finger tapping and gait given features extracted from 2D and 3D pose sequences. For pose estimation, we utilized Mediapipe [47] and MMPose [48], which tracked complementary sets of body keypoints on study subjects without fine-tuning. We used MediaPipe to track 21 keypoints on each hand for finger tapping, and we used MMpose (Halpe26 configuration) to track 26 keypoints on the legs, arms, trunk, and head for gait. To estimate 3D poses, we triangulated 2D poses using the direct linear transformation algorithm. We further improved 3D tracking by interpolating over keypoint outliers and smoothing. We evaluated pose tracking precision via comparison to human keypoint", + "bbox": [ + 75, + 703, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "annotation and via body segment length consistency. Mean errors relative to manual annotations were $21\\mathrm{mm}$ for finger tapping and $56\\mathrm{mm}$ for gait, corresponding to reprojection errors of 22 and 16 pixels, respectively. Tracked body segment lengths were also stable, with standard deviations below $20\\mathrm{mm}$ across all frames in each recording (Supplementary Section 4).", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Features for disease classification", + "text_level": 1, + "bbox": [ + 498, + 205, + 790, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To assess the utility of our dataset and establish benchmarks for future studies, we trained classifiers to distinguish PD from healthy individuals and predict UPDRS score, using spatiotemporal kinematic features extracted from 2D and 3D pose sequences. In this study, we focused on only two activities: index finger tapping and gait, both of which are frequently analyzed in PD research [32, 35, 40, 49-52]. We designed our feature set to enable accurate classification but also include clinical variables that are commonly utilized when scoring UPDRS manually.", + "bbox": [ + 496, + 229, + 890, + 380 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6515f1d61926b6159c0e3f0808e6ffdb09376299c195be0d58589d8b6039dc98.jpg", + "image_caption": [ + "Figure 3. Overview of feature extraction from finger tapping and gait recordings." + ], + "image_footnote": [], + "bbox": [ + 501, + 392, + 890, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1 Index finger tapping", + "text_level": 1, + "bbox": [ + 500, + 756, + 697, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the index finger tapping task, subjects were instructed to perform tapping movements solely with their thumb and index finger while keeping their palm oriented forward. This activity yields valuable data on tremor, rigidity, and the stability of hand posture [35, 53]. For feature extraction, we first segmented pose traces based on tapping events, measuring the angle between the thumb and index fingertip, as well as the distance between them (Figure 3).", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22554", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tapping events were detected as minima in finger-thumb distance traces, i.e. the frames where the index fingertip and thumb made contact.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After detecting the tapping events, the dataset was segmented into windows, each encompassing a sequence of 10 tapping events, in accordance with the UPDRS finger tapping criteria [5] and aligned with methodologies employed in prior research [35]. For each tapping event, separately for 2D and 3D feature sets, we computed three features from the angle $\\theta_{f} = \\cos^{-1}\\left(\\frac{\\vec{w t}_{f} \\cdot \\vec{w i}_{f}}{\\|\\vec{w t}_{f}\\| \\|\\vec{w i}_{f}\\|}\\right)$ , where $f$ is the frame, $\\vec{w t}$ represents the vector from the wrist to the thumb tip, and $\\vec{w i}$ denotes the vector from the wrist to the index fingertip: angular speed, angular acceleration, amplitude. For the 3D features, we also computed additional quantities, as keypoint coordinates are in real-world metric units. We utilized the 3D distance between the thumb tip and index fingertip to get speed relative features, as issues with contact vigor can be a PD indicator [35, 54]. These included wrist cartesian coordinates and finger velocities (opening velocity, closing velocity). Finger velocity, in particular, is regarded as a critical measure for PD detection [54-56]. We then computed seven summary statistics within each tapping window (arithmetic mean, minimum, maximum, median, interquartile range (IQR), standard deviation, and Shannon entropy) and used only these statistics as our final feature set for our baseline PD classifiers. These statistics provide concise summaries of the feature distribution, and previous studies have shown they can be used to discriminate between PD severity levels in finger tapping recordings [35, 57-59].", + "bbox": [ + 75, + 137, + 472, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance this final feature set, we incorporated seven additional metrics capturing dynamic aspects of tapping motions: tapping period, tapping frequency, aperiodicity, number of interruptions, number of freezing events, longest freezing event duration, and complexity of fitting periods. The methodologies for calculating these additional features are detailed in the Supplementary Section 4.2.1. In total, this resulted in a suite of 28 features for 2D and 49 features for 3D poses.", + "bbox": [ + 73, + 542, + 468, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2 Gait", + "text_level": 1, + "bbox": [ + 76, + 707, + 161, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gait analysis is pivotal in tracking the progression of PD and evaluating symptom severity. Characteristically, PD patients display aberrant gait patterns marked by shortened stride length, heightened stride variability, and a reduction in walking speed [51, 60]. Prior research has documented disparities in spatiotemporal and kinematic gait features between PD patients and healthy subjects [52, 61], but these studies have only utilized a limited range of features. Using TULIP, we assessed whether a more comprehensive set of features could enhance gait analysis, and whether gait assessments benefit from 3D vs. 2D measurements.", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the gait measurement task, subjects began from a standing position and then walked 6 meters, turned around, and returned to the starting point. We recorded subject gait for one minute, resulting in several walking cycles for each subject. The methodology for extracting gait features is depicted in Figure 3. Since subjects slowed down and turned around at the end of the recording space, to extract gait features we first segmented each linear walking bout. We then identified the precise timepoints of heel-strike and toe-off events to isolate individual strides. Heel-strike and toe-off events were identified as the extrema in the anterior-posterior trajectories of the heels and toes relative to the hip [62]. We computed two types of gait features: spatiotemporal features, such as double support time (the time that both feet are touching the ground during a gait cycle) and step length; and kinematic features, such as knee angles, etc. For example, we calculated double support time from one foot's heel-strike to the other's toe-off using $t_{ds} = \\left(f_{t_{j+1}} - f_{h_j}\\right) / F_s$ (different feet), where $f_{t_{j+1}}$ denotes the frame number of the $(j+1)$ th toe-off event, $f_{h_j}$ denotes frame number of the $j$ th heel-strike event, and $F_s$ denotes the frame rate. Features derived from 2D poses were calculated in a single side-camera view selected to best capture linear gait. For 3D poses, we derived these features after projecting the poses onto the plane best capturing subject displacement during a linear walking bout. This projection reduces perspective ambiguities, an advantage unique to 3D data. As with finger tapping, we assembled our final feature set for modeling by computing 7 summary statistics, incorporating 'CV (Coefficient of Variation)' instead of 'entropy' of these gait features. The selection of CV in gait analysis was based on its common usage in describing gait variability. Additionally, the amount of variability, such as CV in the movement pattern, has been emphasized in prior research [63].", + "bbox": [ + 496, + 90, + 893, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 500, + 652, + 586, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Clinician agreement", + "text_level": 1, + "bbox": [ + 500, + 679, + 691, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our study, we quantified the consistency and reliability of clinical evaluations using the Intraclass Correlation Coefficient (ICC) and weighted Cohen's Kappa of UPDRS scores. The ICC for mean UPDRS score was consistent across clinicians for the complete study cohort, $\\mathrm{ICC} = 0.92$ (95% CI [0.81, 0.97]). Within the PD patient group, the ICC for mean UPDRS score was 0.87 (95% CI [0.58, 0.97]), with the slight drop in agreement likely due to the increased heterogeneity of PD presentation. Pairwise weighted Kappa values (for ordinal variables) indicated similar levels of consistency for mean UPDRS (0.82 clinician 1 and 2; 0.81 clinician 1 and 3; 0.75 clinician 2 and 3 for the full cohort; 0.81, 0.80, 0.73 within the PD group, respectively). However,", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22555", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/af09b01d7d0602217851bedcafdfea3c17a220321ca2fed0db27665bd55f5170.jpg", + "image_caption": [ + "Figure 4. Schematic of the classification pipeline incorporating LOSO and bias matrices." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 893, + 290 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the reliability of specific UPDRS component scores was more variable. While some components, such as \"Finger tapping - Left hand\" (ICC = 0.89) and \"Postural stability\" (ICC = 0.90) were scored consistently, others exhibited significant discrepancies, such as \"Kinetic tremor - Left hand\" (ICC=0.4) and \"Kinetic tremor - Right hand\" (ICC=0.1). This underscores the inherent complexity of these evaluations and highlights the need for uniformity in clinical assessments.", + "bbox": [ + 75, + 340, + 472, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Disease classification", + "text_level": 1, + "bbox": [ + 76, + 488, + 272, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.1 Modeling details", + "text_level": 1, + "bbox": [ + 76, + 512, + 246, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We used our feature set to train a suite of traditional machine learning models (SVM, Random Forest, AdaBoost, XGBoost, LightGBM, MLP) classifying ordinal UPDRS scores and the overall clinical PD/healthy diagnosis, establishing baseline benchmarks for TULIP. We chose these traditional models because it allowed us to compare to previous studies that have used them with more limited feature sets. Models were trained using a leave-one-subject-out (LOSO) validation scheme for small sample sizes [35, 64], with validation performance calculated using 'gold standard' clinician labels (majority voting over the group of 3 clinicians). For finger tapping, we integrated features from both left and right activities, in line with methodologies established by previous research [35]. To help account for clinician UPDRS scoring variability, our models were trained to predict each of the clinician UPDRS scores separately. These scores were then weighted by clinician bias matrices [65] (calculated only on training samples) and averaged to produce final predicted scores. Detailed modeling pipeline can be shown in Figure 4. Before training, we also refined the feature sets for each activity by removing highly correlated variables. We calculated Pearson's correlation coefficient for all pairs of features and removed one of the pair from the dataset if $r > 0.85$ .", + "bbox": [ + 75, + 537, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.2 Index finger tapping", + "text_level": 1, + "bbox": [ + 500, + 340, + 696, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finger tapping analyses captured consistent periodicity in all subjects, with variability in frequency and amplitude. After removing highly correlated features, we arrived at a diverse set of 20 features for 3D analyses: angular speed (median), number of freezing, angular speed (entropy), wrist movement (maximum), amplitude (median), aperiodicity, amplitude (entropy), angular acceleration (minimum), amplitude (standard deviation), angular speed (minimum), complexity of fitting periods, amplitude (minimum), angular acceleration (maximum), amplitude (IQR), longest freezing duration, wrist movement (mean), wrist movement (minimum), closing velocity (maximum), wrist movement (median), opening velocity (minimum), and 14 features for 2D analyses: longest freezing duration, angular speed (entropy), angular speed (maximum), number of freezing events, amplitude (minimum), aperiodicity, amplitude (median), angular speed (median), amplitude (IQR), angular acceleration (max), amplitude (entropy), complexity of fitting periods, angular speed (minimum), angular acceleration (minimum).", + "bbox": [ + 496, + 371, + 893, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As outlined above, we tested a suite of classifiers on two different tasks: UPDRS score prediction (0-4, ordinal scale) and overall PD/Healthy diagnosis. Among the models tested, a simple MLP neural network using 3D features was most accurate on both UPDRS score prediction (0.69 F1) and overall diagnosis (0.87 F1) as shown in Table 1. We tested previously published approaches for PD classification, including a ResNet50 CNN trained directly on 3D pose sequences and LightGBM using 2D features, on the TULIP finger tapping activity, but they did not perform as well as the 3D feature MLP. In general, models using 2D finger tapping features tended to perform poorly, supporting the case for 3D PD data acquisition and analysis. The poor performance of the ResNet50 pose sequence classifier could stem from the relatively small size of the TULIP dataset.", + "bbox": [ + 496, + 674, + 895, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22556", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f54ded90c8122b31f3bf07c1c4bc5ae04b5a1e1c11e1e075732a3cf17653d919.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 89, + 263, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b1abbe6dc0264a4c7d52d4c59f407b1d2b8212f85ee4f8487e9033ea42a3017e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 90, + 428, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc8c9c23c360909fa9799eef30d088724fa041f5af7e8a7ad2d19beac81ab716.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 90, + 537, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87089bed4e94a084ab1e3b255f120a4d8b6b9441dfd27a2189a5c1b8f6dba71d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 552, + 90, + 622, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb842eb2b4e4685d01be6bee33d48a8018da2fb09c4caf8c735268439ae4ea06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 95, + 883, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/11ab7406c894ef1e85e1485438109ecc75e0cd6c925d769aef6ad346ddff4898.jpg", + "image_caption": [ + "Figure 5. Gait feature extraction. Top: Left knee angles with real images. Bottom: Leg angles with 3D mesh models. Traces show clear differences associated with increasing impairment (higher UPDRS score means more impaired). Points marked as skyblue are occluded in a specific viewpoint. Shaded areas in feature graphs indicate $95\\%$ confidence intervals." + ], + "image_footnote": [], + "bbox": [ + 78, + 234, + 267, + 387 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4741f4b18fcaa71a9aa0976f2a7071059196d9d0164238878e39994d6e8fc9ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 234, + 424, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41846325063f228daa4e22508fe9ebd9aa54aa7bcad9cc33d882868ec0cdeab6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 234, + 529, + 383 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4de52501b256b6dbbfc7199bab3c3de8e48a8c95a5c99db2bf282f5c09c89f28.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 236, + 620, + 381 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/97459782ae9b0c327f94634aedfc383f7cfe72600b1473deaf36a374565fd711.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 243, + 883, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that Mehta et al. [17] use this ResNet50 CNN for sit-stand UPDRS classification and ensemble the ResNet50 with a Graph Convolutional Network. Here we tested the ResNet50 on its own.", + "bbox": [ + 75, + 467, + 468, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We quantified clinician assessment performance by comparing each clinician to the gold standard consensus labels for each subject. On UPDRS scoring, clinician raters were not perfect, achieving F1 scores between 0.7 and 0.87, with our top performing model commensurate with the least accurate rater. For binary PD/Healthy diagnoses, our model, which was trained only on finger tapping features, was nearly as good as one clinician (0.87 F1 for the model, 0.90 F1 for Clinician C). Note that clinicians assigned their overall PD/Healthy scores after viewing all 25 activities.", + "bbox": [ + 75, + 529, + 470, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.3 Gait", + "text_level": 1, + "bbox": [ + 76, + 705, + 163, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our analysis pipeline captures kinematics continuously and accurately at a millisecond resolution. In traces of joint angle (Figure 5), we found evidence for restricted ranges of motion in PD patients, corroborating findings from prior clinical studies [61, 66]. As with finger tapping, we removed highly correlated features, reducing the set to 57 features for 3D and 45 for 2D. The refined feature set captured a variety of step, stride, and cadence measures. Further details on these features are available in Supplementary Section 5.2.2.", + "bbox": [ + 75, + 732, + 468, + 882 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As with finger tapping, we tested classifiers for gait UP-", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DRS score and overall PD/Healthy diagnosis. For gait, a simple Random Forest model using 3D features demonstrated the best performance on UPDRS score prediction (0.72 F1). Again, models using 2D features underperformed. A similar pattern was observed for PD/Healthy classification, although we note that on this task our model performed slightly less well than a previously published Random Forest approach using a smaller set of 3D features.", + "bbox": [ + 496, + 467, + 890, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Clinician UPDRS scoring of gait was less accurate (between 0.67 and 0.73 F1), compared to consensus gold standard labels, than for finger tapping, and our model using 3D features outperformed both Clinician B and Clinician C. This finding suggests that our model, to some extent, effectively mitigates biases present in medical assessments, thus aiding in the diagnostic process and contributing to a more objective scoring system. All approaches performed less well on gait than finger tapping, potentially due to the increased natural heterogeneity of gait. The performance of all tested models can be found in Supplementary Section 5.", + "bbox": [ + 496, + 588, + 892, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 500, + 768, + 612, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here we present the TULIP dataset—a pioneering resource in PD research—as well as baselines for automated PD symptom scoring. TULIP stands out as the first dataset utilizes a markerless multi-camera setup to document both PD patients and healthy subjects, encompassing a comprehensive range of motor examination tasks from the UPDRS, which is the clinical standard for PD assessment. Despite", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22557", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ec4ea9efb330b91d107ffe42999828d1fb274978d77ff11ecdb3ab6491de54a9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TargetTaskFeaturesModelF1 scoreReference
UPDRS PredictionFingertap2D featuresMLP0.37Ours
LightGBM0.36(Islam et al., 2023)
3D featuresMLP0.69Ours
3D posesResNet500.30(Mehta et al., 2021)
VideosClinician A0.80
Clinician B0.70
Clinician C0.87
Gait2D featuresRandom Forest0.41Ours
3D featuresRandom Forest0.72
Random Forest0.34(Rehman et al. 2019)
3D posesResNet500.27(Mehta et al., 2021)
VideosClinician A0.73
Clinician B0.67
Clinician C0.67
PD/Healthy PredictionFingertap2D featuresMLP0.69Ours
LightGBM0.77(Islam et al., 2023)
3D featuresMLP0.87Ours
3D posesResNet500.64(Mehta et al., 2021)
Gait2D featuresSVM0.72Ours
3D featuresRandom Forest0.72
Random Forest0.73(Rehman et al. 2019)
3D posesResNet500.59(Mehta et al., 2021)
All 25 activitiesVideosClinician A1.00
Clinician B0.93
Clinician C0.90
", + "bbox": [ + 78, + 88, + 465, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1. UPDRS scores and holistic PD vs. healthy predictions using gold-standard labels. For UPDRS classification, F1 score denotes weighted F1 score. Clinicians' holistic decisions were not activity-specific and thus are not separated by activity in the PD vs. healthy comparison.", + "bbox": [ + 75, + 323, + 468, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the extensive adoption of UPDRS, current clinical scoring for PD remains semi-subjective, heavily reliant on clinician experience and their interpretation of the rating scale. By curating the TULIP dataset and demonstrating its utility for behavioral analysis in PD patients, we've provided a valuable tool that can propel forward the integration of advanced analytical techniques in clinical settings for PD.", + "bbox": [ + 75, + 426, + 467, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We used TULIP to benchmark models predicting UPDRS scores for gait and finger tapping directly from video recordings. For these baselines, we chose to use temporal, spatial, and kinematic feature extraction approaches to align with previous work and enable more direct comparisons between our 3D approach and existing 2D methods. On TULIP, we found that 3D features significantly outperformed 2D features, underscoring the importance of 3D video measurements. We expect UPDRS score predictions to improve in the future, for instance via networks fit directly to 3D pose time series. Nevertheless, our models achieved results that were commensurate with the accuracy of a clinician subset, demonstrating the potential of 3D computer vision approaches to enable automated and scalable PD assessments.", + "bbox": [ + 75, + 534, + 467, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We note several current limitations and future directions. While the number of recorded subjects in TULIP exceeds that of some popular 3D multi-camera 3D human datasets, such as Human 3.6M, our cohort is still relatively small, especially considering the heterogeneity of PD. TULIP also contains relatively little demographic diversity, which could hinder generalization to more representative patient populations. In addition, our baseline TULIP analyses did not incorporate data from all of the activities we recorded.", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TULIP was also collected in a large room that does not reflect the constraints of a typical clinical exam environment, and in the future it will be important to adapt the behavioral paradigm and acquisition hardware to smaller spaces.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the future, we plan to augment our dataset by incorporating a larger and more diverse set of subjects, coupled with an expanded range of activities aligned closer to activities of daily living. We are also excited by the potential of unsupervised analysis of TULIP data, which could reveal new quantitative signatures of PD independent of UPDRS scoring. Long term, we hope to expand data collection to include movement disorders other than PD, with the overall aim to unearth novel digital behavioral biomarkers.", + "bbox": [ + 496, + 152, + 890, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 301, + 617, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TULIP is the first dataset of its kind to comprehensively record the UPDRS motor examination, providing high-resolution multi-view 3D readouts together with annotations from multiple clinical experts. By leveraging the TULIP dataset, we have shown that UPDRS classifiers incorporating 3D features markedly outperform those based on 2D features, and we have identified interpretable distinctions in movement variables between PD patients and healthy controls in finger tapping and gait activities. TULIP is poised to advance Parkinson's research by aiding the development of more precise, objective, and scalable diagnostics, fostering more effective patient management and more successful treatment discovery.", + "bbox": [ + 496, + 325, + 890, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 532, + 653, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We acknowledge support from the Duke University Department of Neurology and the Duke Gilhuly Accelerator Fund. TWD is an advisor to dannce.ai and Higgs Boson Health. Also we would like to thank Timothy Lindsey, Louis DeFrate, Anshuman Sabath, Brian Lerner, Pranav Manjunath, Sophie Shi, Tianqing Li, and Joshua Wu for helping us recording the data, and three clinicians for labeling the data. We also thank Lisa Gauger and Nicole Calakos for logistical support.", + "bbox": [ + 496, + 556, + 890, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 705, + 596, + 720 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] World Health Organization. Launch of who's parkinson disease technical brief. https://www.who.int/news/item/14-06-2022-launch-of-who-s-parkinson-disease-technical-brief, 2022.1", + "[2] Kenn Freddy Pedersen, Jan Petter Larsen, and Dag Aarsland. Validation of the unified parkinson's disease rating scale (up- drs) section i as a screening and diagnostic instrument for apathy in patients with parkinson's disease. Parkinsonism & related disorders, 14(3):183-186, 2008. 1", + "[3] Neil Ramsay, Angus D Macleod, Guido Alves, Marta Camacho, Lars Forsgren, Rachael A Lawson, Jodi Maple-Grødem, Ole-Bjørn Tysnes, Caroline H Williams-Gray, Ali" + ], + "bbox": [ + 506, + 731, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22558", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "son J Yarnall, et al. Validation of a updrs-/mds-updrs-based definition of functional dependency for parkinson's disease. Parkinsonism & Related Disorders, 76:49-53, 2020.", + "[4] Katsuki Eguchi, Ichigaku Takigawa, Shinichi Shirai, Ikuko Takahashi-Iwata, Masaaki Matsushima, Takahiro Kano, Hiroaki Yaguchi, and Ichiro Yabe. Gait video-based prediction of unified parkinson's disease rating scale score: a retrospective study. BMC neurology, 23(1):358, 2023. 1", + "[5] Christopher G Goetz, Barbara C Tilley, Stephanie R Shaftman, Glenn T Stebbins, Stanley Fahn, Pablo Martinez-Martin, Werner Poewe, Cristina Sampaio, Matthew B Stern, Richard Dodel, et al. Movement disorder society-sponsored revision of the unified parkinson's disease rating scale (mdsupdrs): scale presentation and clinimetric testing results. Movement disorders: official journal of the Movement Disorder Society, 23(15):2129-2170, 2008. 1, 5", + "[6] Stefan Williams, David Wong, Jane E Alty, and Samuel D Relton. Parkinsonian hand or clinician's eye? finger tap bradykinesia interrater reliability for 21 movement disorder experts. Journal of Parkinson's Disease, (Preprint):1-12, 2023. 1", + "[7] Sullafa Kadura and Ruth B Schneider. Moving beyond alerts: Electronic health record strategies to improve inpatient parkinson's disease care. Parkinsonism & Related Disorders, 2023. 1", + "[8] Jing Zhang. Mining imaging and clinical data with machine learning approaches for the diagnosis and early detection of parkinson's disease. NPJ Parkinson's disease, 8(1):13, 2022. 1", + "[9] Ronald B Postuma, Daniela Berg, Matthew Stern, Werner Poewe, C Warren Olanow, Wolfgang Oertel, José Obeso, Kenneth Marek, Irene Litvan, Anthony E Lang, et al. Mds clinical diagnostic criteria for parkinson's disease. Movement disorders, 30(12):1591-1601, 2015. 1", + "[10] Sanghee Moon, Hyun-Je Song, Vibhash D Sharma, Kelly E Lyons, Rajesh Pahwa, Abiodun E Akinwuntan, and Hannes Devos. Classification of parkinson's disease and essential tremor based on balance and gait characteristics from wearable motion sensors via machine learning techniques: a data-driven approach. Journal of neuroengineering and rehabilitation, 17:1-8, 2020. 2", + "[11] Andrea Bandini, Sia Rezaei, Diego L Guarín, Madhura Kulkarni, Derrick Lim, Mark I Boulos, Lorne Zinman, Yana Yunusova, and Babak Taati. A new dataset for facial motion analysis in individuals with neurological disorders. IEEE Journal of Biomedical and Health Informatics, 25(4):1111-1119, 2020. 2", + "[12] Aakash Kaku, Kangning Liu, Avinash Parnandi, Haresh Rengaraj Rajamohan, Kannan Venkataramanan, Anita Venkatesan, Audre Wirtanen, Natasha Pandit, Heidi Schambra, and Carlos Fernandez-Granda. Strokerehab: A benchmark dataset for sub-second action identification. Advances in Neural Information Processing Systems, 35:1671-1684, 2022. 2", + "[13] Marta Isabel ASN Ferreira, Fabio Augusto Barbieri, Vinicius Christianini Moreno, Tiago Penedo, and João Manuel RS Tavares. Machine learning models for parkinson's disease detection and stage classification based on" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "spatial-temporal gait parameters. Gait & Posture, 98:49-55, 2022. 2", + "[14] Ferdous Wahid, Rezaul K Begg, Chris J Hass, Saman Halgamuge, and David C Ackland. Classification of parkinson's disease gait using spatial-temporal gait features. IEEE journal of biomedical and health informatics, 19(6):1794-1802, 2015. 2", + "[15] Yanyi Zhang, Ming Kong, Tianqi Zhao, Wenchen Hong, Qiang Zhu, and Fei Wu. Adhd intelligent auxiliary diagnosis system based on multimodal information fusion. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4494-4496, 2020. 2", + "[16] Mahmoud Seifallahi, Afsoon Hasani Mehraban, James E Galvin, and Behnaz Ghoraani. Alzheimer's disease detection using comprehensive analysis of timed up and go test via kinect v. 2 camera and machine learning. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 30:1589-1600, 2022. 2", + "[17] Deval Mehta, Umar Asif, Tian Hao, Erhan Bilal, Stefan Von Cavallar, Stefan Harrer, and Jeffrey Rogers. Towards automated and marker-less parkinson disease assessment: predicting updrs scores using sit-stand videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3841–3849, 2021. 2, 7", + "[18] Renfei Sun, Kun Hu, Kaylena A Ehgoetz Martens, Markus Hagenbuchner, Ah Chung Tsoi, Mohammed Bennamoun, Simon JG Lewis, and Zhiyong Wang. Higher order polynomial transformer for fine-grained freezing of gait detection. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2", + "[19] Mark Endo, Kathleen L Poston, Edith V Sullivan, Li Fei-Fei, Kilian M Pohl, and Ehsan Adeli. Gaitforemer: Self-supervised pre-training of transformers via human motion forecasting for few-shot gait impairment severity estimation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 130-139. Springer, 2022. 2", + "[20] Mandy Lu, Qingyu Zhao, Kathleen L Poston, Edith V Sullivan, Adolf Pfefferbaum, Marian Shahid, Maya Katz, Leila Montaser Kouhsari, Kevin Schulman, Arnold Milstein, et al. Quantifying parkinson's disease motor severity under uncertainty using mds-updrs videos. Medical image analysis, 73:102179, 2021. 2", + "[21] Mehrbakhsh Nilashi, Othman Ibrahim, Hossein Ahmadi, Leila Shahmoradi, and Mohammadreza Farahmand. A hybrid intelligent system for the prediction of parkinson's disease progression using machine learning techniques. Biocybernetics and Biomedical Engineering, 38(1):1-15, 2018. 2", + "[22] Trevor Exley, Sarah Moudy, Rita M Patterson, Joonghyun Kim, and Mark V Albert. Predicting updrs motor symptoms in individuals with parkinson's disease from force plates using machine learning. IEEE Journal of Biomedical and Health Informatics, 26(7):3486-3494, 2022. 2", + "[23] Max Little, Patrick McSharry, Eric Hunter, Jennifer Spielman, and Lorraine Ramig. Suitability of dysphonia measurements for telemonitoring of parkinson's disease. Nature Precedings, pages 1-1, 2008. 2" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22559", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Athanasios Tsanas, Max Little, Patrick McSharry, and Lorraine Ramig. Accurate telemonitoring of parkinson's disease progression by non-invasive speech tests. Nature Precedings, pages 1-1, 2009. 2", + "[25] Clayton R Pereira, Danilo R Pereira, Francisco A Silva, Joao P Masieiro, Silke AT Weber, Christian Hook, and João P Papa. A new computer vision-based approach to aid the diagnosis of parkinson's disease. Computer Methods and Programs in Biomedicine, 136:79-88, 2016. 2", + "[26] Clayton R Pereira, Silke AT Weber, Christian Hook, Gustavo H Rosa, and Joao P Papa. Deep learning-aided parkinson's disease diagnosis from handwritten dynamics. In 2016 29th SIBGRAPI conference on graphics, patterns and images (SIBGRAPI), pages 340–346. IEEE, 2016.", + "[27] Catherine Taleb, Maha Khachab, Chafic Mokbel, and Laurence Likforman-Sulem. Feature selection for an improved parkinson's disease identification based on handwriting. In 2017 1st International Workshop on Arabic Script Analysis and Recognition (ASAR), pages 52-56. IEEE, 2017. 2", + "[28] Charalamos Sotirakis, Zi Su, Maksymilian A Brzezicki, Niall Conway, Lionel Tarassenko, James J FitzGerald, and Chrystalina A Antoniades. Identification of motor progression in parkinson's disease using wearable sensors and machine learning. npj Parkinson's Disease, 9(1):142, 2023. 2", + "[29] Chariklia Chatzaki, Vasileios Skaramagkas, Nikolaos Tachos, Georgios Christodoulakis, Evangelia Maniadi, Zinovia Kefalopoulou, Dimitrios I Fotiadis, and Manolis Tsiknakis. The smart-insole dataset: Gait analysis using wearable sensors with a focus on elderly and parkinson's patients. Sensors, 21(8):2821, 2021.", + "[30] Marc Bachlin, Meir Plotnik, Daniel Roggen, Inbal Maidan, Jeffrey M Hausdorff, Nir Giladi, and Gerhard Troster. Wearable assistant for parkinson's disease patients with the freezing of gait symptom. IEEE Transactions on Information Technology in Biomedicine, 14(2):436-446, 2009.", + "[31] Jens Barth, Jochen Klucken, Patrick Kugler, Thomas Kammerer, Ralph Steidl, Jürgen Winkler, Joachim Hornegger, and Björn Eskofier. Biometric and mobile gait analysis for early diagnosis and therapy monitoring in parkinson's disease. In 2011 annual international conference of the IEEE engineering in medicine and biology society, pages 868-871. IEEE, 2011.", + "[32] Caroline Ribeiro De Souza, Runfeng Miao, Julia Ávila De Oliveira, Andrea Cristina De Lima-Pardini, Debora Fragoso De Campos, Carla Silva-Batista, Luis Teixeira, Solaiman Shokur, Bouri Mohamed, and Daniel Boari Coelho. A public data set of videos, inertial measurement unit, and clinical scales of freezing of gait in individuals with parkinson's disease during a turning-in-place task. Frontiers in Neuroscience, 16:832463, 2022. 2, 4", + "[33] Mandy Lu, Kathleen Poston, Adolf Pfefferbaum, Edith V Sullivan, Li Fei-Fei, Kilian M Pohl, Juan Carlos Niebles, and Ehsan Adeli. Vision-based estimation of mds-updrs gait scores for assessing parkinson's disease motor severity. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part III 23, pages 637-647. Springer, 2020. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Andrea Sabo, Sina Mehdizadeh, Kimberley-Dale Ng, Andrea Iaboni, and Babak Taati. Assessment of parkinsonian gait in older adults with dementia via human pose tracking in video data. Journal of neuroengineering and rehabilitation, 17(1):1-10, 2020.", + "[35] Md Saiful Islam, Wasifur Rahman, Abdelrahman Abdelkader, Phillip T Yang, Sangwu Lee, Jamie L Adams, Ruth B Schneider, E Dorsey, and Ehsan Hoque. Using ai to measure parkinson's disease severity at home. npj digital medicine, 2023. 4, 5, 6", + "[36] Andrea Sabo, Carolina Gorodetsky, Alfonso Fasano, Andrea Iaboni, and Babak Taati. Concurrent validity of zero instrumented walkway and video-based gait features in adults with parkinson's disease. IEEE Journal of Translational Engineering in Health and Medicine, 10:1-11, 2022.", + "[37] Samuel Rupprechter, Gareth Morinan, Yuwei Peng, Thomas Foltynie, Krista Sibley, Rimona S Weil, Louise-Ann Leyland, Fahd Baig, Francesca Morgante, Ro'ee Gilron, et al. A clinically interpretable computer-vision based method for quantifying gait in parkinson's disease. Sensors, 21(16):5437, 2021. 2", + "[38] Erin Smith, John Bertoni, Danish Bhatti, and Diego Torres-Russotto. Is the updrs a reliable tool for detecting the worse side in parkinson's disease? (1584). *Neurology*, 94:1584, 2020. 2", + "[39] Ashwani Jha, Elisa Menozzi, Rebecca Oyekan, Anna Latorre, Eoin Mulroy, Sebastian R Schreglmann, Cosmin Stamate, Ioannis Daskalopoulos, Stefan Kueppers, Marco Luchini, et al. The cloudupdrs smartphone software in parkinson's study: cross-validation against blinded human raters. npj Parkinson's Disease, 6(1):36, 2020. 2", + "[40] Delaram Safarpour, Marian L Dale, Vrutangkumar V Shah, Lauren Talman, Patricia Carlson-Kuhta, Fay B Horak, and Martina Mancini. Surrogates for rigidity and pigd mds-updrs subscores using wearable sensors. Gait & Posture, 91:186-191, 2022. 2, 4", + "[41] Erika Rovini, Carlo Maremmani, and Filippo Cavallo. How wearable sensors can support parkinson's disease diagnosis and treatment: a systematic review. Frontiers in neuroscience, 11:288959, 2017. 2", + "[42] Mercedes Barrachina-Fernández, Ana María Maitín, Carmen Sánchez-Avila, and Juan Pablo Romero. Wearable technology to detect motor fluctuations in parkinson's disease patients: current state and challenges. Sensors, 21(12):4188, 2021. 2", + "[43] Balasundaram Kadirvelu, Constantinos Gavriel, Sathiji Nageshwaran, Jackson Ping Kei Chan, Suran Nethisinghe, Stavros Athanasopoulos, Valeria Ricotti, Thomas Voit, Paola Giunti, Richard Festenstein, et al. A wearable motion capture suit and machine learning predict disease progression in friedreich's ataxia. Nature Medicine, 29(1):86-94, 2023. 3", + "[44] Benjamin Filtjens, Pieter Ginis, Alice Nieuwboer, Peter Slaets, and Bart Vanrumste. Automated freezing of gait assessment with marker-based motion capture and multistage spatial-temporal graph convolutional neural networks. Journal of NeuroEngineering and Rehabilitation, 19(1):48, 2022." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22560", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Muhammad Hassan Khan, Manuel Schneider, Muhammad Shahid Farid, and Marcin Grzegorzek. Detection of infantile movement disorders in video data using deformable part-based model. Sensors, 18(10):3202, 2018. 3", + "[46] K Severson. campy. https://github.com/ksseverson57/campy, 2022.4.1", + "[47] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for perceiving and processing reality. In Third Workshop on Computer Vision for AR/VR at IEEE Computer Vision and Pattern Recognition (CVPR) 2019, 2019. 4", + "[48] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.4", + "[49] Zhilin Guo, Weiqi Zeng, Taidong Yu, Yan Xu, Yang Xiao, Xuebing Cao, and Zhiguo Cao. Vision-based finger tapping test in patients with parkinson's disease via spatial-temporal 3d hand pose estimation. IEEE Journal of Biomedical and Health Informatics, 26(8):3848-3859, 2022. 4", + "[50] Ahnjili ZhuParris, Eva Thijssen, Willem O Elzinga, Soma Makai-Boloni, Wessel Kraaij, Geert J Groeneveld, and Robert J Doll. Treatment detection and movement disorder society-unified parkinson's disease rating scale, part iii estimation using finger tapping tasks. Movement Disorders, 38(10):1795-1805, 2023.", + "[51] Jeffrey M Hausdorff. Gait dynamics in parkinson's disease: common and distinct behavior among stride length, gait variability, and fractal-like scaling. Chaos: An Interdisciplinary Journal of Nonlinear Science, 19(2), 2009. 5", + "[52] Michele Pistacchi, Manuela Gioulis, Flavio Sanson, Ennio De Giovanni, Giuseppe Filippi, Francesca Rossetto, and Sandro Zambito Marsala. Gait analysis and clinical correlations in early parkinson's disease. *Functional neurology*, 32(1):28, 2017. 4, 5", + "[53] Tianze Yu, Kye Won Park, Martin J McKeown, and Z Jane Wang. Clinically informed automated assessment of finger tapping videos in parkinson's disease. Sensors, 23(22):9149, 2023. 4", + "[54] Kye Won Park, Eun-Jae Lee, Jun Seong Lee, Jinhoon Jeong, Nari Choi, Sungyang Jo, Mina Jung, Ja Yeon Do, Dong-Wha Kang, June-Goo Lee, et al. Machine learning-based automatic rating for cardinal symptoms of parkinson disease. Neurology, 96(13):e1761-e1769, 2021. 5", + "[55] R Agostino, A Curra, M Giovannelli, N Modugno, M Manfredi, and A Berardelli. Impairment of individual finger movements in parkinson's disease. *Movement Disorders*, 18(5):560-565, 2003.", + "[56] Eva Thijssen, Soma Makai-Boloni, Emilie van Brummelen, Jonas den Heijer, Yalcin Yavuz, Robert-Jan Doll, and Geert Jan Groeneveld. A placebo-controlled study to assess the sensitivity of finger tapping to medication effects in parkinson's disease. *Movement Disorders Clinical Practice*, 9(8):1074-1084, 2022. 5", + "[57] M Yokoe, R Okuno, T Hamasaki, Y Kurachi, K Akazawa, and S Sakoda. Opening velocity, a novel parameter, for fin" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ger tapping test in patients with parkinson's disease. Parkinsonism & related disorders, 15(6):440-444, 2009. 5", + "[58] Junjie Li, Huaiyu Zhu, Yun Pan, Haotian Wang, Zhidong Cen, Dehao Yang, and Wei Luo. Three-dimensional pattern features in finger tapping test for patients with parkinson's disease. In 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 3676-3679. IEEE, 2020.", + "[59] Noreen Akram, Haoxuan Li, Aaron Ben-Joseph, Caroline Budu, David A Gallagher, Jonathan P Bestwick, Anette Schrag, Alastair J Noyce, and Cristina Simonet. Developing and assessing a new web-based tapping test for measuring distal movement in parkinson's disease: a distal finger tapping test. Scientific reports, 12(1):386, 2022. 5", + "[60] Kevin J Brusse, Sandy Zimdars, Kathryn R Zalewski, and Teresa M Steffen. Testing functional performance in people with parkinson disease. Physical therapy, 85(2):134-141, 2005. 5", + "[61] Olumide Sofuwa, Alice Nieuwboer, Kaat Desloovere, Annemarie Willems, Fabienne Chavret, and Ilse Jonkers. Quantitative gait analysis in parkinson's disease: comparison with a healthy control group. Archives of physical medicine and rehabilitation, 86(5):1007-1013, 2005. 5, 7", + "[62] JA Zeni Jr, JG Richards, and JS2384115 Higginson. Two simple methods for determining gait events during treadmill and overground walking using kinematic data. Gait & posture, 27(4):710-714, 2008. 5", + "[63] Christopher K Rhea and Adam W Kiefer. Patterned variability in gait behavior: How can it be measured and what does it mean. Gait biometrics: basic patterns, role of neurological disorders and effects of physical activity, pages 17-44, 2014. 5", + "[64] Martin Patrick Pauli, Constantin Pohl, and Martin Golz. Balanced leave-one-subject-out cross-validation for microsleep classification. Current Directions in Biomedical Engineering, 7(2):147-150, 2021. 6", + "[65] Ryutaro Tanno, Ardavan Saeedi, Swami Sankaranarayanan, Daniel C Alexander, and Nathan Silberman. Learning from noisy labels by regularized estimation of annotator confusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11244–11253, 2019. 6", + "[66] Gwyn N Lewis, Winston D Byblow, and Sharon E Walt. Stride length regulation in parkinson's disease: the use of extrinsic, visual cues. Brain, 123(10):2077-2090, 2000. 7", + "[67] CJ Hawley, N Fineberg, AG Roberts, D Baldwin, A Sahadevan, and V Sharman. The use of the simpson angus scale for the assessment of movement disorder: a training guide. International Journal of Psychiatry in Clinical Practice, 7(4):349-2257, 2003. 2", + "[68] K Muller, CK Hemelrijk, J Westerweel, and DSW Tam. Calibration of multiple cameras for large-scale experiments using a freely moving calibration target. *Experiments in Fluids*, 61(1):7, 2020. 1", + "[69] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22561", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "social motion capture. In Proceedings of the IEEE international conference on computer vision, pages 3334-3342, 2015. 7", + "bbox": [ + 109, + 90, + 468, + 132 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "22562", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/532b2907-27ca-48aa-bd4d-3e8291e221ab_model.json b/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/532b2907-27ca-48aa-bd4d-3e8291e221ab_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cce185030ebae8e4153554a321b22dd3a48fd2cc --- /dev/null +++ b/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/532b2907-27ca-48aa-bd4d-3e8291e221ab_model.json @@ -0,0 +1,2138 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.129, + 0.131, + 0.842, + 0.151 + ], + "angle": 0, + "content": "TULIP: Multi-camera 3D Precision Assessment of Parkinson's Disease" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.18, + 0.856, + 0.234 + ], + "angle": 0, + "content": "Kyungdo Kim\\(^{1}\\), Sihan Lyu\\(^{1}\\), Sneha Mantri\\(^{2}\\), Timothy W. Dunn\\(^{1}\\) \nDepartment of \\({}^{1}\\)Biomedical Engineering, \\(^{2}\\)Neurology\\(\\}\\), Duke University, Durham, NC, USA \n{kyungdo.kim, sihan.lyu, sneha.mantri, timothy.dunn}@duke.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.694 + ], + "angle": 0, + "content": "Parkinson's disease (PD) is a devastating movement disorder accelerating in global prevalence, but a lack of precision symptom measurement has made the development of effective therapies challenging. The Unified Parkinson's Disease Rating Scale (UPDRS) is the gold standard for assessing motor symptom severity, yet its manual scoring criteria are vague and subjective, resulting in coarse and noisy clinical assessments. Machine learning approaches have the potential to modernize PD symptom assessments by making them more quantitative, objective, and scalable. However, the lack of benchmark video datasets for PD motor exams hinders model development. Here, we introduce the TULIP dataset to bridge this gap. TULIP emphasizes precision and comprehensiveness, comprising multi-view video recordings (6 cameras) of 25 UPDRS motor exam activities, together with ratings by 3 clinical experts, in a cohort of Parkinson's patients and healthy controls. The multi-view recordings enable 3D reconstructions of body movement that better capture disease signatures than more conventional 2D methods. Using the dataset, we establish a baseline model for predicting UPDRS scores from 3D poses, illustrating how existing diagnostics could be automated. Looking ahead, TULIP could aid the development of new precision diagnostics that transcend UPDRS scores, providing a deeper understanding of PD and its potential treatments." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.724, + 0.21, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Parkinson's disease (PD) is a progressive neurodegenerative disorder leading to a spectrum of motor impairments that differ between individuals and worsen over time. In the last quarter-century, the global prevalence of PD has increased two-fold and continues to rise, now impacting over 8.5 million individuals [1]. Despite decades of research, no cure has emerged, and the focus has shifted towards managing symptoms to improve quality of life." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The current standard for PD clinical assessment, the Movement Disorder Society-Sponsored Revision of the" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.268, + 0.891, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.53, + 0.894, + 0.601 + ], + "angle": 0, + "content": "Figure 1. Landscape of Parkinson's disease datasets. The black icons indicate which activities were recorded. Previous studies recorded a small number of activities, typically either walking, finger tapping, or sitting/standing. TULIP includes 25 different activities, capturing a wider set of movements and postures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Unified Parkinson's Disease Rating Scale (UPDRS), relies on a manual scoring system that can lead to imprecise diagnoses. While the scale is validated [2-4], examiners score subjects coarsely based on vague guidelines. For instance, 'global spontaneity' is rated on an ordinal scale from '1 (Slight)' to '4 (Severe)' [5]. This manual scoring is labor-intensive and a source of variability, and it precludes identification of fine-scale motor pathologies. Numerous studies have identified significant inter-rater differences in UPDRS scoring, with disagreements as large as \\(53\\%\\) [6]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Recent advances in machine learning have greatly enhanced healthcare workflows and patient outcomes across many clinical domains, but applying machine learning approaches to PD remains challenging due to a shortage of training and benchmark datasets. Some approaches have explored modeling PD using electronic health record [7] or neuroimaging data [8], but ultimately such datasets rely on clinical diagnosis of motor symptoms [9]. While small" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22551" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.196 + ], + "angle": 0, + "content": "subsets of symptoms, such as tremor, can be measured precisely with wearable devices [10], video recordings can capture the wide diversity of PD symptoms, which occur on multiple spatiotemporal scales. However, while there are some video-based datasets for other neurological disorders [11, 12], there are currently no benchmark video datasets for PD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.197, + 0.468, + 0.363 + ], + "angle": 0, + "content": "Previous studies have reported PD video analyses, but these studies have been restricted to single-camera recordings, which inherently limit kinematic measurement precision. Precision can be enhanced by analyzing 3D data extracted from multi-camera recordings, but thus far 3D data has only been collected with the aid of markers (i.e., motion capture) [13, 14]. Previous video analyses have also been restricted to a very small subset of UPDRS activities as shown in Figure 1. Past approaches have thus been unable to capture the full spectrum of PD motor deficits, limiting their sensitivity and utility." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.468, + 0.499 + ], + "angle": 0, + "content": "Here, we introduce TULIP (Three-dimensional (3D) Understanding and Learning of Impairments in Parkinson's disease) to address this unmet need. To our knowledge, TULIP is the world's first multi-camera video dataset of PD motor exams to: 1) be available for benchmarking, 2) include all of the body movement tests of UPDRS (TULIP excludes the speech test), and 3) enable markerless 3D kinematic analysis. TULIP also nearly triples the acquisition frame rate of past studies (Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.499, + 0.468, + 0.62 + ], + "angle": 0, + "content": "Using TULIP, we show that UPDRS classifiers built with 3D features significantly outperform those with 2D features. We also identify interpretable distinctions in movement variables between PD patients and healthy controls. We chose the name TULIP, a nod to the floral emblem of PD research and advocacy, to symbolize our goal for this dataset, to foster transformative new machine learning approaches for PD understanding and treatment." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.639, + 0.222, + 0.654 + ], + "angle": 0, + "content": "2. Previous work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.665, + 0.45, + 0.681 + ], + "angle": 0, + "content": "2.1. Machine learning for neurological disorders" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Machine learning is widely used for analyzing motor symptoms in neurological disorders like stroke [12]. Zhang et al. used cameras to detect ADHD by tracking child facial expressions and limb movements [15]. Seifallahi et al. employed Kinect cameras for Alzheimer's gait analysis [16]. In PD, existing computational methods focus on sit-to-stand [17], gait [13, 18] and finger tapping activities [19, 20], which have been the easiest to measure with computer vision. Machine learning has also been pivotal in predicting PD progression and symptom severity. Nilashi et al. proposed a system for predicting UPDRS scores and making holistic decisions only on biomedical voice measurements [21]. Exley et al. used machine learning to build more objective UPDRS motor symptom scores by way of force" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "places [22]. These efforts highlight the evolving landscape of PD research, combining traditional clinical assessments with advanced analytical techniques for better understanding and managing the disease." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.164, + 0.627, + 0.178 + ], + "angle": 0, + "content": "2.2. PD datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.187, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Existing datasets for PD detection and analysis fall into three primary categories: speech datasets, handwriting datasets, and body movement datasets. Speech datasets [23, 24] encompass sustained phonation tests, and voice signal analysis can be employed for PD diagnosis. Handwriting datasets [25-27], gathered in image form, primarily focus on subjects' writing of letters or shapes to assess the impact of PD on fine motor control. Body movement datasets are particularly valuable for PD analysis, as they characterize the motor symptoms and motor dysfunction commonly associated with PD progression and can be used to formulate treatment plans. Body movement data are commonly collected using expensive wearable sensors that only capture a small number of body parts [13, 28-32]. Alternatively, video-based data collected by RGB cameras [20, 33-37] provide a contactless, cost-effective, and comprehensive approach to quantify motor disorders. However, while the UPDRS comprises several dozen motor activities, most video-based data collection is limited to one or two activities, usually gait or index finger tapping. Furthermore, these video datasets use only a single camera and thus are fundamentally limited in their precision when capturing fine-scale features of body movement, an inherently 3D process. Another significant limitation of these datasets is that they are not shared with the community. This makes replication difficult and precludes further algorithmic advances by other resesearchers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.61, + 0.653, + 0.626 + ], + "angle": 0, + "content": "3. TULIP Dataset" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.684, + 0.65 + ], + "angle": 0, + "content": "3.1. Clinical motivation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.795 + ], + "angle": 0, + "content": "The complexity of PD progression and symptomatology necessitates a high-resolution, multi-dimensional dataset of PD body kinematics. Current clinical assessment tools, such as the UPDRS, are coarse, subjective, and noisy. Studies have shown that the UPDRS score for a single exam can vary significantly between raters, with inter-rater Kappa values as low as 0.37, indicating minimal agreement between raters [38, 39]. This underscores a need for more objective, quantitative, and sensitive measures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To address this need, we require a rich data streams that accurately capture the complexities and nuances of PD motor symptoms. Research has shown that wearables [40] and force plates [22] can track some PD symptoms, but they struggle to align with UPDRS guidelines and face hospital implementation challenges [41, 42]. Video recordings, however, can efficiently capture a broad range" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22552" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.891, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.54, + 0.893, + 0.581 + ], + "angle": 0, + "content": "Figure 2. Overview of the TULIP Dataset. In this figure, subject faces are occluded for privacy, and * denotes there are separate left and right activity recordings. In addition to multi-view video recordings, TULIP contains UPDRS scores for each subject. We tracked 2D and 3D poses to build UPDRS classifiers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.744 + ], + "angle": 0, + "content": "of PD movement symptoms and can be more flexibly integrated into clinical settings. For TULIP, we recorded activities using a multi-camera setup to more accurately capture complex 3D movement features. For single-camera recordings, 2D features are measured as projections onto a plane. The resulting 2D features thus vary with perspective and depth, reducing the resolution with which articulated 3D bodies moving freely through 3D space can be measured. We designed TULIP to address this issue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.745, + 0.471, + 0.895 + ], + "angle": 0, + "content": "Marker-based multi-camera motion capture systems are traditionally used to track 3D movement quantities, and recent studies have used such systems for movement disorder analyses [43-45]. But requiring that patients wear markers, together with the size of the motion capture camera array, is not feasible for routine PD assessments. A markerless diagnostic system could easily be deployed in the clinic, or even at home, where it could be used to support decision making or monitor disease progression longitudinally. Furthermore, it is easier to capture facial" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.892, + 0.684 + ], + "angle": 0, + "content": "expressions, a key element of PD assessment, with markerless systems. Markerless 3D analysis could also provide the necessary granularity to identify subtle but clinically significant changes in motor function, thereby enhancing the sensitivity of clinical trials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.67, + 0.725 + ], + "angle": 0, + "content": "3.2. Dataset structure" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.839 + ], + "angle": 0, + "content": "For the TULIP dataset, we enrolled 15 subjects, comprising PD patients and healthy individuals. Ten were clinically diagnosed with PD, while five had no prior PD diagnosis. Subjects demographics are detailed in Supplementary Table 2. The study was approved and conducted in accordance with the ethical guidelines set by the Duke Institutional Review Board." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "TULIP focuses on observational Part III: Motor Examination UPDRS components (Figure 2). With clinician guidance, we recorded 25 total activities from which UPDRS scores could be derived, including unilateral tasks, such" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22553" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "as guided hand movement, index finger tapping, and stepping, and bilateral tasks, such as arm straightening and gait. Twenty-one of these activities are described in the UPDRS guidelines, and 4, including standing on one leg and finger tapping using all fingers, were added to encompass a wider range of body movements, on the advice of clinicians. See Supplementary Section 3 for full activity list and other paradigm details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.222, + 0.236, + 0.237 + ], + "angle": 0, + "content": "3.3. Video collection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.47, + 0.352 + ], + "angle": 0, + "content": "To comprehensively capture the activities enumerated in the UPDRS, we set up the system to record a \\(6.3 \\times 3\\) meter hexagonal space fully capturing subjects in all 6 cameras (1920 × 1200 pixels; 80 fps; Basler acA1920-155uc). Cameras were synchronized via GPIO cables controlled by campy [46]. Intrinsic and extrinsic camera parameters were fit before and after each recording session." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.361, + 0.248, + 0.379 + ], + "angle": 0, + "content": "3.4. UPDRS Labeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.473, + 0.643 + ], + "angle": 0, + "content": "For our study, we recruited three clinical expert neurologists (2 professors, 1 MD fellow) in PD diagnosis and management with over 30 years of PD experience combined. These professionals reviewed the video recordings and scored subjects according to UPDRS guidelines, and further made an overall judgment about whether the subject had PD or was healthy. The evaluations were conducted without access to the subjects' personal or medical history information to ensure unbiased scoring. From the videos, we obtained 29 distinct UPDRS Part III scores summarizing motor function, including bilateral Hand movements, Kinetic body tremor, Resting tremor amplitude arms and elbows, Facial expressions, Gait, Posture, Dyskinesias, and more. The mapping between the 25 recorded activities and the 29 UPDRS scores is delineated in Supplementary Section 3. We derived 'gold-standard' UPDRS labels from clinician scores using majority voting, as per prior research [35]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.657, + 0.252, + 0.673 + ], + "angle": 0, + "content": "4. Analysis Methods" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.238, + 0.696 + ], + "angle": 0, + "content": "4.1. Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We used TULIP to test automated UPDRS scoring models for finger tapping and gait given features extracted from 2D and 3D pose sequences. For pose estimation, we utilized Mediapipe [47] and MMPose [48], which tracked complementary sets of body keypoints on study subjects without fine-tuning. We used MediaPipe to track 21 keypoints on each hand for finger tapping, and we used MMpose (Halpe26 configuration) to track 26 keypoints on the legs, arms, trunk, and head for gait. To estimate 3D poses, we triangulated 2D poses using the direct linear transformation algorithm. We further improved 3D tracking by interpolating over keypoint outliers and smoothing. We evaluated pose tracking precision via comparison to human keypoint" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "annotation and via body segment length consistency. Mean errors relative to manual annotations were \\(21\\mathrm{mm}\\) for finger tapping and \\(56\\mathrm{mm}\\) for gait, corresponding to reprojection errors of 22 and 16 pixels, respectively. Tracked body segment lengths were also stable, with standard deviations below \\(20\\mathrm{mm}\\) across all frames in each recording (Supplementary Section 4)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.207, + 0.791, + 0.221 + ], + "angle": 0, + "content": "4.2. Features for disease classification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.892, + 0.381 + ], + "angle": 0, + "content": "To assess the utility of our dataset and establish benchmarks for future studies, we trained classifiers to distinguish PD from healthy individuals and predict UPDRS score, using spatiotemporal kinematic features extracted from 2D and 3D pose sequences. In this study, we focused on only two activities: index finger tapping and gait, both of which are frequently analyzed in PD research [32, 35, 40, 49-52]. We designed our feature set to enable accurate classification but also include clinical variables that are commonly utilized when scoring UPDRS manually." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.393, + 0.892, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.694, + 0.892, + 0.723 + ], + "angle": 0, + "content": "Figure 3. Overview of feature extraction from finger tapping and gait recordings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.698, + 0.773 + ], + "angle": 0, + "content": "4.2.1 Index finger tapping" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "During the index finger tapping task, subjects were instructed to perform tapping movements solely with their thumb and index finger while keeping their palm oriented forward. This activity yields valuable data on tremor, rigidity, and the stability of hand posture [35, 53]. For feature extraction, we first segmented pose traces based on tapping events, measuring the angle between the thumb and index fingertip, as well as the distance between them (Figure 3)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22554" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "Tapping events were detected as minima in finger-thumb distance traces, i.e. the frames where the index fingertip and thumb made contact." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.473, + 0.543 + ], + "angle": 0, + "content": "After detecting the tapping events, the dataset was segmented into windows, each encompassing a sequence of 10 tapping events, in accordance with the UPDRS finger tapping criteria [5] and aligned with methodologies employed in prior research [35]. For each tapping event, separately for 2D and 3D feature sets, we computed three features from the angle \\(\\theta_{f} = \\cos^{-1}\\left(\\frac{\\vec{w t}_{f} \\cdot \\vec{w i}_{f}}{\\|\\vec{w t}_{f}\\| \\|\\vec{w i}_{f}\\|}\\right)\\), where \\(f\\) is the frame, \\(\\vec{w t}\\) represents the vector from the wrist to the thumb tip, and \\(\\vec{w i}\\) denotes the vector from the wrist to the index fingertip: angular speed, angular acceleration, amplitude. For the 3D features, we also computed additional quantities, as keypoint coordinates are in real-world metric units. We utilized the 3D distance between the thumb tip and index fingertip to get speed relative features, as issues with contact vigor can be a PD indicator [35, 54]. These included wrist cartesian coordinates and finger velocities (opening velocity, closing velocity). Finger velocity, in particular, is regarded as a critical measure for PD detection [54-56]. We then computed seven summary statistics within each tapping window (arithmetic mean, minimum, maximum, median, interquartile range (IQR), standard deviation, and Shannon entropy) and used only these statistics as our final feature set for our baseline PD classifiers. These statistics provide concise summaries of the feature distribution, and previous studies have shown they can be used to discriminate between PD severity levels in finger tapping recordings [35, 57-59]." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.544, + 0.47, + 0.68 + ], + "angle": 0, + "content": "To enhance this final feature set, we incorporated seven additional metrics capturing dynamic aspects of tapping motions: tapping period, tapping frequency, aperiodicity, number of interruptions, number of freezing events, longest freezing event duration, and complexity of fitting periods. The methodologies for calculating these additional features are detailed in the Supplementary Section 4.2.1. In total, this resulted in a suite of 28 features for 2D and 49 features for 3D poses." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.708, + 0.163, + 0.722 + ], + "angle": 0, + "content": "4.2.2 Gait" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Gait analysis is pivotal in tracking the progression of PD and evaluating symptom severity. Characteristically, PD patients display aberrant gait patterns marked by shortened stride length, heightened stride variability, and a reduction in walking speed [51, 60]. Prior research has documented disparities in spatiotemporal and kinematic gait features between PD patients and healthy subjects [52, 61], but these studies have only utilized a limited range of features. Using TULIP, we assessed whether a more comprehensive set of features could enhance gait analysis, and whether gait assessments benefit from 3D vs. 2D measurements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.622 + ], + "angle": 0, + "content": "During the gait measurement task, subjects began from a standing position and then walked 6 meters, turned around, and returned to the starting point. We recorded subject gait for one minute, resulting in several walking cycles for each subject. The methodology for extracting gait features is depicted in Figure 3. Since subjects slowed down and turned around at the end of the recording space, to extract gait features we first segmented each linear walking bout. We then identified the precise timepoints of heel-strike and toe-off events to isolate individual strides. Heel-strike and toe-off events were identified as the extrema in the anterior-posterior trajectories of the heels and toes relative to the hip [62]. We computed two types of gait features: spatiotemporal features, such as double support time (the time that both feet are touching the ground during a gait cycle) and step length; and kinematic features, such as knee angles, etc. For example, we calculated double support time from one foot's heel-strike to the other's toe-off using \\( t_{ds} = \\left(f_{t_{j+1}} - f_{h_j}\\right) / F_s \\) (different feet), where \\( f_{t_{j+1}} \\) denotes the frame number of the \\( (j+1) \\)th toe-off event, \\( f_{h_j} \\) denotes frame number of the \\( j \\)th heel-strike event, and \\( F_s \\) denotes the frame rate. Features derived from 2D poses were calculated in a single side-camera view selected to best capture linear gait. For 3D poses, we derived these features after projecting the poses onto the plane best capturing subject displacement during a linear walking bout. This projection reduces perspective ambiguities, an advantage unique to 3D data. As with finger tapping, we assembled our final feature set for modeling by computing 7 summary statistics, incorporating 'CV (Coefficient of Variation)' instead of 'entropy' of these gait features. The selection of CV in gait analysis was based on its common usage in describing gait variability. Additionally, the amount of variability, such as CV in the movement pattern, has been emphasized in prior research [63]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.654, + 0.587, + 0.669 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.692, + 0.696 + ], + "angle": 0, + "content": "5.1. Clinician agreement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.894, + 0.903 + ], + "angle": 0, + "content": "In our study, we quantified the consistency and reliability of clinical evaluations using the Intraclass Correlation Coefficient (ICC) and weighted Cohen's Kappa of UPDRS scores. The ICC for mean UPDRS score was consistent across clinicians for the complete study cohort, \\(\\mathrm{ICC} = 0.92\\) (95% CI [0.81, 0.97]). Within the PD patient group, the ICC for mean UPDRS score was 0.87 (95% CI [0.58, 0.97]), with the slight drop in agreement likely due to the increased heterogeneity of PD presentation. Pairwise weighted Kappa values (for ordinal variables) indicated similar levels of consistency for mean UPDRS (0.82 clinician 1 and 2; 0.81 clinician 1 and 3; 0.75 clinician 2 and 3 for the full cohort; 0.81, 0.80, 0.73 within the PD group, respectively). However," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22555" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.895, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.3, + 0.747, + 0.314 + ], + "angle": 0, + "content": "Figure 4. Schematic of the classification pipeline incorporating LOSO and bias matrices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.473, + 0.476 + ], + "angle": 0, + "content": "the reliability of specific UPDRS component scores was more variable. While some components, such as \"Finger tapping - Left hand\" (ICC = 0.89) and \"Postural stability\" (ICC = 0.90) were scored consistently, others exhibited significant discrepancies, such as \"Kinetic tremor - Left hand\" (ICC=0.4) and \"Kinetic tremor - Right hand\" (ICC=0.1). This underscores the inherent complexity of these evaluations and highlights the need for uniformity in clinical assessments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.489, + 0.273, + 0.503 + ], + "angle": 0, + "content": "5.2. Disease classification" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.248, + 0.528 + ], + "angle": 0, + "content": "5.2.1 Modeling details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We used our feature set to train a suite of traditional machine learning models (SVM, Random Forest, AdaBoost, XGBoost, LightGBM, MLP) classifying ordinal UPDRS scores and the overall clinical PD/healthy diagnosis, establishing baseline benchmarks for TULIP. We chose these traditional models because it allowed us to compare to previous studies that have used them with more limited feature sets. Models were trained using a leave-one-subject-out (LOSO) validation scheme for small sample sizes [35, 64], with validation performance calculated using 'gold standard' clinician labels (majority voting over the group of 3 clinicians). For finger tapping, we integrated features from both left and right activities, in line with methodologies established by previous research [35]. To help account for clinician UPDRS scoring variability, our models were trained to predict each of the clinician UPDRS scores separately. These scores were then weighted by clinician bias matrices [65] (calculated only on training samples) and averaged to produce final predicted scores. Detailed modeling pipeline can be shown in Figure 4. Before training, we also refined the feature sets for each activity by removing highly correlated variables. We calculated Pearson's correlation coefficient for all pairs of features and removed one of the pair from the dataset if \\( r > 0.85 \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.341, + 0.697, + 0.357 + ], + "angle": 0, + "content": "5.2.2 Index finger tapping" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.895, + 0.674 + ], + "angle": 0, + "content": "Finger tapping analyses captured consistent periodicity in all subjects, with variability in frequency and amplitude. After removing highly correlated features, we arrived at a diverse set of 20 features for 3D analyses: angular speed (median), number of freezing, angular speed (entropy), wrist movement (maximum), amplitude (median), aperiodicity, amplitude (entropy), angular acceleration (minimum), amplitude (standard deviation), angular speed (minimum), complexity of fitting periods, amplitude (minimum), angular acceleration (maximum), amplitude (IQR), longest freezing duration, wrist movement (mean), wrist movement (minimum), closing velocity (maximum), wrist movement (median), opening velocity (minimum), and 14 features for 2D analyses: longest freezing duration, angular speed (entropy), angular speed (maximum), number of freezing events, amplitude (minimum), aperiodicity, amplitude (median), angular speed (median), amplitude (IQR), angular acceleration (max), amplitude (entropy), complexity of fitting periods, angular speed (minimum), angular acceleration (minimum)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.896, + 0.902 + ], + "angle": 0, + "content": "As outlined above, we tested a suite of classifiers on two different tasks: UPDRS score prediction (0-4, ordinal scale) and overall PD/Healthy diagnosis. Among the models tested, a simple MLP neural network using 3D features was most accurate on both UPDRS score prediction (0.69 F1) and overall diagnosis (0.87 F1) as shown in Table 1. We tested previously published approaches for PD classification, including a ResNet50 CNN trained directly on 3D pose sequences and LightGBM using 2D features, on the TULIP finger tapping activity, but they did not perform as well as the 3D feature MLP. In general, models using 2D finger tapping features tended to perform poorly, supporting the case for 3D PD data acquisition and analysis. The poor performance of the ResNet50 pose sequence classifier could stem from the relatively small size of the TULIP dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22556" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.264, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.091, + 0.429, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.092, + 0.538, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.092, + 0.623, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.096, + 0.885, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.236, + 0.268, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.236, + 0.425, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.236, + 0.53, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.237, + 0.621, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.244, + 0.884, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.399, + 0.895, + 0.442 + ], + "angle": 0, + "content": "Figure 5. Gait feature extraction. Top: Left knee angles with real images. Bottom: Leg angles with 3D mesh models. Traces show clear differences associated with increasing impairment (higher UPDRS score means more impaired). Points marked as skyblue are occluded in a specific viewpoint. Shaded areas in feature graphs indicate \\(95\\%\\) confidence intervals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.468, + 0.47, + 0.528 + ], + "angle": 0, + "content": "Note that Mehta et al. [17] use this ResNet50 CNN for sit-stand UPDRS classification and ensemble the ResNet50 with a Graph Convolutional Network. Here we tested the ResNet50 on its own." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.471, + 0.682 + ], + "angle": 0, + "content": "We quantified clinician assessment performance by comparing each clinician to the gold standard consensus labels for each subject. On UPDRS scoring, clinician raters were not perfect, achieving F1 scores between 0.7 and 0.87, with our top performing model commensurate with the least accurate rater. For binary PD/Healthy diagnoses, our model, which was trained only on finger tapping features, was nearly as good as one clinician (0.87 F1 for the model, 0.90 F1 for Clinician C). Note that clinicians assigned their overall PD/Healthy scores after viewing all 25 activities." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.707, + 0.164, + 0.721 + ], + "angle": 0, + "content": "5.2.3 Gait" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.883 + ], + "angle": 0, + "content": "Our analysis pipeline captures kinematics continuously and accurately at a millisecond resolution. In traces of joint angle (Figure 5), we found evidence for restricted ranges of motion in PD patients, corroborating findings from prior clinical studies [61, 66]. As with finger tapping, we removed highly correlated features, reducing the set to 57 features for 3D and 45 for 2D. The refined feature set captured a variety of step, stride, and cadence measures. Further details on these features are available in Supplementary Section 5.2.2." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "As with finger tapping, we tested classifiers for gait UP-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.892, + 0.589 + ], + "angle": 0, + "content": "DRS score and overall PD/Healthy diagnosis. For gait, a simple Random Forest model using 3D features demonstrated the best performance on UPDRS score prediction (0.72 F1). Again, models using 2D features underperformed. A similar pattern was observed for PD/Healthy classification, although we note that on this task our model performed slightly less well than a previously published Random Forest approach using a smaller set of 3D features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.893, + 0.756 + ], + "angle": 0, + "content": "Clinician UPDRS scoring of gait was less accurate (between 0.67 and 0.73 F1), compared to consensus gold standard labels, than for finger tapping, and our model using 3D features outperformed both Clinician B and Clinician C. This finding suggests that our model, to some extent, effectively mitigates biases present in medical assessments, thus aiding in the diagnostic process and contributing to a more objective scoring system. All approaches performed less well on gait than finger tapping, potentially due to the increased natural heterogeneity of gait. The performance of all tested models can be found in Supplementary Section 5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.614, + 0.785 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Here we present the TULIP dataset—a pioneering resource in PD research—as well as baselines for automated PD symptom scoring. TULIP stands out as the first dataset utilizes a markerless multi-camera setup to document both PD patients and healthy subjects, encompassing a comprehensive range of motor examination tasks from the UPDRS, which is the clinical standard for PD assessment. Despite" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22557" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.467, + 0.312 + ], + "angle": 0, + "content": "
TargetTaskFeaturesModelF1 scoreReference
UPDRS PredictionFingertap2D featuresMLP0.37Ours
LightGBM0.36(Islam et al., 2023)
3D featuresMLP0.69Ours
3D posesResNet500.30(Mehta et al., 2021)
VideosClinician A0.80
Clinician B0.70
Clinician C0.87
Gait2D featuresRandom Forest0.41Ours
3D featuresRandom Forest0.72
Random Forest0.34(Rehman et al. 2019)
3D posesResNet500.27(Mehta et al., 2021)
VideosClinician A0.73
Clinician B0.67
Clinician C0.67
PD/Healthy PredictionFingertap2D featuresMLP0.69Ours
LightGBM0.77(Islam et al., 2023)
3D featuresMLP0.87Ours
3D posesResNet500.64(Mehta et al., 2021)
Gait2D featuresSVM0.72Ours
3D featuresRandom Forest0.72
Random Forest0.73(Rehman et al. 2019)
3D posesResNet500.59(Mehta et al., 2021)
All 25 activitiesVideosClinician A1.00
Clinician B0.93
Clinician C0.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.324, + 0.47, + 0.395 + ], + "angle": 0, + "content": "Table 1. UPDRS scores and holistic PD vs. healthy predictions using gold-standard labels. For UPDRS classification, F1 score denotes weighted F1 score. Clinicians' holistic decisions were not activity-specific and thus are not separated by activity in the PD vs. healthy comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.468, + 0.533 + ], + "angle": 0, + "content": "the extensive adoption of UPDRS, current clinical scoring for PD remains semi-subjective, heavily reliant on clinician experience and their interpretation of the rating scale. By curating the TULIP dataset and demonstrating its utility for behavioral analysis in PD patients, we've provided a valuable tool that can propel forward the integration of advanced analytical techniques in clinical settings for PD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.468, + 0.761 + ], + "angle": 0, + "content": "We used TULIP to benchmark models predicting UPDRS scores for gait and finger tapping directly from video recordings. For these baselines, we chose to use temporal, spatial, and kinematic feature extraction approaches to align with previous work and enable more direct comparisons between our 3D approach and existing 2D methods. On TULIP, we found that 3D features significantly outperformed 2D features, underscoring the importance of 3D video measurements. We expect UPDRS score predictions to improve in the future, for instance via networks fit directly to 3D pose time series. Nevertheless, our models achieved results that were commensurate with the accuracy of a clinician subset, demonstrating the potential of 3D computer vision approaches to enable automated and scalable PD assessments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We note several current limitations and future directions. While the number of recorded subjects in TULIP exceeds that of some popular 3D multi-camera 3D human datasets, such as Human 3.6M, our cohort is still relatively small, especially considering the heterogeneity of PD. TULIP also contains relatively little demographic diversity, which could hinder generalization to more representative patient populations. In addition, our baseline TULIP analyses did not incorporate data from all of the activities we recorded." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "TULIP was also collected in a large room that does not reflect the constraints of a typical clinical exam environment, and in the future it will be important to adapt the behavioral paradigm and acquisition hardware to smaller spaces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.892, + 0.288 + ], + "angle": 0, + "content": "In the future, we plan to augment our dataset by incorporating a larger and more diverse set of subjects, coupled with an expanded range of activities aligned closer to activities of daily living. We are also excited by the potential of unsupervised analysis of TULIP data, which could reveal new quantitative signatures of PD independent of UPDRS scoring. Long term, we hope to expand data collection to include movement disorders other than PD, with the overall aim to unearth novel digital behavioral biomarkers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.302, + 0.619, + 0.318 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.524 + ], + "angle": 0, + "content": "TULIP is the first dataset of its kind to comprehensively record the UPDRS motor examination, providing high-resolution multi-view 3D readouts together with annotations from multiple clinical experts. By leveraging the TULIP dataset, we have shown that UPDRS classifiers incorporating 3D features markedly outperform those based on 2D features, and we have identified interpretable distinctions in movement variables between PD patients and healthy controls in finger tapping and gait activities. TULIP is poised to advance Parkinson's research by aiding the development of more precise, objective, and scalable diagnostics, fostering more effective patient management and more successful treatment discovery." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.534, + 0.655, + 0.55 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.892, + 0.693 + ], + "angle": 0, + "content": "We acknowledge support from the Duke University Department of Neurology and the Duke Gilhuly Accelerator Fund. TWD is an advisor to dannce.ai and Higgs Boson Health. Also we would like to thank Timothy Lindsey, Louis DeFrate, Anshuman Sabath, Brian Lerner, Pranav Manjunath, Sophie Shi, Tianqing Li, and Joshua Wu for helping us recording the data, and three clinicians for labeling the data. We also thank Lisa Gauger and Nicole Calakos for logistical support." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.707, + 0.597, + 0.722 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.732, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[1] World Health Organization. Launch of who's parkinson disease technical brief. https://www.who.int/news/item/14-06-2022-launch-of-who-s-parkinson-disease-technical-brief, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.789, + 0.894, + 0.858 + ], + "angle": 0, + "content": "[2] Kenn Freddy Pedersen, Jan Petter Larsen, and Dag Aarsland. Validation of the unified parkinson's disease rating scale (up- drs) section i as a screening and diagnostic instrument for apathy in patients with parkinson's disease. Parkinsonism & related disorders, 14(3):183-186, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.86, + 0.895, + 0.902 + ], + "angle": 0, + "content": "[3] Neil Ramsay, Angus D Macleod, Guido Alves, Marta Camacho, Lars Forsgren, Rachael A Lawson, Jodi Maple-Grødem, Ole-Bjørn Tysnes, Caroline H Williams-Gray, Ali" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.732, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22558" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "son J Yarnall, et al. Validation of a updrs-/mds-updrs-based definition of functional dependency for parkinson's disease. Parkinsonism & Related Disorders, 76:49-53, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.135, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[4] Katsuki Eguchi, Ichigaku Takigawa, Shinichi Shirai, Ikuko Takahashi-Iwata, Masaaki Matsushima, Takahiro Kano, Hiroaki Yaguchi, and Ichiro Yabe. Gait video-based prediction of unified parkinson's disease rating scale score: a retrospective study. BMC neurology, 23(1):358, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.206, + 0.471, + 0.315 + ], + "angle": 0, + "content": "[5] Christopher G Goetz, Barbara C Tilley, Stephanie R Shaftman, Glenn T Stebbins, Stanley Fahn, Pablo Martinez-Martin, Werner Poewe, Cristina Sampaio, Matthew B Stern, Richard Dodel, et al. Movement disorder society-sponsored revision of the unified parkinson's disease rating scale (mdsupdrs): scale presentation and clinimetric testing results. Movement disorders: official journal of the Movement Disorder Society, 23(15):2129-2170, 2008. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.316, + 0.471, + 0.383 + ], + "angle": 0, + "content": "[6] Stefan Williams, David Wong, Jane E Alty, and Samuel D Relton. Parkinsonian hand or clinician's eye? finger tap bradykinesia interrater reliability for 21 movement disorder experts. Journal of Parkinson's Disease, (Preprint):1-12, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.385, + 0.471, + 0.44 + ], + "angle": 0, + "content": "[7] Sullafa Kadura and Ruth B Schneider. Moving beyond alerts: Electronic health record strategies to improve inpatient parkinson's disease care. Parkinsonism & Related Disorders, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.442, + 0.471, + 0.495 + ], + "angle": 0, + "content": "[8] Jing Zhang. Mining imaging and clinical data with machine learning approaches for the diagnosis and early detection of parkinson's disease. NPJ Parkinson's disease, 8(1):13, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.497, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[9] Ronald B Postuma, Daniela Berg, Matthew Stern, Werner Poewe, C Warren Olanow, Wolfgang Oertel, José Obeso, Kenneth Marek, Irene Litvan, Anthony E Lang, et al. Mds clinical diagnostic criteria for parkinson's disease. Movement disorders, 30(12):1591-1601, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.47, + 0.663 + ], + "angle": 0, + "content": "[10] Sanghee Moon, Hyun-Je Song, Vibhash D Sharma, Kelly E Lyons, Rajesh Pahwa, Abiodun E Akinwuntan, and Hannes Devos. Classification of parkinson's disease and essential tremor based on balance and gait characteristics from wearable motion sensors via machine learning techniques: a data-driven approach. Journal of neuroengineering and rehabilitation, 17:1-8, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.47, + 0.747 + ], + "angle": 0, + "content": "[11] Andrea Bandini, Sia Rezaei, Diego L Guarín, Madhura Kulkarni, Derrick Lim, Mark I Boulos, Lorne Zinman, Yana Yunusova, and Babak Taati. A new dataset for facial motion analysis in individuals with neurological disorders. IEEE Journal of Biomedical and Health Informatics, 25(4):1111-1119, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[12] Aakash Kaku, Kangning Liu, Avinash Parnandi, Haresh Rengaraj Rajamohan, Kannan Venkataramanan, Anita Venkatesan, Audre Wirtanen, Natasha Pandit, Heidi Schambra, and Carlos Fernandez-Granda. Strokerehab: A benchmark dataset for sub-second action identification. Advances in Neural Information Processing Systems, 35:1671-1684, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Marta Isabel ASN Ferreira, Fabio Augusto Barbieri, Vinicius Christianini Moreno, Tiago Penedo, and João Manuel RS Tavares. Machine learning models for parkinson's disease detection and stage classification based on" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "spatial-temporal gait parameters. Gait & Posture, 98:49-55, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[14] Ferdous Wahid, Rezaul K Begg, Chris J Hass, Saman Halgamuge, and David C Ackland. Classification of parkinson's disease gait using spatial-temporal gait features. IEEE journal of biomedical and health informatics, 19(6):1794-1802, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.894, + 0.263 + ], + "angle": 0, + "content": "[15] Yanyi Zhang, Ming Kong, Tianqi Zhao, Wenchen Hong, Qiang Zhu, and Fei Wu. Adhd intelligent auxiliary diagnosis system based on multimodal information fusion. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4494-4496, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.265, + 0.894, + 0.347 + ], + "angle": 0, + "content": "[16] Mahmoud Seifallahi, Afsoon Hasani Mehraban, James E Galvin, and Behnaz Ghoraani. Alzheimer's disease detection using comprehensive analysis of timed up and go test via kinect v. 2 camera and machine learning. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 30:1589-1600, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.894, + 0.433 + ], + "angle": 0, + "content": "[17] Deval Mehta, Umar Asif, Tian Hao, Erhan Bilal, Stefan Von Cavallar, Stefan Harrer, and Jeffrey Rogers. Towards automated and marker-less parkinson disease assessment: predicting updrs scores using sit-stand videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3841–3849, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.435, + 0.894, + 0.517 + ], + "angle": 0, + "content": "[18] Renfei Sun, Kun Hu, Kaylena A Ehgoetz Martens, Markus Hagenbuchner, Ah Chung Tsoi, Mohammed Bennamoun, Simon JG Lewis, and Zhiyong Wang. Higher order polynomial transformer for fine-grained freezing of gait detection. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[19] Mark Endo, Kathleen L Poston, Edith V Sullivan, Li Fei-Fei, Kilian M Pohl, and Ehsan Adeli. Gaitforemer: Self-supervised pre-training of transformers via human motion forecasting for few-shot gait impairment severity estimation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 130-139. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[20] Mandy Lu, Qingyu Zhao, Kathleen L Poston, Edith V Sullivan, Adolf Pfefferbaum, Marian Shahid, Maya Katz, Leila Montaser Kouhsari, Kevin Schulman, Arnold Milstein, et al. Quantifying parkinson's disease motor severity under uncertainty using mds-updrs videos. Medical image analysis, 73:102179, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[21] Mehrbakhsh Nilashi, Othman Ibrahim, Hossein Ahmadi, Leila Shahmoradi, and Mohammadreza Farahmand. A hybrid intelligent system for the prediction of parkinson's disease progression using machine learning techniques. Biocybernetics and Biomedical Engineering, 38(1):1-15, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[22] Trevor Exley, Sarah Moudy, Rita M Patterson, Joonghyun Kim, and Mark V Albert. Predicting updrs motor symptoms in individuals with parkinson's disease from force plates using machine learning. IEEE Journal of Biomedical and Health Informatics, 26(7):3486-3494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Max Little, Patrick McSharry, Eric Hunter, Jennifer Spielman, and Lorraine Ramig. Suitability of dysphonia measurements for telemonitoring of parkinson's disease. Nature Precedings, pages 1-1, 2008. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22559" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[24] Athanasios Tsanas, Max Little, Patrick McSharry, and Lorraine Ramig. Accurate telemonitoring of parkinson's disease progression by non-invasive speech tests. Nature Precedings, pages 1-1, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[25] Clayton R Pereira, Danilo R Pereira, Francisco A Silva, Joao P Masieiro, Silke AT Weber, Christian Hook, and João P Papa. A new computer vision-based approach to aid the diagnosis of parkinson's disease. Computer Methods and Programs in Biomedicine, 136:79-88, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[26] Clayton R Pereira, Silke AT Weber, Christian Hook, Gustavo H Rosa, and Joao P Papa. Deep learning-aided parkinson's disease diagnosis from handwritten dynamics. In 2016 29th SIBGRAPI conference on graphics, patterns and images (SIBGRAPI), pages 340–346. IEEE, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.289, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[27] Catherine Taleb, Maha Khachab, Chafic Mokbel, and Laurence Likforman-Sulem. Feature selection for an improved parkinson's disease identification based on handwriting. In 2017 1st International Workshop on Arabic Script Analysis and Recognition (ASAR), pages 52-56. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.468, + 0.427 + ], + "angle": 0, + "content": "[28] Charalamos Sotirakis, Zi Su, Maksymilian A Brzezicki, Niall Conway, Lionel Tarassenko, James J FitzGerald, and Chrystalina A Antoniades. Identification of motor progression in parkinson's disease using wearable sensors and machine learning. npj Parkinson's Disease, 9(1):142, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.468, + 0.51 + ], + "angle": 0, + "content": "[29] Chariklia Chatzaki, Vasileios Skaramagkas, Nikolaos Tachos, Georgios Christodoulakis, Evangelia Maniadi, Zinovia Kefalopoulou, Dimitrios I Fotiadis, and Manolis Tsiknakis. The smart-insole dataset: Gait analysis using wearable sensors with a focus on elderly and parkinson's patients. Sensors, 21(8):2821, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.468, + 0.58 + ], + "angle": 0, + "content": "[30] Marc Bachlin, Meir Plotnik, Daniel Roggen, Inbal Maidan, Jeffrey M Hausdorff, Nir Giladi, and Gerhard Troster. Wearable assistant for parkinson's disease patients with the freezing of gait symptom. IEEE Transactions on Information Technology in Biomedicine, 14(2):436-446, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.582, + 0.468, + 0.676 + ], + "angle": 0, + "content": "[31] Jens Barth, Jochen Klucken, Patrick Kugler, Thomas Kammerer, Ralph Steidl, Jürgen Winkler, Joachim Hornegger, and Björn Eskofier. Biometric and mobile gait analysis for early diagnosis and therapy monitoring in parkinson's disease. In 2011 annual international conference of the IEEE engineering in medicine and biology society, pages 868-871. IEEE, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.468, + 0.788 + ], + "angle": 0, + "content": "[32] Caroline Ribeiro De Souza, Runfeng Miao, Julia Ávila De Oliveira, Andrea Cristina De Lima-Pardini, Debora Fragoso De Campos, Carla Silva-Batista, Luis Teixeira, Solaiman Shokur, Bouri Mohamed, and Daniel Boari Coelho. A public data set of videos, inertial measurement unit, and clinical scales of freezing of gait in individuals with parkinson's disease during a turning-in-place task. Frontiers in Neuroscience, 16:832463, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[33] Mandy Lu, Kathleen Poston, Adolf Pfefferbaum, Edith V Sullivan, Li Fei-Fei, Kilian M Pohl, Juan Carlos Niebles, and Ehsan Adeli. Vision-based estimation of mds-updrs gait scores for assessing parkinson's disease motor severity. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part III 23, pages 637-647. Springer, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[34] Andrea Sabo, Sina Mehdizadeh, Kimberley-Dale Ng, Andrea Iaboni, and Babak Taati. Assessment of parkinsonian gait in older adults with dementia via human pose tracking in video data. Journal of neuroengineering and rehabilitation, 17(1):1-10, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[35] Md Saiful Islam, Wasifur Rahman, Abdelrahman Abdelkader, Phillip T Yang, Sangwu Lee, Jamie L Adams, Ruth B Schneider, E Dorsey, and Ehsan Hoque. Using ai to measure parkinson's disease severity at home. npj digital medicine, 2023. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[36] Andrea Sabo, Carolina Gorodetsky, Alfonso Fasano, Andrea Iaboni, and Babak Taati. Concurrent validity of zero instrumented walkway and video-based gait features in adults with parkinson's disease. IEEE Journal of Translational Engineering in Health and Medicine, 10:1-11, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[37] Samuel Rupprechter, Gareth Morinan, Yuwei Peng, Thomas Foltynie, Krista Sibley, Rimona S Weil, Louise-Ann Leyland, Fahd Baig, Francesca Morgante, Ro'ee Gilron, et al. A clinically interpretable computer-vision based method for quantifying gait in parkinson's disease. Sensors, 21(16):5437, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[38] Erin Smith, John Bertoni, Danish Bhatti, and Diego Torres-Russotto. Is the updrs a reliable tool for detecting the worse side in parkinson's disease? (1584). *Neurology*, 94:1584, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[39] Ashwani Jha, Elisa Menozzi, Rebecca Oyekan, Anna Latorre, Eoin Mulroy, Sebastian R Schreglmann, Cosmin Stamate, Ioannis Daskalopoulos, Stefan Kueppers, Marco Luchini, et al. The cloudupdrs smartphone software in parkinson's study: cross-validation against blinded human raters. npj Parkinson's Disease, 6(1):36, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[40] Delaram Safarpour, Marian L Dale, Vrutangkumar V Shah, Lauren Talman, Patricia Carlson-Kuhta, Fay B Horak, and Martina Mancini. Surrogates for rigidity and pigd mds-updrs subscores using wearable sensors. Gait & Posture, 91:186-191, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[41] Erika Rovini, Carlo Maremmani, and Filippo Cavallo. How wearable sensors can support parkinson's disease diagnosis and treatment: a systematic review. Frontiers in neuroscience, 11:288959, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[42] Mercedes Barrachina-Fernández, Ana María Maitín, Carmen Sánchez-Avila, and Juan Pablo Romero. Wearable technology to detect motor fluctuations in parkinson's disease patients: current state and challenges. Sensors, 21(12):4188, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.733, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[43] Balasundaram Kadirvelu, Constantinos Gavriel, Sathiji Nageshwaran, Jackson Ping Kei Chan, Suran Nethisinghe, Stavros Athanasopoulos, Valeria Ricotti, Thomas Voit, Paola Giunti, Richard Festenstein, et al. A wearable motion capture suit and machine learning predict disease progression in friedreich's ataxia. Nature Medicine, 29(1):86-94, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[44] Benjamin Filtjens, Pieter Ginis, Alice Nieuwboer, Peter Slaets, and Bart Vanrumste. Automated freezing of gait assessment with marker-based motion capture and multistage spatial-temporal graph convolutional neural networks. Journal of NeuroEngineering and Rehabilitation, 19(1):48, 2022." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22560" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[45] Muhammad Hassan Khan, Manuel Schneider, Muhammad Shahid Farid, and Marcin Grzegorzek. Detection of infantile movement disorders in video data using deformable part-based model. Sensors, 18(10):3202, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.176 + ], + "angle": 0, + "content": "[46] K Severson. campy. https://github.com/ksseverson57/campy, 2022.4.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[47] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for perceiving and processing reality. In Third Workshop on Computer Vision for AR/VR at IEEE Computer Vision and Pattern Recognition (CVPR) 2019, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[48] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.321, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[49] Zhilin Guo, Weiqi Zeng, Taidong Yu, Yan Xu, Yang Xiao, Xuebing Cao, and Zhiguo Cao. Vision-based finger tapping test in patients with parkinson's disease via spatial-temporal 3d hand pose estimation. IEEE Journal of Biomedical and Health Informatics, 26(8):3848-3859, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.391, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[50] Ahnjili ZhuParris, Eva Thijssen, Willem O Elzinga, Soma Makai-Boloni, Wessel Kraaij, Geert J Groeneveld, and Robert J Doll. Treatment detection and movement disorder society-unified parkinson's disease rating scale, part iii estimation using finger tapping tasks. Movement Disorders, 38(10):1795-1805, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.476, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[51] Jeffrey M Hausdorff. Gait dynamics in parkinson's disease: common and distinct behavior among stride length, gait variability, and fractal-like scaling. Chaos: An Interdisciplinary Journal of Nonlinear Science, 19(2), 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.533, + 0.469, + 0.601 + ], + "angle": 0, + "content": "[52] Michele Pistacchi, Manuela Gioulis, Flavio Sanson, Ennio De Giovanni, Giuseppe Filippi, Francesca Rossetto, and Sandro Zambito Marsala. Gait analysis and clinical correlations in early parkinson's disease. *Functional neurology*, 32(1):28, 2017. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.603, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[53] Tianze Yu, Kye Won Park, Martin J McKeown, and Z Jane Wang. Clinically informed automated assessment of finger tapping videos in parkinson's disease. Sensors, 23(22):9149, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.66, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[54] Kye Won Park, Eun-Jae Lee, Jun Seong Lee, Jinhoon Jeong, Nari Choi, Sungyang Jo, Mina Jung, Ja Yeon Do, Dong-Wha Kang, June-Goo Lee, et al. Machine learning-based automatic rating for cardinal symptoms of parkinson disease. Neurology, 96(13):e1761-e1769, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.731, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[55] R Agostino, A Curra, M Giovannelli, N Modugno, M Manfredi, and A Berardelli. Impairment of individual finger movements in parkinson's disease. *Movement Disorders*, 18(5):560-565, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.788, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[56] Eva Thijssen, Soma Makai-Boloni, Emilie van Brummelen, Jonas den Heijer, Yalcin Yavuz, Robert-Jan Doll, and Geert Jan Groeneveld. A placebo-controlled study to assess the sensitivity of finger tapping to medication effects in parkinson's disease. *Movement Disorders Clinical Practice*, 9(8):1074-1084, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[57] M Yokoe, R Okuno, T Hamasaki, Y Kurachi, K Akazawa, and S Sakoda. Opening velocity, a novel parameter, for fin" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ger tapping test in patients with parkinson's disease. Parkinsonism & related disorders, 15(6):440-444, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[58] Junjie Li, Huaiyu Zhu, Yun Pan, Haotian Wang, Zhidong Cen, Dehao Yang, and Wei Luo. Three-dimensional pattern features in finger tapping test for patients with parkinson's disease. In 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 3676-3679. IEEE, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[59] Noreen Akram, Haoxuan Li, Aaron Ben-Joseph, Caroline Budu, David A Gallagher, Jonathan P Bestwick, Anette Schrag, Alastair J Noyce, and Cristina Simonet. Developing and assessing a new web-based tapping test for measuring distal movement in parkinson's disease: a distal finger tapping test. Scientific reports, 12(1):386, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[60] Kevin J Brusse, Sandy Zimdars, Kathryn R Zalewski, and Teresa M Steffen. Testing functional performance in people with parkinson disease. Physical therapy, 85(2):134-141, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.348, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[61] Olumide Sofuwa, Alice Nieuwboer, Kaat Desloovere, Annemarie Willems, Fabienne Chavret, and Ilse Jonkers. Quantitative gait analysis in parkinson's disease: comparison with a healthy control group. Archives of physical medicine and rehabilitation, 86(5):1007-1013, 2005. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[62] JA Zeni Jr, JG Richards, and JS2384115 Higginson. Two simple methods for determining gait events during treadmill and overground walking using kinematic data. Gait & posture, 27(4):710-714, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[63] Christopher K Rhea and Adam W Kiefer. Patterned variability in gait behavior: How can it be measured and what does it mean. Gait biometrics: basic patterns, role of neurological disorders and effects of physical activity, pages 17-44, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[64] Martin Patrick Pauli, Constantin Pohl, and Martin Golz. Balanced leave-one-subject-out cross-validation for microsleep classification. Current Directions in Biomedical Engineering, 7(2):147-150, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[65] Ryutaro Tanno, Ardavan Saeedi, Swami Sankaranarayanan, Daniel C Alexander, and Nathan Silberman. Learning from noisy labels by regularized estimation of annotator confusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11244–11253, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.688, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[66] Gwyn N Lewis, Winston D Byblow, and Sharon E Walt. Stride length regulation in parkinson's disease: the use of extrinsic, visual cues. Brain, 123(10):2077-2090, 2000. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[67] CJ Hawley, N Fineberg, AG Roberts, D Baldwin, A Sahadevan, and V Sharman. The use of the simpson angus scale for the assessment of movement disorder: a training guide. International Journal of Psychiatry in Clinical Practice, 7(4):349-2257, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[68] K Muller, CK Hemelrijk, J Westerweel, and DSW Tam. Calibration of multiple cameras for large-scale experiments using a freely moving calibration target. *Experiments in Fluids*, 61(1):7, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[69] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22561" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "social motion capture. In Proceedings of the IEEE international conference on computer vision, pages 3334-3342, 2015. 7" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22562" + } + ] +] \ No newline at end of file diff --git a/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/532b2907-27ca-48aa-bd4d-3e8291e221ab_origin.pdf b/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/532b2907-27ca-48aa-bd4d-3e8291e221ab_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0a15004a15a7328efb1b7d363744992a01a36254 --- /dev/null +++ b/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/532b2907-27ca-48aa-bd4d-3e8291e221ab_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f94173f29f7947d4fc16d152e5a624cc9b3bab8d57491ec06d7f620602ac143b +size 5674230 diff --git a/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/full.md b/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b4ed6c1ee9113d805c24798157179fc39569448d --- /dev/null +++ b/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/full.md @@ -0,0 +1,273 @@ +# TULIP: Multi-camera 3D Precision Assessment of Parkinson's Disease + +Kyungdo Kim $^{1}$ , Sihan Lyu $^{1}$ , Sneha Mantri $^{2}$ , Timothy W. Dunn $^{1}$ +Department of ${}^{1}$ Biomedical Engineering, $^{2}$ Neurology $\}$ , Duke University, Durham, NC, USA +{kyungdo.kim, sihan.lyu, sneha.mantri, timothy.dunn}@duke.edu + +# Abstract + +Parkinson's disease (PD) is a devastating movement disorder accelerating in global prevalence, but a lack of precision symptom measurement has made the development of effective therapies challenging. The Unified Parkinson's Disease Rating Scale (UPDRS) is the gold standard for assessing motor symptom severity, yet its manual scoring criteria are vague and subjective, resulting in coarse and noisy clinical assessments. Machine learning approaches have the potential to modernize PD symptom assessments by making them more quantitative, objective, and scalable. However, the lack of benchmark video datasets for PD motor exams hinders model development. Here, we introduce the TULIP dataset to bridge this gap. TULIP emphasizes precision and comprehensiveness, comprising multi-view video recordings (6 cameras) of 25 UPDRS motor exam activities, together with ratings by 3 clinical experts, in a cohort of Parkinson's patients and healthy controls. The multi-view recordings enable 3D reconstructions of body movement that better capture disease signatures than more conventional 2D methods. Using the dataset, we establish a baseline model for predicting UPDRS scores from 3D poses, illustrating how existing diagnostics could be automated. Looking ahead, TULIP could aid the development of new precision diagnostics that transcend UPDRS scores, providing a deeper understanding of PD and its potential treatments. + +# 1. Introduction + +Parkinson's disease (PD) is a progressive neurodegenerative disorder leading to a spectrum of motor impairments that differ between individuals and worsen over time. In the last quarter-century, the global prevalence of PD has increased two-fold and continues to rise, now impacting over 8.5 million individuals [1]. Despite decades of research, no cure has emerged, and the focus has shifted towards managing symptoms to improve quality of life. + +The current standard for PD clinical assessment, the Movement Disorder Society-Sponsored Revision of the + +![](images/49c569d8bc840333529ab118ef18738df658f4fffb958be4025afcc046429620.jpg) +Figure 1. Landscape of Parkinson's disease datasets. The black icons indicate which activities were recorded. Previous studies recorded a small number of activities, typically either walking, finger tapping, or sitting/standing. TULIP includes 25 different activities, capturing a wider set of movements and postures. + +Unified Parkinson's Disease Rating Scale (UPDRS), relies on a manual scoring system that can lead to imprecise diagnoses. While the scale is validated [2-4], examiners score subjects coarsely based on vague guidelines. For instance, 'global spontaneity' is rated on an ordinal scale from '1 (Slight)' to '4 (Severe)' [5]. This manual scoring is labor-intensive and a source of variability, and it precludes identification of fine-scale motor pathologies. Numerous studies have identified significant inter-rater differences in UPDRS scoring, with disagreements as large as $53\%$ [6]. + +Recent advances in machine learning have greatly enhanced healthcare workflows and patient outcomes across many clinical domains, but applying machine learning approaches to PD remains challenging due to a shortage of training and benchmark datasets. Some approaches have explored modeling PD using electronic health record [7] or neuroimaging data [8], but ultimately such datasets rely on clinical diagnosis of motor symptoms [9]. While small + +subsets of symptoms, such as tremor, can be measured precisely with wearable devices [10], video recordings can capture the wide diversity of PD symptoms, which occur on multiple spatiotemporal scales. However, while there are some video-based datasets for other neurological disorders [11, 12], there are currently no benchmark video datasets for PD. + +Previous studies have reported PD video analyses, but these studies have been restricted to single-camera recordings, which inherently limit kinematic measurement precision. Precision can be enhanced by analyzing 3D data extracted from multi-camera recordings, but thus far 3D data has only been collected with the aid of markers (i.e., motion capture) [13, 14]. Previous video analyses have also been restricted to a very small subset of UPDRS activities as shown in Figure 1. Past approaches have thus been unable to capture the full spectrum of PD motor deficits, limiting their sensitivity and utility. + +Here, we introduce TULIP (Three-dimensional (3D) Understanding and Learning of Impairments in Parkinson's disease) to address this unmet need. To our knowledge, TULIP is the world's first multi-camera video dataset of PD motor exams to: 1) be available for benchmarking, 2) include all of the body movement tests of UPDRS (TULIP excludes the speech test), and 3) enable markerless 3D kinematic analysis. TULIP also nearly triples the acquisition frame rate of past studies (Figure 1). + +Using TULIP, we show that UPDRS classifiers built with 3D features significantly outperform those with 2D features. We also identify interpretable distinctions in movement variables between PD patients and healthy controls. We chose the name TULIP, a nod to the floral emblem of PD research and advocacy, to symbolize our goal for this dataset, to foster transformative new machine learning approaches for PD understanding and treatment. + +# 2. Previous work + +# 2.1. Machine learning for neurological disorders + +Machine learning is widely used for analyzing motor symptoms in neurological disorders like stroke [12]. Zhang et al. used cameras to detect ADHD by tracking child facial expressions and limb movements [15]. Seifallahi et al. employed Kinect cameras for Alzheimer's gait analysis [16]. In PD, existing computational methods focus on sit-to-stand [17], gait [13, 18] and finger tapping activities [19, 20], which have been the easiest to measure with computer vision. Machine learning has also been pivotal in predicting PD progression and symptom severity. Nilashi et al. proposed a system for predicting UPDRS scores and making holistic decisions only on biomedical voice measurements [21]. Exley et al. used machine learning to build more objective UPDRS motor symptom scores by way of force + +places [22]. These efforts highlight the evolving landscape of PD research, combining traditional clinical assessments with advanced analytical techniques for better understanding and managing the disease. + +# 2.2. PD datasets + +Existing datasets for PD detection and analysis fall into three primary categories: speech datasets, handwriting datasets, and body movement datasets. Speech datasets [23, 24] encompass sustained phonation tests, and voice signal analysis can be employed for PD diagnosis. Handwriting datasets [25-27], gathered in image form, primarily focus on subjects' writing of letters or shapes to assess the impact of PD on fine motor control. Body movement datasets are particularly valuable for PD analysis, as they characterize the motor symptoms and motor dysfunction commonly associated with PD progression and can be used to formulate treatment plans. Body movement data are commonly collected using expensive wearable sensors that only capture a small number of body parts [13, 28-32]. Alternatively, video-based data collected by RGB cameras [20, 33-37] provide a contactless, cost-effective, and comprehensive approach to quantify motor disorders. However, while the UPDRS comprises several dozen motor activities, most video-based data collection is limited to one or two activities, usually gait or index finger tapping. Furthermore, these video datasets use only a single camera and thus are fundamentally limited in their precision when capturing fine-scale features of body movement, an inherently 3D process. Another significant limitation of these datasets is that they are not shared with the community. This makes replication difficult and precludes further algorithmic advances by other resesearchers. + +# 3. TULIP Dataset + +# 3.1. Clinical motivation + +The complexity of PD progression and symptomatology necessitates a high-resolution, multi-dimensional dataset of PD body kinematics. Current clinical assessment tools, such as the UPDRS, are coarse, subjective, and noisy. Studies have shown that the UPDRS score for a single exam can vary significantly between raters, with inter-rater Kappa values as low as 0.37, indicating minimal agreement between raters [38, 39]. This underscores a need for more objective, quantitative, and sensitive measures. + +To address this need, we require a rich data streams that accurately capture the complexities and nuances of PD motor symptoms. Research has shown that wearables [40] and force plates [22] can track some PD symptoms, but they struggle to align with UPDRS guidelines and face hospital implementation challenges [41, 42]. Video recordings, however, can efficiently capture a broad range + +![](images/a2e3bbfb6c3f9ddf0254a5dd1f55da5e1cb321922ba322849aa559de62884b58.jpg) +Figure 2. Overview of the TULIP Dataset. In this figure, subject faces are occluded for privacy, and * denotes there are separate left and right activity recordings. In addition to multi-view video recordings, TULIP contains UPDRS scores for each subject. We tracked 2D and 3D poses to build UPDRS classifiers. + +of PD movement symptoms and can be more flexibly integrated into clinical settings. For TULIP, we recorded activities using a multi-camera setup to more accurately capture complex 3D movement features. For single-camera recordings, 2D features are measured as projections onto a plane. The resulting 2D features thus vary with perspective and depth, reducing the resolution with which articulated 3D bodies moving freely through 3D space can be measured. We designed TULIP to address this issue. + +Marker-based multi-camera motion capture systems are traditionally used to track 3D movement quantities, and recent studies have used such systems for movement disorder analyses [43-45]. But requiring that patients wear markers, together with the size of the motion capture camera array, is not feasible for routine PD assessments. A markerless diagnostic system could easily be deployed in the clinic, or even at home, where it could be used to support decision making or monitor disease progression longitudinally. Furthermore, it is easier to capture facial + +expressions, a key element of PD assessment, with markerless systems. Markerless 3D analysis could also provide the necessary granularity to identify subtle but clinically significant changes in motor function, thereby enhancing the sensitivity of clinical trials. + +# 3.2. Dataset structure + +For the TULIP dataset, we enrolled 15 subjects, comprising PD patients and healthy individuals. Ten were clinically diagnosed with PD, while five had no prior PD diagnosis. Subjects demographics are detailed in Supplementary Table 2. The study was approved and conducted in accordance with the ethical guidelines set by the Duke Institutional Review Board. + +TULIP focuses on observational Part III: Motor Examination UPDRS components (Figure 2). With clinician guidance, we recorded 25 total activities from which UPDRS scores could be derived, including unilateral tasks, such + +as guided hand movement, index finger tapping, and stepping, and bilateral tasks, such as arm straightening and gait. Twenty-one of these activities are described in the UPDRS guidelines, and 4, including standing on one leg and finger tapping using all fingers, were added to encompass a wider range of body movements, on the advice of clinicians. See Supplementary Section 3 for full activity list and other paradigm details. + +# 3.3. Video collection + +To comprehensively capture the activities enumerated in the UPDRS, we set up the system to record a $6.3 \times 3$ meter hexagonal space fully capturing subjects in all 6 cameras (1920 × 1200 pixels; 80 fps; Basler acA1920-155uc). Cameras were synchronized via GPIO cables controlled by campy [46]. Intrinsic and extrinsic camera parameters were fit before and after each recording session. + +# 3.4. UPDRS Labeling + +For our study, we recruited three clinical expert neurologists (2 professors, 1 MD fellow) in PD diagnosis and management with over 30 years of PD experience combined. These professionals reviewed the video recordings and scored subjects according to UPDRS guidelines, and further made an overall judgment about whether the subject had PD or was healthy. The evaluations were conducted without access to the subjects' personal or medical history information to ensure unbiased scoring. From the videos, we obtained 29 distinct UPDRS Part III scores summarizing motor function, including bilateral Hand movements, Kinetic body tremor, Resting tremor amplitude arms and elbows, Facial expressions, Gait, Posture, Dyskinesias, and more. The mapping between the 25 recorded activities and the 29 UPDRS scores is delineated in Supplementary Section 3. We derived 'gold-standard' UPDRS labels from clinician scores using majority voting, as per prior research [35]. + +# 4. Analysis Methods + +# 4.1. Pose Estimation + +We used TULIP to test automated UPDRS scoring models for finger tapping and gait given features extracted from 2D and 3D pose sequences. For pose estimation, we utilized Mediapipe [47] and MMPose [48], which tracked complementary sets of body keypoints on study subjects without fine-tuning. We used MediaPipe to track 21 keypoints on each hand for finger tapping, and we used MMpose (Halpe26 configuration) to track 26 keypoints on the legs, arms, trunk, and head for gait. To estimate 3D poses, we triangulated 2D poses using the direct linear transformation algorithm. We further improved 3D tracking by interpolating over keypoint outliers and smoothing. We evaluated pose tracking precision via comparison to human keypoint + +annotation and via body segment length consistency. Mean errors relative to manual annotations were $21\mathrm{mm}$ for finger tapping and $56\mathrm{mm}$ for gait, corresponding to reprojection errors of 22 and 16 pixels, respectively. Tracked body segment lengths were also stable, with standard deviations below $20\mathrm{mm}$ across all frames in each recording (Supplementary Section 4). + +# 4.2. Features for disease classification + +To assess the utility of our dataset and establish benchmarks for future studies, we trained classifiers to distinguish PD from healthy individuals and predict UPDRS score, using spatiotemporal kinematic features extracted from 2D and 3D pose sequences. In this study, we focused on only two activities: index finger tapping and gait, both of which are frequently analyzed in PD research [32, 35, 40, 49-52]. We designed our feature set to enable accurate classification but also include clinical variables that are commonly utilized when scoring UPDRS manually. + +![](images/6515f1d61926b6159c0e3f0808e6ffdb09376299c195be0d58589d8b6039dc98.jpg) +Figure 3. Overview of feature extraction from finger tapping and gait recordings. + +# 4.2.1 Index finger tapping + +During the index finger tapping task, subjects were instructed to perform tapping movements solely with their thumb and index finger while keeping their palm oriented forward. This activity yields valuable data on tremor, rigidity, and the stability of hand posture [35, 53]. For feature extraction, we first segmented pose traces based on tapping events, measuring the angle between the thumb and index fingertip, as well as the distance between them (Figure 3). + +Tapping events were detected as minima in finger-thumb distance traces, i.e. the frames where the index fingertip and thumb made contact. + +After detecting the tapping events, the dataset was segmented into windows, each encompassing a sequence of 10 tapping events, in accordance with the UPDRS finger tapping criteria [5] and aligned with methodologies employed in prior research [35]. For each tapping event, separately for 2D and 3D feature sets, we computed three features from the angle $\theta_{f} = \cos^{-1}\left(\frac{\vec{w t}_{f} \cdot \vec{w i}_{f}}{\|\vec{w t}_{f}\| \|\vec{w i}_{f}\|}\right)$ , where $f$ is the frame, $\vec{w t}$ represents the vector from the wrist to the thumb tip, and $\vec{w i}$ denotes the vector from the wrist to the index fingertip: angular speed, angular acceleration, amplitude. For the 3D features, we also computed additional quantities, as keypoint coordinates are in real-world metric units. We utilized the 3D distance between the thumb tip and index fingertip to get speed relative features, as issues with contact vigor can be a PD indicator [35, 54]. These included wrist cartesian coordinates and finger velocities (opening velocity, closing velocity). Finger velocity, in particular, is regarded as a critical measure for PD detection [54-56]. We then computed seven summary statistics within each tapping window (arithmetic mean, minimum, maximum, median, interquartile range (IQR), standard deviation, and Shannon entropy) and used only these statistics as our final feature set for our baseline PD classifiers. These statistics provide concise summaries of the feature distribution, and previous studies have shown they can be used to discriminate between PD severity levels in finger tapping recordings [35, 57-59]. + +To enhance this final feature set, we incorporated seven additional metrics capturing dynamic aspects of tapping motions: tapping period, tapping frequency, aperiodicity, number of interruptions, number of freezing events, longest freezing event duration, and complexity of fitting periods. The methodologies for calculating these additional features are detailed in the Supplementary Section 4.2.1. In total, this resulted in a suite of 28 features for 2D and 49 features for 3D poses. + +# 4.2.2 Gait + +Gait analysis is pivotal in tracking the progression of PD and evaluating symptom severity. Characteristically, PD patients display aberrant gait patterns marked by shortened stride length, heightened stride variability, and a reduction in walking speed [51, 60]. Prior research has documented disparities in spatiotemporal and kinematic gait features between PD patients and healthy subjects [52, 61], but these studies have only utilized a limited range of features. Using TULIP, we assessed whether a more comprehensive set of features could enhance gait analysis, and whether gait assessments benefit from 3D vs. 2D measurements. + +During the gait measurement task, subjects began from a standing position and then walked 6 meters, turned around, and returned to the starting point. We recorded subject gait for one minute, resulting in several walking cycles for each subject. The methodology for extracting gait features is depicted in Figure 3. Since subjects slowed down and turned around at the end of the recording space, to extract gait features we first segmented each linear walking bout. We then identified the precise timepoints of heel-strike and toe-off events to isolate individual strides. Heel-strike and toe-off events were identified as the extrema in the anterior-posterior trajectories of the heels and toes relative to the hip [62]. We computed two types of gait features: spatiotemporal features, such as double support time (the time that both feet are touching the ground during a gait cycle) and step length; and kinematic features, such as knee angles, etc. For example, we calculated double support time from one foot's heel-strike to the other's toe-off using $t_{ds} = \left(f_{t_{j+1}} - f_{h_j}\right) / F_s$ (different feet), where $f_{t_{j+1}}$ denotes the frame number of the $(j+1)$ th toe-off event, $f_{h_j}$ denotes frame number of the $j$ th heel-strike event, and $F_s$ denotes the frame rate. Features derived from 2D poses were calculated in a single side-camera view selected to best capture linear gait. For 3D poses, we derived these features after projecting the poses onto the plane best capturing subject displacement during a linear walking bout. This projection reduces perspective ambiguities, an advantage unique to 3D data. As with finger tapping, we assembled our final feature set for modeling by computing 7 summary statistics, incorporating 'CV (Coefficient of Variation)' instead of 'entropy' of these gait features. The selection of CV in gait analysis was based on its common usage in describing gait variability. Additionally, the amount of variability, such as CV in the movement pattern, has been emphasized in prior research [63]. + +# 5. Results + +# 5.1. Clinician agreement + +In our study, we quantified the consistency and reliability of clinical evaluations using the Intraclass Correlation Coefficient (ICC) and weighted Cohen's Kappa of UPDRS scores. The ICC for mean UPDRS score was consistent across clinicians for the complete study cohort, $\mathrm{ICC} = 0.92$ (95% CI [0.81, 0.97]). Within the PD patient group, the ICC for mean UPDRS score was 0.87 (95% CI [0.58, 0.97]), with the slight drop in agreement likely due to the increased heterogeneity of PD presentation. Pairwise weighted Kappa values (for ordinal variables) indicated similar levels of consistency for mean UPDRS (0.82 clinician 1 and 2; 0.81 clinician 1 and 3; 0.75 clinician 2 and 3 for the full cohort; 0.81, 0.80, 0.73 within the PD group, respectively). However, + +![](images/af09b01d7d0602217851bedcafdfea3c17a220321ca2fed0db27665bd55f5170.jpg) +Figure 4. Schematic of the classification pipeline incorporating LOSO and bias matrices. + +the reliability of specific UPDRS component scores was more variable. While some components, such as "Finger tapping - Left hand" (ICC = 0.89) and "Postural stability" (ICC = 0.90) were scored consistently, others exhibited significant discrepancies, such as "Kinetic tremor - Left hand" (ICC=0.4) and "Kinetic tremor - Right hand" (ICC=0.1). This underscores the inherent complexity of these evaluations and highlights the need for uniformity in clinical assessments. + +# 5.2. Disease classification + +# 5.2.1 Modeling details + +We used our feature set to train a suite of traditional machine learning models (SVM, Random Forest, AdaBoost, XGBoost, LightGBM, MLP) classifying ordinal UPDRS scores and the overall clinical PD/healthy diagnosis, establishing baseline benchmarks for TULIP. We chose these traditional models because it allowed us to compare to previous studies that have used them with more limited feature sets. Models were trained using a leave-one-subject-out (LOSO) validation scheme for small sample sizes [35, 64], with validation performance calculated using 'gold standard' clinician labels (majority voting over the group of 3 clinicians). For finger tapping, we integrated features from both left and right activities, in line with methodologies established by previous research [35]. To help account for clinician UPDRS scoring variability, our models were trained to predict each of the clinician UPDRS scores separately. These scores were then weighted by clinician bias matrices [65] (calculated only on training samples) and averaged to produce final predicted scores. Detailed modeling pipeline can be shown in Figure 4. Before training, we also refined the feature sets for each activity by removing highly correlated variables. We calculated Pearson's correlation coefficient for all pairs of features and removed one of the pair from the dataset if $r > 0.85$ . + +# 5.2.2 Index finger tapping + +Finger tapping analyses captured consistent periodicity in all subjects, with variability in frequency and amplitude. After removing highly correlated features, we arrived at a diverse set of 20 features for 3D analyses: angular speed (median), number of freezing, angular speed (entropy), wrist movement (maximum), amplitude (median), aperiodicity, amplitude (entropy), angular acceleration (minimum), amplitude (standard deviation), angular speed (minimum), complexity of fitting periods, amplitude (minimum), angular acceleration (maximum), amplitude (IQR), longest freezing duration, wrist movement (mean), wrist movement (minimum), closing velocity (maximum), wrist movement (median), opening velocity (minimum), and 14 features for 2D analyses: longest freezing duration, angular speed (entropy), angular speed (maximum), number of freezing events, amplitude (minimum), aperiodicity, amplitude (median), angular speed (median), amplitude (IQR), angular acceleration (max), amplitude (entropy), complexity of fitting periods, angular speed (minimum), angular acceleration (minimum). + +As outlined above, we tested a suite of classifiers on two different tasks: UPDRS score prediction (0-4, ordinal scale) and overall PD/Healthy diagnosis. Among the models tested, a simple MLP neural network using 3D features was most accurate on both UPDRS score prediction (0.69 F1) and overall diagnosis (0.87 F1) as shown in Table 1. We tested previously published approaches for PD classification, including a ResNet50 CNN trained directly on 3D pose sequences and LightGBM using 2D features, on the TULIP finger tapping activity, but they did not perform as well as the 3D feature MLP. In general, models using 2D finger tapping features tended to perform poorly, supporting the case for 3D PD data acquisition and analysis. The poor performance of the ResNet50 pose sequence classifier could stem from the relatively small size of the TULIP dataset. + +![](images/f54ded90c8122b31f3bf07c1c4bc5ae04b5a1e1c11e1e075732a3cf17653d919.jpg) + +![](images/b1abbe6dc0264a4c7d52d4c59f407b1d2b8212f85ee4f8487e9033ea42a3017e.jpg) + +![](images/fc8c9c23c360909fa9799eef30d088724fa041f5af7e8a7ad2d19beac81ab716.jpg) + +![](images/87089bed4e94a084ab1e3b255f120a4d8b6b9441dfd27a2189a5c1b8f6dba71d.jpg) + +![](images/fb842eb2b4e4685d01be6bee33d48a8018da2fb09c4caf8c735268439ae4ea06.jpg) + +![](images/11ab7406c894ef1e85e1485438109ecc75e0cd6c925d769aef6ad346ddff4898.jpg) +Figure 5. Gait feature extraction. Top: Left knee angles with real images. Bottom: Leg angles with 3D mesh models. Traces show clear differences associated with increasing impairment (higher UPDRS score means more impaired). Points marked as skyblue are occluded in a specific viewpoint. Shaded areas in feature graphs indicate $95\%$ confidence intervals. + +![](images/4741f4b18fcaa71a9aa0976f2a7071059196d9d0164238878e39994d6e8fc9ed.jpg) + +![](images/41846325063f228daa4e22508fe9ebd9aa54aa7bcad9cc33d882868ec0cdeab6.jpg) + +![](images/4de52501b256b6dbbfc7199bab3c3de8e48a8c95a5c99db2bf282f5c09c89f28.jpg) + +![](images/97459782ae9b0c327f94634aedfc383f7cfe72600b1473deaf36a374565fd711.jpg) + +Note that Mehta et al. [17] use this ResNet50 CNN for sit-stand UPDRS classification and ensemble the ResNet50 with a Graph Convolutional Network. Here we tested the ResNet50 on its own. + +We quantified clinician assessment performance by comparing each clinician to the gold standard consensus labels for each subject. On UPDRS scoring, clinician raters were not perfect, achieving F1 scores between 0.7 and 0.87, with our top performing model commensurate with the least accurate rater. For binary PD/Healthy diagnoses, our model, which was trained only on finger tapping features, was nearly as good as one clinician (0.87 F1 for the model, 0.90 F1 for Clinician C). Note that clinicians assigned their overall PD/Healthy scores after viewing all 25 activities. + +# 5.2.3 Gait + +Our analysis pipeline captures kinematics continuously and accurately at a millisecond resolution. In traces of joint angle (Figure 5), we found evidence for restricted ranges of motion in PD patients, corroborating findings from prior clinical studies [61, 66]. As with finger tapping, we removed highly correlated features, reducing the set to 57 features for 3D and 45 for 2D. The refined feature set captured a variety of step, stride, and cadence measures. Further details on these features are available in Supplementary Section 5.2.2. + +As with finger tapping, we tested classifiers for gait UP- + +DRS score and overall PD/Healthy diagnosis. For gait, a simple Random Forest model using 3D features demonstrated the best performance on UPDRS score prediction (0.72 F1). Again, models using 2D features underperformed. A similar pattern was observed for PD/Healthy classification, although we note that on this task our model performed slightly less well than a previously published Random Forest approach using a smaller set of 3D features. + +Clinician UPDRS scoring of gait was less accurate (between 0.67 and 0.73 F1), compared to consensus gold standard labels, than for finger tapping, and our model using 3D features outperformed both Clinician B and Clinician C. This finding suggests that our model, to some extent, effectively mitigates biases present in medical assessments, thus aiding in the diagnostic process and contributing to a more objective scoring system. All approaches performed less well on gait than finger tapping, potentially due to the increased natural heterogeneity of gait. The performance of all tested models can be found in Supplementary Section 5. + +# 6. Discussion + +Here we present the TULIP dataset—a pioneering resource in PD research—as well as baselines for automated PD symptom scoring. TULIP stands out as the first dataset utilizes a markerless multi-camera setup to document both PD patients and healthy subjects, encompassing a comprehensive range of motor examination tasks from the UPDRS, which is the clinical standard for PD assessment. Despite + +
TargetTaskFeaturesModelF1 scoreReference
UPDRS PredictionFingertap2D featuresMLP0.37Ours
LightGBM0.36(Islam et al., 2023)
3D featuresMLP0.69Ours
3D posesResNet500.30(Mehta et al., 2021)
VideosClinician A0.80
Clinician B0.70
Clinician C0.87
Gait2D featuresRandom Forest0.41Ours
3D featuresRandom Forest0.72
Random Forest0.34(Rehman et al. 2019)
3D posesResNet500.27(Mehta et al., 2021)
VideosClinician A0.73
Clinician B0.67
Clinician C0.67
PD/Healthy PredictionFingertap2D featuresMLP0.69Ours
LightGBM0.77(Islam et al., 2023)
3D featuresMLP0.87Ours
3D posesResNet500.64(Mehta et al., 2021)
Gait2D featuresSVM0.72Ours
3D featuresRandom Forest0.72
Random Forest0.73(Rehman et al. 2019)
3D posesResNet500.59(Mehta et al., 2021)
All 25 activitiesVideosClinician A1.00
Clinician B0.93
Clinician C0.90
+ +Table 1. UPDRS scores and holistic PD vs. healthy predictions using gold-standard labels. For UPDRS classification, F1 score denotes weighted F1 score. Clinicians' holistic decisions were not activity-specific and thus are not separated by activity in the PD vs. healthy comparison. + +the extensive adoption of UPDRS, current clinical scoring for PD remains semi-subjective, heavily reliant on clinician experience and their interpretation of the rating scale. By curating the TULIP dataset and demonstrating its utility for behavioral analysis in PD patients, we've provided a valuable tool that can propel forward the integration of advanced analytical techniques in clinical settings for PD. + +We used TULIP to benchmark models predicting UPDRS scores for gait and finger tapping directly from video recordings. For these baselines, we chose to use temporal, spatial, and kinematic feature extraction approaches to align with previous work and enable more direct comparisons between our 3D approach and existing 2D methods. On TULIP, we found that 3D features significantly outperformed 2D features, underscoring the importance of 3D video measurements. We expect UPDRS score predictions to improve in the future, for instance via networks fit directly to 3D pose time series. Nevertheless, our models achieved results that were commensurate with the accuracy of a clinician subset, demonstrating the potential of 3D computer vision approaches to enable automated and scalable PD assessments. + +We note several current limitations and future directions. While the number of recorded subjects in TULIP exceeds that of some popular 3D multi-camera 3D human datasets, such as Human 3.6M, our cohort is still relatively small, especially considering the heterogeneity of PD. TULIP also contains relatively little demographic diversity, which could hinder generalization to more representative patient populations. In addition, our baseline TULIP analyses did not incorporate data from all of the activities we recorded. + +TULIP was also collected in a large room that does not reflect the constraints of a typical clinical exam environment, and in the future it will be important to adapt the behavioral paradigm and acquisition hardware to smaller spaces. + +In the future, we plan to augment our dataset by incorporating a larger and more diverse set of subjects, coupled with an expanded range of activities aligned closer to activities of daily living. We are also excited by the potential of unsupervised analysis of TULIP data, which could reveal new quantitative signatures of PD independent of UPDRS scoring. Long term, we hope to expand data collection to include movement disorders other than PD, with the overall aim to unearth novel digital behavioral biomarkers. + +# 7. Conclusion + +TULIP is the first dataset of its kind to comprehensively record the UPDRS motor examination, providing high-resolution multi-view 3D readouts together with annotations from multiple clinical experts. By leveraging the TULIP dataset, we have shown that UPDRS classifiers incorporating 3D features markedly outperform those based on 2D features, and we have identified interpretable distinctions in movement variables between PD patients and healthy controls in finger tapping and gait activities. TULIP is poised to advance Parkinson's research by aiding the development of more precise, objective, and scalable diagnostics, fostering more effective patient management and more successful treatment discovery. + +# Acknowledgements + +We acknowledge support from the Duke University Department of Neurology and the Duke Gilhuly Accelerator Fund. TWD is an advisor to dannce.ai and Higgs Boson Health. Also we would like to thank Timothy Lindsey, Louis DeFrate, Anshuman Sabath, Brian Lerner, Pranav Manjunath, Sophie Shi, Tianqing Li, and Joshua Wu for helping us recording the data, and three clinicians for labeling the data. We also thank Lisa Gauger and Nicole Calakos for logistical support. + +# References + +[1] World Health Organization. Launch of who's parkinson disease technical brief. https://www.who.int/news/item/14-06-2022-launch-of-who-s-parkinson-disease-technical-brief, 2022.1 +[2] Kenn Freddy Pedersen, Jan Petter Larsen, and Dag Aarsland. Validation of the unified parkinson's disease rating scale (up- drs) section i as a screening and diagnostic instrument for apathy in patients with parkinson's disease. Parkinsonism & related disorders, 14(3):183-186, 2008. 1 +[3] Neil Ramsay, Angus D Macleod, Guido Alves, Marta Camacho, Lars Forsgren, Rachael A Lawson, Jodi Maple-Grødem, Ole-Bjørn Tysnes, Caroline H Williams-Gray, Ali + +son J Yarnall, et al. Validation of a updrs-/mds-updrs-based definition of functional dependency for parkinson's disease. Parkinsonism & Related Disorders, 76:49-53, 2020. +[4] Katsuki Eguchi, Ichigaku Takigawa, Shinichi Shirai, Ikuko Takahashi-Iwata, Masaaki Matsushima, Takahiro Kano, Hiroaki Yaguchi, and Ichiro Yabe. Gait video-based prediction of unified parkinson's disease rating scale score: a retrospective study. BMC neurology, 23(1):358, 2023. 1 +[5] Christopher G Goetz, Barbara C Tilley, Stephanie R Shaftman, Glenn T Stebbins, Stanley Fahn, Pablo Martinez-Martin, Werner Poewe, Cristina Sampaio, Matthew B Stern, Richard Dodel, et al. Movement disorder society-sponsored revision of the unified parkinson's disease rating scale (mdsupdrs): scale presentation and clinimetric testing results. Movement disorders: official journal of the Movement Disorder Society, 23(15):2129-2170, 2008. 1, 5 +[6] Stefan Williams, David Wong, Jane E Alty, and Samuel D Relton. Parkinsonian hand or clinician's eye? finger tap bradykinesia interrater reliability for 21 movement disorder experts. Journal of Parkinson's Disease, (Preprint):1-12, 2023. 1 +[7] Sullafa Kadura and Ruth B Schneider. Moving beyond alerts: Electronic health record strategies to improve inpatient parkinson's disease care. Parkinsonism & Related Disorders, 2023. 1 +[8] Jing Zhang. Mining imaging and clinical data with machine learning approaches for the diagnosis and early detection of parkinson's disease. NPJ Parkinson's disease, 8(1):13, 2022. 1 +[9] Ronald B Postuma, Daniela Berg, Matthew Stern, Werner Poewe, C Warren Olanow, Wolfgang Oertel, José Obeso, Kenneth Marek, Irene Litvan, Anthony E Lang, et al. Mds clinical diagnostic criteria for parkinson's disease. Movement disorders, 30(12):1591-1601, 2015. 1 +[10] Sanghee Moon, Hyun-Je Song, Vibhash D Sharma, Kelly E Lyons, Rajesh Pahwa, Abiodun E Akinwuntan, and Hannes Devos. Classification of parkinson's disease and essential tremor based on balance and gait characteristics from wearable motion sensors via machine learning techniques: a data-driven approach. Journal of neuroengineering and rehabilitation, 17:1-8, 2020. 2 +[11] Andrea Bandini, Sia Rezaei, Diego L Guarín, Madhura Kulkarni, Derrick Lim, Mark I Boulos, Lorne Zinman, Yana Yunusova, and Babak Taati. A new dataset for facial motion analysis in individuals with neurological disorders. IEEE Journal of Biomedical and Health Informatics, 25(4):1111-1119, 2020. 2 +[12] Aakash Kaku, Kangning Liu, Avinash Parnandi, Haresh Rengaraj Rajamohan, Kannan Venkataramanan, Anita Venkatesan, Audre Wirtanen, Natasha Pandit, Heidi Schambra, and Carlos Fernandez-Granda. Strokerehab: A benchmark dataset for sub-second action identification. Advances in Neural Information Processing Systems, 35:1671-1684, 2022. 2 +[13] Marta Isabel ASN Ferreira, Fabio Augusto Barbieri, Vinicius Christianini Moreno, Tiago Penedo, and João Manuel RS Tavares. Machine learning models for parkinson's disease detection and stage classification based on + +spatial-temporal gait parameters. Gait & Posture, 98:49-55, 2022. 2 +[14] Ferdous Wahid, Rezaul K Begg, Chris J Hass, Saman Halgamuge, and David C Ackland. Classification of parkinson's disease gait using spatial-temporal gait features. IEEE journal of biomedical and health informatics, 19(6):1794-1802, 2015. 2 +[15] Yanyi Zhang, Ming Kong, Tianqi Zhao, Wenchen Hong, Qiang Zhu, and Fei Wu. Adhd intelligent auxiliary diagnosis system based on multimodal information fusion. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4494-4496, 2020. 2 +[16] Mahmoud Seifallahi, Afsoon Hasani Mehraban, James E Galvin, and Behnaz Ghoraani. Alzheimer's disease detection using comprehensive analysis of timed up and go test via kinect v. 2 camera and machine learning. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 30:1589-1600, 2022. 2 +[17] Deval Mehta, Umar Asif, Tian Hao, Erhan Bilal, Stefan Von Cavallar, Stefan Harrer, and Jeffrey Rogers. Towards automated and marker-less parkinson disease assessment: predicting updrs scores using sit-stand videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3841–3849, 2021. 2, 7 +[18] Renfei Sun, Kun Hu, Kaylena A Ehgoetz Martens, Markus Hagenbuchner, Ah Chung Tsoi, Mohammed Bennamoun, Simon JG Lewis, and Zhiyong Wang. Higher order polynomial transformer for fine-grained freezing of gait detection. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2 +[19] Mark Endo, Kathleen L Poston, Edith V Sullivan, Li Fei-Fei, Kilian M Pohl, and Ehsan Adeli. Gaitforemer: Self-supervised pre-training of transformers via human motion forecasting for few-shot gait impairment severity estimation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 130-139. Springer, 2022. 2 +[20] Mandy Lu, Qingyu Zhao, Kathleen L Poston, Edith V Sullivan, Adolf Pfefferbaum, Marian Shahid, Maya Katz, Leila Montaser Kouhsari, Kevin Schulman, Arnold Milstein, et al. Quantifying parkinson's disease motor severity under uncertainty using mds-updrs videos. Medical image analysis, 73:102179, 2021. 2 +[21] Mehrbakhsh Nilashi, Othman Ibrahim, Hossein Ahmadi, Leila Shahmoradi, and Mohammadreza Farahmand. A hybrid intelligent system for the prediction of parkinson's disease progression using machine learning techniques. Biocybernetics and Biomedical Engineering, 38(1):1-15, 2018. 2 +[22] Trevor Exley, Sarah Moudy, Rita M Patterson, Joonghyun Kim, and Mark V Albert. Predicting updrs motor symptoms in individuals with parkinson's disease from force plates using machine learning. IEEE Journal of Biomedical and Health Informatics, 26(7):3486-3494, 2022. 2 +[23] Max Little, Patrick McSharry, Eric Hunter, Jennifer Spielman, and Lorraine Ramig. Suitability of dysphonia measurements for telemonitoring of parkinson's disease. Nature Precedings, pages 1-1, 2008. 2 + +[24] Athanasios Tsanas, Max Little, Patrick McSharry, and Lorraine Ramig. Accurate telemonitoring of parkinson's disease progression by non-invasive speech tests. Nature Precedings, pages 1-1, 2009. 2 +[25] Clayton R Pereira, Danilo R Pereira, Francisco A Silva, Joao P Masieiro, Silke AT Weber, Christian Hook, and João P Papa. A new computer vision-based approach to aid the diagnosis of parkinson's disease. Computer Methods and Programs in Biomedicine, 136:79-88, 2016. 2 +[26] Clayton R Pereira, Silke AT Weber, Christian Hook, Gustavo H Rosa, and Joao P Papa. Deep learning-aided parkinson's disease diagnosis from handwritten dynamics. In 2016 29th SIBGRAPI conference on graphics, patterns and images (SIBGRAPI), pages 340–346. IEEE, 2016. +[27] Catherine Taleb, Maha Khachab, Chafic Mokbel, and Laurence Likforman-Sulem. Feature selection for an improved parkinson's disease identification based on handwriting. In 2017 1st International Workshop on Arabic Script Analysis and Recognition (ASAR), pages 52-56. IEEE, 2017. 2 +[28] Charalamos Sotirakis, Zi Su, Maksymilian A Brzezicki, Niall Conway, Lionel Tarassenko, James J FitzGerald, and Chrystalina A Antoniades. Identification of motor progression in parkinson's disease using wearable sensors and machine learning. npj Parkinson's Disease, 9(1):142, 2023. 2 +[29] Chariklia Chatzaki, Vasileios Skaramagkas, Nikolaos Tachos, Georgios Christodoulakis, Evangelia Maniadi, Zinovia Kefalopoulou, Dimitrios I Fotiadis, and Manolis Tsiknakis. The smart-insole dataset: Gait analysis using wearable sensors with a focus on elderly and parkinson's patients. Sensors, 21(8):2821, 2021. +[30] Marc Bachlin, Meir Plotnik, Daniel Roggen, Inbal Maidan, Jeffrey M Hausdorff, Nir Giladi, and Gerhard Troster. Wearable assistant for parkinson's disease patients with the freezing of gait symptom. IEEE Transactions on Information Technology in Biomedicine, 14(2):436-446, 2009. +[31] Jens Barth, Jochen Klucken, Patrick Kugler, Thomas Kammerer, Ralph Steidl, Jürgen Winkler, Joachim Hornegger, and Björn Eskofier. Biometric and mobile gait analysis for early diagnosis and therapy monitoring in parkinson's disease. In 2011 annual international conference of the IEEE engineering in medicine and biology society, pages 868-871. IEEE, 2011. +[32] Caroline Ribeiro De Souza, Runfeng Miao, Julia Ávila De Oliveira, Andrea Cristina De Lima-Pardini, Debora Fragoso De Campos, Carla Silva-Batista, Luis Teixeira, Solaiman Shokur, Bouri Mohamed, and Daniel Boari Coelho. A public data set of videos, inertial measurement unit, and clinical scales of freezing of gait in individuals with parkinson's disease during a turning-in-place task. Frontiers in Neuroscience, 16:832463, 2022. 2, 4 +[33] Mandy Lu, Kathleen Poston, Adolf Pfefferbaum, Edith V Sullivan, Li Fei-Fei, Kilian M Pohl, Juan Carlos Niebles, and Ehsan Adeli. Vision-based estimation of mds-updrs gait scores for assessing parkinson's disease motor severity. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part III 23, pages 637-647. Springer, 2020. 2 + +[34] Andrea Sabo, Sina Mehdizadeh, Kimberley-Dale Ng, Andrea Iaboni, and Babak Taati. Assessment of parkinsonian gait in older adults with dementia via human pose tracking in video data. Journal of neuroengineering and rehabilitation, 17(1):1-10, 2020. +[35] Md Saiful Islam, Wasifur Rahman, Abdelrahman Abdelkader, Phillip T Yang, Sangwu Lee, Jamie L Adams, Ruth B Schneider, E Dorsey, and Ehsan Hoque. Using ai to measure parkinson's disease severity at home. npj digital medicine, 2023. 4, 5, 6 +[36] Andrea Sabo, Carolina Gorodetsky, Alfonso Fasano, Andrea Iaboni, and Babak Taati. Concurrent validity of zero instrumented walkway and video-based gait features in adults with parkinson's disease. IEEE Journal of Translational Engineering in Health and Medicine, 10:1-11, 2022. +[37] Samuel Rupprechter, Gareth Morinan, Yuwei Peng, Thomas Foltynie, Krista Sibley, Rimona S Weil, Louise-Ann Leyland, Fahd Baig, Francesca Morgante, Ro'ee Gilron, et al. A clinically interpretable computer-vision based method for quantifying gait in parkinson's disease. Sensors, 21(16):5437, 2021. 2 +[38] Erin Smith, John Bertoni, Danish Bhatti, and Diego Torres-Russotto. Is the updrs a reliable tool for detecting the worse side in parkinson's disease? (1584). *Neurology*, 94:1584, 2020. 2 +[39] Ashwani Jha, Elisa Menozzi, Rebecca Oyekan, Anna Latorre, Eoin Mulroy, Sebastian R Schreglmann, Cosmin Stamate, Ioannis Daskalopoulos, Stefan Kueppers, Marco Luchini, et al. The cloudupdrs smartphone software in parkinson's study: cross-validation against blinded human raters. npj Parkinson's Disease, 6(1):36, 2020. 2 +[40] Delaram Safarpour, Marian L Dale, Vrutangkumar V Shah, Lauren Talman, Patricia Carlson-Kuhta, Fay B Horak, and Martina Mancini. Surrogates for rigidity and pigd mds-updrs subscores using wearable sensors. Gait & Posture, 91:186-191, 2022. 2, 4 +[41] Erika Rovini, Carlo Maremmani, and Filippo Cavallo. How wearable sensors can support parkinson's disease diagnosis and treatment: a systematic review. Frontiers in neuroscience, 11:288959, 2017. 2 +[42] Mercedes Barrachina-Fernández, Ana María Maitín, Carmen Sánchez-Avila, and Juan Pablo Romero. Wearable technology to detect motor fluctuations in parkinson's disease patients: current state and challenges. Sensors, 21(12):4188, 2021. 2 +[43] Balasundaram Kadirvelu, Constantinos Gavriel, Sathiji Nageshwaran, Jackson Ping Kei Chan, Suran Nethisinghe, Stavros Athanasopoulos, Valeria Ricotti, Thomas Voit, Paola Giunti, Richard Festenstein, et al. A wearable motion capture suit and machine learning predict disease progression in friedreich's ataxia. Nature Medicine, 29(1):86-94, 2023. 3 +[44] Benjamin Filtjens, Pieter Ginis, Alice Nieuwboer, Peter Slaets, and Bart Vanrumste. Automated freezing of gait assessment with marker-based motion capture and multistage spatial-temporal graph convolutional neural networks. Journal of NeuroEngineering and Rehabilitation, 19(1):48, 2022. + +[45] Muhammad Hassan Khan, Manuel Schneider, Muhammad Shahid Farid, and Marcin Grzegorzek. Detection of infantile movement disorders in video data using deformable part-based model. Sensors, 18(10):3202, 2018. 3 +[46] K Severson. campy. https://github.com/ksseverson57/campy, 2022.4.1 +[47] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for perceiving and processing reality. In Third Workshop on Computer Vision for AR/VR at IEEE Computer Vision and Pattern Recognition (CVPR) 2019, 2019. 4 +[48] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.4 +[49] Zhilin Guo, Weiqi Zeng, Taidong Yu, Yan Xu, Yang Xiao, Xuebing Cao, and Zhiguo Cao. Vision-based finger tapping test in patients with parkinson's disease via spatial-temporal 3d hand pose estimation. IEEE Journal of Biomedical and Health Informatics, 26(8):3848-3859, 2022. 4 +[50] Ahnjili ZhuParris, Eva Thijssen, Willem O Elzinga, Soma Makai-Boloni, Wessel Kraaij, Geert J Groeneveld, and Robert J Doll. Treatment detection and movement disorder society-unified parkinson's disease rating scale, part iii estimation using finger tapping tasks. Movement Disorders, 38(10):1795-1805, 2023. +[51] Jeffrey M Hausdorff. Gait dynamics in parkinson's disease: common and distinct behavior among stride length, gait variability, and fractal-like scaling. Chaos: An Interdisciplinary Journal of Nonlinear Science, 19(2), 2009. 5 +[52] Michele Pistacchi, Manuela Gioulis, Flavio Sanson, Ennio De Giovanni, Giuseppe Filippi, Francesca Rossetto, and Sandro Zambito Marsala. Gait analysis and clinical correlations in early parkinson's disease. *Functional neurology*, 32(1):28, 2017. 4, 5 +[53] Tianze Yu, Kye Won Park, Martin J McKeown, and Z Jane Wang. Clinically informed automated assessment of finger tapping videos in parkinson's disease. Sensors, 23(22):9149, 2023. 4 +[54] Kye Won Park, Eun-Jae Lee, Jun Seong Lee, Jinhoon Jeong, Nari Choi, Sungyang Jo, Mina Jung, Ja Yeon Do, Dong-Wha Kang, June-Goo Lee, et al. Machine learning-based automatic rating for cardinal symptoms of parkinson disease. Neurology, 96(13):e1761-e1769, 2021. 5 +[55] R Agostino, A Curra, M Giovannelli, N Modugno, M Manfredi, and A Berardelli. Impairment of individual finger movements in parkinson's disease. *Movement Disorders*, 18(5):560-565, 2003. +[56] Eva Thijssen, Soma Makai-Boloni, Emilie van Brummelen, Jonas den Heijer, Yalcin Yavuz, Robert-Jan Doll, and Geert Jan Groeneveld. A placebo-controlled study to assess the sensitivity of finger tapping to medication effects in parkinson's disease. *Movement Disorders Clinical Practice*, 9(8):1074-1084, 2022. 5 +[57] M Yokoe, R Okuno, T Hamasaki, Y Kurachi, K Akazawa, and S Sakoda. Opening velocity, a novel parameter, for fin + +ger tapping test in patients with parkinson's disease. Parkinsonism & related disorders, 15(6):440-444, 2009. 5 +[58] Junjie Li, Huaiyu Zhu, Yun Pan, Haotian Wang, Zhidong Cen, Dehao Yang, and Wei Luo. Three-dimensional pattern features in finger tapping test for patients with parkinson's disease. In 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 3676-3679. IEEE, 2020. +[59] Noreen Akram, Haoxuan Li, Aaron Ben-Joseph, Caroline Budu, David A Gallagher, Jonathan P Bestwick, Anette Schrag, Alastair J Noyce, and Cristina Simonet. Developing and assessing a new web-based tapping test for measuring distal movement in parkinson's disease: a distal finger tapping test. Scientific reports, 12(1):386, 2022. 5 +[60] Kevin J Brusse, Sandy Zimdars, Kathryn R Zalewski, and Teresa M Steffen. Testing functional performance in people with parkinson disease. Physical therapy, 85(2):134-141, 2005. 5 +[61] Olumide Sofuwa, Alice Nieuwboer, Kaat Desloovere, Annemarie Willems, Fabienne Chavret, and Ilse Jonkers. Quantitative gait analysis in parkinson's disease: comparison with a healthy control group. Archives of physical medicine and rehabilitation, 86(5):1007-1013, 2005. 5, 7 +[62] JA Zeni Jr, JG Richards, and JS2384115 Higginson. Two simple methods for determining gait events during treadmill and overground walking using kinematic data. Gait & posture, 27(4):710-714, 2008. 5 +[63] Christopher K Rhea and Adam W Kiefer. Patterned variability in gait behavior: How can it be measured and what does it mean. Gait biometrics: basic patterns, role of neurological disorders and effects of physical activity, pages 17-44, 2014. 5 +[64] Martin Patrick Pauli, Constantin Pohl, and Martin Golz. Balanced leave-one-subject-out cross-validation for microsleep classification. Current Directions in Biomedical Engineering, 7(2):147-150, 2021. 6 +[65] Ryutaro Tanno, Ardavan Saeedi, Swami Sankaranarayanan, Daniel C Alexander, and Nathan Silberman. Learning from noisy labels by regularized estimation of annotator confusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11244–11253, 2019. 6 +[66] Gwyn N Lewis, Winston D Byblow, and Sharon E Walt. Stride length regulation in parkinson's disease: the use of extrinsic, visual cues. Brain, 123(10):2077-2090, 2000. 7 +[67] CJ Hawley, N Fineberg, AG Roberts, D Baldwin, A Sahadevan, and V Sharman. The use of the simpson angus scale for the assessment of movement disorder: a training guide. International Journal of Psychiatry in Clinical Practice, 7(4):349-2257, 2003. 2 +[68] K Muller, CK Hemelrijk, J Westerweel, and DSW Tam. Calibration of multiple cameras for large-scale experiments using a freely moving calibration target. *Experiments in Fluids*, 61(1):7, 2020. 1 +[69] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for + +social motion capture. In Proceedings of the IEEE international conference on computer vision, pages 3334-3342, 2015. 7 \ No newline at end of file diff --git a/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/images.zip b/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4c632d3ae1e4348c0f0907fc3429f97132b7339b --- /dev/null +++ b/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5dd56b4fc017a6bc589d79fda963bacfbf1dff0abf877610d32e120ad42a1916 +size 478613 diff --git a/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/layout.json b/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3ed9e7c2da2a451765facdc420ef0f41755e164f --- /dev/null +++ b/2024/TULIP_ Multi-camera 3D Precision Assessment of Parkinson's Disease/layout.json @@ -0,0 +1,7022 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 515, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 515, + 119 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 515, + 119 + ], + "type": "text", + "content": "TULIP: Multi-camera 3D Precision Assessment of Parkinson's Disease" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 142, + 523, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 142, + 523, + 185 + ], + "spans": [ + { + "bbox": [ + 70, + 142, + 523, + 185 + ], + "type": "text", + "content": "Kyungdo Kim" + }, + { + "bbox": [ + 70, + 142, + 523, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 142, + 523, + 185 + ], + "type": "text", + "content": ", Sihan Lyu" + }, + { + "bbox": [ + 70, + 142, + 523, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 142, + 523, + 185 + ], + "type": "text", + "content": ", Sneha Mantri" + }, + { + "bbox": [ + 70, + 142, + 523, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 142, + 523, + 185 + ], + "type": "text", + "content": ", Timothy W. Dunn" + }, + { + "bbox": [ + 70, + 142, + 523, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 142, + 523, + 185 + ], + "type": "text", + "content": " \nDepartment of " + }, + { + "bbox": [ + 70, + 142, + 523, + 185 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 70, + 142, + 523, + 185 + ], + "type": "text", + "content": "Biomedical Engineering, " + }, + { + "bbox": [ + 70, + 142, + 523, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 142, + 523, + 185 + ], + "type": "text", + "content": "Neurology" + }, + { + "bbox": [ + 70, + 142, + 523, + 185 + ], + "type": "inline_equation", + "content": "\\}" + }, + { + "bbox": [ + 70, + 142, + 523, + 185 + ], + "type": "text", + "content": ", Duke University, Durham, NC, USA \n{kyungdo.kim, sihan.lyu, sneha.mantri, timothy.dunn}@duke.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 238, + 290, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 549 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 549 + ], + "type": "text", + "content": "Parkinson's disease (PD) is a devastating movement disorder accelerating in global prevalence, but a lack of precision symptom measurement has made the development of effective therapies challenging. The Unified Parkinson's Disease Rating Scale (UPDRS) is the gold standard for assessing motor symptom severity, yet its manual scoring criteria are vague and subjective, resulting in coarse and noisy clinical assessments. Machine learning approaches have the potential to modernize PD symptom assessments by making them more quantitative, objective, and scalable. However, the lack of benchmark video datasets for PD motor exams hinders model development. Here, we introduce the TULIP dataset to bridge this gap. TULIP emphasizes precision and comprehensiveness, comprising multi-view video recordings (6 cameras) of 25 UPDRS motor exam activities, together with ratings by 3 clinical experts, in a cohort of Parkinson's patients and healthy controls. The multi-view recordings enable 3D reconstructions of body movement that better capture disease signatures than more conventional 2D methods. Using the dataset, we establish a baseline model for predicting UPDRS scores from 3D poses, illustrating how existing diagnostics could be automated. Looking ahead, TULIP could aid the development of new precision diagnostics that transcend UPDRS scores, providing a deeper understanding of PD and its potential treatments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 573, + 128, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 128, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 128, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "text", + "content": "Parkinson's disease (PD) is a progressive neurodegenerative disorder leading to a spectrum of motor impairments that differ between individuals and worsen over time. In the last quarter-century, the global prevalence of PD has increased two-fold and continues to rise, now impacting over 8.5 million individuals [1]. Despite decades of research, no cure has emerged, and the focus has shifted towards managing symptoms to improve quality of life." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "The current standard for PD clinical assessment, the Movement Disorder Society-Sponsored Revision of the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 212, + 545, + 411 + ], + "blocks": [ + { + "bbox": [ + 308, + 212, + 545, + 411 + ], + "lines": [ + { + "bbox": [ + 308, + 212, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 308, + 212, + 545, + 411 + ], + "type": "image", + "image_path": "49c569d8bc840333529ab118ef18738df658f4fffb958be4025afcc046429620.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 419, + 547, + 475 + ], + "lines": [ + { + "bbox": [ + 305, + 419, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 419, + 547, + 475 + ], + "type": "text", + "content": "Figure 1. Landscape of Parkinson's disease datasets. The black icons indicate which activities were recorded. Previous studies recorded a small number of activities, typically either walking, finger tapping, or sitting/standing. TULIP includes 25 different activities, capturing a wider set of movements and postures." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "content": "Unified Parkinson's Disease Rating Scale (UPDRS), relies on a manual scoring system that can lead to imprecise diagnoses. While the scale is validated [2-4], examiners score subjects coarsely based on vague guidelines. For instance, 'global spontaneity' is rated on an ordinal scale from '1 (Slight)' to '4 (Severe)' [5]. This manual scoring is labor-intensive and a source of variability, and it precludes identification of fine-scale motor pathologies. Numerous studies have identified significant inter-rater differences in UPDRS scoring, with disagreements as large as " + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "inline_equation", + "content": "53\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "content": " [6]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": "Recent advances in machine learning have greatly enhanced healthcare workflows and patient outcomes across many clinical domains, but applying machine learning approaches to PD remains challenging due to a shortage of training and benchmark datasets. Some approaches have explored modeling PD using electronic health record [7] or neuroimaging data [8], but ultimately such datasets rely on clinical diagnosis of motor symptoms [9]. While small" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22551" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 155 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 155 + ], + "type": "text", + "content": "subsets of symptoms, such as tremor, can be measured precisely with wearable devices [10], video recordings can capture the wide diversity of PD symptoms, which occur on multiple spatiotemporal scales. However, while there are some video-based datasets for other neurological disorders [11, 12], there are currently no benchmark video datasets for PD." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 286, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 286, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 286, + 287 + ], + "type": "text", + "content": "Previous studies have reported PD video analyses, but these studies have been restricted to single-camera recordings, which inherently limit kinematic measurement precision. Precision can be enhanced by analyzing 3D data extracted from multi-camera recordings, but thus far 3D data has only been collected with the aid of markers (i.e., motion capture) [13, 14]. Previous video analyses have also been restricted to a very small subset of UPDRS activities as shown in Figure 1. Past approaches have thus been unable to capture the full spectrum of PD motor deficits, limiting their sensitivity and utility." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 287, + 286, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 286, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 286, + 395 + ], + "type": "text", + "content": "Here, we introduce TULIP (Three-dimensional (3D) Understanding and Learning of Impairments in Parkinson's disease) to address this unmet need. To our knowledge, TULIP is the world's first multi-camera video dataset of PD motor exams to: 1) be available for benchmarking, 2) include all of the body movement tests of UPDRS (TULIP excludes the speech test), and 3) enable markerless 3D kinematic analysis. TULIP also nearly triples the acquisition frame rate of past studies (Figure 1)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 395, + 286, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 395, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 286, + 491 + ], + "type": "text", + "content": "Using TULIP, we show that UPDRS classifiers built with 3D features significantly outperform those with 2D features. We also identify interpretable distinctions in movement variables between PD patients and healthy controls. We chose the name TULIP, a nod to the floral emblem of PD research and advocacy, to symbolize our goal for this dataset, to foster transformative new machine learning approaches for PD understanding and treatment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 506, + 135, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 135, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 135, + 517 + ], + "type": "text", + "content": "2. Previous work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 526, + 275, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 275, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 275, + 539 + ], + "type": "text", + "content": "2.1. Machine learning for neurological disorders" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "content": "Machine learning is widely used for analyzing motor symptoms in neurological disorders like stroke [12]. Zhang et al. used cameras to detect ADHD by tracking child facial expressions and limb movements [15]. Seifallahi et al. employed Kinect cameras for Alzheimer's gait analysis [16]. In PD, existing computational methods focus on sit-to-stand [17], gait [13, 18] and finger tapping activities [19, 20], which have been the easiest to measure with computer vision. Machine learning has also been pivotal in predicting PD progression and symptom severity. Nilashi et al. proposed a system for predicting UPDRS scores and making holistic decisions only on biomedical voice measurements [21]. Exley et al. used machine learning to build more objective UPDRS motor symptom scores by way of force" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "places [22]. These efforts highlight the evolving landscape of PD research, combining traditional clinical assessments with advanced analytical techniques for better understanding and managing the disease." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 129, + 383, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 129, + 383, + 140 + ], + "spans": [ + { + "bbox": [ + 306, + 129, + 383, + 140 + ], + "type": "text", + "content": "2.2. PD datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 148, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 545, + 469 + ], + "type": "text", + "content": "Existing datasets for PD detection and analysis fall into three primary categories: speech datasets, handwriting datasets, and body movement datasets. Speech datasets [23, 24] encompass sustained phonation tests, and voice signal analysis can be employed for PD diagnosis. Handwriting datasets [25-27], gathered in image form, primarily focus on subjects' writing of letters or shapes to assess the impact of PD on fine motor control. Body movement datasets are particularly valuable for PD analysis, as they characterize the motor symptoms and motor dysfunction commonly associated with PD progression and can be used to formulate treatment plans. Body movement data are commonly collected using expensive wearable sensors that only capture a small number of body parts [13, 28-32]. Alternatively, video-based data collected by RGB cameras [20, 33-37] provide a contactless, cost-effective, and comprehensive approach to quantify motor disorders. However, while the UPDRS comprises several dozen motor activities, most video-based data collection is limited to one or two activities, usually gait or index finger tapping. Furthermore, these video datasets use only a single camera and thus are fundamentally limited in their precision when capturing fine-scale features of body movement, an inherently 3D process. Another significant limitation of these datasets is that they are not shared with the community. This makes replication difficult and precludes further algorithmic advances by other resesearchers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 483, + 399, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 399, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 399, + 495 + ], + "type": "text", + "content": "3. TULIP Dataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 503, + 418, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 418, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 418, + 514 + ], + "type": "text", + "content": "3.1. Clinical motivation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 522, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 629 + ], + "type": "text", + "content": "The complexity of PD progression and symptomatology necessitates a high-resolution, multi-dimensional dataset of PD body kinematics. Current clinical assessment tools, such as the UPDRS, are coarse, subjective, and noisy. Studies have shown that the UPDRS score for a single exam can vary significantly between raters, with inter-rater Kappa values as low as 0.37, indicating minimal agreement between raters [38, 39]. This underscores a need for more objective, quantitative, and sensitive measures." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "To address this need, we require a rich data streams that accurately capture the complexities and nuances of PD motor symptoms. Research has shown that wearables [40] and force plates [22] can track some PD symptoms, but they struggle to align with UPDRS guidelines and face hospital implementation challenges [41, 42]. Video recordings, however, can efficiently capture a broad range" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22552" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 545, + 418 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 545, + 418 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 545, + 418 + ], + "type": "image", + "image_path": "a2e3bbfb6c3f9ddf0254a5dd1f55da5e1cb321922ba322849aa559de62884b58.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 427, + 546, + 460 + ], + "lines": [ + { + "bbox": [ + 46, + 427, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 546, + 460 + ], + "type": "text", + "content": "Figure 2. Overview of the TULIP Dataset. In this figure, subject faces are occluded for privacy, and * denotes there are separate left and right activity recordings. In addition to multi-view video recordings, TULIP contains UPDRS scores for each subject. We tracked 2D and 3D poses to build UPDRS classifiers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 482, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 589 + ], + "type": "text", + "content": "of PD movement symptoms and can be more flexibly integrated into clinical settings. For TULIP, we recorded activities using a multi-camera setup to more accurately capture complex 3D movement features. For single-camera recordings, 2D features are measured as projections onto a plane. The resulting 2D features thus vary with perspective and depth, reducing the resolution with which articulated 3D bodies moving freely through 3D space can be measured. We designed TULIP to address this issue." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 590, + 288, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 288, + 708 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 288, + 708 + ], + "type": "text", + "content": "Marker-based multi-camera motion capture systems are traditionally used to track 3D movement quantities, and recent studies have used such systems for movement disorder analyses [43-45]. But requiring that patients wear markers, together with the size of the motion capture camera array, is not feasible for routine PD assessments. A markerless diagnostic system could easily be deployed in the clinic, or even at home, where it could be used to support decision making or monitor disease progression longitudinally. Furthermore, it is easier to capture facial" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 482, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 545, + 541 + ], + "type": "text", + "content": "expressions, a key element of PD assessment, with markerless systems. Markerless 3D analysis could also provide the necessary granularity to identify subtle but clinically significant changes in motor function, thereby enhancing the sensitivity of clinical trials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 563, + 410, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 410, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 410, + 574 + ], + "type": "text", + "content": "3.2. Dataset structure" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 582, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 664 + ], + "type": "text", + "content": "For the TULIP dataset, we enrolled 15 subjects, comprising PD patients and healthy individuals. Ten were clinically diagnosed with PD, while five had no prior PD diagnosis. Subjects demographics are detailed in Supplementary Table 2. The study was approved and conducted in accordance with the ethical guidelines set by the Duke Institutional Review Board." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "TULIP focuses on observational Part III: Motor Examination UPDRS components (Figure 2). With clinician guidance, we recorded 25 total activities from which UPDRS scores could be derived, including unilateral tasks, such" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22553" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "as guided hand movement, index finger tapping, and stepping, and bilateral tasks, such as arm straightening and gait. Twenty-one of these activities are described in the UPDRS guidelines, and 4, including standing on one leg and finger tapping using all fingers, were added to encompass a wider range of body movements, on the advice of clinicians. See Supplementary Section 3 for full activity list and other paradigm details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 175, + 144, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 144, + 187 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 144, + 187 + ], + "type": "text", + "content": "3.3. Video collection" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 194, + 287, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 287, + 278 + ], + "type": "text", + "content": "To comprehensively capture the activities enumerated in the UPDRS, we set up the system to record a " + }, + { + "bbox": [ + 46, + 194, + 287, + 278 + ], + "type": "inline_equation", + "content": "6.3 \\times 3" + }, + { + "bbox": [ + 46, + 194, + 287, + 278 + ], + "type": "text", + "content": " meter hexagonal space fully capturing subjects in all 6 cameras (1920 × 1200 pixels; 80 fps; Basler acA1920-155uc). Cameras were synchronized via GPIO cables controlled by campy [46]. Intrinsic and extrinsic camera parameters were fit before and after each recording session." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 285, + 151, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 151, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 151, + 300 + ], + "type": "text", + "content": "3.4. UPDRS Labeling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 304, + 289, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 289, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 289, + 509 + ], + "type": "text", + "content": "For our study, we recruited three clinical expert neurologists (2 professors, 1 MD fellow) in PD diagnosis and management with over 30 years of PD experience combined. These professionals reviewed the video recordings and scored subjects according to UPDRS guidelines, and further made an overall judgment about whether the subject had PD or was healthy. The evaluations were conducted without access to the subjects' personal or medical history information to ensure unbiased scoring. From the videos, we obtained 29 distinct UPDRS Part III scores summarizing motor function, including bilateral Hand movements, Kinetic body tremor, Resting tremor amplitude arms and elbows, Facial expressions, Gait, Posture, Dyskinesias, and more. The mapping between the 25 recorded activities and the 29 UPDRS scores is delineated in Supplementary Section 3. We derived 'gold-standard' UPDRS labels from clinician scores using majority voting, as per prior research [35]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 520, + 154, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 154, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 154, + 533 + ], + "type": "text", + "content": "4. Analysis Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 539, + 145, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 145, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 145, + 551 + ], + "type": "text", + "content": "4.1. Pose Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": "We used TULIP to test automated UPDRS scoring models for finger tapping and gait given features extracted from 2D and 3D pose sequences. For pose estimation, we utilized Mediapipe [47] and MMPose [48], which tracked complementary sets of body keypoints on study subjects without fine-tuning. We used MediaPipe to track 21 keypoints on each hand for finger tapping, and we used MMpose (Halpe26 configuration) to track 26 keypoints on the legs, arms, trunk, and head for gait. To estimate 3D poses, we triangulated 2D poses using the direct linear transformation algorithm. We further improved 3D tracking by interpolating over keypoint outliers and smoothing. We evaluated pose tracking precision via comparison to human keypoint" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "annotation and via body segment length consistency. Mean errors relative to manual annotations were " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "21\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " for finger tapping and " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "56\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " for gait, corresponding to reprojection errors of 22 and 16 pixels, respectively. Tracked body segment lengths were also stable, with standard deviations below " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "20\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " across all frames in each recording (Supplementary Section 4)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 163, + 484, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 163, + 484, + 175 + ], + "spans": [ + { + "bbox": [ + 305, + 163, + 484, + 175 + ], + "type": "text", + "content": "4.2. Features for disease classification" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 301 + ], + "type": "text", + "content": "To assess the utility of our dataset and establish benchmarks for future studies, we trained classifiers to distinguish PD from healthy individuals and predict UPDRS score, using spatiotemporal kinematic features extracted from 2D and 3D pose sequences. In this study, we focused on only two activities: index finger tapping and gait, both of which are frequently analyzed in PD research [32, 35, 40, 49-52]. We designed our feature set to enable accurate classification but also include clinical variables that are commonly utilized when scoring UPDRS manually." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 311, + 545, + 541 + ], + "blocks": [ + { + "bbox": [ + 307, + 311, + 545, + 541 + ], + "lines": [ + { + "bbox": [ + 307, + 311, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 307, + 311, + 545, + 541 + ], + "type": "image", + "image_path": "6515f1d61926b6159c0e3f0808e6ffdb09376299c195be0d58589d8b6039dc98.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 549, + 545, + 572 + ], + "lines": [ + { + "bbox": [ + 305, + 549, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 305, + 549, + 545, + 572 + ], + "type": "text", + "content": "Figure 3. Overview of feature extraction from finger tapping and gait recordings." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 599, + 427, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 427, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 427, + 612 + ], + "type": "text", + "content": "4.2.1 Index finger tapping" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "During the index finger tapping task, subjects were instructed to perform tapping movements solely with their thumb and index finger while keeping their palm oriented forward. This activity yields valuable data on tremor, rigidity, and the stability of hand posture [35, 53]. For feature extraction, we first segmented pose traces based on tapping events, measuring the angle between the thumb and index fingertip, as well as the distance between them (Figure 3)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22554" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "Tapping events were detected as minima in finger-thumb distance traces, i.e. the frames where the index fingertip and thumb made contact." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 289, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 289, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 289, + 430 + ], + "type": "text", + "content": "After detecting the tapping events, the dataset was segmented into windows, each encompassing a sequence of 10 tapping events, in accordance with the UPDRS finger tapping criteria [5] and aligned with methodologies employed in prior research [35]. For each tapping event, separately for 2D and 3D feature sets, we computed three features from the angle " + }, + { + "bbox": [ + 46, + 109, + 289, + 430 + ], + "type": "inline_equation", + "content": "\\theta_{f} = \\cos^{-1}\\left(\\frac{\\vec{w t}_{f} \\cdot \\vec{w i}_{f}}{\\|\\vec{w t}_{f}\\| \\|\\vec{w i}_{f}\\|}\\right)" + }, + { + "bbox": [ + 46, + 109, + 289, + 430 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 109, + 289, + 430 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 109, + 289, + 430 + ], + "type": "text", + "content": " is the frame, " + }, + { + "bbox": [ + 46, + 109, + 289, + 430 + ], + "type": "inline_equation", + "content": "\\vec{w t}" + }, + { + "bbox": [ + 46, + 109, + 289, + 430 + ], + "type": "text", + "content": " represents the vector from the wrist to the thumb tip, and " + }, + { + "bbox": [ + 46, + 109, + 289, + 430 + ], + "type": "inline_equation", + "content": "\\vec{w i}" + }, + { + "bbox": [ + 46, + 109, + 289, + 430 + ], + "type": "text", + "content": " denotes the vector from the wrist to the index fingertip: angular speed, angular acceleration, amplitude. For the 3D features, we also computed additional quantities, as keypoint coordinates are in real-world metric units. We utilized the 3D distance between the thumb tip and index fingertip to get speed relative features, as issues with contact vigor can be a PD indicator [35, 54]. These included wrist cartesian coordinates and finger velocities (opening velocity, closing velocity). Finger velocity, in particular, is regarded as a critical measure for PD detection [54-56]. We then computed seven summary statistics within each tapping window (arithmetic mean, minimum, maximum, median, interquartile range (IQR), standard deviation, and Shannon entropy) and used only these statistics as our final feature set for our baseline PD classifiers. These statistics provide concise summaries of the feature distribution, and previous studies have shown they can be used to discriminate between PD severity levels in finger tapping recordings [35, 57-59]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 430, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 430, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 45, + 430, + 287, + 538 + ], + "type": "text", + "content": "To enhance this final feature set, we incorporated seven additional metrics capturing dynamic aspects of tapping motions: tapping period, tapping frequency, aperiodicity, number of interruptions, number of freezing events, longest freezing event duration, and complexity of fitting periods. The methodologies for calculating these additional features are detailed in the Supplementary Section 4.2.1. In total, this resulted in a suite of 28 features for 2D and 49 features for 3D poses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 560, + 99, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 99, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 99, + 571 + ], + "type": "text", + "content": "4.2.2 Gait" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "Gait analysis is pivotal in tracking the progression of PD and evaluating symptom severity. Characteristically, PD patients display aberrant gait patterns marked by shortened stride length, heightened stride variability, and a reduction in walking speed [51, 60]. Prior research has documented disparities in spatiotemporal and kinematic gait features between PD patients and healthy subjects [52, 61], but these studies have only utilized a limited range of features. Using TULIP, we assessed whether a more comprehensive set of features could enhance gait analysis, and whether gait assessments benefit from 3D vs. 2D measurements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 492 + ], + "type": "text", + "content": "During the gait measurement task, subjects began from a standing position and then walked 6 meters, turned around, and returned to the starting point. We recorded subject gait for one minute, resulting in several walking cycles for each subject. The methodology for extracting gait features is depicted in Figure 3. Since subjects slowed down and turned around at the end of the recording space, to extract gait features we first segmented each linear walking bout. We then identified the precise timepoints of heel-strike and toe-off events to isolate individual strides. Heel-strike and toe-off events were identified as the extrema in the anterior-posterior trajectories of the heels and toes relative to the hip [62]. We computed two types of gait features: spatiotemporal features, such as double support time (the time that both feet are touching the ground during a gait cycle) and step length; and kinematic features, such as knee angles, etc. For example, we calculated double support time from one foot's heel-strike to the other's toe-off using " + }, + { + "bbox": [ + 304, + 72, + 547, + 492 + ], + "type": "inline_equation", + "content": "t_{ds} = \\left(f_{t_{j+1}} - f_{h_j}\\right) / F_s" + }, + { + "bbox": [ + 304, + 72, + 547, + 492 + ], + "type": "text", + "content": " (different feet), where " + }, + { + "bbox": [ + 304, + 72, + 547, + 492 + ], + "type": "inline_equation", + "content": "f_{t_{j+1}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 492 + ], + "type": "text", + "content": " denotes the frame number of the " + }, + { + "bbox": [ + 304, + 72, + 547, + 492 + ], + "type": "inline_equation", + "content": "(j+1)" + }, + { + "bbox": [ + 304, + 72, + 547, + 492 + ], + "type": "text", + "content": "th toe-off event, " + }, + { + "bbox": [ + 304, + 72, + 547, + 492 + ], + "type": "inline_equation", + "content": "f_{h_j}" + }, + { + "bbox": [ + 304, + 72, + 547, + 492 + ], + "type": "text", + "content": " denotes frame number of the " + }, + { + "bbox": [ + 304, + 72, + 547, + 492 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 72, + 547, + 492 + ], + "type": "text", + "content": "th heel-strike event, and " + }, + { + "bbox": [ + 304, + 72, + 547, + 492 + ], + "type": "inline_equation", + "content": "F_s" + }, + { + "bbox": [ + 304, + 72, + 547, + 492 + ], + "type": "text", + "content": " denotes the frame rate. Features derived from 2D poses were calculated in a single side-camera view selected to best capture linear gait. For 3D poses, we derived these features after projecting the poses onto the plane best capturing subject displacement during a linear walking bout. This projection reduces perspective ambiguities, an advantage unique to 3D data. As with finger tapping, we assembled our final feature set for modeling by computing 7 summary statistics, incorporating 'CV (Coefficient of Variation)' instead of 'entropy' of these gait features. The selection of CV in gait analysis was based on its common usage in describing gait variability. Additionally, the amount of variability, such as CV in the movement pattern, has been emphasized in prior research [63]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 517, + 359, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 359, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 359, + 529 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 538, + 423, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 423, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 423, + 551 + ], + "type": "text", + "content": "5.1. Clinician agreement" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "In our study, we quantified the consistency and reliability of clinical evaluations using the Intraclass Correlation Coefficient (ICC) and weighted Cohen's Kappa of UPDRS scores. The ICC for mean UPDRS score was consistent across clinicians for the complete study cohort, " + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{ICC} = 0.92" + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": " (95% CI [0.81, 0.97]). Within the PD patient group, the ICC for mean UPDRS score was 0.87 (95% CI [0.58, 0.97]), with the slight drop in agreement likely due to the increased heterogeneity of PD presentation. Pairwise weighted Kappa values (for ordinal variables) indicated similar levels of consistency for mean UPDRS (0.82 clinician 1 and 2; 0.81 clinician 1 and 3; 0.75 clinician 2 and 3 for the full cohort; 0.81, 0.80, 0.73 within the PD group, respectively). However," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22555" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 547, + 230 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 230 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 230 + ], + "type": "image", + "image_path": "af09b01d7d0602217851bedcafdfea3c17a220321ca2fed0db27665bd55f5170.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 237, + 457, + 248 + ], + "lines": [ + { + "bbox": [ + 135, + 237, + 457, + 248 + ], + "spans": [ + { + "bbox": [ + 135, + 237, + 457, + 248 + ], + "type": "text", + "content": "Figure 4. Schematic of the classification pipeline incorporating LOSO and bias matrices." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 270, + 289, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 289, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 289, + 376 + ], + "type": "text", + "content": "the reliability of specific UPDRS component scores was more variable. While some components, such as \"Finger tapping - Left hand\" (ICC = 0.89) and \"Postural stability\" (ICC = 0.90) were scored consistently, others exhibited significant discrepancies, such as \"Kinetic tremor - Left hand\" (ICC=0.4) and \"Kinetic tremor - Right hand\" (ICC=0.1). This underscores the inherent complexity of these evaluations and highlights the need for uniformity in clinical assessments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 387, + 167, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 167, + 398 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 167, + 398 + ], + "type": "text", + "content": "5.2. Disease classification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 406, + 151, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 151, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 151, + 418 + ], + "type": "text", + "content": "5.2.1 Modeling details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": "We used our feature set to train a suite of traditional machine learning models (SVM, Random Forest, AdaBoost, XGBoost, LightGBM, MLP) classifying ordinal UPDRS scores and the overall clinical PD/healthy diagnosis, establishing baseline benchmarks for TULIP. We chose these traditional models because it allowed us to compare to previous studies that have used them with more limited feature sets. Models were trained using a leave-one-subject-out (LOSO) validation scheme for small sample sizes [35, 64], with validation performance calculated using 'gold standard' clinician labels (majority voting over the group of 3 clinicians). For finger tapping, we integrated features from both left and right activities, in line with methodologies established by previous research [35]. To help account for clinician UPDRS scoring variability, our models were trained to predict each of the clinician UPDRS scores separately. These scores were then weighted by clinician bias matrices [65] (calculated only on training samples) and averaged to produce final predicted scores. Detailed modeling pipeline can be shown in Figure 4. Before training, we also refined the feature sets for each activity by removing highly correlated variables. We calculated Pearson's correlation coefficient for all pairs of features and removed one of the pair from the dataset if " + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "inline_equation", + "content": "r > 0.85" + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 270, + 426, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 426, + 282 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 426, + 282 + ], + "type": "text", + "content": "5.2.2 Index finger tapping" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 294, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 547, + 533 + ], + "type": "text", + "content": "Finger tapping analyses captured consistent periodicity in all subjects, with variability in frequency and amplitude. After removing highly correlated features, we arrived at a diverse set of 20 features for 3D analyses: angular speed (median), number of freezing, angular speed (entropy), wrist movement (maximum), amplitude (median), aperiodicity, amplitude (entropy), angular acceleration (minimum), amplitude (standard deviation), angular speed (minimum), complexity of fitting periods, amplitude (minimum), angular acceleration (maximum), amplitude (IQR), longest freezing duration, wrist movement (mean), wrist movement (minimum), closing velocity (maximum), wrist movement (median), opening velocity (minimum), and 14 features for 2D analyses: longest freezing duration, angular speed (entropy), angular speed (maximum), number of freezing events, amplitude (minimum), aperiodicity, amplitude (median), angular speed (median), amplitude (IQR), angular acceleration (max), amplitude (entropy), complexity of fitting periods, angular speed (minimum), angular acceleration (minimum)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 534, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 548, + 714 + ], + "type": "text", + "content": "As outlined above, we tested a suite of classifiers on two different tasks: UPDRS score prediction (0-4, ordinal scale) and overall PD/Healthy diagnosis. Among the models tested, a simple MLP neural network using 3D features was most accurate on both UPDRS score prediction (0.69 F1) and overall diagnosis (0.87 F1) as shown in Table 1. We tested previously published approaches for PD classification, including a ResNet50 CNN trained directly on 3D pose sequences and LightGBM using 2D features, on the TULIP finger tapping activity, but they did not perform as well as the 3D feature MLP. In general, models using 2D finger tapping features tended to perform poorly, supporting the case for 3D PD data acquisition and analysis. The poor performance of the ResNet50 pose sequence classifier could stem from the relatively small size of the TULIP dataset." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22556" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 161, + 186 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 161, + 186 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 161, + 186 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 161, + 186 + ], + "type": "image", + "image_path": "f54ded90c8122b31f3bf07c1c4bc5ae04b5a1e1c11e1e075732a3cf17653d919.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 169, + 72, + 262, + 185 + ], + "blocks": [ + { + "bbox": [ + 169, + 72, + 262, + 185 + ], + "lines": [ + { + "bbox": [ + 169, + 72, + 262, + 185 + ], + "spans": [ + { + "bbox": [ + 169, + 72, + 262, + 185 + ], + "type": "image", + "image_path": "b1abbe6dc0264a4c7d52d4c59f407b1d2b8212f85ee4f8487e9033ea42a3017e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 269, + 72, + 329, + 185 + ], + "blocks": [ + { + "bbox": [ + 269, + 72, + 329, + 185 + ], + "lines": [ + { + "bbox": [ + 269, + 72, + 329, + 185 + ], + "spans": [ + { + "bbox": [ + 269, + 72, + 329, + 185 + ], + "type": "image", + "image_path": "fc8c9c23c360909fa9799eef30d088724fa041f5af7e8a7ad2d19beac81ab716.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 338, + 72, + 381, + 185 + ], + "blocks": [ + { + "bbox": [ + 338, + 72, + 381, + 185 + ], + "lines": [ + { + "bbox": [ + 338, + 72, + 381, + 185 + ], + "spans": [ + { + "bbox": [ + 338, + 72, + 381, + 185 + ], + "type": "image", + "image_path": "87089bed4e94a084ab1e3b255f120a4d8b6b9441dfd27a2189a5c1b8f6dba71d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 392, + 76, + 541, + 185 + ], + "blocks": [ + { + "bbox": [ + 392, + 76, + 541, + 185 + ], + "lines": [ + { + "bbox": [ + 392, + 76, + 541, + 185 + ], + "spans": [ + { + "bbox": [ + 392, + 76, + 541, + 185 + ], + "type": "image", + "image_path": "fb842eb2b4e4685d01be6bee33d48a8018da2fb09c4caf8c735268439ae4ea06.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 186, + 164, + 307 + ], + "blocks": [ + { + "bbox": [ + 48, + 186, + 164, + 307 + ], + "lines": [ + { + "bbox": [ + 48, + 186, + 164, + 307 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 164, + 307 + ], + "type": "image", + "image_path": "11ab7406c894ef1e85e1485438109ecc75e0cd6c925d769aef6ad346ddff4898.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 316, + 547, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 316, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 547, + 350 + ], + "type": "text", + "content": "Figure 5. Gait feature extraction. Top: Left knee angles with real images. Bottom: Leg angles with 3D mesh models. Traces show clear differences associated with increasing impairment (higher UPDRS score means more impaired). Points marked as skyblue are occluded in a specific viewpoint. Shaded areas in feature graphs indicate " + }, + { + "bbox": [ + 46, + 316, + 547, + 350 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 316, + 547, + 350 + ], + "type": "text", + "content": " confidence intervals." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 173, + 186, + 260, + 306 + ], + "blocks": [ + { + "bbox": [ + 173, + 186, + 260, + 306 + ], + "lines": [ + { + "bbox": [ + 173, + 186, + 260, + 306 + ], + "spans": [ + { + "bbox": [ + 173, + 186, + 260, + 306 + ], + "type": "image", + "image_path": "4741f4b18fcaa71a9aa0976f2a7071059196d9d0164238878e39994d6e8fc9ed.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 272, + 186, + 324, + 304 + ], + "blocks": [ + { + "bbox": [ + 272, + 186, + 324, + 304 + ], + "lines": [ + { + "bbox": [ + 272, + 186, + 324, + 304 + ], + "spans": [ + { + "bbox": [ + 272, + 186, + 324, + 304 + ], + "type": "image", + "image_path": "41846325063f228daa4e22508fe9ebd9aa54aa7bcad9cc33d882868ec0cdeab6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 343, + 187, + 380, + 302 + ], + "blocks": [ + { + "bbox": [ + 343, + 187, + 380, + 302 + ], + "lines": [ + { + "bbox": [ + 343, + 187, + 380, + 302 + ], + "spans": [ + { + "bbox": [ + 343, + 187, + 380, + 302 + ], + "type": "image", + "image_path": "4de52501b256b6dbbfc7199bab3c3de8e48a8c95a5c99db2bf282f5c09c89f28.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 391, + 193, + 541, + 301 + ], + "blocks": [ + { + "bbox": [ + 391, + 193, + 541, + 301 + ], + "lines": [ + { + "bbox": [ + 391, + 193, + 541, + 301 + ], + "spans": [ + { + "bbox": [ + 391, + 193, + 541, + 301 + ], + "type": "image", + "image_path": "97459782ae9b0c327f94634aedfc383f7cfe72600b1473deaf36a374565fd711.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 370, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 370, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 370, + 287, + 418 + ], + "type": "text", + "content": "Note that Mehta et al. [17] use this ResNet50 CNN for sit-stand UPDRS classification and ensemble the ResNet50 with a Graph Convolutional Network. Here we tested the ResNet50 on its own." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 419, + 288, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 288, + 540 + ], + "type": "text", + "content": "We quantified clinician assessment performance by comparing each clinician to the gold standard consensus labels for each subject. On UPDRS scoring, clinician raters were not perfect, achieving F1 scores between 0.7 and 0.87, with our top performing model commensurate with the least accurate rater. For binary PD/Healthy diagnoses, our model, which was trained only on finger tapping features, was nearly as good as one clinician (0.87 F1 for the model, 0.90 F1 for Clinician C). Note that clinicians assigned their overall PD/Healthy scores after viewing all 25 activities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 559, + 100, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 100, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 100, + 571 + ], + "type": "text", + "content": "5.2.3 Gait" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 580, + 287, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 699 + ], + "type": "text", + "content": "Our analysis pipeline captures kinematics continuously and accurately at a millisecond resolution. In traces of joint angle (Figure 5), we found evidence for restricted ranges of motion in PD patients, corroborating findings from prior clinical studies [61, 66]. As with finger tapping, we removed highly correlated features, reducing the set to 57 features for 3D and 45 for 2D. The refined feature set captured a variety of step, stride, and cadence measures. Further details on these features are available in Supplementary Section 5.2.2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "As with finger tapping, we tested classifiers for gait UP-" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 370, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 545, + 466 + ], + "type": "text", + "content": "DRS score and overall PD/Healthy diagnosis. For gait, a simple Random Forest model using 3D features demonstrated the best performance on UPDRS score prediction (0.72 F1). Again, models using 2D features underperformed. A similar pattern was observed for PD/Healthy classification, although we note that on this task our model performed slightly less well than a previously published Random Forest approach using a smaller set of 3D features." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 466, + 546, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 546, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 546, + 598 + ], + "type": "text", + "content": "Clinician UPDRS scoring of gait was less accurate (between 0.67 and 0.73 F1), compared to consensus gold standard labels, than for finger tapping, and our model using 3D features outperformed both Clinician B and Clinician C. This finding suggests that our model, to some extent, effectively mitigates biases present in medical assessments, thus aiding in the diagnostic process and contributing to a more objective scoring system. All approaches performed less well on gait than finger tapping, potentially due to the increased natural heterogeneity of gait. The performance of all tested models can be found in Supplementary Section 5." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 609, + 375, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 375, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 375, + 621 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Here we present the TULIP dataset—a pioneering resource in PD research—as well as baselines for automated PD symptom scoring. TULIP stands out as the first dataset utilizes a markerless multi-camera setup to document both PD patients and healthy subjects, encompassing a comprehensive range of motor examination tasks from the UPDRS, which is the clinical standard for PD assessment. Despite" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22557" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 285, + 247 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 285, + 247 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 285, + 247 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 285, + 247 + ], + "type": "table", + "html": "
TargetTaskFeaturesModelF1 scoreReference
UPDRS PredictionFingertap2D featuresMLP0.37Ours
LightGBM0.36(Islam et al., 2023)
3D featuresMLP0.69Ours
3D posesResNet500.30(Mehta et al., 2021)
VideosClinician A0.80
Clinician B0.70
Clinician C0.87
Gait2D featuresRandom Forest0.41Ours
3D featuresRandom Forest0.72
Random Forest0.34(Rehman et al. 2019)
3D posesResNet500.27(Mehta et al., 2021)
VideosClinician A0.73
Clinician B0.67
Clinician C0.67
PD/Healthy PredictionFingertap2D featuresMLP0.69Ours
LightGBM0.77(Islam et al., 2023)
3D featuresMLP0.87Ours
3D posesResNet500.64(Mehta et al., 2021)
Gait2D featuresSVM0.72Ours
3D featuresRandom Forest0.72
Random Forest0.73(Rehman et al. 2019)
3D posesResNet500.59(Mehta et al., 2021)
All 25 activitiesVideosClinician A1.00
Clinician B0.93
Clinician C0.90
", + "image_path": "ec4ea9efb330b91d107ffe42999828d1fb274978d77ff11ecdb3ab6491de54a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 256, + 287, + 312 + ], + "lines": [ + { + "bbox": [ + 46, + 256, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 287, + 312 + ], + "type": "text", + "content": "Table 1. UPDRS scores and holistic PD vs. healthy predictions using gold-standard labels. For UPDRS classification, F1 score denotes weighted F1 score. Clinicians' holistic decisions were not activity-specific and thus are not separated by activity in the PD vs. healthy comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 338, + 286, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 286, + 422 + ], + "type": "text", + "content": "the extensive adoption of UPDRS, current clinical scoring for PD remains semi-subjective, heavily reliant on clinician experience and their interpretation of the rating scale. By curating the TULIP dataset and demonstrating its utility for behavioral analysis in PD patients, we've provided a valuable tool that can propel forward the integration of advanced analytical techniques in clinical settings for PD." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 423, + 286, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 286, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 286, + 602 + ], + "type": "text", + "content": "We used TULIP to benchmark models predicting UPDRS scores for gait and finger tapping directly from video recordings. For these baselines, we chose to use temporal, spatial, and kinematic feature extraction approaches to align with previous work and enable more direct comparisons between our 3D approach and existing 2D methods. On TULIP, we found that 3D features significantly outperformed 2D features, underscoring the importance of 3D video measurements. We expect UPDRS score predictions to improve in the future, for instance via networks fit directly to 3D pose time series. Nevertheless, our models achieved results that were commensurate with the accuracy of a clinician subset, demonstrating the potential of 3D computer vision approaches to enable automated and scalable PD assessments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "We note several current limitations and future directions. While the number of recorded subjects in TULIP exceeds that of some popular 3D multi-camera 3D human datasets, such as Human 3.6M, our cohort is still relatively small, especially considering the heterogeneity of PD. TULIP also contains relatively little demographic diversity, which could hinder generalization to more representative patient populations. In addition, our baseline TULIP analyses did not incorporate data from all of the activities we recorded." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "TULIP was also collected in a large room that does not reflect the constraints of a typical clinical exam environment, and in the future it will be important to adapt the behavioral paradigm and acquisition hardware to smaller spaces." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 121, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 228 + ], + "type": "text", + "content": "In the future, we plan to augment our dataset by incorporating a larger and more diverse set of subjects, coupled with an expanded range of activities aligned closer to activities of daily living. We are also excited by the potential of unsupervised analysis of TULIP data, which could reveal new quantitative signatures of PD independent of UPDRS scoring. Long term, we hope to expand data collection to include movement disorders other than PD, with the overall aim to unearth novel digital behavioral biomarkers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 239, + 378, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 378, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 378, + 251 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 258, + 545, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 415 + ], + "type": "text", + "content": "TULIP is the first dataset of its kind to comprehensively record the UPDRS motor examination, providing high-resolution multi-view 3D readouts together with annotations from multiple clinical experts. By leveraging the TULIP dataset, we have shown that UPDRS classifiers incorporating 3D features markedly outperform those based on 2D features, and we have identified interpretable distinctions in movement variables between PD patients and healthy controls in finger tapping and gait activities. TULIP is poised to advance Parkinson's research by aiding the development of more precise, objective, and scalable diagnostics, fostering more effective patient management and more successful treatment discovery." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 422, + 400, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 400, + 435 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 400, + 435 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "text", + "content": "We acknowledge support from the Duke University Department of Neurology and the Duke Gilhuly Accelerator Fund. TWD is an advisor to dannce.ai and Higgs Boson Health. Also we would like to thank Timothy Lindsey, Louis DeFrate, Anshuman Sabath, Brian Lerner, Pranav Manjunath, Sophie Shi, Tianqing Li, and Joshua Wu for helping us recording the data, and three clinicians for labeling the data. We also thank Lisa Gauger and Nicole Calakos for logistical support." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 559, + 365, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 559, + 365, + 571 + ], + "spans": [ + { + "bbox": [ + 306, + 559, + 365, + 571 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 310, + 579, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 310, + 579, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 579, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 310, + 579, + 545, + 622 + ], + "type": "text", + "content": "[1] World Health Organization. Launch of who's parkinson disease technical brief. https://www.who.int/news/item/14-06-2022-launch-of-who-s-parkinson-disease-technical-brief, 2022.1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 624, + 547, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 624, + 547, + 679 + ], + "spans": [ + { + "bbox": [ + 311, + 624, + 547, + 679 + ], + "type": "text", + "content": "[2] Kenn Freddy Pedersen, Jan Petter Larsen, and Dag Aarsland. Validation of the unified parkinson's disease rating scale (up- drs) section i as a screening and diagnostic instrument for apathy in patients with parkinson's disease. Parkinsonism & related disorders, 14(3):183-186, 2008. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 681, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 681, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 311, + 681, + 547, + 714 + ], + "type": "text", + "content": "[3] Neil Ramsay, Angus D Macleod, Guido Alves, Marta Camacho, Lars Forsgren, Rachael A Lawson, Jodi Maple-Grødem, Ole-Bjørn Tysnes, Caroline H Williams-Gray, Ali" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22558" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "son J Yarnall, et al. Validation of a updrs-/mds-updrs-based definition of functional dependency for parkinson's disease. Parkinsonism & Related Disorders, 76:49-53, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 106, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 106, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 287, + 162 + ], + "type": "text", + "content": "[4] Katsuki Eguchi, Ichigaku Takigawa, Shinichi Shirai, Ikuko Takahashi-Iwata, Masaaki Matsushima, Takahiro Kano, Hiroaki Yaguchi, and Ichiro Yabe. Gait video-based prediction of unified parkinson's disease rating scale score: a retrospective study. BMC neurology, 23(1):358, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 163, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 163, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 54, + 163, + 288, + 249 + ], + "type": "text", + "content": "[5] Christopher G Goetz, Barbara C Tilley, Stephanie R Shaftman, Glenn T Stebbins, Stanley Fahn, Pablo Martinez-Martin, Werner Poewe, Cristina Sampaio, Matthew B Stern, Richard Dodel, et al. Movement disorder society-sponsored revision of the unified parkinson's disease rating scale (mdsupdrs): scale presentation and clinimetric testing results. Movement disorders: official journal of the Movement Disorder Society, 23(15):2129-2170, 2008. 1, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 250, + 288, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 250, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 54, + 250, + 288, + 303 + ], + "type": "text", + "content": "[6] Stefan Williams, David Wong, Jane E Alty, and Samuel D Relton. Parkinsonian hand or clinician's eye? finger tap bradykinesia interrater reliability for 21 movement disorder experts. Journal of Parkinson's Disease, (Preprint):1-12, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 304, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 288, + 348 + ], + "type": "text", + "content": "[7] Sullafa Kadura and Ruth B Schneider. Moving beyond alerts: Electronic health record strategies to improve inpatient parkinson's disease care. Parkinsonism & Related Disorders, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 350, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 288, + 392 + ], + "type": "text", + "content": "[8] Jing Zhang. Mining imaging and clinical data with machine learning approaches for the diagnosis and early detection of parkinson's disease. NPJ Parkinson's disease, 8(1):13, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 393, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 393, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 53, + 393, + 288, + 448 + ], + "type": "text", + "content": "[9] Ronald B Postuma, Daniela Berg, Matthew Stern, Werner Poewe, C Warren Olanow, Wolfgang Oertel, José Obeso, Kenneth Marek, Irene Litvan, Anthony E Lang, et al. Mds clinical diagnostic criteria for parkinson's disease. Movement disorders, 30(12):1591-1601, 2015. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 449, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 525 + ], + "type": "text", + "content": "[10] Sanghee Moon, Hyun-Je Song, Vibhash D Sharma, Kelly E Lyons, Rajesh Pahwa, Abiodun E Akinwuntan, and Hannes Devos. Classification of parkinson's disease and essential tremor based on balance and gait characteristics from wearable motion sensors via machine learning techniques: a data-driven approach. Journal of neuroengineering and rehabilitation, 17:1-8, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 526, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 591 + ], + "type": "text", + "content": "[11] Andrea Bandini, Sia Rezaei, Diego L Guarín, Madhura Kulkarni, Derrick Lim, Mark I Boulos, Lorne Zinman, Yana Yunusova, and Babak Taati. A new dataset for facial motion analysis in individuals with neurological disorders. IEEE Journal of Biomedical and Health Informatics, 25(4):1111-1119, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 592, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 668 + ], + "type": "text", + "content": "[12] Aakash Kaku, Kangning Liu, Avinash Parnandi, Haresh Rengaraj Rajamohan, Kannan Venkataramanan, Anita Venkatesan, Audre Wirtanen, Natasha Pandit, Heidi Schambra, and Carlos Fernandez-Granda. Strokerehab: A benchmark dataset for sub-second action identification. Advances in Neural Information Processing Systems, 35:1671-1684, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[13] Marta Isabel ASN Ferreira, Fabio Augusto Barbieri, Vinicius Christianini Moreno, Tiago Penedo, and João Manuel RS Tavares. Machine learning models for parkinson's disease detection and stage classification based on" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "spatial-temporal gait parameters. Gait & Posture, 98:49-55, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "type": "text", + "content": "[14] Ferdous Wahid, Rezaul K Begg, Chris J Hass, Saman Halgamuge, and David C Ackland. Classification of parkinson's disease gait using spatial-temporal gait features. IEEE journal of biomedical and health informatics, 19(6):1794-1802, 2015. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 152, + 547, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 547, + 208 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 547, + 208 + ], + "type": "text", + "content": "[15] Yanyi Zhang, Ming Kong, Tianqi Zhao, Wenchen Hong, Qiang Zhu, and Fei Wu. Adhd intelligent auxiliary diagnosis system based on multimodal information fusion. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4494-4496, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 209, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 547, + 274 + ], + "type": "text", + "content": "[16] Mahmoud Seifallahi, Afsoon Hasani Mehraban, James E Galvin, and Behnaz Ghoraani. Alzheimer's disease detection using comprehensive analysis of timed up and go test via kinect v. 2 camera and machine learning. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 30:1589-1600, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 276, + 547, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 547, + 342 + ], + "type": "text", + "content": "[17] Deval Mehta, Umar Asif, Tian Hao, Erhan Bilal, Stefan Von Cavallar, Stefan Harrer, and Jeffrey Rogers. Towards automated and marker-less parkinson disease assessment: predicting updrs scores using sit-stand videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3841–3849, 2021. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 344, + 547, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 344, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 344, + 547, + 409 + ], + "type": "text", + "content": "[18] Renfei Sun, Kun Hu, Kaylena A Ehgoetz Martens, Markus Hagenbuchner, Ah Chung Tsoi, Mohammed Bennamoun, Simon JG Lewis, and Zhiyong Wang. Higher order polynomial transformer for fine-grained freezing of gait detection. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 411, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 487 + ], + "type": "text", + "content": "[19] Mark Endo, Kathleen L Poston, Edith V Sullivan, Li Fei-Fei, Kilian M Pohl, and Ehsan Adeli. Gaitforemer: Self-supervised pre-training of transformers via human motion forecasting for few-shot gait impairment severity estimation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 130-139. Springer, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 489, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 555 + ], + "type": "text", + "content": "[20] Mandy Lu, Qingyu Zhao, Kathleen L Poston, Edith V Sullivan, Adolf Pfefferbaum, Marian Shahid, Maya Katz, Leila Montaser Kouhsari, Kevin Schulman, Arnold Milstein, et al. Quantifying parkinson's disease motor severity under uncertainty using mds-updrs videos. Medical image analysis, 73:102179, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 556, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 612 + ], + "type": "text", + "content": "[21] Mehrbakhsh Nilashi, Othman Ibrahim, Hossein Ahmadi, Leila Shahmoradi, and Mohammadreza Farahmand. A hybrid intelligent system for the prediction of parkinson's disease progression using machine learning techniques. Biocybernetics and Biomedical Engineering, 38(1):1-15, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "type": "text", + "content": "[22] Trevor Exley, Sarah Moudy, Rita M Patterson, Joonghyun Kim, and Mark V Albert. Predicting updrs motor symptoms in individuals with parkinson's disease from force plates using machine learning. IEEE Journal of Biomedical and Health Informatics, 26(7):3486-3494, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[23] Max Little, Patrick McSharry, Eric Hunter, Jennifer Spielman, and Lorraine Ramig. Suitability of dysphonia measurements for telemonitoring of parkinson's disease. Nature Precedings, pages 1-1, 2008. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22559" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[24] Athanasios Tsanas, Max Little, Patrick McSharry, and Lorraine Ramig. Accurate telemonitoring of parkinson's disease progression by non-invasive speech tests. Nature Precedings, pages 1-1, 2009. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "type": "text", + "content": "[25] Clayton R Pereira, Danilo R Pereira, Francisco A Silva, Joao P Masieiro, Silke AT Weber, Christian Hook, and João P Papa. A new computer vision-based approach to aid the diagnosis of parkinson's disease. Computer Methods and Programs in Biomedicine, 136:79-88, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 173, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 287, + 228 + ], + "type": "text", + "content": "[26] Clayton R Pereira, Silke AT Weber, Christian Hook, Gustavo H Rosa, and Joao P Papa. Deep learning-aided parkinson's disease diagnosis from handwritten dynamics. In 2016 29th SIBGRAPI conference on graphics, patterns and images (SIBGRAPI), pages 340–346. IEEE, 2016." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 282 + ], + "type": "text", + "content": "[27] Catherine Taleb, Maha Khachab, Chafic Mokbel, and Laurence Likforman-Sulem. Feature selection for an improved parkinson's disease identification based on handwriting. In 2017 1st International Workshop on Arabic Script Analysis and Recognition (ASAR), pages 52-56. IEEE, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 284, + 286, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 286, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 286, + 338 + ], + "type": "text", + "content": "[28] Charalamos Sotirakis, Zi Su, Maksymilian A Brzezicki, Niall Conway, Lionel Tarassenko, James J FitzGerald, and Chrystalina A Antoniades. Identification of motor progression in parkinson's disease using wearable sensors and machine learning. npj Parkinson's Disease, 9(1):142, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 338, + 286, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 286, + 403 + ], + "type": "text", + "content": "[29] Chariklia Chatzaki, Vasileios Skaramagkas, Nikolaos Tachos, Georgios Christodoulakis, Evangelia Maniadi, Zinovia Kefalopoulou, Dimitrios I Fotiadis, and Manolis Tsiknakis. The smart-insole dataset: Gait analysis using wearable sensors with a focus on elderly and parkinson's patients. Sensors, 21(8):2821, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 405, + 286, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 286, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 286, + 459 + ], + "type": "text", + "content": "[30] Marc Bachlin, Meir Plotnik, Daniel Roggen, Inbal Maidan, Jeffrey M Hausdorff, Nir Giladi, and Gerhard Troster. Wearable assistant for parkinson's disease patients with the freezing of gait symptom. IEEE Transactions on Information Technology in Biomedicine, 14(2):436-446, 2009." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 460, + 286, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 286, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 286, + 535 + ], + "type": "text", + "content": "[31] Jens Barth, Jochen Klucken, Patrick Kugler, Thomas Kammerer, Ralph Steidl, Jürgen Winkler, Joachim Hornegger, and Björn Eskofier. Biometric and mobile gait analysis for early diagnosis and therapy monitoring in parkinson's disease. In 2011 annual international conference of the IEEE engineering in medicine and biology society, pages 868-871. IEEE, 2011." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 537, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 286, + 624 + ], + "type": "text", + "content": "[32] Caroline Ribeiro De Souza, Runfeng Miao, Julia Ávila De Oliveira, Andrea Cristina De Lima-Pardini, Debora Fragoso De Campos, Carla Silva-Batista, Luis Teixeira, Solaiman Shokur, Bouri Mohamed, and Daniel Boari Coelho. A public data set of videos, inertial measurement unit, and clinical scales of freezing of gait in individuals with parkinson's disease during a turning-in-place task. Frontiers in Neuroscience, 16:832463, 2022. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 625, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 286, + 713 + ], + "type": "text", + "content": "[33] Mandy Lu, Kathleen Poston, Adolf Pfefferbaum, Edith V Sullivan, Li Fei-Fei, Kilian M Pohl, Juan Carlos Niebles, and Ehsan Adeli. Vision-based estimation of mds-updrs gait scores for assessing parkinson's disease motor severity. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part III 23, pages 637-647. Springer, 2020. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[34] Andrea Sabo, Sina Mehdizadeh, Kimberley-Dale Ng, Andrea Iaboni, and Babak Taati. Assessment of parkinsonian gait in older adults with dementia via human pose tracking in video data. Journal of neuroengineering and rehabilitation, 17(1):1-10, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "text", + "content": "[35] Md Saiful Islam, Wasifur Rahman, Abdelrahman Abdelkader, Phillip T Yang, Sangwu Lee, Jamie L Adams, Ruth B Schneider, E Dorsey, and Ehsan Hoque. Using ai to measure parkinson's disease severity at home. npj digital medicine, 2023. 4, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "type": "text", + "content": "[36] Andrea Sabo, Carolina Gorodetsky, Alfonso Fasano, Andrea Iaboni, and Babak Taati. Concurrent validity of zero instrumented walkway and video-based gait features in adults with parkinson's disease. IEEE Journal of Translational Engineering in Health and Medicine, 10:1-11, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 243, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 307 + ], + "type": "text", + "content": "[37] Samuel Rupprechter, Gareth Morinan, Yuwei Peng, Thomas Foltynie, Krista Sibley, Rimona S Weil, Louise-Ann Leyland, Fahd Baig, Francesca Morgante, Ro'ee Gilron, et al. A clinically interpretable computer-vision based method for quantifying gait in parkinson's disease. Sensors, 21(16):5437, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "text", + "content": "[38] Erin Smith, John Bertoni, Danish Bhatti, and Diego Torres-Russotto. Is the updrs a reliable tool for detecting the worse side in parkinson's disease? (1584). *Neurology*, 94:1584, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 354, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 420 + ], + "type": "text", + "content": "[39] Ashwani Jha, Elisa Menozzi, Rebecca Oyekan, Anna Latorre, Eoin Mulroy, Sebastian R Schreglmann, Cosmin Stamate, Ioannis Daskalopoulos, Stefan Kueppers, Marco Luchini, et al. The cloudupdrs smartphone software in parkinson's study: cross-validation against blinded human raters. npj Parkinson's Disease, 6(1):36, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "type": "text", + "content": "[40] Delaram Safarpour, Marian L Dale, Vrutangkumar V Shah, Lauren Talman, Patricia Carlson-Kuhta, Fay B Horak, and Martina Mancini. Surrogates for rigidity and pigd mds-updrs subscores using wearable sensors. Gait & Posture, 91:186-191, 2022. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "text", + "content": "[41] Erika Rovini, Carlo Maremmani, and Filippo Cavallo. How wearable sensors can support parkinson's disease diagnosis and treatment: a systematic review. Frontiers in neuroscience, 11:288959, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "type": "text", + "content": "[42] Mercedes Barrachina-Fernández, Ana María Maitín, Carmen Sánchez-Avila, and Juan Pablo Romero. Wearable technology to detect motor fluctuations in parkinson's disease patients: current state and challenges. Sensors, 21(12):4188, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 580, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 580, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 580, + 545, + 645 + ], + "type": "text", + "content": "[43] Balasundaram Kadirvelu, Constantinos Gavriel, Sathiji Nageshwaran, Jackson Ping Kei Chan, Suran Nethisinghe, Stavros Athanasopoulos, Valeria Ricotti, Thomas Voit, Paola Giunti, Richard Festenstein, et al. A wearable motion capture suit and machine learning predict disease progression in friedreich's ataxia. Nature Medicine, 29(1):86-94, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[44] Benjamin Filtjens, Pieter Ginis, Alice Nieuwboer, Peter Slaets, and Bart Vanrumste. Automated freezing of gait assessment with marker-based motion capture and multistage spatial-temporal graph convolutional neural networks. Journal of NeuroEngineering and Rehabilitation, 19(1):48, 2022." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22560" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[45] Muhammad Hassan Khan, Manuel Schneider, Muhammad Shahid Farid, and Marcin Grzegorzek. Detection of infantile movement disorders in video data using deformable part-based model. Sensors, 18(10):3202, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 139 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 139 + ], + "type": "text", + "content": "[46] K Severson. campy. https://github.com/ksseverson57/campy, 2022.4.1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 287, + 217 + ], + "type": "text", + "content": "[47] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for perceiving and processing reality. In Third Workshop on Computer Vision for AR/VR at IEEE Computer Vision and Pattern Recognition (CVPR) 2019, 2019. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 287, + 252 + ], + "type": "text", + "content": "[48] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 254, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 254, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 254, + 287, + 308 + ], + "type": "text", + "content": "[49] Zhilin Guo, Weiqi Zeng, Taidong Yu, Yan Xu, Yang Xiao, Xuebing Cao, and Zhiguo Cao. Vision-based finger tapping test in patients with parkinson's disease via spatial-temporal 3d hand pose estimation. IEEE Journal of Biomedical and Health Informatics, 26(8):3848-3859, 2022. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 309, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 309, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 287, + 374 + ], + "type": "text", + "content": "[50] Ahnjili ZhuParris, Eva Thijssen, Willem O Elzinga, Soma Makai-Boloni, Wessel Kraaij, Geert J Groeneveld, and Robert J Doll. Treatment detection and movement disorder society-unified parkinson's disease rating scale, part iii estimation using finger tapping tasks. Movement Disorders, 38(10):1795-1805, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 376, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 376, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 49, + 376, + 287, + 420 + ], + "type": "text", + "content": "[51] Jeffrey M Hausdorff. Gait dynamics in parkinson's disease: common and distinct behavior among stride length, gait variability, and fractal-like scaling. Chaos: An Interdisciplinary Journal of Nonlinear Science, 19(2), 2009. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 422, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 422, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 49, + 422, + 287, + 475 + ], + "type": "text", + "content": "[52] Michele Pistacchi, Manuela Gioulis, Flavio Sanson, Ennio De Giovanni, Giuseppe Filippi, Francesca Rossetto, and Sandro Zambito Marsala. Gait analysis and clinical correlations in early parkinson's disease. *Functional neurology*, 32(1):28, 2017. 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 477, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 477, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 49, + 477, + 287, + 520 + ], + "type": "text", + "content": "[53] Tianze Yu, Kye Won Park, Martin J McKeown, and Z Jane Wang. Clinically informed automated assessment of finger tapping videos in parkinson's disease. Sensors, 23(22):9149, 2023. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 522, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 522, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 49, + 522, + 287, + 577 + ], + "type": "text", + "content": "[54] Kye Won Park, Eun-Jae Lee, Jun Seong Lee, Jinhoon Jeong, Nari Choi, Sungyang Jo, Mina Jung, Ja Yeon Do, Dong-Wha Kang, June-Goo Lee, et al. Machine learning-based automatic rating for cardinal symptoms of parkinson disease. Neurology, 96(13):e1761-e1769, 2021. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 578, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 578, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 49, + 578, + 287, + 622 + ], + "type": "text", + "content": "[55] R Agostino, A Curra, M Giovannelli, N Modugno, M Manfredi, and A Berardelli. Impairment of individual finger movements in parkinson's disease. *Movement Disorders*, 18(5):560-565, 2003." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 689 + ], + "type": "text", + "content": "[56] Eva Thijssen, Soma Makai-Boloni, Emilie van Brummelen, Jonas den Heijer, Yalcin Yavuz, Robert-Jan Doll, and Geert Jan Groeneveld. A placebo-controlled study to assess the sensitivity of finger tapping to medication effects in parkinson's disease. *Movement Disorders Clinical Practice*, 9(8):1074-1084, 2022. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "type": "text", + "content": "[57] M Yokoe, R Okuno, T Hamasaki, Y Kurachi, K Akazawa, and S Sakoda. Opening velocity, a novel parameter, for fin" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "ger tapping test in patients with parkinson's disease. Parkinsonism & related disorders, 15(6):440-444, 2009. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 162 + ], + "type": "text", + "content": "[58] Junjie Li, Huaiyu Zhu, Yun Pan, Haotian Wang, Zhidong Cen, Dehao Yang, and Wei Luo. Three-dimensional pattern features in finger tapping test for patients with parkinson's disease. In 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 3676-3679. IEEE, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 229 + ], + "type": "text", + "content": "[59] Noreen Akram, Haoxuan Li, Aaron Ben-Joseph, Caroline Budu, David A Gallagher, Jonathan P Bestwick, Anette Schrag, Alastair J Noyce, and Cristina Simonet. Developing and assessing a new web-based tapping test for measuring distal movement in parkinson's disease: a distal finger tapping test. Scientific reports, 12(1):386, 2022. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 230, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 273 + ], + "type": "text", + "content": "[60] Kevin J Brusse, Sandy Zimdars, Kathryn R Zalewski, and Teresa M Steffen. Testing functional performance in people with parkinson disease. Physical therapy, 85(2):134-141, 2005. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 275, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 275, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 275, + 545, + 330 + ], + "type": "text", + "content": "[61] Olumide Sofuwa, Alice Nieuwboer, Kaat Desloovere, Annemarie Willems, Fabienne Chavret, and Ilse Jonkers. Quantitative gait analysis in parkinson's disease: comparison with a healthy control group. Archives of physical medicine and rehabilitation, 86(5):1007-1013, 2005. 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 331, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 374 + ], + "type": "text", + "content": "[62] JA Zeni Jr, JG Richards, and JS2384115 Higginson. Two simple methods for determining gait events during treadmill and overground walking using kinematic data. Gait & posture, 27(4):710-714, 2008. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 376, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 430 + ], + "type": "text", + "content": "[63] Christopher K Rhea and Adam W Kiefer. Patterned variability in gait behavior: How can it be measured and what does it mean. Gait biometrics: basic patterns, role of neurological disorders and effects of physical activity, pages 17-44, 2014. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 433, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 476 + ], + "type": "text", + "content": "[64] Martin Patrick Pauli, Constantin Pohl, and Martin Golz. Balanced leave-one-subject-out cross-validation for microsleep classification. Current Directions in Biomedical Engineering, 7(2):147-150, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 478, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 542 + ], + "type": "text", + "content": "[65] Ryutaro Tanno, Ardavan Saeedi, Swami Sankaranarayanan, Daniel C Alexander, and Nathan Silberman. Learning from noisy labels by regularized estimation of annotator confusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11244–11253, 2019. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 544, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 578 + ], + "type": "text", + "content": "[66] Gwyn N Lewis, Winston D Byblow, and Sharon E Walt. Stride length regulation in parkinson's disease: the use of extrinsic, visual cues. Brain, 123(10):2077-2090, 2000. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 633 + ], + "type": "text", + "content": "[67] CJ Hawley, N Fineberg, AG Roberts, D Baldwin, A Sahadevan, and V Sharman. The use of the simpson angus scale for the assessment of movement disorder: a training guide. International Journal of Psychiatry in Clinical Practice, 7(4):349-2257, 2003. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "text", + "content": "[68] K Muller, CK Hemelrijk, J Westerweel, and DSW Tam. Calibration of multiple cameras for large-scale experiments using a freely moving calibration target. *Experiments in Fluids*, 61(1):7, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[69] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22561" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "text", + "content": "social motion capture. In Proceedings of the IEEE international conference on computer vision, pages 3334-3342, 2015. 7" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22562" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/52ac8b22-207c-41ff-adcb-c9f4c82bca08_content_list.json b/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/52ac8b22-207c-41ff-adcb-c9f4c82bca08_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..48ffd17f4d12126aa72b89d2d77137937d51e8a4 --- /dev/null +++ b/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/52ac8b22-207c-41ff-adcb-c9f4c82bca08_content_list.json @@ -0,0 +1,1401 @@ +[ + { + "type": "text", + "text": "TULIP: Transformer for Upsampling of LiDAR Point Clouds", + "text_level": 1, + "bbox": [ + 174, + 130, + 794, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bin Yang $^{1}$ , Patrick Pfreundschuh $^{1}$ , Roland Siegwart $^{1}$ , Marco Hutter $^{2}$ , Peyman Moghadam $^{3,4\\dagger}$ , Vaishakh Patil $^{2\\dagger}$ $^{1}$ Autonomous Systems Lab, ETH Zurich, Switzerland, $^{2}$ Robotic Systems Lab, ETH Zurich, Switzerland \n $^{3}$ School of EER, Queensland University of Technology (QUT), Australia, $^{4}$ Data61, CSIRO, Brisbane, Australia", + "bbox": [ + 86, + 179, + 967, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{biyang, patripfr, rolandsi, mahutter, patilv}@ethz.ch, peyman.moghadam@csiro.au", + "bbox": [ + 179, + 236, + 870, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LiDAR Upsampling is a challenging task for the perception systems of robots and autonomous vehicles, due to the sparse and irregular structure of large-scale scene contexts. Recent works propose to solve this problem by converting LiDAR data from 3D Euclidean space into an image super-resolution problem in 2D image space. Although their methods can generate high-resolution range images with fine-grained details, the resulting 3D point clouds often blur out details and predict invalid points. In this paper, we propose TULIP, a new method to reconstruct high-resolution LiDAR point clouds from low-resolution LiDAR input. We also follow a range image-based approach but specifically modify the patch and window geometries of a Swin-Transformer-based network to better fit the characteristics of range images. We conducted several experiments on three public real-world and simulated datasets. TULIP outperforms state-of-the-art methods in all relevant metrics and generates robust and more realistic point clouds than prior works. The code is available at https://github.com/ethz-as1/TULIP.git.", + "bbox": [ + 76, + 318, + 473, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 652, + 209, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Light Detection And Ranging (LiDAR) is one of the most common sensors for perception in various fields of autonomy, such as autonomous driving, and unmanned aerial vehicles (UAV). LiDARs are used to generate 3D point clouds of the scene. These point clouds are essential for mapping, localization, and object detection tasks. However, the accuracy of these tasks often depends on the resolution of the point cloud [53]. Furthermore, the resolution of a LiDAR is inherently associated with increased energy consumption and cost, making its use impractical for various applications. LiDARs also have different vertical and horizontal resolutions. The vertical resolution of rotating 3D LiDARs is typically much lower than the horizontal res", + "bbox": [ + 76, + 676, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e09006fdab21513ee203b9986e2f4dc6a5bb4922c1b25eafbe50755107db8b76.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 506, + 285, + 696, + 387 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/87a798a93a376080981a05716c735f8b529f30f9953116b01f80034c8ffc0275.jpg", + "image_caption": [ + "(b) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 700, + 287, + 885, + 387 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0761cadd0d2db11c46775303ea0145c26832e235fcba08657876c5ea151c4623.jpg", + "image_caption": [ + "(c) SwinIR [29]" + ], + "image_footnote": [], + "bbox": [ + 503, + 405, + 694, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/049793755ba3abf39502d794c1ddf168595b53ca9f34e048d83b236367bbaf22.jpg", + "image_caption": [ + "(d) TULIP (Ours)", + "Figure 1. TULIP generates more realistic LiDAR point clouds from low-resolution LiDAR input (a) and outperforms the state-of-the-art image super-resolution approach [29] in upsampling with fine-grained details of different objects in the scene." + ], + "image_footnote": [], + "bbox": [ + 699, + 402, + 883, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "olution. These limitations make an upsampling technique necessary to increase the resolution of the LiDAR data, especially in the vertical direction. Moreover, upsampling LiDAR data has the potential to counter domain shift problems in LiDAR-based learning methods. Notably, the use of lower resolution LiDAR data in a system that has been trained on higher resolution data leads to a significant drop in performance [3]. Therefore, such upsampling techniques can not only help to mitigate the domain gap by creating a virtual high-resolution sensor that matches the target domain [42, 61] but also reduce the high cost of collecting new LiDAR data, annotating and retraining the methods.", + "bbox": [ + 496, + 583, + 890, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has led to remarkable advancements in LiDAR upsampling techniques in recent years. Several recent methods [42, 56, 58] have focused on learning the upsampling process within the 3D Euclidean space. However, the processing of 3D data in deep learning can be quite resource-intensive. One way to address this challenge is the representation of the point cloud as a range image [18, 26, 43, 49]. This approach facilitates the adoption of well-established image-based super-resolution methods.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Authors share last authorship.", + "bbox": [ + 94, + 886, + 261, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15354", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Such methods mostly employ autoencoder-style networks with convolutional and deconvolutional layers [14, 43, 62]. However, these methods are not directly transferable to range images. Range images generated by the projection of sparse 3D point clouds have distinct sharp edges at the object boundaries. Convolutional neural network-based methods often induce edge smoothing due to the regularization effect, which limits their usability for LiDAR range image processing.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As an alternative, Transformer [48] has shown great success in RGB image super-resolution and produces less blurry outputs [29]. Vision transformers, however, face challenges such as the need for large amounts of training data [9, 29, 35] and the global computation of self-attention, limiting the capture of local information in the range images. To address this, Swin Transformer [33] introduced tokenizing the input image and applying the self-attention mechanism locally in independent windows. While the modifications to Swin Transformer have led to robust training and promising results compared to Vision Transformer, its default structure and settings are not ideally suited for processing LiDAR range images. These range images consist of single-channel data, representing 3D spatial information, as opposed to three-channel color images that depict visual appearance. Furthermore, range images feature large, smooth areas within 3D objects, separated by distinct sharp edges between objects. As a result, minor inaccuracies in the 2D prediction (e.g. blurry output) can lead to significant differences in the projected 3D occupancy. This non-uniform distribution of image details further complicates network training. The challenging details in the range image are concentrated in a few specific pixels, contrasting with RGB images where relevant details are evenly distributed. This makes range images fundamentally different from RGB images. Inspired by the limitations of state-of-the-art image upsampling methods and the potential of Swin-Transformer [33], we propose a novel network named TULIP. Our proposed geometry-aware architecture is tailored to accommodate LiDAR range image data. The proposed method takes a low-resolution range image as input and produces a high-resolution range image, which can then be projected into a 3D point cloud. To accommodate LiDAR range image geometry, we utilize one-line row patches to tokenize the input range image, as opposed to the square patches used for RGB images. This approach helps to preserve vertical information for the upsampling process while capturing boundary discontinuities effectively. This also creates a trade-off between local detail capture and model complexity in the training process. Additionally, we incorporate a non-square window for local self-attention computation. This enhances learning of spatial contexts at different scales in the range image. Further, we extensively train and evaluate TULIP on two real-world and", + "bbox": [ + 75, + 224, + 472, + 893 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "one simulated autonomous driving datasets: KITTI [21], DurLAR [28] and CARLA [26], respectively. Comparisons to existing approaches demonstrate that TULIP outperforms the state-of-the-art methods on all three datasets.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 162, + 650, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image super-resolution has seen significant progress in recent years. The technique aims to construct high-resolution (HR) images from low-resolution (LR) observations, often leveraging advances in Convolutional Neural Networks (CNNs) to enhance the fidelity and detail of images for better visualization and information extraction. Subsequently, several enhanced frameworks have been developed [14, 15, 24, 25, 59].", + "bbox": [ + 496, + 191, + 890, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Driven by their success in the field of natural language processing (NLP) [4, 23], Transformers [48] have been extended to solve a variety of vision-related tasks such as object detection [6] and semantic segmentation [2, 7]. Such Vision Transformers (ViT) [17] excel at learning to focus on relevant image regions by exploring global interactions among different regions. Their impressive performance has led to their adoption in image restoration tasks [9, 11, 29] as well. While they brought significant improvements in RGB image super-resolution, Vision Transformer [17] comes with the drawback of the quadratic computational complexity of self-attention. Additionally, they capture mostly global dependencies within the data and require large amounts of data for training.", + "bbox": [ + 496, + 311, + 890, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, Swin Transformer [33] has been proposed. Unlike the Vision Transformer, which relies on global self-attention across the entire image, the Swin Transformer employs a local window-based attention mechanism and establishes a pyramid-like architecture. This approach processes images hierarchically by gradually merging smaller image patches into larger ones. This strategy enables more effective handling of different scales and facilitates the processing of multi-scale features, making it particularly suitable for range image processing. Furthermore, due to Swin Transformer's hierarchical structure, the increase in the number of model parameters scales linearly with the input image size, which mitigates the challenges associated with handling large range images. Additionally, they require less training data than the classical Vision Transformer, which is beneficial for LiDAR upsampling as LiDAR datasets are typically magnitudes smaller than RGB datasets used for Vision Transformer. The Swin and Vision Transformers have also been used for image super-resolution on omnidirectional panoramic camera images [45, 55]. However, these approaches are specifically designed for the unique characteristics of these cameras (e.g. strong distortions, spherical continuity) that are vastly different from the LiDAR sensing model, which renders them unsuitable for application to range images.", + "bbox": [ + 496, + 523, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15355", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other related works have focussed on upsampling high-density point clouds from low-density point clouds [38, 54, 56, 60, 61]. While those approaches also provide qualitative results for LiDAR point clouds, they focus on increasing the general 3D point density of point clouds of single objects instead of a large scene. Another area closely related to LiDAR upsampling is depth completion [12, 22, 30, 36, 37, 39, 46, 52]. The primary objective of depth completion is to improve sparse depth estimates acquired from LiDAR by integrating data from multiple sensors, predominantly RGB cameras. The result of depth completion methods is a dense depth map that provides depth values to every pixel in the input depth map. Nonetheless, this field slightly diverges from LiDAR upsampling due to its reliance on multimodal sensors and its common deployment within a restricted field of view, making it not directly comparable within the scope of this paper. LiDAR point clouds exhibit specific characteristics, such as a distinct point pattern (due to the stacked lasers in rotating 3D LiDARs) and a decrease in point density with increasing distance from the sensor. Differently from arbitrary point cloud upsampling, LiDAR upsampling tries to mimic a realistic point cloud of a high-resolution LiDAR given the point cloud of low-resolution LiDAR and thus targets a different result than the approaches above. Due to the difficulty and high computational cost of upsampling LiDAR point clouds in 3D space [1, 8, 42], most works [26, 43, 47] represent the point cloud as a range image and perform the LiDAR upsampling task in 2D image space. LiDAR-CNN [47] introduces a CNN-based architecture with semantic and perceptual guidance. The method specifies two further loss functions besides a per-point reconstruction loss to acquire a better synthesis of the high-resolution range image. However, this approach is limited to data with semantic annotations, which drastically reduces its applicability. LiDAR-SR [43] deploys a CNN-based network with U-Net [40] architecture. It additionally uses Monte-Carlo dropout post-processing to reduce the amount of noisy points in the prediction. Approaches based on convolutional operations tend to fail to reconstruct the sharpness in the range image effectively. While the previous approaches directly predict the range image, Implicit LiDAR Network (ILN) [26] uses an implicit neural architecture that learns interpolation weights to fill in new pixels instead of their values directly. Although the method outperforms CNN-based approaches in training speed and preserving the geometrical details in the input, it can still suffer from limited neighboring information in the LiDAR data, especially in areas distant from the sensor due to the extremely sparse input data. Differently from the previous works, our approach builds on Swin Transformer as a backbone and is specifically customized to effectively process range images.", + "bbox": [ + 76, + 90, + 472, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 89, + 635, + 108 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LiDAR upsampling is achieved by upsampling a range image, effectively transforming it from a 3D upsampling problem into a 2D image super-resolution problem [18, 26, 43, 47]. It is, therefore, an evident approach to build on RGB image super-resolution. However, image super-resolution aims to solve a different problem than LiDAR upsampling. Image super-resolution enhances the visual appearance of a low-resolution RGB image by recreating a high-resolution image. While quantitative metrics for evaluating these works exist, there is no one specific correct solution. Differently, LiDAR upsampling tries to recreate a high-resolution, 3D LiDAR sensor output from a low-resolution LiDAR input point cloud. Range images contain more geometrical and spatial contexts rather than visual information, and the scene's geometry strictly dictates the correct solution. The underlying sensing model is also drastically different, as LiDARs are active sensors. The respective range pixel values purely depend on distance, while pixels in an RGB camera depend on many factors, such as scene appearance, lighting conditions, exposure time, and white balance. Furthermore, range images have a highly asymmetric aspect ratio, typically between 1:8 to 1:64, which differs drastically from regular camera images. Accordingly, most works [18, 43, 47] only perform vertical upsampling, in contrast to image super-resolution, which operates in both dimensions. A naive application of state-of-the-art image super-resolution networks [29] to range images does not result in adequate performance (Fig. 1). To this end, we developed TULIP, a novel range image-based LiDAR Upsampling method. Based on the observations above, we specifically adapt a network based on SwinUnet [5] to incorporate range images better. In the following sections, we will describe the technical details of our method, with a special focus on how our approach accommodates range images.", + "bbox": [ + 496, + 117, + 893, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 500, + 648, + 684, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A LiDAR point cloud consists of points captured during one revolution $\\mathcal{P} = \\{\\mathbf{p}_1,\\dots ,\\mathbf{p}_n\\}$ . Each measurement represents a 3D point $\\mathbf{p}_i = \\{x_i,y_i,z_i\\}$ . As input, we assume a low-resolution point cloud $\\mathcal{P}_l$ with $n_l = H_l\\times W_l$ points, where $H_{l}$ and $W_{l}$ correspond to vertical and horizontal resolution. We aim to predict a high-resolution point cloud $\\mathcal{P}_h$ , that has the same field-of-view(FoV) as $\\mathcal{P}_l$ but contains $n_h = \\beta *n_l$ points. In our experiments, we set $\\beta = 4$ . We only increase vertical resolution (amount of LiDAR beams), i.e., $n_h = H_h\\times W_h$ , where $H_{h} = \\beta *H_{l}$ and $W_{h} = W_{l}$ . We project the point cloud into a 2D range image. In this image, each row and column coordinate $v,u$ correspond to the respective elevation and azimuth angles of the LiDAR points, while the pixel value contains the range of the point $r = \\sqrt{x^2 + y^2 + z^2}$ . The image coordinate of a 3D point", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15356", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "can be calculated using a spherical projection model:", + "bbox": [ + 76, + 90, + 428, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{l} u \\\\ v \\end{array} \\right] = \\left[ \\frac {\\frac {W}{2} - \\frac {W}{2 \\pi} a r c t a n (\\frac {y}{x})}{\\Theta_ {m a x} - \\Theta_ {m i n}} * (\\Theta_ {m a x} - a r c t a n (\\frac {z}{\\sqrt {x ^ {2} + y ^ {2}}}) \\right] (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 113, + 468, + 157 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\Theta_{max}$ and $\\Theta_{min}$ correspond to the limits of the vertical field of view of the LiDAR. We formulate the LiDAR upsampling problem as an image upsampling problem, i.e. we aim to predict a high-resolution range image $I_h \\in \\mathbf{R}^{1 \\times H_h \\times W}$ given a low-resolution range image $I_l \\in \\mathbf{R}^{1 \\times H_l \\times W}$ . We can then calculate the high-resolution point cloud $\\mathcal{P}_h$ by inverting the projection in Eq. 1.", + "bbox": [ + 75, + 162, + 468, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Architecture Design and Overview", + "text_level": 1, + "bbox": [ + 76, + 277, + 377, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our network builds upon Swin-Unet [5] which was originally designed for image segmentation. We deploy a U-shaped network structure featuring skip connections that link the encoder and decoder modules. The fundamental building block of TULIP is the Swin Transformer [33]. TULIP first patchifies and maps the input to a high-dimensional feature space. The tokenized low-level feature maps are passed through a multi-stage encoder. Each stage contains a pair of Swin Transformer [33] blocks and a patch merging layer to downsample the resolution by a factor of two and increase the dimension by a factor of four. This step realizes the hierarchic computation and extraction of multi-scale features through encoding. Within each stage, multiple instances of the self-attention mechanism are performed locally in parallel, which is the so-called Window Multi-Head Self Attention (W-MSA). After passing through a two-layer MLP and residual connection with the input, the feature vector is further passed to the second part of the Swin Transformer block, utilizing a Shifted Window Multi-Head Self Attention (SW-MSA) which extends W-MSA by employing a shifted window operation. This enhances the model capacity as it compensates for the lack of interaction between the local windows in W-MSA. The decoder, which realizes the upsampling has a symmetric design to the encoder and operates in a reverse way of the encoder. The resolution of the feature maps is first expanded, and the dimension is reduced by a factor of two accordingly. We feed existing geometrical information via skip connections. The feature maps are transformed into a single-channel, high-resolution range image in the last layer. The reconstruction head comprises a $1 \\times 1$ convolutional layer for feature expansion, followed by a Leaky ReLU activation, a pixel shuffle layer [44], and another $1 \\times 1$ convolutional layer for final projection. As the loss function, we select the pixel-wise L1 loss. Details of the network can be found in the Appendix.", + "bbox": [ + 75, + 300, + 470, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Tokenization", + "text_level": 1, + "bbox": [ + 76, + 847, + 220, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tokenization is done by a patch partition layer that creates an initial feature embedding from the input range image.", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This initial feature representation significantly influences the network's performance. Specifically, the selection of this layer determines how the inherent characteristics, patterns, and relevant information of the LiDAR range image data are encoded into a format that the network can effectively learn from. Most transformer-based RGB image super-resolution approaches utilize a relatively large patch size to tokenize the input [9, 29, 35]. These larger patches enable the construction of a global spatial context, facilitating an understanding of the overall structure and high-level features of the RGB image. These networks employ square-shaped patches and implement upsampling in both vertical and horizontal directions. In contrast, LiDAR upsampling primarily aims to enhance the vertical resolution of the input, given that range images possess properties that are vastly different from RGB images. Furthermore, attention across more pixels in range images is less useful than in RGB images, since there is almost no geometrical relation for the spatial contexts that are far from each other in the 2D range image space. Motivated by this, we propose two adjustments to effectively process range images within a Swin-Unet:", + "bbox": [ + 496, + 90, + 890, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Row-Based Patch Partition: Our model builds row patches with a dimension of $1 \\times 4$ for range image tokenization. The new patch geometry is designed to retain full vertical information while compressing horizontally. This aligns with our objective of conserving and extending vertical details in the range image. Besides that, row patches excel at capturing boundary discontinuities between distinct objects and the background within the scene. Moreover, the selection of the patch size strikes a balance between enhancing model performance and limiting model capacity. Smaller patches can capture more detailed local information, but they also increase the number of model parameters, which can slow down the model's training process.", + "bbox": [ + 496, + 426, + 890, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Circular Padding (CP) in Horizontal Dimension: Instead of adding zero pixels padding, our model uses circular padding in the horizontal dimension. This avoids introducing artificial features and naturally matches the sensor model of a rotating 3D LiDAR. This preserves accurate neighbor relations along the edges, which would otherwise be compromised during the projection to the 2D range image. The Circular Padding (CP) also allows us to fit flexible input sizes without changing the patch geometry.", + "bbox": [ + 496, + 626, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Non-Square window for local self-attention", + "text_level": 1, + "bbox": [ + 498, + 785, + 852, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inside Swin Transformer, input feature maps are window-partitioned along width and height to compute the self-attention locally in each window. In addition to square-shaped RGB images, some recent work applies a square window for panoramic data as well [31, 57]. However, using square windows on range images is unfavorable. On one", + "bbox": [ + 496, + 810, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15357", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/058fa761a4909494e7b5e950bae2e62b7b95ebedb64e4a42db47afd075e1f759.jpg", + "image_caption": [ + "Figure 2. Overview of TULIP: The low-resolution point cloud is spherically projected into the range image. In the training and inference, the range image is tokenized into feature embeddings without compressing the spatial resolution along the vertical axis. We adopt the U-Net-based network presented in [5] for range image upsampling. We follow the hierarchical structure of Swin Transformer [33] for feature extraction and use the non-square window geometry for computing the local self-attention. The reconstruction head generates the high-resolution range image from the latent features and then, the point cloud can be obtained by projecting pixels back into the 3D space." + ], + "image_footnote": [], + "bbox": [ + 138, + 85, + 834, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "hand, small windows can focus more on attention within near areas but lead to increasing computational complexity and can struggle with the reconstruction of objects at a further distance in the scene. On the other hand, a larger square window can blend scene-related contexts at different scales, which can potentially degrade the network performance. In a prior study done by Eskandar et al. [18], they demonstrated a significant enhancement in range image superresolution by separately upsampling the upper and lower portions of the data. In contrast to their approach of utilizing CNN-based shallower and deeper network branches for distinct feature extraction, our proposal involves the direct partitioning of the input using a rectangular window. This window geometry facilitates robust attention interaction among distant points through W-MSA and subtle crossattention between objects at multiple scales via SW-MSA and hierarchical processing. In addition, range images, as typical panorama data, inherently contain abundant information along the horizontal direction. Therefore, directing attention more toward image width rather than height can aid the network in capturing a greater amount of geometrical information. At the same time, the computational complexity remains the same as using a large square window.", + "bbox": [ + 75, + 349, + 472, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3 Further Adaptation and Refinement", + "text_level": 1, + "bbox": [ + 76, + 720, + 382, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Patch Unmerging (PU) and Pixel Shuffle (PS): Similar to the patch merging layer that is responsible for down-sampling at the encoder stage, a mechanism is necessary within the decoder to upsample the feature maps. Conventional CNN-based approaches [43, 50] achieve upsampling through the transposed convolutional operation. As they tend to smooth out image sharpness [29], we reversed the operation of patch merging, re-arranging the channels of patches into their respective positions in a grid to up-sample the spatial resolution of input feature maps, which we", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "denote as patch unmerging. Furthermore, to reconstruct the final range image, we built the reconstruction head upon the upsampled feature maps. The implementation of the reconstruction module is based on the pixel shuffle layer [44].", + "bbox": [ + 496, + 349, + 890, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Monte Carlo Dropout: We additionally use Monte Carlo Dropout [43] to filter unreliable points by thresholding the uncertainty. MC-Dropout executes several feed-forward passes during inference with different active dropouts, which yields a distribution of outputs. We refine the results by removing points with high output variance, as they often come down to noisy, invalid points.", + "bbox": [ + 496, + 412, + 892, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 537, + 633, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 563, + 705, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets: We conduct experiments on three different datasets, that include the two large-scale real-world datasets DurLAR [28] and KITTI [21] (Sec. 4.3.2), as well as on a dataset [26] that was generated using the CARLA simulator [16] (Sec. 4.3.1). We select test sequences that are recorded in different locations than the train set to avoid spatial overlap. We vertically downsample the high-resolution range images with a factor of four by skipping the respective lines to generate low-resolution input images. For ablation studies in Sec. 4.2, we select the KITTI dataset.", + "bbox": [ + 496, + 590, + 890, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details: In all experiments, we use distributed processing with $4 \\times$ GeForce RTX 2080 Ti. For training, we use a fixed batch size of eight per GPU for all datasets. We use the AdamW [34] optimizer with a weight decay factor of 0.01 and a base learning rate of 5e-4.", + "bbox": [ + 496, + 744, + 890, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics: We evaluate the Mean Absolute Error (MAE) for all the pixels in the generated 2-D range images. Additionally, we assess the performance by considering the 3D points reconstructed by our neural networks. The Chamfer Distance (CD) measures the Euclidean distance between", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15358", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5d4a9e49b8b088b9500b3b231376ef94ea57b36bf3533574cf0ffbdb29f35ea7.jpg", + "image_caption": [ + "(a) Scene" + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 187, + 148 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cc353321f5773f05891d97252c6797477d1397fbe84ae08db5338efefc3e2e2d.jpg", + "image_caption": [ + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 187, + 88, + 256, + 148 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ea230d4c9029e503049eacac46d1cb3a88121c05998bf8245211e1e3f1ded68.jpg", + "image_caption": [ + "(c) $1 \\times 4$" + ], + "image_footnote": [], + "bbox": [ + 259, + 88, + 325, + 148 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2b7fb6383d0634125ee16e7befb8e254fbef6b9c57b6d649deb283c5c0736bb6.jpg", + "image_caption": [ + "(d) $2 \\times 2$" + ], + "image_footnote": [], + "bbox": [ + 326, + 88, + 393, + 148 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0fe2aab2c5160df7eff0382e24a75b46a63e6c6d1e6cc172fd70aef00196d387.jpg", + "image_caption": [ + "(e) $4 \\times 4$" + ], + "image_footnote": [], + "bbox": [ + 397, + 88, + 465, + 148 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/988a09b5144fc9df356a255f8d2e51f8348f0fb0fa6b554d4113efefb82ea6cd.jpg", + "image_caption": [ + "Figure 3. Upsampling results with different patch sizes: We observe that square patches $(4 \\times 4, 2 \\times 2)$ blur out edges and create invalid connections between separate walls, while the proposed $1 \\times 4$ patch generates sharp edges and correct discontinuities.", + "(a) Scene", + "Figure 4. Upsampling results with different window geometries: Rectangular windows improve reconstruction of objects at different ranges, indicated by more distinct discontinuities (top) and a better reconstruction of the car (bottom)." + ], + "image_footnote": [], + "bbox": [ + 81, + 223, + 205, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fb9a8f36523583b54143295e7e202e7b6362f22d1ed80c285b642be144e5b77c.jpg", + "image_caption": [ + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 205, + 223, + 290, + 296 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/79e3d4da210bbf932a45af88f201e2d2fdcba66f9a215d34e1a07e1873a40d50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 223, + 377, + 296 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6108bc1888daaeeaca85b93a02fc66c6288af647a327371b8a530e1aab5c9fa9.jpg", + "image_caption": [ + "(c) Rectangular", + "(d) Square" + ], + "image_footnote": [], + "bbox": [ + 377, + 223, + 462, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "two point clouds. We also follow the approach in [26] to evaluate the volumetric occupancy similarity. To do so, the point clouds are voxelized using a voxel size of $0.1m$ . A voxel is classified as occupied for each point cloud if it contains at least one point. We then calculate the Intersection-over-Unit (IoU) based on the occupancy.", + "bbox": [ + 75, + 376, + 468, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 479, + 240, + 493 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Patch size: In Tab. 1 Baseline and Model 1-2, we compare the upsampling results with different patch sizes. We observe that the smaller patch $(2 \\times 2)$ outperforms the larger one $(4 \\times 4)$ while the proposed row patch $(1 \\times 4)$ can further refine the network performance significantly. In particular, the row patch improves $32.6\\%$ in MAE, $12.8\\%$ in IoU, and $28.3\\%$ in CD, compared to a $4 \\times 4$ patch. As visualized in Fig. 3, the row patch better preserves sharpness around edges and corners compared to the larger patch sizes.", + "bbox": [ + 75, + 506, + 468, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Window geometry: To verify the effectiveness of modifying window geometry for range images, we conduct experiments with square and rectangular windows (Tab. 1 Model 2-3. Based on the quantitative results, we can observe the", + "bbox": [ + 76, + 645, + 468, + 705 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a925d8d3961ef23ff474f84556e73bffc17d3c710eff3366fc7aa9c6d1bf3f92.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelPatchWindowCP/PU/PSMCMAE ↓IoU ↑CD ↓
Baseline [5]4 × 4Square××0.71380.32500.1940
Model 12 × 2Square××0.62510.33370.1661
Model 21 × 4Square××0.48140.36670.1391
Model 31 × 4Rectangular××0.42480.40400.1218
Model 41 × 4Rectangular✓×0.42270.40840.1207
TULIP1 × 4Rectangular✓✓0.41850.41740.1207
TULIP-L1 × 4Rectangular✓✓0.37080.43290.0992
ViT-Unet*1 × 4-✓✓1.44840.29480.2665
", + "bbox": [ + 78, + 723, + 470, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Ablation study results for $4\\times$ upsampling on KITTI. Input Resolution: $16\\times 1024$ , Output Resolution: $64\\times 1024$ . *We replaced Swin Transformer blocks with ViT blocks and followed the same training procedure as for TULIP.", + "bbox": [ + 75, + 845, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "following: reconstruction quality produced by local attention within large square windows is inferior to the rectangular window concerning both 2D and 3D evaluation metrics, while they result in the same computational complexity. The underlying reason for this performance difference lies in the perceptive field of the windows. A wider window is more likely to include discontinuities between objects at different scales. Accordingly, the network performs better at separating objects from each other, which can visually be validated by clearer surface boundaries in Fig. 4.", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Further Adaptation and Refinement: Circular Padding (CP) helps maintain continuity and consistency at the edges of the panoramic images, but it has minor contributions to other areas. Patch Unmerging (PU), and Pixel Shuffle (PS) aim at upscaling the spatial resolution with rearrangement, avoiding information loss and parameter increase led by using additional de-convolutional layers. Although they are designed to improve efficiency rather than efficacy of the model, tested with those components (Model 4), a slight improvement in upsampling is still observable. Monte Carlo Dropout [19], as a post-processing step, helps to further reduce ghost points in between objects and we present more details in the Appendix. For TULIP-L, we increase the model capacity, using four Swin-Transformer layers instead of three in both encoder and decoder, which demonstrates an additional performance boost on all metrics.", + "bbox": [ + 496, + 244, + 890, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Transformer Block: From the previous discussion, we infer that Swin Transformer [33] is more advantageous than ViT (Vision Transformer) [17] in LiDAR upsampling. On one hand, prior works [13, 20, 27, 32] have pointed out the inferiority of ViT on smaller datasets due to lack of locality learning and non-overlapping attention. Relevant LiDAR datasets [21, 28] are generally small compared to RGB dataset [41]. On the other hand, by narrowing the patch size, the exponential increase of model capacity hinders sufficient training with ViT on range images, while Swin Transformer expands only linearly. To numerically show the distinction, we trained the network on the same dataset by replacing ViT blocks and applied the same patch size and additional components to the network for a fair comparison. In the last row of Tab. 1, it shows that TULIP significantly outperforms the one with ViT as the backbone.", + "bbox": [ + 496, + 489, + 890, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Benchmark Results", + "text_level": 1, + "bbox": [ + 500, + 741, + 686, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides bilinear interpolation, which computes the interpolation weights from the four nearest neighbors, we evaluate against the state-of-the-art LiDAR upsampling approaches Implicit LiDAR Network (ILN) [26] and LiDAR-SR [43]. We also compare against several image super-resolution works: SRNO [51], and LIIF [10] learn implicit features for interpolation instead of 2D coordinates of new pixels while HAT [9] and Swin-IR [29] are pixel-based approaches that directly reconstruct high-resolution images from low", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15359", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/73ffecc718011df62629a4fef1e3a3186033233b715d03efd24c21b1c8fe3940.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMAE ↓IoU ↑CD ↓
CARLA 4x Output Resolution: 128 × 2048
Bilinear1.81280.13820.7262
SRNO [51]2.46400.13432.0230
HAT [9]1.60320.26980.6337
SWIN-IR [29]1.95600.27180.4840
LIIF [10]0.80640.35020.174
LIDAR-SR [43]0.82160.25810.2044
ILN [26]0.85920.50060.1855
TULIP (Ours)0.76990.51520.1028
TULIP-L (Ours)0.75390.53010.1001
", + "bbox": [ + 76, + 88, + 468, + 273 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c4a4647517818cf9661ca8eef2a4c8cc6b00453556fc47f5ca0aa10b75479bef.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
KITTI 4x Output Resolution: 64 × 1024
Bilinear2.08920.10630.6000
SRNO [51]0.83500.20350.4417
HAT [9]0.68560.20350.2516
SWIN-IR [29]1.29720.27740.7347
LIIF [10]0.61430.32260.1916
LIDAR-SR [43]0.56740.10050.2165
ILN [26]1.05280.33420.2787
TULIP (Ours)0.41850.41740.1207
TULIP-L (Ours)0.37080.43290.0992
", + "bbox": [ + 76, + 275, + 468, + 443 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b1e5ad937115afebcf7e81d4958d105d3df43048a526dfaf39e1cd250a466617.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Quantitative comparison against state-of-the-art LiDAR and image super-resolution methods on different datasets. All methods are trained and evaluated on the same splits." + ], + "table_body": "
DurLAR 4x Output Resolution: 128 × 2048
Bilinear2.43840.12660.6346
SRNO [51]1.53960.15070.5108
HAT [9]1.78200.23530.1973
SWIN-IR [29]1.94160.21570.2279
LIIF [10]1.56720.24690.1548
LIDAR-SR [43]1.53120.13700.1128
ILN [26]1.57200.34300.0893
TULIP (Ours)1.54320.35620.06484
TULIP-L (Ours)1.55920.36540.06346
", + "bbox": [ + 76, + 443, + 468, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "resolution feature embeddings. Similar to our work, they are based on Swin Transformer [33]. We train all methods on the specific datasets using their default parameters.", + "bbox": [ + 75, + 667, + 468, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1 Simulation Results", + "text_level": 1, + "bbox": [ + 76, + 738, + 263, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similar to prior works, we first evaluate the performance on noise-free, simulated data. We train and test all approaches on the dataset presented in [26] that was recorded using the CARLA simulator and select the ground-truth range images of size $128 \\times 2048$ that capture a vertical FoV of $30^{\\circ}$ . We use a (20699/2618) train/test split. We observe that the performance between image super-resolution approaches varies greatly, and LIIF performs the best among them. While LIIF [10] achieves a comparable MAE to the LiDAR up-", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sampling methods, the 3D metrics are significantly worse, indicating that it is unsuitable for the underlying geometric input data. Nevertheless, LIIF [10] still outperforms the Li-DAR upsampling method in LIDAR-SR. The interpolation-based approach in ILN shows decent results for the noise-free data, as indicated by their high IoU. Nevertheless, our approaches achieve the best performance in all evaluated metrics, with the strongest improvement in Chamfer Distance, which validates that our design is well suited for the geometric reconstruction in range images.", + "bbox": [ + 496, + 90, + 893, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df7742b7922515108ac29509113717de52047ce622fdc8a44e8bfea0de0502df.jpg", + "image_caption": [ + "Figure 5. Qualitative results on KITTI: We provide visualizations for the top-4 results from Tab. 2 and Swin-IR [29]. Our approach outperforms other state-of-the-art methods in upsampling realistic point clouds. A: TULIP is the only approach that does not generate invalid ghost points in between the car (right bottom) and the wall behind. It can also be observed that all approaches except for ILN match the characteristic line pattern of the LiDAR on the ground (middle). B: Swin-IR and LIIF generate a blurry reconstruction. TULIP achieves the clearest point cloud and even reconstructs small details such as the car's side mirrors. C: Swin-IR, LIIF, and LiDAR-SR do not clearly separate the car from the wall behind. ILN generates a clear point cloud but does not properly reconstruct the shape of the cars. TULIP achieves the clearest point cloud, with distinct discontinuities between objects, and closely resembles the shape of the three cars." + ], + "image_footnote": [], + "bbox": [ + 496, + 268, + 905, + 700 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15360", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2 Real-World Results", + "text_level": 1, + "bbox": [ + 76, + 90, + 269, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate our approach on real-world data, we downsample the DurLAR and KITTI datasets to a comparable number of frames to the CARLA dataset by temporally skipping frames in the dataset sequences. In particular, we have the following (train/test) split: DurLAR (24372/2508), and KITTI (20000/2500).", + "bbox": [ + 75, + 117, + 468, + 208 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "KITTI: The KITTI dataset [21] was collected using a Velodyne HDL-64E LiDAR with a vertical FoV of $26.8^{\\circ}$ and a resolution of $64 \\times 1024$ . For the image super-resolution approaches, we observe comparable results and trends on KITTI as on the simulated data in Sec.4.3.1. Even though LIDAR-SR [43] is designed for LIDAR upsampling and outperforms image super-resolution approaches in terms of MAE, it achieves the lowest IoU among all approaches, which can be explained by a large amount of invalid floating points between real objects. LIIF [10] achieves better CD and comparable IoU compared to ILN. We observe that the IoU of ILN [26] is drastically lower than on CARLA. The interpolation scheme in ILN is thus not suitable to properly handle the lower resolution data with noise from the real-world sensor. Our approaches significantly outperform all competing methods, especially in terms of IoU as it can better handle discontinuities between objects. As indicated in the visualizations in Fig.5, our approach predicts fewer invalid float points between objects, as well as fewer blurry edges, which results in much clearer point clouds.", + "bbox": [ + 75, + 210, + 470, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DurLAR: The DurLAR dataset [28] was recorded using an Ouster OS1-128 LiDAR, that captures point clouds with a resolution of $128 \\times 2048$ with a $45^{\\circ}$ vertical FoV. Compared to CARLA and KITTI, we see a large performance drop for all approaches in MAE and IoU, which results from the larger maximum range and vertical FoV of the sensor as well as stronger sensor noise. We observe that all approaches designed for LiDAR upsampling clearly improve in Chamfer Distance compared to image super-resolution approaches. ILN and TULIP outperform all methods by a large margin in IoU. TULIP-L achieves the best 3D metrics, which validates that it can also handle the challenging DurLAR data better than prior works.", + "bbox": [ + 75, + 518, + 468, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Range Analysis: As the density of LiDAR point clouds decreases with increasing distance from the sensor, we additionally compare the performance at different ranges for a more fine-grained evaluation. We observe that our approach achieves better or similar performance compared to all approaches in all metrics at all ranges. Up to $30\\mathrm{m}$ , TULIP clearly outperforms the baseline approaches. TULIP handles discontinuities between objects better than other approaches, which can be verified visually through sharper edges and fewer \"floating\" points between objects. This is especially reflected in the high IoU. At ranges above $30\\mathrm{m}$ , our approach performs comparable to other methods.", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f679bc60bb378f808e359d1abf2ee2a83ac8901059ffd4dbbfca9d1bec73be0d.jpg", + "image_caption": [ + "(a) IoU↑" + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 694, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2c3124ec2de544d4df5d2aac0fef9a8d29fd46cad27975674c94019b6e819a8d.jpg", + "image_caption": [ + "(b) Chamfer distance $\\downarrow$", + "Figure 6. 3D error metrics visualized at different ranges for KITTI." + ], + "image_footnote": [], + "bbox": [ + 694, + 88, + 890, + 178 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Above this range, LiDAR data is extremely sparse, meaning that points neighboring in the low-resolution range image are actually distant in the Euclidean space. It is, therefore, not feasible to infer meaningful information about the scene from such neighboring points at a high range, which is also represented in the poor performance of all approaches.", + "bbox": [ + 496, + 224, + 890, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Failure Cases", + "text_level": 1, + "bbox": [ + 500, + 325, + 638, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although TULIP outperforms other state-of-the-art methods both qualitatively and quantitatively, the upsampling quality can still be limited in certain cases. For instance, as shown in Fig 7, our approach shows inferior upsampling results for the specific scene. The irregularity of the scene leads to high uncertainty, which makes the reconstruction noisy. Observing the details in Fig. 7a, the network fails to reconstruct the car at the top of the scene.", + "bbox": [ + 496, + 349, + 890, + 470 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fc987d03c9b37505c3ea807719857ad4aa7a209ce1045ff571c775fc2dd7b00a.jpg", + "image_caption": [ + "(a) Full scene", + "Figure 7. Failure Case: noisy point cloud generated by TULIP and missing reconstruction of the object. (Ours and GT)" + ], + "image_footnote": [], + "bbox": [ + 537, + 489, + 710, + 588 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/07a3bc037bda7aa20ddd231504b007ef05d3b7ece19155f3ba759c7beacf3949.jpg", + "image_caption": [ + "(b) Details" + ], + "image_footnote": [], + "bbox": [ + 736, + 503, + 862, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 662, + 617, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work presents TULIP, a novel method for LiDAR Upsampling that achieves incredible performance in upsampling the range image. Our approach transforms 3D point clouds into 2D range images and performs the upsampling in 2D space. We build upon a Swin-Transformer-based network and specifically modify the patch partition and attention windows to better accommodate the characteristics of range images. Throughout various experiments, testing on three different benchmarks, it shows that TULIP outperforms state-of-the-art methods quantitatively in all evaluation metrics, and quantitatively, TULIP demonstrates more promising results in upsampling the realistic LiDAR data. Acknowledgement: This work was supported by Swiss National Science Foundation's NCCR DFab P3.", + "bbox": [ + 496, + 688, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15361", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anique Akhtar, Zhu Li, Geert Van der Auwera, Li Li, and Jianle Chen. Pu-dense: Sparse tensor-based point cloud geometry upsampling. IEEE Transactions on Image Processing, 31:4133-4148, 2022. 3", + "[2] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. ECCVW What is Motion For?, 2022. 2", + "[3] Borna Bešić, Nikhil Gosala, Daniele Cattaneo, and Abhinav Valada. Unsupervised domain adaptation for lidar panoptic segmentation. IEEE Robotics and Automation Letters, 7(2): 3404-3411, 2022. 1", + "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2", + "[5] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 3, 4, 5, 6", + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2", + "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2", + "[8] Tso Yuan Chen, Ching Chun Hsiao, and Ching-Chun Huang. Density-imbalance-eased lidar point cloud upsampling via feature consistency learning. IEEE Transactions on Intelligent Vehicles, 2022. 3", + "[9] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367-22377, 2023. 2, 4, 6, 7", + "[10] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 6, 7, 8", + "[11] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12312-12321, 2023. 2", + "[12] Xinjing Cheng, Peng Wang, and Ruigang Yang. Depth estimation via affinity learned with convolutional spatial propagation network. In Proceedings of the European conference on computer vision (ECCV), pages 103-119, 2018. 3", + "[13] Yin Dai, Fayu Liu, Weibing Chen, Yue Liu, Lifu Shi, Sheng Liu, Yuhang Zhou, et al. Swin mae: Masked autoencoders" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for small datasets. Computers in Biology and Medicine, 161: 107037, 2023. 6", + "[14] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2", + "[15] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 391-407. Springer, 2016. 2", + "[16] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 5", + "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 6", + "[18] George Eskandar, Sanjeev Sudarsan, Karim Guirguis, Janaranjani Palaniswamy, Bharath Somashekar, and Bin Yang. Hals: A height-aware lidar super-resolution framework for autonomous driving. arXiv preprint arXiv:2202.03901, 2022. 1, 3, 5", + "[19] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pages 1050-1059. PMLR, 2016. 6", + "[20] Hanan Gani, Muzammal Naseer, and Mohammad Yaqub. How to train vision transformer on small-scale datasets? In 33rd British Machine Vision Conference, 2022. 6", + "[21] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 2, 5, 6, 8", + "[22] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Penet: Towards precise and efficient image guided depth completion. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13656-13662. IEEE, 2021. 3", + "[23] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2, 2019. 2", + "[24] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 2", + "[25] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1637-1645, 2016. 2", + "[26] Youngsun Kwon, Minhyuk Sung, and Sung-Eui Yoon. Implicit lidar network: Lidar super-resolution via interpolation weight prediction. In 2022 International Conference on" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15362", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Robotics and Automation (ICRA), pages 8424-8430. IEEE, 2022. 1, 2, 3, 5, 6, 7, 8", + "[27] Seung Hoon Lee, Seunghyun Lee, and Byung Cheol Song. Vision transformer for small-size datasets. arXiv preprint arXiv:2112.13492, 2021. 6", + "[28] L. Li, K.N. Ismail, H.P.H. Shum, and T.P. Breckon. Durlar: A high-fidelity 128-channel lidar dataset with panoramic ambient and reflectivity imagery for multi-modal autonomous driving applications. In Proc. Int. Conf. on 3D Vision. IEEE, 2021. 2, 5, 6, 8", + "[29] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 1, 2, 3, 4, 5, 6, 7", + "[30] Yuankai Lin, Tao Cheng, Qi Zhong, Wending Zhou, and Hua Yang. Dynamic spatial propagation network for depth completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1638-1646, 2022. 3", + "[31] Zhixin Ling, Zhen Xing, Xiangdong Zhou, Manliang Cao, and Guichun Zhou. Panoswin: A pano-style swin transformer for panorama understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2023. 4", + "[32] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco De Nadai. Efficient training of visual transformers with small datasets. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 6", + "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 4, 5, 6, 7", + "[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 5", + "[35] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457-466, 2022. 2, 4", + "[36] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 120-136. Springer, 2020. 3", + "[37] Jiaxiong Qiu, Zhaopeng Cui, Yinda Zhang, Xingdi Zhang, Shuaicheng Liu, Bing Zeng, and Marc Pollefeys. Deepidar: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3313-3322, 2019. 3", + "[38] Shi Qiu, Saeed Anwar, and Nick Barnes. Pu-transformer: Point cloud upsampling transformer. In Proceedings of the Asian Conference on Computer Vision, pages 2475-2493, 2022. 3", + "[39] Kyeongha Rho, Jinsung Ha, and Youngjung Kim. Guideformer: Transformers for image guided depth completion." + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6250-6259, 2022. 3", + "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3", + "[41] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 6", + "[42] Artem Savkin, Yida Wang, Sebastian Wirkert, Nassir Navab, and Federico Tombari. Lidar upsampling with sliced wasserstein distance. IEEE Robotics and Automation Letters, 8(1): 392-399, 2022. 1, 3", + "[43] Tixiao Shan, Jinkun Wang, Fanfei Chen, Paul Szenher, and Brendan Englot. Simulation-based lidar super-resolution for ground vehicles. Robotics and Autonomous Systems, 134: 103647, 2020. 1, 2, 3, 5, 6, 7, 8", + "[44] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 4, 5", + "[45] Xiaopeng Sun, Weiqi Li, Zhenyu Zhang, Qiufang Ma, Xuhan Sheng, Ming Cheng, Haoyu Ma, Shijie Zhao, Jian Zhang, Junlin Li, et al. Opdn: Omnidirectional position-aware deformable network for omnidirectional image superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1293-1301, 2023. 2", + "[46] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing, 30:1116-1129, 2020. 3", + "[47] Larissa T Triess, David Peter, Christoph B Rist, Markus Enzweiler, and J Marius Zöllner. Cnn-based synthesis of realistic high-resolution lidar data. In 2019 IEEE Intelligent Vehicles Symposium (IV), pages 1512-1519. IEEE, 2019. 3", + "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[49] Suhong Wang, Hongqing Wang, Shufeng She, Yanping Zhang, Qingju Qiu, and Zhifeng Xiao. Swin-t-nfc crfs: An encoder-decoder neural model for high-precision UAV positioning via point cloud super resolution and image semantic segmentation. Computer Communications, 197:52–60, 2023. 1", + "[50] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 5" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15363", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Min Wei and Xuesong Zhang. Super-resolution neural operator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18247-18256, 2023. 6, 7", + "[52] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Baobei Xu, Jun Li, and Jian Yang. Rignet: Repetitive image guided network for depth completion. In European Conference on Computer Vision, 2021. 3", + "[53] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1951-1960, 2019. 1", + "[54] Shuquan Ye, Dongdong Chen, Songfang Han, Ziyu Wan, and Jing Liao. Meta-pu: An arbitrary-scale upsampling network for point cloud. IEEE transactions on visualization and computer graphics, 28(9):3206-3218, 2021. 3", + "[55] Fanghua Yu, Xintao Wang, Mingdeng Cao, Gen Li, Ying Shan, and Chao Dong. Osrt: Omnidirectional image superresolution with distortion-aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13283-13292, 2023. 2", + "[56] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-net: Point cloud upsampling network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2790-2799, 2018. 1, 3", + "[57] Heeseung Yun, Sehun Lee, and Gunhee Kim. Panoramic vision transformer for saliency detection in 360 videos. In ECCV, 2022. 4", + "[58] Wentai Zhang, Haoliang Jiang, Zhangsihao Yang, Soji Yamakawa, Kenji Shimada, and Levent Burak Kara. Data-driven upsampling of point clouds. Computer-Aided Design, 112:1-13, 2019. 1", + "[59] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image restoration. TPAMI, 2020. 2", + "[60] Wenbo Zhao, Xianming Liu, Zhiwei Zhong, Junjun Jiang, Wei Gao, Ge Li, and Xiangyang Ji. Self-supervised arbitrary-scale point clouds upsampling via implicit neural representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1999-2007, 2022. 3", + "[61] Yifan Zhao, Le Hui, and Jin Xie. Sspu-net: Self-supervised point cloud upsampling via differentiable rendering. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2214-2223, 2021. 1, 3", + "[62] Örkény Zováthi, Balázs Pálffy, Zsolt Jankó, and Csaba Benedek. St-depthnet: A spatio-temporal deep network for depth completion using a single non-repetitive circular scanning lidar. IEEE Robotics and Automation Letters, 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 782 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15364", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/52ac8b22-207c-41ff-adcb-c9f4c82bca08_model.json b/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/52ac8b22-207c-41ff-adcb-c9f4c82bca08_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c0464c20df4c07b8769ee62e5a087a80bf735e02 --- /dev/null +++ b/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/52ac8b22-207c-41ff-adcb-c9f4c82bca08_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.176, + 0.131, + 0.795, + 0.154 + ], + "angle": 0, + "content": "TULIP: Transformer for Upsampling of LiDAR Point Clouds" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.18, + 0.968, + 0.234 + ], + "angle": 0, + "content": "Bin Yang\\(^{1}\\), Patrick Pfreundschuh\\(^{1}\\), Roland Siegwart\\(^{1}\\), Marco Hutter\\(^{2}\\), Peyman Moghadam\\(^{3,4\\dagger}\\), Vaishakh Patil\\(^{2\\dagger}\\) \n\\(^{1}\\)Autonomous Systems Lab, ETH Zurich, Switzerland, \\(^{2}\\)Robotic Systems Lab, ETH Zurich, Switzerland \n\\(^{3}\\)School of EER, Queensland University of Technology (QUT), Australia, \\(^{4}\\)Data61, CSIRO, Brisbane, Australia" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.237, + 0.871, + 0.252 + ], + "angle": 0, + "content": "{biyang, patripfr, rolandsi, mahutter, patilv}@ethz.ch, peyman.moghadam@csiro.au" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.319, + 0.474, + 0.621 + ], + "angle": 0, + "content": "LiDAR Upsampling is a challenging task for the perception systems of robots and autonomous vehicles, due to the sparse and irregular structure of large-scale scene contexts. Recent works propose to solve this problem by converting LiDAR data from 3D Euclidean space into an image super-resolution problem in 2D image space. Although their methods can generate high-resolution range images with fine-grained details, the resulting 3D point clouds often blur out details and predict invalid points. In this paper, we propose TULIP, a new method to reconstruct high-resolution LiDAR point clouds from low-resolution LiDAR input. We also follow a range image-based approach but specifically modify the patch and window geometries of a Swin-Transformer-based network to better fit the characteristics of range images. We conducted several experiments on three public real-world and simulated datasets. TULIP outperforms state-of-the-art methods in all relevant metrics and generates robust and more realistic point clouds than prior works. The code is available at https://github.com/ethz-as1/TULIP.git." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.653, + 0.21, + 0.669 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.678, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Light Detection And Ranging (LiDAR) is one of the most common sensors for perception in various fields of autonomy, such as autonomous driving, and unmanned aerial vehicles (UAV). LiDARs are used to generate 3D point clouds of the scene. These point clouds are essential for mapping, localization, and object detection tasks. However, the accuracy of these tasks often depends on the resolution of the point cloud [53]. Furthermore, the resolution of a LiDAR is inherently associated with increased energy consumption and cost, making its use impractical for various applications. LiDARs also have different vertical and horizontal resolutions. The vertical resolution of rotating 3D LiDARs is typically much lower than the horizontal res" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.286, + 0.697, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.389, + 0.625, + 0.401 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.288, + 0.886, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.389, + 0.838, + 0.4 + ], + "angle": 0, + "content": "(b) Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.406, + 0.695, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.504, + 0.642, + 0.516 + ], + "angle": 0, + "content": "(c) SwinIR [29]" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.404, + 0.885, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.504, + 0.837, + 0.516 + ], + "angle": 0, + "content": "(d) TULIP (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.518, + 0.892, + 0.573 + ], + "angle": 0, + "content": "Figure 1. TULIP generates more realistic LiDAR point clouds from low-resolution LiDAR input (a) and outperforms the state-of-the-art image super-resolution approach [29] in upsampling with fine-grained details of different objects in the scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.765 + ], + "angle": 0, + "content": "olution. These limitations make an upsampling technique necessary to increase the resolution of the LiDAR data, especially in the vertical direction. Moreover, upsampling LiDAR data has the potential to counter domain shift problems in LiDAR-based learning methods. Notably, the use of lower resolution LiDAR data in a system that has been trained on higher resolution data leads to a significant drop in performance [3]. Therefore, such upsampling techniques can not only help to mitigate the domain gap by creating a virtual high-resolution sensor that matches the target domain [42, 61] but also reduce the high cost of collecting new LiDAR data, annotating and retraining the methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Deep learning has led to remarkable advancements in LiDAR upsampling techniques in recent years. Several recent methods [42, 56, 58] have focused on learning the upsampling process within the 3D Euclidean space. However, the processing of 3D data in deep learning can be quite resource-intensive. One way to address this challenge is the representation of the point cloud as a range image [18, 26, 43, 49]. This approach facilitates the adoption of well-established image-based super-resolution methods." + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.262, + 0.901 + ], + "angle": 0, + "content": "†Authors share last authorship." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15354" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "Such methods mostly employ autoencoder-style networks with convolutional and deconvolutional layers [14, 43, 62]. However, these methods are not directly transferable to range images. Range images generated by the projection of sparse 3D point clouds have distinct sharp edges at the object boundaries. Convolutional neural network-based methods often induce edge smoothing due to the regularization effect, which limits their usability for LiDAR range image processing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.226, + 0.473, + 0.894 + ], + "angle": 0, + "content": "As an alternative, Transformer [48] has shown great success in RGB image super-resolution and produces less blurry outputs [29]. Vision transformers, however, face challenges such as the need for large amounts of training data [9, 29, 35] and the global computation of self-attention, limiting the capture of local information in the range images. To address this, Swin Transformer [33] introduced tokenizing the input image and applying the self-attention mechanism locally in independent windows. While the modifications to Swin Transformer have led to robust training and promising results compared to Vision Transformer, its default structure and settings are not ideally suited for processing LiDAR range images. These range images consist of single-channel data, representing 3D spatial information, as opposed to three-channel color images that depict visual appearance. Furthermore, range images feature large, smooth areas within 3D objects, separated by distinct sharp edges between objects. As a result, minor inaccuracies in the 2D prediction (e.g. blurry output) can lead to significant differences in the projected 3D occupancy. This non-uniform distribution of image details further complicates network training. The challenging details in the range image are concentrated in a few specific pixels, contrasting with RGB images where relevant details are evenly distributed. This makes range images fundamentally different from RGB images. Inspired by the limitations of state-of-the-art image upsampling methods and the potential of Swin-Transformer [33], we propose a novel network named TULIP. Our proposed geometry-aware architecture is tailored to accommodate LiDAR range image data. The proposed method takes a low-resolution range image as input and produces a high-resolution range image, which can then be projected into a 3D point cloud. To accommodate LiDAR range image geometry, we utilize one-line row patches to tokenize the input range image, as opposed to the square patches used for RGB images. This approach helps to preserve vertical information for the upsampling process while capturing boundary discontinuities effectively. This also creates a trade-off between local detail capture and model complexity in the training process. Additionally, we incorporate a non-square window for local self-attention computation. This enhances learning of spatial contexts at different scales in the range image. Further, we extensively train and evaluate TULIP on two real-world and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "one simulated autonomous driving datasets: KITTI [21], DurLAR [28] and CARLA [26], respectively. Comparisons to existing approaches demonstrate that TULIP outperforms the state-of-the-art methods on all three datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.163, + 0.651, + 0.181 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.192, + 0.892, + 0.312 + ], + "angle": 0, + "content": "Image super-resolution has seen significant progress in recent years. The technique aims to construct high-resolution (HR) images from low-resolution (LR) observations, often leveraging advances in Convolutional Neural Networks (CNNs) to enhance the fidelity and detail of images for better visualization and information extraction. Subsequently, several enhanced frameworks have been developed [14, 15, 24, 25, 59]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.892, + 0.523 + ], + "angle": 0, + "content": "Driven by their success in the field of natural language processing (NLP) [4, 23], Transformers [48] have been extended to solve a variety of vision-related tasks such as object detection [6] and semantic segmentation [2, 7]. Such Vision Transformers (ViT) [17] excel at learning to focus on relevant image regions by exploring global interactions among different regions. Their impressive performance has led to their adoption in image restoration tasks [9, 11, 29] as well. While they brought significant improvements in RGB image super-resolution, Vision Transformer [17] comes with the drawback of the quadratic computational complexity of self-attention. Additionally, they capture mostly global dependencies within the data and require large amounts of data for training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.903 + ], + "angle": 0, + "content": "To address these challenges, Swin Transformer [33] has been proposed. Unlike the Vision Transformer, which relies on global self-attention across the entire image, the Swin Transformer employs a local window-based attention mechanism and establishes a pyramid-like architecture. This approach processes images hierarchically by gradually merging smaller image patches into larger ones. This strategy enables more effective handling of different scales and facilitates the processing of multi-scale features, making it particularly suitable for range image processing. Furthermore, due to Swin Transformer's hierarchical structure, the increase in the number of model parameters scales linearly with the input image size, which mitigates the challenges associated with handling large range images. Additionally, they require less training data than the classical Vision Transformer, which is beneficial for LiDAR upsampling as LiDAR datasets are typically magnitudes smaller than RGB datasets used for Vision Transformer. The Swin and Vision Transformers have also been used for image super-resolution on omnidirectional panoramic camera images [45, 55]. However, these approaches are specifically designed for the unique characteristics of these cameras (e.g. strong distortions, spherical continuity) that are vastly different from the LiDAR sensing model, which renders them unsuitable for application to range images." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15355" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.88 + ], + "angle": 0, + "content": "Other related works have focussed on upsampling high-density point clouds from low-density point clouds [38, 54, 56, 60, 61]. While those approaches also provide qualitative results for LiDAR point clouds, they focus on increasing the general 3D point density of point clouds of single objects instead of a large scene. Another area closely related to LiDAR upsampling is depth completion [12, 22, 30, 36, 37, 39, 46, 52]. The primary objective of depth completion is to improve sparse depth estimates acquired from LiDAR by integrating data from multiple sensors, predominantly RGB cameras. The result of depth completion methods is a dense depth map that provides depth values to every pixel in the input depth map. Nonetheless, this field slightly diverges from LiDAR upsampling due to its reliance on multimodal sensors and its common deployment within a restricted field of view, making it not directly comparable within the scope of this paper. LiDAR point clouds exhibit specific characteristics, such as a distinct point pattern (due to the stacked lasers in rotating 3D LiDARs) and a decrease in point density with increasing distance from the sensor. Differently from arbitrary point cloud upsampling, LiDAR upsampling tries to mimic a realistic point cloud of a high-resolution LiDAR given the point cloud of low-resolution LiDAR and thus targets a different result than the approaches above. Due to the difficulty and high computational cost of upsampling LiDAR point clouds in 3D space [1, 8, 42], most works [26, 43, 47] represent the point cloud as a range image and perform the LiDAR upsampling task in 2D image space. LiDAR-CNN [47] introduces a CNN-based architecture with semantic and perceptual guidance. The method specifies two further loss functions besides a per-point reconstruction loss to acquire a better synthesis of the high-resolution range image. However, this approach is limited to data with semantic annotations, which drastically reduces its applicability. LiDAR-SR [43] deploys a CNN-based network with U-Net [40] architecture. It additionally uses Monte-Carlo dropout post-processing to reduce the amount of noisy points in the prediction. Approaches based on convolutional operations tend to fail to reconstruct the sharpness in the range image effectively. While the previous approaches directly predict the range image, Implicit LiDAR Network (ILN) [26] uses an implicit neural architecture that learns interpolation weights to fill in new pixels instead of their values directly. Although the method outperforms CNN-based approaches in training speed and preserving the geometrical details in the input, it can still suffer from limited neighboring information in the LiDAR data, especially in areas distant from the sensor due to the extremely sparse input data. Differently from the previous works, our approach builds on Swin Transformer as a backbone and is specifically customized to effectively process range images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.637, + 0.109 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.118, + 0.895, + 0.635 + ], + "angle": 0, + "content": "LiDAR upsampling is achieved by upsampling a range image, effectively transforming it from a 3D upsampling problem into a 2D image super-resolution problem [18, 26, 43, 47]. It is, therefore, an evident approach to build on RGB image super-resolution. However, image super-resolution aims to solve a different problem than LiDAR upsampling. Image super-resolution enhances the visual appearance of a low-resolution RGB image by recreating a high-resolution image. While quantitative metrics for evaluating these works exist, there is no one specific correct solution. Differently, LiDAR upsampling tries to recreate a high-resolution, 3D LiDAR sensor output from a low-resolution LiDAR input point cloud. Range images contain more geometrical and spatial contexts rather than visual information, and the scene's geometry strictly dictates the correct solution. The underlying sensing model is also drastically different, as LiDARs are active sensors. The respective range pixel values purely depend on distance, while pixels in an RGB camera depend on many factors, such as scene appearance, lighting conditions, exposure time, and white balance. Furthermore, range images have a highly asymmetric aspect ratio, typically between 1:8 to 1:64, which differs drastically from regular camera images. Accordingly, most works [18, 43, 47] only perform vertical upsampling, in contrast to image super-resolution, which operates in both dimensions. A naive application of state-of-the-art image super-resolution networks [29] to range images does not result in adequate performance (Fig. 1). To this end, we developed TULIP, a novel range image-based LiDAR Upsampling method. Based on the observations above, we specifically adapt a network based on SwinUnet [5] to incorporate range images better. In the following sections, we will describe the technical details of our method, with a special focus on how our approach accommodates range images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.685, + 0.665 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "A LiDAR point cloud consists of points captured during one revolution \\(\\mathcal{P} = \\{\\mathbf{p}_1,\\dots ,\\mathbf{p}_n\\}\\). Each measurement represents a 3D point \\(\\mathbf{p}_i = \\{x_i,y_i,z_i\\}\\). As input, we assume a low-resolution point cloud \\(\\mathcal{P}_l\\) with \\(n_l = H_l\\times W_l\\) points, where \\(H_{l}\\) and \\(W_{l}\\) correspond to vertical and horizontal resolution. We aim to predict a high-resolution point cloud \\(\\mathcal{P}_h\\), that has the same field-of-view(FoV) as \\(\\mathcal{P}_l\\) but contains \\(n_h = \\beta *n_l\\) points. In our experiments, we set \\(\\beta = 4\\). We only increase vertical resolution (amount of LiDAR beams), i.e., \\(n_h = H_h\\times W_h\\), where \\(H_{h} = \\beta *H_{l}\\) and \\(W_{h} = W_{l}\\). We project the point cloud into a 2D range image. In this image, each row and column coordinate \\(v,u\\) correspond to the respective elevation and azimuth angles of the LiDAR points, while the pixel value contains the range of the point \\(r = \\sqrt{x^2 + y^2 + z^2}\\). The image coordinate of a 3D point" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "15356" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.429, + 0.108 + ], + "angle": 0, + "content": "can be calculated using a spherical projection model:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.114, + 0.47, + 0.158 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{l} u \\\\ v \\end{array} \\right] = \\left[ \\frac {\\frac {W}{2} - \\frac {W}{2 \\pi} a r c t a n (\\frac {y}{x})}{\\Theta_ {m a x} - \\Theta_ {m i n}} * (\\Theta_ {m a x} - a r c t a n (\\frac {z}{\\sqrt {x ^ {2} + y ^ {2}}}) \\right] (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.164, + 0.47, + 0.271 + ], + "angle": 0, + "content": "\\(\\Theta_{max}\\) and \\(\\Theta_{min}\\) correspond to the limits of the vertical field of view of the LiDAR. We formulate the LiDAR upsampling problem as an image upsampling problem, i.e. we aim to predict a high-resolution range image \\(I_h \\in \\mathbf{R}^{1 \\times H_h \\times W}\\) given a low-resolution range image \\(I_l \\in \\mathbf{R}^{1 \\times H_l \\times W}\\). We can then calculate the high-resolution point cloud \\(\\mathcal{P}_h\\) by inverting the projection in Eq. 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.278, + 0.379, + 0.294 + ], + "angle": 0, + "content": "3.2. Architecture Design and Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.471, + 0.832 + ], + "angle": 0, + "content": "Our network builds upon Swin-Unet [5] which was originally designed for image segmentation. We deploy a U-shaped network structure featuring skip connections that link the encoder and decoder modules. The fundamental building block of TULIP is the Swin Transformer [33]. TULIP first patchifies and maps the input to a high-dimensional feature space. The tokenized low-level feature maps are passed through a multi-stage encoder. Each stage contains a pair of Swin Transformer [33] blocks and a patch merging layer to downsample the resolution by a factor of two and increase the dimension by a factor of four. This step realizes the hierarchic computation and extraction of multi-scale features through encoding. Within each stage, multiple instances of the self-attention mechanism are performed locally in parallel, which is the so-called Window Multi-Head Self Attention (W-MSA). After passing through a two-layer MLP and residual connection with the input, the feature vector is further passed to the second part of the Swin Transformer block, utilizing a Shifted Window Multi-Head Self Attention (SW-MSA) which extends W-MSA by employing a shifted window operation. This enhances the model capacity as it compensates for the lack of interaction between the local windows in W-MSA. The decoder, which realizes the upsampling has a symmetric design to the encoder and operates in a reverse way of the encoder. The resolution of the feature maps is first expanded, and the dimension is reduced by a factor of two accordingly. We feed existing geometrical information via skip connections. The feature maps are transformed into a single-channel, high-resolution range image in the last layer. The reconstruction head comprises a \\(1 \\times 1\\) convolutional layer for feature expansion, followed by a Leaky ReLU activation, a pixel shuffle layer [44], and another \\(1 \\times 1\\) convolutional layer for final projection. As the loss function, we select the pixel-wise L1 loss. Details of the network can be found in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.221, + 0.862 + ], + "angle": 0, + "content": "3.2.1 Tokenization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Tokenization is done by a patch partition layer that creates an initial feature embedding from the input range image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.424 + ], + "angle": 0, + "content": "This initial feature representation significantly influences the network's performance. Specifically, the selection of this layer determines how the inherent characteristics, patterns, and relevant information of the LiDAR range image data are encoded into a format that the network can effectively learn from. Most transformer-based RGB image super-resolution approaches utilize a relatively large patch size to tokenize the input [9, 29, 35]. These larger patches enable the construction of a global spatial context, facilitating an understanding of the overall structure and high-level features of the RGB image. These networks employ square-shaped patches and implement upsampling in both vertical and horizontal directions. In contrast, LiDAR upsampling primarily aims to enhance the vertical resolution of the input, given that range images possess properties that are vastly different from RGB images. Furthermore, attention across more pixels in range images is less useful than in RGB images, since there is almost no geometrical relation for the spatial contexts that are far from each other in the 2D range image space. Motivated by this, we propose two adjustments to effectively process range images within a Swin-Unet:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.892, + 0.625 + ], + "angle": 0, + "content": "Row-Based Patch Partition: Our model builds row patches with a dimension of \\(1 \\times 4\\) for range image tokenization. The new patch geometry is designed to retain full vertical information while compressing horizontally. This aligns with our objective of conserving and extending vertical details in the range image. Besides that, row patches excel at capturing boundary discontinuities between distinct objects and the background within the scene. Moreover, the selection of the patch size strikes a balance between enhancing model performance and limiting model capacity. Smaller patches can capture more detailed local information, but they also increase the number of model parameters, which can slow down the model's training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Circular Padding (CP) in Horizontal Dimension: Instead of adding zero pixels padding, our model uses circular padding in the horizontal dimension. This avoids introducing artificial features and naturally matches the sensor model of a rotating 3D LiDAR. This preserves accurate neighbor relations along the edges, which would otherwise be compromised during the projection to the 2D range image. The Circular Padding (CP) also allows us to fit flexible input sizes without changing the patch geometry." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.786, + 0.854, + 0.801 + ], + "angle": 0, + "content": "3.2.2 Non-Square window for local self-attention" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Inside Swin Transformer, input feature maps are window-partitioned along width and height to compute the self-attention locally in each window. In addition to square-shaped RGB images, some recent work applies a square window for panoramic data as well [31, 57]. However, using square windows on range images is unfavorable. On one" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15357" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.087, + 0.835, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.268, + 0.895, + 0.34 + ], + "angle": 0, + "content": "Figure 2. Overview of TULIP: The low-resolution point cloud is spherically projected into the range image. In the training and inference, the range image is tokenized into feature embeddings without compressing the spatial resolution along the vertical axis. We adopt the U-Net-based network presented in [5] for range image upsampling. We follow the hierarchical structure of Swin Transformer [33] for feature extraction and use the non-square window geometry for computing the local self-attention. The reconstruction head generates the high-resolution range image from the latent features and then, the point cloud can be obtained by projecting pixels back into the 3D space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.473, + 0.699 + ], + "angle": 0, + "content": "hand, small windows can focus more on attention within near areas but lead to increasing computational complexity and can struggle with the reconstruction of objects at a further distance in the scene. On the other hand, a larger square window can blend scene-related contexts at different scales, which can potentially degrade the network performance. In a prior study done by Eskandar et al. [18], they demonstrated a significant enhancement in range image superresolution by separately upsampling the upper and lower portions of the data. In contrast to their approach of utilizing CNN-based shallower and deeper network branches for distinct feature extraction, our proposal involves the direct partitioning of the input using a rectangular window. This window geometry facilitates robust attention interaction among distant points through W-MSA and subtle crossattention between objects at multiple scales via SW-MSA and hierarchical processing. In addition, range images, as typical panorama data, inherently contain abundant information along the horizontal direction. Therefore, directing attention more toward image width rather than height can aid the network in capturing a greater amount of geometrical information. At the same time, the computational complexity remains the same as using a large square window." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.722, + 0.383, + 0.737 + ], + "angle": 0, + "content": "3.2.3 Further Adaptation and Refinement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Patch Unmerging (PU) and Pixel Shuffle (PS): Similar to the patch merging layer that is responsible for down-sampling at the encoder stage, a mechanism is necessary within the decoder to upsample the feature maps. Conventional CNN-based approaches [43, 50] achieve upsampling through the transposed convolutional operation. As they tend to smooth out image sharpness [29], we reversed the operation of patch merging, re-arranging the channels of patches into their respective positions in a grid to up-sample the spatial resolution of input feature maps, which we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.412 + ], + "angle": 0, + "content": "denote as patch unmerging. Furthermore, to reconstruct the final range image, we built the reconstruction head upon the upsampled feature maps. The implementation of the reconstruction module is based on the pixel shuffle layer [44]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.893, + 0.52 + ], + "angle": 0, + "content": "Monte Carlo Dropout: We additionally use Monte Carlo Dropout [43] to filter unreliable points by thresholding the uncertainty. MC-Dropout executes several feed-forward passes during inference with different active dropouts, which yields a distribution of outputs. We refine the results by removing points with high output variance, as they often come down to noisy, invalid points." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.538, + 0.634, + 0.556 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.564, + 0.707, + 0.58 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.892, + 0.743 + ], + "angle": 0, + "content": "Datasets: We conduct experiments on three different datasets, that include the two large-scale real-world datasets DurLAR [28] and KITTI [21] (Sec. 4.3.2), as well as on a dataset [26] that was generated using the CARLA simulator [16] (Sec. 4.3.1). We select test sequences that are recorded in different locations than the train set to avoid spatial overlap. We vertically downsample the high-resolution range images with a factor of four by skipping the respective lines to generate low-resolution input images. For ablation studies in Sec. 4.2, we select the KITTI dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Implementation Details: In all experiments, we use distributed processing with \\(4 \\times\\) GeForce RTX 2080 Ti. For training, we use a fixed batch size of eight per GPU for all datasets. We use the AdamW [34] optimizer with a weight decay factor of 0.01 and a base learning rate of 5e-4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Evaluation metrics: We evaluate the Mean Absolute Error (MAE) for all the pixels in the generated 2-D range images. Additionally, we assess the performance by considering the 3D points reconstructed by our neural networks. The Chamfer Distance (CD) measures the Euclidean distance between" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15358" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.188, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.152, + 0.16, + 0.163 + ], + "angle": 0, + "content": "(a) Scene" + }, + { + "type": "image", + "bbox": [ + 0.189, + 0.089, + 0.258, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.152, + 0.243, + 0.163 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.089, + 0.326, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.152, + 0.317, + 0.163 + ], + "angle": 0, + "content": "(c) \\(1 \\times 4\\)" + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.089, + 0.395, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.152, + 0.387, + 0.163 + ], + "angle": 0, + "content": "(d) \\(2 \\times 2\\)" + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.089, + 0.466, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.152, + 0.455, + 0.163 + ], + "angle": 0, + "content": "(e) \\(4 \\times 4\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.169, + 0.469, + 0.224 + ], + "angle": 0, + "content": "Figure 3. Upsampling results with different patch sizes: We observe that square patches \\((4 \\times 4, 2 \\times 2)\\) blur out edges and create invalid connections between separate walls, while the proposed \\(1 \\times 4\\) patch generates sharp edges and correct discontinuities." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.224, + 0.206, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.3, + 0.169, + 0.311 + ], + "angle": 0, + "content": "(a) Scene" + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.224, + 0.291, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.3, + 0.269, + 0.311 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.224, + 0.378, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.294, + 0.3, + 0.376, + 0.312 + ], + "angle": 0, + "content": "(c) Rectangular" + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.224, + 0.463, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.3, + 0.448, + 0.312 + ], + "angle": 0, + "content": "(d) Square" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.314, + 0.469, + 0.369 + ], + "angle": 0, + "content": "Figure 4. Upsampling results with different window geometries: Rectangular windows improve reconstruction of objects at different ranges, indicated by more distinct discontinuities (top) and a better reconstruction of the car (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.469, + 0.467 + ], + "angle": 0, + "content": "two point clouds. We also follow the approach in [26] to evaluate the volumetric occupancy similarity. To do so, the point clouds are voxelized using a voxel size of \\(0.1m\\). A voxel is classified as occupied for each point cloud if it contains at least one point. We then calculate the Intersection-over-Unit (IoU) based on the occupancy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.48, + 0.241, + 0.494 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.469, + 0.643 + ], + "angle": 0, + "content": "Patch size: In Tab. 1 Baseline and Model 1-2, we compare the upsampling results with different patch sizes. We observe that the smaller patch \\((2 \\times 2)\\) outperforms the larger one \\((4 \\times 4)\\) while the proposed row patch \\((1 \\times 4)\\) can further refine the network performance significantly. In particular, the row patch improves \\(32.6\\%\\) in MAE, \\(12.8\\%\\) in IoU, and \\(28.3\\%\\) in CD, compared to a \\(4 \\times 4\\) patch. As visualized in Fig. 3, the row patch better preserves sharpness around edges and corners compared to the larger patch sizes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.646, + 0.469, + 0.707 + ], + "angle": 0, + "content": "Window geometry: To verify the effectiveness of modifying window geometry for range images, we conduct experiments with square and rectangular windows (Tab. 1 Model 2-3. Based on the quantitative results, we can observe the" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.724, + 0.472, + 0.835 + ], + "angle": 0, + "content": "
ModelPatchWindowCP/PU/PSMCMAE ↓IoU ↑CD ↓
Baseline [5]4 × 4Square××0.71380.32500.1940
Model 12 × 2Square××0.62510.33370.1661
Model 21 × 4Square××0.48140.36670.1391
Model 31 × 4Rectangular××0.42480.40400.1218
Model 41 × 4Rectangular✓×0.42270.40840.1207
TULIP1 × 4Rectangular✓✓0.41850.41740.1207
TULIP-L1 × 4Rectangular✓✓0.37080.43290.0992
ViT-Unet*1 × 4-✓✓1.44840.29480.2665
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.847, + 0.469, + 0.904 + ], + "angle": 0, + "content": "Table 1. Ablation study results for \\(4\\times\\) upsampling on KITTI. Input Resolution: \\(16\\times 1024\\), Output Resolution: \\(64\\times 1024\\). *We replaced Swin Transformer blocks with ViT blocks and followed the same training procedure as for TULIP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "following: reconstruction quality produced by local attention within large square windows is inferior to the rectangular window concerning both 2D and 3D evaluation metrics, while they result in the same computational complexity. The underlying reason for this performance difference lies in the perceptive field of the windows. A wider window is more likely to include discontinuities between objects at different scales. Accordingly, the network performs better at separating objects from each other, which can visually be validated by clearer surface boundaries in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.892, + 0.486 + ], + "angle": 0, + "content": "Further Adaptation and Refinement: Circular Padding (CP) helps maintain continuity and consistency at the edges of the panoramic images, but it has minor contributions to other areas. Patch Unmerging (PU), and Pixel Shuffle (PS) aim at upscaling the spatial resolution with rearrangement, avoiding information loss and parameter increase led by using additional de-convolutional layers. Although they are designed to improve efficiency rather than efficacy of the model, tested with those components (Model 4), a slight improvement in upsampling is still observable. Monte Carlo Dropout [19], as a post-processing step, helps to further reduce ghost points in between objects and we present more details in the Appendix. For TULIP-L, we increase the model capacity, using four Swin-Transformer layers instead of three in both encoder and decoder, which demonstrates an additional performance boost on all metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Transformer Block: From the previous discussion, we infer that Swin Transformer [33] is more advantageous than ViT (Vision Transformer) [17] in LiDAR upsampling. On one hand, prior works [13, 20, 27, 32] have pointed out the inferiority of ViT on smaller datasets due to lack of locality learning and non-overlapping attention. Relevant LiDAR datasets [21, 28] are generally small compared to RGB dataset [41]. On the other hand, by narrowing the patch size, the exponential increase of model capacity hinders sufficient training with ViT on range images, while Swin Transformer expands only linearly. To numerically show the distinction, we trained the network on the same dataset by replacing ViT blocks and applied the same patch size and additional components to the network for a fair comparison. In the last row of Tab. 1, it shows that TULIP significantly outperforms the one with ViT as the backbone." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.687, + 0.756 + ], + "angle": 0, + "content": "4.3. Benchmark Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Besides bilinear interpolation, which computes the interpolation weights from the four nearest neighbors, we evaluate against the state-of-the-art LiDAR upsampling approaches Implicit LiDAR Network (ILN) [26] and LiDAR-SR [43]. We also compare against several image super-resolution works: SRNO [51], and LIIF [10] learn implicit features for interpolation instead of 2D coordinates of new pixels while HAT [9] and Swin-IR [29] are pixel-based approaches that directly reconstruct high-resolution images from low" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15359" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.47, + 0.274 + ], + "angle": 0, + "content": "
ModelMAE ↓IoU ↑CD ↓
CARLA 4x Output Resolution: 128 × 2048
Bilinear1.81280.13820.7262
SRNO [51]2.46400.13432.0230
HAT [9]1.60320.26980.6337
SWIN-IR [29]1.95600.27180.4840
LIIF [10]0.80640.35020.174
LIDAR-SR [43]0.82160.25810.2044
ILN [26]0.85920.50060.1855
TULIP (Ours)0.76990.51520.1028
TULIP-L (Ours)0.75390.53010.1001
" + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.276, + 0.47, + 0.444 + ], + "angle": 0, + "content": "
KITTI 4x Output Resolution: 64 × 1024
Bilinear2.08920.10630.6000
SRNO [51]0.83500.20350.4417
HAT [9]0.68560.20350.2516
SWIN-IR [29]1.29720.27740.7347
LIIF [10]0.61430.32260.1916
LIDAR-SR [43]0.56740.10050.2165
ILN [26]1.05280.33420.2787
TULIP (Ours)0.41850.41740.1207
TULIP-L (Ours)0.37080.43290.0992
" + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.444, + 0.469, + 0.61 + ], + "angle": 0, + "content": "
DurLAR 4x Output Resolution: 128 × 2048
Bilinear2.43840.12660.6346
SRNO [51]1.53960.15070.5108
HAT [9]1.78200.23530.1973
SWIN-IR [29]1.94160.21570.2279
LIIF [10]1.56720.24690.1548
LIDAR-SR [43]1.53120.13700.1128
ILN [26]1.57200.34300.0893
TULIP (Ours)1.54320.35620.06484
TULIP-L (Ours)1.55920.36540.06346
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.612, + 0.47, + 0.655 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison against state-of-the-art LiDAR and image super-resolution methods on different datasets. All methods are trained and evaluated on the same splits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.47, + 0.714 + ], + "angle": 0, + "content": "resolution feature embeddings. Similar to our work, they are based on Swin Transformer [33]. We train all methods on the specific datasets using their default parameters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.739, + 0.264, + 0.753 + ], + "angle": 0, + "content": "4.3.1 Simulation Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Similar to prior works, we first evaluate the performance on noise-free, simulated data. We train and test all approaches on the dataset presented in [26] that was recorded using the CARLA simulator and select the ground-truth range images of size \\(128 \\times 2048\\) that capture a vertical FoV of \\(30^{\\circ}\\). We use a (20699/2618) train/test split. We observe that the performance between image super-resolution approaches varies greatly, and LIIF performs the best among them. While LIIF [10] achieves a comparable MAE to the LiDAR up-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.244 + ], + "angle": 0, + "content": "sampling methods, the 3D metrics are significantly worse, indicating that it is unsuitable for the underlying geometric input data. Nevertheless, LIIF [10] still outperforms the Li-DAR upsampling method in LIDAR-SR. The interpolation-based approach in ILN shows decent results for the noise-free data, as indicated by their high IoU. Nevertheless, our approaches achieve the best performance in all evaluated metrics, with the strongest improvement in Chamfer Distance, which validates that our design is well suited for the geometric reconstruction in range images." + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.269, + 0.906, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.71, + 0.895, + 0.919 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on KITTI: We provide visualizations for the top-4 results from Tab. 2 and Swin-IR [29]. Our approach outperforms other state-of-the-art methods in upsampling realistic point clouds. A: TULIP is the only approach that does not generate invalid ghost points in between the car (right bottom) and the wall behind. It can also be observed that all approaches except for ILN match the characteristic line pattern of the LiDAR on the ground (middle). B: Swin-IR and LIIF generate a blurry reconstruction. TULIP achieves the clearest point cloud and even reconstructs small details such as the car's side mirrors. C: Swin-IR, LIIF, and LiDAR-SR do not clearly separate the car from the wall behind. ILN generates a clear point cloud but does not properly reconstruct the shape of the cars. TULIP achieves the clearest point cloud, with distinct discontinuities between objects, and closely resembles the shape of the three cars." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15360" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.092, + 0.27, + 0.106 + ], + "angle": 0, + "content": "4.3.2 Real-World Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.118, + 0.47, + 0.209 + ], + "angle": 0, + "content": "To evaluate our approach on real-world data, we downsample the DurLAR and KITTI datasets to a comparable number of frames to the CARLA dataset by temporally skipping frames in the dataset sequences. In particular, we have the following (train/test) split: DurLAR (24372/2508), and KITTI (20000/2500)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.212, + 0.471, + 0.515 + ], + "angle": 0, + "content": "KITTI: The KITTI dataset [21] was collected using a Velodyne HDL-64E LiDAR with a vertical FoV of \\(26.8^{\\circ}\\) and a resolution of \\(64 \\times 1024\\). For the image super-resolution approaches, we observe comparable results and trends on KITTI as on the simulated data in Sec.4.3.1. Even though LIDAR-SR [43] is designed for LIDAR upsampling and outperforms image super-resolution approaches in terms of MAE, it achieves the lowest IoU among all approaches, which can be explained by a large amount of invalid floating points between real objects. LIIF [10] achieves better CD and comparable IoU compared to ILN. We observe that the IoU of ILN [26] is drastically lower than on CARLA. The interpolation scheme in ILN is thus not suitable to properly handle the lower resolution data with noise from the real-world sensor. Our approaches significantly outperform all competing methods, especially in terms of IoU as it can better handle discontinuities between objects. As indicated in the visualizations in Fig.5, our approach predicts fewer invalid float points between objects, as well as fewer blurry edges, which results in much clearer point clouds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.47, + 0.716 + ], + "angle": 0, + "content": "DurLAR: The DurLAR dataset [28] was recorded using an Ouster OS1-128 LiDAR, that captures point clouds with a resolution of \\(128 \\times 2048\\) with a \\(45^{\\circ}\\) vertical FoV. Compared to CARLA and KITTI, we see a large performance drop for all approaches in MAE and IoU, which results from the larger maximum range and vertical FoV of the sensor as well as stronger sensor noise. We observe that all approaches designed for LiDAR upsampling clearly improve in Chamfer Distance compared to image super-resolution approaches. ILN and TULIP outperform all methods by a large margin in IoU. TULIP-L achieves the best 3D metrics, which validates that it can also handle the challenging DurLAR data better than prior works." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Range Analysis: As the density of LiDAR point clouds decreases with increasing distance from the sensor, we additionally compare the performance at different ranges for a more fine-grained evaluation. We observe that our approach achieves better or similar performance compared to all approaches in all metrics at all ranges. Up to \\(30\\mathrm{m}\\), TULIP clearly outperforms the baseline approaches. TULIP handles discontinuities between objects better than other approaches, which can be verified visually through sharper edges and fewer \"floating\" points between objects. This is especially reflected in the high IoU. At ranges above \\(30\\mathrm{m}\\), our approach performs comparable to other methods." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.696, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.184, + 0.625, + 0.196 + ], + "angle": 0, + "content": "(a) IoU↑" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.089, + 0.891, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.184, + 0.855, + 0.196 + ], + "angle": 0, + "content": "(b) Chamfer distance \\(\\downarrow\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.205, + 0.892, + 0.22 + ], + "angle": 0, + "content": "Figure 6. 3D error metrics visualized at different ranges for KITTI." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.225, + 0.892, + 0.318 + ], + "angle": 0, + "content": "Above this range, LiDAR data is extremely sparse, meaning that points neighboring in the low-resolution range image are actually distant in the Euclidean space. It is, therefore, not feasible to infer meaningful information about the scene from such neighboring points at a high range, which is also represented in the poor performance of all approaches." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.327, + 0.64, + 0.342 + ], + "angle": 0, + "content": "4.4. Failure Cases" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Although TULIP outperforms other state-of-the-art methods both qualitatively and quantitatively, the upsampling quality can still be limited in certain cases. For instance, as shown in Fig 7, our approach shows inferior upsampling results for the specific scene. The irregularity of the scene leads to high uncertainty, which makes the reconstruction noisy. Observing the details in Fig. 7a, the network fails to reconstruct the car at the top of the scene." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.49, + 0.712, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.591, + 0.658, + 0.603 + ], + "angle": 0, + "content": "(a) Full scene" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.504, + 0.864, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.591, + 0.83, + 0.602 + ], + "angle": 0, + "content": "(b) Details" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.614, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Figure 7. Failure Case: noisy point cloud generated by TULIP and missing reconstruction of the object. (Ours and GT)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.664, + 0.619, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work presents TULIP, a novel method for LiDAR Upsampling that achieves incredible performance in upsampling the range image. Our approach transforms 3D point clouds into 2D range images and performs the upsampling in 2D space. We build upon a Swin-Transformer-based network and specifically modify the patch partition and attention windows to better accommodate the characteristics of range images. Throughout various experiments, testing on three different benchmarks, it shows that TULIP outperforms state-of-the-art methods quantitatively in all evaluation metrics, and quantitatively, TULIP demonstrates more promising results in upsampling the realistic LiDAR data. Acknowledgement: This work was supported by Swiss National Science Foundation's NCCR DFab P3." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15361" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Anique Akhtar, Zhu Li, Geert Van der Auwera, Li Li, and Jianle Chen. Pu-dense: Sparse tensor-based point cloud geometry upsampling. IEEE Transactions on Image Processing, 31:4133-4148, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.471, + 0.214 + ], + "angle": 0, + "content": "[2] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. ECCVW What is Motion For?, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.471, + 0.271 + ], + "angle": 0, + "content": "[3] Borna Bešić, Nikhil Gosala, Daniele Cattaneo, and Abhinav Valada. Unsupervised domain adaptation for lidar panoptic segmentation. IEEE Robotics and Automation Letters, 7(2): 3404-3411, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.471, + 0.413 + ], + "angle": 0, + "content": "[5] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.471, + 0.471 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.471, + 0.599 + ], + "angle": 0, + "content": "[8] Tso Yuan Chen, Ching Chun Hsiao, and Ching-Chun Huang. Density-imbalance-eased lidar point cloud upsampling via feature consistency learning. IEEE Transactions on Intelligent Vehicles, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.602, + 0.471, + 0.671 + ], + "angle": 0, + "content": "[9] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367-22377, 2023. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.673, + 0.471, + 0.741 + ], + "angle": 0, + "content": "[10] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.744, + 0.471, + 0.813 + ], + "angle": 0, + "content": "[11] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12312-12321, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.816, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[12] Xinjing Cheng, Peng Wang, and Ruigang Yang. Depth estimation via affinity learned with convolutional spatial propagation network. In Proceedings of the European conference on computer vision (ECCV), pages 103-119, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[13] Yin Dai, Fayu Liu, Weibing Chen, Yue Liu, Lifu Shi, Sheng Liu, Yuhang Zhou, et al. Swin mae: Masked autoencoders" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "for small datasets. Computers in Biology and Medicine, 161: 107037, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[14] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[15] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 391-407. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[18] George Eskandar, Sanjeev Sudarsan, Karim Guirguis, Janaranjani Palaniswamy, Bharath Somashekar, and Bin Yang. Hals: A height-aware lidar super-resolution framework for autonomous driving. arXiv preprint arXiv:2202.03901, 2022. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[19] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pages 1050-1059. PMLR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[20] Hanan Gani, Muzammal Naseer, and Mohammad Yaqub. How to train vision transformer on small-scale datasets? In 33rd British Machine Vision Conference, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[21] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[22] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Penet: Towards precise and efficient image guided depth completion. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13656-13662. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[23] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[24] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[25] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1637-1645, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Youngsun Kwon, Minhyuk Sung, and Sung-Eui Yoon. Implicit lidar network: Lidar super-resolution via interpolation weight prediction. In 2022 International Conference on" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15362" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Robotics and Automation (ICRA), pages 8424-8430. IEEE, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[27] Seung Hoon Lee, Seunghyun Lee, and Byung Cheol Song. Vision transformer for small-size datasets. arXiv preprint arXiv:2112.13492, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[28] L. Li, K.N. Ismail, H.P.H. Shum, and T.P. Breckon. Durlar: A high-fidelity 128-channel lidar dataset with panoramic ambient and reflectivity imagery for multi-modal autonomous driving applications. In Proc. Int. Conf. on 3D Vision. IEEE, 2021. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[29] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[30] Yuankai Lin, Tao Cheng, Qi Zhong, Wending Zhou, and Hua Yang. Dynamic spatial propagation network for depth completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1638-1646, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.47, + 0.427 + ], + "angle": 0, + "content": "[31] Zhixin Ling, Zhen Xing, Xiangdong Zhou, Manliang Cao, and Guichun Zhou. Panoswin: A pano-style swin transformer for panorama understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[32] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco De Nadai. Efficient training of visual transformers with small datasets. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[35] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457-466, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[36] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 120-136. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.817 + ], + "angle": 0, + "content": "[37] Jiaxiong Qiu, Zhaopeng Cui, Yinda Zhang, Xingdi Zhang, Shuaicheng Liu, Bing Zeng, and Marc Pollefeys. Deepidar: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3313-3322, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[38] Shi Qiu, Saeed Anwar, and Nick Barnes. Pu-transformer: Point cloud upsampling transformer. In Proceedings of the Asian Conference on Computer Vision, pages 2475-2493, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[39] Kyeongha Rho, Jinsung Ha, and Youngjung Kim. Guideformer: Transformers for image guided depth completion." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6250-6259, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.206 + ], + "angle": 0, + "content": "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.894, + 0.276 + ], + "angle": 0, + "content": "[41] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[42] Artem Savkin, Yida Wang, Sebastian Wirkert, Nassir Navab, and Federico Tombari. Lidar upsampling with sliced wasserstein distance. IEEE Robotics and Automation Letters, 8(1): 392-399, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[43] Tixiao Shan, Jinkun Wang, Fanfei Chen, Paul Szenher, and Brendan Englot. Simulation-based lidar super-resolution for ground vehicles. Robotics and Autonomous Systems, 134: 103647, 2020. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[44] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[45] Xiaopeng Sun, Weiqi Li, Zhenyu Zhang, Qiufang Ma, Xuhan Sheng, Ming Cheng, Haoyu Ma, Shijie Zhao, Jian Zhang, Junlin Li, et al. Opdn: Omnidirectional position-aware deformable network for omnidirectional image superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1293-1301, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[46] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing, 30:1116-1129, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[47] Larissa T Triess, David Peter, Christoph B Rist, Markus Enzweiler, and J Marius Zöllner. Cnn-based synthesis of realistic high-resolution lidar data. In 2019 IEEE Intelligent Vehicles Symposium (IV), pages 1512-1519. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.747, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[49] Suhong Wang, Hongqing Wang, Shufeng She, Yanping Zhang, Qingju Qiu, and Zhifeng Xiao. Swin-t-nfc crfs: An encoder-decoder neural model for high-precision UAV positioning via point cloud super resolution and image semantic segmentation. Computer Communications, 197:52–60, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[50] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15363" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[51] Min Wei and Xuesong Zhang. Super-resolution neural operator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18247-18256, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[52] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Baobei Xu, Jun Li, and Jian Yang. Rignet: Repetitive image guided network for depth completion. In European Conference on Computer Vision, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[53] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1951-1960, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.263, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[54] Shuquan Ye, Dongdong Chen, Songfang Han, Ziyu Wan, and Jing Liao. Meta-pu: An arbitrary-scale upsampling network for point cloud. IEEE transactions on visualization and computer graphics, 28(9):3206-3218, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.32, + 0.469, + 0.388 + ], + "angle": 0, + "content": "[55] Fanghua Yu, Xintao Wang, Mingdeng Cao, Gen Li, Ying Shan, and Chao Dong. Osrt: Omnidirectional image superresolution with distortion-aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13283-13292, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.39, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[56] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-net: Point cloud upsampling network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2790-2799, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.447, + 0.468, + 0.487 + ], + "angle": 0, + "content": "[57] Heeseung Yun, Sehun Lee, and Gunhee Kim. Panoramic vision transformer for saliency detection in 360 videos. In ECCV, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.489, + 0.468, + 0.543 + ], + "angle": 0, + "content": "[58] Wentai Zhang, Haoliang Jiang, Zhangsihao Yang, Soji Yamakawa, Kenji Shimada, and Levent Burak Kara. Data-driven upsampling of point clouds. Computer-Aided Design, 112:1-13, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.546, + 0.468, + 0.586 + ], + "angle": 0, + "content": "[59] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image restoration. TPAMI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.589, + 0.469, + 0.67 + ], + "angle": 0, + "content": "[60] Wenbo Zhao, Xianming Liu, Zhiwei Zhong, Junjun Jiang, Wei Gao, Ge Li, and Xiangyang Ji. Self-supervised arbitrary-scale point clouds upsampling via implicit neural representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1999-2007, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.673, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[61] Yifan Zhao, Le Hui, and Jin Xie. Sspu-net: Self-supervised point cloud upsampling via differentiable rendering. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2214-2223, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.73, + 0.469, + 0.784 + ], + "angle": 0, + "content": "[62] Örkény Zováthi, Balázs Pálffy, Zsolt Jankó, and Csaba Benedek. St-depthnet: A spatio-temporal deep network for depth completion using a single non-repetitive circular scanning lidar. IEEE Robotics and Automation Letters, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15364" + } + ] +] \ No newline at end of file diff --git a/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/52ac8b22-207c-41ff-adcb-c9f4c82bca08_origin.pdf b/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/52ac8b22-207c-41ff-adcb-c9f4c82bca08_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ef1474d3b2d36e416bdbc690033e7e8f398332bc --- /dev/null +++ b/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/52ac8b22-207c-41ff-adcb-c9f4c82bca08_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98131b02ee846e6bd8f3d8a3f3ee7d0b6aaffde4a0a94994aa0b08b619db83ec +size 3064801 diff --git a/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/full.md b/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c953aeeee7a0c4a086214d9c2cb0b8e39d0a7b36 --- /dev/null +++ b/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/full.md @@ -0,0 +1,280 @@ +# TULIP: Transformer for Upsampling of LiDAR Point Clouds + +Bin Yang $^{1}$ , Patrick Pfreundschuh $^{1}$ , Roland Siegwart $^{1}$ , Marco Hutter $^{2}$ , Peyman Moghadam $^{3,4\dagger}$ , Vaishakh Patil $^{2\dagger}$ $^{1}$ Autonomous Systems Lab, ETH Zurich, Switzerland, $^{2}$ Robotic Systems Lab, ETH Zurich, Switzerland + $^{3}$ School of EER, Queensland University of Technology (QUT), Australia, $^{4}$ Data61, CSIRO, Brisbane, Australia + +{biyang, patripfr, rolandsi, mahutter, patilv}@ethz.ch, peyman.moghadam@csiro.au + +# Abstract + +LiDAR Upsampling is a challenging task for the perception systems of robots and autonomous vehicles, due to the sparse and irregular structure of large-scale scene contexts. Recent works propose to solve this problem by converting LiDAR data from 3D Euclidean space into an image super-resolution problem in 2D image space. Although their methods can generate high-resolution range images with fine-grained details, the resulting 3D point clouds often blur out details and predict invalid points. In this paper, we propose TULIP, a new method to reconstruct high-resolution LiDAR point clouds from low-resolution LiDAR input. We also follow a range image-based approach but specifically modify the patch and window geometries of a Swin-Transformer-based network to better fit the characteristics of range images. We conducted several experiments on three public real-world and simulated datasets. TULIP outperforms state-of-the-art methods in all relevant metrics and generates robust and more realistic point clouds than prior works. The code is available at https://github.com/ethz-as1/TULIP.git. + +# 1. Introduction + +Light Detection And Ranging (LiDAR) is one of the most common sensors for perception in various fields of autonomy, such as autonomous driving, and unmanned aerial vehicles (UAV). LiDARs are used to generate 3D point clouds of the scene. These point clouds are essential for mapping, localization, and object detection tasks. However, the accuracy of these tasks often depends on the resolution of the point cloud [53]. Furthermore, the resolution of a LiDAR is inherently associated with increased energy consumption and cost, making its use impractical for various applications. LiDARs also have different vertical and horizontal resolutions. The vertical resolution of rotating 3D LiDARs is typically much lower than the horizontal res + +![](images/e09006fdab21513ee203b9986e2f4dc6a5bb4922c1b25eafbe50755107db8b76.jpg) +(a) Input + +![](images/87a798a93a376080981a05716c735f8b529f30f9953116b01f80034c8ffc0275.jpg) +(b) Ground Truth + +![](images/0761cadd0d2db11c46775303ea0145c26832e235fcba08657876c5ea151c4623.jpg) +(c) SwinIR [29] + +![](images/049793755ba3abf39502d794c1ddf168595b53ca9f34e048d83b236367bbaf22.jpg) +(d) TULIP (Ours) +Figure 1. TULIP generates more realistic LiDAR point clouds from low-resolution LiDAR input (a) and outperforms the state-of-the-art image super-resolution approach [29] in upsampling with fine-grained details of different objects in the scene. + +olution. These limitations make an upsampling technique necessary to increase the resolution of the LiDAR data, especially in the vertical direction. Moreover, upsampling LiDAR data has the potential to counter domain shift problems in LiDAR-based learning methods. Notably, the use of lower resolution LiDAR data in a system that has been trained on higher resolution data leads to a significant drop in performance [3]. Therefore, such upsampling techniques can not only help to mitigate the domain gap by creating a virtual high-resolution sensor that matches the target domain [42, 61] but also reduce the high cost of collecting new LiDAR data, annotating and retraining the methods. + +Deep learning has led to remarkable advancements in LiDAR upsampling techniques in recent years. Several recent methods [42, 56, 58] have focused on learning the upsampling process within the 3D Euclidean space. However, the processing of 3D data in deep learning can be quite resource-intensive. One way to address this challenge is the representation of the point cloud as a range image [18, 26, 43, 49]. This approach facilitates the adoption of well-established image-based super-resolution methods. + +Such methods mostly employ autoencoder-style networks with convolutional and deconvolutional layers [14, 43, 62]. However, these methods are not directly transferable to range images. Range images generated by the projection of sparse 3D point clouds have distinct sharp edges at the object boundaries. Convolutional neural network-based methods often induce edge smoothing due to the regularization effect, which limits their usability for LiDAR range image processing. + +As an alternative, Transformer [48] has shown great success in RGB image super-resolution and produces less blurry outputs [29]. Vision transformers, however, face challenges such as the need for large amounts of training data [9, 29, 35] and the global computation of self-attention, limiting the capture of local information in the range images. To address this, Swin Transformer [33] introduced tokenizing the input image and applying the self-attention mechanism locally in independent windows. While the modifications to Swin Transformer have led to robust training and promising results compared to Vision Transformer, its default structure and settings are not ideally suited for processing LiDAR range images. These range images consist of single-channel data, representing 3D spatial information, as opposed to three-channel color images that depict visual appearance. Furthermore, range images feature large, smooth areas within 3D objects, separated by distinct sharp edges between objects. As a result, minor inaccuracies in the 2D prediction (e.g. blurry output) can lead to significant differences in the projected 3D occupancy. This non-uniform distribution of image details further complicates network training. The challenging details in the range image are concentrated in a few specific pixels, contrasting with RGB images where relevant details are evenly distributed. This makes range images fundamentally different from RGB images. Inspired by the limitations of state-of-the-art image upsampling methods and the potential of Swin-Transformer [33], we propose a novel network named TULIP. Our proposed geometry-aware architecture is tailored to accommodate LiDAR range image data. The proposed method takes a low-resolution range image as input and produces a high-resolution range image, which can then be projected into a 3D point cloud. To accommodate LiDAR range image geometry, we utilize one-line row patches to tokenize the input range image, as opposed to the square patches used for RGB images. This approach helps to preserve vertical information for the upsampling process while capturing boundary discontinuities effectively. This also creates a trade-off between local detail capture and model complexity in the training process. Additionally, we incorporate a non-square window for local self-attention computation. This enhances learning of spatial contexts at different scales in the range image. Further, we extensively train and evaluate TULIP on two real-world and + +one simulated autonomous driving datasets: KITTI [21], DurLAR [28] and CARLA [26], respectively. Comparisons to existing approaches demonstrate that TULIP outperforms the state-of-the-art methods on all three datasets. + +# 2. Related Works + +Image super-resolution has seen significant progress in recent years. The technique aims to construct high-resolution (HR) images from low-resolution (LR) observations, often leveraging advances in Convolutional Neural Networks (CNNs) to enhance the fidelity and detail of images for better visualization and information extraction. Subsequently, several enhanced frameworks have been developed [14, 15, 24, 25, 59]. + +Driven by their success in the field of natural language processing (NLP) [4, 23], Transformers [48] have been extended to solve a variety of vision-related tasks such as object detection [6] and semantic segmentation [2, 7]. Such Vision Transformers (ViT) [17] excel at learning to focus on relevant image regions by exploring global interactions among different regions. Their impressive performance has led to their adoption in image restoration tasks [9, 11, 29] as well. While they brought significant improvements in RGB image super-resolution, Vision Transformer [17] comes with the drawback of the quadratic computational complexity of self-attention. Additionally, they capture mostly global dependencies within the data and require large amounts of data for training. + +To address these challenges, Swin Transformer [33] has been proposed. Unlike the Vision Transformer, which relies on global self-attention across the entire image, the Swin Transformer employs a local window-based attention mechanism and establishes a pyramid-like architecture. This approach processes images hierarchically by gradually merging smaller image patches into larger ones. This strategy enables more effective handling of different scales and facilitates the processing of multi-scale features, making it particularly suitable for range image processing. Furthermore, due to Swin Transformer's hierarchical structure, the increase in the number of model parameters scales linearly with the input image size, which mitigates the challenges associated with handling large range images. Additionally, they require less training data than the classical Vision Transformer, which is beneficial for LiDAR upsampling as LiDAR datasets are typically magnitudes smaller than RGB datasets used for Vision Transformer. The Swin and Vision Transformers have also been used for image super-resolution on omnidirectional panoramic camera images [45, 55]. However, these approaches are specifically designed for the unique characteristics of these cameras (e.g. strong distortions, spherical continuity) that are vastly different from the LiDAR sensing model, which renders them unsuitable for application to range images. + +Other related works have focussed on upsampling high-density point clouds from low-density point clouds [38, 54, 56, 60, 61]. While those approaches also provide qualitative results for LiDAR point clouds, they focus on increasing the general 3D point density of point clouds of single objects instead of a large scene. Another area closely related to LiDAR upsampling is depth completion [12, 22, 30, 36, 37, 39, 46, 52]. The primary objective of depth completion is to improve sparse depth estimates acquired from LiDAR by integrating data from multiple sensors, predominantly RGB cameras. The result of depth completion methods is a dense depth map that provides depth values to every pixel in the input depth map. Nonetheless, this field slightly diverges from LiDAR upsampling due to its reliance on multimodal sensors and its common deployment within a restricted field of view, making it not directly comparable within the scope of this paper. LiDAR point clouds exhibit specific characteristics, such as a distinct point pattern (due to the stacked lasers in rotating 3D LiDARs) and a decrease in point density with increasing distance from the sensor. Differently from arbitrary point cloud upsampling, LiDAR upsampling tries to mimic a realistic point cloud of a high-resolution LiDAR given the point cloud of low-resolution LiDAR and thus targets a different result than the approaches above. Due to the difficulty and high computational cost of upsampling LiDAR point clouds in 3D space [1, 8, 42], most works [26, 43, 47] represent the point cloud as a range image and perform the LiDAR upsampling task in 2D image space. LiDAR-CNN [47] introduces a CNN-based architecture with semantic and perceptual guidance. The method specifies two further loss functions besides a per-point reconstruction loss to acquire a better synthesis of the high-resolution range image. However, this approach is limited to data with semantic annotations, which drastically reduces its applicability. LiDAR-SR [43] deploys a CNN-based network with U-Net [40] architecture. It additionally uses Monte-Carlo dropout post-processing to reduce the amount of noisy points in the prediction. Approaches based on convolutional operations tend to fail to reconstruct the sharpness in the range image effectively. While the previous approaches directly predict the range image, Implicit LiDAR Network (ILN) [26] uses an implicit neural architecture that learns interpolation weights to fill in new pixels instead of their values directly. Although the method outperforms CNN-based approaches in training speed and preserving the geometrical details in the input, it can still suffer from limited neighboring information in the LiDAR data, especially in areas distant from the sensor due to the extremely sparse input data. Differently from the previous works, our approach builds on Swin Transformer as a backbone and is specifically customized to effectively process range images. + +# 3. Methodology + +LiDAR upsampling is achieved by upsampling a range image, effectively transforming it from a 3D upsampling problem into a 2D image super-resolution problem [18, 26, 43, 47]. It is, therefore, an evident approach to build on RGB image super-resolution. However, image super-resolution aims to solve a different problem than LiDAR upsampling. Image super-resolution enhances the visual appearance of a low-resolution RGB image by recreating a high-resolution image. While quantitative metrics for evaluating these works exist, there is no one specific correct solution. Differently, LiDAR upsampling tries to recreate a high-resolution, 3D LiDAR sensor output from a low-resolution LiDAR input point cloud. Range images contain more geometrical and spatial contexts rather than visual information, and the scene's geometry strictly dictates the correct solution. The underlying sensing model is also drastically different, as LiDARs are active sensors. The respective range pixel values purely depend on distance, while pixels in an RGB camera depend on many factors, such as scene appearance, lighting conditions, exposure time, and white balance. Furthermore, range images have a highly asymmetric aspect ratio, typically between 1:8 to 1:64, which differs drastically from regular camera images. Accordingly, most works [18, 43, 47] only perform vertical upsampling, in contrast to image super-resolution, which operates in both dimensions. A naive application of state-of-the-art image super-resolution networks [29] to range images does not result in adequate performance (Fig. 1). To this end, we developed TULIP, a novel range image-based LiDAR Upsampling method. Based on the observations above, we specifically adapt a network based on SwinUnet [5] to incorporate range images better. In the following sections, we will describe the technical details of our method, with a special focus on how our approach accommodates range images. + +# 3.1. Problem Definition + +A LiDAR point cloud consists of points captured during one revolution $\mathcal{P} = \{\mathbf{p}_1,\dots ,\mathbf{p}_n\}$ . Each measurement represents a 3D point $\mathbf{p}_i = \{x_i,y_i,z_i\}$ . As input, we assume a low-resolution point cloud $\mathcal{P}_l$ with $n_l = H_l\times W_l$ points, where $H_{l}$ and $W_{l}$ correspond to vertical and horizontal resolution. We aim to predict a high-resolution point cloud $\mathcal{P}_h$ , that has the same field-of-view(FoV) as $\mathcal{P}_l$ but contains $n_h = \beta *n_l$ points. In our experiments, we set $\beta = 4$ . We only increase vertical resolution (amount of LiDAR beams), i.e., $n_h = H_h\times W_h$ , where $H_{h} = \beta *H_{l}$ and $W_{h} = W_{l}$ . We project the point cloud into a 2D range image. In this image, each row and column coordinate $v,u$ correspond to the respective elevation and azimuth angles of the LiDAR points, while the pixel value contains the range of the point $r = \sqrt{x^2 + y^2 + z^2}$ . The image coordinate of a 3D point + +can be calculated using a spherical projection model: + +$$ +\left[ \begin{array}{l} u \\ v \end{array} \right] = \left[ \frac {\frac {W}{2} - \frac {W}{2 \pi} a r c t a n (\frac {y}{x})}{\Theta_ {m a x} - \Theta_ {m i n}} * (\Theta_ {m a x} - a r c t a n (\frac {z}{\sqrt {x ^ {2} + y ^ {2}}}) \right] (1) +$$ + +$\Theta_{max}$ and $\Theta_{min}$ correspond to the limits of the vertical field of view of the LiDAR. We formulate the LiDAR upsampling problem as an image upsampling problem, i.e. we aim to predict a high-resolution range image $I_h \in \mathbf{R}^{1 \times H_h \times W}$ given a low-resolution range image $I_l \in \mathbf{R}^{1 \times H_l \times W}$ . We can then calculate the high-resolution point cloud $\mathcal{P}_h$ by inverting the projection in Eq. 1. + +# 3.2. Architecture Design and Overview + +Our network builds upon Swin-Unet [5] which was originally designed for image segmentation. We deploy a U-shaped network structure featuring skip connections that link the encoder and decoder modules. The fundamental building block of TULIP is the Swin Transformer [33]. TULIP first patchifies and maps the input to a high-dimensional feature space. The tokenized low-level feature maps are passed through a multi-stage encoder. Each stage contains a pair of Swin Transformer [33] blocks and a patch merging layer to downsample the resolution by a factor of two and increase the dimension by a factor of four. This step realizes the hierarchic computation and extraction of multi-scale features through encoding. Within each stage, multiple instances of the self-attention mechanism are performed locally in parallel, which is the so-called Window Multi-Head Self Attention (W-MSA). After passing through a two-layer MLP and residual connection with the input, the feature vector is further passed to the second part of the Swin Transformer block, utilizing a Shifted Window Multi-Head Self Attention (SW-MSA) which extends W-MSA by employing a shifted window operation. This enhances the model capacity as it compensates for the lack of interaction between the local windows in W-MSA. The decoder, which realizes the upsampling has a symmetric design to the encoder and operates in a reverse way of the encoder. The resolution of the feature maps is first expanded, and the dimension is reduced by a factor of two accordingly. We feed existing geometrical information via skip connections. The feature maps are transformed into a single-channel, high-resolution range image in the last layer. The reconstruction head comprises a $1 \times 1$ convolutional layer for feature expansion, followed by a Leaky ReLU activation, a pixel shuffle layer [44], and another $1 \times 1$ convolutional layer for final projection. As the loss function, we select the pixel-wise L1 loss. Details of the network can be found in the Appendix. + +# 3.2.1 Tokenization + +Tokenization is done by a patch partition layer that creates an initial feature embedding from the input range image. + +This initial feature representation significantly influences the network's performance. Specifically, the selection of this layer determines how the inherent characteristics, patterns, and relevant information of the LiDAR range image data are encoded into a format that the network can effectively learn from. Most transformer-based RGB image super-resolution approaches utilize a relatively large patch size to tokenize the input [9, 29, 35]. These larger patches enable the construction of a global spatial context, facilitating an understanding of the overall structure and high-level features of the RGB image. These networks employ square-shaped patches and implement upsampling in both vertical and horizontal directions. In contrast, LiDAR upsampling primarily aims to enhance the vertical resolution of the input, given that range images possess properties that are vastly different from RGB images. Furthermore, attention across more pixels in range images is less useful than in RGB images, since there is almost no geometrical relation for the spatial contexts that are far from each other in the 2D range image space. Motivated by this, we propose two adjustments to effectively process range images within a Swin-Unet: + +Row-Based Patch Partition: Our model builds row patches with a dimension of $1 \times 4$ for range image tokenization. The new patch geometry is designed to retain full vertical information while compressing horizontally. This aligns with our objective of conserving and extending vertical details in the range image. Besides that, row patches excel at capturing boundary discontinuities between distinct objects and the background within the scene. Moreover, the selection of the patch size strikes a balance between enhancing model performance and limiting model capacity. Smaller patches can capture more detailed local information, but they also increase the number of model parameters, which can slow down the model's training process. + +Circular Padding (CP) in Horizontal Dimension: Instead of adding zero pixels padding, our model uses circular padding in the horizontal dimension. This avoids introducing artificial features and naturally matches the sensor model of a rotating 3D LiDAR. This preserves accurate neighbor relations along the edges, which would otherwise be compromised during the projection to the 2D range image. The Circular Padding (CP) also allows us to fit flexible input sizes without changing the patch geometry. + +# 3.2.2 Non-Square window for local self-attention + +Inside Swin Transformer, input feature maps are window-partitioned along width and height to compute the self-attention locally in each window. In addition to square-shaped RGB images, some recent work applies a square window for panoramic data as well [31, 57]. However, using square windows on range images is unfavorable. On one + +![](images/058fa761a4909494e7b5e950bae2e62b7b95ebedb64e4a42db47afd075e1f759.jpg) +Figure 2. Overview of TULIP: The low-resolution point cloud is spherically projected into the range image. In the training and inference, the range image is tokenized into feature embeddings without compressing the spatial resolution along the vertical axis. We adopt the U-Net-based network presented in [5] for range image upsampling. We follow the hierarchical structure of Swin Transformer [33] for feature extraction and use the non-square window geometry for computing the local self-attention. The reconstruction head generates the high-resolution range image from the latent features and then, the point cloud can be obtained by projecting pixels back into the 3D space. + +hand, small windows can focus more on attention within near areas but lead to increasing computational complexity and can struggle with the reconstruction of objects at a further distance in the scene. On the other hand, a larger square window can blend scene-related contexts at different scales, which can potentially degrade the network performance. In a prior study done by Eskandar et al. [18], they demonstrated a significant enhancement in range image superresolution by separately upsampling the upper and lower portions of the data. In contrast to their approach of utilizing CNN-based shallower and deeper network branches for distinct feature extraction, our proposal involves the direct partitioning of the input using a rectangular window. This window geometry facilitates robust attention interaction among distant points through W-MSA and subtle crossattention between objects at multiple scales via SW-MSA and hierarchical processing. In addition, range images, as typical panorama data, inherently contain abundant information along the horizontal direction. Therefore, directing attention more toward image width rather than height can aid the network in capturing a greater amount of geometrical information. At the same time, the computational complexity remains the same as using a large square window. + +# 3.2.3 Further Adaptation and Refinement + +Patch Unmerging (PU) and Pixel Shuffle (PS): Similar to the patch merging layer that is responsible for down-sampling at the encoder stage, a mechanism is necessary within the decoder to upsample the feature maps. Conventional CNN-based approaches [43, 50] achieve upsampling through the transposed convolutional operation. As they tend to smooth out image sharpness [29], we reversed the operation of patch merging, re-arranging the channels of patches into their respective positions in a grid to up-sample the spatial resolution of input feature maps, which we + +denote as patch unmerging. Furthermore, to reconstruct the final range image, we built the reconstruction head upon the upsampled feature maps. The implementation of the reconstruction module is based on the pixel shuffle layer [44]. + +Monte Carlo Dropout: We additionally use Monte Carlo Dropout [43] to filter unreliable points by thresholding the uncertainty. MC-Dropout executes several feed-forward passes during inference with different active dropouts, which yields a distribution of outputs. We refine the results by removing points with high output variance, as they often come down to noisy, invalid points. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets: We conduct experiments on three different datasets, that include the two large-scale real-world datasets DurLAR [28] and KITTI [21] (Sec. 4.3.2), as well as on a dataset [26] that was generated using the CARLA simulator [16] (Sec. 4.3.1). We select test sequences that are recorded in different locations than the train set to avoid spatial overlap. We vertically downsample the high-resolution range images with a factor of four by skipping the respective lines to generate low-resolution input images. For ablation studies in Sec. 4.2, we select the KITTI dataset. + +Implementation Details: In all experiments, we use distributed processing with $4 \times$ GeForce RTX 2080 Ti. For training, we use a fixed batch size of eight per GPU for all datasets. We use the AdamW [34] optimizer with a weight decay factor of 0.01 and a base learning rate of 5e-4. + +Evaluation metrics: We evaluate the Mean Absolute Error (MAE) for all the pixels in the generated 2-D range images. Additionally, we assess the performance by considering the 3D points reconstructed by our neural networks. The Chamfer Distance (CD) measures the Euclidean distance between + +![](images/5d4a9e49b8b088b9500b3b231376ef94ea57b36bf3533574cf0ffbdb29f35ea7.jpg) +(a) Scene + +![](images/cc353321f5773f05891d97252c6797477d1397fbe84ae08db5338efefc3e2e2d.jpg) +(b) GT + +![](images/1ea230d4c9029e503049eacac46d1cb3a88121c05998bf8245211e1e3f1ded68.jpg) +(c) $1 \times 4$ + +![](images/2b7fb6383d0634125ee16e7befb8e254fbef6b9c57b6d649deb283c5c0736bb6.jpg) +(d) $2 \times 2$ + +![](images/0fe2aab2c5160df7eff0382e24a75b46a63e6c6d1e6cc172fd70aef00196d387.jpg) +(e) $4 \times 4$ + +![](images/988a09b5144fc9df356a255f8d2e51f8348f0fb0fa6b554d4113efefb82ea6cd.jpg) +Figure 3. Upsampling results with different patch sizes: We observe that square patches $(4 \times 4, 2 \times 2)$ blur out edges and create invalid connections between separate walls, while the proposed $1 \times 4$ patch generates sharp edges and correct discontinuities. +(a) Scene +Figure 4. Upsampling results with different window geometries: Rectangular windows improve reconstruction of objects at different ranges, indicated by more distinct discontinuities (top) and a better reconstruction of the car (bottom). + +![](images/fb9a8f36523583b54143295e7e202e7b6362f22d1ed80c285b642be144e5b77c.jpg) +(b) GT + +![](images/79e3d4da210bbf932a45af88f201e2d2fdcba66f9a215d34e1a07e1873a40d50.jpg) + +![](images/6108bc1888daaeeaca85b93a02fc66c6288af647a327371b8a530e1aab5c9fa9.jpg) +(c) Rectangular +(d) Square + +two point clouds. We also follow the approach in [26] to evaluate the volumetric occupancy similarity. To do so, the point clouds are voxelized using a voxel size of $0.1m$ . A voxel is classified as occupied for each point cloud if it contains at least one point. We then calculate the Intersection-over-Unit (IoU) based on the occupancy. + +# 4.2. Ablation Studies + +Patch size: In Tab. 1 Baseline and Model 1-2, we compare the upsampling results with different patch sizes. We observe that the smaller patch $(2 \times 2)$ outperforms the larger one $(4 \times 4)$ while the proposed row patch $(1 \times 4)$ can further refine the network performance significantly. In particular, the row patch improves $32.6\%$ in MAE, $12.8\%$ in IoU, and $28.3\%$ in CD, compared to a $4 \times 4$ patch. As visualized in Fig. 3, the row patch better preserves sharpness around edges and corners compared to the larger patch sizes. + +Window geometry: To verify the effectiveness of modifying window geometry for range images, we conduct experiments with square and rectangular windows (Tab. 1 Model 2-3. Based on the quantitative results, we can observe the + +
ModelPatchWindowCP/PU/PSMCMAE ↓IoU ↑CD ↓
Baseline [5]4 × 4Square××0.71380.32500.1940
Model 12 × 2Square××0.62510.33370.1661
Model 21 × 4Square××0.48140.36670.1391
Model 31 × 4Rectangular××0.42480.40400.1218
Model 41 × 4Rectangular✓×0.42270.40840.1207
TULIP1 × 4Rectangular✓✓0.41850.41740.1207
TULIP-L1 × 4Rectangular✓✓0.37080.43290.0992
ViT-Unet*1 × 4-✓✓1.44840.29480.2665
+ +Table 1. Ablation study results for $4\times$ upsampling on KITTI. Input Resolution: $16\times 1024$ , Output Resolution: $64\times 1024$ . *We replaced Swin Transformer blocks with ViT blocks and followed the same training procedure as for TULIP. + +following: reconstruction quality produced by local attention within large square windows is inferior to the rectangular window concerning both 2D and 3D evaluation metrics, while they result in the same computational complexity. The underlying reason for this performance difference lies in the perceptive field of the windows. A wider window is more likely to include discontinuities between objects at different scales. Accordingly, the network performs better at separating objects from each other, which can visually be validated by clearer surface boundaries in Fig. 4. + +Further Adaptation and Refinement: Circular Padding (CP) helps maintain continuity and consistency at the edges of the panoramic images, but it has minor contributions to other areas. Patch Unmerging (PU), and Pixel Shuffle (PS) aim at upscaling the spatial resolution with rearrangement, avoiding information loss and parameter increase led by using additional de-convolutional layers. Although they are designed to improve efficiency rather than efficacy of the model, tested with those components (Model 4), a slight improvement in upsampling is still observable. Monte Carlo Dropout [19], as a post-processing step, helps to further reduce ghost points in between objects and we present more details in the Appendix. For TULIP-L, we increase the model capacity, using four Swin-Transformer layers instead of three in both encoder and decoder, which demonstrates an additional performance boost on all metrics. + +Transformer Block: From the previous discussion, we infer that Swin Transformer [33] is more advantageous than ViT (Vision Transformer) [17] in LiDAR upsampling. On one hand, prior works [13, 20, 27, 32] have pointed out the inferiority of ViT on smaller datasets due to lack of locality learning and non-overlapping attention. Relevant LiDAR datasets [21, 28] are generally small compared to RGB dataset [41]. On the other hand, by narrowing the patch size, the exponential increase of model capacity hinders sufficient training with ViT on range images, while Swin Transformer expands only linearly. To numerically show the distinction, we trained the network on the same dataset by replacing ViT blocks and applied the same patch size and additional components to the network for a fair comparison. In the last row of Tab. 1, it shows that TULIP significantly outperforms the one with ViT as the backbone. + +# 4.3. Benchmark Results + +Besides bilinear interpolation, which computes the interpolation weights from the four nearest neighbors, we evaluate against the state-of-the-art LiDAR upsampling approaches Implicit LiDAR Network (ILN) [26] and LiDAR-SR [43]. We also compare against several image super-resolution works: SRNO [51], and LIIF [10] learn implicit features for interpolation instead of 2D coordinates of new pixels while HAT [9] and Swin-IR [29] are pixel-based approaches that directly reconstruct high-resolution images from low + +
ModelMAE ↓IoU ↑CD ↓
CARLA 4x Output Resolution: 128 × 2048
Bilinear1.81280.13820.7262
SRNO [51]2.46400.13432.0230
HAT [9]1.60320.26980.6337
SWIN-IR [29]1.95600.27180.4840
LIIF [10]0.80640.35020.174
LIDAR-SR [43]0.82160.25810.2044
ILN [26]0.85920.50060.1855
TULIP (Ours)0.76990.51520.1028
TULIP-L (Ours)0.75390.53010.1001
+ +
KITTI 4x Output Resolution: 64 × 1024
Bilinear2.08920.10630.6000
SRNO [51]0.83500.20350.4417
HAT [9]0.68560.20350.2516
SWIN-IR [29]1.29720.27740.7347
LIIF [10]0.61430.32260.1916
LIDAR-SR [43]0.56740.10050.2165
ILN [26]1.05280.33420.2787
TULIP (Ours)0.41850.41740.1207
TULIP-L (Ours)0.37080.43290.0992
+ +
DurLAR 4x Output Resolution: 128 × 2048
Bilinear2.43840.12660.6346
SRNO [51]1.53960.15070.5108
HAT [9]1.78200.23530.1973
SWIN-IR [29]1.94160.21570.2279
LIIF [10]1.56720.24690.1548
LIDAR-SR [43]1.53120.13700.1128
ILN [26]1.57200.34300.0893
TULIP (Ours)1.54320.35620.06484
TULIP-L (Ours)1.55920.36540.06346
+ +Table 2. Quantitative comparison against state-of-the-art LiDAR and image super-resolution methods on different datasets. All methods are trained and evaluated on the same splits. + +resolution feature embeddings. Similar to our work, they are based on Swin Transformer [33]. We train all methods on the specific datasets using their default parameters. + +# 4.3.1 Simulation Results + +Similar to prior works, we first evaluate the performance on noise-free, simulated data. We train and test all approaches on the dataset presented in [26] that was recorded using the CARLA simulator and select the ground-truth range images of size $128 \times 2048$ that capture a vertical FoV of $30^{\circ}$ . We use a (20699/2618) train/test split. We observe that the performance between image super-resolution approaches varies greatly, and LIIF performs the best among them. While LIIF [10] achieves a comparable MAE to the LiDAR up- + +sampling methods, the 3D metrics are significantly worse, indicating that it is unsuitable for the underlying geometric input data. Nevertheless, LIIF [10] still outperforms the Li-DAR upsampling method in LIDAR-SR. The interpolation-based approach in ILN shows decent results for the noise-free data, as indicated by their high IoU. Nevertheless, our approaches achieve the best performance in all evaluated metrics, with the strongest improvement in Chamfer Distance, which validates that our design is well suited for the geometric reconstruction in range images. + +![](images/df7742b7922515108ac29509113717de52047ce622fdc8a44e8bfea0de0502df.jpg) +Figure 5. Qualitative results on KITTI: We provide visualizations for the top-4 results from Tab. 2 and Swin-IR [29]. Our approach outperforms other state-of-the-art methods in upsampling realistic point clouds. A: TULIP is the only approach that does not generate invalid ghost points in between the car (right bottom) and the wall behind. It can also be observed that all approaches except for ILN match the characteristic line pattern of the LiDAR on the ground (middle). B: Swin-IR and LIIF generate a blurry reconstruction. TULIP achieves the clearest point cloud and even reconstructs small details such as the car's side mirrors. C: Swin-IR, LIIF, and LiDAR-SR do not clearly separate the car from the wall behind. ILN generates a clear point cloud but does not properly reconstruct the shape of the cars. TULIP achieves the clearest point cloud, with distinct discontinuities between objects, and closely resembles the shape of the three cars. + +# 4.3.2 Real-World Results + +To evaluate our approach on real-world data, we downsample the DurLAR and KITTI datasets to a comparable number of frames to the CARLA dataset by temporally skipping frames in the dataset sequences. In particular, we have the following (train/test) split: DurLAR (24372/2508), and KITTI (20000/2500). + +KITTI: The KITTI dataset [21] was collected using a Velodyne HDL-64E LiDAR with a vertical FoV of $26.8^{\circ}$ and a resolution of $64 \times 1024$ . For the image super-resolution approaches, we observe comparable results and trends on KITTI as on the simulated data in Sec.4.3.1. Even though LIDAR-SR [43] is designed for LIDAR upsampling and outperforms image super-resolution approaches in terms of MAE, it achieves the lowest IoU among all approaches, which can be explained by a large amount of invalid floating points between real objects. LIIF [10] achieves better CD and comparable IoU compared to ILN. We observe that the IoU of ILN [26] is drastically lower than on CARLA. The interpolation scheme in ILN is thus not suitable to properly handle the lower resolution data with noise from the real-world sensor. Our approaches significantly outperform all competing methods, especially in terms of IoU as it can better handle discontinuities between objects. As indicated in the visualizations in Fig.5, our approach predicts fewer invalid float points between objects, as well as fewer blurry edges, which results in much clearer point clouds. + +DurLAR: The DurLAR dataset [28] was recorded using an Ouster OS1-128 LiDAR, that captures point clouds with a resolution of $128 \times 2048$ with a $45^{\circ}$ vertical FoV. Compared to CARLA and KITTI, we see a large performance drop for all approaches in MAE and IoU, which results from the larger maximum range and vertical FoV of the sensor as well as stronger sensor noise. We observe that all approaches designed for LiDAR upsampling clearly improve in Chamfer Distance compared to image super-resolution approaches. ILN and TULIP outperform all methods by a large margin in IoU. TULIP-L achieves the best 3D metrics, which validates that it can also handle the challenging DurLAR data better than prior works. + +Range Analysis: As the density of LiDAR point clouds decreases with increasing distance from the sensor, we additionally compare the performance at different ranges for a more fine-grained evaluation. We observe that our approach achieves better or similar performance compared to all approaches in all metrics at all ranges. Up to $30\mathrm{m}$ , TULIP clearly outperforms the baseline approaches. TULIP handles discontinuities between objects better than other approaches, which can be verified visually through sharper edges and fewer "floating" points between objects. This is especially reflected in the high IoU. At ranges above $30\mathrm{m}$ , our approach performs comparable to other methods. + +![](images/f679bc60bb378f808e359d1abf2ee2a83ac8901059ffd4dbbfca9d1bec73be0d.jpg) +(a) IoU↑ + +![](images/2c3124ec2de544d4df5d2aac0fef9a8d29fd46cad27975674c94019b6e819a8d.jpg) +(b) Chamfer distance $\downarrow$ +Figure 6. 3D error metrics visualized at different ranges for KITTI. + +Above this range, LiDAR data is extremely sparse, meaning that points neighboring in the low-resolution range image are actually distant in the Euclidean space. It is, therefore, not feasible to infer meaningful information about the scene from such neighboring points at a high range, which is also represented in the poor performance of all approaches. + +# 4.4. Failure Cases + +Although TULIP outperforms other state-of-the-art methods both qualitatively and quantitatively, the upsampling quality can still be limited in certain cases. For instance, as shown in Fig 7, our approach shows inferior upsampling results for the specific scene. The irregularity of the scene leads to high uncertainty, which makes the reconstruction noisy. Observing the details in Fig. 7a, the network fails to reconstruct the car at the top of the scene. + +![](images/fc987d03c9b37505c3ea807719857ad4aa7a209ce1045ff571c775fc2dd7b00a.jpg) +(a) Full scene +Figure 7. Failure Case: noisy point cloud generated by TULIP and missing reconstruction of the object. (Ours and GT) + +![](images/07a3bc037bda7aa20ddd231504b007ef05d3b7ece19155f3ba759c7beacf3949.jpg) +(b) Details + +# 5. Conclusion + +This work presents TULIP, a novel method for LiDAR Upsampling that achieves incredible performance in upsampling the range image. Our approach transforms 3D point clouds into 2D range images and performs the upsampling in 2D space. We build upon a Swin-Transformer-based network and specifically modify the patch partition and attention windows to better accommodate the characteristics of range images. Throughout various experiments, testing on three different benchmarks, it shows that TULIP outperforms state-of-the-art methods quantitatively in all evaluation metrics, and quantitatively, TULIP demonstrates more promising results in upsampling the realistic LiDAR data. Acknowledgement: This work was supported by Swiss National Science Foundation's NCCR DFab P3. + +# References + +[1] Anique Akhtar, Zhu Li, Geert Van der Auwera, Li Li, and Jianle Chen. Pu-dense: Sparse tensor-based point cloud geometry upsampling. IEEE Transactions on Image Processing, 31:4133-4148, 2022. 3 +[2] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. ECCVW What is Motion For?, 2022. 2 +[3] Borna Bešić, Nikhil Gosala, Daniele Cattaneo, and Abhinav Valada. Unsupervised domain adaptation for lidar panoptic segmentation. IEEE Robotics and Automation Letters, 7(2): 3404-3411, 2022. 1 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2 +[5] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 3, 4, 5, 6 +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2 +[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2 +[8] Tso Yuan Chen, Ching Chun Hsiao, and Ching-Chun Huang. Density-imbalance-eased lidar point cloud upsampling via feature consistency learning. IEEE Transactions on Intelligent Vehicles, 2022. 3 +[9] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367-22377, 2023. 2, 4, 6, 7 +[10] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 6, 7, 8 +[11] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12312-12321, 2023. 2 +[12] Xinjing Cheng, Peng Wang, and Ruigang Yang. Depth estimation via affinity learned with convolutional spatial propagation network. In Proceedings of the European conference on computer vision (ECCV), pages 103-119, 2018. 3 +[13] Yin Dai, Fayu Liu, Weibing Chen, Yue Liu, Lifu Shi, Sheng Liu, Yuhang Zhou, et al. Swin mae: Masked autoencoders + +for small datasets. Computers in Biology and Medicine, 161: 107037, 2023. 6 +[14] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2 +[15] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 391-407. Springer, 2016. 2 +[16] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 5 +[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 6 +[18] George Eskandar, Sanjeev Sudarsan, Karim Guirguis, Janaranjani Palaniswamy, Bharath Somashekar, and Bin Yang. Hals: A height-aware lidar super-resolution framework for autonomous driving. arXiv preprint arXiv:2202.03901, 2022. 1, 3, 5 +[19] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pages 1050-1059. PMLR, 2016. 6 +[20] Hanan Gani, Muzammal Naseer, and Mohammad Yaqub. How to train vision transformer on small-scale datasets? In 33rd British Machine Vision Conference, 2022. 6 +[21] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 2, 5, 6, 8 +[22] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Penet: Towards precise and efficient image guided depth completion. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13656-13662. IEEE, 2021. 3 +[23] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2, 2019. 2 +[24] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 2 +[25] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1637-1645, 2016. 2 +[26] Youngsun Kwon, Minhyuk Sung, and Sung-Eui Yoon. Implicit lidar network: Lidar super-resolution via interpolation weight prediction. In 2022 International Conference on + +Robotics and Automation (ICRA), pages 8424-8430. IEEE, 2022. 1, 2, 3, 5, 6, 7, 8 +[27] Seung Hoon Lee, Seunghyun Lee, and Byung Cheol Song. Vision transformer for small-size datasets. arXiv preprint arXiv:2112.13492, 2021. 6 +[28] L. Li, K.N. Ismail, H.P.H. Shum, and T.P. Breckon. Durlar: A high-fidelity 128-channel lidar dataset with panoramic ambient and reflectivity imagery for multi-modal autonomous driving applications. In Proc. Int. Conf. on 3D Vision. IEEE, 2021. 2, 5, 6, 8 +[29] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 1, 2, 3, 4, 5, 6, 7 +[30] Yuankai Lin, Tao Cheng, Qi Zhong, Wending Zhou, and Hua Yang. Dynamic spatial propagation network for depth completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1638-1646, 2022. 3 +[31] Zhixin Ling, Zhen Xing, Xiangdong Zhou, Manliang Cao, and Guichun Zhou. Panoswin: A pano-style swin transformer for panorama understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2023. 4 +[32] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco De Nadai. Efficient training of visual transformers with small datasets. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 6 +[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 4, 5, 6, 7 +[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 5 +[35] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457-466, 2022. 2, 4 +[36] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 120-136. Springer, 2020. 3 +[37] Jiaxiong Qiu, Zhaopeng Cui, Yinda Zhang, Xingdi Zhang, Shuaicheng Liu, Bing Zeng, and Marc Pollefeys. Deepidar: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3313-3322, 2019. 3 +[38] Shi Qiu, Saeed Anwar, and Nick Barnes. Pu-transformer: Point cloud upsampling transformer. In Proceedings of the Asian Conference on Computer Vision, pages 2475-2493, 2022. 3 +[39] Kyeongha Rho, Jinsung Ha, and Youngjung Kim. Guideformer: Transformers for image guided depth completion. + +In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6250-6259, 2022. 3 +[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3 +[41] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 6 +[42] Artem Savkin, Yida Wang, Sebastian Wirkert, Nassir Navab, and Federico Tombari. Lidar upsampling with sliced wasserstein distance. IEEE Robotics and Automation Letters, 8(1): 392-399, 2022. 1, 3 +[43] Tixiao Shan, Jinkun Wang, Fanfei Chen, Paul Szenher, and Brendan Englot. Simulation-based lidar super-resolution for ground vehicles. Robotics and Autonomous Systems, 134: 103647, 2020. 1, 2, 3, 5, 6, 7, 8 +[44] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 4, 5 +[45] Xiaopeng Sun, Weiqi Li, Zhenyu Zhang, Qiufang Ma, Xuhan Sheng, Ming Cheng, Haoyu Ma, Shijie Zhao, Jian Zhang, Junlin Li, et al. Opdn: Omnidirectional position-aware deformable network for omnidirectional image superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1293-1301, 2023. 2 +[46] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing, 30:1116-1129, 2020. 3 +[47] Larissa T Triess, David Peter, Christoph B Rist, Markus Enzweiler, and J Marius Zöllner. Cnn-based synthesis of realistic high-resolution lidar data. In 2019 IEEE Intelligent Vehicles Symposium (IV), pages 1512-1519. IEEE, 2019. 3 +[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[49] Suhong Wang, Hongqing Wang, Shufeng She, Yanping Zhang, Qingju Qiu, and Zhifeng Xiao. Swin-t-nfc crfs: An encoder-decoder neural model for high-precision UAV positioning via point cloud super resolution and image semantic segmentation. Computer Communications, 197:52–60, 2023. 1 +[50] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 5 + +[51] Min Wei and Xuesong Zhang. Super-resolution neural operator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18247-18256, 2023. 6, 7 +[52] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Baobei Xu, Jun Li, and Jian Yang. Rignet: Repetitive image guided network for depth completion. In European Conference on Computer Vision, 2021. 3 +[53] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1951-1960, 2019. 1 +[54] Shuquan Ye, Dongdong Chen, Songfang Han, Ziyu Wan, and Jing Liao. Meta-pu: An arbitrary-scale upsampling network for point cloud. IEEE transactions on visualization and computer graphics, 28(9):3206-3218, 2021. 3 +[55] Fanghua Yu, Xintao Wang, Mingdeng Cao, Gen Li, Ying Shan, and Chao Dong. Osrt: Omnidirectional image superresolution with distortion-aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13283-13292, 2023. 2 +[56] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-net: Point cloud upsampling network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2790-2799, 2018. 1, 3 +[57] Heeseung Yun, Sehun Lee, and Gunhee Kim. Panoramic vision transformer for saliency detection in 360 videos. In ECCV, 2022. 4 +[58] Wentai Zhang, Haoliang Jiang, Zhangsihao Yang, Soji Yamakawa, Kenji Shimada, and Levent Burak Kara. Data-driven upsampling of point clouds. Computer-Aided Design, 112:1-13, 2019. 1 +[59] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image restoration. TPAMI, 2020. 2 +[60] Wenbo Zhao, Xianming Liu, Zhiwei Zhong, Junjun Jiang, Wei Gao, Ge Li, and Xiangyang Ji. Self-supervised arbitrary-scale point clouds upsampling via implicit neural representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1999-2007, 2022. 3 +[61] Yifan Zhao, Le Hui, and Jin Xie. Sspu-net: Self-supervised point cloud upsampling via differentiable rendering. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2214-2223, 2021. 1, 3 +[62] Örkény Zováthi, Balázs Pálffy, Zsolt Jankó, and Csaba Benedek. St-depthnet: A spatio-temporal deep network for depth completion using a single non-repetitive circular scanning lidar. IEEE Robotics and Automation Letters, 2023. 2 \ No newline at end of file diff --git a/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/images.zip b/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..22f9e2245e629360e936fa2512d0e9c5ff4b8bad --- /dev/null +++ b/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd9c1f688c963f6f433a87659516496ca34b9899e3ae72eae5fa8a550c1e3863 +size 591231 diff --git a/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/layout.json b/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d987e6d6367199a51a16a40b19610d2d1fdace29 --- /dev/null +++ b/2024/TULIP_ Transformer for Upsampling of LiDAR Point Clouds/layout.json @@ -0,0 +1,8033 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 107, + 103, + 486, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 103, + 486, + 121 + ], + "spans": [ + { + "bbox": [ + 107, + 103, + 486, + 121 + ], + "type": "text", + "content": "TULIP: Transformer for Upsampling of LiDAR Point Clouds" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "spans": [ + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "text", + "content": "Bin Yang" + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "text", + "content": ", Patrick Pfreundschuh" + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "text", + "content": ", Roland Siegwart" + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "text", + "content": ", Marco Hutter" + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "text", + "content": ", Peyman Moghadam" + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "inline_equation", + "content": "^{3,4\\dagger}" + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "text", + "content": ", Vaishakh Patil" + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "text", + "content": "Autonomous Systems Lab, ETH Zurich, Switzerland, " + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "text", + "content": "Robotic Systems Lab, ETH Zurich, Switzerland \n" + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "text", + "content": "School of EER, Queensland University of Technology (QUT), Australia, " + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 53, + 142, + 592, + 185 + ], + "type": "text", + "content": "Data61, CSIRO, Brisbane, Australia" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 187, + 533, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 187, + 533, + 199 + ], + "spans": [ + { + "bbox": [ + 110, + 187, + 533, + 199 + ], + "type": "text", + "content": "{biyang, patripfr, rolandsi, mahutter, patilv}@ethz.ch, peyman.moghadam@csiro.au" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 252, + 290, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 290, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 290, + 491 + ], + "type": "text", + "content": "LiDAR Upsampling is a challenging task for the perception systems of robots and autonomous vehicles, due to the sparse and irregular structure of large-scale scene contexts. Recent works propose to solve this problem by converting LiDAR data from 3D Euclidean space into an image super-resolution problem in 2D image space. Although their methods can generate high-resolution range images with fine-grained details, the resulting 3D point clouds often blur out details and predict invalid points. In this paper, we propose TULIP, a new method to reconstruct high-resolution LiDAR point clouds from low-resolution LiDAR input. We also follow a range image-based approach but specifically modify the patch and window geometries of a Swin-Transformer-based network to better fit the characteristics of range images. We conducted several experiments on three public real-world and simulated datasets. TULIP outperforms state-of-the-art methods in all relevant metrics and generates robust and more realistic point clouds than prior works. The code is available at https://github.com/ethz-as1/TULIP.git." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 517, + 128, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 128, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 128, + 529 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 536, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 287, + 693 + ], + "type": "text", + "content": "Light Detection And Ranging (LiDAR) is one of the most common sensors for perception in various fields of autonomy, such as autonomous driving, and unmanned aerial vehicles (UAV). LiDARs are used to generate 3D point clouds of the scene. These point clouds are essential for mapping, localization, and object detection tasks. However, the accuracy of these tasks often depends on the resolution of the point cloud [53]. Furthermore, the resolution of a LiDAR is inherently associated with increased energy consumption and cost, making its use impractical for various applications. LiDARs also have different vertical and horizontal resolutions. The vertical resolution of rotating 3D LiDARs is typically much lower than the horizontal res" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 226, + 426, + 307 + ], + "blocks": [ + { + "bbox": [ + 310, + 226, + 426, + 307 + ], + "lines": [ + { + "bbox": [ + 310, + 226, + 426, + 307 + ], + "spans": [ + { + "bbox": [ + 310, + 226, + 426, + 307 + ], + "type": "image", + "image_path": "e09006fdab21513ee203b9986e2f4dc6a5bb4922c1b25eafbe50755107db8b76.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 308, + 382, + 317 + ], + "lines": [ + { + "bbox": [ + 353, + 308, + 382, + 317 + ], + "spans": [ + { + "bbox": [ + 353, + 308, + 382, + 317 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 429, + 228, + 542, + 307 + ], + "blocks": [ + { + "bbox": [ + 429, + 228, + 542, + 307 + ], + "lines": [ + { + "bbox": [ + 429, + 228, + 542, + 307 + ], + "spans": [ + { + "bbox": [ + 429, + 228, + 542, + 307 + ], + "type": "image", + "image_path": "87a798a93a376080981a05716c735f8b529f30f9953116b01f80034c8ffc0275.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 308, + 512, + 316 + ], + "lines": [ + { + "bbox": [ + 456, + 308, + 512, + 316 + ], + "spans": [ + { + "bbox": [ + 456, + 308, + 512, + 316 + ], + "type": "text", + "content": "(b) Ground Truth" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 321, + 425, + 397 + ], + "blocks": [ + { + "bbox": [ + 308, + 321, + 425, + 397 + ], + "lines": [ + { + "bbox": [ + 308, + 321, + 425, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 425, + 397 + ], + "type": "image", + "image_path": "0761cadd0d2db11c46775303ea0145c26832e235fcba08657876c5ea151c4623.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 399, + 392, + 408 + ], + "lines": [ + { + "bbox": [ + 340, + 399, + 392, + 408 + ], + "spans": [ + { + "bbox": [ + 340, + 399, + 392, + 408 + ], + "type": "text", + "content": "(c) SwinIR [29]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 428, + 319, + 541, + 398 + ], + "blocks": [ + { + "bbox": [ + 428, + 319, + 541, + 398 + ], + "lines": [ + { + "bbox": [ + 428, + 319, + 541, + 398 + ], + "spans": [ + { + "bbox": [ + 428, + 319, + 541, + 398 + ], + "type": "image", + "image_path": "049793755ba3abf39502d794c1ddf168595b53ca9f34e048d83b236367bbaf22.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 399, + 512, + 408 + ], + "lines": [ + { + "bbox": [ + 454, + 399, + 512, + 408 + ], + "spans": [ + { + "bbox": [ + 454, + 399, + 512, + 408 + ], + "type": "text", + "content": "(d) TULIP (Ours)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 410, + 545, + 453 + ], + "lines": [ + { + "bbox": [ + 306, + 410, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 410, + 545, + 453 + ], + "type": "text", + "content": "Figure 1. TULIP generates more realistic LiDAR point clouds from low-resolution LiDAR input (a) and outperforms the state-of-the-art image super-resolution approach [29] in upsampling with fine-grained details of different objects in the scene." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 462, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 605 + ], + "type": "text", + "content": "olution. These limitations make an upsampling technique necessary to increase the resolution of the LiDAR data, especially in the vertical direction. Moreover, upsampling LiDAR data has the potential to counter domain shift problems in LiDAR-based learning methods. Notably, the use of lower resolution LiDAR data in a system that has been trained on higher resolution data leads to a significant drop in performance [3]. Therefore, such upsampling techniques can not only help to mitigate the domain gap by creating a virtual high-resolution sensor that matches the target domain [42, 61] but also reduce the high cost of collecting new LiDAR data, annotating and retraining the methods." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Deep learning has led to remarkable advancements in LiDAR upsampling techniques in recent years. Several recent methods [42, 56, 58] have focused on learning the upsampling process within the 3D Euclidean space. However, the processing of 3D data in deep learning can be quite resource-intensive. One way to address this challenge is the representation of the point cloud as a range image [18, 26, 43, 49]. This approach facilitates the adoption of well-established image-based super-resolution methods." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 160, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 160, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 160, + 713 + ], + "type": "text", + "content": "†Authors share last authorship." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15354" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "Such methods mostly employ autoencoder-style networks with convolutional and deconvolutional layers [14, 43, 62]. However, these methods are not directly transferable to range images. Range images generated by the projection of sparse 3D point clouds have distinct sharp edges at the object boundaries. Convolutional neural network-based methods often induce edge smoothing due to the regularization effect, which limits their usability for LiDAR range image processing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 178, + 289, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 178, + 289, + 708 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 289, + 708 + ], + "type": "text", + "content": "As an alternative, Transformer [48] has shown great success in RGB image super-resolution and produces less blurry outputs [29]. Vision transformers, however, face challenges such as the need for large amounts of training data [9, 29, 35] and the global computation of self-attention, limiting the capture of local information in the range images. To address this, Swin Transformer [33] introduced tokenizing the input image and applying the self-attention mechanism locally in independent windows. While the modifications to Swin Transformer have led to robust training and promising results compared to Vision Transformer, its default structure and settings are not ideally suited for processing LiDAR range images. These range images consist of single-channel data, representing 3D spatial information, as opposed to three-channel color images that depict visual appearance. Furthermore, range images feature large, smooth areas within 3D objects, separated by distinct sharp edges between objects. As a result, minor inaccuracies in the 2D prediction (e.g. blurry output) can lead to significant differences in the projected 3D occupancy. This non-uniform distribution of image details further complicates network training. The challenging details in the range image are concentrated in a few specific pixels, contrasting with RGB images where relevant details are evenly distributed. This makes range images fundamentally different from RGB images. Inspired by the limitations of state-of-the-art image upsampling methods and the potential of Swin-Transformer [33], we propose a novel network named TULIP. Our proposed geometry-aware architecture is tailored to accommodate LiDAR range image data. The proposed method takes a low-resolution range image as input and produces a high-resolution range image, which can then be projected into a 3D point cloud. To accommodate LiDAR range image geometry, we utilize one-line row patches to tokenize the input range image, as opposed to the square patches used for RGB images. This approach helps to preserve vertical information for the upsampling process while capturing boundary discontinuities effectively. This also creates a trade-off between local detail capture and model complexity in the training process. Additionally, we incorporate a non-square window for local self-attention computation. This enhances learning of spatial contexts at different scales in the range image. Further, we extensively train and evaluate TULIP on two real-world and" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "one simulated autonomous driving datasets: KITTI [21], DurLAR [28] and CARLA [26], respectively. Comparisons to existing approaches demonstrate that TULIP outperforms the state-of-the-art methods on all three datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 129, + 398, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 129, + 398, + 143 + ], + "spans": [ + { + "bbox": [ + 306, + 129, + 398, + 143 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 152, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 152, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 545, + 247 + ], + "type": "text", + "content": "Image super-resolution has seen significant progress in recent years. The technique aims to construct high-resolution (HR) images from low-resolution (LR) observations, often leveraging advances in Convolutional Neural Networks (CNNs) to enhance the fidelity and detail of images for better visualization and information extraction. Subsequently, several enhanced frameworks have been developed [14, 15, 24, 25, 59]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 247, + 545, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 414 + ], + "type": "text", + "content": "Driven by their success in the field of natural language processing (NLP) [4, 23], Transformers [48] have been extended to solve a variety of vision-related tasks such as object detection [6] and semantic segmentation [2, 7]. Such Vision Transformers (ViT) [17] excel at learning to focus on relevant image regions by exploring global interactions among different regions. Their impressive performance has led to their adoption in image restoration tasks [9, 11, 29] as well. While they brought significant improvements in RGB image super-resolution, Vision Transformer [17] comes with the drawback of the quadratic computational complexity of self-attention. Additionally, they capture mostly global dependencies within the data and require large amounts of data for training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": "To address these challenges, Swin Transformer [33] has been proposed. Unlike the Vision Transformer, which relies on global self-attention across the entire image, the Swin Transformer employs a local window-based attention mechanism and establishes a pyramid-like architecture. This approach processes images hierarchically by gradually merging smaller image patches into larger ones. This strategy enables more effective handling of different scales and facilitates the processing of multi-scale features, making it particularly suitable for range image processing. Furthermore, due to Swin Transformer's hierarchical structure, the increase in the number of model parameters scales linearly with the input image size, which mitigates the challenges associated with handling large range images. Additionally, they require less training data than the classical Vision Transformer, which is beneficial for LiDAR upsampling as LiDAR datasets are typically magnitudes smaller than RGB datasets used for Vision Transformer. The Swin and Vision Transformers have also been used for image super-resolution on omnidirectional panoramic camera images [45, 55]. However, these approaches are specifically designed for the unique characteristics of these cameras (e.g. strong distortions, spherical continuity) that are vastly different from the LiDAR sensing model, which renders them unsuitable for application to range images." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15355" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 696 + ], + "type": "text", + "content": "Other related works have focussed on upsampling high-density point clouds from low-density point clouds [38, 54, 56, 60, 61]. While those approaches also provide qualitative results for LiDAR point clouds, they focus on increasing the general 3D point density of point clouds of single objects instead of a large scene. Another area closely related to LiDAR upsampling is depth completion [12, 22, 30, 36, 37, 39, 46, 52]. The primary objective of depth completion is to improve sparse depth estimates acquired from LiDAR by integrating data from multiple sensors, predominantly RGB cameras. The result of depth completion methods is a dense depth map that provides depth values to every pixel in the input depth map. Nonetheless, this field slightly diverges from LiDAR upsampling due to its reliance on multimodal sensors and its common deployment within a restricted field of view, making it not directly comparable within the scope of this paper. LiDAR point clouds exhibit specific characteristics, such as a distinct point pattern (due to the stacked lasers in rotating 3D LiDARs) and a decrease in point density with increasing distance from the sensor. Differently from arbitrary point cloud upsampling, LiDAR upsampling tries to mimic a realistic point cloud of a high-resolution LiDAR given the point cloud of low-resolution LiDAR and thus targets a different result than the approaches above. Due to the difficulty and high computational cost of upsampling LiDAR point clouds in 3D space [1, 8, 42], most works [26, 43, 47] represent the point cloud as a range image and perform the LiDAR upsampling task in 2D image space. LiDAR-CNN [47] introduces a CNN-based architecture with semantic and perceptual guidance. The method specifies two further loss functions besides a per-point reconstruction loss to acquire a better synthesis of the high-resolution range image. However, this approach is limited to data with semantic annotations, which drastically reduces its applicability. LiDAR-SR [43] deploys a CNN-based network with U-Net [40] architecture. It additionally uses Monte-Carlo dropout post-processing to reduce the amount of noisy points in the prediction. Approaches based on convolutional operations tend to fail to reconstruct the sharpness in the range image effectively. While the previous approaches directly predict the range image, Implicit LiDAR Network (ILN) [26] uses an implicit neural architecture that learns interpolation weights to fill in new pixels instead of their values directly. Although the method outperforms CNN-based approaches in training speed and preserving the geometrical details in the input, it can still suffer from limited neighboring information in the LiDAR data, especially in areas distant from the sensor due to the extremely sparse input data. Differently from the previous works, our approach builds on Swin Transformer as a backbone and is specifically customized to effectively process range images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 306, + 71, + 389, + 86 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 389, + 86 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 389, + 86 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 304, + 93, + 547, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 93, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 93, + 547, + 502 + ], + "type": "text", + "content": "LiDAR upsampling is achieved by upsampling a range image, effectively transforming it from a 3D upsampling problem into a 2D image super-resolution problem [18, 26, 43, 47]. It is, therefore, an evident approach to build on RGB image super-resolution. However, image super-resolution aims to solve a different problem than LiDAR upsampling. Image super-resolution enhances the visual appearance of a low-resolution RGB image by recreating a high-resolution image. While quantitative metrics for evaluating these works exist, there is no one specific correct solution. Differently, LiDAR upsampling tries to recreate a high-resolution, 3D LiDAR sensor output from a low-resolution LiDAR input point cloud. Range images contain more geometrical and spatial contexts rather than visual information, and the scene's geometry strictly dictates the correct solution. The underlying sensing model is also drastically different, as LiDARs are active sensors. The respective range pixel values purely depend on distance, while pixels in an RGB camera depend on many factors, such as scene appearance, lighting conditions, exposure time, and white balance. Furthermore, range images have a highly asymmetric aspect ratio, typically between 1:8 to 1:64, which differs drastically from regular camera images. Accordingly, most works [18, 43, 47] only perform vertical upsampling, in contrast to image super-resolution, which operates in both dimensions. A naive application of state-of-the-art image super-resolution networks [29] to range images does not result in adequate performance (Fig. 1). To this end, we developed TULIP, a novel range image-based LiDAR Upsampling method. Based on the observations above, we specifically adapt a network based on SwinUnet [5] to incorporate range images better. In the following sections, we will describe the technical details of our method, with a special focus on how our approach accommodates range images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 514, + 419, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 419, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 419, + 526 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "A LiDAR point cloud consists of points captured during one revolution " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\mathbf{p}_1,\\dots ,\\mathbf{p}_n\\}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": ". Each measurement represents a 3D point " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i = \\{x_i,y_i,z_i\\}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": ". As input, we assume a low-resolution point cloud " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_l" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "n_l = H_l\\times W_l" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " points, where " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "H_{l}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "W_{l}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " correspond to vertical and horizontal resolution. We aim to predict a high-resolution point cloud " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_h" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": ", that has the same field-of-view(FoV) as " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_l" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " but contains " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "n_h = \\beta *n_l" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " points. In our experiments, we set " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\beta = 4" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": ". We only increase vertical resolution (amount of LiDAR beams), i.e., " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "n_h = H_h\\times W_h" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "H_{h} = \\beta *H_{l}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "W_{h} = W_{l}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": ". We project the point cloud into a 2D range image. In this image, each row and column coordinate " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "v,u" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " correspond to the respective elevation and azimuth angles of the LiDAR points, while the pixel value contains the range of the point " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "r = \\sqrt{x^2 + y^2 + z^2}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": ". The image coordinate of a 3D point" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15356" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 262, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 262, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 262, + 85 + ], + "type": "text", + "content": "can be calculated using a spherical projection model:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 287, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 287, + 125 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{l} u \\\\ v \\end{array} \\right] = \\left[ \\frac {\\frac {W}{2} - \\frac {W}{2 \\pi} a r c t a n (\\frac {y}{x})}{\\Theta_ {m a x} - \\Theta_ {m i n}} * (\\Theta_ {m a x} - a r c t a n (\\frac {z}{\\sqrt {x ^ {2} + y ^ {2}}}) \\right] (1)", + "image_path": "c0ccee30d541c6e1eb4bad72704e2c2ab16e1566b0cd0e3f1983a6ae0ba435e5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 129, + 287, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 129, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 129, + 287, + 214 + ], + "type": "inline_equation", + "content": "\\Theta_{max}" + }, + { + "bbox": [ + 46, + 129, + 287, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 129, + 287, + 214 + ], + "type": "inline_equation", + "content": "\\Theta_{min}" + }, + { + "bbox": [ + 46, + 129, + 287, + 214 + ], + "type": "text", + "content": " correspond to the limits of the vertical field of view of the LiDAR. We formulate the LiDAR upsampling problem as an image upsampling problem, i.e. we aim to predict a high-resolution range image " + }, + { + "bbox": [ + 46, + 129, + 287, + 214 + ], + "type": "inline_equation", + "content": "I_h \\in \\mathbf{R}^{1 \\times H_h \\times W}" + }, + { + "bbox": [ + 46, + 129, + 287, + 214 + ], + "type": "text", + "content": " given a low-resolution range image " + }, + { + "bbox": [ + 46, + 129, + 287, + 214 + ], + "type": "inline_equation", + "content": "I_l \\in \\mathbf{R}^{1 \\times H_l \\times W}" + }, + { + "bbox": [ + 46, + 129, + 287, + 214 + ], + "type": "text", + "content": ". We can then calculate the high-resolution point cloud " + }, + { + "bbox": [ + 46, + 129, + 287, + 214 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_h" + }, + { + "bbox": [ + 46, + 129, + 287, + 214 + ], + "type": "text", + "content": " by inverting the projection in Eq. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 220, + 231, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 220, + 231, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 220, + 231, + 232 + ], + "type": "text", + "content": "3.2. Architecture Design and Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 238, + 288, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 288, + 658 + ], + "type": "text", + "content": "Our network builds upon Swin-Unet [5] which was originally designed for image segmentation. We deploy a U-shaped network structure featuring skip connections that link the encoder and decoder modules. The fundamental building block of TULIP is the Swin Transformer [33]. TULIP first patchifies and maps the input to a high-dimensional feature space. The tokenized low-level feature maps are passed through a multi-stage encoder. Each stage contains a pair of Swin Transformer [33] blocks and a patch merging layer to downsample the resolution by a factor of two and increase the dimension by a factor of four. This step realizes the hierarchic computation and extraction of multi-scale features through encoding. Within each stage, multiple instances of the self-attention mechanism are performed locally in parallel, which is the so-called Window Multi-Head Self Attention (W-MSA). After passing through a two-layer MLP and residual connection with the input, the feature vector is further passed to the second part of the Swin Transformer block, utilizing a Shifted Window Multi-Head Self Attention (SW-MSA) which extends W-MSA by employing a shifted window operation. This enhances the model capacity as it compensates for the lack of interaction between the local windows in W-MSA. The decoder, which realizes the upsampling has a symmetric design to the encoder and operates in a reverse way of the encoder. The resolution of the feature maps is first expanded, and the dimension is reduced by a factor of two accordingly. We feed existing geometrical information via skip connections. The feature maps are transformed into a single-channel, high-resolution range image in the last layer. The reconstruction head comprises a " + }, + { + "bbox": [ + 46, + 238, + 288, + 658 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 238, + 288, + 658 + ], + "type": "text", + "content": " convolutional layer for feature expansion, followed by a Leaky ReLU activation, a pixel shuffle layer [44], and another " + }, + { + "bbox": [ + 46, + 238, + 288, + 658 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 238, + 288, + 658 + ], + "type": "text", + "content": " convolutional layer for final projection. As the loss function, we select the pixel-wise L1 loss. Details of the network can be found in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 671, + 135, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 135, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 135, + 682 + ], + "type": "text", + "content": "3.2.1 Tokenization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Tokenization is done by a patch partition layer that creates an initial feature embedding from the input range image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "text", + "content": "This initial feature representation significantly influences the network's performance. Specifically, the selection of this layer determines how the inherent characteristics, patterns, and relevant information of the LiDAR range image data are encoded into a format that the network can effectively learn from. Most transformer-based RGB image super-resolution approaches utilize a relatively large patch size to tokenize the input [9, 29, 35]. These larger patches enable the construction of a global spatial context, facilitating an understanding of the overall structure and high-level features of the RGB image. These networks employ square-shaped patches and implement upsampling in both vertical and horizontal directions. In contrast, LiDAR upsampling primarily aims to enhance the vertical resolution of the input, given that range images possess properties that are vastly different from RGB images. Furthermore, attention across more pixels in range images is less useful than in RGB images, since there is almost no geometrical relation for the spatial contexts that are far from each other in the 2D range image space. Motivated by this, we propose two adjustments to effectively process range images within a Swin-Unet:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 338, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 495 + ], + "type": "text", + "content": "Row-Based Patch Partition: Our model builds row patches with a dimension of " + }, + { + "bbox": [ + 304, + 338, + 545, + 495 + ], + "type": "inline_equation", + "content": "1 \\times 4" + }, + { + "bbox": [ + 304, + 338, + 545, + 495 + ], + "type": "text", + "content": " for range image tokenization. The new patch geometry is designed to retain full vertical information while compressing horizontally. This aligns with our objective of conserving and extending vertical details in the range image. Besides that, row patches excel at capturing boundary discontinuities between distinct objects and the background within the scene. Moreover, the selection of the patch size strikes a balance between enhancing model performance and limiting model capacity. Smaller patches can capture more detailed local information, but they also increase the number of model parameters, which can slow down the model's training process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 496, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 605 + ], + "type": "text", + "content": "Circular Padding (CP) in Horizontal Dimension: Instead of adding zero pixels padding, our model uses circular padding in the horizontal dimension. This avoids introducing artificial features and naturally matches the sensor model of a rotating 3D LiDAR. This preserves accurate neighbor relations along the edges, which would otherwise be compromised during the projection to the 2D range image. The Circular Padding (CP) also allows us to fit flexible input sizes without changing the patch geometry." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 622, + 522, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 622, + 522, + 634 + ], + "spans": [ + { + "bbox": [ + 305, + 622, + 522, + 634 + ], + "type": "text", + "content": "3.2.2 Non-Square window for local self-attention" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 715 + ], + "type": "text", + "content": "Inside Swin Transformer, input feature maps are window-partitioned along width and height to compute the self-attention locally in each window. In addition to square-shaped RGB images, some recent work applies a square window for panoramic data as well [31, 57]. However, using square windows on range images is unfavorable. On one" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15357" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 68, + 511, + 205 + ], + "blocks": [ + { + "bbox": [ + 85, + 68, + 511, + 205 + ], + "lines": [ + { + "bbox": [ + 85, + 68, + 511, + 205 + ], + "spans": [ + { + "bbox": [ + 85, + 68, + 511, + 205 + ], + "type": "image", + "image_path": "058fa761a4909494e7b5e950bae2e62b7b95ebedb64e4a42db47afd075e1f759.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 212, + 547, + 269 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 547, + 269 + ], + "type": "text", + "content": "Figure 2. Overview of TULIP: The low-resolution point cloud is spherically projected into the range image. In the training and inference, the range image is tokenized into feature embeddings without compressing the spatial resolution along the vertical axis. We adopt the U-Net-based network presented in [5] for range image upsampling. We follow the hierarchical structure of Swin Transformer [33] for feature extraction and use the non-square window geometry for computing the local self-attention. The reconstruction head generates the high-resolution range image from the latent features and then, the point cloud can be obtained by projecting pixels back into the 3D space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 289, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 553 + ], + "type": "text", + "content": "hand, small windows can focus more on attention within near areas but lead to increasing computational complexity and can struggle with the reconstruction of objects at a further distance in the scene. On the other hand, a larger square window can blend scene-related contexts at different scales, which can potentially degrade the network performance. In a prior study done by Eskandar et al. [18], they demonstrated a significant enhancement in range image superresolution by separately upsampling the upper and lower portions of the data. In contrast to their approach of utilizing CNN-based shallower and deeper network branches for distinct feature extraction, our proposal involves the direct partitioning of the input using a rectangular window. This window geometry facilitates robust attention interaction among distant points through W-MSA and subtle crossattention between objects at multiple scales via SW-MSA and hierarchical processing. In addition, range images, as typical panorama data, inherently contain abundant information along the horizontal direction. Therefore, directing attention more toward image width rather than height can aid the network in capturing a greater amount of geometrical information. At the same time, the computational complexity remains the same as using a large square window." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 571, + 234, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 571, + 234, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 234, + 583 + ], + "type": "text", + "content": "3.2.3 Further Adaptation and Refinement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Patch Unmerging (PU) and Pixel Shuffle (PS): Similar to the patch merging layer that is responsible for down-sampling at the encoder stage, a mechanism is necessary within the decoder to upsample the feature maps. Conventional CNN-based approaches [43, 50] achieve upsampling through the transposed convolutional operation. As they tend to smooth out image sharpness [29], we reversed the operation of patch merging, re-arranging the channels of patches into their respective positions in a grid to up-sample the spatial resolution of input feature maps, which we" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 277, + 545, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 326 + ], + "type": "text", + "content": "denote as patch unmerging. Furthermore, to reconstruct the final range image, we built the reconstruction head upon the upsampled feature maps. The implementation of the reconstruction module is based on the pixel shuffle layer [44]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 327, + 546, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 546, + 411 + ], + "type": "text", + "content": "Monte Carlo Dropout: We additionally use Monte Carlo Dropout [43] to filter unreliable points by thresholding the uncertainty. MC-Dropout executes several feed-forward passes during inference with different active dropouts, which yields a distribution of outputs. We refine the results by removing points with high output variance, as they often come down to noisy, invalid points." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 426, + 388, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 388, + 440 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 388, + 440 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 446, + 432, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 446, + 432, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 446, + 432, + 459 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 468, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 588 + ], + "type": "text", + "content": "Datasets: We conduct experiments on three different datasets, that include the two large-scale real-world datasets DurLAR [28] and KITTI [21] (Sec. 4.3.2), as well as on a dataset [26] that was generated using the CARLA simulator [16] (Sec. 4.3.1). We select test sequences that are recorded in different locations than the train set to avoid spatial overlap. We vertically downsample the high-resolution range images with a factor of four by skipping the respective lines to generate low-resolution input images. For ablation studies in Sec. 4.2, we select the KITTI dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 590, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 545, + 651 + ], + "type": "text", + "content": "Implementation Details: In all experiments, we use distributed processing with " + }, + { + "bbox": [ + 304, + 590, + 545, + 651 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 304, + 590, + 545, + 651 + ], + "type": "text", + "content": " GeForce RTX 2080 Ti. For training, we use a fixed batch size of eight per GPU for all datasets. We use the AdamW [34] optimizer with a weight decay factor of 0.01 and a base learning rate of 5e-4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "Evaluation metrics: We evaluate the Mean Absolute Error (MAE) for all the pixels in the generated 2-D range images. Additionally, we assess the performance by considering the 3D points reconstructed by our neural networks. The Chamfer Distance (CD) measures the Euclidean distance between" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15358" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 115, + 118 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 115, + 118 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 115, + 118 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 115, + 118 + ], + "type": "image", + "image_path": "5d4a9e49b8b088b9500b3b231376ef94ea57b36bf3533574cf0ffbdb29f35ea7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 120, + 97, + 129 + ], + "lines": [ + { + "bbox": [ + 66, + 120, + 97, + 129 + ], + "spans": [ + { + "bbox": [ + 66, + 120, + 97, + 129 + ], + "type": "text", + "content": "(a) Scene" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 115, + 70, + 157, + 118 + ], + "blocks": [ + { + "bbox": [ + 115, + 70, + 157, + 118 + ], + "lines": [ + { + "bbox": [ + 115, + 70, + 157, + 118 + ], + "spans": [ + { + "bbox": [ + 115, + 70, + 157, + 118 + ], + "type": "image", + "image_path": "cc353321f5773f05891d97252c6797477d1397fbe84ae08db5338efefc3e2e2d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 120, + 148, + 129 + ], + "lines": [ + { + "bbox": [ + 125, + 120, + 148, + 129 + ], + "spans": [ + { + "bbox": [ + 125, + 120, + 148, + 129 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 159, + 70, + 199, + 118 + ], + "blocks": [ + { + "bbox": [ + 159, + 70, + 199, + 118 + ], + "lines": [ + { + "bbox": [ + 159, + 70, + 199, + 118 + ], + "spans": [ + { + "bbox": [ + 159, + 70, + 199, + 118 + ], + "type": "image", + "image_path": "1ea230d4c9029e503049eacac46d1cb3a88121c05998bf8245211e1e3f1ded68.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 120, + 194, + 129 + ], + "lines": [ + { + "bbox": [ + 163, + 120, + 194, + 129 + ], + "spans": [ + { + "bbox": [ + 163, + 120, + 194, + 129 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 163, + 120, + 194, + 129 + ], + "type": "inline_equation", + "content": "1 \\times 4" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 200, + 70, + 241, + 118 + ], + "blocks": [ + { + "bbox": [ + 200, + 70, + 241, + 118 + ], + "lines": [ + { + "bbox": [ + 200, + 70, + 241, + 118 + ], + "spans": [ + { + "bbox": [ + 200, + 70, + 241, + 118 + ], + "type": "image", + "image_path": "2b7fb6383d0634125ee16e7befb8e254fbef6b9c57b6d649deb283c5c0736bb6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 120, + 236, + 129 + ], + "lines": [ + { + "bbox": [ + 205, + 120, + 236, + 129 + ], + "spans": [ + { + "bbox": [ + 205, + 120, + 236, + 129 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 205, + 120, + 236, + 129 + ], + "type": "inline_equation", + "content": "2 \\times 2" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 243, + 70, + 285, + 118 + ], + "blocks": [ + { + "bbox": [ + 243, + 70, + 285, + 118 + ], + "lines": [ + { + "bbox": [ + 243, + 70, + 285, + 118 + ], + "spans": [ + { + "bbox": [ + 243, + 70, + 285, + 118 + ], + "type": "image", + "image_path": "0fe2aab2c5160df7eff0382e24a75b46a63e6c6d1e6cc172fd70aef00196d387.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 120, + 278, + 129 + ], + "lines": [ + { + "bbox": [ + 247, + 120, + 278, + 129 + ], + "spans": [ + { + "bbox": [ + 247, + 120, + 278, + 129 + ], + "type": "text", + "content": "(e) " + }, + { + "bbox": [ + 247, + 120, + 278, + 129 + ], + "type": "inline_equation", + "content": "4 \\times 4" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 50, + 177, + 126, + 236 + ], + "blocks": [ + { + "bbox": [ + 47, + 133, + 287, + 177 + ], + "lines": [ + { + "bbox": [ + 47, + 133, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 133, + 287, + 177 + ], + "type": "text", + "content": "Figure 3. Upsampling results with different patch sizes: We observe that square patches " + }, + { + "bbox": [ + 47, + 133, + 287, + 177 + ], + "type": "inline_equation", + "content": "(4 \\times 4, 2 \\times 2)" + }, + { + "bbox": [ + 47, + 133, + 287, + 177 + ], + "type": "text", + "content": " blur out edges and create invalid connections between separate walls, while the proposed " + }, + { + "bbox": [ + 47, + 133, + 287, + 177 + ], + "type": "inline_equation", + "content": "1 \\times 4" + }, + { + "bbox": [ + 47, + 133, + 287, + 177 + ], + "type": "text", + "content": " patch generates sharp edges and correct discontinuities." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 177, + 126, + 236 + ], + "lines": [ + { + "bbox": [ + 50, + 177, + 126, + 236 + ], + "spans": [ + { + "bbox": [ + 50, + 177, + 126, + 236 + ], + "type": "image", + "image_path": "988a09b5144fc9df356a255f8d2e51f8348f0fb0fa6b554d4113efefb82ea6cd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 237, + 103, + 246 + ], + "lines": [ + { + "bbox": [ + 72, + 237, + 103, + 246 + ], + "spans": [ + { + "bbox": [ + 72, + 237, + 103, + 246 + ], + "type": "text", + "content": "(a) Scene" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 248, + 287, + 292 + ], + "lines": [ + { + "bbox": [ + 47, + 248, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 287, + 292 + ], + "type": "text", + "content": "Figure 4. Upsampling results with different window geometries: Rectangular windows improve reconstruction of objects at different ranges, indicated by more distinct discontinuities (top) and a better reconstruction of the car (bottom)." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 126, + 177, + 178, + 235 + ], + "blocks": [ + { + "bbox": [ + 126, + 177, + 178, + 235 + ], + "lines": [ + { + "bbox": [ + 126, + 177, + 178, + 235 + ], + "spans": [ + { + "bbox": [ + 126, + 177, + 178, + 235 + ], + "type": "image", + "image_path": "fb9a8f36523583b54143295e7e202e7b6362f22d1ed80c285b642be144e5b77c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 237, + 164, + 246 + ], + "lines": [ + { + "bbox": [ + 141, + 237, + 164, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 237, + 164, + 246 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 179, + 177, + 231, + 235 + ], + "blocks": [ + { + "bbox": [ + 179, + 177, + 231, + 235 + ], + "lines": [ + { + "bbox": [ + 179, + 177, + 231, + 235 + ], + "spans": [ + { + "bbox": [ + 179, + 177, + 231, + 235 + ], + "type": "image", + "image_path": "79e3d4da210bbf932a45af88f201e2d2fdcba66f9a215d34e1a07e1873a40d50.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 231, + 177, + 283, + 236 + ], + "blocks": [ + { + "bbox": [ + 179, + 237, + 230, + 247 + ], + "lines": [ + { + "bbox": [ + 179, + 237, + 230, + 247 + ], + "spans": [ + { + "bbox": [ + 179, + 237, + 230, + 247 + ], + "type": "text", + "content": "(c) Rectangular" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 231, + 177, + 283, + 236 + ], + "lines": [ + { + "bbox": [ + 231, + 177, + 283, + 236 + ], + "spans": [ + { + "bbox": [ + 231, + 177, + 283, + 236 + ], + "type": "image", + "image_path": "6108bc1888daaeeaca85b93a02fc66c6288af647a327371b8a530e1aab5c9fa9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 237, + 274, + 247 + ], + "lines": [ + { + "bbox": [ + 239, + 237, + 274, + 247 + ], + "spans": [ + { + "bbox": [ + 239, + 237, + 274, + 247 + ], + "type": "text", + "content": "(d) Square" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 298, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 287, + 369 + ], + "type": "text", + "content": "two point clouds. We also follow the approach in [26] to evaluate the volumetric occupancy similarity. To do so, the point clouds are voxelized using a voxel size of " + }, + { + "bbox": [ + 46, + 298, + 287, + 369 + ], + "type": "inline_equation", + "content": "0.1m" + }, + { + "bbox": [ + 46, + 298, + 287, + 369 + ], + "type": "text", + "content": ". A voxel is classified as occupied for each point cloud if it contains at least one point. We then calculate the Intersection-over-Unit (IoU) based on the occupancy." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 380, + 147, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 147, + 391 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 147, + 391 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "text", + "content": "Patch size: In Tab. 1 Baseline and Model 1-2, we compare the upsampling results with different patch sizes. We observe that the smaller patch " + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "inline_equation", + "content": "(2 \\times 2)" + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "text", + "content": " outperforms the larger one " + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "inline_equation", + "content": "(4 \\times 4)" + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "text", + "content": " while the proposed row patch " + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "inline_equation", + "content": "(1 \\times 4)" + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "text", + "content": " can further refine the network performance significantly. In particular, the row patch improves " + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "inline_equation", + "content": "32.6\\%" + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "text", + "content": " in MAE, " + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "inline_equation", + "content": "12.8\\%" + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "text", + "content": " in IoU, and " + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "inline_equation", + "content": "28.3\\%" + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "text", + "content": " in CD, compared to a " + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "text", + "content": " patch. As visualized in Fig. 3, the row patch better preserves sharpness around edges and corners compared to the larger patch sizes." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 511, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 287, + 559 + ], + "type": "text", + "content": "Window geometry: To verify the effectiveness of modifying window geometry for range images, we conduct experiments with square and rectangular windows (Tab. 1 Model 2-3. Based on the quantitative results, we can observe the" + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 48, + 573, + 288, + 661 + ], + "blocks": [ + { + "bbox": [ + 48, + 573, + 288, + 661 + ], + "lines": [ + { + "bbox": [ + 48, + 573, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 48, + 573, + 288, + 661 + ], + "type": "table", + "html": "
ModelPatchWindowCP/PU/PSMCMAE ↓IoU ↑CD ↓
Baseline [5]4 × 4Square××0.71380.32500.1940
Model 12 × 2Square××0.62510.33370.1661
Model 21 × 4Square××0.48140.36670.1391
Model 31 × 4Rectangular××0.42480.40400.1218
Model 41 × 4Rectangular✓×0.42270.40840.1207
TULIP1 × 4Rectangular✓✓0.41850.41740.1207
TULIP-L1 × 4Rectangular✓✓0.37080.43290.0992
ViT-Unet*1 × 4-✓✓1.44840.29480.2665
", + "image_path": "a925d8d3961ef23ff474f84556e73bffc17d3c710eff3366fc7aa9c6d1bf3f92.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 670, + 287, + 715 + ], + "lines": [ + { + "bbox": [ + 46, + 670, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 670, + 287, + 715 + ], + "type": "text", + "content": "Table 1. Ablation study results for " + }, + { + "bbox": [ + 46, + 670, + 287, + 715 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 46, + 670, + 287, + 715 + ], + "type": "text", + "content": " upsampling on KITTI. Input Resolution: " + }, + { + "bbox": [ + 46, + 670, + 287, + 715 + ], + "type": "inline_equation", + "content": "16\\times 1024" + }, + { + "bbox": [ + 46, + 670, + 287, + 715 + ], + "type": "text", + "content": ", Output Resolution: " + }, + { + "bbox": [ + 46, + 670, + 287, + 715 + ], + "type": "inline_equation", + "content": "64\\times 1024" + }, + { + "bbox": [ + 46, + 670, + 287, + 715 + ], + "type": "text", + "content": ". *We replaced Swin Transformer blocks with ViT blocks and followed the same training procedure as for TULIP." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "following: reconstruction quality produced by local attention within large square windows is inferior to the rectangular window concerning both 2D and 3D evaluation metrics, while they result in the same computational complexity. The underlying reason for this performance difference lies in the perceptive field of the windows. A wider window is more likely to include discontinuities between objects at different scales. Accordingly, the network performs better at separating objects from each other, which can visually be validated by clearer surface boundaries in Fig. 4." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 194, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 384 + ], + "type": "text", + "content": "Further Adaptation and Refinement: Circular Padding (CP) helps maintain continuity and consistency at the edges of the panoramic images, but it has minor contributions to other areas. Patch Unmerging (PU), and Pixel Shuffle (PS) aim at upscaling the spatial resolution with rearrangement, avoiding information loss and parameter increase led by using additional de-convolutional layers. Although they are designed to improve efficiency rather than efficacy of the model, tested with those components (Model 4), a slight improvement in upsampling is still observable. Monte Carlo Dropout [19], as a post-processing step, helps to further reduce ghost points in between objects and we present more details in the Appendix. For TULIP-L, we increase the model capacity, using four Swin-Transformer layers instead of three in both encoder and decoder, which demonstrates an additional performance boost on all metrics." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 388, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 579 + ], + "type": "text", + "content": "Transformer Block: From the previous discussion, we infer that Swin Transformer [33] is more advantageous than ViT (Vision Transformer) [17] in LiDAR upsampling. On one hand, prior works [13, 20, 27, 32] have pointed out the inferiority of ViT on smaller datasets due to lack of locality learning and non-overlapping attention. Relevant LiDAR datasets [21, 28] are generally small compared to RGB dataset [41]. On the other hand, by narrowing the patch size, the exponential increase of model capacity hinders sufficient training with ViT on range images, while Swin Transformer expands only linearly. To numerically show the distinction, we trained the network on the same dataset by replacing ViT blocks and applied the same patch size and additional components to the network for a fair comparison. In the last row of Tab. 1, it shows that TULIP significantly outperforms the one with ViT as the backbone." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 587, + 420, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 420, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 420, + 598 + ], + "type": "text", + "content": "4.3. Benchmark Results" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Besides bilinear interpolation, which computes the interpolation weights from the four nearest neighbors, we evaluate against the state-of-the-art LiDAR upsampling approaches Implicit LiDAR Network (ILN) [26] and LiDAR-SR [43]. We also compare against several image super-resolution works: SRNO [51], and LIIF [10] learn implicit features for interpolation instead of 2D coordinates of new pixels while HAT [9] and Swin-IR [29] are pixel-based approaches that directly reconstruct high-resolution images from low" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15359" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 287, + 217 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 217 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 217 + ], + "type": "table", + "html": "
ModelMAE ↓IoU ↑CD ↓
CARLA 4x Output Resolution: 128 × 2048
Bilinear1.81280.13820.7262
SRNO [51]2.46400.13432.0230
HAT [9]1.60320.26980.6337
SWIN-IR [29]1.95600.27180.4840
LIIF [10]0.80640.35020.174
LIDAR-SR [43]0.82160.25810.2044
ILN [26]0.85920.50060.1855
TULIP (Ours)0.76990.51520.1028
TULIP-L (Ours)0.75390.53010.1001
", + "image_path": "73ffecc718011df62629a4fef1e3a3186033233b715d03efd24c21b1c8fe3940.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 218, + 287, + 351 + ], + "blocks": [ + { + "bbox": [ + 47, + 218, + 287, + 351 + ], + "lines": [ + { + "bbox": [ + 47, + 218, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 47, + 218, + 287, + 351 + ], + "type": "table", + "html": "
KITTI 4x Output Resolution: 64 × 1024
Bilinear2.08920.10630.6000
SRNO [51]0.83500.20350.4417
HAT [9]0.68560.20350.2516
SWIN-IR [29]1.29720.27740.7347
LIIF [10]0.61430.32260.1916
LIDAR-SR [43]0.56740.10050.2165
ILN [26]1.05280.33420.2787
TULIP (Ours)0.41850.41740.1207
TULIP-L (Ours)0.37080.43290.0992
", + "image_path": "c4a4647517818cf9661ca8eef2a4c8cc6b00453556fc47f5ca0aa10b75479bef.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 47, + 351, + 287, + 483 + ], + "blocks": [ + { + "bbox": [ + 47, + 351, + 287, + 483 + ], + "lines": [ + { + "bbox": [ + 47, + 351, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 351, + 287, + 483 + ], + "type": "table", + "html": "
DurLAR 4x Output Resolution: 128 × 2048
Bilinear2.43840.12660.6346
SRNO [51]1.53960.15070.5108
HAT [9]1.78200.23530.1973
SWIN-IR [29]1.94160.21570.2279
LIIF [10]1.56720.24690.1548
LIDAR-SR [43]1.53120.13700.1128
ILN [26]1.57200.34300.0893
TULIP (Ours)1.54320.35620.06484
TULIP-L (Ours)1.55920.36540.06346
", + "image_path": "b1e5ad937115afebcf7e81d4958d105d3df43048a526dfaf39e1cd250a466617.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 484, + 287, + 518 + ], + "lines": [ + { + "bbox": [ + 47, + 484, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 287, + 518 + ], + "type": "text", + "content": "Table 2. Quantitative comparison against state-of-the-art LiDAR and image super-resolution methods on different datasets. All methods are trained and evaluated on the same splits." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 529, + 287, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 565 + ], + "type": "text", + "content": "resolution feature embeddings. Similar to our work, they are based on Swin Transformer [33]. We train all methods on the specific datasets using their default parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 585, + 161, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 161, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 161, + 596 + ], + "type": "text", + "content": "4.3.1 Simulation Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Similar to prior works, we first evaluate the performance on noise-free, simulated data. We train and test all approaches on the dataset presented in [26] that was recorded using the CARLA simulator and select the ground-truth range images of size " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "128 \\times 2048" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " that capture a vertical FoV of " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": ". We use a (20699/2618) train/test split. We observe that the performance between image super-resolution approaches varies greatly, and LIIF performs the best among them. While LIIF [10] achieves a comparable MAE to the LiDAR up-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": "sampling methods, the 3D metrics are significantly worse, indicating that it is unsuitable for the underlying geometric input data. Nevertheless, LIIF [10] still outperforms the Li-DAR upsampling method in LIDAR-SR. The interpolation-based approach in ILN shows decent results for the noise-free data, as indicated by their high IoU. Nevertheless, our approaches achieve the best performance in all evaluated metrics, with the strongest improvement in Chamfer Distance, which validates that our design is well suited for the geometric reconstruction in range images." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 304, + 213, + 554, + 555 + ], + "blocks": [ + { + "bbox": [ + 304, + 213, + 554, + 555 + ], + "lines": [ + { + "bbox": [ + 304, + 213, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 554, + 555 + ], + "type": "image", + "image_path": "df7742b7922515108ac29509113717de52047ce622fdc8a44e8bfea0de0502df.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 562, + 547, + 727 + ], + "lines": [ + { + "bbox": [ + 304, + 562, + 547, + 727 + ], + "spans": [ + { + "bbox": [ + 304, + 562, + 547, + 727 + ], + "type": "text", + "content": "Figure 5. Qualitative results on KITTI: We provide visualizations for the top-4 results from Tab. 2 and Swin-IR [29]. Our approach outperforms other state-of-the-art methods in upsampling realistic point clouds. A: TULIP is the only approach that does not generate invalid ghost points in between the car (right bottom) and the wall behind. It can also be observed that all approaches except for ILN match the characteristic line pattern of the LiDAR on the ground (middle). B: Swin-IR and LIIF generate a blurry reconstruction. TULIP achieves the clearest point cloud and even reconstructs small details such as the car's side mirrors. C: Swin-IR, LIIF, and LiDAR-SR do not clearly separate the car from the wall behind. ILN generates a clear point cloud but does not properly reconstruct the shape of the cars. TULIP achieves the clearest point cloud, with distinct discontinuities between objects, and closely resembles the shape of the three cars." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15360" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 165, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 165, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 165, + 83 + ], + "type": "text", + "content": "4.3.2 Real-World Results" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 93, + 287, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 93, + 287, + 165 + ], + "spans": [ + { + "bbox": [ + 46, + 93, + 287, + 165 + ], + "type": "text", + "content": "To evaluate our approach on real-world data, we downsample the DurLAR and KITTI datasets to a comparable number of frames to the CARLA dataset by temporally skipping frames in the dataset sequences. In particular, we have the following (train/test) split: DurLAR (24372/2508), and KITTI (20000/2500)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 167, + 288, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 288, + 407 + ], + "type": "text", + "content": "KITTI: The KITTI dataset [21] was collected using a Velodyne HDL-64E LiDAR with a vertical FoV of " + }, + { + "bbox": [ + 46, + 167, + 288, + 407 + ], + "type": "inline_equation", + "content": "26.8^{\\circ}" + }, + { + "bbox": [ + 46, + 167, + 288, + 407 + ], + "type": "text", + "content": " and a resolution of " + }, + { + "bbox": [ + 46, + 167, + 288, + 407 + ], + "type": "inline_equation", + "content": "64 \\times 1024" + }, + { + "bbox": [ + 46, + 167, + 288, + 407 + ], + "type": "text", + "content": ". For the image super-resolution approaches, we observe comparable results and trends on KITTI as on the simulated data in Sec.4.3.1. Even though LIDAR-SR [43] is designed for LIDAR upsampling and outperforms image super-resolution approaches in terms of MAE, it achieves the lowest IoU among all approaches, which can be explained by a large amount of invalid floating points between real objects. LIIF [10] achieves better CD and comparable IoU compared to ILN. We observe that the IoU of ILN [26] is drastically lower than on CARLA. The interpolation scheme in ILN is thus not suitable to properly handle the lower resolution data with noise from the real-world sensor. Our approaches significantly outperform all competing methods, especially in terms of IoU as it can better handle discontinuities between objects. As indicated in the visualizations in Fig.5, our approach predicts fewer invalid float points between objects, as well as fewer blurry edges, which results in much clearer point clouds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": "DurLAR: The DurLAR dataset [28] was recorded using an Ouster OS1-128 LiDAR, that captures point clouds with a resolution of " + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "inline_equation", + "content": "128 \\times 2048" + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": " with a " + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": " vertical FoV. Compared to CARLA and KITTI, we see a large performance drop for all approaches in MAE and IoU, which results from the larger maximum range and vertical FoV of the sensor as well as stronger sensor noise. We observe that all approaches designed for LiDAR upsampling clearly improve in Chamfer Distance compared to image super-resolution approaches. ILN and TULIP outperform all methods by a large margin in IoU. TULIP-L achieves the best 3D metrics, which validates that it can also handle the challenging DurLAR data better than prior works." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Range Analysis: As the density of LiDAR point clouds decreases with increasing distance from the sensor, we additionally compare the performance at different ranges for a more fine-grained evaluation. We observe that our approach achieves better or similar performance compared to all approaches in all metrics at all ranges. Up to " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "30\\mathrm{m}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ", TULIP clearly outperforms the baseline approaches. TULIP handles discontinuities between objects better than other approaches, which can be verified visually through sharper edges and fewer \"floating\" points between objects. This is especially reflected in the high IoU. At ranges above " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "30\\mathrm{m}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ", our approach performs comparable to other methods." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 425, + 141 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 425, + 141 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 425, + 141 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 425, + 141 + ], + "type": "image", + "image_path": "f679bc60bb378f808e359d1abf2ee2a83ac8901059ffd4dbbfca9d1bec73be0d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 145, + 382, + 155 + ], + "lines": [ + { + "bbox": [ + 351, + 145, + 382, + 155 + ], + "spans": [ + { + "bbox": [ + 351, + 145, + 382, + 155 + ], + "type": "text", + "content": "(a) IoU↑" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 425, + 70, + 545, + 141 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 545, + 141 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 545, + 141 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 545, + 141 + ], + "type": "image", + "image_path": "2c3124ec2de544d4df5d2aac0fef9a8d29fd46cad27975674c94019b6e819a8d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 145, + 523, + 155 + ], + "lines": [ + { + "bbox": [ + 447, + 145, + 523, + 155 + ], + "spans": [ + { + "bbox": [ + 447, + 145, + 523, + 155 + ], + "type": "text", + "content": "(b) Chamfer distance " + }, + { + "bbox": [ + 447, + 145, + 523, + 155 + ], + "type": "inline_equation", + "content": "\\downarrow" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 162, + 545, + 174 + ], + "lines": [ + { + "bbox": [ + 305, + 162, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 305, + 162, + 545, + 174 + ], + "type": "text", + "content": "Figure 6. 3D error metrics visualized at different ranges for KITTI." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 178, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 178, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 545, + 251 + ], + "type": "text", + "content": "Above this range, LiDAR data is extremely sparse, meaning that points neighboring in the low-resolution range image are actually distant in the Euclidean space. It is, therefore, not feasible to infer meaningful information about the scene from such neighboring points at a high range, which is also represented in the poor performance of all approaches." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 258, + 391, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 258, + 391, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 391, + 270 + ], + "type": "text", + "content": "4.4. Failure Cases" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 277, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 373 + ], + "type": "text", + "content": "Although TULIP outperforms other state-of-the-art methods both qualitatively and quantitatively, the upsampling quality can still be limited in certain cases. For instance, as shown in Fig 7, our approach shows inferior upsampling results for the specific scene. The irregularity of the scene leads to high uncertainty, which makes the reconstruction noisy. Observing the details in Fig. 7a, the network fails to reconstruct the car at the top of the scene." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 329, + 388, + 435, + 466 + ], + "blocks": [ + { + "bbox": [ + 329, + 388, + 435, + 466 + ], + "lines": [ + { + "bbox": [ + 329, + 388, + 435, + 466 + ], + "spans": [ + { + "bbox": [ + 329, + 388, + 435, + 466 + ], + "type": "image", + "image_path": "fc987d03c9b37505c3ea807719857ad4aa7a209ce1045ff571c775fc2dd7b00a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 468, + 402, + 477 + ], + "lines": [ + { + "bbox": [ + 357, + 468, + 402, + 477 + ], + "spans": [ + { + "bbox": [ + 357, + 468, + 402, + 477 + ], + "type": "text", + "content": "(a) Full scene" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 486, + 545, + 510 + ], + "lines": [ + { + "bbox": [ + 305, + 486, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 486, + 545, + 510 + ], + "type": "text", + "content": "Figure 7. Failure Case: noisy point cloud generated by TULIP and missing reconstruction of the object. (Ours and GT)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 451, + 399, + 528, + 465 + ], + "blocks": [ + { + "bbox": [ + 451, + 399, + 528, + 465 + ], + "lines": [ + { + "bbox": [ + 451, + 399, + 528, + 465 + ], + "spans": [ + { + "bbox": [ + 451, + 399, + 528, + 465 + ], + "type": "image", + "image_path": "07a3bc037bda7aa20ddd231504b007ef05d3b7ece19155f3ba759c7beacf3949.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 468, + 507, + 476 + ], + "lines": [ + { + "bbox": [ + 473, + 468, + 507, + 476 + ], + "spans": [ + { + "bbox": [ + 473, + 468, + 507, + 476 + ], + "type": "text", + "content": "(b) Details" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 525, + 378, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 378, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 378, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 545, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 712 + ], + "type": "text", + "content": "This work presents TULIP, a novel method for LiDAR Upsampling that achieves incredible performance in upsampling the range image. Our approach transforms 3D point clouds into 2D range images and performs the upsampling in 2D space. We build upon a Swin-Transformer-based network and specifically modify the patch partition and attention windows to better accommodate the characteristics of range images. Throughout various experiments, testing on three different benchmarks, it shows that TULIP outperforms state-of-the-art methods quantitatively in all evaluation metrics, and quantitatively, TULIP demonstrates more promising results in upsampling the realistic LiDAR data. Acknowledgement: This work was supported by Swiss National Science Foundation's NCCR DFab P3." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15361" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Anique Akhtar, Zhu Li, Geert Van der Auwera, Li Li, and Jianle Chen. Pu-dense: Sparse tensor-based point cloud geometry upsampling. IEEE Transactions on Image Processing, 31:4133-4148, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "text", + "content": "[2] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. ECCVW What is Motion For?, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 214 + ], + "type": "text", + "content": "[3] Borna Bešić, Nikhil Gosala, Daniele Cattaneo, and Abhinav Valada. Unsupervised domain adaptation for lidar panoptic segmentation. IEEE Robotics and Automation Letters, 7(2): 3404-3411, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 271 + ], + "type": "text", + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 288, + 327 + ], + "type": "text", + "content": "[5] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 3, 4, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 329, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 288, + 373 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 288, + 430 + ], + "type": "text", + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 431, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 288, + 474 + ], + "type": "text", + "content": "[8] Tso Yuan Chen, Ching Chun Hsiao, and Ching-Chun Huang. Density-imbalance-eased lidar point cloud upsampling via feature consistency learning. IEEE Transactions on Intelligent Vehicles, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 476, + 288, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 288, + 531 + ], + "type": "text", + "content": "[9] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367-22377, 2023. 2, 4, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 586 + ], + "type": "text", + "content": "[10] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 589, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 288, + 643 + ], + "type": "text", + "content": "[11] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12312-12321, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[12] Xinjing Cheng, Peng Wang, and Ruigang Yang. Depth estimation via affinity learned with convolutional spatial propagation network. In Proceedings of the European conference on computer vision (ECCV), pages 103-119, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[13] Yin Dai, Fayu Liu, Weibing Chen, Yue Liu, Lifu Shi, Sheng Liu, Yuhang Zhou, et al. Swin mae: Masked autoencoders" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "for small datasets. Computers in Biology and Medicine, 161: 107037, 2023. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[14] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "text", + "content": "[15] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 391-407. Springer, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 198, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 240 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 243, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 308 + ], + "type": "text", + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "type": "text", + "content": "[18] George Eskandar, Sanjeev Sudarsan, Karim Guirguis, Janaranjani Palaniswamy, Bharath Somashekar, and Bin Yang. Hals: A height-aware lidar super-resolution framework for autonomous driving. arXiv preprint arXiv:2202.03901, 2022. 1, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 365, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 408 + ], + "type": "text", + "content": "[19] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pages 1050-1059. PMLR, 2016. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 411, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 443 + ], + "type": "text", + "content": "[20] Hanan Gani, Muzammal Naseer, and Mohammad Yaqub. How to train vision transformer on small-scale datasets? In 33rd British Machine Vision Conference, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "type": "text", + "content": "[21] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 2, 5, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 490, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 543 + ], + "type": "text", + "content": "[22] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Penet: Towards precise and efficient image guided depth completion. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13656-13662. IEEE, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "type": "text", + "content": "[23] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "text", + "content": "[24] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "text", + "content": "[25] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1637-1645, 2016. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[26] Youngsun Kwon, Minhyuk Sung, and Sung-Eui Yoon. Implicit lidar network: Lidar super-resolution via interpolation weight prediction. In 2022 International Conference on" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15362" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "Robotics and Automation (ICRA), pages 8424-8430. IEEE, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "text", + "content": "[27] Seung Hoon Lee, Seunghyun Lee, and Byung Cheol Song. Vision transformer for small-size datasets. arXiv preprint arXiv:2112.13492, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "text", + "content": "[28] L. Li, K.N. Ismail, H.P.H. Shum, and T.P. Breckon. Durlar: A high-fidelity 128-channel lidar dataset with panoramic ambient and reflectivity imagery for multi-modal autonomous driving applications. In Proc. Int. Conf. on 3D Vision. IEEE, 2021. 2, 5, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 239 + ], + "type": "text", + "content": "[29] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "type": "text", + "content": "[30] Yuankai Lin, Tao Cheng, Qi Zhong, Wending Zhou, and Hua Yang. Dynamic spatial propagation network for depth completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1638-1646, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 287, + 338 + ], + "type": "text", + "content": "[31] Zhixin Ling, Zhen Xing, Xiangdong Zhou, Manliang Cao, and Guichun Zhou. Panoswin: A pano-style swin transformer for panorama understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2023. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "type": "text", + "content": "[32] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco De Nadai. Efficient training of visual transformers with small datasets. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 437 + ], + "type": "text", + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "type": "text", + "content": "[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "text", + "content": "[35] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 457-466, 2022. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 581 + ], + "type": "text", + "content": "[36] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 120-136. Springer, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 582, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 647 + ], + "type": "text", + "content": "[37] Jiaxiong Qiu, Zhaopeng Cui, Yinda Zhang, Xingdi Zhang, Shuaicheng Liu, Bing Zeng, and Marc Pollefeys. Deepidar: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3313-3322, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[38] Shi Qiu, Saeed Anwar, and Nick Barnes. Pu-transformer: Point cloud upsampling transformer. In Proceedings of the Asian Conference on Computer Vision, pages 2475-2493, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "text", + "content": "[39] Kyeongha Rho, Jinsung Ha, and Youngjung Kim. Guideformer: Transformers for image guided depth completion." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6250-6259, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 163 + ], + "type": "text", + "content": "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 547, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 218 + ], + "type": "text", + "content": "[41] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "text", + "content": "[42] Artem Savkin, Yida Wang, Sebastian Wirkert, Nassir Navab, and Federico Tombari. Lidar upsampling with sliced wasserstein distance. IEEE Robotics and Automation Letters, 8(1): 392-399, 2022. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[43] Tixiao Shan, Jinkun Wang, Fanfei Chen, Paul Szenher, and Brendan Englot. Simulation-based lidar super-resolution for ground vehicles. Robotics and Autonomous Systems, 134: 103647, 2020. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 310, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 376 + ], + "type": "text", + "content": "[44] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 377, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 453 + ], + "type": "text", + "content": "[45] Xiaopeng Sun, Weiqi Li, Zhenyu Zhang, Qiufang Ma, Xuhan Sheng, Ming Cheng, Haoyu Ma, Shijie Zhao, Jian Zhang, Junlin Li, et al. Opdn: Omnidirectional position-aware deformable network for omnidirectional image superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1293-1301, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "type": "text", + "content": "[46] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing, 30:1116-1129, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "type": "text", + "content": "[47] Larissa T Triess, David Peter, Christoph B Rist, Markus Enzweiler, and J Marius Zöllner. Cnn-based synthesis of realistic high-resolution lidar data. In 2019 IEEE Intelligent Vehicles Symposium (IV), pages 1512-1519. IEEE, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 545, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 590 + ], + "type": "text", + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 591, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 591, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 591, + 545, + 656 + ], + "type": "text", + "content": "[49] Suhong Wang, Hongqing Wang, Shufeng She, Yanping Zhang, Qingju Qiu, and Zhifeng Xiao. Swin-t-nfc crfs: An encoder-decoder neural model for high-precision UAV positioning via point cloud super resolution and image semantic segmentation. Computer Communications, 197:52–60, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[50] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15363" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 620 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[51] Min Wei and Xuesong Zhang. Super-resolution neural operator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18247-18256, 2023. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[52] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Baobei Xu, Jun Li, and Jian Yang. Rignet: Repetitive image guided network for depth completion. In European Conference on Computer Vision, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "type": "text", + "content": "[53] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1951-1960, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 208, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 208, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 208, + 287, + 251 + ], + "type": "text", + "content": "[54] Shuquan Ye, Dongdong Chen, Songfang Han, Ziyu Wan, and Jing Liao. Meta-pu: An arbitrary-scale upsampling network for point cloud. IEEE transactions on visualization and computer graphics, 28(9):3206-3218, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 253, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 253, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 49, + 253, + 287, + 307 + ], + "type": "text", + "content": "[55] Fanghua Yu, Xintao Wang, Mingdeng Cao, Gen Li, Ying Shan, and Chao Dong. Osrt: Omnidirectional image superresolution with distortion-aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13283-13292, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 308, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 308, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 49, + 308, + 287, + 352 + ], + "type": "text", + "content": "[56] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-net: Point cloud upsampling network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2790-2799, 2018. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 354, + 286, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 354, + 286, + 385 + ], + "spans": [ + { + "bbox": [ + 49, + 354, + 286, + 385 + ], + "type": "text", + "content": "[57] Heeseung Yun, Sehun Lee, and Gunhee Kim. Panoramic vision transformer for saliency detection in 360 videos. In ECCV, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 387, + 286, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 387, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 387, + 286, + 430 + ], + "type": "text", + "content": "[58] Wentai Zhang, Haoliang Jiang, Zhangsihao Yang, Soji Yamakawa, Kenji Shimada, and Levent Burak Kara. Data-driven upsampling of point clouds. Computer-Aided Design, 112:1-13, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 432, + 286, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 432, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 432, + 286, + 464 + ], + "type": "text", + "content": "[59] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image restoration. TPAMI, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 466, + 287, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 466, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 49, + 466, + 287, + 530 + ], + "type": "text", + "content": "[60] Wenbo Zhao, Xianming Liu, Zhiwei Zhong, Junjun Jiang, Wei Gao, Ge Li, and Xiangyang Ji. Self-supervised arbitrary-scale point clouds upsampling via implicit neural representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1999-2007, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 533, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 533, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 49, + 533, + 287, + 576 + ], + "type": "text", + "content": "[61] Yifan Zhao, Le Hui, and Jin Xie. Sspu-net: Self-supervised point cloud upsampling via differentiable rendering. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2214-2223, 2021. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 578, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 578, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 49, + 578, + 287, + 620 + ], + "type": "text", + "content": "[62] Örkény Zováthi, Balázs Pálffy, Zsolt Jankó, and Csaba Benedek. St-depthnet: A spatio-temporal deep network for depth completion using a single non-repetitive circular scanning lidar. IEEE Robotics and Automation Letters, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15364" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/TUMTraf V2X Cooperative Perception Dataset/5f2a634b-45a2-4d92-81e8-cda538302a8f_content_list.json b/2024/TUMTraf V2X Cooperative Perception Dataset/5f2a634b-45a2-4d92-81e8-cda538302a8f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9f0132f27c5f107458d914358bb33c8e4c2186f4 --- /dev/null +++ b/2024/TUMTraf V2X Cooperative Perception Dataset/5f2a634b-45a2-4d92-81e8-cda538302a8f_content_list.json @@ -0,0 +1,1538 @@ +[ + { + "type": "text", + "text": "TUMTraf V2X Cooperative Perception Dataset", + "text_level": 1, + "bbox": [ + 246, + 130, + 725, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Walter Zimmer1 Gerhard Arya Wardana1 Suren Sritharan1 Xingcheng Zhou1 Rui Song1,2 Alois C. Knoll1", + "bbox": [ + 161, + 186, + 805, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Technical University of Munich $^{2}$ Fraunhofer IVI", + "bbox": [ + 277, + 243, + 691, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://tum-traffic-dataset.github.io/tumtraf-v2x", + "bbox": [ + 189, + 280, + 772, + 297 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b6e4b17926ba46f919f4c3e6240cd0f154fa469192d50f9b40b3bb990aa8f6b0.jpg", + "image_caption": [ + "Figure 1. Visualization of 3D box labels and tracks in our TUMTraf V2X Cooperative Perception Dataset. The top part shows the labels projected into the four camera images. The part below shows a point cloud from two LiDARs with 3D box labels of the same scene." + ], + "image_footnote": [], + "bbox": [ + 89, + 300, + 877, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 549, + 313, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cooperative perception offers several benefits for enhancing the capabilities of autonomous vehicles and improving road safety. Using roadside sensors in addition to onboard sensors increases reliability and extends the sensor range. External sensors offer higher situational awareness for automated vehicles and prevent occlusions. We propose CoopDet3D, a cooperative multi-modal fusion model, and TUMTraf-V2X, a perception dataset, for the cooperative 3D object detection and tracking task. Our dataset contains 2,000 labeled point clouds and 5,000 labeled images from five roadside and four onboard sensors. It includes 30k 3D boxes with track IDs and precise GPS and IMU data. We labeled nine categories and covered occlusion scenarios with challenging driving maneuvers, like traffic violations, near-miss events, overtaking, and U-turns. Through multiple experiments, we show that our CoopDet3D camera-LiDAR fusion model achieves an increase of +14.36 3D mAP compared to a vehicle camera-LiDAR fusion model. Finally, we make our dataset, model, labeling tool, and devkit publicly available on our website.", + "bbox": [ + 73, + 583, + 472, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 549, + 637, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cooperative perception involves the fusion of onboard sensor data and roadside sensor data, and it offers several advantages for enhancing the capabilities of autonomous vehicles and improving road safety. Using data from multiple sources makes the perception more robust to sensor failures or adverse environmental conditions. Roadside sensors provide an elevated view that helps to detect obstacles early. Moreover, they are also beneficial for precise vehicle localization and reduce the computational load of automated vehicles by offloading some perception tasks to the roadside sensors. Roadside sensors provide a global perspective of the traffic and offer a comprehensive situational awareness when fused with onboard sensor data. There are also fewer false positives or negatives because cooperative perception cross-validates the information from different sensors.", + "bbox": [ + 496, + 577, + 890, + 803 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Infrastructure sensors can share perception-related information with vehicles through V2X. Due to minimal delay, and real-time capabilities, the infrastructure-based perception systems can further enhance the situational awareness and decision-making processes of vehicles.", + "bbox": [ + 496, + 806, + 890, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Intelligent Transportation Systems (ITS) like the Testbed", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22668", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/02f1364aace40a772efc1ca2d10d7f92749a31d5b50eaeb9e952309babba30fe.jpg", + "table_caption": [ + "Table 1. Comparison of 3D cooperative V2X perception datasets with our proposed TUMTraf-V2X Cooperative Perception dataset (I=Infrastructure, V=Vehicle)." + ], + "table_footnote": [ + "$\\dagger$ Image dataset has not been released yet.", + "* Value per vehicle. Multiple Conn. and Autom. Vehicles (CAVs) are used." + ], + "table_body": "
DatasetOPV2V [36]V2XSet [35]V2X-Sim [20]V2V4Real [37]DAIR-V2X-C [41]V2X-Seq (SPD) [43]TUMTraf-V2X (Ours)
Year2022202220222022202220232024
V2XV2VV2V&IV2V&IV2VV2IV2IV2I
Real data---✓✓✓✓
Annotation range120 m120 m70 m200 m280 m280 m200 m
Day & night scenes----✓✓✓
# object classes11151098
Track IDs--✓✓-✓✓
HD Maps✓✓✓✓-✓✓
# of sensors (I | V)- | 6*- | 6*5 | 7- | 8‡2 | 32 | 35 | 4
Available worldwide✓✓✓✓--✓
Traffic violations------✓
Labeled attributes§------✓
OpenLABEL format------✓
# Point Clouds11k11k10k20k39k15k2.0k
# Images44k44k60k40k†39k15k5.0k
# 3D Boxes233k233k26k240k464k10.45k29.38k
LocationCARLACARLACARLAUSAChinaChinaGermany
", + "bbox": [ + 101, + 126, + 867, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Total sensors from 2 CAVs.", + "bbox": [ + 493, + 412, + 651, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$^{\\S}$ Weather, time of day, orientation, #3D points, occlusion, color, sub type", + "bbox": [ + 493, + 426, + 880, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for Autonomous Driving [17] aim to improve safety by providing real-time traffic information. According to [9], testbeds extensively start using LiDAR sensors in their setups to create an accurate live digital twin of the traffic. Connected vehicles get a far-reaching view which enables them to react to breakdowns or accidents early. ITS systems also provide lane and speed recommendations to improve the traffic flow.", + "bbox": [ + 75, + 462, + 468, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key challenge with ego-centric vehicle datasets is that there are many occlusions from a vehicle perspective, e.g., if a large truck in front of the ego vehicle obscures the view. Roadside sensors located at a smart intersection provide a broad overview of the intersection and a full-surround view. Given the immense potential of ITS, there is a specific need for V2X datasets. Despite the high costs associated with collecting and labeling such datasets, this work addresses this challenge as a crucial step toward realizing large-scale ITS implementations.", + "bbox": [ + 75, + 583, + 470, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows:", + "text_level": 1, + "bbox": [ + 76, + 734, + 310, + 748 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We provide a high-quality V2X dataset for the cooperative 3D object detection and tracking task with 2,000 labeled point clouds and 5,000 labeled images. In total, 30k 3D bounding boxes with track IDs were labeled in challenging traffic scenarios like near-miss events, overtaking scenarios, U-turn maneuvers, and traffic violation events.", + "- We open-source our 3D bounding box annotation tool (3DBAT v24.3.2) to label multi-modal V2X datasets.", + "- We propose CoopDet3D, a cooperative 3D object detection model, and show in extensive experiments and abla" + ], + "bbox": [ + 76, + 750, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion studies that it outperforms single view models on our V2X dataset by $+14.3$ 3D mAP.", + "bbox": [ + 509, + 462, + 890, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Finally, we provide a development kit to load the annotations in the widely recognized and standard format OpenLABEL [13], to facilitate a seamless integration and utilization of the dataset. Furthermore, it can preprocess, visualize, and convert labels to and from different dataset formats, and evaluate perception and tracking methods.", + "bbox": [ + 500, + 493, + 890, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 . Related work", + "text_level": 1, + "bbox": [ + 500, + 598, + 643, + 614 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D autonomous driving datasets are mainly categorized based on the viewpoint. Table 1 highlights the main differences between our proposed dataset and other V2X datasets.", + "bbox": [ + 496, + 625, + 893, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Single viewpoint datasets", + "text_level": 1, + "bbox": [ + 500, + 680, + 732, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single viewpoint datasets are obtained from a single point of reference, either an ego-vehicle or roadside infrastructure. Onboard sensor-based datasets like KITTI [12], nuScenes [6], and Waymo [30] contain a diverse set of sensor data collected from a moving vehicle equipped with multiple sensors, including high-resolution cameras, LiDARs, radars, and GPS/INS systems. These datasets are abundant and provide many annotated data, including bounding boxes, track IDs, segmentation masks, and depth maps under different urban driving scenarios.", + "bbox": [ + 496, + 703, + 890, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, roadside sensor-based datasets are in the infancy stage. High-quality multi-modal (camera and LiDAR) datasets are presented in [5, 10, 48], which", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22669", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cbd6b0078709864d3c66c2eb67fc84aca9bfe965f25c700d8998579e28bc9bd6.jpg", + "image_caption": [ + "Figure 2. Our CoopDet3D framework is a multi-modal cooperative fusion system, comprising three distinct fusion pipelines. 1) The roadside camera-LiDAR fusion pipeline fuses three camera images and one LiDAR point cloud by extracting features and transforming them into a BEV representation. 2) The vehicle camera-LiDAR fusion pipeline fuses the vehicle camera feature map with the vehicle point cloud feature map using a convolutional fuser. 3) The vehicle and infrastructure feature maps are then fused by applying an element-wise max-pooling operation (Max Fuser). In the end, we use the TransFusion [1] 3D detection head to obtain 3D bounding box predictions." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 890, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are obtained from Infrastructure Perception Systems (IPS). Similarly, in [40], the authors provide a dataset consisting of only images taken from different viewpoints and under varying traffic conditions. These datasets provide a top-down view of a crowded intersection under different conditions and, as such, can overcome issues such as occlusions created by other vehicles and thereby have a higher number of object labels than onboard sensor-based datasets.", + "bbox": [ + 75, + 436, + 472, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. V2X datasets", + "text_level": 1, + "bbox": [ + 76, + 571, + 218, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "V2X datasets exploit the information from multiple viewpoints to gain additional knowledge regarding the environments. In this way, they overcome the limitations of single viewpoint datasets such as occlusion, limited field of view (FOV), and low point cloud density.", + "bbox": [ + 75, + 595, + 468, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DAIR-V2X dataset family [41] is one of the foremost cooperative multi-modal datasets introduced. It contains three subsets: an intersection, a vehicle, and a cooperative dataset. The cooperative dataset contains 464k 3D box labels belonging to 10 classes, making it one of the largest cooperative datasets. The V2X-Seq dataset [43] extends selected sequences of the DAIR-V2X dataset with track IDs and is partitioned into a sequential perception dataset (SPD) and a trajectory forecasting dataset. Despite these, the lack of specific information, such as the labeling methodology used, the exact models of the sensors deployed, the distribution of the classes, and the scenarios within the dataset, leads to uncertainty in the extendability and application of this dataset in varying conditions.", + "bbox": [ + 75, + 672, + 470, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In V2V4Real [37], the authors propose a multi-modal", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cooperative dataset focusing only on V2V perception. Two vehicles equipped with cameras, LiDAR, and GPS/IMU integration systems are used to collect multi-modal sensor data for diverse scenarios. As opposed to all other V2X datasets, this focuses on V2V perception, and though it is of similar size to other cooperative datasets, it contains fewer classes and 3D bounding box information.", + "bbox": [ + 496, + 436, + 893, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Simulated multi-agent perception datasets have been proposed in [20, 35, 36]. These datasets contain multimodal sensor data (camera and LiDAR) obtained from roadside units (RSUs) and multiple ego vehicles, which enable collaborative perception. They use a combination of simulators such as SUMO [24], CARLA [11], and OpenCDA [36] for flow simulation, data retrieval, and V2X communication. However, the utility of the dataset is still limited due to the simulated nature of the data, and its extendability to real-life applications has not been studied in detail.", + "bbox": [ + 496, + 544, + 893, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. V2X perception models for object detection", + "text_level": 1, + "bbox": [ + 500, + 708, + 870, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The datasets presented above have been used to develop various models for a wide variety of tasks, with the majority focusing on 3D object detection. Different approaches have been taken depending on the availability and challenges, and these methods are grouped based on the number of nodes employed and the modalities used for detection.", + "bbox": [ + 496, + 732, + 890, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most 3D object detection models use multi-modal sensor data obtained from a single point of view, which is often an ego-vehicle. Due to the popularity and abundant availability of onboard datasets [6, 12, 30], most models use images, point clouds, or both modalities. Image-based models were", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22670", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the pioneers in 3D object detection due to their low cost and simplicity, and both vehicular camera-based models [18, 34] and infrastructure camera-based models [39] have been proposed. LiDAR-based 3D object detection models [19, 46, 49] became popular since LiDAR point clouds provide 3D depth information and are robust, especially in adverse weather conditions and nighttime scenarios. Fusion models combine the information obtained from both images and point clouds and have been shown to outperform the prior methods [47]. Single viewpoint fusion models use either vehicular camera and LiDAR [23, 32, 38] or infrastructure camera and LiDAR [47] for 3D object detection.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cooperative perception models, which use data from multiple viewpoints, have been shown to overcome issues related to occlusion, which were often present in vehicular sensor-based models. V2I cooperative perception models [2, 3, 14, 26, 33, 35, 42] use the sensor data from both vehicles and infrastructure and V2V models [15, 29] communicate the sensor data between multiple vehicles. In this work, our cooperative multi-modal dataset is one contribution among others. Thus, while most of the prior works focus on unimodal cooperative perception using either Li-DAR point clouds [2, 7] or camera images [15], we benchmark our dataset with CoopDet3D, a deep fusion based cooperative multi-modal 3D object detection model based on BEVFusion [15] and PillarGrid [2].", + "bbox": [ + 75, + 272, + 472, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. TUMTraf-V2X Dataset", + "text_level": 1, + "bbox": [ + 76, + 496, + 303, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our TUMTraf V2X Cooperative Perception Dataset focuses on challenging traffic scenarios and various day and nighttime scenes. The data is further annotated, emphasizing high-quality labels through careful labeling and high-quality review processes. It also contains dense traffic and fast-moving vehicles, which reveals the specific challenges in cooperative perception, such as pose estimation errors, latency, and synchronization. Furthermore, we provide sensor data from nine different sensors covering the same traffic scenes under diverse weather conditions and lighting variations. The infrastructure sensors are oriented in all four directions of the intersection to get a $360^{\\circ}$ view, which leads to better perception results. Finally, it contains rare events like traffic violations where pedestrians cross the road at a busy four-way intersection while the crossing light is lit red.", + "bbox": [ + 75, + 521, + 472, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Sensor setup", + "text_level": 1, + "bbox": [ + 76, + 757, + 212, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our TUMTraf V2X Cooperative Perception Dataset was recorded on an ITS system with nine sensors.", + "bbox": [ + 75, + 780, + 468, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The infrastructure sensor setup is the following:", + "bbox": [ + 76, + 810, + 393, + 824 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- 1x Ouster LiDAR OS1-64 (gen. 2), 64 vert. layers, $360^{\\circ}$ FOV, below horizon config., $10\\mathrm{cm}$ acc. @ $120\\mathrm{m}$ range", + "- 4x Basler ace acA1920-50gc, $1920 \\times 1200$ , Sony IMX174 with 8 mm lenses" + ], + "bbox": [ + 76, + 825, + 468, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On the vehicle, the following sensors were used:", + "bbox": [ + 76, + 885, + 398, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5315b3b700c301da35141cb641a351776de308830170b77ca6aaa1787ab8ec9f.jpg", + "image_caption": [ + "Figure 3. Demonstration of a possible V2X occlusion scenario. A pedestrian (blue) is crossing the road in front of the ego vehicle. An occluded bicycle is marked in red. The recording vehicle with the sensor setup is shown in the bottom left corner." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 676, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5604719f92ed31b04a8775f70728bc90eb7fb42a5f5a2670f9d35e12a682726a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 88, + 888, + 239 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- 1x Robosense RS-LiDAR-32, 32 vert. layers, $360^{\\circ}$ FOV, 3 cm accuracy @200 m range", + "- 1x Basler ace acA1920-50gc, $1920 \\times 1200$ , Sony IMX174 with $16 \\mathrm{~mm}$ lens", + "- 1x Emlid Reach RS2+ multi-band RTK GNSS receiver", + "- 1x XSENS MTi-30-2A8G4 IMU" + ], + "bbox": [ + 500, + 330, + 890, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Sensor calibration and registration", + "text_level": 1, + "bbox": [ + 500, + 431, + 805, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We synchronize the cameras and LiDARs in the spatial and temporal domain. First, we determine the intrinsic camera parameters and the radial and tangential image distortions by using a checkerboard target. We then calibrate the roadside LiDAR with the roadside cameras by picking 100-point pairs in the point cloud and camera image. Extrinsic parameters (rotation and translation) are calculated by minimizing the reprojection error of 2D-3D point correspondences [25]. We follow the same procedure for onboard camera-LiDAR calibration. Finally, we calibrate the onboard LiDAR to the roadside LiDAR. This spatial registration is done by first estimating a coarse transformation. We pick ten 3D point pairs in each point cloud and minimize their distance using the least squares method. Then, we apply the point-to-point Iterative Closest Point (ICP) algorithm [4] to get the fine transformation between the point clouds.", + "bbox": [ + 496, + 454, + 890, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We label the vehicle and infrastructure point clouds after registering them. The coarse registration was done by measuring the GPS position of the onboard LiDAR and the roadside LiDAR. Then, we transform every 10th onboard point cloud to the coordinate system of the infrastructure point cloud. The fine registration was done by applying the point-to-point ICP to get an accurate V2I transformation matrix. All rotations of the point cloud frames in between are interpolated based on the spherical linear interpolation (SLERP) [28] method:", + "bbox": [ + 496, + 696, + 890, + 847 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS L E R P \\left(q _ {0}, q _ {1}, t\\right) = q _ {0} \\left(q _ {0} ^ {- 1} q _ {1}\\right) ^ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 857, + 890, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $q_{0}$ and $q_{1}$ are the quaternions representing the ro", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22671", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/844b6340366deb9cca125ac434192a9ffb43fee2c45418b3447e1d7f599ac9a3.jpg", + "table_caption": [ + "Table 2. Evaluation results ( $mAP_{BEV}$ and $mAP_{3D}$ ) of CoopDet3D on our TUMTraf-V2X test set in south2 FOV." + ], + "table_footnote": [], + "table_body": "
Config.\\( \\mathbf{mAP_{BEV}}\\uparrow \\)\\( \\mathbf{mAP_{3D}}\\uparrow \\)
DomainModalityEasy↑Mod.↑Hard↑Avg.↑
VehicleCamera46.8331.4737.8230.7730.36
VehicleLiDAR85.3385.2276.8669.0480.11
VehicleCam+LiDAR84.9077.6072.0873.1276.40
Infra.Camera61.9831.1946.7340.4235.04
Infra.LiDAR92.8686.1788.0775.7384.88
Infra.Cam+LiDAR92.9287.9989.0981.6987.01
Coop.Camera68.9445.4142.7657.8345.74
Coop.LiDAR93.9392.6378.0673.9585.86
Coop.Cam+LiDAR94.2293.4288.1779.9490.76
", + "bbox": [ + 76, + 127, + 467, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tations of the start and end frames and $t \\in [0,1]$ . Translation vectors $\\mathbf{T}_0$ and $\\mathbf{T}_1$ were obtained using linear interpolation:", + "bbox": [ + 76, + 290, + 467, + 321 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} (t) = \\mathbf {T} _ {0} + t \\left(\\mathbf {T} _ {1} - \\mathbf {T} _ {0}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 335, + 467, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This dual interpolation strategy ensures that the estimated transformations between the frames are smooth and geometrically accurate, thus adhering closely to the actual movements of the vehicle over time.", + "bbox": [ + 76, + 359, + 467, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Data selection and labeling", + "text_level": 1, + "bbox": [ + 76, + 428, + 321, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We selected the data based on challenging traffic scenarios, like U-turns, tailgate events, and traffic violation maneuvers. Besides the high traffic density of 31 objects per frame, we selected frames with high-class coverage. We selected 700 frames during sunny daytime and 100 frames during cloudy nighttime for labeling. The camera and Li-DAR data were recorded into rosbag files at $15\\mathrm{Hz}$ and $10\\mathrm{Hz}$ , respectively. We extracted and synchronized the data based on ROS [27] timestamps and labeled it with our $3D$ BAT (v24.3.2) annotation tool1. We improved the 3D BAT [45] baseline labeling tool to label 3D objects faster and more precisely with a one-click annotation feature. The annotators were instructed to label traffic participants while examining the images. Objects are still labeled, even if they have no 3D points inside, but are visible in the images. Extremities (e.g., pedestrian limbs) are included in the bounding box, but side mirrors of vehicles aren't. If a pedestrian carries an object, that object is included in the bounding box. If two or more pedestrians are carrying an object, only the box of one will include the object. After labeling, each annotator checked the work of other annotators manually frame-by-frame. When errors were found, the original annotator was notified, and they fixed it. This helps ensure that the labels in our dataset are high quality.", + "bbox": [ + 75, + 452, + 468, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Data structure and format", + "text_level": 1, + "bbox": [ + 76, + 824, + 316, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We record eight different scenes, each 10 sec. long, from vehicle and infrastructure perspectives using nine sensors", + "bbox": [ + 76, + 847, + 467, + 877 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1ad48c9990c6287dd2d933b3b388376109ba81a7b18587a4d6a07459b4ba77fe.jpg", + "table_caption": [ + "Table 3. Evaluation results of infrastructure-only CoopDet3D vs. InfraDet3D [47] on TUMTraf Intersection test set [48]. South 1 and South 2 refer to sensors covering different FOVs." + ], + "table_footnote": [], + "table_body": "
ModelFOVModality\\( \\mathrm{mAP}_{3D} \\uparrow \\)
Easy↑Mod.↑Hard↑Avg.↑
InfraDet3Dsouth 1LiDAR75.8147.6642.1655.21
CoopDet3Dsouth 1LiDAR76.2448.2335.1969.47
InfraDet3Dsouth 2LiDAR38.9246.6043.8643.13
CoopDet3Dsouth 2LiDAR74.9755.5539.9669.94
InfraDet3Dsouth 1Cam+LiDAR67.0831.3835.1744.55
CoopDet3Dsouth 1Cam+LiDAR75.6845.6345.6366.75
InfraDet3Dsouth 2Cam+LiDAR58.3819.7333.0837.06
CoopDet3Dsouth 2Cam+LiDAR74.7353.4641.9666.89
", + "bbox": [ + 500, + 140, + 890, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and split the data into a train (80%), val. (10%), and test (10%) set. We use stratified sampling to distribute all sets' object classes equally (see Fig. 6a). Labels are provided in the ASAM OpenLABEL [13] standard.", + "bbox": [ + 498, + 294, + 890, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Dataset development kit", + "text_level": 1, + "bbox": [ + 500, + 364, + 723, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide a devkit to work with our dataset. In addition to generating the data statistics, it provides modules for multiclass stratified splitting (train/val/test), point cloud registration, loading annotations in OpenLABEL format, evaluation of detection and tracking results, pre-processing steps such as point cloud filtering, and post-processing such as bounding box filtering. The statistics (Fig. 4, 5, and 6) were created using our devkit. It also contains modules to convert the labels from OpenLABEL to KITTI or our custom nuScenes format with timestamps instead of tokens and vice versa. This devkit enables users to migrate their models and make them compatible with our dataset format. We release our devkit2 under the MIT license and the dataset under the Creative Commons (CC) BY-NC-ND 4.0 license.", + "bbox": [ + 498, + 387, + 890, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Benchmark", + "text_level": 1, + "bbox": [ + 500, + 612, + 627, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose CoopDet3D, an extension of BEVFusion [23] and PillarGrid [2] for deep cooperative multi-modal 3D object detection and benchmark it on TUMTraf-V2X mini.", + "bbox": [ + 498, + 638, + 890, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluation metrics", + "text_level": 1, + "bbox": [ + 500, + 693, + 683, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The accuracy is measured in terms of the mean average precision (mAP). Two types of mAP measures are used: BEV mAP considers the BEV center distance, and the results are obtained using the same evaluation methodology used in BEVFusion, which in turn uses the evaluation protocol of nuScenes [6]. Similarly, the 3D mAP measure considers the intersection in 3D, and the results are obtained using the evaluation script of our TUM Traffic dataset devkit. The runtime is evaluated using frames per second (FPS) as the metric and the results were obtained by measuring the time", + "bbox": [ + 498, + 715, + 890, + 867 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1https://github.com/walzimmer/3d-bat", + "bbox": [ + 94, + 886, + 344, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2https://github.com/tum-traffic-dataset/tum-traffic-dataset-dev-kit", + "bbox": [ + 500, + 876, + 888, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22672", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d206a3a08e78e440783e2077d0f0221eee7bbbc5438da77ea290e84e50bcd957.jpg", + "image_caption": [ + "(a) Distribution of objects between day and night." + ], + "image_footnote": [], + "bbox": [ + 84, + 88, + 349, + 250 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/85b64cb2a64f09feb2b1a6bd6d1bd9c48936e20c9660488e11997eb6f363d46e.jpg", + "image_caption": [ + "(b) Avg. and max. num. of 3D points for each class." + ], + "image_footnote": [], + "bbox": [ + 354, + 88, + 620, + 250 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0f371d6d49ce5a390108a7175a9efb49d4fd83275eff48587d85e00fcad5ab47.jpg", + "image_caption": [ + "(c) Avg. and max. track length for all classes." + ], + "image_footnote": [], + "bbox": [ + 625, + 88, + 893, + 250 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/15971227ab788286c7454d3b127c8fa1518b558600cfe37ed1fa3718fd264e43.jpg", + "image_caption": [ + "(a) Visualization of rotations (yaw)." + ], + "image_footnote": [], + "bbox": [ + 84, + 358, + 339, + 542 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3c7804d4733a5c3c0112e6f814c5fee905fa9b156935008a6c3541c55622b450.jpg", + "image_caption": [ + "Figure 4. Our TUMTraf-V2X dataset contains 30k 3D box labels in total and is balanced among nine different object classes. (a) Cars (14,633) and pedestrians (5,276) are highly represented in the dataset. (b) 3D box labels contain, on average 625 points inside which shows the density of the labeled objects. The BUS class has the highest point density. (c) All traffic participants are tracked for $39\\mathrm{m}$ on average. Emergency vehicles have the highest average track length of $128\\mathrm{m}$ , whereas the CAR class contains a max. track length of $174\\mathrm{m}$ .", + "(b) 3D points grouped by distance." + ], + "image_footnote": [], + "bbox": [ + 354, + 356, + 620, + 539 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/380024a0d7512fb60c794f84746c69a477e79068e6d333d946693dc62283c6b5.jpg", + "image_caption": [ + "(c) BEV visualization of tracks." + ], + "image_footnote": [], + "bbox": [ + 630, + 356, + 883, + 542 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/705d9703c938963694243093783f6b3cce1b1911338913a1c662bdbd23c01774.jpg", + "image_caption": [ + "Figure 5. Our dataset was recorded at a crowded intersection with many left and right turns. (a) Most of the vehicles (7,559) are driving in the east direction (0 degrees). (b) 3D boxes were labeled up to $200\\mathrm{m}$ range and are very dense between 10 and $60\\mathrm{m}$ . (c) The visualization of BEV tracks shows where pedestrians and bicycles are crossing the road.", + "(a) Histogram of 3D points." + ], + "image_footnote": [], + "bbox": [ + 84, + 622, + 344, + 806 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f0d78baf01401ab6a2122e2b1a3a6e613b5ea443672373e8254ad13219b9d133.jpg", + "image_caption": [ + "(b) Histogram of 3D box labels." + ], + "image_footnote": [], + "bbox": [ + 357, + 622, + 614, + 806 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/352a55c2997c4ffbacce65d64e73b9216319a1e203e4e1db84322437c40f1269.jpg", + "image_caption": [ + "(c) Histogram of track lengths.", + "Figure 6. (a) Most 3D box labels contain between 1 and 200 3D points inside, with an average of 434 3D points, excluding empty boxes. Objects that were close to both LiDAR sensors even contained up to 18k 3D points. (b) Frames contain between 15 and 45 traffic participants, with an average of 29. (c) Objects were tracked up to $170\\mathrm{m}$ , and the average track length is $39.35\\mathrm{m}$ ." + ], + "image_footnote": [], + "bbox": [ + 620, + 622, + 887, + 808 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22673", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/866fd966ccb2b5e0071274b97dad1103117962b5448090ff4ee0510542ea0a0d.jpg", + "image_caption": [ + "Figure 7. Qualitative results on the TUMTraf-V2X Cooperative Perception test set. The first row shows the inference results of the onboard (vehicle-only) camera-LiDAR fusion with 23 detected objects. In the second row, the results of the cooperative vehicle-infrastructure camera-LiDAR fusion are visualized. Here, all 25 traffic participants could be detected with the support of roadside sensors." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 893, + 289 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/acd460f8b58e7042a21d82314501e603e370e7de61bb737c7a9d3f67141e6066.jpg", + "table_caption": [ + "Table 4. Ablation study on cooperative 3D object detection with 11 combinations of camera and LiDAR backbones. The best tradeoff between speed and accuracy is highlighted in gray. $\\mathrm{ND} =$ non-deterministic, $\\mathrm{TS} =$ TorchSparse, $\\odot =$ retrained. VRAM is in GiB." + ], + "table_footnote": [], + "table_body": "
Backbone Configuration\\( \\mathbf{{mAP}}_{\\mathbf{{BEV}}} \\uparrow \\)\\( \\mathbf{{FPS}} \\uparrow \\)\\( \\mathbf{{VRAM}} \\downarrow \\)
VoxelNet ND + SwinT93.476.306.69
VoxelNet ND + YOLOv8 s92.947.246.39
VoxelNet TS + SwinT93.518.844.61
VoxelNet TS + YOLOv8 s92.9410.664.28
VoxelNet TS + YOLOv8 s ○94.3110.664.28
PointPillars 512 + Swin T94.439.004.94
PointPillars 512 + YOLOv8 s94.2711.144.63
PointPillars 512 + YOLOv8 s ○94.2511.144.63
PointPillars 512_2x + Swin T92.799.064.94
PointPillars 512_2x + YOLOv8 s94.1611.204.63
PointPillars 512_2x + YOLOv8 s ○94.2211.204.63
", + "bbox": [ + 76, + 431, + 465, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "needed by the model to run one full inference, including data preprocessing and voxelization. The first five iterations are skipped as a warmup since they are usually considerably slower than the average. Finally, the complexity of the model is measured in terms of the maximal VRAM usage across all GPUs during training and testing.", + "bbox": [ + 75, + 619, + 468, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. CoopDet3D model", + "text_level": 1, + "bbox": [ + 76, + 724, + 256, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our CoopDet3D uses a BEVFusion-based backbone for camera-LiDAR fusion on the vehicle and the infrastructure sides separately to obtain the vehicle and infrastructure features. The best backbone for image and point cloud feature extraction was chosen through multiple ablation studies. Then, inspired by the method proposed by PillarGrid [2], an element-wise max-pooling operation is proposed to fuse the resulting fused camera-LiDAR features of vehicle and infrastructure together. Finally, the detection head from BEVFusion is used for 3D detection from the fused feature.", + "bbox": [ + 75, + 750, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The architecture of CoopDet3D is shown in Fig. 2.", + "bbox": [ + 498, + 368, + 834, + 382 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, we disable the camera feature extraction nodes and train the LiDAR-only model for 20 epochs. Then, we use pre-trained weights for the cooperative model and fine-tune the entire model for eight further epochs. Hyperparameter tuning revealed that the default hyperparameters of BEV-Fusion [23] gave the best results, and such were not modified. The preprocessing steps are also the same as for BEV-Fusion, but we change the point cloud range to $[-75, 75]$ in the x- and y-scale and $[-8, 0]$ in the z-scale since the dataset used in this case is different. Furthermore, we use 3x NVIDIA RTX 3090 GPUs with 24 GB VRAM for training and a single GPU for evaluation. We open-source our model and provide pre-trained weights3.", + "bbox": [ + 496, + 383, + 890, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Experiments and ablation studies", + "text_level": 1, + "bbox": [ + 498, + 589, + 794, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The objective of these experiments is to highlight the importance of our V2X multi-viewpoint dataset as opposed to single-viewpoint datasets. As such, we conduct multiple experiments and ablation studies with data obtained from each viewpoint and compare the results on the proposed model.", + "bbox": [ + 496, + 612, + 890, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on viewpoint and modality. We conduct multiple experiments with all possible combinations of a) viewpoints: vehicle-only, infrastructure-only, cooperative, and b) modalities: camera-only, LiDAR-only, and camera-LiDAR fusion. Table 2 shows the mAP achieved by CoopDet3D for each of these combinations.", + "bbox": [ + 496, + 696, + 890, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We observe that the results follow a general pattern of cooperative performance being better than infrastructure-only, which is, in turn, better than vehicle-only. Furthermore, fusion models perform better than LiDAR-only models, which in turn are better than camera-only models. Figure 7 shows qualitative results between our vehicle-only", + "bbox": [ + 496, + 786, + 890, + 878 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "3https://github.com/tum-traffic-dataset/coopdet3d", + "bbox": [ + 514, + 886, + 859, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22674", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "camera-LiDAR fusion model and our cooperative vehicle-infrastructure camera-LiDAR fusion model. Again, we observe from these samples that the cooperative perception model is able to detect 25 traffic participants, whereas the vehicle-only model is only able to detect 23 objects due to occlusions and a limited field of view.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on fusion level. We compare our CoopDet3D model to the current SOTA camera-LiDAR fusion method InfraDet3D [47] on the TUMTraf Intersection test set [48]. The proposed method uses deep fusion, whereas the InfraDet3D method is a late fusion method. Table 3 shows the performance of our model against InfraDet3D. The results show that the proposed deep fusion method outperforms the SOTA late fusion model in all metrics, except in the hard difficulty in LiDAR-only mode. Furthermore, Fig. 8 shows two sample images taken during day and nighttime, wherein deep fusion again outperforms late fusion.", + "bbox": [ + 75, + 191, + 468, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We note that these experiments were conducted in an offline setting, disregarding other considerations for simplicity. However, when deploying it in real life, factors such as the transmission bandwidth should also be considered. Since we observed that deep feature fusion generally leads to higher efficacy, the V2I transmissions should contain these features instead of infrastructure bounding boxes.", + "bbox": [ + 75, + 359, + 468, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on backbones. As an ablation study, we present the results of the experiments to find the best backbone and model configuration for the cooperative camera-LiDAR fusion model. For the camera backbone, SwinT [22] and MMYOLO's [8] implementation of YOLOv8 [16] were considered. For the LiDAR backbone, VoxelNet [44] and PointPillars [19] were considered. In addition, VoxelNet was implemented with two different backends, namely SPConv v2 and Torchsparse [31]. For PointPillars, two grid sizes are considered $512 \\times 512$ for both train and test grids (PointPillars 512) and $512 \\times 512$ train grid with $1024 \\times 1024$ test grid (PointPillars 512_2x). The results of these experiments are shown in Table 4.", + "bbox": [ + 75, + 474, + 468, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results show that only models that use any combination of VoxelNet Torchsparse, both PointPillars variants, and YOLOv8 are able to run above 10 FPS. From these configurations, we choose PointPillars 512_2x with YOLOv8 as the best configuration for all the above experiments as it achieves the best results across all the ablation studies. This is a promising result since we also know that this backbone configuration is able to run in real-time (11.2 FPS) on an RTX 3090 without using TensorRT acceleration.", + "bbox": [ + 75, + 672, + 468, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "An interesting observation is that utilizing pre-trained weights for transfer learning of YOLOv8 is not always beneficial, as the results from PointPillars $512 + \\mathrm{YOLOv8}$ show. This is likely because the pre-trained weights were from MS COCO [21], and they have a very different data domain compared to our dataset. Since MS COCO is also", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/193a63668ff5d7a771663d6ae7c6110ce8a11f5a42eea5055b414f37ea7a4100.jpg", + "image_caption": [ + "Figure 8. Qualitative results of our CoopDet3D (left) and the InfraDet3D (right) model on the TUMTraf Intersection test set during day and nighttime. Detected objects marked with a red circle were classified correctly by CoopDet3D." + ], + "image_footnote": [], + "bbox": [ + 509, + 89, + 696, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6ec51ed62d9f340fe69b01b55056f1cbd7bd0fdcfe8e391cf47aed40d835af47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 89, + 885, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "much larger than our dataset in terms of camera images, retraining harms the performance of the model slightly.", + "bbox": [ + 498, + 324, + 890, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In terms of efficiency, the goal of these experiments was to verify that the proposed CoopDet3D model with the best configuration provides the highest accuracy while also being able to run in real-time (minimum of $10\\mathrm{Hz}$ ). Furthermore, it should also be feasible to train the model on a high-performance GPU and perform inference on a mid-range consumer GPU deployable on an edge device. The results concerning the VRAM usage during inference show that the complexity of the model makes this feasible.", + "bbox": [ + 496, + 357, + 890, + 493 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and future work", + "text_level": 1, + "bbox": [ + 500, + 517, + 764, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work proposes the TUMTraf-V2X dataset, a multimodal multi-view V2X dataset for cooperative 3D object detection and tracking. Our dataset focuses on challenging traffic scenarios at an intersection and provides views from the infrastructure and the ego vehicle. To benchmark the dataset, we propose CoopDet3D - a baseline model for cooperative perception. Experiments show that cooperative fusion leads to higher efficacy than its unimodal and single-view camera-LiDAR fusion counterparts. Furthermore, cooperative fusion leads to an improvement of +14.3 3D mAP compared to vehicle-only perception, highlighting the need for V2X datasets. Finally, we provide our 3D BAT v24.3.2 labeling tool and devkit to load, parse, and visualize the dataset. It also includes modules for pre- and postprocessing and evaluation. Future efforts will integrate this platform into online environments, enabling a broader range of infrastructure-based, real-time perception applications.", + "bbox": [ + 496, + 546, + 890, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 827, + 651, + 843 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was supported by the Federal Ministry of Education and Research in Germany within the AUTOtech.agil project, Grant Number: 01IS22088U.", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22675", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers, 2022. 3", + "[2] Zhengwei Bai, Guoyuan Wu, Matthew J Barth, Yongkang Liu, Emrah Akin Sisbot, and Kentaro Oguchi. Pillargrid: Deep learning-based cooperative perception for 3d object detection from onboard-roadside lidar. In IEEE 25th International Conference on Intelligent Transportation Systems (ITSC), pages 1743-1749. IEEE, 2022. 4, 5, 7", + "[3] Zhengwei Bai, Guoyuan Wu, Matthew J Barth, Yongkang Liu, Emrah Akin Sisbot, and Kentaro Oguchi. Vinet: Lightweight, scalable, and heterogeneous cooperative perception for 3d object detection. Mechanical Systems and Signal Processing, 204:110723, 2023. 4", + "[4] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, pages 586-606. Spie, 1992. 4", + "[5] Steffen Busch, Christian Koetsier, Jeldrik Axmann, and Claus Brenner. Lumpi: The leibniz university multiperspective intersection dataset. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 1127-1134. IEEE, 2022. 2", + "[6] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11621-11631, 2020. 2, 3, 5", + "[7] Ziming Chen, Yifeng Shi, and Jinrang Jia. Transiff: An instance-level feature fusion framework for vehicleinfrastructure cooperative 3d detection with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18205-18214, 2023. 4", + "[8] MMYOLO Contributors. MMYOLO: OpenMMLab YOLO series toolbox and benchmark. https://github.com/open-mmlab/mmyolo, 2022.8", + "[9] Christian Creß, Zhenshan Bing, and Alois C. Knoll. Intelligent transportation systems using roadside infrastructure: A literature survey. IEEE Transactions on Intelligent Transportation Systems, pages 1-0, 2023. 2", + "[10] Christian Creß, Walter Zimmer, Leah Strand, Maximilian Fortkord, Siyi Dai, Venkatnarayanan Lakshminarasimhan, and Alois Knoll. A9-dataset: Multi-sensor infrastructure-based dataset for mobility research. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 965-970, 2022. 2", + "[11] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In Proceedings of the 1st Annual Conference on Robot Learning, pages 1-16, 2017. 3", + "[12] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 3", + "[13] Nicco Hagedorn. OpenLABEL Concept Paper. 2, 5", + "[14] Yuze He, Li Ma, Zhehao Jiang, Yi Tang, and Guoliang Xing. Vi-eye: Semantic-based 3d point cloud registration" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for infrastructure-assisted autonomous driving. In Proceedings of the 27th Annual International Conference on Mobile Computing and Networking, pages 573-586, 2021. 4", + "[15] Yue Hu, Yifan Lu, Runsheng Xu, Weidi Xie, Siheng Chen, and Yanfeng Wang. Collaboration helps camera overtake lidar in 3d detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243-9252, 2023. 4", + "[16] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolo, 2023. 8", + "[17] Annkathrin Krammer, Christoph Schöller, Dhiraj Gulati, Venkat Narayanan Lakshminarasimhan, Franz Kurz, Dominik Rosenbaum, Claus Lenz, and Alois Knoll. Providentia-a large-scale sensor system for the assistance of autonomous vehicles and its evaluation. Journal of Field Robotics, pages 1156-1176, 2022. 2", + "[18] Abhinav Kumar, Garrick Brazil, Enrique Corona, Armin Parchami, and Xiaoming Liu. Deviant: Depth equivariant network for monocular 3d object detection. In European Conference on Computer Vision, pages 664-683, 2022. 4", + "[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 4, 8", + "[20] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 7(4):10914-10921, 2022. 2, 3", + "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of European Conference on Computer Vision, ECCV 2014, Part V 13, pages 740-755. Springer, 2014. 8", + "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 8", + "[23] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE Int. Conference on Robotics and Automation (ICRA), pages 2774-2781, 2023. 4, 5, 7", + "[24] Pablo Alvarez Lopez, Michael Behrisch, Laura Bieker-Walz, Jakob Erdmann, Yun-Pang Flotteröd, Robert Hilbrich, Leonhard Lücken, Johannes Rummel, Peter Wagner, and Evamarie Wießner. Microscopic traffic simulation using sumo. In The 21st IEEE International Conference on Intelligent Transportation Systems. IEEE, 2018. 3", + "[25] Kaj Madsen, Hans Bruun Nielsen, and Ole Tingleff. Methods for non-linear least squares problems. 2004. 4", + "[26] Donghao Qiao and Farhana Zulkernine. Cobevfusion: Cooperative perception with lidar-camera bird's-eye view fusion. arXiv preprint arXiv:2310.06008, 2023. 4", + "[27] Morgan Quigley, Ken Conley, Brian Gerkey, Josh Faust, Tully Foote, Jeremy Leibs, Rob Wheeler, Andrew Y Ng," + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22676", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "et al. Ros: an open-source robot operating system. In ICRA workshop on open source software, page 5. Kobe, Japan, 2009. 5", + "[28] Ken Shoemake. Animating rotation with quaternion curves. In Proceedings of the 12th annual conference on Computer graphics and interactive techniques, pages 245-254, 1985. 4", + "[29] Rui Song, Chenwei Liang, Hu Cao, Zhiran Yan, Walter Zimmer, Markus Gross, Andreas Festag, and Alois Knoll. Collaborative semantic occupancy prediction with hybrid feature fusion in connected automated vehicles. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 4", + "[30] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 2, 3", + "[31] Haotian Tang, Shang Yang, Zhijian Liu, Ke Hong, Zhong-ming Yu, Xiuyu Li, Guohao Dai, Yu Wang, and Song Han. TorchSparse++: Efficient Point Cloud Engine. In Computer Vision and Pattern Recognition Workshops CVPRW, 2023. 8", + "[32] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4604-4612, 2020. 4", + "[33] Sizhe Wei, Yuxi Wei, Yue Hu, Yifan Lu, Yiqi Zhong, Siheng Chen, and Ya Zhang. Asynchrony-robust collaborative perception via bird's eye view flow. In Advances in Neural Information Processing Systems, pages 28462–28477. Curran Associates, Inc. 4", + "[34] Zizhang Wu, Yuanzhu Gan, Lei Wang, Guilian Chen, and Jian Pu. Monopgc: Monocular 3d object detection with pixel geometry contexts. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4842-4849, 2023. 4", + "[35] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 2, 3, 4", + "[36] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 3", + "[37] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 2, 3", + "[38] Junjie Yan, Yingfei Liu, Jianjian Sun, Fan Jia, Shuailin Li, Tiancai Wang, and Xiangyu Zhang. Cross modal transformer: Towards fast and robust 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18268-18278, 2023. 4" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Lei Yang, Kaicheng Yu, Tao Tang, Jun Li, Kun Yuan, Li Wang, Xinyu Zhang, and Peng Chen. Bevheight: A robust framework for vision-based roadside 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21611-21620, 2023. 4", + "[40] Xiaoqing Ye, Mao Shu, Hanyu Li, Yifeng Shi, Yingying Li, Guangjie Wang, Xiao Tan, and Errui Ding. Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21341-21350, 2022. 3", + "[41] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21361-21370, 2022. 2, 3", + "[42] Haibao Yu, Yingjuan Tang, Enze Xie, Jilei Mao, Jirui Yuan, Ping Luo, and Zaiqing Nie. Vehicle-infrastructure cooperative 3d object detection via feature flow prediction. arXiv preprint arXiv:2303.10552, 2023. 4", + "[43] Haibao Yu, Wenxian Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xu Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, et al. V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5486-5495, 2023. 2, 3", + "[44] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 8", + "[45] Walter Zimmer, Akshay Rangesh, and Mohan Trivedi. 3d bat: A semi-automatic, web-based 3d annotation toolbox for full-surface, multi-modal data streams. In 2019 IEEE Intelligent Vehicles Symposium (IV), pages 1816-1821. IEEE, 2019. 5", + "[46] Walter Zimmer, Emec Ercelik, Xingcheng Zhou, Xavier Jair Diaz Ortiz, and Alois Knoll. A survey of robust 3d object detection methods in point clouds. arXiv preprint arXiv:2204.00106, 2022. 4", + "[47] Walter Zimmer, Joseph Birkner, Marcel Brucker, Huu Tung Nguyen, Stefan Petrovski, Bohan Wang, and Alois C. Knoll. Infradet3d: Multi-modal 3d object detection based on roadside infrastructure camera and lidar sensors. In 2023 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2023. 4, 5, 8", + "[48] Walter Zimmer, Christian Creß, Huu Tung Nguyen, and Alois C Knoll. Tumtraf intersection dataset: All you need for urban 3d camera-lidar roadside perception [best student paper award]. In 2023 IEEE Intelligent Transportation Systems ITSC. IEEE, 2023. 2, 5, 8", + "[49] Walter Zimmer, Jialong Wu, Xingcheng Zhou, and Alois C Knoll. Real-time and robust 3d object detection with roadside lidars. In Proceedings of the 12th International Scientific Conference on Mobility and Transport: Mobility Innovations for Growing Megacities, pages 199-219. Springer, 2023. 4" + ], + "bbox": [ + 501, + 92, + 893, + 892 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22677", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/TUMTraf V2X Cooperative Perception Dataset/5f2a634b-45a2-4d92-81e8-cda538302a8f_model.json b/2024/TUMTraf V2X Cooperative Perception Dataset/5f2a634b-45a2-4d92-81e8-cda538302a8f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..737db02718e6416eb2e9b677bffded5f62b1ef72 --- /dev/null +++ b/2024/TUMTraf V2X Cooperative Perception Dataset/5f2a634b-45a2-4d92-81e8-cda538302a8f_model.json @@ -0,0 +1,2310 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.247, + 0.131, + 0.726, + 0.154 + ], + "angle": 0, + "content": "TUMTraf V2X Cooperative Perception Dataset" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.188, + 0.806, + 0.227 + ], + "angle": 0, + "content": "Walter Zimmer1 Gerhard Arya Wardana1 Suren Sritharan1 Xingcheng Zhou1 Rui Song1,2 Alois C. Knoll1" + }, + { + "type": "text", + "bbox": [ + 0.279, + 0.244, + 0.692, + 0.264 + ], + "angle": 0, + "content": "\\(^{1}\\)Technical University of Munich \\(^{2}\\)Fraunhofer IVI" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.281, + 0.773, + 0.298 + ], + "angle": 0, + "content": "https://tum-traffic-dataset.github.io/tumtraf-v2x" + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.301, + 0.878, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.509, + 0.893, + 0.539 + ], + "angle": 0, + "content": "Figure 1. Visualization of 3D box labels and tracks in our TUMTraf V2X Cooperative Perception Dataset. The top part shows the labels projected into the four camera images. The part below shows a point cloud from two LiDARs with 3D box labels of the same scene." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.55, + 0.314, + 0.566 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.584, + 0.473, + 0.885 + ], + "angle": 0, + "content": "Cooperative perception offers several benefits for enhancing the capabilities of autonomous vehicles and improving road safety. Using roadside sensors in addition to onboard sensors increases reliability and extends the sensor range. External sensors offer higher situational awareness for automated vehicles and prevent occlusions. We propose CoopDet3D, a cooperative multi-modal fusion model, and TUMTraf-V2X, a perception dataset, for the cooperative 3D object detection and tracking task. Our dataset contains 2,000 labeled point clouds and 5,000 labeled images from five roadside and four onboard sensors. It includes 30k 3D boxes with track IDs and precise GPS and IMU data. We labeled nine categories and covered occlusion scenarios with challenging driving maneuvers, like traffic violations, near-miss events, overtaking, and U-turns. Through multiple experiments, we show that our CoopDet3D camera-LiDAR fusion model achieves an increase of +14.36 3D mAP compared to a vehicle camera-LiDAR fusion model. Finally, we make our dataset, model, labeling tool, and devkit publicly available on our website." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.55, + 0.638, + 0.566 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.804 + ], + "angle": 0, + "content": "Cooperative perception involves the fusion of onboard sensor data and roadside sensor data, and it offers several advantages for enhancing the capabilities of autonomous vehicles and improving road safety. Using data from multiple sources makes the perception more robust to sensor failures or adverse environmental conditions. Roadside sensors provide an elevated view that helps to detect obstacles early. Moreover, they are also beneficial for precise vehicle localization and reduce the computational load of automated vehicles by offloading some perception tasks to the roadside sensors. Roadside sensors provide a global perspective of the traffic and offer a comprehensive situational awareness when fused with onboard sensor data. There are also fewer false positives or negatives because cooperative perception cross-validates the information from different sensors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.808, + 0.892, + 0.883 + ], + "angle": 0, + "content": "Infrastructure sensors can share perception-related information with vehicles through V2X. Due to minimal delay, and real-time capabilities, the infrastructure-based perception systems can further enhance the situational awareness and decision-making processes of vehicles." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Intelligent Transportation Systems (ITS) like the Testbed" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22668" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.088, + 0.895, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison of 3D cooperative V2X perception datasets with our proposed TUMTraf-V2X Cooperative Perception dataset (I=Infrastructure, V=Vehicle)." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.127, + 0.868, + 0.412 + ], + "angle": 0, + "content": "
DatasetOPV2V [36]V2XSet [35]V2X-Sim [20]V2V4Real [37]DAIR-V2X-C [41]V2X-Seq (SPD) [43]TUMTraf-V2X (Ours)
Year2022202220222022202220232024
V2XV2VV2V&IV2V&IV2VV2IV2IV2I
Real data---✓✓✓✓
Annotation range120 m120 m70 m200 m280 m280 m200 m
Day & night scenes----✓✓✓
# object classes11151098
Track IDs--✓✓-✓✓
HD Maps✓✓✓✓-✓✓
# of sensors (I | V)- | 6*- | 6*5 | 7- | 8‡2 | 32 | 35 | 4
Available worldwide✓✓✓✓--✓
Traffic violations------✓
Labeled attributes§------✓
OpenLABEL format------✓
# Point Clouds11k11k10k20k39k15k2.0k
# Images44k44k60k40k†39k15k5.0k
# 3D Boxes233k233k26k240k464k10.45k29.38k
LocationCARLACARLACARLAUSAChinaChinaGermany
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.093, + 0.414, + 0.316, + 0.428 + ], + "angle": 0, + "content": "\\(\\dagger\\) Image dataset has not been released yet." + }, + { + "type": "table_footnote", + "bbox": [ + 0.093, + 0.428, + 0.489, + 0.44 + ], + "angle": 0, + "content": "* Value per vehicle. Multiple Conn. and Autom. Vehicles (CAVs) are used." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.414, + 0.489, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "table_footnote", + "bbox": [ + 0.494, + 0.414, + 0.652, + 0.427 + ], + "angle": 0, + "content": "Total sensors from 2 CAVs." + }, + { + "type": "table_footnote", + "bbox": [ + 0.494, + 0.427, + 0.882, + 0.44 + ], + "angle": 0, + "content": "\\(^{\\S}\\) Weather, time of day, orientation, #3D points, occlusion, color, sub type" + }, + { + "type": "list", + "bbox": [ + 0.494, + 0.414, + 0.882, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.583 + ], + "angle": 0, + "content": "for Autonomous Driving [17] aim to improve safety by providing real-time traffic information. According to [9], testbeds extensively start using LiDAR sensors in their setups to create an accurate live digital twin of the traffic. Connected vehicles get a far-reaching view which enables them to react to breakdowns or accidents early. ITS systems also provide lane and speed recommendations to improve the traffic flow." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.734 + ], + "angle": 0, + "content": "The key challenge with ego-centric vehicle datasets is that there are many occlusions from a vehicle perspective, e.g., if a large truck in front of the ego vehicle obscures the view. Roadside sensors located at a smart intersection provide a broad overview of the intersection and a full-surround view. Given the immense potential of ITS, there is a specific need for V2X datasets. Despite the high costs associated with collecting and labeling such datasets, this work addresses this challenge as a crucial step toward realizing large-scale ITS implementations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.735, + 0.312, + 0.749 + ], + "angle": 0, + "content": "Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.75, + 0.47, + 0.84 + ], + "angle": 0, + "content": "- We provide a high-quality V2X dataset for the cooperative 3D object detection and tracking task with 2,000 labeled point clouds and 5,000 labeled images. In total, 30k 3D bounding boxes with track IDs were labeled in challenging traffic scenarios like near-miss events, overtaking scenarios, U-turn maneuvers, and traffic violation events." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.84, + 0.469, + 0.871 + ], + "angle": 0, + "content": "- We open-source our 3D bounding box annotation tool (3DBAT v24.3.2) to label multi-modal V2X datasets." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- We propose CoopDet3D, a cooperative 3D object detection model, and show in extensive experiments and abla" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.463, + 0.892, + 0.493 + ], + "angle": 0, + "content": "tion studies that it outperforms single view models on our V2X dataset by \\(+14.3\\) 3D mAP." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.494, + 0.892, + 0.584 + ], + "angle": 0, + "content": "- Finally, we provide a development kit to load the annotations in the widely recognized and standard format OpenLABEL [13], to facilitate a seamless integration and utilization of the dataset. Furthermore, it can preprocess, visualize, and convert labels to and from different dataset formats, and evaluate perception and tracking methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.599, + 0.644, + 0.615 + ], + "angle": 0, + "content": "2 . Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.894, + 0.672 + ], + "angle": 0, + "content": "3D autonomous driving datasets are mainly categorized based on the viewpoint. Table 1 highlights the main differences between our proposed dataset and other V2X datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.733, + 0.697 + ], + "angle": 0, + "content": "2.1. Single viewpoint datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Single viewpoint datasets are obtained from a single point of reference, either an ego-vehicle or roadside infrastructure. Onboard sensor-based datasets like KITTI [12], nuScenes [6], and Waymo [30] contain a diverse set of sensor data collected from a moving vehicle equipped with multiple sensors, including high-resolution cameras, LiDARs, radars, and GPS/INS systems. These datasets are abundant and provide many annotated data, including bounding boxes, track IDs, segmentation masks, and depth maps under different urban driving scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "On the other hand, roadside sensor-based datasets are in the infancy stage. High-quality multi-modal (camera and LiDAR) datasets are presented in [5, 10, 48], which" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22669" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.891, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.343, + 0.896, + 0.414 + ], + "angle": 0, + "content": "Figure 2. Our CoopDet3D framework is a multi-modal cooperative fusion system, comprising three distinct fusion pipelines. 1) The roadside camera-LiDAR fusion pipeline fuses three camera images and one LiDAR point cloud by extracting features and transforming them into a BEV representation. 2) The vehicle camera-LiDAR fusion pipeline fuses the vehicle camera feature map with the vehicle point cloud feature map using a convolutional fuser. 3) The vehicle and infrastructure feature maps are then fused by applying an element-wise max-pooling operation (Max Fuser). In the end, we use the TransFusion [1] 3D detection head to obtain 3D bounding box predictions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.473, + 0.558 + ], + "angle": 0, + "content": "are obtained from Infrastructure Perception Systems (IPS). Similarly, in [40], the authors provide a dataset consisting of only images taken from different viewpoints and under varying traffic conditions. These datasets provide a top-down view of a crowded intersection under different conditions and, as such, can overcome issues such as occlusions created by other vehicles and thereby have a higher number of object labels than onboard sensor-based datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.572, + 0.219, + 0.587 + ], + "angle": 0, + "content": "2.2. V2X datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.673 + ], + "angle": 0, + "content": "V2X datasets exploit the information from multiple viewpoints to gain additional knowledge regarding the environments. In this way, they overcome the limitations of single viewpoint datasets such as occlusion, limited field of view (FOV), and low point cloud density." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.471, + 0.884 + ], + "angle": 0, + "content": "DAIR-V2X dataset family [41] is one of the foremost cooperative multi-modal datasets introduced. It contains three subsets: an intersection, a vehicle, and a cooperative dataset. The cooperative dataset contains 464k 3D box labels belonging to 10 classes, making it one of the largest cooperative datasets. The V2X-Seq dataset [43] extends selected sequences of the DAIR-V2X dataset with track IDs and is partitioned into a sequential perception dataset (SPD) and a trajectory forecasting dataset. Despite these, the lack of specific information, such as the labeling methodology used, the exact models of the sensors deployed, the distribution of the classes, and the scenarios within the dataset, leads to uncertainty in the extendability and application of this dataset in varying conditions." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In V2V4Real [37], the authors propose a multi-modal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.895, + 0.544 + ], + "angle": 0, + "content": "cooperative dataset focusing only on V2V perception. Two vehicles equipped with cameras, LiDAR, and GPS/IMU integration systems are used to collect multi-modal sensor data for diverse scenarios. As opposed to all other V2X datasets, this focuses on V2V perception, and though it is of similar size to other cooperative datasets, it contains fewer classes and 3D bounding box information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.895, + 0.697 + ], + "angle": 0, + "content": "Simulated multi-agent perception datasets have been proposed in [20, 35, 36]. These datasets contain multimodal sensor data (camera and LiDAR) obtained from roadside units (RSUs) and multiple ego vehicles, which enable collaborative perception. They use a combination of simulators such as SUMO [24], CARLA [11], and OpenCDA [36] for flow simulation, data retrieval, and V2X communication. However, the utility of the dataset is still limited due to the simulated nature of the data, and its extendability to real-life applications has not been studied in detail." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.709, + 0.871, + 0.726 + ], + "angle": 0, + "content": "2.3. V2X perception models for object detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.824 + ], + "angle": 0, + "content": "The datasets presented above have been used to develop various models for a wide variety of tasks, with the majority focusing on 3D object detection. Different approaches have been taken depending on the availability and challenges, and these methods are grouped based on the number of nodes employed and the modalities used for detection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Most 3D object detection models use multi-modal sensor data obtained from a single point of view, which is often an ego-vehicle. Due to the popularity and abundant availability of onboard datasets [6, 12, 30], most models use images, point clouds, or both modalities. Image-based models were" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22670" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "the pioneers in 3D object detection due to their low cost and simplicity, and both vehicular camera-based models [18, 34] and infrastructure camera-based models [39] have been proposed. LiDAR-based 3D object detection models [19, 46, 49] became popular since LiDAR point clouds provide 3D depth information and are robust, especially in adverse weather conditions and nighttime scenarios. Fusion models combine the information obtained from both images and point clouds and have been shown to outperform the prior methods [47]. Single viewpoint fusion models use either vehicular camera and LiDAR [23, 32, 38] or infrastructure camera and LiDAR [47] for 3D object detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.485 + ], + "angle": 0, + "content": "Cooperative perception models, which use data from multiple viewpoints, have been shown to overcome issues related to occlusion, which were often present in vehicular sensor-based models. V2I cooperative perception models [2, 3, 14, 26, 33, 35, 42] use the sensor data from both vehicles and infrastructure and V2V models [15, 29] communicate the sensor data between multiple vehicles. In this work, our cooperative multi-modal dataset is one contribution among others. Thus, while most of the prior works focus on unimodal cooperative perception using either Li-DAR point clouds [2, 7] or camera images [15], we benchmark our dataset with CoopDet3D, a deep fusion based cooperative multi-modal 3D object detection model based on BEVFusion [15] and PillarGrid [2]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.497, + 0.305, + 0.513 + ], + "angle": 0, + "content": "3. TUMTraf-V2X Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.473, + 0.75 + ], + "angle": 0, + "content": "Our TUMTraf V2X Cooperative Perception Dataset focuses on challenging traffic scenarios and various day and nighttime scenes. The data is further annotated, emphasizing high-quality labels through careful labeling and high-quality review processes. It also contains dense traffic and fast-moving vehicles, which reveals the specific challenges in cooperative perception, such as pose estimation errors, latency, and synchronization. Furthermore, we provide sensor data from nine different sensors covering the same traffic scenes under diverse weather conditions and lighting variations. The infrastructure sensors are oriented in all four directions of the intersection to get a \\(360^{\\circ}\\) view, which leads to better perception results. Finally, it contains rare events like traffic violations where pedestrians cross the road at a busy four-way intersection while the crossing light is lit red." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.758, + 0.214, + 0.773 + ], + "angle": 0, + "content": "3.1. Sensor setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.811 + ], + "angle": 0, + "content": "Our TUMTraf V2X Cooperative Perception Dataset was recorded on an ITS system with nine sensors." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.811, + 0.395, + 0.825 + ], + "angle": 0, + "content": "The infrastructure sensor setup is the following:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.826, + 0.468, + 0.855 + ], + "angle": 0, + "content": "- 1x Ouster LiDAR OS1-64 (gen. 2), 64 vert. layers, \\(360^{\\circ}\\) FOV, below horizon config., \\(10\\mathrm{cm}\\) acc. @ \\(120\\mathrm{m}\\) range" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.469, + 0.884 + ], + "angle": 0, + "content": "- 4x Basler ace acA1920-50gc, \\(1920 \\times 1200\\), Sony IMX174 with 8 mm lenses" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.826, + 0.469, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.399, + 0.901 + ], + "angle": 0, + "content": "On the vehicle, the following sensors were used:" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.677, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.089, + 0.89, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.252, + 0.892, + 0.307 + ], + "angle": 0, + "content": "Figure 3. Demonstration of a possible V2X occlusion scenario. A pedestrian (blue) is crossing the road in front of the ego vehicle. An occluded bicycle is marked in red. The recording vehicle with the sensor setup is shown in the bottom left corner." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.332, + 0.891, + 0.361 + ], + "angle": 0, + "content": "- 1x Robosense RS-LiDAR-32, 32 vert. layers, \\(360^{\\circ}\\) FOV, 3 cm accuracy @200 m range" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.362, + 0.891, + 0.391 + ], + "angle": 0, + "content": "- 1x Basler ace acA1920-50gc, \\(1920 \\times 1200\\), Sony IMX174 with \\(16 \\mathrm{~mm}\\) lens" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.392, + 0.879, + 0.406 + ], + "angle": 0, + "content": "- 1x Emlid Reach RS2+ multi-band RTK GNSS receiver" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.408, + 0.735, + 0.421 + ], + "angle": 0, + "content": "- 1x XSENS MTi-30-2A8G4 IMU" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.332, + 0.891, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.432, + 0.806, + 0.448 + ], + "angle": 0, + "content": "3.2. Sensor calibration and registration" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.697 + ], + "angle": 0, + "content": "We synchronize the cameras and LiDARs in the spatial and temporal domain. First, we determine the intrinsic camera parameters and the radial and tangential image distortions by using a checkerboard target. We then calibrate the roadside LiDAR with the roadside cameras by picking 100-point pairs in the point cloud and camera image. Extrinsic parameters (rotation and translation) are calculated by minimizing the reprojection error of 2D-3D point correspondences [25]. We follow the same procedure for onboard camera-LiDAR calibration. Finally, we calibrate the onboard LiDAR to the roadside LiDAR. This spatial registration is done by first estimating a coarse transformation. We pick ten 3D point pairs in each point cloud and minimize their distance using the least squares method. Then, we apply the point-to-point Iterative Closest Point (ICP) algorithm [4] to get the fine transformation between the point clouds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.848 + ], + "angle": 0, + "content": "We label the vehicle and infrastructure point clouds after registering them. The coarse registration was done by measuring the GPS position of the onboard LiDAR and the roadside LiDAR. Then, we transform every 10th onboard point cloud to the coordinate system of the infrastructure point cloud. The fine registration was done by applying the point-to-point ICP to get an accurate V2I transformation matrix. All rotations of the point cloud frames in between are interpolated based on the spherical linear interpolation (SLERP) [28] method:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.858, + 0.891, + 0.876 + ], + "angle": 0, + "content": "\\[\nS L E R P \\left(q _ {0}, q _ {1}, t\\right) = q _ {0} \\left(q _ {0} ^ {- 1} q _ {1}\\right) ^ {t}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(q_{0}\\) and \\(q_{1}\\) are the quaternions representing the ro" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22671" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.116 + ], + "angle": 0, + "content": "Table 2. Evaluation results (\\(mAP_{BEV}\\) and \\(mAP_{3D}\\)) of CoopDet3D on our TUMTraf-V2X test set in south2 FOV." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.128, + 0.468, + 0.264 + ], + "angle": 0, + "content": "
Config.\\( \\mathbf{mAP_{BEV}}\\uparrow \\)\\( \\mathbf{mAP_{3D}}\\uparrow \\)
DomainModalityEasy↑Mod.↑Hard↑Avg.↑
VehicleCamera46.8331.4737.8230.7730.36
VehicleLiDAR85.3385.2276.8669.0480.11
VehicleCam+LiDAR84.9077.6072.0873.1276.40
Infra.Camera61.9831.1946.7340.4235.04
Infra.LiDAR92.8686.1788.0775.7384.88
Infra.Cam+LiDAR92.9287.9989.0981.6987.01
Coop.Camera68.9445.4142.7657.8345.74
Coop.LiDAR93.9392.6378.0673.9585.86
Coop.Cam+LiDAR94.2293.4288.1779.9490.76
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.291, + 0.468, + 0.322 + ], + "angle": 0, + "content": "tations of the start and end frames and \\( t \\in [0,1] \\). Translation vectors \\( \\mathbf{T}_0 \\) and \\( \\mathbf{T}_1 \\) were obtained using linear interpolation:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.336, + 0.468, + 0.353 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} (t) = \\mathbf {T} _ {0} + t \\left(\\mathbf {T} _ {1} - \\mathbf {T} _ {0}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.36, + 0.468, + 0.419 + ], + "angle": 0, + "content": "This dual interpolation strategy ensures that the estimated transformations between the frames are smooth and geometrically accurate, thus adhering closely to the actual movements of the vehicle over time." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.429, + 0.322, + 0.446 + ], + "angle": 0, + "content": "3.3. Data selection and labeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.47, + 0.816 + ], + "angle": 0, + "content": "We selected the data based on challenging traffic scenarios, like U-turns, tailgate events, and traffic violation maneuvers. Besides the high traffic density of 31 objects per frame, we selected frames with high-class coverage. We selected 700 frames during sunny daytime and 100 frames during cloudy nighttime for labeling. The camera and Li-DAR data were recorded into rosbag files at \\(15\\mathrm{Hz}\\) and \\(10\\mathrm{Hz}\\), respectively. We extracted and synchronized the data based on ROS [27] timestamps and labeled it with our \\(3D\\) BAT (v24.3.2) annotation tool1. We improved the 3D BAT [45] baseline labeling tool to label 3D objects faster and more precisely with a one-click annotation feature. The annotators were instructed to label traffic participants while examining the images. Objects are still labeled, even if they have no 3D points inside, but are visible in the images. Extremities (e.g., pedestrian limbs) are included in the bounding box, but side mirrors of vehicles aren't. If a pedestrian carries an object, that object is included in the bounding box. If two or more pedestrians are carrying an object, only the box of one will include the object. After labeling, each annotator checked the work of other annotators manually frame-by-frame. When errors were found, the original annotator was notified, and they fixed it. This helps ensure that the labels in our dataset are high quality." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.825, + 0.318, + 0.839 + ], + "angle": 0, + "content": "3.4. Data structure and format" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.848, + 0.468, + 0.878 + ], + "angle": 0, + "content": "We record eight different scenes, each 10 sec. long, from vehicle and infrastructure perspectives using nine sensors" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.088, + 0.892, + 0.13 + ], + "angle": 0, + "content": "Table 3. Evaluation results of infrastructure-only CoopDet3D vs. InfraDet3D [47] on TUMTraf Intersection test set [48]. South 1 and South 2 refer to sensors covering different FOVs." + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.141, + 0.892, + 0.267 + ], + "angle": 0, + "content": "
ModelFOVModality\\( \\mathrm{mAP}_{3D} \\uparrow \\)
Easy↑Mod.↑Hard↑Avg.↑
InfraDet3Dsouth 1LiDAR75.8147.6642.1655.21
CoopDet3Dsouth 1LiDAR76.2448.2335.1969.47
InfraDet3Dsouth 2LiDAR38.9246.6043.8643.13
CoopDet3Dsouth 2LiDAR74.9755.5539.9669.94
InfraDet3Dsouth 1Cam+LiDAR67.0831.3835.1744.55
CoopDet3Dsouth 1Cam+LiDAR75.6845.6345.6366.75
InfraDet3Dsouth 2Cam+LiDAR58.3819.7333.0837.06
CoopDet3Dsouth 2Cam+LiDAR74.7353.4641.9666.89
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.295, + 0.892, + 0.355 + ], + "angle": 0, + "content": "and split the data into a train (80%), val. (10%), and test (10%) set. We use stratified sampling to distribute all sets' object classes equally (see Fig. 6a). Labels are provided in the ASAM OpenLABEL [13] standard." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.365, + 0.724, + 0.381 + ], + "angle": 0, + "content": "3.5. Dataset development kit" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.388, + 0.892, + 0.599 + ], + "angle": 0, + "content": "We provide a devkit to work with our dataset. In addition to generating the data statistics, it provides modules for multiclass stratified splitting (train/val/test), point cloud registration, loading annotations in OpenLABEL format, evaluation of detection and tracking results, pre-processing steps such as point cloud filtering, and post-processing such as bounding box filtering. The statistics (Fig. 4, 5, and 6) were created using our devkit. It also contains modules to convert the labels from OpenLABEL to KITTI or our custom nuScenes format with timestamps instead of tokens and vice versa. This devkit enables users to migrate their models and make them compatible with our dataset format. We release our devkit2 under the MIT license and the dataset under the Creative Commons (CC) BY-NC-ND 4.0 license." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.613, + 0.628, + 0.629 + ], + "angle": 0, + "content": "4. Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.639, + 0.892, + 0.684 + ], + "angle": 0, + "content": "We propose CoopDet3D, an extension of BEVFusion [23] and PillarGrid [2] for deep cooperative multi-modal 3D object detection and benchmark it on TUMTraf-V2X mini." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.694, + 0.684, + 0.708 + ], + "angle": 0, + "content": "4.1. Evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.717, + 0.892, + 0.868 + ], + "angle": 0, + "content": "The accuracy is measured in terms of the mean average precision (mAP). Two types of mAP measures are used: BEV mAP considers the BEV center distance, and the results are obtained using the same evaluation methodology used in BEVFusion, which in turn uses the evaluation protocol of nuScenes [6]. Similarly, the 3D mAP measure considers the intersection in 3D, and the results are obtained using the evaluation script of our TUM Traffic dataset devkit. The runtime is evaluated using frames per second (FPS) as the metric and the results were obtained by measuring the time" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.346, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/walzimmer/3d-bat" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.877, + 0.89, + 0.899 + ], + "angle": 0, + "content": "2https://github.com/tum-traffic-dataset/tum-traffic-dataset-dev-kit" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22672" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.089, + 0.351, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.253, + 0.343, + 0.266 + ], + "angle": 0, + "content": "(a) Distribution of objects between day and night." + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.089, + 0.621, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.253, + 0.62, + 0.266 + ], + "angle": 0, + "content": "(b) Avg. and max. num. of 3D points for each class." + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.089, + 0.895, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.253, + 0.875, + 0.266 + ], + "angle": 0, + "content": "(c) Avg. and max. track length for all classes." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.289, + 0.895, + 0.346 + ], + "angle": 0, + "content": "Figure 4. Our TUMTraf-V2X dataset contains 30k 3D box labels in total and is balanced among nine different object classes. (a) Cars (14,633) and pedestrians (5,276) are highly represented in the dataset. (b) 3D box labels contain, on average 625 points inside which shows the density of the labeled objects. The BUS class has the highest point density. (c) All traffic participants are tracked for \\(39\\mathrm{m}\\) on average. Emergency vehicles have the highest average track length of \\(128\\mathrm{m}\\), whereas the CAR class contains a max. track length of \\(174\\mathrm{m}\\)." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.359, + 0.341, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.546, + 0.306, + 0.558 + ], + "angle": 0, + "content": "(a) Visualization of rotations (yaw)." + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.357, + 0.621, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.546, + 0.575, + 0.558 + ], + "angle": 0, + "content": "(b) 3D points grouped by distance." + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.357, + 0.884, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.545, + 0.838, + 0.558 + ], + "angle": 0, + "content": "(c) BEV visualization of tracks." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.569, + 0.894, + 0.611 + ], + "angle": 0, + "content": "Figure 5. Our dataset was recorded at a crowded intersection with many left and right turns. (a) Most of the vehicles (7,559) are driving in the east direction (0 degrees). (b) 3D boxes were labeled up to \\(200\\mathrm{m}\\) range and are very dense between 10 and \\(60\\mathrm{m}\\). (c) The visualization of BEV tracks shows where pedestrians and bicycles are crossing the road." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.623, + 0.345, + 0.807 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.811, + 0.286, + 0.824 + ], + "angle": 0, + "content": "(a) Histogram of 3D points." + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.623, + 0.616, + 0.808 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.811, + 0.568, + 0.824 + ], + "angle": 0, + "content": "(b) Histogram of 3D box labels." + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.623, + 0.888, + 0.809 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.811, + 0.836, + 0.824 + ], + "angle": 0, + "content": "(c) Histogram of track lengths." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.835, + 0.895, + 0.879 + ], + "angle": 0, + "content": "Figure 6. (a) Most 3D box labels contain between 1 and 200 3D points inside, with an average of 434 3D points, excluding empty boxes. Objects that were close to both LiDAR sensors even contained up to 18k 3D points. (b) Frames contain between 15 and 45 traffic participants, with an average of 29. (c) Objects were tracked up to \\(170\\mathrm{m}\\), and the average track length is \\(39.35\\mathrm{m}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22673" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.894, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.301, + 0.893, + 0.345 + ], + "angle": 0, + "content": "Figure 7. Qualitative results on the TUMTraf-V2X Cooperative Perception test set. The first row shows the inference results of the onboard (vehicle-only) camera-LiDAR fusion with 23 detected objects. In the second row, the results of the cooperative vehicle-infrastructure camera-LiDAR fusion are visualized. Here, all 25 traffic participants could be detected with the support of roadside sensors." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.365, + 0.47, + 0.422 + ], + "angle": 0, + "content": "Table 4. Ablation study on cooperative 3D object detection with 11 combinations of camera and LiDAR backbones. The best tradeoff between speed and accuracy is highlighted in gray. \\(\\mathrm{ND} =\\) non-deterministic, \\(\\mathrm{TS} =\\) TorchSparse, \\(\\odot =\\) retrained. VRAM is in GiB." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.433, + 0.467, + 0.591 + ], + "angle": 0, + "content": "
Backbone Configuration\\( \\mathbf{{mAP}}_{\\mathbf{{BEV}}} \\uparrow \\)\\( \\mathbf{{FPS}} \\uparrow \\)\\( \\mathbf{{VRAM}} \\downarrow \\)
VoxelNet ND + SwinT93.476.306.69
VoxelNet ND + YOLOv8 s92.947.246.39
VoxelNet TS + SwinT93.518.844.61
VoxelNet TS + YOLOv8 s92.9410.664.28
VoxelNet TS + YOLOv8 s ○94.3110.664.28
PointPillars 512 + Swin T94.439.004.94
PointPillars 512 + YOLOv8 s94.2711.144.63
PointPillars 512 + YOLOv8 s ○94.2511.144.63
PointPillars 512_2x + Swin T92.799.064.94
PointPillars 512_2x + YOLOv8 s94.1611.204.63
PointPillars 512_2x + YOLOv8 s ○94.2211.204.63
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.47, + 0.711 + ], + "angle": 0, + "content": "needed by the model to run one full inference, including data preprocessing and voxelization. The first five iterations are skipped as a warmup since they are usually considerably slower than the average. Finally, the complexity of the model is measured in terms of the maximal VRAM usage across all GPUs during training and testing." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.258, + 0.74 + ], + "angle": 0, + "content": "4.2. CoopDet3D model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Our CoopDet3D uses a BEVFusion-based backbone for camera-LiDAR fusion on the vehicle and the infrastructure sides separately to obtain the vehicle and infrastructure features. The best backbone for image and point cloud feature extraction was chosen through multiple ablation studies. Then, inspired by the method proposed by PillarGrid [2], an element-wise max-pooling operation is proposed to fuse the resulting fused camera-LiDAR features of vehicle and infrastructure together. Finally, the detection head from BEVFusion is used for 3D detection from the fused feature." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.369, + 0.836, + 0.383 + ], + "angle": 0, + "content": "The architecture of CoopDet3D is shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.581 + ], + "angle": 0, + "content": "First, we disable the camera feature extraction nodes and train the LiDAR-only model for 20 epochs. Then, we use pre-trained weights for the cooperative model and fine-tune the entire model for eight further epochs. Hyperparameter tuning revealed that the default hyperparameters of BEV-Fusion [23] gave the best results, and such were not modified. The preprocessing steps are also the same as for BEV-Fusion, but we change the point cloud range to \\([-75, 75]\\) in the x- and y-scale and \\([-8, 0]\\) in the z-scale since the dataset used in this case is different. Furthermore, we use 3x NVIDIA RTX 3090 GPUs with 24 GB VRAM for training and a single GPU for evaluation. We open-source our model and provide pre-trained weights3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.59, + 0.795, + 0.606 + ], + "angle": 0, + "content": "4.3. Experiments and ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.689 + ], + "angle": 0, + "content": "The objective of these experiments is to highlight the importance of our V2X multi-viewpoint dataset as opposed to single-viewpoint datasets. As such, we conduct multiple experiments and ablation studies with data obtained from each viewpoint and compare the results on the proposed model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.786 + ], + "angle": 0, + "content": "Ablation on viewpoint and modality. We conduct multiple experiments with all possible combinations of a) viewpoints: vehicle-only, infrastructure-only, cooperative, and b) modalities: camera-only, LiDAR-only, and camera-LiDAR fusion. Table 2 shows the mAP achieved by CoopDet3D for each of these combinations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.787, + 0.892, + 0.879 + ], + "angle": 0, + "content": "We observe that the results follow a general pattern of cooperative performance being better than infrastructure-only, which is, in turn, better than vehicle-only. Furthermore, fusion models perform better than LiDAR-only models, which in turn are better than camera-only models. Figure 7 shows qualitative results between our vehicle-only" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.86, + 0.9 + ], + "angle": 0, + "content": "3https://github.com/tum-traffic-dataset/coopdet3d" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22674" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "camera-LiDAR fusion model and our cooperative vehicle-infrastructure camera-LiDAR fusion model. Again, we observe from these samples that the cooperative perception model is able to detect 25 traffic participants, whereas the vehicle-only model is only able to detect 23 objects due to occlusions and a limited field of view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.192, + 0.47, + 0.357 + ], + "angle": 0, + "content": "Ablation on fusion level. We compare our CoopDet3D model to the current SOTA camera-LiDAR fusion method InfraDet3D [47] on the TUMTraf Intersection test set [48]. The proposed method uses deep fusion, whereas the InfraDet3D method is a late fusion method. Table 3 shows the performance of our model against InfraDet3D. The results show that the proposed deep fusion method outperforms the SOTA late fusion model in all metrics, except in the hard difficulty in LiDAR-only mode. Furthermore, Fig. 8 shows two sample images taken during day and nighttime, wherein deep fusion again outperforms late fusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.47, + 0.466 + ], + "angle": 0, + "content": "We note that these experiments were conducted in an offline setting, disregarding other considerations for simplicity. However, when deploying it in real life, factors such as the transmission bandwidth should also be considered. Since we observed that deep feature fusion generally leads to higher efficacy, the V2I transmissions should contain these features instead of infrastructure bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.47, + 0.671 + ], + "angle": 0, + "content": "Ablation on backbones. As an ablation study, we present the results of the experiments to find the best backbone and model configuration for the cooperative camera-LiDAR fusion model. For the camera backbone, SwinT [22] and MMYOLO's [8] implementation of YOLOv8 [16] were considered. For the LiDAR backbone, VoxelNet [44] and PointPillars [19] were considered. In addition, VoxelNet was implemented with two different backends, namely SPConv v2 and Torchsparse [31]. For PointPillars, two grid sizes are considered \\(512 \\times 512\\) for both train and test grids (PointPillars 512) and \\(512 \\times 512\\) train grid with \\(1024 \\times 1024\\) test grid (PointPillars 512_2x). The results of these experiments are shown in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.809 + ], + "angle": 0, + "content": "The results show that only models that use any combination of VoxelNet Torchsparse, both PointPillars variants, and YOLOv8 are able to run above 10 FPS. From these configurations, we choose PointPillars 512_2x with YOLOv8 as the best configuration for all the above experiments as it achieves the best results across all the ablation studies. This is a promising result since we also know that this backbone configuration is able to run in real-time (11.2 FPS) on an RTX 3090 without using TensorRT acceleration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.903 + ], + "angle": 0, + "content": "An interesting observation is that utilizing pre-trained weights for transfer learning of YOLOv8 is not always beneficial, as the results from PointPillars \\(512 + \\mathrm{YOLOv8}\\) show. This is likely because the pre-trained weights were from MS COCO [21], and they have a very different data domain compared to our dataset. Since MS COCO is also" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.09, + 0.697, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.09, + 0.887, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.237, + 0.892, + 0.293 + ], + "angle": 0, + "content": "Figure 8. Qualitative results of our CoopDet3D (left) and the InfraDet3D (right) model on the TUMTraf Intersection test set during day and nighttime. Detected objects marked with a red circle were classified correctly by CoopDet3D." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.325, + 0.892, + 0.355 + ], + "angle": 0, + "content": "much larger than our dataset in terms of camera images, retraining harms the performance of the model slightly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.494 + ], + "angle": 0, + "content": "In terms of efficiency, the goal of these experiments was to verify that the proposed CoopDet3D model with the best configuration provides the highest accuracy while also being able to run in real-time (minimum of \\(10\\mathrm{Hz}\\)). Furthermore, it should also be feasible to train the model on a high-performance GPU and perform inference on a mid-range consumer GPU deployable on an edge device. The results concerning the VRAM usage during inference show that the complexity of the model makes this feasible." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.518, + 0.766, + 0.535 + ], + "angle": 0, + "content": "5. Conclusion and future work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.803 + ], + "angle": 0, + "content": "This work proposes the TUMTraf-V2X dataset, a multimodal multi-view V2X dataset for cooperative 3D object detection and tracking. Our dataset focuses on challenging traffic scenarios at an intersection and provides views from the infrastructure and the ego vehicle. To benchmark the dataset, we propose CoopDet3D - a baseline model for cooperative perception. Experiments show that cooperative fusion leads to higher efficacy than its unimodal and single-view camera-LiDAR fusion counterparts. Furthermore, cooperative fusion leads to an improvement of +14.3 3D mAP compared to vehicle-only perception, highlighting the need for V2X datasets. Finally, we provide our 3D BAT v24.3.2 labeling tool and devkit to load, parse, and visualize the dataset. It also includes modules for pre- and postprocessing and evaluation. Future efforts will integrate this platform into online environments, enabling a broader range of infrastructure-based, real-time perception applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.828, + 0.652, + 0.844 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "This research was supported by the Federal Ministry of Education and Research in Germany within the AUTOtech.agil project, Grant Number: 01IS22088U." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22675" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.471, + 0.254 + ], + "angle": 0, + "content": "[2] Zhengwei Bai, Guoyuan Wu, Matthew J Barth, Yongkang Liu, Emrah Akin Sisbot, and Kentaro Oguchi. Pillargrid: Deep learning-based cooperative perception for 3d object detection from onboard-roadside lidar. In IEEE 25th International Conference on Intelligent Transportation Systems (ITSC), pages 1743-1749. IEEE, 2022. 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.324 + ], + "angle": 0, + "content": "[3] Zhengwei Bai, Guoyuan Wu, Matthew J Barth, Yongkang Liu, Emrah Akin Sisbot, and Kentaro Oguchi. Vinet: Lightweight, scalable, and heterogeneous cooperative perception for 3d object detection. Mechanical Systems and Signal Processing, 204:110723, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[4] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, pages 586-606. Spie, 1992. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.422 + ], + "angle": 0, + "content": "[5] Steffen Busch, Christian Koetsier, Jeldrik Axmann, and Claus Brenner. Lumpi: The leibniz university multiperspective intersection dataset. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 1127-1134. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[6] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11621-11631, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.471, + 0.577 + ], + "angle": 0, + "content": "[7] Ziming Chen, Yifeng Shi, and Jinrang Jia. Transiff: An instance-level feature fusion framework for vehicleinfrastructure cooperative 3d detection with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18205-18214, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.578, + 0.471, + 0.618 + ], + "angle": 0, + "content": "[8] MMYOLO Contributors. MMYOLO: OpenMMLab YOLO series toolbox and benchmark. https://github.com/open-mmlab/mmyolo, 2022.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.62, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[9] Christian Creß, Zhenshan Bing, and Alois C. Knoll. Intelligent transportation systems using roadside infrastructure: A literature survey. IEEE Transactions on Intelligent Transportation Systems, pages 1-0, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.745 + ], + "angle": 0, + "content": "[10] Christian Creß, Walter Zimmer, Leah Strand, Maximilian Fortkord, Siyi Dai, Venkatnarayanan Lakshminarasimhan, and Alois Knoll. A9-dataset: Multi-sensor infrastructure-based dataset for mobility research. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 965-970, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In Proceedings of the 1st Annual Conference on Robot Learning, pages 1-16, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.856 + ], + "angle": 0, + "content": "[12] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.858, + 0.421, + 0.872 + ], + "angle": 0, + "content": "[13] Nicco Hagedorn. OpenLABEL Concept Paper. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Yuze He, Li Ma, Zhehao Jiang, Yi Tang, and Guoliang Xing. Vi-eye: Semantic-based 3d point cloud registration" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "for infrastructure-assisted autonomous driving. In Proceedings of the 27th Annual International Conference on Mobile Computing and Networking, pages 573-586, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[15] Yue Hu, Yifan Lu, Runsheng Xu, Weidi Xie, Siheng Chen, and Yanfeng Wang. Collaboration helps camera overtake lidar in 3d detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243-9252, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[16] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolo, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[17] Annkathrin Krammer, Christoph Schöller, Dhiraj Gulati, Venkat Narayanan Lakshminarasimhan, Franz Kurz, Dominik Rosenbaum, Claus Lenz, and Alois Knoll. Providentia-a large-scale sensor system for the assistance of autonomous vehicles and its evaluation. Journal of Field Robotics, pages 1156-1176, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[18] Abhinav Kumar, Garrick Brazil, Enrique Corona, Armin Parchami, and Xiaoming Liu. Deviant: Depth equivariant network for monocular 3d object detection. In European Conference on Computer Vision, pages 664-683, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[20] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 7(4):10914-10921, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of European Conference on Computer Vision, ECCV 2014, Part V 13, pages 740-755. Springer, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[23] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE Int. Conference on Robotics and Automation (ICRA), pages 2774-2781, 2023. 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[24] Pablo Alvarez Lopez, Michael Behrisch, Laura Bieker-Walz, Jakob Erdmann, Yun-Pang Flotteröd, Robert Hilbrich, Leonhard Lücken, Johannes Rummel, Peter Wagner, and Evamarie Wießner. Microscopic traffic simulation using sumo. In The 21st IEEE International Conference on Intelligent Transportation Systems. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[25] Kaj Madsen, Hans Bruun Nielsen, and Ole Tingleff. Methods for non-linear least squares problems. 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[26] Donghao Qiao and Farhana Zulkernine. Cobevfusion: Cooperative perception with lidar-camera bird's-eye view fusion. arXiv preprint arXiv:2310.06008, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Morgan Quigley, Ken Conley, Brian Gerkey, Josh Faust, Tully Foote, Jeremy Leibs, Rob Wheeler, Andrew Y Ng," + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22676" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "et al. Ros: an open-source robot operating system. In ICRA workshop on open source software, page 5. Kobe, Japan, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[28] Ken Shoemake. Animating rotation with quaternion curves. In Proceedings of the 12th annual conference on Computer graphics and interactive techniques, pages 245-254, 1985. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.471, + 0.258 + ], + "angle": 0, + "content": "[29] Rui Song, Chenwei Liang, Hu Cao, Zhiran Yan, Walter Zimmer, Markus Gross, Andreas Festag, and Alois Knoll. Collaborative semantic occupancy prediction with hybrid feature fusion in connected automated vehicles. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.471, + 0.344 + ], + "angle": 0, + "content": "[30] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[31] Haotian Tang, Shang Yang, Zhijian Liu, Ke Hong, Zhong-ming Yu, Xiuyu Li, Guohao Dai, Yu Wang, and Song Han. TorchSparse++: Efficient Point Cloud Engine. In Computer Vision and Pattern Recognition Workshops CVPRW, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.471, + 0.467 + ], + "angle": 0, + "content": "[32] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4604-4612, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.471, + 0.538 + ], + "angle": 0, + "content": "[33] Sizhe Wei, Yuxi Wei, Yue Hu, Yifan Lu, Yiqi Zhong, Siheng Chen, and Ya Zhang. Asynchrony-robust collaborative perception via bird's eye view flow. In Advances in Neural Information Processing Systems, pages 28462–28477. Curran Associates, Inc. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[34] Zizhang Wu, Yuanzhu Gan, Lei Wang, Guilian Chen, and Jian Pu. Monopgc: Monocular 3d object detection with pixel geometry contexts. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4842-4849, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.471, + 0.678 + ], + "angle": 0, + "content": "[35] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.471, + 0.748 + ], + "angle": 0, + "content": "[36] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[37] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[38] Junjie Yan, Yingfei Liu, Jianjian Sun, Fan Jia, Shuailin Li, Tiancai Wang, and Xiangyu Zhang. Cross modal transformer: Towards fast and robust 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18268-18278, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[39] Lei Yang, Kaicheng Yu, Tao Tang, Jun Li, Kun Yuan, Li Wang, Xinyu Zhang, and Peng Chen. Bevheight: A robust framework for vision-based roadside 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21611-21620, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.247 + ], + "angle": 0, + "content": "[40] Xiaoqing Ye, Mao Shu, Hanyu Li, Yifeng Shi, Yingying Li, Guangjie Wang, Xiao Tan, and Errui Ding. Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21341-21350, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.894, + 0.33 + ], + "angle": 0, + "content": "[41] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21361-21370, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.332, + 0.893, + 0.387 + ], + "angle": 0, + "content": "[42] Haibao Yu, Yingjuan Tang, Enze Xie, Jilei Mao, Jirui Yuan, Ping Luo, and Zaiqing Nie. Vehicle-infrastructure cooperative 3d object detection via feature flow prediction. arXiv preprint arXiv:2303.10552, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.389, + 0.893, + 0.484 + ], + "angle": 0, + "content": "[43] Haibao Yu, Wenxian Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xu Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, et al. V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5486-5495, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.487, + 0.894, + 0.542 + ], + "angle": 0, + "content": "[44] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.543, + 0.893, + 0.611 + ], + "angle": 0, + "content": "[45] Walter Zimmer, Akshay Rangesh, and Mohan Trivedi. 3d bat: A semi-automatic, web-based 3d annotation toolbox for full-surface, multi-modal data streams. In 2019 IEEE Intelligent Vehicles Symposium (IV), pages 1816-1821. IEEE, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.613, + 0.893, + 0.668 + ], + "angle": 0, + "content": "[46] Walter Zimmer, Emec Ercelik, Xingcheng Zhou, Xavier Jair Diaz Ortiz, and Alois Knoll. A survey of robust 3d object detection methods in point clouds. arXiv preprint arXiv:2204.00106, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.67, + 0.893, + 0.739 + ], + "angle": 0, + "content": "[47] Walter Zimmer, Joseph Birkner, Marcel Brucker, Huu Tung Nguyen, Stefan Petrovski, Bohan Wang, and Alois C. Knoll. Infradet3d: Multi-modal 3d object detection based on roadside infrastructure camera and lidar sensors. In 2023 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2023. 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.741, + 0.893, + 0.809 + ], + "angle": 0, + "content": "[48] Walter Zimmer, Christian Creß, Huu Tung Nguyen, and Alois C Knoll. Tumtraf intersection dataset: All you need for urban 3d camera-lidar roadside perception [best student paper award]. In 2023 IEEE Intelligent Transportation Systems ITSC. IEEE, 2023. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.811, + 0.893, + 0.893 + ], + "angle": 0, + "content": "[49] Walter Zimmer, Jialong Wu, Xingcheng Zhou, and Alois C Knoll. Real-time and robust 3d object detection with roadside lidars. In Proceedings of the 12th International Scientific Conference on Mobility and Transport: Mobility Innovations for Growing Megacities, pages 199-219. Springer, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22677" + } + ] +] \ No newline at end of file diff --git a/2024/TUMTraf V2X Cooperative Perception Dataset/5f2a634b-45a2-4d92-81e8-cda538302a8f_origin.pdf b/2024/TUMTraf V2X Cooperative Perception Dataset/5f2a634b-45a2-4d92-81e8-cda538302a8f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..84be854a3a68510d5a434a4c67172bad308dc27f --- /dev/null +++ b/2024/TUMTraf V2X Cooperative Perception Dataset/5f2a634b-45a2-4d92-81e8-cda538302a8f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f988f0b591966e1c81a1449d4c43f76a1c34b510ad04b166b8947a653e3f260 +size 10137358 diff --git a/2024/TUMTraf V2X Cooperative Perception Dataset/full.md b/2024/TUMTraf V2X Cooperative Perception Dataset/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6998ef77705dca0750ece361226796f366feb2f1 --- /dev/null +++ b/2024/TUMTraf V2X Cooperative Perception Dataset/full.md @@ -0,0 +1,300 @@ +# TUMTraf V2X Cooperative Perception Dataset + +Walter Zimmer1 Gerhard Arya Wardana1 Suren Sritharan1 Xingcheng Zhou1 Rui Song1,2 Alois C. Knoll1 + +$^{1}$ Technical University of Munich $^{2}$ Fraunhofer IVI + +https://tum-traffic-dataset.github.io/tumtraf-v2x + +![](images/b6e4b17926ba46f919f4c3e6240cd0f154fa469192d50f9b40b3bb990aa8f6b0.jpg) +Figure 1. Visualization of 3D box labels and tracks in our TUMTraf V2X Cooperative Perception Dataset. The top part shows the labels projected into the four camera images. The part below shows a point cloud from two LiDARs with 3D box labels of the same scene. + +# Abstract + +Cooperative perception offers several benefits for enhancing the capabilities of autonomous vehicles and improving road safety. Using roadside sensors in addition to onboard sensors increases reliability and extends the sensor range. External sensors offer higher situational awareness for automated vehicles and prevent occlusions. We propose CoopDet3D, a cooperative multi-modal fusion model, and TUMTraf-V2X, a perception dataset, for the cooperative 3D object detection and tracking task. Our dataset contains 2,000 labeled point clouds and 5,000 labeled images from five roadside and four onboard sensors. It includes 30k 3D boxes with track IDs and precise GPS and IMU data. We labeled nine categories and covered occlusion scenarios with challenging driving maneuvers, like traffic violations, near-miss events, overtaking, and U-turns. Through multiple experiments, we show that our CoopDet3D camera-LiDAR fusion model achieves an increase of +14.36 3D mAP compared to a vehicle camera-LiDAR fusion model. Finally, we make our dataset, model, labeling tool, and devkit publicly available on our website. + +# 1. Introduction + +Cooperative perception involves the fusion of onboard sensor data and roadside sensor data, and it offers several advantages for enhancing the capabilities of autonomous vehicles and improving road safety. Using data from multiple sources makes the perception more robust to sensor failures or adverse environmental conditions. Roadside sensors provide an elevated view that helps to detect obstacles early. Moreover, they are also beneficial for precise vehicle localization and reduce the computational load of automated vehicles by offloading some perception tasks to the roadside sensors. Roadside sensors provide a global perspective of the traffic and offer a comprehensive situational awareness when fused with onboard sensor data. There are also fewer false positives or negatives because cooperative perception cross-validates the information from different sensors. + +Infrastructure sensors can share perception-related information with vehicles through V2X. Due to minimal delay, and real-time capabilities, the infrastructure-based perception systems can further enhance the situational awareness and decision-making processes of vehicles. + +Intelligent Transportation Systems (ITS) like the Testbed + +Table 1. Comparison of 3D cooperative V2X perception datasets with our proposed TUMTraf-V2X Cooperative Perception dataset (I=Infrastructure, V=Vehicle). + +
DatasetOPV2V [36]V2XSet [35]V2X-Sim [20]V2V4Real [37]DAIR-V2X-C [41]V2X-Seq (SPD) [43]TUMTraf-V2X (Ours)
Year2022202220222022202220232024
V2XV2VV2V&IV2V&IV2VV2IV2IV2I
Real data---✓✓✓✓
Annotation range120 m120 m70 m200 m280 m280 m200 m
Day & night scenes----✓✓✓
# object classes11151098
Track IDs--✓✓-✓✓
HD Maps✓✓✓✓-✓✓
# of sensors (I | V)- | 6*- | 6*5 | 7- | 8‡2 | 32 | 35 | 4
Available worldwide✓✓✓✓--✓
Traffic violations------✓
Labeled attributes§------✓
OpenLABEL format------✓
# Point Clouds11k11k10k20k39k15k2.0k
# Images44k44k60k40k†39k15k5.0k
# 3D Boxes233k233k26k240k464k10.45k29.38k
LocationCARLACARLACARLAUSAChinaChinaGermany
+ +$\dagger$ Image dataset has not been released yet. +* Value per vehicle. Multiple Conn. and Autom. Vehicles (CAVs) are used. + +Total sensors from 2 CAVs. + +$^{\S}$ Weather, time of day, orientation, #3D points, occlusion, color, sub type + +for Autonomous Driving [17] aim to improve safety by providing real-time traffic information. According to [9], testbeds extensively start using LiDAR sensors in their setups to create an accurate live digital twin of the traffic. Connected vehicles get a far-reaching view which enables them to react to breakdowns or accidents early. ITS systems also provide lane and speed recommendations to improve the traffic flow. + +The key challenge with ego-centric vehicle datasets is that there are many occlusions from a vehicle perspective, e.g., if a large truck in front of the ego vehicle obscures the view. Roadside sensors located at a smart intersection provide a broad overview of the intersection and a full-surround view. Given the immense potential of ITS, there is a specific need for V2X datasets. Despite the high costs associated with collecting and labeling such datasets, this work addresses this challenge as a crucial step toward realizing large-scale ITS implementations. + +# Our contributions are as follows: + +- We provide a high-quality V2X dataset for the cooperative 3D object detection and tracking task with 2,000 labeled point clouds and 5,000 labeled images. In total, 30k 3D bounding boxes with track IDs were labeled in challenging traffic scenarios like near-miss events, overtaking scenarios, U-turn maneuvers, and traffic violation events. +- We open-source our 3D bounding box annotation tool (3DBAT v24.3.2) to label multi-modal V2X datasets. +- We propose CoopDet3D, a cooperative 3D object detection model, and show in extensive experiments and abla + +tion studies that it outperforms single view models on our V2X dataset by $+14.3$ 3D mAP. + +- Finally, we provide a development kit to load the annotations in the widely recognized and standard format OpenLABEL [13], to facilitate a seamless integration and utilization of the dataset. Furthermore, it can preprocess, visualize, and convert labels to and from different dataset formats, and evaluate perception and tracking methods. + +# 2 . Related work + +3D autonomous driving datasets are mainly categorized based on the viewpoint. Table 1 highlights the main differences between our proposed dataset and other V2X datasets. + +# 2.1. Single viewpoint datasets + +Single viewpoint datasets are obtained from a single point of reference, either an ego-vehicle or roadside infrastructure. Onboard sensor-based datasets like KITTI [12], nuScenes [6], and Waymo [30] contain a diverse set of sensor data collected from a moving vehicle equipped with multiple sensors, including high-resolution cameras, LiDARs, radars, and GPS/INS systems. These datasets are abundant and provide many annotated data, including bounding boxes, track IDs, segmentation masks, and depth maps under different urban driving scenarios. + +On the other hand, roadside sensor-based datasets are in the infancy stage. High-quality multi-modal (camera and LiDAR) datasets are presented in [5, 10, 48], which + +![](images/cbd6b0078709864d3c66c2eb67fc84aca9bfe965f25c700d8998579e28bc9bd6.jpg) +Figure 2. Our CoopDet3D framework is a multi-modal cooperative fusion system, comprising three distinct fusion pipelines. 1) The roadside camera-LiDAR fusion pipeline fuses three camera images and one LiDAR point cloud by extracting features and transforming them into a BEV representation. 2) The vehicle camera-LiDAR fusion pipeline fuses the vehicle camera feature map with the vehicle point cloud feature map using a convolutional fuser. 3) The vehicle and infrastructure feature maps are then fused by applying an element-wise max-pooling operation (Max Fuser). In the end, we use the TransFusion [1] 3D detection head to obtain 3D bounding box predictions. + +are obtained from Infrastructure Perception Systems (IPS). Similarly, in [40], the authors provide a dataset consisting of only images taken from different viewpoints and under varying traffic conditions. These datasets provide a top-down view of a crowded intersection under different conditions and, as such, can overcome issues such as occlusions created by other vehicles and thereby have a higher number of object labels than onboard sensor-based datasets. + +# 2.2. V2X datasets + +V2X datasets exploit the information from multiple viewpoints to gain additional knowledge regarding the environments. In this way, they overcome the limitations of single viewpoint datasets such as occlusion, limited field of view (FOV), and low point cloud density. + +DAIR-V2X dataset family [41] is one of the foremost cooperative multi-modal datasets introduced. It contains three subsets: an intersection, a vehicle, and a cooperative dataset. The cooperative dataset contains 464k 3D box labels belonging to 10 classes, making it one of the largest cooperative datasets. The V2X-Seq dataset [43] extends selected sequences of the DAIR-V2X dataset with track IDs and is partitioned into a sequential perception dataset (SPD) and a trajectory forecasting dataset. Despite these, the lack of specific information, such as the labeling methodology used, the exact models of the sensors deployed, the distribution of the classes, and the scenarios within the dataset, leads to uncertainty in the extendability and application of this dataset in varying conditions. + +In V2V4Real [37], the authors propose a multi-modal + +cooperative dataset focusing only on V2V perception. Two vehicles equipped with cameras, LiDAR, and GPS/IMU integration systems are used to collect multi-modal sensor data for diverse scenarios. As opposed to all other V2X datasets, this focuses on V2V perception, and though it is of similar size to other cooperative datasets, it contains fewer classes and 3D bounding box information. + +Simulated multi-agent perception datasets have been proposed in [20, 35, 36]. These datasets contain multimodal sensor data (camera and LiDAR) obtained from roadside units (RSUs) and multiple ego vehicles, which enable collaborative perception. They use a combination of simulators such as SUMO [24], CARLA [11], and OpenCDA [36] for flow simulation, data retrieval, and V2X communication. However, the utility of the dataset is still limited due to the simulated nature of the data, and its extendability to real-life applications has not been studied in detail. + +# 2.3. V2X perception models for object detection + +The datasets presented above have been used to develop various models for a wide variety of tasks, with the majority focusing on 3D object detection. Different approaches have been taken depending on the availability and challenges, and these methods are grouped based on the number of nodes employed and the modalities used for detection. + +Most 3D object detection models use multi-modal sensor data obtained from a single point of view, which is often an ego-vehicle. Due to the popularity and abundant availability of onboard datasets [6, 12, 30], most models use images, point clouds, or both modalities. Image-based models were + +the pioneers in 3D object detection due to their low cost and simplicity, and both vehicular camera-based models [18, 34] and infrastructure camera-based models [39] have been proposed. LiDAR-based 3D object detection models [19, 46, 49] became popular since LiDAR point clouds provide 3D depth information and are robust, especially in adverse weather conditions and nighttime scenarios. Fusion models combine the information obtained from both images and point clouds and have been shown to outperform the prior methods [47]. Single viewpoint fusion models use either vehicular camera and LiDAR [23, 32, 38] or infrastructure camera and LiDAR [47] for 3D object detection. + +Cooperative perception models, which use data from multiple viewpoints, have been shown to overcome issues related to occlusion, which were often present in vehicular sensor-based models. V2I cooperative perception models [2, 3, 14, 26, 33, 35, 42] use the sensor data from both vehicles and infrastructure and V2V models [15, 29] communicate the sensor data between multiple vehicles. In this work, our cooperative multi-modal dataset is one contribution among others. Thus, while most of the prior works focus on unimodal cooperative perception using either Li-DAR point clouds [2, 7] or camera images [15], we benchmark our dataset with CoopDet3D, a deep fusion based cooperative multi-modal 3D object detection model based on BEVFusion [15] and PillarGrid [2]. + +# 3. TUMTraf-V2X Dataset + +Our TUMTraf V2X Cooperative Perception Dataset focuses on challenging traffic scenarios and various day and nighttime scenes. The data is further annotated, emphasizing high-quality labels through careful labeling and high-quality review processes. It also contains dense traffic and fast-moving vehicles, which reveals the specific challenges in cooperative perception, such as pose estimation errors, latency, and synchronization. Furthermore, we provide sensor data from nine different sensors covering the same traffic scenes under diverse weather conditions and lighting variations. The infrastructure sensors are oriented in all four directions of the intersection to get a $360^{\circ}$ view, which leads to better perception results. Finally, it contains rare events like traffic violations where pedestrians cross the road at a busy four-way intersection while the crossing light is lit red. + +# 3.1. Sensor setup + +Our TUMTraf V2X Cooperative Perception Dataset was recorded on an ITS system with nine sensors. + +The infrastructure sensor setup is the following: + +- 1x Ouster LiDAR OS1-64 (gen. 2), 64 vert. layers, $360^{\circ}$ FOV, below horizon config., $10\mathrm{cm}$ acc. @ $120\mathrm{m}$ range +- 4x Basler ace acA1920-50gc, $1920 \times 1200$ , Sony IMX174 with 8 mm lenses + +On the vehicle, the following sensors were used: + +![](images/5315b3b700c301da35141cb641a351776de308830170b77ca6aaa1787ab8ec9f.jpg) +Figure 3. Demonstration of a possible V2X occlusion scenario. A pedestrian (blue) is crossing the road in front of the ego vehicle. An occluded bicycle is marked in red. The recording vehicle with the sensor setup is shown in the bottom left corner. + +![](images/5604719f92ed31b04a8775f70728bc90eb7fb42a5f5a2670f9d35e12a682726a.jpg) + +- 1x Robosense RS-LiDAR-32, 32 vert. layers, $360^{\circ}$ FOV, 3 cm accuracy @200 m range +- 1x Basler ace acA1920-50gc, $1920 \times 1200$ , Sony IMX174 with $16 \mathrm{~mm}$ lens +- 1x Emlid Reach RS2+ multi-band RTK GNSS receiver +- 1x XSENS MTi-30-2A8G4 IMU + +# 3.2. Sensor calibration and registration + +We synchronize the cameras and LiDARs in the spatial and temporal domain. First, we determine the intrinsic camera parameters and the radial and tangential image distortions by using a checkerboard target. We then calibrate the roadside LiDAR with the roadside cameras by picking 100-point pairs in the point cloud and camera image. Extrinsic parameters (rotation and translation) are calculated by minimizing the reprojection error of 2D-3D point correspondences [25]. We follow the same procedure for onboard camera-LiDAR calibration. Finally, we calibrate the onboard LiDAR to the roadside LiDAR. This spatial registration is done by first estimating a coarse transformation. We pick ten 3D point pairs in each point cloud and minimize their distance using the least squares method. Then, we apply the point-to-point Iterative Closest Point (ICP) algorithm [4] to get the fine transformation between the point clouds. + +We label the vehicle and infrastructure point clouds after registering them. The coarse registration was done by measuring the GPS position of the onboard LiDAR and the roadside LiDAR. Then, we transform every 10th onboard point cloud to the coordinate system of the infrastructure point cloud. The fine registration was done by applying the point-to-point ICP to get an accurate V2I transformation matrix. All rotations of the point cloud frames in between are interpolated based on the spherical linear interpolation (SLERP) [28] method: + +$$ +S L E R P \left(q _ {0}, q _ {1}, t\right) = q _ {0} \left(q _ {0} ^ {- 1} q _ {1}\right) ^ {t}, \tag {1} +$$ + +where $q_{0}$ and $q_{1}$ are the quaternions representing the ro + +Table 2. Evaluation results ( $mAP_{BEV}$ and $mAP_{3D}$ ) of CoopDet3D on our TUMTraf-V2X test set in south2 FOV. + +
Config.\( \mathbf{mAP_{BEV}}\uparrow \)\( \mathbf{mAP_{3D}}\uparrow \)
DomainModalityEasy↑Mod.↑Hard↑Avg.↑
VehicleCamera46.8331.4737.8230.7730.36
VehicleLiDAR85.3385.2276.8669.0480.11
VehicleCam+LiDAR84.9077.6072.0873.1276.40
Infra.Camera61.9831.1946.7340.4235.04
Infra.LiDAR92.8686.1788.0775.7384.88
Infra.Cam+LiDAR92.9287.9989.0981.6987.01
Coop.Camera68.9445.4142.7657.8345.74
Coop.LiDAR93.9392.6378.0673.9585.86
Coop.Cam+LiDAR94.2293.4288.1779.9490.76
+ +tations of the start and end frames and $t \in [0,1]$ . Translation vectors $\mathbf{T}_0$ and $\mathbf{T}_1$ were obtained using linear interpolation: + +$$ +\mathbf {T} (t) = \mathbf {T} _ {0} + t \left(\mathbf {T} _ {1} - \mathbf {T} _ {0}\right). \tag {2} +$$ + +This dual interpolation strategy ensures that the estimated transformations between the frames are smooth and geometrically accurate, thus adhering closely to the actual movements of the vehicle over time. + +# 3.3. Data selection and labeling + +We selected the data based on challenging traffic scenarios, like U-turns, tailgate events, and traffic violation maneuvers. Besides the high traffic density of 31 objects per frame, we selected frames with high-class coverage. We selected 700 frames during sunny daytime and 100 frames during cloudy nighttime for labeling. The camera and Li-DAR data were recorded into rosbag files at $15\mathrm{Hz}$ and $10\mathrm{Hz}$ , respectively. We extracted and synchronized the data based on ROS [27] timestamps and labeled it with our $3D$ BAT (v24.3.2) annotation tool1. We improved the 3D BAT [45] baseline labeling tool to label 3D objects faster and more precisely with a one-click annotation feature. The annotators were instructed to label traffic participants while examining the images. Objects are still labeled, even if they have no 3D points inside, but are visible in the images. Extremities (e.g., pedestrian limbs) are included in the bounding box, but side mirrors of vehicles aren't. If a pedestrian carries an object, that object is included in the bounding box. If two or more pedestrians are carrying an object, only the box of one will include the object. After labeling, each annotator checked the work of other annotators manually frame-by-frame. When errors were found, the original annotator was notified, and they fixed it. This helps ensure that the labels in our dataset are high quality. + +# 3.4. Data structure and format + +We record eight different scenes, each 10 sec. long, from vehicle and infrastructure perspectives using nine sensors + +Table 3. Evaluation results of infrastructure-only CoopDet3D vs. InfraDet3D [47] on TUMTraf Intersection test set [48]. South 1 and South 2 refer to sensors covering different FOVs. + +
ModelFOVModality\( \mathrm{mAP}_{3D} \uparrow \)
Easy↑Mod.↑Hard↑Avg.↑
InfraDet3Dsouth 1LiDAR75.8147.6642.1655.21
CoopDet3Dsouth 1LiDAR76.2448.2335.1969.47
InfraDet3Dsouth 2LiDAR38.9246.6043.8643.13
CoopDet3Dsouth 2LiDAR74.9755.5539.9669.94
InfraDet3Dsouth 1Cam+LiDAR67.0831.3835.1744.55
CoopDet3Dsouth 1Cam+LiDAR75.6845.6345.6366.75
InfraDet3Dsouth 2Cam+LiDAR58.3819.7333.0837.06
CoopDet3Dsouth 2Cam+LiDAR74.7353.4641.9666.89
+ +and split the data into a train (80%), val. (10%), and test (10%) set. We use stratified sampling to distribute all sets' object classes equally (see Fig. 6a). Labels are provided in the ASAM OpenLABEL [13] standard. + +# 3.5. Dataset development kit + +We provide a devkit to work with our dataset. In addition to generating the data statistics, it provides modules for multiclass stratified splitting (train/val/test), point cloud registration, loading annotations in OpenLABEL format, evaluation of detection and tracking results, pre-processing steps such as point cloud filtering, and post-processing such as bounding box filtering. The statistics (Fig. 4, 5, and 6) were created using our devkit. It also contains modules to convert the labels from OpenLABEL to KITTI or our custom nuScenes format with timestamps instead of tokens and vice versa. This devkit enables users to migrate their models and make them compatible with our dataset format. We release our devkit2 under the MIT license and the dataset under the Creative Commons (CC) BY-NC-ND 4.0 license. + +# 4. Benchmark + +We propose CoopDet3D, an extension of BEVFusion [23] and PillarGrid [2] for deep cooperative multi-modal 3D object detection and benchmark it on TUMTraf-V2X mini. + +# 4.1. Evaluation metrics + +The accuracy is measured in terms of the mean average precision (mAP). Two types of mAP measures are used: BEV mAP considers the BEV center distance, and the results are obtained using the same evaluation methodology used in BEVFusion, which in turn uses the evaluation protocol of nuScenes [6]. Similarly, the 3D mAP measure considers the intersection in 3D, and the results are obtained using the evaluation script of our TUM Traffic dataset devkit. The runtime is evaluated using frames per second (FPS) as the metric and the results were obtained by measuring the time + +![](images/d206a3a08e78e440783e2077d0f0221eee7bbbc5438da77ea290e84e50bcd957.jpg) +(a) Distribution of objects between day and night. + +![](images/85b64cb2a64f09feb2b1a6bd6d1bd9c48936e20c9660488e11997eb6f363d46e.jpg) +(b) Avg. and max. num. of 3D points for each class. + +![](images/0f371d6d49ce5a390108a7175a9efb49d4fd83275eff48587d85e00fcad5ab47.jpg) +(c) Avg. and max. track length for all classes. + +![](images/15971227ab788286c7454d3b127c8fa1518b558600cfe37ed1fa3718fd264e43.jpg) +(a) Visualization of rotations (yaw). + +![](images/3c7804d4733a5c3c0112e6f814c5fee905fa9b156935008a6c3541c55622b450.jpg) +Figure 4. Our TUMTraf-V2X dataset contains 30k 3D box labels in total and is balanced among nine different object classes. (a) Cars (14,633) and pedestrians (5,276) are highly represented in the dataset. (b) 3D box labels contain, on average 625 points inside which shows the density of the labeled objects. The BUS class has the highest point density. (c) All traffic participants are tracked for $39\mathrm{m}$ on average. Emergency vehicles have the highest average track length of $128\mathrm{m}$ , whereas the CAR class contains a max. track length of $174\mathrm{m}$ . +(b) 3D points grouped by distance. + +![](images/380024a0d7512fb60c794f84746c69a477e79068e6d333d946693dc62283c6b5.jpg) +(c) BEV visualization of tracks. + +![](images/705d9703c938963694243093783f6b3cce1b1911338913a1c662bdbd23c01774.jpg) +Figure 5. Our dataset was recorded at a crowded intersection with many left and right turns. (a) Most of the vehicles (7,559) are driving in the east direction (0 degrees). (b) 3D boxes were labeled up to $200\mathrm{m}$ range and are very dense between 10 and $60\mathrm{m}$ . (c) The visualization of BEV tracks shows where pedestrians and bicycles are crossing the road. +(a) Histogram of 3D points. + +![](images/f0d78baf01401ab6a2122e2b1a3a6e613b5ea443672373e8254ad13219b9d133.jpg) +(b) Histogram of 3D box labels. + +![](images/352a55c2997c4ffbacce65d64e73b9216319a1e203e4e1db84322437c40f1269.jpg) +(c) Histogram of track lengths. +Figure 6. (a) Most 3D box labels contain between 1 and 200 3D points inside, with an average of 434 3D points, excluding empty boxes. Objects that were close to both LiDAR sensors even contained up to 18k 3D points. (b) Frames contain between 15 and 45 traffic participants, with an average of 29. (c) Objects were tracked up to $170\mathrm{m}$ , and the average track length is $39.35\mathrm{m}$ . + +![](images/866fd966ccb2b5e0071274b97dad1103117962b5448090ff4ee0510542ea0a0d.jpg) +Figure 7. Qualitative results on the TUMTraf-V2X Cooperative Perception test set. The first row shows the inference results of the onboard (vehicle-only) camera-LiDAR fusion with 23 detected objects. In the second row, the results of the cooperative vehicle-infrastructure camera-LiDAR fusion are visualized. Here, all 25 traffic participants could be detected with the support of roadside sensors. + +Table 4. Ablation study on cooperative 3D object detection with 11 combinations of camera and LiDAR backbones. The best tradeoff between speed and accuracy is highlighted in gray. $\mathrm{ND} =$ non-deterministic, $\mathrm{TS} =$ TorchSparse, $\odot =$ retrained. VRAM is in GiB. + +
Backbone Configuration\( \mathbf{{mAP}}_{\mathbf{{BEV}}} \uparrow \)\( \mathbf{{FPS}} \uparrow \)\( \mathbf{{VRAM}} \downarrow \)
VoxelNet ND + SwinT93.476.306.69
VoxelNet ND + YOLOv8 s92.947.246.39
VoxelNet TS + SwinT93.518.844.61
VoxelNet TS + YOLOv8 s92.9410.664.28
VoxelNet TS + YOLOv8 s ○94.3110.664.28
PointPillars 512 + Swin T94.439.004.94
PointPillars 512 + YOLOv8 s94.2711.144.63
PointPillars 512 + YOLOv8 s ○94.2511.144.63
PointPillars 512_2x + Swin T92.799.064.94
PointPillars 512_2x + YOLOv8 s94.1611.204.63
PointPillars 512_2x + YOLOv8 s ○94.2211.204.63
+ +needed by the model to run one full inference, including data preprocessing and voxelization. The first five iterations are skipped as a warmup since they are usually considerably slower than the average. Finally, the complexity of the model is measured in terms of the maximal VRAM usage across all GPUs during training and testing. + +# 4.2. CoopDet3D model + +Our CoopDet3D uses a BEVFusion-based backbone for camera-LiDAR fusion on the vehicle and the infrastructure sides separately to obtain the vehicle and infrastructure features. The best backbone for image and point cloud feature extraction was chosen through multiple ablation studies. Then, inspired by the method proposed by PillarGrid [2], an element-wise max-pooling operation is proposed to fuse the resulting fused camera-LiDAR features of vehicle and infrastructure together. Finally, the detection head from BEVFusion is used for 3D detection from the fused feature. + +The architecture of CoopDet3D is shown in Fig. 2. + +First, we disable the camera feature extraction nodes and train the LiDAR-only model for 20 epochs. Then, we use pre-trained weights for the cooperative model and fine-tune the entire model for eight further epochs. Hyperparameter tuning revealed that the default hyperparameters of BEV-Fusion [23] gave the best results, and such were not modified. The preprocessing steps are also the same as for BEV-Fusion, but we change the point cloud range to $[-75, 75]$ in the x- and y-scale and $[-8, 0]$ in the z-scale since the dataset used in this case is different. Furthermore, we use 3x NVIDIA RTX 3090 GPUs with 24 GB VRAM for training and a single GPU for evaluation. We open-source our model and provide pre-trained weights3. + +# 4.3. Experiments and ablation studies + +The objective of these experiments is to highlight the importance of our V2X multi-viewpoint dataset as opposed to single-viewpoint datasets. As such, we conduct multiple experiments and ablation studies with data obtained from each viewpoint and compare the results on the proposed model. + +Ablation on viewpoint and modality. We conduct multiple experiments with all possible combinations of a) viewpoints: vehicle-only, infrastructure-only, cooperative, and b) modalities: camera-only, LiDAR-only, and camera-LiDAR fusion. Table 2 shows the mAP achieved by CoopDet3D for each of these combinations. + +We observe that the results follow a general pattern of cooperative performance being better than infrastructure-only, which is, in turn, better than vehicle-only. Furthermore, fusion models perform better than LiDAR-only models, which in turn are better than camera-only models. Figure 7 shows qualitative results between our vehicle-only + +camera-LiDAR fusion model and our cooperative vehicle-infrastructure camera-LiDAR fusion model. Again, we observe from these samples that the cooperative perception model is able to detect 25 traffic participants, whereas the vehicle-only model is only able to detect 23 objects due to occlusions and a limited field of view. + +Ablation on fusion level. We compare our CoopDet3D model to the current SOTA camera-LiDAR fusion method InfraDet3D [47] on the TUMTraf Intersection test set [48]. The proposed method uses deep fusion, whereas the InfraDet3D method is a late fusion method. Table 3 shows the performance of our model against InfraDet3D. The results show that the proposed deep fusion method outperforms the SOTA late fusion model in all metrics, except in the hard difficulty in LiDAR-only mode. Furthermore, Fig. 8 shows two sample images taken during day and nighttime, wherein deep fusion again outperforms late fusion. + +We note that these experiments were conducted in an offline setting, disregarding other considerations for simplicity. However, when deploying it in real life, factors such as the transmission bandwidth should also be considered. Since we observed that deep feature fusion generally leads to higher efficacy, the V2I transmissions should contain these features instead of infrastructure bounding boxes. + +Ablation on backbones. As an ablation study, we present the results of the experiments to find the best backbone and model configuration for the cooperative camera-LiDAR fusion model. For the camera backbone, SwinT [22] and MMYOLO's [8] implementation of YOLOv8 [16] were considered. For the LiDAR backbone, VoxelNet [44] and PointPillars [19] were considered. In addition, VoxelNet was implemented with two different backends, namely SPConv v2 and Torchsparse [31]. For PointPillars, two grid sizes are considered $512 \times 512$ for both train and test grids (PointPillars 512) and $512 \times 512$ train grid with $1024 \times 1024$ test grid (PointPillars 512_2x). The results of these experiments are shown in Table 4. + +The results show that only models that use any combination of VoxelNet Torchsparse, both PointPillars variants, and YOLOv8 are able to run above 10 FPS. From these configurations, we choose PointPillars 512_2x with YOLOv8 as the best configuration for all the above experiments as it achieves the best results across all the ablation studies. This is a promising result since we also know that this backbone configuration is able to run in real-time (11.2 FPS) on an RTX 3090 without using TensorRT acceleration. + +An interesting observation is that utilizing pre-trained weights for transfer learning of YOLOv8 is not always beneficial, as the results from PointPillars $512 + \mathrm{YOLOv8}$ show. This is likely because the pre-trained weights were from MS COCO [21], and they have a very different data domain compared to our dataset. Since MS COCO is also + +![](images/193a63668ff5d7a771663d6ae7c6110ce8a11f5a42eea5055b414f37ea7a4100.jpg) +Figure 8. Qualitative results of our CoopDet3D (left) and the InfraDet3D (right) model on the TUMTraf Intersection test set during day and nighttime. Detected objects marked with a red circle were classified correctly by CoopDet3D. + +![](images/6ec51ed62d9f340fe69b01b55056f1cbd7bd0fdcfe8e391cf47aed40d835af47.jpg) + +much larger than our dataset in terms of camera images, retraining harms the performance of the model slightly. + +In terms of efficiency, the goal of these experiments was to verify that the proposed CoopDet3D model with the best configuration provides the highest accuracy while also being able to run in real-time (minimum of $10\mathrm{Hz}$ ). Furthermore, it should also be feasible to train the model on a high-performance GPU and perform inference on a mid-range consumer GPU deployable on an edge device. The results concerning the VRAM usage during inference show that the complexity of the model makes this feasible. + +# 5. Conclusion and future work + +This work proposes the TUMTraf-V2X dataset, a multimodal multi-view V2X dataset for cooperative 3D object detection and tracking. Our dataset focuses on challenging traffic scenarios at an intersection and provides views from the infrastructure and the ego vehicle. To benchmark the dataset, we propose CoopDet3D - a baseline model for cooperative perception. Experiments show that cooperative fusion leads to higher efficacy than its unimodal and single-view camera-LiDAR fusion counterparts. Furthermore, cooperative fusion leads to an improvement of +14.3 3D mAP compared to vehicle-only perception, highlighting the need for V2X datasets. Finally, we provide our 3D BAT v24.3.2 labeling tool and devkit to load, parse, and visualize the dataset. It also includes modules for pre- and postprocessing and evaluation. Future efforts will integrate this platform into online environments, enabling a broader range of infrastructure-based, real-time perception applications. + +# Acknowledgment + +This research was supported by the Federal Ministry of Education and Research in Germany within the AUTOtech.agil project, Grant Number: 01IS22088U. + +# References + +[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers, 2022. 3 +[2] Zhengwei Bai, Guoyuan Wu, Matthew J Barth, Yongkang Liu, Emrah Akin Sisbot, and Kentaro Oguchi. Pillargrid: Deep learning-based cooperative perception for 3d object detection from onboard-roadside lidar. In IEEE 25th International Conference on Intelligent Transportation Systems (ITSC), pages 1743-1749. IEEE, 2022. 4, 5, 7 +[3] Zhengwei Bai, Guoyuan Wu, Matthew J Barth, Yongkang Liu, Emrah Akin Sisbot, and Kentaro Oguchi. Vinet: Lightweight, scalable, and heterogeneous cooperative perception for 3d object detection. Mechanical Systems and Signal Processing, 204:110723, 2023. 4 +[4] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, pages 586-606. Spie, 1992. 4 +[5] Steffen Busch, Christian Koetsier, Jeldrik Axmann, and Claus Brenner. Lumpi: The leibniz university multiperspective intersection dataset. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 1127-1134. IEEE, 2022. 2 +[6] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11621-11631, 2020. 2, 3, 5 +[7] Ziming Chen, Yifeng Shi, and Jinrang Jia. Transiff: An instance-level feature fusion framework for vehicleinfrastructure cooperative 3d detection with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18205-18214, 2023. 4 +[8] MMYOLO Contributors. MMYOLO: OpenMMLab YOLO series toolbox and benchmark. https://github.com/open-mmlab/mmyolo, 2022.8 +[9] Christian Creß, Zhenshan Bing, and Alois C. Knoll. Intelligent transportation systems using roadside infrastructure: A literature survey. IEEE Transactions on Intelligent Transportation Systems, pages 1-0, 2023. 2 +[10] Christian Creß, Walter Zimmer, Leah Strand, Maximilian Fortkord, Siyi Dai, Venkatnarayanan Lakshminarasimhan, and Alois Knoll. A9-dataset: Multi-sensor infrastructure-based dataset for mobility research. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 965-970, 2022. 2 +[11] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In Proceedings of the 1st Annual Conference on Robot Learning, pages 1-16, 2017. 3 +[12] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 3 +[13] Nicco Hagedorn. OpenLABEL Concept Paper. 2, 5 +[14] Yuze He, Li Ma, Zhehao Jiang, Yi Tang, and Guoliang Xing. Vi-eye: Semantic-based 3d point cloud registration + +for infrastructure-assisted autonomous driving. In Proceedings of the 27th Annual International Conference on Mobile Computing and Networking, pages 573-586, 2021. 4 +[15] Yue Hu, Yifan Lu, Runsheng Xu, Weidi Xie, Siheng Chen, and Yanfeng Wang. Collaboration helps camera overtake lidar in 3d detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243-9252, 2023. 4 +[16] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolo, 2023. 8 +[17] Annkathrin Krammer, Christoph Schöller, Dhiraj Gulati, Venkat Narayanan Lakshminarasimhan, Franz Kurz, Dominik Rosenbaum, Claus Lenz, and Alois Knoll. Providentia-a large-scale sensor system for the assistance of autonomous vehicles and its evaluation. Journal of Field Robotics, pages 1156-1176, 2022. 2 +[18] Abhinav Kumar, Garrick Brazil, Enrique Corona, Armin Parchami, and Xiaoming Liu. Deviant: Depth equivariant network for monocular 3d object detection. In European Conference on Computer Vision, pages 664-683, 2022. 4 +[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 4, 8 +[20] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 7(4):10914-10921, 2022. 2, 3 +[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of European Conference on Computer Vision, ECCV 2014, Part V 13, pages 740-755. Springer, 2014. 8 +[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 8 +[23] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE Int. Conference on Robotics and Automation (ICRA), pages 2774-2781, 2023. 4, 5, 7 +[24] Pablo Alvarez Lopez, Michael Behrisch, Laura Bieker-Walz, Jakob Erdmann, Yun-Pang Flotteröd, Robert Hilbrich, Leonhard Lücken, Johannes Rummel, Peter Wagner, and Evamarie Wießner. Microscopic traffic simulation using sumo. In The 21st IEEE International Conference on Intelligent Transportation Systems. IEEE, 2018. 3 +[25] Kaj Madsen, Hans Bruun Nielsen, and Ole Tingleff. Methods for non-linear least squares problems. 2004. 4 +[26] Donghao Qiao and Farhana Zulkernine. Cobevfusion: Cooperative perception with lidar-camera bird's-eye view fusion. arXiv preprint arXiv:2310.06008, 2023. 4 +[27] Morgan Quigley, Ken Conley, Brian Gerkey, Josh Faust, Tully Foote, Jeremy Leibs, Rob Wheeler, Andrew Y Ng, + +et al. Ros: an open-source robot operating system. In ICRA workshop on open source software, page 5. Kobe, Japan, 2009. 5 +[28] Ken Shoemake. Animating rotation with quaternion curves. In Proceedings of the 12th annual conference on Computer graphics and interactive techniques, pages 245-254, 1985. 4 +[29] Rui Song, Chenwei Liang, Hu Cao, Zhiran Yan, Walter Zimmer, Markus Gross, Andreas Festag, and Alois Knoll. Collaborative semantic occupancy prediction with hybrid feature fusion in connected automated vehicles. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 4 +[30] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 2, 3 +[31] Haotian Tang, Shang Yang, Zhijian Liu, Ke Hong, Zhong-ming Yu, Xiuyu Li, Guohao Dai, Yu Wang, and Song Han. TorchSparse++: Efficient Point Cloud Engine. In Computer Vision and Pattern Recognition Workshops CVPRW, 2023. 8 +[32] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4604-4612, 2020. 4 +[33] Sizhe Wei, Yuxi Wei, Yue Hu, Yifan Lu, Yiqi Zhong, Siheng Chen, and Ya Zhang. Asynchrony-robust collaborative perception via bird's eye view flow. In Advances in Neural Information Processing Systems, pages 28462–28477. Curran Associates, Inc. 4 +[34] Zizhang Wu, Yuanzhu Gan, Lei Wang, Guilian Chen, and Jian Pu. Monopgc: Monocular 3d object detection with pixel geometry contexts. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4842-4849, 2023. 4 +[35] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 2, 3, 4 +[36] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 3 +[37] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 2, 3 +[38] Junjie Yan, Yingfei Liu, Jianjian Sun, Fan Jia, Shuailin Li, Tiancai Wang, and Xiangyu Zhang. Cross modal transformer: Towards fast and robust 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18268-18278, 2023. 4 + +[39] Lei Yang, Kaicheng Yu, Tao Tang, Jun Li, Kun Yuan, Li Wang, Xinyu Zhang, and Peng Chen. Bevheight: A robust framework for vision-based roadside 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21611-21620, 2023. 4 +[40] Xiaoqing Ye, Mao Shu, Hanyu Li, Yifeng Shi, Yingying Li, Guangjie Wang, Xiao Tan, and Errui Ding. Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21341-21350, 2022. 3 +[41] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21361-21370, 2022. 2, 3 +[42] Haibao Yu, Yingjuan Tang, Enze Xie, Jilei Mao, Jirui Yuan, Ping Luo, and Zaiqing Nie. Vehicle-infrastructure cooperative 3d object detection via feature flow prediction. arXiv preprint arXiv:2303.10552, 2023. 4 +[43] Haibao Yu, Wenxian Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xu Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, et al. V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5486-5495, 2023. 2, 3 +[44] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 8 +[45] Walter Zimmer, Akshay Rangesh, and Mohan Trivedi. 3d bat: A semi-automatic, web-based 3d annotation toolbox for full-surface, multi-modal data streams. In 2019 IEEE Intelligent Vehicles Symposium (IV), pages 1816-1821. IEEE, 2019. 5 +[46] Walter Zimmer, Emec Ercelik, Xingcheng Zhou, Xavier Jair Diaz Ortiz, and Alois Knoll. A survey of robust 3d object detection methods in point clouds. arXiv preprint arXiv:2204.00106, 2022. 4 +[47] Walter Zimmer, Joseph Birkner, Marcel Brucker, Huu Tung Nguyen, Stefan Petrovski, Bohan Wang, and Alois C. Knoll. Infradet3d: Multi-modal 3d object detection based on roadside infrastructure camera and lidar sensors. In 2023 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2023. 4, 5, 8 +[48] Walter Zimmer, Christian Creß, Huu Tung Nguyen, and Alois C Knoll. Tumtraf intersection dataset: All you need for urban 3d camera-lidar roadside perception [best student paper award]. In 2023 IEEE Intelligent Transportation Systems ITSC. IEEE, 2023. 2, 5, 8 +[49] Walter Zimmer, Jialong Wu, Xingcheng Zhou, and Alois C Knoll. Real-time and robust 3d object detection with roadside lidars. In Proceedings of the 12th International Scientific Conference on Mobility and Transport: Mobility Innovations for Growing Megacities, pages 199-219. Springer, 2023. 4 \ No newline at end of file diff --git a/2024/TUMTraf V2X Cooperative Perception Dataset/images.zip b/2024/TUMTraf V2X Cooperative Perception Dataset/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5cfb59d90be64ed69d25a8df15a099c8fc51f4f1 --- /dev/null +++ b/2024/TUMTraf V2X Cooperative Perception Dataset/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73165b7d1e249fddb45d39bba887207a591bddcde20430a602bf1162142d2369 +size 973019 diff --git a/2024/TUMTraf V2X Cooperative Perception Dataset/layout.json b/2024/TUMTraf V2X Cooperative Perception Dataset/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e20169785c95a7bb22034dc7f68b7c791a800417 --- /dev/null +++ b/2024/TUMTraf V2X Cooperative Perception Dataset/layout.json @@ -0,0 +1,7797 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 151, + 103, + 444, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 103, + 444, + 121 + ], + "spans": [ + { + "bbox": [ + 151, + 103, + 444, + 121 + ], + "type": "text", + "content": "TUMTraf V2X Cooperative Perception Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 148, + 493, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 148, + 493, + 179 + ], + "spans": [ + { + "bbox": [ + 99, + 148, + 493, + 179 + ], + "type": "text", + "content": "Walter Zimmer1 Gerhard Arya Wardana1 Suren Sritharan1 Xingcheng Zhou1 Rui Song1,2 Alois C. Knoll1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 170, + 193, + 423, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 193, + 423, + 209 + ], + "spans": [ + { + "bbox": [ + 170, + 193, + 423, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 170, + 193, + 423, + 209 + ], + "type": "text", + "content": "Technical University of Munich " + }, + { + "bbox": [ + 170, + 193, + 423, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 170, + 193, + 423, + 209 + ], + "type": "text", + "content": "Fraunhofer IVI" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 222, + 473, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 222, + 473, + 236 + ], + "spans": [ + { + "bbox": [ + 116, + 222, + 473, + 236 + ], + "type": "text", + "content": "https://tum-traffic-dataset.github.io/tumtraf-v2x" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 55, + 238, + 537, + 399 + ], + "blocks": [ + { + "bbox": [ + 55, + 238, + 537, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 238, + 537, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 537, + 399 + ], + "type": "image", + "image_path": "b6e4b17926ba46f919f4c3e6240cd0f154fa469192d50f9b40b3bb990aa8f6b0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 403, + 546, + 426 + ], + "lines": [ + { + "bbox": [ + 46, + 403, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 546, + 426 + ], + "type": "text", + "content": "Figure 1. Visualization of 3D box labels and tracks in our TUMTraf V2X Cooperative Perception Dataset. The top part shows the labels projected into the four camera images. The part below shows a point cloud from two LiDARs with 3D box labels of the same scene." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 435, + 192, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 435, + 192, + 448 + ], + "spans": [ + { + "bbox": [ + 143, + 435, + 192, + 448 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 462, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 462, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 45, + 462, + 289, + 700 + ], + "type": "text", + "content": "Cooperative perception offers several benefits for enhancing the capabilities of autonomous vehicles and improving road safety. Using roadside sensors in addition to onboard sensors increases reliability and extends the sensor range. External sensors offer higher situational awareness for automated vehicles and prevent occlusions. We propose CoopDet3D, a cooperative multi-modal fusion model, and TUMTraf-V2X, a perception dataset, for the cooperative 3D object detection and tracking task. Our dataset contains 2,000 labeled point clouds and 5,000 labeled images from five roadside and four onboard sensors. It includes 30k 3D boxes with track IDs and precise GPS and IMU data. We labeled nine categories and covered occlusion scenarios with challenging driving maneuvers, like traffic violations, near-miss events, overtaking, and U-turns. Through multiple experiments, we show that our CoopDet3D camera-LiDAR fusion model achieves an increase of +14.36 3D mAP compared to a vehicle camera-LiDAR fusion model. Finally, we make our dataset, model, labeling tool, and devkit publicly available on our website." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 435, + 390, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 435, + 390, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 435, + 390, + 448 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 457, + 545, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 636 + ], + "type": "text", + "content": "Cooperative perception involves the fusion of onboard sensor data and roadside sensor data, and it offers several advantages for enhancing the capabilities of autonomous vehicles and improving road safety. Using data from multiple sources makes the perception more robust to sensor failures or adverse environmental conditions. Roadside sensors provide an elevated view that helps to detect obstacles early. Moreover, they are also beneficial for precise vehicle localization and reduce the computational load of automated vehicles by offloading some perception tasks to the roadside sensors. Roadside sensors provide a global perspective of the traffic and offer a comprehensive situational awareness when fused with onboard sensor data. There are also fewer false positives or negatives because cooperative perception cross-validates the information from different sensors." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 639, + 545, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 699 + ], + "type": "text", + "content": "Infrastructure sensors can share perception-related information with vehicles through V2X. Due to minimal delay, and real-time capabilities, the infrastructure-based perception systems can further enhance the situational awareness and decision-making processes of vehicles." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Intelligent Transportation Systems (ITS) like the Testbed" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22668" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 100, + 531, + 326 + ], + "blocks": [ + { + "bbox": [ + 46, + 69, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 46, + 69, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 46, + 69, + 547, + 94 + ], + "type": "text", + "content": "Table 1. Comparison of 3D cooperative V2X perception datasets with our proposed TUMTraf-V2X Cooperative Perception dataset (I=Infrastructure, V=Vehicle)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 100, + 531, + 326 + ], + "lines": [ + { + "bbox": [ + 62, + 100, + 531, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 100, + 531, + 326 + ], + "type": "table", + "html": "
DatasetOPV2V [36]V2XSet [35]V2X-Sim [20]V2V4Real [37]DAIR-V2X-C [41]V2X-Seq (SPD) [43]TUMTraf-V2X (Ours)
Year2022202220222022202220232024
V2XV2VV2V&IV2V&IV2VV2IV2IV2I
Real data---✓✓✓✓
Annotation range120 m120 m70 m200 m280 m280 m200 m
Day & night scenes----✓✓✓
# object classes11151098
Track IDs--✓✓-✓✓
HD Maps✓✓✓✓-✓✓
# of sensors (I | V)- | 6*- | 6*5 | 7- | 8‡2 | 32 | 35 | 4
Available worldwide✓✓✓✓--✓
Traffic violations------✓
Labeled attributes§------✓
OpenLABEL format------✓
# Point Clouds11k11k10k20k39k15k2.0k
# Images44k44k60k40k†39k15k5.0k
# 3D Boxes233k233k26k240k464k10.45k29.38k
LocationCARLACARLACARLAUSAChinaChinaGermany
", + "image_path": "02f1364aace40a772efc1ca2d10d7f92749a31d5b50eaeb9e952309babba30fe.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 56, + 327, + 193, + 338 + ], + "lines": [ + { + "bbox": [ + 56, + 327, + 193, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 193, + 338 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 56, + 327, + 193, + 338 + ], + "type": "text", + "content": " Image dataset has not been released yet." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 56, + 338, + 299, + 348 + ], + "lines": [ + { + "bbox": [ + 56, + 338, + 299, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 299, + 348 + ], + "type": "text", + "content": "* Value per vehicle. Multiple Conn. and Autom. Vehicles (CAVs) are used." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 302, + 327, + 399, + 338 + ], + "lines": [ + { + "bbox": [ + 302, + 327, + 399, + 338 + ], + "spans": [ + { + "bbox": [ + 302, + 327, + 399, + 338 + ], + "type": "text", + "content": "Total sensors from 2 CAVs." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 338, + 539, + 348 + ], + "lines": [ + { + "bbox": [ + 302, + 338, + 539, + 348 + ], + "spans": [ + { + "bbox": [ + 302, + 338, + 539, + 348 + ], + "type": "inline_equation", + "content": "^{\\S}" + }, + { + "bbox": [ + 302, + 338, + 539, + 348 + ], + "type": "text", + "content": " Weather, time of day, orientation, #3D points, occlusion, color, sub type" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 366, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 461 + ], + "type": "text", + "content": "for Autonomous Driving [17] aim to improve safety by providing real-time traffic information. According to [9], testbeds extensively start using LiDAR sensors in their setups to create an accurate live digital twin of the traffic. Connected vehicles get a far-reaching view which enables them to react to breakdowns or accidents early. ITS systems also provide lane and speed recommendations to improve the traffic flow." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 462, + 288, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 581 + ], + "type": "text", + "content": "The key challenge with ego-centric vehicle datasets is that there are many occlusions from a vehicle perspective, e.g., if a large truck in front of the ego vehicle obscures the view. Roadside sensors located at a smart intersection provide a broad overview of the intersection and a full-surround view. Given the immense potential of ITS, there is a specific need for V2X datasets. Despite the high costs associated with collecting and labeling such datasets, this work addresses this challenge as a crucial step toward realizing large-scale ITS implementations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 582, + 190, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 190, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 190, + 593 + ], + "type": "text", + "content": "Our contributions are as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 47, + 594, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 287, + 665 + ], + "type": "text", + "content": "- We provide a high-quality V2X dataset for the cooperative 3D object detection and tracking task with 2,000 labeled point clouds and 5,000 labeled images. In total, 30k 3D bounding boxes with track IDs were labeled in challenging traffic scenarios like near-miss events, overtaking scenarios, U-turn maneuvers, and traffic violation events." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 665, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 689 + ], + "type": "text", + "content": "- We open-source our 3D bounding box annotation tool (3DBAT v24.3.2) to label multi-modal V2X datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "- We propose CoopDet3D, a cooperative 3D object detection model, and show in extensive experiments and abla" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 312, + 366, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 366, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 312, + 366, + 545, + 390 + ], + "type": "text", + "content": "tion studies that it outperforms single view models on our V2X dataset by " + }, + { + "bbox": [ + 312, + 366, + 545, + 390 + ], + "type": "inline_equation", + "content": "+14.3" + }, + { + "bbox": [ + 312, + 366, + 545, + 390 + ], + "type": "text", + "content": " 3D mAP." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 391, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 391, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 391, + 545, + 462 + ], + "type": "text", + "content": "- Finally, we provide a development kit to load the annotations in the widely recognized and standard format OpenLABEL [13], to facilitate a seamless integration and utilization of the dataset. Furthermore, it can preprocess, visualize, and convert labels to and from different dataset formats, and evaluate perception and tracking methods." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 474, + 394, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 474, + 394, + 487 + ], + "spans": [ + { + "bbox": [ + 306, + 474, + 394, + 487 + ], + "type": "text", + "content": "2 . Related work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 495, + 547, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 532 + ], + "type": "text", + "content": "3D autonomous driving datasets are mainly categorized based on the viewpoint. Table 1 highlights the main differences between our proposed dataset and other V2X datasets." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 539, + 448, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 448, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 448, + 552 + ], + "type": "text", + "content": "2.1. Single viewpoint datasets" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 557, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 676 + ], + "type": "text", + "content": "Single viewpoint datasets are obtained from a single point of reference, either an ego-vehicle or roadside infrastructure. Onboard sensor-based datasets like KITTI [12], nuScenes [6], and Waymo [30] contain a diverse set of sensor data collected from a moving vehicle equipped with multiple sensors, including high-resolution cameras, LiDARs, radars, and GPS/INS systems. These datasets are abundant and provide many annotated data, including bounding boxes, track IDs, segmentation masks, and depth maps under different urban driving scenarios." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "On the other hand, roadside sensor-based datasets are in the infancy stage. High-quality multi-modal (camera and LiDAR) datasets are presented in [5, 10, 48], which" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22669" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 545, + 261 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 545, + 261 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 545, + 261 + ], + "type": "image", + "image_path": "cbd6b0078709864d3c66c2eb67fc84aca9bfe965f25c700d8998579e28bc9bd6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 271, + 548, + 327 + ], + "lines": [ + { + "bbox": [ + 46, + 271, + 548, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 548, + 327 + ], + "type": "text", + "content": "Figure 2. Our CoopDet3D framework is a multi-modal cooperative fusion system, comprising three distinct fusion pipelines. 1) The roadside camera-LiDAR fusion pipeline fuses three camera images and one LiDAR point cloud by extracting features and transforming them into a BEV representation. 2) The vehicle camera-LiDAR fusion pipeline fuses the vehicle camera feature map with the vehicle point cloud feature map using a convolutional fuser. 3) The vehicle and infrastructure feature maps are then fused by applying an element-wise max-pooling operation (Max Fuser). In the end, we use the TransFusion [1] 3D detection head to obtain 3D bounding box predictions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 346, + 289, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 289, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 289, + 441 + ], + "type": "text", + "content": "are obtained from Infrastructure Perception Systems (IPS). Similarly, in [40], the authors provide a dataset consisting of only images taken from different viewpoints and under varying traffic conditions. These datasets provide a top-down view of a crowded intersection under different conditions and, as such, can overcome issues such as occlusions created by other vehicles and thereby have a higher number of object labels than onboard sensor-based datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 453, + 134, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 134, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 134, + 464 + ], + "type": "text", + "content": "2.2. V2X datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 472, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 533 + ], + "type": "text", + "content": "V2X datasets exploit the information from multiple viewpoints to gain additional knowledge regarding the environments. In this way, they overcome the limitations of single viewpoint datasets such as occlusion, limited field of view (FOV), and low point cloud density." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 700 + ], + "type": "text", + "content": "DAIR-V2X dataset family [41] is one of the foremost cooperative multi-modal datasets introduced. It contains three subsets: an intersection, a vehicle, and a cooperative dataset. The cooperative dataset contains 464k 3D box labels belonging to 10 classes, making it one of the largest cooperative datasets. The V2X-Seq dataset [43] extends selected sequences of the DAIR-V2X dataset with track IDs and is partitioned into a sequential perception dataset (SPD) and a trajectory forecasting dataset. Despite these, the lack of specific information, such as the labeling methodology used, the exact models of the sensors deployed, the distribution of the classes, and the scenarios within the dataset, leads to uncertainty in the extendability and application of this dataset in varying conditions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "In V2V4Real [37], the authors propose a multi-modal" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 346, + 547, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 547, + 430 + ], + "type": "text", + "content": "cooperative dataset focusing only on V2V perception. Two vehicles equipped with cameras, LiDAR, and GPS/IMU integration systems are used to collect multi-modal sensor data for diverse scenarios. As opposed to all other V2X datasets, this focuses on V2V perception, and though it is of similar size to other cooperative datasets, it contains fewer classes and 3D bounding box information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 431, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 547, + 552 + ], + "type": "text", + "content": "Simulated multi-agent perception datasets have been proposed in [20, 35, 36]. These datasets contain multimodal sensor data (camera and LiDAR) obtained from roadside units (RSUs) and multiple ego vehicles, which enable collaborative perception. They use a combination of simulators such as SUMO [24], CARLA [11], and OpenCDA [36] for flow simulation, data retrieval, and V2X communication. However, the utility of the dataset is still limited due to the simulated nature of the data, and its extendability to real-life applications has not been studied in detail." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 561, + 533, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 561, + 533, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 561, + 533, + 574 + ], + "type": "text", + "content": "2.3. V2X perception models for object detection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 580, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 652 + ], + "type": "text", + "content": "The datasets presented above have been used to develop various models for a wide variety of tasks, with the majority focusing on 3D object detection. Different approaches have been taken depending on the availability and challenges, and these methods are grouped based on the number of nodes employed and the modalities used for detection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Most 3D object detection models use multi-modal sensor data obtained from a single point of view, which is often an ego-vehicle. Due to the popularity and abundant availability of onboard datasets [6, 12, 30], most models use images, point clouds, or both modalities. Image-based models were" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22670" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "the pioneers in 3D object detection due to their low cost and simplicity, and both vehicular camera-based models [18, 34] and infrastructure camera-based models [39] have been proposed. LiDAR-based 3D object detection models [19, 46, 49] became popular since LiDAR point clouds provide 3D depth information and are robust, especially in adverse weather conditions and nighttime scenarios. Fusion models combine the information obtained from both images and point clouds and have been shown to outperform the prior methods [47]. Single viewpoint fusion models use either vehicular camera and LiDAR [23, 32, 38] or infrastructure camera and LiDAR [47] for 3D object detection." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "text", + "content": "Cooperative perception models, which use data from multiple viewpoints, have been shown to overcome issues related to occlusion, which were often present in vehicular sensor-based models. V2I cooperative perception models [2, 3, 14, 26, 33, 35, 42] use the sensor data from both vehicles and infrastructure and V2V models [15, 29] communicate the sensor data between multiple vehicles. In this work, our cooperative multi-modal dataset is one contribution among others. Thus, while most of the prior works focus on unimodal cooperative perception using either Li-DAR point clouds [2, 7] or camera images [15], we benchmark our dataset with CoopDet3D, a deep fusion based cooperative multi-modal 3D object detection model based on BEVFusion [15] and PillarGrid [2]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 393, + 186, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 186, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 186, + 406 + ], + "type": "text", + "content": "3. TUMTraf-V2X Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 413, + 289, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 289, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 289, + 594 + ], + "type": "text", + "content": "Our TUMTraf V2X Cooperative Perception Dataset focuses on challenging traffic scenarios and various day and nighttime scenes. The data is further annotated, emphasizing high-quality labels through careful labeling and high-quality review processes. It also contains dense traffic and fast-moving vehicles, which reveals the specific challenges in cooperative perception, such as pose estimation errors, latency, and synchronization. Furthermore, we provide sensor data from nine different sensors covering the same traffic scenes under diverse weather conditions and lighting variations. The infrastructure sensors are oriented in all four directions of the intersection to get a " + }, + { + "bbox": [ + 46, + 413, + 289, + 594 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 413, + 289, + 594 + ], + "type": "text", + "content": " view, which leads to better perception results. Finally, it contains rare events like traffic violations where pedestrians cross the road at a busy four-way intersection while the crossing light is lit red." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 600, + 130, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 600, + 130, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 600, + 130, + 612 + ], + "type": "text", + "content": "3.1. Sensor setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 287, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 642 + ], + "type": "text", + "content": "Our TUMTraf V2X Cooperative Perception Dataset was recorded on an ITS system with nine sensors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 642, + 241, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 241, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 241, + 653 + ], + "type": "text", + "content": "The infrastructure sensor setup is the following:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 654, + 287, + 700 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 47, + 654, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 286, + 677 + ], + "type": "text", + "content": "- 1x Ouster LiDAR OS1-64 (gen. 2), 64 vert. layers, " + }, + { + "bbox": [ + 47, + 654, + 286, + 677 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 47, + 654, + 286, + 677 + ], + "type": "text", + "content": " FOV, below horizon config., " + }, + { + "bbox": [ + 47, + 654, + 286, + 677 + ], + "type": "inline_equation", + "content": "10\\mathrm{cm}" + }, + { + "bbox": [ + 47, + 654, + 286, + 677 + ], + "type": "text", + "content": " acc. @ " + }, + { + "bbox": [ + 47, + 654, + 286, + 677 + ], + "type": "inline_equation", + "content": "120\\mathrm{m}" + }, + { + "bbox": [ + 47, + 654, + 286, + 677 + ], + "type": "text", + "content": " range" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 700 + ], + "type": "text", + "content": "- 4x Basler ace acA1920-50gc, " + }, + { + "bbox": [ + 47, + 677, + 287, + 700 + ], + "type": "inline_equation", + "content": "1920 \\times 1200" + }, + { + "bbox": [ + 47, + 677, + 287, + 700 + ], + "type": "text", + "content": ", Sony IMX174 with 8 mm lenses" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 701, + 244, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 244, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 244, + 713 + ], + "type": "text", + "content": "On the vehicle, the following sensors were used:" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 414, + 190 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 414, + 190 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 414, + 190 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 414, + 190 + ], + "type": "image", + "image_path": "5315b3b700c301da35141cb641a351776de308830170b77ca6aaa1787ab8ec9f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 199, + 545, + 243 + ], + "lines": [ + { + "bbox": [ + 305, + 199, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 199, + 545, + 243 + ], + "type": "text", + "content": "Figure 3. Demonstration of a possible V2X occlusion scenario. A pedestrian (blue) is crossing the road in front of the ego vehicle. An occluded bicycle is marked in red. The recording vehicle with the sensor setup is shown in the bottom left corner." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 414, + 70, + 544, + 190 + ], + "blocks": [ + { + "bbox": [ + 414, + 70, + 544, + 190 + ], + "lines": [ + { + "bbox": [ + 414, + 70, + 544, + 190 + ], + "spans": [ + { + "bbox": [ + 414, + 70, + 544, + 190 + ], + "type": "image", + "image_path": "5604719f92ed31b04a8775f70728bc90eb7fb42a5f5a2670f9d35e12a682726a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 262, + 545, + 333 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 306, + 262, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 545, + 285 + ], + "type": "text", + "content": "- 1x Robosense RS-LiDAR-32, 32 vert. layers, " + }, + { + "bbox": [ + 306, + 262, + 545, + 285 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 306, + 262, + 545, + 285 + ], + "type": "text", + "content": " FOV, 3 cm accuracy @200 m range" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 286, + 545, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 286, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 306, + 286, + 545, + 309 + ], + "type": "text", + "content": "- 1x Basler ace acA1920-50gc, " + }, + { + "bbox": [ + 306, + 286, + 545, + 309 + ], + "type": "inline_equation", + "content": "1920 \\times 1200" + }, + { + "bbox": [ + 306, + 286, + 545, + 309 + ], + "type": "text", + "content": ", Sony IMX174 with " + }, + { + "bbox": [ + 306, + 286, + 545, + 309 + ], + "type": "inline_equation", + "content": "16 \\mathrm{~mm}" + }, + { + "bbox": [ + 306, + 286, + 545, + 309 + ], + "type": "text", + "content": " lens" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 310, + 537, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 310, + 537, + 321 + ], + "spans": [ + { + "bbox": [ + 306, + 310, + 537, + 321 + ], + "type": "text", + "content": "- 1x Emlid Reach RS2+ multi-band RTK GNSS receiver" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 323, + 449, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 449, + 333 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 449, + 333 + ], + "type": "text", + "content": "- 1x XSENS MTi-30-2A8G4 IMU" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 342, + 493, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 493, + 354 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 493, + 354 + ], + "type": "text", + "content": "3.2. Sensor calibration and registration" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 360, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 552 + ], + "type": "text", + "content": "We synchronize the cameras and LiDARs in the spatial and temporal domain. First, we determine the intrinsic camera parameters and the radial and tangential image distortions by using a checkerboard target. We then calibrate the roadside LiDAR with the roadside cameras by picking 100-point pairs in the point cloud and camera image. Extrinsic parameters (rotation and translation) are calculated by minimizing the reprojection error of 2D-3D point correspondences [25]. We follow the same procedure for onboard camera-LiDAR calibration. Finally, we calibrate the onboard LiDAR to the roadside LiDAR. This spatial registration is done by first estimating a coarse transformation. We pick ten 3D point pairs in each point cloud and minimize their distance using the least squares method. Then, we apply the point-to-point Iterative Closest Point (ICP) algorithm [4] to get the fine transformation between the point clouds." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 552, + 545, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 671 + ], + "type": "text", + "content": "We label the vehicle and infrastructure point clouds after registering them. The coarse registration was done by measuring the GPS position of the onboard LiDAR and the roadside LiDAR. Then, we transform every 10th onboard point cloud to the coordinate system of the infrastructure point cloud. The fine registration was done by applying the point-to-point ICP to get an accurate V2I transformation matrix. All rotations of the point cloud frames in between are interpolated based on the spherical linear interpolation (SLERP) [28] method:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 356, + 679, + 545, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 679, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 356, + 679, + 545, + 693 + ], + "type": "interline_equation", + "content": "S L E R P \\left(q _ {0}, q _ {1}, t\\right) = q _ {0} \\left(q _ {0} ^ {- 1} q _ {1}\\right) ^ {t}, \\tag {1}", + "image_path": "6d7e8672861f4c8de4335375269e3cdf6db45d86bfb888e5496f6afa99ecceae.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "q_{0}" + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "q_{1}" + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": " are the quaternions representing the ro" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22671" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 101, + 286, + 209 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 91 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 91 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 91 + ], + "type": "text", + "content": "Table 2. Evaluation results (" + }, + { + "bbox": [ + 47, + 70, + 287, + 91 + ], + "type": "inline_equation", + "content": "mAP_{BEV}" + }, + { + "bbox": [ + 47, + 70, + 287, + 91 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 70, + 287, + 91 + ], + "type": "inline_equation", + "content": "mAP_{3D}" + }, + { + "bbox": [ + 47, + 70, + 287, + 91 + ], + "type": "text", + "content": ") of CoopDet3D on our TUMTraf-V2X test set in south2 FOV." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 101, + 286, + 209 + ], + "lines": [ + { + "bbox": [ + 47, + 101, + 286, + 209 + ], + "spans": [ + { + "bbox": [ + 47, + 101, + 286, + 209 + ], + "type": "table", + "html": "
Config.\\( \\mathbf{mAP_{BEV}}\\uparrow \\)\\( \\mathbf{mAP_{3D}}\\uparrow \\)
DomainModalityEasy↑Mod.↑Hard↑Avg.↑
VehicleCamera46.8331.4737.8230.7730.36
VehicleLiDAR85.3385.2276.8669.0480.11
VehicleCam+LiDAR84.9077.6072.0873.1276.40
Infra.Camera61.9831.1946.7340.4235.04
Infra.LiDAR92.8686.1788.0775.7384.88
Infra.Cam+LiDAR92.9287.9989.0981.6987.01
Coop.Camera68.9445.4142.7657.8345.74
Coop.LiDAR93.9392.6378.0673.9585.86
Coop.Cam+LiDAR94.2293.4288.1779.9490.76
", + "image_path": "844b6340366deb9cca125ac434192a9ffb43fee2c45418b3447e1d7f599ac9a3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 230, + 286, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 230, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 230, + 286, + 255 + ], + "type": "text", + "content": "tations of the start and end frames and " + }, + { + "bbox": [ + 47, + 230, + 286, + 255 + ], + "type": "inline_equation", + "content": "t \\in [0,1]" + }, + { + "bbox": [ + 47, + 230, + 286, + 255 + ], + "type": "text", + "content": ". Translation vectors " + }, + { + "bbox": [ + 47, + 230, + 286, + 255 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_0" + }, + { + "bbox": [ + 47, + 230, + 286, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 230, + 286, + 255 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_1" + }, + { + "bbox": [ + 47, + 230, + 286, + 255 + ], + "type": "text", + "content": " were obtained using linear interpolation:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 266, + 286, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 266, + 286, + 279 + ], + "spans": [ + { + "bbox": [ + 111, + 266, + 286, + 279 + ], + "type": "interline_equation", + "content": "\\mathbf {T} (t) = \\mathbf {T} _ {0} + t \\left(\\mathbf {T} _ {1} - \\mathbf {T} _ {0}\\right). \\tag {2}", + "image_path": "82e409618ac64a124743e181b4739429b5a2ba2ac90f976befefa9389a3f45b8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 285, + 286, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 286, + 331 + ], + "type": "text", + "content": "This dual interpolation strategy ensures that the estimated transformations between the frames are smooth and geometrically accurate, thus adhering closely to the actual movements of the vehicle over time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 339, + 197, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 339, + 197, + 353 + ], + "spans": [ + { + "bbox": [ + 47, + 339, + 197, + 353 + ], + "type": "text", + "content": "3.3. Data selection and labeling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 358, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 287, + 646 + ], + "type": "text", + "content": "We selected the data based on challenging traffic scenarios, like U-turns, tailgate events, and traffic violation maneuvers. Besides the high traffic density of 31 objects per frame, we selected frames with high-class coverage. We selected 700 frames during sunny daytime and 100 frames during cloudy nighttime for labeling. The camera and Li-DAR data were recorded into rosbag files at " + }, + { + "bbox": [ + 46, + 358, + 287, + 646 + ], + "type": "inline_equation", + "content": "15\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 358, + 287, + 646 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 358, + 287, + 646 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 358, + 287, + 646 + ], + "type": "text", + "content": ", respectively. We extracted and synchronized the data based on ROS [27] timestamps and labeled it with our " + }, + { + "bbox": [ + 46, + 358, + 287, + 646 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 46, + 358, + 287, + 646 + ], + "type": "text", + "content": " BAT (v24.3.2) annotation tool1. We improved the 3D BAT [45] baseline labeling tool to label 3D objects faster and more precisely with a one-click annotation feature. The annotators were instructed to label traffic participants while examining the images. Objects are still labeled, even if they have no 3D points inside, but are visible in the images. Extremities (e.g., pedestrian limbs) are included in the bounding box, but side mirrors of vehicles aren't. If a pedestrian carries an object, that object is included in the bounding box. If two or more pedestrians are carrying an object, only the box of one will include the object. After labeling, each annotator checked the work of other annotators manually frame-by-frame. When errors were found, the original annotator was notified, and they fixed it. This helps ensure that the labels in our dataset are high quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 653, + 194, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 194, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 194, + 664 + ], + "type": "text", + "content": "3.4. Data structure and format" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 671, + 286, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 286, + 695 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 286, + 695 + ], + "type": "text", + "content": "We record eight different scenes, each 10 sec. long, from vehicle and infrastructure perspectives using nine sensors" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 306, + 111, + 545, + 211 + ], + "blocks": [ + { + "bbox": [ + 306, + 69, + 545, + 102 + ], + "lines": [ + { + "bbox": [ + 306, + 69, + 545, + 102 + ], + "spans": [ + { + "bbox": [ + 306, + 69, + 545, + 102 + ], + "type": "text", + "content": "Table 3. Evaluation results of infrastructure-only CoopDet3D vs. InfraDet3D [47] on TUMTraf Intersection test set [48]. South 1 and South 2 refer to sensors covering different FOVs." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 111, + 545, + 211 + ], + "lines": [ + { + "bbox": [ + 306, + 111, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 306, + 111, + 545, + 211 + ], + "type": "table", + "html": "
ModelFOVModality\\( \\mathrm{mAP}_{3D} \\uparrow \\)
Easy↑Mod.↑Hard↑Avg.↑
InfraDet3Dsouth 1LiDAR75.8147.6642.1655.21
CoopDet3Dsouth 1LiDAR76.2448.2335.1969.47
InfraDet3Dsouth 2LiDAR38.9246.6043.8643.13
CoopDet3Dsouth 2LiDAR74.9755.5539.9669.94
InfraDet3Dsouth 1Cam+LiDAR67.0831.3835.1744.55
CoopDet3Dsouth 1Cam+LiDAR75.6845.6345.6366.75
InfraDet3Dsouth 2Cam+LiDAR58.3819.7333.0837.06
CoopDet3Dsouth 2Cam+LiDAR74.7353.4641.9666.89
", + "image_path": "1ad48c9990c6287dd2d933b3b388376109ba81a7b18587a4d6a07459b4ba77fe.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 233, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 233, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 233, + 545, + 281 + ], + "type": "text", + "content": "and split the data into a train (80%), val. (10%), and test (10%) set. We use stratified sampling to distribute all sets' object classes equally (see Fig. 6a). Labels are provided in the ASAM OpenLABEL [13] standard." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 289, + 443, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 289, + 443, + 301 + ], + "spans": [ + { + "bbox": [ + 306, + 289, + 443, + 301 + ], + "type": "text", + "content": "3.5. Dataset development kit" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 307, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 307, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 305, + 307, + 545, + 474 + ], + "type": "text", + "content": "We provide a devkit to work with our dataset. In addition to generating the data statistics, it provides modules for multiclass stratified splitting (train/val/test), point cloud registration, loading annotations in OpenLABEL format, evaluation of detection and tracking results, pre-processing steps such as point cloud filtering, and post-processing such as bounding box filtering. The statistics (Fig. 4, 5, and 6) were created using our devkit. It also contains modules to convert the labels from OpenLABEL to KITTI or our custom nuScenes format with timestamps instead of tokens and vice versa. This devkit enables users to migrate their models and make them compatible with our dataset format. We release our devkit2 under the MIT license and the dataset under the Creative Commons (CC) BY-NC-ND 4.0 license." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 485, + 384, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 485, + 384, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 485, + 384, + 498 + ], + "type": "text", + "content": "4. Benchmark" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 506, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 506, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 305, + 506, + 545, + 541 + ], + "type": "text", + "content": "We propose CoopDet3D, an extension of BEVFusion [23] and PillarGrid [2] for deep cooperative multi-modal 3D object detection and benchmark it on TUMTraf-V2X mini." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 549, + 418, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 549, + 418, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 549, + 418, + 560 + ], + "type": "text", + "content": "4.1. Evaluation metrics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 567, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 567, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 305, + 567, + 545, + 687 + ], + "type": "text", + "content": "The accuracy is measured in terms of the mean average precision (mAP). Two types of mAP measures are used: BEV mAP considers the BEV center distance, and the results are obtained using the same evaluation methodology used in BEVFusion, which in turn uses the evaluation protocol of nuScenes [6]. Similarly, the 3D mAP measure considers the intersection in 3D, and the results are obtained using the evaluation script of our TUM Traffic dataset devkit. The runtime is evaluated using frames per second (FPS) as the metric and the results were obtained by measuring the time" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 211, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 211, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 211, + 712 + ], + "type": "text", + "content": "1https://github.com/walzimmer/3d-bat" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 694, + 544, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 694, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 694, + 544, + 712 + ], + "type": "text", + "content": "2https://github.com/tum-traffic-dataset/tum-traffic-dataset-dev-kit" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22672" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 70, + 214, + 198 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 214, + 198 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 214, + 198 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 214, + 198 + ], + "type": "image", + "image_path": "d206a3a08e78e440783e2077d0f0221eee7bbbc5438da77ea290e84e50bcd957.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 200, + 209, + 210 + ], + "lines": [ + { + "bbox": [ + 49, + 200, + 209, + 210 + ], + "spans": [ + { + "bbox": [ + 49, + 200, + 209, + 210 + ], + "type": "text", + "content": "(a) Distribution of objects between day and night." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 217, + 70, + 380, + 198 + ], + "blocks": [ + { + "bbox": [ + 217, + 70, + 380, + 198 + ], + "lines": [ + { + "bbox": [ + 217, + 70, + 380, + 198 + ], + "spans": [ + { + "bbox": [ + 217, + 70, + 380, + 198 + ], + "type": "image", + "image_path": "85b64cb2a64f09feb2b1a6bd6d1bd9c48936e20c9660488e11997eb6f363d46e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 200, + 379, + 210 + ], + "lines": [ + { + "bbox": [ + 214, + 200, + 379, + 210 + ], + "spans": [ + { + "bbox": [ + 214, + 200, + 379, + 210 + ], + "type": "text", + "content": "(b) Avg. and max. num. of 3D points for each class." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 383, + 70, + 547, + 198 + ], + "blocks": [ + { + "bbox": [ + 383, + 70, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 383, + 70, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 383, + 70, + 547, + 198 + ], + "type": "image", + "image_path": "0f371d6d49ce5a390108a7175a9efb49d4fd83275eff48587d85e00fcad5ab47.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 200, + 535, + 210 + ], + "lines": [ + { + "bbox": [ + 388, + 200, + 535, + 210 + ], + "spans": [ + { + "bbox": [ + 388, + 200, + 535, + 210 + ], + "type": "text", + "content": "(c) Avg. and max. track length for all classes." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 52, + 284, + 208, + 430 + ], + "blocks": [ + { + "bbox": [ + 52, + 284, + 208, + 430 + ], + "lines": [ + { + "bbox": [ + 52, + 284, + 208, + 430 + ], + "spans": [ + { + "bbox": [ + 52, + 284, + 208, + 430 + ], + "type": "image", + "image_path": "15971227ab788286c7454d3b127c8fa1518b558600cfe37ed1fa3718fd264e43.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 432, + 187, + 441 + ], + "lines": [ + { + "bbox": [ + 71, + 432, + 187, + 441 + ], + "spans": [ + { + "bbox": [ + 71, + 432, + 187, + 441 + ], + "type": "text", + "content": "(a) Visualization of rotations (yaw)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 217, + 282, + 380, + 427 + ], + "blocks": [ + { + "bbox": [ + 46, + 228, + 547, + 274 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 547, + 274 + ], + "type": "text", + "content": "Figure 4. Our TUMTraf-V2X dataset contains 30k 3D box labels in total and is balanced among nine different object classes. (a) Cars (14,633) and pedestrians (5,276) are highly represented in the dataset. (b) 3D box labels contain, on average 625 points inside which shows the density of the labeled objects. The BUS class has the highest point density. (c) All traffic participants are tracked for " + }, + { + "bbox": [ + 46, + 228, + 547, + 274 + ], + "type": "inline_equation", + "content": "39\\mathrm{m}" + }, + { + "bbox": [ + 46, + 228, + 547, + 274 + ], + "type": "text", + "content": " on average. Emergency vehicles have the highest average track length of " + }, + { + "bbox": [ + 46, + 228, + 547, + 274 + ], + "type": "inline_equation", + "content": "128\\mathrm{m}" + }, + { + "bbox": [ + 46, + 228, + 547, + 274 + ], + "type": "text", + "content": ", whereas the CAR class contains a max. track length of " + }, + { + "bbox": [ + 46, + 228, + 547, + 274 + ], + "type": "inline_equation", + "content": "174\\mathrm{m}" + }, + { + "bbox": [ + 46, + 228, + 547, + 274 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 217, + 282, + 380, + 427 + ], + "lines": [ + { + "bbox": [ + 217, + 282, + 380, + 427 + ], + "spans": [ + { + "bbox": [ + 217, + 282, + 380, + 427 + ], + "type": "image", + "image_path": "3c7804d4733a5c3c0112e6f814c5fee905fa9b156935008a6c3541c55622b450.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 432, + 351, + 441 + ], + "lines": [ + { + "bbox": [ + 239, + 432, + 351, + 441 + ], + "spans": [ + { + "bbox": [ + 239, + 432, + 351, + 441 + ], + "type": "text", + "content": "(b) 3D points grouped by distance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 386, + 282, + 541, + 430 + ], + "blocks": [ + { + "bbox": [ + 386, + 282, + 541, + 430 + ], + "lines": [ + { + "bbox": [ + 386, + 282, + 541, + 430 + ], + "spans": [ + { + "bbox": [ + 386, + 282, + 541, + 430 + ], + "type": "image", + "image_path": "380024a0d7512fb60c794f84746c69a477e79068e6d333d946693dc62283c6b5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 431, + 512, + 441 + ], + "lines": [ + { + "bbox": [ + 410, + 431, + 512, + 441 + ], + "spans": [ + { + "bbox": [ + 410, + 431, + 512, + 441 + ], + "type": "text", + "content": "(c) BEV visualization of tracks." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 52, + 493, + 211, + 639 + ], + "blocks": [ + { + "bbox": [ + 46, + 450, + 547, + 483 + ], + "lines": [ + { + "bbox": [ + 46, + 450, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 547, + 483 + ], + "type": "text", + "content": "Figure 5. Our dataset was recorded at a crowded intersection with many left and right turns. (a) Most of the vehicles (7,559) are driving in the east direction (0 degrees). (b) 3D boxes were labeled up to " + }, + { + "bbox": [ + 46, + 450, + 547, + 483 + ], + "type": "inline_equation", + "content": "200\\mathrm{m}" + }, + { + "bbox": [ + 46, + 450, + 547, + 483 + ], + "type": "text", + "content": " range and are very dense between 10 and " + }, + { + "bbox": [ + 46, + 450, + 547, + 483 + ], + "type": "inline_equation", + "content": "60\\mathrm{m}" + }, + { + "bbox": [ + 46, + 450, + 547, + 483 + ], + "type": "text", + "content": ". (c) The visualization of BEV tracks shows where pedestrians and bicycles are crossing the road." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 493, + 211, + 639 + ], + "lines": [ + { + "bbox": [ + 52, + 493, + 211, + 639 + ], + "spans": [ + { + "bbox": [ + 52, + 493, + 211, + 639 + ], + "type": "image", + "image_path": "705d9703c938963694243093783f6b3cce1b1911338913a1c662bdbd23c01774.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 642, + 175, + 652 + ], + "lines": [ + { + "bbox": [ + 84, + 642, + 175, + 652 + ], + "spans": [ + { + "bbox": [ + 84, + 642, + 175, + 652 + ], + "type": "text", + "content": "(a) Histogram of 3D points." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 219, + 493, + 376, + 639 + ], + "blocks": [ + { + "bbox": [ + 219, + 493, + 376, + 639 + ], + "lines": [ + { + "bbox": [ + 219, + 493, + 376, + 639 + ], + "spans": [ + { + "bbox": [ + 219, + 493, + 376, + 639 + ], + "type": "image", + "image_path": "f0d78baf01401ab6a2122e2b1a3a6e613b5ea443672373e8254ad13219b9d133.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 243, + 642, + 347, + 652 + ], + "lines": [ + { + "bbox": [ + 243, + 642, + 347, + 652 + ], + "spans": [ + { + "bbox": [ + 243, + 642, + 347, + 652 + ], + "type": "text", + "content": "(b) Histogram of 3D box labels." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 380, + 493, + 543, + 640 + ], + "blocks": [ + { + "bbox": [ + 380, + 493, + 543, + 640 + ], + "lines": [ + { + "bbox": [ + 380, + 493, + 543, + 640 + ], + "spans": [ + { + "bbox": [ + 380, + 493, + 543, + 640 + ], + "type": "image", + "image_path": "352a55c2997c4ffbacce65d64e73b9216319a1e203e4e1db84322437c40f1269.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 642, + 511, + 652 + ], + "lines": [ + { + "bbox": [ + 411, + 642, + 511, + 652 + ], + "spans": [ + { + "bbox": [ + 411, + 642, + 511, + 652 + ], + "type": "text", + "content": "(c) Histogram of track lengths." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 661, + 547, + 696 + ], + "lines": [ + { + "bbox": [ + 46, + 661, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 661, + 547, + 696 + ], + "type": "text", + "content": "Figure 6. (a) Most 3D box labels contain between 1 and 200 3D points inside, with an average of 434 3D points, excluding empty boxes. Objects that were close to both LiDAR sensors even contained up to 18k 3D points. (b) Frames contain between 15 and 45 traffic participants, with an average of 29. (c) Objects were tracked up to " + }, + { + "bbox": [ + 46, + 661, + 547, + 696 + ], + "type": "inline_equation", + "content": "170\\mathrm{m}" + }, + { + "bbox": [ + 46, + 661, + 547, + 696 + ], + "type": "text", + "content": ", and the average track length is " + }, + { + "bbox": [ + 46, + 661, + 547, + 696 + ], + "type": "inline_equation", + "content": "39.35\\mathrm{m}" + }, + { + "bbox": [ + 46, + 661, + 547, + 696 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22673" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 547, + 229 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 547, + 229 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 547, + 229 + ], + "type": "image", + "image_path": "866fd966ccb2b5e0071274b97dad1103117962b5448090ff4ee0510542ea0a0d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 238, + 546, + 273 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 546, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 546, + 273 + ], + "type": "text", + "content": "Figure 7. Qualitative results on the TUMTraf-V2X Cooperative Perception test set. The first row shows the inference results of the onboard (vehicle-only) camera-LiDAR fusion with 23 detected objects. In the second row, the results of the cooperative vehicle-infrastructure camera-LiDAR fusion are visualized. Here, all 25 traffic participants could be detected with the support of roadside sensors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 342, + 285, + 468 + ], + "blocks": [ + { + "bbox": [ + 46, + 289, + 287, + 334 + ], + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 334 + ], + "type": "text", + "content": "Table 4. Ablation study on cooperative 3D object detection with 11 combinations of camera and LiDAR backbones. The best tradeoff between speed and accuracy is highlighted in gray. " + }, + { + "bbox": [ + 46, + 289, + 287, + 334 + ], + "type": "inline_equation", + "content": "\\mathrm{ND} =" + }, + { + "bbox": [ + 46, + 289, + 287, + 334 + ], + "type": "text", + "content": " non-deterministic, " + }, + { + "bbox": [ + 46, + 289, + 287, + 334 + ], + "type": "inline_equation", + "content": "\\mathrm{TS} =" + }, + { + "bbox": [ + 46, + 289, + 287, + 334 + ], + "type": "text", + "content": " TorchSparse, " + }, + { + "bbox": [ + 46, + 289, + 287, + 334 + ], + "type": "inline_equation", + "content": "\\odot =" + }, + { + "bbox": [ + 46, + 289, + 287, + 334 + ], + "type": "text", + "content": " retrained. VRAM is in GiB." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 342, + 285, + 468 + ], + "lines": [ + { + "bbox": [ + 47, + 342, + 285, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 285, + 468 + ], + "type": "table", + "html": "
Backbone Configuration\\( \\mathbf{{mAP}}_{\\mathbf{{BEV}}} \\uparrow \\)\\( \\mathbf{{FPS}} \\uparrow \\)\\( \\mathbf{{VRAM}} \\downarrow \\)
VoxelNet ND + SwinT93.476.306.69
VoxelNet ND + YOLOv8 s92.947.246.39
VoxelNet TS + SwinT93.518.844.61
VoxelNet TS + YOLOv8 s92.9410.664.28
VoxelNet TS + YOLOv8 s ○94.3110.664.28
PointPillars 512 + Swin T94.439.004.94
PointPillars 512 + YOLOv8 s94.2711.144.63
PointPillars 512 + YOLOv8 s ○94.2511.144.63
PointPillars 512_2x + Swin T92.799.064.94
PointPillars 512_2x + YOLOv8 s94.1611.204.63
PointPillars 512_2x + YOLOv8 s ○94.2211.204.63
", + "image_path": "acd460f8b58e7042a21d82314501e603e370e7de61bb737c7a9d3f67141e6066.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 491, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 563 + ], + "type": "text", + "content": "needed by the model to run one full inference, including data preprocessing and voxelization. The first five iterations are skipped as a warmup since they are usually considerably slower than the average. Finally, the complexity of the model is measured in terms of the maximal VRAM usage across all GPUs during training and testing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 574, + 157, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 157, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 157, + 586 + ], + "type": "text", + "content": "4.2. CoopDet3D model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "Our CoopDet3D uses a BEVFusion-based backbone for camera-LiDAR fusion on the vehicle and the infrastructure sides separately to obtain the vehicle and infrastructure features. The best backbone for image and point cloud feature extraction was chosen through multiple ablation studies. Then, inspired by the method proposed by PillarGrid [2], an element-wise max-pooling operation is proposed to fuse the resulting fused camera-LiDAR features of vehicle and infrastructure together. Finally, the detection head from BEVFusion is used for 3D detection from the fused feature." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 292, + 511, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 292, + 511, + 303 + ], + "spans": [ + { + "bbox": [ + 305, + 292, + 511, + 303 + ], + "type": "text", + "content": "The architecture of CoopDet3D is shown in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 304, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 460 + ], + "type": "text", + "content": "First, we disable the camera feature extraction nodes and train the LiDAR-only model for 20 epochs. Then, we use pre-trained weights for the cooperative model and fine-tune the entire model for eight further epochs. Hyperparameter tuning revealed that the default hyperparameters of BEV-Fusion [23] gave the best results, and such were not modified. The preprocessing steps are also the same as for BEV-Fusion, but we change the point cloud range to " + }, + { + "bbox": [ + 304, + 304, + 545, + 460 + ], + "type": "inline_equation", + "content": "[-75, 75]" + }, + { + "bbox": [ + 304, + 304, + 545, + 460 + ], + "type": "text", + "content": " in the x- and y-scale and " + }, + { + "bbox": [ + 304, + 304, + 545, + 460 + ], + "type": "inline_equation", + "content": "[-8, 0]" + }, + { + "bbox": [ + 304, + 304, + 545, + 460 + ], + "type": "text", + "content": " in the z-scale since the dataset used in this case is different. Furthermore, we use 3x NVIDIA RTX 3090 GPUs with 24 GB VRAM for training and a single GPU for evaluation. We open-source our model and provide pre-trained weights3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 467, + 486, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 467, + 486, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 467, + 486, + 479 + ], + "type": "text", + "content": "4.3. Experiments and ablation studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 485, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 545 + ], + "type": "text", + "content": "The objective of these experiments is to highlight the importance of our V2X multi-viewpoint dataset as opposed to single-viewpoint datasets. As such, we conduct multiple experiments and ablation studies with data obtained from each viewpoint and compare the results on the proposed model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 552, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 622 + ], + "type": "text", + "content": "Ablation on viewpoint and modality. We conduct multiple experiments with all possible combinations of a) viewpoints: vehicle-only, infrastructure-only, cooperative, and b) modalities: camera-only, LiDAR-only, and camera-LiDAR fusion. Table 2 shows the mAP achieved by CoopDet3D for each of these combinations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 623, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 623, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 623, + 545, + 696 + ], + "type": "text", + "content": "We observe that the results follow a general pattern of cooperative performance being better than infrastructure-only, which is, in turn, better than vehicle-only. Furthermore, fusion models perform better than LiDAR-only models, which in turn are better than camera-only models. Figure 7 shows qualitative results between our vehicle-only" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 526, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 526, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 526, + 712 + ], + "type": "text", + "content": "3https://github.com/tum-traffic-dataset/coopdet3d" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22674" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "camera-LiDAR fusion model and our cooperative vehicle-infrastructure camera-LiDAR fusion model. Again, we observe from these samples that the cooperative perception model is able to detect 25 traffic participants, whereas the vehicle-only model is only able to detect 23 objects due to occlusions and a limited field of view." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 152, + 287, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 152, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 287, + 282 + ], + "type": "text", + "content": "Ablation on fusion level. We compare our CoopDet3D model to the current SOTA camera-LiDAR fusion method InfraDet3D [47] on the TUMTraf Intersection test set [48]. The proposed method uses deep fusion, whereas the InfraDet3D method is a late fusion method. Table 3 shows the performance of our model against InfraDet3D. The results show that the proposed deep fusion method outperforms the SOTA late fusion model in all metrics, except in the hard difficulty in LiDAR-only mode. Furthermore, Fig. 8 shows two sample images taken during day and nighttime, wherein deep fusion again outperforms late fusion." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "text", + "content": "We note that these experiments were conducted in an offline setting, disregarding other considerations for simplicity. However, when deploying it in real life, factors such as the transmission bandwidth should also be considered. Since we observed that deep feature fusion generally leads to higher efficacy, the V2I transmissions should contain these features instead of infrastructure bounding boxes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 531 + ], + "type": "text", + "content": "Ablation on backbones. As an ablation study, we present the results of the experiments to find the best backbone and model configuration for the cooperative camera-LiDAR fusion model. For the camera backbone, SwinT [22] and MMYOLO's [8] implementation of YOLOv8 [16] were considered. For the LiDAR backbone, VoxelNet [44] and PointPillars [19] were considered. In addition, VoxelNet was implemented with two different backends, namely SPConv v2 and Torchsparse [31]. For PointPillars, two grid sizes are considered " + }, + { + "bbox": [ + 46, + 376, + 287, + 531 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 376, + 287, + 531 + ], + "type": "text", + "content": " for both train and test grids (PointPillars 512) and " + }, + { + "bbox": [ + 46, + 376, + 287, + 531 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 376, + 287, + 531 + ], + "type": "text", + "content": " train grid with " + }, + { + "bbox": [ + 46, + 376, + 287, + 531 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 46, + 376, + 287, + 531 + ], + "type": "text", + "content": " test grid (PointPillars 512_2x). The results of these experiments are shown in Table 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "content": "The results show that only models that use any combination of VoxelNet Torchsparse, both PointPillars variants, and YOLOv8 are able to run above 10 FPS. From these configurations, we choose PointPillars 512_2x with YOLOv8 as the best configuration for all the above experiments as it achieves the best results across all the ablation studies. This is a promising result since we also know that this backbone configuration is able to run in real-time (11.2 FPS) on an RTX 3090 without using TensorRT acceleration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "An interesting observation is that utilizing pre-trained weights for transfer learning of YOLOv8 is not always beneficial, as the results from PointPillars " + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "inline_equation", + "content": "512 + \\mathrm{YOLOv8}" + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": " show. This is likely because the pre-trained weights were from MS COCO [21], and they have a very different data domain compared to our dataset. Since MS COCO is also" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 312, + 71, + 426, + 178 + ], + "blocks": [ + { + "bbox": [ + 312, + 71, + 426, + 178 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 426, + 178 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 426, + 178 + ], + "type": "image", + "image_path": "193a63668ff5d7a771663d6ae7c6110ce8a11f5a42eea5055b414f37ea7a4100.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 187, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 305, + 187, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 187, + 545, + 232 + ], + "type": "text", + "content": "Figure 8. Qualitative results of our CoopDet3D (left) and the InfraDet3D (right) model on the TUMTraf Intersection test set during day and nighttime. Detected objects marked with a red circle were classified correctly by CoopDet3D." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 427, + 71, + 542, + 178 + ], + "blocks": [ + { + "bbox": [ + 427, + 71, + 542, + 178 + ], + "lines": [ + { + "bbox": [ + 427, + 71, + 542, + 178 + ], + "spans": [ + { + "bbox": [ + 427, + 71, + 542, + 178 + ], + "type": "image", + "image_path": "6ec51ed62d9f340fe69b01b55056f1cbd7bd0fdcfe8e391cf47aed40d835af47.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 257, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 257, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 257, + 545, + 281 + ], + "type": "text", + "content": "much larger than our dataset in terms of camera images, retraining harms the performance of the model slightly." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 283, + 545, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 391 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 391 + ], + "type": "text", + "content": "In terms of efficiency, the goal of these experiments was to verify that the proposed CoopDet3D model with the best configuration provides the highest accuracy while also being able to run in real-time (minimum of " + }, + { + "bbox": [ + 304, + 283, + 545, + 391 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 283, + 545, + 391 + ], + "type": "text", + "content": "). Furthermore, it should also be feasible to train the model on a high-performance GPU and perform inference on a mid-range consumer GPU deployable on an edge device. The results concerning the VRAM usage during inference show that the complexity of the model makes this feasible." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 410, + 468, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 410, + 468, + 423 + ], + "spans": [ + { + "bbox": [ + 306, + 410, + 468, + 423 + ], + "type": "text", + "content": "5. Conclusion and future work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 433, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 635 + ], + "type": "text", + "content": "This work proposes the TUMTraf-V2X dataset, a multimodal multi-view V2X dataset for cooperative 3D object detection and tracking. Our dataset focuses on challenging traffic scenarios at an intersection and provides views from the infrastructure and the ego vehicle. To benchmark the dataset, we propose CoopDet3D - a baseline model for cooperative perception. Experiments show that cooperative fusion leads to higher efficacy than its unimodal and single-view camera-LiDAR fusion counterparts. Furthermore, cooperative fusion leads to an improvement of +14.3 3D mAP compared to vehicle-only perception, highlighting the need for V2X datasets. Finally, we provide our 3D BAT v24.3.2 labeling tool and devkit to load, parse, and visualize the dataset. It also includes modules for pre- and postprocessing and evaluation. Future efforts will integrate this platform into online environments, enabling a broader range of infrastructure-based, real-time perception applications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 655, + 399, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 655, + 399, + 668 + ], + "spans": [ + { + "bbox": [ + 306, + 655, + 399, + 668 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "This research was supported by the Federal Ministry of Education and Research in Germany within the AUTOtech.agil project, Grant Number: 01IS22088U." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22675" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 201 + ], + "type": "text", + "content": "[2] Zhengwei Bai, Guoyuan Wu, Matthew J Barth, Yongkang Liu, Emrah Akin Sisbot, and Kentaro Oguchi. Pillargrid: Deep learning-based cooperative perception for 3d object detection from onboard-roadside lidar. In IEEE 25th International Conference on Intelligent Transportation Systems (ITSC), pages 1743-1749. IEEE, 2022. 4, 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "type": "text", + "content": "[3] Zhengwei Bai, Guoyuan Wu, Matthew J Barth, Yongkang Liu, Emrah Akin Sisbot, and Kentaro Oguchi. Vinet: Lightweight, scalable, and heterogeneous cooperative perception for 3d object detection. Mechanical Systems and Signal Processing, 204:110723, 2023. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 290 + ], + "type": "text", + "content": "[4] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, pages 586-606. Spie, 1992. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "type": "text", + "content": "[5] Steffen Busch, Christian Koetsier, Jeldrik Axmann, and Claus Brenner. Lumpi: The leibniz university multiperspective intersection dataset. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 1127-1134. IEEE, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 401 + ], + "type": "text", + "content": "[6] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11621-11631, 2020. 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 402, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 456 + ], + "type": "text", + "content": "[7] Ziming Chen, Yifeng Shi, and Jinrang Jia. Transiff: An instance-level feature fusion framework for vehicleinfrastructure cooperative 3d detection with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18205-18214, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 457, + 288, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 457, + 288, + 489 + ], + "spans": [ + { + "bbox": [ + 53, + 457, + 288, + 489 + ], + "type": "text", + "content": "[8] MMYOLO Contributors. MMYOLO: OpenMMLab YOLO series toolbox and benchmark. https://github.com/open-mmlab/mmyolo, 2022.8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 288, + 534 + ], + "type": "text", + "content": "[9] Christian Creß, Zhenshan Bing, and Alois C. Knoll. Intelligent transportation systems using roadside infrastructure: A literature survey. IEEE Transactions on Intelligent Transportation Systems, pages 1-0, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "text", + "content": "[10] Christian Creß, Walter Zimmer, Leah Strand, Maximilian Fortkord, Siyi Dai, Venkatnarayanan Lakshminarasimhan, and Alois Knoll. A9-dataset: Multi-sensor infrastructure-based dataset for mobility research. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 965-970, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In Proceedings of the 1st Annual Conference on Robot Learning, pages 1-16, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "text", + "content": "[12] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 679, + 257, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 679, + 257, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 679, + 257, + 690 + ], + "type": "text", + "content": "[13] Nicco Hagedorn. OpenLABEL Concept Paper. 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Yuze He, Li Ma, Zhehao Jiang, Yi Tang, and Guoliang Xing. Vi-eye: Semantic-based 3d point cloud registration" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "for infrastructure-assisted autonomous driving. In Proceedings of the 27th Annual International Conference on Mobile Computing and Networking, pages 573-586, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 107, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 160 + ], + "type": "text", + "content": "[15] Yue Hu, Yifan Lu, Runsheng Xu, Weidi Xie, Siheng Chen, and Yanfeng Wang. Collaboration helps camera overtake lidar in 3d detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243-9252, 2023. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 162, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 183 + ], + "type": "text", + "content": "[16] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolo, 2023. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 184, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 248 + ], + "type": "text", + "content": "[17] Annkathrin Krammer, Christoph Schöller, Dhiraj Gulati, Venkat Narayanan Lakshminarasimhan, Franz Kurz, Dominik Rosenbaum, Claus Lenz, and Alois Knoll. Providentia-a large-scale sensor system for the assistance of autonomous vehicles and its evaluation. Journal of Field Robotics, pages 1156-1176, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "type": "text", + "content": "[18] Abhinav Kumar, Garrick Brazil, Enrique Corona, Armin Parchami, and Xiaoming Liu. Deviant: Depth equivariant network for monocular 3d object detection. In European Conference on Computer Vision, pages 664-683, 2022. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "type": "text", + "content": "[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 4, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[20] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 7(4):10914-10921, 2022. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "text", + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of European Conference on Computer Vision, ECCV 2014, Part V 13, pages 740-755. Springer, 2014. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "text", + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 514, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 569 + ], + "type": "text", + "content": "[23] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE Int. Conference on Robotics and Automation (ICRA), pages 2774-2781, 2023. 4, 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 570, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 635 + ], + "type": "text", + "content": "[24] Pablo Alvarez Lopez, Michael Behrisch, Laura Bieker-Walz, Jakob Erdmann, Yun-Pang Flotteröd, Robert Hilbrich, Leonhard Lücken, Johannes Rummel, Peter Wagner, and Evamarie Wießner. Microscopic traffic simulation using sumo. In The 21st IEEE International Conference on Intelligent Transportation Systems. IEEE, 2018. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 635, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 658 + ], + "type": "text", + "content": "[25] Kaj Madsen, Hans Bruun Nielsen, and Ole Tingleff. Methods for non-linear least squares problems. 2004. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "text", + "content": "[26] Donghao Qiao and Farhana Zulkernine. Cobevfusion: Cooperative perception with lidar-camera bird's-eye view fusion. arXiv preprint arXiv:2310.06008, 2023. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[27] Morgan Quigley, Ken Conley, Brian Gerkey, Josh Faust, Tully Foote, Jeremy Leibs, Rob Wheeler, Andrew Y Ng," + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22676" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "et al. Ros: an open-source robot operating system. In ICRA workshop on open source software, page 5. Kobe, Japan, 2009. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 140 + ], + "type": "text", + "content": "[28] Ken Shoemake. Animating rotation with quaternion curves. In Proceedings of the 12th annual conference on Computer graphics and interactive techniques, pages 245-254, 1985. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 288, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 204 + ], + "type": "text", + "content": "[29] Rui Song, Chenwei Liang, Hu Cao, Zhiran Yan, Walter Zimmer, Markus Gross, Andreas Festag, and Alois Knoll. Collaborative semantic occupancy prediction with hybrid feature fusion in connected automated vehicles. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 288, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 272 + ], + "type": "text", + "content": "[30] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "type": "text", + "content": "[31] Haotian Tang, Shang Yang, Zhijian Liu, Ke Hong, Zhong-ming Yu, Xiuyu Li, Guohao Dai, Yu Wang, and Song Han. TorchSparse++: Efficient Point Cloud Engine. In Computer Vision and Pattern Recognition Workshops CVPRW, 2023. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 317, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 317, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 288, + 369 + ], + "type": "text", + "content": "[32] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4604-4612, 2020. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 426 + ], + "type": "text", + "content": "[33] Sizhe Wei, Yuxi Wei, Yue Hu, Yifan Lu, Yiqi Zhong, Siheng Chen, and Ya Zhang. Asynchrony-robust collaborative perception via bird's eye view flow. In Advances in Neural Information Processing Systems, pages 28462–28477. Curran Associates, Inc. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 480 + ], + "type": "text", + "content": "[34] Zizhang Wu, Yuanzhu Gan, Lei Wang, Guilian Chen, and Jian Pu. Monopgc: Monocular 3d object detection with pixel geometry contexts. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4842-4849, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 482, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 288, + 536 + ], + "type": "text", + "content": "[35] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 2, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 537, + 288, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 592 + ], + "type": "text", + "content": "[36] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 658 + ], + "type": "text", + "content": "[37] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[38] Junjie Yan, Yingfei Liu, Jianjian Sun, Fan Jia, Shuailin Li, Tiancai Wang, and Xiangyu Zhang. Cross modal transformer: Towards fast and robust 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18268-18278, 2023. 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 707 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": "[39] Lei Yang, Kaicheng Yu, Tao Tang, Jun Li, Kun Yuan, Li Wang, Xinyu Zhang, and Peng Chen. Bevheight: A robust framework for vision-based roadside 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21611-21620, 2023. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 547, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 195 + ], + "type": "text", + "content": "[40] Xiaoqing Ye, Mao Shu, Hanyu Li, Yifeng Shi, Yingying Li, Guangjie Wang, Xiao Tan, and Errui Ding. Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21341-21350, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 196, + 547, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 547, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 547, + 261 + ], + "type": "text", + "content": "[41] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21361-21370, 2022. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 262, + 546, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 546, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 546, + 306 + ], + "type": "text", + "content": "[42] Haibao Yu, Yingjuan Tang, Enze Xie, Jilei Mao, Jirui Yuan, Ping Luo, and Zaiqing Nie. Vehicle-infrastructure cooperative 3d object detection via feature flow prediction. arXiv preprint arXiv:2303.10552, 2023. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 308, + 546, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 546, + 383 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 546, + 383 + ], + "type": "text", + "content": "[43] Haibao Yu, Wenxian Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xu Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, et al. V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5486-5495, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 385, + 547, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 385, + 547, + 429 + ], + "spans": [ + { + "bbox": [ + 308, + 385, + 547, + 429 + ], + "type": "text", + "content": "[44] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 430, + 546, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 430, + 546, + 483 + ], + "spans": [ + { + "bbox": [ + 308, + 430, + 546, + 483 + ], + "type": "text", + "content": "[45] Walter Zimmer, Akshay Rangesh, and Mohan Trivedi. 3d bat: A semi-automatic, web-based 3d annotation toolbox for full-surface, multi-modal data streams. In 2019 IEEE Intelligent Vehicles Symposium (IV), pages 1816-1821. IEEE, 2019. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 485, + 546, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 485, + 546, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 485, + 546, + 529 + ], + "type": "text", + "content": "[46] Walter Zimmer, Emec Ercelik, Xingcheng Zhou, Xavier Jair Diaz Ortiz, and Alois Knoll. A survey of robust 3d object detection methods in point clouds. arXiv preprint arXiv:2204.00106, 2022. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 530, + 546, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 530, + 546, + 585 + ], + "spans": [ + { + "bbox": [ + 308, + 530, + 546, + 585 + ], + "type": "text", + "content": "[47] Walter Zimmer, Joseph Birkner, Marcel Brucker, Huu Tung Nguyen, Stefan Petrovski, Bohan Wang, and Alois C. Knoll. Infradet3d: Multi-modal 3d object detection based on roadside infrastructure camera and lidar sensors. In 2023 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2023. 4, 5, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 586, + 546, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 586, + 546, + 640 + ], + "spans": [ + { + "bbox": [ + 308, + 586, + 546, + 640 + ], + "type": "text", + "content": "[48] Walter Zimmer, Christian Creß, Huu Tung Nguyen, and Alois C Knoll. Tumtraf intersection dataset: All you need for urban 3d camera-lidar roadside perception [best student paper award]. In 2023 IEEE Intelligent Transportation Systems ITSC. IEEE, 2023. 2, 5, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 642, + 546, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 642, + 546, + 707 + ], + "spans": [ + { + "bbox": [ + 308, + 642, + 546, + 707 + ], + "type": "text", + "content": "[49] Walter Zimmer, Jialong Wu, Xingcheng Zhou, and Alois C Knoll. Real-time and robust 3d object detection with roadside lidars. In Proceedings of the 12th International Scientific Conference on Mobility and Transport: Mobility Innovations for Growing Megacities, pages 199-219. Springer, 2023. 4" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22677" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/992f5837-40b9-47fe-b9bd-9d2e369d7fdb_content_list.json b/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/992f5837-40b9-47fe-b9bd-9d2e369d7fdb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..039246242e9b38d6b72c07e4ee3dce2a17d8f9fd --- /dev/null +++ b/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/992f5837-40b9-47fe-b9bd-9d2e369d7fdb_content_list.json @@ -0,0 +1,1696 @@ +[ + { + "type": "text", + "text": "Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models", + "text_level": 1, + "bbox": [ + 78, + 130, + 893, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pengze Zhang1* Hubery Yin2* Chen Li2 Xiaohua Xie1† \n1Sun Yat-sen University Wechat, Tencent Inc.", + "bbox": [ + 240, + 179, + 736, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhangpz3@mail2.sysu.edu.cn, {hubery, chaselli}@tencent.com, xiexiaoh6@mail.sysu.edu.cn https://pangzecheung.github.io/SingDiffusion/", + "bbox": [ + 104, + 219, + 864, + 251 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/abbc4ae39ef5fb733708e23c1e2760916cab7246fe7d417f5408e108aa3032b2.jpg", + "image_caption": [ + "Figure 1. In this paper, we explore singularities theoretically and propose a plug-and-play module SingDiffusion to address the sampling challenge at the initial singular time step. By integrating this module into existing pre-trained models, our approach effectively tackles the difficulties of generating both dark and bright images, and further enhancing overall image quality as confirmed by quantitative analysis." + ], + "image_footnote": [], + "bbox": [ + 81, + 265, + 885, + 577 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 657, + 313, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most diffusion models assume that the reverse process adheres to a Gaussian distribution. However, this approximation has not been rigorously validated, especially at singularities, where $t = 0$ and $t = 1$ . Improperly dealing with such singularities leads to an average brightness issue in applications, and limits the generation of images with extreme brightness or darkness. We primarily focus on tackling singularities from both theoretical and practical perspectives. Initially, we establish the error bounds for the reverse process approximation, and showcase its Gauss-", + "bbox": [ + 75, + 691, + 470, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sian characteristics at singularity time steps. Based on this theoretical insight, we confirm the singularity at $t = 1$ is conditionally removable while it at $t = 0$ is an inherent property. Upon these significant conclusions, we propose a novel plug-and-play method SingDiffusion to address the initial singular time step sampling, which not only effectively resolves the average brightness issue for a wide range of diffusion models without extra training efforts, but also enhances their generation capability in achieving notable lower FID scores.", + "bbox": [ + 500, + 660, + 893, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 843, + 632, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models, generating samples from initial noise by learning a reverse diffusion process, have achieved remark-", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. This work was done when Pengze Zhang was an intern at WeChat. \n†Corresponding Author.", + "bbox": [ + 76, + 862, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6945", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "able success in multi-modality content generation, such as image generation [4, 5, 26, 30, 35, 46], audio generation [18, 26, 28], and video generation [14]. These achievements owe much to several fundamental theoretical research, namely Denoising Diffusion Probabilistic Modeling (DDPM) [13, 38], Stochastic Differential Equations (SDE) [40], and Ordinary Differential Equations (ODE) [39, 40]. These approaches are all based on the assumption that the reverse diffusion process shares the same functional form as the forward process. Although an indirect validation of this assumption is given by Song et al. [40], it heavily relies on the existence of solutions to the Kolmogorov forward and backward equations [1], which encounters singularities at the endpoints of time intervals where $t = 0$ and $t = 1$ .", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The singularity issue is not only a gap in the theoretical formulation of current diffusion models, but also affects the quality of the content they generate. Current applications simply ignore singularity points in their implementation [27, 31, 42] and restrict the time interval to $[\\varepsilon_1, 1 - \\varepsilon_2]$ . As a result, the average brightness of the generated images typically hovers around 0 [9, 19] (normalizing brightness to $[-1, 1]$ ). For example, as shown in Fig. 1, existing pretrained diffusion models, such as Stable Diffusion 1.5 (SD-1.5) [31] and Stable Diffusion 2.0-base (SD-2.0-base), fail in generating images with pure white or black backgrounds. To address this challenge, Guttenberg et al. [9] add extra offset noise during the training process to allow the network could learn the overall brightness changes of the image. Unfortunately, the offset noise usually disrupts the pre-defined marginal probability distribution and further invalidates the original sampling formula. Lin et al. [19] re-scales the noise schedule to enforce a zero terminal signal-to-noise ratio, and employ the $v$ -prediction technique [36] to circumvent the issue of division by zero at $t = 1$ . However, this method only supports models in $v$ -prediction manner and requires substantial training to fine-tune the entire model. Therefore, it is advantageous to devise a plug-and-play method which effectively deals with the singularity issue for any practicable diffusion model without extra training efforts.", + "bbox": [ + 75, + 306, + 472, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we begin with a theoretical exploration of the singularity issue at the endpoints of time intervals in diffusion model, and then devise a novel plug-and-play module to address the accompanying average brightness problem in image generation. Through establishing mathematical error bounds rigorously, we first prove the approximate Gaussian characteristics of the reverse diffusion process at all sampling time steps, especially where the singularity issue appears. Following this, we conduct a thorough analysis of this approximation in the vicinity of singularities, and arrive at two significant conclusions: 1) by computing the limit of 'zero divided by zero' at $t = 1$ , we confirm the corresponding initial singularity is removable; 2) the singularity at $t = 0$ is an inherent property of diffusion models, thus", + "bbox": [ + 75, + 689, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we should adhere to this form rather than simply avoiding the singularity. Following the aforementioned theoretical analysis, we propose the SingDiffusion method, specifically tailored to address the challenge of sampling during the initial singularity. Especially, this method can seamlessly integrate into existing sampling processes in a plug-and-play manner without requiring additional training efforts.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As demonstrated in Fig. 1 and experiments in the appendix, our novel plug-and-play initial sampling step can effectively resolve the average brightness issue. It can be easily applied to a wide range of diffusion models, including the SD-1.5 and SD-2.0-base with $\\epsilon$ -prediction, SD-2.0 with $v$ -prediction, and various pre-trained models available on the CIVITAI website1, thanks to its one-time training strategy. Furthermore, our method can generally enhance the generative capabilities of these pre-trained diffusion models in notably improving the FID [11] scores at the same CLIP [29] level on the COCO dataset [20].", + "bbox": [ + 496, + 198, + 892, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 375, + 640, + 390 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Reverse process approximation", + "text_level": 1, + "bbox": [ + 500, + 400, + 774, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Denoising Diffusion Probability Models (DDPM) [13, 38] establish a hand-designed forward Markov chain in discrete time, and model the distribution of data in the reverse process. In contrast, Song et al. [40] establish the diffusion model in continuous time, framed as a Stochastic Differential Equation (SDE). Moreover, they reveal an Ordinary Differential Equation (ODE) termed 'probability flow', sharing the same single-time marginal distribution as the original SDE. Notably, ODE also has discrete counterparts known as Denoising Diffusion Implicit Models (DDIM) [39].", + "bbox": [ + 496, + 422, + 890, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The assumption that the reversal of the diffusion process has an identical functional form to the forward process is fundamental in the aforementioned methods. Several studies aim to prove this assumption. Song et al. [40] indirectly substantiate this in continuous time by introducing a reverse SDE. Nevertheless, they do not provide error bounds in discrete-time cases, leaving this assumption unverified at discrete-time steps. Additionally, the treatment of singularities at $t = 0$ and $t = 1$ is not addressed in their work. McAllester et al. [24] provide proof in discrete time, showing the density of the reverse process as a mixture of Gaussian distributions. Despite this, it doesn't qualify as a pure Gaussian distribution. In contrast, to fill this theoretical gap, our approach directly substantiates this assumption by establishing error bounds for the reverse process approximation at both non-singular and singular time steps.", + "bbox": [ + 496, + 575, + 892, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Singularities in diffusion model", + "text_level": 1, + "bbox": [ + 500, + 825, + 777, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several studies have focused on investigating the singularity occurring at $t = 0$ . Song et al. [40] attempt to bypass", + "bbox": [ + 496, + 848, + 890, + 878 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 https://civitai.com/", + "bbox": [ + 517, + 887, + 627, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6946", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "this singularity by initiating their analysis at $t = \\varepsilon > 0$ instead of $t = 0$ . Therefore, this approach didn't effectively address the core singularity problem. Dockhorn et al. [6] propose a diffusion model on the product space of position and velocity, and avoid the singularity through hybrid score matching. Nevertheless, this approach lacks compatibility with DDPM, SDE and ODE due to the incorporation of the velocity space. Lu et al. [22] employ the $x$ -prediction method to mitigate singularity during training, but did not address the singularity issue during the sampling process.", + "bbox": [ + 75, + 90, + 468, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, a comprehensive solution to the singularity at $t = 0$ is still pending. Moreover, the singularity at $t = 1$ remains unexplored. To tackle these issues, we thoroughly investigate the sampling process at both $t = 0$ and $t = 1$ , and offer theoretical solutions.", + "bbox": [ + 75, + 242, + 468, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Average brightness issue", + "text_level": 1, + "bbox": [ + 76, + 325, + 302, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models have shown significant comprehensive quality and controllability in computer vision, including text-to-image generation [5, 12, 26, 30, 31, 35], image editing [3, 10, 16, 25, 33, 43], image-to-image translation [34, 41, 44], surpassing previous generative models [7, 8, 23, 45]. However, most existing diffusion models ignore the sampling at the initial singular time step, resulting in the inability to generate bright and dark images, i.e., the average brightness issue. Adding offset noise [9] and employing $v$ -prediction [19, 36] are two ways to tackle this problem. However, these methods require fine-tuning for each existing model, consuming a substantial amount of time and limiting their applicability. In contrast, we propose a novel plug-and-play solution targeting the core of the average brightness issue, i.e., the singularity at the initial time step. Our method not only empowers the majority of existing pre-trained models to effectively generate images with the desired brightness level, but also significantly enhances their generative capabilities.", + "bbox": [ + 75, + 348, + 470, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 647, + 168, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To facilitate the clarity of our exploration into singularities from theoretical and practical angles, this section is organized as follows: 1) We start by introducing background and symbols in the Preliminaries. 2) Next, we derive error bounds for the reverse process approximation, confirming its Gaussian characteristics at both regular and singular time steps. 3) We then theoretically analyze and handle the sampling at singular time steps, i.e., $t = 0$ and $t = 1$ . 4) Lastly, based on our previous analysis, we propose a plug-and-play method to address initial singularity time step sampling, effectively resolving the average brightness issue.", + "bbox": [ + 75, + 672, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 847, + 217, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the realm of generative models, we are consistently provided with a set of training samples denoted as $\\{y_{i}\\in$", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathbb{R}^d\\}_{i = 1}^N$ , which are inherently characterized by a distribution given by [15]:", + "bbox": [ + 498, + 89, + 890, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np (x, t = 0) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\delta (x - y _ {i}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 132, + 890, + 172 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta (x)$ denotes the Dirac delta function.", + "bbox": [ + 500, + 185, + 794, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a continuous-time Gaussian diffusion process within the interval $0 \\leq t \\leq 1$ . Following [17], the distribution of $x_{t}$ conditioned on $x_{0}$ is written by $p(x_{t} | x_{0}) = \\mathcal{N}(\\alpha_{t} x_{0}, \\sigma_{t}^{2} I)$ , where $\\alpha_{t}$ and $\\sigma_{t}$ are positive scalar functions of $t$ satisfying $\\alpha_{t}^{2} + \\sigma_{t}^{2} = 1$ , and $\\alpha_{t}$ decreases monotonically from 1 to 0 over time $t$ ; the distribution of $x_{t}$ conditioned on $x_{s}$ is represented as $p(x_{t} | x_{s}) = \\mathcal{N}(\\alpha_{t | s} x_{s}, \\sigma_{t | s}^{2} I)$ , where $0 \\leq s < t \\leq 1$ , $\\alpha_{t | s} = \\alpha_{t} / \\alpha_{s}$ , $\\sigma_{t | s}^{2} = 1 - \\alpha_{t | s}^{2}$ . Consequently, the forward process is derived as follows:", + "bbox": [ + 496, + 200, + 890, + 352 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\alpha_ {t | s} x _ {s} + \\sqrt {1 - \\alpha_ {t | s} ^ {2}} z _ {s}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 362, + 890, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the set $\\{z_t\\}_{t=0}^1$ comprises independent standard Gaussian random variables.", + "bbox": [ + 496, + 398, + 890, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a discrete-time diffusion process with $T$ steps, the time $i\\in \\{0,1,\\dots T\\}$ corresponds to $t$ in the continuous case as $i = t\\times T$ . Defining $\\hat{\\beta}_i = 1 - \\hat{\\alpha}_{i|i - 1}^2$ , where $\\hat{\\alpha}_{i|i - 1} = \\alpha_{i / T|(i - 1) / T}$ , and $\\hat{\\mathbf{\\Gamma}}^*$ denotes the symbol in the discrete-time process, the forward process in Eq. 2 can be rewritten as:", + "bbox": [ + 496, + 429, + 890, + 518 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {i} = \\sqrt {1 - \\hat {\\beta} _ {i}} \\hat {x} _ {i - 1} + \\sqrt {\\hat {\\beta} _ {i}} \\hat {z} _ {i - 1}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 518, + 890, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which is equivalent to the forward process outlined in [13].", + "bbox": [ + 498, + 549, + 888, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Taking into account the initial distribution (Eq. 1), the single-time marginal distribution of $x_{t}$ is given by:", + "bbox": [ + 498, + 564, + 890, + 594 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(x _ {t}, t\\right) = \\frac {1}{N} \\sum_ {i} \\left(2 \\pi \\sigma_ {t} ^ {2}\\right) ^ {- \\frac {d}{2}} \\exp \\left(- \\frac {\\left(x _ {t} - \\alpha_ {t} y _ {i}\\right) ^ {2}}{2 \\sigma_ {t} ^ {2}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 604, + 890, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d$ is the dimension of the training samples. As a result, the reverse process can be derived using Bayes' rule:", + "bbox": [ + 498, + 654, + 890, + 684 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np (x _ {s} | x _ {t}) = p (x _ {t} | x _ {s}) \\frac {p (x _ {s} , s)}{p (x _ {t} , t)} = (2 \\pi \\sigma_ {s | t}) ^ {- \\frac {d}{2}}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 695, + 794, + 728 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {i} \\exp \\left(- \\frac {1}{2 \\sigma_ {s | t} ^ {2}} \\left(x _ {s} - \\frac {\\alpha_ {t | s} \\sigma_ {s} ^ {2} x _ {t}}{\\sigma_ {t} ^ {2}} - \\frac {\\alpha_ {s} \\sigma_ {t | s} ^ {2} y _ {i}}{\\sigma_ {t} ^ {2}}\\right) ^ {2}\\right) w _ {i} \\left(x _ {t}, t\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 729, + 890, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sigma_{s|t}^2 = \\sigma_{t|s}^2\\frac{\\sigma_s^2}{\\sigma_t^2}$ , and $w_{i}(x_{t},t) = \\frac{\\exp(-\\frac{(x_{t} - \\alpha_{t}y_{i})^{2}}{2\\sigma_{t}^{2}})}{\\sum_{j}\\exp(-\\frac{(x_{t} - \\alpha_{t}y_{j})^{2}}{2\\sigma_{t}^{2}})}$", + "bbox": [ + 500, + 782, + 888, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Error bound estimation", + "text_level": 1, + "bbox": [ + 500, + 832, + 718, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing diffusion models, such as [13, 38] are based on an assumption that the reverse process in Eq. 5 can be approximated by a Gaussian distribution, when $\\hat{\\beta}_i$ is small, as given", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6947", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "by:", + "bbox": [ + 76, + 92, + 102, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {p} (x _ {s} | x _ {t}) = \\left(2 \\pi \\sigma_ {s | t} ^ {2}\\right) ^ {- \\frac {d}{2}} \\\\ \\exp \\left(- \\frac {1}{2 \\sigma_ {s | t} ^ {2}} \\left(x _ {s} - \\frac {\\alpha_ {t | s} \\sigma_ {s} ^ {2} x _ {t}}{\\sigma_ {t} ^ {2}} - \\frac {\\alpha_ {s} \\sigma_ {t | s} ^ {2} \\bar {y} \\left(x _ {t} , t\\right)}{\\sigma_ {t} ^ {2}}\\right) ^ {2}\\right), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 113, + 467, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{y}(x_t, t) = \\sum_i w_i(x_t, t)y_i$ . However, these studies did not furnish the error bounds to support this assumption. To address this theoretical gap, we estimate the error bound as follows.", + "bbox": [ + 76, + 189, + 467, + 248 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 1 (Error Bound Estimated by $\\sigma_{s|t}$ ). $\\forall s \\in (0,1)$ , $\\exists \\tau \\in (s,1)$ and $C > 0$ , such that $\\forall t \\in (s,\\tau]$ , $\\int_{\\mathbb{R}^d} |p(x_s|x_t) - \\tilde{p}(x_s|x_t)| dx_s < C \\sqrt{\\sigma_{s|t}}$ .", + "bbox": [ + 76, + 257, + 467, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 1 demonstrates that when $\\sigma_{s|t}$ is small, the forward and reverse processes share the same form, i.e., Gaussian distribution. Since $\\hat{\\beta}_i = \\frac{\\sigma_{i / T}}{\\sigma_{(i - 1) / T}}\\sigma_{(i - 1) / T|i / T}$ features a term in the $\\sigma_{s|t}$ form, also supports the inference that this assumption remains valid when $\\hat{\\beta}_i$ is small, as highlighted in [13, 38]. It is worth noting that this error is bounded by $\\sigma_{s|t}$ instead of $\\hat{\\beta}_i$ strictly.", + "bbox": [ + 76, + 310, + 467, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the error bound estimated by $\\sigma_{s|t}$ in Proposition 1 is not sufficient to prove the assumption at the singularity time step $t = 1$ . The reason is that when $t = 1$ , $\\sigma_{s|t}$ approaches 1 as $s \\to 1$ , which is not a small value. Consequently, the error bound at $t = 1$ in Proposition 1 remains non-negligible.", + "bbox": [ + 76, + 420, + 467, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To tackle this issue, we instead utilize $\\alpha_{s}$ to bound the error at time step $t = 1$ , and present a new proposition:", + "bbox": [ + 76, + 511, + 467, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 2 (Error Bound Estimated by $\\alpha_{s}$ ). $\\exists \\nu \\in (0,1)$ and $C > 0$ , such that $\\forall \\nu \\leq s < t \\leq 1$ , $\\int_{\\mathbb{R}^d} |p(x_s|x_t) - \\tilde{p}(x_s|x_t)| dx_s < C \\sqrt{\\alpha_s}$ .", + "bbox": [ + 76, + 547, + 467, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to Proposition 2, setting $t = 1$ , it has $\\alpha_s \\to 0$ as $s \\to 1$ . As a result, the error bound at $t = 1$ assures a small value, affirming the validation of the Gaussian approximation assumption at $t = 1$ .", + "bbox": [ + 76, + 601, + 467, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In sum, through Proposition 1 and 2, we have established that the reverse process of the diffusion model can be approximated by Gaussian distribution across all time steps.", + "bbox": [ + 76, + 662, + 467, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Tackling the singularities", + "text_level": 1, + "bbox": [ + 76, + 715, + 305, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the theoretical foundation provided in Section 3.2, we can delve into the analysis of singular time steps using Eq. 6. It is worth noting that this section will mainly focus on addressing the singularities present in the discrete-time diffusion model [13], while the treatment of the continuous case [40] will be deferred to the appendix.", + "bbox": [ + 76, + 738, + 467, + 828 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 The singularities at $t = 1$", + "text_level": 1, + "bbox": [ + 76, + 845, + 303, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Drawing from Eq. 6, a straightforward way is to train a neural network $\\bar{y}_{\\theta}(x_t,t)$ for estimating $\\bar{y} (x_{t},t)$ , known", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as $x$ -prediction. This approach ensures that the approximated reverse process avoids encountering a singularity at $t = 1$ . However, for stable training [38], the mainstream choice among current diffusion models is $\\epsilon$ -prediction. It utilizes a neural network $\\epsilon_{\\theta}(x_t,t)$ for estimating $\\epsilon (x_{t},t) = \\frac{x_{t} - \\alpha_{t}\\bar{y} (x_{t},t)}{\\sigma_{t}}$ , which will encounter singularity. More specifically, substituting $\\epsilon (x_{t},t)$ for $\\bar{y} (x_{t},t)$ yields the equation:", + "bbox": [ + 500, + 90, + 890, + 198 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {s} = \\frac {\\alpha_ {t | s} \\sigma_ {s} ^ {2}}{\\sigma_ {t} ^ {2}} x _ {t} + \\frac {\\alpha_ {s} \\sigma_ {t | s} ^ {2}}{\\sigma_ {t} ^ {2}} \\frac {x _ {t} - \\sigma_ {t} \\epsilon (x _ {t} , t)}{\\alpha_ {t}} + \\sigma_ {s | t} z _ {t}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 209, + 890, + 246 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At $t = 1$ , the denominator $\\alpha_{t}$ becomes 0, resulting in a division-by-zero singularity. Theoretically, this singularity is removable because the limit exists:", + "bbox": [ + 498, + 255, + 890, + 299 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\lim _ {t \\rightarrow 1 ^ {-}} \\frac {x _ {t} - \\sigma_ {t} \\epsilon (x _ {t} , t)}{\\alpha_ {t}} = \\lim _ {t \\rightarrow 1 ^ {-}} \\bar {y} (x _ {t}, t) = \\frac {1}{N} \\sum_ {i} y _ {i}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 309, + 890, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regrettably, computing this limit during inference is unfeasible, since $\\epsilon(x_1, 1) = x_1$ loses all information of the correct sampling direction. Conversely, $\\bar{y}(x_t, t)$ retains the correct sampling direction for all $t \\in [0, 1]$ . Particularly at $t = 1$ , $\\bar{y}(x_1, 1) = \\frac{1}{N} \\sum_i y_i$ encapsulates the information for the initial inference step. Therefore, leveraging $x$ -prediction at the initial time step proves more advantageous than $\\epsilon$ -prediction.", + "bbox": [ + 498, + 354, + 890, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 The singularities at $t = 0$", + "text_level": 1, + "bbox": [ + 500, + 494, + 725, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When $s = 0$ and $t$ is small, the distribution in Eq. 6 degenerates into a singular distribution, i.e., a Gaussian with zero variance, resembling a Dirac delta:", + "bbox": [ + 500, + 518, + 890, + 563 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p} \\left(x _ {0} \\mid x _ {t}\\right) = \\delta \\left(x _ {0} - y _ {j _ {0}}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 575, + 890, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $j_0 = \\arg \\min_j |x_t - \\alpha_t y_j|$ . This singularity directs the sampling process to converge at the correct point $y_{j_0} = \\bar{y}(x_0, 0)$ . Therefore, the singularity at $t = 0$ is an inherent characteristic of diffusion models that do not require avoidance as long we use suitable sampling techniques. For instance, in the final step of the original DDPM sampling method, it has $x_0 = \\bar{y}(x_t, t)$ . When $t$ is small, $\\bar{y}(x_t, t) \\approx \\bar{y}(x_0, 0)$ , making this process equivalent to Eq. 9. Thus, there is no need to avoid the singularity.", + "bbox": [ + 498, + 602, + 890, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides, we also arrived at this conclusion within the continuous diffusion model, i.e., SDE. More detailed elaboration on this topic can be found in the appendix.", + "bbox": [ + 498, + 747, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. SingDiffusion", + "text_level": 1, + "bbox": [ + 500, + 801, + 640, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to theoretical issues, the singular sampling issue can also cause average brightness issue in applications, as depicted in Fig. 2. This is mainly because most of the existing method sample images starting at time step $1 - \\varepsilon$ using a standard Gaussian distribution, which significantly", + "bbox": [ + 498, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6948", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e037e502645f5b1640c7e7d85edb67b67d942c2b91be2fa47f58febb8dc4e729.jpg", + "image_caption": [ + "Figure 2. Illusion of the gap between the sampling from $x_{1 - \\varepsilon}$ and $x_{1}$ . Due to the lack of consideration of singular time step sampling in most of the existing methods, they will encounter the average brightness issue. To tackle this, we propose a plug-and-play SingDiffusion method (highlighted in red) to bridge this gap." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 472, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "diverges from the true distribution $p(x_{1 - \\varepsilon}, 1 - \\varepsilon)$ . To validate this, we select two images (a black dog against a black background and a white dog against a white background). Then we diffuse them for a time period of $1 - \\varepsilon$ respectively, and calculate the mean value of their latent code. It can be seen that the mean of these two latent codes are -0.14 and +0.28, which are significantly different from the samples obtained from the standard Gaussian distribution. Moreover, as evident from the visualization in Fig. 2, the latent code $(x_{1 - \\varepsilon})$ of the black dog and white dog are notably darker and brighter compared to $x_{1}$ from the Gaussian distribution respectively. Under such distributional differences, according to Proposition 3, employing a Gaussian distribution at $t = 1 - \\epsilon$ is equivalent to generating images towards an average brightness of 0 at $t = 1$ (with grayscale ranging from [-1, 1]). Consequently, current methods encounter challenges in generating dark or bright images.", + "bbox": [ + 75, + 422, + 472, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 3. Setting $x_{1 - \\epsilon} \\sim \\mathcal{N}(0, I)$ is equivalent to sampling the value from standard Gaussian as $\\bar{y}(x_1, 1)$ at $t = 1$ .", + "bbox": [ + 75, + 691, + 470, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on our analysis of sampling at singular time steps in Section 3.3.1, we propose a novel plug-and-play method SingDiffusion to fill the gap at $t = 1$ , thus solving the problem of average brightness issue. Considering a pre-trained model $\\epsilon_{\\theta}(x_t,t)$ afflicted by the singularity due to division by zero, we proceed to train a model using $x$ -prediction at $t = 1$ . The algorithm of our training and sampling process are shown in Algorithm 1 and Algorithm 2. Firstly, for image-prompt data pairs $(x_0,c)$ in the training process, we use a U-net [32] $\\bar{y}_{\\theta}$ to fit $\\bar{y} (x_1,t = 1,c)$ , where $\\bar{y} (x_{t},t,c)$", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Training process of SingDiffusion" + ], + "code_body": "1: repeat \n2: $x_0, c \\sim p(x_0, c)$ , $x_1 \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ \n3: Take gradient descent step on $\\nabla_\\theta \\| \\bar{y}_\\theta(x_1, c) - x_0 \\|^2$ \n4: until converged \nAlgorithm 2 Sampling process of SingDiffusion \n1: $x_1 \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ \n2: $\\varepsilon = 1 / T$ \n3: $x_{1 - \\varepsilon} = \\alpha_{1 - \\varepsilon} \\bar{y}_\\theta(x_1, c) + \\sigma_{1 - \\varepsilon} x_1$ \n4: for $t = 1 - \\varepsilon, \\dots, \\varepsilon$ do \n5: Calculate $x_{t - \\varepsilon}$ using existing sampling algorithms \n6: end for \n7: return $x_0$", + "bbox": [ + 509, + 108, + 890, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is the extension of $\\bar{y}(x_t, t)$ defined in Section 3.2 under the condition $c$ . The loss function can be written as:", + "bbox": [ + 498, + 335, + 890, + 366 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {x _ {0}, c \\sim p (x _ {0}, c), x _ {1} \\sim \\mathcal {N} (\\mathbf {0}, I)} | | \\bar {y} _ {\\theta} (x _ {1}, c) - x _ {0} | | ^ {2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 372, + 890, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As our model is only amortized with respect to text embeddings and not the time step, we omit the variable $t$ for $\\bar{y}_{\\theta}$ and the U-net does not necessitate a complex architecture.", + "bbox": [ + 498, + 397, + 890, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the sampling process, we use the new model $\\bar{y}_{\\theta}$ in the initial time-step with a DDIM scheduler:", + "bbox": [ + 498, + 443, + 890, + 472 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {1 - \\varepsilon} = \\alpha_ {1 - \\varepsilon} \\bar {y} _ {\\theta} (x _ {1}, c) + \\sigma_ {1 - \\varepsilon} x _ {1}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 479, + 890, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This equation guarantees $x_{1 - \\varepsilon}$ adheres to the distribution $p(x_{1 - \\varepsilon}, 1 - \\varepsilon)$ . Following this, we utilize the existing pre-trained model $\\epsilon_{\\theta}(x_t, t, c) / v_{\\theta}(x_t, t, c)$ in $\\epsilon$ -prediction/v-prediction manner to perform the subsequent sampling steps until it generates $x_0$ . It's worth noting that our method is solely involved in the sampling at $t = 1$ , independent of the subsequent sampling process. Consequently, our approach can be once-trained and seamlessly integrated into the majority of diffusion models.", + "bbox": [ + 496, + 503, + 890, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For further improving matching between generated images and input prompts, existing diffusion models typically incorporate classifier-free techniques [26]:", + "bbox": [ + 498, + 638, + 890, + 684 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\no _ {\\text {g u i d a n c e}} = o _ {n e g} + w \\times \\left(o _ {\\text {p o s}} - o _ {n e g}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 691, + 890, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w \\geq 1$ represents the guidance scale, $o$ signifies the output of the diffusion model which can be either $\\bar{y}_{\\theta}$ , $\\epsilon_{\\theta}$ or $v_{\\theta}$ , 'pos' and 'neg' refer to the outputs corresponding to positive and negative prompts respectively. Nevertheless, we notice that when applying this technique at the initial singular time step, the influence of the guidance scale $w$ predominates the results due to the greater directional difference between the negative and positive outputs. To tackle this challenge, we implement a straightforward yet highly effective normalization method:", + "bbox": [ + 496, + 714, + 890, + 864 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {y} _ {\\theta_ {g u i d a n c e}} = \\left[ \\bar {y} _ {\\theta_ {n e g}} + w \\times \\left(\\bar {y} _ {\\theta_ {p o s}} - \\bar {y} _ {\\theta_ {n e g}}\\right) \\right] / w. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 872, + 890, + 890 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6949", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/28e8ccf56423d542704fa8384572c5cfefa5ea8644151d80832b2cb6222e7e75.jpg", + "image_caption": [ + "Figure 3. Comparison of stable diffusion models and SingDiffusion on average brightness issue." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 483, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ae8a6023c25fae9ade665d51f20f6a45f9d77d5be088fa94867cab7bb6bef7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 88, + 890, + 272 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6abdc21c767e15cfe3e79e26587d0f1f3dcb0e67bedaff2900c760d64e205e8c.jpg", + "image_caption": [ + "Figure 4. Comparison of Pareto curves between SingDiffusion, SD-1.5, and SD-2.0-base on 30k COCO images, across various guidance scales in [1.5, 2, 3, 4, 5, 6, 7, 8]." + ], + "image_footnote": [], + "bbox": [ + 89, + 311, + 472, + 459 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7001ca5bf4283e58972562adea7f1031d4116e48dccd8723494eead4b34ce5df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 311, + 885, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 525, + 200, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implement details", + "text_level": 1, + "bbox": [ + 76, + 551, + 251, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To create a versatile plug-and-play model, our model is trained on the Laion2B-en dataset [37], including 2.32 billion text-image pairs. The images are center-cropped to 512 $\\times$ 512. Our $\\bar{y}_{\\theta}$ follows the U-net structure similar to stable diffusion models but with reduced parameters, totaling 140 million. We utilize the AdamW optimizer [21] with a learning rate of 1e-4 to train our $\\bar{y}_{\\theta}$ . The training is executed across 64 Nvidia V100 chips with a batch size of 3072, and completes after 180K steps.", + "bbox": [ + 75, + 575, + 468, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the testing process, all steps, including our initial sampling, are carried out using the default guidance scale of $w = 7.5$ . After our initial sampling at the singular time step, all existing pre-trained diffusion models generate images in the DDIM pipeline, executing 50 steps to produce images following a schedule with a total time of $T = 1000$ .", + "bbox": [ + 75, + 712, + 470, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Average brightness issue", + "text_level": 1, + "bbox": [ + 76, + 815, + 302, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the effectiveness of SingDiffusion in addressing the average brightness issue, we select four extreme prompts, including \"Solid white/black background\", and \"Monochrome line-art logo on a white/black background\".", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b54d322ff8169a8d23c8c89c1e205a9924d087de8ef6656e2fceac5ee232ffcc.jpg", + "table_caption": [ + "Table 1. Comparison of average brightness of 100 generated images between stable diffusion models and our SingDiffusion under different prompt conditions. For 'white'/'black' prompts, higher/ lower average brightness is better." + ], + "table_footnote": [], + "table_body": "
Model"Solid white, background""Solid black, background""Monochrome line-art logo on a white background""Monochrome line-art logo on a black background"
SD-1.5141.4383.09137.95113.66
+ Ours212.593.04223.6811.52
SD-2.0-base150.5299.67136.13104.45
+ Ours227.430.29228.6810.87
", + "bbox": [ + 503, + 592, + 888, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each prompt, we generate 100 images using SD-1.5, SD-2.0-base and SingDiffusion methods, and then calculate their average brightness. The results are shown in Table 1. It is remarkably clear that the stable diffusion methods, whether using prompts with 'black' or 'white' descriptors, tend to generate images with average brightness. However, SingDiffusion, implementing initial singularity sampling, effectively corrects the average brightness of the generated images. For example, under the prompt \"solid black background,\" our method notably lowers the average brightness from 99.67 to 0.29 for images generated by SD-2.0-base.", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6950", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/66cbaa03e2abf489957b7a0e731fd4747bc5afa0fe42b418a0148728ca99ed37.jpg", + "image_caption": [ + "Figure 5. Our method can be trained once and seamlessly integrated into the pre-trained models on CIVITAI in a plug-and-play fashion." + ], + "image_footnote": [], + "bbox": [ + 80, + 95, + 885, + 406 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4d344df5c18aec15bfb8251a9dc2fe47e03930c0b42472b0af7463d866823d57.jpg", + "table_caption": [ + "Table 2. Comparison of stable diffusion model and SingDiffusion on FID score and CLIP score without classifier guidance." + ], + "table_footnote": [], + "table_body": "
ModelSD-1.5SD-2.0-base
FID ↓CLIP ↑FID ↓CLIP ↑
Original31.8626.7025.1727.48
+ SingDiffusion21.0927.7118.0128.23
", + "bbox": [ + 78, + 497, + 464, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, we also provide visualization results for these prompts in Fig. 3. It can be seen that images generated by stable diffusion methods predominantly exhibit a gray tone. In contrast, our SingDiffusion method successfully overcomes this issue, and is capable of generating both dark and bright images.", + "bbox": [ + 75, + 595, + 468, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Improvement on image quality", + "text_level": 1, + "bbox": [ + 76, + 696, + 349, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate our improvements in general image quality, following [35], we randomly select 30k prompts from the COCO dataset [20] as the test set. We employ two metrics for evaluation, including the FID score and CLIP score. Specifically, Fréchet Inception Distance (FID) [11] calculates the Fréchet distance between the real data and the generated data. A lower FID implies more realistic generated data. While the Contrastive Language-Image Pre-training (CLIP) [29] score measures the similarity between the generated images and the given prompts. A higher CLIP score means the generated images better match the input prompts.", + "bbox": [ + 75, + 719, + 468, + 885 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First of all, we compare SingDiffusion with SD-1.5 and", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SD-2.0-base in Table 2 to gauge the model's inherent fitting capability, without using guidance-free techniques. It is evident that SingDiffusion significantly outperforms the existing stable diffusion methods in both FID and CLIP scores. These results highlight the yet-to-be-utilized fitting potential in current stable diffusion models, emphasizing the significance of sampling at the initial singular time step.", + "bbox": [ + 496, + 460, + 890, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, inspired by Imagen [35], we plot CLIP v.s. FID Pareto curves by varying guidance values within the range [1.5, 2, 3, 4, 5, 6, 7, 8] in Fig. 4. SingDiffusion exhibits substantial improvements over stable diffusion models, especially noticeable with smaller guidance scales. As the guidance scale increases, SingDiffusion consistently maintains a lower FID compared to stable diffusion for achieving a similar CLIP score. This emphasizes that our approach not only enhances image realism but also ensures better adherence to the input prompts.", + "bbox": [ + 496, + 566, + 890, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Plug-and-play on other pre-trained models", + "text_level": 1, + "bbox": [ + 500, + 726, + 864, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since our method is extensively trained on the Laion dataset, it can be easily integrated into the majority of existing diffusion models. To validate this, we download several pre-trained models from the CIVITAI website, each specializing in different image domains like anime, animals, clothing, and so on. To facilitate the application of SingDiffusion to these models, we integrate SingDiffusion into the popular Stable Diffusion Web UI [2] system, and sample images with the Euler ancient method after using our initial singularity sampling process. The outcomes, displayed", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6951", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78edc379fc152b6c463ad7bc395ab648d2c66c58beb0bb5d589ab5bb62738140.jpg", + "image_caption": [ + "Figure 6. Ablation study of guidance normalization on SD-1.5 across various guidance scales in [1.5, 2, 3, 4, 5, 6, 7, 8]." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 468, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/97b1e9f6057b36567c309b4036c9656ed998006d5fee2ec5cbb3839e4f57c4cb.jpg", + "image_caption": [ + "Figure 7. Visualization of $\\bar{y}_{\\theta}$ for various prompts." + ], + "image_footnote": [], + "bbox": [ + 81, + 273, + 468, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in Fig. 1 and Fig. 5, showcase SingDiffusion effectively resolving the average brightness issue across all models while preserving their original generative capacities. This demonstrates the adaptability and practicality of our approach.", + "bbox": [ + 75, + 469, + 468, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Effect on classifier guidance normalization", + "text_level": 1, + "bbox": [ + 76, + 540, + 439, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct an ablation study on the guidance normalization operation (Eq. 13), and represent the CLIP v.s. FID Pareto curves in Fig. 6. It can be seen that the method without normalization exhibits inferior results compared with the full method. As the guidance scale increases, the gap in FID between these two methods gradually widens. This phenomenon primarily arises from significant disparities between $\\bar{y}_{\\theta_{pos}}$ and $\\bar{y}_{\\theta_{neg}}$ . According to Eq. 13, larger guidance scales may lead to overflow issues. In contrast, normalizing the results with the guidance scale helps keep the outputs within a typical range, thus restoring the FID score.", + "bbox": [ + 75, + 564, + 468, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Visualization of the $\\bar{y}_{\\theta}$", + "text_level": 1, + "bbox": [ + 76, + 741, + 281, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "According to Eq. 10, our main goal with $\\bar{y}_{\\theta}$ is to model the average image corresponding to each prompt. To confirm this, we employ the prompts presented in Fig. 1 and visualize their corresponding $\\bar{y}_{\\theta}$ . As demonstrated in Fig. 7, it is clear that $\\bar{y}_{\\theta}$ does indeed represent a smoothed average image. For instance, the first four images of the second row resemble average faces, aligning with the input prompt for generating celebrities. Additionally, we notice that $\\bar{y}_{\\theta}$ adapts to the prompt's brightness suggestion,", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96a352894671c981cbe9a42791bbef21584196738259881322cf35fa771ab72f.jpg", + "image_caption": [ + "Figure 8. SingDiffusion integrates seamlessly with ControlNet." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 893, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "appearing brighter or darker accordingly. This highlights our method's ability to effectively capture pertinent lighting conditions, and underscores the significance of the initial singular time step sampling.", + "bbox": [ + 498, + 426, + 892, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Application on ControlNet", + "text_level": 1, + "bbox": [ + 500, + 498, + 740, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our model can seamlessly integrate with existing diffusion model plugins, such as ControlNet [44]. Since our $\\bar{y}_{\\theta}$ is structurally different from ControlNet, ControlNet is adopted after our initial sampling step. The results in Fig. 8 demonstrate that our method is fully compatible with ControlNet and effectively resolves its average brightness issue.", + "bbox": [ + 498, + 522, + 892, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 627, + 617, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we delve into the singularities of time intervals in diffusion models, exploring both theoretical and practical aspects. Firstly, we demonstrate that at both regular and singular time steps, the reverse process can be approximated by a Gaussian distribution. Leveraging these theoretical insights, we conduct an in-depth analysis of singular time step sampling and propose a theoretical solution. Finally, we introduce a novel plug-and-play method SingDiffusion, addressing the initial time-step sampling challenge. Remarkably, this module substantially mitigates the average brightness issue prevalent in most current diffusion models, and also enhances their generative capabilities.", + "bbox": [ + 496, + 652, + 893, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This project is supported by the National Natural Science Foundation of China (62072482, U22A2095)", + "bbox": [ + 498, + 854, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6952", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Brian D.O. Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, pages 313-326, 1982. 2", + "[2] AUTOMATIC1111. Stable diffusion web ui. https://github.com/AUTOMATIC1111/stable-diffusion-webui, 2022.7", + "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18208-18218, 2022. 3", + "[4] Andrew Campbell, Joe Benton, Valentin De Bortoli, Tom Rainforth, George Deligiannidis, and Arnaud Doucet. A continuous time framework for discrete denoising models. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2", + "[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems (NeurIPS), pages 8780-8794, 2021. 2, 3", + "[6] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Score-based generative modeling with critically-damped Langevin diffusion. In International Conference on Learning Representations (ICLR), 2022. 3", + "[7] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 3", + "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems, 2014. 3", + "[9] Nicholas Guttenberg. Diffusion with offset noise. https://www.crosslabs.org/blog/diffusion-with-offset-noise, 2023.2, 3", + "[10] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In International Conference on Learning Representations (ICLR), 2023. 3", + "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems (NeurIPS), 2017. 2, 7", + "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3", + "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), pages 6840-6851, 2020. 2, 3, 4", + "[14] Jonathan Ho, Tim Salimans, Alexey A. Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In ICLR Workshop on Deep Generative Models for Highly Structured Data, 2022. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems (NeurIPS), pages 26565-26577, 2022. 3", + "[16] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6007-6017, 2023. 3", + "[17] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. In Advances in Neural Information Processing Systems (NeurIPS), pages 21696-21707, 2021. 3", + "[18] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations (ICLR), 2021. 2", + "[19] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891, 2023. 2, 3", + "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision (ECCV), pages 740-755, 2014. 2, 7", + "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 6", + "[22] Yubin Lu, Zhongjian Wang, and Guillaume Bal. Mathematical analysis of singularities in the diffusion model under the submanifold assumption. arXiv preprint arXiv:2301.07882, 2023. 3", + "[23] Andreas Lugmayr, Martin Danelljan, Luc Van Gool, and Radu Timofte. Srlflow: Learning the super-resolution space with normalizing flow. In European Conference on Computer Vision (ECCV), page 715-732, 2020. 3", + "[24] David McAllester. On the mathematics of diffusion models. arXiv preprint arXiv:2301.11108, 2023. 2", + "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations (ICLR), 2022. 3", + "[26] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning (ICML), pages 16784-16804, 2022. 2, 3, 5", + "[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2", + "[28] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. Grad-tts: A diffusion probabilistic model for text-to-speech. In International Confer-" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6953", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ence on Machine Learning (ICML), pages 8599-8608, 2021. 2", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 2, 7", + "[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning (ICML), pages 8821-8831, 2021. 2, 3", + "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 3", + "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 5", + "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22500-22510, 2023. 3", + "[34] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH Conference Proceedings, 2022. 3", + "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems (NeurIPS), pages 36479-36494, 2022. 2, 3, 7", + "[36] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations (ICLR), 2022. 2, 3", + "[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems (NeurIPS), pages 25278-25294, 2022. 6", + "[38] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning (ICML), pages 2256-2265, 2015. 2, 3, 4" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2021. 2", + "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021. 2, 4", + "[41] Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. In International Conference on Learning Representations (ICLR), 2023. 3", + "[42] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.2", + "[43] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18381-18391, 2023. 3", + "[44] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3836-3847, 2023. 3, 8", + "[45] Pengze Zhang, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Exploring dual-task correlation for pose guided person image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7713-7722, 2022. 3", + "[46] Pengze Zhang, Hubery Yin, Chen Li, and Xiaohua Xie. Formulating discrete probability flow through optimal transport. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 598 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6954", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/992f5837-40b9-47fe-b9bd-9d2e369d7fdb_model.json b/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/992f5837-40b9-47fe-b9bd-9d2e369d7fdb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..806884ec2f92a7ec56cc5b7a63cb2cd62523e1ef --- /dev/null +++ b/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/992f5837-40b9-47fe-b9bd-9d2e369d7fdb_model.json @@ -0,0 +1,2211 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.131, + 0.895, + 0.155 + ], + "angle": 0, + "content": "Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.18, + 0.737, + 0.218 + ], + "angle": 0, + "content": "Pengze Zhang1* Hubery Yin2* Chen Li2 Xiaohua Xie1† \n1Sun Yat-sen University Wechat, Tencent Inc." + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.22, + 0.865, + 0.252 + ], + "angle": 0, + "content": "zhangpz3@mail2.sysu.edu.cn, {hubery, chaselli}@tencent.com, xiexiaoh6@mail.sysu.edu.cn https://pangzecheung.github.io/SingDiffusion/" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.266, + 0.887, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.59, + 0.895, + 0.634 + ], + "angle": 0, + "content": "Figure 1. In this paper, we explore singularities theoretically and propose a plug-and-play module SingDiffusion to address the sampling challenge at the initial singular time step. By integrating this module into existing pre-trained models, our approach effectively tackles the difficulties of generating both dark and bright images, and further enhancing overall image quality as confirmed by quantitative analysis." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.659, + 0.314, + 0.675 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.693, + 0.471, + 0.844 + ], + "angle": 0, + "content": "Most diffusion models assume that the reverse process adheres to a Gaussian distribution. However, this approximation has not been rigorously validated, especially at singularities, where \\( t = 0 \\) and \\( t = 1 \\). Improperly dealing with such singularities leads to an average brightness issue in applications, and limits the generation of images with extreme brightness or darkness. We primarily focus on tackling singularities from both theoretical and practical perspectives. Initially, we establish the error bounds for the reverse process approximation, and showcase its Gauss-" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.661, + 0.895, + 0.812 + ], + "angle": 0, + "content": "sian characteristics at singularity time steps. Based on this theoretical insight, we confirm the singularity at \\( t = 1 \\) is conditionally removable while it at \\( t = 0 \\) is an inherent property. Upon these significant conclusions, we propose a novel plug-and-play method SingDiffusion to address the initial singular time step sampling, which not only effectively resolves the average brightness issue for a wide range of diffusion models without extra training efforts, but also enhances their generation capability in achieving notable lower FID scores." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.844, + 0.633, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Diffusion models, generating samples from initial noise by learning a reverse diffusion process, have achieved remark-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.863, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. This work was done when Pengze Zhang was an intern at WeChat. \n†Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6945" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "able success in multi-modality content generation, such as image generation [4, 5, 26, 30, 35, 46], audio generation [18, 26, 28], and video generation [14]. These achievements owe much to several fundamental theoretical research, namely Denoising Diffusion Probabilistic Modeling (DDPM) [13, 38], Stochastic Differential Equations (SDE) [40], and Ordinary Differential Equations (ODE) [39, 40]. These approaches are all based on the assumption that the reverse diffusion process shares the same functional form as the forward process. Although an indirect validation of this assumption is given by Song et al. [40], it heavily relies on the existence of solutions to the Kolmogorov forward and backward equations [1], which encounters singularities at the endpoints of time intervals where \\( t = 0 \\) and \\( t = 1 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.473, + 0.686 + ], + "angle": 0, + "content": "The singularity issue is not only a gap in the theoretical formulation of current diffusion models, but also affects the quality of the content they generate. Current applications simply ignore singularity points in their implementation [27, 31, 42] and restrict the time interval to \\([\\varepsilon_1, 1 - \\varepsilon_2]\\). As a result, the average brightness of the generated images typically hovers around 0 [9, 19] (normalizing brightness to \\([-1, 1]\\)). For example, as shown in Fig. 1, existing pretrained diffusion models, such as Stable Diffusion 1.5 (SD-1.5) [31] and Stable Diffusion 2.0-base (SD-2.0-base), fail in generating images with pure white or black backgrounds. To address this challenge, Guttenberg et al. [9] add extra offset noise during the training process to allow the network could learn the overall brightness changes of the image. Unfortunately, the offset noise usually disrupts the pre-defined marginal probability distribution and further invalidates the original sampling formula. Lin et al. [19] re-scales the noise schedule to enforce a zero terminal signal-to-noise ratio, and employ the \\( v \\)-prediction technique [36] to circumvent the issue of division by zero at \\( t = 1 \\). However, this method only supports models in \\( v \\)-prediction manner and requires substantial training to fine-tune the entire model. Therefore, it is advantageous to devise a plug-and-play method which effectively deals with the singularity issue for any practicable diffusion model without extra training efforts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.473, + 0.902 + ], + "angle": 0, + "content": "In this paper, we begin with a theoretical exploration of the singularity issue at the endpoints of time intervals in diffusion model, and then devise a novel plug-and-play module to address the accompanying average brightness problem in image generation. Through establishing mathematical error bounds rigorously, we first prove the approximate Gaussian characteristics of the reverse diffusion process at all sampling time steps, especially where the singularity issue appears. Following this, we conduct a thorough analysis of this approximation in the vicinity of singularities, and arrive at two significant conclusions: 1) by computing the limit of 'zero divided by zero' at \\( t = 1 \\), we confirm the corresponding initial singularity is removable; 2) the singularity at \\( t = 0 \\) is an inherent property of diffusion models, thus" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "we should adhere to this form rather than simply avoiding the singularity. Following the aforementioned theoretical analysis, we propose the SingDiffusion method, specifically tailored to address the challenge of sampling during the initial singularity. Especially, this method can seamlessly integrate into existing sampling processes in a plug-and-play manner without requiring additional training efforts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.893, + 0.364 + ], + "angle": 0, + "content": "As demonstrated in Fig. 1 and experiments in the appendix, our novel plug-and-play initial sampling step can effectively resolve the average brightness issue. It can be easily applied to a wide range of diffusion models, including the SD-1.5 and SD-2.0-base with \\(\\epsilon\\)-prediction, SD-2.0 with \\(v\\)-prediction, and various pre-trained models available on the CIVITAI website1, thanks to its one-time training strategy. Furthermore, our method can generally enhance the generative capabilities of these pre-trained diffusion models in notably improving the FID [11] scores at the same CLIP [29] level on the COCO dataset [20]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.376, + 0.642, + 0.391 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.401, + 0.776, + 0.418 + ], + "angle": 0, + "content": "2.1. Reverse process approximation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.575 + ], + "angle": 0, + "content": "Denoising Diffusion Probability Models (DDPM) [13, 38] establish a hand-designed forward Markov chain in discrete time, and model the distribution of data in the reverse process. In contrast, Song et al. [40] establish the diffusion model in continuous time, framed as a Stochastic Differential Equation (SDE). Moreover, they reveal an Ordinary Differential Equation (ODE) termed 'probability flow', sharing the same single-time marginal distribution as the original SDE. Notably, ODE also has discrete counterparts known as Denoising Diffusion Implicit Models (DDIM) [39]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.893, + 0.818 + ], + "angle": 0, + "content": "The assumption that the reversal of the diffusion process has an identical functional form to the forward process is fundamental in the aforementioned methods. Several studies aim to prove this assumption. Song et al. [40] indirectly substantiate this in continuous time by introducing a reverse SDE. Nevertheless, they do not provide error bounds in discrete-time cases, leaving this assumption unverified at discrete-time steps. Additionally, the treatment of singularities at \\( t = 0 \\) and \\( t = 1 \\) is not addressed in their work. McAllester et al. [24] provide proof in discrete time, showing the density of the reverse process as a mixture of Gaussian distributions. Despite this, it doesn't qualify as a pure Gaussian distribution. In contrast, to fill this theoretical gap, our approach directly substantiates this assumption by establishing error bounds for the reverse process approximation at both non-singular and singular time steps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.826, + 0.779, + 0.841 + ], + "angle": 0, + "content": "2.2. Singularities in diffusion model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.849, + 0.892, + 0.88 + ], + "angle": 0, + "content": "Several studies have focused on investigating the singularity occurring at \\( t = 0 \\). Song et al. [40] attempt to bypass" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.888, + 0.629, + 0.9 + ], + "angle": 0, + "content": "1 https://civitai.com/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6946" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.242 + ], + "angle": 0, + "content": "this singularity by initiating their analysis at \\( t = \\varepsilon > 0 \\) instead of \\( t = 0 \\). Therefore, this approach didn't effectively address the core singularity problem. Dockhorn et al. [6] propose a diffusion model on the product space of position and velocity, and avoid the singularity through hybrid score matching. Nevertheless, this approach lacks compatibility with DDPM, SDE and ODE due to the incorporation of the velocity space. Lu et al. [22] employ the \\( x \\)-prediction method to mitigate singularity during training, but did not address the singularity issue during the sampling process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.47, + 0.318 + ], + "angle": 0, + "content": "Therefore, a comprehensive solution to the singularity at \\( t = 0 \\) is still pending. Moreover, the singularity at \\( t = 1 \\) remains unexplored. To tackle these issues, we thoroughly investigate the sampling process at both \\( t = 0 \\) and \\( t = 1 \\), and offer theoretical solutions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.326, + 0.303, + 0.342 + ], + "angle": 0, + "content": "2.3. Average brightness issue" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.471, + 0.637 + ], + "angle": 0, + "content": "Diffusion models have shown significant comprehensive quality and controllability in computer vision, including text-to-image generation [5, 12, 26, 30, 31, 35], image editing [3, 10, 16, 25, 33, 43], image-to-image translation [34, 41, 44], surpassing previous generative models [7, 8, 23, 45]. However, most existing diffusion models ignore the sampling at the initial singular time step, resulting in the inability to generate bright and dark images, i.e., the average brightness issue. Adding offset noise [9] and employing \\( v \\)-prediction [19, 36] are two ways to tackle this problem. However, these methods require fine-tuning for each existing model, consuming a substantial amount of time and limiting their applicability. In contrast, we propose a novel plug-and-play solution targeting the core of the average brightness issue, i.e., the singularity at the initial time step. Our method not only empowers the majority of existing pre-trained models to effectively generate images with the desired brightness level, but also significantly enhances their generative capabilities." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.169, + 0.663 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.84 + ], + "angle": 0, + "content": "To facilitate the clarity of our exploration into singularities from theoretical and practical angles, this section is organized as follows: 1) We start by introducing background and symbols in the Preliminaries. 2) Next, we derive error bounds for the reverse process approximation, confirming its Gaussian characteristics at both regular and singular time steps. 3) We then theoretically analyze and handle the sampling at singular time steps, i.e., \\( t = 0 \\) and \\( t = 1 \\). 4) Lastly, based on our previous analysis, we propose a plug-and-play method to address initial singularity time step sampling, effectively resolving the average brightness issue." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.218, + 0.862 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "In the realm of generative models, we are consistently provided with a set of training samples denoted as \\(\\{y_{i}\\in\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.122 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^d\\}_{i = 1}^N\\) , which are inherently characterized by a distribution given by [15]:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.133, + 0.892, + 0.174 + ], + "angle": 0, + "content": "\\[\np (x, t = 0) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\delta (x - y _ {i}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.186, + 0.795, + 0.2 + ], + "angle": 0, + "content": "where \\(\\delta (x)\\) denotes the Dirac delta function." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.892, + 0.353 + ], + "angle": 0, + "content": "Consider a continuous-time Gaussian diffusion process within the interval \\(0 \\leq t \\leq 1\\). Following [17], the distribution of \\(x_{t}\\) conditioned on \\(x_{0}\\) is written by \\(p(x_{t} | x_{0}) = \\mathcal{N}(\\alpha_{t} x_{0}, \\sigma_{t}^{2} I)\\), where \\(\\alpha_{t}\\) and \\(\\sigma_{t}\\) are positive scalar functions of \\(t\\) satisfying \\(\\alpha_{t}^{2} + \\sigma_{t}^{2} = 1\\), and \\(\\alpha_{t}\\) decreases monotonically from 1 to 0 over time \\(t\\); the distribution of \\(x_{t}\\) conditioned on \\(x_{s}\\) is represented as \\(p(x_{t} | x_{s}) = \\mathcal{N}(\\alpha_{t | s} x_{s}, \\sigma_{t | s}^{2} I)\\), where \\(0 \\leq s < t \\leq 1\\), \\(\\alpha_{t | s} = \\alpha_{t} / \\alpha_{s}\\), \\(\\sigma_{t | s}^{2} = 1 - \\alpha_{t | s}^{2}\\). Consequently, the forward process is derived as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.363, + 0.892, + 0.389 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\alpha_ {t | s} x _ {s} + \\sqrt {1 - \\alpha_ {t | s} ^ {2}} z _ {s}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.892, + 0.429 + ], + "angle": 0, + "content": "where the set \\(\\{z_t\\}_{t=0}^1\\) comprises independent standard Gaussian random variables." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.519 + ], + "angle": 0, + "content": "For a discrete-time diffusion process with \\(T\\) steps, the time \\(i\\in \\{0,1,\\dots T\\}\\) corresponds to \\(t\\) in the continuous case as \\(i = t\\times T\\). Defining \\(\\hat{\\beta}_i = 1 - \\hat{\\alpha}_{i|i - 1}^2\\), where \\(\\hat{\\alpha}_{i|i - 1} = \\alpha_{i / T|(i - 1) / T}\\), and \\(\\hat{\\mathbf{\\Gamma}}^*\\) denotes the symbol in the discrete-time process, the forward process in Eq. 2 can be rewritten as:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.52, + 0.892, + 0.545 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {i} = \\sqrt {1 - \\hat {\\beta} _ {i}} \\hat {x} _ {i - 1} + \\sqrt {\\hat {\\beta} _ {i}} \\hat {z} _ {i - 1}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.55, + 0.889, + 0.565 + ], + "angle": 0, + "content": "which is equivalent to the forward process outlined in [13]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.565, + 0.892, + 0.595 + ], + "angle": 0, + "content": "Taking into account the initial distribution (Eq. 1), the single-time marginal distribution of \\( x_{t} \\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.606, + 0.892, + 0.643 + ], + "angle": 0, + "content": "\\[\np \\left(x _ {t}, t\\right) = \\frac {1}{N} \\sum_ {i} \\left(2 \\pi \\sigma_ {t} ^ {2}\\right) ^ {- \\frac {d}{2}} \\exp \\left(- \\frac {\\left(x _ {t} - \\alpha_ {t} y _ {i}\\right) ^ {2}}{2 \\sigma_ {t} ^ {2}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.655, + 0.892, + 0.685 + ], + "angle": 0, + "content": "where \\(d\\) is the dimension of the training samples. As a result, the reverse process can be derived using Bayes' rule:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.696, + 0.795, + 0.729 + ], + "angle": 0, + "content": "\\[\np (x _ {s} | x _ {t}) = p (x _ {t} | x _ {s}) \\frac {p (x _ {s} , s)}{p (x _ {t} , t)} = (2 \\pi \\sigma_ {s | t}) ^ {- \\frac {d}{2}}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.731, + 0.892, + 0.783 + ], + "angle": 0, + "content": "\\[\n\\sum_ {i} \\exp \\left(- \\frac {1}{2 \\sigma_ {s | t} ^ {2}} \\left(x _ {s} - \\frac {\\alpha_ {t | s} \\sigma_ {s} ^ {2} x _ {t}}{\\sigma_ {t} ^ {2}} - \\frac {\\alpha_ {s} \\sigma_ {t | s} ^ {2} y _ {i}}{\\sigma_ {t} ^ {2}}\\right) ^ {2}\\right) w _ {i} \\left(x _ {t}, t\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.783, + 0.89, + 0.825 + ], + "angle": 0, + "content": "where \\(\\sigma_{s|t}^2 = \\sigma_{t|s}^2\\frac{\\sigma_s^2}{\\sigma_t^2}\\), and \\(w_{i}(x_{t},t) = \\frac{\\exp(-\\frac{(x_{t} - \\alpha_{t}y_{i})^{2}}{2\\sigma_{t}^{2}})}{\\sum_{j}\\exp(-\\frac{(x_{t} - \\alpha_{t}y_{j})^{2}}{2\\sigma_{t}^{2}})}\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.719, + 0.847 + ], + "angle": 0, + "content": "3.2. Error bound estimation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Existing diffusion models, such as [13, 38] are based on an assumption that the reverse process in Eq. 5 can be approximated by a Gaussian distribution, when \\(\\hat{\\beta}_i\\) is small, as given" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6947" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.093, + 0.103, + 0.106 + ], + "angle": 0, + "content": "by:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.114, + 0.468, + 0.189 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {p} (x _ {s} | x _ {t}) = \\left(2 \\pi \\sigma_ {s | t} ^ {2}\\right) ^ {- \\frac {d}{2}} \\\\ \\exp \\left(- \\frac {1}{2 \\sigma_ {s | t} ^ {2}} \\left(x _ {s} - \\frac {\\alpha_ {t | s} \\sigma_ {s} ^ {2} x _ {t}}{\\sigma_ {t} ^ {2}} - \\frac {\\alpha_ {s} \\sigma_ {t | s} ^ {2} \\bar {y} \\left(x _ {t} , t\\right)}{\\sigma_ {t} ^ {2}}\\right) ^ {2}\\right), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.19, + 0.468, + 0.249 + ], + "angle": 0, + "content": "where \\(\\bar{y}(x_t, t) = \\sum_i w_i(x_t, t)y_i\\). However, these studies did not furnish the error bounds to support this assumption. To address this theoretical gap, we estimate the error bound as follows." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.258, + 0.468, + 0.306 + ], + "angle": 0, + "content": "Proposition 1 (Error Bound Estimated by \\(\\sigma_{s|t}\\)). \\(\\forall s \\in (0,1)\\), \\(\\exists \\tau \\in (s,1)\\) and \\(C > 0\\), such that \\(\\forall t \\in (s,\\tau]\\), \\(\\int_{\\mathbb{R}^d} |p(x_s|x_t) - \\tilde{p}(x_s|x_t)| dx_s < C \\sqrt{\\sigma_{s|t}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.311, + 0.468, + 0.42 + ], + "angle": 0, + "content": "Proposition 1 demonstrates that when \\(\\sigma_{s|t}\\) is small, the forward and reverse processes share the same form, i.e., Gaussian distribution. Since \\(\\hat{\\beta}_i = \\frac{\\sigma_{i / T}}{\\sigma_{(i - 1) / T}}\\sigma_{(i - 1) / T|i / T}\\) features a term in the \\(\\sigma_{s|t}\\) form, also supports the inference that this assumption remains valid when \\(\\hat{\\beta}_i\\) is small, as highlighted in [13, 38]. It is worth noting that this error is bounded by \\(\\sigma_{s|t}\\) instead of \\(\\hat{\\beta}_i\\) strictly." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.421, + 0.468, + 0.51 + ], + "angle": 0, + "content": "However, the error bound estimated by \\(\\sigma_{s|t}\\) in Proposition 1 is not sufficient to prove the assumption at the singularity time step \\(t = 1\\). The reason is that when \\(t = 1\\), \\(\\sigma_{s|t}\\) approaches 1 as \\(s \\to 1\\), which is not a small value. Consequently, the error bound at \\(t = 1\\) in Proposition 1 remains non-negligible." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.512, + 0.468, + 0.542 + ], + "angle": 0, + "content": "To tackle this issue, we instead utilize \\(\\alpha_{s}\\) to bound the error at time step \\(t = 1\\), and present a new proposition:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.549, + 0.468, + 0.596 + ], + "angle": 0, + "content": "Proposition 2 (Error Bound Estimated by \\(\\alpha_{s}\\)). \\(\\exists \\nu \\in (0,1)\\) and \\(C > 0\\), such that \\(\\forall \\nu \\leq s < t \\leq 1\\), \\(\\int_{\\mathbb{R}^d} |p(x_s|x_t) - \\tilde{p}(x_s|x_t)| dx_s < C \\sqrt{\\alpha_s}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.602, + 0.468, + 0.661 + ], + "angle": 0, + "content": "According to Proposition 2, setting \\( t = 1 \\), it has \\( \\alpha_s \\to 0 \\) as \\( s \\to 1 \\). As a result, the error bound at \\( t = 1 \\) assures a small value, affirming the validation of the Gaussian approximation assumption at \\( t = 1 \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.663, + 0.468, + 0.707 + ], + "angle": 0, + "content": "In sum, through Proposition 1 and 2, we have established that the reverse process of the diffusion model can be approximated by Gaussian distribution across all time steps." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.716, + 0.307, + 0.731 + ], + "angle": 0, + "content": "3.3. Tackling the singularities" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.739, + 0.468, + 0.829 + ], + "angle": 0, + "content": "With the theoretical foundation provided in Section 3.2, we can delve into the analysis of singular time steps using Eq. 6. It is worth noting that this section will mainly focus on addressing the singularities present in the discrete-time diffusion model [13], while the treatment of the continuous case [40] will be deferred to the appendix." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.847, + 0.304, + 0.862 + ], + "angle": 0, + "content": "3.3.1 The singularities at \\( t = 1 \\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Drawing from Eq. 6, a straightforward way is to train a neural network \\(\\bar{y}_{\\theta}(x_t,t)\\) for estimating \\(\\bar{y} (x_{t},t)\\), known" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.199 + ], + "angle": 0, + "content": "as \\(x\\)-prediction. This approach ensures that the approximated reverse process avoids encountering a singularity at \\(t = 1\\). However, for stable training [38], the mainstream choice among current diffusion models is \\(\\epsilon\\)-prediction. It utilizes a neural network \\(\\epsilon_{\\theta}(x_t,t)\\) for estimating \\(\\epsilon (x_{t},t) = \\frac{x_{t} - \\alpha_{t}\\bar{y} (x_{t},t)}{\\sigma_{t}}\\), which will encounter singularity. More specifically, substituting \\(\\epsilon (x_{t},t)\\) for \\(\\bar{y} (x_{t},t)\\) yields the equation:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.21, + 0.891, + 0.247 + ], + "angle": 0, + "content": "\\[\nx _ {s} = \\frac {\\alpha_ {t | s} \\sigma_ {s} ^ {2}}{\\sigma_ {t} ^ {2}} x _ {t} + \\frac {\\alpha_ {s} \\sigma_ {t | s} ^ {2}}{\\sigma_ {t} ^ {2}} \\frac {x _ {t} - \\sigma_ {t} \\epsilon (x _ {t} , t)}{\\alpha_ {t}} + \\sigma_ {s | t} z _ {t}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.256, + 0.892, + 0.3 + ], + "angle": 0, + "content": "At \\( t = 1 \\), the denominator \\( \\alpha_{t} \\) becomes 0, resulting in a division-by-zero singularity. Theoretically, this singularity is removable because the limit exists:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.31, + 0.891, + 0.345 + ], + "angle": 0, + "content": "\\[\n\\lim _ {t \\rightarrow 1 ^ {-}} \\frac {x _ {t} - \\sigma_ {t} \\epsilon (x _ {t} , t)}{\\alpha_ {t}} = \\lim _ {t \\rightarrow 1 ^ {-}} \\bar {y} (x _ {t}, t) = \\frac {1}{N} \\sum_ {i} y _ {i}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.356, + 0.892, + 0.476 + ], + "angle": 0, + "content": "Regrettably, computing this limit during inference is unfeasible, since \\(\\epsilon(x_1, 1) = x_1\\) loses all information of the correct sampling direction. Conversely, \\(\\bar{y}(x_t, t)\\) retains the correct sampling direction for all \\(t \\in [0, 1]\\). Particularly at \\(t = 1\\), \\(\\bar{y}(x_1, 1) = \\frac{1}{N} \\sum_i y_i\\) encapsulates the information for the initial inference step. Therefore, leveraging \\(x\\)-prediction at the initial time step proves more advantageous than \\(\\epsilon\\)-prediction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.495, + 0.727, + 0.51 + ], + "angle": 0, + "content": "3.3.2 The singularities at \\( t = 0 \\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.519, + 0.891, + 0.564 + ], + "angle": 0, + "content": "When \\( s = 0 \\) and \\( t \\) is small, the distribution in Eq. 6 degenerates into a singular distribution, i.e., a Gaussian with zero variance, resembling a Dirac delta:" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.576, + 0.891, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\tilde {p} \\left(x _ {0} \\mid x _ {t}\\right) = \\delta \\left(x _ {0} - y _ {j _ {0}}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.603, + 0.892, + 0.747 + ], + "angle": 0, + "content": "where \\( j_0 = \\arg \\min_j |x_t - \\alpha_t y_j| \\). This singularity directs the sampling process to converge at the correct point \\( y_{j_0} = \\bar{y}(x_0, 0) \\). Therefore, the singularity at \\( t = 0 \\) is an inherent characteristic of diffusion models that do not require avoidance as long we use suitable sampling techniques. For instance, in the final step of the original DDPM sampling method, it has \\( x_0 = \\bar{y}(x_t, t) \\). When \\( t \\) is small, \\( \\bar{y}(x_t, t) \\approx \\bar{y}(x_0, 0) \\), making this process equivalent to Eq. 9. Thus, there is no need to avoid the singularity." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.748, + 0.891, + 0.793 + ], + "angle": 0, + "content": "Besides, we also arrived at this conclusion within the continuous diffusion model, i.e., SDE. More detailed elaboration on this topic can be found in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.641, + 0.817 + ], + "angle": 0, + "content": "3.4. SingDiffusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.891, + 0.901 + ], + "angle": 0, + "content": "In addition to theoretical issues, the singular sampling issue can also cause average brightness issue in applications, as depicted in Fig. 2. This is mainly because most of the existing method sample images starting at time step \\(1 - \\varepsilon\\) using a standard Gaussian distribution, which significantly" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6948" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.473, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.327, + 0.47, + 0.398 + ], + "angle": 0, + "content": "Figure 2. Illusion of the gap between the sampling from \\( x_{1 - \\varepsilon} \\) and \\( x_{1} \\). Due to the lack of consideration of singular time step sampling in most of the existing methods, they will encounter the average brightness issue. To tackle this, we propose a plug-and-play SingDiffusion method (highlighted in red) to bridge this gap." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.473, + 0.683 + ], + "angle": 0, + "content": "diverges from the true distribution \\( p(x_{1 - \\varepsilon}, 1 - \\varepsilon) \\). To validate this, we select two images (a black dog against a black background and a white dog against a white background). Then we diffuse them for a time period of \\( 1 - \\varepsilon \\) respectively, and calculate the mean value of their latent code. It can be seen that the mean of these two latent codes are -0.14 and +0.28, which are significantly different from the samples obtained from the standard Gaussian distribution. Moreover, as evident from the visualization in Fig. 2, the latent code \\( (x_{1 - \\varepsilon}) \\) of the black dog and white dog are notably darker and brighter compared to \\( x_{1} \\) from the Gaussian distribution respectively. Under such distributional differences, according to Proposition 3, employing a Gaussian distribution at \\( t = 1 - \\epsilon \\) is equivalent to generating images towards an average brightness of 0 at \\( t = 1 \\) (with grayscale ranging from [-1, 1]). Consequently, current methods encounter challenges in generating dark or bright images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.693, + 0.471, + 0.737 + ], + "angle": 0, + "content": "Proposition 3. Setting \\( x_{1 - \\epsilon} \\sim \\mathcal{N}(0, I) \\) is equivalent to sampling the value from standard Gaussian as \\( \\bar{y}(x_1, 1) \\) at \\( t = 1 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Based on our analysis of sampling at singular time steps in Section 3.3.1, we propose a novel plug-and-play method SingDiffusion to fill the gap at \\( t = 1 \\), thus solving the problem of average brightness issue. Considering a pre-trained model \\( \\epsilon_{\\theta}(x_t,t) \\) afflicted by the singularity due to division by zero, we proceed to train a model using \\( x \\)-prediction at \\( t = 1 \\). The algorithm of our training and sampling process are shown in Algorithm 1 and Algorithm 2. Firstly, for image-prompt data pairs \\( (x_0,c) \\) in the training process, we use a U-net [32] \\( \\bar{y}_{\\theta} \\) to fit \\( \\bar{y} (x_1,t = 1,c) \\), where \\( \\bar{y} (x_{t},t,c) \\)" + }, + { + "type": "code_caption", + "bbox": [ + 0.502, + 0.091, + 0.818, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Training process of SingDiffusion" + }, + { + "type": "algorithm", + "bbox": [ + 0.511, + 0.109, + 0.892, + 0.314 + ], + "angle": 0, + "content": "1: repeat \n2: \\( x_0, c \\sim p(x_0, c) \\), \\( x_1 \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I}) \\) \n3: Take gradient descent step on \\( \\nabla_\\theta \\| \\bar{y}_\\theta(x_1, c) - x_0 \\|^2 \\) \n4: until converged \nAlgorithm 2 Sampling process of SingDiffusion \n1: \\( x_1 \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I}) \\) \n2: \\( \\varepsilon = 1 / T \\) \n3: \\( x_{1 - \\varepsilon} = \\alpha_{1 - \\varepsilon} \\bar{y}_\\theta(x_1, c) + \\sigma_{1 - \\varepsilon} x_1 \\) \n4: for \\( t = 1 - \\varepsilon, \\dots, \\varepsilon \\) do \n5: Calculate \\( x_{t - \\varepsilon} \\) using existing sampling algorithms \n6: end for \n7: return \\( x_0 \\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.337, + 0.892, + 0.367 + ], + "angle": 0, + "content": "is the extension of \\(\\bar{y}(x_t, t)\\) defined in Section 3.2 under the condition \\(c\\). The loss function can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.373, + 0.892, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {x _ {0}, c \\sim p (x _ {0}, c), x _ {1} \\sim \\mathcal {N} (\\mathbf {0}, I)} | | \\bar {y} _ {\\theta} (x _ {1}, c) - x _ {0} | | ^ {2}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.398, + 0.892, + 0.443 + ], + "angle": 0, + "content": "As our model is only amortized with respect to text embeddings and not the time step, we omit the variable \\(t\\) for \\(\\bar{y}_{\\theta}\\) and the U-net does not necessitate a complex architecture." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.444, + 0.892, + 0.473 + ], + "angle": 0, + "content": "In the sampling process, we use the new model \\(\\bar{y}_{\\theta}\\) in the initial time-step with a DDIM scheduler:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.481, + 0.892, + 0.497 + ], + "angle": 0, + "content": "\\[\nx _ {1 - \\varepsilon} = \\alpha_ {1 - \\varepsilon} \\bar {y} _ {\\theta} (x _ {1}, c) + \\sigma_ {1 - \\varepsilon} x _ {1}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.892, + 0.639 + ], + "angle": 0, + "content": "This equation guarantees \\( x_{1 - \\varepsilon} \\) adheres to the distribution \\( p(x_{1 - \\varepsilon}, 1 - \\varepsilon) \\). Following this, we utilize the existing pre-trained model \\( \\epsilon_{\\theta}(x_t, t, c) / v_{\\theta}(x_t, t, c) \\) in \\( \\epsilon \\)-prediction/v-prediction manner to perform the subsequent sampling steps until it generates \\( x_0 \\). It's worth noting that our method is solely involved in the sampling at \\( t = 1 \\), independent of the subsequent sampling process. Consequently, our approach can be once-trained and seamlessly integrated into the majority of diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.64, + 0.892, + 0.685 + ], + "angle": 0, + "content": "For further improving matching between generated images and input prompts, existing diffusion models typically incorporate classifier-free techniques [26]:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.693, + 0.892, + 0.71 + ], + "angle": 0, + "content": "\\[\no _ {\\text {g u i d a n c e}} = o _ {n e g} + w \\times \\left(o _ {\\text {p o s}} - o _ {n e g}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.892, + 0.865 + ], + "angle": 0, + "content": "where \\( w \\geq 1 \\) represents the guidance scale, \\( o \\) signifies the output of the diffusion model which can be either \\( \\bar{y}_{\\theta} \\), \\( \\epsilon_{\\theta} \\) or \\( v_{\\theta} \\), 'pos' and 'neg' refer to the outputs corresponding to positive and negative prompts respectively. Nevertheless, we notice that when applying this technique at the initial singular time step, the influence of the guidance scale \\( w \\) predominates the results due to the greater directional difference between the negative and positive outputs. To tackle this challenge, we implement a straightforward yet highly effective normalization method:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.873, + 0.892, + 0.891 + ], + "angle": 0, + "content": "\\[\n\\bar {y} _ {\\theta_ {g u i d a n c e}} = \\left[ \\bar {y} _ {\\theta_ {n e g}} + w \\times \\left(\\bar {y} _ {\\theta_ {p o s}} - \\bar {y} _ {\\theta_ {n e g}}\\right) \\right] / w. \\tag {13}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6949" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.485, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.891, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.286, + 0.771, + 0.302 + ], + "angle": 0, + "content": "Figure 3. Comparison of stable diffusion models and SingDiffusion on average brightness issue." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.313, + 0.473, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.313, + 0.886, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.473, + 0.893, + 0.501 + ], + "angle": 0, + "content": "Figure 4. Comparison of Pareto curves between SingDiffusion, SD-1.5, and SD-2.0-base on 30k COCO images, across various guidance scales in [1.5, 2, 3, 4, 5, 6, 7, 8]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.526, + 0.202, + 0.544 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.552, + 0.253, + 0.568 + ], + "angle": 0, + "content": "4.1. Implement details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.47, + 0.712 + ], + "angle": 0, + "content": "To create a versatile plug-and-play model, our model is trained on the Laion2B-en dataset [37], including 2.32 billion text-image pairs. The images are center-cropped to 512 \\(\\times\\) 512. Our \\(\\bar{y}_{\\theta}\\) follows the U-net structure similar to stable diffusion models but with reduced parameters, totaling 140 million. We utilize the AdamW optimizer [21] with a learning rate of 1e-4 to train our \\(\\bar{y}_{\\theta}\\). The training is executed across 64 Nvidia V100 chips with a batch size of 3072, and completes after 180K steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.471, + 0.804 + ], + "angle": 0, + "content": "During the testing process, all steps, including our initial sampling, are carried out using the default guidance scale of \\( w = 7.5 \\). After our initial sampling at the singular time step, all existing pre-trained diffusion models generate images in the DDIM pipeline, executing 50 steps to produce images following a schedule with a total time of \\( T = 1000 \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.303, + 0.833 + ], + "angle": 0, + "content": "4.2. Average brightness issue" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To validate the effectiveness of SingDiffusion in addressing the average brightness issue, we select four extreme prompts, including \"Solid white/black background\", and \"Monochrome line-art logo on a white/black background\"." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.526, + 0.893, + 0.582 + ], + "angle": 0, + "content": "Table 1. Comparison of average brightness of 100 generated images between stable diffusion models and our SingDiffusion under different prompt conditions. For 'white'/'black' prompts, higher/ lower average brightness is better." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.593, + 0.89, + 0.698 + ], + "angle": 0, + "content": "
Model"Solid white, background""Solid black, background""Monochrome line-art logo on a white background""Monochrome line-art logo on a black background"
SD-1.5141.4383.09137.95113.66
+ Ours212.593.04223.6811.52
SD-2.0-base150.5299.67136.13104.45
+ Ours227.430.29228.6810.87
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "For each prompt, we generate 100 images using SD-1.5, SD-2.0-base and SingDiffusion methods, and then calculate their average brightness. The results are shown in Table 1. It is remarkably clear that the stable diffusion methods, whether using prompts with 'black' or 'white' descriptors, tend to generate images with average brightness. However, SingDiffusion, implementing initial singularity sampling, effectively corrects the average brightness of the generated images. For example, under the prompt \"solid black background,\" our method notably lowers the average brightness from 99.67 to 0.29 for images generated by SD-2.0-base." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6950" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.096, + 0.887, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.42, + 0.885, + 0.435 + ], + "angle": 0, + "content": "Figure 5. Our method can be trained once and seamlessly integrated into the pre-trained models on CIVITAI in a plug-and-play fashion." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.459, + 0.469, + 0.487 + ], + "angle": 0, + "content": "Table 2. Comparison of stable diffusion model and SingDiffusion on FID score and CLIP score without classifier guidance." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.498, + 0.465, + 0.571 + ], + "angle": 0, + "content": "
ModelSD-1.5SD-2.0-base
FID ↓CLIP ↑FID ↓CLIP ↑
Original31.8626.7025.1727.48
+ SingDiffusion21.0927.7118.0128.23
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.469, + 0.687 + ], + "angle": 0, + "content": "Moreover, we also provide visualization results for these prompts in Fig. 3. It can be seen that images generated by stable diffusion methods predominantly exhibit a gray tone. In contrast, our SingDiffusion method successfully overcomes this issue, and is capable of generating both dark and bright images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.35, + 0.714 + ], + "angle": 0, + "content": "4.3. Improvement on image quality" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.886 + ], + "angle": 0, + "content": "To validate our improvements in general image quality, following [35], we randomly select 30k prompts from the COCO dataset [20] as the test set. We employ two metrics for evaluation, including the FID score and CLIP score. Specifically, Fréchet Inception Distance (FID) [11] calculates the Fréchet distance between the real data and the generated data. A lower FID implies more realistic generated data. While the Contrastive Language-Image Pre-training (CLIP) [29] score measures the similarity between the generated images and the given prompts. A higher CLIP score means the generated images better match the input prompts." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "First of all, we compare SingDiffusion with SD-1.5 and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.566 + ], + "angle": 0, + "content": "SD-2.0-base in Table 2 to gauge the model's inherent fitting capability, without using guidance-free techniques. It is evident that SingDiffusion significantly outperforms the existing stable diffusion methods in both FID and CLIP scores. These results highlight the yet-to-be-utilized fitting potential in current stable diffusion models, emphasizing the significance of sampling at the initial singular time step." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Furthermore, inspired by Imagen [35], we plot CLIP v.s. FID Pareto curves by varying guidance values within the range [1.5, 2, 3, 4, 5, 6, 7, 8] in Fig. 4. SingDiffusion exhibits substantial improvements over stable diffusion models, especially noticeable with smaller guidance scales. As the guidance scale increases, SingDiffusion consistently maintains a lower FID compared to stable diffusion for achieving a similar CLIP score. This emphasizes that our approach not only enhances image realism but also ensures better adherence to the input prompts." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.866, + 0.743 + ], + "angle": 0, + "content": "4.4. Plug-and-play on other pre-trained models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Since our method is extensively trained on the Laion dataset, it can be easily integrated into the majority of existing diffusion models. To validate this, we download several pre-trained models from the CIVITAI website, each specializing in different image domains like anime, animals, clothing, and so on. To facilitate the application of SingDiffusion to these models, we integrate SingDiffusion into the popular Stable Diffusion Web UI [2] system, and sample images with the Euler ancient method after using our initial singularity sampling process. The outcomes, displayed" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6951" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.47, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.233, + 0.47, + 0.261 + ], + "angle": 0, + "content": "Figure 6. Ablation study of guidance normalization on SD-1.5 across various guidance scales in [1.5, 2, 3, 4, 5, 6, 7, 8]." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.275, + 0.47, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.428, + 0.421, + 0.443 + ], + "angle": 0, + "content": "Figure 7. Visualization of \\(\\bar{y}_{\\theta}\\) for various prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.47, + 0.531 + ], + "angle": 0, + "content": "in Fig. 1 and Fig. 5, showcase SingDiffusion effectively resolving the average brightness issue across all models while preserving their original generative capacities. This demonstrates the adaptability and practicality of our approach." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.541, + 0.44, + 0.558 + ], + "angle": 0, + "content": "4.5. Effect on classifier guidance normalization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.47, + 0.732 + ], + "angle": 0, + "content": "We conduct an ablation study on the guidance normalization operation (Eq. 13), and represent the CLIP v.s. FID Pareto curves in Fig. 6. It can be seen that the method without normalization exhibits inferior results compared with the full method. As the guidance scale increases, the gap in FID between these two methods gradually widens. This phenomenon primarily arises from significant disparities between \\(\\bar{y}_{\\theta_{pos}}\\) and \\(\\bar{y}_{\\theta_{neg}}\\). According to Eq. 13, larger guidance scales may lead to overflow issues. In contrast, normalizing the results with the guidance scale helps keep the outputs within a typical range, thus restoring the FID score." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.282, + 0.759 + ], + "angle": 0, + "content": "4.6. Visualization of the \\(\\bar{y}_{\\theta}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "According to Eq. 10, our main goal with \\(\\bar{y}_{\\theta}\\) is to model the average image corresponding to each prompt. To confirm this, we employ the prompts presented in Fig. 1 and visualize their corresponding \\(\\bar{y}_{\\theta}\\). As demonstrated in Fig. 7, it is clear that \\(\\bar{y}_{\\theta}\\) does indeed represent a smoothed average image. For instance, the first four images of the second row resemble average faces, aligning with the input prompt for generating celebrities. Additionally, we notice that \\(\\bar{y}_{\\theta}\\) adapts to the prompt's brightness suggestion," + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.895, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.386, + 0.884, + 0.401 + ], + "angle": 0, + "content": "Figure 8. SingDiffusion integrates seamlessly with ControlNet." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.428, + 0.893, + 0.49 + ], + "angle": 0, + "content": "appearing brighter or darker accordingly. This highlights our method's ability to effectively capture pertinent lighting conditions, and underscores the significance of the initial singular time step sampling." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.499, + 0.741, + 0.515 + ], + "angle": 0, + "content": "4.7. Application on ControlNet" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.523, + 0.893, + 0.614 + ], + "angle": 0, + "content": "Our model can seamlessly integrate with existing diffusion model plugins, such as ControlNet [44]. Since our \\(\\bar{y}_{\\theta}\\) is structurally different from ControlNet, ControlNet is adopted after our initial sampling step. The results in Fig. 8 demonstrate that our method is fully compatible with ControlNet and effectively resolves its average brightness issue." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.628, + 0.619, + 0.644 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.894, + 0.837 + ], + "angle": 0, + "content": "In this study, we delve into the singularities of time intervals in diffusion models, exploring both theoretical and practical aspects. Firstly, we demonstrate that at both regular and singular time steps, the reverse process can be approximated by a Gaussian distribution. Leveraging these theoretical insights, we conduct an in-depth analysis of singular time step sampling and propose a theoretical solution. Finally, we introduce a novel plug-and-play method SingDiffusion, addressing the initial time-step sampling challenge. Remarkably, this module substantially mitigates the average brightness issue prevalent in most current diffusion models, and also enhances their generative capabilities." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This project is supported by the National Natural Science Foundation of China (62072482, U22A2095)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6952" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Brian D.O. Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, pages 313-326, 1982. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[2] AUTOMATIC1111. Stable diffusion web ui. https://github.com/AUTOMATIC1111/stable-diffusion-webui, 2022.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.471, + 0.268 + ], + "angle": 0, + "content": "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18208-18218, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.471, + 0.339 + ], + "angle": 0, + "content": "[4] Andrew Campbell, Joe Benton, Valentin De Bortoli, Tom Rainforth, George Deligiannidis, and Arnaud Doucet. A continuous time framework for discrete denoising models. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems (NeurIPS), pages 8780-8794, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.471, + 0.452 + ], + "angle": 0, + "content": "[6] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Score-based generative modeling with critically-damped Langevin diffusion. In International Conference on Learning Representations (ICLR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.471, + 0.521 + ], + "angle": 0, + "content": "[7] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.468, + 0.62 + ], + "angle": 0, + "content": "[9] Nicholas Guttenberg. Diffusion with offset noise. https://www.crosslabs.org/blog/diffusion-with-offset-noise, 2023.2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.468, + 0.676 + ], + "angle": 0, + "content": "[10] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In International Conference on Learning Representations (ICLR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.468, + 0.746 + ], + "angle": 0, + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems (NeurIPS), 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.468, + 0.789 + ], + "angle": 0, + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), pages 6840-6851, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Tim Salimans, Alexey A. Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In ICLR Workshop on Deep Generative Models for Highly Structured Data, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[15] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems (NeurIPS), pages 26565-26577, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[16] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6007-6017, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[17] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. In Advances in Neural Information Processing Systems (NeurIPS), pages 21696-21707, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[18] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[19] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision (ECCV), pages 740-755, 2014. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[22] Yubin Lu, Zhongjian Wang, and Guillaume Bal. Mathematical analysis of singularities in the diffusion model under the submanifold assumption. arXiv preprint arXiv:2301.07882, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[23] Andreas Lugmayr, Martin Danelljan, Luc Van Gool, and Radu Timofte. Srlflow: Learning the super-resolution space with normalizing flow. In European Conference on Computer Vision (ECCV), page 715-732, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.632 + ], + "angle": 0, + "content": "[24] David McAllester. On the mathematics of diffusion models. arXiv preprint arXiv:2301.11108, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.634, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations (ICLR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[26] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning (ICML), pages 16784-16804, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. Grad-tts: A diffusion probabilistic model for text-to-speech. In International Confer-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6953" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "ence on Machine Learning (ICML), pages 8599-8608, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.207 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning (ICML), pages 8821-8831, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.279, + 0.47, + 0.348 + ], + "angle": 0, + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.469, + 0.405 + ], + "angle": 0, + "content": "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.407, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22500-22510, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.492, + 0.469, + 0.547 + ], + "angle": 0, + "content": "[34] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH Conference Proceedings, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.549, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems (NeurIPS), pages 36479-36494, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.662, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[36] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations (ICLR), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems (NeurIPS), pages 25278-25294, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[38] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning (ICML), pages 2256-2265, 2015. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[41] Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. In International Conference on Learning Representations (ICLR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[42] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[43] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18381-18391, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[44] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3836-3847, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.475, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[45] Pengze Zhang, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Exploring dual-task correlation for pose guided person image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7713-7722, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.545, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[46] Pengze Zhang, Hubery Yin, Chen Li, and Xiaohua Xie. Formulating discrete probability flow through optimal transport. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6954" + } + ] +] \ No newline at end of file diff --git a/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/992f5837-40b9-47fe-b9bd-9d2e369d7fdb_origin.pdf b/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/992f5837-40b9-47fe-b9bd-9d2e369d7fdb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8ea74d87d75d556a2271320d685a22007a240801 --- /dev/null +++ b/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/992f5837-40b9-47fe-b9bd-9d2e369d7fdb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8dcbcdeb8e3527bef7c92a8454e0c42acc8a6791ac191abbf4ba65c330179435 +size 4657515 diff --git a/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/full.md b/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6e8253c13cb8548e02f9ac46a06831eae5e45c1f --- /dev/null +++ b/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/full.md @@ -0,0 +1,349 @@ +# Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models + +Pengze Zhang1* Hubery Yin2* Chen Li2 Xiaohua Xie1† +1Sun Yat-sen University Wechat, Tencent Inc. + +zhangpz3@mail2.sysu.edu.cn, {hubery, chaselli}@tencent.com, xiexiaoh6@mail.sysu.edu.cn https://pangzecheung.github.io/SingDiffusion/ + +![](images/abbc4ae39ef5fb733708e23c1e2760916cab7246fe7d417f5408e108aa3032b2.jpg) +Figure 1. In this paper, we explore singularities theoretically and propose a plug-and-play module SingDiffusion to address the sampling challenge at the initial singular time step. By integrating this module into existing pre-trained models, our approach effectively tackles the difficulties of generating both dark and bright images, and further enhancing overall image quality as confirmed by quantitative analysis. + +# Abstract + +Most diffusion models assume that the reverse process adheres to a Gaussian distribution. However, this approximation has not been rigorously validated, especially at singularities, where $t = 0$ and $t = 1$ . Improperly dealing with such singularities leads to an average brightness issue in applications, and limits the generation of images with extreme brightness or darkness. We primarily focus on tackling singularities from both theoretical and practical perspectives. Initially, we establish the error bounds for the reverse process approximation, and showcase its Gauss- + +sian characteristics at singularity time steps. Based on this theoretical insight, we confirm the singularity at $t = 1$ is conditionally removable while it at $t = 0$ is an inherent property. Upon these significant conclusions, we propose a novel plug-and-play method SingDiffusion to address the initial singular time step sampling, which not only effectively resolves the average brightness issue for a wide range of diffusion models without extra training efforts, but also enhances their generation capability in achieving notable lower FID scores. + +# 1. Introduction + +Diffusion models, generating samples from initial noise by learning a reverse diffusion process, have achieved remark- + +able success in multi-modality content generation, such as image generation [4, 5, 26, 30, 35, 46], audio generation [18, 26, 28], and video generation [14]. These achievements owe much to several fundamental theoretical research, namely Denoising Diffusion Probabilistic Modeling (DDPM) [13, 38], Stochastic Differential Equations (SDE) [40], and Ordinary Differential Equations (ODE) [39, 40]. These approaches are all based on the assumption that the reverse diffusion process shares the same functional form as the forward process. Although an indirect validation of this assumption is given by Song et al. [40], it heavily relies on the existence of solutions to the Kolmogorov forward and backward equations [1], which encounters singularities at the endpoints of time intervals where $t = 0$ and $t = 1$ . + +The singularity issue is not only a gap in the theoretical formulation of current diffusion models, but also affects the quality of the content they generate. Current applications simply ignore singularity points in their implementation [27, 31, 42] and restrict the time interval to $[\varepsilon_1, 1 - \varepsilon_2]$ . As a result, the average brightness of the generated images typically hovers around 0 [9, 19] (normalizing brightness to $[-1, 1]$ ). For example, as shown in Fig. 1, existing pretrained diffusion models, such as Stable Diffusion 1.5 (SD-1.5) [31] and Stable Diffusion 2.0-base (SD-2.0-base), fail in generating images with pure white or black backgrounds. To address this challenge, Guttenberg et al. [9] add extra offset noise during the training process to allow the network could learn the overall brightness changes of the image. Unfortunately, the offset noise usually disrupts the pre-defined marginal probability distribution and further invalidates the original sampling formula. Lin et al. [19] re-scales the noise schedule to enforce a zero terminal signal-to-noise ratio, and employ the $v$ -prediction technique [36] to circumvent the issue of division by zero at $t = 1$ . However, this method only supports models in $v$ -prediction manner and requires substantial training to fine-tune the entire model. Therefore, it is advantageous to devise a plug-and-play method which effectively deals with the singularity issue for any practicable diffusion model without extra training efforts. + +In this paper, we begin with a theoretical exploration of the singularity issue at the endpoints of time intervals in diffusion model, and then devise a novel plug-and-play module to address the accompanying average brightness problem in image generation. Through establishing mathematical error bounds rigorously, we first prove the approximate Gaussian characteristics of the reverse diffusion process at all sampling time steps, especially where the singularity issue appears. Following this, we conduct a thorough analysis of this approximation in the vicinity of singularities, and arrive at two significant conclusions: 1) by computing the limit of 'zero divided by zero' at $t = 1$ , we confirm the corresponding initial singularity is removable; 2) the singularity at $t = 0$ is an inherent property of diffusion models, thus + +we should adhere to this form rather than simply avoiding the singularity. Following the aforementioned theoretical analysis, we propose the SingDiffusion method, specifically tailored to address the challenge of sampling during the initial singularity. Especially, this method can seamlessly integrate into existing sampling processes in a plug-and-play manner without requiring additional training efforts. + +As demonstrated in Fig. 1 and experiments in the appendix, our novel plug-and-play initial sampling step can effectively resolve the average brightness issue. It can be easily applied to a wide range of diffusion models, including the SD-1.5 and SD-2.0-base with $\epsilon$ -prediction, SD-2.0 with $v$ -prediction, and various pre-trained models available on the CIVITAI website1, thanks to its one-time training strategy. Furthermore, our method can generally enhance the generative capabilities of these pre-trained diffusion models in notably improving the FID [11] scores at the same CLIP [29] level on the COCO dataset [20]. + +# 2. Related Work + +# 2.1. Reverse process approximation + +Denoising Diffusion Probability Models (DDPM) [13, 38] establish a hand-designed forward Markov chain in discrete time, and model the distribution of data in the reverse process. In contrast, Song et al. [40] establish the diffusion model in continuous time, framed as a Stochastic Differential Equation (SDE). Moreover, they reveal an Ordinary Differential Equation (ODE) termed 'probability flow', sharing the same single-time marginal distribution as the original SDE. Notably, ODE also has discrete counterparts known as Denoising Diffusion Implicit Models (DDIM) [39]. + +The assumption that the reversal of the diffusion process has an identical functional form to the forward process is fundamental in the aforementioned methods. Several studies aim to prove this assumption. Song et al. [40] indirectly substantiate this in continuous time by introducing a reverse SDE. Nevertheless, they do not provide error bounds in discrete-time cases, leaving this assumption unverified at discrete-time steps. Additionally, the treatment of singularities at $t = 0$ and $t = 1$ is not addressed in their work. McAllester et al. [24] provide proof in discrete time, showing the density of the reverse process as a mixture of Gaussian distributions. Despite this, it doesn't qualify as a pure Gaussian distribution. In contrast, to fill this theoretical gap, our approach directly substantiates this assumption by establishing error bounds for the reverse process approximation at both non-singular and singular time steps. + +# 2.2. Singularities in diffusion model + +Several studies have focused on investigating the singularity occurring at $t = 0$ . Song et al. [40] attempt to bypass + +this singularity by initiating their analysis at $t = \varepsilon > 0$ instead of $t = 0$ . Therefore, this approach didn't effectively address the core singularity problem. Dockhorn et al. [6] propose a diffusion model on the product space of position and velocity, and avoid the singularity through hybrid score matching. Nevertheless, this approach lacks compatibility with DDPM, SDE and ODE due to the incorporation of the velocity space. Lu et al. [22] employ the $x$ -prediction method to mitigate singularity during training, but did not address the singularity issue during the sampling process. + +Therefore, a comprehensive solution to the singularity at $t = 0$ is still pending. Moreover, the singularity at $t = 1$ remains unexplored. To tackle these issues, we thoroughly investigate the sampling process at both $t = 0$ and $t = 1$ , and offer theoretical solutions. + +# 2.3. Average brightness issue + +Diffusion models have shown significant comprehensive quality and controllability in computer vision, including text-to-image generation [5, 12, 26, 30, 31, 35], image editing [3, 10, 16, 25, 33, 43], image-to-image translation [34, 41, 44], surpassing previous generative models [7, 8, 23, 45]. However, most existing diffusion models ignore the sampling at the initial singular time step, resulting in the inability to generate bright and dark images, i.e., the average brightness issue. Adding offset noise [9] and employing $v$ -prediction [19, 36] are two ways to tackle this problem. However, these methods require fine-tuning for each existing model, consuming a substantial amount of time and limiting their applicability. In contrast, we propose a novel plug-and-play solution targeting the core of the average brightness issue, i.e., the singularity at the initial time step. Our method not only empowers the majority of existing pre-trained models to effectively generate images with the desired brightness level, but also significantly enhances their generative capabilities. + +# 3. Method + +To facilitate the clarity of our exploration into singularities from theoretical and practical angles, this section is organized as follows: 1) We start by introducing background and symbols in the Preliminaries. 2) Next, we derive error bounds for the reverse process approximation, confirming its Gaussian characteristics at both regular and singular time steps. 3) We then theoretically analyze and handle the sampling at singular time steps, i.e., $t = 0$ and $t = 1$ . 4) Lastly, based on our previous analysis, we propose a plug-and-play method to address initial singularity time step sampling, effectively resolving the average brightness issue. + +# 3.1. Preliminaries + +In the realm of generative models, we are consistently provided with a set of training samples denoted as $\{y_{i}\in$ + +$\mathbb{R}^d\}_{i = 1}^N$ , which are inherently characterized by a distribution given by [15]: + +$$ +p (x, t = 0) = \frac {1}{N} \sum_ {i = 1} ^ {N} \delta (x - y _ {i}), \tag {1} +$$ + +where $\delta (x)$ denotes the Dirac delta function. + +Consider a continuous-time Gaussian diffusion process within the interval $0 \leq t \leq 1$ . Following [17], the distribution of $x_{t}$ conditioned on $x_{0}$ is written by $p(x_{t} | x_{0}) = \mathcal{N}(\alpha_{t} x_{0}, \sigma_{t}^{2} I)$ , where $\alpha_{t}$ and $\sigma_{t}$ are positive scalar functions of $t$ satisfying $\alpha_{t}^{2} + \sigma_{t}^{2} = 1$ , and $\alpha_{t}$ decreases monotonically from 1 to 0 over time $t$ ; the distribution of $x_{t}$ conditioned on $x_{s}$ is represented as $p(x_{t} | x_{s}) = \mathcal{N}(\alpha_{t | s} x_{s}, \sigma_{t | s}^{2} I)$ , where $0 \leq s < t \leq 1$ , $\alpha_{t | s} = \alpha_{t} / \alpha_{s}$ , $\sigma_{t | s}^{2} = 1 - \alpha_{t | s}^{2}$ . Consequently, the forward process is derived as follows: + +$$ +x _ {t} = \alpha_ {t | s} x _ {s} + \sqrt {1 - \alpha_ {t | s} ^ {2}} z _ {s}, \tag {2} +$$ + +where the set $\{z_t\}_{t=0}^1$ comprises independent standard Gaussian random variables. + +For a discrete-time diffusion process with $T$ steps, the time $i\in \{0,1,\dots T\}$ corresponds to $t$ in the continuous case as $i = t\times T$ . Defining $\hat{\beta}_i = 1 - \hat{\alpha}_{i|i - 1}^2$ , where $\hat{\alpha}_{i|i - 1} = \alpha_{i / T|(i - 1) / T}$ , and $\hat{\mathbf{\Gamma}}^*$ denotes the symbol in the discrete-time process, the forward process in Eq. 2 can be rewritten as: + +$$ +\hat {x} _ {i} = \sqrt {1 - \hat {\beta} _ {i}} \hat {x} _ {i - 1} + \sqrt {\hat {\beta} _ {i}} \hat {z} _ {i - 1}, \tag {3} +$$ + +which is equivalent to the forward process outlined in [13]. + +Taking into account the initial distribution (Eq. 1), the single-time marginal distribution of $x_{t}$ is given by: + +$$ +p \left(x _ {t}, t\right) = \frac {1}{N} \sum_ {i} \left(2 \pi \sigma_ {t} ^ {2}\right) ^ {- \frac {d}{2}} \exp \left(- \frac {\left(x _ {t} - \alpha_ {t} y _ {i}\right) ^ {2}}{2 \sigma_ {t} ^ {2}}\right), \tag {4} +$$ + +where $d$ is the dimension of the training samples. As a result, the reverse process can be derived using Bayes' rule: + +$$ +p (x _ {s} | x _ {t}) = p (x _ {t} | x _ {s}) \frac {p (x _ {s} , s)}{p (x _ {t} , t)} = (2 \pi \sigma_ {s | t}) ^ {- \frac {d}{2}} +$$ + +$$ +\sum_ {i} \exp \left(- \frac {1}{2 \sigma_ {s | t} ^ {2}} \left(x _ {s} - \frac {\alpha_ {t | s} \sigma_ {s} ^ {2} x _ {t}}{\sigma_ {t} ^ {2}} - \frac {\alpha_ {s} \sigma_ {t | s} ^ {2} y _ {i}}{\sigma_ {t} ^ {2}}\right) ^ {2}\right) w _ {i} \left(x _ {t}, t\right), \tag {5} +$$ + +where $\sigma_{s|t}^2 = \sigma_{t|s}^2\frac{\sigma_s^2}{\sigma_t^2}$ , and $w_{i}(x_{t},t) = \frac{\exp(-\frac{(x_{t} - \alpha_{t}y_{i})^{2}}{2\sigma_{t}^{2}})}{\sum_{j}\exp(-\frac{(x_{t} - \alpha_{t}y_{j})^{2}}{2\sigma_{t}^{2}})}$ + +# 3.2. Error bound estimation + +Existing diffusion models, such as [13, 38] are based on an assumption that the reverse process in Eq. 5 can be approximated by a Gaussian distribution, when $\hat{\beta}_i$ is small, as given + +by: + +$$ +\begin{array}{l} \tilde {p} (x _ {s} | x _ {t}) = \left(2 \pi \sigma_ {s | t} ^ {2}\right) ^ {- \frac {d}{2}} \\ \exp \left(- \frac {1}{2 \sigma_ {s | t} ^ {2}} \left(x _ {s} - \frac {\alpha_ {t | s} \sigma_ {s} ^ {2} x _ {t}}{\sigma_ {t} ^ {2}} - \frac {\alpha_ {s} \sigma_ {t | s} ^ {2} \bar {y} \left(x _ {t} , t\right)}{\sigma_ {t} ^ {2}}\right) ^ {2}\right), \tag {6} \\ \end{array} +$$ + +where $\bar{y}(x_t, t) = \sum_i w_i(x_t, t)y_i$ . However, these studies did not furnish the error bounds to support this assumption. To address this theoretical gap, we estimate the error bound as follows. + +Proposition 1 (Error Bound Estimated by $\sigma_{s|t}$ ). $\forall s \in (0,1)$ , $\exists \tau \in (s,1)$ and $C > 0$ , such that $\forall t \in (s,\tau]$ , $\int_{\mathbb{R}^d} |p(x_s|x_t) - \tilde{p}(x_s|x_t)| dx_s < C \sqrt{\sigma_{s|t}}$ . + +Proposition 1 demonstrates that when $\sigma_{s|t}$ is small, the forward and reverse processes share the same form, i.e., Gaussian distribution. Since $\hat{\beta}_i = \frac{\sigma_{i / T}}{\sigma_{(i - 1) / T}}\sigma_{(i - 1) / T|i / T}$ features a term in the $\sigma_{s|t}$ form, also supports the inference that this assumption remains valid when $\hat{\beta}_i$ is small, as highlighted in [13, 38]. It is worth noting that this error is bounded by $\sigma_{s|t}$ instead of $\hat{\beta}_i$ strictly. + +However, the error bound estimated by $\sigma_{s|t}$ in Proposition 1 is not sufficient to prove the assumption at the singularity time step $t = 1$ . The reason is that when $t = 1$ , $\sigma_{s|t}$ approaches 1 as $s \to 1$ , which is not a small value. Consequently, the error bound at $t = 1$ in Proposition 1 remains non-negligible. + +To tackle this issue, we instead utilize $\alpha_{s}$ to bound the error at time step $t = 1$ , and present a new proposition: + +Proposition 2 (Error Bound Estimated by $\alpha_{s}$ ). $\exists \nu \in (0,1)$ and $C > 0$ , such that $\forall \nu \leq s < t \leq 1$ , $\int_{\mathbb{R}^d} |p(x_s|x_t) - \tilde{p}(x_s|x_t)| dx_s < C \sqrt{\alpha_s}$ . + +According to Proposition 2, setting $t = 1$ , it has $\alpha_s \to 0$ as $s \to 1$ . As a result, the error bound at $t = 1$ assures a small value, affirming the validation of the Gaussian approximation assumption at $t = 1$ . + +In sum, through Proposition 1 and 2, we have established that the reverse process of the diffusion model can be approximated by Gaussian distribution across all time steps. + +# 3.3. Tackling the singularities + +With the theoretical foundation provided in Section 3.2, we can delve into the analysis of singular time steps using Eq. 6. It is worth noting that this section will mainly focus on addressing the singularities present in the discrete-time diffusion model [13], while the treatment of the continuous case [40] will be deferred to the appendix. + +# 3.3.1 The singularities at $t = 1$ + +Drawing from Eq. 6, a straightforward way is to train a neural network $\bar{y}_{\theta}(x_t,t)$ for estimating $\bar{y} (x_{t},t)$ , known + +as $x$ -prediction. This approach ensures that the approximated reverse process avoids encountering a singularity at $t = 1$ . However, for stable training [38], the mainstream choice among current diffusion models is $\epsilon$ -prediction. It utilizes a neural network $\epsilon_{\theta}(x_t,t)$ for estimating $\epsilon (x_{t},t) = \frac{x_{t} - \alpha_{t}\bar{y} (x_{t},t)}{\sigma_{t}}$ , which will encounter singularity. More specifically, substituting $\epsilon (x_{t},t)$ for $\bar{y} (x_{t},t)$ yields the equation: + +$$ +x _ {s} = \frac {\alpha_ {t | s} \sigma_ {s} ^ {2}}{\sigma_ {t} ^ {2}} x _ {t} + \frac {\alpha_ {s} \sigma_ {t | s} ^ {2}}{\sigma_ {t} ^ {2}} \frac {x _ {t} - \sigma_ {t} \epsilon (x _ {t} , t)}{\alpha_ {t}} + \sigma_ {s | t} z _ {t}. \tag {7} +$$ + +At $t = 1$ , the denominator $\alpha_{t}$ becomes 0, resulting in a division-by-zero singularity. Theoretically, this singularity is removable because the limit exists: + +$$ +\lim _ {t \rightarrow 1 ^ {-}} \frac {x _ {t} - \sigma_ {t} \epsilon (x _ {t} , t)}{\alpha_ {t}} = \lim _ {t \rightarrow 1 ^ {-}} \bar {y} (x _ {t}, t) = \frac {1}{N} \sum_ {i} y _ {i}. \tag {8} +$$ + +Regrettably, computing this limit during inference is unfeasible, since $\epsilon(x_1, 1) = x_1$ loses all information of the correct sampling direction. Conversely, $\bar{y}(x_t, t)$ retains the correct sampling direction for all $t \in [0, 1]$ . Particularly at $t = 1$ , $\bar{y}(x_1, 1) = \frac{1}{N} \sum_i y_i$ encapsulates the information for the initial inference step. Therefore, leveraging $x$ -prediction at the initial time step proves more advantageous than $\epsilon$ -prediction. + +# 3.3.2 The singularities at $t = 0$ + +When $s = 0$ and $t$ is small, the distribution in Eq. 6 degenerates into a singular distribution, i.e., a Gaussian with zero variance, resembling a Dirac delta: + +$$ +\tilde {p} \left(x _ {0} \mid x _ {t}\right) = \delta \left(x _ {0} - y _ {j _ {0}}\right), \tag {9} +$$ + +where $j_0 = \arg \min_j |x_t - \alpha_t y_j|$ . This singularity directs the sampling process to converge at the correct point $y_{j_0} = \bar{y}(x_0, 0)$ . Therefore, the singularity at $t = 0$ is an inherent characteristic of diffusion models that do not require avoidance as long we use suitable sampling techniques. For instance, in the final step of the original DDPM sampling method, it has $x_0 = \bar{y}(x_t, t)$ . When $t$ is small, $\bar{y}(x_t, t) \approx \bar{y}(x_0, 0)$ , making this process equivalent to Eq. 9. Thus, there is no need to avoid the singularity. + +Besides, we also arrived at this conclusion within the continuous diffusion model, i.e., SDE. More detailed elaboration on this topic can be found in the appendix. + +# 3.4. SingDiffusion + +In addition to theoretical issues, the singular sampling issue can also cause average brightness issue in applications, as depicted in Fig. 2. This is mainly because most of the existing method sample images starting at time step $1 - \varepsilon$ using a standard Gaussian distribution, which significantly + +![](images/e037e502645f5b1640c7e7d85edb67b67d942c2b91be2fa47f58febb8dc4e729.jpg) +Figure 2. Illusion of the gap between the sampling from $x_{1 - \varepsilon}$ and $x_{1}$ . Due to the lack of consideration of singular time step sampling in most of the existing methods, they will encounter the average brightness issue. To tackle this, we propose a plug-and-play SingDiffusion method (highlighted in red) to bridge this gap. + +diverges from the true distribution $p(x_{1 - \varepsilon}, 1 - \varepsilon)$ . To validate this, we select two images (a black dog against a black background and a white dog against a white background). Then we diffuse them for a time period of $1 - \varepsilon$ respectively, and calculate the mean value of their latent code. It can be seen that the mean of these two latent codes are -0.14 and +0.28, which are significantly different from the samples obtained from the standard Gaussian distribution. Moreover, as evident from the visualization in Fig. 2, the latent code $(x_{1 - \varepsilon})$ of the black dog and white dog are notably darker and brighter compared to $x_{1}$ from the Gaussian distribution respectively. Under such distributional differences, according to Proposition 3, employing a Gaussian distribution at $t = 1 - \epsilon$ is equivalent to generating images towards an average brightness of 0 at $t = 1$ (with grayscale ranging from [-1, 1]). Consequently, current methods encounter challenges in generating dark or bright images. + +Proposition 3. Setting $x_{1 - \epsilon} \sim \mathcal{N}(0, I)$ is equivalent to sampling the value from standard Gaussian as $\bar{y}(x_1, 1)$ at $t = 1$ . + +Based on our analysis of sampling at singular time steps in Section 3.3.1, we propose a novel plug-and-play method SingDiffusion to fill the gap at $t = 1$ , thus solving the problem of average brightness issue. Considering a pre-trained model $\epsilon_{\theta}(x_t,t)$ afflicted by the singularity due to division by zero, we proceed to train a model using $x$ -prediction at $t = 1$ . The algorithm of our training and sampling process are shown in Algorithm 1 and Algorithm 2. Firstly, for image-prompt data pairs $(x_0,c)$ in the training process, we use a U-net [32] $\bar{y}_{\theta}$ to fit $\bar{y} (x_1,t = 1,c)$ , where $\bar{y} (x_{t},t,c)$ + +Algorithm 1 Training process of SingDiffusion +1: repeat +2: $x_0, c \sim p(x_0, c)$ , $x_1 \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ +3: Take gradient descent step on $\nabla_\theta \| \bar{y}_\theta(x_1, c) - x_0 \|^2$ +4: until converged +Algorithm 2 Sampling process of SingDiffusion +1: $x_1 \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ +2: $\varepsilon = 1 / T$ +3: $x_{1 - \varepsilon} = \alpha_{1 - \varepsilon} \bar{y}_\theta(x_1, c) + \sigma_{1 - \varepsilon} x_1$ +4: for $t = 1 - \varepsilon, \dots, \varepsilon$ do +5: Calculate $x_{t - \varepsilon}$ using existing sampling algorithms +6: end for +7: return $x_0$ + +is the extension of $\bar{y}(x_t, t)$ defined in Section 3.2 under the condition $c$ . The loss function can be written as: + +$$ +\mathcal {L} = \mathbb {E} _ {x _ {0}, c \sim p (x _ {0}, c), x _ {1} \sim \mathcal {N} (\mathbf {0}, I)} | | \bar {y} _ {\theta} (x _ {1}, c) - x _ {0} | | ^ {2}. \tag {10} +$$ + +As our model is only amortized with respect to text embeddings and not the time step, we omit the variable $t$ for $\bar{y}_{\theta}$ and the U-net does not necessitate a complex architecture. + +In the sampling process, we use the new model $\bar{y}_{\theta}$ in the initial time-step with a DDIM scheduler: + +$$ +x _ {1 - \varepsilon} = \alpha_ {1 - \varepsilon} \bar {y} _ {\theta} (x _ {1}, c) + \sigma_ {1 - \varepsilon} x _ {1}. \tag {11} +$$ + +This equation guarantees $x_{1 - \varepsilon}$ adheres to the distribution $p(x_{1 - \varepsilon}, 1 - \varepsilon)$ . Following this, we utilize the existing pre-trained model $\epsilon_{\theta}(x_t, t, c) / v_{\theta}(x_t, t, c)$ in $\epsilon$ -prediction/v-prediction manner to perform the subsequent sampling steps until it generates $x_0$ . It's worth noting that our method is solely involved in the sampling at $t = 1$ , independent of the subsequent sampling process. Consequently, our approach can be once-trained and seamlessly integrated into the majority of diffusion models. + +For further improving matching between generated images and input prompts, existing diffusion models typically incorporate classifier-free techniques [26]: + +$$ +o _ {\text {g u i d a n c e}} = o _ {n e g} + w \times \left(o _ {\text {p o s}} - o _ {n e g}\right), \tag {12} +$$ + +where $w \geq 1$ represents the guidance scale, $o$ signifies the output of the diffusion model which can be either $\bar{y}_{\theta}$ , $\epsilon_{\theta}$ or $v_{\theta}$ , 'pos' and 'neg' refer to the outputs corresponding to positive and negative prompts respectively. Nevertheless, we notice that when applying this technique at the initial singular time step, the influence of the guidance scale $w$ predominates the results due to the greater directional difference between the negative and positive outputs. To tackle this challenge, we implement a straightforward yet highly effective normalization method: + +$$ +\bar {y} _ {\theta_ {g u i d a n c e}} = \left[ \bar {y} _ {\theta_ {n e g}} + w \times \left(\bar {y} _ {\theta_ {p o s}} - \bar {y} _ {\theta_ {n e g}}\right) \right] / w. \tag {13} +$$ + +![](images/28e8ccf56423d542704fa8384572c5cfefa5ea8644151d80832b2cb6222e7e75.jpg) +Figure 3. Comparison of stable diffusion models and SingDiffusion on average brightness issue. + +![](images/4ae8a6023c25fae9ade665d51f20f6a45f9d77d5be088fa94867cab7bb6bef7f.jpg) + +![](images/6abdc21c767e15cfe3e79e26587d0f1f3dcb0e67bedaff2900c760d64e205e8c.jpg) +Figure 4. Comparison of Pareto curves between SingDiffusion, SD-1.5, and SD-2.0-base on 30k COCO images, across various guidance scales in [1.5, 2, 3, 4, 5, 6, 7, 8]. + +![](images/7001ca5bf4283e58972562adea7f1031d4116e48dccd8723494eead4b34ce5df.jpg) + +# 4. Experiment + +# 4.1. Implement details + +To create a versatile plug-and-play model, our model is trained on the Laion2B-en dataset [37], including 2.32 billion text-image pairs. The images are center-cropped to 512 $\times$ 512. Our $\bar{y}_{\theta}$ follows the U-net structure similar to stable diffusion models but with reduced parameters, totaling 140 million. We utilize the AdamW optimizer [21] with a learning rate of 1e-4 to train our $\bar{y}_{\theta}$ . The training is executed across 64 Nvidia V100 chips with a batch size of 3072, and completes after 180K steps. + +During the testing process, all steps, including our initial sampling, are carried out using the default guidance scale of $w = 7.5$ . After our initial sampling at the singular time step, all existing pre-trained diffusion models generate images in the DDIM pipeline, executing 50 steps to produce images following a schedule with a total time of $T = 1000$ . + +# 4.2. Average brightness issue + +To validate the effectiveness of SingDiffusion in addressing the average brightness issue, we select four extreme prompts, including "Solid white/black background", and "Monochrome line-art logo on a white/black background". + +Table 1. Comparison of average brightness of 100 generated images between stable diffusion models and our SingDiffusion under different prompt conditions. For 'white'/'black' prompts, higher/ lower average brightness is better. + +
Model"Solid white, background""Solid black, background""Monochrome line-art logo on a white background""Monochrome line-art logo on a black background"
SD-1.5141.4383.09137.95113.66
+ Ours212.593.04223.6811.52
SD-2.0-base150.5299.67136.13104.45
+ Ours227.430.29228.6810.87
+ +For each prompt, we generate 100 images using SD-1.5, SD-2.0-base and SingDiffusion methods, and then calculate their average brightness. The results are shown in Table 1. It is remarkably clear that the stable diffusion methods, whether using prompts with 'black' or 'white' descriptors, tend to generate images with average brightness. However, SingDiffusion, implementing initial singularity sampling, effectively corrects the average brightness of the generated images. For example, under the prompt "solid black background," our method notably lowers the average brightness from 99.67 to 0.29 for images generated by SD-2.0-base. + +![](images/66cbaa03e2abf489957b7a0e731fd4747bc5afa0fe42b418a0148728ca99ed37.jpg) +Figure 5. Our method can be trained once and seamlessly integrated into the pre-trained models on CIVITAI in a plug-and-play fashion. + +Table 2. Comparison of stable diffusion model and SingDiffusion on FID score and CLIP score without classifier guidance. + +
ModelSD-1.5SD-2.0-base
FID ↓CLIP ↑FID ↓CLIP ↑
Original31.8626.7025.1727.48
+ SingDiffusion21.0927.7118.0128.23
+ +Moreover, we also provide visualization results for these prompts in Fig. 3. It can be seen that images generated by stable diffusion methods predominantly exhibit a gray tone. In contrast, our SingDiffusion method successfully overcomes this issue, and is capable of generating both dark and bright images. + +# 4.3. Improvement on image quality + +To validate our improvements in general image quality, following [35], we randomly select 30k prompts from the COCO dataset [20] as the test set. We employ two metrics for evaluation, including the FID score and CLIP score. Specifically, Fréchet Inception Distance (FID) [11] calculates the Fréchet distance between the real data and the generated data. A lower FID implies more realistic generated data. While the Contrastive Language-Image Pre-training (CLIP) [29] score measures the similarity between the generated images and the given prompts. A higher CLIP score means the generated images better match the input prompts. + +First of all, we compare SingDiffusion with SD-1.5 and + +SD-2.0-base in Table 2 to gauge the model's inherent fitting capability, without using guidance-free techniques. It is evident that SingDiffusion significantly outperforms the existing stable diffusion methods in both FID and CLIP scores. These results highlight the yet-to-be-utilized fitting potential in current stable diffusion models, emphasizing the significance of sampling at the initial singular time step. + +Furthermore, inspired by Imagen [35], we plot CLIP v.s. FID Pareto curves by varying guidance values within the range [1.5, 2, 3, 4, 5, 6, 7, 8] in Fig. 4. SingDiffusion exhibits substantial improvements over stable diffusion models, especially noticeable with smaller guidance scales. As the guidance scale increases, SingDiffusion consistently maintains a lower FID compared to stable diffusion for achieving a similar CLIP score. This emphasizes that our approach not only enhances image realism but also ensures better adherence to the input prompts. + +# 4.4. Plug-and-play on other pre-trained models + +Since our method is extensively trained on the Laion dataset, it can be easily integrated into the majority of existing diffusion models. To validate this, we download several pre-trained models from the CIVITAI website, each specializing in different image domains like anime, animals, clothing, and so on. To facilitate the application of SingDiffusion to these models, we integrate SingDiffusion into the popular Stable Diffusion Web UI [2] system, and sample images with the Euler ancient method after using our initial singularity sampling process. The outcomes, displayed + +![](images/78edc379fc152b6c463ad7bc395ab648d2c66c58beb0bb5d589ab5bb62738140.jpg) +Figure 6. Ablation study of guidance normalization on SD-1.5 across various guidance scales in [1.5, 2, 3, 4, 5, 6, 7, 8]. + +![](images/97b1e9f6057b36567c309b4036c9656ed998006d5fee2ec5cbb3839e4f57c4cb.jpg) +Figure 7. Visualization of $\bar{y}_{\theta}$ for various prompts. + +in Fig. 1 and Fig. 5, showcase SingDiffusion effectively resolving the average brightness issue across all models while preserving their original generative capacities. This demonstrates the adaptability and practicality of our approach. + +# 4.5. Effect on classifier guidance normalization + +We conduct an ablation study on the guidance normalization operation (Eq. 13), and represent the CLIP v.s. FID Pareto curves in Fig. 6. It can be seen that the method without normalization exhibits inferior results compared with the full method. As the guidance scale increases, the gap in FID between these two methods gradually widens. This phenomenon primarily arises from significant disparities between $\bar{y}_{\theta_{pos}}$ and $\bar{y}_{\theta_{neg}}$ . According to Eq. 13, larger guidance scales may lead to overflow issues. In contrast, normalizing the results with the guidance scale helps keep the outputs within a typical range, thus restoring the FID score. + +# 4.6. Visualization of the $\bar{y}_{\theta}$ + +According to Eq. 10, our main goal with $\bar{y}_{\theta}$ is to model the average image corresponding to each prompt. To confirm this, we employ the prompts presented in Fig. 1 and visualize their corresponding $\bar{y}_{\theta}$ . As demonstrated in Fig. 7, it is clear that $\bar{y}_{\theta}$ does indeed represent a smoothed average image. For instance, the first four images of the second row resemble average faces, aligning with the input prompt for generating celebrities. Additionally, we notice that $\bar{y}_{\theta}$ adapts to the prompt's brightness suggestion, + +![](images/96a352894671c981cbe9a42791bbef21584196738259881322cf35fa771ab72f.jpg) +Figure 8. SingDiffusion integrates seamlessly with ControlNet. + +appearing brighter or darker accordingly. This highlights our method's ability to effectively capture pertinent lighting conditions, and underscores the significance of the initial singular time step sampling. + +# 4.7. Application on ControlNet + +Our model can seamlessly integrate with existing diffusion model plugins, such as ControlNet [44]. Since our $\bar{y}_{\theta}$ is structurally different from ControlNet, ControlNet is adopted after our initial sampling step. The results in Fig. 8 demonstrate that our method is fully compatible with ControlNet and effectively resolves its average brightness issue. + +# 5. Conclusion + +In this study, we delve into the singularities of time intervals in diffusion models, exploring both theoretical and practical aspects. Firstly, we demonstrate that at both regular and singular time steps, the reverse process can be approximated by a Gaussian distribution. Leveraging these theoretical insights, we conduct an in-depth analysis of singular time step sampling and propose a theoretical solution. Finally, we introduce a novel plug-and-play method SingDiffusion, addressing the initial time-step sampling challenge. Remarkably, this module substantially mitigates the average brightness issue prevalent in most current diffusion models, and also enhances their generative capabilities. + +Acknowledgments. This project is supported by the National Natural Science Foundation of China (62072482, U22A2095) + +# References + +[1] Brian D.O. Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, pages 313-326, 1982. 2 +[2] AUTOMATIC1111. Stable diffusion web ui. https://github.com/AUTOMATIC1111/stable-diffusion-webui, 2022.7 +[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18208-18218, 2022. 3 +[4] Andrew Campbell, Joe Benton, Valentin De Bortoli, Tom Rainforth, George Deligiannidis, and Arnaud Doucet. A continuous time framework for discrete denoising models. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2 +[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems (NeurIPS), pages 8780-8794, 2021. 2, 3 +[6] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Score-based generative modeling with critically-damped Langevin diffusion. In International Conference on Learning Representations (ICLR), 2022. 3 +[7] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 3 +[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems, 2014. 3 +[9] Nicholas Guttenberg. Diffusion with offset noise. https://www.crosslabs.org/blog/diffusion-with-offset-noise, 2023.2, 3 +[10] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In International Conference on Learning Representations (ICLR), 2023. 3 +[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems (NeurIPS), 2017. 2, 7 +[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3 +[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), pages 6840-6851, 2020. 2, 3, 4 +[14] Jonathan Ho, Tim Salimans, Alexey A. Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In ICLR Workshop on Deep Generative Models for Highly Structured Data, 2022. 2 + +[15] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems (NeurIPS), pages 26565-26577, 2022. 3 +[16] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6007-6017, 2023. 3 +[17] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. In Advances in Neural Information Processing Systems (NeurIPS), pages 21696-21707, 2021. 3 +[18] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations (ICLR), 2021. 2 +[19] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891, 2023. 2, 3 +[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision (ECCV), pages 740-755, 2014. 2, 7 +[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 6 +[22] Yubin Lu, Zhongjian Wang, and Guillaume Bal. Mathematical analysis of singularities in the diffusion model under the submanifold assumption. arXiv preprint arXiv:2301.07882, 2023. 3 +[23] Andreas Lugmayr, Martin Danelljan, Luc Van Gool, and Radu Timofte. Srlflow: Learning the super-resolution space with normalizing flow. In European Conference on Computer Vision (ECCV), page 715-732, 2020. 3 +[24] David McAllester. On the mathematics of diffusion models. arXiv preprint arXiv:2301.11108, 2023. 2 +[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations (ICLR), 2022. 3 +[26] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning (ICML), pages 16784-16804, 2022. 2, 3, 5 +[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2 +[28] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. Grad-tts: A diffusion probabilistic model for text-to-speech. In International Confer- + +ence on Machine Learning (ICML), pages 8599-8608, 2021. 2 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 2, 7 +[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning (ICML), pages 8821-8831, 2021. 2, 3 +[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 3 +[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 5 +[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22500-22510, 2023. 3 +[34] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH Conference Proceedings, 2022. 3 +[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems (NeurIPS), pages 36479-36494, 2022. 2, 3, 7 +[36] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations (ICLR), 2022. 2, 3 +[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems (NeurIPS), pages 25278-25294, 2022. 6 +[38] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning (ICML), pages 2256-2265, 2015. 2, 3, 4 + +[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2021. 2 +[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021. 2, 4 +[41] Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. In International Conference on Learning Representations (ICLR), 2023. 3 +[42] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.2 +[43] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18381-18391, 2023. 3 +[44] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3836-3847, 2023. 3, 8 +[45] Pengze Zhang, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Exploring dual-task correlation for pose guided person image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7713-7722, 2022. 3 +[46] Pengze Zhang, Hubery Yin, Chen Li, and Xiaohua Xie. Formulating discrete probability flow through optimal transport. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2 \ No newline at end of file diff --git a/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/images.zip b/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2a59d69ff09c65ee760bea12c5d23a7b800eb3a1 --- /dev/null +++ b/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e4a648e42e7d3e78c4e9a0d8580c362bd7b4d7f04f8a3b3a921adbbc4b28c0d +size 738134 diff --git a/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/layout.json b/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..300a5a7ee61d5576751ad5ec61e7562219156e2e --- /dev/null +++ b/2024/Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models/layout.json @@ -0,0 +1,10280 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 547, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 547, + 122 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 547, + 122 + ], + "type": "text", + "content": "Tackling the Singularities at the Endpoints of Time Intervals in Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 142, + 451, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 142, + 451, + 172 + ], + "spans": [ + { + "bbox": [ + 147, + 142, + 451, + 172 + ], + "type": "text", + "content": "Pengze Zhang1* Hubery Yin2* Chen Li2 Xiaohua Xie1† \n1Sun Yat-sen University Wechat, Tencent Inc." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 174, + 529, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 174, + 529, + 199 + ], + "spans": [ + { + "bbox": [ + 64, + 174, + 529, + 199 + ], + "type": "text", + "content": "zhangpz3@mail2.sysu.edu.cn, {hubery, chaselli}@tencent.com, xiexiaoh6@mail.sysu.edu.cn https://pangzecheung.github.io/SingDiffusion/" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 210, + 542, + 457 + ], + "blocks": [ + { + "bbox": [ + 50, + 210, + 542, + 457 + ], + "lines": [ + { + "bbox": [ + 50, + 210, + 542, + 457 + ], + "spans": [ + { + "bbox": [ + 50, + 210, + 542, + 457 + ], + "type": "image", + "image_path": "abbc4ae39ef5fb733708e23c1e2760916cab7246fe7d417f5408e108aa3032b2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 467, + 547, + 502 + ], + "lines": [ + { + "bbox": [ + 46, + 467, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 547, + 502 + ], + "type": "text", + "content": "Figure 1. In this paper, we explore singularities theoretically and propose a plug-and-play module SingDiffusion to address the sampling challenge at the initial singular time step. By integrating this module into existing pre-trained models, our approach effectively tackles the difficulties of generating both dark and bright images, and further enhancing overall image quality as confirmed by quantitative analysis." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 521, + 192, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 521, + 192, + 534 + ], + "spans": [ + { + "bbox": [ + 143, + 521, + 192, + 534 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 548, + 288, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 288, + 668 + ], + "type": "text", + "content": "Most diffusion models assume that the reverse process adheres to a Gaussian distribution. However, this approximation has not been rigorously validated, especially at singularities, where " + }, + { + "bbox": [ + 46, + 548, + 288, + 668 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 46, + 548, + 288, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 548, + 288, + 668 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 46, + 548, + 288, + 668 + ], + "type": "text", + "content": ". Improperly dealing with such singularities leads to an average brightness issue in applications, and limits the generation of images with extreme brightness or darkness. We primarily focus on tackling singularities from both theoretical and practical perspectives. Initially, we establish the error bounds for the reverse process approximation, and showcase its Gauss-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 523, + 547, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 523, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 306, + 523, + 547, + 643 + ], + "type": "text", + "content": "sian characteristics at singularity time steps. Based on this theoretical insight, we confirm the singularity at " + }, + { + "bbox": [ + 306, + 523, + 547, + 643 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 306, + 523, + 547, + 643 + ], + "type": "text", + "content": " is conditionally removable while it at " + }, + { + "bbox": [ + 306, + 523, + 547, + 643 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 306, + 523, + 547, + 643 + ], + "type": "text", + "content": " is an inherent property. Upon these significant conclusions, we propose a novel plug-and-play method SingDiffusion to address the initial singular time step sampling, which not only effectively resolves the average brightness issue for a wide range of diffusion models without extra training efforts, but also enhances their generation capability in achieving notable lower FID scores." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 668, + 387, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 668, + 387, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 668, + 387, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "Diffusion models, generating samples from initial noise by learning a reverse diffusion process, have achieved remark-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 683, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 287, + 713 + ], + "type": "text", + "content": "*Equal contribution. This work was done when Pengze Zhang was an intern at WeChat. \n†Corresponding Author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6945" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "able success in multi-modality content generation, such as image generation [4, 5, 26, 30, 35, 46], audio generation [18, 26, 28], and video generation [14]. These achievements owe much to several fundamental theoretical research, namely Denoising Diffusion Probabilistic Modeling (DDPM) [13, 38], Stochastic Differential Equations (SDE) [40], and Ordinary Differential Equations (ODE) [39, 40]. These approaches are all based on the assumption that the reverse diffusion process shares the same functional form as the forward process. Although an indirect validation of this assumption is given by Song et al. [40], it heavily relies on the existence of solutions to the Kolmogorov forward and backward equations [1], which encounters singularities at the endpoints of time intervals where " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 289, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 289, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 289, + 543 + ], + "type": "text", + "content": "The singularity issue is not only a gap in the theoretical formulation of current diffusion models, but also affects the quality of the content they generate. Current applications simply ignore singularity points in their implementation [27, 31, 42] and restrict the time interval to " + }, + { + "bbox": [ + 46, + 243, + 289, + 543 + ], + "type": "inline_equation", + "content": "[\\varepsilon_1, 1 - \\varepsilon_2]" + }, + { + "bbox": [ + 46, + 243, + 289, + 543 + ], + "type": "text", + "content": ". As a result, the average brightness of the generated images typically hovers around 0 [9, 19] (normalizing brightness to " + }, + { + "bbox": [ + 46, + 243, + 289, + 543 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 46, + 243, + 289, + 543 + ], + "type": "text", + "content": "). For example, as shown in Fig. 1, existing pretrained diffusion models, such as Stable Diffusion 1.5 (SD-1.5) [31] and Stable Diffusion 2.0-base (SD-2.0-base), fail in generating images with pure white or black backgrounds. To address this challenge, Guttenberg et al. [9] add extra offset noise during the training process to allow the network could learn the overall brightness changes of the image. Unfortunately, the offset noise usually disrupts the pre-defined marginal probability distribution and further invalidates the original sampling formula. Lin et al. [19] re-scales the noise schedule to enforce a zero terminal signal-to-noise ratio, and employ the " + }, + { + "bbox": [ + 46, + 243, + 289, + 543 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 243, + 289, + 543 + ], + "type": "text", + "content": "-prediction technique [36] to circumvent the issue of division by zero at " + }, + { + "bbox": [ + 46, + 243, + 289, + 543 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 46, + 243, + 289, + 543 + ], + "type": "text", + "content": ". However, this method only supports models in " + }, + { + "bbox": [ + 46, + 243, + 289, + 543 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 243, + 289, + 543 + ], + "type": "text", + "content": "-prediction manner and requires substantial training to fine-tune the entire model. Therefore, it is advantageous to devise a plug-and-play method which effectively deals with the singularity issue for any practicable diffusion model without extra training efforts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "content": "In this paper, we begin with a theoretical exploration of the singularity issue at the endpoints of time intervals in diffusion model, and then devise a novel plug-and-play module to address the accompanying average brightness problem in image generation. Through establishing mathematical error bounds rigorously, we first prove the approximate Gaussian characteristics of the reverse diffusion process at all sampling time steps, especially where the singularity issue appears. Following this, we conduct a thorough analysis of this approximation in the vicinity of singularities, and arrive at two significant conclusions: 1) by computing the limit of 'zero divided by zero' at " + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "content": ", we confirm the corresponding initial singularity is removable; 2) the singularity at " + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "content": " is an inherent property of diffusion models, thus" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "we should adhere to this form rather than simply avoiding the singularity. Following the aforementioned theoretical analysis, we propose the SingDiffusion method, specifically tailored to address the challenge of sampling during the initial singularity. Especially, this method can seamlessly integrate into existing sampling processes in a plug-and-play manner without requiring additional training efforts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "text", + "content": "As demonstrated in Fig. 1 and experiments in the appendix, our novel plug-and-play initial sampling step can effectively resolve the average brightness issue. It can be easily applied to a wide range of diffusion models, including the SD-1.5 and SD-2.0-base with " + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "text", + "content": "-prediction, SD-2.0 with " + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "text", + "content": "-prediction, and various pre-trained models available on the CIVITAI website1, thanks to its one-time training strategy. Furthermore, our method can generally enhance the generative capabilities of these pre-trained diffusion models in notably improving the FID [11] scores at the same CLIP [29] level on the COCO dataset [20]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 297, + 392, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 392, + 309 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 392, + 309 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 317, + 474, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 317, + 474, + 331 + ], + "spans": [ + { + "bbox": [ + 306, + 317, + 474, + 331 + ], + "type": "text", + "content": "2.1. Reverse process approximation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 335, + 545, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 455 + ], + "type": "text", + "content": "Denoising Diffusion Probability Models (DDPM) [13, 38] establish a hand-designed forward Markov chain in discrete time, and model the distribution of data in the reverse process. In contrast, Song et al. [40] establish the diffusion model in continuous time, framed as a Stochastic Differential Equation (SDE). Moreover, they reveal an Ordinary Differential Equation (ODE) termed 'probability flow', sharing the same single-time marginal distribution as the original SDE. Notably, ODE also has discrete counterparts known as Denoising Diffusion Implicit Models (DDIM) [39]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 456, + 546, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 546, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 546, + 647 + ], + "type": "text", + "content": "The assumption that the reversal of the diffusion process has an identical functional form to the forward process is fundamental in the aforementioned methods. Several studies aim to prove this assumption. Song et al. [40] indirectly substantiate this in continuous time by introducing a reverse SDE. Nevertheless, they do not provide error bounds in discrete-time cases, leaving this assumption unverified at discrete-time steps. Additionally, the treatment of singularities at " + }, + { + "bbox": [ + 304, + 456, + 546, + 647 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 304, + 456, + 546, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 456, + 546, + 647 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 304, + 456, + 546, + 647 + ], + "type": "text", + "content": " is not addressed in their work. McAllester et al. [24] provide proof in discrete time, showing the density of the reverse process as a mixture of Gaussian distributions. Despite this, it doesn't qualify as a pure Gaussian distribution. In contrast, to fill this theoretical gap, our approach directly substantiates this assumption by establishing error bounds for the reverse process approximation at both non-singular and singular time steps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 654, + 476, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 654, + 476, + 666 + ], + "spans": [ + { + "bbox": [ + 306, + 654, + 476, + 666 + ], + "type": "text", + "content": "2.2. Singularities in diffusion model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 672, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 672, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 672, + 545, + 696 + ], + "type": "text", + "content": "Several studies have focused on investigating the singularity occurring at " + }, + { + "bbox": [ + 304, + 672, + 545, + 696 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 304, + 672, + 545, + 696 + ], + "type": "text", + "content": ". Song et al. [40] attempt to bypass" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 703, + 384, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 384, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 384, + 712 + ], + "type": "text", + "content": "1 https://civitai.com/" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6946" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 191 + ], + "type": "text", + "content": "this singularity by initiating their analysis at " + }, + { + "bbox": [ + 46, + 72, + 287, + 191 + ], + "type": "inline_equation", + "content": "t = \\varepsilon > 0" + }, + { + "bbox": [ + 46, + 72, + 287, + 191 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 46, + 72, + 287, + 191 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 46, + 72, + 287, + 191 + ], + "type": "text", + "content": ". Therefore, this approach didn't effectively address the core singularity problem. Dockhorn et al. [6] propose a diffusion model on the product space of position and velocity, and avoid the singularity through hybrid score matching. Nevertheless, this approach lacks compatibility with DDPM, SDE and ODE due to the incorporation of the velocity space. Lu et al. [22] employ the " + }, + { + "bbox": [ + 46, + 72, + 287, + 191 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 72, + 287, + 191 + ], + "type": "text", + "content": "-prediction method to mitigate singularity during training, but did not address the singularity issue during the sampling process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 287, + 251 + ], + "type": "text", + "content": "Therefore, a comprehensive solution to the singularity at " + }, + { + "bbox": [ + 46, + 192, + 287, + 251 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 46, + 192, + 287, + 251 + ], + "type": "text", + "content": " is still pending. Moreover, the singularity at " + }, + { + "bbox": [ + 46, + 192, + 287, + 251 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 46, + 192, + 287, + 251 + ], + "type": "text", + "content": " remains unexplored. To tackle these issues, we thoroughly investigate the sampling process at both " + }, + { + "bbox": [ + 46, + 192, + 287, + 251 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 46, + 192, + 287, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 192, + 287, + 251 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 46, + 192, + 287, + 251 + ], + "type": "text", + "content": ", and offer theoretical solutions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 258, + 185, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 185, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 185, + 270 + ], + "type": "text", + "content": "2.3. Average brightness issue" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 276, + 288, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 288, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 288, + 504 + ], + "type": "text", + "content": "Diffusion models have shown significant comprehensive quality and controllability in computer vision, including text-to-image generation [5, 12, 26, 30, 31, 35], image editing [3, 10, 16, 25, 33, 43], image-to-image translation [34, 41, 44], surpassing previous generative models [7, 8, 23, 45]. However, most existing diffusion models ignore the sampling at the initial singular time step, resulting in the inability to generate bright and dark images, i.e., the average brightness issue. Adding offset noise [9] and employing " + }, + { + "bbox": [ + 46, + 276, + 288, + 504 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 276, + 288, + 504 + ], + "type": "text", + "content": "-prediction [19, 36] are two ways to tackle this problem. However, these methods require fine-tuning for each existing model, consuming a substantial amount of time and limiting their applicability. In contrast, we propose a novel plug-and-play solution targeting the core of the average brightness issue, i.e., the singularity at the initial time step. Our method not only empowers the majority of existing pre-trained models to effectively generate images with the desired brightness level, but also significantly enhances their generative capabilities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 513, + 103, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 103, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 103, + 525 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "text", + "content": "To facilitate the clarity of our exploration into singularities from theoretical and practical angles, this section is organized as follows: 1) We start by introducing background and symbols in the Preliminaries. 2) Next, we derive error bounds for the reverse process approximation, confirming its Gaussian characteristics at both regular and singular time steps. 3) We then theoretically analyze and handle the sampling at singular time steps, i.e., " + }, + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 46, + 533, + 287, + 665 + ], + "type": "text", + "content": ". 4) Lastly, based on our previous analysis, we propose a plug-and-play method to address initial singularity time step sampling, effectively resolving the average brightness issue." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 671, + 133, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 133, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 133, + 682 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "In the realm of generative models, we are consistently provided with a set of training samples denoted as " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{y_{i}\\in" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 71, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d\\}_{i = 1}^N" + }, + { + "bbox": [ + 305, + 71, + 545, + 96 + ], + "type": "text", + "content": " , which are inherently characterized by a distribution given by [15]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 359, + 105, + 545, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 105, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 359, + 105, + 545, + 137 + ], + "type": "interline_equation", + "content": "p (x, t = 0) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\delta (x - y _ {i}), \\tag {1}", + "image_path": "7d8f8b8fbd5cf34c01e281ab3808a0f24d75ee8082329420f0745b636e48e719.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 147, + 486, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 147, + 486, + 158 + ], + "spans": [ + { + "bbox": [ + 306, + 147, + 486, + 158 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 147, + 486, + 158 + ], + "type": "inline_equation", + "content": "\\delta (x)" + }, + { + "bbox": [ + 306, + 147, + 486, + 158 + ], + "type": "text", + "content": " denotes the Dirac delta function." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "content": "Consider a continuous-time Gaussian diffusion process within the interval " + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "inline_equation", + "content": "0 \\leq t \\leq 1" + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "content": ". Following [17], the distribution of " + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "content": " is written by " + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "inline_equation", + "content": "p(x_{t} | x_{0}) = \\mathcal{N}(\\alpha_{t} x_{0}, \\sigma_{t}^{2} I)" + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "content": " are positive scalar functions of " + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "content": " satisfying " + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "inline_equation", + "content": "\\alpha_{t}^{2} + \\sigma_{t}^{2} = 1" + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "content": " decreases monotonically from 1 to 0 over time " + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "content": "; the distribution of " + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "content": " is represented as " + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "inline_equation", + "content": "p(x_{t} | x_{s}) = \\mathcal{N}(\\alpha_{t | s} x_{s}, \\sigma_{t | s}^{2} I)" + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "inline_equation", + "content": "0 \\leq s < t \\leq 1" + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "inline_equation", + "content": "\\alpha_{t | s} = \\alpha_{t} / \\alpha_{s}" + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "inline_equation", + "content": "\\sigma_{t | s}^{2} = 1 - \\alpha_{t | s}^{2}" + }, + { + "bbox": [ + 304, + 159, + 545, + 279 + ], + "type": "text", + "content": ". Consequently, the forward process is derived as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 366, + 287, + 545, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 287, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 366, + 287, + 545, + 308 + ], + "type": "interline_equation", + "content": "x _ {t} = \\alpha_ {t | s} x _ {s} + \\sqrt {1 - \\alpha_ {t | s} ^ {2}} z _ {s}, \\tag {2}", + "image_path": "c48d4f36bb93c2a9ac01014527993735aa7446544ac5b0cb2994c40388d529ff.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 316, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 339 + ], + "type": "text", + "content": "where the set " + }, + { + "bbox": [ + 304, + 316, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\{z_t\\}_{t=0}^1" + }, + { + "bbox": [ + 304, + 316, + 545, + 339 + ], + "type": "text", + "content": " comprises independent standard Gaussian random variables." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "text", + "content": "For a discrete-time diffusion process with " + }, + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "text", + "content": " steps, the time " + }, + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "inline_equation", + "content": "i\\in \\{0,1,\\dots T\\}" + }, + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "text", + "content": " corresponds to " + }, + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "text", + "content": " in the continuous case as " + }, + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "inline_equation", + "content": "i = t\\times T" + }, + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "text", + "content": ". Defining " + }, + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\hat{\\beta}_i = 1 - \\hat{\\alpha}_{i|i - 1}^2" + }, + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\hat{\\alpha}_{i|i - 1} = \\alpha_{i / T|(i - 1) / T}" + }, + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{\\Gamma}}^*" + }, + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "text", + "content": " denotes the symbol in the discrete-time process, the forward process in Eq. 2 can be rewritten as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 359, + 411, + 545, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 411, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 359, + 411, + 545, + 431 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {i} = \\sqrt {1 - \\hat {\\beta} _ {i}} \\hat {x} _ {i - 1} + \\sqrt {\\hat {\\beta} _ {i}} \\hat {z} _ {i - 1}, \\tag {3}", + "image_path": "0cde6b22070f106fb44536ec2a2f910d8c10650594569434004e7bac198b4523.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 435, + 544, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 435, + 544, + 447 + ], + "spans": [ + { + "bbox": [ + 305, + 435, + 544, + 447 + ], + "type": "text", + "content": "which is equivalent to the forward process outlined in [13]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 447, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 447, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 305, + 447, + 545, + 471 + ], + "type": "text", + "content": "Taking into account the initial distribution (Eq. 1), the single-time marginal distribution of " + }, + { + "bbox": [ + 305, + 447, + 545, + 471 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 305, + 447, + 545, + 471 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 479, + 545, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 545, + 509 + ], + "type": "interline_equation", + "content": "p \\left(x _ {t}, t\\right) = \\frac {1}{N} \\sum_ {i} \\left(2 \\pi \\sigma_ {t} ^ {2}\\right) ^ {- \\frac {d}{2}} \\exp \\left(- \\frac {\\left(x _ {t} - \\alpha_ {t} y _ {i}\\right) ^ {2}}{2 \\sigma_ {t} ^ {2}}\\right), \\tag {4}", + "image_path": "fbd5c80dfd351423c046a5f47dcb6cc960067d5ff84b9592c811e276a27dcc89.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 518, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 518, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 305, + 518, + 545, + 542 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 518, + 545, + 542 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 305, + 518, + 545, + 542 + ], + "type": "text", + "content": " is the dimension of the training samples. As a result, the reverse process can be derived using Bayes' rule:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 551, + 486, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 486, + 577 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 486, + 577 + ], + "type": "interline_equation", + "content": "p (x _ {s} | x _ {t}) = p (x _ {t} | x _ {s}) \\frac {p (x _ {s} , s)}{p (x _ {t} , t)} = (2 \\pi \\sigma_ {s | t}) ^ {- \\frac {d}{2}}", + "image_path": "013bc11dca8e07c25de9924b75f66a6366a6b1116b73e6ce845a647468924f93.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 578, + 545, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 545, + 620 + ], + "type": "interline_equation", + "content": "\\sum_ {i} \\exp \\left(- \\frac {1}{2 \\sigma_ {s | t} ^ {2}} \\left(x _ {s} - \\frac {\\alpha_ {t | s} \\sigma_ {s} ^ {2} x _ {t}}{\\sigma_ {t} ^ {2}} - \\frac {\\alpha_ {s} \\sigma_ {t | s} ^ {2} y _ {i}}{\\sigma_ {t} ^ {2}}\\right) ^ {2}\\right) w _ {i} \\left(x _ {t}, t\\right), \\tag {5}", + "image_path": "c586efa711e1e5c362da230a209f3f8e11d0ca3cff33f40b4a5baf09c5b00733.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 620, + 544, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 544, + 653 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 544, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 620, + 544, + 653 + ], + "type": "inline_equation", + "content": "\\sigma_{s|t}^2 = \\sigma_{t|s}^2\\frac{\\sigma_s^2}{\\sigma_t^2}" + }, + { + "bbox": [ + 306, + 620, + 544, + 653 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 620, + 544, + 653 + ], + "type": "inline_equation", + "content": "w_{i}(x_{t},t) = \\frac{\\exp(-\\frac{(x_{t} - \\alpha_{t}y_{i})^{2}}{2\\sigma_{t}^{2}})}{\\sum_{j}\\exp(-\\frac{(x_{t} - \\alpha_{t}y_{j})^{2}}{2\\sigma_{t}^{2}})}" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 659, + 440, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 440, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 440, + 670 + ], + "type": "text", + "content": "3.2. Error bound estimation" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Existing diffusion models, such as [13, 38] are based on an assumption that the reverse process in Eq. 5 can be approximated by a Gaussian distribution, when " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\beta}_i" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " is small, as given" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6947" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 73, + 63, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 73, + 63, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 73, + 63, + 83 + ], + "type": "text", + "content": "by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 90, + 286, + 149 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 90, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 47, + 90, + 286, + 149 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {p} (x _ {s} | x _ {t}) = \\left(2 \\pi \\sigma_ {s | t} ^ {2}\\right) ^ {- \\frac {d}{2}} \\\\ \\exp \\left(- \\frac {1}{2 \\sigma_ {s | t} ^ {2}} \\left(x _ {s} - \\frac {\\alpha_ {t | s} \\sigma_ {s} ^ {2} x _ {t}}{\\sigma_ {t} ^ {2}} - \\frac {\\alpha_ {s} \\sigma_ {t | s} ^ {2} \\bar {y} \\left(x _ {t} , t\\right)}{\\sigma_ {t} ^ {2}}\\right) ^ {2}\\right), \\tag {6} \\\\ \\end{array}", + "image_path": "e18159eae553d2876bd26733804ec9ca28e03cd1ab59c0fdb58778420247a2ed.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 150, + 286, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 150, + 286, + 197 + ], + "spans": [ + { + "bbox": [ + 47, + 150, + 286, + 197 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 150, + 286, + 197 + ], + "type": "inline_equation", + "content": "\\bar{y}(x_t, t) = \\sum_i w_i(x_t, t)y_i" + }, + { + "bbox": [ + 47, + 150, + 286, + 197 + ], + "type": "text", + "content": ". However, these studies did not furnish the error bounds to support this assumption. To address this theoretical gap, we estimate the error bound as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 204, + 286, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 204, + 286, + 242 + ], + "spans": [ + { + "bbox": [ + 47, + 204, + 286, + 242 + ], + "type": "text", + "content": "Proposition 1 (Error Bound Estimated by " + }, + { + "bbox": [ + 47, + 204, + 286, + 242 + ], + "type": "inline_equation", + "content": "\\sigma_{s|t}" + }, + { + "bbox": [ + 47, + 204, + 286, + 242 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 47, + 204, + 286, + 242 + ], + "type": "inline_equation", + "content": "\\forall s \\in (0,1)" + }, + { + "bbox": [ + 47, + 204, + 286, + 242 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 204, + 286, + 242 + ], + "type": "inline_equation", + "content": "\\exists \\tau \\in (s,1)" + }, + { + "bbox": [ + 47, + 204, + 286, + 242 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 204, + 286, + 242 + ], + "type": "inline_equation", + "content": "C > 0" + }, + { + "bbox": [ + 47, + 204, + 286, + 242 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 47, + 204, + 286, + 242 + ], + "type": "inline_equation", + "content": "\\forall t \\in (s,\\tau]" + }, + { + "bbox": [ + 47, + 204, + 286, + 242 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 204, + 286, + 242 + ], + "type": "inline_equation", + "content": "\\int_{\\mathbb{R}^d} |p(x_s|x_t) - \\tilde{p}(x_s|x_t)| dx_s < C \\sqrt{\\sigma_{s|t}}" + }, + { + "bbox": [ + 47, + 204, + 286, + 242 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 246, + 286, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 246, + 286, + 332 + ], + "spans": [ + { + "bbox": [ + 47, + 246, + 286, + 332 + ], + "type": "text", + "content": "Proposition 1 demonstrates that when " + }, + { + "bbox": [ + 47, + 246, + 286, + 332 + ], + "type": "inline_equation", + "content": "\\sigma_{s|t}" + }, + { + "bbox": [ + 47, + 246, + 286, + 332 + ], + "type": "text", + "content": " is small, the forward and reverse processes share the same form, i.e., Gaussian distribution. Since " + }, + { + "bbox": [ + 47, + 246, + 286, + 332 + ], + "type": "inline_equation", + "content": "\\hat{\\beta}_i = \\frac{\\sigma_{i / T}}{\\sigma_{(i - 1) / T}}\\sigma_{(i - 1) / T|i / T}" + }, + { + "bbox": [ + 47, + 246, + 286, + 332 + ], + "type": "text", + "content": " features a term in the " + }, + { + "bbox": [ + 47, + 246, + 286, + 332 + ], + "type": "inline_equation", + "content": "\\sigma_{s|t}" + }, + { + "bbox": [ + 47, + 246, + 286, + 332 + ], + "type": "text", + "content": " form, also supports the inference that this assumption remains valid when " + }, + { + "bbox": [ + 47, + 246, + 286, + 332 + ], + "type": "inline_equation", + "content": "\\hat{\\beta}_i" + }, + { + "bbox": [ + 47, + 246, + 286, + 332 + ], + "type": "text", + "content": " is small, as highlighted in [13, 38]. It is worth noting that this error is bounded by " + }, + { + "bbox": [ + 47, + 246, + 286, + 332 + ], + "type": "inline_equation", + "content": "\\sigma_{s|t}" + }, + { + "bbox": [ + 47, + 246, + 286, + 332 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 47, + 246, + 286, + 332 + ], + "type": "inline_equation", + "content": "\\hat{\\beta}_i" + }, + { + "bbox": [ + 47, + 246, + 286, + 332 + ], + "type": "text", + "content": " strictly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 333, + 286, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 333, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 333, + 286, + 403 + ], + "type": "text", + "content": "However, the error bound estimated by " + }, + { + "bbox": [ + 47, + 333, + 286, + 403 + ], + "type": "inline_equation", + "content": "\\sigma_{s|t}" + }, + { + "bbox": [ + 47, + 333, + 286, + 403 + ], + "type": "text", + "content": " in Proposition 1 is not sufficient to prove the assumption at the singularity time step " + }, + { + "bbox": [ + 47, + 333, + 286, + 403 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 47, + 333, + 286, + 403 + ], + "type": "text", + "content": ". The reason is that when " + }, + { + "bbox": [ + 47, + 333, + 286, + 403 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 47, + 333, + 286, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 333, + 286, + 403 + ], + "type": "inline_equation", + "content": "\\sigma_{s|t}" + }, + { + "bbox": [ + 47, + 333, + 286, + 403 + ], + "type": "text", + "content": " approaches 1 as " + }, + { + "bbox": [ + 47, + 333, + 286, + 403 + ], + "type": "inline_equation", + "content": "s \\to 1" + }, + { + "bbox": [ + 47, + 333, + 286, + 403 + ], + "type": "text", + "content": ", which is not a small value. Consequently, the error bound at " + }, + { + "bbox": [ + 47, + 333, + 286, + 403 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 47, + 333, + 286, + 403 + ], + "type": "text", + "content": " in Proposition 1 remains non-negligible." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 405, + 286, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 405, + 286, + 429 + ], + "spans": [ + { + "bbox": [ + 47, + 405, + 286, + 429 + ], + "type": "text", + "content": "To tackle this issue, we instead utilize " + }, + { + "bbox": [ + 47, + 405, + 286, + 429 + ], + "type": "inline_equation", + "content": "\\alpha_{s}" + }, + { + "bbox": [ + 47, + 405, + 286, + 429 + ], + "type": "text", + "content": " to bound the error at time step " + }, + { + "bbox": [ + 47, + 405, + 286, + 429 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 47, + 405, + 286, + 429 + ], + "type": "text", + "content": ", and present a new proposition:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 434, + 286, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 286, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 286, + 472 + ], + "type": "text", + "content": "Proposition 2 (Error Bound Estimated by " + }, + { + "bbox": [ + 47, + 434, + 286, + 472 + ], + "type": "inline_equation", + "content": "\\alpha_{s}" + }, + { + "bbox": [ + 47, + 434, + 286, + 472 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 47, + 434, + 286, + 472 + ], + "type": "inline_equation", + "content": "\\exists \\nu \\in (0,1)" + }, + { + "bbox": [ + 47, + 434, + 286, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 434, + 286, + 472 + ], + "type": "inline_equation", + "content": "C > 0" + }, + { + "bbox": [ + 47, + 434, + 286, + 472 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 47, + 434, + 286, + 472 + ], + "type": "inline_equation", + "content": "\\forall \\nu \\leq s < t \\leq 1" + }, + { + "bbox": [ + 47, + 434, + 286, + 472 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 434, + 286, + 472 + ], + "type": "inline_equation", + "content": "\\int_{\\mathbb{R}^d} |p(x_s|x_t) - \\tilde{p}(x_s|x_t)| dx_s < C \\sqrt{\\alpha_s}" + }, + { + "bbox": [ + 47, + 434, + 286, + 472 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 476, + 286, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 476, + 286, + 523 + ], + "spans": [ + { + "bbox": [ + 47, + 476, + 286, + 523 + ], + "type": "text", + "content": "According to Proposition 2, setting " + }, + { + "bbox": [ + 47, + 476, + 286, + 523 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 47, + 476, + 286, + 523 + ], + "type": "text", + "content": ", it has " + }, + { + "bbox": [ + 47, + 476, + 286, + 523 + ], + "type": "inline_equation", + "content": "\\alpha_s \\to 0" + }, + { + "bbox": [ + 47, + 476, + 286, + 523 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 476, + 286, + 523 + ], + "type": "inline_equation", + "content": "s \\to 1" + }, + { + "bbox": [ + 47, + 476, + 286, + 523 + ], + "type": "text", + "content": ". As a result, the error bound at " + }, + { + "bbox": [ + 47, + 476, + 286, + 523 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 47, + 476, + 286, + 523 + ], + "type": "text", + "content": " assures a small value, affirming the validation of the Gaussian approximation assumption at " + }, + { + "bbox": [ + 47, + 476, + 286, + 523 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 47, + 476, + 286, + 523 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 525, + 286, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 286, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 286, + 559 + ], + "type": "text", + "content": "In sum, through Proposition 1 and 2, we have established that the reverse process of the diffusion model can be approximated by Gaussian distribution across all time steps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 567, + 187, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 187, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 187, + 578 + ], + "type": "text", + "content": "3.3. Tackling the singularities" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 585, + 286, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 286, + 656 + ], + "type": "text", + "content": "With the theoretical foundation provided in Section 3.2, we can delve into the analysis of singular time steps using Eq. 6. It is worth noting that this section will mainly focus on addressing the singularities present in the discrete-time diffusion model [13], while the treatment of the continuous case [40] will be deferred to the appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 670, + 186, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 186, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 186, + 682 + ], + "type": "text", + "content": "3.3.1 The singularities at " + }, + { + "bbox": [ + 47, + 670, + 186, + 682 + ], + "type": "inline_equation", + "content": "t = 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "Drawing from Eq. 6, a straightforward way is to train a neural network " + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\bar{y}_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": " for estimating " + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\bar{y} (x_{t},t)" + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": ", known" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 545, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 157 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 157 + ], + "type": "text", + "content": "as " + }, + { + "bbox": [ + 306, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 306, + 72, + 545, + 157 + ], + "type": "text", + "content": "-prediction. This approach ensures that the approximated reverse process avoids encountering a singularity at " + }, + { + "bbox": [ + 306, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 306, + 72, + 545, + 157 + ], + "type": "text", + "content": ". However, for stable training [38], the mainstream choice among current diffusion models is " + }, + { + "bbox": [ + 306, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 306, + 72, + 545, + 157 + ], + "type": "text", + "content": "-prediction. It utilizes a neural network " + }, + { + "bbox": [ + 306, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 306, + 72, + 545, + 157 + ], + "type": "text", + "content": " for estimating " + }, + { + "bbox": [ + 306, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "\\epsilon (x_{t},t) = \\frac{x_{t} - \\alpha_{t}\\bar{y} (x_{t},t)}{\\sigma_{t}}" + }, + { + "bbox": [ + 306, + 72, + 545, + 157 + ], + "type": "text", + "content": ", which will encounter singularity. More specifically, substituting " + }, + { + "bbox": [ + 306, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "\\epsilon (x_{t},t)" + }, + { + "bbox": [ + 306, + 72, + 545, + 157 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 306, + 72, + 545, + 157 + ], + "type": "inline_equation", + "content": "\\bar{y} (x_{t},t)" + }, + { + "bbox": [ + 306, + 72, + 545, + 157 + ], + "type": "text", + "content": " yields the equation:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 166, + 545, + 195 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 166, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 166, + 545, + 195 + ], + "type": "interline_equation", + "content": "x _ {s} = \\frac {\\alpha_ {t | s} \\sigma_ {s} ^ {2}}{\\sigma_ {t} ^ {2}} x _ {t} + \\frac {\\alpha_ {s} \\sigma_ {t | s} ^ {2}}{\\sigma_ {t} ^ {2}} \\frac {x _ {t} - \\sigma_ {t} \\epsilon (x _ {t} , t)}{\\alpha_ {t}} + \\sigma_ {s | t} z _ {t}. \\tag {7}", + "image_path": "7df34347f7887ca6468b9f7094c66b39ef97c650cabb2af9fb5780eeab0590fb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 202, + 545, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 202, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 305, + 202, + 545, + 237 + ], + "type": "text", + "content": "At " + }, + { + "bbox": [ + 305, + 202, + 545, + 237 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 305, + 202, + 545, + 237 + ], + "type": "text", + "content": ", the denominator " + }, + { + "bbox": [ + 305, + 202, + 545, + 237 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 305, + 202, + 545, + 237 + ], + "type": "text", + "content": " becomes 0, resulting in a division-by-zero singularity. Theoretically, this singularity is removable because the limit exists:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 245, + 545, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 245, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 245, + 545, + 273 + ], + "type": "interline_equation", + "content": "\\lim _ {t \\rightarrow 1 ^ {-}} \\frac {x _ {t} - \\sigma_ {t} \\epsilon (x _ {t} , t)}{\\alpha_ {t}} = \\lim _ {t \\rightarrow 1 ^ {-}} \\bar {y} (x _ {t}, t) = \\frac {1}{N} \\sum_ {i} y _ {i}. \\tag {8}", + "image_path": "61eb986850c50dabcc409980fffcb6a2cc16cea141e5a5a18f83b2715f036567.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "text", + "content": "Regrettably, computing this limit during inference is unfeasible, since " + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\epsilon(x_1, 1) = x_1" + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "text", + "content": " loses all information of the correct sampling direction. Conversely, " + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\bar{y}(x_t, t)" + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "text", + "content": " retains the correct sampling direction for all " + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "inline_equation", + "content": "t \\in [0, 1]" + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "text", + "content": ". Particularly at " + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\bar{y}(x_1, 1) = \\frac{1}{N} \\sum_i y_i" + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "text", + "content": " encapsulates the information for the initial inference step. Therefore, leveraging " + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "text", + "content": "-prediction at the initial time step proves more advantageous than " + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "text", + "content": "-prediction." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 392, + 444, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 444, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 444, + 403 + ], + "type": "text", + "content": "3.3.2 The singularities at " + }, + { + "bbox": [ + 306, + 392, + 444, + 403 + ], + "type": "inline_equation", + "content": "t = 0" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 411, + 545, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 411, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 545, + 446 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 306, + 411, + 545, + 446 + ], + "type": "inline_equation", + "content": "s = 0" + }, + { + "bbox": [ + 306, + 411, + 545, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 411, + 545, + 446 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 411, + 545, + 446 + ], + "type": "text", + "content": " is small, the distribution in Eq. 6 degenerates into a singular distribution, i.e., a Gaussian with zero variance, resembling a Dirac delta:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 374, + 456, + 545, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 456, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 374, + 456, + 545, + 469 + ], + "type": "interline_equation", + "content": "\\tilde {p} \\left(x _ {0} \\mid x _ {t}\\right) = \\delta \\left(x _ {0} - y _ {j _ {0}}\\right), \\tag {9}", + "image_path": "94240b41ab01243a36ef07d977fddd8c7a24d8e1b34bc29b6c6289bf0c382047.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 477, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 477, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 305, + 477, + 545, + 591 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 477, + 545, + 591 + ], + "type": "inline_equation", + "content": "j_0 = \\arg \\min_j |x_t - \\alpha_t y_j|" + }, + { + "bbox": [ + 305, + 477, + 545, + 591 + ], + "type": "text", + "content": ". This singularity directs the sampling process to converge at the correct point " + }, + { + "bbox": [ + 305, + 477, + 545, + 591 + ], + "type": "inline_equation", + "content": "y_{j_0} = \\bar{y}(x_0, 0)" + }, + { + "bbox": [ + 305, + 477, + 545, + 591 + ], + "type": "text", + "content": ". Therefore, the singularity at " + }, + { + "bbox": [ + 305, + 477, + 545, + 591 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 305, + 477, + 545, + 591 + ], + "type": "text", + "content": " is an inherent characteristic of diffusion models that do not require avoidance as long we use suitable sampling techniques. For instance, in the final step of the original DDPM sampling method, it has " + }, + { + "bbox": [ + 305, + 477, + 545, + 591 + ], + "type": "inline_equation", + "content": "x_0 = \\bar{y}(x_t, t)" + }, + { + "bbox": [ + 305, + 477, + 545, + 591 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 305, + 477, + 545, + 591 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 477, + 545, + 591 + ], + "type": "text", + "content": " is small, " + }, + { + "bbox": [ + 305, + 477, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\bar{y}(x_t, t) \\approx \\bar{y}(x_0, 0)" + }, + { + "bbox": [ + 305, + 477, + 545, + 591 + ], + "type": "text", + "content": ", making this process equivalent to Eq. 9. Thus, there is no need to avoid the singularity." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 592, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 592, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 592, + 545, + 628 + ], + "type": "text", + "content": "Besides, we also arrived at this conclusion within the continuous diffusion model, i.e., SDE. More detailed elaboration on this topic can be found in the appendix." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 635, + 392, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 392, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 392, + 647 + ], + "type": "text", + "content": "3.4. SingDiffusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": "In addition to theoretical issues, the singular sampling issue can also cause average brightness issue in applications, as depicted in Fig. 2. This is mainly because most of the existing method sample images starting at time step " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "1 - \\varepsilon" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": " using a standard Gaussian distribution, which significantly" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6948" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 289, + 250 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 289, + 250 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 289, + 250 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 289, + 250 + ], + "type": "image", + "image_path": "e037e502645f5b1640c7e7d85edb67b67d942c2b91be2fa47f58febb8dc4e729.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 258, + 287, + 315 + ], + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 315 + ], + "type": "text", + "content": "Figure 2. Illusion of the gap between the sampling from " + }, + { + "bbox": [ + 46, + 258, + 287, + 315 + ], + "type": "inline_equation", + "content": "x_{1 - \\varepsilon}" + }, + { + "bbox": [ + 46, + 258, + 287, + 315 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 258, + 287, + 315 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 46, + 258, + 287, + 315 + ], + "type": "text", + "content": ". Due to the lack of consideration of singular time step sampling in most of the existing methods, they will encounter the average brightness issue. To tackle this, we propose a plug-and-play SingDiffusion method (highlighted in red) to bridge this gap." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 335, + 289, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 289, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 289, + 540 + ], + "type": "text", + "content": "diverges from the true distribution " + }, + { + "bbox": [ + 46, + 335, + 289, + 540 + ], + "type": "inline_equation", + "content": "p(x_{1 - \\varepsilon}, 1 - \\varepsilon)" + }, + { + "bbox": [ + 46, + 335, + 289, + 540 + ], + "type": "text", + "content": ". To validate this, we select two images (a black dog against a black background and a white dog against a white background). Then we diffuse them for a time period of " + }, + { + "bbox": [ + 46, + 335, + 289, + 540 + ], + "type": "inline_equation", + "content": "1 - \\varepsilon" + }, + { + "bbox": [ + 46, + 335, + 289, + 540 + ], + "type": "text", + "content": " respectively, and calculate the mean value of their latent code. It can be seen that the mean of these two latent codes are -0.14 and +0.28, which are significantly different from the samples obtained from the standard Gaussian distribution. Moreover, as evident from the visualization in Fig. 2, the latent code " + }, + { + "bbox": [ + 46, + 335, + 289, + 540 + ], + "type": "inline_equation", + "content": "(x_{1 - \\varepsilon})" + }, + { + "bbox": [ + 46, + 335, + 289, + 540 + ], + "type": "text", + "content": " of the black dog and white dog are notably darker and brighter compared to " + }, + { + "bbox": [ + 46, + 335, + 289, + 540 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 46, + 335, + 289, + 540 + ], + "type": "text", + "content": " from the Gaussian distribution respectively. Under such distributional differences, according to Proposition 3, employing a Gaussian distribution at " + }, + { + "bbox": [ + 46, + 335, + 289, + 540 + ], + "type": "inline_equation", + "content": "t = 1 - \\epsilon" + }, + { + "bbox": [ + 46, + 335, + 289, + 540 + ], + "type": "text", + "content": " is equivalent to generating images towards an average brightness of 0 at " + }, + { + "bbox": [ + 46, + 335, + 289, + 540 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 46, + 335, + 289, + 540 + ], + "type": "text", + "content": " (with grayscale ranging from [-1, 1]). Consequently, current methods encounter challenges in generating dark or bright images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 548, + 288, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 288, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 288, + 583 + ], + "type": "text", + "content": "Proposition 3. Setting " + }, + { + "bbox": [ + 46, + 548, + 288, + 583 + ], + "type": "inline_equation", + "content": "x_{1 - \\epsilon} \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 46, + 548, + 288, + 583 + ], + "type": "text", + "content": " is equivalent to sampling the value from standard Gaussian as " + }, + { + "bbox": [ + 46, + 548, + 288, + 583 + ], + "type": "inline_equation", + "content": "\\bar{y}(x_1, 1)" + }, + { + "bbox": [ + 46, + 548, + 288, + 583 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 548, + 288, + 583 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 46, + 548, + 288, + 583 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Based on our analysis of sampling at singular time steps in Section 3.3.1, we propose a novel plug-and-play method SingDiffusion to fill the gap at " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": ", thus solving the problem of average brightness issue. Considering a pre-trained model " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " afflicted by the singularity due to division by zero, we proceed to train a model using " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "-prediction at " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": ". The algorithm of our training and sampling process are shown in Algorithm 1 and Algorithm 2. Firstly, for image-prompt data pairs " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "(x_0,c)" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " in the training process, we use a U-net [32] " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\bar{y}_{\\theta}" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " to fit " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\bar{y} (x_1,t = 1,c)" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\bar{y} (x_{t},t,c)" + } + ] + } + ], + "index": 4 + }, + { + "type": "code", + "bbox": [ + 312, + 86, + 545, + 248 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 500, + 84 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 500, + 84 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 500, + 84 + ], + "type": "text", + "content": "Algorithm 1 Training process of SingDiffusion" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "lines": [ + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "text", + "content": "1: repeat \n2: " + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "inline_equation", + "content": "x_0, c \\sim p(x_0, c)" + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "inline_equation", + "content": "x_1 \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "text", + "content": " \n3: Take gradient descent step on " + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "inline_equation", + "content": "\\nabla_\\theta \\| \\bar{y}_\\theta(x_1, c) - x_0 \\|^2" + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "text", + "content": " \n4: until converged \nAlgorithm 2 Sampling process of SingDiffusion \n1: " + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "inline_equation", + "content": "x_1 \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "text", + "content": " \n2: " + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "inline_equation", + "content": "\\varepsilon = 1 / T" + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "text", + "content": " \n3: " + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "inline_equation", + "content": "x_{1 - \\varepsilon} = \\alpha_{1 - \\varepsilon} \\bar{y}_\\theta(x_1, c) + \\sigma_{1 - \\varepsilon} x_1" + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "text", + "content": " \n4: for " + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "inline_equation", + "content": "t = 1 - \\varepsilon, \\dots, \\varepsilon" + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "text", + "content": " do \n5: Calculate " + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "inline_equation", + "content": "x_{t - \\varepsilon}" + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "text", + "content": " using existing sampling algorithms \n6: end for \n7: return " + }, + { + "bbox": [ + 312, + 86, + 545, + 248 + ], + "type": "inline_equation", + "content": "x_0" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "code_body" + } + ], + "index": 6, + "sub_type": "algorithm" + }, + { + "bbox": [ + 305, + 266, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 545, + 290 + ], + "type": "text", + "content": "is the extension of " + }, + { + "bbox": [ + 305, + 266, + 545, + 290 + ], + "type": "inline_equation", + "content": "\\bar{y}(x_t, t)" + }, + { + "bbox": [ + 305, + 266, + 545, + 290 + ], + "type": "text", + "content": " defined in Section 3.2 under the condition " + }, + { + "bbox": [ + 305, + 266, + 545, + 290 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 266, + 545, + 290 + ], + "type": "text", + "content": ". The loss function can be written as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 319, + 295, + 545, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 295, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 319, + 295, + 545, + 311 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {x _ {0}, c \\sim p (x _ {0}, c), x _ {1} \\sim \\mathcal {N} (\\mathbf {0}, I)} | | \\bar {y} _ {\\theta} (x _ {1}, c) - x _ {0} | | ^ {2}. \\tag {10}", + "image_path": "9091a7d5f3715effed68132d0e4f18ea9dc1fb01e97e0eb9d7404cf23fb51ba7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 315, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 315, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 315, + 545, + 350 + ], + "type": "text", + "content": "As our model is only amortized with respect to text embeddings and not the time step, we omit the variable " + }, + { + "bbox": [ + 305, + 315, + 545, + 350 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 315, + 545, + 350 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 305, + 315, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\bar{y}_{\\theta}" + }, + { + "bbox": [ + 305, + 315, + 545, + 350 + ], + "type": "text", + "content": " and the U-net does not necessitate a complex architecture." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 351, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 351, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 351, + 545, + 374 + ], + "type": "text", + "content": "In the sampling process, we use the new model " + }, + { + "bbox": [ + 305, + 351, + 545, + 374 + ], + "type": "inline_equation", + "content": "\\bar{y}_{\\theta}" + }, + { + "bbox": [ + 305, + 351, + 545, + 374 + ], + "type": "text", + "content": " in the initial time-step with a DDIM scheduler:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 356, + 380, + 545, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 380, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 356, + 380, + 545, + 393 + ], + "type": "interline_equation", + "content": "x _ {1 - \\varepsilon} = \\alpha_ {1 - \\varepsilon} \\bar {y} _ {\\theta} (x _ {1}, c) + \\sigma_ {1 - \\varepsilon} x _ {1}. \\tag {11}", + "image_path": "84b73140b469f0e193b4a90929cebab0dbde83ba24ae1253a33ec4ed2b9fea4e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "text", + "content": "This equation guarantees " + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "inline_equation", + "content": "x_{1 - \\varepsilon}" + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "text", + "content": " adheres to the distribution " + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "inline_equation", + "content": "p(x_{1 - \\varepsilon}, 1 - \\varepsilon)" + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "text", + "content": ". Following this, we utilize the existing pre-trained model " + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t, t, c) / v_{\\theta}(x_t, t, c)" + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "text", + "content": "-prediction/v-prediction manner to perform the subsequent sampling steps until it generates " + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "text", + "content": ". It's worth noting that our method is solely involved in the sampling at " + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "text", + "content": ", independent of the subsequent sampling process. Consequently, our approach can be once-trained and seamlessly integrated into the majority of diffusion models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 506, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 506, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 305, + 506, + 545, + 542 + ], + "type": "text", + "content": "For further improving matching between generated images and input prompts, existing diffusion models typically incorporate classifier-free techniques [26]:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 344, + 548, + 545, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 548, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 344, + 548, + 545, + 562 + ], + "type": "interline_equation", + "content": "o _ {\\text {g u i d a n c e}} = o _ {n e g} + w \\times \\left(o _ {\\text {p o s}} - o _ {n e g}\\right), \\tag {12}", + "image_path": "52a4f0461a82abdbf94312d78d19c490907f22fb52e466a3fe56030e69212471.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 566, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 545, + 685 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 566, + 545, + 685 + ], + "type": "inline_equation", + "content": "w \\geq 1" + }, + { + "bbox": [ + 304, + 566, + 545, + 685 + ], + "type": "text", + "content": " represents the guidance scale, " + }, + { + "bbox": [ + 304, + 566, + 545, + 685 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 304, + 566, + 545, + 685 + ], + "type": "text", + "content": " signifies the output of the diffusion model which can be either " + }, + { + "bbox": [ + 304, + 566, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\bar{y}_{\\theta}" + }, + { + "bbox": [ + 304, + 566, + 545, + 685 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 566, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 566, + 545, + 685 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 566, + 545, + 685 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 304, + 566, + 545, + 685 + ], + "type": "text", + "content": ", 'pos' and 'neg' refer to the outputs corresponding to positive and negative prompts respectively. Nevertheless, we notice that when applying this technique at the initial singular time step, the influence of the guidance scale " + }, + { + "bbox": [ + 304, + 566, + 545, + 685 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 566, + 545, + 685 + ], + "type": "text", + "content": " predominates the results due to the greater directional difference between the negative and positive outputs. To tackle this challenge, we implement a straightforward yet highly effective normalization method:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 691, + 545, + 705 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 691, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 321, + 691, + 545, + 705 + ], + "type": "interline_equation", + "content": "\\bar {y} _ {\\theta_ {g u i d a n c e}} = \\left[ \\bar {y} _ {\\theta_ {n e g}} + w \\times \\left(\\bar {y} _ {\\theta_ {p o s}} - \\bar {y} _ {\\theta_ {n e g}}\\right) \\right] / w. \\tag {13}", + "image_path": "03574525f3a9ea7b4f95455812102d1216e6fc52fe656667f974e3352e2af8d2.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6949" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 296, + 217 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 296, + 217 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 296, + 217 + ], + "type": "image", + "image_path": "28e8ccf56423d542704fa8384572c5cfefa5ea8644151d80832b2cb6222e7e75.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 226, + 471, + 239 + ], + "lines": [ + { + "bbox": [ + 122, + 226, + 471, + 239 + ], + "spans": [ + { + "bbox": [ + 122, + 226, + 471, + 239 + ], + "type": "text", + "content": "Figure 3. Comparison of stable diffusion models and SingDiffusion on average brightness issue." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 545, + 216 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 216 + ], + "type": "image", + "image_path": "4ae8a6023c25fae9ade665d51f20f6a45f9d77d5be088fa94867cab7bb6bef7f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 55, + 247, + 289, + 364 + ], + "blocks": [ + { + "bbox": [ + 55, + 247, + 289, + 364 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 289, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 289, + 364 + ], + "type": "image", + "image_path": "6abdc21c767e15cfe3e79e26587d0f1f3dcb0e67bedaff2900c760d64e205e8c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 374, + 546, + 396 + ], + "lines": [ + { + "bbox": [ + 46, + 374, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 546, + 396 + ], + "type": "text", + "content": "Figure 4. Comparison of Pareto curves between SingDiffusion, SD-1.5, and SD-2.0-base on 30k COCO images, across various guidance scales in [1.5, 2, 3, 4, 5, 6, 7, 8]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 247, + 542, + 364 + ], + "blocks": [ + { + "bbox": [ + 309, + 247, + 542, + 364 + ], + "lines": [ + { + "bbox": [ + 309, + 247, + 542, + 364 + ], + "spans": [ + { + "bbox": [ + 309, + 247, + 542, + 364 + ], + "type": "image", + "image_path": "7001ca5bf4283e58972562adea7f1031d4116e48dccd8723494eead4b34ce5df.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 416, + 123, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 416, + 123, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 123, + 430 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 437, + 154, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 154, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 154, + 449 + ], + "type": "text", + "content": "4.1. Implement details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 456, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 563 + ], + "type": "text", + "content": "To create a versatile plug-and-play model, our model is trained on the Laion2B-en dataset [37], including 2.32 billion text-image pairs. The images are center-cropped to 512 " + }, + { + "bbox": [ + 46, + 456, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 46, + 456, + 287, + 563 + ], + "type": "text", + "content": " 512. Our " + }, + { + "bbox": [ + 46, + 456, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\bar{y}_{\\theta}" + }, + { + "bbox": [ + 46, + 456, + 287, + 563 + ], + "type": "text", + "content": " follows the U-net structure similar to stable diffusion models but with reduced parameters, totaling 140 million. We utilize the AdamW optimizer [21] with a learning rate of 1e-4 to train our " + }, + { + "bbox": [ + 46, + 456, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\bar{y}_{\\theta}" + }, + { + "bbox": [ + 46, + 456, + 287, + 563 + ], + "type": "text", + "content": ". The training is executed across 64 Nvidia V100 chips with a batch size of 3072, and completes after 180K steps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 564, + 288, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 288, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 288, + 636 + ], + "type": "text", + "content": "During the testing process, all steps, including our initial sampling, are carried out using the default guidance scale of " + }, + { + "bbox": [ + 46, + 564, + 288, + 636 + ], + "type": "inline_equation", + "content": "w = 7.5" + }, + { + "bbox": [ + 46, + 564, + 288, + 636 + ], + "type": "text", + "content": ". After our initial sampling at the singular time step, all existing pre-trained diffusion models generate images in the DDIM pipeline, executing 50 steps to produce images following a schedule with a total time of " + }, + { + "bbox": [ + 46, + 564, + 288, + 636 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 46, + 564, + 288, + 636 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 646, + 185, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 185, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 185, + 659 + ], + "type": "text", + "content": "4.2. Average brightness issue" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "To validate the effectiveness of SingDiffusion in addressing the average brightness issue, we select four extreme prompts, including \"Solid white/black background\", and \"Monochrome line-art logo on a white/black background\"." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 308, + 469, + 544, + 552 + ], + "blocks": [ + { + "bbox": [ + 305, + 416, + 546, + 460 + ], + "lines": [ + { + "bbox": [ + 305, + 416, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 416, + 546, + 460 + ], + "type": "text", + "content": "Table 1. Comparison of average brightness of 100 generated images between stable diffusion models and our SingDiffusion under different prompt conditions. For 'white'/'black' prompts, higher/ lower average brightness is better." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 469, + 544, + 552 + ], + "lines": [ + { + "bbox": [ + 308, + 469, + 544, + 552 + ], + "spans": [ + { + "bbox": [ + 308, + 469, + 544, + 552 + ], + "type": "table", + "html": "
Model"Solid white, background""Solid black, background""Monochrome line-art logo on a white background""Monochrome line-art logo on a black background"
SD-1.5141.4383.09137.95113.66
+ Ours212.593.04223.6811.52
SD-2.0-base150.5299.67136.13104.45
+ Ours227.430.29228.6810.87
", + "image_path": "b54d322ff8169a8d23c8c89c1e205a9924d087de8ef6656e2fceac5ee232ffcc.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "For each prompt, we generate 100 images using SD-1.5, SD-2.0-base and SingDiffusion methods, and then calculate their average brightness. The results are shown in Table 1. It is remarkably clear that the stable diffusion methods, whether using prompts with 'black' or 'white' descriptors, tend to generate images with average brightness. However, SingDiffusion, implementing initial singularity sampling, effectively corrects the average brightness of the generated images. For example, under the prompt \"solid black background,\" our method notably lowers the average brightness from 99.67 to 0.29 for images generated by SD-2.0-base." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6950" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 76, + 542, + 322 + ], + "blocks": [ + { + "bbox": [ + 49, + 76, + 542, + 322 + ], + "lines": [ + { + "bbox": [ + 49, + 76, + 542, + 322 + ], + "spans": [ + { + "bbox": [ + 49, + 76, + 542, + 322 + ], + "type": "image", + "image_path": "66cbaa03e2abf489957b7a0e731fd4747bc5afa0fe42b418a0148728ca99ed37.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 332, + 541, + 344 + ], + "lines": [ + { + "bbox": [ + 50, + 332, + 541, + 344 + ], + "spans": [ + { + "bbox": [ + 50, + 332, + 541, + 344 + ], + "type": "text", + "content": "Figure 5. Our method can be trained once and seamlessly integrated into the pre-trained models on CIVITAI in a plug-and-play fashion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 394, + 284, + 452 + ], + "blocks": [ + { + "bbox": [ + 47, + 363, + 287, + 385 + ], + "lines": [ + { + "bbox": [ + 47, + 363, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 287, + 385 + ], + "type": "text", + "content": "Table 2. Comparison of stable diffusion model and SingDiffusion on FID score and CLIP score without classifier guidance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 394, + 284, + 452 + ], + "lines": [ + { + "bbox": [ + 48, + 394, + 284, + 452 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 284, + 452 + ], + "type": "table", + "html": "
ModelSD-1.5SD-2.0-base
FID ↓CLIP ↑FID ↓CLIP ↑
Original31.8626.7025.1727.48
+ SingDiffusion21.0927.7118.0128.23
", + "image_path": "4d344df5c18aec15bfb8251a9dc2fe47e03930c0b42472b0af7463d866823d57.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 472, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 544 + ], + "type": "text", + "content": "Moreover, we also provide visualization results for these prompts in Fig. 3. It can be seen that images generated by stable diffusion methods predominantly exhibit a gray tone. In contrast, our SingDiffusion method successfully overcomes this issue, and is capable of generating both dark and bright images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 552, + 214, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 214, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 214, + 565 + ], + "type": "text", + "content": "4.3. Improvement on image quality" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 701 + ], + "type": "text", + "content": "To validate our improvements in general image quality, following [35], we randomly select 30k prompts from the COCO dataset [20] as the test set. We employ two metrics for evaluation, including the FID score and CLIP score. Specifically, Fréchet Inception Distance (FID) [11] calculates the Fréchet distance between the real data and the generated data. A lower FID implies more realistic generated data. While the Contrastive Language-Image Pre-training (CLIP) [29] score measures the similarity between the generated images and the given prompts. A higher CLIP score means the generated images better match the input prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "First of all, we compare SingDiffusion with SD-1.5 and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 365, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 448 + ], + "type": "text", + "content": "SD-2.0-base in Table 2 to gauge the model's inherent fitting capability, without using guidance-free techniques. It is evident that SingDiffusion significantly outperforms the existing stable diffusion methods in both FID and CLIP scores. These results highlight the yet-to-be-utilized fitting potential in current stable diffusion models, emphasizing the significance of sampling at the initial singular time step." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 449, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 568 + ], + "type": "text", + "content": "Furthermore, inspired by Imagen [35], we plot CLIP v.s. FID Pareto curves by varying guidance values within the range [1.5, 2, 3, 4, 5, 6, 7, 8] in Fig. 4. SingDiffusion exhibits substantial improvements over stable diffusion models, especially noticeable with smaller guidance scales. As the guidance scale increases, SingDiffusion consistently maintains a lower FID compared to stable diffusion for achieving a similar CLIP score. This emphasizes that our approach not only enhances image realism but also ensures better adherence to the input prompts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 575, + 529, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 529, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 529, + 588 + ], + "type": "text", + "content": "4.4. Plug-and-play on other pre-trained models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Since our method is extensively trained on the Laion dataset, it can be easily integrated into the majority of existing diffusion models. To validate this, we download several pre-trained models from the CIVITAI website, each specializing in different image domains like anime, animals, clothing, and so on. To facilitate the application of SingDiffusion to these models, we integrate SingDiffusion into the popular Stable Diffusion Web UI [2] system, and sample images with the Euler ancient method after using our initial singularity sampling process. The outcomes, displayed" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6951" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 287, + 174 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 287, + 174 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 287, + 174 + ], + "type": "image", + "image_path": "78edc379fc152b6c463ad7bc395ab648d2c66c58beb0bb5d589ab5bb62738140.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 184, + 287, + 206 + ], + "lines": [ + { + "bbox": [ + 47, + 184, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 47, + 184, + 287, + 206 + ], + "type": "text", + "content": "Figure 6. Ablation study of guidance normalization on SD-1.5 across various guidance scales in [1.5, 2, 3, 4, 5, 6, 7, 8]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 217, + 287, + 329 + ], + "blocks": [ + { + "bbox": [ + 50, + 217, + 287, + 329 + ], + "lines": [ + { + "bbox": [ + 50, + 217, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 50, + 217, + 287, + 329 + ], + "type": "image", + "image_path": "97b1e9f6057b36567c309b4036c9656ed998006d5fee2ec5cbb3839e4f57c4cb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 338, + 257, + 350 + ], + "lines": [ + { + "bbox": [ + 76, + 338, + 257, + 350 + ], + "spans": [ + { + "bbox": [ + 76, + 338, + 257, + 350 + ], + "type": "text", + "content": "Figure 7. Visualization of " + }, + { + "bbox": [ + 76, + 338, + 257, + 350 + ], + "type": "inline_equation", + "content": "\\bar{y}_{\\theta}" + }, + { + "bbox": [ + 76, + 338, + 257, + 350 + ], + "type": "text", + "content": " for various prompts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 372, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 420 + ], + "type": "text", + "content": "in Fig. 1 and Fig. 5, showcase SingDiffusion effectively resolving the average brightness issue across all models while preserving their original generative capacities. This demonstrates the adaptability and practicality of our approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 428, + 269, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 428, + 269, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 428, + 269, + 441 + ], + "type": "text", + "content": "4.5. Effect on classifier guidance normalization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 447, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 579 + ], + "type": "text", + "content": "We conduct an ablation study on the guidance normalization operation (Eq. 13), and represent the CLIP v.s. FID Pareto curves in Fig. 6. It can be seen that the method without normalization exhibits inferior results compared with the full method. As the guidance scale increases, the gap in FID between these two methods gradually widens. This phenomenon primarily arises from significant disparities between " + }, + { + "bbox": [ + 46, + 447, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\bar{y}_{\\theta_{pos}}" + }, + { + "bbox": [ + 46, + 447, + 287, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 447, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\bar{y}_{\\theta_{neg}}" + }, + { + "bbox": [ + 46, + 447, + 287, + 579 + ], + "type": "text", + "content": ". According to Eq. 13, larger guidance scales may lead to overflow issues. In contrast, normalizing the results with the guidance scale helps keep the outputs within a typical range, thus restoring the FID score." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 587, + 172, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 172, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 172, + 601 + ], + "type": "text", + "content": "4.6. Visualization of the " + }, + { + "bbox": [ + 47, + 587, + 172, + 601 + ], + "type": "inline_equation", + "content": "\\bar{y}_{\\theta}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "According to Eq. 10, our main goal with " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\bar{y}_{\\theta}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " is to model the average image corresponding to each prompt. To confirm this, we employ the prompts presented in Fig. 1 and visualize their corresponding " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\bar{y}_{\\theta}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": ". As demonstrated in Fig. 7, it is clear that " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\bar{y}_{\\theta}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " does indeed represent a smoothed average image. For instance, the first four images of the second row resemble average faces, aligning with the input prompt for generating celebrities. Additionally, we notice that " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\bar{y}_{\\theta}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " adapts to the prompt's brightness suggestion," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 547, + 299 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 547, + 299 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 547, + 299 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 547, + 299 + ], + "type": "image", + "image_path": "96a352894671c981cbe9a42791bbef21584196738259881322cf35fa771ab72f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 305, + 541, + 317 + ], + "lines": [ + { + "bbox": [ + 310, + 305, + 541, + 317 + ], + "spans": [ + { + "bbox": [ + 310, + 305, + 541, + 317 + ], + "type": "text", + "content": "Figure 8. SingDiffusion integrates seamlessly with ControlNet." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 338, + 546, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 338, + 546, + 388 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 546, + 388 + ], + "type": "text", + "content": "appearing brighter or darker accordingly. This highlights our method's ability to effectively capture pertinent lighting conditions, and underscores the significance of the initial singular time step sampling." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 395, + 453, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 453, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 453, + 407 + ], + "type": "text", + "content": "4.7. Application on ControlNet" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 414, + 546, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 414, + 546, + 486 + ], + "spans": [ + { + "bbox": [ + 305, + 414, + 546, + 486 + ], + "type": "text", + "content": "Our model can seamlessly integrate with existing diffusion model plugins, such as ControlNet [44]. Since our " + }, + { + "bbox": [ + 305, + 414, + 546, + 486 + ], + "type": "inline_equation", + "content": "\\bar{y}_{\\theta}" + }, + { + "bbox": [ + 305, + 414, + 546, + 486 + ], + "type": "text", + "content": " is structurally different from ControlNet, ControlNet is adopted after our initial sampling step. The results in Fig. 8 demonstrate that our method is fully compatible with ControlNet and effectively resolves its average brightness issue." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 497, + 378, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 497, + 378, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 497, + 378, + 510 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 517, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 547, + 662 + ], + "type": "text", + "content": "In this study, we delve into the singularities of time intervals in diffusion models, exploring both theoretical and practical aspects. Firstly, we demonstrate that at both regular and singular time steps, the reverse process can be approximated by a Gaussian distribution. Leveraging these theoretical insights, we conduct an in-depth analysis of singular time step sampling and propose a theoretical solution. Finally, we introduce a novel plug-and-play method SingDiffusion, addressing the initial time-step sampling challenge. Remarkably, this module substantially mitigates the average brightness issue prevalent in most current diffusion models, and also enhances their generative capabilities." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "content": "Acknowledgments. This project is supported by the National Natural Science Foundation of China (62072482, U22A2095)" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6952" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Brian D.O. Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, pages 313-326, 1982. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "type": "text", + "content": "[2] AUTOMATIC1111. Stable diffusion web ui. https://github.com/AUTOMATIC1111/stable-diffusion-webui, 2022.7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "text", + "content": "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18208-18218, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 288, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 268 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 268 + ], + "type": "text", + "content": "[4] Andrew Campbell, Joe Benton, Valentin De Bortoli, Tom Rainforth, George Deligiannidis, and Arnaud Doucet. A continuous time framework for discrete denoising models. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 288, + 312 + ], + "type": "text", + "content": "[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems (NeurIPS), pages 8780-8794, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 288, + 357 + ], + "type": "text", + "content": "[6] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Score-based generative modeling with critically-damped Langevin diffusion. In International Conference on Learning Representations (ICLR), 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 358, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 288, + 412 + ], + "type": "text", + "content": "[7] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 458 + ], + "type": "text", + "content": "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems, 2014. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 458, + 286, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 286, + 491 + ], + "type": "text", + "content": "[9] Nicholas Guttenberg. Diffusion with offset noise. https://www.crosslabs.org/blog/diffusion-with-offset-noise, 2023.2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 286, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 286, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 286, + 535 + ], + "type": "text", + "content": "[10] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In International Conference on Learning Representations (ICLR), 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 286, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 286, + 590 + ], + "type": "text", + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems (NeurIPS), 2017. 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 286, + 624 + ], + "type": "text", + "content": "[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 286, + 667 + ], + "type": "text", + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), pages 6840-6851, 2020. 2, 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 286, + 713 + ], + "type": "text", + "content": "[14] Jonathan Ho, Tim Salimans, Alexey A. Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In ICLR Workshop on Deep Generative Models for Highly Structured Data, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[15] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems (NeurIPS), pages 26565-26577, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "text", + "content": "[16] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6007-6017, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[17] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. In Advances in Neural Information Processing Systems (NeurIPS), pages 21696-21707, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "text", + "content": "[18] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "text", + "content": "[19] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv preprint arXiv:2305.08891, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 298, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 352 + ], + "type": "text", + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision (ECCV), pages 740-755, 2014. 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "text", + "content": "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "text", + "content": "[22] Yubin Lu, Zhongjian Wang, and Guillaume Bal. Mathematical analysis of singularities in the diffusion model under the submanifold assumption. arXiv preprint arXiv:2301.07882, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "text", + "content": "[23] Andreas Lugmayr, Martin Danelljan, Luc Van Gool, and Radu Timofte. Srlflow: Learning the super-resolution space with normalizing flow. In European Conference on Computer Vision (ECCV), page 715-732, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 478, + 545, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 500 + ], + "type": "text", + "content": "[24] David McAllester. On the mathematics of diffusion models. arXiv preprint arXiv:2301.11108, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 502, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 502, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 502, + 545, + 555 + ], + "type": "text", + "content": "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations (ICLR), 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 557, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 624 + ], + "type": "text", + "content": "[26] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning (ICML), pages 16784-16804, 2022. 2, 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "text", + "content": "[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[28] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. Grad-tts: A diffusion probabilistic model for text-to-speech. In International Confer-" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6953" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "ence on Machine Learning (ICML), pages 8599-8608, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 163 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 217 + ], + "type": "text", + "content": "[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning (ICML), pages 8821-8831, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 275 + ], + "type": "text", + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 277, + 287, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 320 + ], + "type": "text", + "content": "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 322, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 322, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 322, + 287, + 387 + ], + "type": "text", + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22500-22510, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 389, + 287, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 389, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 48, + 389, + 287, + 433 + ], + "type": "text", + "content": "[34] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH Conference Proceedings, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 434, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 521 + ], + "type": "text", + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems (NeurIPS), pages 36479-36494, 2022. 2, 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 524, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 524, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 524, + 287, + 556 + ], + "type": "text", + "content": "[36] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations (ICLR), 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 558, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 656 + ], + "type": "text", + "content": "[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems (NeurIPS), pages 25278-25294, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[38] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning (ICML), pages 2256-2265, 2015. 2, 3, 4" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 474 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "text", + "content": "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[41] Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. In International Conference on Learning Representations (ICLR), 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "text", + "content": "[42] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 263, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 545, + 328 + ], + "type": "text", + "content": "[43] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18381-18391, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "type": "text", + "content": "[44] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3836-3847, 2023. 3, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 376, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 430 + ], + "type": "text", + "content": "[45] Pengze Zhang, Lingxiao Yang, Jian-Huang Lai, and Xiaohua Xie. Exploring dual-task correlation for pose guided person image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7713-7722, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 431, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 431, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 307, + 431, + 545, + 474 + ], + "type": "text", + "content": "[46] Pengze Zhang, Hubery Yin, Chen Li, and Xiaohua Xie. Formulating discrete probability flow through optimal transport. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6954" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Tactile-Augmented Radiance Fields/f9e9c348-b82c-415a-b07e-5b945a68c375_content_list.json b/2024/Tactile-Augmented Radiance Fields/f9e9c348-b82c-415a-b07e-5b945a68c375_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0e1da795f87b5eae70c3bad4c9830b17f0e8d8ee --- /dev/null +++ b/2024/Tactile-Augmented Radiance Fields/f9e9c348-b82c-415a-b07e-5b945a68c375_content_list.json @@ -0,0 +1,1463 @@ +[ + { + "type": "text", + "text": "Tactile-Augmented Radiance Fields", + "text_level": 1, + "bbox": [ + 305, + 130, + 666, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiming Dou $^{1}$ Fengyu Yang $^{2}$ Yi Liu $^{1}$ Antonio Loquercio $^{3}$ Andrew Owens $^{1}$", + "bbox": [ + 140, + 178, + 803, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ University of Michigan $^{2}$ Yale University $^{3}$ UC Berkeley", + "bbox": [ + 243, + 207, + 725, + 227 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f549b524105136b5fc7c79c0afb9d674e867f0888bda77c6bc2a1700dc4bb4d.jpg", + "image_caption": [ + "Figure 1. Tactile-augmented radiance fields. We capture a tactile-augmented radiance field (TaRF) from photos and sparsely sampled touch probes. To do this, we register the captured visual and tactile signals into a shared 3D space, then train a diffusion model to impute touch at other locations within the scene. Here, we visualize two touch probes and their (color coded) 3D positions in the scene. We also show two touch signals estimated by the diffusion model. The touch signals were collected using a vision-based touch sensor [32] that represents the touch signals as images. Please see our project page for video results." + ], + "image_footnote": [], + "bbox": [ + 78, + 257, + 632, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/163f41aec7e0b4020a7810e4484f53cc9989565a88fdaa2782da932bf26321d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 257, + 890, + 429 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 532, + 313, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a scene representation, which we call a tactile-augmented radiance field (TaRF), that brings vision and touch into a shared 3D space. This representation can be used to estimate the visual and tactile signals for a given 3D position within a scene. We capture a scene's TaRF from a collection of photos and sparsely sampled touch probes. Our approach makes use of two insights: (i) common vision-based touch sensors are built on ordinary cameras and thus can be registered to images using methods from multi-view geometry, and (ii) visually and structurally similar regions of a scene share the same tactile features. We use these insights to register touch signals to a captured visual scene, and to train a conditional diffusion model that, provided with an RGB-D image rendered from a neural radiance field, generates its corresponding tactile signal. To evaluate our approach, we collect a dataset of TaRFs. This dataset contains more touch samples than previous real-world datasets, and it provides spatially aligned visual signals for each captured touch signal. We demonstrate the accuracy of our cross-modal generative model and the utility of the captured visual-tactile data on several downstream tasks. Project page: https://dou-yiming.github.io/TaRF.", + "bbox": [ + 75, + 565, + 473, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 532, + 630, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As humans, our ability to perceive the world relies crucially on cross-modal associations between sight and touch [19, 50]. Tactile sensing provides a detailed understanding of material properties and microgeometry, such as the intricate patterns of bumps on rough surfaces and the complex motions that soft objects make when they deform. This type of understanding, which largely eludes today's computer vision models, is a critical component of applications that require reasoning about physical contact, such as robotic locomotion [3, 24, 31, 34, 37, 38] and manipulation [6, 7, 11, 42, 60], and methods that simulate the behavior of materials [4, 13, 40, 41].", + "bbox": [ + 496, + 563, + 892, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In comparison to many other modalities, collecting tactile data is an expensive and tedious process, since it requires direct physical interaction with the environment. A recent line of work has addressed this problem by having humans or robots probe the environment with touch sensors (see Table 1). Early efforts have been focused on capturing the properties of only a few objects either in simulation [16, 17, 52] or in lab-controlled settings [6, 7, 18, 28, 35, 52, 63], which may not fully convey the diversity of tactile signals in natural environments. Other works have gone beyond a", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26529", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/0cf3144a6253dfddb1641ce1efae8011216f8faaafc4e38dde03f09c7062ccdb.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Dataset comparison. We present the number of real visual-tactile pairs and whether such pairs are visually aligned, i.e., whether the visual image includes an occlusion-free view of the touched surface. *YCB-Slide has real-world touch probes but synthetic images rendered with CAD models of YCB objects on a white background [9]." + ], + "table_body": "
DatasetSamplesAlignedScenarioSource
More Than a Feeling [7]6.5k×TabletopRobot
Feeling of Success [6]9.3k×TabletopRobot
VisGel [35]12k×TabletopRobot
SSVTP [28]4.6k✓TabletopRobot
ObjectFolder 1.0 [16]-✓ObjectSynthetic
ObjectFolder 2.0 [17]-✓ObjectSynthetic
ObjectFolder Real [18]3.7k×ObjectRobot
Burka et al. [5]1.1k×Sub-sceneHuman
Touch and Go [56]13.9k×Sub-sceneHuman
YCB-Slide* [52]-✓ObjectHuman
Touching a NeRF [63]1.2k✓ObjectRobot
TaRF (Ours)19.3k✓Full sceneHuman
", + "bbox": [ + 80, + 88, + 472, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lab setting and have collected touch from real scenes [5, 56]. However, existing datasets lack aligned visual and tactile information, since the touch sensor and the person (or robot) that holds it often occlude large portions of the visual scene (Fig. 2). These datasets also contain only a sparse set of touch signals for each scene, and it is not clear how the sampled touch signals relate to each other in 3D.", + "bbox": [ + 75, + 354, + 468, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present a simple and low-cost procedure to capture quasi-dense, scene-level, and spatially-aligned visual and touch data (Fig. 1). We call the resulting scene representation a tactile-augmented radiance field (TaRF). We remove the need for robotic collection by leveraging a 3D scene representation (a NeRF [39]) to synthesize a view of the surface being touched, which results in spatially aligned visual-tactile data (Fig. 2). We collect this data by mounting a touch sensor to a camera with commonly available materials (Fig. 3). To calibrate the pair of sensors, we take advantage of the fact that popular vision-based touch sensors [25, 26, 32, 48] are built on ordinary cameras. The relative pose between the vision and tactile sensors can thus be estimated using traditional methods from multi-view geometry, such as camera resectioning [20].", + "bbox": [ + 75, + 462, + 468, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We use this procedure to collect a large real-world dataset of aligned visual-tactile data. With this dataset, we train a diffusion model [45, 51] to estimate touch at locations not directly probed by a sensor. In contrast to the recent work of Zhong et al. [63], which also estimates touch from 3D NeRF geometry, we create scene-scale reconstructions, we do not require robotic proprioception, and we use diffusion models [51]. This enables us to obtain tactile data at a much larger scale, and with considerably more diversity. Unlike previous visual-tactile diffusion work [57], we condition the model on spatially aligned visual and depth information, enhancing the generated samples' quality and their usefulness in downstream applications. After training, the diffusion model can be used to predict tactile informa", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/48963dd64dd1561761f3cb4229bcb719dd4ca3e677470e89430566795a3499d9.jpg", + "image_caption": [ + "OF 2.0 [17]" + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 565, + 183 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3df34be0c7d888260703250ad091c94e3386f9e804c9570839d68d4a1c58539d.jpg", + "image_caption": [ + "VisGel [35]" + ], + "image_footnote": [], + "bbox": [ + 566, + 89, + 629, + 184 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ca5bdb3189ea7e7413fe4ab6b4f7dd5ae94e7eb6e988f566e701d29386cb889e.jpg", + "image_caption": [ + "OF Real [18]" + ], + "image_footnote": [], + "bbox": [ + 630, + 89, + 694, + 184 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/66408a8ce54ef733c07222aed6a558248563d7ddf5d651fa8124f3052351a023.jpg", + "image_caption": [ + "SSVTP [28]", + "Figure 2. Visual-tactile examples. In contrast to the visual-tactile data captured in previous work, our approach allows us to sample unobstructed images that are spatially aligned with the touch signal, from arbitrary 3D viewpoints using a NeRF." + ], + "image_footnote": [], + "bbox": [ + 696, + 89, + 758, + 184 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/48eeffb2c37b18e743fe79503e6b089918e7a58f0fd656316a65967904212e1f.jpg", + "image_caption": [ + "TG [56]" + ], + "image_footnote": [], + "bbox": [ + 761, + 89, + 821, + 184 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fc56a6478352a81e2bb1314e0cb0dcb53f113f7f611781af4eef5663883d9afd.jpg", + "image_caption": [ + "TaRF (Ours)" + ], + "image_footnote": [], + "bbox": [ + 823, + 89, + 885, + 184 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion for novel positions in the scene. Analogous to quasi-dense stereo methods [15, 33], the diffusion model effectively propagates sparse touch samples, obtained by probing, to other visually and structurally similar regions of the scene.", + "bbox": [ + 496, + 268, + 890, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our visual-tactile model's ability to accurately perform cross-modal translation using a variety of quality metrics. We also apply it to several downstream tasks, including localizing a touch within a scene and understanding material properties of the touched area. Our experiments suggest:", + "bbox": [ + 496, + 344, + 892, + 435 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Touch signals can be localized in 3D space by exploiting multi-view geometry constraints between sight and touch.", + "- Estimated touch measurements from novel views are not only qualitatively accurate, but also beneficial on downstream tasks.", + "- Cross-modal prediction models can accurately estimate touch from sight for natural scenes.", + "- Visually-acquired 3D scene geometry improves cross-modal prediction." + ], + "bbox": [ + 500, + 436, + 890, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 580, + 640, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual-tactile datasets. Previous work has either used simulators [16, 17] or robotic arms [6, 8, 18, 35, 63] for data generation. Our work is closely related to that of Zhong et al. [63], which uses a NeRF and captured touch data to generate a tactile field for several small objects. They use the proprioception of an expensive robot to spatially align vision and touch. In contrast, we leverage the properties of the tactile sensor and novel view synthesis to use commonly available material (a smartphone and a selfie stick) to align vision and touch. This enables the collection of a larger, scene-level, and more diverse dataset, on which we train a higher-capacity diffusion model (rather than a conditional GAN). Like several previous works [5, 56], we also collect scene-level data. In contrast to them, we spatially align the signals by registering them in a unified 3D representation, thereby increasing the prediction power of the visual-tactile generative model.", + "bbox": [ + 496, + 609, + 892, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Capturing multimodal 3D scenes. Our work is related to methods that capture 3D visual reconstructions of spaces", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26530", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "using RGB-D data [12, 49, 55, 59] and multimodal datasets of paired 3D vision and language [1, 2, 10]. Our work is also related to recent methods that localize objects in NeRFs using joint embeddings between images and language [29] or by semantic segmentation [62]. In contrast to language supervision, touch is tied to a precise position in a scene.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D touch sensing. A variety of works have studied the close relationship between geometry and touch, motivating our use of geometry in imputing touch. Johnson et al. [25, 26] proposed vision-based touch sensing, and showed that highly accurate depth can be estimated from the touch sensor using photometric stereo. Other work has estimated object-scale 3D from touch [54]. By contrast, we combine sparse estimates of touch with quasi-dense tactile signals estimated using generative models.", + "bbox": [ + 75, + 185, + 470, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-modal prediction of touch from sight. Recent work has trained generative models that predict touch from images. Li et al. [35] used a GAN to predict touch for images of a robotic arm, while Gao et al. [18] applied them to objects collected on a turntable. Yang et al. [57] used latent diffusion to predict touch from videos of humans touching objects. Our goal is different from these works: we want to predict touch signals that are spatially aligned with a visual signal, to exploit scene-specific information, and to use geometry. Thus, we use a different architecture and conditioning signal, and fit our model to examples from the same scenes at training and test time. Other work has learned joint embeddings between vision and touch [28, 36, 56, 58, 61].", + "bbox": [ + 75, + 325, + 468, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 549, + 166, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We collect visual and tactile examples from a scene and register them together with a 3D visual reconstruction to build a TaRF. Specifically, we capture a NeRF $F_{\\theta}:(\\mathbf{x},\\mathbf{r})\\mapsto (\\mathbf{c},\\sigma)$ that maps a 3D point $\\mathbf{x} = (x,y,z)$ and viewing direction $\\mathbf{r}$ to its corresponding RGB color $\\mathbf{c}$ and density $\\sigma$ [39]. We associate to the visual representation a touch model $F_{\\phi}:\\mathbf{v}_{t}\\mapsto \\pmb{\\tau}$ that generates the tactile signal that one would obtain by touching at the center of the image $\\mathbf{v}_t$ . In the following, we explain how to estimate $F_{\\theta}$ and $F_{\\phi}$ and put them into the same shared 3D space.", + "bbox": [ + 75, + 574, + 468, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Capturing vision and touch signals", + "text_level": 1, + "bbox": [ + 76, + 734, + 379, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Obtaining a visual 3D reconstruction. We build the visual NeRF, $F_{\\theta}$ , closely following previous work [12, 55]. A human data collector moves through a scene and records a video, covering as much of the space as possible. We then estimate camera pose using structure from motion [47] and create a NeRF using off-the-shelf packages [53]. Additional details are provided in the supplement.", + "bbox": [ + 75, + 760, + 468, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Capturing and registering touch. We simultaneously collect tactile and visual signals by mounting a touch sensor", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/04fa5dd0c1d1252de232aed3efddd62a6bfb07a6e9a07d34030942e3d1e09d7a.jpg", + "image_caption": [ + "Figure 3. Capturing setup. (a) We record paired vision and touch signals using a camera attached to a touch sensor. (b) We estimate the relative pose between the touch sensor and the camera using correspondences between sight and touch." + ], + "image_footnote": [], + "bbox": [ + 504, + 88, + 888, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "on a camera (Fig. 3), obtaining synchronized touch signals $\\{\\tau_i\\}_{i=1}^N$ and video frames $\\mathbf{v}$ . We then estimate the pose of the video frames using off-the-shelf structure from motion methods [47], obtaining poses $\\{p_i^v\\}_{i=1}^N$ . Finally, we use the calibration of the mount to obtain the poses $\\{p_i^t\\}_{i=1}^N$ of the tactile measurements with respect to the scene's global reference frame. As a collection device, we mount an iPhone 14 Pro to one end of a camera rod, and a DIGIT [32] touch sensor to the other end. Note that the devices can be replaced with any RGB-D camera and vision-based tactile sensor.", + "bbox": [ + 498, + 308, + 890, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Capturing setup calibration. To find the relative pose between the camera and the touch sensor (Fig. 3), we exploit the fact that arbitrary viewpoints can be synthesized from $F_{\\theta}$ , and that ubiquitous vision-based touch sensors are based on perspective cameras. In these sensors, an elastomer gel is placed on the lens of a commodity camera, which is illuminated by colored lights. When the gel is pressed into an object, it deforms, and the camera records an image of the deformation; this image is used as the tactile signal. This design allows us to estimate the pose of the tactile sensor through multi-view constraints from visual-tactile correspondences: pixels in visual images and tactile images that are of the same physical point.", + "bbox": [ + 496, + 477, + 890, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We start the calibration process by synthesizing novel views from $F_{\\theta}$ . The views are generated at the camera location $\\{p_i^v\\}_{i = 1}^N$ , but rotated $90^{\\circ}$ on the $x$ -axis. This is because the camera is approximately orthogonal to the touch sensor (see Fig. 3). Then, we manually annotate corresponding pixels between the touch measurements and the generated frames (Fig. 3). To simplify and standardize this process, we place a braille board in each scene and probe it with the touch sensor. This will generate a distinctive touch signal that is easy to localize [23].", + "bbox": [ + 496, + 674, + 890, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We formulate the problem of estimating the six degrees of freedom relative pose $(\\mathbf{R},\\mathbf{t})$ between the touch sensor and the generated frames as a resectioning problem [20]. We use the estimated 3D structure from the NeRF $F_{\\theta}$ to obtain 3D points $\\{\\mathbf{x}_i\\}_{i = 1}^M$ for each of the annotated corre", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26531", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "spondences. Each point has a pixel position $\\mathbf{u}_i\\in \\mathcal{R}^2$ in the touch measurement. We find $(\\mathbf{R},\\mathbf{t})$ by minimizing the reprojection error:", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {R}, \\mathbf {t}} \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\| \\pi (\\mathbf {K} [ \\mathbf {R} \\mid \\mathbf {t} ], \\mathbf {X} _ {i}) - \\mathbf {u} _ {i} \\| _ {1}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 147, + 468, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pi$ projects a 3D point using a given projection matrix, $\\mathbf{K}$ are the known intrinsic of the tactile sensor's camera, and the point $\\mathbf{X}_i$ is in the coordinate system of the generated vision frames. We perform the optimization on 6-15 annotated correspondences from the braille board. For robustness, we compute correspondences from multiple frames. We represent the rotation matrix using quaternions and optimize using nonlinear least-squares. Once we have $(\\mathbf{R},\\mathbf{t})$ with respect to the generated frames, we can derive the relative pose between the camera and the touch sensor.", + "bbox": [ + 75, + 199, + 468, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Imputing the missing touch", + "text_level": 1, + "bbox": [ + 76, + 362, + 323, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use a generative model to estimate the touch signal (represented as an image from a vision-based touch sensor) for other locations within the scene. Specifically, we train a diffusion model $p_{\\phi}(\\pmb{\\tau} \\mid \\mathbf{v}, \\mathbf{d}, \\mathbf{b})$ , where $\\mathbf{v}$ and $\\mathbf{d}$ are images and depth maps extracted from $F_{\\theta}$ (see Fig. 4). We also pass as input to the diffusion model a background image captured by the touch sensor when it is not in contact with anything, denoted as $\\mathbf{b}$ . Although not essential, we have observed that this additional input empirically improves the model's performance (e.g., Fig. 1 the background provides the location of defects in the gel, which appear as black dots). We train the model $p_{\\phi}$ on our entire vision-touch dataset (Sec. 4).", + "bbox": [ + 75, + 385, + 468, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The training of $p_{\\phi}$ is divided into two stages. In the first, we pre-train a cross-modal visual-tactile encoder with self-supervised contrastive learning on our dataset. This stage, initially proposed by [23, 57], is equivalent to the self-supervised encoding pre-training that is common for image generation models [45]. We use a ResNet-50 [21] as the backbone for this contrastive model.", + "bbox": [ + 75, + 566, + 468, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the second stage, we use the contrastive model to generate the input for a conditional latent diffusion model, which is built upon Stable Diffusion [45]. A frozen pretrained VQ-GAN [14] is used to obtain the latent representation with a spatial dimension of $64 \\times 64$ . We start training the diffusion model from scratch and pre-train it on the task of unconditional tactile image generation on the YCB-Slide dataset [52]. After this stage, we train the conditional generative model $p_{\\phi}$ on our spatially aligned visual-tactile dataset, further fine-tuning the contrastive model end-to-end with the generation task.", + "bbox": [ + 75, + 672, + 468, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At inference time, given a novel location in the 3D scene, we first render the visual signals $\\hat{\\mathbf{v}}$ and $\\hat{\\mathbf{d}}$ from NeRF, and then estimate the touch signal $\\hat{\\tau}$ of the position using the diffusion model.", + "bbox": [ + 75, + 839, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/90c5c0e0784f49af905258e53b93778c4f6483f3d7b65ffd45d33327ff63e174.jpg", + "image_caption": [ + "Figure 4. Touch estimation. We estimate the tactile signal for a given touch sensor pose $(\\mathbf{R},\\mathbf{t})$ . To do this, we synthesize a viewpoint from the NeRF, along with a depth map. We use conditional latent diffusion to predict the tactile signal from these inputs." + ], + "image_footnote": [], + "bbox": [ + 498, + 90, + 890, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. A 3D Visual-Tactile Dataset", + "text_level": 1, + "bbox": [ + 498, + 311, + 754, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following, we show the details of the data collection process and statistics of our dataset.", + "bbox": [ + 496, + 338, + 890, + 368 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Data Collection Procedure", + "text_level": 1, + "bbox": [ + 498, + 377, + 738, + 392 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The data collection procedure is divided into two stages. First, we collect multiple views from the scene, capturing enough frames around the areas we plan to touch. During this stage, we collect approximately 500 frames. Next, we collect synchronized visual and touch data, maximizing the geometry and texture being touched. We then estimate the camera location of the vision frames collected in the previous two stages using off-the-shelf mapping tools [47]. After estimating the camera poses for the vision frames, the touch measurements' poses can be derived by using the mount calibration matrix. More details about the pose estimation procedure can be found in the supplement.", + "bbox": [ + 496, + 400, + 890, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we associate each touch sensor with a color image by translating the sensor poses upwards by 0.4 meters and querying the NeRF with such poses. The field of view we use when querying the NeRF is $50^{\\circ}$ . This provides us with approximately 1,500 temporally aligned vision-touch image pairs per scene. Note that this collection procedure is scalable since it does not require specific expertise or equipment and generates abundant scene-level samples.", + "bbox": [ + 496, + 582, + 890, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Dataset Statistics", + "text_level": 1, + "bbox": [ + 500, + 710, + 666, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We collect our data in 13 ordinary scenes including two offices, a workroom, a conference room, a corridor, a tabletop, a corridor, a lounge, a room with various clothes and four outdoor scenes with interesting materials. Typically, we collect 1k to 2k tactile probes in each scene, resulting in a total of 19.3k image pairs in the dataset.", + "bbox": [ + 496, + 734, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Some representative samples from the collected dataset are shown in Fig. 5. Our data includes a large variety of geometry (edges, surfaces, corners, etc.) and texture (plastic, clothes, snow, wood, etc.) of different materials in the scene. During capturing process, the collector will try to", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26532", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0425ab1485baea9d4aa79b2049a478df75308f6518983cc365682f89fecd9677.jpg", + "image_caption": [ + "Figure 5. Representative examples from the captured dataset. Our dataset is obtained from nine everyday scenes, such as offices, classrooms, and kitchens. We show three such scenes in the figure above, together with samples of spatially aligned visual and tactile data. In each scene, 1k to 2k tactile probes were collected, resulting in a total of 19.3k image pairs. The data encompasses diverse geometries (edges, surfaces, corners, etc.) and textures (plastic, clothes, snow, wood, etc.) of various materials. The collector systematically probed different objects, covering areas with distinct geometry and texture using different sensor poses." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 348, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/72ef09a2a86fd21d91441c2fef904784a3aa1d8591bf70e004561ad357c0d92f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 90, + 617, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ffab580b6af5c26b8f31f7f9ce073199f46a20eafd6a214a54dcc49186fff164.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 90, + 890, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "thoroughly probe various objects and cover the interesting areas with more distinguishable geometry and texture with different sensor poses. To the best of our knowledge, our dataset is the first dataset that captures full, scene-scale spatially aligned vision-touch image pairs. We provide more details about the dataset in the supplement.", + "bbox": [ + 75, + 353, + 468, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 457, + 209, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Leveraging the spatially aligned image and touch pairs from our dataset, we first conduct experiments on dense touch estimation. We then show the effectiveness of both the aligned data pairs and the synthesized touch signals by conducting tactile localization and material classification as two downstream tasks.", + "bbox": [ + 75, + 482, + 468, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 583, + 294, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NeRF. We use the Nerfacto method from Nerfstudio [53]. For each scene, we utilize approximately 2,000 images as training set, which thoroughly cover the scene from various view points. We train the network with a base learning rate of $1 \\times 10^{-2}$ using Adam [30] optimizer for 200,000 steps on a single NVIDIA RTX 2080 Ti GPU to achieve optimal performance.", + "bbox": [ + 75, + 609, + 468, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visual-tactile contrastive model. Following prior works [27, 57], we leverage contrastive learning methods to train a ResNet-50 [21] as visual encoder. The visual and tactile encoders share the same architecture but have different weights. We encode visual and tactile data into latent vectors in the resulting shared representation space. We set the dimension of the latent vectors to 32. Similar to CLIP [43], the model is trained on InfoNCE loss obtained from the pairwise dot products of the latent vectors. We train the model for 20 epochs by Adam [30] optimizer with a learning rate of $10^{-4}$ and batch size of 256 on 4 NVIDIA RTX 2080 Ti GPUs.", + "bbox": [ + 75, + 719, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visual-tactile generative model. Our implementation of the diffusion model closely follows Stable Diffusion [46], with the difference that we use a ResNet-50 to generate the visual encoding from RGB-D images for conditioning. Specifically, we also add the RGB-D images rendered from the tactile sensors' poses into the conditioning, which we refer to in Sec. 5.2 as multiscale conditioning. The model is optimized for 30 epochs by Adam [30] optimizer with a base learning rate of $10^{-5}$ . The learning rate is scaled bygpu number $\\times$ batch size. We train the model with batch size of 48 on 4 NVIDIA A40 GPUs. At inference time, the model conducts 200 steps of denoising process with a 7.5 guidance scale. Following prior cross-modal synthesis work [44], we use reranking to improve the prediction quality. We obtain 16 samples from the diffusion model for every instance and re-rank the samples with our pretrained contrastive model. The sample with highest similarity is the final prediction.", + "bbox": [ + 496, + 353, + 890, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Dense Touch Estimation", + "text_level": 1, + "bbox": [ + 500, + 635, + 720, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental setup. We now evaluate the diffusion model's ability to generate touch images. To reduce overlap between the training and test set, we first split the frames into sequences temporally (following previous work [56]). We split them into sequences of 50 touch samples, then divide these sequences into train/validation/test with a ratio of $8/1/1$ . We evaluate the generated samples on Frechet Inception Distance (FID), a standard evaluation metric for cross-modal generation [56]. We also include Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM), though we note that these metrics are highly sensitive to spatial position of the generated content, and can be optimized by models that minimize simple pixelwise losses [22]. We also include CVTP metric proposed by prior work [57], which measures the similarity between visual and tactile embeddings of a contrastive model, analogous to", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26533", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ad2ef9d07b3a69c8c5c5c29ecb00ee7e8d1b2af46f64297a32a167a00d04c60d.jpg", + "image_caption": [ + "Figure 6. Qualitative touch estimation results. Each model is conditioned on the RGB image and depth map rendered from the NeRF (left). The white box indicates the tactile sensor's approximate field of view (which is much smaller than the full conditional image). The G.T. column shows the ground truth touch images measured from a DIGIT sensor. L1 and VisGel often generate blurry textures and inaccurate geometry. By contrast, our model better captures the features of the tactile image, e.g., the rock's microgeometry and complex textures and shapes of furniture. The last row shows two failure cases of our model. In both examples, our model generates a touch image that is geometrically misaligned with the ground truth. All of the examples shown here are at least $10\\mathrm{cm}$ away from any training sample." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 480, + 409 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/033fc97f74dea754889694412bebf42006df17669bb7afa432a61228e20e5808.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 85, + 893, + 410 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CLIP [43] score. We compare against two baselines: VisGel, the approach from Li et. [35], which trains a GAN for touch generation, and $L1$ , a model with the same architecture of VisGel but trained to minimize an L1 loss in pixel space.", + "bbox": [ + 75, + 526, + 468, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. As is shown in Table 2, our approach performs much better on the high-level metrics, with up to $4\\mathrm{x}$ lower FID and $80\\mathrm{x}$ higher CVTP. This indicates that our proposed diffusion model captures the distribution and characteristics of the real tactile data more effectively. On the low-level metrics (PSNR and SSIM), all methods are comparable. In particular, the L1 model slightly outperforms the other methods since the loss it is trained on is highly correlated with low-level, pixel-wise metrics. Fig. 6 qualitatively compares samples from the different models. Indeed, our generated samples exhibit enhanced details in micro-geometry of fabrics and richer textures, including snow, wood and carpeting. However, all methods fail on fine details that are barely visible in the image, such as the tree bark.", + "bbox": [ + 75, + 607, + 468, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation study. We evaluate the importance of the main components of our proposed touch generation approach (Table 3). Removing the conditioning on the RGB image results in the most prominent performance drop. This is expected since RGB image uniquely determines the fine", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/af647b6ca632eec9ee0982b7dd159a14a93593539a35f374f487bacdfde9a6b4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelPSNR ↑SSIM ↑FID ↓CVTP ↑
L124.340.8297.050.01
VisGel [35]23.660.81130.220.03
Ours22.840.7228.970.80
", + "bbox": [ + 500, + 523, + 890, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative results on touch estimation for novel views. While comparable on low-level metrics with the baselines, our approach captures the characteristics of the real tactile data more effectively, resulting in a lower FID score.", + "bbox": [ + 498, + 599, + 890, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "grained details of a tactile image. Removing depth image or contrastive pretraining has small effect on CVTP but results in a drop on FID. Contrastive re-ranking largely improves CVTP, indicating the necessity of obtaining multiple samples from the diffusion model. We also find that multiscale conditioning provide a small benefit on FID and CVTP.", + "bbox": [ + 496, + 667, + 890, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Downstream Task I: Tactile Localization", + "text_level": 1, + "bbox": [ + 498, + 770, + 846, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To help understand the quality of the captured TaRFs, we evaluate the performance of the contrastive model (used for conditioning our diffusion model) on the task of tactile localization. Given a tactile signal, our goal is to find the corresponding regions in a 2D image or in a 3D scene that are associated with it, i.e., we ask the question: what part of this image/scene feel like this? We perform the following", + "bbox": [ + 496, + 794, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26534", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e33279555accbb1569ef76ab055bae8e2f4e64e38f017072f23c0184e3ce9e0d.jpg", + "image_caption": [ + "Figure 7. Tactile localization heatmaps. Given a tactile query image, the heatmap shows the image patches with a higher affinity to this tactile signal, as measured by a contrastive model trained on our dataset. We use a sliding window and compare each extracted patch with the touch signal. In each case, the center patch is the true position. Our model successfully captures the correlation between the two signals. This enables it to localize a variety of touch signals, including fine-grained geometry, e.g., a cable or a keyboard, various types of corners and edges, and large uniform regions, such as a clothing. This ability enables our diffusion model to effectively propagate sparse touch samples to other visually and structurally similar regions of the scene." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 893, + 340 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f93fd144e996ebca674e63ac2ce05f35b23ab1bfdaccb7b75e10370fbc73f6de.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model variationPSNR ↑SSIM ↑FID ↓CVTP ↑
Full22.840.7228.970.80
No RGB conditioning22.130.7034.310.76
No depth conditioning22.570.7133.160.80
No contrastive pretraining22.820.7132.980.79
No re-ranking22.920.7229.460.61
No multiscale23.190.7230.890.77
", + "bbox": [ + 78, + 441, + 468, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study. Since the fine-grained details of touch images can be determined from a RGB image, removing conditioning on the latter results in the largest performance drops. Re-ranking has notable impact on CVTP, indicating the necessity of obtaining multiple samples from the diffusion model.", + "bbox": [ + 75, + 566, + 468, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "evaluations on the test set of our dataset. Note that we run no task-specific training.", + "bbox": [ + 75, + 643, + 468, + 675 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "2D Localization. To determine which part of an image are associated with a given tactile measurement, we follow the same setup of SSVTP [28]. We first split the image into patches and compute their embedding. Then, we generate the tactile embedding of the input touch image. Finally, we compute the pairwise similarities between the tactile and visual embeddings, which we plot as a heatmap. As we can see in Fig. 7, our constrastive encoder can successfully capture the correlations between the visual and tactile data. For instance, the tactile embeddings of edges are associated to edges of similar shape in the visual image. Note that the majority of tactile embeddings are highly ambiguous: all edges with a similar geometry feel the same.", + "bbox": [ + 75, + 703, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D Localization. In 3D, the association of an image to tactile measurements becomes less ambiguous. Indeed, since tactile-visual samples are rotation-dependent, objects with similar shapes but different orientations will generate different tactile measurements. Lifting the task to 3D still does not remove all ambiguities (for example, each side of a rectangular table cannot be precisely localized). Nonetheless, we believe it to be a good fit for a quantitative evaluation since it's rare for two ambiguous parts of the scene to be touched with exactly the same orientation.", + "bbox": [ + 496, + 443, + 890, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use the following experimental setup for 3D localization. Given a tactile image as a query, we compute its distance in embedding space to all visual test images from the same scene. Note that all test images are associated with a 3D location. We define as ground-truth correspondences all test images at a distance of at most $r$ from the 3D location of the test sample. We vary $r$ to account for local ambiguities. As typical in the retrieval literature, we benchmark the performance with metric mean Average Precision (mAP).", + "bbox": [ + 496, + 597, + 890, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We consider three baselines: (1) chance, which randomly selects corresponding samples; (2) real, which uses the contrastive model trained on our dataset; and (3) real + estimated, which trains the contrastive model on both dataset samples and a set of synthetic samples generated via the scenes' NeRF and our touch generation model. Specifically, we render a new image and corresponding touch by interpolating the position of two consecutive frames in the training dataset. This results in a training dataset for the contrastive model that is twice as large.", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26535", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7d1fea4ad270197186abc29fa6700fc6f2927437a3b72b4a18f4b1f2d29f6fe9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Datasetr(m)
0.0010.0050.010.050.1
Chance3.556.8210.2518.2621.33
Real12.1022.9332.1050.3057.15
Real + Est.14.9226.6936.1753.6260.61
", + "bbox": [ + 76, + 88, + 470, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results, presented in Table 4, demonstrate the performance benefit of employing both real and synthetic tactile pairs. Combining synthetic tactile images with the original pairs achieves highest performance on all distance thresholds. Overall, this indicates that touch measurements from novel views are not only qualitatively accurate, but also beneficial for this downstream task.", + "bbox": [ + 75, + 275, + 468, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Downstream Task II: Material Classification", + "text_level": 1, + "bbox": [ + 76, + 392, + 454, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We investigate the efficacy of our visual-tactile dataset for understanding material properties, focusing on the task of material classification. We follow the formulation by Yang et al. [56], which consists of three subtasks: (i) material classification, requiring the distinction of materials among 20 possible classes; (ii) softness classification, a binary problem dividing materials as either hard or soft; and (iii) hardness classification, which requires the classification of materials as either rough or smooth.", + "bbox": [ + 75, + 415, + 468, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We follow the same experimental procedure of [56]: we pretrain a contrastive model on a dataset and perform linear probing on the sub-tasks' training set. Our experiments only vary the pretraining dataset, leaving all architectural choices and hyperparameters the same. We compare against four baselines. A random classifier (chance); the ObjectFolder 2.0 dataset [17]; the VisGel dataset [35]; and the Touch and Go dataset [56]. Note that the touch sensor used in the test data (GelSight) differs from the one used in our dataset (DIGIT). Therefore, we use for pretraining a combination of our dataset and Touch and Go. To ensure a fair comparison, we also compare to the combination of each dataset and Touch and Go.", + "bbox": [ + 75, + 551, + 468, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The findings from this evaluation, as shown in Table 5, suggest that our data improves the effectiveness of the contrastive pretraining objective, even though our data is from a different distribution. Moreover, we find that adding estimated touch probes for pretraining results in a higher performance on all the three tasks, especially the smoothness classification. This indicates that not only does our dataset covers a wide range of materials but also our diffusion model captures the distinguishable and useful patterns of different materials.", + "bbox": [ + 75, + 750, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e87791c10c19195a6d381ea19fd909b3eadcfe79b7e601c059dd87b16e5610e2.jpg", + "table_caption": [ + "Table 4. Quantitative results on 3D tactile localization. We evaluate using mean Average Precision (mAP) as a metric. Training the contrastive model on our dataset of visually aligned real samples together with estimated samples from new locations in the scene results in the highest performance." + ], + "table_footnote": [], + "table_body": "
DatasetMaterialHard/ SoftRough/ Smooth
Chance18.666.156.3
ObjectFolder 2.0 [17]36.272.069.0
VisGel [35]39.169.470.4
Touch and Go [56]54.777.379.4
+ ObjectFolder 2.0 [17]54.687.384.8
+ VisGel [35]53.186.783.6
+ Ours* (Real)57.688.481.7
+ Ours* (Real + Estimated)59.088.786.1
", + "bbox": [ + 500, + 88, + 893, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Material classification. We show the downstream material recognition accuracy of models pre-trained on different datasets. The final rows show the performance when combining different datasets with Touch and Go [56]. * The task-specific training and testing datasets for this task are collected with a Gel-Sight sensor. We note that our data comes from a different distribution, since it is collected with a DIGIT sensor [32].", + "bbox": [ + 496, + 255, + 893, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 385, + 619, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present the TaRF, a scene representation that brings vision and touch into a shared 3D space. This representation enables the generation of touch probes for novel scene locations. To build this representation, we collect the largest dataset of spatially aligned vision and touch probes. We study the utility of both the representation and the dataset in a series of qualitative and quantitative experiments and on two downstream tasks: 3D touch localization and material recognition. Overall, our work makes the first step towards giving current scene representation techniques an understanding of not only how things look, but also how they feel. This capability could be critical in several applications ranging from robotics to the creation of virtual worlds that look and feel like the real world.", + "bbox": [ + 496, + 420, + 893, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Since the touch sensor is based on a highly zoomed-in camera, small (centimeter-scale) errors in SfM or visual-tactile registration can lead to misalignments of several pixels between the views of the NeRF and the touch samples, which can be seen in our TaRFs. Another limitation of the proposed representation is the assumption that the scene's coarse-scale structure does not change when it is touched, an assumption that may be violated for some inelastic surfaces.", + "bbox": [ + 496, + 645, + 893, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. We thank Jeongsoo Park, Ayush Shrivastava, Daniel Geng, Ziyang Chen, Zihao Wei, Zixuan Pan, Chao Feng, Chris Rockwell, Gaurav Kaul and the reviewers for the valuable discussion and feedback. This work was supported by an NSF CAREER Award #2339071, a Sony Research Award, the DARPA Machine Common Sense program, and ONR MURI award N00014-21-1-2801.", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26536", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, 2020. 3", + "[2] Peter Anderson, Ayush Shrivastava, Joanne Truong, Arjun Majumdar, Devi Parikh, Dhruv Batra, and Stefan Lee. Sim-to-real transfer for vision-and-language navigation. In Conference on Robot Learning, pages 671-681. PMLR, 2021. 3", + "[3] Jakub Bednarek, Michal Bednarek, Lorenz Wellhausen, Marco Hutter, and Krzysztof Walas. What am i touching? learning to classify terrain via haptic sensing. In 2019 International Conference on Robotics and Automation (ICRA), pages 7187-7193. IEEE, 2019. 1", + "[4] Katherine L Bouman, Bei Xiao, Peter Battaglia, and William T Freeman. Estimating the material properties of fabric from video. In Proceedings of the IEEE international conference on computer vision, pages 1984-1991, 2013. 1", + "[5] Alexander Burka. Instrumentation, data, and algorithms for visually understanding haptic surface properties. 2018. 2", + "[6] Roberto Calandra, Andrew Owens, Manu Upadhyaya, Wenzhen Yuan, Justin Lin, Edward H. Adelson, and Sergey Levine. The feeling of success: Does touch sensing help predict grasp outcomes? Conference on Robot Learning (CoRL), 2017. 1, 2", + "[7] Roberto Calandra, Andrew Owens, Dinesh Jayaraman, Justin Lin, Wenzhen Yuan, Jitendra Malik, Edward H. Adelson, and Sergey Levine. More than a feeling: Learning to grasp and regrasp using vision and touch. Robotics and Automation Letters (RA-L), 2018. 1, 2", + "[8] Roberto Calandra, Andrew Owens, Dinesh Jayaraman, Justin Lin, Wenzhen Yuan, Jitendra Malik, Edward H. Adelson, and Sergey Levine. More than a feeling: Learning to grasp and regrasp using vision and touch. IEEE Robotics and Automation Letters, 3:3300-3307, 2018. 2", + "[9] Berk Calli, Arjun Singh, Aaron Walsman, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. The ycb object and model set: Towards common benchmarks for manipulation research. In 2015 international conference on advanced robotics (ICAR), pages 510-517. IEEE, 2015. 2", + "[10] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European conference on computer vision, 2020. 3", + "[11] Alex Church, John Lloyd, Raia Hadsell, and Nathan F Lepora. Deep reinforcement learning for tactile robotics: Learning to type on a braille keyboard. IEEE Robotics and Automation Letters, 5(4):6145-6152, 2020. 1", + "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Abe Davis, Justin G Chen, and Frédo Durand. Image-space modal bases for plausible manipulation of objects in video. ACM Transactions on Graphics (TOG), 2015. 1", + "[14] Mucong Ding, Kezhi Kong, Jingling Li, Chen Zhu, John Dickerson, Furong Huang, and Tom Goldstein. Vq-gnn: A universal framework to scale up graph neural networks using vector quantization. Advances in Neural Information Processing Systems, 34:6733-6746, 2021. 4", + "[15] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence, 32(8):1362-1376, 2009. 2", + "[16] Ruohan Gao, Yen-Yu Chang, Shivani Mall, Li Fei-Fei, and Jiajun Wu. Objectfolder: A dataset of objects with implicit visual, auditory, and tactile representations. In CoRL, 2021. 1, 2", + "[17] Ruohan Gao, Zilin Si, Yen-Yu Chang, Samuel Clarke, Jeannette Bohg, Li Fei-Fei, Wenzhen Yuan, and Jiajun Wu. Objectfolder 2.0: A multisensory object dataset for sim2real transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10598–10608, 2022. 1, 2, 8", + "[18] Ruohan Gao, Yiming Dou, Hao Li, Tanmay Agarwal, Jeannette Bohg, Yunzhu Li, Li Fei-Fei, and Jiajun Wu. The objectfolder benchmark: Multisensory learning with neural and real objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17276–17286, 2023. 1, 2, 3", + "[19] Michael SA Graziano and Charles G Gross. The representation of extrapersonal space: A possible role for bimodal, visual-tactile neurons. 1995. 1", + "[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 2, 3", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 5", + "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5", + "[23] Carolina Higuera, Byron Boots, and Mustafa Mukadam. Learning to read braille: Bridging the tactile reality gap with diffusion models. arXiv preprint arXiv:2304.01182, 2023. 3, 4", + "[24] Mark A Hoepflinger, C David Remy, Marco Hutter, Luciano Spinello, and Roland Siegwart. Haptic terrain classification for legged robots. In 2010 IEEE International Conference on Robotics and Automation, pages 2828-2833. IEEE, 2010. 1", + "[25] Micah K Johnson and Edward H Adelson. Retrographic sensing for the measurement of surface texture and shape. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 1070-1077. IEEE, 2009. 2, 3", + "[26] Micah K Johnson, Forrester Cole, Alvin Raj, and Edward H Adelson. Microgeometry capture using an elastomeric sensor. ACM Transactions on Graphics (TOG), 2011. 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26537", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Justin Kerr, Huang Huang, Albert Wilcox, Ryan Hoque, Jeffrey Ichnowski, Roberto Calandra, and Ken Goldberg. Learning self-supervised representations from vision and touch for active sliding perception of deformable surfaces. arXiv preprint arXiv:2209.13042, 2022. 5", + "[28] Justin Kerr, Huang Huang, Albert Wilcox, Ryan Hoque, Jeffrey Ichnowski, Roberto Calandra, and Ken Goldberg. Self-supervised visuo-tactile pretraining to locate and follow garment features. In Robotics: Science and Systems, 2023. 1, 2, 3, 7", + "[29] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3", + "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. *ICLR*, 2015. 5", + "[31] Hendrik Kolvenbach, Christian Bärtschi, Lorenz Wellhausen, Ruben Grandia, and Marco Hutter. Haptic inspection of planetary soils with legged robots. IEEE Robotics and Automation Letters, 4(2):1626-1632, 2019. 1", + "[32] Mike Lambeta, Po-Wei Chou, Stephen Tian, Brian Yang, Benjamin Maloon, Victoria Rose Most, Dave Stroud, Raymond Santos, Ahmad Byagowi, Gregg Kammerer, et al. Digit: A novel design for a low-cost compact high-resolution tactile sensor with application to in-hand manipulation. IEEE Robotics and Automation Letters, 2020. 1, 2, 3, 8", + "[33] Maxime Lhuillier and Long Quan. A quasi-dense approach to surface reconstruction from uncalibrated images. IEEE transactions on pattern analysis and machine intelligence, 27(3):418-433, 2005. 2", + "[34] Hongyu Li, Snehal Dikhale, Soshi Iba, and Nawid Jamali. Vihope: Visuotactile in-hand object 6d pose estimation with shape completion. IEEE Robotics and Automation Letters, 8 (11):6963-6970, 2023. 1", + "[35] Yunzhu Li, Jun-Yan Zhu, Russ Tedrake, and Antonio Torralba. Connecting touch and vision via cross-modal prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10609-10618, 2019. 1, 2, 3, 6, 8", + "[36] Justin Lin, Roberto Calandra, and Sergey Levine. Learning to identify object instances by touch: Tactile recognition via multimodal matching. In 2019 International Conference on Robotics and Automation (ICRA), pages 3644-3650. IEEE, 2019. 3", + "[37] Antonio Loquercio, Ashish Kumar, and Jitendra Malik. Learning visual locomotion with cross-modal supervision. In 2023 IEEE International Conference on Robotics and Automation (ICRA), 2023. 1", + "[38] Gabriel B Margolis, Xiang Fu, Yandong Ji, and Pulkit Agrawal. Learning physically grounded robot vision with active sensing motor policies. In 7th Annual Conference on Robot Learning, 2023. 1", + "[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Andrew Owens, Phillip Isola, Josh McDermott, Antonio Torralba, Edward H Adelson, and William T Freeman. Visually indicated sounds. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2405-2413, 2016. 1", + "[41] Senthil Purushwalkam, Abhinav Gupta, Danny M Kaufman, and Bryan Russell. Bounce and learn: Modeling scene dynamics with real-world bounces. arXiv preprint arXiv:1904.06827, 2019. 1", + "[42] Haozhi Qi, Brent Yi, Sudharshan Suresh, Mike Lambeta, Yi Ma, Roberto Calandra, and Jitendra Malik. General in-hand object rotation with vision and touch. arXiv preprint arXiv:2309.09979, 2023. 1", + "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 5, 6", + "[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 5", + "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 4", + "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5", + "[47] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3, 4", + "[48] Carmelo Sferrazza and Raffaello D'Andrea. Design, motivation and evaluation of a full-resolution optical tactile sensor. Sensors, 19(4):928, 2019. 2", + "[49] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV 2012, ECCV 2012. 3", + "[50] Linda Smith and Michael Gasser. The development of embodied cognition: Six lessons from babies. Artificial life, 2005. 1", + "[51] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2", + "[52] Sudharshan Suresh, Zilin Si, Stuart Anderson, Michael Kaess, and Mustafa Mukadam. Midastouch: Monte-carlo inference over distributions across sliding touch. In Conference on Robot Learning, pages 319-331. PMLR, 2023. 1, 2, 4", + "[53] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David Mcallister, Justin Kerr, and Angjoo Kanazawa. Nerfstudio: A modu" + ], + "bbox": [ + 501, + 92, + 892, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26538", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "lar framework for neural radiance field development. arXiv preprint arXiv:2302.04264, 2023. 3, 5", + "[54] Shaoxiong Wang, Jiajun Wu, Xingyuan Sun, Wenzhen Yuan, William T Freeman, Joshua B Tenenbaum, and Edward H Adelson. 3d shape perception from monocular vision, touch, and shape priors. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018. 3", + "[55] Jianxiong Xiao, Andrew Owens, and Antonio Torralba. Sun3d: A database of big spaces reconstructed using sfm and object labels. 2013. 3", + "[56] Fengyu Yang, Chenyang Ma, Jiacheng Zhang, Jing Zhu, Wenzhen Yuan, and Andrew Owens. Touch and go: Learning from human-collected vision and touch. Neural Information Processing Systems (NeurIPS) - Datasets and Benchmarks Track, 2022. 2, 3, 5, 8", + "[57] Fengyu Yang, Jiacheng Zhang, and Andrew Owens. Generating visual scenes from touch. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22070-22080, 2023. 2, 3, 4, 5", + "[58] Fengyu Yang, Chao Feng, Ziyang Chen, Hyoungseob Park, Daniel Wang, Yiming Dou, Ziyao Zeng, Xien Chen, Rit Gangopadhyay, Andrew Owens, and Alex Wong. Binding touch to everything: Learning unified multimodal tactile representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3", + "[59] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 3", + "[60] Zhao-Heng Yin, Binghao Huang, Yuzhe Qin, Qifeng Chen, and Xiaolong Wang. Rotating without seeing: Towards in-hand dexterity through touch. arXiv preprint arXiv:2303.10880, 2023. 1", + "[61] Wenzhen Yuan, Shaoxiong Wang, Siyuan Dong, and Edward Adelson. Connecting look and feel: Associating the visual and tactile properties of physical materials. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5580-5588, 2017. 3", + "[62] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. 3", + "[63] Shaohong Zhong, Alessandro Albini, Oiwi Parker Jones, Perla Maiolino, and Ingmar Posner. Touching a nerf: Leveraging neural radiance fields for tactile sensory data generation. In Conference on Robot Learning, pages 1618-1628. PMLR, 2023. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 768 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26539", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Tactile-Augmented Radiance Fields/f9e9c348-b82c-415a-b07e-5b945a68c375_model.json b/2024/Tactile-Augmented Radiance Fields/f9e9c348-b82c-415a-b07e-5b945a68c375_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9bf70f31b26b0a71c2c6fe3154aa760d7e649e11 --- /dev/null +++ b/2024/Tactile-Augmented Radiance Fields/f9e9c348-b82c-415a-b07e-5b945a68c375_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.306, + 0.131, + 0.668, + 0.154 + ], + "angle": 0, + "content": "Tactile-Augmented Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.179, + 0.804, + 0.201 + ], + "angle": 0, + "content": "Yiming Dou\\(^{1}\\) Fengyu Yang\\(^{2}\\) Yi Liu\\(^{1}\\) Antonio Loquercio\\(^{3}\\) Andrew Owens\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.208, + 0.726, + 0.228 + ], + "angle": 0, + "content": "\\(^{1}\\)University of Michigan \\(^{2}\\)Yale University \\(^{3}\\)UC Berkeley" + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.258, + 0.633, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.258, + 0.891, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.443, + 0.894, + 0.514 + ], + "angle": 0, + "content": "Figure 1. Tactile-augmented radiance fields. We capture a tactile-augmented radiance field (TaRF) from photos and sparsely sampled touch probes. To do this, we register the captured visual and tactile signals into a shared 3D space, then train a diffusion model to impute touch at other locations within the scene. Here, we visualize two touch probes and their (color coded) 3D positions in the scene. We also show two touch signals estimated by the diffusion model. The touch signals were collected using a vision-based touch sensor [32] that represents the touch signals as images. Please see our project page for video results." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.533, + 0.314, + 0.549 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.474, + 0.886 + ], + "angle": 0, + "content": "We present a scene representation, which we call a tactile-augmented radiance field (TaRF), that brings vision and touch into a shared 3D space. This representation can be used to estimate the visual and tactile signals for a given 3D position within a scene. We capture a scene's TaRF from a collection of photos and sparsely sampled touch probes. Our approach makes use of two insights: (i) common vision-based touch sensors are built on ordinary cameras and thus can be registered to images using methods from multi-view geometry, and (ii) visually and structurally similar regions of a scene share the same tactile features. We use these insights to register touch signals to a captured visual scene, and to train a conditional diffusion model that, provided with an RGB-D image rendered from a neural radiance field, generates its corresponding tactile signal. To evaluate our approach, we collect a dataset of TaRFs. This dataset contains more touch samples than previous real-world datasets, and it provides spatially aligned visual signals for each captured touch signal. We demonstrate the accuracy of our cross-modal generative model and the utility of the captured visual-tactile data on several downstream tasks. Project page: https://dou-yiming.github.io/TaRF." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.533, + 0.632, + 0.549 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.893, + 0.746 + ], + "angle": 0, + "content": "As humans, our ability to perceive the world relies crucially on cross-modal associations between sight and touch [19, 50]. Tactile sensing provides a detailed understanding of material properties and microgeometry, such as the intricate patterns of bumps on rough surfaces and the complex motions that soft objects make when they deform. This type of understanding, which largely eludes today's computer vision models, is a critical component of applications that require reasoning about physical contact, such as robotic locomotion [3, 24, 31, 34, 37, 38] and manipulation [6, 7, 11, 42, 60], and methods that simulate the behavior of materials [4, 13, 40, 41]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.751, + 0.895, + 0.904 + ], + "angle": 0, + "content": "In comparison to many other modalities, collecting tactile data is an expensive and tedious process, since it requires direct physical interaction with the environment. A recent line of work has addressed this problem by having humans or robots probe the environment with touch sensors (see Table 1). Early efforts have been focused on capturing the properties of only a few objects either in simulation [16, 17, 52] or in lab-controlled settings [6, 7, 18, 28, 35, 52, 63], which may not fully convey the diversity of tactile signals in natural environments. Other works have gone beyond a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26529" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.473, + 0.257 + ], + "angle": 0, + "content": "
DatasetSamplesAlignedScenarioSource
More Than a Feeling [7]6.5k×TabletopRobot
Feeling of Success [6]9.3k×TabletopRobot
VisGel [35]12k×TabletopRobot
SSVTP [28]4.6k✓TabletopRobot
ObjectFolder 1.0 [16]-✓ObjectSynthetic
ObjectFolder 2.0 [17]-✓ObjectSynthetic
ObjectFolder Real [18]3.7k×ObjectRobot
Burka et al. [5]1.1k×Sub-sceneHuman
Touch and Go [56]13.9k×Sub-sceneHuman
YCB-Slide* [52]-✓ObjectHuman
Touching a NeRF [63]1.2k✓ObjectRobot
TaRF (Ours)19.3k✓Full sceneHuman
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.258, + 0.47, + 0.342 + ], + "angle": 0, + "content": "Table 1. Dataset comparison. We present the number of real visual-tactile pairs and whether such pairs are visually aligned, i.e., whether the visual image includes an occlusion-free view of the touched surface. *YCB-Slide has real-world touch probes but synthetic images rendered with CAD models of YCB objects on a white background [9]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.47, + 0.461 + ], + "angle": 0, + "content": "lab setting and have collected touch from real scenes [5, 56]. However, existing datasets lack aligned visual and tactile information, since the touch sensor and the person (or robot) that holds it often occlude large portions of the visual scene (Fig. 2). These datasets also contain only a sparse set of touch signals for each scene, and it is not clear how the sampled touch signals relate to each other in 3D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.688 + ], + "angle": 0, + "content": "In this work, we present a simple and low-cost procedure to capture quasi-dense, scene-level, and spatially-aligned visual and touch data (Fig. 1). We call the resulting scene representation a tactile-augmented radiance field (TaRF). We remove the need for robotic collection by leveraging a 3D scene representation (a NeRF [39]) to synthesize a view of the surface being touched, which results in spatially aligned visual-tactile data (Fig. 2). We collect this data by mounting a touch sensor to a camera with commonly available materials (Fig. 3). To calibrate the pair of sensors, we take advantage of the fact that popular vision-based touch sensors [25, 26, 32, 48] are built on ordinary cameras. The relative pose between the vision and tactile sensors can thus be estimated using traditional methods from multi-view geometry, such as camera resectioning [20]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We use this procedure to collect a large real-world dataset of aligned visual-tactile data. With this dataset, we train a diffusion model [45, 51] to estimate touch at locations not directly probed by a sensor. In contrast to the recent work of Zhong et al. [63], which also estimates touch from 3D NeRF geometry, we create scene-scale reconstructions, we do not require robotic proprioception, and we use diffusion models [51]. This enables us to obtain tactile data at a much larger scale, and with considerably more diversity. Unlike previous visual-tactile diffusion work [57], we condition the model on spatially aligned visual and depth information, enhancing the generated samples' quality and their usefulness in downstream applications. After training, the diffusion model can be used to predict tactile informa" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.566, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.187, + 0.562, + 0.198 + ], + "angle": 0, + "content": "OF 2.0 [17]" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.09, + 0.63, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.187, + 0.627, + 0.198 + ], + "angle": 0, + "content": "VisGel [35]" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.09, + 0.695, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.187, + 0.695, + 0.198 + ], + "angle": 0, + "content": "OF Real [18]" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.09, + 0.759, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.187, + 0.756, + 0.198 + ], + "angle": 0, + "content": "SSVTP [28]" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.09, + 0.823, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.187, + 0.812, + 0.198 + ], + "angle": 0, + "content": "TG [56]" + }, + { + "type": "image", + "bbox": [ + 0.825, + 0.09, + 0.887, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.826, + 0.187, + 0.886, + 0.198 + ], + "angle": 0, + "content": "TaRF (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.203, + 0.892, + 0.259 + ], + "angle": 0, + "content": "Figure 2. Visual-tactile examples. In contrast to the visual-tactile data captured in previous work, our approach allows us to sample unobstructed images that are spatially aligned with the touch signal, from arbitrary 3D viewpoints using a NeRF." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.892, + 0.344 + ], + "angle": 0, + "content": "tion for novel positions in the scene. Analogous to quasi-dense stereo methods [15, 33], the diffusion model effectively propagates sparse touch samples, obtained by probing, to other visually and structurally similar regions of the scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.345, + 0.893, + 0.436 + ], + "angle": 0, + "content": "We evaluate our visual-tactile model's ability to accurately perform cross-modal translation using a variety of quality metrics. We also apply it to several downstream tasks, including localizing a touch within a scene and understanding material properties of the touched area. Our experiments suggest:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.437, + 0.892, + 0.467 + ], + "angle": 0, + "content": "- Touch signals can be localized in 3D space by exploiting multi-view geometry constraints between sight and touch." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.468, + 0.892, + 0.511 + ], + "angle": 0, + "content": "- Estimated touch measurements from novel views are not only qualitatively accurate, but also beneficial on downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.513, + 0.892, + 0.542 + ], + "angle": 0, + "content": "- Cross-modal prediction models can accurately estimate touch from sight for natural scenes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.543, + 0.892, + 0.573 + ], + "angle": 0, + "content": "- Visually-acquired 3D scene geometry improves cross-modal prediction." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.437, + 0.892, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.582, + 0.642, + 0.597 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.893, + 0.867 + ], + "angle": 0, + "content": "Visual-tactile datasets. Previous work has either used simulators [16, 17] or robotic arms [6, 8, 18, 35, 63] for data generation. Our work is closely related to that of Zhong et al. [63], which uses a NeRF and captured touch data to generate a tactile field for several small objects. They use the proprioception of an expensive robot to spatially align vision and touch. In contrast, we leverage the properties of the tactile sensor and novel view synthesis to use commonly available material (a smartphone and a selfie stick) to align vision and touch. This enables the collection of a larger, scene-level, and more diverse dataset, on which we train a higher-capacity diffusion model (rather than a conditional GAN). Like several previous works [5, 56], we also collect scene-level data. In contrast to them, we spatially align the signals by registering them in a unified 3D representation, thereby increasing the prediction power of the visual-tactile generative model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Capturing multimodal 3D scenes. Our work is related to methods that capture 3D visual reconstructions of spaces" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26530" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.182 + ], + "angle": 0, + "content": "using RGB-D data [12, 49, 55, 59] and multimodal datasets of paired 3D vision and language [1, 2, 10]. Our work is also related to recent methods that localize objects in NeRFs using joint embeddings between images and language [29] or by semantic segmentation [62]. In contrast to language supervision, touch is tied to a precise position in a scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.186, + 0.471, + 0.322 + ], + "angle": 0, + "content": "3D touch sensing. A variety of works have studied the close relationship between geometry and touch, motivating our use of geometry in imputing touch. Johnson et al. [25, 26] proposed vision-based touch sensing, and showed that highly accurate depth can be estimated from the touch sensor using photometric stereo. Other work has estimated object-scale 3D from touch [54]. By contrast, we combine sparse estimates of touch with quasi-dense tactile signals estimated using generative models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.47, + 0.537 + ], + "angle": 0, + "content": "Cross-modal prediction of touch from sight. Recent work has trained generative models that predict touch from images. Li et al. [35] used a GAN to predict touch for images of a robotic arm, while Gao et al. [18] applied them to objects collected on a turntable. Yang et al. [57] used latent diffusion to predict touch from videos of humans touching objects. Our goal is different from these works: we want to predict touch signals that are spatially aligned with a visual signal, to exploit scene-specific information, and to use geometry. Thus, we use a different architecture and conditioning signal, and fit our model to examples from the same scenes at training and test time. Other work has learned joint embeddings between vision and touch [28, 36, 56, 58, 61]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.55, + 0.168, + 0.565 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.469, + 0.726 + ], + "angle": 0, + "content": "We collect visual and tactile examples from a scene and register them together with a 3D visual reconstruction to build a TaRF. Specifically, we capture a NeRF \\( F_{\\theta}:(\\mathbf{x},\\mathbf{r})\\mapsto (\\mathbf{c},\\sigma) \\) that maps a 3D point \\( \\mathbf{x} = (x,y,z) \\) and viewing direction \\( \\mathbf{r} \\) to its corresponding RGB color \\( \\mathbf{c} \\) and density \\( \\sigma \\) [39]. We associate to the visual representation a touch model \\( F_{\\phi}:\\mathbf{v}_{t}\\mapsto \\pmb{\\tau} \\) that generates the tactile signal that one would obtain by touching at the center of the image \\( \\mathbf{v}_t \\). In the following, we explain how to estimate \\( F_{\\theta} \\) and \\( F_{\\phi} \\) and put them into the same shared 3D space." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.735, + 0.38, + 0.751 + ], + "angle": 0, + "content": "3.1. Capturing vision and touch signals" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.761, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Obtaining a visual 3D reconstruction. We build the visual NeRF, \\(F_{\\theta}\\), closely following previous work [12, 55]. A human data collector moves through a scene and records a video, covering as much of the space as possible. We then estimate camera pose using structure from motion [47] and create a NeRF using off-the-shelf packages [53]. Additional details are provided in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Capturing and registering touch. We simultaneously collect tactile and visual signals by mounting a touch sensor" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.889, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.226, + 0.892, + 0.282 + ], + "angle": 0, + "content": "Figure 3. Capturing setup. (a) We record paired vision and touch signals using a camera attached to a touch sensor. (b) We estimate the relative pose between the touch sensor and the camera using correspondences between sight and touch." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.309, + 0.892, + 0.473 + ], + "angle": 0, + "content": "on a camera (Fig. 3), obtaining synchronized touch signals \\(\\{\\tau_i\\}_{i=1}^N\\) and video frames \\(\\mathbf{v}\\). We then estimate the pose of the video frames using off-the-shelf structure from motion methods [47], obtaining poses \\(\\{p_i^v\\}_{i=1}^N\\). Finally, we use the calibration of the mount to obtain the poses \\(\\{p_i^t\\}_{i=1}^N\\) of the tactile measurements with respect to the scene's global reference frame. As a collection device, we mount an iPhone 14 Pro to one end of a camera rod, and a DIGIT [32] touch sensor to the other end. Note that the devices can be replaced with any RGB-D camera and vision-based tactile sensor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Capturing setup calibration. To find the relative pose between the camera and the touch sensor (Fig. 3), we exploit the fact that arbitrary viewpoints can be synthesized from \\( F_{\\theta} \\), and that ubiquitous vision-based touch sensors are based on perspective cameras. In these sensors, an elastomer gel is placed on the lens of a commodity camera, which is illuminated by colored lights. When the gel is pressed into an object, it deforms, and the camera records an image of the deformation; this image is used as the tactile signal. This design allows us to estimate the pose of the tactile sensor through multi-view constraints from visual-tactile correspondences: pixels in visual images and tactile images that are of the same physical point." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.825 + ], + "angle": 0, + "content": "We start the calibration process by synthesizing novel views from \\(F_{\\theta}\\). The views are generated at the camera location \\(\\{p_i^v\\}_{i = 1}^N\\), but rotated \\(90^{\\circ}\\) on the \\(x\\)-axis. This is because the camera is approximately orthogonal to the touch sensor (see Fig. 3). Then, we manually annotate corresponding pixels between the touch measurements and the generated frames (Fig. 3). To simplify and standardize this process, we place a braille board in each scene and probe it with the touch sensor. This will generate a distinctive touch signal that is easy to localize [23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We formulate the problem of estimating the six degrees of freedom relative pose \\((\\mathbf{R},\\mathbf{t})\\) between the touch sensor and the generated frames as a resectioning problem [20]. We use the estimated 3D structure from the NeRF \\(F_{\\theta}\\) to obtain 3D points \\(\\{\\mathbf{x}_i\\}_{i = 1}^M\\) for each of the annotated corre" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26531" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.136 + ], + "angle": 0, + "content": "spondences. Each point has a pixel position \\(\\mathbf{u}_i\\in \\mathcal{R}^2\\) in the touch measurement. We find \\((\\mathbf{R},\\mathbf{t})\\) by minimizing the reprojection error:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.148, + 0.469, + 0.189 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {R}, \\mathbf {t}} \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\| \\pi (\\mathbf {K} [ \\mathbf {R} \\mid \\mathbf {t} ], \\mathbf {X} _ {i}) - \\mathbf {u} _ {i} \\| _ {1}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.469, + 0.351 + ], + "angle": 0, + "content": "where \\(\\pi\\) projects a 3D point using a given projection matrix, \\(\\mathbf{K}\\) are the known intrinsic of the tactile sensor's camera, and the point \\(\\mathbf{X}_i\\) is in the coordinate system of the generated vision frames. We perform the optimization on 6-15 annotated correspondences from the braille board. For robustness, we compute correspondences from multiple frames. We represent the rotation matrix using quaternions and optimize using nonlinear least-squares. Once we have \\((\\mathbf{R},\\mathbf{t})\\) with respect to the generated frames, we can derive the relative pose between the camera and the touch sensor." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.363, + 0.325, + 0.379 + ], + "angle": 0, + "content": "3.2. Imputing the missing touch" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.469, + 0.567 + ], + "angle": 0, + "content": "We use a generative model to estimate the touch signal (represented as an image from a vision-based touch sensor) for other locations within the scene. Specifically, we train a diffusion model \\( p_{\\phi}(\\pmb{\\tau} \\mid \\mathbf{v}, \\mathbf{d}, \\mathbf{b}) \\), where \\( \\mathbf{v} \\) and \\( \\mathbf{d} \\) are images and depth maps extracted from \\( F_{\\theta} \\) (see Fig. 4). We also pass as input to the diffusion model a background image captured by the touch sensor when it is not in contact with anything, denoted as \\( \\mathbf{b} \\). Although not essential, we have observed that this additional input empirically improves the model's performance (e.g., Fig. 1 the background provides the location of defects in the gel, which appear as black dots). We train the model \\( p_{\\phi} \\) on our entire vision-touch dataset (Sec. 4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.469, + 0.673 + ], + "angle": 0, + "content": "The training of \\( p_{\\phi} \\) is divided into two stages. In the first, we pre-train a cross-modal visual-tactile encoder with self-supervised contrastive learning on our dataset. This stage, initially proposed by [23, 57], is equivalent to the self-supervised encoding pre-training that is common for image generation models [45]. We use a ResNet-50 [21] as the backbone for this contrastive model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.469, + 0.839 + ], + "angle": 0, + "content": "In the second stage, we use the contrastive model to generate the input for a conditional latent diffusion model, which is built upon Stable Diffusion [45]. A frozen pretrained VQ-GAN [14] is used to obtain the latent representation with a spatial dimension of \\(64 \\times 64\\). We start training the diffusion model from scratch and pre-train it on the task of unconditional tactile image generation on the YCB-Slide dataset [52]. After this stage, we train the conditional generative model \\(p_{\\phi}\\) on our spatially aligned visual-tactile dataset, further fine-tuning the contrastive model end-to-end with the generation task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.9 + ], + "angle": 0, + "content": "At inference time, given a novel location in the 3D scene, we first render the visual signals \\(\\hat{\\mathbf{v}}\\) and \\(\\hat{\\mathbf{d}}\\) from NeRF, and then estimate the touch signal \\(\\hat{\\tau}\\) of the position using the diffusion model." + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.245, + 0.892, + 0.301 + ], + "angle": 0, + "content": "Figure 4. Touch estimation. We estimate the tactile signal for a given touch sensor pose \\((\\mathbf{R},\\mathbf{t})\\). To do this, we synthesize a viewpoint from the NeRF, along with a depth map. We use conditional latent diffusion to predict the tactile signal from these inputs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.313, + 0.756, + 0.329 + ], + "angle": 0, + "content": "4. A 3D Visual-Tactile Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.369 + ], + "angle": 0, + "content": "In the following, we show the details of the data collection process and statistics of our dataset." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.378, + 0.74, + 0.393 + ], + "angle": 0, + "content": "4.1. Data Collection Procedure" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.581 + ], + "angle": 0, + "content": "The data collection procedure is divided into two stages. First, we collect multiple views from the scene, capturing enough frames around the areas we plan to touch. During this stage, we collect approximately 500 frames. Next, we collect synchronized visual and touch data, maximizing the geometry and texture being touched. We then estimate the camera location of the vision frames collected in the previous two stages using off-the-shelf mapping tools [47]. After estimating the camera poses for the vision frames, the touch measurements' poses can be derived by using the mount calibration matrix. More details about the pose estimation procedure can be found in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Finally, we associate each touch sensor with a color image by translating the sensor poses upwards by 0.4 meters and querying the NeRF with such poses. The field of view we use when querying the NeRF is \\(50^{\\circ}\\). This provides us with approximately 1,500 temporally aligned vision-touch image pairs per scene. Note that this collection procedure is scalable since it does not require specific expertise or equipment and generates abundant scene-level samples." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.668, + 0.726 + ], + "angle": 0, + "content": "4.2. Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.825 + ], + "angle": 0, + "content": "We collect our data in 13 ordinary scenes including two offices, a workroom, a conference room, a corridor, a tabletop, a corridor, a lounge, a room with various clothes and four outdoor scenes with interesting materials. Typically, we collect 1k to 2k tactile probes in each scene, resulting in a total of 19.3k image pairs in the dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Some representative samples from the collected dataset are shown in Fig. 5. Our data includes a large variety of geometry (edges, surfaces, corners, etc.) and texture (plastic, clothes, snow, wood, etc.) of different materials in the scene. During capturing process, the collector will try to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26532" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.349, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.091, + 0.619, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.091, + 0.891, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.893, + 0.328 + ], + "angle": 0, + "content": "Figure 5. Representative examples from the captured dataset. Our dataset is obtained from nine everyday scenes, such as offices, classrooms, and kitchens. We show three such scenes in the figure above, together with samples of spatially aligned visual and tactile data. In each scene, 1k to 2k tactile probes were collected, resulting in a total of 19.3k image pairs. The data encompasses diverse geometries (edges, surfaces, corners, etc.) and textures (plastic, clothes, snow, wood, etc.) of various materials. The collector systematically probed different objects, covering areas with distinct geometry and texture using different sensor poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.47, + 0.446 + ], + "angle": 0, + "content": "thoroughly probe various objects and cover the interesting areas with more distinguishable geometry and texture with different sensor poses. To the best of our knowledge, our dataset is the first dataset that captures full, scene-scale spatially aligned vision-touch image pairs. We provide more details about the dataset in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.458, + 0.21, + 0.475 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.47, + 0.573 + ], + "angle": 0, + "content": "Leveraging the spatially aligned image and touch pairs from our dataset, we first conduct experiments on dense touch estimation. We then show the effectiveness of both the aligned data pairs and the synthesized touch signals by conducting tactile localization and material classification as two downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.584, + 0.295, + 0.6 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.469, + 0.716 + ], + "angle": 0, + "content": "NeRF. We use the Nerfacto method from Nerfstudio [53]. For each scene, we utilize approximately 2,000 images as training set, which thoroughly cover the scene from various view points. We train the network with a base learning rate of \\(1 \\times 10^{-2}\\) using Adam [30] optimizer for 200,000 steps on a single NVIDIA RTX 2080 Ti GPU to achieve optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Visual-tactile contrastive model. Following prior works [27, 57], we leverage contrastive learning methods to train a ResNet-50 [21] as visual encoder. The visual and tactile encoders share the same architecture but have different weights. We encode visual and tactile data into latent vectors in the resulting shared representation space. We set the dimension of the latent vectors to 32. Similar to CLIP [43], the model is trained on InfoNCE loss obtained from the pairwise dot products of the latent vectors. We train the model for 20 epochs by Adam [30] optimizer with a learning rate of \\(10^{-4}\\) and batch size of 256 on 4 NVIDIA RTX 2080 Ti GPUs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.892, + 0.626 + ], + "angle": 0, + "content": "Visual-tactile generative model. Our implementation of the diffusion model closely follows Stable Diffusion [46], with the difference that we use a ResNet-50 to generate the visual encoding from RGB-D images for conditioning. Specifically, we also add the RGB-D images rendered from the tactile sensors' poses into the conditioning, which we refer to in Sec. 5.2 as multiscale conditioning. The model is optimized for 30 epochs by Adam [30] optimizer with a base learning rate of \\(10^{-5}\\). The learning rate is scaled bygpu number \\(\\times\\) batch size. We train the model with batch size of 48 on 4 NVIDIA A40 GPUs. At inference time, the model conducts 200 steps of denoising process with a 7.5 guidance scale. Following prior cross-modal synthesis work [44], we use reranking to improve the prediction quality. We obtain 16 samples from the diffusion model for every instance and re-rank the samples with our pretrained contrastive model. The sample with highest similarity is the final prediction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.722, + 0.65 + ], + "angle": 0, + "content": "5.2. Dense Touch Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Experimental setup. We now evaluate the diffusion model's ability to generate touch images. To reduce overlap between the training and test set, we first split the frames into sequences temporally (following previous work [56]). We split them into sequences of 50 touch samples, then divide these sequences into train/validation/test with a ratio of \\(8/1/1\\). We evaluate the generated samples on Frechet Inception Distance (FID), a standard evaluation metric for cross-modal generation [56]. We also include Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM), though we note that these metrics are highly sensitive to spatial position of the generated content, and can be optimized by models that minimize simple pixelwise losses [22]. We also include CVTP metric proposed by prior work [57], which measures the similarity between visual and tactile embeddings of a contrastive model, analogous to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26533" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.087, + 0.482, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.087, + 0.895, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.417, + 0.893, + 0.503 + ], + "angle": 0, + "content": "Figure 6. Qualitative touch estimation results. Each model is conditioned on the RGB image and depth map rendered from the NeRF (left). The white box indicates the tactile sensor's approximate field of view (which is much smaller than the full conditional image). The G.T. column shows the ground truth touch images measured from a DIGIT sensor. L1 and VisGel often generate blurry textures and inaccurate geometry. By contrast, our model better captures the features of the tactile image, e.g., the rock's microgeometry and complex textures and shapes of furniture. The last row shows two failure cases of our model. In both examples, our model generates a touch image that is geometrically misaligned with the ground truth. All of the examples shown here are at least \\(10\\mathrm{cm}\\) away from any training sample." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.527, + 0.47, + 0.603 + ], + "angle": 0, + "content": "CLIP [43] score. We compare against two baselines: VisGel, the approach from Li et. [35], which trains a GAN for touch generation, and \\( L1 \\), a model with the same architecture of VisGel but trained to minimize an L1 loss in pixel space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.82 + ], + "angle": 0, + "content": "Results. As is shown in Table 2, our approach performs much better on the high-level metrics, with up to \\(4\\mathrm{x}\\) lower FID and \\(80\\mathrm{x}\\) higher CVTP. This indicates that our proposed diffusion model captures the distribution and characteristics of the real tactile data more effectively. On the low-level metrics (PSNR and SSIM), all methods are comparable. In particular, the L1 model slightly outperforms the other methods since the loss it is trained on is highly correlated with low-level, pixel-wise metrics. Fig. 6 qualitatively compares samples from the different models. Indeed, our generated samples exhibit enhanced details in micro-geometry of fabrics and richer textures, including snow, wood and carpeting. However, all methods fail on fine details that are barely visible in the image, such as the tree bark." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Ablation study. We evaluate the importance of the main components of our proposed touch generation approach (Table 3). Removing the conditioning on the RGB image results in the most prominent performance drop. This is expected since RGB image uniquely determines the fine" + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.524, + 0.892, + 0.598 + ], + "angle": 0, + "content": "
ModelPSNR ↑SSIM ↑FID ↓CVTP ↑
L124.340.8297.050.01
VisGel [35]23.660.81130.220.03
Ours22.840.7228.970.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.601, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Table 2. Quantitative results on touch estimation for novel views. While comparable on low-level metrics with the baselines, our approach captures the characteristics of the real tactile data more effectively, resulting in a lower FID score." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.668, + 0.892, + 0.76 + ], + "angle": 0, + "content": "grained details of a tactile image. Removing depth image or contrastive pretraining has small effect on CVTP but results in a drop on FID. Contrastive re-ranking largely improves CVTP, indicating the necessity of obtaining multiple samples from the diffusion model. We also find that multiscale conditioning provide a small benefit on FID and CVTP." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.771, + 0.848, + 0.787 + ], + "angle": 0, + "content": "5.3. Downstream Task I: Tactile Localization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.893, + 0.903 + ], + "angle": 0, + "content": "To help understand the quality of the captured TaRFs, we evaluate the performance of the contrastive model (used for conditioning our diffusion model) on the task of tactile localization. Given a tactile signal, our goal is to find the corresponding regions in a 2D image or in a 3D scene that are associated with it, i.e., we ask the question: what part of this image/scene feel like this? We perform the following" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26534" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.895, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.346, + 0.895, + 0.43 + ], + "angle": 0, + "content": "Figure 7. Tactile localization heatmaps. Given a tactile query image, the heatmap shows the image patches with a higher affinity to this tactile signal, as measured by a contrastive model trained on our dataset. We use a sliding window and compare each extracted patch with the touch signal. In each case, the center patch is the true position. Our model successfully captures the correlation between the two signals. This enables it to localize a variety of touch signals, including fine-grained geometry, e.g., a cable or a keyboard, various types of corners and edges, and large uniform regions, such as a clothing. This ability enables our diffusion model to effectively propagate sparse touch samples to other visually and structurally similar regions of the scene." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.442, + 0.47, + 0.556 + ], + "angle": 0, + "content": "
Model variationPSNR ↑SSIM ↑FID ↓CVTP ↑
Full22.840.7228.970.80
No RGB conditioning22.130.7034.310.76
No depth conditioning22.570.7133.160.80
No contrastive pretraining22.820.7132.980.79
No re-ranking22.920.7229.460.61
No multiscale23.190.7230.890.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.637 + ], + "angle": 0, + "content": "Table 3. Ablation study. Since the fine-grained details of touch images can be determined from a RGB image, removing conditioning on the latter results in the largest performance drops. Re-ranking has notable impact on CVTP, indicating the necessity of obtaining multiple samples from the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.676 + ], + "angle": 0, + "content": "evaluations on the test set of our dataset. Note that we run no task-specific training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.903 + ], + "angle": 0, + "content": "2D Localization. To determine which part of an image are associated with a given tactile measurement, we follow the same setup of SSVTP [28]. We first split the image into patches and compute their embedding. Then, we generate the tactile embedding of the input touch image. Finally, we compute the pairwise similarities between the tactile and visual embeddings, which we plot as a heatmap. As we can see in Fig. 7, our constrastive encoder can successfully capture the correlations between the visual and tactile data. For instance, the tactile embeddings of edges are associated to edges of similar shape in the visual image. Note that the majority of tactile embeddings are highly ambiguous: all edges with a similar geometry feel the same." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.595 + ], + "angle": 0, + "content": "3D Localization. In 3D, the association of an image to tactile measurements becomes less ambiguous. Indeed, since tactile-visual samples are rotation-dependent, objects with similar shapes but different orientations will generate different tactile measurements. Lifting the task to 3D still does not remove all ambiguities (for example, each side of a rectangular table cannot be precisely localized). Nonetheless, we believe it to be a good fit for a quantitative evaluation since it's rare for two ambiguous parts of the scene to be touched with exactly the same orientation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.748 + ], + "angle": 0, + "content": "We use the following experimental setup for 3D localization. Given a tactile image as a query, we compute its distance in embedding space to all visual test images from the same scene. Note that all test images are associated with a 3D location. We define as ground-truth correspondences all test images at a distance of at most \\( r \\) from the 3D location of the test sample. We vary \\( r \\) to account for local ambiguities. As typical in the retrieval literature, we benchmark the performance with metric mean Average Precision (mAP)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We consider three baselines: (1) chance, which randomly selects corresponding samples; (2) real, which uses the contrastive model trained on our dataset; and (3) real + estimated, which trains the contrastive model on both dataset samples and a set of synthetic samples generated via the scenes' NeRF and our touch generation model. Specifically, we render a new image and corresponding touch by interpolating the position of two consecutive frames in the training dataset. This results in a training dataset for the contrastive model that is twice as large." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26535" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.472, + 0.182 + ], + "angle": 0, + "content": "
Datasetr(m)
0.0010.0050.010.050.1
Chance3.556.8210.2518.2621.33
Real12.1022.9332.1050.3057.15
Real + Est.14.9226.6936.1753.6260.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.188, + 0.47, + 0.26 + ], + "angle": 0, + "content": "Table 4. Quantitative results on 3D tactile localization. We evaluate using mean Average Precision (mAP) as a metric. Training the contrastive model on our dataset of visually aligned real samples together with estimated samples from new locations in the scene results in the highest performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.47, + 0.381 + ], + "angle": 0, + "content": "The results, presented in Table 4, demonstrate the performance benefit of employing both real and synthetic tactile pairs. Combining synthetic tactile images with the original pairs achieves highest performance on all distance thresholds. Overall, this indicates that touch measurements from novel views are not only qualitatively accurate, but also beneficial for this downstream task." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.393, + 0.455, + 0.409 + ], + "angle": 0, + "content": "5.4. Downstream Task II: Material Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.553 + ], + "angle": 0, + "content": "We investigate the efficacy of our visual-tactile dataset for understanding material properties, focusing on the task of material classification. We follow the formulation by Yang et al. [56], which consists of three subtasks: (i) material classification, requiring the distinction of materials among 20 possible classes; (ii) softness classification, a binary problem dividing materials as either hard or soft; and (iii) hardness classification, which requires the classification of materials as either rough or smooth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.749 + ], + "angle": 0, + "content": "We follow the same experimental procedure of [56]: we pretrain a contrastive model on a dataset and perform linear probing on the sub-tasks' training set. Our experiments only vary the pretraining dataset, leaving all architectural choices and hyperparameters the same. We compare against four baselines. A random classifier (chance); the ObjectFolder 2.0 dataset [17]; the VisGel dataset [35]; and the Touch and Go dataset [56]. Note that the touch sensor used in the test data (GelSight) differs from the one used in our dataset (DIGIT). Therefore, we use for pretraining a combination of our dataset and Touch and Go. To ensure a fair comparison, we also compare to the combination of each dataset and Touch and Go." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The findings from this evaluation, as shown in Table 5, suggest that our data improves the effectiveness of the contrastive pretraining objective, even though our data is from a different distribution. Moreover, we find that adding estimated touch probes for pretraining results in a higher performance on all the three tasks, especially the smoothness classification. This indicates that not only does our dataset covers a wide range of materials but also our diffusion model captures the distinguishable and useful patterns of different materials." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.895, + 0.251 + ], + "angle": 0, + "content": "
DatasetMaterialHard/ SoftRough/ Smooth
Chance18.666.156.3
ObjectFolder 2.0 [17]36.272.069.0
VisGel [35]39.169.470.4
Touch and Go [56]54.777.379.4
+ ObjectFolder 2.0 [17]54.687.384.8
+ VisGel [35]53.186.783.6
+ Ours* (Real)57.688.481.7
+ Ours* (Real + Estimated)59.088.786.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.256, + 0.894, + 0.355 + ], + "angle": 0, + "content": "Table 5. Material classification. We show the downstream material recognition accuracy of models pre-trained on different datasets. The final rows show the performance when combining different datasets with Touch and Go [56]. * The task-specific training and testing datasets for this task are collected with a Gel-Sight sensor. We note that our data comes from a different distribution, since it is collected with a DIGIT sensor [32]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.386, + 0.62, + 0.401 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.895, + 0.632 + ], + "angle": 0, + "content": "In this work, we present the TaRF, a scene representation that brings vision and touch into a shared 3D space. This representation enables the generation of touch probes for novel scene locations. To build this representation, we collect the largest dataset of spatially aligned vision and touch probes. We study the utility of both the representation and the dataset in a series of qualitative and quantitative experiments and on two downstream tasks: 3D touch localization and material recognition. Overall, our work makes the first step towards giving current scene representation techniques an understanding of not only how things look, but also how they feel. This capability could be critical in several applications ranging from robotics to the creation of virtual worlds that look and feel like the real world." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.894, + 0.782 + ], + "angle": 0, + "content": "Limitations. Since the touch sensor is based on a highly zoomed-in camera, small (centimeter-scale) errors in SfM or visual-tactile registration can lead to misalignments of several pixels between the views of the NeRF and the touch samples, which can be seen in our TaRFs. Another limitation of the proposed representation is the assumption that the scene's coarse-scale structure does not change when it is touched, an assumption that may be violated for some inelastic surfaces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Acknowledgements. We thank Jeongsoo Park, Ayush Shrivastava, Daniel Geng, Ziyang Chen, Zihao Wei, Zixuan Pan, Chao Feng, Chris Rockwell, Gaurav Kaul and the reviewers for the valuable discussion and feedback. This work was supported by an NSF CAREER Award #2339071, a Sony Research Award, the DARPA Machine Common Sense program, and ONR MURI award N00014-21-1-2801." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26536" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Peter Anderson, Ayush Shrivastava, Joanne Truong, Arjun Majumdar, Devi Parikh, Dhruv Batra, and Stefan Lee. Sim-to-real transfer for vision-and-language navigation. In Conference on Robot Learning, pages 671-681. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.272, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[3] Jakub Bednarek, Michal Bednarek, Lorenz Wellhausen, Marco Hutter, and Krzysztof Walas. What am i touching? learning to classify terrain via haptic sensing. In 2019 International Conference on Robotics and Automation (ICRA), pages 7187-7193. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.471, + 0.4 + ], + "angle": 0, + "content": "[4] Katherine L Bouman, Bei Xiao, Peter Battaglia, and William T Freeman. Estimating the material properties of fabric from video. In Proceedings of the IEEE international conference on computer vision, pages 1984-1991, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.47, + 0.429 + ], + "angle": 0, + "content": "[5] Alexander Burka. Instrumentation, data, and algorithms for visually understanding haptic surface properties. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.432, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[6] Roberto Calandra, Andrew Owens, Manu Upadhyaya, Wenzhen Yuan, Justin Lin, Edward H. Adelson, and Sergey Levine. The feeling of success: Does touch sensing help predict grasp outcomes? Conference on Robot Learning (CoRL), 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.503, + 0.47, + 0.571 + ], + "angle": 0, + "content": "[7] Roberto Calandra, Andrew Owens, Dinesh Jayaraman, Justin Lin, Wenzhen Yuan, Jitendra Malik, Edward H. Adelson, and Sergey Levine. More than a feeling: Learning to grasp and regrasp using vision and touch. Robotics and Automation Letters (RA-L), 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.47, + 0.643 + ], + "angle": 0, + "content": "[8] Roberto Calandra, Andrew Owens, Dinesh Jayaraman, Justin Lin, Wenzhen Yuan, Jitendra Malik, Edward H. Adelson, and Sergey Levine. More than a feeling: Learning to grasp and regrasp using vision and touch. IEEE Robotics and Automation Letters, 3:3300-3307, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.646, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[9] Berk Calli, Arjun Singh, Aaron Walsman, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. The ycb object and model set: Towards common benchmarks for manipulation research. In 2015 international conference on advanced robotics (ICAR), pages 510-517. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[10] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European conference on computer vision, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[11] Alex Church, John Lloyd, Raia Hadsell, and Nathan F Lepora. Deep reinforcement learning for tactile robotics: Learning to type on a braille keyboard. IEEE Robotics and Automation Letters, 5(4):6145-6152, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[13] Abe Davis, Justin G Chen, and Frédo Durand. Image-space modal bases for plausible manipulation of objects in video. ACM Transactions on Graphics (TOG), 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[14] Mucong Ding, Kezhi Kong, Jingling Li, Chen Zhu, John Dickerson, Furong Huang, and Tom Goldstein. Vq-gnn: A universal framework to scale up graph neural networks using vector quantization. Advances in Neural Information Processing Systems, 34:6733-6746, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[15] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence, 32(8):1362-1376, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[16] Ruohan Gao, Yen-Yu Chang, Shivani Mall, Li Fei-Fei, and Jiajun Wu. Objectfolder: A dataset of objects with implicit visual, auditory, and tactile representations. In CoRL, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[17] Ruohan Gao, Zilin Si, Yen-Yu Chang, Samuel Clarke, Jeannette Bohg, Li Fei-Fei, Wenzhen Yuan, and Jiajun Wu. Objectfolder 2.0: A multisensory object dataset for sim2real transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10598–10608, 2022. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[18] Ruohan Gao, Yiming Dou, Hao Li, Tanmay Agarwal, Jeannette Bohg, Yunzhu Li, Li Fei-Fei, and Jiajun Wu. The objectfolder benchmark: Multisensory learning with neural and real objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17276–17286, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[19] Michael SA Graziano and Charles G Gross. The representation of extrapersonal space: A possible role for bimodal, visual-tactile neurons. 1995. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.624, + 0.892, + 0.69 + ], + "angle": 0, + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[23] Carolina Higuera, Byron Boots, and Mustafa Mukadam. Learning to read braille: Bridging the tactile reality gap with diffusion models. arXiv preprint arXiv:2304.01182, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[24] Mark A Hoepflinger, C David Remy, Marco Hutter, Luciano Spinello, and Roland Siegwart. Haptic terrain classification for legged robots. In 2010 IEEE International Conference on Robotics and Automation, pages 2828-2833. IEEE, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.805, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[25] Micah K Johnson and Edward H Adelson. Retrographic sensing for the measurement of surface texture and shape. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 1070-1077. IEEE, 2009. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.861, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Micah K Johnson, Forrester Cole, Alvin Raj, and Edward H Adelson. Microgeometry capture using an elastomeric sensor. ACM Transactions on Graphics (TOG), 2011. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26537" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[27] Justin Kerr, Huang Huang, Albert Wilcox, Ryan Hoque, Jeffrey Ichnowski, Roberto Calandra, and Ken Goldberg. Learning self-supervised representations from vision and touch for active sliding perception of deformable surfaces. arXiv preprint arXiv:2209.13042, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[28] Justin Kerr, Huang Huang, Albert Wilcox, Ryan Hoque, Jeffrey Ichnowski, Roberto Calandra, and Ken Goldberg. Self-supervised visuo-tactile pretraining to locate and follow garment features. In Robotics: Science and Systems, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[29] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.471, + 0.32 + ], + "angle": 0, + "content": "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. *ICLR*, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.471, + 0.376 + ], + "angle": 0, + "content": "[31] Hendrik Kolvenbach, Christian Bärtschi, Lorenz Wellhausen, Ruben Grandia, and Marco Hutter. Haptic inspection of planetary soils with legged robots. IEEE Robotics and Automation Letters, 4(2):1626-1632, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.47, + 0.46 + ], + "angle": 0, + "content": "[32] Mike Lambeta, Po-Wei Chou, Stephen Tian, Brian Yang, Benjamin Maloon, Victoria Rose Most, Dave Stroud, Raymond Santos, Ahmad Byagowi, Gregg Kammerer, et al. Digit: A novel design for a low-cost compact high-resolution tactile sensor with application to in-hand manipulation. IEEE Robotics and Automation Letters, 2020. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.47, + 0.517 + ], + "angle": 0, + "content": "[33] Maxime Lhuillier and Long Quan. A quasi-dense approach to surface reconstruction from uncalibrated images. IEEE transactions on pattern analysis and machine intelligence, 27(3):418-433, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.47, + 0.575 + ], + "angle": 0, + "content": "[34] Hongyu Li, Snehal Dikhale, Soshi Iba, and Nawid Jamali. Vihope: Visuotactile in-hand object 6d pose estimation with shape completion. IEEE Robotics and Automation Letters, 8 (11):6963-6970, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[35] Yunzhu Li, Jun-Yan Zhu, Russ Tedrake, and Antonio Torralba. Connecting touch and vision via cross-modal prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10609-10618, 2019. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[36] Justin Lin, Roberto Calandra, and Sergey Levine. Learning to identify object instances by touch: Tactile recognition via multimodal matching. In 2019 International Conference on Robotics and Automation (ICRA), pages 3644-3650. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[37] Antonio Loquercio, Ashish Kumar, and Jitendra Malik. Learning visual locomotion with cross-modal supervision. In 2023 IEEE International Conference on Robotics and Automation (ICRA), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[38] Gabriel B Margolis, Xiang Fu, Yandong Ji, and Pulkit Agrawal. Learning physically grounded robot vision with active sensing motor policies. In 7th Annual Conference on Robot Learning, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[40] Andrew Owens, Phillip Isola, Josh McDermott, Antonio Torralba, Edward H Adelson, and William T Freeman. Visually indicated sounds. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2405-2413, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.218 + ], + "angle": 0, + "content": "[41] Senthil Purushwalkam, Abhinav Gupta, Danny M Kaufman, and Bryan Russell. Bounce and learn: Modeling scene dynamics with real-world bounces. arXiv preprint arXiv:1904.06827, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.893, + 0.275 + ], + "angle": 0, + "content": "[42] Haozhi Qi, Brent Yi, Sudharshan Suresh, Mike Lambeta, Yi Ma, Roberto Calandra, and Jitendra Malik. General in-hand object rotation with vision and touch. arXiv preprint arXiv:2309.09979, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.893, + 0.347 + ], + "angle": 0, + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.893, + 0.415 + ], + "angle": 0, + "content": "[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.893, + 0.46 + ], + "angle": 0, + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.893, + 0.532 + ], + "angle": 0, + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.893, + 0.574 + ], + "angle": 0, + "content": "[47] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.893, + 0.616 + ], + "angle": 0, + "content": "[48] Carmelo Sferrazza and Raffaello D'Andrea. Design, motivation and evaluation of a full-resolution optical tactile sensor. Sensors, 19(4):928, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.893, + 0.659 + ], + "angle": 0, + "content": "[49] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV 2012, ECCV 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.893, + 0.702 + ], + "angle": 0, + "content": "[50] Linda Smith and Michael Gasser. The development of embodied cognition: Six lessons from babies. Artificial life, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.893, + 0.771 + ], + "angle": 0, + "content": "[51] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.893, + 0.841 + ], + "angle": 0, + "content": "[52] Sudharshan Suresh, Zilin Si, Stuart Anderson, Michael Kaess, and Mustafa Mukadam. Midastouch: Monte-carlo inference over distributions across sliding touch. In Conference on Robot Learning, pages 319-331. PMLR, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[53] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David Mcallister, Justin Kerr, and Angjoo Kanazawa. Nerfstudio: A modu" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26538" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "lar framework for neural radiance field development. arXiv preprint arXiv:2302.04264, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[54] Shaoxiong Wang, Jiajun Wu, Xingyuan Sun, Wenzhen Yuan, William T Freeman, Joshua B Tenenbaum, and Edward H Adelson. 3d shape perception from monocular vision, touch, and shape priors. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[55] Jianxiong Xiao, Andrew Owens, and Antonio Torralba. Sun3d: A database of big spaces reconstructed using sfm and object labels. 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[56] Fengyu Yang, Chenyang Ma, Jiacheng Zhang, Jing Zhu, Wenzhen Yuan, and Andrew Owens. Touch and go: Learning from human-collected vision and touch. Neural Information Processing Systems (NeurIPS) - Datasets and Benchmarks Track, 2022. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[57] Fengyu Yang, Jiacheng Zhang, and Andrew Owens. Generating visual scenes from touch. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22070-22080, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[58] Fengyu Yang, Chao Feng, Ziyang Chen, Hyoungseob Park, Daniel Wang, Yiming Dou, Ziyao Zeng, Xien Chen, Rit Gangopadhyay, Andrew Owens, and Alex Wong. Binding touch to everything: Learning unified multimodal tactile representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[59] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[60] Zhao-Heng Yin, Binghao Huang, Yuzhe Qin, Qifeng Chen, and Xiaolong Wang. Rotating without seeing: Towards in-hand dexterity through touch. arXiv preprint arXiv:2303.10880, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.56, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[61] Wenzhen Yuan, Shaoxiong Wang, Siyuan Dong, and Edward Adelson. Connecting look and feel: Associating the visual and tactile properties of physical materials. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5580-5588, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[62] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.469, + 0.769 + ], + "angle": 0, + "content": "[63] Shaohong Zhong, Alessandro Albini, Oiwi Parker Jones, Perla Maiolino, and Ingmar Posner. Touching a nerf: Leveraging neural radiance fields for tactile sensory data generation. In Conference on Robot Learning, pages 1618-1628. PMLR, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.769 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26539" + } + ] +] \ No newline at end of file diff --git a/2024/Tactile-Augmented Radiance Fields/f9e9c348-b82c-415a-b07e-5b945a68c375_origin.pdf b/2024/Tactile-Augmented Radiance Fields/f9e9c348-b82c-415a-b07e-5b945a68c375_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d05780a9ebc353d492438b226be8e778c1c5f5f9 --- /dev/null +++ b/2024/Tactile-Augmented Radiance Fields/f9e9c348-b82c-415a-b07e-5b945a68c375_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0724060153826acec08e94d44b982d36ad9d5f9fd5720c3c5e48ec43507dceb8 +size 2885352 diff --git a/2024/Tactile-Augmented Radiance Fields/full.md b/2024/Tactile-Augmented Radiance Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..13e88dfa31429e953cb0ce7309fb85a07b39477d --- /dev/null +++ b/2024/Tactile-Augmented Radiance Fields/full.md @@ -0,0 +1,288 @@ +# Tactile-Augmented Radiance Fields + +Yiming Dou $^{1}$ Fengyu Yang $^{2}$ Yi Liu $^{1}$ Antonio Loquercio $^{3}$ Andrew Owens $^{1}$ + +$^{1}$ University of Michigan $^{2}$ Yale University $^{3}$ UC Berkeley + +![](images/0f549b524105136b5fc7c79c0afb9d674e867f0888bda77c6bc2a1700dc4bb4d.jpg) +Figure 1. Tactile-augmented radiance fields. We capture a tactile-augmented radiance field (TaRF) from photos and sparsely sampled touch probes. To do this, we register the captured visual and tactile signals into a shared 3D space, then train a diffusion model to impute touch at other locations within the scene. Here, we visualize two touch probes and their (color coded) 3D positions in the scene. We also show two touch signals estimated by the diffusion model. The touch signals were collected using a vision-based touch sensor [32] that represents the touch signals as images. Please see our project page for video results. + +![](images/163f41aec7e0b4020a7810e4484f53cc9989565a88fdaa2782da932bf26321d3.jpg) + +# Abstract + +We present a scene representation, which we call a tactile-augmented radiance field (TaRF), that brings vision and touch into a shared 3D space. This representation can be used to estimate the visual and tactile signals for a given 3D position within a scene. We capture a scene's TaRF from a collection of photos and sparsely sampled touch probes. Our approach makes use of two insights: (i) common vision-based touch sensors are built on ordinary cameras and thus can be registered to images using methods from multi-view geometry, and (ii) visually and structurally similar regions of a scene share the same tactile features. We use these insights to register touch signals to a captured visual scene, and to train a conditional diffusion model that, provided with an RGB-D image rendered from a neural radiance field, generates its corresponding tactile signal. To evaluate our approach, we collect a dataset of TaRFs. This dataset contains more touch samples than previous real-world datasets, and it provides spatially aligned visual signals for each captured touch signal. We demonstrate the accuracy of our cross-modal generative model and the utility of the captured visual-tactile data on several downstream tasks. Project page: https://dou-yiming.github.io/TaRF. + +# 1. Introduction + +As humans, our ability to perceive the world relies crucially on cross-modal associations between sight and touch [19, 50]. Tactile sensing provides a detailed understanding of material properties and microgeometry, such as the intricate patterns of bumps on rough surfaces and the complex motions that soft objects make when they deform. This type of understanding, which largely eludes today's computer vision models, is a critical component of applications that require reasoning about physical contact, such as robotic locomotion [3, 24, 31, 34, 37, 38] and manipulation [6, 7, 11, 42, 60], and methods that simulate the behavior of materials [4, 13, 40, 41]. + +In comparison to many other modalities, collecting tactile data is an expensive and tedious process, since it requires direct physical interaction with the environment. A recent line of work has addressed this problem by having humans or robots probe the environment with touch sensors (see Table 1). Early efforts have been focused on capturing the properties of only a few objects either in simulation [16, 17, 52] or in lab-controlled settings [6, 7, 18, 28, 35, 52, 63], which may not fully convey the diversity of tactile signals in natural environments. Other works have gone beyond a + +
DatasetSamplesAlignedScenarioSource
More Than a Feeling [7]6.5k×TabletopRobot
Feeling of Success [6]9.3k×TabletopRobot
VisGel [35]12k×TabletopRobot
SSVTP [28]4.6k✓TabletopRobot
ObjectFolder 1.0 [16]-✓ObjectSynthetic
ObjectFolder 2.0 [17]-✓ObjectSynthetic
ObjectFolder Real [18]3.7k×ObjectRobot
Burka et al. [5]1.1k×Sub-sceneHuman
Touch and Go [56]13.9k×Sub-sceneHuman
YCB-Slide* [52]-✓ObjectHuman
Touching a NeRF [63]1.2k✓ObjectRobot
TaRF (Ours)19.3k✓Full sceneHuman
+ +Table 1. Dataset comparison. We present the number of real visual-tactile pairs and whether such pairs are visually aligned, i.e., whether the visual image includes an occlusion-free view of the touched surface. *YCB-Slide has real-world touch probes but synthetic images rendered with CAD models of YCB objects on a white background [9]. + +lab setting and have collected touch from real scenes [5, 56]. However, existing datasets lack aligned visual and tactile information, since the touch sensor and the person (or robot) that holds it often occlude large portions of the visual scene (Fig. 2). These datasets also contain only a sparse set of touch signals for each scene, and it is not clear how the sampled touch signals relate to each other in 3D. + +In this work, we present a simple and low-cost procedure to capture quasi-dense, scene-level, and spatially-aligned visual and touch data (Fig. 1). We call the resulting scene representation a tactile-augmented radiance field (TaRF). We remove the need for robotic collection by leveraging a 3D scene representation (a NeRF [39]) to synthesize a view of the surface being touched, which results in spatially aligned visual-tactile data (Fig. 2). We collect this data by mounting a touch sensor to a camera with commonly available materials (Fig. 3). To calibrate the pair of sensors, we take advantage of the fact that popular vision-based touch sensors [25, 26, 32, 48] are built on ordinary cameras. The relative pose between the vision and tactile sensors can thus be estimated using traditional methods from multi-view geometry, such as camera resectioning [20]. + +We use this procedure to collect a large real-world dataset of aligned visual-tactile data. With this dataset, we train a diffusion model [45, 51] to estimate touch at locations not directly probed by a sensor. In contrast to the recent work of Zhong et al. [63], which also estimates touch from 3D NeRF geometry, we create scene-scale reconstructions, we do not require robotic proprioception, and we use diffusion models [51]. This enables us to obtain tactile data at a much larger scale, and with considerably more diversity. Unlike previous visual-tactile diffusion work [57], we condition the model on spatially aligned visual and depth information, enhancing the generated samples' quality and their usefulness in downstream applications. After training, the diffusion model can be used to predict tactile informa + +![](images/48963dd64dd1561761f3cb4229bcb719dd4ca3e677470e89430566795a3499d9.jpg) +OF 2.0 [17] + +![](images/3df34be0c7d888260703250ad091c94e3386f9e804c9570839d68d4a1c58539d.jpg) +VisGel [35] + +![](images/ca5bdb3189ea7e7413fe4ab6b4f7dd5ae94e7eb6e988f566e701d29386cb889e.jpg) +OF Real [18] + +![](images/66408a8ce54ef733c07222aed6a558248563d7ddf5d651fa8124f3052351a023.jpg) +SSVTP [28] +Figure 2. Visual-tactile examples. In contrast to the visual-tactile data captured in previous work, our approach allows us to sample unobstructed images that are spatially aligned with the touch signal, from arbitrary 3D viewpoints using a NeRF. + +![](images/48eeffb2c37b18e743fe79503e6b089918e7a58f0fd656316a65967904212e1f.jpg) +TG [56] + +![](images/fc56a6478352a81e2bb1314e0cb0dcb53f113f7f611781af4eef5663883d9afd.jpg) +TaRF (Ours) + +tion for novel positions in the scene. Analogous to quasi-dense stereo methods [15, 33], the diffusion model effectively propagates sparse touch samples, obtained by probing, to other visually and structurally similar regions of the scene. + +We evaluate our visual-tactile model's ability to accurately perform cross-modal translation using a variety of quality metrics. We also apply it to several downstream tasks, including localizing a touch within a scene and understanding material properties of the touched area. Our experiments suggest: + +- Touch signals can be localized in 3D space by exploiting multi-view geometry constraints between sight and touch. +- Estimated touch measurements from novel views are not only qualitatively accurate, but also beneficial on downstream tasks. +- Cross-modal prediction models can accurately estimate touch from sight for natural scenes. +- Visually-acquired 3D scene geometry improves cross-modal prediction. + +# 2. Related Work + +Visual-tactile datasets. Previous work has either used simulators [16, 17] or robotic arms [6, 8, 18, 35, 63] for data generation. Our work is closely related to that of Zhong et al. [63], which uses a NeRF and captured touch data to generate a tactile field for several small objects. They use the proprioception of an expensive robot to spatially align vision and touch. In contrast, we leverage the properties of the tactile sensor and novel view synthesis to use commonly available material (a smartphone and a selfie stick) to align vision and touch. This enables the collection of a larger, scene-level, and more diverse dataset, on which we train a higher-capacity diffusion model (rather than a conditional GAN). Like several previous works [5, 56], we also collect scene-level data. In contrast to them, we spatially align the signals by registering them in a unified 3D representation, thereby increasing the prediction power of the visual-tactile generative model. + +Capturing multimodal 3D scenes. Our work is related to methods that capture 3D visual reconstructions of spaces + +using RGB-D data [12, 49, 55, 59] and multimodal datasets of paired 3D vision and language [1, 2, 10]. Our work is also related to recent methods that localize objects in NeRFs using joint embeddings between images and language [29] or by semantic segmentation [62]. In contrast to language supervision, touch is tied to a precise position in a scene. + +3D touch sensing. A variety of works have studied the close relationship between geometry and touch, motivating our use of geometry in imputing touch. Johnson et al. [25, 26] proposed vision-based touch sensing, and showed that highly accurate depth can be estimated from the touch sensor using photometric stereo. Other work has estimated object-scale 3D from touch [54]. By contrast, we combine sparse estimates of touch with quasi-dense tactile signals estimated using generative models. + +Cross-modal prediction of touch from sight. Recent work has trained generative models that predict touch from images. Li et al. [35] used a GAN to predict touch for images of a robotic arm, while Gao et al. [18] applied them to objects collected on a turntable. Yang et al. [57] used latent diffusion to predict touch from videos of humans touching objects. Our goal is different from these works: we want to predict touch signals that are spatially aligned with a visual signal, to exploit scene-specific information, and to use geometry. Thus, we use a different architecture and conditioning signal, and fit our model to examples from the same scenes at training and test time. Other work has learned joint embeddings between vision and touch [28, 36, 56, 58, 61]. + +# 3. Method + +We collect visual and tactile examples from a scene and register them together with a 3D visual reconstruction to build a TaRF. Specifically, we capture a NeRF $F_{\theta}:(\mathbf{x},\mathbf{r})\mapsto (\mathbf{c},\sigma)$ that maps a 3D point $\mathbf{x} = (x,y,z)$ and viewing direction $\mathbf{r}$ to its corresponding RGB color $\mathbf{c}$ and density $\sigma$ [39]. We associate to the visual representation a touch model $F_{\phi}:\mathbf{v}_{t}\mapsto \pmb{\tau}$ that generates the tactile signal that one would obtain by touching at the center of the image $\mathbf{v}_t$ . In the following, we explain how to estimate $F_{\theta}$ and $F_{\phi}$ and put them into the same shared 3D space. + +# 3.1. Capturing vision and touch signals + +Obtaining a visual 3D reconstruction. We build the visual NeRF, $F_{\theta}$ , closely following previous work [12, 55]. A human data collector moves through a scene and records a video, covering as much of the space as possible. We then estimate camera pose using structure from motion [47] and create a NeRF using off-the-shelf packages [53]. Additional details are provided in the supplement. + +Capturing and registering touch. We simultaneously collect tactile and visual signals by mounting a touch sensor + +![](images/04fa5dd0c1d1252de232aed3efddd62a6bfb07a6e9a07d34030942e3d1e09d7a.jpg) +Figure 3. Capturing setup. (a) We record paired vision and touch signals using a camera attached to a touch sensor. (b) We estimate the relative pose between the touch sensor and the camera using correspondences between sight and touch. + +on a camera (Fig. 3), obtaining synchronized touch signals $\{\tau_i\}_{i=1}^N$ and video frames $\mathbf{v}$ . We then estimate the pose of the video frames using off-the-shelf structure from motion methods [47], obtaining poses $\{p_i^v\}_{i=1}^N$ . Finally, we use the calibration of the mount to obtain the poses $\{p_i^t\}_{i=1}^N$ of the tactile measurements with respect to the scene's global reference frame. As a collection device, we mount an iPhone 14 Pro to one end of a camera rod, and a DIGIT [32] touch sensor to the other end. Note that the devices can be replaced with any RGB-D camera and vision-based tactile sensor. + +Capturing setup calibration. To find the relative pose between the camera and the touch sensor (Fig. 3), we exploit the fact that arbitrary viewpoints can be synthesized from $F_{\theta}$ , and that ubiquitous vision-based touch sensors are based on perspective cameras. In these sensors, an elastomer gel is placed on the lens of a commodity camera, which is illuminated by colored lights. When the gel is pressed into an object, it deforms, and the camera records an image of the deformation; this image is used as the tactile signal. This design allows us to estimate the pose of the tactile sensor through multi-view constraints from visual-tactile correspondences: pixels in visual images and tactile images that are of the same physical point. + +We start the calibration process by synthesizing novel views from $F_{\theta}$ . The views are generated at the camera location $\{p_i^v\}_{i = 1}^N$ , but rotated $90^{\circ}$ on the $x$ -axis. This is because the camera is approximately orthogonal to the touch sensor (see Fig. 3). Then, we manually annotate corresponding pixels between the touch measurements and the generated frames (Fig. 3). To simplify and standardize this process, we place a braille board in each scene and probe it with the touch sensor. This will generate a distinctive touch signal that is easy to localize [23]. + +We formulate the problem of estimating the six degrees of freedom relative pose $(\mathbf{R},\mathbf{t})$ between the touch sensor and the generated frames as a resectioning problem [20]. We use the estimated 3D structure from the NeRF $F_{\theta}$ to obtain 3D points $\{\mathbf{x}_i\}_{i = 1}^M$ for each of the annotated corre + +spondences. Each point has a pixel position $\mathbf{u}_i\in \mathcal{R}^2$ in the touch measurement. We find $(\mathbf{R},\mathbf{t})$ by minimizing the reprojection error: + +$$ +\min _ {\mathbf {R}, \mathbf {t}} \frac {1}{M} \sum_ {i = 1} ^ {M} \| \pi (\mathbf {K} [ \mathbf {R} \mid \mathbf {t} ], \mathbf {X} _ {i}) - \mathbf {u} _ {i} \| _ {1}, \tag {1} +$$ + +where $\pi$ projects a 3D point using a given projection matrix, $\mathbf{K}$ are the known intrinsic of the tactile sensor's camera, and the point $\mathbf{X}_i$ is in the coordinate system of the generated vision frames. We perform the optimization on 6-15 annotated correspondences from the braille board. For robustness, we compute correspondences from multiple frames. We represent the rotation matrix using quaternions and optimize using nonlinear least-squares. Once we have $(\mathbf{R},\mathbf{t})$ with respect to the generated frames, we can derive the relative pose between the camera and the touch sensor. + +# 3.2. Imputing the missing touch + +We use a generative model to estimate the touch signal (represented as an image from a vision-based touch sensor) for other locations within the scene. Specifically, we train a diffusion model $p_{\phi}(\pmb{\tau} \mid \mathbf{v}, \mathbf{d}, \mathbf{b})$ , where $\mathbf{v}$ and $\mathbf{d}$ are images and depth maps extracted from $F_{\theta}$ (see Fig. 4). We also pass as input to the diffusion model a background image captured by the touch sensor when it is not in contact with anything, denoted as $\mathbf{b}$ . Although not essential, we have observed that this additional input empirically improves the model's performance (e.g., Fig. 1 the background provides the location of defects in the gel, which appear as black dots). We train the model $p_{\phi}$ on our entire vision-touch dataset (Sec. 4). + +The training of $p_{\phi}$ is divided into two stages. In the first, we pre-train a cross-modal visual-tactile encoder with self-supervised contrastive learning on our dataset. This stage, initially proposed by [23, 57], is equivalent to the self-supervised encoding pre-training that is common for image generation models [45]. We use a ResNet-50 [21] as the backbone for this contrastive model. + +In the second stage, we use the contrastive model to generate the input for a conditional latent diffusion model, which is built upon Stable Diffusion [45]. A frozen pretrained VQ-GAN [14] is used to obtain the latent representation with a spatial dimension of $64 \times 64$ . We start training the diffusion model from scratch and pre-train it on the task of unconditional tactile image generation on the YCB-Slide dataset [52]. After this stage, we train the conditional generative model $p_{\phi}$ on our spatially aligned visual-tactile dataset, further fine-tuning the contrastive model end-to-end with the generation task. + +At inference time, given a novel location in the 3D scene, we first render the visual signals $\hat{\mathbf{v}}$ and $\hat{\mathbf{d}}$ from NeRF, and then estimate the touch signal $\hat{\tau}$ of the position using the diffusion model. + +![](images/90c5c0e0784f49af905258e53b93778c4f6483f3d7b65ffd45d33327ff63e174.jpg) +Figure 4. Touch estimation. We estimate the tactile signal for a given touch sensor pose $(\mathbf{R},\mathbf{t})$ . To do this, we synthesize a viewpoint from the NeRF, along with a depth map. We use conditional latent diffusion to predict the tactile signal from these inputs. + +# 4. A 3D Visual-Tactile Dataset + +In the following, we show the details of the data collection process and statistics of our dataset. + +# 4.1. Data Collection Procedure + +The data collection procedure is divided into two stages. First, we collect multiple views from the scene, capturing enough frames around the areas we plan to touch. During this stage, we collect approximately 500 frames. Next, we collect synchronized visual and touch data, maximizing the geometry and texture being touched. We then estimate the camera location of the vision frames collected in the previous two stages using off-the-shelf mapping tools [47]. After estimating the camera poses for the vision frames, the touch measurements' poses can be derived by using the mount calibration matrix. More details about the pose estimation procedure can be found in the supplement. + +Finally, we associate each touch sensor with a color image by translating the sensor poses upwards by 0.4 meters and querying the NeRF with such poses. The field of view we use when querying the NeRF is $50^{\circ}$ . This provides us with approximately 1,500 temporally aligned vision-touch image pairs per scene. Note that this collection procedure is scalable since it does not require specific expertise or equipment and generates abundant scene-level samples. + +# 4.2. Dataset Statistics + +We collect our data in 13 ordinary scenes including two offices, a workroom, a conference room, a corridor, a tabletop, a corridor, a lounge, a room with various clothes and four outdoor scenes with interesting materials. Typically, we collect 1k to 2k tactile probes in each scene, resulting in a total of 19.3k image pairs in the dataset. + +Some representative samples from the collected dataset are shown in Fig. 5. Our data includes a large variety of geometry (edges, surfaces, corners, etc.) and texture (plastic, clothes, snow, wood, etc.) of different materials in the scene. During capturing process, the collector will try to + +![](images/0425ab1485baea9d4aa79b2049a478df75308f6518983cc365682f89fecd9677.jpg) +Figure 5. Representative examples from the captured dataset. Our dataset is obtained from nine everyday scenes, such as offices, classrooms, and kitchens. We show three such scenes in the figure above, together with samples of spatially aligned visual and tactile data. In each scene, 1k to 2k tactile probes were collected, resulting in a total of 19.3k image pairs. The data encompasses diverse geometries (edges, surfaces, corners, etc.) and textures (plastic, clothes, snow, wood, etc.) of various materials. The collector systematically probed different objects, covering areas with distinct geometry and texture using different sensor poses. + +![](images/72ef09a2a86fd21d91441c2fef904784a3aa1d8591bf70e004561ad357c0d92f.jpg) + +![](images/ffab580b6af5c26b8f31f7f9ce073199f46a20eafd6a214a54dcc49186fff164.jpg) + +thoroughly probe various objects and cover the interesting areas with more distinguishable geometry and texture with different sensor poses. To the best of our knowledge, our dataset is the first dataset that captures full, scene-scale spatially aligned vision-touch image pairs. We provide more details about the dataset in the supplement. + +# 5. Experiments + +Leveraging the spatially aligned image and touch pairs from our dataset, we first conduct experiments on dense touch estimation. We then show the effectiveness of both the aligned data pairs and the synthesized touch signals by conducting tactile localization and material classification as two downstream tasks. + +# 5.1. Implementation Details + +NeRF. We use the Nerfacto method from Nerfstudio [53]. For each scene, we utilize approximately 2,000 images as training set, which thoroughly cover the scene from various view points. We train the network with a base learning rate of $1 \times 10^{-2}$ using Adam [30] optimizer for 200,000 steps on a single NVIDIA RTX 2080 Ti GPU to achieve optimal performance. + +Visual-tactile contrastive model. Following prior works [27, 57], we leverage contrastive learning methods to train a ResNet-50 [21] as visual encoder. The visual and tactile encoders share the same architecture but have different weights. We encode visual and tactile data into latent vectors in the resulting shared representation space. We set the dimension of the latent vectors to 32. Similar to CLIP [43], the model is trained on InfoNCE loss obtained from the pairwise dot products of the latent vectors. We train the model for 20 epochs by Adam [30] optimizer with a learning rate of $10^{-4}$ and batch size of 256 on 4 NVIDIA RTX 2080 Ti GPUs. + +Visual-tactile generative model. Our implementation of the diffusion model closely follows Stable Diffusion [46], with the difference that we use a ResNet-50 to generate the visual encoding from RGB-D images for conditioning. Specifically, we also add the RGB-D images rendered from the tactile sensors' poses into the conditioning, which we refer to in Sec. 5.2 as multiscale conditioning. The model is optimized for 30 epochs by Adam [30] optimizer with a base learning rate of $10^{-5}$ . The learning rate is scaled bygpu number $\times$ batch size. We train the model with batch size of 48 on 4 NVIDIA A40 GPUs. At inference time, the model conducts 200 steps of denoising process with a 7.5 guidance scale. Following prior cross-modal synthesis work [44], we use reranking to improve the prediction quality. We obtain 16 samples from the diffusion model for every instance and re-rank the samples with our pretrained contrastive model. The sample with highest similarity is the final prediction. + +# 5.2. Dense Touch Estimation + +Experimental setup. We now evaluate the diffusion model's ability to generate touch images. To reduce overlap between the training and test set, we first split the frames into sequences temporally (following previous work [56]). We split them into sequences of 50 touch samples, then divide these sequences into train/validation/test with a ratio of $8/1/1$ . We evaluate the generated samples on Frechet Inception Distance (FID), a standard evaluation metric for cross-modal generation [56]. We also include Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM), though we note that these metrics are highly sensitive to spatial position of the generated content, and can be optimized by models that minimize simple pixelwise losses [22]. We also include CVTP metric proposed by prior work [57], which measures the similarity between visual and tactile embeddings of a contrastive model, analogous to + +![](images/ad2ef9d07b3a69c8c5c5c29ecb00ee7e8d1b2af46f64297a32a167a00d04c60d.jpg) +Figure 6. Qualitative touch estimation results. Each model is conditioned on the RGB image and depth map rendered from the NeRF (left). The white box indicates the tactile sensor's approximate field of view (which is much smaller than the full conditional image). The G.T. column shows the ground truth touch images measured from a DIGIT sensor. L1 and VisGel often generate blurry textures and inaccurate geometry. By contrast, our model better captures the features of the tactile image, e.g., the rock's microgeometry and complex textures and shapes of furniture. The last row shows two failure cases of our model. In both examples, our model generates a touch image that is geometrically misaligned with the ground truth. All of the examples shown here are at least $10\mathrm{cm}$ away from any training sample. + +![](images/033fc97f74dea754889694412bebf42006df17669bb7afa432a61228e20e5808.jpg) + +CLIP [43] score. We compare against two baselines: VisGel, the approach from Li et. [35], which trains a GAN for touch generation, and $L1$ , a model with the same architecture of VisGel but trained to minimize an L1 loss in pixel space. + +Results. As is shown in Table 2, our approach performs much better on the high-level metrics, with up to $4\mathrm{x}$ lower FID and $80\mathrm{x}$ higher CVTP. This indicates that our proposed diffusion model captures the distribution and characteristics of the real tactile data more effectively. On the low-level metrics (PSNR and SSIM), all methods are comparable. In particular, the L1 model slightly outperforms the other methods since the loss it is trained on is highly correlated with low-level, pixel-wise metrics. Fig. 6 qualitatively compares samples from the different models. Indeed, our generated samples exhibit enhanced details in micro-geometry of fabrics and richer textures, including snow, wood and carpeting. However, all methods fail on fine details that are barely visible in the image, such as the tree bark. + +Ablation study. We evaluate the importance of the main components of our proposed touch generation approach (Table 3). Removing the conditioning on the RGB image results in the most prominent performance drop. This is expected since RGB image uniquely determines the fine + +
ModelPSNR ↑SSIM ↑FID ↓CVTP ↑
L124.340.8297.050.01
VisGel [35]23.660.81130.220.03
Ours22.840.7228.970.80
+ +Table 2. Quantitative results on touch estimation for novel views. While comparable on low-level metrics with the baselines, our approach captures the characteristics of the real tactile data more effectively, resulting in a lower FID score. + +grained details of a tactile image. Removing depth image or contrastive pretraining has small effect on CVTP but results in a drop on FID. Contrastive re-ranking largely improves CVTP, indicating the necessity of obtaining multiple samples from the diffusion model. We also find that multiscale conditioning provide a small benefit on FID and CVTP. + +# 5.3. Downstream Task I: Tactile Localization + +To help understand the quality of the captured TaRFs, we evaluate the performance of the contrastive model (used for conditioning our diffusion model) on the task of tactile localization. Given a tactile signal, our goal is to find the corresponding regions in a 2D image or in a 3D scene that are associated with it, i.e., we ask the question: what part of this image/scene feel like this? We perform the following + +![](images/e33279555accbb1569ef76ab055bae8e2f4e64e38f017072f23c0184e3ce9e0d.jpg) +Figure 7. Tactile localization heatmaps. Given a tactile query image, the heatmap shows the image patches with a higher affinity to this tactile signal, as measured by a contrastive model trained on our dataset. We use a sliding window and compare each extracted patch with the touch signal. In each case, the center patch is the true position. Our model successfully captures the correlation between the two signals. This enables it to localize a variety of touch signals, including fine-grained geometry, e.g., a cable or a keyboard, various types of corners and edges, and large uniform regions, such as a clothing. This ability enables our diffusion model to effectively propagate sparse touch samples to other visually and structurally similar regions of the scene. + +
Model variationPSNR ↑SSIM ↑FID ↓CVTP ↑
Full22.840.7228.970.80
No RGB conditioning22.130.7034.310.76
No depth conditioning22.570.7133.160.80
No contrastive pretraining22.820.7132.980.79
No re-ranking22.920.7229.460.61
No multiscale23.190.7230.890.77
+ +Table 3. Ablation study. Since the fine-grained details of touch images can be determined from a RGB image, removing conditioning on the latter results in the largest performance drops. Re-ranking has notable impact on CVTP, indicating the necessity of obtaining multiple samples from the diffusion model. + +evaluations on the test set of our dataset. Note that we run no task-specific training. + +2D Localization. To determine which part of an image are associated with a given tactile measurement, we follow the same setup of SSVTP [28]. We first split the image into patches and compute their embedding. Then, we generate the tactile embedding of the input touch image. Finally, we compute the pairwise similarities between the tactile and visual embeddings, which we plot as a heatmap. As we can see in Fig. 7, our constrastive encoder can successfully capture the correlations between the visual and tactile data. For instance, the tactile embeddings of edges are associated to edges of similar shape in the visual image. Note that the majority of tactile embeddings are highly ambiguous: all edges with a similar geometry feel the same. + +3D Localization. In 3D, the association of an image to tactile measurements becomes less ambiguous. Indeed, since tactile-visual samples are rotation-dependent, objects with similar shapes but different orientations will generate different tactile measurements. Lifting the task to 3D still does not remove all ambiguities (for example, each side of a rectangular table cannot be precisely localized). Nonetheless, we believe it to be a good fit for a quantitative evaluation since it's rare for two ambiguous parts of the scene to be touched with exactly the same orientation. + +We use the following experimental setup for 3D localization. Given a tactile image as a query, we compute its distance in embedding space to all visual test images from the same scene. Note that all test images are associated with a 3D location. We define as ground-truth correspondences all test images at a distance of at most $r$ from the 3D location of the test sample. We vary $r$ to account for local ambiguities. As typical in the retrieval literature, we benchmark the performance with metric mean Average Precision (mAP). + +We consider three baselines: (1) chance, which randomly selects corresponding samples; (2) real, which uses the contrastive model trained on our dataset; and (3) real + estimated, which trains the contrastive model on both dataset samples and a set of synthetic samples generated via the scenes' NeRF and our touch generation model. Specifically, we render a new image and corresponding touch by interpolating the position of two consecutive frames in the training dataset. This results in a training dataset for the contrastive model that is twice as large. + +
Datasetr(m)
0.0010.0050.010.050.1
Chance3.556.8210.2518.2621.33
Real12.1022.9332.1050.3057.15
Real + Est.14.9226.6936.1753.6260.61
+ +The results, presented in Table 4, demonstrate the performance benefit of employing both real and synthetic tactile pairs. Combining synthetic tactile images with the original pairs achieves highest performance on all distance thresholds. Overall, this indicates that touch measurements from novel views are not only qualitatively accurate, but also beneficial for this downstream task. + +# 5.4. Downstream Task II: Material Classification + +We investigate the efficacy of our visual-tactile dataset for understanding material properties, focusing on the task of material classification. We follow the formulation by Yang et al. [56], which consists of three subtasks: (i) material classification, requiring the distinction of materials among 20 possible classes; (ii) softness classification, a binary problem dividing materials as either hard or soft; and (iii) hardness classification, which requires the classification of materials as either rough or smooth. + +We follow the same experimental procedure of [56]: we pretrain a contrastive model on a dataset and perform linear probing on the sub-tasks' training set. Our experiments only vary the pretraining dataset, leaving all architectural choices and hyperparameters the same. We compare against four baselines. A random classifier (chance); the ObjectFolder 2.0 dataset [17]; the VisGel dataset [35]; and the Touch and Go dataset [56]. Note that the touch sensor used in the test data (GelSight) differs from the one used in our dataset (DIGIT). Therefore, we use for pretraining a combination of our dataset and Touch and Go. To ensure a fair comparison, we also compare to the combination of each dataset and Touch and Go. + +The findings from this evaluation, as shown in Table 5, suggest that our data improves the effectiveness of the contrastive pretraining objective, even though our data is from a different distribution. Moreover, we find that adding estimated touch probes for pretraining results in a higher performance on all the three tasks, especially the smoothness classification. This indicates that not only does our dataset covers a wide range of materials but also our diffusion model captures the distinguishable and useful patterns of different materials. + +Table 4. Quantitative results on 3D tactile localization. We evaluate using mean Average Precision (mAP) as a metric. Training the contrastive model on our dataset of visually aligned real samples together with estimated samples from new locations in the scene results in the highest performance. + +
DatasetMaterialHard/ SoftRough/ Smooth
Chance18.666.156.3
ObjectFolder 2.0 [17]36.272.069.0
VisGel [35]39.169.470.4
Touch and Go [56]54.777.379.4
+ ObjectFolder 2.0 [17]54.687.384.8
+ VisGel [35]53.186.783.6
+ Ours* (Real)57.688.481.7
+ Ours* (Real + Estimated)59.088.786.1
+ +Table 5. Material classification. We show the downstream material recognition accuracy of models pre-trained on different datasets. The final rows show the performance when combining different datasets with Touch and Go [56]. * The task-specific training and testing datasets for this task are collected with a Gel-Sight sensor. We note that our data comes from a different distribution, since it is collected with a DIGIT sensor [32]. + +# 6. Conclusion + +In this work, we present the TaRF, a scene representation that brings vision and touch into a shared 3D space. This representation enables the generation of touch probes for novel scene locations. To build this representation, we collect the largest dataset of spatially aligned vision and touch probes. We study the utility of both the representation and the dataset in a series of qualitative and quantitative experiments and on two downstream tasks: 3D touch localization and material recognition. Overall, our work makes the first step towards giving current scene representation techniques an understanding of not only how things look, but also how they feel. This capability could be critical in several applications ranging from robotics to the creation of virtual worlds that look and feel like the real world. + +Limitations. Since the touch sensor is based on a highly zoomed-in camera, small (centimeter-scale) errors in SfM or visual-tactile registration can lead to misalignments of several pixels between the views of the NeRF and the touch samples, which can be seen in our TaRFs. Another limitation of the proposed representation is the assumption that the scene's coarse-scale structure does not change when it is touched, an assumption that may be violated for some inelastic surfaces. + +Acknowledgements. We thank Jeongsoo Park, Ayush Shrivastava, Daniel Geng, Ziyang Chen, Zihao Wei, Zixuan Pan, Chao Feng, Chris Rockwell, Gaurav Kaul and the reviewers for the valuable discussion and feedback. This work was supported by an NSF CAREER Award #2339071, a Sony Research Award, the DARPA Machine Common Sense program, and ONR MURI award N00014-21-1-2801. + +# References + +[1] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, 2020. 3 +[2] Peter Anderson, Ayush Shrivastava, Joanne Truong, Arjun Majumdar, Devi Parikh, Dhruv Batra, and Stefan Lee. Sim-to-real transfer for vision-and-language navigation. In Conference on Robot Learning, pages 671-681. PMLR, 2021. 3 +[3] Jakub Bednarek, Michal Bednarek, Lorenz Wellhausen, Marco Hutter, and Krzysztof Walas. What am i touching? learning to classify terrain via haptic sensing. In 2019 International Conference on Robotics and Automation (ICRA), pages 7187-7193. IEEE, 2019. 1 +[4] Katherine L Bouman, Bei Xiao, Peter Battaglia, and William T Freeman. Estimating the material properties of fabric from video. In Proceedings of the IEEE international conference on computer vision, pages 1984-1991, 2013. 1 +[5] Alexander Burka. Instrumentation, data, and algorithms for visually understanding haptic surface properties. 2018. 2 +[6] Roberto Calandra, Andrew Owens, Manu Upadhyaya, Wenzhen Yuan, Justin Lin, Edward H. Adelson, and Sergey Levine. The feeling of success: Does touch sensing help predict grasp outcomes? Conference on Robot Learning (CoRL), 2017. 1, 2 +[7] Roberto Calandra, Andrew Owens, Dinesh Jayaraman, Justin Lin, Wenzhen Yuan, Jitendra Malik, Edward H. Adelson, and Sergey Levine. More than a feeling: Learning to grasp and regrasp using vision and touch. Robotics and Automation Letters (RA-L), 2018. 1, 2 +[8] Roberto Calandra, Andrew Owens, Dinesh Jayaraman, Justin Lin, Wenzhen Yuan, Jitendra Malik, Edward H. Adelson, and Sergey Levine. More than a feeling: Learning to grasp and regrasp using vision and touch. IEEE Robotics and Automation Letters, 3:3300-3307, 2018. 2 +[9] Berk Calli, Arjun Singh, Aaron Walsman, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. The ycb object and model set: Towards common benchmarks for manipulation research. In 2015 international conference on advanced robotics (ICAR), pages 510-517. IEEE, 2015. 2 +[10] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European conference on computer vision, 2020. 3 +[11] Alex Church, John Lloyd, Raia Hadsell, and Nathan F Lepora. Deep reinforcement learning for tactile robotics: Learning to type on a braille keyboard. IEEE Robotics and Automation Letters, 5(4):6145-6152, 2020. 1 +[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 3 + +[13] Abe Davis, Justin G Chen, and Frédo Durand. Image-space modal bases for plausible manipulation of objects in video. ACM Transactions on Graphics (TOG), 2015. 1 +[14] Mucong Ding, Kezhi Kong, Jingling Li, Chen Zhu, John Dickerson, Furong Huang, and Tom Goldstein. Vq-gnn: A universal framework to scale up graph neural networks using vector quantization. Advances in Neural Information Processing Systems, 34:6733-6746, 2021. 4 +[15] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence, 32(8):1362-1376, 2009. 2 +[16] Ruohan Gao, Yen-Yu Chang, Shivani Mall, Li Fei-Fei, and Jiajun Wu. Objectfolder: A dataset of objects with implicit visual, auditory, and tactile representations. In CoRL, 2021. 1, 2 +[17] Ruohan Gao, Zilin Si, Yen-Yu Chang, Samuel Clarke, Jeannette Bohg, Li Fei-Fei, Wenzhen Yuan, and Jiajun Wu. Objectfolder 2.0: A multisensory object dataset for sim2real transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10598–10608, 2022. 1, 2, 8 +[18] Ruohan Gao, Yiming Dou, Hao Li, Tanmay Agarwal, Jeannette Bohg, Yunzhu Li, Li Fei-Fei, and Jiajun Wu. The objectfolder benchmark: Multisensory learning with neural and real objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17276–17286, 2023. 1, 2, 3 +[19] Michael SA Graziano and Charles G Gross. The representation of extrapersonal space: A possible role for bimodal, visual-tactile neurons. 1995. 1 +[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 2, 3 +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 5 +[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5 +[23] Carolina Higuera, Byron Boots, and Mustafa Mukadam. Learning to read braille: Bridging the tactile reality gap with diffusion models. arXiv preprint arXiv:2304.01182, 2023. 3, 4 +[24] Mark A Hoepflinger, C David Remy, Marco Hutter, Luciano Spinello, and Roland Siegwart. Haptic terrain classification for legged robots. In 2010 IEEE International Conference on Robotics and Automation, pages 2828-2833. IEEE, 2010. 1 +[25] Micah K Johnson and Edward H Adelson. Retrographic sensing for the measurement of surface texture and shape. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 1070-1077. IEEE, 2009. 2, 3 +[26] Micah K Johnson, Forrester Cole, Alvin Raj, and Edward H Adelson. Microgeometry capture using an elastomeric sensor. ACM Transactions on Graphics (TOG), 2011. 2, 3 + +[27] Justin Kerr, Huang Huang, Albert Wilcox, Ryan Hoque, Jeffrey Ichnowski, Roberto Calandra, and Ken Goldberg. Learning self-supervised representations from vision and touch for active sliding perception of deformable surfaces. arXiv preprint arXiv:2209.13042, 2022. 5 +[28] Justin Kerr, Huang Huang, Albert Wilcox, Ryan Hoque, Jeffrey Ichnowski, Roberto Calandra, and Ken Goldberg. Self-supervised visuo-tactile pretraining to locate and follow garment features. In Robotics: Science and Systems, 2023. 1, 2, 3, 7 +[29] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3 +[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. *ICLR*, 2015. 5 +[31] Hendrik Kolvenbach, Christian Bärtschi, Lorenz Wellhausen, Ruben Grandia, and Marco Hutter. Haptic inspection of planetary soils with legged robots. IEEE Robotics and Automation Letters, 4(2):1626-1632, 2019. 1 +[32] Mike Lambeta, Po-Wei Chou, Stephen Tian, Brian Yang, Benjamin Maloon, Victoria Rose Most, Dave Stroud, Raymond Santos, Ahmad Byagowi, Gregg Kammerer, et al. Digit: A novel design for a low-cost compact high-resolution tactile sensor with application to in-hand manipulation. IEEE Robotics and Automation Letters, 2020. 1, 2, 3, 8 +[33] Maxime Lhuillier and Long Quan. A quasi-dense approach to surface reconstruction from uncalibrated images. IEEE transactions on pattern analysis and machine intelligence, 27(3):418-433, 2005. 2 +[34] Hongyu Li, Snehal Dikhale, Soshi Iba, and Nawid Jamali. Vihope: Visuotactile in-hand object 6d pose estimation with shape completion. IEEE Robotics and Automation Letters, 8 (11):6963-6970, 2023. 1 +[35] Yunzhu Li, Jun-Yan Zhu, Russ Tedrake, and Antonio Torralba. Connecting touch and vision via cross-modal prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10609-10618, 2019. 1, 2, 3, 6, 8 +[36] Justin Lin, Roberto Calandra, and Sergey Levine. Learning to identify object instances by touch: Tactile recognition via multimodal matching. In 2019 International Conference on Robotics and Automation (ICRA), pages 3644-3650. IEEE, 2019. 3 +[37] Antonio Loquercio, Ashish Kumar, and Jitendra Malik. Learning visual locomotion with cross-modal supervision. In 2023 IEEE International Conference on Robotics and Automation (ICRA), 2023. 1 +[38] Gabriel B Margolis, Xiang Fu, Yandong Ji, and Pulkit Agrawal. Learning physically grounded robot vision with active sensing motor policies. In 7th Annual Conference on Robot Learning, 2023. 1 +[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3 + +[40] Andrew Owens, Phillip Isola, Josh McDermott, Antonio Torralba, Edward H Adelson, and William T Freeman. Visually indicated sounds. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2405-2413, 2016. 1 +[41] Senthil Purushwalkam, Abhinav Gupta, Danny M Kaufman, and Bryan Russell. Bounce and learn: Modeling scene dynamics with real-world bounces. arXiv preprint arXiv:1904.06827, 2019. 1 +[42] Haozhi Qi, Brent Yi, Sudharshan Suresh, Mike Lambeta, Yi Ma, Roberto Calandra, and Jitendra Malik. General in-hand object rotation with vision and touch. arXiv preprint arXiv:2309.09979, 2023. 1 +[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 5, 6 +[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 5 +[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 4 +[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5 +[47] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3, 4 +[48] Carmelo Sferrazza and Raffaello D'Andrea. Design, motivation and evaluation of a full-resolution optical tactile sensor. Sensors, 19(4):928, 2019. 2 +[49] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV 2012, ECCV 2012. 3 +[50] Linda Smith and Michael Gasser. The development of embodied cognition: Six lessons from babies. Artificial life, 2005. 1 +[51] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2 +[52] Sudharshan Suresh, Zilin Si, Stuart Anderson, Michael Kaess, and Mustafa Mukadam. Midastouch: Monte-carlo inference over distributions across sliding touch. In Conference on Robot Learning, pages 319-331. PMLR, 2023. 1, 2, 4 +[53] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David Mcallister, Justin Kerr, and Angjoo Kanazawa. Nerfstudio: A modu + +lar framework for neural radiance field development. arXiv preprint arXiv:2302.04264, 2023. 3, 5 +[54] Shaoxiong Wang, Jiajun Wu, Xingyuan Sun, Wenzhen Yuan, William T Freeman, Joshua B Tenenbaum, and Edward H Adelson. 3d shape perception from monocular vision, touch, and shape priors. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018. 3 +[55] Jianxiong Xiao, Andrew Owens, and Antonio Torralba. Sun3d: A database of big spaces reconstructed using sfm and object labels. 2013. 3 +[56] Fengyu Yang, Chenyang Ma, Jiacheng Zhang, Jing Zhu, Wenzhen Yuan, and Andrew Owens. Touch and go: Learning from human-collected vision and touch. Neural Information Processing Systems (NeurIPS) - Datasets and Benchmarks Track, 2022. 2, 3, 5, 8 +[57] Fengyu Yang, Jiacheng Zhang, and Andrew Owens. Generating visual scenes from touch. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22070-22080, 2023. 2, 3, 4, 5 +[58] Fengyu Yang, Chao Feng, Ziyang Chen, Hyoungseob Park, Daniel Wang, Yiming Dou, Ziyao Zeng, Xien Chen, Rit Gangopadhyay, Andrew Owens, and Alex Wong. Binding touch to everything: Learning unified multimodal tactile representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3 +[59] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 3 +[60] Zhao-Heng Yin, Binghao Huang, Yuzhe Qin, Qifeng Chen, and Xiaolong Wang. Rotating without seeing: Towards in-hand dexterity through touch. arXiv preprint arXiv:2303.10880, 2023. 1 +[61] Wenzhen Yuan, Shaoxiong Wang, Siyuan Dong, and Edward Adelson. Connecting look and feel: Associating the visual and tactile properties of physical materials. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5580-5588, 2017. 3 +[62] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. 3 +[63] Shaohong Zhong, Alessandro Albini, Oiwi Parker Jones, Perla Maiolino, and Ingmar Posner. Touching a nerf: Leveraging neural radiance fields for tactile sensory data generation. In Conference on Robot Learning, pages 1618-1628. PMLR, 2023. 1, 2 \ No newline at end of file diff --git a/2024/Tactile-Augmented Radiance Fields/images.zip b/2024/Tactile-Augmented Radiance Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..65744ab623653af95001bf5f3fb7dbd32cc4b60d --- /dev/null +++ b/2024/Tactile-Augmented Radiance Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59d6b08114ec1baaa80cfe1d1f2b2ff1c82d3a346e79f15f2dfd12140f489294 +size 606523 diff --git a/2024/Tactile-Augmented Radiance Fields/layout.json b/2024/Tactile-Augmented Radiance Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4771d500b326dae731e9cb38aae88308fc72c8eb --- /dev/null +++ b/2024/Tactile-Augmented Radiance Fields/layout.json @@ -0,0 +1,8071 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 187, + 103, + 408, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 103, + 408, + 121 + ], + "spans": [ + { + "bbox": [ + 187, + 103, + 408, + 121 + ], + "type": "text", + "content": "Tactile-Augmented Radiance Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 141, + 492, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 141, + 492, + 159 + ], + "spans": [ + { + "bbox": [ + 86, + 141, + 492, + 159 + ], + "type": "text", + "content": "Yiming Dou" + }, + { + "bbox": [ + 86, + 141, + 492, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 141, + 492, + 159 + ], + "type": "text", + "content": " Fengyu Yang" + }, + { + "bbox": [ + 86, + 141, + 492, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 86, + 141, + 492, + 159 + ], + "type": "text", + "content": " Yi Liu" + }, + { + "bbox": [ + 86, + 141, + 492, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 141, + 492, + 159 + ], + "type": "text", + "content": " Antonio Loquercio" + }, + { + "bbox": [ + 86, + 141, + 492, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 86, + 141, + 492, + 159 + ], + "type": "text", + "content": " Andrew Owens" + }, + { + "bbox": [ + 86, + 141, + 492, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 164, + 444, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 164, + 444, + 180 + ], + "spans": [ + { + "bbox": [ + 149, + 164, + 444, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 164, + 444, + 180 + ], + "type": "text", + "content": "University of Michigan " + }, + { + "bbox": [ + 149, + 164, + 444, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 164, + 444, + 180 + ], + "type": "text", + "content": "Yale University " + }, + { + "bbox": [ + 149, + 164, + 444, + 180 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 149, + 164, + 444, + 180 + ], + "type": "text", + "content": "UC Berkeley" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 204, + 387, + 339 + ], + "blocks": [ + { + "bbox": [ + 48, + 204, + 387, + 339 + ], + "lines": [ + { + "bbox": [ + 48, + 204, + 387, + 339 + ], + "spans": [ + { + "bbox": [ + 48, + 204, + 387, + 339 + ], + "type": "image", + "image_path": "0f549b524105136b5fc7c79c0afb9d674e867f0888bda77c6bc2a1700dc4bb4d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 350, + 547, + 407 + ], + "lines": [ + { + "bbox": [ + 45, + 350, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 45, + 350, + 547, + 407 + ], + "type": "text", + "content": "Figure 1. Tactile-augmented radiance fields. We capture a tactile-augmented radiance field (TaRF) from photos and sparsely sampled touch probes. To do this, we register the captured visual and tactile signals into a shared 3D space, then train a diffusion model to impute touch at other locations within the scene. Here, we visualize two touch probes and their (color coded) 3D positions in the scene. We also show two touch signals estimated by the diffusion model. The touch signals were collected using a vision-based touch sensor [32] that represents the touch signals as images. Please see our project page for video results." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 388, + 204, + 545, + 340 + ], + "blocks": [ + { + "bbox": [ + 388, + 204, + 545, + 340 + ], + "lines": [ + { + "bbox": [ + 388, + 204, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 388, + 204, + 545, + 340 + ], + "type": "image", + "image_path": "163f41aec7e0b4020a7810e4484f53cc9989565a88fdaa2782da932bf26321d3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 422, + 192, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 422, + 192, + 434 + ], + "spans": [ + { + "bbox": [ + 143, + 422, + 192, + 434 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 448, + 290, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 290, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 290, + 701 + ], + "type": "text", + "content": "We present a scene representation, which we call a tactile-augmented radiance field (TaRF), that brings vision and touch into a shared 3D space. This representation can be used to estimate the visual and tactile signals for a given 3D position within a scene. We capture a scene's TaRF from a collection of photos and sparsely sampled touch probes. Our approach makes use of two insights: (i) common vision-based touch sensors are built on ordinary cameras and thus can be registered to images using methods from multi-view geometry, and (ii) visually and structurally similar regions of a scene share the same tactile features. We use these insights to register touch signals to a captured visual scene, and to train a conditional diffusion model that, provided with an RGB-D image rendered from a neural radiance field, generates its corresponding tactile signal. To evaluate our approach, we collect a dataset of TaRFs. This dataset contains more touch samples than previous real-world datasets, and it provides spatially aligned visual signals for each captured touch signal. We demonstrate the accuracy of our cross-modal generative model and the utility of the captured visual-tactile data on several downstream tasks. Project page: https://dou-yiming.github.io/TaRF." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 422, + 386, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 386, + 434 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 386, + 434 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 446, + 546, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 546, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 546, + 590 + ], + "type": "text", + "content": "As humans, our ability to perceive the world relies crucially on cross-modal associations between sight and touch [19, 50]. Tactile sensing provides a detailed understanding of material properties and microgeometry, such as the intricate patterns of bumps on rough surfaces and the complex motions that soft objects make when they deform. This type of understanding, which largely eludes today's computer vision models, is a critical component of applications that require reasoning about physical contact, such as robotic locomotion [3, 24, 31, 34, 37, 38] and manipulation [6, 7, 11, 42, 60], and methods that simulate the behavior of materials [4, 13, 40, 41]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "In comparison to many other modalities, collecting tactile data is an expensive and tedious process, since it requires direct physical interaction with the environment. A recent line of work has addressed this problem by having humans or robots probe the environment with touch sensors (see Table 1). Early efforts have been focused on capturing the properties of only a few objects either in simulation [16, 17, 52] or in lab-controlled settings [6, 7, 18, 28, 35, 52, 63], which may not fully convey the diversity of tactile signals in natural environments. Other works have gone beyond a" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26529" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 289, + 203 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 289, + 203 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 289, + 203 + ], + "type": "table", + "html": "
DatasetSamplesAlignedScenarioSource
More Than a Feeling [7]6.5k×TabletopRobot
Feeling of Success [6]9.3k×TabletopRobot
VisGel [35]12k×TabletopRobot
SSVTP [28]4.6k✓TabletopRobot
ObjectFolder 1.0 [16]-✓ObjectSynthetic
ObjectFolder 2.0 [17]-✓ObjectSynthetic
ObjectFolder Real [18]3.7k×ObjectRobot
Burka et al. [5]1.1k×Sub-sceneHuman
Touch and Go [56]13.9k×Sub-sceneHuman
YCB-Slide* [52]-✓ObjectHuman
Touching a NeRF [63]1.2k✓ObjectRobot
TaRF (Ours)19.3k✓Full sceneHuman
", + "image_path": "0cf3144a6253dfddb1641ce1efae8011216f8faaafc4e38dde03f09c7062ccdb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 204, + 287, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 287, + 270 + ], + "type": "text", + "content": "Table 1. Dataset comparison. We present the number of real visual-tactile pairs and whether such pairs are visually aligned, i.e., whether the visual image includes an occlusion-free view of the touched surface. *YCB-Slide has real-world touch probes but synthetic images rendered with CAD models of YCB objects on a white background [9]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "text", + "content": "lab setting and have collected touch from real scenes [5, 56]. However, existing datasets lack aligned visual and tactile information, since the touch sensor and the person (or robot) that holds it often occlude large portions of the visual scene (Fig. 2). These datasets also contain only a sparse set of touch signals for each scene, and it is not clear how the sampled touch signals relate to each other in 3D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 366, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 544 + ], + "type": "text", + "content": "In this work, we present a simple and low-cost procedure to capture quasi-dense, scene-level, and spatially-aligned visual and touch data (Fig. 1). We call the resulting scene representation a tactile-augmented radiance field (TaRF). We remove the need for robotic collection by leveraging a 3D scene representation (a NeRF [39]) to synthesize a view of the surface being touched, which results in spatially aligned visual-tactile data (Fig. 2). We collect this data by mounting a touch sensor to a camera with commonly available materials (Fig. 3). To calibrate the pair of sensors, we take advantage of the fact that popular vision-based touch sensors [25, 26, 32, 48] are built on ordinary cameras. The relative pose between the vision and tactile sensors can thus be estimated using traditional methods from multi-view geometry, such as camera resectioning [20]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "We use this procedure to collect a large real-world dataset of aligned visual-tactile data. With this dataset, we train a diffusion model [45, 51] to estimate touch at locations not directly probed by a sensor. In contrast to the recent work of Zhong et al. [63], which also estimates touch from 3D NeRF geometry, we create scene-scale reconstructions, we do not require robotic proprioception, and we use diffusion models [51]. This enables us to obtain tactile data at a much larger scale, and with considerably more diversity. Unlike previous visual-tactile diffusion work [57], we condition the model on spatially aligned visual and depth information, enhancing the generated samples' quality and their usefulness in downstream applications. After training, the diffusion model can be used to predict tactile informa" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 346, + 145 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 346, + 145 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 346, + 145 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 346, + 145 + ], + "type": "image", + "image_path": "48963dd64dd1561761f3cb4229bcb719dd4ca3e677470e89430566795a3499d9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 148, + 343, + 156 + ], + "lines": [ + { + "bbox": [ + 309, + 148, + 343, + 156 + ], + "spans": [ + { + "bbox": [ + 309, + 148, + 343, + 156 + ], + "type": "text", + "content": "OF 2.0 [17]" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 347, + 71, + 385, + 146 + ], + "blocks": [ + { + "bbox": [ + 347, + 71, + 385, + 146 + ], + "lines": [ + { + "bbox": [ + 347, + 71, + 385, + 146 + ], + "spans": [ + { + "bbox": [ + 347, + 71, + 385, + 146 + ], + "type": "image", + "image_path": "3df34be0c7d888260703250ad091c94e3386f9e804c9570839d68d4a1c58539d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 148, + 383, + 156 + ], + "lines": [ + { + "bbox": [ + 349, + 148, + 383, + 156 + ], + "spans": [ + { + "bbox": [ + 349, + 148, + 383, + 156 + ], + "type": "text", + "content": "VisGel [35]" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 386, + 71, + 425, + 146 + ], + "blocks": [ + { + "bbox": [ + 386, + 71, + 425, + 146 + ], + "lines": [ + { + "bbox": [ + 386, + 71, + 425, + 146 + ], + "spans": [ + { + "bbox": [ + 386, + 71, + 425, + 146 + ], + "type": "image", + "image_path": "ca5bdb3189ea7e7413fe4ab6b4f7dd5ae94e7eb6e988f566e701d29386cb889e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 148, + 425, + 156 + ], + "lines": [ + { + "bbox": [ + 387, + 148, + 425, + 156 + ], + "spans": [ + { + "bbox": [ + 387, + 148, + 425, + 156 + ], + "type": "text", + "content": "OF Real [18]" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 426, + 71, + 464, + 146 + ], + "blocks": [ + { + "bbox": [ + 426, + 71, + 464, + 146 + ], + "lines": [ + { + "bbox": [ + 426, + 71, + 464, + 146 + ], + "spans": [ + { + "bbox": [ + 426, + 71, + 464, + 146 + ], + "type": "image", + "image_path": "66408a8ce54ef733c07222aed6a558248563d7ddf5d651fa8124f3052351a023.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 428, + 148, + 462, + 156 + ], + "lines": [ + { + "bbox": [ + 428, + 148, + 462, + 156 + ], + "spans": [ + { + "bbox": [ + 428, + 148, + 462, + 156 + ], + "type": "text", + "content": "SSVTP [28]" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 160, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 305, + 160, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 160, + 545, + 205 + ], + "type": "text", + "content": "Figure 2. Visual-tactile examples. In contrast to the visual-tactile data captured in previous work, our approach allows us to sample unobstructed images that are spatially aligned with the touch signal, from arbitrary 3D viewpoints using a NeRF." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 466, + 71, + 503, + 146 + ], + "blocks": [ + { + "bbox": [ + 466, + 71, + 503, + 146 + ], + "lines": [ + { + "bbox": [ + 466, + 71, + 503, + 146 + ], + "spans": [ + { + "bbox": [ + 466, + 71, + 503, + 146 + ], + "type": "image", + "image_path": "48eeffb2c37b18e743fe79503e6b089918e7a58f0fd656316a65967904212e1f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 148, + 496, + 156 + ], + "lines": [ + { + "bbox": [ + 471, + 148, + 496, + 156 + ], + "spans": [ + { + "bbox": [ + 471, + 148, + 496, + 156 + ], + "type": "text", + "content": "TG [56]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 504, + 71, + 542, + 146 + ], + "blocks": [ + { + "bbox": [ + 504, + 71, + 542, + 146 + ], + "lines": [ + { + "bbox": [ + 504, + 71, + 542, + 146 + ], + "spans": [ + { + "bbox": [ + 504, + 71, + 542, + 146 + ], + "type": "image", + "image_path": "fc56a6478352a81e2bb1314e0cb0dcb53f113f7f611781af4eef5663883d9afd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 505, + 148, + 542, + 156 + ], + "lines": [ + { + "bbox": [ + 505, + 148, + 542, + 156 + ], + "spans": [ + { + "bbox": [ + 505, + 148, + 542, + 156 + ], + "type": "text", + "content": "TaRF (Ours)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 213, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 545, + 272 + ], + "type": "text", + "content": "tion for novel positions in the scene. Analogous to quasi-dense stereo methods [15, 33], the diffusion model effectively propagates sparse touch samples, obtained by probing, to other visually and structurally similar regions of the scene." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 273, + 546, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 546, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 546, + 345 + ], + "type": "text", + "content": "We evaluate our visual-tactile model's ability to accurately perform cross-modal translation using a variety of quality metrics. We also apply it to several downstream tasks, including localizing a touch within a scene and understanding material properties of the touched area. Our experiments suggest:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 346, + 545, + 453 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 306, + 346, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 545, + 369 + ], + "type": "text", + "content": "- Touch signals can be localized in 3D space by exploiting multi-view geometry constraints between sight and touch." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 370, + 545, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 545, + 404 + ], + "type": "text", + "content": "- Estimated touch measurements from novel views are not only qualitatively accurate, but also beneficial on downstream tasks." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 406, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 545, + 429 + ], + "type": "text", + "content": "- Cross-modal prediction models can accurately estimate touch from sight for natural scenes." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 430, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 545, + 453 + ], + "type": "text", + "content": "- Visually-acquired 3D scene geometry improves cross-modal prediction." + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 460, + 392, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 392, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 392, + 472 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 483, + 546, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 546, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 546, + 686 + ], + "type": "text", + "content": "Visual-tactile datasets. Previous work has either used simulators [16, 17] or robotic arms [6, 8, 18, 35, 63] for data generation. Our work is closely related to that of Zhong et al. [63], which uses a NeRF and captured touch data to generate a tactile field for several small objects. They use the proprioception of an expensive robot to spatially align vision and touch. In contrast, we leverage the properties of the tactile sensor and novel view synthesis to use commonly available material (a smartphone and a selfie stick) to align vision and touch. This enables the collection of a larger, scene-level, and more diverse dataset, on which we train a higher-capacity diffusion model (rather than a conditional GAN). Like several previous works [5, 56], we also collect scene-level data. In contrast to them, we spatially align the signals by registering them in a unified 3D representation, thereby increasing the prediction power of the visual-tactile generative model." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Capturing multimodal 3D scenes. Our work is related to methods that capture 3D visual reconstructions of spaces" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26530" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "using RGB-D data [12, 49, 55, 59] and multimodal datasets of paired 3D vision and language [1, 2, 10]. Our work is also related to recent methods that localize objects in NeRFs using joint embeddings between images and language [29] or by semantic segmentation [62]. In contrast to language supervision, touch is tied to a precise position in a scene." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 147, + 288, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 147, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 288, + 255 + ], + "type": "text", + "content": "3D touch sensing. A variety of works have studied the close relationship between geometry and touch, motivating our use of geometry in imputing touch. Johnson et al. [25, 26] proposed vision-based touch sensing, and showed that highly accurate depth can be estimated from the touch sensor using photometric stereo. Other work has estimated object-scale 3D from touch [54]. By contrast, we combine sparse estimates of touch with quasi-dense tactile signals estimated using generative models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 258, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 425 + ], + "type": "text", + "content": "Cross-modal prediction of touch from sight. Recent work has trained generative models that predict touch from images. Li et al. [35] used a GAN to predict touch for images of a robotic arm, while Gao et al. [18] applied them to objects collected on a turntable. Yang et al. [57] used latent diffusion to predict touch from videos of humans touching objects. Our goal is different from these works: we want to predict touch signals that are spatially aligned with a visual signal, to exploit scene-specific information, and to use geometry. Thus, we use a different architecture and conditioning signal, and fit our model to examples from the same scenes at training and test time. Other work has learned joint embeddings between vision and touch [28, 36, 56, 58, 61]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 435, + 102, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 435, + 102, + 447 + ], + "spans": [ + { + "bbox": [ + 47, + 435, + 102, + 447 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "text", + "content": "We collect visual and tactile examples from a scene and register them together with a 3D visual reconstruction to build a TaRF. Specifically, we capture a NeRF " + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "inline_equation", + "content": "F_{\\theta}:(\\mathbf{x},\\mathbf{r})\\mapsto (\\mathbf{c},\\sigma)" + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "text", + "content": " that maps a 3D point " + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = (x,y,z)" + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "text", + "content": " and viewing direction " + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "text", + "content": " to its corresponding RGB color " + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "text", + "content": " [39]. We associate to the visual representation a touch model " + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "inline_equation", + "content": "F_{\\phi}:\\mathbf{v}_{t}\\mapsto \\pmb{\\tau}" + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "text", + "content": " that generates the tactile signal that one would obtain by touching at the center of the image " + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t" + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "text", + "content": ". In the following, we explain how to estimate " + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "inline_equation", + "content": "F_{\\phi}" + }, + { + "bbox": [ + 46, + 455, + 287, + 574 + ], + "type": "text", + "content": " and put them into the same shared 3D space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 582, + 232, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 232, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 232, + 594 + ], + "type": "text", + "content": "3.1. Capturing vision and touch signals" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 602, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 602, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 602, + 287, + 686 + ], + "type": "text", + "content": "Obtaining a visual 3D reconstruction. We build the visual NeRF, " + }, + { + "bbox": [ + 46, + 602, + 287, + 686 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 46, + 602, + 287, + 686 + ], + "type": "text", + "content": ", closely following previous work [12, 55]. A human data collector moves through a scene and records a video, covering as much of the space as possible. We then estimate camera pose using structure from motion [47] and create a NeRF using off-the-shelf packages [53]. Additional details are provided in the supplement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Capturing and registering touch. We simultaneously collect tactile and visual signals by mounting a touch sensor" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 544, + 169 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 544, + 169 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 544, + 169 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 544, + 169 + ], + "type": "image", + "image_path": "04fa5dd0c1d1252de232aed3efddd62a6bfb07a6e9a07d34030942e3d1e09d7a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 178, + 545, + 223 + ], + "lines": [ + { + "bbox": [ + 305, + 178, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 545, + 223 + ], + "type": "text", + "content": "Figure 3. Capturing setup. (a) We record paired vision and touch signals using a camera attached to a touch sensor. (b) We estimate the relative pose between the touch sensor and the camera using correspondences between sight and touch." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 244, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 244, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 244, + 545, + 374 + ], + "type": "text", + "content": "on a camera (Fig. 3), obtaining synchronized touch signals " + }, + { + "bbox": [ + 305, + 244, + 545, + 374 + ], + "type": "inline_equation", + "content": "\\{\\tau_i\\}_{i=1}^N" + }, + { + "bbox": [ + 305, + 244, + 545, + 374 + ], + "type": "text", + "content": " and video frames " + }, + { + "bbox": [ + 305, + 244, + 545, + 374 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 305, + 244, + 545, + 374 + ], + "type": "text", + "content": ". We then estimate the pose of the video frames using off-the-shelf structure from motion methods [47], obtaining poses " + }, + { + "bbox": [ + 305, + 244, + 545, + 374 + ], + "type": "inline_equation", + "content": "\\{p_i^v\\}_{i=1}^N" + }, + { + "bbox": [ + 305, + 244, + 545, + 374 + ], + "type": "text", + "content": ". Finally, we use the calibration of the mount to obtain the poses " + }, + { + "bbox": [ + 305, + 244, + 545, + 374 + ], + "type": "inline_equation", + "content": "\\{p_i^t\\}_{i=1}^N" + }, + { + "bbox": [ + 305, + 244, + 545, + 374 + ], + "type": "text", + "content": " of the tactile measurements with respect to the scene's global reference frame. As a collection device, we mount an iPhone 14 Pro to one end of a camera rod, and a DIGIT [32] touch sensor to the other end. Note that the devices can be replaced with any RGB-D camera and vision-based tactile sensor." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 378, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 533 + ], + "type": "text", + "content": "Capturing setup calibration. To find the relative pose between the camera and the touch sensor (Fig. 3), we exploit the fact that arbitrary viewpoints can be synthesized from " + }, + { + "bbox": [ + 304, + 378, + 545, + 533 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 304, + 378, + 545, + 533 + ], + "type": "text", + "content": ", and that ubiquitous vision-based touch sensors are based on perspective cameras. In these sensors, an elastomer gel is placed on the lens of a commodity camera, which is illuminated by colored lights. When the gel is pressed into an object, it deforms, and the camera records an image of the deformation; this image is used as the tactile signal. This design allows us to estimate the pose of the tactile sensor through multi-view constraints from visual-tactile correspondences: pixels in visual images and tactile images that are of the same physical point." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 653 + ], + "type": "text", + "content": "We start the calibration process by synthesizing novel views from " + }, + { + "bbox": [ + 304, + 534, + 545, + 653 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 304, + 534, + 545, + 653 + ], + "type": "text", + "content": ". The views are generated at the camera location " + }, + { + "bbox": [ + 304, + 534, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\{p_i^v\\}_{i = 1}^N" + }, + { + "bbox": [ + 304, + 534, + 545, + 653 + ], + "type": "text", + "content": ", but rotated " + }, + { + "bbox": [ + 304, + 534, + 545, + 653 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 304, + 534, + 545, + 653 + ], + "type": "text", + "content": " on the " + }, + { + "bbox": [ + 304, + 534, + 545, + 653 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 534, + 545, + 653 + ], + "type": "text", + "content": "-axis. This is because the camera is approximately orthogonal to the touch sensor (see Fig. 3). Then, we manually annotate corresponding pixels between the touch measurements and the generated frames (Fig. 3). To simplify and standardize this process, we place a braille board in each scene and probe it with the touch sensor. This will generate a distinctive touch signal that is easy to localize [23]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "We formulate the problem of estimating the six degrees of freedom relative pose " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{R},\\mathbf{t})" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " between the touch sensor and the generated frames as a resectioning problem [20]. We use the estimated 3D structure from the NeRF " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " to obtain 3D points " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i\\}_{i = 1}^M" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " for each of the annotated corre" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26531" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "spondences. Each point has a pixel position " + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i\\in \\mathcal{R}^2" + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": " in the touch measurement. We find " + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "(\\mathbf{R},\\mathbf{t})" + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": " by minimizing the reprojection error:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 88, + 117, + 287, + 149 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 117, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 88, + 117, + 287, + 149 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {R}, \\mathbf {t}} \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\| \\pi (\\mathbf {K} [ \\mathbf {R} \\mid \\mathbf {t} ], \\mathbf {X} _ {i}) - \\mathbf {u} _ {i} \\| _ {1}, \\tag {1}", + "image_path": "3e1c028f4a56eda7d5bd515beb3c0522b83a3207dfaef8b9a4ea11762809726e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 158, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 287, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 158, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 158, + 287, + 277 + ], + "type": "text", + "content": " projects a 3D point using a given projection matrix, " + }, + { + "bbox": [ + 46, + 158, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 158, + 287, + 277 + ], + "type": "text", + "content": " are the known intrinsic of the tactile sensor's camera, and the point " + }, + { + "bbox": [ + 46, + 158, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 46, + 158, + 287, + 277 + ], + "type": "text", + "content": " is in the coordinate system of the generated vision frames. We perform the optimization on 6-15 annotated correspondences from the braille board. For robustness, we compute correspondences from multiple frames. We represent the rotation matrix using quaternions and optimize using nonlinear least-squares. Once we have " + }, + { + "bbox": [ + 46, + 158, + 287, + 277 + ], + "type": "inline_equation", + "content": "(\\mathbf{R},\\mathbf{t})" + }, + { + "bbox": [ + 46, + 158, + 287, + 277 + ], + "type": "text", + "content": " with respect to the generated frames, we can derive the relative pose between the camera and the touch sensor." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 287, + 198, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 287, + 198, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 287, + 198, + 300 + ], + "type": "text", + "content": "3.2. Imputing the missing touch" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 305, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 287, + 449 + ], + "type": "text", + "content": "We use a generative model to estimate the touch signal (represented as an image from a vision-based touch sensor) for other locations within the scene. Specifically, we train a diffusion model " + }, + { + "bbox": [ + 46, + 305, + 287, + 449 + ], + "type": "inline_equation", + "content": "p_{\\phi}(\\pmb{\\tau} \\mid \\mathbf{v}, \\mathbf{d}, \\mathbf{b})" + }, + { + "bbox": [ + 46, + 305, + 287, + 449 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 305, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 305, + 287, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 305, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 46, + 305, + 287, + 449 + ], + "type": "text", + "content": " are images and depth maps extracted from " + }, + { + "bbox": [ + 46, + 305, + 287, + 449 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 46, + 305, + 287, + 449 + ], + "type": "text", + "content": " (see Fig. 4). We also pass as input to the diffusion model a background image captured by the touch sensor when it is not in contact with anything, denoted as " + }, + { + "bbox": [ + 46, + 305, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 46, + 305, + 287, + 449 + ], + "type": "text", + "content": ". Although not essential, we have observed that this additional input empirically improves the model's performance (e.g., Fig. 1 the background provides the location of defects in the gel, which appear as black dots). We train the model " + }, + { + "bbox": [ + 46, + 305, + 287, + 449 + ], + "type": "inline_equation", + "content": "p_{\\phi}" + }, + { + "bbox": [ + 46, + 305, + 287, + 449 + ], + "type": "text", + "content": " on our entire vision-touch dataset (Sec. 4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 533 + ], + "type": "text", + "content": "The training of " + }, + { + "bbox": [ + 46, + 449, + 287, + 533 + ], + "type": "inline_equation", + "content": "p_{\\phi}" + }, + { + "bbox": [ + 46, + 449, + 287, + 533 + ], + "type": "text", + "content": " is divided into two stages. In the first, we pre-train a cross-modal visual-tactile encoder with self-supervised contrastive learning on our dataset. This stage, initially proposed by [23, 57], is equivalent to the self-supervised encoding pre-training that is common for image generation models [45]. We use a ResNet-50 [21] as the backbone for this contrastive model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "text", + "content": "In the second stage, we use the contrastive model to generate the input for a conditional latent diffusion model, which is built upon Stable Diffusion [45]. A frozen pretrained VQ-GAN [14] is used to obtain the latent representation with a spatial dimension of " + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "text", + "content": ". We start training the diffusion model from scratch and pre-train it on the task of unconditional tactile image generation on the YCB-Slide dataset [52]. After this stage, we train the conditional generative model " + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "inline_equation", + "content": "p_{\\phi}" + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "text", + "content": " on our spatially aligned visual-tactile dataset, further fine-tuning the contrastive model end-to-end with the generation task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": "At inference time, given a novel location in the 3D scene, we first render the visual signals " + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}" + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{d}}" + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": " from NeRF, and then estimate the touch signal " + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}" + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": " of the position using the diffusion model." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 305, + 72, + 545, + 185 + ], + "blocks": [ + { + "bbox": [ + 305, + 72, + 545, + 185 + ], + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 185 + ], + "type": "image", + "image_path": "90c5c0e0784f49af905258e53b93778c4f6483f3d7b65ffd45d33327ff63e174.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 194, + 545, + 238 + ], + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 238 + ], + "type": "text", + "content": "Figure 4. Touch estimation. We estimate the tactile signal for a given touch sensor pose " + }, + { + "bbox": [ + 304, + 194, + 545, + 238 + ], + "type": "inline_equation", + "content": "(\\mathbf{R},\\mathbf{t})" + }, + { + "bbox": [ + 304, + 194, + 545, + 238 + ], + "type": "text", + "content": ". To do this, we synthesize a viewpoint from the NeRF, along with a depth map. We use conditional latent diffusion to predict the tactile signal from these inputs." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 247, + 462, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 462, + 260 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 462, + 260 + ], + "type": "text", + "content": "4. A 3D Visual-Tactile Dataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 268, + 545, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 292 + ], + "type": "text", + "content": "In the following, we show the details of the data collection process and statistics of our dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 299, + 452, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 299, + 452, + 311 + ], + "spans": [ + { + "bbox": [ + 305, + 299, + 452, + 311 + ], + "type": "text", + "content": "4.1. Data Collection Procedure" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 317, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 460 + ], + "type": "text", + "content": "The data collection procedure is divided into two stages. First, we collect multiple views from the scene, capturing enough frames around the areas we plan to touch. During this stage, we collect approximately 500 frames. Next, we collect synchronized visual and touch data, maximizing the geometry and texture being touched. We then estimate the camera location of the vision frames collected in the previous two stages using off-the-shelf mapping tools [47]. After estimating the camera poses for the vision frames, the touch measurements' poses can be derived by using the mount calibration matrix. More details about the pose estimation procedure can be found in the supplement." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 461, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 556 + ], + "type": "text", + "content": "Finally, we associate each touch sensor with a color image by translating the sensor poses upwards by 0.4 meters and querying the NeRF with such poses. The field of view we use when querying the NeRF is " + }, + { + "bbox": [ + 304, + 461, + 545, + 556 + ], + "type": "inline_equation", + "content": "50^{\\circ}" + }, + { + "bbox": [ + 304, + 461, + 545, + 556 + ], + "type": "text", + "content": ". This provides us with approximately 1,500 temporally aligned vision-touch image pairs per scene. Note that this collection procedure is scalable since it does not require specific expertise or equipment and generates abundant scene-level samples." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 563, + 408, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 408, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 408, + 574 + ], + "type": "text", + "content": "4.2. Dataset Statistics" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "text", + "content": "We collect our data in 13 ordinary scenes including two offices, a workroom, a conference room, a corridor, a tabletop, a corridor, a lounge, a room with various clothes and four outdoor scenes with interesting materials. Typically, we collect 1k to 2k tactile probes in each scene, resulting in a total of 19.3k image pairs in the dataset." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Some representative samples from the collected dataset are shown in Fig. 5. Our data includes a large variety of geometry (edges, surfaces, corners, etc.) and texture (plastic, clothes, snow, wood, etc.) of different materials in the scene. During capturing process, the collector will try to" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26532" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 213, + 194 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 213, + 194 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 213, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 213, + 194 + ], + "type": "image", + "image_path": "0425ab1485baea9d4aa79b2049a478df75308f6518983cc365682f89fecd9677.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 204, + 546, + 259 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 546, + 259 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 546, + 259 + ], + "type": "text", + "content": "Figure 5. Representative examples from the captured dataset. Our dataset is obtained from nine everyday scenes, such as offices, classrooms, and kitchens. We show three such scenes in the figure above, together with samples of spatially aligned visual and tactile data. In each scene, 1k to 2k tactile probes were collected, resulting in a total of 19.3k image pairs. The data encompasses diverse geometries (edges, surfaces, corners, etc.) and textures (plastic, clothes, snow, wood, etc.) of various materials. The collector systematically probed different objects, covering areas with distinct geometry and texture using different sensor poses." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 214, + 72, + 378, + 194 + ], + "blocks": [ + { + "bbox": [ + 214, + 72, + 378, + 194 + ], + "lines": [ + { + "bbox": [ + 214, + 72, + 378, + 194 + ], + "spans": [ + { + "bbox": [ + 214, + 72, + 378, + 194 + ], + "type": "image", + "image_path": "72ef09a2a86fd21d91441c2fef904784a3aa1d8591bf70e004561ad357c0d92f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 380, + 72, + 545, + 194 + ], + "blocks": [ + { + "bbox": [ + 380, + 72, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 380, + 72, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 380, + 72, + 545, + 194 + ], + "type": "image", + "image_path": "ffab580b6af5c26b8f31f7f9ce073199f46a20eafd6a214a54dcc49186fff164.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 280, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 287, + 353 + ], + "type": "text", + "content": "thoroughly probe various objects and cover the interesting areas with more distinguishable geometry and texture with different sensor poses. To the best of our knowledge, our dataset is the first dataset that captures full, scene-scale spatially aligned vision-touch image pairs. We provide more details about the dataset in the supplement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 362, + 128, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 362, + 128, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 128, + 376 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 382, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 287, + 453 + ], + "type": "text", + "content": "Leveraging the spatially aligned image and touch pairs from our dataset, we first conduct experiments on dense touch estimation. We then show the effectiveness of both the aligned data pairs and the synthesized touch signals by conducting tactile localization and material classification as two downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 462, + 180, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 462, + 180, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 462, + 180, + 475 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": "NeRF. We use the Nerfacto method from Nerfstudio [53]. For each scene, we utilize approximately 2,000 images as training set, which thoroughly cover the scene from various view points. We train the network with a base learning rate of " + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-2}" + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": " using Adam [30] optimizer for 200,000 steps on a single NVIDIA RTX 2080 Ti GPU to achieve optimal performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 712 + ], + "type": "text", + "content": "Visual-tactile contrastive model. Following prior works [27, 57], we leverage contrastive learning methods to train a ResNet-50 [21] as visual encoder. The visual and tactile encoders share the same architecture but have different weights. We encode visual and tactile data into latent vectors in the resulting shared representation space. We set the dimension of the latent vectors to 32. Similar to CLIP [43], the model is trained on InfoNCE loss obtained from the pairwise dot products of the latent vectors. We train the model for 20 epochs by Adam [30] optimizer with a learning rate of " + }, + { + "bbox": [ + 46, + 570, + 287, + 712 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 570, + 287, + 712 + ], + "type": "text", + "content": " and batch size of 256 on 4 NVIDIA RTX 2080 Ti GPUs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 280, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 545, + 495 + ], + "type": "text", + "content": "Visual-tactile generative model. Our implementation of the diffusion model closely follows Stable Diffusion [46], with the difference that we use a ResNet-50 to generate the visual encoding from RGB-D images for conditioning. Specifically, we also add the RGB-D images rendered from the tactile sensors' poses into the conditioning, which we refer to in Sec. 5.2 as multiscale conditioning. The model is optimized for 30 epochs by Adam [30] optimizer with a base learning rate of " + }, + { + "bbox": [ + 304, + 280, + 545, + 495 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 304, + 280, + 545, + 495 + ], + "type": "text", + "content": ". The learning rate is scaled bygpu number " + }, + { + "bbox": [ + 304, + 280, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 304, + 280, + 545, + 495 + ], + "type": "text", + "content": " batch size. We train the model with batch size of 48 on 4 NVIDIA A40 GPUs. At inference time, the model conducts 200 steps of denoising process with a 7.5 guidance scale. Following prior cross-modal synthesis work [44], we use reranking to improve the prediction quality. We obtain 16 samples from the diffusion model for every instance and re-rank the samples with our pretrained contrastive model. The sample with highest similarity is the final prediction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 503, + 441, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 441, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 441, + 514 + ], + "type": "text", + "content": "5.2. Dense Touch Estimation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "Experimental setup. We now evaluate the diffusion model's ability to generate touch images. To reduce overlap between the training and test set, we first split the frames into sequences temporally (following previous work [56]). We split them into sequences of 50 touch samples, then divide these sequences into train/validation/test with a ratio of " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "8/1/1" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": ". We evaluate the generated samples on Frechet Inception Distance (FID), a standard evaluation metric for cross-modal generation [56]. We also include Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM), though we note that these metrics are highly sensitive to spatial position of the generated content, and can be optimized by models that minimize simple pixelwise losses [22]. We also include CVTP metric proposed by prior work [57], which measures the similarity between visual and tactile embeddings of a contrastive model, analogous to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26533" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 294, + 324 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 294, + 324 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 294, + 324 + ], + "type": "image", + "image_path": "ad2ef9d07b3a69c8c5c5c29ecb00ee7e8d1b2af46f64297a32a167a00d04c60d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 330, + 546, + 398 + ], + "lines": [ + { + "bbox": [ + 46, + 330, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 546, + 398 + ], + "type": "text", + "content": "Figure 6. Qualitative touch estimation results. Each model is conditioned on the RGB image and depth map rendered from the NeRF (left). The white box indicates the tactile sensor's approximate field of view (which is much smaller than the full conditional image). The G.T. column shows the ground truth touch images measured from a DIGIT sensor. L1 and VisGel often generate blurry textures and inaccurate geometry. By contrast, our model better captures the features of the tactile image, e.g., the rock's microgeometry and complex textures and shapes of furniture. The last row shows two failure cases of our model. In both examples, our model generates a touch image that is geometrically misaligned with the ground truth. All of the examples shown here are at least " + }, + { + "bbox": [ + 46, + 330, + 546, + 398 + ], + "type": "inline_equation", + "content": "10\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 330, + 546, + 398 + ], + "type": "text", + "content": " away from any training sample." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 299, + 68, + 547, + 325 + ], + "blocks": [ + { + "bbox": [ + 299, + 68, + 547, + 325 + ], + "lines": [ + { + "bbox": [ + 299, + 68, + 547, + 325 + ], + "spans": [ + { + "bbox": [ + 299, + 68, + 547, + 325 + ], + "type": "image", + "image_path": "033fc97f74dea754889694412bebf42006df17669bb7afa432a61228e20e5808.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 417, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 417, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 417, + 287, + 477 + ], + "type": "text", + "content": "CLIP [43] score. We compare against two baselines: VisGel, the approach from Li et. [35], which trains a GAN for touch generation, and " + }, + { + "bbox": [ + 46, + 417, + 287, + 477 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 46, + 417, + 287, + 477 + ], + "type": "text", + "content": ", a model with the same architecture of VisGel but trained to minimize an L1 loss in pixel space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 481, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 649 + ], + "type": "text", + "content": "Results. As is shown in Table 2, our approach performs much better on the high-level metrics, with up to " + }, + { + "bbox": [ + 46, + 481, + 287, + 649 + ], + "type": "inline_equation", + "content": "4\\mathrm{x}" + }, + { + "bbox": [ + 46, + 481, + 287, + 649 + ], + "type": "text", + "content": " lower FID and " + }, + { + "bbox": [ + 46, + 481, + 287, + 649 + ], + "type": "inline_equation", + "content": "80\\mathrm{x}" + }, + { + "bbox": [ + 46, + 481, + 287, + 649 + ], + "type": "text", + "content": " higher CVTP. This indicates that our proposed diffusion model captures the distribution and characteristics of the real tactile data more effectively. On the low-level metrics (PSNR and SSIM), all methods are comparable. In particular, the L1 model slightly outperforms the other methods since the loss it is trained on is highly correlated with low-level, pixel-wise metrics. Fig. 6 qualitatively compares samples from the different models. Indeed, our generated samples exhibit enhanced details in micro-geometry of fabrics and richer textures, including snow, wood and carpeting. However, all methods fail on fine details that are barely visible in the image, such as the tree bark." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Ablation study. We evaluate the importance of the main components of our proposed touch generation approach (Table 3). Removing the conditioning on the RGB image results in the most prominent performance drop. This is expected since RGB image uniquely determines the fine" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 306, + 415, + 545, + 473 + ], + "blocks": [ + { + "bbox": [ + 306, + 415, + 545, + 473 + ], + "lines": [ + { + "bbox": [ + 306, + 415, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 545, + 473 + ], + "type": "table", + "html": "
ModelPSNR ↑SSIM ↑FID ↓CVTP ↑
L124.340.8297.050.01
VisGel [35]23.660.81130.220.03
Ours22.840.7228.970.80
", + "image_path": "af647b6ca632eec9ee0982b7dd159a14a93593539a35f374f487bacdfde9a6b4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 475, + 545, + 521 + ], + "lines": [ + { + "bbox": [ + 305, + 475, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 475, + 545, + 521 + ], + "type": "text", + "content": "Table 2. Quantitative results on touch estimation for novel views. While comparable on low-level metrics with the baselines, our approach captures the characteristics of the real tactile data more effectively, resulting in a lower FID score." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 529, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 601 + ], + "type": "text", + "content": "grained details of a tactile image. Removing depth image or contrastive pretraining has small effect on CVTP but results in a drop on FID. Contrastive re-ranking largely improves CVTP, indicating the necessity of obtaining multiple samples from the diffusion model. We also find that multiscale conditioning provide a small benefit on FID and CVTP." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 610, + 518, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 610, + 518, + 623 + ], + "spans": [ + { + "bbox": [ + 305, + 610, + 518, + 623 + ], + "type": "text", + "content": "5.3. Downstream Task I: Tactile Localization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "text", + "content": "To help understand the quality of the captured TaRFs, we evaluate the performance of the contrastive model (used for conditioning our diffusion model) on the task of tactile localization. Given a tactile signal, our goal is to find the corresponding regions in a 2D image or in a 3D scene that are associated with it, i.e., we ask the question: what part of this image/scene feel like this? We perform the following" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26534" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 547, + 270 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 547, + 270 + ], + "type": "image", + "image_path": "e33279555accbb1569ef76ab055bae8e2f4e64e38f017072f23c0184e3ce9e0d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 274, + 547, + 340 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 547, + 340 + ], + "type": "text", + "content": "Figure 7. Tactile localization heatmaps. Given a tactile query image, the heatmap shows the image patches with a higher affinity to this tactile signal, as measured by a contrastive model trained on our dataset. We use a sliding window and compare each extracted patch with the touch signal. In each case, the center patch is the true position. Our model successfully captures the correlation between the two signals. This enables it to localize a variety of touch signals, including fine-grained geometry, e.g., a cable or a keyboard, various types of corners and edges, and large uniform regions, such as a clothing. This ability enables our diffusion model to effectively propagate sparse touch samples to other visually and structurally similar regions of the scene." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 350, + 287, + 440 + ], + "blocks": [ + { + "bbox": [ + 48, + 350, + 287, + 440 + ], + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 440 + ], + "type": "table", + "html": "
Model variationPSNR ↑SSIM ↑FID ↓CVTP ↑
Full22.840.7228.970.80
No RGB conditioning22.130.7034.310.76
No depth conditioning22.570.7133.160.80
No contrastive pretraining22.820.7132.980.79
No re-ranking22.920.7229.460.61
No multiscale23.190.7230.890.77
", + "image_path": "f93fd144e996ebca674e63ac2ce05f35b23ab1bfdaccb7b75e10370fbc73f6de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 287, + 504 + ], + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 504 + ], + "type": "text", + "content": "Table 3. Ablation study. Since the fine-grained details of touch images can be determined from a RGB image, removing conditioning on the latter results in the largest performance drops. Re-ranking has notable impact on CVTP, indicating the necessity of obtaining multiple samples from the diffusion model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 510, + 287, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 535 + ], + "type": "text", + "content": "evaluations on the test set of our dataset. Note that we run no task-specific training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": "2D Localization. To determine which part of an image are associated with a given tactile measurement, we follow the same setup of SSVTP [28]. We first split the image into patches and compute their embedding. Then, we generate the tactile embedding of the input touch image. Finally, we compute the pairwise similarities between the tactile and visual embeddings, which we plot as a heatmap. As we can see in Fig. 7, our constrastive encoder can successfully capture the correlations between the visual and tactile data. For instance, the tactile embeddings of edges are associated to edges of similar shape in the visual image. Note that the majority of tactile embeddings are highly ambiguous: all edges with a similar geometry feel the same." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "text", + "content": "3D Localization. In 3D, the association of an image to tactile measurements becomes less ambiguous. Indeed, since tactile-visual samples are rotation-dependent, objects with similar shapes but different orientations will generate different tactile measurements. Lifting the task to 3D still does not remove all ambiguities (for example, each side of a rectangular table cannot be precisely localized). Nonetheless, we believe it to be a good fit for a quantitative evaluation since it's rare for two ambiguous parts of the scene to be touched with exactly the same orientation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "text", + "content": "We use the following experimental setup for 3D localization. Given a tactile image as a query, we compute its distance in embedding space to all visual test images from the same scene. Note that all test images are associated with a 3D location. We define as ground-truth correspondences all test images at a distance of at most " + }, + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "text", + "content": " from the 3D location of the test sample. We vary " + }, + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "text", + "content": " to account for local ambiguities. As typical in the retrieval literature, we benchmark the performance with metric mean Average Precision (mAP)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "We consider three baselines: (1) chance, which randomly selects corresponding samples; (2) real, which uses the contrastive model trained on our dataset; and (3) real + estimated, which trains the contrastive model on both dataset samples and a set of synthetic samples generated via the scenes' NeRF and our touch generation model. Specifically, we render a new image and corresponding touch by interpolating the position of two consecutive frames in the training dataset. This results in a training dataset for the contrastive model that is twice as large." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26535" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 288, + 144 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 144 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 144 + ], + "type": "table", + "html": "
Datasetr(m)
0.0010.0050.010.050.1
Chance3.556.8210.2518.2621.33
Real12.1022.9332.1050.3057.15
Real + Est.14.9226.6936.1753.6260.61
", + "image_path": "7d1fea4ad270197186abc29fa6700fc6f2927437a3b72b4a18f4b1f2d29f6fe9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 218, + 287, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 287, + 301 + ], + "type": "text", + "content": "The results, presented in Table 4, demonstrate the performance benefit of employing both real and synthetic tactile pairs. Combining synthetic tactile images with the original pairs achieves highest performance on all distance thresholds. Overall, this indicates that touch measurements from novel views are not only qualitatively accurate, but also beneficial for this downstream task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 311, + 278, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 278, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 278, + 323 + ], + "type": "text", + "content": "5.4. Downstream Task II: Material Classification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 329, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 437 + ], + "type": "text", + "content": "We investigate the efficacy of our visual-tactile dataset for understanding material properties, focusing on the task of material classification. We follow the formulation by Yang et al. [56], which consists of three subtasks: (i) material classification, requiring the distinction of materials among 20 possible classes; (ii) softness classification, a binary problem dividing materials as either hard or soft; and (iii) hardness classification, which requires the classification of materials as either rough or smooth." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 593 + ], + "type": "text", + "content": "We follow the same experimental procedure of [56]: we pretrain a contrastive model on a dataset and perform linear probing on the sub-tasks' training set. Our experiments only vary the pretraining dataset, leaving all architectural choices and hyperparameters the same. We compare against four baselines. A random classifier (chance); the ObjectFolder 2.0 dataset [17]; the VisGel dataset [35]; and the Touch and Go dataset [56]. Note that the touch sensor used in the test data (GelSight) differs from the one used in our dataset (DIGIT). Therefore, we use for pretraining a combination of our dataset and Touch and Go. To ensure a fair comparison, we also compare to the combination of each dataset and Touch and Go." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "The findings from this evaluation, as shown in Table 5, suggest that our data improves the effectiveness of the contrastive pretraining objective, even though our data is from a different distribution. Moreover, we find that adding estimated touch probes for pretraining results in a higher performance on all the three tasks, especially the smoothness classification. This indicates that not only does our dataset covers a wide range of materials but also our diffusion model captures the distinguishable and useful patterns of different materials." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 547, + 198 + ], + "blocks": [ + { + "bbox": [ + 46, + 148, + 287, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 148, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 287, + 205 + ], + "type": "text", + "content": "Table 4. Quantitative results on 3D tactile localization. We evaluate using mean Average Precision (mAP) as a metric. Training the contrastive model on our dataset of visually aligned real samples together with estimated samples from new locations in the scene results in the highest performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 70, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 547, + 198 + ], + "type": "table", + "html": "
DatasetMaterialHard/ SoftRough/ Smooth
Chance18.666.156.3
ObjectFolder 2.0 [17]36.272.069.0
VisGel [35]39.169.470.4
Touch and Go [56]54.777.379.4
+ ObjectFolder 2.0 [17]54.687.384.8
+ VisGel [35]53.186.783.6
+ Ours* (Real)57.688.481.7
+ Ours* (Real + Estimated)59.088.786.1
", + "image_path": "e87791c10c19195a6d381ea19fd909b3eadcfe79b7e601c059dd87b16e5610e2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 202, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 304, + 202, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 547, + 281 + ], + "type": "text", + "content": "Table 5. Material classification. We show the downstream material recognition accuracy of models pre-trained on different datasets. The final rows show the performance when combining different datasets with Touch and Go [56]. * The task-specific training and testing datasets for this task are collected with a Gel-Sight sensor. We note that our data comes from a different distribution, since it is collected with a DIGIT sensor [32]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 305, + 379, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 305, + 379, + 317 + ], + "spans": [ + { + "bbox": [ + 306, + 305, + 379, + 317 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 333, + 547, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 547, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 547, + 500 + ], + "type": "text", + "content": "In this work, we present the TaRF, a scene representation that brings vision and touch into a shared 3D space. This representation enables the generation of touch probes for novel scene locations. To build this representation, we collect the largest dataset of spatially aligned vision and touch probes. We study the utility of both the representation and the dataset in a series of qualitative and quantitative experiments and on two downstream tasks: 3D touch localization and material recognition. Overall, our work makes the first step towards giving current scene representation techniques an understanding of not only how things look, but also how they feel. This capability could be critical in several applications ranging from robotics to the creation of virtual worlds that look and feel like the real world." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 511, + 547, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 511, + 547, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 547, + 619 + ], + "type": "text", + "content": "Limitations. Since the touch sensor is based on a highly zoomed-in camera, small (centimeter-scale) errors in SfM or visual-tactile registration can lead to misalignments of several pixels between the views of the NeRF and the touch samples, which can be seen in our TaRFs. Another limitation of the proposed representation is the assumption that the scene's coarse-scale structure does not change when it is touched, an assumption that may be violated for some inelastic surfaces." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Acknowledgements. We thank Jeongsoo Park, Ayush Shrivastava, Daniel Geng, Ziyang Chen, Zihao Wei, Zixuan Pan, Chao Feng, Chris Rockwell, Gaurav Kaul and the reviewers for the valuable discussion and feedback. This work was supported by an NSF CAREER Award #2339071, a Sony Research Award, the DARPA Machine Common Sense program, and ONR MURI award N00014-21-1-2801." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26536" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 212 + ], + "type": "text", + "content": "[2] Peter Anderson, Ayush Shrivastava, Joanne Truong, Arjun Majumdar, Devi Parikh, Dhruv Batra, and Stefan Lee. Sim-to-real transfer for vision-and-language navigation. In Conference on Robot Learning, pages 671-681. PMLR, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 215, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 215, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 287, + 270 + ], + "type": "text", + "content": "[3] Jakub Bednarek, Michal Bednarek, Lorenz Wellhausen, Marco Hutter, and Krzysztof Walas. What am i touching? learning to classify terrain via haptic sensing. In 2019 International Conference on Robotics and Automation (ICRA), pages 7187-7193. IEEE, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "type": "text", + "content": "[4] Katherine L Bouman, Bei Xiao, Peter Battaglia, and William T Freeman. Estimating the material properties of fabric from video. In Proceedings of the IEEE international conference on computer vision, pages 1984-1991, 2013. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 318, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 339 + ], + "type": "text", + "content": "[5] Alexander Burka. Instrumentation, data, and algorithms for visually understanding haptic surface properties. 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 342, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 287, + 396 + ], + "type": "text", + "content": "[6] Roberto Calandra, Andrew Owens, Manu Upadhyaya, Wenzhen Yuan, Justin Lin, Edward H. Adelson, and Sergey Levine. The feeling of success: Does touch sensing help predict grasp outcomes? Conference on Robot Learning (CoRL), 2017. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 398, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 398, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 398, + 287, + 452 + ], + "type": "text", + "content": "[7] Roberto Calandra, Andrew Owens, Dinesh Jayaraman, Justin Lin, Wenzhen Yuan, Jitendra Malik, Edward H. Adelson, and Sergey Levine. More than a feeling: Learning to grasp and regrasp using vision and touch. Robotics and Automation Letters (RA-L), 2018. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "type": "text", + "content": "[8] Roberto Calandra, Andrew Owens, Dinesh Jayaraman, Justin Lin, Wenzhen Yuan, Jitendra Malik, Edward H. Adelson, and Sergey Levine. More than a feeling: Learning to grasp and regrasp using vision and touch. IEEE Robotics and Automation Letters, 3:3300-3307, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 511, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 511, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 53, + 511, + 287, + 566 + ], + "type": "text", + "content": "[9] Berk Calli, Arjun Singh, Aaron Walsman, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. The ycb object and model set: Towards common benchmarks for manipulation research. In 2015 international conference on advanced robotics (ICAR), pages 510-517. IEEE, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "text", + "content": "[10] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European conference on computer vision, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[11] Alex Church, John Lloyd, Raia Hadsell, and Nathan F Lepora. Deep reinforcement learning for tactile robotics: Learning to type on a braille keyboard. IEEE Robotics and Automation Letters, 5(4):6145-6152, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[13] Abe Davis, Justin G Chen, and Frédo Durand. Image-space modal bases for plausible manipulation of objects in video. ACM Transactions on Graphics (TOG), 2015. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[14] Mucong Ding, Kezhi Kong, Jingling Li, Chen Zhu, John Dickerson, Furong Huang, and Tom Goldstein. Vq-gnn: A universal framework to scale up graph neural networks using vector quantization. Advances in Neural Information Processing Systems, 34:6733-6746, 2021. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 162, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 204 + ], + "type": "text", + "content": "[15] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence, 32(8):1362-1376, 2009. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 206, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 248 + ], + "type": "text", + "content": "[16] Ruohan Gao, Yen-Yu Chang, Shivani Mall, Li Fei-Fei, and Jiajun Wu. Objectfolder: A dataset of objects with implicit visual, auditory, and tactile representations. In CoRL, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 250, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 315 + ], + "type": "text", + "content": "[17] Ruohan Gao, Zilin Si, Yen-Yu Chang, Samuel Clarke, Jeannette Bohg, Li Fei-Fei, Wenzhen Yuan, and Jiajun Wu. Objectfolder 2.0: A multisensory object dataset for sim2real transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10598–10608, 2022. 1, 2, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 316, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 380 + ], + "type": "text", + "content": "[18] Ruohan Gao, Yiming Dou, Hao Li, Tanmay Agarwal, Jeannette Bohg, Yunzhu Li, Li Fei-Fei, and Jiajun Wu. The objectfolder benchmark: Multisensory learning with neural and real objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17276–17286, 2023. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "text", + "content": "[19] Michael SA Graziano and Charles G Gross. The representation of extrapersonal space: A possible role for bimodal, visual-tactile neurons. 1995. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 416, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 447 + ], + "type": "text", + "content": "[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 494, + 545, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 494, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 308, + 494, + 545, + 546 + ], + "type": "text", + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 548, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 590 + ], + "type": "text", + "content": "[23] Carolina Higuera, Byron Boots, and Mustafa Mukadam. Learning to read braille: Bridging the tactile reality gap with diffusion models. arXiv preprint arXiv:2304.01182, 2023. 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 593, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 635 + ], + "type": "text", + "content": "[24] Mark A Hoepflinger, C David Remy, Marco Hutter, Luciano Spinello, and Roland Siegwart. Haptic terrain classification for legged robots. In 2010 IEEE International Conference on Robotics and Automation, pages 2828-2833. IEEE, 2010. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 637, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 637, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 637, + 545, + 680 + ], + "type": "text", + "content": "[25] Micah K Johnson and Edward H Adelson. Retrographic sensing for the measurement of surface texture and shape. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 1070-1077. IEEE, 2009. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[26] Micah K Johnson, Forrester Cole, Alvin Raj, and Edward H Adelson. Microgeometry capture using an elastomeric sensor. ACM Transactions on Graphics (TOG), 2011. 2, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26537" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[27] Justin Kerr, Huang Huang, Albert Wilcox, Ryan Hoque, Jeffrey Ichnowski, Roberto Calandra, and Ken Goldberg. Learning self-supervised representations from vision and touch for active sliding perception of deformable surfaces. arXiv preprint arXiv:2209.13042, 2022. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "type": "text", + "content": "[28] Justin Kerr, Huang Huang, Albert Wilcox, Ryan Hoque, Jeffrey Ichnowski, Roberto Calandra, and Ken Goldberg. Self-supervised visuo-tactile pretraining to locate and follow garment features. In Robotics: Science and Systems, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 229 + ], + "type": "text", + "content": "[29] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 253 + ], + "type": "text", + "content": "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. *ICLR*, 2015. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 288, + 297 + ], + "type": "text", + "content": "[31] Hendrik Kolvenbach, Christian Bärtschi, Lorenz Wellhausen, Ruben Grandia, and Marco Hutter. Haptic inspection of planetary soils with legged robots. IEEE Robotics and Automation Letters, 4(2):1626-1632, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 364 + ], + "type": "text", + "content": "[32] Mike Lambeta, Po-Wei Chou, Stephen Tian, Brian Yang, Benjamin Maloon, Victoria Rose Most, Dave Stroud, Raymond Santos, Ahmad Byagowi, Gregg Kammerer, et al. Digit: A novel design for a low-cost compact high-resolution tactile sensor with application to in-hand manipulation. IEEE Robotics and Automation Letters, 2020. 1, 2, 3, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "type": "text", + "content": "[33] Maxime Lhuillier and Long Quan. A quasi-dense approach to surface reconstruction from uncalibrated images. IEEE transactions on pattern analysis and machine intelligence, 27(3):418-433, 2005. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "type": "text", + "content": "[34] Hongyu Li, Snehal Dikhale, Soshi Iba, and Nawid Jamali. Vihope: Visuotactile in-hand object 6d pose estimation with shape completion. IEEE Robotics and Automation Letters, 8 (11):6963-6970, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "text", + "content": "[35] Yunzhu Li, Jun-Yan Zhu, Russ Tedrake, and Antonio Torralba. Connecting touch and vision via cross-modal prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10609-10618, 2019. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "text", + "content": "[36] Justin Lin, Roberto Calandra, and Sergey Levine. Learning to identify object instances by touch: Tactile recognition via multimodal matching. In 2019 International Conference on Robotics and Automation (ICRA), pages 3644-3650. IEEE, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "type": "text", + "content": "[37] Antonio Loquercio, Ashish Kumar, and Jitendra Malik. Learning visual locomotion with cross-modal supervision. In 2023 IEEE International Conference on Robotics and Automation (ICRA), 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "type": "text", + "content": "[38] Gabriel B Margolis, Xiang Fu, Yandong Ji, and Pulkit Agrawal. Learning physically grounded robot vision with active sensing motor policies. In 7th Annual Conference on Robot Learning, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "text", + "content": "[40] Andrew Owens, Phillip Isola, Josh McDermott, Antonio Torralba, Edward H Adelson, and William T Freeman. Visually indicated sounds. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2405-2413, 2016. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 546, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 172 + ], + "type": "text", + "content": "[41] Senthil Purushwalkam, Abhinav Gupta, Danny M Kaufman, and Bryan Russell. Bounce and learn: Modeling scene dynamics with real-world bounces. arXiv preprint arXiv:1904.06827, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 174, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 546, + 217 + ], + "type": "text", + "content": "[42] Haozhi Qi, Brent Yi, Sudharshan Suresh, Mike Lambeta, Yi Ma, Roberto Calandra, and Jitendra Malik. General in-hand object rotation with vision and touch. arXiv preprint arXiv:2309.09979, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 546, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 546, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 546, + 274 + ], + "type": "text", + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 276, + 546, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 546, + 328 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 546, + 328 + ], + "type": "text", + "content": "[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 331, + 546, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 546, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 546, + 364 + ], + "type": "text", + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 365, + 546, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 546, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 546, + 421 + ], + "type": "text", + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 422, + 546, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 546, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 546, + 454 + ], + "type": "text", + "content": "[47] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 456, + 546, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 546, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 546, + 487 + ], + "type": "text", + "content": "[48] Carmelo Sferrazza and Raffaello D'Andrea. Design, motivation and evaluation of a full-resolution optical tactile sensor. Sensors, 19(4):928, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 489, + 546, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 546, + 521 + ], + "type": "text", + "content": "[49] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV 2012, ECCV 2012. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 523, + 546, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 546, + 555 + ], + "type": "text", + "content": "[50] Linda Smith and Michael Gasser. The development of embodied cognition: Six lessons from babies. Artificial life, 2005. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 557, + 546, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 546, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 546, + 610 + ], + "type": "text", + "content": "[51] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 613, + 546, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 546, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 546, + 666 + ], + "type": "text", + "content": "[52] Sudharshan Suresh, Zilin Si, Stuart Anderson, Michael Kaess, and Mustafa Mukadam. Midastouch: Monte-carlo inference over distributions across sliding touch. In Conference on Robot Learning, pages 319-331. PMLR, 2023. 1, 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 669, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 546, + 713 + ], + "type": "text", + "content": "[53] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David Mcallister, Justin Kerr, and Angjoo Kanazawa. Nerfstudio: A modu" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26538" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 609 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "lar framework for neural radiance field development. arXiv preprint arXiv:2302.04264, 2023. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "text", + "content": "[54] Shaoxiong Wang, Jiajun Wu, Xingyuan Sun, Wenzhen Yuan, William T Freeman, Joshua B Tenenbaum, and Edward H Adelson. 3d shape perception from monocular vision, touch, and shape priors. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "text", + "content": "[55] Jianxiong Xiao, Andrew Owens, and Antonio Torralba. Sun3d: A database of big spaces reconstructed using sfm and object labels. 2013. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "text", + "content": "[56] Fengyu Yang, Chenyang Ma, Jiacheng Zhang, Jing Zhu, Wenzhen Yuan, and Andrew Owens. Touch and go: Learning from human-collected vision and touch. Neural Information Processing Systems (NeurIPS) - Datasets and Benchmarks Track, 2022. 2, 3, 5, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "text", + "content": "[57] Fengyu Yang, Jiacheng Zhang, and Andrew Owens. Generating visual scenes from touch. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22070-22080, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 352 + ], + "type": "text", + "content": "[58] Fengyu Yang, Chao Feng, Ziyang Chen, Hyoungseob Park, Daniel Wang, Yiming Dou, Ziyao Zeng, Xien Chen, Rit Gangopadhyay, Andrew Owens, and Alex Wong. Binding touch to everything: Learning unified multimodal tactile representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 353, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 397 + ], + "type": "text", + "content": "[59] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "type": "text", + "content": "[60] Zhao-Heng Yin, Binghao Huang, Yuzhe Qin, Qifeng Chen, and Xiaolong Wang. Rotating without seeing: Towards in-hand dexterity through touch. arXiv preprint arXiv:2303.10880, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 443, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 443, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 443, + 287, + 497 + ], + "type": "text", + "content": "[61] Wenzhen Yuan, Shaoxiong Wang, Siyuan Dong, and Edward Adelson. Connecting look and feel: Associating the visual and tactile properties of physical materials. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5580-5588, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 498, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 553 + ], + "type": "text", + "content": "[62] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 554, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 287, + 609 + ], + "type": "text", + "content": "[63] Shaohong Zhong, Alessandro Albini, Oiwi Parker Jones, Perla Maiolino, and Ingmar Posner. Touching a nerf: Leveraging neural radiance fields for tactile sensory data generation. In Conference on Robot Learning, pages 1618-1628. PMLR, 2023. 1, 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26539" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/7aad987a-39f2-4096-8aa1-a539bbaacb9c_content_list.json b/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/7aad987a-39f2-4096-8aa1-a539bbaacb9c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5f1d7df5b1c4ce9115c41ae22a93d44cd2f2a19d --- /dev/null +++ b/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/7aad987a-39f2-4096-8aa1-a539bbaacb9c_content_list.json @@ -0,0 +1,1943 @@ +[ + { + "type": "text", + "text": "Tailored Visions: Enhancing Text-to-Image Generation with Personalized Prompt Rewriting", + "text_level": 1, + "bbox": [ + 116, + 128, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zijie Chen*,1,2 Lichao Zhang*,2 Fangsheng Weng3 Lili Pan†,4 Zhenzhong Lan†,2", + "bbox": [ + 151, + 200, + 841, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Zhejiang University $^{2}$ Westlake University $^{3}$ Scietrain", + "bbox": [ + 264, + 227, + 728, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4University of Electronic Science and Technology of China", + "bbox": [ + 264, + 246, + 730, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chenzijie, zhanglichao, lanzhenzhong}@westlake.edu.cn, lilipan@uestc.edu.cn", + "bbox": [ + 153, + 266, + 834, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 315, + 313, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite significant progress in the field, it is still challenging to create personalized visual representations that align closely with the desires and preferences of individual users. This process requires users to articulate their ideas in words that are both comprehensible to the models and accurately capture their vision, posing difficulties for many users. In this paper, we tackle this challenge by leveraging historical user interactions with the system to enhance user prompts. We propose a novel approach that involves rewriting user prompts based on a newly collected large-scale text-to-image dataset with over 300k prompts from 3115 users. Our rewriting model enhances the expressiveness and alignment of user prompts with their intended visual outputs. Experimental results demonstrate the superiority of our methods over baseline approaches, as evidenced in our new offline evaluation method and online tests. Our code and dataset are available at https://github.com/zzjchen/Tailored-Visions", + "bbox": [ + 75, + 348, + 473, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 651, + 209, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Increasingly large and powerful foundation models [1, 4, 20] are trained through self-supervised learning. These large pretrained models (LPMs) serve as efficient compressors [3], condensing vast amounts of internet data. This compression enables the convenient extraction of the knowledge encoded within these models via natural language descriptions. Despite being in its infancy, this approach exhibits potentials to surpass traditional search engines as a superior source for knowledge and information acquisition.", + "bbox": [ + 75, + 676, + 468, + 827 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Akin to refining queries for search engines, prompts given to LPMs must also be carefully crafted. However,", + "bbox": [ + 76, + 829, + 468, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "User History", + "text_level": 1, + "bbox": [ + 549, + 316, + 609, + 328 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0c29a6ba42cd59e518f9ac1365305c8bd34906d820f1116331ceb3d1bfa4d039.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 343, + 563, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Italianate House with Beautiful Breathing on the Shores of Lake Como-Oil Painting in the Style of Claude Monet trending in artstation", + "bbox": [ + 566, + 347, + 653, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/10570ac38f7617b9c4a0994a2766e3fe4276a1ea9fe548e83013a5277b28f9e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 391, + 563, + 436 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A Magical Forest in the morning, oil painting in the style of Claude Monet trending in artstation.", + "bbox": [ + 566, + 401, + 653, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8380a50d87796cb66afde073133bb1fec76ccf743209e413d9d3129e2b856bb3.jpg", + "image_caption": [ + "Figure 1. Comparison between our personalized prompt rewriting method and the standard prompt rewriting method. Our technique excels at incorporating user preferences, such as \"oil paintings by artists,\" while methods that lack a historical context frequently generate content that may not align with the user's desires." + ], + "image_footnote": [], + "bbox": [ + 506, + 438, + 565, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tulip bouquet. Oil Painting, bright color, highly detailed, Vincent Van Gogh, trending in artstation", + "bbox": [ + 566, + 448, + 653, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/38706e5319762fe75a8b8a748a09d4d3f7f651e880a7b9d017df2d10e4fe1cef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 489, + 565, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Still life of a hamburger and some fruits on a napkin, detailed oil painting in the style of Paul Cezanme, trenging in artstation", + "bbox": [ + 566, + 492, + 653, + 525 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generated Results", + "text_level": 1, + "bbox": [ + 727, + 318, + 810, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Original Prompt: Walking street at night", + "bbox": [ + 686, + 330, + 743, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/73eae4661ebe7256dc79bff46275a5f0d7b28713abbde65dd5a0d28a81923a3c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 344, + 754, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4a0599bb42bdcdc2fcd27370ad9bfe24218cbb225a153e8b9698b529de7daf36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 330, + 874, + 402 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "t... a mesmerizing walking street at night...with a touch of Italianate c-harm, reminiscent of Claude Monet's oil painting style.", + "bbox": [ + 661, + 402, + 769, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/82dc172a8227b2aae5153f43cc4fe88ca96bf3f9d4ef04ae2a26072561e5bf2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 428, + 753, + 486 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "et an astronaut, inspired by Claude tce cMonet_rides a horse in a captivati ng scene,..., fusion of magic and natury...", + "bbox": [ + 753, + 402, + 879, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/00d34ff0a9222c9fc756649c2f67cf6d26e16eaf3252350fe05784d545d3a372.jpg", + "image_caption": [ + "W/o History" + ], + "image_footnote": [], + "bbox": [ + 790, + 428, + 864, + 487 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "An astronaut, adorned in futuristic gear, gracefully maneuvers on horseback through an otherwise landscape bathed in the ethereal glow of distant stars, capturing the juxtaposition of advanced space exploration and timeless equestrian elegance.", + "bbox": [ + 774, + 486, + 879, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the complexity of prompts, the unpredictability of model responses compared to traditional search engines present unique challenges. Significant research efforts [10, 29] have been made to comprehend how LPMs react to various prompts, with some studies examining the feasibility of rewriting prompts for specificity. However, without access to users' personal data and behavior, tailoring the prompt to meet the user's needs accurately remains challenging.", + "bbox": [ + 496, + 657, + 893, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our research addresses this issue by integrating user preference information into prompt rewriting. The primary obstacle in personalized query rewriting is the absence of a dataset containing text-to-image prompts with personalized information. To overcome this, we have assembled a large dataset encompassing over 300k text-to-image histories from 3,115 users. We rewrite user prompts using their query history, although we had limited access to personal", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 808, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 94, + 871, + 207, + 885 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 98, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7727", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "information, leaving room for further research. Another significant challenge is the evaluation of rewritten queries. To evaluate their efficacy, we've developed a new offline method that uses multiple metrics to measure how well our rewriting models can recover the original user query from the ChatGPT-shortened version.", + "bbox": [ + 75, + 90, + 468, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our paper's contributions are threefold:", + "bbox": [ + 96, + 181, + 359, + 196 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We have compiled a large Personalized Image Prompt (PIP) dataset and made it publicly available to aid future research in the field.", + "2. We experimented with two query rewriting techniques and proposed a new query evaluation method to assess their performance.", + "3. We propose a new benchmark for personalized text-to-image generation, which promotes the standardization of this field." + ], + "bbox": [ + 76, + 196, + 468, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While there is still a considerable distance to cover before we can create a perfect prompt encapsulating both the user's requirements and the model's capabilities, we believe our research provides a critical stepping stone in this ongoing exploration.", + "bbox": [ + 75, + 332, + 468, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 416, + 218, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section provides an overview of prior work on text-to-image generation, personalization for such generation, and prompt rewriting. It's important to note that our review is aimed more at offering sufficient background knowledge rather than exhaustive coverage of all related works.", + "bbox": [ + 75, + 435, + 468, + 510 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 76, + 518, + 313, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large text-to-image generation models can generate high-fidelity image synthesis and achieve a deep level of language understanding. DALL-E [17] uses a VQ-VAE transformer-based method to learn a visual codebook in the first stage and then trains autoregressive transformers on sequences of text tokens followed by image tokens in the second stage. DALL-E2 [18] introduces latent diffusion models to generate various images by conditioning on CLIP text latents and CLIP image embeddings generated by a prior model. Imagen [22] discovers that a larger language model with more parameters trained on text-only data improves the quality of text-to-image generation. Late developments like stable-diffusion (SD) [20] proposes to generate images effectively in latent space significantly lowering computational costs. Furthermore, SD designs a conditional mechanism to complete class-conditional, text-to-image and layout-to-image models. Furthermore, ControlNet [28] accomplishes certain function by conditioning on multi-modal data, e.g., edge, sketching, pose, segmentation, depth etc., which unavoidably involving additional condition-generation modalities.", + "bbox": [ + 75, + 537, + 468, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these models' ability to generate high-fidelity images, they often fail to meet the precise needs of the users. Text-to-image generation is more like a game of chance.", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Personalization for Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 498, + 90, + 883, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, personalization approaches based on text-to-image models have taken a set of images of a concept and generated variations of the concept. Specifically, some methods optimize a set of text embeddings. For example, [2] involves pseudo-word embeddings by a set encoder to provide personalization and Textual inversion [4] composes the concept into language sentences and performed as a personalized creation. Some methods finetune the diffusion model. For example, DreamBooth [21] finetunes the text-to-image diffusion model with shared parallel branches. To speed up, CustomDiffusion [12] reduces the amount of finetuned parameters, and Tewel et al. [25] locks the subject's cross-attention key to its superordinate category to align with visual concepts. Moreover, an additional encoder is trained to map concept images to its textual representation by [5] and [24].", + "bbox": [ + 496, + 109, + 890, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing studies have three key limitations: they demand extra images and fine-tuning of text-to-image models with limited scope for new concepts; they can't learn from user interaction history and need detailed user prompts; and there's a lack of public, personalized text-to-image datasets that truly reflect user preferences.", + "bbox": [ + 496, + 352, + 890, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Prompt Rewriting", + "text_level": 1, + "bbox": [ + 498, + 449, + 676, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, researchers have found that optimizing prompts can boost the performance of LLMs on several NLP tasks and even search systems. For examples, Guo et al. [6] connect the LLM with evolutionary algorithms to generate an optimized prompt from parent prompts, without any gradient calculation. Yang et al. [27] propose to use LLM as an optimizer by generating new prompts based on a trajectory of previously generated prompts in each optimization step with the objective of maximizing the accuracy of the task. In the work [30], LLMs serve as models for engineering work like inference, scoring, and resampling. In search systems, LLMs are used to generate query expansion terms by [10], while they are used to reformulate query by Wang et al. [26] instead.", + "bbox": [ + 496, + 469, + 890, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For T2I generation, one relatively close work, SUR-adapter [29] learns to align the representations between simple prompt and complex prompt by an adapter. [7] optimize prompt through general rewriting. However, Neither of above works utilized personalized information for improving prompts.", + "bbox": [ + 496, + 680, + 890, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Personalized Image-Prompt (PIP) Dataset", + "text_level": 1, + "bbox": [ + 498, + 782, + 874, + 799 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Dataset Collection", + "text_level": 1, + "bbox": [ + 500, + 806, + 676, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Personalized Image-Prompt (PIP) dataset is the first large-scale personalized generated image-text dataset. The original data are collected from a public website that we", + "bbox": [ + 496, + 830, + 890, + 877 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ https://zmrj.art/", + "bbox": [ + 517, + 885, + 609, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7728", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b53d3f0ace518c1792c225648fb3133c3e9b4ff4a03ff39e3ee74274fc8120f6.jpg", + "image_caption": [ + "Figure 2. Dataset creation process. We split our dataset into training and testing sets and summarize each prompts in the test set using ChatGPT." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 467, + 268 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/90524e7cc1928da278f0c921c62a80a62038caa90a31f265e961d6ddc427daa0.jpg", + "image_caption": [ + "Figure 3. Dataset statistics and distribution. Left: Proportion of users based on the varying number of historical prompts they have. Note that each user has a minimum of 18 historical prompts, as we have excluded those with fewer prompts from the dataset. Right: Proportion of prompts based on their varying lengths. Best view in color." + ], + "image_footnote": [], + "bbox": [ + 84, + 348, + 465, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "host to provide open-domain text-to-image generation to users. PIP dataset includes 3115 users and 300,237 text-to-image histories generated by these users using SD v1-5 [20] and an internal fine-tuned version of SD v1-5. Each user in PIP has created 18 or more images and provided at least 12 different prompts.", + "bbox": [ + 75, + 628, + 468, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 illustrates the process of creating the dataset. For each individual user, we randomly choose two prompts to serve as test prompts, with the remaining prompts allocated as training prompts (historical user query). The purpose of using random selection instead of the most recent generated prompts is to enhance the diversity of our test data. Subsequently, we employ ChatGPT to condense the test prompts, ensuring they only include the primary object or scene, as depicted in Figure 2. We shorten the prompts into three scales, i.e., contain only nouns, noun phases or short sentences respectively.", + "bbox": [ + 75, + 719, + 470, + 885 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the ensuing experiment, each test prompt in the test", + "bbox": [ + 96, + 885, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "set will be considered as the input prompt $x_{t}$ for every user $u$ , with the original prompts serving as the ground truth that reflect the user's authentic preferences. The remaining prompts are utilized as training samples.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The PIP dataset consists of 300,237 image-prompt pairs, personally categorized by 3,115 users. These pairs are divided into 294,007 training samples and 6,230 test samples.", + "bbox": [ + 496, + 152, + 890, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Dataset Statistics and Distribution", + "text_level": 1, + "bbox": [ + 500, + 205, + 799, + 219 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we showcase the data statistics that depict the quality and diversity of PIP. We specifically illustrate data distributions of the number of prompts and prompt length for each user, and delve deeper into the content of the prompt through a word cloud representation.", + "bbox": [ + 496, + 227, + 890, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each data sample contains a prompt, the generated images,UserID, Image size, and URL, as illustrated Figure 4.", + "bbox": [ + 498, + 303, + 890, + 334 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/43077457f808c0e52fea4fda2ab5a8692fd0a2fb55acddd97f10219425179767.jpg", + "image_caption": [ + "Figure 4. Two examples from a user history, containing Image, Prompt, User ID, Image size and URL." + ], + "image_footnote": [], + "bbox": [ + 503, + 348, + 883, + 502 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bd813f8f9f0a153cb777a4c569c6f59f29379514ba524dddfae443e8c807904e.jpg", + "image_caption": [ + "Figure 5. Word cloud visualization of top 250 keywords sampled from the PIP dataset." + ], + "image_footnote": [], + "bbox": [ + 504, + 568, + 887, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In PIP dataset, each user contributes at least 18 images, as depicted in the left section of Figure 3. This results in a long-tail distribution. The prompts have an average word count of 27.53. The length of the prompts, ranging from 1 to 284 words, also follows a long-tail distribution, as seen in the right section of Figure 3. Despite the presence of about 2500 prompts that exceed the 75-word limit of SD, we retain them to maintain the integrity of user preferences.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7729", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dbd0a4e9b0786cf271544e01d273a106606844b5a3aa170e58345b53f3985e9c.jpg", + "image_caption": [ + "Figure 6. Text similarity between user histories and user preferences. There is a great deal of diversity in user preferences in the PIP dataset. In the zoomed-in version of the similarity map on the left, users' history prompts are highly similar to their preferences." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 467, + 218 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 5 presents the 250 most frequently used words or phrases. The frequency of these words is determined by the highest TF-IDF value across all users. The word cloud reveals that these words describe various image attributes, such as objects, styles, quality, and colors. This variety underlines the high diversity present within the prompt content of the PIP dataset.", + "bbox": [ + 75, + 311, + 468, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, we visualize the text similarities between the histories and preferences of 100 users in Figure 6. For each user $u$ in PIP dataset, we summarize his preference $P_{u}$ into 5 phrases from his history prompts using ChatGPT. For two users $u, v$ , the similarity of $u$ 's histories and $v$ 's preferences is defined as the mean of text similarities between $u$ 's history prompts and $v$ 's preference $P_{v}$ . The text similarity is calculated using GTR-T5-large [14]. Furthermore, we visualize the text similarities between the history prompts of a random user and 100 user preferences in the left part of Figure 6. This shows that text-to-image users have different preferences and the preference $p_{u}$ we summarized successfully captures the key feature the user prefers.", + "bbox": [ + 75, + 419, + 472, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 628, + 261, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present two metrics to evaluate prompt rewriting methods in terms of how the rewritten results are aligned to users' preferences, namely Preference Matching Score (PMS) and Image-Align.", + "bbox": [ + 75, + 654, + 468, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preference Matching Score (PMS). PMS calculates the CLIPScore [8] between generated image and user's preference $P_{u}$ . It measures how the generated image aligns with the user's preference.", + "bbox": [ + 75, + 715, + 468, + 776 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P M S} = \\frac {w}{N} \\sum_ {u = 1} ^ {N} \\max \\left(\\cos \\left(\\mathbf {E m} \\left(I _ {u} ^ {\\prime}\\right), \\mathbf {E m} \\left(P _ {u}\\right)\\right), 0\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 785, + 468, + 828 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{u}$ is the user $u$ 's preference, $I_{u}^{\\prime}$ is the generated image correspondingly, $N$ is total user number (i.e. 3115). $\\mathbf{Em}$ means the embedding extracted by using CLIP. $w = 2.5$ is a scaling constant.", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image-Align. It measures the similarity between the generated image and the ground-truth image. Image-Align quantifies how closely the current created image aligns with the user's truly saved image. The similarity between two images are calculated using CLIP [16].", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Apart from these metrics, we also adopt ROUGE-L to evaluate prompt rewriting methods in our experiment. Calculating ROUGE-L between the rewritten prompt against the original prompt measures the ability of prompt rewriting methods to recover the original prompt. We set $\\beta = 5$ to emphasize the recall of generated prompts.", + "bbox": [ + 496, + 167, + 892, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Personalized Prompt Rewriting", + "text_level": 1, + "bbox": [ + 498, + 272, + 790, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The basic idea of our personalization method is to rewrite the input prompt, considering user preferences gleaned from past user interactions. The full pipeline of our Personalized Prompt Rewriting (Personalized PR) method is depicted in the left part of Figure 7. If a user $u$ input a prompt $x_{t}$ , a retriever $\\mathbf{Ret}(x_{t},\\mathcal{Q}_{t})$ retrieves prompts from the user's historical prompt set $\\mathcal{Q}_t$ , using $x_{t}$ as a query. Based on the retrieval result $\\mathcal{R}_t = \\mathbf{Ret}(x_t,\\mathcal{Q}_t)$ , the rewritten Rew rewrites the input prompt to generate a personalized prompt $x_{t}^{\\prime} = \\mathbf{Rew}(x_{t},\\mathcal{R}_{t})$ . Finally, the text-to-image generation model $\\mathbf{G}$ produces the image $I_{t}^{\\prime} = \\mathbf{G}\\left(x_{t}^{\\prime},\\epsilon\\right)$ from the rewritten prompt, where $\\epsilon$ is a random noise vector.", + "bbox": [ + 496, + 297, + 892, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Retrieval and Ranking", + "text_level": 1, + "bbox": [ + 498, + 491, + 712, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the retrieval stage, given the input prompt $x_{t}$ , the retriever $\\mathbf{Ret}(x_t,\\mathcal{Q}_t)$ retrieves relevant prompts from historical prompt set $\\mathcal{Q}_t$ , using $x_{t}$ as a query.", + "bbox": [ + 496, + 513, + 890, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By analyzing user prompts, we have noticed that users tend to construct prompts that involve objects, their attributes, and the relationships between objects. In the works [23], [11], a query for image retrieval is defined to include objects, attributes of objects, and relations between objects. Inspired by this, we suspect users have the habit of using attributes and some objects, such as background, to express their preferences. To confirm this, we visualize the word cloud of the top 250 frequent words in the text prompts of all users, as shown in the right part of Figure 5. In Figure 5, we find some attributes, such as \"cute\", \"golden\", and \"beautiful\" appear in prompts with high frequency, as well as some objects, such as \"mountain\", \"sea,\" and \"sky\". Intuitively, we can use the current prompt $x_{t}$ to locate the relevant history prompts that include the same or similar attributes or objects.", + "bbox": [ + 496, + 560, + 892, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To locate relevant prompts, two retrieval methods are used: dense and sparse. In dense retrieval, we choose the prompt $x_{t}$ and calculate its textual embedding $\\mathbf{Em}(x_t)$ using CLIP's text encoder, also the text encoder in Stable Dif", + "bbox": [ + 496, + 801, + 890, + 862 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "The image generated from the prompt $x_{t}$ is denoted as $I_{t}$ , where $I_{t} = \\mathbf{G}(x_{t},\\epsilon)$ .", + "bbox": [ + 500, + 873, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7730", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/daab4cf2bf36d4a86877b4f34983a4edecc7278d55fcd4614047885933afa765.jpg", + "image_caption": [ + "Figure 7. Left: Pipeline of Personalized Prompt Rewriting (Personalized PR), including Retriever, Rewrter and T2I model to generate personalized images from user histories. Right: Illustration of context-dependent prompt rewriting. We present a specific example for better understanding the procedure of context-dependent prompt rewriting." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 285, + 333 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/aaa4eb49f2d6b4ce451c9064dbc9c17bfb61e5eeb5f8192a8b7feee3dfaf2a47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 87, + 888, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "fusion [20]. We suspect the prompts with similar visual attributes and objects will be close to each other in the text embedding space. To confirm this, we visualize some retrieval results in Figure 8. The three nearest neighbors of $\\mathbf{Em}(x_t)$ are prompts that are semantically related. For example, if the input prompt is \"Hobbit homes\", the three most relevant prompts would include the words \"village\", \"city\", and \"house\". This dense retrieval method is also referred to as embedding-based retrieval (EBR). In sparse retrieval, we use BM25 to locate relevant prompts that include the same visual attributes and objects.", + "bbox": [ + 75, + 422, + 472, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the above retrieval, we rank relevant prompts in EBR-based or BM25-based ranking, depending on the retrieval ways. In EBR-based ranking, we rank the relevant prompts based on their embedding similarity with the query $x_{t}$ . For similarity measuring, we choose cosine similarity as it is a commonly used similarity measure in embedding learning. In BM25-based ranking, BM25 scores are used for similarity measures. Consequently, we obtain the top $k$ relevant user queries $\\mathcal{R}_t = \\{r_1,\\dots,r_k\\}$ .", + "bbox": [ + 75, + 590, + 468, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Rewriting", + "text_level": 1, + "bbox": [ + 76, + 741, + 192, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The procedure of context-independent rewriting leverages pertinent queries $\\mathcal{R}_t = \\{r_1, \\dots, r_k\\}$ , and employs ChatGPT to encapsulate user preferences and rewrite the prompt directly. These queries $\\mathcal{R}_t$ are organized based on their relevance to $x_t$ .", + "bbox": [ + 75, + 762, + 468, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the context-dependent scenario, we initially create a collection of demonstration examples $\\mathcal{E} = \\{e_1,\\dots,e_L\\}$ using manual design. We then select a small subset of these examples to serve as demonstrations for each rewriting task.", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the issue of order sensitivity in in-context learning, as highlighted in the study [13], we arrange the demonstration examples in a descending sequence based on their proximity to the input prompt $x_{t}$ . The in-context rewriting process we employ is illustrated in the right section of Figure 7.", + "bbox": [ + 496, + 422, + 893, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 498, + 515, + 625, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We carried out experiments for prompt rewriting methods on our PIP dataset. We validate our method through both offline and online evaluation. And we further analyze the number of historical prompts for best extracting the users' preference by ablating top retrieval. For offline evaluation, we use the aforementioned three metrics: PMS, ImageAlign, and ROUGE-L.", + "bbox": [ + 496, + 541, + 890, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For online evaluation, we carried out single blind experiment for recently active users on our website. Real user feedback is collected to evaluate our method.", + "bbox": [ + 496, + 647, + 893, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 704, + 718, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Details of our Personalized PR method are as follows. In retrieval, we choose the relevant text prompt number as $k = 3$ . We use ChatGPT [15] as our rewritten. An input example for context-independent rewriting is shown in Table 1. For in-context rewriting, we set $L = 5$ and randomly select one demonstration example for each rewriting task, unless otherwise specified.", + "bbox": [ + 496, + 724, + 890, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unless other specified, all the experiments use EBR to", + "bbox": [ + 517, + 832, + 890, + 849 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "†The input for in-context rewriting is only different from context-independent rewriting in terms of the presence of demonstration example. See supplementary for detail.", + "bbox": [ + 498, + 861, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7731", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ed392e07c2ad5a2addc8c1a4d07be7e27a5118f71453e511774688a7bee3a7aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 89, + 880, + 263 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/716a9b1917b36ca46d889228cfa2866a6964b8dac09a833351e2fec31edd9248.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 282, + 555, + 433 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/16c63df41afbc5349041e50b0b29241f68276a46429ca7d08c6884fbf8d6a6ff.jpg", + "image_caption": [ + "Figure 8. Qualitative analysis of personalized retrieval and rewriting." + ], + "image_footnote": [], + "bbox": [ + 84, + 444, + 883, + 588 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/056a3f22eed6b764c316e0b468df60c7008e78f729c4523bd2746f47d49882a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 282, + 880, + 429 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/59d1473b9693830ffc8e7411cdb60b42cdec05cb5e0446ef4126920faa795816.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 444, + 880, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "retrieve historical prompts, and one-shot in-context learning to rewrite the shortened prompt.", + "bbox": [ + 76, + 647, + 468, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ChatGPT Input Template for Context-independent Rewriting", + "text_level": 1, + "bbox": [ + 89, + 696, + 459, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt in text-to-image generation describes the detailed attributes of the object user plans to draw. User's preference in text-to-image generation is shown in history prompts.", + "bbox": [ + 89, + 718, + 460, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given 3 history prompts, your task is to rewrite the current prompt so that it matches the user's preference. The rewritten prompt should retain primary objects in the original prompt and conform to the user's preference. Please avoid being too diffused and restrict your output within 70 words.", + "bbox": [ + 89, + 758, + 459, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The history prompts are: $\\{\\mathcal{R}_t\\}$", + "bbox": [ + 89, + 824, + 266, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The current prompt is: $\\{x_{t}\\}$", + "bbox": [ + 89, + 839, + 250, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The rewritten prompt (one sentence less than 70 words) is:", + "bbox": [ + 89, + 853, + 418, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Input template for context-independent rewriting.", + "bbox": [ + 99, + 882, + 444, + 896 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use Stable Diffusion (SD) v1-5 [19] as our text-to-image generation model for all methods. SD v1-5 is sampled using PNDM scheduler in 50 steps and setting the classifier-free guidance scale to 7.0.", + "bbox": [ + 498, + 646, + 890, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison Baselines", + "text_level": 1, + "bbox": [ + 500, + 720, + 705, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with two baseline methods, namely Prompt [7] and General Prompt Rewriting (General PR). Both are general text-to-image prompt rewriting methods completely overlooking users' preferences.", + "bbox": [ + 496, + 746, + 890, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Promptist [7]. Promptist uses the GPT-2 as the base model and has trained it on a self-collected text-to-image prompt dataset with 360K prompts, by using supervised fine-tuning method and reinforcement learning.", + "bbox": [ + 496, + 806, + 890, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "General PR. We prompt ChatGPT [15] to perform general prompt rewriting in a normal manner. In practice, we input", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7732", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ChatGPT with the current prompt $x_{t}$ without any user historical information. Apart from the absence of histories, everything else of the prompt template for ChatGPT remains the same as the template in Table 1.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 76, + 159, + 269, + 174 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To illustrate the effectiveness of our retrieval methods, we provide visualizations of the retrieved relevant image-prompt pairs in the left section of Figure 8. These examples are drawn from the experimental results of three different test prompts from three users. The retrieved histories exhibit a high degree of similarity with their corresponding queries in terms of objects, attributes, as well as the overall style or mood of the image.", + "bbox": [ + 75, + 181, + 467, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This effectively demonstrates the proficiency of our retriever in sourcing relevant user histories, thereby providing a robust reference for our rewriter to carry out personalized prompt rewriting.", + "bbox": [ + 75, + 301, + 467, + 362 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the right section of Figure 8, we display the generation outcomes of both \"Shortened Prompt\" and \"Personalized Prompt\". The \"Shortened Prompt\" column exhibits the results produced from the shortened prompts, while the \"Personalized Prompt\" column features the rewritten prompts of our method along with their corresponding generated images. It's evident from these displays that images our method generates are more inclined towards user preferences based on their histories, a testament to the expressive power of our rewritten prompt. For instance, when the query \"Hobbit homes\" is used (as seen in the first row), we observe the user's preferred style across three images, all capturing the mood of mountainous scenery and depicting the Hobbit homes within a consistent landscape.", + "bbox": [ + 75, + 363, + 467, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4.Quantitative Comparison", + "text_level": 1, + "bbox": [ + 76, + 580, + 308, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further examine the effectiveness of our Personalized PR method, we conduct offline and online quantitative evaluation, comparing different settings to the baseline methods without enhanced prompt rewritten on the test samples we created.", + "bbox": [ + 75, + 604, + 467, + 679 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/efd4f5abcdbd1a1a2bb877b1a57b6fc82dd3c215eb4e7d28317e9b5f14dd8c8d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRetrieverPMS ↑Image-Align↑ROUGE-L↑
Shortened Prompt-0.55670.62720.3268
Promptist [7]-0.58580.64810.2947
General PR-0.59960.59120.2082
Personalized PRBM250.61250.65810.3942
EBR0.60830.64850.4137
Personalized PR+ ICLBM250.62530.64560.4417
EBR0.61790.67960.4686
", + "bbox": [ + 81, + 686, + 467, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Offline Test. Table 2 showcases the numerical results comparing various retrieval and rewriting configurations with shortened prompts. We can see that Prompt [7] and General PR are slightly better than 'Shortened Prompt' in terms", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of PMS and Image-Align. Without the users' histories, these large language models perform prompt rewriting in arbitrary pathways. Thus, the rewritten prompts these baseline methods generate could totally deviate from the users' preferences. Such hallucination scenarios result in the decrease in terms of ROUGE-L for both Prompt [7] and General PR. This indicates that although general prompt rewriting methods appear to be refining the prompts, their results do not align with users' preferences well. Our method outperforms all baseline methods, i.e. Prompt [7] and General PR, on all metrics.", + "bbox": [ + 496, + 90, + 890, + 256 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A comparison between BM25 and EBR reveals that dense retrieval generally outperforms sparse retrieval, although the difference is relatively small, indicating that both methods can produce satisfactory results.", + "bbox": [ + 496, + 257, + 890, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Further, when contrasting context-independent rewriting with in-context rewriting, it's evident that ICL produces superior outcomes. By integrating ICL with EBR, we achieve absolute improvements of $14.2\\%$ , $6.9\\%$ and $5.6\\%$ in terms of ROUGE-L, PMS and Text-Align metrics respectively. This underscores the exceptional performance of personalized prompting.", + "bbox": [ + 496, + 316, + 890, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To check how sensitive our prompt rewriting methods with respect to the length of the prompts before rewriting, we assess our rewriting method using two additional types of prompts with shorter lengths, namely \"Noun Phrase\" and \"Noun\", as detailed in Table 3.", + "bbox": [ + 496, + 422, + 890, + 497 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/53605a8a04efb4544a1e4b9fde26de10876c124d83d85d32753b7eb3ca27a60c.jpg", + "table_caption": [ + "Table 2. Comparison results of different variants of our method with the baseline. Evidently, our method using EBR retriever (top-3 retrieval) and 1-shot ICL can achieve most best results." + ], + "table_footnote": [], + "table_body": "
Prompt TypeMethodPMS ↑Image-Align↑ROUGE-L↑
NounOriginal0.55370.60870.1770
Personalized PR0.61420.64780.2804
Noun PhraseOriginal0.55540.61460.2459
Personalized PR0.61680.65340.3387
Short SentenceOriginal0.55670.62720.3268
Personalized PR0.61790.67960.4686
", + "bbox": [ + 504, + 507, + 890, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Performance with respect to different prompt lengths, i.e. Noun, Noun Phrase and Short Sentence. Our \"Personalized PR\" equipped with top-3 dense retriever, and 1-shot ICL consistently enhances results, even with only nouns or noun phrases.", + "bbox": [ + 496, + 603, + 890, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These two prompts are derived using spaCy [9] from our dataset, adhering to the principle of minimizing word count while maintaining the main entities. The results displayed in Table 3 show that across all three shortened scale, \"Personalized PR\" outperforms the baseline across all metrics, demonstrating the effectiveness of our methods in recovering user preferences.", + "bbox": [ + 496, + 674, + 890, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Online Test. To further validate our approach, we have carried out a single-blinded online evaluation on our website mentioned above. Active users recently in the website are randomly selected to participate in the online test. Upon each prompt input by a participant, there's an equal chance of generating an image using either \"Original Prompt\" or our method's \"Personalized Prompt\". Participants can choose to \"Save\" or \"Delete\" each generated image based", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7733", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4dfb5d41b9245ab81ef5cad70471c6ce74710280455d16d04d90e61f114b7193.jpg", + "image_caption": [ + "Figure 9. Illustration of our online evaluation methods and the improvement of Save Rate from personalized Prompt." + ], + "image_footnote": [], + "bbox": [ + 81, + 93, + 251, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9819b24a6b39f99a5060e3c58cd2cc319cabda9fe82545fc51b0a288e94476c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 93, + 468, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on their preference. We use their \"Save\" actions to assess our method's effectiveness. In the online test, 247 users generated 905 images, with 433 from \"Original Prompt\" and 472 from \"Personalized Prompt\". The results of the online evaluation, as shown in Figure 9, indicate that users prefer images generated using the \"Personalized Prompt\" over the \"Original Prompt\", with a $17.1\\%$ increase in \"Save\" actions. This suggests that our method aligns better with user preferences, affirming its effectiveness. We anticipate even better results in real-world scenarios with more user information and an improved text-to-image generation method.", + "bbox": [ + 75, + 315, + 470, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 491, + 230, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we ablate $k$ relevant historical prompts used to rewrite personalized prompts and the number of demonstration examples used for in-context learning. We evaluate each experiment on the same setup as in Section 5.1.", + "bbox": [ + 75, + 515, + 468, + 575 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4d1f9a18aa185c0102f569b5e37a972c1e4a0da387ea31137bfadf068ca723e8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRetrieval Top-kPMS ↑Image-Align ↑ROUGE-L ↑
Personalized PR + ICL10.60570.67510.4539
30.61790.67960.4686
50.62040.67480.4474
70.62650.66510.4592
", + "bbox": [ + 81, + 589, + 468, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Retrieval Top-k Ablation. As shown in Table 4, when using 3 most relevant historical prompts for rewriting, we can obtain most best results among all evaluation metrics. We analyze that too more historical prompts could provide redundant information and also too long prompt input to ChatGPT worsens the rewriting performance. Therefore, we choose 3 as the number of retrieval results, a balanced manner between performance and efficiency.", + "bbox": [ + 75, + 703, + 468, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of ICL Demonstrations Ablation. Table 5 shows that 1-shot setting, i.e., given 1 demonstration example for in-context learning, can achieve the best results in general among all the four evaluation metrics regarding both BM25 and EBR. This demonstrates that our prompt rewriting tem-", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/25c3c76b59abeebd107d9ba4624310686d91fb5262cef1dbcfd3df4e218e98b7.jpg", + "table_caption": [ + "Table 4. Ablation for Retrieval Top-k. We empirically conduct experiments with respect to $k \\in \\{ 1,3,5,7\\}$ ,keeping the configuration of DENSE retriever and 1-shot ICL." + ], + "table_footnote": [], + "table_body": "
Method (Retrieval)ICL ShotPMS ↑Image-Align ↑ROUGE-L ↑
Personalized PR (BM25)10.62530.64560.4417
30.62890.65800.4381
50.62360.65710.4226
Personalized PR (EBR)10.61790.67960.4686
30.62740.67080.4354
50.62420.67240.4439
", + "bbox": [ + 504, + 88, + 890, + 170 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation for Number of ICL Demonstrations. We experiment with consistent top-3 retrieval both on BM25 and EBR. When we set ICL shot as 1 and use EBR retriever, we observe more superior results appearing.", + "bbox": [ + 498, + 180, + 890, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "plate is efficient and effective enough for extracting the personalized preference from numerous historical data of each user.", + "bbox": [ + 498, + 261, + 890, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 498, + 333, + 769, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, this study has underscored the significance of harnessing historical user behaviors to construct personalized AI content generation. Our strategy aims to refine user prompts by capitalizing on previous user interactions with the system. We have introduced an innovative technique that entails reconfiguring user prompts based on a newly developed, large-scale text-to-image dataset encompassing over 300,000 prompts from 3,115 distinct users. This methodology has proven to augment the expressiveness of user prompts and ensure their alignment with the desired visual outputs. Our empirical results have underscored the supremacy of our techniques over conventional methods. This superiority was corroborated through our novel offline evaluation method and online tests, thereby affirming the efficacy of our approach.", + "bbox": [ + 496, + 363, + 890, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although the outcomes are encouraging, considerable research still lies ahead in this domain. In the realm of personalized text-to-image generation, the incorporation of more personal details like a user's age and gender could bolster performance. The methodologies used could be extended to other LPMs, including LLMs. Additionally, techniques for enhancing search engines, e.g. data source purification and ranking optimization, could be assimilated into these models.", + "bbox": [ + 496, + 594, + 890, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We are confident that our contributions represent advancements towards a more personalized and user-focused artificial intelligence.", + "bbox": [ + 498, + 734, + 890, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported in part by the Research Center for Industries of the Future at Westlake University and Westlake Education Foundation (Grant No. WU2023C017), National Natural Science Foundation of China (Grant Nos. 62171111, 62006245), and Open Research Fund from State Key Laboratory of High Performance Computing of China (Grant No. 202201-15).", + "bbox": [ + 496, + 803, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7734", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1", + "[2] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. \"this is my unicorn, fluffy\": Personalizing frozen vision-language representations. In European Conference on Computer Vision, pages 558-577. Springer, 2022. 2", + "[3] Grégoire Deletang, Anian Ruoss, Paul-Ambroise Duquenne, Elliot Catt, Tim Genewein, Christopher Mattern, Jordi Grau-Moya, Li Kevin Wenliang, Matthew Aitchison, Laurent Orseau, et al. Language modeling is compression. arXiv preprint arXiv:2309.10668, 2023. 1", + "[4] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1, 2", + "[5] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 2", + "[6] Qingyan Guo, Rui Wang, Junliang Guo, Bei Li, Kaitao Song, Xu Tan, Guoqing Liu, Jiang Bian, and Yujiu Yang. Connecting large language models with evolutionary algorithms yields powerful prompt optimizers, 2023. 2", + "[7] Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. arXiv preprint arXiv:2212.09611, 2022. 2, 6, 7, 1", + "[8] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 4", + "[9] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python. 2020. 7", + "[10] Rolf Jagerman, Honglei Zhuang, Zhen Qin, Xuanhui Wang, and Michael Bendersky. Query expansion by prompting large language models. arXiv preprint arXiv:2305.03653, 2023. 1, 2", + "[11] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 4", + "[12] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2", + "[13] Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. What makes" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "good in-context examples for gpt-3? arXiv preprint arXiv:2101.06804, 2021.5", + "[14] Jianmo Ni, Chen Qu, Jing Lu, Zhuyun Dai, Gustavo Hernández Abrego, Ji Ma, Vincent Y Zhao, Yi Luan, Keith B Hall, Ming-Wei Chang, et al. Large dual encoders are generalizable retrievers. arXiv preprint arXiv:2112.07899, 2021. 4", + "[15] OpenAI. Chatgpt. https://chat.openai.com. 5, 6, 3", + "[16] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4", + "[17] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2", + "[18] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[19] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. Stable diffusion v1-5. https://huggingface.co/runwayml/stable-diffusion-v1-5.6", + "[20] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3, 5", + "[21] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2", + "[22] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2", + "[23] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 4", + "[24] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 2", + "[25] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7735", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Xiao Wang, Sean MacAvaney, Craig Macdonald, and Iadh Ounis. Generative query reformulation for effective adhoc search. arXiv preprint arXiv:2308.00415, 2023. 2", + "[27] Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V Le, Denny Zhou, and Xinyun Chen. Large language models as optimizers. arXiv preprint arXiv:2309.03409, 2023. 2", + "[28] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2", + "[29] Shanshan Zhong, Zhongzhan Huang, Wushao Wen, Jinghui Qin, and Liang Lin. Sur-adapter: Enhancing text-to-image pre-trained diffusion models with large language models. arXiv preprint arXiv:2305.05189, 2023. 1, 2", + "[30] Yongchao Zhou, Andrei Loan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. arXiv preprint arXiv:2211.01910, 2022. 2" + ], + "bbox": [ + 78, + 90, + 467, + 345 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7736", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/7aad987a-39f2-4096-8aa1-a539bbaacb9c_model.json b/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/7aad987a-39f2-4096-8aa1-a539bbaacb9c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2240bd33483f203ca5615b717be13e98d978b335 --- /dev/null +++ b/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/7aad987a-39f2-4096-8aa1-a539bbaacb9c_model.json @@ -0,0 +1,2332 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.809, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.13, + 0.856, + 0.177 + ], + "angle": 0, + "content": "Tailored Visions: Enhancing Text-to-Image Generation with Personalized Prompt Rewriting" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.202, + 0.842, + 0.224 + ], + "angle": 0, + "content": "Zijie Chen*,1,2 Lichao Zhang*,2 Fangsheng Weng3 Lili Pan†,4 Zhenzhong Lan†,2" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.228, + 0.729, + 0.247 + ], + "angle": 0, + "content": "\\(^{1}\\)Zhejiang University \\(^{2}\\)Westlake University \\(^{3}\\)Scietrain" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.247, + 0.731, + 0.265 + ], + "angle": 0, + "content": "4University of Electronic Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.267, + 0.835, + 0.281 + ], + "angle": 0, + "content": "{chenzijie, zhanglichao, lanzhenzhong}@westlake.edu.cn, lilipan@uestc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.316, + 0.314, + 0.332 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.474, + 0.621 + ], + "angle": 0, + "content": "Despite significant progress in the field, it is still challenging to create personalized visual representations that align closely with the desires and preferences of individual users. This process requires users to articulate their ideas in words that are both comprehensible to the models and accurately capture their vision, posing difficulties for many users. In this paper, we tackle this challenge by leveraging historical user interactions with the system to enhance user prompts. We propose a novel approach that involves rewriting user prompts based on a newly collected large-scale text-to-image dataset with over 300k prompts from 3115 users. Our rewriting model enhances the expressiveness and alignment of user prompts with their intended visual outputs. Experimental results demonstrate the superiority of our methods over baseline approaches, as evidenced in our new offline evaluation method and online tests. Our code and dataset are available at https://github.com/zzjchen/Tailored-Visions" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.652, + 0.21, + 0.668 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.678, + 0.47, + 0.828 + ], + "angle": 0, + "content": "Increasingly large and powerful foundation models [1, 4, 20] are trained through self-supervised learning. These large pretrained models (LPMs) serve as efficient compressors [3], condensing vast amounts of internet data. This compression enables the convenient extraction of the knowledge encoded within these models via natural language descriptions. Despite being in its infancy, this approach exhibits potentials to surpass traditional search engines as a superior source for knowledge and information acquisition." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.83, + 0.47, + 0.86 + ], + "angle": 0, + "content": "Akin to refining queries for search engines, prompts given to LPMs must also be carefully crafted. However," + }, + { + "type": "title", + "bbox": [ + 0.55, + 0.318, + 0.611, + 0.329 + ], + "angle": 0, + "content": "User History" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.344, + 0.565, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.567, + 0.348, + 0.655, + 0.386 + ], + "angle": 0, + "content": "Large Italianate House with Beautiful Breathing on the Shores of Lake Como-Oil Painting in the Style of Claude Monet trending in artstation" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.392, + 0.565, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.567, + 0.402, + 0.654, + 0.428 + ], + "angle": 0, + "content": "A Magical Forest in the morning, oil painting in the style of Claude Monet trending in artstation." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.439, + 0.566, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.568, + 0.449, + 0.654, + 0.476 + ], + "angle": 0, + "content": "Tulip bouquet. Oil Painting, bright color, highly detailed, Vincent Van Gogh, trending in artstation" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.49, + 0.566, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.568, + 0.493, + 0.654, + 0.526 + ], + "angle": 0, + "content": "Still life of a hamburger and some fruits on a napkin, detailed oil painting in the style of Paul Cezanme, trenging in artstation" + }, + { + "type": "title", + "bbox": [ + 0.728, + 0.319, + 0.811, + 0.329 + ], + "angle": 0, + "content": "Generated Results" + }, + { + "type": "text", + "bbox": [ + 0.687, + 0.331, + 0.744, + 0.345 + ], + "angle": 0, + "content": "Original Prompt: Walking street at night" + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.345, + 0.755, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.331, + 0.875, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.662, + 0.404, + 0.77, + 0.427 + ], + "angle": 0, + "content": "t... a mesmerizing walking street at night...with a touch of Italianate c-harm, reminiscent of Claude Monet's oil painting style." + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.429, + 0.754, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.754, + 0.404, + 0.88, + 0.427 + ], + "angle": 0, + "content": "et an astronaut, inspired by Claude tce cMonet_rides a horse in a captivati ng scene,..., fusion of magic and natury..." + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.429, + 0.866, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.878, + 0.439, + 0.888, + 0.478 + ], + "angle": 0, + "content": "W/o History" + }, + { + "type": "text", + "bbox": [ + 0.776, + 0.487, + 0.88, + 0.535 + ], + "angle": 0, + "content": "An astronaut, adorned in futuristic gear, gracefully maneuvers on horseback through an otherwise landscape bathed in the ethereal glow of distant stars, capturing the juxtaposition of advanced space exploration and timeless equestrian elegance." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.566, + 0.895, + 0.637 + ], + "angle": 0, + "content": "Figure 1. Comparison between our personalized prompt rewriting method and the standard prompt rewriting method. Our technique excels at incorporating user preferences, such as \"oil paintings by artists,\" while methods that lack a historical context frequently generate content that may not align with the user's desires." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.894, + 0.779 + ], + "angle": 0, + "content": "the complexity of prompts, the unpredictability of model responses compared to traditional search engines present unique challenges. Significant research efforts [10, 29] have been made to comprehend how LPMs react to various prompts, with some studies examining the feasibility of rewriting prompts for specificity. However, without access to users' personal data and behavior, tailoring the prompt to meet the user's needs accurately remains challenging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Our research addresses this issue by integrating user preference information into prompt rewriting. The primary obstacle in personalized query rewriting is the absence of a dataset containing text-to-image prompts with personalized information. To overcome this, we have assembled a large dataset encompassing over 300k text-to-image histories from 3,115 users. We rewrite user prompts using their query history, although we had limited access to personal" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.872, + 0.208, + 0.886 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.872, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7727" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.181 + ], + "angle": 0, + "content": "information, leaving room for further research. Another significant challenge is the evaluation of rewritten queries. To evaluate their efficacy, we've developed a new offline method that uses multiple metrics to measure how well our rewriting models can recover the original user query from the ChatGPT-shortened version." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.183, + 0.361, + 0.197 + ], + "angle": 0, + "content": "Our paper's contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.198, + 0.469, + 0.242 + ], + "angle": 0, + "content": "1. We have compiled a large Personalized Image Prompt (PIP) dataset and made it publicly available to aid future research in the field." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.243, + 0.469, + 0.287 + ], + "angle": 0, + "content": "2. We experimented with two query rewriting techniques and proposed a new query evaluation method to assess their performance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.288, + 0.469, + 0.331 + ], + "angle": 0, + "content": "3. We propose a new benchmark for personalized text-to-image generation, which promotes the standardization of this field." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.198, + 0.469, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.469, + 0.409 + ], + "angle": 0, + "content": "While there is still a considerable distance to cover before we can create a perfect prompt encapsulating both the user's requirements and the model's capabilities, we believe our research provides a critical stepping stone in this ongoing exploration." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.417, + 0.22, + 0.433 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.436, + 0.469, + 0.511 + ], + "angle": 0, + "content": "This section provides an overview of prior work on text-to-image generation, personalization for such generation, and prompt rewriting. It's important to note that our review is aimed more at offering sufficient background knowledge rather than exhaustive coverage of all related works." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.519, + 0.314, + 0.535 + ], + "angle": 0, + "content": "2.1. Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.469, + 0.854 + ], + "angle": 0, + "content": "Large text-to-image generation models can generate high-fidelity image synthesis and achieve a deep level of language understanding. DALL-E [17] uses a VQ-VAE transformer-based method to learn a visual codebook in the first stage and then trains autoregressive transformers on sequences of text tokens followed by image tokens in the second stage. DALL-E2 [18] introduces latent diffusion models to generate various images by conditioning on CLIP text latents and CLIP image embeddings generated by a prior model. Imagen [22] discovers that a larger language model with more parameters trained on text-only data improves the quality of text-to-image generation. Late developments like stable-diffusion (SD) [20] proposes to generate images effectively in latent space significantly lowering computational costs. Furthermore, SD designs a conditional mechanism to complete class-conditional, text-to-image and layout-to-image models. Furthermore, ControlNet [28] accomplishes certain function by conditioning on multi-modal data, e.g., edge, sketching, pose, segmentation, depth etc., which unavoidably involving additional condition-generation modalities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Despite these models' ability to generate high-fidelity images, they often fail to meet the precise needs of the users. Text-to-image generation is more like a game of chance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.091, + 0.885, + 0.107 + ], + "angle": 0, + "content": "2.2. Personalization for Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.111, + 0.892, + 0.352 + ], + "angle": 0, + "content": "Recently, personalization approaches based on text-to-image models have taken a set of images of a concept and generated variations of the concept. Specifically, some methods optimize a set of text embeddings. For example, [2] involves pseudo-word embeddings by a set encoder to provide personalization and Textual inversion [4] composes the concept into language sentences and performed as a personalized creation. Some methods finetune the diffusion model. For example, DreamBooth [21] finetunes the text-to-image diffusion model with shared parallel branches. To speed up, CustomDiffusion [12] reduces the amount of finetuned parameters, and Tewel et al. [25] locks the subject's cross-attention key to its superordinate category to align with visual concepts. Moreover, an additional encoder is trained to map concept images to its textual representation by [5] and [24]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Existing studies have three key limitations: they demand extra images and fine-tuning of text-to-image models with limited scope for new concepts; they can't learn from user interaction history and need detailed user prompts; and there's a lack of public, personalized text-to-image datasets that truly reflect user preferences." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.45, + 0.677, + 0.466 + ], + "angle": 0, + "content": "2.3. Prompt Rewriting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.68 + ], + "angle": 0, + "content": "Recently, researchers have found that optimizing prompts can boost the performance of LLMs on several NLP tasks and even search systems. For examples, Guo et al. [6] connect the LLM with evolutionary algorithms to generate an optimized prompt from parent prompts, without any gradient calculation. Yang et al. [27] propose to use LLM as an optimizer by generating new prompts based on a trajectory of previously generated prompts in each optimization step with the objective of maximizing the accuracy of the task. In the work [30], LLMs serve as models for engineering work like inference, scoring, and resampling. In search systems, LLMs are used to generate query expansion terms by [10], while they are used to reformulate query by Wang et al. [26] instead." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.681, + 0.892, + 0.772 + ], + "angle": 0, + "content": "For T2I generation, one relatively close work, SUR-adapter [29] learns to align the representations between simple prompt and complex prompt by an adapter. [7] optimize prompt through general rewriting. However, Neither of above works utilized personalized information for improving prompts." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.783, + 0.875, + 0.8 + ], + "angle": 0, + "content": "3. Personalized Image-Prompt (PIP) Dataset" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.808, + 0.677, + 0.823 + ], + "angle": 0, + "content": "3.1. Dataset Collection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.831, + 0.892, + 0.878 + ], + "angle": 0, + "content": "The Personalized Image-Prompt (PIP) dataset is the first large-scale personalized generated image-text dataset. The original data are collected from a public website that we" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.886, + 0.611, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\) https://zmrj.art/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7728" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.471, + 0.334 + ], + "angle": 0, + "content": "Figure 2. Dataset creation process. We split our dataset into training and testing sets and summarize each prompts in the test set using ChatGPT." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.349, + 0.466, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.518, + 0.471, + 0.603 + ], + "angle": 0, + "content": "Figure 3. Dataset statistics and distribution. Left: Proportion of users based on the varying number of historical prompts they have. Note that each user has a minimum of 18 historical prompts, as we have excluded those with fewer prompts from the dataset. Right: Proportion of prompts based on their varying lengths. Best view in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.719 + ], + "angle": 0, + "content": "host to provide open-domain text-to-image generation to users. PIP dataset includes 3115 users and 300,237 text-to-image histories generated by these users using SD v1-5 [20] and an internal fine-tuned version of SD v1-5. Each user in PIP has created 18 or more images and provided at least 12 different prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.886 + ], + "angle": 0, + "content": "Figure 2 illustrates the process of creating the dataset. For each individual user, we randomly choose two prompts to serve as test prompts, with the remaining prompts allocated as training prompts (historical user query). The purpose of using random selection instead of the most recent generated prompts is to enhance the diversity of our test data. Subsequently, we employ ChatGPT to condense the test prompts, ensuring they only include the primary object or scene, as depicted in Figure 2. We shorten the prompts into three scales, i.e., contain only nouns, noun phases or short sentences respectively." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In the ensuing experiment, each test prompt in the test" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "set will be considered as the input prompt \\( x_{t} \\) for every user \\( u \\), with the original prompts serving as the ground truth that reflect the user's authentic preferences. The remaining prompts are utilized as training samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.892, + 0.198 + ], + "angle": 0, + "content": "The PIP dataset consists of 300,237 image-prompt pairs, personally categorized by 3,115 users. These pairs are divided into 294,007 training samples and 6,230 test samples." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.206, + 0.8, + 0.22 + ], + "angle": 0, + "content": "3.2. Dataset Statistics and Distribution" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.304 + ], + "angle": 0, + "content": "In this section, we showcase the data statistics that depict the quality and diversity of PIP. We specifically illustrate data distributions of the number of prompts and prompt length for each user, and delve deeper into the content of the prompt through a word cloud representation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.304, + 0.892, + 0.335 + ], + "angle": 0, + "content": "Each data sample contains a prompt, the generated images,UserID, Image size, and URL, as illustrated Figure 4." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.349, + 0.885, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.516, + 0.892, + 0.543 + ], + "angle": 0, + "content": "Figure 4. Two examples from a user history, containing Image, Prompt, User ID, Image size and URL." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.569, + 0.888, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.745, + 0.892, + 0.772 + ], + "angle": 0, + "content": "Figure 5. Word cloud visualization of top 250 keywords sampled from the PIP dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In PIP dataset, each user contributes at least 18 images, as depicted in the left section of Figure 3. This results in a long-tail distribution. The prompts have an average word count of 27.53. The length of the prompts, ranging from 1 to 284 words, also follows a long-tail distribution, as seen in the right section of Figure 3. Despite the presence of about 2500 prompts that exceed the 75-word limit of SD, we retain them to maintain the integrity of user preferences." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7729" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.468, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.228, + 0.472, + 0.287 + ], + "angle": 0, + "content": "Figure 6. Text similarity between user histories and user preferences. There is a great deal of diversity in user preferences in the PIP dataset. In the zoomed-in version of the similarity map on the left, users' history prompts are highly similar to their preferences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.47, + 0.418 + ], + "angle": 0, + "content": "Figure 5 presents the 250 most frequently used words or phrases. The frequency of these words is determined by the highest TF-IDF value across all users. The word cloud reveals that these words describe various image attributes, such as objects, styles, quality, and colors. This variety underlines the high diversity present within the prompt content of the PIP dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.42, + 0.473, + 0.618 + ], + "angle": 0, + "content": "In addition, we visualize the text similarities between the histories and preferences of 100 users in Figure 6. For each user \\( u \\) in PIP dataset, we summarize his preference \\( P_{u} \\) into 5 phrases from his history prompts using ChatGPT. For two users \\( u, v \\), the similarity of \\( u \\)'s histories and \\( v \\)'s preferences is defined as the mean of text similarities between \\( u \\)'s history prompts and \\( v \\)'s preference \\( P_{v} \\). The text similarity is calculated using GTR-T5-large [14]. Furthermore, we visualize the text similarities between the history prompts of a random user and 100 user preferences in the left part of Figure 6. This shows that text-to-image users have different preferences and the preference \\( p_{u} \\) we summarized successfully captures the key feature the user prefers." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.63, + 0.263, + 0.645 + ], + "angle": 0, + "content": "3.3. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.47, + 0.716 + ], + "angle": 0, + "content": "We present two metrics to evaluate prompt rewriting methods in terms of how the rewritten results are aligned to users' preferences, namely Preference Matching Score (PMS) and Image-Align." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.47, + 0.777 + ], + "angle": 0, + "content": "Preference Matching Score (PMS). PMS calculates the CLIPScore [8] between generated image and user's preference \\( P_{u} \\). It measures how the generated image aligns with the user's preference." + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.786, + 0.47, + 0.829 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P M S} = \\frac {w}{N} \\sum_ {u = 1} ^ {N} \\max \\left(\\cos \\left(\\mathbf {E m} \\left(I _ {u} ^ {\\prime}\\right), \\mathbf {E m} \\left(P _ {u}\\right)\\right), 0\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\( P_{u} \\) is the user \\( u \\)'s preference, \\( I_{u}^{\\prime} \\) is the generated image correspondingly, \\( N \\) is total user number (i.e. 3115). \\( \\mathbf{Em} \\) means the embedding extracted by using CLIP. \\( w = 2.5 \\) is a scaling constant." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "Image-Align. It measures the similarity between the generated image and the ground-truth image. Image-Align quantifies how closely the current created image aligns with the user's truly saved image. The similarity between two images are calculated using CLIP [16]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.893, + 0.259 + ], + "angle": 0, + "content": "Apart from these metrics, we also adopt ROUGE-L to evaluate prompt rewriting methods in our experiment. Calculating ROUGE-L between the rewritten prompt against the original prompt measures the ability of prompt rewriting methods to recover the original prompt. We set \\(\\beta = 5\\) to emphasize the recall of generated prompts." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.273, + 0.791, + 0.29 + ], + "angle": 0, + "content": "4. Personalized Prompt Rewriting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.298, + 0.893, + 0.482 + ], + "angle": 0, + "content": "The basic idea of our personalization method is to rewrite the input prompt, considering user preferences gleaned from past user interactions. The full pipeline of our Personalized Prompt Rewriting (Personalized PR) method is depicted in the left part of Figure 7. If a user \\( u \\) input a prompt \\( x_{t} \\), a retriever \\( \\mathbf{Ret}(x_{t},\\mathcal{Q}_{t}) \\) retrieves prompts from the user's historical prompt set \\( \\mathcal{Q}_t \\), using \\( x_{t} \\) as a query. Based on the retrieval result \\( \\mathcal{R}_t = \\mathbf{Ret}(x_t,\\mathcal{Q}_t) \\), the rewritten Rew rewrites the input prompt to generate a personalized prompt \\( x_{t}^{\\prime} = \\mathbf{Rew}(x_{t},\\mathcal{R}_{t}) \\). Finally, the text-to-image generation model \\( \\mathbf{G} \\) produces the image \\( I_{t}^{\\prime} = \\mathbf{G}\\left(x_{t}^{\\prime},\\epsilon\\right) \\) from the rewritten prompt, where \\( \\epsilon \\) is a random noise vector." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.492, + 0.713, + 0.508 + ], + "angle": 0, + "content": "4.1. Retrieval and Ranking" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.892, + 0.56 + ], + "angle": 0, + "content": "In the retrieval stage, given the input prompt \\( x_{t} \\), the retriever \\( \\mathbf{Ret}(x_t,\\mathcal{Q}_t) \\) retrieves relevant prompts from historical prompt set \\( \\mathcal{Q}_t \\), using \\( x_{t} \\) as a query." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.893, + 0.802 + ], + "angle": 0, + "content": "By analyzing user prompts, we have noticed that users tend to construct prompts that involve objects, their attributes, and the relationships between objects. In the works [23], [11], a query for image retrieval is defined to include objects, attributes of objects, and relations between objects. Inspired by this, we suspect users have the habit of using attributes and some objects, such as background, to express their preferences. To confirm this, we visualize the word cloud of the top 250 frequent words in the text prompts of all users, as shown in the right part of Figure 5. In Figure 5, we find some attributes, such as \"cute\", \"golden\", and \"beautiful\" appear in prompts with high frequency, as well as some objects, such as \"mountain\", \"sea,\" and \"sky\". Intuitively, we can use the current prompt \\( x_{t} \\) to locate the relevant history prompts that include the same or similar attributes or objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.803, + 0.892, + 0.863 + ], + "angle": 0, + "content": "To locate relevant prompts, two retrieval methods are used: dense and sparse. In dense retrieval, we choose the prompt \\( x_{t} \\) and calculate its textual embedding \\( \\mathbf{Em}(x_t) \\) using CLIP's text encoder, also the text encoder in Stable Dif" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The image generated from the prompt \\( x_{t} \\) is denoted as \\( I_{t} \\), where \\( I_{t} = \\mathbf{G}(x_{t},\\epsilon) \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7730" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.287, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.088, + 0.889, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.359, + 0.895, + 0.403 + ], + "angle": 0, + "content": "Figure 7. Left: Pipeline of Personalized Prompt Rewriting (Personalized PR), including Retriever, Rewrter and T2I model to generate personalized images from user histories. Right: Illustration of context-dependent prompt rewriting. We present a specific example for better understanding the procedure of context-dependent prompt rewriting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.473, + 0.59 + ], + "angle": 0, + "content": "fusion [20]. We suspect the prompts with similar visual attributes and objects will be close to each other in the text embedding space. To confirm this, we visualize some retrieval results in Figure 8. The three nearest neighbors of \\(\\mathbf{Em}(x_t)\\) are prompts that are semantically related. For example, if the input prompt is \"Hobbit homes\", the three most relevant prompts would include the words \"village\", \"city\", and \"house\". This dense retrieval method is also referred to as embedding-based retrieval (EBR). In sparse retrieval, we use BM25 to locate relevant prompts that include the same visual attributes and objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.47, + 0.729 + ], + "angle": 0, + "content": "In the above retrieval, we rank relevant prompts in EBR-based or BM25-based ranking, depending on the retrieval ways. In EBR-based ranking, we rank the relevant prompts based on their embedding similarity with the query \\( x_{t} \\). For similarity measuring, we choose cosine similarity as it is a commonly used similarity measure in embedding learning. In BM25-based ranking, BM25 scores are used for similarity measures. Consequently, we obtain the top \\( k \\) relevant user queries \\( \\mathcal{R}_t = \\{r_1,\\dots,r_k\\} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.194, + 0.759 + ], + "angle": 0, + "content": "4.2. Rewriting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.47, + 0.839 + ], + "angle": 0, + "content": "The procedure of context-independent rewriting leverages pertinent queries \\(\\mathcal{R}_t = \\{r_1, \\dots, r_k\\}\\), and employs ChatGPT to encapsulate user preferences and rewrite the prompt directly. These queries \\(\\mathcal{R}_t\\) are organized based on their relevance to \\(x_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In the context-dependent scenario, we initially create a collection of demonstration examples \\(\\mathcal{E} = \\{e_1,\\dots,e_L\\}\\) using manual design. We then select a small subset of these examples to serve as demonstrations for each rewriting task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.895, + 0.502 + ], + "angle": 0, + "content": "Given the issue of order sensitivity in in-context learning, as highlighted in the study [13], we arrange the demonstration examples in a descending sequence based on their proximity to the input prompt \\( x_{t} \\). The in-context rewriting process we employ is illustrated in the right section of Figure 7." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.516, + 0.627, + 0.533 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.648 + ], + "angle": 0, + "content": "We carried out experiments for prompt rewriting methods on our PIP dataset. We validate our method through both offline and online evaluation. And we further analyze the number of historical prompts for best extracting the users' preference by ablating top retrieval. For offline evaluation, we use the aforementioned three metrics: PMS, ImageAlign, and ROUGE-L." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.895, + 0.694 + ], + "angle": 0, + "content": "For online evaluation, we carried out single blind experiment for recently active users on our website. Real user feedback is collected to evaluate our method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.705, + 0.719, + 0.722 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.726, + 0.892, + 0.832 + ], + "angle": 0, + "content": "Details of our Personalized PR method are as follows. In retrieval, we choose the relevant text prompt number as \\( k = 3 \\). We use ChatGPT [15] as our rewritten. An input example for context-independent rewriting is shown in Table 1. For in-context rewriting, we set \\( L = 5 \\) and randomly select one demonstration example for each rewriting task, unless otherwise specified." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.833, + 0.892, + 0.85 + ], + "angle": 0, + "content": "Unless other specified, all the experiments use EBR to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.499, + 0.862, + 0.892, + 0.901 + ], + "angle": 0, + "content": "†The input for in-context rewriting is only different from context-independent rewriting in terms of the presence of demonstration example. See supplementary for detail." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "7731" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.882, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.283, + 0.557, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.445, + 0.885, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.283, + 0.881, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.445, + 0.881, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.606, + 0.688, + 0.621 + ], + "angle": 0, + "content": "Figure 8. Qualitative analysis of personalized retrieval and rewriting." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.648, + 0.47, + 0.678 + ], + "angle": 0, + "content": "retrieve historical prompts, and one-shot in-context learning to rewrite the shortened prompt." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.46, + 0.712 + ], + "angle": 0, + "content": "ChatGPT Input Template for Context-independent Rewriting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.461, + 0.758 + ], + "angle": 0, + "content": "Prompt in text-to-image generation describes the detailed attributes of the object user plans to draw. User's preference in text-to-image generation is shown in history prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.759, + 0.46, + 0.824 + ], + "angle": 0, + "content": "Given 3 history prompts, your task is to rewrite the current prompt so that it matches the user's preference. The rewritten prompt should retain primary objects in the original prompt and conform to the user's preference. Please avoid being too diffused and restrict your output within 70 words." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.825, + 0.267, + 0.839 + ], + "angle": 0, + "content": "The history prompts are: \\(\\{\\mathcal{R}_t\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.25, + 0.854 + ], + "angle": 0, + "content": "The current prompt is: \\(\\{x_{t}\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.854, + 0.419, + 0.867 + ], + "angle": 0, + "content": "The rewritten prompt (one sentence less than 70 words) is:" + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.883, + 0.446, + 0.897 + ], + "angle": 0, + "content": "Table 1. Input template for context-independent rewriting." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.647, + 0.892, + 0.708 + ], + "angle": 0, + "content": "We use Stable Diffusion (SD) v1-5 [19] as our text-to-image generation model for all methods. SD v1-5 is sampled using PNDM scheduler in 50 steps and setting the classifier-free guidance scale to 7.0." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.721, + 0.707, + 0.737 + ], + "angle": 0, + "content": "5.2. Comparison Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.892, + 0.807 + ], + "angle": 0, + "content": "We compare our method with two baseline methods, namely Prompt [7] and General Prompt Rewriting (General PR). Both are general text-to-image prompt rewriting methods completely overlooking users' preferences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.808, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Promptist [7]. Promptist uses the GPT-2 as the base model and has trained it on a self-collected text-to-image prompt dataset with 360K prompts, by using supervised fine-tuning method and reinforcement learning." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "General PR. We prompt ChatGPT [15] to perform general prompt rewriting in a normal manner. In practice, we input" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7732" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "ChatGPT with the current prompt \\( x_{t} \\) without any user historical information. Apart from the absence of histories, everything else of the prompt template for ChatGPT remains the same as the template in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.16, + 0.27, + 0.175 + ], + "angle": 0, + "content": "5.3. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.468, + 0.303 + ], + "angle": 0, + "content": "To illustrate the effectiveness of our retrieval methods, we provide visualizations of the retrieved relevant image-prompt pairs in the left section of Figure 8. These examples are drawn from the experimental results of three different test prompts from three users. The retrieved histories exhibit a high degree of similarity with their corresponding queries in terms of objects, attributes, as well as the overall style or mood of the image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.468, + 0.363 + ], + "angle": 0, + "content": "This effectively demonstrates the proficiency of our retriever in sourcing relevant user histories, thereby providing a robust reference for our rewriter to carry out personalized prompt rewriting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.468, + 0.575 + ], + "angle": 0, + "content": "In the right section of Figure 8, we display the generation outcomes of both \"Shortened Prompt\" and \"Personalized Prompt\". The \"Shortened Prompt\" column exhibits the results produced from the shortened prompts, while the \"Personalized Prompt\" column features the rewritten prompts of our method along with their corresponding generated images. It's evident from these displays that images our method generates are more inclined towards user preferences based on their histories, a testament to the expressive power of our rewritten prompt. For instance, when the query \"Hobbit homes\" is used (as seen in the first row), we observe the user's preferred style across three images, all capturing the mood of mountainous scenery and depicting the Hobbit homes within a consistent landscape." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.582, + 0.31, + 0.598 + ], + "angle": 0, + "content": "5.4.Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.468, + 0.68 + ], + "angle": 0, + "content": "To further examine the effectiveness of our Personalized PR method, we conduct offline and online quantitative evaluation, comparing different settings to the baseline methods without enhanced prompt rewritten on the test samples we created." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.688, + 0.468, + 0.787 + ], + "angle": 0, + "content": "
MethodRetrieverPMS ↑Image-Align↑ROUGE-L↑
Shortened Prompt-0.55670.62720.3268
Promptist [7]-0.58580.64810.2947
General PR-0.59960.59120.2082
Personalized PRBM250.61250.65810.3942
EBR0.60830.64850.4137
Personalized PR+ ICLBM250.62530.64560.4417
EBR0.61790.67960.4686
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.793, + 0.468, + 0.834 + ], + "angle": 0, + "content": "Table 2. Comparison results of different variants of our method with the baseline. Evidently, our method using EBR retriever (top-3 retrieval) and 1-shot ICL can achieve most best results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Offline Test. Table 2 showcases the numerical results comparing various retrieval and rewriting configurations with shortened prompts. We can see that Prompt [7] and General PR are slightly better than 'Shortened Prompt' in terms" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.257 + ], + "angle": 0, + "content": "of PMS and Image-Align. Without the users' histories, these large language models perform prompt rewriting in arbitrary pathways. Thus, the rewritten prompts these baseline methods generate could totally deviate from the users' preferences. Such hallucination scenarios result in the decrease in terms of ROUGE-L for both Prompt [7] and General PR. This indicates that although general prompt rewriting methods appear to be refining the prompts, their results do not align with users' preferences well. Our method outperforms all baseline methods, i.e. Prompt [7] and General PR, on all metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.892, + 0.318 + ], + "angle": 0, + "content": "A comparison between BM25 and EBR reveals that dense retrieval generally outperforms sparse retrieval, although the difference is relatively small, indicating that both methods can produce satisfactory results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.424 + ], + "angle": 0, + "content": "Further, when contrasting context-independent rewriting with in-context rewriting, it's evident that ICL produces superior outcomes. By integrating ICL with EBR, we achieve absolute improvements of \\(14.2\\%\\), \\(6.9\\%\\) and \\(5.6\\%\\) in terms of ROUGE-L, PMS and Text-Align metrics respectively. This underscores the exceptional performance of personalized prompting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.498 + ], + "angle": 0, + "content": "To check how sensitive our prompt rewriting methods with respect to the length of the prompts before rewriting, we assess our rewriting method using two additional types of prompts with shorter lengths, namely \"Noun Phrase\" and \"Noun\", as detailed in Table 3." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.508, + 0.892, + 0.593 + ], + "angle": 0, + "content": "
Prompt TypeMethodPMS ↑Image-Align↑ROUGE-L↑
NounOriginal0.55370.60870.1770
Personalized PR0.61420.64780.2804
Noun PhraseOriginal0.55540.61460.2459
Personalized PR0.61680.65340.3387
Short SentenceOriginal0.55670.62720.3268
Personalized PR0.61790.67960.4686
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.604, + 0.892, + 0.66 + ], + "angle": 0, + "content": "Table 3. Performance with respect to different prompt lengths, i.e. Noun, Noun Phrase and Short Sentence. Our \"Personalized PR\" equipped with top-3 dense retriever, and 1-shot ICL consistently enhances results, even with only nouns or noun phrases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.78 + ], + "angle": 0, + "content": "These two prompts are derived using spaCy [9] from our dataset, adhering to the principle of minimizing word count while maintaining the main entities. The results displayed in Table 3 show that across all three shortened scale, \"Personalized PR\" outperforms the baseline across all metrics, demonstrating the effectiveness of our methods in recovering user preferences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Online Test. To further validate our approach, we have carried out a single-blinded online evaluation on our website mentioned above. Active users recently in the website are randomly selected to participate in the online test. Upon each prompt input by a participant, there's an equal chance of generating an image using either \"Original Prompt\" or our method's \"Personalized Prompt\". Participants can choose to \"Save\" or \"Delete\" each generated image based" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7733" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.094, + 0.252, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.094, + 0.47, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.278, + 0.47, + 0.307 + ], + "angle": 0, + "content": "Figure 9. Illustration of our online evaluation methods and the improvement of Save Rate from personalized Prompt." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.316, + 0.471, + 0.482 + ], + "angle": 0, + "content": "on their preference. We use their \"Save\" actions to assess our method's effectiveness. In the online test, 247 users generated 905 images, with 433 from \"Original Prompt\" and 472 from \"Personalized Prompt\". The results of the online evaluation, as shown in Figure 9, indicate that users prefer images generated using the \"Personalized Prompt\" over the \"Original Prompt\", with a \\(17.1\\%\\) increase in \"Save\" actions. This suggests that our method aligns better with user preferences, affirming its effectiveness. We anticipate even better results in real-world scenarios with more user information and an improved text-to-image generation method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.492, + 0.231, + 0.509 + ], + "angle": 0, + "content": "5.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.47, + 0.577 + ], + "angle": 0, + "content": "In this section, we ablate \\( k \\) relevant historical prompts used to rewrite personalized prompts and the number of demonstration examples used for in-context learning. We evaluate each experiment on the same setup as in Section 5.1." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.59, + 0.469, + 0.645 + ], + "angle": 0, + "content": "
MethodRetrieval Top-kPMS ↑Image-Align ↑ROUGE-L ↑
Personalized PR + ICL10.60570.67510.4539
30.61790.67960.4686
50.62040.67480.4474
70.62650.66510.4592
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.656, + 0.469, + 0.697 + ], + "angle": 0, + "content": "Table 4. Ablation for Retrieval Top-k. We empirically conduct experiments with respect to \\( k \\in \\{ 1,3,5,7\\} \\) ,keeping the configuration of DENSE retriever and 1-shot ICL." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Retrieval Top-k Ablation. As shown in Table 4, when using 3 most relevant historical prompts for rewriting, we can obtain most best results among all evaluation metrics. We analyze that too more historical prompts could provide redundant information and also too long prompt input to ChatGPT worsens the rewriting performance. Therefore, we choose 3 as the number of retrieval results, a balanced manner between performance and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Number of ICL Demonstrations Ablation. Table 5 shows that 1-shot setting, i.e., given 1 demonstration example for in-context learning, can achieve the best results in general among all the four evaluation metrics regarding both BM25 and EBR. This demonstrates that our prompt rewriting tem-" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.892, + 0.171 + ], + "angle": 0, + "content": "
Method (Retrieval)ICL ShotPMS ↑Image-Align ↑ROUGE-L ↑
Personalized PR (BM25)10.62530.64560.4417
30.62890.65800.4381
50.62360.65710.4226
Personalized PR (EBR)10.61790.67960.4686
30.62740.67080.4354
50.62420.67240.4439
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.181, + 0.892, + 0.238 + ], + "angle": 0, + "content": "Table 5. Ablation for Number of ICL Demonstrations. We experiment with consistent top-3 retrieval both on BM25 and EBR. When we set ICL shot as 1 and use EBR retriever, we observe more superior results appearing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.262, + 0.892, + 0.306 + ], + "angle": 0, + "content": "plate is efficient and effective enough for extracting the personalized preference from numerous historical data of each user." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.334, + 0.77, + 0.35 + ], + "angle": 0, + "content": "6. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.59 + ], + "angle": 0, + "content": "In conclusion, this study has underscored the significance of harnessing historical user behaviors to construct personalized AI content generation. Our strategy aims to refine user prompts by capitalizing on previous user interactions with the system. We have introduced an innovative technique that entails reconfiguring user prompts based on a newly developed, large-scale text-to-image dataset encompassing over 300,000 prompts from 3,115 distinct users. This methodology has proven to augment the expressiveness of user prompts and ensure their alignment with the desired visual outputs. Our empirical results have underscored the supremacy of our techniques over conventional methods. This superiority was corroborated through our novel offline evaluation method and online tests, thereby affirming the efficacy of our approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.73 + ], + "angle": 0, + "content": "Although the outcomes are encouraging, considerable research still lies ahead in this domain. In the realm of personalized text-to-image generation, the incorporation of more personal details like a user's age and gender could bolster performance. The methodologies used could be extended to other LPMs, including LLMs. Additionally, techniques for enhancing search engines, e.g. data source purification and ranking optimization, could be assimilated into these models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.735, + 0.892, + 0.78 + ], + "angle": 0, + "content": "We are confident that our contributions represent advancements towards a more personalized and user-focused artificial intelligence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.804, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported in part by the Research Center for Industries of the Future at Westlake University and Westlake Education Foundation (Grant No. WU2023C017), National Natural Science Foundation of China (Grant Nos. 62171111, 62006245), and Open Research Fund from State Key Laboratory of High Performance Computing of China (Grant No. 202201-15)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7734" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. \"this is my unicorn, fluffy\": Personalizing frozen vision-language representations. In European Conference on Computer Vision, pages 558-577. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.471, + 0.327 + ], + "angle": 0, + "content": "[3] Grégoire Deletang, Anian Ruoss, Paul-Ambroise Duquenne, Elliot Catt, Tim Genewein, Christopher Mattern, Jordi Grau-Moya, Li Kevin Wenliang, Matthew Aitchison, Laurent Orseau, et al. Language modeling is compression. arXiv preprint arXiv:2309.10668, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.471, + 0.398 + ], + "angle": 0, + "content": "[4] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.471, + 0.456 + ], + "angle": 0, + "content": "[5] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.458, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[6] Qingyan Guo, Rui Wang, Junliang Guo, Bei Li, Kaitao Song, Xu Tan, Guoqing Liu, Jiang Bian, and Yujiu Yang. Connecting large language models with evolutionary algorithms yields powerful prompt optimizers, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.471, + 0.557 + ], + "angle": 0, + "content": "[7] Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. arXiv preprint arXiv:2212.09611, 2022. 2, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[8] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[9] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python. 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.47, + 0.714 + ], + "angle": 0, + "content": "[10] Rolf Jagerman, Honglei Zhuang, Zhen Qin, Xuanhui Wang, and Michael Bendersky. Query expansion by prompting large language models. arXiv preprint arXiv:2305.03653, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.799 + ], + "angle": 0, + "content": "[11] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[12] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[13] Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. What makes" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "good in-context examples for gpt-3? arXiv preprint arXiv:2101.06804, 2021.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[14] Jianmo Ni, Chen Qu, Jing Lu, Zhuyun Dai, Gustavo Hernández Abrego, Ji Ma, Vincent Y Zhao, Yi Luan, Keith B Hall, Ming-Wei Chang, et al. Large dual encoders are generalizable retrievers. arXiv preprint arXiv:2112.07899, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[15] OpenAI. Chatgpt. https://chat.openai.com. 5, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[16] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[17] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[18] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[19] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. Stable diffusion v1-5. https://huggingface.co/runwayml/stable-diffusion-v1-5.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.492, + 0.894, + 0.561 + ], + "angle": 0, + "content": "[20] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.563, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[21] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[22] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[23] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[24] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.847, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7735" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[26] Xiao Wang, Sean MacAvaney, Craig Macdonald, and Iadh Ounis. Generative query reformulation for effective adhoc search. arXiv preprint arXiv:2308.00415, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.189 + ], + "angle": 0, + "content": "[27] Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V Le, Denny Zhou, and Xinyun Chen. Large language models as optimizers. arXiv preprint arXiv:2309.03409, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[28] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.468, + 0.29 + ], + "angle": 0, + "content": "[29] Shanshan Zhong, Zhongzhan Huang, Wushao Wen, Jinghui Qin, and Liang Lin. Sur-adapter: Enhancing text-to-image pre-trained diffusion models with large language models. arXiv preprint arXiv:2305.05189, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.468, + 0.346 + ], + "angle": 0, + "content": "[30] Yongchao Zhou, Andrei Loan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. arXiv preprint arXiv:2211.01910, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7736" + } + ] +] \ No newline at end of file diff --git a/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/7aad987a-39f2-4096-8aa1-a539bbaacb9c_origin.pdf b/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/7aad987a-39f2-4096-8aa1-a539bbaacb9c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cfe5b3b18609be1bd5788a69b7d34ba9010b64a6 --- /dev/null +++ b/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/7aad987a-39f2-4096-8aa1-a539bbaacb9c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b891992fc8cf743287e30f7e7db597582b6995aa87171b846f3b5b27be996244 +size 3114604 diff --git a/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/full.md b/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6b4d1105f6ce4ead120b9a56f7e29a4e3cc53ec7 --- /dev/null +++ b/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/full.md @@ -0,0 +1,337 @@ +# Tailored Visions: Enhancing Text-to-Image Generation with Personalized Prompt Rewriting + +Zijie Chen*,1,2 Lichao Zhang*,2 Fangsheng Weng3 Lili Pan†,4 Zhenzhong Lan†,2 + +$^{1}$ Zhejiang University $^{2}$ Westlake University $^{3}$ Scietrain + +4University of Electronic Science and Technology of China + +{chenzijie, zhanglichao, lanzhenzhong}@westlake.edu.cn, lilipan@uestc.edu.cn + +# Abstract + +Despite significant progress in the field, it is still challenging to create personalized visual representations that align closely with the desires and preferences of individual users. This process requires users to articulate their ideas in words that are both comprehensible to the models and accurately capture their vision, posing difficulties for many users. In this paper, we tackle this challenge by leveraging historical user interactions with the system to enhance user prompts. We propose a novel approach that involves rewriting user prompts based on a newly collected large-scale text-to-image dataset with over 300k prompts from 3115 users. Our rewriting model enhances the expressiveness and alignment of user prompts with their intended visual outputs. Experimental results demonstrate the superiority of our methods over baseline approaches, as evidenced in our new offline evaluation method and online tests. Our code and dataset are available at https://github.com/zzjchen/Tailored-Visions + +# 1. Introduction + +Increasingly large and powerful foundation models [1, 4, 20] are trained through self-supervised learning. These large pretrained models (LPMs) serve as efficient compressors [3], condensing vast amounts of internet data. This compression enables the convenient extraction of the knowledge encoded within these models via natural language descriptions. Despite being in its infancy, this approach exhibits potentials to surpass traditional search engines as a superior source for knowledge and information acquisition. + +Akin to refining queries for search engines, prompts given to LPMs must also be carefully crafted. However, + +# User History + +![](images/0c29a6ba42cd59e518f9ac1365305c8bd34906d820f1116331ceb3d1bfa4d039.jpg) + +Large Italianate House with Beautiful Breathing on the Shores of Lake Como-Oil Painting in the Style of Claude Monet trending in artstation + +![](images/10570ac38f7617b9c4a0994a2766e3fe4276a1ea9fe548e83013a5277b28f9e8.jpg) + +A Magical Forest in the morning, oil painting in the style of Claude Monet trending in artstation. + +![](images/8380a50d87796cb66afde073133bb1fec76ccf743209e413d9d3129e2b856bb3.jpg) +Figure 1. Comparison between our personalized prompt rewriting method and the standard prompt rewriting method. Our technique excels at incorporating user preferences, such as "oil paintings by artists," while methods that lack a historical context frequently generate content that may not align with the user's desires. + +Tulip bouquet. Oil Painting, bright color, highly detailed, Vincent Van Gogh, trending in artstation + +![](images/38706e5319762fe75a8b8a748a09d4d3f7f651e880a7b9d017df2d10e4fe1cef.jpg) + +Still life of a hamburger and some fruits on a napkin, detailed oil painting in the style of Paul Cezanme, trenging in artstation + +# Generated Results + +Original Prompt: Walking street at night + +![](images/73eae4661ebe7256dc79bff46275a5f0d7b28713abbde65dd5a0d28a81923a3c.jpg) + +![](images/4a0599bb42bdcdc2fcd27370ad9bfe24218cbb225a153e8b9698b529de7daf36.jpg) + +t... a mesmerizing walking street at night...with a touch of Italianate c-harm, reminiscent of Claude Monet's oil painting style. + +![](images/82dc172a8227b2aae5153f43cc4fe88ca96bf3f9d4ef04ae2a26072561e5bf2e.jpg) + +et an astronaut, inspired by Claude tce cMonet_rides a horse in a captivati ng scene,..., fusion of magic and natury... + +![](images/00d34ff0a9222c9fc756649c2f67cf6d26e16eaf3252350fe05784d545d3a372.jpg) +W/o History + +An astronaut, adorned in futuristic gear, gracefully maneuvers on horseback through an otherwise landscape bathed in the ethereal glow of distant stars, capturing the juxtaposition of advanced space exploration and timeless equestrian elegance. + +the complexity of prompts, the unpredictability of model responses compared to traditional search engines present unique challenges. Significant research efforts [10, 29] have been made to comprehend how LPMs react to various prompts, with some studies examining the feasibility of rewriting prompts for specificity. However, without access to users' personal data and behavior, tailoring the prompt to meet the user's needs accurately remains challenging. + +Our research addresses this issue by integrating user preference information into prompt rewriting. The primary obstacle in personalized query rewriting is the absence of a dataset containing text-to-image prompts with personalized information. To overcome this, we have assembled a large dataset encompassing over 300k text-to-image histories from 3,115 users. We rewrite user prompts using their query history, although we had limited access to personal + +information, leaving room for further research. Another significant challenge is the evaluation of rewritten queries. To evaluate their efficacy, we've developed a new offline method that uses multiple metrics to measure how well our rewriting models can recover the original user query from the ChatGPT-shortened version. + +Our paper's contributions are threefold: + +1. We have compiled a large Personalized Image Prompt (PIP) dataset and made it publicly available to aid future research in the field. +2. We experimented with two query rewriting techniques and proposed a new query evaluation method to assess their performance. +3. We propose a new benchmark for personalized text-to-image generation, which promotes the standardization of this field. + +While there is still a considerable distance to cover before we can create a perfect prompt encapsulating both the user's requirements and the model's capabilities, we believe our research provides a critical stepping stone in this ongoing exploration. + +# 2. Related Work + +This section provides an overview of prior work on text-to-image generation, personalization for such generation, and prompt rewriting. It's important to note that our review is aimed more at offering sufficient background knowledge rather than exhaustive coverage of all related works. + +# 2.1. Text-to-Image Generation + +Large text-to-image generation models can generate high-fidelity image synthesis and achieve a deep level of language understanding. DALL-E [17] uses a VQ-VAE transformer-based method to learn a visual codebook in the first stage and then trains autoregressive transformers on sequences of text tokens followed by image tokens in the second stage. DALL-E2 [18] introduces latent diffusion models to generate various images by conditioning on CLIP text latents and CLIP image embeddings generated by a prior model. Imagen [22] discovers that a larger language model with more parameters trained on text-only data improves the quality of text-to-image generation. Late developments like stable-diffusion (SD) [20] proposes to generate images effectively in latent space significantly lowering computational costs. Furthermore, SD designs a conditional mechanism to complete class-conditional, text-to-image and layout-to-image models. Furthermore, ControlNet [28] accomplishes certain function by conditioning on multi-modal data, e.g., edge, sketching, pose, segmentation, depth etc., which unavoidably involving additional condition-generation modalities. + +Despite these models' ability to generate high-fidelity images, they often fail to meet the precise needs of the users. Text-to-image generation is more like a game of chance. + +# 2.2. Personalization for Text-to-Image Generation + +Recently, personalization approaches based on text-to-image models have taken a set of images of a concept and generated variations of the concept. Specifically, some methods optimize a set of text embeddings. For example, [2] involves pseudo-word embeddings by a set encoder to provide personalization and Textual inversion [4] composes the concept into language sentences and performed as a personalized creation. Some methods finetune the diffusion model. For example, DreamBooth [21] finetunes the text-to-image diffusion model with shared parallel branches. To speed up, CustomDiffusion [12] reduces the amount of finetuned parameters, and Tewel et al. [25] locks the subject's cross-attention key to its superordinate category to align with visual concepts. Moreover, an additional encoder is trained to map concept images to its textual representation by [5] and [24]. + +Existing studies have three key limitations: they demand extra images and fine-tuning of text-to-image models with limited scope for new concepts; they can't learn from user interaction history and need detailed user prompts; and there's a lack of public, personalized text-to-image datasets that truly reflect user preferences. + +# 2.3. Prompt Rewriting + +Recently, researchers have found that optimizing prompts can boost the performance of LLMs on several NLP tasks and even search systems. For examples, Guo et al. [6] connect the LLM with evolutionary algorithms to generate an optimized prompt from parent prompts, without any gradient calculation. Yang et al. [27] propose to use LLM as an optimizer by generating new prompts based on a trajectory of previously generated prompts in each optimization step with the objective of maximizing the accuracy of the task. In the work [30], LLMs serve as models for engineering work like inference, scoring, and resampling. In search systems, LLMs are used to generate query expansion terms by [10], while they are used to reformulate query by Wang et al. [26] instead. + +For T2I generation, one relatively close work, SUR-adapter [29] learns to align the representations between simple prompt and complex prompt by an adapter. [7] optimize prompt through general rewriting. However, Neither of above works utilized personalized information for improving prompts. + +# 3. Personalized Image-Prompt (PIP) Dataset + +# 3.1. Dataset Collection + +The Personalized Image-Prompt (PIP) dataset is the first large-scale personalized generated image-text dataset. The original data are collected from a public website that we + +![](images/b53d3f0ace518c1792c225648fb3133c3e9b4ff4a03ff39e3ee74274fc8120f6.jpg) +Figure 2. Dataset creation process. We split our dataset into training and testing sets and summarize each prompts in the test set using ChatGPT. + +![](images/90524e7cc1928da278f0c921c62a80a62038caa90a31f265e961d6ddc427daa0.jpg) +Figure 3. Dataset statistics and distribution. Left: Proportion of users based on the varying number of historical prompts they have. Note that each user has a minimum of 18 historical prompts, as we have excluded those with fewer prompts from the dataset. Right: Proportion of prompts based on their varying lengths. Best view in color. + +host to provide open-domain text-to-image generation to users. PIP dataset includes 3115 users and 300,237 text-to-image histories generated by these users using SD v1-5 [20] and an internal fine-tuned version of SD v1-5. Each user in PIP has created 18 or more images and provided at least 12 different prompts. + +Figure 2 illustrates the process of creating the dataset. For each individual user, we randomly choose two prompts to serve as test prompts, with the remaining prompts allocated as training prompts (historical user query). The purpose of using random selection instead of the most recent generated prompts is to enhance the diversity of our test data. Subsequently, we employ ChatGPT to condense the test prompts, ensuring they only include the primary object or scene, as depicted in Figure 2. We shorten the prompts into three scales, i.e., contain only nouns, noun phases or short sentences respectively. + +In the ensuing experiment, each test prompt in the test + +set will be considered as the input prompt $x_{t}$ for every user $u$ , with the original prompts serving as the ground truth that reflect the user's authentic preferences. The remaining prompts are utilized as training samples. + +The PIP dataset consists of 300,237 image-prompt pairs, personally categorized by 3,115 users. These pairs are divided into 294,007 training samples and 6,230 test samples. + +# 3.2. Dataset Statistics and Distribution + +In this section, we showcase the data statistics that depict the quality and diversity of PIP. We specifically illustrate data distributions of the number of prompts and prompt length for each user, and delve deeper into the content of the prompt through a word cloud representation. + +Each data sample contains a prompt, the generated images,UserID, Image size, and URL, as illustrated Figure 4. + +![](images/43077457f808c0e52fea4fda2ab5a8692fd0a2fb55acddd97f10219425179767.jpg) +Figure 4. Two examples from a user history, containing Image, Prompt, User ID, Image size and URL. + +![](images/bd813f8f9f0a153cb777a4c569c6f59f29379514ba524dddfae443e8c807904e.jpg) +Figure 5. Word cloud visualization of top 250 keywords sampled from the PIP dataset. + +In PIP dataset, each user contributes at least 18 images, as depicted in the left section of Figure 3. This results in a long-tail distribution. The prompts have an average word count of 27.53. The length of the prompts, ranging from 1 to 284 words, also follows a long-tail distribution, as seen in the right section of Figure 3. Despite the presence of about 2500 prompts that exceed the 75-word limit of SD, we retain them to maintain the integrity of user preferences. + +![](images/dbd0a4e9b0786cf271544e01d273a106606844b5a3aa170e58345b53f3985e9c.jpg) +Figure 6. Text similarity between user histories and user preferences. There is a great deal of diversity in user preferences in the PIP dataset. In the zoomed-in version of the similarity map on the left, users' history prompts are highly similar to their preferences. + +Figure 5 presents the 250 most frequently used words or phrases. The frequency of these words is determined by the highest TF-IDF value across all users. The word cloud reveals that these words describe various image attributes, such as objects, styles, quality, and colors. This variety underlines the high diversity present within the prompt content of the PIP dataset. + +In addition, we visualize the text similarities between the histories and preferences of 100 users in Figure 6. For each user $u$ in PIP dataset, we summarize his preference $P_{u}$ into 5 phrases from his history prompts using ChatGPT. For two users $u, v$ , the similarity of $u$ 's histories and $v$ 's preferences is defined as the mean of text similarities between $u$ 's history prompts and $v$ 's preference $P_{v}$ . The text similarity is calculated using GTR-T5-large [14]. Furthermore, we visualize the text similarities between the history prompts of a random user and 100 user preferences in the left part of Figure 6. This shows that text-to-image users have different preferences and the preference $p_{u}$ we summarized successfully captures the key feature the user prefers. + +# 3.3. Evaluation Metrics + +We present two metrics to evaluate prompt rewriting methods in terms of how the rewritten results are aligned to users' preferences, namely Preference Matching Score (PMS) and Image-Align. + +Preference Matching Score (PMS). PMS calculates the CLIPScore [8] between generated image and user's preference $P_{u}$ . It measures how the generated image aligns with the user's preference. + +$$ +\mathbf {P M S} = \frac {w}{N} \sum_ {u = 1} ^ {N} \max \left(\cos \left(\mathbf {E m} \left(I _ {u} ^ {\prime}\right), \mathbf {E m} \left(P _ {u}\right)\right), 0\right) \tag {1} +$$ + +where $P_{u}$ is the user $u$ 's preference, $I_{u}^{\prime}$ is the generated image correspondingly, $N$ is total user number (i.e. 3115). $\mathbf{Em}$ means the embedding extracted by using CLIP. $w = 2.5$ is a scaling constant. + +Image-Align. It measures the similarity between the generated image and the ground-truth image. Image-Align quantifies how closely the current created image aligns with the user's truly saved image. The similarity between two images are calculated using CLIP [16]. + +Apart from these metrics, we also adopt ROUGE-L to evaluate prompt rewriting methods in our experiment. Calculating ROUGE-L between the rewritten prompt against the original prompt measures the ability of prompt rewriting methods to recover the original prompt. We set $\beta = 5$ to emphasize the recall of generated prompts. + +# 4. Personalized Prompt Rewriting + +The basic idea of our personalization method is to rewrite the input prompt, considering user preferences gleaned from past user interactions. The full pipeline of our Personalized Prompt Rewriting (Personalized PR) method is depicted in the left part of Figure 7. If a user $u$ input a prompt $x_{t}$ , a retriever $\mathbf{Ret}(x_{t},\mathcal{Q}_{t})$ retrieves prompts from the user's historical prompt set $\mathcal{Q}_t$ , using $x_{t}$ as a query. Based on the retrieval result $\mathcal{R}_t = \mathbf{Ret}(x_t,\mathcal{Q}_t)$ , the rewritten Rew rewrites the input prompt to generate a personalized prompt $x_{t}^{\prime} = \mathbf{Rew}(x_{t},\mathcal{R}_{t})$ . Finally, the text-to-image generation model $\mathbf{G}$ produces the image $I_{t}^{\prime} = \mathbf{G}\left(x_{t}^{\prime},\epsilon\right)$ from the rewritten prompt, where $\epsilon$ is a random noise vector. + +# 4.1. Retrieval and Ranking + +In the retrieval stage, given the input prompt $x_{t}$ , the retriever $\mathbf{Ret}(x_t,\mathcal{Q}_t)$ retrieves relevant prompts from historical prompt set $\mathcal{Q}_t$ , using $x_{t}$ as a query. + +By analyzing user prompts, we have noticed that users tend to construct prompts that involve objects, their attributes, and the relationships between objects. In the works [23], [11], a query for image retrieval is defined to include objects, attributes of objects, and relations between objects. Inspired by this, we suspect users have the habit of using attributes and some objects, such as background, to express their preferences. To confirm this, we visualize the word cloud of the top 250 frequent words in the text prompts of all users, as shown in the right part of Figure 5. In Figure 5, we find some attributes, such as "cute", "golden", and "beautiful" appear in prompts with high frequency, as well as some objects, such as "mountain", "sea," and "sky". Intuitively, we can use the current prompt $x_{t}$ to locate the relevant history prompts that include the same or similar attributes or objects. + +To locate relevant prompts, two retrieval methods are used: dense and sparse. In dense retrieval, we choose the prompt $x_{t}$ and calculate its textual embedding $\mathbf{Em}(x_t)$ using CLIP's text encoder, also the text encoder in Stable Dif + +![](images/daab4cf2bf36d4a86877b4f34983a4edecc7278d55fcd4614047885933afa765.jpg) +Figure 7. Left: Pipeline of Personalized Prompt Rewriting (Personalized PR), including Retriever, Rewrter and T2I model to generate personalized images from user histories. Right: Illustration of context-dependent prompt rewriting. We present a specific example for better understanding the procedure of context-dependent prompt rewriting. + +![](images/aaa4eb49f2d6b4ce451c9064dbc9c17bfb61e5eeb5f8192a8b7feee3dfaf2a47.jpg) + +fusion [20]. We suspect the prompts with similar visual attributes and objects will be close to each other in the text embedding space. To confirm this, we visualize some retrieval results in Figure 8. The three nearest neighbors of $\mathbf{Em}(x_t)$ are prompts that are semantically related. For example, if the input prompt is "Hobbit homes", the three most relevant prompts would include the words "village", "city", and "house". This dense retrieval method is also referred to as embedding-based retrieval (EBR). In sparse retrieval, we use BM25 to locate relevant prompts that include the same visual attributes and objects. + +In the above retrieval, we rank relevant prompts in EBR-based or BM25-based ranking, depending on the retrieval ways. In EBR-based ranking, we rank the relevant prompts based on their embedding similarity with the query $x_{t}$ . For similarity measuring, we choose cosine similarity as it is a commonly used similarity measure in embedding learning. In BM25-based ranking, BM25 scores are used for similarity measures. Consequently, we obtain the top $k$ relevant user queries $\mathcal{R}_t = \{r_1,\dots,r_k\}$ . + +# 4.2. Rewriting + +The procedure of context-independent rewriting leverages pertinent queries $\mathcal{R}_t = \{r_1, \dots, r_k\}$ , and employs ChatGPT to encapsulate user preferences and rewrite the prompt directly. These queries $\mathcal{R}_t$ are organized based on their relevance to $x_t$ . + +In the context-dependent scenario, we initially create a collection of demonstration examples $\mathcal{E} = \{e_1,\dots,e_L\}$ using manual design. We then select a small subset of these examples to serve as demonstrations for each rewriting task. + +Given the issue of order sensitivity in in-context learning, as highlighted in the study [13], we arrange the demonstration examples in a descending sequence based on their proximity to the input prompt $x_{t}$ . The in-context rewriting process we employ is illustrated in the right section of Figure 7. + +# 5. Experiment + +We carried out experiments for prompt rewriting methods on our PIP dataset. We validate our method through both offline and online evaluation. And we further analyze the number of historical prompts for best extracting the users' preference by ablating top retrieval. For offline evaluation, we use the aforementioned three metrics: PMS, ImageAlign, and ROUGE-L. + +For online evaluation, we carried out single blind experiment for recently active users on our website. Real user feedback is collected to evaluate our method. + +# 5.1. Implementation Details + +Details of our Personalized PR method are as follows. In retrieval, we choose the relevant text prompt number as $k = 3$ . We use ChatGPT [15] as our rewritten. An input example for context-independent rewriting is shown in Table 1. For in-context rewriting, we set $L = 5$ and randomly select one demonstration example for each rewriting task, unless otherwise specified. + +Unless other specified, all the experiments use EBR to + +![](images/ed392e07c2ad5a2addc8c1a4d07be7e27a5118f71453e511774688a7bee3a7aa.jpg) + +![](images/716a9b1917b36ca46d889228cfa2866a6964b8dac09a833351e2fec31edd9248.jpg) + +![](images/16c63df41afbc5349041e50b0b29241f68276a46429ca7d08c6884fbf8d6a6ff.jpg) +Figure 8. Qualitative analysis of personalized retrieval and rewriting. + +![](images/056a3f22eed6b764c316e0b468df60c7008e78f729c4523bd2746f47d49882a9.jpg) + +![](images/59d1473b9693830ffc8e7411cdb60b42cdec05cb5e0446ef4126920faa795816.jpg) + +retrieve historical prompts, and one-shot in-context learning to rewrite the shortened prompt. + +# ChatGPT Input Template for Context-independent Rewriting + +Prompt in text-to-image generation describes the detailed attributes of the object user plans to draw. User's preference in text-to-image generation is shown in history prompts. + +Given 3 history prompts, your task is to rewrite the current prompt so that it matches the user's preference. The rewritten prompt should retain primary objects in the original prompt and conform to the user's preference. Please avoid being too diffused and restrict your output within 70 words. + +The history prompts are: $\{\mathcal{R}_t\}$ + +The current prompt is: $\{x_{t}\}$ + +The rewritten prompt (one sentence less than 70 words) is: + +Table 1. Input template for context-independent rewriting. + +We use Stable Diffusion (SD) v1-5 [19] as our text-to-image generation model for all methods. SD v1-5 is sampled using PNDM scheduler in 50 steps and setting the classifier-free guidance scale to 7.0. + +# 5.2. Comparison Baselines + +We compare our method with two baseline methods, namely Prompt [7] and General Prompt Rewriting (General PR). Both are general text-to-image prompt rewriting methods completely overlooking users' preferences. + +Promptist [7]. Promptist uses the GPT-2 as the base model and has trained it on a self-collected text-to-image prompt dataset with 360K prompts, by using supervised fine-tuning method and reinforcement learning. + +General PR. We prompt ChatGPT [15] to perform general prompt rewriting in a normal manner. In practice, we input + +ChatGPT with the current prompt $x_{t}$ without any user historical information. Apart from the absence of histories, everything else of the prompt template for ChatGPT remains the same as the template in Table 1. + +# 5.3. Qualitative Analysis + +To illustrate the effectiveness of our retrieval methods, we provide visualizations of the retrieved relevant image-prompt pairs in the left section of Figure 8. These examples are drawn from the experimental results of three different test prompts from three users. The retrieved histories exhibit a high degree of similarity with their corresponding queries in terms of objects, attributes, as well as the overall style or mood of the image. + +This effectively demonstrates the proficiency of our retriever in sourcing relevant user histories, thereby providing a robust reference for our rewriter to carry out personalized prompt rewriting. + +In the right section of Figure 8, we display the generation outcomes of both "Shortened Prompt" and "Personalized Prompt". The "Shortened Prompt" column exhibits the results produced from the shortened prompts, while the "Personalized Prompt" column features the rewritten prompts of our method along with their corresponding generated images. It's evident from these displays that images our method generates are more inclined towards user preferences based on their histories, a testament to the expressive power of our rewritten prompt. For instance, when the query "Hobbit homes" is used (as seen in the first row), we observe the user's preferred style across three images, all capturing the mood of mountainous scenery and depicting the Hobbit homes within a consistent landscape. + +# 5.4.Quantitative Comparison + +To further examine the effectiveness of our Personalized PR method, we conduct offline and online quantitative evaluation, comparing different settings to the baseline methods without enhanced prompt rewritten on the test samples we created. + +
MethodRetrieverPMS ↑Image-Align↑ROUGE-L↑
Shortened Prompt-0.55670.62720.3268
Promptist [7]-0.58580.64810.2947
General PR-0.59960.59120.2082
Personalized PRBM250.61250.65810.3942
EBR0.60830.64850.4137
Personalized PR+ ICLBM250.62530.64560.4417
EBR0.61790.67960.4686
+ +Offline Test. Table 2 showcases the numerical results comparing various retrieval and rewriting configurations with shortened prompts. We can see that Prompt [7] and General PR are slightly better than 'Shortened Prompt' in terms + +of PMS and Image-Align. Without the users' histories, these large language models perform prompt rewriting in arbitrary pathways. Thus, the rewritten prompts these baseline methods generate could totally deviate from the users' preferences. Such hallucination scenarios result in the decrease in terms of ROUGE-L for both Prompt [7] and General PR. This indicates that although general prompt rewriting methods appear to be refining the prompts, their results do not align with users' preferences well. Our method outperforms all baseline methods, i.e. Prompt [7] and General PR, on all metrics. + +A comparison between BM25 and EBR reveals that dense retrieval generally outperforms sparse retrieval, although the difference is relatively small, indicating that both methods can produce satisfactory results. + +Further, when contrasting context-independent rewriting with in-context rewriting, it's evident that ICL produces superior outcomes. By integrating ICL with EBR, we achieve absolute improvements of $14.2\%$ , $6.9\%$ and $5.6\%$ in terms of ROUGE-L, PMS and Text-Align metrics respectively. This underscores the exceptional performance of personalized prompting. + +To check how sensitive our prompt rewriting methods with respect to the length of the prompts before rewriting, we assess our rewriting method using two additional types of prompts with shorter lengths, namely "Noun Phrase" and "Noun", as detailed in Table 3. + +Table 2. Comparison results of different variants of our method with the baseline. Evidently, our method using EBR retriever (top-3 retrieval) and 1-shot ICL can achieve most best results. + +
Prompt TypeMethodPMS ↑Image-Align↑ROUGE-L↑
NounOriginal0.55370.60870.1770
Personalized PR0.61420.64780.2804
Noun PhraseOriginal0.55540.61460.2459
Personalized PR0.61680.65340.3387
Short SentenceOriginal0.55670.62720.3268
Personalized PR0.61790.67960.4686
+ +Table 3. Performance with respect to different prompt lengths, i.e. Noun, Noun Phrase and Short Sentence. Our "Personalized PR" equipped with top-3 dense retriever, and 1-shot ICL consistently enhances results, even with only nouns or noun phrases. + +These two prompts are derived using spaCy [9] from our dataset, adhering to the principle of minimizing word count while maintaining the main entities. The results displayed in Table 3 show that across all three shortened scale, "Personalized PR" outperforms the baseline across all metrics, demonstrating the effectiveness of our methods in recovering user preferences. + +Online Test. To further validate our approach, we have carried out a single-blinded online evaluation on our website mentioned above. Active users recently in the website are randomly selected to participate in the online test. Upon each prompt input by a participant, there's an equal chance of generating an image using either "Original Prompt" or our method's "Personalized Prompt". Participants can choose to "Save" or "Delete" each generated image based + +![](images/4dfb5d41b9245ab81ef5cad70471c6ce74710280455d16d04d90e61f114b7193.jpg) +Figure 9. Illustration of our online evaluation methods and the improvement of Save Rate from personalized Prompt. + +![](images/9819b24a6b39f99a5060e3c58cd2cc319cabda9fe82545fc51b0a288e94476c8.jpg) + +on their preference. We use their "Save" actions to assess our method's effectiveness. In the online test, 247 users generated 905 images, with 433 from "Original Prompt" and 472 from "Personalized Prompt". The results of the online evaluation, as shown in Figure 9, indicate that users prefer images generated using the "Personalized Prompt" over the "Original Prompt", with a $17.1\%$ increase in "Save" actions. This suggests that our method aligns better with user preferences, affirming its effectiveness. We anticipate even better results in real-world scenarios with more user information and an improved text-to-image generation method. + +# 5.5. Ablation Study + +In this section, we ablate $k$ relevant historical prompts used to rewrite personalized prompts and the number of demonstration examples used for in-context learning. We evaluate each experiment on the same setup as in Section 5.1. + +
MethodRetrieval Top-kPMS ↑Image-Align ↑ROUGE-L ↑
Personalized PR + ICL10.60570.67510.4539
30.61790.67960.4686
50.62040.67480.4474
70.62650.66510.4592
+ +Retrieval Top-k Ablation. As shown in Table 4, when using 3 most relevant historical prompts for rewriting, we can obtain most best results among all evaluation metrics. We analyze that too more historical prompts could provide redundant information and also too long prompt input to ChatGPT worsens the rewriting performance. Therefore, we choose 3 as the number of retrieval results, a balanced manner between performance and efficiency. + +Number of ICL Demonstrations Ablation. Table 5 shows that 1-shot setting, i.e., given 1 demonstration example for in-context learning, can achieve the best results in general among all the four evaluation metrics regarding both BM25 and EBR. This demonstrates that our prompt rewriting tem- + +Table 4. Ablation for Retrieval Top-k. We empirically conduct experiments with respect to $k \in \{ 1,3,5,7\}$ ,keeping the configuration of DENSE retriever and 1-shot ICL. + +
Method (Retrieval)ICL ShotPMS ↑Image-Align ↑ROUGE-L ↑
Personalized PR (BM25)10.62530.64560.4417
30.62890.65800.4381
50.62360.65710.4226
Personalized PR (EBR)10.61790.67960.4686
30.62740.67080.4354
50.62420.67240.4439
+ +Table 5. Ablation for Number of ICL Demonstrations. We experiment with consistent top-3 retrieval both on BM25 and EBR. When we set ICL shot as 1 and use EBR retriever, we observe more superior results appearing. + +plate is efficient and effective enough for extracting the personalized preference from numerous historical data of each user. + +# 6. Conclusion and Future Work + +In conclusion, this study has underscored the significance of harnessing historical user behaviors to construct personalized AI content generation. Our strategy aims to refine user prompts by capitalizing on previous user interactions with the system. We have introduced an innovative technique that entails reconfiguring user prompts based on a newly developed, large-scale text-to-image dataset encompassing over 300,000 prompts from 3,115 distinct users. This methodology has proven to augment the expressiveness of user prompts and ensure their alignment with the desired visual outputs. Our empirical results have underscored the supremacy of our techniques over conventional methods. This superiority was corroborated through our novel offline evaluation method and online tests, thereby affirming the efficacy of our approach. + +Although the outcomes are encouraging, considerable research still lies ahead in this domain. In the realm of personalized text-to-image generation, the incorporation of more personal details like a user's age and gender could bolster performance. The methodologies used could be extended to other LPMs, including LLMs. Additionally, techniques for enhancing search engines, e.g. data source purification and ranking optimization, could be assimilated into these models. + +We are confident that our contributions represent advancements towards a more personalized and user-focused artificial intelligence. + +Acknowledgements. This work is supported in part by the Research Center for Industries of the Future at Westlake University and Westlake Education Foundation (Grant No. WU2023C017), National Natural Science Foundation of China (Grant Nos. 62171111, 62006245), and Open Research Fund from State Key Laboratory of High Performance Computing of China (Grant No. 202201-15). + +# References + +[1] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1 +[2] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. "this is my unicorn, fluffy": Personalizing frozen vision-language representations. In European Conference on Computer Vision, pages 558-577. Springer, 2022. 2 +[3] Grégoire Deletang, Anian Ruoss, Paul-Ambroise Duquenne, Elliot Catt, Tim Genewein, Christopher Mattern, Jordi Grau-Moya, Li Kevin Wenliang, Matthew Aitchison, Laurent Orseau, et al. Language modeling is compression. arXiv preprint arXiv:2309.10668, 2023. 1 +[4] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1, 2 +[5] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 2 +[6] Qingyan Guo, Rui Wang, Junliang Guo, Bei Li, Kaitao Song, Xu Tan, Guoqing Liu, Jiang Bian, and Yujiu Yang. Connecting large language models with evolutionary algorithms yields powerful prompt optimizers, 2023. 2 +[7] Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. arXiv preprint arXiv:2212.09611, 2022. 2, 6, 7, 1 +[8] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 4 +[9] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python. 2020. 7 +[10] Rolf Jagerman, Honglei Zhuang, Zhen Qin, Xuanhui Wang, and Michael Bendersky. Query expansion by prompting large language models. arXiv preprint arXiv:2305.03653, 2023. 1, 2 +[11] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 4 +[12] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2 +[13] Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. What makes + +good in-context examples for gpt-3? arXiv preprint arXiv:2101.06804, 2021.5 +[14] Jianmo Ni, Chen Qu, Jing Lu, Zhuyun Dai, Gustavo Hernández Abrego, Ji Ma, Vincent Y Zhao, Yi Luan, Keith B Hall, Ming-Wei Chang, et al. Large dual encoders are generalizable retrievers. arXiv preprint arXiv:2112.07899, 2021. 4 +[15] OpenAI. Chatgpt. https://chat.openai.com. 5, 6, 3 +[16] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4 +[17] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2 +[18] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[19] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. Stable diffusion v1-5. https://huggingface.co/runwayml/stable-diffusion-v1-5.6 +[20] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3, 5 +[21] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2 +[22] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2 +[23] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 4 +[24] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 2 +[25] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2 + +[26] Xiao Wang, Sean MacAvaney, Craig Macdonald, and Iadh Ounis. Generative query reformulation for effective adhoc search. arXiv preprint arXiv:2308.00415, 2023. 2 +[27] Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V Le, Denny Zhou, and Xinyun Chen. Large language models as optimizers. arXiv preprint arXiv:2309.03409, 2023. 2 +[28] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2 +[29] Shanshan Zhong, Zhongzhan Huang, Wushao Wen, Jinghui Qin, and Liang Lin. Sur-adapter: Enhancing text-to-image pre-trained diffusion models with large language models. arXiv preprint arXiv:2305.05189, 2023. 1, 2 +[30] Yongchao Zhou, Andrei Loan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. arXiv preprint arXiv:2211.01910, 2022. 2 \ No newline at end of file diff --git a/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/images.zip b/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..67995bd0b7bf5a3885e4860e17ea4898a827cf53 --- /dev/null +++ b/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69eccc0256459a133e1d34187bb341f502f4a934e9ee092c74b44f72633c57e6 +size 761035 diff --git a/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/layout.json b/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..efc4befe80e67ce7cb540c27e16b08cb8ae4e816 --- /dev/null +++ b/2024/Tailored Visions_ Enhancing Text-to-Image Generation with Personalized Prompt Rewriting/layout.json @@ -0,0 +1,8474 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 71, + 102, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 102, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 71, + 102, + 523, + 140 + ], + "type": "text", + "content": "Tailored Visions: Enhancing Text-to-Image Generation with Personalized Prompt Rewriting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 159, + 515, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 159, + 515, + 177 + ], + "spans": [ + { + "bbox": [ + 93, + 159, + 515, + 177 + ], + "type": "text", + "content": "Zijie Chen*,1,2 Lichao Zhang*,2 Fangsheng Weng3 Lili Pan†,4 Zhenzhong Lan†,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 162, + 180, + 446, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 180, + 446, + 195 + ], + "spans": [ + { + "bbox": [ + 162, + 180, + 446, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 180, + 446, + 195 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 162, + 180, + 446, + 195 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 162, + 180, + 446, + 195 + ], + "type": "text", + "content": "Westlake University " + }, + { + "bbox": [ + 162, + 180, + 446, + 195 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 162, + 180, + 446, + 195 + ], + "type": "text", + "content": "Scietrain" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 162, + 195, + 447, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 195, + 447, + 209 + ], + "spans": [ + { + "bbox": [ + 162, + 195, + 447, + 209 + ], + "type": "text", + "content": "4University of Electronic Science and Technology of China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 94, + 211, + 511, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 211, + 511, + 222 + ], + "spans": [ + { + "bbox": [ + 94, + 211, + 511, + 222 + ], + "type": "text", + "content": "{chenzijie, zhanglichao, lanzhenzhong}@westlake.edu.cn, lilipan@uestc.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 250, + 192, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 250, + 192, + 262 + ], + "spans": [ + { + "bbox": [ + 143, + 250, + 192, + 262 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 276, + 290, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 290, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 290, + 491 + ], + "type": "text", + "content": "Despite significant progress in the field, it is still challenging to create personalized visual representations that align closely with the desires and preferences of individual users. This process requires users to articulate their ideas in words that are both comprehensible to the models and accurately capture their vision, posing difficulties for many users. In this paper, we tackle this challenge by leveraging historical user interactions with the system to enhance user prompts. We propose a novel approach that involves rewriting user prompts based on a newly collected large-scale text-to-image dataset with over 300k prompts from 3115 users. Our rewriting model enhances the expressiveness and alignment of user prompts with their intended visual outputs. Experimental results demonstrate the superiority of our methods over baseline approaches, as evidenced in our new offline evaluation method and online tests. Our code and dataset are available at https://github.com/zzjchen/Tailored-Visions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 516, + 128, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 516, + 128, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 516, + 128, + 529 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 536, + 287, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 655 + ], + "type": "text", + "content": "Increasingly large and powerful foundation models [1, 4, 20] are trained through self-supervised learning. These large pretrained models (LPMs) serve as efficient compressors [3], condensing vast amounts of internet data. This compression enables the convenient extraction of the knowledge encoded within these models via natural language descriptions. Despite being in its infancy, this approach exhibits potentials to surpass traditional search engines as a superior source for knowledge and information acquisition." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 657, + 287, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 287, + 681 + ], + "type": "text", + "content": "Akin to refining queries for search engines, prompts given to LPMs must also be carefully crafted. However," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 336, + 251, + 373, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 251, + 373, + 260 + ], + "spans": [ + { + "bbox": [ + 336, + 251, + 373, + 260 + ], + "type": "text", + "content": "User History" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 272, + 345, + 308 + ], + "blocks": [ + { + "bbox": [ + 310, + 272, + 345, + 308 + ], + "lines": [ + { + "bbox": [ + 310, + 272, + 345, + 308 + ], + "spans": [ + { + "bbox": [ + 310, + 272, + 345, + 308 + ], + "type": "image", + "image_path": "0c29a6ba42cd59e518f9ac1365305c8bd34906d820f1116331ceb3d1bfa4d039.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 347, + 275, + 400, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 275, + 400, + 305 + ], + "spans": [ + { + "bbox": [ + 347, + 275, + 400, + 305 + ], + "type": "text", + "content": "Large Italianate House with Beautiful Breathing on the Shores of Lake Como-Oil Painting in the Style of Claude Monet trending in artstation" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 310, + 310, + 345, + 346 + ], + "blocks": [ + { + "bbox": [ + 310, + 310, + 345, + 346 + ], + "lines": [ + { + "bbox": [ + 310, + 310, + 345, + 346 + ], + "spans": [ + { + "bbox": [ + 310, + 310, + 345, + 346 + ], + "type": "image", + "image_path": "10570ac38f7617b9c4a0994a2766e3fe4276a1ea9fe548e83013a5277b28f9e8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 347, + 318, + 400, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 318, + 400, + 338 + ], + "spans": [ + { + "bbox": [ + 347, + 318, + 400, + 338 + ], + "type": "text", + "content": "A Magical Forest in the morning, oil painting in the style of Claude Monet trending in artstation." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 310, + 347, + 346, + 384 + ], + "blocks": [ + { + "bbox": [ + 310, + 347, + 346, + 384 + ], + "lines": [ + { + "bbox": [ + 310, + 347, + 346, + 384 + ], + "spans": [ + { + "bbox": [ + 310, + 347, + 346, + 384 + ], + "type": "image", + "image_path": "8380a50d87796cb66afde073133bb1fec76ccf743209e413d9d3129e2b856bb3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 448, + 547, + 504 + ], + "lines": [ + { + "bbox": [ + 305, + 448, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 305, + 448, + 547, + 504 + ], + "type": "text", + "content": "Figure 1. Comparison between our personalized prompt rewriting method and the standard prompt rewriting method. Our technique excels at incorporating user preferences, such as \"oil paintings by artists,\" while methods that lack a historical context frequently generate content that may not align with the user's desires." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 347, + 355, + 400, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 355, + 400, + 376 + ], + "spans": [ + { + "bbox": [ + 347, + 355, + 400, + 376 + ], + "type": "text", + "content": "Tulip bouquet. Oil Painting, bright color, highly detailed, Vincent Van Gogh, trending in artstation" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 310, + 388, + 346, + 423 + ], + "blocks": [ + { + "bbox": [ + 310, + 388, + 346, + 423 + ], + "lines": [ + { + "bbox": [ + 310, + 388, + 346, + 423 + ], + "spans": [ + { + "bbox": [ + 310, + 388, + 346, + 423 + ], + "type": "image", + "image_path": "38706e5319762fe75a8b8a748a09d4d3f7f651e880a7b9d017df2d10e4fe1cef.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 347, + 390, + 400, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 390, + 400, + 416 + ], + "spans": [ + { + "bbox": [ + 347, + 390, + 400, + 416 + ], + "type": "text", + "content": "Still life of a hamburger and some fruits on a napkin, detailed oil painting in the style of Paul Cezanme, trenging in artstation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 445, + 252, + 496, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 445, + 252, + 496, + 260 + ], + "spans": [ + { + "bbox": [ + 445, + 252, + 496, + 260 + ], + "type": "text", + "content": "Generated Results" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 420, + 262, + 455, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 420, + 262, + 455, + 273 + ], + "spans": [ + { + "bbox": [ + 420, + 262, + 455, + 273 + ], + "type": "text", + "content": "Original Prompt: Walking street at night" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 414, + 273, + 462, + 319 + ], + "blocks": [ + { + "bbox": [ + 414, + 273, + 462, + 319 + ], + "lines": [ + { + "bbox": [ + 414, + 273, + 462, + 319 + ], + "spans": [ + { + "bbox": [ + 414, + 273, + 462, + 319 + ], + "type": "image", + "image_path": "73eae4661ebe7256dc79bff46275a5f0d7b28713abbde65dd5a0d28a81923a3c.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 462, + 262, + 535, + 319 + ], + "blocks": [ + { + "bbox": [ + 462, + 262, + 535, + 319 + ], + "lines": [ + { + "bbox": [ + 462, + 262, + 535, + 319 + ], + "spans": [ + { + "bbox": [ + 462, + 262, + 535, + 319 + ], + "type": "image", + "image_path": "4a0599bb42bdcdc2fcd27370ad9bfe24218cbb225a153e8b9698b529de7daf36.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 405, + 319, + 471, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 319, + 471, + 338 + ], + "spans": [ + { + "bbox": [ + 405, + 319, + 471, + 338 + ], + "type": "text", + "content": "t... a mesmerizing walking street at night...with a touch of Italianate c-harm, reminiscent of Claude Monet's oil painting style." + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 414, + 339, + 461, + 385 + ], + "blocks": [ + { + "bbox": [ + 414, + 339, + 461, + 385 + ], + "lines": [ + { + "bbox": [ + 414, + 339, + 461, + 385 + ], + "spans": [ + { + "bbox": [ + 414, + 339, + 461, + 385 + ], + "type": "image", + "image_path": "82dc172a8227b2aae5153f43cc4fe88ca96bf3f9d4ef04ae2a26072561e5bf2e.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 461, + 319, + 538, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 461, + 319, + 538, + 338 + ], + "spans": [ + { + "bbox": [ + 461, + 319, + 538, + 338 + ], + "type": "text", + "content": "et an astronaut, inspired by Claude tce cMonet_rides a horse in a captivati ng scene,..., fusion of magic and natury..." + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 484, + 339, + 529, + 386 + ], + "blocks": [ + { + "bbox": [ + 484, + 339, + 529, + 386 + ], + "lines": [ + { + "bbox": [ + 484, + 339, + 529, + 386 + ], + "spans": [ + { + "bbox": [ + 484, + 339, + 529, + 386 + ], + "type": "image", + "image_path": "00d34ff0a9222c9fc756649c2f67cf6d26e16eaf3252350fe05784d545d3a372.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 537, + 347, + 543, + 378 + ], + "lines": [ + { + "bbox": [ + 537, + 347, + 543, + 378 + ], + "spans": [ + { + "bbox": [ + 537, + 347, + 543, + 378 + ], + "type": "text", + "content": "W/o History" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 474, + 385, + 538, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 385, + 538, + 423 + ], + "spans": [ + { + "bbox": [ + 474, + 385, + 538, + 423 + ], + "type": "text", + "content": "An astronaut, adorned in futuristic gear, gracefully maneuvers on horseback through an otherwise landscape bathed in the ethereal glow of distant stars, capturing the juxtaposition of advanced space exploration and timeless equestrian elegance." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 521, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 616 + ], + "type": "text", + "content": "the complexity of prompts, the unpredictability of model responses compared to traditional search engines present unique challenges. Significant research efforts [10, 29] have been made to comprehend how LPMs react to various prompts, with some studies examining the feasibility of rewriting prompts for specificity. However, without access to users' personal data and behavior, tailoring the prompt to meet the user's needs accurately remains challenging." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Our research addresses this issue by integrating user preference information into prompt rewriting. The primary obstacle in personalized query rewriting is the absence of a dataset containing text-to-image prompts with personalized information. To overcome this, we have assembled a large dataset encompassing over 300k text-to-image histories from 3,115 users. We rewrite user prompts using their query history, although we had limited access to personal" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 690, + 127, + 701 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 690, + 127, + 701 + ], + "spans": [ + { + "bbox": [ + 58, + 690, + 127, + 701 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 60, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 60, + 702, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7727" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "type": "text", + "content": "information, leaving room for further research. Another significant challenge is the evaluation of rewritten queries. To evaluate their efficacy, we've developed a new offline method that uses multiple metrics to measure how well our rewriting models can recover the original user query from the ChatGPT-shortened version." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 144, + 220, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 144, + 220, + 156 + ], + "spans": [ + { + "bbox": [ + 59, + 144, + 220, + 156 + ], + "type": "text", + "content": "Our paper's contributions are threefold:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 156, + 287, + 262 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 156, + 287, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 287, + 191 + ], + "type": "text", + "content": "1. We have compiled a large Personalized Image Prompt (PIP) dataset and made it publicly available to aid future research in the field." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 192, + 287, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 192, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 287, + 227 + ], + "type": "text", + "content": "2. We experimented with two query rewriting techniques and proposed a new query evaluation method to assess their performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 228, + 287, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 287, + 262 + ], + "type": "text", + "content": "3. We propose a new benchmark for personalized text-to-image generation, which promotes the standardization of this field." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 263, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 287, + 323 + ], + "type": "text", + "content": "While there is still a considerable distance to cover before we can create a perfect prompt encapsulating both the user's requirements and the model's capabilities, we believe our research provides a critical stepping stone in this ongoing exploration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 330, + 134, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 330, + 134, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 330, + 134, + 342 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 345, + 287, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 345, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 287, + 404 + ], + "type": "text", + "content": "This section provides an overview of prior work on text-to-image generation, personalization for such generation, and prompt rewriting. It's important to note that our review is aimed more at offering sufficient background knowledge rather than exhaustive coverage of all related works." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 411, + 192, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 192, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 192, + 423 + ], + "type": "text", + "content": "2.1. Text-to-Image Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 426, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 676 + ], + "type": "text", + "content": "Large text-to-image generation models can generate high-fidelity image synthesis and achieve a deep level of language understanding. DALL-E [17] uses a VQ-VAE transformer-based method to learn a visual codebook in the first stage and then trains autoregressive transformers on sequences of text tokens followed by image tokens in the second stage. DALL-E2 [18] introduces latent diffusion models to generate various images by conditioning on CLIP text latents and CLIP image embeddings generated by a prior model. Imagen [22] discovers that a larger language model with more parameters trained on text-only data improves the quality of text-to-image generation. Late developments like stable-diffusion (SD) [20] proposes to generate images effectively in latent space significantly lowering computational costs. Furthermore, SD designs a conditional mechanism to complete class-conditional, text-to-image and layout-to-image models. Furthermore, ControlNet [28] accomplishes certain function by conditioning on multi-modal data, e.g., edge, sketching, pose, segmentation, depth etc., which unavoidably involving additional condition-generation modalities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Despite these models' ability to generate high-fidelity images, they often fail to meet the precise needs of the users. Text-to-image generation is more like a game of chance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 541, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 541, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 541, + 84 + ], + "type": "text", + "content": "2.2. Personalization for Text-to-Image Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 87, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 87, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 87, + 545, + 278 + ], + "type": "text", + "content": "Recently, personalization approaches based on text-to-image models have taken a set of images of a concept and generated variations of the concept. Specifically, some methods optimize a set of text embeddings. For example, [2] involves pseudo-word embeddings by a set encoder to provide personalization and Textual inversion [4] composes the concept into language sentences and performed as a personalized creation. Some methods finetune the diffusion model. For example, DreamBooth [21] finetunes the text-to-image diffusion model with shared parallel branches. To speed up, CustomDiffusion [12] reduces the amount of finetuned parameters, and Tewel et al. [25] locks the subject's cross-attention key to its superordinate category to align with visual concepts. Moreover, an additional encoder is trained to map concept images to its textual representation by [5] and [24]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 279, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 350 + ], + "type": "text", + "content": "Existing studies have three key limitations: they demand extra images and fine-tuning of text-to-image models with limited scope for new concepts; they can't learn from user interaction history and need detailed user prompts; and there's a lack of public, personalized text-to-image datasets that truly reflect user preferences." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 356, + 414, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 356, + 414, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 356, + 414, + 369 + ], + "type": "text", + "content": "2.3. Prompt Rewriting" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 372, + 545, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 538 + ], + "type": "text", + "content": "Recently, researchers have found that optimizing prompts can boost the performance of LLMs on several NLP tasks and even search systems. For examples, Guo et al. [6] connect the LLM with evolutionary algorithms to generate an optimized prompt from parent prompts, without any gradient calculation. Yang et al. [27] propose to use LLM as an optimizer by generating new prompts based on a trajectory of previously generated prompts in each optimization step with the objective of maximizing the accuracy of the task. In the work [30], LLMs serve as models for engineering work like inference, scoring, and resampling. In search systems, LLMs are used to generate query expansion terms by [10], while they are used to reformulate query by Wang et al. [26] instead." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 539, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 545, + 611 + ], + "type": "text", + "content": "For T2I generation, one relatively close work, SUR-adapter [29] learns to align the representations between simple prompt and complex prompt by an adapter. [7] optimize prompt through general rewriting. However, Neither of above works utilized personalized information for improving prompts." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 620, + 535, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 620, + 535, + 633 + ], + "spans": [ + { + "bbox": [ + 305, + 620, + 535, + 633 + ], + "type": "text", + "content": "3. Personalized Image-Prompt (PIP) Dataset" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 639, + 414, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 414, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 414, + 651 + ], + "type": "text", + "content": "3.1. Dataset Collection" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 658, + 545, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 658, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 304, + 658, + 545, + 695 + ], + "type": "text", + "content": "The Personalized Image-Prompt (PIP) dataset is the first large-scale personalized generated image-text dataset. The original data are collected from a public website that we" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 701, + 373, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 373, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 373, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 317, + 701, + 373, + 713 + ], + "type": "text", + "content": " https://zmrj.art/" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7728" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 286, + 213 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 213 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 213 + ], + "type": "image", + "image_path": "b53d3f0ace518c1792c225648fb3133c3e9b4ff4a03ff39e3ee74274fc8120f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 229, + 288, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 288, + 264 + ], + "type": "text", + "content": "Figure 2. Dataset creation process. We split our dataset into training and testing sets and summarize each prompts in the test set using ChatGPT." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 276, + 285, + 390 + ], + "blocks": [ + { + "bbox": [ + 52, + 276, + 285, + 390 + ], + "lines": [ + { + "bbox": [ + 52, + 276, + 285, + 390 + ], + "spans": [ + { + "bbox": [ + 52, + 276, + 285, + 390 + ], + "type": "image", + "image_path": "90524e7cc1928da278f0c921c62a80a62038caa90a31f265e961d6ddc427daa0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 410, + 288, + 477 + ], + "lines": [ + { + "bbox": [ + 46, + 410, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 288, + 477 + ], + "type": "text", + "content": "Figure 3. Dataset statistics and distribution. Left: Proportion of users based on the varying number of historical prompts they have. Note that each user has a minimum of 18 historical prompts, as we have excluded those with fewer prompts from the dataset. Right: Proportion of prompts based on their varying lengths. Best view in color." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "type": "text", + "content": "host to provide open-domain text-to-image generation to users. PIP dataset includes 3115 users and 300,237 text-to-image histories generated by these users using SD v1-5 [20] and an internal fine-tuned version of SD v1-5. Each user in PIP has created 18 or more images and provided at least 12 different prompts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 288, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 701 + ], + "type": "text", + "content": "Figure 2 illustrates the process of creating the dataset. For each individual user, we randomly choose two prompts to serve as test prompts, with the remaining prompts allocated as training prompts (historical user query). The purpose of using random selection instead of the most recent generated prompts is to enhance the diversity of our test data. Subsequently, we employ ChatGPT to condense the test prompts, ensuring they only include the primary object or scene, as depicted in Figure 2. We shorten the prompts into three scales, i.e., contain only nouns, noun phases or short sentences respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "content": "In the ensuing experiment, each test prompt in the test" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "set will be considered as the input prompt " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " for every user " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ", with the original prompts serving as the ground truth that reflect the user's authentic preferences. The remaining prompts are utilized as training samples." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 121, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 156 + ], + "type": "text", + "content": "The PIP dataset consists of 300,237 image-prompt pairs, personally categorized by 3,115 users. These pairs are divided into 294,007 training samples and 6,230 test samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 163, + 489, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 163, + 489, + 174 + ], + "spans": [ + { + "bbox": [ + 306, + 163, + 489, + 174 + ], + "type": "text", + "content": "3.2. Dataset Statistics and Distribution" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 180, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 240 + ], + "type": "text", + "content": "In this section, we showcase the data statistics that depict the quality and diversity of PIP. We specifically illustrate data distributions of the number of prompts and prompt length for each user, and delve deeper into the content of the prompt through a word cloud representation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 240, + 545, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 240, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 305, + 240, + 545, + 265 + ], + "type": "text", + "content": "Each data sample contains a prompt, the generated images,UserID, Image size, and URL, as illustrated Figure 4." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 276, + 541, + 398 + ], + "blocks": [ + { + "bbox": [ + 308, + 276, + 541, + 398 + ], + "lines": [ + { + "bbox": [ + 308, + 276, + 541, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 541, + 398 + ], + "type": "image", + "image_path": "43077457f808c0e52fea4fda2ab5a8692fd0a2fb55acddd97f10219425179767.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 408, + 545, + 430 + ], + "lines": [ + { + "bbox": [ + 306, + 408, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 408, + 545, + 430 + ], + "type": "text", + "content": "Figure 4. Two examples from a user history, containing Image, Prompt, User ID, Image size and URL." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 450, + 543, + 582 + ], + "blocks": [ + { + "bbox": [ + 309, + 450, + 543, + 582 + ], + "lines": [ + { + "bbox": [ + 309, + 450, + 543, + 582 + ], + "spans": [ + { + "bbox": [ + 309, + 450, + 543, + 582 + ], + "type": "image", + "image_path": "bd813f8f9f0a153cb777a4c569c6f59f29379514ba524dddfae443e8c807904e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 590, + 545, + 611 + ], + "lines": [ + { + "bbox": [ + 306, + 590, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 590, + 545, + 611 + ], + "type": "text", + "content": "Figure 5. Word cloud visualization of top 250 keywords sampled from the PIP dataset." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "In PIP dataset, each user contributes at least 18 images, as depicted in the left section of Figure 3. This results in a long-tail distribution. The prompts have an average word count of 27.53. The length of the prompts, ranging from 1 to 284 words, also follows a long-tail distribution, as seen in the right section of Figure 3. Despite the presence of about 2500 prompts that exceed the 75-word limit of SD, we retain them to maintain the integrity of user preferences." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7729" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 286, + 173 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 286, + 173 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 286, + 173 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 286, + 173 + ], + "type": "image", + "image_path": "dbd0a4e9b0786cf271544e01d273a106606844b5a3aa170e58345b53f3985e9c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 180, + 288, + 227 + ], + "lines": [ + { + "bbox": [ + 46, + 180, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 288, + 227 + ], + "type": "text", + "content": "Figure 6. Text similarity between user histories and user preferences. There is a great deal of diversity in user preferences in the PIP dataset. In the zoomed-in version of the similarity map on the left, users' history prompts are highly similar to their preferences." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 247, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 287, + 331 + ], + "type": "text", + "content": "Figure 5 presents the 250 most frequently used words or phrases. The frequency of these words is determined by the highest TF-IDF value across all users. The word cloud reveals that these words describe various image attributes, such as objects, styles, quality, and colors. This variety underlines the high diversity present within the prompt content of the PIP dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "text", + "content": "In addition, we visualize the text similarities between the histories and preferences of 100 users in Figure 6. For each user " + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "text", + "content": " in PIP dataset, we summarize his preference " + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "inline_equation", + "content": "P_{u}" + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "text", + "content": " into 5 phrases from his history prompts using ChatGPT. For two users " + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "inline_equation", + "content": "u, v" + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "text", + "content": ", the similarity of " + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "text", + "content": "'s histories and " + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "text", + "content": "'s preferences is defined as the mean of text similarities between " + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "text", + "content": "'s history prompts and " + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "text", + "content": "'s preference " + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "inline_equation", + "content": "P_{v}" + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "text", + "content": ". The text similarity is calculated using GTR-T5-large [14]. Furthermore, we visualize the text similarities between the history prompts of a random user and 100 user preferences in the left part of Figure 6. This shows that text-to-image users have different preferences and the preference " + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "inline_equation", + "content": "p_{u}" + }, + { + "bbox": [ + 46, + 332, + 289, + 489 + ], + "type": "text", + "content": " we summarized successfully captures the key feature the user prefers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 498, + 160, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 160, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 160, + 510 + ], + "type": "text", + "content": "3.3. Evaluation Metrics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 518, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 567 + ], + "type": "text", + "content": "We present two metrics to evaluate prompt rewriting methods in terms of how the rewritten results are aligned to users' preferences, namely Preference Matching Score (PMS) and Image-Align." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 567, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 615 + ], + "type": "text", + "content": "Preference Matching Score (PMS). PMS calculates the CLIPScore [8] between generated image and user's preference " + }, + { + "bbox": [ + 46, + 567, + 287, + 615 + ], + "type": "inline_equation", + "content": "P_{u}" + }, + { + "bbox": [ + 46, + 567, + 287, + 615 + ], + "type": "text", + "content": ". It measures how the generated image aligns with the user's preference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 622, + 287, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 622, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 622, + 287, + 656 + ], + "type": "interline_equation", + "content": "\\mathbf {P M S} = \\frac {w}{N} \\sum_ {u = 1} ^ {N} \\max \\left(\\cos \\left(\\mathbf {E m} \\left(I _ {u} ^ {\\prime}\\right), \\mathbf {E m} \\left(P _ {u}\\right)\\right), 0\\right) \\tag {1}", + "image_path": "6a91120ce4d6d315e1ff4ce84c8d36060191c6cae03c21f0dcc07968b8388490.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "P_{u}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " is the user " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "'s preference, " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "I_{u}^{\\prime}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " is the generated image correspondingly, " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " is total user number (i.e. 3115). " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Em}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " means the embedding extracted by using CLIP. " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "w = 2.5" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " is a scaling constant." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "Image-Align. It measures the similarity between the generated image and the ground-truth image. Image-Align quantifies how closely the current created image aligns with the user's truly saved image. The similarity between two images are calculated using CLIP [16]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 133, + 546, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 205 + ], + "type": "text", + "content": "Apart from these metrics, we also adopt ROUGE-L to evaluate prompt rewriting methods in our experiment. Calculating ROUGE-L between the rewritten prompt against the original prompt measures the ability of prompt rewriting methods to recover the original prompt. We set " + }, + { + "bbox": [ + 304, + 133, + 546, + 205 + ], + "type": "inline_equation", + "content": "\\beta = 5" + }, + { + "bbox": [ + 304, + 133, + 546, + 205 + ], + "type": "text", + "content": " to emphasize the recall of generated prompts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 216, + 484, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 216, + 484, + 229 + ], + "spans": [ + { + "bbox": [ + 305, + 216, + 484, + 229 + ], + "type": "text", + "content": "4. Personalized Prompt Rewriting" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "text", + "content": "The basic idea of our personalization method is to rewrite the input prompt, considering user preferences gleaned from past user interactions. The full pipeline of our Personalized Prompt Rewriting (Personalized PR) method is depicted in the left part of Figure 7. If a user " + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "text", + "content": " input a prompt " + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "text", + "content": ", a retriever " + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{Ret}(x_{t},\\mathcal{Q}_{t})" + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "text", + "content": " retrieves prompts from the user's historical prompt set " + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_t" + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "text", + "content": ", using " + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "text", + "content": " as a query. Based on the retrieval result " + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t = \\mathbf{Ret}(x_t,\\mathcal{Q}_t)" + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "text", + "content": ", the rewritten Rew rewrites the input prompt to generate a personalized prompt " + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "inline_equation", + "content": "x_{t}^{\\prime} = \\mathbf{Rew}(x_{t},\\mathcal{R}_{t})" + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "text", + "content": ". Finally, the text-to-image generation model " + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "text", + "content": " produces the image " + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "inline_equation", + "content": "I_{t}^{\\prime} = \\mathbf{G}\\left(x_{t}^{\\prime},\\epsilon\\right)" + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "text", + "content": " from the rewritten prompt, where " + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 236, + 546, + 381 + ], + "type": "text", + "content": " is a random noise vector." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 389, + 436, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 389, + 436, + 402 + ], + "spans": [ + { + "bbox": [ + 305, + 389, + 436, + 402 + ], + "type": "text", + "content": "4.1. Retrieval and Ranking" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 407, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 443 + ], + "type": "text", + "content": "In the retrieval stage, given the input prompt " + }, + { + "bbox": [ + 304, + 407, + 545, + 443 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 407, + 545, + 443 + ], + "type": "text", + "content": ", the retriever " + }, + { + "bbox": [ + 304, + 407, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{Ret}(x_t,\\mathcal{Q}_t)" + }, + { + "bbox": [ + 304, + 407, + 545, + 443 + ], + "type": "text", + "content": " retrieves relevant prompts from historical prompt set " + }, + { + "bbox": [ + 304, + 407, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_t" + }, + { + "bbox": [ + 304, + 407, + 545, + 443 + ], + "type": "text", + "content": ", using " + }, + { + "bbox": [ + 304, + 407, + 545, + 443 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 407, + 545, + 443 + ], + "type": "text", + "content": " as a query." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 444, + 546, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 546, + 635 + ], + "type": "text", + "content": "By analyzing user prompts, we have noticed that users tend to construct prompts that involve objects, their attributes, and the relationships between objects. In the works [23], [11], a query for image retrieval is defined to include objects, attributes of objects, and relations between objects. Inspired by this, we suspect users have the habit of using attributes and some objects, such as background, to express their preferences. To confirm this, we visualize the word cloud of the top 250 frequent words in the text prompts of all users, as shown in the right part of Figure 5. In Figure 5, we find some attributes, such as \"cute\", \"golden\", and \"beautiful\" appear in prompts with high frequency, as well as some objects, such as \"mountain\", \"sea,\" and \"sky\". Intuitively, we can use the current prompt " + }, + { + "bbox": [ + 304, + 444, + 546, + 635 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 444, + 546, + 635 + ], + "type": "text", + "content": " to locate the relevant history prompts that include the same or similar attributes or objects." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 635, + 545, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 545, + 683 + ], + "type": "text", + "content": "To locate relevant prompts, two retrieval methods are used: dense and sparse. In dense retrieval, we choose the prompt " + }, + { + "bbox": [ + 304, + 635, + 545, + 683 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 635, + 545, + 683 + ], + "type": "text", + "content": " and calculate its textual embedding " + }, + { + "bbox": [ + 304, + 635, + 545, + 683 + ], + "type": "inline_equation", + "content": "\\mathbf{Em}(x_t)" + }, + { + "bbox": [ + 304, + 635, + 545, + 683 + ], + "type": "text", + "content": " using CLIP's text encoder, also the text encoder in Stable Dif" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 692, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 692, + 545, + 713 + ], + "type": "text", + "content": "The image generated from the prompt " + }, + { + "bbox": [ + 306, + 692, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 306, + 692, + 545, + 713 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 306, + 692, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 306, + 692, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 306, + 692, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{t} = \\mathbf{G}(x_{t},\\epsilon)" + }, + { + "bbox": [ + 306, + 692, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7730" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 175, + 264 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 175, + 264 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 175, + 264 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 175, + 264 + ], + "type": "image", + "image_path": "daab4cf2bf36d4a86877b4f34983a4edecc7278d55fcd4614047885933afa765.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 284, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 284, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 547, + 319 + ], + "type": "text", + "content": "Figure 7. Left: Pipeline of Personalized Prompt Rewriting (Personalized PR), including Retriever, Rewrter and T2I model to generate personalized images from user histories. Right: Illustration of context-dependent prompt rewriting. We present a specific example for better understanding the procedure of context-dependent prompt rewriting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 209, + 69, + 544, + 260 + ], + "blocks": [ + { + "bbox": [ + 209, + 69, + 544, + 260 + ], + "lines": [ + { + "bbox": [ + 209, + 69, + 544, + 260 + ], + "spans": [ + { + "bbox": [ + 209, + 69, + 544, + 260 + ], + "type": "image", + "image_path": "aaa4eb49f2d6b4ce451c9064dbc9c17bfb61e5eeb5f8192a8b7feee3dfaf2a47.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 335, + 289, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 289, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 289, + 467 + ], + "type": "text", + "content": "fusion [20]. We suspect the prompts with similar visual attributes and objects will be close to each other in the text embedding space. To confirm this, we visualize some retrieval results in Figure 8. The three nearest neighbors of " + }, + { + "bbox": [ + 46, + 335, + 289, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{Em}(x_t)" + }, + { + "bbox": [ + 46, + 335, + 289, + 467 + ], + "type": "text", + "content": " are prompts that are semantically related. For example, if the input prompt is \"Hobbit homes\", the three most relevant prompts would include the words \"village\", \"city\", and \"house\". This dense retrieval method is also referred to as embedding-based retrieval (EBR). In sparse retrieval, we use BM25 to locate relevant prompts that include the same visual attributes and objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 468, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 577 + ], + "type": "text", + "content": "In the above retrieval, we rank relevant prompts in EBR-based or BM25-based ranking, depending on the retrieval ways. In EBR-based ranking, we rank the relevant prompts based on their embedding similarity with the query " + }, + { + "bbox": [ + 46, + 468, + 287, + 577 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 468, + 287, + 577 + ], + "type": "text", + "content": ". For similarity measuring, we choose cosine similarity as it is a commonly used similarity measure in embedding learning. In BM25-based ranking, BM25 scores are used for similarity measures. Consequently, we obtain the top " + }, + { + "bbox": [ + 46, + 468, + 287, + 577 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 468, + 287, + 577 + ], + "type": "text", + "content": " relevant user queries " + }, + { + "bbox": [ + 46, + 468, + 287, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t = \\{r_1,\\dots,r_k\\}" + }, + { + "bbox": [ + 46, + 468, + 287, + 577 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 587, + 118, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 118, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 118, + 601 + ], + "type": "text", + "content": "4.2. Rewriting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 604, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 287, + 664 + ], + "type": "text", + "content": "The procedure of context-independent rewriting leverages pertinent queries " + }, + { + "bbox": [ + 46, + 604, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t = \\{r_1, \\dots, r_k\\}" + }, + { + "bbox": [ + 46, + 604, + 287, + 664 + ], + "type": "text", + "content": ", and employs ChatGPT to encapsulate user preferences and rewrite the prompt directly. These queries " + }, + { + "bbox": [ + 46, + 604, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t" + }, + { + "bbox": [ + 46, + 604, + 287, + 664 + ], + "type": "text", + "content": " are organized based on their relevance to " + }, + { + "bbox": [ + 46, + 604, + 287, + 664 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 46, + 604, + 287, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "In the context-dependent scenario, we initially create a collection of demonstration examples " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{e_1,\\dots,e_L\\}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " using manual design. We then select a small subset of these examples to serve as demonstrations for each rewriting task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 335, + 547, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 547, + 397 + ], + "type": "text", + "content": "Given the issue of order sensitivity in in-context learning, as highlighted in the study [13], we arrange the demonstration examples in a descending sequence based on their proximity to the input prompt " + }, + { + "bbox": [ + 304, + 335, + 547, + 397 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 335, + 547, + 397 + ], + "type": "text", + "content": ". The in-context rewriting process we employ is illustrated in the right section of Figure 7." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 408, + 383, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 408, + 383, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 408, + 383, + 422 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "content": "We carried out experiments for prompt rewriting methods on our PIP dataset. We validate our method through both offline and online evaluation. And we further analyze the number of historical prompts for best extracting the users' preference by ablating top retrieval. For offline evaluation, we use the aforementioned three metrics: PMS, ImageAlign, and ROUGE-L." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 513, + 547, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 547, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 547, + 549 + ], + "type": "text", + "content": "For online evaluation, we carried out single blind experiment for recently active users on our website. Real user feedback is collected to evaluate our method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 558, + 440, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 558, + 440, + 571 + ], + "spans": [ + { + "bbox": [ + 305, + 558, + 440, + 571 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 574, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 545, + 658 + ], + "type": "text", + "content": "Details of our Personalized PR method are as follows. In retrieval, we choose the relevant text prompt number as " + }, + { + "bbox": [ + 304, + 574, + 545, + 658 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 304, + 574, + 545, + 658 + ], + "type": "text", + "content": ". We use ChatGPT [15] as our rewritten. An input example for context-independent rewriting is shown in Table 1. For in-context rewriting, we set " + }, + { + "bbox": [ + 304, + 574, + 545, + 658 + ], + "type": "inline_equation", + "content": "L = 5" + }, + { + "bbox": [ + 304, + 574, + 545, + 658 + ], + "type": "text", + "content": " and randomly select one demonstration example for each rewriting task, unless otherwise specified." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 659, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 545, + 673 + ], + "type": "text", + "content": "Unless other specified, all the experiments use EBR to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 305, + 682, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 682, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 682, + 545, + 713 + ], + "type": "text", + "content": "†The input for in-context rewriting is only different from context-independent rewriting in terms of the presence of demonstration example. See supplementary for detail." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "7731" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 539, + 209 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 539, + 209 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 539, + 209 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 539, + 209 + ], + "type": "image", + "image_path": "ed392e07c2ad5a2addc8c1a4d07be7e27a5118f71453e511774688a7bee3a7aa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 224, + 340, + 343 + ], + "blocks": [ + { + "bbox": [ + 51, + 224, + 340, + 343 + ], + "lines": [ + { + "bbox": [ + 51, + 224, + 340, + 343 + ], + "spans": [ + { + "bbox": [ + 51, + 224, + 340, + 343 + ], + "type": "image", + "image_path": "716a9b1917b36ca46d889228cfa2866a6964b8dac09a833351e2fec31edd9248.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 52, + 352, + 541, + 466 + ], + "blocks": [ + { + "bbox": [ + 52, + 352, + 541, + 466 + ], + "lines": [ + { + "bbox": [ + 52, + 352, + 541, + 466 + ], + "spans": [ + { + "bbox": [ + 52, + 352, + 541, + 466 + ], + "type": "image", + "image_path": "16c63df41afbc5349041e50b0b29241f68276a46429ca7d08c6884fbf8d6a6ff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 479, + 421, + 491 + ], + "lines": [ + { + "bbox": [ + 171, + 479, + 421, + 491 + ], + "spans": [ + { + "bbox": [ + 171, + 479, + 421, + 491 + ], + "type": "text", + "content": "Figure 8. Qualitative analysis of personalized retrieval and rewriting." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 355, + 224, + 539, + 340 + ], + "blocks": [ + { + "bbox": [ + 355, + 224, + 539, + 340 + ], + "lines": [ + { + "bbox": [ + 355, + 224, + 539, + 340 + ], + "spans": [ + { + "bbox": [ + 355, + 224, + 539, + 340 + ], + "type": "image", + "image_path": "056a3f22eed6b764c316e0b468df60c7008e78f729c4523bd2746f47d49882a9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 356, + 352, + 539, + 468 + ], + "blocks": [ + { + "bbox": [ + 356, + 352, + 539, + 468 + ], + "lines": [ + { + "bbox": [ + 356, + 352, + 539, + 468 + ], + "spans": [ + { + "bbox": [ + 356, + 352, + 539, + 468 + ], + "type": "image", + "image_path": "59d1473b9693830ffc8e7411cdb60b42cdec05cb5e0446ef4126920faa795816.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 513, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 287, + 536 + ], + "type": "text", + "content": "retrieve historical prompts, and one-shot in-context learning to rewrite the shortened prompt." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 552, + 281, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 281, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 281, + 563 + ], + "type": "text", + "content": "ChatGPT Input Template for Context-independent Rewriting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 569, + 282, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 282, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 282, + 600 + ], + "type": "text", + "content": "Prompt in text-to-image generation describes the detailed attributes of the object user plans to draw. User's preference in text-to-image generation is shown in history prompts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 601, + 281, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 281, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 281, + 652 + ], + "type": "text", + "content": "Given 3 history prompts, your task is to rewrite the current prompt so that it matches the user's preference. The rewritten prompt should retain primary objects in the original prompt and conform to the user's preference. Please avoid being too diffused and restrict your output within 70 words." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 163, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 163, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 163, + 664 + ], + "type": "text", + "content": "The history prompts are: " + }, + { + "bbox": [ + 55, + 653, + 163, + 664 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{R}_t\\}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 153, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 153, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 153, + 676 + ], + "type": "text", + "content": "The current prompt is: " + }, + { + "bbox": [ + 55, + 665, + 153, + 676 + ], + "type": "inline_equation", + "content": "\\{x_{t}\\}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 676, + 256, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 676, + 256, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 676, + 256, + 686 + ], + "type": "text", + "content": "The rewritten prompt (one sentence less than 70 words) is:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 699, + 272, + 710 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 699, + 272, + 710 + ], + "spans": [ + { + "bbox": [ + 61, + 699, + 272, + 710 + ], + "type": "text", + "content": "Table 1. Input template for context-independent rewriting." + } + ] + } + ], + "index": 13, + "type": "text" + }, + { + "bbox": [ + 305, + 512, + 545, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 512, + 545, + 560 + ], + "spans": [ + { + "bbox": [ + 305, + 512, + 545, + 560 + ], + "type": "text", + "content": "We use Stable Diffusion (SD) v1-5 [19] as our text-to-image generation model for all methods. SD v1-5 is sampled using PNDM scheduler in 50 steps and setting the classifier-free guidance scale to 7.0." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 571, + 432, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 571, + 432, + 583 + ], + "spans": [ + { + "bbox": [ + 306, + 571, + 432, + 583 + ], + "type": "text", + "content": "5.2. Comparison Baselines" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 591, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 545, + 639 + ], + "type": "text", + "content": "We compare our method with two baseline methods, namely Prompt [7] and General Prompt Rewriting (General PR). Both are general text-to-image prompt rewriting methods completely overlooking users' preferences." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 688 + ], + "type": "text", + "content": "Promptist [7]. Promptist uses the GPT-2 as the base model and has trained it on a self-collected text-to-image prompt dataset with 360K prompts, by using supervised fine-tuning method and reinforcement learning." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "General PR. We prompt ChatGPT [15] to perform general prompt rewriting in a normal manner. In practice, we input" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7732" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "ChatGPT with the current prompt " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " without any user historical information. Apart from the absence of histories, everything else of the prompt template for ChatGPT remains the same as the template in Table 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 126, + 165, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 126, + 165, + 138 + ], + "spans": [ + { + "bbox": [ + 47, + 126, + 165, + 138 + ], + "type": "text", + "content": "5.3. Qualitative Analysis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 144, + 286, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 286, + 239 + ], + "type": "text", + "content": "To illustrate the effectiveness of our retrieval methods, we provide visualizations of the retrieved relevant image-prompt pairs in the left section of Figure 8. These examples are drawn from the experimental results of three different test prompts from three users. The retrieved histories exhibit a high degree of similarity with their corresponding queries in terms of objects, attributes, as well as the overall style or mood of the image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 239, + 286, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 286, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 286, + 287 + ], + "type": "text", + "content": "This effectively demonstrates the proficiency of our retriever in sourcing relevant user histories, thereby providing a robust reference for our rewriter to carry out personalized prompt rewriting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 288, + 286, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 286, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 286, + 455 + ], + "type": "text", + "content": "In the right section of Figure 8, we display the generation outcomes of both \"Shortened Prompt\" and \"Personalized Prompt\". The \"Shortened Prompt\" column exhibits the results produced from the shortened prompts, while the \"Personalized Prompt\" column features the rewritten prompts of our method along with their corresponding generated images. It's evident from these displays that images our method generates are more inclined towards user preferences based on their histories, a testament to the expressive power of our rewritten prompt. For instance, when the query \"Hobbit homes\" is used (as seen in the first row), we observe the user's preferred style across three images, all capturing the mood of mountainous scenery and depicting the Hobbit homes within a consistent landscape." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 460, + 189, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 189, + 473 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 189, + 473 + ], + "type": "text", + "content": "5.4.Quantitative Comparison" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 479, + 286, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 286, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 286, + 538 + ], + "type": "text", + "content": "To further examine the effectiveness of our Personalized PR method, we conduct offline and online quantitative evaluation, comparing different settings to the baseline methods without enhanced prompt rewritten on the test samples we created." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 50, + 544, + 286, + 623 + ], + "blocks": [ + { + "bbox": [ + 50, + 544, + 286, + 623 + ], + "lines": [ + { + "bbox": [ + 50, + 544, + 286, + 623 + ], + "spans": [ + { + "bbox": [ + 50, + 544, + 286, + 623 + ], + "type": "table", + "html": "
MethodRetrieverPMS ↑Image-Align↑ROUGE-L↑
Shortened Prompt-0.55670.62720.3268
Promptist [7]-0.58580.64810.2947
General PR-0.59960.59120.2082
Personalized PRBM250.61250.65810.3942
EBR0.60830.64850.4137
Personalized PR+ ICLBM250.62530.64560.4417
EBR0.61790.67960.4686
", + "image_path": "efd4f5abcdbd1a1a2bb877b1a57b6fc82dd3c215eb4e7d28317e9b5f14dd8c8d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Offline Test. Table 2 showcases the numerical results comparing various retrieval and rewriting configurations with shortened prompts. We can see that Prompt [7] and General PR are slightly better than 'Shortened Prompt' in terms" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "of PMS and Image-Align. Without the users' histories, these large language models perform prompt rewriting in arbitrary pathways. Thus, the rewritten prompts these baseline methods generate could totally deviate from the users' preferences. Such hallucination scenarios result in the decrease in terms of ROUGE-L for both Prompt [7] and General PR. This indicates that although general prompt rewriting methods appear to be refining the prompts, their results do not align with users' preferences well. Our method outperforms all baseline methods, i.e. Prompt [7] and General PR, on all metrics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 204, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 545, + 251 + ], + "type": "text", + "content": "A comparison between BM25 and EBR reveals that dense retrieval generally outperforms sparse retrieval, although the difference is relatively small, indicating that both methods can produce satisfactory results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 251, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 335 + ], + "type": "text", + "content": "Further, when contrasting context-independent rewriting with in-context rewriting, it's evident that ICL produces superior outcomes. By integrating ICL with EBR, we achieve absolute improvements of " + }, + { + "bbox": [ + 304, + 251, + 545, + 335 + ], + "type": "inline_equation", + "content": "14.2\\%" + }, + { + "bbox": [ + 304, + 251, + 545, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 251, + 545, + 335 + ], + "type": "inline_equation", + "content": "6.9\\%" + }, + { + "bbox": [ + 304, + 251, + 545, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 251, + 545, + 335 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 304, + 251, + 545, + 335 + ], + "type": "text", + "content": " in terms of ROUGE-L, PMS and Text-Align metrics respectively. This underscores the exceptional performance of personalized prompting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 335, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 394 + ], + "type": "text", + "content": "To check how sensitive our prompt rewriting methods with respect to the length of the prompts before rewriting, we assess our rewriting method using two additional types of prompts with shorter lengths, namely \"Noun Phrase\" and \"Noun\", as detailed in Table 3." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 309, + 402, + 545, + 469 + ], + "blocks": [ + { + "bbox": [ + 47, + 628, + 286, + 660 + ], + "lines": [ + { + "bbox": [ + 47, + 628, + 286, + 660 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 286, + 660 + ], + "type": "text", + "content": "Table 2. Comparison results of different variants of our method with the baseline. Evidently, our method using EBR retriever (top-3 retrieval) and 1-shot ICL can achieve most best results." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 402, + 545, + 469 + ], + "lines": [ + { + "bbox": [ + 309, + 402, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 309, + 402, + 545, + 469 + ], + "type": "table", + "html": "
Prompt TypeMethodPMS ↑Image-Align↑ROUGE-L↑
NounOriginal0.55370.60870.1770
Personalized PR0.61420.64780.2804
Noun PhraseOriginal0.55540.61460.2459
Personalized PR0.61680.65340.3387
Short SentenceOriginal0.55670.62720.3268
Personalized PR0.61790.67960.4686
", + "image_path": "53605a8a04efb4544a1e4b9fde26de10876c124d83d85d32753b7eb3ca27a60c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 478, + 545, + 522 + ], + "lines": [ + { + "bbox": [ + 304, + 478, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 545, + 522 + ], + "type": "text", + "content": "Table 3. Performance with respect to different prompt lengths, i.e. Noun, Noun Phrase and Short Sentence. Our \"Personalized PR\" equipped with top-3 dense retriever, and 1-shot ICL consistently enhances results, even with only nouns or noun phrases." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "type": "text", + "content": "These two prompts are derived using spaCy [9] from our dataset, adhering to the principle of minimizing word count while maintaining the main entities. The results displayed in Table 3 show that across all three shortened scale, \"Personalized PR\" outperforms the baseline across all metrics, demonstrating the effectiveness of our methods in recovering user preferences." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Online Test. To further validate our approach, we have carried out a single-blinded online evaluation on our website mentioned above. Active users recently in the website are randomly selected to participate in the online test. Upon each prompt input by a participant, there's an equal chance of generating an image using either \"Original Prompt\" or our method's \"Personalized Prompt\". Participants can choose to \"Save\" or \"Delete\" each generated image based" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7733" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 74, + 154, + 205 + ], + "blocks": [ + { + "bbox": [ + 50, + 74, + 154, + 205 + ], + "lines": [ + { + "bbox": [ + 50, + 74, + 154, + 205 + ], + "spans": [ + { + "bbox": [ + 50, + 74, + 154, + 205 + ], + "type": "image", + "image_path": "4dfb5d41b9245ab81ef5cad70471c6ce74710280455d16d04d90e61f114b7193.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 287, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 243 + ], + "type": "text", + "content": "Figure 9. Illustration of our online evaluation methods and the improvement of Save Rate from personalized Prompt." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 74, + 287, + 198 + ], + "blocks": [ + { + "bbox": [ + 170, + 74, + 287, + 198 + ], + "lines": [ + { + "bbox": [ + 170, + 74, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 170, + 74, + 287, + 198 + ], + "type": "image", + "image_path": "9819b24a6b39f99a5060e3c58cd2cc319cabda9fe82545fc51b0a288e94476c8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 250, + 288, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 250, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 288, + 381 + ], + "type": "text", + "content": "on their preference. We use their \"Save\" actions to assess our method's effectiveness. In the online test, 247 users generated 905 images, with 433 from \"Original Prompt\" and 472 from \"Personalized Prompt\". The results of the online evaluation, as shown in Figure 9, indicate that users prefer images generated using the \"Personalized Prompt\" over the \"Original Prompt\", with a " + }, + { + "bbox": [ + 46, + 250, + 288, + 381 + ], + "type": "inline_equation", + "content": "17.1\\%" + }, + { + "bbox": [ + 46, + 250, + 288, + 381 + ], + "type": "text", + "content": " increase in \"Save\" actions. This suggests that our method aligns better with user preferences, affirming its effectiveness. We anticipate even better results in real-world scenarios with more user information and an improved text-to-image generation method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 389, + 141, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 389, + 141, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 141, + 403 + ], + "type": "text", + "content": "5.5. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 408, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 287, + 456 + ], + "type": "text", + "content": "In this section, we ablate " + }, + { + "bbox": [ + 46, + 408, + 287, + 456 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 408, + 287, + 456 + ], + "type": "text", + "content": " relevant historical prompts used to rewrite personalized prompts and the number of demonstration examples used for in-context learning. We evaluate each experiment on the same setup as in Section 5.1." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 50, + 467, + 287, + 510 + ], + "blocks": [ + { + "bbox": [ + 50, + 467, + 287, + 510 + ], + "lines": [ + { + "bbox": [ + 50, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 50, + 467, + 287, + 510 + ], + "type": "table", + "html": "
MethodRetrieval Top-kPMS ↑Image-Align ↑ROUGE-L ↑
Personalized PR + ICL10.60570.67510.4539
30.61790.67960.4686
50.62040.67480.4474
70.62650.66510.4592
", + "image_path": "4d1f9a18aa185c0102f569b5e37a972c1e4a0da387ea31137bfadf068ca723e8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 652 + ], + "type": "text", + "content": "Retrieval Top-k Ablation. As shown in Table 4, when using 3 most relevant historical prompts for rewriting, we can obtain most best results among all evaluation metrics. We analyze that too more historical prompts could provide redundant information and also too long prompt input to ChatGPT worsens the rewriting performance. Therefore, we choose 3 as the number of retrieval results, a balanced manner between performance and efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Number of ICL Demonstrations Ablation. Table 5 shows that 1-shot setting, i.e., given 1 demonstration example for in-context learning, can achieve the best results in general among all the four evaluation metrics regarding both BM25 and EBR. This demonstrates that our prompt rewriting tem-" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 135 + ], + "blocks": [ + { + "bbox": [ + 46, + 519, + 287, + 552 + ], + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 552 + ], + "type": "text", + "content": "Table 4. Ablation for Retrieval Top-k. We empirically conduct experiments with respect to " + }, + { + "bbox": [ + 46, + 519, + 287, + 552 + ], + "type": "inline_equation", + "content": "k \\in \\{ 1,3,5,7\\}" + }, + { + "bbox": [ + 46, + 519, + 287, + 552 + ], + "type": "text", + "content": " ,keeping the configuration of DENSE retriever and 1-shot ICL." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 545, + 135 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 135 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 135 + ], + "type": "table", + "html": "
Method (Retrieval)ICL ShotPMS ↑Image-Align ↑ROUGE-L ↑
Personalized PR (BM25)10.62530.64560.4417
30.62890.65800.4381
50.62360.65710.4226
Personalized PR (EBR)10.61790.67960.4686
30.62740.67080.4354
50.62420.67240.4439
", + "image_path": "25c3c76b59abeebd107d9ba4624310686d91fb5262cef1dbcfd3df4e218e98b7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 143, + 545, + 188 + ], + "lines": [ + { + "bbox": [ + 305, + 143, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 305, + 143, + 545, + 188 + ], + "type": "text", + "content": "Table 5. Ablation for Number of ICL Demonstrations. We experiment with consistent top-3 retrieval both on BM25 and EBR. When we set ICL shot as 1 and use EBR retriever, we observe more superior results appearing." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 207, + 545, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 207, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 305, + 207, + 545, + 242 + ], + "type": "text", + "content": "plate is efficient and effective enough for extracting the personalized preference from numerous historical data of each user." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 264, + 471, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 264, + 471, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 264, + 471, + 277 + ], + "type": "text", + "content": "6. Conclusion and Future Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 288, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 467 + ], + "type": "text", + "content": "In conclusion, this study has underscored the significance of harnessing historical user behaviors to construct personalized AI content generation. Our strategy aims to refine user prompts by capitalizing on previous user interactions with the system. We have introduced an innovative technique that entails reconfiguring user prompts based on a newly developed, large-scale text-to-image dataset encompassing over 300,000 prompts from 3,115 distinct users. This methodology has proven to augment the expressiveness of user prompts and ensure their alignment with the desired visual outputs. Our empirical results have underscored the supremacy of our techniques over conventional methods. This superiority was corroborated through our novel offline evaluation method and online tests, thereby affirming the efficacy of our approach." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 471, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 578 + ], + "type": "text", + "content": "Although the outcomes are encouraging, considerable research still lies ahead in this domain. In the realm of personalized text-to-image generation, the incorporation of more personal details like a user's age and gender could bolster performance. The methodologies used could be extended to other LPMs, including LLMs. Additionally, techniques for enhancing search engines, e.g. data source purification and ranking optimization, could be assimilated into these models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 582, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 545, + 617 + ], + "type": "text", + "content": "We are confident that our contributions represent advancements towards a more personalized and user-focused artificial intelligence." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 636, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 636, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 636, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work is supported in part by the Research Center for Industries of the Future at Westlake University and Westlake Education Foundation (Grant No. WU2023C017), National Natural Science Foundation of China (Grant Nos. 62171111, 62006245), and Open Research Fund from State Key Laboratory of High Performance Computing of China (Grant No. 202201-15)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7734" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 201 + ], + "type": "text", + "content": "[2] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. \"this is my unicorn, fluffy\": Personalizing frozen vision-language representations. In European Conference on Computer Vision, pages 558-577. Springer, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 288, + 258 + ], + "type": "text", + "content": "[3] Grégoire Deletang, Anian Ruoss, Paul-Ambroise Duquenne, Elliot Catt, Tim Genewein, Christopher Mattern, Jordi Grau-Moya, Li Kevin Wenliang, Matthew Aitchison, Laurent Orseau, et al. Language modeling is compression. arXiv preprint arXiv:2309.10668, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 260, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 288, + 315 + ], + "type": "text", + "content": "[4] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 317, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 288, + 361 + ], + "type": "text", + "content": "[5] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 362, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 362, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 362, + 287, + 407 + ], + "type": "text", + "content": "[6] Qingyan Guo, Rui Wang, Junliang Guo, Bei Li, Kaitao Song, Xu Tan, Guoqing Liu, Jiang Bian, and Yujiu Yang. Connecting large language models with evolutionary algorithms yields powerful prompt optimizers, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 408, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 288, + 441 + ], + "type": "text", + "content": "[7] Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. arXiv preprint arXiv:2212.09611, 2022. 2, 6, 7, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "type": "text", + "content": "[8] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 520 + ], + "type": "text", + "content": "[9] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python. 2020. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 565 + ], + "type": "text", + "content": "[10] Rolf Jagerman, Honglei Zhuang, Zhen Qin, Xuanhui Wang, and Michael Bendersky. Query expansion by prompting large language models. arXiv preprint arXiv:2305.03653, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 632 + ], + "type": "text", + "content": "[11] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[12] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "text", + "content": "[13] Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. What makes" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "good in-context examples for gpt-3? arXiv preprint arXiv:2101.06804, 2021.5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "text", + "content": "[14] Jianmo Ni, Chen Qu, Jing Lu, Zhuyun Dai, Gustavo Hernández Abrego, Ji Ma, Vincent Y Zhao, Yi Luan, Keith B Hall, Ming-Wei Chang, et al. Large dual encoders are generalizable retrievers. arXiv preprint arXiv:2112.07899, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 173 + ], + "type": "text", + "content": "[15] OpenAI. Chatgpt. https://chat.openai.com. 5, 6, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 240 + ], + "type": "text", + "content": "[16] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 243, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 296 + ], + "type": "text", + "content": "[17] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "type": "text", + "content": "[18] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 344, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 545, + 387 + ], + "type": "text", + "content": "[19] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. Stable diffusion v1-5. https://huggingface.co/runwayml/stable-diffusion-v1-5.6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 389, + 547, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 389, + 547, + 444 + ], + "spans": [ + { + "bbox": [ + 307, + 389, + 547, + 444 + ], + "type": "text", + "content": "[20] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 445, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 510 + ], + "type": "text", + "content": "[21] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 512, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 578 + ], + "type": "text", + "content": "[22] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 579, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 634 + ], + "type": "text", + "content": "[23] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 669 + ], + "type": "text", + "content": "[24] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[25] Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. Key-locked rank one editing for text-to-image personalization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7735" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 274 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[26] Xiao Wang, Sean MacAvaney, Craig Macdonald, and Iadh Ounis. Generative query reformulation for effective adhoc search. arXiv preprint arXiv:2308.00415, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 149 + ], + "type": "text", + "content": "[27] Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V Le, Denny Zhou, and Xinyun Chen. Large language models as optimizers. arXiv preprint arXiv:2309.03409, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "type": "text", + "content": "[28] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 229 + ], + "type": "text", + "content": "[29] Shanshan Zhong, Zhongzhan Huang, Wushao Wen, Jinghui Qin, and Liang Lin. Sur-adapter: Enhancing text-to-image pre-trained diffusion models with large language models. arXiv preprint arXiv:2305.05189, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "type": "text", + "content": "[30] Yongchao Zhou, Andrei Loan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. arXiv preprint arXiv:2211.01910, 2022. 2" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "7736" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/695a2b86-fc93-420e-a626-d23f402233d8_content_list.json b/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/695a2b86-fc93-420e-a626-d23f402233d8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..28df05eb057605c7d4b8c51d9089d293061bcbc9 --- /dev/null +++ b/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/695a2b86-fc93-420e-a626-d23f402233d8_content_list.json @@ -0,0 +1,1590 @@ +[ + { + "type": "text", + "text": "Taming Mode Collapse in Score Distillation for Text-to-3D Generation", + "text_level": 1, + "bbox": [ + 132, + 130, + 839, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Peihao Wang $^{1}$ , Dejia Xu $^{1}$ , Zhiwen Fan $^{1}$ , Dilin Wang $^{2}$ , Sreyas Mohan $^{2}$ , Forrest Iandola $^{2}$ , Rakesh Ranjan $^{2}$ , Yilei Li $^{2}$ , Qiang Liu $^{1}$ , Zhangyang Wang $^{1}$ , Vikas Chandra $^{2}$ $^{1}$ The University of Texas at Austin, $^{2}$ Meta Reality Labs", + "bbox": [ + 138, + 179, + 828, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{peihaowang,iejia, zhiwenfan, atlaswang}@utexas.edu, lqiang@cs.utexas.edu \n{wdilin, sreyasmohan, fni, rakeshr, yileil, vchandra}@meta.com \nvita-group.github.io/3D-Mode-Collapse/", + "bbox": [ + 156, + 236, + 810, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the remarkable performance of score distillation in text-to-3D generation, such techniques notoriously suffer from view inconsistency issues, also known as \"Janus\" artifact, where the generated objects fake each view with multiple front faces. Although empirically effective methods have approached this problem via score debiasing or prompt engineering, a more rigorous perspective to explain and tackle this problem remains elusive. In this paper, we reveal that the existing score distillation-based text-to-3D generation frameworks degenerate to maximal likelihood seeking on each view independently and thus suffer from the mode collapse problem, manifesting as the Janus artifact in practice. To tame mode collapse, we improve score distillation by re-establishing the entropy term in the corresponding variational objective, which is applied to the distribution of rendered images. Maximizing the entropy encourages diversity among different views in generated 3D assets, thereby mitigating the Janus problem. Based on this new objective, we derive a new update rule for 3D score distillation, dubbed Entropic Score Distillation (ESD). We theoretically reveal that ESD can be simplified and implemented by just adopting the classifier-free guidance trick upon variational score distillation. Although embarrassingly straightforward, our extensive experiments demonstrate that ESD can be an effective treatment for Janus artifacts in score distillation.", + "bbox": [ + 75, + 353, + 473, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 760, + 209, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in text-to-3D technology have attracted considerable attention, particularly for its pivotal role in automating high-quality 3D content. This is especially crucial in fields such as virtual reality and gaming, where 3D content forms the bedrock. While numerous techniques are available, the prevailing text-to-3D approach is based on", + "bbox": [ + 75, + 785, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "score distillation [31], popularized by DreamFusion and its follow-up works [4, 19, 26, 50, 54, 56].", + "bbox": [ + 500, + 323, + 890, + 352 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Score distillation leverages a pre-trained 2D diffusion model to sample over the 3D parameter space (i.e. Neural Radiance Fields (NeRF) [27]) such that views rendered from a random angle satisfy the statistics of the image distribution. This algorithm is implemented by backpropagating the estimated score of each view via the chain rule. Despite the notable progress achieved with score distillation-based approaches, it is widely observed that 3D content generated using score distillation suffers from the Janus problem [12], referring to the artifacts that generated 3D objects contain multiple canonical views (see Fig. 1).", + "bbox": [ + 496, + 353, + 893, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To understand this drawback of score distillation, we draw the theoretical connection between the Janus problem and mode collapse, a statistical term describing a distribution concentrating on the high-density area while losing information about the probability tail. We first uncover that the optimization of existing score distillation-based text-to-3D generation degenerates to a maximum likelihood objective, making it susceptible to model collapse. As pre-trained diffusion models are biased to frequently encountered views [12], this oversight leads all views opt to convergence toward the point with the highest likelihood, manifesting as the Janus artifact in practical applications. The main limitation of current methods is that their distillation objectives solely maximize the likelihood of each view independently, without considering the diversity between different views.", + "bbox": [ + 496, + 518, + 895, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the aforementioned issue, we propose a principled approach Entropic Score Distillation (ESD), which regularizes the score distillation process by entropy maximization of the rendered image distribution, thereby enhancing the diversity of views in generated 3D assets and alleviating the Janus problem. Our derived ESD update admits a simple form as a weighted combination of scores for pre-trained image distribution and rendered image distribution. Com", + "bbox": [ + 496, + 746, + 893, + 867 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1For example, it is common that a frontal view of a cat is more likely to be sampled from latent diffusion models than the back view.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during an internship with Meta.", + "bbox": [ + 94, + 886, + 331, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9037", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3411a3abeb213c21e1da738c11e93e90d67e603c19a36d3e0a541118c29a28e5.jpg", + "image_caption": [ + "Our results", + "A cat with tiger stripes" + ], + "image_footnote": [], + "bbox": [ + 117, + 113, + 589, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6631107c8c53aff27d4e5038ded3ab7ee579a8639c8e6e44361baae353a32eb7.jpg", + "image_caption": [ + "\"Janus\" results" + ], + "image_footnote": [], + "bbox": [ + 609, + 113, + 846, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/877e227111e2f1ba9f60321151fe52c25097464e4f4e231917a57f21bb286d2d.jpg", + "image_caption": [ + "A bald eagle carved out of wood" + ], + "image_footnote": [], + "bbox": [ + 117, + 220, + 589, + 314 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/20786928d71d09f8973962fdde74c821b03167203b9d3f153ca22eceab7b24c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 220, + 846, + 314 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6756dc1ea598b650b86c8f6e4cad871f23715a945a68e3be55c8e8a445796d19.jpg", + "image_caption": [ + "A chimpanzee dressed like Henry VIII king of England" + ], + "image_footnote": [], + "bbox": [ + 117, + 330, + 589, + 422 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b259af55d7fb3fa251ea81c9ee10ba0b7455a24070247d764489869f837707b9.jpg", + "image_caption": [ + "Front View", + "Back View", + "Figure 1. A Preview of Qualitative Results. We present the front and back views of objects synthesized by VSD (ProlificDreamer) on the right two columns, and four views of our generated results on the left. VSD suffers from \"Janus\" problem, where both front and back views contain a frontal face of the targeted object, while our method effectively mitigates this artifact. Please refer to more results in Appendix D." + ], + "image_footnote": [], + "bbox": [ + 609, + 330, + 846, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pared with Score Distillation Sampling (SDS) [31], our ESD involves the score of the rendered image distribution, serving to maximize the entropy of the rendered image distribution. Unlike Variational Score Distillation (VSD) [56], the learned score function of the rendered image distribution does not depend on the camera pose. This subtle difference has a more profound impact, as we show the score function of rendered images modeled by VSD corresponds to an objective with fixed entropy, thereby having no influence on view variety. In contrast, ESD optimizes for a Kullback-Leibler divergence with a non-constant entropy term parameterized by the 3D model, leading to an effect that encourages diversity among different views.", + "bbox": [ + 75, + 501, + 470, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In practice, we find it challenging to optimize the score of the rendered image distribution without conditioning on the camera pose. To facilitate training, we discover that the gradient from the entropy can be decomposed into a combination of scores: one depends on the camera pose, and the other independent of it, with a coefficient interacting between these two terms. Through this theoretical establishment, we are able to adopt a handy implementation of ESD by Classifier Free Guidance (CFG) trick [10] where conditional and unconditional scores are trained alternatively and mixed during inference.", + "bbox": [ + 75, + 700, + 470, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through extensive experiments with our proposed ESD, we demonstrate its efficacy in alleviating the Janus problem", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and its significant advantages in improving 3D generation quality when compared to the baseline methods [31, 56] and other remedy techniques [2, 12]. As a side contribution, we also borrow two inception scores [36] to evaluate text-to-3D results and numerically probe model collapse in score distillation. We show these two metrics can effectively characterize the quality and diversity of views, highly matching our qualitative observations.", + "bbox": [ + 496, + 501, + 893, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 500, + 638, + 627, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion Models", + "text_level": 1, + "bbox": [ + 500, + 664, + 666, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models, as demonstrated by various works [11, 44, 46, 48], have shown to be highly effective in text-to-image generation. Technically, a diffusion model learns to gradually transform a normal distribution $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ to the target distribution $p_{data}(\\boldsymbol {x}|\\boldsymbol {y})$ where $\\pmb{y}$ denotes the text prompt embeddings. The sampling trajectory is determined by a forward process with the conditional probability $p_t(\\pmb {x}_t|\\pmb {x}_0) = \\mathcal{N}(\\pmb {x}_t|\\alpha_t\\pmb {x}_0,\\sigma_t^2\\pmb {I})$ ,where $\\pmb {x}_t\\in \\mathbb{R}^D$ represents the sample at time $t\\in [0,T]$ , and $\\alpha_{t},\\sigma_{t} > 0$ are time-dependent diffusion coefficients. Consequently, the distribution at time $t$ can be formulated as $p_t(\\pmb {x}_t|\\pmb {y}) =$ $\\int p_{data}(\\pmb {x}_0|\\pmb {y})\\mathcal{N}(\\pmb {x}_t|\\alpha_t\\pmb {x}_0,\\sigma_t^2\\pmb {I})d\\pmb {x}_0.$ Diffusion models generate samples through a reverse process starting from Gaussian noises, which can be described by the ODE:", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9038", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/21ba407548e7b038693c647e4ca8d1f10418ff0ad9162d38d5595fcaba68d9bc.jpg", + "image_caption": [ + "Figure 2. Illustration of the effect of entropy regularization. Learned image distributions often exhibit a higher probability mass for objects' frontal faces. Pure maximal likelihood seeking is opt to mode collapse (Sec. 3). Adding entropy regularization can expand the support of fitted distribution $q_{t}^{\\theta}(\\boldsymbol{x}|\\boldsymbol{y})$ with mode-covering behavior (Sec. 4)." + ], + "image_footnote": [], + "bbox": [ + 83, + 85, + 442, + 244 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/916da4f02dfc3f104e790dd200d7ab14c69a42883c50231afaaad73cf2b61d27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 85, + 854, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathrm{d}\\pmb{x}_t / \\mathrm{d}t = -\\nabla_{\\pmb{x}}\\log p_t(\\pmb{x}_t)$ with the boundary condition $\\pmb{x}_T\\sim \\mathcal{N}(\\pmb {0},\\pmb {I})$ [22, 45, 48]. Such a process requires the computation of score function $\\nabla_{\\pmb{x}}\\log p_t(\\pmb{x}_t)$ which is often obtained by fitting a time-conditioned noise estimator $\\epsilon_{\\phi}:\\mathbb{R}^{D}\\to \\mathbb{R}^{D}$ using score matching loss [15, 47, 52].", + "bbox": [ + 75, + 304, + 470, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Text-to-3D Score Distillation", + "text_level": 1, + "bbox": [ + 76, + 392, + 331, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Score distillation based 3D asset generation requires representing 3D scenes as learnable parameters $\\pmb{\\theta} \\in \\mathbb{R}^{N}$ equipped with a differentiable renderer $g(\\pmb{\\theta}, \\pmb{c}): \\mathbb{R}^{N} \\to \\mathbb{R}^{D}$ that projects 3D scene $\\pmb{\\theta}$ into images with respect to the camera pose $\\pmb{c}$ . Here $N, D$ are the dimensions of the 3D parameter space and rendered images, respectively. Neural radiance fields (NeRF) [27] are often employed as the underlying 3D representation for its capability of modeling complex scenes.", + "bbox": [ + 75, + 416, + 468, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent works [4, 14, 19, 26, 31, 50, 54-56] demonstrate the feasibility of using a pretrained 2D diffusion model to guide 3D object creation. Below, we elaborate on two score distillation schemes, adopted therein: Score Distillation Sampling (SDS) [31] and Variational Score Distillation (VSD) [56].", + "bbox": [ + 75, + 537, + 470, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Score Distillation Sampling (SDS). SDS updates the 3D parameter $\\theta$ as follows:", + "bbox": [ + 75, + 651, + 468, + 681 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {\\theta}} J _ {S D S} (\\boldsymbol {\\theta}) = - \\mathbb {E} \\left[ \\omega (t) \\frac {\\partial g (\\boldsymbol {\\theta} , \\boldsymbol {c})}{\\partial \\boldsymbol {\\theta}} \\left(\\sigma_ {t} \\nabla \\log p _ {t} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {y}\\right) - \\boldsymbol {\\epsilon}\\right) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 694, + 483, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the expectation is taken over timestep $t\\sim \\mathcal{U}[0,T]$ Gaussian noises $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ and camera pose $\\pmb {c}\\sim p_c(\\pmb {c})$ Here is $\\nabla \\log p$ is a pre-trained diffusion model $\\epsilon_{\\phi}(\\pmb {x},t,\\pmb {y})$ and $\\pmb{x}_t$ is a noisy version of the rendering given by camera pose $\\pmb {c}$ $x_{t} = \\alpha_{t}g(\\pmb {\\theta},\\pmb {c}) + \\sigma_{t}\\pmb{\\epsilon}$ .Updating $\\pmb{\\theta}$ as in Eq. (1) has been shown to minimize the evidence lower bound (ELBO) for the rendered images, see Wang et al. [54], Xu et al. [59].", + "bbox": [ + 75, + 756, + 470, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Variational Score Distillation (VSD). VSD [56] is introduced in ProlificDreamer, VSD improves upon SDS by deriving the following Wasserstein gradient flow [51]:", + "bbox": [ + 500, + 304, + 893, + 349 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\boldsymbol {\\theta}} J _ {V S D} (\\boldsymbol {\\theta}) = - \\mathbb {E} \\left[ \\omega (t) \\frac {\\partial g (\\boldsymbol {\\theta} , \\boldsymbol {c})}{\\partial \\boldsymbol {\\theta}} \\left(\\sigma_ {t} \\nabla \\log p _ {t} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}\\right) \\right. \\\\ \\left. - \\sigma_ {t} \\nabla \\log q _ {t} (\\boldsymbol {x} _ {t} | \\boldsymbol {c})\\right) \\Bigg ]. \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 359, + 890, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarly, $\\pmb{x}_t = \\alpha_t g(\\pmb{\\theta}, \\pmb{c}) + \\sigma_t \\pmb{\\epsilon}$ is the noisy observation of the rendered image. In contrast to SDS, VSD introduces a new score function of the noisy rendered images conditioned on the camera pose $\\mathbf{c}$ . To obtain this score, Wang et al. [56] fine-tunes a diffusion model using images rendered from the 3D scene as follows:", + "bbox": [ + 498, + 438, + 892, + 527 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {\\psi}} \\mathbb {E} \\left[ \\omega (t) \\| \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\psi}} \\left(\\alpha_ {t} g (\\boldsymbol {\\theta}, \\boldsymbol {c}) + \\sigma_ {t} \\boldsymbol {\\epsilon}, t, \\boldsymbol {c}, \\boldsymbol {y}\\right) - \\boldsymbol {\\epsilon} \\| _ {2} ^ {2} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 539, + 890, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\psi}(\\boldsymbol{x},t,\\boldsymbol{c},\\boldsymbol{y})$ is the noise estimator of $\\nabla \\log q_t(\\boldsymbol{x}_t|\\boldsymbol{c})$ as in diffusion models. As proposed in ProlificDreamer, $\\psi$ is parameterized by LoRA [13] and initialized from a pre-trained diffusion model same as $\\nabla \\log p_t$ .", + "bbox": [ + 498, + 574, + 893, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Revealing Mode Collapse in Score Distillation", + "text_level": 1, + "bbox": [ + 500, + 648, + 890, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite the remarkable performance of SDS and VSD in 3D asset generation, it is widely observed that the synthesized objects suffer from \"Janus\" artifacts. Janus artifacts refer to the generated 3D scene containing multiple canonical views (the most representative perspective of the object such as the frontal face). In earlier works, Hong et al. [12] and Huang et al. [14] attribute this problem to unimodality of the learned 2D image distribution since the training data for the diffusion models are naturally biased to the most commonly seen views per each category. In this section, we examine extant distillation schemes from a statistical view, which has been overlooked in previous literature.", + "bbox": [ + 496, + 672, + 890, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In principle, natural 2D images can be seen as random projections of 3D scenes. Score distillation matches the image distribution generated by randomly sampled views with a", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2Without special specification, expectations are taken over all relevant random variables and Jacobian matrices are transposed by default.", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9039", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "text-conditioned image distribution to recover the underlying 3D representation. Hence, Janus artifact, in which each view becomes uniform and identical to the most commonly seen views, can be interpreted as a manifestation of distribution collapse to samples within the high-density region. Such distribution degeneration essentially corresponds to the statistical phenomenon mode collapse, which happens when an optimized distribution fails to characterize the data diversity and concentrates on a single type of output [1, 7, 25, 36, 49].", + "bbox": [ + 75, + 90, + 472, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Below we theoretically reveal why SDS and VSD are prone to mode collapse. As shown in Poole et al. [31], Wang et al. [56], SDS and VSD equals to the gradient of the following Kullback-Leibler (KL) divergence, i.e., $J_{SDS}(\\pmb{\\theta}) = J_{VSD}(\\pmb{\\theta}) = J_{KL}(\\pmb{\\theta})$ up to an additive constant:", + "bbox": [ + 76, + 227, + 472, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nJ _ {K L} (\\boldsymbol {\\theta}) = \\mathbb {E} \\left[ \\Omega (t) \\mathcal {D} _ {\\mathrm {K L}} \\left(q _ {t} ^ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {t} | \\boldsymbol {c}, \\boldsymbol {y}\\right) \\| p _ {t} \\left(\\boldsymbol {x} _ {t} | \\boldsymbol {y}\\right)\\right) \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 311, + 470, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Omega(t) = \\omega(t)\\sigma_t / \\alpha_t$ and the expectation is taken over $t \\sim \\mathcal{U}[0,T]$ and $\\pmb{c} \\sim p_c(\\pmb{c})$ . Here $p_t(\\pmb{x}_t|\\pmb{y}) = \\int p_0(\\pmb{x}_0|\\pmb{y})\\mathcal{N}(\\pmb{x}_t|\\alpha_t\\pmb{x}_0,\\sigma_t^2\\pmb{I})d\\pmb{x}_0$ is the image distribution perturbed by Gaussian noises, while $q_t^\\theta (\\pmb{x}_t|\\pmb{c},\\pmb{y}) = \\int q_0^\\theta (\\pmb{x}_0|\\pmb{c})\\mathcal{N}(\\pmb{x}_t|\\alpha_t\\pmb{x}_0,\\sigma_t^2\\pmb{I})d\\pmb{x}_0$ models the image distribution generated by 3D parameter $\\pmb{\\theta}$ with respect to camera pose $\\pmb{c}$ and diffused by Gaussian distribution. As shown by Wang et al. [56], $J_{KL}(\\pmb{\\theta}) = 0$ implies $q_0^\\theta (\\pmb{x}_0|\\pmb{c}) = p(\\pmb{x}_0|\\pmb{y})$ , i.e., the distribution of synthesized views satisfy the text-conditioned image distribution.", + "bbox": [ + 75, + 340, + 472, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, it has not escaped from our notice that $q_0^\\theta(\\boldsymbol{x}_0|\\boldsymbol{c}) = \\delta(\\boldsymbol{x}_0 - g(\\boldsymbol{\\theta},\\boldsymbol{c}))$ is a Dirac distribution for both SDS and VSD. This causes the original KL divergence minimization (Eq. 4) degenerate to a Maximal Likelihood Estimation (MLE) problem:", + "bbox": [ + 76, + 492, + 472, + 568 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} J _ {K L} (\\boldsymbol {\\theta}) = \\underbrace {- \\mathbb {E} \\left[ \\Omega (t) \\mathbb {E} _ {\\boldsymbol {x} _ {t} \\sim q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {c} , \\boldsymbol {y})} \\log p _ {t} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}) \\right]} _ {J _ {M L E} (\\boldsymbol {\\theta})} \\tag {5} \\\\ - \\underbrace {\\mathbb {E} \\left[ \\Omega (t) H \\left[ q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {c} , \\boldsymbol {y}) \\right] \\right]} _ {\\text {c o n s t .}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 575, + 468, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H[q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{y})] = -\\mathbb{E}_{\\boldsymbol{x}_t \\sim q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{c}, \\boldsymbol{y})}[\\log q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{c}, \\boldsymbol{y})]$ denotes the entropy of $q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{y})$ , which turns out to be a constant because $q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{c}, \\boldsymbol{y}) = \\mathcal{N}(\\boldsymbol{x}_t | \\alpha_t g(\\boldsymbol{\\theta}, \\boldsymbol{c}), \\sigma_t^2 \\boldsymbol{I})$ which has fixed entropy once $t, \\boldsymbol{\\theta}$ and $\\boldsymbol{c}$ have been specified. See full derivation in Appendix A.1.", + "bbox": [ + 75, + 670, + 468, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that Eq. 5 signifies $J_{KL}(\\pmb{\\theta}) = J_{MLE}(\\pmb{\\theta})$ up to an additive constant, hence $J_{KL}(\\pmb{\\theta})$ shares all minima with $J_{MLE}(\\pmb{\\theta})$ . It is known that likelihood maximization is more prone to mode collapse. Intuitively, minimizing $J_{MLE}(\\pmb{\\theta})$ seeks each view independently to have the maximum log-likelihood on the image distribution $p(\\pmb{x}_0|\\pmb{y})$ . Since $p(\\pmb{x}_0|\\pmb{y})$ is usually unimodal and peaks at the canonical view, each view of the scene will collapse to the same local minimum, resulting in Janus artifact (see Fig. 2). We postulate that the existing distillation strategies may be inherently limited", + "bbox": [ + 75, + 750, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a33b82e50ad9746c4eb02291f66fe6c94ddbb4dd5361afb564e95a855af58acb.jpg", + "image_caption": [ + "Figure 3. Gaussian Example. To illustrate the effects of entropy regularization, we leverage SDS, VSD and ESD to fit a 2D Gaussian distribution. The blue points are sampled from the ground-truth distribution while the orange points are from the fitted distribution." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 887, + 204 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by their log-likelihood seeking behaviors, which are more susceptible to mode collapse, especially with biased image distributions.", + "bbox": [ + 496, + 275, + 890, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Entropy Regularized Score Distillation", + "text_level": 1, + "bbox": [ + 498, + 335, + 848, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Entropic Score Distillation", + "text_level": 1, + "bbox": [ + 498, + 372, + 740, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we highlight the importance of the entropy in score distillation. It is known that higher entropy implies the corresponding distribution could cover a larger support of the ambient space and thus increase the sample diversity. In Eq. 5, the entropy term is shown to diminish in the training objective, which causes each generated view to lack diversity and collapse to a single image with the highest likelihood.", + "bbox": [ + 496, + 396, + 890, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this end, we propose to bring in an entropy regularization to $J_{MLE}(\\pmb{\\theta})$ for boosting the view diversity. Since $q_t^\\pmb{\\theta}(\\pmb{x}_t|\\pmb{c},\\pmb{y})$ has constant entropy, we regularize entropy for the distribution $q_t^\\pmb{\\theta}(\\pmb{x}_t|\\pmb{y}) = \\int q_t^\\pmb{\\theta}(\\pmb{x}_t|\\pmb{c},\\pmb{y})p_c(\\pmb{c})d\\pmb{c}$ , which can be simulated by randomly sampling views from the 3D parameter $\\pmb{\\theta}$ . Consider the following objective:", + "bbox": [ + 496, + 502, + 893, + 593 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} J _ {E n t} (\\boldsymbol {\\theta}, \\lambda) = - \\mathbb {E} \\left[ \\Omega (t) \\mathbb {E} _ {\\boldsymbol {x} _ {t} \\sim q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {c}, \\boldsymbol {y})} \\log p _ {t} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}) \\right] \\tag {7} \\\\ - \\lambda \\mathbb {E} \\left[ \\Omega (t) H \\left[ q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}) \\right] \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 604, + 890, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyper-parameter controlling the regularization strength. We note that without $H[q_t^\\theta (\\pmb {x}_t|\\pmb {y})]$ , each view is optimized independently and implicitly regularized by the underlying parameterization. However, upon imposing $H[q_t^\\theta (\\pmb {x}_t|\\pmb {y})]$ , all views become explicitly correlated with each other, as they collectively contribute to the entropy computation. Intuitively, $J_{Ent}(\\pmb {\\theta},\\lambda) = J_{MLE}(\\pmb {\\theta})-$ $\\lambda \\mathbb{E}[\\Omega (t)H[q_t^\\theta (\\pmb {x}_t|\\pmb {y})]]$ seeks the maximal log-likelihood for each view while simultaneously enlarging the entropy for distribution $q_{t}^{\\theta}(\\pmb{x}_{t}|\\pmb{y})$ , which spans the support and encourages diversity across the rendered views. To gain more insights, we present the following theoretical results:", + "bbox": [ + 496, + 662, + 893, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. For any $\\lambda \\in \\mathbb{R}$ and $\\pmb{\\theta} \\in \\mathbb{R}^{D}$ , we have $J_{Ent}(\\pmb{\\theta},\\lambda) = \\lambda \\mathbb{E}_t[\\Omega(t)\\mathcal{D}_{\\mathrm{KL}}(q_t^\\pmb{\\theta}(\\pmb{x}_t|\\pmb{y})||p_t(\\pmb{x}_t|\\pmb{y}))] + (1 - \\lambda)\\mathbb{E}_{t,\\pmb{c}}[\\Omega(t)\\mathcal{D}_{\\mathrm{KL}}(q_t^\\pmb{\\theta}(\\pmb{x}_t|\\pmb{c},\\pmb{y})||p_t(\\pmb{x}_t|\\pmb{y}))] + const.$", + "bbox": [ + 498, + 854, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9040", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 ESD: Entropic score distillation for text-to-3D generation" + ], + "code_body": "Input: A diffusion model $\\epsilon_{\\phi}(\\pmb {x},t,\\pmb {y})$ ; learnable 3D parameter $\\theta$ ; coefficient $\\lambda$ ; text prompt $\\pmb{y}$ \nInitialize $\\psi$ for another diffusion model $\\epsilon_{\\psi}(\\pmb {x},t,\\pmb {y})$ with the parameter $\\phi$ specified in diffusion model $\\epsilon_{\\phi}(\\pmb {x},t,\\pmb {y})$ , parame \nterized with LoRA. \nwhile not converged do \nRandomly sample a camera pose $c\\sim p_c$ and render a view $\\pmb {x}_0 = g(\\pmb {\\theta},\\pmb {c})$ from $\\pmb{\\theta}$ Sample a $t\\sim \\mathcal{U}[0,T]$ and add Gaussian noise $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ .. $\\pmb{x}_{t} = \\alpha_{t}\\pmb{x}_{0} + \\sigma_{t}\\pmb{\\epsilon}$ $\\pmb {\\theta}\\gets \\pmb {\\theta} + \\eta_1\\left[\\omega (t)\\frac{\\partial g(\\pmb{\\theta},\\pmb{c})}{\\partial\\pmb{\\theta}} (\\epsilon_{\\phi}(\\pmb{x}_{t},t,\\pmb {y}) - \\lambda \\epsilon_{\\psi}(\\pmb{x}_{t},t,\\emptyset ,\\pmb {y}) - (1 - \\lambda)\\epsilon_{\\psi}(\\pmb{x}_{t},t,\\pmb {c},\\pmb {y})\\right]$ (6) With probability $1 - p_{\\emptyset}$ $\\psi \\gets \\psi -\\eta_2\\nabla_{\\psi}\\left[\\omega (t)\\| \\epsilon_{\\psi}(\\pmb {x}_t,t,\\pmb {c},\\pmb {y}) - \\epsilon \\| _2^2\\right]$ Otherwise, $\\psi \\gets \\psi -\\eta_2\\nabla_{\\psi}\\left[\\omega (t)\\| \\epsilon_{\\psi}(\\pmb {x}_t,t,\\emptyset ,\\pmb {y}) - \\epsilon \\| _2^2\\right]$ \nend while \nReturn $\\theta$", + "bbox": [ + 91, + 109, + 893, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We prove Theorem 1 in Appendix A.3. Theorem 1 implies that $J_{Ent}(\\theta, \\lambda)$ essentially equal to a combination of two types of KL divergences, where the former one minimizes the distribution discrepancy between $q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{y})$ and $p_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{y})$ which marginalizes the camera pose within $q_t^\\theta$ , while the latter is the original KL divergence $J_{KL}(\\theta)$ adopted by SDS and VSD which takes expectation over $\\boldsymbol{c}$ out of KL divergence.", + "bbox": [ + 75, + 315, + 472, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we derive the gradient of $J_{Ent}(\\theta, \\lambda)$ that will be backpropagated to update the 3D representation. It can be obtained by path derivative and reparameterization trick:", + "bbox": [ + 75, + 436, + 472, + 481 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\boldsymbol {\\theta}} J _ {E n t} (\\boldsymbol {\\theta}, \\lambda) = - \\mathbb {E} \\left[ \\omega (t) \\frac {\\partial g (\\boldsymbol {\\theta} , \\boldsymbol {c})}{\\partial \\boldsymbol {\\theta}} \\left(\\sigma_ {t} \\nabla \\log p _ {t} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {y}\\right) \\right. \\right. \\tag {8} \\\\ \\left. - \\lambda \\sigma_ {t} \\nabla \\log q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}) \\right]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 502, + 468, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The full derivation is deferred to Appendix A.2. We name this update rule as Entropic Score Distillation (ESD). Note that ESD differs from VSD as its second score function does not depend on the camera pose.", + "bbox": [ + 75, + 595, + 468, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Classifier-Free Guidance Trick", + "text_level": 1, + "bbox": [ + 76, + 665, + 349, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to SDS and VSD, we approximate $\\nabla \\log p_t(\\pmb{x}_t|\\pmb{y})$ via a pre-trained diffusion model $\\epsilon_{\\phi}(\\pmb{x}_t,t,\\pmb{y})$ . However, $\\nabla \\log q_t^\\theta (\\pmb{x}|\\pmb{y})$ is not readily available. We found that directly fine-tuning a pre-trained diffusion model using rendered images to approximate $\\nabla \\log q_t^\\theta (\\pmb {x}|\\pmb {y})$ , akin to Prolific-Dreamer, does not yield robust performance. We postulate this difficulty arises from the removal of the camera condition, increasing the complexity of the distribution to be fitted.", + "bbox": [ + 75, + 688, + 472, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To tackle this problem, we recall the result in Theorem 1 that $J_{Ent}(\\theta, \\lambda)$ can be written in terms of two KL divergence losses. Therefore, its gradient can be decomposed as a weighted combination of their gradients, which correspond to unconditional and conditional score functions in terms of", + "bbox": [ + 75, + 824, + 472, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the camera pose $c$ , respectively:", + "bbox": [ + 498, + 315, + 714, + 330 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\boldsymbol {\\theta}} J _ {E n t} (\\boldsymbol {\\theta}, \\lambda) = - \\mathbb {E} \\left[ \\omega (t) \\frac {\\partial g (\\boldsymbol {\\theta} , \\boldsymbol {c})}{\\partial \\boldsymbol {\\theta}} \\left(\\sigma_ {t} \\nabla \\log p _ {t} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {y}\\right) \\right. \\right. \\tag {9} \\\\ \\left. \\left. - \\lambda \\sigma_ {t} \\nabla \\log q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}) - (1 - \\lambda) \\sigma_ {t} \\nabla \\log q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {c}, \\boldsymbol {y})\\right) \\right]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 338, + 892, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We formally prove Eq. 9 in Appendix A.3. With the above formulation, ESD can be implemented via the Classifier-Free Guidance (CFG) trick, which was initially proposed to balance the variety and quality of text-conditionally generated images from diffusion models [10]. Algorithm 1 outlines the computation paradigm of ESD, in which we surrogate score functions in Eq. 9 with pre-trained and fine-tuned diffusion models (see Eq. 6), and takes random turns with a probability $p_{\\emptyset}$ to balance the training of conditional and unconditional score functions, as suggested by Ho and Salimans [10].", + "bbox": [ + 496, + 430, + 893, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Discussion", + "text_level": 1, + "bbox": [ + 500, + 590, + 617, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In VSD, the camera-conditioned score is believed to play a significant role in facilitating visual quality. Intuitively, such conditioning can equip the tuned diffusion model with multi-view priors [20]. Also, Hertz et al. [8] suggests such a method can be useful to stabilize the update of the implicit parameters. However, ESD counters this argument by suggesting that the camera condition might not always be advantageous, particularly when the particle size is reduced to one. In such cases, the resulting KL divergence provably degenerates to a likelihood maximization algorithm vulnerable to mode collapse.", + "bbox": [ + 496, + 613, + 893, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is noteworthy that, even though their subtle differences in implementation, the optimization objectives of ESD and VSD are fundamentally different (see Sec. 4.1). ESD sets itself apart from VSD by incorporating entropy regularization, a crucial feature absent in VSD, aiming to augment diversity across views. Despite originating from distinct objectives, our theoretical establishment allows for a straightforward implementation of ESD based on VSD using the CFG trick.", + "bbox": [ + 496, + 779, + 895, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9041", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cd94397d6e389ce5637fe38b0607cdd5ba3aa76d321d0907cbd75e123487ad50.jpg", + "image_caption": [ + "Figure 4. Qualitative Results. Our proposed outperforms all baselines in terms of better geometry and well-constructed texture details. Our results deliver photo-realistic and diverse rendered views, while baseline methods more or less suffer from the Janus problem. Best view in an electronic copy." + ], + "image_footnote": [], + "bbox": [ + 119, + 89, + 851, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We provide an illustrative example by leveraging SDS, VSD and ESD (with different $\\lambda$ 's) to fit a 2D Gaussian distribution in Fig. 3. With SDS and VSD, all samples are converged to the high-density area while ESD recovers the entire support of the distribution. We provide more details and examples in Appendix B.", + "bbox": [ + 75, + 517, + 470, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Other Related Work", + "text_level": 1, + "bbox": [ + 76, + 623, + 272, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-to-Image Diffusion Model. Text-to-image diffusion models [32, 33] are cornerstone components of text-to-3D generation. It involves text embedding conditioning into the iterative denoising process. Equipped with large-scale image-text paired datasets, many works [29, 33, 35] scale up to tackle text-to-image generation. Among them, latent diffusion models attracted great interest in the open-source community since they reduced the computation cost by diffusing in the low-resolution latent space instead of directly in the pixel space. In addition, text-to-image diffusion models have also found applications in various computer vision tasks, including text-to-3D [31, 43], image-to-3D [59], text-to-svq [17], text-to-video [18, 42], etc.", + "bbox": [ + 75, + 648, + 470, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Generation with 2D Priors. Well-annotated 3D data requires immense effort to collect. Instead, a line of research studies on how to learn 3D generative models using 2D su", + "bbox": [ + 75, + 854, + 472, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pervision. Early attempts, including pi-GAN [34], EG3D [3], GRAF [37], GIRAFFE [30], adopt adversarial loss between the rendered images and natural images. DreamField [16] leverages CLIP to align NeRF with text prompts. More recently, with the rapid development of text-to-image diffusion models, diffusion-based image priors have attracted increasing interest, and score distillation has then become the dominant technique. Pioneer works DreamFusion [31] and ProlificDreamer [56] have been introduced in detail in Sec. 2. Their concurrent work SJC [54] derives the score Jacobian chaining method from another theoretical viewpoint of Perturb and Average Scoring. Even though diffusion models directly trained with 3D data nowadays demonstrate largely improved results [21, 41], score distillation still plays a pivotal role in ensuring view consistency.", + "bbox": [ + 496, + 517, + 893, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Techniques to Improve Score Distillation. Providing the empirical promise of score distillation, there have been numerous techniques proposed to improve its effectiveness. Magic3D [19] and Fantasia3D [4] utilize mesh and DMTet [40] to disentangle the optimization of geometry and texture. TextMesh [50] and 3DFuse [38] use depth-conditioned text-to-image diffusion priors that support geometry-aware texturing. Score debiasing[12] and Perp-Neg [2] study to refine the text prompts for a better 3D generation. DreamTime [14] and RED-Diff [24] investigate the timestep scheduling in the", + "bbox": [ + 496, + 750, + 895, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9042", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d398962526ba43d25c7ee07cc30b8700a417c90f1923353349837d2ec2520567.jpg", + "image_caption": [ + "Figure 5. Qualitative Results. We combine our proposed ESD with timestep scheduling in DreamTime [14] and compare it against baseline methods. Prompt: A caramic lion." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 459, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "score distillation process. HIFA [60] adopts multiple diffusion steps for distillation. Score distillation also works with auxiliary losses, including CLIP loss [59] and adversarial loss [5, 39].", + "bbox": [ + 75, + 537, + 470, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 613, + 264, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we introduce four information-theoretic metrics to numerically evaluate the generated 3D results with a particular focus on identifying Janus artifacts or mode collapse. The metrics we propose comprehensively cover four aspects: 1) the relevance with the text prompts, 2) distribution fitness, 3) rendering quality, and 4) view diversity.", + "bbox": [ + 75, + 638, + 472, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CLIP Distance. We compute the average distance between rendered images and the text embedding to reflect the relevance between generated results and the specified text prompt. Specifically, we render $N$ views from the generated 3D representations, and for each view, we obtain an embedding vector through the image encoder of a CLIP model [53]. In the meantime, we compute the text embedding utilizing the text encoder. The CLIP distance is computed as the one minus cosine similarity between the image embeddings and text embeddings averaged over all views.", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Frechet inception distance (FID). As shown in Sec. 3 and 4, score distillation essentially matches distributions via KL divergence. Hence, it becomes reasonable to employ FID to measure the distance between the image distribution $q^{\\theta}(\\pmb{x}_0|\\pmb{y})$ generated by randomly rendering 3D representation and the text-conditioned image distribution $p(\\pmb{x}_0|\\pmb{y})$ modeled by a diffusion model. We sample $N$ images using pre-trained latent diffusion model given text prompts as the ground truth image dataset, and render $N$ views uniformly distributed over a unit sphere from the optimized 3D scene as the generated image dataset. Then standard FID [9] is computed between these two sets of images. Note that FID is known to be effective in quantitatively identifying mode collapse.", + "bbox": [ + 496, + 90, + 893, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inception Quality and Variety. Thanks to our established connection with mode collapse, we know that Janus problem is due to a lack of sample diversity. Inspired by Inception Score (IS) [36], we utilize entropy-related metrics to reflect the generated image quality and diversity. We propose Inception Quality (IQ) and Inception Variety (IV), formulated as below:", + "bbox": [ + 496, + 320, + 893, + 426 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nI Q (\\boldsymbol {\\theta}) = \\mathbb {E} _ {\\boldsymbol {c}} \\left[ H \\left[ p _ {c l s} (\\boldsymbol {y} | g (\\boldsymbol {\\theta}, \\boldsymbol {c})) \\right] \\right], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 438, + 890, + 454 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nI V (\\boldsymbol {\\theta}) = H \\left[ \\mathbb {E} _ {\\boldsymbol {c}} \\left[ p _ {c l s} (\\boldsymbol {y} | g (\\boldsymbol {\\theta}, \\boldsymbol {c}) \\right] \\right], \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 455, + 890, + 473 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $p_{cls}(\\boldsymbol{y}|\\boldsymbol{x})$ is a pre-trained classifier. IQ computes the average entropy of the label logits predicted for all rendered views, while IV computes the entropy of the averaged label logits of all rendered views. Intuitively, the smaller IQ means highly confident classification results on rendered views, which also indicates better visual quality of generated 3D assets. In the meanwhile, the higher IV signifies that each rendered view is likely to have a distinct label prediction, meaning the 3D creation has higher view diversity. Note that IV upper bounds IQ due to Jensen inequality. So we can define Inception Gain $IG = (IV - IQ) / IQ$ , which characterizes the information gain brought by knowing where the camera pose is, namely the improvement of distinguishability among different views.", + "bbox": [ + 496, + 484, + 893, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7. Experiments", + "text_level": 1, + "bbox": [ + 500, + 709, + 633, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Settings. In this section, we empirically validate the effectiveness of our proposal. The chosen prompts are targeted at objects with clearly defined canonical views, posing a challenge for existing methods. Our baseline approaches include SDS (DreamFusion) [31] and VSD (ProlificDreamer) [56], as well as two methods dedicated to solving Janus problem: Debiased-SDS [12] and Perp-Neg [2]. For fair comparison, all experiments are benchmarked under the open-source threestudio framework. Geometry refinement [56] is adopted for all distillation schemes. Please refer to Appendix C for more implementation details.", + "bbox": [ + 496, + 734, + 895, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9043", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/818f7bf7fa32591a0ca0eb1e8e55e6964ac177856f04e013e3e4e1a7b9560901.jpg", + "image_caption": [ + "$\\lambda = 1.0$" + ], + "image_footnote": [], + "bbox": [ + 117, + 112, + 357, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/67d990746d35d70bff52a1f66c12fcbada67207c870ca4b0d6c2823254550fa8.jpg", + "image_caption": [ + "$\\lambda = 0.5$", + "Figure 6. Ablation Studies on $\\lambda$ . We investigate the choice of different entropy regularization strength $\\lambda$ . Prompt: Michelangelo-style statue of dog reading news on a cellphone." + ], + "image_footnote": [], + "bbox": [ + 362, + 112, + 602, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c8f6ac5623ee5032e1fbb441d7bd99d83d49dd39c2c2f303233787d1eeb342f2.jpg", + "image_caption": [ + "$\\lambda = 0.0$" + ], + "image_footnote": [], + "bbox": [ + 607, + 111, + 846, + 205 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6cf0732716333822d2e60ac4e46ce259dcd64041f25246c916a8243019c649b8.jpg", + "table_caption": [ + "Table 1. Quantitative Comparisons. $(\\downarrow)$ means the lower the better, and $(\\uparrow)$ means the higher the better." + ], + "table_footnote": [], + "table_body": "
CLIP (↓)FID (↓)IQ (↓)IV (↑)IG (↑)SR (↑)
SDS0.737291.8604.2954.85520.12315.00%
VSD0.725265.1413.1493.57120.13719.17%
ESD0.714235.9153.1354.03140.32755.83%
", + "bbox": [ + 81, + 285, + 467, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Comparison. We present qualitative comparisons in Fig. 4. We encourage interested readers to Appendix D for more results and our project page for videos. It is clearly shown that our proposed ESD delivers more precise geometry with the Janus problem alleviated. In comparison, the results presented by SDS and VSD all contain more or less corrupted geometry with multi-face structures. Debiased-SDS and Perp-Neg are shown to be effective for some text prompts, while not so consistent as ESD. Additionally, we find that ESD can work particularly well when combined with the time-prioritized scheduling proposed in DreamTime [14], as shown in Fig. 5. This means ESD is orthogonal to many other methods and can cooperate with them to further reduce Janus artifacts.", + "bbox": [ + 75, + 357, + 468, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative Comparison. With metrics proposed in Sec. 6, we numerically evaluate our method and baselines across 120 text prompts provided in [58]. We additionally involve Successful generation Rate (SR) based on human evaluation. The results are presented in Tab. 1. We observe that among all metrics, ESD reaches the best CLIP score, FID, and IG. More importantly, ESD achieves the optimal balance between view quality and diversity as shown by IQ and IV. Whereas, SDS suffers from low image quality with high IQ and VSD is limited by insufficient view variety with low IV. The superior IG of ESD indicates that views inside the generated scene are distinguishable rather than collapsing to be the same. We defer the breakdown table for numerical evaluation on examples in Fig. 4, human evaluation criteria, and the standard deviation of metrics to Appendix E.", + "bbox": [ + 75, + 580, + 468, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Studies We conduct ablation studies on the choice of $\\lambda$ (i.e. CFG weights) in Fig. 6. We demonstrate that $\\lambda$ can adjust ESD's preference toward view- quality or diversity. When set to one, the produced Janus-free result albeit with fewer realistic details in the textures. Conversely,", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c1c5500abdd311042b5fd6bbeed3589c5562d03e2c38ef76c80578d5a2ae186c.jpg", + "image_caption": [ + "ESD w/ CFG", + "Step $= 25\\mathrm{k}$", + "Figure 7. Ablation on Implementations. The successfully generated result is obtained via our suggested CFG trick while the diverged result is yielded by fitting the unconditioned score function in Eq. 8 via LoRA. Prompt: an elephant skull." + ], + "image_footnote": [], + "bbox": [ + 501, + 263, + 581, + 324 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fc7f3df07ccbf19d99e77b818a19c42c2ccb27970da651e31af9b96f5c7b26ce.jpg", + "image_caption": [ + "ESD w/ fine-tuned unconditional LoRA", + "Step $= 18\\mathrm{k}$", + "Step $= 20\\mathrm{k}$", + "Step $= 22\\mathrm{k}$" + ], + "image_footnote": [], + "bbox": [ + 588, + 263, + 815, + 324 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9226588b631f302947307531af3bccd832931adc667f129e3907da71c64f7363.jpg", + "image_caption": [ + "Step $= 25\\mathrm{k}$" + ], + "image_footnote": [], + "bbox": [ + 821, + 263, + 888, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "when set to zero, ESD equates to VSD, and the Janus problem emerges again. We empirically find that choosing $\\lambda$ around 0.5 yields the best result, balancing fine-grained textures and well-constructed geometry. We also implement ESD by directly fitting the score function $\\nabla \\log q_t^\\theta (\\pmb{x}_t|\\pmb{y})$ without camera pose conditioning to validate the suggested implementation by CFG trick. We show in Fig. 7 that this optimization scheme is unstable. As training proceeds, the gradient explodes, and the optimized texture overflows.", + "bbox": [ + 496, + 412, + 892, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 561, + 617, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we reveal that existing score distillation methods degenerate to maximal likelihood seeking on each view independently, leading to the mode collapse problem. We identify that re-establishing the entropy term in the variational objective brings a new update rule, called Entropic Score Distillation (ESD), which is theoretically equivalent to adopting classifier-free guidance trick upon variational score distillation. ESD maximizes the entropy of the rendered image distribution, encouraging diversity across views and mitigating the Janus problem.", + "bbox": [ + 496, + 585, + 892, + 737 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 756, + 633, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "P Wang is sincerely grateful for constructive feedback regarding this manuscript from Zhaoyang Lv, Xiaoyu Xiang, Amit Kumar, Jinhui Xiong, and Varun Nagaraja. P Wang also thanks Ruisi Cai for providing decent visual materials for illustration purposes. Any statements, opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of their employers or the supporting entities.", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9044", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017. 4", + "[2] Mohammadreza Armandpour, Huangjie Zheng, Ali Sadeghian, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968, 2023. 2, 6, 7, 8", + "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 6", + "[4] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 1, 3, 6", + "[5] Yiwen Chen, Chi Zhang, Xiaofeng Yang, Zhongang Cai, Gang Yu, Lei Yang, and Guosheng Lin. It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473, 2023. 7", + "[6] Giannis Daras, Kulin Shah, Yuval Dagan, Aravind Gollakota, Alex Dimakis, and Adam Klivans. Ambient diffusion: Learning clean distributions from corrupted data. Advances in Neural Information Processing Systems, 36, 2024. 6", + "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 4", + "[8] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. arXiv preprint arXiv:2304.07090, 2023. 5", + "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7", + "[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 5", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[12] Susung Hong, Donghoon Ahn, and Seungryong Kim. Debi-asing scores and prompts of 2d diffusion for robust text-to-3d generation. arXiv preprint arXiv:2303.15413, 2023. 1, 2, 3, 6, 7, 8", + "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3", + "[14] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023. 3, 6, 7, 8" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005. 3", + "[16] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 6", + "[17] Ajay Jain, Amber Xie, and Pieter Abbeel. Vectorfusion: Text-to-svq by abstracting pixel-based diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1911-1920, 2023. 6", + "[18] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 6", + "[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 1, 3, 6", + "[20] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023.5", + "[21] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 6", + "[22] Ziming Liu, Di Luo, Yilun Xu, Tommi Jaakkola, and Max Tegmark. Genphys: From physical processes to generative models. arXiv preprint arXiv:2304.02637, 2023. 3", + "[23] Jonathan Lorraine, Kevin Xie, Xiaohui Zeng, Chen-Hsuan Lin, Towaki Takikawa, Nicholas Sharp, Tsung-Yi Lin, Ming-Yu Liu, Sanja Fidler, and James Lucas. Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349, 2023.8", + "[24] Morteza Mardani, Jiaming Song, Jan Kautz, and Arash Vahdat. A variational perspective on solving inverse problems with diffusion models. arXiv preprint arXiv:2305.04391, 2023.6", + "[25] Luke Metz, Ben Poole, David Pfau, and Jascha Sohl-Dickstein. Unrolled generative adversarial networks. arXiv preprint arXiv:1611.02163, 2016. 4", + "[26] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 1, 3", + "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 3" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9045", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022.6", + "[29] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 6", + "[30] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 6", + "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 3, 4, 6, 7, 8", + "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 6", + "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 6", + "[34] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 6", + "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems, 2022. 6", + "[36] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 2, 4, 7, 8", + "[37] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 6", + "[38] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Jaehoon Ko, Hyeonsu Kim, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937, 2023. 6", + "[39] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Dynamic portrait editing by learning 4d gan from 2d diffusion-based editor. arXiv preprint arXiv:2305.20082, 2023. 7", + "[40] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 6" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 6", + "[42] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.6", + "[43] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023.6", + "[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2", + "[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3", + "[46] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2", + "[47] Yang Song, Sahaj Garg, Jiaxin Shi, and Stefano Ermon. Sliced score matching: A scalable approach to density and score estimation. In Uncertainty in Artificial Intelligence, pages 574-584. PMLR, 2020. 3", + "[48] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 3", + "[49] Akash Srivastava, Lazar Valkov, Chris Russell, Michael U Gutmann, and Charles Sutton. Veegan: Reducing mode collapse in gans using implicit variational learning. Advances in neural information processing systems, 30, 2017. 4", + "[50] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439, 2023. 1, 3, 6", + "[51] Cédric Villani et al. Optimal transport: old and new. Springer, 2009. 3", + "[52] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 3", + "[53] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 7", + "[54] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 1, 3, 6", + "[55] Peihao Wang, Zhiwen Fan, Dejia Xu, Dilin Wang, Sreyas Mohan, Forrest Iandola, Rakesh Ranjan, Yilei Li, Qiang Liu," + ], + "bbox": [ + 501, + 90, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9046", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zhangyang Wang, et al. Steindreamer: Variance reduction for text-to-3d score distillation via stein identity. arXiv preprint arXiv:2401.00604, 2023.", + "[56] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 2, 3, 4, 6, 7, 8", + "[57] Zhendong Wang, Yifan Jiang, Huangjie Zheng, Peihao Wang, Pengcheng He, Zhangyang Wang, Weizhu Chen, Mingyuan Zhou, et al. Patch diffusion: Faster and more data-efficient training of diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 6", + "[58] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. arXiv preprint arXiv:2401.04092, 2024. 8, 7", + "[59] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360 \\{\\neg\\} views. arXiv preprint arXiv:2211.16431, 2022. 3, 6, 7", + "[60] Joseph Zhu and Peiye Zhuang. Hifa: High-fidelity text-to-3d with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 7" + ], + "bbox": [ + 78, + 90, + 470, + 429 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9047", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/695a2b86-fc93-420e-a626-d23f402233d8_model.json b/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/695a2b86-fc93-420e-a626-d23f402233d8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..398c29e1d942ad148ccf97643c214058105fba19 --- /dev/null +++ b/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/695a2b86-fc93-420e-a626-d23f402233d8_model.json @@ -0,0 +1,2389 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.133, + 0.131, + 0.84, + 0.152 + ], + "angle": 0, + "content": "Taming Mode Collapse in Score Distillation for Text-to-3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.18, + 0.83, + 0.234 + ], + "angle": 0, + "content": "Peihao Wang\\(^{1}\\), Dejia Xu\\(^{1}\\), Zhiwen Fan\\(^{1}\\), Dilin Wang\\(^{2}\\), Sreyas Mohan\\(^{2}\\), Forrest Iandola\\(^{2}\\), Rakesh Ranjan\\(^{2}\\), Yilei Li\\(^{2}\\), Qiang Liu\\(^{1}\\), Zhangyang Wang\\(^{1}\\), Vikas Chandra\\(^{2}\\) \n\\(^{1}\\)The University of Texas at Austin, \\(^{2}\\)Meta Reality Labs" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.237, + 0.812, + 0.287 + ], + "angle": 0, + "content": "{peihaowang,iejia, zhiwenfan, atlaswang}@utexas.edu, lqiang@cs.utexas.edu \n{wdilin, sreyasmohan, fni, rakeshr, yileil, vchandra}@meta.com \nvita-group.github.io/3D-Mode-Collapse/" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.474, + 0.732 + ], + "angle": 0, + "content": "Despite the remarkable performance of score distillation in text-to-3D generation, such techniques notoriously suffer from view inconsistency issues, also known as \"Janus\" artifact, where the generated objects fake each view with multiple front faces. Although empirically effective methods have approached this problem via score debiasing or prompt engineering, a more rigorous perspective to explain and tackle this problem remains elusive. In this paper, we reveal that the existing score distillation-based text-to-3D generation frameworks degenerate to maximal likelihood seeking on each view independently and thus suffer from the mode collapse problem, manifesting as the Janus artifact in practice. To tame mode collapse, we improve score distillation by re-establishing the entropy term in the corresponding variational objective, which is applied to the distribution of rendered images. Maximizing the entropy encourages diversity among different views in generated 3D assets, thereby mitigating the Janus problem. Based on this new objective, we derive a new update rule for 3D score distillation, dubbed Entropic Score Distillation (ESD). We theoretically reveal that ESD can be simplified and implemented by just adopting the classifier-free guidance trick upon variational score distillation. Although embarrassingly straightforward, our extensive experiments demonstrate that ESD can be an effective treatment for Janus artifacts in score distillation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.761, + 0.21, + 0.777 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.786, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Recent advancements in text-to-3D technology have attracted considerable attention, particularly for its pivotal role in automating high-quality 3D content. This is especially crucial in fields such as virtual reality and gaming, where 3D content forms the bedrock. While numerous techniques are available, the prevailing text-to-3D approach is based on" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.324, + 0.892, + 0.353 + ], + "angle": 0, + "content": "score distillation [31], popularized by DreamFusion and its follow-up works [4, 19, 26, 50, 54, 56]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.895, + 0.519 + ], + "angle": 0, + "content": "Score distillation leverages a pre-trained 2D diffusion model to sample over the 3D parameter space (i.e. Neural Radiance Fields (NeRF) [27]) such that views rendered from a random angle satisfy the statistics of the image distribution. This algorithm is implemented by backpropagating the estimated score of each view via the chain rule. Despite the notable progress achieved with score distillation-based approaches, it is widely observed that 3D content generated using score distillation suffers from the Janus problem [12], referring to the artifacts that generated 3D objects contain multiple canonical views (see Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.897, + 0.746 + ], + "angle": 0, + "content": "To understand this drawback of score distillation, we draw the theoretical connection between the Janus problem and mode collapse, a statistical term describing a distribution concentrating on the high-density area while losing information about the probability tail. We first uncover that the optimization of existing score distillation-based text-to-3D generation degenerates to a maximum likelihood objective, making it susceptible to model collapse. As pre-trained diffusion models are biased to frequently encountered views [12], this oversight leads all views opt to convergence toward the point with the highest likelihood, manifesting as the Janus artifact in practical applications. The main limitation of current methods is that their distillation objectives solely maximize the likelihood of each view independently, without considering the diversity between different views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.895, + 0.868 + ], + "angle": 0, + "content": "To address the aforementioned issue, we propose a principled approach Entropic Score Distillation (ESD), which regularizes the score distillation process by entropy maximization of the rendered image distribution, thereby enhancing the diversity of views in generated 3D assets and alleviating the Janus problem. Our derived ESD update admits a simple form as a weighted combination of scores for pre-trained image distribution and rendered image distribution. Com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1For example, it is common that a frontal view of a cat is more likely to be sampled from latent diffusion models than the back view." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.332, + 0.9 + ], + "angle": 0, + "content": "*Work done during an internship with Meta." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9037" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.317, + 0.097, + 0.396, + 0.11 + ], + "angle": 0, + "content": "Our results" + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.114, + 0.591, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.208, + 0.417, + 0.22 + ], + "angle": 0, + "content": "A cat with tiger stripes" + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.096, + 0.78, + 0.11 + ], + "angle": 0, + "content": "\"Janus\" results" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.114, + 0.848, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.222, + 0.591, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.317, + 0.446, + 0.33 + ], + "angle": 0, + "content": "A bald eagle carved out of wood" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.221, + 0.848, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.332, + 0.591, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.427, + 0.517, + 0.441 + ], + "angle": 0, + "content": "A chimpanzee dressed like Henry VIII king of England" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.332, + 0.848, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.427, + 0.704, + 0.439 + ], + "angle": 0, + "content": "Front View" + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.427, + 0.82, + 0.439 + ], + "angle": 0, + "content": "Back View" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.447, + 0.893, + 0.488 + ], + "angle": 0, + "content": "Figure 1. A Preview of Qualitative Results. We present the front and back views of objects synthesized by VSD (ProlificDreamer) on the right two columns, and four views of our generated results on the left. VSD suffers from \"Janus\" problem, where both front and back views contain a frontal face of the targeted object, while our method effectively mitigates this artifact. Please refer to more results in Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.471, + 0.698 + ], + "angle": 0, + "content": "pared with Score Distillation Sampling (SDS) [31], our ESD involves the score of the rendered image distribution, serving to maximize the entropy of the rendered image distribution. Unlike Variational Score Distillation (VSD) [56], the learned score function of the rendered image distribution does not depend on the camera pose. This subtle difference has a more profound impact, as we show the score function of rendered images modeled by VSD corresponds to an objective with fixed entropy, thereby having no influence on view variety. In contrast, ESD optimizes for a Kullback-Leibler divergence with a non-constant entropy term parameterized by the 3D model, leading to an effect that encourages diversity among different views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.471, + 0.867 + ], + "angle": 0, + "content": "In practice, we find it challenging to optimize the score of the rendered image distribution without conditioning on the camera pose. To facilitate training, we discover that the gradient from the entropy can be decomposed into a combination of scores: one depends on the camera pose, and the other independent of it, with a coefficient interacting between these two terms. Through this theoretical establishment, we are able to adopt a handy implementation of ESD by Classifier Free Guidance (CFG) trick [10] where conditional and unconditional scores are trained alternatively and mixed during inference." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Through extensive experiments with our proposed ESD, we demonstrate its efficacy in alleviating the Janus problem" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.895, + 0.623 + ], + "angle": 0, + "content": "and its significant advantages in improving 3D generation quality when compared to the baseline methods [31, 56] and other remedy techniques [2, 12]. As a side contribution, we also borrow two inception scores [36] to evaluate text-to-3D results and numerically probe model collapse in score distillation. We show these two metrics can effectively characterize the quality and diversity of views, highly matching our qualitative observations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.64, + 0.628, + 0.657 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.668, + 0.68 + ], + "angle": 0, + "content": "2.1. Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Diffusion models, as demonstrated by various works [11, 44, 46, 48], have shown to be highly effective in text-to-image generation. Technically, a diffusion model learns to gradually transform a normal distribution \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) to the target distribution \\(p_{data}(\\boldsymbol {x}|\\boldsymbol {y})\\) where \\(\\pmb{y}\\) denotes the text prompt embeddings. The sampling trajectory is determined by a forward process with the conditional probability \\(p_t(\\pmb {x}_t|\\pmb {x}_0) = \\mathcal{N}(\\pmb {x}_t|\\alpha_t\\pmb {x}_0,\\sigma_t^2\\pmb {I})\\) ,where \\(\\pmb {x}_t\\in \\mathbb{R}^D\\) represents the sample at time \\(t\\in [0,T]\\) , and \\(\\alpha_{t},\\sigma_{t} > 0\\) are time-dependent diffusion coefficients. Consequently, the distribution at time \\(t\\) can be formulated as \\(p_t(\\pmb {x}_t|\\pmb {y}) =\\) \\(\\int p_{data}(\\pmb {x}_0|\\pmb {y})\\mathcal{N}(\\pmb {x}_t|\\alpha_t\\pmb {x}_0,\\sigma_t^2\\pmb {I})d\\pmb {x}_0.\\) Diffusion models generate samples through a reverse process starting from Gaussian noises, which can be described by the ODE:" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9038" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.086, + 0.443, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.471, + 0.087, + 0.855, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.249, + 0.895, + 0.292 + ], + "angle": 0, + "content": "Figure 2. Illustration of the effect of entropy regularization. Learned image distributions often exhibit a higher probability mass for objects' frontal faces. Pure maximal likelihood seeking is opt to mode collapse (Sec. 3). Adding entropy regularization can expand the support of fitted distribution \\( q_{t}^{\\theta}(\\boldsymbol{x}|\\boldsymbol{y}) \\) with mode-covering behavior (Sec. 4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.471, + 0.381 + ], + "angle": 0, + "content": "\\(\\mathrm{d}\\pmb{x}_t / \\mathrm{d}t = -\\nabla_{\\pmb{x}}\\log p_t(\\pmb{x}_t)\\) with the boundary condition \\(\\pmb{x}_T\\sim \\mathcal{N}(\\pmb {0},\\pmb {I})\\) [22, 45, 48]. Such a process requires the computation of score function \\(\\nabla_{\\pmb{x}}\\log p_t(\\pmb{x}_t)\\) which is often obtained by fitting a time-conditioned noise estimator \\(\\epsilon_{\\phi}:\\mathbb{R}^{D}\\to \\mathbb{R}^{D}\\) using score matching loss [15, 47, 52]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.393, + 0.332, + 0.407 + ], + "angle": 0, + "content": "2.2. Text-to-3D Score Distillation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.47, + 0.538 + ], + "angle": 0, + "content": "Score distillation based 3D asset generation requires representing 3D scenes as learnable parameters \\(\\pmb{\\theta} \\in \\mathbb{R}^{N}\\) equipped with a differentiable renderer \\(g(\\pmb{\\theta}, \\pmb{c}): \\mathbb{R}^{N} \\to \\mathbb{R}^{D}\\) that projects 3D scene \\(\\pmb{\\theta}\\) into images with respect to the camera pose \\(\\pmb{c}\\). Here \\(N, D\\) are the dimensions of the 3D parameter space and rendered images, respectively. Neural radiance fields (NeRF) [27] are often employed as the underlying 3D representation for its capability of modeling complex scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.471, + 0.63 + ], + "angle": 0, + "content": "Recent works [4, 14, 19, 26, 31, 50, 54-56] demonstrate the feasibility of using a pretrained 2D diffusion model to guide 3D object creation. Below, we elaborate on two score distillation schemes, adopted therein: Score Distillation Sampling (SDS) [31] and Variational Score Distillation (VSD) [56]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.469, + 0.682 + ], + "angle": 0, + "content": "Score Distillation Sampling (SDS). SDS updates the 3D parameter \\(\\theta\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.695, + 0.484, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {\\theta}} J _ {S D S} (\\boldsymbol {\\theta}) = - \\mathbb {E} \\left[ \\omega (t) \\frac {\\partial g (\\boldsymbol {\\theta} , \\boldsymbol {c})}{\\partial \\boldsymbol {\\theta}} \\left(\\sigma_ {t} \\nabla \\log p _ {t} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {y}\\right) - \\boldsymbol {\\epsilon}\\right) \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.471, + 0.863 + ], + "angle": 0, + "content": "where the expectation is taken over timestep \\(t\\sim \\mathcal{U}[0,T]\\) Gaussian noises \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) and camera pose \\(\\pmb {c}\\sim p_c(\\pmb {c})\\) Here is \\(\\nabla \\log p\\) is a pre-trained diffusion model \\(\\epsilon_{\\phi}(\\pmb {x},t,\\pmb {y})\\) and \\(\\pmb{x}_t\\) is a noisy version of the rendering given by camera pose \\(\\pmb {c}\\) \\(x_{t} = \\alpha_{t}g(\\pmb {\\theta},\\pmb {c}) + \\sigma_{t}\\pmb{\\epsilon}\\) .Updating \\(\\pmb{\\theta}\\) as in Eq. (1) has been shown to minimize the evidence lower bound (ELBO) for the rendered images, see Wang et al. [54], Xu et al. [59]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.305, + 0.894, + 0.351 + ], + "angle": 0, + "content": "Variational Score Distillation (VSD). VSD [56] is introduced in ProlificDreamer, VSD improves upon SDS by deriving the following Wasserstein gradient flow [51]:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.36, + 0.892, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\boldsymbol {\\theta}} J _ {V S D} (\\boldsymbol {\\theta}) = - \\mathbb {E} \\left[ \\omega (t) \\frac {\\partial g (\\boldsymbol {\\theta} , \\boldsymbol {c})}{\\partial \\boldsymbol {\\theta}} \\left(\\sigma_ {t} \\nabla \\log p _ {t} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}\\right) \\right. \\\\ \\left. - \\sigma_ {t} \\nabla \\log q _ {t} (\\boldsymbol {x} _ {t} | \\boldsymbol {c})\\right) \\Bigg ]. \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.439, + 0.893, + 0.529 + ], + "angle": 0, + "content": "Similarly, \\( \\pmb{x}_t = \\alpha_t g(\\pmb{\\theta}, \\pmb{c}) + \\sigma_t \\pmb{\\epsilon} \\) is the noisy observation of the rendered image. In contrast to SDS, VSD introduces a new score function of the noisy rendered images conditioned on the camera pose \\( \\mathbf{c} \\). To obtain this score, Wang et al. [56] fine-tunes a diffusion model using images rendered from the 3D scene as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.54, + 0.892, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {\\psi}} \\mathbb {E} \\left[ \\omega (t) \\| \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\psi}} \\left(\\alpha_ {t} g (\\boldsymbol {\\theta}, \\boldsymbol {c}) + \\sigma_ {t} \\boldsymbol {\\epsilon}, t, \\boldsymbol {c}, \\boldsymbol {y}\\right) - \\boldsymbol {\\epsilon} \\| _ {2} ^ {2} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.575, + 0.894, + 0.637 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\psi}(\\boldsymbol{x},t,\\boldsymbol{c},\\boldsymbol{y})\\) is the noise estimator of \\(\\nabla \\log q_t(\\boldsymbol{x}_t|\\boldsymbol{c})\\) as in diffusion models. As proposed in ProlificDreamer, \\(\\psi\\) is parameterized by LoRA [13] and initialized from a pre-trained diffusion model same as \\(\\nabla \\log p_t\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.892, + 0.667 + ], + "angle": 0, + "content": "3. Revealing Mode Collapse in Score Distillation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Despite the remarkable performance of SDS and VSD in 3D asset generation, it is widely observed that the synthesized objects suffer from \"Janus\" artifacts. Janus artifacts refer to the generated 3D scene containing multiple canonical views (the most representative perspective of the object such as the frontal face). In earlier works, Hong et al. [12] and Huang et al. [14] attribute this problem to unimodality of the learned 2D image distribution since the training data for the diffusion models are naturally biased to the most commonly seen views per each category. In this section, we examine extant distillation schemes from a statistical view, which has been overlooked in previous literature." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In principle, natural 2D images can be seen as random projections of 3D scenes. Score distillation matches the image distribution generated by randomly sampled views with a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.901 + ], + "angle": 0, + "content": "2Without special specification, expectations are taken over all relevant random variables and Jacobian matrices are transposed by default." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9039" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.227 + ], + "angle": 0, + "content": "text-conditioned image distribution to recover the underlying 3D representation. Hence, Janus artifact, in which each view becomes uniform and identical to the most commonly seen views, can be interpreted as a manifestation of distribution collapse to samples within the high-density region. Such distribution degeneration essentially corresponds to the statistical phenomenon mode collapse, which happens when an optimized distribution fails to characterize the data diversity and concentrates on a single type of output [1, 7, 25, 36, 49]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.228, + 0.473, + 0.304 + ], + "angle": 0, + "content": "Below we theoretically reveal why SDS and VSD are prone to mode collapse. As shown in Poole et al. [31], Wang et al. [56], SDS and VSD equals to the gradient of the following Kullback-Leibler (KL) divergence, i.e., \\( J_{SDS}(\\pmb{\\theta}) = J_{VSD}(\\pmb{\\theta}) = J_{KL}(\\pmb{\\theta}) \\) up to an additive constant:" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.313, + 0.471, + 0.332 + ], + "angle": 0, + "content": "\\[\nJ _ {K L} (\\boldsymbol {\\theta}) = \\mathbb {E} \\left[ \\Omega (t) \\mathcal {D} _ {\\mathrm {K L}} \\left(q _ {t} ^ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {t} | \\boldsymbol {c}, \\boldsymbol {y}\\right) \\| p _ {t} \\left(\\boldsymbol {x} _ {t} | \\boldsymbol {y}\\right)\\right) \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.473, + 0.492 + ], + "angle": 0, + "content": "where \\(\\Omega(t) = \\omega(t)\\sigma_t / \\alpha_t\\) and the expectation is taken over \\(t \\sim \\mathcal{U}[0,T]\\) and \\(\\pmb{c} \\sim p_c(\\pmb{c})\\). Here \\(p_t(\\pmb{x}_t|\\pmb{y}) = \\int p_0(\\pmb{x}_0|\\pmb{y})\\mathcal{N}(\\pmb{x}_t|\\alpha_t\\pmb{x}_0,\\sigma_t^2\\pmb{I})d\\pmb{x}_0\\) is the image distribution perturbed by Gaussian noises, while \\(q_t^\\theta (\\pmb{x}_t|\\pmb{c},\\pmb{y}) = \\int q_0^\\theta (\\pmb{x}_0|\\pmb{c})\\mathcal{N}(\\pmb{x}_t|\\alpha_t\\pmb{x}_0,\\sigma_t^2\\pmb{I})d\\pmb{x}_0\\) models the image distribution generated by 3D parameter \\(\\pmb{\\theta}\\) with respect to camera pose \\(\\pmb{c}\\) and diffused by Gaussian distribution. As shown by Wang et al. [56], \\(J_{KL}(\\pmb{\\theta}) = 0\\) implies \\(q_0^\\theta (\\pmb{x}_0|\\pmb{c}) = p(\\pmb{x}_0|\\pmb{y})\\), i.e., the distribution of synthesized views satisfy the text-conditioned image distribution." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.493, + 0.473, + 0.569 + ], + "angle": 0, + "content": "However, it has not escaped from our notice that \\( q_0^\\theta(\\boldsymbol{x}_0|\\boldsymbol{c}) = \\delta(\\boldsymbol{x}_0 - g(\\boldsymbol{\\theta},\\boldsymbol{c})) \\) is a Dirac distribution for both SDS and VSD. This causes the original KL divergence minimization (Eq. 4) degenerate to a Maximal Likelihood Estimation (MLE) problem:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.577, + 0.47, + 0.661 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} J _ {K L} (\\boldsymbol {\\theta}) = \\underbrace {- \\mathbb {E} \\left[ \\Omega (t) \\mathbb {E} _ {\\boldsymbol {x} _ {t} \\sim q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {c} , \\boldsymbol {y})} \\log p _ {t} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}) \\right]} _ {J _ {M L E} (\\boldsymbol {\\theta})} \\tag {5} \\\\ - \\underbrace {\\mathbb {E} \\left[ \\Omega (t) H \\left[ q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {c} , \\boldsymbol {y}) \\right] \\right]} _ {\\text {c o n s t .}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.749 + ], + "angle": 0, + "content": "where \\( H[q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{y})] = -\\mathbb{E}_{\\boldsymbol{x}_t \\sim q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{c}, \\boldsymbol{y})}[\\log q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{c}, \\boldsymbol{y})] \\) denotes the entropy of \\( q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{y}) \\), which turns out to be a constant because \\( q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{c}, \\boldsymbol{y}) = \\mathcal{N}(\\boldsymbol{x}_t | \\alpha_t g(\\boldsymbol{\\theta}, \\boldsymbol{c}), \\sigma_t^2 \\boldsymbol{I}) \\) which has fixed entropy once \\( t, \\boldsymbol{\\theta} \\) and \\( \\boldsymbol{c} \\) have been specified. See full derivation in Appendix A.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Note that Eq. 5 signifies \\( J_{KL}(\\pmb{\\theta}) = J_{MLE}(\\pmb{\\theta}) \\) up to an additive constant, hence \\( J_{KL}(\\pmb{\\theta}) \\) shares all minima with \\( J_{MLE}(\\pmb{\\theta}) \\). It is known that likelihood maximization is more prone to mode collapse. Intuitively, minimizing \\( J_{MLE}(\\pmb{\\theta}) \\) seeks each view independently to have the maximum log-likelihood on the image distribution \\( p(\\pmb{x}_0|\\pmb{y}) \\). Since \\( p(\\pmb{x}_0|\\pmb{y}) \\) is usually unimodal and peaks at the canonical view, each view of the scene will collapse to the same local minimum, resulting in Janus artifact (see Fig. 2). We postulate that the existing distillation strategies may be inherently limited" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.092, + 0.888, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.206, + 0.892, + 0.262 + ], + "angle": 0, + "content": "Figure 3. Gaussian Example. To illustrate the effects of entropy regularization, we leverage SDS, VSD and ESD to fit a 2D Gaussian distribution. The blue points are sampled from the ground-truth distribution while the orange points are from the fitted distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.276, + 0.892, + 0.32 + ], + "angle": 0, + "content": "by their log-likelihood seeking behaviors, which are more susceptible to mode collapse, especially with biased image distributions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.337, + 0.849, + 0.354 + ], + "angle": 0, + "content": "4. Entropy Regularized Score Distillation" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.373, + 0.741, + 0.389 + ], + "angle": 0, + "content": "4.1. Entropic Score Distillation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.502 + ], + "angle": 0, + "content": "In this section, we highlight the importance of the entropy in score distillation. It is known that higher entropy implies the corresponding distribution could cover a larger support of the ambient space and thus increase the sample diversity. In Eq. 5, the entropy term is shown to diminish in the training objective, which causes each generated view to lack diversity and collapse to a single image with the highest likelihood." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.894, + 0.594 + ], + "angle": 0, + "content": "To this end, we propose to bring in an entropy regularization to \\(J_{MLE}(\\pmb{\\theta})\\) for boosting the view diversity. Since \\(q_t^\\pmb{\\theta}(\\pmb{x}_t|\\pmb{c},\\pmb{y})\\) has constant entropy, we regularize entropy for the distribution \\(q_t^\\pmb{\\theta}(\\pmb{x}_t|\\pmb{y}) = \\int q_t^\\pmb{\\theta}(\\pmb{x}_t|\\pmb{c},\\pmb{y})p_c(\\pmb{c})d\\pmb{c}\\), which can be simulated by randomly sampling views from the 3D parameter \\(\\pmb{\\theta}\\). Consider the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.605, + 0.892, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} J _ {E n t} (\\boldsymbol {\\theta}, \\lambda) = - \\mathbb {E} \\left[ \\Omega (t) \\mathbb {E} _ {\\boldsymbol {x} _ {t} \\sim q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {c}, \\boldsymbol {y})} \\log p _ {t} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}) \\right] \\tag {7} \\\\ - \\lambda \\mathbb {E} \\left[ \\Omega (t) H \\left[ q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}) \\right] \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.895, + 0.844 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyper-parameter controlling the regularization strength. We note that without \\(H[q_t^\\theta (\\pmb {x}_t|\\pmb {y})]\\) , each view is optimized independently and implicitly regularized by the underlying parameterization. However, upon imposing \\(H[q_t^\\theta (\\pmb {x}_t|\\pmb {y})]\\) , all views become explicitly correlated with each other, as they collectively contribute to the entropy computation. Intuitively, \\(J_{Ent}(\\pmb {\\theta},\\lambda) = J_{MLE}(\\pmb {\\theta})-\\) \\(\\lambda \\mathbb{E}[\\Omega (t)H[q_t^\\theta (\\pmb {x}_t|\\pmb {y})]]\\) seeks the maximal log-likelihood for each view while simultaneously enlarging the entropy for distribution \\(q_{t}^{\\theta}(\\pmb{x}_{t}|\\pmb{y})\\) , which spans the support and encourages diversity across the rendered views. To gain more insights, we present the following theoretical results:" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.855, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Theorem 1. For any \\(\\lambda \\in \\mathbb{R}\\) and \\(\\pmb{\\theta} \\in \\mathbb{R}^{D}\\), we have \\(J_{Ent}(\\pmb{\\theta},\\lambda) = \\lambda \\mathbb{E}_t[\\Omega(t)\\mathcal{D}_{\\mathrm{KL}}(q_t^\\pmb{\\theta}(\\pmb{x}_t|\\pmb{y})||p_t(\\pmb{x}_t|\\pmb{y}))] + (1 - \\lambda)\\mathbb{E}_{t,\\pmb{c}}[\\Omega(t)\\mathcal{D}_{\\mathrm{KL}}(q_t^\\pmb{\\theta}(\\pmb{x}_t|\\pmb{c},\\pmb{y})||p_t(\\pmb{x}_t|\\pmb{y}))] + const.\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9040" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.079, + 0.092, + 0.545, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 ESD: Entropic score distillation for text-to-3D generation" + }, + { + "type": "algorithm", + "bbox": [ + 0.092, + 0.111, + 0.895, + 0.289 + ], + "angle": 0, + "content": "Input: A diffusion model \\(\\epsilon_{\\phi}(\\pmb {x},t,\\pmb {y})\\) ; learnable 3D parameter \\(\\theta\\) ; coefficient \\(\\lambda\\) ; text prompt \\(\\pmb{y}\\) \nInitialize \\(\\psi\\) for another diffusion model \\(\\epsilon_{\\psi}(\\pmb {x},t,\\pmb {y})\\) with the parameter \\(\\phi\\) specified in diffusion model \\(\\epsilon_{\\phi}(\\pmb {x},t,\\pmb {y})\\) , parame \nterized with LoRA. \nwhile not converged do \nRandomly sample a camera pose \\(c\\sim p_c\\) and render a view \\(\\pmb {x}_0 = g(\\pmb {\\theta},\\pmb {c})\\) from \\(\\pmb{\\theta}\\) Sample a \\(t\\sim \\mathcal{U}[0,T]\\) and add Gaussian noise \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) .. \\(\\pmb{x}_{t} = \\alpha_{t}\\pmb{x}_{0} + \\sigma_{t}\\pmb{\\epsilon}\\) \n\\(\\pmb {\\theta}\\gets \\pmb {\\theta} + \\eta_1\\left[\\omega (t)\\frac{\\partial g(\\pmb{\\theta},\\pmb{c})}{\\partial\\pmb{\\theta}} (\\epsilon_{\\phi}(\\pmb{x}_{t},t,\\pmb {y}) - \\lambda \\epsilon_{\\psi}(\\pmb{x}_{t},t,\\emptyset ,\\pmb {y}) - (1 - \\lambda)\\epsilon_{\\psi}(\\pmb{x}_{t},t,\\pmb {c},\\pmb {y})\\right]\\) (6) With probability \\(1 - p_{\\emptyset}\\) \\(\\psi \\gets \\psi -\\eta_2\\nabla_{\\psi}\\left[\\omega (t)\\| \\epsilon_{\\psi}(\\pmb {x}_t,t,\\pmb {c},\\pmb {y}) - \\epsilon \\| _2^2\\right]\\) Otherwise, \\(\\psi \\gets \\psi -\\eta_2\\nabla_{\\psi}\\left[\\omega (t)\\| \\epsilon_{\\psi}(\\pmb {x}_t,t,\\emptyset ,\\pmb {y}) - \\epsilon \\| _2^2\\right]\\) \nend while \nReturn \\(\\theta\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.316, + 0.473, + 0.436 + ], + "angle": 0, + "content": "We prove Theorem 1 in Appendix A.3. Theorem 1 implies that \\(J_{Ent}(\\theta, \\lambda)\\) essentially equal to a combination of two types of KL divergences, where the former one minimizes the distribution discrepancy between \\(q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{y})\\) and \\(p_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{y})\\) which marginalizes the camera pose within \\(q_t^\\theta\\), while the latter is the original KL divergence \\(J_{KL}(\\theta)\\) adopted by SDS and VSD which takes expectation over \\(\\boldsymbol{c}\\) out of KL divergence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.437, + 0.473, + 0.482 + ], + "angle": 0, + "content": "Next, we derive the gradient of \\( J_{Ent}(\\theta, \\lambda) \\) that will be backpropagated to update the 3D representation. It can be obtained by path derivative and reparameterization trick:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.503, + 0.47, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\boldsymbol {\\theta}} J _ {E n t} (\\boldsymbol {\\theta}, \\lambda) = - \\mathbb {E} \\left[ \\omega (t) \\frac {\\partial g (\\boldsymbol {\\theta} , \\boldsymbol {c})}{\\partial \\boldsymbol {\\theta}} \\left(\\sigma_ {t} \\nabla \\log p _ {t} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {y}\\right) \\right. \\right. \\tag {8} \\\\ \\left. - \\lambda \\sigma_ {t} \\nabla \\log q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}) \\right]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.658 + ], + "angle": 0, + "content": "The full derivation is deferred to Appendix A.2. We name this update rule as Entropic Score Distillation (ESD). Note that ESD differs from VSD as its second score function does not depend on the camera pose." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.35, + 0.681 + ], + "angle": 0, + "content": "4.2. Classifier-Free Guidance Trick" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.824 + ], + "angle": 0, + "content": "Similar to SDS and VSD, we approximate \\(\\nabla \\log p_t(\\pmb{x}_t|\\pmb{y})\\) via a pre-trained diffusion model \\(\\epsilon_{\\phi}(\\pmb{x}_t,t,\\pmb{y})\\). However, \\(\\nabla \\log q_t^\\theta (\\pmb{x}|\\pmb{y})\\) is not readily available. We found that directly fine-tuning a pre-trained diffusion model using rendered images to approximate \\(\\nabla \\log q_t^\\theta (\\pmb {x}|\\pmb {y})\\), akin to Prolific-Dreamer, does not yield robust performance. We postulate this difficulty arises from the removal of the camera condition, increasing the complexity of the distribution to be fitted." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.9 + ], + "angle": 0, + "content": "To tackle this problem, we recall the result in Theorem 1 that \\( J_{Ent}(\\theta, \\lambda) \\) can be written in terms of two KL divergence losses. Therefore, its gradient can be decomposed as a weighted combination of their gradients, which correspond to unconditional and conditional score functions in terms of" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.316, + 0.715, + 0.332 + ], + "angle": 0, + "content": "the camera pose \\(c\\), respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.339, + 0.893, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\boldsymbol {\\theta}} J _ {E n t} (\\boldsymbol {\\theta}, \\lambda) = - \\mathbb {E} \\left[ \\omega (t) \\frac {\\partial g (\\boldsymbol {\\theta} , \\boldsymbol {c})}{\\partial \\boldsymbol {\\theta}} \\left(\\sigma_ {t} \\nabla \\log p _ {t} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {y}\\right) \\right. \\right. \\tag {9} \\\\ \\left. \\left. - \\lambda \\sigma_ {t} \\nabla \\log q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}) - (1 - \\lambda) \\sigma_ {t} \\nabla \\log q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {c}, \\boldsymbol {y})\\right) \\right]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.895, + 0.584 + ], + "angle": 0, + "content": "We formally prove Eq. 9 in Appendix A.3. With the above formulation, ESD can be implemented via the Classifier-Free Guidance (CFG) trick, which was initially proposed to balance the variety and quality of text-conditionally generated images from diffusion models [10]. Algorithm 1 outlines the computation paradigm of ESD, in which we surrogate score functions in Eq. 9 with pre-trained and fine-tuned diffusion models (see Eq. 6), and takes random turns with a probability \\( p_{\\emptyset} \\) to balance the training of conditional and unconditional score functions, as suggested by Ho and Salimans [10]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.591, + 0.618, + 0.605 + ], + "angle": 0, + "content": "4.3. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.78 + ], + "angle": 0, + "content": "In VSD, the camera-conditioned score is believed to play a significant role in facilitating visual quality. Intuitively, such conditioning can equip the tuned diffusion model with multi-view priors [20]. Also, Hertz et al. [8] suggests such a method can be useful to stabilize the update of the implicit parameters. However, ESD counters this argument by suggesting that the camera condition might not always be advantageous, particularly when the particle size is reduced to one. In such cases, the resulting KL divergence provably degenerates to a likelihood maximization algorithm vulnerable to mode collapse." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.897, + 0.901 + ], + "angle": 0, + "content": "It is noteworthy that, even though their subtle differences in implementation, the optimization objectives of ESD and VSD are fundamentally different (see Sec. 4.1). ESD sets itself apart from VSD by incorporating entropy regularization, a crucial feature absent in VSD, aiming to augment diversity across views. Despite originating from distinct objectives, our theoretical establishment allows for a straightforward implementation of ESD based on VSD using the CFG trick." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9041" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.09, + 0.852, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.462, + 0.895, + 0.504 + ], + "angle": 0, + "content": "Figure 4. Qualitative Results. Our proposed outperforms all baselines in terms of better geometry and well-constructed texture details. Our results deliver photo-realistic and diverse rendered views, while baseline methods more or less suffer from the Janus problem. Best view in an electronic copy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.471, + 0.61 + ], + "angle": 0, + "content": "We provide an illustrative example by leveraging SDS, VSD and ESD (with different \\(\\lambda\\)'s) to fit a 2D Gaussian distribution in Fig. 3. With SDS and VSD, all samples are converged to the high-density area while ESD recovers the entire support of the distribution. We provide more details and examples in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.624, + 0.274, + 0.64 + ], + "angle": 0, + "content": "5. Other Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.65, + 0.471, + 0.846 + ], + "angle": 0, + "content": "Text-to-Image Diffusion Model. Text-to-image diffusion models [32, 33] are cornerstone components of text-to-3D generation. It involves text embedding conditioning into the iterative denoising process. Equipped with large-scale image-text paired datasets, many works [29, 33, 35] scale up to tackle text-to-image generation. Among them, latent diffusion models attracted great interest in the open-source community since they reduced the computation cost by diffusing in the low-resolution latent space instead of directly in the pixel space. In addition, text-to-image diffusion models have also found applications in various computer vision tasks, including text-to-3D [31, 43], image-to-3D [59], text-to-svq [17], text-to-video [18, 42], etc." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.473, + 0.901 + ], + "angle": 0, + "content": "3D Generation with 2D Priors. Well-annotated 3D data requires immense effort to collect. Instead, a line of research studies on how to learn 3D generative models using 2D su" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.895, + 0.745 + ], + "angle": 0, + "content": "pervision. Early attempts, including pi-GAN [34], EG3D [3], GRAF [37], GIRAFFE [30], adopt adversarial loss between the rendered images and natural images. DreamField [16] leverages CLIP to align NeRF with text prompts. More recently, with the rapid development of text-to-image diffusion models, diffusion-based image priors have attracted increasing interest, and score distillation has then become the dominant technique. Pioneer works DreamFusion [31] and ProlificDreamer [56] have been introduced in detail in Sec. 2. Their concurrent work SJC [54] derives the score Jacobian chaining method from another theoretical viewpoint of Perturb and Average Scoring. Even though diffusion models directly trained with 3D data nowadays demonstrate largely improved results [21, 41], score distillation still plays a pivotal role in ensuring view consistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Techniques to Improve Score Distillation. Providing the empirical promise of score distillation, there have been numerous techniques proposed to improve its effectiveness. Magic3D [19] and Fantasia3D [4] utilize mesh and DMTet [40] to disentangle the optimization of geometry and texture. TextMesh [50] and 3DFuse [38] use depth-conditioned text-to-image diffusion priors that support geometry-aware texturing. Score debiasing[12] and Perp-Neg [2] study to refine the text prompts for a better 3D generation. DreamTime [14] and RED-Diff [24] investigate the timestep scheduling in the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9042" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.089, + 0.46, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.482, + 0.472, + 0.525 + ], + "angle": 0, + "content": "Figure 5. Qualitative Results. We combine our proposed ESD with timestep scheduling in DreamTime [14] and compare it against baseline methods. Prompt: A caramic lion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.472, + 0.6 + ], + "angle": 0, + "content": "score distillation process. HIFA [60] adopts multiple diffusion steps for distillation. Score distillation also works with auxiliary losses, including CLIP loss [59] and adversarial loss [5, 39]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.614, + 0.265, + 0.63 + ], + "angle": 0, + "content": "6. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.473, + 0.732 + ], + "angle": 0, + "content": "In this section, we introduce four information-theoretic metrics to numerically evaluate the generated 3D results with a particular focus on identifying Janus artifacts or mode collapse. The metrics we propose comprehensively cover four aspects: 1) the relevance with the text prompts, 2) distribution fitness, 3) rendering quality, and 4) view diversity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "CLIP Distance. We compute the average distance between rendered images and the text embedding to reflect the relevance between generated results and the specified text prompt. Specifically, we render \\( N \\) views from the generated 3D representations, and for each view, we obtain an embedding vector through the image encoder of a CLIP model [53]. In the meantime, we compute the text embedding utilizing the text encoder. The CLIP distance is computed as the one minus cosine similarity between the image embeddings and text embeddings averaged over all views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.304 + ], + "angle": 0, + "content": "Frechet inception distance (FID). As shown in Sec. 3 and 4, score distillation essentially matches distributions via KL divergence. Hence, it becomes reasonable to employ FID to measure the distance between the image distribution \\( q^{\\theta}(\\pmb{x}_0|\\pmb{y}) \\) generated by randomly rendering 3D representation and the text-conditioned image distribution \\( p(\\pmb{x}_0|\\pmb{y}) \\) modeled by a diffusion model. We sample \\( N \\) images using pre-trained latent diffusion model given text prompts as the ground truth image dataset, and render \\( N \\) views uniformly distributed over a unit sphere from the optimized 3D scene as the generated image dataset. Then standard FID [9] is computed between these two sets of images. Note that FID is known to be effective in quantitatively identifying mode collapse." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.895, + 0.427 + ], + "angle": 0, + "content": "Inception Quality and Variety. Thanks to our established connection with mode collapse, we know that Janus problem is due to a lack of sample diversity. Inspired by Inception Score (IS) [36], we utilize entropy-related metrics to reflect the generated image quality and diversity. We propose Inception Quality (IQ) and Inception Variety (IV), formulated as below:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.439, + 0.892, + 0.455 + ], + "angle": 0, + "content": "\\[\nI Q (\\boldsymbol {\\theta}) = \\mathbb {E} _ {\\boldsymbol {c}} \\left[ H \\left[ p _ {c l s} (\\boldsymbol {y} | g (\\boldsymbol {\\theta}, \\boldsymbol {c})) \\right] \\right], \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.457, + 0.892, + 0.474 + ], + "angle": 0, + "content": "\\[\nI V (\\boldsymbol {\\theta}) = H \\left[ \\mathbb {E} _ {\\boldsymbol {c}} \\left[ p _ {c l s} (\\boldsymbol {y} | g (\\boldsymbol {\\theta}, \\boldsymbol {c}) \\right] \\right], \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.895, + 0.697 + ], + "angle": 0, + "content": "where \\( p_{cls}(\\boldsymbol{y}|\\boldsymbol{x}) \\) is a pre-trained classifier. IQ computes the average entropy of the label logits predicted for all rendered views, while IV computes the entropy of the averaged label logits of all rendered views. Intuitively, the smaller IQ means highly confident classification results on rendered views, which also indicates better visual quality of generated 3D assets. In the meanwhile, the higher IV signifies that each rendered view is likely to have a distinct label prediction, meaning the 3D creation has higher view diversity. Note that IV upper bounds IQ due to Jensen inequality. So we can define Inception Gain \\( IG = (IV - IQ) / IQ \\), which characterizes the information gain brought by knowing where the camera pose is, namely the improvement of distinguishability among different views." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.634, + 0.727 + ], + "angle": 0, + "content": "7. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Settings. In this section, we empirically validate the effectiveness of our proposal. The chosen prompts are targeted at objects with clearly defined canonical views, posing a challenge for existing methods. Our baseline approaches include SDS (DreamFusion) [31] and VSD (ProlificDreamer) [56], as well as two methods dedicated to solving Janus problem: Debiased-SDS [12] and Perp-Neg [2]. For fair comparison, all experiments are benchmarked under the open-source threestudio framework. Geometry refinement [56] is adopted for all distillation schemes. Please refer to Appendix C for more implementation details." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9043" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.097, + 0.27, + 0.111 + ], + "angle": 0, + "content": "\\(\\lambda = 1.0\\)" + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.113, + 0.358, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.097, + 0.515, + 0.11 + ], + "angle": 0, + "content": "\\(\\lambda = 0.5\\)" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.113, + 0.603, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.097, + 0.765, + 0.11 + ], + "angle": 0, + "content": "\\(\\lambda = 0.0\\)" + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.112, + 0.848, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.213, + 0.89, + 0.24 + ], + "angle": 0, + "content": "Figure 6. Ablation Studies on \\(\\lambda\\). We investigate the choice of different entropy regularization strength \\(\\lambda\\). Prompt: Michelangelo-style statue of dog reading news on a cellphone." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.253, + 0.468, + 0.281 + ], + "angle": 0, + "content": "Table 1. Quantitative Comparisons. \\((\\downarrow)\\) means the lower the better, and \\((\\uparrow)\\) means the higher the better." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.286, + 0.468, + 0.345 + ], + "angle": 0, + "content": "
CLIP (↓)FID (↓)IQ (↓)IV (↑)IG (↑)SR (↑)
SDS0.737291.8604.2954.85520.12315.00%
VSD0.725265.1413.1493.57120.13719.17%
ESD0.714235.9153.1354.03140.32755.83%
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.568 + ], + "angle": 0, + "content": "Qualitative Comparison. We present qualitative comparisons in Fig. 4. We encourage interested readers to Appendix D for more results and our project page for videos. It is clearly shown that our proposed ESD delivers more precise geometry with the Janus problem alleviated. In comparison, the results presented by SDS and VSD all contain more or less corrupted geometry with multi-face structures. Debiased-SDS and Perp-Neg are shown to be effective for some text prompts, while not so consistent as ESD. Additionally, we find that ESD can work particularly well when combined with the time-prioritized scheduling proposed in DreamTime [14], as shown in Fig. 5. This means ESD is orthogonal to many other methods and can cooperate with them to further reduce Janus artifacts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.47, + 0.807 + ], + "angle": 0, + "content": "Quantitative Comparison. With metrics proposed in Sec. 6, we numerically evaluate our method and baselines across 120 text prompts provided in [58]. We additionally involve Successful generation Rate (SR) based on human evaluation. The results are presented in Tab. 1. We observe that among all metrics, ESD reaches the best CLIP score, FID, and IG. More importantly, ESD achieves the optimal balance between view quality and diversity as shown by IQ and IV. Whereas, SDS suffers from low image quality with high IQ and VSD is limited by insufficient view variety with low IV. The superior IG of ESD indicates that views inside the generated scene are distinguishable rather than collapsing to be the same. We defer the breakdown table for numerical evaluation on examples in Fig. 4, human evaluation criteria, and the standard deviation of metrics to Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Ablation Studies We conduct ablation studies on the choice of \\(\\lambda\\) (i.e. CFG weights) in Fig. 6. We demonstrate that \\(\\lambda\\) can adjust ESD's preference toward view- quality or diversity. When set to one, the produced Janus-free result albeit with fewer realistic details in the textures. Conversely," + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.253, + 0.568, + 0.262 + ], + "angle": 0, + "content": "ESD w/ CFG" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.264, + 0.583, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.326, + 0.563, + 0.334 + ], + "angle": 0, + "content": "Step \\(= 25\\mathrm{k}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.253, + 0.814, + 0.262 + ], + "angle": 0, + "content": "ESD w/ fine-tuned unconditional LoRA" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.264, + 0.816, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.326, + 0.65, + 0.335 + ], + "angle": 0, + "content": "Step \\(= 18\\mathrm{k}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.326, + 0.725, + 0.335 + ], + "angle": 0, + "content": "Step \\(= 20\\mathrm{k}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.325, + 0.798, + 0.334 + ], + "angle": 0, + "content": "Step \\(= 22\\mathrm{k}\\)" + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.264, + 0.889, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.831, + 0.326, + 0.873, + 0.334 + ], + "angle": 0, + "content": "Step \\(= 25\\mathrm{k}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.343, + 0.894, + 0.399 + ], + "angle": 0, + "content": "Figure 7. Ablation on Implementations. The successfully generated result is obtained via our suggested CFG trick while the diverged result is yielded by fitting the unconditioned score function in Eq. 8 via LoRA. Prompt: an elephant skull." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.893, + 0.548 + ], + "angle": 0, + "content": "when set to zero, ESD equates to VSD, and the Janus problem emerges again. We empirically find that choosing \\(\\lambda\\) around 0.5 yields the best result, balancing fine-grained textures and well-constructed geometry. We also implement ESD by directly fitting the score function \\(\\nabla \\log q_t^\\theta (\\pmb{x}_t|\\pmb{y})\\) without camera pose conditioning to validate the suggested implementation by CFG trick. We show in Fig. 7 that this optimization scheme is unstable. As training proceeds, the gradient explodes, and the optimized texture overflows." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.562, + 0.618, + 0.577 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.893, + 0.738 + ], + "angle": 0, + "content": "In this paper, we reveal that existing score distillation methods degenerate to maximal likelihood seeking on each view independently, leading to the mode collapse problem. We identify that re-establishing the entropy term in the variational objective brings a new update rule, called Entropic Score Distillation (ESD), which is theoretically equivalent to adopting classifier-free guidance trick upon variational score distillation. ESD maximizes the entropy of the rendered image distribution, encouraging diversity across views and mitigating the Janus problem." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.634, + 0.771 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.894, + 0.901 + ], + "angle": 0, + "content": "P Wang is sincerely grateful for constructive feedback regarding this manuscript from Zhaoyang Lv, Xiaoyu Xiang, Amit Kumar, Jinhui Xiong, and Varun Nagaraja. P Wang also thanks Ruisi Cai for providing decent visual materials for illustration purposes. Any statements, opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of their employers or the supporting entities." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9044" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.472, + 0.169 + ], + "angle": 0, + "content": "[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Mohammadreza Armandpour, Huangjie Zheng, Ali Sadeghian, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968, 2023. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.472, + 0.327 + ], + "angle": 0, + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.328, + 0.472, + 0.384 + ], + "angle": 0, + "content": "[4] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.472, + 0.442 + ], + "angle": 0, + "content": "[5] Yiwen Chen, Chi Zhang, Xiaofeng Yang, Zhongang Cai, Gang Yu, Lei Yang, and Guosheng Lin. It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.472, + 0.499 + ], + "angle": 0, + "content": "[6] Giannis Daras, Kulin Shah, Yuval Dagan, Aravind Gollakota, Alex Dimakis, and Adam Klivans. Ambient diffusion: Learning clean distributions from corrupted data. Advances in Neural Information Processing Systems, 36, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.472, + 0.557 + ], + "angle": 0, + "content": "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.558, + 0.472, + 0.586 + ], + "angle": 0, + "content": "[8] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. arXiv preprint arXiv:2304.07090, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.588, + 0.472, + 0.655 + ], + "angle": 0, + "content": "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.47, + 0.686 + ], + "angle": 0, + "content": "[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.47, + 0.729 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[12] Susung Hong, Donghoon Ahn, and Seungryong Kim. Debi-asing scores and prompts of 2d diffusion for robust text-to-3d generation. arXiv preprint arXiv:2303.15413, 2023. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023. 3, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.135 + ], + "angle": 0, + "content": "[15] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[16] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.894, + 0.265 + ], + "angle": 0, + "content": "[17] Ajay Jain, Amber Xie, and Pieter Abbeel. Vectorfusion: Text-to-svq by abstracting pixel-based diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1911-1920, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.267, + 0.894, + 0.336 + ], + "angle": 0, + "content": "[18] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.339, + 0.894, + 0.422 + ], + "angle": 0, + "content": "[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.424, + 0.894, + 0.479 + ], + "angle": 0, + "content": "[20] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.482, + 0.894, + 0.538 + ], + "angle": 0, + "content": "[21] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.541, + 0.894, + 0.582 + ], + "angle": 0, + "content": "[22] Ziming Liu, Di Luo, Yilun Xu, Tommi Jaakkola, and Max Tegmark. Genphys: From physical processes to generative models. arXiv preprint arXiv:2304.02637, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.584, + 0.894, + 0.653 + ], + "angle": 0, + "content": "[23] Jonathan Lorraine, Kevin Xie, Xiaohui Zeng, Chen-Hsuan Lin, Towaki Takikawa, Nicholas Sharp, Tsung-Yi Lin, Ming-Yu Liu, Sanja Fidler, and James Lucas. Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.657, + 0.894, + 0.711 + ], + "angle": 0, + "content": "[24] Morteza Mardani, Jiaming Song, Jan Kautz, and Arash Vahdat. A variational perspective on solving inverse problems with diffusion models. arXiv preprint arXiv:2305.04391, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.715, + 0.894, + 0.757 + ], + "angle": 0, + "content": "[25] Luke Metz, Ben Poole, David Pfau, and Jascha Sohl-Dickstein. Unrolled generative adversarial networks. arXiv preprint arXiv:1611.02163, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.759, + 0.894, + 0.829 + ], + "angle": 0, + "content": "[26] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9045" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.146 + ], + "angle": 0, + "content": "[28] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.472, + 0.217 + ], + "angle": 0, + "content": "[29] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.472, + 0.273 + ], + "angle": 0, + "content": "[30] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.472, + 0.315 + ], + "angle": 0, + "content": "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.472, + 0.383 + ], + "angle": 0, + "content": "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.472, + 0.439 + ], + "angle": 0, + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.472, + 0.496 + ], + "angle": 0, + "content": "[34] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.472, + 0.594 + ], + "angle": 0, + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.472, + 0.649 + ], + "angle": 0, + "content": "[36] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.472, + 0.705 + ], + "angle": 0, + "content": "[37] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.472, + 0.774 + ], + "angle": 0, + "content": "[38] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Jaehoon Ko, Hyeonsu Kim, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.472, + 0.831 + ], + "angle": 0, + "content": "[39] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Dynamic portrait editing by learning 4d gan from 2d diffusion-based editor. arXiv preprint arXiv:2305.20082, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[40] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.134 + ], + "angle": 0, + "content": "[41] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[42] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.894, + 0.274 + ], + "angle": 0, + "content": "[43] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.276, + 0.894, + 0.343 + ], + "angle": 0, + "content": "[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.346, + 0.894, + 0.387 + ], + "angle": 0, + "content": "[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.893, + 0.432 + ], + "angle": 0, + "content": "[46] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.893, + 0.488 + ], + "angle": 0, + "content": "[47] Yang Song, Sahaj Garg, Jiaxin Shi, and Stefano Ermon. Sliced score matching: A scalable approach to density and score estimation. In Uncertainty in Artificial Intelligence, pages 574-584. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.894, + 0.545 + ], + "angle": 0, + "content": "[48] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.894, + 0.602 + ], + "angle": 0, + "content": "[49] Akash Srivastava, Lazar Valkov, Chris Russell, Michael U Gutmann, and Charles Sutton. Veegan: Reducing mode collapse in gans using implicit variational learning. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.894, + 0.657 + ], + "angle": 0, + "content": "[50] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.893, + 0.686 + ], + "angle": 0, + "content": "[51] Cédric Villani et al. Optimal transport: old and new. Springer, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.894, + 0.73 + ], + "angle": 0, + "content": "[52] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.894, + 0.801 + ], + "angle": 0, + "content": "[53] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[54] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[55] Peihao Wang, Zhiwen Fan, Dejia Xu, Dilin Wang, Sreyas Mohan, Forrest Iandola, Rakesh Ranjan, Yilei Li, Qiang Liu," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "9046" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Zhangyang Wang, et al. Steindreamer: Variance reduction for text-to-3d score distillation via stein identity. arXiv preprint arXiv:2401.00604, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.203 + ], + "angle": 0, + "content": "[56] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.472, + 0.275 + ], + "angle": 0, + "content": "[57] Zhendong Wang, Yifan Jiang, Huangjie Zheng, Peihao Wang, Pengcheng He, Zhangyang Wang, Weizhu Chen, Mingyuan Zhou, et al. Patch diffusion: Faster and more data-efficient training of diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[58] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. arXiv preprint arXiv:2401.04092, 2024. 8, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.472, + 0.387 + ], + "angle": 0, + "content": "[59] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360 \\{\\neg\\} views. arXiv preprint arXiv:2211.16431, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.472, + 0.43 + ], + "angle": 0, + "content": "[60] Joseph Zhu and Peiye Zhuang. Hifa: High-fidelity text-to-3d with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9047" + } + ] +] \ No newline at end of file diff --git a/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/695a2b86-fc93-420e-a626-d23f402233d8_origin.pdf b/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/695a2b86-fc93-420e-a626-d23f402233d8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..639fbd73d7266ba2f03d353642649f016630b139 --- /dev/null +++ b/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/695a2b86-fc93-420e-a626-d23f402233d8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80c337106c2df3db0338c8e52fd083016c4d5cd8dc3a8d2e449a7d799d8ef68e +size 4331778 diff --git a/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/full.md b/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..31c73de7641c1fc1fb8f030770bf1164538cc1ac --- /dev/null +++ b/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/full.md @@ -0,0 +1,334 @@ +# Taming Mode Collapse in Score Distillation for Text-to-3D Generation + +Peihao Wang $^{1}$ , Dejia Xu $^{1}$ , Zhiwen Fan $^{1}$ , Dilin Wang $^{2}$ , Sreyas Mohan $^{2}$ , Forrest Iandola $^{2}$ , Rakesh Ranjan $^{2}$ , Yilei Li $^{2}$ , Qiang Liu $^{1}$ , Zhangyang Wang $^{1}$ , Vikas Chandra $^{2}$ $^{1}$ The University of Texas at Austin, $^{2}$ Meta Reality Labs + +{peihaowang,iejia, zhiwenfan, atlaswang}@utexas.edu, lqiang@cs.utexas.edu +{wdilin, sreyasmohan, fni, rakeshr, yileil, vchandra}@meta.com +vita-group.github.io/3D-Mode-Collapse/ + +# Abstract + +Despite the remarkable performance of score distillation in text-to-3D generation, such techniques notoriously suffer from view inconsistency issues, also known as "Janus" artifact, where the generated objects fake each view with multiple front faces. Although empirically effective methods have approached this problem via score debiasing or prompt engineering, a more rigorous perspective to explain and tackle this problem remains elusive. In this paper, we reveal that the existing score distillation-based text-to-3D generation frameworks degenerate to maximal likelihood seeking on each view independently and thus suffer from the mode collapse problem, manifesting as the Janus artifact in practice. To tame mode collapse, we improve score distillation by re-establishing the entropy term in the corresponding variational objective, which is applied to the distribution of rendered images. Maximizing the entropy encourages diversity among different views in generated 3D assets, thereby mitigating the Janus problem. Based on this new objective, we derive a new update rule for 3D score distillation, dubbed Entropic Score Distillation (ESD). We theoretically reveal that ESD can be simplified and implemented by just adopting the classifier-free guidance trick upon variational score distillation. Although embarrassingly straightforward, our extensive experiments demonstrate that ESD can be an effective treatment for Janus artifacts in score distillation. + +# 1. Introduction + +Recent advancements in text-to-3D technology have attracted considerable attention, particularly for its pivotal role in automating high-quality 3D content. This is especially crucial in fields such as virtual reality and gaming, where 3D content forms the bedrock. While numerous techniques are available, the prevailing text-to-3D approach is based on + +score distillation [31], popularized by DreamFusion and its follow-up works [4, 19, 26, 50, 54, 56]. + +Score distillation leverages a pre-trained 2D diffusion model to sample over the 3D parameter space (i.e. Neural Radiance Fields (NeRF) [27]) such that views rendered from a random angle satisfy the statistics of the image distribution. This algorithm is implemented by backpropagating the estimated score of each view via the chain rule. Despite the notable progress achieved with score distillation-based approaches, it is widely observed that 3D content generated using score distillation suffers from the Janus problem [12], referring to the artifacts that generated 3D objects contain multiple canonical views (see Fig. 1). + +To understand this drawback of score distillation, we draw the theoretical connection between the Janus problem and mode collapse, a statistical term describing a distribution concentrating on the high-density area while losing information about the probability tail. We first uncover that the optimization of existing score distillation-based text-to-3D generation degenerates to a maximum likelihood objective, making it susceptible to model collapse. As pre-trained diffusion models are biased to frequently encountered views [12], this oversight leads all views opt to convergence toward the point with the highest likelihood, manifesting as the Janus artifact in practical applications. The main limitation of current methods is that their distillation objectives solely maximize the likelihood of each view independently, without considering the diversity between different views. + +To address the aforementioned issue, we propose a principled approach Entropic Score Distillation (ESD), which regularizes the score distillation process by entropy maximization of the rendered image distribution, thereby enhancing the diversity of views in generated 3D assets and alleviating the Janus problem. Our derived ESD update admits a simple form as a weighted combination of scores for pre-trained image distribution and rendered image distribution. Com + +![](images/3411a3abeb213c21e1da738c11e93e90d67e603c19a36d3e0a541118c29a28e5.jpg) +Our results +A cat with tiger stripes + +![](images/6631107c8c53aff27d4e5038ded3ab7ee579a8639c8e6e44361baae353a32eb7.jpg) +"Janus" results + +![](images/877e227111e2f1ba9f60321151fe52c25097464e4f4e231917a57f21bb286d2d.jpg) +A bald eagle carved out of wood + +![](images/20786928d71d09f8973962fdde74c821b03167203b9d3f153ca22eceab7b24c9.jpg) + +![](images/6756dc1ea598b650b86c8f6e4cad871f23715a945a68e3be55c8e8a445796d19.jpg) +A chimpanzee dressed like Henry VIII king of England + +![](images/b259af55d7fb3fa251ea81c9ee10ba0b7455a24070247d764489869f837707b9.jpg) +Front View +Back View +Figure 1. A Preview of Qualitative Results. We present the front and back views of objects synthesized by VSD (ProlificDreamer) on the right two columns, and four views of our generated results on the left. VSD suffers from "Janus" problem, where both front and back views contain a frontal face of the targeted object, while our method effectively mitigates this artifact. Please refer to more results in Appendix D. + +pared with Score Distillation Sampling (SDS) [31], our ESD involves the score of the rendered image distribution, serving to maximize the entropy of the rendered image distribution. Unlike Variational Score Distillation (VSD) [56], the learned score function of the rendered image distribution does not depend on the camera pose. This subtle difference has a more profound impact, as we show the score function of rendered images modeled by VSD corresponds to an objective with fixed entropy, thereby having no influence on view variety. In contrast, ESD optimizes for a Kullback-Leibler divergence with a non-constant entropy term parameterized by the 3D model, leading to an effect that encourages diversity among different views. + +In practice, we find it challenging to optimize the score of the rendered image distribution without conditioning on the camera pose. To facilitate training, we discover that the gradient from the entropy can be decomposed into a combination of scores: one depends on the camera pose, and the other independent of it, with a coefficient interacting between these two terms. Through this theoretical establishment, we are able to adopt a handy implementation of ESD by Classifier Free Guidance (CFG) trick [10] where conditional and unconditional scores are trained alternatively and mixed during inference. + +Through extensive experiments with our proposed ESD, we demonstrate its efficacy in alleviating the Janus problem + +and its significant advantages in improving 3D generation quality when compared to the baseline methods [31, 56] and other remedy techniques [2, 12]. As a side contribution, we also borrow two inception scores [36] to evaluate text-to-3D results and numerically probe model collapse in score distillation. We show these two metrics can effectively characterize the quality and diversity of views, highly matching our qualitative observations. + +# 2. Background + +# 2.1. Diffusion Models + +Diffusion models, as demonstrated by various works [11, 44, 46, 48], have shown to be highly effective in text-to-image generation. Technically, a diffusion model learns to gradually transform a normal distribution $\mathcal{N}(\mathbf{0},\mathbf{I})$ to the target distribution $p_{data}(\boldsymbol {x}|\boldsymbol {y})$ where $\pmb{y}$ denotes the text prompt embeddings. The sampling trajectory is determined by a forward process with the conditional probability $p_t(\pmb {x}_t|\pmb {x}_0) = \mathcal{N}(\pmb {x}_t|\alpha_t\pmb {x}_0,\sigma_t^2\pmb {I})$ ,where $\pmb {x}_t\in \mathbb{R}^D$ represents the sample at time $t\in [0,T]$ , and $\alpha_{t},\sigma_{t} > 0$ are time-dependent diffusion coefficients. Consequently, the distribution at time $t$ can be formulated as $p_t(\pmb {x}_t|\pmb {y}) =$ $\int p_{data}(\pmb {x}_0|\pmb {y})\mathcal{N}(\pmb {x}_t|\alpha_t\pmb {x}_0,\sigma_t^2\pmb {I})d\pmb {x}_0.$ Diffusion models generate samples through a reverse process starting from Gaussian noises, which can be described by the ODE: + +![](images/21ba407548e7b038693c647e4ca8d1f10418ff0ad9162d38d5595fcaba68d9bc.jpg) +Figure 2. Illustration of the effect of entropy regularization. Learned image distributions often exhibit a higher probability mass for objects' frontal faces. Pure maximal likelihood seeking is opt to mode collapse (Sec. 3). Adding entropy regularization can expand the support of fitted distribution $q_{t}^{\theta}(\boldsymbol{x}|\boldsymbol{y})$ with mode-covering behavior (Sec. 4). + +![](images/916da4f02dfc3f104e790dd200d7ab14c69a42883c50231afaaad73cf2b61d27.jpg) + +$\mathrm{d}\pmb{x}_t / \mathrm{d}t = -\nabla_{\pmb{x}}\log p_t(\pmb{x}_t)$ with the boundary condition $\pmb{x}_T\sim \mathcal{N}(\pmb {0},\pmb {I})$ [22, 45, 48]. Such a process requires the computation of score function $\nabla_{\pmb{x}}\log p_t(\pmb{x}_t)$ which is often obtained by fitting a time-conditioned noise estimator $\epsilon_{\phi}:\mathbb{R}^{D}\to \mathbb{R}^{D}$ using score matching loss [15, 47, 52]. + +# 2.2. Text-to-3D Score Distillation + +Score distillation based 3D asset generation requires representing 3D scenes as learnable parameters $\pmb{\theta} \in \mathbb{R}^{N}$ equipped with a differentiable renderer $g(\pmb{\theta}, \pmb{c}): \mathbb{R}^{N} \to \mathbb{R}^{D}$ that projects 3D scene $\pmb{\theta}$ into images with respect to the camera pose $\pmb{c}$ . Here $N, D$ are the dimensions of the 3D parameter space and rendered images, respectively. Neural radiance fields (NeRF) [27] are often employed as the underlying 3D representation for its capability of modeling complex scenes. + +Recent works [4, 14, 19, 26, 31, 50, 54-56] demonstrate the feasibility of using a pretrained 2D diffusion model to guide 3D object creation. Below, we elaborate on two score distillation schemes, adopted therein: Score Distillation Sampling (SDS) [31] and Variational Score Distillation (VSD) [56]. + +Score Distillation Sampling (SDS). SDS updates the 3D parameter $\theta$ as follows: + +$$ +\nabla_ {\boldsymbol {\theta}} J _ {S D S} (\boldsymbol {\theta}) = - \mathbb {E} \left[ \omega (t) \frac {\partial g (\boldsymbol {\theta} , \boldsymbol {c})}{\partial \boldsymbol {\theta}} \left(\sigma_ {t} \nabla \log p _ {t} \left(\boldsymbol {x} _ {t} \mid \boldsymbol {y}\right) - \boldsymbol {\epsilon}\right) \right], \tag {1} +$$ + +where the expectation is taken over timestep $t\sim \mathcal{U}[0,T]$ Gaussian noises $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ and camera pose $\pmb {c}\sim p_c(\pmb {c})$ Here is $\nabla \log p$ is a pre-trained diffusion model $\epsilon_{\phi}(\pmb {x},t,\pmb {y})$ and $\pmb{x}_t$ is a noisy version of the rendering given by camera pose $\pmb {c}$ $x_{t} = \alpha_{t}g(\pmb {\theta},\pmb {c}) + \sigma_{t}\pmb{\epsilon}$ .Updating $\pmb{\theta}$ as in Eq. (1) has been shown to minimize the evidence lower bound (ELBO) for the rendered images, see Wang et al. [54], Xu et al. [59]. + +Variational Score Distillation (VSD). VSD [56] is introduced in ProlificDreamer, VSD improves upon SDS by deriving the following Wasserstein gradient flow [51]: + +$$ +\begin{array}{l} \nabla_ {\boldsymbol {\theta}} J _ {V S D} (\boldsymbol {\theta}) = - \mathbb {E} \left[ \omega (t) \frac {\partial g (\boldsymbol {\theta} , \boldsymbol {c})}{\partial \boldsymbol {\theta}} \left(\sigma_ {t} \nabla \log p _ {t} (\boldsymbol {x} _ {t} | \boldsymbol {y}\right) \right. \\ \left. - \sigma_ {t} \nabla \log q _ {t} (\boldsymbol {x} _ {t} | \boldsymbol {c})\right) \Bigg ]. \tag {2} \\ \end{array} +$$ + +Similarly, $\pmb{x}_t = \alpha_t g(\pmb{\theta}, \pmb{c}) + \sigma_t \pmb{\epsilon}$ is the noisy observation of the rendered image. In contrast to SDS, VSD introduces a new score function of the noisy rendered images conditioned on the camera pose $\mathbf{c}$ . To obtain this score, Wang et al. [56] fine-tunes a diffusion model using images rendered from the 3D scene as follows: + +$$ +\min _ {\boldsymbol {\psi}} \mathbb {E} \left[ \omega (t) \| \boldsymbol {\epsilon} _ {\boldsymbol {\psi}} \left(\alpha_ {t} g (\boldsymbol {\theta}, \boldsymbol {c}) + \sigma_ {t} \boldsymbol {\epsilon}, t, \boldsymbol {c}, \boldsymbol {y}\right) - \boldsymbol {\epsilon} \| _ {2} ^ {2} \right], \tag {3} +$$ + +where $\epsilon_{\psi}(\boldsymbol{x},t,\boldsymbol{c},\boldsymbol{y})$ is the noise estimator of $\nabla \log q_t(\boldsymbol{x}_t|\boldsymbol{c})$ as in diffusion models. As proposed in ProlificDreamer, $\psi$ is parameterized by LoRA [13] and initialized from a pre-trained diffusion model same as $\nabla \log p_t$ . + +# 3. Revealing Mode Collapse in Score Distillation + +Despite the remarkable performance of SDS and VSD in 3D asset generation, it is widely observed that the synthesized objects suffer from "Janus" artifacts. Janus artifacts refer to the generated 3D scene containing multiple canonical views (the most representative perspective of the object such as the frontal face). In earlier works, Hong et al. [12] and Huang et al. [14] attribute this problem to unimodality of the learned 2D image distribution since the training data for the diffusion models are naturally biased to the most commonly seen views per each category. In this section, we examine extant distillation schemes from a statistical view, which has been overlooked in previous literature. + +In principle, natural 2D images can be seen as random projections of 3D scenes. Score distillation matches the image distribution generated by randomly sampled views with a + +text-conditioned image distribution to recover the underlying 3D representation. Hence, Janus artifact, in which each view becomes uniform and identical to the most commonly seen views, can be interpreted as a manifestation of distribution collapse to samples within the high-density region. Such distribution degeneration essentially corresponds to the statistical phenomenon mode collapse, which happens when an optimized distribution fails to characterize the data diversity and concentrates on a single type of output [1, 7, 25, 36, 49]. + +Below we theoretically reveal why SDS and VSD are prone to mode collapse. As shown in Poole et al. [31], Wang et al. [56], SDS and VSD equals to the gradient of the following Kullback-Leibler (KL) divergence, i.e., $J_{SDS}(\pmb{\theta}) = J_{VSD}(\pmb{\theta}) = J_{KL}(\pmb{\theta})$ up to an additive constant: + +$$ +J _ {K L} (\boldsymbol {\theta}) = \mathbb {E} \left[ \Omega (t) \mathcal {D} _ {\mathrm {K L}} \left(q _ {t} ^ {\boldsymbol {\theta}} \left(\boldsymbol {x} _ {t} | \boldsymbol {c}, \boldsymbol {y}\right) \| p _ {t} \left(\boldsymbol {x} _ {t} | \boldsymbol {y}\right)\right) \right], \tag {4} +$$ + +where $\Omega(t) = \omega(t)\sigma_t / \alpha_t$ and the expectation is taken over $t \sim \mathcal{U}[0,T]$ and $\pmb{c} \sim p_c(\pmb{c})$ . Here $p_t(\pmb{x}_t|\pmb{y}) = \int p_0(\pmb{x}_0|\pmb{y})\mathcal{N}(\pmb{x}_t|\alpha_t\pmb{x}_0,\sigma_t^2\pmb{I})d\pmb{x}_0$ is the image distribution perturbed by Gaussian noises, while $q_t^\theta (\pmb{x}_t|\pmb{c},\pmb{y}) = \int q_0^\theta (\pmb{x}_0|\pmb{c})\mathcal{N}(\pmb{x}_t|\alpha_t\pmb{x}_0,\sigma_t^2\pmb{I})d\pmb{x}_0$ models the image distribution generated by 3D parameter $\pmb{\theta}$ with respect to camera pose $\pmb{c}$ and diffused by Gaussian distribution. As shown by Wang et al. [56], $J_{KL}(\pmb{\theta}) = 0$ implies $q_0^\theta (\pmb{x}_0|\pmb{c}) = p(\pmb{x}_0|\pmb{y})$ , i.e., the distribution of synthesized views satisfy the text-conditioned image distribution. + +However, it has not escaped from our notice that $q_0^\theta(\boldsymbol{x}_0|\boldsymbol{c}) = \delta(\boldsymbol{x}_0 - g(\boldsymbol{\theta},\boldsymbol{c}))$ is a Dirac distribution for both SDS and VSD. This causes the original KL divergence minimization (Eq. 4) degenerate to a Maximal Likelihood Estimation (MLE) problem: + +$$ +\begin{array}{l} J _ {K L} (\boldsymbol {\theta}) = \underbrace {- \mathbb {E} \left[ \Omega (t) \mathbb {E} _ {\boldsymbol {x} _ {t} \sim q _ {t} ^ {\boldsymbol {\theta}} (\boldsymbol {x} _ {t} | \boldsymbol {c} , \boldsymbol {y})} \log p _ {t} (\boldsymbol {x} _ {t} | \boldsymbol {y}) \right]} _ {J _ {M L E} (\boldsymbol {\theta})} \tag {5} \\ - \underbrace {\mathbb {E} \left[ \Omega (t) H \left[ q _ {t} ^ {\boldsymbol {\theta}} (\boldsymbol {x} _ {t} | \boldsymbol {c} , \boldsymbol {y}) \right] \right]} _ {\text {c o n s t .}}, \\ \end{array} +$$ + +where $H[q_t^\theta(\boldsymbol{x}_t|\boldsymbol{y})] = -\mathbb{E}_{\boldsymbol{x}_t \sim q_t^\theta(\boldsymbol{x}_t|\boldsymbol{c}, \boldsymbol{y})}[\log q_t^\theta(\boldsymbol{x}_t|\boldsymbol{c}, \boldsymbol{y})]$ denotes the entropy of $q_t^\theta(\boldsymbol{x}_t|\boldsymbol{y})$ , which turns out to be a constant because $q_t^\theta(\boldsymbol{x}_t|\boldsymbol{c}, \boldsymbol{y}) = \mathcal{N}(\boldsymbol{x}_t | \alpha_t g(\boldsymbol{\theta}, \boldsymbol{c}), \sigma_t^2 \boldsymbol{I})$ which has fixed entropy once $t, \boldsymbol{\theta}$ and $\boldsymbol{c}$ have been specified. See full derivation in Appendix A.1. + +Note that Eq. 5 signifies $J_{KL}(\pmb{\theta}) = J_{MLE}(\pmb{\theta})$ up to an additive constant, hence $J_{KL}(\pmb{\theta})$ shares all minima with $J_{MLE}(\pmb{\theta})$ . It is known that likelihood maximization is more prone to mode collapse. Intuitively, minimizing $J_{MLE}(\pmb{\theta})$ seeks each view independently to have the maximum log-likelihood on the image distribution $p(\pmb{x}_0|\pmb{y})$ . Since $p(\pmb{x}_0|\pmb{y})$ is usually unimodal and peaks at the canonical view, each view of the scene will collapse to the same local minimum, resulting in Janus artifact (see Fig. 2). We postulate that the existing distillation strategies may be inherently limited + +![](images/a33b82e50ad9746c4eb02291f66fe6c94ddbb4dd5361afb564e95a855af58acb.jpg) +Figure 3. Gaussian Example. To illustrate the effects of entropy regularization, we leverage SDS, VSD and ESD to fit a 2D Gaussian distribution. The blue points are sampled from the ground-truth distribution while the orange points are from the fitted distribution. + +by their log-likelihood seeking behaviors, which are more susceptible to mode collapse, especially with biased image distributions. + +# 4. Entropy Regularized Score Distillation + +# 4.1. Entropic Score Distillation + +In this section, we highlight the importance of the entropy in score distillation. It is known that higher entropy implies the corresponding distribution could cover a larger support of the ambient space and thus increase the sample diversity. In Eq. 5, the entropy term is shown to diminish in the training objective, which causes each generated view to lack diversity and collapse to a single image with the highest likelihood. + +To this end, we propose to bring in an entropy regularization to $J_{MLE}(\pmb{\theta})$ for boosting the view diversity. Since $q_t^\pmb{\theta}(\pmb{x}_t|\pmb{c},\pmb{y})$ has constant entropy, we regularize entropy for the distribution $q_t^\pmb{\theta}(\pmb{x}_t|\pmb{y}) = \int q_t^\pmb{\theta}(\pmb{x}_t|\pmb{c},\pmb{y})p_c(\pmb{c})d\pmb{c}$ , which can be simulated by randomly sampling views from the 3D parameter $\pmb{\theta}$ . Consider the following objective: + +$$ +\begin{array}{l} J _ {E n t} (\boldsymbol {\theta}, \lambda) = - \mathbb {E} \left[ \Omega (t) \mathbb {E} _ {\boldsymbol {x} _ {t} \sim q _ {t} ^ {\boldsymbol {\theta}} (\boldsymbol {x} _ {t} | \boldsymbol {c}, \boldsymbol {y})} \log p _ {t} (\boldsymbol {x} _ {t} | \boldsymbol {y}) \right] \tag {7} \\ - \lambda \mathbb {E} \left[ \Omega (t) H \left[ q _ {t} ^ {\boldsymbol {\theta}} (\boldsymbol {x} _ {t} | \boldsymbol {y}) \right] \right], \\ \end{array} +$$ + +where $\lambda$ is a hyper-parameter controlling the regularization strength. We note that without $H[q_t^\theta (\pmb {x}_t|\pmb {y})]$ , each view is optimized independently and implicitly regularized by the underlying parameterization. However, upon imposing $H[q_t^\theta (\pmb {x}_t|\pmb {y})]$ , all views become explicitly correlated with each other, as they collectively contribute to the entropy computation. Intuitively, $J_{Ent}(\pmb {\theta},\lambda) = J_{MLE}(\pmb {\theta})-$ $\lambda \mathbb{E}[\Omega (t)H[q_t^\theta (\pmb {x}_t|\pmb {y})]]$ seeks the maximal log-likelihood for each view while simultaneously enlarging the entropy for distribution $q_{t}^{\theta}(\pmb{x}_{t}|\pmb{y})$ , which spans the support and encourages diversity across the rendered views. To gain more insights, we present the following theoretical results: + +Theorem 1. For any $\lambda \in \mathbb{R}$ and $\pmb{\theta} \in \mathbb{R}^{D}$ , we have $J_{Ent}(\pmb{\theta},\lambda) = \lambda \mathbb{E}_t[\Omega(t)\mathcal{D}_{\mathrm{KL}}(q_t^\pmb{\theta}(\pmb{x}_t|\pmb{y})||p_t(\pmb{x}_t|\pmb{y}))] + (1 - \lambda)\mathbb{E}_{t,\pmb{c}}[\Omega(t)\mathcal{D}_{\mathrm{KL}}(q_t^\pmb{\theta}(\pmb{x}_t|\pmb{c},\pmb{y})||p_t(\pmb{x}_t|\pmb{y}))] + const.$ + +Algorithm 1 ESD: Entropic score distillation for text-to-3D generation +Input: A diffusion model $\epsilon_{\phi}(\pmb {x},t,\pmb {y})$ ; learnable 3D parameter $\theta$ ; coefficient $\lambda$ ; text prompt $\pmb{y}$ +Initialize $\psi$ for another diffusion model $\epsilon_{\psi}(\pmb {x},t,\pmb {y})$ with the parameter $\phi$ specified in diffusion model $\epsilon_{\phi}(\pmb {x},t,\pmb {y})$ , parame +terized with LoRA. +while not converged do +Randomly sample a camera pose $c\sim p_c$ and render a view $\pmb {x}_0 = g(\pmb {\theta},\pmb {c})$ from $\pmb{\theta}$ Sample a $t\sim \mathcal{U}[0,T]$ and add Gaussian noise $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ .. $\pmb{x}_{t} = \alpha_{t}\pmb{x}_{0} + \sigma_{t}\pmb{\epsilon}$ $\pmb {\theta}\gets \pmb {\theta} + \eta_1\left[\omega (t)\frac{\partial g(\pmb{\theta},\pmb{c})}{\partial\pmb{\theta}} (\epsilon_{\phi}(\pmb{x}_{t},t,\pmb {y}) - \lambda \epsilon_{\psi}(\pmb{x}_{t},t,\emptyset ,\pmb {y}) - (1 - \lambda)\epsilon_{\psi}(\pmb{x}_{t},t,\pmb {c},\pmb {y})\right]$ (6) With probability $1 - p_{\emptyset}$ $\psi \gets \psi -\eta_2\nabla_{\psi}\left[\omega (t)\| \epsilon_{\psi}(\pmb {x}_t,t,\pmb {c},\pmb {y}) - \epsilon \| _2^2\right]$ Otherwise, $\psi \gets \psi -\eta_2\nabla_{\psi}\left[\omega (t)\| \epsilon_{\psi}(\pmb {x}_t,t,\emptyset ,\pmb {y}) - \epsilon \| _2^2\right]$ +end while +Return $\theta$ + +We prove Theorem 1 in Appendix A.3. Theorem 1 implies that $J_{Ent}(\theta, \lambda)$ essentially equal to a combination of two types of KL divergences, where the former one minimizes the distribution discrepancy between $q_t^\theta(\boldsymbol{x}_t|\boldsymbol{y})$ and $p_t^\theta(\boldsymbol{x}_t|\boldsymbol{y})$ which marginalizes the camera pose within $q_t^\theta$ , while the latter is the original KL divergence $J_{KL}(\theta)$ adopted by SDS and VSD which takes expectation over $\boldsymbol{c}$ out of KL divergence. + +Next, we derive the gradient of $J_{Ent}(\theta, \lambda)$ that will be backpropagated to update the 3D representation. It can be obtained by path derivative and reparameterization trick: + +$$ +\begin{array}{l} \nabla_ {\boldsymbol {\theta}} J _ {E n t} (\boldsymbol {\theta}, \lambda) = - \mathbb {E} \left[ \omega (t) \frac {\partial g (\boldsymbol {\theta} , \boldsymbol {c})}{\partial \boldsymbol {\theta}} \left(\sigma_ {t} \nabla \log p _ {t} \left(\boldsymbol {x} _ {t} \mid \boldsymbol {y}\right) \right. \right. \tag {8} \\ \left. - \lambda \sigma_ {t} \nabla \log q _ {t} ^ {\boldsymbol {\theta}} (\boldsymbol {x} _ {t} | \boldsymbol {y}) \right]. \\ \end{array} +$$ + +The full derivation is deferred to Appendix A.2. We name this update rule as Entropic Score Distillation (ESD). Note that ESD differs from VSD as its second score function does not depend on the camera pose. + +# 4.2. Classifier-Free Guidance Trick + +Similar to SDS and VSD, we approximate $\nabla \log p_t(\pmb{x}_t|\pmb{y})$ via a pre-trained diffusion model $\epsilon_{\phi}(\pmb{x}_t,t,\pmb{y})$ . However, $\nabla \log q_t^\theta (\pmb{x}|\pmb{y})$ is not readily available. We found that directly fine-tuning a pre-trained diffusion model using rendered images to approximate $\nabla \log q_t^\theta (\pmb {x}|\pmb {y})$ , akin to Prolific-Dreamer, does not yield robust performance. We postulate this difficulty arises from the removal of the camera condition, increasing the complexity of the distribution to be fitted. + +To tackle this problem, we recall the result in Theorem 1 that $J_{Ent}(\theta, \lambda)$ can be written in terms of two KL divergence losses. Therefore, its gradient can be decomposed as a weighted combination of their gradients, which correspond to unconditional and conditional score functions in terms of + +the camera pose $c$ , respectively: + +$$ +\begin{array}{l} \nabla_ {\boldsymbol {\theta}} J _ {E n t} (\boldsymbol {\theta}, \lambda) = - \mathbb {E} \left[ \omega (t) \frac {\partial g (\boldsymbol {\theta} , \boldsymbol {c})}{\partial \boldsymbol {\theta}} \left(\sigma_ {t} \nabla \log p _ {t} \left(\boldsymbol {x} _ {t} \mid \boldsymbol {y}\right) \right. \right. \tag {9} \\ \left. \left. - \lambda \sigma_ {t} \nabla \log q _ {t} ^ {\boldsymbol {\theta}} (\boldsymbol {x} _ {t} | \boldsymbol {y}) - (1 - \lambda) \sigma_ {t} \nabla \log q _ {t} ^ {\boldsymbol {\theta}} (\boldsymbol {x} _ {t} | \boldsymbol {c}, \boldsymbol {y})\right) \right]. \\ \end{array} +$$ + +We formally prove Eq. 9 in Appendix A.3. With the above formulation, ESD can be implemented via the Classifier-Free Guidance (CFG) trick, which was initially proposed to balance the variety and quality of text-conditionally generated images from diffusion models [10]. Algorithm 1 outlines the computation paradigm of ESD, in which we surrogate score functions in Eq. 9 with pre-trained and fine-tuned diffusion models (see Eq. 6), and takes random turns with a probability $p_{\emptyset}$ to balance the training of conditional and unconditional score functions, as suggested by Ho and Salimans [10]. + +# 4.3. Discussion + +In VSD, the camera-conditioned score is believed to play a significant role in facilitating visual quality. Intuitively, such conditioning can equip the tuned diffusion model with multi-view priors [20]. Also, Hertz et al. [8] suggests such a method can be useful to stabilize the update of the implicit parameters. However, ESD counters this argument by suggesting that the camera condition might not always be advantageous, particularly when the particle size is reduced to one. In such cases, the resulting KL divergence provably degenerates to a likelihood maximization algorithm vulnerable to mode collapse. + +It is noteworthy that, even though their subtle differences in implementation, the optimization objectives of ESD and VSD are fundamentally different (see Sec. 4.1). ESD sets itself apart from VSD by incorporating entropy regularization, a crucial feature absent in VSD, aiming to augment diversity across views. Despite originating from distinct objectives, our theoretical establishment allows for a straightforward implementation of ESD based on VSD using the CFG trick. + +![](images/cd94397d6e389ce5637fe38b0607cdd5ba3aa76d321d0907cbd75e123487ad50.jpg) +Figure 4. Qualitative Results. Our proposed outperforms all baselines in terms of better geometry and well-constructed texture details. Our results deliver photo-realistic and diverse rendered views, while baseline methods more or less suffer from the Janus problem. Best view in an electronic copy. + +We provide an illustrative example by leveraging SDS, VSD and ESD (with different $\lambda$ 's) to fit a 2D Gaussian distribution in Fig. 3. With SDS and VSD, all samples are converged to the high-density area while ESD recovers the entire support of the distribution. We provide more details and examples in Appendix B. + +# 5. Other Related Work + +Text-to-Image Diffusion Model. Text-to-image diffusion models [32, 33] are cornerstone components of text-to-3D generation. It involves text embedding conditioning into the iterative denoising process. Equipped with large-scale image-text paired datasets, many works [29, 33, 35] scale up to tackle text-to-image generation. Among them, latent diffusion models attracted great interest in the open-source community since they reduced the computation cost by diffusing in the low-resolution latent space instead of directly in the pixel space. In addition, text-to-image diffusion models have also found applications in various computer vision tasks, including text-to-3D [31, 43], image-to-3D [59], text-to-svq [17], text-to-video [18, 42], etc. + +3D Generation with 2D Priors. Well-annotated 3D data requires immense effort to collect. Instead, a line of research studies on how to learn 3D generative models using 2D su + +pervision. Early attempts, including pi-GAN [34], EG3D [3], GRAF [37], GIRAFFE [30], adopt adversarial loss between the rendered images and natural images. DreamField [16] leverages CLIP to align NeRF with text prompts. More recently, with the rapid development of text-to-image diffusion models, diffusion-based image priors have attracted increasing interest, and score distillation has then become the dominant technique. Pioneer works DreamFusion [31] and ProlificDreamer [56] have been introduced in detail in Sec. 2. Their concurrent work SJC [54] derives the score Jacobian chaining method from another theoretical viewpoint of Perturb and Average Scoring. Even though diffusion models directly trained with 3D data nowadays demonstrate largely improved results [21, 41], score distillation still plays a pivotal role in ensuring view consistency. + +Techniques to Improve Score Distillation. Providing the empirical promise of score distillation, there have been numerous techniques proposed to improve its effectiveness. Magic3D [19] and Fantasia3D [4] utilize mesh and DMTet [40] to disentangle the optimization of geometry and texture. TextMesh [50] and 3DFuse [38] use depth-conditioned text-to-image diffusion priors that support geometry-aware texturing. Score debiasing[12] and Perp-Neg [2] study to refine the text prompts for a better 3D generation. DreamTime [14] and RED-Diff [24] investigate the timestep scheduling in the + +![](images/d398962526ba43d25c7ee07cc30b8700a417c90f1923353349837d2ec2520567.jpg) +Figure 5. Qualitative Results. We combine our proposed ESD with timestep scheduling in DreamTime [14] and compare it against baseline methods. Prompt: A caramic lion. + +score distillation process. HIFA [60] adopts multiple diffusion steps for distillation. Score distillation also works with auxiliary losses, including CLIP loss [59] and adversarial loss [5, 39]. + +# 6. Evaluation Metrics + +In this section, we introduce four information-theoretic metrics to numerically evaluate the generated 3D results with a particular focus on identifying Janus artifacts or mode collapse. The metrics we propose comprehensively cover four aspects: 1) the relevance with the text prompts, 2) distribution fitness, 3) rendering quality, and 4) view diversity. + +CLIP Distance. We compute the average distance between rendered images and the text embedding to reflect the relevance between generated results and the specified text prompt. Specifically, we render $N$ views from the generated 3D representations, and for each view, we obtain an embedding vector through the image encoder of a CLIP model [53]. In the meantime, we compute the text embedding utilizing the text encoder. The CLIP distance is computed as the one minus cosine similarity between the image embeddings and text embeddings averaged over all views. + +Frechet inception distance (FID). As shown in Sec. 3 and 4, score distillation essentially matches distributions via KL divergence. Hence, it becomes reasonable to employ FID to measure the distance between the image distribution $q^{\theta}(\pmb{x}_0|\pmb{y})$ generated by randomly rendering 3D representation and the text-conditioned image distribution $p(\pmb{x}_0|\pmb{y})$ modeled by a diffusion model. We sample $N$ images using pre-trained latent diffusion model given text prompts as the ground truth image dataset, and render $N$ views uniformly distributed over a unit sphere from the optimized 3D scene as the generated image dataset. Then standard FID [9] is computed between these two sets of images. Note that FID is known to be effective in quantitatively identifying mode collapse. + +Inception Quality and Variety. Thanks to our established connection with mode collapse, we know that Janus problem is due to a lack of sample diversity. Inspired by Inception Score (IS) [36], we utilize entropy-related metrics to reflect the generated image quality and diversity. We propose Inception Quality (IQ) and Inception Variety (IV), formulated as below: + +$$ +I Q (\boldsymbol {\theta}) = \mathbb {E} _ {\boldsymbol {c}} \left[ H \left[ p _ {c l s} (\boldsymbol {y} | g (\boldsymbol {\theta}, \boldsymbol {c})) \right] \right], \tag {10} +$$ + +$$ +I V (\boldsymbol {\theta}) = H \left[ \mathbb {E} _ {\boldsymbol {c}} \left[ p _ {c l s} (\boldsymbol {y} | g (\boldsymbol {\theta}, \boldsymbol {c}) \right] \right], \tag {11} +$$ + +where $p_{cls}(\boldsymbol{y}|\boldsymbol{x})$ is a pre-trained classifier. IQ computes the average entropy of the label logits predicted for all rendered views, while IV computes the entropy of the averaged label logits of all rendered views. Intuitively, the smaller IQ means highly confident classification results on rendered views, which also indicates better visual quality of generated 3D assets. In the meanwhile, the higher IV signifies that each rendered view is likely to have a distinct label prediction, meaning the 3D creation has higher view diversity. Note that IV upper bounds IQ due to Jensen inequality. So we can define Inception Gain $IG = (IV - IQ) / IQ$ , which characterizes the information gain brought by knowing where the camera pose is, namely the improvement of distinguishability among different views. + +# 7. Experiments + +Settings. In this section, we empirically validate the effectiveness of our proposal. The chosen prompts are targeted at objects with clearly defined canonical views, posing a challenge for existing methods. Our baseline approaches include SDS (DreamFusion) [31] and VSD (ProlificDreamer) [56], as well as two methods dedicated to solving Janus problem: Debiased-SDS [12] and Perp-Neg [2]. For fair comparison, all experiments are benchmarked under the open-source threestudio framework. Geometry refinement [56] is adopted for all distillation schemes. Please refer to Appendix C for more implementation details. + +![](images/818f7bf7fa32591a0ca0eb1e8e55e6964ac177856f04e013e3e4e1a7b9560901.jpg) +$\lambda = 1.0$ + +![](images/67d990746d35d70bff52a1f66c12fcbada67207c870ca4b0d6c2823254550fa8.jpg) +$\lambda = 0.5$ +Figure 6. Ablation Studies on $\lambda$ . We investigate the choice of different entropy regularization strength $\lambda$ . Prompt: Michelangelo-style statue of dog reading news on a cellphone. + +![](images/c8f6ac5623ee5032e1fbb441d7bd99d83d49dd39c2c2f303233787d1eeb342f2.jpg) +$\lambda = 0.0$ + +Table 1. Quantitative Comparisons. $(\downarrow)$ means the lower the better, and $(\uparrow)$ means the higher the better. + +
CLIP (↓)FID (↓)IQ (↓)IV (↑)IG (↑)SR (↑)
SDS0.737291.8604.2954.85520.12315.00%
VSD0.725265.1413.1493.57120.13719.17%
ESD0.714235.9153.1354.03140.32755.83%
+ +Qualitative Comparison. We present qualitative comparisons in Fig. 4. We encourage interested readers to Appendix D for more results and our project page for videos. It is clearly shown that our proposed ESD delivers more precise geometry with the Janus problem alleviated. In comparison, the results presented by SDS and VSD all contain more or less corrupted geometry with multi-face structures. Debiased-SDS and Perp-Neg are shown to be effective for some text prompts, while not so consistent as ESD. Additionally, we find that ESD can work particularly well when combined with the time-prioritized scheduling proposed in DreamTime [14], as shown in Fig. 5. This means ESD is orthogonal to many other methods and can cooperate with them to further reduce Janus artifacts. + +Quantitative Comparison. With metrics proposed in Sec. 6, we numerically evaluate our method and baselines across 120 text prompts provided in [58]. We additionally involve Successful generation Rate (SR) based on human evaluation. The results are presented in Tab. 1. We observe that among all metrics, ESD reaches the best CLIP score, FID, and IG. More importantly, ESD achieves the optimal balance between view quality and diversity as shown by IQ and IV. Whereas, SDS suffers from low image quality with high IQ and VSD is limited by insufficient view variety with low IV. The superior IG of ESD indicates that views inside the generated scene are distinguishable rather than collapsing to be the same. We defer the breakdown table for numerical evaluation on examples in Fig. 4, human evaluation criteria, and the standard deviation of metrics to Appendix E. + +Ablation Studies We conduct ablation studies on the choice of $\lambda$ (i.e. CFG weights) in Fig. 6. We demonstrate that $\lambda$ can adjust ESD's preference toward view- quality or diversity. When set to one, the produced Janus-free result albeit with fewer realistic details in the textures. Conversely, + +![](images/c1c5500abdd311042b5fd6bbeed3589c5562d03e2c38ef76c80578d5a2ae186c.jpg) +ESD w/ CFG +Step $= 25\mathrm{k}$ +Figure 7. Ablation on Implementations. The successfully generated result is obtained via our suggested CFG trick while the diverged result is yielded by fitting the unconditioned score function in Eq. 8 via LoRA. Prompt: an elephant skull. + +![](images/fc7f3df07ccbf19d99e77b818a19c42c2ccb27970da651e31af9b96f5c7b26ce.jpg) +ESD w/ fine-tuned unconditional LoRA +Step $= 18\mathrm{k}$ +Step $= 20\mathrm{k}$ +Step $= 22\mathrm{k}$ + +![](images/9226588b631f302947307531af3bccd832931adc667f129e3907da71c64f7363.jpg) +Step $= 25\mathrm{k}$ + +when set to zero, ESD equates to VSD, and the Janus problem emerges again. We empirically find that choosing $\lambda$ around 0.5 yields the best result, balancing fine-grained textures and well-constructed geometry. We also implement ESD by directly fitting the score function $\nabla \log q_t^\theta (\pmb{x}_t|\pmb{y})$ without camera pose conditioning to validate the suggested implementation by CFG trick. We show in Fig. 7 that this optimization scheme is unstable. As training proceeds, the gradient explodes, and the optimized texture overflows. + +# 8. Conclusion + +In this paper, we reveal that existing score distillation methods degenerate to maximal likelihood seeking on each view independently, leading to the mode collapse problem. We identify that re-establishing the entropy term in the variational objective brings a new update rule, called Entropic Score Distillation (ESD), which is theoretically equivalent to adopting classifier-free guidance trick upon variational score distillation. ESD maximizes the entropy of the rendered image distribution, encouraging diversity across views and mitigating the Janus problem. + +# Acknowledgments + +P Wang is sincerely grateful for constructive feedback regarding this manuscript from Zhaoyang Lv, Xiaoyu Xiang, Amit Kumar, Jinhui Xiong, and Varun Nagaraja. P Wang also thanks Ruisi Cai for providing decent visual materials for illustration purposes. Any statements, opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of their employers or the supporting entities. + +# References + +[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017. 4 +[2] Mohammadreza Armandpour, Huangjie Zheng, Ali Sadeghian, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968, 2023. 2, 6, 7, 8 +[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 6 +[4] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 1, 3, 6 +[5] Yiwen Chen, Chi Zhang, Xiaofeng Yang, Zhongang Cai, Gang Yu, Lei Yang, and Guosheng Lin. It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473, 2023. 7 +[6] Giannis Daras, Kulin Shah, Yuval Dagan, Aravind Gollakota, Alex Dimakis, and Adam Klivans. Ambient diffusion: Learning clean distributions from corrupted data. Advances in Neural Information Processing Systems, 36, 2024. 6 +[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 4 +[8] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. arXiv preprint arXiv:2304.07090, 2023. 5 +[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7 +[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 5 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[12] Susung Hong, Donghoon Ahn, and Seungryong Kim. Debi-asing scores and prompts of 2d diffusion for robust text-to-3d generation. arXiv preprint arXiv:2303.15413, 2023. 1, 2, 3, 6, 7, 8 +[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3 +[14] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023. 3, 6, 7, 8 + +[15] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005. 3 +[16] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 6 +[17] Ajay Jain, Amber Xie, and Pieter Abbeel. Vectorfusion: Text-to-svq by abstracting pixel-based diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1911-1920, 2023. 6 +[18] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 6 +[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 1, 3, 6 +[20] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023.5 +[21] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 6 +[22] Ziming Liu, Di Luo, Yilun Xu, Tommi Jaakkola, and Max Tegmark. Genphys: From physical processes to generative models. arXiv preprint arXiv:2304.02637, 2023. 3 +[23] Jonathan Lorraine, Kevin Xie, Xiaohui Zeng, Chen-Hsuan Lin, Towaki Takikawa, Nicholas Sharp, Tsung-Yi Lin, Ming-Yu Liu, Sanja Fidler, and James Lucas. Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349, 2023.8 +[24] Morteza Mardani, Jiaming Song, Jan Kautz, and Arash Vahdat. A variational perspective on solving inverse problems with diffusion models. arXiv preprint arXiv:2305.04391, 2023.6 +[25] Luke Metz, Ben Poole, David Pfau, and Jascha Sohl-Dickstein. Unrolled generative adversarial networks. arXiv preprint arXiv:1611.02163, 2016. 4 +[26] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 1, 3 +[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 3 + +[28] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022.6 +[29] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 6 +[30] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 6 +[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 3, 4, 6, 7, 8 +[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 6 +[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 6 +[34] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 6 +[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems, 2022. 6 +[36] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 2, 4, 7, 8 +[37] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 6 +[38] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Jaehoon Ko, Hyeonsu Kim, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937, 2023. 6 +[39] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Dynamic portrait editing by learning 4d gan from 2d diffusion-based editor. arXiv preprint arXiv:2305.20082, 2023. 7 +[40] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 6 + +[41] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 6 +[42] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.6 +[43] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023.6 +[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2 +[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3 +[46] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2 +[47] Yang Song, Sahaj Garg, Jiaxin Shi, and Stefano Ermon. Sliced score matching: A scalable approach to density and score estimation. In Uncertainty in Artificial Intelligence, pages 574-584. PMLR, 2020. 3 +[48] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 3 +[49] Akash Srivastava, Lazar Valkov, Chris Russell, Michael U Gutmann, and Charles Sutton. Veegan: Reducing mode collapse in gans using implicit variational learning. Advances in neural information processing systems, 30, 2017. 4 +[50] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439, 2023. 1, 3, 6 +[51] Cédric Villani et al. Optimal transport: old and new. Springer, 2009. 3 +[52] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 3 +[53] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 7 +[54] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 1, 3, 6 +[55] Peihao Wang, Zhiwen Fan, Dejia Xu, Dilin Wang, Sreyas Mohan, Forrest Iandola, Rakesh Ranjan, Yilei Li, Qiang Liu, + +Zhangyang Wang, et al. Steindreamer: Variance reduction for text-to-3d score distillation via stein identity. arXiv preprint arXiv:2401.00604, 2023. +[56] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 2, 3, 4, 6, 7, 8 +[57] Zhendong Wang, Yifan Jiang, Huangjie Zheng, Peihao Wang, Pengcheng He, Zhangyang Wang, Weizhu Chen, Mingyuan Zhou, et al. Patch diffusion: Faster and more data-efficient training of diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 6 +[58] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. arXiv preprint arXiv:2401.04092, 2024. 8, 7 +[59] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360 \{\neg\} views. arXiv preprint arXiv:2211.16431, 2022. 3, 6, 7 +[60] Joseph Zhu and Peiye Zhuang. Hifa: High-fidelity text-to-3d with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 7 \ No newline at end of file diff --git a/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/images.zip b/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5a408eda1610886935456df4007de27cf1ab49c2 --- /dev/null +++ b/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f44ac3d76cbaafe3f6224c6590114038a7356c22eb65dc9f97e9d57408ca0717 +size 492320 diff --git a/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/layout.json b/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b5eed0edad0b1161dcd81527c7437332014de425 --- /dev/null +++ b/2024/Taming Mode Collapse in Score Distillation for Text-to-3D Generation/layout.json @@ -0,0 +1,9987 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 81, + 103, + 514, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 103, + 514, + 120 + ], + "spans": [ + { + "bbox": [ + 81, + 103, + 514, + 120 + ], + "type": "text", + "content": "Taming Mode Collapse in Score Distillation for Text-to-3D Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "spans": [ + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "text", + "content": "Peihao Wang" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "text", + "content": ", Dejia Xu" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "text", + "content": ", Zhiwen Fan" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "text", + "content": ", Dilin Wang" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "text", + "content": ", Sreyas Mohan" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "text", + "content": ", Forrest Iandola" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "text", + "content": ", Rakesh Ranjan" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "text", + "content": ", Yilei Li" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "text", + "content": ", Qiang Liu" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "text", + "content": ", Zhangyang Wang" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "text", + "content": ", Vikas Chandra" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "text", + "content": "The University of Texas at Austin, " + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 142, + 507, + 185 + ], + "type": "text", + "content": "Meta Reality Labs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 96, + 187, + 496, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 187, + 496, + 227 + ], + "spans": [ + { + "bbox": [ + 96, + 187, + 496, + 227 + ], + "type": "text", + "content": "{peihaowang,iejia, zhiwenfan, atlaswang}@utexas.edu, lqiang@cs.utexas.edu \n{wdilin, sreyasmohan, fni, rakeshr, yileil, vchandra}@meta.com \nvita-group.github.io/3D-Mode-Collapse/" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 280, + 290, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 290, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 290, + 579 + ], + "type": "text", + "content": "Despite the remarkable performance of score distillation in text-to-3D generation, such techniques notoriously suffer from view inconsistency issues, also known as \"Janus\" artifact, where the generated objects fake each view with multiple front faces. Although empirically effective methods have approached this problem via score debiasing or prompt engineering, a more rigorous perspective to explain and tackle this problem remains elusive. In this paper, we reveal that the existing score distillation-based text-to-3D generation frameworks degenerate to maximal likelihood seeking on each view independently and thus suffer from the mode collapse problem, manifesting as the Janus artifact in practice. To tame mode collapse, we improve score distillation by re-establishing the entropy term in the corresponding variational objective, which is applied to the distribution of rendered images. Maximizing the entropy encourages diversity among different views in generated 3D assets, thereby mitigating the Janus problem. Based on this new objective, we derive a new update rule for 3D score distillation, dubbed Entropic Score Distillation (ESD). We theoretically reveal that ESD can be simplified and implemented by just adopting the classifier-free guidance trick upon variational score distillation. Although embarrassingly straightforward, our extensive experiments demonstrate that ESD can be an effective treatment for Janus artifacts in score distillation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 602, + 128, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 602, + 128, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 602, + 128, + 615 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 622, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 622, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 622, + 287, + 696 + ], + "type": "text", + "content": "Recent advancements in text-to-3D technology have attracted considerable attention, particularly for its pivotal role in automating high-quality 3D content. This is especially crucial in fields such as virtual reality and gaming, where 3D content forms the bedrock. While numerous techniques are available, the prevailing text-to-3D approach is based on" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 256, + 545, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 256, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 306, + 256, + 545, + 279 + ], + "type": "text", + "content": "score distillation [31], popularized by DreamFusion and its follow-up works [4, 19, 26, 50, 54, 56]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 280, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 547, + 411 + ], + "type": "text", + "content": "Score distillation leverages a pre-trained 2D diffusion model to sample over the 3D parameter space (i.e. Neural Radiance Fields (NeRF) [27]) such that views rendered from a random angle satisfy the statistics of the image distribution. This algorithm is implemented by backpropagating the estimated score of each view via the chain rule. Despite the notable progress achieved with score distillation-based approaches, it is widely observed that 3D content generated using score distillation suffers from the Janus problem [12], referring to the artifacts that generated 3D objects contain multiple canonical views (see Fig. 1)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 411, + 548, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 548, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 548, + 590 + ], + "type": "text", + "content": "To understand this drawback of score distillation, we draw the theoretical connection between the Janus problem and mode collapse, a statistical term describing a distribution concentrating on the high-density area while losing information about the probability tail. We first uncover that the optimization of existing score distillation-based text-to-3D generation degenerates to a maximum likelihood objective, making it susceptible to model collapse. As pre-trained diffusion models are biased to frequently encountered views [12], this oversight leads all views opt to convergence toward the point with the highest likelihood, manifesting as the Janus artifact in practical applications. The main limitation of current methods is that their distillation objectives solely maximize the likelihood of each view independently, without considering the diversity between different views." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 591, + 547, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 547, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 547, + 687 + ], + "type": "text", + "content": "To address the aforementioned issue, we propose a principled approach Entropic Score Distillation (ESD), which regularizes the score distillation process by entropy maximization of the rendered image distribution, thereby enhancing the diversity of views in generated 3D assets and alleviating the Janus problem. Our derived ESD update admits a simple form as a weighted combination of scores for pre-trained image distribution and rendered image distribution. Com" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "1For example, it is common that a frontal view of a cat is more likely to be sampled from latent diffusion models than the back view." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 58, + 702, + 203, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 203, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 203, + 712 + ], + "type": "text", + "content": "*Work done during an internship with Meta." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9037" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 90, + 361, + 163 + ], + "blocks": [ + { + "bbox": [ + 194, + 76, + 242, + 87 + ], + "lines": [ + { + "bbox": [ + 194, + 76, + 242, + 87 + ], + "spans": [ + { + "bbox": [ + 194, + 76, + 242, + 87 + ], + "type": "text", + "content": "Our results" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 72, + 90, + 361, + 163 + ], + "lines": [ + { + "bbox": [ + 72, + 90, + 361, + 163 + ], + "spans": [ + { + "bbox": [ + 72, + 90, + 361, + 163 + ], + "type": "image", + "image_path": "3411a3abeb213c21e1da738c11e93e90d67e603c19a36d3e0a541118c29a28e5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 164, + 255, + 174 + ], + "lines": [ + { + "bbox": [ + 174, + 164, + 255, + 174 + ], + "spans": [ + { + "bbox": [ + 174, + 164, + 255, + 174 + ], + "type": "text", + "content": "A cat with tiger stripes" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 373, + 90, + 518, + 163 + ], + "blocks": [ + { + "bbox": [ + 413, + 76, + 477, + 87 + ], + "lines": [ + { + "bbox": [ + 413, + 76, + 477, + 87 + ], + "spans": [ + { + "bbox": [ + 413, + 76, + 477, + 87 + ], + "type": "text", + "content": "\"Janus\" results" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 373, + 90, + 518, + 163 + ], + "lines": [ + { + "bbox": [ + 373, + 90, + 518, + 163 + ], + "spans": [ + { + "bbox": [ + 373, + 90, + 518, + 163 + ], + "type": "image", + "image_path": "6631107c8c53aff27d4e5038ded3ab7ee579a8639c8e6e44361baae353a32eb7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 72, + 175, + 361, + 249 + ], + "blocks": [ + { + "bbox": [ + 72, + 175, + 361, + 249 + ], + "lines": [ + { + "bbox": [ + 72, + 175, + 361, + 249 + ], + "spans": [ + { + "bbox": [ + 72, + 175, + 361, + 249 + ], + "type": "image", + "image_path": "877e227111e2f1ba9f60321151fe52c25097464e4f4e231917a57f21bb286d2d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 251, + 272, + 261 + ], + "lines": [ + { + "bbox": [ + 157, + 251, + 272, + 261 + ], + "spans": [ + { + "bbox": [ + 157, + 251, + 272, + 261 + ], + "type": "text", + "content": "A bald eagle carved out of wood" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 373, + 175, + 518, + 249 + ], + "blocks": [ + { + "bbox": [ + 373, + 175, + 518, + 249 + ], + "lines": [ + { + "bbox": [ + 373, + 175, + 518, + 249 + ], + "spans": [ + { + "bbox": [ + 373, + 175, + 518, + 249 + ], + "type": "image", + "image_path": "20786928d71d09f8973962fdde74c821b03167203b9d3f153ca22eceab7b24c9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 72, + 262, + 361, + 335 + ], + "blocks": [ + { + "bbox": [ + 72, + 262, + 361, + 335 + ], + "lines": [ + { + "bbox": [ + 72, + 262, + 361, + 335 + ], + "spans": [ + { + "bbox": [ + 72, + 262, + 361, + 335 + ], + "type": "image", + "image_path": "6756dc1ea598b650b86c8f6e4cad871f23715a945a68e3be55c8e8a445796d19.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 338, + 316, + 349 + ], + "lines": [ + { + "bbox": [ + 120, + 338, + 316, + 349 + ], + "spans": [ + { + "bbox": [ + 120, + 338, + 316, + 349 + ], + "type": "text", + "content": "A chimpanzee dressed like Henry VIII king of England" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 373, + 262, + 518, + 336 + ], + "blocks": [ + { + "bbox": [ + 373, + 262, + 518, + 336 + ], + "lines": [ + { + "bbox": [ + 373, + 262, + 518, + 336 + ], + "spans": [ + { + "bbox": [ + 373, + 262, + 518, + 336 + ], + "type": "image", + "image_path": "b259af55d7fb3fa251ea81c9ee10ba0b7455a24070247d764489869f837707b9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 338, + 430, + 347 + ], + "lines": [ + { + "bbox": [ + 389, + 338, + 430, + 347 + ], + "spans": [ + { + "bbox": [ + 389, + 338, + 430, + 347 + ], + "type": "text", + "content": "Front View" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 462, + 338, + 501, + 347 + ], + "lines": [ + { + "bbox": [ + 462, + 338, + 501, + 347 + ], + "spans": [ + { + "bbox": [ + 462, + 338, + 501, + 347 + ], + "type": "text", + "content": "Back View" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 354, + 546, + 386 + ], + "lines": [ + { + "bbox": [ + 46, + 354, + 546, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 546, + 386 + ], + "type": "text", + "content": "Figure 1. A Preview of Qualitative Results. We present the front and back views of objects synthesized by VSD (ProlificDreamer) on the right two columns, and four views of our generated results on the left. VSD suffers from \"Janus\" problem, where both front and back views contain a frontal face of the targeted object, while our method effectively mitigates this artifact. Please refer to more results in Appendix D." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 397, + 288, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 288, + 552 + ], + "type": "text", + "content": "pared with Score Distillation Sampling (SDS) [31], our ESD involves the score of the rendered image distribution, serving to maximize the entropy of the rendered image distribution. Unlike Variational Score Distillation (VSD) [56], the learned score function of the rendered image distribution does not depend on the camera pose. This subtle difference has a more profound impact, as we show the score function of rendered images modeled by VSD corresponds to an objective with fixed entropy, thereby having no influence on view variety. In contrast, ESD optimizes for a Kullback-Leibler divergence with a non-constant entropy term parameterized by the 3D model, leading to an effect that encourages diversity among different views." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 555, + 288, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 288, + 686 + ], + "type": "text", + "content": "In practice, we find it challenging to optimize the score of the rendered image distribution without conditioning on the camera pose. To facilitate training, we discover that the gradient from the entropy can be decomposed into a combination of scores: one depends on the camera pose, and the other independent of it, with a coefficient interacting between these two terms. Through this theoretical establishment, we are able to adopt a handy implementation of ESD by Classifier Free Guidance (CFG) trick [10] where conditional and unconditional scores are trained alternatively and mixed during inference." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "Through extensive experiments with our proposed ESD, we demonstrate its efficacy in alleviating the Janus problem" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 397, + 547, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 547, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 547, + 493 + ], + "type": "text", + "content": "and its significant advantages in improving 3D generation quality when compared to the baseline methods [31, 56] and other remedy techniques [2, 12]. As a side contribution, we also borrow two inception scores [36] to evaluate text-to-3D results and numerically probe model collapse in score distillation. We show these two metrics can effectively characterize the quality and diversity of views, highly matching our qualitative observations." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 506, + 384, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 384, + 520 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 384, + 520 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 526, + 408, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 408, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 408, + 538 + ], + "type": "text", + "content": "2.1. Diffusion Models" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "Diffusion models, as demonstrated by various works [11, 44, 46, 48], have shown to be highly effective in text-to-image generation. Technically, a diffusion model learns to gradually transform a normal distribution " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " to the target distribution " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "p_{data}(\\boldsymbol {x}|\\boldsymbol {y})" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " denotes the text prompt embeddings. The sampling trajectory is determined by a forward process with the conditional probability " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "p_t(\\pmb {x}_t|\\pmb {x}_0) = \\mathcal{N}(\\pmb {x}_t|\\alpha_t\\pmb {x}_0,\\sigma_t^2\\pmb {I})" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " ,where " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\pmb {x}_t\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " represents the sample at time " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "t\\in [0,T]" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{t},\\sigma_{t} > 0" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " are time-dependent diffusion coefficients. Consequently, the distribution at time " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " can be formulated as " + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "p_t(\\pmb {x}_t|\\pmb {y}) =" + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\int p_{data}(\\pmb {x}_0|\\pmb {y})\\mathcal{N}(\\pmb {x}_t|\\alpha_t\\pmb {x}_0,\\sigma_t^2\\pmb {I})d\\pmb {x}_0." + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": " Diffusion models generate samples through a reverse process starting from Gaussian noises, which can be described by the ODE:" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9038" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 68, + 271, + 194 + ], + "blocks": [ + { + "bbox": [ + 51, + 68, + 271, + 194 + ], + "lines": [ + { + "bbox": [ + 51, + 68, + 271, + 194 + ], + "spans": [ + { + "bbox": [ + 51, + 68, + 271, + 194 + ], + "type": "image", + "image_path": "21ba407548e7b038693c647e4ca8d1f10418ff0ad9162d38d5595fcaba68d9bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 197, + 547, + 231 + ], + "lines": [ + { + "bbox": [ + 46, + 197, + 547, + 231 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 547, + 231 + ], + "type": "text", + "content": "Figure 2. Illustration of the effect of entropy regularization. Learned image distributions often exhibit a higher probability mass for objects' frontal faces. Pure maximal likelihood seeking is opt to mode collapse (Sec. 3). Adding entropy regularization can expand the support of fitted distribution " + }, + { + "bbox": [ + 46, + 197, + 547, + 231 + ], + "type": "inline_equation", + "content": "q_{t}^{\\theta}(\\boldsymbol{x}|\\boldsymbol{y})" + }, + { + "bbox": [ + 46, + 197, + 547, + 231 + ], + "type": "text", + "content": " with mode-covering behavior (Sec. 4)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 288, + 68, + 523, + 194 + ], + "blocks": [ + { + "bbox": [ + 288, + 68, + 523, + 194 + ], + "lines": [ + { + "bbox": [ + 288, + 68, + 523, + 194 + ], + "spans": [ + { + "bbox": [ + 288, + 68, + 523, + 194 + ], + "type": "image", + "image_path": "916da4f02dfc3f104e790dd200d7ab14c69a42883c50231afaaad73cf2b61d27.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 241, + 288, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 288, + 301 + ], + "type": "inline_equation", + "content": "\\mathrm{d}\\pmb{x}_t / \\mathrm{d}t = -\\nabla_{\\pmb{x}}\\log p_t(\\pmb{x}_t)" + }, + { + "bbox": [ + 46, + 241, + 288, + 301 + ], + "type": "text", + "content": " with the boundary condition " + }, + { + "bbox": [ + 46, + 241, + 288, + 301 + ], + "type": "inline_equation", + "content": "\\pmb{x}_T\\sim \\mathcal{N}(\\pmb {0},\\pmb {I})" + }, + { + "bbox": [ + 46, + 241, + 288, + 301 + ], + "type": "text", + "content": " [22, 45, 48]. Such a process requires the computation of score function " + }, + { + "bbox": [ + 46, + 241, + 288, + 301 + ], + "type": "inline_equation", + "content": "\\nabla_{\\pmb{x}}\\log p_t(\\pmb{x}_t)" + }, + { + "bbox": [ + 46, + 241, + 288, + 301 + ], + "type": "text", + "content": " which is often obtained by fitting a time-conditioned noise estimator " + }, + { + "bbox": [ + 46, + 241, + 288, + 301 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}:\\mathbb{R}^{D}\\to \\mathbb{R}^{D}" + }, + { + "bbox": [ + 46, + 241, + 288, + 301 + ], + "type": "text", + "content": " using score matching loss [15, 47, 52]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 311, + 203, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 203, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 203, + 322 + ], + "type": "text", + "content": "2.2. Text-to-3D Score Distillation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "content": "Score distillation based 3D asset generation requires representing 3D scenes as learnable parameters " + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} \\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "content": " equipped with a differentiable renderer " + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "inline_equation", + "content": "g(\\pmb{\\theta}, \\pmb{c}): \\mathbb{R}^{N} \\to \\mathbb{R}^{D}" + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "content": " that projects 3D scene " + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "content": " into images with respect to the camera pose " + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\pmb{c}" + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "inline_equation", + "content": "N, D" + }, + { + "bbox": [ + 46, + 330, + 287, + 426 + ], + "type": "text", + "content": " are the dimensions of the 3D parameter space and rendered images, respectively. Neural radiance fields (NeRF) [27] are often employed as the underlying 3D representation for its capability of modeling complex scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 288, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 498 + ], + "type": "text", + "content": "Recent works [4, 14, 19, 26, 31, 50, 54-56] demonstrate the feasibility of using a pretrained 2D diffusion model to guide 3D object creation. Below, we elaborate on two score distillation schemes, adopted therein: Score Distillation Sampling (SDS) [31] and Variational Score Distillation (VSD) [56]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 516, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 287, + 540 + ], + "type": "text", + "content": "Score Distillation Sampling (SDS). SDS updates the 3D parameter " + }, + { + "bbox": [ + 46, + 516, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 516, + 287, + 540 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 550, + 296, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 296, + 587 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 296, + 587 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {\\theta}} J _ {S D S} (\\boldsymbol {\\theta}) = - \\mathbb {E} \\left[ \\omega (t) \\frac {\\partial g (\\boldsymbol {\\theta} , \\boldsymbol {c})}{\\partial \\boldsymbol {\\theta}} \\left(\\sigma_ {t} \\nabla \\log p _ {t} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {y}\\right) - \\boldsymbol {\\epsilon}\\right) \\right], \\tag {1}", + "image_path": "c5db0a9aabaa67b17c057b6107f3e5e204bf9d7a041cc7dfbf28c032e260138f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "text", + "content": "where the expectation is taken over timestep " + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "inline_equation", + "content": "t\\sim \\mathcal{U}[0,T]" + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "text", + "content": " Gaussian noises " + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "text", + "content": " and camera pose " + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "inline_equation", + "content": "\\pmb {c}\\sim p_c(\\pmb {c})" + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "text", + "content": " Here is " + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "inline_equation", + "content": "\\nabla \\log p" + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "text", + "content": " is a pre-trained diffusion model " + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\pmb {x},t,\\pmb {y})" + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "text", + "content": " is a noisy version of the rendering given by camera pose " + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "inline_equation", + "content": "\\pmb {c}" + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "inline_equation", + "content": "x_{t} = \\alpha_{t}g(\\pmb {\\theta},\\pmb {c}) + \\sigma_{t}\\pmb{\\epsilon}" + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "text", + "content": " .Updating " + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 46, + 599, + 288, + 683 + ], + "type": "text", + "content": " as in Eq. (1) has been shown to minimize the evidence lower bound (ELBO) for the rendered images, see Wang et al. [54], Xu et al. [59]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 241, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 241, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 241, + 547, + 277 + ], + "type": "text", + "content": "Variational Score Distillation (VSD). VSD [56] is introduced in ProlificDreamer, VSD improves upon SDS by deriving the following Wasserstein gradient flow [51]:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 285, + 545, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 285, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 314, + 285, + 545, + 339 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\boldsymbol {\\theta}} J _ {V S D} (\\boldsymbol {\\theta}) = - \\mathbb {E} \\left[ \\omega (t) \\frac {\\partial g (\\boldsymbol {\\theta} , \\boldsymbol {c})}{\\partial \\boldsymbol {\\theta}} \\left(\\sigma_ {t} \\nabla \\log p _ {t} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}\\right) \\right. \\\\ \\left. - \\sigma_ {t} \\nabla \\log q _ {t} (\\boldsymbol {x} _ {t} | \\boldsymbol {c})\\right) \\Bigg ]. \\tag {2} \\\\ \\end{array}", + "image_path": "a49d03fc76e48f094d1877c6a871e9dc2c506983e8ddcc85f435ae8a01eab9be.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 347, + 546, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 347, + 546, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 546, + 418 + ], + "type": "text", + "content": "Similarly, " + }, + { + "bbox": [ + 305, + 347, + 546, + 418 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t = \\alpha_t g(\\pmb{\\theta}, \\pmb{c}) + \\sigma_t \\pmb{\\epsilon}" + }, + { + "bbox": [ + 305, + 347, + 546, + 418 + ], + "type": "text", + "content": " is the noisy observation of the rendered image. In contrast to SDS, VSD introduces a new score function of the noisy rendered images conditioned on the camera pose " + }, + { + "bbox": [ + 305, + 347, + 546, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 305, + 347, + 546, + 418 + ], + "type": "text", + "content": ". To obtain this score, Wang et al. [56] fine-tunes a diffusion model using images rendered from the 3D scene as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 427, + 545, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 427, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 318, + 427, + 545, + 447 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {\\psi}} \\mathbb {E} \\left[ \\omega (t) \\| \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\psi}} \\left(\\alpha_ {t} g (\\boldsymbol {\\theta}, \\boldsymbol {c}) + \\sigma_ {t} \\boldsymbol {\\epsilon}, t, \\boldsymbol {c}, \\boldsymbol {y}\\right) - \\boldsymbol {\\epsilon} \\| _ {2} ^ {2} \\right], \\tag {3}", + "image_path": "59845fee54d24a893d2de9bcef432a2bc2998fa74c7f81d3c4ff67ca887ae085.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 455, + 547, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 455, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 305, + 455, + 547, + 504 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 455, + 547, + 504 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}(\\boldsymbol{x},t,\\boldsymbol{c},\\boldsymbol{y})" + }, + { + "bbox": [ + 305, + 455, + 547, + 504 + ], + "type": "text", + "content": " is the noise estimator of " + }, + { + "bbox": [ + 305, + 455, + 547, + 504 + ], + "type": "inline_equation", + "content": "\\nabla \\log q_t(\\boldsymbol{x}_t|\\boldsymbol{c})" + }, + { + "bbox": [ + 305, + 455, + 547, + 504 + ], + "type": "text", + "content": " as in diffusion models. As proposed in ProlificDreamer, " + }, + { + "bbox": [ + 305, + 455, + 547, + 504 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 305, + 455, + 547, + 504 + ], + "type": "text", + "content": " is parameterized by LoRA [13] and initialized from a pre-trained diffusion model same as " + }, + { + "bbox": [ + 305, + 455, + 547, + 504 + ], + "type": "inline_equation", + "content": "\\nabla \\log p_t" + }, + { + "bbox": [ + 305, + 455, + 547, + 504 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 514, + 545, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 545, + 528 + ], + "type": "text", + "content": "3. Revealing Mode Collapse in Score Distillation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "type": "text", + "content": "Despite the remarkable performance of SDS and VSD in 3D asset generation, it is widely observed that the synthesized objects suffer from \"Janus\" artifacts. Janus artifacts refer to the generated 3D scene containing multiple canonical views (the most representative perspective of the object such as the frontal face). In earlier works, Hong et al. [12] and Huang et al. [14] attribute this problem to unimodality of the learned 2D image distribution since the training data for the diffusion models are naturally biased to the most commonly seen views per each category. In this section, we examine extant distillation schemes from a statistical view, which has been overlooked in previous literature." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "In principle, natural 2D images can be seen as random projections of 3D scenes. Score distillation matches the image distribution generated by randomly sampled views with a" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "2Without special specification, expectations are taken over all relevant random variables and Jacobian matrices are transposed by default." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9039" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "content": "text-conditioned image distribution to recover the underlying 3D representation. Hence, Janus artifact, in which each view becomes uniform and identical to the most commonly seen views, can be interpreted as a manifestation of distribution collapse to samples within the high-density region. Such distribution degeneration essentially corresponds to the statistical phenomenon mode collapse, which happens when an optimized distribution fails to characterize the data diversity and concentrates on a single type of output [1, 7, 25, 36, 49]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 180, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 180, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 180, + 289, + 240 + ], + "type": "text", + "content": "Below we theoretically reveal why SDS and VSD are prone to mode collapse. As shown in Poole et al. [31], Wang et al. [56], SDS and VSD equals to the gradient of the following Kullback-Leibler (KL) divergence, i.e., " + }, + { + "bbox": [ + 47, + 180, + 289, + 240 + ], + "type": "inline_equation", + "content": "J_{SDS}(\\pmb{\\theta}) = J_{VSD}(\\pmb{\\theta}) = J_{KL}(\\pmb{\\theta})" + }, + { + "bbox": [ + 47, + 180, + 289, + 240 + ], + "type": "text", + "content": " up to an additive constant:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 247, + 288, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 247, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 59, + 247, + 288, + 262 + ], + "type": "interline_equation", + "content": "J _ {K L} (\\boldsymbol {\\theta}) = \\mathbb {E} \\left[ \\Omega (t) \\mathcal {D} _ {\\mathrm {K L}} \\left(q _ {t} ^ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {t} | \\boldsymbol {c}, \\boldsymbol {y}\\right) \\| p _ {t} \\left(\\boldsymbol {x} _ {t} | \\boldsymbol {y}\\right)\\right) \\right], \\tag {4}", + "image_path": "458c25096af4b77cb8eba42f90a2eb5c93a1812c61020c84c00f6c439148c708.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "inline_equation", + "content": "\\Omega(t) = \\omega(t)\\sigma_t / \\alpha_t" + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "text", + "content": " and the expectation is taken over " + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "inline_equation", + "content": "t \\sim \\mathcal{U}[0,T]" + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "inline_equation", + "content": "\\pmb{c} \\sim p_c(\\pmb{c})" + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "inline_equation", + "content": "p_t(\\pmb{x}_t|\\pmb{y}) = \\int p_0(\\pmb{x}_0|\\pmb{y})\\mathcal{N}(\\pmb{x}_t|\\alpha_t\\pmb{x}_0,\\sigma_t^2\\pmb{I})d\\pmb{x}_0" + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "text", + "content": " is the image distribution perturbed by Gaussian noises, while " + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "inline_equation", + "content": "q_t^\\theta (\\pmb{x}_t|\\pmb{c},\\pmb{y}) = \\int q_0^\\theta (\\pmb{x}_0|\\pmb{c})\\mathcal{N}(\\pmb{x}_t|\\alpha_t\\pmb{x}_0,\\sigma_t^2\\pmb{I})d\\pmb{x}_0" + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "text", + "content": " models the image distribution generated by 3D parameter " + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "text", + "content": " with respect to camera pose " + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "inline_equation", + "content": "\\pmb{c}" + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "text", + "content": " and diffused by Gaussian distribution. As shown by Wang et al. [56], " + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "inline_equation", + "content": "J_{KL}(\\pmb{\\theta}) = 0" + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "text", + "content": " implies " + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "inline_equation", + "content": "q_0^\\theta (\\pmb{x}_0|\\pmb{c}) = p(\\pmb{x}_0|\\pmb{y})" + }, + { + "bbox": [ + 46, + 270, + 289, + 389 + ], + "type": "text", + "content": ", i.e., the distribution of synthesized views satisfy the text-conditioned image distribution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 390, + 289, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 390, + 289, + 450 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 289, + 450 + ], + "type": "text", + "content": "However, it has not escaped from our notice that " + }, + { + "bbox": [ + 47, + 390, + 289, + 450 + ], + "type": "inline_equation", + "content": "q_0^\\theta(\\boldsymbol{x}_0|\\boldsymbol{c}) = \\delta(\\boldsymbol{x}_0 - g(\\boldsymbol{\\theta},\\boldsymbol{c}))" + }, + { + "bbox": [ + 47, + 390, + 289, + 450 + ], + "type": "text", + "content": " is a Dirac distribution for both SDS and VSD. This causes the original KL divergence minimization (Eq. 4) degenerate to a Maximal Likelihood Estimation (MLE) problem:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 456, + 287, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 456, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 58, + 456, + 287, + 523 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} J _ {K L} (\\boldsymbol {\\theta}) = \\underbrace {- \\mathbb {E} \\left[ \\Omega (t) \\mathbb {E} _ {\\boldsymbol {x} _ {t} \\sim q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {c} , \\boldsymbol {y})} \\log p _ {t} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}) \\right]} _ {J _ {M L E} (\\boldsymbol {\\theta})} \\tag {5} \\\\ - \\underbrace {\\mathbb {E} \\left[ \\Omega (t) H \\left[ q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {c} , \\boldsymbol {y}) \\right] \\right]} _ {\\text {c o n s t .}}, \\\\ \\end{array}", + "image_path": "ecc468171810bb3fda5783705fc8174ce5dea685ca4c5af8f91d7d374a194efd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 531, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 531, + 287, + 593 + ], + "type": "inline_equation", + "content": "H[q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{y})] = -\\mathbb{E}_{\\boldsymbol{x}_t \\sim q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{c}, \\boldsymbol{y})}[\\log q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{c}, \\boldsymbol{y})]" + }, + { + "bbox": [ + 46, + 531, + 287, + 593 + ], + "type": "text", + "content": " denotes the entropy of " + }, + { + "bbox": [ + 46, + 531, + 287, + 593 + ], + "type": "inline_equation", + "content": "q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{y})" + }, + { + "bbox": [ + 46, + 531, + 287, + 593 + ], + "type": "text", + "content": ", which turns out to be a constant because " + }, + { + "bbox": [ + 46, + 531, + 287, + 593 + ], + "type": "inline_equation", + "content": "q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{c}, \\boldsymbol{y}) = \\mathcal{N}(\\boldsymbol{x}_t | \\alpha_t g(\\boldsymbol{\\theta}, \\boldsymbol{c}), \\sigma_t^2 \\boldsymbol{I})" + }, + { + "bbox": [ + 46, + 531, + 287, + 593 + ], + "type": "text", + "content": " which has fixed entropy once " + }, + { + "bbox": [ + 46, + 531, + 287, + 593 + ], + "type": "inline_equation", + "content": "t, \\boldsymbol{\\theta}" + }, + { + "bbox": [ + 46, + 531, + 287, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 531, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\boldsymbol{c}" + }, + { + "bbox": [ + 46, + 531, + 287, + 593 + ], + "type": "text", + "content": " have been specified. See full derivation in Appendix A.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": "Note that Eq. 5 signifies " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "J_{KL}(\\pmb{\\theta}) = J_{MLE}(\\pmb{\\theta})" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": " up to an additive constant, hence " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "J_{KL}(\\pmb{\\theta})" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": " shares all minima with " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "J_{MLE}(\\pmb{\\theta})" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": ". It is known that likelihood maximization is more prone to mode collapse. Intuitively, minimizing " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "J_{MLE}(\\pmb{\\theta})" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": " seeks each view independently to have the maximum log-likelihood on the image distribution " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_0|\\pmb{y})" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_0|\\pmb{y})" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": " is usually unimodal and peaks at the canonical view, each view of the scene will collapse to the same local minimum, resulting in Janus artifact (see Fig. 2). We postulate that the existing distillation strategies may be inherently limited" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 543, + 162 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 543, + 162 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 543, + 162 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 543, + 162 + ], + "type": "image", + "image_path": "a33b82e50ad9746c4eb02291f66fe6c94ddbb4dd5361afb564e95a855af58acb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 163, + 545, + 207 + ], + "lines": [ + { + "bbox": [ + 304, + 163, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 545, + 207 + ], + "type": "text", + "content": "Figure 3. Gaussian Example. To illustrate the effects of entropy regularization, we leverage SDS, VSD and ESD to fit a 2D Gaussian distribution. The blue points are sampled from the ground-truth distribution while the orange points are from the fitted distribution." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 218, + 545, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 545, + 253 + ], + "type": "text", + "content": "by their log-likelihood seeking behaviors, which are more susceptible to mode collapse, especially with biased image distributions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 266, + 519, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 519, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 519, + 280 + ], + "type": "text", + "content": "4. Entropy Regularized Score Distillation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 295, + 453, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 295, + 453, + 308 + ], + "spans": [ + { + "bbox": [ + 305, + 295, + 453, + 308 + ], + "type": "text", + "content": "4.1. Entropic Score Distillation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 314, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 397 + ], + "type": "text", + "content": "In this section, we highlight the importance of the entropy in score distillation. It is known that higher entropy implies the corresponding distribution could cover a larger support of the ambient space and thus increase the sample diversity. In Eq. 5, the entropy term is shown to diminish in the training objective, which causes each generated view to lack diversity and collapse to a single image with the highest likelihood." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 398, + 547, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 547, + 470 + ], + "type": "text", + "content": "To this end, we propose to bring in an entropy regularization to " + }, + { + "bbox": [ + 304, + 398, + 547, + 470 + ], + "type": "inline_equation", + "content": "J_{MLE}(\\pmb{\\theta})" + }, + { + "bbox": [ + 304, + 398, + 547, + 470 + ], + "type": "text", + "content": " for boosting the view diversity. Since " + }, + { + "bbox": [ + 304, + 398, + 547, + 470 + ], + "type": "inline_equation", + "content": "q_t^\\pmb{\\theta}(\\pmb{x}_t|\\pmb{c},\\pmb{y})" + }, + { + "bbox": [ + 304, + 398, + 547, + 470 + ], + "type": "text", + "content": " has constant entropy, we regularize entropy for the distribution " + }, + { + "bbox": [ + 304, + 398, + 547, + 470 + ], + "type": "inline_equation", + "content": "q_t^\\pmb{\\theta}(\\pmb{x}_t|\\pmb{y}) = \\int q_t^\\pmb{\\theta}(\\pmb{x}_t|\\pmb{c},\\pmb{y})p_c(\\pmb{c})d\\pmb{c}" + }, + { + "bbox": [ + 304, + 398, + 547, + 470 + ], + "type": "text", + "content": ", which can be simulated by randomly sampling views from the 3D parameter " + }, + { + "bbox": [ + 304, + 398, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 304, + 398, + 547, + 470 + ], + "type": "text", + "content": ". Consider the following objective:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 479, + 545, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 479, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 310, + 479, + 545, + 516 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} J _ {E n t} (\\boldsymbol {\\theta}, \\lambda) = - \\mathbb {E} \\left[ \\Omega (t) \\mathbb {E} _ {\\boldsymbol {x} _ {t} \\sim q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {c}, \\boldsymbol {y})} \\log p _ {t} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}) \\right] \\tag {7} \\\\ - \\lambda \\mathbb {E} \\left[ \\Omega (t) H \\left[ q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}) \\right] \\right], \\\\ \\end{array}", + "image_path": "b1828c02b220ade694b11365c1fd1aa91610b8cd739ff0717186a698bc8a3089.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 525, + 547, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 547, + 668 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 525, + 547, + 668 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 525, + 547, + 668 + ], + "type": "text", + "content": " is a hyper-parameter controlling the regularization strength. We note that without " + }, + { + "bbox": [ + 304, + 525, + 547, + 668 + ], + "type": "inline_equation", + "content": "H[q_t^\\theta (\\pmb {x}_t|\\pmb {y})]" + }, + { + "bbox": [ + 304, + 525, + 547, + 668 + ], + "type": "text", + "content": " , each view is optimized independently and implicitly regularized by the underlying parameterization. However, upon imposing " + }, + { + "bbox": [ + 304, + 525, + 547, + 668 + ], + "type": "inline_equation", + "content": "H[q_t^\\theta (\\pmb {x}_t|\\pmb {y})]" + }, + { + "bbox": [ + 304, + 525, + 547, + 668 + ], + "type": "text", + "content": " , all views become explicitly correlated with each other, as they collectively contribute to the entropy computation. Intuitively, " + }, + { + "bbox": [ + 304, + 525, + 547, + 668 + ], + "type": "inline_equation", + "content": "J_{Ent}(\\pmb {\\theta},\\lambda) = J_{MLE}(\\pmb {\\theta})-" + }, + { + "bbox": [ + 304, + 525, + 547, + 668 + ], + "type": "inline_equation", + "content": "\\lambda \\mathbb{E}[\\Omega (t)H[q_t^\\theta (\\pmb {x}_t|\\pmb {y})]]" + }, + { + "bbox": [ + 304, + 525, + 547, + 668 + ], + "type": "text", + "content": " seeks the maximal log-likelihood for each view while simultaneously enlarging the entropy for distribution " + }, + { + "bbox": [ + 304, + 525, + 547, + 668 + ], + "type": "inline_equation", + "content": "q_{t}^{\\theta}(\\pmb{x}_{t}|\\pmb{y})" + }, + { + "bbox": [ + 304, + 525, + 547, + 668 + ], + "type": "text", + "content": " , which spans the support and encourages diversity across the rendered views. To gain more insights, we present the following theoretical results:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "content": "Theorem 1. For any " + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\lambda \\in \\mathbb{R}" + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "inline_equation", + "content": "J_{Ent}(\\pmb{\\theta},\\lambda) = \\lambda \\mathbb{E}_t[\\Omega(t)\\mathcal{D}_{\\mathrm{KL}}(q_t^\\pmb{\\theta}(\\pmb{x}_t|\\pmb{y})||p_t(\\pmb{x}_t|\\pmb{y}))] + (1 - \\lambda)\\mathbb{E}_{t,\\pmb{c}}[\\Omega(t)\\mathcal{D}_{\\mathrm{KL}}(q_t^\\pmb{\\theta}(\\pmb{x}_t|\\pmb{c},\\pmb{y})||p_t(\\pmb{x}_t|\\pmb{y}))] + const." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9040" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 56, + 87, + 547, + 228 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 333, + 84 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 333, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 333, + 84 + ], + "type": "text", + "content": "Algorithm 1 ESD: Entropic score distillation for text-to-3D generation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "text", + "content": "Input: A diffusion model " + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\pmb {x},t,\\pmb {y})" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "text", + "content": " ; learnable 3D parameter " + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "text", + "content": " ; coefficient " + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "text", + "content": " ; text prompt " + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "text", + "content": " \nInitialize " + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "text", + "content": " for another diffusion model " + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}(\\pmb {x},t,\\pmb {y})" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "text", + "content": " with the parameter " + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "text", + "content": " specified in diffusion model " + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\pmb {x},t,\\pmb {y})" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "text", + "content": " , parame \nterized with LoRA. \nwhile not converged do \nRandomly sample a camera pose " + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "c\\sim p_c" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "text", + "content": " and render a view " + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\pmb {x}_0 = g(\\pmb {\\theta},\\pmb {c})" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "text", + "content": " Sample a " + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "t\\sim \\mathcal{U}[0,T]" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "text", + "content": " and add Gaussian noise " + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "text", + "content": " .. " + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t} = \\alpha_{t}\\pmb{x}_{0} + \\sigma_{t}\\pmb{\\epsilon}" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\pmb {\\theta}\\gets \\pmb {\\theta} + \\eta_1\\left[\\omega (t)\\frac{\\partial g(\\pmb{\\theta},\\pmb{c})}{\\partial\\pmb{\\theta}} (\\epsilon_{\\phi}(\\pmb{x}_{t},t,\\pmb {y}) - \\lambda \\epsilon_{\\psi}(\\pmb{x}_{t},t,\\emptyset ,\\pmb {y}) - (1 - \\lambda)\\epsilon_{\\psi}(\\pmb{x}_{t},t,\\pmb {c},\\pmb {y})\\right]" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "text", + "content": " (6) With probability " + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "1 - p_{\\emptyset}" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\psi \\gets \\psi -\\eta_2\\nabla_{\\psi}\\left[\\omega (t)\\| \\epsilon_{\\psi}(\\pmb {x}_t,t,\\pmb {c},\\pmb {y}) - \\epsilon \\| _2^2\\right]" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "text", + "content": " Otherwise, " + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\psi \\gets \\psi -\\eta_2\\nabla_{\\psi}\\left[\\omega (t)\\| \\epsilon_{\\psi}(\\pmb {x}_t,t,\\emptyset ,\\pmb {y}) - \\epsilon \\| _2^2\\right]" + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "text", + "content": " \nend while \nReturn " + }, + { + "bbox": [ + 56, + 87, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\theta" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "type": "text", + "content": "We prove Theorem 1 in Appendix A.3. Theorem 1 implies that " + }, + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "type": "inline_equation", + "content": "J_{Ent}(\\theta, \\lambda)" + }, + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "type": "text", + "content": " essentially equal to a combination of two types of KL divergences, where the former one minimizes the distribution discrepancy between " + }, + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "type": "inline_equation", + "content": "q_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{y})" + }, + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "type": "inline_equation", + "content": "p_t^\\theta(\\boldsymbol{x}_t|\\boldsymbol{y})" + }, + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "type": "text", + "content": " which marginalizes the camera pose within " + }, + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "type": "inline_equation", + "content": "q_t^\\theta" + }, + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "type": "text", + "content": ", while the latter is the original KL divergence " + }, + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "type": "inline_equation", + "content": "J_{KL}(\\theta)" + }, + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "type": "text", + "content": " adopted by SDS and VSD which takes expectation over " + }, + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "type": "inline_equation", + "content": "\\boldsymbol{c}" + }, + { + "bbox": [ + 46, + 250, + 289, + 345 + ], + "type": "text", + "content": " out of KL divergence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 346, + 289, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 289, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 289, + 381 + ], + "type": "text", + "content": "Next, we derive the gradient of " + }, + { + "bbox": [ + 46, + 346, + 289, + 381 + ], + "type": "inline_equation", + "content": "J_{Ent}(\\theta, \\lambda)" + }, + { + "bbox": [ + 46, + 346, + 289, + 381 + ], + "type": "text", + "content": " that will be backpropagated to update the 3D representation. It can be obtained by path derivative and reparameterization trick:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 398, + 287, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 398, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 52, + 398, + 287, + 466 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\boldsymbol {\\theta}} J _ {E n t} (\\boldsymbol {\\theta}, \\lambda) = - \\mathbb {E} \\left[ \\omega (t) \\frac {\\partial g (\\boldsymbol {\\theta} , \\boldsymbol {c})}{\\partial \\boldsymbol {\\theta}} \\left(\\sigma_ {t} \\nabla \\log p _ {t} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {y}\\right) \\right. \\right. \\tag {8} \\\\ \\left. - \\lambda \\sigma_ {t} \\nabla \\log q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}) \\right]. \\\\ \\end{array}", + "image_path": "1c81de195e55745b833b274fdc286b86b09828fa27004977f00088a6a2ea3d4b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 472, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 521 + ], + "type": "text", + "content": "The full derivation is deferred to Appendix A.2. We name this update rule as Entropic Score Distillation (ESD). Note that ESD differs from VSD as its second score function does not depend on the camera pose." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 527, + 214, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 214, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 214, + 539 + ], + "type": "text", + "content": "4.2. Classifier-Free Guidance Trick" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 289, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 652 + ], + "type": "text", + "content": "Similar to SDS and VSD, we approximate " + }, + { + "bbox": [ + 46, + 545, + 289, + 652 + ], + "type": "inline_equation", + "content": "\\nabla \\log p_t(\\pmb{x}_t|\\pmb{y})" + }, + { + "bbox": [ + 46, + 545, + 289, + 652 + ], + "type": "text", + "content": " via a pre-trained diffusion model " + }, + { + "bbox": [ + 46, + 545, + 289, + 652 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\pmb{x}_t,t,\\pmb{y})" + }, + { + "bbox": [ + 46, + 545, + 289, + 652 + ], + "type": "text", + "content": ". However, " + }, + { + "bbox": [ + 46, + 545, + 289, + 652 + ], + "type": "inline_equation", + "content": "\\nabla \\log q_t^\\theta (\\pmb{x}|\\pmb{y})" + }, + { + "bbox": [ + 46, + 545, + 289, + 652 + ], + "type": "text", + "content": " is not readily available. We found that directly fine-tuning a pre-trained diffusion model using rendered images to approximate " + }, + { + "bbox": [ + 46, + 545, + 289, + 652 + ], + "type": "inline_equation", + "content": "\\nabla \\log q_t^\\theta (\\pmb {x}|\\pmb {y})" + }, + { + "bbox": [ + 46, + 545, + 289, + 652 + ], + "type": "text", + "content": ", akin to Prolific-Dreamer, does not yield robust performance. We postulate this difficulty arises from the removal of the camera condition, increasing the complexity of the distribution to be fitted." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 289, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 712 + ], + "type": "text", + "content": "To tackle this problem, we recall the result in Theorem 1 that " + }, + { + "bbox": [ + 46, + 653, + 289, + 712 + ], + "type": "inline_equation", + "content": "J_{Ent}(\\theta, \\lambda)" + }, + { + "bbox": [ + 46, + 653, + 289, + 712 + ], + "type": "text", + "content": " can be written in terms of two KL divergence losses. Therefore, its gradient can be decomposed as a weighted combination of their gradients, which correspond to unconditional and conditional score functions in terms of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 250, + 437, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 250, + 437, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 250, + 437, + 262 + ], + "type": "text", + "content": "the camera pose " + }, + { + "bbox": [ + 305, + 250, + 437, + 262 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 250, + 437, + 262 + ], + "type": "text", + "content": ", respectively:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 268, + 546, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 268, + 546, + 335 + ], + "spans": [ + { + "bbox": [ + 312, + 268, + 546, + 335 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\boldsymbol {\\theta}} J _ {E n t} (\\boldsymbol {\\theta}, \\lambda) = - \\mathbb {E} \\left[ \\omega (t) \\frac {\\partial g (\\boldsymbol {\\theta} , \\boldsymbol {c})}{\\partial \\boldsymbol {\\theta}} \\left(\\sigma_ {t} \\nabla \\log p _ {t} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {y}\\right) \\right. \\right. \\tag {9} \\\\ \\left. \\left. - \\lambda \\sigma_ {t} \\nabla \\log q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}) - (1 - \\lambda) \\sigma_ {t} \\nabla \\log q _ {t} ^ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t} | \\boldsymbol {c}, \\boldsymbol {y})\\right) \\right]. \\\\ \\end{array}", + "image_path": "a8d5c226f087346617fde2e644981bbbbc050c35e71625e75534a8b794692254.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 341, + 547, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 547, + 462 + ], + "type": "text", + "content": "We formally prove Eq. 9 in Appendix A.3. With the above formulation, ESD can be implemented via the Classifier-Free Guidance (CFG) trick, which was initially proposed to balance the variety and quality of text-conditionally generated images from diffusion models [10]. Algorithm 1 outlines the computation paradigm of ESD, in which we surrogate score functions in Eq. 9 with pre-trained and fine-tuned diffusion models (see Eq. 6), and takes random turns with a probability " + }, + { + "bbox": [ + 304, + 341, + 547, + 462 + ], + "type": "inline_equation", + "content": "p_{\\emptyset}" + }, + { + "bbox": [ + 304, + 341, + 547, + 462 + ], + "type": "text", + "content": " to balance the training of conditional and unconditional score functions, as suggested by Ho and Salimans [10]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 468, + 378, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 378, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 378, + 479 + ], + "type": "text", + "content": "4.3. Discussion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 486, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 617 + ], + "type": "text", + "content": "In VSD, the camera-conditioned score is believed to play a significant role in facilitating visual quality. Intuitively, such conditioning can equip the tuned diffusion model with multi-view priors [20]. Also, Hertz et al. [8] suggests such a method can be useful to stabilize the update of the implicit parameters. However, ESD counters this argument by suggesting that the camera condition might not always be advantageous, particularly when the particle size is reduced to one. In such cases, the resulting KL divergence provably degenerates to a likelihood maximization algorithm vulnerable to mode collapse." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 548, + 713 + ], + "type": "text", + "content": "It is noteworthy that, even though their subtle differences in implementation, the optimization objectives of ESD and VSD are fundamentally different (see Sec. 4.1). ESD sets itself apart from VSD by incorporating entropy regularization, a crucial feature absent in VSD, aiming to augment diversity across views. Despite originating from distinct objectives, our theoretical establishment allows for a straightforward implementation of ESD based on VSD using the CFG trick." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9041" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 71, + 521, + 355 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 521, + 355 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 521, + 355 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 521, + 355 + ], + "type": "image", + "image_path": "cd94397d6e389ce5637fe38b0607cdd5ba3aa76d321d0907cbd75e123487ad50.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 365, + 547, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 547, + 399 + ], + "type": "text", + "content": "Figure 4. Qualitative Results. Our proposed outperforms all baselines in terms of better geometry and well-constructed texture details. Our results deliver photo-realistic and diverse rendered views, while baseline methods more or less suffer from the Janus problem. Best view in an electronic copy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 410, + 288, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 288, + 483 + ], + "type": "text", + "content": "We provide an illustrative example by leveraging SDS, VSD and ESD (with different " + }, + { + "bbox": [ + 46, + 410, + 288, + 483 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 410, + 288, + 483 + ], + "type": "text", + "content": "'s) to fit a 2D Gaussian distribution in Fig. 3. With SDS and VSD, all samples are converged to the high-density area while ESD recovers the entire support of the distribution. We provide more details and examples in Appendix B." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 494, + 167, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 494, + 167, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 494, + 167, + 506 + ], + "type": "text", + "content": "5. Other Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 514, + 288, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 288, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 288, + 670 + ], + "type": "text", + "content": "Text-to-Image Diffusion Model. Text-to-image diffusion models [32, 33] are cornerstone components of text-to-3D generation. It involves text embedding conditioning into the iterative denoising process. Equipped with large-scale image-text paired datasets, many works [29, 33, 35] scale up to tackle text-to-image generation. Among them, latent diffusion models attracted great interest in the open-source community since they reduced the computation cost by diffusing in the low-resolution latent space instead of directly in the pixel space. In addition, text-to-image diffusion models have also found applications in various computer vision tasks, including text-to-3D [31, 43], image-to-3D [59], text-to-svq [17], text-to-video [18, 42], etc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 713 + ], + "type": "text", + "content": "3D Generation with 2D Priors. Well-annotated 3D data requires immense effort to collect. Instead, a line of research studies on how to learn 3D generative models using 2D su" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 410, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 547, + 590 + ], + "type": "text", + "content": "pervision. Early attempts, including pi-GAN [34], EG3D [3], GRAF [37], GIRAFFE [30], adopt adversarial loss between the rendered images and natural images. DreamField [16] leverages CLIP to align NeRF with text prompts. More recently, with the rapid development of text-to-image diffusion models, diffusion-based image priors have attracted increasing interest, and score distillation has then become the dominant technique. Pioneer works DreamFusion [31] and ProlificDreamer [56] have been introduced in detail in Sec. 2. Their concurrent work SJC [54] derives the score Jacobian chaining method from another theoretical viewpoint of Perturb and Average Scoring. Even though diffusion models directly trained with 3D data nowadays demonstrate largely improved results [21, 41], score distillation still plays a pivotal role in ensuring view consistency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 594, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 713 + ], + "type": "text", + "content": "Techniques to Improve Score Distillation. Providing the empirical promise of score distillation, there have been numerous techniques proposed to improve its effectiveness. Magic3D [19] and Fantasia3D [4] utilize mesh and DMTet [40] to disentangle the optimization of geometry and texture. TextMesh [50] and 3DFuse [38] use depth-conditioned text-to-image diffusion priors that support geometry-aware texturing. Score debiasing[12] and Perp-Neg [2] study to refine the text prompts for a better 3D generation. DreamTime [14] and RED-Diff [24] investigate the timestep scheduling in the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9042" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 281, + 373 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 281, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 281, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 281, + 373 + ], + "type": "image", + "image_path": "d398962526ba43d25c7ee07cc30b8700a417c90f1923353349837d2ec2520567.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 381, + 288, + 415 + ], + "lines": [ + { + "bbox": [ + 46, + 381, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 288, + 415 + ], + "type": "text", + "content": "Figure 5. Qualitative Results. We combine our proposed ESD with timestep scheduling in DreamTime [14] and compare it against baseline methods. Prompt: A caramic lion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 426, + 288, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 475 + ], + "type": "text", + "content": "score distillation process. HIFA [60] adopts multiple diffusion steps for distillation. Score distillation also works with auxiliary losses, including CLIP loss [59] and adversarial loss [5, 39]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 486, + 162, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 486, + 162, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 486, + 162, + 498 + ], + "type": "text", + "content": "6. Evaluation Metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 506, + 289, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 289, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 289, + 579 + ], + "type": "text", + "content": "In this section, we introduce four information-theoretic metrics to numerically evaluate the generated 3D results with a particular focus on identifying Janus artifacts or mode collapse. The metrics we propose comprehensively cover four aspects: 1) the relevance with the text prompts, 2) distribution fitness, 3) rendering quality, and 4) view diversity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "CLIP Distance. We compute the average distance between rendered images and the text embedding to reflect the relevance between generated results and the specified text prompt. Specifically, we render " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " views from the generated 3D representations, and for each view, we obtain an embedding vector through the image encoder of a CLIP model [53]. In the meantime, we compute the text embedding utilizing the text encoder. The CLIP distance is computed as the one minus cosine similarity between the image embeddings and text embeddings averaged over all views." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": "Frechet inception distance (FID). As shown in Sec. 3 and 4, score distillation essentially matches distributions via KL divergence. Hence, it becomes reasonable to employ FID to measure the distance between the image distribution " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "q^{\\theta}(\\pmb{x}_0|\\pmb{y})" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " generated by randomly rendering 3D representation and the text-conditioned image distribution " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_0|\\pmb{y})" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " modeled by a diffusion model. We sample " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " images using pre-trained latent diffusion model given text prompts as the ground truth image dataset, and render " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " views uniformly distributed over a unit sphere from the optimized 3D scene as the generated image dataset. Then standard FID [9] is computed between these two sets of images. Note that FID is known to be effective in quantitatively identifying mode collapse." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 254, + 547, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 547, + 338 + ], + "type": "text", + "content": "Inception Quality and Variety. Thanks to our established connection with mode collapse, we know that Janus problem is due to a lack of sample diversity. Inspired by Inception Score (IS) [36], we utilize entropy-related metrics to reflect the generated image quality and diversity. We propose Inception Quality (IQ) and Inception Variety (IV), formulated as below:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 354, + 347, + 545, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 347, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 354, + 347, + 545, + 360 + ], + "type": "interline_equation", + "content": "I Q (\\boldsymbol {\\theta}) = \\mathbb {E} _ {\\boldsymbol {c}} \\left[ H \\left[ p _ {c l s} (\\boldsymbol {y} | g (\\boldsymbol {\\theta}, \\boldsymbol {c})) \\right] \\right], \\tag {10}", + "image_path": "8d77b2914916841c38f51975d123f1789e45deccca7d3a8ff27af5465dd5c000.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 356, + 361, + 545, + 375 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 361, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 356, + 361, + 545, + 375 + ], + "type": "interline_equation", + "content": "I V (\\boldsymbol {\\theta}) = H \\left[ \\mathbb {E} _ {\\boldsymbol {c}} \\left[ p _ {c l s} (\\boldsymbol {y} | g (\\boldsymbol {\\theta}, \\boldsymbol {c}) \\right] \\right], \\tag {11}", + "image_path": "3d6273b6a2e661c8babd940c664bad7bc15a63ffddd0ebd66e28ac312e55eeec.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 384, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 547, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 384, + 547, + 552 + ], + "type": "inline_equation", + "content": "p_{cls}(\\boldsymbol{y}|\\boldsymbol{x})" + }, + { + "bbox": [ + 304, + 384, + 547, + 552 + ], + "type": "text", + "content": " is a pre-trained classifier. IQ computes the average entropy of the label logits predicted for all rendered views, while IV computes the entropy of the averaged label logits of all rendered views. Intuitively, the smaller IQ means highly confident classification results on rendered views, which also indicates better visual quality of generated 3D assets. In the meanwhile, the higher IV signifies that each rendered view is likely to have a distinct label prediction, meaning the 3D creation has higher view diversity. Note that IV upper bounds IQ due to Jensen inequality. So we can define Inception Gain " + }, + { + "bbox": [ + 304, + 384, + 547, + 552 + ], + "type": "inline_equation", + "content": "IG = (IV - IQ) / IQ" + }, + { + "bbox": [ + 304, + 384, + 547, + 552 + ], + "type": "text", + "content": ", which characterizes the information gain brought by knowing where the camera pose is, namely the improvement of distinguishability among different views." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 562, + 388, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 388, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 388, + 575 + ], + "type": "text", + "content": "7. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 714 + ], + "type": "text", + "content": "Settings. In this section, we empirically validate the effectiveness of our proposal. The chosen prompts are targeted at objects with clearly defined canonical views, posing a challenge for existing methods. Our baseline approaches include SDS (DreamFusion) [31] and VSD (ProlificDreamer) [56], as well as two methods dedicated to solving Janus problem: Debiased-SDS [12] and Perp-Neg [2]. For fair comparison, all experiments are benchmarked under the open-source threestudio framework. Geometry refinement [56] is adopted for all distillation schemes. Please refer to Appendix C for more implementation details." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9043" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 89, + 219, + 163 + ], + "blocks": [ + { + "bbox": [ + 127, + 76, + 165, + 87 + ], + "lines": [ + { + "bbox": [ + 127, + 76, + 165, + 87 + ], + "spans": [ + { + "bbox": [ + 127, + 76, + 165, + 87 + ], + "type": "inline_equation", + "content": "\\lambda = 1.0" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 72, + 89, + 219, + 163 + ], + "lines": [ + { + "bbox": [ + 72, + 89, + 219, + 163 + ], + "spans": [ + { + "bbox": [ + 72, + 89, + 219, + 163 + ], + "type": "image", + "image_path": "818f7bf7fa32591a0ca0eb1e8e55e6964ac177856f04e013e3e4e1a7b9560901.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 222, + 89, + 369, + 163 + ], + "blocks": [ + { + "bbox": [ + 278, + 76, + 315, + 87 + ], + "lines": [ + { + "bbox": [ + 278, + 76, + 315, + 87 + ], + "spans": [ + { + "bbox": [ + 278, + 76, + 315, + 87 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 222, + 89, + 369, + 163 + ], + "lines": [ + { + "bbox": [ + 222, + 89, + 369, + 163 + ], + "spans": [ + { + "bbox": [ + 222, + 89, + 369, + 163 + ], + "type": "image", + "image_path": "67d990746d35d70bff52a1f66c12fcbada67207c870ca4b0d6c2823254550fa8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 168, + 544, + 190 + ], + "lines": [ + { + "bbox": [ + 46, + 168, + 544, + 190 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 544, + 190 + ], + "type": "text", + "content": "Figure 6. Ablation Studies on " + }, + { + "bbox": [ + 46, + 168, + 544, + 190 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 168, + 544, + 190 + ], + "type": "text", + "content": ". We investigate the choice of different entropy regularization strength " + }, + { + "bbox": [ + 46, + 168, + 544, + 190 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 168, + 544, + 190 + ], + "type": "text", + "content": ". Prompt: Michelangelo-style statue of dog reading news on a cellphone." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 372, + 88, + 518, + 163 + ], + "blocks": [ + { + "bbox": [ + 431, + 76, + 468, + 87 + ], + "lines": [ + { + "bbox": [ + 431, + 76, + 468, + 87 + ], + "spans": [ + { + "bbox": [ + 431, + 76, + 468, + 87 + ], + "type": "inline_equation", + "content": "\\lambda = 0.0" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 372, + 88, + 518, + 163 + ], + "lines": [ + { + "bbox": [ + 372, + 88, + 518, + 163 + ], + "spans": [ + { + "bbox": [ + 372, + 88, + 518, + 163 + ], + "type": "image", + "image_path": "c8f6ac5623ee5032e1fbb441d7bd99d83d49dd39c2c2f303233787d1eeb342f2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 50, + 226, + 286, + 273 + ], + "blocks": [ + { + "bbox": [ + 47, + 200, + 286, + 222 + ], + "lines": [ + { + "bbox": [ + 47, + 200, + 286, + 222 + ], + "spans": [ + { + "bbox": [ + 47, + 200, + 286, + 222 + ], + "type": "text", + "content": "Table 1. Quantitative Comparisons. " + }, + { + "bbox": [ + 47, + 200, + 286, + 222 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 47, + 200, + 286, + 222 + ], + "type": "text", + "content": " means the lower the better, and " + }, + { + "bbox": [ + 47, + 200, + 286, + 222 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 47, + 200, + 286, + 222 + ], + "type": "text", + "content": " means the higher the better." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 226, + 286, + 273 + ], + "lines": [ + { + "bbox": [ + 50, + 226, + 286, + 273 + ], + "spans": [ + { + "bbox": [ + 50, + 226, + 286, + 273 + ], + "type": "table", + "html": "
CLIP (↓)FID (↓)IQ (↓)IV (↑)IG (↑)SR (↑)
SDS0.737291.8604.2954.85520.12315.00%
VSD0.725265.1413.1493.57120.13719.17%
ESD0.714235.9153.1354.03140.32755.83%
", + "image_path": "6cf0732716333822d2e60ac4e46ce259dcd64041f25246c916a8243019c649b8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 283, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 449 + ], + "type": "text", + "content": "Qualitative Comparison. We present qualitative comparisons in Fig. 4. We encourage interested readers to Appendix D for more results and our project page for videos. It is clearly shown that our proposed ESD delivers more precise geometry with the Janus problem alleviated. In comparison, the results presented by SDS and VSD all contain more or less corrupted geometry with multi-face structures. Debiased-SDS and Perp-Neg are shown to be effective for some text prompts, while not so consistent as ESD. Additionally, we find that ESD can work particularly well when combined with the time-prioritized scheduling proposed in DreamTime [14], as shown in Fig. 5. This means ESD is orthogonal to many other methods and can cooperate with them to further reduce Janus artifacts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 460, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 639 + ], + "type": "text", + "content": "Quantitative Comparison. With metrics proposed in Sec. 6, we numerically evaluate our method and baselines across 120 text prompts provided in [58]. We additionally involve Successful generation Rate (SR) based on human evaluation. The results are presented in Tab. 1. We observe that among all metrics, ESD reaches the best CLIP score, FID, and IG. More importantly, ESD achieves the optimal balance between view quality and diversity as shown by IQ and IV. Whereas, SDS suffers from low image quality with high IQ and VSD is limited by insufficient view variety with low IV. The superior IG of ESD indicates that views inside the generated scene are distinguishable rather than collapsing to be the same. We defer the breakdown table for numerical evaluation on examples in Fig. 4, human evaluation criteria, and the standard deviation of metrics to Appendix E." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Ablation Studies We conduct ablation studies on the choice of " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " (i.e. CFG weights) in Fig. 6. We demonstrate that " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " can adjust ESD's preference toward view- quality or diversity. When set to one, the produced Janus-free result albeit with fewer realistic details in the textures. Conversely," + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 209, + 356, + 257 + ], + "blocks": [ + { + "bbox": [ + 315, + 200, + 347, + 207 + ], + "lines": [ + { + "bbox": [ + 315, + 200, + 347, + 207 + ], + "spans": [ + { + "bbox": [ + 315, + 200, + 347, + 207 + ], + "type": "text", + "content": "ESD w/ CFG" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 209, + 356, + 257 + ], + "lines": [ + { + "bbox": [ + 307, + 209, + 356, + 257 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 356, + 257 + ], + "type": "image", + "image_path": "c1c5500abdd311042b5fd6bbeed3589c5562d03e2c38ef76c80578d5a2ae186c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 258, + 344, + 264 + ], + "lines": [ + { + "bbox": [ + 318, + 258, + 344, + 264 + ], + "spans": [ + { + "bbox": [ + 318, + 258, + 344, + 264 + ], + "type": "text", + "content": "Step " + }, + { + "bbox": [ + 318, + 258, + 344, + 264 + ], + "type": "inline_equation", + "content": "= 25\\mathrm{k}" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 271, + 547, + 316 + ], + "lines": [ + { + "bbox": [ + 306, + 271, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 271, + 547, + 316 + ], + "type": "text", + "content": "Figure 7. Ablation on Implementations. The successfully generated result is obtained via our suggested CFG trick while the diverged result is yielded by fitting the unconditioned score function in Eq. 8 via LoRA. Prompt: an elephant skull." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 360, + 209, + 499, + 257 + ], + "blocks": [ + { + "bbox": [ + 406, + 200, + 498, + 207 + ], + "lines": [ + { + "bbox": [ + 406, + 200, + 498, + 207 + ], + "spans": [ + { + "bbox": [ + 406, + 200, + 498, + 207 + ], + "type": "text", + "content": "ESD w/ fine-tuned unconditional LoRA" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 360, + 209, + 499, + 257 + ], + "lines": [ + { + "bbox": [ + 360, + 209, + 499, + 257 + ], + "spans": [ + { + "bbox": [ + 360, + 209, + 499, + 257 + ], + "type": "image", + "image_path": "fc7f3df07ccbf19d99e77b818a19c42c2ccb27970da651e31af9b96f5c7b26ce.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 258, + 397, + 265 + ], + "lines": [ + { + "bbox": [ + 372, + 258, + 397, + 265 + ], + "spans": [ + { + "bbox": [ + 372, + 258, + 397, + 265 + ], + "type": "text", + "content": "Step " + }, + { + "bbox": [ + 372, + 258, + 397, + 265 + ], + "type": "inline_equation", + "content": "= 18\\mathrm{k}" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 258, + 443, + 265 + ], + "lines": [ + { + "bbox": [ + 417, + 258, + 443, + 265 + ], + "spans": [ + { + "bbox": [ + 417, + 258, + 443, + 265 + ], + "type": "text", + "content": "Step " + }, + { + "bbox": [ + 417, + 258, + 443, + 265 + ], + "type": "inline_equation", + "content": "= 20\\mathrm{k}" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 462, + 257, + 488, + 264 + ], + "lines": [ + { + "bbox": [ + 462, + 257, + 488, + 264 + ], + "spans": [ + { + "bbox": [ + 462, + 257, + 488, + 264 + ], + "type": "text", + "content": "Step " + }, + { + "bbox": [ + 462, + 257, + 488, + 264 + ], + "type": "inline_equation", + "content": "= 22\\mathrm{k}" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 503, + 209, + 544, + 257 + ], + "blocks": [ + { + "bbox": [ + 503, + 209, + 544, + 257 + ], + "lines": [ + { + "bbox": [ + 503, + 209, + 544, + 257 + ], + "spans": [ + { + "bbox": [ + 503, + 209, + 544, + 257 + ], + "type": "image", + "image_path": "9226588b631f302947307531af3bccd832931adc667f129e3907da71c64f7363.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 508, + 258, + 534, + 264 + ], + "lines": [ + { + "bbox": [ + 508, + 258, + 534, + 264 + ], + "spans": [ + { + "bbox": [ + 508, + 258, + 534, + 264 + ], + "type": "text", + "content": "Step " + }, + { + "bbox": [ + 508, + 258, + 534, + 264 + ], + "type": "inline_equation", + "content": "= 25\\mathrm{k}" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 327, + 546, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 546, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 546, + 434 + ], + "type": "text", + "content": "when set to zero, ESD equates to VSD, and the Janus problem emerges again. We empirically find that choosing " + }, + { + "bbox": [ + 304, + 327, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 327, + 546, + 434 + ], + "type": "text", + "content": " around 0.5 yields the best result, balancing fine-grained textures and well-constructed geometry. We also implement ESD by directly fitting the score function " + }, + { + "bbox": [ + 304, + 327, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\nabla \\log q_t^\\theta (\\pmb{x}_t|\\pmb{y})" + }, + { + "bbox": [ + 304, + 327, + 546, + 434 + ], + "type": "text", + "content": " without camera pose conditioning to validate the suggested implementation by CFG trick. We show in Fig. 7 that this optimization scheme is unstable. As training proceeds, the gradient explodes, and the optimized texture overflows." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 445, + 378, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 378, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 378, + 456 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 464, + 546, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 546, + 584 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 546, + 584 + ], + "type": "text", + "content": "In this paper, we reveal that existing score distillation methods degenerate to maximal likelihood seeking on each view independently, leading to the mode collapse problem. We identify that re-establishing the entropy term in the variational objective brings a new update rule, called Entropic Score Distillation (ESD), which is theoretically equivalent to adopting classifier-free guidance trick upon variational score distillation. ESD maximizes the entropy of the rendered image distribution, encouraging diversity across views and mitigating the Janus problem." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 599, + 388, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 388, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 388, + 610 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "P Wang is sincerely grateful for constructive feedback regarding this manuscript from Zhaoyang Lv, Xiaoyu Xiang, Amit Kumar, Jinhui Xiong, and Varun Nagaraja. P Wang also thanks Ruisi Cai for providing decent visual materials for illustration purposes. Any statements, opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of their employers or the supporting entities." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9044" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 133 + ], + "type": "text", + "content": "[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Mohammadreza Armandpour, Huangjie Zheng, Ali Sadeghian, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968, 2023. 2, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 258 + ], + "type": "text", + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 259, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 259, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 259, + 288, + 304 + ], + "type": "text", + "content": "[4] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.* arXiv preprint arXiv:2303.13873, 2023. 1, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 304, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 288, + 350 + ], + "type": "text", + "content": "[5] Yiwen Chen, Chi Zhang, Xiaofeng Yang, Zhongang Cai, Gang Yu, Lei Yang, and Guosheng Lin. It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473, 2023. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 288, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 288, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 288, + 395 + ], + "type": "text", + "content": "[6] Giannis Daras, Kulin Shah, Yuval Dagan, Aravind Gollakota, Alex Dimakis, and Adam Klivans. Ambient diffusion: Learning clean distributions from corrupted data. Advances in Neural Information Processing Systems, 36, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 396, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 288, + 441 + ], + "type": "text", + "content": "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 441, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 288, + 464 + ], + "type": "text", + "content": "[8] Amir Hertz, Kfir Aberman, and Daniel Cohen-Or. Delta denoising score. arXiv preprint arXiv:2304.07090, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 465, + 288, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 288, + 518 + ], + "type": "text", + "content": "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 520, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 543 + ], + "type": "text", + "content": "[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "text", + "content": "[12] Susung Hong, Donghoon Ahn, and Seungryong Kim. Debi-asing scores and prompts of 2d diffusion for robust text-to-3d generation. arXiv preprint arXiv:2303.15413, 2023. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[14] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[15] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 108, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 547, + 162 + ], + "type": "text", + "content": "[16] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 164, + 547, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 547, + 209 + ], + "type": "text", + "content": "[17] Ajay Jain, Amber Xie, and Pieter Abbeel. Vectorfusion: Text-to-svq by abstracting pixel-based diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1911-1920, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 211, + 547, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 211, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 308, + 211, + 547, + 266 + ], + "type": "text", + "content": "[18] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 268, + 547, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 268, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 308, + 268, + 547, + 334 + ], + "type": "text", + "content": "[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 1, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 335, + 547, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 335, + 547, + 379 + ], + "spans": [ + { + "bbox": [ + 308, + 335, + 547, + 379 + ], + "type": "text", + "content": "[20] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023.5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 381, + 547, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 381, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 381, + 547, + 426 + ], + "type": "text", + "content": "[21] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 428, + 547, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 428, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 308, + 428, + 547, + 460 + ], + "type": "text", + "content": "[22] Ziming Liu, Di Luo, Yilun Xu, Tommi Jaakkola, and Max Tegmark. Genphys: From physical processes to generative models. arXiv preprint arXiv:2304.02637, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 462, + 547, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 462, + 547, + 517 + ], + "spans": [ + { + "bbox": [ + 308, + 462, + 547, + 517 + ], + "type": "text", + "content": "[23] Jonathan Lorraine, Kevin Xie, Xiaohui Zeng, Chen-Hsuan Lin, Towaki Takikawa, Nicholas Sharp, Tsung-Yi Lin, Ming-Yu Liu, Sanja Fidler, and James Lucas. Att3d: Amortized text-to-3d object synthesis. arXiv preprint arXiv:2306.07349, 2023.8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 520, + 547, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 520, + 547, + 563 + ], + "spans": [ + { + "bbox": [ + 308, + 520, + 547, + 563 + ], + "type": "text", + "content": "[24] Morteza Mardani, Jiaming Song, Jan Kautz, and Arash Vahdat. A variational perspective on solving inverse problems with diffusion models. arXiv preprint arXiv:2305.04391, 2023.6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 566, + 547, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 566, + 547, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 566, + 547, + 599 + ], + "type": "text", + "content": "[25] Luke Metz, Ben Poole, David Pfau, and Jascha Sohl-Dickstein. Unrolled generative adversarial networks. arXiv preprint arXiv:1611.02163, 2016. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 601, + 547, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 547, + 656 + ], + "type": "text", + "content": "[26] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 547, + 713 + ], + "type": "text", + "content": "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9045" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 115 + ], + "type": "text", + "content": "[28] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022.6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "type": "text", + "content": "[29] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "type": "text", + "content": "[30] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 249 + ], + "type": "text", + "content": "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 288, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 288, + 303 + ], + "type": "text", + "content": "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 288, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 288, + 347 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 288, + 347 + ], + "type": "text", + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 288, + 392 + ], + "type": "text", + "content": "[34] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 394, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 288, + 470 + ], + "type": "text", + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "text", + "content": "[36] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 2, 4, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "type": "text", + "content": "[37] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "type": "text", + "content": "[38] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Jaehoon Ko, Hyeonsu Kim, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 658 + ], + "type": "text", + "content": "[39] Ruizhi Shao, Jingxiang Sun, Cheng Peng, Zerong Zheng, Boyao Zhou, Hongwen Zhang, and Yebin Liu. Control4d: Dynamic portrait editing by learning 4d gan from 2d diffusion-based editor. arXiv preprint arXiv:2305.20082, 2023. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[40] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 72, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 547, + 106 + ], + "type": "text", + "content": "[41] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 161 + ], + "type": "text", + "content": "[42] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "text", + "content": "[43] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023.6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 218, + 547, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 218, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 218, + 547, + 271 + ], + "type": "text", + "content": "[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 274, + 547, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 547, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 547, + 306 + ], + "type": "text", + "content": "[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 308, + 546, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 546, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 546, + 342 + ], + "type": "text", + "content": "[46] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 343, + 546, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 546, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 546, + 386 + ], + "type": "text", + "content": "[47] Yang Song, Sahaj Garg, Jiaxin Shi, and Stefano Ermon. Sliced score matching: A scalable approach to density and score estimation. In Uncertainty in Artificial Intelligence, pages 574-584. PMLR, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 388, + 547, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 547, + 431 + ], + "type": "text", + "content": "[48] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 433, + 547, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 547, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 547, + 476 + ], + "type": "text", + "content": "[49] Akash Srivastava, Lazar Valkov, Chris Russell, Michael U Gutmann, and Charles Sutton. Veegan: Reducing mode collapse in gans using implicit variational learning. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 478, + 547, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 547, + 520 + ], + "type": "text", + "content": "[50] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439, 2023. 1, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 522, + 546, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 546, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 546, + 543 + ], + "type": "text", + "content": "[51] Cédric Villani et al. Optimal transport: old and new. Springer, 2009. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 545, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 547, + 578 + ], + "type": "text", + "content": "[52] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 579, + 547, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 547, + 634 + ], + "type": "text", + "content": "[53] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 635, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 547, + 689 + ], + "type": "text", + "content": "[54] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 1, 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "text", + "content": "[55] Peihao Wang, Zhiwen Fan, Dejia Xu, Dilin Wang, Sreyas Mohan, Forrest Iandola, Rakesh Ranjan, Yilei Li, Qiang Liu," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9046" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 340 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Zhangyang Wang, et al. Steindreamer: Variance reduction for text-to-3d score distillation via stein identity. arXiv preprint arXiv:2401.00604, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 160 + ], + "type": "text", + "content": "[56] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "text", + "content": "[57] Zhendong Wang, Yifan Jiang, Huangjie Zheng, Peihao Wang, Pengcheng He, Zhangyang Wang, Weizhu Chen, Mingyuan Zhou, et al. Patch diffusion: Faster and more data-efficient training of diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 288, + 262 + ], + "type": "text", + "content": "[58] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. arXiv preprint arXiv:2401.04092, 2024. 8, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 306 + ], + "type": "text", + "content": "[59] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360 \\{\\neg\\} views. arXiv preprint arXiv:2211.16431, 2022. 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 288, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 288, + 340 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 288, + 340 + ], + "type": "text", + "content": "[60] Joseph Zhu and Peiye Zhuang. Hifa: High-fidelity text-to-3d with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 7" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9047" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Taming Self-Training for Open-Vocabulary Object Detection/08e6cd2d-6b02-4147-b946-b39f423452c4_content_list.json b/2024/Taming Self-Training for Open-Vocabulary Object Detection/08e6cd2d-6b02-4147-b946-b39f423452c4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..21ed5c12ca2d1a38de81f6a730a35c335e9396e6 --- /dev/null +++ b/2024/Taming Self-Training for Open-Vocabulary Object Detection/08e6cd2d-6b02-4147-b946-b39f423452c4_content_list.json @@ -0,0 +1,1340 @@ +[ + { + "type": "text", + "text": "Taming Self-Training for Open-Vocabulary Object Detection", + "text_level": 1, + "bbox": [ + 179, + 130, + 790, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shiyu Zhao $^{1,*}$ Samuel Schulter $^{2}$ Long Zhao $^{3}$ Zhixing Zhang $^{1}$ Vijay Kumar B G $^{2}$ Yumin Suh $^{2}$ Manmohan Chandraker $^{2,4}$ Dimitris N. Metaxas $^{1}$", + "bbox": [ + 143, + 179, + 848, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Rutgers University $^{2}$ NEC Laboratories America $^{3}$ Google Research $^{4}$ UC San Diego", + "bbox": [ + 104, + 232, + 859, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent studies have shown promising performance in open-vocabulary object detection (OVD) by utilizing pseudo labels (PLs) from pretrained vision and language models (VLMs). However, teacher-student self-training, a powerful and widely used paradigm to leverage PLs, is rarely explored for OVD. This work identifies two challenges of using self-training in OVD: noisy PLs from VLMs and frequent distribution changes of PLs. To address these challenges, we propose SAS-Det that tames self-training for OVD from two key perspectives. First, we present a split-and-fusion (SAF) head that splits a standard detection into an open-branch and a closed-branch. This design can reduce noisy supervision from pseudo boxes. Moreover, the two branches learn complementary knowledge from different training data, significantly enhancing performance when fused together. Second, in our view, unlike in closed-set tasks, the PL distributions in OVD are solely determined by the teacher model. We introduce a periodic update strategy to decrease the number of updates to the teacher, thereby decreasing the frequency of changes in PL distributions, which stabilizes the training process. Extensive experiments demonstrate SAS-Det is both efficient and effective. SAS-Det outperforms recent models of the same scale by a clear margin and achieves 37.4 AP50 and 29.1 APr on novel categories of the COCO and LVIS benchmarks, respectively. Code is available at https://github.com/xiaofeng94/SAS-Det.", + "bbox": [ + 75, + 318, + 473, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 752, + 209, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional closed-set object detectors [4, 13, 32] are restricted to detecting objects with a limited number of categories. Increasing the size of detection vocabularies usually requires heavy human labor to collect annotated data. With the recent advent of strong vision and language models (VLMs) [20, 30], open-vocabulary object detection", + "bbox": [ + 75, + 777, + 468, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(OVD) [11] provides an alternative direction to approach this challenge. Typically, OVD detectors are trained with annotations of base categories and expected to generalize to novel categories with the power of pretrained VLMs.", + "bbox": [ + 500, + 287, + 890, + 348 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One promising thread of recent studies for OVD [8, 9, 42, 49] leverages VLMs to obtain pseudo labels (PLs) beyond base categories. But they rarely explore self-training, a powerful and widely used schema for utilizing PLs in closed-set tasks [37, 38, 44, 45]. We investigate this and find the vanilla self-training approach does not improve OVD performance due to the following challenges.", + "bbox": [ + 496, + 351, + 892, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "First, the typical self-training in closed-set tasks sets a confidence threshold to remove noisy PLs based on the fact that the quality of PLs is positively correlated to their confidences. However, VLMs employed in OVD are pretrained for image-level alignment with texts instead of instance-level object detection that requires the localization ability. Thus, the confidence score from pretrained VLMs is usually not a good indicator for the quality of box locations (i.e., pseudo boxes) provided by PLs. For example, prior studies [11, 49] show that CLIP [30] tends to output imperfect object boxes as predictions with high confidence. Recent methods for OVD [9, 49] just apply thresholding to VLMs' confidence scores and ignore the poor quality of pseudo boxes, which provides noisy supervision to the model. This issue becomes even worse when self-training is applied directly, since the noise accumulates which degrades the performance on novel categories. Moreover, these methods handle noisy PLs in the same way as ground truth of base categories during training, which further decreases the performance on base categories [9, 23].", + "bbox": [ + 496, + 459, + 892, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Second, self-training for closed-set object detection [19, 38, 45] usually follows an online teacher-student manner. In each training iteration, the teacher generates PLs, and the student is trained with a mixture of ground truth and PLs. Then, the teacher is updated by the student with exponential moving average (EMA). However, we find such EMA updates degrade OVD models (see Table 4). Our hypothesis is that, unlike closed-set tasks, OVD provides no ground truth for target categories, and thus, the supervision for target cat", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Part of this work was done during an internship at NEC Laboratories America. Correspondence to: Shiyu Zhao sz553@rutgers.edu", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13938", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "egories is fully decided by the distribution of PLs predicted by the teacher. Hence, the EMA updates change the distribution of PLs in each iteration, unstabilizing the training.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose Self-training And Split-and-fusion head for open-vocabulary Detection (SAS-Det) to tame self-training for OVD. First, we present a split-and-fusion (SAF) head to handle the noise of PLs. The SAF head splits the standard detection head into two branches: the closed-branch and the open-branch, which are fused at inference. The closed-branch, akin to the standard detection head, comprises a classification module and a box refinement module. It is supervised solely by ground truth from base categories, mitigating the impact of noisy PLs on the performance of base categories. The open-branch is a classification module supervised by class labels of both ground truth and PLs. It acquires complementary knowledge to the closed-branch and can significantly boost the performance when fused with the closed-branch. Moreover, this design circumvents noisy locations of pseudo boxes, reducing the accumulation of noise during self-training.", + "bbox": [ + 73, + 137, + 468, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, instead of adopting the vanilla EMA update, we reduce the number of the updates and periodically update the teacher by the student. The quality of our PLs improves along with the periodic updates, and our final PLs are better than those of prior PL-based methods [9, 49] that introduce external handcrafted steps. Fig. 1 shows the key differences.", + "bbox": [ + 75, + 393, + 468, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed SAS-Det outperforms recent OVD models of the same scale by a clear margin on two popular benchmarks, i.e., COCO and LVIS. Without extra handcrafted steps, our pseudo labeling is more efficient than prior methods, i.e., nearly 4 times faster than PB-OVD [9] and 3 times faster than VL-PLM [49]. Extensive ablation studies demonstrate the effectiveness of the proposed components. On COCO, the thresholding of vanilla self-training decreases the performance of novel categories by 3.6 AP, and the EMA update decreases the performance by 6.9 AP. Instead, SAS-Det eliminates the degradation with two separate detection heads and the periodic update. The fusion of the SAF head boosts the performance by 6.0 AP.", + "bbox": [ + 73, + 484, + 468, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this work are summarized as follows. (1) We show two challenges of applying self-training to OVD and propose two simple but effective solutions, i.e., using different detection heads to mitigate the noise in PLs, and using periodic updates to reduce frequency of changes in PLs' distributions. (2) The proposed SAF head for OVD handles the noisy boxes of PLs and enables fusion to improve the performance. (3) We present the leading performance on COCO and LVIS under widely used OVD settings and provide detailed analysis of the proposed SAS-Det.", + "bbox": [ + 75, + 680, + 468, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 844, + 218, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-language models (VLMs). VLMs are trained to learn the alignment between images and text in a com", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mon embedding space. CLIP [30] and ALIGN [14] use contrastive losses to learn such alignment on large-scale noisy image-text pairs from the Internet. ALBEF [20] introduces multi-modal fusion and additional self-supervised objectives. SIMLA [16] employs a single stream architecture to achieve multi-level image-text alignment. FDT [6] learns shared discrete tokens as the embedding space. These VLMs achieve impressive zero-shot performance on image classification. But due to the gap between the pretraining and detection tasks, VLMs have limited abilities in object detection. In this work, we attempt to close the gap via PLs. There are studies [15, 21, 34] focusing on aligning any text phrases with objects. But they require visual grounding data that are more expensive than detection annotations. Our work scales up the vocabulary size for object detection without requiring such costly data.", + "bbox": [ + 496, + 90, + 890, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-vocabulary object detection (OVD). Zero-shot object detection methods [3, 31, 46, 52] increase the vocabulary size but with limited accuracy. Motivated by the strong zero-shot abilities of VLMs, recent efforts focus on OVD. Finetuning-based methods [17, 18, 28, 48, 50] add detection heads onto pretrained VLMs and then finetune the detector with concepts of base categories. Such methods are simple but may forget the knowledge learned in the pretraining [2]. Distillation-based methods [7, 11, 40, 41] introduce additional distillation loss functions that force the output of a detector to be close to that of a VLM and thus avoid forgetting. However, since the distillation losses are not designed for the detection task, they may conflict with detection objectives due to gradient conflicts that are a common issue for multi-task models. Methods like Feng et al. [8], Gao et al. [9], Wu et al. [42], Zhao et al. [49] create pseudo labels (PL) of novel concepts as supervision, and do not require extra losses, which sidesteps both catastrophic forgetting and gradient conflicts. But these methods need handcrafted steps to generate high quality PLs, e.g. multiple runs of box regression [49], activation maps [9] from Grad-CAM [35], image retrieval [8], or multi-stage training [42]. In this work, we address the two challenges of using self-training for OVD and enable an efficient end-to-end pseudo labeling pipeline. Besides, we point out the noise due to the poor locations of PLs and introduce the SAF head to handle such noise.", + "bbox": [ + 496, + 337, + 890, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-training for object detection. Weakly-supervised object detection methods [27, 33, 38, 39] usually explore online self-training by distilling the knowledge from the model itself. Recent semi-supervised object detection methods [19, 26, 45] adopt a teacher-student design, where the teacher is an exponential moving average (EMA) of the student. Self-training has been widely explored in the above fields but rarely in OVD. Unlike semi-supervised object detection, OVD encounters two challenges to use self-training, i.e., more noisy PLs, and larger changes in PLs' distributions. This work proposes the SAF head to address", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13939", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0b926dd32fbc74393bc3209e36a96ad73775294503a7653fe6185d1ed49dc07a.jpg", + "image_caption": [ + "Figure 1. Left: Prior PL-based methods for OVD rely on handcrafted heuristics to leverage a frozen VLM for offline pseudo labels. This is usually inefficient and does not allow for improving PLs throughout training. Right: We customize self-training and finetune VLMs for OVD, which allows efficient on-the-fly computation of PLs that can be improved throughout training." + ], + "image_footnote": [], + "bbox": [ + 78, + 92, + 488, + 275 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6c4e1bc716bd1f5a42b1ef780b8a951b2cd24bb09b6560a4379b5f8c9161d600.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 93, + 890, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the noise, and adopts periodic updates to reduce the frequency of changes in PLs' distribution.", + "bbox": [ + 75, + 349, + 468, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 398, + 184, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In open-vocabulary detection, an object detector is trained with bounding boxes and class labels of base categories $\\mathcal{C}^B$ . At inference, the detector is used for detecting objects of open concepts including $\\mathcal{C}^B$ and novel categories $\\mathcal{C}^N$ , where $\\mathcal{C}^B \\cap \\mathcal{C}^N = \\emptyset$ . To make such detection practical, recent studies [8, 11, 23, 48] adopt extra data (e.g., image-text pairs and image-level tags) and/or external VLMs. We follow their settings and leverage the pretrained CLIP [30] to build our OVD detector.", + "bbox": [ + 75, + 424, + 470, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Adapting CLIP to OVD", + "text_level": 1, + "bbox": [ + 76, + 571, + 295, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce how to adapt the ResNet-based CLIP into a Faster R-CNN [32] (C4) detector. For simplicity, we do not use FPN [25], but it can be incorporated with learnable gating from Flamingo [1]. Similarly, ViT-based CLIP models can be adapted to detectors following [22].", + "bbox": [ + 75, + 597, + 468, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Region proposals from an external RPN. CLIP is pretrained via image-text alignment. As shown in Sect. 4.3, finetuning pretrained backbones to get region proposals decreases the performance, probably because such finetuning breaks the image-text alignment learned in the pretraining. To address this problem, F-VLM [18] freezes the pretrained backbone and only finetunes detection heads. But this solution limits the capacity of the detector. Unlike F-VLM, we follow the approach of Singh et al. [36], Zhong et al. [50] and employ an external RPN to generate region proposals, which is trained with ground truth boxes of base categories. Prior studies adopt RPN mainly to accelerate inference [36] or to improve region recognition [50]. By contrast, we aim to leverage the RPN to preserve the knowledge learned in the pretraining for better self-training.", + "bbox": [ + 75, + 674, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text embeddings as the classifier. For the $i$ -th region proposal from the external RPN, we apply RoIAlign [13] on the 4th feature maps of CLIP's ResNet to get the proposal features. Then, the features are fed to the last ResNet block and the attention pooling of CLIP to get the region embedding $r_i$ , which is later used for classification. Following prior studies [11, 23, 50], we convert a set of given concepts with prompt engineering into CLIP text embeddings, which act as classifiers, as shown in Fig. 2b. A fixed all-zero embedding is adopted for the background category. Assuming $t_c$ is the embedding of the $c$ -th category, the probability of $r_i$ to be classified as the $c$ -th category is,", + "bbox": [ + 496, + 349, + 890, + 532 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {i, c} = \\frac {\\exp \\left(\\langle r _ {i} , t _ {c} \\rangle / \\tau\\right)}{\\sum_ {j = 0} ^ {C - 1} \\exp \\left(\\langle r _ {i} , t _ {j} \\rangle / \\tau\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 542, + 890, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\langle \\cdot, \\cdot \\rangle$ denotes the cosine distance, $C$ denotes the vocabulary size, and $\\tau$ denotes the temperature. In our detector, $r_i$ may be further fed to a box refinement module to predict the box shift based on the region proposal. With the above adoptions, our initial detector gains zero-shot detection ability to some extent. Please refer to the supplement for quantitative evaluations.", + "bbox": [ + 496, + 592, + 890, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Taming Self-Training", + "text_level": 1, + "bbox": [ + 500, + 709, + 700, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although self-training has been widely explored in closed-set object detection, using it for OVD presents two challenges, i.e., noisy PLs and frequent changes of PLs' distributions. In the following, we describe our self-training pipeline and how to address the two challenges.", + "bbox": [ + 496, + 733, + 890, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-training pipeline. Fig. 2a illustrates the pipeline of our self-training with a teacher-student manner. Training images are first fed into the RPN to obtain region proposals. Then, the teacher model runs inference on those proposals, where the resulting predictions with confidence above a threshold are selected as PLs. The student model adopts", + "bbox": [ + 496, + 810, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13940", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5cc705d581966c09fa8f8f0efe3217590b1081c40432853b709efcf57da96eaa.jpg", + "image_caption": [ + "Figure 2. (a) Pipeline of our self-training. The teacher and the student are models of the same architecture. They are initialized with the same pretrained CLIP model. The teacher generates PLs that are used to train the student, and the student updates the teacher periodically. (b) Structure of our detector. The proposed SAF head is put on top of a CLIP image encoder. The open- and closed-branches take the text embeddings from a CLIP text encoder as classifier." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 576, + 321 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9e0e3fcd09dafb0c8586f19bd00496071e6818c8d807ca8caa442d1f7b226e05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 89, + 888, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the same region proposals, and is supervised with both base ground truth and PLs generated by the teacher. The teacher is periodically updated with the parameters of the student model.", + "bbox": [ + 75, + 417, + 468, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Handling noise in PLs. The noise of PLs from VLMs exists in both classification and localization. It is a common practice to filter PLs with classification confidence, but this only addresses noise in classification. To reduce the noise in locations of PLs, we exclude the noisy boxes of PLs from the training. That is, the classification loss is calculated on class labels of both ground truth and PLs, but the box regression loss is calculated only on ground truth boxes. Besides, to mitigate the impact of PLs on the performance of base categories, we propose SAF head and elaborate it in Sect. 3.3.", + "bbox": [ + 75, + 484, + 468, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reducing changes in PLs' distributions with periodic updates. The teacher model is updated with the student to enable self-training. The exponential moving average (EMA) is a widely used approach for object detection [19, 26, 45], which updates the teacher in every iteration. But we observed empirically that it does not benefit OVD (Table 4). We hypothesize that, unlike semi-supervised object detection, OVD has no ground truth for the target categories. Thus, the distribution of the target data (i.e., PLs) is fully determined by the teacher. The EMA update changes the PLs' distribution in each iteration and leads to a constantly shifting training target that has been shown hard to optimize (i.e., destabilizing the training) in deep Q-learning [29]. As a solution, we periodically update the teacher after a set number of iterations to maintain consistent distributions for PLs between updates. We call this", + "bbox": [ + 75, + 657, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "strategy as the periodic update and show it outperforms the EMA update by a large margin in Sect. 4.3.", + "bbox": [ + 498, + 417, + 888, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Split-and-Fusion (SAF) Head", + "text_level": 1, + "bbox": [ + 500, + 463, + 759, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our SAF head first splits a detection head into two branches, i.e. \"closed-branch\" and \"open-branch\", with the goal to better handle noisy PLs during training. At inference, predictions from both heads are fused to boost the performance.", + "bbox": [ + 498, + 488, + 890, + 563 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Splitting the detection head. The closed-branch follows a standard detection head with a classification module and a class-agnostic box refinement module. The former module classifies region proposals based on Eq. 1, and the latter one refines the proposal boxes for better locations. We train the closed-branch with boxes and class labels of ground truth for $\\mathcal{C}^B$ using standard detection losses, which include a cross entropy loss for classification and a box regression loss for localization. Since no PLs are used to train the closed-branch, the noise of PLs is unlike to impact its performance on $\\mathcal{C}^B$ . Moreover, as shown in Gu et al. [11], Zareian et al. [48], box regression modules trained on $\\mathcal{C}^B$ can generalize to novel categories that are unseen during training. Therefore, the closed-branch is able to provide generalized boxes, as well.", + "bbox": [ + 496, + 566, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The open-branch only contains a classification module. It is trained using only the cross-entropy loss with class labels from both $\\mathcal{C}^B$ and PLs, hence, learning broader concepts beyond $\\mathcal{C}^B$ . Unlike distillation losses [7, 11], all losses for the two branches are originally designed for detection and are unlikely to conflict with each other. When generating PLs, we use the classification scores from the", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13941", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "open-branch and bounding boxes from the closed-branch. Boxes of PLs are not directly used in our losses but are involved to select foreground proposal candidates for the classification loss.", + "bbox": [ + 75, + 90, + 467, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fusing complementary predictions. The open- and closed-branches are trained in different ways and learn complementary knowledge. Therefore, we fuse their predictions with the geometric mean at inference time. Specifically, assuming $p_{i,c}^{\\mathrm{open}}$ and $p_{i,c}^{\\mathrm{closed}}$ are prediction scores of the open- and closed-branches, respectively, the final score is calculated as", + "bbox": [ + 75, + 151, + 467, + 255 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {i, c} ^ {\\text {f u s e d}} = \\left\\{ \\begin{array}{l l} \\left(p _ {i, c} ^ {\\text {c l o s e d}}\\right) ^ {(1 - \\alpha)} \\cdot \\left(p _ {i, c} ^ {\\text {o p e n}}\\right) ^ {\\alpha}, & \\text {i f} i \\in \\mathcal {C} ^ {B} \\\\ \\left(p _ {i, c} ^ {\\text {c l o s e d}}\\right) ^ {\\alpha} \\cdot \\left(p _ {i, c} ^ {\\text {o p e n}}\\right) ^ {(1 - \\alpha)}, & \\text {i f} i \\in \\mathcal {C} ^ {N} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 263, + 467, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha \\in [0,1]$ balances the two branches. The indices $i, c$ are the same as in Eq. 1. We keep a list of known base categories $\\mathcal{C}^B$ and take other categories beyond the list as novel. It's important to note that fusion is a common strategy. For instance, F-VLM [18] employs score fusion to enhance the location quality of CLIP's predictions. In contrast, our approach focuses on fusing the complementary predictions from the two branches. The primary contribution of our method lies in learning what to fuse rather than the fusion process itself.", + "bbox": [ + 75, + 311, + 467, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 474, + 207, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 76, + 500, + 251, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on the two popular OVD benchmarks COCO-OVD and LVIS-OVD, based on COCO [24] and LVIS [12], respectively. (1) COCO-OVD: 65 out of 80 COCO categories are divided into 48 base categories and 17 novel categories. We report results on the validation set. (2) LVIS-OVD: 337 rare categories are used as the novel concepts and the remaining 866 categories (frequent and common) as the base concepts. We report results on the whole LVIS validation set.", + "bbox": [ + 75, + 522, + 467, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. Following prior studies [11, 41, 48, 50], we report AP at IoU threshold of 0.5 for COCO-OVD. $\\mathrm{AP}_{50}^{novel}$ , $\\mathrm{AP}_{50}^{base}$ and $\\mathrm{AP}_{50}^{all}$ are the metrics for the novel, the base and all (novel + base) categories, respectively. For LVIS-OVD, we report the mean AP averaged on IoUs from 0.5 to 0.95. $\\mathrm{AP}_r$ , $\\mathrm{AP}_c$ , $\\mathrm{AP}_f$ and AP are the metrics for rare, common, frequent, and all categories.", + "bbox": [ + 75, + 657, + 467, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We built our detector on Faster R-CNN with the CLIP version of ResNet (RN50-C4 or RN50x4-C4) as the backbone. Following prior studies [47, 49, 51], we assumed the novel concepts were available during training and took them as our open concepts, but no annotations of novel concepts are used. Our method was implemented on Detectron2 [43] and trained with 8 NVIDIA A6000 GPUs. We trained our models with the $1 \\times$ schedule (90k iterations) for COCO-OVD and the $2 \\times$ schedule for", + "bbox": [ + 75, + 763, + 467, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LVIS-OVD. The batch size was 16 with an initial learning rate of 0.002. The default data augmentation of Detectron2 was applied. Loss terms were equally weighted. By default, the teacher models were updated three times, usually together with the decreases of the learning rate.", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with the Existing Methods", + "text_level": 1, + "bbox": [ + 498, + 176, + 834, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "COCO-OVD. We compare our method with prior work on COCO in Table 1. With the same backbone and detector, SAS-Det outperforms the most recent method BARON [41] by $4.3\\mathrm{AP}_{50}^{novel}$ on novel categories and $3.7\\mathrm{AP}_{50}^{base}$ on base categories. Probably, BARON distills knowledge from VLMs, but distillation may lead to gradient conflicts and degrade the performance [49]. Compared to VLDet [23], SAS-Det gains an improvement of $7.9\\mathrm{AP}_{50}^{base}$ . It is likely that VLDet employs noisy pseudo labels to train the detection head, impacting the detection of base classes in both classification and localization. In contrast, the two-branch design of our SAF head allows to train the closed-branch with just ground-truth of base classes and reduces the impact of the noise from pseudo labels. For more analysis on where our improvements come, please refer to Sect. 4.3 and Table 3. The first block of Table 1 provides results for methods with stronger backbones or detector architectures. When compared with those methods, SAS-Det still achieves the leading performance. Those results clearly demonstrate the effectiveness of SAS-Det.", + "bbox": [ + 496, + 200, + 890, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LVIS-OVD. We provide the main results on LVIS in Table 2. When using ResNet50 as the backbone, SAS-Det achieves similar performance as the recent method DetPro [7]. DetPro proposes learnable prompts to generate better text embeddings as the classifier for OVD. Our method is orthogonal and can leverage DetPro's prompts for further improvement. The first block of Table 2 also shows that SAS-Det outperforms ViLD by a large margin on novel categories (indicated by $\\mathrm{AP}_r$ ) but gets lower $\\mathrm{AP}_f$ for base categories. This is also observed in the recent method BARON [41]. The performance gap is probably due to the training setup. ViLD is trained for 8 times more iterations and adopts a strong data augmentation method [10], both of which benefit detection on base categories. In contrast, BARON and our approach adopt shorter training and standard data augmentations. When replacing the pretrained CLIP ResNet50 with ResNet50x4 [30] as the backbone, we improve $\\mathrm{AP}_r$ by 8.2 from 20.9 to 29.1, which demonstrates SAS-Det scales up nicely with stronger pretrained VLM backbones.", + "bbox": [ + 496, + 503, + 890, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 816, + 663, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, the default baseline shares the same architecture and the training as our SAS-Det (with RN50-C4 as the backbone), except that the SAF head is replaced with a single detection head. The localization module of the detec", + "bbox": [ + 498, + 839, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13942", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/98b104303f5583abde06d92387f2bd32c50dd0da45ffd112f09e91a3cdd4ca61.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTraining SetupBackboneDetector\\(AP^{novel}\\)\\(AP^{base}\\)\\(AP^{all}\\)
ViLD [11]16×+LSJRN50-FPNFaster R-CNN27.659.551.2
VL-PLM [49]1×+DefaultRN50-FPNFaster R-CNN32.354.048.3
F-VLM [18]0.5×+LSJRN50-FPNFaster R-CNN28.0-39.6
OV-DETR [47](Not Given)RN50Deform. DETR29.461.052.7
CORA [42]3×+DefaultRN50DAB-DETR35.135.535.4
OVR-CNN [48](Not Given)RN50-C4Faster R-CNN22.846.039.9
RegionCLIP [50]1×+DefaultRN50-C4Faster R-CNN26.854.847.5
Detic [51]1×+DefaultRN50-C4Faster R-CNN27.851.145.0
PB-OVD [9]6×+DefaultRN50-C4Faster R-CNN30.846.142.1
VLDet [23]1×+LSJRN50-C4Faster R-CNN32.050.645.8
BARON [41]1×+DefaultRN50-C4Faster R-CNN33.154.849.1
SAS-Det (Ours)1×+DefaultRN50-C4Faster R-CNN37.458.553.0
", + "bbox": [ + 151, + 88, + 820, + 306 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6ac22963d90b65c019fa23bc3656f64e11893c09d4e8b7d0acb73846d3af22bd.jpg", + "table_caption": [ + "Table 1. Comparison with recent methods on COCO-OVD. We group methods into two blocks. The first block contains methods using stronger backbones or detector architectures than ours. The second block contains models of the same scale as ours. Training setup indicates training iterations $(\\mathrm{N}\\times)$ and data augmentations. Large Scale Jittering (LSJ) [10] is a stronger data augmentation than Detectron2's default." + ], + "table_footnote": [], + "table_body": "
MethodTraining SetupBackboneDetectorAPrAPcAPfAP
ViLD [11]16×+LSJRN50-FPNFaster R-CNN16.726.534.227.8
DetPro [7]2×+DefaultRN50-FPNFaster R-CNN20.827.832.428.4
F-VLM [18]9×+LSJRN50-FPNFaster R-CNN18.6--24.2
BARON [41]2×+DefaultRN50-C4Faster R-CNN17.325.631.026.3
SAS-Det (Ours)2×+DefaultRN50-C4Faster R-CNN20.926.131.627.4
ViLD [11]16×+LSJRN152-FPNFaster R-CNN19.827.134.528.7
OWL-ViT [28]12×+LSJViT-L/14OWL-ViT25.6--34.7
F-VLM [18]9×+LSJRN50x4-FPNFaster R-CNN26.3--28.5
CORA [42]3×+DefaultRN50x4DAB-DETR22.2---
SAS-Det (Ours)2×+DefaultRN50x4-C4Faster R-CNN29.132.436.833.5
", + "bbox": [ + 151, + 375, + 818, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison with recent methods on LVIS-OVD. We group methods based on the scale of backbones. Training setup contains training iterations $(\\mathrm{N}\\times)$ and data augmentations. LSJ [10] is a stronger data augmentation than Detectron2's default.", + "bbox": [ + 75, + 573, + 892, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion head is trained with base ground truth boxes only, but the classification module is trained with class labels of both ground truth and PLs. Thus, the baseline provides a naive solution to exclude noisy pseudo boxes from the training. All evaluations are conducted on COCO.", + "bbox": [ + 75, + 628, + 470, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "External RPN. We leverage an external RPN to generate region proposals so that finetuning focuses on region-text alignment that is similar to the pretraining task. In this way, the detector is unlikely to forget the knowledge obtained in the pretraining. To validate the effectiveness of the external RPN, we follow F-VLM [18] to train a detector without an external RPN (See the supplement for more details). Table 3 compares baseline with the detector in the row of (1). As shown, without the external RPN, the performance drops on both novel and base categories.", + "bbox": [ + 75, + 710, + 468, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Removing boxes of PLs from training. Our baseline handles location noise of PLs' boxes by directly excluding them", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "from the training. As shown in Table 3, baseline outperforms (2)'s detector. The only difference between them is the removal of pseudo boxes from training. This result clearly shows that simply removing the pseudo boxes is an effective way to deal with location noise.", + "bbox": [ + 496, + 628, + 893, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Splitting the detection head. As shown in Table 3, compared to baseline, the open-branch of the proposed SAF head in (5) achieves similar performance on novel categories. This indicates that the split handles the noise in PLs' location as well as the naive solution of baseline. This is plausible because both solutions exclude the noisy boxes of PLs from training. Additionally, the open-branch gains 1.7 $\\mathrm{AP}_{50}^{base}$ on base categories. Compared to (3)'s, the closed-branch in (4) gain improvements in terms of both $\\mathrm{AP}_{50}^{novel}$ and $\\mathrm{AP}_{50}^{base}$ . Note that (4)'s and (3)'s heads are trained with the same data. Based on those results, the split benefits OVD beyond handling the location noise. One advantage", + "bbox": [ + 496, + 719, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13943", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a28e0195b712f8c587344d3d09bda332198e8d7f63c25dc933ca7e49390e52e2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Ablation\\(AP_{50}^{novel}\\)\\(AP_{50}^{base}\\)\\(AP_{50}^{all}\\)
Baseline31.455.749.4
(1) No external RPN, train the backbone for region proposals(-6.0) 25.453.446.1
(2) Noisy boxes of PLs as supervision for box regression(-3.6) 27.855.248.0
(3) No PLs, train with base data only(-8.2) 23.256.948.1
(4) W/ SAF head, use closed-branch's predictions (\\(p_{i,c}^{closed}\\) in Eq. 2)(-5.8) 25.657.949.5
(5) W/ SAF head, use open-branch's predictions (\\(p_{i,c}^{open}\\) in Eq. 2)(+0.5) 31.957.450.7
(6) W/ SAF head, use fused predictions (\\(p_{i,c}^{fused}\\) in Eq. 2)(+6.0) 37.458.553.0
", + "bbox": [ + 147, + 88, + 823, + 227 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9ced908a91ec468f26453f288d32a4ce1b18b8c118ca1d792e92eba2eb392555.jpg", + "table_caption": [ + "Table 3. Ablation studies to analyze the effect of each component of SAS-Det on COCO-OVD. Results (4), (5), and (6) denote different outputs from the same model." + ], + "table_footnote": [], + "table_body": "
Update strategy\\( AP_{50}^{novel} \\)\\( AP_{50}^{base} \\)\\( AP_{50}^{all} \\)
Baseline (w/ periodic update)31.455.749.4
(7) No teacher and no pseudo labels(-8.2) 23.256.948.1
(8) Not update the teacher(-1.8) 29.655.949.1
(9) Take the EMA of the student as the teacher(-6.9) 24.553.946.2
(10) Replace the teacher with the student every iteration(-23.4) 8.053.341.5
", + "bbox": [ + 184, + 279, + 785, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparison of different strategies for updating the teacher in self-training on COCO-OVD.", + "bbox": [ + 187, + 398, + 781, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "is that, due to the split, the closed-branch is less likely to be influenced by the noise of PLs and learn how to better localize objects.", + "bbox": [ + 75, + 439, + 468, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fusing predictions. Our SAF head fuses predictions of the open-branch and the closed-branch at inference time. As shown in Table 3, (6)'s outperforms all the others by a large margin with the prediction fusion. Note that (4)'s, (5)'s and (6)'s refer to the different outputs of the same model. The improvement can be attributed to the fact that the two branches are trained on different sets of data and learn complementary knowledge. The closed-branch is trained with class labels and boxes of base categories. The open-branch is trained with class labels of more vocabularies including base and novel classes. Thus, the former learns more about how to localize objects and how to detect base classes. The latter learns more about how to detect new objects, which complements the former.", + "bbox": [ + 75, + 488, + 468, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Update strategies for the teacher model. The teacher is updated with the student model to improve PLs during finetuning. We evaluated several strategies for updating the teacher model in Table 4 and summarize our findings as follows. First, the EMA update in (9), which is shown effective and widely used in semi-supervised object detection [19, 26, 45], is just slightly better than training without PLs in (7). It is worse than no update to the teacher (i.e. no self-training) in (8). Second, if we replace the teacher with the student every iteration as (10), $\\mathrm{AP}_{50}^{novel}$ drops to 8.0 that is much lower than that number in (7). These findings indicate that it's harmful to change the teacher frequently. We hypothesize that frequent updates change the distribution of", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PLs too often and makes the training unstable. Last but not least, by reducing the number of updates, our periodic update significantly outperforms the EMA update in OVD.", + "bbox": [ + 496, + 439, + 890, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The number of updates to the teacher model. We trained our detectors with different numbers of updates to teacher models. Please see the supplement for how we distribute the updates. As shown in Table 5, too many updates, e.g., 8 or 4 updates, lead to performance drops mainly due to the following. First, similar as the aforementioned EMA update, too many updates change the distribution of PLs too often and make the training unstable. Second, the more updates, the earlier an update happens. However, the student model is not well trained at the early stage of the training and thus is not good enough to update the teacher. Table 5 shows that 2 and 3 updates achieve similar performance. But we set 3 updates as default to include as many updates as possible.", + "bbox": [ + 496, + 484, + 892, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Further Analysis", + "text_level": 1, + "bbox": [ + 500, + 694, + 666, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on the retained 15 COCO categories. Following prior studies [47, 49, 51], SAS-Det assumes novel concepts are available during training and achieves good performance. A natural question is if SAS-Det generalizes to alien concepts that are unknown before the evaluation. To answer the question, we follow the evaluation protocol of MEDet [5] where an OVD detector is evaluated on the whole COCO validation set with all 80 COCO concepts. $\\mathrm{AP}_{50}^{retain}$ , $\\mathrm{AP}_{50}^{novel}$ and $\\mathrm{AP}_{50}^{coco}$ are reported as APs averaged on 15 retained, 17 novel, and all 80 COCO categories, respectively. As shown in Table 6, SAS-Det achieves leading performance by a clear margin, which indicates the good", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13944", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cea351ccf90b526b49b4eced7c1fa065342cce78457928ab399be8b4cbfb3e63.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
# Update\\(AP_{50}^{novel}\\)
827.6
430.6
3 (baseline)31.4
231.6
130.9
0 (No update)29.6
", + "bbox": [ + 168, + 89, + 356, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/741c4188c8a53a2cb7c3200ff4d285cdc73d68873475c97539e068f05db18e2f.jpg", + "image_caption": [ + "Figure 3. Quality of PLs during training." + ], + "image_footnote": [], + "bbox": [ + 102, + 270, + 454, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "generalization capability of the proposed method.", + "bbox": [ + 75, + 455, + 403, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PL's quality during self-training. We compare our PLs and the prior state-of-the-art PLs from VL-PLM [49] on the COCO validation set in terms of PLs' quality. Following VL-PLM, we take $\\mathrm{AP}_{50}^{novel}$ as the metric and evaluate our teacher models during training. As illustrated in Fig. 3, the quality of our initial PLs is not as good as VL-PLM's. But ours get close to VL-PLM's after two updates and become the better by the third update. This is because the teacher is initialized with CLIP, which is not pretrained for detection, and thus cannot provide high quality PLs. With updates, the teacher is equipped with the knowledge about detection and generates promising PLs.", + "bbox": [ + 75, + 470, + 468, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Time cost of pseudo labeling. We run two major PL-based methods [9, 49] on our computational environment and compare them with our pseudo labeling in terms of the time cost in Table 7.", + "bbox": [ + 75, + 652, + 468, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reducing the time cost of external RPN. There is a simple solution to remove the extra computational cost of external RPN adopted by SAS-Det. We first generate pseudo labels with SAS-Det, and then follow VL-PLM [49] to train a Faster R-CNN based on ResNet50 for OVD without the external RPN (\"Faster R-CNN + PLs\"). As shown in Table 8, \"Faster R-CNN + PLs\" achieves slightly better or close performance as SAS-Det.", + "bbox": [ + 75, + 712, + 468, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 844, + 194, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we highlight two challenges associated with applying self-training to OVD: 1) noisy PLs from pretrained", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/21e7c2c69201978f23bec2b91ba02aa5a26236985956650644d55fbbe3a8e82f.jpg", + "table_caption": [ + "Table 5. Performance with varied numbers of updates to the teacher model." + ], + "table_footnote": [], + "table_body": "
MethodAPretain50APnovel50APcoco50
OVR-CNN [48]11.522.938.1
Detic-80 [51]11.527.338.3
MEDet [5]18.632.642.4
Ours23.037.147.2
", + "bbox": [ + 514, + 89, + 859, + 181 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ac657bb1dbbde35875466f8cce8e021f3432dd9ac4bd75e19dd78df55e8dd902.jpg", + "table_caption": [ + "Table 6. Generalization capability of OVD methods on retained 15 COCO categories. Text embeddings of 80 COCO categories are used as the classifier. Numbers are from MEDet [5]." + ], + "table_footnote": [], + "table_body": "
MethodTime (s)APnovel50
PB-OVD [9]0.484818.7
VL-PLM [49]0.445625.5
Ours0.130826.7
", + "bbox": [ + 557, + 265, + 823, + 343 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1bf6c72ec183987a499d721b569c02d8157e4fdf2e67d69d58d78fd836b5e478.jpg", + "table_caption": [ + "Table 7. Mean time cost to get PLs per image. The quality of PLs on the validation set of COCO-OVD are provided for reference. We report the quality of our PLs after the last update." + ], + "table_footnote": [], + "table_body": "
Method\\( AP_{50}^{novel} \\)\\( AP_{50}^{base} \\)\\( AP_{50}^{all} \\)
SAS-Det37.458.553.0
Faster R-CNN + PLs38.158.353.0
", + "bbox": [ + 517, + 452, + 875, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Training Faster R-CNN with PLs from SAS-Det on COCO-OVD", + "bbox": [ + 498, + 525, + 890, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "VLMs and 2) frequent changes of PLs' distributions. To address the challenges, we introduce a split-and-fusion (SAF) head and implement periodic updates. The SAF head splits a standard detection head into an open-branch and a closed-branch. The open-branch is trained with the class labels of ground truth and PLs. The closed-branch is trained with both class labels and boxes of ground truth. This approach effectively eliminates the location noise of PLs during training. Furthermore, the SAF head acquires complementary knowledge from different sets of training data, enabling fusion to enhance performance. The periodic updates reduce the number of updates to the teacher models, thereby decreasing the frequency of changes in PLs' distributions. We demonstrate that the proposed method SAS-Det is both efficient and effective. Our pseudo labeling is much faster than prior PL-based methods [9, 49]. SAS-Det also outperforms recent models on both COCO and LVIS for OVD.", + "bbox": [ + 496, + 582, + 890, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: This research project has been partially funded by research grants to Dimitris N. Metaxas through NSF: 2310966, 2235405, 2212301, 2003874, and FA9550-23-1-0417.", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13945", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 3", + "[2] Relja Arandjelović, Alex Andonian, Arthur Mensch, Olivier J Henaff, Jean-Baptiste Alayrac, and Andrew Zisserman. Three ways to improve feature alignment for open vocabulary detection, 2022. 2", + "[3] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In ECCV, pages 384-400, 2018. 2", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-End Object Detection with Transformers. In ECCV, 2020. 1", + "[5] Peixian Chen, Kekai Sheng, Mengdan Zhang, Yunhang Shen, Ke Li, and Chunhua Shen. Open vocabulary object detection with proposal mining and prediction equalization, 2022. 7, 8", + "[6] Yuxiao Chen, Jianbo Yuan, Yu Tian, Shijie Geng, Xinyu Li, Ding Zhou, Dimitris N Metaxas, and Hongxia Yang. Revisiting multimodal representation in contrastive learning: from patch and token embeddings to finite discrete tokens. In CVPR, 2023. 2", + "[7] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In CVPR, pages 14084-14093, 2022. 2, 4, 5, 6", + "[8] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In ECCV, pages 701-717. Springer, 2022. 1, 2, 3", + "[9] Mingfei Gao, Chen Xing, Juan Carlos Niebles, Junnan Li, Ran Xu, Wenhao Liu, and Caiming Xiong. Open vocabulary object detection with pseudo bounding-box labels. In ECCV, pages 266-282. Springer, 2022. 1, 2, 6, 8", + "[10] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, pages 2918-2928, 2021. 5, 6", + "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary Object Detection via Vision and Language Knowledge Distillation. In ICLR, 2022. 1, 2, 3, 4, 5, 6", + "[12] Agrim Gupta, Piotr Dólar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 5", + "[13] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 1, 3", + "[14] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. In ICML, 2021. 2", + "[15] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR - Modulated detection for end-to-end multi-modal understanding." + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 2", + "[16] Zaid Khan, Vijay Kumar B.G., Xiang Yu, Samuel Schulter, Manmohan Chandraker, and Yun Fu. Single-stream multilevel alignment for vision-language pretraining. In ECCV, pages 735-751. Springer, 2022. 2", + "[17] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Region-aware pretraining for open-vocabulary object detection with vision transformers. In CVPR, pages 11144-11154, 2023. 2", + "[18] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. In ICLR, 2023. 2, 3, 5, 6", + "[19] Gang Li, Xiang Li, Yujie Wang, Yichao Wu, Ding Liang, and Shanshan Zhang. Pseudo labeling and consistency training for semi-supervised object detection. In ECCV, pages 457-472. Springer, 2022. 1, 2, 4, 7", + "[20] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 1, 2", + "[21] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, pages 10965-10975, 2022. 2", + "[22] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, pages 280-296. Springer, 2022. 3", + "[23] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In ICLR, 2023. 1, 3, 5, 6", + "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 5", + "[25] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature Pyramid Networks for Object Detection. In CVPR, 2017. 3", + "[26] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In ICLR, 2021. 2, 4, 7", + "[27] Zhixiang Min, Bingbing Zhuang, Samuel Schulter, Buyu Liu, Enrique Dunn, and Manmohan Chandraker. Neurocs: Neural nocs supervision for monocular 3d object localization. In CVPR, pages 21404-21414, 2023. 2", + "[28] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection with vision transformers. In ECCV. Springer, 2022. 2, 6", + "[29] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529-533, 2015. 4" + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13946", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 5", + "[31] Shafin Rahman, Salman Khan, and Nick Barnes. Improved visual-semantic alignment for zero-shot object detection. In AAAI, pages 11932–11939, 2020. 2", + "[32] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In NeurIPS, 2015. 1, 3", + "[33] Zhongzheng Ren, Zhiding Yu, Xiaodong Yang, Ming-Yu Liu, Yong Jae Lee, Alexander G Schwing, and Jan Kautz. Instance-aware, context-focused, and memory-efficient weakly supervised object detection. In CVPR, pages 10598-10607, 2020. 2", + "[34] Samuel Schulter, Yumin Suh, Konstantinos M Dafnis, Zhixing Zhang, Shiyu Zhao, Dimitris Metaxas, et al. Omnilabel: A challenging benchmark for language-based object detection. ICCV, 2023. 2", + "[35] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2", + "[36] Bharat Singh, Hengduo Li, Abhishek Sharma, and Larry S Davis. R-FCN-3000 at 30fps: Decoupling detection and classification. In CVPR, pages 1081–1090, 2018. 3", + "[37] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. In arXiv:2005.04757, 2020. 1", + "[38] Peng Tang, Xinggang Wang, Xiang Bai, and Wenyu Liu. Multiple instance detection network with online instance classifier refinement. In CVPR, pages 2843-2851, 2017. 1, 2", + "[39] Fang Wan, Chang Liu, Wei Ke, Xiangyang Ji, Jianbin Jiao, and Qixiang Ye. C-MIL: Continuation multiple instance learning for weakly supervised object detection. In CVPR, pages 2199-2208, 2019. 2", + "[40] Luting Wang, Yi Liu, Penghui Du, Zihan Ding, Yue Liao, Qiaosong Qi, Biaolong Chen, and Si Liu. Object-aware distillation pyramid for open-vocabulary object detection. In CVPR, pages 11186-11196, 2023. 2", + "[41] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In CVPR, 2023. 2, 5, 6", + "[42] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. CORA: Adapting CLIP for Open-Vocabulary Detection with Region Prompting and Anchor Pre-Matching. In CVPR, 2023. 1, 2, 6", + "[43] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.5" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, pages 10687-10698, 2020. 1", + "[45] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In ICCV, pages 3060-3069, 2021. 1, 2, 4, 7", + "[46] Caixia Yan, Xiaojun Chang, Minnan Luo, Huan Liu, Xiaoqin Zhang, and Qinghua Zheng. Semantics-guided contrastive network for zero-shot object detection. IEEE TPAMI, 2022. 2", + "[47] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In ECCV, pages 106-122. Springer, 2022. 5, 6, 7", + "[48] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-Vocabulary Object Detection Using Captions. In CVPR, 2021. 2, 3, 4, 5, 6, 8", + "[49] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, BG Vijay Kumar, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris N Metaxas. Exploiting unlabeled data with vision and language models for object detection. In ECCV, pages 159-175. Springer, 2022. 1, 2, 5, 6, 7, 8", + "[50] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. RegionCLIP: Region-based Language-Image Pretraining. In CVPR, 2022. 2, 3, 5, 6", + "[51] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, pages 350-368. Springer, 2022. 5, 6, 7, 8", + "[52] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Don't even look once: Synthesizing features for zero-shot detection. In CVPR, pages 11693-11702, 2020. 2" + ], + "bbox": [ + 501, + 92, + 890, + 571 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13947", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Taming Self-Training for Open-Vocabulary Object Detection/08e6cd2d-6b02-4147-b946-b39f423452c4_model.json b/2024/Taming Self-Training for Open-Vocabulary Object Detection/08e6cd2d-6b02-4147-b946-b39f423452c4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..23d01ba8b0c1c9361255d1691a905685388203e9 --- /dev/null +++ b/2024/Taming Self-Training for Open-Vocabulary Object Detection/08e6cd2d-6b02-4147-b946-b39f423452c4_model.json @@ -0,0 +1,1903 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.131, + 0.792, + 0.154 + ], + "angle": 0, + "content": "Taming Self-Training for Open-Vocabulary Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.18, + 0.849, + 0.219 + ], + "angle": 0, + "content": "Shiyu Zhao\\(^{1,*}\\) Samuel Schulter\\(^{2}\\) Long Zhao\\(^{3}\\) Zhixing Zhang\\(^{1}\\) Vijay Kumar B G\\(^{2}\\) Yumin Suh\\(^{2}\\) Manmohan Chandraker\\(^{2,4}\\) Dimitris N. Metaxas\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.233, + 0.86, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\) Rutgers University \\(^{2}\\) NEC Laboratories America \\(^{3}\\) Google Research \\(^{4}\\) UC San Diego" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.726 + ], + "angle": 0, + "content": "Recent studies have shown promising performance in open-vocabulary object detection (OVD) by utilizing pseudo labels (PLs) from pretrained vision and language models (VLMs). However, teacher-student self-training, a powerful and widely used paradigm to leverage PLs, is rarely explored for OVD. This work identifies two challenges of using self-training in OVD: noisy PLs from VLMs and frequent distribution changes of PLs. To address these challenges, we propose SAS-Det that tames self-training for OVD from two key perspectives. First, we present a split-and-fusion (SAF) head that splits a standard detection into an open-branch and a closed-branch. This design can reduce noisy supervision from pseudo boxes. Moreover, the two branches learn complementary knowledge from different training data, significantly enhancing performance when fused together. Second, in our view, unlike in closed-set tasks, the PL distributions in OVD are solely determined by the teacher model. We introduce a periodic update strategy to decrease the number of updates to the teacher, thereby decreasing the frequency of changes in PL distributions, which stabilizes the training process. Extensive experiments demonstrate SAS-Det is both efficient and effective. SAS-Det outperforms recent models of the same scale by a clear margin and achieves 37.4 AP50 and 29.1 APr on novel categories of the COCO and LVIS benchmarks, respectively. Code is available at https://github.com/xiaofeng94/SAS-Det." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.753, + 0.21, + 0.768 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.778, + 0.47, + 0.869 + ], + "angle": 0, + "content": "Traditional closed-set object detectors [4, 13, 32] are restricted to detecting objects with a limited number of categories. Increasing the size of detection vocabularies usually requires heavy human labor to collect annotated data. With the recent advent of strong vision and language models (VLMs) [20, 30], open-vocabulary object detection" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.288, + 0.892, + 0.349 + ], + "angle": 0, + "content": "(OVD) [11] provides an alternative direction to approach this challenge. Typically, OVD detectors are trained with annotations of base categories and expected to generalize to novel categories with the power of pretrained VLMs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.893, + 0.457 + ], + "angle": 0, + "content": "One promising thread of recent studies for OVD [8, 9, 42, 49] leverages VLMs to obtain pseudo labels (PLs) beyond base categories. But they rarely explore self-training, a powerful and widely used schema for utilizing PLs in closed-set tasks [37, 38, 44, 45]. We investigate this and find the vanilla self-training approach does not improve OVD performance due to the following challenges." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.893, + 0.763 + ], + "angle": 0, + "content": "First, the typical self-training in closed-set tasks sets a confidence threshold to remove noisy PLs based on the fact that the quality of PLs is positively correlated to their confidences. However, VLMs employed in OVD are pretrained for image-level alignment with texts instead of instance-level object detection that requires the localization ability. Thus, the confidence score from pretrained VLMs is usually not a good indicator for the quality of box locations (i.e., pseudo boxes) provided by PLs. For example, prior studies [11, 49] show that CLIP [30] tends to output imperfect object boxes as predictions with high confidence. Recent methods for OVD [9, 49] just apply thresholding to VLMs' confidence scores and ignore the poor quality of pseudo boxes, which provides noisy supervision to the model. This issue becomes even worse when self-training is applied directly, since the noise accumulates which degrades the performance on novel categories. Moreover, these methods handle noisy PLs in the same way as ground truth of base categories during training, which further decreases the performance on base categories [9, 23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Second, self-training for closed-set object detection [19, 38, 45] usually follows an online teacher-student manner. In each training iteration, the teacher generates PLs, and the student is trained with a mixture of ground truth and PLs. Then, the teacher is updated by the student with exponential moving average (EMA). However, we find such EMA updates degrade OVD models (see Table 4). Our hypothesis is that, unlike closed-set tasks, OVD provides no ground truth for target categories, and thus, the supervision for target cat" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "* Part of this work was done during an internship at NEC Laboratories America. Correspondence to: Shiyu Zhao sz553@rutgers.edu" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13938" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "egories is fully decided by the distribution of PLs predicted by the teacher. Hence, the EMA updates change the distribution of PLs in each iteration, unstabilizing the training." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.138, + 0.47, + 0.394 + ], + "angle": 0, + "content": "In this paper, we propose Self-training And Split-and-fusion head for open-vocabulary Detection (SAS-Det) to tame self-training for OVD. First, we present a split-and-fusion (SAF) head to handle the noise of PLs. The SAF head splits the standard detection head into two branches: the closed-branch and the open-branch, which are fused at inference. The closed-branch, akin to the standard detection head, comprises a classification module and a box refinement module. It is supervised solely by ground truth from base categories, mitigating the impact of noisy PLs on the performance of base categories. The open-branch is a classification module supervised by class labels of both ground truth and PLs. It acquires complementary knowledge to the closed-branch and can significantly boost the performance when fused with the closed-branch. Moreover, this design circumvents noisy locations of pseudo boxes, reducing the accumulation of noise during self-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.47, + 0.485 + ], + "angle": 0, + "content": "Second, instead of adopting the vanilla EMA update, we reduce the number of the updates and periodically update the teacher by the student. The quality of our PLs improves along with the periodic updates, and our final PLs are better than those of prior PL-based methods [9, 49] that introduce external handcrafted steps. Fig. 1 shows the key differences." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.485, + 0.47, + 0.68 + ], + "angle": 0, + "content": "The proposed SAS-Det outperforms recent OVD models of the same scale by a clear margin on two popular benchmarks, i.e., COCO and LVIS. Without extra handcrafted steps, our pseudo labeling is more efficient than prior methods, i.e., nearly 4 times faster than PB-OVD [9] and 3 times faster than VL-PLM [49]. Extensive ablation studies demonstrate the effectiveness of the proposed components. On COCO, the thresholding of vanilla self-training decreases the performance of novel categories by 3.6 AP, and the EMA update decreases the performance by 6.9 AP. Instead, SAS-Det eliminates the degradation with two separate detection heads and the periodic update. The fusion of the SAF head boosts the performance by 6.0 AP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.833 + ], + "angle": 0, + "content": "The contributions of this work are summarized as follows. (1) We show two challenges of applying self-training to OVD and propose two simple but effective solutions, i.e., using different detection heads to mitigate the noise in PLs, and using periodic updates to reduce frequency of changes in PLs' distributions. (2) The proposed SAF head for OVD handles the noisy boxes of PLs and enables fusion to improve the performance. (3) We present the leading performance on COCO and LVIS under widely used OVD settings and provide detailed analysis of the proposed SAS-Det." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.22, + 0.86 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Vision-language models (VLMs). VLMs are trained to learn the alignment between images and text in a com" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.333 + ], + "angle": 0, + "content": "mon embedding space. CLIP [30] and ALIGN [14] use contrastive losses to learn such alignment on large-scale noisy image-text pairs from the Internet. ALBEF [20] introduces multi-modal fusion and additional self-supervised objectives. SIMLA [16] employs a single stream architecture to achieve multi-level image-text alignment. FDT [6] learns shared discrete tokens as the embedding space. These VLMs achieve impressive zero-shot performance on image classification. But due to the gap between the pretraining and detection tasks, VLMs have limited abilities in object detection. In this work, we attempt to close the gap via PLs. There are studies [15, 21, 34] focusing on aligning any text phrases with objects. But they require visual grounding data that are more expensive than detection annotations. Our work scales up the vocabulary size for object detection without requiring such costly data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.73 + ], + "angle": 0, + "content": "Open-vocabulary object detection (OVD). Zero-shot object detection methods [3, 31, 46, 52] increase the vocabulary size but with limited accuracy. Motivated by the strong zero-shot abilities of VLMs, recent efforts focus on OVD. Finetuning-based methods [17, 18, 28, 48, 50] add detection heads onto pretrained VLMs and then finetune the detector with concepts of base categories. Such methods are simple but may forget the knowledge learned in the pretraining [2]. Distillation-based methods [7, 11, 40, 41] introduce additional distillation loss functions that force the output of a detector to be close to that of a VLM and thus avoid forgetting. However, since the distillation losses are not designed for the detection task, they may conflict with detection objectives due to gradient conflicts that are a common issue for multi-task models. Methods like Feng et al. [8], Gao et al. [9], Wu et al. [42], Zhao et al. [49] create pseudo labels (PL) of novel concepts as supervision, and do not require extra losses, which sidesteps both catastrophic forgetting and gradient conflicts. But these methods need handcrafted steps to generate high quality PLs, e.g. multiple runs of box regression [49], activation maps [9] from Grad-CAM [35], image retrieval [8], or multi-stage training [42]. In this work, we address the two challenges of using self-training for OVD and enable an efficient end-to-end pseudo labeling pipeline. Besides, we point out the noise due to the poor locations of PLs and introduce the SAF head to handle such noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Self-training for object detection. Weakly-supervised object detection methods [27, 33, 38, 39] usually explore online self-training by distilling the knowledge from the model itself. Recent semi-supervised object detection methods [19, 26, 45] adopt a teacher-student design, where the teacher is an exponential moving average (EMA) of the student. Self-training has been widely explored in the above fields but rarely in OVD. Unlike semi-supervised object detection, OVD encounters two challenges to use self-training, i.e., more noisy PLs, and larger changes in PLs' distributions. This work proposes the SAF head to address" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13939" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.093, + 0.49, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.094, + 0.891, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.286, + 0.893, + 0.329 + ], + "angle": 0, + "content": "Figure 1. Left: Prior PL-based methods for OVD rely on handcrafted heuristics to leverage a frozen VLM for offline pseudo labels. This is usually inefficient and does not allow for improving PLs throughout training. Right: We customize self-training and finetune VLMs for OVD, which allows efficient on-the-fly computation of PLs that can be improved throughout training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.47, + 0.381 + ], + "angle": 0, + "content": "the noise, and adopts periodic updates to reduce the frequency of changes in PLs' distribution." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.399, + 0.185, + 0.416 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.471, + 0.56 + ], + "angle": 0, + "content": "In open-vocabulary detection, an object detector is trained with bounding boxes and class labels of base categories \\(\\mathcal{C}^B\\). At inference, the detector is used for detecting objects of open concepts including \\(\\mathcal{C}^B\\) and novel categories \\(\\mathcal{C}^N\\), where \\(\\mathcal{C}^B \\cap \\mathcal{C}^N = \\emptyset\\). To make such detection practical, recent studies [8, 11, 23, 48] adopt extra data (e.g., image-text pairs and image-level tags) and/or external VLMs. We follow their settings and leverage the pretrained CLIP [30] to build our OVD detector." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.296, + 0.589 + ], + "angle": 0, + "content": "3.1. Adapting CLIP to OVD" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.469, + 0.673 + ], + "angle": 0, + "content": "In this section, we introduce how to adapt the ResNet-based CLIP into a Faster R-CNN [32] (C4) detector. For simplicity, we do not use FPN [25], but it can be incorporated with learnable gating from Flamingo [1]. Similarly, ViT-based CLIP models can be adapted to detectors following [22]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Region proposals from an external RPN. CLIP is pretrained via image-text alignment. As shown in Sect. 4.3, finetuning pretrained backbones to get region proposals decreases the performance, probably because such finetuning breaks the image-text alignment learned in the pretraining. To address this problem, F-VLM [18] freezes the pretrained backbone and only finetunes detection heads. But this solution limits the capacity of the detector. Unlike F-VLM, we follow the approach of Singh et al. [36], Zhong et al. [50] and employ an external RPN to generate region proposals, which is trained with ground truth boxes of base categories. Prior studies adopt RPN mainly to accelerate inference [36] or to improve region recognition [50]. By contrast, we aim to leverage the RPN to preserve the knowledge learned in the pretraining for better self-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Text embeddings as the classifier. For the \\(i\\)-th region proposal from the external RPN, we apply RoIAlign [13] on the 4th feature maps of CLIP's ResNet to get the proposal features. Then, the features are fed to the last ResNet block and the attention pooling of CLIP to get the region embedding \\(r_i\\), which is later used for classification. Following prior studies [11, 23, 50], we convert a set of given concepts with prompt engineering into CLIP text embeddings, which act as classifiers, as shown in Fig. 2b. A fixed all-zero embedding is adopted for the background category. Assuming \\(t_c\\) is the embedding of the \\(c\\)-th category, the probability of \\(r_i\\) to be classified as the \\(c\\)-th category is," + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.544, + 0.892, + 0.582 + ], + "angle": 0, + "content": "\\[\np _ {i, c} = \\frac {\\exp \\left(\\langle r _ {i} , t _ {c} \\rangle / \\tau\\right)}{\\sum_ {j = 0} ^ {C - 1} \\exp \\left(\\langle r _ {i} , t _ {j} \\rangle / \\tau\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.892, + 0.7 + ], + "angle": 0, + "content": "where \\(\\langle \\cdot, \\cdot \\rangle\\) denotes the cosine distance, \\(C\\) denotes the vocabulary size, and \\(\\tau\\) denotes the temperature. In our detector, \\(r_i\\) may be further fed to a box refinement module to predict the box shift based on the region proposal. With the above adoptions, our initial detector gains zero-shot detection ability to some extent. Please refer to the supplement for quantitative evaluations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.702, + 0.727 + ], + "angle": 0, + "content": "3.2. Taming Self-Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Although self-training has been widely explored in closed-set object detection, using it for OVD presents two challenges, i.e., noisy PLs and frequent changes of PLs' distributions. In the following, we describe our self-training pipeline and how to address the two challenges." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Self-training pipeline. Fig. 2a illustrates the pipeline of our self-training with a teacher-student manner. Training images are first fed into the RPN to obtain region proposals. Then, the teacher model runs inference on those proposals, where the resulting predictions with confidence above a threshold are selected as PLs. The student model adopts" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13940" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.578, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.09, + 0.89, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.336, + 0.893, + 0.392 + ], + "angle": 0, + "content": "Figure 2. (a) Pipeline of our self-training. The teacher and the student are models of the same architecture. They are initialized with the same pretrained CLIP model. The teacher generates PLs that are used to train the student, and the student updates the teacher periodically. (b) Structure of our detector. The proposed SAF head is put on top of a CLIP image encoder. The open- and closed-branches take the text embeddings from a CLIP text encoder as classifier." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.47, + 0.478 + ], + "angle": 0, + "content": "the same region proposals, and is supervised with both base ground truth and PLs generated by the teacher. The teacher is periodically updated with the parameters of the student model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.47, + 0.651 + ], + "angle": 0, + "content": "Handling noise in PLs. The noise of PLs from VLMs exists in both classification and localization. It is a common practice to filter PLs with classification confidence, but this only addresses noise in classification. To reduce the noise in locations of PLs, we exclude the noisy boxes of PLs from the training. That is, the classification loss is calculated on class labels of both ground truth and PLs, but the box regression loss is calculated only on ground truth boxes. Besides, to mitigate the impact of PLs on the performance of base categories, we propose SAF head and elaborate it in Sect. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Reducing changes in PLs' distributions with periodic updates. The teacher model is updated with the student to enable self-training. The exponential moving average (EMA) is a widely used approach for object detection [19, 26, 45], which updates the teacher in every iteration. But we observed empirically that it does not benefit OVD (Table 4). We hypothesize that, unlike semi-supervised object detection, OVD has no ground truth for the target categories. Thus, the distribution of the target data (i.e., PLs) is fully determined by the teacher. The EMA update changes the PLs' distribution in each iteration and leads to a constantly shifting training target that has been shown hard to optimize (i.e., destabilizing the training) in deep Q-learning [29]. As a solution, we periodically update the teacher after a set number of iterations to maintain consistent distributions for PLs between updates. We call this" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.419, + 0.89, + 0.451 + ], + "angle": 0, + "content": "strategy as the periodic update and show it outperforms the EMA update by a large margin in Sect. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.464, + 0.761, + 0.48 + ], + "angle": 0, + "content": "3.3. Split-and-Fusion (SAF) Head" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.489, + 0.891, + 0.564 + ], + "angle": 0, + "content": "Our SAF head first splits a detection head into two branches, i.e. \"closed-branch\" and \"open-branch\", with the goal to better handle noisy PLs during training. At inference, predictions from both heads are fused to boost the performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Splitting the detection head. The closed-branch follows a standard detection head with a classification module and a class-agnostic box refinement module. The former module classifies region proposals based on Eq. 1, and the latter one refines the proposal boxes for better locations. We train the closed-branch with boxes and class labels of ground truth for \\(\\mathcal{C}^B\\) using standard detection losses, which include a cross entropy loss for classification and a box regression loss for localization. Since no PLs are used to train the closed-branch, the noise of PLs is unlike to impact its performance on \\(\\mathcal{C}^B\\). Moreover, as shown in Gu et al. [11], Zareian et al. [48], box regression modules trained on \\(\\mathcal{C}^B\\) can generalize to novel categories that are unseen during training. Therefore, the closed-branch is able to provide generalized boxes, as well." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The open-branch only contains a classification module. It is trained using only the cross-entropy loss with class labels from both \\(\\mathcal{C}^B\\) and PLs, hence, learning broader concepts beyond \\(\\mathcal{C}^B\\). Unlike distillation losses [7, 11], all losses for the two branches are originally designed for detection and are unlikely to conflict with each other. When generating PLs, we use the classification scores from the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13941" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.151 + ], + "angle": 0, + "content": "open-branch and bounding boxes from the closed-branch. Boxes of PLs are not directly used in our losses but are involved to select foreground proposal candidates for the classification loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.468, + 0.256 + ], + "angle": 0, + "content": "Fusing complementary predictions. The open- and closed-branches are trained in different ways and learn complementary knowledge. Therefore, we fuse their predictions with the geometric mean at inference time. Specifically, assuming \\( p_{i,c}^{\\mathrm{open}} \\) and \\( p_{i,c}^{\\mathrm{closed}} \\) are prediction scores of the open- and closed-branches, respectively, the final score is calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.264, + 0.468, + 0.304 + ], + "angle": 0, + "content": "\\[\np _ {i, c} ^ {\\text {f u s e d}} = \\left\\{ \\begin{array}{l l} \\left(p _ {i, c} ^ {\\text {c l o s e d}}\\right) ^ {(1 - \\alpha)} \\cdot \\left(p _ {i, c} ^ {\\text {o p e n}}\\right) ^ {\\alpha}, & \\text {i f} i \\in \\mathcal {C} ^ {B} \\\\ \\left(p _ {i, c} ^ {\\text {c l o s e d}}\\right) ^ {\\alpha} \\cdot \\left(p _ {i, c} ^ {\\text {o p e n}}\\right) ^ {(1 - \\alpha)}, & \\text {i f} i \\in \\mathcal {C} ^ {N} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.468, + 0.463 + ], + "angle": 0, + "content": "where \\(\\alpha \\in [0,1]\\) balances the two branches. The indices \\(i, c\\) are the same as in Eq. 1. We keep a list of known base categories \\(\\mathcal{C}^B\\) and take other categories beyond the list as novel. It's important to note that fusion is a common strategy. For instance, F-VLM [18] employs score fusion to enhance the location quality of CLIP's predictions. In contrast, our approach focuses on fusing the complementary predictions from the two branches. The primary contribution of our method lies in learning what to fuse rather than the fusion process itself." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.476, + 0.209, + 0.492 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.5, + 0.253, + 0.516 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.468, + 0.658 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on the two popular OVD benchmarks COCO-OVD and LVIS-OVD, based on COCO [24] and LVIS [12], respectively. (1) COCO-OVD: 65 out of 80 COCO categories are divided into 48 base categories and 17 novel categories. We report results on the validation set. (2) LVIS-OVD: 337 rare categories are used as the novel concepts and the remaining 866 categories (frequent and common) as the base concepts. We report results on the whole LVIS validation set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.468, + 0.764 + ], + "angle": 0, + "content": "Evaluation metrics. Following prior studies [11, 41, 48, 50], we report AP at IoU threshold of 0.5 for COCO-OVD. \\(\\mathrm{AP}_{50}^{novel}\\), \\(\\mathrm{AP}_{50}^{base}\\) and \\(\\mathrm{AP}_{50}^{all}\\) are the metrics for the novel, the base and all (novel + base) categories, respectively. For LVIS-OVD, we report the mean AP averaged on IoUs from 0.5 to 0.95. \\(\\mathrm{AP}_r\\), \\(\\mathrm{AP}_c\\), \\(\\mathrm{AP}_f\\) and AP are the metrics for rare, common, frequent, and all categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Implementation details. We built our detector on Faster R-CNN with the CLIP version of ResNet (RN50-C4 or RN50x4-C4) as the backbone. Following prior studies [47, 49, 51], we assumed the novel concepts were available during training and took them as our open concepts, but no annotations of novel concepts are used. Our method was implemented on Detectron2 [43] and trained with 8 NVIDIA A6000 GPUs. We trained our models with the \\(1 \\times\\) schedule (90k iterations) for COCO-OVD and the \\(2 \\times\\) schedule for" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "LVIS-OVD. The batch size was 16 with an initial learning rate of 0.002. The default data augmentation of Detectron2 was applied. Loss terms were equally weighted. By default, the teacher models were updated three times, usually together with the decreases of the learning rate." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.178, + 0.836, + 0.194 + ], + "angle": 0, + "content": "4.2. Comparison with the Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.892, + 0.502 + ], + "angle": 0, + "content": "COCO-OVD. We compare our method with prior work on COCO in Table 1. With the same backbone and detector, SAS-Det outperforms the most recent method BARON [41] by \\(4.3\\mathrm{AP}_{50}^{novel}\\) on novel categories and \\(3.7\\mathrm{AP}_{50}^{base}\\) on base categories. Probably, BARON distills knowledge from VLMs, but distillation may lead to gradient conflicts and degrade the performance [49]. Compared to VLDet [23], SAS-Det gains an improvement of \\(7.9\\mathrm{AP}_{50}^{base}\\). It is likely that VLDet employs noisy pseudo labels to train the detection head, impacting the detection of base classes in both classification and localization. In contrast, the two-branch design of our SAF head allows to train the closed-branch with just ground-truth of base classes and reduces the impact of the noise from pseudo labels. For more analysis on where our improvements come, please refer to Sect. 4.3 and Table 3. The first block of Table 1 provides results for methods with stronger backbones or detector architectures. When compared with those methods, SAS-Det still achieves the leading performance. Those results clearly demonstrate the effectiveness of SAS-Det." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.892, + 0.805 + ], + "angle": 0, + "content": "LVIS-OVD. We provide the main results on LVIS in Table 2. When using ResNet50 as the backbone, SAS-Det achieves similar performance as the recent method DetPro [7]. DetPro proposes learnable prompts to generate better text embeddings as the classifier for OVD. Our method is orthogonal and can leverage DetPro's prompts for further improvement. The first block of Table 2 also shows that SAS-Det outperforms ViLD by a large margin on novel categories (indicated by \\(\\mathrm{AP}_r\\)) but gets lower \\(\\mathrm{AP}_f\\) for base categories. This is also observed in the recent method BARON [41]. The performance gap is probably due to the training setup. ViLD is trained for 8 times more iterations and adopts a strong data augmentation method [10], both of which benefit detection on base categories. In contrast, BARON and our approach adopt shorter training and standard data augmentations. When replacing the pretrained CLIP ResNet50 with ResNet50x4 [30] as the backbone, we improve \\(\\mathrm{AP}_r\\) by 8.2 from 20.9 to 29.1, which demonstrates SAS-Det scales up nicely with stronger pretrained VLM backbones." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.665, + 0.832 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "In this section, the default baseline shares the same architecture and the training as our SAS-Det (with RN50-C4 as the backbone), except that the SAF head is replaced with a single detection head. The localization module of the detec" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13942" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.153, + 0.089, + 0.821, + 0.308 + ], + "angle": 0, + "content": "
MethodTraining SetupBackboneDetector\\(AP^{novel}\\)\\(AP^{base}\\)\\(AP^{all}\\)
ViLD [11]16×+LSJRN50-FPNFaster R-CNN27.659.551.2
VL-PLM [49]1×+DefaultRN50-FPNFaster R-CNN32.354.048.3
F-VLM [18]0.5×+LSJRN50-FPNFaster R-CNN28.0-39.6
OV-DETR [47](Not Given)RN50Deform. DETR29.461.052.7
CORA [42]3×+DefaultRN50DAB-DETR35.135.535.4
OVR-CNN [48](Not Given)RN50-C4Faster R-CNN22.846.039.9
RegionCLIP [50]1×+DefaultRN50-C4Faster R-CNN26.854.847.5
Detic [51]1×+DefaultRN50-C4Faster R-CNN27.851.145.0
PB-OVD [9]6×+DefaultRN50-C4Faster R-CNN30.846.142.1
VLDet [23]1×+LSJRN50-C4Faster R-CNN32.050.645.8
BARON [41]1×+DefaultRN50-C4Faster R-CNN33.154.849.1
SAS-Det (Ours)1×+DefaultRN50-C4Faster R-CNN37.458.553.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.319, + 0.895, + 0.364 + ], + "angle": 0, + "content": "Table 1. Comparison with recent methods on COCO-OVD. We group methods into two blocks. The first block contains methods using stronger backbones or detector architectures than ours. The second block contains models of the same scale as ours. Training setup indicates training iterations \\((\\mathrm{N}\\times)\\) and data augmentations. Large Scale Jittering (LSJ) [10] is a stronger data augmentation than Detectron2's default." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.375, + 0.819, + 0.564 + ], + "angle": 0, + "content": "
MethodTraining SetupBackboneDetectorAPrAPcAPfAP
ViLD [11]16×+LSJRN50-FPNFaster R-CNN16.726.534.227.8
DetPro [7]2×+DefaultRN50-FPNFaster R-CNN20.827.832.428.4
F-VLM [18]9×+LSJRN50-FPNFaster R-CNN18.6--24.2
BARON [41]2×+DefaultRN50-C4Faster R-CNN17.325.631.026.3
SAS-Det (Ours)2×+DefaultRN50-C4Faster R-CNN20.926.131.627.4
ViLD [11]16×+LSJRN152-FPNFaster R-CNN19.827.134.528.7
OWL-ViT [28]12×+LSJViT-L/14OWL-ViT25.6--34.7
F-VLM [18]9×+LSJRN50x4-FPNFaster R-CNN26.3--28.5
CORA [42]3×+DefaultRN50x4DAB-DETR22.2---
SAS-Det (Ours)2×+DefaultRN50x4-C4Faster R-CNN29.132.436.833.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.574, + 0.893, + 0.604 + ], + "angle": 0, + "content": "Table 2. Comparison with recent methods on LVIS-OVD. We group methods based on the scale of backbones. Training setup contains training iterations \\((\\mathrm{N}\\times)\\) and data augmentations. LSJ [10] is a stronger data augmentation than Detectron2's default." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.471, + 0.705 + ], + "angle": 0, + "content": "tion head is trained with base ground truth boxes only, but the classification module is trained with class labels of both ground truth and PLs. Thus, the baseline provides a naive solution to exclude noisy pseudo boxes from the training. All evaluations are conducted on COCO." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.47, + 0.864 + ], + "angle": 0, + "content": "External RPN. We leverage an external RPN to generate region proposals so that finetuning focuses on region-text alignment that is similar to the pretraining task. In this way, the detector is unlikely to forget the knowledge obtained in the pretraining. To validate the effectiveness of the external RPN, we follow F-VLM [18] to train a detector without an external RPN (See the supplement for more details). Table 3 compares baseline with the detector in the row of (1). As shown, without the external RPN, the performance drops on both novel and base categories." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Removing boxes of PLs from training. Our baseline handles location noise of PLs' boxes by directly excluding them" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.706 + ], + "angle": 0, + "content": "from the training. As shown in Table 3, baseline outperforms (2)'s detector. The only difference between them is the removal of pseudo boxes from training. This result clearly shows that simply removing the pseudo boxes is an effective way to deal with location noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Splitting the detection head. As shown in Table 3, compared to baseline, the open-branch of the proposed SAF head in (5) achieves similar performance on novel categories. This indicates that the split handles the noise in PLs' location as well as the naive solution of baseline. This is plausible because both solutions exclude the noisy boxes of PLs from training. Additionally, the open-branch gains 1.7 \\(\\mathrm{AP}_{50}^{base}\\) on base categories. Compared to (3)'s, the closed-branch in (4) gain improvements in terms of both \\(\\mathrm{AP}_{50}^{novel}\\) and \\(\\mathrm{AP}_{50}^{base}\\). Note that (4)'s and (3)'s heads are trained with the same data. Based on those results, the split benefits OVD beyond handling the location noise. One advantage" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13943" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.148, + 0.089, + 0.825, + 0.228 + ], + "angle": 0, + "content": "
Ablation\\(AP_{50}^{novel}\\)\\(AP_{50}^{base}\\)\\(AP_{50}^{all}\\)
Baseline31.455.749.4
(1) No external RPN, train the backbone for region proposals(-6.0) 25.453.446.1
(2) Noisy boxes of PLs as supervision for box regression(-3.6) 27.855.248.0
(3) No PLs, train with base data only(-8.2) 23.256.948.1
(4) W/ SAF head, use closed-branch's predictions (\\(p_{i,c}^{closed}\\) in Eq. 2)(-5.8) 25.657.949.5
(5) W/ SAF head, use open-branch's predictions (\\(p_{i,c}^{open}\\) in Eq. 2)(+0.5) 31.957.450.7
(6) W/ SAF head, use fused predictions (\\(p_{i,c}^{fused}\\) in Eq. 2)(+6.0) 37.458.553.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.239, + 0.895, + 0.268 + ], + "angle": 0, + "content": "Table 3. Ablation studies to analyze the effect of each component of SAS-Det on COCO-OVD. Results (4), (5), and (6) denote different outputs from the same model." + }, + { + "type": "table", + "bbox": [ + 0.185, + 0.28, + 0.787, + 0.389 + ], + "angle": 0, + "content": "
Update strategy\\( AP_{50}^{novel} \\)\\( AP_{50}^{base} \\)\\( AP_{50}^{all} \\)
Baseline (w/ periodic update)31.455.749.4
(7) No teacher and no pseudo labels(-8.2) 23.256.948.1
(8) Not update the teacher(-1.8) 29.655.949.1
(9) Take the EMA of the student as the teacher(-6.9) 24.553.946.2
(10) Replace the teacher with the student every iteration(-23.4) 8.053.341.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.188, + 0.399, + 0.782, + 0.413 + ], + "angle": 0, + "content": "Table 4. Comparison of different strategies for updating the teacher in self-training on COCO-OVD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.485 + ], + "angle": 0, + "content": "is that, due to the split, the closed-branch is less likely to be influenced by the noise of PLs and learn how to better localize objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.7 + ], + "angle": 0, + "content": "Fusing predictions. Our SAF head fuses predictions of the open-branch and the closed-branch at inference time. As shown in Table 3, (6)'s outperforms all the others by a large margin with the prediction fusion. Note that (4)'s, (5)'s and (6)'s refer to the different outputs of the same model. The improvement can be attributed to the fact that the two branches are trained on different sets of data and learn complementary knowledge. The closed-branch is trained with class labels and boxes of base categories. The open-branch is trained with class labels of more vocabularies including base and novel classes. Thus, the former learns more about how to localize objects and how to detect base classes. The latter learns more about how to detect new objects, which complements the former." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Update strategies for the teacher model. The teacher is updated with the student model to improve PLs during finetuning. We evaluated several strategies for updating the teacher model in Table 4 and summarize our findings as follows. First, the EMA update in (9), which is shown effective and widely used in semi-supervised object detection [19, 26, 45], is just slightly better than training without PLs in (7). It is worse than no update to the teacher (i.e. no self-training) in (8). Second, if we replace the teacher with the student every iteration as (10), \\(\\mathrm{AP}_{50}^{novel}\\) drops to 8.0 that is much lower than that number in (7). These findings indicate that it's harmful to change the teacher frequently. We hypothesize that frequent updates change the distribution of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.485 + ], + "angle": 0, + "content": "PLs too often and makes the training unstable. Last but not least, by reducing the number of updates, our periodic update significantly outperforms the EMA update in OVD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.893, + 0.683 + ], + "angle": 0, + "content": "The number of updates to the teacher model. We trained our detectors with different numbers of updates to teacher models. Please see the supplement for how we distribute the updates. As shown in Table 5, too many updates, e.g., 8 or 4 updates, lead to performance drops mainly due to the following. First, similar as the aforementioned EMA update, too many updates change the distribution of PLs too often and make the training unstable. Second, the more updates, the earlier an update happens. However, the student model is not well trained at the early stage of the training and thus is not good enough to update the teacher. Table 5 shows that 2 and 3 updates achieve similar performance. But we set 3 updates as default to include as many updates as possible." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.667, + 0.712 + ], + "angle": 0, + "content": "4.4. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation on the retained 15 COCO categories. Following prior studies [47, 49, 51], SAS-Det assumes novel concepts are available during training and achieves good performance. A natural question is if SAS-Det generalizes to alien concepts that are unknown before the evaluation. To answer the question, we follow the evaluation protocol of MEDet [5] where an OVD detector is evaluated on the whole COCO validation set with all 80 COCO concepts. \\(\\mathrm{AP}_{50}^{retain}\\), \\(\\mathrm{AP}_{50}^{novel}\\) and \\(\\mathrm{AP}_{50}^{coco}\\) are reported as APs averaged on 15 retained, 17 novel, and all 80 COCO categories, respectively. As shown in Table 6, SAS-Det achieves leading performance by a clear margin, which indicates the good" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13944" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.169, + 0.09, + 0.357, + 0.213 + ], + "angle": 0, + "content": "
# Update\\(AP_{50}^{novel}\\)
827.6
430.6
3 (baseline)31.4
231.6
130.9
0 (No update)29.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.223, + 0.449, + 0.25 + ], + "angle": 0, + "content": "Table 5. Performance with varied numbers of updates to the teacher model." + }, + { + "type": "image", + "bbox": [ + 0.103, + 0.271, + 0.455, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.415, + 0.405, + 0.43 + ], + "angle": 0, + "content": "Figure 3. Quality of PLs during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.405, + 0.471 + ], + "angle": 0, + "content": "generalization capability of the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.652 + ], + "angle": 0, + "content": "PL's quality during self-training. We compare our PLs and the prior state-of-the-art PLs from VL-PLM [49] on the COCO validation set in terms of PLs' quality. Following VL-PLM, we take \\(\\mathrm{AP}_{50}^{novel}\\) as the metric and evaluate our teacher models during training. As illustrated in Fig. 3, the quality of our initial PLs is not as good as VL-PLM's. But ours get close to VL-PLM's after two updates and become the better by the third update. This is because the teacher is initialized with CLIP, which is not pretrained for detection, and thus cannot provide high quality PLs. With updates, the teacher is equipped with the knowledge about detection and generates promising PLs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.47, + 0.711 + ], + "angle": 0, + "content": "Time cost of pseudo labeling. We run two major PL-based methods [9, 49] on our computational environment and compare them with our pseudo labeling in terms of the time cost in Table 7." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.469, + 0.833 + ], + "angle": 0, + "content": "Reducing the time cost of external RPN. There is a simple solution to remove the extra computational cost of external RPN adopted by SAS-Det. We first generate pseudo labels with SAS-Det, and then follow VL-PLM [49] to train a Faster R-CNN based on ResNet50 for OVD without the external RPN (\"Faster R-CNN + PLs\"). As shown in Table 8, \"Faster R-CNN + PLs\" achieves slightly better or close performance as SAS-Det." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.195, + 0.86 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In this paper, we highlight two challenges associated with applying self-training to OVD: 1) noisy PLs from pretrained" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.09, + 0.861, + 0.183 + ], + "angle": 0, + "content": "
MethodAPretain50APnovel50APcoco50
OVR-CNN [48]11.522.938.1
Detic-80 [51]11.527.338.3
MEDet [5]18.632.642.4
Ours23.037.147.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.481, + 0.193, + 0.894, + 0.235 + ], + "angle": 0, + "content": "Table 6. Generalization capability of OVD methods on retained 15 COCO categories. Text embeddings of 80 COCO categories are used as the classifier. Numbers are from MEDet [5]." + }, + { + "type": "table", + "bbox": [ + 0.558, + 0.266, + 0.824, + 0.344 + ], + "angle": 0, + "content": "
MethodTime (s)APnovel50
PB-OVD [9]0.484818.7
VL-PLM [49]0.445625.5
Ours0.130826.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.483, + 0.354, + 0.896, + 0.397 + ], + "angle": 0, + "content": "Table 7. Mean time cost to get PLs per image. The quality of PLs on the validation set of COCO-OVD are provided for reference. We report the quality of our PLs after the last update." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.453, + 0.876, + 0.516 + ], + "angle": 0, + "content": "
Method\\( AP_{50}^{novel} \\)\\( AP_{50}^{base} \\)\\( AP_{50}^{all} \\)
SAS-Det37.458.553.0
Faster R-CNN + PLs38.158.353.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.526, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Table 8. Training Faster R-CNN with PLs from SAS-Det on COCO-OVD" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.838 + ], + "angle": 0, + "content": "VLMs and 2) frequent changes of PLs' distributions. To address the challenges, we introduce a split-and-fusion (SAF) head and implement periodic updates. The SAF head splits a standard detection head into an open-branch and a closed-branch. The open-branch is trained with the class labels of ground truth and PLs. The closed-branch is trained with both class labels and boxes of ground truth. This approach effectively eliminates the location noise of PLs during training. Furthermore, the SAF head acquires complementary knowledge from different sets of training data, enabling fusion to enhance performance. The periodic updates reduce the number of updates to the teacher models, thereby decreasing the frequency of changes in PLs' distributions. We demonstrate that the proposed method SAS-Det is both efficient and effective. Our pseudo labeling is much faster than prior PL-based methods [9, 49]. SAS-Det also outperforms recent models on both COCO and LVIS for OVD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgments: This research project has been partially funded by research grants to Dimitris N. Metaxas through NSF: 2310966, 2235405, 2212301, 2003874, and FA9550-23-1-0417." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13945" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Relja Arandjelović, Alex Andonian, Arthur Mensch, Olivier J Henaff, Jean-Baptiste Alayrac, and Andrew Zisserman. Three ways to improve feature alignment for open vocabulary detection, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.284 + ], + "angle": 0, + "content": "[3] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In ECCV, pages 384-400, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.326 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-End Object Detection with Transformers. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[5] Peixian Chen, Kekai Sheng, Mengdan Zhang, Yunhang Shen, Ke Li, and Chunhua Shen. Open vocabulary object detection with proposal mining and prediction equalization, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.381, + 0.471, + 0.45 + ], + "angle": 0, + "content": "[6] Yuxiao Chen, Jianbo Yuan, Yu Tian, Shijie Geng, Xinyu Li, Ding Zhou, Dimitris N Metaxas, and Hongxia Yang. Revisiting multimodal representation in contrastive learning: from patch and token embeddings to finite discrete tokens. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.451, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[7] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In CVPR, pages 14084-14093, 2022. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[8] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In ECCV, pages 701-717. Springer, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[9] Mingfei Gao, Chen Xing, Juan Carlos Niebles, Junnan Li, Ran Xu, Wenhao Liu, and Caiming Xiong. Open vocabulary object detection with pseudo bounding-box labels. In ECCV, pages 266-282. Springer, 2022. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.62, + 0.471, + 0.676 + ], + "angle": 0, + "content": "[10] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, pages 2918-2928, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.677, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary Object Detection via Vision and Language Knowledge Distillation. In ICLR, 2022. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.719, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[12] Agrim Gupta, Piotr Dólar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.761, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[13] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.789, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[14] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.859, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[15] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR - Modulated detection for end-to-end multi-modal understanding." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[16] Zaid Khan, Vijay Kumar B.G., Xiang Yu, Samuel Schulter, Manmohan Chandraker, and Yun Fu. Single-stream multilevel alignment for vision-language pretraining. In ECCV, pages 735-751. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.177, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[17] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Region-aware pretraining for open-vocabulary object detection with vision transformers. In CVPR, pages 11144-11154, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[18] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. In ICLR, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[19] Gang Li, Xiang Li, Yujie Wang, Yichao Wu, Ding Liang, and Shanshan Zhang. Pseudo labeling and consistency training for semi-supervised object detection. In ECCV, pages 457-472. Springer, 2022. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[20] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[21] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, pages 10965-10975, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.497 + ], + "angle": 0, + "content": "[22] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, pages 280-296. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[23] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In ICLR, 2023. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.608, + 0.892, + 0.651 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature Pyramid Networks for Object Detection. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[26] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In ICLR, 2021. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[27] Zhixiang Min, Bingbing Zhuang, Samuel Schulter, Buyu Liu, Enrique Dunn, and Manmohan Chandraker. Neurocs: Neural nocs supervision for monocular 3d object localization. In CVPR, pages 21404-21414, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[28] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection with vision transformers. In ECCV. Springer, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[29] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529-533, 2015. 4" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13946" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.175 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.472, + 0.22 + ], + "angle": 0, + "content": "[31] Shafin Rahman, Salman Khan, and Nick Barnes. Improved visual-semantic alignment for zero-shot object detection. In AAAI, pages 11932–11939, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.47, + 0.264 + ], + "angle": 0, + "content": "[32] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In NeurIPS, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.267, + 0.47, + 0.335 + ], + "angle": 0, + "content": "[33] Zhongzheng Ren, Zhiding Yu, Xiaodong Yang, Ming-Yu Liu, Yong Jae Lee, Alexander G Schwing, and Jan Kautz. Instance-aware, context-focused, and memory-efficient weakly supervised object detection. In CVPR, pages 10598-10607, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.338, + 0.47, + 0.393 + ], + "angle": 0, + "content": "[34] Samuel Schulter, Yumin Suh, Konstantinos M Dafnis, Zhixing Zhang, Shiyu Zhao, Dimitris Metaxas, et al. Omnilabel: A challenging benchmark for language-based object detection. ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.396, + 0.47, + 0.478 + ], + "angle": 0, + "content": "[35] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.481, + 0.47, + 0.523 + ], + "angle": 0, + "content": "[36] Bharat Singh, Hengduo Li, Abhishek Sharma, and Larry S Davis. R-FCN-3000 at 30fps: Decoupling detection and classification. In CVPR, pages 1081–1090, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.47, + 0.58 + ], + "angle": 0, + "content": "[37] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. In arXiv:2005.04757, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.584, + 0.47, + 0.638 + ], + "angle": 0, + "content": "[38] Peng Tang, Xinggang Wang, Xiang Bai, and Wenyu Liu. Multiple instance detection network with online instance classifier refinement. In CVPR, pages 2843-2851, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.642, + 0.47, + 0.697 + ], + "angle": 0, + "content": "[39] Fang Wan, Chang Liu, Wei Ke, Xiangyang Ji, Jianbin Jiao, and Qixiang Ye. C-MIL: Continuation multiple instance learning for weakly supervised object detection. In CVPR, pages 2199-2208, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.47, + 0.755 + ], + "angle": 0, + "content": "[40] Luting Wang, Yi Liu, Penghui Du, Zihan Ding, Yue Liao, Qiaosong Qi, Biaolong Chen, and Si Liu. Object-aware distillation pyramid for open-vocabulary object detection. In CVPR, pages 11186-11196, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.47, + 0.799 + ], + "angle": 0, + "content": "[41] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In CVPR, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.856 + ], + "angle": 0, + "content": "[42] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. CORA: Adapting CLIP for Open-Vocabulary Detection with Region Prompting and Anchor Pre-Matching. In CVPR, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[43] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[44] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, pages 10687-10698, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[45] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In ICCV, pages 3060-3069, 2021. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[46] Caixia Yan, Xiaojun Chang, Minnan Luo, Huan Liu, Xiaoqin Zhang, and Qinghua Zheng. Semantics-guided contrastive network for zero-shot object detection. IEEE TPAMI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[47] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In ECCV, pages 106-122. Springer, 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[48] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-Vocabulary Object Detection Using Captions. In CVPR, 2021. 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[49] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, BG Vijay Kumar, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris N Metaxas. Exploiting unlabeled data with vision and language models for object detection. In ECCV, pages 159-175. Springer, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[50] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. RegionCLIP: Region-based Language-Image Pretraining. In CVPR, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[51] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, pages 350-368. Springer, 2022. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[52] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Don't even look once: Synthesizing features for zero-shot detection. In CVPR, pages 11693-11702, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13947" + } + ] +] \ No newline at end of file diff --git a/2024/Taming Self-Training for Open-Vocabulary Object Detection/08e6cd2d-6b02-4147-b946-b39f423452c4_origin.pdf b/2024/Taming Self-Training for Open-Vocabulary Object Detection/08e6cd2d-6b02-4147-b946-b39f423452c4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8244b33c48b4c5ea1e6d48042c5bb607e36ffff7 --- /dev/null +++ b/2024/Taming Self-Training for Open-Vocabulary Object Detection/08e6cd2d-6b02-4147-b946-b39f423452c4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1190262e432755dd0da60605e63abba09bf9702ae9cdf96705df38e21011ebf +size 1086168 diff --git a/2024/Taming Self-Training for Open-Vocabulary Object Detection/full.md b/2024/Taming Self-Training for Open-Vocabulary Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2f712eb0dda14f800820a3aa8f00e777fff10fed --- /dev/null +++ b/2024/Taming Self-Training for Open-Vocabulary Object Detection/full.md @@ -0,0 +1,259 @@ +# Taming Self-Training for Open-Vocabulary Object Detection + +Shiyu Zhao $^{1,*}$ Samuel Schulter $^{2}$ Long Zhao $^{3}$ Zhixing Zhang $^{1}$ Vijay Kumar B G $^{2}$ Yumin Suh $^{2}$ Manmohan Chandraker $^{2,4}$ Dimitris N. Metaxas $^{1}$ + +$^{1}$ Rutgers University $^{2}$ NEC Laboratories America $^{3}$ Google Research $^{4}$ UC San Diego + +# Abstract + +Recent studies have shown promising performance in open-vocabulary object detection (OVD) by utilizing pseudo labels (PLs) from pretrained vision and language models (VLMs). However, teacher-student self-training, a powerful and widely used paradigm to leverage PLs, is rarely explored for OVD. This work identifies two challenges of using self-training in OVD: noisy PLs from VLMs and frequent distribution changes of PLs. To address these challenges, we propose SAS-Det that tames self-training for OVD from two key perspectives. First, we present a split-and-fusion (SAF) head that splits a standard detection into an open-branch and a closed-branch. This design can reduce noisy supervision from pseudo boxes. Moreover, the two branches learn complementary knowledge from different training data, significantly enhancing performance when fused together. Second, in our view, unlike in closed-set tasks, the PL distributions in OVD are solely determined by the teacher model. We introduce a periodic update strategy to decrease the number of updates to the teacher, thereby decreasing the frequency of changes in PL distributions, which stabilizes the training process. Extensive experiments demonstrate SAS-Det is both efficient and effective. SAS-Det outperforms recent models of the same scale by a clear margin and achieves 37.4 AP50 and 29.1 APr on novel categories of the COCO and LVIS benchmarks, respectively. Code is available at https://github.com/xiaofeng94/SAS-Det. + +# 1. Introduction + +Traditional closed-set object detectors [4, 13, 32] are restricted to detecting objects with a limited number of categories. Increasing the size of detection vocabularies usually requires heavy human labor to collect annotated data. With the recent advent of strong vision and language models (VLMs) [20, 30], open-vocabulary object detection + +(OVD) [11] provides an alternative direction to approach this challenge. Typically, OVD detectors are trained with annotations of base categories and expected to generalize to novel categories with the power of pretrained VLMs. + +One promising thread of recent studies for OVD [8, 9, 42, 49] leverages VLMs to obtain pseudo labels (PLs) beyond base categories. But they rarely explore self-training, a powerful and widely used schema for utilizing PLs in closed-set tasks [37, 38, 44, 45]. We investigate this and find the vanilla self-training approach does not improve OVD performance due to the following challenges. + +First, the typical self-training in closed-set tasks sets a confidence threshold to remove noisy PLs based on the fact that the quality of PLs is positively correlated to their confidences. However, VLMs employed in OVD are pretrained for image-level alignment with texts instead of instance-level object detection that requires the localization ability. Thus, the confidence score from pretrained VLMs is usually not a good indicator for the quality of box locations (i.e., pseudo boxes) provided by PLs. For example, prior studies [11, 49] show that CLIP [30] tends to output imperfect object boxes as predictions with high confidence. Recent methods for OVD [9, 49] just apply thresholding to VLMs' confidence scores and ignore the poor quality of pseudo boxes, which provides noisy supervision to the model. This issue becomes even worse when self-training is applied directly, since the noise accumulates which degrades the performance on novel categories. Moreover, these methods handle noisy PLs in the same way as ground truth of base categories during training, which further decreases the performance on base categories [9, 23]. + +Second, self-training for closed-set object detection [19, 38, 45] usually follows an online teacher-student manner. In each training iteration, the teacher generates PLs, and the student is trained with a mixture of ground truth and PLs. Then, the teacher is updated by the student with exponential moving average (EMA). However, we find such EMA updates degrade OVD models (see Table 4). Our hypothesis is that, unlike closed-set tasks, OVD provides no ground truth for target categories, and thus, the supervision for target cat + +egories is fully decided by the distribution of PLs predicted by the teacher. Hence, the EMA updates change the distribution of PLs in each iteration, unstabilizing the training. + +In this paper, we propose Self-training And Split-and-fusion head for open-vocabulary Detection (SAS-Det) to tame self-training for OVD. First, we present a split-and-fusion (SAF) head to handle the noise of PLs. The SAF head splits the standard detection head into two branches: the closed-branch and the open-branch, which are fused at inference. The closed-branch, akin to the standard detection head, comprises a classification module and a box refinement module. It is supervised solely by ground truth from base categories, mitigating the impact of noisy PLs on the performance of base categories. The open-branch is a classification module supervised by class labels of both ground truth and PLs. It acquires complementary knowledge to the closed-branch and can significantly boost the performance when fused with the closed-branch. Moreover, this design circumvents noisy locations of pseudo boxes, reducing the accumulation of noise during self-training. + +Second, instead of adopting the vanilla EMA update, we reduce the number of the updates and periodically update the teacher by the student. The quality of our PLs improves along with the periodic updates, and our final PLs are better than those of prior PL-based methods [9, 49] that introduce external handcrafted steps. Fig. 1 shows the key differences. + +The proposed SAS-Det outperforms recent OVD models of the same scale by a clear margin on two popular benchmarks, i.e., COCO and LVIS. Without extra handcrafted steps, our pseudo labeling is more efficient than prior methods, i.e., nearly 4 times faster than PB-OVD [9] and 3 times faster than VL-PLM [49]. Extensive ablation studies demonstrate the effectiveness of the proposed components. On COCO, the thresholding of vanilla self-training decreases the performance of novel categories by 3.6 AP, and the EMA update decreases the performance by 6.9 AP. Instead, SAS-Det eliminates the degradation with two separate detection heads and the periodic update. The fusion of the SAF head boosts the performance by 6.0 AP. + +The contributions of this work are summarized as follows. (1) We show two challenges of applying self-training to OVD and propose two simple but effective solutions, i.e., using different detection heads to mitigate the noise in PLs, and using periodic updates to reduce frequency of changes in PLs' distributions. (2) The proposed SAF head for OVD handles the noisy boxes of PLs and enables fusion to improve the performance. (3) We present the leading performance on COCO and LVIS under widely used OVD settings and provide detailed analysis of the proposed SAS-Det. + +# 2. Related Work + +Vision-language models (VLMs). VLMs are trained to learn the alignment between images and text in a com + +mon embedding space. CLIP [30] and ALIGN [14] use contrastive losses to learn such alignment on large-scale noisy image-text pairs from the Internet. ALBEF [20] introduces multi-modal fusion and additional self-supervised objectives. SIMLA [16] employs a single stream architecture to achieve multi-level image-text alignment. FDT [6] learns shared discrete tokens as the embedding space. These VLMs achieve impressive zero-shot performance on image classification. But due to the gap between the pretraining and detection tasks, VLMs have limited abilities in object detection. In this work, we attempt to close the gap via PLs. There are studies [15, 21, 34] focusing on aligning any text phrases with objects. But they require visual grounding data that are more expensive than detection annotations. Our work scales up the vocabulary size for object detection without requiring such costly data. + +Open-vocabulary object detection (OVD). Zero-shot object detection methods [3, 31, 46, 52] increase the vocabulary size but with limited accuracy. Motivated by the strong zero-shot abilities of VLMs, recent efforts focus on OVD. Finetuning-based methods [17, 18, 28, 48, 50] add detection heads onto pretrained VLMs and then finetune the detector with concepts of base categories. Such methods are simple but may forget the knowledge learned in the pretraining [2]. Distillation-based methods [7, 11, 40, 41] introduce additional distillation loss functions that force the output of a detector to be close to that of a VLM and thus avoid forgetting. However, since the distillation losses are not designed for the detection task, they may conflict with detection objectives due to gradient conflicts that are a common issue for multi-task models. Methods like Feng et al. [8], Gao et al. [9], Wu et al. [42], Zhao et al. [49] create pseudo labels (PL) of novel concepts as supervision, and do not require extra losses, which sidesteps both catastrophic forgetting and gradient conflicts. But these methods need handcrafted steps to generate high quality PLs, e.g. multiple runs of box regression [49], activation maps [9] from Grad-CAM [35], image retrieval [8], or multi-stage training [42]. In this work, we address the two challenges of using self-training for OVD and enable an efficient end-to-end pseudo labeling pipeline. Besides, we point out the noise due to the poor locations of PLs and introduce the SAF head to handle such noise. + +Self-training for object detection. Weakly-supervised object detection methods [27, 33, 38, 39] usually explore online self-training by distilling the knowledge from the model itself. Recent semi-supervised object detection methods [19, 26, 45] adopt a teacher-student design, where the teacher is an exponential moving average (EMA) of the student. Self-training has been widely explored in the above fields but rarely in OVD. Unlike semi-supervised object detection, OVD encounters two challenges to use self-training, i.e., more noisy PLs, and larger changes in PLs' distributions. This work proposes the SAF head to address + +![](images/0b926dd32fbc74393bc3209e36a96ad73775294503a7653fe6185d1ed49dc07a.jpg) +Figure 1. Left: Prior PL-based methods for OVD rely on handcrafted heuristics to leverage a frozen VLM for offline pseudo labels. This is usually inefficient and does not allow for improving PLs throughout training. Right: We customize self-training and finetune VLMs for OVD, which allows efficient on-the-fly computation of PLs that can be improved throughout training. + +![](images/6c4e1bc716bd1f5a42b1ef780b8a951b2cd24bb09b6560a4379b5f8c9161d600.jpg) + +the noise, and adopts periodic updates to reduce the frequency of changes in PLs' distribution. + +# 3. Approach + +In open-vocabulary detection, an object detector is trained with bounding boxes and class labels of base categories $\mathcal{C}^B$ . At inference, the detector is used for detecting objects of open concepts including $\mathcal{C}^B$ and novel categories $\mathcal{C}^N$ , where $\mathcal{C}^B \cap \mathcal{C}^N = \emptyset$ . To make such detection practical, recent studies [8, 11, 23, 48] adopt extra data (e.g., image-text pairs and image-level tags) and/or external VLMs. We follow their settings and leverage the pretrained CLIP [30] to build our OVD detector. + +# 3.1. Adapting CLIP to OVD + +In this section, we introduce how to adapt the ResNet-based CLIP into a Faster R-CNN [32] (C4) detector. For simplicity, we do not use FPN [25], but it can be incorporated with learnable gating from Flamingo [1]. Similarly, ViT-based CLIP models can be adapted to detectors following [22]. + +Region proposals from an external RPN. CLIP is pretrained via image-text alignment. As shown in Sect. 4.3, finetuning pretrained backbones to get region proposals decreases the performance, probably because such finetuning breaks the image-text alignment learned in the pretraining. To address this problem, F-VLM [18] freezes the pretrained backbone and only finetunes detection heads. But this solution limits the capacity of the detector. Unlike F-VLM, we follow the approach of Singh et al. [36], Zhong et al. [50] and employ an external RPN to generate region proposals, which is trained with ground truth boxes of base categories. Prior studies adopt RPN mainly to accelerate inference [36] or to improve region recognition [50]. By contrast, we aim to leverage the RPN to preserve the knowledge learned in the pretraining for better self-training. + +Text embeddings as the classifier. For the $i$ -th region proposal from the external RPN, we apply RoIAlign [13] on the 4th feature maps of CLIP's ResNet to get the proposal features. Then, the features are fed to the last ResNet block and the attention pooling of CLIP to get the region embedding $r_i$ , which is later used for classification. Following prior studies [11, 23, 50], we convert a set of given concepts with prompt engineering into CLIP text embeddings, which act as classifiers, as shown in Fig. 2b. A fixed all-zero embedding is adopted for the background category. Assuming $t_c$ is the embedding of the $c$ -th category, the probability of $r_i$ to be classified as the $c$ -th category is, + +$$ +p _ {i, c} = \frac {\exp \left(\langle r _ {i} , t _ {c} \rangle / \tau\right)}{\sum_ {j = 0} ^ {C - 1} \exp \left(\langle r _ {i} , t _ {j} \rangle / \tau\right)}, \tag {1} +$$ + +where $\langle \cdot, \cdot \rangle$ denotes the cosine distance, $C$ denotes the vocabulary size, and $\tau$ denotes the temperature. In our detector, $r_i$ may be further fed to a box refinement module to predict the box shift based on the region proposal. With the above adoptions, our initial detector gains zero-shot detection ability to some extent. Please refer to the supplement for quantitative evaluations. + +# 3.2. Taming Self-Training + +Although self-training has been widely explored in closed-set object detection, using it for OVD presents two challenges, i.e., noisy PLs and frequent changes of PLs' distributions. In the following, we describe our self-training pipeline and how to address the two challenges. + +Self-training pipeline. Fig. 2a illustrates the pipeline of our self-training with a teacher-student manner. Training images are first fed into the RPN to obtain region proposals. Then, the teacher model runs inference on those proposals, where the resulting predictions with confidence above a threshold are selected as PLs. The student model adopts + +![](images/5cc705d581966c09fa8f8f0efe3217590b1081c40432853b709efcf57da96eaa.jpg) +Figure 2. (a) Pipeline of our self-training. The teacher and the student are models of the same architecture. They are initialized with the same pretrained CLIP model. The teacher generates PLs that are used to train the student, and the student updates the teacher periodically. (b) Structure of our detector. The proposed SAF head is put on top of a CLIP image encoder. The open- and closed-branches take the text embeddings from a CLIP text encoder as classifier. + +![](images/9e0e3fcd09dafb0c8586f19bd00496071e6818c8d807ca8caa442d1f7b226e05.jpg) + +the same region proposals, and is supervised with both base ground truth and PLs generated by the teacher. The teacher is periodically updated with the parameters of the student model. + +Handling noise in PLs. The noise of PLs from VLMs exists in both classification and localization. It is a common practice to filter PLs with classification confidence, but this only addresses noise in classification. To reduce the noise in locations of PLs, we exclude the noisy boxes of PLs from the training. That is, the classification loss is calculated on class labels of both ground truth and PLs, but the box regression loss is calculated only on ground truth boxes. Besides, to mitigate the impact of PLs on the performance of base categories, we propose SAF head and elaborate it in Sect. 3.3. + +Reducing changes in PLs' distributions with periodic updates. The teacher model is updated with the student to enable self-training. The exponential moving average (EMA) is a widely used approach for object detection [19, 26, 45], which updates the teacher in every iteration. But we observed empirically that it does not benefit OVD (Table 4). We hypothesize that, unlike semi-supervised object detection, OVD has no ground truth for the target categories. Thus, the distribution of the target data (i.e., PLs) is fully determined by the teacher. The EMA update changes the PLs' distribution in each iteration and leads to a constantly shifting training target that has been shown hard to optimize (i.e., destabilizing the training) in deep Q-learning [29]. As a solution, we periodically update the teacher after a set number of iterations to maintain consistent distributions for PLs between updates. We call this + +strategy as the periodic update and show it outperforms the EMA update by a large margin in Sect. 4.3. + +# 3.3. Split-and-Fusion (SAF) Head + +Our SAF head first splits a detection head into two branches, i.e. "closed-branch" and "open-branch", with the goal to better handle noisy PLs during training. At inference, predictions from both heads are fused to boost the performance. + +Splitting the detection head. The closed-branch follows a standard detection head with a classification module and a class-agnostic box refinement module. The former module classifies region proposals based on Eq. 1, and the latter one refines the proposal boxes for better locations. We train the closed-branch with boxes and class labels of ground truth for $\mathcal{C}^B$ using standard detection losses, which include a cross entropy loss for classification and a box regression loss for localization. Since no PLs are used to train the closed-branch, the noise of PLs is unlike to impact its performance on $\mathcal{C}^B$ . Moreover, as shown in Gu et al. [11], Zareian et al. [48], box regression modules trained on $\mathcal{C}^B$ can generalize to novel categories that are unseen during training. Therefore, the closed-branch is able to provide generalized boxes, as well. + +The open-branch only contains a classification module. It is trained using only the cross-entropy loss with class labels from both $\mathcal{C}^B$ and PLs, hence, learning broader concepts beyond $\mathcal{C}^B$ . Unlike distillation losses [7, 11], all losses for the two branches are originally designed for detection and are unlikely to conflict with each other. When generating PLs, we use the classification scores from the + +open-branch and bounding boxes from the closed-branch. Boxes of PLs are not directly used in our losses but are involved to select foreground proposal candidates for the classification loss. + +Fusing complementary predictions. The open- and closed-branches are trained in different ways and learn complementary knowledge. Therefore, we fuse their predictions with the geometric mean at inference time. Specifically, assuming $p_{i,c}^{\mathrm{open}}$ and $p_{i,c}^{\mathrm{closed}}$ are prediction scores of the open- and closed-branches, respectively, the final score is calculated as + +$$ +p _ {i, c} ^ {\text {f u s e d}} = \left\{ \begin{array}{l l} \left(p _ {i, c} ^ {\text {c l o s e d}}\right) ^ {(1 - \alpha)} \cdot \left(p _ {i, c} ^ {\text {o p e n}}\right) ^ {\alpha}, & \text {i f} i \in \mathcal {C} ^ {B} \\ \left(p _ {i, c} ^ {\text {c l o s e d}}\right) ^ {\alpha} \cdot \left(p _ {i, c} ^ {\text {o p e n}}\right) ^ {(1 - \alpha)}, & \text {i f} i \in \mathcal {C} ^ {N} \end{array} \right. \tag {2} +$$ + +where $\alpha \in [0,1]$ balances the two branches. The indices $i, c$ are the same as in Eq. 1. We keep a list of known base categories $\mathcal{C}^B$ and take other categories beyond the list as novel. It's important to note that fusion is a common strategy. For instance, F-VLM [18] employs score fusion to enhance the location quality of CLIP's predictions. In contrast, our approach focuses on fusing the complementary predictions from the two branches. The primary contribution of our method lies in learning what to fuse rather than the fusion process itself. + +# 4. Experiments + +# 4.1. Experiment Setup + +Datasets. We conduct experiments on the two popular OVD benchmarks COCO-OVD and LVIS-OVD, based on COCO [24] and LVIS [12], respectively. (1) COCO-OVD: 65 out of 80 COCO categories are divided into 48 base categories and 17 novel categories. We report results on the validation set. (2) LVIS-OVD: 337 rare categories are used as the novel concepts and the remaining 866 categories (frequent and common) as the base concepts. We report results on the whole LVIS validation set. + +Evaluation metrics. Following prior studies [11, 41, 48, 50], we report AP at IoU threshold of 0.5 for COCO-OVD. $\mathrm{AP}_{50}^{novel}$ , $\mathrm{AP}_{50}^{base}$ and $\mathrm{AP}_{50}^{all}$ are the metrics for the novel, the base and all (novel + base) categories, respectively. For LVIS-OVD, we report the mean AP averaged on IoUs from 0.5 to 0.95. $\mathrm{AP}_r$ , $\mathrm{AP}_c$ , $\mathrm{AP}_f$ and AP are the metrics for rare, common, frequent, and all categories. + +Implementation details. We built our detector on Faster R-CNN with the CLIP version of ResNet (RN50-C4 or RN50x4-C4) as the backbone. Following prior studies [47, 49, 51], we assumed the novel concepts were available during training and took them as our open concepts, but no annotations of novel concepts are used. Our method was implemented on Detectron2 [43] and trained with 8 NVIDIA A6000 GPUs. We trained our models with the $1 \times$ schedule (90k iterations) for COCO-OVD and the $2 \times$ schedule for + +LVIS-OVD. The batch size was 16 with an initial learning rate of 0.002. The default data augmentation of Detectron2 was applied. Loss terms were equally weighted. By default, the teacher models were updated three times, usually together with the decreases of the learning rate. + +# 4.2. Comparison with the Existing Methods + +COCO-OVD. We compare our method with prior work on COCO in Table 1. With the same backbone and detector, SAS-Det outperforms the most recent method BARON [41] by $4.3\mathrm{AP}_{50}^{novel}$ on novel categories and $3.7\mathrm{AP}_{50}^{base}$ on base categories. Probably, BARON distills knowledge from VLMs, but distillation may lead to gradient conflicts and degrade the performance [49]. Compared to VLDet [23], SAS-Det gains an improvement of $7.9\mathrm{AP}_{50}^{base}$ . It is likely that VLDet employs noisy pseudo labels to train the detection head, impacting the detection of base classes in both classification and localization. In contrast, the two-branch design of our SAF head allows to train the closed-branch with just ground-truth of base classes and reduces the impact of the noise from pseudo labels. For more analysis on where our improvements come, please refer to Sect. 4.3 and Table 3. The first block of Table 1 provides results for methods with stronger backbones or detector architectures. When compared with those methods, SAS-Det still achieves the leading performance. Those results clearly demonstrate the effectiveness of SAS-Det. + +LVIS-OVD. We provide the main results on LVIS in Table 2. When using ResNet50 as the backbone, SAS-Det achieves similar performance as the recent method DetPro [7]. DetPro proposes learnable prompts to generate better text embeddings as the classifier for OVD. Our method is orthogonal and can leverage DetPro's prompts for further improvement. The first block of Table 2 also shows that SAS-Det outperforms ViLD by a large margin on novel categories (indicated by $\mathrm{AP}_r$ ) but gets lower $\mathrm{AP}_f$ for base categories. This is also observed in the recent method BARON [41]. The performance gap is probably due to the training setup. ViLD is trained for 8 times more iterations and adopts a strong data augmentation method [10], both of which benefit detection on base categories. In contrast, BARON and our approach adopt shorter training and standard data augmentations. When replacing the pretrained CLIP ResNet50 with ResNet50x4 [30] as the backbone, we improve $\mathrm{AP}_r$ by 8.2 from 20.9 to 29.1, which demonstrates SAS-Det scales up nicely with stronger pretrained VLM backbones. + +# 4.3. Ablation Studies + +In this section, the default baseline shares the same architecture and the training as our SAS-Det (with RN50-C4 as the backbone), except that the SAF head is replaced with a single detection head. The localization module of the detec + +
MethodTraining SetupBackboneDetector\(AP^{novel}\)\(AP^{base}\)\(AP^{all}\)
ViLD [11]16×+LSJRN50-FPNFaster R-CNN27.659.551.2
VL-PLM [49]1×+DefaultRN50-FPNFaster R-CNN32.354.048.3
F-VLM [18]0.5×+LSJRN50-FPNFaster R-CNN28.0-39.6
OV-DETR [47](Not Given)RN50Deform. DETR29.461.052.7
CORA [42]3×+DefaultRN50DAB-DETR35.135.535.4
OVR-CNN [48](Not Given)RN50-C4Faster R-CNN22.846.039.9
RegionCLIP [50]1×+DefaultRN50-C4Faster R-CNN26.854.847.5
Detic [51]1×+DefaultRN50-C4Faster R-CNN27.851.145.0
PB-OVD [9]6×+DefaultRN50-C4Faster R-CNN30.846.142.1
VLDet [23]1×+LSJRN50-C4Faster R-CNN32.050.645.8
BARON [41]1×+DefaultRN50-C4Faster R-CNN33.154.849.1
SAS-Det (Ours)1×+DefaultRN50-C4Faster R-CNN37.458.553.0
+ +Table 1. Comparison with recent methods on COCO-OVD. We group methods into two blocks. The first block contains methods using stronger backbones or detector architectures than ours. The second block contains models of the same scale as ours. Training setup indicates training iterations $(\mathrm{N}\times)$ and data augmentations. Large Scale Jittering (LSJ) [10] is a stronger data augmentation than Detectron2's default. + +
MethodTraining SetupBackboneDetectorAPrAPcAPfAP
ViLD [11]16×+LSJRN50-FPNFaster R-CNN16.726.534.227.8
DetPro [7]2×+DefaultRN50-FPNFaster R-CNN20.827.832.428.4
F-VLM [18]9×+LSJRN50-FPNFaster R-CNN18.6--24.2
BARON [41]2×+DefaultRN50-C4Faster R-CNN17.325.631.026.3
SAS-Det (Ours)2×+DefaultRN50-C4Faster R-CNN20.926.131.627.4
ViLD [11]16×+LSJRN152-FPNFaster R-CNN19.827.134.528.7
OWL-ViT [28]12×+LSJViT-L/14OWL-ViT25.6--34.7
F-VLM [18]9×+LSJRN50x4-FPNFaster R-CNN26.3--28.5
CORA [42]3×+DefaultRN50x4DAB-DETR22.2---
SAS-Det (Ours)2×+DefaultRN50x4-C4Faster R-CNN29.132.436.833.5
+ +Table 2. Comparison with recent methods on LVIS-OVD. We group methods based on the scale of backbones. Training setup contains training iterations $(\mathrm{N}\times)$ and data augmentations. LSJ [10] is a stronger data augmentation than Detectron2's default. + +tion head is trained with base ground truth boxes only, but the classification module is trained with class labels of both ground truth and PLs. Thus, the baseline provides a naive solution to exclude noisy pseudo boxes from the training. All evaluations are conducted on COCO. + +External RPN. We leverage an external RPN to generate region proposals so that finetuning focuses on region-text alignment that is similar to the pretraining task. In this way, the detector is unlikely to forget the knowledge obtained in the pretraining. To validate the effectiveness of the external RPN, we follow F-VLM [18] to train a detector without an external RPN (See the supplement for more details). Table 3 compares baseline with the detector in the row of (1). As shown, without the external RPN, the performance drops on both novel and base categories. + +Removing boxes of PLs from training. Our baseline handles location noise of PLs' boxes by directly excluding them + +from the training. As shown in Table 3, baseline outperforms (2)'s detector. The only difference between them is the removal of pseudo boxes from training. This result clearly shows that simply removing the pseudo boxes is an effective way to deal with location noise. + +Splitting the detection head. As shown in Table 3, compared to baseline, the open-branch of the proposed SAF head in (5) achieves similar performance on novel categories. This indicates that the split handles the noise in PLs' location as well as the naive solution of baseline. This is plausible because both solutions exclude the noisy boxes of PLs from training. Additionally, the open-branch gains 1.7 $\mathrm{AP}_{50}^{base}$ on base categories. Compared to (3)'s, the closed-branch in (4) gain improvements in terms of both $\mathrm{AP}_{50}^{novel}$ and $\mathrm{AP}_{50}^{base}$ . Note that (4)'s and (3)'s heads are trained with the same data. Based on those results, the split benefits OVD beyond handling the location noise. One advantage + +
Ablation\(AP_{50}^{novel}\)\(AP_{50}^{base}\)\(AP_{50}^{all}\)
Baseline31.455.749.4
(1) No external RPN, train the backbone for region proposals(-6.0) 25.453.446.1
(2) Noisy boxes of PLs as supervision for box regression(-3.6) 27.855.248.0
(3) No PLs, train with base data only(-8.2) 23.256.948.1
(4) W/ SAF head, use closed-branch's predictions (\(p_{i,c}^{closed}\) in Eq. 2)(-5.8) 25.657.949.5
(5) W/ SAF head, use open-branch's predictions (\(p_{i,c}^{open}\) in Eq. 2)(+0.5) 31.957.450.7
(6) W/ SAF head, use fused predictions (\(p_{i,c}^{fused}\) in Eq. 2)(+6.0) 37.458.553.0
+ +Table 3. Ablation studies to analyze the effect of each component of SAS-Det on COCO-OVD. Results (4), (5), and (6) denote different outputs from the same model. + +
Update strategy\( AP_{50}^{novel} \)\( AP_{50}^{base} \)\( AP_{50}^{all} \)
Baseline (w/ periodic update)31.455.749.4
(7) No teacher and no pseudo labels(-8.2) 23.256.948.1
(8) Not update the teacher(-1.8) 29.655.949.1
(9) Take the EMA of the student as the teacher(-6.9) 24.553.946.2
(10) Replace the teacher with the student every iteration(-23.4) 8.053.341.5
+ +Table 4. Comparison of different strategies for updating the teacher in self-training on COCO-OVD. + +is that, due to the split, the closed-branch is less likely to be influenced by the noise of PLs and learn how to better localize objects. + +Fusing predictions. Our SAF head fuses predictions of the open-branch and the closed-branch at inference time. As shown in Table 3, (6)'s outperforms all the others by a large margin with the prediction fusion. Note that (4)'s, (5)'s and (6)'s refer to the different outputs of the same model. The improvement can be attributed to the fact that the two branches are trained on different sets of data and learn complementary knowledge. The closed-branch is trained with class labels and boxes of base categories. The open-branch is trained with class labels of more vocabularies including base and novel classes. Thus, the former learns more about how to localize objects and how to detect base classes. The latter learns more about how to detect new objects, which complements the former. + +Update strategies for the teacher model. The teacher is updated with the student model to improve PLs during finetuning. We evaluated several strategies for updating the teacher model in Table 4 and summarize our findings as follows. First, the EMA update in (9), which is shown effective and widely used in semi-supervised object detection [19, 26, 45], is just slightly better than training without PLs in (7). It is worse than no update to the teacher (i.e. no self-training) in (8). Second, if we replace the teacher with the student every iteration as (10), $\mathrm{AP}_{50}^{novel}$ drops to 8.0 that is much lower than that number in (7). These findings indicate that it's harmful to change the teacher frequently. We hypothesize that frequent updates change the distribution of + +PLs too often and makes the training unstable. Last but not least, by reducing the number of updates, our periodic update significantly outperforms the EMA update in OVD. + +The number of updates to the teacher model. We trained our detectors with different numbers of updates to teacher models. Please see the supplement for how we distribute the updates. As shown in Table 5, too many updates, e.g., 8 or 4 updates, lead to performance drops mainly due to the following. First, similar as the aforementioned EMA update, too many updates change the distribution of PLs too often and make the training unstable. Second, the more updates, the earlier an update happens. However, the student model is not well trained at the early stage of the training and thus is not good enough to update the teacher. Table 5 shows that 2 and 3 updates achieve similar performance. But we set 3 updates as default to include as many updates as possible. + +# 4.4. Further Analysis + +Evaluation on the retained 15 COCO categories. Following prior studies [47, 49, 51], SAS-Det assumes novel concepts are available during training and achieves good performance. A natural question is if SAS-Det generalizes to alien concepts that are unknown before the evaluation. To answer the question, we follow the evaluation protocol of MEDet [5] where an OVD detector is evaluated on the whole COCO validation set with all 80 COCO concepts. $\mathrm{AP}_{50}^{retain}$ , $\mathrm{AP}_{50}^{novel}$ and $\mathrm{AP}_{50}^{coco}$ are reported as APs averaged on 15 retained, 17 novel, and all 80 COCO categories, respectively. As shown in Table 6, SAS-Det achieves leading performance by a clear margin, which indicates the good + +
# Update\(AP_{50}^{novel}\)
827.6
430.6
3 (baseline)31.4
231.6
130.9
0 (No update)29.6
+ +![](images/741c4188c8a53a2cb7c3200ff4d285cdc73d68873475c97539e068f05db18e2f.jpg) +Figure 3. Quality of PLs during training. + +generalization capability of the proposed method. + +PL's quality during self-training. We compare our PLs and the prior state-of-the-art PLs from VL-PLM [49] on the COCO validation set in terms of PLs' quality. Following VL-PLM, we take $\mathrm{AP}_{50}^{novel}$ as the metric and evaluate our teacher models during training. As illustrated in Fig. 3, the quality of our initial PLs is not as good as VL-PLM's. But ours get close to VL-PLM's after two updates and become the better by the third update. This is because the teacher is initialized with CLIP, which is not pretrained for detection, and thus cannot provide high quality PLs. With updates, the teacher is equipped with the knowledge about detection and generates promising PLs. + +Time cost of pseudo labeling. We run two major PL-based methods [9, 49] on our computational environment and compare them with our pseudo labeling in terms of the time cost in Table 7. + +Reducing the time cost of external RPN. There is a simple solution to remove the extra computational cost of external RPN adopted by SAS-Det. We first generate pseudo labels with SAS-Det, and then follow VL-PLM [49] to train a Faster R-CNN based on ResNet50 for OVD without the external RPN ("Faster R-CNN + PLs"). As shown in Table 8, "Faster R-CNN + PLs" achieves slightly better or close performance as SAS-Det. + +# 5. Conclusion + +In this paper, we highlight two challenges associated with applying self-training to OVD: 1) noisy PLs from pretrained + +Table 5. Performance with varied numbers of updates to the teacher model. + +
MethodAPretain50APnovel50APcoco50
OVR-CNN [48]11.522.938.1
Detic-80 [51]11.527.338.3
MEDet [5]18.632.642.4
Ours23.037.147.2
+ +Table 6. Generalization capability of OVD methods on retained 15 COCO categories. Text embeddings of 80 COCO categories are used as the classifier. Numbers are from MEDet [5]. + +
MethodTime (s)APnovel50
PB-OVD [9]0.484818.7
VL-PLM [49]0.445625.5
Ours0.130826.7
+ +Table 7. Mean time cost to get PLs per image. The quality of PLs on the validation set of COCO-OVD are provided for reference. We report the quality of our PLs after the last update. + +
Method\( AP_{50}^{novel} \)\( AP_{50}^{base} \)\( AP_{50}^{all} \)
SAS-Det37.458.553.0
Faster R-CNN + PLs38.158.353.0
+ +Table 8. Training Faster R-CNN with PLs from SAS-Det on COCO-OVD + +VLMs and 2) frequent changes of PLs' distributions. To address the challenges, we introduce a split-and-fusion (SAF) head and implement periodic updates. The SAF head splits a standard detection head into an open-branch and a closed-branch. The open-branch is trained with the class labels of ground truth and PLs. The closed-branch is trained with both class labels and boxes of ground truth. This approach effectively eliminates the location noise of PLs during training. Furthermore, the SAF head acquires complementary knowledge from different sets of training data, enabling fusion to enhance performance. The periodic updates reduce the number of updates to the teacher models, thereby decreasing the frequency of changes in PLs' distributions. We demonstrate that the proposed method SAS-Det is both efficient and effective. Our pseudo labeling is much faster than prior PL-based methods [9, 49]. SAS-Det also outperforms recent models on both COCO and LVIS for OVD. + +Acknowledgments: This research project has been partially funded by research grants to Dimitris N. Metaxas through NSF: 2310966, 2235405, 2212301, 2003874, and FA9550-23-1-0417. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 3 +[2] Relja Arandjelović, Alex Andonian, Arthur Mensch, Olivier J Henaff, Jean-Baptiste Alayrac, and Andrew Zisserman. Three ways to improve feature alignment for open vocabulary detection, 2022. 2 +[3] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In ECCV, pages 384-400, 2018. 2 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-End Object Detection with Transformers. In ECCV, 2020. 1 +[5] Peixian Chen, Kekai Sheng, Mengdan Zhang, Yunhang Shen, Ke Li, and Chunhua Shen. Open vocabulary object detection with proposal mining and prediction equalization, 2022. 7, 8 +[6] Yuxiao Chen, Jianbo Yuan, Yu Tian, Shijie Geng, Xinyu Li, Ding Zhou, Dimitris N Metaxas, and Hongxia Yang. Revisiting multimodal representation in contrastive learning: from patch and token embeddings to finite discrete tokens. In CVPR, 2023. 2 +[7] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In CVPR, pages 14084-14093, 2022. 2, 4, 5, 6 +[8] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In ECCV, pages 701-717. Springer, 2022. 1, 2, 3 +[9] Mingfei Gao, Chen Xing, Juan Carlos Niebles, Junnan Li, Ran Xu, Wenhao Liu, and Caiming Xiong. Open vocabulary object detection with pseudo bounding-box labels. In ECCV, pages 266-282. Springer, 2022. 1, 2, 6, 8 +[10] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, pages 2918-2928, 2021. 5, 6 +[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary Object Detection via Vision and Language Knowledge Distillation. In ICLR, 2022. 1, 2, 3, 4, 5, 6 +[12] Agrim Gupta, Piotr Dólar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 5 +[13] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 1, 3 +[14] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. In ICML, 2021. 2 +[15] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR - Modulated detection for end-to-end multi-modal understanding. + +In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 2 +[16] Zaid Khan, Vijay Kumar B.G., Xiang Yu, Samuel Schulter, Manmohan Chandraker, and Yun Fu. Single-stream multilevel alignment for vision-language pretraining. In ECCV, pages 735-751. Springer, 2022. 2 +[17] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Region-aware pretraining for open-vocabulary object detection with vision transformers. In CVPR, pages 11144-11154, 2023. 2 +[18] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. In ICLR, 2023. 2, 3, 5, 6 +[19] Gang Li, Xiang Li, Yujie Wang, Yichao Wu, Ding Liang, and Shanshan Zhang. Pseudo labeling and consistency training for semi-supervised object detection. In ECCV, pages 457-472. Springer, 2022. 1, 2, 4, 7 +[20] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 1, 2 +[21] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, pages 10965-10975, 2022. 2 +[22] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, pages 280-296. Springer, 2022. 3 +[23] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In ICLR, 2023. 1, 3, 5, 6 +[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 5 +[25] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature Pyramid Networks for Object Detection. In CVPR, 2017. 3 +[26] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In ICLR, 2021. 2, 4, 7 +[27] Zhixiang Min, Bingbing Zhuang, Samuel Schulter, Buyu Liu, Enrique Dunn, and Manmohan Chandraker. Neurocs: Neural nocs supervision for monocular 3d object localization. In CVPR, pages 21404-21414, 2023. 2 +[28] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection with vision transformers. In ECCV. Springer, 2022. 2, 6 +[29] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529-533, 2015. 4 + +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 5 +[31] Shafin Rahman, Salman Khan, and Nick Barnes. Improved visual-semantic alignment for zero-shot object detection. In AAAI, pages 11932–11939, 2020. 2 +[32] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In NeurIPS, 2015. 1, 3 +[33] Zhongzheng Ren, Zhiding Yu, Xiaodong Yang, Ming-Yu Liu, Yong Jae Lee, Alexander G Schwing, and Jan Kautz. Instance-aware, context-focused, and memory-efficient weakly supervised object detection. In CVPR, pages 10598-10607, 2020. 2 +[34] Samuel Schulter, Yumin Suh, Konstantinos M Dafnis, Zhixing Zhang, Shiyu Zhao, Dimitris Metaxas, et al. Omnilabel: A challenging benchmark for language-based object detection. ICCV, 2023. 2 +[35] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2 +[36] Bharat Singh, Hengduo Li, Abhishek Sharma, and Larry S Davis. R-FCN-3000 at 30fps: Decoupling detection and classification. In CVPR, pages 1081–1090, 2018. 3 +[37] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. In arXiv:2005.04757, 2020. 1 +[38] Peng Tang, Xinggang Wang, Xiang Bai, and Wenyu Liu. Multiple instance detection network with online instance classifier refinement. In CVPR, pages 2843-2851, 2017. 1, 2 +[39] Fang Wan, Chang Liu, Wei Ke, Xiangyang Ji, Jianbin Jiao, and Qixiang Ye. C-MIL: Continuation multiple instance learning for weakly supervised object detection. In CVPR, pages 2199-2208, 2019. 2 +[40] Luting Wang, Yi Liu, Penghui Du, Zihan Ding, Yue Liao, Qiaosong Qi, Biaolong Chen, and Si Liu. Object-aware distillation pyramid for open-vocabulary object detection. In CVPR, pages 11186-11196, 2023. 2 +[41] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In CVPR, 2023. 2, 5, 6 +[42] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. CORA: Adapting CLIP for Open-Vocabulary Detection with Region Prompting and Anchor Pre-Matching. In CVPR, 2023. 1, 2, 6 +[43] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.5 + +[44] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, pages 10687-10698, 2020. 1 +[45] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In ICCV, pages 3060-3069, 2021. 1, 2, 4, 7 +[46] Caixia Yan, Xiaojun Chang, Minnan Luo, Huan Liu, Xiaoqin Zhang, and Qinghua Zheng. Semantics-guided contrastive network for zero-shot object detection. IEEE TPAMI, 2022. 2 +[47] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In ECCV, pages 106-122. Springer, 2022. 5, 6, 7 +[48] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-Vocabulary Object Detection Using Captions. In CVPR, 2021. 2, 3, 4, 5, 6, 8 +[49] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, BG Vijay Kumar, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris N Metaxas. Exploiting unlabeled data with vision and language models for object detection. In ECCV, pages 159-175. Springer, 2022. 1, 2, 5, 6, 7, 8 +[50] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. RegionCLIP: Region-based Language-Image Pretraining. In CVPR, 2022. 2, 3, 5, 6 +[51] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, pages 350-368. Springer, 2022. 5, 6, 7, 8 +[52] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Don't even look once: Synthesizing features for zero-shot detection. In CVPR, pages 11693-11702, 2020. 2 \ No newline at end of file diff --git a/2024/Taming Self-Training for Open-Vocabulary Object Detection/images.zip b/2024/Taming Self-Training for Open-Vocabulary Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f34f891d6aed47c011634541fe88c9398974d577 --- /dev/null +++ b/2024/Taming Self-Training for Open-Vocabulary Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d73468289bdc790343bae373241bbfad2ce03aa06b647a00812955f66793835 +size 559512 diff --git a/2024/Taming Self-Training for Open-Vocabulary Object Detection/layout.json b/2024/Taming Self-Training for Open-Vocabulary Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..173c21543b661ea270102b4f2f70933a3d4fe300 --- /dev/null +++ b/2024/Taming Self-Training for Open-Vocabulary Object Detection/layout.json @@ -0,0 +1,7029 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 484, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 484, + 121 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 484, + 121 + ], + "type": "text", + "content": "Taming Self-Training for Open-Vocabulary Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "spans": [ + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "type": "text", + "content": "Shiyu Zhao" + }, + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "type": "text", + "content": " Samuel Schulter" + }, + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "type": "text", + "content": " Long Zhao" + }, + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "type": "text", + "content": " Zhixing Zhang" + }, + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "type": "text", + "content": " Vijay Kumar B G" + }, + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "type": "text", + "content": " Yumin Suh" + }, + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "type": "text", + "content": " Manmohan Chandraker" + }, + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "type": "text", + "content": " Dimitris N. Metaxas" + }, + { + "bbox": [ + 88, + 142, + 519, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 184, + 526, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 184, + 526, + 200 + ], + "spans": [ + { + "bbox": [ + 64, + 184, + 526, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 184, + 526, + 200 + ], + "type": "text", + "content": " Rutgers University " + }, + { + "bbox": [ + 64, + 184, + 526, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 184, + 526, + 200 + ], + "type": "text", + "content": " NEC Laboratories America " + }, + { + "bbox": [ + 64, + 184, + 526, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 184, + 526, + 200 + ], + "type": "text", + "content": " Google Research " + }, + { + "bbox": [ + 64, + 184, + 526, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 64, + 184, + 526, + 200 + ], + "type": "text", + "content": " UC San Diego" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 252, + 290, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 574 + ], + "type": "text", + "content": "Recent studies have shown promising performance in open-vocabulary object detection (OVD) by utilizing pseudo labels (PLs) from pretrained vision and language models (VLMs). However, teacher-student self-training, a powerful and widely used paradigm to leverage PLs, is rarely explored for OVD. This work identifies two challenges of using self-training in OVD: noisy PLs from VLMs and frequent distribution changes of PLs. To address these challenges, we propose SAS-Det that tames self-training for OVD from two key perspectives. First, we present a split-and-fusion (SAF) head that splits a standard detection into an open-branch and a closed-branch. This design can reduce noisy supervision from pseudo boxes. Moreover, the two branches learn complementary knowledge from different training data, significantly enhancing performance when fused together. Second, in our view, unlike in closed-set tasks, the PL distributions in OVD are solely determined by the teacher model. We introduce a periodic update strategy to decrease the number of updates to the teacher, thereby decreasing the frequency of changes in PL distributions, which stabilizes the training process. Extensive experiments demonstrate SAS-Det is both efficient and effective. SAS-Det outperforms recent models of the same scale by a clear margin and achieves 37.4 AP50 and 29.1 APr on novel categories of the COCO and LVIS benchmarks, respectively. Code is available at https://github.com/xiaofeng94/SAS-Det." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 596, + 128, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 128, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 128, + 608 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 616, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 688 + ], + "type": "text", + "content": "Traditional closed-set object detectors [4, 13, 32] are restricted to detecting objects with a limited number of categories. Increasing the size of detection vocabularies usually requires heavy human labor to collect annotated data. With the recent advent of strong vision and language models (VLMs) [20, 30], open-vocabulary object detection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 228, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 545, + 276 + ], + "type": "text", + "content": "(OVD) [11] provides an alternative direction to approach this challenge. Typically, OVD detectors are trained with annotations of base categories and expected to generalize to novel categories with the power of pretrained VLMs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 278, + 546, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 546, + 361 + ], + "type": "text", + "content": "One promising thread of recent studies for OVD [8, 9, 42, 49] leverages VLMs to obtain pseudo labels (PLs) beyond base categories. But they rarely explore self-training, a powerful and widely used schema for utilizing PLs in closed-set tasks [37, 38, 44, 45]. We investigate this and find the vanilla self-training approach does not improve OVD performance due to the following challenges." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 364, + 546, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 546, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 546, + 604 + ], + "type": "text", + "content": "First, the typical self-training in closed-set tasks sets a confidence threshold to remove noisy PLs based on the fact that the quality of PLs is positively correlated to their confidences. However, VLMs employed in OVD are pretrained for image-level alignment with texts instead of instance-level object detection that requires the localization ability. Thus, the confidence score from pretrained VLMs is usually not a good indicator for the quality of box locations (i.e., pseudo boxes) provided by PLs. For example, prior studies [11, 49] show that CLIP [30] tends to output imperfect object boxes as predictions with high confidence. Recent methods for OVD [9, 49] just apply thresholding to VLMs' confidence scores and ignore the poor quality of pseudo boxes, which provides noisy supervision to the model. This issue becomes even worse when self-training is applied directly, since the noise accumulates which degrades the performance on novel categories. Moreover, these methods handle noisy PLs in the same way as ground truth of base categories during training, which further decreases the performance on base categories [9, 23]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "Second, self-training for closed-set object detection [19, 38, 45] usually follows an online teacher-student manner. In each training iteration, the teacher generates PLs, and the student is trained with a mixture of ground truth and PLs. Then, the teacher is updated by the student with exponential moving average (EMA). However, we find such EMA updates degrade OVD models (see Table 4). Our hypothesis is that, unlike closed-set tasks, OVD provides no ground truth for target categories, and thus, the supervision for target cat" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "* Part of this work was done during an internship at NEC Laboratories America. Correspondence to: Shiyu Zhao sz553@rutgers.edu" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13938" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "egories is fully decided by the distribution of PLs predicted by the teacher. Hence, the EMA updates change the distribution of PLs in each iteration, unstabilizing the training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 109, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 109, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 45, + 109, + 287, + 312 + ], + "type": "text", + "content": "In this paper, we propose Self-training And Split-and-fusion head for open-vocabulary Detection (SAS-Det) to tame self-training for OVD. First, we present a split-and-fusion (SAF) head to handle the noise of PLs. The SAF head splits the standard detection head into two branches: the closed-branch and the open-branch, which are fused at inference. The closed-branch, akin to the standard detection head, comprises a classification module and a box refinement module. It is supervised solely by ground truth from base categories, mitigating the impact of noisy PLs on the performance of base categories. The open-branch is a classification module supervised by class labels of both ground truth and PLs. It acquires complementary knowledge to the closed-branch and can significantly boost the performance when fused with the closed-branch. Moreover, this design circumvents noisy locations of pseudo boxes, reducing the accumulation of noise during self-training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "text", + "content": "Second, instead of adopting the vanilla EMA update, we reduce the number of the updates and periodically update the teacher by the student. The quality of our PLs improves along with the periodic updates, and our final PLs are better than those of prior PL-based methods [9, 49] that introduce external handcrafted steps. Fig. 1 shows the key differences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 384, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 384, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 45, + 384, + 287, + 538 + ], + "type": "text", + "content": "The proposed SAS-Det outperforms recent OVD models of the same scale by a clear margin on two popular benchmarks, i.e., COCO and LVIS. Without extra handcrafted steps, our pseudo labeling is more efficient than prior methods, i.e., nearly 4 times faster than PB-OVD [9] and 3 times faster than VL-PLM [49]. Extensive ablation studies demonstrate the effectiveness of the proposed components. On COCO, the thresholding of vanilla self-training decreases the performance of novel categories by 3.6 AP, and the EMA update decreases the performance by 6.9 AP. Instead, SAS-Det eliminates the degradation with two separate detection heads and the periodic update. The fusion of the SAF head boosts the performance by 6.0 AP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 539, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 659 + ], + "type": "text", + "content": "The contributions of this work are summarized as follows. (1) We show two challenges of applying self-training to OVD and propose two simple but effective solutions, i.e., using different detection heads to mitigate the noise in PLs, and using periodic updates to reduce frequency of changes in PLs' distributions. (2) The proposed SAF head for OVD handles the noisy boxes of PLs and enables fusion to improve the performance. (3) We present the leading performance on COCO and LVIS under widely used OVD settings and provide detailed analysis of the proposed SAS-Det." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Vision-language models (VLMs). VLMs are trained to learn the alignment between images and text in a com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": "mon embedding space. CLIP [30] and ALIGN [14] use contrastive losses to learn such alignment on large-scale noisy image-text pairs from the Internet. ALBEF [20] introduces multi-modal fusion and additional self-supervised objectives. SIMLA [16] employs a single stream architecture to achieve multi-level image-text alignment. FDT [6] learns shared discrete tokens as the embedding space. These VLMs achieve impressive zero-shot performance on image classification. But due to the gap between the pretraining and detection tasks, VLMs have limited abilities in object detection. In this work, we attempt to close the gap via PLs. There are studies [15, 21, 34] focusing on aligning any text phrases with objects. But they require visual grounding data that are more expensive than detection annotations. Our work scales up the vocabulary size for object detection without requiring such costly data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 267, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 578 + ], + "type": "text", + "content": "Open-vocabulary object detection (OVD). Zero-shot object detection methods [3, 31, 46, 52] increase the vocabulary size but with limited accuracy. Motivated by the strong zero-shot abilities of VLMs, recent efforts focus on OVD. Finetuning-based methods [17, 18, 28, 48, 50] add detection heads onto pretrained VLMs and then finetune the detector with concepts of base categories. Such methods are simple but may forget the knowledge learned in the pretraining [2]. Distillation-based methods [7, 11, 40, 41] introduce additional distillation loss functions that force the output of a detector to be close to that of a VLM and thus avoid forgetting. However, since the distillation losses are not designed for the detection task, they may conflict with detection objectives due to gradient conflicts that are a common issue for multi-task models. Methods like Feng et al. [8], Gao et al. [9], Wu et al. [42], Zhao et al. [49] create pseudo labels (PL) of novel concepts as supervision, and do not require extra losses, which sidesteps both catastrophic forgetting and gradient conflicts. But these methods need handcrafted steps to generate high quality PLs, e.g. multiple runs of box regression [49], activation maps [9] from Grad-CAM [35], image retrieval [8], or multi-stage training [42]. In this work, we address the two challenges of using self-training for OVD and enable an efficient end-to-end pseudo labeling pipeline. Besides, we point out the noise due to the poor locations of PLs and introduce the SAF head to handle such noise." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Self-training for object detection. Weakly-supervised object detection methods [27, 33, 38, 39] usually explore online self-training by distilling the knowledge from the model itself. Recent semi-supervised object detection methods [19, 26, 45] adopt a teacher-student design, where the teacher is an exponential moving average (EMA) of the student. Self-training has been widely explored in the above fields but rarely in OVD. Unlike semi-supervised object detection, OVD encounters two challenges to use self-training, i.e., more noisy PLs, and larger changes in PLs' distributions. This work proposes the SAF head to address" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13939" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 73, + 299, + 218 + ], + "blocks": [ + { + "bbox": [ + 48, + 73, + 299, + 218 + ], + "lines": [ + { + "bbox": [ + 48, + 73, + 299, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 299, + 218 + ], + "type": "image", + "image_path": "0b926dd32fbc74393bc3209e36a96ad73775294503a7653fe6185d1ed49dc07a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 226, + 546, + 260 + ], + "lines": [ + { + "bbox": [ + 46, + 226, + 546, + 260 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 546, + 260 + ], + "type": "text", + "content": "Figure 1. Left: Prior PL-based methods for OVD rely on handcrafted heuristics to leverage a frozen VLM for offline pseudo labels. This is usually inefficient and does not allow for improving PLs throughout training. Right: We customize self-training and finetune VLMs for OVD, which allows efficient on-the-fly computation of PLs that can be improved throughout training." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 311, + 74, + 545, + 216 + ], + "blocks": [ + { + "bbox": [ + 311, + 74, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 311, + 74, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 311, + 74, + 545, + 216 + ], + "type": "image", + "image_path": "6c4e1bc716bd1f5a42b1ef780b8a951b2cd24bb09b6560a4379b5f8c9161d600.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "text", + "content": "the noise, and adopts periodic updates to reduce the frequency of changes in PLs' distribution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 316, + 113, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 113, + 329 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 113, + 329 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 336, + 288, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 288, + 443 + ], + "type": "text", + "content": "In open-vocabulary detection, an object detector is trained with bounding boxes and class labels of base categories " + }, + { + "bbox": [ + 46, + 336, + 288, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^B" + }, + { + "bbox": [ + 46, + 336, + 288, + 443 + ], + "type": "text", + "content": ". At inference, the detector is used for detecting objects of open concepts including " + }, + { + "bbox": [ + 46, + 336, + 288, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^B" + }, + { + "bbox": [ + 46, + 336, + 288, + 443 + ], + "type": "text", + "content": " and novel categories " + }, + { + "bbox": [ + 46, + 336, + 288, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^N" + }, + { + "bbox": [ + 46, + 336, + 288, + 443 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 336, + 288, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^B \\cap \\mathcal{C}^N = \\emptyset" + }, + { + "bbox": [ + 46, + 336, + 288, + 443 + ], + "type": "text", + "content": ". To make such detection practical, recent studies [8, 11, 23, 48] adopt extra data (e.g., image-text pairs and image-level tags) and/or external VLMs. We follow their settings and leverage the pretrained CLIP [30] to build our OVD detector." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 453, + 181, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 181, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 181, + 466 + ], + "type": "text", + "content": "3.1. Adapting CLIP to OVD" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 473, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 533 + ], + "type": "text", + "content": "In this section, we introduce how to adapt the ResNet-based CLIP into a Faster R-CNN [32] (C4) detector. For simplicity, we do not use FPN [25], but it can be incorporated with learnable gating from Flamingo [1]. Similarly, ViT-based CLIP models can be adapted to detectors following [22]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "Region proposals from an external RPN. CLIP is pretrained via image-text alignment. As shown in Sect. 4.3, finetuning pretrained backbones to get region proposals decreases the performance, probably because such finetuning breaks the image-text alignment learned in the pretraining. To address this problem, F-VLM [18] freezes the pretrained backbone and only finetunes detection heads. But this solution limits the capacity of the detector. Unlike F-VLM, we follow the approach of Singh et al. [36], Zhong et al. [50] and employ an external RPN to generate region proposals, which is trained with ground truth boxes of base categories. Prior studies adopt RPN mainly to accelerate inference [36] or to improve region recognition [50]. By contrast, we aim to leverage the RPN to preserve the knowledge learned in the pretraining for better self-training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 277, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 422 + ], + "type": "text", + "content": "Text embeddings as the classifier. For the " + }, + { + "bbox": [ + 304, + 277, + 545, + 422 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 277, + 545, + 422 + ], + "type": "text", + "content": "-th region proposal from the external RPN, we apply RoIAlign [13] on the 4th feature maps of CLIP's ResNet to get the proposal features. Then, the features are fed to the last ResNet block and the attention pooling of CLIP to get the region embedding " + }, + { + "bbox": [ + 304, + 277, + 545, + 422 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 304, + 277, + 545, + 422 + ], + "type": "text", + "content": ", which is later used for classification. Following prior studies [11, 23, 50], we convert a set of given concepts with prompt engineering into CLIP text embeddings, which act as classifiers, as shown in Fig. 2b. A fixed all-zero embedding is adopted for the background category. Assuming " + }, + { + "bbox": [ + 304, + 277, + 545, + 422 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 304, + 277, + 545, + 422 + ], + "type": "text", + "content": " is the embedding of the " + }, + { + "bbox": [ + 304, + 277, + 545, + 422 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 277, + 545, + 422 + ], + "type": "text", + "content": "-th category, the probability of " + }, + { + "bbox": [ + 304, + 277, + 545, + 422 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 304, + 277, + 545, + 422 + ], + "type": "text", + "content": " to be classified as the " + }, + { + "bbox": [ + 304, + 277, + 545, + 422 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 277, + 545, + 422 + ], + "type": "text", + "content": "-th category is," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 361, + 430, + 545, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 430, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 361, + 430, + 545, + 460 + ], + "type": "interline_equation", + "content": "p _ {i, c} = \\frac {\\exp \\left(\\langle r _ {i} , t _ {c} \\rangle / \\tau\\right)}{\\sum_ {j = 0} ^ {C - 1} \\exp \\left(\\langle r _ {i} , t _ {j} \\rangle / \\tau\\right)}, \\tag {1}", + "image_path": "1209ea5f962fe68d1d510b5fd15973c0c1fb162c518f043f4c555bdde554cc7d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 469, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 545, + 554 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 469, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle" + }, + { + "bbox": [ + 304, + 469, + 545, + 554 + ], + "type": "text", + "content": " denotes the cosine distance, " + }, + { + "bbox": [ + 304, + 469, + 545, + 554 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 469, + 545, + 554 + ], + "type": "text", + "content": " denotes the vocabulary size, and " + }, + { + "bbox": [ + 304, + 469, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 469, + 545, + 554 + ], + "type": "text", + "content": " denotes the temperature. In our detector, " + }, + { + "bbox": [ + 304, + 469, + 545, + 554 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 304, + 469, + 545, + 554 + ], + "type": "text", + "content": " may be further fed to a box refinement module to predict the box shift based on the region proposal. With the above adoptions, our initial detector gains zero-shot detection ability to some extent. Please refer to the supplement for quantitative evaluations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 562, + 429, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 429, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 429, + 575 + ], + "type": "text", + "content": "3.2. Taming Self-Training" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 581, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 640 + ], + "type": "text", + "content": "Although self-training has been widely explored in closed-set object detection, using it for OVD presents two challenges, i.e., noisy PLs and frequent changes of PLs' distributions. In the following, we describe our self-training pipeline and how to address the two challenges." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "content": "Self-training pipeline. Fig. 2a illustrates the pipeline of our self-training with a teacher-student manner. Training images are first fed into the RPN to obtain region proposals. Then, the teacher model runs inference on those proposals, where the resulting predictions with confidence above a threshold are selected as PLs. The student model adopts" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13940" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 353, + 255 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 353, + 255 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 353, + 255 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 353, + 255 + ], + "type": "image", + "image_path": "5cc705d581966c09fa8f8f0efe3217590b1081c40432853b709efcf57da96eaa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 266, + 546, + 310 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 546, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 546, + 310 + ], + "type": "text", + "content": "Figure 2. (a) Pipeline of our self-training. The teacher and the student are models of the same architecture. They are initialized with the same pretrained CLIP model. The teacher generates PLs that are used to train the student, and the student updates the teacher periodically. (b) Structure of our detector. The proposed SAF head is put on top of a CLIP image encoder. The open- and closed-branches take the text embeddings from a CLIP text encoder as classifier." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 356, + 71, + 544, + 255 + ], + "blocks": [ + { + "bbox": [ + 356, + 71, + 544, + 255 + ], + "lines": [ + { + "bbox": [ + 356, + 71, + 544, + 255 + ], + "spans": [ + { + "bbox": [ + 356, + 71, + 544, + 255 + ], + "type": "image", + "image_path": "9e0e3fcd09dafb0c8586f19bd00496071e6818c8d807ca8caa442d1f7b226e05.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 331, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 378 + ], + "type": "text", + "content": "the same region proposals, and is supervised with both base ground truth and PLs generated by the teacher. The teacher is periodically updated with the parameters of the student model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 384, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 515 + ], + "type": "text", + "content": "Handling noise in PLs. The noise of PLs from VLMs exists in both classification and localization. It is a common practice to filter PLs with classification confidence, but this only addresses noise in classification. To reduce the noise in locations of PLs, we exclude the noisy boxes of PLs from the training. That is, the classification loss is calculated on class labels of both ground truth and PLs, but the box regression loss is calculated only on ground truth boxes. Besides, to mitigate the impact of PLs on the performance of base categories, we propose SAF head and elaborate it in Sect. 3.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "type": "text", + "content": "Reducing changes in PLs' distributions with periodic updates. The teacher model is updated with the student to enable self-training. The exponential moving average (EMA) is a widely used approach for object detection [19, 26, 45], which updates the teacher in every iteration. But we observed empirically that it does not benefit OVD (Table 4). We hypothesize that, unlike semi-supervised object detection, OVD has no ground truth for the target categories. Thus, the distribution of the target data (i.e., PLs) is fully determined by the teacher. The EMA update changes the PLs' distribution in each iteration and leads to a constantly shifting training target that has been shown hard to optimize (i.e., destabilizing the training) in deep Q-learning [29]. As a solution, we periodically update the teacher after a set number of iterations to maintain consistent distributions for PLs between updates. We call this" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 331, + 544, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 544, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 544, + 357 + ], + "type": "text", + "content": "strategy as the periodic update and show it outperforms the EMA update by a large margin in Sect. 4.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 367, + 465, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 367, + 465, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 367, + 465, + 380 + ], + "type": "text", + "content": "3.3. Split-and-Fusion (SAF) Head" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 387, + 545, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 387, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 305, + 387, + 545, + 446 + ], + "type": "text", + "content": "Our SAF head first splits a detection head into two branches, i.e. \"closed-branch\" and \"open-branch\", with the goal to better handle noisy PLs during training. At inference, predictions from both heads are fused to boost the performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "text", + "content": "Splitting the detection head. The closed-branch follows a standard detection head with a classification module and a class-agnostic box refinement module. The former module classifies region proposals based on Eq. 1, and the latter one refines the proposal boxes for better locations. We train the closed-branch with boxes and class labels of ground truth for " + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^B" + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "text", + "content": " using standard detection losses, which include a cross entropy loss for classification and a box regression loss for localization. Since no PLs are used to train the closed-branch, the noise of PLs is unlike to impact its performance on " + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^B" + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "text", + "content": ". Moreover, as shown in Gu et al. [11], Zareian et al. [48], box regression modules trained on " + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^B" + }, + { + "bbox": [ + 304, + 449, + 545, + 628 + ], + "type": "text", + "content": " can generalize to novel categories that are unseen during training. Therefore, the closed-branch is able to provide generalized boxes, as well." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "The open-branch only contains a classification module. It is trained using only the cross-entropy loss with class labels from both " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^B" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and PLs, hence, learning broader concepts beyond " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^B" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". Unlike distillation losses [7, 11], all losses for the two branches are originally designed for detection and are unlikely to conflict with each other. When generating PLs, we use the classification scores from the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13941" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 119 + ], + "type": "text", + "content": "open-branch and bounding boxes from the closed-branch. Boxes of PLs are not directly used in our losses but are involved to select foreground proposal candidates for the classification loss." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 286, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 286, + 202 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 286, + 202 + ], + "type": "text", + "content": "Fusing complementary predictions. The open- and closed-branches are trained in different ways and learn complementary knowledge. Therefore, we fuse their predictions with the geometric mean at inference time. Specifically, assuming " + }, + { + "bbox": [ + 46, + 120, + 286, + 202 + ], + "type": "inline_equation", + "content": "p_{i,c}^{\\mathrm{open}}" + }, + { + "bbox": [ + 46, + 120, + 286, + 202 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 120, + 286, + 202 + ], + "type": "inline_equation", + "content": "p_{i,c}^{\\mathrm{closed}}" + }, + { + "bbox": [ + 46, + 120, + 286, + 202 + ], + "type": "text", + "content": " are prediction scores of the open- and closed-branches, respectively, the final score is calculated as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 209, + 286, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 209, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 209, + 286, + 240 + ], + "type": "interline_equation", + "content": "p _ {i, c} ^ {\\text {f u s e d}} = \\left\\{ \\begin{array}{l l} \\left(p _ {i, c} ^ {\\text {c l o s e d}}\\right) ^ {(1 - \\alpha)} \\cdot \\left(p _ {i, c} ^ {\\text {o p e n}}\\right) ^ {\\alpha}, & \\text {i f} i \\in \\mathcal {C} ^ {B} \\\\ \\left(p _ {i, c} ^ {\\text {c l o s e d}}\\right) ^ {\\alpha} \\cdot \\left(p _ {i, c} ^ {\\text {o p e n}}\\right) ^ {(1 - \\alpha)}, & \\text {i f} i \\in \\mathcal {C} ^ {N} \\end{array} \\right. \\tag {2}", + "image_path": "7a6ae07ecd73537965c1e2774330988ae9d82e69cc3a75d8e1bc0bb1ff253b2a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "type": "text", + "content": " balances the two branches. The indices " + }, + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "type": "inline_equation", + "content": "i, c" + }, + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "type": "text", + "content": " are the same as in Eq. 1. We keep a list of known base categories " + }, + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^B" + }, + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "type": "text", + "content": " and take other categories beyond the list as novel. It's important to note that fusion is a common strategy. For instance, F-VLM [18] employs score fusion to enhance the location quality of CLIP's predictions. In contrast, our approach focuses on fusing the complementary predictions from the two branches. The primary contribution of our method lies in learning what to fuse rather than the fusion process itself." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 376, + 127, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 127, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 127, + 389 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 396, + 154, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 154, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 154, + 408 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 414, + 286, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 286, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 286, + 521 + ], + "type": "text", + "content": "Datasets. We conduct experiments on the two popular OVD benchmarks COCO-OVD and LVIS-OVD, based on COCO [24] and LVIS [12], respectively. (1) COCO-OVD: 65 out of 80 COCO categories are divided into 48 base categories and 17 novel categories. We report results on the validation set. (2) LVIS-OVD: 337 rare categories are used as the novel concepts and the remaining 866 categories (frequent and common) as the base concepts. We report results on the whole LVIS validation set." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "type": "text", + "content": "Evaluation metrics. Following prior studies [11, 41, 48, 50], we report AP at IoU threshold of 0.5 for COCO-OVD. " + }, + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}^{novel}" + }, + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}^{base}" + }, + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}^{all}" + }, + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "type": "text", + "content": " are the metrics for the novel, the base and all (novel + base) categories, respectively. For LVIS-OVD, we report the mean AP averaged on IoUs from 0.5 to 0.95. " + }, + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_c" + }, + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_f" + }, + { + "bbox": [ + 46, + 521, + 286, + 605 + ], + "type": "text", + "content": " and AP are the metrics for rare, common, frequent, and all categories." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 712 + ], + "type": "text", + "content": "Implementation details. We built our detector on Faster R-CNN with the CLIP version of ResNet (RN50-C4 or RN50x4-C4) as the backbone. Following prior studies [47, 49, 51], we assumed the novel concepts were available during training and took them as our open concepts, but no annotations of novel concepts are used. Our method was implemented on Detectron2 [43] and trained with 8 NVIDIA A6000 GPUs. We trained our models with the " + }, + { + "bbox": [ + 46, + 605, + 286, + 712 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 46, + 605, + 286, + 712 + ], + "type": "text", + "content": " schedule (90k iterations) for COCO-OVD and the " + }, + { + "bbox": [ + 46, + 605, + 286, + 712 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 46, + 605, + 286, + 712 + ], + "type": "text", + "content": " schedule for" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "LVIS-OVD. The batch size was 16 with an initial learning rate of 0.002. The default data augmentation of Detectron2 was applied. Loss terms were equally weighted. By default, the teacher models were updated three times, usually together with the decreases of the learning rate." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 140, + 511, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 140, + 511, + 153 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 511, + 153 + ], + "type": "text", + "content": "4.2. Comparison with the Existing Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 159, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 397 + ], + "type": "text", + "content": "COCO-OVD. We compare our method with prior work on COCO in Table 1. With the same backbone and detector, SAS-Det outperforms the most recent method BARON [41] by " + }, + { + "bbox": [ + 304, + 159, + 545, + 397 + ], + "type": "inline_equation", + "content": "4.3\\mathrm{AP}_{50}^{novel}" + }, + { + "bbox": [ + 304, + 159, + 545, + 397 + ], + "type": "text", + "content": " on novel categories and " + }, + { + "bbox": [ + 304, + 159, + 545, + 397 + ], + "type": "inline_equation", + "content": "3.7\\mathrm{AP}_{50}^{base}" + }, + { + "bbox": [ + 304, + 159, + 545, + 397 + ], + "type": "text", + "content": " on base categories. Probably, BARON distills knowledge from VLMs, but distillation may lead to gradient conflicts and degrade the performance [49]. Compared to VLDet [23], SAS-Det gains an improvement of " + }, + { + "bbox": [ + 304, + 159, + 545, + 397 + ], + "type": "inline_equation", + "content": "7.9\\mathrm{AP}_{50}^{base}" + }, + { + "bbox": [ + 304, + 159, + 545, + 397 + ], + "type": "text", + "content": ". It is likely that VLDet employs noisy pseudo labels to train the detection head, impacting the detection of base classes in both classification and localization. In contrast, the two-branch design of our SAF head allows to train the closed-branch with just ground-truth of base classes and reduces the impact of the noise from pseudo labels. For more analysis on where our improvements come, please refer to Sect. 4.3 and Table 3. The first block of Table 1 provides results for methods with stronger backbones or detector architectures. When compared with those methods, SAS-Det still achieves the leading performance. Those results clearly demonstrate the effectiveness of SAS-Det." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 399, + 545, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 637 + ], + "type": "text", + "content": "LVIS-OVD. We provide the main results on LVIS in Table 2. When using ResNet50 as the backbone, SAS-Det achieves similar performance as the recent method DetPro [7]. DetPro proposes learnable prompts to generate better text embeddings as the classifier for OVD. Our method is orthogonal and can leverage DetPro's prompts for further improvement. The first block of Table 2 also shows that SAS-Det outperforms ViLD by a large margin on novel categories (indicated by " + }, + { + "bbox": [ + 304, + 399, + 545, + 637 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 304, + 399, + 545, + 637 + ], + "type": "text", + "content": ") but gets lower " + }, + { + "bbox": [ + 304, + 399, + 545, + 637 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_f" + }, + { + "bbox": [ + 304, + 399, + 545, + 637 + ], + "type": "text", + "content": " for base categories. This is also observed in the recent method BARON [41]. The performance gap is probably due to the training setup. ViLD is trained for 8 times more iterations and adopts a strong data augmentation method [10], both of which benefit detection on base categories. In contrast, BARON and our approach adopt shorter training and standard data augmentations. When replacing the pretrained CLIP ResNet50 with ResNet50x4 [30] as the backbone, we improve " + }, + { + "bbox": [ + 304, + 399, + 545, + 637 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 304, + 399, + 545, + 637 + ], + "type": "text", + "content": " by 8.2 from 20.9 to 29.1, which demonstrates SAS-Det scales up nicely with stronger pretrained VLM backbones." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 647, + 406, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 406, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 406, + 658 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "type": "text", + "content": "In this section, the default baseline shares the same architecture and the training as our SAS-Det (with RN50-C4 as the backbone), except that the SAF head is replaced with a single detection head. The localization module of the detec" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13942" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 70, + 502, + 243 + ], + "blocks": [ + { + "bbox": [ + 93, + 70, + 502, + 243 + ], + "lines": [ + { + "bbox": [ + 93, + 70, + 502, + 243 + ], + "spans": [ + { + "bbox": [ + 93, + 70, + 502, + 243 + ], + "type": "table", + "html": "
MethodTraining SetupBackboneDetector\\(AP^{novel}\\)\\(AP^{base}\\)\\(AP^{all}\\)
ViLD [11]16×+LSJRN50-FPNFaster R-CNN27.659.551.2
VL-PLM [49]1×+DefaultRN50-FPNFaster R-CNN32.354.048.3
F-VLM [18]0.5×+LSJRN50-FPNFaster R-CNN28.0-39.6
OV-DETR [47](Not Given)RN50Deform. DETR29.461.052.7
CORA [42]3×+DefaultRN50DAB-DETR35.135.535.4
OVR-CNN [48](Not Given)RN50-C4Faster R-CNN22.846.039.9
RegionCLIP [50]1×+DefaultRN50-C4Faster R-CNN26.854.847.5
Detic [51]1×+DefaultRN50-C4Faster R-CNN27.851.145.0
PB-OVD [9]6×+DefaultRN50-C4Faster R-CNN30.846.142.1
VLDet [23]1×+LSJRN50-C4Faster R-CNN32.050.645.8
BARON [41]1×+DefaultRN50-C4Faster R-CNN33.154.849.1
SAS-Det (Ours)1×+DefaultRN50-C4Faster R-CNN37.458.553.0
", + "image_path": "98b104303f5583abde06d92387f2bd32c50dd0da45ffd112f09e91a3cdd4ca61.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 93, + 297, + 501, + 446 + ], + "blocks": [ + { + "bbox": [ + 46, + 252, + 547, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 252, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 547, + 288 + ], + "type": "text", + "content": "Table 1. Comparison with recent methods on COCO-OVD. We group methods into two blocks. The first block contains methods using stronger backbones or detector architectures than ours. The second block contains models of the same scale as ours. Training setup indicates training iterations " + }, + { + "bbox": [ + 46, + 252, + 547, + 288 + ], + "type": "inline_equation", + "content": "(\\mathrm{N}\\times)" + }, + { + "bbox": [ + 46, + 252, + 547, + 288 + ], + "type": "text", + "content": " and data augmentations. Large Scale Jittering (LSJ) [10] is a stronger data augmentation than Detectron2's default." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 297, + 501, + 446 + ], + "lines": [ + { + "bbox": [ + 93, + 297, + 501, + 446 + ], + "spans": [ + { + "bbox": [ + 93, + 297, + 501, + 446 + ], + "type": "table", + "html": "
MethodTraining SetupBackboneDetectorAPrAPcAPfAP
ViLD [11]16×+LSJRN50-FPNFaster R-CNN16.726.534.227.8
DetPro [7]2×+DefaultRN50-FPNFaster R-CNN20.827.832.428.4
F-VLM [18]9×+LSJRN50-FPNFaster R-CNN18.6--24.2
BARON [41]2×+DefaultRN50-C4Faster R-CNN17.325.631.026.3
SAS-Det (Ours)2×+DefaultRN50-C4Faster R-CNN20.926.131.627.4
ViLD [11]16×+LSJRN152-FPNFaster R-CNN19.827.134.528.7
OWL-ViT [28]12×+LSJViT-L/14OWL-ViT25.6--34.7
F-VLM [18]9×+LSJRN50x4-FPNFaster R-CNN26.3--28.5
CORA [42]3×+DefaultRN50x4DAB-DETR22.2---
SAS-Det (Ours)2×+DefaultRN50x4-C4Faster R-CNN29.132.436.833.5
", + "image_path": "6ac22963d90b65c019fa23bc3656f64e11893c09d4e8b7d0acb73846d3af22bd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 454, + 546, + 478 + ], + "lines": [ + { + "bbox": [ + 46, + 454, + 546, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 546, + 478 + ], + "type": "text", + "content": "Table 2. Comparison with recent methods on LVIS-OVD. We group methods based on the scale of backbones. Training setup contains training iterations " + }, + { + "bbox": [ + 46, + 454, + 546, + 478 + ], + "type": "inline_equation", + "content": "(\\mathrm{N}\\times)" + }, + { + "bbox": [ + 46, + 454, + 546, + 478 + ], + "type": "text", + "content": " and data augmentations. LSJ [10] is a stronger data augmentation than Detectron2's default." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 498, + 288, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 558 + ], + "type": "text", + "content": "tion head is trained with base ground truth boxes only, but the classification module is trained with class labels of both ground truth and PLs. Thus, the baseline provides a naive solution to exclude noisy pseudo boxes from the training. All evaluations are conducted on COCO." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 563, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 684 + ], + "type": "text", + "content": "External RPN. We leverage an external RPN to generate region proposals so that finetuning focuses on region-text alignment that is similar to the pretraining task. In this way, the detector is unlikely to forget the knowledge obtained in the pretraining. To validate the effectiveness of the external RPN, we follow F-VLM [18] to train a detector without an external RPN (See the supplement for more details). Table 3 compares baseline with the detector in the row of (1). As shown, without the external RPN, the performance drops on both novel and base categories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Removing boxes of PLs from training. Our baseline handles location noise of PLs' boxes by directly excluding them" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 498, + 547, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 559 + ], + "type": "text", + "content": "from the training. As shown in Table 3, baseline outperforms (2)'s detector. The only difference between them is the removal of pseudo boxes from training. This result clearly shows that simply removing the pseudo boxes is an effective way to deal with location noise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": "Splitting the detection head. As shown in Table 3, compared to baseline, the open-branch of the proposed SAF head in (5) achieves similar performance on novel categories. This indicates that the split handles the noise in PLs' location as well as the naive solution of baseline. This is plausible because both solutions exclude the noisy boxes of PLs from training. Additionally, the open-branch gains 1.7 " + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}^{base}" + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": " on base categories. Compared to (3)'s, the closed-branch in (4) gain improvements in terms of both " + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}^{novel}" + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}^{base}" + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": ". Note that (4)'s and (3)'s heads are trained with the same data. Based on those results, the split benefits OVD beyond handling the location noise. One advantage" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13943" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 90, + 70, + 504, + 180 + ], + "blocks": [ + { + "bbox": [ + 90, + 70, + 504, + 180 + ], + "lines": [ + { + "bbox": [ + 90, + 70, + 504, + 180 + ], + "spans": [ + { + "bbox": [ + 90, + 70, + 504, + 180 + ], + "type": "table", + "html": "
Ablation\\(AP_{50}^{novel}\\)\\(AP_{50}^{base}\\)\\(AP_{50}^{all}\\)
Baseline31.455.749.4
(1) No external RPN, train the backbone for region proposals(-6.0) 25.453.446.1
(2) Noisy boxes of PLs as supervision for box regression(-3.6) 27.855.248.0
(3) No PLs, train with base data only(-8.2) 23.256.948.1
(4) W/ SAF head, use closed-branch's predictions (\\(p_{i,c}^{closed}\\) in Eq. 2)(-5.8) 25.657.949.5
(5) W/ SAF head, use open-branch's predictions (\\(p_{i,c}^{open}\\) in Eq. 2)(+0.5) 31.957.450.7
(6) W/ SAF head, use fused predictions (\\(p_{i,c}^{fused}\\) in Eq. 2)(+6.0) 37.458.553.0
", + "image_path": "a28e0195b712f8c587344d3d09bda332198e8d7f63c25dc933ca7e49390e52e2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 113, + 221, + 481, + 308 + ], + "blocks": [ + { + "bbox": [ + 46, + 189, + 547, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 189, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 547, + 212 + ], + "type": "text", + "content": "Table 3. Ablation studies to analyze the effect of each component of SAS-Det on COCO-OVD. Results (4), (5), and (6) denote different outputs from the same model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 113, + 221, + 481, + 308 + ], + "lines": [ + { + "bbox": [ + 113, + 221, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 113, + 221, + 481, + 308 + ], + "type": "table", + "html": "
Update strategy\\( AP_{50}^{novel} \\)\\( AP_{50}^{base} \\)\\( AP_{50}^{all} \\)
Baseline (w/ periodic update)31.455.749.4
(7) No teacher and no pseudo labels(-8.2) 23.256.948.1
(8) Not update the teacher(-1.8) 29.655.949.1
(9) Take the EMA of the student as the teacher(-6.9) 24.553.946.2
(10) Replace the teacher with the student every iteration(-23.4) 8.053.341.5
", + "image_path": "9ced908a91ec468f26453f288d32a4ce1b18b8c118ca1d792e92eba2eb392555.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 316, + 478, + 327 + ], + "lines": [ + { + "bbox": [ + 115, + 316, + 478, + 327 + ], + "spans": [ + { + "bbox": [ + 115, + 316, + 478, + 327 + ], + "type": "text", + "content": "Table 4. Comparison of different strategies for updating the teacher in self-training on COCO-OVD." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 348, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 384 + ], + "type": "text", + "content": "is that, due to the split, the closed-branch is less likely to be influenced by the noise of PLs and learn how to better localize objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 387, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 554 + ], + "type": "text", + "content": "Fusing predictions. Our SAF head fuses predictions of the open-branch and the closed-branch at inference time. As shown in Table 3, (6)'s outperforms all the others by a large margin with the prediction fusion. Note that (4)'s, (5)'s and (6)'s refer to the different outputs of the same model. The improvement can be attributed to the fact that the two branches are trained on different sets of data and learn complementary knowledge. The closed-branch is trained with class labels and boxes of base categories. The open-branch is trained with class labels of more vocabularies including base and novel classes. Thus, the former learns more about how to localize objects and how to detect base classes. The latter learns more about how to detect new objects, which complements the former." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "Update strategies for the teacher model. The teacher is updated with the student model to improve PLs during finetuning. We evaluated several strategies for updating the teacher model in Table 4 and summarize our findings as follows. First, the EMA update in (9), which is shown effective and widely used in semi-supervised object detection [19, 26, 45], is just slightly better than training without PLs in (7). It is worse than no update to the teacher (i.e. no self-training) in (8). Second, if we replace the teacher with the student every iteration as (10), " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}^{novel}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " drops to 8.0 that is much lower than that number in (7). These findings indicate that it's harmful to change the teacher frequently. We hypothesize that frequent updates change the distribution of" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 348, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 384 + ], + "type": "text", + "content": "PLs too often and makes the training unstable. Last but not least, by reducing the number of updates, our periodic update significantly outperforms the EMA update in OVD." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 384, + 546, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 546, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 546, + 540 + ], + "type": "text", + "content": "The number of updates to the teacher model. We trained our detectors with different numbers of updates to teacher models. Please see the supplement for how we distribute the updates. As shown in Table 5, too many updates, e.g., 8 or 4 updates, lead to performance drops mainly due to the following. First, similar as the aforementioned EMA update, too many updates change the distribution of PLs too often and make the training unstable. Second, the more updates, the earlier an update happens. However, the student model is not well trained at the early stage of the training and thus is not good enough to update the teacher. Table 5 shows that 2 and 3 updates achieve similar performance. But we set 3 updates as default to include as many updates as possible." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 550, + 408, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 408, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 408, + 563 + ], + "type": "text", + "content": "4.4. Further Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Evaluation on the retained 15 COCO categories. Following prior studies [47, 49, 51], SAS-Det assumes novel concepts are available during training and achieves good performance. A natural question is if SAS-Det generalizes to alien concepts that are unknown before the evaluation. To answer the question, we follow the evaluation protocol of MEDet [5] where an OVD detector is evaluated on the whole COCO validation set with all 80 COCO concepts. " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}^{retain}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}^{novel}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}^{coco}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " are reported as APs averaged on 15 retained, 17 novel, and all 80 COCO categories, respectively. As shown in Table 6, SAS-Det achieves leading performance by a clear margin, which indicates the good" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13944" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 103, + 71, + 218, + 168 + ], + "blocks": [ + { + "bbox": [ + 103, + 71, + 218, + 168 + ], + "lines": [ + { + "bbox": [ + 103, + 71, + 218, + 168 + ], + "spans": [ + { + "bbox": [ + 103, + 71, + 218, + 168 + ], + "type": "table", + "html": "
# Update\\(AP_{50}^{novel}\\)
827.6
430.6
3 (baseline)31.4
231.6
130.9
0 (No update)29.6
", + "image_path": "cea351ccf90b526b49b4eced7c1fa065342cce78457928ab399be8b4cbfb3e63.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 214, + 278, + 319 + ], + "blocks": [ + { + "bbox": [ + 63, + 214, + 278, + 319 + ], + "lines": [ + { + "bbox": [ + 63, + 214, + 278, + 319 + ], + "spans": [ + { + "bbox": [ + 63, + 214, + 278, + 319 + ], + "type": "image", + "image_path": "741c4188c8a53a2cb7c3200ff4d285cdc73d68873475c97539e068f05db18e2f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 328, + 247, + 340 + ], + "lines": [ + { + "bbox": [ + 97, + 328, + 247, + 340 + ], + "spans": [ + { + "bbox": [ + 97, + 328, + 247, + 340 + ], + "type": "text", + "content": "Figure 3. Quality of PLs during training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 247, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 247, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 247, + 373 + ], + "type": "text", + "content": "generalization capability of the proposed method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 373, + 287, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 516 + ], + "type": "text", + "content": "PL's quality during self-training. We compare our PLs and the prior state-of-the-art PLs from VL-PLM [49] on the COCO validation set in terms of PLs' quality. Following VL-PLM, we take " + }, + { + "bbox": [ + 46, + 373, + 287, + 516 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}^{novel}" + }, + { + "bbox": [ + 46, + 373, + 287, + 516 + ], + "type": "text", + "content": " as the metric and evaluate our teacher models during training. As illustrated in Fig. 3, the quality of our initial PLs is not as good as VL-PLM's. But ours get close to VL-PLM's after two updates and become the better by the third update. This is because the teacher is initialized with CLIP, which is not pretrained for detection, and thus cannot provide high quality PLs. With updates, the teacher is equipped with the knowledge about detection and generates promising PLs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 517, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 563 + ], + "type": "text", + "content": "Time cost of pseudo labeling. We run two major PL-based methods [9, 49] on our computational environment and compare them with our pseudo labeling in terms of the time cost in Table 7." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 564, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 287, + 659 + ], + "type": "text", + "content": "Reducing the time cost of external RPN. There is a simple solution to remove the extra computational cost of external RPN adopted by SAS-Det. We first generate pseudo labels with SAS-Det, and then follow VL-PLM [49] to train a Faster R-CNN based on ResNet50 for OVD without the external RPN (\"Faster R-CNN + PLs\"). As shown in Table 8, \"Faster R-CNN + PLs\" achieves slightly better or close performance as SAS-Det." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "In this paper, we highlight two challenges associated with applying self-training to OVD: 1) noisy PLs from pretrained" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 526, + 144 + ], + "blocks": [ + { + "bbox": [ + 46, + 176, + 274, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 176, + 274, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 274, + 198 + ], + "type": "text", + "content": "Table 5. Performance with varied numbers of updates to the teacher model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 71, + 526, + 144 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 526, + 144 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 526, + 144 + ], + "type": "table", + "html": "
MethodAPretain50APnovel50APcoco50
OVR-CNN [48]11.522.938.1
Detic-80 [51]11.527.338.3
MEDet [5]18.632.642.4
Ours23.037.147.2
", + "image_path": "21e7c2c69201978f23bec2b91ba02aa5a26236985956650644d55fbbe3a8e82f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 341, + 210, + 504, + 272 + ], + "blocks": [ + { + "bbox": [ + 294, + 152, + 547, + 186 + ], + "lines": [ + { + "bbox": [ + 294, + 152, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 294, + 152, + 547, + 186 + ], + "type": "text", + "content": "Table 6. Generalization capability of OVD methods on retained 15 COCO categories. Text embeddings of 80 COCO categories are used as the classifier. Numbers are from MEDet [5]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 341, + 210, + 504, + 272 + ], + "lines": [ + { + "bbox": [ + 341, + 210, + 504, + 272 + ], + "spans": [ + { + "bbox": [ + 341, + 210, + 504, + 272 + ], + "type": "table", + "html": "
MethodTime (s)APnovel50
PB-OVD [9]0.484818.7
VL-PLM [49]0.445625.5
Ours0.130826.7
", + "image_path": "ac657bb1dbbde35875466f8cce8e021f3432dd9ac4bd75e19dd78df55e8dd902.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 317, + 358, + 536, + 408 + ], + "blocks": [ + { + "bbox": [ + 295, + 280, + 548, + 314 + ], + "lines": [ + { + "bbox": [ + 295, + 280, + 548, + 314 + ], + "spans": [ + { + "bbox": [ + 295, + 280, + 548, + 314 + ], + "type": "text", + "content": "Table 7. Mean time cost to get PLs per image. The quality of PLs on the validation set of COCO-OVD are provided for reference. We report the quality of our PLs after the last update." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 358, + 536, + 408 + ], + "lines": [ + { + "bbox": [ + 317, + 358, + 536, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 358, + 536, + 408 + ], + "type": "table", + "html": "
Method\\( AP_{50}^{novel} \\)\\( AP_{50}^{base} \\)\\( AP_{50}^{all} \\)
SAS-Det37.458.553.0
Faster R-CNN + PLs38.158.353.0
", + "image_path": "1bf6c72ec183987a499d721b569c02d8157e4fdf2e67d69d58d78fd836b5e478.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 416, + 545, + 437 + ], + "lines": [ + { + "bbox": [ + 305, + 416, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 416, + 545, + 437 + ], + "type": "text", + "content": "Table 8. Training Faster R-CNN with PLs from SAS-Det on COCO-OVD" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 461, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 663 + ], + "type": "text", + "content": "VLMs and 2) frequent changes of PLs' distributions. To address the challenges, we introduce a split-and-fusion (SAF) head and implement periodic updates. The SAF head splits a standard detection head into an open-branch and a closed-branch. The open-branch is trained with the class labels of ground truth and PLs. The closed-branch is trained with both class labels and boxes of ground truth. This approach effectively eliminates the location noise of PLs during training. Furthermore, the SAF head acquires complementary knowledge from different sets of training data, enabling fusion to enhance performance. The periodic updates reduce the number of updates to the teacher models, thereby decreasing the frequency of changes in PLs' distributions. We demonstrate that the proposed method SAS-Det is both efficient and effective. Our pseudo labeling is much faster than prior PL-based methods [9, 49]. SAS-Det also outperforms recent models on both COCO and LVIS for OVD." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgments: This research project has been partially funded by research grants to Dimitris N. Metaxas through NSF: 2310966, 2235405, 2212301, 2003874, and FA9550-23-1-0417." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13945" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "text", + "content": "[2] Relja Arandjelović, Alex Andonian, Arthur Mensch, Olivier J Henaff, Jean-Baptiste Alayrac, and Andrew Zisserman. Three ways to improve feature alignment for open vocabulary detection, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "type": "text", + "content": "[3] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In ECCV, pages 384-400, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 258 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-End Object Detection with Transformers. In ECCV, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 300 + ], + "type": "text", + "content": "[5] Peixian Chen, Kekai Sheng, Mengdan Zhang, Yunhang Shen, Ke Li, and Chunhua Shen. Open vocabulary object detection with proposal mining and prediction equalization, 2022. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 301, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 301, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 301, + 288, + 356 + ], + "type": "text", + "content": "[6] Yuxiao Chen, Jianbo Yuan, Yu Tian, Shijie Geng, Xinyu Li, Ding Zhou, Dimitris N Metaxas, and Hongxia Yang. Revisiting multimodal representation in contrastive learning: from patch and token embeddings to finite discrete tokens. In CVPR, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "type": "text", + "content": "[7] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In CVPR, pages 14084-14093, 2022. 2, 4, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "type": "text", + "content": "[8] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In ECCV, pages 701-717. Springer, 2022. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "text", + "content": "[9] Mingfei Gao, Chen Xing, Juan Carlos Niebles, Junnan Li, Ran Xu, Wenhao Liu, and Caiming Xiong. Open vocabulary object detection with pseudo bounding-box labels. In ECCV, pages 266-282. Springer, 2022. 1, 2, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 288, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 288, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 288, + 535 + ], + "type": "text", + "content": "[10] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, pages 2918-2928, 2021. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 568 + ], + "type": "text", + "content": "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary Object Detection via Vision and Language Knowledge Distillation. In ICLR, 2022. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 569, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 288, + 602 + ], + "type": "text", + "content": "[12] Agrim Gupta, Piotr Dólar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 602, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 624 + ], + "type": "text", + "content": "[13] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "text", + "content": "[14] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[15] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR - Modulated detection for end-to-end multi-modal understanding." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 95, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 140 + ], + "type": "text", + "content": "[16] Zaid Khan, Vijay Kumar B.G., Xiang Yu, Samuel Schulter, Manmohan Chandraker, and Yun Fu. Single-stream multilevel alignment for vision-language pretraining. In ECCV, pages 735-751. Springer, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 140, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 173 + ], + "type": "text", + "content": "[17] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Region-aware pretraining for open-vocabulary object detection with vision transformers. In CVPR, pages 11144-11154, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 173, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 216 + ], + "type": "text", + "content": "[18] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. In ICLR, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 217, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 261 + ], + "type": "text", + "content": "[19] Gang Li, Xiang Li, Yujie Wang, Yichao Wu, Ding Liang, and Shanshan Zhang. Pseudo labeling and consistency training for semi-supervised object detection. In ECCV, pages 457-472. Springer, 2022. 1, 2, 4, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "text", + "content": "[20] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 305, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 358 + ], + "type": "text", + "content": "[21] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, pages 10965-10975, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 360, + 545, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 393 + ], + "type": "text", + "content": "[22] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, pages 280-296. Springer, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 393, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 437 + ], + "type": "text", + "content": "[23] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In ICLR, 2023. 1, 3, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 481, + 545, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 481, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 308, + 481, + 545, + 515 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature Pyramid Networks for Object Detection. In CVPR, 2017. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "text", + "content": "[26] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In ICLR, 2021. 2, 4, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "type": "text", + "content": "[27] Zhixiang Min, Bingbing Zhuang, Samuel Schulter, Buyu Liu, Enrique Dunn, and Manmohan Chandraker. Neurocs: Neural nocs supervision for monocular 3d object localization. In CVPR, pages 21404-21414, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 603, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 658 + ], + "type": "text", + "content": "[28] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection with vision transformers. In ECCV. Springer, 2022. 2, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 658, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 714 + ], + "type": "text", + "content": "[29] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529-533, 2015. 4" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13946" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 174 + ], + "type": "text", + "content": "[31] Shafin Rahman, Salman Khan, and Nick Barnes. Improved visual-semantic alignment for zero-shot object detection. In AAAI, pages 11932–11939, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 176, + 287, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 176, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 287, + 209 + ], + "type": "text", + "content": "[32] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In NeurIPS, 2015. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 211, + 287, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 211, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 48, + 211, + 287, + 265 + ], + "type": "text", + "content": "[33] Zhongzheng Ren, Zhiding Yu, Xiaodong Yang, Ming-Yu Liu, Yong Jae Lee, Alexander G Schwing, and Jan Kautz. Instance-aware, context-focused, and memory-efficient weakly supervised object detection. In CVPR, pages 10598-10607, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 267, + 287, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 267, + 287, + 311 + ], + "spans": [ + { + "bbox": [ + 48, + 267, + 287, + 311 + ], + "type": "text", + "content": "[34] Samuel Schulter, Yumin Suh, Konstantinos M Dafnis, Zhixing Zhang, Shiyu Zhao, Dimitris Metaxas, et al. Omnilabel: A challenging benchmark for language-based object detection. ICCV, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 313, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 313, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 48, + 313, + 287, + 378 + ], + "type": "text", + "content": "[35] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 380, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 380, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 380, + 287, + 414 + ], + "type": "text", + "content": "[36] Bharat Singh, Hengduo Li, Abhishek Sharma, and Larry S Davis. R-FCN-3000 at 30fps: Decoupling detection and classification. In CVPR, pages 1081–1090, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 416, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 459 + ], + "type": "text", + "content": "[37] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. In arXiv:2005.04757, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 462, + 287, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 462, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 48, + 462, + 287, + 505 + ], + "type": "text", + "content": "[38] Peng Tang, Xinggang Wang, Xiang Bai, and Wenyu Liu. Multiple instance detection network with online instance classifier refinement. In CVPR, pages 2843-2851, 2017. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 508, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 508, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 508, + 287, + 552 + ], + "type": "text", + "content": "[39] Fang Wan, Chang Liu, Wei Ke, Xiangyang Ji, Jianbin Jiao, and Qixiang Ye. C-MIL: Continuation multiple instance learning for weakly supervised object detection. In CVPR, pages 2199-2208, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 554, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 287, + 597 + ], + "type": "text", + "content": "[40] Luting Wang, Yi Liu, Penghui Du, Zihan Ding, Yue Liao, Qiaosong Qi, Biaolong Chen, and Si Liu. Object-aware distillation pyramid for open-vocabulary object detection. In CVPR, pages 11186-11196, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "type": "text", + "content": "[41] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In CVPR, 2023. 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "text", + "content": "[42] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. CORA: Adapting CLIP for Open-Vocabulary Detection with Region Prompting and Anchor Pre-Matching. In CVPR, 2023. 1, 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "text", + "content": "[43] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 453 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[44] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, pages 10687-10698, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[45] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In ICCV, pages 3060-3069, 2021. 1, 2, 4, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "text", + "content": "[46] Caixia Yan, Xiaojun Chang, Minnan Luo, Huan Liu, Xiaoqin Zhang, and Qinghua Zheng. Semantics-guided contrastive network for zero-shot object detection. IEEE TPAMI, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "text", + "content": "[47] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In ECCV, pages 106-122. Springer, 2022. 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "type": "text", + "content": "[48] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-Vocabulary Object Detection Using Captions. In CVPR, 2021. 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 319 + ], + "type": "text", + "content": "[49] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, BG Vijay Kumar, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris N Metaxas. Exploiting unlabeled data with vision and language models for object detection. In ECCV, pages 159-175. Springer, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "type": "text", + "content": "[50] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. RegionCLIP: Region-based Language-Image Pretraining. In CVPR, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "text", + "content": "[51] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, pages 350-368. Springer, 2022. 5, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "type": "text", + "content": "[52] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Don't even look once: Synthesizing features for zero-shot detection. In CVPR, pages 11693-11702, 2020. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13947" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/3259ff45-5fcb-4059-9736-190d9ce81cb7_content_list.json b/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/3259ff45-5fcb-4059-9736-190d9ce81cb7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..009700273eeba0182e1ddd41636607c023993417 --- /dev/null +++ b/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/3259ff45-5fcb-4059-9736-190d9ce81cb7_content_list.json @@ -0,0 +1,1673 @@ +[ + { + "type": "text", + "text": "Taming Stable Diffusion for Text to $360^{\\circ}$ Panorama Image Generation", + "text_level": 1, + "bbox": [ + 132, + 130, + 836, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cheng Zhang $^{1,3}$ Qianyi Wu $^{1}$ Camilo Cruz Gambardella $^{1,3}$ Xiaoshui Huang $^{2*}$ \nDinh Phung $^{1}$ Wanli Ouyang $^{2}$ Jianfei Cai $^{1*}$", + "bbox": [ + 156, + 179, + 820, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Monash University 2Shanghai AI Laboratory 3Building 4.0 CRC, Caulfield East, Victoria, Australia", + "bbox": [ + 86, + 215, + 880, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8e1203debcca5709d928d249739fd1c3c9569ac0f14258ee85a50ee0a40877ee.jpg", + "image_caption": [ + "Figure 1. Our PanFusion can generate realistic and consistent $360^{\\circ}$ horizontal by $180^{\\circ}$ vertical FOV panoramas from a single text prompt, compared to the limited FOV of current state-of-the-art method MVDiffusion [47]. Left: PanFusion addresses the problem of repetitive elements (duplicated \"ceiling fans\") and inconsistency (the ceiling and wall in the center) of MVDiffusion. Right: While trained mostly on indoor scenes, PanFusion can generalize well to out-of-domain outdoor prompts with more reasonable layout." + ], + "image_footnote": [], + "bbox": [ + 76, + 273, + 486, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/abb1b99aa30d27822aa4ee9ff979ff1bcbe7557aa62a3b856476ba66d5e8e7b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 273, + 885, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 607, + 313, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative models, e.g., Stable Diffusion, have enabled the creation of photorealistic images from text prompts. Yet, the generation of 360-degree panorama images from text remains a challenge, particularly due to the dearth of paired text-panorama data and the domain gap between panorama and perspective images. In this paper, we introduce a novel dual-branch diffusion model named PanFusion to generate a 360-degree image from a text prompt. We leverage the stable diffusion model as one branch to provide prior knowledge in natural image generation and register it to another panorama branch for holistic image generation. We propose a unique cross-attention mechanism with projection awareness to minimize distortion during the collaborative denoising process. Our experiments validate that PanFusion surpasses existing methods and, thanks to its dual-branch structure, can integrate additional constraints like room layout for customized panorama outputs.", + "bbox": [ + 75, + 623, + 473, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 606, + 630, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating a $360^{\\circ}$ panorama image from textual prompts is a nascent yet pivotal frontier in computer vision, with profound implications for applications that require extensive environmental representation, such as environmental lighting [1, 4, 51], VR/AR [60, 61], autonomous driving [59], and visual navigation [17]. Despite recent significant strides in text-to-image synthesis, the leap to generating full $360^{\\circ}$ horizontal by $180^{\\circ}$ vertical field-of-view (FOV) panorama remains challenging.", + "bbox": [ + 496, + 638, + 890, + 775 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are two major hurdles for achieving this goal. The first hurdle is data scarcity. The availability of text-to-panorama image pairs [20, 47] is significantly less compared with the abundance of text-to-common image pairs [38, 39]. The dearth of data complicates the training and finetuning of generative models. The second hurdle lies in the geometric and domain variations. Panorama images are distinct not only in their aspect ratio $(2:1)$ but", + "bbox": [ + 496, + 780, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 888, + 220, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6347", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "also in the underlying equirectangular projection (ERP) geometry [58]. This differs significantly from typical square images of the perspective projection that are used in most generative model training [38, 39].", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate the scarcity of panorama-specific training data, the previous solutions follow a common principle that leverages the prior knowledge of the pre-trained generative model [17, 20, 47]. However, taming powerful models like stable diffusion [31, 34] to generate high-fidelity panorama images remains a non-trivial task. Early attempts turn to formulate the 360-degree generation as an iterative image inpainting or warping process [17, 20]. Such solutions suffer from error accumulation and fail to handle loop closure [47]. To address this, MVDiffusion [47] proposes to produce multiple perspective images simultaneously by introducing a correspondence-aware attention module to facilitate multiview consistency, and then stitch together the perspective images to form a complete panorama. Despite the improved performance, the pixel-level consistency between neighboring perspectives in MVDiffusion cannot ensure global consistency, often resulting in repetitive elements or semantic inconsistency, as illustrated in Fig. 1.", + "bbox": [ + 75, + 154, + 470, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, in this paper, we propose a novel dual-branch diffusion model called PanFusion that is tailored to address the limitations of prior models for high-quality text to $360^{\\circ}$ panorama image generation. Specifically, PanFusion is designed to operate in both panorama and perspective domains, employing a global branch for creating a coherent panoramic \"canvas\" and a local branch that focuses on rendering detail-rich multiview perspectives. The local-global synergy of PanFusion significantly improves the resulting panoramas against the prevalent issues of error propagation and visual inconsistency that prior models have struggled with. To enhance the synergy between the two branches, we further propose an Equirectangular-Perspective Projection Attention (EPPA) mechanism, which respects the equirectangular projection for maintaining geometric integrity throughout the generation process. In addition, our adoption of parametric mapping for positional encoding is another leap forward, enhancing the model's spatial awareness and further ensuring the consistency of the generated panorama. Moreover, the panorama branch of PanFusion can be easily leveraged to accommodate supplementary control inputs at the panorama level, such as room layout, allowing for the creation of images that adhere to precise spatial conditions. We summarize our primary contributions as follows.", + "bbox": [ + 75, + 429, + 470, + 806 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We pioneer a dual-branch diffusion model PanFusion, harnessing both the global panorama and local perspective latent domains, to generate high-quality, consistent $360^{\\circ}$ pancreas from text prompts.", + "- To enhance the interaction between the two branches, we introduce an \"equirectangular-perspective projection at" + ], + "bbox": [ + 76, + 810, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tention” mechanism that establishes a novel correspondence between global panorama and local perspective branches, addressing the unique projection challenges of panorama synthesis.", + "bbox": [ + 509, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Our PanFusion not only surpasses prior models in quality and consistency but also supports extended control over the generation process with the inclusion of room layout. Extensive experimental results demonstrate the superiority of our proposed framework.", + "bbox": [ + 500, + 152, + 890, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 239, + 640, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models. In recent years, diffusion models [6, 12, 41, 42, 44, 62] have taken the world of image generation by storm, as they have become faster [15, 19, 42] and more capable in terms of image quality and resolution [28, 31, 36]. This success has prompted the development of various applications for diffusion models, such as text-to-image [25, 28, 31, 36], image-conditioned generation [24, 64], in-painting [21, 35] and subject-driven generation [10, 33]. Most of these applications try to exploit the prior knowledge of a pre-trained diffusion model to mitigate the scarcity of task-specific data, by either finetuning with techniques like LoRA [14], or introducing auxiliary modules to distill the knowledge. We also adopt the same principle to harness the power of a pre-trained latent diffusion model [31] for panorama image generation.", + "bbox": [ + 498, + 266, + 890, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Panorama generation. Panorama image generation encompasses different settings, including panorama outpainting and text-to-panorama generation. Panorama outpainting [1, 5, 26, 51, 54, 56] focus on generating a 360-degree panorama from a partial input image. Various methods, such as StyleLight [51] and BIPS [26], have addressed specific use cases, focusing on HDR environment lighting and robotic guidance scenarios. Recent works [54, 56] have improved realism with diffusion models, but often lack the exploitation of rich prior information from pre-trained models, limiting generalization. On the other hand, recent developments in generative models have opened new frontiers in synthesizing immersive visual content from textual input [8, 13, 43, 52, 53, 55, 56, 61, 63]. As an image-based representation, generating panorama from text has gained much attention. Text2Light [4] adopt the VQGAN [7] structure to synthesize an HDR panorama image from text. To generate in arbitrary resolution with pre-trained diffusion models, DiffCollage [65], MultiDiffusion [2] and SyncDiffusion [16] propose to fuse the diffusion paths, while PanoGen [17] solves by iteratively inpainting. However, they failed to model the equirectangular projection of 360-degree panoramas. Lu et al. [20] adopts an autoregressive framework, but suffers from inefficient issues. MVDiffusion [47] designs a correspondence-aware attention module to produce multi-view images simultaneously that can be stitched together but results in repetitive elements and in", + "bbox": [ + 498, + 494, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6348", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ed3a739337d81ae64bd181e429d79a0db48cff0fc1f855a18fdb90f214bf4498.jpg", + "image_caption": [ + "Figure 2. Our proposed dual-branch PanFusion pipeline. The panorama branch (upper) provides global layout guidance and registers the perspective information to get seamless panorama output. The perspective branch (lower) harnesses the rich prior knowledge of Stable Diffusion (SD) and provides guidance to alleviate distortion under perspective projection. Both branches employ the same UNet backbone with shared weights, while finetuned with separate LoRA layers. Equirectangular-Perspective Projection Attention (EPPA) modules are plugged into different layers of the UNet to pass information between the two branches." + ], + "image_footnote": [], + "bbox": [ + 89, + 79, + 875, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "consistency. In contrast, our proposed PanFusion, a dual-branch framework, addresses the limitations of existing methods by considering both global panorama views and local perspective views, providing a comprehensive solution for text-driven 360-degree panorama image generation.", + "bbox": [ + 75, + 385, + 470, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 472, + 169, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 497, + 205, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First introduced in [41], Diffusion models [12, 42, 44] aim to generate images from a noise distribution by iterative denoising with a learned prior distribution. However, the early diffusion models operate in the image space, which is of high dimension and complex. We have recently witnessed the huge success of latent diffusion models [31] that first transform an image $x$ to a latent representation $z$ with a learned encoder $\\mathcal{E}$ and then train a UNet [32] model $\\epsilon_{\\theta}$ parameterized by $\\theta$ for the reverse process in the latent space, formulating the training objective as:", + "bbox": [ + 75, + 520, + 468, + 672 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {\\mathcal {E} (x), t, \\epsilon , y} [ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\tau (y)) | | _ {2} ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 679, + 468, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau(y)$ is an encoding of input condition $y$ (e.g., text prompt or image), $z_{t}$ is the latent map at time step $t$ , and $\\epsilon$ is sampled from a Gaussian noise. To sample from the model, $z_{t}$ is first initialized from standard Gaussian distribution, then the reverse process is applied iteratively to generate $z_{0}$ , which is finally decoded into image space with a decoder $\\mathcal{D}$ .", + "bbox": [ + 75, + 703, + 470, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Dual-Branch Diffusion Model", + "text_level": 1, + "bbox": [ + 76, + 816, + 341, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Directly employing pre-trained latent diffusion models [31], e.g., Stable Diffusion (SD) [34], to generate panorama from multiple perspective images, either in an iterative manner [9, 13, 17] or in a synchronized way [2, 47], would", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fail to handle loop closure [47] or produce repetitive elements (Fig. 1) due to lack of global understanding. To address this issue, we propose a dual-branch diffusion model, which consists of a panorama branch and a perspective branch both based on the UNet of SD as shown in Fig. 2. The panorama branch is designed to provide global layout guidance and to register the perspective information to get the final panorama without stitching, while the perspective branch is designed to exploit the rich perspective image-generation capabilities of SD and to provide guidance to alleviate distortion under perspective projection. The two branches work together during the diffusion process to generate a denoised panorama latent map. Finally, this latent map runs through the pre-trained decoder $\\mathcal{D}$ of SD to produce the final panorama image.", + "bbox": [ + 496, + 383, + 890, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Panorama Branch. Given a text prompt $y$ , our goal is to generate a panorama $x \\in \\mathbb{R}^{3 \\times H \\times W}$ , where $W = 2H$ , $H = 512$ . To account for the difference between the target resolution and the one that SD is trained on, we introduce LoRA [14] layers to adapt the model to the new resolution. SD with LoRA can already serve as a strong baseline for panorama generation, but the results are not loop-consistent. Previous works [8, 56] have attempted to overcome this problem using 90 degree rotations of the latent map at each diffusion step. However, the seams are still obvious [56]. On close inspection of the SD model we have observed that the loop inconsistency is mainly caused by the convolutional layers in the UNet backbone, due to the lack of a mechanism to pass information between the two ends of the panorama image. Therefore, we introduce an adaptation to the UNet by adding additional circular padding [40, 54, 66] before each convolutional layer, and then cropping the output feature maps to the original size. In addition, we also add circular padding to the latent map before decoding to", + "bbox": [ + 496, + 613, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6349", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mitigate the less apparent loop inconsistency caused by the decoder. The combination of the techniques outlined above - latent rotation and circular padding - enable the generation of loop-consistent results with negligible computational cost, and can thus serve as another strong baseline. However, these alone do not make full use of the perspective generation capabilities that SD possess.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Perspective Branch. The perspective branch aims to exploit the outstanding capabilities that SD has shown in the generation perspective images. Since it does not generate panoramas directly, it can operate at a lower resolution. Specifically, we set its resolution to $H / 2 \\times H / 2$ and add LoRA layers to adapt to the new resolution. To evenly distribute perspective cameras and fully cover the panorama image, we sample $N = 20$ cameras with poses $R^i \\in SO(3), i \\in 1, \\dots, N$ on an icosahedron, similar to [27, 30], and set the FOV $= 90^\\circ$ . We input the same prompt $y$ to both branches, and leave it to the model to decide how to exploit the prompt.", + "bbox": [ + 75, + 198, + 470, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Joint Latent Map Initialization. Previous work [22] has shown that the latent map initialization can affect the layout of the generated image, i.e., the initial noise can be modified to manipulate the layout of the generated image. We find this is particularly important for multi-image generation with correspondence, as the overlapping regions in different views tend to generate different elements if noise is sampled independently. Under this observation, we propose to jointly sample noise for the panorama and perspective latent maps by projecting noise from the panorama to the perspective. To be more specific, we first initialize panorama latent map $z_{T}^{*} \\in \\mathbb{R}^{4 \\times H / f \\times W / f}$ as Gaussian noise, then initialize perspective latent maps with $z_{T}^{i} = \\mathrm{P}(z_{T}^{*}, R^{i}, \\mathrm{FOV}, (H / 2f, H / 2f)), z_{T}^{i} \\in \\mathbb{R}^{4 \\times H / 2f \\times H / 2f}$ , where $\\mathrm{P}(\\cdot)$ is the function that projects $z_{T}^{*}$ to a perspective view, and $f$ is the down-sampling factor the encoder $\\mathcal{E}$ used in SD. We use nearest interpolation for projection as it introduces fewer artifacts than bilinear interpolation.", + "bbox": [ + 75, + 380, + 470, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. EPP Attention", + "text_level": 1, + "bbox": [ + 76, + 664, + 225, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To pass the guidance between the perspective and panorama branches, a naive solution is to project the feature maps in different layers from one branch to the other, similar to [47]. However, this will introduce information loss during interpolation and will restrict the receptive field to a small region around corresponding pixels. Instead, we propose an Equirectangular-Perspective Projection Attention (EPPA) module to implicitly pass the guidance in a cross-attention way as shown in Fig. 2. The EPPA operates on feature maps in different layers of two UNet branches, denoted as $F^{*} \\in \\mathbb{R}^{c \\times h \\times w}$ and $F^{i} \\in \\mathbb{R}^{c \\times h / 2 \\times h / 2}$ , where $c$ is the channel dimension, $h$ and $w$ are the height and width of the feature maps, respectively. In addition, while cross-attention bypasses interpolation and provides benefits from", + "bbox": [ + 75, + 688, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/50449cba423ab2c448e7b4f1528b67d19e45599cbe2fb2d383ac966358af1826.jpg", + "image_caption": [ + "Figure 3. Equirectangular-Perspective Projection (EPP) Attention. As EPP attention module is designed to be bijective to pass information in both directions, we only illustrated the direction of registering perspective information to panorama." + ], + "image_footnote": [], + "bbox": [ + 503, + 88, + 885, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the global receptive field of the attention mechanism, it is unaware of the projection between different formats. To address this issue, as shown in Fig. 3, we introduce two key components: EPP spherical positional encoding and EPP attention mask.", + "bbox": [ + 496, + 361, + 890, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "EPP Spherical Positional Encoding. Since the EPPA module tries to associate two different formats, it is important to add positional information in the same space so that the attention mechanism can learn the correspondence. We achieve this by introducing Spherical Positional Encoding (SPE) [4, 23] to the EPPA module. The $\\mathrm{SPE}(\\theta ,\\phi) = (\\gamma (\\theta),\\gamma (\\phi))$ function maps polar coordinates $(\\theta ,\\phi)$ to a higher dimensional space $\\mathbb{R}^{4L}$ with Fourier positional encoding:", + "bbox": [ + 496, + 436, + 890, + 574 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma (\\theta) = \\left[ \\sin \\left(2 ^ {0} \\pi \\theta\\right), \\cos \\left(2 ^ {0} \\pi \\theta\\right), \\dots , \\sin \\left(2 ^ {L - 1} \\pi \\theta\\right), \\cos \\left(2 ^ {L - 1} \\pi \\theta\\right) \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 579, + 890, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here we set $L = c / 4$ so that $\\mathrm{SPE}(\\theta, \\phi) \\in \\mathbb{R}^c$ . In the EPPA module, we apply SPE by first computing an SPE map for the panorama feature map, then project it to each perspective feature map with projection function $\\mathrm{P}(\\cdot)$ , so that corresponding pixels in different formats share the same SPE vector. Finally, the SPE maps are added to the feature maps accordingly and go through a linear layer to get the query $Q$ and key $K$ , following a matrix product to get the affinity matrix $A$ . Take the perspective-to-panorama direction as an example as shown in Fig. 3, where $Q \\in \\mathbb{R}^{c \\times h \\times w}$ and $K \\in \\mathbb{R}^{N \\times c \\times h / 2 \\times h / 2}$ are reshaped and multiplied to get the affinity matrix $A \\in \\mathbb{R}^{hw \\times Nh^2 / 4}$ .", + "bbox": [ + 496, + 611, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "EPP Attention Mask. In addition to the SPE, we also propose an EPP attention mask to encourage the attention mechanism to focus around the corresponding pixels, inspired by [48]. For example, for a target pixel in the panorama feature map as in Fig. 3, we would like to focus on registering the information from corresponding source pixels in the perspective feature maps. We achieve", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6350", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "this by enhancing the affinity matrix $A$ with a soft mask $M \\in \\mathbb{R}^{hw \\times Nh^2 / 4}$ highlighting the corresponding pixels. To generate $M$ , we first use $\\mathrm{P}(\\cdot)$ to project a binary mask $M_{j,k}$ for each pixel $(j, k)$ in panorama feature map to each perspective view $i$ . Then we apply a Gaussian kernel to smooth the masks and normalized them to $[-1, 1]$ . The masks are stacked and reshaped into $M$ , which is then added to $A$ to get the enhanced affinity matrix $A'$ . The rest goes the same as the vanilla attention mechanism, where a softmax function is applied to $A'$ to get the attention weights, which are finally multiplied with the value $V$ to get the output.", + "bbox": [ + 75, + 90, + 472, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by [47, 64], we add zero-initialized $1 \\times 1$ convolutional layers to the output of cross-attention and add it as a residual to the target feature map. This ensures the UNet stays unmodified at the beginning of training and can be gradually adapted to the EPPA modules. We add independent EPPA modules after each down-sampling layer and before each up-sampling layer of UNet to connect two branches, detailed in Sec. A of the supplementary material. Considering that the guidance is passed in both directions following the same equirectangular-perspective projection rule, which is bijective in nature, we share the weights of the EPPA modules in the two directions.", + "bbox": [ + 75, + 258, + 472, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Layout-Conditioned Generation", + "text_level": 1, + "bbox": [ + 76, + 448, + 361, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One of the most important applications for panorama generation is to generate panorama according to a given room layout. This is particularly useful for panorama novel view synthesis [57] and can potentially benefit indoor 3D scene generation [8, 13]. However, this problem is not well researched for diffusion-based panorama generation, largely due to the difficulty of introducing layout constraints while exploiting the rich prior knowledge of SD in perspective format at the same time. For panorama generation from multi-view [17, 47], one naive solution is to project layout condition into different views to locally condition the generation of perspective images. Instead, for our dual-branch diffusion model, we can naturally exploit the global nature of the panorama branch to enforce a much stronger layout constraint. Specifically, we render the layout condition as a distance map, then use it as the input of a ControlNet [64] to condition the panorama branch.", + "bbox": [ + 75, + 470, + 472, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training", + "text_level": 1, + "bbox": [ + 76, + 736, + 181, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we employ the same technique for latent map initialization to jointly sample noise for the panorama and perspective view, which we denote as $\\epsilon^{*}$ and $\\epsilon^i$ , respectively. Given a GT panorama $x^{*}$ , we apply supervision on the panorama branch with the loss in Eq. (1):", + "bbox": [ + 75, + 758, + 468, + 835 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {*} = \\mathbb {E} _ {\\mathcal {E} \\left(x ^ {*}\\right), t, \\epsilon^ {*}, y} \\left[ \\| \\epsilon^ {*} - \\epsilon_ {\\theta} ^ {*} \\left(z _ {t} ^ {*}, t, \\tau (y)\\right) \\| _ {2} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 844, + 468, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}^{*}$ is the predicted noise from the panorama branch. To encourage the synchronization between the", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "two branches, we also apply supervision on the perspective branch with the following loss:", + "bbox": [ + 496, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {i} = \\mathbb {E} _ {\\mathcal {E} (x ^ {i}), t, \\epsilon^ {i}, y} \\left[ | | \\epsilon^ {i} - \\epsilon_ {\\theta} ^ {i} \\left(z _ {t} ^ {i}, t, \\tau (y)\\right) | | _ {2} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 133, + 890, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x^{i}$ is a perspective image projected from $x^{*}$ and $\\epsilon_{\\theta}^{i}$ is the predicted noise from the perspective branch. We joint train the EPPA modules and the LoRA layers in the two branches by combining the above two losses as $\\mathcal{L} = \\mathcal{L}^{*} + \\frac{1}{N}\\sum_{i = 1}^{N}\\mathcal{L}^{i}$ . Note that the SD UNet blocks remain frozen.", + "bbox": [ + 496, + 164, + 893, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 498, + 257, + 624, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 498, + 282, + 689, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We follow the MVDiffusion [47] to use the Matterport3D dataset [3], which has 10,800 panoramic images with 2,295 room layout annotations. We employ BLIP-2 [18] to generate a short description for each image.", + "bbox": [ + 496, + 306, + 890, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. For text-conditioned generation, the training and inference schedules are kept the same as MVDiffusion [47] to make a fair comparison. For textlayout conditioned generation, we train the additional ControlNet with other parameters fixed.", + "bbox": [ + 496, + 369, + 890, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following previous works, we evaluate image quality in panorama [4] and perspective [47] domain. For layout-conditioned generation, we propose a new metric to evaluate how well the generated panorama follows input layout. Specifically, we use the following metrics:", + "bbox": [ + 496, + 446, + 890, + 521 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Panorama. We follow Text2Light [4] to report Fréchet Inception Distance (FID) [11] and Inception Score (IS) [37] on panoramas to measure realism and diversity. Additionally, CLIP Score (CS) [29] is used to evaluate the text-image consistency. While FID is widely used for image generation, it relies on an Inception network [46] trained on perspective images, thus less applicable for panorama images. Therefore, a variant of FID customized for panorama, Fréchet Auto-Encoder Distance (FAED) [26], is used to better compare the realism.", + "- Perspective. To simulate the real-world scenario where the user can freely navigate a panorama by viewing from different perspective views, we also report FID and IS for 20 randomly sampled views to compare with methods that generate $180^{\\circ}$ vertical FOV. We also follow MVDiffusion [47] to report FID, IS and CS scores on 8 horizontally sampled views. It is worth noting that this group of metrics favors MVDiffusion by measuring its direct outputs, while ours involve interpolation for perspective views.", + "- Layout Consistency. We propose a layout consistency metric, which employs a layout estimation network HorizonNet [45] to estimate the room layout from the generated panorama and then compute its 2D IoU and 3D IoU [45] with the input layout condition." + ], + "bbox": [ + 496, + 522, + 890, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "See Supplementary Sec. B for more details of the above.", + "bbox": [ + 500, + 885, + 872, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6351", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ace065b904e2da79b70c150698d52595c2f7162accc98e4577db23613d95579b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 78, + 493, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9c6aa4413a5ce6428bba4458d6f3de59a21235d03f2896d0f2fffedfad6475b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 155, + 493, + 231 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e087e8bfc1296b7c0043d3d0fa9bf6e0df571ff89d63f0e7b89ba191cfc6720f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 231, + 493, + 310 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f90956f5d9c5e72ddbc2d51ebed4048f9aa9b2251863ff240383eadc6345ec31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 310, + 493, + 383 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/633b687c007e1a1c7fd7b368033f464ceb895b7b3761fd7bfdc843689302f493.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 385, + 493, + 462 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/661a6782d8476d2a9ff693eaf74ef5079a39218adec91cdd6627b0833aeee0b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 462, + 493, + 537 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1bbd72bb06e8a7461e2b28f8ea9afbe41f59e274056ced04d70d72e4b794b9fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 537, + 493, + 614 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/124f5afc5b44742917491feaae341ca5c2862d25df2767fb691a557c99e83082.jpg", + "image_caption": [ + "\"A living room with a chandelier.\"" + ], + "image_footnote": [], + "bbox": [ + 101, + 614, + 493, + 691 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/934bd18a6bedfd75bbb98b8071b1afa3e82dcfe8e48899c86ae01caf02822919.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 78, + 887, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c51d90771ea05f1c23acdec07726f51d57969b665cc0105875e5a3af11335b18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 155, + 887, + 232 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/02f192f1c9331f0adf31e1fa81289385a76683188a686edd236b3ce53b5ca05b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 232, + 887, + 310 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ae6ab66321c5989c965c17d714d555741c0d2fd7918fecb68528dd3d02eae89.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 310, + 887, + 383 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/78369b2b8cbc8431420a113b7c53f80a95993ac00e3f3c5bb27bd9cbef850005.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 385, + 887, + 462 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e99228e06b04e5be2e579c0cff6de6915e415079096c1590aa83ba055bb4e214.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 462, + 887, + 537 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f28a8b7af8b21980302e3cbeb2fefad2cd02af5f1a814c0c9fc9ebf9f5a7c228.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 537, + 887, + 614 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8e013f9e1fb5c70e1bdd78571bc309db501bb50ea13d2b56a6b422bbdc390457.jpg", + "image_caption": [ + "\"A bedroom with a bed and a table.\"", + "Figure 4. Qualitative comparisons of text-conditioned panorama generation. We show panoramas cropped to the vertical FoV of MVDiffusion [47]. Below each panorama, we show 4 evenly spaced perspective projections, with the first view crossing the left and right boundaries. We highlight the loop inconsistency, distorted lines and repetitive objects and unreasonable furniture layout of baseline methods with corresponding colors of boxes, which are addressed by our method. More results can be found in Sec. E of the supplementary." + ], + "image_footnote": [], + "bbox": [ + 496, + 614, + 887, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with Previous Methods", + "text_level": 1, + "bbox": [ + 75, + 782, + 395, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our PanFusion with the following baselines (see Supplementary Sec. B for details).", + "bbox": [ + 75, + 808, + 467, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- MVDiffusion [47] utilizes a multi-view diffusion model to generate 8 horizontal views that can be stitched into a panorama with vertical FOV of $90^{\\circ}$ . It requires separate prompts for training while providing an option to generate", + "bbox": [ + 75, + 839, + 467, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "from a single prompt.", + "bbox": [ + 500, + 785, + 645, + 799 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Text2Light [4] generates a $180^{\\circ}$ vertical FOV panorama from a text prompt in a two-stage auto-regressive manner.", + "- $SD + LoRA$ is our baseline model that uses LoRA [14] to finetune a Stable Diffusion [31] on panorama images.", + "- Pano Branch is SD+LoRA with additional modifications as described in Sec. 3.2 to ensure loop consistency." + ], + "bbox": [ + 500, + 799, + 890, + 888 + ], + "page_idx": 5 + }, + { + "type": "aside_text", + "text": "PanFusion (Ours) SD+LoRA MVDiffusion Text2Light", + "bbox": [ + 81, + 130, + 96, + 652 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6352", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/42b3e4e2af64bc38fb10c52c12d5f09764faac51360467db5988198904e651aa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPanorama20 ViewsHorizontal 8 Views [47]
FAED ↓FID ↓IS ↑CS ↑FID ↓IS ↑FID ↓IS ↑CS ↑
Text2Light [4]97.2476.53.6027.4836.255.6743.664.9225.88
MVDiffusion [47]------25.276.9026.34
SD+LoRA [14, 31]7.1951.694.4028.8319.326.9020.686.4824.77
Pano Branch7.9050.404.5428.6720.107.0620.566.3724.85
PanFusion (Ours)6.0446.474.3628.5817.046.8519.886.5024.98
", + "bbox": [ + 148, + 75, + 816, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/40c16d3582734f542585fe17be6a2c1752272ae3355bb8f80de6eb3ffd352fcf.jpg", + "image_caption": [ + "Figure 5. Ablation study. Artifacts are highlighted with red boxes and projected to perspective views. Prompt: \"A hallway in a hotel.\"" + ], + "image_footnote": [], + "bbox": [ + 80, + 263, + 890, + 450 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d5f5e3e8628aaa8131651ab22dc0e516c250e055090b999baccbf536c36d0606.jpg", + "table_caption": [ + "Table 1. Comparison with SoTA methods. We evaluate the panorama image quality with Fréchet Auto-Encoder Distance (FAED), Fréchet Inception Distance (FID), Inception Score (IS), and CLIP Score (CS). In addition, we evaluate the perspective image quality in two settings. We first randomly sample 20 views, which is the closest to the real-world scenario where the user can freely navigate the panorama to view the scene from different perspectives. We then follow MVDiffusion [47] to horizontally sample 8 evenly spaced views." + ], + "table_footnote": [], + "table_body": "
MethodPanorama20 Views
FAED ↓FID ↓IS ↑CS ↑FID ↓IS ↑
Pano Branch7.9050.404.5428.6620.107.06
Ours-joint7.7559.434.3427.9429.686.82
Ours-SPE6.6349.554.6528.6620.827.00
Ours-mask6.7745.494.3528.6615.816.81
Ours-bijective7.3648.354.5828.6618.787.12
Ours6.0446.474.3628.5817.046.85
", + "bbox": [ + 81, + 481, + 465, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study. We compare the ablated versions of our method on both panorama and perspective domains. Here, “-” indicates removing the subsequent component.", + "bbox": [ + 75, + 594, + 468, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. Tab. 1 shows the quantitative comparison results. Here, we assign the highest value to realism in image generation, measured in FAED and FID. On these two metrics, our method outperforms baseline methods in both panorama and perspective. For IS, our method's performance is slightly lower than baselines. This is likely due to the fact that IS evaluates diversity of objects in generated images, using a classifier, and our model, unlike the baselines, tends to not generate unexpected objects. Similarly, it is possible to say that baseline models present slightly higher CS is due to the repetition of objects reinforcing alignment with prompts. Considering SD+LoRA is superior to Pano Branch on FAED and is on par in other metrics, we only qualitatively compare with SD+LoRA below.", + "bbox": [ + 75, + 655, + 468, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 4 shows the qualitative comparison results. Loop inconsistency can be observed for", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text2Light and SD+LoRA due to lack of message passing between left-right boundaries. They also suffer from distorted lines in perspective views, meaning the generated panorama does not follow the correct equirectangular projection. MVDiffusion on the other hand suffers from repetitive objects and unreasonable furniture layout, which is likely due to the lack of global context. Our method generates the most realistic scenes and aligns to text condition the best, also with less distortion in perspective views.", + "bbox": [ + 496, + 483, + 890, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 632, + 653, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Sec. 4.2 and Tab. 1, we show that our full model outperforms Pano Branch, the baseline model of our method without the perspective branch. Here, as shown in Tab. 2 and Fig. 5, we further conduct an ablation study to validate the effectiveness of each component in our method. For a consistent comparison, we keep the layout similar between different ablated versions by sampling the same noise for latent map initialization, exploiting the observation of [22].", + "bbox": [ + 496, + 657, + 890, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Joint latent map initialization. We ablate the joint latent map initialization by initializing the latent maps of the panorama and perspective branches separately (Ours-joint). Significant performance drop can be observed in all metrics and qualitative results, demonstrating the importance of joint latent map initialization. Interestingly, Ours-joint is even worse than Pano Branch in FID. It is likely due to that joint latent map initialization helps the corresponding", + "bbox": [ + 496, + 779, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6353", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b57889ee13bce4ebcd5d22f7cbaf577de7e4aab8985a68acbdd81aee1158fa63.jpg", + "image_caption": [ + "MVDiffusion" + ], + "image_footnote": [], + "bbox": [ + 104, + 75, + 419, + 135 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8312357515e061e142b64a63e35a959d260adb3ad1e47cd2a642184bc88e89c2.jpg", + "image_caption": [ + "Ours", + "\"A hallway in a building.\"" + ], + "image_footnote": [], + "bbox": [ + 104, + 135, + 418, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d7d388b8919fd8609b7529f550620f5ba905b8872d8bec11693fae51ba7ba152.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 75, + 496, + 135 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3191492e99170d3c2c89a76c59c093a394d500973fa4a5187bd38382149c4934.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 135, + 495, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab1ead31dc4bc0626532ab269220ad8d341f6f952fc1df6da75bc40229850700.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 75, + 813, + 135 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b34ff3731efe3f8f10ab37b0f854bd9bef7f5516236a1b9082ad9d4a0ec6627.jpg", + "image_caption": [ + "\"A living room with a view of the ocean.\"" + ], + "image_footnote": [], + "bbox": [ + 500, + 135, + 815, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/84ba38cc3abca5f015b2e71d32686c1be76bbdd7fa3d542299f3609348748b4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 815, + 75, + 890, + 135 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/44435942303edbbab2d31f069f1e90d43eaaeaaaef328ee0f3b87668fe9454ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 815, + 135, + 890, + 194 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8e6213f6690ac1e8df1f0eea249417c13472acec173670c5252bce6b1ff880f4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLayout ConsistencyHorizontal 8 Views [47]
3D IoU ↑2D IoU ↑FID ↓IS ↑CS ↑
MVDiffusion [47]61.0664.4328.835.6026.79
PanFusion (Ours)68.4671.8222.585.1026.04
GT image74.3177.15---
", + "bbox": [ + 78, + 263, + 465, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Layout-conditioned comparisons. We evaluate layout consistency between the condition and the layout of each generated image extracted by a layout estimator HorizonNet [45]. \"GT image\" indicates the upper bound of the layout estimator.", + "bbox": [ + 75, + 340, + 468, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pixels to share a similar noise distribution from the beginning of the diffusion process, which is essential for EPPA to align the content of overlapped regions.", + "bbox": [ + 75, + 411, + 468, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "EPP SPE and attention mask. We ablate spherical positional encoding (Ours-SPE) and attention mask (Ours-mask) from EPPA module. From Tab. 2, we can see that missing SPE hurts FAED and FID, which is likely due to that SPE helps the model to learn the relative position of the pixels between the two branches. Missing attention mask gets FID better, but it hurts FAED, which is more accurate in evaluating the panorama quality as it is customized for the target dataset. Both result in clear artifacts around point light sources, inconsistent textures of the floor, and distortions in the highlighted projections, as shown in Fig. 5.", + "bbox": [ + 75, + 457, + 468, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Bijective EPPA. We ablate the bijective EPPA (Ours-bijective) by using separate parameters for the two directions in the EPPA module. Both FAED and FID get worse for Ours-bijective. In addition, the ablated version struggles to generate consistent textures for the floor and ceiling for the two directions of the hallway in Fig. 5. On the contrary, our full model generates floor and ceiling with consistent style, showing a better global understanding of the scene.", + "bbox": [ + 75, + 623, + 468, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Application: Layout-Conditioned Generation", + "text_level": 1, + "bbox": [ + 76, + 756, + 460, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To showcase the advantage of our method in generating panorama images with additional layout conditions, we build a baseline model by adding a ControlNet to MVDiffusion as described in Sec. 3.4. We render the layout condition into a distance map and then project it to perspective views to condition its generation of multi-view images. The training settings are kept the same as our PanFusion. As shown in Tab. 3, our method outperforms the baseline", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3669e5d95b40ce62785b7bd191bfad3e9c21d937d360d78f89652501cd50204.jpg", + "image_caption": [ + "Figure 6. Layout-conditioned generation comparisons. We showcase how layout-conditioned panorama generation can benefit from our dual-branch structure. The input layout is drawn on the generated panorama as yellow lines, while the panorama is cropped to the FoV of MVDiffusion and projected to one view to highlight the layout consistency.", + "Figure 7. Failure case. Our method sometimes generates indoor scenes with no entrance." + ], + "image_footnote": [], + "bbox": [ + 503, + 263, + 696, + 339 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d5674ed2fa49f2cf3face45ffd5ec237046424fe671b0ce18c2e3244ff50dd00.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 263, + 888, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model in layout consistency while keeping the realism advantage of perspective projections. We visualize the layout condition as wireframes overlaid on generated panorama images in Fig. 6, where we can see our generated panorama images follow their layout conditions better, especially as highlighted in perspective views. More details can be found in Sec. F of the supplementary material.", + "bbox": [ + 498, + 388, + 890, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 510, + 617, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have proposed PanFusion, a novel text-to- $360^{\\circ}$ -panorama image generation method that can generate high-quality panorama images from a single text prompt. Particularly, a dual-branch diffusion architecture has been introduced to harness prior knowledge of Stable Diffusion in the perspective domain while addressing the repetitive elements and inconsistency issues observed in previous works. An EPPA module has been further introduced to enhance the information passing between the two branches. We have also extended our PanFusion for the application of layout-conditioned panorama generation. Comprehensive experiments have demonstrated that PanFusion can generate high-quality panorama images with better realism and layout consistency than previous methods.", + "bbox": [ + 496, + 537, + 890, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Although the dual-branch architecture of Pan-Fusion combines the advantages in both panorama and perspective domains, it comes with a cost of higher computational complexity. Additionally, our method sometimes fails to generate entrance for indoor scenes, as shown in Fig. 7, which is essential for use cases like virtual tour.", + "bbox": [ + 496, + 750, + 890, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement: This research is supported by Building 4.0 CRC and the National Key R&D Program of China (NO.2022ZD0160101), and was partially done at Shanghai AI Laboratory.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6354", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Naofumi Akimoto, Yuhi Matsuo, and Yoshimitsu Aoki. *Diverse plausible 360-degree image outpainting for efficient 3dgc background creation*. In *CVPR*, pages 11441–11450, 2022. 1, 2", + "[2] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In ICML, pages 1737-1752. PMLR, 2023. 2, 3", + "[3] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. 3DV, 2017. 5, 1, 2", + "[4] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM TOG, 41(6):1-16, 2022. 1, 2, 4, 5, 6, 7", + "[5] Mohammad Reza Karimi Dastjerdi, Yannick Hold-Geoffroy, Jonathan Eisenmann, Siavash Khodadadeh, and Jean-François Lalonde. Guided co-modulated gan for $360^{\\circ}$ field of view extrapolation. In 3DV, pages 475–485. IEEE, 2022. 2", + "[6] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, pages 8780-8794, 2021. 2", + "[7] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2", + "[8] Chuan Fang, Xiaotao Hu, Kunming Luo, and Ping Tan. Ctrlroom: Controllable text-to-3d room meshes generation with layout constraints. arXiv preprint arXiv:2310.03602, 2023. 2, 3, 5", + "[9] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. arXiv preprint arXiv:2302.01133, 2023. 3", + "[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2", + "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, pages 6626-6637, 2017. 5", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 2, 3", + "[13] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In ICCV, pages 7909-7920, 2023. 2, 3, 5", + "[14] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3, 6, 7, 4", + "[15] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, pages 26565-26577, 2022. 2" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Yuseung Lee, Kunho Kim, Hyunjin Kim, and Minhyuk Sung. Syncdiffusion: Coherent montage via synchronized joint diffusions. In NeurIPS, 2023. 2", + "[17] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. In NeurlPS, 2023. 1, 2, 3, 5", + "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 5, 1", + "[19] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, pages 5775-5787, 2022. 2", + "[20] Zhuqiang Lu, Kun Hu, Chaoyue Wang, Lei Bai, and Zhiyong Wang. Autoregressive omni-aware outpainting for open-vocabulary 360-degree image generation. arXiv preprint arXiv:2309.03467, 2023.1, 2", + "[21] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 2", + "[22] Jiafeng Mao, Xueting Wang, and Kiyoharu Aizawa. Guided image synthesis via initial image editing in diffusion model. In ACM MM, pages 5321-5329. ACM, 2023. 4, 7", + "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421. Springer, 2020. 4", + "[24] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2", + "[25] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, pages 16784-16804. PMLR, 2022. 2", + "[26] Changgyoon Oh, Wonjune Cho, Yujeong Chae, Daehee Park, Lin Wang, and Kuk-Jin Yoon. Bips: Bi-modal indoor panorama synthesis via residual depth-aided adversarial learning. In ECCV, pages 352–371. Springer, 2022. 2, 5, 1", + "[27] Chi-Han Peng and Jiayao Zhang. High-resolution depth estimation for 360deg panoramicas through perspective and panoramic depth images registration. In WACV, pages 3116-3125, 2023. 4", + "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 5, 4" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6355", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Manuel Rey-Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution 360deg monocular depth estimation. In CVPR, pages 3762-3772, 2022. 4", + "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 6, 7, 1, 4", + "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 3", + "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 2", + "[34] RunwayML. Stable diffusion. https://github.com/runwayml/stable-diffusion, 2021.2,3", + "[35] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2", + "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, pages 36479-36494, 2022. 2", + "[37] Tim Salimans, Ian J. Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. In NeurIPS, pages 2226-2234, 2016. 5", + "[38] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 2", + "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. In NeurIPS, pages 25278–25294, 2022. 1, 2", + "[40] Ka Chun Shum, Hong-Wing Pang, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Conditional 360-degree image synthesis for immersive indoor scene decoration. In ICCV, pages 4478–4488, 2023. 3", + "[41] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2, 3", + "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3, 1", + "[43] Liangchen Song, Liangliang Cao, Hongyu Xu, Kai Kang, Feng Tang, Junsong Yuan, and Zhao Yang. Roomdreamer: Text-driven 3d indoor scene synthesis with coherent geometry and texture. In ACM MM, pages 6898-6906. ACM, 2023. 2, 4" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, pages 11895–11907, 2019. 2, 3", + "[45] Cheng Sun, Chi-Wei Hsiao, Min Sun, and Hwann-Tzong Chen. Horizonnet: Learning room layout with 1d representation and pano stretch data augmentation. In CVPR, pages 1047–1056, 2019. 5, 8", + "[46] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 5", + "[47] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. In NeurIPS, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[48] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, JiaBin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In CVPR, pages 16773- 16783,2023.4", + "[49] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.1", + "[50] Fu-En Wang, Yu-Hsuan Yeh, Min Sun, Wei-Chen Chiu, and Yi-Hsuan Tsai. Layouttmp3d: Layout annotation of matter-port3d. arXiv preprint arXiv:2003.13516, 2020. 1", + "[51] Guangcong Wang, Yinuo Yang, Chen Change Loy, and Zwei Liu. Stylelight: HDR panorama generation for lighting estimation and editing. In ECCV, pages 477-492. Springer, 2022. 1, 2", + "[52] Guangcong Wang, Peng Wang, Zhaoxi Chen, Wenping Wang, Chen Change Loy, and Ziwei Liu. Perf: Panoramic neural radiance field from a single panorama. arXiv preprint arXiv:2310.16831, 2023. 2", + "[53] Hai Wang, Xiaoyu Xiang, Yuchen Fan, and Jing-Hao Xue. Customizing 360-degree panoramas through text-to-image diffusion models. In WACV, 2024. 2", + "[54] Jionghao Wang, Ziyu Chen, Jun Ling, Rong Xie, and Li Song. 360-degree panorama generation from few unregistered nfov images. In ACM MM, pages 6811-6821. ACM, 2023. 2, 3", + "[55] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 2", + "[56] Tianhao Wu, Chuanxia Zheng, and Tat-Jen Cham. IPO-ldm: Depth-aided 360-degree indoor rgb panorama outpainting via latent diffusion model. arXiv preprint arXiv:2307.03177, 2023. 2, 3", + "[57] Jiale Xu, Jia Zheng, Yanyu Xu, Rui Tang, and Shenghua Gao. Layout-guided novel view synthesis from a single indoor panorama. In CVPR, pages 16438-16447, 2021. 5", + "[58] Mai Xu, Chen Li, Shanyi Zhang, and Patrick Le Callet. State-of-the-art in 360 video/image processing: Perception, assessment and compression. JSTSP, 14(1):5-26, 2020. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6356", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Jian-Ru Xue, Jian-Wu Fang, and Pu Zhang. A survey of scene understanding by event reasoning in autonomous driving. MIR, 15(3):249-266, 2018. 1", + "[60] Bangbang Yang, Yinda Zhang, Yijin Li, Zhaopeng Cui, Sean Fanello, Hujun Bao, and Guofeng Zhang. Neural rendering in a room: amodal 3d understanding and free-viewpoint rendering for the closed scene composed of pre-captured objects. ACM TOG, 41(4):1-10, 2022. 1", + "[61] Bangbang Yang, Wenqi Dong, Lin Ma, Wenbo Hu, Xiao Liu, Zhaopeng Cui, and Yuewen Ma. Dreamspace: Dreaming your room space with text-driven panoramic texture propagation. arXiv preprint arXiv:2310.13119, 2023. 1, 2", + "[62] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM CSUR, 2022. 2", + "[63] Jason J. Yu, Fereshteh Forghani, Konstantinos G. Derpanis, and Marcus A. Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In ICCV, 2023. 2", + "[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 2, 5", + "[65] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming yu Liu. Diffcollage: Parallel generation of large content with diffusion models. In CVPR, 2023. 2", + "[66] Chuanqing Zhuang, Zhengda Lu, Yiqun Wang, Jun Xiao, and Ying Wang. Acdnet: Adaptively combined dilated convolution for monocular panorama depth estimation. In AAAI, pages 3653-3661, 2022. 3", + "[67] Chuhang Zou, Jheng-Wei Su, Chi-Han Peng, Alex Colburn, Qi Shan, Peter Wonka, Hung-Kuo Chu, and Derek Hoiem. Manhattan room layout reconstruction from a single 360 image: A comparative study of state-of-the-art methods. IJCV, 129:1410-1431, 2021. 1" + ], + "bbox": [ + 78, + 90, + 468, + 585 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6357", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/3259ff45-5fcb-4059-9736-190d9ce81cb7_model.json b/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/3259ff45-5fcb-4059-9736-190d9ce81cb7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5f32d6e8d0941c587dce12c9daf8e5efcf97ea7c --- /dev/null +++ b/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/3259ff45-5fcb-4059-9736-190d9ce81cb7_model.json @@ -0,0 +1,2466 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.133, + 0.131, + 0.838, + 0.154 + ], + "angle": 0, + "content": "Taming Stable Diffusion for Text to \\(360^{\\circ}\\) Panorama Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.18, + 0.821, + 0.217 + ], + "angle": 0, + "content": "Cheng Zhang\\(^{1,3}\\) Qianyi Wu\\(^{1}\\) Camilo Cruz Gambardella\\(^{1,3}\\) Xiaoshui Huang\\(^{2*}\\) \nDinh Phung\\(^{1}\\) Wanli Ouyang\\(^{2}\\) Jianfei Cai\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.217, + 0.882, + 0.235 + ], + "angle": 0, + "content": "1Monash University 2Shanghai AI Laboratory 3Building 4.0 CRC, Caulfield East, Victoria, Australia" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.275, + 0.488, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.275, + 0.886, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.537, + 0.893, + 0.593 + ], + "angle": 0, + "content": "Figure 1. Our PanFusion can generate realistic and consistent \\(360^{\\circ}\\) horizontal by \\(180^{\\circ}\\) vertical FOV panoramas from a single text prompt, compared to the limited FOV of current state-of-the-art method MVDiffusion [47]. Left: PanFusion addresses the problem of repetitive elements (duplicated \"ceiling fans\") and inconsistency (the ceiling and wall in the center) of MVDiffusion. Right: While trained mostly on indoor scenes, PanFusion can generalize well to out-of-domain outdoor prompts with more reasonable layout." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.608, + 0.314, + 0.622 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.474, + 0.88 + ], + "angle": 0, + "content": "Generative models, e.g., Stable Diffusion, have enabled the creation of photorealistic images from text prompts. Yet, the generation of 360-degree panorama images from text remains a challenge, particularly due to the dearth of paired text-panorama data and the domain gap between panorama and perspective images. In this paper, we introduce a novel dual-branch diffusion model named PanFusion to generate a 360-degree image from a text prompt. We leverage the stable diffusion model as one branch to provide prior knowledge in natural image generation and register it to another panorama branch for holistic image generation. We propose a unique cross-attention mechanism with projection awareness to minimize distortion during the collaborative denoising process. Our experiments validate that PanFusion surpasses existing methods and, thanks to its dual-branch structure, can integrate additional constraints like room layout for customized panorama outputs." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.607, + 0.631, + 0.623 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.776 + ], + "angle": 0, + "content": "Creating a \\(360^{\\circ}\\) panorama image from textual prompts is a nascent yet pivotal frontier in computer vision, with profound implications for applications that require extensive environmental representation, such as environmental lighting [1, 4, 51], VR/AR [60, 61], autonomous driving [59], and visual navigation [17]. Despite recent significant strides in text-to-image synthesis, the leap to generating full \\(360^{\\circ}\\) horizontal by \\(180^{\\circ}\\) vertical field-of-view (FOV) panorama remains challenging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.893, + 0.903 + ], + "angle": 0, + "content": "There are two major hurdles for achieving this goal. The first hurdle is data scarcity. The availability of text-to-panorama image pairs [20, 47] is significantly less compared with the abundance of text-to-common image pairs [38, 39]. The dearth of data complicates the training and finetuning of generative models. The second hurdle lies in the geometric and domain variations. Panorama images are distinct not only in their aspect ratio \\((2:1)\\) but" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.889, + 0.222, + 0.902 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6347" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "also in the underlying equirectangular projection (ERP) geometry [58]. This differs significantly from typical square images of the perspective projection that are used in most generative model training [38, 39]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.471, + 0.428 + ], + "angle": 0, + "content": "To mitigate the scarcity of panorama-specific training data, the previous solutions follow a common principle that leverages the prior knowledge of the pre-trained generative model [17, 20, 47]. However, taming powerful models like stable diffusion [31, 34] to generate high-fidelity panorama images remains a non-trivial task. Early attempts turn to formulate the 360-degree generation as an iterative image inpainting or warping process [17, 20]. Such solutions suffer from error accumulation and fail to handle loop closure [47]. To address this, MVDiffusion [47] proposes to produce multiple perspective images simultaneously by introducing a correspondence-aware attention module to facilitate multiview consistency, and then stitch together the perspective images to form a complete panorama. Despite the improved performance, the pixel-level consistency between neighboring perspectives in MVDiffusion cannot ensure global consistency, often resulting in repetitive elements or semantic inconsistency, as illustrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.471, + 0.807 + ], + "angle": 0, + "content": "Therefore, in this paper, we propose a novel dual-branch diffusion model called PanFusion that is tailored to address the limitations of prior models for high-quality text to \\(360^{\\circ}\\) panorama image generation. Specifically, PanFusion is designed to operate in both panorama and perspective domains, employing a global branch for creating a coherent panoramic \"canvas\" and a local branch that focuses on rendering detail-rich multiview perspectives. The local-global synergy of PanFusion significantly improves the resulting panoramas against the prevalent issues of error propagation and visual inconsistency that prior models have struggled with. To enhance the synergy between the two branches, we further propose an Equirectangular-Perspective Projection Attention (EPPA) mechanism, which respects the equirectangular projection for maintaining geometric integrity throughout the generation process. In addition, our adoption of parametric mapping for positional encoding is another leap forward, enhancing the model's spatial awareness and further ensuring the consistency of the generated panorama. Moreover, the panorama branch of PanFusion can be easily leveraged to accommodate supplementary control inputs at the panorama level, such as room layout, allowing for the creation of images that adhere to precise spatial conditions. We summarize our primary contributions as follows." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.811, + 0.468, + 0.87 + ], + "angle": 0, + "content": "- We pioneer a dual-branch diffusion model PanFusion, harnessing both the global panorama and local perspective latent domains, to generate high-quality, consistent \\(360^{\\circ}\\) pancreas from text prompts." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.468, + 0.9 + ], + "angle": 0, + "content": "- To enhance the interaction between the two branches, we introduce an \"equirectangular-perspective projection at" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.811, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "tention” mechanism that establishes a novel correspondence between global panorama and local perspective branches, addressing the unique projection challenges of panorama synthesis." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.153, + 0.892, + 0.228 + ], + "angle": 0, + "content": "- Our PanFusion not only surpasses prior models in quality and consistency but also supports extended control over the generation process with the inclusion of room layout. Extensive experimental results demonstrate the superiority of our proposed framework." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.241, + 0.642, + 0.258 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.267, + 0.892, + 0.493 + ], + "angle": 0, + "content": "Diffusion models. In recent years, diffusion models [6, 12, 41, 42, 44, 62] have taken the world of image generation by storm, as they have become faster [15, 19, 42] and more capable in terms of image quality and resolution [28, 31, 36]. This success has prompted the development of various applications for diffusion models, such as text-to-image [25, 28, 31, 36], image-conditioned generation [24, 64], in-painting [21, 35] and subject-driven generation [10, 33]. Most of these applications try to exploit the prior knowledge of a pre-trained diffusion model to mitigate the scarcity of task-specific data, by either finetuning with techniques like LoRA [14], or introducing auxiliary modules to distill the knowledge. We also adopt the same principle to harness the power of a pre-trained latent diffusion model [31] for panorama image generation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.495, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Panorama generation. Panorama image generation encompasses different settings, including panorama outpainting and text-to-panorama generation. Panorama outpainting [1, 5, 26, 51, 54, 56] focus on generating a 360-degree panorama from a partial input image. Various methods, such as StyleLight [51] and BIPS [26], have addressed specific use cases, focusing on HDR environment lighting and robotic guidance scenarios. Recent works [54, 56] have improved realism with diffusion models, but often lack the exploitation of rich prior information from pre-trained models, limiting generalization. On the other hand, recent developments in generative models have opened new frontiers in synthesizing immersive visual content from textual input [8, 13, 43, 52, 53, 55, 56, 61, 63]. As an image-based representation, generating panorama from text has gained much attention. Text2Light [4] adopt the VQGAN [7] structure to synthesize an HDR panorama image from text. To generate in arbitrary resolution with pre-trained diffusion models, DiffCollage [65], MultiDiffusion [2] and SyncDiffusion [16] propose to fuse the diffusion paths, while PanoGen [17] solves by iteratively inpainting. However, they failed to model the equirectangular projection of 360-degree panoramas. Lu et al. [20] adopts an autoregressive framework, but suffers from inefficient issues. MVDiffusion [47] designs a correspondence-aware attention module to produce multi-view images simultaneously that can be stitched together but results in repetitive elements and in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6348" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.08, + 0.877, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.302, + 0.895, + 0.373 + ], + "angle": 0, + "content": "Figure 2. Our proposed dual-branch PanFusion pipeline. The panorama branch (upper) provides global layout guidance and registers the perspective information to get seamless panorama output. The perspective branch (lower) harnesses the rich prior knowledge of Stable Diffusion (SD) and provides guidance to alleviate distortion under perspective projection. Both branches employ the same UNet backbone with shared weights, while finetuned with separate LoRA layers. Equirectangular-Perspective Projection Attention (EPPA) modules are plugged into different layers of the UNet to pass information between the two branches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.471, + 0.462 + ], + "angle": 0, + "content": "consistency. In contrast, our proposed PanFusion, a dual-branch framework, addresses the limitations of existing methods by considering both global panorama views and local perspective views, providing a comprehensive solution for text-driven 360-degree panorama image generation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.473, + 0.17, + 0.488 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.206, + 0.513 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.673 + ], + "angle": 0, + "content": "First introduced in [41], Diffusion models [12, 42, 44] aim to generate images from a noise distribution by iterative denoising with a learned prior distribution. However, the early diffusion models operate in the image space, which is of high dimension and complex. We have recently witnessed the huge success of latent diffusion models [31] that first transform an image \\( x \\) to a latent representation \\( z \\) with a learned encoder \\( \\mathcal{E} \\) and then train a UNet [32] model \\( \\epsilon_{\\theta} \\) parameterized by \\( \\theta \\) for the reverse process in the latent space, formulating the training objective as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.68, + 0.47, + 0.698 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {\\mathcal {E} (x), t, \\epsilon , y} [ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\tau (y)) | | _ {2} ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.809 + ], + "angle": 0, + "content": "where \\(\\tau(y)\\) is an encoding of input condition \\(y\\) (e.g., text prompt or image), \\(z_{t}\\) is the latent map at time step \\(t\\), and \\(\\epsilon\\) is sampled from a Gaussian noise. To sample from the model, \\(z_{t}\\) is first initialized from standard Gaussian distribution, then the reverse process is applied iteratively to generate \\(z_{0}\\), which is finally decoded into image space with a decoder \\(\\mathcal{D}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.342, + 0.832 + ], + "angle": 0, + "content": "3.2. Dual-Branch Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Directly employing pre-trained latent diffusion models [31], e.g., Stable Diffusion (SD) [34], to generate panorama from multiple perspective images, either in an iterative manner [9, 13, 17] or in a synchronized way [2, 47], would" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.612 + ], + "angle": 0, + "content": "fail to handle loop closure [47] or produce repetitive elements (Fig. 1) due to lack of global understanding. To address this issue, we propose a dual-branch diffusion model, which consists of a panorama branch and a perspective branch both based on the UNet of SD as shown in Fig. 2. The panorama branch is designed to provide global layout guidance and to register the perspective information to get the final panorama without stitching, while the perspective branch is designed to exploit the rich perspective image-generation capabilities of SD and to provide guidance to alleviate distortion under perspective projection. The two branches work together during the diffusion process to generate a denoised panorama latent map. Finally, this latent map runs through the pre-trained decoder \\(\\mathcal{D}\\) of SD to produce the final panorama image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Panorama Branch. Given a text prompt \\( y \\), our goal is to generate a panorama \\( x \\in \\mathbb{R}^{3 \\times H \\times W} \\), where \\( W = 2H \\), \\( H = 512 \\). To account for the difference between the target resolution and the one that SD is trained on, we introduce LoRA [14] layers to adapt the model to the new resolution. SD with LoRA can already serve as a strong baseline for panorama generation, but the results are not loop-consistent. Previous works [8, 56] have attempted to overcome this problem using 90 degree rotations of the latent map at each diffusion step. However, the seams are still obvious [56]. On close inspection of the SD model we have observed that the loop inconsistency is mainly caused by the convolutional layers in the UNet backbone, due to the lack of a mechanism to pass information between the two ends of the panorama image. Therefore, we introduce an adaptation to the UNet by adding additional circular padding [40, 54, 66] before each convolutional layer, and then cropping the output feature maps to the original size. In addition, we also add circular padding to the latent map before decoding to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6349" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "mitigate the less apparent loop inconsistency caused by the decoder. The combination of the techniques outlined above - latent rotation and circular padding - enable the generation of loop-consistent results with negligible computational cost, and can thus serve as another strong baseline. However, these alone do not make full use of the perspective generation capabilities that SD possess." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.471, + 0.38 + ], + "angle": 0, + "content": "Perspective Branch. The perspective branch aims to exploit the outstanding capabilities that SD has shown in the generation perspective images. Since it does not generate panoramas directly, it can operate at a lower resolution. Specifically, we set its resolution to \\( H / 2 \\times H / 2 \\) and add LoRA layers to adapt to the new resolution. To evenly distribute perspective cameras and fully cover the panorama image, we sample \\( N = 20 \\) cameras with poses \\( R^i \\in SO(3), i \\in 1, \\dots, N \\) on an icosahedron, similar to [27, 30], and set the FOV \\( = 90^\\circ \\). We input the same prompt \\( y \\) to both branches, and leave it to the model to decide how to exploit the prompt." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.471, + 0.653 + ], + "angle": 0, + "content": "Joint Latent Map Initialization. Previous work [22] has shown that the latent map initialization can affect the layout of the generated image, i.e., the initial noise can be modified to manipulate the layout of the generated image. We find this is particularly important for multi-image generation with correspondence, as the overlapping regions in different views tend to generate different elements if noise is sampled independently. Under this observation, we propose to jointly sample noise for the panorama and perspective latent maps by projecting noise from the panorama to the perspective. To be more specific, we first initialize panorama latent map \\( z_{T}^{*} \\in \\mathbb{R}^{4 \\times H / f \\times W / f} \\) as Gaussian noise, then initialize perspective latent maps with \\( z_{T}^{i} = \\mathrm{P}(z_{T}^{*}, R^{i}, \\mathrm{FOV}, (H / 2f, H / 2f)), z_{T}^{i} \\in \\mathbb{R}^{4 \\times H / 2f \\times H / 2f} \\), where \\( \\mathrm{P}(\\cdot) \\) is the function that projects \\( z_{T}^{*} \\) to a perspective view, and \\( f \\) is the down-sampling factor the encoder \\( \\mathcal{E} \\) used in SD. We use nearest interpolation for projection as it introduces fewer artifacts than bilinear interpolation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.665, + 0.226, + 0.68 + ], + "angle": 0, + "content": "3.3. EPP Attention" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To pass the guidance between the perspective and panorama branches, a naive solution is to project the feature maps in different layers from one branch to the other, similar to [47]. However, this will introduce information loss during interpolation and will restrict the receptive field to a small region around corresponding pixels. Instead, we propose an Equirectangular-Perspective Projection Attention (EPPA) module to implicitly pass the guidance in a cross-attention way as shown in Fig. 2. The EPPA operates on feature maps in different layers of two UNet branches, denoted as \\( F^{*} \\in \\mathbb{R}^{c \\times h \\times w} \\) and \\( F^{i} \\in \\mathbb{R}^{c \\times h / 2 \\times h / 2} \\), where \\( c \\) is the channel dimension, \\( h \\) and \\( w \\) are the height and width of the feature maps, respectively. In addition, while cross-attention bypasses interpolation and provides benefits from" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.089, + 0.887, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.292, + 0.892, + 0.348 + ], + "angle": 0, + "content": "Figure 3. Equirectangular-Perspective Projection (EPP) Attention. As EPP attention module is designed to be bijective to pass information in both directions, we only illustrated the direction of registering perspective information to panorama." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.892, + 0.436 + ], + "angle": 0, + "content": "the global receptive field of the attention mechanism, it is unaware of the projection between different formats. To address this issue, as shown in Fig. 3, we introduce two key components: EPP spherical positional encoding and EPP attention mask." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.892, + 0.575 + ], + "angle": 0, + "content": "EPP Spherical Positional Encoding. Since the EPPA module tries to associate two different formats, it is important to add positional information in the same space so that the attention mechanism can learn the correspondence. We achieve this by introducing Spherical Positional Encoding (SPE) [4, 23] to the EPPA module. The \\(\\mathrm{SPE}(\\theta ,\\phi) = (\\gamma (\\theta),\\gamma (\\phi))\\) function maps polar coordinates \\((\\theta ,\\phi)\\) to a higher dimensional space \\(\\mathbb{R}^{4L}\\) with Fourier positional encoding:" + }, + { + "type": "equation", + "bbox": [ + 0.505, + 0.58, + 0.891, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\gamma (\\theta) = \\left[ \\sin \\left(2 ^ {0} \\pi \\theta\\right), \\cos \\left(2 ^ {0} \\pi \\theta\\right), \\dots , \\sin \\left(2 ^ {L - 1} \\pi \\theta\\right), \\cos \\left(2 ^ {L - 1} \\pi \\theta\\right) \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Here we set \\( L = c / 4 \\) so that \\( \\mathrm{SPE}(\\theta, \\phi) \\in \\mathbb{R}^c \\). In the EPPA module, we apply SPE by first computing an SPE map for the panorama feature map, then project it to each perspective feature map with projection function \\( \\mathrm{P}(\\cdot) \\), so that corresponding pixels in different formats share the same SPE vector. Finally, the SPE maps are added to the feature maps accordingly and go through a linear layer to get the query \\( Q \\) and key \\( K \\), following a matrix product to get the affinity matrix \\( A \\). Take the perspective-to-panorama direction as an example as shown in Fig. 3, where \\( Q \\in \\mathbb{R}^{c \\times h \\times w} \\) and \\( K \\in \\mathbb{R}^{N \\times c \\times h / 2 \\times h / 2} \\) are reshaped and multiplied to get the affinity matrix \\( A \\in \\mathbb{R}^{hw \\times Nh^2 / 4} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "EPP Attention Mask. In addition to the SPE, we also propose an EPP attention mask to encourage the attention mechanism to focus around the corresponding pixels, inspired by [48]. For example, for a target pixel in the panorama feature map as in Fig. 3, we would like to focus on registering the information from corresponding source pixels in the perspective feature maps. We achieve" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6350" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.259 + ], + "angle": 0, + "content": "this by enhancing the affinity matrix \\( A \\) with a soft mask \\( M \\in \\mathbb{R}^{hw \\times Nh^2 / 4} \\) highlighting the corresponding pixels. To generate \\( M \\), we first use \\( \\mathrm{P}(\\cdot) \\) to project a binary mask \\( M_{j,k} \\) for each pixel \\( (j, k) \\) in panorama feature map to each perspective view \\( i \\). Then we apply a Gaussian kernel to smooth the masks and normalized them to \\( [-1, 1] \\). The masks are stacked and reshaped into \\( M \\), which is then added to \\( A \\) to get the enhanced affinity matrix \\( A' \\). The rest goes the same as the vanilla attention mechanism, where a softmax function is applied to \\( A' \\) to get the attention weights, which are finally multiplied with the value \\( V \\) to get the output." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.473, + 0.44 + ], + "angle": 0, + "content": "Inspired by [47, 64], we add zero-initialized \\(1 \\times 1\\) convolutional layers to the output of cross-attention and add it as a residual to the target feature map. This ensures the UNet stays unmodified at the beginning of training and can be gradually adapted to the EPPA modules. We add independent EPPA modules after each down-sampling layer and before each up-sampling layer of UNet to connect two branches, detailed in Sec. A of the supplementary material. Considering that the guidance is passed in both directions following the same equirectangular-perspective projection rule, which is bijective in nature, we share the weights of the EPPA modules in the two directions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.449, + 0.362, + 0.464 + ], + "angle": 0, + "content": "3.4. Layout-Conditioned Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.473, + 0.729 + ], + "angle": 0, + "content": "One of the most important applications for panorama generation is to generate panorama according to a given room layout. This is particularly useful for panorama novel view synthesis [57] and can potentially benefit indoor 3D scene generation [8, 13]. However, this problem is not well researched for diffusion-based panorama generation, largely due to the difficulty of introducing layout constraints while exploiting the rich prior knowledge of SD in perspective format at the same time. For panorama generation from multi-view [17, 47], one naive solution is to project layout condition into different views to locally condition the generation of perspective images. Instead, for our dual-branch diffusion model, we can naturally exploit the global nature of the panorama branch to enforce a much stronger layout constraint. Specifically, we render the layout condition as a distance map, then use it as the input of a ControlNet [64] to condition the panorama branch." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.737, + 0.182, + 0.753 + ], + "angle": 0, + "content": "3.5. Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.76, + 0.47, + 0.837 + ], + "angle": 0, + "content": "During training, we employ the same technique for latent map initialization to jointly sample noise for the panorama and perspective view, which we denote as \\(\\epsilon^{*}\\) and \\(\\epsilon^i\\), respectively. Given a GT panorama \\(x^{*}\\), we apply supervision on the panorama branch with the loss in Eq. (1):" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.845, + 0.469, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {*} = \\mathbb {E} _ {\\mathcal {E} \\left(x ^ {*}\\right), t, \\epsilon^ {*}, y} \\left[ \\| \\epsilon^ {*} - \\epsilon_ {\\theta} ^ {*} \\left(z _ {t} ^ {*}, t, \\tau (y)\\right) \\| _ {2} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}^{*}\\) is the predicted noise from the panorama branch. To encourage the synchronization between the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "two branches, we also apply supervision on the perspective branch with the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.135, + 0.892, + 0.154 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {i} = \\mathbb {E} _ {\\mathcal {E} (x ^ {i}), t, \\epsilon^ {i}, y} \\left[ | | \\epsilon^ {i} - \\epsilon_ {\\theta} ^ {i} \\left(z _ {t} ^ {i}, t, \\tau (y)\\right) | | _ {2} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.165, + 0.894, + 0.245 + ], + "angle": 0, + "content": "where \\( x^{i} \\) is a perspective image projected from \\( x^{*} \\) and \\( \\epsilon_{\\theta}^{i} \\) is the predicted noise from the perspective branch. We joint train the EPPA modules and the LoRA layers in the two branches by combining the above two losses as \\( \\mathcal{L} = \\mathcal{L}^{*} + \\frac{1}{N}\\sum_{i = 1}^{N}\\mathcal{L}^{i} \\). Note that the SD UNet blocks remain frozen." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.258, + 0.625, + 0.275 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.284, + 0.691, + 0.3 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.892, + 0.369 + ], + "angle": 0, + "content": "Dataset. We follow the MVDiffusion [47] to use the Matterport3D dataset [3], which has 10,800 panoramic images with 2,295 room layout annotations. We employ BLIP-2 [18] to generate a short description for each image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.892, + 0.445 + ], + "angle": 0, + "content": "Implementation Details. For text-conditioned generation, the training and inference schedules are kept the same as MVDiffusion [47] to make a fair comparison. For textlayout conditioned generation, we train the additional ControlNet with other parameters fixed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.522 + ], + "angle": 0, + "content": "Evaluation Metrics. Following previous works, we evaluate image quality in panorama [4] and perspective [47] domain. For layout-conditioned generation, we propose a new metric to evaluate how well the generated panorama follows input layout. Specifically, we use the following metrics:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.674 + ], + "angle": 0, + "content": "- Panorama. We follow Text2Light [4] to report Fréchet Inception Distance (FID) [11] and Inception Score (IS) [37] on panoramas to measure realism and diversity. Additionally, CLIP Score (CS) [29] is used to evaluate the text-image consistency. While FID is widely used for image generation, it relies on an Inception network [46] trained on perspective images, thus less applicable for panorama images. Therefore, a variant of FID customized for panorama, Fréchet Auto-Encoder Distance (FAED) [26], is used to better compare the realism." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.809 + ], + "angle": 0, + "content": "- Perspective. To simulate the real-world scenario where the user can freely navigate a panorama by viewing from different perspective views, we also report FID and IS for 20 randomly sampled views to compare with methods that generate \\(180^{\\circ}\\) vertical FOV. We also follow MVDiffusion [47] to report FID, IS and CS scores on 8 horizontally sampled views. It is worth noting that this group of metrics favors MVDiffusion by measuring its direct outputs, while ours involve interpolation for perspective views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.884 + ], + "angle": 0, + "content": "- Layout Consistency. We propose a layout consistency metric, which employs a layout estimation network HorizonNet [45] to estimate the room layout from the generated panorama and then compute its 2D IoU and 3D IoU [45] with the input layout condition." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.874, + 0.901 + ], + "angle": 0, + "content": "See Supplementary Sec. B for more details of the above." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6351" + } + ], + [ + { + "type": "aside_text", + "bbox": [ + 0.082, + 0.131, + 0.098, + 0.653 + ], + "angle": 270, + "content": "PanFusion (Ours) SD+LoRA MVDiffusion Text2Light" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.079, + 0.495, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.156, + 0.495, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.232, + 0.495, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.311, + 0.495, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.386, + 0.495, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.463, + 0.495, + 0.538 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.539, + 0.495, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.616, + 0.495, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.695, + 0.395, + 0.708 + ], + "angle": 0, + "content": "\"A living room with a chandelier.\"" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.079, + 0.888, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.156, + 0.888, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.233, + 0.888, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.311, + 0.888, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.386, + 0.888, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.463, + 0.888, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.539, + 0.888, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.616, + 0.888, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.695, + 0.794, + 0.708 + ], + "angle": 0, + "content": "\"A bedroom with a bed and a table.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.713, + 0.892, + 0.77 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons of text-conditioned panorama generation. We show panoramas cropped to the vertical FoV of MVDiffusion [47]. Below each panorama, we show 4 evenly spaced perspective projections, with the first view crossing the left and right boundaries. We highlight the loop inconsistency, distorted lines and repetitive objects and unreasonable furniture layout of baseline methods with corresponding colors of boxes, which are addressed by our method. More results can be found in Sec. E of the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.784, + 0.396, + 0.8 + ], + "angle": 0, + "content": "4.2. Comparisons with Previous Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.809, + 0.468, + 0.839 + ], + "angle": 0, + "content": "Baselines. We compare our PanFusion with the following baselines (see Supplementary Sec. B for details)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.902 + ], + "angle": 0, + "content": "- MVDiffusion [47] utilizes a multi-view diffusion model to generate 8 horizontal views that can be stitched into a panorama with vertical FOV of \\(90^{\\circ}\\). It requires separate prompts for training while providing an option to generate" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.786, + 0.646, + 0.8 + ], + "angle": 0, + "content": "from a single prompt." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.8, + 0.891, + 0.83 + ], + "angle": 0, + "content": "- Text2Light [4] generates a \\(180^{\\circ}\\) vertical FOV panorama from a text prompt in a two-stage auto-regressive manner." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.831, + 0.891, + 0.861 + ], + "angle": 0, + "content": "- \\( SD + LoRA \\) is our baseline model that uses LoRA [14] to finetune a Stable Diffusion [31] on panorama images." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.861, + 0.891, + 0.89 + ], + "angle": 0, + "content": "- Pano Branch is SD+LoRA with additional modifications as described in Sec. 3.2 to ensure loop consistency." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.8, + 0.891, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6352" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.149, + 0.076, + 0.817, + 0.196 + ], + "angle": 0, + "content": "
MethodPanorama20 ViewsHorizontal 8 Views [47]
FAED ↓FID ↓IS ↑CS ↑FID ↓IS ↑FID ↓IS ↑CS ↑
Text2Light [4]97.2476.53.6027.4836.255.6743.664.9225.88
MVDiffusion [47]------25.276.9026.34
SD+LoRA [14, 31]7.1951.694.4028.8319.326.9020.686.4824.77
Pano Branch7.9050.404.5428.6720.107.0620.566.3724.85
PanFusion (Ours)6.0446.474.3628.5817.046.8519.886.5024.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.201, + 0.895, + 0.258 + ], + "angle": 0, + "content": "Table 1. Comparison with SoTA methods. We evaluate the panorama image quality with Fréchet Auto-Encoder Distance (FAED), Fréchet Inception Distance (FID), Inception Score (IS), and CLIP Score (CS). In addition, we evaluate the perspective image quality in two settings. We first randomly sample 20 views, which is the closest to the real-world scenario where the user can freely navigate the panorama to view the scene from different perspectives. We then follow MVDiffusion [47] to horizontally sample 8 evenly spaced views." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.264, + 0.892, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.455, + 0.879, + 0.471 + ], + "angle": 0, + "content": "Figure 5. Ablation study. Artifacts are highlighted with red boxes and projected to perspective views. Prompt: \"A hallway in a hotel.\"" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.482, + 0.466, + 0.594 + ], + "angle": 0, + "content": "
MethodPanorama20 Views
FAED ↓FID ↓IS ↑CS ↑FID ↓IS ↑
Pano Branch7.9050.404.5428.6620.107.06
Ours-joint7.7559.434.3427.9429.686.82
Ours-SPE6.6349.554.6528.6620.827.00
Ours-mask6.7745.494.3528.6615.816.81
Ours-bijective7.3648.354.5828.6618.787.12
Ours6.0446.474.3628.5817.046.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.595, + 0.47, + 0.637 + ], + "angle": 0, + "content": "Table 2. Ablation study. We compare the ablated versions of our method on both panorama and perspective domains. Here, “-” indicates removing the subsequent component." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Quantitative Results. Tab. 1 shows the quantitative comparison results. Here, we assign the highest value to realism in image generation, measured in FAED and FID. On these two metrics, our method outperforms baseline methods in both panorama and perspective. For IS, our method's performance is slightly lower than baselines. This is likely due to the fact that IS evaluates diversity of objects in generated images, using a classifier, and our model, unlike the baselines, tends to not generate unexpected objects. Similarly, it is possible to say that baseline models present slightly higher CS is due to the repetition of objects reinforcing alignment with prompts. Considering SD+LoRA is superior to Pano Branch on FAED and is on par in other metrics, we only qualitatively compare with SD+LoRA below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 4 shows the qualitative comparison results. Loop inconsistency can be observed for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.892, + 0.621 + ], + "angle": 0, + "content": "Text2Light and SD+LoRA due to lack of message passing between left-right boundaries. They also suffer from distorted lines in perspective views, meaning the generated panorama does not follow the correct equirectangular projection. MVDiffusion on the other hand suffers from repetitive objects and unreasonable furniture layout, which is likely due to the lack of global context. Our method generates the most realistic scenes and aligns to text condition the best, also with less distortion in perspective views." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.633, + 0.655, + 0.65 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.779 + ], + "angle": 0, + "content": "In Sec. 4.2 and Tab. 1, we show that our full model outperforms Pano Branch, the baseline model of our method without the perspective branch. Here, as shown in Tab. 2 and Fig. 5, we further conduct an ablation study to validate the effectiveness of each component in our method. For a consistent comparison, we keep the layout similar between different ablated versions by sampling the same noise for latent map initialization, exploiting the observation of [22]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Joint latent map initialization. We ablate the joint latent map initialization by initializing the latent maps of the panorama and perspective branches separately (Ours-joint). Significant performance drop can be observed in all metrics and qualitative results, demonstrating the importance of joint latent map initialization. Interestingly, Ours-joint is even worse than Pano Branch in FID. It is likely due to that joint latent map initialization helps the corresponding" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6353" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.077, + 0.101, + 0.136 + ], + "angle": 270, + "content": "MVDiffusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.155, + 0.099, + 0.177 + ], + "angle": 270, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.077, + 0.42, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.136, + 0.419, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.199, + 0.375, + 0.212 + ], + "angle": 0, + "content": "\"A hallway in a building.\"" + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.077, + 0.497, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.136, + 0.496, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.077, + 0.814, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.136, + 0.816, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.816, + 0.077, + 0.892, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.816, + 0.136, + 0.892, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.199, + 0.812, + 0.212 + ], + "angle": 0, + "content": "\"A living room with a view of the ocean.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.217, + 0.894, + 0.259 + ], + "angle": 0, + "content": "Figure 6. Layout-conditioned generation comparisons. We showcase how layout-conditioned panorama generation can benefit from our dual-branch structure. The input layout is drawn on the generated panorama as yellow lines, while the panorama is cropped to the FoV of MVDiffusion and projected to one view to highlight the layout consistency." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.264, + 0.466, + 0.34 + ], + "angle": 0, + "content": "
MethodLayout ConsistencyHorizontal 8 Views [47]
3D IoU ↑2D IoU ↑FID ↓IS ↑CS ↑
MVDiffusion [47]61.0664.4328.835.6026.79
PanFusion (Ours)68.4671.8222.585.1026.04
GT image74.3177.15---
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.342, + 0.469, + 0.397 + ], + "angle": 0, + "content": "Table 3. Layout-conditioned comparisons. We evaluate layout consistency between the condition and the layout of each generated image extracted by a layout estimator HorizonNet [45]. \"GT image\" indicates the upper bound of the layout estimator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.469, + 0.457 + ], + "angle": 0, + "content": "pixels to share a similar noise distribution from the beginning of the diffusion process, which is essential for EPPA to align the content of overlapped regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.469, + 0.623 + ], + "angle": 0, + "content": "EPP SPE and attention mask. We ablate spherical positional encoding (Ours-SPE) and attention mask (Ours-mask) from EPPA module. From Tab. 2, we can see that missing SPE hurts FAED and FID, which is likely due to that SPE helps the model to learn the relative position of the pixels between the two branches. Missing attention mask gets FID better, but it hurts FAED, which is more accurate in evaluating the panorama quality as it is customized for the target dataset. Both result in clear artifacts around point light sources, inconsistent textures of the floor, and distortions in the highlighted projections, as shown in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.469, + 0.746 + ], + "angle": 0, + "content": "Bijective EPPA. We ablate the bijective EPPA (Ours-bijective) by using separate parameters for the two directions in the EPPA module. Both FAED and FID get worse for Ours-bijective. In addition, the ablated version struggles to generate consistent textures for the floor and ceiling for the two directions of the hallway in Fig. 5. On the contrary, our full model generates floor and ceiling with consistent style, showing a better global understanding of the scene." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.462, + 0.773 + ], + "angle": 0, + "content": "4.4. Application: Layout-Conditioned Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To showcase the advantage of our method in generating panorama images with additional layout conditions, we build a baseline model by adding a ControlNet to MVDiffusion as described in Sec. 3.4. We render the layout condition into a distance map and then project it to perspective views to condition its generation of multi-view images. The training settings are kept the same as our PanFusion. As shown in Tab. 3, our method outperforms the baseline" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.265, + 0.697, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.265, + 0.89, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.345, + 0.892, + 0.372 + ], + "angle": 0, + "content": "Figure 7. Failure case. Our method sometimes generates indoor scenes with no entrance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.39, + 0.892, + 0.495 + ], + "angle": 0, + "content": "model in layout consistency while keeping the realism advantage of perspective projections. We visualize the layout condition as wireframes overlaid on generated panorama images in Fig. 6, where we can see our generated panorama images follow their layout conditions better, especially as highlighted in perspective views. More details can be found in Sec. F of the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.511, + 0.618, + 0.527 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.748 + ], + "angle": 0, + "content": "In this paper, we have proposed PanFusion, a novel text-to- \\(360^{\\circ}\\)-panorama image generation method that can generate high-quality panorama images from a single text prompt. Particularly, a dual-branch diffusion architecture has been introduced to harness prior knowledge of Stable Diffusion in the perspective domain while addressing the repetitive elements and inconsistency issues observed in previous works. An EPPA module has been further introduced to enhance the information passing between the two branches. We have also extended our PanFusion for the application of layout-conditioned panorama generation. Comprehensive experiments have demonstrated that PanFusion can generate high-quality panorama images with better realism and layout consistency than previous methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Limitations. Although the dual-branch architecture of Pan-Fusion combines the advantages in both panorama and perspective domains, it comes with a cost of higher computational complexity. Additionally, our method sometimes fails to generate entrance for indoor scenes, as shown in Fig. 7, which is essential for use cases like virtual tour." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement: This research is supported by Building 4.0 CRC and the National Key R&D Program of China (NO.2022ZD0160101), and was partially done at Shanghai AI Laboratory." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6354" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Naofumi Akimoto, Yuhi Matsuo, and Yoshimitsu Aoki. *Diverse plausible 360-degree image outpainting for efficient 3dgc background creation*. In *CVPR*, pages 11441–11450, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.468, + 0.215 + ], + "angle": 0, + "content": "[2] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In ICML, pages 1737-1752. PMLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.218, + 0.468, + 0.272 + ], + "angle": 0, + "content": "[3] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. 3DV, 2017. 5, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.276, + 0.468, + 0.316 + ], + "angle": 0, + "content": "[4] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM TOG, 41(6):1-16, 2022. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.32, + 0.468, + 0.387 + ], + "angle": 0, + "content": "[5] Mohammad Reza Karimi Dastjerdi, Yannick Hold-Geoffroy, Jonathan Eisenmann, Siavash Khodadadeh, and Jean-François Lalonde. Guided co-modulated gan for \\(360^{\\circ}\\) field of view extrapolation. In 3DV, pages 475–485. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.392, + 0.468, + 0.433 + ], + "angle": 0, + "content": "[6] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, pages 8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.436, + 0.468, + 0.478 + ], + "angle": 0, + "content": "[7] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.468, + 0.535 + ], + "angle": 0, + "content": "[8] Chuan Fang, Xiaotao Hu, Kunming Luo, and Ping Tan. Ctrlroom: Controllable text-to-3d room meshes generation with layout constraints. arXiv preprint arXiv:2310.03602, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.538, + 0.468, + 0.58 + ], + "angle": 0, + "content": "[9] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. arXiv preprint arXiv:2302.01133, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.583, + 0.468, + 0.638 + ], + "angle": 0, + "content": "[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.641, + 0.468, + 0.696 + ], + "angle": 0, + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, pages 6626-6637, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.699, + 0.468, + 0.739 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.468, + 0.797 + ], + "angle": 0, + "content": "[13] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In ICCV, pages 7909-7920, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.801, + 0.468, + 0.856 + ], + "angle": 0, + "content": "[14] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3, 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[15] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, pages 26565-26577, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[16] Yuseung Lee, Kunho Kim, Hyunjin Kim, and Minhyuk Sung. Syncdiffusion: Coherent montage via synchronized joint diffusions. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[17] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. In NeurlPS, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.239, + 0.892, + 0.293 + ], + "angle": 0, + "content": "[19] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, pages 5775-5787, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.297, + 0.892, + 0.351 + ], + "angle": 0, + "content": "[20] Zhuqiang Lu, Kun Hu, Chaoyue Wang, Lei Bai, and Zhiyong Wang. Autoregressive omni-aware outpainting for open-vocabulary 360-degree image generation. arXiv preprint arXiv:2309.03467, 2023.1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.354, + 0.892, + 0.408 + ], + "angle": 0, + "content": "[21] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.412, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[22] Jiafeng Mao, Xueting Wang, and Kiyoharu Aizawa. Guided image synthesis via initial image editing in diffusion model. In ACM MM, pages 5321-5329. ACM, 2023. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.457, + 0.892, + 0.512 + ], + "angle": 0, + "content": "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421. Springer, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.515, + 0.892, + 0.57 + ], + "angle": 0, + "content": "[24] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.572, + 0.892, + 0.641 + ], + "angle": 0, + "content": "[25] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, pages 16784-16804. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.644, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[26] Changgyoon Oh, Wonjune Cho, Yujeong Chae, Daehee Park, Lin Wang, and Kuk-Jin Yoon. Bips: Bi-modal indoor panorama synthesis via residual depth-aided adversarial learning. In ECCV, pages 352–371. Springer, 2022. 2, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[27] Chi-Han Peng and Jiayao Zhang. High-resolution depth estimation for 360deg panoramicas through perspective and panoramic depth images registration. In WACV, pages 3116-3125, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 5, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "6355" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[30] Manuel Rey-Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution 360deg monocular depth estimation. In CVPR, pages 3762-3772, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 6, 7, 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.469, + 0.291 + ], + "angle": 0, + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.468, + 0.32 + ], + "angle": 0, + "content": "[34] RunwayML. Stable diffusion. https://github.com/runwayml/stable-diffusion, 2021.2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.469, + 0.376 + ], + "angle": 0, + "content": "[35] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.469, + 0.474 + ], + "angle": 0, + "content": "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, pages 36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.469, + 0.518 + ], + "angle": 0, + "content": "[37] Tim Salimans, Ian J. Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. In NeurIPS, pages 2226-2234, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[38] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. In NeurIPS, pages 25278–25294, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[40] Ka Chun Shum, Hong-Wing Pang, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Conditional 360-degree image synthesis for immersive indoor scene decoration. In ICCV, pages 4478–4488, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[41] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[43] Liangchen Song, Liangliang Cao, Hongyu Xu, Kai Kang, Feng Tang, Junsong Yuan, and Zhao Yang. Roomdreamer: Text-driven 3d indoor scene synthesis with coherent geometry and texture. In ACM MM, pages 6898-6906. ACM, 2023. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[44] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, pages 11895–11907, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[45] Cheng Sun, Chi-Wei Hsiao, Min Sun, and Hwann-Tzong Chen. Horizonnet: Learning room layout with 1d representation and pano stretch data augmentation. In CVPR, pages 1047–1056, 2019. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[46] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.252, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[47] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. In NeurIPS, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.364 + ], + "angle": 0, + "content": "[48] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, JiaBin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In CVPR, pages 16773- 16783,2023.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.368, + 0.892, + 0.436 + ], + "angle": 0, + "content": "[49] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.439, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[50] Fu-En Wang, Yu-Hsuan Yeh, Min Sun, Wei-Chen Chiu, and Yi-Hsuan Tsai. Layouttmp3d: Layout annotation of matter-port3d. arXiv preprint arXiv:2003.13516, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[51] Guangcong Wang, Yinuo Yang, Chen Change Loy, and Zwei Liu. Stylelight: HDR panorama generation for lighting estimation and editing. In ECCV, pages 477-492. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.541, + 0.892, + 0.595 + ], + "angle": 0, + "content": "[52] Guangcong Wang, Peng Wang, Zhaoxi Chen, Wenping Wang, Chen Change Loy, and Ziwei Liu. Perf: Panoramic neural radiance field from a single panorama. arXiv preprint arXiv:2310.16831, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.599, + 0.892, + 0.639 + ], + "angle": 0, + "content": "[53] Hai Wang, Xiaoyu Xiang, Yuchen Fan, and Jing-Hao Xue. Customizing 360-degree panoramas through text-to-image diffusion models. In WACV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.642, + 0.892, + 0.696 + ], + "angle": 0, + "content": "[54] Jionghao Wang, Ziyu Chen, Jun Ling, Rong Xie, and Li Song. 360-degree panorama generation from few unregistered nfov images. In ACM MM, pages 6811-6821. ACM, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.7, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[55] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.758, + 0.892, + 0.812 + ], + "angle": 0, + "content": "[56] Tianhao Wu, Chuanxia Zheng, and Tat-Jen Cham. IPO-ldm: Depth-aided 360-degree indoor rgb panorama outpainting via latent diffusion model. arXiv preprint arXiv:2307.03177, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[57] Jiale Xu, Jia Zheng, Yanyu Xu, Rui Tang, and Shenghua Gao. Layout-guided novel view synthesis from a single indoor panorama. In CVPR, pages 16438-16447, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[58] Mai Xu, Chen Li, Shanyi Zhang, and Patrick Le Callet. State-of-the-art in 360 video/image processing: Perception, assessment and compression. JSTSP, 14(1):5-26, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6356" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[59] Jian-Ru Xue, Jian-Wu Fang, and Pu Zhang. A survey of scene understanding by event reasoning in autonomous driving. MIR, 15(3):249-266, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[60] Bangbang Yang, Yinda Zhang, Yijin Li, Zhaopeng Cui, Sean Fanello, Hujun Bao, and Guofeng Zhang. Neural rendering in a room: amodal 3d understanding and free-viewpoint rendering for the closed scene composed of pre-captured objects. ACM TOG, 41(4):1-10, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[61] Bangbang Yang, Wenqi Dong, Lin Ma, Wenbo Hu, Xiao Liu, Zhaopeng Cui, and Yuewen Ma. Dreamspace: Dreaming your room space with text-driven panoramic texture propagation. arXiv preprint arXiv:2310.13119, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.263, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[62] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM CSUR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.32, + 0.469, + 0.373 + ], + "angle": 0, + "content": "[63] Jason J. Yu, Fereshteh Forghani, Konstantinos G. Derpanis, and Marcus A. Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.376, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.419, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[65] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming yu Liu. Diffcollage: Parallel generation of large content with diffusion models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.462, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[66] Chuanqing Zhuang, Zhengda Lu, Yiqun Wang, Jun Xiao, and Ying Wang. Acdnet: Adaptively combined dilated convolution for monocular panorama depth estimation. In AAAI, pages 3653-3661, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.519, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[67] Chuhang Zou, Jheng-Wei Su, Chi-Han Peng, Alex Colburn, Qi Shan, Peter Wonka, Hung-Kuo Chu, and Derek Hoiem. Manhattan room layout reconstruction from a single 360 image: A comparative study of state-of-the-art methods. IJCV, 129:1410-1431, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6357" + } + ] +] \ No newline at end of file diff --git a/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/3259ff45-5fcb-4059-9736-190d9ce81cb7_origin.pdf b/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/3259ff45-5fcb-4059-9736-190d9ce81cb7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d95e5482e666cf372ac6ad6f3b5d5fb60bd83e0f --- /dev/null +++ b/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/3259ff45-5fcb-4059-9736-190d9ce81cb7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e4670d39568afa8d5282f37040862069ebb8c58ca2a1014ebdcfca61772e5f5 +size 8670936 diff --git a/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/full.md b/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d8bd5225af6a077a67a11f8986ef5d40dbe59e32 --- /dev/null +++ b/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/full.md @@ -0,0 +1,326 @@ +# Taming Stable Diffusion for Text to $360^{\circ}$ Panorama Image Generation + +Cheng Zhang $^{1,3}$ Qianyi Wu $^{1}$ Camilo Cruz Gambardella $^{1,3}$ Xiaoshui Huang $^{2*}$ +Dinh Phung $^{1}$ Wanli Ouyang $^{2}$ Jianfei Cai $^{1*}$ + +1Monash University 2Shanghai AI Laboratory 3Building 4.0 CRC, Caulfield East, Victoria, Australia + +![](images/8e1203debcca5709d928d249739fd1c3c9569ac0f14258ee85a50ee0a40877ee.jpg) +Figure 1. Our PanFusion can generate realistic and consistent $360^{\circ}$ horizontal by $180^{\circ}$ vertical FOV panoramas from a single text prompt, compared to the limited FOV of current state-of-the-art method MVDiffusion [47]. Left: PanFusion addresses the problem of repetitive elements (duplicated "ceiling fans") and inconsistency (the ceiling and wall in the center) of MVDiffusion. Right: While trained mostly on indoor scenes, PanFusion can generalize well to out-of-domain outdoor prompts with more reasonable layout. + +![](images/abb1b99aa30d27822aa4ee9ff979ff1bcbe7557aa62a3b856476ba66d5e8e7b4.jpg) + +# Abstract + +Generative models, e.g., Stable Diffusion, have enabled the creation of photorealistic images from text prompts. Yet, the generation of 360-degree panorama images from text remains a challenge, particularly due to the dearth of paired text-panorama data and the domain gap between panorama and perspective images. In this paper, we introduce a novel dual-branch diffusion model named PanFusion to generate a 360-degree image from a text prompt. We leverage the stable diffusion model as one branch to provide prior knowledge in natural image generation and register it to another panorama branch for holistic image generation. We propose a unique cross-attention mechanism with projection awareness to minimize distortion during the collaborative denoising process. Our experiments validate that PanFusion surpasses existing methods and, thanks to its dual-branch structure, can integrate additional constraints like room layout for customized panorama outputs. + +# 1. Introduction + +Creating a $360^{\circ}$ panorama image from textual prompts is a nascent yet pivotal frontier in computer vision, with profound implications for applications that require extensive environmental representation, such as environmental lighting [1, 4, 51], VR/AR [60, 61], autonomous driving [59], and visual navigation [17]. Despite recent significant strides in text-to-image synthesis, the leap to generating full $360^{\circ}$ horizontal by $180^{\circ}$ vertical field-of-view (FOV) panorama remains challenging. + +There are two major hurdles for achieving this goal. The first hurdle is data scarcity. The availability of text-to-panorama image pairs [20, 47] is significantly less compared with the abundance of text-to-common image pairs [38, 39]. The dearth of data complicates the training and finetuning of generative models. The second hurdle lies in the geometric and domain variations. Panorama images are distinct not only in their aspect ratio $(2:1)$ but + +also in the underlying equirectangular projection (ERP) geometry [58]. This differs significantly from typical square images of the perspective projection that are used in most generative model training [38, 39]. + +To mitigate the scarcity of panorama-specific training data, the previous solutions follow a common principle that leverages the prior knowledge of the pre-trained generative model [17, 20, 47]. However, taming powerful models like stable diffusion [31, 34] to generate high-fidelity panorama images remains a non-trivial task. Early attempts turn to formulate the 360-degree generation as an iterative image inpainting or warping process [17, 20]. Such solutions suffer from error accumulation and fail to handle loop closure [47]. To address this, MVDiffusion [47] proposes to produce multiple perspective images simultaneously by introducing a correspondence-aware attention module to facilitate multiview consistency, and then stitch together the perspective images to form a complete panorama. Despite the improved performance, the pixel-level consistency between neighboring perspectives in MVDiffusion cannot ensure global consistency, often resulting in repetitive elements or semantic inconsistency, as illustrated in Fig. 1. + +Therefore, in this paper, we propose a novel dual-branch diffusion model called PanFusion that is tailored to address the limitations of prior models for high-quality text to $360^{\circ}$ panorama image generation. Specifically, PanFusion is designed to operate in both panorama and perspective domains, employing a global branch for creating a coherent panoramic "canvas" and a local branch that focuses on rendering detail-rich multiview perspectives. The local-global synergy of PanFusion significantly improves the resulting panoramas against the prevalent issues of error propagation and visual inconsistency that prior models have struggled with. To enhance the synergy between the two branches, we further propose an Equirectangular-Perspective Projection Attention (EPPA) mechanism, which respects the equirectangular projection for maintaining geometric integrity throughout the generation process. In addition, our adoption of parametric mapping for positional encoding is another leap forward, enhancing the model's spatial awareness and further ensuring the consistency of the generated panorama. Moreover, the panorama branch of PanFusion can be easily leveraged to accommodate supplementary control inputs at the panorama level, such as room layout, allowing for the creation of images that adhere to precise spatial conditions. We summarize our primary contributions as follows. + +- We pioneer a dual-branch diffusion model PanFusion, harnessing both the global panorama and local perspective latent domains, to generate high-quality, consistent $360^{\circ}$ pancreas from text prompts. +- To enhance the interaction between the two branches, we introduce an "equirectangular-perspective projection at + +tention” mechanism that establishes a novel correspondence between global panorama and local perspective branches, addressing the unique projection challenges of panorama synthesis. + +- Our PanFusion not only surpasses prior models in quality and consistency but also supports extended control over the generation process with the inclusion of room layout. Extensive experimental results demonstrate the superiority of our proposed framework. + +# 2. Related Work + +Diffusion models. In recent years, diffusion models [6, 12, 41, 42, 44, 62] have taken the world of image generation by storm, as they have become faster [15, 19, 42] and more capable in terms of image quality and resolution [28, 31, 36]. This success has prompted the development of various applications for diffusion models, such as text-to-image [25, 28, 31, 36], image-conditioned generation [24, 64], in-painting [21, 35] and subject-driven generation [10, 33]. Most of these applications try to exploit the prior knowledge of a pre-trained diffusion model to mitigate the scarcity of task-specific data, by either finetuning with techniques like LoRA [14], or introducing auxiliary modules to distill the knowledge. We also adopt the same principle to harness the power of a pre-trained latent diffusion model [31] for panorama image generation. + +Panorama generation. Panorama image generation encompasses different settings, including panorama outpainting and text-to-panorama generation. Panorama outpainting [1, 5, 26, 51, 54, 56] focus on generating a 360-degree panorama from a partial input image. Various methods, such as StyleLight [51] and BIPS [26], have addressed specific use cases, focusing on HDR environment lighting and robotic guidance scenarios. Recent works [54, 56] have improved realism with diffusion models, but often lack the exploitation of rich prior information from pre-trained models, limiting generalization. On the other hand, recent developments in generative models have opened new frontiers in synthesizing immersive visual content from textual input [8, 13, 43, 52, 53, 55, 56, 61, 63]. As an image-based representation, generating panorama from text has gained much attention. Text2Light [4] adopt the VQGAN [7] structure to synthesize an HDR panorama image from text. To generate in arbitrary resolution with pre-trained diffusion models, DiffCollage [65], MultiDiffusion [2] and SyncDiffusion [16] propose to fuse the diffusion paths, while PanoGen [17] solves by iteratively inpainting. However, they failed to model the equirectangular projection of 360-degree panoramas. Lu et al. [20] adopts an autoregressive framework, but suffers from inefficient issues. MVDiffusion [47] designs a correspondence-aware attention module to produce multi-view images simultaneously that can be stitched together but results in repetitive elements and in + +![](images/ed3a739337d81ae64bd181e429d79a0db48cff0fc1f855a18fdb90f214bf4498.jpg) +Figure 2. Our proposed dual-branch PanFusion pipeline. The panorama branch (upper) provides global layout guidance and registers the perspective information to get seamless panorama output. The perspective branch (lower) harnesses the rich prior knowledge of Stable Diffusion (SD) and provides guidance to alleviate distortion under perspective projection. Both branches employ the same UNet backbone with shared weights, while finetuned with separate LoRA layers. Equirectangular-Perspective Projection Attention (EPPA) modules are plugged into different layers of the UNet to pass information between the two branches. + +consistency. In contrast, our proposed PanFusion, a dual-branch framework, addresses the limitations of existing methods by considering both global panorama views and local perspective views, providing a comprehensive solution for text-driven 360-degree panorama image generation. + +# 3. Method + +# 3.1. Preliminary + +First introduced in [41], Diffusion models [12, 42, 44] aim to generate images from a noise distribution by iterative denoising with a learned prior distribution. However, the early diffusion models operate in the image space, which is of high dimension and complex. We have recently witnessed the huge success of latent diffusion models [31] that first transform an image $x$ to a latent representation $z$ with a learned encoder $\mathcal{E}$ and then train a UNet [32] model $\epsilon_{\theta}$ parameterized by $\theta$ for the reverse process in the latent space, formulating the training objective as: + +$$ +\mathcal {L} = \mathbb {E} _ {\mathcal {E} (x), t, \epsilon , y} [ | | \epsilon - \epsilon_ {\theta} (z _ {t}, t, \tau (y)) | | _ {2} ], \tag {1} +$$ + +where $\tau(y)$ is an encoding of input condition $y$ (e.g., text prompt or image), $z_{t}$ is the latent map at time step $t$ , and $\epsilon$ is sampled from a Gaussian noise. To sample from the model, $z_{t}$ is first initialized from standard Gaussian distribution, then the reverse process is applied iteratively to generate $z_{0}$ , which is finally decoded into image space with a decoder $\mathcal{D}$ . + +# 3.2. Dual-Branch Diffusion Model + +Directly employing pre-trained latent diffusion models [31], e.g., Stable Diffusion (SD) [34], to generate panorama from multiple perspective images, either in an iterative manner [9, 13, 17] or in a synchronized way [2, 47], would + +fail to handle loop closure [47] or produce repetitive elements (Fig. 1) due to lack of global understanding. To address this issue, we propose a dual-branch diffusion model, which consists of a panorama branch and a perspective branch both based on the UNet of SD as shown in Fig. 2. The panorama branch is designed to provide global layout guidance and to register the perspective information to get the final panorama without stitching, while the perspective branch is designed to exploit the rich perspective image-generation capabilities of SD and to provide guidance to alleviate distortion under perspective projection. The two branches work together during the diffusion process to generate a denoised panorama latent map. Finally, this latent map runs through the pre-trained decoder $\mathcal{D}$ of SD to produce the final panorama image. + +Panorama Branch. Given a text prompt $y$ , our goal is to generate a panorama $x \in \mathbb{R}^{3 \times H \times W}$ , where $W = 2H$ , $H = 512$ . To account for the difference between the target resolution and the one that SD is trained on, we introduce LoRA [14] layers to adapt the model to the new resolution. SD with LoRA can already serve as a strong baseline for panorama generation, but the results are not loop-consistent. Previous works [8, 56] have attempted to overcome this problem using 90 degree rotations of the latent map at each diffusion step. However, the seams are still obvious [56]. On close inspection of the SD model we have observed that the loop inconsistency is mainly caused by the convolutional layers in the UNet backbone, due to the lack of a mechanism to pass information between the two ends of the panorama image. Therefore, we introduce an adaptation to the UNet by adding additional circular padding [40, 54, 66] before each convolutional layer, and then cropping the output feature maps to the original size. In addition, we also add circular padding to the latent map before decoding to + +mitigate the less apparent loop inconsistency caused by the decoder. The combination of the techniques outlined above - latent rotation and circular padding - enable the generation of loop-consistent results with negligible computational cost, and can thus serve as another strong baseline. However, these alone do not make full use of the perspective generation capabilities that SD possess. + +Perspective Branch. The perspective branch aims to exploit the outstanding capabilities that SD has shown in the generation perspective images. Since it does not generate panoramas directly, it can operate at a lower resolution. Specifically, we set its resolution to $H / 2 \times H / 2$ and add LoRA layers to adapt to the new resolution. To evenly distribute perspective cameras and fully cover the panorama image, we sample $N = 20$ cameras with poses $R^i \in SO(3), i \in 1, \dots, N$ on an icosahedron, similar to [27, 30], and set the FOV $= 90^\circ$ . We input the same prompt $y$ to both branches, and leave it to the model to decide how to exploit the prompt. + +Joint Latent Map Initialization. Previous work [22] has shown that the latent map initialization can affect the layout of the generated image, i.e., the initial noise can be modified to manipulate the layout of the generated image. We find this is particularly important for multi-image generation with correspondence, as the overlapping regions in different views tend to generate different elements if noise is sampled independently. Under this observation, we propose to jointly sample noise for the panorama and perspective latent maps by projecting noise from the panorama to the perspective. To be more specific, we first initialize panorama latent map $z_{T}^{*} \in \mathbb{R}^{4 \times H / f \times W / f}$ as Gaussian noise, then initialize perspective latent maps with $z_{T}^{i} = \mathrm{P}(z_{T}^{*}, R^{i}, \mathrm{FOV}, (H / 2f, H / 2f)), z_{T}^{i} \in \mathbb{R}^{4 \times H / 2f \times H / 2f}$ , where $\mathrm{P}(\cdot)$ is the function that projects $z_{T}^{*}$ to a perspective view, and $f$ is the down-sampling factor the encoder $\mathcal{E}$ used in SD. We use nearest interpolation for projection as it introduces fewer artifacts than bilinear interpolation. + +# 3.3. EPP Attention + +To pass the guidance between the perspective and panorama branches, a naive solution is to project the feature maps in different layers from one branch to the other, similar to [47]. However, this will introduce information loss during interpolation and will restrict the receptive field to a small region around corresponding pixels. Instead, we propose an Equirectangular-Perspective Projection Attention (EPPA) module to implicitly pass the guidance in a cross-attention way as shown in Fig. 2. The EPPA operates on feature maps in different layers of two UNet branches, denoted as $F^{*} \in \mathbb{R}^{c \times h \times w}$ and $F^{i} \in \mathbb{R}^{c \times h / 2 \times h / 2}$ , where $c$ is the channel dimension, $h$ and $w$ are the height and width of the feature maps, respectively. In addition, while cross-attention bypasses interpolation and provides benefits from + +![](images/50449cba423ab2c448e7b4f1528b67d19e45599cbe2fb2d383ac966358af1826.jpg) +Figure 3. Equirectangular-Perspective Projection (EPP) Attention. As EPP attention module is designed to be bijective to pass information in both directions, we only illustrated the direction of registering perspective information to panorama. + +the global receptive field of the attention mechanism, it is unaware of the projection between different formats. To address this issue, as shown in Fig. 3, we introduce two key components: EPP spherical positional encoding and EPP attention mask. + +EPP Spherical Positional Encoding. Since the EPPA module tries to associate two different formats, it is important to add positional information in the same space so that the attention mechanism can learn the correspondence. We achieve this by introducing Spherical Positional Encoding (SPE) [4, 23] to the EPPA module. The $\mathrm{SPE}(\theta ,\phi) = (\gamma (\theta),\gamma (\phi))$ function maps polar coordinates $(\theta ,\phi)$ to a higher dimensional space $\mathbb{R}^{4L}$ with Fourier positional encoding: + +$$ +\gamma (\theta) = \left[ \sin \left(2 ^ {0} \pi \theta\right), \cos \left(2 ^ {0} \pi \theta\right), \dots , \sin \left(2 ^ {L - 1} \pi \theta\right), \cos \left(2 ^ {L - 1} \pi \theta\right) \right]. \tag {2} +$$ + +Here we set $L = c / 4$ so that $\mathrm{SPE}(\theta, \phi) \in \mathbb{R}^c$ . In the EPPA module, we apply SPE by first computing an SPE map for the panorama feature map, then project it to each perspective feature map with projection function $\mathrm{P}(\cdot)$ , so that corresponding pixels in different formats share the same SPE vector. Finally, the SPE maps are added to the feature maps accordingly and go through a linear layer to get the query $Q$ and key $K$ , following a matrix product to get the affinity matrix $A$ . Take the perspective-to-panorama direction as an example as shown in Fig. 3, where $Q \in \mathbb{R}^{c \times h \times w}$ and $K \in \mathbb{R}^{N \times c \times h / 2 \times h / 2}$ are reshaped and multiplied to get the affinity matrix $A \in \mathbb{R}^{hw \times Nh^2 / 4}$ . + +EPP Attention Mask. In addition to the SPE, we also propose an EPP attention mask to encourage the attention mechanism to focus around the corresponding pixels, inspired by [48]. For example, for a target pixel in the panorama feature map as in Fig. 3, we would like to focus on registering the information from corresponding source pixels in the perspective feature maps. We achieve + +this by enhancing the affinity matrix $A$ with a soft mask $M \in \mathbb{R}^{hw \times Nh^2 / 4}$ highlighting the corresponding pixels. To generate $M$ , we first use $\mathrm{P}(\cdot)$ to project a binary mask $M_{j,k}$ for each pixel $(j, k)$ in panorama feature map to each perspective view $i$ . Then we apply a Gaussian kernel to smooth the masks and normalized them to $[-1, 1]$ . The masks are stacked and reshaped into $M$ , which is then added to $A$ to get the enhanced affinity matrix $A'$ . The rest goes the same as the vanilla attention mechanism, where a softmax function is applied to $A'$ to get the attention weights, which are finally multiplied with the value $V$ to get the output. + +Inspired by [47, 64], we add zero-initialized $1 \times 1$ convolutional layers to the output of cross-attention and add it as a residual to the target feature map. This ensures the UNet stays unmodified at the beginning of training and can be gradually adapted to the EPPA modules. We add independent EPPA modules after each down-sampling layer and before each up-sampling layer of UNet to connect two branches, detailed in Sec. A of the supplementary material. Considering that the guidance is passed in both directions following the same equirectangular-perspective projection rule, which is bijective in nature, we share the weights of the EPPA modules in the two directions. + +# 3.4. Layout-Conditioned Generation + +One of the most important applications for panorama generation is to generate panorama according to a given room layout. This is particularly useful for panorama novel view synthesis [57] and can potentially benefit indoor 3D scene generation [8, 13]. However, this problem is not well researched for diffusion-based panorama generation, largely due to the difficulty of introducing layout constraints while exploiting the rich prior knowledge of SD in perspective format at the same time. For panorama generation from multi-view [17, 47], one naive solution is to project layout condition into different views to locally condition the generation of perspective images. Instead, for our dual-branch diffusion model, we can naturally exploit the global nature of the panorama branch to enforce a much stronger layout constraint. Specifically, we render the layout condition as a distance map, then use it as the input of a ControlNet [64] to condition the panorama branch. + +# 3.5. Training + +During training, we employ the same technique for latent map initialization to jointly sample noise for the panorama and perspective view, which we denote as $\epsilon^{*}$ and $\epsilon^i$ , respectively. Given a GT panorama $x^{*}$ , we apply supervision on the panorama branch with the loss in Eq. (1): + +$$ +\mathcal {L} ^ {*} = \mathbb {E} _ {\mathcal {E} \left(x ^ {*}\right), t, \epsilon^ {*}, y} \left[ \| \epsilon^ {*} - \epsilon_ {\theta} ^ {*} \left(z _ {t} ^ {*}, t, \tau (y)\right) \| _ {2} \right], \tag {3} +$$ + +where $\epsilon_{\theta}^{*}$ is the predicted noise from the panorama branch. To encourage the synchronization between the + +two branches, we also apply supervision on the perspective branch with the following loss: + +$$ +\mathcal {L} ^ {i} = \mathbb {E} _ {\mathcal {E} (x ^ {i}), t, \epsilon^ {i}, y} \left[ | | \epsilon^ {i} - \epsilon_ {\theta} ^ {i} \left(z _ {t} ^ {i}, t, \tau (y)\right) | | _ {2} \right], \tag {4} +$$ + +where $x^{i}$ is a perspective image projected from $x^{*}$ and $\epsilon_{\theta}^{i}$ is the predicted noise from the perspective branch. We joint train the EPPA modules and the LoRA layers in the two branches by combining the above two losses as $\mathcal{L} = \mathcal{L}^{*} + \frac{1}{N}\sum_{i = 1}^{N}\mathcal{L}^{i}$ . Note that the SD UNet blocks remain frozen. + +# 4. Experiment + +# 4.1. Experimental Setup + +Dataset. We follow the MVDiffusion [47] to use the Matterport3D dataset [3], which has 10,800 panoramic images with 2,295 room layout annotations. We employ BLIP-2 [18] to generate a short description for each image. + +Implementation Details. For text-conditioned generation, the training and inference schedules are kept the same as MVDiffusion [47] to make a fair comparison. For textlayout conditioned generation, we train the additional ControlNet with other parameters fixed. + +Evaluation Metrics. Following previous works, we evaluate image quality in panorama [4] and perspective [47] domain. For layout-conditioned generation, we propose a new metric to evaluate how well the generated panorama follows input layout. Specifically, we use the following metrics: + +- Panorama. We follow Text2Light [4] to report Fréchet Inception Distance (FID) [11] and Inception Score (IS) [37] on panoramas to measure realism and diversity. Additionally, CLIP Score (CS) [29] is used to evaluate the text-image consistency. While FID is widely used for image generation, it relies on an Inception network [46] trained on perspective images, thus less applicable for panorama images. Therefore, a variant of FID customized for panorama, Fréchet Auto-Encoder Distance (FAED) [26], is used to better compare the realism. +- Perspective. To simulate the real-world scenario where the user can freely navigate a panorama by viewing from different perspective views, we also report FID and IS for 20 randomly sampled views to compare with methods that generate $180^{\circ}$ vertical FOV. We also follow MVDiffusion [47] to report FID, IS and CS scores on 8 horizontally sampled views. It is worth noting that this group of metrics favors MVDiffusion by measuring its direct outputs, while ours involve interpolation for perspective views. +- Layout Consistency. We propose a layout consistency metric, which employs a layout estimation network HorizonNet [45] to estimate the room layout from the generated panorama and then compute its 2D IoU and 3D IoU [45] with the input layout condition. + +See Supplementary Sec. B for more details of the above. + +![](images/ace065b904e2da79b70c150698d52595c2f7162accc98e4577db23613d95579b.jpg) + +![](images/9c6aa4413a5ce6428bba4458d6f3de59a21235d03f2896d0f2fffedfad6475b0.jpg) + +![](images/e087e8bfc1296b7c0043d3d0fa9bf6e0df571ff89d63f0e7b89ba191cfc6720f.jpg) + +![](images/f90956f5d9c5e72ddbc2d51ebed4048f9aa9b2251863ff240383eadc6345ec31.jpg) + +![](images/633b687c007e1a1c7fd7b368033f464ceb895b7b3761fd7bfdc843689302f493.jpg) + +![](images/661a6782d8476d2a9ff693eaf74ef5079a39218adec91cdd6627b0833aeee0b9.jpg) + +![](images/1bbd72bb06e8a7461e2b28f8ea9afbe41f59e274056ced04d70d72e4b794b9fe.jpg) + +![](images/124f5afc5b44742917491feaae341ca5c2862d25df2767fb691a557c99e83082.jpg) +"A living room with a chandelier." + +![](images/934bd18a6bedfd75bbb98b8071b1afa3e82dcfe8e48899c86ae01caf02822919.jpg) + +![](images/c51d90771ea05f1c23acdec07726f51d57969b665cc0105875e5a3af11335b18.jpg) + +![](images/02f192f1c9331f0adf31e1fa81289385a76683188a686edd236b3ce53b5ca05b.jpg) + +![](images/4ae6ab66321c5989c965c17d714d555741c0d2fd7918fecb68528dd3d02eae89.jpg) + +![](images/78369b2b8cbc8431420a113b7c53f80a95993ac00e3f3c5bb27bd9cbef850005.jpg) + +![](images/e99228e06b04e5be2e579c0cff6de6915e415079096c1590aa83ba055bb4e214.jpg) + +![](images/f28a8b7af8b21980302e3cbeb2fefad2cd02af5f1a814c0c9fc9ebf9f5a7c228.jpg) + +![](images/8e013f9e1fb5c70e1bdd78571bc309db501bb50ea13d2b56a6b422bbdc390457.jpg) +"A bedroom with a bed and a table." +Figure 4. Qualitative comparisons of text-conditioned panorama generation. We show panoramas cropped to the vertical FoV of MVDiffusion [47]. Below each panorama, we show 4 evenly spaced perspective projections, with the first view crossing the left and right boundaries. We highlight the loop inconsistency, distorted lines and repetitive objects and unreasonable furniture layout of baseline methods with corresponding colors of boxes, which are addressed by our method. More results can be found in Sec. E of the supplementary. + +# 4.2. Comparisons with Previous Methods + +Baselines. We compare our PanFusion with the following baselines (see Supplementary Sec. B for details). + +- MVDiffusion [47] utilizes a multi-view diffusion model to generate 8 horizontal views that can be stitched into a panorama with vertical FOV of $90^{\circ}$ . It requires separate prompts for training while providing an option to generate + +from a single prompt. + +- Text2Light [4] generates a $180^{\circ}$ vertical FOV panorama from a text prompt in a two-stage auto-regressive manner. +- $SD + LoRA$ is our baseline model that uses LoRA [14] to finetune a Stable Diffusion [31] on panorama images. +- Pano Branch is SD+LoRA with additional modifications as described in Sec. 3.2 to ensure loop consistency. + +
MethodPanorama20 ViewsHorizontal 8 Views [47]
FAED ↓FID ↓IS ↑CS ↑FID ↓IS ↑FID ↓IS ↑CS ↑
Text2Light [4]97.2476.53.6027.4836.255.6743.664.9225.88
MVDiffusion [47]------25.276.9026.34
SD+LoRA [14, 31]7.1951.694.4028.8319.326.9020.686.4824.77
Pano Branch7.9050.404.5428.6720.107.0620.566.3724.85
PanFusion (Ours)6.0446.474.3628.5817.046.8519.886.5024.98
+ +![](images/40c16d3582734f542585fe17be6a2c1752272ae3355bb8f80de6eb3ffd352fcf.jpg) +Figure 5. Ablation study. Artifacts are highlighted with red boxes and projected to perspective views. Prompt: "A hallway in a hotel." + +Table 1. Comparison with SoTA methods. We evaluate the panorama image quality with Fréchet Auto-Encoder Distance (FAED), Fréchet Inception Distance (FID), Inception Score (IS), and CLIP Score (CS). In addition, we evaluate the perspective image quality in two settings. We first randomly sample 20 views, which is the closest to the real-world scenario where the user can freely navigate the panorama to view the scene from different perspectives. We then follow MVDiffusion [47] to horizontally sample 8 evenly spaced views. + +
MethodPanorama20 Views
FAED ↓FID ↓IS ↑CS ↑FID ↓IS ↑
Pano Branch7.9050.404.5428.6620.107.06
Ours-joint7.7559.434.3427.9429.686.82
Ours-SPE6.6349.554.6528.6620.827.00
Ours-mask6.7745.494.3528.6615.816.81
Ours-bijective7.3648.354.5828.6618.787.12
Ours6.0446.474.3628.5817.046.85
+ +Table 2. Ablation study. We compare the ablated versions of our method on both panorama and perspective domains. Here, “-” indicates removing the subsequent component. + +Quantitative Results. Tab. 1 shows the quantitative comparison results. Here, we assign the highest value to realism in image generation, measured in FAED and FID. On these two metrics, our method outperforms baseline methods in both panorama and perspective. For IS, our method's performance is slightly lower than baselines. This is likely due to the fact that IS evaluates diversity of objects in generated images, using a classifier, and our model, unlike the baselines, tends to not generate unexpected objects. Similarly, it is possible to say that baseline models present slightly higher CS is due to the repetition of objects reinforcing alignment with prompts. Considering SD+LoRA is superior to Pano Branch on FAED and is on par in other metrics, we only qualitatively compare with SD+LoRA below. + +Qualitative Results. Fig. 4 shows the qualitative comparison results. Loop inconsistency can be observed for + +Text2Light and SD+LoRA due to lack of message passing between left-right boundaries. They also suffer from distorted lines in perspective views, meaning the generated panorama does not follow the correct equirectangular projection. MVDiffusion on the other hand suffers from repetitive objects and unreasonable furniture layout, which is likely due to the lack of global context. Our method generates the most realistic scenes and aligns to text condition the best, also with less distortion in perspective views. + +# 4.3. Ablation Study + +In Sec. 4.2 and Tab. 1, we show that our full model outperforms Pano Branch, the baseline model of our method without the perspective branch. Here, as shown in Tab. 2 and Fig. 5, we further conduct an ablation study to validate the effectiveness of each component in our method. For a consistent comparison, we keep the layout similar between different ablated versions by sampling the same noise for latent map initialization, exploiting the observation of [22]. + +Joint latent map initialization. We ablate the joint latent map initialization by initializing the latent maps of the panorama and perspective branches separately (Ours-joint). Significant performance drop can be observed in all metrics and qualitative results, demonstrating the importance of joint latent map initialization. Interestingly, Ours-joint is even worse than Pano Branch in FID. It is likely due to that joint latent map initialization helps the corresponding + +![](images/b57889ee13bce4ebcd5d22f7cbaf577de7e4aab8985a68acbdd81aee1158fa63.jpg) +MVDiffusion + +![](images/8312357515e061e142b64a63e35a959d260adb3ad1e47cd2a642184bc88e89c2.jpg) +Ours +"A hallway in a building." + +![](images/d7d388b8919fd8609b7529f550620f5ba905b8872d8bec11693fae51ba7ba152.jpg) + +![](images/3191492e99170d3c2c89a76c59c093a394d500973fa4a5187bd38382149c4934.jpg) + +![](images/ab1ead31dc4bc0626532ab269220ad8d341f6f952fc1df6da75bc40229850700.jpg) + +![](images/4b34ff3731efe3f8f10ab37b0f854bd9bef7f5516236a1b9082ad9d4a0ec6627.jpg) +"A living room with a view of the ocean." + +![](images/84ba38cc3abca5f015b2e71d32686c1be76bbdd7fa3d542299f3609348748b4c.jpg) + +![](images/44435942303edbbab2d31f069f1e90d43eaaeaaaef328ee0f3b87668fe9454ff.jpg) + +
MethodLayout ConsistencyHorizontal 8 Views [47]
3D IoU ↑2D IoU ↑FID ↓IS ↑CS ↑
MVDiffusion [47]61.0664.4328.835.6026.79
PanFusion (Ours)68.4671.8222.585.1026.04
GT image74.3177.15---
+ +Table 3. Layout-conditioned comparisons. We evaluate layout consistency between the condition and the layout of each generated image extracted by a layout estimator HorizonNet [45]. "GT image" indicates the upper bound of the layout estimator. + +pixels to share a similar noise distribution from the beginning of the diffusion process, which is essential for EPPA to align the content of overlapped regions. + +EPP SPE and attention mask. We ablate spherical positional encoding (Ours-SPE) and attention mask (Ours-mask) from EPPA module. From Tab. 2, we can see that missing SPE hurts FAED and FID, which is likely due to that SPE helps the model to learn the relative position of the pixels between the two branches. Missing attention mask gets FID better, but it hurts FAED, which is more accurate in evaluating the panorama quality as it is customized for the target dataset. Both result in clear artifacts around point light sources, inconsistent textures of the floor, and distortions in the highlighted projections, as shown in Fig. 5. + +Bijective EPPA. We ablate the bijective EPPA (Ours-bijective) by using separate parameters for the two directions in the EPPA module. Both FAED and FID get worse for Ours-bijective. In addition, the ablated version struggles to generate consistent textures for the floor and ceiling for the two directions of the hallway in Fig. 5. On the contrary, our full model generates floor and ceiling with consistent style, showing a better global understanding of the scene. + +# 4.4. Application: Layout-Conditioned Generation + +To showcase the advantage of our method in generating panorama images with additional layout conditions, we build a baseline model by adding a ControlNet to MVDiffusion as described in Sec. 3.4. We render the layout condition into a distance map and then project it to perspective views to condition its generation of multi-view images. The training settings are kept the same as our PanFusion. As shown in Tab. 3, our method outperforms the baseline + +![](images/f3669e5d95b40ce62785b7bd191bfad3e9c21d937d360d78f89652501cd50204.jpg) +Figure 6. Layout-conditioned generation comparisons. We showcase how layout-conditioned panorama generation can benefit from our dual-branch structure. The input layout is drawn on the generated panorama as yellow lines, while the panorama is cropped to the FoV of MVDiffusion and projected to one view to highlight the layout consistency. +Figure 7. Failure case. Our method sometimes generates indoor scenes with no entrance. + +![](images/d5674ed2fa49f2cf3face45ffd5ec237046424fe671b0ce18c2e3244ff50dd00.jpg) + +model in layout consistency while keeping the realism advantage of perspective projections. We visualize the layout condition as wireframes overlaid on generated panorama images in Fig. 6, where we can see our generated panorama images follow their layout conditions better, especially as highlighted in perspective views. More details can be found in Sec. F of the supplementary material. + +# 5. Conclusion + +In this paper, we have proposed PanFusion, a novel text-to- $360^{\circ}$ -panorama image generation method that can generate high-quality panorama images from a single text prompt. Particularly, a dual-branch diffusion architecture has been introduced to harness prior knowledge of Stable Diffusion in the perspective domain while addressing the repetitive elements and inconsistency issues observed in previous works. An EPPA module has been further introduced to enhance the information passing between the two branches. We have also extended our PanFusion for the application of layout-conditioned panorama generation. Comprehensive experiments have demonstrated that PanFusion can generate high-quality panorama images with better realism and layout consistency than previous methods. + +Limitations. Although the dual-branch architecture of Pan-Fusion combines the advantages in both panorama and perspective domains, it comes with a cost of higher computational complexity. Additionally, our method sometimes fails to generate entrance for indoor scenes, as shown in Fig. 7, which is essential for use cases like virtual tour. + +Acknowledgement: This research is supported by Building 4.0 CRC and the National Key R&D Program of China (NO.2022ZD0160101), and was partially done at Shanghai AI Laboratory. + +# References + +[1] Naofumi Akimoto, Yuhi Matsuo, and Yoshimitsu Aoki. *Diverse plausible 360-degree image outpainting for efficient 3dgc background creation*. In *CVPR*, pages 11441–11450, 2022. 1, 2 +[2] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In ICML, pages 1737-1752. PMLR, 2023. 2, 3 +[3] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. 3DV, 2017. 5, 1, 2 +[4] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM TOG, 41(6):1-16, 2022. 1, 2, 4, 5, 6, 7 +[5] Mohammad Reza Karimi Dastjerdi, Yannick Hold-Geoffroy, Jonathan Eisenmann, Siavash Khodadadeh, and Jean-François Lalonde. Guided co-modulated gan for $360^{\circ}$ field of view extrapolation. In 3DV, pages 475–485. IEEE, 2022. 2 +[6] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, pages 8780-8794, 2021. 2 +[7] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2 +[8] Chuan Fang, Xiaotao Hu, Kunming Luo, and Ping Tan. Ctrlroom: Controllable text-to-3d room meshes generation with layout constraints. arXiv preprint arXiv:2310.03602, 2023. 2, 3, 5 +[9] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. arXiv preprint arXiv:2302.01133, 2023. 3 +[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2 +[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, pages 6626-6637, 2017. 5 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 2, 3 +[13] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In ICCV, pages 7909-7920, 2023. 2, 3, 5 +[14] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3, 6, 7, 4 +[15] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, pages 26565-26577, 2022. 2 + +[16] Yuseung Lee, Kunho Kim, Hyunjin Kim, and Minhyuk Sung. Syncdiffusion: Coherent montage via synchronized joint diffusions. In NeurIPS, 2023. 2 +[17] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. In NeurlPS, 2023. 1, 2, 3, 5 +[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 5, 1 +[19] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, pages 5775-5787, 2022. 2 +[20] Zhuqiang Lu, Kun Hu, Chaoyue Wang, Lei Bai, and Zhiyong Wang. Autoregressive omni-aware outpainting for open-vocabulary 360-degree image generation. arXiv preprint arXiv:2309.03467, 2023.1, 2 +[21] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 2 +[22] Jiafeng Mao, Xueting Wang, and Kiyoharu Aizawa. Guided image synthesis via initial image editing in diffusion model. In ACM MM, pages 5321-5329. ACM, 2023. 4, 7 +[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421. Springer, 2020. 4 +[24] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2 +[25] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, pages 16784-16804. PMLR, 2022. 2 +[26] Changgyoon Oh, Wonjune Cho, Yujeong Chae, Daehee Park, Lin Wang, and Kuk-Jin Yoon. Bips: Bi-modal indoor panorama synthesis via residual depth-aided adversarial learning. In ECCV, pages 352–371. Springer, 2022. 2, 5, 1 +[27] Chi-Han Peng and Jiayao Zhang. High-resolution depth estimation for 360deg panoramicas through perspective and panoramic depth images registration. In WACV, pages 3116-3125, 2023. 4 +[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 5, 4 + +[30] Manuel Rey-Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution 360deg monocular depth estimation. In CVPR, pages 3762-3772, 2022. 4 +[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 6, 7, 1, 4 +[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 3 +[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 2 +[34] RunwayML. Stable diffusion. https://github.com/runwayml/stable-diffusion, 2021.2,3 +[35] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2 +[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, pages 36479-36494, 2022. 2 +[37] Tim Salimans, Ian J. Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. In NeurIPS, pages 2226-2234, 2016. 5 +[38] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 2 +[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. In NeurIPS, pages 25278–25294, 2022. 1, 2 +[40] Ka Chun Shum, Hong-Wing Pang, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Conditional 360-degree image synthesis for immersive indoor scene decoration. In ICCV, pages 4478–4488, 2023. 3 +[41] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2, 3 +[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3, 1 +[43] Liangchen Song, Liangliang Cao, Hongyu Xu, Kai Kang, Feng Tang, Junsong Yuan, and Zhao Yang. Roomdreamer: Text-driven 3d indoor scene synthesis with coherent geometry and texture. In ACM MM, pages 6898-6906. ACM, 2023. 2, 4 + +[44] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, pages 11895–11907, 2019. 2, 3 +[45] Cheng Sun, Chi-Wei Hsiao, Min Sun, and Hwann-Tzong Chen. Horizonnet: Learning room layout with 1d representation and pano stretch data augmentation. In CVPR, pages 1047–1056, 2019. 5, 8 +[46] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 5 +[47] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. In NeurIPS, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[48] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, JiaBin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In CVPR, pages 16773- 16783,2023.4 +[49] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.1 +[50] Fu-En Wang, Yu-Hsuan Yeh, Min Sun, Wei-Chen Chiu, and Yi-Hsuan Tsai. Layouttmp3d: Layout annotation of matter-port3d. arXiv preprint arXiv:2003.13516, 2020. 1 +[51] Guangcong Wang, Yinuo Yang, Chen Change Loy, and Zwei Liu. Stylelight: HDR panorama generation for lighting estimation and editing. In ECCV, pages 477-492. Springer, 2022. 1, 2 +[52] Guangcong Wang, Peng Wang, Zhaoxi Chen, Wenping Wang, Chen Change Loy, and Ziwei Liu. Perf: Panoramic neural radiance field from a single panorama. arXiv preprint arXiv:2310.16831, 2023. 2 +[53] Hai Wang, Xiaoyu Xiang, Yuchen Fan, and Jing-Hao Xue. Customizing 360-degree panoramas through text-to-image diffusion models. In WACV, 2024. 2 +[54] Jionghao Wang, Ziyu Chen, Jun Ling, Rong Xie, and Li Song. 360-degree panorama generation from few unregistered nfov images. In ACM MM, pages 6811-6821. ACM, 2023. 2, 3 +[55] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 2 +[56] Tianhao Wu, Chuanxia Zheng, and Tat-Jen Cham. IPO-ldm: Depth-aided 360-degree indoor rgb panorama outpainting via latent diffusion model. arXiv preprint arXiv:2307.03177, 2023. 2, 3 +[57] Jiale Xu, Jia Zheng, Yanyu Xu, Rui Tang, and Shenghua Gao. Layout-guided novel view synthesis from a single indoor panorama. In CVPR, pages 16438-16447, 2021. 5 +[58] Mai Xu, Chen Li, Shanyi Zhang, and Patrick Le Callet. State-of-the-art in 360 video/image processing: Perception, assessment and compression. JSTSP, 14(1):5-26, 2020. 2 + +[59] Jian-Ru Xue, Jian-Wu Fang, and Pu Zhang. A survey of scene understanding by event reasoning in autonomous driving. MIR, 15(3):249-266, 2018. 1 +[60] Bangbang Yang, Yinda Zhang, Yijin Li, Zhaopeng Cui, Sean Fanello, Hujun Bao, and Guofeng Zhang. Neural rendering in a room: amodal 3d understanding and free-viewpoint rendering for the closed scene composed of pre-captured objects. ACM TOG, 41(4):1-10, 2022. 1 +[61] Bangbang Yang, Wenqi Dong, Lin Ma, Wenbo Hu, Xiao Liu, Zhaopeng Cui, and Yuewen Ma. Dreamspace: Dreaming your room space with text-driven panoramic texture propagation. arXiv preprint arXiv:2310.13119, 2023. 1, 2 +[62] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM CSUR, 2022. 2 +[63] Jason J. Yu, Fereshteh Forghani, Konstantinos G. Derpanis, and Marcus A. Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In ICCV, 2023. 2 +[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 2, 5 +[65] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming yu Liu. Diffcollage: Parallel generation of large content with diffusion models. In CVPR, 2023. 2 +[66] Chuanqing Zhuang, Zhengda Lu, Yiqun Wang, Jun Xiao, and Ying Wang. Acdnet: Adaptively combined dilated convolution for monocular panorama depth estimation. In AAAI, pages 3653-3661, 2022. 3 +[67] Chuhang Zou, Jheng-Wei Su, Chi-Han Peng, Alex Colburn, Qi Shan, Peter Wonka, Hung-Kuo Chu, and Derek Hoiem. Manhattan room layout reconstruction from a single 360 image: A comparative study of state-of-the-art methods. IJCV, 129:1410-1431, 2021. 1 \ No newline at end of file diff --git a/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/images.zip b/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6e8c0eb107859c0cd7b2300c1a2935be789e938a --- /dev/null +++ b/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56d0a67c721a244945f8123bf14e172005687839886301766cf5403ec2a93d1d +size 844545 diff --git a/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/layout.json b/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..44c9c7af1b7fa035db74d1bbf5f3cd728e2cff87 --- /dev/null +++ b/2024/Taming Stable Diffusion for Text to 360 Panorama Image Generation/layout.json @@ -0,0 +1,9506 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 81, + 103, + 512, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 103, + 512, + 121 + ], + "spans": [ + { + "bbox": [ + 81, + 103, + 512, + 121 + ], + "type": "text", + "content": "Taming Stable Diffusion for Text to " + }, + { + "bbox": [ + 81, + 103, + 512, + 121 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 81, + 103, + 512, + 121 + ], + "type": "text", + "content": " Panorama Image Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 142, + 502, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 142, + 502, + 171 + ], + "spans": [ + { + "bbox": [ + 96, + 142, + 502, + 171 + ], + "type": "text", + "content": "Cheng Zhang" + }, + { + "bbox": [ + 96, + 142, + 502, + 171 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 96, + 142, + 502, + 171 + ], + "type": "text", + "content": " Qianyi Wu" + }, + { + "bbox": [ + 96, + 142, + 502, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 142, + 502, + 171 + ], + "type": "text", + "content": " Camilo Cruz Gambardella" + }, + { + "bbox": [ + 96, + 142, + 502, + 171 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 96, + 142, + 502, + 171 + ], + "type": "text", + "content": " Xiaoshui Huang" + }, + { + "bbox": [ + 96, + 142, + 502, + 171 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 96, + 142, + 502, + 171 + ], + "type": "text", + "content": " \nDinh Phung" + }, + { + "bbox": [ + 96, + 142, + 502, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 142, + 502, + 171 + ], + "type": "text", + "content": " Wanli Ouyang" + }, + { + "bbox": [ + 96, + 142, + 502, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 142, + 502, + 171 + ], + "type": "text", + "content": " Jianfei Cai" + }, + { + "bbox": [ + 96, + 142, + 502, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 171, + 539, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 539, + 186 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 539, + 186 + ], + "type": "text", + "content": "1Monash University 2Shanghai AI Laboratory 3Building 4.0 CRC, Caulfield East, Victoria, Australia" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 47, + 217, + 298, + 415 + ], + "blocks": [ + { + "bbox": [ + 47, + 217, + 298, + 415 + ], + "lines": [ + { + "bbox": [ + 47, + 217, + 298, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 298, + 415 + ], + "type": "image", + "image_path": "8e1203debcca5709d928d249739fd1c3c9569ac0f14258ee85a50ee0a40877ee.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 425, + 546, + 469 + ], + "lines": [ + { + "bbox": [ + 45, + 425, + 546, + 469 + ], + "spans": [ + { + "bbox": [ + 45, + 425, + 546, + 469 + ], + "type": "text", + "content": "Figure 1. Our PanFusion can generate realistic and consistent " + }, + { + "bbox": [ + 45, + 425, + 546, + 469 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 45, + 425, + 546, + 469 + ], + "type": "text", + "content": " horizontal by " + }, + { + "bbox": [ + 45, + 425, + 546, + 469 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 45, + 425, + 546, + 469 + ], + "type": "text", + "content": " vertical FOV panoramas from a single text prompt, compared to the limited FOV of current state-of-the-art method MVDiffusion [47]. Left: PanFusion addresses the problem of repetitive elements (duplicated \"ceiling fans\") and inconsistency (the ceiling and wall in the center) of MVDiffusion. Right: While trained mostly on indoor scenes, PanFusion can generalize well to out-of-domain outdoor prompts with more reasonable layout." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 300, + 217, + 542, + 420 + ], + "blocks": [ + { + "bbox": [ + 300, + 217, + 542, + 420 + ], + "lines": [ + { + "bbox": [ + 300, + 217, + 542, + 420 + ], + "spans": [ + { + "bbox": [ + 300, + 217, + 542, + 420 + ], + "type": "image", + "image_path": "abb1b99aa30d27822aa4ee9ff979ff1bcbe7557aa62a3b856476ba66d5e8e7b4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 481, + 192, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 481, + 192, + 492 + ], + "spans": [ + { + "bbox": [ + 143, + 481, + 192, + 492 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 494, + 290, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 290, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 290, + 696 + ], + "type": "text", + "content": "Generative models, e.g., Stable Diffusion, have enabled the creation of photorealistic images from text prompts. Yet, the generation of 360-degree panorama images from text remains a challenge, particularly due to the dearth of paired text-panorama data and the domain gap between panorama and perspective images. In this paper, we introduce a novel dual-branch diffusion model named PanFusion to generate a 360-degree image from a text prompt. We leverage the stable diffusion model as one branch to provide prior knowledge in natural image generation and register it to another panorama branch for holistic image generation. We propose a unique cross-attention mechanism with projection awareness to minimize distortion during the collaborative denoising process. Our experiments validate that PanFusion surpasses existing methods and, thanks to its dual-branch structure, can integrate additional constraints like room layout for customized panorama outputs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 480, + 386, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 480, + 386, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 480, + 386, + 493 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 506, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 614 + ], + "type": "text", + "content": "Creating a " + }, + { + "bbox": [ + 304, + 506, + 545, + 614 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 506, + 545, + 614 + ], + "type": "text", + "content": " panorama image from textual prompts is a nascent yet pivotal frontier in computer vision, with profound implications for applications that require extensive environmental representation, such as environmental lighting [1, 4, 51], VR/AR [60, 61], autonomous driving [59], and visual navigation [17]. Despite recent significant strides in text-to-image synthesis, the leap to generating full " + }, + { + "bbox": [ + 304, + 506, + 545, + 614 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 506, + 545, + 614 + ], + "type": "text", + "content": " horizontal by " + }, + { + "bbox": [ + 304, + 506, + 545, + 614 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 304, + 506, + 545, + 614 + ], + "type": "text", + "content": " vertical field-of-view (FOV) panorama remains challenging." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 618, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 546, + 715 + ], + "type": "text", + "content": "There are two major hurdles for achieving this goal. The first hurdle is data scarcity. The availability of text-to-panorama image pairs [20, 47] is significantly less compared with the abundance of text-to-common image pairs [38, 39]. The dearth of data complicates the training and finetuning of generative models. The second hurdle lies in the geometric and domain variations. Panorama images are distinct not only in their aspect ratio " + }, + { + "bbox": [ + 304, + 618, + 546, + 715 + ], + "type": "inline_equation", + "content": "(2:1)" + }, + { + "bbox": [ + 304, + 618, + 546, + 715 + ], + "type": "text", + "content": " but" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 704, + 135, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 704, + 135, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 704, + 135, + 714 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6347" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "also in the underlying equirectangular projection (ERP) geometry [58]. This differs significantly from typical square images of the perspective projection that are used in most generative model training [38, 39]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 288, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 288, + 338 + ], + "type": "text", + "content": "To mitigate the scarcity of panorama-specific training data, the previous solutions follow a common principle that leverages the prior knowledge of the pre-trained generative model [17, 20, 47]. However, taming powerful models like stable diffusion [31, 34] to generate high-fidelity panorama images remains a non-trivial task. Early attempts turn to formulate the 360-degree generation as an iterative image inpainting or warping process [17, 20]. Such solutions suffer from error accumulation and fail to handle loop closure [47]. To address this, MVDiffusion [47] proposes to produce multiple perspective images simultaneously by introducing a correspondence-aware attention module to facilitate multiview consistency, and then stitch together the perspective images to form a complete panorama. Despite the improved performance, the pixel-level consistency between neighboring perspectives in MVDiffusion cannot ensure global consistency, often resulting in repetitive elements or semantic inconsistency, as illustrated in Fig. 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 340, + 288, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 288, + 639 + ], + "type": "text", + "content": "Therefore, in this paper, we propose a novel dual-branch diffusion model called PanFusion that is tailored to address the limitations of prior models for high-quality text to " + }, + { + "bbox": [ + 46, + 340, + 288, + 639 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 340, + 288, + 639 + ], + "type": "text", + "content": " panorama image generation. Specifically, PanFusion is designed to operate in both panorama and perspective domains, employing a global branch for creating a coherent panoramic \"canvas\" and a local branch that focuses on rendering detail-rich multiview perspectives. The local-global synergy of PanFusion significantly improves the resulting panoramas against the prevalent issues of error propagation and visual inconsistency that prior models have struggled with. To enhance the synergy between the two branches, we further propose an Equirectangular-Perspective Projection Attention (EPPA) mechanism, which respects the equirectangular projection for maintaining geometric integrity throughout the generation process. In addition, our adoption of parametric mapping for positional encoding is another leap forward, enhancing the model's spatial awareness and further ensuring the consistency of the generated panorama. Moreover, the panorama branch of PanFusion can be easily leveraged to accommodate supplementary control inputs at the panorama level, such as room layout, allowing for the creation of images that adhere to precise spatial conditions. We summarize our primary contributions as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 642, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 642, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 286, + 689 + ], + "type": "text", + "content": "- We pioneer a dual-branch diffusion model PanFusion, harnessing both the global panorama and local perspective latent domains, to generate high-quality, consistent " + }, + { + "bbox": [ + 47, + 642, + 286, + 689 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 47, + 642, + 286, + 689 + ], + "type": "text", + "content": " pancreas from text prompts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 712 + ], + "type": "text", + "content": "- To enhance the interaction between the two branches, we introduce an \"equirectangular-perspective projection at" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 312, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 545, + 120 + ], + "type": "text", + "content": "tention” mechanism that establishes a novel correspondence between global panorama and local perspective branches, addressing the unique projection challenges of panorama synthesis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 121, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 121, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 121, + 545, + 180 + ], + "type": "text", + "content": "- Our PanFusion not only surpasses prior models in quality and consistency but also supports extended control over the generation process with the inclusion of room layout. Extensive experimental results demonstrate the superiority of our proposed framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 190, + 392, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 190, + 392, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 392, + 204 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 211, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 211, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 305, + 211, + 545, + 390 + ], + "type": "text", + "content": "Diffusion models. In recent years, diffusion models [6, 12, 41, 42, 44, 62] have taken the world of image generation by storm, as they have become faster [15, 19, 42] and more capable in terms of image quality and resolution [28, 31, 36]. This success has prompted the development of various applications for diffusion models, such as text-to-image [25, 28, 31, 36], image-conditioned generation [24, 64], in-painting [21, 35] and subject-driven generation [10, 33]. Most of these applications try to exploit the prior knowledge of a pre-trained diffusion model to mitigate the scarcity of task-specific data, by either finetuning with techniques like LoRA [14], or introducing auxiliary modules to distill the knowledge. We also adopt the same principle to harness the power of a pre-trained latent diffusion model [31] for panorama image generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 392, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 392, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 545, + 713 + ], + "type": "text", + "content": "Panorama generation. Panorama image generation encompasses different settings, including panorama outpainting and text-to-panorama generation. Panorama outpainting [1, 5, 26, 51, 54, 56] focus on generating a 360-degree panorama from a partial input image. Various methods, such as StyleLight [51] and BIPS [26], have addressed specific use cases, focusing on HDR environment lighting and robotic guidance scenarios. Recent works [54, 56] have improved realism with diffusion models, but often lack the exploitation of rich prior information from pre-trained models, limiting generalization. On the other hand, recent developments in generative models have opened new frontiers in synthesizing immersive visual content from textual input [8, 13, 43, 52, 53, 55, 56, 61, 63]. As an image-based representation, generating panorama from text has gained much attention. Text2Light [4] adopt the VQGAN [7] structure to synthesize an HDR panorama image from text. To generate in arbitrary resolution with pre-trained diffusion models, DiffCollage [65], MultiDiffusion [2] and SyncDiffusion [16] propose to fuse the diffusion paths, while PanoGen [17] solves by iteratively inpainting. However, they failed to model the equirectangular projection of 360-degree panoramas. Lu et al. [20] adopts an autoregressive framework, but suffers from inefficient issues. MVDiffusion [47] designs a correspondence-aware attention module to produce multi-view images simultaneously that can be stitched together but results in repetitive elements and in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6348" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 63, + 536, + 236 + ], + "blocks": [ + { + "bbox": [ + 55, + 63, + 536, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 63, + 536, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 63, + 536, + 236 + ], + "type": "image", + "image_path": "ed3a739337d81ae64bd181e429d79a0db48cff0fc1f855a18fdb90f214bf4498.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 547, + 295 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 547, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 547, + 295 + ], + "type": "text", + "content": "Figure 2. Our proposed dual-branch PanFusion pipeline. The panorama branch (upper) provides global layout guidance and registers the perspective information to get seamless panorama output. The perspective branch (lower) harnesses the rich prior knowledge of Stable Diffusion (SD) and provides guidance to alleviate distortion under perspective projection. Both branches employ the same UNet backbone with shared weights, while finetuned with separate LoRA layers. Equirectangular-Perspective Projection Attention (EPPA) modules are plugged into different layers of the UNet to pass information between the two branches." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 305, + 288, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 288, + 365 + ], + "type": "text", + "content": "consistency. In contrast, our proposed PanFusion, a dual-branch framework, addresses the limitations of existing methods by considering both global panorama views and local perspective views, providing a comprehensive solution for text-driven 360-degree panorama image generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 374, + 104, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 374, + 104, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 104, + 386 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 394, + 126, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 126, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 126, + 406 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 412, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 533 + ], + "type": "text", + "content": "First introduced in [41], Diffusion models [12, 42, 44] aim to generate images from a noise distribution by iterative denoising with a learned prior distribution. However, the early diffusion models operate in the image space, which is of high dimension and complex. We have recently witnessed the huge success of latent diffusion models [31] that first transform an image " + }, + { + "bbox": [ + 46, + 412, + 287, + 533 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 412, + 287, + 533 + ], + "type": "text", + "content": " to a latent representation " + }, + { + "bbox": [ + 46, + 412, + 287, + 533 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 412, + 287, + 533 + ], + "type": "text", + "content": " with a learned encoder " + }, + { + "bbox": [ + 46, + 412, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 412, + 287, + 533 + ], + "type": "text", + "content": " and then train a UNet [32] model " + }, + { + "bbox": [ + 46, + 412, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 412, + 287, + 533 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 46, + 412, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 412, + 287, + 533 + ], + "type": "text", + "content": " for the reverse process in the latent space, formulating the training objective as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 538, + 287, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 538, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 85, + 538, + 287, + 552 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {\\mathcal {E} (x), t, \\epsilon , y} [ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, \\tau (y)) | | _ {2} ], \\tag {1}", + "image_path": "6055d8de7a254810fdf58f33c852cbd1af8d6a1e89ad005030ff0ad1d956d64c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "inline_equation", + "content": "\\tau(y)" + }, + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "text", + "content": " is an encoding of input condition " + }, + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "text", + "content": " (e.g., text prompt or image), " + }, + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "text", + "content": " is the latent map at time step " + }, + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "text", + "content": " is sampled from a Gaussian noise. To sample from the model, " + }, + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "text", + "content": " is first initialized from standard Gaussian distribution, then the reverse process is applied iteratively to generate " + }, + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "text", + "content": ", which is finally decoded into image space with a decoder " + }, + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 557, + 288, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 647, + 209, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 209, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 209, + 658 + ], + "type": "text", + "content": "3.2. Dual-Branch Diffusion Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Directly employing pre-trained latent diffusion models [31], e.g., Stable Diffusion (SD) [34], to generate panorama from multiple perspective images, either in an iterative manner [9, 13, 17] or in a synchronized way [2, 47], would" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 304, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 484 + ], + "type": "text", + "content": "fail to handle loop closure [47] or produce repetitive elements (Fig. 1) due to lack of global understanding. To address this issue, we propose a dual-branch diffusion model, which consists of a panorama branch and a perspective branch both based on the UNet of SD as shown in Fig. 2. The panorama branch is designed to provide global layout guidance and to register the perspective information to get the final panorama without stitching, while the perspective branch is designed to exploit the rich perspective image-generation capabilities of SD and to provide guidance to alleviate distortion under perspective projection. The two branches work together during the diffusion process to generate a denoised panorama latent map. Finally, this latent map runs through the pre-trained decoder " + }, + { + "bbox": [ + 304, + 304, + 545, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 304, + 545, + 484 + ], + "type": "text", + "content": " of SD to produce the final panorama image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "type": "text", + "content": "Panorama Branch. Given a text prompt " + }, + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "type": "text", + "content": ", our goal is to generate a panorama " + }, + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "type": "inline_equation", + "content": "W = 2H" + }, + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "type": "inline_equation", + "content": "H = 512" + }, + { + "bbox": [ + 304, + 486, + 546, + 714 + ], + "type": "text", + "content": ". To account for the difference between the target resolution and the one that SD is trained on, we introduce LoRA [14] layers to adapt the model to the new resolution. SD with LoRA can already serve as a strong baseline for panorama generation, but the results are not loop-consistent. Previous works [8, 56] have attempted to overcome this problem using 90 degree rotations of the latent map at each diffusion step. However, the seams are still obvious [56]. On close inspection of the SD model we have observed that the loop inconsistency is mainly caused by the convolutional layers in the UNet backbone, due to the lack of a mechanism to pass information between the two ends of the panorama image. Therefore, we introduce an adaptation to the UNet by adding additional circular padding [40, 54, 66] before each convolutional layer, and then cropping the output feature maps to the original size. In addition, we also add circular padding to the latent map before decoding to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6349" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "mitigate the less apparent loop inconsistency caused by the decoder. The combination of the techniques outlined above - latent rotation and circular padding - enable the generation of loop-consistent results with negligible computational cost, and can thus serve as another strong baseline. However, these alone do not make full use of the perspective generation capabilities that SD possess." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 288, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 288, + 300 + ], + "type": "text", + "content": "Perspective Branch. The perspective branch aims to exploit the outstanding capabilities that SD has shown in the generation perspective images. Since it does not generate panoramas directly, it can operate at a lower resolution. Specifically, we set its resolution to " + }, + { + "bbox": [ + 46, + 157, + 288, + 300 + ], + "type": "inline_equation", + "content": "H / 2 \\times H / 2" + }, + { + "bbox": [ + 46, + 157, + 288, + 300 + ], + "type": "text", + "content": " and add LoRA layers to adapt to the new resolution. To evenly distribute perspective cameras and fully cover the panorama image, we sample " + }, + { + "bbox": [ + 46, + 157, + 288, + 300 + ], + "type": "inline_equation", + "content": "N = 20" + }, + { + "bbox": [ + 46, + 157, + 288, + 300 + ], + "type": "text", + "content": " cameras with poses " + }, + { + "bbox": [ + 46, + 157, + 288, + 300 + ], + "type": "inline_equation", + "content": "R^i \\in SO(3), i \\in 1, \\dots, N" + }, + { + "bbox": [ + 46, + 157, + 288, + 300 + ], + "type": "text", + "content": " on an icosahedron, similar to [27, 30], and set the FOV " + }, + { + "bbox": [ + 46, + 157, + 288, + 300 + ], + "type": "inline_equation", + "content": "= 90^\\circ" + }, + { + "bbox": [ + 46, + 157, + 288, + 300 + ], + "type": "text", + "content": ". We input the same prompt " + }, + { + "bbox": [ + 46, + 157, + 288, + 300 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 157, + 288, + 300 + ], + "type": "text", + "content": " to both branches, and leave it to the model to decide how to exploit the prompt." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 301, + 288, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 288, + 517 + ], + "type": "text", + "content": "Joint Latent Map Initialization. Previous work [22] has shown that the latent map initialization can affect the layout of the generated image, i.e., the initial noise can be modified to manipulate the layout of the generated image. We find this is particularly important for multi-image generation with correspondence, as the overlapping regions in different views tend to generate different elements if noise is sampled independently. Under this observation, we propose to jointly sample noise for the panorama and perspective latent maps by projecting noise from the panorama to the perspective. To be more specific, we first initialize panorama latent map " + }, + { + "bbox": [ + 46, + 301, + 288, + 517 + ], + "type": "inline_equation", + "content": "z_{T}^{*} \\in \\mathbb{R}^{4 \\times H / f \\times W / f}" + }, + { + "bbox": [ + 46, + 301, + 288, + 517 + ], + "type": "text", + "content": " as Gaussian noise, then initialize perspective latent maps with " + }, + { + "bbox": [ + 46, + 301, + 288, + 517 + ], + "type": "inline_equation", + "content": "z_{T}^{i} = \\mathrm{P}(z_{T}^{*}, R^{i}, \\mathrm{FOV}, (H / 2f, H / 2f)), z_{T}^{i} \\in \\mathbb{R}^{4 \\times H / 2f \\times H / 2f}" + }, + { + "bbox": [ + 46, + 301, + 288, + 517 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 301, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\mathrm{P}(\\cdot)" + }, + { + "bbox": [ + 46, + 301, + 288, + 517 + ], + "type": "text", + "content": " is the function that projects " + }, + { + "bbox": [ + 46, + 301, + 288, + 517 + ], + "type": "inline_equation", + "content": "z_{T}^{*}" + }, + { + "bbox": [ + 46, + 301, + 288, + 517 + ], + "type": "text", + "content": " to a perspective view, and " + }, + { + "bbox": [ + 46, + 301, + 288, + 517 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 301, + 288, + 517 + ], + "type": "text", + "content": " is the down-sampling factor the encoder " + }, + { + "bbox": [ + 46, + 301, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 301, + 288, + 517 + ], + "type": "text", + "content": " used in SD. We use nearest interpolation for projection as it introduces fewer artifacts than bilinear interpolation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 526, + 138, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 138, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 138, + 538 + ], + "type": "text", + "content": "3.3. EPP Attention" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "content": "To pass the guidance between the perspective and panorama branches, a naive solution is to project the feature maps in different layers from one branch to the other, similar to [47]. However, this will introduce information loss during interpolation and will restrict the receptive field to a small region around corresponding pixels. Instead, we propose an Equirectangular-Perspective Projection Attention (EPPA) module to implicitly pass the guidance in a cross-attention way as shown in Fig. 2. The EPPA operates on feature maps in different layers of two UNet branches, denoted as " + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "inline_equation", + "content": "F^{*} \\in \\mathbb{R}^{c \\times h \\times w}" + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "inline_equation", + "content": "F^{i} \\in \\mathbb{R}^{c \\times h / 2 \\times h / 2}" + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "content": " is the channel dimension, " + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "content": " are the height and width of the feature maps, respectively. In addition, while cross-attention bypasses interpolation and provides benefits from" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 308, + 70, + 542, + 225 + ], + "blocks": [ + { + "bbox": [ + 308, + 70, + 542, + 225 + ], + "lines": [ + { + "bbox": [ + 308, + 70, + 542, + 225 + ], + "spans": [ + { + "bbox": [ + 308, + 70, + 542, + 225 + ], + "type": "image", + "image_path": "50449cba423ab2c448e7b4f1528b67d19e45599cbe2fb2d383ac966358af1826.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 231, + 545, + 275 + ], + "lines": [ + { + "bbox": [ + 305, + 231, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 305, + 231, + 545, + 275 + ], + "type": "text", + "content": "Figure 3. Equirectangular-Perspective Projection (EPP) Attention. As EPP attention module is designed to be bijective to pass information in both directions, we only illustrated the direction of registering perspective information to panorama." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 286, + 545, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 545, + 345 + ], + "type": "text", + "content": "the global receptive field of the attention mechanism, it is unaware of the projection between different formats. To address this issue, as shown in Fig. 3, we introduce two key components: EPP spherical positional encoding and EPP attention mask." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 346, + 545, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 455 + ], + "type": "text", + "content": "EPP Spherical Positional Encoding. Since the EPPA module tries to associate two different formats, it is important to add positional information in the same space so that the attention mechanism can learn the correspondence. We achieve this by introducing Spherical Positional Encoding (SPE) [4, 23] to the EPPA module. The " + }, + { + "bbox": [ + 304, + 346, + 545, + 455 + ], + "type": "inline_equation", + "content": "\\mathrm{SPE}(\\theta ,\\phi) = (\\gamma (\\theta),\\gamma (\\phi))" + }, + { + "bbox": [ + 304, + 346, + 545, + 455 + ], + "type": "text", + "content": " function maps polar coordinates " + }, + { + "bbox": [ + 304, + 346, + 545, + 455 + ], + "type": "inline_equation", + "content": "(\\theta ,\\phi)" + }, + { + "bbox": [ + 304, + 346, + 545, + 455 + ], + "type": "text", + "content": " to a higher dimensional space " + }, + { + "bbox": [ + 304, + 346, + 545, + 455 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{4L}" + }, + { + "bbox": [ + 304, + 346, + 545, + 455 + ], + "type": "text", + "content": " with Fourier positional encoding:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 309, + 459, + 545, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 459, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 309, + 459, + 545, + 484 + ], + "type": "interline_equation", + "content": "\\gamma (\\theta) = \\left[ \\sin \\left(2 ^ {0} \\pi \\theta\\right), \\cos \\left(2 ^ {0} \\pi \\theta\\right), \\dots , \\sin \\left(2 ^ {L - 1} \\pi \\theta\\right), \\cos \\left(2 ^ {L - 1} \\pi \\theta\\right) \\right]. \\tag {2}", + "image_path": "cb5badc1bf489a5b9dede8d320073786fee329de641d67836bd61a012812b915.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": "Here we set " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "L = c / 4" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": " so that " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{SPE}(\\theta, \\phi) \\in \\mathbb{R}^c" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": ". In the EPPA module, we apply SPE by first computing an SPE map for the panorama feature map, then project it to each perspective feature map with projection function " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{P}(\\cdot)" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": ", so that corresponding pixels in different formats share the same SPE vector. Finally, the SPE maps are added to the feature maps accordingly and go through a linear layer to get the query " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": " and key " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": ", following a matrix product to get the affinity matrix " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": ". Take the perspective-to-panorama direction as an example as shown in Fig. 3, where " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{c \\times h \\times w}" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{N \\times c \\times h / 2 \\times h / 2}" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": " are reshaped and multiplied to get the affinity matrix " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{hw \\times Nh^2 / 4}" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "EPP Attention Mask. In addition to the SPE, we also propose an EPP attention mask to encourage the attention mechanism to focus around the corresponding pixels, inspired by [48]. For example, for a target pixel in the panorama feature map as in Fig. 3, we would like to focus on registering the information from corresponding source pixels in the perspective feature maps. We achieve" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6350" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": "this by enhancing the affinity matrix " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": " with a soft mask " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{hw \\times Nh^2 / 4}" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": " highlighting the corresponding pixels. To generate " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": ", we first use " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "\\mathrm{P}(\\cdot)" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": " to project a binary mask " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "M_{j,k}" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": " for each pixel " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "(j, k)" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": " in panorama feature map to each perspective view " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": ". Then we apply a Gaussian kernel to smooth the masks and normalized them to " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": ". The masks are stacked and reshaped into " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": ", which is then added to " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": " to get the enhanced affinity matrix " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "A'" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": ". The rest goes the same as the vanilla attention mechanism, where a softmax function is applied to " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "A'" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": " to get the attention weights, which are finally multiplied with the value " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": " to get the output." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 289, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 348 + ], + "type": "text", + "content": "Inspired by [47, 64], we add zero-initialized " + }, + { + "bbox": [ + 46, + 205, + 289, + 348 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 205, + 289, + 348 + ], + "type": "text", + "content": " convolutional layers to the output of cross-attention and add it as a residual to the target feature map. This ensures the UNet stays unmodified at the beginning of training and can be gradually adapted to the EPPA modules. We add independent EPPA modules after each down-sampling layer and before each up-sampling layer of UNet to connect two branches, detailed in Sec. A of the supplementary material. Considering that the guidance is passed in both directions following the same equirectangular-perspective projection rule, which is bijective in nature, we share the weights of the EPPA modules in the two directions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 355, + 221, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 355, + 221, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 355, + 221, + 367 + ], + "type": "text", + "content": "3.4. Layout-Conditioned Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 373, + 289, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 289, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 289, + 577 + ], + "type": "text", + "content": "One of the most important applications for panorama generation is to generate panorama according to a given room layout. This is particularly useful for panorama novel view synthesis [57] and can potentially benefit indoor 3D scene generation [8, 13]. However, this problem is not well researched for diffusion-based panorama generation, largely due to the difficulty of introducing layout constraints while exploiting the rich prior knowledge of SD in perspective format at the same time. For panorama generation from multi-view [17, 47], one naive solution is to project layout condition into different views to locally condition the generation of perspective images. Instead, for our dual-branch diffusion model, we can naturally exploit the global nature of the panorama branch to enforce a much stronger layout constraint. Specifically, we render the layout condition as a distance map, then use it as the input of a ControlNet [64] to condition the panorama branch." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 583, + 111, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 583, + 111, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 583, + 111, + 596 + ], + "type": "text", + "content": "3.5. Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 601, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 662 + ], + "type": "text", + "content": "During training, we employ the same technique for latent map initialization to jointly sample noise for the panorama and perspective view, which we denote as " + }, + { + "bbox": [ + 46, + 601, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\epsilon^{*}" + }, + { + "bbox": [ + 46, + 601, + 287, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 601, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\epsilon^i" + }, + { + "bbox": [ + 46, + 601, + 287, + 662 + ], + "type": "text", + "content": ", respectively. Given a GT panorama " + }, + { + "bbox": [ + 46, + 601, + 287, + 662 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 46, + 601, + 287, + 662 + ], + "type": "text", + "content": ", we apply supervision on the panorama branch with the loss in Eq. (1):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 669, + 287, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 669, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 76, + 669, + 287, + 683 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {*} = \\mathbb {E} _ {\\mathcal {E} \\left(x ^ {*}\\right), t, \\epsilon^ {*}, y} \\left[ \\| \\epsilon^ {*} - \\epsilon_ {\\theta} ^ {*} \\left(z _ {t} ^ {*}, t, \\tau (y)\\right) \\| _ {2} \\right], \\tag {3}", + "image_path": "c50512ced72760baf9e17af822ea2518804125ddd6b305b498295be577016a2a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{*}" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": " is the predicted noise from the panorama branch. To encourage the synchronization between the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": "two branches, we also apply supervision on the perspective branch with the following loss:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 336, + 106, + 545, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 106, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 336, + 106, + 545, + 121 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {i} = \\mathbb {E} _ {\\mathcal {E} (x ^ {i}), t, \\epsilon^ {i}, y} \\left[ | | \\epsilon^ {i} - \\epsilon_ {\\theta} ^ {i} \\left(z _ {t} ^ {i}, t, \\tau (y)\\right) | | _ {2} \\right], \\tag {4}", + "image_path": "6a61f4e611d689c9a318a6b74ba7becee355acc50f14474b1f9310cd6dcedad6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 130, + 547, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 130, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 130, + 547, + 194 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 130, + 547, + 194 + ], + "type": "inline_equation", + "content": "x^{i}" + }, + { + "bbox": [ + 304, + 130, + 547, + 194 + ], + "type": "text", + "content": " is a perspective image projected from " + }, + { + "bbox": [ + 304, + 130, + 547, + 194 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 304, + 130, + 547, + 194 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 130, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{i}" + }, + { + "bbox": [ + 304, + 130, + 547, + 194 + ], + "type": "text", + "content": " is the predicted noise from the perspective branch. We joint train the EPPA modules and the LoRA layers in the two branches by combining the above two losses as " + }, + { + "bbox": [ + 304, + 130, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}^{*} + \\frac{1}{N}\\sum_{i = 1}^{N}\\mathcal{L}^{i}" + }, + { + "bbox": [ + 304, + 130, + 547, + 194 + ], + "type": "text", + "content": ". Note that the SD UNet blocks remain frozen." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 204, + 382, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 204, + 382, + 217 + ], + "spans": [ + { + "bbox": [ + 305, + 204, + 382, + 217 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 224, + 422, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 224, + 422, + 237 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 422, + 237 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 243, + 545, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 292 + ], + "type": "text", + "content": "Dataset. We follow the MVDiffusion [47] to use the Matterport3D dataset [3], which has 10,800 panoramic images with 2,295 room layout annotations. We employ BLIP-2 [18] to generate a short description for each image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 293, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 352 + ], + "type": "text", + "content": "Implementation Details. For text-conditioned generation, the training and inference schedules are kept the same as MVDiffusion [47] to make a fair comparison. For textlayout conditioned generation, we train the additional ControlNet with other parameters fixed." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 354, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 413 + ], + "type": "text", + "content": "Evaluation Metrics. Following previous works, we evaluate image quality in panorama [4] and perspective [47] domain. For layout-conditioned generation, we propose a new metric to evaluate how well the generated panorama follows input layout. Specifically, we use the following metrics:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 414, + 545, + 700 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 304, + 414, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 533 + ], + "type": "text", + "content": "- Panorama. We follow Text2Light [4] to report Fréchet Inception Distance (FID) [11] and Inception Score (IS) [37] on panoramas to measure realism and diversity. Additionally, CLIP Score (CS) [29] is used to evaluate the text-image consistency. While FID is widely used for image generation, it relies on an Inception network [46] trained on perspective images, thus less applicable for panorama images. Therefore, a variant of FID customized for panorama, Fréchet Auto-Encoder Distance (FAED) [26], is used to better compare the realism." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "content": "- Perspective. To simulate the real-world scenario where the user can freely navigate a panorama by viewing from different perspective views, we also report FID and IS for 20 randomly sampled views to compare with methods that generate " + }, + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "content": " vertical FOV. We also follow MVDiffusion [47] to report FID, IS and CS scores on 8 horizontally sampled views. It is worth noting that this group of metrics favors MVDiffusion by measuring its direct outputs, while ours involve interpolation for perspective views." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 641, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 700 + ], + "type": "text", + "content": "- Layout Consistency. We propose a layout consistency metric, which employs a layout estimation network HorizonNet [45] to estimate the room layout from the generated panorama and then compute its 2D IoU and 3D IoU [45] with the input layout condition." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 701, + 534, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 534, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 534, + 713 + ], + "type": "text", + "content": "See Supplementary Sec. B for more details of the above." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6351" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 62, + 302, + 123 + ], + "blocks": [ + { + "bbox": [ + 62, + 62, + 302, + 123 + ], + "lines": [ + { + "bbox": [ + 62, + 62, + 302, + 123 + ], + "spans": [ + { + "bbox": [ + 62, + 62, + 302, + 123 + ], + "type": "image", + "image_path": "ace065b904e2da79b70c150698d52595c2f7162accc98e4577db23613d95579b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 123, + 302, + 183 + ], + "blocks": [ + { + "bbox": [ + 62, + 123, + 302, + 183 + ], + "lines": [ + { + "bbox": [ + 62, + 123, + 302, + 183 + ], + "spans": [ + { + "bbox": [ + 62, + 123, + 302, + 183 + ], + "type": "image", + "image_path": "9c6aa4413a5ce6428bba4458d6f3de59a21235d03f2896d0f2fffedfad6475b0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 61, + 183, + 302, + 246 + ], + "blocks": [ + { + "bbox": [ + 61, + 183, + 302, + 246 + ], + "lines": [ + { + "bbox": [ + 61, + 183, + 302, + 246 + ], + "spans": [ + { + "bbox": [ + 61, + 183, + 302, + 246 + ], + "type": "image", + "image_path": "e087e8bfc1296b7c0043d3d0fa9bf6e0df571ff89d63f0e7b89ba191cfc6720f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 62, + 246, + 302, + 304 + ], + "blocks": [ + { + "bbox": [ + 62, + 246, + 302, + 304 + ], + "lines": [ + { + "bbox": [ + 62, + 246, + 302, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 246, + 302, + 304 + ], + "type": "image", + "image_path": "f90956f5d9c5e72ddbc2d51ebed4048f9aa9b2251863ff240383eadc6345ec31.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 62, + 305, + 302, + 366 + ], + "blocks": [ + { + "bbox": [ + 62, + 305, + 302, + 366 + ], + "lines": [ + { + "bbox": [ + 62, + 305, + 302, + 366 + ], + "spans": [ + { + "bbox": [ + 62, + 305, + 302, + 366 + ], + "type": "image", + "image_path": "633b687c007e1a1c7fd7b368033f464ceb895b7b3761fd7bfdc843689302f493.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 62, + 366, + 302, + 426 + ], + "blocks": [ + { + "bbox": [ + 62, + 366, + 302, + 426 + ], + "lines": [ + { + "bbox": [ + 62, + 366, + 302, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 366, + 302, + 426 + ], + "type": "image", + "image_path": "661a6782d8476d2a9ff693eaf74ef5079a39218adec91cdd6627b0833aeee0b9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 62, + 426, + 302, + 487 + ], + "blocks": [ + { + "bbox": [ + 62, + 426, + 302, + 487 + ], + "lines": [ + { + "bbox": [ + 62, + 426, + 302, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 302, + 487 + ], + "type": "image", + "image_path": "1bbd72bb06e8a7461e2b28f8ea9afbe41f59e274056ced04d70d72e4b794b9fe.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 62, + 487, + 302, + 548 + ], + "blocks": [ + { + "bbox": [ + 62, + 487, + 302, + 548 + ], + "lines": [ + { + "bbox": [ + 62, + 487, + 302, + 548 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 302, + 548 + ], + "type": "image", + "image_path": "124f5afc5b44742917491feaae341ca5c2862d25df2767fb691a557c99e83082.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 550, + 241, + 560 + ], + "lines": [ + { + "bbox": [ + 123, + 550, + 241, + 560 + ], + "spans": [ + { + "bbox": [ + 123, + 550, + 241, + 560 + ], + "type": "text", + "content": "\"A living room with a chandelier.\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 304, + 62, + 543, + 123 + ], + "blocks": [ + { + "bbox": [ + 304, + 62, + 543, + 123 + ], + "lines": [ + { + "bbox": [ + 304, + 62, + 543, + 123 + ], + "spans": [ + { + "bbox": [ + 304, + 62, + 543, + 123 + ], + "type": "image", + "image_path": "934bd18a6bedfd75bbb98b8071b1afa3e82dcfe8e48899c86ae01caf02822919.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 304, + 123, + 543, + 184 + ], + "blocks": [ + { + "bbox": [ + 304, + 123, + 543, + 184 + ], + "lines": [ + { + "bbox": [ + 304, + 123, + 543, + 184 + ], + "spans": [ + { + "bbox": [ + 304, + 123, + 543, + 184 + ], + "type": "image", + "image_path": "c51d90771ea05f1c23acdec07726f51d57969b665cc0105875e5a3af11335b18.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 304, + 184, + 543, + 246 + ], + "blocks": [ + { + "bbox": [ + 304, + 184, + 543, + 246 + ], + "lines": [ + { + "bbox": [ + 304, + 184, + 543, + 246 + ], + "spans": [ + { + "bbox": [ + 304, + 184, + 543, + 246 + ], + "type": "image", + "image_path": "02f192f1c9331f0adf31e1fa81289385a76683188a686edd236b3ce53b5ca05b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 304, + 246, + 543, + 304 + ], + "blocks": [ + { + "bbox": [ + 304, + 246, + 543, + 304 + ], + "lines": [ + { + "bbox": [ + 304, + 246, + 543, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 543, + 304 + ], + "type": "image", + "image_path": "4ae6ab66321c5989c965c17d714d555741c0d2fd7918fecb68528dd3d02eae89.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 304, + 305, + 543, + 366 + ], + "blocks": [ + { + "bbox": [ + 304, + 305, + 543, + 366 + ], + "lines": [ + { + "bbox": [ + 304, + 305, + 543, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 543, + 366 + ], + "type": "image", + "image_path": "78369b2b8cbc8431420a113b7c53f80a95993ac00e3f3c5bb27bd9cbef850005.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 304, + 366, + 543, + 426 + ], + "blocks": [ + { + "bbox": [ + 304, + 366, + 543, + 426 + ], + "lines": [ + { + "bbox": [ + 304, + 366, + 543, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 543, + 426 + ], + "type": "image", + "image_path": "e99228e06b04e5be2e579c0cff6de6915e415079096c1590aa83ba055bb4e214.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 304, + 426, + 543, + 487 + ], + "blocks": [ + { + "bbox": [ + 304, + 426, + 543, + 487 + ], + "lines": [ + { + "bbox": [ + 304, + 426, + 543, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 543, + 487 + ], + "type": "image", + "image_path": "f28a8b7af8b21980302e3cbeb2fefad2cd02af5f1a814c0c9fc9ebf9f5a7c228.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 304, + 487, + 543, + 548 + ], + "blocks": [ + { + "bbox": [ + 304, + 487, + 543, + 548 + ], + "lines": [ + { + "bbox": [ + 304, + 487, + 543, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 543, + 548 + ], + "type": "image", + "image_path": "8e013f9e1fb5c70e1bdd78571bc309db501bb50ea13d2b56a6b422bbdc390457.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 550, + 485, + 560 + ], + "lines": [ + { + "bbox": [ + 362, + 550, + 485, + 560 + ], + "spans": [ + { + "bbox": [ + 362, + 550, + 485, + 560 + ], + "type": "text", + "content": "\"A bedroom with a bed and a table.\"" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 564, + 545, + 609 + ], + "lines": [ + { + "bbox": [ + 45, + 564, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 45, + 564, + 545, + 609 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons of text-conditioned panorama generation. We show panoramas cropped to the vertical FoV of MVDiffusion [47]. Below each panorama, we show 4 evenly spaced perspective projections, with the first view crossing the left and right boundaries. We highlight the loop inconsistency, distorted lines and repetitive objects and unreasonable furniture layout of baseline methods with corresponding colors of boxes, which are addressed by our method. More results can be found in Sec. E of the supplementary." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 620, + 242, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 620, + 242, + 633 + ], + "spans": [ + { + "bbox": [ + 46, + 620, + 242, + 633 + ], + "type": "text", + "content": "4.2. Comparisons with Previous Methods" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 640, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 640, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 640, + 286, + 664 + ], + "type": "text", + "content": "Baselines. We compare our PanFusion with the following baselines (see Supplementary Sec. B for details)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": "- MVDiffusion [47] utilizes a multi-view diffusion model to generate 8 horizontal views that can be stitched into a panorama with vertical FOV of " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": ". It requires separate prompts for training while providing an option to generate" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 622, + 395, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 622, + 395, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 622, + 395, + 633 + ], + "type": "text", + "content": "from a single prompt." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 633, + 545, + 704 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 306, + 633, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 545, + 657 + ], + "type": "text", + "content": "- Text2Light [4] generates a " + }, + { + "bbox": [ + 306, + 633, + 545, + 657 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 306, + 633, + 545, + 657 + ], + "type": "text", + "content": " vertical FOV panorama from a text prompt in a two-stage auto-regressive manner." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 658, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 545, + 681 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 306, + 658, + 545, + 681 + ], + "type": "inline_equation", + "content": "SD + LoRA" + }, + { + "bbox": [ + 306, + 658, + 545, + 681 + ], + "type": "text", + "content": " is our baseline model that uses LoRA [14] to finetune a Stable Diffusion [31] on panorama images." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 681, + 545, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 681, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 306, + 681, + 545, + 704 + ], + "type": "text", + "content": "- Pano Branch is SD+LoRA with additional modifications as described in Sec. 3.2 to ensure loop consistency." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 50, + 103, + 59, + 517 + ], + "type": "aside_text", + "angle": 270, + "lines": [ + { + "bbox": [ + 50, + 103, + 59, + 517 + ], + "spans": [ + { + "bbox": [ + 50, + 103, + 59, + 517 + ], + "type": "text", + "content": "PanFusion (Ours) SD+LoRA MVDiffusion Text2Light" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6352" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 91, + 60, + 500, + 155 + ], + "blocks": [ + { + "bbox": [ + 91, + 60, + 500, + 155 + ], + "lines": [ + { + "bbox": [ + 91, + 60, + 500, + 155 + ], + "spans": [ + { + "bbox": [ + 91, + 60, + 500, + 155 + ], + "type": "table", + "html": "
MethodPanorama20 ViewsHorizontal 8 Views [47]
FAED ↓FID ↓IS ↑CS ↑FID ↓IS ↑FID ↓IS ↑CS ↑
Text2Light [4]97.2476.53.6027.4836.255.6743.664.9225.88
MVDiffusion [47]------25.276.9026.34
SD+LoRA [14, 31]7.1951.694.4028.8319.326.9020.686.4824.77
Pano Branch7.9050.404.5428.6720.107.0620.566.3724.85
PanFusion (Ours)6.0446.474.3628.5817.046.8519.886.5024.98
", + "image_path": "42b3e4e2af64bc38fb10c52c12d5f09764faac51360467db5988198904e651aa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 209, + 545, + 357 + ], + "blocks": [ + { + "bbox": [ + 49, + 209, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 49, + 209, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 545, + 357 + ], + "type": "image", + "image_path": "40c16d3582734f542585fe17be6a2c1752272ae3355bb8f80de6eb3ffd352fcf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 360, + 537, + 373 + ], + "lines": [ + { + "bbox": [ + 54, + 360, + 537, + 373 + ], + "spans": [ + { + "bbox": [ + 54, + 360, + 537, + 373 + ], + "type": "text", + "content": "Figure 5. Ablation study. Artifacts are highlighted with red boxes and projected to perspective views. Prompt: \"A hallway in a hotel.\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 381, + 285, + 470 + ], + "blocks": [ + { + "bbox": [ + 46, + 159, + 547, + 204 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 547, + 204 + ], + "type": "text", + "content": "Table 1. Comparison with SoTA methods. We evaluate the panorama image quality with Fréchet Auto-Encoder Distance (FAED), Fréchet Inception Distance (FID), Inception Score (IS), and CLIP Score (CS). In addition, we evaluate the perspective image quality in two settings. We first randomly sample 20 views, which is the closest to the real-world scenario where the user can freely navigate the panorama to view the scene from different perspectives. We then follow MVDiffusion [47] to horizontally sample 8 evenly spaced views." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 381, + 285, + 470 + ], + "lines": [ + { + "bbox": [ + 50, + 381, + 285, + 470 + ], + "spans": [ + { + "bbox": [ + 50, + 381, + 285, + 470 + ], + "type": "table", + "html": "
MethodPanorama20 Views
FAED ↓FID ↓IS ↑CS ↑FID ↓IS ↑
Pano Branch7.9050.404.5428.6620.107.06
Ours-joint7.7559.434.3427.9429.686.82
Ours-SPE6.6349.554.6528.6620.827.00
Ours-mask6.7745.494.3528.6615.816.81
Ours-bijective7.3648.354.5828.6618.787.12
Ours6.0446.474.3628.5817.046.85
", + "image_path": "d5f5e3e8628aaa8131651ab22dc0e516c250e055090b999baccbf536c36d0606.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 471, + 287, + 504 + ], + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 504 + ], + "type": "text", + "content": "Table 2. Ablation study. We compare the ablated versions of our method on both panorama and perspective domains. Here, “-” indicates removing the subsequent component." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 519, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 687 + ], + "type": "text", + "content": "Quantitative Results. Tab. 1 shows the quantitative comparison results. Here, we assign the highest value to realism in image generation, measured in FAED and FID. On these two metrics, our method outperforms baseline methods in both panorama and perspective. For IS, our method's performance is slightly lower than baselines. This is likely due to the fact that IS evaluates diversity of objects in generated images, using a classifier, and our model, unlike the baselines, tends to not generate unexpected objects. Similarly, it is possible to say that baseline models present slightly higher CS is due to the repetition of objects reinforcing alignment with prompts. Considering SD+LoRA is superior to Pano Branch on FAED and is on par in other metrics, we only qualitatively compare with SD+LoRA below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Qualitative Results. Fig. 4 shows the qualitative comparison results. Loop inconsistency can be observed for" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 383, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 545, + 491 + ], + "type": "text", + "content": "Text2Light and SD+LoRA due to lack of message passing between left-right boundaries. They also suffer from distorted lines in perspective views, meaning the generated panorama does not follow the correct equirectangular projection. MVDiffusion on the other hand suffers from repetitive objects and unreasonable furniture layout, which is likely due to the lack of global context. Our method generates the most realistic scenes and aligns to text condition the best, also with less distortion in perspective views." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 501, + 400, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 501, + 400, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 501, + 400, + 514 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 616 + ], + "type": "text", + "content": "In Sec. 4.2 and Tab. 1, we show that our full model outperforms Pano Branch, the baseline model of our method without the perspective branch. Here, as shown in Tab. 2 and Fig. 5, we further conduct an ablation study to validate the effectiveness of each component in our method. For a consistent comparison, we keep the layout similar between different ablated versions by sampling the same noise for latent map initialization, exploiting the observation of [22]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": "Joint latent map initialization. We ablate the joint latent map initialization by initializing the latent maps of the panorama and perspective branches separately (Ours-joint). Significant performance drop can be observed in all metrics and qualitative results, demonstrating the importance of joint latent map initialization. Interestingly, Ours-joint is even worse than Pano Branch in FID. It is likely due to that joint latent map initialization helps the corresponding" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6353" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 64, + 60, + 257, + 107 + ], + "blocks": [ + { + "bbox": [ + 52, + 60, + 61, + 107 + ], + "lines": [ + { + "bbox": [ + 52, + 60, + 61, + 107 + ], + "spans": [ + { + "bbox": [ + 52, + 60, + 61, + 107 + ], + "type": "text", + "content": "MVDiffusion" + } + ] + } + ], + "index": 0, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 60, + 257, + 107 + ], + "lines": [ + { + "bbox": [ + 64, + 60, + 257, + 107 + ], + "spans": [ + { + "bbox": [ + 64, + 60, + 257, + 107 + ], + "type": "image", + "image_path": "b57889ee13bce4ebcd5d22f7cbaf577de7e4aab8985a68acbdd81aee1158fa63.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 64, + 107, + 256, + 154 + ], + "blocks": [ + { + "bbox": [ + 53, + 122, + 60, + 140 + ], + "lines": [ + { + "bbox": [ + 53, + 122, + 60, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 122, + 60, + 140 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 1, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 107, + 256, + 154 + ], + "lines": [ + { + "bbox": [ + 64, + 107, + 256, + 154 + ], + "spans": [ + { + "bbox": [ + 64, + 107, + 256, + 154 + ], + "type": "image", + "image_path": "8312357515e061e142b64a63e35a959d260adb3ad1e47cd2a642184bc88e89c2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 157, + 229, + 167 + ], + "lines": [ + { + "bbox": [ + 138, + 157, + 229, + 167 + ], + "spans": [ + { + "bbox": [ + 138, + 157, + 229, + 167 + ], + "type": "text", + "content": "\"A hallway in a building.\"" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 257, + 60, + 304, + 107 + ], + "blocks": [ + { + "bbox": [ + 257, + 60, + 304, + 107 + ], + "lines": [ + { + "bbox": [ + 257, + 60, + 304, + 107 + ], + "spans": [ + { + "bbox": [ + 257, + 60, + 304, + 107 + ], + "type": "image", + "image_path": "d7d388b8919fd8609b7529f550620f5ba905b8872d8bec11693fae51ba7ba152.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 257, + 107, + 303, + 154 + ], + "blocks": [ + { + "bbox": [ + 257, + 107, + 303, + 154 + ], + "lines": [ + { + "bbox": [ + 257, + 107, + 303, + 154 + ], + "spans": [ + { + "bbox": [ + 257, + 107, + 303, + 154 + ], + "type": "image", + "image_path": "3191492e99170d3c2c89a76c59c093a394d500973fa4a5187bd38382149c4934.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 60, + 498, + 107 + ], + "blocks": [ + { + "bbox": [ + 306, + 60, + 498, + 107 + ], + "lines": [ + { + "bbox": [ + 306, + 60, + 498, + 107 + ], + "spans": [ + { + "bbox": [ + 306, + 60, + 498, + 107 + ], + "type": "image", + "image_path": "ab1ead31dc4bc0626532ab269220ad8d341f6f952fc1df6da75bc40229850700.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 107, + 499, + 154 + ], + "blocks": [ + { + "bbox": [ + 306, + 107, + 499, + 154 + ], + "lines": [ + { + "bbox": [ + 306, + 107, + 499, + 154 + ], + "spans": [ + { + "bbox": [ + 306, + 107, + 499, + 154 + ], + "type": "image", + "image_path": "4b34ff3731efe3f8f10ab37b0f854bd9bef7f5516236a1b9082ad9d4a0ec6627.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 157, + 496, + 167 + ], + "lines": [ + { + "bbox": [ + 354, + 157, + 496, + 167 + ], + "spans": [ + { + "bbox": [ + 354, + 157, + 496, + 167 + ], + "type": "text", + "content": "\"A living room with a view of the ocean.\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 499, + 60, + 545, + 107 + ], + "blocks": [ + { + "bbox": [ + 499, + 60, + 545, + 107 + ], + "lines": [ + { + "bbox": [ + 499, + 60, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 499, + 60, + 545, + 107 + ], + "type": "image", + "image_path": "84ba38cc3abca5f015b2e71d32686c1be76bbdd7fa3d542299f3609348748b4c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 499, + 107, + 545, + 154 + ], + "blocks": [ + { + "bbox": [ + 499, + 107, + 545, + 154 + ], + "lines": [ + { + "bbox": [ + 499, + 107, + 545, + 154 + ], + "spans": [ + { + "bbox": [ + 499, + 107, + 545, + 154 + ], + "type": "image", + "image_path": "44435942303edbbab2d31f069f1e90d43eaaeaaaef328ee0f3b87668fe9454ff.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 48, + 209, + 285, + 269 + ], + "blocks": [ + { + "bbox": [ + 48, + 209, + 285, + 269 + ], + "lines": [ + { + "bbox": [ + 48, + 209, + 285, + 269 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 285, + 269 + ], + "type": "table", + "html": "
MethodLayout ConsistencyHorizontal 8 Views [47]
3D IoU ↑2D IoU ↑FID ↓IS ↑CS ↑
MVDiffusion [47]61.0664.4328.835.6026.79
PanFusion (Ours)68.4671.8222.585.1026.04
GT image74.3177.15---
", + "image_path": "8e6213f6690ac1e8df1f0eea249417c13472acec173670c5252bce6b1ff880f4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 270, + 287, + 314 + ], + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 314 + ], + "type": "text", + "content": "Table 3. Layout-conditioned comparisons. We evaluate layout consistency between the condition and the layout of each generated image extracted by a layout estimator HorizonNet [45]. \"GT image\" indicates the upper bound of the layout estimator." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 326, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 361 + ], + "type": "text", + "content": "pixels to share a similar noise distribution from the beginning of the diffusion process, which is essential for EPPA to align the content of overlapped regions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 362, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 287, + 493 + ], + "type": "text", + "content": "EPP SPE and attention mask. We ablate spherical positional encoding (Ours-SPE) and attention mask (Ours-mask) from EPPA module. From Tab. 2, we can see that missing SPE hurts FAED and FID, which is likely due to that SPE helps the model to learn the relative position of the pixels between the two branches. Missing attention mask gets FID better, but it hurts FAED, which is more accurate in evaluating the panorama quality as it is customized for the target dataset. Both result in clear artifacts around point light sources, inconsistent textures of the floor, and distortions in the highlighted projections, as shown in Fig. 5." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 494, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 287, + 590 + ], + "type": "text", + "content": "Bijective EPPA. We ablate the bijective EPPA (Ours-bijective) by using separate parameters for the two directions in the EPPA module. Both FAED and FID get worse for Ours-bijective. In addition, the ablated version struggles to generate consistent textures for the floor and ceiling for the two directions of the hallway in Fig. 5. On the contrary, our full model generates floor and ceiling with consistent style, showing a better global understanding of the scene." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 599, + 282, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 282, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 282, + 612 + ], + "type": "text", + "content": "4.4. Application: Layout-Conditioned Generation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "To showcase the advantage of our method in generating panorama images with additional layout conditions, we build a baseline model by adding a ControlNet to MVDiffusion as described in Sec. 3.4. We render the layout condition into a distance map and then project it to perspective views to condition its generation of multi-view images. The training settings are kept the same as our PanFusion. As shown in Tab. 3, our method outperforms the baseline" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 308, + 209, + 426, + 269 + ], + "blocks": [ + { + "bbox": [ + 46, + 171, + 547, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 547, + 205 + ], + "type": "text", + "content": "Figure 6. Layout-conditioned generation comparisons. We showcase how layout-conditioned panorama generation can benefit from our dual-branch structure. The input layout is drawn on the generated panorama as yellow lines, while the panorama is cropped to the FoV of MVDiffusion and projected to one view to highlight the layout consistency." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 209, + 426, + 269 + ], + "lines": [ + { + "bbox": [ + 308, + 209, + 426, + 269 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 426, + 269 + ], + "type": "image", + "image_path": "f3669e5d95b40ce62785b7bd191bfad3e9c21d937d360d78f89652501cd50204.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 273, + 545, + 294 + ], + "lines": [ + { + "bbox": [ + 306, + 273, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 306, + 273, + 545, + 294 + ], + "type": "text", + "content": "Figure 7. Failure case. Our method sometimes generates indoor scenes with no entrance." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 427, + 209, + 544, + 269 + ], + "blocks": [ + { + "bbox": [ + 427, + 209, + 544, + 269 + ], + "lines": [ + { + "bbox": [ + 427, + 209, + 544, + 269 + ], + "spans": [ + { + "bbox": [ + 427, + 209, + 544, + 269 + ], + "type": "image", + "image_path": "d5674ed2fa49f2cf3face45ffd5ec237046424fe671b0ce18c2e3244ff50dd00.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 308, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 308, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 545, + 392 + ], + "type": "text", + "content": "model in layout consistency while keeping the realism advantage of perspective projections. We visualize the layout condition as wireframes overlaid on generated panorama images in Fig. 6, where we can see our generated panorama images follow their layout conditions better, especially as highlighted in perspective views. More details can be found in Sec. F of the supplementary material." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 404, + 378, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 404, + 378, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 404, + 378, + 417 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 426, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 592 + ], + "type": "text", + "content": "In this paper, we have proposed PanFusion, a novel text-to- " + }, + { + "bbox": [ + 304, + 426, + 545, + 592 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 426, + 545, + 592 + ], + "type": "text", + "content": "-panorama image generation method that can generate high-quality panorama images from a single text prompt. Particularly, a dual-branch diffusion architecture has been introduced to harness prior knowledge of Stable Diffusion in the perspective domain while addressing the repetitive elements and inconsistency issues observed in previous works. An EPPA module has been further introduced to enhance the information passing between the two branches. We have also extended our PanFusion for the application of layout-conditioned panorama generation. Comprehensive experiments have demonstrated that PanFusion can generate high-quality panorama images with better realism and layout consistency than previous methods." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 594, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 664 + ], + "type": "text", + "content": "Limitations. Although the dual-branch architecture of Pan-Fusion combines the advantages in both panorama and perspective domains, it comes with a cost of higher computational complexity. Additionally, our method sometimes fails to generate entrance for indoor scenes, as shown in Fig. 7, which is essential for use cases like virtual tour." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement: This research is supported by Building 4.0 CRC and the National Key R&D Program of China (NO.2022ZD0160101), and was partially done at Shanghai AI Laboratory." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6354" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Naofumi Akimoto, Yuhi Matsuo, and Yoshimitsu Aoki. *Diverse plausible 360-degree image outpainting for efficient 3dgc background creation*. In *CVPR*, pages 11441–11450, 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 286, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 286, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 286, + 170 + ], + "type": "text", + "content": "[2] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In ICML, pages 1737-1752. PMLR, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 172, + 286, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 172, + 286, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 172, + 286, + 215 + ], + "type": "text", + "content": "[3] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. 3DV, 2017. 5, 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 218, + 286, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 218, + 286, + 250 + ], + "spans": [ + { + "bbox": [ + 53, + 218, + 286, + 250 + ], + "type": "text", + "content": "[4] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM TOG, 41(6):1-16, 2022. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 253, + 286, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 253, + 286, + 306 + ], + "spans": [ + { + "bbox": [ + 53, + 253, + 286, + 306 + ], + "type": "text", + "content": "[5] Mohammad Reza Karimi Dastjerdi, Yannick Hold-Geoffroy, Jonathan Eisenmann, Siavash Khodadadeh, and Jean-François Lalonde. Guided co-modulated gan for " + }, + { + "bbox": [ + 53, + 253, + 286, + 306 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 53, + 253, + 286, + 306 + ], + "type": "text", + "content": " field of view extrapolation. In 3DV, pages 475–485. IEEE, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 310, + 286, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 310, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 53, + 310, + 286, + 342 + ], + "type": "text", + "content": "[6] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, pages 8780-8794, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 345, + 286, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 345, + 286, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 345, + 286, + 378 + ], + "type": "text", + "content": "[7] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 286, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 286, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 286, + 423 + ], + "type": "text", + "content": "[8] Chuan Fang, Xiaotao Hu, Kunming Luo, and Ping Tan. Ctrlroom: Controllable text-to-3d room meshes generation with layout constraints. arXiv preprint arXiv:2310.03602, 2023. 2, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 426, + 286, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 286, + 459 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 286, + 459 + ], + "type": "text", + "content": "[9] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. arXiv preprint arXiv:2302.01133, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 461, + 286, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 461, + 286, + 505 + ], + "spans": [ + { + "bbox": [ + 48, + 461, + 286, + 505 + ], + "type": "text", + "content": "[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 507, + 286, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 507, + 286, + 551 + ], + "spans": [ + { + "bbox": [ + 48, + 507, + 286, + 551 + ], + "type": "text", + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, pages 6626-6637, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 553, + 286, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 553, + 286, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 553, + 286, + 585 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 588, + 286, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 286, + 631 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 286, + 631 + ], + "type": "text", + "content": "[13] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In ICCV, pages 7909-7920, 2023. 2, 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 634, + 286, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 634, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 634, + 286, + 677 + ], + "type": "text", + "content": "[14] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3, 6, 7, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "text", + "content": "[15] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, pages 26565-26577, 2022. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Yuseung Lee, Kunho Kim, Hyunjin Kim, and Minhyuk Sung. Syncdiffusion: Coherent montage via synchronized joint diffusions. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[17] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. In NeurlPS, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "type": "text", + "content": "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 5, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 189, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 189, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 308, + 189, + 545, + 232 + ], + "type": "text", + "content": "[19] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, pages 5775-5787, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 235, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 235, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 235, + 545, + 277 + ], + "type": "text", + "content": "[20] Zhuqiang Lu, Kun Hu, Chaoyue Wang, Lei Bai, and Zhiyong Wang. Autoregressive omni-aware outpainting for open-vocabulary 360-degree image generation. arXiv preprint arXiv:2309.03467, 2023.1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 280, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 545, + 323 + ], + "type": "text", + "content": "[21] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 326, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 326, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 326, + 545, + 358 + ], + "type": "text", + "content": "[22] Jiafeng Mao, Xueting Wang, and Kiyoharu Aizawa. Guided image synthesis via initial image editing in diffusion model. In ACM MM, pages 5321-5329. ACM, 2023. 4, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 361, + 545, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 405 + ], + "type": "text", + "content": "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421. Springer, 2020. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 407, + 545, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 407, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 307, + 407, + 545, + 451 + ], + "type": "text", + "content": "[24] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 453, + 545, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 453, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 307, + 453, + 545, + 507 + ], + "type": "text", + "content": "[25] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, pages 16784-16804. PMLR, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "type": "text", + "content": "[26] Changgyoon Oh, Wonjune Cho, Yujeong Chae, Daehee Park, Lin Wang, and Kuk-Jin Yoon. Bips: Bi-modal indoor panorama synthesis via residual depth-aided adversarial learning. In ECCV, pages 352–371. Springer, 2022. 2, 5, 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 555, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 598 + ], + "type": "text", + "content": "[27] Chi-Han Peng and Jiayao Zhang. High-resolution depth estimation for 360deg panoramicas through perspective and panoramic depth images registration. In WACV, pages 3116-3125, 2023. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 601, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 655 + ], + "type": "text", + "content": "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 5, 4" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6355" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[30] Manuel Rey-Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution 360deg monocular depth estimation. In CVPR, pages 3762-3772, 2022. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 6, 7, 1, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "text", + "content": "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 230 + ], + "type": "text", + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 286, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 286, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 286, + 253 + ], + "type": "text", + "content": "[34] RunwayML. Stable diffusion. https://github.com/runwayml/stable-diffusion, 2021.2,3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 255, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 287, + 297 + ], + "type": "text", + "content": "[35] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 299, + 287, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 375 + ], + "type": "text", + "content": "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, pages 36479-36494, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 377, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 410 + ], + "type": "text", + "content": "[37] Tim Salimans, Ian J. Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. In NeurIPS, pages 2226-2234, 2016. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 465 + ], + "type": "text", + "content": "[38] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 554 + ], + "type": "text", + "content": "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. In NeurIPS, pages 25278–25294, 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 600 + ], + "type": "text", + "content": "[40] Ka Chun Shum, Hong-Wing Pang, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Conditional 360-degree image synthesis for immersive indoor scene decoration. In ICCV, pages 4478–4488, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "type": "text", + "content": "[41] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 657 + ], + "type": "text", + "content": "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3, 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[43] Liangchen Song, Liangliang Cao, Hongyu Xu, Kai Kang, Feng Tang, Junsong Yuan, and Zhao Yang. Roomdreamer: Text-driven 3d indoor scene synthesis with coherent geometry and texture. In ACM MM, pages 6898-6906. ACM, 2023. 2, 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[44] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, pages 11895–11907, 2019. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 151 + ], + "type": "text", + "content": "[45] Cheng Sun, Chi-Wei Hsiao, Min Sun, and Hwann-Tzong Chen. Horizonnet: Learning room layout with 1d representation and pano stretch data augmentation. In CVPR, pages 1047–1056, 2019. 5, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 153, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 196 + ], + "type": "text", + "content": "[46] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 199, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 199, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 199, + 545, + 243 + ], + "type": "text", + "content": "[47] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. In NeurIPS, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 245, + 545, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 288 + ], + "type": "text", + "content": "[48] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, JiaBin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In CVPR, pages 16773- 16783,2023.4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 291, + 545, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 291, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 307, + 291, + 545, + 345 + ], + "type": "text", + "content": "[49] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 347, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 347, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 347, + 545, + 380 + ], + "type": "text", + "content": "[50] Fu-En Wang, Yu-Hsuan Yeh, Min Sun, Wei-Chen Chiu, and Yi-Hsuan Tsai. Layouttmp3d: Layout annotation of matter-port3d. arXiv preprint arXiv:2003.13516, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 382, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 425 + ], + "type": "text", + "content": "[51] Guangcong Wang, Yinuo Yang, Chen Change Loy, and Zwei Liu. Stylelight: HDR panorama generation for lighting estimation and editing. In ECCV, pages 477-492. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 428, + 545, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 428, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 307, + 428, + 545, + 471 + ], + "type": "text", + "content": "[52] Guangcong Wang, Peng Wang, Zhaoxi Chen, Wenping Wang, Chen Change Loy, and Ziwei Liu. Perf: Panoramic neural radiance field from a single panorama. arXiv preprint arXiv:2310.16831, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 474, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 474, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 474, + 545, + 506 + ], + "type": "text", + "content": "[53] Hai Wang, Xiaoyu Xiang, Yuchen Fan, and Jing-Hao Xue. Customizing 360-degree panoramas through text-to-image diffusion models. In WACV, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 508, + 545, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 508, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 307, + 508, + 545, + 551 + ], + "type": "text", + "content": "[54] Jionghao Wang, Ziyu Chen, Jun Ling, Rong Xie, and Li Song. 360-degree panorama generation from few unregistered nfov images. In ACM MM, pages 6811-6821. ACM, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 554, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 554, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 307, + 554, + 545, + 597 + ], + "type": "text", + "content": "[55] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 600, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 600, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 600, + 545, + 643 + ], + "type": "text", + "content": "[56] Tianhao Wu, Chuanxia Zheng, and Tat-Jen Cham. IPO-ldm: Depth-aided 360-degree indoor rgb panorama outpainting via latent diffusion model. arXiv preprint arXiv:2307.03177, 2023. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "text", + "content": "[57] Jiale Xu, Jia Zheng, Yanyu Xu, Rui Tang, and Shenghua Gao. Layout-guided novel view synthesis from a single indoor panorama. In CVPR, pages 16438-16447, 2021. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[58] Mai Xu, Chen Li, Shanyi Zhang, and Patrick Le Callet. State-of-the-art in 360 video/image processing: Perception, assessment and compression. JSTSP, 14(1):5-26, 2020. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6356" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 464 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[59] Jian-Ru Xue, Jian-Wu Fang, and Pu Zhang. A survey of scene understanding by event reasoning in autonomous driving. MIR, 15(3):249-266, 2018. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 162 + ], + "type": "text", + "content": "[60] Bangbang Yang, Yinda Zhang, Yijin Li, Zhaopeng Cui, Sean Fanello, Hujun Bao, and Guofeng Zhang. Neural rendering in a room: amodal 3d understanding and free-viewpoint rendering for the closed scene composed of pre-captured objects. ACM TOG, 41(4):1-10, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "type": "text", + "content": "[61] Bangbang Yang, Wenqi Dong, Lin Ma, Wenbo Hu, Xiao Liu, Zhaopeng Cui, and Yuewen Ma. Dreamspace: Dreaming your room space with text-driven panoramic texture propagation. arXiv preprint arXiv:2310.13119, 2023. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 208, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 208, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 208, + 287, + 251 + ], + "type": "text", + "content": "[62] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM CSUR, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 253, + 287, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 253, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 49, + 253, + 287, + 295 + ], + "type": "text", + "content": "[63] Jason J. Yu, Fereshteh Forghani, Konstantinos G. Derpanis, and Marcus A. Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In ICCV, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 297, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 297, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 297, + 287, + 330 + ], + "type": "text", + "content": "[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 331, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 331, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 49, + 331, + 287, + 364 + ], + "type": "text", + "content": "[65] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming yu Liu. Diffcollage: Parallel generation of large content with diffusion models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 365, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 287, + 408 + ], + "type": "text", + "content": "[66] Chuanqing Zhuang, Zhengda Lu, Yiqun Wang, Jun Xiao, and Ying Wang. Acdnet: Adaptively combined dilated convolution for monocular panorama depth estimation. In AAAI, pages 3653-3661, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 411, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 411, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 411, + 287, + 464 + ], + "type": "text", + "content": "[67] Chuhang Zou, Jheng-Wei Su, Chi-Han Peng, Alex Colburn, Qi Shan, Peter Wonka, Hung-Kuo Chu, and Derek Hoiem. Manhattan room layout reconstruction from a single 360 image: A comparative study of state-of-the-art methods. IJCV, 129:1410-1431, 2021. 1" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6357" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/301b4dc8-6f1d-458c-a509-0fe38145fc73_content_list.json b/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/301b4dc8-6f1d-458c-a509-0fe38145fc73_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..84b34a511c2575806b6edea32e3bdded0febfcb8 --- /dev/null +++ b/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/301b4dc8-6f1d-458c-a509-0fe38145fc73_content_list.json @@ -0,0 +1,1533 @@ +[ + { + "type": "text", + "text": "Taming the Tail in Class-Conditional GANs: Knowledge Sharing via Unconditional Training at Lower Resolutions", + "text_level": 1, + "bbox": [ + 140, + 130, + 828, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Saeed Khorram1, Mingqi Jiang1, Mohamad Shahbazi2, Mohamad H. Danesh3, Li Fuxin1 \n1Oregon State University 2 ETH Zürich 3 McGill University \n{khorrams, jiangmi, lif}@oregonstate.edu \nmshahbazi@vision.ee.ethz.ch, mohamad.danesh@mail.mcgill.ca", + "bbox": [ + 130, + 202, + 836, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 309, + 313, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite extensive research on training generative adversarial networks (GANs) with limited training data, learning to generate images from long-tailed training distributions remains fairly unexplored. In the presence of imbalanced multi-class training data, GANs tend to favor classes with more samples, leading to the generation of low quality and less diverse samples in tail classes. In this study, we aim to improve the training of class-conditional GANs with long-tailed data. We propose a straightforward yet effective method for knowledge sharing, allowing tail classes to borrow from the rich information from classes with more abundant training data. More concretely, we propose modifications to existing class-conditional GAN architectures to ensure that the lower-resolution layers of the generator are trained entirely unconditionally while reserving class-conditional generation for the higher-resolution layers. Experiments on several long-tail benchmarks and GAN architectures demonstrate a significant improvement over existing methods in both the diversity and fidelity of the generated images. The code is available at https://github.com/khorrams/utlo.", + "bbox": [ + 76, + 342, + 473, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 675, + 207, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the past few years, research on Generative Adversarial Networks (GANs) [7] has led to remarkable advances in generating realistic images [2, 17, 41]. Conditional GANs [29] (cGANs) have garnered particular attention due to their ability to accept user inputs which can additionally guide the generation process. They enable a wide range of applications such as class-conditional image generation [42], image manipulation [43], image-to-image translation [12], super-resolution [22] and text-to-image synthesis [52].", + "bbox": [ + 76, + 702, + 468, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the advances, past research on cGANs has focused primarily on learning from balanced data. In real-world scenarios, however, data often follows a power-law distribution, with a few classes dominating most of the", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f4a0dd369a59fafe23dd9933222693c0433260cfcca2151847975791f6d3ba9.jpg", + "image_caption": [ + "All (2) Training Images in Tail Classes" + ], + "image_footnote": [], + "bbox": [ + 517, + 325, + 874, + 397 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dccb4b8415eb372b06fec69c1ebfed3f30347bf85f4d5dd403cc004810a74dbc.jpg", + "image_caption": [ + "Baseline (FID: 9.8)" + ], + "image_footnote": [], + "bbox": [ + 517, + 414, + 874, + 553 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a40cfaf4559f03714d37f685a7014e6dbc4fe2f6fdb51d0c41d69ea5c81b594.jpg", + "image_caption": [ + "Baseline + Ours (FID: 6.6)", + "Figure 1. Generating images from rare tail classes in the Flowers-LT with only two training images. Our proposed approach allows for a more diverse set of features such as backgrounds, colors, poses, and object layouts to be infused into the tail classes." + ], + "image_footnote": [], + "bbox": [ + 517, + 571, + 874, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "training data. This is referred to as the \"long-tail (LT) problem\" [55], where most of the training data come from a few classes (referred to as the \"head\" of the distribution). In contrast, data from a large number of classes rarely occur (referred to as the \"tail\" of the distribution). This imbalance in data poses a challenge to the effective learning of tail classes, as standard machine learning algorithms tend", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 808, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7580", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to favor densely sampled regions of the input distribution.", + "bbox": [ + 76, + 90, + 460, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although there has been a surge of interest in the long-tail problem in recent years, the focus has been primarily on recognition tasks [49]. Generative learning on the long-tail data, particularly cGANs which is the focus of this work, remains underexplored. Recently, [36, 37] identified mode collapse on tail classes as a result of the spectral norm explosion in class-conditional BatchNorms and the collapse of the latents in the StyleGANs $\\mathcal{W}$ space. While techniques such as regularization, normalization, and/or classbalancing techniques [49] might improve the performance of long-tail generative models to some extent, their effectiveness is limited. The primary challenge in the long-tail problem is the under-representation of tail classes in the training data, e.g., in Fig. 1 each tail class contains merely 2 training images, resulting in insufficient observation of the tail classes by the learning algorithm.", + "bbox": [ + 75, + 108, + 468, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, we propose to leverage the abundant information in the head classes, with the aim to learn and infuse knowledge from the head classes into the tail ones, thereby enriching the training distribution for the latter. However, it is very difficult to explicitly disentangle the class-specific information pertaining to head classes from class-independent information shared between head and tail classes. We make a key observation that the similarity between head and tail instances is often higher at lower resolutions: discriminative features such as shape or specific texture are usually unveiled at higher resolutions. Conversely, information from the lower resolutions tends to be more class-independent such as background, configuration, or direction of objects (Fig. 6), thus can be shared between the head and tail classes.", + "bbox": [ + 75, + 351, + 470, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building upon this observation, we propose Unconditional Training at LOWER resolution (UTLO), a novel approach for training cGANs in the long-tail setup. In addition to the standard conditional GAN objective, UTLO trains the intermediate low-resolution output of the generator with an unconditional GAN objective. The unconditional GAN objective encourages the learning of low-resolution features common to both the head and tail classes, which are infused into the subsequent layers of the generator, particularly benefiting the under-represented tail classes.", + "bbox": [ + 75, + 580, + 468, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through extensive experiments and analysis on several long-tailed datasets, we demonstrate that using UTLO to combine high-resolution conditional and low-resolution unconditional training effectively facilitates knowledge sharing between head and tail classes, thereby improving the overall generative modeling of cGANs in the long-tail setup. Due to the strong class imbalance in the long-tailed data, naive usage of existing GAN metrics can be misleading. To mitigate this issue, we propose a few practices to adapt commonly-used GAN metrics to the long-tail setup.", + "bbox": [ + 75, + 732, + 468, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Below, we highlight our main contributions:", + "bbox": [ + 96, + 885, + 388, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose UTLO, a novel knowledge-sharing framework tailored for training cGANs in the long-tailed setup. UTLO allows infusing information from the head classes to the tail classes via an additional unconditional objective applied to the low-resolution part of the cGAN generator. To the best of our knowledge, this work is the first to demonstrate that not all layers in a cGAN need to receive class-conditional information, i.e., a cGAN featuring a partially unconditional generator.", + "- We present a set of practices and metrics designed to adapt the commonly-used GAN evaluation metrics for long-tail setups, enabling a more precise evaluation of the image generation quality.", + "- Through extensive experiments across multiple benchmarks and architectures, we validate the effectiveness of our proposed method in improving the training of cGANs in the long-tail setup, achieving state-of-the-art results across several long-tail datasets." + ], + "bbox": [ + 500, + 90, + 890, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 376, + 640, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conditional Generative Adversarial Networks The initial work on conditional image generation using GANs [7] was presented by [29], in which the class conditions were concatenated to the inputs of the generator and discriminator networks. The AC-GAN [33] proposed the use of auxiliary classification in the discriminator. [2] introduced BigGAN and set a milestone in large-scale and high-resolution conditional image generation. The authors of StyleGAN [17, 18] first map the input noise and class condition to a latent style space, which is then passed to the multiple layers of the generator for image synthesis. More recently, an extension of the StyleGAN called StyleGAN-XL [41] has become the state-of-the-art in conditional image generation on several datasets and outperforms more complex and time-consuming approaches such as diffusion models [6].", + "bbox": [ + 496, + 402, + 890, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "GAN Regularization under Limited Data Training GANs under limited data is challenging as the discriminator can memorize the training samples, resulting in the collapse of the training or quality degradation of the generated images. Recently, data augmentation and regularization techniques have been incorporated into GAN training as a means to mitigate this problem [15, 17, 46, 53, 57, 58]. Additionally, [47] introduced a regularization term to the GAN objective, which tracks the predictions of the discriminator for real and generated images using separate moving averages. [21] employed off-the-shelf vision models in an ensemble of discriminators, demonstrating improved performance in both limited-data and large-scale GAN training.", + "bbox": [ + 496, + 628, + 890, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the context of few-shot image generation, FastGAN [23] proposed a lightweight architecture with a self-supervised discriminator. [34] introduced cross-domain distance consistency in order to transfer diversity from a source domain to a target domain. Subsequently, [19] proposed", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7581", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a latent-mixup strategy to smooth the latent space through controlled latent interpolation.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, [42, 57] have observed that class-conditioning can cause mode collapse in the limited data regime. Their proposed work learns from limited but balanced/unlabeled data, which is different from our setup: heavily imbalanced long-tail data. [42] proposed Transitional-cGAN, a training strategy that starts with unconditional GAN objective and injects class-conditional information during a transition period. This approach was found to be effective in mitigating mode-collapse in the limited data regime. Our main idea of separating pathways between low-resolution and high-resolution is novel w.r.t. [42]: to promote knowledge sharing, we modify the cGAN architecture so that the lower-resolution part of the generator is entirely unconditional while only the higher-resolution part is conditional and received class-conditional information.", + "bbox": [ + 75, + 125, + 470, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Long-tail Recognition To address the long-tail recognition problem [49, 55], previous research has primarily focused on techniques such as class re-balancing [3, 28, 38], learning algorithm and model design [13, 14, 59], and information augmentation [24, 45, 48]. More related to our work, re-sampling [8, 56] has been widely used to handle class imbalance problems. This involves balancing the number of samples used during training through over-sampling and/or under-sampling. Oversampling increases the frequency of tail classes [9, 31] while under-sampling reduces the imbalance by reducing the frequency of head class instances [25, 50]. Note that our research diverges from prior studies on long-tail recognition by addressing the more challenging task of generative modeling of the long-tail data, as opposed to the conventional long-tail classification task.", + "bbox": [ + 75, + 354, + 470, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training GANs on Long-tail Data Recent years have seen a growing interest in learning from long-tail data. However, most focus has been directed toward recognition tasks, with a limited number of studies addressing the development of generative models for long-tail data. GAMO [31] uses adversarial training to over-sample from minority classes. [35] introduced class-balancing regularization to the GAN objective, which leverages the predictions of a pretrained classifier to improve the generation of underrepresented classes in an unconditional setting. However, this method trains an unconditional GAN and is restricted to having access to a pre-trained classifier on the long-tail data to guide the training. The most closely related work to ours is the Group Spectral Regularization (GSR) [36] and NoisyTwins [37] regularizations. The authors identified mode collapse on tail classes as a result of the spectral norm explosion in class-conditional BatchNorms and the collapse of the latents in the StyleGANs $\\mathcal{W}$ space, respectively. In contrast, we present a novel approach for sharing knowledge between head and tail classes to enhance long-tail learning, which is orthogonal to regularization techniques. In addition, while", + "bbox": [ + 75, + 583, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[37] is restricted to StyleGANs, our framework can be extended to different GAN architectures.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Taming the Tail in cGANs", + "text_level": 1, + "bbox": [ + 498, + 135, + 743, + 152 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background: Conditional Generative Adversarial Networks", + "text_level": 1, + "bbox": [ + 498, + 160, + 890, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conditional Generative Adversarial Networks (cGANs) are suitable for many applications due to providing user control over the generated samples at inference time. This is particularly helpful in the long-tail setup, where it is desirable to explicitly generate samples from rare (tail) classes. In contrast, unconditional GANs are likely to generate samples that track the training distribution, resulting in most samples coming from common (head) classes. To train a cGAN over a dataset containing $n$ instances $\\mathbb{X} = \\{\\pmb{x}_1,\\dots ,\\pmb{x}_n\\}$ and their corresponding labels $\\mathbb{Y} = \\{y_{1},\\ldots ,y_{n}\\}$ , we can formulate the adversarial training objective that alternatively optimizes the following loss terms,", + "bbox": [ + 496, + 199, + 890, + 380 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} ^ {D} = \\mathbb {E} _ {\\boldsymbol {x}, y} [ f _ {D} (- D (\\boldsymbol {x} | y)) ] + \\mathbb {E} _ {\\boldsymbol {z}, y} [ f _ {D} (D (G (\\boldsymbol {z}, y))) ]\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 391, + 879, + 410 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} ^ {G} = \\mathbb {E} _ {\\boldsymbol {z}, y} [ f _ {G} (- D (G (\\boldsymbol {z}, y))) ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 412, + 890, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $G$ and $D$ are the generator and discriminator networks with their corresponding loss functions $f_{G}$ and $f_{D}$ [27]; $\\pmb{x}|y\\sim p_{\\mathrm{data}}(\\pmb{x}|y)$ is real data drawn from class $y$ . The generator $G$ conditions on both $z\\sim p_z(z)$ , a random noise vector from a prior distribution over latent space, and $y$ , the class conditioning vector. This adversarial training objective encourages the cGAN to reach an equilibrium between the generator and the discriminator, resulting in the generator producing realistic samples indistinguishable from real data by the discriminator.", + "bbox": [ + 496, + 443, + 890, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Caveat of Training cGANs on Long-tail Data", + "text_level": 1, + "bbox": [ + 498, + 604, + 879, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training GANs on long-tail data is challenging due to the inherent difficulty in modeling the rare examples that comprise the tail of the data distribution. This skewness in the data distribution hinders balanced learning across classes. In particular, the discriminator does not see enough examples from the tail of the distribution during training, which can lead to poor discriminative signals for the generator. As a result, the generator learns to generate a small subset of the possible outputs that fools the inadequately learned discriminator – exhibiting a classic “mode collapse” scenario.", + "bbox": [ + 496, + 628, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although regularizations introduced in [36, 37] can help mitigate the collapse of learning in tail classes that contain a sufficiently large number of samples, we still observe mode collapse occurring when a very limited number of samples are in tail classes, which is inherently challenging. This can be seen in Fig. 2 for tail classes of CIFAR100-LT with as few as 5 training instances. Although early-stopping can be adopted to obtain reasonable performance, the key in boost-", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7582", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b5a8d2f28344ede7ebda1ecd806c5c7e43db2d70a1b4ea0f75bfcfdff7bb5bef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 95, + 459, + 306 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8ffed90ca950e0de7fb240c83639adec6c3eb98e8554a766c9bcc84b33693056.jpg", + "image_caption": [ + "Figure 2. Convergence of different methods on CIFAR100-LT $(\\rho = 100)$ , where tail classes have as few as 5 training examples. Incorporating our framework into the baseline alleviates overfitting as a result of knowledge sharing from head to rare tail classes." + ], + "image_footnote": [], + "bbox": [ + 78, + 313, + 176, + 395 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/db32d36e6b935cb2edc49caa32715f170289f72b0b7654b8866f9873146b8867.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 314, + 274, + 395 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f621dd8311f986bceac51b1440c3dac89ae767d025c387d52ce32b7eb67a7541.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 314, + 370, + 395 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dfe73fe5f9a38b618e16649cb2fceeedf5d855d97c7662c26767554f88425339.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 314, + 468, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing the performance would be an approach to infuse additional diversities into the tail classes rather than only relying on techniques such as regularization.", + "bbox": [ + 75, + 468, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Knowledge Sharing via Unconditional Training at Lower Resolutions (UTLO)", + "text_level": 1, + "bbox": [ + 76, + 527, + 468, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We devise a novel generative adversarial network (GAN) training scheme for long-tail data that aims to share knowledge from predominant head classes to rare tail classes, assuming some of the tail and head classes are at least coarsely similar. We propose a framework that uses unconditional learning in low-resolution images / features at an intermediate layer of the generator using the unconditional GAN objective. This helps with learning universal features from both head and tail classes. The subsequent layers, responsible for introducing finer details at higher resolutions, are trained using a class-conditional GAN objective.", + "bbox": [ + 75, + 566, + 468, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our proposed framework, we perform conditional training on the high-resolution image output from the final layer of the generator network. This is built upon the features learned at lower layers through the use of an unconditional objective, inheriting features primarily from head classes. The conditioning on the class labels gives control at the inference time to explicitly generate images from tail classes, meeting our design desideratum. The combination of unconditional and conditional objectives using our proposed method enables knowledge sharing, which, in turn, improves the quality of GAN training on long-tail data.", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our framework is general and applicable to many GAN architectures. In this section, we illustrate the application of our framework on StyleGAN2 with adaptive data augmentation (ADA) [17], a state-of-the-art and solid baseline for training GANs, particularly in the limited data regime. In the following, we demonstrate the necessary modifications to the architecture of the generator and discriminator networks. Although we show the necessary modifications for StyleGAN2-ADA, our method can be easily extended to other GAN architectures such as FastGAN [23], which is also used in our experiments (see Section 4).", + "bbox": [ + 496, + 90, + 893, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Modifying the Generator", + "text_level": 1, + "bbox": [ + 498, + 273, + 728, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the generator design of the StyleGAN2-ADA [17, 18], the class-conditional information, combined with the latent vector, is first embedded in the style space $w$ using the style-mapping network $G_{\\mathrm{map}}$ . The style vectors are then broadcasted across the layers of the synthesis network $G_{\\mathrm{syn}}$ in order to generate diverse images. To meet our design desideratum, we dissect the synthesis network into two subnetworks $G_{\\mathrm{syn}} = G_h \\circ G_l$ where $G_l$ represents the earlier part of the network that produces intermediate features and/or images at a low resolution $L \\times L$ . $G_h$ on the other hand, represents the latter part of $G_{\\mathrm{syn}}$ that generates the output image at the high resolution $H \\times H$ where $H > L$ .", + "bbox": [ + 496, + 296, + 890, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To block the flow of class-conditional information to the lower layers of $G_{\\mathrm{syn}}$ , we generate separate $w$ vectors, one containing the class-conditional embeddings, referred to as $w_{z,y}$ , while for the other one, referred to as $w_{z}$ , class information was set to zero. Note that both vectors share the same latent variable $z$ and mapping network $G_{\\mathrm{map}}$ parameters, and the only difference is the presence of class-conditional information. The lower layers of the synthesis network $G_{l}$ are conditioned on $w_{z}$ while the subsequent layers at higher resolutions receive $w_{z,y}$ (see Fig. 3).", + "bbox": [ + 496, + 478, + 892, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generators typically follow a network design that gradually increases the resolution of intermediate features and/or images as the network progresses. This allows us to select a desired low-resolution, such as $8 \\times 8$ or $16 \\times 16$ , during training. Recent generator designs often incorporate skip connections [16] or residual connections [30] to improve gradient flow. The generator of StyleGAN2-ADA uses skip connections, which explicitly generates intermediate RGB images at each layer. We take this low-resolution image as the input to the discriminator. Note that in the case of a generator design that uses residual connections, a $1 \\times 1$ convolutional layer can be incorporated to convert the residual channels to RGB channels.", + "bbox": [ + 496, + 628, + 893, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In a forward pass through the generator, images at low and high resolutions can be generated simultaneously without additional overhead. By inputting a latent code $z \\sim p_{z}(z)$ and a target label $y$ , we can obtain an unconditional synthetic image $\\hat{x}_l \\in \\mathbb{R}^{3 \\times L \\times L}$ along with a conditional", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7583", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8e61a754a7a7b98da3145e27a96323c25a3973fae410e01287b27271836d3112.jpg", + "image_caption": [ + "Figure 3. The proposed framework, UTLO, illustrated for the StyleGAN2-ADA architecture. Low and high resolution image pathways are used for unconditional and class-conditional objectives, respectively. $z$ is the input latent code, $y$ indicates the class embeddings, and $c$ is a constant input. Separate style vectors, $w_{z}$ (class-independent) and $w_{z,y}$ (class-conditional), are generated using the same $z$ and a shared style-mapping network $G_{\\mathrm{map}}$ which are then passed to $G_{l}$ and $G_{h}$ , respectively. The high-resolution generated image $\\hat{x} \\in \\mathbb{R}^{3\\times H\\times H}$ is passed through the discriminator $D = D_{l} \\circ D_{h}$ to calculate the conditional objective $\\mathcal{L}_c$ while the low-resolution image $\\hat{x}_l \\in \\mathbb{R}^{3\\times L\\times L}$ is passed only through $D_{l}$ to calculate the unconditional objective $\\mathcal{L}_{uc}$ . The final objective $\\mathcal{L}$ is the combination of the two. While $D_{l}$ is shared, two separate prediction heads (FC layers) are used for unconditional and conditional objectives. The fromRGB is designed to increase the dimensionality of RGB channels to match the input channels of the $D_{l}$ ." + ], + "image_footnote": [], + "bbox": [ + 181, + 87, + 794, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "high-resolution synthetic image $\\hat{\\pmb{x}}\\in \\mathbb{R}^{3\\times H\\times H}$ at the output. While $\\hat{\\pmb{x}}$ is trained to be from class $y$ , $\\hat{\\pmb{x}}_l$ has no class-specific constraints. At inference time, the intermediate images can be discarded.", + "bbox": [ + 75, + 381, + 470, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Modifying the Discriminator", + "text_level": 1, + "bbox": [ + 76, + 453, + 331, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Opposite to the generator design, discriminators gradually reduce the resolution of the intermediate features as the network progresses. In order to pass images to an intermediate layer of the discriminator, we dissect it into two sub-networks. Formally, we define the discriminator as $D = D_{l} \\circ D_{h}$ where $D_{h}$ represents the earlier part of the original network that takes the images at training resolution $\\mathbb{R}^{3 \\times H \\times H}$ . On the other hand, $D_{l}$ represents the latter part of the discriminator taking a lower resolution input $\\mathbb{R}^{C \\times L \\times L}$ with $C$ channels.", + "bbox": [ + 75, + 477, + 468, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The default discriminator design in StyleGAN2-ADA uses residual connections. This does not allow direct passing of RGB inputs to the intermediate layers. To overcome this, we add a $1 \\times 1$ convolutional layer, referred to as \"fromRGB\", that increases the channels of the RGB images to match $C$ , the number of input channels of $D_{l}$ .", + "bbox": [ + 75, + 628, + 468, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More concretely, $D_{l}$ is a shared network in both the unconditional low-resolution and class-conditional high-resolution image pathways (see Fig. 3). To obtain separate unconditional and conditional predictions, we use two different fully-connected (FC) layers at the final layer of $D_{l}$ . The output of the $D_{h}$ and the low-res image $\\hat{\\boldsymbol{x}}_l$ are passed through the $D_{l}$ which are then used for calculating the class-conditional loss $\\mathcal{L}_c$ and unconditional loss $\\mathcal{L}_{uc}$ , respectively.", + "bbox": [ + 75, + 719, + 470, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note, when passing real images to $D_{l}$ for the unconditional loss, they are first downsized to the low-resolution $L \\times L$ using bi-linear interpolation to be comparable with", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the ones generated from $G_{l}$ . They are then passed through the fromRGB layer and consequently $D_{l}$ . For the class-conditional loss, the original images are passed to $D$ without any modification.", + "bbox": [ + 496, + 382, + 892, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3 Final Training Objective", + "text_level": 1, + "bbox": [ + 498, + 446, + 725, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final objective of our framework is simply a combination of the conditional and unconditional GAN losses, using a weighting parameter $\\lambda$ . To put it formally we have,", + "bbox": [ + 496, + 470, + 892, + 516 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {D} = \\mathcal {L} _ {c} ^ {D} + \\lambda \\cdot \\mathcal {L} _ {u c} ^ {D}, \\mathcal {L} ^ {G} = \\mathcal {L} _ {c} ^ {G} + \\lambda \\cdot \\mathcal {L} _ {u c} ^ {G} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 523, + 892, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_c^D$ and $\\mathcal{L}_c^G$ represent the class-conditional losses for the discriminator and generator, respectively (as outlined in Eq. 3.1), and $\\mathcal{L}_{uc}^D$ and $\\mathcal{L}_{uc}^G$ represent their unconditional counterparts, where $\\pmb{x}$ is sampled from $p_{\\mathrm{data}}(\\pmb{x})$ instead of $p_{\\mathrm{data}}(\\pmb{x}|y)$ . The overall framework of our proposed method is depicted in Fig. 3.", + "bbox": [ + 496, + 551, + 893, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 656, + 633, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 500, + 681, + 581, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We use 6 different long-tailed datasets in our experiments: CIFAR10 and CIFAR100[20], LSUN [51], Flowers [32], iNaturalist2019 [11], and AnimalFaces [44]. This selection is intended to encompass a wide spectrum of image domains, dataset sizes, resolutions, and imbalance ratios (denoted as $\\rho$ ). $\\rho$ represents the ratio of the number of training samples between the classes with the most and the least number of examples. To ensure the tail classes stay few-shot, we keep the number of images in the smallest tail classes under 50 [49]. A detailed description of these datasets is provided in Supp. A.", + "bbox": [ + 496, + 704, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. In order to demonstrate the effectiveness of our proposed method, we use various cGAN architectures cov", + "bbox": [ + 496, + 869, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7584", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ering different generator and discriminator designs and data augmentation pipelines:", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- StyleGAN2 with Adaptive Data Augmentation (ADA) [17]. We also integrate Transitional training [42], Group Spectral Regularization (GSR) [36], and NoisyTwins [37] to this baseline.", + "- Projected GAN (PGAN) [40]: we use projected discriminator with Differentiable Augmentation (DA) [57]. We also add GSR [36] to this baseline." + ], + "bbox": [ + 76, + 121, + 468, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For comprehensive details on the baselines, hyperparameters, and implementation please refer to Supp. F.", + "bbox": [ + 75, + 228, + 468, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation Metrics For Long-Tail Datasets", + "text_level": 1, + "bbox": [ + 76, + 263, + 441, + 279 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluating the images generated by GANs on a long-tailed setup poses challenges, primarily due to the imbalanced data and access to a limited number of samples for the tail classes. For our experiments, we employ widely-used GAN metrics: Fréchet Inception Distance (FID) [10], and Kernel Inception Distance (KID) [1]. Following [10, 17], we report all metrics by computing statistics between $50k$ generated images and all available training images. In the following, we present a set of practices for adapting the aforementioned metrics to the long-tail setup.", + "bbox": [ + 75, + 282, + 468, + 433 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, when a larger and (more) balanced dataset following the long-tail training data distribution is available, it is used for metric calculation, e.g., full CIFAR10 before artificial imbalance. For naturally imbalanced datasets such as iNaturalist2019, we use the unmodified training set to calculate metrics. Note that when generating images for evaluation, we sample from the same distribution as the available real dataset, which may be imbalanced. While FID and KID indicate how well the real data distribution matches the generated images across all classes, we suggest additional metrics to mitigate the disproportionate influence of head classes in data statistic calculations.", + "bbox": [ + 75, + 434, + 468, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In response to the varied number of training instances across classes in a long-tail setup, we propose evaluating metrics specifically on few-shot (FS) categories [49]. We calculate the FID and KID for the few-shot subset and refer to them as \"FID-FS\" and \"KID-FS\", respectively. This is tailored to evaluate the quality of generated samples in the tail classes. Contrasting with the standard FID/KID, we maintain an equal number of real images across all classes during our FID-FS/KID-FS calculation. This emphasizes learning quality on tail classes, irrespective of any imbalances that might be present in the few-shot subset. Reporting both standard and few-shot metrics provides a more comprehensive evaluation of long-tail learning, considering the performance of both head and tail classes.", + "bbox": [ + 75, + 614, + 468, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results", + "text_level": 1, + "bbox": [ + 76, + 830, + 171, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Below, we report the results obtained across different benchmarks and cGAN architectures. We pick the best model for reporting results as the one with the lowest FID", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/626f4b8be4441d4ef817a8a56269e8a7cd570dd8a4ee28e69e72f18789b515e5.jpg", + "table_caption": [ + "Table 1. Our proposed method, UTLO, outperforms the baselines in terms of quantitative image quality metrics on the AnimalFaces-LT dataset (64 × 64 resolution)." + ], + "table_footnote": [], + "table_body": "
MethodsFID ↓FID-FS ↓KID ↓KID-FS ↓ ×1000
StyleGAN2-ADA UnCond. [17]39.4104.117.327.6
StyleGAN2-ADA [17]51.487.124.735.9
+ Transitional [42]62.199.038.548.9
+ GSR[36]39.267.221.232.7
+ NoisyTwins [37]29.450.216.721.2
+ UTLO (Ours)26.248.412.619.6
", + "bbox": [ + 501, + 137, + 888, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FS from two independent runs. Note that our method exhibits significantly less reliance on early stopping compared to the baselines, as illustrated in Fig. 2.", + "bbox": [ + 496, + 247, + 890, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "AnimalFaces-LT This benchmark has a naturally occurring imbalance across its classes. We compare our method against the StyleGAN2-ADA benchmarks in terms of quantitative image quality metrics in Table. 1. Our method outperforms the baselines across all metrics, as demonstrated in the results. For a visual comparison against the baselines, please refer to Supp. H.", + "bbox": [ + 496, + 292, + 890, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation on the Choice of Low-resolution for Unconditional Training ( $res_{\\mathrm{uc}}$ ). In our proposed method, one of the hyperparameter choices is to select an intermediate low-resolution $res_{\\mathrm{uc}}$ for unconditional training. All layers with equal or lower resolution than $res_{\\mathrm{uc}}$ do not receive class-conditional information. An unconditional GAN objective is added over the images and/or features at $res_{\\mathrm{uc}}$ . To study the impact of $res_{\\mathrm{uc}}$ , we conducted an ablation study on the AnimalFaces-LT dataset, containing images at $64 \\times 64$ . Table. 2 presents the results for selecting $res_{\\mathrm{uc}}$ from resolutions lower than $64 \\times 64$ , namely $8 \\times 8$ , $16 \\times 16$ , and $32 \\times 32$ . Studying the obtained results, we observe that resolutions of $8 \\times 8$ and 16 achieve relatively close performance. Conversely, the performance diminishes when layers up to $32 \\times 32$ are trained unconditionally. This is anticipated as the AnimalFaces-LT is at $64 \\times 64$ , leaving only one up-sampling layer to learn the class-conditional information which is shown to be insufficient. In all our experiments, we use $res_{\\mathrm{uc}}$ of $8 \\times 8$ unless stated otherwise. Additional visual analysis on the role of $res_{\\mathrm{uc}}$ as well as an ablation on the choice of unconditional training objective weight $\\lambda$ (see Eq. 2), need for unconditional layers in the discriminator, and distinction from \"coarse-to-fine\" training strategies are presented in the Supp. C.", + "bbox": [ + 496, + 400, + 892, + 762 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e7ae853e3efe14146179baee13f165141f1dc424c854feef213faa5baac3dec6.jpg", + "table_caption": [ + "Table 2. Ablation on choice of low resolution for unconditional training $(res_{\\mathrm{uc}})$ ." + ], + "table_footnote": [], + "table_body": "
resucFID ↓FID-FS ↓KID ↓KID-FS ↓ ×1000
8 × 826.248.412.619.6
16 × 1627.550.313.720.8
32 × 3238.064.923.334.3
", + "bbox": [ + 558, + 801, + 828, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Can any low-resolution be the starting point for highres generations? In general, our observations indicate that", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7585", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "high-res images are guided, yet not entirely restricted by the low-res images, as shown in Fig. 4. Each row in this figure presents a set of different high-resolution images that are generated from the same unconditional low-resolution image. This illustration reveals that high-resolution images can substantially differ from low-resolution ones, both in terms of color and texture. This suggests that subsequent conditional blocks are capable of significant modification in the background or foreground.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Importance of FS metrics: comparison against unconditional training. We also include results from training an unconditional StyleGAN2-ADA model [17] in Table 1, referred to as StyleGAN2-ADA UnCond. The unconditional model generates samples that follow the training distribution, which is mainly dominated by head classes. This bias favors unconditional learning in terms of the FID/KID metrics, which do not consider the skewness in the data distribution. Nevertheless, it can be observed the UnCond. baselines significantly perform worse in terms of few-shot metrics FID-FS/KID-FS. This suggests solely relying on FID/KID can be misleading. We strongly recommend including few-shot metrics, i.e., FID-FS and KID-FS, when evaluating cGANs on long-tailed datasets.", + "bbox": [ + 75, + 229, + 472, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CIFAR10-LT In this benchmark, our method is compared with various baselines with imbalanced ratios $\\rho$ of 50 and 100. This indicates the least frequent tail class includes only 100 and 50 training instances, respectively. Table. 3 quantitatively assesses the quality of the samples generated by our method and the baselines. All the baselines use the StyleGAN2-ADA [17] architecture. Our proposed method consistently outperforms the baselines across all metrics. This demonstrates that beyond improving generative learning across all classes (FID & KID), UTLO can also notably improve the learning from tail classes (FID-FS & KID-FS).", + "bbox": [ + 75, + 441, + 472, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Imbalance Ratio $\\rho$ . Not surprisingly, all methods demonstrate improved performance with lower values of $\\rho$ , as seen in Table. 3. Further, we notice the benefits of knowledge sharing via UTLO become more pronounced as the imbalance in the dataset increases, i.e., learning tail classes becomes more challenging. It is also worth mentioning that the FID-FS and KID-FS metrics are critical to represent the quality of the few-shot classes as those are under-represented in the dataset-wide metrics that as", + "bbox": [ + 75, + 611, + 472, + 747 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3943075715ef5bc49e78c3680a9f262e33371ff05544fb0c5c983e9f3116dd1a.jpg", + "image_caption": [ + "Figure 4. Different class-conditional images generated given the same unconditional low-resolution images (left-most column)." + ], + "image_footnote": [], + "bbox": [ + 81, + 762, + 465, + 875 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1fc6ccf5edfaeb26f45dace050c479aec7898a13728dec981d3ee90010bc7ef6.jpg", + "table_caption": [ + "Table 3. Quantitative comparison of the generated images from UTLO against baselines on CIFAR10-LT dataset with different imbalance ratios $\\rho$ . UTLO shows substantial improvements on few-shot (FS) metrics compared to StyleGAN2-ADA [17] while outperforming existing methods including NoisyTwins [37]." + ], + "table_footnote": [], + "table_body": "
ρMethodsFID ↓FID-FS ↓KID ↓KID-FS ↓ ×1000
50StyleGAN2-ADA [17]6.521.42.49.0
+ Transitional [42]9.417.74.78.6
+ GSR [36]6.421.32.38.1
+ NoisyTwins [37]6.212.22.45.0
+ UTLO (ours)6.111.82.44.8
100StyleGAN2-ADA [17]9.024.24.09.7
+ Transitional [42]11.320.65.49.2
+ GSR [36]8.424.33.911.8
+ NoisyTwins [37]7.114.12.95.9
+ UTLO (Ours)6.813.42.85.4
", + "bbox": [ + 501, + 165, + 890, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sume balanced sampling. For instance, at $\\rho = 50$ , the FID and KID metrics obtained from UTLO show little changes, while FID-FS and KID-FS exhibit improvements close to $50\\%$ in comparison to GSR.", + "bbox": [ + 496, + 320, + 893, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Knowledge-sharing at Low Resolutions. As explained in Sec. 3, our proposed generator builds on top of low-resolution (e.g., $8 \\times 8$ ) unconditionally trained images which are subsequently used to generate conditional images at higher resolutions (e.g., $32 \\times 32$ ). Several low-resolution images are shown in Fig. 6 along with their high-resolution conditional images generated for the head and tail classes of the dataset. It can be seen that conditional images generated from the tail classes evidently share certain features from the head classes. This demonstrates that knowledge sharing at low resolutions is an effective approach for infusing information from head classes to tail ones. A quantitative analysis of knowledge-sharing is presented in Supp. D. For additional qualitative examples see Supp. H.", + "bbox": [ + 496, + 383, + 892, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CIFAR100-LT In addition to CIFAR10-LT, we evaluate our method and various baselines on the more challenging CIFAR100-LT dataset, wherein the tail classes include as few as five training instances. The quantitative results, shown in Table. 4, reveal that our method outperforms the baselines in a benchmark where a large number of classes with high diversity are present. In Fig. 2, we demonstrate the FID-FS curve during the course of the training along with the generated samples from the tail classes of the CIFAR100-LT dataset. This shows the effectiveness of our proposed method in addressing mode collapse while regularization methods diverge and require early stopping. Further visual comparisons of our proposed method against the baselines for CIFAR100-LT can be found in Supp. H.", + "bbox": [ + 496, + 595, + 893, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LSUN5-LT We include LSUN5-LT as a higher resolution $(128 \\times 128)$ and highly-imbalanced $(\\rho = 1000)$ benchmark. In addition, we use Projected GAN with FastGAN[26] generator and DA [40]) as the baseline model. This is a different generator and discriminator design compared to StyleGAN2 along with a different aug-", + "bbox": [ + 496, + 809, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7586", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/53b872f9bca3a0fb6d14f9a056b4e5f8007f9be82edcc09e257c54b1baf3a5a2.jpg", + "image_caption": [ + "Figure 5. Generated images from LSUN5-LT dataset. Despite only 50 training instances for the tail class kitchen, the proposed UTLO framework produces diverse, high-fidelity images." + ], + "image_footnote": [], + "bbox": [ + 158, + 88, + 812, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7713869e15067a4a9545890df6157f99a4a8b0f945bda653bd9787fd68d4f8b.jpg", + "image_caption": [ + "Figure 6. Knowledge sharing from head to tail classes in CIFAR10-LT dataset $(\\rho = 100)$ using UTLO. The conditional images generated from the head (middle columns) and tail (right columns) classes share and are built on top of the same low-resolution (unconditional) images (left columns). Low-resolution images $(8 \\times 8)$ are upsampled to that of CIFAR10-LT $(32 \\times 32)$ for improved visualization." + ], + "image_footnote": [], + "bbox": [ + 78, + 308, + 467, + 406 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/46c11c06eaf4c3be3653ed8d917e5277b9dfff7d259b00082bc951b8f6c51796.jpg", + "table_caption": [ + "Table 4. Comparison of our proposed UTLO method with StyleGAN2-ADA baselines on the CIFAR100-LT dataset." + ], + "table_footnote": [], + "table_body": "
MethodsFID ↓FID-FS ↓KID ↓KID-FS ↓
×1000
StyleGAN2-ADA [17]10.824.95.19.3
+ Transitional [42]10.623.74.28.5
+ GSR[36]11.125.05.08.2
+ NoisyTwins [37]10.122.55.07.9
+ UTLO (Ours)9.921.84.67.5
", + "bbox": [ + 76, + 539, + 467, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mentation method. In conditional FastGAN implementation, the class-specific information is injected using class-conditional batch normalization, differing from the style mapping network utilized in StyleGAN2. UTLO can be readily adapted to this design by using standard batch normalization layers at low resolutions and utilizing class-conditional batch normalization only at higher resolutions. However, it is noteworthy that methods such as [37] are restricted to StyleGANs and cannot be applied in this context.", + "bbox": [ + 75, + 641, + 468, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table. 5 compares the results obtained from our method against the baseline and with the addition of GSR [36]. Our method consistently surpasses the baselines by around $30\\%$ across all metrics. As previously shown for the CIFAR10-LT (Table 3), UTLO not only improved on the tail classes but also significantly improved the dataset-wide FID/KID due to the high imbalance ratio $(\\rho = 1000)$ .", + "bbox": [ + 75, + 777, + 470, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative results in Fig. 5 contrast the generated", + "bbox": [ + 96, + 885, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aee78efda9f27f54a9388f3d2f85f13203c65346c21f24f22de75b9b3ba58564.jpg", + "table_caption": [ + "Table 5. Evaluating the quality of generated images from the proposed method and comparing against baselines on LSUN5-LT." + ], + "table_footnote": [], + "table_body": "
MethodsFID ↓FID-FS ↓KID ↓KID-FS ↓ ×1000
PGAN (FastGAN)+DA [40]15.060.24.652.7
+ GSR [36]15.763.75.758.0
+ UTLO (Ours)10.943.63.535.3
", + "bbox": [ + 503, + 342, + 888, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "images from our method against the baselines across all classes. For the kitchen class which only includes 50 training examples, the baselines struggle to effectively learn from the tail class whereas UTLO succeeds in generating diverse and high-quality images. More visual examples from our method and baselines are presented in Supp. H.", + "bbox": [ + 496, + 417, + 890, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to space constraints, additional results and analysis are provided in the supplementary material. e.g., Analysis of the role of weighted sampling (Supp. B.), The distinction of training class-conditional GANs in the Long-tailed setup vs. Limited-data setup (Supp. E.), and additional evaluation on Flowers-LT and iNaturalist2019 datasets (Supp. G).", + "bbox": [ + 496, + 510, + 892, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 611, + 619, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed UTLO, a novel framework designed to improve the training of cGANs on long-tailed data. Inspired by the observation that head and tail classes often have more similarities at lower resolutions, our method facilitates knowledge sharing from head to tail classes using unconditional training at lower resolutions. The proposed method enriches the limited training distribution of the tail classes and effectively addresses mode collapse, leading to significant improvement in image generation for tail classes. We have also introduced the FID-FS/KID-FS metric, an adaptation of widely-used GAN metrics, specifically tailored for tail classes. We hope our findings are useful for future work in long-tail learning.", + "bbox": [ + 496, + 635, + 890, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 843, + 668, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is partially supported by the National Science Foundation (NSF) under grant 1751412 and 1927564.", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7587", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In International Conference on Learning Representations, 2018. 6", + "[2] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. 1, 2", + "[3] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3", + "[4] Jiequan Cui, Shu Liu, Zhuotao Tian, Zhisheng Zhong, and Jiaya Jia. Result: Residual learning for long-tailed recognition. *IEEE transactions on pattern analysis and machine intelligence*, 2021. 12", + "[5] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge J. Belongie. Class-balanced loss based on effective number of samples. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 9268-9277. Computer Vision Foundation / IEEE, 2019. 12", + "[6] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 8780-8794, 2021. 2", + "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, A aron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2014. 1, 2", + "[8] Hao Guo and Song Wang. Long-tailed multi-label visual recognition by collaborative training on uniform and rebalanced samplings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15089-15098, 2021. 3", + "[9] Hui Han, Wen-Yuan Wang, and Bing-Huan Mao. Borderline-smote: a new over-sampling method in imbalanced data sets learning. In International conference on intelligent computing, pages 878-887. Springer, 2005. 3", + "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[11] Grant Van Horn and Oisin Mac Aodha. The inaturalist 2019 competition dataset. https://www.kaggle.com/c/inaturalist-2019-fgvc6.5, 12", + "[12] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 1" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 3", + "[14] Bingyi Kang, Yu Li, Sai Nan Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In International Conference on Learning Representations, 2021. 3", + "[15] Tejan Karmali, Rishubh Parihar, Susmit Agrawal, Harsh Rangwani, Varun Jampani, Maneesh Kumar Singh, and R. Venkatesh Babu. Hierarchical semantic regularization of latent spaces in stylegans. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 443-459. Springer, 2022. 2", + "[16] Animesh Karnewar and Oliver Wang. MSG-GAN: multiscale gradients for generative adversarial networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 7796-7805. Computer Vision Foundation / IEEE, 2020. 4", + "[17] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. 1, 2, 4, 6, 7, 8, 12, 13, 16, 17, 18", + "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 8107-8116. Computer Vision Foundation / IEEE, 2020. 2, 4, 18", + "[19] Chaerin Kong, Jeesoo Kim, Donghoon Han, and Nojun Kwak. Few-shot image generation with mixup-based distance learning. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 563-580. Springer, 2022. 2", + "[20] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. preprint, 2009. 5, 12", + "[21] Nupur Kumari, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Ensembling off-the-shelf models for gan training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10651-10662, 2022. 2", + "[22] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 1", + "[23] Bingchen Liu, Yizhe Zhu, Kunpeng Song, and A. Elgammal. Towards faster and stabilized gan training for high-fidelity few-shot image synthesis. iclr, 2021. 2, 4, 18", + "[24] Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, and Wenhui Li. Deep representation learning on long-tailed data:" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7588", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "A learnable embedding augmentation perspective. Computer Vision And Pattern Recognition, 2020. 3", + "[25] Shudong Liu and Ke Zhang. Under-sampling and feature selection algorithms for s2smlp. IEEE Access, 8:191803-191814, 2020. 3", + "[26] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 7", + "[27] Mario Lucic, Karol Kurach, Marcin Michalski, Sylvain Gelly, and Olivier Bousquet. Are gans created equal? a large-scale study. Advances in neural information processing systems, 31, 2018. 3", + "[28] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In International Conference on Learning Representations, 2021. 3", + "[29] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv: Arxiv-1411.1784, 2014. 1, 2", + "[30] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. ICLR, 2018. 4", + "[31] S. S. Mullick, Shounak Datta, and Swagatam Das. Generative adversarial minority oversampling. IEEE International Conference On Computer Vision, 2019. 3", + "[32] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5, 12", + "[33] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, pages 2642-2651. PMLR, 2017. 2", + "[34] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A. Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10743-10752, 2021. 2", + "[35] Harsh Rangwani, Konda Reddy Mopuri, and R. Venkatesh Babu. Class balancing gan with a classifier in the loop. In Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence, pages 1618-1627. PMLR, 2021. 3", + "[36] Harsh Rangwani, Naman Jaswani, Tejan Karmali, Varun Jampani, and R. Venkatesh Babu. Improving gans for long-tailed data through group spectral regularization. European Conference On Computer Vision, 2022. 2, 3, 6, 7, 8, 12, 13, 16, 18, 19", + "[37] Harsh Rangwani, Lavish Bansal, Kartik Sharma, Tejan Karmali, Varun Jampani, and R Venkatesh Babu. Noisytwins: Class-consistent and diverse image generation through stylegans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5987-5996, 2023. 2, 3, 6, 7, 8, 18, 19" + ], + "bbox": [ + 78, + 92, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Jiawei Ren, Cunjun Yu, shunan sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and hongsheng Li. Balanced meta-softmax for long-tailed visual recognition. In Advances in Neural Information Processing Systems, pages 4175–4186. Curran Associates, Inc., 2020. 3", + "[39] Shibani Santurkar, Ludwig Schmidt, and A. Madry. A classification-based study of covariate shift in gan distributions. International Conference On Machine Learning, 2017. 12", + "[40] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster. Advances in Neural Information Processing Systems, 34:17480-17492, 2021. 6, 7, 8, 18", + "[41] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleganxI: Scaling stylegan to large diverse datasets. In SIGGRAPH '22: Special Interest Group on Computer Graphics and Interactive Techniques Conference, Vancouver, BC, Canada, August 7 - 11, 2022, pages 49:1-49:10. ACM, 2022. 1, 2, 15, 16", + "[42] Mohamad Shahbazi, Martin Danelljan, Danda Pani Paudel, and Luc Van Gool. Collapse by conditioning: Training class-conditional GANs with limited data. In International Conference on Learning Representations, 2022. 1, 3, 6, 7, 8, 16, 18", + "[43] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 1", + "[44] Zhangzhang Si and Song-Chun Zhu. Learning hybrid image templates (hit) by information projection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(7):1354-1367, 2012. 5, 12", + "[45] Kaihua Tang, Jianqiang Huang, and Hanwang Zhang. Long-tailed classification by keeping the good and removing the bad momentum causal effect. In Advances in Neural Information Processing Systems, pages 1513-1524. Curran Associates, Inc., 2020. 3", + "[46] Ngoc-Trung Tran, Viet-Hung Tran, Ngoc-Bao Nguyen, Trung-Kien Nguyen, and Ngai-Man Cheung. On data augmentation for gan training. IEEE Transactions on Image Processing, 30:1882-1897, 2021. 2", + "[47] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7921-7931, 2021. 2", + "[48] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X. Yu. Long-tailed recognition by routing diverse distribution-aware experts. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3", + "[49] Lu Yang, He Jiang, Qing Song, and Jun Guo. A survey on long-tailed visual recognition. International Journal of Computer Vision, 130:1837-1872, 2022. 2, 3, 5, 6, 12", + "[50] Baofeng Yao and Lei Wang. An improved under-sampling imbalanced classification algorithm. In 2021 13th International Conference on Measuring Technology and Mechatronics Automation (ICMTMA), pages 775-779. IEEE, 2021. 3", + "[51] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset" + ], + "bbox": [ + 501, + 92, + 892, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7589", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 5, 12", + "[52] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 1", + "[53] Han Zhang, Zizhao Zhang, Augustus Odena, and Honglak Lee. Consistency regularization for generative adversarial networks. In International Conference on Learning Representations, 2019. 2, 12", + "[54] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. IEEE/cvf Conference On Computer Vision And Pattern Recognition, 2018. 16", + "[55] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. arXiv preprint arXiv: Arxiv-2110.04596, 2021. 1, 3", + "[56] Zizhao Zhang and Tomas Pfister. Learning fast sample re-weighting without reward data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 725-734, 2021. 3", + "[57] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. Advances in Neural Information Processing Systems, 33:7559-7570, 2020. 2, 3, 6, 12, 16, 18", + "[58] Zhengli Zhao, Zizhao Zhang, Ting Chen, Sameer Singh, and Han Zhang. Image augmentations for gan training. arXiv preprint arXiv: 2006.02595, 2020. 2", + "[59] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 16489-16498. Computer Vision Foundation / IEEE, 2021. 3" + ], + "bbox": [ + 78, + 92, + 470, + 585 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7590", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/301b4dc8-6f1d-458c-a509-0fe38145fc73_model.json b/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/301b4dc8-6f1d-458c-a509-0fe38145fc73_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b4bb878cac25ed9dce2ab460cc3a43b51780af94 --- /dev/null +++ b/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/301b4dc8-6f1d-458c-a509-0fe38145fc73_model.json @@ -0,0 +1,2268 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.809, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.142, + 0.131, + 0.83, + 0.175 + ], + "angle": 0, + "content": "Taming the Tail in Class-Conditional GANs: Knowledge Sharing via Unconditional Training at Lower Resolutions" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.203, + 0.838, + 0.275 + ], + "angle": 0, + "content": "Saeed Khorram1, Mingqi Jiang1, Mohamad Shahbazi2, Mohamad H. Danesh3, Li Fuxin1 \n1Oregon State University 2 ETH Zürich 3 McGill University \n{khorrams, jiangmi, lif}@oregonstate.edu \nmshahbazi@vision.ee.ethz.ch, mohamad.danesh@mail.mcgill.ca" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.31, + 0.314, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.343, + 0.474, + 0.66 + ], + "angle": 0, + "content": "Despite extensive research on training generative adversarial networks (GANs) with limited training data, learning to generate images from long-tailed training distributions remains fairly unexplored. In the presence of imbalanced multi-class training data, GANs tend to favor classes with more samples, leading to the generation of low quality and less diverse samples in tail classes. In this study, we aim to improve the training of class-conditional GANs with long-tailed data. We propose a straightforward yet effective method for knowledge sharing, allowing tail classes to borrow from the rich information from classes with more abundant training data. More concretely, we propose modifications to existing class-conditional GAN architectures to ensure that the lower-resolution layers of the generator are trained entirely unconditionally while reserving class-conditional generation for the higher-resolution layers. Experiments on several long-tail benchmarks and GAN architectures demonstrate a significant improvement over existing methods in both the diversity and fidelity of the generated images. The code is available at https://github.com/khorrams/utlo." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.676, + 0.208, + 0.691 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.703, + 0.469, + 0.839 + ], + "angle": 0, + "content": "In the past few years, research on Generative Adversarial Networks (GANs) [7] has led to remarkable advances in generating realistic images [2, 17, 41]. Conditional GANs [29] (cGANs) have garnered particular attention due to their ability to accept user inputs which can additionally guide the generation process. They enable a wide range of applications such as class-conditional image generation [42], image manipulation [43], image-to-image translation [12], super-resolution [22] and text-to-image synthesis [52]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Despite the advances, past research on cGANs has focused primarily on learning from balanced data. In real-world scenarios, however, data often follows a power-law distribution, with a few classes dominating most of the" + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.311, + 0.826, + 0.326 + ], + "angle": 0, + "content": "All (2) Training Images in Tail Classes" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.327, + 0.875, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.632, + 0.401, + 0.762, + 0.414 + ], + "angle": 0, + "content": "Baseline (FID: 9.8)" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.415, + 0.875, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.559, + 0.782, + 0.572 + ], + "angle": 0, + "content": "Baseline + Ours (FID: 6.6)" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.572, + 0.875, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.714, + 0.892, + 0.77 + ], + "angle": 0, + "content": "Figure 1. Generating images from rare tail classes in the Flowers-LT with only two training images. Our proposed approach allows for a more diverse set of features such as backgrounds, colors, poses, and object layouts to be infused into the tail classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "training data. This is referred to as the \"long-tail (LT) problem\" [55], where most of the training data come from a few classes (referred to as the \"head\" of the distribution). In contrast, data from a large number of classes rarely occur (referred to as the \"tail\" of the distribution). This imbalance in data poses a challenge to the effective learning of tail classes, as standard machine learning algorithms tend" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7580" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.461, + 0.107 + ], + "angle": 0, + "content": "to favor densely sampled regions of the input distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.109, + 0.47, + 0.35 + ], + "angle": 0, + "content": "Although there has been a surge of interest in the long-tail problem in recent years, the focus has been primarily on recognition tasks [49]. Generative learning on the long-tail data, particularly cGANs which is the focus of this work, remains underexplored. Recently, [36, 37] identified mode collapse on tail classes as a result of the spectral norm explosion in class-conditional BatchNorms and the collapse of the latents in the StyleGANs \\(\\mathcal{W}\\) space. While techniques such as regularization, normalization, and/or classbalancing techniques [49] might improve the performance of long-tail generative models to some extent, their effectiveness is limited. The primary challenge in the long-tail problem is the under-representation of tail classes in the training data, e.g., in Fig. 1 each tail class contains merely 2 training images, resulting in insufficient observation of the tail classes by the learning algorithm." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.471, + 0.577 + ], + "angle": 0, + "content": "To address this challenge, we propose to leverage the abundant information in the head classes, with the aim to learn and infuse knowledge from the head classes into the tail ones, thereby enriching the training distribution for the latter. However, it is very difficult to explicitly disentangle the class-specific information pertaining to head classes from class-independent information shared between head and tail classes. We make a key observation that the similarity between head and tail instances is often higher at lower resolutions: discriminative features such as shape or specific texture are usually unveiled at higher resolutions. Conversely, information from the lower resolutions tends to be more class-independent such as background, configuration, or direction of objects (Fig. 6), thus can be shared between the head and tail classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.47, + 0.732 + ], + "angle": 0, + "content": "Building upon this observation, we propose Unconditional Training at LOWER resolution (UTLO), a novel approach for training cGANs in the long-tail setup. In addition to the standard conditional GAN objective, UTLO trains the intermediate low-resolution output of the generator with an unconditional GAN objective. The unconditional GAN objective encourages the learning of low-resolution features common to both the head and tail classes, which are infused into the subsequent layers of the generator, particularly benefiting the under-represented tail classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.884 + ], + "angle": 0, + "content": "Through extensive experiments and analysis on several long-tailed datasets, we demonstrate that using UTLO to combine high-resolution conditional and low-resolution unconditional training effectively facilitates knowledge sharing between head and tail classes, thereby improving the overall generative modeling of cGANs in the long-tail setup. Due to the strong class imbalance in the long-tailed data, naive usage of existing GAN metrics can be misleading. To mitigate this issue, we propose a few practices to adapt commonly-used GAN metrics to the long-tail setup." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.39, + 0.901 + ], + "angle": 0, + "content": "Below, we highlight our main contributions:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "- We propose UTLO, a novel knowledge-sharing framework tailored for training cGANs in the long-tailed setup. UTLO allows infusing information from the head classes to the tail classes via an additional unconditional objective applied to the low-resolution part of the cGAN generator. To the best of our knowledge, this work is the first to demonstrate that not all layers in a cGAN need to receive class-conditional information, i.e., a cGAN featuring a partially unconditional generator." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.228, + 0.892, + 0.288 + ], + "angle": 0, + "content": "- We present a set of practices and metrics designed to adapt the commonly-used GAN evaluation metrics for long-tail setups, enabling a more precise evaluation of the image generation quality." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.289, + 0.892, + 0.363 + ], + "angle": 0, + "content": "- Through extensive experiments across multiple benchmarks and architectures, we validate the effectiveness of our proposed method in improving the training of cGANs in the long-tail setup, achieving state-of-the-art results across several long-tail datasets." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.377, + 0.642, + 0.393 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Conditional Generative Adversarial Networks The initial work on conditional image generation using GANs [7] was presented by [29], in which the class conditions were concatenated to the inputs of the generator and discriminator networks. The AC-GAN [33] proposed the use of auxiliary classification in the discriminator. [2] introduced BigGAN and set a milestone in large-scale and high-resolution conditional image generation. The authors of StyleGAN [17, 18] first map the input noise and class condition to a latent style space, which is then passed to the multiple layers of the generator for image synthesis. More recently, an extension of the StyleGAN called StyleGAN-XL [41] has become the state-of-the-art in conditional image generation on several datasets and outperforms more complex and time-consuming approaches such as diffusion models [6]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.825 + ], + "angle": 0, + "content": "GAN Regularization under Limited Data Training GANs under limited data is challenging as the discriminator can memorize the training samples, resulting in the collapse of the training or quality degradation of the generated images. Recently, data augmentation and regularization techniques have been incorporated into GAN training as a means to mitigate this problem [15, 17, 46, 53, 57, 58]. Additionally, [47] introduced a regularization term to the GAN objective, which tracks the predictions of the discriminator for real and generated images using separate moving averages. [21] employed off-the-shelf vision models in an ensemble of discriminators, demonstrating improved performance in both limited-data and large-scale GAN training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the context of few-shot image generation, FastGAN [23] proposed a lightweight architecture with a self-supervised discriminator. [34] introduced cross-domain distance consistency in order to transfer diversity from a source domain to a target domain. Subsequently, [19] proposed" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7581" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "a latent-mixup strategy to smooth the latent space through controlled latent interpolation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.125, + 0.471, + 0.351 + ], + "angle": 0, + "content": "Recently, [42, 57] have observed that class-conditioning can cause mode collapse in the limited data regime. Their proposed work learns from limited but balanced/unlabeled data, which is different from our setup: heavily imbalanced long-tail data. [42] proposed Transitional-cGAN, a training strategy that starts with unconditional GAN objective and injects class-conditional information during a transition period. This approach was found to be effective in mitigating mode-collapse in the limited data regime. Our main idea of separating pathways between low-resolution and high-resolution is novel w.r.t. [42]: to promote knowledge sharing, we modify the cGAN architecture so that the lower-resolution part of the generator is entirely unconditional while only the higher-resolution part is conditional and received class-conditional information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.471, + 0.581 + ], + "angle": 0, + "content": "Long-tail Recognition To address the long-tail recognition problem [49, 55], previous research has primarily focused on techniques such as class re-balancing [3, 28, 38], learning algorithm and model design [13, 14, 59], and information augmentation [24, 45, 48]. More related to our work, re-sampling [8, 56] has been widely used to handle class imbalance problems. This involves balancing the number of samples used during training through over-sampling and/or under-sampling. Oversampling increases the frequency of tail classes [9, 31] while under-sampling reduces the imbalance by reducing the frequency of head class instances [25, 50]. Note that our research diverges from prior studies on long-tail recognition by addressing the more challenging task of generative modeling of the long-tail data, as opposed to the conventional long-tail classification task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Training GANs on Long-tail Data Recent years have seen a growing interest in learning from long-tail data. However, most focus has been directed toward recognition tasks, with a limited number of studies addressing the development of generative models for long-tail data. GAMO [31] uses adversarial training to over-sample from minority classes. [35] introduced class-balancing regularization to the GAN objective, which leverages the predictions of a pretrained classifier to improve the generation of underrepresented classes in an unconditional setting. However, this method trains an unconditional GAN and is restricted to having access to a pre-trained classifier on the long-tail data to guide the training. The most closely related work to ours is the Group Spectral Regularization (GSR) [36] and NoisyTwins [37] regularizations. The authors identified mode collapse on tail classes as a result of the spectral norm explosion in class-conditional BatchNorms and the collapse of the latents in the StyleGANs \\(\\mathcal{W}\\) space, respectively. In contrast, we present a novel approach for sharing knowledge between head and tail classes to enhance long-tail learning, which is orthogonal to regularization techniques. In addition, while" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "[37] is restricted to StyleGANs, our framework can be extended to different GAN architectures." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.136, + 0.744, + 0.154 + ], + "angle": 0, + "content": "3. Taming the Tail in cGANs" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.161, + 0.892, + 0.192 + ], + "angle": 0, + "content": "3.1. Background: Conditional Generative Adversarial Networks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.892, + 0.381 + ], + "angle": 0, + "content": "Conditional Generative Adversarial Networks (cGANs) are suitable for many applications due to providing user control over the generated samples at inference time. This is particularly helpful in the long-tail setup, where it is desirable to explicitly generate samples from rare (tail) classes. In contrast, unconditional GANs are likely to generate samples that track the training distribution, resulting in most samples coming from common (head) classes. To train a cGAN over a dataset containing \\( n \\) instances \\( \\mathbb{X} = \\{\\pmb{x}_1,\\dots ,\\pmb{x}_n\\} \\) and their corresponding labels \\( \\mathbb{Y} = \\{y_{1},\\ldots ,y_{n}\\} \\), we can formulate the adversarial training objective that alternatively optimizes the following loss terms," + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.392, + 0.88, + 0.411 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} ^ {D} = \\mathbb {E} _ {\\boldsymbol {x}, y} [ f _ {D} (- D (\\boldsymbol {x} | y)) ] + \\mathbb {E} _ {\\boldsymbol {z}, y} [ f _ {D} (D (G (\\boldsymbol {z}, y))) ]\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.414, + 0.892, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} ^ {G} = \\mathbb {E} _ {\\boldsymbol {z}, y} [ f _ {G} (- D (G (\\boldsymbol {z}, y))) ] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.595 + ], + "angle": 0, + "content": "where \\(G\\) and \\(D\\) are the generator and discriminator networks with their corresponding loss functions \\(f_{G}\\) and \\(f_{D}\\) [27]; \\(\\pmb{x}|y\\sim p_{\\mathrm{data}}(\\pmb{x}|y)\\) is real data drawn from class \\(y\\). The generator \\(G\\) conditions on both \\(z\\sim p_z(z)\\), a random noise vector from a prior distribution over latent space, and \\(y\\), the class conditioning vector. This adversarial training objective encourages the cGAN to reach an equilibrium between the generator and the discriminator, resulting in the generator producing realistic samples indistinguishable from real data by the discriminator." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.605, + 0.88, + 0.621 + ], + "angle": 0, + "content": "3.2. Caveat of Training cGANs on Long-tail Data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Training GANs on long-tail data is challenging due to the inherent difficulty in modeling the rare examples that comprise the tail of the data distribution. This skewness in the data distribution hinders balanced learning across classes. In particular, the discriminator does not see enough examples from the tail of the distribution during training, which can lead to poor discriminative signals for the generator. As a result, the generator learns to generate a small subset of the possible outputs that fools the inadequately learned discriminator – exhibiting a classic “mode collapse” scenario." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Although regularizations introduced in [36, 37] can help mitigate the collapse of learning in tail classes that contain a sufficiently large number of samples, we still observe mode collapse occurring when a very limited number of samples are in tail classes, which is inherently challenging. This can be seen in Fig. 2 for tail classes of CIFAR100-LT with as few as 5 training instances. Although early-stopping can be adopted to obtain reasonable performance, the key in boost-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7582" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.097, + 0.46, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.314, + 0.177, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.315, + 0.275, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.315, + 0.371, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.315, + 0.469, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.398, + 0.47, + 0.455 + ], + "angle": 0, + "content": "Figure 2. Convergence of different methods on CIFAR100-LT \\((\\rho = 100)\\), where tail classes have as few as 5 training examples. Incorporating our framework into the baseline alleviates overfitting as a result of knowledge sharing from head to rare tail classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.515 + ], + "angle": 0, + "content": "ing the performance would be an approach to infuse additional diversities into the tail classes rather than only relying on techniques such as regularization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.528, + 0.47, + 0.559 + ], + "angle": 0, + "content": "3.3. Knowledge Sharing via Unconditional Training at Lower Resolutions (UTLO)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.734 + ], + "angle": 0, + "content": "We devise a novel generative adversarial network (GAN) training scheme for long-tail data that aims to share knowledge from predominant head classes to rare tail classes, assuming some of the tail and head classes are at least coarsely similar. We propose a framework that uses unconditional learning in low-resolution images / features at an intermediate layer of the generator using the unconditional GAN objective. This helps with learning universal features from both head and tail classes. The subsequent layers, responsible for introducing finer details at higher resolutions, are trained using a class-conditional GAN objective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In our proposed framework, we perform conditional training on the high-resolution image output from the final layer of the generator network. This is built upon the features learned at lower layers through the use of an unconditional objective, inheriting features primarily from head classes. The conditioning on the class labels gives control at the inference time to explicitly generate images from tail classes, meeting our design desideratum. The combination of unconditional and conditional objectives using our proposed method enables knowledge sharing, which, in turn, improves the quality of GAN training on long-tail data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.259 + ], + "angle": 0, + "content": "Our framework is general and applicable to many GAN architectures. In this section, we illustrate the application of our framework on StyleGAN2 with adaptive data augmentation (ADA) [17], a state-of-the-art and solid baseline for training GANs, particularly in the limited data regime. In the following, we demonstrate the necessary modifications to the architecture of the generator and discriminator networks. Although we show the necessary modifications for StyleGAN2-ADA, our method can be easily extended to other GAN architectures such as FastGAN [23], which is also used in our experiments (see Section 4)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.274, + 0.73, + 0.289 + ], + "angle": 0, + "content": "3.3.1 Modifying the Generator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.892, + 0.478 + ], + "angle": 0, + "content": "In the generator design of the StyleGAN2-ADA [17, 18], the class-conditional information, combined with the latent vector, is first embedded in the style space \\( w \\) using the style-mapping network \\( G_{\\mathrm{map}} \\). The style vectors are then broadcasted across the layers of the synthesis network \\( G_{\\mathrm{syn}} \\) in order to generate diverse images. To meet our design desideratum, we dissect the synthesis network into two subnetworks \\( G_{\\mathrm{syn}} = G_h \\circ G_l \\) where \\( G_l \\) represents the earlier part of the network that produces intermediate features and/or images at a low resolution \\( L \\times L \\). \\( G_h \\) on the other hand, represents the latter part of \\( G_{\\mathrm{syn}} \\) that generates the output image at the high resolution \\( H \\times H \\) where \\( H > L \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.893, + 0.63 + ], + "angle": 0, + "content": "To block the flow of class-conditional information to the lower layers of \\( G_{\\mathrm{syn}} \\), we generate separate \\( w \\) vectors, one containing the class-conditional embeddings, referred to as \\( w_{z,y} \\), while for the other one, referred to as \\( w_{z} \\), class information was set to zero. Note that both vectors share the same latent variable \\( z \\) and mapping network \\( G_{\\mathrm{map}} \\) parameters, and the only difference is the presence of class-conditional information. The lower layers of the synthesis network \\( G_{l} \\) are conditioned on \\( w_{z} \\) while the subsequent layers at higher resolutions receive \\( w_{z,y} \\) (see Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.894, + 0.825 + ], + "angle": 0, + "content": "Generators typically follow a network design that gradually increases the resolution of intermediate features and/or images as the network progresses. This allows us to select a desired low-resolution, such as \\(8 \\times 8\\) or \\(16 \\times 16\\), during training. Recent generator designs often incorporate skip connections [16] or residual connections [30] to improve gradient flow. The generator of StyleGAN2-ADA uses skip connections, which explicitly generates intermediate RGB images at each layer. We take this low-resolution image as the input to the discriminator. Note that in the case of a generator design that uses residual connections, a \\(1 \\times 1\\) convolutional layer can be incorporated to convert the residual channels to RGB channels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "In a forward pass through the generator, images at low and high resolutions can be generated simultaneously without additional overhead. By inputting a latent code \\( z \\sim p_{z}(z) \\) and a target label \\( y \\), we can obtain an unconditional synthetic image \\( \\hat{x}_l \\in \\mathbb{R}^{3 \\times L \\times L} \\) along with a conditional" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7583" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.182, + 0.088, + 0.795, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.259, + 0.895, + 0.373 + ], + "angle": 0, + "content": "Figure 3. The proposed framework, UTLO, illustrated for the StyleGAN2-ADA architecture. Low and high resolution image pathways are used for unconditional and class-conditional objectives, respectively. \\( z \\) is the input latent code, \\( y \\) indicates the class embeddings, and \\( c \\) is a constant input. Separate style vectors, \\( w_{z} \\) (class-independent) and \\( w_{z,y} \\) (class-conditional), are generated using the same \\( z \\) and a shared style-mapping network \\( G_{\\mathrm{map}} \\) which are then passed to \\( G_{l} \\) and \\( G_{h} \\), respectively. The high-resolution generated image \\( \\hat{x} \\in \\mathbb{R}^{3\\times H\\times H} \\) is passed through the discriminator \\( D = D_{l} \\circ D_{h} \\) to calculate the conditional objective \\( \\mathcal{L}_c \\) while the low-resolution image \\( \\hat{x}_l \\in \\mathbb{R}^{3\\times L\\times L} \\) is passed only through \\( D_{l} \\) to calculate the unconditional objective \\( \\mathcal{L}_{uc} \\). The final objective \\( \\mathcal{L} \\) is the combination of the two. While \\( D_{l} \\) is shared, two separate prediction heads (FC layers) are used for unconditional and conditional objectives. The fromRGB is designed to increase the dimensionality of RGB channels to match the input channels of the \\( D_{l} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.472, + 0.445 + ], + "angle": 0, + "content": "high-resolution synthetic image \\(\\hat{\\pmb{x}}\\in \\mathbb{R}^{3\\times H\\times H}\\) at the output. While \\(\\hat{\\pmb{x}}\\) is trained to be from class \\(y\\), \\(\\hat{\\pmb{x}}_l\\) has no class-specific constraints. At inference time, the intermediate images can be discarded." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.454, + 0.333, + 0.469 + ], + "angle": 0, + "content": "3.3.2 Modifying the Discriminator" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.47, + 0.627 + ], + "angle": 0, + "content": "Opposite to the generator design, discriminators gradually reduce the resolution of the intermediate features as the network progresses. In order to pass images to an intermediate layer of the discriminator, we dissect it into two sub-networks. Formally, we define the discriminator as \\( D = D_{l} \\circ D_{h} \\) where \\( D_{h} \\) represents the earlier part of the original network that takes the images at training resolution \\( \\mathbb{R}^{3 \\times H \\times H} \\). On the other hand, \\( D_{l} \\) represents the latter part of the discriminator taking a lower resolution input \\( \\mathbb{R}^{C \\times L \\times L} \\) with \\( C \\) channels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.719 + ], + "angle": 0, + "content": "The default discriminator design in StyleGAN2-ADA uses residual connections. This does not allow direct passing of RGB inputs to the intermediate layers. To overcome this, we add a \\(1 \\times 1\\) convolutional layer, referred to as \"fromRGB\", that increases the channels of the RGB images to match \\(C\\), the number of input channels of \\(D_{l}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.856 + ], + "angle": 0, + "content": "More concretely, \\( D_{l} \\) is a shared network in both the unconditional low-resolution and class-conditional high-resolution image pathways (see Fig. 3). To obtain separate unconditional and conditional predictions, we use two different fully-connected (FC) layers at the final layer of \\( D_{l} \\). The output of the \\( D_{h} \\) and the low-res image \\( \\hat{\\boldsymbol{x}}_l \\) are passed through the \\( D_{l} \\) which are then used for calculating the class-conditional loss \\( \\mathcal{L}_c \\) and unconditional loss \\( \\mathcal{L}_{uc} \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Note, when passing real images to \\( D_{l} \\) for the unconditional loss, they are first downsized to the low-resolution \\( L \\times L \\) using bi-linear interpolation to be comparable with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.893, + 0.444 + ], + "angle": 0, + "content": "the ones generated from \\( G_{l} \\). They are then passed through the fromRGB layer and consequently \\( D_{l} \\). For the class-conditional loss, the original images are passed to \\( D \\) without any modification." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.448, + 0.727, + 0.464 + ], + "angle": 0, + "content": "3.3.3 Final Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.893, + 0.517 + ], + "angle": 0, + "content": "The final objective of our framework is simply a combination of the conditional and unconditional GAN losses, using a weighting parameter \\(\\lambda\\). To put it formally we have," + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.524, + 0.893, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {D} = \\mathcal {L} _ {c} ^ {D} + \\lambda \\cdot \\mathcal {L} _ {u c} ^ {D}, \\mathcal {L} ^ {G} = \\mathcal {L} _ {c} ^ {G} + \\lambda \\cdot \\mathcal {L} _ {u c} ^ {G} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.894, + 0.645 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_c^D\\) and \\(\\mathcal{L}_c^G\\) represent the class-conditional losses for the discriminator and generator, respectively (as outlined in Eq. 3.1), and \\(\\mathcal{L}_{uc}^D\\) and \\(\\mathcal{L}_{uc}^G\\) represent their unconditional counterparts, where \\(\\pmb{x}\\) is sampled from \\(p_{\\mathrm{data}}(\\pmb{x})\\) instead of \\(p_{\\mathrm{data}}(\\pmb{x}|y)\\). The overall framework of our proposed method is depicted in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.657, + 0.634, + 0.674 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.583, + 0.699 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.871 + ], + "angle": 0, + "content": "Datasets. We use 6 different long-tailed datasets in our experiments: CIFAR10 and CIFAR100[20], LSUN [51], Flowers [32], iNaturalist2019 [11], and AnimalFaces [44]. This selection is intended to encompass a wide spectrum of image domains, dataset sizes, resolutions, and imbalance ratios (denoted as \\(\\rho\\)). \\(\\rho\\) represents the ratio of the number of training samples between the classes with the most and the least number of examples. To ensure the tail classes stay few-shot, we keep the number of images in the smallest tail classes under 50 [49]. A detailed description of these datasets is provided in Supp. A." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Baselines. In order to demonstrate the effectiveness of our proposed method, we use various cGAN architectures cov" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7584" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "ering different generator and discriminator designs and data augmentation pipelines:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.122, + 0.47, + 0.181 + ], + "angle": 0, + "content": "- StyleGAN2 with Adaptive Data Augmentation (ADA) [17]. We also integrate Transitional training [42], Group Spectral Regularization (GSR) [36], and NoisyTwins [37] to this baseline." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.183, + 0.469, + 0.227 + ], + "angle": 0, + "content": "- Projected GAN (PGAN) [40]: we use projected discriminator with Differentiable Augmentation (DA) [57]. We also add GSR [36] to this baseline." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.122, + 0.47, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.469, + 0.259 + ], + "angle": 0, + "content": "For comprehensive details on the baselines, hyperparameters, and implementation please refer to Supp. F." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.264, + 0.442, + 0.28 + ], + "angle": 0, + "content": "4.2. Evaluation Metrics For Long-Tail Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.469, + 0.434 + ], + "angle": 0, + "content": "Evaluating the images generated by GANs on a long-tailed setup poses challenges, primarily due to the imbalanced data and access to a limited number of samples for the tail classes. For our experiments, we employ widely-used GAN metrics: Fréchet Inception Distance (FID) [10], and Kernel Inception Distance (KID) [1]. Following [10, 17], we report all metrics by computing statistics between \\(50k\\) generated images and all available training images. In the following, we present a set of practices for adapting the aforementioned metrics to the long-tail setup." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.435, + 0.469, + 0.614 + ], + "angle": 0, + "content": "First, when a larger and (more) balanced dataset following the long-tail training data distribution is available, it is used for metric calculation, e.g., full CIFAR10 before artificial imbalance. For naturally imbalanced datasets such as iNaturalist2019, we use the unmodified training set to calculate metrics. Note that when generating images for evaluation, we sample from the same distribution as the available real dataset, which may be imbalanced. While FID and KID indicate how well the real data distribution matches the generated images across all classes, we suggest additional metrics to mitigate the disproportionate influence of head classes in data statistic calculations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.616, + 0.469, + 0.826 + ], + "angle": 0, + "content": "In response to the varied number of training instances across classes in a long-tail setup, we propose evaluating metrics specifically on few-shot (FS) categories [49]. We calculate the FID and KID for the few-shot subset and refer to them as \"FID-FS\" and \"KID-FS\", respectively. This is tailored to evaluate the quality of generated samples in the tail classes. Contrasting with the standard FID/KID, we maintain an equal number of real images across all classes during our FID-FS/KID-FS calculation. This emphasizes learning quality on tail classes, irrespective of any imbalances that might be present in the few-shot subset. Reporting both standard and few-shot metrics provides a more comprehensive evaluation of long-tail learning, considering the performance of both head and tail classes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.172, + 0.847 + ], + "angle": 0, + "content": "4.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Below, we report the results obtained across different benchmarks and cGAN architectures. We pick the best model for reporting results as the one with the lowest FID" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.131 + ], + "angle": 0, + "content": "Table 1. Our proposed method, UTLO, outperforms the baselines in terms of quantitative image quality metrics on the AnimalFaces-LT dataset (64 × 64 resolution)." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.138, + 0.89, + 0.237 + ], + "angle": 0, + "content": "
MethodsFID ↓FID-FS ↓KID ↓KID-FS ↓ ×1000
StyleGAN2-ADA UnCond. [17]39.4104.117.327.6
StyleGAN2-ADA [17]51.487.124.735.9
+ Transitional [42]62.199.038.548.9
+ GSR[36]39.267.221.232.7
+ NoisyTwins [37]29.450.216.721.2
+ UTLO (Ours)26.248.412.619.6
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.248, + 0.892, + 0.293 + ], + "angle": 0, + "content": "FS from two independent runs. Note that our method exhibits significantly less reliance on early stopping compared to the baselines, as illustrated in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.294, + 0.892, + 0.4 + ], + "angle": 0, + "content": "AnimalFaces-LT This benchmark has a naturally occurring imbalance across its classes. We compare our method against the StyleGAN2-ADA benchmarks in terms of quantitative image quality metrics in Table. 1. Our method outperforms the baselines across all metrics, as demonstrated in the results. For a visual comparison against the baselines, please refer to Supp. H." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.893, + 0.763 + ], + "angle": 0, + "content": "Ablation on the Choice of Low-resolution for Unconditional Training (\\( res_{\\mathrm{uc}} \\)). In our proposed method, one of the hyperparameter choices is to select an intermediate low-resolution \\( res_{\\mathrm{uc}} \\) for unconditional training. All layers with equal or lower resolution than \\( res_{\\mathrm{uc}} \\) do not receive class-conditional information. An unconditional GAN objective is added over the images and/or features at \\( res_{\\mathrm{uc}} \\). To study the impact of \\( res_{\\mathrm{uc}} \\), we conducted an ablation study on the AnimalFaces-LT dataset, containing images at \\( 64 \\times 64 \\). Table. 2 presents the results for selecting \\( res_{\\mathrm{uc}} \\) from resolutions lower than \\( 64 \\times 64 \\), namely \\( 8 \\times 8 \\), \\( 16 \\times 16 \\), and \\( 32 \\times 32 \\). Studying the obtained results, we observe that resolutions of \\( 8 \\times 8 \\) and 16 achieve relatively close performance. Conversely, the performance diminishes when layers up to \\( 32 \\times 32 \\) are trained unconditionally. This is anticipated as the AnimalFaces-LT is at \\( 64 \\times 64 \\), leaving only one up-sampling layer to learn the class-conditional information which is shown to be insufficient. In all our experiments, we use \\( res_{\\mathrm{uc}} \\) of \\( 8 \\times 8 \\) unless stated otherwise. Additional visual analysis on the role of \\( res_{\\mathrm{uc}} \\) as well as an ablation on the choice of unconditional training objective weight \\( \\lambda \\) (see Eq. 2), need for unconditional layers in the discriminator, and distinction from \"coarse-to-fine\" training strategies are presented in the Supp. C." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.773, + 0.892, + 0.801 + ], + "angle": 0, + "content": "Table 2. Ablation on choice of low resolution for unconditional training \\( (res_{\\mathrm{uc}}) \\)." + }, + { + "type": "table", + "bbox": [ + 0.56, + 0.803, + 0.83, + 0.864 + ], + "angle": 0, + "content": "
resucFID ↓FID-FS ↓KID ↓KID-FS ↓ ×1000
8 × 826.248.412.619.6
16 × 1627.550.313.720.8
32 × 3238.064.923.334.3
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Can any low-resolution be the starting point for highres generations? In general, our observations indicate that" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7585" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "high-res images are guided, yet not entirely restricted by the low-res images, as shown in Fig. 4. Each row in this figure presents a set of different high-resolution images that are generated from the same unconditional low-resolution image. This illustration reveals that high-resolution images can substantially differ from low-resolution ones, both in terms of color and texture. This suggests that subsequent conditional blocks are capable of significant modification in the background or foreground." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.473, + 0.442 + ], + "angle": 0, + "content": "Importance of FS metrics: comparison against unconditional training. We also include results from training an unconditional StyleGAN2-ADA model [17] in Table 1, referred to as StyleGAN2-ADA UnCond. The unconditional model generates samples that follow the training distribution, which is mainly dominated by head classes. This bias favors unconditional learning in terms of the FID/KID metrics, which do not consider the skewness in the data distribution. Nevertheless, it can be observed the UnCond. baselines significantly perform worse in terms of few-shot metrics FID-FS/KID-FS. This suggests solely relying on FID/KID can be misleading. We strongly recommend including few-shot metrics, i.e., FID-FS and KID-FS, when evaluating cGANs on long-tailed datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.473, + 0.61 + ], + "angle": 0, + "content": "CIFAR10-LT In this benchmark, our method is compared with various baselines with imbalanced ratios \\(\\rho\\) of 50 and 100. This indicates the least frequent tail class includes only 100 and 50 training instances, respectively. Table. 3 quantitatively assesses the quality of the samples generated by our method and the baselines. All the baselines use the StyleGAN2-ADA [17] architecture. Our proposed method consistently outperforms the baselines across all metrics. This demonstrates that beyond improving generative learning across all classes (FID & KID), UTLO can also notably improve the learning from tail classes (FID-FS & KID-FS)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.473, + 0.748 + ], + "angle": 0, + "content": "Effect of Imbalance Ratio \\(\\rho\\). Not surprisingly, all methods demonstrate improved performance with lower values of \\(\\rho\\), as seen in Table. 3. Further, we notice the benefits of knowledge sharing via UTLO become more pronounced as the imbalance in the dataset increases, i.e., learning tail classes becomes more challenging. It is also worth mentioning that the FID-FS and KID-FS metrics are critical to represent the quality of the few-shot classes as those are under-represented in the dataset-wide metrics that as" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.763, + 0.466, + 0.876 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.878, + 0.468, + 0.906 + ], + "angle": 0, + "content": "Figure 4. Different class-conditional images generated given the same unconditional low-resolution images (left-most column)." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.09, + 0.895, + 0.16 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison of the generated images from UTLO against baselines on CIFAR10-LT dataset with different imbalance ratios \\(\\rho\\). UTLO shows substantial improvements on few-shot (FS) metrics compared to StyleGAN2-ADA [17] while outperforming existing methods including NoisyTwins [37]." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.166, + 0.892, + 0.313 + ], + "angle": 0, + "content": "
ρMethodsFID ↓FID-FS ↓KID ↓KID-FS ↓ ×1000
50StyleGAN2-ADA [17]6.521.42.49.0
+ Transitional [42]9.417.74.78.6
+ GSR [36]6.421.32.38.1
+ NoisyTwins [37]6.212.22.45.0
+ UTLO (ours)6.111.82.44.8
100StyleGAN2-ADA [17]9.024.24.09.7
+ Transitional [42]11.320.65.49.2
+ GSR [36]8.424.33.911.8
+ NoisyTwins [37]7.114.12.95.9
+ UTLO (Ours)6.813.42.85.4
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.895, + 0.382 + ], + "angle": 0, + "content": "sume balanced sampling. For instance, at \\(\\rho = 50\\), the FID and KID metrics obtained from UTLO show little changes, while FID-FS and KID-FS exhibit improvements close to \\(50\\%\\) in comparison to GSR." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.893, + 0.596 + ], + "angle": 0, + "content": "Knowledge-sharing at Low Resolutions. As explained in Sec. 3, our proposed generator builds on top of low-resolution (e.g., \\(8 \\times 8\\)) unconditionally trained images which are subsequently used to generate conditional images at higher resolutions (e.g., \\(32 \\times 32\\)). Several low-resolution images are shown in Fig. 6 along with their high-resolution conditional images generated for the head and tail classes of the dataset. It can be seen that conditional images generated from the tail classes evidently share certain features from the head classes. This demonstrates that knowledge sharing at low resolutions is an effective approach for infusing information from head classes to tail ones. A quantitative analysis of knowledge-sharing is presented in Supp. D. For additional qualitative examples see Supp. H." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.895, + 0.809 + ], + "angle": 0, + "content": "CIFAR100-LT In addition to CIFAR10-LT, we evaluate our method and various baselines on the more challenging CIFAR100-LT dataset, wherein the tail classes include as few as five training instances. The quantitative results, shown in Table. 4, reveal that our method outperforms the baselines in a benchmark where a large number of classes with high diversity are present. In Fig. 2, we demonstrate the FID-FS curve during the course of the training along with the generated samples from the tail classes of the CIFAR100-LT dataset. This shows the effectiveness of our proposed method in addressing mode collapse while regularization methods diverge and require early stopping. Further visual comparisons of our proposed method against the baselines for CIFAR100-LT can be found in Supp. H." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.902 + ], + "angle": 0, + "content": "LSUN5-LT We include LSUN5-LT as a higher resolution \\((128 \\times 128)\\) and highly-imbalanced \\((\\rho = 1000)\\) benchmark. In addition, we use Projected GAN with FastGAN[26] generator and DA [40]) as the baseline model. This is a different generator and discriminator design compared to StyleGAN2 along with a different aug-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7586" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.089, + 0.813, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.265, + 0.895, + 0.296 + ], + "angle": 0, + "content": "Figure 5. Generated images from LSUN5-LT dataset. Despite only 50 training instances for the tail class kitchen, the proposed UTLO framework produces diverse, high-fidelity images." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.309, + 0.468, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.409, + 0.47, + 0.506 + ], + "angle": 0, + "content": "Figure 6. Knowledge sharing from head to tail classes in CIFAR10-LT dataset \\((\\rho = 100)\\) using UTLO. The conditional images generated from the head (middle columns) and tail (right columns) classes share and are built on top of the same low-resolution (unconditional) images (left columns). Low-resolution images \\((8 \\times 8)\\) are upsampled to that of CIFAR10-LT \\((32 \\times 32)\\) for improved visualization." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.509, + 0.47, + 0.539 + ], + "angle": 0, + "content": "Table 4. Comparison of our proposed UTLO method with StyleGAN2-ADA baselines on the CIFAR100-LT dataset." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.54, + 0.468, + 0.635 + ], + "angle": 0, + "content": "
MethodsFID ↓FID-FS ↓KID ↓KID-FS ↓
×1000
StyleGAN2-ADA [17]10.824.95.19.3
+ Transitional [42]10.623.74.28.5
+ GSR[36]11.125.05.08.2
+ NoisyTwins [37]10.122.55.07.9
+ UTLO (Ours)9.921.84.67.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.779 + ], + "angle": 0, + "content": "mentation method. In conditional FastGAN implementation, the class-specific information is injected using class-conditional batch normalization, differing from the style mapping network utilized in StyleGAN2. UTLO can be readily adapted to this design by using standard batch normalization layers at low resolutions and utilizing class-conditional batch normalization only at higher resolutions. However, it is noteworthy that methods such as [37] are restricted to StyleGANs and cannot be applied in this context." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.471, + 0.884 + ], + "angle": 0, + "content": "Table. 5 compares the results obtained from our method against the baseline and with the addition of GSR [36]. Our method consistently surpasses the baselines by around \\(30\\%\\) across all metrics. As previously shown for the CIFAR10-LT (Table 3), UTLO not only improved on the tail classes but also significantly improved the dataset-wide FID/KID due to the high imbalance ratio \\((\\rho = 1000)\\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Qualitative results in Fig. 5 contrast the generated" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.309, + 0.892, + 0.338 + ], + "angle": 0, + "content": "Table 5. Evaluating the quality of generated images from the proposed method and comparing against baselines on LSUN5-LT." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.343, + 0.89, + 0.409 + ], + "angle": 0, + "content": "
MethodsFID ↓FID-FS ↓KID ↓KID-FS ↓ ×1000
PGAN (FastGAN)+DA [40]15.060.24.652.7
+ GSR [36]15.763.75.758.0
+ UTLO (Ours)10.943.63.535.3
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.509 + ], + "angle": 0, + "content": "images from our method against the baselines across all classes. For the kitchen class which only includes 50 training examples, the baselines struggle to effectively learn from the tail class whereas UTLO succeeds in generating diverse and high-quality images. More visual examples from our method and baselines are presented in Supp. H." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.893, + 0.603 + ], + "angle": 0, + "content": "Due to space constraints, additional results and analysis are provided in the supplementary material. e.g., Analysis of the role of weighted sampling (Supp. B.), The distinction of training class-conditional GANs in the Long-tailed setup vs. Limited-data setup (Supp. E.), and additional evaluation on Flowers-LT and iNaturalist2019 datasets (Supp. G)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.612, + 0.62, + 0.629 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.833 + ], + "angle": 0, + "content": "In this paper, we proposed UTLO, a novel framework designed to improve the training of cGANs on long-tailed data. Inspired by the observation that head and tail classes often have more similarities at lower resolutions, our method facilitates knowledge sharing from head to tail classes using unconditional training at lower resolutions. The proposed method enriches the limited training distribution of the tail classes and effectively addresses mode collapse, leading to significant improvement in image generation for tail classes. We have also introduced the FID-FS/KID-FS metric, an adaptation of widely-used GAN metrics, specifically tailored for tail classes. We hope our findings are useful for future work in long-tail learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.844, + 0.669, + 0.861 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work is partially supported by the National Science Foundation (NSF) under grant 1751412 and 1927564." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "7587" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.47, + 0.295 + ], + "angle": 0, + "content": "[3] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.354 + ], + "angle": 0, + "content": "[4] Jiequan Cui, Shu Liu, Zhuotao Tian, Zhisheng Zhong, and Jiaya Jia. Result: Residual learning for long-tailed recognition. *IEEE transactions on pattern analysis and machine intelligence*, 2021. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.47, + 0.437 + ], + "angle": 0, + "content": "[5] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge J. Belongie. Class-balanced loss based on effective number of samples. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 9268-9277. Computer Vision Foundation / IEEE, 2019. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.47, + 0.521 + ], + "angle": 0, + "content": "[6] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.47, + 0.591 + ], + "angle": 0, + "content": "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, A aron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.47, + 0.661 + ], + "angle": 0, + "content": "[8] Hao Guo and Song Wang. Long-tailed multi-label visual recognition by collaborative training on uniform and rebalanced samplings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15089-15098, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.663, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[9] Hui Han, Wen-Yuan Wang, and Bing-Huan Mao. Borderline-smote: a new over-sampling method in imbalanced data sets learning. In International conference on intelligent computing, pages 878-887. Springer, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[11] Grant Van Horn and Oisin Mac Aodha. The inaturalist 2019 competition dataset. https://www.kaggle.com/c/inaturalist-2019-fgvc6.5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[13] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[14] Bingyi Kang, Yu Li, Sai Nan Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[15] Tejan Karmali, Rishubh Parihar, Susmit Agrawal, Harsh Rangwani, Varun Jampani, Maneesh Kumar Singh, and R. Venkatesh Babu. Hierarchical semantic regularization of latent spaces in stylegans. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 443-459. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[16] Animesh Karnewar and Oliver Wang. MSG-GAN: multiscale gradients for generative adversarial networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 7796-7805. Computer Vision Foundation / IEEE, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[17] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. 1, 2, 4, 6, 7, 8, 12, 13, 16, 17, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 8107-8116. Computer Vision Foundation / IEEE, 2020. 2, 4, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.574, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[19] Chaerin Kong, Jeesoo Kim, Donghoon Han, and Nojun Kwak. Few-shot image generation with mixup-based distance learning. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 563-580. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.645, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[20] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. preprint, 2009. 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[21] Nupur Kumari, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Ensembling off-the-shelf models for gan training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10651-10662, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[22] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.83, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[23] Bingchen Liu, Yizhe Zhu, Kunpeng Song, and A. Elgammal. Towards faster and stabilized gan training for high-fidelity few-shot image synthesis. iclr, 2021. 2, 4, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[24] Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, and Wenhui Li. Deep representation learning on long-tailed data:" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7588" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "A learnable embedding augmentation perspective. Computer Vision And Pattern Recognition, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.471, + 0.162 + ], + "angle": 0, + "content": "[25] Shudong Liu and Ke Zhang. Under-sampling and feature selection algorithms for s2smlp. IEEE Access, 8:191803-191814, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.471, + 0.233 + ], + "angle": 0, + "content": "[26] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.288 + ], + "angle": 0, + "content": "[27] Mario Lucic, Karol Kurach, Marcin Michalski, Sylvain Gelly, and Olivier Bousquet. Are gans created equal? a large-scale study. Advances in neural information processing systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.289, + 0.471, + 0.344 + ], + "angle": 0, + "content": "[28] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.385 + ], + "angle": 0, + "content": "[29] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv: Arxiv-1411.1784, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.471, + 0.427 + ], + "angle": 0, + "content": "[30] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. ICLR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.471, + 0.47 + ], + "angle": 0, + "content": "[31] S. S. Mullick, Shounak Datta, and Swagatam Das. Generative adversarial minority oversampling. IEEE International Conference On Computer Vision, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.471, + 0.525 + ], + "angle": 0, + "content": "[32] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.471, + 0.595 + ], + "angle": 0, + "content": "[33] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, pages 2642-2651. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.596, + 0.471, + 0.678 + ], + "angle": 0, + "content": "[34] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A. Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10743-10752, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.471, + 0.747 + ], + "angle": 0, + "content": "[35] Harsh Rangwani, Konda Reddy Mopuri, and R. Venkatesh Babu. Class balancing gan with a classifier in the loop. In Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence, pages 1618-1627. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.817 + ], + "angle": 0, + "content": "[36] Harsh Rangwani, Naman Jaswani, Tejan Karmali, Varun Jampani, and R. Venkatesh Babu. Improving gans for long-tailed data through group spectral regularization. European Conference On Computer Vision, 2022. 2, 3, 6, 7, 8, 12, 13, 16, 18, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[37] Harsh Rangwani, Lavish Bansal, Kartik Sharma, Tejan Karmali, Varun Jampani, and R Venkatesh Babu. Noisytwins: Class-consistent and diverse image generation through stylegans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5987-5996, 2023. 2, 3, 6, 7, 8, 18, 19" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[38] Jiawei Ren, Cunjun Yu, shunan sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and hongsheng Li. Balanced meta-softmax for long-tailed visual recognition. In Advances in Neural Information Processing Systems, pages 4175–4186. Curran Associates, Inc., 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.217 + ], + "angle": 0, + "content": "[39] Shibani Santurkar, Ludwig Schmidt, and A. Madry. A classification-based study of covariate shift in gan distributions. International Conference On Machine Learning, 2017. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.893, + 0.273 + ], + "angle": 0, + "content": "[40] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster. Advances in Neural Information Processing Systems, 34:17480-17492, 2021. 6, 7, 8, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.893, + 0.355 + ], + "angle": 0, + "content": "[41] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleganxI: Scaling stylegan to large diverse datasets. In SIGGRAPH '22: Special Interest Group on Computer Graphics and Interactive Techniques Conference, Vancouver, BC, Canada, August 7 - 11, 2022, pages 49:1-49:10. ACM, 2022. 1, 2, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.893, + 0.425 + ], + "angle": 0, + "content": "[42] Mohamad Shahbazi, Martin Danelljan, Danda Pani Paudel, and Luc Van Gool. Collapse by conditioning: Training class-conditional GANs with limited data. In International Conference on Learning Representations, 2022. 1, 3, 6, 7, 8, 16, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.426, + 0.893, + 0.455 + ], + "angle": 0, + "content": "[43] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.893, + 0.51 + ], + "angle": 0, + "content": "[44] Zhangzhang Si and Song-Chun Zhu. Learning hybrid image templates (hit) by information projection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(7):1354-1367, 2012. 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.893, + 0.58 + ], + "angle": 0, + "content": "[45] Kaihua Tang, Jianqiang Huang, and Hanwang Zhang. Long-tailed classification by keeping the good and removing the bad momentum causal effect. In Advances in Neural Information Processing Systems, pages 1513-1524. Curran Associates, Inc., 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.893, + 0.636 + ], + "angle": 0, + "content": "[46] Ngoc-Trung Tran, Viet-Hung Tran, Ngoc-Bao Nguyen, Trung-Kien Nguyen, and Ngai-Man Cheung. On data augmentation for gan training. IEEE Transactions on Image Processing, 30:1882-1897, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.636, + 0.893, + 0.704 + ], + "angle": 0, + "content": "[47] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7921-7931, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.705, + 0.893, + 0.775 + ], + "angle": 0, + "content": "[48] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X. Yu. Long-tailed recognition by routing diverse distribution-aware experts. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.893, + 0.817 + ], + "angle": 0, + "content": "[49] Lu Yang, He Jiang, Qing Song, and Jun Guo. A survey on long-tailed visual recognition. International Journal of Computer Vision, 130:1837-1872, 2022. 2, 3, 5, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.893, + 0.873 + ], + "angle": 0, + "content": "[50] Baofeng Yao and Lei Wang. An improved under-sampling imbalanced classification algorithm. In 2021 13th International Conference on Measuring Technology and Mechatronics Automation (ICMTMA), pages 775-779. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.893, + 0.903 + ], + "angle": 0, + "content": "[51] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7589" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[52] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.47, + 0.26 + ], + "angle": 0, + "content": "[53] Han Zhang, Zizhao Zhang, Augustus Odena, and Honglak Lee. Consistency regularization for generative adversarial networks. In International Conference on Learning Representations, 2019. 2, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.471, + 0.318 + ], + "angle": 0, + "content": "[54] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. IEEE/cvf Conference On Computer Vision And Pattern Recognition, 2018. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.36 + ], + "angle": 0, + "content": "[55] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. arXiv preprint arXiv: Arxiv-2110.04596, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[56] Zizhao Zhang and Tomas Pfister. Learning fast sample re-weighting without reward data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 725-734, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[57] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. Advances in Neural Information Processing Systems, 33:7559-7570, 2020. 2, 3, 6, 12, 16, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.47, + 0.516 + ], + "angle": 0, + "content": "[58] Zhengli Zhao, Zizhao Zhang, Ting Chen, Sameer Singh, and Han Zhang. Image augmentations for gan training. arXiv preprint arXiv: 2006.02595, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[59] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 16489-16498. Computer Vision Foundation / IEEE, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7590" + } + ] +] \ No newline at end of file diff --git a/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/301b4dc8-6f1d-458c-a509-0fe38145fc73_origin.pdf b/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/301b4dc8-6f1d-458c-a509-0fe38145fc73_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bf20062f95ff2cea07e1286455f4b50770e3754b --- /dev/null +++ b/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/301b4dc8-6f1d-458c-a509-0fe38145fc73_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70a90bbe77572b421b6315d736cf489612bcd8b79d948695dd0e1a066c991a5e +size 6915749 diff --git a/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/full.md b/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..701ec306a1862e1f25dad7e1de46be09ef06c228 --- /dev/null +++ b/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/full.md @@ -0,0 +1,308 @@ +# Taming the Tail in Class-Conditional GANs: Knowledge Sharing via Unconditional Training at Lower Resolutions + +Saeed Khorram1, Mingqi Jiang1, Mohamad Shahbazi2, Mohamad H. Danesh3, Li Fuxin1 +1Oregon State University 2 ETH Zürich 3 McGill University +{khorrams, jiangmi, lif}@oregonstate.edu +mshahbazi@vision.ee.ethz.ch, mohamad.danesh@mail.mcgill.ca + +# Abstract + +Despite extensive research on training generative adversarial networks (GANs) with limited training data, learning to generate images from long-tailed training distributions remains fairly unexplored. In the presence of imbalanced multi-class training data, GANs tend to favor classes with more samples, leading to the generation of low quality and less diverse samples in tail classes. In this study, we aim to improve the training of class-conditional GANs with long-tailed data. We propose a straightforward yet effective method for knowledge sharing, allowing tail classes to borrow from the rich information from classes with more abundant training data. More concretely, we propose modifications to existing class-conditional GAN architectures to ensure that the lower-resolution layers of the generator are trained entirely unconditionally while reserving class-conditional generation for the higher-resolution layers. Experiments on several long-tail benchmarks and GAN architectures demonstrate a significant improvement over existing methods in both the diversity and fidelity of the generated images. The code is available at https://github.com/khorrams/utlo. + +# 1. Introduction + +In the past few years, research on Generative Adversarial Networks (GANs) [7] has led to remarkable advances in generating realistic images [2, 17, 41]. Conditional GANs [29] (cGANs) have garnered particular attention due to their ability to accept user inputs which can additionally guide the generation process. They enable a wide range of applications such as class-conditional image generation [42], image manipulation [43], image-to-image translation [12], super-resolution [22] and text-to-image synthesis [52]. + +Despite the advances, past research on cGANs has focused primarily on learning from balanced data. In real-world scenarios, however, data often follows a power-law distribution, with a few classes dominating most of the + +![](images/6f4a0dd369a59fafe23dd9933222693c0433260cfcca2151847975791f6d3ba9.jpg) +All (2) Training Images in Tail Classes + +![](images/dccb4b8415eb372b06fec69c1ebfed3f30347bf85f4d5dd403cc004810a74dbc.jpg) +Baseline (FID: 9.8) + +![](images/1a40cfaf4559f03714d37f685a7014e6dbc4fe2f6fdb51d0c41d69ea5c81b594.jpg) +Baseline + Ours (FID: 6.6) +Figure 1. Generating images from rare tail classes in the Flowers-LT with only two training images. Our proposed approach allows for a more diverse set of features such as backgrounds, colors, poses, and object layouts to be infused into the tail classes. + +training data. This is referred to as the "long-tail (LT) problem" [55], where most of the training data come from a few classes (referred to as the "head" of the distribution). In contrast, data from a large number of classes rarely occur (referred to as the "tail" of the distribution). This imbalance in data poses a challenge to the effective learning of tail classes, as standard machine learning algorithms tend + +to favor densely sampled regions of the input distribution. + +Although there has been a surge of interest in the long-tail problem in recent years, the focus has been primarily on recognition tasks [49]. Generative learning on the long-tail data, particularly cGANs which is the focus of this work, remains underexplored. Recently, [36, 37] identified mode collapse on tail classes as a result of the spectral norm explosion in class-conditional BatchNorms and the collapse of the latents in the StyleGANs $\mathcal{W}$ space. While techniques such as regularization, normalization, and/or classbalancing techniques [49] might improve the performance of long-tail generative models to some extent, their effectiveness is limited. The primary challenge in the long-tail problem is the under-representation of tail classes in the training data, e.g., in Fig. 1 each tail class contains merely 2 training images, resulting in insufficient observation of the tail classes by the learning algorithm. + +To address this challenge, we propose to leverage the abundant information in the head classes, with the aim to learn and infuse knowledge from the head classes into the tail ones, thereby enriching the training distribution for the latter. However, it is very difficult to explicitly disentangle the class-specific information pertaining to head classes from class-independent information shared between head and tail classes. We make a key observation that the similarity between head and tail instances is often higher at lower resolutions: discriminative features such as shape or specific texture are usually unveiled at higher resolutions. Conversely, information from the lower resolutions tends to be more class-independent such as background, configuration, or direction of objects (Fig. 6), thus can be shared between the head and tail classes. + +Building upon this observation, we propose Unconditional Training at LOWER resolution (UTLO), a novel approach for training cGANs in the long-tail setup. In addition to the standard conditional GAN objective, UTLO trains the intermediate low-resolution output of the generator with an unconditional GAN objective. The unconditional GAN objective encourages the learning of low-resolution features common to both the head and tail classes, which are infused into the subsequent layers of the generator, particularly benefiting the under-represented tail classes. + +Through extensive experiments and analysis on several long-tailed datasets, we demonstrate that using UTLO to combine high-resolution conditional and low-resolution unconditional training effectively facilitates knowledge sharing between head and tail classes, thereby improving the overall generative modeling of cGANs in the long-tail setup. Due to the strong class imbalance in the long-tailed data, naive usage of existing GAN metrics can be misleading. To mitigate this issue, we propose a few practices to adapt commonly-used GAN metrics to the long-tail setup. + +Below, we highlight our main contributions: + +- We propose UTLO, a novel knowledge-sharing framework tailored for training cGANs in the long-tailed setup. UTLO allows infusing information from the head classes to the tail classes via an additional unconditional objective applied to the low-resolution part of the cGAN generator. To the best of our knowledge, this work is the first to demonstrate that not all layers in a cGAN need to receive class-conditional information, i.e., a cGAN featuring a partially unconditional generator. +- We present a set of practices and metrics designed to adapt the commonly-used GAN evaluation metrics for long-tail setups, enabling a more precise evaluation of the image generation quality. +- Through extensive experiments across multiple benchmarks and architectures, we validate the effectiveness of our proposed method in improving the training of cGANs in the long-tail setup, achieving state-of-the-art results across several long-tail datasets. + +# 2. Related Work + +Conditional Generative Adversarial Networks The initial work on conditional image generation using GANs [7] was presented by [29], in which the class conditions were concatenated to the inputs of the generator and discriminator networks. The AC-GAN [33] proposed the use of auxiliary classification in the discriminator. [2] introduced BigGAN and set a milestone in large-scale and high-resolution conditional image generation. The authors of StyleGAN [17, 18] first map the input noise and class condition to a latent style space, which is then passed to the multiple layers of the generator for image synthesis. More recently, an extension of the StyleGAN called StyleGAN-XL [41] has become the state-of-the-art in conditional image generation on several datasets and outperforms more complex and time-consuming approaches such as diffusion models [6]. + +GAN Regularization under Limited Data Training GANs under limited data is challenging as the discriminator can memorize the training samples, resulting in the collapse of the training or quality degradation of the generated images. Recently, data augmentation and regularization techniques have been incorporated into GAN training as a means to mitigate this problem [15, 17, 46, 53, 57, 58]. Additionally, [47] introduced a regularization term to the GAN objective, which tracks the predictions of the discriminator for real and generated images using separate moving averages. [21] employed off-the-shelf vision models in an ensemble of discriminators, demonstrating improved performance in both limited-data and large-scale GAN training. + +In the context of few-shot image generation, FastGAN [23] proposed a lightweight architecture with a self-supervised discriminator. [34] introduced cross-domain distance consistency in order to transfer diversity from a source domain to a target domain. Subsequently, [19] proposed + +a latent-mixup strategy to smooth the latent space through controlled latent interpolation. + +Recently, [42, 57] have observed that class-conditioning can cause mode collapse in the limited data regime. Their proposed work learns from limited but balanced/unlabeled data, which is different from our setup: heavily imbalanced long-tail data. [42] proposed Transitional-cGAN, a training strategy that starts with unconditional GAN objective and injects class-conditional information during a transition period. This approach was found to be effective in mitigating mode-collapse in the limited data regime. Our main idea of separating pathways between low-resolution and high-resolution is novel w.r.t. [42]: to promote knowledge sharing, we modify the cGAN architecture so that the lower-resolution part of the generator is entirely unconditional while only the higher-resolution part is conditional and received class-conditional information. + +Long-tail Recognition To address the long-tail recognition problem [49, 55], previous research has primarily focused on techniques such as class re-balancing [3, 28, 38], learning algorithm and model design [13, 14, 59], and information augmentation [24, 45, 48]. More related to our work, re-sampling [8, 56] has been widely used to handle class imbalance problems. This involves balancing the number of samples used during training through over-sampling and/or under-sampling. Oversampling increases the frequency of tail classes [9, 31] while under-sampling reduces the imbalance by reducing the frequency of head class instances [25, 50]. Note that our research diverges from prior studies on long-tail recognition by addressing the more challenging task of generative modeling of the long-tail data, as opposed to the conventional long-tail classification task. + +Training GANs on Long-tail Data Recent years have seen a growing interest in learning from long-tail data. However, most focus has been directed toward recognition tasks, with a limited number of studies addressing the development of generative models for long-tail data. GAMO [31] uses adversarial training to over-sample from minority classes. [35] introduced class-balancing regularization to the GAN objective, which leverages the predictions of a pretrained classifier to improve the generation of underrepresented classes in an unconditional setting. However, this method trains an unconditional GAN and is restricted to having access to a pre-trained classifier on the long-tail data to guide the training. The most closely related work to ours is the Group Spectral Regularization (GSR) [36] and NoisyTwins [37] regularizations. The authors identified mode collapse on tail classes as a result of the spectral norm explosion in class-conditional BatchNorms and the collapse of the latents in the StyleGANs $\mathcal{W}$ space, respectively. In contrast, we present a novel approach for sharing knowledge between head and tail classes to enhance long-tail learning, which is orthogonal to regularization techniques. In addition, while + +[37] is restricted to StyleGANs, our framework can be extended to different GAN architectures. + +# 3. Taming the Tail in cGANs + +# 3.1. Background: Conditional Generative Adversarial Networks + +Conditional Generative Adversarial Networks (cGANs) are suitable for many applications due to providing user control over the generated samples at inference time. This is particularly helpful in the long-tail setup, where it is desirable to explicitly generate samples from rare (tail) classes. In contrast, unconditional GANs are likely to generate samples that track the training distribution, resulting in most samples coming from common (head) classes. To train a cGAN over a dataset containing $n$ instances $\mathbb{X} = \{\pmb{x}_1,\dots ,\pmb{x}_n\}$ and their corresponding labels $\mathbb{Y} = \{y_{1},\ldots ,y_{n}\}$ , we can formulate the adversarial training objective that alternatively optimizes the following loss terms, + +$$ +\mathcal {L} _ {c} ^ {D} = \mathbb {E} _ {\boldsymbol {x}, y} [ f _ {D} (- D (\boldsymbol {x} | y)) ] + \mathbb {E} _ {\boldsymbol {z}, y} [ f _ {D} (D (G (\boldsymbol {z}, y))) ] +$$ + +$$ +\mathcal {L} _ {c} ^ {G} = \mathbb {E} _ {\boldsymbol {z}, y} [ f _ {G} (- D (G (\boldsymbol {z}, y))) ] \tag {1} +$$ + +where $G$ and $D$ are the generator and discriminator networks with their corresponding loss functions $f_{G}$ and $f_{D}$ [27]; $\pmb{x}|y\sim p_{\mathrm{data}}(\pmb{x}|y)$ is real data drawn from class $y$ . The generator $G$ conditions on both $z\sim p_z(z)$ , a random noise vector from a prior distribution over latent space, and $y$ , the class conditioning vector. This adversarial training objective encourages the cGAN to reach an equilibrium between the generator and the discriminator, resulting in the generator producing realistic samples indistinguishable from real data by the discriminator. + +# 3.2. Caveat of Training cGANs on Long-tail Data + +Training GANs on long-tail data is challenging due to the inherent difficulty in modeling the rare examples that comprise the tail of the data distribution. This skewness in the data distribution hinders balanced learning across classes. In particular, the discriminator does not see enough examples from the tail of the distribution during training, which can lead to poor discriminative signals for the generator. As a result, the generator learns to generate a small subset of the possible outputs that fools the inadequately learned discriminator – exhibiting a classic “mode collapse” scenario. + +Although regularizations introduced in [36, 37] can help mitigate the collapse of learning in tail classes that contain a sufficiently large number of samples, we still observe mode collapse occurring when a very limited number of samples are in tail classes, which is inherently challenging. This can be seen in Fig. 2 for tail classes of CIFAR100-LT with as few as 5 training instances. Although early-stopping can be adopted to obtain reasonable performance, the key in boost- + +![](images/b5a8d2f28344ede7ebda1ecd806c5c7e43db2d70a1b4ea0f75bfcfdff7bb5bef.jpg) + +![](images/8ffed90ca950e0de7fb240c83639adec6c3eb98e8554a766c9bcc84b33693056.jpg) +Figure 2. Convergence of different methods on CIFAR100-LT $(\rho = 100)$ , where tail classes have as few as 5 training examples. Incorporating our framework into the baseline alleviates overfitting as a result of knowledge sharing from head to rare tail classes. + +![](images/db32d36e6b935cb2edc49caa32715f170289f72b0b7654b8866f9873146b8867.jpg) + +![](images/f621dd8311f986bceac51b1440c3dac89ae767d025c387d52ce32b7eb67a7541.jpg) + +![](images/dfe73fe5f9a38b618e16649cb2fceeedf5d855d97c7662c26767554f88425339.jpg) + +ing the performance would be an approach to infuse additional diversities into the tail classes rather than only relying on techniques such as regularization. + +# 3.3. Knowledge Sharing via Unconditional Training at Lower Resolutions (UTLO) + +We devise a novel generative adversarial network (GAN) training scheme for long-tail data that aims to share knowledge from predominant head classes to rare tail classes, assuming some of the tail and head classes are at least coarsely similar. We propose a framework that uses unconditional learning in low-resolution images / features at an intermediate layer of the generator using the unconditional GAN objective. This helps with learning universal features from both head and tail classes. The subsequent layers, responsible for introducing finer details at higher resolutions, are trained using a class-conditional GAN objective. + +In our proposed framework, we perform conditional training on the high-resolution image output from the final layer of the generator network. This is built upon the features learned at lower layers through the use of an unconditional objective, inheriting features primarily from head classes. The conditioning on the class labels gives control at the inference time to explicitly generate images from tail classes, meeting our design desideratum. The combination of unconditional and conditional objectives using our proposed method enables knowledge sharing, which, in turn, improves the quality of GAN training on long-tail data. + +Our framework is general and applicable to many GAN architectures. In this section, we illustrate the application of our framework on StyleGAN2 with adaptive data augmentation (ADA) [17], a state-of-the-art and solid baseline for training GANs, particularly in the limited data regime. In the following, we demonstrate the necessary modifications to the architecture of the generator and discriminator networks. Although we show the necessary modifications for StyleGAN2-ADA, our method can be easily extended to other GAN architectures such as FastGAN [23], which is also used in our experiments (see Section 4). + +# 3.3.1 Modifying the Generator + +In the generator design of the StyleGAN2-ADA [17, 18], the class-conditional information, combined with the latent vector, is first embedded in the style space $w$ using the style-mapping network $G_{\mathrm{map}}$ . The style vectors are then broadcasted across the layers of the synthesis network $G_{\mathrm{syn}}$ in order to generate diverse images. To meet our design desideratum, we dissect the synthesis network into two subnetworks $G_{\mathrm{syn}} = G_h \circ G_l$ where $G_l$ represents the earlier part of the network that produces intermediate features and/or images at a low resolution $L \times L$ . $G_h$ on the other hand, represents the latter part of $G_{\mathrm{syn}}$ that generates the output image at the high resolution $H \times H$ where $H > L$ . + +To block the flow of class-conditional information to the lower layers of $G_{\mathrm{syn}}$ , we generate separate $w$ vectors, one containing the class-conditional embeddings, referred to as $w_{z,y}$ , while for the other one, referred to as $w_{z}$ , class information was set to zero. Note that both vectors share the same latent variable $z$ and mapping network $G_{\mathrm{map}}$ parameters, and the only difference is the presence of class-conditional information. The lower layers of the synthesis network $G_{l}$ are conditioned on $w_{z}$ while the subsequent layers at higher resolutions receive $w_{z,y}$ (see Fig. 3). + +Generators typically follow a network design that gradually increases the resolution of intermediate features and/or images as the network progresses. This allows us to select a desired low-resolution, such as $8 \times 8$ or $16 \times 16$ , during training. Recent generator designs often incorporate skip connections [16] or residual connections [30] to improve gradient flow. The generator of StyleGAN2-ADA uses skip connections, which explicitly generates intermediate RGB images at each layer. We take this low-resolution image as the input to the discriminator. Note that in the case of a generator design that uses residual connections, a $1 \times 1$ convolutional layer can be incorporated to convert the residual channels to RGB channels. + +In a forward pass through the generator, images at low and high resolutions can be generated simultaneously without additional overhead. By inputting a latent code $z \sim p_{z}(z)$ and a target label $y$ , we can obtain an unconditional synthetic image $\hat{x}_l \in \mathbb{R}^{3 \times L \times L}$ along with a conditional + +![](images/8e61a754a7a7b98da3145e27a96323c25a3973fae410e01287b27271836d3112.jpg) +Figure 3. The proposed framework, UTLO, illustrated for the StyleGAN2-ADA architecture. Low and high resolution image pathways are used for unconditional and class-conditional objectives, respectively. $z$ is the input latent code, $y$ indicates the class embeddings, and $c$ is a constant input. Separate style vectors, $w_{z}$ (class-independent) and $w_{z,y}$ (class-conditional), are generated using the same $z$ and a shared style-mapping network $G_{\mathrm{map}}$ which are then passed to $G_{l}$ and $G_{h}$ , respectively. The high-resolution generated image $\hat{x} \in \mathbb{R}^{3\times H\times H}$ is passed through the discriminator $D = D_{l} \circ D_{h}$ to calculate the conditional objective $\mathcal{L}_c$ while the low-resolution image $\hat{x}_l \in \mathbb{R}^{3\times L\times L}$ is passed only through $D_{l}$ to calculate the unconditional objective $\mathcal{L}_{uc}$ . The final objective $\mathcal{L}$ is the combination of the two. While $D_{l}$ is shared, two separate prediction heads (FC layers) are used for unconditional and conditional objectives. The fromRGB is designed to increase the dimensionality of RGB channels to match the input channels of the $D_{l}$ . + +high-resolution synthetic image $\hat{\pmb{x}}\in \mathbb{R}^{3\times H\times H}$ at the output. While $\hat{\pmb{x}}$ is trained to be from class $y$ , $\hat{\pmb{x}}_l$ has no class-specific constraints. At inference time, the intermediate images can be discarded. + +# 3.3.2 Modifying the Discriminator + +Opposite to the generator design, discriminators gradually reduce the resolution of the intermediate features as the network progresses. In order to pass images to an intermediate layer of the discriminator, we dissect it into two sub-networks. Formally, we define the discriminator as $D = D_{l} \circ D_{h}$ where $D_{h}$ represents the earlier part of the original network that takes the images at training resolution $\mathbb{R}^{3 \times H \times H}$ . On the other hand, $D_{l}$ represents the latter part of the discriminator taking a lower resolution input $\mathbb{R}^{C \times L \times L}$ with $C$ channels. + +The default discriminator design in StyleGAN2-ADA uses residual connections. This does not allow direct passing of RGB inputs to the intermediate layers. To overcome this, we add a $1 \times 1$ convolutional layer, referred to as "fromRGB", that increases the channels of the RGB images to match $C$ , the number of input channels of $D_{l}$ . + +More concretely, $D_{l}$ is a shared network in both the unconditional low-resolution and class-conditional high-resolution image pathways (see Fig. 3). To obtain separate unconditional and conditional predictions, we use two different fully-connected (FC) layers at the final layer of $D_{l}$ . The output of the $D_{h}$ and the low-res image $\hat{\boldsymbol{x}}_l$ are passed through the $D_{l}$ which are then used for calculating the class-conditional loss $\mathcal{L}_c$ and unconditional loss $\mathcal{L}_{uc}$ , respectively. + +Note, when passing real images to $D_{l}$ for the unconditional loss, they are first downsized to the low-resolution $L \times L$ using bi-linear interpolation to be comparable with + +the ones generated from $G_{l}$ . They are then passed through the fromRGB layer and consequently $D_{l}$ . For the class-conditional loss, the original images are passed to $D$ without any modification. + +# 3.3.3 Final Training Objective + +The final objective of our framework is simply a combination of the conditional and unconditional GAN losses, using a weighting parameter $\lambda$ . To put it formally we have, + +$$ +\mathcal {L} ^ {D} = \mathcal {L} _ {c} ^ {D} + \lambda \cdot \mathcal {L} _ {u c} ^ {D}, \mathcal {L} ^ {G} = \mathcal {L} _ {c} ^ {G} + \lambda \cdot \mathcal {L} _ {u c} ^ {G} \tag {2} +$$ + +where $\mathcal{L}_c^D$ and $\mathcal{L}_c^G$ represent the class-conditional losses for the discriminator and generator, respectively (as outlined in Eq. 3.1), and $\mathcal{L}_{uc}^D$ and $\mathcal{L}_{uc}^G$ represent their unconditional counterparts, where $\pmb{x}$ is sampled from $p_{\mathrm{data}}(\pmb{x})$ instead of $p_{\mathrm{data}}(\pmb{x}|y)$ . The overall framework of our proposed method is depicted in Fig. 3. + +# 4. Experiments + +# 4.1. Setup + +Datasets. We use 6 different long-tailed datasets in our experiments: CIFAR10 and CIFAR100[20], LSUN [51], Flowers [32], iNaturalist2019 [11], and AnimalFaces [44]. This selection is intended to encompass a wide spectrum of image domains, dataset sizes, resolutions, and imbalance ratios (denoted as $\rho$ ). $\rho$ represents the ratio of the number of training samples between the classes with the most and the least number of examples. To ensure the tail classes stay few-shot, we keep the number of images in the smallest tail classes under 50 [49]. A detailed description of these datasets is provided in Supp. A. + +Baselines. In order to demonstrate the effectiveness of our proposed method, we use various cGAN architectures cov + +ering different generator and discriminator designs and data augmentation pipelines: + +- StyleGAN2 with Adaptive Data Augmentation (ADA) [17]. We also integrate Transitional training [42], Group Spectral Regularization (GSR) [36], and NoisyTwins [37] to this baseline. +- Projected GAN (PGAN) [40]: we use projected discriminator with Differentiable Augmentation (DA) [57]. We also add GSR [36] to this baseline. + +For comprehensive details on the baselines, hyperparameters, and implementation please refer to Supp. F. + +# 4.2. Evaluation Metrics For Long-Tail Datasets + +Evaluating the images generated by GANs on a long-tailed setup poses challenges, primarily due to the imbalanced data and access to a limited number of samples for the tail classes. For our experiments, we employ widely-used GAN metrics: Fréchet Inception Distance (FID) [10], and Kernel Inception Distance (KID) [1]. Following [10, 17], we report all metrics by computing statistics between $50k$ generated images and all available training images. In the following, we present a set of practices for adapting the aforementioned metrics to the long-tail setup. + +First, when a larger and (more) balanced dataset following the long-tail training data distribution is available, it is used for metric calculation, e.g., full CIFAR10 before artificial imbalance. For naturally imbalanced datasets such as iNaturalist2019, we use the unmodified training set to calculate metrics. Note that when generating images for evaluation, we sample from the same distribution as the available real dataset, which may be imbalanced. While FID and KID indicate how well the real data distribution matches the generated images across all classes, we suggest additional metrics to mitigate the disproportionate influence of head classes in data statistic calculations. + +In response to the varied number of training instances across classes in a long-tail setup, we propose evaluating metrics specifically on few-shot (FS) categories [49]. We calculate the FID and KID for the few-shot subset and refer to them as "FID-FS" and "KID-FS", respectively. This is tailored to evaluate the quality of generated samples in the tail classes. Contrasting with the standard FID/KID, we maintain an equal number of real images across all classes during our FID-FS/KID-FS calculation. This emphasizes learning quality on tail classes, irrespective of any imbalances that might be present in the few-shot subset. Reporting both standard and few-shot metrics provides a more comprehensive evaluation of long-tail learning, considering the performance of both head and tail classes. + +# 4.3. Results + +Below, we report the results obtained across different benchmarks and cGAN architectures. We pick the best model for reporting results as the one with the lowest FID + +Table 1. Our proposed method, UTLO, outperforms the baselines in terms of quantitative image quality metrics on the AnimalFaces-LT dataset (64 × 64 resolution). + +
MethodsFID ↓FID-FS ↓KID ↓KID-FS ↓ ×1000
StyleGAN2-ADA UnCond. [17]39.4104.117.327.6
StyleGAN2-ADA [17]51.487.124.735.9
+ Transitional [42]62.199.038.548.9
+ GSR[36]39.267.221.232.7
+ NoisyTwins [37]29.450.216.721.2
+ UTLO (Ours)26.248.412.619.6
+ +FS from two independent runs. Note that our method exhibits significantly less reliance on early stopping compared to the baselines, as illustrated in Fig. 2. + +AnimalFaces-LT This benchmark has a naturally occurring imbalance across its classes. We compare our method against the StyleGAN2-ADA benchmarks in terms of quantitative image quality metrics in Table. 1. Our method outperforms the baselines across all metrics, as demonstrated in the results. For a visual comparison against the baselines, please refer to Supp. H. + +Ablation on the Choice of Low-resolution for Unconditional Training ( $res_{\mathrm{uc}}$ ). In our proposed method, one of the hyperparameter choices is to select an intermediate low-resolution $res_{\mathrm{uc}}$ for unconditional training. All layers with equal or lower resolution than $res_{\mathrm{uc}}$ do not receive class-conditional information. An unconditional GAN objective is added over the images and/or features at $res_{\mathrm{uc}}$ . To study the impact of $res_{\mathrm{uc}}$ , we conducted an ablation study on the AnimalFaces-LT dataset, containing images at $64 \times 64$ . Table. 2 presents the results for selecting $res_{\mathrm{uc}}$ from resolutions lower than $64 \times 64$ , namely $8 \times 8$ , $16 \times 16$ , and $32 \times 32$ . Studying the obtained results, we observe that resolutions of $8 \times 8$ and 16 achieve relatively close performance. Conversely, the performance diminishes when layers up to $32 \times 32$ are trained unconditionally. This is anticipated as the AnimalFaces-LT is at $64 \times 64$ , leaving only one up-sampling layer to learn the class-conditional information which is shown to be insufficient. In all our experiments, we use $res_{\mathrm{uc}}$ of $8 \times 8$ unless stated otherwise. Additional visual analysis on the role of $res_{\mathrm{uc}}$ as well as an ablation on the choice of unconditional training objective weight $\lambda$ (see Eq. 2), need for unconditional layers in the discriminator, and distinction from "coarse-to-fine" training strategies are presented in the Supp. C. + +Table 2. Ablation on choice of low resolution for unconditional training $(res_{\mathrm{uc}})$ . + +
resucFID ↓FID-FS ↓KID ↓KID-FS ↓ ×1000
8 × 826.248.412.619.6
16 × 1627.550.313.720.8
32 × 3238.064.923.334.3
+ +Can any low-resolution be the starting point for highres generations? In general, our observations indicate that + +high-res images are guided, yet not entirely restricted by the low-res images, as shown in Fig. 4. Each row in this figure presents a set of different high-resolution images that are generated from the same unconditional low-resolution image. This illustration reveals that high-resolution images can substantially differ from low-resolution ones, both in terms of color and texture. This suggests that subsequent conditional blocks are capable of significant modification in the background or foreground. + +Importance of FS metrics: comparison against unconditional training. We also include results from training an unconditional StyleGAN2-ADA model [17] in Table 1, referred to as StyleGAN2-ADA UnCond. The unconditional model generates samples that follow the training distribution, which is mainly dominated by head classes. This bias favors unconditional learning in terms of the FID/KID metrics, which do not consider the skewness in the data distribution. Nevertheless, it can be observed the UnCond. baselines significantly perform worse in terms of few-shot metrics FID-FS/KID-FS. This suggests solely relying on FID/KID can be misleading. We strongly recommend including few-shot metrics, i.e., FID-FS and KID-FS, when evaluating cGANs on long-tailed datasets. + +CIFAR10-LT In this benchmark, our method is compared with various baselines with imbalanced ratios $\rho$ of 50 and 100. This indicates the least frequent tail class includes only 100 and 50 training instances, respectively. Table. 3 quantitatively assesses the quality of the samples generated by our method and the baselines. All the baselines use the StyleGAN2-ADA [17] architecture. Our proposed method consistently outperforms the baselines across all metrics. This demonstrates that beyond improving generative learning across all classes (FID & KID), UTLO can also notably improve the learning from tail classes (FID-FS & KID-FS). + +Effect of Imbalance Ratio $\rho$ . Not surprisingly, all methods demonstrate improved performance with lower values of $\rho$ , as seen in Table. 3. Further, we notice the benefits of knowledge sharing via UTLO become more pronounced as the imbalance in the dataset increases, i.e., learning tail classes becomes more challenging. It is also worth mentioning that the FID-FS and KID-FS metrics are critical to represent the quality of the few-shot classes as those are under-represented in the dataset-wide metrics that as + +![](images/3943075715ef5bc49e78c3680a9f262e33371ff05544fb0c5c983e9f3116dd1a.jpg) +Figure 4. Different class-conditional images generated given the same unconditional low-resolution images (left-most column). + +Table 3. Quantitative comparison of the generated images from UTLO against baselines on CIFAR10-LT dataset with different imbalance ratios $\rho$ . UTLO shows substantial improvements on few-shot (FS) metrics compared to StyleGAN2-ADA [17] while outperforming existing methods including NoisyTwins [37]. + +
ρMethodsFID ↓FID-FS ↓KID ↓KID-FS ↓ ×1000
50StyleGAN2-ADA [17]6.521.42.49.0
+ Transitional [42]9.417.74.78.6
+ GSR [36]6.421.32.38.1
+ NoisyTwins [37]6.212.22.45.0
+ UTLO (ours)6.111.82.44.8
100StyleGAN2-ADA [17]9.024.24.09.7
+ Transitional [42]11.320.65.49.2
+ GSR [36]8.424.33.911.8
+ NoisyTwins [37]7.114.12.95.9
+ UTLO (Ours)6.813.42.85.4
+ +sume balanced sampling. For instance, at $\rho = 50$ , the FID and KID metrics obtained from UTLO show little changes, while FID-FS and KID-FS exhibit improvements close to $50\%$ in comparison to GSR. + +Knowledge-sharing at Low Resolutions. As explained in Sec. 3, our proposed generator builds on top of low-resolution (e.g., $8 \times 8$ ) unconditionally trained images which are subsequently used to generate conditional images at higher resolutions (e.g., $32 \times 32$ ). Several low-resolution images are shown in Fig. 6 along with their high-resolution conditional images generated for the head and tail classes of the dataset. It can be seen that conditional images generated from the tail classes evidently share certain features from the head classes. This demonstrates that knowledge sharing at low resolutions is an effective approach for infusing information from head classes to tail ones. A quantitative analysis of knowledge-sharing is presented in Supp. D. For additional qualitative examples see Supp. H. + +CIFAR100-LT In addition to CIFAR10-LT, we evaluate our method and various baselines on the more challenging CIFAR100-LT dataset, wherein the tail classes include as few as five training instances. The quantitative results, shown in Table. 4, reveal that our method outperforms the baselines in a benchmark where a large number of classes with high diversity are present. In Fig. 2, we demonstrate the FID-FS curve during the course of the training along with the generated samples from the tail classes of the CIFAR100-LT dataset. This shows the effectiveness of our proposed method in addressing mode collapse while regularization methods diverge and require early stopping. Further visual comparisons of our proposed method against the baselines for CIFAR100-LT can be found in Supp. H. + +LSUN5-LT We include LSUN5-LT as a higher resolution $(128 \times 128)$ and highly-imbalanced $(\rho = 1000)$ benchmark. In addition, we use Projected GAN with FastGAN[26] generator and DA [40]) as the baseline model. This is a different generator and discriminator design compared to StyleGAN2 along with a different aug- + +![](images/53b872f9bca3a0fb6d14f9a056b4e5f8007f9be82edcc09e257c54b1baf3a5a2.jpg) +Figure 5. Generated images from LSUN5-LT dataset. Despite only 50 training instances for the tail class kitchen, the proposed UTLO framework produces diverse, high-fidelity images. + +![](images/b7713869e15067a4a9545890df6157f99a4a8b0f945bda653bd9787fd68d4f8b.jpg) +Figure 6. Knowledge sharing from head to tail classes in CIFAR10-LT dataset $(\rho = 100)$ using UTLO. The conditional images generated from the head (middle columns) and tail (right columns) classes share and are built on top of the same low-resolution (unconditional) images (left columns). Low-resolution images $(8 \times 8)$ are upsampled to that of CIFAR10-LT $(32 \times 32)$ for improved visualization. + +Table 4. Comparison of our proposed UTLO method with StyleGAN2-ADA baselines on the CIFAR100-LT dataset. + +
MethodsFID ↓FID-FS ↓KID ↓KID-FS ↓
×1000
StyleGAN2-ADA [17]10.824.95.19.3
+ Transitional [42]10.623.74.28.5
+ GSR[36]11.125.05.08.2
+ NoisyTwins [37]10.122.55.07.9
+ UTLO (Ours)9.921.84.67.5
+ +mentation method. In conditional FastGAN implementation, the class-specific information is injected using class-conditional batch normalization, differing from the style mapping network utilized in StyleGAN2. UTLO can be readily adapted to this design by using standard batch normalization layers at low resolutions and utilizing class-conditional batch normalization only at higher resolutions. However, it is noteworthy that methods such as [37] are restricted to StyleGANs and cannot be applied in this context. + +Table. 5 compares the results obtained from our method against the baseline and with the addition of GSR [36]. Our method consistently surpasses the baselines by around $30\%$ across all metrics. As previously shown for the CIFAR10-LT (Table 3), UTLO not only improved on the tail classes but also significantly improved the dataset-wide FID/KID due to the high imbalance ratio $(\rho = 1000)$ . + +Qualitative results in Fig. 5 contrast the generated + +Table 5. Evaluating the quality of generated images from the proposed method and comparing against baselines on LSUN5-LT. + +
MethodsFID ↓FID-FS ↓KID ↓KID-FS ↓ ×1000
PGAN (FastGAN)+DA [40]15.060.24.652.7
+ GSR [36]15.763.75.758.0
+ UTLO (Ours)10.943.63.535.3
+ +images from our method against the baselines across all classes. For the kitchen class which only includes 50 training examples, the baselines struggle to effectively learn from the tail class whereas UTLO succeeds in generating diverse and high-quality images. More visual examples from our method and baselines are presented in Supp. H. + +Due to space constraints, additional results and analysis are provided in the supplementary material. e.g., Analysis of the role of weighted sampling (Supp. B.), The distinction of training class-conditional GANs in the Long-tailed setup vs. Limited-data setup (Supp. E.), and additional evaluation on Flowers-LT and iNaturalist2019 datasets (Supp. G). + +# 5. Conclusion + +In this paper, we proposed UTLO, a novel framework designed to improve the training of cGANs on long-tailed data. Inspired by the observation that head and tail classes often have more similarities at lower resolutions, our method facilitates knowledge sharing from head to tail classes using unconditional training at lower resolutions. The proposed method enriches the limited training distribution of the tail classes and effectively addresses mode collapse, leading to significant improvement in image generation for tail classes. We have also introduced the FID-FS/KID-FS metric, an adaptation of widely-used GAN metrics, specifically tailored for tail classes. We hope our findings are useful for future work in long-tail learning. + +# Acknowledgements + +This work is partially supported by the National Science Foundation (NSF) under grant 1751412 and 1927564. + +# References + +[1] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In International Conference on Learning Representations, 2018. 6 +[2] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. 1, 2 +[3] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3 +[4] Jiequan Cui, Shu Liu, Zhuotao Tian, Zhisheng Zhong, and Jiaya Jia. Result: Residual learning for long-tailed recognition. *IEEE transactions on pattern analysis and machine intelligence*, 2021. 12 +[5] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge J. Belongie. Class-balanced loss based on effective number of samples. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 9268-9277. Computer Vision Foundation / IEEE, 2019. 12 +[6] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 8780-8794, 2021. 2 +[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, A aron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2014. 1, 2 +[8] Hao Guo and Song Wang. Long-tailed multi-label visual recognition by collaborative training on uniform and rebalanced samplings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15089-15098, 2021. 3 +[9] Hui Han, Wen-Yuan Wang, and Bing-Huan Mao. Borderline-smote: a new over-sampling method in imbalanced data sets learning. In International conference on intelligent computing, pages 878-887. Springer, 2005. 3 +[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[11] Grant Van Horn and Oisin Mac Aodha. The inaturalist 2019 competition dataset. https://www.kaggle.com/c/inaturalist-2019-fgvc6.5, 12 +[12] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 1 + +[13] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 3 +[14] Bingyi Kang, Yu Li, Sai Nan Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In International Conference on Learning Representations, 2021. 3 +[15] Tejan Karmali, Rishubh Parihar, Susmit Agrawal, Harsh Rangwani, Varun Jampani, Maneesh Kumar Singh, and R. Venkatesh Babu. Hierarchical semantic regularization of latent spaces in stylegans. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 443-459. Springer, 2022. 2 +[16] Animesh Karnewar and Oliver Wang. MSG-GAN: multiscale gradients for generative adversarial networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 7796-7805. Computer Vision Foundation / IEEE, 2020. 4 +[17] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. 1, 2, 4, 6, 7, 8, 12, 13, 16, 17, 18 +[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 8107-8116. Computer Vision Foundation / IEEE, 2020. 2, 4, 18 +[19] Chaerin Kong, Jeesoo Kim, Donghoon Han, and Nojun Kwak. Few-shot image generation with mixup-based distance learning. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 563-580. Springer, 2022. 2 +[20] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. preprint, 2009. 5, 12 +[21] Nupur Kumari, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Ensembling off-the-shelf models for gan training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10651-10662, 2022. 2 +[22] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 1 +[23] Bingchen Liu, Yizhe Zhu, Kunpeng Song, and A. Elgammal. Towards faster and stabilized gan training for high-fidelity few-shot image synthesis. iclr, 2021. 2, 4, 18 +[24] Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, and Wenhui Li. Deep representation learning on long-tailed data: + +A learnable embedding augmentation perspective. Computer Vision And Pattern Recognition, 2020. 3 +[25] Shudong Liu and Ke Zhang. Under-sampling and feature selection algorithms for s2smlp. IEEE Access, 8:191803-191814, 2020. 3 +[26] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 7 +[27] Mario Lucic, Karol Kurach, Marcin Michalski, Sylvain Gelly, and Olivier Bousquet. Are gans created equal? a large-scale study. Advances in neural information processing systems, 31, 2018. 3 +[28] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In International Conference on Learning Representations, 2021. 3 +[29] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv: Arxiv-1411.1784, 2014. 1, 2 +[30] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. ICLR, 2018. 4 +[31] S. S. Mullick, Shounak Datta, and Swagatam Das. Generative adversarial minority oversampling. IEEE International Conference On Computer Vision, 2019. 3 +[32] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5, 12 +[33] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, pages 2642-2651. PMLR, 2017. 2 +[34] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A. Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10743-10752, 2021. 2 +[35] Harsh Rangwani, Konda Reddy Mopuri, and R. Venkatesh Babu. Class balancing gan with a classifier in the loop. In Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence, pages 1618-1627. PMLR, 2021. 3 +[36] Harsh Rangwani, Naman Jaswani, Tejan Karmali, Varun Jampani, and R. Venkatesh Babu. Improving gans for long-tailed data through group spectral regularization. European Conference On Computer Vision, 2022. 2, 3, 6, 7, 8, 12, 13, 16, 18, 19 +[37] Harsh Rangwani, Lavish Bansal, Kartik Sharma, Tejan Karmali, Varun Jampani, and R Venkatesh Babu. Noisytwins: Class-consistent and diverse image generation through stylegans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5987-5996, 2023. 2, 3, 6, 7, 8, 18, 19 + +[38] Jiawei Ren, Cunjun Yu, shunan sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and hongsheng Li. Balanced meta-softmax for long-tailed visual recognition. In Advances in Neural Information Processing Systems, pages 4175–4186. Curran Associates, Inc., 2020. 3 +[39] Shibani Santurkar, Ludwig Schmidt, and A. Madry. A classification-based study of covariate shift in gan distributions. International Conference On Machine Learning, 2017. 12 +[40] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster. Advances in Neural Information Processing Systems, 34:17480-17492, 2021. 6, 7, 8, 18 +[41] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleganxI: Scaling stylegan to large diverse datasets. In SIGGRAPH '22: Special Interest Group on Computer Graphics and Interactive Techniques Conference, Vancouver, BC, Canada, August 7 - 11, 2022, pages 49:1-49:10. ACM, 2022. 1, 2, 15, 16 +[42] Mohamad Shahbazi, Martin Danelljan, Danda Pani Paudel, and Luc Van Gool. Collapse by conditioning: Training class-conditional GANs with limited data. In International Conference on Learning Representations, 2022. 1, 3, 6, 7, 8, 16, 18 +[43] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 1 +[44] Zhangzhang Si and Song-Chun Zhu. Learning hybrid image templates (hit) by information projection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(7):1354-1367, 2012. 5, 12 +[45] Kaihua Tang, Jianqiang Huang, and Hanwang Zhang. Long-tailed classification by keeping the good and removing the bad momentum causal effect. In Advances in Neural Information Processing Systems, pages 1513-1524. Curran Associates, Inc., 2020. 3 +[46] Ngoc-Trung Tran, Viet-Hung Tran, Ngoc-Bao Nguyen, Trung-Kien Nguyen, and Ngai-Man Cheung. On data augmentation for gan training. IEEE Transactions on Image Processing, 30:1882-1897, 2021. 2 +[47] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7921-7931, 2021. 2 +[48] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X. Yu. Long-tailed recognition by routing diverse distribution-aware experts. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3 +[49] Lu Yang, He Jiang, Qing Song, and Jun Guo. A survey on long-tailed visual recognition. International Journal of Computer Vision, 130:1837-1872, 2022. 2, 3, 5, 6, 12 +[50] Baofeng Yao and Lei Wang. An improved under-sampling imbalanced classification algorithm. In 2021 13th International Conference on Measuring Technology and Mechatronics Automation (ICMTMA), pages 775-779. IEEE, 2021. 3 +[51] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset + +using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 5, 12 +[52] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 1 +[53] Han Zhang, Zizhao Zhang, Augustus Odena, and Honglak Lee. Consistency regularization for generative adversarial networks. In International Conference on Learning Representations, 2019. 2, 12 +[54] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. IEEE/cvf Conference On Computer Vision And Pattern Recognition, 2018. 16 +[55] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. arXiv preprint arXiv: Arxiv-2110.04596, 2021. 1, 3 +[56] Zizhao Zhang and Tomas Pfister. Learning fast sample re-weighting without reward data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 725-734, 2021. 3 +[57] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. Advances in Neural Information Processing Systems, 33:7559-7570, 2020. 2, 3, 6, 12, 16, 18 +[58] Zhengli Zhao, Zizhao Zhang, Ting Chen, Sameer Singh, and Han Zhang. Image augmentations for gan training. arXiv preprint arXiv: 2006.02595, 2020. 2 +[59] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 16489-16498. Computer Vision Foundation / IEEE, 2021. 3 \ No newline at end of file diff --git a/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/images.zip b/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..89e4b3261eb9204b671ca823704796c8f1fb8250 --- /dev/null +++ b/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e15859e08440df18636991f3d18d329d2929cda844e3ec8fee012c170b9f6891 +size 472552 diff --git a/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/layout.json b/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..60ce906e6372d194b058da230d9524d4bdc5b2e8 --- /dev/null +++ b/2024/Taming the Tail in Class-Conditional GANs_ Knowledge Sharing via Unconditional Training at Lower Resolutions/layout.json @@ -0,0 +1,9498 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 507, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 507, + 138 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 507, + 138 + ], + "type": "text", + "content": "Taming the Tail in Class-Conditional GANs: Knowledge Sharing via Unconditional Training at Lower Resolutions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 160, + 512, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 160, + 512, + 217 + ], + "spans": [ + { + "bbox": [ + 80, + 160, + 512, + 217 + ], + "type": "text", + "content": "Saeed Khorram1, Mingqi Jiang1, Mohamad Shahbazi2, Mohamad H. Danesh3, Li Fuxin1 \n1Oregon State University 2 ETH Zürich 3 McGill University \n{khorrams, jiangmi, lif}@oregonstate.edu \nmshahbazi@vision.ee.ethz.ch, mohamad.danesh@mail.mcgill.ca" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 245, + 192, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 245, + 192, + 257 + ], + "spans": [ + { + "bbox": [ + 143, + 245, + 192, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 271, + 290, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 271, + 290, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 290, + 522 + ], + "type": "text", + "content": "Despite extensive research on training generative adversarial networks (GANs) with limited training data, learning to generate images from long-tailed training distributions remains fairly unexplored. In the presence of imbalanced multi-class training data, GANs tend to favor classes with more samples, leading to the generation of low quality and less diverse samples in tail classes. In this study, we aim to improve the training of class-conditional GANs with long-tailed data. We propose a straightforward yet effective method for knowledge sharing, allowing tail classes to borrow from the rich information from classes with more abundant training data. More concretely, we propose modifications to existing class-conditional GAN architectures to ensure that the lower-resolution layers of the generator are trained entirely unconditionally while reserving class-conditional generation for the higher-resolution layers. Experiments on several long-tail benchmarks and GAN architectures demonstrate a significant improvement over existing methods in both the diversity and fidelity of the generated images. The code is available at https://github.com/khorrams/utlo." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 535, + 127, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 127, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 127, + 547 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 556, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 556, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 556, + 287, + 664 + ], + "type": "text", + "content": "In the past few years, research on Generative Adversarial Networks (GANs) [7] has led to remarkable advances in generating realistic images [2, 17, 41]. Conditional GANs [29] (cGANs) have garnered particular attention due to their ability to accept user inputs which can additionally guide the generation process. They enable a wide range of applications such as class-conditional image generation [42], image manipulation [43], image-to-image translation [12], super-resolution [22] and text-to-image synthesis [52]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "Despite the advances, past research on cGANs has focused primarily on learning from balanced data. In real-world scenarios, however, data often follows a power-law distribution, with a few classes dominating most of the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 258, + 535, + 315 + ], + "blocks": [ + { + "bbox": [ + 347, + 246, + 505, + 258 + ], + "lines": [ + { + "bbox": [ + 347, + 246, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 347, + 246, + 505, + 258 + ], + "type": "text", + "content": "All (2) Training Images in Tail Classes" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 258, + 535, + 315 + ], + "lines": [ + { + "bbox": [ + 317, + 258, + 535, + 315 + ], + "spans": [ + { + "bbox": [ + 317, + 258, + 535, + 315 + ], + "type": "image", + "image_path": "6f4a0dd369a59fafe23dd9933222693c0433260cfcca2151847975791f6d3ba9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 328, + 535, + 438 + ], + "blocks": [ + { + "bbox": [ + 386, + 317, + 466, + 327 + ], + "lines": [ + { + "bbox": [ + 386, + 317, + 466, + 327 + ], + "spans": [ + { + "bbox": [ + 386, + 317, + 466, + 327 + ], + "type": "text", + "content": "Baseline (FID: 9.8)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 328, + 535, + 438 + ], + "lines": [ + { + "bbox": [ + 317, + 328, + 535, + 438 + ], + "spans": [ + { + "bbox": [ + 317, + 328, + 535, + 438 + ], + "type": "image", + "image_path": "dccb4b8415eb372b06fec69c1ebfed3f30347bf85f4d5dd403cc004810a74dbc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 453, + 535, + 563 + ], + "blocks": [ + { + "bbox": [ + 369, + 442, + 478, + 453 + ], + "lines": [ + { + "bbox": [ + 369, + 442, + 478, + 453 + ], + "spans": [ + { + "bbox": [ + 369, + 442, + 478, + 453 + ], + "type": "text", + "content": "Baseline + Ours (FID: 6.6)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 453, + 535, + 563 + ], + "lines": [ + { + "bbox": [ + 317, + 453, + 535, + 563 + ], + "spans": [ + { + "bbox": [ + 317, + 453, + 535, + 563 + ], + "type": "image", + "image_path": "1a40cfaf4559f03714d37f685a7014e6dbc4fe2f6fdb51d0c41d69ea5c81b594.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 565, + 545, + 609 + ], + "lines": [ + { + "bbox": [ + 305, + 565, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 305, + 565, + 545, + 609 + ], + "type": "text", + "content": "Figure 1. Generating images from rare tail classes in the Flowers-LT with only two training images. Our proposed approach allows for a more diverse set of features such as backgrounds, colors, poses, and object layouts to be infused into the tail classes." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "training data. This is referred to as the \"long-tail (LT) problem\" [55], where most of the training data come from a few classes (referred to as the \"head\" of the distribution). In contrast, data from a large number of classes rarely occur (referred to as the \"tail\" of the distribution). This imbalance in data poses a challenge to the effective learning of tail classes, as standard machine learning algorithms tend" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7580" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 282, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 282, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 282, + 84 + ], + "type": "text", + "content": "to favor densely sampled regions of the input distribution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 86, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 86, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 86, + 287, + 277 + ], + "type": "text", + "content": "Although there has been a surge of interest in the long-tail problem in recent years, the focus has been primarily on recognition tasks [49]. Generative learning on the long-tail data, particularly cGANs which is the focus of this work, remains underexplored. Recently, [36, 37] identified mode collapse on tail classes as a result of the spectral norm explosion in class-conditional BatchNorms and the collapse of the latents in the StyleGANs " + }, + { + "bbox": [ + 46, + 86, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 86, + 287, + 277 + ], + "type": "text", + "content": " space. While techniques such as regularization, normalization, and/or classbalancing techniques [49] might improve the performance of long-tail generative models to some extent, their effectiveness is limited. The primary challenge in the long-tail problem is the under-representation of tail classes in the training data, e.g., in Fig. 1 each tail class contains merely 2 training images, resulting in insufficient observation of the tail classes by the learning algorithm." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 278, + 288, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 288, + 456 + ], + "type": "text", + "content": "To address this challenge, we propose to leverage the abundant information in the head classes, with the aim to learn and infuse knowledge from the head classes into the tail ones, thereby enriching the training distribution for the latter. However, it is very difficult to explicitly disentangle the class-specific information pertaining to head classes from class-independent information shared between head and tail classes. We make a key observation that the similarity between head and tail instances is often higher at lower resolutions: discriminative features such as shape or specific texture are usually unveiled at higher resolutions. Conversely, information from the lower resolutions tends to be more class-independent such as background, configuration, or direction of objects (Fig. 6), thus can be shared between the head and tail classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 460, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 579 + ], + "type": "text", + "content": "Building upon this observation, we propose Unconditional Training at LOWER resolution (UTLO), a novel approach for training cGANs in the long-tail setup. In addition to the standard conditional GAN objective, UTLO trains the intermediate low-resolution output of the generator with an unconditional GAN objective. The unconditional GAN objective encourages the learning of low-resolution features common to both the head and tail classes, which are infused into the subsequent layers of the generator, particularly benefiting the under-represented tail classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 700 + ], + "type": "text", + "content": "Through extensive experiments and analysis on several long-tailed datasets, we demonstrate that using UTLO to combine high-resolution conditional and low-resolution unconditional training effectively facilitates knowledge sharing between head and tail classes, thereby improving the overall generative modeling of cGANs in the long-tail setup. Due to the strong class imbalance in the long-tailed data, naive usage of existing GAN metrics can be misleading. To mitigate this issue, we propose a few practices to adapt commonly-used GAN metrics to the long-tail setup." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 238, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 238, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 238, + 713 + ], + "type": "text", + "content": "Below, we highlight our main contributions:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 545, + 287 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "text", + "content": "- We propose UTLO, a novel knowledge-sharing framework tailored for training cGANs in the long-tailed setup. UTLO allows infusing information from the head classes to the tail classes via an additional unconditional objective applied to the low-resolution part of the cGAN generator. To the best of our knowledge, this work is the first to demonstrate that not all layers in a cGAN need to receive class-conditional information, i.e., a cGAN featuring a partially unconditional generator." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 180, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 545, + 228 + ], + "type": "text", + "content": "- We present a set of practices and metrics designed to adapt the commonly-used GAN evaluation metrics for long-tail setups, enabling a more precise evaluation of the image generation quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 228, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 545, + 287 + ], + "type": "text", + "content": "- Through extensive experiments across multiple benchmarks and architectures, we validate the effectiveness of our proposed method in improving the training of cGANs in the long-tail setup, achieving state-of-the-art results across several long-tail datasets." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 298, + 392, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 298, + 392, + 311 + ], + "spans": [ + { + "bbox": [ + 306, + 298, + 392, + 311 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 319, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 498 + ], + "type": "text", + "content": "Conditional Generative Adversarial Networks The initial work on conditional image generation using GANs [7] was presented by [29], in which the class conditions were concatenated to the inputs of the generator and discriminator networks. The AC-GAN [33] proposed the use of auxiliary classification in the discriminator. [2] introduced BigGAN and set a milestone in large-scale and high-resolution conditional image generation. The authors of StyleGAN [17, 18] first map the input noise and class condition to a latent style space, which is then passed to the multiple layers of the generator for image synthesis. More recently, an extension of the StyleGAN called StyleGAN-XL [41] has become the state-of-the-art in conditional image generation on several datasets and outperforms more complex and time-consuming approaches such as diffusion models [6]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 653 + ], + "type": "text", + "content": "GAN Regularization under Limited Data Training GANs under limited data is challenging as the discriminator can memorize the training samples, resulting in the collapse of the training or quality degradation of the generated images. Recently, data augmentation and regularization techniques have been incorporated into GAN training as a means to mitigate this problem [15, 17, 46, 53, 57, 58]. Additionally, [47] introduced a regularization term to the GAN objective, which tracks the predictions of the discriminator for real and generated images using separate moving averages. [21] employed off-the-shelf vision models in an ensemble of discriminators, demonstrating improved performance in both limited-data and large-scale GAN training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "In the context of few-shot image generation, FastGAN [23] proposed a lightweight architecture with a self-supervised discriminator. [34] introduced cross-domain distance consistency in order to transfer diversity from a source domain to a target domain. Subsequently, [19] proposed" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "7581" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "a latent-mixup strategy to smooth the latent space through controlled latent interpolation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 99, + 288, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 99, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 99, + 288, + 277 + ], + "type": "text", + "content": "Recently, [42, 57] have observed that class-conditioning can cause mode collapse in the limited data regime. Their proposed work learns from limited but balanced/unlabeled data, which is different from our setup: heavily imbalanced long-tail data. [42] proposed Transitional-cGAN, a training strategy that starts with unconditional GAN objective and injects class-conditional information during a transition period. This approach was found to be effective in mitigating mode-collapse in the limited data regime. Our main idea of separating pathways between low-resolution and high-resolution is novel w.r.t. [42]: to promote knowledge sharing, we modify the cGAN architecture so that the lower-resolution part of the generator is entirely unconditional while only the higher-resolution part is conditional and received class-conditional information." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 281, + 288, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 288, + 460 + ], + "type": "text", + "content": "Long-tail Recognition To address the long-tail recognition problem [49, 55], previous research has primarily focused on techniques such as class re-balancing [3, 28, 38], learning algorithm and model design [13, 14, 59], and information augmentation [24, 45, 48]. More related to our work, re-sampling [8, 56] has been widely used to handle class imbalance problems. This involves balancing the number of samples used during training through over-sampling and/or under-sampling. Oversampling increases the frequency of tail classes [9, 31] while under-sampling reduces the imbalance by reducing the frequency of head class instances [25, 50]. Note that our research diverges from prior studies on long-tail recognition by addressing the more challenging task of generative modeling of the long-tail data, as opposed to the conventional long-tail classification task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "content": "Training GANs on Long-tail Data Recent years have seen a growing interest in learning from long-tail data. However, most focus has been directed toward recognition tasks, with a limited number of studies addressing the development of generative models for long-tail data. GAMO [31] uses adversarial training to over-sample from minority classes. [35] introduced class-balancing regularization to the GAN objective, which leverages the predictions of a pretrained classifier to improve the generation of underrepresented classes in an unconditional setting. However, this method trains an unconditional GAN and is restricted to having access to a pre-trained classifier on the long-tail data to guide the training. The most closely related work to ours is the Group Spectral Regularization (GSR) [36] and NoisyTwins [37] regularizations. The authors identified mode collapse on tail classes as a result of the spectral norm explosion in class-conditional BatchNorms and the collapse of the latents in the StyleGANs " + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "content": " space, respectively. In contrast, we present a novel approach for sharing knowledge between head and tail classes to enhance long-tail learning, which is orthogonal to regularization techniques. In addition, while" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "[37] is restricted to StyleGANs, our framework can be extended to different GAN architectures." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 107, + 455, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 107, + 455, + 121 + ], + "spans": [ + { + "bbox": [ + 305, + 107, + 455, + 121 + ], + "type": "text", + "content": "3. Taming the Tail in cGANs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 127, + 545, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 127, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 127, + 545, + 152 + ], + "type": "text", + "content": "3.1. Background: Conditional Generative Adversarial Networks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 158, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 545, + 301 + ], + "type": "text", + "content": "Conditional Generative Adversarial Networks (cGANs) are suitable for many applications due to providing user control over the generated samples at inference time. This is particularly helpful in the long-tail setup, where it is desirable to explicitly generate samples from rare (tail) classes. In contrast, unconditional GANs are likely to generate samples that track the training distribution, resulting in most samples coming from common (head) classes. To train a cGAN over a dataset containing " + }, + { + "bbox": [ + 304, + 158, + 545, + 301 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 158, + 545, + 301 + ], + "type": "text", + "content": " instances " + }, + { + "bbox": [ + 304, + 158, + 545, + 301 + ], + "type": "inline_equation", + "content": "\\mathbb{X} = \\{\\pmb{x}_1,\\dots ,\\pmb{x}_n\\}" + }, + { + "bbox": [ + 304, + 158, + 545, + 301 + ], + "type": "text", + "content": " and their corresponding labels " + }, + { + "bbox": [ + 304, + 158, + 545, + 301 + ], + "type": "inline_equation", + "content": "\\mathbb{Y} = \\{y_{1},\\ldots ,y_{n}\\}" + }, + { + "bbox": [ + 304, + 158, + 545, + 301 + ], + "type": "text", + "content": ", we can formulate the adversarial training objective that alternatively optimizes the following loss terms," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 310, + 538, + 325 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 538, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 538, + 325 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} ^ {D} = \\mathbb {E} _ {\\boldsymbol {x}, y} [ f _ {D} (- D (\\boldsymbol {x} | y)) ] + \\mathbb {E} _ {\\boldsymbol {z}, y} [ f _ {D} (D (G (\\boldsymbol {z}, y))) ]", + "image_path": "89f030675824ecfaedceae6d78577553de26155be1bf4bfb0dbcc38efa87f0d6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 327, + 545, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 545, + 342 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} ^ {G} = \\mathbb {E} _ {\\boldsymbol {z}, y} [ f _ {G} (- D (G (\\boldsymbol {z}, y))) ] \\tag {1}", + "image_path": "453df00832bfa26d813605ac0f479ab18f701c1d9c5d694d68c9c5e810ffd0fb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "text", + "content": " are the generator and discriminator networks with their corresponding loss functions " + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "inline_equation", + "content": "f_{G}" + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "inline_equation", + "content": "f_{D}" + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "text", + "content": " [27]; " + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\pmb{x}|y\\sim p_{\\mathrm{data}}(\\pmb{x}|y)" + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "text", + "content": " is real data drawn from class " + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "text", + "content": ". The generator " + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "text", + "content": " conditions on both " + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "inline_equation", + "content": "z\\sim p_z(z)" + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "text", + "content": ", a random noise vector from a prior distribution over latent space, and " + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 351, + 545, + 471 + ], + "type": "text", + "content": ", the class conditioning vector. This adversarial training objective encourages the cGAN to reach an equilibrium between the generator and the discriminator, resulting in the generator producing realistic samples indistinguishable from real data by the discriminator." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 479, + 538, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 479, + 538, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 479, + 538, + 491 + ], + "type": "text", + "content": "3.2. Caveat of Training cGANs on Long-tail Data" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 617 + ], + "type": "text", + "content": "Training GANs on long-tail data is challenging due to the inherent difficulty in modeling the rare examples that comprise the tail of the data distribution. This skewness in the data distribution hinders balanced learning across classes. In particular, the discriminator does not see enough examples from the tail of the distribution during training, which can lead to poor discriminative signals for the generator. As a result, the generator learns to generate a small subset of the possible outputs that fools the inadequately learned discriminator – exhibiting a classic “mode collapse” scenario." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Although regularizations introduced in [36, 37] can help mitigate the collapse of learning in tail classes that contain a sufficiently large number of samples, we still observe mode collapse occurring when a very limited number of samples are in tail classes, which is inherently challenging. This can be seen in Fig. 2 for tail classes of CIFAR100-LT with as few as 5 training instances. Although early-stopping can be adopted to obtain reasonable performance, the key in boost-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7582" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 76, + 281, + 243 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 281, + 243 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 281, + 243 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 281, + 243 + ], + "type": "image", + "image_path": "b5a8d2f28344ede7ebda1ecd806c5c7e43db2d70a1b4ea0f75bfcfdff7bb5bef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 248, + 108, + 313 + ], + "blocks": [ + { + "bbox": [ + 48, + 248, + 108, + 313 + ], + "lines": [ + { + "bbox": [ + 48, + 248, + 108, + 313 + ], + "spans": [ + { + "bbox": [ + 48, + 248, + 108, + 313 + ], + "type": "image", + "image_path": "8ffed90ca950e0de7fb240c83639adec6c3eb98e8554a766c9bcc84b33693056.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 315, + 287, + 360 + ], + "lines": [ + { + "bbox": [ + 47, + 315, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 315, + 287, + 360 + ], + "type": "text", + "content": "Figure 2. Convergence of different methods on CIFAR100-LT " + }, + { + "bbox": [ + 47, + 315, + 287, + 360 + ], + "type": "inline_equation", + "content": "(\\rho = 100)" + }, + { + "bbox": [ + 47, + 315, + 287, + 360 + ], + "type": "text", + "content": ", where tail classes have as few as 5 training examples. Incorporating our framework into the baseline alleviates overfitting as a result of knowledge sharing from head to rare tail classes." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 108, + 249, + 168, + 313 + ], + "blocks": [ + { + "bbox": [ + 108, + 249, + 168, + 313 + ], + "lines": [ + { + "bbox": [ + 108, + 249, + 168, + 313 + ], + "spans": [ + { + "bbox": [ + 108, + 249, + 168, + 313 + ], + "type": "image", + "image_path": "db32d36e6b935cb2edc49caa32715f170289f72b0b7654b8866f9873146b8867.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 249, + 227, + 313 + ], + "blocks": [ + { + "bbox": [ + 168, + 249, + 227, + 313 + ], + "lines": [ + { + "bbox": [ + 168, + 249, + 227, + 313 + ], + "spans": [ + { + "bbox": [ + 168, + 249, + 227, + 313 + ], + "type": "image", + "image_path": "f621dd8311f986bceac51b1440c3dac89ae767d025c387d52ce32b7eb67a7541.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 227, + 249, + 287, + 313 + ], + "blocks": [ + { + "bbox": [ + 227, + 249, + 287, + 313 + ], + "lines": [ + { + "bbox": [ + 227, + 249, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 227, + 249, + 287, + 313 + ], + "type": "image", + "image_path": "dfe73fe5f9a38b618e16649cb2fceeedf5d855d97c7662c26767554f88425339.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 407 + ], + "type": "text", + "content": "ing the performance would be an approach to infuse additional diversities into the tail classes rather than only relying on techniques such as regularization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 418, + 287, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 287, + 442 + ], + "type": "text", + "content": "3.3. Knowledge Sharing via Unconditional Training at Lower Resolutions (UTLO)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "content": "We devise a novel generative adversarial network (GAN) training scheme for long-tail data that aims to share knowledge from predominant head classes to rare tail classes, assuming some of the tail and head classes are at least coarsely similar. We propose a framework that uses unconditional learning in low-resolution images / features at an intermediate layer of the generator using the unconditional GAN objective. This helps with learning universal features from both head and tail classes. The subsequent layers, responsible for introducing finer details at higher resolutions, are trained using a class-conditional GAN objective." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "In our proposed framework, we perform conditional training on the high-resolution image output from the final layer of the generator network. This is built upon the features learned at lower layers through the use of an unconditional objective, inheriting features primarily from head classes. The conditioning on the class labels gives control at the inference time to explicitly generate images from tail classes, meeting our design desideratum. The combination of unconditional and conditional objectives using our proposed method enables knowledge sharing, which, in turn, improves the quality of GAN training on long-tail data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": "Our framework is general and applicable to many GAN architectures. In this section, we illustrate the application of our framework on StyleGAN2 with adaptive data augmentation (ADA) [17], a state-of-the-art and solid baseline for training GANs, particularly in the limited data regime. In the following, we demonstrate the necessary modifications to the architecture of the generator and discriminator networks. Although we show the necessary modifications for StyleGAN2-ADA, our method can be easily extended to other GAN architectures such as FastGAN [23], which is also used in our experiments (see Section 4)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 217, + 446, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 217, + 446, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 446, + 228 + ], + "type": "text", + "content": "3.3.1 Modifying the Generator" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "text", + "content": "In the generator design of the StyleGAN2-ADA [17, 18], the class-conditional information, combined with the latent vector, is first embedded in the style space " + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "text", + "content": " using the style-mapping network " + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{map}}" + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "text", + "content": ". The style vectors are then broadcasted across the layers of the synthesis network " + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{syn}}" + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "text", + "content": " in order to generate diverse images. To meet our design desideratum, we dissect the synthesis network into two subnetworks " + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{syn}} = G_h \\circ G_l" + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "inline_equation", + "content": "G_l" + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "text", + "content": " represents the earlier part of the network that produces intermediate features and/or images at a low resolution " + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "inline_equation", + "content": "L \\times L" + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "inline_equation", + "content": "G_h" + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "text", + "content": " on the other hand, represents the latter part of " + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{syn}}" + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "text", + "content": " that generates the output image at the high resolution " + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "inline_equation", + "content": "H \\times H" + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "inline_equation", + "content": "H > L" + }, + { + "bbox": [ + 304, + 235, + 545, + 378 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "text", + "content": "To block the flow of class-conditional information to the lower layers of " + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{syn}}" + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "text", + "content": ", we generate separate " + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "text", + "content": " vectors, one containing the class-conditional embeddings, referred to as " + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "inline_equation", + "content": "w_{z,y}" + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "text", + "content": ", while for the other one, referred to as " + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "inline_equation", + "content": "w_{z}" + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "text", + "content": ", class information was set to zero. Note that both vectors share the same latent variable " + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "text", + "content": " and mapping network " + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{map}}" + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "text", + "content": " parameters, and the only difference is the presence of class-conditional information. The lower layers of the synthesis network " + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "inline_equation", + "content": "G_{l}" + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "text", + "content": " are conditioned on " + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "inline_equation", + "content": "w_{z}" + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "text", + "content": " while the subsequent layers at higher resolutions receive " + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "inline_equation", + "content": "w_{z,y}" + }, + { + "bbox": [ + 304, + 379, + 546, + 498 + ], + "type": "text", + "content": " (see Fig. 3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "text", + "content": "Generators typically follow a network design that gradually increases the resolution of intermediate features and/or images as the network progresses. This allows us to select a desired low-resolution, such as " + }, + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "text", + "content": ", during training. Recent generator designs often incorporate skip connections [16] or residual connections [30] to improve gradient flow. The generator of StyleGAN2-ADA uses skip connections, which explicitly generates intermediate RGB images at each layer. We take this low-resolution image as the input to the discriminator. Note that in the case of a generator design that uses residual connections, a " + }, + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "text", + "content": " convolutional layer can be incorporated to convert the residual channels to RGB channels." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "In a forward pass through the generator, images at low and high resolutions can be generated simultaneously without additional overhead. By inputting a latent code " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "z \\sim p_{z}(z)" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": " and a target label " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": ", we can obtain an unconditional synthetic image " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{x}_l \\in \\mathbb{R}^{3 \\times L \\times L}" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": " along with a conditional" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7583" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 69, + 486, + 198 + ], + "blocks": [ + { + "bbox": [ + 111, + 69, + 486, + 198 + ], + "lines": [ + { + "bbox": [ + 111, + 69, + 486, + 198 + ], + "spans": [ + { + "bbox": [ + 111, + 69, + 486, + 198 + ], + "type": "image", + "image_path": "8e61a754a7a7b98da3145e27a96323c25a3973fae410e01287b27271836d3112.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": "Figure 3. The proposed framework, UTLO, illustrated for the StyleGAN2-ADA architecture. Low and high resolution image pathways are used for unconditional and class-conditional objectives, respectively. " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": " is the input latent code, " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": " indicates the class embeddings, and " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": " is a constant input. Separate style vectors, " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "w_{z}" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": " (class-independent) and " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "w_{z,y}" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": " (class-conditional), are generated using the same " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": " and a shared style-mapping network " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{map}}" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": " which are then passed to " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "G_{l}" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "G_{h}" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": ", respectively. The high-resolution generated image " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "\\hat{x} \\in \\mathbb{R}^{3\\times H\\times H}" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": " is passed through the discriminator " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "D = D_{l} \\circ D_{h}" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": " to calculate the conditional objective " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": " while the low-resolution image " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "\\hat{x}_l \\in \\mathbb{R}^{3\\times L\\times L}" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": " is passed only through " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "D_{l}" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": " to calculate the unconditional objective " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{uc}" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": ". The final objective " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": " is the combination of the two. While " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "D_{l}" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": " is shared, two separate prediction heads (FC layers) are used for unconditional and conditional objectives. The fromRGB is designed to increase the dimensionality of RGB channels to match the input channels of the " + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "inline_equation", + "content": "D_{l}" + }, + { + "bbox": [ + 46, + 205, + 547, + 295 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 302, + 288, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 288, + 352 + ], + "type": "text", + "content": "high-resolution synthetic image " + }, + { + "bbox": [ + 46, + 302, + 288, + 352 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}\\in \\mathbb{R}^{3\\times H\\times H}" + }, + { + "bbox": [ + 46, + 302, + 288, + 352 + ], + "type": "text", + "content": " at the output. While " + }, + { + "bbox": [ + 46, + 302, + 288, + 352 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}" + }, + { + "bbox": [ + 46, + 302, + 288, + 352 + ], + "type": "text", + "content": " is trained to be from class " + }, + { + "bbox": [ + 46, + 302, + 288, + 352 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 302, + 288, + 352 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 302, + 288, + 352 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_l" + }, + { + "bbox": [ + 46, + 302, + 288, + 352 + ], + "type": "text", + "content": " has no class-specific constraints. At inference time, the intermediate images can be discarded." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 359, + 203, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 359, + 203, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 359, + 203, + 371 + ], + "type": "text", + "content": "3.3.2 Modifying the Discriminator" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 378, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 496 + ], + "type": "text", + "content": "Opposite to the generator design, discriminators gradually reduce the resolution of the intermediate features as the network progresses. In order to pass images to an intermediate layer of the discriminator, we dissect it into two sub-networks. Formally, we define the discriminator as " + }, + { + "bbox": [ + 46, + 378, + 287, + 496 + ], + "type": "inline_equation", + "content": "D = D_{l} \\circ D_{h}" + }, + { + "bbox": [ + 46, + 378, + 287, + 496 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 378, + 287, + 496 + ], + "type": "inline_equation", + "content": "D_{h}" + }, + { + "bbox": [ + 46, + 378, + 287, + 496 + ], + "type": "text", + "content": " represents the earlier part of the original network that takes the images at training resolution " + }, + { + "bbox": [ + 46, + 378, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{3 \\times H \\times H}" + }, + { + "bbox": [ + 46, + 378, + 287, + 496 + ], + "type": "text", + "content": ". On the other hand, " + }, + { + "bbox": [ + 46, + 378, + 287, + 496 + ], + "type": "inline_equation", + "content": "D_{l}" + }, + { + "bbox": [ + 46, + 378, + 287, + 496 + ], + "type": "text", + "content": " represents the latter part of the discriminator taking a lower resolution input " + }, + { + "bbox": [ + 46, + 378, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C \\times L \\times L}" + }, + { + "bbox": [ + 46, + 378, + 287, + 496 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 378, + 287, + 496 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 378, + 287, + 496 + ], + "type": "text", + "content": " channels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "type": "text", + "content": "The default discriminator design in StyleGAN2-ADA uses residual connections. This does not allow direct passing of RGB inputs to the intermediate layers. To overcome this, we add a " + }, + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "type": "text", + "content": " convolutional layer, referred to as \"fromRGB\", that increases the channels of the RGB images to match " + }, + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "type": "text", + "content": ", the number of input channels of " + }, + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "type": "inline_equation", + "content": "D_{l}" + }, + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 677 + ], + "type": "text", + "content": "More concretely, " + }, + { + "bbox": [ + 46, + 570, + 288, + 677 + ], + "type": "inline_equation", + "content": "D_{l}" + }, + { + "bbox": [ + 46, + 570, + 288, + 677 + ], + "type": "text", + "content": " is a shared network in both the unconditional low-resolution and class-conditional high-resolution image pathways (see Fig. 3). To obtain separate unconditional and conditional predictions, we use two different fully-connected (FC) layers at the final layer of " + }, + { + "bbox": [ + 46, + 570, + 288, + 677 + ], + "type": "inline_equation", + "content": "D_{l}" + }, + { + "bbox": [ + 46, + 570, + 288, + 677 + ], + "type": "text", + "content": ". The output of the " + }, + { + "bbox": [ + 46, + 570, + 288, + 677 + ], + "type": "inline_equation", + "content": "D_{h}" + }, + { + "bbox": [ + 46, + 570, + 288, + 677 + ], + "type": "text", + "content": " and the low-res image " + }, + { + "bbox": [ + 46, + 570, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{x}}_l" + }, + { + "bbox": [ + 46, + 570, + 288, + 677 + ], + "type": "text", + "content": " are passed through the " + }, + { + "bbox": [ + 46, + 570, + 288, + 677 + ], + "type": "inline_equation", + "content": "D_{l}" + }, + { + "bbox": [ + 46, + 570, + 288, + 677 + ], + "type": "text", + "content": " which are then used for calculating the class-conditional loss " + }, + { + "bbox": [ + 46, + 570, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 570, + 288, + 677 + ], + "type": "text", + "content": " and unconditional loss " + }, + { + "bbox": [ + 46, + 570, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{uc}" + }, + { + "bbox": [ + 46, + 570, + 288, + 677 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "Note, when passing real images to " + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "D_{l}" + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": " for the unconditional loss, they are first downsized to the low-resolution " + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "L \\times L" + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": " using bi-linear interpolation to be comparable with" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 303, + 546, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 546, + 351 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 546, + 351 + ], + "type": "text", + "content": "the ones generated from " + }, + { + "bbox": [ + 304, + 303, + 546, + 351 + ], + "type": "inline_equation", + "content": "G_{l}" + }, + { + "bbox": [ + 304, + 303, + 546, + 351 + ], + "type": "text", + "content": ". They are then passed through the fromRGB layer and consequently " + }, + { + "bbox": [ + 304, + 303, + 546, + 351 + ], + "type": "inline_equation", + "content": "D_{l}" + }, + { + "bbox": [ + 304, + 303, + 546, + 351 + ], + "type": "text", + "content": ". For the class-conditional loss, the original images are passed to " + }, + { + "bbox": [ + 304, + 303, + 546, + 351 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 303, + 546, + 351 + ], + "type": "text", + "content": " without any modification." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 354, + 444, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 444, + 367 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 444, + 367 + ], + "type": "text", + "content": "3.3.3 Final Training Objective" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 373, + 546, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 546, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 546, + 409 + ], + "type": "text", + "content": "The final objective of our framework is simply a combination of the conditional and unconditional GAN losses, using a weighting parameter " + }, + { + "bbox": [ + 304, + 373, + 546, + 409 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 373, + 546, + 409 + ], + "type": "text", + "content": ". To put it formally we have," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 338, + 415, + 546, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 415, + 546, + 430 + ], + "spans": [ + { + "bbox": [ + 338, + 415, + 546, + 430 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {D} = \\mathcal {L} _ {c} ^ {D} + \\lambda \\cdot \\mathcal {L} _ {u c} ^ {D}, \\mathcal {L} ^ {G} = \\mathcal {L} _ {c} ^ {G} + \\lambda \\cdot \\mathcal {L} _ {u c} ^ {G} \\tag {2}", + "image_path": "bdd373dd038986dfc7ff4059f963a68630d773a58023870990c8cbabc1bdd6dc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c^D" + }, + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c^G" + }, + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "text", + "content": " represent the class-conditional losses for the discriminator and generator, respectively (as outlined in Eq. 3.1), and " + }, + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{uc}^D" + }, + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{uc}^G" + }, + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "text", + "content": " represent their unconditional counterparts, where " + }, + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "text", + "content": " is sampled from " + }, + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{data}}(\\pmb{x})" + }, + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{data}}(\\pmb{x}|y)" + }, + { + "bbox": [ + 304, + 437, + 547, + 510 + ], + "type": "text", + "content": ". The overall framework of our proposed method is depicted in Fig. 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 520, + 388, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 520, + 388, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 520, + 388, + 533 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 540, + 356, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 356, + 553 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 356, + 553 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": "Datasets. We use 6 different long-tailed datasets in our experiments: CIFAR10 and CIFAR100[20], LSUN [51], Flowers [32], iNaturalist2019 [11], and AnimalFaces [44]. This selection is intended to encompass a wide spectrum of image domains, dataset sizes, resolutions, and imbalance ratios (denoted as " + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": " represents the ratio of the number of training samples between the classes with the most and the least number of examples. To ensure the tail classes stay few-shot, we keep the number of images in the smallest tail classes under 50 [49]. A detailed description of these datasets is provided in Supp. A." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": "Baselines. In order to demonstrate the effectiveness of our proposed method, we use various cGAN architectures cov" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7584" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "ering different generator and discriminator designs and data augmentation pipelines:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 96, + 287, + 179 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 47, + 96, + 287, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 96, + 287, + 143 + ], + "spans": [ + { + "bbox": [ + 47, + 96, + 287, + 143 + ], + "type": "text", + "content": "- StyleGAN2 with Adaptive Data Augmentation (ADA) [17]. We also integrate Transitional training [42], Group Spectral Regularization (GSR) [36], and NoisyTwins [37] to this baseline." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 144, + 287, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 144, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 47, + 144, + 287, + 179 + ], + "type": "text", + "content": "- Projected GAN (PGAN) [40]: we use projected discriminator with Differentiable Augmentation (DA) [57]. We also add GSR [36] to this baseline." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 181, + 287, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 287, + 205 + ], + "type": "text", + "content": "For comprehensive details on the baselines, hyperparameters, and implementation please refer to Supp. F." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 209, + 270, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 209, + 270, + 221 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 270, + 221 + ], + "type": "text", + "content": "4.2. Evaluation Metrics For Long-Tail Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 224, + 287, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 287, + 343 + ], + "type": "text", + "content": "Evaluating the images generated by GANs on a long-tailed setup poses challenges, primarily due to the imbalanced data and access to a limited number of samples for the tail classes. For our experiments, we employ widely-used GAN metrics: Fréchet Inception Distance (FID) [10], and Kernel Inception Distance (KID) [1]. Following [10, 17], we report all metrics by computing statistics between " + }, + { + "bbox": [ + 46, + 224, + 287, + 343 + ], + "type": "inline_equation", + "content": "50k" + }, + { + "bbox": [ + 46, + 224, + 287, + 343 + ], + "type": "text", + "content": " generated images and all available training images. In the following, we present a set of practices for adapting the aforementioned metrics to the long-tail setup." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 344, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 344, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 287, + 486 + ], + "type": "text", + "content": "First, when a larger and (more) balanced dataset following the long-tail training data distribution is available, it is used for metric calculation, e.g., full CIFAR10 before artificial imbalance. For naturally imbalanced datasets such as iNaturalist2019, we use the unmodified training set to calculate metrics. Note that when generating images for evaluation, we sample from the same distribution as the available real dataset, which may be imbalanced. While FID and KID indicate how well the real data distribution matches the generated images across all classes, we suggest additional metrics to mitigate the disproportionate influence of head classes in data statistic calculations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 487, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 654 + ], + "type": "text", + "content": "In response to the varied number of training instances across classes in a long-tail setup, we propose evaluating metrics specifically on few-shot (FS) categories [49]. We calculate the FID and KID for the few-shot subset and refer to them as \"FID-FS\" and \"KID-FS\", respectively. This is tailored to evaluate the quality of generated samples in the tail classes. Contrasting with the standard FID/KID, we maintain an equal number of real images across all classes during our FID-FS/KID-FS calculation. This emphasizes learning quality on tail classes, irrespective of any imbalances that might be present in the few-shot subset. Reporting both standard and few-shot metrics provides a more comprehensive evaluation of long-tail learning, considering the performance of both head and tail classes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 658, + 105, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 105, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 105, + 670 + ], + "type": "text", + "content": "4.3. Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Below, we report the results obtained across different benchmarks and cGAN architectures. We pick the best model for reporting results as the one with the lowest FID" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 307, + 109, + 544, + 187 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 103 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 103 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 103 + ], + "type": "text", + "content": "Table 1. Our proposed method, UTLO, outperforms the baselines in terms of quantitative image quality metrics on the AnimalFaces-LT dataset (64 × 64 resolution)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 109, + 544, + 187 + ], + "lines": [ + { + "bbox": [ + 307, + 109, + 544, + 187 + ], + "spans": [ + { + "bbox": [ + 307, + 109, + 544, + 187 + ], + "type": "table", + "html": "
MethodsFID ↓FID-FS ↓KID ↓KID-FS ↓ ×1000
StyleGAN2-ADA UnCond. [17]39.4104.117.327.6
StyleGAN2-ADA [17]51.487.124.735.9
+ Transitional [42]62.199.038.548.9
+ GSR[36]39.267.221.232.7
+ NoisyTwins [37]29.450.216.721.2
+ UTLO (Ours)26.248.412.619.6
", + "image_path": "626f4b8be4441d4ef817a8a56269e8a7cd570dd8a4ee28e69e72f18789b515e5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 196, + 545, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 196, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 304, + 196, + 545, + 232 + ], + "type": "text", + "content": "FS from two independent runs. Note that our method exhibits significantly less reliance on early stopping compared to the baselines, as illustrated in Fig. 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 232, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 316 + ], + "type": "text", + "content": "AnimalFaces-LT This benchmark has a naturally occurring imbalance across its classes. We compare our method against the StyleGAN2-ADA benchmarks in terms of quantitative image quality metrics in Table. 1. Our method outperforms the baselines across all metrics, as demonstrated in the results. For a visual comparison against the baselines, please refer to Supp. H." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": "Ablation on the Choice of Low-resolution for Unconditional Training (" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "res_{\\mathrm{uc}}" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": "). In our proposed method, one of the hyperparameter choices is to select an intermediate low-resolution " + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "res_{\\mathrm{uc}}" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": " for unconditional training. All layers with equal or lower resolution than " + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "res_{\\mathrm{uc}}" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": " do not receive class-conditional information. An unconditional GAN objective is added over the images and/or features at " + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "res_{\\mathrm{uc}}" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": ". To study the impact of " + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "res_{\\mathrm{uc}}" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": ", we conducted an ablation study on the AnimalFaces-LT dataset, containing images at " + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": ". Table. 2 presents the results for selecting " + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "res_{\\mathrm{uc}}" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": " from resolutions lower than " + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": ", namely " + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": ". Studying the obtained results, we observe that resolutions of " + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": " and 16 achieve relatively close performance. Conversely, the performance diminishes when layers up to " + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": " are trained unconditionally. This is anticipated as the AnimalFaces-LT is at " + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": ", leaving only one up-sampling layer to learn the class-conditional information which is shown to be insufficient. In all our experiments, we use " + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "res_{\\mathrm{uc}}" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": " unless stated otherwise. Additional visual analysis on the role of " + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "res_{\\mathrm{uc}}" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": " as well as an ablation on the choice of unconditional training objective weight " + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 317, + 546, + 604 + ], + "type": "text", + "content": " (see Eq. 2), need for unconditional layers in the discriminator, and distinction from \"coarse-to-fine\" training strategies are presented in the Supp. C." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 342, + 635, + 507, + 684 + ], + "blocks": [ + { + "bbox": [ + 305, + 612, + 545, + 634 + ], + "lines": [ + { + "bbox": [ + 305, + 612, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 305, + 612, + 545, + 634 + ], + "type": "text", + "content": "Table 2. Ablation on choice of low resolution for unconditional training " + }, + { + "bbox": [ + 305, + 612, + 545, + 634 + ], + "type": "inline_equation", + "content": "(res_{\\mathrm{uc}})" + }, + { + "bbox": [ + 305, + 612, + 545, + 634 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 342, + 635, + 507, + 684 + ], + "lines": [ + { + "bbox": [ + 342, + 635, + 507, + 684 + ], + "spans": [ + { + "bbox": [ + 342, + 635, + 507, + 684 + ], + "type": "table", + "html": "
resucFID ↓FID-FS ↓KID ↓KID-FS ↓ ×1000
8 × 826.248.412.619.6
16 × 1627.550.313.720.8
32 × 3238.064.923.334.3
", + "image_path": "e7ae853e3efe14146179baee13f165141f1dc424c854feef213faa5baac3dec6.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Can any low-resolution be the starting point for highres generations? In general, our observations indicate that" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7585" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "high-res images are guided, yet not entirely restricted by the low-res images, as shown in Fig. 4. Each row in this figure presents a set of different high-resolution images that are generated from the same unconditional low-resolution image. This illustration reveals that high-resolution images can substantially differ from low-resolution ones, both in terms of color and texture. This suggests that subsequent conditional blocks are capable of significant modification in the background or foreground." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "type": "text", + "content": "Importance of FS metrics: comparison against unconditional training. We also include results from training an unconditional StyleGAN2-ADA model [17] in Table 1, referred to as StyleGAN2-ADA UnCond. The unconditional model generates samples that follow the training distribution, which is mainly dominated by head classes. This bias favors unconditional learning in terms of the FID/KID metrics, which do not consider the skewness in the data distribution. Nevertheless, it can be observed the UnCond. baselines significantly perform worse in terms of few-shot metrics FID-FS/KID-FS. This suggests solely relying on FID/KID can be misleading. We strongly recommend including few-shot metrics, i.e., FID-FS and KID-FS, when evaluating cGANs on long-tailed datasets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 350, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 289, + 483 + ], + "type": "text", + "content": "CIFAR10-LT In this benchmark, our method is compared with various baselines with imbalanced ratios " + }, + { + "bbox": [ + 46, + 350, + 289, + 483 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 350, + 289, + 483 + ], + "type": "text", + "content": " of 50 and 100. This indicates the least frequent tail class includes only 100 and 50 training instances, respectively. Table. 3 quantitatively assesses the quality of the samples generated by our method and the baselines. All the baselines use the StyleGAN2-ADA [17] architecture. Our proposed method consistently outperforms the baselines across all metrics. This demonstrates that beyond improving generative learning across all classes (FID & KID), UTLO can also notably improve the learning from tail classes (FID-FS & KID-FS)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 484, + 289, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 289, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 289, + 592 + ], + "type": "text", + "content": "Effect of Imbalance Ratio " + }, + { + "bbox": [ + 46, + 484, + 289, + 592 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 484, + 289, + 592 + ], + "type": "text", + "content": ". Not surprisingly, all methods demonstrate improved performance with lower values of " + }, + { + "bbox": [ + 46, + 484, + 289, + 592 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 484, + 289, + 592 + ], + "type": "text", + "content": ", as seen in Table. 3. Further, we notice the benefits of knowledge sharing via UTLO become more pronounced as the imbalance in the dataset increases, i.e., learning tail classes becomes more challenging. It is also worth mentioning that the FID-FS and KID-FS metrics are critical to represent the quality of the few-shot classes as those are under-represented in the dataset-wide metrics that as" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 604, + 285, + 693 + ], + "blocks": [ + { + "bbox": [ + 50, + 604, + 285, + 693 + ], + "lines": [ + { + "bbox": [ + 50, + 604, + 285, + 693 + ], + "spans": [ + { + "bbox": [ + 50, + 604, + 285, + 693 + ], + "type": "image", + "image_path": "3943075715ef5bc49e78c3680a9f262e33371ff05544fb0c5c983e9f3116dd1a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 695, + 286, + 717 + ], + "lines": [ + { + "bbox": [ + 47, + 695, + 286, + 717 + ], + "spans": [ + { + "bbox": [ + 47, + 695, + 286, + 717 + ], + "type": "text", + "content": "Figure 4. Different class-conditional images generated given the same unconditional low-resolution images (left-most column)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 131, + 545, + 247 + ], + "blocks": [ + { + "bbox": [ + 304, + 71, + 547, + 126 + ], + "lines": [ + { + "bbox": [ + 304, + 71, + 547, + 126 + ], + "spans": [ + { + "bbox": [ + 304, + 71, + 547, + 126 + ], + "type": "text", + "content": "Table 3. Quantitative comparison of the generated images from UTLO against baselines on CIFAR10-LT dataset with different imbalance ratios " + }, + { + "bbox": [ + 304, + 71, + 547, + 126 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 71, + 547, + 126 + ], + "type": "text", + "content": ". UTLO shows substantial improvements on few-shot (FS) metrics compared to StyleGAN2-ADA [17] while outperforming existing methods including NoisyTwins [37]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 131, + 545, + 247 + ], + "lines": [ + { + "bbox": [ + 307, + 131, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 307, + 131, + 545, + 247 + ], + "type": "table", + "html": "
ρMethodsFID ↓FID-FS ↓KID ↓KID-FS ↓ ×1000
50StyleGAN2-ADA [17]6.521.42.49.0
+ Transitional [42]9.417.74.78.6
+ GSR [36]6.421.32.38.1
+ NoisyTwins [37]6.212.22.45.0
+ UTLO (ours)6.111.82.44.8
100StyleGAN2-ADA [17]9.024.24.09.7
+ Transitional [42]11.320.65.49.2
+ GSR [36]8.424.33.911.8
+ NoisyTwins [37]7.114.12.95.9
+ UTLO (Ours)6.813.42.85.4
", + "image_path": "1fc6ccf5edfaeb26f45dace050c479aec7898a13728dec981d3ee90010bc7ef6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 254, + 547, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 547, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 547, + 302 + ], + "type": "text", + "content": "sume balanced sampling. For instance, at " + }, + { + "bbox": [ + 304, + 254, + 547, + 302 + ], + "type": "inline_equation", + "content": "\\rho = 50" + }, + { + "bbox": [ + 304, + 254, + 547, + 302 + ], + "type": "text", + "content": ", the FID and KID metrics obtained from UTLO show little changes, while FID-FS and KID-FS exhibit improvements close to " + }, + { + "bbox": [ + 304, + 254, + 547, + 302 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 254, + 547, + 302 + ], + "type": "text", + "content": " in comparison to GSR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 304, + 546, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 546, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 546, + 472 + ], + "type": "text", + "content": "Knowledge-sharing at Low Resolutions. As explained in Sec. 3, our proposed generator builds on top of low-resolution (e.g., " + }, + { + "bbox": [ + 304, + 304, + 546, + 472 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 304, + 304, + 546, + 472 + ], + "type": "text", + "content": ") unconditionally trained images which are subsequently used to generate conditional images at higher resolutions (e.g., " + }, + { + "bbox": [ + 304, + 304, + 546, + 472 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 304, + 546, + 472 + ], + "type": "text", + "content": "). Several low-resolution images are shown in Fig. 6 along with their high-resolution conditional images generated for the head and tail classes of the dataset. It can be seen that conditional images generated from the tail classes evidently share certain features from the head classes. This demonstrates that knowledge sharing at low resolutions is an effective approach for infusing information from head classes to tail ones. A quantitative analysis of knowledge-sharing is presented in Supp. D. For additional qualitative examples see Supp. H." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "type": "text", + "content": "CIFAR100-LT In addition to CIFAR10-LT, we evaluate our method and various baselines on the more challenging CIFAR100-LT dataset, wherein the tail classes include as few as five training instances. The quantitative results, shown in Table. 4, reveal that our method outperforms the baselines in a benchmark where a large number of classes with high diversity are present. In Fig. 2, we demonstrate the FID-FS curve during the course of the training along with the generated samples from the tail classes of the CIFAR100-LT dataset. This shows the effectiveness of our proposed method in addressing mode collapse while regularization methods diverge and require early stopping. Further visual comparisons of our proposed method against the baselines for CIFAR100-LT can be found in Supp. H." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": "LSUN5-LT We include LSUN5-LT as a higher resolution " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "(128 \\times 128)" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": " and highly-imbalanced " + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "inline_equation", + "content": "(\\rho = 1000)" + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": " benchmark. In addition, we use Projected GAN with FastGAN[26] generator and DA [40]) as the baseline model. This is a different generator and discriminator design compared to StyleGAN2 along with a different aug-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7586" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 70, + 497, + 208 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 497, + 208 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 497, + 208 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 497, + 208 + ], + "type": "image", + "image_path": "53b872f9bca3a0fb6d14f9a056b4e5f8007f9be82edcc09e257c54b1baf3a5a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 547, + 234 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 547, + 234 + ], + "type": "text", + "content": "Figure 5. Generated images from LSUN5-LT dataset. Despite only 50 training instances for the tail class kitchen, the proposed UTLO framework produces diverse, high-fidelity images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 244, + 286, + 322 + ], + "blocks": [ + { + "bbox": [ + 48, + 244, + 286, + 322 + ], + "lines": [ + { + "bbox": [ + 48, + 244, + 286, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 244, + 286, + 322 + ], + "type": "image", + "image_path": "b7713869e15067a4a9545890df6157f99a4a8b0f945bda653bd9787fd68d4f8b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 323, + 287, + 400 + ], + "lines": [ + { + "bbox": [ + 46, + 323, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 287, + 400 + ], + "type": "text", + "content": "Figure 6. Knowledge sharing from head to tail classes in CIFAR10-LT dataset " + }, + { + "bbox": [ + 46, + 323, + 287, + 400 + ], + "type": "inline_equation", + "content": "(\\rho = 100)" + }, + { + "bbox": [ + 46, + 323, + 287, + 400 + ], + "type": "text", + "content": " using UTLO. The conditional images generated from the head (middle columns) and tail (right columns) classes share and are built on top of the same low-resolution (unconditional) images (left columns). Low-resolution images " + }, + { + "bbox": [ + 46, + 323, + 287, + 400 + ], + "type": "inline_equation", + "content": "(8 \\times 8)" + }, + { + "bbox": [ + 46, + 323, + 287, + 400 + ], + "type": "text", + "content": " are upsampled to that of CIFAR10-LT " + }, + { + "bbox": [ + 46, + 323, + 287, + 400 + ], + "type": "inline_equation", + "content": "(32 \\times 32)" + }, + { + "bbox": [ + 46, + 323, + 287, + 400 + ], + "type": "text", + "content": " for improved visualization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 47, + 427, + 286, + 502 + ], + "blocks": [ + { + "bbox": [ + 47, + 403, + 287, + 426 + ], + "lines": [ + { + "bbox": [ + 47, + 403, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 287, + 426 + ], + "type": "text", + "content": "Table 4. Comparison of our proposed UTLO method with StyleGAN2-ADA baselines on the CIFAR100-LT dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 427, + 286, + 502 + ], + "lines": [ + { + "bbox": [ + 47, + 427, + 286, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 427, + 286, + 502 + ], + "type": "table", + "html": "
MethodsFID ↓FID-FS ↓KID ↓KID-FS ↓
×1000
StyleGAN2-ADA [17]10.824.95.19.3
+ Transitional [42]10.623.74.28.5
+ GSR[36]11.125.05.08.2
+ NoisyTwins [37]10.122.55.07.9
+ UTLO (Ours)9.921.84.67.5
", + "image_path": "46c11c06eaf4c3be3653ed8d917e5277b9dfff7d259b00082bc951b8f6c51796.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 508, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 616 + ], + "type": "text", + "content": "mentation method. In conditional FastGAN implementation, the class-specific information is injected using class-conditional batch normalization, differing from the style mapping network utilized in StyleGAN2. UTLO can be readily adapted to this design by using standard batch normalization layers at low resolutions and utilizing class-conditional batch normalization only at higher resolutions. However, it is noteworthy that methods such as [37] are restricted to StyleGANs and cannot be applied in this context." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 616, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 288, + 700 + ], + "type": "text", + "content": "Table. 5 compares the results obtained from our method against the baseline and with the addition of GSR [36]. Our method consistently surpasses the baselines by around " + }, + { + "bbox": [ + 46, + 616, + 288, + 700 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 616, + 288, + 700 + ], + "type": "text", + "content": " across all metrics. As previously shown for the CIFAR10-LT (Table 3), UTLO not only improved on the tail classes but also significantly improved the dataset-wide FID/KID due to the high imbalance ratio " + }, + { + "bbox": [ + 46, + 616, + 288, + 700 + ], + "type": "inline_equation", + "content": "(\\rho = 1000)" + }, + { + "bbox": [ + 46, + 616, + 288, + 700 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "content": "Qualitative results in Fig. 5 contrast the generated" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 271, + 544, + 323 + ], + "blocks": [ + { + "bbox": [ + 305, + 244, + 545, + 267 + ], + "lines": [ + { + "bbox": [ + 305, + 244, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 305, + 244, + 545, + 267 + ], + "type": "text", + "content": "Table 5. Evaluating the quality of generated images from the proposed method and comparing against baselines on LSUN5-LT." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 271, + 544, + 323 + ], + "lines": [ + { + "bbox": [ + 308, + 271, + 544, + 323 + ], + "spans": [ + { + "bbox": [ + 308, + 271, + 544, + 323 + ], + "type": "table", + "html": "
MethodsFID ↓FID-FS ↓KID ↓KID-FS ↓ ×1000
PGAN (FastGAN)+DA [40]15.060.24.652.7
+ GSR [36]15.763.75.758.0
+ UTLO (Ours)10.943.63.535.3
", + "image_path": "aee78efda9f27f54a9388f3d2f85f13203c65346c21f24f22de75b9b3ba58564.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 331, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 403 + ], + "type": "text", + "content": "images from our method against the baselines across all classes. For the kitchen class which only includes 50 training examples, the baselines struggle to effectively learn from the tail class whereas UTLO succeeds in generating diverse and high-quality images. More visual examples from our method and baselines are presented in Supp. H." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 404, + 546, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 546, + 477 + ], + "type": "text", + "content": "Due to space constraints, additional results and analysis are provided in the supplementary material. e.g., Analysis of the role of weighted sampling (Supp. B.), The distinction of training class-conditional GANs in the Long-tailed setup vs. Limited-data setup (Supp. E.), and additional evaluation on Flowers-LT and iNaturalist2019 datasets (Supp. G)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 484, + 379, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 484, + 379, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 484, + 379, + 498 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 503, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 659 + ], + "type": "text", + "content": "In this paper, we proposed UTLO, a novel framework designed to improve the training of cGANs on long-tailed data. Inspired by the observation that head and tail classes often have more similarities at lower resolutions, our method facilitates knowledge sharing from head to tail classes using unconditional training at lower resolutions. The proposed method enriches the limited training distribution of the tail classes and effectively addresses mode collapse, leading to significant improvement in image generation for tail classes. We have also introduced the FID-FS/KID-FS metric, an adaptation of widely-used GAN metrics, specifically tailored for tail classes. We hope our findings are useful for future work in long-tail learning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 668, + 409, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 668, + 409, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 668, + 409, + 681 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "This work is partially supported by the National Science Foundation (NSF) under grant 1751412 and 1927564." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "7587" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "type": "text", + "content": "[2] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 233 + ], + "type": "text", + "content": "[3] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 280 + ], + "type": "text", + "content": "[4] Jiequan Cui, Shu Liu, Zhuotao Tian, Zhisheng Zhong, and Jiaya Jia. Result: Residual learning for long-tailed recognition. *IEEE transactions on pattern analysis and machine intelligence*, 2021. 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 281, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 287, + 346 + ], + "type": "text", + "content": "[5] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge J. Belongie. Class-balanced loss based on effective number of samples. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 9268-9277. Computer Vision Foundation / IEEE, 2019. 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 347, + 287, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 287, + 412 + ], + "type": "text", + "content": "[6] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 8780-8794, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 414, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 468 + ], + "type": "text", + "content": "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, A aron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2014. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 469, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 287, + 523 + ], + "type": "text", + "content": "[8] Hao Guo and Song Wang. Long-tailed multi-label visual recognition by collaborative training on uniform and rebalanced samplings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15089-15098, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 287, + 568 + ], + "type": "text", + "content": "[9] Hui Han, Wen-Yuan Wang, and Bing-Huan Mao. Borderline-smote: a new over-sampling method in imbalanced data sets learning. In International conference on intelligent computing, pages 878-887. Springer, 2005. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "type": "text", + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "text", + "content": "[11] Grant Van Horn and Oisin Mac Aodha. The inaturalist 2019 competition dataset. https://www.kaggle.com/c/inaturalist-2019-fgvc6.5, 12" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "text", + "content": "[13] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 184 + ], + "type": "text", + "content": "[14] Bingyi Kang, Yu Li, Sai Nan Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 186, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 260 + ], + "type": "text", + "content": "[15] Tejan Karmali, Rishubh Parihar, Susmit Agrawal, Harsh Rangwani, Varun Jampani, Maneesh Kumar Singh, and R. Venkatesh Babu. Hierarchical semantic regularization of latent spaces in stylegans. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 443-459. Springer, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 263, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 545, + 328 + ], + "type": "text", + "content": "[16] Animesh Karnewar and Oliver Wang. MSG-GAN: multiscale gradients for generative adversarial networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 7796-7805. Computer Vision Foundation / IEEE, 2020. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "type": "text", + "content": "[17] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. 1, 2, 4, 6, 7, 8, 12, 13, 16, 17, 18" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 387, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 387, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 545, + 453 + ], + "type": "text", + "content": "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 8107-8116. Computer Vision Foundation / IEEE, 2020. 2, 4, 18" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 454, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 454, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 454, + 545, + 510 + ], + "type": "text", + "content": "[19] Chaerin Kong, Jeesoo Kim, Donghoon Han, and Nojun Kwak. Few-shot image generation with mixup-based distance learning. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 563-580. Springer, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 510, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 510, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 510, + 545, + 533 + ], + "type": "text", + "content": "[20] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. preprint, 2009. 5, 12" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "text", + "content": "[21] Nupur Kumari, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Ensembling off-the-shelf models for gan training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10651-10662, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 579, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 654 + ], + "type": "text", + "content": "[22] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 657, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 657, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 657, + 545, + 689 + ], + "type": "text", + "content": "[23] Bingchen Liu, Yizhe Zhu, Kunpeng Song, and A. Elgammal. Towards faster and stabilized gan training for high-fidelity few-shot image synthesis. iclr, 2021. 2, 4, 18" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "text", + "content": "[24] Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, and Wenhui Li. Deep representation learning on long-tailed data:" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7588" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "A learnable embedding augmentation perspective. Computer Vision And Pattern Recognition, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 288, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 288, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 288, + 128 + ], + "type": "text", + "content": "[25] Shudong Liu and Ke Zhang. Under-sampling and feature selection algorithms for s2smlp. IEEE Access, 8:191803-191814, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "text", + "content": "[26] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "type": "text", + "content": "[27] Mario Lucic, Karol Kurach, Marcin Michalski, Sylvain Gelly, and Olivier Bousquet. Are gans created equal? a large-scale study. Advances in neural information processing systems, 31, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 288, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 272 + ], + "type": "text", + "content": "[28] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 304 + ], + "type": "text", + "content": "[29] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv: Arxiv-1411.1784, 2014. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 288, + 338 + ], + "type": "text", + "content": "[30] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. ICLR, 2018. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 338, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 372 + ], + "type": "text", + "content": "[31] S. S. Mullick, Shounak Datta, and Swagatam Das. Generative adversarial minority oversampling. IEEE International Conference On Computer Vision, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 372, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 288, + 415 + ], + "type": "text", + "content": "[32] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 416, + 288, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 288, + 471 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 288, + 471 + ], + "type": "text", + "content": "[33] Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier gans. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, pages 2642-2651. PMLR, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 472, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 472, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 288, + 536 + ], + "type": "text", + "content": "[34] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A. Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10743-10752, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 537, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 591 + ], + "type": "text", + "content": "[35] Harsh Rangwani, Konda Reddy Mopuri, and R. Venkatesh Babu. Class balancing gan with a classifier in the loop. In Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence, pages 1618-1627. PMLR, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 647 + ], + "type": "text", + "content": "[36] Harsh Rangwani, Naman Jaswani, Tejan Karmali, Varun Jampani, and R. Venkatesh Babu. Improving gans for long-tailed data through group spectral regularization. European Conference On Computer Vision, 2022. 2, 3, 6, 7, 8, 12, 13, 16, 18, 19" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "text", + "content": "[37] Harsh Rangwani, Lavish Bansal, Kartik Sharma, Tejan Karmali, Varun Jampani, and R Venkatesh Babu. Noisytwins: Class-consistent and diverse image generation through stylegans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5987-5996, 2023. 2, 3, 6, 7, 8, 18, 19" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 715 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": "[38] Jiawei Ren, Cunjun Yu, shunan sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and hongsheng Li. Balanced meta-softmax for long-tailed visual recognition. In Advances in Neural Information Processing Systems, pages 4175–4186. Curran Associates, Inc., 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 546, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 171 + ], + "type": "text", + "content": "[39] Shibani Santurkar, Ludwig Schmidt, and A. Madry. A classification-based study of covariate shift in gan distributions. International Conference On Machine Learning, 2017. 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 172, + 546, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 546, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 546, + 216 + ], + "type": "text", + "content": "[40] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster. Advances in Neural Information Processing Systems, 34:17480-17492, 2021. 6, 7, 8, 18" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 217, + 546, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 546, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 546, + 281 + ], + "type": "text", + "content": "[41] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleganxI: Scaling stylegan to large diverse datasets. In SIGGRAPH '22: Special Interest Group on Computer Graphics and Interactive Techniques Conference, Vancouver, BC, Canada, August 7 - 11, 2022, pages 49:1-49:10. ACM, 2022. 1, 2, 15, 16" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 282, + 546, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 546, + 336 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 546, + 336 + ], + "type": "text", + "content": "[42] Mohamad Shahbazi, Martin Danelljan, Danda Pani Paudel, and Luc Van Gool. Collapse by conditioning: Training class-conditional GANs with limited data. In International Conference on Learning Representations, 2022. 1, 3, 6, 7, 8, 16, 18" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 337, + 546, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 337, + 546, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 546, + 360 + ], + "type": "text", + "content": "[43] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 360, + 546, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 546, + 403 + ], + "type": "text", + "content": "[44] Zhangzhang Si and Song-Chun Zhu. Learning hybrid image templates (hit) by information projection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(7):1354-1367, 2012. 5, 12" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 404, + 546, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 546, + 459 + ], + "type": "text", + "content": "[45] Kaihua Tang, Jianqiang Huang, and Hanwang Zhang. Long-tailed classification by keeping the good and removing the bad momentum causal effect. In Advances in Neural Information Processing Systems, pages 1513-1524. Curran Associates, Inc., 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 459, + 546, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 546, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 546, + 503 + ], + "type": "text", + "content": "[46] Ngoc-Trung Tran, Viet-Hung Tran, Ngoc-Bao Nguyen, Trung-Kien Nguyen, and Ngai-Man Cheung. On data augmentation for gan training. IEEE Transactions on Image Processing, 30:1882-1897, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 503, + 546, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 503, + 546, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 503, + 546, + 557 + ], + "type": "text", + "content": "[47] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7921-7931, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 558, + 546, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 558, + 546, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 558, + 546, + 613 + ], + "type": "text", + "content": "[48] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X. Yu. Long-tailed recognition by routing diverse distribution-aware experts. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 614, + 546, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 546, + 647 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 546, + 647 + ], + "type": "text", + "content": "[49] Lu Yang, He Jiang, Qing Song, and Jun Guo. A survey on long-tailed visual recognition. International Journal of Computer Vision, 130:1837-1872, 2022. 2, 3, 5, 6, 12" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 647, + 546, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 546, + 691 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 546, + 691 + ], + "type": "text", + "content": "[50] Baofeng Yao and Lei Wang. An improved under-sampling imbalanced classification algorithm. In 2021 13th International Conference on Measuring Technology and Mechatronics Automation (ICMTMA), pages 775-779. IEEE, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 546, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 546, + 715 + ], + "type": "text", + "content": "[51] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7589" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 464 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 5, 12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 161 + ], + "type": "text", + "content": "[52] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 205 + ], + "type": "text", + "content": "[53] Han Zhang, Zizhao Zhang, Augustus Odena, and Honglak Lee. Consistency regularization for generative adversarial networks. In International Conference on Learning Representations, 2019. 2, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 288, + 251 + ], + "type": "text", + "content": "[54] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. IEEE/cvf Conference On Computer Vision And Pattern Recognition, 2018. 16" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 285 + ], + "type": "text", + "content": "[55] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. arXiv preprint arXiv: Arxiv-2110.04596, 2021. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 287, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 330 + ], + "type": "text", + "content": "[56] Zizhao Zhang and Tomas Pfister. Learning fast sample re-weighting without reward data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 725-734, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "text", + "content": "[57] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. Advances in Neural Information Processing Systems, 33:7559-7570, 2020. 2, 3, 6, 12, 16, 18" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 376, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 408 + ], + "type": "text", + "content": "[58] Zhengli Zhao, Zizhao Zhang, Ting Chen, Sameer Singh, and Han Zhang. Image augmentations for gan training. arXiv preprint arXiv: 2006.02595, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 464 + ], + "type": "text", + "content": "[59] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 16489-16498. Computer Vision Foundation / IEEE, 2021. 3" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7590" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/b33baf4c-8c46-4e13-a308-0177b1e2eee6_content_list.json b/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/b33baf4c-8c46-4e13-a308-0177b1e2eee6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1525fd01cd74fff7929503849a3f12515b057888 --- /dev/null +++ b/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/b33baf4c-8c46-4e13-a308-0177b1e2eee6_content_list.json @@ -0,0 +1,1609 @@ +[ + { + "type": "text", + "text": "Targeted Representation Alignment for Open-World Semi-Supervised Learning", + "text_level": 1, + "bbox": [ + 84, + 130, + 885, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruixuan Xiao $^{1}$ , Lei Feng $^{2}$ , Kai Tang $^{1}$ , Junbo Zhao $^{1}$ , Yixuan Li $^{3}$ , Gang Chen $^{1}$ , Haobo Wang $^{1*}$", + "bbox": [ + 125, + 179, + 864, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Zhejiang University, China", + "bbox": [ + 380, + 199, + 602, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Singapore University of Technology and Design, Singapore", + "bbox": [ + 254, + 215, + 730, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3University of Wisconsin-Madison, USA", + "bbox": [ + 330, + 234, + 653, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xiaoruixuan,tk0819,j.zhao,cg,wanghaobo}@zju.edu.cn,lfengqaq@gmail.com,sharonli@cs.wisc.edu", + "bbox": [ + 86, + 253, + 890, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-world Semi-Supervised Learning aims to classify unlabeled samples utilizing information from labeled data, while unlabeled samples are not only from the labeled known categories but also from novel categories previously unseen. Despite the promise, current approaches solely rely on hazardous similarity-based clustering algorithms and give unlabeled samples free rein to spontaneously group into distinct novel class clusters. Nevertheless, due to the absence of novel class supervision, these methods typically suffer from the representation collapse dilemma—features of different novel categories can get closely intertwined and indistinguishable, even collapsing into the same cluster and leading to degraded performance. To alleviate this, we propose a novel framework TRAILER which targets to attain an optimal feature arrangement revealed by the recently uncovered neural collapse phenomenon. To fulfill this, we adopt targeted prototypes that are pre-assigned uniformly with maximum separation and then progressively align the representations to them. To further tackle the potential downsides of such stringent alignment, we encapsulate a sample-target allocation mechanism with coarse-to-fine refinery that is able to infer label assignments with high quality. Extensive experiments demonstrate that TRAILER outperforms current state-of-the-art methods on generic and fine-grained benchmarks. The code is available at https://github.com/Justherozen/TRAILER.", + "bbox": [ + 76, + 335, + 473, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 758, + 209, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The resounding achievements of deep learning heavily rely on large-scale and accurately-annotated training data, which is labor-intensive and time-consuming. To alleviate this, semi-supervised learning (SSL) has emerged as a prominent alternative [4, 55, 60], which exploits unlabeled data for boosted performance. Despite the success, current SSL", + "bbox": [ + 75, + 784, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2b29116574fc6e9c93ae3ff3977346e698f8b29a8a029afeaadff6616eb96bc.jpg", + "image_caption": [ + "Figure 1. Left: Scenario of open-world SSL where unlabeled samples are from known and novel classes. Right: Visualization without and with our targeted calibration mechanism. Without calibration, representations tend to get jumbled and inseparable, as in the left where dog and cat categories converge into the same cluster." + ], + "image_footnote": [], + "bbox": [ + 521, + 304, + 870, + 449 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods are primarily based on the closed-world assumption that labeled and unlabeled data share the same predefined label space. Nevertheless, this assumption rarely holds in real-world applications, where the model may encounter novel classes previously unseen in unlabeled data. To relax this assumption, novel category discovery (NCD) was first proposed [15, 21, 22] which assumes all unlabeled data strictly belong to disjoint novel classes. Recently, a more practical setting, open-world semi-supervised learning (open-world SSL), also referred to as generalized category discovery (GCD), has received huge attention [5, 46, 50] by generalizing standard SSL and NCD. It assumes that unlabeled data contain both known classes from the labeled data and similar yet distinct novel classes.", + "bbox": [ + 496, + 564, + 892, + 775 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The goal of open-world SSL is to classify the already-known classes from the labeled data, as well as to discover new classes in unlabeled data without annotations and assign instances to them. Nascent efforts have been made to address this practical yet under-investigated task [20, 33, 41, 50]. The essence of these recent works is to rashly rely on the similarity-based [5, 41] or contrastive-based [46] clustering algorithm in feature space to spontaneously", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 1, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23072", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "group different samples into separate clusters. Some methods further integrate self-training techniques with generated pseudo-labels for boosted performance [20, 33]. However, due to the missing supervision and semantic shift for novel categories, the self-organized clustering step can get overfitted and biased towards the known classes, leading to a representation collapse dilemma—the feature distance between one novel category and another category may get too close and finally become inseparable, even converging within a single cluster and thus jeopardizing generalization performance, as illustrated in the left of Figure 1.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, a recent study has unveiled a feature-classifier alignment phenomenon termed neural collapse (NC) [38]. It indicates that under ideal conditions, the last-layer features of the same class tend to collapse into their within-class mean. Meanwhile, these within-class means of all classes will be aligned with classifiers and formed as a simplex equiangular tight frame (ETF) at the end of training, as elaborated in Section 3.1. Intuitively, such an elegant NC phenomenon affords an optimal feature structure with minimized within-class variance and maximized between-class variance [36], coinciding with what we fervently desire. Nevertheless, in open-world SSL, the absence of novel class annotation can hinder the representation learning and thus disrupt the natural induction of NC.", + "bbox": [ + 75, + 257, + 472, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In an effort to salvage this and accomplish the appealing feature arrangement of NC, we propose a novel framework, Targeted Representation AIIgnment for open-world semisupervised LEaRning (dubbed TRAILER). Key to our method, TRAILER leverages a targeted classifier to progressively align embeddings to the pre-assigned ETF structure, eventually fulfilling the optimal state of uniformly and maximally separated embeddings. Despite the promise, this calibration procedure is still susceptible to false alignment stemming from unreliable pseudo-labels. To mitigate this, we integrate a hierarchical sample-target allocation strategy that comprises two sequential steps: (i)-rough sample assignment based on the optimal transport mechanism; (ii)-label refinery by casting the known-novel separation problem to a specific weakly-supervised paradigm termed PU learning [2, 13, 62]. Such an allocation mechanism can provide more precise pseudo-labels and thus facilitate representation calibration. Through our visualized results in Figure 1 and Section 4.3, TRAILER does indeed induce discriminative and non-skewed embeddings. Empirically, TRAILER surpasses state-of-the-art approaches on different evaluation benchmarks, e.g., improves the best baseline by $3.1\\%$ , $2.9\\%$ , and $1.5\\%$ overall accuracy on CIFAR-10, CIFAR-100 and ImageNet-100 datasets, respectively.", + "bbox": [ + 75, + 468, + 472, + 833 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 844, + 218, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Closed-World Semi-Supervised Learning has been widely explored [4, 44, 55, 66] to learn from limited la", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "beled and massive unlabeled data. Pseudo labeling [31, 42] and consistency regularization [30, 44, 55] are two prominent techniques. The former is based on self-training that utilizes model predictions as optimization targets [40, 56], while the latter enforces model to produce consistent outputs for different views of the same image [3, 4, 43, 55]. However, these methods are primarily based on the closed-world assumption that labeled and unlabeled data share the same pre-defined label space, which limits their real-world application where unlabeled samples from previously unseen classes may emerge. Open-set SSL [19, 24, 25, 45] relaxes this by simply rejecting all unseen class instances, but still fails to classify them into distinct novel categories.", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-World Semi-Supervised Learning further accommodates real-world scenarios [5, 20, 46]. It is also known as generalized category discovery (GCD) [50] and is a natural extension of novel category discovery (NCD). NCD is first introduced in [21] and improved in later works [15, 22] to discover new categories by leveraging the knowledge from labeled known categories. However, NCD assumes that all unlabeled samples belong to novel classes and thus fails to recognize known class samples. Most recently, open-world SSL generalizes NCD to identify unlabeled samples from both known and novel classes [5]. To tackle this, several incipient solutions have been proposed [20, 59], which typically involve similarity-based representation learning and then involve sample clustering to generate pseudo-labels for self-training. ORCA [5] is the pioneering work that integrates an uncertainty-based margin loss to control the intraclass variance of different classes. OpenCon [46] further derives a semi-supervised constrastive mechanism to produce compact feature space. Despite the promise, these methods solely rely on spontaneous clustering based on sample similarity, overlooking potential risks that representations of different classes possibly converge into the same cluster, thus leading to the representation collapse dilemma.", + "bbox": [ + 496, + 306, + 893, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Collapse describes an elegant alignment phenomenon [38], which states that the last-layer features will collapse into their within-class centers and these centers together with classifiers will eventually form a simplex equiangular tight frame (ETF). Stemming from the appealing symmetry, a plethora of studies have been triggered to explain this phenomenon theoretically [26, 48, 58, 65]. NC is proved to be the global optimality under ideal conditions with the CE [18, 26] and the MSE [23, 64] loss functions. The NC phenomenon has also been investigated in some specific scenarios like imbalanced learning [54, 57, 63] noisy label learning [37, 52], transfer learning [17] and incremental learning [58]. In our work, we investigate the opportunities of inducing such optimal feature arrangement of NC through targeted calibration for open-world SSL.", + "bbox": [ + 496, + 674, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23073", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/206d0955d948800be5240746976bb41019e819af1a2cd14fbb43518183211c80.jpg", + "image_caption": [ + "Figure 2. Overview of TRAILER. For representation calibration, we leverage the targeted classifier $\\phi^{etf}$ to align the feature towards the optimal arrangement of ETF progressively while maintaining the within-class compactness with prototypical clustering. For sample-target allocation, the unlabeled samples are provided with reliable pseudo-labels with rough assignment and subsequent label refinery." + ], + "image_footnote": [], + "bbox": [ + 116, + 89, + 330, + 303 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/08e30f5355d1bb8d4a427b7ffb323661198a6beb378faa3dffb95265494d562f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 88, + 852, + 304 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 382, + 169, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Definition. In open-world SSL, we assume that the training dataset $\\mathcal{D} = \\mathcal{D}_l \\cup \\mathcal{D}_u$ comprises two subsets: a labeled set $\\mathcal{D}_l = \\{x_i, y_i\\}_{i=1}^m$ where the input image $x_i \\in \\mathbb{R}^d$ and the ground-truth label $y_i$ come from a set of known classes $\\mathcal{C}_{known}$ , and an unlabeled set $\\mathcal{D}_u = \\{x_i\\}_{i=1}^n$ where the potential ground-truth label is from a set of classes $\\mathcal{C}_{all}$ . In closed-world SSL, the unlabeled samples strictly come from known classes, i.e., $\\mathcal{C}_{all} = \\mathcal{C}_{known}$ . As for open-world SSL which is more practical, the unlabeled samples can come from either known classes $\\mathcal{C}_{known}$ or novel classes, i.e., $\\mathcal{C}_{known} \\subset \\mathcal{C}_{all}$ . We consider $\\mathcal{C}_{novel} = \\mathcal{C}_{all} \\backslash \\mathcal{C}_{known}$ as the set of novel classes. Following previous setting [5, 15, 46], we assume the number of novel classes $|\\mathcal{C}_{novel}|$ is known apriori, which can also be estimated using the off-the-shelf methods [21, 50] as discussed in Section 4.3.", + "bbox": [ + 75, + 407, + 473, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning Goal. For classification problem with $K = |\\mathcal{C}_{all}|$ classes, open-world SSL aims to obtain a model $f_{\\theta,\\phi} : \\mathbb{R}^d \\mapsto \\mathbb{R}^K$ that recognizes samples from both known and novel classes based on $\\mathcal{D}_l$ and $\\mathcal{D}_u$ . Specifically, $f_{\\theta,\\phi} = h_\\phi \\circ g_\\theta$ can be decomposed into a representation backbone $g_\\theta$ and a classification head $h_\\phi$ with weights $\\phi = [\\phi_1, \\dots, \\phi_k] \\in \\mathbb{R}^{P \\times K}$ , for which we adopt a prototypical one following [15] with logit $h(z) = \\mathrm{Norm}(z)^\\top \\mathrm{Norm}(\\phi)$ . Formally, given the weakly and strongly augmented view $x_i$ and $x_i'$ of the same image in batch $\\mathcal{B}$ , we define $z_i, z_i' \\in \\mathbb{R}^P$ as the last-layer $P$ -dimensional features and $p_i = \\mathrm{softmax}(h(z_i))$ as output probability. Normalization for features is denoted as $\\mathrm{Norm}(z) = z / ||z||_2$ and omitted in the subsequent.", + "bbox": [ + 76, + 654, + 472, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Warm-up Phase. Similar to previous works [5, 33], we also include a commonly adopted warm-up strategy at", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the start of training. The overall objectives to cluster known classes and discover potential novel samples consist of three losses: 1) the cross-entropy loss (CE) $\\mathcal{L}_{ce}$ on labeled samples $\\mathcal{D}_l$ which is elaborated later in Section 3.1; 2) the self-supervised contrastive loss $\\mathcal{L}_{con} = -\\log \\frac{\\exp(z_i^\\top z_i' / \\tau)}{\\sum_{j=1}^n \\exp(z_i^\\top z_j' / \\tau)}$ on all samples; and 3) the entropy regularization $\\mathcal{L}_{ent} = \\sum_{k=1}^{K} \\overline{p}^k \\log (\\overline{p}^k)$ for preventing the trivial solution of assigning all instances to the same class, where $\\overline{p} = \\frac{1}{|\\mathcal{B}|} \\sum_{x_i \\in \\mathcal{B}} p_i$ is the average probability of the batch. The overall losses for warm-up are formulated,", + "bbox": [ + 496, + 383, + 893, + 547 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {w a r m}} = \\mathcal {L} _ {c e} + \\mathcal {L} _ {\\text {c o n}} + \\alpha \\mathcal {L} _ {\\text {e n t}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 556, + 890, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha$ is a weighting parameter. After warm-up, we also would like to group unlabeled samples into distinct clusters and assign them pseudo-labels for self-training. As mentioned earlier, the crucial challenge is how to abstain from the representation collapse dilemma and obtain discriminative features without novel class annotations. To fulfill this, we first draw inspiration from neural collapse phenomenon.", + "bbox": [ + 496, + 582, + 893, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Neural Collapse for Representation Calibration", + "text_level": 1, + "bbox": [ + 498, + 695, + 890, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent works have disclosed the Neural Collapse phenomenon [17, 38] which corresponds to an optimal feature-classifier alignment towards a simplex equiangular tight frame (ETF). This offers a uniformly and equally separated feature arrangement that is desirable for open-world SSL.", + "bbox": [ + 496, + 718, + 893, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1 (Simplex ETF) A Simplex equiangular tight frame (ETF) refers to a collection of $K$ equal-length and maximally-equiangular $P$ -dimensional embedding vectors $\\mathbf{E} = [e_1, \\dots, e_K] \\in \\mathbb{R}^{P \\times K}$ which satisfies:", + "bbox": [ + 496, + 801, + 893, + 863 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} = \\sqrt {\\frac {K}{K - 1}} \\mathbf {U} \\left(\\mathbf {I} _ {K} - \\frac {1}{K} \\mathbf {1} _ {K} \\mathbf {1} _ {K} ^ {\\top}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 869, + 890, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23074", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{I}_K$ is the identity matrix, $\\mathbf{1}_K$ is an all-ones vector, and $\\mathbf{U} \\in \\mathbb{R}^{P \\times K}$ ( $P \\geq K$ ) allows a rotation.", + "bbox": [ + 76, + 90, + 472, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All vectors in a simplex ETF $\\mathbf{E}$ have an equal $\\ell_2$ norm and the same pair-wise maximal equiangular angle $-\\frac{1}{K - 1}$ [57],", + "bbox": [ + 76, + 128, + 468, + 162 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne _ {k _ {1}} ^ {\\top} e _ {k _ {2}} = \\frac {K}{K - 1} \\delta_ {k _ {1}, k _ {2}} - \\frac {1}{K - 1}, \\forall k _ {1}, k _ {2} \\in [ 1, K ] \\quad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 176, + 468, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\delta_{k_1,k_2} = 1$ when $k_{1} = k_{2}$ and 0 otherwise. The NC phenomenon can be described by the following properties under ideal conditions: (1) The last-layer features converge to their within-class means, and the within-class means of all classes converge to different vertices of a simplex ETF. (2) The normalized classifier weights converge to the same simplex ETF. Notably, this process is implicit and natural since it is induced with common optimizers such as SGD, rather than enforced by an explicit objective.", + "bbox": [ + 76, + 212, + 468, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As discussed above, NC has provided an optimal ETF structure for feature arrangement with minimal within-class covariance and maximum between-class covariance (i.e., maximized Fisher discriminant ratio [16]), which coincides with our goal of deriving discriminative and non-skewed embeddings. However, NC is delicate and can fail to be naturally induced under imperfect conditions [37, 57]. For open-world SSL, the missing novel class annotations can also lead to biased features and break the NC optimality.", + "bbox": [ + 75, + 348, + 468, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To remedy this, we resort to effectuating an explicit NC and fulfilling the appealing feature arrangement with representation calibration, which is achieved by leveraging a targeted prototypical classifier $h_{\\text{etf}}$ . In specific, $h_{\\text{etf}}$ is preassigned with weights $\\phi^{\\text{etf}} \\in \\mathbb{R}^{P \\times K}$ , which is a random simplex ETF with optimal structure as Definition 1 and is fixed rather than learnable during training. By assuming a known label $\\hat{y}_i$ for each sample $x_i$ , the standard CE loss $\\mathcal{L}_{ce}$ is then replaced as $\\mathcal{L}_{\\text{etf}}$ , which is formulated,", + "bbox": [ + 75, + 484, + 468, + 621 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e t f} \\left(x _ {i}, \\hat {y} _ {i}\\right) = - \\log \\frac {\\exp \\left(z _ {i} ^ {\\top} \\cdot \\phi_ {\\hat {y} _ {i}} ^ {e t f} / \\tau\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(z _ {i} ^ {\\top} \\cdot \\phi_ {k} ^ {e t f} / \\tau\\right)} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 627, + 468, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature parameter. Through $\\mathcal{L}_{\\text{etf}}$ , the feature $z_{i}$ is drawn closer to its corresponding target $\\phi_{\\hat{y}_i}^{\\text{etf}}$ and pulled away from other class vectors of ETF structure.", + "bbox": [ + 76, + 685, + 467, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the labeled sample $(x_{i},y_{i})$ , we directly set $\\hat{y}_i = y_i$ . As for unlabeled data without ground-truth labels, we design a sample-target allocation strategy to generate pseudolabels $\\hat{y}$ , whose details are shown in Section 3.2.", + "bbox": [ + 76, + 732, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Progressive Alignment. Notably, if we directly optimize the loss $\\mathcal{L}_{\\text{etf}}$ after the warm-up phase, the provisionally clustered samples would be forcibly dragged to the corresponding targeted vector $\\phi_{\\hat{y}}^{\\text{etf}}$ . As a result, the already-learned clustering compactness can get compromised and the within-cluster feature distribution can get jumbled.", + "bbox": [ + 76, + 809, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this, we intend to align the representations in a progressive manner to preserve the compactness within each cluster during calibration. Formally, given sample $x_{i}$ with pseudo-label $\\hat{y}_i$ , we also maintain the authentic prototypes $\\mu_k = \\mathrm{Norm}(\\mathrm{Avg}_{\\hat{y}_i = k}(z_i))$ for each class $1 \\leq k \\leq K$ . To preserve the compactness of cluster structure, a prototypical clustering loss is also integrated,", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p r o t o}} \\left(z _ {i}, \\hat {y} _ {i}\\right) = - \\log \\frac {\\exp \\left(z _ {i} ^ {\\top} \\cdot \\mu_ {\\hat {y} _ {i}} / \\tau\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(z _ {i} ^ {\\top} \\cdot \\mu_ {k} / \\tau\\right)} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 207, + 890, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall classification loss is then formulated,", + "bbox": [ + 500, + 255, + 823, + 270 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} = \\lambda \\mathcal {L} _ {p r o t o} + (1 - \\lambda) \\mathcal {L} _ {e t f} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 285, + 890, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is a calibration strength parameter that is gradually ramped down from 1 to 0. With such a progressive manner, the embeddings are cautiously and smoothly drawn closer to the targeted optimal arrangement of ETF while the within cluster compactness is preserved.", + "bbox": [ + 496, + 315, + 890, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Hierarchical Sample-Target Allocation", + "text_level": 1, + "bbox": [ + 500, + 400, + 833, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In previous steps, we involve targeted classifiers to regulate representation learning. Nevertheless, our stringent calibration still holds unexpected risks. Lacking supervision, the calibration process for unlabeled samples largely relies on the quality of generated pseudo-labels. When the pseudolabels are unreliable, the feature alignment procedure can be misguided towards the wrong cluster. To overcome this, we introduce our hierarchical sample-target allocation mechanism with course-to-fine assignment in what follows.", + "bbox": [ + 496, + 422, + 890, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rough Assignment through Optimal Transport. For pseudo-label generation, one may adopt model prediction $\\arg \\max_{1\\leq k\\leq K}h^{k}(z_{i})$ directly. However, this may lead to degenerate solutions where data points of different categories map into a single category [1], thus resulting in representation collapse. To mitigate this, we adopt a widely-used equipartition constraint [6, 15] for rough label assignment, which is attained by solving an optimal transport problem.", + "bbox": [ + 496, + 579, + 890, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, given a batch $\\mathcal{B}$ of $b$ unlabeled samples with the logit matrix $\\mathbf{P} = [h(z_1),\\dots,h(z_b)]$ whose columns are output logits by $h_{\\mathrm{etf}}$ and the label-assignment matrix $\\mathbf{Q} = [q_1,\\dots,q_b]\\in [0,1]^{K\\times b}$ whose columns are the to-beassigned soft labels $q_{i}$ . $\\mathbf{Q}$ is estimated by,", + "bbox": [ + 496, + 698, + 890, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {Q} = \\max _ {\\mathbf {Q} \\in \\Gamma} \\operatorname {T r} (\\mathbf {Q} ^ {\\top} \\mathbf {P}) + \\epsilon \\mathrm {H} (\\mathbf {Q}) \\\\ \\text {s . t .} \\Gamma = \\left\\{\\mathbf {Q} \\in \\mathbb {R} _ {+} ^ {K \\times b} \\mid \\mathbf {Q} \\mathbf {1} _ {b} = \\frac {1}{K} \\mathbf {1} _ {k}, \\mathbf {Q} ^ {\\top} \\mathbf {1} _ {K} = \\frac {1}{b} \\mathbf {1} _ {b} \\right\\} \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 779, + 890, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{H}(\\cdot)$ is the entropy function, $\\epsilon$ is a hyperparameter and $\\operatorname{Tr}(\\cdot)$ is the trace function. The solution of Eq. (7) is obtained using the Sinkhorn-Knopp algorithm [11]. Please refer to Appendix A for more optimization details.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23075", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Label Refinery with Known-Novel Separation. After the rough assignment, we obtain unpolished soft labels $q \\in [0,1]^K$ . However, the aforementioned skewed predictions towards known classes stemming from overfitting remain unresolved. To rectify this, we prioritize the subtask of precisely demarcating known and novel classes for label refinery, which is achieved by including an auxiliary binary head $h_{aux}$ and reformulating this binary task as the classical positive unlabeled (PU) learning problem [2, 13, 62].", + "bbox": [ + 75, + 90, + 467, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PU learning tackles a special case of binary classification where the model has access to a labeled set containing merely positive samples and an unlabeled set with both positive and negative samples. Thus, the task of distinguishing known and novel classes can be intuitively converted to PU learning by regarding known and novel as positive and negative respectively. A flurry of methods has been designed for PU learning [10, 27, 34], among which we adopt a prior-free variational algorithm [8] with loss $\\mathcal{L}_{aux}$ to optimize the newly devised $h_{aux}$ . See Appendix A for more details.", + "bbox": [ + 75, + 227, + 467, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After that, we elicit the binary prediction $\\hat{y}^{aux}$ for $h_{aux}$ , where 1 and 0 indicate prediction of known and novel respectively. Then we leverage $\\hat{y}^{aux}$ to refine the label $q$ ,", + "bbox": [ + 76, + 378, + 467, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq ^ {j} = \\left\\{ \\begin{array}{l l} q ^ {j} & \\text {i f} \\mathbb {I} (j \\in \\mathcal {C} _ {\\text {k n o w n}}) = \\hat {y} ^ {\\text {a u x}} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 431, + 467, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $q^{j}$ is the $j$ -th entry of $q$ , $1 \\leq j \\leq K$ and $\\mathbb{I}(\\cdot)$ is the indicator function. That is, we retain the known/novel class components of soft label $q$ based on prediction $\\hat{y}^{aux}$ and discard the remaining inconsistent components. The refined $q$ are then utilized for the final hard pseudo-label generation $\\hat{y}_{i} = \\arg \\max_{1 \\leq k \\leq K} q^{k}$ in $\\mathcal{L}_{cls}$ of Eq. (6).", + "bbox": [ + 75, + 481, + 467, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pseudo-Label Filtering. While the proposed sample-target allocation mechanism can enhance the quality of pseudo-labels, the remaining obstinate noisy pseudo-labels can still impede model training, especially in the early stages when the model output is not accurate enough. To neutralize this, we further utilize a high-confidence selection to select clean pseudo-labels from noisy ones for $\\mathcal{D}_u$ . Formally, for each class $1 \\leq j \\leq K$ , we select a clean subset $\\mathcal{D}_{sel}^j$ from $\\mathcal{D}_u^j = \\{(x, \\hat{y}) \\in \\mathcal{D}_u | \\hat{y} = j\\}$ class by class with the highest confidence scores in the first $R$ percent,", + "bbox": [ + 75, + 590, + 467, + 741 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {s e l} ^ {j} = \\left\\{\\left(x, \\hat {y}\\right) \\in \\mathcal {D} _ {u} ^ {j} \\mid r a n k \\left(q ^ {j}\\right) < R \\% \\right\\} \\tag{9}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 752, + 467, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then the integral selected set is merged as $\\mathcal{D}_{sel} = \\cup_{j=1}^{K} \\mathcal{D}_{sel}^{j}$ . The classification loss in Eq. (6) is then merely valid on $\\mathcal{D}_{sel}$ . In practice, we set a small $R$ to ensure the precision of selected samples at the beginning and gradually increase it to integrate more clean samples. Finally, the overall training loss is given by,", + "bbox": [ + 75, + 780, + 467, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {a u x}} + \\mathcal {L} _ {\\text {c o n}} + \\alpha \\mathcal {L} _ {\\text {e n t}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 883, + 467, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{con}$ and $\\mathcal{L}_{ent}$ are described in Eq. (1). Through this procedure, Our sample-target allocation mechanism is able to provide pseudo-labels with high quality and guide the representation calibration toward the correct targets. The overview of TRAILER is illustrated in Figure 2.", + "bbox": [ + 498, + 90, + 890, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 181, + 622, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present the main results and part of the ablation results to verify the effectiveness of TRAILER. More experimental details and results, such as the results with less labeled data, can be found in Appendix.", + "bbox": [ + 498, + 207, + 890, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 500, + 277, + 580, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We first evaluate TRAILER on three datasets CIFAR-10, CIFAR-100 [29] and ImageNet-100, which is randomly sub-sampled with 100 classes from ImageNet [12] following [5]. For each dataset, We first divide classes into the first $50\\%$ known and the rest $50\\%$ novel classes following [5]. Then we randomly select a portion $(50\\%)$ of samples from the known classes as labeled $\\mathcal{D}_l$ , and the remaining along with all novel class samples are regarded as unlabeled $\\mathcal{D}_u$ . We also conduct experiments with different novel class ratios and on the fine-grained semantic shift benchmarks (SSB) [51], including CUB-200 [53], Stanford Cars [28], FGVC-Aircraft [35], and a naturally long-tailed Herbarium 19 [47] datasets, as discussed in Section 4.3.", + "bbox": [ + 498, + 301, + 890, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare TRAILER with various open-world SSL baselines, including ORCA [5], GCD [50], OpenCon [46], NACH [20], OpenLDN [41], GPC [61] and OpenNCD [33]. We also provide comparisons with SSL, open-set SSL, and NCD baselines following [20, 33]. For traditional and open-set SSL baselines including FixMatch [44], $\\mathrm{DS}^3\\mathrm{L}$ [19], and CGDL [45], we extend them to be applicable to novel classes by selecting samples with lower confidence scores as novel and applying $k$ -means clustering to obtain results. For NCD methods including DTC [21], RankStats [22], and $\\mathrm{UNO+}$ [15], they are extended to classify known classes by using Hungarian algorithm to match some discovered classes with known classes. For performances of baselines, we directly adopt reported results from receptive papers or from [33], except OpenLDN was retested under transductive setting. All compared methods are implemented based on pre-trained ResNet using SimCLR [9] following [5], except DTC and OpenCon which have specialized pre-training procedure. We run all experiments three times and report the averaged results.", + "bbox": [ + 498, + 517, + 890, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following [5, 33], we evaluate the transductive performance on both known and novel classes. For novel classes, we utilize the Hungarian algorithm to derive the optimal assignments and then calculate accuracy", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23076", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/234479152dab275c8f9c4fbdb13b05c185d3b8f44d9997f30d85fef57283ae3f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsCIFAR-10CIFAR-100ImageNet-100
KnownNovelAllKnownNovelAllKnownNovelAll
FixMatch [44]71.550.449.539.623.520.365.836.734.9
DS3L [19]77.645.340.255.123.724.071.232.530.8
CGDL [45]72.344.639.749.322.523.567.333.831.9
DTC [21]53.939.538.331.322.918.325.620.821.3
RankStats [22]86.681.082.936.428.423.147.328.740.3
SimCLR [9]58.363.451.728.621.122.339.535.736.9
ORCA [5]88.290.489.766.943.048.189.172.177.8
GCD [50]78.479.779.168.533.545.282.358.368.2
OpenLDN [41]92.386.888.667.340.449.474.545.755.3
OpenCon [46]90.491.189.369.147.852.790.680.883.8
NACH [20]89.592.291.368.747.052.191.075.579.6
OpenNCD [33]88.490.690.169.743.449.390.077.581.6
TRAILER (Ours)93.495.094.469.748.755.691.482.485.3
", + "bbox": [ + 112, + 88, + 859, + 319 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/74c5753d05df0f0bd5cd1b33e2be4bbe0ebbe97eb79cbb253838e98cdf3dbe2e.jpg", + "table_caption": [ + "Table 1. Accuracy comparison of known, novel, and all classes on CIFAR-10, CIFAR-100 and ImaneNet-100 dataset. The dataset is composed of $50\\%$ known classes and $50\\%$ novel classes, with $50\\%$ of the known classes labeled. Bold entries indicate superior results." + ], + "table_footnote": [], + "table_body": "
AblationCIFAR-10CIFAR-100
KnownNovelAllKnownNovelAll
TRAILER93.495.094.469.748.755.6
w/o Rep Calibration94.158.970.669.214.933.4
with λ = 093.991.892.570.246.654.6
w/o Rough Assign92.992.392.569.337.948.3
w/o Label Refinery95.289.691.570.345.453.7
w/o PL Filtering92.290.591.069.943.652.2
", + "bbox": [ + 83, + 380, + 460, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Ablation results on CIFAR-10 and CIFAR-100. 'Rep' and 'PL' indicate representation and pseudo-label respectively.", + "bbox": [ + 75, + 521, + 467, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\frac{1}{n}\\sum_{i = 1}^{n}\\mathbb{I}(y_i = \\text{perm}(\\hat{y}_i))$ where $\\text{perm}(\\cdot)$ is the matching permutation. The overall joint accuracy is also measured by Hungarian match using both known and novel classes.", + "bbox": [ + 76, + 575, + 468, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. For CIFAR experiments, we use ResNet-18 as the backbone. The model is trained for 200 epochs with warm-up of 15 epochs. We employ the Adam optimizer with a batch size of 200 and an initial learning rate of $5e^{-4}$ , which decays by a cosine scheduler. The temperature $\\tau$ is set as 0.3 and 0.5 for CIFAR-10 and CIFAR-100. We linearly ramp down $\\lambda$ from 1 to 0 and ramp up selection ratio $R\\%$ from 0.3 to 0.9 for pseudo-label filtering. The label refinery is enabled after training $h_{aux}$ for 15 epochs. For the optimization of Sinkhorn-Knopp, we adopt hyperparameters from [15] and set $\\epsilon = 0.05$ with 3 iterations. We put more implementation details for ImageNet-100 and fine-grained dataset in Appendix A.", + "bbox": [ + 75, + 641, + 468, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Reults", + "text_level": 1, + "bbox": [ + 76, + 847, + 207, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 shows the comparison results on the CIFAR-10, CIFAR-100, and ImageNet-100 datasets, including", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d80f3225708c2a7771dd47ba19da40f81c48c8372df5f7c1b33659e51222812d.jpg", + "image_caption": [ + "(a) Without calibration." + ], + "image_footnote": [], + "bbox": [ + 504, + 383, + 692, + 498 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9793e6453ee8e88732ea8c24047e8d31aa01041aaf25485ae5e180474f2aadd5.jpg", + "image_caption": [ + "(b) With calibration.", + "Figure 3. T-SNE feature visualization on CIFAR-10. Different colors represent the corresponding ground-truth classes. Without calibration, the dog and cat classes collapse into the same cluster." + ], + "image_footnote": [], + "bbox": [ + 702, + 385, + 887, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the accuracy of the 'Known', 'Novel', and 'All' classes. TRAILER significantly outperforms all the rivals by a considerable margin in terms of overall accuracy on different datasets. Specifically, on CIFAR-10, TRAILER improves upon the best baseline by $1.1\\%$ , $2.8\\%$ , and $3.1\\%$ for the 'Known', 'Novel', and 'All' classes respectively. Moreover, on the more challenging CIFAR-100 and ImageNet-100 datasets with larger label space, while most baselines demonstrate a notable performance drop, TRAILER shows robustness and reaches an extraordinary trade-off between the known and novel classes. It consistently maintains superior performance with $2.9\\%$ and $1.5\\%$ overall enhancement. These results empirically demonstrate the efficacy of our proposed TRAILER for the task of open-world SSL.", + "bbox": [ + 496, + 595, + 890, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Analysis", + "text_level": 1, + "bbox": [ + 500, + 816, + 601, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on Semantic Shift Benchmarks. We further conduct experiments on fine-grained semantic shift benchmarks (SSB) [51] following previous work [50, 61]. For fair comparisons, we adopt the same training and compar", + "bbox": [ + 498, + 839, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23077", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/eb77879e2befc76d3d760c2b0d79007a9d80afb9320522477bc69eee5ee1ac86.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsCUBStanford CarsFGVC-AircraftHerbarium 19
KnownNovelAllKnownNovelAllKnownNovelAllKnownNovelAll
k-means38.932.134.310.613.812.814.416.816.012.912.812.9
RankStats+ [22]51.624.233.361.812.128.336.422.226.955.812.827.9
UNO+ [15]49.028.135.170.518.635.556.432.240.353.714.728.3
ORCA [5]45.630.235.350.110.723.531.817.122.030.915.520.9
GCD [50]56.648.751.357.629.939.041.146.945.051.027.035.4
OpenCon [46]63.852.154.778.632.749.1---58.928.639.3
GPC [61]58.253.155.459.232.842.842.547.946.351.727.936.5
TRAILER (Ours)71.361.965.171.747.655.462.650.554.557.037.844.5
", + "bbox": [ + 109, + 88, + 861, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Results on the Semantic Shift Benchmarks. Bold entries indicate superior results and blank ones indicate results are not provided.", + "bbox": [ + 76, + 260, + 890, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee64ebf940375ccd6ac8ad37a41a8d6be115f5df96dec19ffde11e9399fc5440.jpg", + "image_caption": [ + "(a) Binary accuracy of $h_{aux}$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 299, + 269, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2672ce8cfb7404415f6ba6e8148ac5c272989e7bc84c0f88fca993d2e70743d2.jpg", + "image_caption": [ + "(b) Pseudo-label precision.", + "Figure 4. (a) Accuracy of binary prediction between known and novel classes in first 100 epochs. (b) Per-class pseudo-label precision for unlabeled samples on CIFAR-10 with and without label refinery. Class indexes 0-4 are known while 5-9 are novel." + ], + "image_footnote": [], + "bbox": [ + 277, + 299, + 467, + 415 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "isons protocol from [50] and utilize ViT-B/16 backbone [14] with DINO pre-trained weights [7]. As shown in Table 3, it can be observed that TRAILER exceeds the baselines by a substantial lead in all metrics, with improvements of $8.8\\%$ , $14.8\\%$ , $2.6\\%$ , and $9.2\\%$ for 'Novel' accuracy on these datasets. These results consistently demonstrate its superior performance, even under challenging fine-grained tasks.", + "bbox": [ + 75, + 527, + 468, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benefits of Representation Calibration. To explore the benefits of representation calibration, we conduct ablations and equip TRAILER with different representation learning strategies: 1) TRAILER w/o Rep Calibration which dismiss the feature alignment of the targeted classifier and instead utilize the vanilla learnable classifier; 2) TRAILER with $\\lambda = 0$ which discards the progressive manner for direct calibration. From Table 2, although TRAILER w/o Rep Calibration remains competitive on known classes, it suffers from a calamitous performance plunge for novel classes, which indicates the efficacy of representation calibration. TRAILER with $\\lambda = 0$ achieves decent performance, but still underperforms TRAILER perceptibly, indicating the advantage of the progressive manner.", + "bbox": [ + 75, + 657, + 467, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further visualize the representation with and without representation calibration on CIFAR-10 using t-SNE [49]", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a0d947f2818acd11e2df042fe87a7e951f7a25822e96b25c783e6ac549d63e9.jpg", + "image_caption": [ + "Figure 5. Confusion matrix between known and novel classes on CIFAR-10. TRAILER can alleviate biased predictions." + ], + "image_footnote": [], + "bbox": [ + 511, + 300, + 870, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in Figure 3 and in Appendix B. It can be shown that representation calibration facilitates inducing compact and nonskewed representations, discarding which can lead to representation collapse—the representations are biased towards known classes, where the dog class (novel, brown) and part of horse class samples (novel, grey) collapse into the cluster of the cat class (known, red), as shown in Figure 3a.", + "bbox": [ + 496, + 474, + 890, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Sample-Target Allocation. Next, we ablate the effectiveness of the sample-target allocation mechanism in Table 2. For the coarse-level assignment, we test the variant TRAILER w/o Rough Assign which sticks to the vanilla softmax function for rough label assignment of $q$ and suffers from rigorous performance degradation, especially on CIFAR-100. For the fine-level refinery, we test the variant TRAILER w/o Label Refinery which discards the refinery procedure of auxiliary head $h_{aux}$ . Although such a variant exhibits slightly better performance for the known classes, it significantly lags behind TRAILER for the novel classes.", + "bbox": [ + 496, + 598, + 890, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate it more intuitively, we visualize the accuracy of binary prediction for $h_{aux}$ between known and novel classes in Figure 4a. We can find that $h_{aux}$ quickly achieves high prediction accuracies and such accurate separation between known and novel indeed refines the rough assignment supplementarily, finally enhancing the pseudo-label precision on 9 of 10 classes, as shown in Figure 4b. Such a mechanism also alleviates the biased prediction, as shown in Figure 5 where a non-trivial amount of novel samples are", + "bbox": [ + 496, + 765, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23078", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/58b98580f4eedf51fa54f741d2dba1253968974d3bdf02a037548209fb8c581f.jpg", + "image_caption": [ + "(a) Equiangularity" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 269, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/412f447defbe88d237d77246a05c22bc5320c1a434804aeba54c83ec05d0de39.jpg", + "image_caption": [ + "(b) Maximal separation", + "Figure 6. Visualizations on the (a) equiangularity and (b) maximal separation properties of neural collapse on CIFAR-10 in the first 100 epochs. Representation calibration enables TRAILER to achieve feature arrangement with equal and maximal separation." + ], + "image_footnote": [], + "bbox": [ + 274, + 89, + 464, + 205 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3a09091dea07e146cbdad5cea38044afd4246003d82c5756cb2f4b92459b4565.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AblationEstimation of KKnownNovelAll
TRAILERKnown (100)69.748.755.6
ORCAKnown (100)66.943.048.1
TRAILEREstimated (124)71.144.353.3
ORCAEstimated (124)66.340.046.4
", + "bbox": [ + 84, + 305, + 460, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablations with unknown class number $K$ on CIFAR-100.", + "bbox": [ + 76, + 401, + 467, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "misidentified as known (the number in red) for baselines.", + "bbox": [ + 76, + 444, + 452, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Empirical Visualizations for Neural Collapse. In the previous section, we have provided the embedding visualization. Here we supply some quantitative results in addition. Recall that two core properties of NC are equiangularity and maximal separation. Given the centered class-means $\\hat{z}^k$ for features from all classes $1 \\leq k \\leq K$ , for equiangularity, we calculate the standard deviation of the cosines between each pair $\\mathrm{Std}_{k \\neq k'}(\\cos (\\hat{z}^k, \\hat{z}^{k'}))$ . For maximal separation, we calculate the average value $\\mathrm{Avg}_{k \\neq k'}(\\cos (\\hat{z}^k, \\hat{z}^{k'}) + \\frac{1}{K - 1})$ since the optimal pair-wise angle is $-\\frac{1}{K - 1}$ as discussed in Eq. (3). The trends of these two metrics on the CIFAR-10 dataset with and without representation calibration are shown in Figure 6. Notably, equipped with representation calibration, the standard deviations of the cosines and shifted average cosine values approach zero, which indicates that TRAILER achieves the engaging feature structure that is close to neural collapse.", + "bbox": [ + 75, + 482, + 468, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unknown Number of Novel Classes. The previous experiments follow previous protocols and assume that the class number is known apriori [5, 15, 46], which is sometimes impractical in real-world applications and brings challenges for both TRAILER and the other baselines. A simple resolution is to first use estimation methods [32, 39] to derive the number of classes before deployment, for example, $k$ -estimation with Brent's algorithm from [50] or cross-validating using labeled probe set from [5, 21, 46]. On", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a3d4c1dd03d8f9f2a88416addad0a78fd8eb2a950719e383e1aea43362dd5291.jpg", + "image_caption": [ + "(a) Accuracy for known classes.", + "Figure 7. Accuracy comparisons for known and novel classes on CIFAR-10 with different novel class ratios ranging from 0.1 to 0.9." + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 692, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/489b6e23a2ef86d0b8cea91eac19ad5b617b4c676c71bd08bb0860e23c9536dd.jpg", + "image_caption": [ + "(b) Accuracy for novel classes." + ], + "image_footnote": [], + "bbox": [ + 699, + 90, + 887, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CIFAR-100 the former gets exactly 100 while the latter derives 124 and was adopted by us for further exploration. We then utilize this estimation result to rerun the experiment by initializing the corresponding number of classifier vectors. As shown in Table 4, TRAILER is able to achieve competitive performance even when the number of classes is not known in advance. We further provide additional results with different estimated class numbers $K$ from 80 to 130 in Appendix B to demonstrate the robustness of TRAILER.", + "bbox": [ + 498, + 291, + 890, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Novel Class Ratio. Here we further evaluate the performance of TRAILER on the CIFAR-10 dataset with different ratios of novel classes from 0.1 to 0.9. As shown in Figure 7, we can see that TRAILER displays robust performance when the ratio of novel classes grows larger from 0.1 to 0.9 and consistently manifests superior performance than other baselines on both known and novel classes.", + "bbox": [ + 498, + 450, + 890, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 573, + 617, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel framework TRAILER for open-world SSL. We first take inspiration from the recently discovered neural collapse phenomenon and intend to attain its appealing feature arrangement with minimal within-class and maximum between-class covariance. To achieve this, we adopt a targeted classifier and align representations towards its pre-assigned optimal structure in a progressive manner. To mitigate the potential risk of misalignment, we tackle the sample-target assignment for unlabeled samples with a coarse-to-fine allocation mechanism. Comprehensive experiments show that TRAILER enhances the baselines by a notable margin on both generic and fine-grained benchmarks. We hope our work can inspire future research to leverage the appealing properties of the neural collapse phenomenon for tasks in open-world scenarios.", + "bbox": [ + 496, + 599, + 890, + 827 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 844, + 678, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the Pioneer R&D Program of Zhejiang (No. 2024C01035).", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23079", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR. OpenReview.net, 2020. 4", + "[2] Jessa Bekker and Jesse Davis. Learning from positive and unlabeled data: a survey. Mach. Learn., 109(4):719-760, 2020. 2, 5", + "[3] David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution alignment and augmentation anchoring. CoRR, abs/1911.09785, 2019. 2", + "[4] David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In NeurIPS, pages 5050-5060, 2019. 1, 2", + "[5] Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning. In ICLR. OpenReview.net, 2022. 1, 2, 3, 5, 6, 7, 8", + "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 4", + "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9630-9640. IEEE, 2021. 7", + "[8] Hui Chen, Fangqing Liu, Yin Wang, Liyue Zhao, and Hao Wu. A variational approach for learning from positive and unlabeled data. In NeurIPS 2020, 2020. 5", + "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. PMLR, 2020. 5, 6", + "[10] Xuxi Chen, Wuyang Chen, Tianlong Chen, Ye Yuan, Chen Gong, Kewei Chen, and Zhangyang Wang. Self-pu: Self boosted and calibrated positive-unlabeled training. In ICML, pages 1510-1519. PMLR, 2020. 5", + "[11] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NeurIPS, pages 2292-2300, 2013. 4", + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE Computer Society, 2009. 5", + "[13] François Denis. PAC learning from positive statistical queries. In Algorithmic Learning Theory, 9th International Conference, ALT '98, Otzenhausen, Germany, October 8-10, 1998, Proceedings, pages 112-126. Springer, 1998. 2, 5", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR. OpenReview.net, 2021. 7", + "[15] Enrico Fini, Enver Sangineto, Stephane Lathuilière, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "novel class discovery. In ICCV, pages 9264-9272. IEEE, 2021. 1, 2, 3, 4, 5, 6, 7, 8", + "[16] Ronald A Fisher. The use of multiple measurements in taxonomic problems. Annals of eugenics, 7(2):179-188, 1936. 4", + "[17] Tomer Galanti, András György, and Marcus Hutter. On the role of neural collapse in transfer learning. In ICLR. Open-Review.net, 2022. 2, 3", + "[18] Florian Graf, Christoph D. Hofer, Marc Niethammer, and Roland Kwitt. Dissecting supervised constrastive learning. In ICML, pages 3821-3830. PMLR, 2021. 2", + "[19] Lan-Zhe Guo, Zhenyu Zhang, Yuan Jiang, Yu-Feng Li, and Zhi-Hua Zhou. Safe deep semi-supervised learning for unseen-class unlabeled data. In ICML, pages 3897-3906. PMLR, 2020. 2, 5, 6", + "[20] Lan-Zhe Guo, Yi-Ge Zhang, Zhi-Fan Wu, Jie-Jing Shao, and Yu-Feng Li. Robust semi-supervised learning when not all classes have labels. In NeurIPS, 2022. 1, 2, 5, 6", + "[21] Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer clustering. In ICCV, pages 8400-8408. IEEE, 2019. 1, 2, 3, 5, 6, 8", + "[22] Kai Han, Sylvestre-Alvise Rebuffi, Sébastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Automatically discovering and learning new visual categories with ranking statistics. In ICLR. OpenReview.net, 2020. 1, 2, 5, 6, 7", + "[23] X. Y. Han, Vardan Papyan, and David L. Donoho. Neural collapse under MSE loss: Proximity to and dynamics on the central path. In ICLR. OpenReview.net, 2022. 2", + "[24] Zongyan Han, Zhenyong Fu, Shuo Chen, and Jian Yang. Contrastive embedding for generalized zero-shot learning. In CVPR, pages 2371-2381. Computer Vision Foundation / IEEE, 2021. 2", + "[25] Zongyan Han, Zhenyong Fu, Shuo Chen, and Jian Yang. Semantic contrastive embedding for generalized zero-shot learning. Int. J. Comput. Vis., 130(11):2606-2622, 2022. 2", + "[26] Wenlong Ji, Yiping Lu, Yiliang Zhang, Zhun Deng, and Weijie J. Su. An unconstrained layer-peeled perspective on neural collapse. In ICLR. OpenReview.net, 2022. 2", + "[27] Ryuichi Kiryo, Gang Niu, Marthinus Christoffel du Plessis, and Masashi Sugiyama. Positive-unlabeled learning with non-negative risk estimator. In NeurIPS, pages 1675-1685, 2017. 5", + "[28] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV, pages 554–561. IEEE Computer Society, 2013. 5", + "[29] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Master's thesis, Department of Computer Science, University of Toronto, 2009. 5", + "[30] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In ICLR. OpenReview.net, 2017. 2", + "[31] Dong-Hyun Lee. Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks. ICML 2013 Workshop : Challenges in Representation Learning (WREPL), 2013. 2", + "[32] Collin Leiber, Lena G. M. Bauer, Benjamin Schelling, Christian Böhm, and Claudia Plant. Dip-based deep embedded" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23080", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "clustering with k-estimation. In KDD, pages 903-913. ACM, 2021. 8", + "[33] Jiaming Liu, Yangqiming Wang, Tongze Zhang, Yulu Fan, Qinli Yang, and Junming Shao. Open-world semi-supervised novel class discovery. In *IJCAI*, pages 4002–4010. ijcai.org, 2023. 1, 2, 3, 5, 6", + "[34] Chuan Luo, Pu Zhao, Chen Chen, Bo Qiao, Chao Du, Hongyu Zhang, Wei Wu, Shaowei Cai, Bing He, Saravanakumar Rajmohan, and Qingwei Lin. PULNS: positive unlabeled learning with effective negative sample selector. In AAAI, pages 8784-8792. AAAI Press, 2021. 5", + "[35] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. CoRR, abs/1306.5151, 2013. 5", + "[36] Aleix M. Martínez and Avinash C. Kak. PCA versus LDA. IEEE Trans. Pattern Anal. Mach. Intell., 23(2):228-233, 2001. 2", + "[37] Duc Anh Nguyen, Ron Levie, Julian Lienen, Eyke Hüllermeier, and Gitta Kutyniok. Memorization-dilation: Modeling neural collapse under noise. In ICLR. OpenReview.net, 2023. 2, 4", + "[38] Vardan Papyan, X. Y. Han, and David L. Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 117(40):24652-24663, 2020. 2, 3", + "[39] Dan Pelleg and Andrew W. Moore. X-means: Extending k-means with efficient estimation of the number of clusters. In ICML, pages 727-734. Morgan Kaufmann, 2000. 8", + "[40] Zhongzheng Ren, Raymond A. Yeh, and Alexander G. Schwing. Not all unlabeled data are equal: Learning to weight data in semi-supervised learning. In NeurIPS, 2020. 2", + "[41] Mamshad Nayeem Rizve, Navid Kardan, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Openldn: Learning to discover novel classes for open-world semi-supervised learning. In ECCV, pages 382-401. Springer, 2022. 1, 5, 6", + "[42] Chuck Rosenberg, Martial Hebert, and Henry Schneiderman. Semi-supervised self-training of object detection models. In 7th IEEE Workshop on Applications of Computer Vision / IEEE Workshop on Motion and Video Computing (WACV/MOTION 2005), 5-7 January 2005, Breckenridge, CO, USA, pages 29-36. IEEE Computer Society, 2005. 2", + "[43] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In NeurIPS, pages 1163-1171, 2016. 2", + "[44] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 2, 5, 6", + "[45] Xin Sun, Zhenning Yang, Chi Zhang, Keck Voon Ling, and Guohao Peng. Conditional gaussian distribution learning for open set recognition. In CVPR, pages 13477-13486. Computer Vision Foundation / IEEE, 2020. 2, 5, 6", + "[46] Yiyou Sun and Yixuan Li. Opencon: Open-world contrastive learning. Trans. Mach. Learn. Res., 2023, 2023. 1, 2, 3, 5, 6, 7, 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Kiat Chuan Tan, Yulong Liu, Barbara Ambrose, Melissa Tulig, and Serge J. Belongie. The herbarium challenge 2019 dataset. CoRR, abs/1906.05372, 2019. 5", + "[48] Christos Thrampoulidis, Ganesh Ramachandra Kini, Vala Vakilian, and Tina Behnia. Imbalance trouble: Revisiting neural-collapse geometry. In NeurIPS, 2022. 2", + "[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7", + "[50] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. In CVPR, pages 7482-7491. IEEE, 2022. 1, 2, 3, 5, 6, 7, 8", + "[51] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In ICLR. OpenReview.net, 2022. 5, 6", + "[52] Jiaheng Wei, Zhaowei Zhu, Hao Cheng, Tongliang Liu, Gang Niu, and Yang Liu. Learning with noisy labels revisited: A study using real-world human annotations. In ICLR. OpenReview.net, 2022. 2", + "[53] P. Welinder, S. Branson, T. Mita, C. Wah, F. Schroff, S. Belongie, and P. Perona. Caltech-UCSD Birds 200. Technical Report CNS-TR-2010-001, California Institute of Technology, 2010. 5", + "[54] Liang Xie, Yibo Yang, Deng Cai, and Xiaofei He. Neural collapse inspired attraction-repulsion-balanced loss for imbalanced learning. Neurocomputing, 527:60-70, 2023. 2", + "[55] Qizhe Xie, Zihang Dai, Eduard H. Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. In NeurIPS, 2020. 1, 2", + "[56] Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In ICML, pages 11525-11536. PMLR, 2021. 2", + "[57] Yibo Yang, Shixiang Chen, Xiangtai Li, Liang Xie, Zhouchen Lin, and Dacheng Tao. Inducing neural collapse in imbalanced learning: Do we really need a learnable classifier at the end of deep neural network? In NeurIPS, 2022. 2, 4", + "[58] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip H. S. Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In ICLR. OpenReview.net, 2023. 2", + "[59] Bo Ye, Kai Gan, Tong Wei, and Min-Ling Zhang. Bridging the gap: Learning pace synchronization for open-world semi-supervised learning. CoRR, abs/2309.11930, 2023. 2", + "[60] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In NeurIPS 2021, pages 18408-18419, 2021. 1", + "[61] Bingchen Zhao, Xin Wen, and Kai Han. Learning semi-supervised gaussian mixture models for generalized category discovery. In ICCV, pages 16623-16633, 2023. 5, 6, 7", + "[62] Yunrui Zhao, Qianqian Xu, Yangbangyan Jiang, Peisong Wen, and Qingming Huang. Dist-pu: Positive-unlabeled learning from a label distribution perspective. In CVPR, pages 14441-14450. IEEE, 2022. 2, 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23081", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Zhisheng Zhong, Jiequan Cui, Yibo Yang, Xiaoyang Wu, Xiaojuan Qi, Xiangyu Zhang, and Jiaya Jia. Understanding imbalanced semantic segmentation through neural collapse. In CVPR, pages 19550-19559. IEEE, 2023. 2", + "[64] Jinxin Zhou, Xiao Li, Tianyu Ding, Chong You, Qing Qu, and Zhihui Zhu. On the optimization landscape of neural collapse under MSE loss: Global optimality with unconstrained features. In ICML, pages 27179-27202. PMLR, 2022. 2", + "[65] Zhihui Zhu, Tianyu Ding, Jinxin Zhou, Xiao Li, Chong You, Jeremias Sulam, and Qing Qu. A geometric analysis of neural collapse with unconstrained features. In NeurIPS, pages 29820-29834, 2021. 2", + "[66] Zhaowei Zhu, Tianyi Luo, and Yang Liu. The rich get richer: Disparate impact of semi-supervised learning. In ICLR. OpenReview.net, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 303 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23082", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/b33baf4c-8c46-4e13-a308-0177b1e2eee6_model.json b/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/b33baf4c-8c46-4e13-a308-0177b1e2eee6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8dd7b3fb1d882912527dd052c2026891da870e31 --- /dev/null +++ b/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/b33baf4c-8c46-4e13-a308-0177b1e2eee6_model.json @@ -0,0 +1,2389 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.002, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.085, + 0.131, + 0.887, + 0.155 + ], + "angle": 0, + "content": "Targeted Representation Alignment for Open-World Semi-Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.18, + 0.865, + 0.2 + ], + "angle": 0, + "content": "Ruixuan Xiao\\(^{1}\\), Lei Feng\\(^{2}\\), Kai Tang\\(^{1}\\), Junbo Zhao\\(^{1}\\), Yixuan Li\\(^{3}\\), Gang Chen\\(^{1}\\), Haobo Wang\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.2, + 0.604, + 0.217 + ], + "angle": 0, + "content": "1Zhejiang University, China" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.217, + 0.732, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\)Singapore University of Technology and Design, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.235, + 0.655, + 0.252 + ], + "angle": 0, + "content": "3University of Wisconsin-Madison, USA" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.255, + 0.892, + 0.27 + ], + "angle": 0, + "content": "{xiaoruixuan,tk0819,j.zhao,cg,wanghaobo}@zju.edu.cn,lfengqaq@gmail.com,sharonli@cs.wisc.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.337, + 0.474, + 0.73 + ], + "angle": 0, + "content": "Open-world Semi-Supervised Learning aims to classify unlabeled samples utilizing information from labeled data, while unlabeled samples are not only from the labeled known categories but also from novel categories previously unseen. Despite the promise, current approaches solely rely on hazardous similarity-based clustering algorithms and give unlabeled samples free rein to spontaneously group into distinct novel class clusters. Nevertheless, due to the absence of novel class supervision, these methods typically suffer from the representation collapse dilemma—features of different novel categories can get closely intertwined and indistinguishable, even collapsing into the same cluster and leading to degraded performance. To alleviate this, we propose a novel framework TRAILER which targets to attain an optimal feature arrangement revealed by the recently uncovered neural collapse phenomenon. To fulfill this, we adopt targeted prototypes that are pre-assigned uniformly with maximum separation and then progressively align the representations to them. To further tackle the potential downsides of such stringent alignment, we encapsulate a sample-target allocation mechanism with coarse-to-fine refinery that is able to infer label assignments with high quality. Extensive experiments demonstrate that TRAILER outperforms current state-of-the-art methods on generic and fine-grained benchmarks. The code is available at https://github.com/Justherozen/TRAILER." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.759, + 0.21, + 0.775 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.785, + 0.47, + 0.876 + ], + "angle": 0, + "content": "The resounding achievements of deep learning heavily rely on large-scale and accurately-annotated training data, which is labor-intensive and time-consuming. To alleviate this, semi-supervised learning (SSL) has emerged as a prominent alternative [4, 55, 60], which exploits unlabeled data for boosted performance. Despite the success, current SSL" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.305, + 0.871, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.461, + 0.893, + 0.531 + ], + "angle": 0, + "content": "Figure 1. Left: Scenario of open-world SSL where unlabeled samples are from known and novel classes. Right: Visualization without and with our targeted calibration mechanism. Without calibration, representations tend to get jumbled and inseparable, as in the left where dog and cat categories converge into the same cluster." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.893, + 0.776 + ], + "angle": 0, + "content": "methods are primarily based on the closed-world assumption that labeled and unlabeled data share the same predefined label space. Nevertheless, this assumption rarely holds in real-world applications, where the model may encounter novel classes previously unseen in unlabeled data. To relax this assumption, novel category discovery (NCD) was first proposed [15, 21, 22] which assumes all unlabeled data strictly belong to disjoint novel classes. Recently, a more practical setting, open-world semi-supervised learning (open-world SSL), also referred to as generalized category discovery (GCD), has received huge attention [5, 46, 50] by generalizing standard SSL and NCD. It assumes that unlabeled data contain both known classes from the labeled data and similar yet distinct novel classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The goal of open-world SSL is to classify the already-known classes from the labeled data, as well as to discover new classes in unlabeled data without annotations and assign instances to them. Nascent efforts have been made to address this practical yet under-investigated task [20, 33, 41, 50]. The essence of these recent works is to rashly rely on the similarity-based [5, 41] or contrastive-based [46] clustering algorithm in feature space to spontaneously" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23072" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "group different samples into separate clusters. Some methods further integrate self-training techniques with generated pseudo-labels for boosted performance [20, 33]. However, due to the missing supervision and semantic shift for novel categories, the self-organized clustering step can get overfitted and biased towards the known classes, leading to a representation collapse dilemma—the feature distance between one novel category and another category may get too close and finally become inseparable, even converging within a single cluster and thus jeopardizing generalization performance, as illustrated in the left of Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.469 + ], + "angle": 0, + "content": "On the other hand, a recent study has unveiled a feature-classifier alignment phenomenon termed neural collapse (NC) [38]. It indicates that under ideal conditions, the last-layer features of the same class tend to collapse into their within-class mean. Meanwhile, these within-class means of all classes will be aligned with classifiers and formed as a simplex equiangular tight frame (ETF) at the end of training, as elaborated in Section 3.1. Intuitively, such an elegant NC phenomenon affords an optimal feature structure with minimized within-class variance and maximized between-class variance [36], coinciding with what we fervently desire. Nevertheless, in open-world SSL, the absence of novel class annotation can hinder the representation learning and thus disrupt the natural induction of NC." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.473, + 0.834 + ], + "angle": 0, + "content": "In an effort to salvage this and accomplish the appealing feature arrangement of NC, we propose a novel framework, Targeted Representation AIIgnment for open-world semisupervised LEaRning (dubbed TRAILER). Key to our method, TRAILER leverages a targeted classifier to progressively align embeddings to the pre-assigned ETF structure, eventually fulfilling the optimal state of uniformly and maximally separated embeddings. Despite the promise, this calibration procedure is still susceptible to false alignment stemming from unreliable pseudo-labels. To mitigate this, we integrate a hierarchical sample-target allocation strategy that comprises two sequential steps: (i)-rough sample assignment based on the optimal transport mechanism; (ii)-label refinery by casting the known-novel separation problem to a specific weakly-supervised paradigm termed PU learning [2, 13, 62]. Such an allocation mechanism can provide more precise pseudo-labels and thus facilitate representation calibration. Through our visualized results in Figure 1 and Section 4.3, TRAILER does indeed induce discriminative and non-skewed embeddings. Empirically, TRAILER surpasses state-of-the-art approaches on different evaluation benchmarks, e.g., improves the best baseline by \\(3.1\\%\\), \\(2.9\\%\\), and \\(1.5\\%\\) overall accuracy on CIFAR-10, CIFAR-100 and ImageNet-100 datasets, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.22, + 0.86 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Closed-World Semi-Supervised Learning has been widely explored [4, 44, 55, 66] to learn from limited la" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.289 + ], + "angle": 0, + "content": "beled and massive unlabeled data. Pseudo labeling [31, 42] and consistency regularization [30, 44, 55] are two prominent techniques. The former is based on self-training that utilizes model predictions as optimization targets [40, 56], while the latter enforces model to produce consistent outputs for different views of the same image [3, 4, 43, 55]. However, these methods are primarily based on the closed-world assumption that labeled and unlabeled data share the same pre-defined label space, which limits their real-world application where unlabeled samples from previously unseen classes may emerge. Open-set SSL [19, 24, 25, 45] relaxes this by simply rejecting all unseen class instances, but still fails to classify them into distinct novel categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.895, + 0.656 + ], + "angle": 0, + "content": "Open-World Semi-Supervised Learning further accommodates real-world scenarios [5, 20, 46]. It is also known as generalized category discovery (GCD) [50] and is a natural extension of novel category discovery (NCD). NCD is first introduced in [21] and improved in later works [15, 22] to discover new categories by leveraging the knowledge from labeled known categories. However, NCD assumes that all unlabeled samples belong to novel classes and thus fails to recognize known class samples. Most recently, open-world SSL generalizes NCD to identify unlabeled samples from both known and novel classes [5]. To tackle this, several incipient solutions have been proposed [20, 59], which typically involve similarity-based representation learning and then involve sample clustering to generate pseudo-labels for self-training. ORCA [5] is the pioneering work that integrates an uncertainty-based margin loss to control the intraclass variance of different classes. OpenCon [46] further derives a semi-supervised constrastive mechanism to produce compact feature space. Despite the promise, these methods solely rely on spontaneous clustering based on sample similarity, overlooking potential risks that representations of different classes possibly converge into the same cluster, thus leading to the representation collapse dilemma." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Neural Collapse describes an elegant alignment phenomenon [38], which states that the last-layer features will collapse into their within-class centers and these centers together with classifiers will eventually form a simplex equiangular tight frame (ETF). Stemming from the appealing symmetry, a plethora of studies have been triggered to explain this phenomenon theoretically [26, 48, 58, 65]. NC is proved to be the global optimality under ideal conditions with the CE [18, 26] and the MSE [23, 64] loss functions. The NC phenomenon has also been investigated in some specific scenarios like imbalanced learning [54, 57, 63] noisy label learning [37, 52], transfer learning [17] and incremental learning [58]. In our work, we investigate the opportunities of inducing such optimal feature arrangement of NC through targeted calibration for open-world SSL." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23073" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.09, + 0.331, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.089, + 0.854, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.316, + 0.895, + 0.361 + ], + "angle": 0, + "content": "Figure 2. Overview of TRAILER. For representation calibration, we leverage the targeted classifier \\(\\phi^{etf}\\) to align the feature towards the optimal arrangement of ETF progressively while maintaining the within-class compactness with prototypical clustering. For sample-target allocation, the unlabeled samples are provided with reliable pseudo-labels with rough assignment and subsequent label refinery." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.383, + 0.17, + 0.399 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.474, + 0.637 + ], + "angle": 0, + "content": "Problem Definition. In open-world SSL, we assume that the training dataset \\(\\mathcal{D} = \\mathcal{D}_l \\cup \\mathcal{D}_u\\) comprises two subsets: a labeled set \\(\\mathcal{D}_l = \\{x_i, y_i\\}_{i=1}^m\\) where the input image \\(x_i \\in \\mathbb{R}^d\\) and the ground-truth label \\(y_i\\) come from a set of known classes \\(\\mathcal{C}_{known}\\), and an unlabeled set \\(\\mathcal{D}_u = \\{x_i\\}_{i=1}^n\\) where the potential ground-truth label is from a set of classes \\(\\mathcal{C}_{all}\\). In closed-world SSL, the unlabeled samples strictly come from known classes, i.e., \\(\\mathcal{C}_{all} = \\mathcal{C}_{known}\\). As for open-world SSL which is more practical, the unlabeled samples can come from either known classes \\(\\mathcal{C}_{known}\\) or novel classes, i.e., \\(\\mathcal{C}_{known} \\subset \\mathcal{C}_{all}\\). We consider \\(\\mathcal{C}_{novel} = \\mathcal{C}_{all} \\backslash \\mathcal{C}_{known}\\) as the set of novel classes. Following previous setting [5, 15, 46], we assume the number of novel classes \\(|\\mathcal{C}_{novel}|\\) is known apriori, which can also be estimated using the off-the-shelf methods [21, 50] as discussed in Section 4.3." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.655, + 0.473, + 0.854 + ], + "angle": 0, + "content": "Learning Goal. For classification problem with \\( K = |\\mathcal{C}_{all}| \\) classes, open-world SSL aims to obtain a model \\( f_{\\theta,\\phi} : \\mathbb{R}^d \\mapsto \\mathbb{R}^K \\) that recognizes samples from both known and novel classes based on \\( \\mathcal{D}_l \\) and \\( \\mathcal{D}_u \\). Specifically, \\( f_{\\theta,\\phi} = h_\\phi \\circ g_\\theta \\) can be decomposed into a representation backbone \\( g_\\theta \\) and a classification head \\( h_\\phi \\) with weights \\( \\phi = [\\phi_1, \\dots, \\phi_k] \\in \\mathbb{R}^{P \\times K} \\), for which we adopt a prototypical one following [15] with logit \\( h(z) = \\mathrm{Norm}(z)^\\top \\mathrm{Norm}(\\phi) \\). Formally, given the weakly and strongly augmented view \\( x_i \\) and \\( x_i' \\) of the same image in batch \\( \\mathcal{B} \\), we define \\( z_i, z_i' \\in \\mathbb{R}^P \\) as the last-layer \\( P \\)-dimensional features and \\( p_i = \\mathrm{softmax}(h(z_i)) \\) as output probability. Normalization for features is denoted as \\( \\mathrm{Norm}(z) = z / ||z||_2 \\) and omitted in the subsequent." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Warm-up Phase. Similar to previous works [5, 33], we also include a commonly adopted warm-up strategy at" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.895, + 0.549 + ], + "angle": 0, + "content": "the start of training. The overall objectives to cluster known classes and discover potential novel samples consist of three losses: 1) the cross-entropy loss (CE) \\(\\mathcal{L}_{ce}\\) on labeled samples \\(\\mathcal{D}_l\\) which is elaborated later in Section 3.1; 2) the self-supervised contrastive loss \\(\\mathcal{L}_{con} = -\\log \\frac{\\exp(z_i^\\top z_i' / \\tau)}{\\sum_{j=1}^n \\exp(z_i^\\top z_j' / \\tau)}\\) on all samples; and 3) the entropy regularization \\(\\mathcal{L}_{ent} = \\sum_{k=1}^{K} \\overline{p}^k \\log (\\overline{p}^k)\\) for preventing the trivial solution of assigning all instances to the same class, where \\(\\overline{p} = \\frac{1}{|\\mathcal{B}|} \\sum_{x_i \\in \\mathcal{B}} p_i\\) is the average probability of the batch. The overall losses for warm-up are formulated," + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.558, + 0.892, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {w a r m}} = \\mathcal {L} _ {c e} + \\mathcal {L} _ {\\text {c o n}} + \\alpha \\mathcal {L} _ {\\text {e n t}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.895, + 0.69 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a weighting parameter. After warm-up, we also would like to group unlabeled samples into distinct clusters and assign them pseudo-labels for self-training. As mentioned earlier, the crucial challenge is how to abstain from the representation collapse dilemma and obtain discriminative features without novel class annotations. To fulfill this, we first draw inspiration from neural collapse phenomenon." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.696, + 0.892, + 0.712 + ], + "angle": 0, + "content": "3.1. Neural Collapse for Representation Calibration" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.894, + 0.795 + ], + "angle": 0, + "content": "Recent works have disclosed the Neural Collapse phenomenon [17, 38] which corresponds to an optimal feature-classifier alignment towards a simplex equiangular tight frame (ETF). This offers a uniformly and equally separated feature arrangement that is desirable for open-world SSL." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.802, + 0.895, + 0.864 + ], + "angle": 0, + "content": "Definition 1 (Simplex ETF) A Simplex equiangular tight frame (ETF) refers to a collection of \\(K\\) equal-length and maximally-equiangular \\(P\\)-dimensional embedding vectors \\(\\mathbf{E} = [e_1, \\dots, e_K] \\in \\mathbb{R}^{P \\times K}\\) which satisfies:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.87, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} = \\sqrt {\\frac {K}{K - 1}} \\mathbf {U} \\left(\\mathbf {I} _ {K} - \\frac {1}{K} \\mathbf {1} _ {K} \\mathbf {1} _ {K} ^ {\\top}\\right) \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23074" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.473, + 0.123 + ], + "angle": 0, + "content": "where \\(\\mathbf{I}_K\\) is the identity matrix, \\(\\mathbf{1}_K\\) is an all-ones vector, and \\(\\mathbf{U} \\in \\mathbb{R}^{P \\times K}\\) (\\(P \\geq K\\)) allows a rotation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.13, + 0.47, + 0.163 + ], + "angle": 0, + "content": "All vectors in a simplex ETF \\(\\mathbf{E}\\) have an equal \\(\\ell_2\\) norm and the same pair-wise maximal equiangular angle \\(-\\frac{1}{K - 1}\\) [57]," + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.178, + 0.47, + 0.208 + ], + "angle": 0, + "content": "\\[\ne _ {k _ {1}} ^ {\\top} e _ {k _ {2}} = \\frac {K}{K - 1} \\delta_ {k _ {1}, k _ {2}} - \\frac {1}{K - 1}, \\forall k _ {1}, k _ {2} \\in [ 1, K ] \\quad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.213, + 0.47, + 0.349 + ], + "angle": 0, + "content": "where \\(\\delta_{k_1,k_2} = 1\\) when \\(k_{1} = k_{2}\\) and 0 otherwise. The NC phenomenon can be described by the following properties under ideal conditions: (1) The last-layer features converge to their within-class means, and the within-class means of all classes converge to different vertices of a simplex ETF. (2) The normalized classifier weights converge to the same simplex ETF. Notably, this process is implicit and natural since it is induced with common optimizers such as SGD, rather than enforced by an explicit objective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.469, + 0.485 + ], + "angle": 0, + "content": "As discussed above, NC has provided an optimal ETF structure for feature arrangement with minimal within-class covariance and maximum between-class covariance (i.e., maximized Fisher discriminant ratio [16]), which coincides with our goal of deriving discriminative and non-skewed embeddings. However, NC is delicate and can fail to be naturally induced under imperfect conditions [37, 57]. For open-world SSL, the missing novel class annotations can also lead to biased features and break the NC optimality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.469, + 0.622 + ], + "angle": 0, + "content": "To remedy this, we resort to effectuating an explicit NC and fulfilling the appealing feature arrangement with representation calibration, which is achieved by leveraging a targeted prototypical classifier \\( h_{\\text{etf}} \\). In specific, \\( h_{\\text{etf}} \\) is preassigned with weights \\( \\phi^{\\text{etf}} \\in \\mathbb{R}^{P \\times K} \\), which is a random simplex ETF with optimal structure as Definition 1 and is fixed rather than learnable during training. By assuming a known label \\( \\hat{y}_i \\) for each sample \\( x_i \\), the standard CE loss \\( \\mathcal{L}_{ce} \\) is then replaced as \\( \\mathcal{L}_{\\text{etf}} \\), which is formulated," + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.628, + 0.47, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e t f} \\left(x _ {i}, \\hat {y} _ {i}\\right) = - \\log \\frac {\\exp \\left(z _ {i} ^ {\\top} \\cdot \\phi_ {\\hat {y} _ {i}} ^ {e t f} / \\tau\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(z _ {i} ^ {\\top} \\cdot \\phi_ {k} ^ {e t f} / \\tau\\right)} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.686, + 0.468, + 0.732 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature parameter. Through \\(\\mathcal{L}_{\\text{etf}}\\), the feature \\(z_{i}\\) is drawn closer to its corresponding target \\(\\phi_{\\hat{y}_i}^{\\text{etf}}\\) and pulled away from other class vectors of ETF structure." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.733, + 0.469, + 0.793 + ], + "angle": 0, + "content": "For the labeled sample \\((x_{i},y_{i})\\), we directly set \\(\\hat{y}_i = y_i\\). As for unlabeled data without ground-truth labels, we design a sample-target allocation strategy to generate pseudolabels \\(\\hat{y}\\), whose details are shown in Section 3.2." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Progressive Alignment. Notably, if we directly optimize the loss \\(\\mathcal{L}_{\\text{etf}}\\) after the warm-up phase, the provisionally clustered samples would be forcibly dragged to the corresponding targeted vector \\(\\phi_{\\hat{y}}^{\\text{etf}}\\). As a result, the already-learned clustering compactness can get compromised and the within-cluster feature distribution can get jumbled." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "To address this, we intend to align the representations in a progressive manner to preserve the compactness within each cluster during calibration. Formally, given sample \\( x_{i} \\) with pseudo-label \\( \\hat{y}_i \\), we also maintain the authentic prototypes \\( \\mu_k = \\mathrm{Norm}(\\mathrm{Avg}_{\\hat{y}_i = k}(z_i)) \\) for each class \\( 1 \\leq k \\leq K \\). To preserve the compactness of cluster structure, a prototypical clustering loss is also integrated," + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.208, + 0.892, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p r o t o}} \\left(z _ {i}, \\hat {y} _ {i}\\right) = - \\log \\frac {\\exp \\left(z _ {i} ^ {\\top} \\cdot \\mu_ {\\hat {y} _ {i}} / \\tau\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(z _ {i} ^ {\\top} \\cdot \\mu_ {k} / \\tau\\right)} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.256, + 0.825, + 0.271 + ], + "angle": 0, + "content": "The overall classification loss is then formulated," + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.286, + 0.892, + 0.303 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} = \\lambda \\mathcal {L} _ {p r o t o} + (1 - \\lambda) \\mathcal {L} _ {e t f} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.316, + 0.892, + 0.392 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a calibration strength parameter that is gradually ramped down from 1 to 0. With such a progressive manner, the embeddings are cautiously and smoothly drawn closer to the targeted optimal arrangement of ETF while the within cluster compactness is preserved." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.401, + 0.834, + 0.417 + ], + "angle": 0, + "content": "3.2. Hierarchical Sample-Target Allocation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.561 + ], + "angle": 0, + "content": "In previous steps, we involve targeted classifiers to regulate representation learning. Nevertheless, our stringent calibration still holds unexpected risks. Lacking supervision, the calibration process for unlabeled samples largely relies on the quality of generated pseudo-labels. When the pseudolabels are unreliable, the feature alignment procedure can be misguided towards the wrong cluster. To overcome this, we introduce our hierarchical sample-target allocation mechanism with course-to-fine assignment in what follows." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.699 + ], + "angle": 0, + "content": "Rough Assignment through Optimal Transport. For pseudo-label generation, one may adopt model prediction \\(\\arg \\max_{1\\leq k\\leq K}h^{k}(z_{i})\\) directly. However, this may lead to degenerate solutions where data points of different categories map into a single category [1], thus resulting in representation collapse. To mitigate this, we adopt a widely-used equipartition constraint [6, 15] for rough label assignment, which is attained by solving an optimal transport problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.699, + 0.892, + 0.772 + ], + "angle": 0, + "content": "Formally, given a batch \\(\\mathcal{B}\\) of \\(b\\) unlabeled samples with the logit matrix \\(\\mathbf{P} = [h(z_1),\\dots,h(z_b)]\\) whose columns are output logits by \\(h_{\\mathrm{etf}}\\) and the label-assignment matrix \\(\\mathbf{Q} = [q_1,\\dots,q_b]\\in [0,1]^{K\\times b}\\) whose columns are the to-beassigned soft labels \\(q_{i}\\). \\(\\mathbf{Q}\\) is estimated by," + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.78, + 0.892, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {Q} = \\max _ {\\mathbf {Q} \\in \\Gamma} \\operatorname {T r} (\\mathbf {Q} ^ {\\top} \\mathbf {P}) + \\epsilon \\mathrm {H} (\\mathbf {Q}) \\\\ \\text {s . t .} \\Gamma = \\left\\{\\mathbf {Q} \\in \\mathbb {R} _ {+} ^ {K \\times b} \\mid \\mathbf {Q} \\mathbf {1} _ {b} = \\frac {1}{K} \\mathbf {1} _ {k}, \\mathbf {Q} ^ {\\top} \\mathbf {1} _ {K} = \\frac {1}{b} \\mathbf {1} _ {b} \\right\\} \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathrm{H}(\\cdot)\\) is the entropy function, \\(\\epsilon\\) is a hyperparameter and \\(\\operatorname{Tr}(\\cdot)\\) is the trace function. The solution of Eq. (7) is obtained using the Sinkhorn-Knopp algorithm [11]. Please refer to Appendix A for more optimization details." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23075" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.227 + ], + "angle": 0, + "content": "Label Refinery with Known-Novel Separation. After the rough assignment, we obtain unpolished soft labels \\(q \\in [0,1]^K\\). However, the aforementioned skewed predictions towards known classes stemming from overfitting remain unresolved. To rectify this, we prioritize the subtask of precisely demarcating known and novel classes for label refinery, which is achieved by including an auxiliary binary head \\(h_{aux}\\) and reformulating this binary task as the classical positive unlabeled (PU) learning problem [2, 13, 62]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.468, + 0.378 + ], + "angle": 0, + "content": "PU learning tackles a special case of binary classification where the model has access to a labeled set containing merely positive samples and an unlabeled set with both positive and negative samples. Thus, the task of distinguishing known and novel classes can be intuitively converted to PU learning by regarding known and novel as positive and negative respectively. A flurry of methods has been designed for PU learning [10, 27, 34], among which we adopt a prior-free variational algorithm [8] with loss \\(\\mathcal{L}_{aux}\\) to optimize the newly devised \\(h_{aux}\\). See Appendix A for more details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.379, + 0.468, + 0.424 + ], + "angle": 0, + "content": "After that, we elicit the binary prediction \\(\\hat{y}^{aux}\\) for \\(h_{aux}\\), where 1 and 0 indicate prediction of known and novel respectively. Then we leverage \\(\\hat{y}^{aux}\\) to refine the label \\(q\\)," + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.433, + 0.468, + 0.473 + ], + "angle": 0, + "content": "\\[\nq ^ {j} = \\left\\{ \\begin{array}{l l} q ^ {j} & \\text {i f} \\mathbb {I} (j \\in \\mathcal {C} _ {\\text {k n o w n}}) = \\hat {y} ^ {\\text {a u x}} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.468, + 0.573 + ], + "angle": 0, + "content": "where \\(q^{j}\\) is the \\(j\\)-th entry of \\(q\\), \\(1 \\leq j \\leq K\\) and \\(\\mathbb{I}(\\cdot)\\) is the indicator function. That is, we retain the known/novel class components of soft label \\(q\\) based on prediction \\(\\hat{y}^{aux}\\) and discard the remaining inconsistent components. The refined \\(q\\) are then utilized for the final hard pseudo-label generation \\(\\hat{y}_{i} = \\arg \\max_{1 \\leq k \\leq K} q^{k}\\) in \\(\\mathcal{L}_{cls}\\) of Eq. (6)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.468, + 0.742 + ], + "angle": 0, + "content": "Pseudo-Label Filtering. While the proposed sample-target allocation mechanism can enhance the quality of pseudo-labels, the remaining obstinate noisy pseudo-labels can still impede model training, especially in the early stages when the model output is not accurate enough. To neutralize this, we further utilize a high-confidence selection to select clean pseudo-labels from noisy ones for \\(\\mathcal{D}_u\\). Formally, for each class \\(1 \\leq j \\leq K\\), we select a clean subset \\(\\mathcal{D}_{sel}^j\\) from \\(\\mathcal{D}_u^j = \\{(x, \\hat{y}) \\in \\mathcal{D}_u | \\hat{y} = j\\}\\) class by class with the highest confidence scores in the first \\(R\\) percent," + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.753, + 0.468, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {s e l} ^ {j} = \\left\\{\\left(x, \\hat {y}\\right) \\in \\mathcal {D} _ {u} ^ {j} \\mid r a n k \\left(q ^ {j}\\right) < R \\% \\right\\} \\tag{9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.468, + 0.871 + ], + "angle": 0, + "content": "Then the integral selected set is merged as \\(\\mathcal{D}_{sel} = \\cup_{j=1}^{K} \\mathcal{D}_{sel}^{j}\\). The classification loss in Eq. (6) is then merely valid on \\(\\mathcal{D}_{sel}\\). In practice, we set a small \\(R\\) to ensure the precision of selected samples at the beginning and gradually increase it to integrate more clean samples. Finally, the overall training loss is given by," + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.884, + 0.468, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {a u x}} + \\mathcal {L} _ {\\text {c o n}} + \\alpha \\mathcal {L} _ {\\text {e n t}} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.891, + 0.167 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{con}\\) and \\(\\mathcal{L}_{ent}\\) are described in Eq. (1). Through this procedure, Our sample-target allocation mechanism is able to provide pseudo-labels with high quality and guide the representation calibration toward the correct targets. The overview of TRAILER is illustrated in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.182, + 0.623, + 0.199 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.208, + 0.891, + 0.268 + ], + "angle": 0, + "content": "In this section, we present the main results and part of the ablation results to verify the effectiveness of TRAILER. More experimental details and results, such as the results with less labeled data, can be found in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.279, + 0.581, + 0.294 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.302, + 0.892, + 0.498 + ], + "angle": 0, + "content": "Datasets. We first evaluate TRAILER on three datasets CIFAR-10, CIFAR-100 [29] and ImageNet-100, which is randomly sub-sampled with 100 classes from ImageNet [12] following [5]. For each dataset, We first divide classes into the first \\(50\\%\\) known and the rest \\(50\\%\\) novel classes following [5]. Then we randomly select a portion \\((50\\%)\\) of samples from the known classes as labeled \\(\\mathcal{D}_l\\), and the remaining along with all novel class samples are regarded as unlabeled \\(\\mathcal{D}_u\\). We also conduct experiments with different novel class ratios and on the fine-grained semantic shift benchmarks (SSB) [51], including CUB-200 [53], Stanford Cars [28], FGVC-Aircraft [35], and a naturally long-tailed Herbarium 19 [47] datasets, as discussed in Section 4.3." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.518, + 0.892, + 0.821 + ], + "angle": 0, + "content": "Baselines. We compare TRAILER with various open-world SSL baselines, including ORCA [5], GCD [50], OpenCon [46], NACH [20], OpenLDN [41], GPC [61] and OpenNCD [33]. We also provide comparisons with SSL, open-set SSL, and NCD baselines following [20, 33]. For traditional and open-set SSL baselines including FixMatch [44], \\(\\mathrm{DS}^3\\mathrm{L}\\) [19], and CGDL [45], we extend them to be applicable to novel classes by selecting samples with lower confidence scores as novel and applying \\(k\\)-means clustering to obtain results. For NCD methods including DTC [21], RankStats [22], and \\(\\mathrm{UNO+}\\) [15], they are extended to classify known classes by using Hungarian algorithm to match some discovered classes with known classes. For performances of baselines, we directly adopt reported results from receptive papers or from [33], except OpenLDN was retested under transductive setting. All compared methods are implemented based on pre-trained ResNet using SimCLR [9] following [5], except DTC and OpenCon which have specialized pre-training procedure. We run all experiments three times and report the averaged results." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. Following [5, 33], we evaluate the transductive performance on both known and novel classes. For novel classes, we utilize the Hungarian algorithm to derive the optimal assignments and then calculate accuracy" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23076" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.113, + 0.089, + 0.86, + 0.32 + ], + "angle": 0, + "content": "
MethodsCIFAR-10CIFAR-100ImageNet-100
KnownNovelAllKnownNovelAllKnownNovelAll
FixMatch [44]71.550.449.539.623.520.365.836.734.9
DS3L [19]77.645.340.255.123.724.071.232.530.8
CGDL [45]72.344.639.749.322.523.567.333.831.9
DTC [21]53.939.538.331.322.918.325.620.821.3
RankStats [22]86.681.082.936.428.423.147.328.740.3
SimCLR [9]58.363.451.728.621.122.339.535.736.9
ORCA [5]88.290.489.766.943.048.189.172.177.8
GCD [50]78.479.779.168.533.545.282.358.368.2
OpenLDN [41]92.386.888.667.340.449.474.545.755.3
OpenCon [46]90.491.189.369.147.852.790.680.883.8
NACH [20]89.592.291.368.747.052.191.075.579.6
OpenNCD [33]88.490.690.169.743.449.390.077.581.6
TRAILER (Ours)93.495.094.469.748.755.691.482.485.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.33, + 0.893, + 0.36 + ], + "angle": 0, + "content": "Table 1. Accuracy comparison of known, novel, and all classes on CIFAR-10, CIFAR-100 and ImaneNet-100 dataset. The dataset is composed of \\(50\\%\\) known classes and \\(50\\%\\) novel classes, with \\(50\\%\\) of the known classes labeled. Bold entries indicate superior results." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.381, + 0.462, + 0.512 + ], + "angle": 0, + "content": "
AblationCIFAR-10CIFAR-100
KnownNovelAllKnownNovelAll
TRAILER93.495.094.469.748.755.6
w/o Rep Calibration94.158.970.669.214.933.4
with λ = 093.991.892.570.246.654.6
w/o Rough Assign92.992.392.569.337.948.3
w/o Label Refinery95.289.691.570.345.453.7
w/o PL Filtering92.290.591.069.943.652.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.522, + 0.468, + 0.551 + ], + "angle": 0, + "content": "Table 2. Ablation results on CIFAR-10 and CIFAR-100. 'Rep' and 'PL' indicate representation and pseudo-label respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.576, + 0.469, + 0.623 + ], + "angle": 0, + "content": "\\(\\frac{1}{n}\\sum_{i = 1}^{n}\\mathbb{I}(y_i = \\text{perm}(\\hat{y}_i))\\) where \\(\\text{perm}(\\cdot)\\) is the matching permutation. The overall joint accuracy is also measured by Hungarian match using both known and novel classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.837 + ], + "angle": 0, + "content": "Implementation Details. For CIFAR experiments, we use ResNet-18 as the backbone. The model is trained for 200 epochs with warm-up of 15 epochs. We employ the Adam optimizer with a batch size of 200 and an initial learning rate of \\(5e^{-4}\\), which decays by a cosine scheduler. The temperature \\(\\tau\\) is set as 0.3 and 0.5 for CIFAR-10 and CIFAR-100. We linearly ramp down \\(\\lambda\\) from 1 to 0 and ramp up selection ratio \\(R\\%\\) from 0.3 to 0.9 for pseudo-label filtering. The label refinery is enabled after training \\(h_{aux}\\) for 15 epochs. For the optimization of Sinkhorn-Knopp, we adopt hyperparameters from [15] and set \\(\\epsilon = 0.05\\) with 3 iterations. We put more implementation details for ImageNet-100 and fine-grained dataset in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.208, + 0.862 + ], + "angle": 0, + "content": "4.2. Main Reults" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Table 1 shows the comparison results on the CIFAR-10, CIFAR-100, and ImageNet-100 datasets, including" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.384, + 0.693, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.504, + 0.66, + 0.516 + ], + "angle": 0, + "content": "(a) Without calibration." + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.386, + 0.888, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.504, + 0.851, + 0.516 + ], + "angle": 0, + "content": "(b) With calibration." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.528, + 0.892, + 0.571 + ], + "angle": 0, + "content": "Figure 3. T-SNE feature visualization on CIFAR-10. Different colors represent the corresponding ground-truth classes. Without calibration, the dog and cat classes collapse into the same cluster." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.809 + ], + "angle": 0, + "content": "the accuracy of the 'Known', 'Novel', and 'All' classes. TRAILER significantly outperforms all the rivals by a considerable margin in terms of overall accuracy on different datasets. Specifically, on CIFAR-10, TRAILER improves upon the best baseline by \\(1.1\\%\\), \\(2.8\\%\\), and \\(3.1\\%\\) for the 'Known', 'Novel', and 'All' classes respectively. Moreover, on the more challenging CIFAR-100 and ImageNet-100 datasets with larger label space, while most baselines demonstrate a notable performance drop, TRAILER shows robustness and reaches an extraordinary trade-off between the known and novel classes. It consistently maintains superior performance with \\(2.9\\%\\) and \\(1.5\\%\\) overall enhancement. These results empirically demonstrate the efficacy of our proposed TRAILER for the task of open-world SSL." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.602, + 0.833 + ], + "angle": 0, + "content": "4.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Results on Semantic Shift Benchmarks. We further conduct experiments on fine-grained semantic shift benchmarks (SSB) [51] following previous work [50, 61]. For fair comparisons, we adopt the same training and compar" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23077" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.862, + 0.251 + ], + "angle": 0, + "content": "
MethodsCUBStanford CarsFGVC-AircraftHerbarium 19
KnownNovelAllKnownNovelAllKnownNovelAllKnownNovelAll
k-means38.932.134.310.613.812.814.416.816.012.912.812.9
RankStats+ [22]51.624.233.361.812.128.336.422.226.955.812.827.9
UNO+ [15]49.028.135.170.518.635.556.432.240.353.714.728.3
ORCA [5]45.630.235.350.110.723.531.817.122.030.915.520.9
GCD [50]56.648.751.357.629.939.041.146.945.051.027.035.4
OpenCon [46]63.852.154.778.632.749.1---58.928.639.3
GPC [61]58.253.155.459.232.842.842.547.946.351.727.936.5
TRAILER (Ours)71.361.965.171.747.655.462.650.554.557.037.844.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.261, + 0.892, + 0.275 + ], + "angle": 0, + "content": "Table 3. Results on the Semantic Shift Benchmarks. Bold entries indicate superior results and blank ones indicate results are not provided." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.3, + 0.27, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.421, + 0.252, + 0.433 + ], + "angle": 0, + "content": "(a) Binary accuracy of \\( h_{aux} \\)." + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.3, + 0.468, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.421, + 0.443, + 0.433 + ], + "angle": 0, + "content": "(b) Pseudo-label precision." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.501 + ], + "angle": 0, + "content": "Figure 4. (a) Accuracy of binary prediction between known and novel classes in first 100 epochs. (b) Per-class pseudo-label precision for unlabeled samples on CIFAR-10 with and without label refinery. Class indexes 0-4 are known while 5-9 are novel." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.47, + 0.636 + ], + "angle": 0, + "content": "isons protocol from [50] and utilize ViT-B/16 backbone [14] with DINO pre-trained weights [7]. As shown in Table 3, it can be observed that TRAILER exceeds the baselines by a substantial lead in all metrics, with improvements of \\(8.8\\%\\), \\(14.8\\%\\), \\(2.6\\%\\), and \\(9.2\\%\\) for 'Novel' accuracy on these datasets. These results consistently demonstrate its superior performance, even under challenging fine-grained tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.468, + 0.869 + ], + "angle": 0, + "content": "Benefits of Representation Calibration. To explore the benefits of representation calibration, we conduct ablations and equip TRAILER with different representation learning strategies: 1) TRAILER w/o Rep Calibration which dismiss the feature alignment of the targeted classifier and instead utilize the vanilla learnable classifier; 2) TRAILER with \\(\\lambda = 0\\) which discards the progressive manner for direct calibration. From Table 2, although TRAILER w/o Rep Calibration remains competitive on known classes, it suffers from a calamitous performance plunge for novel classes, which indicates the efficacy of representation calibration. TRAILER with \\(\\lambda = 0\\) achieves decent performance, but still underperforms TRAILER perceptibly, indicating the advantage of the progressive manner." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We further visualize the representation with and without representation calibration on CIFAR-10 using t-SNE [49]" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.301, + 0.871, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.421, + 0.892, + 0.448 + ], + "angle": 0, + "content": "Figure 5. Confusion matrix between known and novel classes on CIFAR-10. TRAILER can alleviate biased predictions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.892, + 0.581 + ], + "angle": 0, + "content": "in Figure 3 and in Appendix B. It can be shown that representation calibration facilitates inducing compact and nonskewed representations, discarding which can lead to representation collapse—the representations are biased towards known classes, where the dog class (novel, brown) and part of horse class samples (novel, grey) collapse into the cluster of the cat class (known, red), as shown in Figure 3a." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Effect of Sample-Target Allocation. Next, we ablate the effectiveness of the sample-target allocation mechanism in Table 2. For the coarse-level assignment, we test the variant TRAILER w/o Rough Assign which sticks to the vanilla softmax function for rough label assignment of \\( q \\) and suffers from rigorous performance degradation, especially on CIFAR-100. For the fine-level refinery, we test the variant TRAILER w/o Label Refinery which discards the refinery procedure of auxiliary head \\( h_{aux} \\). Although such a variant exhibits slightly better performance for the known classes, it significantly lags behind TRAILER for the novel classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To demonstrate it more intuitively, we visualize the accuracy of binary prediction for \\( h_{aux} \\) between known and novel classes in Figure 4a. We can find that \\( h_{aux} \\) quickly achieves high prediction accuracies and such accurate separation between known and novel indeed refines the rough assignment supplementarily, finally enhancing the pseudo-label precision on 9 of 10 classes, as shown in Figure 4b. Such a mechanism also alleviates the biased prediction, as shown in Figure 5 where a non-trivial amount of novel samples are" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "23078" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.27, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.211, + 0.224, + 0.223 + ], + "angle": 0, + "content": "(a) Equiangularity" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.09, + 0.465, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.211, + 0.434, + 0.223 + ], + "angle": 0, + "content": "(b) Maximal separation" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.235, + 0.47, + 0.29 + ], + "angle": 0, + "content": "Figure 6. Visualizations on the (a) equiangularity and (b) maximal separation properties of neural collapse on CIFAR-10 in the first 100 epochs. Representation calibration enables TRAILER to achieve feature arrangement with equal and maximal separation." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.306, + 0.462, + 0.392 + ], + "angle": 0, + "content": "
AblationEstimation of KKnownNovelAll
TRAILERKnown (100)69.748.755.6
ORCAKnown (100)66.943.048.1
TRAILEREstimated (124)71.144.353.3
ORCAEstimated (124)66.340.046.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.402, + 0.468, + 0.416 + ], + "angle": 0, + "content": "Table 4. Ablations with unknown class number \\( K \\) on CIFAR-100." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.445, + 0.454, + 0.46 + ], + "angle": 0, + "content": "misidentified as known (the number in red) for baselines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.47, + 0.743 + ], + "angle": 0, + "content": "Empirical Visualizations for Neural Collapse. In the previous section, we have provided the embedding visualization. Here we supply some quantitative results in addition. Recall that two core properties of NC are equiangularity and maximal separation. Given the centered class-means \\(\\hat{z}^k\\) for features from all classes \\(1 \\leq k \\leq K\\), for equiangularity, we calculate the standard deviation of the cosines between each pair \\(\\mathrm{Std}_{k \\neq k'}(\\cos (\\hat{z}^k, \\hat{z}^{k'}))\\). For maximal separation, we calculate the average value \\(\\mathrm{Avg}_{k \\neq k'}(\\cos (\\hat{z}^k, \\hat{z}^{k'}) + \\frac{1}{K - 1})\\) since the optimal pair-wise angle is \\(-\\frac{1}{K - 1}\\) as discussed in Eq. (3). The trends of these two metrics on the CIFAR-10 dataset with and without representation calibration are shown in Figure 6. Notably, equipped with representation calibration, the standard deviations of the cosines and shifted average cosine values approach zero, which indicates that TRAILER achieves the engaging feature structure that is close to neural collapse." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Unknown Number of Novel Classes. The previous experiments follow previous protocols and assume that the class number is known apriori [5, 15, 46], which is sometimes impractical in real-world applications and brings challenges for both TRAILER and the other baselines. A simple resolution is to first use estimation methods [32, 39] to derive the number of classes before deployment, for example, \\( k \\)-estimation with Brent's algorithm from [50] or cross-validating using labeled probe set from [5, 21, 46]. On" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.693, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.211, + 0.682, + 0.222 + ], + "angle": 0, + "content": "(a) Accuracy for known classes." + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.091, + 0.888, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.211, + 0.876, + 0.222 + ], + "angle": 0, + "content": "(b) Accuracy for novel classes." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.235, + 0.892, + 0.263 + ], + "angle": 0, + "content": "Figure 7. Accuracy comparisons for known and novel classes on CIFAR-10 with different novel class ratios ranging from 0.1 to 0.9." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.292, + 0.892, + 0.428 + ], + "angle": 0, + "content": "CIFAR-100 the former gets exactly 100 while the latter derives 124 and was adopted by us for further exploration. We then utilize this estimation result to rerun the experiment by initializing the corresponding number of classifier vectors. As shown in Table 4, TRAILER is able to achieve competitive performance even when the number of classes is not known in advance. We further provide additional results with different estimated class numbers \\( K \\) from 80 to 130 in Appendix B to demonstrate the robustness of TRAILER." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.452, + 0.892, + 0.557 + ], + "angle": 0, + "content": "Effect of Novel Class Ratio. Here we further evaluate the performance of TRAILER on the CIFAR-10 dataset with different ratios of novel classes from 0.1 to 0.9. As shown in Figure 7, we can see that TRAILER displays robust performance when the ratio of novel classes grows larger from 0.1 to 0.9 and consistently manifests superior performance than other baselines on both known and novel classes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.619, + 0.59 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.892, + 0.828 + ], + "angle": 0, + "content": "In this work, we propose a novel framework TRAILER for open-world SSL. We first take inspiration from the recently discovered neural collapse phenomenon and intend to attain its appealing feature arrangement with minimal within-class and maximum between-class covariance. To achieve this, we adopt a targeted classifier and align representations towards its pre-assigned optimal structure in a progressive manner. To mitigate the potential risk of misalignment, we tackle the sample-target assignment for unlabeled samples with a coarse-to-fine allocation mechanism. Comprehensive experiments show that TRAILER enhances the baselines by a notable margin on both generic and fine-grained benchmarks. We hope our work can inspire future research to leverage the appealing properties of the neural collapse phenomenon for tasks in open-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.679, + 0.862 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work is supported by the Pioneer R&D Program of Zhejiang (No. 2024C01035)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23079" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.157 + ], + "angle": 0, + "content": "[1] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR. OpenReview.net, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.469, + 0.2 + ], + "angle": 0, + "content": "[2] Jessa Bekker and Jesse Davis. Learning from positive and unlabeled data: a survey. Mach. Learn., 109(4):719-760, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.469, + 0.27 + ], + "angle": 0, + "content": "[3] David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution alignment and augmentation anchoring. CoRR, abs/1911.09785, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[4] David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In NeurIPS, pages 5050-5060, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[5] Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning. In ICLR. OpenReview.net, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.469, + 0.428 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.469, + 0.486 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9630-9640. IEEE, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.469, + 0.528 + ], + "angle": 0, + "content": "[8] Hui Chen, Fangqing Liu, Yin Wang, Liyue Zhao, and Hao Wu. A variational approach for learning from positive and unlabeled data. In NeurIPS 2020, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.469, + 0.585 + ], + "angle": 0, + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. PMLR, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[10] Xuxi Chen, Wuyang Chen, Tianlong Chen, Ye Yuan, Chen Gong, Kewei Chen, and Zhangyang Wang. Self-pu: Self boosted and calibrated positive-unlabeled training. In ICML, pages 1510-1519. PMLR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[11] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NeurIPS, pages 2292-2300, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE Computer Society, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[13] François Denis. PAC learning from positive statistical queries. In Algorithmic Learning Theory, 9th International Conference, ALT '98, Otzenhausen, Germany, October 8-10, 1998, Proceedings, pages 112-126. Springer, 1998. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR. OpenReview.net, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Enrico Fini, Enver Sangineto, Stephane Lathuilière, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "novel class discovery. In ICCV, pages 9264-9272. IEEE, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[16] Ronald A Fisher. The use of multiple measurements in taxonomic problems. Annals of eugenics, 7(2):179-188, 1936. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[17] Tomer Galanti, András György, and Marcus Hutter. On the role of neural collapse in transfer learning. In ICLR. Open-Review.net, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[18] Florian Graf, Christoph D. Hofer, Marc Niethammer, and Roland Kwitt. Dissecting supervised constrastive learning. In ICML, pages 3821-3830. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.251, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[19] Lan-Zhe Guo, Zhenyu Zhang, Yuan Jiang, Yu-Feng Li, and Zhi-Hua Zhou. Safe deep semi-supervised learning for unseen-class unlabeled data. In ICML, pages 3897-3906. PMLR, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[20] Lan-Zhe Guo, Yi-Ge Zhang, Zhi-Fan Wu, Jie-Jing Shao, and Yu-Feng Li. Robust semi-supervised learning when not all classes have labels. In NeurIPS, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.35, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[21] Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer clustering. In ICCV, pages 8400-8408. IEEE, 2019. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[22] Kai Han, Sylvestre-Alvise Rebuffi, Sébastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Automatically discovering and learning new visual categories with ranking statistics. In ICLR. OpenReview.net, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[23] X. Y. Han, Vardan Papyan, and David L. Donoho. Neural collapse under MSE loss: Proximity to and dynamics on the central path. In ICLR. OpenReview.net, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[24] Zongyan Han, Zhenyong Fu, Shuo Chen, and Jian Yang. Contrastive embedding for generalized zero-shot learning. In CVPR, pages 2371-2381. Computer Vision Foundation / IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[25] Zongyan Han, Zhenyong Fu, Shuo Chen, and Jian Yang. Semantic contrastive embedding for generalized zero-shot learning. Int. J. Comput. Vis., 130(11):2606-2622, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[26] Wenlong Ji, Yiping Lu, Yiliang Zhang, Zhun Deng, and Weijie J. Su. An unconstrained layer-peeled perspective on neural collapse. In ICLR. OpenReview.net, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[27] Ryuichi Kiryo, Gang Niu, Marthinus Christoffel du Plessis, and Masashi Sugiyama. Positive-unlabeled learning with non-negative risk estimator. In NeurIPS, pages 1675-1685, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[28] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV, pages 554–561. IEEE Computer Society, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[29] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Master's thesis, Department of Computer Science, University of Toronto, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[30] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In ICLR. OpenReview.net, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[31] Dong-Hyun Lee. Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks. ICML 2013 Workshop : Challenges in Representation Learning (WREPL), 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Collin Leiber, Lena G. M. Bauer, Benjamin Schelling, Christian Böhm, and Claudia Plant. Dip-based deep embedded" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23080" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "clustering with k-estimation. In KDD, pages 903-913. ACM, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.175 + ], + "angle": 0, + "content": "[33] Jiaming Liu, Yangqiming Wang, Tongze Zhang, Yulu Fan, Qinli Yang, and Junming Shao. Open-world semi-supervised novel class discovery. In *IJCAI*, pages 4002–4010. ijcai.org, 2023. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[34] Chuan Luo, Pu Zhao, Chen Chen, Bo Qiao, Chao Du, Hongyu Zhang, Wei Wu, Shaowei Cai, Bing He, Saravanakumar Rajmohan, and Qingwei Lin. PULNS: positive unlabeled learning with effective negative sample selector. In AAAI, pages 8784-8792. AAAI Press, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[35] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. CoRR, abs/1306.5151, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[36] Aleix M. Martínez and Avinash C. Kak. PCA versus LDA. IEEE Trans. Pattern Anal. Mach. Intell., 23(2):228-233, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[37] Duc Anh Nguyen, Ron Levie, Julian Lienen, Eyke Hüllermeier, and Gitta Kutyniok. Memorization-dilation: Modeling neural collapse under noise. In ICLR. OpenReview.net, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[38] Vardan Papyan, X. Y. Han, and David L. Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 117(40):24652-24663, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.483 + ], + "angle": 0, + "content": "[39] Dan Pelleg and Andrew W. Moore. X-means: Extending k-means with efficient estimation of the number of clusters. In ICML, pages 727-734. Morgan Kaufmann, 2000. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[40] Zhongzheng Ren, Raymond A. Yeh, and Alexander G. Schwing. Not all unlabeled data are equal: Learning to weight data in semi-supervised learning. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[41] Mamshad Nayeem Rizve, Navid Kardan, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Openldn: Learning to discover novel classes for open-world semi-supervised learning. In ECCV, pages 382-401. Springer, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.678 + ], + "angle": 0, + "content": "[42] Chuck Rosenberg, Martial Hebert, and Henry Schneiderman. Semi-supervised self-training of object detection models. In 7th IEEE Workshop on Applications of Computer Vision / IEEE Workshop on Motion and Video Computing (WACV/MOTION 2005), 5-7 January 2005, Breckenridge, CO, USA, pages 29-36. IEEE Computer Society, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.733 + ], + "angle": 0, + "content": "[43] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In NeurIPS, pages 1163-1171, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[44] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[45] Xin Sun, Zhenning Yang, Chi Zhang, Keck Voon Ling, and Guohao Peng. Conditional gaussian distribution learning for open set recognition. In CVPR, pages 13477-13486. Computer Vision Foundation / IEEE, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[46] Yiyou Sun and Yixuan Li. Opencon: Open-world contrastive learning. Trans. Mach. Learn. Res., 2023, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[47] Kiat Chuan Tan, Yulong Liu, Barbara Ambrose, Melissa Tulig, and Serge J. Belongie. The herbarium challenge 2019 dataset. CoRR, abs/1906.05372, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[48] Christos Thrampoulidis, Ganesh Ramachandra Kini, Vala Vakilian, and Tina Behnia. Imbalance trouble: Revisiting neural-collapse geometry. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[50] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. In CVPR, pages 7482-7491. IEEE, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[51] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In ICLR. OpenReview.net, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[52] Jiaheng Wei, Zhaowei Zhu, Hao Cheng, Tongliang Liu, Gang Niu, and Yang Liu. Learning with noisy labels revisited: A study using real-world human annotations. In ICLR. OpenReview.net, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[53] P. Welinder, S. Branson, T. Mita, C. Wah, F. Schroff, S. Belongie, and P. Perona. Caltech-UCSD Birds 200. Technical Report CNS-TR-2010-001, California Institute of Technology, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.421, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[54] Liang Xie, Yibo Yang, Deng Cai, and Xiaofei He. Neural collapse inspired attraction-repulsion-balanced loss for imbalanced learning. Neurocomputing, 527:60-70, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.463, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[55] Qizhe Xie, Zihang Dai, Eduard H. Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. In NeurIPS, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.506, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[56] Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In ICML, pages 11525-11536. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[57] Yibo Yang, Shixiang Chen, Xiangtai Li, Liang Xie, Zhouchen Lin, and Dacheng Tao. Inducing neural collapse in imbalanced learning: Do we really need a learnable classifier at the end of deep neural network? In NeurIPS, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[58] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip H. S. Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In ICLR. OpenReview.net, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[59] Bo Ye, Kai Gan, Tong Wei, and Min-Ling Zhang. Bridging the gap: Learning pace synchronization for open-world semi-supervised learning. CoRR, abs/2309.11930, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[60] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In NeurIPS 2021, pages 18408-18419, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[61] Bingchen Zhao, Xin Wen, and Kai Han. Learning semi-supervised gaussian mixture models for generalized category discovery. In ICCV, pages 16623-16633, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[62] Yunrui Zhao, Qianqian Xu, Yangbangyan Jiang, Peisong Wen, and Qingming Huang. Dist-pu: Positive-unlabeled learning from a label distribution perspective. In CVPR, pages 14441-14450. IEEE, 2022. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23081" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[63] Zhisheng Zhong, Jiequan Cui, Yibo Yang, Xiaoyang Wu, Xiaojuan Qi, Xiangyu Zhang, and Jiaya Jia. Understanding imbalanced semantic segmentation through neural collapse. In CVPR, pages 19550-19559. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[64] Jinxin Zhou, Xiao Li, Tianyu Ding, Chong You, Qing Qu, and Zhihui Zhu. On the optimization landscape of neural collapse under MSE loss: Global optimality with unconstrained features. In ICML, pages 27179-27202. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.47, + 0.261 + ], + "angle": 0, + "content": "[65] Zhihui Zhu, Tianyu Ding, Jinxin Zhou, Xiao Li, Chong You, Jeremias Sulam, and Qing Qu. A geometric analysis of neural collapse with unconstrained features. In NeurIPS, pages 29820-29834, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.47, + 0.304 + ], + "angle": 0, + "content": "[66] Zhaowei Zhu, Tianyi Luo, and Yang Liu. The rich get richer: Disparate impact of semi-supervised learning. In ICLR. OpenReview.net, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "23082" + } + ] +] \ No newline at end of file diff --git a/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/b33baf4c-8c46-4e13-a308-0177b1e2eee6_origin.pdf b/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/b33baf4c-8c46-4e13-a308-0177b1e2eee6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..44a1ef373d81a3b276cc9f39d6f49e120a2fc1bf --- /dev/null +++ b/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/b33baf4c-8c46-4e13-a308-0177b1e2eee6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9fbaa8c2b22aeda560c87a0e1366c2a2cfe4b5f4988e2c78d9ea4ae810097f45 +size 4179694 diff --git a/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/full.md b/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f0965fa673a4de969d68f372c349f15c4e83a941 --- /dev/null +++ b/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/full.md @@ -0,0 +1,329 @@ +# Targeted Representation Alignment for Open-World Semi-Supervised Learning + +Ruixuan Xiao $^{1}$ , Lei Feng $^{2}$ , Kai Tang $^{1}$ , Junbo Zhao $^{1}$ , Yixuan Li $^{3}$ , Gang Chen $^{1}$ , Haobo Wang $^{1*}$ + +1Zhejiang University, China + +$^{2}$ Singapore University of Technology and Design, Singapore + +3University of Wisconsin-Madison, USA + +{xiaoruixuan,tk0819,j.zhao,cg,wanghaobo}@zju.edu.cn,lfengqaq@gmail.com,sharonli@cs.wisc.edu + +# Abstract + +Open-world Semi-Supervised Learning aims to classify unlabeled samples utilizing information from labeled data, while unlabeled samples are not only from the labeled known categories but also from novel categories previously unseen. Despite the promise, current approaches solely rely on hazardous similarity-based clustering algorithms and give unlabeled samples free rein to spontaneously group into distinct novel class clusters. Nevertheless, due to the absence of novel class supervision, these methods typically suffer from the representation collapse dilemma—features of different novel categories can get closely intertwined and indistinguishable, even collapsing into the same cluster and leading to degraded performance. To alleviate this, we propose a novel framework TRAILER which targets to attain an optimal feature arrangement revealed by the recently uncovered neural collapse phenomenon. To fulfill this, we adopt targeted prototypes that are pre-assigned uniformly with maximum separation and then progressively align the representations to them. To further tackle the potential downsides of such stringent alignment, we encapsulate a sample-target allocation mechanism with coarse-to-fine refinery that is able to infer label assignments with high quality. Extensive experiments demonstrate that TRAILER outperforms current state-of-the-art methods on generic and fine-grained benchmarks. The code is available at https://github.com/Justherozen/TRAILER. + +# 1. Introduction + +The resounding achievements of deep learning heavily rely on large-scale and accurately-annotated training data, which is labor-intensive and time-consuming. To alleviate this, semi-supervised learning (SSL) has emerged as a prominent alternative [4, 55, 60], which exploits unlabeled data for boosted performance. Despite the success, current SSL + +![](images/d2b29116574fc6e9c93ae3ff3977346e698f8b29a8a029afeaadff6616eb96bc.jpg) +Figure 1. Left: Scenario of open-world SSL where unlabeled samples are from known and novel classes. Right: Visualization without and with our targeted calibration mechanism. Without calibration, representations tend to get jumbled and inseparable, as in the left where dog and cat categories converge into the same cluster. + +methods are primarily based on the closed-world assumption that labeled and unlabeled data share the same predefined label space. Nevertheless, this assumption rarely holds in real-world applications, where the model may encounter novel classes previously unseen in unlabeled data. To relax this assumption, novel category discovery (NCD) was first proposed [15, 21, 22] which assumes all unlabeled data strictly belong to disjoint novel classes. Recently, a more practical setting, open-world semi-supervised learning (open-world SSL), also referred to as generalized category discovery (GCD), has received huge attention [5, 46, 50] by generalizing standard SSL and NCD. It assumes that unlabeled data contain both known classes from the labeled data and similar yet distinct novel classes. + +The goal of open-world SSL is to classify the already-known classes from the labeled data, as well as to discover new classes in unlabeled data without annotations and assign instances to them. Nascent efforts have been made to address this practical yet under-investigated task [20, 33, 41, 50]. The essence of these recent works is to rashly rely on the similarity-based [5, 41] or contrastive-based [46] clustering algorithm in feature space to spontaneously + +group different samples into separate clusters. Some methods further integrate self-training techniques with generated pseudo-labels for boosted performance [20, 33]. However, due to the missing supervision and semantic shift for novel categories, the self-organized clustering step can get overfitted and biased towards the known classes, leading to a representation collapse dilemma—the feature distance between one novel category and another category may get too close and finally become inseparable, even converging within a single cluster and thus jeopardizing generalization performance, as illustrated in the left of Figure 1. + +On the other hand, a recent study has unveiled a feature-classifier alignment phenomenon termed neural collapse (NC) [38]. It indicates that under ideal conditions, the last-layer features of the same class tend to collapse into their within-class mean. Meanwhile, these within-class means of all classes will be aligned with classifiers and formed as a simplex equiangular tight frame (ETF) at the end of training, as elaborated in Section 3.1. Intuitively, such an elegant NC phenomenon affords an optimal feature structure with minimized within-class variance and maximized between-class variance [36], coinciding with what we fervently desire. Nevertheless, in open-world SSL, the absence of novel class annotation can hinder the representation learning and thus disrupt the natural induction of NC. + +In an effort to salvage this and accomplish the appealing feature arrangement of NC, we propose a novel framework, Targeted Representation AIIgnment for open-world semisupervised LEaRning (dubbed TRAILER). Key to our method, TRAILER leverages a targeted classifier to progressively align embeddings to the pre-assigned ETF structure, eventually fulfilling the optimal state of uniformly and maximally separated embeddings. Despite the promise, this calibration procedure is still susceptible to false alignment stemming from unreliable pseudo-labels. To mitigate this, we integrate a hierarchical sample-target allocation strategy that comprises two sequential steps: (i)-rough sample assignment based on the optimal transport mechanism; (ii)-label refinery by casting the known-novel separation problem to a specific weakly-supervised paradigm termed PU learning [2, 13, 62]. Such an allocation mechanism can provide more precise pseudo-labels and thus facilitate representation calibration. Through our visualized results in Figure 1 and Section 4.3, TRAILER does indeed induce discriminative and non-skewed embeddings. Empirically, TRAILER surpasses state-of-the-art approaches on different evaluation benchmarks, e.g., improves the best baseline by $3.1\%$ , $2.9\%$ , and $1.5\%$ overall accuracy on CIFAR-10, CIFAR-100 and ImageNet-100 datasets, respectively. + +# 2. Related Work + +Closed-World Semi-Supervised Learning has been widely explored [4, 44, 55, 66] to learn from limited la + +beled and massive unlabeled data. Pseudo labeling [31, 42] and consistency regularization [30, 44, 55] are two prominent techniques. The former is based on self-training that utilizes model predictions as optimization targets [40, 56], while the latter enforces model to produce consistent outputs for different views of the same image [3, 4, 43, 55]. However, these methods are primarily based on the closed-world assumption that labeled and unlabeled data share the same pre-defined label space, which limits their real-world application where unlabeled samples from previously unseen classes may emerge. Open-set SSL [19, 24, 25, 45] relaxes this by simply rejecting all unseen class instances, but still fails to classify them into distinct novel categories. + +Open-World Semi-Supervised Learning further accommodates real-world scenarios [5, 20, 46]. It is also known as generalized category discovery (GCD) [50] and is a natural extension of novel category discovery (NCD). NCD is first introduced in [21] and improved in later works [15, 22] to discover new categories by leveraging the knowledge from labeled known categories. However, NCD assumes that all unlabeled samples belong to novel classes and thus fails to recognize known class samples. Most recently, open-world SSL generalizes NCD to identify unlabeled samples from both known and novel classes [5]. To tackle this, several incipient solutions have been proposed [20, 59], which typically involve similarity-based representation learning and then involve sample clustering to generate pseudo-labels for self-training. ORCA [5] is the pioneering work that integrates an uncertainty-based margin loss to control the intraclass variance of different classes. OpenCon [46] further derives a semi-supervised constrastive mechanism to produce compact feature space. Despite the promise, these methods solely rely on spontaneous clustering based on sample similarity, overlooking potential risks that representations of different classes possibly converge into the same cluster, thus leading to the representation collapse dilemma. + +Neural Collapse describes an elegant alignment phenomenon [38], which states that the last-layer features will collapse into their within-class centers and these centers together with classifiers will eventually form a simplex equiangular tight frame (ETF). Stemming from the appealing symmetry, a plethora of studies have been triggered to explain this phenomenon theoretically [26, 48, 58, 65]. NC is proved to be the global optimality under ideal conditions with the CE [18, 26] and the MSE [23, 64] loss functions. The NC phenomenon has also been investigated in some specific scenarios like imbalanced learning [54, 57, 63] noisy label learning [37, 52], transfer learning [17] and incremental learning [58]. In our work, we investigate the opportunities of inducing such optimal feature arrangement of NC through targeted calibration for open-world SSL. + +![](images/206d0955d948800be5240746976bb41019e819af1a2cd14fbb43518183211c80.jpg) +Figure 2. Overview of TRAILER. For representation calibration, we leverage the targeted classifier $\phi^{etf}$ to align the feature towards the optimal arrangement of ETF progressively while maintaining the within-class compactness with prototypical clustering. For sample-target allocation, the unlabeled samples are provided with reliable pseudo-labels with rough assignment and subsequent label refinery. + +![](images/08e30f5355d1bb8d4a427b7ffb323661198a6beb378faa3dffb95265494d562f.jpg) + +# 3. Method + +Problem Definition. In open-world SSL, we assume that the training dataset $\mathcal{D} = \mathcal{D}_l \cup \mathcal{D}_u$ comprises two subsets: a labeled set $\mathcal{D}_l = \{x_i, y_i\}_{i=1}^m$ where the input image $x_i \in \mathbb{R}^d$ and the ground-truth label $y_i$ come from a set of known classes $\mathcal{C}_{known}$ , and an unlabeled set $\mathcal{D}_u = \{x_i\}_{i=1}^n$ where the potential ground-truth label is from a set of classes $\mathcal{C}_{all}$ . In closed-world SSL, the unlabeled samples strictly come from known classes, i.e., $\mathcal{C}_{all} = \mathcal{C}_{known}$ . As for open-world SSL which is more practical, the unlabeled samples can come from either known classes $\mathcal{C}_{known}$ or novel classes, i.e., $\mathcal{C}_{known} \subset \mathcal{C}_{all}$ . We consider $\mathcal{C}_{novel} = \mathcal{C}_{all} \backslash \mathcal{C}_{known}$ as the set of novel classes. Following previous setting [5, 15, 46], we assume the number of novel classes $|\mathcal{C}_{novel}|$ is known apriori, which can also be estimated using the off-the-shelf methods [21, 50] as discussed in Section 4.3. + +Learning Goal. For classification problem with $K = |\mathcal{C}_{all}|$ classes, open-world SSL aims to obtain a model $f_{\theta,\phi} : \mathbb{R}^d \mapsto \mathbb{R}^K$ that recognizes samples from both known and novel classes based on $\mathcal{D}_l$ and $\mathcal{D}_u$ . Specifically, $f_{\theta,\phi} = h_\phi \circ g_\theta$ can be decomposed into a representation backbone $g_\theta$ and a classification head $h_\phi$ with weights $\phi = [\phi_1, \dots, \phi_k] \in \mathbb{R}^{P \times K}$ , for which we adopt a prototypical one following [15] with logit $h(z) = \mathrm{Norm}(z)^\top \mathrm{Norm}(\phi)$ . Formally, given the weakly and strongly augmented view $x_i$ and $x_i'$ of the same image in batch $\mathcal{B}$ , we define $z_i, z_i' \in \mathbb{R}^P$ as the last-layer $P$ -dimensional features and $p_i = \mathrm{softmax}(h(z_i))$ as output probability. Normalization for features is denoted as $\mathrm{Norm}(z) = z / ||z||_2$ and omitted in the subsequent. + +Warm-up Phase. Similar to previous works [5, 33], we also include a commonly adopted warm-up strategy at + +the start of training. The overall objectives to cluster known classes and discover potential novel samples consist of three losses: 1) the cross-entropy loss (CE) $\mathcal{L}_{ce}$ on labeled samples $\mathcal{D}_l$ which is elaborated later in Section 3.1; 2) the self-supervised contrastive loss $\mathcal{L}_{con} = -\log \frac{\exp(z_i^\top z_i' / \tau)}{\sum_{j=1}^n \exp(z_i^\top z_j' / \tau)}$ on all samples; and 3) the entropy regularization $\mathcal{L}_{ent} = \sum_{k=1}^{K} \overline{p}^k \log (\overline{p}^k)$ for preventing the trivial solution of assigning all instances to the same class, where $\overline{p} = \frac{1}{|\mathcal{B}|} \sum_{x_i \in \mathcal{B}} p_i$ is the average probability of the batch. The overall losses for warm-up are formulated, + +$$ +\mathcal {L} _ {\text {w a r m}} = \mathcal {L} _ {c e} + \mathcal {L} _ {\text {c o n}} + \alpha \mathcal {L} _ {\text {e n t}} \tag {1} +$$ + +where $\alpha$ is a weighting parameter. After warm-up, we also would like to group unlabeled samples into distinct clusters and assign them pseudo-labels for self-training. As mentioned earlier, the crucial challenge is how to abstain from the representation collapse dilemma and obtain discriminative features without novel class annotations. To fulfill this, we first draw inspiration from neural collapse phenomenon. + +# 3.1. Neural Collapse for Representation Calibration + +Recent works have disclosed the Neural Collapse phenomenon [17, 38] which corresponds to an optimal feature-classifier alignment towards a simplex equiangular tight frame (ETF). This offers a uniformly and equally separated feature arrangement that is desirable for open-world SSL. + +Definition 1 (Simplex ETF) A Simplex equiangular tight frame (ETF) refers to a collection of $K$ equal-length and maximally-equiangular $P$ -dimensional embedding vectors $\mathbf{E} = [e_1, \dots, e_K] \in \mathbb{R}^{P \times K}$ which satisfies: + +$$ +\mathbf {E} = \sqrt {\frac {K}{K - 1}} \mathbf {U} \left(\mathbf {I} _ {K} - \frac {1}{K} \mathbf {1} _ {K} \mathbf {1} _ {K} ^ {\top}\right) \tag {2} +$$ + +where $\mathbf{I}_K$ is the identity matrix, $\mathbf{1}_K$ is an all-ones vector, and $\mathbf{U} \in \mathbb{R}^{P \times K}$ ( $P \geq K$ ) allows a rotation. + +All vectors in a simplex ETF $\mathbf{E}$ have an equal $\ell_2$ norm and the same pair-wise maximal equiangular angle $-\frac{1}{K - 1}$ [57], + +$$ +e _ {k _ {1}} ^ {\top} e _ {k _ {2}} = \frac {K}{K - 1} \delta_ {k _ {1}, k _ {2}} - \frac {1}{K - 1}, \forall k _ {1}, k _ {2} \in [ 1, K ] \quad (3) +$$ + +where $\delta_{k_1,k_2} = 1$ when $k_{1} = k_{2}$ and 0 otherwise. The NC phenomenon can be described by the following properties under ideal conditions: (1) The last-layer features converge to their within-class means, and the within-class means of all classes converge to different vertices of a simplex ETF. (2) The normalized classifier weights converge to the same simplex ETF. Notably, this process is implicit and natural since it is induced with common optimizers such as SGD, rather than enforced by an explicit objective. + +As discussed above, NC has provided an optimal ETF structure for feature arrangement with minimal within-class covariance and maximum between-class covariance (i.e., maximized Fisher discriminant ratio [16]), which coincides with our goal of deriving discriminative and non-skewed embeddings. However, NC is delicate and can fail to be naturally induced under imperfect conditions [37, 57]. For open-world SSL, the missing novel class annotations can also lead to biased features and break the NC optimality. + +To remedy this, we resort to effectuating an explicit NC and fulfilling the appealing feature arrangement with representation calibration, which is achieved by leveraging a targeted prototypical classifier $h_{\text{etf}}$ . In specific, $h_{\text{etf}}$ is preassigned with weights $\phi^{\text{etf}} \in \mathbb{R}^{P \times K}$ , which is a random simplex ETF with optimal structure as Definition 1 and is fixed rather than learnable during training. By assuming a known label $\hat{y}_i$ for each sample $x_i$ , the standard CE loss $\mathcal{L}_{ce}$ is then replaced as $\mathcal{L}_{\text{etf}}$ , which is formulated, + +$$ +\mathcal {L} _ {e t f} \left(x _ {i}, \hat {y} _ {i}\right) = - \log \frac {\exp \left(z _ {i} ^ {\top} \cdot \phi_ {\hat {y} _ {i}} ^ {e t f} / \tau\right)}{\sum_ {k = 1} ^ {K} \exp \left(z _ {i} ^ {\top} \cdot \phi_ {k} ^ {e t f} / \tau\right)} \tag {4} +$$ + +where $\tau$ is the temperature parameter. Through $\mathcal{L}_{\text{etf}}$ , the feature $z_{i}$ is drawn closer to its corresponding target $\phi_{\hat{y}_i}^{\text{etf}}$ and pulled away from other class vectors of ETF structure. + +For the labeled sample $(x_{i},y_{i})$ , we directly set $\hat{y}_i = y_i$ . As for unlabeled data without ground-truth labels, we design a sample-target allocation strategy to generate pseudolabels $\hat{y}$ , whose details are shown in Section 3.2. + +Progressive Alignment. Notably, if we directly optimize the loss $\mathcal{L}_{\text{etf}}$ after the warm-up phase, the provisionally clustered samples would be forcibly dragged to the corresponding targeted vector $\phi_{\hat{y}}^{\text{etf}}$ . As a result, the already-learned clustering compactness can get compromised and the within-cluster feature distribution can get jumbled. + +To address this, we intend to align the representations in a progressive manner to preserve the compactness within each cluster during calibration. Formally, given sample $x_{i}$ with pseudo-label $\hat{y}_i$ , we also maintain the authentic prototypes $\mu_k = \mathrm{Norm}(\mathrm{Avg}_{\hat{y}_i = k}(z_i))$ for each class $1 \leq k \leq K$ . To preserve the compactness of cluster structure, a prototypical clustering loss is also integrated, + +$$ +\mathcal {L} _ {\text {p r o t o}} \left(z _ {i}, \hat {y} _ {i}\right) = - \log \frac {\exp \left(z _ {i} ^ {\top} \cdot \mu_ {\hat {y} _ {i}} / \tau\right)}{\sum_ {k = 1} ^ {K} \exp \left(z _ {i} ^ {\top} \cdot \mu_ {k} / \tau\right)} \tag {5} +$$ + +The overall classification loss is then formulated, + +$$ +\mathcal {L} _ {c l s} = \lambda \mathcal {L} _ {p r o t o} + (1 - \lambda) \mathcal {L} _ {e t f} \tag {6} +$$ + +where $\lambda$ is a calibration strength parameter that is gradually ramped down from 1 to 0. With such a progressive manner, the embeddings are cautiously and smoothly drawn closer to the targeted optimal arrangement of ETF while the within cluster compactness is preserved. + +# 3.2. Hierarchical Sample-Target Allocation + +In previous steps, we involve targeted classifiers to regulate representation learning. Nevertheless, our stringent calibration still holds unexpected risks. Lacking supervision, the calibration process for unlabeled samples largely relies on the quality of generated pseudo-labels. When the pseudolabels are unreliable, the feature alignment procedure can be misguided towards the wrong cluster. To overcome this, we introduce our hierarchical sample-target allocation mechanism with course-to-fine assignment in what follows. + +Rough Assignment through Optimal Transport. For pseudo-label generation, one may adopt model prediction $\arg \max_{1\leq k\leq K}h^{k}(z_{i})$ directly. However, this may lead to degenerate solutions where data points of different categories map into a single category [1], thus resulting in representation collapse. To mitigate this, we adopt a widely-used equipartition constraint [6, 15] for rough label assignment, which is attained by solving an optimal transport problem. + +Formally, given a batch $\mathcal{B}$ of $b$ unlabeled samples with the logit matrix $\mathbf{P} = [h(z_1),\dots,h(z_b)]$ whose columns are output logits by $h_{\mathrm{etf}}$ and the label-assignment matrix $\mathbf{Q} = [q_1,\dots,q_b]\in [0,1]^{K\times b}$ whose columns are the to-beassigned soft labels $q_{i}$ . $\mathbf{Q}$ is estimated by, + +$$ +\begin{array}{l} \mathbf {Q} = \max _ {\mathbf {Q} \in \Gamma} \operatorname {T r} (\mathbf {Q} ^ {\top} \mathbf {P}) + \epsilon \mathrm {H} (\mathbf {Q}) \\ \text {s . t .} \Gamma = \left\{\mathbf {Q} \in \mathbb {R} _ {+} ^ {K \times b} \mid \mathbf {Q} \mathbf {1} _ {b} = \frac {1}{K} \mathbf {1} _ {k}, \mathbf {Q} ^ {\top} \mathbf {1} _ {K} = \frac {1}{b} \mathbf {1} _ {b} \right\} \tag {7} \\ \end{array} +$$ + +where $\mathrm{H}(\cdot)$ is the entropy function, $\epsilon$ is a hyperparameter and $\operatorname{Tr}(\cdot)$ is the trace function. The solution of Eq. (7) is obtained using the Sinkhorn-Knopp algorithm [11]. Please refer to Appendix A for more optimization details. + +Label Refinery with Known-Novel Separation. After the rough assignment, we obtain unpolished soft labels $q \in [0,1]^K$ . However, the aforementioned skewed predictions towards known classes stemming from overfitting remain unresolved. To rectify this, we prioritize the subtask of precisely demarcating known and novel classes for label refinery, which is achieved by including an auxiliary binary head $h_{aux}$ and reformulating this binary task as the classical positive unlabeled (PU) learning problem [2, 13, 62]. + +PU learning tackles a special case of binary classification where the model has access to a labeled set containing merely positive samples and an unlabeled set with both positive and negative samples. Thus, the task of distinguishing known and novel classes can be intuitively converted to PU learning by regarding known and novel as positive and negative respectively. A flurry of methods has been designed for PU learning [10, 27, 34], among which we adopt a prior-free variational algorithm [8] with loss $\mathcal{L}_{aux}$ to optimize the newly devised $h_{aux}$ . See Appendix A for more details. + +After that, we elicit the binary prediction $\hat{y}^{aux}$ for $h_{aux}$ , where 1 and 0 indicate prediction of known and novel respectively. Then we leverage $\hat{y}^{aux}$ to refine the label $q$ , + +$$ +q ^ {j} = \left\{ \begin{array}{l l} q ^ {j} & \text {i f} \mathbb {I} (j \in \mathcal {C} _ {\text {k n o w n}}) = \hat {y} ^ {\text {a u x}} \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {8} +$$ + +where $q^{j}$ is the $j$ -th entry of $q$ , $1 \leq j \leq K$ and $\mathbb{I}(\cdot)$ is the indicator function. That is, we retain the known/novel class components of soft label $q$ based on prediction $\hat{y}^{aux}$ and discard the remaining inconsistent components. The refined $q$ are then utilized for the final hard pseudo-label generation $\hat{y}_{i} = \arg \max_{1 \leq k \leq K} q^{k}$ in $\mathcal{L}_{cls}$ of Eq. (6). + +Pseudo-Label Filtering. While the proposed sample-target allocation mechanism can enhance the quality of pseudo-labels, the remaining obstinate noisy pseudo-labels can still impede model training, especially in the early stages when the model output is not accurate enough. To neutralize this, we further utilize a high-confidence selection to select clean pseudo-labels from noisy ones for $\mathcal{D}_u$ . Formally, for each class $1 \leq j \leq K$ , we select a clean subset $\mathcal{D}_{sel}^j$ from $\mathcal{D}_u^j = \{(x, \hat{y}) \in \mathcal{D}_u | \hat{y} = j\}$ class by class with the highest confidence scores in the first $R$ percent, + +$$ +\mathcal {D} _ {s e l} ^ {j} = \left\{\left(x, \hat {y}\right) \in \mathcal {D} _ {u} ^ {j} \mid r a n k \left(q ^ {j}\right) < R \% \right\} \tag{9} +$$ + +Then the integral selected set is merged as $\mathcal{D}_{sel} = \cup_{j=1}^{K} \mathcal{D}_{sel}^{j}$ . The classification loss in Eq. (6) is then merely valid on $\mathcal{D}_{sel}$ . In practice, we set a small $R$ to ensure the precision of selected samples at the beginning and gradually increase it to integrate more clean samples. Finally, the overall training loss is given by, + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {c l s}} + \mathcal {L} _ {\text {a u x}} + \mathcal {L} _ {\text {c o n}} + \alpha \mathcal {L} _ {\text {e n t}} \tag {10} +$$ + +where $\mathcal{L}_{con}$ and $\mathcal{L}_{ent}$ are described in Eq. (1). Through this procedure, Our sample-target allocation mechanism is able to provide pseudo-labels with high quality and guide the representation calibration toward the correct targets. The overview of TRAILER is illustrated in Figure 2. + +# 4. Experiment + +In this section, we present the main results and part of the ablation results to verify the effectiveness of TRAILER. More experimental details and results, such as the results with less labeled data, can be found in Appendix. + +# 4.1. Setup + +Datasets. We first evaluate TRAILER on three datasets CIFAR-10, CIFAR-100 [29] and ImageNet-100, which is randomly sub-sampled with 100 classes from ImageNet [12] following [5]. For each dataset, We first divide classes into the first $50\%$ known and the rest $50\%$ novel classes following [5]. Then we randomly select a portion $(50\%)$ of samples from the known classes as labeled $\mathcal{D}_l$ , and the remaining along with all novel class samples are regarded as unlabeled $\mathcal{D}_u$ . We also conduct experiments with different novel class ratios and on the fine-grained semantic shift benchmarks (SSB) [51], including CUB-200 [53], Stanford Cars [28], FGVC-Aircraft [35], and a naturally long-tailed Herbarium 19 [47] datasets, as discussed in Section 4.3. + +Baselines. We compare TRAILER with various open-world SSL baselines, including ORCA [5], GCD [50], OpenCon [46], NACH [20], OpenLDN [41], GPC [61] and OpenNCD [33]. We also provide comparisons with SSL, open-set SSL, and NCD baselines following [20, 33]. For traditional and open-set SSL baselines including FixMatch [44], $\mathrm{DS}^3\mathrm{L}$ [19], and CGDL [45], we extend them to be applicable to novel classes by selecting samples with lower confidence scores as novel and applying $k$ -means clustering to obtain results. For NCD methods including DTC [21], RankStats [22], and $\mathrm{UNO+}$ [15], they are extended to classify known classes by using Hungarian algorithm to match some discovered classes with known classes. For performances of baselines, we directly adopt reported results from receptive papers or from [33], except OpenLDN was retested under transductive setting. All compared methods are implemented based on pre-trained ResNet using SimCLR [9] following [5], except DTC and OpenCon which have specialized pre-training procedure. We run all experiments three times and report the averaged results. + +Evaluation Metrics. Following [5, 33], we evaluate the transductive performance on both known and novel classes. For novel classes, we utilize the Hungarian algorithm to derive the optimal assignments and then calculate accuracy + +
MethodsCIFAR-10CIFAR-100ImageNet-100
KnownNovelAllKnownNovelAllKnownNovelAll
FixMatch [44]71.550.449.539.623.520.365.836.734.9
DS3L [19]77.645.340.255.123.724.071.232.530.8
CGDL [45]72.344.639.749.322.523.567.333.831.9
DTC [21]53.939.538.331.322.918.325.620.821.3
RankStats [22]86.681.082.936.428.423.147.328.740.3
SimCLR [9]58.363.451.728.621.122.339.535.736.9
ORCA [5]88.290.489.766.943.048.189.172.177.8
GCD [50]78.479.779.168.533.545.282.358.368.2
OpenLDN [41]92.386.888.667.340.449.474.545.755.3
OpenCon [46]90.491.189.369.147.852.790.680.883.8
NACH [20]89.592.291.368.747.052.191.075.579.6
OpenNCD [33]88.490.690.169.743.449.390.077.581.6
TRAILER (Ours)93.495.094.469.748.755.691.482.485.3
+ +Table 1. Accuracy comparison of known, novel, and all classes on CIFAR-10, CIFAR-100 and ImaneNet-100 dataset. The dataset is composed of $50\%$ known classes and $50\%$ novel classes, with $50\%$ of the known classes labeled. Bold entries indicate superior results. + +
AblationCIFAR-10CIFAR-100
KnownNovelAllKnownNovelAll
TRAILER93.495.094.469.748.755.6
w/o Rep Calibration94.158.970.669.214.933.4
with λ = 093.991.892.570.246.654.6
w/o Rough Assign92.992.392.569.337.948.3
w/o Label Refinery95.289.691.570.345.453.7
w/o PL Filtering92.290.591.069.943.652.2
+ +Table 2. Ablation results on CIFAR-10 and CIFAR-100. 'Rep' and 'PL' indicate representation and pseudo-label respectively. + +$\frac{1}{n}\sum_{i = 1}^{n}\mathbb{I}(y_i = \text{perm}(\hat{y}_i))$ where $\text{perm}(\cdot)$ is the matching permutation. The overall joint accuracy is also measured by Hungarian match using both known and novel classes. + +Implementation Details. For CIFAR experiments, we use ResNet-18 as the backbone. The model is trained for 200 epochs with warm-up of 15 epochs. We employ the Adam optimizer with a batch size of 200 and an initial learning rate of $5e^{-4}$ , which decays by a cosine scheduler. The temperature $\tau$ is set as 0.3 and 0.5 for CIFAR-10 and CIFAR-100. We linearly ramp down $\lambda$ from 1 to 0 and ramp up selection ratio $R\%$ from 0.3 to 0.9 for pseudo-label filtering. The label refinery is enabled after training $h_{aux}$ for 15 epochs. For the optimization of Sinkhorn-Knopp, we adopt hyperparameters from [15] and set $\epsilon = 0.05$ with 3 iterations. We put more implementation details for ImageNet-100 and fine-grained dataset in Appendix A. + +# 4.2. Main Reults + +Table 1 shows the comparison results on the CIFAR-10, CIFAR-100, and ImageNet-100 datasets, including + +![](images/d80f3225708c2a7771dd47ba19da40f81c48c8372df5f7c1b33659e51222812d.jpg) +(a) Without calibration. + +![](images/9793e6453ee8e88732ea8c24047e8d31aa01041aaf25485ae5e180474f2aadd5.jpg) +(b) With calibration. +Figure 3. T-SNE feature visualization on CIFAR-10. Different colors represent the corresponding ground-truth classes. Without calibration, the dog and cat classes collapse into the same cluster. + +the accuracy of the 'Known', 'Novel', and 'All' classes. TRAILER significantly outperforms all the rivals by a considerable margin in terms of overall accuracy on different datasets. Specifically, on CIFAR-10, TRAILER improves upon the best baseline by $1.1\%$ , $2.8\%$ , and $3.1\%$ for the 'Known', 'Novel', and 'All' classes respectively. Moreover, on the more challenging CIFAR-100 and ImageNet-100 datasets with larger label space, while most baselines demonstrate a notable performance drop, TRAILER shows robustness and reaches an extraordinary trade-off between the known and novel classes. It consistently maintains superior performance with $2.9\%$ and $1.5\%$ overall enhancement. These results empirically demonstrate the efficacy of our proposed TRAILER for the task of open-world SSL. + +# 4.3. Analysis + +Results on Semantic Shift Benchmarks. We further conduct experiments on fine-grained semantic shift benchmarks (SSB) [51] following previous work [50, 61]. For fair comparisons, we adopt the same training and compar + +
MethodsCUBStanford CarsFGVC-AircraftHerbarium 19
KnownNovelAllKnownNovelAllKnownNovelAllKnownNovelAll
k-means38.932.134.310.613.812.814.416.816.012.912.812.9
RankStats+ [22]51.624.233.361.812.128.336.422.226.955.812.827.9
UNO+ [15]49.028.135.170.518.635.556.432.240.353.714.728.3
ORCA [5]45.630.235.350.110.723.531.817.122.030.915.520.9
GCD [50]56.648.751.357.629.939.041.146.945.051.027.035.4
OpenCon [46]63.852.154.778.632.749.1---58.928.639.3
GPC [61]58.253.155.459.232.842.842.547.946.351.727.936.5
TRAILER (Ours)71.361.965.171.747.655.462.650.554.557.037.844.5
+ +Table 3. Results on the Semantic Shift Benchmarks. Bold entries indicate superior results and blank ones indicate results are not provided. + +![](images/ee64ebf940375ccd6ac8ad37a41a8d6be115f5df96dec19ffde11e9399fc5440.jpg) +(a) Binary accuracy of $h_{aux}$ . + +![](images/2672ce8cfb7404415f6ba6e8148ac5c272989e7bc84c0f88fca993d2e70743d2.jpg) +(b) Pseudo-label precision. +Figure 4. (a) Accuracy of binary prediction between known and novel classes in first 100 epochs. (b) Per-class pseudo-label precision for unlabeled samples on CIFAR-10 with and without label refinery. Class indexes 0-4 are known while 5-9 are novel. + +isons protocol from [50] and utilize ViT-B/16 backbone [14] with DINO pre-trained weights [7]. As shown in Table 3, it can be observed that TRAILER exceeds the baselines by a substantial lead in all metrics, with improvements of $8.8\%$ , $14.8\%$ , $2.6\%$ , and $9.2\%$ for 'Novel' accuracy on these datasets. These results consistently demonstrate its superior performance, even under challenging fine-grained tasks. + +Benefits of Representation Calibration. To explore the benefits of representation calibration, we conduct ablations and equip TRAILER with different representation learning strategies: 1) TRAILER w/o Rep Calibration which dismiss the feature alignment of the targeted classifier and instead utilize the vanilla learnable classifier; 2) TRAILER with $\lambda = 0$ which discards the progressive manner for direct calibration. From Table 2, although TRAILER w/o Rep Calibration remains competitive on known classes, it suffers from a calamitous performance plunge for novel classes, which indicates the efficacy of representation calibration. TRAILER with $\lambda = 0$ achieves decent performance, but still underperforms TRAILER perceptibly, indicating the advantage of the progressive manner. + +We further visualize the representation with and without representation calibration on CIFAR-10 using t-SNE [49] + +![](images/1a0d947f2818acd11e2df042fe87a7e951f7a25822e96b25c783e6ac549d63e9.jpg) +Figure 5. Confusion matrix between known and novel classes on CIFAR-10. TRAILER can alleviate biased predictions. + +in Figure 3 and in Appendix B. It can be shown that representation calibration facilitates inducing compact and nonskewed representations, discarding which can lead to representation collapse—the representations are biased towards known classes, where the dog class (novel, brown) and part of horse class samples (novel, grey) collapse into the cluster of the cat class (known, red), as shown in Figure 3a. + +Effect of Sample-Target Allocation. Next, we ablate the effectiveness of the sample-target allocation mechanism in Table 2. For the coarse-level assignment, we test the variant TRAILER w/o Rough Assign which sticks to the vanilla softmax function for rough label assignment of $q$ and suffers from rigorous performance degradation, especially on CIFAR-100. For the fine-level refinery, we test the variant TRAILER w/o Label Refinery which discards the refinery procedure of auxiliary head $h_{aux}$ . Although such a variant exhibits slightly better performance for the known classes, it significantly lags behind TRAILER for the novel classes. + +To demonstrate it more intuitively, we visualize the accuracy of binary prediction for $h_{aux}$ between known and novel classes in Figure 4a. We can find that $h_{aux}$ quickly achieves high prediction accuracies and such accurate separation between known and novel indeed refines the rough assignment supplementarily, finally enhancing the pseudo-label precision on 9 of 10 classes, as shown in Figure 4b. Such a mechanism also alleviates the biased prediction, as shown in Figure 5 where a non-trivial amount of novel samples are + +![](images/58b98580f4eedf51fa54f741d2dba1253968974d3bdf02a037548209fb8c581f.jpg) +(a) Equiangularity + +![](images/412f447defbe88d237d77246a05c22bc5320c1a434804aeba54c83ec05d0de39.jpg) +(b) Maximal separation +Figure 6. Visualizations on the (a) equiangularity and (b) maximal separation properties of neural collapse on CIFAR-10 in the first 100 epochs. Representation calibration enables TRAILER to achieve feature arrangement with equal and maximal separation. + +
AblationEstimation of KKnownNovelAll
TRAILERKnown (100)69.748.755.6
ORCAKnown (100)66.943.048.1
TRAILEREstimated (124)71.144.353.3
ORCAEstimated (124)66.340.046.4
+ +Table 4. Ablations with unknown class number $K$ on CIFAR-100. + +misidentified as known (the number in red) for baselines. + +Empirical Visualizations for Neural Collapse. In the previous section, we have provided the embedding visualization. Here we supply some quantitative results in addition. Recall that two core properties of NC are equiangularity and maximal separation. Given the centered class-means $\hat{z}^k$ for features from all classes $1 \leq k \leq K$ , for equiangularity, we calculate the standard deviation of the cosines between each pair $\mathrm{Std}_{k \neq k'}(\cos (\hat{z}^k, \hat{z}^{k'}))$ . For maximal separation, we calculate the average value $\mathrm{Avg}_{k \neq k'}(\cos (\hat{z}^k, \hat{z}^{k'}) + \frac{1}{K - 1})$ since the optimal pair-wise angle is $-\frac{1}{K - 1}$ as discussed in Eq. (3). The trends of these two metrics on the CIFAR-10 dataset with and without representation calibration are shown in Figure 6. Notably, equipped with representation calibration, the standard deviations of the cosines and shifted average cosine values approach zero, which indicates that TRAILER achieves the engaging feature structure that is close to neural collapse. + +Unknown Number of Novel Classes. The previous experiments follow previous protocols and assume that the class number is known apriori [5, 15, 46], which is sometimes impractical in real-world applications and brings challenges for both TRAILER and the other baselines. A simple resolution is to first use estimation methods [32, 39] to derive the number of classes before deployment, for example, $k$ -estimation with Brent's algorithm from [50] or cross-validating using labeled probe set from [5, 21, 46]. On + +![](images/a3d4c1dd03d8f9f2a88416addad0a78fd8eb2a950719e383e1aea43362dd5291.jpg) +(a) Accuracy for known classes. +Figure 7. Accuracy comparisons for known and novel classes on CIFAR-10 with different novel class ratios ranging from 0.1 to 0.9. + +![](images/489b6e23a2ef86d0b8cea91eac19ad5b617b4c676c71bd08bb0860e23c9536dd.jpg) +(b) Accuracy for novel classes. + +CIFAR-100 the former gets exactly 100 while the latter derives 124 and was adopted by us for further exploration. We then utilize this estimation result to rerun the experiment by initializing the corresponding number of classifier vectors. As shown in Table 4, TRAILER is able to achieve competitive performance even when the number of classes is not known in advance. We further provide additional results with different estimated class numbers $K$ from 80 to 130 in Appendix B to demonstrate the robustness of TRAILER. + +Effect of Novel Class Ratio. Here we further evaluate the performance of TRAILER on the CIFAR-10 dataset with different ratios of novel classes from 0.1 to 0.9. As shown in Figure 7, we can see that TRAILER displays robust performance when the ratio of novel classes grows larger from 0.1 to 0.9 and consistently manifests superior performance than other baselines on both known and novel classes. + +# 5. Conclusion + +In this work, we propose a novel framework TRAILER for open-world SSL. We first take inspiration from the recently discovered neural collapse phenomenon and intend to attain its appealing feature arrangement with minimal within-class and maximum between-class covariance. To achieve this, we adopt a targeted classifier and align representations towards its pre-assigned optimal structure in a progressive manner. To mitigate the potential risk of misalignment, we tackle the sample-target assignment for unlabeled samples with a coarse-to-fine allocation mechanism. Comprehensive experiments show that TRAILER enhances the baselines by a notable margin on both generic and fine-grained benchmarks. We hope our work can inspire future research to leverage the appealing properties of the neural collapse phenomenon for tasks in open-world scenarios. + +# 6. Acknowledgments + +This work is supported by the Pioneer R&D Program of Zhejiang (No. 2024C01035). + +# References + +[1] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR. OpenReview.net, 2020. 4 +[2] Jessa Bekker and Jesse Davis. Learning from positive and unlabeled data: a survey. Mach. Learn., 109(4):719-760, 2020. 2, 5 +[3] David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution alignment and augmentation anchoring. CoRR, abs/1911.09785, 2019. 2 +[4] David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In NeurIPS, pages 5050-5060, 2019. 1, 2 +[5] Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning. In ICLR. OpenReview.net, 2022. 1, 2, 3, 5, 6, 7, 8 +[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 4 +[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9630-9640. IEEE, 2021. 7 +[8] Hui Chen, Fangqing Liu, Yin Wang, Liyue Zhao, and Hao Wu. A variational approach for learning from positive and unlabeled data. In NeurIPS 2020, 2020. 5 +[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. PMLR, 2020. 5, 6 +[10] Xuxi Chen, Wuyang Chen, Tianlong Chen, Ye Yuan, Chen Gong, Kewei Chen, and Zhangyang Wang. Self-pu: Self boosted and calibrated positive-unlabeled training. In ICML, pages 1510-1519. PMLR, 2020. 5 +[11] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NeurIPS, pages 2292-2300, 2013. 4 +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE Computer Society, 2009. 5 +[13] François Denis. PAC learning from positive statistical queries. In Algorithmic Learning Theory, 9th International Conference, ALT '98, Otzenhausen, Germany, October 8-10, 1998, Proceedings, pages 112-126. Springer, 1998. 2, 5 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR. OpenReview.net, 2021. 7 +[15] Enrico Fini, Enver Sangineto, Stephane Lathuilière, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for + +novel class discovery. In ICCV, pages 9264-9272. IEEE, 2021. 1, 2, 3, 4, 5, 6, 7, 8 +[16] Ronald A Fisher. The use of multiple measurements in taxonomic problems. Annals of eugenics, 7(2):179-188, 1936. 4 +[17] Tomer Galanti, András György, and Marcus Hutter. On the role of neural collapse in transfer learning. In ICLR. Open-Review.net, 2022. 2, 3 +[18] Florian Graf, Christoph D. Hofer, Marc Niethammer, and Roland Kwitt. Dissecting supervised constrastive learning. In ICML, pages 3821-3830. PMLR, 2021. 2 +[19] Lan-Zhe Guo, Zhenyu Zhang, Yuan Jiang, Yu-Feng Li, and Zhi-Hua Zhou. Safe deep semi-supervised learning for unseen-class unlabeled data. In ICML, pages 3897-3906. PMLR, 2020. 2, 5, 6 +[20] Lan-Zhe Guo, Yi-Ge Zhang, Zhi-Fan Wu, Jie-Jing Shao, and Yu-Feng Li. Robust semi-supervised learning when not all classes have labels. In NeurIPS, 2022. 1, 2, 5, 6 +[21] Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer clustering. In ICCV, pages 8400-8408. IEEE, 2019. 1, 2, 3, 5, 6, 8 +[22] Kai Han, Sylvestre-Alvise Rebuffi, Sébastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Automatically discovering and learning new visual categories with ranking statistics. In ICLR. OpenReview.net, 2020. 1, 2, 5, 6, 7 +[23] X. Y. Han, Vardan Papyan, and David L. Donoho. Neural collapse under MSE loss: Proximity to and dynamics on the central path. In ICLR. OpenReview.net, 2022. 2 +[24] Zongyan Han, Zhenyong Fu, Shuo Chen, and Jian Yang. Contrastive embedding for generalized zero-shot learning. In CVPR, pages 2371-2381. Computer Vision Foundation / IEEE, 2021. 2 +[25] Zongyan Han, Zhenyong Fu, Shuo Chen, and Jian Yang. Semantic contrastive embedding for generalized zero-shot learning. Int. J. Comput. Vis., 130(11):2606-2622, 2022. 2 +[26] Wenlong Ji, Yiping Lu, Yiliang Zhang, Zhun Deng, and Weijie J. Su. An unconstrained layer-peeled perspective on neural collapse. In ICLR. OpenReview.net, 2022. 2 +[27] Ryuichi Kiryo, Gang Niu, Marthinus Christoffel du Plessis, and Masashi Sugiyama. Positive-unlabeled learning with non-negative risk estimator. In NeurIPS, pages 1675-1685, 2017. 5 +[28] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV, pages 554–561. IEEE Computer Society, 2013. 5 +[29] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Master's thesis, Department of Computer Science, University of Toronto, 2009. 5 +[30] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In ICLR. OpenReview.net, 2017. 2 +[31] Dong-Hyun Lee. Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks. ICML 2013 Workshop : Challenges in Representation Learning (WREPL), 2013. 2 +[32] Collin Leiber, Lena G. M. Bauer, Benjamin Schelling, Christian Böhm, and Claudia Plant. Dip-based deep embedded + +clustering with k-estimation. In KDD, pages 903-913. ACM, 2021. 8 +[33] Jiaming Liu, Yangqiming Wang, Tongze Zhang, Yulu Fan, Qinli Yang, and Junming Shao. Open-world semi-supervised novel class discovery. In *IJCAI*, pages 4002–4010. ijcai.org, 2023. 1, 2, 3, 5, 6 +[34] Chuan Luo, Pu Zhao, Chen Chen, Bo Qiao, Chao Du, Hongyu Zhang, Wei Wu, Shaowei Cai, Bing He, Saravanakumar Rajmohan, and Qingwei Lin. PULNS: positive unlabeled learning with effective negative sample selector. In AAAI, pages 8784-8792. AAAI Press, 2021. 5 +[35] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. CoRR, abs/1306.5151, 2013. 5 +[36] Aleix M. Martínez and Avinash C. Kak. PCA versus LDA. IEEE Trans. Pattern Anal. Mach. Intell., 23(2):228-233, 2001. 2 +[37] Duc Anh Nguyen, Ron Levie, Julian Lienen, Eyke Hüllermeier, and Gitta Kutyniok. Memorization-dilation: Modeling neural collapse under noise. In ICLR. OpenReview.net, 2023. 2, 4 +[38] Vardan Papyan, X. Y. Han, and David L. Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 117(40):24652-24663, 2020. 2, 3 +[39] Dan Pelleg and Andrew W. Moore. X-means: Extending k-means with efficient estimation of the number of clusters. In ICML, pages 727-734. Morgan Kaufmann, 2000. 8 +[40] Zhongzheng Ren, Raymond A. Yeh, and Alexander G. Schwing. Not all unlabeled data are equal: Learning to weight data in semi-supervised learning. In NeurIPS, 2020. 2 +[41] Mamshad Nayeem Rizve, Navid Kardan, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Openldn: Learning to discover novel classes for open-world semi-supervised learning. In ECCV, pages 382-401. Springer, 2022. 1, 5, 6 +[42] Chuck Rosenberg, Martial Hebert, and Henry Schneiderman. Semi-supervised self-training of object detection models. In 7th IEEE Workshop on Applications of Computer Vision / IEEE Workshop on Motion and Video Computing (WACV/MOTION 2005), 5-7 January 2005, Breckenridge, CO, USA, pages 29-36. IEEE Computer Society, 2005. 2 +[43] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In NeurIPS, pages 1163-1171, 2016. 2 +[44] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 2, 5, 6 +[45] Xin Sun, Zhenning Yang, Chi Zhang, Keck Voon Ling, and Guohao Peng. Conditional gaussian distribution learning for open set recognition. In CVPR, pages 13477-13486. Computer Vision Foundation / IEEE, 2020. 2, 5, 6 +[46] Yiyou Sun and Yixuan Li. Opencon: Open-world contrastive learning. Trans. Mach. Learn. Res., 2023, 2023. 1, 2, 3, 5, 6, 7, 8 + +[47] Kiat Chuan Tan, Yulong Liu, Barbara Ambrose, Melissa Tulig, and Serge J. Belongie. The herbarium challenge 2019 dataset. CoRR, abs/1906.05372, 2019. 5 +[48] Christos Thrampoulidis, Ganesh Ramachandra Kini, Vala Vakilian, and Tina Behnia. Imbalance trouble: Revisiting neural-collapse geometry. In NeurIPS, 2022. 2 +[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7 +[50] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. In CVPR, pages 7482-7491. IEEE, 2022. 1, 2, 3, 5, 6, 7, 8 +[51] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In ICLR. OpenReview.net, 2022. 5, 6 +[52] Jiaheng Wei, Zhaowei Zhu, Hao Cheng, Tongliang Liu, Gang Niu, and Yang Liu. Learning with noisy labels revisited: A study using real-world human annotations. In ICLR. OpenReview.net, 2022. 2 +[53] P. Welinder, S. Branson, T. Mita, C. Wah, F. Schroff, S. Belongie, and P. Perona. Caltech-UCSD Birds 200. Technical Report CNS-TR-2010-001, California Institute of Technology, 2010. 5 +[54] Liang Xie, Yibo Yang, Deng Cai, and Xiaofei He. Neural collapse inspired attraction-repulsion-balanced loss for imbalanced learning. Neurocomputing, 527:60-70, 2023. 2 +[55] Qizhe Xie, Zihang Dai, Eduard H. Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. In NeurIPS, 2020. 1, 2 +[56] Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In ICML, pages 11525-11536. PMLR, 2021. 2 +[57] Yibo Yang, Shixiang Chen, Xiangtai Li, Liang Xie, Zhouchen Lin, and Dacheng Tao. Inducing neural collapse in imbalanced learning: Do we really need a learnable classifier at the end of deep neural network? In NeurIPS, 2022. 2, 4 +[58] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip H. S. Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In ICLR. OpenReview.net, 2023. 2 +[59] Bo Ye, Kai Gan, Tong Wei, and Min-Ling Zhang. Bridging the gap: Learning pace synchronization for open-world semi-supervised learning. CoRR, abs/2309.11930, 2023. 2 +[60] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In NeurIPS 2021, pages 18408-18419, 2021. 1 +[61] Bingchen Zhao, Xin Wen, and Kai Han. Learning semi-supervised gaussian mixture models for generalized category discovery. In ICCV, pages 16623-16633, 2023. 5, 6, 7 +[62] Yunrui Zhao, Qianqian Xu, Yangbangyan Jiang, Peisong Wen, and Qingming Huang. Dist-pu: Positive-unlabeled learning from a label distribution perspective. In CVPR, pages 14441-14450. IEEE, 2022. 2, 5 + +[63] Zhisheng Zhong, Jiequan Cui, Yibo Yang, Xiaoyang Wu, Xiaojuan Qi, Xiangyu Zhang, and Jiaya Jia. Understanding imbalanced semantic segmentation through neural collapse. In CVPR, pages 19550-19559. IEEE, 2023. 2 +[64] Jinxin Zhou, Xiao Li, Tianyu Ding, Chong You, Qing Qu, and Zhihui Zhu. On the optimization landscape of neural collapse under MSE loss: Global optimality with unconstrained features. In ICML, pages 27179-27202. PMLR, 2022. 2 +[65] Zhihui Zhu, Tianyu Ding, Jinxin Zhou, Xiao Li, Chong You, Jeremias Sulam, and Qing Qu. A geometric analysis of neural collapse with unconstrained features. In NeurIPS, pages 29820-29834, 2021. 2 +[66] Zhaowei Zhu, Tianyi Luo, and Yang Liu. The rich get richer: Disparate impact of semi-supervised learning. In ICLR. OpenReview.net, 2022. 2 \ No newline at end of file diff --git a/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/images.zip b/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8292f67da3e4145d560f8f2723d9bc965ccc5da9 --- /dev/null +++ b/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4911656314c6340d7e87ad1c88a7f251ec8ba8f26ddb745bee13efb56ba9e33 +size 560983 diff --git a/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/layout.json b/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c8d6175ad2cd57747cd84990c5f9b644ad24919e --- /dev/null +++ b/2024/Targeted Representation Alignment for Open-World Semi-Supervised Learning/layout.json @@ -0,0 +1,10523 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 103, + 542, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 103, + 542, + 122 + ], + "spans": [ + { + "bbox": [ + 52, + 103, + 542, + 122 + ], + "type": "text", + "content": "Targeted Representation Alignment for Open-World Semi-Supervised Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 142, + 529, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 529, + 158 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 529, + 158 + ], + "type": "text", + "content": "Ruixuan Xiao" + }, + { + "bbox": [ + 77, + 142, + 529, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 529, + 158 + ], + "type": "text", + "content": ", Lei Feng" + }, + { + "bbox": [ + 77, + 142, + 529, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 142, + 529, + 158 + ], + "type": "text", + "content": ", Kai Tang" + }, + { + "bbox": [ + 77, + 142, + 529, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 529, + 158 + ], + "type": "text", + "content": ", Junbo Zhao" + }, + { + "bbox": [ + 77, + 142, + 529, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 529, + 158 + ], + "type": "text", + "content": ", Yixuan Li" + }, + { + "bbox": [ + 77, + 142, + 529, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 142, + 529, + 158 + ], + "type": "text", + "content": ", Gang Chen" + }, + { + "bbox": [ + 77, + 142, + 529, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 529, + 158 + ], + "type": "text", + "content": ", Haobo Wang" + }, + { + "bbox": [ + 77, + 142, + 529, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 233, + 158, + 369, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 158, + 369, + 171 + ], + "spans": [ + { + "bbox": [ + 233, + 158, + 369, + 171 + ], + "type": "text", + "content": "1Zhejiang University, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 156, + 171, + 447, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 171, + 447, + 186 + ], + "spans": [ + { + "bbox": [ + 156, + 171, + 447, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 156, + 171, + 447, + 186 + ], + "type": "text", + "content": "Singapore University of Technology and Design, Singapore" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 202, + 186, + 400, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 186, + 400, + 199 + ], + "spans": [ + { + "bbox": [ + 202, + 186, + 400, + 199 + ], + "type": "text", + "content": "3University of Wisconsin-Madison, USA" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 201, + 545, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 201, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 201, + 545, + 213 + ], + "type": "text", + "content": "{xiaoruixuan,tk0819,j.zhao,cg,wanghaobo}@zju.edu.cn,lfengqaq@gmail.com,sharonli@cs.wisc.edu" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 266, + 290, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 290, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 290, + 578 + ], + "type": "text", + "content": "Open-world Semi-Supervised Learning aims to classify unlabeled samples utilizing information from labeled data, while unlabeled samples are not only from the labeled known categories but also from novel categories previously unseen. Despite the promise, current approaches solely rely on hazardous similarity-based clustering algorithms and give unlabeled samples free rein to spontaneously group into distinct novel class clusters. Nevertheless, due to the absence of novel class supervision, these methods typically suffer from the representation collapse dilemma—features of different novel categories can get closely intertwined and indistinguishable, even collapsing into the same cluster and leading to degraded performance. To alleviate this, we propose a novel framework TRAILER which targets to attain an optimal feature arrangement revealed by the recently uncovered neural collapse phenomenon. To fulfill this, we adopt targeted prototypes that are pre-assigned uniformly with maximum separation and then progressively align the representations to them. To further tackle the potential downsides of such stringent alignment, we encapsulate a sample-target allocation mechanism with coarse-to-fine refinery that is able to infer label assignments with high quality. Extensive experiments demonstrate that TRAILER outperforms current state-of-the-art methods on generic and fine-grained benchmarks. The code is available at https://github.com/Justherozen/TRAILER." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 621, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 621, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 621, + 287, + 693 + ], + "type": "text", + "content": "The resounding achievements of deep learning heavily rely on large-scale and accurately-annotated training data, which is labor-intensive and time-consuming. To alleviate this, semi-supervised learning (SSL) has emerged as a prominent alternative [4, 55, 60], which exploits unlabeled data for boosted performance. Despite the success, current SSL" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 319, + 241, + 533, + 356 + ], + "blocks": [ + { + "bbox": [ + 319, + 241, + 533, + 356 + ], + "lines": [ + { + "bbox": [ + 319, + 241, + 533, + 356 + ], + "spans": [ + { + "bbox": [ + 319, + 241, + 533, + 356 + ], + "type": "image", + "image_path": "d2b29116574fc6e9c93ae3ff3977346e698f8b29a8a029afeaadff6616eb96bc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 365, + 546, + 420 + ], + "lines": [ + { + "bbox": [ + 305, + 365, + 546, + 420 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 546, + 420 + ], + "type": "text", + "content": "Figure 1. Left: Scenario of open-world SSL where unlabeled samples are from known and novel classes. Right: Visualization without and with our targeted calibration mechanism. Without calibration, representations tend to get jumbled and inseparable, as in the left where dog and cat categories converge into the same cluster." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 447, + 546, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 546, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 546, + 614 + ], + "type": "text", + "content": "methods are primarily based on the closed-world assumption that labeled and unlabeled data share the same predefined label space. Nevertheless, this assumption rarely holds in real-world applications, where the model may encounter novel classes previously unseen in unlabeled data. To relax this assumption, novel category discovery (NCD) was first proposed [15, 21, 22] which assumes all unlabeled data strictly belong to disjoint novel classes. Recently, a more practical setting, open-world semi-supervised learning (open-world SSL), also referred to as generalized category discovery (GCD), has received huge attention [5, 46, 50] by generalizing standard SSL and NCD. It assumes that unlabeled data contain both known classes from the labeled data and similar yet distinct novel classes." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "The goal of open-world SSL is to classify the already-known classes from the labeled data, as well as to discover new classes in unlabeled data without annotations and assign instances to them. Nascent efforts have been made to address this practical yet under-investigated task [20, 33, 41, 50]. The essence of these recent works is to rashly rely on the similarity-based [5, 41] or contrastive-based [46] clustering algorithm in feature space to spontaneously" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23072" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "group different samples into separate clusters. Some methods further integrate self-training techniques with generated pseudo-labels for boosted performance [20, 33]. However, due to the missing supervision and semantic shift for novel categories, the self-organized clustering step can get overfitted and biased towards the known classes, leading to a representation collapse dilemma—the feature distance between one novel category and another category may get too close and finally become inseparable, even converging within a single cluster and thus jeopardizing generalization performance, as illustrated in the left of Figure 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 371 + ], + "type": "text", + "content": "On the other hand, a recent study has unveiled a feature-classifier alignment phenomenon termed neural collapse (NC) [38]. It indicates that under ideal conditions, the last-layer features of the same class tend to collapse into their within-class mean. Meanwhile, these within-class means of all classes will be aligned with classifiers and formed as a simplex equiangular tight frame (ETF) at the end of training, as elaborated in Section 3.1. Intuitively, such an elegant NC phenomenon affords an optimal feature structure with minimized within-class variance and maximized between-class variance [36], coinciding with what we fervently desire. Nevertheless, in open-world SSL, the absence of novel class annotation can hinder the representation learning and thus disrupt the natural induction of NC." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 371, + 289, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 289, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 289, + 660 + ], + "type": "text", + "content": "In an effort to salvage this and accomplish the appealing feature arrangement of NC, we propose a novel framework, Targeted Representation AIIgnment for open-world semisupervised LEaRning (dubbed TRAILER). Key to our method, TRAILER leverages a targeted classifier to progressively align embeddings to the pre-assigned ETF structure, eventually fulfilling the optimal state of uniformly and maximally separated embeddings. Despite the promise, this calibration procedure is still susceptible to false alignment stemming from unreliable pseudo-labels. To mitigate this, we integrate a hierarchical sample-target allocation strategy that comprises two sequential steps: (i)-rough sample assignment based on the optimal transport mechanism; (ii)-label refinery by casting the known-novel separation problem to a specific weakly-supervised paradigm termed PU learning [2, 13, 62]. Such an allocation mechanism can provide more precise pseudo-labels and thus facilitate representation calibration. Through our visualized results in Figure 1 and Section 4.3, TRAILER does indeed induce discriminative and non-skewed embeddings. Empirically, TRAILER surpasses state-of-the-art approaches on different evaluation benchmarks, e.g., improves the best baseline by " + }, + { + "bbox": [ + 46, + 371, + 289, + 660 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 46, + 371, + 289, + 660 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 371, + 289, + 660 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 46, + 371, + 289, + 660 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 371, + 289, + 660 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 46, + 371, + 289, + 660 + ], + "type": "text", + "content": " overall accuracy on CIFAR-10, CIFAR-100 and ImageNet-100 datasets, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Closed-World Semi-Supervised Learning has been widely explored [4, 44, 55, 66] to learn from limited la" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "beled and massive unlabeled data. Pseudo labeling [31, 42] and consistency regularization [30, 44, 55] are two prominent techniques. The former is based on self-training that utilizes model predictions as optimization targets [40, 56], while the latter enforces model to produce consistent outputs for different views of the same image [3, 4, 43, 55]. However, these methods are primarily based on the closed-world assumption that labeled and unlabeled data share the same pre-defined label space, which limits their real-world application where unlabeled samples from previously unseen classes may emerge. Open-set SSL [19, 24, 25, 45] relaxes this by simply rejecting all unseen class instances, but still fails to classify them into distinct novel categories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 243, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 547, + 519 + ], + "type": "text", + "content": "Open-World Semi-Supervised Learning further accommodates real-world scenarios [5, 20, 46]. It is also known as generalized category discovery (GCD) [50] and is a natural extension of novel category discovery (NCD). NCD is first introduced in [21] and improved in later works [15, 22] to discover new categories by leveraging the knowledge from labeled known categories. However, NCD assumes that all unlabeled samples belong to novel classes and thus fails to recognize known class samples. Most recently, open-world SSL generalizes NCD to identify unlabeled samples from both known and novel classes [5]. To tackle this, several incipient solutions have been proposed [20, 59], which typically involve similarity-based representation learning and then involve sample clustering to generate pseudo-labels for self-training. ORCA [5] is the pioneering work that integrates an uncertainty-based margin loss to control the intraclass variance of different classes. OpenCon [46] further derives a semi-supervised constrastive mechanism to produce compact feature space. Despite the promise, these methods solely rely on spontaneous clustering based on sample similarity, overlooking potential risks that representations of different classes possibly converge into the same cluster, thus leading to the representation collapse dilemma." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 534, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 548, + 715 + ], + "type": "text", + "content": "Neural Collapse describes an elegant alignment phenomenon [38], which states that the last-layer features will collapse into their within-class centers and these centers together with classifiers will eventually form a simplex equiangular tight frame (ETF). Stemming from the appealing symmetry, a plethora of studies have been triggered to explain this phenomenon theoretically [26, 48, 58, 65]. NC is proved to be the global optimality under ideal conditions with the CE [18, 26] and the MSE [23, 64] loss functions. The NC phenomenon has also been investigated in some specific scenarios like imbalanced learning [54, 57, 63] noisy label learning [37, 52], transfer learning [17] and incremental learning [58]. In our work, we investigate the opportunities of inducing such optimal feature arrangement of NC through targeted calibration for open-world SSL." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23073" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 71, + 202, + 240 + ], + "blocks": [ + { + "bbox": [ + 71, + 71, + 202, + 240 + ], + "lines": [ + { + "bbox": [ + 71, + 71, + 202, + 240 + ], + "spans": [ + { + "bbox": [ + 71, + 71, + 202, + 240 + ], + "type": "image", + "image_path": "206d0955d948800be5240746976bb41019e819af1a2cd14fbb43518183211c80.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 250, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 250, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 547, + 285 + ], + "type": "text", + "content": "Figure 2. Overview of TRAILER. For representation calibration, we leverage the targeted classifier " + }, + { + "bbox": [ + 46, + 250, + 547, + 285 + ], + "type": "inline_equation", + "content": "\\phi^{etf}" + }, + { + "bbox": [ + 46, + 250, + 547, + 285 + ], + "type": "text", + "content": " to align the feature towards the optimal arrangement of ETF progressively while maintaining the within-class compactness with prototypical clustering. For sample-target allocation, the unlabeled samples are provided with reliable pseudo-labels with rough assignment and subsequent label refinery." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 203, + 70, + 522, + 241 + ], + "blocks": [ + { + "bbox": [ + 203, + 70, + 522, + 241 + ], + "lines": [ + { + "bbox": [ + 203, + 70, + 522, + 241 + ], + "spans": [ + { + "bbox": [ + 203, + 70, + 522, + 241 + ], + "type": "image", + "image_path": "08e30f5355d1bb8d4a427b7ffb323661198a6beb378faa3dffb95265494d562f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 303, + 104, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 104, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 104, + 316 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "text", + "content": "Problem Definition. In open-world SSL, we assume that the training dataset " + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\mathcal{D}_l \\cup \\mathcal{D}_u" + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "text", + "content": " comprises two subsets: a labeled set " + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_l = \\{x_i, y_i\\}_{i=1}^m" + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "text", + "content": " where the input image " + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "inline_equation", + "content": "x_i \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "text", + "content": " and the ground-truth label " + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "text", + "content": " come from a set of known classes " + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{known}" + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "text", + "content": ", and an unlabeled set " + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u = \\{x_i\\}_{i=1}^n" + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "text", + "content": " where the potential ground-truth label is from a set of classes " + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{all}" + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "text", + "content": ". In closed-world SSL, the unlabeled samples strictly come from known classes, i.e., " + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{all} = \\mathcal{C}_{known}" + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "text", + "content": ". As for open-world SSL which is more practical, the unlabeled samples can come from either known classes " + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{known}" + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "text", + "content": " or novel classes, i.e., " + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{known} \\subset \\mathcal{C}_{all}" + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "text", + "content": ". We consider " + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{novel} = \\mathcal{C}_{all} \\backslash \\mathcal{C}_{known}" + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "text", + "content": " as the set of novel classes. Following previous setting [5, 15, 46], we assume the number of novel classes " + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}_{novel}|" + }, + { + "bbox": [ + 46, + 323, + 290, + 504 + ], + "type": "text", + "content": " is known apriori, which can also be estimated using the off-the-shelf methods [21, 50] as discussed in Section 4.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "content": "Learning Goal. For classification problem with " + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "inline_equation", + "content": "K = |\\mathcal{C}_{all}|" + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "content": " classes, open-world SSL aims to obtain a model " + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "inline_equation", + "content": "f_{\\theta,\\phi} : \\mathbb{R}^d \\mapsto \\mathbb{R}^K" + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "content": " that recognizes samples from both known and novel classes based on " + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_l" + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u" + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "inline_equation", + "content": "f_{\\theta,\\phi} = h_\\phi \\circ g_\\theta" + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "content": " can be decomposed into a representation backbone " + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "inline_equation", + "content": "g_\\theta" + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "content": " and a classification head " + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "inline_equation", + "content": "h_\\phi" + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "content": " with weights " + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "inline_equation", + "content": "\\phi = [\\phi_1, \\dots, \\phi_k] \\in \\mathbb{R}^{P \\times K}" + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "content": ", for which we adopt a prototypical one following [15] with logit " + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "inline_equation", + "content": "h(z) = \\mathrm{Norm}(z)^\\top \\mathrm{Norm}(\\phi)" + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "content": ". Formally, given the weakly and strongly augmented view " + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "inline_equation", + "content": "x_i'" + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "content": " of the same image in batch " + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "content": ", we define " + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "inline_equation", + "content": "z_i, z_i' \\in \\mathbb{R}^P" + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "content": " as the last-layer " + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "content": "-dimensional features and " + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "inline_equation", + "content": "p_i = \\mathrm{softmax}(h(z_i))" + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "content": " as output probability. Normalization for features is denoted as " + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{Norm}(z) = z / ||z||_2" + }, + { + "bbox": [ + 47, + 518, + 289, + 676 + ], + "type": "text", + "content": " and omitted in the subsequent." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Warm-up Phase. Similar to previous works [5, 33], we also include a commonly adopted warm-up strategy at" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 304, + 547, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 547, + 434 + ], + "type": "text", + "content": "the start of training. The overall objectives to cluster known classes and discover potential novel samples consist of three losses: 1) the cross-entropy loss (CE) " + }, + { + "bbox": [ + 304, + 304, + 547, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 304, + 304, + 547, + 434 + ], + "type": "text", + "content": " on labeled samples " + }, + { + "bbox": [ + 304, + 304, + 547, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_l" + }, + { + "bbox": [ + 304, + 304, + 547, + 434 + ], + "type": "text", + "content": " which is elaborated later in Section 3.1; 2) the self-supervised contrastive loss " + }, + { + "bbox": [ + 304, + 304, + 547, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con} = -\\log \\frac{\\exp(z_i^\\top z_i' / \\tau)}{\\sum_{j=1}^n \\exp(z_i^\\top z_j' / \\tau)}" + }, + { + "bbox": [ + 304, + 304, + 547, + 434 + ], + "type": "text", + "content": " on all samples; and 3) the entropy regularization " + }, + { + "bbox": [ + 304, + 304, + 547, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ent} = \\sum_{k=1}^{K} \\overline{p}^k \\log (\\overline{p}^k)" + }, + { + "bbox": [ + 304, + 304, + 547, + 434 + ], + "type": "text", + "content": " for preventing the trivial solution of assigning all instances to the same class, where " + }, + { + "bbox": [ + 304, + 304, + 547, + 434 + ], + "type": "inline_equation", + "content": "\\overline{p} = \\frac{1}{|\\mathcal{B}|} \\sum_{x_i \\in \\mathcal{B}} p_i" + }, + { + "bbox": [ + 304, + 304, + 547, + 434 + ], + "type": "text", + "content": " is the average probability of the batch. The overall losses for warm-up are formulated," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 361, + 441, + 545, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 441, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 361, + 441, + 545, + 454 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {w a r m}} = \\mathcal {L} _ {c e} + \\mathcal {L} _ {\\text {c o n}} + \\alpha \\mathcal {L} _ {\\text {e n t}} \\tag {1}", + "image_path": "919be89a5de9d5d339529c66ae4d66e2b6f91e9fd3bdb426694ab4ab778726c8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 461, + 547, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 547, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 547, + 546 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 461, + 547, + 546 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 461, + 547, + 546 + ], + "type": "text", + "content": " is a weighting parameter. After warm-up, we also would like to group unlabeled samples into distinct clusters and assign them pseudo-labels for self-training. As mentioned earlier, the crucial challenge is how to abstain from the representation collapse dilemma and obtain discriminative features without novel class annotations. To fulfill this, we first draw inspiration from neural collapse phenomenon." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 551, + 545, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 551, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 305, + 551, + 545, + 563 + ], + "type": "text", + "content": "3.1. Neural Collapse for Representation Calibration" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 569, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 547, + 629 + ], + "type": "text", + "content": "Recent works have disclosed the Neural Collapse phenomenon [17, 38] which corresponds to an optimal feature-classifier alignment towards a simplex equiangular tight frame (ETF). This offers a uniformly and equally separated feature arrangement that is desirable for open-world SSL." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 635, + 547, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 547, + 684 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 547, + 684 + ], + "type": "text", + "content": "Definition 1 (Simplex ETF) A Simplex equiangular tight frame (ETF) refers to a collection of " + }, + { + "bbox": [ + 304, + 635, + 547, + 684 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 635, + 547, + 684 + ], + "type": "text", + "content": " equal-length and maximally-equiangular " + }, + { + "bbox": [ + 304, + 635, + 547, + 684 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 635, + 547, + 684 + ], + "type": "text", + "content": "-dimensional embedding vectors " + }, + { + "bbox": [ + 304, + 635, + 547, + 684 + ], + "type": "inline_equation", + "content": "\\mathbf{E} = [e_1, \\dots, e_K] \\in \\mathbb{R}^{P \\times K}" + }, + { + "bbox": [ + 304, + 635, + 547, + 684 + ], + "type": "text", + "content": " which satisfies:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 350, + 689, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 689, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 350, + 689, + 545, + 717 + ], + "type": "interline_equation", + "content": "\\mathbf {E} = \\sqrt {\\frac {K}{K - 1}} \\mathbf {U} \\left(\\mathbf {I} _ {K} - \\frac {1}{K} \\mathbf {1} _ {K} \\mathbf {1} _ {K} ^ {\\top}\\right) \\tag {2}", + "image_path": "5f4a47f749c61c0e3e3690824ee25f78988f12b31a20942a6494360361741287.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23074" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_K" + }, + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "text", + "content": " is the identity matrix, " + }, + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_K" + }, + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "text", + "content": " is an all-ones vector, and " + }, + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{U} \\in \\mathbb{R}^{P \\times K}" + }, + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "inline_equation", + "content": "P \\geq K" + }, + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "text", + "content": ") allows a rotation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 102, + 287, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 102, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 287, + 129 + ], + "type": "text", + "content": "All vectors in a simplex ETF " + }, + { + "bbox": [ + 47, + 102, + 287, + 129 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 47, + 102, + 287, + 129 + ], + "type": "text", + "content": " have an equal " + }, + { + "bbox": [ + 47, + 102, + 287, + 129 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 47, + 102, + 287, + 129 + ], + "type": "text", + "content": " norm and the same pair-wise maximal equiangular angle " + }, + { + "bbox": [ + 47, + 102, + 287, + 129 + ], + "type": "inline_equation", + "content": "-\\frac{1}{K - 1}" + }, + { + "bbox": [ + 47, + 102, + 287, + 129 + ], + "type": "text", + "content": " [57]," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 140, + 287, + 164 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 140, + 287, + 164 + ], + "spans": [ + { + "bbox": [ + 58, + 140, + 287, + 164 + ], + "type": "interline_equation", + "content": "e _ {k _ {1}} ^ {\\top} e _ {k _ {2}} = \\frac {K}{K - 1} \\delta_ {k _ {1}, k _ {2}} - \\frac {1}{K - 1}, \\forall k _ {1}, k _ {2} \\in [ 1, K ] \\quad (3)", + "image_path": "b8503974034c8645539e9dff2e31f11bda5d79858680fb8a20826a4697097724.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 168, + 287, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 168, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 287, + 276 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 168, + 287, + 276 + ], + "type": "inline_equation", + "content": "\\delta_{k_1,k_2} = 1" + }, + { + "bbox": [ + 47, + 168, + 287, + 276 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 47, + 168, + 287, + 276 + ], + "type": "inline_equation", + "content": "k_{1} = k_{2}" + }, + { + "bbox": [ + 47, + 168, + 287, + 276 + ], + "type": "text", + "content": " and 0 otherwise. The NC phenomenon can be described by the following properties under ideal conditions: (1) The last-layer features converge to their within-class means, and the within-class means of all classes converge to different vertices of a simplex ETF. (2) The normalized classifier weights converge to the same simplex ETF. Notably, this process is implicit and natural since it is induced with common optimizers such as SGD, rather than enforced by an explicit objective." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 276, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 287, + 384 + ], + "type": "text", + "content": "As discussed above, NC has provided an optimal ETF structure for feature arrangement with minimal within-class covariance and maximum between-class covariance (i.e., maximized Fisher discriminant ratio [16]), which coincides with our goal of deriving discriminative and non-skewed embeddings. However, NC is delicate and can fail to be naturally induced under imperfect conditions [37, 57]. For open-world SSL, the missing novel class annotations can also lead to biased features and break the NC optimality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 384, + 287, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 492 + ], + "type": "text", + "content": "To remedy this, we resort to effectuating an explicit NC and fulfilling the appealing feature arrangement with representation calibration, which is achieved by leveraging a targeted prototypical classifier " + }, + { + "bbox": [ + 46, + 384, + 287, + 492 + ], + "type": "inline_equation", + "content": "h_{\\text{etf}}" + }, + { + "bbox": [ + 46, + 384, + 287, + 492 + ], + "type": "text", + "content": ". In specific, " + }, + { + "bbox": [ + 46, + 384, + 287, + 492 + ], + "type": "inline_equation", + "content": "h_{\\text{etf}}" + }, + { + "bbox": [ + 46, + 384, + 287, + 492 + ], + "type": "text", + "content": " is preassigned with weights " + }, + { + "bbox": [ + 46, + 384, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\phi^{\\text{etf}} \\in \\mathbb{R}^{P \\times K}" + }, + { + "bbox": [ + 46, + 384, + 287, + 492 + ], + "type": "text", + "content": ", which is a random simplex ETF with optimal structure as Definition 1 and is fixed rather than learnable during training. By assuming a known label " + }, + { + "bbox": [ + 46, + 384, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 46, + 384, + 287, + 492 + ], + "type": "text", + "content": " for each sample " + }, + { + "bbox": [ + 46, + 384, + 287, + 492 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 46, + 384, + 287, + 492 + ], + "type": "text", + "content": ", the standard CE loss " + }, + { + "bbox": [ + 46, + 384, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 46, + 384, + 287, + 492 + ], + "type": "text", + "content": " is then replaced as " + }, + { + "bbox": [ + 46, + 384, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{etf}}" + }, + { + "bbox": [ + 46, + 384, + 287, + 492 + ], + "type": "text", + "content": ", which is formulated," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 497, + 287, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 497, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 63, + 497, + 287, + 538 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e t f} \\left(x _ {i}, \\hat {y} _ {i}\\right) = - \\log \\frac {\\exp \\left(z _ {i} ^ {\\top} \\cdot \\phi_ {\\hat {y} _ {i}} ^ {e t f} / \\tau\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(z _ {i} ^ {\\top} \\cdot \\phi_ {k} ^ {e t f} / \\tau\\right)} \\tag {4}", + "image_path": "7f6cf3684cef9a2e8b14a2eab60447f03bc7495d727a913015d1884a2b291806.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 543, + 286, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 543, + 286, + 579 + ], + "spans": [ + { + "bbox": [ + 47, + 543, + 286, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 543, + 286, + 579 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 543, + 286, + 579 + ], + "type": "text", + "content": " is the temperature parameter. Through " + }, + { + "bbox": [ + 47, + 543, + 286, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{etf}}" + }, + { + "bbox": [ + 47, + 543, + 286, + 579 + ], + "type": "text", + "content": ", the feature " + }, + { + "bbox": [ + 47, + 543, + 286, + 579 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 47, + 543, + 286, + 579 + ], + "type": "text", + "content": " is drawn closer to its corresponding target " + }, + { + "bbox": [ + 47, + 543, + 286, + 579 + ], + "type": "inline_equation", + "content": "\\phi_{\\hat{y}_i}^{\\text{etf}}" + }, + { + "bbox": [ + 47, + 543, + 286, + 579 + ], + "type": "text", + "content": " and pulled away from other class vectors of ETF structure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 580, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 287, + 628 + ], + "type": "text", + "content": "For the labeled sample " + }, + { + "bbox": [ + 47, + 580, + 287, + 628 + ], + "type": "inline_equation", + "content": "(x_{i},y_{i})" + }, + { + "bbox": [ + 47, + 580, + 287, + 628 + ], + "type": "text", + "content": ", we directly set " + }, + { + "bbox": [ + 47, + 580, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\hat{y}_i = y_i" + }, + { + "bbox": [ + 47, + 580, + 287, + 628 + ], + "type": "text", + "content": ". As for unlabeled data without ground-truth labels, we design a sample-target allocation strategy to generate pseudolabels " + }, + { + "bbox": [ + 47, + 580, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 47, + 580, + 287, + 628 + ], + "type": "text", + "content": ", whose details are shown in Section 3.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "text", + "content": "Progressive Alignment. Notably, if we directly optimize the loss " + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{etf}}" + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "text", + "content": " after the warm-up phase, the provisionally clustered samples would be forcibly dragged to the corresponding targeted vector " + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\phi_{\\hat{y}}^{\\text{etf}}" + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "text", + "content": ". As a result, the already-learned clustering compactness can get compromised and the within-cluster feature distribution can get jumbled." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "To address this, we intend to align the representations in a progressive manner to preserve the compactness within each cluster during calibration. Formally, given sample " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " with pseudo-label " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", we also maintain the authentic prototypes " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mu_k = \\mathrm{Norm}(\\mathrm{Avg}_{\\hat{y}_i = k}(z_i))" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " for each class " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "1 \\leq k \\leq K" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ". To preserve the compactness of cluster structure, a prototypical clustering loss is also integrated," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 329, + 164, + 545, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 164, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 329, + 164, + 545, + 196 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p r o t o}} \\left(z _ {i}, \\hat {y} _ {i}\\right) = - \\log \\frac {\\exp \\left(z _ {i} ^ {\\top} \\cdot \\mu_ {\\hat {y} _ {i}} / \\tau\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(z _ {i} ^ {\\top} \\cdot \\mu_ {k} / \\tau\\right)} \\tag {5}", + "image_path": "e633fb151357887f1424e1374fc27ff7ce47b24d034dedd57d3fc9f6ae6f3828.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 202, + 504, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 202, + 504, + 214 + ], + "spans": [ + { + "bbox": [ + 306, + 202, + 504, + 214 + ], + "type": "text", + "content": "The overall classification loss is then formulated," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 362, + 226, + 545, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 226, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 362, + 226, + 545, + 239 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} = \\lambda \\mathcal {L} _ {p r o t o} + (1 - \\lambda) \\mathcal {L} _ {e t f} \\tag {6}", + "image_path": "4ede97c10f84b4b54d0a112be672c33596133438a860315b0e116373086f2dae.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 250, + 545, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 250, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 304, + 250, + 545, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 250, + 545, + 310 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 250, + 545, + 310 + ], + "type": "text", + "content": " is a calibration strength parameter that is gradually ramped down from 1 to 0. With such a progressive manner, the embeddings are cautiously and smoothly drawn closer to the targeted optimal arrangement of ETF while the within cluster compactness is preserved." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 317, + 510, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 317, + 510, + 330 + ], + "spans": [ + { + "bbox": [ + 306, + 317, + 510, + 330 + ], + "type": "text", + "content": "3.2. Hierarchical Sample-Target Allocation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 335, + 545, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 444 + ], + "type": "text", + "content": "In previous steps, we involve targeted classifiers to regulate representation learning. Nevertheless, our stringent calibration still holds unexpected risks. Lacking supervision, the calibration process for unlabeled samples largely relies on the quality of generated pseudo-labels. When the pseudolabels are unreliable, the feature alignment procedure can be misguided towards the wrong cluster. To overcome this, we introduce our hierarchical sample-target allocation mechanism with course-to-fine assignment in what follows." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 459, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 553 + ], + "type": "text", + "content": "Rough Assignment through Optimal Transport. For pseudo-label generation, one may adopt model prediction " + }, + { + "bbox": [ + 304, + 459, + 545, + 553 + ], + "type": "inline_equation", + "content": "\\arg \\max_{1\\leq k\\leq K}h^{k}(z_{i})" + }, + { + "bbox": [ + 304, + 459, + 545, + 553 + ], + "type": "text", + "content": " directly. However, this may lead to degenerate solutions where data points of different categories map into a single category [1], thus resulting in representation collapse. To mitigate this, we adopt a widely-used equipartition constraint [6, 15] for rough label assignment, which is attained by solving an optimal transport problem." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 553, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 545, + 611 + ], + "type": "text", + "content": "Formally, given a batch " + }, + { + "bbox": [ + 304, + 553, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 553, + 545, + 611 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 553, + 545, + 611 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 553, + 545, + 611 + ], + "type": "text", + "content": " unlabeled samples with the logit matrix " + }, + { + "bbox": [ + 304, + 553, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = [h(z_1),\\dots,h(z_b)]" + }, + { + "bbox": [ + 304, + 553, + 545, + 611 + ], + "type": "text", + "content": " whose columns are output logits by " + }, + { + "bbox": [ + 304, + 553, + 545, + 611 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{etf}}" + }, + { + "bbox": [ + 304, + 553, + 545, + 611 + ], + "type": "text", + "content": " and the label-assignment matrix " + }, + { + "bbox": [ + 304, + 553, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = [q_1,\\dots,q_b]\\in [0,1]^{K\\times b}" + }, + { + "bbox": [ + 304, + 553, + 545, + 611 + ], + "type": "text", + "content": " whose columns are the to-beassigned soft labels " + }, + { + "bbox": [ + 304, + 553, + 545, + 611 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 304, + 553, + 545, + 611 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 553, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 304, + 553, + 545, + 611 + ], + "type": "text", + "content": " is estimated by," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 617, + 545, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 617, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 320, + 617, + 545, + 659 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {Q} = \\max _ {\\mathbf {Q} \\in \\Gamma} \\operatorname {T r} (\\mathbf {Q} ^ {\\top} \\mathbf {P}) + \\epsilon \\mathrm {H} (\\mathbf {Q}) \\\\ \\text {s . t .} \\Gamma = \\left\\{\\mathbf {Q} \\in \\mathbb {R} _ {+} ^ {K \\times b} \\mid \\mathbf {Q} \\mathbf {1} _ {b} = \\frac {1}{K} \\mathbf {1} _ {k}, \\mathbf {Q} ^ {\\top} \\mathbf {1} _ {K} = \\frac {1}{b} \\mathbf {1} _ {b} \\right\\} \\tag {7} \\\\ \\end{array}", + "image_path": "bbb9d11b5b9d884ecde0d70b8012da7333f0e5dc7861f24d87bfb1d6ea0a96fc.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{H}(\\cdot)" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " is the entropy function, " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " is a hyperparameter and " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\operatorname{Tr}(\\cdot)" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " is the trace function. The solution of Eq. (7) is obtained using the Sinkhorn-Knopp algorithm [11]. Please refer to Appendix A for more optimization details." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23075" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 179 + ], + "type": "text", + "content": "Label Refinery with Known-Novel Separation. After the rough assignment, we obtain unpolished soft labels " + }, + { + "bbox": [ + 46, + 72, + 286, + 179 + ], + "type": "inline_equation", + "content": "q \\in [0,1]^K" + }, + { + "bbox": [ + 46, + 72, + 286, + 179 + ], + "type": "text", + "content": ". However, the aforementioned skewed predictions towards known classes stemming from overfitting remain unresolved. To rectify this, we prioritize the subtask of precisely demarcating known and novel classes for label refinery, which is achieved by including an auxiliary binary head " + }, + { + "bbox": [ + 46, + 72, + 286, + 179 + ], + "type": "inline_equation", + "content": "h_{aux}" + }, + { + "bbox": [ + 46, + 72, + 286, + 179 + ], + "type": "text", + "content": " and reformulating this binary task as the classical positive unlabeled (PU) learning problem [2, 13, 62]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 286, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 286, + 299 + ], + "type": "text", + "content": "PU learning tackles a special case of binary classification where the model has access to a labeled set containing merely positive samples and an unlabeled set with both positive and negative samples. Thus, the task of distinguishing known and novel classes can be intuitively converted to PU learning by regarding known and novel as positive and negative respectively. A flurry of methods has been designed for PU learning [10, 27, 34], among which we adopt a prior-free variational algorithm [8] with loss " + }, + { + "bbox": [ + 46, + 180, + 286, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{aux}" + }, + { + "bbox": [ + 46, + 180, + 286, + 299 + ], + "type": "text", + "content": " to optimize the newly devised " + }, + { + "bbox": [ + 46, + 180, + 286, + 299 + ], + "type": "inline_equation", + "content": "h_{aux}" + }, + { + "bbox": [ + 46, + 180, + 286, + 299 + ], + "type": "text", + "content": ". See Appendix A for more details." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 300, + 286, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 286, + 335 + ], + "type": "text", + "content": "After that, we elicit the binary prediction " + }, + { + "bbox": [ + 47, + 300, + 286, + 335 + ], + "type": "inline_equation", + "content": "\\hat{y}^{aux}" + }, + { + "bbox": [ + 47, + 300, + 286, + 335 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 300, + 286, + 335 + ], + "type": "inline_equation", + "content": "h_{aux}" + }, + { + "bbox": [ + 47, + 300, + 286, + 335 + ], + "type": "text", + "content": ", where 1 and 0 indicate prediction of known and novel respectively. Then we leverage " + }, + { + "bbox": [ + 47, + 300, + 286, + 335 + ], + "type": "inline_equation", + "content": "\\hat{y}^{aux}" + }, + { + "bbox": [ + 47, + 300, + 286, + 335 + ], + "type": "text", + "content": " to refine the label " + }, + { + "bbox": [ + 47, + 300, + 286, + 335 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 47, + 300, + 286, + 335 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 342, + 286, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 342, + 286, + 374 + ], + "spans": [ + { + "bbox": [ + 89, + 342, + 286, + 374 + ], + "type": "interline_equation", + "content": "q ^ {j} = \\left\\{ \\begin{array}{l l} q ^ {j} & \\text {i f} \\mathbb {I} (j \\in \\mathcal {C} _ {\\text {k n o w n}}) = \\hat {y} ^ {\\text {a u x}} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {8}", + "image_path": "dcb72da3e32627f985344c61d702bfa697acd8d9425a687b24eb294bb592cabd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "inline_equation", + "content": "q^{j}" + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "text", + "content": "-th entry of " + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "inline_equation", + "content": "1 \\leq j \\leq K" + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(\\cdot)" + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "text", + "content": " is the indicator function. That is, we retain the known/novel class components of soft label " + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "text", + "content": " based on prediction " + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "inline_equation", + "content": "\\hat{y}^{aux}" + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "text", + "content": " and discard the remaining inconsistent components. The refined " + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "text", + "content": " are then utilized for the final hard pseudo-label generation " + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "inline_equation", + "content": "\\hat{y}_{i} = \\arg \\max_{1 \\leq k \\leq K} q^{k}" + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "text", + "content": " of Eq. (6)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 468, + 286, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 286, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 286, + 587 + ], + "type": "text", + "content": "Pseudo-Label Filtering. While the proposed sample-target allocation mechanism can enhance the quality of pseudo-labels, the remaining obstinate noisy pseudo-labels can still impede model training, especially in the early stages when the model output is not accurate enough. To neutralize this, we further utilize a high-confidence selection to select clean pseudo-labels from noisy ones for " + }, + { + "bbox": [ + 46, + 468, + 286, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u" + }, + { + "bbox": [ + 46, + 468, + 286, + 587 + ], + "type": "text", + "content": ". Formally, for each class " + }, + { + "bbox": [ + 46, + 468, + 286, + 587 + ], + "type": "inline_equation", + "content": "1 \\leq j \\leq K" + }, + { + "bbox": [ + 46, + 468, + 286, + 587 + ], + "type": "text", + "content": ", we select a clean subset " + }, + { + "bbox": [ + 46, + 468, + 286, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sel}^j" + }, + { + "bbox": [ + 46, + 468, + 286, + 587 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 468, + 286, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u^j = \\{(x, \\hat{y}) \\in \\mathcal{D}_u | \\hat{y} = j\\}" + }, + { + "bbox": [ + 46, + 468, + 286, + 587 + ], + "type": "text", + "content": " class by class with the highest confidence scores in the first " + }, + { + "bbox": [ + 46, + 468, + 286, + 587 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 468, + 286, + 587 + ], + "type": "text", + "content": " percent," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 596, + 286, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 596, + 286, + 611 + ], + "spans": [ + { + "bbox": [ + 84, + 596, + 286, + 611 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {s e l} ^ {j} = \\left\\{\\left(x, \\hat {y}\\right) \\in \\mathcal {D} _ {u} ^ {j} \\mid r a n k \\left(q ^ {j}\\right) < R \\% \\right\\} \\tag{9}", + "image_path": "072fe1f69390ec770e323daaeae4ceca4bbf34b22e63a6d006528febab3bcfbe.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 618, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 286, + 689 + ], + "type": "text", + "content": "Then the integral selected set is merged as " + }, + { + "bbox": [ + 46, + 618, + 286, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sel} = \\cup_{j=1}^{K} \\mathcal{D}_{sel}^{j}" + }, + { + "bbox": [ + 46, + 618, + 286, + 689 + ], + "type": "text", + "content": ". The classification loss in Eq. (6) is then merely valid on " + }, + { + "bbox": [ + 46, + 618, + 286, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sel}" + }, + { + "bbox": [ + 46, + 618, + 286, + 689 + ], + "type": "text", + "content": ". In practice, we set a small " + }, + { + "bbox": [ + 46, + 618, + 286, + 689 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 618, + 286, + 689 + ], + "type": "text", + "content": " to ensure the precision of selected samples at the beginning and gradually increase it to integrate more clean samples. Finally, the overall training loss is given by," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 86, + 700, + 286, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 700, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 86, + 700, + 286, + 712 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {a u x}} + \\mathcal {L} _ {\\text {c o n}} + \\alpha \\mathcal {L} _ {\\text {e n t}} \\tag {10}", + "image_path": "c8e90a0a8f7e7345c18536de287863e4d5605018d09955f64ff0b9678808dc8f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ent}" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": " are described in Eq. (1). Through this procedure, Our sample-target allocation mechanism is able to provide pseudo-labels with high quality and guide the representation calibration toward the correct targets. The overview of TRAILER is illustrated in Figure 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 144, + 381, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 144, + 381, + 157 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 381, + 157 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 164, + 545, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 164, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 305, + 164, + 545, + 212 + ], + "type": "text", + "content": "In this section, we present the main results and part of the ablation results to verify the effectiveness of TRAILER. More experimental details and results, such as the results with less labeled data, can be found in Appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 220, + 355, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 220, + 355, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 355, + 232 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 239, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 239, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 545, + 394 + ], + "type": "text", + "content": "Datasets. We first evaluate TRAILER on three datasets CIFAR-10, CIFAR-100 [29] and ImageNet-100, which is randomly sub-sampled with 100 classes from ImageNet [12] following [5]. For each dataset, We first divide classes into the first " + }, + { + "bbox": [ + 305, + 239, + 545, + 394 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 305, + 239, + 545, + 394 + ], + "type": "text", + "content": " known and the rest " + }, + { + "bbox": [ + 305, + 239, + 545, + 394 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 305, + 239, + 545, + 394 + ], + "type": "text", + "content": " novel classes following [5]. Then we randomly select a portion " + }, + { + "bbox": [ + 305, + 239, + 545, + 394 + ], + "type": "inline_equation", + "content": "(50\\%)" + }, + { + "bbox": [ + 305, + 239, + 545, + 394 + ], + "type": "text", + "content": " of samples from the known classes as labeled " + }, + { + "bbox": [ + 305, + 239, + 545, + 394 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_l" + }, + { + "bbox": [ + 305, + 239, + 545, + 394 + ], + "type": "text", + "content": ", and the remaining along with all novel class samples are regarded as unlabeled " + }, + { + "bbox": [ + 305, + 239, + 545, + 394 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u" + }, + { + "bbox": [ + 305, + 239, + 545, + 394 + ], + "type": "text", + "content": ". We also conduct experiments with different novel class ratios and on the fine-grained semantic shift benchmarks (SSB) [51], including CUB-200 [53], Stanford Cars [28], FGVC-Aircraft [35], and a naturally long-tailed Herbarium 19 [47] datasets, as discussed in Section 4.3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 410, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 410, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 305, + 410, + 545, + 650 + ], + "type": "text", + "content": "Baselines. We compare TRAILER with various open-world SSL baselines, including ORCA [5], GCD [50], OpenCon [46], NACH [20], OpenLDN [41], GPC [61] and OpenNCD [33]. We also provide comparisons with SSL, open-set SSL, and NCD baselines following [20, 33]. For traditional and open-set SSL baselines including FixMatch [44], " + }, + { + "bbox": [ + 305, + 410, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\mathrm{DS}^3\\mathrm{L}" + }, + { + "bbox": [ + 305, + 410, + 545, + 650 + ], + "type": "text", + "content": " [19], and CGDL [45], we extend them to be applicable to novel classes by selecting samples with lower confidence scores as novel and applying " + }, + { + "bbox": [ + 305, + 410, + 545, + 650 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 410, + 545, + 650 + ], + "type": "text", + "content": "-means clustering to obtain results. For NCD methods including DTC [21], RankStats [22], and " + }, + { + "bbox": [ + 305, + 410, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\mathrm{UNO+}" + }, + { + "bbox": [ + 305, + 410, + 545, + 650 + ], + "type": "text", + "content": " [15], they are extended to classify known classes by using Hungarian algorithm to match some discovered classes with known classes. For performances of baselines, we directly adopt reported results from receptive papers or from [33], except OpenLDN was retested under transductive setting. All compared methods are implemented based on pre-trained ResNet using SimCLR [9] following [5], except DTC and OpenCon which have specialized pre-training procedure. We run all experiments three times and report the averaged results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. Following [5, 33], we evaluate the transductive performance on both known and novel classes. For novel classes, we utilize the Hungarian algorithm to derive the optimal assignments and then calculate accuracy" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23076" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 526, + 253 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 526, + 253 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 526, + 253 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 526, + 253 + ], + "type": "table", + "html": "
MethodsCIFAR-10CIFAR-100ImageNet-100
KnownNovelAllKnownNovelAllKnownNovelAll
FixMatch [44]71.550.449.539.623.520.365.836.734.9
DS3L [19]77.645.340.255.123.724.071.232.530.8
CGDL [45]72.344.639.749.322.523.567.333.831.9
DTC [21]53.939.538.331.322.918.325.620.821.3
RankStats [22]86.681.082.936.428.423.147.328.740.3
SimCLR [9]58.363.451.728.621.122.339.535.736.9
ORCA [5]88.290.489.766.943.048.189.172.177.8
GCD [50]78.479.779.168.533.545.282.358.368.2
OpenLDN [41]92.386.888.667.340.449.474.545.755.3
OpenCon [46]90.491.189.369.147.852.790.680.883.8
NACH [20]89.592.291.368.747.052.191.075.579.6
OpenNCD [33]88.490.690.169.743.449.390.077.581.6
TRAILER (Ours)93.495.094.469.748.755.691.482.485.3
", + "image_path": "234479152dab275c8f9c4fbdb13b05c185d3b8f44d9997f30d85fef57283ae3f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 301, + 282, + 405 + ], + "blocks": [ + { + "bbox": [ + 46, + 261, + 546, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 546, + 285 + ], + "type": "text", + "content": "Table 1. Accuracy comparison of known, novel, and all classes on CIFAR-10, CIFAR-100 and ImaneNet-100 dataset. The dataset is composed of " + }, + { + "bbox": [ + 46, + 261, + 546, + 285 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 261, + 546, + 285 + ], + "type": "text", + "content": " known classes and " + }, + { + "bbox": [ + 46, + 261, + 546, + 285 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 261, + 546, + 285 + ], + "type": "text", + "content": " novel classes, with " + }, + { + "bbox": [ + 46, + 261, + 546, + 285 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 261, + 546, + 285 + ], + "type": "text", + "content": " of the known classes labeled. Bold entries indicate superior results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 301, + 282, + 405 + ], + "lines": [ + { + "bbox": [ + 51, + 301, + 282, + 405 + ], + "spans": [ + { + "bbox": [ + 51, + 301, + 282, + 405 + ], + "type": "table", + "html": "
AblationCIFAR-10CIFAR-100
KnownNovelAllKnownNovelAll
TRAILER93.495.094.469.748.755.6
w/o Rep Calibration94.158.970.669.214.933.4
with λ = 093.991.892.570.246.654.6
w/o Rough Assign92.992.392.569.337.948.3
w/o Label Refinery95.289.691.570.345.453.7
w/o PL Filtering92.290.591.069.943.652.2
", + "image_path": "74c5753d05df0f0bd5cd1b33e2be4bbe0ebbe97eb79cbb253838e98cdf3dbe2e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 413, + 286, + 436 + ], + "lines": [ + { + "bbox": [ + 46, + 413, + 286, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 286, + 436 + ], + "type": "text", + "content": "Table 2. Ablation results on CIFAR-10 and CIFAR-100. 'Rep' and 'PL' indicate representation and pseudo-label respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 456, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\frac{1}{n}\\sum_{i = 1}^{n}\\mathbb{I}(y_i = \\text{perm}(\\hat{y}_i))" + }, + { + "bbox": [ + 47, + 456, + 287, + 493 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 47, + 456, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\text{perm}(\\cdot)" + }, + { + "bbox": [ + 47, + 456, + 287, + 493 + ], + "type": "text", + "content": " is the matching permutation. The overall joint accuracy is also measured by Hungarian match using both known and novel classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 508, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 662 + ], + "type": "text", + "content": "Implementation Details. For CIFAR experiments, we use ResNet-18 as the backbone. The model is trained for 200 epochs with warm-up of 15 epochs. We employ the Adam optimizer with a batch size of 200 and an initial learning rate of " + }, + { + "bbox": [ + 46, + 508, + 287, + 662 + ], + "type": "inline_equation", + "content": "5e^{-4}" + }, + { + "bbox": [ + 46, + 508, + 287, + 662 + ], + "type": "text", + "content": ", which decays by a cosine scheduler. The temperature " + }, + { + "bbox": [ + 46, + 508, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 508, + 287, + 662 + ], + "type": "text", + "content": " is set as 0.3 and 0.5 for CIFAR-10 and CIFAR-100. We linearly ramp down " + }, + { + "bbox": [ + 46, + 508, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 508, + 287, + 662 + ], + "type": "text", + "content": " from 1 to 0 and ramp up selection ratio " + }, + { + "bbox": [ + 46, + 508, + 287, + 662 + ], + "type": "inline_equation", + "content": "R\\%" + }, + { + "bbox": [ + 46, + 508, + 287, + 662 + ], + "type": "text", + "content": " from 0.3 to 0.9 for pseudo-label filtering. The label refinery is enabled after training " + }, + { + "bbox": [ + 46, + 508, + 287, + 662 + ], + "type": "inline_equation", + "content": "h_{aux}" + }, + { + "bbox": [ + 46, + 508, + 287, + 662 + ], + "type": "text", + "content": " for 15 epochs. For the optimization of Sinkhorn-Knopp, we adopt hyperparameters from [15] and set " + }, + { + "bbox": [ + 46, + 508, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.05" + }, + { + "bbox": [ + 46, + 508, + 287, + 662 + ], + "type": "text", + "content": " with 3 iterations. We put more implementation details for ImageNet-100 and fine-grained dataset in Appendix A." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 671, + 127, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 127, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 127, + 682 + ], + "type": "text", + "content": "4.2. Main Reults" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Table 1 shows the comparison results on the CIFAR-10, CIFAR-100, and ImageNet-100 datasets, including" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 304, + 424, + 395 + ], + "blocks": [ + { + "bbox": [ + 309, + 304, + 424, + 395 + ], + "lines": [ + { + "bbox": [ + 309, + 304, + 424, + 395 + ], + "spans": [ + { + "bbox": [ + 309, + 304, + 424, + 395 + ], + "type": "image", + "image_path": "d80f3225708c2a7771dd47ba19da40f81c48c8372df5f7c1b33659e51222812d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 399, + 403, + 408 + ], + "lines": [ + { + "bbox": [ + 328, + 399, + 403, + 408 + ], + "spans": [ + { + "bbox": [ + 328, + 399, + 403, + 408 + ], + "type": "text", + "content": "(a) Without calibration." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 430, + 305, + 543, + 395 + ], + "blocks": [ + { + "bbox": [ + 430, + 305, + 543, + 395 + ], + "lines": [ + { + "bbox": [ + 430, + 305, + 543, + 395 + ], + "spans": [ + { + "bbox": [ + 430, + 305, + 543, + 395 + ], + "type": "image", + "image_path": "9793e6453ee8e88732ea8c24047e8d31aa01041aaf25485ae5e180474f2aadd5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 399, + 520, + 408 + ], + "lines": [ + { + "bbox": [ + 454, + 399, + 520, + 408 + ], + "spans": [ + { + "bbox": [ + 454, + 399, + 520, + 408 + ], + "type": "text", + "content": "(b) With calibration." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 418, + 545, + 452 + ], + "lines": [ + { + "bbox": [ + 305, + 418, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 545, + 452 + ], + "type": "text", + "content": "Figure 3. T-SNE feature visualization on CIFAR-10. Different colors represent the corresponding ground-truth classes. Without calibration, the dog and cat classes collapse into the same cluster." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 472, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 640 + ], + "type": "text", + "content": "the accuracy of the 'Known', 'Novel', and 'All' classes. TRAILER significantly outperforms all the rivals by a considerable margin in terms of overall accuracy on different datasets. Specifically, on CIFAR-10, TRAILER improves upon the best baseline by " + }, + { + "bbox": [ + 304, + 472, + 545, + 640 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 304, + 472, + 545, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 472, + 545, + 640 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 304, + 472, + 545, + 640 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 472, + 545, + 640 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 304, + 472, + 545, + 640 + ], + "type": "text", + "content": " for the 'Known', 'Novel', and 'All' classes respectively. Moreover, on the more challenging CIFAR-100 and ImageNet-100 datasets with larger label space, while most baselines demonstrate a notable performance drop, TRAILER shows robustness and reaches an extraordinary trade-off between the known and novel classes. It consistently maintains superior performance with " + }, + { + "bbox": [ + 304, + 472, + 545, + 640 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 304, + 472, + 545, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 472, + 545, + 640 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 304, + 472, + 545, + 640 + ], + "type": "text", + "content": " overall enhancement. These results empirically demonstrate the efficacy of our proposed TRAILER for the task of open-world SSL." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 647, + 368, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 368, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 368, + 659 + ], + "type": "text", + "content": "4.3. Analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": "Results on Semantic Shift Benchmarks. We further conduct experiments on fine-grained semantic shift benchmarks (SSB) [51] following previous work [50, 61]. For fair comparisons, we adopt the same training and compar" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23077" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 527, + 198 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 527, + 198 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 527, + 198 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 527, + 198 + ], + "type": "table", + "html": "
MethodsCUBStanford CarsFGVC-AircraftHerbarium 19
KnownNovelAllKnownNovelAllKnownNovelAllKnownNovelAll
k-means38.932.134.310.613.812.814.416.816.012.912.812.9
RankStats+ [22]51.624.233.361.812.128.336.422.226.955.812.827.9
UNO+ [15]49.028.135.170.518.635.556.432.240.353.714.728.3
ORCA [5]45.630.235.350.110.723.531.817.122.030.915.520.9
GCD [50]56.648.751.357.629.939.041.146.945.051.027.035.4
OpenCon [46]63.852.154.778.632.749.1---58.928.639.3
GPC [61]58.253.155.459.232.842.842.547.946.351.727.936.5
TRAILER (Ours)71.361.965.171.747.655.462.650.554.557.037.844.5
", + "image_path": "eb77879e2befc76d3d760c2b0d79007a9d80afb9320522477bc69eee5ee1ac86.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 206, + 545, + 217 + ], + "lines": [ + { + "bbox": [ + 47, + 206, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 545, + 217 + ], + "type": "text", + "content": "Table 3. Results on the Semantic Shift Benchmarks. Bold entries indicate superior results and blank ones indicate results are not provided." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 237, + 165, + 329 + ], + "blocks": [ + { + "bbox": [ + 50, + 237, + 165, + 329 + ], + "lines": [ + { + "bbox": [ + 50, + 237, + 165, + 329 + ], + "spans": [ + { + "bbox": [ + 50, + 237, + 165, + 329 + ], + "type": "image", + "image_path": "ee64ebf940375ccd6ac8ad37a41a8d6be115f5df96dec19ffde11e9399fc5440.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 333, + 154, + 342 + ], + "lines": [ + { + "bbox": [ + 60, + 333, + 154, + 342 + ], + "spans": [ + { + "bbox": [ + 60, + 333, + 154, + 342 + ], + "type": "text", + "content": "(a) Binary accuracy of " + }, + { + "bbox": [ + 60, + 333, + 154, + 342 + ], + "type": "inline_equation", + "content": "h_{aux}" + }, + { + "bbox": [ + 60, + 333, + 154, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 170, + 237, + 286, + 329 + ], + "blocks": [ + { + "bbox": [ + 170, + 237, + 286, + 329 + ], + "lines": [ + { + "bbox": [ + 170, + 237, + 286, + 329 + ], + "spans": [ + { + "bbox": [ + 170, + 237, + 286, + 329 + ], + "type": "image", + "image_path": "2672ce8cfb7404415f6ba6e8148ac5c272989e7bc84c0f88fca993d2e70743d2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 333, + 271, + 342 + ], + "lines": [ + { + "bbox": [ + 184, + 333, + 271, + 342 + ], + "spans": [ + { + "bbox": [ + 184, + 333, + 271, + 342 + ], + "type": "text", + "content": "(b) Pseudo-label precision." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 352, + 287, + 396 + ], + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 396 + ], + "type": "text", + "content": "Figure 4. (a) Accuracy of binary prediction between known and novel classes in first 100 epochs. (b) Per-class pseudo-label precision for unlabeled samples on CIFAR-10 with and without label refinery. Class indexes 0-4 are known while 5-9 are novel." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "text", + "content": "isons protocol from [50] and utilize ViT-B/16 backbone [14] with DINO pre-trained weights [7]. As shown in Table 3, it can be observed that TRAILER exceeds the baselines by a substantial lead in all metrics, with improvements of " + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "inline_equation", + "content": "8.8\\%" + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "inline_equation", + "content": "14.8\\%" + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "inline_equation", + "content": "9.2\\%" + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "text", + "content": " for 'Novel' accuracy on these datasets. These results consistently demonstrate its superior performance, even under challenging fine-grained tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 521, + 286, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 286, + 688 + ], + "type": "text", + "content": "Benefits of Representation Calibration. To explore the benefits of representation calibration, we conduct ablations and equip TRAILER with different representation learning strategies: 1) TRAILER w/o Rep Calibration which dismiss the feature alignment of the targeted classifier and instead utilize the vanilla learnable classifier; 2) TRAILER with " + }, + { + "bbox": [ + 46, + 521, + 286, + 688 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 46, + 521, + 286, + 688 + ], + "type": "text", + "content": " which discards the progressive manner for direct calibration. From Table 2, although TRAILER w/o Rep Calibration remains competitive on known classes, it suffers from a calamitous performance plunge for novel classes, which indicates the efficacy of representation calibration. TRAILER with " + }, + { + "bbox": [ + 46, + 521, + 286, + 688 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 46, + 521, + 286, + 688 + ], + "type": "text", + "content": " achieves decent performance, but still underperforms TRAILER perceptibly, indicating the advantage of the progressive manner." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "We further visualize the representation with and without representation calibration on CIFAR-10 using t-SNE [49]" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 238, + 533, + 324 + ], + "blocks": [ + { + "bbox": [ + 313, + 238, + 533, + 324 + ], + "lines": [ + { + "bbox": [ + 313, + 238, + 533, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 533, + 324 + ], + "type": "image", + "image_path": "1a0d947f2818acd11e2df042fe87a7e951f7a25822e96b25c783e6ac549d63e9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 333, + 545, + 354 + ], + "lines": [ + { + "bbox": [ + 306, + 333, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 306, + 333, + 545, + 354 + ], + "type": "text", + "content": "Figure 5. Confusion matrix between known and novel classes on CIFAR-10. TRAILER can alleviate biased predictions." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "type": "text", + "content": "in Figure 3 and in Appendix B. It can be shown that representation calibration facilitates inducing compact and nonskewed representations, discarding which can lead to representation collapse—the representations are biased towards known classes, where the dog class (novel, brown) and part of horse class samples (novel, grey) collapse into the cluster of the cat class (known, red), as shown in Figure 3a." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": "Effect of Sample-Target Allocation. Next, we ablate the effectiveness of the sample-target allocation mechanism in Table 2. For the coarse-level assignment, we test the variant TRAILER w/o Rough Assign which sticks to the vanilla softmax function for rough label assignment of " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": " and suffers from rigorous performance degradation, especially on CIFAR-100. For the fine-level refinery, we test the variant TRAILER w/o Label Refinery which discards the refinery procedure of auxiliary head " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "h_{aux}" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": ". Although such a variant exhibits slightly better performance for the known classes, it significantly lags behind TRAILER for the novel classes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": "To demonstrate it more intuitively, we visualize the accuracy of binary prediction for " + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "inline_equation", + "content": "h_{aux}" + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": " between known and novel classes in Figure 4a. We can find that " + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "inline_equation", + "content": "h_{aux}" + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": " quickly achieves high prediction accuracies and such accurate separation between known and novel indeed refines the rough assignment supplementarily, finally enhancing the pseudo-label precision on 9 of 10 classes, as shown in Figure 4b. Such a mechanism also alleviates the biased prediction, as shown in Figure 5 where a non-trivial amount of novel samples are" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23078" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 165, + 163 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 165, + 163 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 165, + 163 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 165, + 163 + ], + "type": "image", + "image_path": "58b98580f4eedf51fa54f741d2dba1253968974d3bdf02a037548209fb8c581f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 167, + 137, + 176 + ], + "lines": [ + { + "bbox": [ + 78, + 167, + 137, + 176 + ], + "spans": [ + { + "bbox": [ + 78, + 167, + 137, + 176 + ], + "type": "text", + "content": "(a) Equiangularity" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 71, + 284, + 163 + ], + "blocks": [ + { + "bbox": [ + 168, + 71, + 284, + 163 + ], + "lines": [ + { + "bbox": [ + 168, + 71, + 284, + 163 + ], + "spans": [ + { + "bbox": [ + 168, + 71, + 284, + 163 + ], + "type": "image", + "image_path": "412f447defbe88d237d77246a05c22bc5320c1a434804aeba54c83ec05d0de39.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 167, + 265, + 176 + ], + "lines": [ + { + "bbox": [ + 189, + 167, + 265, + 176 + ], + "spans": [ + { + "bbox": [ + 189, + 167, + 265, + 176 + ], + "type": "text", + "content": "(b) Maximal separation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 186, + 287, + 229 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 287, + 229 + ], + "type": "text", + "content": "Figure 6. Visualizations on the (a) equiangularity and (b) maximal separation properties of neural collapse on CIFAR-10 in the first 100 epochs. Representation calibration enables TRAILER to achieve feature arrangement with equal and maximal separation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 52, + 242, + 282, + 310 + ], + "blocks": [ + { + "bbox": [ + 52, + 242, + 282, + 310 + ], + "lines": [ + { + "bbox": [ + 52, + 242, + 282, + 310 + ], + "spans": [ + { + "bbox": [ + 52, + 242, + 282, + 310 + ], + "type": "table", + "html": "
AblationEstimation of KKnownNovelAll
TRAILERKnown (100)69.748.755.6
ORCAKnown (100)66.943.048.1
TRAILEREstimated (124)71.144.353.3
ORCAEstimated (124)66.340.046.4
", + "image_path": "3a09091dea07e146cbdad5cea38044afd4246003d82c5756cb2f4b92459b4565.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 318, + 286, + 329 + ], + "lines": [ + { + "bbox": [ + 47, + 318, + 286, + 329 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 286, + 329 + ], + "type": "text", + "content": "Table 4. Ablations with unknown class number " + }, + { + "bbox": [ + 47, + 318, + 286, + 329 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 318, + 286, + 329 + ], + "type": "text", + "content": " on CIFAR-100." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 352, + 277, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 352, + 277, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 352, + 277, + 364 + ], + "type": "text", + "content": "misidentified as known (the number in red) for baselines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 382, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 287, + 588 + ], + "type": "text", + "content": "Empirical Visualizations for Neural Collapse. In the previous section, we have provided the embedding visualization. Here we supply some quantitative results in addition. Recall that two core properties of NC are equiangularity and maximal separation. Given the centered class-means " + }, + { + "bbox": [ + 46, + 382, + 287, + 588 + ], + "type": "inline_equation", + "content": "\\hat{z}^k" + }, + { + "bbox": [ + 46, + 382, + 287, + 588 + ], + "type": "text", + "content": " for features from all classes " + }, + { + "bbox": [ + 46, + 382, + 287, + 588 + ], + "type": "inline_equation", + "content": "1 \\leq k \\leq K" + }, + { + "bbox": [ + 46, + 382, + 287, + 588 + ], + "type": "text", + "content": ", for equiangularity, we calculate the standard deviation of the cosines between each pair " + }, + { + "bbox": [ + 46, + 382, + 287, + 588 + ], + "type": "inline_equation", + "content": "\\mathrm{Std}_{k \\neq k'}(\\cos (\\hat{z}^k, \\hat{z}^{k'}))" + }, + { + "bbox": [ + 46, + 382, + 287, + 588 + ], + "type": "text", + "content": ". For maximal separation, we calculate the average value " + }, + { + "bbox": [ + 46, + 382, + 287, + 588 + ], + "type": "inline_equation", + "content": "\\mathrm{Avg}_{k \\neq k'}(\\cos (\\hat{z}^k, \\hat{z}^{k'}) + \\frac{1}{K - 1})" + }, + { + "bbox": [ + 46, + 382, + 287, + 588 + ], + "type": "text", + "content": " since the optimal pair-wise angle is " + }, + { + "bbox": [ + 46, + 382, + 287, + 588 + ], + "type": "inline_equation", + "content": "-\\frac{1}{K - 1}" + }, + { + "bbox": [ + 46, + 382, + 287, + 588 + ], + "type": "text", + "content": " as discussed in Eq. (3). The trends of these two metrics on the CIFAR-10 dataset with and without representation calibration are shown in Figure 6. Notably, equipped with representation calibration, the standard deviations of the cosines and shifted average cosine values approach zero, which indicates that TRAILER achieves the engaging feature structure that is close to neural collapse." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "Unknown Number of Novel Classes. The previous experiments follow previous protocols and assume that the class number is known apriori [5, 15, 46], which is sometimes impractical in real-world applications and brings challenges for both TRAILER and the other baselines. A simple resolution is to first use estimation methods [32, 39] to derive the number of classes before deployment, for example, " + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "-estimation with Brent's algorithm from [50] or cross-validating using labeled probe set from [5, 21, 46]. On" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 424, + 163 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 424, + 163 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 424, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 424, + 163 + ], + "type": "image", + "image_path": "a3d4c1dd03d8f9f2a88416addad0a78fd8eb2a950719e383e1aea43362dd5291.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 167, + 417, + 175 + ], + "lines": [ + { + "bbox": [ + 315, + 167, + 417, + 175 + ], + "spans": [ + { + "bbox": [ + 315, + 167, + 417, + 175 + ], + "type": "text", + "content": "(a) Accuracy for known classes." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 186, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 305, + 186, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 545, + 208 + ], + "type": "text", + "content": "Figure 7. Accuracy comparisons for known and novel classes on CIFAR-10 with different novel class ratios ranging from 0.1 to 0.9." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 428, + 72, + 543, + 163 + ], + "blocks": [ + { + "bbox": [ + 428, + 72, + 543, + 163 + ], + "lines": [ + { + "bbox": [ + 428, + 72, + 543, + 163 + ], + "spans": [ + { + "bbox": [ + 428, + 72, + 543, + 163 + ], + "type": "image", + "image_path": "489b6e23a2ef86d0b8cea91eac19ad5b617b4c676c71bd08bb0860e23c9536dd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 167, + 536, + 175 + ], + "lines": [ + { + "bbox": [ + 436, + 167, + 536, + 175 + ], + "spans": [ + { + "bbox": [ + 436, + 167, + 536, + 175 + ], + "type": "text", + "content": "(b) Accuracy for novel classes." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 231, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 231, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 231, + 545, + 338 + ], + "type": "text", + "content": "CIFAR-100 the former gets exactly 100 while the latter derives 124 and was adopted by us for further exploration. We then utilize this estimation result to rerun the experiment by initializing the corresponding number of classifier vectors. As shown in Table 4, TRAILER is able to achieve competitive performance even when the number of classes is not known in advance. We further provide additional results with different estimated class numbers " + }, + { + "bbox": [ + 305, + 231, + 545, + 338 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 231, + 545, + 338 + ], + "type": "text", + "content": " from 80 to 130 in Appendix B to demonstrate the robustness of TRAILER." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 357, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 357, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 305, + 357, + 545, + 441 + ], + "type": "text", + "content": "Effect of Novel Class Ratio. Here we further evaluate the performance of TRAILER on the CIFAR-10 dataset with different ratios of novel classes from 0.1 to 0.9. As shown in Figure 7, we can see that TRAILER displays robust performance when the ratio of novel classes grows larger from 0.1 to 0.9 and consistently manifests superior performance than other baselines on both known and novel classes." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 454, + 378, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 378, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 378, + 467 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 475, + 545, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 655 + ], + "type": "text", + "content": "In this work, we propose a novel framework TRAILER for open-world SSL. We first take inspiration from the recently discovered neural collapse phenomenon and intend to attain its appealing feature arrangement with minimal within-class and maximum between-class covariance. To achieve this, we adopt a targeted classifier and align representations towards its pre-assigned optimal structure in a progressive manner. To mitigate the potential risk of misalignment, we tackle the sample-target assignment for unlabeled samples with a coarse-to-fine allocation mechanism. Comprehensive experiments show that TRAILER enhances the baselines by a notable margin on both generic and fine-grained benchmarks. We hope our work can inspire future research to leverage the appealing properties of the neural collapse phenomenon for tasks in open-world scenarios." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 669, + 415, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 415, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 415, + 682 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "This work is supported by the Pioneer R&D Program of Zhejiang (No. 2024C01035)." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23079" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR. OpenReview.net, 2020. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 287, + 158 + ], + "type": "text", + "content": "[2] Jessa Bekker and Jesse Davis. Learning from positive and unlabeled data: a survey. Mach. Learn., 109(4):719-760, 2020. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 287, + 213 + ], + "type": "text", + "content": "[3] David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution alignment and augmentation anchoring. CoRR, abs/1911.09785, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 287, + 259 + ], + "type": "text", + "content": "[4] David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In NeurIPS, pages 5050-5060, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 287, + 293 + ], + "type": "text", + "content": "[5] Kaidi Cao, Maria Brbic, and Jure Leskovec. Open-world semi-supervised learning. In ICLR. OpenReview.net, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 295, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 338 + ], + "type": "text", + "content": "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 340, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 287, + 384 + ], + "type": "text", + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9630-9640. IEEE, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 386, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 287, + 418 + ], + "type": "text", + "content": "[8] Hui Chen, Fangqing Liu, Yin Wang, Liyue Zhao, and Hao Wu. A variational approach for learning from positive and unlabeled data. In NeurIPS 2020, 2020. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 420, + 287, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 287, + 463 + ], + "type": "text", + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. PMLR, 2020. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 465, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 287, + 509 + ], + "type": "text", + "content": "[10] Xuxi Chen, Wuyang Chen, Tianlong Chen, Ye Yuan, Chen Gong, Kewei Chen, and Zhangyang Wang. Self-pu: Self boosted and calibrated positive-unlabeled training. In ICML, pages 1510-1519. PMLR, 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 510, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 533 + ], + "type": "text", + "content": "[11] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NeurIPS, pages 2292-2300, 2013. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 534, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 576 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE Computer Society, 2009. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 623 + ], + "type": "text", + "content": "[13] François Denis. PAC learning from positive statistical queries. In Algorithmic Learning Theory, 9th International Conference, ALT '98, Otzenhausen, Germany, October 8-10, 1998, Proceedings, pages 112-126. Springer, 1998. 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR. OpenReview.net, 2021. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[15] Enrico Fini, Enver Sangineto, Stephane Lathuilière, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "novel class discovery. In ICCV, pages 9264-9272. IEEE, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "text", + "content": "[16] Ronald A Fisher. The use of multiple measurements in taxonomic problems. Annals of eugenics, 7(2):179-188, 1936. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 130, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 162 + ], + "type": "text", + "content": "[17] Tomer Galanti, András György, and Marcus Hutter. On the role of neural collapse in transfer learning. In ICLR. Open-Review.net, 2022. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 164, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 196 + ], + "type": "text", + "content": "[18] Florian Graf, Christoph D. Hofer, Marc Niethammer, and Roland Kwitt. Dissecting supervised constrastive learning. In ICML, pages 3821-3830. PMLR, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 198, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 240 + ], + "type": "text", + "content": "[19] Lan-Zhe Guo, Zhenyu Zhang, Yuan Jiang, Yu-Feng Li, and Zhi-Hua Zhou. Safe deep semi-supervised learning for unseen-class unlabeled data. In ICML, pages 3897-3906. PMLR, 2020. 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 243, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 274 + ], + "type": "text", + "content": "[20] Lan-Zhe Guo, Yi-Ge Zhang, Zhi-Fan Wu, Jie-Jing Shao, and Yu-Feng Li. Robust semi-supervised learning when not all classes have labels. In NeurIPS, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 277, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 319 + ], + "type": "text", + "content": "[21] Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer clustering. In ICCV, pages 8400-8408. IEEE, 2019. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "type": "text", + "content": "[22] Kai Han, Sylvestre-Alvise Rebuffi, Sébastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Automatically discovering and learning new visual categories with ranking statistics. In ICLR. OpenReview.net, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 366, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 398 + ], + "type": "text", + "content": "[23] X. Y. Han, Vardan Papyan, and David L. Donoho. Neural collapse under MSE loss: Proximity to and dynamics on the central path. In ICLR. OpenReview.net, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 399, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 441 + ], + "type": "text", + "content": "[24] Zongyan Han, Zhenyong Fu, Shuo Chen, and Jian Yang. Contrastive embedding for generalized zero-shot learning. In CVPR, pages 2371-2381. Computer Vision Foundation / IEEE, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "type": "text", + "content": "[25] Zongyan Han, Zhenyong Fu, Shuo Chen, and Jian Yang. Semantic contrastive embedding for generalized zero-shot learning. Int. J. Comput. Vis., 130(11):2606-2622, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "text", + "content": "[26] Wenlong Ji, Yiping Lu, Yiliang Zhang, Zhun Deng, and Weijie J. Su. An unconstrained layer-peeled perspective on neural collapse. In ICLR. OpenReview.net, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 512, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 553 + ], + "type": "text", + "content": "[27] Ryuichi Kiryo, Gang Niu, Marthinus Christoffel du Plessis, and Masashi Sugiyama. Positive-unlabeled learning with non-negative risk estimator. In NeurIPS, pages 1675-1685, 2017. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 556, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 589 + ], + "type": "text", + "content": "[28] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV, pages 554–561. IEEE Computer Society, 2013. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 591, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 545, + 623 + ], + "type": "text", + "content": "[29] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Master's thesis, Department of Computer Science, University of Toronto, 2009. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 624, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 645 + ], + "type": "text", + "content": "[30] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In ICLR. OpenReview.net, 2017. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[31] Dong-Hyun Lee. Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks. ICML 2013 Workshop : Challenges in Representation Learning (WREPL), 2013. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[32] Collin Leiber, Lena G. M. Bauer, Benjamin Schelling, Christian Böhm, and Claudia Plant. Dip-based deep embedded" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23080" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "text", + "content": "clustering with k-estimation. In KDD, pages 903-913. ACM, 2021. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "type": "text", + "content": "[33] Jiaming Liu, Yangqiming Wang, Tongze Zhang, Yulu Fan, Qinli Yang, and Junming Shao. Open-world semi-supervised novel class discovery. In *IJCAI*, pages 4002–4010. ijcai.org, 2023. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "type": "text", + "content": "[34] Chuan Luo, Pu Zhao, Chen Chen, Bo Qiao, Chao Du, Hongyu Zhang, Wei Wu, Shaowei Cai, Bing He, Saravanakumar Rajmohan, and Qingwei Lin. PULNS: positive unlabeled learning with effective negative sample selector. In AAAI, pages 8784-8792. AAAI Press, 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 227 + ], + "type": "text", + "content": "[35] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. CoRR, abs/1306.5151, 2013. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 259 + ], + "type": "text", + "content": "[36] Aleix M. Martínez and Avinash C. Kak. PCA versus LDA. IEEE Trans. Pattern Anal. Mach. Intell., 23(2):228-233, 2001. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[37] Duc Anh Nguyen, Ron Levie, Julian Lienen, Eyke Hüllermeier, and Gitta Kutyniok. Memorization-dilation: Modeling neural collapse under noise. In ICLR. OpenReview.net, 2023. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "type": "text", + "content": "[38] Vardan Papyan, X. Y. Han, and David L. Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 117(40):24652-24663, 2020. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 350, + 287, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 382 + ], + "type": "text", + "content": "[39] Dan Pelleg and Andrew W. Moore. X-means: Extending k-means with efficient estimation of the number of clusters. In ICML, pages 727-734. Morgan Kaufmann, 2000. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 383, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 425 + ], + "type": "text", + "content": "[40] Zhongzheng Ren, Raymond A. Yeh, and Alexander G. Schwing. Not all unlabeled data are equal: Learning to weight data in semi-supervised learning. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "type": "text", + "content": "[41] Mamshad Nayeem Rizve, Navid Kardan, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Openldn: Learning to discover novel classes for open-world semi-supervised learning. In ECCV, pages 382-401. Springer, 2022. 1, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 471, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 536 + ], + "type": "text", + "content": "[42] Chuck Rosenberg, Martial Hebert, and Henry Schneiderman. Semi-supervised self-training of object detection models. In 7th IEEE Workshop on Applications of Computer Vision / IEEE Workshop on Motion and Video Computing (WACV/MOTION 2005), 5-7 January 2005, Breckenridge, CO, USA, pages 29-36. IEEE Computer Society, 2005. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 537, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 580 + ], + "type": "text", + "content": "[43] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In NeurIPS, pages 1163-1171, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "text", + "content": "[44] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "type": "text", + "content": "[45] Xin Sun, Zhenning Yang, Chi Zhang, Keck Voon Ling, and Guohao Peng. Conditional gaussian distribution learning for open set recognition. In CVPR, pages 13477-13486. Computer Vision Foundation / IEEE, 2020. 2, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[46] Yiyou Sun and Yixuan Li. Opencon: Open-world contrastive learning. Trans. Mach. Learn. Res., 2023, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[47] Kiat Chuan Tan, Yulong Liu, Barbara Ambrose, Melissa Tulig, and Serge J. Belongie. The herbarium challenge 2019 dataset. CoRR, abs/1906.05372, 2019. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[48] Christos Thrampoulidis, Ganesh Ramachandra Kini, Vala Vakilian, and Tina Behnia. Imbalance trouble: Revisiting neural-collapse geometry. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "[50] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. In CVPR, pages 7482-7491. IEEE, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "text", + "content": "[51] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In ICLR. OpenReview.net, 2022. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "text", + "content": "[52] Jiaheng Wei, Zhaowei Zhu, Hao Cheng, Tongliang Liu, Gang Niu, and Yang Liu. Learning with noisy labels revisited: A study using real-world human annotations. In ICLR. OpenReview.net, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 287, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 331 + ], + "type": "text", + "content": "[53] P. Welinder, S. Branson, T. Mita, C. Wah, F. Schroff, S. Belongie, and P. Perona. Caltech-UCSD Birds 200. Technical Report CNS-TR-2010-001, California Institute of Technology, 2010. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "type": "text", + "content": "[54] Liang Xie, Yibo Yang, Deng Cai, and Xiaofei He. Neural collapse inspired attraction-repulsion-balanced loss for imbalanced learning. Neurocomputing, 527:60-70, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 366, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 366, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 366, + 545, + 399 + ], + "type": "text", + "content": "[55] Qizhe Xie, Zihang Dai, Eduard H. Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. In NeurIPS, 2020. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 400, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 400, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 400, + 545, + 443 + ], + "type": "text", + "content": "[56] Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In ICML, pages 11525-11536. PMLR, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "text", + "content": "[57] Yibo Yang, Shixiang Chen, Xiangtai Li, Liang Xie, Zhouchen Lin, and Dacheng Tao. Inducing neural collapse in imbalanced learning: Do we really need a learnable classifier at the end of deep neural network? In NeurIPS, 2022. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "text", + "content": "[58] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip H. S. Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class-incremental learning. In ICLR. OpenReview.net, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "text", + "content": "[59] Bo Ye, Kai Gan, Tong Wei, and Min-Ling Zhang. Bridging the gap: Learning pace synchronization for open-world semi-supervised learning. CoRR, abs/2309.11930, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "type": "text", + "content": "[60] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In NeurIPS 2021, pages 18408-18419, 2021. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 635, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 668 + ], + "type": "text", + "content": "[61] Bingchen Zhao, Xin Wen, and Kai Han. Learning semi-supervised gaussian mixture models for generalized category discovery. In ICCV, pages 16623-16633, 2023. 5, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[62] Yunrui Zhao, Qianqian Xu, Yangbangyan Jiang, Peisong Wen, and Qingming Huang. Dist-pu: Positive-unlabeled learning from a label distribution perspective. In CVPR, pages 14441-14450. IEEE, 2022. 2, 5" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23081" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 240 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[63] Zhisheng Zhong, Jiequan Cui, Yibo Yang, Xiaoyang Wu, Xiaojuan Qi, Xiangyu Zhang, and Jiaya Jia. Understanding imbalanced semantic segmentation through neural collapse. In CVPR, pages 19550-19559. IEEE, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[64] Jinxin Zhou, Xiao Li, Tianyu Ding, Chong You, Qing Qu, and Zhihui Zhu. On the optimization landscape of neural collapse under MSE loss: Global optimality with unconstrained features. In ICML, pages 27179-27202. PMLR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "text", + "content": "[65] Zhihui Zhu, Tianyu Ding, Jinxin Zhou, Xiao Li, Chong You, Jeremias Sulam, and Qing Qu. A geometric analysis of neural collapse with unconstrained features. In NeurIPS, pages 29820-29834, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 240 + ], + "type": "text", + "content": "[66] Zhaowei Zhu, Tianyi Luo, and Yang Liu. The rich get richer: Disparate impact of semi-supervised learning. In ICLR. OpenReview.net, 2022. 2" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23082" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/ef2d4502-c3bc-43a6-a846-89e23e77cec3_content_list.json b/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/ef2d4502-c3bc-43a6-a846-89e23e77cec3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9ceceef8506ad12ca417a69ec826ca8510d34d89 --- /dev/null +++ b/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/ef2d4502-c3bc-43a6-a846-89e23e77cec3_content_list.json @@ -0,0 +1,1535 @@ +[ + { + "type": "text", + "text": "Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning", + "text_level": 1, + "bbox": [ + 78, + 130, + 893, + 155 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xialei Liu $^{2,1,*}$ Jiang-Tian Zhai $^{1,*}$ Andrew D. Bagdanov $^{3}$ Ke Li $^{4}$ Ming-Ming Cheng $^{2,1,\\dagger}$", + "bbox": [ + 91, + 178, + 883, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ VCIP, CS, Nankai University $^{2}$ NKIARI, Shenzhen Futian $^{3}$ MICC, University of Florence $^{4}$ Tencent Youtu Lab {xialei, cmm}@nankai.edu.cn, {jtzhai30,tristanli.sh}@gmail.com, andrew.bagdanov@unifi.it", + "bbox": [ + 91, + 200, + 861, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Exemplar-free Class Incremental Learning (EFCIL) aims to sequentially learn tasks with access only to data from the current one. EFCIL is of interest because it mitigates concerns about privacy and long-term storage of data, while at the same time alleviating the problem of catastrophic forgetting in incremental learning. In this work, we introduce task-adaptive saliency for EFCIL and propose a new framework, which we call Task-Adaptive Saliency Supervision (TASS), for mitigating the negative effects of saliency drift between different tasks. We first apply boundary-guided saliency to maintain task adaptivity and plasticity on model attention. Besides, we introduce task-agnostic low-level signals as auxiliary supervision to increase the stability of model attention. Finally, we introduce a module for injecting and recovering saliency noise to increase the robustness of saliency preservation. Our experiments demonstrate that our method can better preserve saliency maps across tasks and achieve state-of-the-art results on the CIFAR-100, Tiny-ImageNet, and ImageNet-Subset EFCIL benchmarks. Code is available at https://github.com/scok30/tass.", + "bbox": [ + 75, + 300, + 473, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 645, + 209, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks achieve state-of-the-art performance on many computer vision tasks. However, most of these tasks consider a static world in which tasks are well-defined, stationary, and all training data is available in a single training session. The real world consists of dynamically changing environments and data distributions, which - especially given the computational burden of training large CNNs - has led to renewed interest in learning new tasks incrementally while avoiding catastrophic forgetting [14, 34].", + "bbox": [ + 75, + 670, + 468, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Class Incremental Learning (CIL) [1, 33] is a scenario that considers the possibility of adding new classes to already-trained models. Most CIL methods rely on a memory buffer to store exemplars from past tasks [2, 10, 37, 42].", + "bbox": [ + 75, + 806, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd5dab2fb9d2796dc711f57546f48ac75b4224e5bfdec6811129787de84e050a.jpg", + "image_caption": [ + "Figure 1. We propose the TASS method, which can be directly applied to many recent exemplar-free class incremental learning methods, resulting in a significant improvement in EFCIL classification accuracy and a reduction in catastrophic forgetting." + ], + "image_footnote": [], + "bbox": [ + 501, + 267, + 890, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we consider Exemplar-Free Class Incremental Learning (EFCIL), which is a more challenging setting in which no data from previous tasks is retained. This is a realistic scenario and of great interest due to privacy concerns or restrictions on the long-term storage of data. The inability to retain examples from past tasks, however, significantly exacerbates the problem of catastrophic forgetting.", + "bbox": [ + 496, + 491, + 892, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are several recent works that consider the EFCIL problem. DeepInversion [45] inverts trained networks from random noise to generate images as exemplars and mixes them with current task samples for training. SDC [46] updates prototypes of each learned class by hypothesizing that semantic drift of classes from previous tasks can be approximated and estimated using new data. Other previous works propose representation learning methods for overcoming catastrophic forgetting [51, 52]. As pointed out in IL2A [51], learning better representations can reduce representation bias when transferred to new tasks. Incorporating self-supervised learning tasks, such as Barlow Twins [36] and rotation prediction [52], has also been proposed to achieve more stable representations and alleviate forgetting.", + "bbox": [ + 496, + 598, + 893, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CNNs naturally learn to attend to features that are discriminative for the tasks they are trained to solve. Catastrophic forgetting also occurs in EFCIL due to the model's attention to salient features drifting to features specific to the new task. Standard regularization approaches do little", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The first two authors contribute equally.", + "bbox": [ + 93, + 875, + 313, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23954", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to prevent this saliency drift when learning new tasks. One direct method of regularizing saliency is to apply distillation on saliency maps of old samples [11]. However, this is complicated by the inability to save samples from previous tasks in the EFCIL setting. Another method is to apply saliency distillation between current task samples and previous task attention [9]. This method however suffers from the semantic gap between current and old classes when enforcing saliency consistency.", + "bbox": [ + 75, + 90, + 468, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A lack of saliency regularization may lead to attention drifting toward the background in future tasks causing forgetting. Besides, simply applying distillation on attention [9] fails to offer plasticity and is susceptible to attention forgetting, which is a crucial factor of knowledge forgetting. In comparison, our Task-Adaptive Saliency Supervision (TASS) approach aims at keeping saliency focused on incrementally learned tasks (for more details, see Section 4), while maintaining its plasticity and stability. It improves many previous EFCIL methods with higher performance by supervising their attention, as shown in Figure 1.", + "bbox": [ + 75, + 229, + 468, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, TASS integrates three components to address this issue. Firstly, we apply dilated boundary maps to prevent saliency drift across object boundaries at intermediate layers. Since saliency drift typically occurs across tasks, encouraging the model to focus on significant foreground regions through dilated boundary supervision reduces the likelihood of saliency shifting toward the background, allowing the model to adaptively select attention areas within the foreground that are relevant to the task. Secondly, to simultaneously enhance the stability of the model's attention across tasks, we add a task-agnostic low-level auxiliary supervision task to the class-incremental framework, which is closely related to our core EFCIL task since image classification has been shown to help models localize the most salient areas within an image. Finally, we propose a module to inject saliency noise into certain feature channels and train the network to denoise them, helping the network further resist attention drift across tasks.", + "bbox": [ + 75, + 398, + 468, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this work are: (i) We provide new insight into task-adaptive saliency supervision under EFCIL settings; we also show the negative effect of methods with no or trivial saliency supervision, which illustrates the superiority of our method and motivates the need for saliency drift mitigation in EFCIL. (ii) We propose the Task-Adaptive Saliency Supervision (TASS) with three components that combine to mitigate the saliency drift problem. (iii) We show that TASS can be easily integrated into other state-of-the-art methods, such as MUC [28], IL2A [51], PASS [52], SSRE [53], leading to significant performance gains. (iv) Our experiments demonstrate that TASS outperforms all existing EFCIL methods and even several exemplar-based methods on the CIFAR-100, TinyImageNet, and ImageNet-Subset EFCIL benchmarks.", + "bbox": [ + 75, + 674, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first discuss previous work on incremental learning from the recent literature and then describe work on EFCIL.", + "bbox": [ + 498, + 114, + 890, + 143 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Incremental Learning", + "text_level": 1, + "bbox": [ + 500, + 152, + 705, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A variety of methods have been proposed for incremental learning in the past few years [1, 7]. Recent works can be coarsely grouped into three categories: replay-based, regularization-based, and parameter-isolation methods. Replay-based methods mitigate the task-recency bias by retaining training samples from previous tasks [37, 48]. In addition to replaying samples, BiC [42], PODNet [10], iCaRL [37] and LTCIL [27] apply a distillation loss to prevent forgetting and enhance model stability. GEM [30], AGEM [4], and MER [38] exploit past-task exemplars by modifying gradients on current training samples to match old samples. Rehearsal may cause models to overfit to stored samples. Regularization-based approaches such as LwF [23], EWC, R-EWC [18, 26], and DMC [49] offer ways to learn better representations while leaving enough plasticity for adaptation to new tasks. Parameter-isolation methods [32, 44] use models with different computational graphs for each task. With the help of growing models, new model branches mitigate catastrophic forgetting at the cost of more parameters and computational cost. It is also widely studied in other fields, such as semantic segmentation [3, 25, 43, 47] and object detection [13, 29].", + "bbox": [ + 496, + 175, + 890, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As for saliency-guided incremental learning, LwM [9] regularizes the saliency activations of previous classes on current task data. However, there exist semantic gaps between current classes and old ones, which results in an inaccurate distillation target for preserving saliency activations on old samples. RRR [11] directly saves the GradCAM saliency activations of each sample in a replay buffer and applies distillation to memorize this old knowledge, which requires storing additional samples during incremental learning. Despite these initial works on saliency-guided incremental learning, saliency drift still remains problematic and leads to catastrophic forgetting.", + "bbox": [ + 496, + 508, + 890, + 689 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Exemplar-free Class Incremental Learning", + "text_level": 1, + "bbox": [ + 500, + 696, + 864, + 713 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to conventional class incremental learning, exemplar-free class incremental learning is more appropriate for applications where training data is sensitive and may not be stored in perpetuity. DAFL [5] uses a GAN to generate synthetic samples from past tasks as an alternative to storing actual data. DeepInversion, which inverts trained networks using random noise to generate images, is another popular EFCIL method [45]. Always Be Dreaming further improves on DeepInversion for EFCIL [41]. SDC attempts to overcome the problems caused by semantic drift when training new tasks on old class samples [46]. It directly estimates prototypes of each learned class to use in", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23955", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bd01f5a77150e3636fa64febeed139f3cdf717582aa24e4555829237e5c64cf2.jpg", + "image_caption": [ + "Figure 2. Overall framework of Task-Adaptive Saliency Supervision (TASS). We apply a low-level model to generate saliency and boundary maps. The boundary map is dilated and downsampled to provide supervision at different stages of the encoder. A decoder is attached after the encoder for low-level distillation, which serves as stationary task-agnostic saliency guidance. To prevent saliency drift in later training phases, we introduce saliency noise into each encoder stage. The model is trained to denoise and reduce the saliency drift on current data in future phases. TASS can be integrated into an EFCIL approach to provide robust saliency guidance across incremental tasks." + ], + "image_footnote": [], + "bbox": [ + 166, + 92, + 810, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a nearest class mean classifier. PASS [52] and IL2A [51] are prototype-based replay methods for efficient and effective EFCIL. Since these prototypes are features computed from past training samples, image exemplars need not be retained. SSRE [53] introduces a re-parameterization method to trade-off between old and new knowledge. Our task-Adaptive Saliency Supervision (TASS) method uses three new components to reduce saliency drift in EFCIL and is complementary to several of the approaches mentioned above.", + "bbox": [ + 75, + 547, + 472, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Task-Adaptive Saliency Supervision", + "text_level": 1, + "bbox": [ + 76, + 715, + 401, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first define the Exemplar-Free Class Incremental Learning (EFCIL) scenario. Then we describe our TASS approach including dilated boundary supervision, auxiliary low-level supervision, and saliency noise injection. Our overall framework is illustrated in Figure 2.", + "bbox": [ + 75, + 742, + 468, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Exemplar-free Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 76, + 830, + 444, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Class-incremental learning aims to sequentially learn tasks consisting of disjoint classes of samples. Let $t \\in \\{1, 2, \\dots, T\\}$ denote the incremental learning tasks. The training data $D_{t}$", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for each task contains classes $C_t$ with $N_t$ training samples $\\{(x_t^i, y_t^i)\\}_{i=1}^{N_t}$ , where $x_t^i$ are images and $y_t^i \\in C_t$ are their labels.", + "bbox": [ + 498, + 547, + 890, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most deep networks applied to class-incremental learning can be split into two components: a feature extractor $F_{\\theta}$ and a common classifier $G_{\\phi}$ which grows with each new task $t + 1$ to include classes $C_{t + 1}$ . The feature extractor $F_{\\theta}$ first maps the input $x$ to a deep feature vector $z = F_{\\theta}(x) \\in \\mathbb{R}^{d}$ , and then the unified classifier $G_{\\phi}(z) \\in \\mathbb{R}^{|C_t|}$ is a probability distribution over classes $C_t$ used to make predictions on input $x$ .", + "bbox": [ + 496, + 593, + 892, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Class-incremental learning requires that the model be capable of correctly classifying all samples from previous tasks at any training task – that is, when learning task $t$ , the model must not forget how to classify samples from classes from tasks $t' < t$ . Exemplar-free class-incremental learning additionally restricts models to learn each new task without access to samples from previous ones. This typically leads to learning objectives that minimize a loss function $\\mathcal{L}$ defined on current training data $D_t$ :", + "bbox": [ + 496, + 713, + 893, + 849 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} ^ {\\mathrm {C I L}} (x, y) = \\mathcal {L} _ {\\mathrm {c e}} \\left(G _ {\\phi_ {t}} \\left(F _ {\\theta_ {t}} (x)\\right), y\\right) + \\mathcal {L} _ {t} ^ {\\mathrm {M}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 858, + 890, + 877 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{ce}}$ is the standard cross-entropy classification loss", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23956", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/64df8aa965c826eaa1f69e82985a9c539221e82d9ef4029bea46b31917fe3830.jpg", + "image_caption": [ + "Figure 3. We dilate the boundary map and apply a binary cross entropy loss at three stages in the CNN backbone to prevent mid-level attention from drifting into boundary regions." + ], + "image_footnote": [], + "bbox": [ + 86, + 93, + 467, + 170 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $\\mathcal{L}_t^{\\mathrm{M}}$ is a method-specific loss that mitigates forgetting during incremental learning. Note that without $\\mathcal{L}_t^{\\mathrm{M}}$ , Eq. 1 reduces to fine-tuning on task $t$ .", + "bbox": [ + 75, + 242, + 467, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Boundary-guided Mid-level Saliency Drift Regularization", + "text_level": 1, + "bbox": [ + 76, + 295, + 467, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Simply distilling attention between tasks does not consider task-adaptive attention. A model generates low-level representations of each input image $x$ (saliency and boundary maps in our experiments). We use CSNet [6] to generate salient regions and boundary maps, as it is lightweight and efficient, but any model producing saliency maps can be used in our framework (we explore other options in the Supplementary Material). To introduce plasticity on attention at these intermediate layers in the backbone, we use the generated boundary maps as a type of adaptive supervision as shown in Figure 3. We add a penalty term on object boundaries to avoid drift into the background. We first use 0.5 as the threshold to binarize the generated boundary map and dilate the maps by:", + "bbox": [ + 75, + 333, + 467, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nB _ {d} (x) = \\operatorname {D i l a t e} \\left(A _ {b} (x), d\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 551, + 467, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $A_{b}(x)$ is the generated boundary map of image $x$ , which is converted from the saliency map with a Laplacian filter, and $d$ denotes the dilation radius applied on the boundary map for controlling the strictness of boundary-guided saliency.", + "bbox": [ + 75, + 575, + 467, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rather than use a decoder at each layer as done for the low-level saliency maps described above, the mid-level saliency maps of our model are generated using Grad-CAM [39] at three stages of the CNN backbone (see Figure 2). We also experiment with several other methods for generating student saliency maps and report on them in the Supplementary Material. The dilated generated boundary map $B_{d}(x)$ is downsampled to match the feature map dimensions at these three stages in order to compare with the Grad-CAM generated saliency boundary maps. We use the binary cross entropy loss for supervision on dilated boundary regions. This loss is defined as:", + "bbox": [ + 75, + 651, + 467, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} ^ {\\mathrm {d b s}} (x) = - \\frac {\\sum_ {j = 1} ^ {N} B _ {d} (x , j) \\log (1 - S (x , j))}{\\sum_ {j = 1} ^ {N} B _ {d} (x , j)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 837, + 467, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S(x, j)$ denotes the saliency map of our model on image $x$ at pixel $j$ , $B_d(x, j)$ is the dilated generated boundary map at pixel $j$ , and $N$ is the number of pixels in $x$ . We compute this loss only within dilated boundary regions (i.e. where $B_d(x, j) = 1$ ). This loss helps the student saliency map avoid intersecting with the dilated teacher boundary region.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Auxiliary Low-level Supervisions for EFCIL", + "text_level": 1, + "bbox": [ + 498, + 214, + 874, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose to learn stable features from low-level stationary tasks shared across all incremental classification tasks during class-incremental learning. Low-level vision tasks like salient object detection require useful representations of input images. By learning these feature representations across tasks, the model can focus on key areas of input images and exploit learned, stable features with less representation drift since the low-level features change very little between tasks.", + "bbox": [ + 496, + 239, + 890, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Saliency map prediction is relevant to image classification since the foreground largely determines the results, while the background is comparatively less important. When learning new tasks with new classes, the background of images of new classes may contain new visual concepts that introduce undesirable noise and lead to forgetting of essential previous knowledge. The effectiveness of saliency features for learning classification tasks was demonstrated by Saliency Guided Training [17]. Additional supervision of salient region boundaries can aid salient object detection tasks for both segmentation and localization [12, 21, 24, 31, 50]. The positive interaction between these two tasks brings richer attention to features relevant to the main classification task. It can provide positive guidance in the form of stationary knowledge across class-incremental tasks. Some examples are illustrated in Figure 5.", + "bbox": [ + 496, + 378, + 890, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We incorporate low-level vision tasks into the network as an auxiliary supervision for enriching task-agnostic attention. The boundary map is computed with a Laplacian filter over the estimated saliency map. We add a decoder $D_{\\psi}$ [22] after the backbone $F_{\\theta}$ to predict low-level saliency and boundary maps for input images. The average L2 distance between the prediction and target is used as a low-level saliency distillation loss:", + "bbox": [ + 496, + 638, + 890, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} ^ {\\mathrm {l m s}} (x) = \\frac {| | D _ {\\psi} (F _ {\\theta} (x)) - A (x) | | _ {2}}{\\sqrt {N}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 773, + 890, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $A(x)$ denotes the target low-level maps on input $x$ , consisting of a saliency map $A_{s}(x)$ and a boundary map $A_{b}(x)$ . $D_{\\psi}(F_{\\theta}(x))$ are combined saliency and boundary maps produced by the decoder, and $N$ is the number of pixels in the saliency maps.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1https://github.com/jacobgil/pytorch-grad-cam", + "bbox": [ + 94, + 886, + 408, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23957", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 TASS Pseudocode" + ], + "code_body": "Input: The number of tasks $T$ , training samples $D_{t} = \\{(x_{i},y_{i})\\}_{i=1}^{N}$ of task $t$ , initial parameters $\\Theta^0 = \\{\\theta_0,\\phi_0,\\psi_0\\}$ containing parameters of feature extractor $F_{\\theta}$ , classifier $G_{\\phi}$ , and low-level decoder $D_{\\psi}$ . \nOutput: Model $\\Theta^T$ \n1: for $t \\in \\{1,2,\\dots,T\\}$ do \n2: $\\Theta^t \\gets \\Theta^{t-1}$ \n3: while not converged do \n4: Sample $(x,y)$ from $D_{t}$ \n5: $\\mathcal{L}_t^{\\mathrm{CIL}} \\gets \\text{SaliencyNoiseInjection}(x,y)$ \n6: $\\mathcal{L}_t^{\\mathrm{Ims}} \\gets \\text{LowLevelMultitask}(x,A(x))$ \n7: $S \\gets \\text{ComputeGradCAMSaliency}(x,y)$ \n8: $\\mathcal{L}_t^{\\mathrm{obs}} \\gets \\text{DilatedBoundarySupervision}(S,A(x))$ \n9: update $\\Theta^t$ by minimizing $\\mathcal{L}_t^{\\mathrm{all}}$ from Eq. 5 \n10: end while \n11: end for", + "bbox": [ + 76, + 109, + 468, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Saliency Noise Injection", + "text_level": 1, + "bbox": [ + 76, + 375, + 295, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although we apply low-level distillation and dilated boundary supervision to maintain saliency representations across tasks, the model can still forget saliency on samples from previous tasks. To address this, we force the model to be able to recover correct saliency maps from injected saliency noise.", + "bbox": [ + 75, + 398, + 468, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At each task there is no available training data from previous or future tasks, and therefore we cannot directly know saliency drift on these samples. Instead of supervising the model with ground-truth saliency drift signals, we introduce saliency noise on random feature channels. We use a random ellipse to approximate the potential saliency drift in future tasks and the model is trained to denoise within each stage. Therefore the model can learn to effectively reduce real saliency drift.", + "bbox": [ + 75, + 489, + 468, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We generate elliptical noise using a very simple approach. There are six parameter dimensions: the center coordinate $(x, y)$ , the major and minor axis lengths $(a, b)$ , the rotation angle $\\alpha$ , and the mask weight $w$ . A detailed explanation of this process is given in the Supplementary Material. With the help of dilated boundary supervision, each stage learns to eliminate this additional saliency noise and this aids generalization for future tasks and mitigates saliency forgetting in previous ones.", + "bbox": [ + 75, + 625, + 468, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Learning Objective and Training Algorithm", + "text_level": 1, + "bbox": [ + 76, + 768, + 449, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall learning objective combines the low-level multi-task learning, dilated boundary supervision, and random saliency noise injection modules:", + "bbox": [ + 75, + 792, + 468, + 837 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} ^ {\\text {a l l}} = \\mathcal {L} _ {t} ^ {\\text {C I L}} + \\mathcal {L} _ {t} ^ {\\text {l m s}} + \\mathcal {L} _ {t} ^ {\\text {d b s}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 843, + 468, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparing this loss with Eq. 1, for TASS $\\mathcal{L}_t^{\\mathrm{M}} = \\mathcal{L}_t^{\\mathrm{CIL}} + \\mathcal{L}_t^{\\mathrm{Ins}}$ , thus incorporating saliency-aware supervision with", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the cross-entropy loss. The entire process is detailed in Algorithm 1.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 498, + 137, + 702, + 154 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section we first describe our experimental setup and then compare TASS to state-of-the-art methods on several EFCIL benchmarks. In Section 4.3 we give further analysis of the various components of TASS.", + "bbox": [ + 496, + 162, + 890, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 498, + 234, + 689, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow standard experimental protocols for EFCIL on three benchmark datasets.", + "bbox": [ + 496, + 258, + 890, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We perform experiments on CIFAR-100 [19], Tiny-ImageNet [20], and ImageNet-Subset [8]. For most experiments, we train the model on half of the classes for the first task, and then equally distribute the remaining classes across each of the subsequent tasks. The convention we use is: $F + C \\times T$ means that the first task contains $F$ classes, and the next $T$ tasks each contain $C$ classes. This is a common configuration for EFCIL used in both PASS [52] and SSRE [53].", + "bbox": [ + 496, + 290, + 890, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "State-of-the-art methods. Since we focus on EFCIL, we mainly compare with exemplar-free state-of-the-art approaches: SSRE [53], PASS [52], IL2A [51], EWC [18], LwF-MC [37], and MUC [28]. To demonstrate the effectiveness of our method, we also compare its performance with several exemplar-based methods like iCaRL (both nearest-mean and CNN) [37], EEIL [2], and LUCIR [16]. We also compare with RRR [11] integrated with SSRE, which focuses on preserving saliency using exemplar replay.", + "bbox": [ + 496, + 426, + 890, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details and metrics. We use ResNet-18 [15] as a feature extraction backbone. This is the same base network used in SSRE [53] and PASS [52], two state-of-the-art EFCIL approaches. We use the decoder in [22] to estimate low-level student saliency maps. All experiments are trained from scratch using Adam for 100 epochs with an initial learning rate 0.001. The learning rate is reduced by a factor 10 at epochs 45 and 90. For exemplar-based approaches, we use herding [37] to select and store 20 samples per class following common settings [16, 37]. We implement RRR [11] with SSRE to fairly compare it with TASS. For dilated boundary supervision, we set $d$ of the three mid-level boundary dilation stages to be $5\\%$ , $10\\%$ and $15\\%$ of the image size.", + "bbox": [ + 496, + 579, + 890, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We report three common metrics for class incremental learning: the average and last top-1 accuracy, as well as average forgetting for all classes learned up to task $t$ . Denoting by $Acc_{i}$ the accuracy over all learned classes up to and including task $i$ , the average accuracy is defined as $Avg = \\frac{\\sum_{i=1}^{T} Acc_{i}}{T}$ , and the last accuracy is $Acc_{T}$ . Letting $a_{m,n}$ denotes the accuracy of task $n$ after learning task $m$ ,", + "bbox": [ + 496, + 791, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23958", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d2236872246007ece57198f19ec99914f05881bf6337ff2b434f07552ee305fc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetCIFAR-100Tiny-ImageNet
Method5 tasks10 tasks20 tasks5 tasks10 tasks20 tasks
MUC38.4519.5715.6518.9515.479.14
+TASS49.17 (+10.72)40.34 (+20.77)37.86 (+22.21)32.47 (+13.46)30.13 (+14.66)27.70 (+18.56)
IL2A55.1345.3245.2436.7734.5328.68
+TASS58.74 (+3.61)53.24 (+7.92)53.07 (+7.83)42.49 (+5.72)41.34 (+6.81)40.59 (+11.91)
PASS55.6749.0348.4841.5839.2832.78
+TASS59.10 (+3.43)54.45 (+5.42)52.37 (+3.89)44.05 (+2.47)43.06 (+3.78)42.57 (+9.79)
SSRE56.3355.0150.4741.4540.0739.25
+TASS59.26 (+2.93)57.93 (+2.92)53.78 (+3.31)44.13 (+2.68)43.86 (+3.79)43.55 (+4.30)
", + "bbox": [ + 169, + 88, + 799, + 242 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/da3b221466ebca517a18a97dc9959b6368b8ae3525006ec34fd60d28900aa8dc.jpg", + "table_caption": [ + "Table 1. Performance gain in top-1 accuracy by applying TASS to other EFCIL methods in a plug-and-play way. Absolute gains are indicated in (red)." + ], + "table_footnote": [], + "table_body": "
DatasetCIFAR100TinyImageNet
Setting5 tasks10 tasks20 tasks5 tasks10 tasks20 tasks
MethodAvg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓
E=20iCaRL-CNN†51.0740.1242.1348.6639.6545.6944.4335.4743.5434.6422.3136.8931.1521.1036.7027.9020.4645.12
iCaRL-NCM†58.5649.7424.9054.1945.1328.3250.5140.6835.5345.8633.4527.1543.2933.7528.8938.0428.8937.40
LUCIR†63.7855.0621.0062.3950.1425.1259.0748.7828.6549.1537.0920.6148.5236.8022.2542.8332.5533.74
EEIL†60.3752.3523.3656.0547.6726.6552.3441.5932.4047.1234.2425.5645.0134.2625.9140.5030.1435.04
RRR†66.4357.2218.0565.7855.7418.5962.4351.3518.4051.2042.2316.6749.5440.1221.6447.4635.5429.10
E=20LwFMC45.9336.1744.2327.4350.4717.0420.0715.8855.4629.1217.1254.2623.1012.3354.3717.438.7563.54
EWC16.049.3260.1714.708.4762.5314.128.2363.8918.8012.7167.5515.7710.1270.2312.398.4275.54
MUC49.4238.4540.2830.1919.5747.5621.2715.6552.6532.5817.9851.4626.6114.5450.2121.9512.7058.00
IL2A63.2255.1323.7857.6545.3230.4154.9045.2430.8448.1736.1425.4342.1035.2328.3236.7928.7435.46
PASS63.4755.6725.2061.8449.0330.2558.0948.4830.6149.5541.5818.0447.2939.2823.1142.0732.7830.55
SSRE65.8856.3318.3765.0455.0119.4861.7050.4718.3750.3941.6717.2548.9339.8922.5048.1739.7626.74
TASS (Ours)68.7559.2616.4267.4257.9317.7862.7653.7817.7855.1244.1315.4054.2143.8618.4752.7943.5522.51
", + "bbox": [ + 81, + 294, + 888, + 502 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/24dd2ee15a58623054fbaecef4e9a1454c6a855330e260c0244d3a36b3dafc57.jpg", + "table_caption": [ + "Table 2. Average, last top-1 accuracy, and forgetting on CIFAR-100 with different numbers of tasks. Replay-based methods storing 20 exemplars from each previous class are indicate with †. The best overall results are in bold. We run all experiments three times and report the mean for all metrics." + ], + "table_footnote": [], + "table_body": "
DatasetImageNet-Subset
Setting5 tasks10 tasks20 tasks
MethodAvg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓
LwF_TC34.8624.1049.3631.1820.0153.0427.5417.4256.07
MUC40.6527.8947.1335.0722.6552.1031.4420.1253.85
PASS63.1252.6122.4761.8050.4423.5755.2346.0726.73
SSRE69.5458.4617.2267.6957.5118.6061.2350.0523.22
TASS (Ours)74.3263.1414.3772.6057.9316.0968.7957.6018.41
", + "bbox": [ + 81, + 563, + 483, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Average, last top-1 accuracy, and forgetting on ImageNet-Subset for different numbers of tasks. We run all experiments three times and report means for all metrics.", + "bbox": [ + 75, + 679, + 470, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the forgetting measure $f_{k}^{i}$ of task $i$ after learning task $k$ is computed as $f_{k}^{i} = \\max_{t\\in 1,2,\\ldots ,k - 1}(a_{t,i} - a_{k,i})$ . The average forgetting $F_{k}$ is defined as $F_{k} = \\frac{1}{k - 1}\\sum_{i = 1}^{k - 1}f_{k}^{i}$ .", + "bbox": [ + 75, + 751, + 468, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with the State-of-the-art", + "text_level": 1, + "bbox": [ + 76, + 814, + 398, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare TASS with the state-of-the-art on CIFAR-100 in Table 2. TASS outperforms all exemplar-free approaches. For exemplar-based methods like iCaRL [37], EEIL [2], and LUCIR [16], our method still has significantly better", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "performance. On longer sequences (i.e. 10 and 20 tasks), our method significantly reduces forgetting when learning new classes compared to other EFCIL methods. TASS outperforms the best method SSRE by about $3\\%$ on the last task. This performance improvement can be also observed in terms of average forgetting.", + "bbox": [ + 496, + 566, + 890, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As we see in Table 3 and Figure 4 for Tiny-ImageNet and ImageNet-Subset, although our method has similar top-1 accuracy on the first task in Figure 4, it has better performance at most intermediate tasks and also the final one. For longer sequences in Figure 4, the gap between our method and the best baseline is largely consistent, showing the effectiveness of our method at mitigating forgetting. The performance gain in Table 3 is larger on Tiny-ImageNet and ImageNet-Subset compared to CIFAR100, and this demonstrates that our method generalizes to datasets with larger images and object scales. It is worth mentioning that TASS also produces results with smaller variance. We believe this to be due to TASS reducing saliency drift to background regions, which may include random noise.", + "bbox": [ + 496, + 657, + 892, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Plug-and-play with other EFCIL methods. Some existing EFCIL methods, like PASS [52], IL2A [51] and", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23959", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2786bca8e8231d2a3146062959cbdc48efc3ddaf9cc1549f1e2f25be0b9c2f1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 95, + 352, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c79aabdedbb864260f53e8611d4a22a80a3622f33ba36d1e3dc940a1bb95f0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 95, + 602, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/caf86358f27db210f3dd0db47cff36829e498d5eb40ded3758f7de87657ec83d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 95, + 846, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e101741421122b30b8d6925fd5d3306b006c8a595aec14bbcf7e44176725f797.jpg", + "image_caption": [ + "Figure 4. Results on Tiny-ImageNet and ImageNet-Subset for different numbers of tasks. Our method outperforms others, especially on longer task sequences (i.e. more, but smaller, tasks)." + ], + "image_footnote": [], + "bbox": [ + 116, + 233, + 352, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/873b053d588788e54782dde677d13d6f5aa5f43bcda3b49cd9e13fc67e69ab18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 233, + 599, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/420c6ef776df5e91f24ac5dd6ba74001c0d9d6279dde2fe2db7aa838ef0b2432.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 234, + 846, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SSRE [53], focus on reducing forgetting via embedding regularization. Considering the importance of saliency to image classification, it is natural to consider whether TASS can be integrated into these methods. The results in Table 1 show the performance gain brought by this integration. Adding TASS doubles the performance for MUC in many cases and significantly improves IL2A and PASS. When we incorporate it into the best baseline SSRE, it yields a consistent gain of about $3\\%$ . These results clearly show that TASS, by explicitly mitigating saliency drift, is complementary to other methods in relieving forgetting. They additionally demonstrate the significance of saliency drift as a cause of catastrophic forgetting in EFCIL.", + "bbox": [ + 75, + 422, + 472, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Additional Analysis", + "text_level": 1, + "bbox": [ + 76, + 632, + 266, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section we take a deeper look at the method we propose. If not specified, the results are produced using TASS integrated into SSRE [53].", + "bbox": [ + 75, + 656, + 468, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study. We performed ablations using the 10-task setting on CIFAR-100 (see Table 4). We ablate on both PASS [52] and SSRE [53]. Low-level multi-task supervision is crucial and improves by $2.2\\%$ (PASS) and $1.2\\%$ (SSRE). Dilated boundary supervision further boosts performance by about $1 - 2\\%$ . Saliency noise injection is also helpful for both methods and improves by $1.5\\%$ for PASS. In total, TASS improves baselines by $5.5\\%$ and $2.9\\%$ points, respectively. Note that SSRE is the previous state-of-the-art method and TASS outperforms it by a large margin.", + "bbox": [ + 75, + 702, + 468, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Low-level Multi-task Supervision. To analyze the effect of our proposed low-level saliency supervision, we performed experiments on ImageNet-Subset in the 5 and 10", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5934431c6f6d9ae66447ff3504b4ae8e29656772529af4e56389f12b3bbe6419.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method & TasksL1msL1bsSNIAccuracy
Baseline (PASS)49.0
Variants✓51.2
✓✓53.0
✓✓✓54.5
Baseline (SSRE)55.0
Variants✓56.2
✓✓57.3
✓✓✓57.9
", + "bbox": [ + 550, + 420, + 844, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablations on each TASS component. Experiments are on CIFAR-100 in the 10-task setting and we report the top-1 accuracy in % for TASS integrated into PASS and SSRE. $\\mathcal{L}_{\\mathrm{ubs}}$ (Eq. 3), $\\mathcal{L}_{\\mathrm{lms}}$ (Eq. 4), and SNI denote the three TASS components: Dilated Boundary Supervision, Low-level Multi-task Supervision, and Saliency Noise Injection.", + "bbox": [ + 498, + 597, + 890, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "task settings. We first plot the loss across tasks in Figure 5(c). After learning to predict boundary and saliency maps in the first task, the network maintains good performance for the rest of the 5 task sequence. This shows that the low-level tasks are stable during continual learning. Furthermore, we visualize the results of saliency and boundary map prediction during incremental learning in Figure 5(a-b). We give some examples of predicted boundary and saliency maps after learning different tasks. Although CIL involves samples of different classes, we see that the low-level outputs are relatively stable and class agnostic. Since the model is able to stably predict these low-level features across tasks, it therefore can preserve useful prior knowl", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23960", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1476576f2e990c0f32b9db0382198a45131e1d2325330f4a31b6fe8934e42324.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 91, + 90, + 356, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d28aeb54d478b5f5b10363ffd05810ec4f00605d06636e532b5bd843197fbe54.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 367, + 90, + 632, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7a8e433045b1b2e3c350917ccc79574eaa71dcb785bece9d4d91380e06f7038a.jpg", + "image_caption": [ + "(c)", + "Figure 5. Visualization of the saliency (a) and boundary (b) maps from our student encoder-decoder network with original images from different tasks at different stages of incremental learning. Our method produces stable low-level results while reducing forgetting in classification. (c) The MAE loss between the student and teacher network across tasks." + ], + "image_footnote": [], + "bbox": [ + 640, + 90, + 880, + 265 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "edge for continual learning.", + "bbox": [ + 76, + 356, + 261, + 371 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d80d5d29036d1a6bed5d6ec7b3252daf6d1d55bbfdf98e5431c69b34af60b49c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Metric & MethodAvg↑Last↑
FeTrIL [35]65.2056.34
SOPE [54]65.8456.80
PRAKA [40]68.8659.20
TASS (SSRE)67.4257.93
TASS (PRAKA)69.7060.04
", + "bbox": [ + 166, + 396, + 380, + 494 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/88971ce58b618a4287ff3abb6cd393ddd2a73e4b64abf8e17e7cd64837d0ff9f.jpg", + "table_caption": [ + "Table 5. Average and last accuracy on CIFAR-100 10-task setting." + ], + "table_footnote": [], + "table_body": "
Metric & MethodAvg↑
SOPE [54]60.20
FeTrIL [35]65.00
TASS (FeTrIL)66.03
", + "bbox": [ + 187, + 556, + 359, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TASS with more methods and benchmarks. Due to the strong generalization capability of our method, we have also applied our TASS paradigm to PRAKA [40], resulting in further performance improvements. We give specific experimental comparisons in Table 5. We also compare with FeTrIL [35] in Table 5 and Table 6 above. Experiments on ImageNet-Full show consistent improvement in Table 6 above.", + "bbox": [ + 75, + 666, + 468, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative analysis of saliency drift. We measure the intersection over union (IoU) between the self-attention map of the last layer in DINO and the Grad-CAM saliency maps for both SSRE and $\\mathrm{SSRE + TASS}$ . As shown in Table 7, TASS significantly reduces saliency drift in the student model, which again shows the effectiveness of our method for maintaining saliency during CIL.", + "bbox": [ + 75, + 794, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f288d85502710137669e6c3b1c41e84abc590e6a27833de14c27a4d0c2858362.jpg", + "table_caption": [ + "Table 6. Average accuracy on the ImageNet-Full 10-task setting." + ], + "table_footnote": [], + "table_body": "
Task14710
SSRE47.450.156.678.5
SSRE+TASS75.282.388.590.1
", + "bbox": [ + 553, + 352, + 838, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Average IoU (\\%) between DINO self-attention maps and student model saliency in the CIFAR-100 10-task.", + "bbox": [ + 500, + 409, + 890, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 464, + 625, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose an approach of task-adaptive saliency guidance for EFCIL. The insight behind TASS is to guide the model to focus on salient regions and inhibit saliency drift. We show that robust saliency guidance is crucial to mitigating forgetting across tasks. Experiments demonstrate that TASS is effective and surpasses the state-of-the-art. TASS can be easily combined with other methods, leading to large performance gains over baselines. Qualitative results also show that low-level tasks are stable across different tasks, resulting in less forgetting.", + "bbox": [ + 496, + 496, + 890, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. In this work, we introduce the auxiliary saliency knowledge across tasks for better CIL performance, which may raise the concern of unfair comparison. However, we believe it is important to explore external knowledge to make the CIL system applicable in real applications. Additionally, our method can be easily integrated with other methods and the low-level auxiliary knowledge takes negligible cost and is easy to obtain. We will explore leveraging other forms of knowledge in future work, such as pre-trained visual and large language models.", + "bbox": [ + 496, + 652, + 890, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work is funded by NSFC (NO. 62206135, 62225604), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), and the Fundamental Research Funds for the Central Universities (Nankai University, 070-63233085). Computation is supported by the Supercomputing Center of Nankai University.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23961", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eden Belouadah, Adrian Popescu, and Ioannis Kanellos. A comprehensive study of class incremental learning algorithms for visual tasks. Neural Networks, 135:38-54, 2021. 1, 2", + "[2] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In ECCV, 2018. 1, 5, 6", + "[3] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9233-9242, 2020. 2", + "[4] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. In ICLR, 2019. 2", + "[5] Hanting Chen, Yunhe Wang, Chang Xu, Zhaohui Yang, Chuanjian Liu, Boxin Shi, Chunjing Xu, Chao Xu, and Qi Tian. Data-free learning of student networks. In ICCV, 2019. 2", + "[6] Ming-Ming Cheng, Shanghua Gao, Ali Borji, Yong-Qiang Tan, Zheng Lin, and Meng Wang. A highly efficient model to study the semantics of salient object detection. IEEE TPAMI, 2021. 4", + "[7] Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. IEEE TPAMI, 2021. 2", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5", + "[9] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In CVPR, 2019. 2", + "[10] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020. 1, 2", + "[11] Sayna Ebrahimi, Suzanne Petryk, Akash Gokul, William Gan, Joseph E. Gonzalez, Marcus Rohrbach, and trevor darrell. Remembering for the right reasons: Explanations reduce catastrophic forgetting. In ICLR, 2021. 2, 5", + "[12] Deng-Ping Fan, Ge-Peng Ji, Peng Xu, Ming-Ming Cheng, Christos Sakaridis, and Luc Van Gool. Advances in deep concealed scene understanding. Visual Intelligence, 1(1):16, 2023. 4", + "[13] Tao Feng, Mang Wang, and Hangjie Yuan. Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9427-9436, 2022. 2", + "[14] Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv preprint arXiv:1312.6211, 2013. 1", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun." + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Deep residual learning for image recognition. In CVPR, 2016.5", + "[16] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 5, 6", + "[17] Aya Abdelsalam Ismail, Hector Corrada Bravo, and Soheil Feizi. Improving deep learning interpretability by saliency guided training. NeurIPS, 2021. 4", + "[18] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 2017. 2, 5", + "[19] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, 2009. 5", + "[20] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 5", + "[21] Hao Li, Dingwen Zhang, Nian Liu, Lechao Cheng, Yalun Dai, Chao Zhang, Xinggang Wang, and Junwei Han. Boosting low-data instance segmentation by unsupervised pretraining with saliency prompt. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15485-15494, 2023. 4", + "[22] Xiangtai Li, Ansheng You, Zhen Zhu, Houlong Zhao, Maoke Yang, Kuiyuan Yang, and Yunhai Tong. Semantic flow for fast and accurate scene parsing. In ECCV, 2020. 4, 5", + "[23] Zhizhong Li and Derek Hoiem. Learning without forgetting. In ECCV, 2016. 2", + "[24] Liqiang Lin, Pengdi Huang, Chi-Wing Fu, Kai Xu, Hao Zhang, and Hui Huang. On learning the right attention point for feature enhancement. Science China Information Sciences, 66(1):112107, 2023. 4", + "[25] Zheng Lin, Zhao Zhang, Zi-Yue Zhu, Deng-Ping Fan, and Xia-Lei Liu. Sequential interactive image segmentation. Computational Visual Media, 9(4):753-765, 2023. 2", + "[26] Xialei Liu, Marc Masana, Luis Herranz, Joost Van de Weijer, Antonio M Lopez, and Andrew D Bagdanov. Rotate your networks: Better weight consolidation and less catastrophic forgetting. In ICPR, 2018. 2", + "[27] Xialei Liu, Yu-Song Hu, Xu-Sheng Cao, Andrew D Bagdanov, Ke Li, and Ming-Ming Cheng. Long-tailed class incremental learning. In European Conference on Computer Vision, pages 495-512. Springer, 2022. 2", + "[28] Yu Liu, Sarah Parisot, Gregory Slabaugh, Xu Jia, Ales Leonardis, and Tinne Tuytelaars. More classifiers, less forgetting: A generic multi-classifier paradigm for incremental learning. In ECCV. Springer, 2020. 2, 5", + "[29] Yuyang Liu, Yang Cong, Dipam Goswami, Xialei Liu, and Joost van de Weijer. Augmented box replay: Overcoming foreground shift for incremental object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11367-11377, 2023. 2", + "[30] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. NeurIPS, 2017. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23962", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Zhouzhou Ma, Guanghua Gu, and Wenrui Zhao. Self-attention guidance based crowd localization and counting. Machine Intelligence Research, pages 1–17, 2024. 4", + "[32] Arun Mallya and Svetlana Lazebnik. Packet: Adding multiple tasks to a single network by iterative pruning. In CVPR, 2018. 2", + "[33] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost van de Weijer. Class incremental learning: survey and performance evaluation on image classification. IEEE TPAMI, 2022. 1", + "[34] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of learning and motivation*, pages 109–165. Elsevier, 1989. 1", + "[35] Grégoire Petit, Adrian Popescu, Hugo Schindler, David Picard, and Bertrand Delezoise. Fetril: Feature translation for exemplar-free class-incremental learning. In WACV, 2023. 8", + "[36] Quang Pham, Chenghao Liu, and Steven Hoi. Dualnet: Continual learning, fast and slow. NeurIPS, 2021. 1", + "[37] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017. 1, 2, 5, 6", + "[38] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. In ICLR, 2019. 2", + "[39] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 4", + "[40] Wuxuan Shi and Mang Ye. Prototype reminiscence and augmented asymmetric knowledge aggregation for nonexemplar class-incremental learning. In ICCV, 2023. 8", + "[41] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In ICCV, 2021. 2", + "[42] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019. 1, 2", + "[43] Jia-Wen Xiao, Chang-Bin Zhang, Jiekang Feng, Xialei Liu, Joost van de Weijer, and Ming-Ming Cheng. Endpoints weight fusion for class incremental semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7204-7213, 2023. 2", + "[44] Ju Xu and Zhanxing Zhu. Reinforced continual learning. NeurIPS, 2018. 2", + "[45] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In CVPR, 2020. 1, 2", + "[46] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In CVPR, 2020. 1, 2" + ], + "bbox": [ + 78, + 90, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Lu Yu, Xialei Liu, and Joost Van de Weijer. Self-training for class-incremental semantic segmentation. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2", + "[48] Jiang-Tian Zhai, Xialei Liu, Andrew D Bagdanov, Ke Li, and Ming-Ming Cheng. Masked autoencoders are efficient class incremental learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19104-19113, 2023. 2", + "[49] Junting Zhang, Jie Zhang, Shalini Ghosh, Dawei Li, Serafettin Tasci, Larry Heck, Heming Zhang, and C-C Jay Kuo. Class-incremental learning via deep model consolidation. In WACV, 2020. 2", + "[50] Jia-Xing Zhao, Jiang-Jiang Liu, Deng-Ping Fan, Yang Cao, Jufeng Yang, and Ming-Ming Cheng. Egnet: Edge guidance network for salient object detection. In ICCV, 2019. 4", + "[51] Fei Zhu, Zhen Cheng, Xu-yao Zhang, and Cheng-lin Liu. Class-incremental learning via dual augmentation. NeurIPS, 2021. 1, 2, 3, 5, 6", + "[52] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In CVPR, 2021. 1, 2, 3, 5, 6, 7", + "[53] Kai Zhu, Wei Zhai, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Self-sustaining representation expansion for nonexemplar class-incremental learning. In CVPR, 2022. 2, 3, 5, 7", + "[54] Kai Zhu, Kecheng Zheng, Ruili Feng, Deli Zhao, Yang Cao, and Zheng-Jun Zha. Self-organizing pathway expansion for non-exemplar class-incremental learning. In ICCV, 2023. 8" + ], + "bbox": [ + 501, + 92, + 890, + 488 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23963", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/ef2d4502-c3bc-43a6-a846-89e23e77cec3_model.json b/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/ef2d4502-c3bc-43a6-a846-89e23e77cec3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4f642ce19d3e4caa52d2fe4c83b3e6dae1d2c762 --- /dev/null +++ b/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/ef2d4502-c3bc-43a6-a846-89e23e77cec3_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.131, + 0.895, + 0.156 + ], + "angle": 0, + "content": "Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.179, + 0.884, + 0.201 + ], + "angle": 0, + "content": "Xialei Liu \\(^{2,1,*}\\) Jiang-Tian Zhai \\(^{1,*}\\) Andrew D. Bagdanov \\(^{3}\\) Ke Li \\(^{4}\\) Ming-Ming Cheng \\(^{2,1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.201, + 0.862, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)VCIP, CS, Nankai University \\(^{2}\\)NKIARI, Shenzhen Futian \\(^{3}\\)MICC, University of Florence \\(^{4}\\)Tencent Youtu Lab {xialei, cmm}@nankai.edu.cn, {jtzhai30,tristanli.sh}@gmail.com, andrew.bagdanov@unifi.it" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.619 + ], + "angle": 0, + "content": "Exemplar-free Class Incremental Learning (EFCIL) aims to sequentially learn tasks with access only to data from the current one. EFCIL is of interest because it mitigates concerns about privacy and long-term storage of data, while at the same time alleviating the problem of catastrophic forgetting in incremental learning. In this work, we introduce task-adaptive saliency for EFCIL and propose a new framework, which we call Task-Adaptive Saliency Supervision (TASS), for mitigating the negative effects of saliency drift between different tasks. We first apply boundary-guided saliency to maintain task adaptivity and plasticity on model attention. Besides, we introduce task-agnostic low-level signals as auxiliary supervision to increase the stability of model attention. Finally, we introduce a module for injecting and recovering saliency noise to increase the robustness of saliency preservation. Our experiments demonstrate that our method can better preserve saliency maps across tasks and achieve state-of-the-art results on the CIFAR-100, Tiny-ImageNet, and ImageNet-Subset EFCIL benchmarks. Code is available at https://github.com/scok30/tass." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.646, + 0.21, + 0.661 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.807 + ], + "angle": 0, + "content": "Deep neural networks achieve state-of-the-art performance on many computer vision tasks. However, most of these tasks consider a static world in which tasks are well-defined, stationary, and all training data is available in a single training session. The real world consists of dynamically changing environments and data distributions, which - especially given the computational burden of training large CNNs - has led to renewed interest in learning new tasks incrementally while avoiding catastrophic forgetting [14, 34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.807, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Class Incremental Learning (CIL) [1, 33] is a scenario that considers the possibility of adding new classes to already-trained models. Most CIL methods rely on a memory buffer to store exemplars from past tasks [2, 10, 37, 42]." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.268, + 0.891, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.415, + 0.894, + 0.471 + ], + "angle": 0, + "content": "Figure 1. We propose the TASS method, which can be directly applied to many recent exemplar-free class incremental learning methods, resulting in a significant improvement in EFCIL classification accuracy and a reduction in catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.893, + 0.598 + ], + "angle": 0, + "content": "In this paper, we consider Exemplar-Free Class Incremental Learning (EFCIL), which is a more challenging setting in which no data from previous tasks is retained. This is a realistic scenario and of great interest due to privacy concerns or restrictions on the long-term storage of data. The inability to retain examples from past tasks, however, significantly exacerbates the problem of catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.825 + ], + "angle": 0, + "content": "There are several recent works that consider the EFCIL problem. DeepInversion [45] inverts trained networks from random noise to generate images as exemplars and mixes them with current task samples for training. SDC [46] updates prototypes of each learned class by hypothesizing that semantic drift of classes from previous tasks can be approximated and estimated using new data. Other previous works propose representation learning methods for overcoming catastrophic forgetting [51, 52]. As pointed out in IL2A [51], learning better representations can reduce representation bias when transferred to new tasks. Incorporating self-supervised learning tasks, such as Barlow Twins [36] and rotation prediction [52], has also been proposed to achieve more stable representations and alleviate forgetting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "CNNs naturally learn to attend to features that are discriminative for the tasks they are trained to solve. Catastrophic forgetting also occurs in EFCIL due to the model's attention to salient features drifting to features specific to the new task. Standard regularization approaches do little" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.875, + 0.315, + 0.888 + ], + "angle": 0, + "content": "*The first two authors contribute equally." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.875, + 0.315, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23954" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.228 + ], + "angle": 0, + "content": "to prevent this saliency drift when learning new tasks. One direct method of regularizing saliency is to apply distillation on saliency maps of old samples [11]. However, this is complicated by the inability to save samples from previous tasks in the EFCIL setting. Another method is to apply saliency distillation between current task samples and previous task attention [9]. This method however suffers from the semantic gap between current and old classes when enforcing saliency consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.47, + 0.397 + ], + "angle": 0, + "content": "A lack of saliency regularization may lead to attention drifting toward the background in future tasks causing forgetting. Besides, simply applying distillation on attention [9] fails to offer plasticity and is susceptible to attention forgetting, which is a crucial factor of knowledge forgetting. In comparison, our Task-Adaptive Saliency Supervision (TASS) approach aims at keeping saliency focused on incrementally learned tasks (for more details, see Section 4), while maintaining its plasticity and stability. It improves many previous EFCIL methods with higher performance by supervising their attention, as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.47, + 0.671 + ], + "angle": 0, + "content": "Specifically, TASS integrates three components to address this issue. Firstly, we apply dilated boundary maps to prevent saliency drift across object boundaries at intermediate layers. Since saliency drift typically occurs across tasks, encouraging the model to focus on significant foreground regions through dilated boundary supervision reduces the likelihood of saliency shifting toward the background, allowing the model to adaptively select attention areas within the foreground that are relevant to the task. Secondly, to simultaneously enhance the stability of the model's attention across tasks, we add a task-agnostic low-level auxiliary supervision task to the class-incremental framework, which is closely related to our core EFCIL task since image classification has been shown to help models localize the most salient areas within an image. Finally, we propose a module to inject saliency noise into certain feature channels and train the network to denoise them, helping the network further resist attention drift across tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The main contributions of this work are: (i) We provide new insight into task-adaptive saliency supervision under EFCIL settings; we also show the negative effect of methods with no or trivial saliency supervision, which illustrates the superiority of our method and motivates the need for saliency drift mitigation in EFCIL. (ii) We propose the Task-Adaptive Saliency Supervision (TASS) with three components that combine to mitigate the saliency drift problem. (iii) We show that TASS can be easily integrated into other state-of-the-art methods, such as MUC [28], IL2A [51], PASS [52], SSRE [53], leading to significant performance gains. (iv) Our experiments demonstrate that TASS outperforms all existing EFCIL methods and even several exemplar-based methods on the CIFAR-100, TinyImageNet, and ImageNet-Subset EFCIL benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.115, + 0.892, + 0.145 + ], + "angle": 0, + "content": "We first discuss previous work on incremental learning from the recent literature and then describe work on EFCIL." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.153, + 0.707, + 0.169 + ], + "angle": 0, + "content": "2.1. Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.176, + 0.892, + 0.508 + ], + "angle": 0, + "content": "A variety of methods have been proposed for incremental learning in the past few years [1, 7]. Recent works can be coarsely grouped into three categories: replay-based, regularization-based, and parameter-isolation methods. Replay-based methods mitigate the task-recency bias by retaining training samples from previous tasks [37, 48]. In addition to replaying samples, BiC [42], PODNet [10], iCaRL [37] and LTCIL [27] apply a distillation loss to prevent forgetting and enhance model stability. GEM [30], AGEM [4], and MER [38] exploit past-task exemplars by modifying gradients on current training samples to match old samples. Rehearsal may cause models to overfit to stored samples. Regularization-based approaches such as LwF [23], EWC, R-EWC [18, 26], and DMC [49] offer ways to learn better representations while leaving enough plasticity for adaptation to new tasks. Parameter-isolation methods [32, 44] use models with different computational graphs for each task. With the help of growing models, new model branches mitigate catastrophic forgetting at the cost of more parameters and computational cost. It is also widely studied in other fields, such as semantic segmentation [3, 25, 43, 47] and object detection [13, 29]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.69 + ], + "angle": 0, + "content": "As for saliency-guided incremental learning, LwM [9] regularizes the saliency activations of previous classes on current task data. However, there exist semantic gaps between current classes and old ones, which results in an inaccurate distillation target for preserving saliency activations on old samples. RRR [11] directly saves the GradCAM saliency activations of each sample in a replay buffer and applies distillation to memorize this old knowledge, which requires storing additional samples during incremental learning. Despite these initial works on saliency-guided incremental learning, saliency drift still remains problematic and leads to catastrophic forgetting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.866, + 0.714 + ], + "angle": 0, + "content": "2.2. Exemplar-free Class Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Compared to conventional class incremental learning, exemplar-free class incremental learning is more appropriate for applications where training data is sensitive and may not be stored in perpetuity. DAFL [5] uses a GAN to generate synthetic samples from past tasks as an alternative to storing actual data. DeepInversion, which inverts trained networks using random noise to generate images, is another popular EFCIL method [45]. Always Be Dreaming further improves on DeepInversion for EFCIL [41]. SDC attempts to overcome the problems caused by semantic drift when training new tasks on old class samples [46]. It directly estimates prototypes of each learned class to use in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "23955" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.168, + 0.093, + 0.812, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.452, + 0.895, + 0.522 + ], + "angle": 0, + "content": "Figure 2. Overall framework of Task-Adaptive Saliency Supervision (TASS). We apply a low-level model to generate saliency and boundary maps. The boundary map is dilated and downsampled to provide supervision at different stages of the encoder. A decoder is attached after the encoder for low-level distillation, which serves as stationary task-agnostic saliency guidance. To prevent saliency drift in later training phases, we introduce saliency noise into each encoder stage. The model is trained to denoise and reduce the saliency drift on current data in future phases. TASS can be integrated into an EFCIL approach to provide robust saliency guidance across incremental tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.473, + 0.701 + ], + "angle": 0, + "content": "a nearest class mean classifier. PASS [52] and IL2A [51] are prototype-based replay methods for efficient and effective EFCIL. Since these prototypes are features computed from past training samples, image exemplars need not be retained. SSRE [53] introduces a re-parameterization method to trade-off between old and new knowledge. Our task-Adaptive Saliency Supervision (TASS) method uses three new components to reduce saliency drift in EFCIL and is complementary to several of the approaches mentioned above." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.717, + 0.403, + 0.736 + ], + "angle": 0, + "content": "3. Task-Adaptive Saliency Supervision" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.47, + 0.819 + ], + "angle": 0, + "content": "We first define the Exemplar-Free Class Incremental Learning (EFCIL) scenario. Then we describe our TASS approach including dilated boundary supervision, auxiliary low-level supervision, and saliency noise injection. Our overall framework is illustrated in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.446, + 0.848 + ], + "angle": 0, + "content": "3.1. Exemplar-free Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Class-incremental learning aims to sequentially learn tasks consisting of disjoint classes of samples. Let \\( t \\in \\{1, 2, \\dots, T\\} \\) denote the incremental learning tasks. The training data \\( D_{t} \\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.549, + 0.892, + 0.592 + ], + "angle": 0, + "content": "for each task contains classes \\( C_t \\) with \\( N_t \\) training samples \\( \\{(x_t^i, y_t^i)\\}_{i=1}^{N_t} \\), where \\( x_t^i \\) are images and \\( y_t^i \\in C_t \\) are their labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.893, + 0.713 + ], + "angle": 0, + "content": "Most deep networks applied to class-incremental learning can be split into two components: a feature extractor \\( F_{\\theta} \\) and a common classifier \\( G_{\\phi} \\) which grows with each new task \\( t + 1 \\) to include classes \\( C_{t + 1} \\). The feature extractor \\( F_{\\theta} \\) first maps the input \\( x \\) to a deep feature vector \\( z = F_{\\theta}(x) \\in \\mathbb{R}^{d} \\), and then the unified classifier \\( G_{\\phi}(z) \\in \\mathbb{R}^{|C_t|} \\) is a probability distribution over classes \\( C_t \\) used to make predictions on input \\( x \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.895, + 0.851 + ], + "angle": 0, + "content": "Class-incremental learning requires that the model be capable of correctly classifying all samples from previous tasks at any training task – that is, when learning task \\( t \\), the model must not forget how to classify samples from classes from tasks \\( t' < t \\). Exemplar-free class-incremental learning additionally restricts models to learn each new task without access to samples from previous ones. This typically leads to learning objectives that minimize a loss function \\( \\mathcal{L} \\) defined on current training data \\( D_t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.859, + 0.892, + 0.878 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} ^ {\\mathrm {C I L}} (x, y) = \\mathcal {L} _ {\\mathrm {c e}} \\left(G _ {\\phi_ {t}} \\left(F _ {\\theta_ {t}} (x)\\right), y\\right) + \\mathcal {L} _ {t} ^ {\\mathrm {M}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{ce}}\\) is the standard cross-entropy classification loss" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23956" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.094, + 0.468, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.179, + 0.47, + 0.222 + ], + "angle": 0, + "content": "Figure 3. We dilate the boundary map and apply a binary cross entropy loss at three stages in the CNN backbone to prevent mid-level attention from drifting into boundary regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.468, + 0.289 + ], + "angle": 0, + "content": "and \\(\\mathcal{L}_t^{\\mathrm{M}}\\) is a method-specific loss that mitigates forgetting during incremental learning. Note that without \\(\\mathcal{L}_t^{\\mathrm{M}}\\), Eq. 1 reduces to fine-tuning on task \\(t\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.296, + 0.468, + 0.327 + ], + "angle": 0, + "content": "3.2. Boundary-guided Mid-level Saliency Drift Regularization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.468, + 0.546 + ], + "angle": 0, + "content": "Simply distilling attention between tasks does not consider task-adaptive attention. A model generates low-level representations of each input image \\( x \\) (saliency and boundary maps in our experiments). We use CSNet [6] to generate salient regions and boundary maps, as it is lightweight and efficient, but any model producing saliency maps can be used in our framework (we explore other options in the Supplementary Material). To introduce plasticity on attention at these intermediate layers in the backbone, we use the generated boundary maps as a type of adaptive supervision as shown in Figure 3. We add a penalty term on object boundaries to avoid drift into the background. We first use 0.5 as the threshold to binarize the generated boundary map and dilate the maps by:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.553, + 0.468, + 0.569 + ], + "angle": 0, + "content": "\\[\nB _ {d} (x) = \\operatorname {D i l a t e} \\left(A _ {b} (x), d\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.468, + 0.65 + ], + "angle": 0, + "content": "where \\( A_{b}(x) \\) is the generated boundary map of image \\( x \\), which is converted from the saliency map with a Laplacian filter, and \\( d \\) denotes the dilation radius applied on the boundary map for controlling the strictness of boundary-guided saliency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.468, + 0.832 + ], + "angle": 0, + "content": "Rather than use a decoder at each layer as done for the low-level saliency maps described above, the mid-level saliency maps of our model are generated using Grad-CAM [39] at three stages of the CNN backbone (see Figure 2). We also experiment with several other methods for generating student saliency maps and report on them in the Supplementary Material. The dilated generated boundary map \\( B_{d}(x) \\) is downsampled to match the feature map dimensions at these three stages in order to compare with the Grad-CAM generated saliency boundary maps. We use the binary cross entropy loss for supervision on dilated boundary regions. This loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.838, + 0.468, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} ^ {\\mathrm {d b s}} (x) = - \\frac {\\sum_ {j = 1} ^ {N} B _ {d} (x , j) \\log (1 - S (x , j))}{\\sum_ {j = 1} ^ {N} B _ {d} (x , j)}, \\tag {3}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.409, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/jacobgil/pytorch-grad-cam" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "where \\( S(x, j) \\) denotes the saliency map of our model on image \\( x \\) at pixel \\( j \\), \\( B_d(x, j) \\) is the dilated generated boundary map at pixel \\( j \\), and \\( N \\) is the number of pixels in \\( x \\). We compute this loss only within dilated boundary regions (i.e. where \\( B_d(x, j) = 1 \\)). This loss helps the student saliency map avoid intersecting with the dilated teacher boundary region." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.215, + 0.875, + 0.231 + ], + "angle": 0, + "content": "3.3. Auxiliary Low-level Supervisions for EFCIL" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.241, + 0.892, + 0.376 + ], + "angle": 0, + "content": "We propose to learn stable features from low-level stationary tasks shared across all incremental classification tasks during class-incremental learning. Low-level vision tasks like salient object detection require useful representations of input images. By learning these feature representations across tasks, the model can focus on key areas of input images and exploit learned, stable features with less representation drift since the low-level features change very little between tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.892, + 0.636 + ], + "angle": 0, + "content": "Saliency map prediction is relevant to image classification since the foreground largely determines the results, while the background is comparatively less important. When learning new tasks with new classes, the background of images of new classes may contain new visual concepts that introduce undesirable noise and lead to forgetting of essential previous knowledge. The effectiveness of saliency features for learning classification tasks was demonstrated by Saliency Guided Training [17]. Additional supervision of salient region boundaries can aid salient object detection tasks for both segmentation and localization [12, 21, 24, 31, 50]. The positive interaction between these two tasks brings richer attention to features relevant to the main classification task. It can provide positive guidance in the form of stationary knowledge across class-incremental tasks. Some examples are illustrated in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.76 + ], + "angle": 0, + "content": "We incorporate low-level vision tasks into the network as an auxiliary supervision for enriching task-agnostic attention. The boundary map is computed with a Laplacian filter over the estimated saliency map. We add a decoder \\(D_{\\psi}\\) [22] after the backbone \\(F_{\\theta}\\) to predict low-level saliency and boundary maps for input images. The average L2 distance between the prediction and target is used as a low-level saliency distillation loss:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.775, + 0.891, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} ^ {\\mathrm {l m s}} (x) = \\frac {| | D _ {\\psi} (F _ {\\theta} (x)) - A (x) | | _ {2}}{\\sqrt {N}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( A(x) \\) denotes the target low-level maps on input \\( x \\), consisting of a saliency map \\( A_{s}(x) \\) and a boundary map \\( A_{b}(x) \\). \\( D_{\\psi}(F_{\\theta}(x)) \\) are combined saliency and boundary maps produced by the decoder, and \\( N \\) is the number of pixels in the saliency maps." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23957" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.081, + 0.092, + 0.292, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 TASS Pseudocode" + }, + { + "type": "algorithm", + "bbox": [ + 0.078, + 0.11, + 0.47, + 0.352 + ], + "angle": 0, + "content": "Input: The number of tasks \\(T\\), training samples \\(D_{t} = \\{(x_{i},y_{i})\\}_{i=1}^{N}\\) of task \\(t\\), initial parameters \\(\\Theta^0 = \\{\\theta_0,\\phi_0,\\psi_0\\}\\) containing parameters of feature extractor \\(F_{\\theta}\\), classifier \\(G_{\\phi}\\), and low-level decoder \\(D_{\\psi}\\). \nOutput: Model \\(\\Theta^T\\) \n1: for \\(t \\in \\{1,2,\\dots,T\\}\\) do \n2: \\(\\Theta^t \\gets \\Theta^{t-1}\\) \n3: while not converged do \n4: Sample \\((x,y)\\) from \\(D_{t}\\) \n5: \\(\\mathcal{L}_t^{\\mathrm{CIL}} \\gets \\text{SaliencyNoiseInjection}(x,y)\\) \n6: \\(\\mathcal{L}_t^{\\mathrm{Ims}} \\gets \\text{LowLevelMultitask}(x,A(x))\\) \n7: \\(S \\gets \\text{ComputeGradCAMSaliency}(x,y)\\) \n8: \\(\\mathcal{L}_t^{\\mathrm{obs}} \\gets \\text{DilatedBoundarySupervision}(S,A(x))\\) \n9: update \\(\\Theta^t\\) by minimizing \\(\\mathcal{L}_t^{\\mathrm{all}}\\) from Eq. 5 \n10: end while \n11: end for" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.376, + 0.296, + 0.392 + ], + "angle": 0, + "content": "3.4. Saliency Noise Injection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.469, + 0.489 + ], + "angle": 0, + "content": "Although we apply low-level distillation and dilated boundary supervision to maintain saliency representations across tasks, the model can still forget saliency on samples from previous tasks. To address this, we force the model to be able to recover correct saliency maps from injected saliency noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.469, + 0.626 + ], + "angle": 0, + "content": "At each task there is no available training data from previous or future tasks, and therefore we cannot directly know saliency drift on these samples. Instead of supervising the model with ground-truth saliency drift signals, we introduce saliency noise on random feature channels. We use a random ellipse to approximate the potential saliency drift in future tasks and the model is trained to denoise within each stage. Therefore the model can learn to effectively reduce real saliency drift." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.469, + 0.762 + ], + "angle": 0, + "content": "We generate elliptical noise using a very simple approach. There are six parameter dimensions: the center coordinate \\((x, y)\\), the major and minor axis lengths \\((a, b)\\), the rotation angle \\(\\alpha\\), and the mask weight \\(w\\). A detailed explanation of this process is given in the Supplementary Material. With the help of dilated boundary supervision, each stage learns to eliminate this additional saliency noise and this aids generalization for future tasks and mitigates saliency forgetting in previous ones." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.769, + 0.45, + 0.786 + ], + "angle": 0, + "content": "3.5. Learning Objective and Training Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.469, + 0.838 + ], + "angle": 0, + "content": "The overall learning objective combines the low-level multi-task learning, dilated boundary supervision, and random saliency noise injection modules:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.844, + 0.469, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} ^ {\\text {a l l}} = \\mathcal {L} _ {t} ^ {\\text {C I L}} + \\mathcal {L} _ {t} ^ {\\text {l m s}} + \\mathcal {L} _ {t} ^ {\\text {d b s}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Comparing this loss with Eq. 1, for TASS \\(\\mathcal{L}_t^{\\mathrm{M}} = \\mathcal{L}_t^{\\mathrm{CIL}} + \\mathcal{L}_t^{\\mathrm{Ins}}\\), thus incorporating saliency-aware supervision with" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.891, + 0.122 + ], + "angle": 0, + "content": "the cross-entropy loss. The entire process is detailed in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.138, + 0.704, + 0.155 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.164, + 0.892, + 0.224 + ], + "angle": 0, + "content": "In this section we first describe our experimental setup and then compare TASS to state-of-the-art methods on several EFCIL benchmarks. In Section 4.3 we give further analysis of the various components of TASS." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.236, + 0.691, + 0.253 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.26, + 0.891, + 0.289 + ], + "angle": 0, + "content": "We follow standard experimental protocols for EFCIL on three benchmark datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.892, + 0.427 + ], + "angle": 0, + "content": "Datasets. We perform experiments on CIFAR-100 [19], Tiny-ImageNet [20], and ImageNet-Subset [8]. For most experiments, we train the model on half of the classes for the first task, and then equally distribute the remaining classes across each of the subsequent tasks. The convention we use is: \\( F + C \\times T \\) means that the first task contains \\( F \\) classes, and the next \\( T \\) tasks each contain \\( C \\) classes. This is a common configuration for EFCIL used in both PASS [52] and SSRE [53]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.892, + 0.579 + ], + "angle": 0, + "content": "State-of-the-art methods. Since we focus on EFCIL, we mainly compare with exemplar-free state-of-the-art approaches: SSRE [53], PASS [52], IL2A [51], EWC [18], LwF-MC [37], and MUC [28]. To demonstrate the effectiveness of our method, we also compare its performance with several exemplar-based methods like iCaRL (both nearest-mean and CNN) [37], EEIL [2], and LUCIR [16]. We also compare with RRR [11] integrated with SSRE, which focuses on preserving saliency using exemplar replay." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.79 + ], + "angle": 0, + "content": "Implementation details and metrics. We use ResNet-18 [15] as a feature extraction backbone. This is the same base network used in SSRE [53] and PASS [52], two state-of-the-art EFCIL approaches. We use the decoder in [22] to estimate low-level student saliency maps. All experiments are trained from scratch using Adam for 100 epochs with an initial learning rate 0.001. The learning rate is reduced by a factor 10 at epochs 45 and 90. For exemplar-based approaches, we use herding [37] to select and store 20 samples per class following common settings [16, 37]. We implement RRR [11] with SSRE to fairly compare it with TASS. For dilated boundary supervision, we set \\(d\\) of the three mid-level boundary dilation stages to be \\(5\\%\\), \\(10\\%\\) and \\(15\\%\\) of the image size." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.792, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We report three common metrics for class incremental learning: the average and last top-1 accuracy, as well as average forgetting for all classes learned up to task \\( t \\). Denoting by \\( Acc_{i} \\) the accuracy over all learned classes up to and including task \\( i \\), the average accuracy is defined as \\( Avg = \\frac{\\sum_{i=1}^{T} Acc_{i}}{T} \\), and the last accuracy is \\( Acc_{T} \\). Letting \\( a_{m,n} \\) denotes the accuracy of task \\( n \\) after learning task \\( m \\)," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23958" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.171, + 0.089, + 0.8, + 0.243 + ], + "angle": 0, + "content": "
DatasetCIFAR-100Tiny-ImageNet
Method5 tasks10 tasks20 tasks5 tasks10 tasks20 tasks
MUC38.4519.5715.6518.9515.479.14
+TASS49.17 (+10.72)40.34 (+20.77)37.86 (+22.21)32.47 (+13.46)30.13 (+14.66)27.70 (+18.56)
IL2A55.1345.3245.2436.7734.5328.68
+TASS58.74 (+3.61)53.24 (+7.92)53.07 (+7.83)42.49 (+5.72)41.34 (+6.81)40.59 (+11.91)
PASS55.6749.0348.4841.5839.2832.78
+TASS59.10 (+3.43)54.45 (+5.42)52.37 (+3.89)44.05 (+2.47)43.06 (+3.78)42.57 (+9.79)
SSRE56.3355.0150.4741.4540.0739.25
+TASS59.26 (+2.93)57.93 (+2.92)53.78 (+3.31)44.13 (+2.68)43.86 (+3.79)43.55 (+4.30)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.254, + 0.895, + 0.281 + ], + "angle": 0, + "content": "Table 1. Performance gain in top-1 accuracy by applying TASS to other EFCIL methods in a plug-and-play way. Absolute gains are indicated in (red)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.295, + 0.89, + 0.503 + ], + "angle": 0, + "content": "
DatasetCIFAR100TinyImageNet
Setting5 tasks10 tasks20 tasks5 tasks10 tasks20 tasks
MethodAvg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓
E=20iCaRL-CNN†51.0740.1242.1348.6639.6545.6944.4335.4743.5434.6422.3136.8931.1521.1036.7027.9020.4645.12
iCaRL-NCM†58.5649.7424.9054.1945.1328.3250.5140.6835.5345.8633.4527.1543.2933.7528.8938.0428.8937.40
LUCIR†63.7855.0621.0062.3950.1425.1259.0748.7828.6549.1537.0920.6148.5236.8022.2542.8332.5533.74
EEIL†60.3752.3523.3656.0547.6726.6552.3441.5932.4047.1234.2425.5645.0134.2625.9140.5030.1435.04
RRR†66.4357.2218.0565.7855.7418.5962.4351.3518.4051.2042.2316.6749.5440.1221.6447.4635.5429.10
E=20LwFMC45.9336.1744.2327.4350.4717.0420.0715.8855.4629.1217.1254.2623.1012.3354.3717.438.7563.54
EWC16.049.3260.1714.708.4762.5314.128.2363.8918.8012.7167.5515.7710.1270.2312.398.4275.54
MUC49.4238.4540.2830.1919.5747.5621.2715.6552.6532.5817.9851.4626.6114.5450.2121.9512.7058.00
IL2A63.2255.1323.7857.6545.3230.4154.9045.2430.8448.1736.1425.4342.1035.2328.3236.7928.7435.46
PASS63.4755.6725.2061.8449.0330.2558.0948.4830.6149.5541.5818.0447.2939.2823.1142.0732.7830.55
SSRE65.8856.3318.3765.0455.0119.4861.7050.4718.3750.3941.6717.2548.9339.8922.5048.1739.7626.74
TASS (Ours)68.7559.2616.4267.4257.9317.7862.7653.7817.7855.1244.1315.4054.2143.8618.4752.7943.5522.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.513, + 0.895, + 0.556 + ], + "angle": 0, + "content": "Table 2. Average, last top-1 accuracy, and forgetting on CIFAR-100 with different numbers of tasks. Replay-based methods storing 20 exemplars from each previous class are indicate with †. The best overall results are in bold. We run all experiments three times and report the mean for all metrics." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.564, + 0.485, + 0.67 + ], + "angle": 0, + "content": "
DatasetImageNet-Subset
Setting5 tasks10 tasks20 tasks
MethodAvg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓
LwF_TC34.8624.1049.3631.1820.0153.0427.5417.4256.07
MUC40.6527.8947.1335.0722.6552.1031.4420.1253.85
PASS63.1252.6122.4761.8050.4423.5755.2346.0726.73
SSRE69.5458.4617.2267.6957.5118.6061.2350.0523.22
TASS (Ours)74.3263.1414.3772.6057.9316.0968.7957.6018.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.68, + 0.472, + 0.722 + ], + "angle": 0, + "content": "Table 3. Average, last top-1 accuracy, and forgetting on ImageNet-Subset for different numbers of tasks. We run all experiments three times and report means for all metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.752, + 0.47, + 0.803 + ], + "angle": 0, + "content": "the forgetting measure \\( f_{k}^{i} \\) of task \\( i \\) after learning task \\( k \\) is computed as \\( f_{k}^{i} = \\max_{t\\in 1,2,\\ldots ,k - 1}(a_{t,i} - a_{k,i}) \\). The average forgetting \\( F_{k} \\) is defined as \\( F_{k} = \\frac{1}{k - 1}\\sum_{i = 1}^{k - 1}f_{k}^{i} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.4, + 0.831 + ], + "angle": 0, + "content": "4.2. Comparison with the State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We compare TASS with the state-of-the-art on CIFAR-100 in Table 2. TASS outperforms all exemplar-free approaches. For exemplar-based methods like iCaRL [37], EEIL [2], and LUCIR [16], our method still has significantly better" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.658 + ], + "angle": 0, + "content": "performance. On longer sequences (i.e. 10 and 20 tasks), our method significantly reduces forgetting when learning new classes compared to other EFCIL methods. TASS outperforms the best method SSRE by about \\(3\\%\\) on the last task. This performance improvement can be also observed in terms of average forgetting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.893, + 0.87 + ], + "angle": 0, + "content": "As we see in Table 3 and Figure 4 for Tiny-ImageNet and ImageNet-Subset, although our method has similar top-1 accuracy on the first task in Figure 4, it has better performance at most intermediate tasks and also the final one. For longer sequences in Figure 4, the gap between our method and the best baseline is largely consistent, showing the effectiveness of our method at mitigating forgetting. The performance gain in Table 3 is larger on Tiny-ImageNet and ImageNet-Subset compared to CIFAR100, and this demonstrates that our method generalizes to datasets with larger images and object scales. It is worth mentioning that TASS also produces results with smaller variance. We believe this to be due to TASS reducing saliency drift to background regions, which may include random noise." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Plug-and-play with other EFCIL methods. Some existing EFCIL methods, like PASS [52], IL2A [51] and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23959" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.096, + 0.353, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.096, + 0.603, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.097, + 0.848, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.234, + 0.353, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.234, + 0.601, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.235, + 0.848, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.37, + 0.894, + 0.397 + ], + "angle": 0, + "content": "Figure 4. Results on Tiny-ImageNet and ImageNet-Subset for different numbers of tasks. Our method outperforms others, especially on longer task sequences (i.e. more, but smaller, tasks)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.473, + 0.621 + ], + "angle": 0, + "content": "SSRE [53], focus on reducing forgetting via embedding regularization. Considering the importance of saliency to image classification, it is natural to consider whether TASS can be integrated into these methods. The results in Table 1 show the performance gain brought by this integration. Adding TASS doubles the performance for MUC in many cases and significantly improves IL2A and PASS. When we incorporate it into the best baseline SSRE, it yields a consistent gain of about \\(3\\%\\). These results clearly show that TASS, by explicitly mitigating saliency drift, is complementary to other methods in relieving forgetting. They additionally demonstrate the significance of saliency drift as a cause of catastrophic forgetting in EFCIL." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.633, + 0.267, + 0.649 + ], + "angle": 0, + "content": "4.3. Additional Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.702 + ], + "angle": 0, + "content": "In this section we take a deeper look at the method we propose. If not specified, the results are produced using TASS integrated into SSRE [53]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Ablation Study. We performed ablations using the 10-task setting on CIFAR-100 (see Table 4). We ablate on both PASS [52] and SSRE [53]. Low-level multi-task supervision is crucial and improves by \\(2.2\\%\\) (PASS) and \\(1.2\\%\\) (SSRE). Dilated boundary supervision further boosts performance by about \\(1 - 2\\%\\). Saliency noise injection is also helpful for both methods and improves by \\(1.5\\%\\) for PASS. In total, TASS improves baselines by \\(5.5\\%\\) and \\(2.9\\%\\) points, respectively. Note that SSRE is the previous state-of-the-art method and TASS outperforms it by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Low-level Multi-task Supervision. To analyze the effect of our proposed low-level saliency supervision, we performed experiments on ImageNet-Subset in the 5 and 10" + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.421, + 0.846, + 0.589 + ], + "angle": 0, + "content": "
Method & TasksL1msL1bsSNIAccuracy
Baseline (PASS)49.0
Variants✓51.2
✓✓53.0
✓✓✓54.5
Baseline (SSRE)55.0
Variants✓56.2
✓✓57.3
✓✓✓57.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.598, + 0.892, + 0.682 + ], + "angle": 0, + "content": "Table 4. Ablations on each TASS component. Experiments are on CIFAR-100 in the 10-task setting and we report the top-1 accuracy in % for TASS integrated into PASS and SSRE. \\(\\mathcal{L}_{\\mathrm{ubs}}\\) (Eq. 3), \\(\\mathcal{L}_{\\mathrm{lms}}\\) (Eq. 4), and SNI denote the three TASS components: Dilated Boundary Supervision, Low-level Multi-task Supervision, and Saliency Noise Injection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "task settings. We first plot the loss across tasks in Figure 5(c). After learning to predict boundary and saliency maps in the first task, the network maintains good performance for the rest of the 5 task sequence. This shows that the low-level tasks are stable during continual learning. Furthermore, we visualize the results of saliency and boundary map prediction during incremental learning in Figure 5(a-b). We give some examples of predicted boundary and saliency maps after learning different tasks. Although CIL involves samples of different classes, we see that the low-level outputs are relatively stable and class agnostic. Since the model is able to stably predict these low-level features across tasks, it therefore can preserve useful prior knowl" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23960" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.092, + 0.357, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.267, + 0.227, + 0.28 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.092, + 0.633, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.267, + 0.503, + 0.28 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.092, + 0.882, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.268, + 0.781, + 0.281 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.298, + 0.895, + 0.34 + ], + "angle": 0, + "content": "Figure 5. Visualization of the saliency (a) and boundary (b) maps from our student encoder-decoder network with original images from different tasks at different stages of incremental learning. Our method produces stable low-level results while reducing forgetting in classification. (c) The MAE loss between the student and teacher network across tasks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.357, + 0.262, + 0.372 + ], + "angle": 0, + "content": "edge for continual learning." + }, + { + "type": "table", + "bbox": [ + 0.167, + 0.397, + 0.382, + 0.495 + ], + "angle": 0, + "content": "
Metric & MethodAvg↑Last↑
FeTrIL [35]65.2056.34
SOPE [54]65.8456.80
PRAKA [40]68.8659.20
TASS (SSRE)67.4257.93
TASS (PRAKA)69.7060.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.505, + 0.468, + 0.519 + ], + "angle": 0, + "content": "Table 5. Average and last accuracy on CIFAR-100 10-task setting." + }, + { + "type": "table", + "bbox": [ + 0.189, + 0.557, + 0.361, + 0.624 + ], + "angle": 0, + "content": "
Metric & MethodAvg↑
SOPE [54]60.20
FeTrIL [35]65.00
TASS (FeTrIL)66.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.634, + 0.465, + 0.648 + ], + "angle": 0, + "content": "Table 6. Average accuracy on the ImageNet-Full 10-task setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.47, + 0.788 + ], + "angle": 0, + "content": "TASS with more methods and benchmarks. Due to the strong generalization capability of our method, we have also applied our TASS paradigm to PRAKA [40], resulting in further performance improvements. We give specific experimental comparisons in Table 5. We also compare with FeTrIL [35] in Table 5 and Table 6 above. Experiments on ImageNet-Full show consistent improvement in Table 6 above." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Quantitative analysis of saliency drift. We measure the intersection over union (IoU) between the self-attention map of the last layer in DINO and the Grad-CAM saliency maps for both SSRE and \\(\\mathrm{SSRE + TASS}\\). As shown in Table 7, TASS significantly reduces saliency drift in the student model, which again shows the effectiveness of our method for maintaining saliency during CIL." + }, + { + "type": "table", + "bbox": [ + 0.555, + 0.353, + 0.839, + 0.4 + ], + "angle": 0, + "content": "
Task14710
SSRE47.450.156.678.5
SSRE+TASS75.282.388.590.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.41, + 0.892, + 0.438 + ], + "angle": 0, + "content": "Table 7. Average IoU (\\%) between DINO self-attention maps and student model saliency in the CIFAR-100 10-task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.465, + 0.627, + 0.481 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.892, + 0.648 + ], + "angle": 0, + "content": "In this paper, we propose an approach of task-adaptive saliency guidance for EFCIL. The insight behind TASS is to guide the model to focus on salient regions and inhibit saliency drift. We show that robust saliency guidance is crucial to mitigating forgetting across tasks. Experiments demonstrate that TASS is effective and surpasses the state-of-the-art. TASS can be easily combined with other methods, leading to large performance gains over baselines. Qualitative results also show that low-level tasks are stable across different tasks, resulting in less forgetting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.892, + 0.804 + ], + "angle": 0, + "content": "Limitations and Future Work. In this work, we introduce the auxiliary saliency knowledge across tasks for better CIL performance, which may raise the concern of unfair comparison. However, we believe it is important to explore external knowledge to make the CIL system applicable in real applications. Additionally, our method can be easily integrated with other methods and the low-level auxiliary knowledge takes negligible cost and is easy to obtain. We will explore leveraging other forms of knowledge in future work, such as pre-trained visual and large language models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments This work is funded by NSFC (NO. 62206135, 62225604), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), and the Fundamental Research Funds for the Central Universities (Nankai University, 070-63233085). Computation is supported by the Supercomputing Center of Nankai University." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23961" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Eden Belouadah, Adrian Popescu, and Ioannis Kanellos. A comprehensive study of class incremental learning algorithms for visual tasks. Neural Networks, 135:38-54, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.47, + 0.214 + ], + "angle": 0, + "content": "[2] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In ECCV, 2018. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.47, + 0.285 + ], + "angle": 0, + "content": "[3] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9233-9242, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.47, + 0.327 + ], + "angle": 0, + "content": "[4] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.47, + 0.384 + ], + "angle": 0, + "content": "[5] Hanting Chen, Yunhe Wang, Chang Xu, Zhaohui Yang, Chuanjian Liu, Boxin Shi, Chunjing Xu, Chao Xu, and Qi Tian. Data-free learning of student networks. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.47, + 0.441 + ], + "angle": 0, + "content": "[6] Ming-Ming Cheng, Shanghua Gao, Ali Borji, Yong-Qiang Tan, Zheng Lin, and Meng Wang. A highly efficient model to study the semantics of salient object detection. IEEE TPAMI, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.47, + 0.499 + ], + "angle": 0, + "content": "[7] Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. IEEE TPAMI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.47, + 0.542 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.544, + 0.47, + 0.585 + ], + "angle": 0, + "content": "[9] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.588, + 0.47, + 0.641 + ], + "angle": 0, + "content": "[10] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.645, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[11] Sayna Ebrahimi, Suzanne Petryk, Akash Gokul, William Gan, Joseph E. Gonzalez, Marcus Rohrbach, and trevor darrell. Remembering for the right reasons: Explanations reduce catastrophic forgetting. In ICLR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.702, + 0.47, + 0.756 + ], + "angle": 0, + "content": "[12] Deng-Ping Fan, Ge-Peng Ji, Peng Xu, Ming-Ming Cheng, Christos Sakaridis, and Luc Van Gool. Advances in deep concealed scene understanding. Visual Intelligence, 1(1):16, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.759, + 0.47, + 0.828 + ], + "angle": 0, + "content": "[13] Tao Feng, Mang Wang, and Hangjie Yuan. Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9427-9436, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.83, + 0.47, + 0.885 + ], + "angle": 0, + "content": "[14] Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv preprint arXiv:1312.6211, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.887, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Deep residual learning for image recognition. In CVPR, 2016.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[16] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.892, + 0.208 + ], + "angle": 0, + "content": "[17] Aya Abdelsalam Ismail, Hector Corrada Bravo, and Soheil Feizi. Improving deep learning interpretability by saliency guided training. NeurIPS, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.211, + 0.892, + 0.292 + ], + "angle": 0, + "content": "[18] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.296, + 0.892, + 0.334 + ], + "angle": 0, + "content": "[19] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.339, + 0.892, + 0.367 + ], + "angle": 0, + "content": "[20] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.369, + 0.892, + 0.451 + ], + "angle": 0, + "content": "[21] Hao Li, Dingwen Zhang, Nian Liu, Lechao Cheng, Yalun Dai, Chao Zhang, Xinggang Wang, and Junwei Han. Boosting low-data instance segmentation by unsupervised pretraining with saliency prompt. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15485-15494, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.454, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[22] Xiangtai Li, Ansheng You, Zhen Zhu, Houlong Zhao, Maoke Yang, Kuiyuan Yang, and Yunhai Tong. Semantic flow for fast and accurate scene parsing. In ECCV, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[23] Zhizhong Li and Derek Hoiem. Learning without forgetting. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.528, + 0.892, + 0.582 + ], + "angle": 0, + "content": "[24] Liqiang Lin, Pengdi Huang, Chi-Wing Fu, Kai Xu, Hao Zhang, and Hui Huang. On learning the right attention point for feature enhancement. Science China Information Sciences, 66(1):112107, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.585, + 0.892, + 0.627 + ], + "angle": 0, + "content": "[25] Zheng Lin, Zhao Zhang, Zi-Yue Zhu, Deng-Ping Fan, and Xia-Lei Liu. Sequential interactive image segmentation. Computational Visual Media, 9(4):753-765, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.629, + 0.892, + 0.684 + ], + "angle": 0, + "content": "[26] Xialei Liu, Marc Masana, Luis Herranz, Joost Van de Weijer, Antonio M Lopez, and Andrew D Bagdanov. Rotate your networks: Better weight consolidation and less catastrophic forgetting. In ICPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[27] Xialei Liu, Yu-Song Hu, Xu-Sheng Cao, Andrew D Bagdanov, Ke Li, and Ming-Ming Cheng. Long-tailed class incremental learning. In European Conference on Computer Vision, pages 495-512. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[28] Yu Liu, Sarah Parisot, Gregory Slabaugh, Xu Jia, Ales Leonardis, and Tinne Tuytelaars. More classifiers, less forgetting: A generic multi-classifier paradigm for incremental learning. In ECCV. Springer, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[29] Yuyang Liu, Yang Cong, Dipam Goswami, Xialei Liu, and Joost van de Weijer. Augmented box replay: Overcoming foreground shift for incremental object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11367-11377, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. NeurIPS, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23962" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[31] Zhouzhou Ma, Guanghua Gu, and Wenrui Zhao. Self-attention guidance based crowd localization and counting. Machine Intelligence Research, pages 1–17, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.176 + ], + "angle": 0, + "content": "[32] Arun Mallya and Svetlana Lazebnik. Packet: Adding multiple tasks to a single network by iterative pruning. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.181, + 0.468, + 0.234 + ], + "angle": 0, + "content": "[33] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost van de Weijer. Class incremental learning: survey and performance evaluation on image classification. IEEE TPAMI, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.239, + 0.468, + 0.292 + ], + "angle": 0, + "content": "[34] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of learning and motivation*, pages 109–165. Elsevier, 1989. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.296, + 0.468, + 0.337 + ], + "angle": 0, + "content": "[35] Grégoire Petit, Adrian Popescu, Hugo Schindler, David Picard, and Bertrand Delezoise. Fetril: Feature translation for exemplar-free class-incremental learning. In WACV, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.34, + 0.468, + 0.367 + ], + "angle": 0, + "content": "[36] Quang Pham, Chenghao Liu, and Steven Hoi. Dualnet: Continual learning, fast and slow. NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.369, + 0.468, + 0.422 + ], + "angle": 0, + "content": "[37] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.427, + 0.468, + 0.48 + ], + "angle": 0, + "content": "[38] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.485, + 0.468, + 0.539 + ], + "angle": 0, + "content": "[39] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.542, + 0.468, + 0.583 + ], + "angle": 0, + "content": "[40] Wuxuan Shi and Mang Ye. Prototype reminiscence and augmented asymmetric knowledge aggregation for nonexemplar class-incremental learning. In ICCV, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.586, + 0.468, + 0.639 + ], + "angle": 0, + "content": "[41] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.644, + 0.468, + 0.684 + ], + "angle": 0, + "content": "[42] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.687, + 0.468, + 0.755 + ], + "angle": 0, + "content": "[43] Jia-Wen Xiao, Chang-Bin Zhang, Jiekang Feng, Xialei Liu, Joost van de Weijer, and Ming-Ming Cheng. Endpoints weight fusion for class incremental semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7204-7213, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.759, + 0.468, + 0.785 + ], + "angle": 0, + "content": "[44] Ju Xu and Zhanxing Zhu. Reinforced continual learning. NeurIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.789, + 0.468, + 0.842 + ], + "angle": 0, + "content": "[45] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[46] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In CVPR, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[47] Lu Yu, Xialei Liu, and Joost Van de Weijer. Self-training for class-incremental semantic segmentation. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[48] Jiang-Tian Zhai, Xialei Liu, Andrew D Bagdanov, Ke Li, and Ming-Ming Cheng. Masked autoencoders are efficient class incremental learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19104-19113, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[49] Junting Zhang, Jie Zhang, Shalini Ghosh, Dawei Li, Serafettin Tasci, Larry Heck, Heming Zhang, and C-C Jay Kuo. Class-incremental learning via deep model consolidation. In WACV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[50] Jia-Xing Zhao, Jiang-Jiang Liu, Deng-Ping Fan, Yang Cao, Jufeng Yang, and Ming-Ming Cheng. Egnet: Edge guidance network for salient object detection. In ICCV, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.306, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[51] Fei Zhu, Zhen Cheng, Xu-yao Zhang, and Cheng-lin Liu. Class-incremental learning via dual augmentation. NeurIPS, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[52] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In CVPR, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[53] Kai Zhu, Wei Zhai, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Self-sustaining representation expansion for nonexemplar class-incremental learning. In CVPR, 2022. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[54] Kai Zhu, Kecheng Zheng, Ruili Feng, Deli Zhao, Yang Cao, and Zheng-Jun Zha. Self-organizing pathway expansion for non-exemplar class-incremental learning. In ICCV, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23963" + } + ] +] \ No newline at end of file diff --git a/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/ef2d4502-c3bc-43a6-a846-89e23e77cec3_origin.pdf b/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/ef2d4502-c3bc-43a6-a846-89e23e77cec3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9df2850decc86b9f48b7f3e0a828cec5f795c120 --- /dev/null +++ b/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/ef2d4502-c3bc-43a6-a846-89e23e77cec3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91f44c856a0596cf8efd1b1873f76483ffb15ee046c3c43e035a2ced00f2647f +size 2834651 diff --git a/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/full.md b/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e43b707f1f6ded1bc32d37e8a324784bc11388cc --- /dev/null +++ b/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/full.md @@ -0,0 +1,309 @@ +# Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning + +Xialei Liu $^{2,1,*}$ Jiang-Tian Zhai $^{1,*}$ Andrew D. Bagdanov $^{3}$ Ke Li $^{4}$ Ming-Ming Cheng $^{2,1,\dagger}$ + +$^{1}$ VCIP, CS, Nankai University $^{2}$ NKIARI, Shenzhen Futian $^{3}$ MICC, University of Florence $^{4}$ Tencent Youtu Lab {xialei, cmm}@nankai.edu.cn, {jtzhai30,tristanli.sh}@gmail.com, andrew.bagdanov@unifi.it + +# Abstract + +Exemplar-free Class Incremental Learning (EFCIL) aims to sequentially learn tasks with access only to data from the current one. EFCIL is of interest because it mitigates concerns about privacy and long-term storage of data, while at the same time alleviating the problem of catastrophic forgetting in incremental learning. In this work, we introduce task-adaptive saliency for EFCIL and propose a new framework, which we call Task-Adaptive Saliency Supervision (TASS), for mitigating the negative effects of saliency drift between different tasks. We first apply boundary-guided saliency to maintain task adaptivity and plasticity on model attention. Besides, we introduce task-agnostic low-level signals as auxiliary supervision to increase the stability of model attention. Finally, we introduce a module for injecting and recovering saliency noise to increase the robustness of saliency preservation. Our experiments demonstrate that our method can better preserve saliency maps across tasks and achieve state-of-the-art results on the CIFAR-100, Tiny-ImageNet, and ImageNet-Subset EFCIL benchmarks. Code is available at https://github.com/scok30/tass. + +# 1. Introduction + +Deep neural networks achieve state-of-the-art performance on many computer vision tasks. However, most of these tasks consider a static world in which tasks are well-defined, stationary, and all training data is available in a single training session. The real world consists of dynamically changing environments and data distributions, which - especially given the computational burden of training large CNNs - has led to renewed interest in learning new tasks incrementally while avoiding catastrophic forgetting [14, 34]. + +Class Incremental Learning (CIL) [1, 33] is a scenario that considers the possibility of adding new classes to already-trained models. Most CIL methods rely on a memory buffer to store exemplars from past tasks [2, 10, 37, 42]. + +![](images/fd5dab2fb9d2796dc711f57546f48ac75b4224e5bfdec6811129787de84e050a.jpg) +Figure 1. We propose the TASS method, which can be directly applied to many recent exemplar-free class incremental learning methods, resulting in a significant improvement in EFCIL classification accuracy and a reduction in catastrophic forgetting. + +In this paper, we consider Exemplar-Free Class Incremental Learning (EFCIL), which is a more challenging setting in which no data from previous tasks is retained. This is a realistic scenario and of great interest due to privacy concerns or restrictions on the long-term storage of data. The inability to retain examples from past tasks, however, significantly exacerbates the problem of catastrophic forgetting. + +There are several recent works that consider the EFCIL problem. DeepInversion [45] inverts trained networks from random noise to generate images as exemplars and mixes them with current task samples for training. SDC [46] updates prototypes of each learned class by hypothesizing that semantic drift of classes from previous tasks can be approximated and estimated using new data. Other previous works propose representation learning methods for overcoming catastrophic forgetting [51, 52]. As pointed out in IL2A [51], learning better representations can reduce representation bias when transferred to new tasks. Incorporating self-supervised learning tasks, such as Barlow Twins [36] and rotation prediction [52], has also been proposed to achieve more stable representations and alleviate forgetting. + +CNNs naturally learn to attend to features that are discriminative for the tasks they are trained to solve. Catastrophic forgetting also occurs in EFCIL due to the model's attention to salient features drifting to features specific to the new task. Standard regularization approaches do little + +to prevent this saliency drift when learning new tasks. One direct method of regularizing saliency is to apply distillation on saliency maps of old samples [11]. However, this is complicated by the inability to save samples from previous tasks in the EFCIL setting. Another method is to apply saliency distillation between current task samples and previous task attention [9]. This method however suffers from the semantic gap between current and old classes when enforcing saliency consistency. + +A lack of saliency regularization may lead to attention drifting toward the background in future tasks causing forgetting. Besides, simply applying distillation on attention [9] fails to offer plasticity and is susceptible to attention forgetting, which is a crucial factor of knowledge forgetting. In comparison, our Task-Adaptive Saliency Supervision (TASS) approach aims at keeping saliency focused on incrementally learned tasks (for more details, see Section 4), while maintaining its plasticity and stability. It improves many previous EFCIL methods with higher performance by supervising their attention, as shown in Figure 1. + +Specifically, TASS integrates three components to address this issue. Firstly, we apply dilated boundary maps to prevent saliency drift across object boundaries at intermediate layers. Since saliency drift typically occurs across tasks, encouraging the model to focus on significant foreground regions through dilated boundary supervision reduces the likelihood of saliency shifting toward the background, allowing the model to adaptively select attention areas within the foreground that are relevant to the task. Secondly, to simultaneously enhance the stability of the model's attention across tasks, we add a task-agnostic low-level auxiliary supervision task to the class-incremental framework, which is closely related to our core EFCIL task since image classification has been shown to help models localize the most salient areas within an image. Finally, we propose a module to inject saliency noise into certain feature channels and train the network to denoise them, helping the network further resist attention drift across tasks. + +The main contributions of this work are: (i) We provide new insight into task-adaptive saliency supervision under EFCIL settings; we also show the negative effect of methods with no or trivial saliency supervision, which illustrates the superiority of our method and motivates the need for saliency drift mitigation in EFCIL. (ii) We propose the Task-Adaptive Saliency Supervision (TASS) with three components that combine to mitigate the saliency drift problem. (iii) We show that TASS can be easily integrated into other state-of-the-art methods, such as MUC [28], IL2A [51], PASS [52], SSRE [53], leading to significant performance gains. (iv) Our experiments demonstrate that TASS outperforms all existing EFCIL methods and even several exemplar-based methods on the CIFAR-100, TinyImageNet, and ImageNet-Subset EFCIL benchmarks. + +# 2. Related Work + +We first discuss previous work on incremental learning from the recent literature and then describe work on EFCIL. + +# 2.1. Incremental Learning + +A variety of methods have been proposed for incremental learning in the past few years [1, 7]. Recent works can be coarsely grouped into three categories: replay-based, regularization-based, and parameter-isolation methods. Replay-based methods mitigate the task-recency bias by retaining training samples from previous tasks [37, 48]. In addition to replaying samples, BiC [42], PODNet [10], iCaRL [37] and LTCIL [27] apply a distillation loss to prevent forgetting and enhance model stability. GEM [30], AGEM [4], and MER [38] exploit past-task exemplars by modifying gradients on current training samples to match old samples. Rehearsal may cause models to overfit to stored samples. Regularization-based approaches such as LwF [23], EWC, R-EWC [18, 26], and DMC [49] offer ways to learn better representations while leaving enough plasticity for adaptation to new tasks. Parameter-isolation methods [32, 44] use models with different computational graphs for each task. With the help of growing models, new model branches mitigate catastrophic forgetting at the cost of more parameters and computational cost. It is also widely studied in other fields, such as semantic segmentation [3, 25, 43, 47] and object detection [13, 29]. + +As for saliency-guided incremental learning, LwM [9] regularizes the saliency activations of previous classes on current task data. However, there exist semantic gaps between current classes and old ones, which results in an inaccurate distillation target for preserving saliency activations on old samples. RRR [11] directly saves the GradCAM saliency activations of each sample in a replay buffer and applies distillation to memorize this old knowledge, which requires storing additional samples during incremental learning. Despite these initial works on saliency-guided incremental learning, saliency drift still remains problematic and leads to catastrophic forgetting. + +# 2.2. Exemplar-free Class Incremental Learning + +Compared to conventional class incremental learning, exemplar-free class incremental learning is more appropriate for applications where training data is sensitive and may not be stored in perpetuity. DAFL [5] uses a GAN to generate synthetic samples from past tasks as an alternative to storing actual data. DeepInversion, which inverts trained networks using random noise to generate images, is another popular EFCIL method [45]. Always Be Dreaming further improves on DeepInversion for EFCIL [41]. SDC attempts to overcome the problems caused by semantic drift when training new tasks on old class samples [46]. It directly estimates prototypes of each learned class to use in + +![](images/bd01f5a77150e3636fa64febeed139f3cdf717582aa24e4555829237e5c64cf2.jpg) +Figure 2. Overall framework of Task-Adaptive Saliency Supervision (TASS). We apply a low-level model to generate saliency and boundary maps. The boundary map is dilated and downsampled to provide supervision at different stages of the encoder. A decoder is attached after the encoder for low-level distillation, which serves as stationary task-agnostic saliency guidance. To prevent saliency drift in later training phases, we introduce saliency noise into each encoder stage. The model is trained to denoise and reduce the saliency drift on current data in future phases. TASS can be integrated into an EFCIL approach to provide robust saliency guidance across incremental tasks. + +a nearest class mean classifier. PASS [52] and IL2A [51] are prototype-based replay methods for efficient and effective EFCIL. Since these prototypes are features computed from past training samples, image exemplars need not be retained. SSRE [53] introduces a re-parameterization method to trade-off between old and new knowledge. Our task-Adaptive Saliency Supervision (TASS) method uses three new components to reduce saliency drift in EFCIL and is complementary to several of the approaches mentioned above. + +# 3. Task-Adaptive Saliency Supervision + +We first define the Exemplar-Free Class Incremental Learning (EFCIL) scenario. Then we describe our TASS approach including dilated boundary supervision, auxiliary low-level supervision, and saliency noise injection. Our overall framework is illustrated in Figure 2. + +# 3.1. Exemplar-free Class-Incremental Learning + +Class-incremental learning aims to sequentially learn tasks consisting of disjoint classes of samples. Let $t \in \{1, 2, \dots, T\}$ denote the incremental learning tasks. The training data $D_{t}$ + +for each task contains classes $C_t$ with $N_t$ training samples $\{(x_t^i, y_t^i)\}_{i=1}^{N_t}$ , where $x_t^i$ are images and $y_t^i \in C_t$ are their labels. + +Most deep networks applied to class-incremental learning can be split into two components: a feature extractor $F_{\theta}$ and a common classifier $G_{\phi}$ which grows with each new task $t + 1$ to include classes $C_{t + 1}$ . The feature extractor $F_{\theta}$ first maps the input $x$ to a deep feature vector $z = F_{\theta}(x) \in \mathbb{R}^{d}$ , and then the unified classifier $G_{\phi}(z) \in \mathbb{R}^{|C_t|}$ is a probability distribution over classes $C_t$ used to make predictions on input $x$ . + +Class-incremental learning requires that the model be capable of correctly classifying all samples from previous tasks at any training task – that is, when learning task $t$ , the model must not forget how to classify samples from classes from tasks $t' < t$ . Exemplar-free class-incremental learning additionally restricts models to learn each new task without access to samples from previous ones. This typically leads to learning objectives that minimize a loss function $\mathcal{L}$ defined on current training data $D_t$ : + +$$ +\mathcal {L} _ {t} ^ {\mathrm {C I L}} (x, y) = \mathcal {L} _ {\mathrm {c e}} \left(G _ {\phi_ {t}} \left(F _ {\theta_ {t}} (x)\right), y\right) + \mathcal {L} _ {t} ^ {\mathrm {M}}, \tag {1} +$$ + +where $\mathcal{L}_{\mathrm{ce}}$ is the standard cross-entropy classification loss + +![](images/64df8aa965c826eaa1f69e82985a9c539221e82d9ef4029bea46b31917fe3830.jpg) +Figure 3. We dilate the boundary map and apply a binary cross entropy loss at three stages in the CNN backbone to prevent mid-level attention from drifting into boundary regions. + +and $\mathcal{L}_t^{\mathrm{M}}$ is a method-specific loss that mitigates forgetting during incremental learning. Note that without $\mathcal{L}_t^{\mathrm{M}}$ , Eq. 1 reduces to fine-tuning on task $t$ . + +# 3.2. Boundary-guided Mid-level Saliency Drift Regularization + +Simply distilling attention between tasks does not consider task-adaptive attention. A model generates low-level representations of each input image $x$ (saliency and boundary maps in our experiments). We use CSNet [6] to generate salient regions and boundary maps, as it is lightweight and efficient, but any model producing saliency maps can be used in our framework (we explore other options in the Supplementary Material). To introduce plasticity on attention at these intermediate layers in the backbone, we use the generated boundary maps as a type of adaptive supervision as shown in Figure 3. We add a penalty term on object boundaries to avoid drift into the background. We first use 0.5 as the threshold to binarize the generated boundary map and dilate the maps by: + +$$ +B _ {d} (x) = \operatorname {D i l a t e} \left(A _ {b} (x), d\right), \tag {2} +$$ + +where $A_{b}(x)$ is the generated boundary map of image $x$ , which is converted from the saliency map with a Laplacian filter, and $d$ denotes the dilation radius applied on the boundary map for controlling the strictness of boundary-guided saliency. + +Rather than use a decoder at each layer as done for the low-level saliency maps described above, the mid-level saliency maps of our model are generated using Grad-CAM [39] at three stages of the CNN backbone (see Figure 2). We also experiment with several other methods for generating student saliency maps and report on them in the Supplementary Material. The dilated generated boundary map $B_{d}(x)$ is downsampled to match the feature map dimensions at these three stages in order to compare with the Grad-CAM generated saliency boundary maps. We use the binary cross entropy loss for supervision on dilated boundary regions. This loss is defined as: + +$$ +\mathcal {L} _ {t} ^ {\mathrm {d b s}} (x) = - \frac {\sum_ {j = 1} ^ {N} B _ {d} (x , j) \log (1 - S (x , j))}{\sum_ {j = 1} ^ {N} B _ {d} (x , j)}, \tag {3} +$$ + +where $S(x, j)$ denotes the saliency map of our model on image $x$ at pixel $j$ , $B_d(x, j)$ is the dilated generated boundary map at pixel $j$ , and $N$ is the number of pixels in $x$ . We compute this loss only within dilated boundary regions (i.e. where $B_d(x, j) = 1$ ). This loss helps the student saliency map avoid intersecting with the dilated teacher boundary region. + +# 3.3. Auxiliary Low-level Supervisions for EFCIL + +We propose to learn stable features from low-level stationary tasks shared across all incremental classification tasks during class-incremental learning. Low-level vision tasks like salient object detection require useful representations of input images. By learning these feature representations across tasks, the model can focus on key areas of input images and exploit learned, stable features with less representation drift since the low-level features change very little between tasks. + +Saliency map prediction is relevant to image classification since the foreground largely determines the results, while the background is comparatively less important. When learning new tasks with new classes, the background of images of new classes may contain new visual concepts that introduce undesirable noise and lead to forgetting of essential previous knowledge. The effectiveness of saliency features for learning classification tasks was demonstrated by Saliency Guided Training [17]. Additional supervision of salient region boundaries can aid salient object detection tasks for both segmentation and localization [12, 21, 24, 31, 50]. The positive interaction between these two tasks brings richer attention to features relevant to the main classification task. It can provide positive guidance in the form of stationary knowledge across class-incremental tasks. Some examples are illustrated in Figure 5. + +We incorporate low-level vision tasks into the network as an auxiliary supervision for enriching task-agnostic attention. The boundary map is computed with a Laplacian filter over the estimated saliency map. We add a decoder $D_{\psi}$ [22] after the backbone $F_{\theta}$ to predict low-level saliency and boundary maps for input images. The average L2 distance between the prediction and target is used as a low-level saliency distillation loss: + +$$ +\mathcal {L} _ {t} ^ {\mathrm {l m s}} (x) = \frac {| | D _ {\psi} (F _ {\theta} (x)) - A (x) | | _ {2}}{\sqrt {N}}, \tag {4} +$$ + +where $A(x)$ denotes the target low-level maps on input $x$ , consisting of a saliency map $A_{s}(x)$ and a boundary map $A_{b}(x)$ . $D_{\psi}(F_{\theta}(x))$ are combined saliency and boundary maps produced by the decoder, and $N$ is the number of pixels in the saliency maps. + +Algorithm 1 TASS Pseudocode +Input: The number of tasks $T$ , training samples $D_{t} = \{(x_{i},y_{i})\}_{i=1}^{N}$ of task $t$ , initial parameters $\Theta^0 = \{\theta_0,\phi_0,\psi_0\}$ containing parameters of feature extractor $F_{\theta}$ , classifier $G_{\phi}$ , and low-level decoder $D_{\psi}$ . +Output: Model $\Theta^T$ +1: for $t \in \{1,2,\dots,T\}$ do +2: $\Theta^t \gets \Theta^{t-1}$ +3: while not converged do +4: Sample $(x,y)$ from $D_{t}$ +5: $\mathcal{L}_t^{\mathrm{CIL}} \gets \text{SaliencyNoiseInjection}(x,y)$ +6: $\mathcal{L}_t^{\mathrm{Ims}} \gets \text{LowLevelMultitask}(x,A(x))$ +7: $S \gets \text{ComputeGradCAMSaliency}(x,y)$ +8: $\mathcal{L}_t^{\mathrm{obs}} \gets \text{DilatedBoundarySupervision}(S,A(x))$ +9: update $\Theta^t$ by minimizing $\mathcal{L}_t^{\mathrm{all}}$ from Eq. 5 +10: end while +11: end for + +# 3.4. Saliency Noise Injection + +Although we apply low-level distillation and dilated boundary supervision to maintain saliency representations across tasks, the model can still forget saliency on samples from previous tasks. To address this, we force the model to be able to recover correct saliency maps from injected saliency noise. + +At each task there is no available training data from previous or future tasks, and therefore we cannot directly know saliency drift on these samples. Instead of supervising the model with ground-truth saliency drift signals, we introduce saliency noise on random feature channels. We use a random ellipse to approximate the potential saliency drift in future tasks and the model is trained to denoise within each stage. Therefore the model can learn to effectively reduce real saliency drift. + +We generate elliptical noise using a very simple approach. There are six parameter dimensions: the center coordinate $(x, y)$ , the major and minor axis lengths $(a, b)$ , the rotation angle $\alpha$ , and the mask weight $w$ . A detailed explanation of this process is given in the Supplementary Material. With the help of dilated boundary supervision, each stage learns to eliminate this additional saliency noise and this aids generalization for future tasks and mitigates saliency forgetting in previous ones. + +# 3.5. Learning Objective and Training Algorithm + +The overall learning objective combines the low-level multi-task learning, dilated boundary supervision, and random saliency noise injection modules: + +$$ +\mathcal {L} _ {t} ^ {\text {a l l}} = \mathcal {L} _ {t} ^ {\text {C I L}} + \mathcal {L} _ {t} ^ {\text {l m s}} + \mathcal {L} _ {t} ^ {\text {d b s}}. \tag {5} +$$ + +Comparing this loss with Eq. 1, for TASS $\mathcal{L}_t^{\mathrm{M}} = \mathcal{L}_t^{\mathrm{CIL}} + \mathcal{L}_t^{\mathrm{Ins}}$ , thus incorporating saliency-aware supervision with + +the cross-entropy loss. The entire process is detailed in Algorithm 1. + +# 4. Experimental Results + +In this section we first describe our experimental setup and then compare TASS to state-of-the-art methods on several EFCIL benchmarks. In Section 4.3 we give further analysis of the various components of TASS. + +# 4.1. Experimental Setup + +We follow standard experimental protocols for EFCIL on three benchmark datasets. + +Datasets. We perform experiments on CIFAR-100 [19], Tiny-ImageNet [20], and ImageNet-Subset [8]. For most experiments, we train the model on half of the classes for the first task, and then equally distribute the remaining classes across each of the subsequent tasks. The convention we use is: $F + C \times T$ means that the first task contains $F$ classes, and the next $T$ tasks each contain $C$ classes. This is a common configuration for EFCIL used in both PASS [52] and SSRE [53]. + +State-of-the-art methods. Since we focus on EFCIL, we mainly compare with exemplar-free state-of-the-art approaches: SSRE [53], PASS [52], IL2A [51], EWC [18], LwF-MC [37], and MUC [28]. To demonstrate the effectiveness of our method, we also compare its performance with several exemplar-based methods like iCaRL (both nearest-mean and CNN) [37], EEIL [2], and LUCIR [16]. We also compare with RRR [11] integrated with SSRE, which focuses on preserving saliency using exemplar replay. + +Implementation details and metrics. We use ResNet-18 [15] as a feature extraction backbone. This is the same base network used in SSRE [53] and PASS [52], two state-of-the-art EFCIL approaches. We use the decoder in [22] to estimate low-level student saliency maps. All experiments are trained from scratch using Adam for 100 epochs with an initial learning rate 0.001. The learning rate is reduced by a factor 10 at epochs 45 and 90. For exemplar-based approaches, we use herding [37] to select and store 20 samples per class following common settings [16, 37]. We implement RRR [11] with SSRE to fairly compare it with TASS. For dilated boundary supervision, we set $d$ of the three mid-level boundary dilation stages to be $5\%$ , $10\%$ and $15\%$ of the image size. + +We report three common metrics for class incremental learning: the average and last top-1 accuracy, as well as average forgetting for all classes learned up to task $t$ . Denoting by $Acc_{i}$ the accuracy over all learned classes up to and including task $i$ , the average accuracy is defined as $Avg = \frac{\sum_{i=1}^{T} Acc_{i}}{T}$ , and the last accuracy is $Acc_{T}$ . Letting $a_{m,n}$ denotes the accuracy of task $n$ after learning task $m$ , + +
DatasetCIFAR-100Tiny-ImageNet
Method5 tasks10 tasks20 tasks5 tasks10 tasks20 tasks
MUC38.4519.5715.6518.9515.479.14
+TASS49.17 (+10.72)40.34 (+20.77)37.86 (+22.21)32.47 (+13.46)30.13 (+14.66)27.70 (+18.56)
IL2A55.1345.3245.2436.7734.5328.68
+TASS58.74 (+3.61)53.24 (+7.92)53.07 (+7.83)42.49 (+5.72)41.34 (+6.81)40.59 (+11.91)
PASS55.6749.0348.4841.5839.2832.78
+TASS59.10 (+3.43)54.45 (+5.42)52.37 (+3.89)44.05 (+2.47)43.06 (+3.78)42.57 (+9.79)
SSRE56.3355.0150.4741.4540.0739.25
+TASS59.26 (+2.93)57.93 (+2.92)53.78 (+3.31)44.13 (+2.68)43.86 (+3.79)43.55 (+4.30)
+ +Table 1. Performance gain in top-1 accuracy by applying TASS to other EFCIL methods in a plug-and-play way. Absolute gains are indicated in (red). + +
DatasetCIFAR100TinyImageNet
Setting5 tasks10 tasks20 tasks5 tasks10 tasks20 tasks
MethodAvg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓
E=20iCaRL-CNN†51.0740.1242.1348.6639.6545.6944.4335.4743.5434.6422.3136.8931.1521.1036.7027.9020.4645.12
iCaRL-NCM†58.5649.7424.9054.1945.1328.3250.5140.6835.5345.8633.4527.1543.2933.7528.8938.0428.8937.40
LUCIR†63.7855.0621.0062.3950.1425.1259.0748.7828.6549.1537.0920.6148.5236.8022.2542.8332.5533.74
EEIL†60.3752.3523.3656.0547.6726.6552.3441.5932.4047.1234.2425.5645.0134.2625.9140.5030.1435.04
RRR†66.4357.2218.0565.7855.7418.5962.4351.3518.4051.2042.2316.6749.5440.1221.6447.4635.5429.10
E=20LwFMC45.9336.1744.2327.4350.4717.0420.0715.8855.4629.1217.1254.2623.1012.3354.3717.438.7563.54
EWC16.049.3260.1714.708.4762.5314.128.2363.8918.8012.7167.5515.7710.1270.2312.398.4275.54
MUC49.4238.4540.2830.1919.5747.5621.2715.6552.6532.5817.9851.4626.6114.5450.2121.9512.7058.00
IL2A63.2255.1323.7857.6545.3230.4154.9045.2430.8448.1736.1425.4342.1035.2328.3236.7928.7435.46
PASS63.4755.6725.2061.8449.0330.2558.0948.4830.6149.5541.5818.0447.2939.2823.1142.0732.7830.55
SSRE65.8856.3318.3765.0455.0119.4861.7050.4718.3750.3941.6717.2548.9339.8922.5048.1739.7626.74
TASS (Ours)68.7559.2616.4267.4257.9317.7862.7653.7817.7855.1244.1315.4054.2143.8618.4752.7943.5522.51
+ +Table 2. Average, last top-1 accuracy, and forgetting on CIFAR-100 with different numbers of tasks. Replay-based methods storing 20 exemplars from each previous class are indicate with †. The best overall results are in bold. We run all experiments three times and report the mean for all metrics. + +
DatasetImageNet-Subset
Setting5 tasks10 tasks20 tasks
MethodAvg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓
LwF_TC34.8624.1049.3631.1820.0153.0427.5417.4256.07
MUC40.6527.8947.1335.0722.6552.1031.4420.1253.85
PASS63.1252.6122.4761.8050.4423.5755.2346.0726.73
SSRE69.5458.4617.2267.6957.5118.6061.2350.0523.22
TASS (Ours)74.3263.1414.3772.6057.9316.0968.7957.6018.41
+ +Table 3. Average, last top-1 accuracy, and forgetting on ImageNet-Subset for different numbers of tasks. We run all experiments three times and report means for all metrics. + +the forgetting measure $f_{k}^{i}$ of task $i$ after learning task $k$ is computed as $f_{k}^{i} = \max_{t\in 1,2,\ldots ,k - 1}(a_{t,i} - a_{k,i})$ . The average forgetting $F_{k}$ is defined as $F_{k} = \frac{1}{k - 1}\sum_{i = 1}^{k - 1}f_{k}^{i}$ . + +# 4.2. Comparison with the State-of-the-art + +We compare TASS with the state-of-the-art on CIFAR-100 in Table 2. TASS outperforms all exemplar-free approaches. For exemplar-based methods like iCaRL [37], EEIL [2], and LUCIR [16], our method still has significantly better + +performance. On longer sequences (i.e. 10 and 20 tasks), our method significantly reduces forgetting when learning new classes compared to other EFCIL methods. TASS outperforms the best method SSRE by about $3\%$ on the last task. This performance improvement can be also observed in terms of average forgetting. + +As we see in Table 3 and Figure 4 for Tiny-ImageNet and ImageNet-Subset, although our method has similar top-1 accuracy on the first task in Figure 4, it has better performance at most intermediate tasks and also the final one. For longer sequences in Figure 4, the gap between our method and the best baseline is largely consistent, showing the effectiveness of our method at mitigating forgetting. The performance gain in Table 3 is larger on Tiny-ImageNet and ImageNet-Subset compared to CIFAR100, and this demonstrates that our method generalizes to datasets with larger images and object scales. It is worth mentioning that TASS also produces results with smaller variance. We believe this to be due to TASS reducing saliency drift to background regions, which may include random noise. + +Plug-and-play with other EFCIL methods. Some existing EFCIL methods, like PASS [52], IL2A [51] and + +![](images/2786bca8e8231d2a3146062959cbdc48efc3ddaf9cc1549f1e2f25be0b9c2f1b.jpg) + +![](images/3c79aabdedbb864260f53e8611d4a22a80a3622f33ba36d1e3dc940a1bb95f0a.jpg) + +![](images/caf86358f27db210f3dd0db47cff36829e498d5eb40ded3758f7de87657ec83d.jpg) + +![](images/e101741421122b30b8d6925fd5d3306b006c8a595aec14bbcf7e44176725f797.jpg) +Figure 4. Results on Tiny-ImageNet and ImageNet-Subset for different numbers of tasks. Our method outperforms others, especially on longer task sequences (i.e. more, but smaller, tasks). + +![](images/873b053d588788e54782dde677d13d6f5aa5f43bcda3b49cd9e13fc67e69ab18.jpg) + +![](images/420c6ef776df5e91f24ac5dd6ba74001c0d9d6279dde2fe2db7aa838ef0b2432.jpg) + +SSRE [53], focus on reducing forgetting via embedding regularization. Considering the importance of saliency to image classification, it is natural to consider whether TASS can be integrated into these methods. The results in Table 1 show the performance gain brought by this integration. Adding TASS doubles the performance for MUC in many cases and significantly improves IL2A and PASS. When we incorporate it into the best baseline SSRE, it yields a consistent gain of about $3\%$ . These results clearly show that TASS, by explicitly mitigating saliency drift, is complementary to other methods in relieving forgetting. They additionally demonstrate the significance of saliency drift as a cause of catastrophic forgetting in EFCIL. + +# 4.3. Additional Analysis + +In this section we take a deeper look at the method we propose. If not specified, the results are produced using TASS integrated into SSRE [53]. + +Ablation Study. We performed ablations using the 10-task setting on CIFAR-100 (see Table 4). We ablate on both PASS [52] and SSRE [53]. Low-level multi-task supervision is crucial and improves by $2.2\%$ (PASS) and $1.2\%$ (SSRE). Dilated boundary supervision further boosts performance by about $1 - 2\%$ . Saliency noise injection is also helpful for both methods and improves by $1.5\%$ for PASS. In total, TASS improves baselines by $5.5\%$ and $2.9\%$ points, respectively. Note that SSRE is the previous state-of-the-art method and TASS outperforms it by a large margin. + +Low-level Multi-task Supervision. To analyze the effect of our proposed low-level saliency supervision, we performed experiments on ImageNet-Subset in the 5 and 10 + +
Method & TasksL1msL1bsSNIAccuracy
Baseline (PASS)49.0
Variants✓51.2
✓✓53.0
✓✓✓54.5
Baseline (SSRE)55.0
Variants✓56.2
✓✓57.3
✓✓✓57.9
+ +Table 4. Ablations on each TASS component. Experiments are on CIFAR-100 in the 10-task setting and we report the top-1 accuracy in % for TASS integrated into PASS and SSRE. $\mathcal{L}_{\mathrm{ubs}}$ (Eq. 3), $\mathcal{L}_{\mathrm{lms}}$ (Eq. 4), and SNI denote the three TASS components: Dilated Boundary Supervision, Low-level Multi-task Supervision, and Saliency Noise Injection. + +task settings. We first plot the loss across tasks in Figure 5(c). After learning to predict boundary and saliency maps in the first task, the network maintains good performance for the rest of the 5 task sequence. This shows that the low-level tasks are stable during continual learning. Furthermore, we visualize the results of saliency and boundary map prediction during incremental learning in Figure 5(a-b). We give some examples of predicted boundary and saliency maps after learning different tasks. Although CIL involves samples of different classes, we see that the low-level outputs are relatively stable and class agnostic. Since the model is able to stably predict these low-level features across tasks, it therefore can preserve useful prior knowl + +![](images/1476576f2e990c0f32b9db0382198a45131e1d2325330f4a31b6fe8934e42324.jpg) +(a) + +![](images/d28aeb54d478b5f5b10363ffd05810ec4f00605d06636e532b5bd843197fbe54.jpg) +(b) + +![](images/7a8e433045b1b2e3c350917ccc79574eaa71dcb785bece9d4d91380e06f7038a.jpg) +(c) +Figure 5. Visualization of the saliency (a) and boundary (b) maps from our student encoder-decoder network with original images from different tasks at different stages of incremental learning. Our method produces stable low-level results while reducing forgetting in classification. (c) The MAE loss between the student and teacher network across tasks. + +edge for continual learning. + +
Metric & MethodAvg↑Last↑
FeTrIL [35]65.2056.34
SOPE [54]65.8456.80
PRAKA [40]68.8659.20
TASS (SSRE)67.4257.93
TASS (PRAKA)69.7060.04
+ +Table 5. Average and last accuracy on CIFAR-100 10-task setting. + +
Metric & MethodAvg↑
SOPE [54]60.20
FeTrIL [35]65.00
TASS (FeTrIL)66.03
+ +TASS with more methods and benchmarks. Due to the strong generalization capability of our method, we have also applied our TASS paradigm to PRAKA [40], resulting in further performance improvements. We give specific experimental comparisons in Table 5. We also compare with FeTrIL [35] in Table 5 and Table 6 above. Experiments on ImageNet-Full show consistent improvement in Table 6 above. + +Quantitative analysis of saliency drift. We measure the intersection over union (IoU) between the self-attention map of the last layer in DINO and the Grad-CAM saliency maps for both SSRE and $\mathrm{SSRE + TASS}$ . As shown in Table 7, TASS significantly reduces saliency drift in the student model, which again shows the effectiveness of our method for maintaining saliency during CIL. + +Table 6. Average accuracy on the ImageNet-Full 10-task setting. + +
Task14710
SSRE47.450.156.678.5
SSRE+TASS75.282.388.590.1
+ +Table 7. Average IoU (\%) between DINO self-attention maps and student model saliency in the CIFAR-100 10-task. + +# 5. Conclusions + +In this paper, we propose an approach of task-adaptive saliency guidance for EFCIL. The insight behind TASS is to guide the model to focus on salient regions and inhibit saliency drift. We show that robust saliency guidance is crucial to mitigating forgetting across tasks. Experiments demonstrate that TASS is effective and surpasses the state-of-the-art. TASS can be easily combined with other methods, leading to large performance gains over baselines. Qualitative results also show that low-level tasks are stable across different tasks, resulting in less forgetting. + +Limitations and Future Work. In this work, we introduce the auxiliary saliency knowledge across tasks for better CIL performance, which may raise the concern of unfair comparison. However, we believe it is important to explore external knowledge to make the CIL system applicable in real applications. Additionally, our method can be easily integrated with other methods and the low-level auxiliary knowledge takes negligible cost and is easy to obtain. We will explore leveraging other forms of knowledge in future work, such as pre-trained visual and large language models. + +Acknowledgments This work is funded by NSFC (NO. 62206135, 62225604), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), and the Fundamental Research Funds for the Central Universities (Nankai University, 070-63233085). Computation is supported by the Supercomputing Center of Nankai University. + +# References + +[1] Eden Belouadah, Adrian Popescu, and Ioannis Kanellos. A comprehensive study of class incremental learning algorithms for visual tasks. Neural Networks, 135:38-54, 2021. 1, 2 +[2] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In ECCV, 2018. 1, 5, 6 +[3] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9233-9242, 2020. 2 +[4] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. In ICLR, 2019. 2 +[5] Hanting Chen, Yunhe Wang, Chang Xu, Zhaohui Yang, Chuanjian Liu, Boxin Shi, Chunjing Xu, Chao Xu, and Qi Tian. Data-free learning of student networks. In ICCV, 2019. 2 +[6] Ming-Ming Cheng, Shanghua Gao, Ali Borji, Yong-Qiang Tan, Zheng Lin, and Meng Wang. A highly efficient model to study the semantics of salient object detection. IEEE TPAMI, 2021. 4 +[7] Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. IEEE TPAMI, 2021. 2 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5 +[9] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In CVPR, 2019. 2 +[10] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020. 1, 2 +[11] Sayna Ebrahimi, Suzanne Petryk, Akash Gokul, William Gan, Joseph E. Gonzalez, Marcus Rohrbach, and trevor darrell. Remembering for the right reasons: Explanations reduce catastrophic forgetting. In ICLR, 2021. 2, 5 +[12] Deng-Ping Fan, Ge-Peng Ji, Peng Xu, Ming-Ming Cheng, Christos Sakaridis, and Luc Van Gool. Advances in deep concealed scene understanding. Visual Intelligence, 1(1):16, 2023. 4 +[13] Tao Feng, Mang Wang, and Hangjie Yuan. Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9427-9436, 2022. 2 +[14] Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv preprint arXiv:1312.6211, 2013. 1 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. + +Deep residual learning for image recognition. In CVPR, 2016.5 +[16] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 5, 6 +[17] Aya Abdelsalam Ismail, Hector Corrada Bravo, and Soheil Feizi. Improving deep learning interpretability by saliency guided training. NeurIPS, 2021. 4 +[18] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 2017. 2, 5 +[19] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, 2009. 5 +[20] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 5 +[21] Hao Li, Dingwen Zhang, Nian Liu, Lechao Cheng, Yalun Dai, Chao Zhang, Xinggang Wang, and Junwei Han. Boosting low-data instance segmentation by unsupervised pretraining with saliency prompt. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15485-15494, 2023. 4 +[22] Xiangtai Li, Ansheng You, Zhen Zhu, Houlong Zhao, Maoke Yang, Kuiyuan Yang, and Yunhai Tong. Semantic flow for fast and accurate scene parsing. In ECCV, 2020. 4, 5 +[23] Zhizhong Li and Derek Hoiem. Learning without forgetting. In ECCV, 2016. 2 +[24] Liqiang Lin, Pengdi Huang, Chi-Wing Fu, Kai Xu, Hao Zhang, and Hui Huang. On learning the right attention point for feature enhancement. Science China Information Sciences, 66(1):112107, 2023. 4 +[25] Zheng Lin, Zhao Zhang, Zi-Yue Zhu, Deng-Ping Fan, and Xia-Lei Liu. Sequential interactive image segmentation. Computational Visual Media, 9(4):753-765, 2023. 2 +[26] Xialei Liu, Marc Masana, Luis Herranz, Joost Van de Weijer, Antonio M Lopez, and Andrew D Bagdanov. Rotate your networks: Better weight consolidation and less catastrophic forgetting. In ICPR, 2018. 2 +[27] Xialei Liu, Yu-Song Hu, Xu-Sheng Cao, Andrew D Bagdanov, Ke Li, and Ming-Ming Cheng. Long-tailed class incremental learning. In European Conference on Computer Vision, pages 495-512. Springer, 2022. 2 +[28] Yu Liu, Sarah Parisot, Gregory Slabaugh, Xu Jia, Ales Leonardis, and Tinne Tuytelaars. More classifiers, less forgetting: A generic multi-classifier paradigm for incremental learning. In ECCV. Springer, 2020. 2, 5 +[29] Yuyang Liu, Yang Cong, Dipam Goswami, Xialei Liu, and Joost van de Weijer. Augmented box replay: Overcoming foreground shift for incremental object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11367-11377, 2023. 2 +[30] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. NeurIPS, 2017. 2 + +[31] Zhouzhou Ma, Guanghua Gu, and Wenrui Zhao. Self-attention guidance based crowd localization and counting. Machine Intelligence Research, pages 1–17, 2024. 4 +[32] Arun Mallya and Svetlana Lazebnik. Packet: Adding multiple tasks to a single network by iterative pruning. In CVPR, 2018. 2 +[33] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost van de Weijer. Class incremental learning: survey and performance evaluation on image classification. IEEE TPAMI, 2022. 1 +[34] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of learning and motivation*, pages 109–165. Elsevier, 1989. 1 +[35] Grégoire Petit, Adrian Popescu, Hugo Schindler, David Picard, and Bertrand Delezoise. Fetril: Feature translation for exemplar-free class-incremental learning. In WACV, 2023. 8 +[36] Quang Pham, Chenghao Liu, and Steven Hoi. Dualnet: Continual learning, fast and slow. NeurIPS, 2021. 1 +[37] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017. 1, 2, 5, 6 +[38] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. In ICLR, 2019. 2 +[39] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 4 +[40] Wuxuan Shi and Mang Ye. Prototype reminiscence and augmented asymmetric knowledge aggregation for nonexemplar class-incremental learning. In ICCV, 2023. 8 +[41] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In ICCV, 2021. 2 +[42] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019. 1, 2 +[43] Jia-Wen Xiao, Chang-Bin Zhang, Jiekang Feng, Xialei Liu, Joost van de Weijer, and Ming-Ming Cheng. Endpoints weight fusion for class incremental semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7204-7213, 2023. 2 +[44] Ju Xu and Zhanxing Zhu. Reinforced continual learning. NeurIPS, 2018. 2 +[45] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In CVPR, 2020. 1, 2 +[46] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In CVPR, 2020. 1, 2 + +[47] Lu Yu, Xialei Liu, and Joost Van de Weijer. Self-training for class-incremental semantic segmentation. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2 +[48] Jiang-Tian Zhai, Xialei Liu, Andrew D Bagdanov, Ke Li, and Ming-Ming Cheng. Masked autoencoders are efficient class incremental learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19104-19113, 2023. 2 +[49] Junting Zhang, Jie Zhang, Shalini Ghosh, Dawei Li, Serafettin Tasci, Larry Heck, Heming Zhang, and C-C Jay Kuo. Class-incremental learning via deep model consolidation. In WACV, 2020. 2 +[50] Jia-Xing Zhao, Jiang-Jiang Liu, Deng-Ping Fan, Yang Cao, Jufeng Yang, and Ming-Ming Cheng. Egnet: Edge guidance network for salient object detection. In ICCV, 2019. 4 +[51] Fei Zhu, Zhen Cheng, Xu-yao Zhang, and Cheng-lin Liu. Class-incremental learning via dual augmentation. NeurIPS, 2021. 1, 2, 3, 5, 6 +[52] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In CVPR, 2021. 1, 2, 3, 5, 6, 7 +[53] Kai Zhu, Wei Zhai, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Self-sustaining representation expansion for nonexemplar class-incremental learning. In CVPR, 2022. 2, 3, 5, 7 +[54] Kai Zhu, Kecheng Zheng, Ruili Feng, Deli Zhao, Yang Cao, and Zheng-Jun Zha. Self-organizing pathway expansion for non-exemplar class-incremental learning. In ICCV, 2023. 8 \ No newline at end of file diff --git a/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/images.zip b/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..59c20ae376f3c37cd3873a93960e59bf101fc2a2 --- /dev/null +++ b/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd3638f590100dfbe24783a30674738fc211f7d41fba58dea8c2d071347c7089 +size 618523 diff --git a/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/layout.json b/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4af3dcd3f7dba0b96c99e7c0f610b8177da74fd4 --- /dev/null +++ b/2024/Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning/layout.json @@ -0,0 +1,8852 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 547, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 547, + 123 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 547, + 123 + ], + "type": "text", + "content": "Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 141, + 541, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 541, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 541, + 159 + ], + "type": "text", + "content": "Xialei Liu " + }, + { + "bbox": [ + 56, + 141, + 541, + 159 + ], + "type": "inline_equation", + "content": "^{2,1,*}" + }, + { + "bbox": [ + 56, + 141, + 541, + 159 + ], + "type": "text", + "content": " Jiang-Tian Zhai " + }, + { + "bbox": [ + 56, + 141, + 541, + 159 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 56, + 141, + 541, + 159 + ], + "type": "text", + "content": " Andrew D. Bagdanov " + }, + { + "bbox": [ + 56, + 141, + 541, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 56, + 141, + 541, + 159 + ], + "type": "text", + "content": " Ke Li " + }, + { + "bbox": [ + 56, + 141, + 541, + 159 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 56, + 141, + 541, + 159 + ], + "type": "text", + "content": " Ming-Ming Cheng " + }, + { + "bbox": [ + 56, + 141, + 541, + 159 + ], + "type": "inline_equation", + "content": "^{2,1,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 159, + 527, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 527, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 527, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 159, + 527, + 186 + ], + "type": "text", + "content": "VCIP, CS, Nankai University " + }, + { + "bbox": [ + 56, + 159, + 527, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 56, + 159, + 527, + 186 + ], + "type": "text", + "content": "NKIARI, Shenzhen Futian " + }, + { + "bbox": [ + 56, + 159, + 527, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 56, + 159, + 527, + 186 + ], + "type": "text", + "content": "MICC, University of Florence " + }, + { + "bbox": [ + 56, + 159, + 527, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 56, + 159, + 527, + 186 + ], + "type": "text", + "content": "Tencent Youtu Lab {xialei, cmm}@nankai.edu.cn, {jtzhai30,tristanli.sh}@gmail.com, andrew.bagdanov@unifi.it" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 238, + 290, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 490 + ], + "type": "text", + "content": "Exemplar-free Class Incremental Learning (EFCIL) aims to sequentially learn tasks with access only to data from the current one. EFCIL is of interest because it mitigates concerns about privacy and long-term storage of data, while at the same time alleviating the problem of catastrophic forgetting in incremental learning. In this work, we introduce task-adaptive saliency for EFCIL and propose a new framework, which we call Task-Adaptive Saliency Supervision (TASS), for mitigating the negative effects of saliency drift between different tasks. We first apply boundary-guided saliency to maintain task adaptivity and plasticity on model attention. Besides, we introduce task-agnostic low-level signals as auxiliary supervision to increase the stability of model attention. Finally, we introduce a module for injecting and recovering saliency noise to increase the robustness of saliency preservation. Our experiments demonstrate that our method can better preserve saliency maps across tasks and achieve state-of-the-art results on the CIFAR-100, Tiny-ImageNet, and ImageNet-Subset EFCIL benchmarks. Code is available at https://github.com/scok30/tass." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 511, + 128, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 128, + 523 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 128, + 523 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 531, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 639 + ], + "type": "text", + "content": "Deep neural networks achieve state-of-the-art performance on many computer vision tasks. However, most of these tasks consider a static world in which tasks are well-defined, stationary, and all training data is available in a single training session. The real world consists of dynamically changing environments and data distributions, which - especially given the computational burden of training large CNNs - has led to renewed interest in learning new tasks incrementally while avoiding catastrophic forgetting [14, 34]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 639, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 287, + 687 + ], + "type": "text", + "content": "Class Incremental Learning (CIL) [1, 33] is a scenario that considers the possibility of adding new classes to already-trained models. Most CIL methods rely on a memory buffer to store exemplars from past tasks [2, 10, 37, 42]." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 212, + 545, + 327 + ], + "blocks": [ + { + "bbox": [ + 307, + 212, + 545, + 327 + ], + "lines": [ + { + "bbox": [ + 307, + 212, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 545, + 327 + ], + "type": "image", + "image_path": "fd5dab2fb9d2796dc711f57546f48ac75b4224e5bfdec6811129787de84e050a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 328, + 547, + 373 + ], + "lines": [ + { + "bbox": [ + 305, + 328, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 328, + 547, + 373 + ], + "type": "text", + "content": "Figure 1. We propose the TASS method, which can be directly applied to many recent exemplar-free class incremental learning methods, resulting in a significant improvement in EFCIL classification accuracy and a reduction in catastrophic forgetting." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 389, + 546, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 546, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 546, + 473 + ], + "type": "text", + "content": "In this paper, we consider Exemplar-Free Class Incremental Learning (EFCIL), which is a more challenging setting in which no data from previous tasks is retained. This is a realistic scenario and of great interest due to privacy concerns or restrictions on the long-term storage of data. The inability to retain examples from past tasks, however, significantly exacerbates the problem of catastrophic forgetting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 474, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 653 + ], + "type": "text", + "content": "There are several recent works that consider the EFCIL problem. DeepInversion [45] inverts trained networks from random noise to generate images as exemplars and mixes them with current task samples for training. SDC [46] updates prototypes of each learned class by hypothesizing that semantic drift of classes from previous tasks can be approximated and estimated using new data. Other previous works propose representation learning methods for overcoming catastrophic forgetting [51, 52]. As pointed out in IL2A [51], learning better representations can reduce representation bias when transferred to new tasks. Incorporating self-supervised learning tasks, such as Barlow Twins [36] and rotation prediction [52], has also been proposed to achieve more stable representations and alleviate forgetting." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "CNNs naturally learn to attend to features that are discriminative for the tasks they are trained to solve. Catastrophic forgetting also occurs in EFCIL due to the model's attention to salient features drifting to features specific to the new task. Standard regularization approaches do little" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 693, + 192, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 693, + 192, + 703 + ], + "spans": [ + { + "bbox": [ + 57, + 693, + 192, + 703 + ], + "type": "text", + "content": "*The first two authors contribute equally." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23954" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "text", + "content": "to prevent this saliency drift when learning new tasks. One direct method of regularizing saliency is to apply distillation on saliency maps of old samples [11]. However, this is complicated by the inability to save samples from previous tasks in the EFCIL setting. Another method is to apply saliency distillation between current task samples and previous task attention [9]. This method however suffers from the semantic gap between current and old classes when enforcing saliency consistency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 314 + ], + "type": "text", + "content": "A lack of saliency regularization may lead to attention drifting toward the background in future tasks causing forgetting. Besides, simply applying distillation on attention [9] fails to offer plasticity and is susceptible to attention forgetting, which is a crucial factor of knowledge forgetting. In comparison, our Task-Adaptive Saliency Supervision (TASS) approach aims at keeping saliency focused on incrementally learned tasks (for more details, see Section 4), while maintaining its plasticity and stability. It improves many previous EFCIL methods with higher performance by supervising their attention, as shown in Figure 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 531 + ], + "type": "text", + "content": "Specifically, TASS integrates three components to address this issue. Firstly, we apply dilated boundary maps to prevent saliency drift across object boundaries at intermediate layers. Since saliency drift typically occurs across tasks, encouraging the model to focus on significant foreground regions through dilated boundary supervision reduces the likelihood of saliency shifting toward the background, allowing the model to adaptively select attention areas within the foreground that are relevant to the task. Secondly, to simultaneously enhance the stability of the model's attention across tasks, we add a task-agnostic low-level auxiliary supervision task to the class-incremental framework, which is closely related to our core EFCIL task since image classification has been shown to help models localize the most salient areas within an image. Finally, we propose a module to inject saliency noise into certain feature channels and train the network to denoise them, helping the network further resist attention drift across tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "The main contributions of this work are: (i) We provide new insight into task-adaptive saliency supervision under EFCIL settings; we also show the negative effect of methods with no or trivial saliency supervision, which illustrates the superiority of our method and motivates the need for saliency drift mitigation in EFCIL. (ii) We propose the Task-Adaptive Saliency Supervision (TASS) with three components that combine to mitigate the saliency drift problem. (iii) We show that TASS can be easily integrated into other state-of-the-art methods, such as MUC [28], IL2A [51], PASS [52], SSRE [53], leading to significant performance gains. (iv) Our experiments demonstrate that TASS outperforms all existing EFCIL methods and even several exemplar-based methods on the CIFAR-100, TinyImageNet, and ImageNet-Subset EFCIL benchmarks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 91, + 545, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 545, + 114 + ], + "type": "text", + "content": "We first discuss previous work on incremental learning from the recent literature and then describe work on EFCIL." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 121, + 432, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 121, + 432, + 133 + ], + "spans": [ + { + "bbox": [ + 306, + 121, + 432, + 133 + ], + "type": "text", + "content": "2.1. Incremental Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 139, + 545, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 139, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 139, + 545, + 402 + ], + "type": "text", + "content": "A variety of methods have been proposed for incremental learning in the past few years [1, 7]. Recent works can be coarsely grouped into three categories: replay-based, regularization-based, and parameter-isolation methods. Replay-based methods mitigate the task-recency bias by retaining training samples from previous tasks [37, 48]. In addition to replaying samples, BiC [42], PODNet [10], iCaRL [37] and LTCIL [27] apply a distillation loss to prevent forgetting and enhance model stability. GEM [30], AGEM [4], and MER [38] exploit past-task exemplars by modifying gradients on current training samples to match old samples. Rehearsal may cause models to overfit to stored samples. Regularization-based approaches such as LwF [23], EWC, R-EWC [18, 26], and DMC [49] offer ways to learn better representations while leaving enough plasticity for adaptation to new tasks. Parameter-isolation methods [32, 44] use models with different computational graphs for each task. With the help of growing models, new model branches mitigate catastrophic forgetting at the cost of more parameters and computational cost. It is also widely studied in other fields, such as semantic segmentation [3, 25, 43, 47] and object detection [13, 29]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 403, + 545, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 546 + ], + "type": "text", + "content": "As for saliency-guided incremental learning, LwM [9] regularizes the saliency activations of previous classes on current task data. However, there exist semantic gaps between current classes and old ones, which results in an inaccurate distillation target for preserving saliency activations on old samples. RRR [11] directly saves the GradCAM saliency activations of each sample in a replay buffer and applies distillation to memorize this old knowledge, which requires storing additional samples during incremental learning. Despite these initial works on saliency-guided incremental learning, saliency drift still remains problematic and leads to catastrophic forgetting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 552, + 529, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 529, + 565 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 529, + 565 + ], + "type": "text", + "content": "2.2. Exemplar-free Class Incremental Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Compared to conventional class incremental learning, exemplar-free class incremental learning is more appropriate for applications where training data is sensitive and may not be stored in perpetuity. DAFL [5] uses a GAN to generate synthetic samples from past tasks as an alternative to storing actual data. DeepInversion, which inverts trained networks using random noise to generate images, is another popular EFCIL method [45]. Always Be Dreaming further improves on DeepInversion for EFCIL [41]. SDC attempts to overcome the problems caused by semantic drift when training new tasks on old class samples [46]. It directly estimates prototypes of each learned class to use in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "23955" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 73, + 496, + 350 + ], + "blocks": [ + { + "bbox": [ + 102, + 73, + 496, + 350 + ], + "lines": [ + { + "bbox": [ + 102, + 73, + 496, + 350 + ], + "spans": [ + { + "bbox": [ + 102, + 73, + 496, + 350 + ], + "type": "image", + "image_path": "bd01f5a77150e3636fa64febeed139f3cdf717582aa24e4555829237e5c64cf2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 357, + 547, + 413 + ], + "lines": [ + { + "bbox": [ + 46, + 357, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 547, + 413 + ], + "type": "text", + "content": "Figure 2. Overall framework of Task-Adaptive Saliency Supervision (TASS). We apply a low-level model to generate saliency and boundary maps. The boundary map is dilated and downsampled to provide supervision at different stages of the encoder. A decoder is attached after the encoder for low-level distillation, which serves as stationary task-agnostic saliency guidance. To prevent saliency drift in later training phases, we introduce saliency noise into each encoder stage. The model is trained to denoise and reduce the saliency drift on current data in future phases. TASS can be integrated into an EFCIL approach to provide robust saliency guidance across incremental tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 434, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 289, + 555 + ], + "type": "text", + "content": "a nearest class mean classifier. PASS [52] and IL2A [51] are prototype-based replay methods for efficient and effective EFCIL. Since these prototypes are features computed from past training samples, image exemplars need not be retained. SSRE [53] introduces a re-parameterization method to trade-off between old and new knowledge. Our task-Adaptive Saliency Supervision (TASS) method uses three new components to reduce saliency drift in EFCIL and is complementary to several of the approaches mentioned above." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 567, + 246, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 246, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 246, + 582 + ], + "type": "text", + "content": "3. Task-Adaptive Saliency Supervision" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 588, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 287, + 648 + ], + "type": "text", + "content": "We first define the Exemplar-Free Class Incremental Learning (EFCIL) scenario. Then we describe our TASS approach including dilated boundary supervision, auxiliary low-level supervision, and saliency noise injection. Our overall framework is illustrated in Figure 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 272, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 272, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 272, + 671 + ], + "type": "text", + "content": "3.1. Exemplar-free Class-Incremental Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "Class-incremental learning aims to sequentially learn tasks consisting of disjoint classes of samples. Let " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "t \\in \\{1, 2, \\dots, T\\}" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " denote the incremental learning tasks. The training data " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "D_{t}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 434, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 434, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 434, + 545, + 468 + ], + "type": "text", + "content": "for each task contains classes " + }, + { + "bbox": [ + 305, + 434, + 545, + 468 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 305, + 434, + 545, + 468 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 434, + 545, + 468 + ], + "type": "inline_equation", + "content": "N_t" + }, + { + "bbox": [ + 305, + 434, + 545, + 468 + ], + "type": "text", + "content": " training samples " + }, + { + "bbox": [ + 305, + 434, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\{(x_t^i, y_t^i)\\}_{i=1}^{N_t}" + }, + { + "bbox": [ + 305, + 434, + 545, + 468 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 434, + 545, + 468 + ], + "type": "inline_equation", + "content": "x_t^i" + }, + { + "bbox": [ + 305, + 434, + 545, + 468 + ], + "type": "text", + "content": " are images and " + }, + { + "bbox": [ + 305, + 434, + 545, + 468 + ], + "type": "inline_equation", + "content": "y_t^i \\in C_t" + }, + { + "bbox": [ + 305, + 434, + 545, + 468 + ], + "type": "text", + "content": " are their labels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "text", + "content": "Most deep networks applied to class-incremental learning can be split into two components: a feature extractor " + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "text", + "content": " and a common classifier " + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "inline_equation", + "content": "G_{\\phi}" + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "text", + "content": " which grows with each new task " + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "text", + "content": " to include classes " + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "inline_equation", + "content": "C_{t + 1}" + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "text", + "content": ". The feature extractor " + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "text", + "content": " first maps the input " + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "text", + "content": " to a deep feature vector " + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "inline_equation", + "content": "z = F_{\\theta}(x) \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "text", + "content": ", and then the unified classifier " + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "inline_equation", + "content": "G_{\\phi}(z) \\in \\mathbb{R}^{|C_t|}" + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "text", + "content": " is a probability distribution over classes " + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "text", + "content": " used to make predictions on input " + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 470, + 546, + 564 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "type": "text", + "content": "Class-incremental learning requires that the model be capable of correctly classifying all samples from previous tasks at any training task – that is, when learning task " + }, + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "type": "text", + "content": ", the model must not forget how to classify samples from classes from tasks " + }, + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "type": "inline_equation", + "content": "t' < t" + }, + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "type": "text", + "content": ". Exemplar-free class-incremental learning additionally restricts models to learn each new task without access to samples from previous ones. This typically leads to learning objectives that minimize a loss function " + }, + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "type": "text", + "content": " defined on current training data " + }, + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "type": "inline_equation", + "content": "D_t" + }, + { + "bbox": [ + 304, + 565, + 547, + 673 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 335, + 680, + 545, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 680, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 335, + 680, + 545, + 695 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} ^ {\\mathrm {C I L}} (x, y) = \\mathcal {L} _ {\\mathrm {c e}} \\left(G _ {\\phi_ {t}} \\left(F _ {\\theta_ {t}} (x)\\right), y\\right) + \\mathcal {L} _ {t} ^ {\\mathrm {M}}, \\tag {1}", + "image_path": "4e164f796f58a035e540a6b918bcaf8ba6ee68a7b172cd755088be5aa97ab28e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ce}}" + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": " is the standard cross-entropy classification loss" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23956" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 74, + 286, + 135 + ], + "blocks": [ + { + "bbox": [ + 53, + 74, + 286, + 135 + ], + "lines": [ + { + "bbox": [ + 53, + 74, + 286, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 74, + 286, + 135 + ], + "type": "image", + "image_path": "64df8aa965c826eaa1f69e82985a9c539221e82d9ef4029bea46b31917fe3830.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 141, + 287, + 175 + ], + "lines": [ + { + "bbox": [ + 46, + 141, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 141, + 287, + 175 + ], + "type": "text", + "content": "Figure 3. We dilate the boundary map and apply a binary cross entropy loss at three stages in the CNN backbone to prevent mid-level attention from drifting into boundary regions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 286, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 286, + 228 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 46, + 192, + 286, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t^{\\mathrm{M}}" + }, + { + "bbox": [ + 46, + 192, + 286, + 228 + ], + "type": "text", + "content": " is a method-specific loss that mitigates forgetting during incremental learning. Note that without " + }, + { + "bbox": [ + 46, + 192, + 286, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t^{\\mathrm{M}}" + }, + { + "bbox": [ + 46, + 192, + 286, + 228 + ], + "type": "text", + "content": ", Eq. 1 reduces to fine-tuning on task " + }, + { + "bbox": [ + 46, + 192, + 286, + 228 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 192, + 286, + 228 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 234, + 286, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 234, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 234, + 286, + 258 + ], + "type": "text", + "content": "3.2. Boundary-guided Mid-level Saliency Drift Regularization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 264, + 286, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 286, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 286, + 432 + ], + "type": "text", + "content": "Simply distilling attention between tasks does not consider task-adaptive attention. A model generates low-level representations of each input image " + }, + { + "bbox": [ + 46, + 264, + 286, + 432 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 264, + 286, + 432 + ], + "type": "text", + "content": " (saliency and boundary maps in our experiments). We use CSNet [6] to generate salient regions and boundary maps, as it is lightweight and efficient, but any model producing saliency maps can be used in our framework (we explore other options in the Supplementary Material). To introduce plasticity on attention at these intermediate layers in the backbone, we use the generated boundary maps as a type of adaptive supervision as shown in Figure 3. We add a penalty term on object boundaries to avoid drift into the background. We first use 0.5 as the threshold to binarize the generated boundary map and dilate the maps by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 437, + 286, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 437, + 286, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 286, + 450 + ], + "type": "interline_equation", + "content": "B _ {d} (x) = \\operatorname {D i l a t e} \\left(A _ {b} (x), d\\right), \\tag {2}", + "image_path": "022533a6d257e2af63d3ac66058f2d51b3d213336ea7414919ba46d5b5d8fed4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 456, + 286, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 286, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 456, + 286, + 514 + ], + "type": "inline_equation", + "content": "A_{b}(x)" + }, + { + "bbox": [ + 46, + 456, + 286, + 514 + ], + "type": "text", + "content": " is the generated boundary map of image " + }, + { + "bbox": [ + 46, + 456, + 286, + 514 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 456, + 286, + 514 + ], + "type": "text", + "content": ", which is converted from the saliency map with a Laplacian filter, and " + }, + { + "bbox": [ + 46, + 456, + 286, + 514 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 456, + 286, + 514 + ], + "type": "text", + "content": " denotes the dilation radius applied on the boundary map for controlling the strictness of boundary-guided saliency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 516, + 286, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 286, + 658 + ], + "type": "text", + "content": "Rather than use a decoder at each layer as done for the low-level saliency maps described above, the mid-level saliency maps of our model are generated using Grad-CAM [39] at three stages of the CNN backbone (see Figure 2). We also experiment with several other methods for generating student saliency maps and report on them in the Supplementary Material. The dilated generated boundary map " + }, + { + "bbox": [ + 46, + 516, + 286, + 658 + ], + "type": "inline_equation", + "content": "B_{d}(x)" + }, + { + "bbox": [ + 46, + 516, + 286, + 658 + ], + "type": "text", + "content": " is downsampled to match the feature map dimensions at these three stages in order to compare with the Grad-CAM generated saliency boundary maps. We use the binary cross entropy loss for supervision on dilated boundary regions. This loss is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 663, + 286, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 663, + 286, + 696 + ], + "spans": [ + { + "bbox": [ + 61, + 663, + 286, + 696 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} ^ {\\mathrm {d b s}} (x) = - \\frac {\\sum_ {j = 1} ^ {N} B _ {d} (x , j) \\log (1 - S (x , j))}{\\sum_ {j = 1} ^ {N} B _ {d} (x , j)}, \\tag {3}", + "image_path": "fb10aacf2bcc91c61f74b8712cc937236d5c23c2458f0ef5087d3b64cad78fd7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "S(x, j)" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " denotes the saliency map of our model on image " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "B_d(x, j)" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " is the dilated generated boundary map at pixel " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " is the number of pixels in " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ". We compute this loss only within dilated boundary regions (i.e. where " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "B_d(x, j) = 1" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "). This loss helps the student saliency map avoid intersecting with the dilated teacher boundary region." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 170, + 535, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 170, + 535, + 182 + ], + "spans": [ + { + "bbox": [ + 305, + 170, + 535, + 182 + ], + "type": "text", + "content": "3.3. Auxiliary Low-level Supervisions for EFCIL" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 190, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 545, + 297 + ], + "type": "text", + "content": "We propose to learn stable features from low-level stationary tasks shared across all incremental classification tasks during class-incremental learning. Low-level vision tasks like salient object detection require useful representations of input images. By learning these feature representations across tasks, the model can focus on key areas of input images and exploit learned, stable features with less representation drift since the low-level features change very little between tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 300, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 503 + ], + "type": "text", + "content": "Saliency map prediction is relevant to image classification since the foreground largely determines the results, while the background is comparatively less important. When learning new tasks with new classes, the background of images of new classes may contain new visual concepts that introduce undesirable noise and lead to forgetting of essential previous knowledge. The effectiveness of saliency features for learning classification tasks was demonstrated by Saliency Guided Training [17]. Additional supervision of salient region boundaries can aid salient object detection tasks for both segmentation and localization [12, 21, 24, 31, 50]. The positive interaction between these two tasks brings richer attention to features relevant to the main classification task. It can provide positive guidance in the form of stationary knowledge across class-incremental tasks. Some examples are illustrated in Figure 5." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 506, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 601 + ], + "type": "text", + "content": "We incorporate low-level vision tasks into the network as an auxiliary supervision for enriching task-agnostic attention. The boundary map is computed with a Laplacian filter over the estimated saliency map. We add a decoder " + }, + { + "bbox": [ + 304, + 506, + 545, + 601 + ], + "type": "inline_equation", + "content": "D_{\\psi}" + }, + { + "bbox": [ + 304, + 506, + 545, + 601 + ], + "type": "text", + "content": " [22] after the backbone " + }, + { + "bbox": [ + 304, + 506, + 545, + 601 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 304, + 506, + 545, + 601 + ], + "type": "text", + "content": " to predict low-level saliency and boundary maps for input images. The average L2 distance between the prediction and target is used as a low-level saliency distillation loss:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 347, + 613, + 545, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 613, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 347, + 613, + 545, + 640 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} ^ {\\mathrm {l m s}} (x) = \\frac {| | D _ {\\psi} (F _ {\\theta} (x)) - A (x) | | _ {2}}{\\sqrt {N}}, \\tag {4}", + "image_path": "0d7ce7788ebf8f3e5a43a75ebe586c3a85b2db7205462389e180568142b2ca45.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "A(x)" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " denotes the target low-level maps on input " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": ", consisting of a saliency map " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "A_{s}(x)" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " and a boundary map " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "A_{b}(x)" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "D_{\\psi}(F_{\\theta}(x))" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " are combined saliency and boundary maps produced by the decoder, and " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " is the number of pixels in the saliency maps." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 250, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 250, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 250, + 712 + ], + "type": "text", + "content": "1https://github.com/jacobgil/pytorch-grad-cam" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23957" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 47, + 87, + 287, + 278 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 178, + 83 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 178, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 178, + 83 + ], + "type": "text", + "content": "Algorithm 1 TASS Pseudocode" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "lines": [ + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": "Input: The number of tasks " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": ", training samples " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "D_{t} = \\{(x_{i},y_{i})\\}_{i=1}^{N}" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": " of task " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": ", initial parameters " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\Theta^0 = \\{\\theta_0,\\phi_0,\\psi_0\\}" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": " containing parameters of feature extractor " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": ", classifier " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "G_{\\phi}" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": ", and low-level decoder " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "D_{\\psi}" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": ". \nOutput: Model " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\Theta^T" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": " \n1: for " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "t \\in \\{1,2,\\dots,T\\}" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": " do \n2: " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\Theta^t \\gets \\Theta^{t-1}" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": " \n3: while not converged do \n4: Sample " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t^{\\mathrm{CIL}} \\gets \\text{SaliencyNoiseInjection}(x,y)" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t^{\\mathrm{Ims}} \\gets \\text{LowLevelMultitask}(x,A(x))" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "S \\gets \\text{ComputeGradCAMSaliency}(x,y)" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t^{\\mathrm{obs}} \\gets \\text{DilatedBoundarySupervision}(S,A(x))" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": " \n9: update " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": " by minimizing " + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t^{\\mathrm{all}}" + }, + { + "bbox": [ + 47, + 87, + 287, + 278 + ], + "type": "text", + "content": " from Eq. 5 \n10: end while \n11: end for" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 47, + 297, + 181, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 181, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 181, + 310 + ], + "type": "text", + "content": "3.4. Saliency Noise Injection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 387 + ], + "type": "text", + "content": "Although we apply low-level distillation and dilated boundary supervision to maintain saliency representations across tasks, the model can still forget saliency on samples from previous tasks. To address this, we force the model to be able to recover correct saliency maps from injected saliency noise." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 388, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 495 + ], + "type": "text", + "content": "At each task there is no available training data from previous or future tasks, and therefore we cannot directly know saliency drift on these samples. Instead of supervising the model with ground-truth saliency drift signals, we introduce saliency noise on random feature channels. We use a random ellipse to approximate the potential saliency drift in future tasks and the model is trained to denoise within each stage. Therefore the model can learn to effectively reduce real saliency drift." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 495, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 603 + ], + "type": "text", + "content": "We generate elliptical noise using a very simple approach. There are six parameter dimensions: the center coordinate " + }, + { + "bbox": [ + 46, + 495, + 287, + 603 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 46, + 495, + 287, + 603 + ], + "type": "text", + "content": ", the major and minor axis lengths " + }, + { + "bbox": [ + 46, + 495, + 287, + 603 + ], + "type": "inline_equation", + "content": "(a, b)" + }, + { + "bbox": [ + 46, + 495, + 287, + 603 + ], + "type": "text", + "content": ", the rotation angle " + }, + { + "bbox": [ + 46, + 495, + 287, + 603 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 495, + 287, + 603 + ], + "type": "text", + "content": ", and the mask weight " + }, + { + "bbox": [ + 46, + 495, + 287, + 603 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 495, + 287, + 603 + ], + "type": "text", + "content": ". A detailed explanation of this process is given in the Supplementary Material. With the help of dilated boundary supervision, each stage learns to eliminate this additional saliency noise and this aids generalization for future tasks and mitigates saliency forgetting in previous ones." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 609, + 275, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 275, + 622 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 275, + 622 + ], + "type": "text", + "content": "3.5. Learning Objective and Training Algorithm" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 628, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 663 + ], + "type": "text", + "content": "The overall learning objective combines the low-level multi-task learning, dilated boundary supervision, and random saliency noise injection modules:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 668, + 287, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 668, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 287, + 683 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} ^ {\\text {a l l}} = \\mathcal {L} _ {t} ^ {\\text {C I L}} + \\mathcal {L} _ {t} ^ {\\text {l m s}} + \\mathcal {L} _ {t} ^ {\\text {d b s}}. \\tag {5}", + "image_path": "a3e404b5367fe575ee1a11599f5479e9fec003138f92bddc6462c66ad7ec07f5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Comparing this loss with Eq. 1, for TASS " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t^{\\mathrm{M}} = \\mathcal{L}_t^{\\mathrm{CIL}} + \\mathcal{L}_t^{\\mathrm{Ins}}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": ", thus incorporating saliency-aware supervision with" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "the cross-entropy loss. The entire process is detailed in Algorithm 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 109, + 430, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 109, + 430, + 122 + ], + "spans": [ + { + "bbox": [ + 305, + 109, + 430, + 122 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 129, + 545, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 129, + 545, + 177 + ], + "spans": [ + { + "bbox": [ + 304, + 129, + 545, + 177 + ], + "type": "text", + "content": "In this section we first describe our experimental setup and then compare TASS to state-of-the-art methods on several EFCIL benchmarks. In Section 4.3 we give further analysis of the various components of TASS." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 186, + 422, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 186, + 422, + 200 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 422, + 200 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 205, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 228 + ], + "type": "text", + "content": "We follow standard experimental protocols for EFCIL on three benchmark datasets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "text", + "content": "Datasets. We perform experiments on CIFAR-100 [19], Tiny-ImageNet [20], and ImageNet-Subset [8]. For most experiments, we train the model on half of the classes for the first task, and then equally distribute the remaining classes across each of the subsequent tasks. The convention we use is: " + }, + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "inline_equation", + "content": "F + C \\times T" + }, + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "text", + "content": " means that the first task contains " + }, + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "text", + "content": " classes, and the next " + }, + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "text", + "content": " tasks each contain " + }, + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 230, + 545, + 338 + ], + "type": "text", + "content": " classes. This is a common configuration for EFCIL used in both PASS [52] and SSRE [53]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 338, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 458 + ], + "type": "text", + "content": "State-of-the-art methods. Since we focus on EFCIL, we mainly compare with exemplar-free state-of-the-art approaches: SSRE [53], PASS [52], IL2A [51], EWC [18], LwF-MC [37], and MUC [28]. To demonstrate the effectiveness of our method, we also compare its performance with several exemplar-based methods like iCaRL (both nearest-mean and CNN) [37], EEIL [2], and LUCIR [16]. We also compare with RRR [11] integrated with SSRE, which focuses on preserving saliency using exemplar replay." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 459, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 625 + ], + "type": "text", + "content": "Implementation details and metrics. We use ResNet-18 [15] as a feature extraction backbone. This is the same base network used in SSRE [53] and PASS [52], two state-of-the-art EFCIL approaches. We use the decoder in [22] to estimate low-level student saliency maps. All experiments are trained from scratch using Adam for 100 epochs with an initial learning rate 0.001. The learning rate is reduced by a factor 10 at epochs 45 and 90. For exemplar-based approaches, we use herding [37] to select and store 20 samples per class following common settings [16, 37]. We implement RRR [11] with SSRE to fairly compare it with TASS. For dilated boundary supervision, we set " + }, + { + "bbox": [ + 304, + 459, + 545, + 625 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 459, + 545, + 625 + ], + "type": "text", + "content": " of the three mid-level boundary dilation stages to be " + }, + { + "bbox": [ + 304, + 459, + 545, + 625 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 459, + 545, + 625 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 459, + 545, + 625 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 459, + 545, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 459, + 545, + 625 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 459, + 545, + 625 + ], + "type": "text", + "content": " of the image size." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "text", + "content": "We report three common metrics for class incremental learning: the average and last top-1 accuracy, as well as average forgetting for all classes learned up to task " + }, + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "text", + "content": ". Denoting by " + }, + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "inline_equation", + "content": "Acc_{i}" + }, + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "text", + "content": " the accuracy over all learned classes up to and including task " + }, + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "text", + "content": ", the average accuracy is defined as " + }, + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "inline_equation", + "content": "Avg = \\frac{\\sum_{i=1}^{T} Acc_{i}}{T}" + }, + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "text", + "content": ", and the last accuracy is " + }, + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "inline_equation", + "content": "Acc_{T}" + }, + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "text", + "content": ". Letting " + }, + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "inline_equation", + "content": "a_{m,n}" + }, + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "text", + "content": " denotes the accuracy of task " + }, + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "text", + "content": " after learning task " + }, + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 627, + 545, + 713 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23958" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 104, + 70, + 489, + 192 + ], + "blocks": [ + { + "bbox": [ + 104, + 70, + 489, + 192 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 489, + 192 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 489, + 192 + ], + "type": "table", + "html": "
DatasetCIFAR-100Tiny-ImageNet
Method5 tasks10 tasks20 tasks5 tasks10 tasks20 tasks
MUC38.4519.5715.6518.9515.479.14
+TASS49.17 (+10.72)40.34 (+20.77)37.86 (+22.21)32.47 (+13.46)30.13 (+14.66)27.70 (+18.56)
IL2A55.1345.3245.2436.7734.5328.68
+TASS58.74 (+3.61)53.24 (+7.92)53.07 (+7.83)42.49 (+5.72)41.34 (+6.81)40.59 (+11.91)
PASS55.6749.0348.4841.5839.2832.78
+TASS59.10 (+3.43)54.45 (+5.42)52.37 (+3.89)44.05 (+2.47)43.06 (+3.78)42.57 (+9.79)
SSRE56.3355.0150.4741.4540.0739.25
+TASS59.26 (+2.93)57.93 (+2.92)53.78 (+3.31)44.13 (+2.68)43.86 (+3.79)43.55 (+4.30)
", + "image_path": "d2236872246007ece57198f19ec99914f05881bf6337ff2b434f07552ee305fc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 233, + 544, + 398 + ], + "blocks": [ + { + "bbox": [ + 46, + 201, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 222 + ], + "type": "text", + "content": "Table 1. Performance gain in top-1 accuracy by applying TASS to other EFCIL methods in a plug-and-play way. Absolute gains are indicated in (red)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 233, + 544, + 398 + ], + "lines": [ + { + "bbox": [ + 50, + 233, + 544, + 398 + ], + "spans": [ + { + "bbox": [ + 50, + 233, + 544, + 398 + ], + "type": "table", + "html": "
DatasetCIFAR100TinyImageNet
Setting5 tasks10 tasks20 tasks5 tasks10 tasks20 tasks
MethodAvg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓
E=20iCaRL-CNN†51.0740.1242.1348.6639.6545.6944.4335.4743.5434.6422.3136.8931.1521.1036.7027.9020.4645.12
iCaRL-NCM†58.5649.7424.9054.1945.1328.3250.5140.6835.5345.8633.4527.1543.2933.7528.8938.0428.8937.40
LUCIR†63.7855.0621.0062.3950.1425.1259.0748.7828.6549.1537.0920.6148.5236.8022.2542.8332.5533.74
EEIL†60.3752.3523.3656.0547.6726.6552.3441.5932.4047.1234.2425.5645.0134.2625.9140.5030.1435.04
RRR†66.4357.2218.0565.7855.7418.5962.4351.3518.4051.2042.2316.6749.5440.1221.6447.4635.5429.10
E=20LwFMC45.9336.1744.2327.4350.4717.0420.0715.8855.4629.1217.1254.2623.1012.3354.3717.438.7563.54
EWC16.049.3260.1714.708.4762.5314.128.2363.8918.8012.7167.5515.7710.1270.2312.398.4275.54
MUC49.4238.4540.2830.1919.5747.5621.2715.6552.6532.5817.9851.4626.6114.5450.2121.9512.7058.00
IL2A63.2255.1323.7857.6545.3230.4154.9045.2430.8448.1736.1425.4342.1035.2328.3236.7928.7435.46
PASS63.4755.6725.2061.8449.0330.2558.0948.4830.6149.5541.5818.0447.2939.2823.1142.0732.7830.55
SSRE65.8856.3318.3765.0455.0119.4861.7050.4718.3750.3941.6717.2548.9339.8922.5048.1739.7626.74
TASS (Ours)68.7559.2616.4267.4257.9317.7862.7653.7817.7855.1244.1315.4054.2143.8618.4752.7943.5522.51
", + "image_path": "da3b221466ebca517a18a97dc9959b6368b8ae3525006ec34fd60d28900aa8dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 446, + 296, + 530 + ], + "blocks": [ + { + "bbox": [ + 46, + 406, + 547, + 440 + ], + "lines": [ + { + "bbox": [ + 46, + 406, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 547, + 440 + ], + "type": "text", + "content": "Table 2. Average, last top-1 accuracy, and forgetting on CIFAR-100 with different numbers of tasks. Replay-based methods storing 20 exemplars from each previous class are indicate with †. The best overall results are in bold. We run all experiments three times and report the mean for all metrics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 446, + 296, + 530 + ], + "lines": [ + { + "bbox": [ + 50, + 446, + 296, + 530 + ], + "spans": [ + { + "bbox": [ + 50, + 446, + 296, + 530 + ], + "type": "table", + "html": "
DatasetImageNet-Subset
Setting5 tasks10 tasks20 tasks
MethodAvg↑Last↑F↓Avg↑Last↑F↓Avg↑Last↑F↓
LwF_TC34.8624.1049.3631.1820.0153.0427.5417.4256.07
MUC40.6527.8947.1335.0722.6552.1031.4420.1253.85
PASS63.1252.6122.4761.8050.4423.5755.2346.0726.73
SSRE69.5458.4617.2267.6957.5118.6061.2350.0523.22
TASS (Ours)74.3263.1414.3772.6057.9316.0968.7957.6018.41
", + "image_path": "24dd2ee15a58623054fbaecef4e9a1454c6a855330e260c0244d3a36b3dafc57.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 538, + 288, + 571 + ], + "lines": [ + { + "bbox": [ + 46, + 538, + 288, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 288, + 571 + ], + "type": "text", + "content": "Table 3. Average, last top-1 accuracy, and forgetting on ImageNet-Subset for different numbers of tasks. We run all experiments three times and report means for all metrics." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 595, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 595, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 595, + 287, + 635 + ], + "type": "text", + "content": "the forgetting measure " + }, + { + "bbox": [ + 46, + 595, + 287, + 635 + ], + "type": "inline_equation", + "content": "f_{k}^{i}" + }, + { + "bbox": [ + 46, + 595, + 287, + 635 + ], + "type": "text", + "content": " of task " + }, + { + "bbox": [ + 46, + 595, + 287, + 635 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 595, + 287, + 635 + ], + "type": "text", + "content": " after learning task " + }, + { + "bbox": [ + 46, + 595, + 287, + 635 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 595, + 287, + 635 + ], + "type": "text", + "content": " is computed as " + }, + { + "bbox": [ + 46, + 595, + 287, + 635 + ], + "type": "inline_equation", + "content": "f_{k}^{i} = \\max_{t\\in 1,2,\\ldots ,k - 1}(a_{t,i} - a_{k,i})" + }, + { + "bbox": [ + 46, + 595, + 287, + 635 + ], + "type": "text", + "content": ". The average forgetting " + }, + { + "bbox": [ + 46, + 595, + 287, + 635 + ], + "type": "inline_equation", + "content": "F_{k}" + }, + { + "bbox": [ + 46, + 595, + 287, + 635 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 46, + 595, + 287, + 635 + ], + "type": "inline_equation", + "content": "F_{k} = \\frac{1}{k - 1}\\sum_{i = 1}^{k - 1}f_{k}^{i}" + }, + { + "bbox": [ + 46, + 595, + 287, + 635 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 645, + 244, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 244, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 244, + 658 + ], + "type": "text", + "content": "4.2. Comparison with the State-of-the-art" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "We compare TASS with the state-of-the-art on CIFAR-100 in Table 2. TASS outperforms all exemplar-free approaches. For exemplar-based methods like iCaRL [37], EEIL [2], and LUCIR [16], our method still has significantly better" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 449, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 521 + ], + "type": "text", + "content": "performance. On longer sequences (i.e. 10 and 20 tasks), our method significantly reduces forgetting when learning new classes compared to other EFCIL methods. TASS outperforms the best method SSRE by about " + }, + { + "bbox": [ + 304, + 449, + 545, + 521 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 304, + 449, + 545, + 521 + ], + "type": "text", + "content": " on the last task. This performance improvement can be also observed in terms of average forgetting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "type": "text", + "content": "As we see in Table 3 and Figure 4 for Tiny-ImageNet and ImageNet-Subset, although our method has similar top-1 accuracy on the first task in Figure 4, it has better performance at most intermediate tasks and also the final one. For longer sequences in Figure 4, the gap between our method and the best baseline is largely consistent, showing the effectiveness of our method at mitigating forgetting. The performance gain in Table 3 is larger on Tiny-ImageNet and ImageNet-Subset compared to CIFAR100, and this demonstrates that our method generalizes to datasets with larger images and object scales. It is worth mentioning that TASS also produces results with smaller variance. We believe this to be due to TASS reducing saliency drift to background regions, which may include random noise." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "Plug-and-play with other EFCIL methods. Some existing EFCIL methods, like PASS [52], IL2A [51] and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23959" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 76, + 216, + 183 + ], + "blocks": [ + { + "bbox": [ + 71, + 76, + 216, + 183 + ], + "lines": [ + { + "bbox": [ + 71, + 76, + 216, + 183 + ], + "spans": [ + { + "bbox": [ + 71, + 76, + 216, + 183 + ], + "type": "image", + "image_path": "2786bca8e8231d2a3146062959cbdc48efc3ddaf9cc1549f1e2f25be0b9c2f1b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 76, + 369, + 182 + ], + "blocks": [ + { + "bbox": [ + 220, + 76, + 369, + 182 + ], + "lines": [ + { + "bbox": [ + 220, + 76, + 369, + 182 + ], + "spans": [ + { + "bbox": [ + 220, + 76, + 369, + 182 + ], + "type": "image", + "image_path": "3c79aabdedbb864260f53e8611d4a22a80a3622f33ba36d1e3dc940a1bb95f0a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 373, + 76, + 518, + 182 + ], + "blocks": [ + { + "bbox": [ + 373, + 76, + 518, + 182 + ], + "lines": [ + { + "bbox": [ + 373, + 76, + 518, + 182 + ], + "spans": [ + { + "bbox": [ + 373, + 76, + 518, + 182 + ], + "type": "image", + "image_path": "caf86358f27db210f3dd0db47cff36829e498d5eb40ded3758f7de87657ec83d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 71, + 185, + 216, + 291 + ], + "blocks": [ + { + "bbox": [ + 71, + 185, + 216, + 291 + ], + "lines": [ + { + "bbox": [ + 71, + 185, + 216, + 291 + ], + "spans": [ + { + "bbox": [ + 71, + 185, + 216, + 291 + ], + "type": "image", + "image_path": "e101741421122b30b8d6925fd5d3306b006c8a595aec14bbcf7e44176725f797.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 293, + 547, + 314 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 547, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 547, + 314 + ], + "type": "text", + "content": "Figure 4. Results on Tiny-ImageNet and ImageNet-Subset for different numbers of tasks. Our method outperforms others, especially on longer task sequences (i.e. more, but smaller, tasks)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 223, + 185, + 367, + 290 + ], + "blocks": [ + { + "bbox": [ + 223, + 185, + 367, + 290 + ], + "lines": [ + { + "bbox": [ + 223, + 185, + 367, + 290 + ], + "spans": [ + { + "bbox": [ + 223, + 185, + 367, + 290 + ], + "type": "image", + "image_path": "873b053d588788e54782dde677d13d6f5aa5f43bcda3b49cd9e13fc67e69ab18.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 374, + 186, + 518, + 291 + ], + "blocks": [ + { + "bbox": [ + 374, + 186, + 518, + 291 + ], + "lines": [ + { + "bbox": [ + 374, + 186, + 518, + 291 + ], + "spans": [ + { + "bbox": [ + 374, + 186, + 518, + 291 + ], + "type": "image", + "image_path": "420c6ef776df5e91f24ac5dd6ba74001c0d9d6279dde2fe2db7aa838ef0b2432.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 335, + 289, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 289, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 289, + 491 + ], + "type": "text", + "content": "SSRE [53], focus on reducing forgetting via embedding regularization. Considering the importance of saliency to image classification, it is natural to consider whether TASS can be integrated into these methods. The results in Table 1 show the performance gain brought by this integration. Adding TASS doubles the performance for MUC in many cases and significantly improves IL2A and PASS. When we incorporate it into the best baseline SSRE, it yields a consistent gain of about " + }, + { + "bbox": [ + 46, + 335, + 289, + 491 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 46, + 335, + 289, + 491 + ], + "type": "text", + "content": ". These results clearly show that TASS, by explicitly mitigating saliency drift, is complementary to other methods in relieving forgetting. They additionally demonstrate the significance of saliency drift as a cause of catastrophic forgetting in EFCIL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 501, + 163, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 163, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 163, + 514 + ], + "type": "text", + "content": "4.3. Additional Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 520, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 555 + ], + "type": "text", + "content": "In this section we take a deeper look at the method we propose. If not specified, the results are produced using TASS integrated into SSRE [53]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 556, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 677 + ], + "type": "text", + "content": "Ablation Study. We performed ablations using the 10-task setting on CIFAR-100 (see Table 4). We ablate on both PASS [52] and SSRE [53]. Low-level multi-task supervision is crucial and improves by " + }, + { + "bbox": [ + 46, + 556, + 287, + 677 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 46, + 556, + 287, + 677 + ], + "type": "text", + "content": " (PASS) and " + }, + { + "bbox": [ + 46, + 556, + 287, + 677 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 46, + 556, + 287, + 677 + ], + "type": "text", + "content": " (SSRE). Dilated boundary supervision further boosts performance by about " + }, + { + "bbox": [ + 46, + 556, + 287, + 677 + ], + "type": "inline_equation", + "content": "1 - 2\\%" + }, + { + "bbox": [ + 46, + 556, + 287, + 677 + ], + "type": "text", + "content": ". Saliency noise injection is also helpful for both methods and improves by " + }, + { + "bbox": [ + 46, + 556, + 287, + 677 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 46, + 556, + 287, + 677 + ], + "type": "text", + "content": " for PASS. In total, TASS improves baselines by " + }, + { + "bbox": [ + 46, + 556, + 287, + 677 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 46, + 556, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 556, + 287, + 677 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 46, + 556, + 287, + 677 + ], + "type": "text", + "content": " points, respectively. Note that SSRE is the previous state-of-the-art method and TASS outperforms it by a large margin." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Low-level Multi-task Supervision. To analyze the effect of our proposed low-level saliency supervision, we performed experiments on ImageNet-Subset in the 5 and 10" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 337, + 333, + 517, + 466 + ], + "blocks": [ + { + "bbox": [ + 337, + 333, + 517, + 466 + ], + "lines": [ + { + "bbox": [ + 337, + 333, + 517, + 466 + ], + "spans": [ + { + "bbox": [ + 337, + 333, + 517, + 466 + ], + "type": "table", + "html": "
Method & TasksL1msL1bsSNIAccuracy
Baseline (PASS)49.0
Variants✓51.2
✓✓53.0
✓✓✓54.5
Baseline (SSRE)55.0
Variants✓56.2
✓✓57.3
✓✓✓57.9
", + "image_path": "5934431c6f6d9ae66447ff3504b4ae8e29656772529af4e56389f12b3bbe6419.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 473, + 545, + 540 + ], + "lines": [ + { + "bbox": [ + 305, + 473, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 473, + 545, + 540 + ], + "type": "text", + "content": "Table 4. Ablations on each TASS component. Experiments are on CIFAR-100 in the 10-task setting and we report the top-1 accuracy in % for TASS integrated into PASS and SSRE. " + }, + { + "bbox": [ + 305, + 473, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ubs}}" + }, + { + "bbox": [ + 305, + 473, + 545, + 540 + ], + "type": "text", + "content": " (Eq. 3), " + }, + { + "bbox": [ + 305, + 473, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{lms}}" + }, + { + "bbox": [ + 305, + 473, + 545, + 540 + ], + "type": "text", + "content": " (Eq. 4), and SNI denote the three TASS components: Dilated Boundary Supervision, Low-level Multi-task Supervision, and Saliency Noise Injection." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "task settings. We first plot the loss across tasks in Figure 5(c). After learning to predict boundary and saliency maps in the first task, the network maintains good performance for the rest of the 5 task sequence. This shows that the low-level tasks are stable during continual learning. Furthermore, we visualize the results of saliency and boundary map prediction during incremental learning in Figure 5(a-b). We give some examples of predicted boundary and saliency maps after learning different tasks. Although CIL involves samples of different classes, we see that the low-level outputs are relatively stable and class agnostic. Since the model is able to stably predict these low-level features across tasks, it therefore can preserve useful prior knowl" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23960" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 218, + 207 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 218, + 207 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 218, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 218, + 207 + ], + "type": "image", + "image_path": "1476576f2e990c0f32b9db0382198a45131e1d2325330f4a31b6fe8934e42324.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 211, + 138, + 221 + ], + "lines": [ + { + "bbox": [ + 126, + 211, + 138, + 221 + ], + "spans": [ + { + "bbox": [ + 126, + 211, + 138, + 221 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 72, + 387, + 207 + ], + "blocks": [ + { + "bbox": [ + 225, + 72, + 387, + 207 + ], + "lines": [ + { + "bbox": [ + 225, + 72, + 387, + 207 + ], + "spans": [ + { + "bbox": [ + 225, + 72, + 387, + 207 + ], + "type": "image", + "image_path": "d28aeb54d478b5f5b10363ffd05810ec4f00605d06636e532b5bd843197fbe54.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 294, + 211, + 307, + 221 + ], + "lines": [ + { + "bbox": [ + 294, + 211, + 307, + 221 + ], + "spans": [ + { + "bbox": [ + 294, + 211, + 307, + 221 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 392, + 72, + 539, + 210 + ], + "blocks": [ + { + "bbox": [ + 392, + 72, + 539, + 210 + ], + "lines": [ + { + "bbox": [ + 392, + 72, + 539, + 210 + ], + "spans": [ + { + "bbox": [ + 392, + 72, + 539, + 210 + ], + "type": "image", + "image_path": "7a8e433045b1b2e3c350917ccc79574eaa71dcb785bece9d4d91380e06f7038a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 212, + 477, + 222 + ], + "lines": [ + { + "bbox": [ + 465, + 212, + 477, + 222 + ], + "spans": [ + { + "bbox": [ + 465, + 212, + 477, + 222 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 236, + 547, + 269 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 547, + 269 + ], + "type": "text", + "content": "Figure 5. Visualization of the saliency (a) and boundary (b) maps from our student encoder-decoder network with original images from different tasks at different stages of incremental learning. Our method produces stable low-level results while reducing forgetting in classification. (c) The MAE loss between the student and teacher network across tasks." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 282, + 160, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 160, + 294 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 160, + 294 + ], + "type": "text", + "content": "edge for continual learning." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 102, + 314, + 233, + 392 + ], + "blocks": [ + { + "bbox": [ + 102, + 314, + 233, + 392 + ], + "lines": [ + { + "bbox": [ + 102, + 314, + 233, + 392 + ], + "spans": [ + { + "bbox": [ + 102, + 314, + 233, + 392 + ], + "type": "table", + "html": "
Metric & MethodAvg↑Last↑
FeTrIL [35]65.2056.34
SOPE [54]65.8456.80
PRAKA [40]68.8659.20
TASS (SSRE)67.4257.93
TASS (PRAKA)69.7060.04
", + "image_path": "d80d5d29036d1a6bed5d6ec7b3252daf6d1d55bbfdf98e5431c69b34af60b49c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 115, + 441, + 220, + 494 + ], + "blocks": [ + { + "bbox": [ + 47, + 399, + 286, + 411 + ], + "lines": [ + { + "bbox": [ + 47, + 399, + 286, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 286, + 411 + ], + "type": "text", + "content": "Table 5. Average and last accuracy on CIFAR-100 10-task setting." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 115, + 441, + 220, + 494 + ], + "lines": [ + { + "bbox": [ + 115, + 441, + 220, + 494 + ], + "spans": [ + { + "bbox": [ + 115, + 441, + 220, + 494 + ], + "type": "table", + "html": "
Metric & MethodAvg↑
SOPE [54]60.20
FeTrIL [35]65.00
TASS (FeTrIL)66.03
", + "image_path": "88971ce58b618a4287ff3abb6cd393ddd2a73e4b64abf8e17e7cd64837d0ff9f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 528, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 624 + ], + "type": "text", + "content": "TASS with more methods and benchmarks. Due to the strong generalization capability of our method, we have also applied our TASS paradigm to PRAKA [40], resulting in further performance improvements. We give specific experimental comparisons in Table 5. We also compare with FeTrIL [35] in Table 5 and Table 6 above. Experiments on ImageNet-Full show consistent improvement in Table 6 above." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": "Quantitative analysis of saliency drift. We measure the intersection over union (IoU) between the self-attention map of the last layer in DINO and the Grad-CAM saliency maps for both SSRE and " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{SSRE + TASS}" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": ". As shown in Table 7, TASS significantly reduces saliency drift in the student model, which again shows the effectiveness of our method for maintaining saliency during CIL." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 339, + 279, + 513, + 316 + ], + "blocks": [ + { + "bbox": [ + 49, + 502, + 284, + 513 + ], + "lines": [ + { + "bbox": [ + 49, + 502, + 284, + 513 + ], + "spans": [ + { + "bbox": [ + 49, + 502, + 284, + 513 + ], + "type": "text", + "content": "Table 6. Average accuracy on the ImageNet-Full 10-task setting." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 339, + 279, + 513, + 316 + ], + "lines": [ + { + "bbox": [ + 339, + 279, + 513, + 316 + ], + "spans": [ + { + "bbox": [ + 339, + 279, + 513, + 316 + ], + "type": "table", + "html": "
Task14710
SSRE47.450.156.678.5
SSRE+TASS75.282.388.590.1
", + "image_path": "f288d85502710137669e6c3b1c41e84abc590e6a27833de14c27a4d0c2858362.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 324, + 545, + 346 + ], + "lines": [ + { + "bbox": [ + 306, + 324, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 324, + 545, + 346 + ], + "type": "text", + "content": "Table 7. Average IoU (\\%) between DINO self-attention maps and student model saliency in the CIFAR-100 10-task." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 368, + 383, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 368, + 383, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 368, + 383, + 380 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 393, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 545, + 513 + ], + "type": "text", + "content": "In this paper, we propose an approach of task-adaptive saliency guidance for EFCIL. The insight behind TASS is to guide the model to focus on salient regions and inhibit saliency drift. We show that robust saliency guidance is crucial to mitigating forgetting across tasks. Experiments demonstrate that TASS is effective and surpasses the state-of-the-art. TASS can be easily combined with other methods, leading to large performance gains over baselines. Qualitative results also show that low-level tasks are stable across different tasks, resulting in less forgetting." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 517, + 545, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 636 + ], + "type": "text", + "content": "Limitations and Future Work. In this work, we introduce the auxiliary saliency knowledge across tasks for better CIL performance, which may raise the concern of unfair comparison. However, we believe it is important to explore external knowledge to make the CIL system applicable in real applications. Additionally, our method can be easily integrated with other methods and the low-level auxiliary knowledge takes negligible cost and is easy to obtain. We will explore leveraging other forms of knowledge in future work, such as pre-trained visual and large language models." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgments This work is funded by NSFC (NO. 62206135, 62225604), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), and the Fundamental Research Funds for the Central Universities (Nankai University, 070-63233085). Computation is supported by the Supercomputing Center of Nankai University." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23961" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Eden Belouadah, Adrian Popescu, and Ioannis Kanellos. A comprehensive study of class incremental learning algorithms for visual tasks. Neural Networks, 135:38-54, 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "type": "text", + "content": "[2] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In ECCV, 2018. 1, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 225 + ], + "type": "text", + "content": "[3] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9233-9242, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 227, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 287, + 258 + ], + "type": "text", + "content": "[4] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. In ICLR, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 287, + 304 + ], + "type": "text", + "content": "[5] Hanting Chen, Yunhe Wang, Chang Xu, Zhaohui Yang, Chuanjian Liu, Boxin Shi, Chunjing Xu, Chao Xu, and Qi Tian. Data-free learning of student networks. In ICCV, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 306, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 287, + 349 + ], + "type": "text", + "content": "[6] Ming-Ming Cheng, Shanghua Gao, Ali Borji, Yong-Qiang Tan, Zheng Lin, and Meng Wang. A highly efficient model to study the semantics of salient object detection. IEEE TPAMI, 2021. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 351, + 287, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 395 + ], + "type": "text", + "content": "[7] Matthias Delange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Greg Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. IEEE TPAMI, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 396, + 287, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 287, + 429 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 430, + 287, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 287, + 463 + ], + "type": "text", + "content": "[9] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In CVPR, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 465, + 287, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 287, + 507 + ], + "type": "text", + "content": "[10] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "text", + "content": "[11] Sayna Ebrahimi, Suzanne Petryk, Akash Gokul, William Gan, Joseph E. Gonzalez, Marcus Rohrbach, and trevor darrell. Remembering for the right reasons: Explanations reduce catastrophic forgetting. In ICLR, 2021. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 287, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 598 + ], + "type": "text", + "content": "[12] Deng-Ping Fan, Ge-Peng Ji, Peng Xu, Ming-Ming Cheng, Christos Sakaridis, and Luc Van Gool. Advances in deep concealed scene understanding. Visual Intelligence, 1(1):16, 2023. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "text", + "content": "[13] Tao Feng, Mang Wang, and Hangjie Yuan. Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9427-9436, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 657, + 287, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 700 + ], + "type": "text", + "content": "[14] Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv preprint arXiv:1312.6211, 2013. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 702, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 287, + 713 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Deep residual learning for image recognition. In CVPR, 2016.5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 97, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 129 + ], + "type": "text", + "content": "[16] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, 2019. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 132, + 545, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 132, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 307, + 132, + 545, + 164 + ], + "type": "text", + "content": "[17] Aya Abdelsalam Ismail, Hector Corrada Bravo, and Soheil Feizi. Improving deep learning interpretability by saliency guided training. NeurIPS, 2021. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 167, + 545, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 167, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 307, + 167, + 545, + 231 + ], + "type": "text", + "content": "[18] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 2017. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 234, + 545, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 234, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 545, + 264 + ], + "type": "text", + "content": "[19] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, 2009. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 268, + 545, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 268, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 307, + 268, + 545, + 290 + ], + "type": "text", + "content": "[20] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 292, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 292, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 292, + 545, + 357 + ], + "type": "text", + "content": "[21] Hao Li, Dingwen Zhang, Nian Liu, Lechao Cheng, Yalun Dai, Chao Zhang, Xinggang Wang, and Junwei Han. Boosting low-data instance segmentation by unsupervised pretraining with saliency prompt. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15485-15494, 2023. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 359, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 359, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 359, + 545, + 392 + ], + "type": "text", + "content": "[22] Xiangtai Li, Ansheng You, Zhen Zhu, Houlong Zhao, Maoke Yang, Kuiyuan Yang, and Yunhai Tong. Semantic flow for fast and accurate scene parsing. In ECCV, 2020. 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 394, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 415 + ], + "type": "text", + "content": "[23] Zhizhong Li and Derek Hoiem. Learning without forgetting. In ECCV, 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 418, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 418, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 545, + 460 + ], + "type": "text", + "content": "[24] Liqiang Lin, Pengdi Huang, Chi-Wing Fu, Kai Xu, Hao Zhang, and Hui Huang. On learning the right attention point for feature enhancement. Science China Information Sciences, 66(1):112107, 2023. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 463, + 545, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 463, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 463, + 545, + 496 + ], + "type": "text", + "content": "[25] Zheng Lin, Zhao Zhang, Zi-Yue Zhu, Deng-Ping Fan, and Xia-Lei Liu. Sequential interactive image segmentation. Computational Visual Media, 9(4):753-765, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 498, + 545, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 541 + ], + "type": "text", + "content": "[26] Xialei Liu, Marc Masana, Luis Herranz, Joost Van de Weijer, Antonio M Lopez, and Andrew D Bagdanov. Rotate your networks: Better weight consolidation and less catastrophic forgetting. In ICPR, 2018. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 544, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 587 + ], + "type": "text", + "content": "[27] Xialei Liu, Yu-Song Hu, Xu-Sheng Cao, Andrew D Bagdanov, Ke Li, and Ming-Ming Cheng. Long-tailed class incremental learning. In European Conference on Computer Vision, pages 495-512. Springer, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "type": "text", + "content": "[28] Yu Liu, Sarah Parisot, Gregory Slabaugh, Xu Jia, Ales Leonardis, and Tinne Tuytelaars. More classifiers, less forgetting: A generic multi-classifier paradigm for incremental learning. In ECCV. Springer, 2020. 2, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[29] Yuyang Liu, Yang Cong, Dipam Goswami, Xialei Liu, and Joost van de Weijer. Augmented box replay: Overcoming foreground shift for incremental object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11367-11377, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[30] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. NeurIPS, 2017. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23962" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[31] Zhouzhou Ma, Guanghua Gu, and Wenrui Zhao. Self-attention guidance based crowd localization and counting. Machine Intelligence Research, pages 1–17, 2024. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 139 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 139 + ], + "type": "text", + "content": "[32] Arun Mallya and Svetlana Lazebnik. Packet: Adding multiple tasks to a single network by iterative pruning. In CVPR, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 143, + 286, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 143, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 143, + 286, + 185 + ], + "type": "text", + "content": "[33] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost van de Weijer. Class incremental learning: survey and performance evaluation on image classification. IEEE TPAMI, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 189, + 286, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 189, + 286, + 231 + ], + "spans": [ + { + "bbox": [ + 49, + 189, + 286, + 231 + ], + "type": "text", + "content": "[34] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of learning and motivation*, pages 109–165. Elsevier, 1989. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 234, + 286, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 234, + 286, + 266 + ], + "spans": [ + { + "bbox": [ + 49, + 234, + 286, + 266 + ], + "type": "text", + "content": "[35] Grégoire Petit, Adrian Popescu, Hugo Schindler, David Picard, and Bertrand Delezoise. Fetril: Feature translation for exemplar-free class-incremental learning. In WACV, 2023. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 269, + 286, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 269, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 49, + 269, + 286, + 290 + ], + "type": "text", + "content": "[36] Quang Pham, Chenghao Liu, and Steven Hoi. Dualnet: Continual learning, fast and slow. NeurIPS, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 292, + 286, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 292, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 49, + 292, + 286, + 334 + ], + "type": "text", + "content": "[37] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017. 1, 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 338, + 286, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 338, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 49, + 338, + 286, + 380 + ], + "type": "text", + "content": "[38] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. In ICLR, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 384, + 286, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 384, + 286, + 426 + ], + "spans": [ + { + "bbox": [ + 49, + 384, + 286, + 426 + ], + "type": "text", + "content": "[39] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 429, + 286, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 429, + 286, + 461 + ], + "spans": [ + { + "bbox": [ + 49, + 429, + 286, + 461 + ], + "type": "text", + "content": "[40] Wuxuan Shi and Mang Ye. Prototype reminiscence and augmented asymmetric knowledge aggregation for nonexemplar class-incremental learning. In ICCV, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 464, + 286, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 464, + 286, + 506 + ], + "spans": [ + { + "bbox": [ + 49, + 464, + 286, + 506 + ], + "type": "text", + "content": "[41] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In ICCV, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 510, + 286, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 510, + 286, + 541 + ], + "spans": [ + { + "bbox": [ + 49, + 510, + 286, + 541 + ], + "type": "text", + "content": "[42] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 544, + 286, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 544, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 49, + 544, + 286, + 597 + ], + "type": "text", + "content": "[43] Jia-Wen Xiao, Chang-Bin Zhang, Jiekang Feng, Xialei Liu, Joost van de Weijer, and Ming-Ming Cheng. Endpoints weight fusion for class incremental semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7204-7213, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 601, + 286, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 601, + 286, + 621 + ], + "spans": [ + { + "bbox": [ + 49, + 601, + 286, + 621 + ], + "type": "text", + "content": "[44] Ju Xu and Zhanxing Zhu. Reinforced continual learning. NeurIPS, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 624, + 286, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 286, + 666 + ], + "type": "text", + "content": "[45] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "type": "text", + "content": "[46] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 387 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[47] Lu Yu, Xialei Liu, and Joost Van de Weijer. Self-training for class-incremental semantic segmentation. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[48] Jiang-Tian Zhai, Xialei Liu, Andrew D Bagdanov, Ke Li, and Ming-Ming Cheng. Masked autoencoders are efficient class incremental learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19104-19113, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 163, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 205 + ], + "type": "text", + "content": "[49] Junting Zhang, Jie Zhang, Shalini Ghosh, Dawei Li, Serafettin Tasci, Larry Heck, Heming Zhang, and C-C Jay Kuo. Class-incremental learning via deep model consolidation. In WACV, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 208, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 239 + ], + "type": "text", + "content": "[50] Jia-Xing Zhao, Jiang-Jiang Liu, Deng-Ping Fan, Yang Cao, Jufeng Yang, and Ming-Ming Cheng. Egnet: Edge guidance network for salient object detection. In ICCV, 2019. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 242, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 242, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 242, + 545, + 274 + ], + "type": "text", + "content": "[51] Fei Zhu, Zhen Cheng, Xu-yao Zhang, and Cheng-lin Liu. Class-incremental learning via dual augmentation. NeurIPS, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "text", + "content": "[52] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In CVPR, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 310, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 352 + ], + "type": "text", + "content": "[53] Kai Zhu, Wei Zhai, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Self-sustaining representation expansion for nonexemplar class-incremental learning. In CVPR, 2022. 2, 3, 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 387 + ], + "type": "text", + "content": "[54] Kai Zhu, Kecheng Zheng, Ruili Feng, Deli Zhao, Yang Cao, and Zheng-Jun Zha. Self-organizing pathway expansion for non-exemplar class-incremental learning. In ICCV, 2023. 8" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23963" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Task-Aware Encoder Control for Deep Video Compression/b978be40-5174-4abc-9649-d7daa2d37c89_content_list.json b/2024/Task-Aware Encoder Control for Deep Video Compression/b978be40-5174-4abc-9649-d7daa2d37c89_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..09f4085a0c9b48f3f71caba04537d7cb19620e2a --- /dev/null +++ b/2024/Task-Aware Encoder Control for Deep Video Compression/b978be40-5174-4abc-9649-d7daa2d37c89_content_list.json @@ -0,0 +1,1495 @@ +[ + { + "type": "text", + "text": "Task-Aware Encoder Control for Deep Video Compression", + "text_level": 1, + "bbox": [ + 189, + 130, + 782, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xingtong Ge $^{1,2*}$ Jixiang Luo $^{2}$ Xinjie Zhang $^{3}$ Tongda Xu $^{4}$ Guo Lu $^{5}$ \nDailan He $^{6}$ Jing Geng $^{1\\ddagger}$ Yan Wang $^{4}$ Jun Zhang $^{3}$ Hongwei Qin $^{2}$ $^{1}$ Beijing Institute of Technology $^{2}$ SenseTime Research \n $^{3}$ The Hong Kong University of Science and Technology \n $^{4}$ Institute for AI Industry Research (AIR), Tsinghua University \n $^{5}$ Shanghai Jiaotong University $^{6}$ The Chinese University of Hong Kong", + "bbox": [ + 147, + 178, + 821, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prior research on deep video compression (DVC) for machine tasks typically necessitates training a unique codec for each specific task, mandating a dedicated decoder per task. In contrast, traditional video codecs employ a flexible encoder controller, enabling the adaptation of a single codec to different tasks through mechanisms like mode prediction. Drawing inspiration from this, we introduce an innovative encoder controller for deep video compression for machines. This controller features a mode prediction and a Group of Pictures (GoP) selection module. Our approach centralizes control at the encoding stage, allowing for adaptable encoder adjustments across different tasks, such as detection and tracking, while maintaining compatibility with a standard pre-trained DVC decoder. Empirical evidence demonstrates that our method is applicable across multiple tasks with various existing pre-trained DVCs. Moreover, extensive experiments demonstrate that our method outperforms previous DVC by about $25\\%$ bitrate for different tasks, with only one pre-trained decoder.", + "bbox": [ + 73, + 353, + 473, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 686, + 209, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past decades, video analysis techniques have proliferated across a variety of fields, including smart cities, autonomous vehicles, and traffic surveillance. For these applications, videos are often compressed before being transmitted to cloud-based systems for further machine vision analyses, such as object detection and tracking. However, current video compression techniques, which range from traditional codecs [2, 33, 42] to recent learning-based codecs [14, 18, 24, 25, 29], primarily cater to the human vi", + "bbox": [ + 75, + 713, + 468, + 849 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a105509d85e278cad2fb12c7a0de63003730f785de90480a600d7fe39344373c.jpg", + "image_caption": [ + "Figure 1. (a) Mainstream video codec that serves the human viewing. (b) Our controlled video codecs for machine vision with fixed decoder. (c) Other video codec for machine vision with one-to-one encoders and decoders." + ], + "image_footnote": [], + "bbox": [ + 503, + 320, + 890, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sual system, as shown in Fig. 1(a). This specificity leads to inefficiencies, as machine vision tasks typically focus on selective semantic details and regions within the video frames, rather than the whole frames.", + "bbox": [ + 496, + 521, + 892, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle this problem, recent studies [1, 4, 11-13, 20, 38, 40, 44, 47] have explored scalable compression for multiple tasks through different layers in image and video coding. They propose a base layer dedicated to machine vision, with an enhancement layer containing additional information for human vision. Additionally, the latest work [36] has tried to compress semantic features at the encoder side, which serve as side information to supplement decoded images with more semantic details for machine vision at lower bitrates. Despite these advances, these methods require individually customized CODECs when applied to different downstream tasks, as the \"One-to-one CODECs\" shown in Fig. 1(c). In contrast, traditional CODECs can control the encoding process for different objectives, like PSNR and SSIM optimization, using a single decoder. Inspired by this feature, in this paper, we focus on how to use one pre-trained DVC-decoder to support both human and multiple machine vision tasks.", + "bbox": [ + 496, + 582, + 893, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In both residual [14, 15, 24] and conditional [18, 19, 29] DVC series, the encoded bitstreams predominantly comprise two parts: motion and residual/contextual informa", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work was done when Xingtong Ge interned at SenseTime Research", + "bbox": [ + 93, + 861, + 460, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ xingtong.ge@gmail.com", + "bbox": [ + 96, + 875, + 233, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Corresponding author, janegeng@bit.edu.cn", + "bbox": [ + 96, + 887, + 334, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26036", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion. It is observed that a substantial portion of the bitrate, often exceeding $80\\%$ in high bitrate models, is allocated to the residual/contextual information. This allocation is primarily for high-quality frame reconstruction. However, this approach is inappropriate for downstream vision tasks such as tracking and detection, which primarily concentrate on the objects and their movements rather than the full frames.", + "bbox": [ + 75, + 90, + 467, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate this issue, we introduce a dynamic vision mode prediction (DVMP) module specifically tailored for machine vision tasks. This module optimizes the entropy coding process by dynamically predicting the skip/no-skip coding mode for each feature element, based on its relevance to machine vision. Specifically, it takes hyperprior data from motion or residual/contextual features within the DVC framework as the input and assesses the utility of each feature element for machine vision. By replacing non-essential elements with their predicted mean value derived from the hyperprior network, our method effectively reduces bit consumption and circumvent the entropy decoding step for these elements to expedite the decoding process. In this case, we generate a novel frame type that significantly reduces bitrate usage compared to traditional P frames in the Group of Pictures (GoP) structure, while maintaining the information valuable for machine vision tasks. This new frame is designated as the $P_{m}$ frame.", + "bbox": [ + 75, + 199, + 467, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "At the same time, it is important for DVC to maintain reconstruction quality because there exists significant reference relationship between frames, which is a crucial difference between image and video compression. Since our $P_{m}$ frame degrades the fidelity of reconstruction compared to the $P$ frame, we explore the reorganization of the Group of Pictures (GoP) structure. To be specific, we firstly explore to use a hand-crafted GoP structure, which already achieves significant rate-precision improvements. In the revised structure, both $P$ frames and $P_{m}$ frames derive their prediction from the preceding $I / P$ frame. Further, we introduce a GoP selection network, which can dynamically determine the GoP structure during the encoding process and achieves a better rate-precision trade-off.", + "bbox": [ + 75, + 473, + 467, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We choose three representative residual and conditional DVC methods as touchstones in our proposed video coding for machine vision framework. It's worth mentioning that our method utilizes pre-trained DVCs without altering the weights of the original encoders and decoders. Our control happens at encoding stage, leaving the decoders' architecture intact. Moreover, when human viewing is required, we can switch back to the original encoding procedure to restore reconstruction performance. In essence, using the proposed method, we can control the encoder of a DVC to adapt for both machine and human vision requirements.", + "bbox": [ + 75, + 686, + 467, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of our work are summarised as follows: (1) Built upon the mainstream DVC CODECs, we propose a novel video coding for machine vision framework", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that controls the encoder to adapt for different downstream vision tasks, such as video object detection and multi-object tracking. (2) We employ a dynamic vision mode prediction approach to refine the original P frame, effectively reducing the bitrate while preserving critical information pertinent to vision tasks. Furthermore, we utilize a GoP selection strategy to dynamically forecast the coding GoP structure during the encoding stage, which controls the bitstream for better rate-precision trade-off in downstream vision tasks. (3) Experiments show that our controlling method is novel and flexible for different DVC CODECs, achieving up to more than $25\\%$ bitrate savings in various downstream tasks.", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 287, + 648, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Video Compression", + "text_level": 1, + "bbox": [ + 500, + 311, + 684, + 329 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous video codecs are designed to remove spatial-temporal redundancies effectively. Traditional video coding standards, including H.264/AVC [42], H.265/HEVC [33], and H.266/VVC [2], significantly increase the compression efficiency of images and videos. Recently, a number of learning-based video codecs [14, 18, 24, 25, 29, 35] have been proposed, designed based on residual coding or conditional coding, achieved better and better pixel-wise signal quality metrics, e.g., PSNR and MS-SSIM [41], which mainly serve the human visual experience. Recently, there are also some generative video coding methods [27, 45] that mainly consider visual comfort and perceptual quality.", + "bbox": [ + 496, + 337, + 890, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Compression for Machine Vision", + "text_level": 1, + "bbox": [ + 500, + 529, + 785, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To facilitate the efficiency of machine vision tasks, early research works were devoted to extracting visual features from signals. Early standards, such as CDVA [8] and CDVS [7], suggest the pre-extraction and transportation of image keypoints to facilitate image indexing or retrieval tasks. With the development of deep learning, some studies began to explore connections between compression and downstream tasks, some studies [1, 3, 4, 11, 13, 20, 38, 40, 44, 47] also focus on the joint optimization of image compression and downstream machine vision tasks by introducing a rate-distortion optimization strategy guided by downstream tasks or by adding a task-specific feature encoding stream. For instance, Torfason et al. [38] introduced a method for executing image understanding tasks, such as classification and segmentation, on compressed outputs from learning-based image compression techniques. Furthermore, Lu et al. [26] enhanced traditional CODECs with a preprocessing step, thereby improving codec performance for downstream vision tasks. The field has also witnessed the emergence of self-supervised representation learning methods aimed at deriving compact semantic representations. Dubois et al. [9] presented a theoretical framework suggesting that the distortion term in the lossy rate", + "bbox": [ + 496, + 553, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26037", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/44edc55052840a0107689515e6c08b00e26a7d95015b77a3dae11631b6ec9bd2.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 135, + 94, + 540, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/69851fbeec1e47931de7146feb70572da047965607c43313fbdc543c62f0f00e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 108, + 816, + 189 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e855742e4276dd2605f2f1d2e1cde33146445af297fd7cc6a43ee77e846518a5.jpg", + "image_caption": [ + "(b)", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 148, + 227, + 710, + 329 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/868e6e6aa89606bc49050a15b87cbea5a532facb32768c0fcad6e441f1e409ab.jpg", + "image_caption": [ + "GoP Feature Extractor", + "Figure 2. (a) Overview of our \"Controlling DVC for Machine\" framework. Given an input GoP, we firstly use GoP Selection network to predict the GoP structure, then the predicted structure controls the encoding procedure to encode input frames for machine vision tasks. (b) The \"0\" element controls encoder to use DVMP. (c) The GoP Selection network, including the pre-analysis stage and GoP prediction stage." + ], + "image_footnote": [], + "bbox": [ + 709, + 213, + 825, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "distortion trade-off for image classification could be approximated by a contrastive learning objective. Feng et al. [10] proposed a method to learn a unified feature representation for AI tasks from unlabeled data. These methodologies necessitate fine-tuning the downstream models to adapt to the learned features.", + "bbox": [ + 75, + 416, + 468, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite advancements in image coding for machine applications, reconstructing high-fidelity video from extracted features remains a formidable challenge due to the critical inter-frame reference relationships. Addressing the requirements of both machine and human vision, Tian et al. [36] introduced a self-supervised edge representation as a semantic intermediary, which preserves the video's semantic structure. Furthermore, Lin et al. [21] developed a scalable video coding framework tailored for machines, segregating semantic features for machine analysis and human viewing into separate bitstreams.", + "bbox": [ + 75, + 507, + 468, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most existing approaches [21, 36, 37] necessitate the use of distinct encoders and decoders for various tasks, which complicates and burdens the coding pipelines. Our methodology advances the \"Coding for Machine\" paradigm by focusing on the encoder side. We present a versatile framework capable of accommodating both human and machine vision demands without the need to modify the pre-trained decoders of existing video codecs.", + "bbox": [ + 75, + 672, + 470, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Compressed Video Analysis/Understanding", + "text_level": 1, + "bbox": [ + 76, + 801, + 446, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are also amounts of works performing video analysis tasks [17, 23, 39], such as image recognition, action recognition [31, 43] and multiple object tracking (MOT) [6, 16], in the compressed video domain. However, these methods focus on developing video analysis models that better lever", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "age the partially decoded video stream, such as the motion vector. In contrast, our work focuses on the coding procedure, specifically the encoding procedure.", + "bbox": [ + 498, + 416, + 892, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 473, + 589, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 498, + 609, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our coding framework is illustrated in Fig. 2 (a). Let $\\mathcal{X} = \\{\\pmb{x}_1,\\dots ,\\pmb{x}_T\\}$ represent the video sequence. The framework categorizes the $P$ frames into two types: original $P$ frames and machine vision-specific $P_{m}$ frames. The categorization of frame $\\pmb{x}_t$ is determined by the GoP Selection Network, where a value of 1 signifies a $P$ frame and 0 a $P_{m}$ frame. The $P_{m}$ frames, designed to reduce the bitrate while preserving critical information for future vision tasks, are generated from the preceding $P$ frames using Dynamic Vision Mode Prediction (DVMP). To maintain the quality of the decoded video sequence, each $P$ frame is computed from the previous $P$ frame with superior visual quality, rather than from a $P_{m}$ frame. In scenarios requiring coding for downstream vision tasks, a machine-centric control applies the GoP selection and DVMP, altering the encoding structure to \"I, $P_{m},P,\\dots$ \". Conversely, for human viewing, a human-centric control restores the encoding structure to its original form, \"I,P,P,P,,..", + "bbox": [ + 496, + 521, + 892, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Dynamic Vision Mode Prediction", + "text_level": 1, + "bbox": [ + 500, + 801, + 790, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our coding framework, we aim to minimize the video sequence bitrate. Initial analyses of residual video CODECs reveal that they allocate a substantial portion of the bitrate—exceeding $80\\%$ in high bitrate configurations—to encode and transmit residual information. This approach en", + "bbox": [ + 496, + 825, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26038", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/57c3885ba644d58e3c8ff878e11daa3fc9de30a0238beeea7ef83aed31901a54.jpg", + "image_caption": [ + "Figure 3. Hyper-prior guided Dynamic Vision Mode Prediction network." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 450, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sures high-quality visuals for human perception, as measured by metrics such as PSNR or MS-SSIM [41]. However, this method introduces significant redundancy, especially since downstream tasks like video object detection or tracking predominantly concentrate on specific regions of interest rather than the entirety of the frame. Consequently, it is crucial for our study to develop an approach that reduces this redundancy in the coding bitstream.", + "bbox": [ + 75, + 308, + 467, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by Hu et al., who developed a mode prediction technique specifically designed for the human visual system within a slimmable encoder and decoder, we introduce the Dynamic Vision Mode Prediction (DVMP) module. This innovative module autonomously selects the optimal coding mode and decides whether each feature element should be encoded and transmitted, as illustrated in Fig. 3.", + "bbox": [ + 75, + 430, + 467, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To elucidate the functionality of our proposed Dynamic Vision Mode Prediction (DVMP) module, we utilize the encoded residual feature $m_{t}$ as an illustrative example. Assume $m_{t}$ possesses dimensions $c \\times h \\times w$ , indicating $c$ channels each with a spatial dimension of $h \\times w$ . The hyperprior network then predicts the mean and variance for each element, resulting in dimensions of $2c \\times h \\times w$ . The mode prediction network's architecture, detailed in the upper branch, consists of two convolution layers and three ResBlocks. To render the \"mode prediction\" process differentiable, we employ the Gumbel Softmax technique during training to ascertain the skip mode for each encoded residual feature element. This module generates a binary mask for the feature map $m_{t}$ , where a \"1\" denotes elements imperative for machine vision that require entropy coding, and a \"0\" signifies elements either irrelevant for machine vision or accurately predictable by the hyperprior network, thus streamlining the entropy coding process and bitrate reduction. Note that the showed architecture is not suitable for prior models with autoregressive components, like which in DCVC [18]. However, we can extend the module into an autoregressive way to enable the support for DCVC. The detailed structure is shown in the supplementary material.", + "bbox": [ + 75, + 537, + 467, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To optimize the DVMP module, the loss function can be", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/87de36ec866ecd879792b0c7bc1ff9fce01cf868895a9e0624f3c005a1254f0d.jpg", + "image_caption": [ + "Figure 4. Different coding GoP structures. The original structure consists of I and P frames. In the middle, the hand-crafted structure for machine vision consists of I, P and $P_{m}$ frames which are arranged alternately. In the predicted (GoP selected) structure, the type of frames are predicted by the GoP selection network, targeting on better bitrate and machine vision performance trade-off." + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 890, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "given by:", + "bbox": [ + 500, + 318, + 565, + 332 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m} = R + \\lambda_ {m} \\mathcal {L} _ {t} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 335, + 890, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $R$ denotes the coding bitrate and $\\mathcal{L}_t$ denotes the loss of downstream vision task, respectively. $\\lambda_{m}$ is a hyperparameter used to control the trade-off.", + "bbox": [ + 496, + 359, + 890, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. DivGoP $\\rightarrow$ GoP Selection", + "text_level": 1, + "bbox": [ + 500, + 417, + 728, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Obtaining the $P_{m}$ frames, we firstly use a hand-crafted structure in a GoP, where the $P_{m}$ frames and $P$ frames are arranged alternately (referred as \"DivGoP\" and shown in Fig.4). It is observed that this arrangement can bring a certain degree of bitrate saving in detection, since the $P_{m}$ frames cost much less bitrate than $P$ frames while maintaining the motion and machine vision information. Experimental findings, as depicted in the \"FVC DivGoP\" curve within Fig. 5, demonstrate that the \"I, $P_{m}, P, \\cdots$ \" structure markedly achieve about $18.6\\%$ bitrate saving.", + "bbox": [ + 496, + 441, + 890, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The \"DivGoP\" configuration, while methodically structured, may not be universally optimal due to the variability in motion and object information across different videos, which could necessitate distinct GoP structures. For instance, videos characterized by rapid target movement might demand a higher frequency of $P$ frames to preserve the quality of reconstruction, as an excess of $P_{m}$ frames may lead to degradation that negatively affects subsequent frames. Conversely, in scenarios where the camera remains static or the target movement is minimal, the introduction of more $P_{m}$ frames could be feasible without significantly impacting the reconstruction quality of following frames. Stimulated by this notion, our research endeavors to tailor the GoP structure dynamically, enhancing codec performance for machine vision tasks. To approach this issue, we adopt FVC [14] as the foundational codec, select a GoP size of 10, and address the detection task utilizing the MOT Dataset. The optimization problem is thus defined:", + "bbox": [ + 496, + 594, + 892, + 866 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\theta} {\\arg \\min } R (\\theta) + \\lambda_ {p} L _ {\\det } (\\theta) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 878, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26039", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $R$ denotes the bitrate, $L_{det}$ denotes the detection loss, $\\lambda_p$ is a weighting factor that balances the effects of bitrate and detection loss, and $\\theta$ denotes the setting of GoP structure here.", + "bbox": [ + 75, + 90, + 468, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Obtaining a GoP, each of the 9 predicted frames can be either $P$ frame or $P_{m}$ frame, which can be dynamically decided. To explore the optimal GoP structure for target function 2, and evaluate the performance gap between DivGoP and optimal GoP structure, we use a deep-first-searching (DFS) algorithm, where we use different $\\lambda$ for different bitrate models. The DFS algorithm searches for the structure that makes the target function smallest for each GoP. Results are shown in the \"FVC DFS\" curve of Fig.5. The application of the DFS algorithm led to a substantial increase in Bpp-mAP performance,, achieving about $32.3\\%$ bitrate saving in terms of the BD-RATE metric, while the \"Div-GoP\" method achieves about $18.6\\%$ bitrate saving. Similar conclusions can also be observed on DCVC [18] and TCM [29], reinforcing our hypothesis that distinct videos necessitate unique GoP structures tailored for machine vision tasks. However, the DFS algorithm is extremely time consuming, and the labels do not exist during real inference process, making the DFS not available to be applied in real time encoding. How to further find an \"adaptive GoP structure\" in an available time complexity is worth exploring.", + "bbox": [ + 75, + 154, + 470, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From this perspective, we introduce our GoP selection method to adaptively predict the GoP structure at encoding stage, which has a low time complexity and is available in application. Illustrated in Fig.2 (c), our GoP selection network operates in two phases: the pre-analysis and the GoP prediction stages. Initially, input frames undergo preanalysis, where a detector identifies objects and generates an overlapping mask for each frame, while the RAFT [34] optical flow estimation method calculates optical flow features for each predicted frame using its adjacent predecessor. These features are then refined by masking with the object bounding box mask to produce masked optical flow features, alongside motion feature priors (mean and variance, averaged by channel) derived from the codec's motion encoder. Subsequently, these processed features are input into the GoP prediction network, which employs a feature extractor with convolution blocks for downsampling and aggregation, followed by adaptive average pooling and linear layers that output a normalized $s_{\\text{logit}}$ for this GoP, which comprises two elements that sum to one. The averaged detection confidence information from the detector is also injected into the linear layers as meta data.", + "bbox": [ + 75, + 474, + 468, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Upon acquiring the $s_{\\mathrm{logit}}$ , we employ the Gumbel softmax sampling technique throughout the training phase to iteratively generate a binary value for each frame. This process culminates in the formation of a GoP structure vector, where the binary value—0 for a $P_{m}$ frame and 1 for a $P$ frame—specifies the type of each frame within the GoP.", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2ed356220b088e2248ecfe355c84f5531407d576b4f81e64620881d5621c41dc.jpg", + "image_caption": [ + "Figure 5. Left: Using DFS to search for the near-optimal GoP structure for Bpp-mAP. Right: Results of simply fine-tuning FVC" + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 697, + 218 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a94aa44b7630bcc5bf4f5233448509af4db142350cd6c437bbc3c956deef741d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 89, + 890, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference stage, we apply a softmax operation on $s_{\\mathrm{logit}}$ to derive probabilities, which are then used to evenly distribute 1s and 0s, thereby determining the actual encoded GoP structure. Similar to our mode prediction network, the loss function for optimizing GoP structure selection network is given by:", + "bbox": [ + 498, + 282, + 890, + 373 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g} = \\bar {R} + \\lambda_ {g} \\bar {\\mathcal {L}} _ {t} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 391, + 890, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\bar{R}$ and $\\bar{\\mathcal{L}}_t$ denote the GoP-wise average coding bitrate and average loss of downstream vision task.", + "bbox": [ + 498, + 417, + 890, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 465, + 632, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe the dataset settings, downstream task models, and evaluation metrics. We then provide empirical evidence underscoring the effectiveness of our proposed method across multiple tasks and CODECs. Lastly, we validate the contributions of the mode prediction and GoP selection modules to performance improvements through ablation studies.", + "bbox": [ + 498, + 489, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. For multi-object tracking (MOT) task, we choose MOT17 dataset. For video object detection (VOD) task, we adopt the Imagenet VID dataset [28] for training and evaluation. For video action recognition (VAR) task, we adopt UCF101[32] dataset. For the data processing procedure, we downsample the short side of the evaluation videos to 512 for MOT17 dataset and keep the original frame ratios for both training and evaluation. For the evaluation of VOD, we follow the most VOD works to resize the frames to $512 \\times 512$ for training and $576 \\times 576$ for evaluation.", + "bbox": [ + 498, + 597, + 890, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Downstream Task Models. For MOT task, we adipot the popular ByteTrack [46] method as the backbone model. The model weights are provided by their official framework. For VOD task, we adipot the detector YOLOv [30], and the model weights are also provided by its official framework. For the VAR task, we adipot the TSM [22] model as the recognition network, the weights are provided by mmaction2 [5]. During training and evaluation of our proposed framework, the weights of the downstream task models are fixed, without any fine-tuning.", + "bbox": [ + 498, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26040", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5f27cbeccdaea6c972b064752557970cd151ab1783920f8dbdf7fe833161431f.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 356, + 265 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6339a848c9c0d0b8aafaaa4ceee93a618d0f788d98965b1f5946a1e864d01bed.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 356, + 90, + 629, + 266 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/393fcbc5c3537694ff4da72484cd5e1f43ccde4a5bf417156831021d1558e4ca.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 629, + 90, + 900, + 266 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7f1c9c2addf59ee8646531406ddd2c9ed7dc895a7ae754517308db8b67a0173a.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 80, + 286, + 354, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/85c59007fe6b17fa1a765799494cd46b181e3955d57534432f7127b717baa2c1.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 354, + 286, + 627, + 462 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ff2b4c14824dae2fe4f19ec9fbc47f428c02035dbb41de37c665f0fc56796b9a.jpg", + "image_caption": [ + "(f)", + "Figure 6. (a) Bpp-MOTA curves on MOT Dataset. (b) Bpp-mAP curves on MOT Dataset. (c) Bpp-MOTP curves on MOT Dataset. (d) Bpp-FN curves on MOT Dataset. (e) Bpp-mAP50 curves on ImageNet VID Dataset using YOLOv [30]. (f) Bpp-Top1 Accuracy curves on UCF101 Dataset." + ], + "image_footnote": [], + "bbox": [ + 629, + 286, + 901, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For compression, We use bpp (bits per pixel) to measure the average number of bits for one pixel. For the MOT task, we adopt the MOTA (multiple object tracking accuracy), MOTP (multiple object tracking precision), FN (false negative detection number) and mAP (mean average precision) metrics to evaluate the tracking and detection performance. For the VOD task, we adopt the mAP50 metric to evaluate the detection performance, following most VOD methods. For the VAR task, we adopt the Top1 Accuracy as the performance indicator.", + "bbox": [ + 75, + 551, + 468, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our whole framework is implemented on PyTorch with CUDA support and trained on eight V100 or A100 GPUs. We use the Adam optimizer by setting the learning rate as 0.0001, $\\beta_{1}$ as 0.9 and $\\beta_{2}$ as 0.999, respectively. The whole training process has the following stages: Firstly, we use the loss function 1 to train the mode prediction network and obtain the $P_{m}$ frames. This stage we use single frame training. Note that for TCM, DVMP is applied to both motion and contextual information. For DCVC and FVC, DVMP is exclusively utilized for motion information, and all residual/contextual information is omitted. After that, we introduce the GoP selection network in a GOP length unit. Using the GoP size of 32", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on MOT and VAR tasks, we decide the GoP into 2 mini GoPs with size of 16, and the predicted vector of each mini GoP has a length of 15. For VOD task, we directly set the GoP size as 16. Using multi-frame training and loss function 3, we fix the weights of mode prediction network and train GoP selection network for about 100 epochs on MOT half-train dataset and 20 epochs on ImageNet VID training dataset, respectively for multi-object tracking and video object detection tasks. The whole training procedure takes about 2 days for each task.", + "bbox": [ + 496, + 551, + 892, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Multi-Object Tracking", + "text_level": 1, + "bbox": [ + 500, + 710, + 712, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 6, we present a comparison between our method and various codecs, specifically DCVC, FVC, and TCM. It is evident that our approach yields a markedly improved tradeoff in terms of Bpp-MOTA, Bpp-mAP, Bpp-MOTP and Bpp-FN metrics. When compared to the baseline codecs FVC, DCVC and TCM, our control method results in bitrate savings of approximately $27.54\\%$ , $41.82\\%$ and $12.64\\%$ , respectively, for equivalent MOTA metrics. With DCVC established as the anchor baseline, Table 1 displays the BD-RATE values, where our controlled TCM, controlled DCVC and controlled FVC codecs demonstrate the best, second-", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26041", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/404416386c280a5d2d2b2418a7ce691eee76d85e1afed2fb87244770c7914eac.jpg", + "image_caption": [ + "Figure 7. Video Reconstruction Quality in terms of Bpp-PSNR on HEVC Class B dataset (Left) and Class C dataset (Right)" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 277, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/39c4a621b853dc3fa1fc774816057b2c4c8788d14f1102c083c4ac1cedd67c96.jpg", + "image_caption": [ + "Figure 8. Left: Ablation study for Mode Prediction network. Right: Ablation study for GoP Selection network" + ], + "image_footnote": [], + "bbox": [ + 282, + 90, + 473, + 215 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d4b5da86dbae6e382526bf12c2bf6ba2eade776f61402bc697df2e526973d8c1.jpg", + "table_caption": [ + "Table 1. BD-RATE results on multi-object tracking." + ], + "table_footnote": [], + "table_body": "
MethodMOTAmAPmAP50MOTPFNAverage
DCVC [18]0.00.00.00.00.00.0
Ours+DCVC-41.82-39.43-40.60-37.73-41.74-40.26
FVC [14]-5.32-9.75-14.53-1.39-6.20-7.44
Ours+FVC-34.28-31.27-32.09-35.02-32.89-33.11
TCM [29]-25.19-32.34-31.02-39.85-26.44-30.97
Ours+TCM-40.82-45.10-46.15-51.66-38.98-44.54
HEVC [33]-33.88-31.35-40.4315.80-34.02-24.78
", + "bbox": [ + 80, + 292, + 467, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "best and third-best performance, respectively. Additionally, we contrast our method with the task-oriented \"One-to-one Codes\" of VCM, with detailed results provided in the supplementary materials.", + "bbox": [ + 75, + 426, + 467, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Video Object Detection", + "text_level": 1, + "bbox": [ + 76, + 497, + 290, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the evaluation of the video object detection task, we juxtapose our method with established codecs such as FVC, and TCM, with the comparative results illustrated in Fig. 6(e). Our method is noted to achieve a superior tradeoff, particularly in the Bpp-mAP50 metric, when applied to the downstream model YOLOv. For instance, in comparison to the FVC and TCM baseline codecs, our method facilitates bitrate reductions of approximately $21.76\\%$ and $20.42\\%$ , respectively, while maintaining the same mAP50 metric. The uncompressed video represents the upper bound for the mAP50 metric, recorded at 76.4, and is delineated by the grey line within the figure.", + "bbox": [ + 75, + 520, + 467, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Video Action Recognition", + "text_level": 1, + "bbox": [ + 76, + 710, + 308, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted an evaluation of our method on the UCF101 dataset, benchmarking against TCM and FVC CODECs. The outcomes of this comparison are depicted in Fig. 6(f). For the task of video action recognition, rather than initiating a new set of network trainings, we employed pre-trained mode prediction networks previously utilized for the MOT task and use the \"DivGoP\" structure for validation. The original videos were compressed using the various CODECs, and the resulting decoded videos served as inputs to the TSM model as cited in Lin et al. [22]. The experimental results indicate that our method maintains efficacy in this", + "bbox": [ + 75, + 734, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75f089c755a2ddff914016c4f10bb666a8d8e7096d9a74361da105160f9f9cf1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 89, + 696, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb9e6c73e1c8d49c94adc78b78b8e57d02f69306ef795f1380b24c24853aefc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 89, + 890, + 215 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "context, securing approximately $21.81\\%$ and $26.03\\%$ in bitrate savings for TCM and FVC, respectively.", + "bbox": [ + 498, + 281, + 888, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Video Reconstruction Quality", + "text_level": 1, + "bbox": [ + 500, + 324, + 764, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When it comes to the performance of reconstructing the video for human viewing, the situation is straightforward for conventional video codecs, no matter learned or traditional ones: just encode the frames, transmit and decode the compressed bitstream, and measure their quality using appropriate metrics (PSNR, MS-SSIM, etc). In our method, there're two options: Control the encoding procedure for human viewing, which is the ideal way for best frame reconstruction quality, or just continue the encoding procedure used for machine vision tasks.", + "bbox": [ + 496, + 349, + 890, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that our method does not change the weights of the original encoder and decoder in pre-trained DVC, so using this option can directly recover the human viewing quality to original best performance of the CODECs. On the other hand, we also evaluate the human viewing quality of the encoded bitstream for machine, using \"DivGop\" encoding structure on HEVC Class B and Class C test datasets. These are referred to as \"FVC DivGoP\" and \"TCM DivGoP\" in Fig.7. Results show that our control for machine does decrease the video reconstruction quality of the original TCM codec by about 1dB in PSNR. Besides, it is also observed that the PSNR of residual codec FVC drops more sharply than that of conditional CODEs like TCM. However, these are acceptable since we can always adjust the encoder to compress the video for human viewing when it's needed.", + "bbox": [ + 496, + 501, + 890, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 739, + 651, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mode Prediction. In our method, we propose mode prediction network to decide the coding modes of motion and residual/contextual feature elements for machine vision, mainly targeting on reducing the bitrate and keep the crucial information for important objects in the frame. Moreover, mainstream DVC, including FVC, DCVC, TCM and others, use motion estimation and warping operation to generate a rough predicted frame, which we call $P_r$ frame here. The $P_r$ frame also meets our requirement of \"consuming", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26042", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ab69aa52beebe4eaa050237683d1ccddd4a3c99725541767d4e7a6108e9ad335.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Input imageContextual skip mode maskMasked motion featureMasked Contextual feature
", + "bbox": [ + 83, + 89, + 462, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 9. Visualization results of the mode prediction networks.", + "bbox": [ + 81, + 238, + 460, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lower code streams but retaining object and motion information\" to a certain extent. So comparing it with the $P_{m}$ frame used in our method is also a point worthy of attention. In \"DivGoP\" structure, we use the $P_{r}$ frame to replace the original $P_{m}$ frames, and the GoP structure will be $[I, P_{r}, P, P_{r}, \\ldots]$ . As shown in the left figure in Fig.8, based on DCVC and FVC, we compare the structure using $P_{r}$ frames with the original \"DivGoP\" structure using $P_{m}$ frames, corresponding to the \"DCVC/FVC DivGoP w $P_{r}$ \" and \"DCVC/FVC DivGoP\" curves respectively. It is observed that the \"DivGoP w $P_{r}$ \" curve also achieves about $28.68\\%$ and $16.24\\%$ bitrate savings for DCVC and FVC, respectively. But our \"DivGoP\" curve shows better trade-off, by achieving $32.67\\%$ and $19.43\\%$ bitrate savings.", + "bbox": [ + 75, + 266, + 468, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GoP Selection. We prove the effectiveness of Gop Selection module in terms of rate-precision performance on MOT dataset. As shown in the right figure in Fig. 8, we compare the Bpp-mAP performance of \"DCVC/FVC Gop Selection\" and \"DCVC/FVC DivGoP\" structure on MOT Dataset. Using DCVC as an example, experiment results show that our Gop Selection achieves about $39.43\\%$ bitrate saving, which does show better trade-off than that $32.67\\%$ bitrate saving of hand-crafted \"DivGoP\" structure.", + "bbox": [ + 75, + 477, + 468, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Simply fine-tuning DVC for machine vision tasks. We employed the loss function $L_{f} = R + \\lambda_{1}MSE + \\lambda_{2}L_{det}$ to fine-tune the FVC encoder and the entire FVC codec for the MOT task. As shown in Fig. 5 right, while simple finetuning improves accuracy, it also notably increases the bitrate. This phenomenon could be attributed to the degradation in the fidelity of decoded frames when optimized for downstream tasks. Due to the inter-frame reference nature of video codecs, this degradation is propagated across frames, resulting in a high bitrate consumption. This observation aligns with findings reported in DeepSVC [21], indicating that such task-specific performance can not be achieved with simply finetuning a DVC. This also reflects the superiority of our method.", + "bbox": [ + 75, + 613, + 467, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization Results. We present the visualization results of our mode prediction network, as shown in Fig. 9. The first column shows the input images, the second column shows the skip mode masks for contextual features (averaged by channel), and the third and fourth columns are", + "bbox": [ + 75, + 824, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bc35d74ba1c37644a2c7d503bb8a9d818fd6371d99c7820e6ffb7911afc0abba.jpg", + "table_caption": [ + "Table 2. Encoding Latency vs. Performance." + ], + "table_footnote": [], + "table_body": "
MethodEncoding Latency(s)BD-BR Performance
DCVC1.6750.0
Ours+DCVC1.738-40.26
FVC0.116-7.44
Ours+FVC0.159-33.11
", + "bbox": [ + 501, + 90, + 883, + 175 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "masked motion and contextual features, respectively, where the red color indicates large value and green color indicates small value. It is observed that the predicted coding mode mask for contextual features kept the regions of the small objects in the frames, and masked the conspicuous objects which are easy to recognize and the background regions. This shows the network's preference for objects in the frame. As for the motion information, it is observed that the predicted masks kept most of the elements of moving objects, which are crucial for detection and tracking.", + "bbox": [ + 496, + 186, + 890, + 337 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Complexity and Encoding Latency Analysis. The parameter numbers of our dynamic vision mode prediction network and GoP selection network are 1.48M and 5.36M. Since our method controls the video encoder during the encoding phase, such an operation brings encoding latency. To measure the impact on the encoding phase, we include a comparison of average encoding latency versus performance on MOT Dataset $(960 \\times 512)$ in Tab. 2. While our method brings an increase in encoding latency (1.738s vs. 1.675s), the improvement in R-P performance is substantial. Besides, the decoding procedure and decoder side models are not changed, which brings convenience to the deployment of the decoder and support for multiple tasks.", + "bbox": [ + 496, + 339, + 890, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 554, + 617, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a flexible framework for deep video compression models, where the pre-trained encoders can be controlled to change the encoding pipeline for machine vision tasks, achieving significant better bppprecision trade-off without changing the decoders or decoding procedure. The controlling methods are attributed to two main developments, including dynamic vision mode prediction network and GoP structure selection network. Experiment results show that our framework is general for residual and conditional deep video CODECs and different downstream vision tasks like detection, tracking and action recognition. Our comprehensive evaluation, conducted against existing deep video CODECs, confirms the superior performance of our models, establishing a new benchmark for future deep video coding for machine. The effectiveness of each component is clearly evidenced through extensive ablation studies.", + "bbox": [ + 496, + 580, + 890, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by the National Natural Science Fund of China (Project No. 42201461) and the General Research Fund (Project No. 16209622) from the Hong Kong Research Grants Council.", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26043", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mohammad Akbari, Jie Liang, and Jingning Han. Dsslic: Deep semantic segmentation-based layered image compression. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2042-2046. IEEE, 2019. 1, 2", + "[2] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J. Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 1, 2", + "[3] Zhuo Chen, Kui Fan, Shiqi Wang, Lingyu Duan, Weisi Lin, and Alex Chichung Kot. Toward intelligent sensing: Intermediate deep feature compression. IEEE Transactions on Image Processing, 29:2230-2243, 2020. 2", + "[4] Hyomin Choi and Ivan V Bajic. Scalable image coding for humans and machines. IEEE Transactions on Image Processing, 31:2739-2754, 2022. 1, 2", + "[5] MMAction2 Contributors. Openmmlab's next generation video understanding toolbox and benchmark. https://github.com/open-mmlab/mmaction2, 2020.5", + "[6] Yan Dai, Ziyu Hu, Shuqi Zhang, and Lianjun Liu. A survey of detection-based video multi-object tracking. Displays, page 102317, 2022. 3", + "[7] Ling-Yu Duan, Tiejun Huang, and Wen Gao. Overview of the mpeg cdvs standard. In 2015 Data Compression Conference, pages 323-332, 2015. 2", + "[8] Ling-Yu Duan, Yihang Lou, Yan Bai, Tiejun Huang, Wen Gao, Vijay Chandrasekhar, Jie Lin, Shiqi Wang, and Alex Chichung Kot. Compact descriptors for video analysis: The emerging mpg standard. IEEE MultiMedia, 26(2): 44-54, 2019. 2", + "[9] Yann Dubois, Benjamin Bloem-Reddy, Karen Ullrich, and Chris J Maddison. Lossy compression for lossless prediction. Advances in Neural Information Processing Systems, 34:14014–14028, 2021. 2", + "[10] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In European Conference on Computer Vision, pages 510-528. Springer, 2022. 3", + "[11] Kristian Fischer, Fabian Brand, and André Kaup. Boosting neural image compression for machines using latent space masking. arXiv preprint arXiv:2112.08168, 2021. 1, 2", + "[12] Hadi Hadizadeh and Ivan V Bajić. Learned scalable video coding for humans and machines. arXiv preprint arXiv:2307.08978, 2023.", + "[13] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 1, 2", + "[14] Zhihao Hu, Guo Lu, and Dong Xu. Fvc: A new framework towards deep video compression in feature space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1502-1511, 2021. 1, 2, 4, 7" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Zhihao Hu, Guo Lu, Jinyang Guo, Shan Liu, Wei Jiang, and Dong Xu. Coarse-to-fine deep video coding with hyperprior-guided mode prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5921-5930, 2022. 1", + "[16] Sayed Hossein Khatoonabadi and Ivan V Bajic. Video object tracking in the compressed domain using spatio-temporal markov random fields. IEEE transactions on image processing, 22(1):300-313, 2012. 3", + "[17] Congcong Li, Xinyao Wang, Longyin Wen, Dxiang Hong, Tiejian Luo, and Libo Zhang. End-to-end compressed video representation learning for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13967-13976, 2022. 3", + "[18] Jiahao Li, Bin Li, and Yan Lu. Deep contextual video compression. Advances in Neural Information Processing Systems, 34:18114-18125, 2021. 1, 2, 4, 5, 7", + "[19] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22616-22626, 2023. 1", + "[20] Xin Li, Jun Shi, and Zhibo Chen. Task-driven semantic coding via reinforcement learning. IEEE Transactions on Image Processing, 30:6307-6320, 2021. 1, 2", + "[21] Hongbin Lin, Bolin Chen, Zhichen Zhang, Jielian Lin, Xu Wang, and Tiesong Zhao. Deep scalable video coding for both machine and human vision. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9205-9214, 2023. 3, 8", + "[22] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7083-7093, 2019. 5, 7", + "[23] Qiankun Liu, Bin Liu, Yue Wu, Weihai Li, and Nenghai Yu. Real-time online multi-object tracking in compressed domain. arXiv preprint arXiv:2204.02081, 2022. 3", + "[24] Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, and Zhiyong Gao. Dvc: An end-to-end deep video compression framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11006-11015, 2019. 1, 2", + "[25] Guo Lu, Xiaoyun Zhang, Wanli Ouyang, Li Chen, Zhiyong Gao, and Dong Xu. An end-to-end learning framework for video compression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3292-3308, 2021. 1, 2", + "[26] Guo Lu, Xingtong Ge, Tianxiong Zhong, Jing Geng, and Qiang Hu. Preprocessing enhanced image compression for machine vision. arXiv preprint arXiv:2206.05650, 2022. 2", + "[27] Fabian Mentzer, Eirikur Agustsson, Johannes Balle, David Minnen, Nick Johnston, and George Toderici. Neural video compression using gans for detail synthesis and propagation. In European Conference on Computer Vision, pages 562-578. Springer, 2022. 2", + "[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26044", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5", + "[29] Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. Temporal context mining for learned video compression. IEEE Transactions on Multimedia, 2022. 1, 2, 5, 7", + "[30] Yuheng Shi, Naiyan Wang, and Xiaojie Guo. Yolov: making still image object detectors great at video object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2254-2262, 2023. 5, 6", + "[31] Zheng Shou, Xudong Lin, Yannis Kalantidis, Laura Sevilla-Lara, Marcus Rohrbach, Shih-Fu Chang, and Zhicheng Yan. Dmc-net: Generating discriminative motion cues for fast compressed video action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1268-1277, 2019. 3", + "[32] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5", + "[33] Gary J. Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on Circuits and Systems for Video Technology, 22(12):1649-1668, 2012. 1, 2, 7", + "[34] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 5", + "[35] Yuan Tian, Guo Lu, Xiongkuo Min, Zhaohui Che, Guang-tao Zhai, Guodong Guo, and Zhiyong Gao. Self-conditioned probabilistic learning of video rescaling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4490-4499, 2021. 2", + "[36] Yuan Tian, Guo Lu, Guangtao Zhai, and Zhiyong Gao. Nonsementics suppressed mask learning for unsupervised video semantic compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13610-13622, 2023. 1, 3", + "[37] Yuan Tian, Guo Lu, Yichao Yan, Guangtao Zhai, Li Chen, and Zhiyong Gao. A coding framework and benchmark towards low-bitrate video understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3", + "[38] Robert Torfason, Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Towards image understanding from deep compression without decoding. arXiv preprint arXiv:1803.06131, 2018. 1, 2", + "[39] Shiyao Wang, Hongchao Lu, and Zhidong Deng. Fast object detection in compressed video. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7104-7113, 2019. 3", + "[40] Shurun Wang, Shiqi Wang, Wenhan Yang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. Towards analysis-friendly face representation with scalable feature and texture compression. IEEE Transactions on Multimedia, 2021. 1, 2", + "[41] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402. IEEE, 2003. 2, 4" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] T. Wiegand, G.J. Sullivan, G. Bjontegaard, and A. Luthra. Overview of the h.264/avc video coding standard. IEEE Transactions on Circuits and Systems for Video Technology, 13(7):560-576, 2003. 1, 2", + "[43] Chao-Yuan Wu, Manzil Zaheer, Hexiang Hu, R Manmatha, Alexander J Smola, and Philipp Krahenbuhl. Compressed video action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6026-6035, 2018. 3", + "[44] Ning Yan, Changsheng Gao, Dong Liu, Houqiang Li, Li Li, and Feng Wu. Sssic: semantics-to-signal scalable image coding with learned structural representations. IEEE Transactions on Image Processing, 30:8939-8954, 2021. 1, 2", + "[45] Ren Yang, Luc Van Gool, and Radu Timofte. Perceptual learned video compression with recurrent conditional gan. arXiv preprint arXiv:2109.03082, 1, 2021. 2", + "[46] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In European Conference on Computer Vision, pages 1-21. Springer, 2022. 5", + "[47] Ezgi Özyilkan, Mateen Ulhaq, Hyomin Choi, and Fabien Racapé. Learned disentangled latent representations for scalable image coding for humans and machines. In 2023 Data Compression Conference (DCC), pages 42-51, 2023. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 445 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26045", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Task-Aware Encoder Control for Deep Video Compression/b978be40-5174-4abc-9649-d7daa2d37c89_model.json b/2024/Task-Aware Encoder Control for Deep Video Compression/b978be40-5174-4abc-9649-d7daa2d37c89_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c1e866cbd4fca55a74d987459004525153a24ae8 --- /dev/null +++ b/2024/Task-Aware Encoder Control for Deep Video Compression/b978be40-5174-4abc-9649-d7daa2d37c89_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.19, + 0.131, + 0.783, + 0.154 + ], + "angle": 0, + "content": "Task-Aware Encoder Control for Deep Video Compression" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.179, + 0.823, + 0.291 + ], + "angle": 0, + "content": "Xingtong Ge\\(^{1,2*}\\) Jixiang Luo\\(^{2}\\) Xinjie Zhang\\(^{3}\\) Tongda Xu\\(^{4}\\) Guo Lu\\(^{5}\\) \nDailan He\\(^{6}\\) Jing Geng\\(^{1\\ddagger}\\) Yan Wang\\(^{4}\\) Jun Zhang\\(^{3}\\) Hongwei Qin\\(^{2}\\) \n\\(^{1}\\)Beijing Institute of Technology \\(^{2}\\)SenseTime Research \n\\(^{3}\\)The Hong Kong University of Science and Technology \n\\(^{4}\\)Institute for AI Industry Research (AIR), Tsinghua University \n\\(^{5}\\)Shanghai Jiaotong University \\(^{6}\\)The Chinese University of Hong Kong" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.354, + 0.474, + 0.645 + ], + "angle": 0, + "content": "Prior research on deep video compression (DVC) for machine tasks typically necessitates training a unique codec for each specific task, mandating a dedicated decoder per task. In contrast, traditional video codecs employ a flexible encoder controller, enabling the adaptation of a single codec to different tasks through mechanisms like mode prediction. Drawing inspiration from this, we introduce an innovative encoder controller for deep video compression for machines. This controller features a mode prediction and a Group of Pictures (GoP) selection module. Our approach centralizes control at the encoding stage, allowing for adaptable encoder adjustments across different tasks, such as detection and tracking, while maintaining compatibility with a standard pre-trained DVC decoder. Empirical evidence demonstrates that our method is applicable across multiple tasks with various existing pre-trained DVCs. Moreover, extensive experiments demonstrate that our method outperforms previous DVC by about \\(25\\%\\) bitrate for different tasks, with only one pre-trained decoder." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.688, + 0.21, + 0.705 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.47, + 0.851 + ], + "angle": 0, + "content": "Over the past decades, video analysis techniques have proliferated across a variety of fields, including smart cities, autonomous vehicles, and traffic surveillance. For these applications, videos are often compressed before being transmitted to cloud-based systems for further machine vision analyses, such as object detection and tracking. However, current video compression techniques, which range from traditional codecs [2, 33, 42] to recent learning-based codecs [14, 18, 24, 25, 29], primarily cater to the human vi" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.321, + 0.891, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.457, + 0.895, + 0.514 + ], + "angle": 0, + "content": "Figure 1. (a) Mainstream video codec that serves the human viewing. (b) Our controlled video codecs for machine vision with fixed decoder. (c) Other video codec for machine vision with one-to-one encoders and decoders." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.893, + 0.583 + ], + "angle": 0, + "content": "sual system, as shown in Fig. 1(a). This specificity leads to inefficiencies, as machine vision tasks typically focus on selective semantic details and regions within the video frames, rather than the whole frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.895, + 0.854 + ], + "angle": 0, + "content": "To tackle this problem, recent studies [1, 4, 11-13, 20, 38, 40, 44, 47] have explored scalable compression for multiple tasks through different layers in image and video coding. They propose a base layer dedicated to machine vision, with an enhancement layer containing additional information for human vision. Additionally, the latest work [36] has tried to compress semantic features at the encoder side, which serve as side information to supplement decoded images with more semantic details for machine vision at lower bitrates. Despite these advances, these methods require individually customized CODECs when applied to different downstream tasks, as the \"One-to-one CODECs\" shown in Fig. 1(c). In contrast, traditional CODECs can control the encoding process for different objectives, like PSNR and SSIM optimization, using a single decoder. Inspired by this feature, in this paper, we focus on how to use one pre-trained DVC-decoder to support both human and multiple machine vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In both residual [14, 15, 24] and conditional [18, 19, 29] DVC series, the encoded bitstreams predominantly comprise two parts: motion and residual/contextual informa" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.862, + 0.461, + 0.876 + ], + "angle": 0, + "content": "*Work was done when Xingtong Ge interned at SenseTime Research" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.876, + 0.235, + 0.888 + ], + "angle": 0, + "content": "\\(\\dagger\\) xingtong.ge@gmail.com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.336, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\)Corresponding author, janegeng@bit.edu.cn" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.862, + 0.461, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26036" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.198 + ], + "angle": 0, + "content": "tion. It is observed that a substantial portion of the bitrate, often exceeding \\(80\\%\\) in high bitrate models, is allocated to the residual/contextual information. This allocation is primarily for high-quality frame reconstruction. However, this approach is inappropriate for downstream vision tasks such as tracking and detection, which primarily concentrate on the objects and their movements rather than the full frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.468, + 0.472 + ], + "angle": 0, + "content": "To mitigate this issue, we introduce a dynamic vision mode prediction (DVMP) module specifically tailored for machine vision tasks. This module optimizes the entropy coding process by dynamically predicting the skip/no-skip coding mode for each feature element, based on its relevance to machine vision. Specifically, it takes hyperprior data from motion or residual/contextual features within the DVC framework as the input and assesses the utility of each feature element for machine vision. By replacing non-essential elements with their predicted mean value derived from the hyperprior network, our method effectively reduces bit consumption and circumvent the entropy decoding step for these elements to expedite the decoding process. In this case, we generate a novel frame type that significantly reduces bitrate usage compared to traditional P frames in the Group of Pictures (GoP) structure, while maintaining the information valuable for machine vision tasks. This new frame is designated as the \\( P_{m} \\) frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.468, + 0.685 + ], + "angle": 0, + "content": "At the same time, it is important for DVC to maintain reconstruction quality because there exists significant reference relationship between frames, which is a crucial difference between image and video compression. Since our \\(P_{m}\\) frame degrades the fidelity of reconstruction compared to the \\(P\\) frame, we explore the reorganization of the Group of Pictures (GoP) structure. To be specific, we firstly explore to use a hand-crafted GoP structure, which already achieves significant rate-precision improvements. In the revised structure, both \\(P\\) frames and \\(P_{m}\\) frames derive their prediction from the preceding \\(I / P\\) frame. Further, we introduce a GoP selection network, which can dynamically determine the GoP structure during the encoding process and achieves a better rate-precision trade-off." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.468, + 0.853 + ], + "angle": 0, + "content": "We choose three representative residual and conditional DVC methods as touchstones in our proposed video coding for machine vision framework. It's worth mentioning that our method utilizes pre-trained DVCs without altering the weights of the original encoders and decoders. Our control happens at encoding stage, leaving the decoders' architecture intact. Moreover, when human viewing is required, we can switch back to the original encoding procedure to restore reconstruction performance. In essence, using the proposed method, we can control the encoder of a DVC to adapt for both machine and human vision requirements." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The main contributions of our work are summarised as follows: (1) Built upon the mainstream DVC CODECs, we propose a novel video coding for machine vision framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "that controls the encoder to adapt for different downstream vision tasks, such as video object detection and multi-object tracking. (2) We employ a dynamic vision mode prediction approach to refine the original P frame, effectively reducing the bitrate while preserving critical information pertinent to vision tasks. Furthermore, we utilize a GoP selection strategy to dynamically forecast the coding GoP structure during the encoding stage, which controls the bitstream for better rate-precision trade-off in downstream vision tasks. (3) Experiments show that our controlling method is novel and flexible for different DVC CODECs, achieving up to more than \\(25\\%\\) bitrate savings in various downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.288, + 0.65, + 0.304 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.313, + 0.686, + 0.33 + ], + "angle": 0, + "content": "2.1. Video Compression" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.519 + ], + "angle": 0, + "content": "Previous video codecs are designed to remove spatial-temporal redundancies effectively. Traditional video coding standards, including H.264/AVC [42], H.265/HEVC [33], and H.266/VVC [2], significantly increase the compression efficiency of images and videos. Recently, a number of learning-based video codecs [14, 18, 24, 25, 29, 35] have been proposed, designed based on residual coding or conditional coding, achieved better and better pixel-wise signal quality metrics, e.g., PSNR and MS-SSIM [41], which mainly serve the human visual experience. Recently, there are also some generative video coding methods [27, 45] that mainly consider visual comfort and perceptual quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.53, + 0.787, + 0.545 + ], + "angle": 0, + "content": "2.2. Compression for Machine Vision" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To facilitate the efficiency of machine vision tasks, early research works were devoted to extracting visual features from signals. Early standards, such as CDVA [8] and CDVS [7], suggest the pre-extraction and transportation of image keypoints to facilitate image indexing or retrieval tasks. With the development of deep learning, some studies began to explore connections between compression and downstream tasks, some studies [1, 3, 4, 11, 13, 20, 38, 40, 44, 47] also focus on the joint optimization of image compression and downstream machine vision tasks by introducing a rate-distortion optimization strategy guided by downstream tasks or by adding a task-specific feature encoding stream. For instance, Torfason et al. [38] introduced a method for executing image understanding tasks, such as classification and segmentation, on compressed outputs from learning-based image compression techniques. Furthermore, Lu et al. [26] enhanced traditional CODECs with a preprocessing step, thereby improving codec performance for downstream vision tasks. The field has also witnessed the emergence of self-supervised representation learning methods aimed at deriving compact semantic representations. Dubois et al. [9] presented a theoretical framework suggesting that the distortion term in the lossy rate" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "26037" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.095, + 0.541, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.214, + 0.377, + 0.225 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.109, + 0.818, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.212, + 0.705, + 0.224 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.228, + 0.712, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.332, + 0.525, + 0.343 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.214, + 0.826, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.333, + 0.826, + 0.343 + ], + "angle": 0, + "content": "GoP Feature Extractor" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.359, + 0.895, + 0.403 + ], + "angle": 0, + "content": "Figure 2. (a) Overview of our \"Controlling DVC for Machine\" framework. Given an input GoP, we firstly use GoP Selection network to predict the GoP structure, then the predicted structure controls the encoding procedure to encode input frames for machine vision tasks. (b) The \"0\" element controls encoder to use DVMP. (c) The GoP Selection network, including the pre-analysis stage and GoP prediction stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.47, + 0.506 + ], + "angle": 0, + "content": "distortion trade-off for image classification could be approximated by a contrastive learning objective. Feng et al. [10] proposed a method to learn a unified feature representation for AI tasks from unlabeled data. These methodologies necessitate fine-tuning the downstream models to adapt to the learned features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.673 + ], + "angle": 0, + "content": "Despite advancements in image coding for machine applications, reconstructing high-fidelity video from extracted features remains a formidable challenge due to the critical inter-frame reference relationships. Addressing the requirements of both machine and human vision, Tian et al. [36] introduced a self-supervised edge representation as a semantic intermediary, which preserves the video's semantic structure. Furthermore, Lin et al. [21] developed a scalable video coding framework tailored for machines, segregating semantic features for machine analysis and human viewing into separate bitstreams." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.794 + ], + "angle": 0, + "content": "Most existing approaches [21, 36, 37] necessitate the use of distinct encoders and decoders for various tasks, which complicates and burdens the coding pipelines. Our methodology advances the \"Coding for Machine\" paradigm by focusing on the encoder side. We present a versatile framework capable of accommodating both human and machine vision demands without the need to modify the pre-trained decoders of existing video codecs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.447, + 0.819 + ], + "angle": 0, + "content": "2.3. Compressed Video Analysis/Understanding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.901 + ], + "angle": 0, + "content": "There are also amounts of works performing video analysis tasks [17, 23, 39], such as image recognition, action recognition [31, 43] and multiple object tracking (MOT) [6, 16], in the compressed video domain. However, these methods focus on developing video analysis models that better lever" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.417, + 0.893, + 0.463 + ], + "angle": 0, + "content": "age the partially decoded video stream, such as the motion vector. In contrast, our work focuses on the coding procedure, specifically the encoding procedure." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.474, + 0.591, + 0.489 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.499, + 0.61, + 0.514 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.893, + 0.794 + ], + "angle": 0, + "content": "Our coding framework is illustrated in Fig. 2 (a). Let \\(\\mathcal{X} = \\{\\pmb{x}_1,\\dots ,\\pmb{x}_T\\}\\) represent the video sequence. The framework categorizes the \\(P\\) frames into two types: original \\(P\\) frames and machine vision-specific \\(P_{m}\\) frames. The categorization of frame \\(\\pmb{x}_t\\) is determined by the GoP Selection Network, where a value of 1 signifies a \\(P\\) frame and 0 a \\(P_{m}\\) frame. The \\(P_{m}\\) frames, designed to reduce the bitrate while preserving critical information for future vision tasks, are generated from the preceding \\(P\\) frames using Dynamic Vision Mode Prediction (DVMP). To maintain the quality of the decoded video sequence, each \\(P\\) frame is computed from the previous \\(P\\) frame with superior visual quality, rather than from a \\(P_{m}\\) frame. In scenarios requiring coding for downstream vision tasks, a machine-centric control applies the GoP selection and DVMP, altering the encoding structure to \"I, \\(P_{m},P,\\dots\\)\". Conversely, for human viewing, a human-centric control restores the encoding structure to its original form, \"I,P,P,P,,.." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.791, + 0.818 + ], + "angle": 0, + "content": "3.2. Dynamic Vision Mode Prediction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In our coding framework, we aim to minimize the video sequence bitrate. Initial analyses of residual video CODECs reveal that they allocate a substantial portion of the bitrate—exceeding \\(80\\%\\) in high bitrate configurations—to encode and transmit residual information. This approach en" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26038" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.452, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.252, + 0.47, + 0.279 + ], + "angle": 0, + "content": "Figure 3. Hyper-prior guided Dynamic Vision Mode Prediction network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.468, + 0.43 + ], + "angle": 0, + "content": "sures high-quality visuals for human perception, as measured by metrics such as PSNR or MS-SSIM [41]. However, this method introduces significant redundancy, especially since downstream tasks like video object detection or tracking predominantly concentrate on specific regions of interest rather than the entirety of the frame. Consequently, it is crucial for our study to develop an approach that reduces this redundancy in the coding bitstream." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.468, + 0.536 + ], + "angle": 0, + "content": "Inspired by Hu et al., who developed a mode prediction technique specifically designed for the human visual system within a slimmable encoder and decoder, we introduce the Dynamic Vision Mode Prediction (DVMP) module. This innovative module autonomously selects the optimal coding mode and decides whether each feature element should be encoded and transmitted, as illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.468, + 0.884 + ], + "angle": 0, + "content": "To elucidate the functionality of our proposed Dynamic Vision Mode Prediction (DVMP) module, we utilize the encoded residual feature \\( m_{t} \\) as an illustrative example. Assume \\( m_{t} \\) possesses dimensions \\( c \\times h \\times w \\), indicating \\( c \\) channels each with a spatial dimension of \\( h \\times w \\). The hyperprior network then predicts the mean and variance for each element, resulting in dimensions of \\( 2c \\times h \\times w \\). The mode prediction network's architecture, detailed in the upper branch, consists of two convolution layers and three ResBlocks. To render the \"mode prediction\" process differentiable, we employ the Gumbel Softmax technique during training to ascertain the skip mode for each encoded residual feature element. This module generates a binary mask for the feature map \\( m_{t} \\), where a \"1\" denotes elements imperative for machine vision that require entropy coding, and a \"0\" signifies elements either irrelevant for machine vision or accurately predictable by the hyperprior network, thus streamlining the entropy coding process and bitrate reduction. Note that the showed architecture is not suitable for prior models with autoregressive components, like which in DCVC [18]. However, we can extend the module into an autoregressive way to enable the support for DCVC. The detailed structure is shown in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To optimize the DVMP module, the loss function can be" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.891, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.218, + 0.892, + 0.301 + ], + "angle": 0, + "content": "Figure 4. Different coding GoP structures. The original structure consists of I and P frames. In the middle, the hand-crafted structure for machine vision consists of I, P and \\(P_{m}\\) frames which are arranged alternately. In the predicted (GoP selected) structure, the type of frames are predicted by the GoP selection network, targeting on better bitrate and machine vision performance trade-off." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.319, + 0.566, + 0.333 + ], + "angle": 0, + "content": "given by:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.337, + 0.892, + 0.352 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m} = R + \\lambda_ {m} \\mathcal {L} _ {t} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.892, + 0.407 + ], + "angle": 0, + "content": "where \\(R\\) denotes the coding bitrate and \\(\\mathcal{L}_t\\) denotes the loss of downstream vision task, respectively. \\(\\lambda_{m}\\) is a hyperparameter used to control the trade-off." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.419, + 0.73, + 0.433 + ], + "angle": 0, + "content": "3.3. DivGoP \\(\\rightarrow\\) GoP Selection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.892, + 0.594 + ], + "angle": 0, + "content": "Obtaining the \\(P_{m}\\) frames, we firstly use a hand-crafted structure in a GoP, where the \\(P_{m}\\) frames and \\(P\\) frames are arranged alternately (referred as \"DivGoP\" and shown in Fig.4). It is observed that this arrangement can bring a certain degree of bitrate saving in detection, since the \\(P_{m}\\) frames cost much less bitrate than \\(P\\) frames while maintaining the motion and machine vision information. Experimental findings, as depicted in the \"FVC DivGoP\" curve within Fig. 5, demonstrate that the \"I, \\(P_{m}, P, \\cdots\\)\" structure markedly achieve about \\(18.6\\%\\) bitrate saving." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.893, + 0.867 + ], + "angle": 0, + "content": "The \"DivGoP\" configuration, while methodically structured, may not be universally optimal due to the variability in motion and object information across different videos, which could necessitate distinct GoP structures. For instance, videos characterized by rapid target movement might demand a higher frequency of \\( P \\) frames to preserve the quality of reconstruction, as an excess of \\( P_{m} \\) frames may lead to degradation that negatively affects subsequent frames. Conversely, in scenarios where the camera remains static or the target movement is minimal, the introduction of more \\( P_{m} \\) frames could be feasible without significantly impacting the reconstruction quality of following frames. Stimulated by this notion, our research endeavors to tailor the GoP structure dynamically, enhancing codec performance for machine vision tasks. To approach this issue, we adopt FVC [14] as the foundational codec, select a GoP size of 10, and address the detection task utilizing the MOT Dataset. The optimization problem is thus defined:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.88, + 0.891, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\underset {\\theta} {\\arg \\min } R (\\theta) + \\lambda_ {p} L _ {\\det } (\\theta) \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26039" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "where \\( R \\) denotes the bitrate, \\( L_{det} \\) denotes the detection loss, \\( \\lambda_p \\) is a weighting factor that balances the effects of bitrate and detection loss, and \\( \\theta \\) denotes the setting of GoP structure here." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.471, + 0.473 + ], + "angle": 0, + "content": "Obtaining a GoP, each of the 9 predicted frames can be either \\( P \\) frame or \\( P_{m} \\) frame, which can be dynamically decided. To explore the optimal GoP structure for target function 2, and evaluate the performance gap between DivGoP and optimal GoP structure, we use a deep-first-searching (DFS) algorithm, where we use different \\( \\lambda \\) for different bitrate models. The DFS algorithm searches for the structure that makes the target function smallest for each GoP. Results are shown in the \"FVC DFS\" curve of Fig.5. The application of the DFS algorithm led to a substantial increase in Bpp-mAP performance,, achieving about \\( 32.3\\% \\) bitrate saving in terms of the BD-RATE metric, while the \"Div-GoP\" method achieves about \\( 18.6\\% \\) bitrate saving. Similar conclusions can also be observed on DCVC [18] and TCM [29], reinforcing our hypothesis that distinct videos necessitate unique GoP structures tailored for machine vision tasks. However, the DFS algorithm is extremely time consuming, and the labels do not exist during real inference process, making the DFS not available to be applied in real time encoding. How to further find an \"adaptive GoP structure\" in an available time complexity is worth exploring." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.807 + ], + "angle": 0, + "content": "From this perspective, we introduce our GoP selection method to adaptively predict the GoP structure at encoding stage, which has a low time complexity and is available in application. Illustrated in Fig.2 (c), our GoP selection network operates in two phases: the pre-analysis and the GoP prediction stages. Initially, input frames undergo preanalysis, where a detector identifies objects and generates an overlapping mask for each frame, while the RAFT [34] optical flow estimation method calculates optical flow features for each predicted frame using its adjacent predecessor. These features are then refined by masking with the object bounding box mask to produce masked optical flow features, alongside motion feature priors (mean and variance, averaged by channel) derived from the codec's motion encoder. Subsequently, these processed features are input into the GoP prediction network, which employs a feature extractor with convolution blocks for downsampling and aggregation, followed by adaptive average pooling and linear layers that output a normalized \\(s_{\\text{logit}}\\) for this GoP, which comprises two elements that sum to one. The averaged detection confidence information from the detector is also injected into the linear layers as meta data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Upon acquiring the \\( s_{\\mathrm{logit}} \\), we employ the Gumbel softmax sampling technique throughout the training phase to iteratively generate a binary value for each frame. This process culminates in the formation of a GoP structure vector, where the binary value—0 for a \\( P_{m} \\) frame and 1 for a \\( P \\) frame—specifies the type of each frame within the GoP." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.698, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.09, + 0.892, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.23, + 0.892, + 0.259 + ], + "angle": 0, + "content": "Figure 5. Left: Using DFS to search for the near-optimal GoP structure for Bpp-mAP. Right: Results of simply fine-tuning FVC" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.284, + 0.892, + 0.374 + ], + "angle": 0, + "content": "During inference stage, we apply a softmax operation on \\( s_{\\mathrm{logit}} \\) to derive probabilities, which are then used to evenly distribute 1s and 0s, thereby determining the actual encoded GoP structure. Similar to our mode prediction network, the loss function for optimizing GoP structure selection network is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.392, + 0.891, + 0.41 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g} = \\bar {R} + \\lambda_ {g} \\bar {\\mathcal {L}} _ {t} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.418, + 0.891, + 0.449 + ], + "angle": 0, + "content": "where the \\(\\bar{R}\\) and \\(\\bar{\\mathcal{L}}_t\\) denote the GoP-wise average coding bitrate and average loss of downstream vision task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.466, + 0.633, + 0.483 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.491, + 0.892, + 0.596 + ], + "angle": 0, + "content": "In this section, we describe the dataset settings, downstream task models, and evaluation metrics. We then provide empirical evidence underscoring the effectiveness of our proposed method across multiple tasks and CODECs. Lastly, we validate the contributions of the mode prediction and GoP selection modules to performance improvements through ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.598, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Datasets. For multi-object tracking (MOT) task, we choose MOT17 dataset. For video object detection (VOD) task, we adopt the Imagenet VID dataset [28] for training and evaluation. For video action recognition (VAR) task, we adopt UCF101[32] dataset. For the data processing procedure, we downsample the short side of the evaluation videos to 512 for MOT17 dataset and keep the original frame ratios for both training and evaluation. For the evaluation of VOD, we follow the most VOD works to resize the frames to \\(512 \\times 512\\) for training and \\(576 \\times 576\\) for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Downstream Task Models. For MOT task, we adipot the popular ByteTrack [46] method as the backbone model. The model weights are provided by their official framework. For VOD task, we adipot the detector YOLOv [30], and the model weights are also provided by its official framework. For the VAR task, we adipot the TSM [22] model as the recognition network, the weights are provided by mmaction2 [5]. During training and evaluation of our proposed framework, the weights of the downstream task models are fixed, without any fine-tuning." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26040" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.357, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.273, + 0.225, + 0.284 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.091, + 0.63, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.273, + 0.499, + 0.283 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.091, + 0.901, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.273, + 0.772, + 0.283 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.287, + 0.356, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.468, + 0.225, + 0.479 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.287, + 0.628, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.469, + 0.499, + 0.479 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.287, + 0.902, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.468, + 0.772, + 0.479 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.492, + 0.893, + 0.534 + ], + "angle": 0, + "content": "Figure 6. (a) Bpp-MOTA curves on MOT Dataset. (b) Bpp-mAP curves on MOT Dataset. (c) Bpp-MOTP curves on MOT Dataset. (d) Bpp-FN curves on MOT Dataset. (e) Bpp-mAP50 curves on ImageNet VID Dataset using YOLOv [30]. (f) Bpp-Top1 Accuracy curves on UCF101 Dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.704 + ], + "angle": 0, + "content": "Evaluation Metrics. For compression, We use bpp (bits per pixel) to measure the average number of bits for one pixel. For the MOT task, we adopt the MOTA (multiple object tracking accuracy), MOTP (multiple object tracking precision), FN (false negative detection number) and mAP (mean average precision) metrics to evaluate the tracking and detection performance. For the VOD task, we adopt the mAP50 metric to evaluate the detection performance, following most VOD methods. For the VAR task, we adopt the Top1 Accuracy as the performance indicator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. Our whole framework is implemented on PyTorch with CUDA support and trained on eight V100 or A100 GPUs. We use the Adam optimizer by setting the learning rate as 0.0001, \\(\\beta_{1}\\) as 0.9 and \\(\\beta_{2}\\) as 0.999, respectively. The whole training process has the following stages: Firstly, we use the loss function 1 to train the mode prediction network and obtain the \\(P_{m}\\) frames. This stage we use single frame training. Note that for TCM, DVMP is applied to both motion and contextual information. For DCVC and FVC, DVMP is exclusively utilized for motion information, and all residual/contextual information is omitted. After that, we introduce the GoP selection network in a GOP length unit. Using the GoP size of 32" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.893, + 0.705 + ], + "angle": 0, + "content": "on MOT and VAR tasks, we decide the GoP into 2 mini GoPs with size of 16, and the predicted vector of each mini GoP has a length of 15. For VOD task, we directly set the GoP size as 16. Using multi-frame training and loss function 3, we fix the weights of mode prediction network and train GoP selection network for about 100 epochs on MOT half-train dataset and 20 epochs on ImageNet VID training dataset, respectively for multi-object tracking and video object detection tasks. The whole training procedure takes about 2 days for each task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.714, + 0.729 + ], + "angle": 0, + "content": "4.1. Multi-Object Tracking" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In Fig. 6, we present a comparison between our method and various codecs, specifically DCVC, FVC, and TCM. It is evident that our approach yields a markedly improved tradeoff in terms of Bpp-MOTA, Bpp-mAP, Bpp-MOTP and Bpp-FN metrics. When compared to the baseline codecs FVC, DCVC and TCM, our control method results in bitrate savings of approximately \\(27.54\\%\\), \\(41.82\\%\\) and \\(12.64\\%\\), respectively, for equivalent MOTA metrics. With DCVC established as the anchor baseline, Table 1 displays the BD-RATE values, where our controlled TCM, controlled DCVC and controlled FVC codecs demonstrate the best, second-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26041" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.278, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.091, + 0.474, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.226, + 0.47, + 0.254 + ], + "angle": 0, + "content": "Figure 7. Video Reconstruction Quality in terms of Bpp-PSNR on HEVC Class B dataset (Left) and Class C dataset (Right)" + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.269, + 0.427, + 0.283 + ], + "angle": 0, + "content": "Table 1. BD-RATE results on multi-object tracking." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.294, + 0.468, + 0.403 + ], + "angle": 0, + "content": "
MethodMOTAmAPmAP50MOTPFNAverage
DCVC [18]0.00.00.00.00.00.0
Ours+DCVC-41.82-39.43-40.60-37.73-41.74-40.26
FVC [14]-5.32-9.75-14.53-1.39-6.20-7.44
Ours+FVC-34.28-31.27-32.09-35.02-32.89-33.11
TCM [29]-25.19-32.34-31.02-39.85-26.44-30.97
Ours+TCM-40.82-45.10-46.15-51.66-38.98-44.54
HEVC [33]-33.88-31.35-40.4315.80-34.02-24.78
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.468, + 0.489 + ], + "angle": 0, + "content": "best and third-best performance, respectively. Additionally, we contrast our method with the task-oriented \"One-to-one Codes\" of VCM, with detailed results provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.292, + 0.513 + ], + "angle": 0, + "content": "4.2. Video Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.468, + 0.703 + ], + "angle": 0, + "content": "For the evaluation of the video object detection task, we juxtapose our method with established codecs such as FVC, and TCM, with the comparative results illustrated in Fig. 6(e). Our method is noted to achieve a superior tradeoff, particularly in the Bpp-mAP50 metric, when applied to the downstream model YOLOv. For instance, in comparison to the FVC and TCM baseline codecs, our method facilitates bitrate reductions of approximately \\(21.76\\%\\) and \\(20.42\\%\\), respectively, while maintaining the same mAP50 metric. The uncompressed video represents the upper bound for the mAP50 metric, recorded at 76.4, and is delineated by the grey line within the figure." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.31, + 0.728 + ], + "angle": 0, + "content": "4.3. Video Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We conducted an evaluation of our method on the UCF101 dataset, benchmarking against TCM and FVC CODECs. The outcomes of this comparison are depicted in Fig. 6(f). For the task of video action recognition, rather than initiating a new set of network trainings, we employed pre-trained mode prediction networks previously utilized for the MOT task and use the \"DivGoP\" structure for validation. The original videos were compressed using the various CODECs, and the resulting decoded videos served as inputs to the TSM model as cited in Lin et al. [22]. The experimental results indicate that our method maintains efficacy in this" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.697, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.09, + 0.892, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.229, + 0.89, + 0.257 + ], + "angle": 0, + "content": "Figure 8. Left: Ablation study for Mode Prediction network. Right: Ablation study for GoP Selection network" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.282, + 0.89, + 0.313 + ], + "angle": 0, + "content": "context, securing approximately \\(21.81\\%\\) and \\(26.03\\%\\) in bitrate savings for TCM and FVC, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.325, + 0.765, + 0.342 + ], + "angle": 0, + "content": "4.4. Video Reconstruction Quality" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.5 + ], + "angle": 0, + "content": "When it comes to the performance of reconstructing the video for human viewing, the situation is straightforward for conventional video codecs, no matter learned or traditional ones: just encode the frames, transmit and decode the compressed bitstream, and measure their quality using appropriate metrics (PSNR, MS-SSIM, etc). In our method, there're two options: Control the encoding procedure for human viewing, which is the ideal way for best frame reconstruction quality, or just continue the encoding procedure used for machine vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.729 + ], + "angle": 0, + "content": "Note that our method does not change the weights of the original encoder and decoder in pre-trained DVC, so using this option can directly recover the human viewing quality to original best performance of the CODECs. On the other hand, we also evaluate the human viewing quality of the encoded bitstream for machine, using \"DivGop\" encoding structure on HEVC Class B and Class C test datasets. These are referred to as \"FVC DivGoP\" and \"TCM DivGoP\" in Fig.7. Results show that our control for machine does decrease the video reconstruction quality of the original TCM codec by about 1dB in PSNR. Besides, it is also observed that the PSNR of residual codec FVC drops more sharply than that of conditional CODEs like TCM. However, these are acceptable since we can always adjust the encoder to compress the video for human viewing when it's needed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.652, + 0.757 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Mode Prediction. In our method, we propose mode prediction network to decide the coding modes of motion and residual/contextual feature elements for machine vision, mainly targeting on reducing the bitrate and keep the crucial information for important objects in the frame. Moreover, mainstream DVC, including FVC, DCVC, TCM and others, use motion estimation and warping operation to generate a rough predicted frame, which we call \\( P_r \\) frame here. The \\( P_r \\) frame also meets our requirement of \"consuming" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26042" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.563, + 0.065, + 0.828, + 0.078 + ], + "angle": 0, + "content": "Table 2. Encoding Latency vs. Performance." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.09, + 0.464, + 0.226 + ], + "angle": 0, + "content": "
Input imageContextual skip mode maskMasked motion featureMasked Contextual feature
" + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.239, + 0.462, + 0.253 + ], + "angle": 0, + "content": "Figure 9. Visualization results of the mode prediction networks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.267, + 0.47, + 0.477 + ], + "angle": 0, + "content": "lower code streams but retaining object and motion information\" to a certain extent. So comparing it with the \\(P_{m}\\) frame used in our method is also a point worthy of attention. In \"DivGoP\" structure, we use the \\(P_{r}\\) frame to replace the original \\(P_{m}\\) frames, and the GoP structure will be \\([I, P_{r}, P, P_{r}, \\ldots]\\). As shown in the left figure in Fig.8, based on DCVC and FVC, we compare the structure using \\(P_{r}\\) frames with the original \"DivGoP\" structure using \\(P_{m}\\) frames, corresponding to the \"DCVC/FVC DivGoP w \\(P_{r}\\)\" and \"DCVC/FVC DivGoP\" curves respectively. It is observed that the \"DivGoP w \\(P_{r}\\)\" curve also achieves about \\(28.68\\%\\) and \\(16.24\\%\\) bitrate savings for DCVC and FVC, respectively. But our \"DivGoP\" curve shows better trade-off, by achieving \\(32.67\\%\\) and \\(19.43\\%\\) bitrate savings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.47, + 0.612 + ], + "angle": 0, + "content": "GoP Selection. We prove the effectiveness of Gop Selection module in terms of rate-precision performance on MOT dataset. As shown in the right figure in Fig. 8, we compare the Bpp-mAP performance of \"DCVC/FVC Gop Selection\" and \"DCVC/FVC DivGoP\" structure on MOT Dataset. Using DCVC as an example, experiment results show that our Gop Selection achieves about \\(39.43\\%\\) bitrate saving, which does show better trade-off than that \\(32.67\\%\\) bitrate saving of hand-crafted \"DivGoP\" structure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.468, + 0.824 + ], + "angle": 0, + "content": "Simply fine-tuning DVC for machine vision tasks. We employed the loss function \\( L_{f} = R + \\lambda_{1}MSE + \\lambda_{2}L_{det} \\) to fine-tune the FVC encoder and the entire FVC codec for the MOT task. As shown in Fig. 5 right, while simple finetuning improves accuracy, it also notably increases the bitrate. This phenomenon could be attributed to the degradation in the fidelity of decoded frames when optimized for downstream tasks. Due to the inter-frame reference nature of video codecs, this degradation is propagated across frames, resulting in a high bitrate consumption. This observation aligns with findings reported in DeepSVC [21], indicating that such task-specific performance can not be achieved with simply finetuning a DVC. This also reflects the superiority of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Visualization Results. We present the visualization results of our mode prediction network, as shown in Fig. 9. The first column shows the input images, the second column shows the skip mode masks for contextual features (averaged by channel), and the third and fourth columns are" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.091, + 0.885, + 0.176 + ], + "angle": 0, + "content": "
MethodEncoding Latency(s)BD-BR Performance
DCVC1.6750.0
Ours+DCVC1.738-40.26
FVC0.116-7.44
Ours+FVC0.159-33.11
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.188, + 0.892, + 0.338 + ], + "angle": 0, + "content": "masked motion and contextual features, respectively, where the red color indicates large value and green color indicates small value. It is observed that the predicted coding mode mask for contextual features kept the regions of the small objects in the frames, and masked the conspicuous objects which are easy to recognize and the background regions. This shows the network's preference for objects in the frame. As for the motion information, it is observed that the predicted masks kept most of the elements of moving objects, which are crucial for detection and tracking." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.892, + 0.536 + ], + "angle": 0, + "content": "Complexity and Encoding Latency Analysis. The parameter numbers of our dynamic vision mode prediction network and GoP selection network are 1.48M and 5.36M. Since our method controls the video encoder during the encoding phase, such an operation brings encoding latency. To measure the impact on the encoding phase, we include a comparison of average encoding latency versus performance on MOT Dataset \\((960 \\times 512)\\) in Tab. 2. While our method brings an increase in encoding latency (1.738s vs. 1.675s), the improvement in R-P performance is substantial. Besides, the decoding procedure and decoder side models are not changed, which brings convenience to the deployment of the decoder and support for multiple tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.555, + 0.618, + 0.57 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.837 + ], + "angle": 0, + "content": "In this paper, we propose a flexible framework for deep video compression models, where the pre-trained encoders can be controlled to change the encoding pipeline for machine vision tasks, achieving significant better bppprecision trade-off without changing the decoders or decoding procedure. The controlling methods are attributed to two main developments, including dynamic vision mode prediction network and GoP structure selection network. Experiment results show that our framework is general for residual and conditional deep video CODECs and different downstream vision tasks like detection, tracking and action recognition. Our comprehensive evaluation, conducted against existing deep video CODECs, confirms the superior performance of our models, establishing a new benchmark for future deep video coding for machine. The effectiveness of each component is clearly evidenced through extensive ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by the National Natural Science Fund of China (Project No. 42201461) and the General Research Fund (Project No. 16209622) from the Hong Kong Research Grants Council." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26043" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Mohammad Akbari, Jie Liang, and Jingning Han. Dsslic: Deep semantic segmentation-based layered image compression. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2042-2046. IEEE, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J. Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.471, + 0.313 + ], + "angle": 0, + "content": "[3] Zhuo Chen, Kui Fan, Shiqi Wang, Lingyu Duan, Weisi Lin, and Alex Chichung Kot. Toward intelligent sensing: Intermediate deep feature compression. IEEE Transactions on Image Processing, 29:2230-2243, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.314, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[4] Hyomin Choi and Ivan V Bajic. Scalable image coding for humans and machines. IEEE Transactions on Image Processing, 31:2739-2754, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.4 + ], + "angle": 0, + "content": "[5] MMAction2 Contributors. Openmmlab's next generation video understanding toolbox and benchmark. https://github.com/open-mmlab/mmaction2, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.47, + 0.444 + ], + "angle": 0, + "content": "[6] Yan Dai, Ziyu Hu, Shuqi Zhang, and Lianjun Liu. A survey of detection-based video multi-object tracking. Displays, page 102317, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.471, + 0.487 + ], + "angle": 0, + "content": "[7] Ling-Yu Duan, Tiejun Huang, and Wen Gao. Overview of the mpeg cdvs standard. In 2015 Data Compression Conference, pages 323-332, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.471, + 0.557 + ], + "angle": 0, + "content": "[8] Ling-Yu Duan, Yihang Lou, Yan Bai, Tiejun Huang, Wen Gao, Vijay Chandrasekhar, Jie Lin, Shiqi Wang, and Alex Chichung Kot. Compact descriptors for video analysis: The emerging mpg standard. IEEE MultiMedia, 26(2): 44-54, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.471, + 0.614 + ], + "angle": 0, + "content": "[9] Yann Dubois, Benjamin Bloem-Reddy, Karen Ullrich, and Chris J Maddison. Lossy compression for lossless prediction. Advances in Neural Information Processing Systems, 34:14014–14028, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.47, + 0.686 + ], + "angle": 0, + "content": "[10] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In European Conference on Computer Vision, pages 510-528. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.47, + 0.73 + ], + "angle": 0, + "content": "[11] Kristian Fischer, Fabian Brand, and André Kaup. Boosting neural image compression for machines using latent space masking. arXiv preprint arXiv:2112.08168, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[12] Hadi Hadizadeh and Ivan V Bajić. Learned scalable video coding for humans and machines. arXiv preprint arXiv:2307.08978, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[13] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[14] Zhihao Hu, Guo Lu, and Dong Xu. Fvc: A new framework towards deep video compression in feature space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1502-1511, 2021. 1, 2, 4, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[15] Zhihao Hu, Guo Lu, Jinyang Guo, Shan Liu, Wei Jiang, and Dong Xu. Coarse-to-fine deep video coding with hyperprior-guided mode prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5921-5930, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[16] Sayed Hossein Khatoonabadi and Ivan V Bajic. Video object tracking in the compressed domain using spatio-temporal markov random fields. IEEE transactions on image processing, 22(1):300-313, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.894, + 0.302 + ], + "angle": 0, + "content": "[17] Congcong Li, Xinyao Wang, Longyin Wen, Dxiang Hong, Tiejian Luo, and Libo Zhang. End-to-end compressed video representation learning for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13967-13976, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.893, + 0.347 + ], + "angle": 0, + "content": "[18] Jiahao Li, Bin Li, and Yan Lu. Deep contextual video compression. Advances in Neural Information Processing Systems, 34:18114-18125, 2021. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.893, + 0.403 + ], + "angle": 0, + "content": "[19] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22616-22626, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.893, + 0.446 + ], + "angle": 0, + "content": "[20] Xin Li, Jun Shi, and Zhibo Chen. Task-driven semantic coding via reinforcement learning. IEEE Transactions on Image Processing, 30:6307-6320, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.894, + 0.517 + ], + "angle": 0, + "content": "[21] Hongbin Lin, Bolin Chen, Zhichen Zhang, Jielian Lin, Xu Wang, and Tiesong Zhao. Deep scalable video coding for both machine and human vision. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9205-9214, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.894, + 0.574 + ], + "angle": 0, + "content": "[22] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7083-7093, 2019. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.893, + 0.617 + ], + "angle": 0, + "content": "[23] Qiankun Liu, Bin Liu, Yue Wu, Weihai Li, and Nenghai Yu. Real-time online multi-object tracking in compressed domain. arXiv preprint arXiv:2204.02081, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.893, + 0.687 + ], + "angle": 0, + "content": "[24] Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, and Zhiyong Gao. Dvc: An end-to-end deep video compression framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11006-11015, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.893, + 0.744 + ], + "angle": 0, + "content": "[25] Guo Lu, Xiaoyun Zhang, Wanli Ouyang, Li Chen, Zhiyong Gao, and Dong Xu. An end-to-end learning framework for video compression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3292-3308, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.893, + 0.788 + ], + "angle": 0, + "content": "[26] Guo Lu, Xingtong Ge, Tianxiong Zhong, Jing Geng, and Qiang Hu. Preprocessing enhanced image compression for machine vision. arXiv preprint arXiv:2206.05650, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.893, + 0.858 + ], + "angle": 0, + "content": "[27] Fabian Mentzer, Eirikur Agustsson, Johannes Balle, David Minnen, Nick Johnston, and George Toderici. Neural video compression using gans for detail synthesis and propagation. In European Conference on Computer Vision, pages 562-578. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26044" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.471, + 0.12 + ], + "angle": 0, + "content": "scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[29] Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. Temporal context mining for learned video compression. IEEE Transactions on Multimedia, 2022. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[30] Yuheng Shi, Naiyan Wang, and Xiaojie Guo. Yolov: making still image object detectors great at video object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2254-2262, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[31] Zheng Shou, Xudong Lin, Yannis Kalantidis, Laura Sevilla-Lara, Marcus Rohrbach, Shih-Fu Chang, and Zhicheng Yan. Dmc-net: Generating discriminative motion cues for fast compressed video action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1268-1277, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.47, + 0.344 + ], + "angle": 0, + "content": "[32] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.413 + ], + "angle": 0, + "content": "[33] Gary J. Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on Circuits and Systems for Video Technology, 22(12):1649-1668, 2012. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[34] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.471, + 0.552 + ], + "angle": 0, + "content": "[35] Yuan Tian, Guo Lu, Xiongkuo Min, Zhaohui Che, Guang-tao Zhai, Guodong Guo, and Zhiyong Gao. Self-conditioned probabilistic learning of video rescaling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4490-4499, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.47, + 0.621 + ], + "angle": 0, + "content": "[36] Yuan Tian, Guo Lu, Guangtao Zhai, and Zhiyong Gao. Nonsementics suppressed mask learning for unsupervised video semantic compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13610-13622, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.47, + 0.678 + ], + "angle": 0, + "content": "[37] Yuan Tian, Guo Lu, Yichao Yan, Guangtao Zhai, Li Chen, and Zhiyong Gao. A coding framework and benchmark towards low-bitrate video understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.47, + 0.734 + ], + "angle": 0, + "content": "[38] Robert Torfason, Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Towards image understanding from deep compression without decoding. arXiv preprint arXiv:1803.06131, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.471, + 0.79 + ], + "angle": 0, + "content": "[39] Shiyao Wang, Hongchao Lu, and Zhidong Deng. Fast object detection in compressed video. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7104-7113, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[40] Shurun Wang, Shiqi Wang, Wenhan Yang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. Towards analysis-friendly face representation with scalable feature and texture compression. IEEE Transactions on Multimedia, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[41] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402. IEEE, 2003. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[42] T. Wiegand, G.J. Sullivan, G. Bjontegaard, and A. Luthra. Overview of the h.264/avc video coding standard. IEEE Transactions on Circuits and Systems for Video Technology, 13(7):560-576, 2003. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[43] Chao-Yuan Wu, Manzil Zaheer, Hexiang Hu, R Manmatha, Alexander J Smola, and Philipp Krahenbuhl. Compressed video action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6026-6035, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[44] Ning Yan, Changsheng Gao, Dong Liu, Houqiang Li, Li Li, and Feng Wu. Sssic: semantics-to-signal scalable image coding with learned structural representations. IEEE Transactions on Image Processing, 30:8939-8954, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[45] Ren Yang, Luc Van Gool, and Radu Timofte. Perceptual learned video compression with recurrent conditional gan. arXiv preprint arXiv:2109.03082, 1, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[46] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In European Conference on Computer Vision, pages 1-21. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[47] Ezgi Özyilkan, Mateen Ulhaq, Hyomin Choi, and Fabien Racapé. Learned disentangled latent representations for scalable image coding for humans and machines. In 2023 Data Compression Conference (DCC), pages 42-51, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26045" + } + ] +] \ No newline at end of file diff --git a/2024/Task-Aware Encoder Control for Deep Video Compression/b978be40-5174-4abc-9649-d7daa2d37c89_origin.pdf b/2024/Task-Aware Encoder Control for Deep Video Compression/b978be40-5174-4abc-9649-d7daa2d37c89_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..eb27a7606f32f28aa9b1c5364624ac052160d3e8 --- /dev/null +++ b/2024/Task-Aware Encoder Control for Deep Video Compression/b978be40-5174-4abc-9649-d7daa2d37c89_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd5dbd439608a8dcca99e7ac5ad963bab20d025cb6c1971d0f9b402f1071920b +size 1570722 diff --git a/2024/Task-Aware Encoder Control for Deep Video Compression/full.md b/2024/Task-Aware Encoder Control for Deep Video Compression/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a7a7fc25d5508e934836aa0f8a1143227ec79984 --- /dev/null +++ b/2024/Task-Aware Encoder Control for Deep Video Compression/full.md @@ -0,0 +1,287 @@ +# Task-Aware Encoder Control for Deep Video Compression + +Xingtong Ge $^{1,2*}$ Jixiang Luo $^{2}$ Xinjie Zhang $^{3}$ Tongda Xu $^{4}$ Guo Lu $^{5}$ +Dailan He $^{6}$ Jing Geng $^{1\ddagger}$ Yan Wang $^{4}$ Jun Zhang $^{3}$ Hongwei Qin $^{2}$ $^{1}$ Beijing Institute of Technology $^{2}$ SenseTime Research + $^{3}$ The Hong Kong University of Science and Technology + $^{4}$ Institute for AI Industry Research (AIR), Tsinghua University + $^{5}$ Shanghai Jiaotong University $^{6}$ The Chinese University of Hong Kong + +# Abstract + +Prior research on deep video compression (DVC) for machine tasks typically necessitates training a unique codec for each specific task, mandating a dedicated decoder per task. In contrast, traditional video codecs employ a flexible encoder controller, enabling the adaptation of a single codec to different tasks through mechanisms like mode prediction. Drawing inspiration from this, we introduce an innovative encoder controller for deep video compression for machines. This controller features a mode prediction and a Group of Pictures (GoP) selection module. Our approach centralizes control at the encoding stage, allowing for adaptable encoder adjustments across different tasks, such as detection and tracking, while maintaining compatibility with a standard pre-trained DVC decoder. Empirical evidence demonstrates that our method is applicable across multiple tasks with various existing pre-trained DVCs. Moreover, extensive experiments demonstrate that our method outperforms previous DVC by about $25\%$ bitrate for different tasks, with only one pre-trained decoder. + +# 1. Introduction + +Over the past decades, video analysis techniques have proliferated across a variety of fields, including smart cities, autonomous vehicles, and traffic surveillance. For these applications, videos are often compressed before being transmitted to cloud-based systems for further machine vision analyses, such as object detection and tracking. However, current video compression techniques, which range from traditional codecs [2, 33, 42] to recent learning-based codecs [14, 18, 24, 25, 29], primarily cater to the human vi + +![](images/a105509d85e278cad2fb12c7a0de63003730f785de90480a600d7fe39344373c.jpg) +Figure 1. (a) Mainstream video codec that serves the human viewing. (b) Our controlled video codecs for machine vision with fixed decoder. (c) Other video codec for machine vision with one-to-one encoders and decoders. + +sual system, as shown in Fig. 1(a). This specificity leads to inefficiencies, as machine vision tasks typically focus on selective semantic details and regions within the video frames, rather than the whole frames. + +To tackle this problem, recent studies [1, 4, 11-13, 20, 38, 40, 44, 47] have explored scalable compression for multiple tasks through different layers in image and video coding. They propose a base layer dedicated to machine vision, with an enhancement layer containing additional information for human vision. Additionally, the latest work [36] has tried to compress semantic features at the encoder side, which serve as side information to supplement decoded images with more semantic details for machine vision at lower bitrates. Despite these advances, these methods require individually customized CODECs when applied to different downstream tasks, as the "One-to-one CODECs" shown in Fig. 1(c). In contrast, traditional CODECs can control the encoding process for different objectives, like PSNR and SSIM optimization, using a single decoder. Inspired by this feature, in this paper, we focus on how to use one pre-trained DVC-decoder to support both human and multiple machine vision tasks. + +In both residual [14, 15, 24] and conditional [18, 19, 29] DVC series, the encoded bitstreams predominantly comprise two parts: motion and residual/contextual informa + +tion. It is observed that a substantial portion of the bitrate, often exceeding $80\%$ in high bitrate models, is allocated to the residual/contextual information. This allocation is primarily for high-quality frame reconstruction. However, this approach is inappropriate for downstream vision tasks such as tracking and detection, which primarily concentrate on the objects and their movements rather than the full frames. + +To mitigate this issue, we introduce a dynamic vision mode prediction (DVMP) module specifically tailored for machine vision tasks. This module optimizes the entropy coding process by dynamically predicting the skip/no-skip coding mode for each feature element, based on its relevance to machine vision. Specifically, it takes hyperprior data from motion or residual/contextual features within the DVC framework as the input and assesses the utility of each feature element for machine vision. By replacing non-essential elements with their predicted mean value derived from the hyperprior network, our method effectively reduces bit consumption and circumvent the entropy decoding step for these elements to expedite the decoding process. In this case, we generate a novel frame type that significantly reduces bitrate usage compared to traditional P frames in the Group of Pictures (GoP) structure, while maintaining the information valuable for machine vision tasks. This new frame is designated as the $P_{m}$ frame. + +At the same time, it is important for DVC to maintain reconstruction quality because there exists significant reference relationship between frames, which is a crucial difference between image and video compression. Since our $P_{m}$ frame degrades the fidelity of reconstruction compared to the $P$ frame, we explore the reorganization of the Group of Pictures (GoP) structure. To be specific, we firstly explore to use a hand-crafted GoP structure, which already achieves significant rate-precision improvements. In the revised structure, both $P$ frames and $P_{m}$ frames derive their prediction from the preceding $I / P$ frame. Further, we introduce a GoP selection network, which can dynamically determine the GoP structure during the encoding process and achieves a better rate-precision trade-off. + +We choose three representative residual and conditional DVC methods as touchstones in our proposed video coding for machine vision framework. It's worth mentioning that our method utilizes pre-trained DVCs without altering the weights of the original encoders and decoders. Our control happens at encoding stage, leaving the decoders' architecture intact. Moreover, when human viewing is required, we can switch back to the original encoding procedure to restore reconstruction performance. In essence, using the proposed method, we can control the encoder of a DVC to adapt for both machine and human vision requirements. + +The main contributions of our work are summarised as follows: (1) Built upon the mainstream DVC CODECs, we propose a novel video coding for machine vision framework + +that controls the encoder to adapt for different downstream vision tasks, such as video object detection and multi-object tracking. (2) We employ a dynamic vision mode prediction approach to refine the original P frame, effectively reducing the bitrate while preserving critical information pertinent to vision tasks. Furthermore, we utilize a GoP selection strategy to dynamically forecast the coding GoP structure during the encoding stage, which controls the bitstream for better rate-precision trade-off in downstream vision tasks. (3) Experiments show that our controlling method is novel and flexible for different DVC CODECs, achieving up to more than $25\%$ bitrate savings in various downstream tasks. + +# 2. Related Works + +# 2.1. Video Compression + +Previous video codecs are designed to remove spatial-temporal redundancies effectively. Traditional video coding standards, including H.264/AVC [42], H.265/HEVC [33], and H.266/VVC [2], significantly increase the compression efficiency of images and videos. Recently, a number of learning-based video codecs [14, 18, 24, 25, 29, 35] have been proposed, designed based on residual coding or conditional coding, achieved better and better pixel-wise signal quality metrics, e.g., PSNR and MS-SSIM [41], which mainly serve the human visual experience. Recently, there are also some generative video coding methods [27, 45] that mainly consider visual comfort and perceptual quality. + +# 2.2. Compression for Machine Vision + +To facilitate the efficiency of machine vision tasks, early research works were devoted to extracting visual features from signals. Early standards, such as CDVA [8] and CDVS [7], suggest the pre-extraction and transportation of image keypoints to facilitate image indexing or retrieval tasks. With the development of deep learning, some studies began to explore connections between compression and downstream tasks, some studies [1, 3, 4, 11, 13, 20, 38, 40, 44, 47] also focus on the joint optimization of image compression and downstream machine vision tasks by introducing a rate-distortion optimization strategy guided by downstream tasks or by adding a task-specific feature encoding stream. For instance, Torfason et al. [38] introduced a method for executing image understanding tasks, such as classification and segmentation, on compressed outputs from learning-based image compression techniques. Furthermore, Lu et al. [26] enhanced traditional CODECs with a preprocessing step, thereby improving codec performance for downstream vision tasks. The field has also witnessed the emergence of self-supervised representation learning methods aimed at deriving compact semantic representations. Dubois et al. [9] presented a theoretical framework suggesting that the distortion term in the lossy rate + +![](images/44edc55052840a0107689515e6c08b00e26a7d95015b77a3dae11631b6ec9bd2.jpg) +(a) + +![](images/69851fbeec1e47931de7146feb70572da047965607c43313fbdc543c62f0f00e.jpg) + +![](images/e855742e4276dd2605f2f1d2e1cde33146445af297fd7cc6a43ee77e846518a5.jpg) +(b) +(c) + +![](images/868e6e6aa89606bc49050a15b87cbea5a532facb32768c0fcad6e441f1e409ab.jpg) +GoP Feature Extractor +Figure 2. (a) Overview of our "Controlling DVC for Machine" framework. Given an input GoP, we firstly use GoP Selection network to predict the GoP structure, then the predicted structure controls the encoding procedure to encode input frames for machine vision tasks. (b) The "0" element controls encoder to use DVMP. (c) The GoP Selection network, including the pre-analysis stage and GoP prediction stage. + +distortion trade-off for image classification could be approximated by a contrastive learning objective. Feng et al. [10] proposed a method to learn a unified feature representation for AI tasks from unlabeled data. These methodologies necessitate fine-tuning the downstream models to adapt to the learned features. + +Despite advancements in image coding for machine applications, reconstructing high-fidelity video from extracted features remains a formidable challenge due to the critical inter-frame reference relationships. Addressing the requirements of both machine and human vision, Tian et al. [36] introduced a self-supervised edge representation as a semantic intermediary, which preserves the video's semantic structure. Furthermore, Lin et al. [21] developed a scalable video coding framework tailored for machines, segregating semantic features for machine analysis and human viewing into separate bitstreams. + +Most existing approaches [21, 36, 37] necessitate the use of distinct encoders and decoders for various tasks, which complicates and burdens the coding pipelines. Our methodology advances the "Coding for Machine" paradigm by focusing on the encoder side. We present a versatile framework capable of accommodating both human and machine vision demands without the need to modify the pre-trained decoders of existing video codecs. + +# 2.3. Compressed Video Analysis/Understanding + +There are also amounts of works performing video analysis tasks [17, 23, 39], such as image recognition, action recognition [31, 43] and multiple object tracking (MOT) [6, 16], in the compressed video domain. However, these methods focus on developing video analysis models that better lever + +age the partially decoded video stream, such as the motion vector. In contrast, our work focuses on the coding procedure, specifically the encoding procedure. + +# 3. Method + +# 3.1. Overview + +Our coding framework is illustrated in Fig. 2 (a). Let $\mathcal{X} = \{\pmb{x}_1,\dots ,\pmb{x}_T\}$ represent the video sequence. The framework categorizes the $P$ frames into two types: original $P$ frames and machine vision-specific $P_{m}$ frames. The categorization of frame $\pmb{x}_t$ is determined by the GoP Selection Network, where a value of 1 signifies a $P$ frame and 0 a $P_{m}$ frame. The $P_{m}$ frames, designed to reduce the bitrate while preserving critical information for future vision tasks, are generated from the preceding $P$ frames using Dynamic Vision Mode Prediction (DVMP). To maintain the quality of the decoded video sequence, each $P$ frame is computed from the previous $P$ frame with superior visual quality, rather than from a $P_{m}$ frame. In scenarios requiring coding for downstream vision tasks, a machine-centric control applies the GoP selection and DVMP, altering the encoding structure to "I, $P_{m},P,\dots$ ". Conversely, for human viewing, a human-centric control restores the encoding structure to its original form, "I,P,P,P,,.. + +# 3.2. Dynamic Vision Mode Prediction + +In our coding framework, we aim to minimize the video sequence bitrate. Initial analyses of residual video CODECs reveal that they allocate a substantial portion of the bitrate—exceeding $80\%$ in high bitrate configurations—to encode and transmit residual information. This approach en + +![](images/57c3885ba644d58e3c8ff878e11daa3fc9de30a0238beeea7ef83aed31901a54.jpg) +Figure 3. Hyper-prior guided Dynamic Vision Mode Prediction network. + +sures high-quality visuals for human perception, as measured by metrics such as PSNR or MS-SSIM [41]. However, this method introduces significant redundancy, especially since downstream tasks like video object detection or tracking predominantly concentrate on specific regions of interest rather than the entirety of the frame. Consequently, it is crucial for our study to develop an approach that reduces this redundancy in the coding bitstream. + +Inspired by Hu et al., who developed a mode prediction technique specifically designed for the human visual system within a slimmable encoder and decoder, we introduce the Dynamic Vision Mode Prediction (DVMP) module. This innovative module autonomously selects the optimal coding mode and decides whether each feature element should be encoded and transmitted, as illustrated in Fig. 3. + +To elucidate the functionality of our proposed Dynamic Vision Mode Prediction (DVMP) module, we utilize the encoded residual feature $m_{t}$ as an illustrative example. Assume $m_{t}$ possesses dimensions $c \times h \times w$ , indicating $c$ channels each with a spatial dimension of $h \times w$ . The hyperprior network then predicts the mean and variance for each element, resulting in dimensions of $2c \times h \times w$ . The mode prediction network's architecture, detailed in the upper branch, consists of two convolution layers and three ResBlocks. To render the "mode prediction" process differentiable, we employ the Gumbel Softmax technique during training to ascertain the skip mode for each encoded residual feature element. This module generates a binary mask for the feature map $m_{t}$ , where a "1" denotes elements imperative for machine vision that require entropy coding, and a "0" signifies elements either irrelevant for machine vision or accurately predictable by the hyperprior network, thus streamlining the entropy coding process and bitrate reduction. Note that the showed architecture is not suitable for prior models with autoregressive components, like which in DCVC [18]. However, we can extend the module into an autoregressive way to enable the support for DCVC. The detailed structure is shown in the supplementary material. + +To optimize the DVMP module, the loss function can be + +![](images/87de36ec866ecd879792b0c7bc1ff9fce01cf868895a9e0624f3c005a1254f0d.jpg) +Figure 4. Different coding GoP structures. The original structure consists of I and P frames. In the middle, the hand-crafted structure for machine vision consists of I, P and $P_{m}$ frames which are arranged alternately. In the predicted (GoP selected) structure, the type of frames are predicted by the GoP selection network, targeting on better bitrate and machine vision performance trade-off. + +given by: + +$$ +\mathcal {L} _ {m} = R + \lambda_ {m} \mathcal {L} _ {t} \tag {1} +$$ + +where $R$ denotes the coding bitrate and $\mathcal{L}_t$ denotes the loss of downstream vision task, respectively. $\lambda_{m}$ is a hyperparameter used to control the trade-off. + +# 3.3. DivGoP $\rightarrow$ GoP Selection + +Obtaining the $P_{m}$ frames, we firstly use a hand-crafted structure in a GoP, where the $P_{m}$ frames and $P$ frames are arranged alternately (referred as "DivGoP" and shown in Fig.4). It is observed that this arrangement can bring a certain degree of bitrate saving in detection, since the $P_{m}$ frames cost much less bitrate than $P$ frames while maintaining the motion and machine vision information. Experimental findings, as depicted in the "FVC DivGoP" curve within Fig. 5, demonstrate that the "I, $P_{m}, P, \cdots$ " structure markedly achieve about $18.6\%$ bitrate saving. + +The "DivGoP" configuration, while methodically structured, may not be universally optimal due to the variability in motion and object information across different videos, which could necessitate distinct GoP structures. For instance, videos characterized by rapid target movement might demand a higher frequency of $P$ frames to preserve the quality of reconstruction, as an excess of $P_{m}$ frames may lead to degradation that negatively affects subsequent frames. Conversely, in scenarios where the camera remains static or the target movement is minimal, the introduction of more $P_{m}$ frames could be feasible without significantly impacting the reconstruction quality of following frames. Stimulated by this notion, our research endeavors to tailor the GoP structure dynamically, enhancing codec performance for machine vision tasks. To approach this issue, we adopt FVC [14] as the foundational codec, select a GoP size of 10, and address the detection task utilizing the MOT Dataset. The optimization problem is thus defined: + +$$ +\underset {\theta} {\arg \min } R (\theta) + \lambda_ {p} L _ {\det } (\theta) \tag {2} +$$ + +where $R$ denotes the bitrate, $L_{det}$ denotes the detection loss, $\lambda_p$ is a weighting factor that balances the effects of bitrate and detection loss, and $\theta$ denotes the setting of GoP structure here. + +Obtaining a GoP, each of the 9 predicted frames can be either $P$ frame or $P_{m}$ frame, which can be dynamically decided. To explore the optimal GoP structure for target function 2, and evaluate the performance gap between DivGoP and optimal GoP structure, we use a deep-first-searching (DFS) algorithm, where we use different $\lambda$ for different bitrate models. The DFS algorithm searches for the structure that makes the target function smallest for each GoP. Results are shown in the "FVC DFS" curve of Fig.5. The application of the DFS algorithm led to a substantial increase in Bpp-mAP performance,, achieving about $32.3\%$ bitrate saving in terms of the BD-RATE metric, while the "Div-GoP" method achieves about $18.6\%$ bitrate saving. Similar conclusions can also be observed on DCVC [18] and TCM [29], reinforcing our hypothesis that distinct videos necessitate unique GoP structures tailored for machine vision tasks. However, the DFS algorithm is extremely time consuming, and the labels do not exist during real inference process, making the DFS not available to be applied in real time encoding. How to further find an "adaptive GoP structure" in an available time complexity is worth exploring. + +From this perspective, we introduce our GoP selection method to adaptively predict the GoP structure at encoding stage, which has a low time complexity and is available in application. Illustrated in Fig.2 (c), our GoP selection network operates in two phases: the pre-analysis and the GoP prediction stages. Initially, input frames undergo preanalysis, where a detector identifies objects and generates an overlapping mask for each frame, while the RAFT [34] optical flow estimation method calculates optical flow features for each predicted frame using its adjacent predecessor. These features are then refined by masking with the object bounding box mask to produce masked optical flow features, alongside motion feature priors (mean and variance, averaged by channel) derived from the codec's motion encoder. Subsequently, these processed features are input into the GoP prediction network, which employs a feature extractor with convolution blocks for downsampling and aggregation, followed by adaptive average pooling and linear layers that output a normalized $s_{\text{logit}}$ for this GoP, which comprises two elements that sum to one. The averaged detection confidence information from the detector is also injected into the linear layers as meta data. + +Upon acquiring the $s_{\mathrm{logit}}$ , we employ the Gumbel softmax sampling technique throughout the training phase to iteratively generate a binary value for each frame. This process culminates in the formation of a GoP structure vector, where the binary value—0 for a $P_{m}$ frame and 1 for a $P$ frame—specifies the type of each frame within the GoP. + +![](images/2ed356220b088e2248ecfe355c84f5531407d576b4f81e64620881d5621c41dc.jpg) +Figure 5. Left: Using DFS to search for the near-optimal GoP structure for Bpp-mAP. Right: Results of simply fine-tuning FVC + +![](images/a94aa44b7630bcc5bf4f5233448509af4db142350cd6c437bbc3c956deef741d.jpg) + +During inference stage, we apply a softmax operation on $s_{\mathrm{logit}}$ to derive probabilities, which are then used to evenly distribute 1s and 0s, thereby determining the actual encoded GoP structure. Similar to our mode prediction network, the loss function for optimizing GoP structure selection network is given by: + +$$ +\mathcal {L} _ {g} = \bar {R} + \lambda_ {g} \bar {\mathcal {L}} _ {t} \tag {3} +$$ + +where the $\bar{R}$ and $\bar{\mathcal{L}}_t$ denote the GoP-wise average coding bitrate and average loss of downstream vision task. + +# 4. Experiments + +In this section, we describe the dataset settings, downstream task models, and evaluation metrics. We then provide empirical evidence underscoring the effectiveness of our proposed method across multiple tasks and CODECs. Lastly, we validate the contributions of the mode prediction and GoP selection modules to performance improvements through ablation studies. + +Datasets. For multi-object tracking (MOT) task, we choose MOT17 dataset. For video object detection (VOD) task, we adopt the Imagenet VID dataset [28] for training and evaluation. For video action recognition (VAR) task, we adopt UCF101[32] dataset. For the data processing procedure, we downsample the short side of the evaluation videos to 512 for MOT17 dataset and keep the original frame ratios for both training and evaluation. For the evaluation of VOD, we follow the most VOD works to resize the frames to $512 \times 512$ for training and $576 \times 576$ for evaluation. + +Downstream Task Models. For MOT task, we adipot the popular ByteTrack [46] method as the backbone model. The model weights are provided by their official framework. For VOD task, we adipot the detector YOLOv [30], and the model weights are also provided by its official framework. For the VAR task, we adipot the TSM [22] model as the recognition network, the weights are provided by mmaction2 [5]. During training and evaluation of our proposed framework, the weights of the downstream task models are fixed, without any fine-tuning. + +![](images/5f27cbeccdaea6c972b064752557970cd151ab1783920f8dbdf7fe833161431f.jpg) +(a) + +![](images/6339a848c9c0d0b8aafaaa4ceee93a618d0f788d98965b1f5946a1e864d01bed.jpg) +(b) + +![](images/393fcbc5c3537694ff4da72484cd5e1f43ccde4a5bf417156831021d1558e4ca.jpg) +(c) + +![](images/7f1c9c2addf59ee8646531406ddd2c9ed7dc895a7ae754517308db8b67a0173a.jpg) +(d) + +![](images/85c59007fe6b17fa1a765799494cd46b181e3955d57534432f7127b717baa2c1.jpg) +(e) + +![](images/ff2b4c14824dae2fe4f19ec9fbc47f428c02035dbb41de37c665f0fc56796b9a.jpg) +(f) +Figure 6. (a) Bpp-MOTA curves on MOT Dataset. (b) Bpp-mAP curves on MOT Dataset. (c) Bpp-MOTP curves on MOT Dataset. (d) Bpp-FN curves on MOT Dataset. (e) Bpp-mAP50 curves on ImageNet VID Dataset using YOLOv [30]. (f) Bpp-Top1 Accuracy curves on UCF101 Dataset. + +Evaluation Metrics. For compression, We use bpp (bits per pixel) to measure the average number of bits for one pixel. For the MOT task, we adopt the MOTA (multiple object tracking accuracy), MOTP (multiple object tracking precision), FN (false negative detection number) and mAP (mean average precision) metrics to evaluate the tracking and detection performance. For the VOD task, we adopt the mAP50 metric to evaluate the detection performance, following most VOD methods. For the VAR task, we adopt the Top1 Accuracy as the performance indicator. + +Implementation Details. Our whole framework is implemented on PyTorch with CUDA support and trained on eight V100 or A100 GPUs. We use the Adam optimizer by setting the learning rate as 0.0001, $\beta_{1}$ as 0.9 and $\beta_{2}$ as 0.999, respectively. The whole training process has the following stages: Firstly, we use the loss function 1 to train the mode prediction network and obtain the $P_{m}$ frames. This stage we use single frame training. Note that for TCM, DVMP is applied to both motion and contextual information. For DCVC and FVC, DVMP is exclusively utilized for motion information, and all residual/contextual information is omitted. After that, we introduce the GoP selection network in a GOP length unit. Using the GoP size of 32 + +on MOT and VAR tasks, we decide the GoP into 2 mini GoPs with size of 16, and the predicted vector of each mini GoP has a length of 15. For VOD task, we directly set the GoP size as 16. Using multi-frame training and loss function 3, we fix the weights of mode prediction network and train GoP selection network for about 100 epochs on MOT half-train dataset and 20 epochs on ImageNet VID training dataset, respectively for multi-object tracking and video object detection tasks. The whole training procedure takes about 2 days for each task. + +# 4.1. Multi-Object Tracking + +In Fig. 6, we present a comparison between our method and various codecs, specifically DCVC, FVC, and TCM. It is evident that our approach yields a markedly improved tradeoff in terms of Bpp-MOTA, Bpp-mAP, Bpp-MOTP and Bpp-FN metrics. When compared to the baseline codecs FVC, DCVC and TCM, our control method results in bitrate savings of approximately $27.54\%$ , $41.82\%$ and $12.64\%$ , respectively, for equivalent MOTA metrics. With DCVC established as the anchor baseline, Table 1 displays the BD-RATE values, where our controlled TCM, controlled DCVC and controlled FVC codecs demonstrate the best, second- + +![](images/404416386c280a5d2d2b2418a7ce691eee76d85e1afed2fb87244770c7914eac.jpg) +Figure 7. Video Reconstruction Quality in terms of Bpp-PSNR on HEVC Class B dataset (Left) and Class C dataset (Right) + +![](images/39c4a621b853dc3fa1fc774816057b2c4c8788d14f1102c083c4ac1cedd67c96.jpg) +Figure 8. Left: Ablation study for Mode Prediction network. Right: Ablation study for GoP Selection network + +Table 1. BD-RATE results on multi-object tracking. + +
MethodMOTAmAPmAP50MOTPFNAverage
DCVC [18]0.00.00.00.00.00.0
Ours+DCVC-41.82-39.43-40.60-37.73-41.74-40.26
FVC [14]-5.32-9.75-14.53-1.39-6.20-7.44
Ours+FVC-34.28-31.27-32.09-35.02-32.89-33.11
TCM [29]-25.19-32.34-31.02-39.85-26.44-30.97
Ours+TCM-40.82-45.10-46.15-51.66-38.98-44.54
HEVC [33]-33.88-31.35-40.4315.80-34.02-24.78
+ +best and third-best performance, respectively. Additionally, we contrast our method with the task-oriented "One-to-one Codes" of VCM, with detailed results provided in the supplementary materials. + +# 4.2. Video Object Detection + +For the evaluation of the video object detection task, we juxtapose our method with established codecs such as FVC, and TCM, with the comparative results illustrated in Fig. 6(e). Our method is noted to achieve a superior tradeoff, particularly in the Bpp-mAP50 metric, when applied to the downstream model YOLOv. For instance, in comparison to the FVC and TCM baseline codecs, our method facilitates bitrate reductions of approximately $21.76\%$ and $20.42\%$ , respectively, while maintaining the same mAP50 metric. The uncompressed video represents the upper bound for the mAP50 metric, recorded at 76.4, and is delineated by the grey line within the figure. + +# 4.3. Video Action Recognition + +We conducted an evaluation of our method on the UCF101 dataset, benchmarking against TCM and FVC CODECs. The outcomes of this comparison are depicted in Fig. 6(f). For the task of video action recognition, rather than initiating a new set of network trainings, we employed pre-trained mode prediction networks previously utilized for the MOT task and use the "DivGoP" structure for validation. The original videos were compressed using the various CODECs, and the resulting decoded videos served as inputs to the TSM model as cited in Lin et al. [22]. The experimental results indicate that our method maintains efficacy in this + +![](images/75f089c755a2ddff914016c4f10bb666a8d8e7096d9a74361da105160f9f9cf1.jpg) + +![](images/cb9e6c73e1c8d49c94adc78b78b8e57d02f69306ef795f1380b24c24853aefc4.jpg) + +context, securing approximately $21.81\%$ and $26.03\%$ in bitrate savings for TCM and FVC, respectively. + +# 4.4. Video Reconstruction Quality + +When it comes to the performance of reconstructing the video for human viewing, the situation is straightforward for conventional video codecs, no matter learned or traditional ones: just encode the frames, transmit and decode the compressed bitstream, and measure their quality using appropriate metrics (PSNR, MS-SSIM, etc). In our method, there're two options: Control the encoding procedure for human viewing, which is the ideal way for best frame reconstruction quality, or just continue the encoding procedure used for machine vision tasks. + +Note that our method does not change the weights of the original encoder and decoder in pre-trained DVC, so using this option can directly recover the human viewing quality to original best performance of the CODECs. On the other hand, we also evaluate the human viewing quality of the encoded bitstream for machine, using "DivGop" encoding structure on HEVC Class B and Class C test datasets. These are referred to as "FVC DivGoP" and "TCM DivGoP" in Fig.7. Results show that our control for machine does decrease the video reconstruction quality of the original TCM codec by about 1dB in PSNR. Besides, it is also observed that the PSNR of residual codec FVC drops more sharply than that of conditional CODEs like TCM. However, these are acceptable since we can always adjust the encoder to compress the video for human viewing when it's needed. + +# 4.5. Ablation Study + +Mode Prediction. In our method, we propose mode prediction network to decide the coding modes of motion and residual/contextual feature elements for machine vision, mainly targeting on reducing the bitrate and keep the crucial information for important objects in the frame. Moreover, mainstream DVC, including FVC, DCVC, TCM and others, use motion estimation and warping operation to generate a rough predicted frame, which we call $P_r$ frame here. The $P_r$ frame also meets our requirement of "consuming + +
Input imageContextual skip mode maskMasked motion featureMasked Contextual feature
+ +Figure 9. Visualization results of the mode prediction networks. + +lower code streams but retaining object and motion information" to a certain extent. So comparing it with the $P_{m}$ frame used in our method is also a point worthy of attention. In "DivGoP" structure, we use the $P_{r}$ frame to replace the original $P_{m}$ frames, and the GoP structure will be $[I, P_{r}, P, P_{r}, \ldots]$ . As shown in the left figure in Fig.8, based on DCVC and FVC, we compare the structure using $P_{r}$ frames with the original "DivGoP" structure using $P_{m}$ frames, corresponding to the "DCVC/FVC DivGoP w $P_{r}$ " and "DCVC/FVC DivGoP" curves respectively. It is observed that the "DivGoP w $P_{r}$ " curve also achieves about $28.68\%$ and $16.24\%$ bitrate savings for DCVC and FVC, respectively. But our "DivGoP" curve shows better trade-off, by achieving $32.67\%$ and $19.43\%$ bitrate savings. + +GoP Selection. We prove the effectiveness of Gop Selection module in terms of rate-precision performance on MOT dataset. As shown in the right figure in Fig. 8, we compare the Bpp-mAP performance of "DCVC/FVC Gop Selection" and "DCVC/FVC DivGoP" structure on MOT Dataset. Using DCVC as an example, experiment results show that our Gop Selection achieves about $39.43\%$ bitrate saving, which does show better trade-off than that $32.67\%$ bitrate saving of hand-crafted "DivGoP" structure. + +Simply fine-tuning DVC for machine vision tasks. We employed the loss function $L_{f} = R + \lambda_{1}MSE + \lambda_{2}L_{det}$ to fine-tune the FVC encoder and the entire FVC codec for the MOT task. As shown in Fig. 5 right, while simple finetuning improves accuracy, it also notably increases the bitrate. This phenomenon could be attributed to the degradation in the fidelity of decoded frames when optimized for downstream tasks. Due to the inter-frame reference nature of video codecs, this degradation is propagated across frames, resulting in a high bitrate consumption. This observation aligns with findings reported in DeepSVC [21], indicating that such task-specific performance can not be achieved with simply finetuning a DVC. This also reflects the superiority of our method. + +Visualization Results. We present the visualization results of our mode prediction network, as shown in Fig. 9. The first column shows the input images, the second column shows the skip mode masks for contextual features (averaged by channel), and the third and fourth columns are + +Table 2. Encoding Latency vs. Performance. + +
MethodEncoding Latency(s)BD-BR Performance
DCVC1.6750.0
Ours+DCVC1.738-40.26
FVC0.116-7.44
Ours+FVC0.159-33.11
+ +masked motion and contextual features, respectively, where the red color indicates large value and green color indicates small value. It is observed that the predicted coding mode mask for contextual features kept the regions of the small objects in the frames, and masked the conspicuous objects which are easy to recognize and the background regions. This shows the network's preference for objects in the frame. As for the motion information, it is observed that the predicted masks kept most of the elements of moving objects, which are crucial for detection and tracking. + +Complexity and Encoding Latency Analysis. The parameter numbers of our dynamic vision mode prediction network and GoP selection network are 1.48M and 5.36M. Since our method controls the video encoder during the encoding phase, such an operation brings encoding latency. To measure the impact on the encoding phase, we include a comparison of average encoding latency versus performance on MOT Dataset $(960 \times 512)$ in Tab. 2. While our method brings an increase in encoding latency (1.738s vs. 1.675s), the improvement in R-P performance is substantial. Besides, the decoding procedure and decoder side models are not changed, which brings convenience to the deployment of the decoder and support for multiple tasks. + +# 5. Conclusion + +In this paper, we propose a flexible framework for deep video compression models, where the pre-trained encoders can be controlled to change the encoding pipeline for machine vision tasks, achieving significant better bppprecision trade-off without changing the decoders or decoding procedure. The controlling methods are attributed to two main developments, including dynamic vision mode prediction network and GoP structure selection network. Experiment results show that our framework is general for residual and conditional deep video CODECs and different downstream vision tasks like detection, tracking and action recognition. Our comprehensive evaluation, conducted against existing deep video CODECs, confirms the superior performance of our models, establishing a new benchmark for future deep video coding for machine. The effectiveness of each component is clearly evidenced through extensive ablation studies. + +Acknowledgments. This work was supported by the National Natural Science Fund of China (Project No. 42201461) and the General Research Fund (Project No. 16209622) from the Hong Kong Research Grants Council. + +# References + +[1] Mohammad Akbari, Jie Liang, and Jingning Han. Dsslic: Deep semantic segmentation-based layered image compression. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2042-2046. IEEE, 2019. 1, 2 +[2] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J. Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 1, 2 +[3] Zhuo Chen, Kui Fan, Shiqi Wang, Lingyu Duan, Weisi Lin, and Alex Chichung Kot. Toward intelligent sensing: Intermediate deep feature compression. IEEE Transactions on Image Processing, 29:2230-2243, 2020. 2 +[4] Hyomin Choi and Ivan V Bajic. Scalable image coding for humans and machines. IEEE Transactions on Image Processing, 31:2739-2754, 2022. 1, 2 +[5] MMAction2 Contributors. Openmmlab's next generation video understanding toolbox and benchmark. https://github.com/open-mmlab/mmaction2, 2020.5 +[6] Yan Dai, Ziyu Hu, Shuqi Zhang, and Lianjun Liu. A survey of detection-based video multi-object tracking. Displays, page 102317, 2022. 3 +[7] Ling-Yu Duan, Tiejun Huang, and Wen Gao. Overview of the mpeg cdvs standard. In 2015 Data Compression Conference, pages 323-332, 2015. 2 +[8] Ling-Yu Duan, Yihang Lou, Yan Bai, Tiejun Huang, Wen Gao, Vijay Chandrasekhar, Jie Lin, Shiqi Wang, and Alex Chichung Kot. Compact descriptors for video analysis: The emerging mpg standard. IEEE MultiMedia, 26(2): 44-54, 2019. 2 +[9] Yann Dubois, Benjamin Bloem-Reddy, Karen Ullrich, and Chris J Maddison. Lossy compression for lossless prediction. Advances in Neural Information Processing Systems, 34:14014–14028, 2021. 2 +[10] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In European Conference on Computer Vision, pages 510-528. Springer, 2022. 3 +[11] Kristian Fischer, Fabian Brand, and André Kaup. Boosting neural image compression for machines using latent space masking. arXiv preprint arXiv:2112.08168, 2021. 1, 2 +[12] Hadi Hadizadeh and Ivan V Bajić. Learned scalable video coding for humans and machines. arXiv preprint arXiv:2307.08978, 2023. +[13] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 1, 2 +[14] Zhihao Hu, Guo Lu, and Dong Xu. Fvc: A new framework towards deep video compression in feature space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1502-1511, 2021. 1, 2, 4, 7 + +[15] Zhihao Hu, Guo Lu, Jinyang Guo, Shan Liu, Wei Jiang, and Dong Xu. Coarse-to-fine deep video coding with hyperprior-guided mode prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5921-5930, 2022. 1 +[16] Sayed Hossein Khatoonabadi and Ivan V Bajic. Video object tracking in the compressed domain using spatio-temporal markov random fields. IEEE transactions on image processing, 22(1):300-313, 2012. 3 +[17] Congcong Li, Xinyao Wang, Longyin Wen, Dxiang Hong, Tiejian Luo, and Libo Zhang. End-to-end compressed video representation learning for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13967-13976, 2022. 3 +[18] Jiahao Li, Bin Li, and Yan Lu. Deep contextual video compression. Advances in Neural Information Processing Systems, 34:18114-18125, 2021. 1, 2, 4, 5, 7 +[19] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22616-22626, 2023. 1 +[20] Xin Li, Jun Shi, and Zhibo Chen. Task-driven semantic coding via reinforcement learning. IEEE Transactions on Image Processing, 30:6307-6320, 2021. 1, 2 +[21] Hongbin Lin, Bolin Chen, Zhichen Zhang, Jielian Lin, Xu Wang, and Tiesong Zhao. Deep scalable video coding for both machine and human vision. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9205-9214, 2023. 3, 8 +[22] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7083-7093, 2019. 5, 7 +[23] Qiankun Liu, Bin Liu, Yue Wu, Weihai Li, and Nenghai Yu. Real-time online multi-object tracking in compressed domain. arXiv preprint arXiv:2204.02081, 2022. 3 +[24] Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, and Zhiyong Gao. Dvc: An end-to-end deep video compression framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11006-11015, 2019. 1, 2 +[25] Guo Lu, Xiaoyun Zhang, Wanli Ouyang, Li Chen, Zhiyong Gao, and Dong Xu. An end-to-end learning framework for video compression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3292-3308, 2021. 1, 2 +[26] Guo Lu, Xingtong Ge, Tianxiong Zhong, Jing Geng, and Qiang Hu. Preprocessing enhanced image compression for machine vision. arXiv preprint arXiv:2206.05650, 2022. 2 +[27] Fabian Mentzer, Eirikur Agustsson, Johannes Balle, David Minnen, Nick Johnston, and George Toderici. Neural video compression using gans for detail synthesis and propagation. In European Conference on Computer Vision, pages 562-578. Springer, 2022. 2 +[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large + +scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5 +[29] Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. Temporal context mining for learned video compression. IEEE Transactions on Multimedia, 2022. 1, 2, 5, 7 +[30] Yuheng Shi, Naiyan Wang, and Xiaojie Guo. Yolov: making still image object detectors great at video object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2254-2262, 2023. 5, 6 +[31] Zheng Shou, Xudong Lin, Yannis Kalantidis, Laura Sevilla-Lara, Marcus Rohrbach, Shih-Fu Chang, and Zhicheng Yan. Dmc-net: Generating discriminative motion cues for fast compressed video action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1268-1277, 2019. 3 +[32] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5 +[33] Gary J. Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on Circuits and Systems for Video Technology, 22(12):1649-1668, 2012. 1, 2, 7 +[34] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 5 +[35] Yuan Tian, Guo Lu, Xiongkuo Min, Zhaohui Che, Guang-tao Zhai, Guodong Guo, and Zhiyong Gao. Self-conditioned probabilistic learning of video rescaling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4490-4499, 2021. 2 +[36] Yuan Tian, Guo Lu, Guangtao Zhai, and Zhiyong Gao. Nonsementics suppressed mask learning for unsupervised video semantic compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13610-13622, 2023. 1, 3 +[37] Yuan Tian, Guo Lu, Yichao Yan, Guangtao Zhai, Li Chen, and Zhiyong Gao. A coding framework and benchmark towards low-bitrate video understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3 +[38] Robert Torfason, Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Towards image understanding from deep compression without decoding. arXiv preprint arXiv:1803.06131, 2018. 1, 2 +[39] Shiyao Wang, Hongchao Lu, and Zhidong Deng. Fast object detection in compressed video. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7104-7113, 2019. 3 +[40] Shurun Wang, Shiqi Wang, Wenhan Yang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. Towards analysis-friendly face representation with scalable feature and texture compression. IEEE Transactions on Multimedia, 2021. 1, 2 +[41] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402. IEEE, 2003. 2, 4 + +[42] T. Wiegand, G.J. Sullivan, G. Bjontegaard, and A. Luthra. Overview of the h.264/avc video coding standard. IEEE Transactions on Circuits and Systems for Video Technology, 13(7):560-576, 2003. 1, 2 +[43] Chao-Yuan Wu, Manzil Zaheer, Hexiang Hu, R Manmatha, Alexander J Smola, and Philipp Krahenbuhl. Compressed video action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6026-6035, 2018. 3 +[44] Ning Yan, Changsheng Gao, Dong Liu, Houqiang Li, Li Li, and Feng Wu. Sssic: semantics-to-signal scalable image coding with learned structural representations. IEEE Transactions on Image Processing, 30:8939-8954, 2021. 1, 2 +[45] Ren Yang, Luc Van Gool, and Radu Timofte. Perceptual learned video compression with recurrent conditional gan. arXiv preprint arXiv:2109.03082, 1, 2021. 2 +[46] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In European Conference on Computer Vision, pages 1-21. Springer, 2022. 5 +[47] Ezgi Özyilkan, Mateen Ulhaq, Hyomin Choi, and Fabien Racapé. Learned disentangled latent representations for scalable image coding for humans and machines. In 2023 Data Compression Conference (DCC), pages 42-51, 2023. 1, 2 \ No newline at end of file diff --git a/2024/Task-Aware Encoder Control for Deep Video Compression/images.zip b/2024/Task-Aware Encoder Control for Deep Video Compression/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..007bf336cc218a88207adad361107132e8d4df65 --- /dev/null +++ b/2024/Task-Aware Encoder Control for Deep Video Compression/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:669bc077a01e74dc5036fe08e17c3354cb1cdd73d487d4776bd3311148274eda +size 479586 diff --git a/2024/Task-Aware Encoder Control for Deep Video Compression/layout.json b/2024/Task-Aware Encoder Control for Deep Video Compression/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..47d8414f5ef65a0800c677005516211ab83422c1 --- /dev/null +++ b/2024/Task-Aware Encoder Control for Deep Video Compression/layout.json @@ -0,0 +1,8697 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 103, + 479, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 103, + 479, + 121 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 479, + 121 + ], + "type": "text", + "content": "Task-Aware Encoder Control for Deep Video Compression" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "spans": [ + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "text", + "content": "Xingtong Ge" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "text", + "content": " Jixiang Luo" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "text", + "content": " Xinjie Zhang" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "text", + "content": " Tongda Xu" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "text", + "content": " Guo Lu" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "text", + "content": " \nDailan He" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "text", + "content": " Jing Geng" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "inline_equation", + "content": "^{1\\ddagger}" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "text", + "content": " Yan Wang" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "text", + "content": " Jun Zhang" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "text", + "content": " Hongwei Qin" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "text", + "content": "Beijing Institute of Technology " + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "text", + "content": "SenseTime Research \n" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "text", + "content": "The Hong Kong University of Science and Technology \n" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "text", + "content": "Institute for AI Industry Research (AIR), Tsinghua University \n" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "text", + "content": "Shanghai Jiaotong University " + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 90, + 141, + 503, + 230 + ], + "type": "text", + "content": "The Chinese University of Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 280, + 290, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 280, + 290, + 510 + ], + "spans": [ + { + "bbox": [ + 45, + 280, + 290, + 510 + ], + "type": "text", + "content": "Prior research on deep video compression (DVC) for machine tasks typically necessitates training a unique codec for each specific task, mandating a dedicated decoder per task. In contrast, traditional video codecs employ a flexible encoder controller, enabling the adaptation of a single codec to different tasks through mechanisms like mode prediction. Drawing inspiration from this, we introduce an innovative encoder controller for deep video compression for machines. This controller features a mode prediction and a Group of Pictures (GoP) selection module. Our approach centralizes control at the encoding stage, allowing for adaptable encoder adjustments across different tasks, such as detection and tracking, while maintaining compatibility with a standard pre-trained DVC decoder. Empirical evidence demonstrates that our method is applicable across multiple tasks with various existing pre-trained DVCs. Moreover, extensive experiments demonstrate that our method outperforms previous DVC by about " + }, + { + "bbox": [ + 45, + 280, + 290, + 510 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 45, + 280, + 290, + 510 + ], + "type": "text", + "content": " bitrate for different tasks, with only one pre-trained decoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 544, + 128, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 128, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 128, + 558 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 565, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 287, + 673 + ], + "type": "text", + "content": "Over the past decades, video analysis techniques have proliferated across a variety of fields, including smart cities, autonomous vehicles, and traffic surveillance. For these applications, videos are often compressed before being transmitted to cloud-based systems for further machine vision analyses, such as object detection and tracking. However, current video compression techniques, which range from traditional codecs [2, 33, 42] to recent learning-based codecs [14, 18, 24, 25, 29], primarily cater to the human vi" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 254, + 545, + 354 + ], + "blocks": [ + { + "bbox": [ + 308, + 254, + 545, + 354 + ], + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 354 + ], + "type": "image", + "image_path": "a105509d85e278cad2fb12c7a0de63003730f785de90480a600d7fe39344373c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 361, + 547, + 407 + ], + "lines": [ + { + "bbox": [ + 304, + 361, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 547, + 407 + ], + "type": "text", + "content": "Figure 1. (a) Mainstream video codec that serves the human viewing. (b) Our controlled video codecs for machine vision with fixed decoder. (c) Other video codec for machine vision with one-to-one encoders and decoders." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 413, + 546, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 546, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 546, + 461 + ], + "type": "text", + "content": "sual system, as shown in Fig. 1(a). This specificity leads to inefficiencies, as machine vision tasks typically focus on selective semantic details and regions within the video frames, rather than the whole frames." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 461, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 547, + 676 + ], + "type": "text", + "content": "To tackle this problem, recent studies [1, 4, 11-13, 20, 38, 40, 44, 47] have explored scalable compression for multiple tasks through different layers in image and video coding. They propose a base layer dedicated to machine vision, with an enhancement layer containing additional information for human vision. Additionally, the latest work [36] has tried to compress semantic features at the encoder side, which serve as side information to supplement decoded images with more semantic details for machine vision at lower bitrates. Despite these advances, these methods require individually customized CODECs when applied to different downstream tasks, as the \"One-to-one CODECs\" shown in Fig. 1(c). In contrast, traditional CODECs can control the encoding process for different objectives, like PSNR and SSIM optimization, using a single decoder. Inspired by this feature, in this paper, we focus on how to use one pre-trained DVC-decoder to support both human and multiple machine vision tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "In both residual [14, 15, 24] and conditional [18, 19, 29] DVC series, the encoded bitstreams predominantly comprise two parts: motion and residual/contextual informa" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 682, + 282, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 682, + 282, + 693 + ], + "spans": [ + { + "bbox": [ + 57, + 682, + 282, + 693 + ], + "type": "text", + "content": "*Work was done when Xingtong Ge interned at SenseTime Research" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 693, + 143, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 693, + 143, + 703 + ], + "spans": [ + { + "bbox": [ + 59, + 693, + 143, + 703 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 59, + 693, + 143, + 703 + ], + "type": "text", + "content": " xingtong.ge@gmail.com" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 703, + 205, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 205, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 205, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 59, + 703, + 205, + 712 + ], + "type": "text", + "content": "Corresponding author, janegeng@bit.edu.cn" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26036" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": "tion. It is observed that a substantial portion of the bitrate, often exceeding " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": " in high bitrate models, is allocated to the residual/contextual information. This allocation is primarily for high-quality frame reconstruction. However, this approach is inappropriate for downstream vision tasks such as tracking and detection, which primarily concentrate on the objects and their movements rather than the full frames." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 158, + 286, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 286, + 373 + ], + "type": "text", + "content": "To mitigate this issue, we introduce a dynamic vision mode prediction (DVMP) module specifically tailored for machine vision tasks. This module optimizes the entropy coding process by dynamically predicting the skip/no-skip coding mode for each feature element, based on its relevance to machine vision. Specifically, it takes hyperprior data from motion or residual/contextual features within the DVC framework as the input and assesses the utility of each feature element for machine vision. By replacing non-essential elements with their predicted mean value derived from the hyperprior network, our method effectively reduces bit consumption and circumvent the entropy decoding step for these elements to expedite the decoding process. In this case, we generate a novel frame type that significantly reduces bitrate usage compared to traditional P frames in the Group of Pictures (GoP) structure, while maintaining the information valuable for machine vision tasks. This new frame is designated as the " + }, + { + "bbox": [ + 46, + 158, + 286, + 373 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 46, + 158, + 286, + 373 + ], + "type": "text", + "content": " frame." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 375, + 286, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 286, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 286, + 542 + ], + "type": "text", + "content": "At the same time, it is important for DVC to maintain reconstruction quality because there exists significant reference relationship between frames, which is a crucial difference between image and video compression. Since our " + }, + { + "bbox": [ + 46, + 375, + 286, + 542 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 46, + 375, + 286, + 542 + ], + "type": "text", + "content": " frame degrades the fidelity of reconstruction compared to the " + }, + { + "bbox": [ + 46, + 375, + 286, + 542 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 375, + 286, + 542 + ], + "type": "text", + "content": " frame, we explore the reorganization of the Group of Pictures (GoP) structure. To be specific, we firstly explore to use a hand-crafted GoP structure, which already achieves significant rate-precision improvements. In the revised structure, both " + }, + { + "bbox": [ + 46, + 375, + 286, + 542 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 375, + 286, + 542 + ], + "type": "text", + "content": " frames and " + }, + { + "bbox": [ + 46, + 375, + 286, + 542 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 46, + 375, + 286, + 542 + ], + "type": "text", + "content": " frames derive their prediction from the preceding " + }, + { + "bbox": [ + 46, + 375, + 286, + 542 + ], + "type": "inline_equation", + "content": "I / P" + }, + { + "bbox": [ + 46, + 375, + 286, + 542 + ], + "type": "text", + "content": " frame. Further, we introduce a GoP selection network, which can dynamically determine the GoP structure during the encoding process and achieves a better rate-precision trade-off." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 544, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 286, + 675 + ], + "type": "text", + "content": "We choose three representative residual and conditional DVC methods as touchstones in our proposed video coding for machine vision framework. It's worth mentioning that our method utilizes pre-trained DVCs without altering the weights of the original encoders and decoders. Our control happens at encoding stage, leaving the decoders' architecture intact. Moreover, when human viewing is required, we can switch back to the original encoding procedure to restore reconstruction performance. In essence, using the proposed method, we can control the encoder of a DVC to adapt for both machine and human vision requirements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "The main contributions of our work are summarised as follows: (1) Built upon the mainstream DVC CODECs, we propose a novel video coding for machine vision framework" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "that controls the encoder to adapt for different downstream vision tasks, such as video object detection and multi-object tracking. (2) We employ a dynamic vision mode prediction approach to refine the original P frame, effectively reducing the bitrate while preserving critical information pertinent to vision tasks. Furthermore, we utilize a GoP selection strategy to dynamically forecast the coding GoP structure during the encoding stage, which controls the bitstream for better rate-precision trade-off in downstream vision tasks. (3) Experiments show that our controlling method is novel and flexible for different DVC CODECs, achieving up to more than " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " bitrate savings in various downstream tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 228, + 397, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 397, + 240 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 397, + 240 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 247, + 419, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 247, + 419, + 261 + ], + "spans": [ + { + "bbox": [ + 306, + 247, + 419, + 261 + ], + "type": "text", + "content": "2.1. Video Compression" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 267, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 411 + ], + "type": "text", + "content": "Previous video codecs are designed to remove spatial-temporal redundancies effectively. Traditional video coding standards, including H.264/AVC [42], H.265/HEVC [33], and H.266/VVC [2], significantly increase the compression efficiency of images and videos. Recently, a number of learning-based video codecs [14, 18, 24, 25, 29, 35] have been proposed, designed based on residual coding or conditional coding, achieved better and better pixel-wise signal quality metrics, e.g., PSNR and MS-SSIM [41], which mainly serve the human visual experience. Recently, there are also some generative video coding methods [27, 45] that mainly consider visual comfort and perceptual quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 419, + 481, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 419, + 481, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 419, + 481, + 431 + ], + "type": "text", + "content": "2.2. Compression for Machine Vision" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": "To facilitate the efficiency of machine vision tasks, early research works were devoted to extracting visual features from signals. Early standards, such as CDVA [8] and CDVS [7], suggest the pre-extraction and transportation of image keypoints to facilitate image indexing or retrieval tasks. With the development of deep learning, some studies began to explore connections between compression and downstream tasks, some studies [1, 3, 4, 11, 13, 20, 38, 40, 44, 47] also focus on the joint optimization of image compression and downstream machine vision tasks by introducing a rate-distortion optimization strategy guided by downstream tasks or by adding a task-specific feature encoding stream. For instance, Torfason et al. [38] introduced a method for executing image understanding tasks, such as classification and segmentation, on compressed outputs from learning-based image compression techniques. Furthermore, Lu et al. [26] enhanced traditional CODECs with a preprocessing step, thereby improving codec performance for downstream vision tasks. The field has also witnessed the emergence of self-supervised representation learning methods aimed at deriving compact semantic representations. Dubois et al. [9] presented a theoretical framework suggesting that the distortion term in the lossy rate" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26037" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 75, + 331, + 168 + ], + "blocks": [ + { + "bbox": [ + 83, + 75, + 331, + 168 + ], + "lines": [ + { + "bbox": [ + 83, + 75, + 331, + 168 + ], + "spans": [ + { + "bbox": [ + 83, + 75, + 331, + 168 + ], + "type": "image", + "image_path": "44edc55052840a0107689515e6c08b00e26a7d95015b77a3dae11631b6ec9bd2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 169, + 230, + 178 + ], + "lines": [ + { + "bbox": [ + 220, + 169, + 230, + 178 + ], + "spans": [ + { + "bbox": [ + 220, + 169, + 230, + 178 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 352, + 86, + 500, + 150 + ], + "blocks": [ + { + "bbox": [ + 352, + 86, + 500, + 150 + ], + "lines": [ + { + "bbox": [ + 352, + 86, + 500, + 150 + ], + "spans": [ + { + "bbox": [ + 352, + 86, + 500, + 150 + ], + "type": "image", + "image_path": "69851fbeec1e47931de7146feb70572da047965607c43313fbdc543c62f0f00e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 91, + 180, + 435, + 261 + ], + "blocks": [ + { + "bbox": [ + 421, + 167, + 431, + 177 + ], + "lines": [ + { + "bbox": [ + 421, + 167, + 431, + 177 + ], + "spans": [ + { + "bbox": [ + 421, + 167, + 431, + 177 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 91, + 180, + 435, + 261 + ], + "lines": [ + { + "bbox": [ + 91, + 180, + 435, + 261 + ], + "spans": [ + { + "bbox": [ + 91, + 180, + 435, + 261 + ], + "type": "image", + "image_path": "e855742e4276dd2605f2f1d2e1cde33146445af297fd7cc6a43ee77e846518a5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 262, + 321, + 271 + ], + "lines": [ + { + "bbox": [ + 311, + 262, + 321, + 271 + ], + "spans": [ + { + "bbox": [ + 311, + 262, + 321, + 271 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 434, + 169, + 505, + 262 + ], + "blocks": [ + { + "bbox": [ + 434, + 169, + 505, + 262 + ], + "lines": [ + { + "bbox": [ + 434, + 169, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 434, + 169, + 505, + 262 + ], + "type": "image", + "image_path": "868e6e6aa89606bc49050a15b87cbea5a532facb32768c0fcad6e441f1e409ab.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 263, + 505, + 271 + ], + "lines": [ + { + "bbox": [ + 445, + 263, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 445, + 263, + 505, + 271 + ], + "type": "text", + "content": "GoP Feature Extractor" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 284, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 284, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 547, + 319 + ], + "type": "text", + "content": "Figure 2. (a) Overview of our \"Controlling DVC for Machine\" framework. Given an input GoP, we firstly use GoP Selection network to predict the GoP structure, then the predicted structure controls the encoding procedure to encode input frames for machine vision tasks. (b) The \"0\" element controls encoder to use DVMP. (c) The GoP Selection network, including the pre-analysis stage and GoP prediction stage." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 330, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 287, + 400 + ], + "type": "text", + "content": "distortion trade-off for image classification could be approximated by a contrastive learning objective. Feng et al. [10] proposed a method to learn a unified feature representation for AI tasks from unlabeled data. These methodologies necessitate fine-tuning the downstream models to adapt to the learned features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 402, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 533 + ], + "type": "text", + "content": "Despite advancements in image coding for machine applications, reconstructing high-fidelity video from extracted features remains a formidable challenge due to the critical inter-frame reference relationships. Addressing the requirements of both machine and human vision, Tian et al. [36] introduced a self-supervised edge representation as a semantic intermediary, which preserves the video's semantic structure. Furthermore, Lin et al. [21] developed a scalable video coding framework tailored for machines, segregating semantic features for machine analysis and human viewing into separate bitstreams." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 533, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 628 + ], + "type": "text", + "content": "Most existing approaches [21, 36, 37] necessitate the use of distinct encoders and decoders for various tasks, which complicates and burdens the coding pipelines. Our methodology advances the \"Coding for Machine\" paradigm by focusing on the encoder side. We present a versatile framework capable of accommodating both human and machine vision demands without the need to modify the pre-trained decoders of existing video codecs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 635, + 273, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 273, + 648 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 273, + 648 + ], + "type": "text", + "content": "2.3. Compressed Video Analysis/Understanding" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "There are also amounts of works performing video analysis tasks [17, 23, 39], such as image recognition, action recognition [31, 43] and multiple object tracking (MOT) [6, 16], in the compressed video domain. However, these methods focus on developing video analysis models that better lever" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 330, + 546, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 330, + 546, + 366 + ], + "spans": [ + { + "bbox": [ + 305, + 330, + 546, + 366 + ], + "type": "text", + "content": "age the partially decoded video stream, such as the motion vector. In contrast, our work focuses on the coding procedure, specifically the encoding procedure." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 375, + 361, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 375, + 361, + 387 + ], + "spans": [ + { + "bbox": [ + 306, + 375, + 361, + 387 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 395, + 373, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 373, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 373, + 407 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "text", + "content": "Our coding framework is illustrated in Fig. 2 (a). Let " + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{\\pmb{x}_1,\\dots ,\\pmb{x}_T\\}" + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "text", + "content": " represent the video sequence. The framework categorizes the " + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "text", + "content": " frames into two types: original " + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "text", + "content": " frames and machine vision-specific " + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "text", + "content": " frames. The categorization of frame " + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "text", + "content": " is determined by the GoP Selection Network, where a value of 1 signifies a " + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "text", + "content": " frame and 0 a " + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "text", + "content": " frame. The " + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "text", + "content": " frames, designed to reduce the bitrate while preserving critical information for future vision tasks, are generated from the preceding " + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "text", + "content": " frames using Dynamic Vision Mode Prediction (DVMP). To maintain the quality of the decoded video sequence, each " + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "text", + "content": " frame is computed from the previous " + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "text", + "content": " frame with superior visual quality, rather than from a " + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "text", + "content": " frame. In scenarios requiring coding for downstream vision tasks, a machine-centric control applies the GoP selection and DVMP, altering the encoding structure to \"I, " + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "inline_equation", + "content": "P_{m},P,\\dots" + }, + { + "bbox": [ + 304, + 413, + 546, + 628 + ], + "type": "text", + "content": "\". Conversely, for human viewing, a human-centric control restores the encoding structure to its original form, \"I,P,P,P,,.." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 635, + 484, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 484, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 484, + 647 + ], + "type": "text", + "content": "3.2. Dynamic Vision Mode Prediction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "content": "In our coding framework, we aim to minimize the video sequence bitrate. Initial analyses of residual video CODECs reveal that they allocate a substantial portion of the bitrate—exceeding " + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "content": " in high bitrate configurations—to encode and transmit residual information. This approach en" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26038" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 276, + 191 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 276, + 191 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 276, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 276, + 191 + ], + "type": "image", + "image_path": "57c3885ba644d58e3c8ff878e11daa3fc9de30a0238beeea7ef83aed31901a54.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 199, + 287, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 199, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 287, + 220 + ], + "type": "text", + "content": "Figure 3. Hyper-prior guided Dynamic Vision Mode Prediction network." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 244, + 286, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 244, + 286, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 286, + 340 + ], + "type": "text", + "content": "sures high-quality visuals for human perception, as measured by metrics such as PSNR or MS-SSIM [41]. However, this method introduces significant redundancy, especially since downstream tasks like video object detection or tracking predominantly concentrate on specific regions of interest rather than the entirety of the frame. Consequently, it is crucial for our study to develop an approach that reduces this redundancy in the coding bitstream." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 341, + 286, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 286, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 286, + 424 + ], + "type": "text", + "content": "Inspired by Hu et al., who developed a mode prediction technique specifically designed for the human visual system within a slimmable encoder and decoder, we introduce the Dynamic Vision Mode Prediction (DVMP) module. This innovative module autonomously selects the optimal coding mode and decides whether each feature element should be encoded and transmitted, as illustrated in Fig. 3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 286, + 700 + ], + "type": "text", + "content": "To elucidate the functionality of our proposed Dynamic Vision Mode Prediction (DVMP) module, we utilize the encoded residual feature " + }, + { + "bbox": [ + 46, + 426, + 286, + 700 + ], + "type": "inline_equation", + "content": "m_{t}" + }, + { + "bbox": [ + 46, + 426, + 286, + 700 + ], + "type": "text", + "content": " as an illustrative example. Assume " + }, + { + "bbox": [ + 46, + 426, + 286, + 700 + ], + "type": "inline_equation", + "content": "m_{t}" + }, + { + "bbox": [ + 46, + 426, + 286, + 700 + ], + "type": "text", + "content": " possesses dimensions " + }, + { + "bbox": [ + 46, + 426, + 286, + 700 + ], + "type": "inline_equation", + "content": "c \\times h \\times w" + }, + { + "bbox": [ + 46, + 426, + 286, + 700 + ], + "type": "text", + "content": ", indicating " + }, + { + "bbox": [ + 46, + 426, + 286, + 700 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 426, + 286, + 700 + ], + "type": "text", + "content": " channels each with a spatial dimension of " + }, + { + "bbox": [ + 46, + 426, + 286, + 700 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 46, + 426, + 286, + 700 + ], + "type": "text", + "content": ". The hyperprior network then predicts the mean and variance for each element, resulting in dimensions of " + }, + { + "bbox": [ + 46, + 426, + 286, + 700 + ], + "type": "inline_equation", + "content": "2c \\times h \\times w" + }, + { + "bbox": [ + 46, + 426, + 286, + 700 + ], + "type": "text", + "content": ". The mode prediction network's architecture, detailed in the upper branch, consists of two convolution layers and three ResBlocks. To render the \"mode prediction\" process differentiable, we employ the Gumbel Softmax technique during training to ascertain the skip mode for each encoded residual feature element. This module generates a binary mask for the feature map " + }, + { + "bbox": [ + 46, + 426, + 286, + 700 + ], + "type": "inline_equation", + "content": "m_{t}" + }, + { + "bbox": [ + 46, + 426, + 286, + 700 + ], + "type": "text", + "content": ", where a \"1\" denotes elements imperative for machine vision that require entropy coding, and a \"0\" signifies elements either irrelevant for machine vision or accurately predictable by the hyperprior network, thus streamlining the entropy coding process and bitrate reduction. Note that the showed architecture is not suitable for prior models with autoregressive components, like which in DCVC [18]. However, we can extend the module into an autoregressive way to enable the support for DCVC. The detailed structure is shown in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "To optimize the DVMP module, the loss function can be" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 545, + 163 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 163 + ], + "type": "image", + "image_path": "87de36ec866ecd879792b0c7bc1ff9fce01cf868895a9e0624f3c005a1254f0d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "lines": [ + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "type": "text", + "content": "Figure 4. Different coding GoP structures. The original structure consists of I and P frames. In the middle, the hand-crafted structure for machine vision consists of I, P and " + }, + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 304, + 172, + 545, + 238 + ], + "type": "text", + "content": " frames which are arranged alternately. In the predicted (GoP selected) structure, the type of frames are predicted by the GoP selection network, targeting on better bitrate and machine vision performance trade-off." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 252, + 346, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 252, + 346, + 263 + ], + "spans": [ + { + "bbox": [ + 306, + 252, + 346, + 263 + ], + "type": "text", + "content": "given by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 383, + 266, + 545, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 266, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 383, + 266, + 545, + 278 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m} = R + \\lambda_ {m} \\mathcal {L} _ {t} \\tag {1}", + "image_path": "6b02f8d266e0707df8127c4bd74b803b73e2bdb982e7a6dde1bc79827a95611e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 285, + 545, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 322 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 285, + 545, + 322 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 285, + 545, + 322 + ], + "type": "text", + "content": " denotes the coding bitrate and " + }, + { + "bbox": [ + 304, + 285, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 304, + 285, + 545, + 322 + ], + "type": "text", + "content": " denotes the loss of downstream vision task, respectively. " + }, + { + "bbox": [ + 304, + 285, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 304, + 285, + 545, + 322 + ], + "type": "text", + "content": " is a hyperparameter used to control the trade-off." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 331, + 446, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 331, + 446, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 446, + 342 + ], + "type": "text", + "content": "3.3. DivGoP " + }, + { + "bbox": [ + 306, + 331, + 446, + 342 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 306, + 331, + 446, + 342 + ], + "type": "text", + "content": " GoP Selection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "text", + "content": "Obtaining the " + }, + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "text", + "content": " frames, we firstly use a hand-crafted structure in a GoP, where the " + }, + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "text", + "content": " frames and " + }, + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "text", + "content": " frames are arranged alternately (referred as \"DivGoP\" and shown in Fig.4). It is observed that this arrangement can bring a certain degree of bitrate saving in detection, since the " + }, + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "text", + "content": " frames cost much less bitrate than " + }, + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "text", + "content": " frames while maintaining the motion and machine vision information. Experimental findings, as depicted in the \"FVC DivGoP\" curve within Fig. 5, demonstrate that the \"I, " + }, + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "inline_equation", + "content": "P_{m}, P, \\cdots" + }, + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "text", + "content": "\" structure markedly achieve about " + }, + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "inline_equation", + "content": "18.6\\%" + }, + { + "bbox": [ + 304, + 350, + 545, + 470 + ], + "type": "text", + "content": " bitrate saving." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 471, + 546, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 546, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 546, + 686 + ], + "type": "text", + "content": "The \"DivGoP\" configuration, while methodically structured, may not be universally optimal due to the variability in motion and object information across different videos, which could necessitate distinct GoP structures. For instance, videos characterized by rapid target movement might demand a higher frequency of " + }, + { + "bbox": [ + 304, + 471, + 546, + 686 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 471, + 546, + 686 + ], + "type": "text", + "content": " frames to preserve the quality of reconstruction, as an excess of " + }, + { + "bbox": [ + 304, + 471, + 546, + 686 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 304, + 471, + 546, + 686 + ], + "type": "text", + "content": " frames may lead to degradation that negatively affects subsequent frames. Conversely, in scenarios where the camera remains static or the target movement is minimal, the introduction of more " + }, + { + "bbox": [ + 304, + 471, + 546, + 686 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 304, + 471, + 546, + 686 + ], + "type": "text", + "content": " frames could be feasible without significantly impacting the reconstruction quality of following frames. Stimulated by this notion, our research endeavors to tailor the GoP structure dynamically, enhancing codec performance for machine vision tasks. To approach this issue, we adopt FVC [14] as the foundational codec, select a GoP size of 10, and address the detection task utilizing the MOT Dataset. The optimization problem is thus defined:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 367, + 696, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 696, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 367, + 696, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\underset {\\theta} {\\arg \\min } R (\\theta) + \\lambda_ {p} L _ {\\det } (\\theta) \\tag {2}", + "image_path": "49ae8bb7fe8d361fabeab22eadb941db3df8ed2a713ad4f4d86fae23f55859a6.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26039" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": " denotes the bitrate, " + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "inline_equation", + "content": "L_{det}" + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": " denotes the detection loss, " + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "inline_equation", + "content": "\\lambda_p" + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": " is a weighting factor that balances the effects of bitrate and detection loss, and " + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": " denotes the setting of GoP structure here." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 288, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 288, + 374 + ], + "type": "text", + "content": "Obtaining a GoP, each of the 9 predicted frames can be either " + }, + { + "bbox": [ + 46, + 122, + 288, + 374 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 122, + 288, + 374 + ], + "type": "text", + "content": " frame or " + }, + { + "bbox": [ + 46, + 122, + 288, + 374 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 46, + 122, + 288, + 374 + ], + "type": "text", + "content": " frame, which can be dynamically decided. To explore the optimal GoP structure for target function 2, and evaluate the performance gap between DivGoP and optimal GoP structure, we use a deep-first-searching (DFS) algorithm, where we use different " + }, + { + "bbox": [ + 46, + 122, + 288, + 374 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 122, + 288, + 374 + ], + "type": "text", + "content": " for different bitrate models. The DFS algorithm searches for the structure that makes the target function smallest for each GoP. Results are shown in the \"FVC DFS\" curve of Fig.5. The application of the DFS algorithm led to a substantial increase in Bpp-mAP performance,, achieving about " + }, + { + "bbox": [ + 46, + 122, + 288, + 374 + ], + "type": "inline_equation", + "content": "32.3\\%" + }, + { + "bbox": [ + 46, + 122, + 288, + 374 + ], + "type": "text", + "content": " bitrate saving in terms of the BD-RATE metric, while the \"Div-GoP\" method achieves about " + }, + { + "bbox": [ + 46, + 122, + 288, + 374 + ], + "type": "inline_equation", + "content": "18.6\\%" + }, + { + "bbox": [ + 46, + 122, + 288, + 374 + ], + "type": "text", + "content": " bitrate saving. Similar conclusions can also be observed on DCVC [18] and TCM [29], reinforcing our hypothesis that distinct videos necessitate unique GoP structures tailored for machine vision tasks. However, the DFS algorithm is extremely time consuming, and the labels do not exist during real inference process, making the DFS not available to be applied in real time encoding. How to further find an \"adaptive GoP structure\" in an available time complexity is worth exploring." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 376, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 639 + ], + "type": "text", + "content": "From this perspective, we introduce our GoP selection method to adaptively predict the GoP structure at encoding stage, which has a low time complexity and is available in application. Illustrated in Fig.2 (c), our GoP selection network operates in two phases: the pre-analysis and the GoP prediction stages. Initially, input frames undergo preanalysis, where a detector identifies objects and generates an overlapping mask for each frame, while the RAFT [34] optical flow estimation method calculates optical flow features for each predicted frame using its adjacent predecessor. These features are then refined by masking with the object bounding box mask to produce masked optical flow features, alongside motion feature priors (mean and variance, averaged by channel) derived from the codec's motion encoder. Subsequently, these processed features are input into the GoP prediction network, which employs a feature extractor with convolution blocks for downsampling and aggregation, followed by adaptive average pooling and linear layers that output a normalized " + }, + { + "bbox": [ + 46, + 376, + 287, + 639 + ], + "type": "inline_equation", + "content": "s_{\\text{logit}}" + }, + { + "bbox": [ + 46, + 376, + 287, + 639 + ], + "type": "text", + "content": " for this GoP, which comprises two elements that sum to one. The averaged detection confidence information from the detector is also injected into the linear layers as meta data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Upon acquiring the " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{logit}}" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": ", we employ the Gumbel softmax sampling technique throughout the training phase to iteratively generate a binary value for each frame. This process culminates in the formation of a GoP structure vector, where the binary value—0 for a " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": " frame and 1 for a " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": " frame—specifies the type of each frame within the GoP." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 427, + 173 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 427, + 173 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 427, + 173 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 427, + 173 + ], + "type": "image", + "image_path": "2ed356220b088e2248ecfe355c84f5531407d576b4f81e64620881d5621c41dc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 182, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 305, + 182, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 182, + 545, + 205 + ], + "type": "text", + "content": "Figure 5. Left: Using DFS to search for the near-optimal GoP structure for Bpp-mAP. Right: Results of simply fine-tuning FVC" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 427, + 71, + 545, + 172 + ], + "blocks": [ + { + "bbox": [ + 427, + 71, + 545, + 172 + ], + "lines": [ + { + "bbox": [ + 427, + 71, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 427, + 71, + 545, + 172 + ], + "type": "image", + "image_path": "a94aa44b7630bcc5bf4f5233448509af4db142350cd6c437bbc3c956deef741d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 224, + 545, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 224, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 545, + 296 + ], + "type": "text", + "content": "During inference stage, we apply a softmax operation on " + }, + { + "bbox": [ + 305, + 224, + 545, + 296 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{logit}}" + }, + { + "bbox": [ + 305, + 224, + 545, + 296 + ], + "type": "text", + "content": " to derive probabilities, which are then used to evenly distribute 1s and 0s, thereby determining the actual encoded GoP structure. Similar to our mode prediction network, the loss function for optimizing GoP structure selection network is given by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 386, + 310, + 545, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 310, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 386, + 310, + 545, + 324 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g} = \\bar {R} + \\lambda_ {g} \\bar {\\mathcal {L}} _ {t} \\tag {3}", + "image_path": "10d09e0a554743c5e37f2fffa35d5668e24fa53c299d1caf38098abcd29aca73.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 331, + 545, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 545, + 355 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 305, + 331, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\bar{R}" + }, + { + "bbox": [ + 305, + 331, + 545, + 355 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 331, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{L}}_t" + }, + { + "bbox": [ + 305, + 331, + 545, + 355 + ], + "type": "text", + "content": " denote the GoP-wise average coding bitrate and average loss of downstream vision task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 369, + 387, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 387, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 387, + 382 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 388, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 388, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 545, + 472 + ], + "type": "text", + "content": "In this section, we describe the dataset settings, downstream task models, and evaluation metrics. We then provide empirical evidence underscoring the effectiveness of our proposed method across multiple tasks and CODECs. Lastly, we validate the contributions of the mode prediction and GoP selection modules to performance improvements through ablation studies." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 473, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 473, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 305, + 473, + 545, + 593 + ], + "type": "text", + "content": "Datasets. For multi-object tracking (MOT) task, we choose MOT17 dataset. For video object detection (VOD) task, we adopt the Imagenet VID dataset [28] for training and evaluation. For video action recognition (VAR) task, we adopt UCF101[32] dataset. For the data processing procedure, we downsample the short side of the evaluation videos to 512 for MOT17 dataset and keep the original frame ratios for both training and evaluation. For the evaluation of VOD, we follow the most VOD works to resize the frames to " + }, + { + "bbox": [ + 305, + 473, + 545, + 593 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 305, + 473, + 545, + 593 + ], + "type": "text", + "content": " for training and " + }, + { + "bbox": [ + 305, + 473, + 545, + 593 + ], + "type": "inline_equation", + "content": "576 \\times 576" + }, + { + "bbox": [ + 305, + 473, + 545, + 593 + ], + "type": "text", + "content": " for evaluation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 545, + 713 + ], + "type": "text", + "content": "Downstream Task Models. For MOT task, we adipot the popular ByteTrack [46] method as the backbone model. The model weights are provided by their official framework. For VOD task, we adipot the detector YOLOv [30], and the model weights are also provided by its official framework. For the VAR task, we adipot the TSM [22] model as the recognition network, the weights are provided by mmaction2 [5]. During training and evaluation of our proposed framework, the weights of the downstream task models are fixed, without any fine-tuning." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26040" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 218, + 210 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 218, + 210 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 218, + 210 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 218, + 210 + ], + "type": "image", + "image_path": "5f27cbeccdaea6c972b064752557970cd151ab1783920f8dbdf7fe833161431f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 216, + 137, + 224 + ], + "lines": [ + { + "bbox": [ + 127, + 216, + 137, + 224 + ], + "spans": [ + { + "bbox": [ + 127, + 216, + 137, + 224 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 218, + 72, + 385, + 211 + ], + "blocks": [ + { + "bbox": [ + 218, + 72, + 385, + 211 + ], + "lines": [ + { + "bbox": [ + 218, + 72, + 385, + 211 + ], + "spans": [ + { + "bbox": [ + 218, + 72, + 385, + 211 + ], + "type": "image", + "image_path": "6339a848c9c0d0b8aafaaa4ceee93a618d0f788d98965b1f5946a1e864d01bed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 294, + 216, + 305, + 224 + ], + "lines": [ + { + "bbox": [ + 294, + 216, + 305, + 224 + ], + "spans": [ + { + "bbox": [ + 294, + 216, + 305, + 224 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 385, + 72, + 551, + 211 + ], + "blocks": [ + { + "bbox": [ + 385, + 72, + 551, + 211 + ], + "lines": [ + { + "bbox": [ + 385, + 72, + 551, + 211 + ], + "spans": [ + { + "bbox": [ + 385, + 72, + 551, + 211 + ], + "type": "image", + "image_path": "393fcbc5c3537694ff4da72484cd5e1f43ccde4a5bf417156831021d1558e4ca.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 216, + 472, + 224 + ], + "lines": [ + { + "bbox": [ + 462, + 216, + 472, + 224 + ], + "spans": [ + { + "bbox": [ + 462, + 216, + 472, + 224 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 49, + 227, + 217, + 365 + ], + "blocks": [ + { + "bbox": [ + 49, + 227, + 217, + 365 + ], + "lines": [ + { + "bbox": [ + 49, + 227, + 217, + 365 + ], + "spans": [ + { + "bbox": [ + 49, + 227, + 217, + 365 + ], + "type": "image", + "image_path": "7f1c9c2addf59ee8646531406ddd2c9ed7dc895a7ae754517308db8b67a0173a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 370, + 137, + 379 + ], + "lines": [ + { + "bbox": [ + 127, + 370, + 137, + 379 + ], + "spans": [ + { + "bbox": [ + 127, + 370, + 137, + 379 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 217, + 227, + 384, + 366 + ], + "blocks": [ + { + "bbox": [ + 217, + 227, + 384, + 366 + ], + "lines": [ + { + "bbox": [ + 217, + 227, + 384, + 366 + ], + "spans": [ + { + "bbox": [ + 217, + 227, + 384, + 366 + ], + "type": "image", + "image_path": "85c59007fe6b17fa1a765799494cd46b181e3955d57534432f7127b717baa2c1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 294, + 371, + 305, + 379 + ], + "lines": [ + { + "bbox": [ + 294, + 371, + 305, + 379 + ], + "spans": [ + { + "bbox": [ + 294, + 371, + 305, + 379 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 385, + 227, + 552, + 366 + ], + "blocks": [ + { + "bbox": [ + 385, + 227, + 552, + 366 + ], + "lines": [ + { + "bbox": [ + 385, + 227, + 552, + 366 + ], + "spans": [ + { + "bbox": [ + 385, + 227, + 552, + 366 + ], + "type": "image", + "image_path": "ff2b4c14824dae2fe4f19ec9fbc47f428c02035dbb41de37c665f0fc56796b9a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 370, + 472, + 379 + ], + "lines": [ + { + "bbox": [ + 462, + 370, + 472, + 379 + ], + "spans": [ + { + "bbox": [ + 462, + 370, + 472, + 379 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 389, + 546, + 422 + ], + "lines": [ + { + "bbox": [ + 46, + 389, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 546, + 422 + ], + "type": "text", + "content": "Figure 6. (a) Bpp-MOTA curves on MOT Dataset. (b) Bpp-mAP curves on MOT Dataset. (c) Bpp-MOTP curves on MOT Dataset. (d) Bpp-FN curves on MOT Dataset. (e) Bpp-mAP50 curves on ImageNet VID Dataset using YOLOv [30]. (f) Bpp-Top1 Accuracy curves on UCF101 Dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 437, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 557 + ], + "type": "text", + "content": "Evaluation Metrics. For compression, We use bpp (bits per pixel) to measure the average number of bits for one pixel. For the MOT task, we adopt the MOTA (multiple object tracking accuracy), MOTP (multiple object tracking precision), FN (false negative detection number) and mAP (mean average precision) metrics to evaluate the tracking and detection performance. For the VOD task, we adopt the mAP50 metric to evaluate the detection performance, following most VOD methods. For the VAR task, we adopt the Top1 Accuracy as the performance indicator." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "Implementation Details. Our whole framework is implemented on PyTorch with CUDA support and trained on eight V100 or A100 GPUs. We use the Adam optimizer by setting the learning rate as 0.0001, " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " as 0.9 and " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " as 0.999, respectively. The whole training process has the following stages: Firstly, we use the loss function 1 to train the mode prediction network and obtain the " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " frames. This stage we use single frame training. Note that for TCM, DVMP is applied to both motion and contextual information. For DCVC and FVC, DVMP is exclusively utilized for motion information, and all residual/contextual information is omitted. After that, we introduce the GoP selection network in a GOP length unit. Using the GoP size of 32" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 437, + 546, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 558 + ], + "type": "text", + "content": "on MOT and VAR tasks, we decide the GoP into 2 mini GoPs with size of 16, and the predicted vector of each mini GoP has a length of 15. For VOD task, we directly set the GoP size as 16. Using multi-frame training and loss function 3, we fix the weights of mode prediction network and train GoP selection network for about 100 epochs on MOT half-train dataset and 20 epochs on ImageNet VID training dataset, respectively for multi-object tracking and video object detection tasks. The whole training procedure takes about 2 days for each task." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 563, + 436, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 436, + 577 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 436, + 577 + ], + "type": "text", + "content": "4.1. Multi-Object Tracking" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "In Fig. 6, we present a comparison between our method and various codecs, specifically DCVC, FVC, and TCM. It is evident that our approach yields a markedly improved tradeoff in terms of Bpp-MOTA, Bpp-mAP, Bpp-MOTP and Bpp-FN metrics. When compared to the baseline codecs FVC, DCVC and TCM, our control method results in bitrate savings of approximately " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "27.54\\%" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "41.82\\%" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "12.64\\%" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": ", respectively, for equivalent MOTA metrics. With DCVC established as the anchor baseline, Table 1 displays the BD-RATE values, where our controlled TCM, controlled DCVC and controlled FVC codecs demonstrate the best, second-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26041" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 170, + 171 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 170, + 171 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 170, + 171 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 170, + 171 + ], + "type": "image", + "image_path": "404416386c280a5d2d2b2418a7ce691eee76d85e1afed2fb87244770c7914eac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 178, + 287, + 201 + ], + "lines": [ + { + "bbox": [ + 47, + 178, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 287, + 201 + ], + "type": "text", + "content": "Figure 7. Video Reconstruction Quality in terms of Bpp-PSNR on HEVC Class B dataset (Left) and Class C dataset (Right)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 72, + 290, + 171 + ], + "blocks": [ + { + "bbox": [ + 173, + 72, + 290, + 171 + ], + "lines": [ + { + "bbox": [ + 173, + 72, + 290, + 171 + ], + "spans": [ + { + "bbox": [ + 173, + 72, + 290, + 171 + ], + "type": "image", + "image_path": "39c4a621b853dc3fa1fc774816057b2c4c8788d14f1102c083c4ac1cedd67c96.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 181, + 544, + 203 + ], + "lines": [ + { + "bbox": [ + 305, + 181, + 544, + 203 + ], + "spans": [ + { + "bbox": [ + 305, + 181, + 544, + 203 + ], + "type": "text", + "content": "Figure 8. Left: Ablation study for Mode Prediction network. Right: Ablation study for GoP Selection network" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 232, + 286, + 319 + ], + "blocks": [ + { + "bbox": [ + 72, + 213, + 261, + 224 + ], + "lines": [ + { + "bbox": [ + 72, + 213, + 261, + 224 + ], + "spans": [ + { + "bbox": [ + 72, + 213, + 261, + 224 + ], + "type": "text", + "content": "Table 1. BD-RATE results on multi-object tracking." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 232, + 286, + 319 + ], + "lines": [ + { + "bbox": [ + 49, + 232, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 232, + 286, + 319 + ], + "type": "table", + "html": "
MethodMOTAmAPmAP50MOTPFNAverage
DCVC [18]0.00.00.00.00.00.0
Ours+DCVC-41.82-39.43-40.60-37.73-41.74-40.26
FVC [14]-5.32-9.75-14.53-1.39-6.20-7.44
Ours+FVC-34.28-31.27-32.09-35.02-32.89-33.11
TCM [29]-25.19-32.34-31.02-39.85-26.44-30.97
Ours+TCM-40.82-45.10-46.15-51.66-38.98-44.54
HEVC [33]-33.88-31.35-40.4315.80-34.02-24.78
", + "image_path": "d4b5da86dbae6e382526bf12c2bf6ba2eade776f61402bc697df2e526973d8c1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 338, + 286, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 286, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 286, + 387 + ], + "type": "text", + "content": "best and third-best performance, respectively. Additionally, we contrast our method with the task-oriented \"One-to-one Codes\" of VCM, with detailed results provided in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 394, + 178, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 178, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 178, + 406 + ], + "type": "text", + "content": "4.2. Video Object Detection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 412, + 286, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 286, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 286, + 556 + ], + "type": "text", + "content": "For the evaluation of the video object detection task, we juxtapose our method with established codecs such as FVC, and TCM, with the comparative results illustrated in Fig. 6(e). Our method is noted to achieve a superior tradeoff, particularly in the Bpp-mAP50 metric, when applied to the downstream model YOLOv. For instance, in comparison to the FVC and TCM baseline codecs, our method facilitates bitrate reductions of approximately " + }, + { + "bbox": [ + 46, + 412, + 286, + 556 + ], + "type": "inline_equation", + "content": "21.76\\%" + }, + { + "bbox": [ + 46, + 412, + 286, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 412, + 286, + 556 + ], + "type": "inline_equation", + "content": "20.42\\%" + }, + { + "bbox": [ + 46, + 412, + 286, + 556 + ], + "type": "text", + "content": ", respectively, while maintaining the same mAP50 metric. The uncompressed video represents the upper bound for the mAP50 metric, recorded at 76.4, and is delineated by the grey line within the figure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 563, + 189, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 189, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 189, + 576 + ], + "type": "text", + "content": "4.3. Video Action Recognition" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": "We conducted an evaluation of our method on the UCF101 dataset, benchmarking against TCM and FVC CODECs. The outcomes of this comparison are depicted in Fig. 6(f). For the task of video action recognition, rather than initiating a new set of network trainings, we employed pre-trained mode prediction networks previously utilized for the MOT task and use the \"DivGoP\" structure for validation. The original videos were compressed using the various CODECs, and the resulting decoded videos served as inputs to the TSM model as cited in Lin et al. [22]. The experimental results indicate that our method maintains efficacy in this" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 426, + 171 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 426, + 171 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 426, + 171 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 426, + 171 + ], + "type": "image", + "image_path": "75f089c755a2ddff914016c4f10bb666a8d8e7096d9a74361da105160f9f9cf1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 427, + 71, + 545, + 171 + ], + "blocks": [ + { + "bbox": [ + 427, + 71, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 427, + 71, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 427, + 71, + 545, + 171 + ], + "type": "image", + "image_path": "cb9e6c73e1c8d49c94adc78b78b8e57d02f69306ef795f1380b24c24853aefc4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 223, + 544, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 223, + 544, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 223, + 544, + 247 + ], + "type": "text", + "content": "context, securing approximately " + }, + { + "bbox": [ + 305, + 223, + 544, + 247 + ], + "type": "inline_equation", + "content": "21.81\\%" + }, + { + "bbox": [ + 305, + 223, + 544, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 223, + 544, + 247 + ], + "type": "inline_equation", + "content": "26.03\\%" + }, + { + "bbox": [ + 305, + 223, + 544, + 247 + ], + "type": "text", + "content": " in bitrate savings for TCM and FVC, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 257, + 468, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 257, + 468, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 257, + 468, + 270 + ], + "type": "text", + "content": "4.4. Video Reconstruction Quality" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 277, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 396 + ], + "type": "text", + "content": "When it comes to the performance of reconstructing the video for human viewing, the situation is straightforward for conventional video codecs, no matter learned or traditional ones: just encode the frames, transmit and decode the compressed bitstream, and measure their quality using appropriate metrics (PSNR, MS-SSIM, etc). In our method, there're two options: Control the encoding procedure for human viewing, which is the ideal way for best frame reconstruction quality, or just continue the encoding procedure used for machine vision tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 397, + 545, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 577 + ], + "type": "text", + "content": "Note that our method does not change the weights of the original encoder and decoder in pre-trained DVC, so using this option can directly recover the human viewing quality to original best performance of the CODECs. On the other hand, we also evaluate the human viewing quality of the encoded bitstream for machine, using \"DivGop\" encoding structure on HEVC Class B and Class C test datasets. These are referred to as \"FVC DivGoP\" and \"TCM DivGoP\" in Fig.7. Results show that our control for machine does decrease the video reconstruction quality of the original TCM codec by about 1dB in PSNR. Besides, it is also observed that the PSNR of residual codec FVC drops more sharply than that of conditional CODEs like TCM. However, these are acceptable since we can always adjust the encoder to compress the video for human viewing when it's needed." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 586, + 399, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 399, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 399, + 599 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Mode Prediction. In our method, we propose mode prediction network to decide the coding modes of motion and residual/contextual feature elements for machine vision, mainly targeting on reducing the bitrate and keep the crucial information for important objects in the frame. Moreover, mainstream DVC, including FVC, DCVC, TCM and others, use motion estimation and warping operation to generate a rough predicted frame, which we call " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "P_r" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " frame here. The " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "P_r" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " frame also meets our requirement of \"consuming" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26042" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 71, + 283, + 178 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 283, + 178 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 283, + 178 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 283, + 178 + ], + "type": "table", + "html": "
Input imageContextual skip mode maskMasked motion featureMasked Contextual feature
", + "image_path": "ab69aa52beebe4eaa050237683d1ccddd4a3c99725541767d4e7a6108e9ad335.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 50, + 189, + 282, + 200 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 189, + 282, + 200 + ], + "spans": [ + { + "bbox": [ + 50, + 189, + 282, + 200 + ], + "type": "text", + "content": "Figure 9. Visualization results of the mode prediction networks." + } + ] + } + ], + "index": 2, + "type": "text" + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "text", + "content": "lower code streams but retaining object and motion information\" to a certain extent. So comparing it with the " + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "text", + "content": " frame used in our method is also a point worthy of attention. In \"DivGoP\" structure, we use the " + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "inline_equation", + "content": "P_{r}" + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "text", + "content": " frame to replace the original " + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "text", + "content": " frames, and the GoP structure will be " + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "inline_equation", + "content": "[I, P_{r}, P, P_{r}, \\ldots]" + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "text", + "content": ". As shown in the left figure in Fig.8, based on DCVC and FVC, we compare the structure using " + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "inline_equation", + "content": "P_{r}" + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "text", + "content": " frames with the original \"DivGoP\" structure using " + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "text", + "content": " frames, corresponding to the \"DCVC/FVC DivGoP w " + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "inline_equation", + "content": "P_{r}" + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "text", + "content": "\" and \"DCVC/FVC DivGoP\" curves respectively. It is observed that the \"DivGoP w " + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "inline_equation", + "content": "P_{r}" + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "text", + "content": "\" curve also achieves about " + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "inline_equation", + "content": "28.68\\%" + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "inline_equation", + "content": "16.24\\%" + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "text", + "content": " bitrate savings for DCVC and FVC, respectively. But our \"DivGoP\" curve shows better trade-off, by achieving " + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "inline_equation", + "content": "32.67\\%" + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "inline_equation", + "content": "19.43\\%" + }, + { + "bbox": [ + 46, + 211, + 287, + 377 + ], + "type": "text", + "content": " bitrate savings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "text", + "content": "GoP Selection. We prove the effectiveness of Gop Selection module in terms of rate-precision performance on MOT dataset. As shown in the right figure in Fig. 8, we compare the Bpp-mAP performance of \"DCVC/FVC Gop Selection\" and \"DCVC/FVC DivGoP\" structure on MOT Dataset. Using DCVC as an example, experiment results show that our Gop Selection achieves about " + }, + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "inline_equation", + "content": "39.43\\%" + }, + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "text", + "content": " bitrate saving, which does show better trade-off than that " + }, + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "inline_equation", + "content": "32.67\\%" + }, + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "text", + "content": " bitrate saving of hand-crafted \"DivGoP\" structure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 486, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 286, + 652 + ], + "type": "text", + "content": "Simply fine-tuning DVC for machine vision tasks. We employed the loss function " + }, + { + "bbox": [ + 46, + 486, + 286, + 652 + ], + "type": "inline_equation", + "content": "L_{f} = R + \\lambda_{1}MSE + \\lambda_{2}L_{det}" + }, + { + "bbox": [ + 46, + 486, + 286, + 652 + ], + "type": "text", + "content": " to fine-tune the FVC encoder and the entire FVC codec for the MOT task. As shown in Fig. 5 right, while simple finetuning improves accuracy, it also notably increases the bitrate. This phenomenon could be attributed to the degradation in the fidelity of decoded frames when optimized for downstream tasks. Due to the inter-frame reference nature of video codecs, this degradation is propagated across frames, resulting in a high bitrate consumption. This observation aligns with findings reported in DeepSVC [21], indicating that such task-specific performance can not be achieved with simply finetuning a DVC. This also reflects the superiority of our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": "Visualization Results. We present the visualization results of our mode prediction network, as shown in Fig. 9. The first column shows the input images, the second column shows the skip mode masks for contextual features (averaged by channel), and the third and fourth columns are" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 72, + 541, + 139 + ], + "blocks": [ + { + "bbox": [ + 344, + 51, + 506, + 61 + ], + "lines": [ + { + "bbox": [ + 344, + 51, + 506, + 61 + ], + "spans": [ + { + "bbox": [ + 344, + 51, + 506, + 61 + ], + "type": "text", + "content": "Table 2. Encoding Latency vs. Performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 72, + 541, + 139 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 541, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 541, + 139 + ], + "type": "table", + "html": "
MethodEncoding Latency(s)BD-BR Performance
DCVC1.6750.0
Ours+DCVC1.738-40.26
FVC0.116-7.44
Ours+FVC0.159-33.11
", + "image_path": "bc35d74ba1c37644a2c7d503bb8a9d818fd6371d99c7820e6ffb7911afc0abba.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 148, + 545, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 545, + 267 + ], + "type": "text", + "content": "masked motion and contextual features, respectively, where the red color indicates large value and green color indicates small value. It is observed that the predicted coding mode mask for contextual features kept the regions of the small objects in the frames, and masked the conspicuous objects which are easy to recognize and the background regions. This shows the network's preference for objects in the frame. As for the motion information, it is observed that the predicted masks kept most of the elements of moving objects, which are crucial for detection and tracking." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 269, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 545, + 424 + ], + "type": "text", + "content": "Complexity and Encoding Latency Analysis. The parameter numbers of our dynamic vision mode prediction network and GoP selection network are 1.48M and 5.36M. Since our method controls the video encoder during the encoding phase, such an operation brings encoding latency. To measure the impact on the encoding phase, we include a comparison of average encoding latency versus performance on MOT Dataset " + }, + { + "bbox": [ + 304, + 269, + 545, + 424 + ], + "type": "inline_equation", + "content": "(960 \\times 512)" + }, + { + "bbox": [ + 304, + 269, + 545, + 424 + ], + "type": "text", + "content": " in Tab. 2. While our method brings an increase in encoding latency (1.738s vs. 1.675s), the improvement in R-P performance is substantial. Besides, the decoding procedure and decoder side models are not changed, which brings convenience to the deployment of the decoder and support for multiple tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 439, + 378, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 378, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 378, + 451 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 460, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 662 + ], + "type": "text", + "content": "In this paper, we propose a flexible framework for deep video compression models, where the pre-trained encoders can be controlled to change the encoding pipeline for machine vision tasks, achieving significant better bppprecision trade-off without changing the decoders or decoding procedure. The controlling methods are attributed to two main developments, including dynamic vision mode prediction network and GoP structure selection network. Experiment results show that our framework is general for residual and conditional deep video CODECs and different downstream vision tasks like detection, tracking and action recognition. Our comprehensive evaluation, conducted against existing deep video CODECs, confirms the superior performance of our models, establishing a new benchmark for future deep video coding for machine. The effectiveness of each component is clearly evidenced through extensive ablation studies." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by the National Natural Science Fund of China (Project No. 42201461) and the General Research Fund (Project No. 16209622) from the Hong Kong Research Grants Council." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26043" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Mohammad Akbari, Jie Liang, and Jingning Han. Dsslic: Deep semantic segmentation-based layered image compression. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2042-2046. IEEE, 2019. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 203 + ], + "type": "text", + "content": "[2] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J. Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 288, + 247 + ], + "type": "text", + "content": "[3] Zhuo Chen, Kui Fan, Shiqi Wang, Lingyu Duan, Weisi Lin, and Alex Chichung Kot. Toward intelligent sensing: Intermediate deep feature compression. IEEE Transactions on Image Processing, 29:2230-2243, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 248, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 248, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 248, + 288, + 282 + ], + "type": "text", + "content": "[4] Hyomin Choi and Ivan V Bajic. Scalable image coding for humans and machines. IEEE Transactions on Image Processing, 31:2739-2754, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 316 + ], + "type": "text", + "content": "[5] MMAction2 Contributors. Openmmlab's next generation video understanding toolbox and benchmark. https://github.com/open-mmlab/mmaction2, 2020.5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 351 + ], + "type": "text", + "content": "[6] Yan Dai, Ziyu Hu, Shuqi Zhang, and Lianjun Liu. A survey of detection-based video multi-object tracking. Displays, page 102317, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 352, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 288, + 385 + ], + "type": "text", + "content": "[7] Ling-Yu Duan, Tiejun Huang, and Wen Gao. Overview of the mpeg cdvs standard. In 2015 Data Compression Conference, pages 323-332, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 386, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 288, + 441 + ], + "type": "text", + "content": "[8] Ling-Yu Duan, Yihang Lou, Yan Bai, Tiejun Huang, Wen Gao, Vijay Chandrasekhar, Jie Lin, Shiqi Wang, and Alex Chichung Kot. Compact descriptors for video analysis: The emerging mpg standard. IEEE MultiMedia, 26(2): 44-54, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 442, + 288, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 288, + 486 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 288, + 486 + ], + "type": "text", + "content": "[9] Yann Dubois, Benjamin Bloem-Reddy, Karen Ullrich, and Chris J Maddison. Lossy compression for lossless prediction. Advances in Neural Information Processing Systems, 34:14014–14028, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 487, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 543 + ], + "type": "text", + "content": "[10] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In European Conference on Computer Vision, pages 510-528. Springer, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 578 + ], + "type": "text", + "content": "[11] Kristian Fischer, Fabian Brand, and André Kaup. Boosting neural image compression for machines using latent space masking. arXiv preprint arXiv:2112.08168, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "type": "text", + "content": "[12] Hadi Hadizadeh and Ivan V Bajić. Learned scalable video coding for humans and machines. arXiv preprint arXiv:2307.08978, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "text", + "content": "[13] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 714 + ], + "type": "text", + "content": "[14] Zhihao Hu, Guo Lu, and Dong Xu. Fvc: A new framework towards deep video compression in feature space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1502-1511, 2021. 1, 2, 4, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[15] Zhihao Hu, Guo Lu, Jinyang Guo, Shan Liu, Wei Jiang, and Dong Xu. Coarse-to-fine deep video coding with hyperprior-guided mode prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5921-5930, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "type": "text", + "content": "[16] Sayed Hossein Khatoonabadi and Ivan V Bajic. Video object tracking in the compressed domain using spatio-temporal markov random fields. IEEE transactions on image processing, 22(1):300-313, 2012. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 239 + ], + "type": "text", + "content": "[17] Congcong Li, Xinyao Wang, Longyin Wen, Dxiang Hong, Tiejian Luo, and Libo Zhang. End-to-end compressed video representation learning for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13967-13976, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 240, + 546, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 546, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 546, + 274 + ], + "type": "text", + "content": "[18] Jiahao Li, Bin Li, and Yan Lu. Deep contextual video compression. Advances in Neural Information Processing Systems, 34:18114-18125, 2021. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 275, + 546, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 546, + 319 + ], + "type": "text", + "content": "[19] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22616-22626, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 319, + 546, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 546, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 546, + 353 + ], + "type": "text", + "content": "[20] Xin Li, Jun Shi, and Zhibo Chen. Task-driven semantic coding via reinforcement learning. IEEE Transactions on Image Processing, 30:6307-6320, 2021. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 354, + 547, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 547, + 409 + ], + "type": "text", + "content": "[21] Hongbin Lin, Bolin Chen, Zhichen Zhang, Jielian Lin, Xu Wang, and Tiesong Zhao. Deep scalable video coding for both machine and human vision. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9205-9214, 2023. 3, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 410, + 547, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 547, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 547, + 454 + ], + "type": "text", + "content": "[22] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7083-7093, 2019. 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 455, + 546, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 546, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 546, + 488 + ], + "type": "text", + "content": "[23] Qiankun Liu, Bin Liu, Yue Wu, Weihai Li, and Nenghai Yu. Real-time online multi-object tracking in compressed domain. arXiv preprint arXiv:2204.02081, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 489, + 546, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 546, + 544 + ], + "type": "text", + "content": "[24] Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, and Zhiyong Gao. Dvc: An end-to-end deep video compression framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11006-11015, 2019. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 545, + 546, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 546, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 546, + 589 + ], + "type": "text", + "content": "[25] Guo Lu, Xiaoyun Zhang, Wanli Ouyang, Li Chen, Zhiyong Gao, and Dong Xu. An end-to-end learning framework for video compression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3292-3308, 2021. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 590, + 546, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 546, + 624 + ], + "type": "text", + "content": "[26] Guo Lu, Xingtong Ge, Tianxiong Zhong, Jing Geng, and Qiang Hu. Preprocessing enhanced image compression for machine vision. arXiv preprint arXiv:2206.05650, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 624, + 546, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 546, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 546, + 679 + ], + "type": "text", + "content": "[27] Fabian Mentzer, Eirikur Agustsson, Johannes Balle, David Minnen, Nick Johnston, and George Toderici. Neural video compression using gans for detail synthesis and propagation. In European Conference on Computer Vision, pages 562-578. Springer, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "type": "text", + "content": "[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26044" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "type": "text", + "content": "scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "text", + "content": "[29] Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. Temporal context mining for learned video compression. IEEE Transactions on Multimedia, 2022. 1, 2, 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[30] Yuheng Shi, Naiyan Wang, and Xiaojie Guo. Yolov: making still image object detectors great at video object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2254-2262, 2023. 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 173, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 288, + 239 + ], + "type": "text", + "content": "[31] Zheng Shou, Xudong Lin, Yannis Kalantidis, Laura Sevilla-Lara, Marcus Rohrbach, Shih-Fu Chang, and Zhicheng Yan. Dmc-net: Generating discriminative motion cues for fast compressed video action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1268-1277, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 272 + ], + "type": "text", + "content": "[32] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 327 + ], + "type": "text", + "content": "[33] Gary J. Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on Circuits and Systems for Video Technology, 22(12):1649-1668, 2012. 1, 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 381 + ], + "type": "text", + "content": "[34] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "type": "text", + "content": "[35] Yuan Tian, Guo Lu, Xiongkuo Min, Zhaohui Che, Guang-tao Zhai, Guodong Guo, and Zhiyong Gao. Self-conditioned probabilistic learning of video rescaling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4490-4499, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "text", + "content": "[36] Yuan Tian, Guo Lu, Guangtao Zhai, and Zhiyong Gao. Nonsementics suppressed mask learning for unsupervised video semantic compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13610-13622, 2023. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "text", + "content": "[37] Yuan Tian, Guo Lu, Yichao Yan, Guangtao Zhai, Li Chen, and Zhiyong Gao. A coding framework and benchmark towards low-bitrate video understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 581 + ], + "type": "text", + "content": "[38] Robert Torfason, Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Towards image understanding from deep compression without decoding. arXiv preprint arXiv:1803.06131, 2018. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 582, + 288, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 288, + 625 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 288, + 625 + ], + "type": "text", + "content": "[39] Shiyao Wang, Hongchao Lu, and Zhidong Deng. Fast object detection in compressed video. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7104-7113, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "text", + "content": "[40] Shurun Wang, Shiqi Wang, Wenhan Yang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. Towards analysis-friendly face representation with scalable feature and texture compression. IEEE Transactions on Multimedia, 2021. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[41] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402. IEEE, 2003. 2, 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 353 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[42] T. Wiegand, G.J. Sullivan, G. Bjontegaard, and A. Luthra. Overview of the h.264/avc video coding standard. IEEE Transactions on Circuits and Systems for Video Technology, 13(7):560-576, 2003. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[43] Chao-Yuan Wu, Manzil Zaheer, Hexiang Hu, R Manmatha, Alexander J Smola, and Philipp Krahenbuhl. Compressed video action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6026-6035, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 217 + ], + "type": "text", + "content": "[44] Ning Yan, Changsheng Gao, Dong Liu, Houqiang Li, Li Li, and Feng Wu. Sssic: semantics-to-signal scalable image coding with learned structural representations. IEEE Transactions on Image Processing, 30:8939-8954, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "text", + "content": "[45] Ren Yang, Luc Van Gool, and Radu Timofte. Perceptual learned video compression with recurrent conditional gan. arXiv preprint arXiv:2109.03082, 1, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 252, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 308 + ], + "type": "text", + "content": "[46] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In European Conference on Computer Vision, pages 1-21. Springer, 2022. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 353 + ], + "type": "text", + "content": "[47] Ezgi Özyilkan, Mateen Ulhaq, Hyomin Choi, and Fabien Racapé. Learned disentangled latent representations for scalable image coding for humans and machines. In 2023 Data Compression Conference (DCC), pages 42-51, 2023. 1, 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26045" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/5ebc706d-8b34-4357-9c79-967e39bda2a5_content_list.json b/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/5ebc706d-8b34-4357-9c79-967e39bda2a5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b818b582efd7d78f9241e41453f12d14c699d78f --- /dev/null +++ b/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/5ebc706d-8b34-4357-9c79-967e39bda2a5_content_list.json @@ -0,0 +1,1400 @@ +[ + { + "type": "text", + "text": "Task-conditioned adaptation of visual features in multi-task policy learning", + "text_level": 1, + "bbox": [ + 107, + 130, + 864, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pierre Marza", + "bbox": [ + 153, + 178, + 272, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Laetitia Matignon", + "bbox": [ + 312, + 180, + 464, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Olivier Simonin1", + "bbox": [ + 504, + 180, + 640, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Christian Wolf3", + "bbox": [ + 683, + 180, + 808, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ INSA Lyon $^{2}$ UCBL $^{3}$ Naver Labs Europe", + "bbox": [ + 323, + 199, + 642, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{pierre.marza, olivier.simonin}@insa-lyon.fr", + "bbox": [ + 292, + 219, + 679, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "laetitia.matignon@univ-lyon1.fr,christian.wolf@naverlabs.com", + "bbox": [ + 217, + 237, + 751, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project Page: https://pierremarza.github.io/projects/task_conditioned_adaptation/", + "bbox": [ + 217, + 253, + 746, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Successfully addressing a wide variety of tasks is a core ability of autonomous agents, requiring flexibly adapting the underlying decision-making strategies and, as we argue in this work, also adapting the perception modules. An analogical argument would be the human visual system, which uses top-down signals to focus attention determined by the current task. Similarly, we adapt pre-trained large vision models conditioned on specific downstream tasks in the context of multi-task policy learning. We introduce task-conditioned adapters that do not require finetuning any pretrained weights, combined with a single policy trained with behavior cloning and capable of addressing multiple tasks. We condition the visual adapters on task embeddings, which can be selected at inference if the task is known, or alternatively inferred from a set of example demonstrations. To this end, we propose a new optimization-based estimator. We evaluate the method on a wide variety of tasks from the CortexBench benchmark and show that, compared to existing work, it can be addressed with a single policy. In particular, we demonstrate that adapting visual features is a key design choice and that the method generalizes to unseen tasks given a few demonstrations.", + "bbox": [ + 75, + 335, + 473, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 694, + 209, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision is one of the most important modalities for agents interacting with the world and is almost indispensable for dexterous manipulation or locomotion, as no other sensor can provide information as rich and versatile. The inherent flexibility of the sensor comes with a high price, the high dimensionality of the information, and the complexity of the processes necessary to extract useful information. Humans and other biological agents are capable of adapting their perception systems to the task at hand. There is indeed evidence for bottom-up and top-down processes in human vision, with the latter guiding attention to regions determined by the requirements of the task [3, 6].", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0db39667673aa3f9b61b4a250567188503d70940baee6b7524540aaa7fac4382.jpg", + "image_caption": [ + "Figure 1. Task-conditioned adaptation: A single policy can be trained to address multiple heterogenous tasks including manipulation, legged motion etc., and few-shot learning is possible to address tasks given as demonstrations but unseen during training. A key element is the task-conditioned adaptation of visual features." + ], + "image_footnote": [], + "bbox": [ + 504, + 303, + 883, + 563 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There is a growing need for a similar versatility in artificial systems, and general neural networks have been trained from large-scale data in different domains such as natural language processing (NLP), computer vision (CV), and, more recently, robotics. A single general vision model coupled with a neural policy would be an appealing choice if it could allow an easy generalization to new domains or tasks. The wide adoption of attention mechanisms in several domains has made it easier for trained models to adapt their behavior to the requirements of different tasks without changing parameters, and it has been shown that attention plays a crucial role in specialization on a specific instance in language models [34] and vision and language models [15]. However, even powerful generally pre-trained models can benefit from parameter adaptations to specific tasks, e.i", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17847", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cba8313c5b16ab1c5e27e215cffd0c7ef51f03e4965562ea1992de7bd8389ae6.jpg", + "image_caption": [ + "Figure 2. Considered tasks: We train the method on a set $T^k$ of known tasks and evaluate it either on the same set, with the task known (Known task setting), or in a Few-shot setting, where a new unseen task from a set $T^u$ is inferred from a few demonstrations." + ], + "image_footnote": [], + "bbox": [ + 81, + 85, + 890, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ther through fine-tuning [12] or by adding additional trained adapter layers to a frozen model [4].", + "bbox": [ + 75, + 321, + 468, + 353 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In robotics, prior work on the generalization capabilities of agents has focused on large-scale end-to-end training [29] or, targeting vision specifically, on pre-trained visual models required to generalize to various different policies [19-21]. In this work, we argue and will show that a single policy can be trained for a large number of different tasks including manipulation, locomotion, and that the adaptation of visual features is highly beneficial, beyond the inherent adaptation capabilities of attention-based models. In particular, different tasks require diverse types of invariances and symmetries. While in principle it should be possible to learn to disentangle a sufficiently wide set of factors of variation in a captured representation such that it optimally performs on a wide variety of tasks, we will show that this is not the case for the arguably dominant pretraining method, masked auto-encoding (MAE) [9].", + "bbox": [ + 75, + 354, + 468, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose task-conditioned adaptation, which allows leveraging the high-quality representations of generally pretrained large vision models, while keeping the required flexibility to address a wide variety of tasks, and also new (unseen) tasks. We introduce a set of task-conditioned visual adapters that can be inserted inside a pre-trained visual Transformer [33]-based backbone. The task is characterized by an embedding space, which is learned from supervision during training. We show that this embedding space captures regularities of tasks and demonstrate this with few-shot capabilities: the single policy and (adapted) visual representations can address new unseen tasks, whose embedding is estimated from a few demonstrations (cf. Figure 1).", + "bbox": [ + 75, + 597, + 468, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this work can be summarized as follows: (i) task-conditioned visual adapters to flexibly modulate visual features to a specific task; (ii) a single multitask policy solving tasks with different embodiments and environments; (iii) a task embedding optimization procedure based on a few demonstrations of a new task (unseen at training time) to adapt the model in a few-shot manner with", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "out any weight fine-tuning; (iv) quantitative and qualitative results assessing the gain brought by the different novelties.", + "bbox": [ + 498, + 321, + 890, + 353 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 371, + 640, + 386 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pre-trained visual representations for robotics — Backbone models pre-trained on large and diverse data have shown great promises in NLP [2, 14, 18, 24], CV [5, 9, 22, 26], and more recently, robotics [16, 19-21, 23, 27, 35, 36]. Parisi et al. [23] study visual pre-training methods for visuomotor control, showing the quality of self-supervised representations. Nair et al. [21] introduce R3M, a general vision model pre-trained on egocentric video data to capture temporal dynamics and semantic features, improving downstream manipulation performance. Radosavovic et al. [27] employ the well-known MAE framework [9] to pre-train a single vision encoder applied to robots in the real world. Ma et al. [19] pre-train a visual model with a self-supervised value function objective on egocentric human videos to improve control policies. Finally, recent work [16, 35, 36] has shown the great promise of pre-trained models, either CLIP-based [16] or self-supervised [35, 36], in visual navigation.", + "bbox": [ + 496, + 398, + 892, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The most related work is [20] which studies the impact of pre-trained vision models on a diversity of tasks gathered in a benchmark named Cortexbench. They introduce a Vision Transformer (ViT) [7] backbone, VC-1, pre-trained from a set of out-of-domain datasets with a focus on egocentric visual frames. We believe that visual features should be task-dependent and study how the VC-1 model can be adapted to improve the performance of a single multi-task policy on a subset of Cortexbench tasks, which is also different from previous work focusing on single-task policies.", + "bbox": [ + 496, + 657, + 892, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformer adapters — Transferring Transformer [33]-based pre-trained models to new tasks or domains is an important topic. Methods involving adapter modules were introduced in NLP [10, 11, 25] to allow a fast and parameter-efficient transfer. Recent works employ the same methods in robotics [17, 31]. Sharma", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17848", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f641e1ddeb7b58befeb79d7c4099749ed52bc10fe1bbd85da85f676fd64b7fa6.jpg", + "image_caption": [ + "(a) Training the policy and adapters." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 467, + 258 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/edb6d490a2fe28cf06f5cc707a835bf3e488d51864155936ffd6091c91adb05b.jpg", + "image_caption": [ + "(b) Optimization of the task embedding in the $\\overline{\\mathrm{Few - shot}}$ setting." + ], + "image_footnote": [], + "bbox": [ + 81, + 290, + 464, + 458 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2735fe26590ee2eb8f3bde4ef465e576c72d1da755a31f49695f2e9200d452cf.jpg", + "image_caption": [ + "(c) Inference for the settings: $\\boxed{\\mathrm{Known~task}}$ and $\\boxed{\\mathrm{Few-shot}}$ .", + "Figure 3. Method overview: (a) the adapted policy is trained with behavior cloning from expert demonstrations and given a visual encoder pre-trained with MAE. The model is conditioned on a task embedding learned from ground-truth 1-in-K task identifiers. (b) In the Few-shot case, a task embedding is estimated by optimization, maximizing the likelihood of a given demonstration of an unknown task. (c) Inference uses a task embedding given in the Known task case, or estimated in the Few-shot case." + ], + "image_footnote": [], + "bbox": [ + 81, + 486, + 464, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "et al. [31] insert visual adapters in a Vision Transformer (ViT) [7] pre-trained model. They introduce different types of adapter blocks (\"bottom\", \"middle\", \"top\") located at diverse places in the visual model and show that combining them improves performance. Liang et al. [17] also show the", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "positive impact of inserting task-specific adapters, trained with imitation learning, in a pre-trained Transformer-based model to adapt to robotics tasks.", + "bbox": [ + 496, + 90, + 890, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While prior work learns a specific set of adapters for each task, we argue that tasks share similarities and explore these regularities with a single set of task-conditioned adapters.", + "bbox": [ + 496, + 136, + 890, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-task robotics policies — Having a single policy performing a wide range of tasks is a long-standing problem in robotics. With Deep Learning-based solutions becoming more popular, some prior work focuses on training multitask neural agents. Approaches like BC-Z [13], RT-1 [1], RT-2 [38] or Gato [30] study the scaling abilities of neural models to large-scale datasets. Trained generalist agents show strong performance on a wide set of tasks, and can generalize to some extent to novel tasks.", + "bbox": [ + 496, + 181, + 892, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, our work leverages pre-trained vision models but does not assume access to a large set of robotics data. Instead, we focus on how to adapt visual features with reasonable computation requirements and train a single multitask policy from only a few expert demonstrations. We also show promising few-shot adaptation to new unknown tasks without requiring very diverse training data. Somewhat related to our work is also TD-MPC2 [8], which introduces a model-based RL algorithm to learn general world models and studies task embeddings to condition a multi-task policy. However, the latter does not act from vision while we specifically study how to modulate visual features conditioned by a task embedding.", + "bbox": [ + 496, + 316, + 893, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Task-conditioned adaptation", + "text_level": 1, + "bbox": [ + 500, + 527, + 764, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All tasks considered in this work are sequential decision-making problems, where at each discrete timestep $t$ an agent receives the last 3 visual frames as an observation $\\mathbf{v}_t \\in \\mathbb{R}^{3 \\times h \\times w \\times 3}$ , where $h$ and $w$ are the height and width of images, and a proprioception input $\\mathbf{p}_t \\in \\mathbb{R}^{d_a}$ , and predicts a continuous action $\\hat{\\mathbf{a}}_t \\in \\mathbb{R}^{d_a}$ , where $d_a$ is the dimension of the action space, which depends on the task at hand. We are provided with a training dataset of expert demonstrations to train a single policy, and for inference we study two different setups: (Known task), where we a priori know the task to be executed, and (Few-shot), where the trained policy must be adapted to a new unseen task without fine-tuning only given a small set of demonstrations.", + "bbox": [ + 496, + 551, + 890, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Known tasks — Following [20], we consider $K = 12$ robotics tasks from 3 benchmarks, Adroit [28], Deepmind control suite [32] and MetaWorld [37]. The set of all known tasks is denoted as $T^{k} = \\{t_{i}^{k}\\}_{[i = 1..K]}$ , where $t_{i}^{k}$ is a 1-in-K vector encoding a known task, and is illustrated in Figure 2.", + "bbox": [ + 496, + 750, + 890, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unknown tasks - The ability of our method to adapt to new skills is evaluated on a set of $U = 15$ tasks from MetaWorld [37], for which we artificially generate demonstrations with a process described in section 4. The set of all unknown tasks is denoted as $T^{u} = \\{t_{i}^{u}\\}_{[i = 1..U]}$ , where $t_i^u$", + "bbox": [ + 496, + 825, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17849", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4565124186cd3888adcdbe4ca6628aee8d234a72351e7dbb0760b040656e1cb4.jpg", + "image_caption": [ + "Figure 4. (Known task) — Qualitative results: Three successful policy rollouts on known tasks from the test set. The multi-task approach performs well on a variety of diverse tasks while being trained on a limited set of demonstrations." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 893, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is a 1-in-U vector encoding an unknown task, and is illustrated in Figure 2. Most importantly, $T^k \\cap T^u = \\emptyset$ .", + "bbox": [ + 75, + 425, + 470, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Base agent architecture", + "text_level": 1, + "bbox": [ + 76, + 467, + 294, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following a large body of work in end-to-end training for robotics, the agent directly maps pixels to actions and decomposes into a visual encoder and a policy.", + "bbox": [ + 75, + 491, + 470, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual encoder without adapters — following [20], the visual encoder, denoted as $\\phi$ , is a ViT model [7] pretrained with masked auto-encoding (MAE). We keep pretrained weights from VC-1 in [20], which are publicly available. However, we change the way the representation is collected from the pre-trained model. Unlike [20], the representation is not taken as the embedding of the 'CLS' token, which we consider to be undertrained by the MAE pretext task. Instead, we train a fully-connected layer $\\psi$ to aggregate all the token representations of the last layer of the ViT except the 'CLS' token. The visual observation $\\mathbf{v}_t$ associated with timestep $t$ is thus encoded as", + "bbox": [ + 75, + 537, + 472, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {t} = \\psi [ \\phi (\\mathbf {v} _ {t}; \\theta_ {\\phi}); \\theta_ {\\psi} ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 732, + 468, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_{\\phi}$ and $\\theta_{\\psi}$ are weights parametrizing $\\phi$ and $\\psi$ respectively. $\\mathbf{r}_t \\in \\mathbb{R}^{3 \\times d_r}$ as it contains the $d_r$ -dim encoding of each of the 3 last visual frames processed as a data batch, where $d_r$ is the output dimension of $\\psi$ .", + "bbox": [ + 75, + 763, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As this will be relevant later, we recall here that a ViT $\\phi$ is composed of a sequence of $N_{l}$ self-attention blocks, where $\\phi_l$ is the layer at index $l$ . If we denote the internal hidden representation predicted at layer $l$ as $\\mathbf{s}_t^l$ , and omit the weights of $\\phi_l$ for simplicity, we have,", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {t} ^ {l} = \\phi_ {l} \\left(\\mathbf {s} _ {t} ^ {l - 1}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 439, + 890, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{s}_t^0 = \\mathbf{v}_t$", + "bbox": [ + 500, + 465, + 596, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Single-task policy — following [20], the policy $\\pi$ is implemented as an MLP predicting actions from the input which is a concatenation of the current frame, two frame differences and the proprioception input $\\mathbf{p_t}$ ,", + "bbox": [ + 498, + 482, + 893, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {a}} _ {t} = \\pi \\left(\\left[ \\mathbf {r} _ {t, 1} - \\mathbf {r} _ {t, 0}, \\mathbf {r} _ {t, 2} - \\mathbf {r} _ {t, 1}, \\mathbf {r} _ {t, 2}, \\mathbf {p} _ {\\mathbf {t}} \\right]; \\theta_ {\\pi}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 554, + 890, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[ ]$ is the concatenation operator and $\\theta_{\\pi}$ are weights parametrizing $\\pi$ .", + "bbox": [ + 498, + 592, + 890, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Adaptation", + "text_level": 1, + "bbox": [ + 500, + 635, + 624, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our key contributions are visual adapter modules along with a multi-task policy, which are all conditioned on the task at hand. This is done with a specific task embedding for each task, taken from an embedding space of dimension $d_{e}$ , which is aimed to have sufficient regularities to enable few-show generalization to unseen tasks. Importantly, the different adapters and the multi-task policy are conditioned on the same task embedding, leading to a common and shared embedding space. For the known task setting, where the ground-truth label of the task is available, the task embedding is projected from a 1-in-K vector with a linear function $g$ trained jointly with the adapters and the policy with the downstream loss (imitation learning). In the Few-shot setting, at test time a new unknown task is described with a few demonstrations, and a task embedding is estimated through optimization, as will be detailed in subsection 3.4.", + "bbox": [ + 496, + 657, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17850", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/2dbc6f2a9060acbe244681c03ec3ddc1897124ffca0e10029a68124388166e05.jpg", + "table_caption": [ + "Table 1. Known task — Impact of visual adapters: Validation and test performance on known tasks of different baselines highlighting the gain brought by adapters. Both middle and top adapters bring a boost in performance, and conditioning them on the learned task embedding increases performance. Our multi-task policy outperforms single-task policies with VC-1 non-adapted features. MT $\\pi$ : multitask policy - NC: Non-conditioned - C: Conditioned - Task emb.: whether to input at evaluation time, either the learned task embedding and chosen from ground-truth (L), a random vector as the task embedding (Rd), or a randomly picked task embedding among the set of $K = 12$ embeddings (RdP) - Benchmarks avg: average performance across the 3 considered benchmarks (Adroit, DMC, MetaWorld) - Tasks avg: average performance across all 12 known tasks. Performance is reported as mean ± std over 3 training runs (seeds)." + ], + "table_footnote": [], + "table_body": "
MT Adapters TaskMulti-task performance
πMid.Top emb.AdroitDMCMetaWorldBenchmarks avgTasks avg
ValTestValTestValTestValTestValTest
(a) [20]---N/A44.0 ± 1.138.3 ± 2.549.6 ± 0.548.0 ± 0.353.5 ± 2.147.8 ± 2.049.1 ± 0.544.7 ± 0.350.3 ± 0.9
(b)✓--L36.3 ± 1.733.0 ± 4.055.3 ± 1.454.1 ± 0.341.7 ± 1.634.7 ± 0.944.4 ± 1.040.6 ± 1.846.5 ± 1.1
(c)✓NC-L40.2 ± 1.237.3 ± 2.854.0 ± 1.554.8 ± 1.945.8 ± 4.536.3 ± 2.546.7 ± 1.642.8 ± 1.948.3 ± 1.9
(d)✓C-L42.0 ± 2.543.8 ± 2.259.1 ± 1.358.8 ± 0.348.6 ± 4.840.8 ± 3.049.9 ± 2.147.8 ± 1.451.9 ± 2.1
(e)✓CNCL44.3 ± 1.243.2 ± 1.560.5 ± 0.560.3 ± 2.558.6 ± 1.648.4 ± 1.954.5 ± 0.750.6 ± 0.857.0 ± 0.7
(f)✓CCL42.0 ± 0.842.3 ± 1.059.9 ± 0.960.0 ± 0.565.3 ± 1.054.5 ± 3.355.8 ± 0.152.3 ± 1.059.2 ± 0.1
(g)✓CCRd4.2 ± 4.01.3 ± 0.910.3 ± 0.78.5 ± 1.11.3 ± 0.90.1 ± 0.15.3 ± 0.83.3 ± 0.25.5 ± 0.1
(h)✓CCRdP0.7 ± 0.93.2 ± 2.55.7 ± 1.29.5 ± 6.10.9 ± 0.60.3 ± 0.42.4 ± 0.54.3 ± 2.22.9 ± 0.4
", + "bbox": [ + 78, + 196, + 890, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 3 outlines the architecture, its details will be given in the supplementary material.", + "bbox": [ + 75, + 410, + 468, + 440 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conditioned on a task embedding we denote as $\\mathbf{e}$ , the proposed adaptations are based on \"middle\" and \"top\" adapters following [10, 31].", + "bbox": [ + 75, + 441, + 468, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Middle adapters — we add one trainable adapter after each ViT block to modulate its output. We introduce a set of middle adapters $A = \\{\\alpha_l\\}_{l=1..N_l}$ , where $\\alpha_l$ is a 2-layer MLP. In the modified visual encoder $\\phi^m$ , each adapter modulates the output of the corresponding self-attention block and is conditioned on the task embedding $\\mathbf{e}$ . Its output is combined with the one of the self-attention layer through a residual connection. If we denote the internal hidden representation predicted at layer $l$ as $\\mathbf{s}_t^{m,l}$ , and omit references to the weights of $\\phi_l$ and $\\alpha_l$ as in eq. (2) for simplicity, the associated forward pass of a given layer becomes,", + "bbox": [ + 75, + 486, + 470, + 652 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {s} _ {t} ^ {m, l} = \\phi_ {l} ^ {m} \\left(\\mathbf {s} _ {t} ^ {m, (l - 1)}\\right) (4) \\\\ = \\phi_ {l} \\left(\\mathbf {s} _ {t} ^ {m, (l - 1)}\\right) + \\alpha_ {l} \\left(\\phi_ {l} \\left(\\mathbf {s} _ {t} ^ {m, (l - 1)}\\right), \\mathbf {e}\\right). (5) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 664, + 468, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Top adapter — A top adapter $\\tau$ , also conditioned on the task at hand, is added after the ViT model, to transform the output of the aggregation layer $\\psi$ to be fed to the multi-task policy (presented below). $\\tau$ has the same architecture as a single middle adapter $\\alpha_{i}$ . The prediction of $\\mathbf{r}_t^m$ , equivalent to $\\mathbf{r}_t$ in the non-adapted case, can be written as,", + "bbox": [ + 75, + 717, + 468, + 808 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {t} ^ {m} = \\tau \\left[ \\psi \\left[ \\phi^ {m} (\\mathbf {v} _ {t}, \\mathbf {e}; \\theta_ {\\phi}, \\theta_ {A}); \\theta_ {\\psi} \\right], \\mathbf {e}; \\theta_ {\\tau} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 819, + 468, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta_{A}$ and $\\theta_{\\tau}$ are the weights parametrizing the middle adapters (each middle adapter has a different set of weights) and the top adapter respectively.", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-task policy — We keep the architecture of the single-task policy in eq. (3), as in [20]. However, instead of re-training a policy for each downstream task of interest, we train a single multi-task policy $\\pi^m$ , whose action space is the union of the action spaces of the different tasks. During training we apply a masking procedure on the output, considering only the actions possible for the task at hand.", + "bbox": [ + 496, + 410, + 890, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let's denote $\\tilde{\\mathbf{r}}_t^m$ as the input to the policy derived from the adapted representation $\\mathbf{r}_t^m$ and the proprioception input $\\mathbf{p_t}$ as done in eq. (3). The conditioning on the task is done by concatenating $\\tilde{\\mathbf{r}}_t^m$ with the task embedding $\\mathbf{e}$ , giving", + "bbox": [ + 496, + 517, + 890, + 578 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {a}} _ {t} = \\pi^ {m} \\left(\\left[ \\tilde {\\mathbf {r}} _ {t} ^ {m}, \\mathbf {e} \\right], \\theta_ {\\pi^ {m}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 590, + 890, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta_{\\pi^m}$ are weights parametrizing $\\pi^m$", + "bbox": [ + 500, + 612, + 779, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training", + "text_level": 1, + "bbox": [ + 500, + 636, + 604, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train the model by keeping the weights of the pre-trained vision-encoder model $\\theta_{\\phi}$ frozen, only the weights of the adapter modules $(\\theta_A,\\theta_\\tau)$ , aggregation layer $(\\theta_{\\psi})$ , embedding layer $(\\theta_g)$ and multi-task policy $(\\theta_{\\pi^m})$ are trained, cf. Figure 3a. Let's denote by $\\Theta = \\{\\theta_A,\\theta_\\tau ,\\theta_\\psi ,\\theta_g,\\theta_{\\pi^m}\\}$ the set of optimized weights. We train with imitation learning, more specifically Behavior Cloning (BC): for each known task $t_i^k$ , we have access to a set of $N_{i}$ expert trajectories that are composed of $T_{i}$ discrete steps, including expert actions. The optimization problem is given as", + "bbox": [ + 496, + 659, + 890, + 810 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\Theta} = \\arg \\min _ {\\Theta} \\sum_ {i = 1} ^ {K} \\sum_ {n = 1} ^ {N _ {i}} \\sum_ {t = 1} ^ {T _ {i}} \\mathcal {L} \\left(\\hat {\\mathbf {a}} _ {t} ^ {i, n}, \\mathbf {a} _ {t} ^ {i, n}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 818, + 890, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{a}}_t^{i,n}$ and $\\mathbf{a}_t^{i,n}$ are the predicted and ground-truth actions for a given step in a trajectory, and $\\mathcal{L}$ is the Mean", + "bbox": [ + 496, + 868, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17851", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/021926cb4c3f4b4f761e69a332cce286f538b40adbb423ac5b644d57c3bb7916.jpg", + "image_caption": [ + "Figure 5. Known task — Per-task performance of policies in Table 1: single-task policies (row (a)), our approach without any adapter (row (b)) and with conditioned middle and top adapters (row (f)). The adapters lead to a performance gain on most tasks, and our multi-task solution is competitive with single-task policies. Colored bars and error bars respectively show mean and std over 3 training runs (seeds)." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 883, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Squared Error loss.", + "bbox": [ + 76, + 385, + 207, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Few-shot adaption to new tasks", + "text_level": 1, + "bbox": [ + 76, + 407, + 352, + 424 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the Few-shot setting, the task embedding $\\mathbf{e}$ is unknown at inference and needs to be estimated from a set of $N_{d}$ example demonstrations $\\mathcal{D} = \\{d_n\\}_{[n = 1..N_d]}$ where $d_{n} = \\{(v_{t}^{n*},p_{t}^{n*},a_{t}^{n*})\\}_{[t = 1..T_{d}]}$ is composed of observations and actions, with $T_{d}$ being the length of each demonstration. We exploit the conditioning property of the policy itself to estimate the embedding $\\hat{\\mathbf{e}}$ as the one which obtains the highest probability of the demonstration actions, when the policy is applied to the demonstration inputs, i.e.", + "bbox": [ + 75, + 430, + 470, + 568 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {e}} = \\arg \\min _ {\\mathbf {e}} \\sum_ {n = 1} ^ {N _ {d}} \\sum_ {t = 1} ^ {T _ {d}} \\mathcal {L} \\left(\\pi^ {m} \\left(\\left[ \\tilde {\\mathbf {r}} _ {t} ^ {m n *}, \\mathbf {e} \\right], \\theta_ {\\pi^ {m}}\\right), \\mathbf {a} _ {t} ^ {n *}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 574, + 468, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{r}}^{mn*}$ is the representation extracted from the demonstration input $(\\mathbf{v}_t^{n*},\\mathbf{p}_t^{n*})$ , and which itself depends on $\\mathbf{e}$ (not made explicit in the notation). The minimization is carried out with SGD from an embedding initialized to zero.", + "bbox": [ + 75, + 621, + 468, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 694, + 209, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training — all variants involving adapters and/or a multitask policy (rows (b)-(f) in Table 1) were trained for 50 epochs with behavior cloning, cf. §3.3, following training hyper-parameters in [20]. Between 20 and 95 expert trajectories are available depending on the task. We used the datasets of trajectories from [20].", + "bbox": [ + 75, + 719, + 468, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation — to better handle possible overfit on hyper-parameter selection, our evaluation setup is slightly different from [20] as we perform 100 validation rollouts to select the best checkpoint of each model, and then test the chosen model on 100 test rollouts. For our multi-task policy, the best checkpoint is the one with the highest average", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "validation performance across all tasks. Single-task policies are validated only on the task they were trained on, giving them an advantage, and for this reason, they are reported as \"soft upper bounds\". We report the average performance and standard deviation among 3 trained models (3 random seeds) for each variant as mean ± std (Table 1).", + "bbox": [ + 496, + 385, + 890, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In total, we conduct evaluations of our method on 27 different tasks, 12 known and 15 unknown, with varying environments, embodiments, and required sub-skills. This allows to evaluate the adaptation and generalization abilities of the multi-task policy.", + "bbox": [ + 496, + 476, + 890, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics — Following [20], we consider a rollout success (1 if the task was completed properly, 0 otherwise) for tasks in the Adroit and MetaWorld benchmarks, and report the normalized return for DMC. Episodes have a maximum length of 1000 steps and each step reward is comprised between 0 and 100 in DMC, normalization is therefore done by dividing the agent's return by 10. For all tasks, performance is averaged across rollouts.", + "bbox": [ + 496, + 551, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Known task — Impact of visual adapters — Table 1 presents a detailed comparison of different methods on the known task setting. The baseline in row (a) follows the setup in [20] to train single-task policies (one per task) from non-conditioned VC-1 features. For this variant, we use the representation of the 'CLS' token as the vector fed to the policy as done in [20], while all other baselines use our proposed token aggregation layer.", + "bbox": [ + 496, + 672, + 890, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Row (b) is our multi-task policy without any adapter. As expected there is a performance drop compared to the specialized policies in row (a), as the problem to solve has become more difficult. Adding adapters and conditioning them on the task embedding, shown in rows (c)-(f), brings a boost in performance, both for middle and top adapters. In particular, conditioning adds a further boost compared to", + "bbox": [ + 496, + 795, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17852", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1e78ac63be7b007999df7e11223b10b164079f4d191bbea07c9579f7b4617fa6.jpg", + "image_caption": [ + "Figure 6. Few-shot — Per-task performance: After optimizing the task embedding for each task from 5 demonstrations, our method can adapt to many of them (without finetuning). Colored bars and error bars respectively show mean and std over 3 training runs (seeds)." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 890, + 310 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "non-conditioned adapters, with all choices enabled, row (f), obtaining the best average performance.", + "bbox": [ + 75, + 376, + 468, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rows (g) and (h) are ablation experiments evaluating the impact of choosing random task embeddings, row (g), or of taking a random choice between the 12 learned embeddings, row (h). In both cases, the performance collapses.", + "bbox": [ + 75, + 409, + 470, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A particularly important conclusion that can be drawn from the experiments outlined in Table 1 is that the proposed multi-task approach (row (f)) outperforms the single-task policies without adapters (row (a)). This shows that a multi-task policy can perform well on a series of tasks while being trained on a limited set of demonstrations. Figure 4 presents 3 successful test rollouts of our multi-task approach on diverse known tasks.", + "bbox": [ + 75, + 470, + 468, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5 visualizes the per-task test performance on known tasks of single-task policies (row (a) in Table 1), our approach without any adapter (row (b) in Table 1) and with conditioned middle and top adapters (row (f) in Table 1). The proposed adapters lead to a performance gain on most tasks compared with the solution without adapters, and the multi-task solution is competitive with single-task policies, even outperforming them on half the tasks.", + "bbox": [ + 75, + 594, + 468, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Known task Additional ablation studies are presented in the supplementary material (Sections B and C). Section B shows that conditioning the policy when using task-conditioned adapters is not necessary, that our aggregation layer works better than using the 'CLS' token as input to the policy and further highlights the impact of both middle and top adapters. Section C shows that our adapters improve visual embeddings extracted by two state-of-the-art pre-trained backbones, i.e. PVR [23] and MVP [27], confirming the conclusions drawn from experiments on VC-1.", + "bbox": [ + 75, + 717, + 468, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Known task — Visualizing the influence of task-conditioned adaptation — Figures 2 and 3 in the sup-", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "plementary material visualize the attention map of the last VC-1 layer. They show that adding adapters focuses attention on objects of interest and task-conditioning improves attention, mainly towards goals.", + "bbox": [ + 496, + 376, + 890, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Few-shot adaptation to new tasks without finetuning model weights is an important ability of any general policy, which we evaluate with the following experiments: for each task within a set of unknown tasks (cf. Figure 2), we collect only 5 demonstrations used to optimize a task embedding specific to this task with the method detailed in section §3.4. We then evaluate the method conditioned on the optimized embedding on 100 test rollouts.", + "bbox": [ + 496, + 439, + 892, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To generate the set $T^u$ of unknown tasks, we select tasks from the MetaWorld dataset that do not belong to CortexBench, and are thus not part of the set of training known tasks $T^k$ . We collect demonstrations using single-task policies from TD-MPC2 [8] that were specifically trained on each task of MetaWorld independently. To ensure high-quality demonstrations, we only consider tasks where TD-MPC2 policies reach a success rate higher than $95\\%$ . Furthermore, to be compatible with the setup in CortexBench authors, in particular, to keep the same camera locations, we filter out the tasks where the goal is not always visible in the camera FOV. This leads to a set of 15 unknown tasks that are quite different from the tasks in the training set $T^k$ as they involve different objects (handle press, faucet, plate, door, window, etc.) and types of manipulation (sliding an object, lowering a press, opening a window, etc.). Each collected demonstration is a sequence of visual frames, proprioception inputs, and expert actions. The optimization of the task embedding is performed independently for each task (cf. §3.4). We use the AdamW optimizer and a learning rate of $1e - 1$ during the task embedding search.", + "bbox": [ + 496, + 564, + 893, + 882 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 6 presents the per-task performance in this set", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17853", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7750e91419ec948d0b65300110badacf47393c40682a25a32738fd99c81c80ae.jpg", + "image_caption": [ + "(a) Successful rollouts" + ], + "image_footnote": [], + "bbox": [ + 120, + 89, + 851, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a2493e3a4364821fc20776ca06aa3ddbed9d05f6c3b5bb273bec996ecd85ef8.jpg", + "image_caption": [ + "(b) Failed rollouts", + "Figure 7. Few-shot Qualitative results: (a) The policy tackles new tasks involving objects and/or manipulation requirements unseen during training. (b) In the first row (button-press-wall task), it performs the task correctly until the end where it fails to properly push the button fully. In the second row (push-back task), it properly moves the cube but fails to bring it to the goal position (green dot)." + ], + "image_footnote": [], + "bbox": [ + 120, + 273, + 849, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ting. Despite the large variations between the new tasks in $T^u$ and the ones in the training set $(T^k)$ , the multi-task policy can adapt to many of them, without requiring any weight finetuning. Interestingly, the method performs particularly well on the **Drawer Close** task, which could be related to the presence of the time-inversed **Drawer Open** task in the training set. This provides some evidence that the method can exploit regularities between tasks, which seem to be captured by the task embedding space, making it possible to generalize to unseen variations. Figure 7 (a) shows qualitative examples of successful rollouts on the new tasks. The policy manipulates new objects (faucet, plate, window) and performs new moves (rotating the faucet or sliding the plate) not seen during training.", + "bbox": [ + 75, + 535, + 472, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, Figure 7 (b) shows failure cases on unknown tasks. As seen on the first row, the policy avoids the wall, reaches the button, and starts pushing it, but fails to push it fully. This particular behavior was also observed on other rollouts, explaining the $0\\%$ success on this button-push-wall task while mastering a part of the required sub-skills. On the second row, the policy is able to move the cube but fails to bring it to the goal location (green dot). This gives some indication of the difficulty of the few-shot generalization case: exploiting regularities in the task space requires that", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tasks be performed more than just approximately, as often the success metric is sparse, and rollouts only count to the metric when they are executed fully and correctly.", + "bbox": [ + 498, + 535, + 890, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 609, + 619, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Perception and action are closely tied together, and studies of human cognition have shown that a priori knowledge about a downstream task guides the visual system. We follow this direction in the context of artificial agents by introducing task-conditioned adapters that modulate the visual features of a pre-trained neural backbone. Such adapters, conditioned on a learned task embedding, improve the performance of a multi-task policy across benchmarks and embodiments. Even more interesting is the use of task embeddings to adapt in a few-shot manner, i.e. from a small set of demonstrations, to new tasks unseen at training time. We propose an optimization procedure to estimate a new task embedding and achieve generalization to unseen tasks, involving new objects and manipulation sub-skills, providing evidence for regularities in the learned embedding space.", + "bbox": [ + 496, + 641, + 890, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement - We thank ANR for support through AI-chair grant \"Remember\" (ANR-20-CHIA-0018).", + "bbox": [ + 498, + 871, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17854", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv, 2022. 3", + "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020. 2", + "[3] T. J. Buschman and E. K. Miller. Top-down versus bottom-up control of attention in the prefrontal and posterior parietal cortices. science, 2007. 1", + "[4] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In NeurIPS, 2022. 2", + "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2", + "[6] M. Corbetta and G. L. Shulman. Control of goal-directed and stimulus-driven attention in the brain. Nature Reviews Neuroscience, 2002. 1", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 2, 3, 4", + "[8] Nicklas Hansen, Hao Su, and Xiaolong Wang. Td-mpc2: Scalable, robust world models for continuous control. arXiv, 2023. 3, 7", + "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2", + "[10] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 2, 5", + "[11] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2021. 2", + "[12] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2", + "[13] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In CoRL, 2022. 3", + "[14] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019. 2", + "[15] C. Kervadec, T. Jaunet, G. Antipov, M. Baccouche, R. Vuillemot, and C. Wolf. How Transferrable are Reasoning Patterns in VQA? In CVPR, 2021. 1" + ], + "bbox": [ + 76, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: CLIP embeddings for embodied AI. In CVPR, 2022. 2", + "[17] Anthony Liang, Ishika Singh, Karl Pertsch, and Jesse Thomason. Transformer adapters for robot learning. In CoRL Workshop on Pre-training Robot Learning, 2022. 2, 3", + "[18] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized BERT pretraining approach. arXiv, 2019. 2", + "[19] Yecheng Jason Ma, Shagun Sodhani, Dinesh Jayaraman, Osbert Bastani, Vikash Kumar, and Amy Zhang. Vip: Towards universal visual reward and representation via value-implicit pre-training. In ICLR, 2022. 2", + "[20] Arjun Majumdar, Karmesh Yadav, Sergio Arnaud, Yecheng Jason Ma, Claire Chen, Sneha Silwal, Aryan Jain, Vincent-Pierre Berges, Pieter Abbeel, Jitendra Malik, et al. Where are we in the search for an artificial visual cortex for embodied intelligence? arXiv, 2023. 2, 3, 4, 5, 6", + "[21] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. In CoRL, 2023. 2", + "[22] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv, 2023. 2", + "[23] Simone Parisi, Aravind Rajeswaran, Senthil Purushwalkam, and Abhinav Gupta. The unsurprising effectiveness of pretrained vision models for control. In ICML, 2022. 2, 7", + "[24] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In NAACL-HLT, 2018. 2", + "[25] Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. Adapterfusion: Nondestructive task composition for transfer learning. In EACL, 2021. 2", + "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2", + "[27] Ilija Radosavovic, Tete Xiao, Stephen James, Pieter Abbeel, Jitendra Malik, and Trevor Darrell. Real-world robot learning with masked visual pre-training. In CoRL, 2023. 2, 7", + "[28] Aravind Rajeswaran, Vikash Kumar, Abhishek Gupta, Giulia Vezzani, John Schulman, Emanuel Todorov, and Sergey Levine. Learning complex dexterous manipulation with deep reinforcement learning and demonstrations. RSS, 2018. 3", + "[29] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, Tom Eccles, Jake Bruce, Ali Razavi, Ashley Edwards, Nicolas Heess, Yutian Chen, Raia Hadsell, Oriol Vinyals, Mahyar Bordbar, and Nando de Freitas. A Generalist Agent. TMLR, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17855", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv, 2022. 3", + "[31] Mohit Sharma, Claudio Fantacci, Yuxiang Zhou, Skanda Koppula, Nicolas Heess, Jon Scholz, and Yusuf Aytar. Lossless adaptation of pretrained vision models for robotic manipulation. In ICLR, 2022. 2, 3, 5", + "[32] Yuval Tassa, Yotam Doron, Alistair Muldal, Tom Erez, Yazhe Li, Diego de Las Casas, David Budden, Abbas Abdelmaleki, Josh Merel, Andrew Lefrancq, et al. Deepmind control suite. arXiv, 2018. 3", + "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 2", + "[34] Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 2019. 1", + "[35] Karmesh Yadav, Ram Ramrakhya, Arjun Majumdar, Vincent-Pierre Berges, Sachit Kuhar, Dhruv Batra, Alexei Baevski, and Oleksandr Maksymets. Offline visual representation learning for embodied navigation. arXiv, 2022. 2", + "[36] Karmesh Yadav, Arjun Majumdar, Ram Ramrakhya, Naoki Yokoyama, Alexei Baevski, Zsolt Kira, Oleksandr Maksymets, and Dhruv Batra. Ovrl-v2: A simple state-of-art baseline forImagenav and objectnav. arXiv, 2023. 2", + "[37] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In CoRL, 2020. 3", + "[38] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023. 3" + ], + "bbox": [ + 78, + 90, + 470, + 599 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17856", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/5ebc706d-8b34-4357-9c79-967e39bda2a5_model.json b/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/5ebc706d-8b34-4357-9c79-967e39bda2a5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..994eb2b06c99ba2edd31a7b86bd4c1da1b20a793 --- /dev/null +++ b/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/5ebc706d-8b34-4357-9c79-967e39bda2a5_model.json @@ -0,0 +1,1870 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.131, + 0.865, + 0.155 + ], + "angle": 0, + "content": "Task-conditioned adaptation of visual features in multi-task policy learning" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.179, + 0.273, + 0.198 + ], + "angle": 0, + "content": "Pierre Marza" + }, + { + "type": "text", + "bbox": [ + 0.313, + 0.181, + 0.465, + 0.199 + ], + "angle": 0, + "content": "Laetitia Matignon" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.181, + 0.641, + 0.198 + ], + "angle": 0, + "content": "Olivier Simonin1" + }, + { + "type": "text", + "bbox": [ + 0.684, + 0.181, + 0.809, + 0.198 + ], + "angle": 0, + "content": "Christian Wolf3" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.2, + 0.643, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)INSA Lyon \\(^{2}\\)UCBL \\(^{3}\\)Naver Labs Europe" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.22, + 0.68, + 0.234 + ], + "angle": 0, + "content": "{pierre.marza, olivier.simonin}@insa-lyon.fr" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.238, + 0.752, + 0.251 + ], + "angle": 0, + "content": "laetitia.matignon@univ-lyon1.fr,christian.wolf@naverlabs.com" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.254, + 0.748, + 0.27 + ], + "angle": 0, + "content": "Project Page: https://pierremarza.github.io/projects/task_conditioned_adaptation/" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.474, + 0.669 + ], + "angle": 0, + "content": "Successfully addressing a wide variety of tasks is a core ability of autonomous agents, requiring flexibly adapting the underlying decision-making strategies and, as we argue in this work, also adapting the perception modules. An analogical argument would be the human visual system, which uses top-down signals to focus attention determined by the current task. Similarly, we adapt pre-trained large vision models conditioned on specific downstream tasks in the context of multi-task policy learning. We introduce task-conditioned adapters that do not require finetuning any pretrained weights, combined with a single policy trained with behavior cloning and capable of addressing multiple tasks. We condition the visual adapters on task embeddings, which can be selected at inference if the task is known, or alternatively inferred from a set of example demonstrations. To this end, we propose a new optimization-based estimator. We evaluate the method on a wide variety of tasks from the CortexBench benchmark and show that, compared to existing work, it can be addressed with a single policy. In particular, we demonstrate that adapting visual features is a key design choice and that the method generalizes to unseen tasks given a few demonstrations." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.695, + 0.21, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Vision is one of the most important modalities for agents interacting with the world and is almost indispensable for dexterous manipulation or locomotion, as no other sensor can provide information as rich and versatile. The inherent flexibility of the sensor comes with a high price, the high dimensionality of the information, and the complexity of the processes necessary to extract useful information. Humans and other biological agents are capable of adapting their perception systems to the task at hand. There is indeed evidence for bottom-up and top-down processes in human vision, with the latter guiding attention to regions determined by the requirements of the task [3, 6]." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.304, + 0.885, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.575, + 0.894, + 0.646 + ], + "angle": 0, + "content": "Figure 1. Task-conditioned adaptation: A single policy can be trained to address multiple heterogenous tasks including manipulation, legged motion etc., and few-shot learning is possible to address tasks given as demonstrations but unseen during training. A key element is the task-conditioned adaptation of visual features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "There is a growing need for a similar versatility in artificial systems, and general neural networks have been trained from large-scale data in different domains such as natural language processing (NLP), computer vision (CV), and, more recently, robotics. A single general vision model coupled with a neural policy would be an appealing choice if it could allow an easy generalization to new domains or tasks. The wide adoption of attention mechanisms in several domains has made it easier for trained models to adapt their behavior to the requirements of different tasks without changing parameters, and it has been shown that attention plays a crucial role in specialization on a specific instance in language models [34] and vision and language models [15]. However, even powerful generally pre-trained models can benefit from parameter adaptations to specific tasks, e.i" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17847" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.086, + 0.891, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.268, + 0.893, + 0.297 + ], + "angle": 0, + "content": "Figure 2. Considered tasks: We train the method on a set \\( T^k \\) of known tasks and evaluate it either on the same set, with the task known (Known task setting), or in a Few-shot setting, where a new unseen task from a set \\( T^u \\) is inferred from a few demonstrations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.47, + 0.354 + ], + "angle": 0, + "content": "ther through fine-tuning [12] or by adding additional trained adapter layers to a frozen model [4]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.47, + 0.597 + ], + "angle": 0, + "content": "In robotics, prior work on the generalization capabilities of agents has focused on large-scale end-to-end training [29] or, targeting vision specifically, on pre-trained visual models required to generalize to various different policies [19-21]. In this work, we argue and will show that a single policy can be trained for a large number of different tasks including manipulation, locomotion, and that the adaptation of visual features is highly beneficial, beyond the inherent adaptation capabilities of attention-based models. In particular, different tasks require diverse types of invariances and symmetries. While in principle it should be possible to learn to disentangle a sufficiently wide set of factors of variation in a captured representation such that it optimally performs on a wide variety of tasks, we will show that this is not the case for the arguably dominant pretraining method, masked auto-encoding (MAE) [9]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.794 + ], + "angle": 0, + "content": "We propose task-conditioned adaptation, which allows leveraging the high-quality representations of generally pretrained large vision models, while keeping the required flexibility to address a wide variety of tasks, and also new (unseen) tasks. We introduce a set of task-conditioned visual adapters that can be inserted inside a pre-trained visual Transformer [33]-based backbone. The task is characterized by an embedding space, which is learned from supervision during training. We show that this embedding space captures regularities of tasks and demonstrate this with few-shot capabilities: the single policy and (adapted) visual representations can address new unseen tasks, whose embedding is estimated from a few demonstrations (cf. Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The contributions of this work can be summarized as follows: (i) task-conditioned visual adapters to flexibly modulate visual features to a specific task; (ii) a single multitask policy solving tasks with different embodiments and environments; (iii) a task embedding optimization procedure based on a few demonstrations of a new task (unseen at training time) to adapt the model in a few-shot manner with" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.323, + 0.892, + 0.354 + ], + "angle": 0, + "content": "out any weight fine-tuning; (iv) quantitative and qualitative results assessing the gain brought by the different novelties." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.372, + 0.642, + 0.387 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.893, + 0.656 + ], + "angle": 0, + "content": "Pre-trained visual representations for robotics — Backbone models pre-trained on large and diverse data have shown great promises in NLP [2, 14, 18, 24], CV [5, 9, 22, 26], and more recently, robotics [16, 19-21, 23, 27, 35, 36]. Parisi et al. [23] study visual pre-training methods for visuomotor control, showing the quality of self-supervised representations. Nair et al. [21] introduce R3M, a general vision model pre-trained on egocentric video data to capture temporal dynamics and semantic features, improving downstream manipulation performance. Radosavovic et al. [27] employ the well-known MAE framework [9] to pre-train a single vision encoder applied to robots in the real world. Ma et al. [19] pre-train a visual model with a self-supervised value function objective on egocentric human videos to improve control policies. Finally, recent work [16, 35, 36] has shown the great promise of pre-trained models, either CLIP-based [16] or self-supervised [35, 36], in visual navigation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.893, + 0.809 + ], + "angle": 0, + "content": "The most related work is [20] which studies the impact of pre-trained vision models on a diversity of tasks gathered in a benchmark named Cortexbench. They introduce a Vision Transformer (ViT) [7] backbone, VC-1, pre-trained from a set of out-of-domain datasets with a focus on egocentric visual frames. We believe that visual features should be task-dependent and study how the VC-1 model can be adapted to improve the performance of a single multi-task policy on a subset of Cortexbench tasks, which is also different from previous work focusing on single-task policies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Transformer adapters — Transferring Transformer [33]-based pre-trained models to new tasks or domains is an important topic. Methods involving adapter modules were introduced in NLP [10, 11, 25] to allow a fast and parameter-efficient transfer. Recent works employ the same methods in robotics [17, 31]. Sharma" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17848" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.468, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.265, + 0.382, + 0.28 + ], + "angle": 0, + "content": "(a) Training the policy and adapters." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.291, + 0.465, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.464, + 0.466, + 0.479 + ], + "angle": 0, + "content": "(b) Optimization of the task embedding in the \\(\\overline{\\mathrm{Few - shot}}\\) setting." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.487, + 0.465, + 0.654 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.657, + 0.446, + 0.672 + ], + "angle": 0, + "content": "(c) Inference for the settings: \\(\\boxed{\\mathrm{Known~task}}\\) and \\(\\boxed{\\mathrm{Few-shot}}\\)." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Figure 3. Method overview: (a) the adapted policy is trained with behavior cloning from expert demonstrations and given a visual encoder pre-trained with MAE. The model is conditioned on a task embedding learned from ground-truth 1-in-K task identifiers. (b) In the Few-shot case, a task embedding is estimated by optimization, maximizing the likelihood of a given demonstration of an unknown task. (c) Inference uses a task embedding given in the Known task case, or estimated in the Few-shot case." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "et al. [31] insert visual adapters in a Vision Transformer (ViT) [7] pre-trained model. They introduce different types of adapter blocks (\"bottom\", \"middle\", \"top\") located at diverse places in the visual model and show that combining them improves performance. Liang et al. [17] also show the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "positive impact of inserting task-specific adapters, trained with imitation learning, in a pre-trained Transformer-based model to adapt to robotics tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.182 + ], + "angle": 0, + "content": "While prior work learns a specific set of adapters for each task, we argue that tasks share similarities and explore these regularities with a single set of task-conditioned adapters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.893, + 0.317 + ], + "angle": 0, + "content": "Multi-task robotics policies — Having a single policy performing a wide range of tasks is a long-standing problem in robotics. With Deep Learning-based solutions becoming more popular, some prior work focuses on training multitask neural agents. Approaches like BC-Z [13], RT-1 [1], RT-2 [38] or Gato [30] study the scaling abilities of neural models to large-scale datasets. Trained generalist agents show strong performance on a wide set of tasks, and can generalize to some extent to novel tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.895, + 0.515 + ], + "angle": 0, + "content": "In contrast, our work leverages pre-trained vision models but does not assume access to a large set of robotics data. Instead, we focus on how to adapt visual features with reasonable computation requirements and train a single multitask policy from only a few expert demonstrations. We also show promising few-shot adaptation to new unknown tasks without requiring very diverse training data. Somewhat related to our work is also TD-MPC2 [8], which introduces a model-based RL algorithm to learn general world models and studies task embeddings to condition a multi-task policy. However, the latter does not act from vision while we specifically study how to modulate visual features conditioned by a task embedding." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.528, + 0.766, + 0.546 + ], + "angle": 0, + "content": "3. Task-conditioned adaptation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.749 + ], + "angle": 0, + "content": "All tasks considered in this work are sequential decision-making problems, where at each discrete timestep \\( t \\) an agent receives the last 3 visual frames as an observation \\( \\mathbf{v}_t \\in \\mathbb{R}^{3 \\times h \\times w \\times 3} \\), where \\( h \\) and \\( w \\) are the height and width of images, and a proprioception input \\( \\mathbf{p}_t \\in \\mathbb{R}^{d_a} \\), and predicts a continuous action \\( \\hat{\\mathbf{a}}_t \\in \\mathbb{R}^{d_a} \\), where \\( d_a \\) is the dimension of the action space, which depends on the task at hand. We are provided with a training dataset of expert demonstrations to train a single policy, and for inference we study two different setups: (Known task), where we a priori know the task to be executed, and (Few-shot), where the trained policy must be adapted to a new unseen task without fine-tuning only given a small set of demonstrations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.826 + ], + "angle": 0, + "content": "Known tasks — Following [20], we consider \\( K = 12 \\) robotics tasks from 3 benchmarks, Adroit [28], Deepmind control suite [32] and MetaWorld [37]. The set of all known tasks is denoted as \\( T^{k} = \\{t_{i}^{k}\\}_{[i = 1..K]} \\), where \\( t_{i}^{k} \\) is a 1-in-K vector encoding a known task, and is illustrated in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Unknown tasks - The ability of our method to adapt to new skills is evaluated on a set of \\( U = 15 \\) tasks from MetaWorld [37], for which we artificially generate demonstrations with a process described in section 4. The set of all unknown tasks is denoted as \\( T^{u} = \\{t_{i}^{u}\\}_{[i = 1..U]} \\), where \\( t_i^u \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17849" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.895, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.37, + 0.895, + 0.4 + ], + "angle": 0, + "content": "Figure 4. (Known task) — Qualitative results: Three successful policy rollouts on known tasks from the test set. The multi-task approach performs well on a variety of diverse tasks while being trained on a limited set of demonstrations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.471, + 0.456 + ], + "angle": 0, + "content": "is a 1-in-U vector encoding an unknown task, and is illustrated in Figure 2. Most importantly, \\( T^k \\cap T^u = \\emptyset \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.468, + 0.295, + 0.485 + ], + "angle": 0, + "content": "3.1. Base agent architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.471, + 0.539 + ], + "angle": 0, + "content": "Following a large body of work in end-to-end training for robotics, the agent directly maps pixels to actions and decomposes into a visual encoder and a policy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.473, + 0.721 + ], + "angle": 0, + "content": "Visual encoder without adapters — following [20], the visual encoder, denoted as \\(\\phi\\), is a ViT model [7] pretrained with masked auto-encoding (MAE). We keep pretrained weights from VC-1 in [20], which are publicly available. However, we change the way the representation is collected from the pre-trained model. Unlike [20], the representation is not taken as the embedding of the 'CLS' token, which we consider to be undertrained by the MAE pretext task. Instead, we train a fully-connected layer \\(\\psi\\) to aggregate all the token representations of the last layer of the ViT except the 'CLS' token. The visual observation \\(\\mathbf{v}_t\\) associated with timestep \\(t\\) is thus encoded as" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.733, + 0.47, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {t} = \\psi [ \\phi (\\mathbf {v} _ {t}; \\theta_ {\\phi}); \\theta_ {\\psi} ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.47, + 0.825 + ], + "angle": 0, + "content": "where \\(\\theta_{\\phi}\\) and \\(\\theta_{\\psi}\\) are weights parametrizing \\(\\phi\\) and \\(\\psi\\) respectively. \\(\\mathbf{r}_t \\in \\mathbb{R}^{3 \\times d_r}\\) as it contains the \\(d_r\\)-dim encoding of each of the 3 last visual frames processed as a data batch, where \\(d_r\\) is the output dimension of \\(\\psi\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "As this will be relevant later, we recall here that a ViT \\(\\phi\\) is composed of a sequence of \\(N_{l}\\) self-attention blocks, where \\(\\phi_l\\) is the layer at index \\(l\\). If we denote the internal hidden representation predicted at layer \\(l\\) as \\(\\mathbf{s}_t^l\\), and omit the weights of \\(\\phi_l\\) for simplicity, we have," + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.44, + 0.892, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {t} ^ {l} = \\phi_ {l} \\left(\\mathbf {s} _ {t} ^ {l - 1}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.466, + 0.597, + 0.483 + ], + "angle": 0, + "content": "where \\(\\mathbf{s}_t^0 = \\mathbf{v}_t\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.483, + 0.894, + 0.545 + ], + "angle": 0, + "content": "Single-task policy — following [20], the policy \\(\\pi\\) is implemented as an MLP predicting actions from the input which is a concatenation of the current frame, two frame differences and the proprioception input \\(\\mathbf{p_t}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.555, + 0.892, + 0.583 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {a}} _ {t} = \\pi \\left(\\left[ \\mathbf {r} _ {t, 1} - \\mathbf {r} _ {t, 0}, \\mathbf {r} _ {t, 2} - \\mathbf {r} _ {t, 1}, \\mathbf {r} _ {t, 2}, \\mathbf {p} _ {\\mathbf {t}} \\right]; \\theta_ {\\pi}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.593, + 0.892, + 0.626 + ], + "angle": 0, + "content": "where \\([ ]\\) is the concatenation operator and \\(\\theta_{\\pi}\\) are weights parametrizing \\(\\pi\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.625, + 0.652 + ], + "angle": 0, + "content": "3.2. Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Our key contributions are visual adapter modules along with a multi-task policy, which are all conditioned on the task at hand. This is done with a specific task embedding for each task, taken from an embedding space of dimension \\( d_{e} \\), which is aimed to have sufficient regularities to enable few-show generalization to unseen tasks. Importantly, the different adapters and the multi-task policy are conditioned on the same task embedding, leading to a common and shared embedding space. For the known task setting, where the ground-truth label of the task is available, the task embedding is projected from a 1-in-K vector with a linear function \\( g \\) trained jointly with the adapters and the policy with the downstream loss (imitation learning). In the Few-shot setting, at test time a new unknown task is described with a few demonstrations, and a task embedding is estimated through optimization, as will be detailed in subsection 3.4." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17850" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.189 + ], + "angle": 0, + "content": "Table 1. Known task — Impact of visual adapters: Validation and test performance on known tasks of different baselines highlighting the gain brought by adapters. Both middle and top adapters bring a boost in performance, and conditioning them on the learned task embedding increases performance. Our multi-task policy outperforms single-task policies with VC-1 non-adapted features. MT \\(\\pi\\): multitask policy - NC: Non-conditioned - C: Conditioned - Task emb.: whether to input at evaluation time, either the learned task embedding and chosen from ground-truth (L), a random vector as the task embedding (Rd), or a randomly picked task embedding among the set of \\(K = 12\\) embeddings (RdP) - Benchmarks avg: average performance across the 3 considered benchmarks (Adroit, DMC, MetaWorld) - Tasks avg: average performance across all 12 known tasks. Performance is reported as mean ± std over 3 training runs (seeds)." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.198, + 0.891, + 0.389 + ], + "angle": 0, + "content": "
MT Adapters TaskMulti-task performance
πMid.Top emb.AdroitDMCMetaWorldBenchmarks avgTasks avg
ValTestValTestValTestValTestValTest
(a) [20]---N/A44.0 ± 1.138.3 ± 2.549.6 ± 0.548.0 ± 0.353.5 ± 2.147.8 ± 2.049.1 ± 0.544.7 ± 0.350.3 ± 0.9
(b)✓--L36.3 ± 1.733.0 ± 4.055.3 ± 1.454.1 ± 0.341.7 ± 1.634.7 ± 0.944.4 ± 1.040.6 ± 1.846.5 ± 1.1
(c)✓NC-L40.2 ± 1.237.3 ± 2.854.0 ± 1.554.8 ± 1.945.8 ± 4.536.3 ± 2.546.7 ± 1.642.8 ± 1.948.3 ± 1.9
(d)✓C-L42.0 ± 2.543.8 ± 2.259.1 ± 1.358.8 ± 0.348.6 ± 4.840.8 ± 3.049.9 ± 2.147.8 ± 1.451.9 ± 2.1
(e)✓CNCL44.3 ± 1.243.2 ± 1.560.5 ± 0.560.3 ± 2.558.6 ± 1.648.4 ± 1.954.5 ± 0.750.6 ± 0.857.0 ± 0.7
(f)✓CCL42.0 ± 0.842.3 ± 1.059.9 ± 0.960.0 ± 0.565.3 ± 1.054.5 ± 3.355.8 ± 0.152.3 ± 1.059.2 ± 0.1
(g)✓CCRd4.2 ± 4.01.3 ± 0.910.3 ± 0.78.5 ± 1.11.3 ± 0.90.1 ± 0.15.3 ± 0.83.3 ± 0.25.5 ± 0.1
(h)✓CCRdP0.7 ± 0.93.2 ± 2.55.7 ± 1.29.5 ± 6.10.9 ± 0.60.3 ± 0.42.4 ± 0.54.3 ± 2.22.9 ± 0.4
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.47, + 0.441 + ], + "angle": 0, + "content": "Figure 3 outlines the architecture, its details will be given in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.47, + 0.487 + ], + "angle": 0, + "content": "Conditioned on a task embedding we denote as \\(\\mathbf{e}\\), the proposed adaptations are based on \"middle\" and \"top\" adapters following [10, 31]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.471, + 0.653 + ], + "angle": 0, + "content": "Middle adapters — we add one trainable adapter after each ViT block to modulate its output. We introduce a set of middle adapters \\( A = \\{\\alpha_l\\}_{l=1..N_l} \\), where \\( \\alpha_l \\) is a 2-layer MLP. In the modified visual encoder \\( \\phi^m \\), each adapter modulates the output of the corresponding self-attention block and is conditioned on the task embedding \\( \\mathbf{e} \\). Its output is combined with the one of the self-attention layer through a residual connection. If we denote the internal hidden representation predicted at layer \\( l \\) as \\( \\mathbf{s}_t^{m,l} \\), and omit references to the weights of \\( \\phi_l \\) and \\( \\alpha_l \\) as in eq. (2) for simplicity, the associated forward pass of a given layer becomes," + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.665, + 0.469, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {s} _ {t} ^ {m, l} = \\phi_ {l} ^ {m} \\left(\\mathbf {s} _ {t} ^ {m, (l - 1)}\\right) (4) \\\\ = \\phi_ {l} \\left(\\mathbf {s} _ {t} ^ {m, (l - 1)}\\right) + \\alpha_ {l} \\left(\\phi_ {l} \\left(\\mathbf {s} _ {t} ^ {m, (l - 1)}\\right), \\mathbf {e}\\right). (5) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Top adapter — A top adapter \\(\\tau\\), also conditioned on the task at hand, is added after the ViT model, to transform the output of the aggregation layer \\(\\psi\\) to be fed to the multi-task policy (presented below). \\(\\tau\\) has the same architecture as a single middle adapter \\(\\alpha_{i}\\). The prediction of \\(\\mathbf{r}_t^m\\), equivalent to \\(\\mathbf{r}_t\\) in the non-adapted case, can be written as," + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.82, + 0.47, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {t} ^ {m} = \\tau \\left[ \\psi \\left[ \\phi^ {m} (\\mathbf {v} _ {t}, \\mathbf {e}; \\theta_ {\\phi}, \\theta_ {A}); \\theta_ {\\psi} \\right], \\mathbf {e}; \\theta_ {\\tau} \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(\\theta_{A}\\) and \\(\\theta_{\\tau}\\) are the weights parametrizing the middle adapters (each middle adapter has a different set of weights) and the top adapter respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.517 + ], + "angle": 0, + "content": "Multi-task policy — We keep the architecture of the single-task policy in eq. (3), as in [20]. However, instead of re-training a policy for each downstream task of interest, we train a single multi-task policy \\(\\pi^m\\), whose action space is the union of the action spaces of the different tasks. During training we apply a masking procedure on the output, considering only the actions possible for the task at hand." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.579 + ], + "angle": 0, + "content": "Let's denote \\(\\tilde{\\mathbf{r}}_t^m\\) as the input to the policy derived from the adapted representation \\(\\mathbf{r}_t^m\\) and the proprioception input \\(\\mathbf{p_t}\\) as done in eq. (3). The conditioning on the task is done by concatenating \\(\\tilde{\\mathbf{r}}_t^m\\) with the task embedding \\(\\mathbf{e}\\), giving" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.592, + 0.891, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {a}} _ {t} = \\pi^ {m} \\left(\\left[ \\tilde {\\mathbf {r}} _ {t} ^ {m}, \\mathbf {e} \\right], \\theta_ {\\pi^ {m}}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.613, + 0.78, + 0.629 + ], + "angle": 0, + "content": "where \\(\\theta_{\\pi^m}\\) are weights parametrizing \\(\\pi^m\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.637, + 0.605, + 0.653 + ], + "angle": 0, + "content": "3.3. Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.811 + ], + "angle": 0, + "content": "We train the model by keeping the weights of the pre-trained vision-encoder model \\(\\theta_{\\phi}\\) frozen, only the weights of the adapter modules \\((\\theta_A,\\theta_\\tau)\\), aggregation layer \\((\\theta_{\\psi})\\), embedding layer \\((\\theta_g)\\) and multi-task policy \\((\\theta_{\\pi^m})\\) are trained, cf. Figure 3a. Let's denote by \\(\\Theta = \\{\\theta_A,\\theta_\\tau ,\\theta_\\psi ,\\theta_g,\\theta_{\\pi^m}\\}\\) the set of optimized weights. We train with imitation learning, more specifically Behavior Cloning (BC): for each known task \\(t_i^k\\), we have access to a set of \\(N_{i}\\) expert trajectories that are composed of \\(T_{i}\\) discrete steps, including expert actions. The optimization problem is given as" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.819, + 0.892, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\hat {\\Theta} = \\arg \\min _ {\\Theta} \\sum_ {i = 1} ^ {K} \\sum_ {n = 1} ^ {N _ {i}} \\sum_ {t = 1} ^ {T _ {i}} \\mathcal {L} \\left(\\hat {\\mathbf {a}} _ {t} ^ {i, n}, \\mathbf {a} _ {t} ^ {i, n}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.869, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{a}}_t^{i,n}\\) and \\(\\mathbf{a}_t^{i,n}\\) are the predicted and ground-truth actions for a given step in a trajectory, and \\(\\mathcal{L}\\) is the Mean" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17851" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.885, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.317, + 0.895, + 0.361 + ], + "angle": 0, + "content": "Figure 5. Known task — Per-task performance of policies in Table 1: single-task policies (row (a)), our approach without any adapter (row (b)) and with conditioned middle and top adapters (row (f)). The adapters lead to a performance gain on most tasks, and our multi-task solution is competitive with single-task policies. Colored bars and error bars respectively show mean and std over 3 training runs (seeds)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.386, + 0.208, + 0.401 + ], + "angle": 0, + "content": "Squared Error loss." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.408, + 0.354, + 0.425 + ], + "angle": 0, + "content": "3.4. Few-shot adaption to new tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.471, + 0.569 + ], + "angle": 0, + "content": "For the Few-shot setting, the task embedding \\(\\mathbf{e}\\) is unknown at inference and needs to be estimated from a set of \\(N_{d}\\) example demonstrations \\(\\mathcal{D} = \\{d_n\\}_{[n = 1..N_d]}\\) where \\(d_{n} = \\{(v_{t}^{n*},p_{t}^{n*},a_{t}^{n*})\\}_{[t = 1..T_{d}]}\\) is composed of observations and actions, with \\(T_{d}\\) being the length of each demonstration. We exploit the conditioning property of the policy itself to estimate the embedding \\(\\hat{\\mathbf{e}}\\) as the one which obtains the highest probability of the demonstration actions, when the policy is applied to the demonstration inputs, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.575, + 0.47, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {e}} = \\arg \\min _ {\\mathbf {e}} \\sum_ {n = 1} ^ {N _ {d}} \\sum_ {t = 1} ^ {T _ {d}} \\mathcal {L} \\left(\\pi^ {m} \\left(\\left[ \\tilde {\\mathbf {r}} _ {t} ^ {m n *}, \\mathbf {e} \\right], \\theta_ {\\pi^ {m}}\\right), \\mathbf {a} _ {t} ^ {n *}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.47, + 0.684 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{r}}^{mn*}\\) is the representation extracted from the demonstration input \\((\\mathbf{v}_t^{n*},\\mathbf{p}_t^{n*})\\), and which itself depends on \\(\\mathbf{e}\\) (not made explicit in the notation). The minimization is carried out with SGD from an embedding initialized to zero." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.21, + 0.712 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Training — all variants involving adapters and/or a multitask policy (rows (b)-(f) in Table 1) were trained for 50 epochs with behavior cloning, cf. §3.3, following training hyper-parameters in [20]. Between 20 and 95 expert trajectories are available depending on the task. We used the datasets of trajectories from [20]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Evaluation — to better handle possible overfit on hyper-parameter selection, our evaluation setup is slightly different from [20] as we perform 100 validation rollouts to select the best checkpoint of each model, and then test the chosen model on 100 test rollouts. For our multi-task policy, the best checkpoint is the one with the highest average" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.892, + 0.476 + ], + "angle": 0, + "content": "validation performance across all tasks. Single-task policies are validated only on the task they were trained on, giving them an advantage, and for this reason, they are reported as \"soft upper bounds\". We report the average performance and standard deviation among 3 trained models (3 random seeds) for each variant as mean ± std (Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.553 + ], + "angle": 0, + "content": "In total, we conduct evaluations of our method on 27 different tasks, 12 known and 15 unknown, with varying environments, embodiments, and required sub-skills. This allows to evaluate the adaptation and generalization abilities of the multi-task policy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Evaluation metrics — Following [20], we consider a rollout success (1 if the task was completed properly, 0 otherwise) for tasks in the Adroit and MetaWorld benchmarks, and report the normalized return for DMC. Episodes have a maximum length of 1000 steps and each step reward is comprised between 0 and 100 in DMC, normalization is therefore done by dividing the agent's return by 10. For all tasks, performance is averaged across rollouts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Known task — Impact of visual adapters — Table 1 presents a detailed comparison of different methods on the known task setting. The baseline in row (a) follows the setup in [20] to train single-task policies (one per task) from non-conditioned VC-1 features. For this variant, we use the representation of the 'CLS' token as the vector fed to the policy as done in [20], while all other baselines use our proposed token aggregation layer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Row (b) is our multi-task policy without any adapter. As expected there is a performance drop compared to the specialized policies in row (a), as the problem to solve has become more difficult. Adding adapters and conditioning them on the task embedding, shown in rows (c)-(f), brings a boost in performance, both for middle and top adapters. In particular, conditioning adds a further boost compared to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17852" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.322, + 0.895, + 0.355 + ], + "angle": 0, + "content": "Figure 6. Few-shot — Per-task performance: After optimizing the task embedding for each task from 5 demonstrations, our method can adapt to many of them (without finetuning). Colored bars and error bars respectively show mean and std over 3 training runs (seeds)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.47, + 0.408 + ], + "angle": 0, + "content": "non-conditioned adapters, with all choices enabled, row (f), obtaining the best average performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.471, + 0.47 + ], + "angle": 0, + "content": "Rows (g) and (h) are ablation experiments evaluating the impact of choosing random task embeddings, row (g), or of taking a random choice between the 12 learned embeddings, row (h). In both cases, the performance collapses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.593 + ], + "angle": 0, + "content": "A particularly important conclusion that can be drawn from the experiments outlined in Table 1 is that the proposed multi-task approach (row (f)) outperforms the single-task policies without adapters (row (a)). This shows that a multi-task policy can perform well on a series of tasks while being trained on a limited set of demonstrations. Figure 4 presents 3 successful test rollouts of our multi-task approach on diverse known tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.47, + 0.716 + ], + "angle": 0, + "content": "Figure 5 visualizes the per-task test performance on known tasks of single-task policies (row (a) in Table 1), our approach without any adapter (row (b) in Table 1) and with conditioned middle and top adapters (row (f) in Table 1). The proposed adapters lead to a performance gain on most tasks compared with the solution without adapters, and the multi-task solution is competitive with single-task policies, even outperforming them on half the tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.869 + ], + "angle": 0, + "content": "Known task Additional ablation studies are presented in the supplementary material (Sections B and C). Section B shows that conditioning the policy when using task-conditioned adapters is not necessary, that our aggregation layer works better than using the 'CLS' token as input to the policy and further highlights the impact of both middle and top adapters. Section C shows that our adapters improve visual embeddings extracted by two state-of-the-art pre-trained backbones, i.e. PVR [23] and MVP [27], confirming the conclusions drawn from experiments on VC-1." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Known task — Visualizing the influence of task-conditioned adaptation — Figures 2 and 3 in the sup-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.892, + 0.438 + ], + "angle": 0, + "content": "plementary material visualize the attention map of the last VC-1 layer. They show that adding adapters focuses attention on objects of interest and task-conditioning improves attention, mainly towards goals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.893, + 0.562 + ], + "angle": 0, + "content": "Few-shot adaptation to new tasks without finetuning model weights is an important ability of any general policy, which we evaluate with the following experiments: for each task within a set of unknown tasks (cf. Figure 2), we collect only 5 demonstrations used to optimize a task embedding specific to this task with the method detailed in section §3.4. We then evaluate the method conditioned on the optimized embedding on 100 test rollouts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.894, + 0.883 + ], + "angle": 0, + "content": "To generate the set \\( T^u \\) of unknown tasks, we select tasks from the MetaWorld dataset that do not belong to CortexBench, and are thus not part of the set of training known tasks \\( T^k \\). We collect demonstrations using single-task policies from TD-MPC2 [8] that were specifically trained on each task of MetaWorld independently. To ensure high-quality demonstrations, we only consider tasks where TD-MPC2 policies reach a success rate higher than \\( 95\\% \\). Furthermore, to be compatible with the setup in CortexBench authors, in particular, to keep the same camera locations, we filter out the tasks where the goal is not always visible in the camera FOV. This leads to a set of 15 unknown tasks that are quite different from the tasks in the training set \\( T^k \\) as they involve different objects (handle press, faucet, plate, door, window, etc.) and types of manipulation (sliding an object, lowering a press, opening a window, etc.). Each collected demonstration is a sequence of visual frames, proprioception inputs, and expert actions. The optimization of the task embedding is performed independently for each task (cf. §3.4). We use the AdamW optimizer and a learning rate of \\( 1e - 1 \\) during the task embedding search." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Figure 6 presents the per-task performance in this set" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17853" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.09, + 0.852, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.256, + 0.561, + 0.271 + ], + "angle": 0, + "content": "(a) Successful rollouts" + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.274, + 0.851, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.441, + 0.546, + 0.456 + ], + "angle": 0, + "content": "(b) Failed rollouts" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.467, + 0.895, + 0.511 + ], + "angle": 0, + "content": "Figure 7. Few-shot Qualitative results: (a) The policy tackles new tasks involving objects and/or manipulation requirements unseen during training. (b) In the first row (button-press-wall task), it performs the task correctly until the end where it fails to properly push the button fully. In the second row (push-back task), it properly moves the cube but fails to bring it to the goal position (green dot)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.473, + 0.748 + ], + "angle": 0, + "content": "ting. Despite the large variations between the new tasks in \\( T^u \\) and the ones in the training set \\( (T^k) \\), the multi-task policy can adapt to many of them, without requiring any weight finetuning. Interestingly, the method performs particularly well on the **Drawer Close** task, which could be related to the presence of the time-inversed **Drawer Open** task in the training set. This provides some evidence that the method can exploit regularities between tasks, which seem to be captured by the task embedding space, making it possible to generalize to unseen variations. Figure 7 (a) shows qualitative examples of successful rollouts on the new tasks. The policy manipulates new objects (faucet, plate, window) and performs new moves (rotating the faucet or sliding the plate) not seen during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Finally, Figure 7 (b) shows failure cases on unknown tasks. As seen on the first row, the policy avoids the wall, reaches the button, and starts pushing it, but fails to push it fully. This particular behavior was also observed on other rollouts, explaining the \\(0\\%\\) success on this button-push-wall task while mastering a part of the required sub-skills. On the second row, the policy is able to move the cube but fails to bring it to the goal location (green dot). This gives some indication of the difficulty of the few-shot generalization case: exploiting regularities in the task space requires that" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.536, + 0.892, + 0.584 + ], + "angle": 0, + "content": "tasks be performed more than just approximately, as often the success metric is sparse, and rollouts only count to the metric when they are executed fully and correctly." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.611, + 0.62, + 0.627 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Perception and action are closely tied together, and studies of human cognition have shown that a priori knowledge about a downstream task guides the visual system. We follow this direction in the context of artificial agents by introducing task-conditioned adapters that modulate the visual features of a pre-trained neural backbone. Such adapters, conditioned on a learned task embedding, improve the performance of a multi-task policy across benchmarks and embodiments. Even more interesting is the use of task embeddings to adapt in a few-shot manner, i.e. from a small set of demonstrations, to new tasks unseen at training time. We propose an optimization procedure to estimate a new task embedding and achieve generalization to unseen tasks, involving new objects and manipulation sub-skills, providing evidence for regularities in the learned embedding space." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Acknowledgement - We thank ANR for support through AI-chair grant \"Remember\" (ANR-20-CHIA-0018)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17854" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.47, + 0.283 + ], + "angle": 0, + "content": "[3] T. J. Buschman and E. K. Miller. Top-down versus bottom-up control of attention in the prefrontal and posterior parietal cortices. science, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.283, + 0.47, + 0.338 + ], + "angle": 0, + "content": "[4] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.339, + 0.47, + 0.382 + ], + "angle": 0, + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.423 + ], + "angle": 0, + "content": "[6] M. Corbetta and G. L. Shulman. Control of goal-directed and stimulus-driven attention in the brain. Nature Reviews Neuroscience, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.47, + 0.506 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.47, + 0.549 + ], + "angle": 0, + "content": "[8] Nicklas Hansen, Hao Su, and Xiaolong Wang. Td-mpc2: Scalable, robust world models for continuous control. arXiv, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.47, + 0.592 + ], + "angle": 0, + "content": "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.592, + 0.47, + 0.648 + ], + "angle": 0, + "content": "[10] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.649, + 0.468, + 0.691 + ], + "angle": 0, + "content": "[11] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.691, + 0.468, + 0.746 + ], + "angle": 0, + "content": "[12] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.746, + 0.468, + 0.803 + ], + "angle": 0, + "content": "[13] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In CoRL, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.804, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[14] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.859, + 0.468, + 0.902 + ], + "angle": 0, + "content": "[15] C. Kervadec, T. Jaunet, G. Antipov, M. Baccouche, R. Vuillemot, and C. Wolf. How Transferrable are Reasoning Patterns in VQA? In CVPR, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[16] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: CLIP embeddings for embodied AI. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[17] Anthony Liang, Ishika Singh, Karl Pertsch, and Jesse Thomason. Transformer adapters for robot learning. In CoRL Workshop on Pre-training Robot Learning, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[18] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized BERT pretraining approach. arXiv, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[19] Yecheng Jason Ma, Shagun Sodhani, Dinesh Jayaraman, Osbert Bastani, Vikash Kumar, and Amy Zhang. Vip: Towards universal visual reward and representation via value-implicit pre-training. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[20] Arjun Majumdar, Karmesh Yadav, Sergio Arnaud, Yecheng Jason Ma, Claire Chen, Sneha Silwal, Aryan Jain, Vincent-Pierre Berges, Pieter Abbeel, Jitendra Malik, et al. Where are we in the search for an artificial visual cortex for embodied intelligence? arXiv, 2023. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[21] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. In CoRL, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[22] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[23] Simone Parisi, Aravind Rajeswaran, Senthil Purushwalkam, and Abhinav Gupta. The unsurprising effectiveness of pretrained vision models for control. In ICML, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[24] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In NAACL-HLT, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[25] Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. Adapterfusion: Nondestructive task composition for transfer learning. In EACL, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.635, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.705, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[27] Ilija Radosavovic, Tete Xiao, Stephen James, Pieter Abbeel, Jitendra Malik, and Trevor Darrell. Real-world robot learning with masked visual pre-training. In CoRL, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[28] Aravind Rajeswaran, Vikash Kumar, Abhishek Gupta, Giulia Vezzani, John Schulman, Emanuel Todorov, and Sergey Levine. Learning complex dexterous manipulation with deep reinforcement learning and demonstrations. RSS, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[29] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, Tom Eccles, Jake Bruce, Ali Razavi, Ashley Edwards, Nicolas Heess, Yutian Chen, Raia Hadsell, Oriol Vinyals, Mahyar Bordbar, and Nando de Freitas. A Generalist Agent. TMLR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17855" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[30] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[31] Mohit Sharma, Claudio Fantacci, Yuxiang Zhou, Skanda Koppula, Nicolas Heess, Jon Scholz, and Yusuf Aytar. Lossless adaptation of pretrained vision models for robotic manipulation. In ICLR, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.261 + ], + "angle": 0, + "content": "[32] Yuval Tassa, Yotam Doron, Alistair Muldal, Tom Erez, Yazhe Li, Diego de Las Casas, David Budden, Abbas Abdelmaleki, Josh Merel, Andrew Lefrancq, et al. Deepmind control suite. arXiv, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.471, + 0.304 + ], + "angle": 0, + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.471, + 0.374 + ], + "angle": 0, + "content": "[34] Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.471, + 0.432 + ], + "angle": 0, + "content": "[35] Karmesh Yadav, Ram Ramrakhya, Arjun Majumdar, Vincent-Pierre Berges, Sachit Kuhar, Dhruv Batra, Alexei Baevski, and Oleksandr Maksymets. Offline visual representation learning for embodied navigation. arXiv, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.471, + 0.488 + ], + "angle": 0, + "content": "[36] Karmesh Yadav, Arjun Majumdar, Ram Ramrakhya, Naoki Yokoyama, Alexei Baevski, Zsolt Kira, Oleksandr Maksymets, and Dhruv Batra. Ovrl-v2: A simple state-of-art baseline forImagenav and objectnav. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.471, + 0.544 + ], + "angle": 0, + "content": "[37] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In CoRL, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.546, + 0.471, + 0.601 + ], + "angle": 0, + "content": "[38] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17856" + } + ] +] \ No newline at end of file diff --git a/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/5ebc706d-8b34-4357-9c79-967e39bda2a5_origin.pdf b/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/5ebc706d-8b34-4357-9c79-967e39bda2a5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4e8f1d17a51dd1722f7729d22b2b546770e0db1c --- /dev/null +++ b/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/5ebc706d-8b34-4357-9c79-967e39bda2a5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:afec3a937c4328cc12a29883ce06998fb525185fb63e750d0826088f0ceec174 +size 8660134 diff --git a/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/full.md b/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8c89926691940c5b897313bf0def6b4313f6cb0f --- /dev/null +++ b/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/full.md @@ -0,0 +1,274 @@ +# Task-conditioned adaptation of visual features in multi-task policy learning + +Pierre Marza + +Laetitia Matignon + +Olivier Simonin1 + +Christian Wolf3 + +$^{1}$ INSA Lyon $^{2}$ UCBL $^{3}$ Naver Labs Europe + +{pierre.marza, olivier.simonin}@insa-lyon.fr + +laetitia.matignon@univ-lyon1.fr,christian.wolf@naverlabs.com + +Project Page: https://pierremarza.github.io/projects/task_conditioned_adaptation/ + +# Abstract + +Successfully addressing a wide variety of tasks is a core ability of autonomous agents, requiring flexibly adapting the underlying decision-making strategies and, as we argue in this work, also adapting the perception modules. An analogical argument would be the human visual system, which uses top-down signals to focus attention determined by the current task. Similarly, we adapt pre-trained large vision models conditioned on specific downstream tasks in the context of multi-task policy learning. We introduce task-conditioned adapters that do not require finetuning any pretrained weights, combined with a single policy trained with behavior cloning and capable of addressing multiple tasks. We condition the visual adapters on task embeddings, which can be selected at inference if the task is known, or alternatively inferred from a set of example demonstrations. To this end, we propose a new optimization-based estimator. We evaluate the method on a wide variety of tasks from the CortexBench benchmark and show that, compared to existing work, it can be addressed with a single policy. In particular, we demonstrate that adapting visual features is a key design choice and that the method generalizes to unseen tasks given a few demonstrations. + +# 1. Introduction + +Vision is one of the most important modalities for agents interacting with the world and is almost indispensable for dexterous manipulation or locomotion, as no other sensor can provide information as rich and versatile. The inherent flexibility of the sensor comes with a high price, the high dimensionality of the information, and the complexity of the processes necessary to extract useful information. Humans and other biological agents are capable of adapting their perception systems to the task at hand. There is indeed evidence for bottom-up and top-down processes in human vision, with the latter guiding attention to regions determined by the requirements of the task [3, 6]. + +![](images/0db39667673aa3f9b61b4a250567188503d70940baee6b7524540aaa7fac4382.jpg) +Figure 1. Task-conditioned adaptation: A single policy can be trained to address multiple heterogenous tasks including manipulation, legged motion etc., and few-shot learning is possible to address tasks given as demonstrations but unseen during training. A key element is the task-conditioned adaptation of visual features. + +There is a growing need for a similar versatility in artificial systems, and general neural networks have been trained from large-scale data in different domains such as natural language processing (NLP), computer vision (CV), and, more recently, robotics. A single general vision model coupled with a neural policy would be an appealing choice if it could allow an easy generalization to new domains or tasks. The wide adoption of attention mechanisms in several domains has made it easier for trained models to adapt their behavior to the requirements of different tasks without changing parameters, and it has been shown that attention plays a crucial role in specialization on a specific instance in language models [34] and vision and language models [15]. However, even powerful generally pre-trained models can benefit from parameter adaptations to specific tasks, e.i + +![](images/cba8313c5b16ab1c5e27e215cffd0c7ef51f03e4965562ea1992de7bd8389ae6.jpg) +Figure 2. Considered tasks: We train the method on a set $T^k$ of known tasks and evaluate it either on the same set, with the task known (Known task setting), or in a Few-shot setting, where a new unseen task from a set $T^u$ is inferred from a few demonstrations. + +ther through fine-tuning [12] or by adding additional trained adapter layers to a frozen model [4]. + +In robotics, prior work on the generalization capabilities of agents has focused on large-scale end-to-end training [29] or, targeting vision specifically, on pre-trained visual models required to generalize to various different policies [19-21]. In this work, we argue and will show that a single policy can be trained for a large number of different tasks including manipulation, locomotion, and that the adaptation of visual features is highly beneficial, beyond the inherent adaptation capabilities of attention-based models. In particular, different tasks require diverse types of invariances and symmetries. While in principle it should be possible to learn to disentangle a sufficiently wide set of factors of variation in a captured representation such that it optimally performs on a wide variety of tasks, we will show that this is not the case for the arguably dominant pretraining method, masked auto-encoding (MAE) [9]. + +We propose task-conditioned adaptation, which allows leveraging the high-quality representations of generally pretrained large vision models, while keeping the required flexibility to address a wide variety of tasks, and also new (unseen) tasks. We introduce a set of task-conditioned visual adapters that can be inserted inside a pre-trained visual Transformer [33]-based backbone. The task is characterized by an embedding space, which is learned from supervision during training. We show that this embedding space captures regularities of tasks and demonstrate this with few-shot capabilities: the single policy and (adapted) visual representations can address new unseen tasks, whose embedding is estimated from a few demonstrations (cf. Figure 1). + +The contributions of this work can be summarized as follows: (i) task-conditioned visual adapters to flexibly modulate visual features to a specific task; (ii) a single multitask policy solving tasks with different embodiments and environments; (iii) a task embedding optimization procedure based on a few demonstrations of a new task (unseen at training time) to adapt the model in a few-shot manner with + +out any weight fine-tuning; (iv) quantitative and qualitative results assessing the gain brought by the different novelties. + +# 2. Related Work + +Pre-trained visual representations for robotics — Backbone models pre-trained on large and diverse data have shown great promises in NLP [2, 14, 18, 24], CV [5, 9, 22, 26], and more recently, robotics [16, 19-21, 23, 27, 35, 36]. Parisi et al. [23] study visual pre-training methods for visuomotor control, showing the quality of self-supervised representations. Nair et al. [21] introduce R3M, a general vision model pre-trained on egocentric video data to capture temporal dynamics and semantic features, improving downstream manipulation performance. Radosavovic et al. [27] employ the well-known MAE framework [9] to pre-train a single vision encoder applied to robots in the real world. Ma et al. [19] pre-train a visual model with a self-supervised value function objective on egocentric human videos to improve control policies. Finally, recent work [16, 35, 36] has shown the great promise of pre-trained models, either CLIP-based [16] or self-supervised [35, 36], in visual navigation. + +The most related work is [20] which studies the impact of pre-trained vision models on a diversity of tasks gathered in a benchmark named Cortexbench. They introduce a Vision Transformer (ViT) [7] backbone, VC-1, pre-trained from a set of out-of-domain datasets with a focus on egocentric visual frames. We believe that visual features should be task-dependent and study how the VC-1 model can be adapted to improve the performance of a single multi-task policy on a subset of Cortexbench tasks, which is also different from previous work focusing on single-task policies. + +Transformer adapters — Transferring Transformer [33]-based pre-trained models to new tasks or domains is an important topic. Methods involving adapter modules were introduced in NLP [10, 11, 25] to allow a fast and parameter-efficient transfer. Recent works employ the same methods in robotics [17, 31]. Sharma + +![](images/f641e1ddeb7b58befeb79d7c4099749ed52bc10fe1bbd85da85f676fd64b7fa6.jpg) +(a) Training the policy and adapters. + +![](images/edb6d490a2fe28cf06f5cc707a835bf3e488d51864155936ffd6091c91adb05b.jpg) +(b) Optimization of the task embedding in the $\overline{\mathrm{Few - shot}}$ setting. + +![](images/2735fe26590ee2eb8f3bde4ef465e576c72d1da755a31f49695f2e9200d452cf.jpg) +(c) Inference for the settings: $\boxed{\mathrm{Known~task}}$ and $\boxed{\mathrm{Few-shot}}$ . +Figure 3. Method overview: (a) the adapted policy is trained with behavior cloning from expert demonstrations and given a visual encoder pre-trained with MAE. The model is conditioned on a task embedding learned from ground-truth 1-in-K task identifiers. (b) In the Few-shot case, a task embedding is estimated by optimization, maximizing the likelihood of a given demonstration of an unknown task. (c) Inference uses a task embedding given in the Known task case, or estimated in the Few-shot case. + +et al. [31] insert visual adapters in a Vision Transformer (ViT) [7] pre-trained model. They introduce different types of adapter blocks ("bottom", "middle", "top") located at diverse places in the visual model and show that combining them improves performance. Liang et al. [17] also show the + +positive impact of inserting task-specific adapters, trained with imitation learning, in a pre-trained Transformer-based model to adapt to robotics tasks. + +While prior work learns a specific set of adapters for each task, we argue that tasks share similarities and explore these regularities with a single set of task-conditioned adapters. + +Multi-task robotics policies — Having a single policy performing a wide range of tasks is a long-standing problem in robotics. With Deep Learning-based solutions becoming more popular, some prior work focuses on training multitask neural agents. Approaches like BC-Z [13], RT-1 [1], RT-2 [38] or Gato [30] study the scaling abilities of neural models to large-scale datasets. Trained generalist agents show strong performance on a wide set of tasks, and can generalize to some extent to novel tasks. + +In contrast, our work leverages pre-trained vision models but does not assume access to a large set of robotics data. Instead, we focus on how to adapt visual features with reasonable computation requirements and train a single multitask policy from only a few expert demonstrations. We also show promising few-shot adaptation to new unknown tasks without requiring very diverse training data. Somewhat related to our work is also TD-MPC2 [8], which introduces a model-based RL algorithm to learn general world models and studies task embeddings to condition a multi-task policy. However, the latter does not act from vision while we specifically study how to modulate visual features conditioned by a task embedding. + +# 3. Task-conditioned adaptation + +All tasks considered in this work are sequential decision-making problems, where at each discrete timestep $t$ an agent receives the last 3 visual frames as an observation $\mathbf{v}_t \in \mathbb{R}^{3 \times h \times w \times 3}$ , where $h$ and $w$ are the height and width of images, and a proprioception input $\mathbf{p}_t \in \mathbb{R}^{d_a}$ , and predicts a continuous action $\hat{\mathbf{a}}_t \in \mathbb{R}^{d_a}$ , where $d_a$ is the dimension of the action space, which depends on the task at hand. We are provided with a training dataset of expert demonstrations to train a single policy, and for inference we study two different setups: (Known task), where we a priori know the task to be executed, and (Few-shot), where the trained policy must be adapted to a new unseen task without fine-tuning only given a small set of demonstrations. + +Known tasks — Following [20], we consider $K = 12$ robotics tasks from 3 benchmarks, Adroit [28], Deepmind control suite [32] and MetaWorld [37]. The set of all known tasks is denoted as $T^{k} = \{t_{i}^{k}\}_{[i = 1..K]}$ , where $t_{i}^{k}$ is a 1-in-K vector encoding a known task, and is illustrated in Figure 2. + +Unknown tasks - The ability of our method to adapt to new skills is evaluated on a set of $U = 15$ tasks from MetaWorld [37], for which we artificially generate demonstrations with a process described in section 4. The set of all unknown tasks is denoted as $T^{u} = \{t_{i}^{u}\}_{[i = 1..U]}$ , where $t_i^u$ + +![](images/4565124186cd3888adcdbe4ca6628aee8d234a72351e7dbb0760b040656e1cb4.jpg) +Figure 4. (Known task) — Qualitative results: Three successful policy rollouts on known tasks from the test set. The multi-task approach performs well on a variety of diverse tasks while being trained on a limited set of demonstrations. + +is a 1-in-U vector encoding an unknown task, and is illustrated in Figure 2. Most importantly, $T^k \cap T^u = \emptyset$ . + +# 3.1. Base agent architecture + +Following a large body of work in end-to-end training for robotics, the agent directly maps pixels to actions and decomposes into a visual encoder and a policy. + +Visual encoder without adapters — following [20], the visual encoder, denoted as $\phi$ , is a ViT model [7] pretrained with masked auto-encoding (MAE). We keep pretrained weights from VC-1 in [20], which are publicly available. However, we change the way the representation is collected from the pre-trained model. Unlike [20], the representation is not taken as the embedding of the 'CLS' token, which we consider to be undertrained by the MAE pretext task. Instead, we train a fully-connected layer $\psi$ to aggregate all the token representations of the last layer of the ViT except the 'CLS' token. The visual observation $\mathbf{v}_t$ associated with timestep $t$ is thus encoded as + +$$ +\mathbf {r} _ {t} = \psi [ \phi (\mathbf {v} _ {t}; \theta_ {\phi}); \theta_ {\psi} ], \tag {1} +$$ + +where $\theta_{\phi}$ and $\theta_{\psi}$ are weights parametrizing $\phi$ and $\psi$ respectively. $\mathbf{r}_t \in \mathbb{R}^{3 \times d_r}$ as it contains the $d_r$ -dim encoding of each of the 3 last visual frames processed as a data batch, where $d_r$ is the output dimension of $\psi$ . + +As this will be relevant later, we recall here that a ViT $\phi$ is composed of a sequence of $N_{l}$ self-attention blocks, where $\phi_l$ is the layer at index $l$ . If we denote the internal hidden representation predicted at layer $l$ as $\mathbf{s}_t^l$ , and omit the weights of $\phi_l$ for simplicity, we have, + +$$ +\mathbf {s} _ {t} ^ {l} = \phi_ {l} \left(\mathbf {s} _ {t} ^ {l - 1}\right), \tag {2} +$$ + +where $\mathbf{s}_t^0 = \mathbf{v}_t$ + +Single-task policy — following [20], the policy $\pi$ is implemented as an MLP predicting actions from the input which is a concatenation of the current frame, two frame differences and the proprioception input $\mathbf{p_t}$ , + +$$ +\hat {\mathbf {a}} _ {t} = \pi \left(\left[ \mathbf {r} _ {t, 1} - \mathbf {r} _ {t, 0}, \mathbf {r} _ {t, 2} - \mathbf {r} _ {t, 1}, \mathbf {r} _ {t, 2}, \mathbf {p} _ {\mathbf {t}} \right]; \theta_ {\pi}\right), \tag {3} +$$ + +where $[ ]$ is the concatenation operator and $\theta_{\pi}$ are weights parametrizing $\pi$ . + +# 3.2. Adaptation + +Our key contributions are visual adapter modules along with a multi-task policy, which are all conditioned on the task at hand. This is done with a specific task embedding for each task, taken from an embedding space of dimension $d_{e}$ , which is aimed to have sufficient regularities to enable few-show generalization to unseen tasks. Importantly, the different adapters and the multi-task policy are conditioned on the same task embedding, leading to a common and shared embedding space. For the known task setting, where the ground-truth label of the task is available, the task embedding is projected from a 1-in-K vector with a linear function $g$ trained jointly with the adapters and the policy with the downstream loss (imitation learning). In the Few-shot setting, at test time a new unknown task is described with a few demonstrations, and a task embedding is estimated through optimization, as will be detailed in subsection 3.4. + +Table 1. Known task — Impact of visual adapters: Validation and test performance on known tasks of different baselines highlighting the gain brought by adapters. Both middle and top adapters bring a boost in performance, and conditioning them on the learned task embedding increases performance. Our multi-task policy outperforms single-task policies with VC-1 non-adapted features. MT $\pi$ : multitask policy - NC: Non-conditioned - C: Conditioned - Task emb.: whether to input at evaluation time, either the learned task embedding and chosen from ground-truth (L), a random vector as the task embedding (Rd), or a randomly picked task embedding among the set of $K = 12$ embeddings (RdP) - Benchmarks avg: average performance across the 3 considered benchmarks (Adroit, DMC, MetaWorld) - Tasks avg: average performance across all 12 known tasks. Performance is reported as mean ± std over 3 training runs (seeds). + +
MT Adapters TaskMulti-task performance
πMid.Top emb.AdroitDMCMetaWorldBenchmarks avgTasks avg
ValTestValTestValTestValTestValTest
(a) [20]---N/A44.0 ± 1.138.3 ± 2.549.6 ± 0.548.0 ± 0.353.5 ± 2.147.8 ± 2.049.1 ± 0.544.7 ± 0.350.3 ± 0.9
(b)✓--L36.3 ± 1.733.0 ± 4.055.3 ± 1.454.1 ± 0.341.7 ± 1.634.7 ± 0.944.4 ± 1.040.6 ± 1.846.5 ± 1.1
(c)✓NC-L40.2 ± 1.237.3 ± 2.854.0 ± 1.554.8 ± 1.945.8 ± 4.536.3 ± 2.546.7 ± 1.642.8 ± 1.948.3 ± 1.9
(d)✓C-L42.0 ± 2.543.8 ± 2.259.1 ± 1.358.8 ± 0.348.6 ± 4.840.8 ± 3.049.9 ± 2.147.8 ± 1.451.9 ± 2.1
(e)✓CNCL44.3 ± 1.243.2 ± 1.560.5 ± 0.560.3 ± 2.558.6 ± 1.648.4 ± 1.954.5 ± 0.750.6 ± 0.857.0 ± 0.7
(f)✓CCL42.0 ± 0.842.3 ± 1.059.9 ± 0.960.0 ± 0.565.3 ± 1.054.5 ± 3.355.8 ± 0.152.3 ± 1.059.2 ± 0.1
(g)✓CCRd4.2 ± 4.01.3 ± 0.910.3 ± 0.78.5 ± 1.11.3 ± 0.90.1 ± 0.15.3 ± 0.83.3 ± 0.25.5 ± 0.1
(h)✓CCRdP0.7 ± 0.93.2 ± 2.55.7 ± 1.29.5 ± 6.10.9 ± 0.60.3 ± 0.42.4 ± 0.54.3 ± 2.22.9 ± 0.4
+ +Figure 3 outlines the architecture, its details will be given in the supplementary material. + +Conditioned on a task embedding we denote as $\mathbf{e}$ , the proposed adaptations are based on "middle" and "top" adapters following [10, 31]. + +Middle adapters — we add one trainable adapter after each ViT block to modulate its output. We introduce a set of middle adapters $A = \{\alpha_l\}_{l=1..N_l}$ , where $\alpha_l$ is a 2-layer MLP. In the modified visual encoder $\phi^m$ , each adapter modulates the output of the corresponding self-attention block and is conditioned on the task embedding $\mathbf{e}$ . Its output is combined with the one of the self-attention layer through a residual connection. If we denote the internal hidden representation predicted at layer $l$ as $\mathbf{s}_t^{m,l}$ , and omit references to the weights of $\phi_l$ and $\alpha_l$ as in eq. (2) for simplicity, the associated forward pass of a given layer becomes, + +$$ +\begin{array}{l} \mathbf {s} _ {t} ^ {m, l} = \phi_ {l} ^ {m} \left(\mathbf {s} _ {t} ^ {m, (l - 1)}\right) (4) \\ = \phi_ {l} \left(\mathbf {s} _ {t} ^ {m, (l - 1)}\right) + \alpha_ {l} \left(\phi_ {l} \left(\mathbf {s} _ {t} ^ {m, (l - 1)}\right), \mathbf {e}\right). (5) \\ \end{array} +$$ + +Top adapter — A top adapter $\tau$ , also conditioned on the task at hand, is added after the ViT model, to transform the output of the aggregation layer $\psi$ to be fed to the multi-task policy (presented below). $\tau$ has the same architecture as a single middle adapter $\alpha_{i}$ . The prediction of $\mathbf{r}_t^m$ , equivalent to $\mathbf{r}_t$ in the non-adapted case, can be written as, + +$$ +\mathbf {r} _ {t} ^ {m} = \tau \left[ \psi \left[ \phi^ {m} (\mathbf {v} _ {t}, \mathbf {e}; \theta_ {\phi}, \theta_ {A}); \theta_ {\psi} \right], \mathbf {e}; \theta_ {\tau} \right], \tag {6} +$$ + +where $\theta_{A}$ and $\theta_{\tau}$ are the weights parametrizing the middle adapters (each middle adapter has a different set of weights) and the top adapter respectively. + +Multi-task policy — We keep the architecture of the single-task policy in eq. (3), as in [20]. However, instead of re-training a policy for each downstream task of interest, we train a single multi-task policy $\pi^m$ , whose action space is the union of the action spaces of the different tasks. During training we apply a masking procedure on the output, considering only the actions possible for the task at hand. + +Let's denote $\tilde{\mathbf{r}}_t^m$ as the input to the policy derived from the adapted representation $\mathbf{r}_t^m$ and the proprioception input $\mathbf{p_t}$ as done in eq. (3). The conditioning on the task is done by concatenating $\tilde{\mathbf{r}}_t^m$ with the task embedding $\mathbf{e}$ , giving + +$$ +\hat {\mathbf {a}} _ {t} = \pi^ {m} \left(\left[ \tilde {\mathbf {r}} _ {t} ^ {m}, \mathbf {e} \right], \theta_ {\pi^ {m}}\right), \tag {7} +$$ + +where $\theta_{\pi^m}$ are weights parametrizing $\pi^m$ + +# 3.3. Training + +We train the model by keeping the weights of the pre-trained vision-encoder model $\theta_{\phi}$ frozen, only the weights of the adapter modules $(\theta_A,\theta_\tau)$ , aggregation layer $(\theta_{\psi})$ , embedding layer $(\theta_g)$ and multi-task policy $(\theta_{\pi^m})$ are trained, cf. Figure 3a. Let's denote by $\Theta = \{\theta_A,\theta_\tau ,\theta_\psi ,\theta_g,\theta_{\pi^m}\}$ the set of optimized weights. We train with imitation learning, more specifically Behavior Cloning (BC): for each known task $t_i^k$ , we have access to a set of $N_{i}$ expert trajectories that are composed of $T_{i}$ discrete steps, including expert actions. The optimization problem is given as + +$$ +\hat {\Theta} = \arg \min _ {\Theta} \sum_ {i = 1} ^ {K} \sum_ {n = 1} ^ {N _ {i}} \sum_ {t = 1} ^ {T _ {i}} \mathcal {L} \left(\hat {\mathbf {a}} _ {t} ^ {i, n}, \mathbf {a} _ {t} ^ {i, n}\right), \tag {8} +$$ + +where $\hat{\mathbf{a}}_t^{i,n}$ and $\mathbf{a}_t^{i,n}$ are the predicted and ground-truth actions for a given step in a trajectory, and $\mathcal{L}$ is the Mean + +![](images/021926cb4c3f4b4f761e69a332cce286f538b40adbb423ac5b644d57c3bb7916.jpg) +Figure 5. Known task — Per-task performance of policies in Table 1: single-task policies (row (a)), our approach without any adapter (row (b)) and with conditioned middle and top adapters (row (f)). The adapters lead to a performance gain on most tasks, and our multi-task solution is competitive with single-task policies. Colored bars and error bars respectively show mean and std over 3 training runs (seeds). + +Squared Error loss. + +# 3.4. Few-shot adaption to new tasks + +For the Few-shot setting, the task embedding $\mathbf{e}$ is unknown at inference and needs to be estimated from a set of $N_{d}$ example demonstrations $\mathcal{D} = \{d_n\}_{[n = 1..N_d]}$ where $d_{n} = \{(v_{t}^{n*},p_{t}^{n*},a_{t}^{n*})\}_{[t = 1..T_{d}]}$ is composed of observations and actions, with $T_{d}$ being the length of each demonstration. We exploit the conditioning property of the policy itself to estimate the embedding $\hat{\mathbf{e}}$ as the one which obtains the highest probability of the demonstration actions, when the policy is applied to the demonstration inputs, i.e. + +$$ +\hat {\mathbf {e}} = \arg \min _ {\mathbf {e}} \sum_ {n = 1} ^ {N _ {d}} \sum_ {t = 1} ^ {T _ {d}} \mathcal {L} \left(\pi^ {m} \left(\left[ \tilde {\mathbf {r}} _ {t} ^ {m n *}, \mathbf {e} \right], \theta_ {\pi^ {m}}\right), \mathbf {a} _ {t} ^ {n *}\right), \tag {9} +$$ + +where $\tilde{\mathbf{r}}^{mn*}$ is the representation extracted from the demonstration input $(\mathbf{v}_t^{n*},\mathbf{p}_t^{n*})$ , and which itself depends on $\mathbf{e}$ (not made explicit in the notation). The minimization is carried out with SGD from an embedding initialized to zero. + +# 4. Experiments + +Training — all variants involving adapters and/or a multitask policy (rows (b)-(f) in Table 1) were trained for 50 epochs with behavior cloning, cf. §3.3, following training hyper-parameters in [20]. Between 20 and 95 expert trajectories are available depending on the task. We used the datasets of trajectories from [20]. + +Evaluation — to better handle possible overfit on hyper-parameter selection, our evaluation setup is slightly different from [20] as we perform 100 validation rollouts to select the best checkpoint of each model, and then test the chosen model on 100 test rollouts. For our multi-task policy, the best checkpoint is the one with the highest average + +validation performance across all tasks. Single-task policies are validated only on the task they were trained on, giving them an advantage, and for this reason, they are reported as "soft upper bounds". We report the average performance and standard deviation among 3 trained models (3 random seeds) for each variant as mean ± std (Table 1). + +In total, we conduct evaluations of our method on 27 different tasks, 12 known and 15 unknown, with varying environments, embodiments, and required sub-skills. This allows to evaluate the adaptation and generalization abilities of the multi-task policy. + +Evaluation metrics — Following [20], we consider a rollout success (1 if the task was completed properly, 0 otherwise) for tasks in the Adroit and MetaWorld benchmarks, and report the normalized return for DMC. Episodes have a maximum length of 1000 steps and each step reward is comprised between 0 and 100 in DMC, normalization is therefore done by dividing the agent's return by 10. For all tasks, performance is averaged across rollouts. + +Known task — Impact of visual adapters — Table 1 presents a detailed comparison of different methods on the known task setting. The baseline in row (a) follows the setup in [20] to train single-task policies (one per task) from non-conditioned VC-1 features. For this variant, we use the representation of the 'CLS' token as the vector fed to the policy as done in [20], while all other baselines use our proposed token aggregation layer. + +Row (b) is our multi-task policy without any adapter. As expected there is a performance drop compared to the specialized policies in row (a), as the problem to solve has become more difficult. Adding adapters and conditioning them on the task embedding, shown in rows (c)-(f), brings a boost in performance, both for middle and top adapters. In particular, conditioning adds a further boost compared to + +![](images/1e78ac63be7b007999df7e11223b10b164079f4d191bbea07c9579f7b4617fa6.jpg) +Figure 6. Few-shot — Per-task performance: After optimizing the task embedding for each task from 5 demonstrations, our method can adapt to many of them (without finetuning). Colored bars and error bars respectively show mean and std over 3 training runs (seeds). + +non-conditioned adapters, with all choices enabled, row (f), obtaining the best average performance. + +Rows (g) and (h) are ablation experiments evaluating the impact of choosing random task embeddings, row (g), or of taking a random choice between the 12 learned embeddings, row (h). In both cases, the performance collapses. + +A particularly important conclusion that can be drawn from the experiments outlined in Table 1 is that the proposed multi-task approach (row (f)) outperforms the single-task policies without adapters (row (a)). This shows that a multi-task policy can perform well on a series of tasks while being trained on a limited set of demonstrations. Figure 4 presents 3 successful test rollouts of our multi-task approach on diverse known tasks. + +Figure 5 visualizes the per-task test performance on known tasks of single-task policies (row (a) in Table 1), our approach without any adapter (row (b) in Table 1) and with conditioned middle and top adapters (row (f) in Table 1). The proposed adapters lead to a performance gain on most tasks compared with the solution without adapters, and the multi-task solution is competitive with single-task policies, even outperforming them on half the tasks. + +Known task Additional ablation studies are presented in the supplementary material (Sections B and C). Section B shows that conditioning the policy when using task-conditioned adapters is not necessary, that our aggregation layer works better than using the 'CLS' token as input to the policy and further highlights the impact of both middle and top adapters. Section C shows that our adapters improve visual embeddings extracted by two state-of-the-art pre-trained backbones, i.e. PVR [23] and MVP [27], confirming the conclusions drawn from experiments on VC-1. + +Known task — Visualizing the influence of task-conditioned adaptation — Figures 2 and 3 in the sup- + +plementary material visualize the attention map of the last VC-1 layer. They show that adding adapters focuses attention on objects of interest and task-conditioning improves attention, mainly towards goals. + +Few-shot adaptation to new tasks without finetuning model weights is an important ability of any general policy, which we evaluate with the following experiments: for each task within a set of unknown tasks (cf. Figure 2), we collect only 5 demonstrations used to optimize a task embedding specific to this task with the method detailed in section §3.4. We then evaluate the method conditioned on the optimized embedding on 100 test rollouts. + +To generate the set $T^u$ of unknown tasks, we select tasks from the MetaWorld dataset that do not belong to CortexBench, and are thus not part of the set of training known tasks $T^k$ . We collect demonstrations using single-task policies from TD-MPC2 [8] that were specifically trained on each task of MetaWorld independently. To ensure high-quality demonstrations, we only consider tasks where TD-MPC2 policies reach a success rate higher than $95\%$ . Furthermore, to be compatible with the setup in CortexBench authors, in particular, to keep the same camera locations, we filter out the tasks where the goal is not always visible in the camera FOV. This leads to a set of 15 unknown tasks that are quite different from the tasks in the training set $T^k$ as they involve different objects (handle press, faucet, plate, door, window, etc.) and types of manipulation (sliding an object, lowering a press, opening a window, etc.). Each collected demonstration is a sequence of visual frames, proprioception inputs, and expert actions. The optimization of the task embedding is performed independently for each task (cf. §3.4). We use the AdamW optimizer and a learning rate of $1e - 1$ during the task embedding search. + +Figure 6 presents the per-task performance in this set + +![](images/7750e91419ec948d0b65300110badacf47393c40682a25a32738fd99c81c80ae.jpg) +(a) Successful rollouts + +![](images/9a2493e3a4364821fc20776ca06aa3ddbed9d05f6c3b5bb273bec996ecd85ef8.jpg) +(b) Failed rollouts +Figure 7. Few-shot Qualitative results: (a) The policy tackles new tasks involving objects and/or manipulation requirements unseen during training. (b) In the first row (button-press-wall task), it performs the task correctly until the end where it fails to properly push the button fully. In the second row (push-back task), it properly moves the cube but fails to bring it to the goal position (green dot). + +ting. Despite the large variations between the new tasks in $T^u$ and the ones in the training set $(T^k)$ , the multi-task policy can adapt to many of them, without requiring any weight finetuning. Interestingly, the method performs particularly well on the **Drawer Close** task, which could be related to the presence of the time-inversed **Drawer Open** task in the training set. This provides some evidence that the method can exploit regularities between tasks, which seem to be captured by the task embedding space, making it possible to generalize to unseen variations. Figure 7 (a) shows qualitative examples of successful rollouts on the new tasks. The policy manipulates new objects (faucet, plate, window) and performs new moves (rotating the faucet or sliding the plate) not seen during training. + +Finally, Figure 7 (b) shows failure cases on unknown tasks. As seen on the first row, the policy avoids the wall, reaches the button, and starts pushing it, but fails to push it fully. This particular behavior was also observed on other rollouts, explaining the $0\%$ success on this button-push-wall task while mastering a part of the required sub-skills. On the second row, the policy is able to move the cube but fails to bring it to the goal location (green dot). This gives some indication of the difficulty of the few-shot generalization case: exploiting regularities in the task space requires that + +tasks be performed more than just approximately, as often the success metric is sparse, and rollouts only count to the metric when they are executed fully and correctly. + +# 5. Conclusion + +Perception and action are closely tied together, and studies of human cognition have shown that a priori knowledge about a downstream task guides the visual system. We follow this direction in the context of artificial agents by introducing task-conditioned adapters that modulate the visual features of a pre-trained neural backbone. Such adapters, conditioned on a learned task embedding, improve the performance of a multi-task policy across benchmarks and embodiments. Even more interesting is the use of task embeddings to adapt in a few-shot manner, i.e. from a small set of demonstrations, to new tasks unseen at training time. We propose an optimization procedure to estimate a new task embedding and achieve generalization to unseen tasks, involving new objects and manipulation sub-skills, providing evidence for regularities in the learned embedding space. + +Acknowledgement - We thank ANR for support through AI-chair grant "Remember" (ANR-20-CHIA-0018). + +# References + +[1] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv, 2022. 3 +[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020. 2 +[3] T. J. Buschman and E. K. Miller. Top-down versus bottom-up control of attention in the prefrontal and posterior parietal cortices. science, 2007. 1 +[4] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In NeurIPS, 2022. 2 +[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2 +[6] M. Corbetta and G. L. Shulman. Control of goal-directed and stimulus-driven attention in the brain. Nature Reviews Neuroscience, 2002. 1 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 2, 3, 4 +[8] Nicklas Hansen, Hao Su, and Xiaolong Wang. Td-mpc2: Scalable, robust world models for continuous control. arXiv, 2023. 3, 7 +[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2 +[10] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 2, 5 +[11] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2021. 2 +[12] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2 +[13] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In CoRL, 2022. 3 +[14] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019. 2 +[15] C. Kervadec, T. Jaunet, G. Antipov, M. Baccouche, R. Vuillemot, and C. Wolf. How Transferrable are Reasoning Patterns in VQA? In CVPR, 2021. 1 + +[16] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: CLIP embeddings for embodied AI. In CVPR, 2022. 2 +[17] Anthony Liang, Ishika Singh, Karl Pertsch, and Jesse Thomason. Transformer adapters for robot learning. In CoRL Workshop on Pre-training Robot Learning, 2022. 2, 3 +[18] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized BERT pretraining approach. arXiv, 2019. 2 +[19] Yecheng Jason Ma, Shagun Sodhani, Dinesh Jayaraman, Osbert Bastani, Vikash Kumar, and Amy Zhang. Vip: Towards universal visual reward and representation via value-implicit pre-training. In ICLR, 2022. 2 +[20] Arjun Majumdar, Karmesh Yadav, Sergio Arnaud, Yecheng Jason Ma, Claire Chen, Sneha Silwal, Aryan Jain, Vincent-Pierre Berges, Pieter Abbeel, Jitendra Malik, et al. Where are we in the search for an artificial visual cortex for embodied intelligence? arXiv, 2023. 2, 3, 4, 5, 6 +[21] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. In CoRL, 2023. 2 +[22] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv, 2023. 2 +[23] Simone Parisi, Aravind Rajeswaran, Senthil Purushwalkam, and Abhinav Gupta. The unsurprising effectiveness of pretrained vision models for control. In ICML, 2022. 2, 7 +[24] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In NAACL-HLT, 2018. 2 +[25] Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. Adapterfusion: Nondestructive task composition for transfer learning. In EACL, 2021. 2 +[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2 +[27] Ilija Radosavovic, Tete Xiao, Stephen James, Pieter Abbeel, Jitendra Malik, and Trevor Darrell. Real-world robot learning with masked visual pre-training. In CoRL, 2023. 2, 7 +[28] Aravind Rajeswaran, Vikash Kumar, Abhishek Gupta, Giulia Vezzani, John Schulman, Emanuel Todorov, and Sergey Levine. Learning complex dexterous manipulation with deep reinforcement learning and demonstrations. RSS, 2018. 3 +[29] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, Tom Eccles, Jake Bruce, Ali Razavi, Ashley Edwards, Nicolas Heess, Yutian Chen, Raia Hadsell, Oriol Vinyals, Mahyar Bordbar, and Nando de Freitas. A Generalist Agent. TMLR, 2022. 2 + +[30] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv, 2022. 3 +[31] Mohit Sharma, Claudio Fantacci, Yuxiang Zhou, Skanda Koppula, Nicolas Heess, Jon Scholz, and Yusuf Aytar. Lossless adaptation of pretrained vision models for robotic manipulation. In ICLR, 2022. 2, 3, 5 +[32] Yuval Tassa, Yotam Doron, Alistair Muldal, Tom Erez, Yazhe Li, Diego de Las Casas, David Budden, Abbas Abdelmaleki, Josh Merel, Andrew Lefrancq, et al. Deepmind control suite. arXiv, 2018. 3 +[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 2 +[34] Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 2019. 1 +[35] Karmesh Yadav, Ram Ramrakhya, Arjun Majumdar, Vincent-Pierre Berges, Sachit Kuhar, Dhruv Batra, Alexei Baevski, and Oleksandr Maksymets. Offline visual representation learning for embodied navigation. arXiv, 2022. 2 +[36] Karmesh Yadav, Arjun Majumdar, Ram Ramrakhya, Naoki Yokoyama, Alexei Baevski, Zsolt Kira, Oleksandr Maksymets, and Dhruv Batra. Ovrl-v2: A simple state-of-art baseline forImagenav and objectnav. arXiv, 2023. 2 +[37] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In CoRL, 2020. 3 +[38] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023. 3 \ No newline at end of file diff --git a/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/images.zip b/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5e6bbcac43bad21b877dd4e9740b4402d04d6888 --- /dev/null +++ b/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:429d2cf49c469572b7114d3cd08ecefa2a25caddd69678e027de3a9fa78dcc36 +size 717912 diff --git a/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/layout.json b/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3e64b7b8dd0bda1d62aa5999b3a30f7d0301667a --- /dev/null +++ b/2024/Task-Conditioned Adaptation of Visual Features in Multi-Task Policy Learning/layout.json @@ -0,0 +1,7804 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 529, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 529, + 122 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 529, + 122 + ], + "type": "text", + "content": "Task-conditioned adaptation of visual features in multi-task policy learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 141, + 167, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 141, + 167, + 156 + ], + "spans": [ + { + "bbox": [ + 94, + 141, + 167, + 156 + ], + "type": "text", + "content": "Pierre Marza" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 191, + 143, + 284, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 191, + 143, + 284, + 157 + ], + "spans": [ + { + "bbox": [ + 191, + 143, + 284, + 157 + ], + "type": "text", + "content": "Laetitia Matignon" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 309, + 143, + 392, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 143, + 392, + 156 + ], + "spans": [ + { + "bbox": [ + 309, + 143, + 392, + 156 + ], + "type": "text", + "content": "Olivier Simonin1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 418, + 143, + 495, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 143, + 495, + 156 + ], + "spans": [ + { + "bbox": [ + 418, + 143, + 495, + 156 + ], + "type": "text", + "content": "Christian Wolf3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 198, + 158, + 393, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 158, + 393, + 171 + ], + "spans": [ + { + "bbox": [ + 198, + 158, + 393, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 198, + 158, + 393, + 171 + ], + "type": "text", + "content": "INSA Lyon " + }, + { + "bbox": [ + 198, + 158, + 393, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 198, + 158, + 393, + 171 + ], + "type": "text", + "content": "UCBL " + }, + { + "bbox": [ + 198, + 158, + 393, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 198, + 158, + 393, + 171 + ], + "type": "text", + "content": "Naver Labs Europe" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 179, + 174, + 416, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 174, + 416, + 185 + ], + "spans": [ + { + "bbox": [ + 179, + 174, + 416, + 185 + ], + "type": "text", + "content": "{pierre.marza, olivier.simonin}@insa-lyon.fr" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 188, + 460, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 188, + 460, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 188, + 460, + 198 + ], + "type": "text", + "content": "laetitia.matignon@univ-lyon1.fr,christian.wolf@naverlabs.com" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 201, + 457, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 201, + 457, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 201, + 457, + 213 + ], + "type": "text", + "content": "Project Page: https://pierremarza.github.io/projects/task_conditioned_adaptation/" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 266, + 290, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 529 + ], + "type": "text", + "content": "Successfully addressing a wide variety of tasks is a core ability of autonomous agents, requiring flexibly adapting the underlying decision-making strategies and, as we argue in this work, also adapting the perception modules. An analogical argument would be the human visual system, which uses top-down signals to focus attention determined by the current task. Similarly, we adapt pre-trained large vision models conditioned on specific downstream tasks in the context of multi-task policy learning. We introduce task-conditioned adapters that do not require finetuning any pretrained weights, combined with a single policy trained with behavior cloning and capable of addressing multiple tasks. We condition the visual adapters on task embeddings, which can be selected at inference if the task is known, or alternatively inferred from a set of example demonstrations. To this end, we propose a new optimization-based estimator. We evaluate the method on a wide variety of tasks from the CortexBench benchmark and show that, compared to existing work, it can be addressed with a single policy. In particular, we demonstrate that adapting visual features is a key design choice and that the method generalizes to unseen tasks given a few demonstrations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 550, + 128, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 128, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 128, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Vision is one of the most important modalities for agents interacting with the world and is almost indispensable for dexterous manipulation or locomotion, as no other sensor can provide information as rich and versatile. The inherent flexibility of the sensor comes with a high price, the high dimensionality of the information, and the complexity of the processes necessary to extract useful information. Humans and other biological agents are capable of adapting their perception systems to the task at hand. There is indeed evidence for bottom-up and top-down processes in human vision, with the latter guiding attention to regions determined by the requirements of the task [3, 6]." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 309, + 240, + 541, + 446 + ], + "blocks": [ + { + "bbox": [ + 309, + 240, + 541, + 446 + ], + "lines": [ + { + "bbox": [ + 309, + 240, + 541, + 446 + ], + "spans": [ + { + "bbox": [ + 309, + 240, + 541, + 446 + ], + "type": "image", + "image_path": "0db39667673aa3f9b61b4a250567188503d70940baee6b7524540aaa7fac4382.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 455, + 547, + 511 + ], + "lines": [ + { + "bbox": [ + 305, + 455, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 305, + 455, + 547, + 511 + ], + "type": "text", + "content": "Figure 1. Task-conditioned adaptation: A single policy can be trained to address multiple heterogenous tasks including manipulation, legged motion etc., and few-shot learning is possible to address tasks given as demonstrations but unseen during training. A key element is the task-conditioned adaptation of visual features." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "There is a growing need for a similar versatility in artificial systems, and general neural networks have been trained from large-scale data in different domains such as natural language processing (NLP), computer vision (CV), and, more recently, robotics. A single general vision model coupled with a neural policy would be an appealing choice if it could allow an easy generalization to new domains or tasks. The wide adoption of attention mechanisms in several domains has made it easier for trained models to adapt their behavior to the requirements of different tasks without changing parameters, and it has been shown that attention plays a crucial role in specialization on a specific instance in language models [34] and vision and language models [15]. However, even powerful generally pre-trained models can benefit from parameter adaptations to specific tasks, e.i" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17847" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 68, + 545, + 203 + ], + "blocks": [ + { + "bbox": [ + 50, + 68, + 545, + 203 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 545, + 203 + ], + "type": "image", + "image_path": "cba8313c5b16ab1c5e27e215cffd0c7ef51f03e4965562ea1992de7bd8389ae6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 212, + 546, + 235 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 546, + 235 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 546, + 235 + ], + "type": "text", + "content": "Figure 2. Considered tasks: We train the method on a set " + }, + { + "bbox": [ + 46, + 212, + 546, + 235 + ], + "type": "inline_equation", + "content": "T^k" + }, + { + "bbox": [ + 46, + 212, + 546, + 235 + ], + "type": "text", + "content": " of known tasks and evaluate it either on the same set, with the task known (Known task setting), or in a Few-shot setting, where a new unseen task from a set " + }, + { + "bbox": [ + 46, + 212, + 546, + 235 + ], + "type": "inline_equation", + "content": "T^u" + }, + { + "bbox": [ + 46, + 212, + 546, + 235 + ], + "type": "text", + "content": " is inferred from a few demonstrations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 255, + 287, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 280 + ], + "type": "text", + "content": "ther through fine-tuning [12] or by adding additional trained adapter layers to a frozen model [4]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "content": "In robotics, prior work on the generalization capabilities of agents has focused on large-scale end-to-end training [29] or, targeting vision specifically, on pre-trained visual models required to generalize to various different policies [19-21]. In this work, we argue and will show that a single policy can be trained for a large number of different tasks including manipulation, locomotion, and that the adaptation of visual features is highly beneficial, beyond the inherent adaptation capabilities of attention-based models. In particular, different tasks require diverse types of invariances and symmetries. While in principle it should be possible to learn to disentangle a sufficiently wide set of factors of variation in a captured representation such that it optimally performs on a wide variety of tasks, we will show that this is not the case for the arguably dominant pretraining method, masked auto-encoding (MAE) [9]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 473, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 628 + ], + "type": "text", + "content": "We propose task-conditioned adaptation, which allows leveraging the high-quality representations of generally pretrained large vision models, while keeping the required flexibility to address a wide variety of tasks, and also new (unseen) tasks. We introduce a set of task-conditioned visual adapters that can be inserted inside a pre-trained visual Transformer [33]-based backbone. The task is characterized by an embedding space, which is learned from supervision during training. We show that this embedding space captures regularities of tasks and demonstrate this with few-shot capabilities: the single policy and (adapted) visual representations can address new unseen tasks, whose embedding is estimated from a few demonstrations (cf. Figure 1)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "The contributions of this work can be summarized as follows: (i) task-conditioned visual adapters to flexibly modulate visual features to a specific task; (ii) a single multitask policy solving tasks with different embodiments and environments; (iii) a task embedding optimization procedure based on a few demonstrations of a new task (unseen at training time) to adapt the model in a few-shot manner with" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 255, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 255, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 545, + 280 + ], + "type": "text", + "content": "out any weight fine-tuning; (iv) quantitative and qualitative results assessing the gain brought by the different novelties." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 294, + 392, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 294, + 392, + 306 + ], + "spans": [ + { + "bbox": [ + 306, + 294, + 392, + 306 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 316, + 546, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 546, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 546, + 519 + ], + "type": "text", + "content": "Pre-trained visual representations for robotics — Backbone models pre-trained on large and diverse data have shown great promises in NLP [2, 14, 18, 24], CV [5, 9, 22, 26], and more recently, robotics [16, 19-21, 23, 27, 35, 36]. Parisi et al. [23] study visual pre-training methods for visuomotor control, showing the quality of self-supervised representations. Nair et al. [21] introduce R3M, a general vision model pre-trained on egocentric video data to capture temporal dynamics and semantic features, improving downstream manipulation performance. Radosavovic et al. [27] employ the well-known MAE framework [9] to pre-train a single vision encoder applied to robots in the real world. Ma et al. [19] pre-train a visual model with a self-supervised value function objective on egocentric human videos to improve control policies. Finally, recent work [16, 35, 36] has shown the great promise of pre-trained models, either CLIP-based [16] or self-supervised [35, 36], in visual navigation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 521, + 546, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 640 + ], + "type": "text", + "content": "The most related work is [20] which studies the impact of pre-trained vision models on a diversity of tasks gathered in a benchmark named Cortexbench. They introduce a Vision Transformer (ViT) [7] backbone, VC-1, pre-trained from a set of out-of-domain datasets with a focus on egocentric visual frames. We believe that visual features should be task-dependent and study how the VC-1 model can be adapted to improve the performance of a single multi-task policy on a subset of Cortexbench tasks, which is also different from previous work focusing on single-task policies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "Transformer adapters — Transferring Transformer [33]-based pre-trained models to new tasks or domains is an important topic. Methods involving adapter modules were introduced in NLP [10, 11, 25] to allow a fast and parameter-efficient transfer. Recent works employ the same methods in robotics [17, 31]. Sharma" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17848" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 286, + 205 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 286, + 205 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 286, + 205 + ], + "type": "image", + "image_path": "f641e1ddeb7b58befeb79d7c4099749ed52bc10fe1bbd85da85f676fd64b7fa6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 209, + 233, + 221 + ], + "lines": [ + { + "bbox": [ + 102, + 209, + 233, + 221 + ], + "spans": [ + { + "bbox": [ + 102, + 209, + 233, + 221 + ], + "type": "text", + "content": "(a) Training the policy and adapters." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 230, + 284, + 363 + ], + "blocks": [ + { + "bbox": [ + 50, + 230, + 284, + 363 + ], + "lines": [ + { + "bbox": [ + 50, + 230, + 284, + 363 + ], + "spans": [ + { + "bbox": [ + 50, + 230, + 284, + 363 + ], + "type": "image", + "image_path": "edb6d490a2fe28cf06f5cc707a835bf3e488d51864155936ffd6091c91adb05b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 367, + 285, + 379 + ], + "lines": [ + { + "bbox": [ + 50, + 367, + 285, + 379 + ], + "spans": [ + { + "bbox": [ + 50, + 367, + 285, + 379 + ], + "type": "text", + "content": "(b) Optimization of the task embedding in the " + }, + { + "bbox": [ + 50, + 367, + 285, + 379 + ], + "type": "inline_equation", + "content": "\\overline{\\mathrm{Few - shot}}" + }, + { + "bbox": [ + 50, + 367, + 285, + 379 + ], + "type": "text", + "content": " setting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 385, + 284, + 517 + ], + "blocks": [ + { + "bbox": [ + 50, + 385, + 284, + 517 + ], + "lines": [ + { + "bbox": [ + 50, + 385, + 284, + 517 + ], + "spans": [ + { + "bbox": [ + 50, + 385, + 284, + 517 + ], + "type": "image", + "image_path": "2735fe26590ee2eb8f3bde4ef465e576c72d1da755a31f49695f2e9200d452cf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 520, + 272, + 532 + ], + "lines": [ + { + "bbox": [ + 61, + 520, + 272, + 532 + ], + "spans": [ + { + "bbox": [ + 61, + 520, + 272, + 532 + ], + "type": "text", + "content": "(c) Inference for the settings: " + }, + { + "bbox": [ + 61, + 520, + 272, + 532 + ], + "type": "inline_equation", + "content": "\\boxed{\\mathrm{Known~task}}" + }, + { + "bbox": [ + 61, + 520, + 272, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 520, + 272, + 532 + ], + "type": "inline_equation", + "content": "\\boxed{\\mathrm{Few-shot}}" + }, + { + "bbox": [ + 61, + 520, + 272, + 532 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 540, + 287, + 628 + ], + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 628 + ], + "type": "text", + "content": "Figure 3. Method overview: (a) the adapted policy is trained with behavior cloning from expert demonstrations and given a visual encoder pre-trained with MAE. The model is conditioned on a task embedding learned from ground-truth 1-in-K task identifiers. (b) In the Few-shot case, a task embedding is estimated by optimization, maximizing the likelihood of a given demonstration of an unknown task. (c) Inference uses a task embedding given in the Known task case, or estimated in the Few-shot case." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "et al. [31] insert visual adapters in a Vision Transformer (ViT) [7] pre-trained model. They introduce different types of adapter blocks (\"bottom\", \"middle\", \"top\") located at diverse places in the visual model and show that combining them improves performance. Liang et al. [17] also show the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": "positive impact of inserting task-specific adapters, trained with imitation learning, in a pre-trained Transformer-based model to adapt to robotics tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 108, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 144 + ], + "type": "text", + "content": "While prior work learns a specific set of adapters for each task, we argue that tasks share similarities and explore these regularities with a single set of task-conditioned adapters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 144, + 546, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 546, + 251 + ], + "type": "text", + "content": "Multi-task robotics policies — Having a single policy performing a wide range of tasks is a long-standing problem in robotics. With Deep Learning-based solutions becoming more popular, some prior work focuses on training multitask neural agents. Approaches like BC-Z [13], RT-1 [1], RT-2 [38] or Gato [30] study the scaling abilities of neural models to large-scale datasets. Trained generalist agents show strong performance on a wide set of tasks, and can generalize to some extent to novel tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 251, + 547, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 547, + 407 + ], + "type": "text", + "content": "In contrast, our work leverages pre-trained vision models but does not assume access to a large set of robotics data. Instead, we focus on how to adapt visual features with reasonable computation requirements and train a single multitask policy from only a few expert demonstrations. We also show promising few-shot adaptation to new unknown tasks without requiring very diverse training data. Somewhat related to our work is also TD-MPC2 [8], which introduces a model-based RL algorithm to learn general world models and studies task embeddings to condition a multi-task policy. However, the latter does not act from vision while we specifically study how to modulate visual features conditioned by a task embedding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 418, + 468, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 468, + 432 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 468, + 432 + ], + "type": "text", + "content": "3. Task-conditioned adaptation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 437, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 593 + ], + "type": "text", + "content": "All tasks considered in this work are sequential decision-making problems, where at each discrete timestep " + }, + { + "bbox": [ + 304, + 437, + 545, + 593 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 437, + 545, + 593 + ], + "type": "text", + "content": " an agent receives the last 3 visual frames as an observation " + }, + { + "bbox": [ + 304, + 437, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t \\in \\mathbb{R}^{3 \\times h \\times w \\times 3}" + }, + { + "bbox": [ + 304, + 437, + 545, + 593 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 437, + 545, + 593 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 437, + 545, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 437, + 545, + 593 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 437, + 545, + 593 + ], + "type": "text", + "content": " are the height and width of images, and a proprioception input " + }, + { + "bbox": [ + 304, + 437, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_t \\in \\mathbb{R}^{d_a}" + }, + { + "bbox": [ + 304, + 437, + 545, + 593 + ], + "type": "text", + "content": ", and predicts a continuous action " + }, + { + "bbox": [ + 304, + 437, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}_t \\in \\mathbb{R}^{d_a}" + }, + { + "bbox": [ + 304, + 437, + 545, + 593 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 437, + 545, + 593 + ], + "type": "inline_equation", + "content": "d_a" + }, + { + "bbox": [ + 304, + 437, + 545, + 593 + ], + "type": "text", + "content": " is the dimension of the action space, which depends on the task at hand. We are provided with a training dataset of expert demonstrations to train a single policy, and for inference we study two different setups: (Known task), where we a priori know the task to be executed, and (Few-shot), where the trained policy must be adapted to a new unseen task without fine-tuning only given a small set of demonstrations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 654 + ], + "type": "text", + "content": "Known tasks — Following [20], we consider " + }, + { + "bbox": [ + 304, + 594, + 545, + 654 + ], + "type": "inline_equation", + "content": "K = 12" + }, + { + "bbox": [ + 304, + 594, + 545, + 654 + ], + "type": "text", + "content": " robotics tasks from 3 benchmarks, Adroit [28], Deepmind control suite [32] and MetaWorld [37]. The set of all known tasks is denoted as " + }, + { + "bbox": [ + 304, + 594, + 545, + 654 + ], + "type": "inline_equation", + "content": "T^{k} = \\{t_{i}^{k}\\}_{[i = 1..K]}" + }, + { + "bbox": [ + 304, + 594, + 545, + 654 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 594, + 545, + 654 + ], + "type": "inline_equation", + "content": "t_{i}^{k}" + }, + { + "bbox": [ + 304, + 594, + 545, + 654 + ], + "type": "text", + "content": " is a 1-in-K vector encoding a known task, and is illustrated in Figure 2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": "Unknown tasks - The ability of our method to adapt to new skills is evaluated on a set of " + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "inline_equation", + "content": "U = 15" + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": " tasks from MetaWorld [37], for which we artificially generate demonstrations with a process described in section 4. The set of all unknown tasks is denoted as " + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "inline_equation", + "content": "T^{u} = \\{t_{i}^{u}\\}_{[i = 1..U]}" + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "inline_equation", + "content": "t_i^u" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17849" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 547, + 286 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 547, + 286 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 547, + 286 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 547, + 286 + ], + "type": "image", + "image_path": "4565124186cd3888adcdbe4ca6628aee8d234a72351e7dbb0760b040656e1cb4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 293, + 547, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 547, + 316 + ], + "type": "text", + "content": "Figure 4. (Known task) — Qualitative results: Three successful policy rollouts on known tasks from the test set. The multi-task approach performs well on a variety of diverse tasks while being trained on a limited set of demonstrations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 337, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 288, + 361 + ], + "type": "text", + "content": "is a 1-in-U vector encoding an unknown task, and is illustrated in Figure 2. Most importantly, " + }, + { + "bbox": [ + 46, + 337, + 288, + 361 + ], + "type": "inline_equation", + "content": "T^k \\cap T^u = \\emptyset" + }, + { + "bbox": [ + 46, + 337, + 288, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 370, + 180, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 180, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 180, + 384 + ], + "type": "text", + "content": "3.1. Base agent architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 389, + 288, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 288, + 426 + ], + "type": "text", + "content": "Following a large body of work in end-to-end training for robotics, the agent directly maps pixels to actions and decomposes into a visual encoder and a policy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 289, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 571 + ], + "type": "text", + "content": "Visual encoder without adapters — following [20], the visual encoder, denoted as " + }, + { + "bbox": [ + 46, + 426, + 289, + 571 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 426, + 289, + 571 + ], + "type": "text", + "content": ", is a ViT model [7] pretrained with masked auto-encoding (MAE). We keep pretrained weights from VC-1 in [20], which are publicly available. However, we change the way the representation is collected from the pre-trained model. Unlike [20], the representation is not taken as the embedding of the 'CLS' token, which we consider to be undertrained by the MAE pretext task. Instead, we train a fully-connected layer " + }, + { + "bbox": [ + 46, + 426, + 289, + 571 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 426, + 289, + 571 + ], + "type": "text", + "content": " to aggregate all the token representations of the last layer of the ViT except the 'CLS' token. The visual observation " + }, + { + "bbox": [ + 46, + 426, + 289, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t" + }, + { + "bbox": [ + 46, + 426, + 289, + 571 + ], + "type": "text", + "content": " associated with timestep " + }, + { + "bbox": [ + 46, + 426, + 289, + 571 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 426, + 289, + 571 + ], + "type": "text", + "content": " is thus encoded as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 118, + 580, + 287, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 580, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 118, + 580, + 287, + 596 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {t} = \\psi [ \\phi (\\mathbf {v} _ {t}; \\theta_ {\\phi}); \\theta_ {\\psi} ], \\tag {1}", + "image_path": "5a9f4a787c538e8105d3f0511d72c6565081014cbbbe8649ede10a221f7c397d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\theta_{\\phi}" + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\theta_{\\psi}" + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "text", + "content": " are weights parametrizing " + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "text", + "content": " respectively. " + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_t \\in \\mathbb{R}^{3 \\times d_r}" + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "text", + "content": " as it contains the " + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "inline_equation", + "content": "d_r" + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "text", + "content": "-dim encoding of each of the 3 last visual frames processed as a data batch, where " + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "inline_equation", + "content": "d_r" + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "text", + "content": " is the output dimension of " + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "As this will be relevant later, we recall here that a ViT " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " is composed of a sequence of " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "N_{l}" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " self-attention blocks, where " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\phi_l" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " is the layer at index " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": ". If we denote the internal hidden representation predicted at layer " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t^l" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": ", and omit the weights of " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\phi_l" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " for simplicity, we have," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 394, + 348, + 545, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 348, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 394, + 348, + 545, + 364 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {t} ^ {l} = \\phi_ {l} \\left(\\mathbf {s} _ {t} ^ {l - 1}\\right), \\tag {2}", + "image_path": "10c2cebfacbe1f05ea6267521db284137107ed7a0214edb7b41301e2c483ce56.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 369, + 365, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 365, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 365, + 382 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 369, + 365, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t^0 = \\mathbf{v}_t" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 382, + 547, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 382, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 382, + 547, + 431 + ], + "type": "text", + "content": "Single-task policy — following [20], the policy " + }, + { + "bbox": [ + 305, + 382, + 547, + 431 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 305, + 382, + 547, + 431 + ], + "type": "text", + "content": " is implemented as an MLP predicting actions from the input which is a concatenation of the current frame, two frame differences and the proprioception input " + }, + { + "bbox": [ + 305, + 382, + 547, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{p_t}" + }, + { + "bbox": [ + 305, + 382, + 547, + 431 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 332, + 439, + 545, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 439, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 332, + 439, + 545, + 461 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {a}} _ {t} = \\pi \\left(\\left[ \\mathbf {r} _ {t, 1} - \\mathbf {r} _ {t, 0}, \\mathbf {r} _ {t, 2} - \\mathbf {r} _ {t, 1}, \\mathbf {r} _ {t, 2}, \\mathbf {p} _ {\\mathbf {t}} \\right]; \\theta_ {\\pi}\\right), \\tag {3}", + "image_path": "97ae920d02acf6f8d7d81332e3faf2e0c5641b77d3552e40b6b292a2fb640fe6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 469, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 469, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 305, + 469, + 545, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 469, + 545, + 495 + ], + "type": "inline_equation", + "content": "[ ]" + }, + { + "bbox": [ + 305, + 469, + 545, + 495 + ], + "type": "text", + "content": " is the concatenation operator and " + }, + { + "bbox": [ + 305, + 469, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\theta_{\\pi}" + }, + { + "bbox": [ + 305, + 469, + 545, + 495 + ], + "type": "text", + "content": " are weights parametrizing " + }, + { + "bbox": [ + 305, + 469, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 305, + 469, + 545, + 495 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 503, + 382, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 382, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 382, + 516 + ], + "type": "text", + "content": "3.2. Adaptation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": "Our key contributions are visual adapter modules along with a multi-task policy, which are all conditioned on the task at hand. This is done with a specific task embedding for each task, taken from an embedding space of dimension " + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "inline_equation", + "content": "d_{e}" + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": ", which is aimed to have sufficient regularities to enable few-show generalization to unseen tasks. Importantly, the different adapters and the multi-task policy are conditioned on the same task embedding, leading to a common and shared embedding space. For the known task setting, where the ground-truth label of the task is available, the task embedding is projected from a 1-in-K vector with a linear function " + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": " trained jointly with the adapters and the policy with the downstream loss (imitation learning). In the Few-shot setting, at test time a new unknown task is described with a few demonstrations, and a task embedding is estimated through optimization, as will be detailed in subsection 3.4." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17850" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 156, + 545, + 308 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 149 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 149 + ], + "type": "text", + "content": "Table 1. Known task — Impact of visual adapters: Validation and test performance on known tasks of different baselines highlighting the gain brought by adapters. Both middle and top adapters bring a boost in performance, and conditioning them on the learned task embedding increases performance. Our multi-task policy outperforms single-task policies with VC-1 non-adapted features. MT " + }, + { + "bbox": [ + 46, + 70, + 547, + 149 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 70, + 547, + 149 + ], + "type": "text", + "content": ": multitask policy - NC: Non-conditioned - C: Conditioned - Task emb.: whether to input at evaluation time, either the learned task embedding and chosen from ground-truth (L), a random vector as the task embedding (Rd), or a randomly picked task embedding among the set of " + }, + { + "bbox": [ + 46, + 70, + 547, + 149 + ], + "type": "inline_equation", + "content": "K = 12" + }, + { + "bbox": [ + 46, + 70, + 547, + 149 + ], + "type": "text", + "content": " embeddings (RdP) - Benchmarks avg: average performance across the 3 considered benchmarks (Adroit, DMC, MetaWorld) - Tasks avg: average performance across all 12 known tasks. Performance is reported as mean ± std over 3 training runs (seeds)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 156, + 545, + 308 + ], + "lines": [ + { + "bbox": [ + 48, + 156, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 156, + 545, + 308 + ], + "type": "table", + "html": "
MT Adapters TaskMulti-task performance
πMid.Top emb.AdroitDMCMetaWorldBenchmarks avgTasks avg
ValTestValTestValTestValTestValTest
(a) [20]---N/A44.0 ± 1.138.3 ± 2.549.6 ± 0.548.0 ± 0.353.5 ± 2.147.8 ± 2.049.1 ± 0.544.7 ± 0.350.3 ± 0.9
(b)✓--L36.3 ± 1.733.0 ± 4.055.3 ± 1.454.1 ± 0.341.7 ± 1.634.7 ± 0.944.4 ± 1.040.6 ± 1.846.5 ± 1.1
(c)✓NC-L40.2 ± 1.237.3 ± 2.854.0 ± 1.554.8 ± 1.945.8 ± 4.536.3 ± 2.546.7 ± 1.642.8 ± 1.948.3 ± 1.9
(d)✓C-L42.0 ± 2.543.8 ± 2.259.1 ± 1.358.8 ± 0.348.6 ± 4.840.8 ± 3.049.9 ± 2.147.8 ± 1.451.9 ± 2.1
(e)✓CNCL44.3 ± 1.243.2 ± 1.560.5 ± 0.560.3 ± 2.558.6 ± 1.648.4 ± 1.954.5 ± 0.750.6 ± 0.857.0 ± 0.7
(f)✓CCL42.0 ± 0.842.3 ± 1.059.9 ± 0.960.0 ± 0.565.3 ± 1.054.5 ± 3.355.8 ± 0.152.3 ± 1.059.2 ± 0.1
(g)✓CCRd4.2 ± 4.01.3 ± 0.910.3 ± 0.78.5 ± 1.11.3 ± 0.90.1 ± 0.15.3 ± 0.83.3 ± 0.25.5 ± 0.1
(h)✓CCRdP0.7 ± 0.93.2 ± 2.55.7 ± 1.29.5 ± 6.10.9 ± 0.60.3 ± 0.42.4 ± 0.54.3 ± 2.22.9 ± 0.4
", + "image_path": "2dbc6f2a9060acbe244681c03ec3ddc1897124ffca0e10029a68124388166e05.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 325, + 287, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 287, + 349 + ], + "type": "text", + "content": "Figure 3 outlines the architecture, its details will be given in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 287, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 385 + ], + "type": "text", + "content": "Conditioned on a task embedding we denote as " + }, + { + "bbox": [ + 46, + 350, + 287, + 385 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 46, + 350, + 287, + 385 + ], + "type": "text", + "content": ", the proposed adaptations are based on \"middle\" and \"top\" adapters following [10, 31]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "content": "Middle adapters — we add one trainable adapter after each ViT block to modulate its output. We introduce a set of middle adapters " + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "inline_equation", + "content": "A = \\{\\alpha_l\\}_{l=1..N_l}" + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\alpha_l" + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "content": " is a 2-layer MLP. In the modified visual encoder " + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\phi^m" + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "content": ", each adapter modulates the output of the corresponding self-attention block and is conditioned on the task embedding " + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "content": ". Its output is combined with the one of the self-attention layer through a residual connection. If we denote the internal hidden representation predicted at layer " + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t^{m,l}" + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "content": ", and omit references to the weights of " + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\phi_l" + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\alpha_l" + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "content": " as in eq. (2) for simplicity, the associated forward pass of a given layer becomes," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 526, + 287, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 526, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 80, + 526, + 287, + 559 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {s} _ {t} ^ {m, l} = \\phi_ {l} ^ {m} \\left(\\mathbf {s} _ {t} ^ {m, (l - 1)}\\right) (4) \\\\ = \\phi_ {l} \\left(\\mathbf {s} _ {t} ^ {m, (l - 1)}\\right) + \\alpha_ {l} \\left(\\phi_ {l} \\left(\\mathbf {s} _ {t} ^ {m, (l - 1)}\\right), \\mathbf {e}\\right). (5) \\\\ \\end{array}", + "image_path": "7704ee0c848f11d38f49f89c683cf34dc72d09f025df9bd9ef7f493b77f0ea7b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "type": "text", + "content": "Top adapter — A top adapter " + }, + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "type": "text", + "content": ", also conditioned on the task at hand, is added after the ViT model, to transform the output of the aggregation layer " + }, + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "type": "text", + "content": " to be fed to the multi-task policy (presented below). " + }, + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "type": "text", + "content": " has the same architecture as a single middle adapter " + }, + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "type": "text", + "content": ". The prediction of " + }, + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_t^m" + }, + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "type": "text", + "content": ", equivalent to " + }, + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_t" + }, + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "type": "text", + "content": " in the non-adapted case, can be written as," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 81, + 649, + 287, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 649, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 81, + 649, + 287, + 670 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {t} ^ {m} = \\tau \\left[ \\psi \\left[ \\phi^ {m} (\\mathbf {v} _ {t}, \\mathbf {e}; \\theta_ {\\phi}, \\theta_ {A}); \\theta_ {\\psi} \\right], \\mathbf {e}; \\theta_ {\\tau} \\right], \\tag {6}", + "image_path": "ad30f0f100d50ff4f32a5c43252b98ad368f95ccd09aa3ff822203e1ff8e4991.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{A}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{\\tau}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " are the weights parametrizing the middle adapters (each middle adapter has a different set of weights) and the top adapter respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 325, + 545, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 409 + ], + "type": "text", + "content": "Multi-task policy — We keep the architecture of the single-task policy in eq. (3), as in [20]. However, instead of re-training a policy for each downstream task of interest, we train a single multi-task policy " + }, + { + "bbox": [ + 304, + 325, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\pi^m" + }, + { + "bbox": [ + 304, + 325, + 545, + 409 + ], + "type": "text", + "content": ", whose action space is the union of the action spaces of the different tasks. During training we apply a masking procedure on the output, considering only the actions possible for the task at hand." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 410, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 458 + ], + "type": "text", + "content": "Let's denote " + }, + { + "bbox": [ + 304, + 410, + 545, + 458 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{r}}_t^m" + }, + { + "bbox": [ + 304, + 410, + 545, + 458 + ], + "type": "text", + "content": " as the input to the policy derived from the adapted representation " + }, + { + "bbox": [ + 304, + 410, + 545, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_t^m" + }, + { + "bbox": [ + 304, + 410, + 545, + 458 + ], + "type": "text", + "content": " and the proprioception input " + }, + { + "bbox": [ + 304, + 410, + 545, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{p_t}" + }, + { + "bbox": [ + 304, + 410, + 545, + 458 + ], + "type": "text", + "content": " as done in eq. (3). The conditioning on the task is done by concatenating " + }, + { + "bbox": [ + 304, + 410, + 545, + 458 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{r}}_t^m" + }, + { + "bbox": [ + 304, + 410, + 545, + 458 + ], + "type": "text", + "content": " with the task embedding " + }, + { + "bbox": [ + 304, + 410, + 545, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 304, + 410, + 545, + 458 + ], + "type": "text", + "content": ", giving" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 373, + 468, + 545, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 468, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 373, + 468, + 545, + 482 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {a}} _ {t} = \\pi^ {m} \\left(\\left[ \\tilde {\\mathbf {r}} _ {t} ^ {m}, \\mathbf {e} \\right], \\theta_ {\\pi^ {m}}\\right), \\tag {7}", + "image_path": "8d9617e24265885e3e1003cd79539622f3bc9e09c5575f1975b3aef0acea41b5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 485, + 477, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 485, + 477, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 485, + 477, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 485, + 477, + 498 + ], + "type": "inline_equation", + "content": "\\theta_{\\pi^m}" + }, + { + "bbox": [ + 306, + 485, + 477, + 498 + ], + "type": "text", + "content": " are weights parametrizing " + }, + { + "bbox": [ + 306, + 485, + 477, + 498 + ], + "type": "inline_equation", + "content": "\\pi^m" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 504, + 370, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 504, + 370, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 504, + 370, + 517 + ], + "type": "text", + "content": "3.3. Training" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "text", + "content": "We train the model by keeping the weights of the pre-trained vision-encoder model " + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\theta_{\\phi}" + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "text", + "content": " frozen, only the weights of the adapter modules " + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "inline_equation", + "content": "(\\theta_A,\\theta_\\tau)" + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "text", + "content": ", aggregation layer " + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "inline_equation", + "content": "(\\theta_{\\psi})" + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "text", + "content": ", embedding layer " + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "inline_equation", + "content": "(\\theta_g)" + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "text", + "content": " and multi-task policy " + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "inline_equation", + "content": "(\\theta_{\\pi^m})" + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "text", + "content": " are trained, cf. Figure 3a. Let's denote by " + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\Theta = \\{\\theta_A,\\theta_\\tau ,\\theta_\\psi ,\\theta_g,\\theta_{\\pi^m}\\}" + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "text", + "content": " the set of optimized weights. We train with imitation learning, more specifically Behavior Cloning (BC): for each known task " + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "inline_equation", + "content": "t_i^k" + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "text", + "content": ", we have access to a set of " + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "text", + "content": " expert trajectories that are composed of " + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "text", + "content": " discrete steps, including expert actions. The optimization problem is given as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 345, + 648, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 648, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 345, + 648, + 545, + 681 + ], + "type": "interline_equation", + "content": "\\hat {\\Theta} = \\arg \\min _ {\\Theta} \\sum_ {i = 1} ^ {K} \\sum_ {n = 1} ^ {N _ {i}} \\sum_ {t = 1} ^ {T _ {i}} \\mathcal {L} \\left(\\hat {\\mathbf {a}} _ {t} ^ {i, n}, \\mathbf {a} _ {t} ^ {i, n}\\right), \\tag {8}", + "image_path": "1d08bc9589931718b5bf5316897912d29c93dc87fe39eb6618a760cb6030a644.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 688, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 688, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 688, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 688, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}_t^{i,n}" + }, + { + "bbox": [ + 304, + 688, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 688, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t^{i,n}" + }, + { + "bbox": [ + 304, + 688, + 545, + 714 + ], + "type": "text", + "content": " are the predicted and ground-truth actions for a given step in a trajectory, and " + }, + { + "bbox": [ + 304, + 688, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 688, + 545, + 714 + ], + "type": "text", + "content": " is the Mean" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17851" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 541, + 240 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 541, + 240 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 541, + 240 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 541, + 240 + ], + "type": "image", + "image_path": "021926cb4c3f4b4f761e69a332cce286f538b40adbb423ac5b644d57c3bb7916.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 547, + 285 + ], + "type": "text", + "content": "Figure 5. Known task — Per-task performance of policies in Table 1: single-task policies (row (a)), our approach without any adapter (row (b)) and with conditioned middle and top adapters (row (f)). The adapters lead to a performance gain on most tasks, and our multi-task solution is competitive with single-task policies. Colored bars and error bars respectively show mean and std over 3 training runs (seeds)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 305, + 127, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 305, + 127, + 317 + ], + "spans": [ + { + "bbox": [ + 47, + 305, + 127, + 317 + ], + "type": "text", + "content": "Squared Error loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 323, + 216, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 216, + 336 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 216, + 336 + ], + "type": "text", + "content": "3.4. Few-shot adaption to new tasks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 341, + 288, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 288, + 450 + ], + "type": "text", + "content": "For the Few-shot setting, the task embedding " + }, + { + "bbox": [ + 46, + 341, + 288, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 46, + 341, + 288, + 450 + ], + "type": "text", + "content": " is unknown at inference and needs to be estimated from a set of " + }, + { + "bbox": [ + 46, + 341, + 288, + 450 + ], + "type": "inline_equation", + "content": "N_{d}" + }, + { + "bbox": [ + 46, + 341, + 288, + 450 + ], + "type": "text", + "content": " example demonstrations " + }, + { + "bbox": [ + 46, + 341, + 288, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{d_n\\}_{[n = 1..N_d]}" + }, + { + "bbox": [ + 46, + 341, + 288, + 450 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 341, + 288, + 450 + ], + "type": "inline_equation", + "content": "d_{n} = \\{(v_{t}^{n*},p_{t}^{n*},a_{t}^{n*})\\}_{[t = 1..T_{d}]}" + }, + { + "bbox": [ + 46, + 341, + 288, + 450 + ], + "type": "text", + "content": " is composed of observations and actions, with " + }, + { + "bbox": [ + 46, + 341, + 288, + 450 + ], + "type": "inline_equation", + "content": "T_{d}" + }, + { + "bbox": [ + 46, + 341, + 288, + 450 + ], + "type": "text", + "content": " being the length of each demonstration. We exploit the conditioning property of the policy itself to estimate the embedding " + }, + { + "bbox": [ + 46, + 341, + 288, + 450 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{e}}" + }, + { + "bbox": [ + 46, + 341, + 288, + 450 + ], + "type": "text", + "content": " as the one which obtains the highest probability of the demonstration actions, when the policy is applied to the demonstration inputs, i.e." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 455, + 287, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 287, + 488 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {e}} = \\arg \\min _ {\\mathbf {e}} \\sum_ {n = 1} ^ {N _ {d}} \\sum_ {t = 1} ^ {T _ {d}} \\mathcal {L} \\left(\\pi^ {m} \\left(\\left[ \\tilde {\\mathbf {r}} _ {t} ^ {m n *}, \\mathbf {e} \\right], \\theta_ {\\pi^ {m}}\\right), \\mathbf {a} _ {t} ^ {n *}\\right), \\tag {9}", + "image_path": "c43180c5d2f28741580b8f33853b595bb217b7b644d0e466bc1ebe9b7f24ef99.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 492, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 287, + 541 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 492, + 287, + 541 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{r}}^{mn*}" + }, + { + "bbox": [ + 46, + 492, + 287, + 541 + ], + "type": "text", + "content": " is the representation extracted from the demonstration input " + }, + { + "bbox": [ + 46, + 492, + 287, + 541 + ], + "type": "inline_equation", + "content": "(\\mathbf{v}_t^{n*},\\mathbf{p}_t^{n*})" + }, + { + "bbox": [ + 46, + 492, + 287, + 541 + ], + "type": "text", + "content": ", and which itself depends on " + }, + { + "bbox": [ + 46, + 492, + 287, + 541 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 46, + 492, + 287, + 541 + ], + "type": "text", + "content": " (not made explicit in the notation). The minimization is carried out with SGD from an embedding initialized to zero." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 550, + 128, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 128, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 128, + 563 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 641 + ], + "type": "text", + "content": "Training — all variants involving adapters and/or a multitask policy (rows (b)-(f) in Table 1) were trained for 50 epochs with behavior cloning, cf. §3.3, following training hyper-parameters in [20]. Between 20 and 95 expert trajectories are available depending on the task. We used the datasets of trajectories from [20]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Evaluation — to better handle possible overfit on hyper-parameter selection, our evaluation setup is slightly different from [20] as we perform 100 validation rollouts to select the best checkpoint of each model, and then test the chosen model on 100 test rollouts. For our multi-task policy, the best checkpoint is the one with the highest average" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 305, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 545, + 376 + ], + "type": "text", + "content": "validation performance across all tasks. Single-task policies are validated only on the task they were trained on, giving them an advantage, and for this reason, they are reported as \"soft upper bounds\". We report the average performance and standard deviation among 3 trained models (3 random seeds) for each variant as mean ± std (Table 1)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 377, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 437 + ], + "type": "text", + "content": "In total, we conduct evaluations of our method on 27 different tasks, 12 known and 15 unknown, with varying environments, embodiments, and required sub-skills. This allows to evaluate the adaptation and generalization abilities of the multi-task policy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "content": "Evaluation metrics — Following [20], we consider a rollout success (1 if the task was completed properly, 0 otherwise) for tasks in the Adroit and MetaWorld benchmarks, and report the normalized return for DMC. Episodes have a maximum length of 1000 steps and each step reward is comprised between 0 and 100 in DMC, normalization is therefore done by dividing the agent's return by 10. For all tasks, performance is averaged across rollouts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 533, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 629 + ], + "type": "text", + "content": "Known task — Impact of visual adapters — Table 1 presents a detailed comparison of different methods on the known task setting. The baseline in row (a) follows the setup in [20] to train single-task policies (one per task) from non-conditioned VC-1 features. For this variant, we use the representation of the 'CLS' token as the vector fed to the policy as done in [20], while all other baselines use our proposed token aggregation layer." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": "Row (b) is our multi-task policy without any adapter. As expected there is a performance drop compared to the specialized policies in row (a), as the problem to solve has become more difficult. Adding adapters and conditioning them on the task embedding, shown in rows (c)-(f), brings a boost in performance, both for middle and top adapters. In particular, conditioning adds a further boost compared to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17852" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 545, + 246 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 246 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 246 + ], + "type": "image", + "image_path": "1e78ac63be7b007999df7e11223b10b164079f4d191bbea07c9579f7b4617fa6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 547, + 281 + ], + "type": "text", + "content": "Figure 6. Few-shot — Per-task performance: After optimizing the task embedding for each task from 5 demonstrations, our method can adapt to many of them (without finetuning). Colored bars and error bars respectively show mean and std over 3 training runs (seeds)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 298, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 287, + 323 + ], + "type": "text", + "content": "non-conditioned adapters, with all choices enabled, row (f), obtaining the best average performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 324, + 288, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 288, + 372 + ], + "type": "text", + "content": "Rows (g) and (h) are ablation experiments evaluating the impact of choosing random task embeddings, row (g), or of taking a random choice between the 12 learned embeddings, row (h). In both cases, the performance collapses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "text", + "content": "A particularly important conclusion that can be drawn from the experiments outlined in Table 1 is that the proposed multi-task approach (row (f)) outperforms the single-task policies without adapters (row (a)). This shows that a multi-task policy can perform well on a series of tasks while being trained on a limited set of demonstrations. Figure 4 presents 3 successful test rollouts of our multi-task approach on diverse known tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 471, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 567 + ], + "type": "text", + "content": "Figure 5 visualizes the per-task test performance on known tasks of single-task policies (row (a) in Table 1), our approach without any adapter (row (b) in Table 1) and with conditioned middle and top adapters (row (f) in Table 1). The proposed adapters lead to a performance gain on most tasks compared with the solution without adapters, and the multi-task solution is competitive with single-task policies, even outperforming them on half the tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 568, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 688 + ], + "type": "text", + "content": "Known task Additional ablation studies are presented in the supplementary material (Sections B and C). Section B shows that conditioning the policy when using task-conditioned adapters is not necessary, that our aggregation layer works better than using the 'CLS' token as input to the policy and further highlights the impact of both middle and top adapters. Section C shows that our adapters improve visual embeddings extracted by two state-of-the-art pre-trained backbones, i.e. PVR [23] and MVP [27], confirming the conclusions drawn from experiments on VC-1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Known task — Visualizing the influence of task-conditioned adaptation — Figures 2 and 3 in the sup-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 298, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 545, + 346 + ], + "type": "text", + "content": "plementary material visualize the attention map of the last VC-1 layer. They show that adding adapters focuses attention on objects of interest and task-conditioning improves attention, mainly towards goals." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 348, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 546, + 445 + ], + "type": "text", + "content": "Few-shot adaptation to new tasks without finetuning model weights is an important ability of any general policy, which we evaluate with the following experiments: for each task within a set of unknown tasks (cf. Figure 2), we collect only 5 demonstrations used to optimize a task embedding specific to this task with the method detailed in section §3.4. We then evaluate the method conditioned on the optimized embedding on 100 test rollouts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 447, + 547, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 547, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 547, + 699 + ], + "type": "text", + "content": "To generate the set " + }, + { + "bbox": [ + 304, + 447, + 547, + 699 + ], + "type": "inline_equation", + "content": "T^u" + }, + { + "bbox": [ + 304, + 447, + 547, + 699 + ], + "type": "text", + "content": " of unknown tasks, we select tasks from the MetaWorld dataset that do not belong to CortexBench, and are thus not part of the set of training known tasks " + }, + { + "bbox": [ + 304, + 447, + 547, + 699 + ], + "type": "inline_equation", + "content": "T^k" + }, + { + "bbox": [ + 304, + 447, + 547, + 699 + ], + "type": "text", + "content": ". We collect demonstrations using single-task policies from TD-MPC2 [8] that were specifically trained on each task of MetaWorld independently. To ensure high-quality demonstrations, we only consider tasks where TD-MPC2 policies reach a success rate higher than " + }, + { + "bbox": [ + 304, + 447, + 547, + 699 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 447, + 547, + 699 + ], + "type": "text", + "content": ". Furthermore, to be compatible with the setup in CortexBench authors, in particular, to keep the same camera locations, we filter out the tasks where the goal is not always visible in the camera FOV. This leads to a set of 15 unknown tasks that are quite different from the tasks in the training set " + }, + { + "bbox": [ + 304, + 447, + 547, + 699 + ], + "type": "inline_equation", + "content": "T^k" + }, + { + "bbox": [ + 304, + 447, + 547, + 699 + ], + "type": "text", + "content": " as they involve different objects (handle press, faucet, plate, door, window, etc.) and types of manipulation (sliding an object, lowering a press, opening a window, etc.). Each collected demonstration is a sequence of visual frames, proprioception inputs, and expert actions. The optimization of the task embedding is performed independently for each task (cf. §3.4). We use the AdamW optimizer and a learning rate of " + }, + { + "bbox": [ + 304, + 447, + 547, + 699 + ], + "type": "inline_equation", + "content": "1e - 1" + }, + { + "bbox": [ + 304, + 447, + 547, + 699 + ], + "type": "text", + "content": " during the task embedding search." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Figure 6 presents the per-task performance in this set" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17853" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 71, + 521, + 201 + ], + "blocks": [ + { + "bbox": [ + 74, + 71, + 521, + 201 + ], + "lines": [ + { + "bbox": [ + 74, + 71, + 521, + 201 + ], + "spans": [ + { + "bbox": [ + 74, + 71, + 521, + 201 + ], + "type": "image", + "image_path": "7750e91419ec948d0b65300110badacf47393c40682a25a32738fd99c81c80ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 202, + 343, + 214 + ], + "lines": [ + { + "bbox": [ + 251, + 202, + 343, + 214 + ], + "spans": [ + { + "bbox": [ + 251, + 202, + 343, + 214 + ], + "type": "text", + "content": "(a) Successful rollouts" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 74, + 217, + 520, + 346 + ], + "blocks": [ + { + "bbox": [ + 74, + 217, + 520, + 346 + ], + "lines": [ + { + "bbox": [ + 74, + 217, + 520, + 346 + ], + "spans": [ + { + "bbox": [ + 74, + 217, + 520, + 346 + ], + "type": "image", + "image_path": "9a2493e3a4364821fc20776ca06aa3ddbed9d05f6c3b5bb273bec996ecd85ef8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 349, + 334, + 361 + ], + "lines": [ + { + "bbox": [ + 260, + 349, + 334, + 361 + ], + "spans": [ + { + "bbox": [ + 260, + 349, + 334, + 361 + ], + "type": "text", + "content": "(b) Failed rollouts" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 369, + 547, + 404 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 547, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 547, + 404 + ], + "type": "text", + "content": "Figure 7. Few-shot Qualitative results: (a) The policy tackles new tasks involving objects and/or manipulation requirements unseen during training. (b) In the first row (button-press-wall task), it performs the task correctly until the end where it fails to properly push the button fully. In the second row (push-back task), it properly moves the cube but fails to bring it to the goal position (green dot)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 424, + 289, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 289, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 289, + 592 + ], + "type": "text", + "content": "ting. Despite the large variations between the new tasks in " + }, + { + "bbox": [ + 46, + 424, + 289, + 592 + ], + "type": "inline_equation", + "content": "T^u" + }, + { + "bbox": [ + 46, + 424, + 289, + 592 + ], + "type": "text", + "content": " and the ones in the training set " + }, + { + "bbox": [ + 46, + 424, + 289, + 592 + ], + "type": "inline_equation", + "content": "(T^k)" + }, + { + "bbox": [ + 46, + 424, + 289, + 592 + ], + "type": "text", + "content": ", the multi-task policy can adapt to many of them, without requiring any weight finetuning. Interestingly, the method performs particularly well on the **Drawer Close** task, which could be related to the presence of the time-inversed **Drawer Open** task in the training set. This provides some evidence that the method can exploit regularities between tasks, which seem to be captured by the task embedding space, making it possible to generalize to unseen variations. Figure 7 (a) shows qualitative examples of successful rollouts on the new tasks. The policy manipulates new objects (faucet, plate, window) and performs new moves (rotating the faucet or sliding the plate) not seen during training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Finally, Figure 7 (b) shows failure cases on unknown tasks. As seen on the first row, the policy avoids the wall, reaches the button, and starts pushing it, but fails to push it fully. This particular behavior was also observed on other rollouts, explaining the " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " success on this button-push-wall task while mastering a part of the required sub-skills. On the second row, the policy is able to move the cube but fails to bring it to the goal location (green dot). This gives some indication of the difficulty of the few-shot generalization case: exploiting regularities in the task space requires that" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 424, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 424, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 305, + 424, + 545, + 462 + ], + "type": "text", + "content": "tasks be performed more than just approximately, as often the success metric is sparse, and rollouts only count to the metric when they are executed fully and correctly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 483, + 379, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 379, + 496 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 379, + 496 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 508, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 688 + ], + "type": "text", + "content": "Perception and action are closely tied together, and studies of human cognition have shown that a priori knowledge about a downstream task guides the visual system. We follow this direction in the context of artificial agents by introducing task-conditioned adapters that modulate the visual features of a pre-trained neural backbone. Such adapters, conditioned on a learned task embedding, improve the performance of a multi-task policy across benchmarks and embodiments. Even more interesting is the use of task embeddings to adapt in a few-shot manner, i.e. from a small set of demonstrations, to new tasks unseen at training time. We propose an optimization procedure to estimate a new task embedding and achieve generalization to unseen tasks, involving new objects and manipulation sub-skills, providing evidence for regularities in the learned embedding space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 690, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 690, + 545, + 714 + ], + "type": "text", + "content": "Acknowledgement - We thank ANR for support through AI-chair grant \"Remember\" (ANR-20-CHIA-0018)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17854" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "text", + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 224 + ], + "type": "text", + "content": "[3] T. J. Buschman and E. K. Miller. Top-down versus bottom-up control of attention in the prefrontal and posterior parietal cortices. science, 2007. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 267 + ], + "type": "text", + "content": "[4] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 268, + 287, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 268, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 268, + 287, + 302 + ], + "type": "text", + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "type": "text", + "content": "[6] M. Corbetta and G. L. Shulman. Control of goal-directed and stimulus-driven attention in the brain. Nature Reviews Neuroscience, 2002. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 287, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 400 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 402, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 287, + 434 + ], + "type": "text", + "content": "[8] Nicklas Hansen, Hao Su, and Xiaolong Wang. Td-mpc2: Scalable, robust world models for continuous control. arXiv, 2023. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 435, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 287, + 468 + ], + "type": "text", + "content": "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 468, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 287, + 513 + ], + "type": "text", + "content": "[10] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 514, + 286, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 286, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 286, + 547 + ], + "type": "text", + "content": "[11] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 547, + 286, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 547, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 547, + 286, + 590 + ], + "type": "text", + "content": "[12] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 590, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 286, + 635 + ], + "type": "text", + "content": "[13] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In CoRL, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 636, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 636, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 47, + 636, + 286, + 678 + ], + "type": "text", + "content": "[14] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 680, + 286, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 680, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 680, + 286, + 714 + ], + "type": "text", + "content": "[15] C. Kervadec, T. Jaunet, G. Antipov, M. Baccouche, R. Vuillemot, and C. Wolf. How Transferrable are Reasoning Patterns in VQA? In CVPR, 2021. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: CLIP embeddings for embodied AI. In CVPR, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 106, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 149 + ], + "type": "text", + "content": "[17] Anthony Liang, Ishika Singh, Karl Pertsch, and Jesse Thomason. Transformer adapters for robot learning. In CoRL Workshop on Pre-training Robot Learning, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "text", + "content": "[18] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized BERT pretraining approach. arXiv, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 194, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 239 + ], + "type": "text", + "content": "[19] Yecheng Jason Ma, Shagun Sodhani, Dinesh Jayaraman, Osbert Bastani, Vikash Kumar, and Amy Zhang. Vip: Towards universal visual reward and representation via value-implicit pre-training. In ICLR, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 239, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 294 + ], + "type": "text", + "content": "[20] Arjun Majumdar, Karmesh Yadav, Sergio Arnaud, Yecheng Jason Ma, Claire Chen, Sneha Silwal, Aryan Jain, Vincent-Pierre Berges, Pieter Abbeel, Jitendra Malik, et al. Where are we in the search for an artificial visual cortex for embodied intelligence? arXiv, 2023. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 294, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 327 + ], + "type": "text", + "content": "[21] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. In CoRL, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 381 + ], + "type": "text", + "content": "[22] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "type": "text", + "content": "[23] Simone Parisi, Aravind Rajeswaran, Senthil Purushwalkam, and Abhinav Gupta. The unsurprising effectiveness of pretrained vision models for control. In ICML, 2022. 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 415, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 459 + ], + "type": "text", + "content": "[24] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In NAACL-HLT, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 459, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 545, + 502 + ], + "type": "text", + "content": "[25] Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. Adapterfusion: Nondestructive task composition for transfer learning. In EACL, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 502, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 502, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 502, + 545, + 558 + ], + "type": "text", + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 558, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 558, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 558, + 545, + 592 + ], + "type": "text", + "content": "[27] Ilija Radosavovic, Tete Xiao, Stephen James, Pieter Abbeel, Jitendra Malik, and Trevor Darrell. Real-world robot learning with masked visual pre-training. In CoRL, 2023. 2, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "type": "text", + "content": "[28] Aravind Rajeswaran, Vikash Kumar, Abhishek Gupta, Giulia Vezzani, John Schulman, Emanuel Todorov, and Sergey Levine. Learning complex dexterous manipulation with deep reinforcement learning and demonstrations. RSS, 2018. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 636, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 714 + ], + "type": "text", + "content": "[29] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, Tom Eccles, Jake Bruce, Ali Razavi, Ashley Edwards, Nicolas Heess, Yutian Chen, Raia Hadsell, Oriol Vinyals, Mahyar Bordbar, and Nando de Freitas. A Generalist Agent. TMLR, 2022. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17855" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 475 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[30] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[31] Mohit Sharma, Claudio Fantacci, Yuxiang Zhou, Skanda Koppula, Nicolas Heess, Jon Scholz, and Yusuf Aytar. Lossless adaptation of pretrained vision models for robotic manipulation. In ICLR, 2022. 2, 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "type": "text", + "content": "[32] Yuval Tassa, Yotam Doron, Alistair Muldal, Tom Erez, Yazhe Li, Diego de Las Casas, David Budden, Abbas Abdelmaleki, Josh Merel, Andrew Lefrancq, et al. Deepmind control suite. arXiv, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 288, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 288, + 240 + ], + "type": "text", + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 288, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 288, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 288, + 296 + ], + "type": "text", + "content": "[34] Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 288, + 342 + ], + "type": "text", + "content": "[35] Karmesh Yadav, Ram Ramrakhya, Arjun Majumdar, Vincent-Pierre Berges, Sachit Kuhar, Dhruv Batra, Alexei Baevski, and Oleksandr Maksymets. Offline visual representation learning for embodied navigation. arXiv, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 342, + 288, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 386 + ], + "type": "text", + "content": "[36] Karmesh Yadav, Arjun Majumdar, Ram Ramrakhya, Naoki Yokoyama, Alexei Baevski, Zsolt Kira, Oleksandr Maksymets, and Dhruv Batra. Ovrl-v2: A simple state-of-art baseline forImagenav and objectnav. arXiv, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 387, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 288, + 430 + ], + "type": "text", + "content": "[37] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In CoRL, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 432, + 288, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 432, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 432, + 288, + 475 + ], + "type": "text", + "content": "[38] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023. 3" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17856" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Task-Customized Mixture of Adapters for General Image Fusion/7c9be988-491f-4562-822c-11cfddbdaa7d_content_list.json b/2024/Task-Customized Mixture of Adapters for General Image Fusion/7c9be988-491f-4562-822c-11cfddbdaa7d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0f02a7a3330afc9f65c7235a45d8dfbd357ea7f1 --- /dev/null +++ b/2024/Task-Customized Mixture of Adapters for General Image Fusion/7c9be988-491f-4562-822c-11cfddbdaa7d_content_list.json @@ -0,0 +1,1623 @@ +[ + { + "type": "text", + "text": "Task-Customized Mixture of Adapters for General Image Fusion", + "text_level": 1, + "bbox": [ + 158, + 130, + 813, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pengfei Zhu, Yang Sun, Bing Cao*, Qinghua Hu \nTianjin Key Lab of Machine Learning, College of Intelligence and Computing, Tianjin University, China \n{zhupengfei, yangsun, caobing, huqinghua}@tju.edu.cn", + "bbox": [ + 86, + 180, + 911, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "General image fusion aims at integrating important information from multi-source images. However, due to the significant cross-task gap, the respective fusion mechanism varies considerably in practice, resulting in limited performance across subtasks. To handle this problem, we propose a novel task-customized mixture of adapters (TC-MoA) for general image fusion, adaptively prompting various fusion tasks in a unified model. We borrow the insight from the mixture of experts (MoE), taking the experts as efficient tuning adapters to prompt a pre-trained foundation model. These adapters are shared across different tasks and constrained by mutual information regularization, ensuring compatibility with different tasks while complementarity for multi-source images. The task-specific routing networks customize these adapters to extract task-specific information from different sources with dynamic dominant intensity, performing adaptive visual feature prompt fusion. Notably, our TC-MoA controls the dominant intensity bias for different fusion tasks, successfully unifying multiple fusion tasks in a single model. Extensive experiments show that TC-MoA outperforms the competing approaches in learning commonalities while retaining compatibility for general image fusion (multi-modal, multi-exposure, and multifocus), and also demonstrating striking controllability on more generalization experiments. The code is available at https://github.com/YangSun22/TC-MoA.", + "bbox": [ + 76, + 301, + 473, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 726, + 209, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image fusion aims to integrate complementary information from multi-source images captured by different sensors in the same scene onto a single image. It is usually used to enhance important information and visual quality [14]. Currently, general image fusion mainly includes multi-modal, multi-exposure, and multi-focus image fusion, etc. Fusion tasks exhibit diverse fusion mechanisms. The Multi-Exposure image Fusion (MEF) focuses on integrat", + "bbox": [ + 76, + 752, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4faa96e52ba1405a5a98f0214853b96206b8e12d16bae11f5e7239a828b33a78.jpg", + "image_caption": [ + "Figure 1. Prompt can adaptively select the complementary information from multi-source features. The dominant intensity bias varies according to the task, which is reflected by the different shades of colors." + ], + "image_footnote": [], + "bbox": [ + 501, + 267, + 890, + 469 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing images with multiple exposure levels into a high-quality full exposure image [40]. Each source image contributes its own illumination and structural information to the fused image [28]. The Visible-Infrared image Fusion (VIF) is a subfield of the Multi-Modal Fusion (MMF) [8, 16, 48] that aims at fusing complementary information of infrared and visible modalities to produce robust and informative fused images [25]. Infrared images provide more intensity information, while visible images offer more texture and gradient information. The Multi-Focus image Fusion (MFF) aims at generating an all-in-focus image from a series of partially focused images [23]. A patch of multi-focus fused image typically corresponds to only one source image. Consequently, it can be observed that the MEF and VIF tasks involve relatively balanced fusion, while MFF is an extremely unbalanced task that tends to exhibit polarized choices.", + "bbox": [ + 496, + 535, + 893, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid development of deep learning techniques, image fusion has achieved great progress in recent years [31, 41], while most existing methods [28, 54, 55] focus on single image fusion scenario alone. Task-specific methods [28, 36] often employ task-specific strategies such as designing complex task-biased networks or utilizing task-specific fusion loss functions, resulting in the weak generalization to other tasks [41]. Considering that different", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7099", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fusion tasks have similar objectives, i.e., integrating complementary information from multiple source images, some recently proposed approaches [20, 41, 51] attempted to conduct various fusion tasks using a unified model, forming general image fusion. These methods, however, are either enmeshed in dominant-task bias [51] or multi-task commonality [20, 41] at the expense of individuality, leading to unsatisfactory performance. This motivates us to explore a more compatible fusion paradigm that is adaptively and dynamically compatible with different fusion scenarios.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To handle this challenge, inspired by the impressive representation capability of pre-trained foundation models, we introduce the foundation model as a fixed encoder to extract complementary features for multi-source images. Different from most existing methods, we borrow insight from the mixture of experts (MoE) [24, 33], taking each expert as an efficient tuning adapter to perform adaptive visual feature prompt fusion based on the foundation model. The task-specific routing networks customize these adapters to generate task-specific fusion prompt for different sources, forming a novel task-customized mixture of adapters (TC-MoA) architecture. An mutual information regularization is further developed to constrain the prompt, which guarantees complementarity for diverse sources. It is worth noting that the prompt have significant task bias and dominant intensity gaps. As shown in Fig. 1, the prompt of MFF has a greater color difference than VIF and MEF, meaning that feature selection has more bipolarity on dominant intensity bias. Our model effectively perceives the fusion intensity bias among varied fusion tasks in a single model and is therefore compatible with a wider range of fusion tasks. Extensive experiments verify our superiority over the competing methods in general image fusion, including multi-modal, multi-exposure, and multi-focus fusion. More importantly, our TC-MoA even shows unprecedented controllability and generalizability for unseen fusion tasks, fully demonstrating our potential in broader scenarios. We summarize our main contributions as follows:", + "bbox": [ + 75, + 244, + 472, + 667 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a unified general image fusion model, providing a new task-customized mixture of adapters (TC-MoA) for adaptive multi-source image fusion (benefiting from the dynamically aggregating effective information from the respective modalities).", + "- We propose a mutual information regularization for adapters, which allows our model to more accurately identify the dominant intensity of different source images.", + "- To the best of our knowledge, we for the first time propose an MoE-based flexible adapter for the foundation model in general image fusion. By only adding $2.8\\%$ of learnable parameters, our model copes with numerous fusion tasks. Extensive experiments demonstrate our superiority against the competing methods, while showing significant controllability and generalizability." + ], + "bbox": [ + 76, + 674, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Fusion. Image fusion focuses on generating a fused image containing complementary information from different source images. Some early methods [16, 22, 31] tackled their respective task by leveraging CNN. Then, GAN-based [43, 46] and Transformer-based methods [13, 30, 36] have been proposed to improve the fusion quality. Furthermore, high-level tasks [3, 21, 34] are also introduced to guide the fusion of images. Feature decomposition [52, 53] for high-low frequency has also gained significant attention as a research direction. Different from these methods, some most recent methods focus on general image fusion that aims to address multiple fusion tasks in a single model. Zhang et al. [51] performed a supervised training framework for MFF, which was generalized to other tasks by adjusting fusion conditions. Xu et al. [41, 42] appraised the quantity and quality of information from source images or features with the image quality assessment (IQA) to decide the fusion paradigm. PMGI [45] and SwinFusion [26] incorporated a unified fusion framework and loss function, and separate model training for individual tasks. Liang et al. [20] performed a self-supervised fusion framework to learn fusion commonality while ignoring task-specific individuality. In this work, we accommodate diverse fusion tasks by dynamically customizing the mixture of adapters, rather than suffering individuality for cross-task commonality.", + "bbox": [ + 496, + 119, + 893, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Parameter-Efficient Fine-Tuning. To efficiently adapt pre-trained models to the respective downstream application tasks, some Parameter-Efficient Fine-Tuning (PEFT) studies have been proposed. PEFT can be mainly divided into the Adapter [10, 19, 32] and the Prefix Tuning [11, 18]. Jia et al. [11] introduced prompt tuning to the ViT structure. Chen et al. [4] proposed a new adapter by scaling the original features instead of summing. Lian et al. [19] bridged the gap between the pre-trained features and the downstream task features by linearly varying the original features. In this paper, different from these methods that focus on tuning high-level tasks, we for the first time generalize the powerful foundational model to general image fusion.", + "bbox": [ + 496, + 502, + 893, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mixture of Experts. Shazeer et al. [33] first proposed the mixture of experts (MoE) to increase model capacity without increasing computational complexity. Based on this, Lepikhin et al. [15] and Fedus et al. [7] integrated MoE with the transformer structure, further pushing the upper limit of network capacity. In addition, MoE has been validated as effective in other challenges. Ma et al. [24] tackled the multi-task problem by designing multi-gates MoE. Mustafa et al. [29] exploited MoE to train the contrastive learning-based vision-language foundation model. Inspired by this, we take each expert as an adapter, forming a task-customized mixture of adapters for tuning our general image fusion framework.", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7100", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4afc4a966748aec168d990d926fbd73cbbe07952b6361c58cf20fcdcf0c7bd31.jpg", + "image_caption": [ + "Figure 2. An overview of our proposed TC-MoA method. Our approach gradually modulates the fusion results by inserting TC-MoA into the frozen ViT backbone. TC-MoA generates task-specific prompt through a task-specific router bank and an shared adapter bank. The fusion layer utilizes prompt as scale and source-specific embeddings as biases to obtain fusion images." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 893, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 398, + 169, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 422, + 187, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose a task-customized mixture of adapters for general image fusion, which is a novel parameter-efficient fine-tuning method to jointly and adaptively prompt various fusion tasks. Given a pair of source images $\\{(X,Y)|X,Y\\in \\mathbb{R}^{H\\times W\\times 3}\\}$ , the network integrates complementary information from different sources to obtain the fused image $I_{Fusion}\\in \\mathbb{R}^{H\\times W\\times 3}$ . As shown in Fig. 2, we feed the source images into the ViT network and obtain tokens of the source images through the patch embedding layer. ViT consists of an encoder for feature extraction and a decoder for image reconstruction, both of which are composed of transformer blocks. A TC-MoA is inserted into each $\\tau$ ( $\\tau = 4$ ) transformer blocks in both the encoder and decoder. Each TC-MoA consists of a task-specific router bank, a task-shared adapter bank, and a fusion layer. Our network gradually modulates the fusion results through these TC-MoA.", + "bbox": [ + 75, + 446, + 473, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Task-Customized Mixture of Adapters", + "text_level": 1, + "bbox": [ + 76, + 710, + 410, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, each TC-MoA consists of a task-specific router bank $\\{G^{\\mathcal{V}}, G^{E}, G^{F}\\}$ , a task-shared adapter bank $\\{A_{1}, \\ldots, A_{N}\\}$ and a prompting fusion layer $\\mathcal{F}$ . The TC-MoA includes two main stages: prompt generation and prompt-driven fusion. For ease of expression, we take the VIF as an example, assuming that the input comes from the VIF dataset and uses $G$ to denote $G^{V}$ .", + "bbox": [ + 75, + 734, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt Generation. First, we obtain multi-source features for subsequent processing. The network structure before the $j$ -th TC-MoA is defined as $E_{j}(\\cdot)$ , and the extracted features are defined as $f_{x}$ and $f_{y}$ , where $f \\in \\mathbb{R}^{pH \\times pW \\times C}$ .", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We concatenate $f_{x}$ and $f_{y}$ as a feature representation of pairs of multi-source tokens. This allows tokens from different sources to exchange information in the subsequent network. However, directly computing on the high-dimensional concatenated features will bring amounts of unnecessary parameters. Therefore, we use $L(\\cdot)$ to perform feature dimension reduction and obtain the processed multiple-source features $\\Phi$ as follows,", + "bbox": [ + 496, + 398, + 893, + 521 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {x} = E _ {j} (X), f _ {y} = E _ {j} (Y) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 534, + 890, + 551 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi = L \\left(\\operatorname {C a t} \\left(f _ {x}, f _ {y}\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 566, + 890, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Cat(\\cdot)$ represents the concatenation of features, $L(\\cdot)$ consists of a linear layer and normalization layers. Next, depending on the task to which $\\Phi$ belongs, we select a task-specific router from the router bank to customize the routing schemes, i.e., which adapters in the adapter bank should be input for each token pair.", + "bbox": [ + 496, + 592, + 893, + 684 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (x) = \\operatorname {S o f t m a x} (T o p K (x \\cdot W _ {g} +\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 700, + 756, + 717 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} (0, 1) \\cdot \\text {S o f t p l u s} (x \\cdot W _ {\\text {n o i s e}})) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 719, + 890, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $TopK(\\cdot)$ keeps only the top $K(K = 2)$ values, setting the rest to $-\\infty$ (after Softmax( $\\cdot$ ), the value becomes 0). $W_{g}$ and $W_{noise}$ are learnable parameters. The customization routing schemes vary for tasks, as evidenced by the proportion of the number of times that different adapters have been routed in Fig. 2 (b). After that, we weight sum the output from the adapters to obtain the prompt. Each router masters a task-biased appetite for customizing a suitable mixture of adapters, and each adapter generates the prompt. The task-customized multi-source prompt is then", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7101", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "calculated as follows,", + "bbox": [ + 76, + 90, + 223, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np r o m p t = G A P (S i g m o i d \\left(\\sum_ {i = 1} ^ {N} G (\\Phi) _ {i} \\cdot A _ {i} (\\Phi)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 112, + 468, + 154 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $GAP(\\cdot)$ represents the global average pooling, $i$ is the index of adapters and $G(\\cdot)_i$ is the routing value for the $i$ -th adapter. The task-customized prompt, denoted as prompt $\\in \\mathbb{R}^{pH\\times pW\\times 2}$ , is composed of prompt $_x$ and prompt $_y$ in $\\mathbb{R}^{pH\\times pW\\times 1}$ and has a value range of $(0,1)$ .", + "bbox": [ + 75, + 160, + 468, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompt-Driven Fusion. The task-customized prompt is constrained by mutual information regularization (MIR), which guarantees complementarity for diverse sources. Thus the prompt can be taken as an evaluation of the proportion of important information in each source. By dot-multiplication of multi-source features and prompt, we retain complementary information while removing redundant information. Afterward, considering that the feature representation should contain a source-correlated bias (e.g. visible or infrared images), we introduce input-independent learnable parameters for each source, i.e., source embedding $S$ . After feature prompting and shifting, we obtain the refined source features $(h_x, h_y)$ as follows,", + "bbox": [ + 75, + 237, + 468, + 434 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {x} = \\operatorname {p r o m p t} _ {x} \\cdot f _ {x} + S _ {x}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 441, + 361, + 457 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {y} = \\text {p r o m p t} _ {y} \\cdot f _ {y} + S _ {y} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 460, + 468, + 476 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {T C - M o A} = \\mathcal {F} \\left(h _ {x} + h _ {y}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 479, + 361, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S_{x}$ and $S_{y}$ represent the source embeddings of visible and infrared images in VIF respectively. $\\mathcal{F}(\\cdot)$ fuses the refined multi-source features using an additive operation and then passes them through a set of convolutional layers. These layers introduce local receptive fields to reduce the checkerboard artifacts and align the solution space of the subsequent transformer blocks. Ultimately, we obtain a fusion feature through the task-customized prompt. To encourage the model to extract important information gradually, the features input into the next transformer block are processed as follows,", + "bbox": [ + 75, + 501, + 468, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {x} ^ {\\prime} = \\lambda_ {f} f _ {x} + (1 - \\lambda_ {f}) f _ {T C - M o A} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 674, + 468, + 698 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {y} ^ {\\prime} = \\lambda_ {f} f _ {y} + (1 - \\lambda_ {f}) f _ {T C - M o A}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 695, + 385, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{f}$ is a learnable parameter initialized to 0.5.", + "bbox": [ + 75, + 717, + 416, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mutual Information Regularization. In order to ensure that the model dynamically retains complementary information while discarding redundant information of multisource features, we impose regularization constraints on the prompt. Assuming that the information representation of features varies linearly, we define the MIR as follows,", + "bbox": [ + 75, + 733, + 468, + 823 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min \\left| p r o m p t _ {x} + p r o m p t _ {y} - 1 \\right| \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 832, + 468, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The MIR allows the model to accurately identify the dominant intensity of different sources, which is positively correlated with the information content of the sources.", + "bbox": [ + 75, + 854, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Task-Customized Loss Function", + "text_level": 1, + "bbox": [ + 500, + 90, + 782, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to accommodating the individuality of different tasks in our network structure, we also customize unsupervised loss functions for each fusion task. We add $\\mathcal{L}_{aux}$ to the loss function of each task to ensure the training of TC-MoA and $\\mathcal{L}_{aux}$ is the auxiliary loss in [33] to avoid adapters learning unbalanced. On the other hand, in order to generate high-quality fusion images, we impose constraints on the structural information ( $\\mathcal{L}_{ssim}$ ), intensity information ( $\\mathcal{L}_{Pixel}$ ) for VIF loss function, and gradient information ( $\\mathcal{L}_{Grad}$ ) of the fusion images for different fusion tasks.", + "bbox": [ + 496, + 113, + 890, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the VIF task, our objective is to retain the most pronounced high-frequency and low-frequency information from the source images in the fused image. Thus, we design the $\\mathcal{L}_{MaxPixel}$ and $\\mathcal{L}_{MaxGrad}$ . To avoid confusing gradients, we retain the sign of the gradient values in all loss functions related to gradient information. For the MEF task, we consider that the luminance of the fused images should be at an average level with all the gradient information. Thus we design the loss functions for MEF with $\\mathcal{L}_{AvgPixel}$ and $L_{MaxGrad}$ . Additionally, we adopt $\\mathcal{L}_{mefssim}$ [28] which is specially designed for the MEF task, instead of the SSIM loss function. For the MFF task, we believe that each patch of the fused image should only depend on a single source image with the maximum gradient. This is to prevent the objects' edges in defocused images from being preserved, thereby affecting the quality of the fused image. For this purpose, we select only one of the source images to compute the loss function for each patch in the image, i.e. $\\mathcal{L}_{MaskPixel}$ and $\\mathcal{L}_{MaskGrad}$ . Please refer to the supplementary material for details of the loss functions.", + "bbox": [ + 496, + 265, + 892, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 580, + 633, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 500, + 604, + 700, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments in three image fusion scenarios: visible-infrared, multi-focus, and multi-exposure fusion. For VIF, we evaluate our model on the LLVIP [12] dataset. The training set contains 12025 image pairs, and we randomly select 70 samples out of the test set for evaluation. For MEF, we employ the SCIE [2] dataset (589 pairs) for training and MEFB [49] dataset (100 pairs) for testing, we utilize the most underexposed and overexposed images from sequences in SCIE as inputs. For MFF, we train our model on the RealMFF [47] and the MFI-WHU [46] datasets and follow the test setting in MFIF [50].", + "bbox": [ + 496, + 628, + 890, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation Details. Our experiments are performed on a server with $8 \\times$ NVIDIA RTX A6000 GPUs. Although our model supports inputs of arbitrary size, we ensure that each fusion task receives an equal amount of data per iteration by randomly cropping all samples to a size of $448 \\times 448$ . We employ the pre-trained MAE-large model [9] with GAN loss as our backbone. We train 20 epochs jointly", + "bbox": [ + 496, + 795, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7102", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/27d2a3d40010f0ab9be2cbdcfa8c5ed021a7fcc149dcfc74ed11beae7a6e065b.jpg", + "image_caption": [ + "Visible" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 176, + 175 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eb0594ec9d6039a539fccf7a93413b0a5c8e0fa7f561249c6ea914e976a7b715.jpg", + "image_caption": [ + "Infrared" + ], + "image_footnote": [], + "bbox": [ + 176, + 88, + 272, + 175 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0f3833092c0eb472b9063c6a6daeeb52cb53160277f34b0f5879ddd9cc1a5b32.jpg", + "image_caption": [ + "DenseFuse" + ], + "image_footnote": [], + "bbox": [ + 272, + 88, + 370, + 175 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8abaf1c91d80e98c37755be470db609e0684b583735d697bb7cb646559c1d955.jpg", + "image_caption": [ + "TarDAL" + ], + "image_footnote": [], + "bbox": [ + 370, + 89, + 467, + 175 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5cf1986a470e004dbb5bf2de1fd8247f6168f457ca1553d4db13851cb4ae8014.jpg", + "image_caption": [ + "CDDFuse" + ], + "image_footnote": [], + "bbox": [ + 80, + 181, + 174, + 268 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ee448dbc20e52058528f110d63dcfd9e15910c68c0c5ab910a98a7cfc276115c.jpg", + "image_caption": [ + "DDEM" + ], + "image_footnote": [], + "bbox": [ + 176, + 181, + 272, + 268 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d49a02e470178b1823db51cb56f7b70f6ebbdf9a1cc61a22fd1698a2406a5f3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 181, + 370, + 268 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6bf3b35b064e8153d4fd139e3077d40fd225d6d30000cfbc8c0eea7341576391.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 181, + 467, + 268 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/506129fe1c4083657ac6aeb11d8cbd4ea14e6c4a031c8724befa6a420ce8df9e.jpg", + "image_caption": [ + "U2Fusion", + "Figure 3. Qualitative comparisons of various methods in VIF task." + ], + "image_footnote": [], + "bbox": [ + 80, + 276, + 174, + 362 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b60ad65da7359c50b66439811fcd2aef35b8cc08bddd23ad6ca42b4d8ec92858.jpg", + "image_caption": [ + "DeFusion" + ], + "image_footnote": [], + "bbox": [ + 176, + 276, + 272, + 362 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/57fc4e3f61e7f8354da3d26091edb1927f131c230570e5ec4882ba247d56f957.jpg", + "image_caption": [ + "TC-MoA" + ], + "image_footnote": [], + "bbox": [ + 272, + 276, + 370, + 362 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f528e94175affc3405cf93e164287e6bc7f49316a7eab0424ee6860c9e4db105.jpg", + "image_caption": [ + "TC-MoA $^{YUV}$" + ], + "image_footnote": [], + "bbox": [ + 370, + 276, + 467, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with all the fusion tasks, and the batch size for each fusion task is set to 3. We adopt the AdamW optimizer with the initial learning rate of $1.5 \\times 10^{-4}$ . To ensure training stability, we apply the Exponential Moving Average (EMA) to routers and adapters optimization. Each TC-MoA consists of $N(N = 4)$ adapters embedded in different layers, but only the top $K(K = 2)$ adapters are activated.", + "bbox": [ + 75, + 407, + 467, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. Since the commonly used evaluation metrics for fusion tasks are not exactly the same, we customize a set of metrics for each task. Existing metrics can be mainly divided into the following four categories [40]. The information theory-based metrics: entropy (EN), peak signal-to-noise ratio (PSNR), mutual information (MI), normalized MI (NMI), Nonlinear correlation information entropy $(Q_{ncie})$ , feature mutual information (FMI). The image feature-based metrics: phase congruency-based metric $Q_{p}$ , standard deviation (SD), gradient-based metric $Q_{abf}$ and $Q_{g}$ . The structural similarity-based metrics: $Q_{c}$ [5], $Q_{w}$ , $Q_{s}$ [38], structural similarity (SSIM), multiscale structural similarity (MS-SSIM [37]) and MEF-SSIM [27]. The human perception inspired metrics: visual information fidelity (VIF), $Q_{cb}$ and $Q_{cv}$ . Detailed information on metrics can be found in [25] and [23].", + "bbox": [ + 75, + 513, + 467, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation on Multi-Modal Fusion", + "text_level": 1, + "bbox": [ + 76, + 768, + 379, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we compare our TC-MoA with ten task-specific image fusion methods and four general image fusion methods on the LLVIP dataset.", + "bbox": [ + 75, + 792, + 467, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Comparisons. We evaluate the fusion performance on 8 quantitative metrics, as shown in Table 1. Our method outperforms all the general image fusion methods, demonstrating our superior compatibility across multi-", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b9bffd31a2927fc9eac665236bcc437052b8d7f2f1099967b6840f949a7372d2.jpg", + "table_caption": [ + "Table 1. Quantitative results of the VIF task in LLVIP dataset. Boldface in red and boldface in blue show the best and second-best values, respectively. The underline represents the best value in the general image fusion methods." + ], + "table_footnote": [], + "table_body": "
Method\\(VIF\\)\\(Q_c\\)ENSD\\(Q_{cv}\\)↓MS-SSIMFMI\\(Q_w\\)
Densefuse [31]0.5450.5336.8309.381817.2130.8780.8760.622
AUIF [53]0.4020.3706.1377.7691087.5690.7840.8690.530
DIDFuse [52]0.3660.3485.9917.765897.0070.7670.8630.479
TarDAL [21]0.5500.5627.3479.609549.1770.8640.8600.628
YDTR [36]0.4860.5246.6388.810883.3330.8350.8760.614
RFN-Nest [17]0.4970.4567.0529.609857.1570.8620.8710.385
SwinFuse [39]0.3990.3215.8787.4571306.6520.7340.8700.425
UMF-CMGR [6]0.4140.4796.6078.520810.6700.8010.8760.561
DDFM [55]0.5880.5617.0699.696760.0060.9080.8780.663
CDDFuse [54]0.6940.6457.3429.733495.4730.9330.8830.830
SDNet [44]0.5270.5756.8979.318936.3890.8780.8720.749
IFCNN [51]0.6790.6347.2239.662521.7410.9460.8820.856
U2Fusion [41]0.5030.4926.6478.789857.4550.8780.8700.695
DeFusion [20]0.6060.6067.2169.701532.0920.8900.8800.676
TC-MoA0.7260.6377.4289.805423.7730.9490.8860.858
", + "bbox": [ + 501, + 145, + 890, + 327 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c9e8bdca026c628c10df2448b000b22284878ee08ca1d76581973a61bed20679.jpg", + "table_caption": [ + "Table 2. Quantitative results of the MEF task in MEFB [49]." + ], + "table_footnote": [], + "table_body": "
MethodPsnr\\(Q_c\\)\\(Q_p\\)\\(Q_g\\)\\(Q_s\\)MEF-SSIMFMI\\(Q_w\\)
Deepfuse [31]57.1040.4260.3520.3250.6410.8970.8730.548
MEF-GAN [43]56.9470.3090.1240.2460.4870.7720.8460.300
MEFNet [28]56.5940.6560.5950.5650.8380.9140.8900.866
IFCNN [51]57.1950.5530.5620.4780.7200.9430.8820.834
FusionDN [42]56.9770.5000.5040.4340.6720.9240.8770.776
PMGI [45]57.1170.4890.5250.4420.6660.9360.8850.804
U2Fusion [41]57.0550.4570.5050.4150.5850.9300.8820.787
DeFusion [20]57.1310.5390.3780.3760.7510.9020.8770.733
TC-MoA57.2130.5780.5980.5280.7670.9640.8880.845
", + "bbox": [ + 500, + 345, + 888, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ple fusion tasks. For most task-specific image fusion methods, TC-MoA also performs remarkable improvements, although these methods deploy complex task-specific designs. Specifically, our method achieves a significant advantage in VIF, $Q_{cv}$ , and EN metrics correlated to human perception or information theory. This indicates that our fused image conforms more to human visual perception and contains more information from the source images. The results on MS-SSIM, SD, and $Q_{w}$ metrics also explain that our fusion results possess a sufficient amount of structural information and gradient information.", + "bbox": [ + 496, + 476, + 890, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative Comparisons. As shown in Fig. 3, our model outperforms the competing methods in terms of visual quality. For instance, the contours of the tree in dark areas are well depicted and the background texture is clearer. It is worth noting that our model has the capability to directly generate color images, or we can also follow other approaches to conduct on gray images and colorize the results to obtain $\\mathrm{TC - MoA}^{YUV}$ . Compared with the competing methods, our model exhibits higher contrast between foreground and background, and more saturated colors. This comparison demonstrates our model is effective in generating fused images with superior perceptual quality.", + "bbox": [ + 496, + 642, + 890, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Evaluation on Multi-Exposure Fusion", + "text_level": 1, + "bbox": [ + 500, + 832, + 823, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we compare TC-MoA with three task-specific MEF methods i.e. Deepfuse[31], MEF-GAN[43] and MEF-Net [28], and five general image fusion meth", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7103", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3be55776a0e0e294babd779c2eb66bfff6b8bb8f12b1e6269c4b6c3ed99a5da9.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons in MEF task." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 468, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ods i.e. IFCNN [51], FusionDN [42], PMGI [45], U2Fusion [41], and DeFusion [20].", + "bbox": [ + 76, + 354, + 468, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparisons. We employ 8 quantitative metrics to evaluate our model and the competing methods on MEFB, as presented in Table 2. Our model achieves the SoTA performance in the general image fusion methods and achieves competitive results in task-specific methods. For example, our model significantly improves the MEF-SSIM scores due to the compatibility of TC-MoA with diverse tasks, enabling task-specific optimization while reducing task conflicts. The MEF-SSIM focuses on the structure and contrast distortion of images, and $Q_{p}$ measures the phase congruency of the source and fused images. The highest MEF-SSIM, PSNR, and $Q_{p}$ values indicate our effectiveness in the preservation of structural information, detailed features, and image quality in the fused images.", + "bbox": [ + 75, + 385, + 468, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparisons. As shown in Fig. 4, our method preserves the most texture details and color in both high and low illumination regions. Specifically, in the high illumination region, our model effectively retains more structural information of the clouds around the sun. In the low illumination region, the colors of the fonts on the hot air balloon are completely confused in the PMGI and U2Fusion methods. As a comparison, our method maintains detailed information while keeping more accurate color information. In fact, there is no objective standard for the brightness of the fused image for multi-exposure images. To solve this problem, our method makes the fusion of images controllable by modulating the prompt. The image of \"TC-MoA Light\" shows the result of this modulation. Detailed information about the modulation can be found in Sec. 4.6.", + "bbox": [ + 75, + 595, + 470, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Evaluation on Multi-Focus Fusion", + "text_level": 1, + "bbox": [ + 76, + 832, + 375, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare TC-MoA with two task-specific MFF methods i.e. IFCNN [51] (supervised on MFF) and MFF-GAN [46], and four general image fusion methods i.e. FusionDN [42],", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4baabd277aba43943df99b68d36323a697a056c98a527e3edc084397c757357c.jpg", + "image_caption": [ + "Figure 5. Qualitative comparisons in MFF task." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 890, + 304 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b3ffc08056e5adbdcd2eafeeed22b5871d3060c181dbc0673815b4a6fa0d559c.jpg", + "table_caption": [ + "Table 3. Quantitative results of the MFF task." + ], + "table_footnote": [], + "table_body": "
MethodNMIQncieMIQcbQcv↓MS-SSIMFMIQw
IFCNN [51]0.8470.8256.4950.69144.3730.9910.8810.912
MFF-GAN [46]0.7280.8205.6890.60972.4600.9770.8760.877
FusionDN [42]0.6750.8185.4490.519178.4910.9420.8640.794
PMGI [45]0.7020.8195.5220.542140.9570.9230.8660.575
U2Fusion [41]0.6700.8185.3290.530142.3250.9480.8670.830
DeFusion [20]0.7680.8215.8380.62583.1950.9600.8700.685
TC-MoA0.8750.8276.6950.71836.5120.9900.8810.891
", + "bbox": [ + 500, + 343, + 890, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PMGI [45], U2Fusion [41], and DeFusion [20].", + "bbox": [ + 500, + 455, + 813, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparisons. We employ 8 quantitative metrics on the dataset provided in the [50], as illustrated in Table 3. Our method shows competitive performance compared to most image fusion methods. For general image fusion, our model achieves superior performance across all metrics. Additionally, TC-MoA directly generates the fused image without predicting the decision map. Nevertheless, we still significantly outperform IFCNN in NMI, MI, Qcb, and Qcv metrics. This indicates that our fusion results, compared to supervised method, retain more source image details and demonstrate remarkable human visual perception.", + "bbox": [ + 496, + 470, + 890, + 637 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparisons. As depicted in Fig. 5, our fused image exhibits superior consistency in terms of texture and color, surpassing that of the other methods. U2Fusion exhibits color deviations in the far-focus region, while IFCNN distorts the near-focus image. Furthermore, these methods are unable to effectively remove the distortion around the flowers, while we introduce an MFF-specific loss, which achieves the most visually appealing results for the flower.", + "bbox": [ + 496, + 638, + 890, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 770, + 653, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct ablation studies to verify the effect of our TC-MoA and network architecture.", + "bbox": [ + 496, + 794, + 888, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TC-MoA. To verify the effectiveness of TC-MoA, we remove it from our framework. As shown in the first and second rows of Table 4, the models trained by multi-task training outperform those trained for a single task, based on 5 commonly used metrics. Interestingly though, this", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7104", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/aa7a9ed72a98fb2ac302948e04ecba4335435af80f8d898f2a0ff6225863360a.jpg", + "table_caption": [ + "Table 4. Ablation studies on TC-MoA. S and M represent single task training and multi-task training, while SA and MA denote single adapter and multiple adapters respectively." + ], + "table_footnote": [], + "table_body": "
TaskSMSAMA\\(Q_{\\text{abf}}\\)\\(Q_p\\)FMI\\(Q_c\\)SSIM
VIF✓✓0.59840.40730.88570.63350.4540
✓✓0.59970.41160.88620.63570.4544
✓✓0.60070.41190.88620.63650.4550
MEF✓✓0.63850.59160.88750.57540.4115
✓✓0.63620.58990.88750.57650.4106
✓✓0.64490.59800.88830.57760.4116
MFF✓✓0.65170.67330.87970.77020.6774
✓✓0.65620.67990.88060.77420.6792
✓✓0.65680.68110.88080.77550.6794
", + "bbox": [ + 78, + 131, + 467, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rule reverses in the case of MEF, indicating that inter-task competition occurs for models trained directly with multiple tasks. Comparatively, the model incorporating the multiple adapters improves the performance across all metrics, suggesting that TC-MoA is dynamically compatible with different tasks.", + "bbox": [ + 75, + 263, + 467, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adapter-based Fine-tuning. We performed ablation experiments on the backbone and adapter methods, as shown in Table 5. The fine-tuning method of \"FrozenBackbone\" represents freezing the entire backbone and only unfreezing the final linear layer. Compared to the approach of \"Frozen-Backbone\", our TC-MoA achieves significant fusion performance improvements by introducing a small number of learnable parameters. More importantly, our method outperforms the existing adapter-based visual fine-tuning approach AdaptFormer [4] on both base and large versions of pre-trained models, demonstrating the superiority of the TC-MoA structure in fusion tasks. In addition to this, Table 9 shows our base version achieves competitive performance on the VIF task, which can be further enhanced by the large version.", + "bbox": [ + 75, + 354, + 467, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Analysis and Discussion", + "text_level": 1, + "bbox": [ + 76, + 590, + 295, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficiency. We have observed the speed issues brought about by the massive parameters of the pre-trained models. To address this, we have optimized the ViT architecture by the shifted windows, which will be detailed in Sec. 8. After optimization, for multi-task with arbitrary-size inputs, Table 6 shows the base and large versions of the model have accelerated by $178\\%$ and $167\\%$ , resulting in acceptable FPS (Frames Per Second) compared with other methods. The \"Frozen\" refers to the vanilla ViT architecture.", + "bbox": [ + 75, + 613, + 467, + 748 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np r o m p t _ {x} ^ {\\prime} = \\mu + \\alpha (p r o m p t _ {x} - \\mu) + \\beta \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 758, + 467, + 785 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np r o m p t _ {y} = \\mu + \\alpha (p r o m p t _ {y} - \\mu) - \\beta\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 785, + 403, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prompt Controllability. Employing the formula noted in Eq. (8), we manipulate both the scale and shift of the prompt provided to the trained model, where $\\mu = 0.5$ represents the mean, while $\\alpha$ and $\\beta$ denote the scaling and shifting factors, respectively. We conduct the manipulation on some samples and present an example in Fig. 6 (a).", + "bbox": [ + 75, + 809, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6c9d4ca09556d98fbc0af57cc2cf3021827da587988b2530c77ebf3690e52c9a.jpg", + "table_caption": [ + "Table 5. Ablation studies on fine-tuning. $\\mathbb{B}$ and $\\mathbb{L}$ represent the use of ViT-base and ViT-large structures respectively." + ], + "table_footnote": [], + "table_body": "
ArchitectureVIFMEFMFF
\\( Q_{abf} \\)\\( Q_p \\)SSIM\\( Q_{abf} \\)\\( Q_p \\)SSIM\\( Q_{abf} \\)\\( Q_p \\)SSIM
FrozenBackbone\\( _\\text{B} \\)0.1750.0890.2750.2500.1390.2270.2950.1890.466
AdaptFormet\\( _\\text{B} \\)0.5310.3360.4270.5740.5040.3770.5790.5610.651
TC-Mo\\( A_\\text{B} \\)0.5760.3960.4500.6040.5400.3940.6120.6190.660
FrozenBackbone\\( _\\text{L} \\)0.3300.2380.3630.4520.4080.3330.4500.4280.596
AdaptFormet\\( _\\text{L} \\)0.5760.3920.4460.6160.5630.3960.6400.6500.671
TC-Mo\\( A_\\text{L} \\)0.6010.4120.4550.6450.5980.4120.6570.6810.679
", + "bbox": [ + 503, + 117, + 890, + 210 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f7c8c4fae4e6e4c7ae4e35d5761e6c57ee47f86eed902ee05716fd89360b59c8.jpg", + "table_caption": [ + "Table 6. Efficiency and scalability comparisons of approaches." + ], + "table_footnote": [], + "table_body": "
U2FDDFMCDDFrozenBTC-MoABFrozenLTC-MoAL
Total Params (M)1.32-1.78111.65115.40339.12348.70
Trainable Params (M)1.32-1.78-3.87-9.58
FPS (VIF 640×512)4.720.012.521.173.330.601.60
FPS (MEF arbitrary-size)2.21--1.373.580.711.86
FPS (MFF arbitrary-size)1.86--1.393.980.711.95
", + "bbox": [ + 503, + 231, + 890, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Obviously, as the value of $\\beta$ increases, the brightness and texture of the fused image increasingly resemble the source image $X$ , and inversely, the source image $Y$ . Thus, $\\beta$ can be considered as a shifting of the dominant intensity bias. Moreover, a higher value of $\\alpha$ tends to favor one source image over the other in the image patches, which can be considered as a scaling of the dominant intensity deviation. For example, outputs with $\\beta = -0.3$ are more similar to the under-exposure image globally. As $\\alpha$ increases from 0 to 3, the building areas of the fused image gradually become brighter, but the brightness of the clouds remains basically unchanged. This suggests that changing the scale factor of the prompt motivates the model to select different regions. Overall, this experiment demonstrates that the prompt is controllable and can be linearly transformed to obtain fuse images in different degrees.", + "bbox": [ + 496, + 309, + 890, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalizability. Our method is highly generalizable to tasks with similar fusion rules by zero-shot fusion. Specifically, we take samples from medical image fusion dataset [1] and pan-sharpening dataset generated by the Quickbird satellite, which are fused based on VIF's paradigm. We obtain a base fused image with $\\alpha = 1$ and $\\beta = 0$ , as shown in Fig. 6 (b) and (c). However, without any prior knowledge of these unknown tasks, the output fused image is an under-performing image, missing many high and low-frequency information. Interestingly, by manipulating the prompt, we can find a suitable fusion rule for these new tasks, obtaining reasonable fusion results.", + "bbox": [ + 496, + 551, + 890, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Routers Controllability. As shown in Fig. 7, the images generated by the MFF router are darker than VIF in areas without gradient guidance. The MFF selects the reference sources based on gradients, it tends to average the information from sources without gradient guidance. In contrast, the VIF router tends to preserve the maximum intensity information. Moreover, at the bottom of the image, the MFF router tends to preserve the gradient information depending on a single source, while the VIF router preserves the gradient information from both sources. Additionally, by manipulating the combination of different routers, our model", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7105", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e23e1eab439c3742adbc4d23d079208fcc54fcc98b946502f79a914d6d7ca9e.jpg", + "image_caption": [ + "Figure 6. Visualisation of controllability and generalization of our method." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 885, + 323 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c946d116226f28e58b216f15b60c42222f9c3108b2a24acfbf30c8889e77b2cb.jpg", + "image_caption": [ + "Figure 7. Visualisation of different routers on medical images." + ], + "image_footnote": [], + "bbox": [ + 81, + 345, + 464, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "copes with controlling the dynamic changing fusion degree, obtaining more controllable results.", + "bbox": [ + 76, + 530, + 468, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyperparameters. We conducted hyperparameters analysis on two aspects: the number of experts $(N)$ and the interval between two TC-MoAs $(\\tau)$ . We utilize the ViT-large as our backbone, which consists of 24 transformer blocks in the encoder and 8 in the decoder. If $\\tau = 4$ , our model contains 8 instances of TC-MoA. We employ two widely-used metrics $Q_{abf}$ and SSIM to inform the selection of hyperparameters. As presented in Fig. 8, as $N$ increases, the performance initially increases but eventually declines, with the peak performance detected at $N = 4$ . This suggests that increasing the number of experts does not necessarily enhance the model's performance. Similarly, when $\\tau = 4$ , the model demonstrates the best performance across multitask. Hence, our experiments are conducted under $N = 4$ and $\\tau = 4$ .", + "bbox": [ + 76, + 561, + 468, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 800, + 194, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a task-customized mixture of adapters for general image fusion. With the help of TCMoA, different fusion tasks customize their respective mixture of adapters to obtain prompt guidance for multiple-source image fusion in a unified model. To ensure compat", + "bbox": [ + 76, + 825, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9638f7640b7903690f40bacc9fd574573f9c7514bd2dfea28243dd51395cee1a.jpg", + "image_caption": [ + "Figure 8. Experiments on the hyperparameters." + ], + "image_footnote": [], + "bbox": [ + 511, + 345, + 887, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ibility with different tasks while maintaining complementarity for multi-source images, we further propose mutual information regularization to constrain these adapters. Experimental results have shown that TC-MoA achieves advanced performance in general image fusion against the competing methods. In addition, TC-MoA demonstrates strong prompt controllability and router controllability to perform flexible manipulation on the fused images.", + "bbox": [ + 496, + 529, + 890, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 664, + 679, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was sponsored in part by the National Key R&D Program of China 2022ZD0116500, in part by the National Natural Science Foundation of China (62222608, 62106171, U23B2049, 61925602), in part by the Haihe Lab of ITAI under Grant 22HHXCJC00002, in part by the Tianjin Natural Science Foundation under Grant 21JCYBJC00580, in part by Tianjin Natural Science Funds for Distinguished Young Scholar under Grant 23JCJQJC00270, and in part by the Key Laboratory of Big Data Intelligent Computing, Chongqing University of Posts and Telecommunications under Grant BDIC-2023-A-008. This work was also sponsored by CAAI-CANN Open Fund, developed on OpenI Community. The authors appreciate the suggestions from CVPR anonymous peer reviewers.", + "bbox": [ + 496, + 688, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7106", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aanlib. Ihttp://www.med.harvard.edu/AANLIB/home.html, (Accessed 5 January 2020). 7", + "[2] Jianrui Cai, Shuhang Gu, and Lei Zhang. Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing, 27(4):2049-2062, 2018. 4", + "[3] Bing Cao, Yiming Sun, Pengfei Zhu, and Qinghua Hu. Multi-modal gated mixture of local-to-global experts for dynamic image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 23555-23564, 2023. 2, 5", + "[4] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 2, 7", + "[5] Nedeljko Cvejic, Artur Loza, David Bull, and Nishan Canagarajah. A similarity metric for assessment of image fusion algorithms. International journal of signal processing, 2(3): 178-182, 2005. 5", + "[6] Wang Di, Liu Jinyuan, Fan Xin, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In International Joint Conference on Artificial Intelligence (IJCAI), 2022. 5", + "[7] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. The Journal of Machine Learning Research, 23(1):5232-5270, 2022. 2", + "[8] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification with dynamic evidential fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2551-2566, 2023. 1", + "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 4", + "[10] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2", + "[11] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2", + "[12] Xinyu Jia, Chuang Zhu, Minzhen Li, Wenqi Tang, and Wenli Zhou. Llvip: A visible-infrared paired dataset for low-light vision. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3496-3504, 2021. 4", + "[13] Xin Jin, Xiuliang Xi, Ding Zhou, Xiaoxuan Ren, Jie Yang, and Qian Jiang. An unsupervised multi-focus image fusion method based on transformer and u-net. IET Image Processing, 17(3):733-746, 2023. 2", + "[14] Harpreet Kaur, Deepika Koundal, and Virender Kadyan. Image fusion techniques: a survey. Archives of computational methods in Engineering, 28:4425-4447, 2021. 1" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. CoRR, abs/2006.16668, 2020. 2", + "[16] Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images. IEEE Transactions on Image Processing, 28(5):2614-2623, 2018. 1, 2", + "[17] Hui Li, Xiao-Jun Wu, and Josef Kittler. Rfn-nest: An end-to-end residual fusion network for infrared and visible images. Information Fusion, 73:72-86, 2021. 5", + "[18] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2", + "[19] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems, 35:109-123, 2022. 2", + "[20] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Fusion from decomposition: A self-supervised decomposition approach for image fusion. In European Conference on Computer Vision, pages 719-735. Springer, 2022. 2, 5, 6", + "[21] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5811, 2022. 2, 5", + "[22] Yu Liu, Xun Chen, Hu Peng, and Zengfu Wang. Multi-focus image fusion with a deep convolutional neural network. Information Fusion, 36:191-207, 2017. 2", + "[23] Yu Liu, Lei Wang, Juan Cheng, Chang Li, and Xun Chen. Multi-focus image fusion: A survey of the state of the art. Information Fusion, 64:71-91, 2020. 1, 5", + "[24] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2", + "[25] Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible image fusion methods and applications: A survey. Information fusion, 45:153-178, 2019. 1, 5", + "[26] Jiayi Ma, Linfeng Tang, Fan Fan, Jun Huang, Xiaoguang Mei, and Yong Ma. Swinfusion: Cross-domain long-range learning for general image fusion via swim transformer. IEEE/CAA Journal of Automatica Sinica, 9(7):1200-1217, 2022. 2", + "[27] Kede Ma, Kai Zeng, and Zhou Wang. Perceptual quality assessment for multi-exposure image fusion. IEEE Transactions on Image Processing, 24(11):3345-3356, 2015. 5", + "[28] Kede Ma, Zhengfang Duanmu, Hanwei Zhu, Yuming Fang, and Zhou Wang. Deep guided learning for fast multi-exposure image fusion. IEEE Transactions on Image Processing, 29:2808-2819, 2019. 1, 4, 5" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7107", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with limoe: the language-image mixture of experts. Advances in Neural Information Processing Systems, 35:9564-9576, 2022. 2", + "[30] Linhao Qu, Shaolei Liu, Manning Wang, and Zhijian Song. Transmef: A transformer-based multi-exposure image fusion framework using self-supervised multi-task learning. In Proceedings of the AAAI conference on artificial intelligence, pages 2126-2134, 2022. 2", + "[31] K Ram Prabhakar, V Sai Srikar, and R Venkatesh Babu. Deepfuse: A deep unsupervised approach for exposure fusion with extreme exposure image pairs. In Proceedings of the IEEE international conference on computer vision, pages 4714-4722, 2017. 1, 2, 5", + "[32] Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual adapters. Advances in neural information processing systems, 30, 2017. 2", + "[33] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 2, 4", + "[34] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Det-fusion: A detection-driven infrared and visible image fusion network. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4003-4011, 2022. 2", + "[35] Linfeng Tang, Jiteng Yuan, and Jiayi Ma. Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network. Information Fusion, 82:28-42, 2022. 1", + "[36] Wei Tang, Fazhi He, and Yu Liu. Ydtr: Infrared and visible image fusion via y-shape dynamic transformer. IEEE Transactions on Multimedia, 2022. 1, 2, 5", + "[37] Z. Wang, E.P. Simoncelli, and A.C. Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402 Vol.2, 2003. 5", + "[38] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5", + "[39] Zhishe Wang, Yanlin Chen, Wenyu Shao, Hui Li, and Lei Zhang. Swinfuse: A residual swim transformer fusion network for infrared and visible images. IEEE Transactions on Instrumentation and Measurement, 71:1-12, 2022. 5", + "[40] Fang Xu, Jinghong Liu, Yueming Song, Hui Sun, and Xuan Wang. Multi-exposure image fusion techniques: A comprehensive review. Remote Sensing, 14(3):771, 2022. 1, 5", + "[41] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2020. 1, 2, 5, 6", + "[42] Han Xu, Jiayi Ma, Zhuliang Le, Junjun Jiang, and Xiaojie Guo. Fusiondn: A unified densely connected network for image fusion. In Proceedings of the AAAI conference on artificial intelligence, pages 12484-12491, 2020. 2, 5, 6" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Han Xu, Jiayi Ma, and Xiao-Ping Zhang. Mef-gan: Multiexposure image fusion via generative adversarial networks. IEEE Transactions on Image Processing, 29:7203-7216, 2020. 2, 5", + "[44] Hao Zhang and Jiayi Ma. Sdnet: A versatile squeeze-and-decomposition network for real-time image fusion. International Journal of Computer Vision, 129:2761-2785, 2021. 5", + "[45] Hao Zhang, Han Xu, Yang Xiao, Xiaojie Guo, and Jiayi Ma. Rethinking the image fusion: A fast unified image fusion network based on proportional maintenance of gradient and intensity. In Proceedings of the AAAI conference on artificial intelligence, pages 12797-12804, 2020. 2, 5, 6", + "[46] Hao Zhang, Zhuliang Le, Zhenfeng Shao, Han Xu, and Jiayi Ma. Mff-gan: An unsupervised generative adversarial network with adaptive and gradient joint constraints for multifocus image fusion. Information Fusion, 66:40-53, 2021. 2, 4, 6", + "[47] Juncheng Zhang, Qingmin Liao, Shaojun Liu, Haoyu Ma, Wenming Yang, and Jing-Hao Xue. Real-mff: A large realistic multi-focus image dataset with ground truth. Pattern Recognition Letters, 138:370-377, 2020. 4", + "[48] Qingyang Zhang, Haitao Wu, Changqing Zhang, Qinghua Hu, Huazhu Fu, Joey Tianyi Zhou, and Xi Peng. Provable dynamic fusion for low-quality multimodal data. In International Conference on Machine Learning, 2023. 1", + "[49] Xingchen Zhang. Benchmarking and comparing multi-exposure image fusion algorithms. Information Fusion, 74: 111-131, 2021. 4, 5", + "[50] Xingchen Zhang. Deep learning-based multi-focus image fusion: A survey and a comparative study. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 4, 6", + "[51] Yu Zhang, Yu Liu, Peng Sun, Han Yan, Xiaolin Zhao, and Li Zhang. Ifcnn: A general image fusion framework based on convolutional neural network. Information Fusion, 54: 99-118, 2020. 2, 5, 6", + "[52] Zixiang Zhao, Shuang Xu, Chunxia Zhang, Junmin Liu, Pengfei Li, and Jiangshe Zhang. Didfuse: Deep image decomposition for infrared and visible image fusion. arXiv preprint arXiv:2003.09210, 2020. 2, 5", + "[53] Zixiang Zhao, Shuang Xu, Jiangshe Zhang, Chengyang Liang, Chunxia Zhang, and Junmin Liu. Efficient and model-based infrared and visible image fusion via algorithm unrolling. IEEE Transactions on Circuits and Systems for Video Technology, 32(3):1186-1196, 2021. 2, 5", + "[54] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5906–5916, 2023. 1, 5", + "[55] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. Ddfm: denoising diffusion model for multi-modality image fusion. arXiv preprint arXiv:2303.06840, 2023. 1, 5" + ], + "bbox": [ + 501, + 92, + 890, + 868 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7108", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Task-Customized Mixture of Adapters for General Image Fusion/7c9be988-491f-4562-822c-11cfddbdaa7d_model.json b/2024/Task-Customized Mixture of Adapters for General Image Fusion/7c9be988-491f-4562-822c-11cfddbdaa7d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..92f3a56be13ae915796adb4ee7d0e1ca661c3c4d --- /dev/null +++ b/2024/Task-Customized Mixture of Adapters for General Image Fusion/7c9be988-491f-4562-822c-11cfddbdaa7d_model.json @@ -0,0 +1,2343 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.159, + 0.131, + 0.815, + 0.154 + ], + "angle": 0, + "content": "Task-Customized Mixture of Adapters for General Image Fusion" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.181, + 0.912, + 0.234 + ], + "angle": 0, + "content": "Pengfei Zhu, Yang Sun, Bing Cao*, Qinghua Hu \nTianjin Key Lab of Machine Learning, College of Intelligence and Computing, Tianjin University, China \n{zhupengfei, yangsun, caobing, huqinghua}@tju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.302, + 0.474, + 0.695 + ], + "angle": 0, + "content": "General image fusion aims at integrating important information from multi-source images. However, due to the significant cross-task gap, the respective fusion mechanism varies considerably in practice, resulting in limited performance across subtasks. To handle this problem, we propose a novel task-customized mixture of adapters (TC-MoA) for general image fusion, adaptively prompting various fusion tasks in a unified model. We borrow the insight from the mixture of experts (MoE), taking the experts as efficient tuning adapters to prompt a pre-trained foundation model. These adapters are shared across different tasks and constrained by mutual information regularization, ensuring compatibility with different tasks while complementarity for multi-source images. The task-specific routing networks customize these adapters to extract task-specific information from different sources with dynamic dominant intensity, performing adaptive visual feature prompt fusion. Notably, our TC-MoA controls the dominant intensity bias for different fusion tasks, successfully unifying multiple fusion tasks in a single model. Extensive experiments show that TC-MoA outperforms the competing approaches in learning commonalities while retaining compatibility for general image fusion (multi-modal, multi-exposure, and multifocus), and also demonstrating striking controllability on more generalization experiments. The code is available at https://github.com/YangSun22/TC-MoA." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.727, + 0.21, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.753, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Image fusion aims to integrate complementary information from multi-source images captured by different sensors in the same scene onto a single image. It is usually used to enhance important information and visual quality [14]. Currently, general image fusion mainly includes multi-modal, multi-exposure, and multi-focus image fusion, etc. Fusion tasks exhibit diverse fusion mechanisms. The Multi-Exposure image Fusion (MEF) focuses on integrat" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.268, + 0.892, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.472, + 0.892, + 0.526 + ], + "angle": 0, + "content": "Figure 1. Prompt can adaptively select the complementary information from multi-source features. The dominant intensity bias varies according to the task, which is reflected by the different shades of colors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.895, + 0.779 + ], + "angle": 0, + "content": "ing images with multiple exposure levels into a high-quality full exposure image [40]. Each source image contributes its own illumination and structural information to the fused image [28]. The Visible-Infrared image Fusion (VIF) is a subfield of the Multi-Modal Fusion (MMF) [8, 16, 48] that aims at fusing complementary information of infrared and visible modalities to produce robust and informative fused images [25]. Infrared images provide more intensity information, while visible images offer more texture and gradient information. The Multi-Focus image Fusion (MFF) aims at generating an all-in-focus image from a series of partially focused images [23]. A patch of multi-focus fused image typically corresponds to only one source image. Consequently, it can be observed that the MEF and VIF tasks involve relatively balanced fusion, while MFF is an extremely unbalanced task that tends to exhibit polarized choices." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "With the rapid development of deep learning techniques, image fusion has achieved great progress in recent years [31, 41], while most existing methods [28, 54, 55] focus on single image fusion scenario alone. Task-specific methods [28, 36] often employ task-specific strategies such as designing complex task-biased networks or utilizing task-specific fusion loss functions, resulting in the weak generalization to other tasks [41]. Considering that different" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7099" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "fusion tasks have similar objectives, i.e., integrating complementary information from multiple source images, some recently proposed approaches [20, 41, 51] attempted to conduct various fusion tasks using a unified model, forming general image fusion. These methods, however, are either enmeshed in dominant-task bias [51] or multi-task commonality [20, 41] at the expense of individuality, leading to unsatisfactory performance. This motivates us to explore a more compatible fusion paradigm that is adaptively and dynamically compatible with different fusion scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.245, + 0.473, + 0.668 + ], + "angle": 0, + "content": "To handle this challenge, inspired by the impressive representation capability of pre-trained foundation models, we introduce the foundation model as a fixed encoder to extract complementary features for multi-source images. Different from most existing methods, we borrow insight from the mixture of experts (MoE) [24, 33], taking each expert as an efficient tuning adapter to perform adaptive visual feature prompt fusion based on the foundation model. The task-specific routing networks customize these adapters to generate task-specific fusion prompt for different sources, forming a novel task-customized mixture of adapters (TC-MoA) architecture. An mutual information regularization is further developed to constrain the prompt, which guarantees complementarity for diverse sources. It is worth noting that the prompt have significant task bias and dominant intensity gaps. As shown in Fig. 1, the prompt of MFF has a greater color difference than VIF and MEF, meaning that feature selection has more bipolarity on dominant intensity bias. Our model effectively perceives the fusion intensity bias among varied fusion tasks in a single model and is therefore compatible with a wider range of fusion tasks. Extensive experiments verify our superiority over the competing methods in general image fusion, including multi-modal, multi-exposure, and multi-focus fusion. More importantly, our TC-MoA even shows unprecedented controllability and generalizability for unseen fusion tasks, fully demonstrating our potential in broader scenarios. We summarize our main contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.675, + 0.47, + 0.749 + ], + "angle": 0, + "content": "- We propose a unified general image fusion model, providing a new task-customized mixture of adapters (TC-MoA) for adaptive multi-source image fusion (benefiting from the dynamically aggregating effective information from the respective modalities)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.75, + 0.47, + 0.796 + ], + "angle": 0, + "content": "- We propose a mutual information regularization for adapters, which allows our model to more accurately identify the dominant intensity of different source images." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.797, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- To the best of our knowledge, we for the first time propose an MoE-based flexible adapter for the foundation model in general image fusion. By only adding \\(2.8\\%\\) of learnable parameters, our model copes with numerous fusion tasks. Extensive experiments demonstrate our superiority against the competing methods, while showing significant controllability and generalizability." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.675, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.121, + 0.895, + 0.5 + ], + "angle": 0, + "content": "Image Fusion. Image fusion focuses on generating a fused image containing complementary information from different source images. Some early methods [16, 22, 31] tackled their respective task by leveraging CNN. Then, GAN-based [43, 46] and Transformer-based methods [13, 30, 36] have been proposed to improve the fusion quality. Furthermore, high-level tasks [3, 21, 34] are also introduced to guide the fusion of images. Feature decomposition [52, 53] for high-low frequency has also gained significant attention as a research direction. Different from these methods, some most recent methods focus on general image fusion that aims to address multiple fusion tasks in a single model. Zhang et al. [51] performed a supervised training framework for MFF, which was generalized to other tasks by adjusting fusion conditions. Xu et al. [41, 42] appraised the quantity and quality of information from source images or features with the image quality assessment (IQA) to decide the fusion paradigm. PMGI [45] and SwinFusion [26] incorporated a unified fusion framework and loss function, and separate model training for individual tasks. Liang et al. [20] performed a self-supervised fusion framework to learn fusion commonality while ignoring task-specific individuality. In this work, we accommodate diverse fusion tasks by dynamically customizing the mixture of adapters, rather than suffering individuality for cross-task commonality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.895, + 0.701 + ], + "angle": 0, + "content": "Parameter-Efficient Fine-Tuning. To efficiently adapt pre-trained models to the respective downstream application tasks, some Parameter-Efficient Fine-Tuning (PEFT) studies have been proposed. PEFT can be mainly divided into the Adapter [10, 19, 32] and the Prefix Tuning [11, 18]. Jia et al. [11] introduced prompt tuning to the ViT structure. Chen et al. [4] proposed a new adapter by scaling the original features instead of summing. Lian et al. [19] bridged the gap between the pre-trained features and the downstream task features by linearly varying the original features. In this paper, different from these methods that focus on tuning high-level tasks, we for the first time generalize the powerful foundational model to general image fusion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Mixture of Experts. Shazeer et al. [33] first proposed the mixture of experts (MoE) to increase model capacity without increasing computational complexity. Based on this, Lepikhin et al. [15] and Fedus et al. [7] integrated MoE with the transformer structure, further pushing the upper limit of network capacity. In addition, MoE has been validated as effective in other challenges. Ma et al. [24] tackled the multi-task problem by designing multi-gates MoE. Mustafa et al. [29] exploited MoE to train the contrastive learning-based vision-language foundation model. Inspired by this, we take each expert as an adapter, forming a task-customized mixture of adapters for tuning our general image fusion framework." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7100" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.895, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.344, + 0.895, + 0.388 + ], + "angle": 0, + "content": "Figure 2. An overview of our proposed TC-MoA method. Our approach gradually modulates the fusion results by inserting TC-MoA into the frozen ViT backbone. TC-MoA generates task-specific prompt through a task-specific router bank and an shared adapter bank. The fusion layer utilizes prompt as scale and source-specific embeddings as biases to obtain fusion images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.399, + 0.17, + 0.415 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.423, + 0.188, + 0.438 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.474, + 0.705 + ], + "angle": 0, + "content": "In this paper, we propose a task-customized mixture of adapters for general image fusion, which is a novel parameter-efficient fine-tuning method to jointly and adaptively prompt various fusion tasks. Given a pair of source images \\(\\{(X,Y)|X,Y\\in \\mathbb{R}^{H\\times W\\times 3}\\}\\), the network integrates complementary information from different sources to obtain the fused image \\(I_{Fusion}\\in \\mathbb{R}^{H\\times W\\times 3}\\). As shown in Fig. 2, we feed the source images into the ViT network and obtain tokens of the source images through the patch embedding layer. ViT consists of an encoder for feature extraction and a decoder for image reconstruction, both of which are composed of transformer blocks. A TC-MoA is inserted into each \\(\\tau\\) (\\(\\tau = 4\\)) transformer blocks in both the encoder and decoder. Each TC-MoA consists of a task-specific router bank, a task-shared adapter bank, and a fusion layer. Our network gradually modulates the fusion results through these TC-MoA." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.411, + 0.729 + ], + "angle": 0, + "content": "3.2. Task-Customized Mixture of Adapters" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.84 + ], + "angle": 0, + "content": "As shown in Fig. 2, each TC-MoA consists of a task-specific router bank \\(\\{G^{\\mathcal{V}}, G^{E}, G^{F}\\}\\), a task-shared adapter bank \\(\\{A_{1}, \\ldots, A_{N}\\}\\) and a prompting fusion layer \\(\\mathcal{F}\\). The TC-MoA includes two main stages: prompt generation and prompt-driven fusion. For ease of expression, we take the VIF as an example, assuming that the input comes from the VIF dataset and uses \\(G\\) to denote \\(G^{V}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Prompt Generation. First, we obtain multi-source features for subsequent processing. The network structure before the \\(j\\)-th TC-MoA is defined as \\(E_{j}(\\cdot)\\), and the extracted features are defined as \\(f_{x}\\) and \\(f_{y}\\), where \\(f \\in \\mathbb{R}^{pH \\times pW \\times C}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.895, + 0.522 + ], + "angle": 0, + "content": "We concatenate \\( f_{x} \\) and \\( f_{y} \\) as a feature representation of pairs of multi-source tokens. This allows tokens from different sources to exchange information in the subsequent network. However, directly computing on the high-dimensional concatenated features will bring amounts of unnecessary parameters. Therefore, we use \\( L(\\cdot) \\) to perform feature dimension reduction and obtain the processed multiple-source features \\( \\Phi \\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.535, + 0.892, + 0.553 + ], + "angle": 0, + "content": "\\[\nf _ {x} = E _ {j} (X), f _ {y} = E _ {j} (Y) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.567, + 0.892, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\Phi = L \\left(\\operatorname {C a t} \\left(f _ {x}, f _ {y}\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.894, + 0.685 + ], + "angle": 0, + "content": "where \\(Cat(\\cdot)\\) represents the concatenation of features, \\(L(\\cdot)\\) consists of a linear layer and normalization layers. Next, depending on the task to which \\(\\Phi\\) belongs, we select a task-specific router from the router bank to customize the routing schemes, i.e., which adapters in the adapter bank should be input for each token pair." + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.701, + 0.758, + 0.718 + ], + "angle": 0, + "content": "\\[\nG (x) = \\operatorname {S o f t m a x} (T o p K (x \\cdot W _ {g} +\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.72, + 0.892, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} (0, 1) \\cdot \\text {S o f t p l u s} (x \\cdot W _ {\\text {n o i s e}})) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where \\(TopK(\\cdot)\\) keeps only the top \\(K(K = 2)\\) values, setting the rest to \\(-\\infty\\) (after Softmax(\\(\\cdot\\)), the value becomes 0). \\(W_{g}\\) and \\(W_{noise}\\) are learnable parameters. The customization routing schemes vary for tasks, as evidenced by the proportion of the number of times that different adapters have been routed in Fig. 2 (b). After that, we weight sum the output from the adapters to obtain the prompt. Each router masters a task-biased appetite for customizing a suitable mixture of adapters, and each adapter generates the prompt. The task-customized multi-source prompt is then" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7101" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.224, + 0.106 + ], + "angle": 0, + "content": "calculated as follows," + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.113, + 0.47, + 0.155 + ], + "angle": 0, + "content": "\\[\np r o m p t = G A P (S i g m o i d \\left(\\sum_ {i = 1} ^ {N} G (\\Phi) _ {i} \\cdot A _ {i} (\\Phi)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.161, + 0.469, + 0.237 + ], + "angle": 0, + "content": "where \\(GAP(\\cdot)\\) represents the global average pooling, \\(i\\) is the index of adapters and \\(G(\\cdot)_i\\) is the routing value for the \\(i\\)-th adapter. The task-customized prompt, denoted as prompt \\(\\in \\mathbb{R}^{pH\\times pW\\times 2}\\), is composed of prompt\\(_x\\) and prompt\\(_y\\) in \\(\\mathbb{R}^{pH\\times pW\\times 1}\\) and has a value range of \\((0,1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.238, + 0.47, + 0.435 + ], + "angle": 0, + "content": "Prompt-Driven Fusion. The task-customized prompt is constrained by mutual information regularization (MIR), which guarantees complementarity for diverse sources. Thus the prompt can be taken as an evaluation of the proportion of important information in each source. By dot-multiplication of multi-source features and prompt, we retain complementary information while removing redundant information. Afterward, considering that the feature representation should contain a source-correlated bias (e.g. visible or infrared images), we introduce input-independent learnable parameters for each source, i.e., source embedding \\( S \\). After feature prompting and shifting, we obtain the refined source features \\( (h_x, h_y) \\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.442, + 0.362, + 0.458 + ], + "angle": 0, + "content": "\\[\nh _ {x} = \\operatorname {p r o m p t} _ {x} \\cdot f _ {x} + S _ {x}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.461, + 0.469, + 0.477 + ], + "angle": 0, + "content": "\\[\nh _ {y} = \\text {p r o m p t} _ {y} \\cdot f _ {y} + S _ {y} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.48, + 0.362, + 0.496 + ], + "angle": 0, + "content": "\\[\nf _ {T C - M o A} = \\mathcal {F} \\left(h _ {x} + h _ {y}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.469, + 0.668 + ], + "angle": 0, + "content": "where \\( S_{x} \\) and \\( S_{y} \\) represent the source embeddings of visible and infrared images in VIF respectively. \\( \\mathcal{F}(\\cdot) \\) fuses the refined multi-source features using an additive operation and then passes them through a set of convolutional layers. These layers introduce local receptive fields to reduce the checkerboard artifacts and align the solution space of the subsequent transformer blocks. Ultimately, we obtain a fusion feature through the task-customized prompt. To encourage the model to extract important information gradually, the features input into the next transformer block are processed as follows," + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.675, + 0.469, + 0.699 + ], + "angle": 0, + "content": "\\[\nf _ {x} ^ {\\prime} = \\lambda_ {f} f _ {x} + (1 - \\lambda_ {f}) f _ {T C - M o A} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.696, + 0.387, + 0.712 + ], + "angle": 0, + "content": "\\[\nf _ {y} ^ {\\prime} = \\lambda_ {f} f _ {y} + (1 - \\lambda_ {f}) f _ {T C - M o A}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.418, + 0.733 + ], + "angle": 0, + "content": "where \\(\\lambda_{f}\\) is a learnable parameter initialized to 0.5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Mutual Information Regularization. In order to ensure that the model dynamically retains complementary information while discarding redundant information of multisource features, we impose regularization constraints on the prompt. Assuming that the information representation of features varies linearly, we define the MIR as follows," + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.833, + 0.469, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\min \\left| p r o m p t _ {x} + p r o m p t _ {y} - 1 \\right| \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.9 + ], + "angle": 0, + "content": "The MIR allows the model to accurately identify the dominant intensity of different sources, which is positively correlated with the information content of the sources." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.783, + 0.106 + ], + "angle": 0, + "content": "3.3. Task-Customized Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.265 + ], + "angle": 0, + "content": "In addition to accommodating the individuality of different tasks in our network structure, we also customize unsupervised loss functions for each fusion task. We add \\(\\mathcal{L}_{aux}\\) to the loss function of each task to ensure the training of TC-MoA and \\(\\mathcal{L}_{aux}\\) is the auxiliary loss in [33] to avoid adapters learning unbalanced. On the other hand, in order to generate high-quality fusion images, we impose constraints on the structural information (\\(\\mathcal{L}_{ssim}\\)), intensity information (\\(\\mathcal{L}_{Pixel}\\)) for VIF loss function, and gradient information (\\(\\mathcal{L}_{Grad}\\)) of the fusion images for different fusion tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.893, + 0.568 + ], + "angle": 0, + "content": "For the VIF task, our objective is to retain the most pronounced high-frequency and low-frequency information from the source images in the fused image. Thus, we design the \\(\\mathcal{L}_{MaxPixel}\\) and \\(\\mathcal{L}_{MaxGrad}\\). To avoid confusing gradients, we retain the sign of the gradient values in all loss functions related to gradient information. For the MEF task, we consider that the luminance of the fused images should be at an average level with all the gradient information. Thus we design the loss functions for MEF with \\(\\mathcal{L}_{AvgPixel}\\) and \\(L_{MaxGrad}\\). Additionally, we adopt \\(\\mathcal{L}_{mefssim}\\) [28] which is specially designed for the MEF task, instead of the SSIM loss function. For the MFF task, we believe that each patch of the fused image should only depend on a single source image with the maximum gradient. This is to prevent the objects' edges in defocused images from being preserved, thereby affecting the quality of the fused image. For this purpose, we select only one of the source images to compute the loss function for each patch in the image, i.e. \\(\\mathcal{L}_{MaskPixel}\\) and \\(\\mathcal{L}_{MaskGrad}\\). Please refer to the supplementary material for details of the loss functions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.581, + 0.634, + 0.598 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.701, + 0.623 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Datasets. We conduct experiments in three image fusion scenarios: visible-infrared, multi-focus, and multi-exposure fusion. For VIF, we evaluate our model on the LLVIP [12] dataset. The training set contains 12025 image pairs, and we randomly select 70 samples out of the test set for evaluation. For MEF, we employ the SCIE [2] dataset (589 pairs) for training and MEFB [49] dataset (100 pairs) for testing, we utilize the most underexposed and overexposed images from sequences in SCIE as inputs. For MFF, we train our model on the RealMFF [47] and the MFI-WHU [46] datasets and follow the test setting in MFIF [50]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. Our experiments are performed on a server with \\(8 \\times\\) NVIDIA RTX A6000 GPUs. Although our model supports inputs of arbitrary size, we ensure that each fusion task receives an equal amount of data per iteration by randomly cropping all samples to a size of \\(448 \\times 448\\). We employ the pre-trained MAE-large model [9] with GAN loss as our backbone. We train 20 epochs jointly" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7102" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.177, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.089, + 0.273, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.089, + 0.371, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.09, + 0.468, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.176, + 0.142, + 0.183 + ], + "angle": 0, + "content": "Visible" + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.176, + 0.242, + 0.183 + ], + "angle": 0, + "content": "Infrared" + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.176, + 0.343, + 0.183 + ], + "angle": 0, + "content": "DenseFuse" + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.176, + 0.434, + 0.183 + ], + "angle": 0, + "content": "TarDAL" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.183, + 0.176, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.177, + 0.183, + 0.273, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.183, + 0.371, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.183, + 0.468, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.27, + 0.149, + 0.276 + ], + "angle": 0, + "content": "CDDFuse" + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.27, + 0.24, + 0.276 + ], + "angle": 0, + "content": "DDEM" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.277, + 0.176, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.364, + 0.149, + 0.371 + ], + "angle": 0, + "content": "U2Fusion" + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.277, + 0.273, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.364, + 0.245, + 0.371 + ], + "angle": 0, + "content": "DeFusion" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.277, + 0.371, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.305, + 0.364, + 0.34, + 0.371 + ], + "angle": 0, + "content": "TC-MoA" + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.277, + 0.468, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.364, + 0.443, + 0.371 + ], + "angle": 0, + "content": "TC-MoA\\(^{YUV}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.381, + 0.468, + 0.395 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons of various methods in VIF task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.468, + 0.514 + ], + "angle": 0, + "content": "with all the fusion tasks, and the batch size for each fusion task is set to 3. We adopt the AdamW optimizer with the initial learning rate of \\(1.5 \\times 10^{-4}\\). To ensure training stability, we apply the Exponential Moving Average (EMA) to routers and adapters optimization. Each TC-MoA consists of \\(N(N = 4)\\) adapters embedded in different layers, but only the top \\(K(K = 2)\\) adapters are activated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.468, + 0.756 + ], + "angle": 0, + "content": "Evaluation Metrics. Since the commonly used evaluation metrics for fusion tasks are not exactly the same, we customize a set of metrics for each task. Existing metrics can be mainly divided into the following four categories [40]. The information theory-based metrics: entropy (EN), peak signal-to-noise ratio (PSNR), mutual information (MI), normalized MI (NMI), Nonlinear correlation information entropy \\((Q_{ncie})\\), feature mutual information (FMI). The image feature-based metrics: phase congruency-based metric \\(Q_{p}\\), standard deviation (SD), gradient-based metric \\(Q_{abf}\\) and \\(Q_{g}\\). The structural similarity-based metrics: \\(Q_{c}\\) [5], \\(Q_{w}\\), \\(Q_{s}\\) [38], structural similarity (SSIM), multiscale structural similarity (MS-SSIM [37]) and MEF-SSIM [27]. The human perception inspired metrics: visual information fidelity (VIF), \\(Q_{cb}\\) and \\(Q_{cv}\\). Detailed information on metrics can be found in [25] and [23]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.769, + 0.38, + 0.784 + ], + "angle": 0, + "content": "4.2. Evaluation on Multi-Modal Fusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.794, + 0.468, + 0.838 + ], + "angle": 0, + "content": "In this section, we compare our TC-MoA with ten task-specific image fusion methods and four general image fusion methods on the LLVIP dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Quantitative Comparisons. We evaluate the fusion performance on 8 quantitative metrics, as shown in Table 1. Our method outperforms all the general image fusion methods, demonstrating our superior compatibility across multi-" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.144 + ], + "angle": 0, + "content": "Table 1. Quantitative results of the VIF task in LLVIP dataset. Boldface in red and boldface in blue show the best and second-best values, respectively. The underline represents the best value in the general image fusion methods." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.146, + 0.892, + 0.328 + ], + "angle": 0, + "content": "
Method\\(VIF\\)\\(Q_c\\)ENSD\\(Q_{cv}\\)↓MS-SSIMFMI\\(Q_w\\)
Densefuse [31]0.5450.5336.8309.381817.2130.8780.8760.622
AUIF [53]0.4020.3706.1377.7691087.5690.7840.8690.530
DIDFuse [52]0.3660.3485.9917.765897.0070.7670.8630.479
TarDAL [21]0.5500.5627.3479.609549.1770.8640.8600.628
YDTR [36]0.4860.5246.6388.810883.3330.8350.8760.614
RFN-Nest [17]0.4970.4567.0529.609857.1570.8620.8710.385
SwinFuse [39]0.3990.3215.8787.4571306.6520.7340.8700.425
UMF-CMGR [6]0.4140.4796.6078.520810.6700.8010.8760.561
DDFM [55]0.5880.5617.0699.696760.0060.9080.8780.663
CDDFuse [54]0.6940.6457.3429.733495.4730.9330.8830.830
SDNet [44]0.5270.5756.8979.318936.3890.8780.8720.749
IFCNN [51]0.6790.6347.2239.662521.7410.9460.8820.856
U2Fusion [41]0.5030.4926.6478.789857.4550.8780.8700.695
DeFusion [20]0.6060.6067.2169.701532.0920.8900.8800.676
TC-MoA0.7260.6377.4289.805423.7730.9490.8860.858
" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.332, + 0.875, + 0.345 + ], + "angle": 0, + "content": "Table 2. Quantitative results of the MEF task in MEFB [49]." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.347, + 0.89, + 0.469 + ], + "angle": 0, + "content": "
MethodPsnr\\(Q_c\\)\\(Q_p\\)\\(Q_g\\)\\(Q_s\\)MEF-SSIMFMI\\(Q_w\\)
Deepfuse [31]57.1040.4260.3520.3250.6410.8970.8730.548
MEF-GAN [43]56.9470.3090.1240.2460.4870.7720.8460.300
MEFNet [28]56.5940.6560.5950.5650.8380.9140.8900.866
IFCNN [51]57.1950.5530.5620.4780.7200.9430.8820.834
FusionDN [42]56.9770.5000.5040.4340.6720.9240.8770.776
PMGI [45]57.1170.4890.5250.4420.6660.9360.8850.804
U2Fusion [41]57.0550.4570.5050.4150.5850.9300.8820.787
DeFusion [20]57.1310.5390.3780.3760.7510.9020.8770.733
TC-MoA57.2130.5780.5980.5280.7670.9640.8880.845
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.642 + ], + "angle": 0, + "content": "ple fusion tasks. For most task-specific image fusion methods, TC-MoA also performs remarkable improvements, although these methods deploy complex task-specific designs. Specifically, our method achieves a significant advantage in VIF, \\( Q_{cv} \\), and EN metrics correlated to human perception or information theory. This indicates that our fused image conforms more to human visual perception and contains more information from the source images. The results on MS-SSIM, SD, and \\( Q_{w} \\) metrics also explain that our fusion results possess a sufficient amount of structural information and gradient information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Qualitative Comparisons. As shown in Fig. 3, our model outperforms the competing methods in terms of visual quality. For instance, the contours of the tree in dark areas are well depicted and the background texture is clearer. It is worth noting that our model has the capability to directly generate color images, or we can also follow other approaches to conduct on gray images and colorize the results to obtain \\(\\mathrm{TC - MoA}^{YUV}\\). Compared with the competing methods, our model exhibits higher contrast between foreground and background, and more saturated colors. This comparison demonstrates our model is effective in generating fused images with superior perceptual quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.825, + 0.849 + ], + "angle": 0, + "content": "4.3. Evaluation on Multi-Exposure Fusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we compare TC-MoA with three task-specific MEF methods i.e. Deepfuse[31], MEF-GAN[43] and MEF-Net [28], and five general image fusion meth" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7103" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.47, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.335, + 0.416, + 0.348 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons in MEF task." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.355, + 0.469, + 0.385 + ], + "angle": 0, + "content": "ods i.e. IFCNN [51], FusionDN [42], PMGI [45], U2Fusion [41], and DeFusion [20]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.47, + 0.596 + ], + "angle": 0, + "content": "Quantitative Comparisons. We employ 8 quantitative metrics to evaluate our model and the competing methods on MEFB, as presented in Table 2. Our model achieves the SoTA performance in the general image fusion methods and achieves competitive results in task-specific methods. For example, our model significantly improves the MEF-SSIM scores due to the compatibility of TC-MoA with diverse tasks, enabling task-specific optimization while reducing task conflicts. The MEF-SSIM focuses on the structure and contrast distortion of images, and \\( Q_{p} \\) measures the phase congruency of the source and fused images. The highest MEF-SSIM, PSNR, and \\( Q_{p} \\) values indicate our effectiveness in the preservation of structural information, detailed features, and image quality in the fused images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.471, + 0.822 + ], + "angle": 0, + "content": "Qualitative Comparisons. As shown in Fig. 4, our method preserves the most texture details and color in both high and low illumination regions. Specifically, in the high illumination region, our model effectively retains more structural information of the clouds around the sun. In the low illumination region, the colors of the fonts on the hot air balloon are completely confused in the PMGI and U2Fusion methods. As a comparison, our method maintains detailed information while keeping more accurate color information. In fact, there is no objective standard for the brightness of the fused image for multi-exposure images. To solve this problem, our method makes the fusion of images controllable by modulating the prompt. The image of \"TC-MoA Light\" shows the result of this modulation. Detailed information about the modulation can be found in Sec. 4.6." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.376, + 0.848 + ], + "angle": 0, + "content": "4.4. Evaluation on Multi-Focus Fusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We compare TC-MoA with two task-specific MFF methods i.e. IFCNN [51] (supervised on MFF) and MFF-GAN [46], and four general image fusion methods i.e. FusionDN [42]," + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.892, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.311, + 0.836, + 0.325 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons in MFF task." + }, + { + "type": "table_caption", + "bbox": [ + 0.56, + 0.329, + 0.832, + 0.342 + ], + "angle": 0, + "content": "Table 3. Quantitative results of the MFF task." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.344, + 0.892, + 0.447 + ], + "angle": 0, + "content": "
MethodNMIQncieMIQcbQcv↓MS-SSIMFMIQw
IFCNN [51]0.8470.8256.4950.69144.3730.9910.8810.912
MFF-GAN [46]0.7280.8205.6890.60972.4600.9770.8760.877
FusionDN [42]0.6750.8185.4490.519178.4910.9420.8640.794
PMGI [45]0.7020.8195.5220.542140.9570.9230.8660.575
U2Fusion [41]0.6700.8185.3290.530142.3250.9480.8670.830
DeFusion [20]0.7680.8215.8380.62583.1950.9600.8700.685
TC-MoA0.8750.8276.6950.71836.5120.9900.8810.891
" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.456, + 0.815, + 0.47 + ], + "angle": 0, + "content": "PMGI [45], U2Fusion [41], and DeFusion [20]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.638 + ], + "angle": 0, + "content": "Quantitative Comparisons. We employ 8 quantitative metrics on the dataset provided in the [50], as illustrated in Table 3. Our method shows competitive performance compared to most image fusion methods. For general image fusion, our model achieves superior performance across all metrics. Additionally, TC-MoA directly generates the fused image without predicting the decision map. Nevertheless, we still significantly outperform IFCNN in NMI, MI, Qcb, and Qcv metrics. This indicates that our fusion results, compared to supervised method, retain more source image details and demonstrate remarkable human visual perception." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.759 + ], + "angle": 0, + "content": "Qualitative Comparisons. As depicted in Fig. 5, our fused image exhibits superior consistency in terms of texture and color, surpassing that of the other methods. U2Fusion exhibits color deviations in the far-focus region, while IFCNN distorts the near-focus image. Furthermore, these methods are unable to effectively remove the distortion around the flowers, while we introduce an MFF-specific loss, which achieves the most visually appealing results for the flower." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.654, + 0.787 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.89, + 0.823 + ], + "angle": 0, + "content": "We conduct ablation studies to verify the effect of our TC-MoA and network architecture." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "TC-MoA. To verify the effectiveness of TC-MoA, we remove it from our framework. As shown in the first and second rows of Table 4, the models trained by multi-task training outperform those trained for a single task, based on 5 commonly used metrics. Interestingly though, this" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7104" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.131 + ], + "angle": 0, + "content": "Table 4. Ablation studies on TC-MoA. S and M represent single task training and multi-task training, while SA and MA denote single adapter and multiple adapters respectively." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.132, + 0.468, + 0.25 + ], + "angle": 0, + "content": "
TaskSMSAMA\\(Q_{\\text{abf}}\\)\\(Q_p\\)FMI\\(Q_c\\)SSIM
VIF✓✓0.59840.40730.88570.63350.4540
✓✓0.59970.41160.88620.63570.4544
✓✓0.60070.41190.88620.63650.4550
MEF✓✓0.63850.59160.88750.57540.4115
✓✓0.63620.58990.88750.57650.4106
✓✓0.64490.59800.88830.57760.4116
MFF✓✓0.65170.67330.87970.77020.6774
✓✓0.65620.67990.88060.77420.6792
✓✓0.65680.68110.88080.77550.6794
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.265, + 0.468, + 0.354 + ], + "angle": 0, + "content": "rule reverses in the case of MEF, indicating that inter-task competition occurs for models trained directly with multiple tasks. Comparatively, the model incorporating the multiple adapters improves the performance across all metrics, suggesting that TC-MoA is dynamically compatible with different tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.468, + 0.581 + ], + "angle": 0, + "content": "Adapter-based Fine-tuning. We performed ablation experiments on the backbone and adapter methods, as shown in Table 5. The fine-tuning method of \"FrozenBackbone\" represents freezing the entire backbone and only unfreezing the final linear layer. Compared to the approach of \"Frozen-Backbone\", our TC-MoA achieves significant fusion performance improvements by introducing a small number of learnable parameters. More importantly, our method outperforms the existing adapter-based visual fine-tuning approach AdaptFormer [4] on both base and large versions of pre-trained models, demonstrating the superiority of the TC-MoA structure in fusion tasks. In addition to this, Table 9 shows our base version achieves competitive performance on the VIF task, which can be further enhanced by the large version." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.296, + 0.607 + ], + "angle": 0, + "content": "4.6. Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.468, + 0.749 + ], + "angle": 0, + "content": "Efficiency. We have observed the speed issues brought about by the massive parameters of the pre-trained models. To address this, we have optimized the ViT architecture by the shifted windows, which will be detailed in Sec. 8. After optimization, for multi-task with arbitrary-size inputs, Table 6 shows the base and large versions of the model have accelerated by \\(178\\%\\) and \\(167\\%\\), resulting in acceptable FPS (Frames Per Second) compared with other methods. The \"Frozen\" refers to the vanilla ViT architecture." + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.759, + 0.468, + 0.786 + ], + "angle": 0, + "content": "\\[\np r o m p t _ {x} ^ {\\prime} = \\mu + \\alpha (p r o m p t _ {x} - \\mu) + \\beta \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.786, + 0.405, + 0.801 + ], + "angle": 0, + "content": "\\[\np r o m p t _ {y} = \\mu + \\alpha (p r o m p t _ {y} - \\mu) - \\beta\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Prompt Controllability. Employing the formula noted in Eq. (8), we manipulate both the scale and shift of the prompt provided to the trained model, where \\(\\mu = 0.5\\) represents the mean, while \\(\\alpha\\) and \\(\\beta\\) denote the scaling and shifting factors, respectively. We conduct the manipulation on some samples and present an example in Fig. 6 (a)." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.089, + 0.892, + 0.117 + ], + "angle": 0, + "content": "Table 5. Ablation studies on fine-tuning. \\( \\mathbb{B} \\) and \\( \\mathbb{L} \\) represent the use of ViT-base and ViT-large structures respectively." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.118, + 0.892, + 0.212 + ], + "angle": 0, + "content": "
ArchitectureVIFMEFMFF
\\( Q_{abf} \\)\\( Q_p \\)SSIM\\( Q_{abf} \\)\\( Q_p \\)SSIM\\( Q_{abf} \\)\\( Q_p \\)SSIM
FrozenBackbone\\( _\\text{B} \\)0.1750.0890.2750.2500.1390.2270.2950.1890.466
AdaptFormet\\( _\\text{B} \\)0.5310.3360.4270.5740.5040.3770.5790.5610.651
TC-Mo\\( A_\\text{B} \\)0.5760.3960.4500.6040.5400.3940.6120.6190.660
FrozenBackbone\\( _\\text{L} \\)0.3300.2380.3630.4520.4080.3330.4500.4280.596
AdaptFormet\\( _\\text{L} \\)0.5760.3920.4460.6160.5630.3960.6400.6500.671
TC-Mo\\( A_\\text{L} \\)0.6010.4120.4550.6450.5980.4120.6570.6810.679
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.218, + 0.882, + 0.231 + ], + "angle": 0, + "content": "Table 6. Efficiency and scalability comparisons of approaches." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.232, + 0.891, + 0.299 + ], + "angle": 0, + "content": "
U2FDDFMCDDFrozenBTC-MoABFrozenLTC-MoAL
Total Params (M)1.32-1.78111.65115.40339.12348.70
Trainable Params (M)1.32-1.78-3.87-9.58
FPS (VIF 640×512)4.720.012.521.173.330.601.60
FPS (MEF arbitrary-size)2.21--1.373.580.711.86
FPS (MFF arbitrary-size)1.86--1.393.980.711.95
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.892, + 0.552 + ], + "angle": 0, + "content": "Obviously, as the value of \\(\\beta\\) increases, the brightness and texture of the fused image increasingly resemble the source image \\(X\\), and inversely, the source image \\(Y\\). Thus, \\(\\beta\\) can be considered as a shifting of the dominant intensity bias. Moreover, a higher value of \\(\\alpha\\) tends to favor one source image over the other in the image patches, which can be considered as a scaling of the dominant intensity deviation. For example, outputs with \\(\\beta = -0.3\\) are more similar to the under-exposure image globally. As \\(\\alpha\\) increases from 0 to 3, the building areas of the fused image gradually become brighter, but the brightness of the clouds remains basically unchanged. This suggests that changing the scale factor of the prompt motivates the model to select different regions. Overall, this experiment demonstrates that the prompt is controllable and can be linearly transformed to obtain fuse images in different degrees." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Generalizability. Our method is highly generalizable to tasks with similar fusion rules by zero-shot fusion. Specifically, we take samples from medical image fusion dataset [1] and pan-sharpening dataset generated by the Quickbird satellite, which are fused based on VIF's paradigm. We obtain a base fused image with \\(\\alpha = 1\\) and \\(\\beta = 0\\), as shown in Fig. 6 (b) and (c). However, without any prior knowledge of these unknown tasks, the output fused image is an under-performing image, missing many high and low-frequency information. Interestingly, by manipulating the prompt, we can find a suitable fusion rule for these new tasks, obtaining reasonable fusion results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Routers Controllability. As shown in Fig. 7, the images generated by the MFF router are darker than VIF in areas without gradient guidance. The MFF selects the reference sources based on gradients, it tends to average the information from sources without gradient guidance. In contrast, the VIF router tends to preserve the maximum intensity information. Moreover, at the bottom of the image, the MFF router tends to preserve the gradient information depending on a single source, while the VIF router preserves the gradient information from both sources. Additionally, by manipulating the combination of different routers, our model" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7105" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.887, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.325, + 0.706, + 0.339 + ], + "angle": 0, + "content": "Figure 6. Visualisation of controllability and generalization of our method." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.346, + 0.465, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.505, + 0.458, + 0.52 + ], + "angle": 0, + "content": "Figure 7. Visualisation of different routers on medical images." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.531, + 0.469, + 0.561 + ], + "angle": 0, + "content": "copes with controlling the dynamic changing fusion degree, obtaining more controllable results." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.562, + 0.47, + 0.787 + ], + "angle": 0, + "content": "Hyperparameters. We conducted hyperparameters analysis on two aspects: the number of experts \\((N)\\) and the interval between two TC-MoAs \\((\\tau)\\). We utilize the ViT-large as our backbone, which consists of 24 transformer blocks in the encoder and 8 in the decoder. If \\(\\tau = 4\\), our model contains 8 instances of TC-MoA. We employ two widely-used metrics \\(Q_{abf}\\) and SSIM to inform the selection of hyperparameters. As presented in Fig. 8, as \\(N\\) increases, the performance initially increases but eventually declines, with the peak performance detected at \\(N = 4\\). This suggests that increasing the number of experts does not necessarily enhance the model's performance. Similarly, when \\(\\tau = 4\\), the model demonstrates the best performance across multitask. Hence, our experiments are conducted under \\(N = 4\\) and \\(\\tau = 4\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.801, + 0.196, + 0.816 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose a task-customized mixture of adapters for general image fusion. With the help of TCMoA, different fusion tasks customize their respective mixture of adapters to obtain prompt guidance for multiple-source image fusion in a unified model. To ensure compat" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.346, + 0.888, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.505, + 0.837, + 0.52 + ], + "angle": 0, + "content": "Figure 8. Experiments on the hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.892, + 0.652 + ], + "angle": 0, + "content": "ibility with different tasks while maintaining complementarity for multi-source images, we further propose mutual information regularization to constrain these adapters. Experimental results have shown that TC-MoA achieves advanced performance in general image fusion against the competing methods. In addition, TC-MoA demonstrates strong prompt controllability and router controllability to perform flexible manipulation on the fused images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.681, + 0.682 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.901 + ], + "angle": 0, + "content": "This work was sponsored in part by the National Key R&D Program of China 2022ZD0116500, in part by the National Natural Science Foundation of China (62222608, 62106171, U23B2049, 61925602), in part by the Haihe Lab of ITAI under Grant 22HHXCJC00002, in part by the Tianjin Natural Science Foundation under Grant 21JCYBJC00580, in part by Tianjin Natural Science Funds for Distinguished Young Scholar under Grant 23JCJQJC00270, and in part by the Key Laboratory of Big Data Intelligent Computing, Chongqing University of Posts and Telecommunications under Grant BDIC-2023-A-008. This work was also sponsored by CAAI-CANN Open Fund, developed on OpenI Community. The authors appreciate the suggestions from CVPR anonymous peer reviewers." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7106" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Aanlib. Ihttp://www.med.harvard.edu/AANLIB/home.html, (Accessed 5 January 2020). 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.468, + 0.197 + ], + "angle": 0, + "content": "[2] Jianrui Cai, Shuhang Gu, and Lei Zhang. Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing, 27(4):2049-2062, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.201, + 0.468, + 0.268 + ], + "angle": 0, + "content": "[3] Bing Cao, Yiming Sun, Pengfei Zhu, and Qinghua Hu. Multi-modal gated mixture of local-to-global experts for dynamic image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 23555-23564, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.271, + 0.468, + 0.338 + ], + "angle": 0, + "content": "[4] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.341, + 0.468, + 0.394 + ], + "angle": 0, + "content": "[5] Nedeljko Cvejic, Artur Loza, David Bull, and Nishan Canagarajah. A similarity metric for assessment of image fusion algorithms. International journal of signal processing, 2(3): 178-182, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.398, + 0.468, + 0.452 + ], + "angle": 0, + "content": "[6] Wang Di, Liu Jinyuan, Fan Xin, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In International Joint Conference on Artificial Intelligence (IJCAI), 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.453, + 0.468, + 0.507 + ], + "angle": 0, + "content": "[7] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. The Journal of Machine Learning Research, 23(1):5232-5270, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.509, + 0.468, + 0.564 + ], + "angle": 0, + "content": "[8] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification with dynamic evidential fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2551-2566, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.566, + 0.468, + 0.632 + ], + "angle": 0, + "content": "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.468, + 0.689 + ], + "angle": 0, + "content": "[10] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.468, + 0.746 + ], + "angle": 0, + "content": "[11] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.468, + 0.802 + ], + "angle": 0, + "content": "[12] Xinyu Jia, Chuang Zhu, Minzhen Li, Wenqi Tang, and Wenli Zhou. Llvip: A visible-infrared paired dataset for low-light vision. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3496-3504, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.468, + 0.858 + ], + "angle": 0, + "content": "[13] Xin Jin, Xiuliang Xi, Ding Zhou, Xiaoxuan Ren, Jie Yang, and Qian Jiang. An unsupervised multi-focus image fusion method based on transformer and u-net. IET Image Processing, 17(3):733-746, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[14] Harpreet Kaur, Deepika Koundal, and Virender Kadyan. Image fusion techniques: a survey. Archives of computational methods in Engineering, 28:4425-4447, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[15] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. CoRR, abs/2006.16668, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[16] Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images. IEEE Transactions on Image Processing, 28(5):2614-2623, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.209, + 0.892, + 0.249 + ], + "angle": 0, + "content": "[17] Hui Li, Xiao-Jun Wu, and Josef Kittler. Rfn-nest: An end-to-end residual fusion network for infrared and visible images. Information Fusion, 73:72-86, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.253, + 0.892, + 0.293 + ], + "angle": 0, + "content": "[18] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.297, + 0.892, + 0.352 + ], + "angle": 0, + "content": "[19] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems, 35:109-123, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.355, + 0.892, + 0.41 + ], + "angle": 0, + "content": "[20] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Fusion from decomposition: A self-supervised decomposition approach for image fusion. In European Conference on Computer Vision, pages 719-735. Springer, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.413, + 0.892, + 0.508 + ], + "angle": 0, + "content": "[21] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5811, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[22] Yu Liu, Xun Chen, Hu Peng, and Zengfu Wang. Multi-focus image fusion with a deep convolutional neural network. Information Fusion, 36:191-207, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.556, + 0.892, + 0.597 + ], + "angle": 0, + "content": "[23] Yu Liu, Lei Wang, Juan Cheng, Chang Li, and Xun Chen. Multi-focus image fusion: A survey of the state of the art. Information Fusion, 64:71-91, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.601, + 0.892, + 0.682 + ], + "angle": 0, + "content": "[24] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.686, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[25] Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible image fusion methods and applications: A survey. Information fusion, 45:153-178, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.73, + 0.892, + 0.797 + ], + "angle": 0, + "content": "[26] Jiayi Ma, Linfeng Tang, Fan Fan, Jun Huang, Xiaoguang Mei, and Yong Ma. Swinfusion: Cross-domain long-range learning for general image fusion via swim transformer. IEEE/CAA Journal of Automatica Sinica, 9(7):1200-1217, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[27] Kede Ma, Kai Zeng, and Zhou Wang. Perceptual quality assessment for multi-exposure image fusion. IEEE Transactions on Image Processing, 24(11):3345-3356, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Kede Ma, Zhengfang Duanmu, Hanwei Zhu, Yuming Fang, and Zhou Wang. Deep guided learning for fast multi-exposure image fusion. IEEE Transactions on Image Processing, 29:2808-2819, 2019. 1, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7107" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[29] Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with limoe: the language-image mixture of experts. Advances in Neural Information Processing Systems, 35:9564-9576, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[30] Linhao Qu, Shaolei Liu, Manning Wang, and Zhijian Song. Transmef: A transformer-based multi-exposure image fusion framework using self-supervised multi-task learning. In Proceedings of the AAAI conference on artificial intelligence, pages 2126-2134, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[31] K Ram Prabhakar, V Sai Srikar, and R Venkatesh Babu. Deepfuse: A deep unsupervised approach for exposure fusion with extreme exposure image pairs. In Proceedings of the IEEE international conference on computer vision, pages 4714-4722, 2017. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.301, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[32] Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual adapters. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.357, + 0.47, + 0.425 + ], + "angle": 0, + "content": "[33] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.47, + 0.482 + ], + "angle": 0, + "content": "[34] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Det-fusion: A detection-driven infrared and visible image fusion network. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4003-4011, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.47, + 0.538 + ], + "angle": 0, + "content": "[35] Linfeng Tang, Jiteng Yuan, and Jiayi Ma. Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network. Information Fusion, 82:28-42, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.47, + 0.58 + ], + "angle": 0, + "content": "[36] Wei Tang, Fazhi He, and Yu Liu. Ydtr: Infrared and visible image fusion via y-shape dynamic transformer. IEEE Transactions on Multimedia, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.47, + 0.636 + ], + "angle": 0, + "content": "[37] Z. Wang, E.P. Simoncelli, and A.C. Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402 Vol.2, 2003. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.47, + 0.691 + ], + "angle": 0, + "content": "[38] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.47, + 0.747 + ], + "angle": 0, + "content": "[39] Zhishe Wang, Yanlin Chen, Wenyu Shao, Hui Li, and Lei Zhang. Swinfuse: A residual swim transformer fusion network for infrared and visible images. IEEE Transactions on Instrumentation and Measurement, 71:1-12, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[40] Fang Xu, Jinghong Liu, Yueming Song, Hui Sun, and Xuan Wang. Multi-exposure image fusion techniques: A comprehensive review. Remote Sensing, 14(3):771, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[41] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[42] Han Xu, Jiayi Ma, Zhuliang Le, Junjun Jiang, and Xiaojie Guo. Fusiondn: A unified densely connected network for image fusion. In Proceedings of the AAAI conference on artificial intelligence, pages 12484-12491, 2020. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[43] Han Xu, Jiayi Ma, and Xiao-Ping Zhang. Mef-gan: Multiexposure image fusion via generative adversarial networks. IEEE Transactions on Image Processing, 29:7203-7216, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[44] Hao Zhang and Jiayi Ma. Sdnet: A versatile squeeze-and-decomposition network for real-time image fusion. International Journal of Computer Vision, 129:2761-2785, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[45] Hao Zhang, Han Xu, Yang Xiao, Xiaojie Guo, and Jiayi Ma. Rethinking the image fusion: A fast unified image fusion network based on proportional maintenance of gradient and intensity. In Proceedings of the AAAI conference on artificial intelligence, pages 12797-12804, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.262, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[46] Hao Zhang, Zhuliang Le, Zhenfeng Shao, Han Xu, and Jiayi Ma. Mff-gan: An unsupervised generative adversarial network with adaptive and gradient joint constraints for multifocus image fusion. Information Fusion, 66:40-53, 2021. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.332, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[47] Juncheng Zhang, Qingmin Liao, Shaojun Liu, Haoyu Ma, Wenming Yang, and Jing-Hao Xue. Real-mff: A large realistic multi-focus image dataset with ground truth. Pattern Recognition Letters, 138:370-377, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.39, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[48] Qingyang Zhang, Haitao Wu, Changqing Zhang, Qinghua Hu, Huazhu Fu, Joey Tianyi Zhou, and Xi Peng. Provable dynamic fusion for low-quality multimodal data. In International Conference on Machine Learning, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.446, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[49] Xingchen Zhang. Benchmarking and comparing multi-exposure image fusion algorithms. Information Fusion, 74: 111-131, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[50] Xingchen Zhang. Deep learning-based multi-focus image fusion: A survey and a comparative study. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[51] Yu Zhang, Yu Liu, Peng Sun, Han Yan, Xiaolin Zhao, and Li Zhang. Ifcnn: A general image fusion framework based on convolutional neural network. Information Fusion, 54: 99-118, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.588, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[52] Zixiang Zhao, Shuang Xu, Chunxia Zhang, Junmin Liu, Pengfei Li, and Jiangshe Zhang. Didfuse: Deep image decomposition for infrared and visible image fusion. arXiv preprint arXiv:2003.09210, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.645, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[53] Zixiang Zhao, Shuang Xu, Jiangshe Zhang, Chengyang Liang, Chunxia Zhang, and Junmin Liu. Efficient and model-based infrared and visible image fusion via algorithm unrolling. IEEE Transactions on Circuits and Systems for Video Technology, 32(3):1186-1196, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.715, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[54] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5906–5916, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.8, + 0.892, + 0.869 + ], + "angle": 0, + "content": "[55] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. Ddfm: denoising diffusion model for multi-modality image fusion. arXiv preprint arXiv:2303.06840, 2023. 1, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.869 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7108" + } + ] +] \ No newline at end of file diff --git a/2024/Task-Customized Mixture of Adapters for General Image Fusion/7c9be988-491f-4562-822c-11cfddbdaa7d_origin.pdf b/2024/Task-Customized Mixture of Adapters for General Image Fusion/7c9be988-491f-4562-822c-11cfddbdaa7d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..495916025aca2eff2f0df59c9824e562d3785bab --- /dev/null +++ b/2024/Task-Customized Mixture of Adapters for General Image Fusion/7c9be988-491f-4562-822c-11cfddbdaa7d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dec147c432e8b55c2b542e568f9399049400831c154164dfea8d0abddf1af75a +size 4412935 diff --git a/2024/Task-Customized Mixture of Adapters for General Image Fusion/full.md b/2024/Task-Customized Mixture of Adapters for General Image Fusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6360392054785ef0742d7082b87efeaf1ff0661b --- /dev/null +++ b/2024/Task-Customized Mixture of Adapters for General Image Fusion/full.md @@ -0,0 +1,342 @@ +# Task-Customized Mixture of Adapters for General Image Fusion + +Pengfei Zhu, Yang Sun, Bing Cao*, Qinghua Hu +Tianjin Key Lab of Machine Learning, College of Intelligence and Computing, Tianjin University, China +{zhupengfei, yangsun, caobing, huqinghua}@tju.edu.cn + +# Abstract + +General image fusion aims at integrating important information from multi-source images. However, due to the significant cross-task gap, the respective fusion mechanism varies considerably in practice, resulting in limited performance across subtasks. To handle this problem, we propose a novel task-customized mixture of adapters (TC-MoA) for general image fusion, adaptively prompting various fusion tasks in a unified model. We borrow the insight from the mixture of experts (MoE), taking the experts as efficient tuning adapters to prompt a pre-trained foundation model. These adapters are shared across different tasks and constrained by mutual information regularization, ensuring compatibility with different tasks while complementarity for multi-source images. The task-specific routing networks customize these adapters to extract task-specific information from different sources with dynamic dominant intensity, performing adaptive visual feature prompt fusion. Notably, our TC-MoA controls the dominant intensity bias for different fusion tasks, successfully unifying multiple fusion tasks in a single model. Extensive experiments show that TC-MoA outperforms the competing approaches in learning commonalities while retaining compatibility for general image fusion (multi-modal, multi-exposure, and multifocus), and also demonstrating striking controllability on more generalization experiments. The code is available at https://github.com/YangSun22/TC-MoA. + +# 1. Introduction + +Image fusion aims to integrate complementary information from multi-source images captured by different sensors in the same scene onto a single image. It is usually used to enhance important information and visual quality [14]. Currently, general image fusion mainly includes multi-modal, multi-exposure, and multi-focus image fusion, etc. Fusion tasks exhibit diverse fusion mechanisms. The Multi-Exposure image Fusion (MEF) focuses on integrat + +![](images/4faa96e52ba1405a5a98f0214853b96206b8e12d16bae11f5e7239a828b33a78.jpg) +Figure 1. Prompt can adaptively select the complementary information from multi-source features. The dominant intensity bias varies according to the task, which is reflected by the different shades of colors. + +ing images with multiple exposure levels into a high-quality full exposure image [40]. Each source image contributes its own illumination and structural information to the fused image [28]. The Visible-Infrared image Fusion (VIF) is a subfield of the Multi-Modal Fusion (MMF) [8, 16, 48] that aims at fusing complementary information of infrared and visible modalities to produce robust and informative fused images [25]. Infrared images provide more intensity information, while visible images offer more texture and gradient information. The Multi-Focus image Fusion (MFF) aims at generating an all-in-focus image from a series of partially focused images [23]. A patch of multi-focus fused image typically corresponds to only one source image. Consequently, it can be observed that the MEF and VIF tasks involve relatively balanced fusion, while MFF is an extremely unbalanced task that tends to exhibit polarized choices. + +With the rapid development of deep learning techniques, image fusion has achieved great progress in recent years [31, 41], while most existing methods [28, 54, 55] focus on single image fusion scenario alone. Task-specific methods [28, 36] often employ task-specific strategies such as designing complex task-biased networks or utilizing task-specific fusion loss functions, resulting in the weak generalization to other tasks [41]. Considering that different + +fusion tasks have similar objectives, i.e., integrating complementary information from multiple source images, some recently proposed approaches [20, 41, 51] attempted to conduct various fusion tasks using a unified model, forming general image fusion. These methods, however, are either enmeshed in dominant-task bias [51] or multi-task commonality [20, 41] at the expense of individuality, leading to unsatisfactory performance. This motivates us to explore a more compatible fusion paradigm that is adaptively and dynamically compatible with different fusion scenarios. + +To handle this challenge, inspired by the impressive representation capability of pre-trained foundation models, we introduce the foundation model as a fixed encoder to extract complementary features for multi-source images. Different from most existing methods, we borrow insight from the mixture of experts (MoE) [24, 33], taking each expert as an efficient tuning adapter to perform adaptive visual feature prompt fusion based on the foundation model. The task-specific routing networks customize these adapters to generate task-specific fusion prompt for different sources, forming a novel task-customized mixture of adapters (TC-MoA) architecture. An mutual information regularization is further developed to constrain the prompt, which guarantees complementarity for diverse sources. It is worth noting that the prompt have significant task bias and dominant intensity gaps. As shown in Fig. 1, the prompt of MFF has a greater color difference than VIF and MEF, meaning that feature selection has more bipolarity on dominant intensity bias. Our model effectively perceives the fusion intensity bias among varied fusion tasks in a single model and is therefore compatible with a wider range of fusion tasks. Extensive experiments verify our superiority over the competing methods in general image fusion, including multi-modal, multi-exposure, and multi-focus fusion. More importantly, our TC-MoA even shows unprecedented controllability and generalizability for unseen fusion tasks, fully demonstrating our potential in broader scenarios. We summarize our main contributions as follows: + +- We propose a unified general image fusion model, providing a new task-customized mixture of adapters (TC-MoA) for adaptive multi-source image fusion (benefiting from the dynamically aggregating effective information from the respective modalities). +- We propose a mutual information regularization for adapters, which allows our model to more accurately identify the dominant intensity of different source images. +- To the best of our knowledge, we for the first time propose an MoE-based flexible adapter for the foundation model in general image fusion. By only adding $2.8\%$ of learnable parameters, our model copes with numerous fusion tasks. Extensive experiments demonstrate our superiority against the competing methods, while showing significant controllability and generalizability. + +# 2. Related Work + +Image Fusion. Image fusion focuses on generating a fused image containing complementary information from different source images. Some early methods [16, 22, 31] tackled their respective task by leveraging CNN. Then, GAN-based [43, 46] and Transformer-based methods [13, 30, 36] have been proposed to improve the fusion quality. Furthermore, high-level tasks [3, 21, 34] are also introduced to guide the fusion of images. Feature decomposition [52, 53] for high-low frequency has also gained significant attention as a research direction. Different from these methods, some most recent methods focus on general image fusion that aims to address multiple fusion tasks in a single model. Zhang et al. [51] performed a supervised training framework for MFF, which was generalized to other tasks by adjusting fusion conditions. Xu et al. [41, 42] appraised the quantity and quality of information from source images or features with the image quality assessment (IQA) to decide the fusion paradigm. PMGI [45] and SwinFusion [26] incorporated a unified fusion framework and loss function, and separate model training for individual tasks. Liang et al. [20] performed a self-supervised fusion framework to learn fusion commonality while ignoring task-specific individuality. In this work, we accommodate diverse fusion tasks by dynamically customizing the mixture of adapters, rather than suffering individuality for cross-task commonality. + +Parameter-Efficient Fine-Tuning. To efficiently adapt pre-trained models to the respective downstream application tasks, some Parameter-Efficient Fine-Tuning (PEFT) studies have been proposed. PEFT can be mainly divided into the Adapter [10, 19, 32] and the Prefix Tuning [11, 18]. Jia et al. [11] introduced prompt tuning to the ViT structure. Chen et al. [4] proposed a new adapter by scaling the original features instead of summing. Lian et al. [19] bridged the gap between the pre-trained features and the downstream task features by linearly varying the original features. In this paper, different from these methods that focus on tuning high-level tasks, we for the first time generalize the powerful foundational model to general image fusion. + +Mixture of Experts. Shazeer et al. [33] first proposed the mixture of experts (MoE) to increase model capacity without increasing computational complexity. Based on this, Lepikhin et al. [15] and Fedus et al. [7] integrated MoE with the transformer structure, further pushing the upper limit of network capacity. In addition, MoE has been validated as effective in other challenges. Ma et al. [24] tackled the multi-task problem by designing multi-gates MoE. Mustafa et al. [29] exploited MoE to train the contrastive learning-based vision-language foundation model. Inspired by this, we take each expert as an adapter, forming a task-customized mixture of adapters for tuning our general image fusion framework. + +![](images/4afc4a966748aec168d990d926fbd73cbbe07952b6361c58cf20fcdcf0c7bd31.jpg) +Figure 2. An overview of our proposed TC-MoA method. Our approach gradually modulates the fusion results by inserting TC-MoA into the frozen ViT backbone. TC-MoA generates task-specific prompt through a task-specific router bank and an shared adapter bank. The fusion layer utilizes prompt as scale and source-specific embeddings as biases to obtain fusion images. + +# 3. Method + +# 3.1. Overview + +In this paper, we propose a task-customized mixture of adapters for general image fusion, which is a novel parameter-efficient fine-tuning method to jointly and adaptively prompt various fusion tasks. Given a pair of source images $\{(X,Y)|X,Y\in \mathbb{R}^{H\times W\times 3}\}$ , the network integrates complementary information from different sources to obtain the fused image $I_{Fusion}\in \mathbb{R}^{H\times W\times 3}$ . As shown in Fig. 2, we feed the source images into the ViT network and obtain tokens of the source images through the patch embedding layer. ViT consists of an encoder for feature extraction and a decoder for image reconstruction, both of which are composed of transformer blocks. A TC-MoA is inserted into each $\tau$ ( $\tau = 4$ ) transformer blocks in both the encoder and decoder. Each TC-MoA consists of a task-specific router bank, a task-shared adapter bank, and a fusion layer. Our network gradually modulates the fusion results through these TC-MoA. + +# 3.2. Task-Customized Mixture of Adapters + +As shown in Fig. 2, each TC-MoA consists of a task-specific router bank $\{G^{\mathcal{V}}, G^{E}, G^{F}\}$ , a task-shared adapter bank $\{A_{1}, \ldots, A_{N}\}$ and a prompting fusion layer $\mathcal{F}$ . The TC-MoA includes two main stages: prompt generation and prompt-driven fusion. For ease of expression, we take the VIF as an example, assuming that the input comes from the VIF dataset and uses $G$ to denote $G^{V}$ . + +Prompt Generation. First, we obtain multi-source features for subsequent processing. The network structure before the $j$ -th TC-MoA is defined as $E_{j}(\cdot)$ , and the extracted features are defined as $f_{x}$ and $f_{y}$ , where $f \in \mathbb{R}^{pH \times pW \times C}$ . + +We concatenate $f_{x}$ and $f_{y}$ as a feature representation of pairs of multi-source tokens. This allows tokens from different sources to exchange information in the subsequent network. However, directly computing on the high-dimensional concatenated features will bring amounts of unnecessary parameters. Therefore, we use $L(\cdot)$ to perform feature dimension reduction and obtain the processed multiple-source features $\Phi$ as follows, + +$$ +f _ {x} = E _ {j} (X), f _ {y} = E _ {j} (Y) \tag {1} +$$ + +$$ +\Phi = L \left(\operatorname {C a t} \left(f _ {x}, f _ {y}\right)\right) \tag {2} +$$ + +where $Cat(\cdot)$ represents the concatenation of features, $L(\cdot)$ consists of a linear layer and normalization layers. Next, depending on the task to which $\Phi$ belongs, we select a task-specific router from the router bank to customize the routing schemes, i.e., which adapters in the adapter bank should be input for each token pair. + +$$ +G (x) = \operatorname {S o f t m a x} (T o p K (x \cdot W _ {g} + +$$ + +$$ +\mathcal {N} (0, 1) \cdot \text {S o f t p l u s} (x \cdot W _ {\text {n o i s e}})) \tag {3} +$$ + +where $TopK(\cdot)$ keeps only the top $K(K = 2)$ values, setting the rest to $-\infty$ (after Softmax( $\cdot$ ), the value becomes 0). $W_{g}$ and $W_{noise}$ are learnable parameters. The customization routing schemes vary for tasks, as evidenced by the proportion of the number of times that different adapters have been routed in Fig. 2 (b). After that, we weight sum the output from the adapters to obtain the prompt. Each router masters a task-biased appetite for customizing a suitable mixture of adapters, and each adapter generates the prompt. The task-customized multi-source prompt is then + +calculated as follows, + +$$ +p r o m p t = G A P (S i g m o i d \left(\sum_ {i = 1} ^ {N} G (\Phi) _ {i} \cdot A _ {i} (\Phi)\right) \tag {4} +$$ + +where $GAP(\cdot)$ represents the global average pooling, $i$ is the index of adapters and $G(\cdot)_i$ is the routing value for the $i$ -th adapter. The task-customized prompt, denoted as prompt $\in \mathbb{R}^{pH\times pW\times 2}$ , is composed of prompt $_x$ and prompt $_y$ in $\mathbb{R}^{pH\times pW\times 1}$ and has a value range of $(0,1)$ . + +Prompt-Driven Fusion. The task-customized prompt is constrained by mutual information regularization (MIR), which guarantees complementarity for diverse sources. Thus the prompt can be taken as an evaluation of the proportion of important information in each source. By dot-multiplication of multi-source features and prompt, we retain complementary information while removing redundant information. Afterward, considering that the feature representation should contain a source-correlated bias (e.g. visible or infrared images), we introduce input-independent learnable parameters for each source, i.e., source embedding $S$ . After feature prompting and shifting, we obtain the refined source features $(h_x, h_y)$ as follows, + +$$ +h _ {x} = \operatorname {p r o m p t} _ {x} \cdot f _ {x} + S _ {x} +$$ + +$$ +h _ {y} = \text {p r o m p t} _ {y} \cdot f _ {y} + S _ {y} \tag {5} +$$ + +$$ +f _ {T C - M o A} = \mathcal {F} \left(h _ {x} + h _ {y}\right) +$$ + +where $S_{x}$ and $S_{y}$ represent the source embeddings of visible and infrared images in VIF respectively. $\mathcal{F}(\cdot)$ fuses the refined multi-source features using an additive operation and then passes them through a set of convolutional layers. These layers introduce local receptive fields to reduce the checkerboard artifacts and align the solution space of the subsequent transformer blocks. Ultimately, we obtain a fusion feature through the task-customized prompt. To encourage the model to extract important information gradually, the features input into the next transformer block are processed as follows, + +$$ +f _ {x} ^ {\prime} = \lambda_ {f} f _ {x} + (1 - \lambda_ {f}) f _ {T C - M o A} \tag {6} +$$ + +$$ +f _ {y} ^ {\prime} = \lambda_ {f} f _ {y} + (1 - \lambda_ {f}) f _ {T C - M o A} +$$ + +where $\lambda_{f}$ is a learnable parameter initialized to 0.5. + +Mutual Information Regularization. In order to ensure that the model dynamically retains complementary information while discarding redundant information of multisource features, we impose regularization constraints on the prompt. Assuming that the information representation of features varies linearly, we define the MIR as follows, + +$$ +\min \left| p r o m p t _ {x} + p r o m p t _ {y} - 1 \right| \tag {7} +$$ + +The MIR allows the model to accurately identify the dominant intensity of different sources, which is positively correlated with the information content of the sources. + +# 3.3. Task-Customized Loss Function + +In addition to accommodating the individuality of different tasks in our network structure, we also customize unsupervised loss functions for each fusion task. We add $\mathcal{L}_{aux}$ to the loss function of each task to ensure the training of TC-MoA and $\mathcal{L}_{aux}$ is the auxiliary loss in [33] to avoid adapters learning unbalanced. On the other hand, in order to generate high-quality fusion images, we impose constraints on the structural information ( $\mathcal{L}_{ssim}$ ), intensity information ( $\mathcal{L}_{Pixel}$ ) for VIF loss function, and gradient information ( $\mathcal{L}_{Grad}$ ) of the fusion images for different fusion tasks. + +For the VIF task, our objective is to retain the most pronounced high-frequency and low-frequency information from the source images in the fused image. Thus, we design the $\mathcal{L}_{MaxPixel}$ and $\mathcal{L}_{MaxGrad}$ . To avoid confusing gradients, we retain the sign of the gradient values in all loss functions related to gradient information. For the MEF task, we consider that the luminance of the fused images should be at an average level with all the gradient information. Thus we design the loss functions for MEF with $\mathcal{L}_{AvgPixel}$ and $L_{MaxGrad}$ . Additionally, we adopt $\mathcal{L}_{mefssim}$ [28] which is specially designed for the MEF task, instead of the SSIM loss function. For the MFF task, we believe that each patch of the fused image should only depend on a single source image with the maximum gradient. This is to prevent the objects' edges in defocused images from being preserved, thereby affecting the quality of the fused image. For this purpose, we select only one of the source images to compute the loss function for each patch in the image, i.e. $\mathcal{L}_{MaskPixel}$ and $\mathcal{L}_{MaskGrad}$ . Please refer to the supplementary material for details of the loss functions. + +# 4. Experiments + +# 4.1. Experimental Setting + +Datasets. We conduct experiments in three image fusion scenarios: visible-infrared, multi-focus, and multi-exposure fusion. For VIF, we evaluate our model on the LLVIP [12] dataset. The training set contains 12025 image pairs, and we randomly select 70 samples out of the test set for evaluation. For MEF, we employ the SCIE [2] dataset (589 pairs) for training and MEFB [49] dataset (100 pairs) for testing, we utilize the most underexposed and overexposed images from sequences in SCIE as inputs. For MFF, we train our model on the RealMFF [47] and the MFI-WHU [46] datasets and follow the test setting in MFIF [50]. + +Implementation Details. Our experiments are performed on a server with $8 \times$ NVIDIA RTX A6000 GPUs. Although our model supports inputs of arbitrary size, we ensure that each fusion task receives an equal amount of data per iteration by randomly cropping all samples to a size of $448 \times 448$ . We employ the pre-trained MAE-large model [9] with GAN loss as our backbone. We train 20 epochs jointly + +![](images/27d2a3d40010f0ab9be2cbdcfa8c5ed021a7fcc149dcfc74ed11beae7a6e065b.jpg) +Visible + +![](images/eb0594ec9d6039a539fccf7a93413b0a5c8e0fa7f561249c6ea914e976a7b715.jpg) +Infrared + +![](images/0f3833092c0eb472b9063c6a6daeeb52cb53160277f34b0f5879ddd9cc1a5b32.jpg) +DenseFuse + +![](images/8abaf1c91d80e98c37755be470db609e0684b583735d697bb7cb646559c1d955.jpg) +TarDAL + +![](images/5cf1986a470e004dbb5bf2de1fd8247f6168f457ca1553d4db13851cb4ae8014.jpg) +CDDFuse + +![](images/ee448dbc20e52058528f110d63dcfd9e15910c68c0c5ab910a98a7cfc276115c.jpg) +DDEM + +![](images/d49a02e470178b1823db51cb56f7b70f6ebbdf9a1cc61a22fd1698a2406a5f3d.jpg) + +![](images/6bf3b35b064e8153d4fd139e3077d40fd225d6d30000cfbc8c0eea7341576391.jpg) + +![](images/506129fe1c4083657ac6aeb11d8cbd4ea14e6c4a031c8724befa6a420ce8df9e.jpg) +U2Fusion +Figure 3. Qualitative comparisons of various methods in VIF task. + +![](images/b60ad65da7359c50b66439811fcd2aef35b8cc08bddd23ad6ca42b4d8ec92858.jpg) +DeFusion + +![](images/57fc4e3f61e7f8354da3d26091edb1927f131c230570e5ec4882ba247d56f957.jpg) +TC-MoA + +![](images/f528e94175affc3405cf93e164287e6bc7f49316a7eab0424ee6860c9e4db105.jpg) +TC-MoA $^{YUV}$ + +with all the fusion tasks, and the batch size for each fusion task is set to 3. We adopt the AdamW optimizer with the initial learning rate of $1.5 \times 10^{-4}$ . To ensure training stability, we apply the Exponential Moving Average (EMA) to routers and adapters optimization. Each TC-MoA consists of $N(N = 4)$ adapters embedded in different layers, but only the top $K(K = 2)$ adapters are activated. + +Evaluation Metrics. Since the commonly used evaluation metrics for fusion tasks are not exactly the same, we customize a set of metrics for each task. Existing metrics can be mainly divided into the following four categories [40]. The information theory-based metrics: entropy (EN), peak signal-to-noise ratio (PSNR), mutual information (MI), normalized MI (NMI), Nonlinear correlation information entropy $(Q_{ncie})$ , feature mutual information (FMI). The image feature-based metrics: phase congruency-based metric $Q_{p}$ , standard deviation (SD), gradient-based metric $Q_{abf}$ and $Q_{g}$ . The structural similarity-based metrics: $Q_{c}$ [5], $Q_{w}$ , $Q_{s}$ [38], structural similarity (SSIM), multiscale structural similarity (MS-SSIM [37]) and MEF-SSIM [27]. The human perception inspired metrics: visual information fidelity (VIF), $Q_{cb}$ and $Q_{cv}$ . Detailed information on metrics can be found in [25] and [23]. + +# 4.2. Evaluation on Multi-Modal Fusion + +In this section, we compare our TC-MoA with ten task-specific image fusion methods and four general image fusion methods on the LLVIP dataset. + +Quantitative Comparisons. We evaluate the fusion performance on 8 quantitative metrics, as shown in Table 1. Our method outperforms all the general image fusion methods, demonstrating our superior compatibility across multi- + +Table 1. Quantitative results of the VIF task in LLVIP dataset. Boldface in red and boldface in blue show the best and second-best values, respectively. The underline represents the best value in the general image fusion methods. + +
Method\(VIF\)\(Q_c\)ENSD\(Q_{cv}\)↓MS-SSIMFMI\(Q_w\)
Densefuse [31]0.5450.5336.8309.381817.2130.8780.8760.622
AUIF [53]0.4020.3706.1377.7691087.5690.7840.8690.530
DIDFuse [52]0.3660.3485.9917.765897.0070.7670.8630.479
TarDAL [21]0.5500.5627.3479.609549.1770.8640.8600.628
YDTR [36]0.4860.5246.6388.810883.3330.8350.8760.614
RFN-Nest [17]0.4970.4567.0529.609857.1570.8620.8710.385
SwinFuse [39]0.3990.3215.8787.4571306.6520.7340.8700.425
UMF-CMGR [6]0.4140.4796.6078.520810.6700.8010.8760.561
DDFM [55]0.5880.5617.0699.696760.0060.9080.8780.663
CDDFuse [54]0.6940.6457.3429.733495.4730.9330.8830.830
SDNet [44]0.5270.5756.8979.318936.3890.8780.8720.749
IFCNN [51]0.6790.6347.2239.662521.7410.9460.8820.856
U2Fusion [41]0.5030.4926.6478.789857.4550.8780.8700.695
DeFusion [20]0.6060.6067.2169.701532.0920.8900.8800.676
TC-MoA0.7260.6377.4289.805423.7730.9490.8860.858
+ +Table 2. Quantitative results of the MEF task in MEFB [49]. + +
MethodPsnr\(Q_c\)\(Q_p\)\(Q_g\)\(Q_s\)MEF-SSIMFMI\(Q_w\)
Deepfuse [31]57.1040.4260.3520.3250.6410.8970.8730.548
MEF-GAN [43]56.9470.3090.1240.2460.4870.7720.8460.300
MEFNet [28]56.5940.6560.5950.5650.8380.9140.8900.866
IFCNN [51]57.1950.5530.5620.4780.7200.9430.8820.834
FusionDN [42]56.9770.5000.5040.4340.6720.9240.8770.776
PMGI [45]57.1170.4890.5250.4420.6660.9360.8850.804
U2Fusion [41]57.0550.4570.5050.4150.5850.9300.8820.787
DeFusion [20]57.1310.5390.3780.3760.7510.9020.8770.733
TC-MoA57.2130.5780.5980.5280.7670.9640.8880.845
+ +ple fusion tasks. For most task-specific image fusion methods, TC-MoA also performs remarkable improvements, although these methods deploy complex task-specific designs. Specifically, our method achieves a significant advantage in VIF, $Q_{cv}$ , and EN metrics correlated to human perception or information theory. This indicates that our fused image conforms more to human visual perception and contains more information from the source images. The results on MS-SSIM, SD, and $Q_{w}$ metrics also explain that our fusion results possess a sufficient amount of structural information and gradient information. + +Qualitative Comparisons. As shown in Fig. 3, our model outperforms the competing methods in terms of visual quality. For instance, the contours of the tree in dark areas are well depicted and the background texture is clearer. It is worth noting that our model has the capability to directly generate color images, or we can also follow other approaches to conduct on gray images and colorize the results to obtain $\mathrm{TC - MoA}^{YUV}$ . Compared with the competing methods, our model exhibits higher contrast between foreground and background, and more saturated colors. This comparison demonstrates our model is effective in generating fused images with superior perceptual quality. + +# 4.3. Evaluation on Multi-Exposure Fusion + +In this section, we compare TC-MoA with three task-specific MEF methods i.e. Deepfuse[31], MEF-GAN[43] and MEF-Net [28], and five general image fusion meth + +![](images/3be55776a0e0e294babd779c2eb66bfff6b8bb8f12b1e6269c4b6c3ed99a5da9.jpg) +Figure 4. Qualitative comparisons in MEF task. + +ods i.e. IFCNN [51], FusionDN [42], PMGI [45], U2Fusion [41], and DeFusion [20]. + +Quantitative Comparisons. We employ 8 quantitative metrics to evaluate our model and the competing methods on MEFB, as presented in Table 2. Our model achieves the SoTA performance in the general image fusion methods and achieves competitive results in task-specific methods. For example, our model significantly improves the MEF-SSIM scores due to the compatibility of TC-MoA with diverse tasks, enabling task-specific optimization while reducing task conflicts. The MEF-SSIM focuses on the structure and contrast distortion of images, and $Q_{p}$ measures the phase congruency of the source and fused images. The highest MEF-SSIM, PSNR, and $Q_{p}$ values indicate our effectiveness in the preservation of structural information, detailed features, and image quality in the fused images. + +Qualitative Comparisons. As shown in Fig. 4, our method preserves the most texture details and color in both high and low illumination regions. Specifically, in the high illumination region, our model effectively retains more structural information of the clouds around the sun. In the low illumination region, the colors of the fonts on the hot air balloon are completely confused in the PMGI and U2Fusion methods. As a comparison, our method maintains detailed information while keeping more accurate color information. In fact, there is no objective standard for the brightness of the fused image for multi-exposure images. To solve this problem, our method makes the fusion of images controllable by modulating the prompt. The image of "TC-MoA Light" shows the result of this modulation. Detailed information about the modulation can be found in Sec. 4.6. + +# 4.4. Evaluation on Multi-Focus Fusion + +We compare TC-MoA with two task-specific MFF methods i.e. IFCNN [51] (supervised on MFF) and MFF-GAN [46], and four general image fusion methods i.e. FusionDN [42], + +![](images/4baabd277aba43943df99b68d36323a697a056c98a527e3edc084397c757357c.jpg) +Figure 5. Qualitative comparisons in MFF task. + +Table 3. Quantitative results of the MFF task. + +
MethodNMIQncieMIQcbQcv↓MS-SSIMFMIQw
IFCNN [51]0.8470.8256.4950.69144.3730.9910.8810.912
MFF-GAN [46]0.7280.8205.6890.60972.4600.9770.8760.877
FusionDN [42]0.6750.8185.4490.519178.4910.9420.8640.794
PMGI [45]0.7020.8195.5220.542140.9570.9230.8660.575
U2Fusion [41]0.6700.8185.3290.530142.3250.9480.8670.830
DeFusion [20]0.7680.8215.8380.62583.1950.9600.8700.685
TC-MoA0.8750.8276.6950.71836.5120.9900.8810.891
+ +PMGI [45], U2Fusion [41], and DeFusion [20]. + +Quantitative Comparisons. We employ 8 quantitative metrics on the dataset provided in the [50], as illustrated in Table 3. Our method shows competitive performance compared to most image fusion methods. For general image fusion, our model achieves superior performance across all metrics. Additionally, TC-MoA directly generates the fused image without predicting the decision map. Nevertheless, we still significantly outperform IFCNN in NMI, MI, Qcb, and Qcv metrics. This indicates that our fusion results, compared to supervised method, retain more source image details and demonstrate remarkable human visual perception. + +Qualitative Comparisons. As depicted in Fig. 5, our fused image exhibits superior consistency in terms of texture and color, surpassing that of the other methods. U2Fusion exhibits color deviations in the far-focus region, while IFCNN distorts the near-focus image. Furthermore, these methods are unable to effectively remove the distortion around the flowers, while we introduce an MFF-specific loss, which achieves the most visually appealing results for the flower. + +# 4.5. Ablation Study + +We conduct ablation studies to verify the effect of our TC-MoA and network architecture. + +TC-MoA. To verify the effectiveness of TC-MoA, we remove it from our framework. As shown in the first and second rows of Table 4, the models trained by multi-task training outperform those trained for a single task, based on 5 commonly used metrics. Interestingly though, this + +Table 4. Ablation studies on TC-MoA. S and M represent single task training and multi-task training, while SA and MA denote single adapter and multiple adapters respectively. + +
TaskSMSAMA\(Q_{\text{abf}}\)\(Q_p\)FMI\(Q_c\)SSIM
VIF✓✓0.59840.40730.88570.63350.4540
✓✓0.59970.41160.88620.63570.4544
✓✓0.60070.41190.88620.63650.4550
MEF✓✓0.63850.59160.88750.57540.4115
✓✓0.63620.58990.88750.57650.4106
✓✓0.64490.59800.88830.57760.4116
MFF✓✓0.65170.67330.87970.77020.6774
✓✓0.65620.67990.88060.77420.6792
✓✓0.65680.68110.88080.77550.6794
+ +rule reverses in the case of MEF, indicating that inter-task competition occurs for models trained directly with multiple tasks. Comparatively, the model incorporating the multiple adapters improves the performance across all metrics, suggesting that TC-MoA is dynamically compatible with different tasks. + +Adapter-based Fine-tuning. We performed ablation experiments on the backbone and adapter methods, as shown in Table 5. The fine-tuning method of "FrozenBackbone" represents freezing the entire backbone and only unfreezing the final linear layer. Compared to the approach of "Frozen-Backbone", our TC-MoA achieves significant fusion performance improvements by introducing a small number of learnable parameters. More importantly, our method outperforms the existing adapter-based visual fine-tuning approach AdaptFormer [4] on both base and large versions of pre-trained models, demonstrating the superiority of the TC-MoA structure in fusion tasks. In addition to this, Table 9 shows our base version achieves competitive performance on the VIF task, which can be further enhanced by the large version. + +# 4.6. Analysis and Discussion + +Efficiency. We have observed the speed issues brought about by the massive parameters of the pre-trained models. To address this, we have optimized the ViT architecture by the shifted windows, which will be detailed in Sec. 8. After optimization, for multi-task with arbitrary-size inputs, Table 6 shows the base and large versions of the model have accelerated by $178\%$ and $167\%$ , resulting in acceptable FPS (Frames Per Second) compared with other methods. The "Frozen" refers to the vanilla ViT architecture. + +$$ +p r o m p t _ {x} ^ {\prime} = \mu + \alpha (p r o m p t _ {x} - \mu) + \beta \tag {8} +$$ + +$$ +p r o m p t _ {y} = \mu + \alpha (p r o m p t _ {y} - \mu) - \beta +$$ + +Prompt Controllability. Employing the formula noted in Eq. (8), we manipulate both the scale and shift of the prompt provided to the trained model, where $\mu = 0.5$ represents the mean, while $\alpha$ and $\beta$ denote the scaling and shifting factors, respectively. We conduct the manipulation on some samples and present an example in Fig. 6 (a). + +Table 5. Ablation studies on fine-tuning. $\mathbb{B}$ and $\mathbb{L}$ represent the use of ViT-base and ViT-large structures respectively. + +
ArchitectureVIFMEFMFF
\( Q_{abf} \)\( Q_p \)SSIM\( Q_{abf} \)\( Q_p \)SSIM\( Q_{abf} \)\( Q_p \)SSIM
FrozenBackbone\( _\text{B} \)0.1750.0890.2750.2500.1390.2270.2950.1890.466
AdaptFormet\( _\text{B} \)0.5310.3360.4270.5740.5040.3770.5790.5610.651
TC-Mo\( A_\text{B} \)0.5760.3960.4500.6040.5400.3940.6120.6190.660
FrozenBackbone\( _\text{L} \)0.3300.2380.3630.4520.4080.3330.4500.4280.596
AdaptFormet\( _\text{L} \)0.5760.3920.4460.6160.5630.3960.6400.6500.671
TC-Mo\( A_\text{L} \)0.6010.4120.4550.6450.5980.4120.6570.6810.679
+ +Table 6. Efficiency and scalability comparisons of approaches. + +
U2FDDFMCDDFrozenBTC-MoABFrozenLTC-MoAL
Total Params (M)1.32-1.78111.65115.40339.12348.70
Trainable Params (M)1.32-1.78-3.87-9.58
FPS (VIF 640×512)4.720.012.521.173.330.601.60
FPS (MEF arbitrary-size)2.21--1.373.580.711.86
FPS (MFF arbitrary-size)1.86--1.393.980.711.95
+ +Obviously, as the value of $\beta$ increases, the brightness and texture of the fused image increasingly resemble the source image $X$ , and inversely, the source image $Y$ . Thus, $\beta$ can be considered as a shifting of the dominant intensity bias. Moreover, a higher value of $\alpha$ tends to favor one source image over the other in the image patches, which can be considered as a scaling of the dominant intensity deviation. For example, outputs with $\beta = -0.3$ are more similar to the under-exposure image globally. As $\alpha$ increases from 0 to 3, the building areas of the fused image gradually become brighter, but the brightness of the clouds remains basically unchanged. This suggests that changing the scale factor of the prompt motivates the model to select different regions. Overall, this experiment demonstrates that the prompt is controllable and can be linearly transformed to obtain fuse images in different degrees. + +Generalizability. Our method is highly generalizable to tasks with similar fusion rules by zero-shot fusion. Specifically, we take samples from medical image fusion dataset [1] and pan-sharpening dataset generated by the Quickbird satellite, which are fused based on VIF's paradigm. We obtain a base fused image with $\alpha = 1$ and $\beta = 0$ , as shown in Fig. 6 (b) and (c). However, without any prior knowledge of these unknown tasks, the output fused image is an under-performing image, missing many high and low-frequency information. Interestingly, by manipulating the prompt, we can find a suitable fusion rule for these new tasks, obtaining reasonable fusion results. + +Routers Controllability. As shown in Fig. 7, the images generated by the MFF router are darker than VIF in areas without gradient guidance. The MFF selects the reference sources based on gradients, it tends to average the information from sources without gradient guidance. In contrast, the VIF router tends to preserve the maximum intensity information. Moreover, at the bottom of the image, the MFF router tends to preserve the gradient information depending on a single source, while the VIF router preserves the gradient information from both sources. Additionally, by manipulating the combination of different routers, our model + +![](images/8e23e1eab439c3742adbc4d23d079208fcc54fcc98b946502f79a914d6d7ca9e.jpg) +Figure 6. Visualisation of controllability and generalization of our method. + +![](images/c946d116226f28e58b216f15b60c42222f9c3108b2a24acfbf30c8889e77b2cb.jpg) +Figure 7. Visualisation of different routers on medical images. + +copes with controlling the dynamic changing fusion degree, obtaining more controllable results. + +Hyperparameters. We conducted hyperparameters analysis on two aspects: the number of experts $(N)$ and the interval between two TC-MoAs $(\tau)$ . We utilize the ViT-large as our backbone, which consists of 24 transformer blocks in the encoder and 8 in the decoder. If $\tau = 4$ , our model contains 8 instances of TC-MoA. We employ two widely-used metrics $Q_{abf}$ and SSIM to inform the selection of hyperparameters. As presented in Fig. 8, as $N$ increases, the performance initially increases but eventually declines, with the peak performance detected at $N = 4$ . This suggests that increasing the number of experts does not necessarily enhance the model's performance. Similarly, when $\tau = 4$ , the model demonstrates the best performance across multitask. Hence, our experiments are conducted under $N = 4$ and $\tau = 4$ . + +# 5. Conclusion + +In this paper, we propose a task-customized mixture of adapters for general image fusion. With the help of TCMoA, different fusion tasks customize their respective mixture of adapters to obtain prompt guidance for multiple-source image fusion in a unified model. To ensure compat + +![](images/9638f7640b7903690f40bacc9fd574573f9c7514bd2dfea28243dd51395cee1a.jpg) +Figure 8. Experiments on the hyperparameters. + +ibility with different tasks while maintaining complementarity for multi-source images, we further propose mutual information regularization to constrain these adapters. Experimental results have shown that TC-MoA achieves advanced performance in general image fusion against the competing methods. In addition, TC-MoA demonstrates strong prompt controllability and router controllability to perform flexible manipulation on the fused images. + +# 6. Acknowledgement + +This work was sponsored in part by the National Key R&D Program of China 2022ZD0116500, in part by the National Natural Science Foundation of China (62222608, 62106171, U23B2049, 61925602), in part by the Haihe Lab of ITAI under Grant 22HHXCJC00002, in part by the Tianjin Natural Science Foundation under Grant 21JCYBJC00580, in part by Tianjin Natural Science Funds for Distinguished Young Scholar under Grant 23JCJQJC00270, and in part by the Key Laboratory of Big Data Intelligent Computing, Chongqing University of Posts and Telecommunications under Grant BDIC-2023-A-008. This work was also sponsored by CAAI-CANN Open Fund, developed on OpenI Community. The authors appreciate the suggestions from CVPR anonymous peer reviewers. + +# References + +[1] Aanlib. Ihttp://www.med.harvard.edu/AANLIB/home.html, (Accessed 5 January 2020). 7 +[2] Jianrui Cai, Shuhang Gu, and Lei Zhang. Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing, 27(4):2049-2062, 2018. 4 +[3] Bing Cao, Yiming Sun, Pengfei Zhu, and Qinghua Hu. Multi-modal gated mixture of local-to-global experts for dynamic image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 23555-23564, 2023. 2, 5 +[4] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 2, 7 +[5] Nedeljko Cvejic, Artur Loza, David Bull, and Nishan Canagarajah. A similarity metric for assessment of image fusion algorithms. International journal of signal processing, 2(3): 178-182, 2005. 5 +[6] Wang Di, Liu Jinyuan, Fan Xin, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In International Joint Conference on Artificial Intelligence (IJCAI), 2022. 5 +[7] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. The Journal of Machine Learning Research, 23(1):5232-5270, 2022. 2 +[8] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification with dynamic evidential fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2551-2566, 2023. 1 +[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 4 +[10] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2 +[11] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2 +[12] Xinyu Jia, Chuang Zhu, Minzhen Li, Wenqi Tang, and Wenli Zhou. Llvip: A visible-infrared paired dataset for low-light vision. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3496-3504, 2021. 4 +[13] Xin Jin, Xiuliang Xi, Ding Zhou, Xiaoxuan Ren, Jie Yang, and Qian Jiang. An unsupervised multi-focus image fusion method based on transformer and u-net. IET Image Processing, 17(3):733-746, 2023. 2 +[14] Harpreet Kaur, Deepika Koundal, and Virender Kadyan. Image fusion techniques: a survey. Archives of computational methods in Engineering, 28:4425-4447, 2021. 1 + +[15] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. CoRR, abs/2006.16668, 2020. 2 +[16] Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images. IEEE Transactions on Image Processing, 28(5):2614-2623, 2018. 1, 2 +[17] Hui Li, Xiao-Jun Wu, and Josef Kittler. Rfn-nest: An end-to-end residual fusion network for infrared and visible images. Information Fusion, 73:72-86, 2021. 5 +[18] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2 +[19] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems, 35:109-123, 2022. 2 +[20] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Fusion from decomposition: A self-supervised decomposition approach for image fusion. In European Conference on Computer Vision, pages 719-735. Springer, 2022. 2, 5, 6 +[21] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5811, 2022. 2, 5 +[22] Yu Liu, Xun Chen, Hu Peng, and Zengfu Wang. Multi-focus image fusion with a deep convolutional neural network. Information Fusion, 36:191-207, 2017. 2 +[23] Yu Liu, Lei Wang, Juan Cheng, Chang Li, and Xun Chen. Multi-focus image fusion: A survey of the state of the art. Information Fusion, 64:71-91, 2020. 1, 5 +[24] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2 +[25] Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible image fusion methods and applications: A survey. Information fusion, 45:153-178, 2019. 1, 5 +[26] Jiayi Ma, Linfeng Tang, Fan Fan, Jun Huang, Xiaoguang Mei, and Yong Ma. Swinfusion: Cross-domain long-range learning for general image fusion via swim transformer. IEEE/CAA Journal of Automatica Sinica, 9(7):1200-1217, 2022. 2 +[27] Kede Ma, Kai Zeng, and Zhou Wang. Perceptual quality assessment for multi-exposure image fusion. IEEE Transactions on Image Processing, 24(11):3345-3356, 2015. 5 +[28] Kede Ma, Zhengfang Duanmu, Hanwei Zhu, Yuming Fang, and Zhou Wang. Deep guided learning for fast multi-exposure image fusion. IEEE Transactions on Image Processing, 29:2808-2819, 2019. 1, 4, 5 + +[29] Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with limoe: the language-image mixture of experts. Advances in Neural Information Processing Systems, 35:9564-9576, 2022. 2 +[30] Linhao Qu, Shaolei Liu, Manning Wang, and Zhijian Song. Transmef: A transformer-based multi-exposure image fusion framework using self-supervised multi-task learning. In Proceedings of the AAAI conference on artificial intelligence, pages 2126-2134, 2022. 2 +[31] K Ram Prabhakar, V Sai Srikar, and R Venkatesh Babu. Deepfuse: A deep unsupervised approach for exposure fusion with extreme exposure image pairs. In Proceedings of the IEEE international conference on computer vision, pages 4714-4722, 2017. 1, 2, 5 +[32] Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual adapters. Advances in neural information processing systems, 30, 2017. 2 +[33] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 2, 4 +[34] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Det-fusion: A detection-driven infrared and visible image fusion network. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4003-4011, 2022. 2 +[35] Linfeng Tang, Jiteng Yuan, and Jiayi Ma. Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network. Information Fusion, 82:28-42, 2022. 1 +[36] Wei Tang, Fazhi He, and Yu Liu. Ydtr: Infrared and visible image fusion via y-shape dynamic transformer. IEEE Transactions on Multimedia, 2022. 1, 2, 5 +[37] Z. Wang, E.P. Simoncelli, and A.C. Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402 Vol.2, 2003. 5 +[38] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5 +[39] Zhishe Wang, Yanlin Chen, Wenyu Shao, Hui Li, and Lei Zhang. Swinfuse: A residual swim transformer fusion network for infrared and visible images. IEEE Transactions on Instrumentation and Measurement, 71:1-12, 2022. 5 +[40] Fang Xu, Jinghong Liu, Yueming Song, Hui Sun, and Xuan Wang. Multi-exposure image fusion techniques: A comprehensive review. Remote Sensing, 14(3):771, 2022. 1, 5 +[41] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2020. 1, 2, 5, 6 +[42] Han Xu, Jiayi Ma, Zhuliang Le, Junjun Jiang, and Xiaojie Guo. Fusiondn: A unified densely connected network for image fusion. In Proceedings of the AAAI conference on artificial intelligence, pages 12484-12491, 2020. 2, 5, 6 + +[43] Han Xu, Jiayi Ma, and Xiao-Ping Zhang. Mef-gan: Multiexposure image fusion via generative adversarial networks. IEEE Transactions on Image Processing, 29:7203-7216, 2020. 2, 5 +[44] Hao Zhang and Jiayi Ma. Sdnet: A versatile squeeze-and-decomposition network for real-time image fusion. International Journal of Computer Vision, 129:2761-2785, 2021. 5 +[45] Hao Zhang, Han Xu, Yang Xiao, Xiaojie Guo, and Jiayi Ma. Rethinking the image fusion: A fast unified image fusion network based on proportional maintenance of gradient and intensity. In Proceedings of the AAAI conference on artificial intelligence, pages 12797-12804, 2020. 2, 5, 6 +[46] Hao Zhang, Zhuliang Le, Zhenfeng Shao, Han Xu, and Jiayi Ma. Mff-gan: An unsupervised generative adversarial network with adaptive and gradient joint constraints for multifocus image fusion. Information Fusion, 66:40-53, 2021. 2, 4, 6 +[47] Juncheng Zhang, Qingmin Liao, Shaojun Liu, Haoyu Ma, Wenming Yang, and Jing-Hao Xue. Real-mff: A large realistic multi-focus image dataset with ground truth. Pattern Recognition Letters, 138:370-377, 2020. 4 +[48] Qingyang Zhang, Haitao Wu, Changqing Zhang, Qinghua Hu, Huazhu Fu, Joey Tianyi Zhou, and Xi Peng. Provable dynamic fusion for low-quality multimodal data. In International Conference on Machine Learning, 2023. 1 +[49] Xingchen Zhang. Benchmarking and comparing multi-exposure image fusion algorithms. Information Fusion, 74: 111-131, 2021. 4, 5 +[50] Xingchen Zhang. Deep learning-based multi-focus image fusion: A survey and a comparative study. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 4, 6 +[51] Yu Zhang, Yu Liu, Peng Sun, Han Yan, Xiaolin Zhao, and Li Zhang. Ifcnn: A general image fusion framework based on convolutional neural network. Information Fusion, 54: 99-118, 2020. 2, 5, 6 +[52] Zixiang Zhao, Shuang Xu, Chunxia Zhang, Junmin Liu, Pengfei Li, and Jiangshe Zhang. Didfuse: Deep image decomposition for infrared and visible image fusion. arXiv preprint arXiv:2003.09210, 2020. 2, 5 +[53] Zixiang Zhao, Shuang Xu, Jiangshe Zhang, Chengyang Liang, Chunxia Zhang, and Junmin Liu. Efficient and model-based infrared and visible image fusion via algorithm unrolling. IEEE Transactions on Circuits and Systems for Video Technology, 32(3):1186-1196, 2021. 2, 5 +[54] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5906–5916, 2023. 1, 5 +[55] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. Ddfm: denoising diffusion model for multi-modality image fusion. arXiv preprint arXiv:2303.06840, 2023. 1, 5 \ No newline at end of file diff --git a/2024/Task-Customized Mixture of Adapters for General Image Fusion/images.zip b/2024/Task-Customized Mixture of Adapters for General Image Fusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c618c84192a5b70290b817b74214bb01fabcb8d1 --- /dev/null +++ b/2024/Task-Customized Mixture of Adapters for General Image Fusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf5a08b04e7c247b2749e72fe13cfd26e75aee08dfe1ab73c40861232805005e +size 950060 diff --git a/2024/Task-Customized Mixture of Adapters for General Image Fusion/layout.json b/2024/Task-Customized Mixture of Adapters for General Image Fusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..231fffaadc19418e8bb608f51c1ff3e5a122bc8a --- /dev/null +++ b/2024/Task-Customized Mixture of Adapters for General Image Fusion/layout.json @@ -0,0 +1,9293 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 103, + 498, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 103, + 498, + 121 + ], + "spans": [ + { + "bbox": [ + 97, + 103, + 498, + 121 + ], + "type": "text", + "content": "Task-Customized Mixture of Adapters for General Image Fusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 143, + 558, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 143, + 558, + 185 + ], + "spans": [ + { + "bbox": [ + 53, + 143, + 558, + 185 + ], + "type": "text", + "content": "Pengfei Zhu, Yang Sun, Bing Cao*, Qinghua Hu \nTianjin Key Lab of Machine Learning, College of Intelligence and Computing, Tianjin University, China \n{zhupengfei, yangsun, caobing, huqinghua}@tju.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 239, + 290, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 290, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 290, + 550 + ], + "type": "text", + "content": "General image fusion aims at integrating important information from multi-source images. However, due to the significant cross-task gap, the respective fusion mechanism varies considerably in practice, resulting in limited performance across subtasks. To handle this problem, we propose a novel task-customized mixture of adapters (TC-MoA) for general image fusion, adaptively prompting various fusion tasks in a unified model. We borrow the insight from the mixture of experts (MoE), taking the experts as efficient tuning adapters to prompt a pre-trained foundation model. These adapters are shared across different tasks and constrained by mutual information regularization, ensuring compatibility with different tasks while complementarity for multi-source images. The task-specific routing networks customize these adapters to extract task-specific information from different sources with dynamic dominant intensity, performing adaptive visual feature prompt fusion. Notably, our TC-MoA controls the dominant intensity bias for different fusion tasks, successfully unifying multiple fusion tasks in a single model. Extensive experiments show that TC-MoA outperforms the competing approaches in learning commonalities while retaining compatibility for general image fusion (multi-modal, multi-exposure, and multifocus), and also demonstrating striking controllability on more generalization experiments. The code is available at https://github.com/YangSun22/TC-MoA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 575, + 128, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 128, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 128, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 596, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 287, + 693 + ], + "type": "text", + "content": "Image fusion aims to integrate complementary information from multi-source images captured by different sensors in the same scene onto a single image. It is usually used to enhance important information and visual quality [14]. Currently, general image fusion mainly includes multi-modal, multi-exposure, and multi-focus image fusion, etc. Fusion tasks exhibit diverse fusion mechanisms. The Multi-Exposure image Fusion (MEF) focuses on integrat" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 212, + 545, + 372 + ], + "blocks": [ + { + "bbox": [ + 307, + 212, + 545, + 372 + ], + "lines": [ + { + "bbox": [ + 307, + 212, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 545, + 372 + ], + "type": "image", + "image_path": "4faa96e52ba1405a5a98f0214853b96206b8e12d16bae11f5e7239a828b33a78.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 373, + 545, + 416 + ], + "lines": [ + { + "bbox": [ + 306, + 373, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 545, + 416 + ], + "type": "text", + "content": "Figure 1. Prompt can adaptively select the complementary information from multi-source features. The dominant intensity bias varies according to the task, which is reflected by the different shades of colors." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 424, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 547, + 616 + ], + "type": "text", + "content": "ing images with multiple exposure levels into a high-quality full exposure image [40]. Each source image contributes its own illumination and structural information to the fused image [28]. The Visible-Infrared image Fusion (VIF) is a subfield of the Multi-Modal Fusion (MMF) [8, 16, 48] that aims at fusing complementary information of infrared and visible modalities to produce robust and informative fused images [25]. Infrared images provide more intensity information, while visible images offer more texture and gradient information. The Multi-Focus image Fusion (MFF) aims at generating an all-in-focus image from a series of partially focused images [23]. A patch of multi-focus fused image typically corresponds to only one source image. Consequently, it can be observed that the MEF and VIF tasks involve relatively balanced fusion, while MFF is an extremely unbalanced task that tends to exhibit polarized choices." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "With the rapid development of deep learning techniques, image fusion has achieved great progress in recent years [31, 41], while most existing methods [28, 54, 55] focus on single image fusion scenario alone. Task-specific methods [28, 36] often employ task-specific strategies such as designing complex task-biased networks or utilizing task-specific fusion loss functions, resulting in the weak generalization to other tasks [41]. Considering that different" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7099" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "fusion tasks have similar objectives, i.e., integrating complementary information from multiple source images, some recently proposed approaches [20, 41, 51] attempted to conduct various fusion tasks using a unified model, forming general image fusion. These methods, however, are either enmeshed in dominant-task bias [51] or multi-task commonality [20, 41] at the expense of individuality, leading to unsatisfactory performance. This motivates us to explore a more compatible fusion paradigm that is adaptively and dynamically compatible with different fusion scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 529 + ], + "type": "text", + "content": "To handle this challenge, inspired by the impressive representation capability of pre-trained foundation models, we introduce the foundation model as a fixed encoder to extract complementary features for multi-source images. Different from most existing methods, we borrow insight from the mixture of experts (MoE) [24, 33], taking each expert as an efficient tuning adapter to perform adaptive visual feature prompt fusion based on the foundation model. The task-specific routing networks customize these adapters to generate task-specific fusion prompt for different sources, forming a novel task-customized mixture of adapters (TC-MoA) architecture. An mutual information regularization is further developed to constrain the prompt, which guarantees complementarity for diverse sources. It is worth noting that the prompt have significant task bias and dominant intensity gaps. As shown in Fig. 1, the prompt of MFF has a greater color difference than VIF and MEF, meaning that feature selection has more bipolarity on dominant intensity bias. Our model effectively perceives the fusion intensity bias among varied fusion tasks in a single model and is therefore compatible with a wider range of fusion tasks. Extensive experiments verify our superiority over the competing methods in general image fusion, including multi-modal, multi-exposure, and multi-focus fusion. More importantly, our TC-MoA even shows unprecedented controllability and generalizability for unseen fusion tasks, fully demonstrating our potential in broader scenarios. We summarize our main contributions as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 534, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 534, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 287, + 593 + ], + "type": "text", + "content": "- We propose a unified general image fusion model, providing a new task-customized mixture of adapters (TC-MoA) for adaptive multi-source image fusion (benefiting from the dynamically aggregating effective information from the respective modalities)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 594, + 287, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 287, + 630 + ], + "type": "text", + "content": "- We propose a mutual information regularization for adapters, which allows our model to more accurately identify the dominant intensity of different source images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 631, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 631, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 631, + 287, + 714 + ], + "type": "text", + "content": "- To the best of our knowledge, we for the first time propose an MoE-based flexible adapter for the foundation model in general image fusion. By only adding " + }, + { + "bbox": [ + 47, + 631, + 287, + 714 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 47, + 631, + 287, + 714 + ], + "type": "text", + "content": " of learnable parameters, our model copes with numerous fusion tasks. Extensive experiments demonstrate our superiority against the competing methods, while showing significant controllability and generalizability." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 95, + 547, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 547, + 396 + ], + "type": "text", + "content": "Image Fusion. Image fusion focuses on generating a fused image containing complementary information from different source images. Some early methods [16, 22, 31] tackled their respective task by leveraging CNN. Then, GAN-based [43, 46] and Transformer-based methods [13, 30, 36] have been proposed to improve the fusion quality. Furthermore, high-level tasks [3, 21, 34] are also introduced to guide the fusion of images. Feature decomposition [52, 53] for high-low frequency has also gained significant attention as a research direction. Different from these methods, some most recent methods focus on general image fusion that aims to address multiple fusion tasks in a single model. Zhang et al. [51] performed a supervised training framework for MFF, which was generalized to other tasks by adjusting fusion conditions. Xu et al. [41, 42] appraised the quantity and quality of information from source images or features with the image quality assessment (IQA) to decide the fusion paradigm. PMGI [45] and SwinFusion [26] incorporated a unified fusion framework and loss function, and separate model training for individual tasks. Liang et al. [20] performed a self-supervised fusion framework to learn fusion commonality while ignoring task-specific individuality. In this work, we accommodate diverse fusion tasks by dynamically customizing the mixture of adapters, rather than suffering individuality for cross-task commonality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 398, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 547, + 555 + ], + "type": "text", + "content": "Parameter-Efficient Fine-Tuning. To efficiently adapt pre-trained models to the respective downstream application tasks, some Parameter-Efficient Fine-Tuning (PEFT) studies have been proposed. PEFT can be mainly divided into the Adapter [10, 19, 32] and the Prefix Tuning [11, 18]. Jia et al. [11] introduced prompt tuning to the ViT structure. Chen et al. [4] proposed a new adapter by scaling the original features instead of summing. Lian et al. [19] bridged the gap between the pre-trained features and the downstream task features by linearly varying the original features. In this paper, different from these methods that focus on tuning high-level tasks, we for the first time generalize the powerful foundational model to general image fusion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "Mixture of Experts. Shazeer et al. [33] first proposed the mixture of experts (MoE) to increase model capacity without increasing computational complexity. Based on this, Lepikhin et al. [15] and Fedus et al. [7] integrated MoE with the transformer structure, further pushing the upper limit of network capacity. In addition, MoE has been validated as effective in other challenges. Ma et al. [24] tackled the multi-task problem by designing multi-gates MoE. Mustafa et al. [29] exploited MoE to train the contrastive learning-based vision-language foundation model. Inspired by this, we take each expert as an adapter, forming a task-customized mixture of adapters for tuning our general image fusion framework." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7100" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 547, + 269 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 547, + 269 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 547, + 269 + ], + "type": "image", + "image_path": "4afc4a966748aec168d990d926fbd73cbbe07952b6361c58cf20fcdcf0c7bd31.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 272, + 547, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 272, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 547, + 307 + ], + "type": "text", + "content": "Figure 2. An overview of our proposed TC-MoA method. Our approach gradually modulates the fusion results by inserting TC-MoA into the frozen ViT backbone. TC-MoA generates task-specific prompt through a task-specific router bank and an shared adapter bank. The fusion layer utilizes prompt as scale and source-specific embeddings as biases to obtain fusion images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 316, + 104, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 104, + 328 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 104, + 328 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 335, + 115, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 115, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 115, + 346 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 290, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 290, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 290, + 558 + ], + "type": "text", + "content": "In this paper, we propose a task-customized mixture of adapters for general image fusion, which is a novel parameter-efficient fine-tuning method to jointly and adaptively prompt various fusion tasks. Given a pair of source images " + }, + { + "bbox": [ + 46, + 354, + 290, + 558 + ], + "type": "inline_equation", + "content": "\\{(X,Y)|X,Y\\in \\mathbb{R}^{H\\times W\\times 3}\\}" + }, + { + "bbox": [ + 46, + 354, + 290, + 558 + ], + "type": "text", + "content": ", the network integrates complementary information from different sources to obtain the fused image " + }, + { + "bbox": [ + 46, + 354, + 290, + 558 + ], + "type": "inline_equation", + "content": "I_{Fusion}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 46, + 354, + 290, + 558 + ], + "type": "text", + "content": ". As shown in Fig. 2, we feed the source images into the ViT network and obtain tokens of the source images through the patch embedding layer. ViT consists of an encoder for feature extraction and a decoder for image reconstruction, both of which are composed of transformer blocks. A TC-MoA is inserted into each " + }, + { + "bbox": [ + 46, + 354, + 290, + 558 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 354, + 290, + 558 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 354, + 290, + 558 + ], + "type": "inline_equation", + "content": "\\tau = 4" + }, + { + "bbox": [ + 46, + 354, + 290, + 558 + ], + "type": "text", + "content": ") transformer blocks in both the encoder and decoder. Each TC-MoA consists of a task-specific router bank, a task-shared adapter bank, and a fusion layer. Our network gradually modulates the fusion results through these TC-MoA." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 563, + 251, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 251, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 251, + 577 + ], + "type": "text", + "content": "3.2. Task-Customized Mixture of Adapters" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "text", + "content": "As shown in Fig. 2, each TC-MoA consists of a task-specific router bank " + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\{G^{\\mathcal{V}}, G^{E}, G^{F}\\}" + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "text", + "content": ", a task-shared adapter bank " + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\{A_{1}, \\ldots, A_{N}\\}" + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "text", + "content": " and a prompting fusion layer " + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "text", + "content": ". The TC-MoA includes two main stages: prompt generation and prompt-driven fusion. For ease of expression, we take the VIF as an example, assuming that the input comes from the VIF dataset and uses " + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "text", + "content": " to denote " + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "inline_equation", + "content": "G^{V}" + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Prompt Generation. First, we obtain multi-source features for subsequent processing. The network structure before the " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "-th TC-MoA is defined as " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "E_{j}(\\cdot)" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": ", and the extracted features are defined as " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{x}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{y}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "f \\in \\mathbb{R}^{pH \\times pW \\times C}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 316, + 547, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 413 + ], + "type": "text", + "content": "We concatenate " + }, + { + "bbox": [ + 304, + 316, + 547, + 413 + ], + "type": "inline_equation", + "content": "f_{x}" + }, + { + "bbox": [ + 304, + 316, + 547, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 316, + 547, + 413 + ], + "type": "inline_equation", + "content": "f_{y}" + }, + { + "bbox": [ + 304, + 316, + 547, + 413 + ], + "type": "text", + "content": " as a feature representation of pairs of multi-source tokens. This allows tokens from different sources to exchange information in the subsequent network. However, directly computing on the high-dimensional concatenated features will bring amounts of unnecessary parameters. Therefore, we use " + }, + { + "bbox": [ + 304, + 316, + 547, + 413 + ], + "type": "inline_equation", + "content": "L(\\cdot)" + }, + { + "bbox": [ + 304, + 316, + 547, + 413 + ], + "type": "text", + "content": " to perform feature dimension reduction and obtain the processed multiple-source features " + }, + { + "bbox": [ + 304, + 316, + 547, + 413 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 316, + 547, + 413 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 370, + 423, + 545, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 423, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 370, + 423, + 545, + 437 + ], + "type": "interline_equation", + "content": "f _ {x} = E _ {j} (X), f _ {y} = E _ {j} (Y) \\tag {1}", + "image_path": "fc16c3f361a2b619a6e61c1eabebcfef478c3adb1cb4125cb82870d356779be5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 382, + 449, + 545, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 449, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 382, + 449, + 545, + 463 + ], + "type": "interline_equation", + "content": "\\Phi = L \\left(\\operatorname {C a t} \\left(f _ {x}, f _ {y}\\right)\\right) \\tag {2}", + "image_path": "b61a8e02649f37227729004a5ae1ff399143816aad14e7227ca2b04cf429f8f1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 469, + 547, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 547, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 547, + 542 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 469, + 547, + 542 + ], + "type": "inline_equation", + "content": "Cat(\\cdot)" + }, + { + "bbox": [ + 304, + 469, + 547, + 542 + ], + "type": "text", + "content": " represents the concatenation of features, " + }, + { + "bbox": [ + 304, + 469, + 547, + 542 + ], + "type": "inline_equation", + "content": "L(\\cdot)" + }, + { + "bbox": [ + 304, + 469, + 547, + 542 + ], + "type": "text", + "content": " consists of a linear layer and normalization layers. Next, depending on the task to which " + }, + { + "bbox": [ + 304, + 469, + 547, + 542 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 469, + 547, + 542 + ], + "type": "text", + "content": " belongs, we select a task-specific router from the router bank to customize the routing schemes, i.e., which adapters in the adapter bank should be input for each token pair." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 555, + 463, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 555, + 463, + 568 + ], + "spans": [ + { + "bbox": [ + 315, + 555, + 463, + 568 + ], + "type": "interline_equation", + "content": "G (x) = \\operatorname {S o f t m a x} (T o p K (x \\cdot W _ {g} +", + "image_path": "6ea65e0a2fa0bfb88b0138fe2a32e61d88b75ac0f35a9d0cc96c66978fae32b5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 384, + 570, + 545, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 570, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 384, + 570, + 545, + 583 + ], + "type": "interline_equation", + "content": "\\mathcal {N} (0, 1) \\cdot \\text {S o f t p l u s} (x \\cdot W _ {\\text {n o i s e}})) \\tag {3}", + "image_path": "7934b5816d46a4e3aa0c0898f5e7789794a09f8cde4354ec0f42c458f3f0ce50.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "TopK(\\cdot)" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " keeps only the top " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "K(K = 2)" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " values, setting the rest to " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "-\\infty" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " (after Softmax(" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "), the value becomes 0). " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "W_{g}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "W_{noise}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " are learnable parameters. The customization routing schemes vary for tasks, as evidenced by the proportion of the number of times that different adapters have been routed in Fig. 2 (b). After that, we weight sum the output from the adapters to obtain the prompt. Each router masters a task-biased appetite for customizing a suitable mixture of adapters, and each adapter generates the prompt. The task-customized multi-source prompt is then" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7101" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 137, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 137, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 137, + 83 + ], + "type": "text", + "content": "calculated as follows," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 89, + 287, + 122 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 89, + 287, + 122 + ], + "spans": [ + { + "bbox": [ + 63, + 89, + 287, + 122 + ], + "type": "interline_equation", + "content": "p r o m p t = G A P (S i g m o i d \\left(\\sum_ {i = 1} ^ {N} G (\\Phi) _ {i} \\cdot A _ {i} (\\Phi)\\right) \\tag {4}", + "image_path": "fbc58a4a93557696a4c67f91e4ebe8a392cfd1811927e3a8fef007ebf7da80eb.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "spans": [ + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "inline_equation", + "content": "GAP(\\cdot)" + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "text", + "content": " represents the global average pooling, " + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "text", + "content": " is the index of adapters and " + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "inline_equation", + "content": "G(\\cdot)_i" + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "text", + "content": " is the routing value for the " + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "text", + "content": "-th adapter. The task-customized prompt, denoted as prompt " + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "inline_equation", + "content": "\\in \\mathbb{R}^{pH\\times pW\\times 2}" + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "text", + "content": ", is composed of prompt" + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "inline_equation", + "content": "_x" + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "text", + "content": " and prompt" + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "inline_equation", + "content": "_y" + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{pH\\times pW\\times 1}" + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "text", + "content": " and has a value range of " + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 46, + 127, + 287, + 187 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 188, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 188, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 287, + 344 + ], + "type": "text", + "content": "Prompt-Driven Fusion. The task-customized prompt is constrained by mutual information regularization (MIR), which guarantees complementarity for diverse sources. Thus the prompt can be taken as an evaluation of the proportion of important information in each source. By dot-multiplication of multi-source features and prompt, we retain complementary information while removing redundant information. Afterward, considering that the feature representation should contain a source-correlated bias (e.g. visible or infrared images), we introduce input-independent learnable parameters for each source, i.e., source embedding " + }, + { + "bbox": [ + 46, + 188, + 287, + 344 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 188, + 287, + 344 + ], + "type": "text", + "content": ". After feature prompting and shifting, we obtain the refined source features " + }, + { + "bbox": [ + 46, + 188, + 287, + 344 + ], + "type": "inline_equation", + "content": "(h_x, h_y)" + }, + { + "bbox": [ + 46, + 188, + 287, + 344 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 350, + 221, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 350, + 221, + 362 + ], + "spans": [ + { + "bbox": [ + 115, + 350, + 221, + 362 + ], + "type": "interline_equation", + "content": "h _ {x} = \\operatorname {p r o m p t} _ {x} \\cdot f _ {x} + S _ {x}", + "image_path": "30c073cd909f493b34d2d5608ba2db27d19cc6f97a54ff3319fc4c4b998a4a10.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 365, + 287, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 365, + 287, + 377 + ], + "spans": [ + { + "bbox": [ + 117, + 365, + 287, + 377 + ], + "type": "interline_equation", + "content": "h _ {y} = \\text {p r o m p t} _ {y} \\cdot f _ {y} + S _ {y} \\tag {5}", + "image_path": "1fffe59f5667bb46818bc2e2aba0964a8b1f47807ad610cc41823be1c6264eb4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 380, + 221, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 380, + 221, + 392 + ], + "spans": [ + { + "bbox": [ + 113, + 380, + 221, + 392 + ], + "type": "interline_equation", + "content": "f _ {T C - M o A} = \\mathcal {F} \\left(h _ {x} + h _ {y}\\right)", + "image_path": "5c4da43dfecdfd732c6bd2710d4c646cb5a78c6fb89b122d3ba8c084cb58d1e8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 397, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 287, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 397, + 287, + 529 + ], + "type": "inline_equation", + "content": "S_{x}" + }, + { + "bbox": [ + 46, + 397, + 287, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 397, + 287, + 529 + ], + "type": "inline_equation", + "content": "S_{y}" + }, + { + "bbox": [ + 46, + 397, + 287, + 529 + ], + "type": "text", + "content": " represent the source embeddings of visible and infrared images in VIF respectively. " + }, + { + "bbox": [ + 46, + 397, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 46, + 397, + 287, + 529 + ], + "type": "text", + "content": " fuses the refined multi-source features using an additive operation and then passes them through a set of convolutional layers. These layers introduce local receptive fields to reduce the checkerboard artifacts and align the solution space of the subsequent transformer blocks. Ultimately, we obtain a fusion feature through the task-customized prompt. To encourage the model to extract important information gradually, the features input into the next transformer block are processed as follows," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 96, + 534, + 287, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 534, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 96, + 534, + 287, + 553 + ], + "type": "interline_equation", + "content": "f _ {x} ^ {\\prime} = \\lambda_ {f} f _ {x} + (1 - \\lambda_ {f}) f _ {T C - M o A} \\tag {6}", + "image_path": "8f2f15c811344415cc367686ad419020ecd2bf7714855ef3af7b243b1afe742b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 99, + 551, + 236, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 551, + 236, + 563 + ], + "spans": [ + { + "bbox": [ + 99, + 551, + 236, + 563 + ], + "type": "interline_equation", + "content": "f _ {y} ^ {\\prime} = \\lambda_ {f} f _ {y} + (1 - \\lambda_ {f}) f _ {T C - M o A}", + "image_path": "320a89012fcfc75218989403ddf1a10fd552b0ace6fe0bf585571709a42dcb8b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 568, + 255, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 255, + 580 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 255, + 580 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 568, + 255, + 580 + ], + "type": "inline_equation", + "content": "\\lambda_{f}" + }, + { + "bbox": [ + 46, + 568, + 255, + 580 + ], + "type": "text", + "content": " is a learnable parameter initialized to 0.5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 581, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 652 + ], + "type": "text", + "content": "Mutual Information Regularization. In order to ensure that the model dynamically retains complementary information while discarding redundant information of multisource features, we impose regularization constraints on the prompt. Assuming that the information representation of features varies linearly, we define the MIR as follows," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 102, + 659, + 287, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 659, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 102, + 659, + 287, + 673 + ], + "type": "interline_equation", + "content": "\\min \\left| p r o m p t _ {x} + p r o m p t _ {y} - 1 \\right| \\tag {7}", + "image_path": "f0440560033b58c4be01e639011b4b3edc7343416b6f1dc044d25ac99c0d9d1d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "content": "The MIR allows the model to accurately identify the dominant intensity of different sources, which is positively correlated with the information content of the sources." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 479, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 479, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 479, + 83 + ], + "type": "text", + "content": "3.3. Task-Customized Loss Function" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "text", + "content": "In addition to accommodating the individuality of different tasks in our network structure, we also customize unsupervised loss functions for each fusion task. We add " + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{aux}" + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "text", + "content": " to the loss function of each task to ensure the training of TC-MoA and " + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{aux}" + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "text", + "content": " is the auxiliary loss in [33] to avoid adapters learning unbalanced. On the other hand, in order to generate high-quality fusion images, we impose constraints on the structural information (" + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ssim}" + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "text", + "content": "), intensity information (" + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Pixel}" + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "text", + "content": ") for VIF loss function, and gradient information (" + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Grad}" + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "text", + "content": ") of the fusion images for different fusion tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 210, + 546, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 546, + 449 + ], + "type": "text", + "content": "For the VIF task, our objective is to retain the most pronounced high-frequency and low-frequency information from the source images in the fused image. Thus, we design the " + }, + { + "bbox": [ + 304, + 210, + 546, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MaxPixel}" + }, + { + "bbox": [ + 304, + 210, + 546, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 210, + 546, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MaxGrad}" + }, + { + "bbox": [ + 304, + 210, + 546, + 449 + ], + "type": "text", + "content": ". To avoid confusing gradients, we retain the sign of the gradient values in all loss functions related to gradient information. For the MEF task, we consider that the luminance of the fused images should be at an average level with all the gradient information. Thus we design the loss functions for MEF with " + }, + { + "bbox": [ + 304, + 210, + 546, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AvgPixel}" + }, + { + "bbox": [ + 304, + 210, + 546, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 210, + 546, + 449 + ], + "type": "inline_equation", + "content": "L_{MaxGrad}" + }, + { + "bbox": [ + 304, + 210, + 546, + 449 + ], + "type": "text", + "content": ". Additionally, we adopt " + }, + { + "bbox": [ + 304, + 210, + 546, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mefssim}" + }, + { + "bbox": [ + 304, + 210, + 546, + 449 + ], + "type": "text", + "content": " [28] which is specially designed for the MEF task, instead of the SSIM loss function. For the MFF task, we believe that each patch of the fused image should only depend on a single source image with the maximum gradient. This is to prevent the objects' edges in defocused images from being preserved, thereby affecting the quality of the fused image. For this purpose, we select only one of the source images to compute the loss function for each patch in the image, i.e. " + }, + { + "bbox": [ + 304, + 210, + 546, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MaskPixel}" + }, + { + "bbox": [ + 304, + 210, + 546, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 210, + 546, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MaskGrad}" + }, + { + "bbox": [ + 304, + 210, + 546, + 449 + ], + "type": "text", + "content": ". Please refer to the supplementary material for details of the loss functions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 460, + 388, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 388, + 473 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 388, + 473 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 479, + 429, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 429, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 429, + 493 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "type": "text", + "content": "Datasets. We conduct experiments in three image fusion scenarios: visible-infrared, multi-focus, and multi-exposure fusion. For VIF, we evaluate our model on the LLVIP [12] dataset. The training set contains 12025 image pairs, and we randomly select 70 samples out of the test set for evaluation. For MEF, we employ the SCIE [2] dataset (589 pairs) for training and MEFB [49] dataset (100 pairs) for testing, we utilize the most underexposed and overexposed images from sequences in SCIE as inputs. For MFF, we train our model on the RealMFF [47] and the MFI-WHU [46] datasets and follow the test setting in MFIF [50]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "type": "text", + "content": "Implementation Details. Our experiments are performed on a server with " + }, + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "type": "text", + "content": " NVIDIA RTX A6000 GPUs. Although our model supports inputs of arbitrary size, we ensure that each fusion task receives an equal amount of data per iteration by randomly cropping all samples to a size of " + }, + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "type": "text", + "content": ". We employ the pre-trained MAE-large model [9] with GAN loss as our backbone. We train 20 epochs jointly" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7102" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 108, + 139 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 108, + 139 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 108, + 139 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 108, + 139 + ], + "type": "image", + "image_path": "27d2a3d40010f0ab9be2cbdcfa8c5ed021a7fcc149dcfc74ed11beae7a6e065b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 139, + 86, + 144 + ], + "lines": [ + { + "bbox": [ + 69, + 139, + 86, + 144 + ], + "spans": [ + { + "bbox": [ + 69, + 139, + 86, + 144 + ], + "type": "text", + "content": "Visible" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 108, + 70, + 167, + 139 + ], + "blocks": [ + { + "bbox": [ + 108, + 70, + 167, + 139 + ], + "lines": [ + { + "bbox": [ + 108, + 70, + 167, + 139 + ], + "spans": [ + { + "bbox": [ + 108, + 70, + 167, + 139 + ], + "type": "image", + "image_path": "eb0594ec9d6039a539fccf7a93413b0a5c8e0fa7f561249c6ea914e976a7b715.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 139, + 148, + 144 + ], + "lines": [ + { + "bbox": [ + 128, + 139, + 148, + 144 + ], + "spans": [ + { + "bbox": [ + 128, + 139, + 148, + 144 + ], + "type": "text", + "content": "Infrared" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 167, + 70, + 227, + 139 + ], + "blocks": [ + { + "bbox": [ + 167, + 70, + 227, + 139 + ], + "lines": [ + { + "bbox": [ + 167, + 70, + 227, + 139 + ], + "spans": [ + { + "bbox": [ + 167, + 70, + 227, + 139 + ], + "type": "image", + "image_path": "0f3833092c0eb472b9063c6a6daeeb52cb53160277f34b0f5879ddd9cc1a5b32.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 139, + 209, + 144 + ], + "lines": [ + { + "bbox": [ + 184, + 139, + 209, + 144 + ], + "spans": [ + { + "bbox": [ + 184, + 139, + 209, + 144 + ], + "type": "text", + "content": "DenseFuse" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 227, + 71, + 286, + 139 + ], + "blocks": [ + { + "bbox": [ + 227, + 71, + 286, + 139 + ], + "lines": [ + { + "bbox": [ + 227, + 71, + 286, + 139 + ], + "spans": [ + { + "bbox": [ + 227, + 71, + 286, + 139 + ], + "type": "image", + "image_path": "8abaf1c91d80e98c37755be470db609e0684b583735d697bb7cb646559c1d955.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 139, + 265, + 144 + ], + "lines": [ + { + "bbox": [ + 247, + 139, + 265, + 144 + ], + "spans": [ + { + "bbox": [ + 247, + 139, + 265, + 144 + ], + "type": "text", + "content": "TarDAL" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 49, + 144, + 107, + 213 + ], + "blocks": [ + { + "bbox": [ + 49, + 144, + 107, + 213 + ], + "lines": [ + { + "bbox": [ + 49, + 144, + 107, + 213 + ], + "spans": [ + { + "bbox": [ + 49, + 144, + 107, + 213 + ], + "type": "image", + "image_path": "5cf1986a470e004dbb5bf2de1fd8247f6168f457ca1553d4db13851cb4ae8014.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 213, + 91, + 218 + ], + "lines": [ + { + "bbox": [ + 66, + 213, + 91, + 218 + ], + "spans": [ + { + "bbox": [ + 66, + 213, + 91, + 218 + ], + "type": "text", + "content": "CDDFuse" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 108, + 144, + 167, + 213 + ], + "blocks": [ + { + "bbox": [ + 108, + 144, + 167, + 213 + ], + "lines": [ + { + "bbox": [ + 108, + 144, + 167, + 213 + ], + "spans": [ + { + "bbox": [ + 108, + 144, + 167, + 213 + ], + "type": "image", + "image_path": "ee448dbc20e52058528f110d63dcfd9e15910c68c0c5ab910a98a7cfc276115c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 213, + 146, + 218 + ], + "lines": [ + { + "bbox": [ + 128, + 213, + 146, + 218 + ], + "spans": [ + { + "bbox": [ + 128, + 213, + 146, + 218 + ], + "type": "text", + "content": "DDEM" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 167, + 144, + 227, + 213 + ], + "blocks": [ + { + "bbox": [ + 167, + 144, + 227, + 213 + ], + "lines": [ + { + "bbox": [ + 167, + 144, + 227, + 213 + ], + "spans": [ + { + "bbox": [ + 167, + 144, + 227, + 213 + ], + "type": "image", + "image_path": "d49a02e470178b1823db51cb56f7b70f6ebbdf9a1cc61a22fd1698a2406a5f3d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 227, + 144, + 286, + 213 + ], + "blocks": [ + { + "bbox": [ + 227, + 144, + 286, + 213 + ], + "lines": [ + { + "bbox": [ + 227, + 144, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 227, + 144, + 286, + 213 + ], + "type": "image", + "image_path": "6bf3b35b064e8153d4fd139e3077d40fd225d6d30000cfbc8c0eea7341576391.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 49, + 219, + 107, + 287 + ], + "blocks": [ + { + "bbox": [ + 49, + 219, + 107, + 287 + ], + "lines": [ + { + "bbox": [ + 49, + 219, + 107, + 287 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 107, + 287 + ], + "type": "image", + "image_path": "506129fe1c4083657ac6aeb11d8cbd4ea14e6c4a031c8724befa6a420ce8df9e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 288, + 91, + 293 + ], + "lines": [ + { + "bbox": [ + 67, + 288, + 91, + 293 + ], + "spans": [ + { + "bbox": [ + 67, + 288, + 91, + 293 + ], + "type": "text", + "content": "U2Fusion" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 301, + 286, + 312 + ], + "lines": [ + { + "bbox": [ + 47, + 301, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 286, + 312 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons of various methods in VIF task." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 108, + 219, + 167, + 287 + ], + "blocks": [ + { + "bbox": [ + 108, + 219, + 167, + 287 + ], + "lines": [ + { + "bbox": [ + 108, + 219, + 167, + 287 + ], + "spans": [ + { + "bbox": [ + 108, + 219, + 167, + 287 + ], + "type": "image", + "image_path": "b60ad65da7359c50b66439811fcd2aef35b8cc08bddd23ad6ca42b4d8ec92858.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 288, + 149, + 293 + ], + "lines": [ + { + "bbox": [ + 126, + 288, + 149, + 293 + ], + "spans": [ + { + "bbox": [ + 126, + 288, + 149, + 293 + ], + "type": "text", + "content": "DeFusion" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 167, + 219, + 227, + 287 + ], + "blocks": [ + { + "bbox": [ + 167, + 219, + 227, + 287 + ], + "lines": [ + { + "bbox": [ + 167, + 219, + 227, + 287 + ], + "spans": [ + { + "bbox": [ + 167, + 219, + 227, + 287 + ], + "type": "image", + "image_path": "57fc4e3f61e7f8354da3d26091edb1927f131c230570e5ec4882ba247d56f957.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 288, + 208, + 293 + ], + "lines": [ + { + "bbox": [ + 186, + 288, + 208, + 293 + ], + "spans": [ + { + "bbox": [ + 186, + 288, + 208, + 293 + ], + "type": "text", + "content": "TC-MoA" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 227, + 219, + 286, + 287 + ], + "blocks": [ + { + "bbox": [ + 227, + 219, + 286, + 287 + ], + "lines": [ + { + "bbox": [ + 227, + 219, + 286, + 287 + ], + "spans": [ + { + "bbox": [ + 227, + 219, + 286, + 287 + ], + "type": "image", + "image_path": "f528e94175affc3405cf93e164287e6bc7f49316a7eab0424ee6860c9e4db105.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 288, + 271, + 293 + ], + "lines": [ + { + "bbox": [ + 239, + 288, + 271, + 293 + ], + "spans": [ + { + "bbox": [ + 239, + 288, + 271, + 293 + ], + "type": "text", + "content": "TC-MoA" + }, + { + "bbox": [ + 239, + 288, + 271, + 293 + ], + "type": "inline_equation", + "content": "^{YUV}" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 323, + 286, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 286, + 407 + ], + "type": "text", + "content": "with all the fusion tasks, and the batch size for each fusion task is set to 3. We adopt the AdamW optimizer with the initial learning rate of " + }, + { + "bbox": [ + 46, + 323, + 286, + 407 + ], + "type": "inline_equation", + "content": "1.5 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 323, + 286, + 407 + ], + "type": "text", + "content": ". To ensure training stability, we apply the Exponential Moving Average (EMA) to routers and adapters optimization. Each TC-MoA consists of " + }, + { + "bbox": [ + 46, + 323, + 286, + 407 + ], + "type": "inline_equation", + "content": "N(N = 4)" + }, + { + "bbox": [ + 46, + 323, + 286, + 407 + ], + "type": "text", + "content": " adapters embedded in different layers, but only the top " + }, + { + "bbox": [ + 46, + 323, + 286, + 407 + ], + "type": "inline_equation", + "content": "K(K = 2)" + }, + { + "bbox": [ + 46, + 323, + 286, + 407 + ], + "type": "text", + "content": " adapters are activated." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "text", + "content": "Evaluation Metrics. Since the commonly used evaluation metrics for fusion tasks are not exactly the same, we customize a set of metrics for each task. Existing metrics can be mainly divided into the following four categories [40]. The information theory-based metrics: entropy (EN), peak signal-to-noise ratio (PSNR), mutual information (MI), normalized MI (NMI), Nonlinear correlation information entropy " + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "inline_equation", + "content": "(Q_{ncie})" + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "text", + "content": ", feature mutual information (FMI). The image feature-based metrics: phase congruency-based metric " + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "inline_equation", + "content": "Q_{p}" + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "text", + "content": ", standard deviation (SD), gradient-based metric " + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "inline_equation", + "content": "Q_{abf}" + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "inline_equation", + "content": "Q_{g}" + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "text", + "content": ". The structural similarity-based metrics: " + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "inline_equation", + "content": "Q_{c}" + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "text", + "content": " [5], " + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "inline_equation", + "content": "Q_{w}" + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "inline_equation", + "content": "Q_{s}" + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "text", + "content": " [38], structural similarity (SSIM), multiscale structural similarity (MS-SSIM [37]) and MEF-SSIM [27]. The human perception inspired metrics: visual information fidelity (VIF), " + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "inline_equation", + "content": "Q_{cb}" + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "inline_equation", + "content": "Q_{cv}" + }, + { + "bbox": [ + 46, + 407, + 286, + 598 + ], + "type": "text", + "content": ". Detailed information on metrics can be found in [25] and [23]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 47, + 609, + 232, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 232, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 232, + 620 + ], + "type": "text", + "content": "4.2. Evaluation on Multi-Modal Fusion" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 628, + 286, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 286, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 286, + 663 + ], + "type": "text", + "content": "In this section, we compare our TC-MoA with ten task-specific image fusion methods and four general image fusion methods on the LLVIP dataset." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Quantitative Comparisons. We evaluate the fusion performance on 8 quantitative metrics, as shown in Table 1. Our method outperforms all the general image fusion methods, demonstrating our superior compatibility across multi-" + } + ] + } + ], + "index": 27 + }, + { + "type": "table", + "bbox": [ + 307, + 115, + 545, + 259 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "type": "text", + "content": "Table 1. Quantitative results of the VIF task in LLVIP dataset. Boldface in red and boldface in blue show the best and second-best values, respectively. The underline represents the best value in the general image fusion methods." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 115, + 545, + 259 + ], + "lines": [ + { + "bbox": [ + 307, + 115, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 115, + 545, + 259 + ], + "type": "table", + "html": "
Method\\(VIF\\)\\(Q_c\\)ENSD\\(Q_{cv}\\)↓MS-SSIMFMI\\(Q_w\\)
Densefuse [31]0.5450.5336.8309.381817.2130.8780.8760.622
AUIF [53]0.4020.3706.1377.7691087.5690.7840.8690.530
DIDFuse [52]0.3660.3485.9917.765897.0070.7670.8630.479
TarDAL [21]0.5500.5627.3479.609549.1770.8640.8600.628
YDTR [36]0.4860.5246.6388.810883.3330.8350.8760.614
RFN-Nest [17]0.4970.4567.0529.609857.1570.8620.8710.385
SwinFuse [39]0.3990.3215.8787.4571306.6520.7340.8700.425
UMF-CMGR [6]0.4140.4796.6078.520810.6700.8010.8760.561
DDFM [55]0.5880.5617.0699.696760.0060.9080.8780.663
CDDFuse [54]0.6940.6457.3429.733495.4730.9330.8830.830
SDNet [44]0.5270.5756.8979.318936.3890.8780.8720.749
IFCNN [51]0.6790.6347.2239.662521.7410.9460.8820.856
U2Fusion [41]0.5030.4926.6478.789857.4550.8780.8700.695
DeFusion [20]0.6060.6067.2169.701532.0920.8900.8800.676
TC-MoA0.7260.6377.4289.805423.7730.9490.8860.858
", + "image_path": "b9bffd31a2927fc9eac665236bcc437052b8d7f2f1099967b6840f949a7372d2.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_body" + } + ], + "index": 29 + }, + { + "type": "table", + "bbox": [ + 306, + 274, + 544, + 371 + ], + "blocks": [ + { + "bbox": [ + 315, + 262, + 535, + 273 + ], + "lines": [ + { + "bbox": [ + 315, + 262, + 535, + 273 + ], + "spans": [ + { + "bbox": [ + 315, + 262, + 535, + 273 + ], + "type": "text", + "content": "Table 2. Quantitative results of the MEF task in MEFB [49]." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 274, + 544, + 371 + ], + "lines": [ + { + "bbox": [ + 306, + 274, + 544, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 544, + 371 + ], + "type": "table", + "html": "
MethodPsnr\\(Q_c\\)\\(Q_p\\)\\(Q_g\\)\\(Q_s\\)MEF-SSIMFMI\\(Q_w\\)
Deepfuse [31]57.1040.4260.3520.3250.6410.8970.8730.548
MEF-GAN [43]56.9470.3090.1240.2460.4870.7720.8460.300
MEFNet [28]56.5940.6560.5950.5650.8380.9140.8900.866
IFCNN [51]57.1950.5530.5620.4780.7200.9430.8820.834
FusionDN [42]56.9770.5000.5040.4340.6720.9240.8770.776
PMGI [45]57.1170.4890.5250.4420.6660.9360.8850.804
U2Fusion [41]57.0550.4570.5050.4150.5850.9300.8820.787
DeFusion [20]57.1310.5390.3780.3760.7510.9020.8770.733
TC-MoA57.2130.5780.5980.5280.7670.9640.8880.845
", + "image_path": "c9e8bdca026c628c10df2448b000b22284878ee08ca1d76581973a61bed20679.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 377, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 508 + ], + "type": "text", + "content": "ple fusion tasks. For most task-specific image fusion methods, TC-MoA also performs remarkable improvements, although these methods deploy complex task-specific designs. Specifically, our method achieves a significant advantage in VIF, " + }, + { + "bbox": [ + 304, + 377, + 545, + 508 + ], + "type": "inline_equation", + "content": "Q_{cv}" + }, + { + "bbox": [ + 304, + 377, + 545, + 508 + ], + "type": "text", + "content": ", and EN metrics correlated to human perception or information theory. This indicates that our fused image conforms more to human visual perception and contains more information from the source images. The results on MS-SSIM, SD, and " + }, + { + "bbox": [ + 304, + 377, + 545, + 508 + ], + "type": "inline_equation", + "content": "Q_{w}" + }, + { + "bbox": [ + 304, + 377, + 545, + 508 + ], + "type": "text", + "content": " metrics also explain that our fusion results possess a sufficient amount of structural information and gradient information." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 509, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 652 + ], + "type": "text", + "content": "Qualitative Comparisons. As shown in Fig. 3, our model outperforms the competing methods in terms of visual quality. For instance, the contours of the tree in dark areas are well depicted and the background texture is clearer. It is worth noting that our model has the capability to directly generate color images, or we can also follow other approaches to conduct on gray images and colorize the results to obtain " + }, + { + "bbox": [ + 304, + 509, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{TC - MoA}^{YUV}" + }, + { + "bbox": [ + 304, + 509, + 545, + 652 + ], + "type": "text", + "content": ". Compared with the competing methods, our model exhibits higher contrast between foreground and background, and more saturated colors. This comparison demonstrates our model is effective in generating fused images with superior perceptual quality." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 306, + 659, + 504, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 504, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 504, + 672 + ], + "type": "text", + "content": "4.3. Evaluation on Multi-Exposure Fusion" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "In this section, we compare TC-MoA with three task-specific MEF methods i.e. Deepfuse[31], MEF-GAN[43] and MEF-Net [28], and five general image fusion meth" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7103" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 287, + 263 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 287, + 263 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 287, + 263 + ], + "type": "image", + "image_path": "3be55776a0e0e294babd779c2eb66bfff6b8bb8f12b1e6269c4b6c3ed99a5da9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 265, + 254, + 275 + ], + "lines": [ + { + "bbox": [ + 80, + 265, + 254, + 275 + ], + "spans": [ + { + "bbox": [ + 80, + 265, + 254, + 275 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons in MEF task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 281, + 287, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 287, + 304 + ], + "type": "text", + "content": "ods i.e. IFCNN [51], FusionDN [42], PMGI [45], U2Fusion [41], and DeFusion [20]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 305, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 287, + 472 + ], + "type": "text", + "content": "Quantitative Comparisons. We employ 8 quantitative metrics to evaluate our model and the competing methods on MEFB, as presented in Table 2. Our model achieves the SoTA performance in the general image fusion methods and achieves competitive results in task-specific methods. For example, our model significantly improves the MEF-SSIM scores due to the compatibility of TC-MoA with diverse tasks, enabling task-specific optimization while reducing task conflicts. The MEF-SSIM focuses on the structure and contrast distortion of images, and " + }, + { + "bbox": [ + 46, + 305, + 287, + 472 + ], + "type": "inline_equation", + "content": "Q_{p}" + }, + { + "bbox": [ + 46, + 305, + 287, + 472 + ], + "type": "text", + "content": " measures the phase congruency of the source and fused images. The highest MEF-SSIM, PSNR, and " + }, + { + "bbox": [ + 46, + 305, + 287, + 472 + ], + "type": "inline_equation", + "content": "Q_{p}" + }, + { + "bbox": [ + 46, + 305, + 287, + 472 + ], + "type": "text", + "content": " values indicate our effectiveness in the preservation of structural information, detailed features, and image quality in the fused images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 472, + 288, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 288, + 651 + ], + "type": "text", + "content": "Qualitative Comparisons. As shown in Fig. 4, our method preserves the most texture details and color in both high and low illumination regions. Specifically, in the high illumination region, our model effectively retains more structural information of the clouds around the sun. In the low illumination region, the colors of the fonts on the hot air balloon are completely confused in the PMGI and U2Fusion methods. As a comparison, our method maintains detailed information while keeping more accurate color information. In fact, there is no objective standard for the brightness of the fused image for multi-exposure images. To solve this problem, our method makes the fusion of images controllable by modulating the prompt. The image of \"TC-MoA Light\" shows the result of this modulation. Detailed information about the modulation can be found in Sec. 4.6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 659, + 230, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 230, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 230, + 671 + ], + "type": "text", + "content": "4.4. Evaluation on Multi-Focus Fusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "We compare TC-MoA with two task-specific MFF methods i.e. IFCNN [51] (supervised on MFF) and MFF-GAN [46], and four general image fusion methods i.e. FusionDN [42]," + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 545, + 241 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 241 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 241 + ], + "type": "image", + "image_path": "4baabd277aba43943df99b68d36323a697a056c98a527e3edc084397c757357c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 246, + 511, + 257 + ], + "lines": [ + { + "bbox": [ + 339, + 246, + 511, + 257 + ], + "spans": [ + { + "bbox": [ + 339, + 246, + 511, + 257 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons in MFF task." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 306, + 272, + 545, + 354 + ], + "blocks": [ + { + "bbox": [ + 342, + 260, + 509, + 270 + ], + "lines": [ + { + "bbox": [ + 342, + 260, + 509, + 270 + ], + "spans": [ + { + "bbox": [ + 342, + 260, + 509, + 270 + ], + "type": "text", + "content": "Table 3. Quantitative results of the MFF task." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 272, + 545, + 354 + ], + "lines": [ + { + "bbox": [ + 306, + 272, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 306, + 272, + 545, + 354 + ], + "type": "table", + "html": "
MethodNMIQncieMIQcbQcv↓MS-SSIMFMIQw
IFCNN [51]0.8470.8256.4950.69144.3730.9910.8810.912
MFF-GAN [46]0.7280.8205.6890.60972.4600.9770.8760.877
FusionDN [42]0.6750.8185.4490.519178.4910.9420.8640.794
PMGI [45]0.7020.8195.5220.542140.9570.9230.8660.575
U2Fusion [41]0.6700.8185.3290.530142.3250.9480.8670.830
DeFusion [20]0.7680.8215.8380.62583.1950.9600.8700.685
TC-MoA0.8750.8276.6950.71836.5120.9900.8810.891
", + "image_path": "b3ffc08056e5adbdcd2eafeeed22b5871d3060c181dbc0673815b4a6fa0d559c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 361, + 498, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 498, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 498, + 372 + ], + "type": "text", + "content": "PMGI [45], U2Fusion [41], and DeFusion [20]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 373, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 505 + ], + "type": "text", + "content": "Quantitative Comparisons. We employ 8 quantitative metrics on the dataset provided in the [50], as illustrated in Table 3. Our method shows competitive performance compared to most image fusion methods. For general image fusion, our model achieves superior performance across all metrics. Additionally, TC-MoA directly generates the fused image without predicting the decision map. Nevertheless, we still significantly outperform IFCNN in NMI, MI, Qcb, and Qcv metrics. This indicates that our fusion results, compared to supervised method, retain more source image details and demonstrate remarkable human visual perception." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 506, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 601 + ], + "type": "text", + "content": "Qualitative Comparisons. As depicted in Fig. 5, our fused image exhibits superior consistency in terms of texture and color, surpassing that of the other methods. U2Fusion exhibits color deviations in the far-focus region, while IFCNN distorts the near-focus image. Furthermore, these methods are unable to effectively remove the distortion around the flowers, while we introduce an MFF-specific loss, which achieves the most visually appealing results for the flower." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 610, + 400, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 400, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 400, + 623 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 544, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 544, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 544, + 651 + ], + "type": "text", + "content": "We conduct ablation studies to verify the effect of our TC-MoA and network architecture." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "TC-MoA. To verify the effectiveness of TC-MoA, we remove it from our framework. As shown in the first and second rows of Table 4, the models trained by multi-task training outperform those trained for a single task, based on 5 commonly used metrics. Interestingly though, this" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7104" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 104, + 286, + 198 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 103 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 103 + ], + "type": "text", + "content": "Table 4. Ablation studies on TC-MoA. S and M represent single task training and multi-task training, while SA and MA denote single adapter and multiple adapters respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 104, + 286, + 198 + ], + "lines": [ + { + "bbox": [ + 48, + 104, + 286, + 198 + ], + "spans": [ + { + "bbox": [ + 48, + 104, + 286, + 198 + ], + "type": "table", + "html": "
TaskSMSAMA\\(Q_{\\text{abf}}\\)\\(Q_p\\)FMI\\(Q_c\\)SSIM
VIF✓✓0.59840.40730.88570.63350.4540
✓✓0.59970.41160.88620.63570.4544
✓✓0.60070.41190.88620.63650.4550
MEF✓✓0.63850.59160.88750.57540.4115
✓✓0.63620.58990.88750.57650.4106
✓✓0.64490.59800.88830.57760.4116
MFF✓✓0.65170.67330.87970.77020.6774
✓✓0.65620.67990.88060.77420.6792
✓✓0.65680.68110.88080.77550.6794
", + "image_path": "aa7a9ed72a98fb2ac302948e04ecba4335435af80f8d898f2a0ff6225863360a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 209, + 286, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 286, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 286, + 280 + ], + "type": "text", + "content": "rule reverses in the case of MEF, indicating that inter-task competition occurs for models trained directly with multiple tasks. Comparatively, the model incorporating the multiple adapters improves the performance across all metrics, suggesting that TC-MoA is dynamically compatible with different tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 286, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 286, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 286, + 460 + ], + "type": "text", + "content": "Adapter-based Fine-tuning. We performed ablation experiments on the backbone and adapter methods, as shown in Table 5. The fine-tuning method of \"FrozenBackbone\" represents freezing the entire backbone and only unfreezing the final linear layer. Compared to the approach of \"Frozen-Backbone\", our TC-MoA achieves significant fusion performance improvements by introducing a small number of learnable parameters. More importantly, our method outperforms the existing adapter-based visual fine-tuning approach AdaptFormer [4] on both base and large versions of pre-trained models, demonstrating the superiority of the TC-MoA structure in fusion tasks. In addition to this, Table 9 shows our base version achieves competitive performance on the VIF task, which can be further enhanced by the large version." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 468, + 181, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 181, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 181, + 480 + ], + "type": "text", + "content": "4.6. Analysis and Discussion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 486, + 286, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 286, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 286, + 593 + ], + "type": "text", + "content": "Efficiency. We have observed the speed issues brought about by the massive parameters of the pre-trained models. To address this, we have optimized the ViT architecture by the shifted windows, which will be detailed in Sec. 8. After optimization, for multi-task with arbitrary-size inputs, Table 6 shows the base and large versions of the model have accelerated by " + }, + { + "bbox": [ + 46, + 486, + 286, + 593 + ], + "type": "inline_equation", + "content": "178\\%" + }, + { + "bbox": [ + 46, + 486, + 286, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 286, + 593 + ], + "type": "inline_equation", + "content": "167\\%" + }, + { + "bbox": [ + 46, + 486, + 286, + 593 + ], + "type": "text", + "content": ", resulting in acceptable FPS (Frames Per Second) compared with other methods. The \"Frozen\" refers to the vanilla ViT architecture." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 601, + 286, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 601, + 286, + 622 + ], + "spans": [ + { + "bbox": [ + 85, + 601, + 286, + 622 + ], + "type": "interline_equation", + "content": "p r o m p t _ {x} ^ {\\prime} = \\mu + \\alpha (p r o m p t _ {x} - \\mu) + \\beta \\tag {8}", + "image_path": "cc3d90bd1b5d77736db3247e6a59c6242f5600621cdc6c475633850731d7aaf6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 87, + 622, + 247, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 622, + 247, + 634 + ], + "spans": [ + { + "bbox": [ + 87, + 622, + 247, + 634 + ], + "type": "interline_equation", + "content": "p r o m p t _ {y} = \\mu + \\alpha (p r o m p t _ {y} - \\mu) - \\beta", + "image_path": "3762954183771d1a722eea1cfe630034e2db483d1d03406792bec3862b573de6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": "Prompt Controllability. Employing the formula noted in Eq. (8), we manipulate both the scale and shift of the prompt provided to the trained model, where " + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mu = 0.5" + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": " represents the mean, while " + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": " denote the scaling and shifting factors, respectively. We conduct the manipulation on some samples and present an example in Fig. 6 (a)." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 93, + 545, + 167 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 545, + 92 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 92 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 92 + ], + "type": "text", + "content": "Table 5. Ablation studies on fine-tuning. " + }, + { + "bbox": [ + 306, + 70, + 545, + 92 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 306, + 70, + 545, + 92 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 70, + 545, + 92 + ], + "type": "inline_equation", + "content": "\\mathbb{L}" + }, + { + "bbox": [ + 306, + 70, + 545, + 92 + ], + "type": "text", + "content": " represent the use of ViT-base and ViT-large structures respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 93, + 545, + 167 + ], + "lines": [ + { + "bbox": [ + 308, + 93, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 308, + 93, + 545, + 167 + ], + "type": "table", + "html": "
ArchitectureVIFMEFMFF
\\( Q_{abf} \\)\\( Q_p \\)SSIM\\( Q_{abf} \\)\\( Q_p \\)SSIM\\( Q_{abf} \\)\\( Q_p \\)SSIM
FrozenBackbone\\( _\\text{B} \\)0.1750.0890.2750.2500.1390.2270.2950.1890.466
AdaptFormet\\( _\\text{B} \\)0.5310.3360.4270.5740.5040.3770.5790.5610.651
TC-Mo\\( A_\\text{B} \\)0.5760.3960.4500.6040.5400.3940.6120.6190.660
FrozenBackbone\\( _\\text{L} \\)0.3300.2380.3630.4520.4080.3330.4500.4280.596
AdaptFormet\\( _\\text{L} \\)0.5760.3920.4460.6160.5630.3960.6400.6500.671
TC-Mo\\( A_\\text{L} \\)0.6010.4120.4550.6450.5980.4120.6570.6810.679
", + "image_path": "6c9d4ca09556d98fbc0af57cc2cf3021827da587988b2530c77ebf3690e52c9a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 183, + 545, + 236 + ], + "blocks": [ + { + "bbox": [ + 312, + 172, + 539, + 182 + ], + "lines": [ + { + "bbox": [ + 312, + 172, + 539, + 182 + ], + "spans": [ + { + "bbox": [ + 312, + 172, + 539, + 182 + ], + "type": "text", + "content": "Table 6. Efficiency and scalability comparisons of approaches." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 183, + 545, + 236 + ], + "lines": [ + { + "bbox": [ + 308, + 183, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 308, + 183, + 545, + 236 + ], + "type": "table", + "html": "
U2FDDFMCDDFrozenBTC-MoABFrozenLTC-MoAL
Total Params (M)1.32-1.78111.65115.40339.12348.70
Trainable Params (M)1.32-1.78-3.87-9.58
FPS (VIF 640×512)4.720.012.521.173.330.601.60
FPS (MEF arbitrary-size)2.21--1.373.580.711.86
FPS (MFF arbitrary-size)1.86--1.393.980.711.95
", + "image_path": "f7c8c4fae4e6e4c7ae4e35d5761e6c57ee47f86eed902ee05716fd89360b59c8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 245, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 545, + 437 + ], + "type": "text", + "content": "Obviously, as the value of " + }, + { + "bbox": [ + 304, + 245, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 245, + 545, + 437 + ], + "type": "text", + "content": " increases, the brightness and texture of the fused image increasingly resemble the source image " + }, + { + "bbox": [ + 304, + 245, + 545, + 437 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 245, + 545, + 437 + ], + "type": "text", + "content": ", and inversely, the source image " + }, + { + "bbox": [ + 304, + 245, + 545, + 437 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 245, + 545, + 437 + ], + "type": "text", + "content": ". Thus, " + }, + { + "bbox": [ + 304, + 245, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 245, + 545, + 437 + ], + "type": "text", + "content": " can be considered as a shifting of the dominant intensity bias. Moreover, a higher value of " + }, + { + "bbox": [ + 304, + 245, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 245, + 545, + 437 + ], + "type": "text", + "content": " tends to favor one source image over the other in the image patches, which can be considered as a scaling of the dominant intensity deviation. For example, outputs with " + }, + { + "bbox": [ + 304, + 245, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\beta = -0.3" + }, + { + "bbox": [ + 304, + 245, + 545, + 437 + ], + "type": "text", + "content": " are more similar to the under-exposure image globally. As " + }, + { + "bbox": [ + 304, + 245, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 245, + 545, + 437 + ], + "type": "text", + "content": " increases from 0 to 3, the building areas of the fused image gradually become brighter, but the brightness of the clouds remains basically unchanged. This suggests that changing the scale factor of the prompt motivates the model to select different regions. Overall, this experiment demonstrates that the prompt is controllable and can be linearly transformed to obtain fuse images in different degrees." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 437, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 581 + ], + "type": "text", + "content": "Generalizability. Our method is highly generalizable to tasks with similar fusion rules by zero-shot fusion. Specifically, we take samples from medical image fusion dataset [1] and pan-sharpening dataset generated by the Quickbird satellite, which are fused based on VIF's paradigm. We obtain a base fused image with " + }, + { + "bbox": [ + 304, + 437, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 304, + 437, + 545, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 437, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\beta = 0" + }, + { + "bbox": [ + 304, + 437, + 545, + 581 + ], + "type": "text", + "content": ", as shown in Fig. 6 (b) and (c). However, without any prior knowledge of these unknown tasks, the output fused image is an under-performing image, missing many high and low-frequency information. Interestingly, by manipulating the prompt, we can find a suitable fusion rule for these new tasks, obtaining reasonable fusion results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Routers Controllability. As shown in Fig. 7, the images generated by the MFF router are darker than VIF in areas without gradient guidance. The MFF selects the reference sources based on gradients, it tends to average the information from sources without gradient guidance. In contrast, the VIF router tends to preserve the maximum intensity information. Moreover, at the bottom of the image, the MFF router tends to preserve the gradient information depending on a single source, while the VIF router preserves the gradient information from both sources. Additionally, by manipulating the combination of different routers, our model" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7105" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 542, + 256 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 542, + 256 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 542, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 542, + 256 + ], + "type": "image", + "image_path": "8e23e1eab439c3742adbc4d23d079208fcc54fcc98b946502f79a914d6d7ca9e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 257, + 432, + 268 + ], + "lines": [ + { + "bbox": [ + 160, + 257, + 432, + 268 + ], + "spans": [ + { + "bbox": [ + 160, + 257, + 432, + 268 + ], + "type": "text", + "content": "Figure 6. Visualisation of controllability and generalization of our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 274, + 284, + 395 + ], + "blocks": [ + { + "bbox": [ + 50, + 274, + 284, + 395 + ], + "lines": [ + { + "bbox": [ + 50, + 274, + 284, + 395 + ], + "spans": [ + { + "bbox": [ + 50, + 274, + 284, + 395 + ], + "type": "image", + "image_path": "c946d116226f28e58b216f15b60c42222f9c3108b2a24acfbf30c8889e77b2cb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 399, + 280, + 411 + ], + "lines": [ + { + "bbox": [ + 53, + 399, + 280, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 280, + 411 + ], + "type": "text", + "content": "Figure 7. Visualisation of different routers on medical images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 420, + 287, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 287, + 444 + ], + "type": "text", + "content": "copes with controlling the dynamic changing fusion degree, obtaining more controllable results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "text", + "content": "Hyperparameters. We conducted hyperparameters analysis on two aspects: the number of experts " + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "inline_equation", + "content": "(N)" + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "text", + "content": " and the interval between two TC-MoAs " + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "inline_equation", + "content": "(\\tau)" + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "text", + "content": ". We utilize the ViT-large as our backbone, which consists of 24 transformer blocks in the encoder and 8 in the decoder. If " + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "inline_equation", + "content": "\\tau = 4" + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "text", + "content": ", our model contains 8 instances of TC-MoA. We employ two widely-used metrics " + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "inline_equation", + "content": "Q_{abf}" + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "text", + "content": " and SSIM to inform the selection of hyperparameters. As presented in Fig. 8, as " + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "text", + "content": " increases, the performance initially increases but eventually declines, with the peak performance detected at " + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "text", + "content": ". This suggests that increasing the number of experts does not necessarily enhance the model's performance. Similarly, when " + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "inline_equation", + "content": "\\tau = 4" + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "text", + "content": ", the model demonstrates the best performance across multitask. Hence, our experiments are conducted under " + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "inline_equation", + "content": "\\tau = 4" + }, + { + "bbox": [ + 47, + 445, + 287, + 623 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 634, + 119, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 119, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 119, + 646 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "text", + "content": "In this paper, we propose a task-customized mixture of adapters for general image fusion. With the help of TCMoA, different fusion tasks customize their respective mixture of adapters to obtain prompt guidance for multiple-source image fusion in a unified model. To ensure compat" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 274, + 543, + 392 + ], + "blocks": [ + { + "bbox": [ + 313, + 274, + 543, + 392 + ], + "lines": [ + { + "bbox": [ + 313, + 274, + 543, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 543, + 392 + ], + "type": "image", + "image_path": "9638f7640b7903690f40bacc9fd574573f9c7514bd2dfea28243dd51395cee1a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 399, + 512, + 411 + ], + "lines": [ + { + "bbox": [ + 339, + 399, + 512, + 411 + ], + "spans": [ + { + "bbox": [ + 339, + 399, + 512, + 411 + ], + "type": "text", + "content": "Figure 8. Experiments on the hyperparameters." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 419, + 545, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 545, + 516 + ], + "type": "text", + "content": "ibility with different tasks while maintaining complementarity for multi-source images, we further propose mutual information regularization to constrain these adapters. Experimental results have shown that TC-MoA achieves advanced performance in general image fusion against the competing methods. In addition, TC-MoA demonstrates strong prompt controllability and router controllability to perform flexible manipulation on the fused images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 526, + 416, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 416, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 416, + 540 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": "This work was sponsored in part by the National Key R&D Program of China 2022ZD0116500, in part by the National Natural Science Foundation of China (62222608, 62106171, U23B2049, 61925602), in part by the Haihe Lab of ITAI under Grant 22HHXCJC00002, in part by the Tianjin Natural Science Foundation under Grant 21JCYBJC00580, in part by Tianjin Natural Science Funds for Distinguished Young Scholar under Grant 23JCJQJC00270, and in part by the Key Laboratory of Big Data Intelligent Computing, Chongqing University of Posts and Telecommunications under Grant BDIC-2023-A-008. This work was also sponsored by CAAI-CANN Open Fund, developed on OpenI Community. The authors appreciate the suggestions from CVPR anonymous peer reviewers." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7106" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Aanlib. Ihttp://www.med.harvard.edu/AANLIB/home.html, (Accessed 5 January 2020). 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 156 + ], + "type": "text", + "content": "[2] Jianrui Cai, Shuhang Gu, and Lei Zhang. Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing, 27(4):2049-2062, 2018. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 159, + 286, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 159, + 286, + 212 + ], + "spans": [ + { + "bbox": [ + 54, + 159, + 286, + 212 + ], + "type": "text", + "content": "[3] Bing Cao, Yiming Sun, Pengfei Zhu, and Qinghua Hu. Multi-modal gated mixture of local-to-global experts for dynamic image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 23555-23564, 2023. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 214, + 286, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 214, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 54, + 214, + 286, + 267 + ], + "type": "text", + "content": "[4] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 286, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 286, + 312 + ], + "type": "text", + "content": "[5] Nedeljko Cvejic, Artur Loza, David Bull, and Nishan Canagarajah. A similarity metric for assessment of image fusion algorithms. International journal of signal processing, 2(3): 178-182, 2005. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 315, + 286, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 315, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 315, + 286, + 357 + ], + "type": "text", + "content": "[6] Wang Di, Liu Jinyuan, Fan Xin, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In International Joint Conference on Artificial Intelligence (IJCAI), 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 358, + 286, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 286, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 286, + 401 + ], + "type": "text", + "content": "[7] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. The Journal of Machine Learning Research, 23(1):5232-5270, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 286, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 286, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 286, + 446 + ], + "type": "text", + "content": "[8] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification with dynamic evidential fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2551-2566, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 448, + 286, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 448, + 286, + 500 + ], + "spans": [ + { + "bbox": [ + 53, + 448, + 286, + 500 + ], + "type": "text", + "content": "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 502, + 286, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 286, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 286, + 545 + ], + "type": "text", + "content": "[10] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 547, + 286, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 286, + 590 + ], + "type": "text", + "content": "[11] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 286, + 635 + ], + "type": "text", + "content": "[12] Xinyu Jia, Chuang Zhu, Minzhen Li, Wenqi Tang, and Wenli Zhou. Llvip: A visible-infrared paired dataset for low-light vision. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3496-3504, 2021. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 286, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 286, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 286, + 679 + ], + "type": "text", + "content": "[13] Xin Jin, Xiuliang Xi, Ding Zhou, Xiaoxuan Ren, Jie Yang, and Qian Jiang. An unsupervised multi-focus image fusion method based on transformer and u-net. IET Image Processing, 17(3):733-746, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "type": "text", + "content": "[14] Harpreet Kaur, Deepika Koundal, and Virender Kadyan. Image fusion techniques: a survey. Archives of computational methods in Engineering, 28:4425-4447, 2021. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[15] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. CoRR, abs/2006.16668, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "text", + "content": "[16] Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images. IEEE Transactions on Image Processing, 28(5):2614-2623, 2018. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 165, + 545, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 165, + 545, + 197 + ], + "spans": [ + { + "bbox": [ + 308, + 165, + 545, + 197 + ], + "type": "text", + "content": "[17] Hui Li, Xiao-Jun Wu, and Josef Kittler. Rfn-nest: An end-to-end residual fusion network for infrared and visible images. Information Fusion, 73:72-86, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 200, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 200, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 308, + 200, + 545, + 232 + ], + "type": "text", + "content": "[18] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 235, + 545, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 235, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 308, + 235, + 545, + 278 + ], + "type": "text", + "content": "[19] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems, 35:109-123, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 281, + 545, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 281, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 308, + 281, + 545, + 324 + ], + "type": "text", + "content": "[20] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Fusion from decomposition: A self-supervised decomposition approach for image fusion. In European Conference on Computer Vision, pages 719-735. Springer, 2022. 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 327, + 545, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 402 + ], + "type": "text", + "content": "[21] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5811, 2022. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 405, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 437 + ], + "type": "text", + "content": "[22] Yu Liu, Xun Chen, Hu Peng, and Zengfu Wang. Multi-focus image fusion with a deep convolutional neural network. Information Fusion, 36:191-207, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 440, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 440, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 308, + 440, + 545, + 472 + ], + "type": "text", + "content": "[23] Yu Liu, Lei Wang, Juan Cheng, Chang Li, and Xun Chen. Multi-focus image fusion: A survey of the state of the art. Information Fusion, 64:71-91, 2020. 1, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 475, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 545, + 540 + ], + "type": "text", + "content": "[24] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 543, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 543, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 308, + 543, + 545, + 575 + ], + "type": "text", + "content": "[25] Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible image fusion methods and applications: A survey. Information fusion, 45:153-178, 2019. 1, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 578, + 545, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 545, + 631 + ], + "type": "text", + "content": "[26] Jiayi Ma, Linfeng Tang, Fan Fan, Jun Huang, Xiaoguang Mei, and Yong Ma. Swinfusion: Cross-domain long-range learning for general image fusion via swim transformer. IEEE/CAA Journal of Automatica Sinica, 9(7):1200-1217, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 635, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 667 + ], + "type": "text", + "content": "[27] Kede Ma, Kai Zeng, and Zhou Wang. Perceptual quality assessment for multi-exposure image fusion. IEEE Transactions on Image Processing, 24(11):3345-3356, 2015. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[28] Kede Ma, Zhengfang Duanmu, Hanwei Zhu, Yuming Fang, and Zhou Wang. Deep guided learning for fast multi-exposure image fusion. IEEE Transactions on Image Processing, 29:2808-2819, 2019. 1, 4, 5" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7107" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[29] Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with limoe: the language-image mixture of experts. Advances in Neural Information Processing Systems, 35:9564-9576, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "type": "text", + "content": "[30] Linhao Qu, Shaolei Liu, Manning Wang, and Zhijian Song. Transmef: A transformer-based multi-exposure image fusion framework using self-supervised multi-task learning. In Proceedings of the AAAI conference on artificial intelligence, pages 2126-2134, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "type": "text", + "content": "[31] K Ram Prabhakar, V Sai Srikar, and R Venkatesh Babu. Deepfuse: A deep unsupervised approach for exposure fusion with extreme exposure image pairs. In Proceedings of the IEEE international conference on computer vision, pages 4714-4722, 2017. 1, 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 238, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 238, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 238, + 287, + 282 + ], + "type": "text", + "content": "[32] Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual adapters. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 282, + 287, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 282, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 282, + 287, + 336 + ], + "type": "text", + "content": "[33] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 337, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 287, + 381 + ], + "type": "text", + "content": "[34] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Det-fusion: A detection-driven infrared and visible image fusion network. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4003-4011, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "type": "text", + "content": "[35] Linfeng Tang, Jiteng Yuan, and Jiayi Ma. Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network. Information Fusion, 82:28-42, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "text", + "content": "[36] Wei Tang, Fazhi He, and Yu Liu. Ydtr: Infrared and visible image fusion via y-shape dynamic transformer. IEEE Transactions on Multimedia, 2022. 1, 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "type": "text", + "content": "[37] Z. Wang, E.P. Simoncelli, and A.C. Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402 Vol.2, 2003. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "type": "text", + "content": "[38] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "text", + "content": "[39] Zhishe Wang, Yanlin Chen, Wenyu Shao, Hui Li, and Lei Zhang. Swinfuse: A residual swim transformer fusion network for infrared and visible images. IEEE Transactions on Instrumentation and Measurement, 71:1-12, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "type": "text", + "content": "[40] Fang Xu, Jinghong Liu, Yueming Song, Hui Sun, and Xuan Wang. Multi-exposure image fusion techniques: A comprehensive review. Remote Sensing, 14(3):771, 2022. 1, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "text", + "content": "[41] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[42] Han Xu, Jiayi Ma, Zhuliang Le, Junjun Jiang, and Xiaojie Guo. Fusiondn: A unified densely connected network for image fusion. In Proceedings of the AAAI conference on artificial intelligence, pages 12484-12491, 2020. 2, 5, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 688 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[43] Han Xu, Jiayi Ma, and Xiao-Ping Zhang. Mef-gan: Multiexposure image fusion via generative adversarial networks. IEEE Transactions on Image Processing, 29:7203-7216, 2020. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[44] Hao Zhang and Jiayi Ma. Sdnet: A versatile squeeze-and-decomposition network for real-time image fusion. International Journal of Computer Vision, 129:2761-2785, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[45] Hao Zhang, Han Xu, Yang Xiao, Xiaojie Guo, and Jiayi Ma. Rethinking the image fusion: A fast unified image fusion network based on proportional maintenance of gradient and intensity. In Proceedings of the AAAI conference on artificial intelligence, pages 12797-12804, 2020. 2, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 207, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 207, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 207, + 545, + 261 + ], + "type": "text", + "content": "[46] Hao Zhang, Zhuliang Le, Zhenfeng Shao, Han Xu, and Jiayi Ma. Mff-gan: An unsupervised generative adversarial network with adaptive and gradient joint constraints for multifocus image fusion. Information Fusion, 66:40-53, 2021. 2, 4, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 262, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 262, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 262, + 545, + 307 + ], + "type": "text", + "content": "[47] Juncheng Zhang, Qingmin Liao, Shaojun Liu, Haoyu Ma, Wenming Yang, and Jing-Hao Xue. Real-mff: A large realistic multi-focus image dataset with ground truth. Pattern Recognition Letters, 138:370-377, 2020. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 308, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 352 + ], + "type": "text", + "content": "[48] Qingyang Zhang, Haitao Wu, Changqing Zhang, Qinghua Hu, Huazhu Fu, Joey Tianyi Zhou, and Xi Peng. Provable dynamic fusion for low-quality multimodal data. In International Conference on Machine Learning, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 353, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 353, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 353, + 545, + 385 + ], + "type": "text", + "content": "[49] Xingchen Zhang. Benchmarking and comparing multi-exposure image fusion algorithms. Information Fusion, 74: 111-131, 2021. 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 387, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 387, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 545, + 420 + ], + "type": "text", + "content": "[50] Xingchen Zhang. Deep learning-based multi-focus image fusion: A survey and a comparative study. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 420, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 420, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 420, + 545, + 464 + ], + "type": "text", + "content": "[51] Yu Zhang, Yu Liu, Peng Sun, Han Yan, Xiaolin Zhao, and Li Zhang. Ifcnn: A general image fusion framework based on convolutional neural network. Information Fusion, 54: 99-118, 2020. 2, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 465, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 465, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 465, + 545, + 510 + ], + "type": "text", + "content": "[52] Zixiang Zhao, Shuang Xu, Chunxia Zhang, Junmin Liu, Pengfei Li, and Jiangshe Zhang. Didfuse: Deep image decomposition for infrared and visible image fusion. arXiv preprint arXiv:2003.09210, 2020. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 510, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 510, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 308, + 510, + 545, + 564 + ], + "type": "text", + "content": "[53] Zixiang Zhao, Shuang Xu, Jiangshe Zhang, Chengyang Liang, Chunxia Zhang, and Junmin Liu. Efficient and model-based infrared and visible image fusion via algorithm unrolling. IEEE Transactions on Circuits and Systems for Video Technology, 32(3):1186-1196, 2021. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 566, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 566, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 566, + 545, + 632 + ], + "type": "text", + "content": "[54] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5906–5916, 2023. 1, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 633, + 545, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 633, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 308, + 633, + 545, + 688 + ], + "type": "text", + "content": "[55] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. Ddfm: denoising diffusion model for multi-modality image fusion. arXiv preprint arXiv:2303.06840, 2023. 1, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7108" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/a6a833ff-f398-48fe-909b-70ff81bf1b94_content_list.json b/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/a6a833ff-f398-48fe-909b-70ff81bf1b94_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6a7443a9a321528f91c35a256c4aca4613b2bb7b --- /dev/null +++ b/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/a6a833ff-f398-48fe-909b-70ff81bf1b94_content_list.json @@ -0,0 +1,1661 @@ +[ + { + "type": "text", + "text": "Task-Driven Exploration: Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection", + "text_level": 1, + "bbox": [ + 120, + 128, + 851, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jin Yang, Ping Wei*, Huan Li, Ziyang Ren \nNational Key Laboratory of Human-Machine Hybrid Augmented Intelligence Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University \n{jin.yang, lh875056558, rzyrzy}@stu.xjtu.edu.cn, pingwei@xjtu.edu.cn", + "bbox": [ + 176, + 203, + 790, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video moment retrieval and highlight detection are two highly valuable tasks in video understanding, but until recently they have been jointly studied. Although existing studies have made impressive advancement recently, they predominantly follow the data-driven bottom-up paradigm. Such paradigm overlooks task-specific and inter-task effects, resulting in poor model performance. In this paper, we propose a novel task-driven top-down framework TaskWeave for joint moment retrieval and highlight detection. The framework introduces a task-decoupled unit to capture task-specific and common representations. To investigate the interplay between the two tasks, we propose an inter-task feedback mechanism, which transforms the results of one task as guiding masks to assist the other task. Different from existing methods, we present a task-dependent joint loss function to optimize the model. Comprehensive experiments and in-depth ablation studies on QVHighlights, TVSum, and CharadesSTA datasets corroborate the effectiveness and flexibility of the proposed framework. Codes are available at github.com/EdenGabriel/TaskWeave.", + "bbox": [ + 75, + 340, + 473, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 686, + 209, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As videos are prevailing in a wide range of applications, the diversity and massive scales of video content have posed unprecedented challenges in finding relevant moments based on user queries. To this end, the moment retrieval (MR) [60, 63, 65] and highlight detection (HD) [4, 56] tasks have emerged recently. MR aims to retrieve video moments that are relevant to the given query [14, 62]. HD aims to predict the clip-level saliency scores in the video [44].", + "bbox": [ + 75, + 710, + 468, + 847 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As MR and HD tasks are closely related, they have been jointly addressed and achieved breakthroughs recently", + "bbox": [ + 75, + 847, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/66e4867e1d41709697bfc3ea6505e5271921a04678a42e0b1c7df877dbbf6dce.jpg", + "image_caption": [ + "Figure 1. Although the positive temporal intervals of moment retrieval and highlight detection exhibit high overlap, they pursue different objectives." + ], + "image_footnote": [], + "bbox": [ + 503, + 310, + 893, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[20, 25, 29, 32, 36, 53]. The existing joint approaches in general utilize a shared backbone to learn the multimodal features as the common representations for MR and HD. Then a MR prediction head is employed for moments localization and a HD prediction head predicts saliency scores. These methods adhere to the bottom-up, data-driven paradigm, i.e. they capture common features from the input data and then utilize the features for different tasks. The effectiveness of these methods is built upon the premise of the high correlation between MR and HD. As illustrated in Fig. 1, the task MR and HD share the identical query and video inputs, additionally exhibiting a substantial temporal overlap between their respective positive temporal intervals.", + "bbox": [ + 496, + 546, + 893, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, the bottom-up, data-driven paradigm tends to excessively rely on the common features, but overlooks the inherent specific characteristics of MR and HD. This tendency might simplify the joint modeling as a problem of feature fusion, without considering the interplay between the two tasks. The distinct objectives pursued by MR and HD rely on distinct task-specific characteristics. Unfortunately, existing methods overlook the specificity.", + "bbox": [ + 496, + 746, + 895, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the aforementioned issues, we believe it is essential to leverage the fundamental multi-modal data to", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 101, + 887, + 230, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18308", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mine commonalities across tasks (bottom-up), while also strengthen the awareness of task-specific characteristics (top-down). To this end, we propose a novel paradigm TaskWeave from a task-driven perspective. The key idea is to jointly address the tasks MR and HD by considering the commonality, specificity, and interplay of MR and HD.", + "bbox": [ + 75, + 90, + 467, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To effectively capture the commonality and specificity, we design a task-decoupled unit, i.e. a shared expert to capture common features and two task-specific experts to acquire the distinct characteristics. In order to investigate the interplay between MR and HD in-depth, we design an inter-task feedback mechanism. It converts the predictions of MR/HD into mask information, which are fed back to the input of the HD/MR prediction head. Furthermore, we introduce a principled task-dependent joint loss in which the task-specific weights are dynamically adjusted, rather than manually tuned.", + "bbox": [ + 75, + 181, + 467, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct experiments on the QVHighlights [25] dataset to validate the effectiveness of the proposed method. Moreover, we also conduct experiments for two individual tasks on the datasets Charades-STA [14] (moment retrieval) and TVSUM [43] (highlight detection). The proposed approach outperforms the existing methods.", + "bbox": [ + 75, + 347, + 467, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key contributions of this paper are three folds.", + "bbox": [ + 94, + 438, + 429, + 453 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. It proposes a novel task-driven, top-down framework for joint moment retrieval and highlight detection.", + "2. It introduces a task-decoupled unit, an inter-task feedback mechanism, and a principled task-dependent loss.", + "3. It achieves state-of-the-art performance on three datasets. The ablation study validates the methods." + ], + "bbox": [ + 88, + 454, + 467, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 558, + 217, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We review the related work from four aspects.", + "bbox": [ + 96, + 584, + 401, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moment Retrieval. Existing approaches for MR mainly include two groups. One group follows a two-stage procedure [1, 14, 60, 65], which involves generating candidate temporal intervals and ranking them based on the correlation with the query. The other group directly regresses the temporal interval based on the aligned visual-text features [8, 37, 58, 59, 61]. Moreover, most datasets provide only one moment annotation for each video-query pair [1, 14], which does not align with real-world scenarios.", + "bbox": [ + 75, + 599, + 467, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Highlight Detection. The saliency score in the highlight detection represents the relevance of a video clip to the given query. Most prior highlight detection benchmark datasets are query-agnostic [15, 44]. The saliency scores for video clips remained constant regardless of the query. As a result, some previous approaches treated HD as a solely visual task [4, 44, 49, 56].", + "bbox": [ + 75, + 734, + 467, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MR and HD tasks are traditionally studied separately. They have been jointly addressed recently as the introduction of QVHighlights dataset [25]. QVHighlights provides multiple moment-annotations for each query and ensures", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "these moments are uniformly distributed throughout the video. It also provides query-dependent highlightness annotations. With QVHighlights, the model Moment-DETR [25] is proposed for joint MR and HD. Following Moment-DETR, a growing number of approaches have been proposed to accomplish the joint task [20, 29, 32, 36]. However, these methods adopt the data-driven and bottom-up paradigm. Different from them, we propose a novel task-driven and top-down paradigm.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Transformers. Transformer-based models [2] have brought huge achievement in both the image and video related domains [5, 11, 27, 28, 54, 55]. One of the most well-known methods is DETR [5], which regards object detection as a set prediction problem. Its end-to-end prediction procedure eliminates the intermediate or post-processing steps. On the other hand, some studies have employed cross-attention mechanism [35, 36, 50] to inject multi-modal data into the Transformer architecture. In this paper, we also adopt a DETR-like architecture [26, 31]. However, different from those methods, we incorporate diverse network architectures for feature extraction. Our focus lies in the collaboration of the network architectures.", + "bbox": [ + 496, + 227, + 890, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-task Learning. Multi-task learning (MTL) aims to train a single model for multiple tasks [7]. The most straightforward approach is to utilize a shared backbone to extract common features, which relies on the data-driven manner. However, it often leads to suboptimal performance for unrelated tasks and lacks flexibility. In response, some studies introduce MOE [19] and MMOE [33], which utilize a set of the shared experts in place of the shared bottom layer. To mitigate the seesaw phenomenon of multi-task learning, PLE [45] employs separate experts for each task while still retaining the shared experts. From the perspective of MTL, existing approaches for joint MR and HD follow the shared bottom paradigm. Moreover, directly applying MTL methods to joint MR and HD does not yield favorable results. Existing MTL methods are complex, therefore direct usage of them leads to a sharp increase in model complexity. This might decrease the performance. To this end, we aim to design an effective task-driven framework to jointly address MR and HD.", + "bbox": [ + 496, + 425, + 890, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 729, + 633, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 753, + 609, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a text query with $W$ words and an untrimmed video composed of $N$ clips, the objective of the joint moment retrieval (MR) and highlight detection (HD) is to localize the center coordinate $q_{c}$ and width $q_{w}$ of temporal intervals that are relevant to the text query, in addition to ranking clip-wise saliency scores.", + "bbox": [ + 496, + 777, + 890, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Architecture Overview. Given the high correlation between moment retrieval and highlight detection tasks, the", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18309", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a101a6befdf96406fa8311a1c3ba8d48453ca0993b6bbf611c746faf88eff16d.jpg", + "image_caption": [ + "Figure 2. The overall pipeline of the proposed task-driven model TaskWeave. We propose the task-decoupled unit to capture task-specific and common features. Various experts can adopt different network implementations, showcasing the flexibility of the model. Inter-task feedback mechanism is designed to investigate the influence between both tasks. There are two feedback manners: Moment-guided and Highlightness-guided feedback. The principled task-dependent joint loss is introduced for jointly optimize the model." + ], + "image_footnote": [], + "bbox": [ + 81, + 92, + 893, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "most intuitive approach is using a shared backbone in conjunction with two task-specific prediction heads. It is a data-driven paradigm that is commonly employed by previous methods, for its simplicity and ease of implementation. However, the task-specific characteristics are inherently present, since MR and HD have distinct objectives. Additionally, the interplay between MR and HD should also be considered, which can enhance the model's performance.", + "bbox": [ + 75, + 395, + 468, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall pipeline of our approach is illustrated in Fig. 2. We employed the frozen video/text-encoder backbones to extract the video/text features, while ensuring their dimensions remained consistent at $D$ by the projection. These features are utilized for multi-modal fusion through methods such as cross-attention or concatenation, resulting in query-related video representations $Z \\in \\mathbb{R}^{N \\times D}$ . In our paper, we extract these representations through the cross-attention. After cross-attention, inspired by [41, 46], a 1D Max Pooling (MaxPool1d) with the kernel size 5, stride 1 and padding 2 is utilized to eliminate the rank loss problem in the attention mechanism. These representations are fed into the task-decoupled unit to capture the task-related features $X_{\\mathrm{mr}} \\in \\mathbb{R}^{N \\times D}$ , $X_{\\mathrm{hd}} \\in \\mathbb{R}^{N \\times D}$ . Then, we employ the task-specific decoders with inter-task feedback mechanism to make predictions for moments localization and clip-wise saliency scores. We introduce the principled task-dependent loss to jointly optimize the model.", + "bbox": [ + 75, + 517, + 470, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Task-decoupled Unit", + "text_level": 1, + "bbox": [ + 76, + 800, + 272, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since moment retrieval and highlight detection have distinct objectives, the specificity of each task should be considered, rather than solely focusing on their commonalities. For this purpose, we propose a task-decoupled unit from a task-driven perspective to capture the task-related features,", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which involves task-specific features and common features.", + "bbox": [ + 498, + 393, + 888, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The task-decoupled unit is depicted in Fig. 2. Inspired by the attention mechanism [2], the query-related video representation $Z$ is initially fed into two task-specific mappers, mr-mapper and hd-mapper. Each mapper is implemented using one-layer feed-forward network. The output of each task-specific mapper is directed into both the respective task-specific expert network (MR-specific Expert or HD-specific Expert) and the Shared Expert network.", + "bbox": [ + 496, + 410, + 890, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thanks to the design of our task-decoupled framework, each expert can employ various networks, such as convolutional network, Transformer, and feed-forward network. It offers more flexibility in addressing the joint MR and HD tasks. In our study, we also investigate the influence of different expert networks, referring to Sec. 4 for more details. Consequently, we can capture the task-related features $X_{\\mathrm{mr}}$ and $X_{\\mathrm{hd}}$ via element-wise product between the output of the task-specific expert and that of the shared expert.", + "bbox": [ + 496, + 532, + 892, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The MR-specific feature $X_{\\mathrm{mr}}$ can be computed as:", + "bbox": [ + 519, + 669, + 854, + 684 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nX _ {\\mathrm {m r}} = \\mathcal {P} _ {\\mathrm {m r}} \\left(\\mathcal {M} _ {\\mathrm {m r}} (Z)\\right) \\otimes \\mathcal {S} \\left(\\mathcal {M} _ {\\mathrm {m r}} (Z)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 698, + 890, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{M}_{\\mathrm{mr}}(\\cdot)$ refers to the mr-mapper operation. $S(\\cdot)$ is the shared expert calculation. $\\otimes$ denotes the element-wise product. $\\mathcal{P}_{\\mathrm{mr}}(\\cdot)$ is the mr-specific expert calculation. The calculation method for the HD-specific feature $X_{\\mathrm{hd}}$ follows the similar approach.", + "bbox": [ + 496, + 728, + 890, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Inter-task Feedback Mechanism", + "text_level": 1, + "bbox": [ + 500, + 816, + 782, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is imperative to investigate the interplay between the moment retrieval and highlight detection tasks for their significant correlation. However, existing methods overlook this aspect and employ two decoders for direct prediction.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18310", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/65981cfcc0b0602aa6a2ab62b305e89cea7e8cb62f8892edc4cdd6f90957a42c.jpg", + "image_caption": [ + "(a) Moment-guided feedback" + ], + "image_footnote": [], + "bbox": [ + 76, + 93, + 467, + 218 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e16f690698437a1b2a93900d4c076fbda4288dc5896ed400196f377db6a77e3b.jpg", + "image_caption": [ + "(b) Highlightness-guided feedback", + "Figure 3. Illustration of the inter-task feedback mechanism. (a) moment-guided feedback manner. (b) highlightness-guided feedback manner." + ], + "image_footnote": [], + "bbox": [ + 81, + 244, + 467, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To resolve this issue, we propose an inter-task feedback mechanism, which contains two task-specific decoders (Transformer decoder for MR and lightweight decoder for HD) and two feedback manners (moment-guided feedback and highlightness-guided feedback). The output of one task is transformed into the mask to assist another task in the feedback mechanism. All components of the inter-task feedback mechanism is detailed as follows.", + "bbox": [ + 75, + 507, + 468, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer Decoder for MR. Previous studies have demonstrated that it is effective to employ dynamic anchors as queries for the decoder within the DETR-like structure [26, 31]. They iteratively update anchor boxes, thereby transforming the procedure for updating queries into the cascaded soft ROI-pooling. In our work, the Transformer decoder for MR follows the work [31]. Each moment query is represented by its temporal center coordinate $q_{c}$ and width $q_{w}$ . The output of the Transformer decoder for MR is denoted as $Y_{\\mathrm{mr}} \\in \\mathbb{R}^{N_q \\times 2}$ . $N_{q}$ is the number of moment queries and set to 10 in our work.", + "bbox": [ + 75, + 628, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lightweight Decoder for HD. The most straightforward method for the HD decoder would be utilizing one or more fully-connected layers, as seen in prior studies such as Moment-DETR [25] and UMT [32]. However, such design overlooks the diversity of video-query pairs [36], which provides identical criterias for the saliency score prediction of each video-query pair. Although QD-DETR [36] pro", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "poses a global saliency token to predict the saliency scores, it remains highly coupled with the encoder.", + "bbox": [ + 498, + 90, + 888, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Different from those methods, we introduce CNN structures in the decoder. CNNs can capture local details, ensuring accurate saliency predictions across varying queries within the same video. The structure of the lightweight decoder for HD prediction is illustrated in Fig. 3, which alternates between Linear and Conv1d layers. We denote the input of the HD decoder as $\\mathcal{J} \\in \\mathbb{R}^{N \\times d}$ . The saliency prediction $Y_{\\mathrm{hd}} \\in \\mathbb{R}^{N \\times 1}$ can be computed as follows:", + "bbox": [ + 496, + 119, + 890, + 241 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {\\mathrm {h d}} = D _ {\\mathrm {H D}} (\\mathcal {J}) + \\frac {\\mathbb {I} _ {i = 1} ^ {N} \\left(\\sum_ {k = 1} ^ {d} \\mathcal {J} ^ {i k}\\right)}{\\sqrt {d}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 247, + 890, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{I}_{i = 1}^{N}(\\cdot)$ represents the traversal operation. $D_{\\mathrm{HD}}(\\cdot)$ refers to the decoder network operation.", + "bbox": [ + 498, + 295, + 888, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moment-guided Feedback. We propose a moment-guided feedback manner to investigate the influence of MR on HD. Our key focus is on how to effectively utilize the results of moment retrieval for highlight detection. The implementation procedure of moment-guided feedback is illustrated in Fig. 3 (a).", + "bbox": [ + 496, + 327, + 890, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We transform the tentatively predicted moment results from the \"(center, width)\" format to the \"(start, end)\" format. These moment results are fed into the Moment-to-Mask (M2M) converter to generate moment-aware masks $\\xi_{\\mathrm{m2m}}$ . We initialize a clip-wise vector with a length of $N$ for each moment prediction. The indices in these vectors correspond to the clip indexes within the video. We binarize these vectors based on the moment results. We can obtain the clip indexes that are included in the moments through the prior information and set the values at those positions in the vector to 1, while the rest to 0. The obtained $N_q$ binary vectors are then summed and normalized with L2 norm.", + "bbox": [ + 496, + 417, + 892, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The obtained moment-aware masks $\\xi_{\\mathrm{m2m}}$ integrate the prediction results from $N_{q}$ queries and provide guidance for the highlight detection. In the moment-guided feedback manner, the input $\\mathcal{I}$ of the HD decoder is updated by $X_{\\mathrm{hd}} + X_{\\mathrm{hd}}*\\xi_{\\mathrm{m2m}}$", + "bbox": [ + 496, + 598, + 890, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Highlightness-guided Feedback. Similar to the moment-guided feedback, we introduce a highlightness-guided feedback manner with the Highlightness-to-Mask (H2M) converter, to explore the influence of HD on MR. As illustrated in Fig. 3 (b), the saliency score vector is binarized in the H2M to obtain the highlightness-aware mask $\\xi_{\\mathrm{h2m}}$ . Values in the saliency score vector that are below the mean of the vector are set to 0, while those above are set to 1. Then $\\xi_{\\mathrm{h2m}}$ would provide guidance for moment retrieval. In this feedback manner, the input of the MR decoder is updated by $X_{\\mathrm{mr}} + X_{\\mathrm{mr}} * \\xi_{\\mathrm{h2m}}$ .", + "bbox": [ + 496, + 672, + 890, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Task-dependent Joint Loss", + "text_level": 1, + "bbox": [ + 500, + 847, + 741, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A direct way to optimize the joint task is to sum the respective task-specific loss functions. However, it neglects", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18311", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the variations in magnitudes of different task losses, leading to dominance of one task. Existing methods [20, 25, 29, 32, 36] utilize manually-set weights for the weighted sum of task losses, which limit the task learning as tasks evolve at different rates. To address this problem, inspired by the studies using dynamic weights [10, 16, 22, 23, 30] based on task learning stages, we introduce the task-dependent joint loss as a more effective and flexible solution.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MR Loss. For moment retrieval, we define the MR likelihood using a Gaussian distribution as shown in Eq. (3), with its mean determined by the output $f^{\\theta_{\\mathrm{mr}}}(x)$ of the neural network with $\\theta_{\\mathrm{mr}}$ . $x$ is the input of the neural network.", + "bbox": [ + 76, + 215, + 468, + 275 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\left(Y _ {\\mathrm {m r}} \\mid f ^ {\\theta_ {\\mathrm {m r}}} (x)\\right) = \\mathcal {N} \\left(f ^ {\\theta_ {\\mathrm {m r}}} (x), \\delta_ {\\mathrm {m r}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 297, + 468, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta_{\\mathrm{mr}}$ is a learnable parameter that quantifies the uncertainty of moment retrieval. The negative log-likelihood can be derived as follows:", + "bbox": [ + 76, + 333, + 468, + 377 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} - \\log p \\left(Y _ {\\mathrm {m r}} \\mid f ^ {\\theta_ {\\mathrm {m r}}} (x)\\right) \\\\ \\propto \\frac {1}{2 \\delta_ {\\mathrm {m r}} {} ^ {2}} \\left\\| Y _ {\\mathrm {m r}} - f ^ {\\theta_ {\\mathrm {m r}}} (x) \\right\\| ^ {2} + \\log \\delta_ {\\mathrm {m r}}. \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 392, + 468, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\left\\| Y_{\\mathrm{mr}} - f^{\\theta_{\\mathrm{mr}}}(x)\\right\\|^2$ measures the offset between the model's predicted value and the ground-truth value. Therefore, inspired by Eq. (4), the MR loss function $\\mathcal{L}_{\\mathrm{mr}}(\\theta_{\\mathrm{mr}},\\delta_{\\mathrm{mr}})$ can be defined as:", + "bbox": [ + 76, + 467, + 468, + 531 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {m r}} \\left(\\theta_ {\\mathrm {m r}}, \\delta_ {\\mathrm {m r}}\\right) = \\frac {1}{2 \\delta_ {\\mathrm {m r}} ^ {2}} \\mathcal {L} \\left(\\theta_ {\\mathrm {m r}}\\right) + \\log \\delta_ {\\mathrm {m r}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 547, + 468, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following existing approaches [25, 32, 36], $\\mathcal{L}(\\theta_{\\mathrm{mr}})$ consists of three components: the L1 loss $L_{\\mathrm{L1}}$ , the generalized IoU loss [40] $L_{\\mathrm{gIoU}}$ , and the cross-entropy loss $L_{\\mathrm{BCE}}$ . $L_{\\mathrm{L1}}$ and $L_{\\mathrm{gIoU}}$ are employed to calculate the mean absolute error and gIoU deviation between ground-truth moments and predicted moments, respectively. $L_{\\mathrm{BCE}}$ is used to classify whether the predicted moments belong to the foreground or background. In summary, $\\mathcal{L}(\\theta_{\\mathrm{mr}}) = L_{\\mathrm{L1}} + L_{\\mathrm{gIoU}} + L_{\\mathrm{BCE}}$ .", + "bbox": [ + 76, + 599, + 468, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "HD Loss. For highlight detection, inspired by the Boltzmann distribution, the HD likelihood is modeled using a softmax function applied to the scaled model output similar to [23], as Eq. (6). The learnable scaling factor is denoted as $\\delta_{\\mathrm{hd}}$ . $f^{\\theta_{\\mathrm{hd}}}(x)$ is the output of the network with $\\theta_{\\mathrm{hd}}$ on input $x$ .", + "bbox": [ + 76, + 724, + 468, + 814 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\left(Y _ {\\mathrm {h d}} \\mid f ^ {\\theta_ {\\mathrm {h d}}} (x), \\delta_ {\\mathrm {h d}}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {1}{\\delta_ {\\mathrm {h d}} 2} f ^ {\\theta_ {\\mathrm {h d}}} (x)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 832, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to Eq. (4), the negative log-likelihood of Eq. (6)", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "can be calculated as follows:", + "bbox": [ + 500, + 90, + 691, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} - \\log p \\left(Y _ {\\mathrm {h d}} = y \\mid f ^ {\\theta_ {\\mathrm {h d}}} (x), \\delta_ {\\mathrm {h d}}\\right) \\\\ = \\frac {1}{\\delta_ {\\mathrm {h d}} ^ {2}} \\log \\frac {\\sum_ {y ^ {\\prime}} \\exp \\left(f _ {y ^ {\\prime}} ^ {\\theta_ {\\mathrm {h d}}} (x)\\right)}{\\exp \\left(f _ {y} ^ {\\theta_ {\\mathrm {h d}}} (x)\\right)} \\\\ + \\log \\frac {\\sum_ {y ^ {\\prime}} \\exp \\left(\\frac {1}{\\delta_ {\\mathrm {h d}} {} ^ {2}} f _ {y ^ {\\prime}} ^ {\\theta_ {\\mathrm {h d}}} (x)\\right)}{\\left(\\sum_ {y ^ {\\prime}} \\exp \\left(f _ {y ^ {\\prime}} ^ {\\theta_ {\\mathrm {h d}}} (x)\\right)\\right) ^ {\\frac {1}{\\delta_ {\\mathrm {h d}} {} ^ {2}}}} \\\\ \\approx - \\frac {1}{\\delta_ {\\mathrm {h d}} ^ {2}} \\log \\operatorname {s o f t m a x} \\left(Y _ {\\mathrm {h d}}, f ^ {\\theta_ {\\mathrm {h d}}} (x)\\right) + \\log \\delta_ {\\mathrm {h d}}, \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 114, + 890, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{y}^{\\theta_{\\mathrm{hd}}}(x)$ refers to the y-th value of the $f^{\\theta_{\\mathrm{hd}}}(x)$ . We introduce a simplify assumption $\\sum_{y'} \\exp \\left( \\frac{1}{\\delta_{\\mathrm{hd}}^2} f_{y'}^{\\theta_{\\mathrm{hd}}}(x) \\right) \\approx \\left( \\sum_{y'} \\exp \\left( f_{y'}^{\\theta_{\\mathrm{hd}}}(x) \\right) \\right)^{\\frac{1}{\\delta_{\\mathrm{hd}}^2}}$ when $\\delta_{\\mathrm{hd}} \\to 1$ .", + "bbox": [ + 498, + 291, + 890, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- log softmax $\\left(Y_{\\mathrm{hd}},f^{\\theta_{\\mathrm{hd}}}(x)\\right)$ of Eq. (7) represents the cross-entropy classification loss of $Y_{\\mathrm{hd}}$ . Inspired on this term, we generalize this expression to other classification losses. The HD loss function $\\mathcal{L}_{\\mathrm{hd}}(\\theta_{\\mathrm{hd}},\\delta_{\\mathrm{hd}})$ can be formulated as:", + "bbox": [ + 498, + 364, + 890, + 438 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {h d}} \\left(\\theta_ {\\mathrm {h d}}, \\delta_ {\\mathrm {h d}}\\right) = \\frac {1}{\\delta_ {\\mathrm {h d}} ^ {2}} \\mathcal {L} \\left(\\theta_ {\\mathrm {h d}}\\right) + \\log \\delta_ {\\mathrm {h d}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 449, + 890, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consistent with prior approaches [25, 36], $\\mathcal{L}(\\theta_{\\mathrm{hd}})$ includes the hinge loss $L_{\\mathrm{hinge}}$ , negative video-query pairs loss $L_{\\mathrm{neg}}$ and rank-aware contrastive loss [17] $L_{\\mathrm{cont}}$ , i.e. $\\mathcal{L}(\\theta_{\\mathrm{hd}}) = L_{\\mathrm{hinge}} + L_{\\mathrm{neg}} + L_{\\mathrm{cont}}$ . $L_{\\mathrm{hinge}}$ is computed between two pairs of positive and negative clips, with its margin 0.2 to maintain consistency with [25] for fairness. $L_{\\mathrm{neg}}$ and $L_{\\mathrm{cont}}$ from [36] are used to reduce the saliency of negative pairs.", + "bbox": [ + 498, + 491, + 890, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can obtain final task-dependent joint loss $\\mathcal{L}_{\\mathrm{joint}}$ through integrating Eq. (5) and Eq. (8):", + "bbox": [ + 498, + 598, + 888, + 628 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {j o i n t}} = \\mathcal {L} _ {\\mathrm {m r}} \\left(\\theta_ {\\mathrm {m r}}, \\delta_ {\\mathrm {m r}}\\right) + \\mathcal {L} _ {\\mathrm {h d}} \\left(\\theta_ {\\mathrm {h d}}, \\delta_ {\\mathrm {h d}}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 641, + 890, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 670, + 630, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 696, + 689, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Extensive experiments are conducted on three benchmark datasets: QVHighlights [25], TVSum [43], and Charades-STA [14]. QVHighlights is currently the sole publicly dataset for joint moment retrieval and highlight detection tasks. It provides 10,310 queries associated with 18,367 moments, with an average of 1.8 disjoint moments per query. In contrast to other MR datasets with one-to-one query-moment mappings, QVHighlights aligns more closely with real-world scenarios. Each video in the dataset comprises 75 clips, each of which is 2s-long.", + "bbox": [ + 496, + 719, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also utilize two task-specific datasets, Charades-STA [14] for MR and TVSum [43] for HD, to evaluate the model.", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18312", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b8a57e58b4746c7686e7e944c28f2aa3970d1e7aa90aa362c3467e5f39154590.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVT VU GA MS PK PR FM BK BT DS Avg.
sLSTM [64]ECCV'1641.1 46.2 46.3 47.7 44.8 46.1 45.2 40.6 47.1 45.5 45.1
SG [34]CVPR'1742.3 47.2 47.5 48.9 45.6 47.3 46.4 41.7 48.3 46.6 46.2
LIM-S [51]CVPR'1955.9 42.9 61.2 54.0 60.4 47.5 43.2 66.3 69.1 62.6 56.3
Trailer [47]ECCV'2061.3 54.6 65.7 60.8 59.1 70.1 58.2 64.7 65.6 68.1 62.8
SL-Module [52]ICCV'2186.5 68.7 74.9 86.2 79.0 63.2 58.9 72.6 78.9 64.0 73.3
MINI-Net† [18]ECCV'2080.6 68.3 78.2 81.8 78.1 65.8 57.8 75.0 80.2 65.5 73.2
TCG† [57]ICCV'2185.0 71.4 81.9 78.6 80.2 75.5 71.6 77.3 78.6 68.1 76.8
Joint-VA† [3]ICCV'2183.7 57.3 78.5 86.1 80.1 69.2 70.0 73.0 97.4 67.5 76.3
UMT† [32]CVPR'2287.5 81.5 88.2 78.8 81.4 87.0 76.0 86.9 84.4 79.6 83.1
QD-DETR [36]CVPR'2388.2 87.4 85.6 85.0 85.8 86.9 76.4 91.3 89.2 73.7 85.0
UniVTG‡ [29]ICCV'2383.9 85.1 89.0 80.1 84.6 81.4 70.9 91.7 73.5 69.3 81.0
TaskWeave(Ours)88.2 90.8 93.3 87.5 87.0 82.0 80.9 92.9 89.5 81.2 87.3
", + "bbox": [ + 81, + 88, + 545, + 337 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f2e93bd6a286065241b2de12cf31ea1a53b339346ac12a74c7b6dfc90f5efe4d.jpg", + "table_caption": [ + "Table 1. Experimental results $(\\%)$ on TVSum. $\\dagger$ means including audio modality. $\\ddagger$ means following the pretrain-finetune paradigm." + ], + "table_footnote": [], + "table_body": "
MethodMRHD
R1mAP≥ Very Good
@0.5@0.7@0.5@0.75Avg.mAP HIT@1
Moment-DETR [25]NIPS'2153.9434.84--32.2035.6555.55
UMT†[32]CVPR'2260.2644.26--38.5939.8564.19
QD-DETR [36]CVPR'2362.6846.6662.2341.8241.2239.1363.03
EaTR [20]ICCV'2361.3645.7961.8641.9141.7437.1558.65
UniVTG‡[29]ICCV'2359.74---36.1338.8361.81
TaskWeave(Ours)64.2650.0665.3946.4745.3839.2863.68
", + "bbox": [ + 81, + 405, + 535, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Charades-STA provides 16,128 query-moment pairs. TVSum comprises videos from 10 domains, with each domain containing 5 videos.", + "bbox": [ + 75, + 643, + 468, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We utilize the same evaluation metrics used in prior approaches [25, 32, 36]. For QVHighlights, mean average precision (mAP) with different tIoU thresholds 0.5, 0.75, the average mAP over [0.5:0.05:0.95], and Recall@1 with tIoU 0.5, 0.7 are utilized for MR evaluation. mAP and HIT@1 are used for HD evaluation, where a clip is considered as a true positive when it achieves a \"Very Good\" [25] saliency score. For Charades-STA, we utilize Recall@1 with tIoU thresholds 0.5, 0.7. For TVSum, Top-5 mAP is adopted.", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cef49f43e5d2dbbb82611e99ae6658513341a060a7fe729dfa87280e95ffdffb.jpg", + "table_caption": [ + "Table 3. Experimental results (\\%) on QVHighlights val split. † means including audio modality. ‡ means following the pretrain-finetune paradigm." + ], + "table_footnote": [], + "table_body": "
BackboneMethodR1@0.5R1@0.7
VGGSAP [9]AAAI'1927.4213.36
SM-RL [48]CVPR'1924.3611.17
2D-TAN [65]AAAI'2040.9422.85
FVMR [13]CVPR'2124.3611.17
UMT† [32]CVPR'2248.3129.25
QD-DETR [36]CVPR'2352.7731.13
TaskWeave(Ours)56.5133.66
I3DCTRL [14]ICCV'1723.638.89
MAN [60]CVPR'1946.5322.72
VSLNet [61]ACL'2047.3130.19
QD-DETR [36]CVPR'2350.6731.02
TaskWeave(Ours)53.3631.4
", + "bbox": [ + 573, + 88, + 883, + 338 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b68bc165132fe2d750b6b911fb3695b189f1044f3cb8ef8def412e057b3af69f.jpg", + "table_caption": [ + "Table 2. Experimental results $(\\%)$ on Charades-STA test split. $\\dagger$ means including audio modality. $\\ddagger$ means following the pretrain-finetune paradigm." + ], + "table_footnote": [], + "table_body": "
Task-decoupled UnitInter-task feedbackJoint lossMRHD
R1 @0.7Avg. mAPmAP HIT@1
46.2641.038.9462.84
✓47.8743.2438.5861.81
✓✓49.2945.1238.9662.0
✓✓✓50.0645.3839.2863.68
", + "bbox": [ + 568, + 441, + 887, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. The ablation results $(\\%)$ of the components of our proposed method.", + "bbox": [ + 560, + 588, + 890, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 643, + 718, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Feature representations. For OVHighlights, we employ the pre-trained SlowFast [12] and CLIP [39] backbones to extract video features, following [25, 32, 36] for fairness. For Charedes-STA, we leverage the official release of VGG [42] and I3D [6] features as video embeddings. For TVSum, we extract video features by the I3D [6] pre-trained on Kinetics 400 [21]. For QVHighlights and TVSum, we use CLIP [39] to extract text features, while using GloVe [38] text embeddings for Charades-STA.", + "bbox": [ + 496, + 670, + 890, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training settings. We leverage AdamW [24] optimizer with 1e-4 learning rate and 1e-4 weight decay. We train 200, 100, and 2000 epochs with batch size 32, 32 and 2 for QVHighlights, Charades-STA and TVSum, respectively. All Transformer layers follow the consistent configuration, including sinusoidal positional encodings, 8 at", + "bbox": [ + 496, + 809, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18313", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/25aa03576deb1abb81c7806be9f7edbba3057faaeff947754a473dae3a433ad5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IndexMR-expertHR expertMRHDR
R1 @0.7Avg. mAPmAPHIT@1
(a)Iden.Iden.44.7740.2338.3860.19
(b)Line.Line.47.5543.6538.3260.77
(c)CNNCNN49.2343.7338.8260.77
(d)Trans.Trans.48.3242.8538.6161.10
(e)Trans.Iden.47.142.138.761.68
(f)CNNIden.49.2945.1238.9662.0
", + "bbox": [ + 83, + 88, + 364, + 258 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b7fd42d00f55d9a8d6ada1fe5406e91070f210c0a77c3ffd152a667119cce8d4.jpg", + "table_caption": [ + "Table 5. Flexibility validation $(\\%)$ of the taskdecoupled unit. \"Iden:\" identity mapping; \"Trans\":Transformer; \"Line\":linear layer." + ], + "table_footnote": [], + "table_body": "
TypeMRHD
R1 @0.7Avg. mAPmAP HIT@1
Sum47.6844.2738.760.32
Weighted Sum48.3243.2138.3161.35
Ours50.0645.3839.2863.68
", + "bbox": [ + 383, + 132, + 637, + 258 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c6617c555f9c4b4369360c0b39fc489e1183b62960d4e8883076fe43efbaaa3c.jpg", + "table_caption": [ + "Table 6. Effectiveness justification $(\\%)$ of the task-dependent joint loss." + ], + "table_footnote": [], + "table_body": "
TypeMRHD
R1 @0.7Avg. mAPmAPHIT@1
MR2HD50.0645.3838.7362.84
HD2MR48.6544.7839.2863.68
Bi-MRHD48.8445.138.3661.81
MR-HD50.045.338.9662.45
HD-MR49.144.8739.0359.94
", + "bbox": [ + 648, + 103, + 888, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Comparing various inter-task feedback combinations $(\\%)$ on QVHighlights val split.", + "bbox": [ + 642, + 267, + 890, + 310 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tention heads, and a dropout rate of 0.1. Multi-modal fusion employs a 2-layer Transformer with cross attention, where video features serve as the query and text features serve as key and value. For Eq. (9), we define $\\gamma_{\\mathrm{mr}}, \\gamma_{\\mathrm{hd}}$ as $\\log \\delta_{\\mathrm{mr}}^2, \\log \\delta_{\\mathrm{hd}}^2$ , respectively. Therefore, Eq. (9) can be rewritten as $\\mathcal{L}_{joint} = \\exp(-\\gamma_{\\mathrm{mr}})\\mathcal{L}(\\theta_{\\mathrm{mr}}) + 2\\exp(-\\gamma_{\\mathrm{hd}})\\mathcal{L}(\\theta_{\\mathrm{hd}}) + \\gamma_{\\mathrm{mr}} + \\gamma_{\\mathrm{hd}}$ . $\\gamma_{\\mathrm{mr}}$ and $\\gamma_{\\mathrm{hd}}$ are learnable parameters. They are initialized to 0. To further stabilize the training, the model EMA strategy has also undertaken. All experiments are conducted with Pytorch v1.13.1 on a single NVIDIA RTX 3090.", + "bbox": [ + 75, + 335, + 472, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-arts", + "text_level": 1, + "bbox": [ + 76, + 513, + 377, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on TVSum. In Tab. 1, we compare our TaskWeave with existing state-of-the-arts (SOTA) methods, where the methods with $\\dagger$ are incorporate with audio features. UniVTG [29] follows pretrain-then-finetune paradigm, therefore we use its non-pretrained results for a fair comparison. We observe that: i) our methods outperforms SOTA methods in 9 out of the 10 categories; ii) TaskWeave significantly outperforms all methods in terms of Avg. mAP, with a remarkable $2.71\\%$ improvement over the previous SOTA method.", + "bbox": [ + 75, + 537, + 468, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Charades-STA. In Tab. 2, we utilize different backbones to demonstrate the effectiveness of our TaskWeave. We compare our results with other methods for MR only (white background) and methods for joint MR and HD (gray background). It results in significant improvements, i.e., $+7.09\\%$ in R1@0.5 and $+8.13\\%$ in R1@0.7.", + "bbox": [ + 75, + 688, + 468, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on QVHighlights. As shown in Tab. 3, we present the performance comparisons with existing methods. We report the MR results with the MR-guided feedback, while HD results with the HD-guided feedback. We observe that: i) our TaskWeave surpasses SOTA methods by a large margin in all moment retrieval metrics, with a remarkable $8.72\\%$ improvement in Avg. mAP; ii) our method achieves the second-best performance on the HD. However,", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It's worth noting that UMT [32] incorporates audio features, while our method does not. When compared to other methods without audio features, our method outperforms them all. In summary, these results validate the effectiveness of our task-driven framework.", + "bbox": [ + 496, + 335, + 893, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We provide visualization examples in Fig. 4 for the qualitative analysis among our TaskWeave, Moment-DETR [25], and QD-DETR [36]. Given diverse queries for a video, TaskWeave precisely localizes moments for queries and presents high IoU with the ground-truth. Moment-DETR misses some instances, and QD-DETR exhibits lower retrieval accuracy. Moreover, our method obtains higher saliency scores for relevant clips in response to the query.", + "bbox": [ + 496, + 412, + 893, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies and Discussions", + "text_level": 1, + "bbox": [ + 500, + 544, + 790, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present some ablation studies and discussions about TaskWeave, with all experiments are conducted on the QVHighlights val split [25].", + "bbox": [ + 496, + 566, + 890, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of components. We conduct ablation studies on each component of TaskWeave, as shown in Tab. 4. Sequentially employing the task-decoupled unit and our proposed joint loss contributes a $5.5\\%$ improvement in average mAP and a $6.5\\%$ increase in R1@0.7 for MR. Utilizing all components, we observe that a significant $10.7\\%$ improvement in average mAP of MR and a $1.3\\%$ enhancement in HIT@1 of HD. These results indicate the effectiveness of the proposed components in out TaskWeave.", + "bbox": [ + 496, + 613, + 893, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Flexibility of the task-decoupled unit. To demonstrate the flexibility of the task-decoupled unit, we investigate the performance of applying different networks within various experts, as shown in Tab. 5. The shared expert utilize a fixed configuration with 2 Transformer layers. We provide 4 different methods for each task-specific expert, including the feed-forward network (one layer Linear), the identity mapping, CNN (composed of one depthwise convolution layer with the kernel/stride/padding of $5/1/2$ and 1D convolution), and the Transformer. Due to space constraints, we present", + "bbox": [ + 496, + 750, + 895, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18314", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a4aa83bb45accdc89c01decbb04416267c07dffc42087a05418c66c8287babe.jpg", + "image_caption": [ + "Query1: A guy with grey top is showing a box filled with rubbish. Query2: A man shaves his beard looks in a mirror and talks to the camera.", + "Figure 4. Qualitative results on the QVHighlights for Ground-Truth, Moment-DETR, QD-DETR and our method. The predicted moments and saliency scores are illustrated through intervals and lines." + ], + "image_footnote": [], + "bbox": [ + 76, + 107, + 893, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "results of 6 out of 16 combinations. The results in Tab. 5 are obtained by TaskWeave without inter-task feedback.", + "bbox": [ + 75, + 378, + 468, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From this table, we can observe several interesting facts. First, different task-specific experts have different performance. The need for experts is reflected in (a). Second, although Transformer has made significant progress in visual, it's not a panacea. Results in (d) are lower compared to (b) and (c), we believe that the high computational complexity of Transformer decreases the performance. Third, the task-specific expert should be designed based on the task objective. For instance, (c) and (f) perform well on the MR because they focus on local features, which is important for localizing moments. Finally, we also find evidence that MR and HD are highly related. Improvement in one task enhances the other (compare (c) and (f)), while a decline in one limits the other's performance (compare (d) and (e)). In this paper, our mr-specific expert utilizes CNN and hd-specific expert is implemented by identity mapping.", + "bbox": [ + 75, + 411, + 468, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Task-dependent joint loss. Tab. 6 shows the performance of TaskWeave with different losses. \"Sum\" and \"Weighted Sum\" refer to the fusion manner of loss for different tasks, respectively. Weights in \"Weighted Sum\" are consistent with the existing methods [25, 36] for fairness. Comparing \"Sum\" and \"Weighted Sum\", we find that \"Weighted Sum\" can better balance the performance of the two tasks. However, it is obvious that the proposed task-dependent joint loss is optimal.", + "bbox": [ + 75, + 655, + 468, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inter-task interactions. We believe that each task requires learning before providing effective feedback. Therefore, in our inter-task feedback mechanism, the feedback process starts when the model is trained to half of the max epoch. For brevity, we write Moment/Highlightness-guided feedback as MgF/HgF. In Tab. 7, we explore five feedback manners as follows: \"MR2HD\" (with MgF only),", + "bbox": [ + 75, + 795, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "\"HD2MR\" (with HgF only), \"Bi-MRHD\" (with MgF and HgF simultaneously), \"MR-HD\" (MgF first, then HgF), and \"HD-MR\" (HgF first, then MgF).", + "bbox": [ + 496, + 377, + 890, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We find that the inter-task feedback makes both tasks gain simultaneously. The performance of the model with \"HD2MR\" is slightly worse, this is because HD focuses on more refined moments than MR. Ground-truth annotations and the results of \"MR-HD\" and \"HD-MR\" also illustrate this fact. The gain brought by \"Bi-MRHD\" is limited, which we believe that it's a natural result of feedback not always being effective. In general, the inter-task feedback not only contributes to bring gains for both tasks but also helps to understand the characteristics of each task.", + "bbox": [ + 496, + 422, + 892, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 585, + 619, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes a novel task-driven paradigm for addressing joint moment retrieval and highlight detection. Different from existing data-driven methods, we utilize the task-decoupled unit to capture the task-specific and common features, respectively. We also explore different network architecture for moment retrieval and highlight detection. We design the inter-task feedback mechanism to in-depth investigate the interplay between both tasks. Different from prior methods, we introduce the principled joint loss to optimize the model. The effectiveness, flexibility, and superiority of the proposed method have been demonstrated on three benchmark datasets.", + "bbox": [ + 496, + 612, + 890, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 806, + 660, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was supported by the National Natural Science Foundation of China (No. U23B2060, 62088102), the Youth Innovation Team of Shaanxi Universities, and the Fundamental Research Funds for the Central Universities.", + "bbox": [ + 496, + 830, + 890, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18315", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2", + "[2] Niki Parmar Jakob Uszkoreit Llion Jones Aidan N. Gomez Lukasz Kaiser Ashish Vaswani, Noam Shazeer and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5999-6009, 2017. 2, 3", + "[3] Taiwanbat Badamdorj, Mrigank Rochan, Yang Wang, and Li Cheng. Joint visual and audio learning for video highlight detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8127-8137, 2021. 6", + "[4] Taiwanbat Badamdorj, Mrigank Rochan, Yang Wang, and Li Cheng. Contrastive learning for unsupervised video highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14042-14052, 2022. 1, 2", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 2", + "[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6", + "[7] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 2", + "[8] Jingyuan Chen, Lin Ma, Xinpeng Chen, Zequn Jie, and Jiebo Luo. Localizing natural language in videos. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8175-8182, 2019. 2", + "[9] Shaoxiang Chen and Yu-Gang Jiang. Semantic proposal for activity localization in videos via sentence query. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8199-8206, 2019. 6", + "[10] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018. 5", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[12] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 6", + "[13] Junyu Gao and Changsheng Xu. Fast video moment retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1523-1532, 2021. 6" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 2, 5, 6", + "[15] Ana Garcia del Molino and Michael Gygli. PhD-gifs: personalized highlight detection for automatic gif creation. In Proceedings of the 26th ACM international conference on Multimedia, pages 600-608, 2018. 2", + "[16] Michelle Guo, Albert Haque, De-An Huang, Serena Yeung, and Li Fei-Fei. Dynamic task prioritization for multitask learning. In Proceedings of the European conference on computer vision (ECCV), pages 270–287, 2018. 5", + "[17] David T Hoffmann, Nadine Behrmann, Juergen Gall, Thomas Brox, and Mehdi Noroozi. Ranking info noise contrastive estimation: Boosting contrastive learning via ranked positives. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 897–905, 2022. 5", + "[18] Fa-Ting Hong, Xuanteng Huang, Wei-Hong Li, and Wei-Shi Zheng. Mini-net: Multiple instance ranking network for video highlight detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 345-360. Springer, 2020. 6", + "[19] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 2", + "[20] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13846-13856, 2023. 1, 2, 5, 6", + "[21] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6", + "[22] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 5", + "[23] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482-7491, 2018. 5", + "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[25] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34: 11846-11858, 2021. 1, 2, 4, 5, 6, 7, 8", + "[26] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 2, 4", + "[27] Huan Li, Ping Wei, Jiapeng Li, Zeyu Ma, Jiahui Shang, and Nanning Zheng. Asymmetric relation consistency reasoning" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18316", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for video relation grounding. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 2", + "[28] Huan Li, Ping Wei, Zeyu Ma, and Nanning Zheng. Inverse compositional learning for weakly-supervised relation grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15477-15487, 2023. 2", + "[29] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified videolanguage temporal grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2794-2804, 2023. 1, 2, 5, 6, 7", + "[30] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1871-1880, 2019. 5", + "[31] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2, 4", + "[32] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3042-3051, 2022. 1, 2, 4, 5, 6, 7", + "[33] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2", + "[34] Behrooz Mahasseni, Michael Lam, and Sinisa Todorovic. Unsupervised video summarization with adversarial LSTM networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 202-211, 2017. 6", + "[35] Otniel-Bogdan Mercea, Lukas Riesch, A Koepke, and Zeynep Akata. Audio-visual generalised zero-shot learning with cross-modal attention and language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10553–10563, 2022. 2", + "[36] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23023-23033, 2023. 1, 2, 4, 5, 6, 7, 8", + "[37] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810-10819, 2020. 2", + "[38] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6", + "[40] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5", + "[41] Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, and Dacheng Tao. Tridet: Temporal action detection with relative boundary modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18857-18866, 2023. 3", + "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6", + "[43] Yale Song, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaimes. Tvsum: Summarizing web videos using titles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5179-5187, 2015. 2, 5", + "[44] Min Sun, Ali Farhadi, and Steve Seitz. Ranking domain-specific highlights by analyzing edited videos. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13, pages 787-802. Springer, 2014. 1, 2", + "[45] Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. Progressive layered extraction (ple): A novel multi-task learning (mtl) model for personalized recommendations. In Proceedings of the 14th ACM Conference on Recommender Systems, pages 269-278, 2020. 2", + "[46] Tuan N Tang, Kwonyoung Kim, and Kwanghoon Sohn. Temporalmaxer: Maximize temporal context with only max pooling for temporal action localization. arXiv preprint arXiv:2303.09055, 2023. 3", + "[47] Lezi Wang, Dong Liu, Rohit Puri, and Dimitris N Metaxas. Learning trailer moments in full-length movies with co-contrastive attention. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVIII 16, pages 300–316. Springer, 2020. 6", + "[48] Weining Wang, Yan Huang, and Liang Wang. Language-driven temporal activity localization: A semantic matching reinforcement learning model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 334-343, 2019. 6", + "[49] Fanyue Wei, Biao Wang, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Learning pixel-level distinctions for video highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3073-3082, 2022. 2", + "[50] Xi Wei, Tianzhu Zhang, Yan Li, Yongdong Zhang, and Feng Wu. Multi-modality cross attention network for image and sentence matching. In Proceedings of the IEEE/CVF con" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18317", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on computer vision and pattern recognition, pages 10941-10950, 2020. 2", + "[51] Bo Xiong, Yannis Kalantidis, Deepi Ghadiyaram, and Kristen Grauman. Less is more: Learning highlight detection from video duration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1258-1267, 2019. 6", + "[52] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7970-7979, 2021. 6", + "[53] Yifang Xu, Yunzhuo Sun, Yang Li, Yilei Shi, Xiaoxiang Zhu, and Sidan Du. Mh-detr: Video moment and highlight detection with cross-modal transformer. arXiv preprint arXiv:2305.00355, 2023. 1", + "[54] Jin Yang, Ping Wei, Ziyang Ren, and Nanning Zheng. Gated multi-scale transformer for temporal action localization. IEEE Transactions on Multimedia, 2023. 2", + "[55] Jin Yang, Ping Wei, and Nanning Zheng. Cross time-frequency transformer for temporal action localization. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2", + "[56] Ting Yao, Tao Mei, and Yong Rui. Highlight detection with pairwise deep ranking for first-person video summarization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 982–990, 2016. 1, 2", + "[57] Qinghao Ye, Xiyue Shen, Yuan Gao, Zirui Wang, Qi Bi, Ping Li, and Guang Yang. Temporal cue guided video highlight detection with low-rank audio-visual fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7950–7959, 2021. 6", + "[58] Yitian Yuan, Tao Mei, and Wenwu Zhu. To find where you talk: Temporal sentence localization in video with attention based location regression. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9159–9166, 2019. 2", + "[59] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 2", + "[60] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1247-1257, 2019. 1, 2, 6", + "[61] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6543-6554, 2020. 2, 6", + "[62] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Temporal sentence grounding in videos: A survey and future directions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10443-10465, 2023. 1" + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Temporal sentence grounding in videos: a survey and future directions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1", + "[64] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. Video summarization with long short-term memory. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 766-782. Springer, 2016. 6", + "[65] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12870-12877, 2020. 1, 2, 6" + ], + "bbox": [ + 501, + 92, + 890, + 287 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18318", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/a6a833ff-f398-48fe-909b-70ff81bf1b94_model.json b/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/a6a833ff-f398-48fe-909b-70ff81bf1b94_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8e68e08aa9cb207fdba09f9d4bbfcb4c5a4e72b3 --- /dev/null +++ b/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/a6a833ff-f398-48fe-909b-70ff81bf1b94_model.json @@ -0,0 +1,2422 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.121, + 0.13, + 0.852, + 0.177 + ], + "angle": 0, + "content": "Task-Driven Exploration: Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.204, + 0.792, + 0.275 + ], + "angle": 0, + "content": "Jin Yang, Ping Wei*, Huan Li, Ziyang Ren \nNational Key Laboratory of Human-Machine Hybrid Augmented Intelligence Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University \n{jin.yang, lh875056558, rzyrzy}@stu.xjtu.edu.cn, pingwei@xjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.474, + 0.659 + ], + "angle": 0, + "content": "Video moment retrieval and highlight detection are two highly valuable tasks in video understanding, but until recently they have been jointly studied. Although existing studies have made impressive advancement recently, they predominantly follow the data-driven bottom-up paradigm. Such paradigm overlooks task-specific and inter-task effects, resulting in poor model performance. In this paper, we propose a novel task-driven top-down framework TaskWeave for joint moment retrieval and highlight detection. The framework introduces a task-decoupled unit to capture task-specific and common representations. To investigate the interplay between the two tasks, we propose an inter-task feedback mechanism, which transforms the results of one task as guiding masks to assist the other task. Different from existing methods, we present a task-dependent joint loss function to optimize the model. Comprehensive experiments and in-depth ablation studies on QVHighlights, TVSum, and CharadesSTA datasets corroborate the effectiveness and flexibility of the proposed framework. Codes are available at github.com/EdenGabriel/TaskWeave." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.687, + 0.21, + 0.703 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.47, + 0.848 + ], + "angle": 0, + "content": "As videos are prevailing in a wide range of applications, the diversity and massive scales of video content have posed unprecedented challenges in finding relevant moments based on user queries. To this end, the moment retrieval (MR) [60, 63, 65] and highlight detection (HD) [4, 56] tasks have emerged recently. MR aims to retrieve video moments that are relevant to the given query [14, 62]. HD aims to predict the clip-level saliency scores in the video [44]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.848, + 0.47, + 0.879 + ], + "angle": 0, + "content": "As MR and HD tasks are closely related, they have been jointly addressed and achieved breakthroughs recently" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.311, + 0.895, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.474, + 0.895, + 0.517 + ], + "angle": 0, + "content": "Figure 1. Although the positive temporal intervals of moment retrieval and highlight detection exhibit high overlap, they pursue different objectives." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.895, + 0.746 + ], + "angle": 0, + "content": "[20, 25, 29, 32, 36, 53]. The existing joint approaches in general utilize a shared backbone to learn the multimodal features as the common representations for MR and HD. Then a MR prediction head is employed for moments localization and a HD prediction head predicts saliency scores. These methods adhere to the bottom-up, data-driven paradigm, i.e. they capture common features from the input data and then utilize the features for different tasks. The effectiveness of these methods is built upon the premise of the high correlation between MR and HD. As illustrated in Fig. 1, the task MR and HD share the identical query and video inputs, additionally exhibiting a substantial temporal overlap between their respective positive temporal intervals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.896, + 0.868 + ], + "angle": 0, + "content": "However, the bottom-up, data-driven paradigm tends to excessively rely on the common features, but overlooks the inherent specific characteristics of MR and HD. This tendency might simplify the joint modeling as a problem of feature fusion, without considering the interplay between the two tasks. The distinct objectives pursued by MR and HD rely on distinct task-specific characteristics. Unfortunately, existing methods overlook the specificity." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To address the aforementioned issues, we believe it is essential to leverage the fundamental multi-modal data to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.232, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18308" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.182 + ], + "angle": 0, + "content": "mine commonalities across tasks (bottom-up), while also strengthen the awareness of task-specific characteristics (top-down). To this end, we propose a novel paradigm TaskWeave from a task-driven perspective. The key idea is to jointly address the tasks MR and HD by considering the commonality, specificity, and interplay of MR and HD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.468, + 0.346 + ], + "angle": 0, + "content": "To effectively capture the commonality and specificity, we design a task-decoupled unit, i.e. a shared expert to capture common features and two task-specific experts to acquire the distinct characteristics. In order to investigate the interplay between MR and HD in-depth, we design an inter-task feedback mechanism. It converts the predictions of MR/HD into mask information, which are fed back to the input of the HD/MR prediction head. Furthermore, we introduce a principled task-dependent joint loss in which the task-specific weights are dynamically adjusted, rather than manually tuned." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.468, + 0.437 + ], + "angle": 0, + "content": "We conduct experiments on the QVHighlights [25] dataset to validate the effectiveness of the proposed method. Moreover, we also conduct experiments for two individual tasks on the datasets Charades-STA [14] (moment retrieval) and TVSUM [43] (highlight detection). The proposed approach outperforms the existing methods." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.439, + 0.431, + 0.454 + ], + "angle": 0, + "content": "The key contributions of this paper are three folds." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.455, + 0.468, + 0.483 + ], + "angle": 0, + "content": "1. It proposes a novel task-driven, top-down framework for joint moment retrieval and highlight detection." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.485, + 0.468, + 0.514 + ], + "angle": 0, + "content": "2. It introduces a task-decoupled unit, an inter-task feedback mechanism, and a principled task-dependent loss." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.515, + 0.468, + 0.544 + ], + "angle": 0, + "content": "3. It achieves state-of-the-art performance on three datasets. The ablation study validates the methods." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.455, + 0.468, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.559, + 0.218, + 0.573 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.585, + 0.402, + 0.599 + ], + "angle": 0, + "content": "We review the related work from four aspects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.468, + 0.734 + ], + "angle": 0, + "content": "Moment Retrieval. Existing approaches for MR mainly include two groups. One group follows a two-stage procedure [1, 14, 60, 65], which involves generating candidate temporal intervals and ranking them based on the correlation with the query. The other group directly regresses the temporal interval based on the aligned visual-text features [8, 37, 58, 59, 61]. Moreover, most datasets provide only one moment annotation for each video-query pair [1, 14], which does not align with real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.839 + ], + "angle": 0, + "content": "Highlight Detection. The saliency score in the highlight detection represents the relevance of a video clip to the given query. Most prior highlight detection benchmark datasets are query-agnostic [15, 44]. The saliency scores for video clips remained constant regardless of the query. As a result, some previous approaches treated HD as a solely visual task [4, 44, 49, 56]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "MR and HD tasks are traditionally studied separately. They have been jointly addressed recently as the introduction of QVHighlights dataset [25]. QVHighlights provides multiple moment-annotations for each query and ensures" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "these moments are uniformly distributed throughout the video. It also provides query-dependent highlightness annotations. With QVHighlights, the model Moment-DETR [25] is proposed for joint MR and HD. Following Moment-DETR, a growing number of approaches have been proposed to accomplish the joint task [20, 29, 32, 36]. However, these methods adopt the data-driven and bottom-up paradigm. Different from them, we propose a novel task-driven and top-down paradigm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.424 + ], + "angle": 0, + "content": "Vision Transformers. Transformer-based models [2] have brought huge achievement in both the image and video related domains [5, 11, 27, 28, 54, 55]. One of the most well-known methods is DETR [5], which regards object detection as a set prediction problem. Its end-to-end prediction procedure eliminates the intermediate or post-processing steps. On the other hand, some studies have employed cross-attention mechanism [35, 36, 50] to inject multi-modal data into the Transformer architecture. In this paper, we also adopt a DETR-like architecture [26, 31]. However, different from those methods, we incorporate diverse network architectures for feature extraction. Our focus lies in the collaboration of the network architectures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.892, + 0.712 + ], + "angle": 0, + "content": "Multi-task Learning. Multi-task learning (MTL) aims to train a single model for multiple tasks [7]. The most straightforward approach is to utilize a shared backbone to extract common features, which relies on the data-driven manner. However, it often leads to suboptimal performance for unrelated tasks and lacks flexibility. In response, some studies introduce MOE [19] and MMOE [33], which utilize a set of the shared experts in place of the shared bottom layer. To mitigate the seesaw phenomenon of multi-task learning, PLE [45] employs separate experts for each task while still retaining the shared experts. From the perspective of MTL, existing approaches for joint MR and HD follow the shared bottom paradigm. Moreover, directly applying MTL methods to joint MR and HD does not yield favorable results. Existing MTL methods are complex, therefore direct usage of them leads to a sharp increase in model complexity. This might decrease the performance. To this end, we aim to design an effective task-driven framework to jointly address MR and HD." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.73, + 0.634, + 0.746 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.61, + 0.769 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Given a text query with \\( W \\) words and an untrimmed video composed of \\( N \\) clips, the objective of the joint moment retrieval (MR) and highlight detection (HD) is to localize the center coordinate \\( q_{c} \\) and width \\( q_{w} \\) of temporal intervals that are relevant to the text query, in addition to ranking clip-wise saliency scores." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Architecture Overview. Given the high correlation between moment retrieval and highlight detection tasks, the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18309" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.093, + 0.895, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.313, + 0.895, + 0.371 + ], + "angle": 0, + "content": "Figure 2. The overall pipeline of the proposed task-driven model TaskWeave. We propose the task-decoupled unit to capture task-specific and common features. Various experts can adopt different network implementations, showcasing the flexibility of the model. Inter-task feedback mechanism is designed to investigate the influence between both tasks. There are two feedback manners: Moment-guided and Highlightness-guided feedback. The principled task-dependent joint loss is introduced for jointly optimize the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.47, + 0.517 + ], + "angle": 0, + "content": "most intuitive approach is using a shared backbone in conjunction with two task-specific prediction heads. It is a data-driven paradigm that is commonly employed by previous methods, for its simplicity and ease of implementation. However, the task-specific characteristics are inherently present, since MR and HD have distinct objectives. Additionally, the interplay between MR and HD should also be considered, which can enhance the model's performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.471, + 0.79 + ], + "angle": 0, + "content": "The overall pipeline of our approach is illustrated in Fig. 2. We employed the frozen video/text-encoder backbones to extract the video/text features, while ensuring their dimensions remained consistent at \\(D\\) by the projection. These features are utilized for multi-modal fusion through methods such as cross-attention or concatenation, resulting in query-related video representations \\(Z \\in \\mathbb{R}^{N \\times D}\\). In our paper, we extract these representations through the cross-attention. After cross-attention, inspired by [41, 46], a 1D Max Pooling (MaxPool1d) with the kernel size 5, stride 1 and padding 2 is utilized to eliminate the rank loss problem in the attention mechanism. These representations are fed into the task-decoupled unit to capture the task-related features \\(X_{\\mathrm{mr}} \\in \\mathbb{R}^{N \\times D}\\), \\(X_{\\mathrm{hd}} \\in \\mathbb{R}^{N \\times D}\\). Then, we employ the task-specific decoders with inter-task feedback mechanism to make predictions for moments localization and clip-wise saliency scores. We introduce the principled task-dependent loss to jointly optimize the model." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.801, + 0.273, + 0.817 + ], + "angle": 0, + "content": "3.2. Task-decoupled Unit" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Since moment retrieval and highlight detection have distinct objectives, the specificity of each task should be considered, rather than solely focusing on their commonalities. For this purpose, we propose a task-decoupled unit from a task-driven perspective to capture the task-related features," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.395, + 0.89, + 0.41 + ], + "angle": 0, + "content": "which involves task-specific features and common features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.532 + ], + "angle": 0, + "content": "The task-decoupled unit is depicted in Fig. 2. Inspired by the attention mechanism [2], the query-related video representation \\( Z \\) is initially fed into two task-specific mappers, mr-mapper and hd-mapper. Each mapper is implemented using one-layer feed-forward network. The output of each task-specific mapper is directed into both the respective task-specific expert network (MR-specific Expert or HD-specific Expert) and the Shared Expert network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.893, + 0.669 + ], + "angle": 0, + "content": "Thanks to the design of our task-decoupled framework, each expert can employ various networks, such as convolutional network, Transformer, and feed-forward network. It offers more flexibility in addressing the joint MR and HD tasks. In our study, we also investigate the influence of different expert networks, referring to Sec. 4 for more details. Consequently, we can capture the task-related features \\( X_{\\mathrm{mr}} \\) and \\( X_{\\mathrm{hd}} \\) via element-wise product between the output of the task-specific expert and that of the shared expert." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.67, + 0.855, + 0.685 + ], + "angle": 0, + "content": "The MR-specific feature \\(X_{\\mathrm{mr}}\\) can be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.699, + 0.892, + 0.716 + ], + "angle": 0, + "content": "\\[\nX _ {\\mathrm {m r}} = \\mathcal {P} _ {\\mathrm {m r}} \\left(\\mathcal {M} _ {\\mathrm {m r}} (Z)\\right) \\otimes \\mathcal {S} \\left(\\mathcal {M} _ {\\mathrm {m r}} (Z)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.729, + 0.892, + 0.805 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}_{\\mathrm{mr}}(\\cdot)\\) refers to the mr-mapper operation. \\(S(\\cdot)\\) is the shared expert calculation. \\(\\otimes\\) denotes the element-wise product. \\(\\mathcal{P}_{\\mathrm{mr}}(\\cdot)\\) is the mr-specific expert calculation. The calculation method for the HD-specific feature \\(X_{\\mathrm{hd}}\\) follows the similar approach." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.783, + 0.832 + ], + "angle": 0, + "content": "3.3. Inter-task Feedback Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "It is imperative to investigate the interplay between the moment retrieval and highlight detection tasks for their significant correlation. However, existing methods overlook this aspect and employ two decoders for direct prediction." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18310" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.094, + 0.468, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.222, + 0.358, + 0.234 + ], + "angle": 0, + "content": "(a) Moment-guided feedback" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.245, + 0.468, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.199, + 0.413, + 0.373, + 0.425 + ], + "angle": 0, + "content": "(b) Highlightness-guided feedback" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.479 + ], + "angle": 0, + "content": "Figure 3. Illustration of the inter-task feedback mechanism. (a) moment-guided feedback manner. (b) highlightness-guided feedback manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.469, + 0.627 + ], + "angle": 0, + "content": "To resolve this issue, we propose an inter-task feedback mechanism, which contains two task-specific decoders (Transformer decoder for MR and lightweight decoder for HD) and two feedback manners (moment-guided feedback and highlightness-guided feedback). The output of one task is transformed into the mask to assist another task in the feedback mechanism. All components of the inter-task feedback mechanism is detailed as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.794 + ], + "angle": 0, + "content": "Transformer Decoder for MR. Previous studies have demonstrated that it is effective to employ dynamic anchors as queries for the decoder within the DETR-like structure [26, 31]. They iteratively update anchor boxes, thereby transforming the procedure for updating queries into the cascaded soft ROI-pooling. In our work, the Transformer decoder for MR follows the work [31]. Each moment query is represented by its temporal center coordinate \\( q_{c} \\) and width \\( q_{w} \\). The output of the Transformer decoder for MR is denoted as \\( Y_{\\mathrm{mr}} \\in \\mathbb{R}^{N_q \\times 2} \\). \\( N_{q} \\) is the number of moment queries and set to 10 in our work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Lightweight Decoder for HD. The most straightforward method for the HD decoder would be utilizing one or more fully-connected layers, as seen in prior studies such as Moment-DETR [25] and UMT [32]. However, such design overlooks the diversity of video-query pairs [36], which provides identical criterias for the saliency score prediction of each video-query pair. Although QD-DETR [36] pro" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "poses a global saliency token to predict the saliency scores, it remains highly coupled with the encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.121, + 0.892, + 0.242 + ], + "angle": 0, + "content": "Different from those methods, we introduce CNN structures in the decoder. CNNs can capture local details, ensuring accurate saliency predictions across varying queries within the same video. The structure of the lightweight decoder for HD prediction is illustrated in Fig. 3, which alternates between Linear and Conv1d layers. We denote the input of the HD decoder as \\(\\mathcal{J} \\in \\mathbb{R}^{N \\times d}\\). The saliency prediction \\(Y_{\\mathrm{hd}} \\in \\mathbb{R}^{N \\times 1}\\) can be computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.248, + 0.892, + 0.291 + ], + "angle": 0, + "content": "\\[\nY _ {\\mathrm {h d}} = D _ {\\mathrm {H D}} (\\mathcal {J}) + \\frac {\\mathbb {I} _ {i = 1} ^ {N} \\left(\\sum_ {k = 1} ^ {d} \\mathcal {J} ^ {i k}\\right)}{\\sqrt {d}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.296, + 0.89, + 0.327 + ], + "angle": 0, + "content": "where \\(\\mathbb{I}_{i = 1}^{N}(\\cdot)\\) represents the traversal operation. \\(D_{\\mathrm{HD}}(\\cdot)\\) refers to the decoder network operation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.892, + 0.417 + ], + "angle": 0, + "content": "Moment-guided Feedback. We propose a moment-guided feedback manner to investigate the influence of MR on HD. Our key focus is on how to effectively utilize the results of moment retrieval for highlight detection. The implementation procedure of moment-guided feedback is illustrated in Fig. 3 (a)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.893, + 0.598 + ], + "angle": 0, + "content": "We transform the tentatively predicted moment results from the \"(center, width)\" format to the \"(start, end)\" format. These moment results are fed into the Moment-to-Mask (M2M) converter to generate moment-aware masks \\(\\xi_{\\mathrm{m2m}}\\). We initialize a clip-wise vector with a length of \\(N\\) for each moment prediction. The indices in these vectors correspond to the clip indexes within the video. We binarize these vectors based on the moment results. We can obtain the clip indexes that are included in the moments through the prior information and set the values at those positions in the vector to 1, while the rest to 0. The obtained \\(N_q\\) binary vectors are then summed and normalized with L2 norm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.674 + ], + "angle": 0, + "content": "The obtained moment-aware masks \\(\\xi_{\\mathrm{m2m}}\\) integrate the prediction results from \\(N_{q}\\) queries and provide guidance for the highlight detection. In the moment-guided feedback manner, the input \\(\\mathcal{I}\\) of the HD decoder is updated by \\(X_{\\mathrm{hd}} + X_{\\mathrm{hd}}*\\xi_{\\mathrm{m2m}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.841 + ], + "angle": 0, + "content": "Highlightness-guided Feedback. Similar to the moment-guided feedback, we introduce a highlightness-guided feedback manner with the Highlightness-to-Mask (H2M) converter, to explore the influence of HD on MR. As illustrated in Fig. 3 (b), the saliency score vector is binarized in the H2M to obtain the highlightness-aware mask \\(\\xi_{\\mathrm{h2m}}\\). Values in the saliency score vector that are below the mean of the vector are set to 0, while those above are set to 1. Then \\(\\xi_{\\mathrm{h2m}}\\) would provide guidance for moment retrieval. In this feedback manner, the input of the MR decoder is updated by \\(X_{\\mathrm{mr}} + X_{\\mathrm{mr}} * \\xi_{\\mathrm{h2m}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.742, + 0.864 + ], + "angle": 0, + "content": "3.4. Task-dependent Joint Loss" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "A direct way to optimize the joint task is to sum the respective task-specific loss functions. However, it neglects" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18311" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.211 + ], + "angle": 0, + "content": "the variations in magnitudes of different task losses, leading to dominance of one task. Existing methods [20, 25, 29, 32, 36] utilize manually-set weights for the weighted sum of task losses, which limit the task learning as tasks evolve at different rates. To address this problem, inspired by the studies using dynamic weights [10, 16, 22, 23, 30] based on task learning stages, we introduce the task-dependent joint loss as a more effective and flexible solution." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.216, + 0.47, + 0.276 + ], + "angle": 0, + "content": "MR Loss. For moment retrieval, we define the MR likelihood using a Gaussian distribution as shown in Eq. (3), with its mean determined by the output \\( f^{\\theta_{\\mathrm{mr}}}(x) \\) of the neural network with \\( \\theta_{\\mathrm{mr}} \\). \\( x \\) is the input of the neural network." + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.299, + 0.469, + 0.318 + ], + "angle": 0, + "content": "\\[\np \\left(Y _ {\\mathrm {m r}} \\mid f ^ {\\theta_ {\\mathrm {m r}}} (x)\\right) = \\mathcal {N} \\left(f ^ {\\theta_ {\\mathrm {m r}}} (x), \\delta_ {\\mathrm {m r}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.334, + 0.469, + 0.378 + ], + "angle": 0, + "content": "where \\(\\delta_{\\mathrm{mr}}\\) is a learnable parameter that quantifies the uncertainty of moment retrieval. The negative log-likelihood can be derived as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.393, + 0.469, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} - \\log p \\left(Y _ {\\mathrm {m r}} \\mid f ^ {\\theta_ {\\mathrm {m r}}} (x)\\right) \\\\ \\propto \\frac {1}{2 \\delta_ {\\mathrm {m r}} {} ^ {2}} \\left\\| Y _ {\\mathrm {m r}} - f ^ {\\theta_ {\\mathrm {m r}}} (x) \\right\\| ^ {2} + \\log \\delta_ {\\mathrm {m r}}. \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.468, + 0.469, + 0.532 + ], + "angle": 0, + "content": "\\(\\left\\| Y_{\\mathrm{mr}} - f^{\\theta_{\\mathrm{mr}}}(x)\\right\\|^2\\) measures the offset between the model's predicted value and the ground-truth value. Therefore, inspired by Eq. (4), the MR loss function \\(\\mathcal{L}_{\\mathrm{mr}}(\\theta_{\\mathrm{mr}},\\delta_{\\mathrm{mr}})\\) can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.548, + 0.469, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {m r}} \\left(\\theta_ {\\mathrm {m r}}, \\delta_ {\\mathrm {m r}}\\right) = \\frac {1}{2 \\delta_ {\\mathrm {m r}} ^ {2}} \\mathcal {L} \\left(\\theta_ {\\mathrm {m r}}\\right) + \\log \\delta_ {\\mathrm {m r}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.601, + 0.469, + 0.722 + ], + "angle": 0, + "content": "Following existing approaches [25, 32, 36], \\(\\mathcal{L}(\\theta_{\\mathrm{mr}})\\) consists of three components: the L1 loss \\(L_{\\mathrm{L1}}\\), the generalized IoU loss [40] \\(L_{\\mathrm{gIoU}}\\), and the cross-entropy loss \\(L_{\\mathrm{BCE}}\\). \\(L_{\\mathrm{L1}}\\) and \\(L_{\\mathrm{gIoU}}\\) are employed to calculate the mean absolute error and gIoU deviation between ground-truth moments and predicted moments, respectively. \\(L_{\\mathrm{BCE}}\\) is used to classify whether the predicted moments belong to the foreground or background. In summary, \\(\\mathcal{L}(\\theta_{\\mathrm{mr}}) = L_{\\mathrm{L1}} + L_{\\mathrm{gIoU}} + L_{\\mathrm{BCE}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.725, + 0.469, + 0.815 + ], + "angle": 0, + "content": "HD Loss. For highlight detection, inspired by the Boltzmann distribution, the HD likelihood is modeled using a softmax function applied to the scaled model output similar to [23], as Eq. (6). The learnable scaling factor is denoted as \\(\\delta_{\\mathrm{hd}}\\). \\(f^{\\theta_{\\mathrm{hd}}}(x)\\) is the output of the network with \\(\\theta_{\\mathrm{hd}}\\) on input \\(x\\)." + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.833, + 0.469, + 0.867 + ], + "angle": 0, + "content": "\\[\np \\left(Y _ {\\mathrm {h d}} \\mid f ^ {\\theta_ {\\mathrm {h d}}} (x), \\delta_ {\\mathrm {h d}}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {1}{\\delta_ {\\mathrm {h d}} 2} f ^ {\\theta_ {\\mathrm {h d}}} (x)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Similar to Eq. (4), the negative log-likelihood of Eq. (6)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.692, + 0.105 + ], + "angle": 0, + "content": "can be calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.116, + 0.891, + 0.292 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} - \\log p \\left(Y _ {\\mathrm {h d}} = y \\mid f ^ {\\theta_ {\\mathrm {h d}}} (x), \\delta_ {\\mathrm {h d}}\\right) \\\\ = \\frac {1}{\\delta_ {\\mathrm {h d}} ^ {2}} \\log \\frac {\\sum_ {y ^ {\\prime}} \\exp \\left(f _ {y ^ {\\prime}} ^ {\\theta_ {\\mathrm {h d}}} (x)\\right)}{\\exp \\left(f _ {y} ^ {\\theta_ {\\mathrm {h d}}} (x)\\right)} \\\\ + \\log \\frac {\\sum_ {y ^ {\\prime}} \\exp \\left(\\frac {1}{\\delta_ {\\mathrm {h d}} {} ^ {2}} f _ {y ^ {\\prime}} ^ {\\theta_ {\\mathrm {h d}}} (x)\\right)}{\\left(\\sum_ {y ^ {\\prime}} \\exp \\left(f _ {y ^ {\\prime}} ^ {\\theta_ {\\mathrm {h d}}} (x)\\right)\\right) ^ {\\frac {1}{\\delta_ {\\mathrm {h d}} {} ^ {2}}}} \\\\ \\approx - \\frac {1}{\\delta_ {\\mathrm {h d}} ^ {2}} \\log \\operatorname {s o f t m a x} \\left(Y _ {\\mathrm {h d}}, f ^ {\\theta_ {\\mathrm {h d}}} (x)\\right) + \\log \\delta_ {\\mathrm {h d}}, \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.292, + 0.891, + 0.365 + ], + "angle": 0, + "content": "where \\(f_{y}^{\\theta_{\\mathrm{hd}}}(x)\\) refers to the y-th value of the \\(f^{\\theta_{\\mathrm{hd}}}(x)\\). We introduce a simplify assumption \\(\\sum_{y'} \\exp \\left( \\frac{1}{\\delta_{\\mathrm{hd}}^2} f_{y'}^{\\theta_{\\mathrm{hd}}}(x) \\right) \\approx \\left( \\sum_{y'} \\exp \\left( f_{y'}^{\\theta_{\\mathrm{hd}}}(x) \\right) \\right)^{\\frac{1}{\\delta_{\\mathrm{hd}}^2}}\\) when \\(\\delta_{\\mathrm{hd}} \\to 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.365, + 0.891, + 0.439 + ], + "angle": 0, + "content": "- log softmax \\(\\left(Y_{\\mathrm{hd}},f^{\\theta_{\\mathrm{hd}}}(x)\\right)\\) of Eq. (7) represents the cross-entropy classification loss of \\(Y_{\\mathrm{hd}}\\). Inspired on this term, we generalize this expression to other classification losses. The HD loss function \\(\\mathcal{L}_{\\mathrm{hd}}(\\theta_{\\mathrm{hd}},\\delta_{\\mathrm{hd}})\\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.45, + 0.891, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {h d}} \\left(\\theta_ {\\mathrm {h d}}, \\delta_ {\\mathrm {h d}}\\right) = \\frac {1}{\\delta_ {\\mathrm {h d}} ^ {2}} \\mathcal {L} \\left(\\theta_ {\\mathrm {h d}}\\right) + \\log \\delta_ {\\mathrm {h d}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.492, + 0.891, + 0.598 + ], + "angle": 0, + "content": "Consistent with prior approaches [25, 36], \\(\\mathcal{L}(\\theta_{\\mathrm{hd}})\\) includes the hinge loss \\(L_{\\mathrm{hinge}}\\), negative video-query pairs loss \\(L_{\\mathrm{neg}}\\) and rank-aware contrastive loss [17] \\(L_{\\mathrm{cont}}\\), i.e. \\(\\mathcal{L}(\\theta_{\\mathrm{hd}}) = L_{\\mathrm{hinge}} + L_{\\mathrm{neg}} + L_{\\mathrm{cont}}\\). \\(L_{\\mathrm{hinge}}\\) is computed between two pairs of positive and negative clips, with its margin 0.2 to maintain consistency with [25] for fairness. \\(L_{\\mathrm{neg}}\\) and \\(L_{\\mathrm{cont}}\\) from [36] are used to reduce the saliency of negative pairs." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.599, + 0.89, + 0.629 + ], + "angle": 0, + "content": "We can obtain final task-dependent joint loss \\(\\mathcal{L}_{\\mathrm{joint}}\\) through integrating Eq. (5) and Eq. (8):" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.642, + 0.891, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {j o i n t}} = \\mathcal {L} _ {\\mathrm {m r}} \\left(\\theta_ {\\mathrm {m r}}, \\delta_ {\\mathrm {m r}}\\right) + \\mathcal {L} _ {\\mathrm {h d}} \\left(\\theta_ {\\mathrm {h d}}, \\delta_ {\\mathrm {h d}}\\right). \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.671, + 0.632, + 0.688 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.69, + 0.712 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Datasets. Extensive experiments are conducted on three benchmark datasets: QVHighlights [25], TVSum [43], and Charades-STA [14]. QVHighlights is currently the sole publicly dataset for joint moment retrieval and highlight detection tasks. It provides 10,310 queries associated with 18,367 moments, with an average of 1.8 disjoint moments per query. In contrast to other MR datasets with one-to-one query-moment mappings, QVHighlights aligns more closely with real-world scenarios. Each video in the dataset comprises 75 clips, each of which is 2s-long." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "We also utilize two task-specific datasets, Charades-STA [14] for MR and TVSum [43] for HD, to evaluate the model." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18312" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.546, + 0.338 + ], + "angle": 0, + "content": "
MethodVT VU GA MS PK PR FM BK BT DS Avg.
sLSTM [64]ECCV'1641.1 46.2 46.3 47.7 44.8 46.1 45.2 40.6 47.1 45.5 45.1
SG [34]CVPR'1742.3 47.2 47.5 48.9 45.6 47.3 46.4 41.7 48.3 46.6 46.2
LIM-S [51]CVPR'1955.9 42.9 61.2 54.0 60.4 47.5 43.2 66.3 69.1 62.6 56.3
Trailer [47]ECCV'2061.3 54.6 65.7 60.8 59.1 70.1 58.2 64.7 65.6 68.1 62.8
SL-Module [52]ICCV'2186.5 68.7 74.9 86.2 79.0 63.2 58.9 72.6 78.9 64.0 73.3
MINI-Net† [18]ECCV'2080.6 68.3 78.2 81.8 78.1 65.8 57.8 75.0 80.2 65.5 73.2
TCG† [57]ICCV'2185.0 71.4 81.9 78.6 80.2 75.5 71.6 77.3 78.6 68.1 76.8
Joint-VA† [3]ICCV'2183.7 57.3 78.5 86.1 80.1 69.2 70.0 73.0 97.4 67.5 76.3
UMT† [32]CVPR'2287.5 81.5 88.2 78.8 81.4 87.0 76.0 86.9 84.4 79.6 83.1
QD-DETR [36]CVPR'2388.2 87.4 85.6 85.0 85.8 86.9 76.4 91.3 89.2 73.7 85.0
UniVTG‡ [29]ICCV'2383.9 85.1 89.0 80.1 84.6 81.4 70.9 91.7 73.5 69.3 81.0
TaskWeave(Ours)88.2 90.8 93.3 87.5 87.0 82.0 80.9 92.9 89.5 81.2 87.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.35, + 0.53, + 0.38 + ], + "angle": 0, + "content": "Table 1. Experimental results \\((\\%)\\) on TVSum. \\(\\dagger\\) means including audio modality. \\(\\ddagger\\) means following the pretrain-finetune paradigm." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.406, + 0.536, + 0.579 + ], + "angle": 0, + "content": "
MethodMRHD
R1mAP≥ Very Good
@0.5@0.7@0.5@0.75Avg.mAP HIT@1
Moment-DETR [25]NIPS'2153.9434.84--32.2035.6555.55
UMT†[32]CVPR'2260.2644.26--38.5939.8564.19
QD-DETR [36]CVPR'2362.6846.6662.2341.8241.2239.1363.03
EaTR [20]ICCV'2361.3645.7961.8641.9141.7437.1558.65
UniVTG‡[29]ICCV'2359.74---36.1338.8361.81
TaskWeave(Ours)64.2650.0665.3946.4745.3839.2863.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.589, + 0.529, + 0.619 + ], + "angle": 0, + "content": "Table 3. Experimental results (\\%) on QVHighlights val split. † means including audio modality. ‡ means following the pretrain-finetune paradigm." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.69 + ], + "angle": 0, + "content": "Charades-STA provides 16,128 query-moment pairs. TVSum comprises videos from 10 domains, with each domain containing 5 videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. We utilize the same evaluation metrics used in prior approaches [25, 32, 36]. For QVHighlights, mean average precision (mAP) with different tIoU thresholds 0.5, 0.75, the average mAP over [0.5:0.05:0.95], and Recall@1 with tIoU 0.5, 0.7 are utilized for MR evaluation. mAP and HIT@1 are used for HD evaluation, where a clip is considered as a true positive when it achieves a \"Very Good\" [25] saliency score. For Charades-STA, we utilize Recall@1 with tIoU thresholds 0.5, 0.7. For TVSum, Top-5 mAP is adopted." + }, + { + "type": "table", + "bbox": [ + 0.575, + 0.089, + 0.885, + 0.339 + ], + "angle": 0, + "content": "
BackboneMethodR1@0.5R1@0.7
VGGSAP [9]AAAI'1927.4213.36
SM-RL [48]CVPR'1924.3611.17
2D-TAN [65]AAAI'2040.9422.85
FVMR [13]CVPR'2124.3611.17
UMT† [32]CVPR'2248.3129.25
QD-DETR [36]CVPR'2352.7731.13
TaskWeave(Ours)56.5133.66
I3DCTRL [14]ICCV'1723.638.89
MAN [60]CVPR'1946.5322.72
VSLNet [61]ACL'2047.3130.19
QD-DETR [36]CVPR'2350.6731.02
TaskWeave(Ours)53.3631.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.561, + 0.35, + 0.892, + 0.393 + ], + "angle": 0, + "content": "Table 2. Experimental results \\((\\%)\\) on Charades-STA test split. \\(\\dagger\\) means including audio modality. \\(\\ddagger\\) means following the pretrain-finetune paradigm." + }, + { + "type": "table", + "bbox": [ + 0.57, + 0.443, + 0.888, + 0.579 + ], + "angle": 0, + "content": "
Task-decoupled UnitInter-task feedbackJoint lossMRHD
R1 @0.7Avg. mAPmAP HIT@1
46.2641.038.9462.84
✓47.8743.2438.5861.81
✓✓49.2945.1238.9662.0
✓✓✓50.0645.3839.2863.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.561, + 0.589, + 0.892, + 0.618 + ], + "angle": 0, + "content": "Table 4. The ablation results \\((\\%)\\) of the components of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.644, + 0.719, + 0.66 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Feature representations. For OVHighlights, we employ the pre-trained SlowFast [12] and CLIP [39] backbones to extract video features, following [25, 32, 36] for fairness. For Charedes-STA, we leverage the official release of VGG [42] and I3D [6] features as video embeddings. For TVSum, we extract video features by the I3D [6] pre-trained on Kinetics 400 [21]. For QVHighlights and TVSum, we use CLIP [39] to extract text features, while using GloVe [38] text embeddings for Charades-STA." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Training settings. We leverage AdamW [24] optimizer with 1e-4 learning rate and 1e-4 weight decay. We train 200, 100, and 2000 epochs with batch size 32, 32 and 2 for QVHighlights, Charades-STA and TVSum, respectively. All Transformer layers follow the consistent configuration, including sinusoidal positional encodings, 8 at" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18313" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.365, + 0.259 + ], + "angle": 0, + "content": "
IndexMR-expertHR expertMRHDR
R1 @0.7Avg. mAPmAPHIT@1
(a)Iden.Iden.44.7740.2338.3860.19
(b)Line.Line.47.5543.6538.3260.77
(c)CNNCNN49.2343.7338.8260.77
(d)Trans.Trans.48.3242.8538.6161.10
(e)Trans.Iden.47.142.138.761.68
(f)CNNIden.49.2945.1238.9662.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.268, + 0.368, + 0.311 + ], + "angle": 0, + "content": "Table 5. Flexibility validation \\((\\%)\\) of the taskdecoupled unit. \"Iden:\" identity mapping; \"Trans\":Transformer; \"Line\":linear layer." + }, + { + "type": "table", + "bbox": [ + 0.384, + 0.133, + 0.638, + 0.26 + ], + "angle": 0, + "content": "
TypeMRHD
R1 @0.7Avg. mAPmAP HIT@1
Sum47.6844.2738.760.32
Weighted Sum48.3243.2138.3161.35
Ours50.0645.3839.2863.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.381, + 0.268, + 0.631, + 0.297 + ], + "angle": 0, + "content": "Table 6. Effectiveness justification \\((\\%)\\) of the task-dependent joint loss." + }, + { + "type": "table", + "bbox": [ + 0.649, + 0.104, + 0.889, + 0.259 + ], + "angle": 0, + "content": "
TypeMRHD
R1 @0.7Avg. mAPmAPHIT@1
MR2HD50.0645.3838.7362.84
HD2MR48.6544.7839.2863.68
Bi-MRHD48.8445.138.3661.81
MR-HD50.045.338.9662.45
HD-MR49.144.8739.0359.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.643, + 0.268, + 0.892, + 0.311 + ], + "angle": 0, + "content": "Table 7. Comparing various inter-task feedback combinations \\((\\%)\\) on QVHighlights val split." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.473, + 0.504 + ], + "angle": 0, + "content": "tention heads, and a dropout rate of 0.1. Multi-modal fusion employs a 2-layer Transformer with cross attention, where video features serve as the query and text features serve as key and value. For Eq. (9), we define \\(\\gamma_{\\mathrm{mr}}, \\gamma_{\\mathrm{hd}}\\) as \\(\\log \\delta_{\\mathrm{mr}}^2, \\log \\delta_{\\mathrm{hd}}^2\\), respectively. Therefore, Eq. (9) can be rewritten as \\(\\mathcal{L}_{joint} = \\exp(-\\gamma_{\\mathrm{mr}})\\mathcal{L}(\\theta_{\\mathrm{mr}}) + 2\\exp(-\\gamma_{\\mathrm{hd}})\\mathcal{L}(\\theta_{\\mathrm{hd}}) + \\gamma_{\\mathrm{mr}} + \\gamma_{\\mathrm{hd}}\\). \\(\\gamma_{\\mathrm{mr}}\\) and \\(\\gamma_{\\mathrm{hd}}\\) are learnable parameters. They are initialized to 0. To further stabilize the training, the model EMA strategy has also undertaken. All experiments are conducted with Pytorch v1.13.1 on a single NVIDIA RTX 3090." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.514, + 0.379, + 0.53 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.688 + ], + "angle": 0, + "content": "Results on TVSum. In Tab. 1, we compare our TaskWeave with existing state-of-the-arts (SOTA) methods, where the methods with \\(\\dagger\\) are incorporate with audio features. UniVTG [29] follows pretrain-then-finetune paradigm, therefore we use its non-pretrained results for a fair comparison. We observe that: i) our methods outperforms SOTA methods in 9 out of the 10 categories; ii) TaskWeave significantly outperforms all methods in terms of Avg. mAP, with a remarkable \\(2.71\\%\\) improvement over the previous SOTA method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Results on Charades-STA. In Tab. 2, we utilize different backbones to demonstrate the effectiveness of our TaskWeave. We compare our results with other methods for MR only (white background) and methods for joint MR and HD (gray background). It results in significant improvements, i.e., \\(+7.09\\%\\) in R1@0.5 and \\(+8.13\\%\\) in R1@0.7." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Results on QVHighlights. As shown in Tab. 3, we present the performance comparisons with existing methods. We report the MR results with the MR-guided feedback, while HD results with the HD-guided feedback. We observe that: i) our TaskWeave surpasses SOTA methods by a large margin in all moment retrieval metrics, with a remarkable \\(8.72\\%\\) improvement in Avg. mAP; ii) our method achieves the second-best performance on the HD. However," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.895, + 0.412 + ], + "angle": 0, + "content": "It's worth noting that UMT [32] incorporates audio features, while our method does not. When compared to other methods without audio features, our method outperforms them all. In summary, these results validate the effectiveness of our task-driven framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.895, + 0.535 + ], + "angle": 0, + "content": "We provide visualization examples in Fig. 4 for the qualitative analysis among our TaskWeave, Moment-DETR [25], and QD-DETR [36]. Given diverse queries for a video, TaskWeave precisely localizes moments for queries and presents high IoU with the ground-truth. Moment-DETR misses some instances, and QD-DETR exhibits lower retrieval accuracy. Moreover, our method obtains higher saliency scores for relevant clips in response to the query." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.792, + 0.56 + ], + "angle": 0, + "content": "4.4. Ablation Studies and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.613 + ], + "angle": 0, + "content": "We present some ablation studies and discussions about TaskWeave, with all experiments are conducted on the QVHighlights val split [25]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.749 + ], + "angle": 0, + "content": "Ablation of components. We conduct ablation studies on each component of TaskWeave, as shown in Tab. 4. Sequentially employing the task-decoupled unit and our proposed joint loss contributes a \\(5.5\\%\\) improvement in average mAP and a \\(6.5\\%\\) increase in R1@0.7 for MR. Utilizing all components, we observe that a significant \\(10.7\\%\\) improvement in average mAP of MR and a \\(1.3\\%\\) enhancement in HIT@1 of HD. These results indicate the effectiveness of the proposed components in out TaskWeave." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Flexibility of the task-decoupled unit. To demonstrate the flexibility of the task-decoupled unit, we investigate the performance of applying different networks within various experts, as shown in Tab. 5. The shared expert utilize a fixed configuration with 2 Transformer layers. We provide 4 different methods for each task-specific expert, including the feed-forward network (one layer Linear), the identity mapping, CNN (composed of one depthwise convolution layer with the kernel/stride/padding of \\(5/1/2\\) and 1D convolution), and the Transformer. Due to space constraints, we present" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18314" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.092, + 0.851, + 0.105 + ], + "angle": 0, + "content": "Query1: A guy with grey top is showing a box filled with rubbish. Query2: A man shaves his beard looks in a mirror and talks to the camera." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.108, + 0.895, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.323, + 0.893, + 0.352 + ], + "angle": 0, + "content": "Figure 4. Qualitative results on the QVHighlights for Ground-Truth, Moment-DETR, QD-DETR and our method. The predicted moments and saliency scores are illustrated through intervals and lines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.47, + 0.408 + ], + "angle": 0, + "content": "results of 6 out of 16 combinations. The results in Tab. 5 are obtained by TaskWeave without inter-task feedback." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.47, + 0.653 + ], + "angle": 0, + "content": "From this table, we can observe several interesting facts. First, different task-specific experts have different performance. The need for experts is reflected in (a). Second, although Transformer has made significant progress in visual, it's not a panacea. Results in (d) are lower compared to (b) and (c), we believe that the high computational complexity of Transformer decreases the performance. Third, the task-specific expert should be designed based on the task objective. For instance, (c) and (f) perform well on the MR because they focus on local features, which is important for localizing moments. Finally, we also find evidence that MR and HD are highly related. Improvement in one task enhances the other (compare (c) and (f)), while a decline in one limits the other's performance (compare (d) and (e)). In this paper, our mr-specific expert utilizes CNN and hd-specific expert is implemented by identity mapping." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.793 + ], + "angle": 0, + "content": "Task-dependent joint loss. Tab. 6 shows the performance of TaskWeave with different losses. \"Sum\" and \"Weighted Sum\" refer to the fusion manner of loss for different tasks, respectively. Weights in \"Weighted Sum\" are consistent with the existing methods [25, 36] for fairness. Comparing \"Sum\" and \"Weighted Sum\", we find that \"Weighted Sum\" can better balance the performance of the two tasks. However, it is obvious that the proposed task-dependent joint loss is optimal." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Inter-task interactions. We believe that each task requires learning before providing effective feedback. Therefore, in our inter-task feedback mechanism, the feedback process starts when the model is trained to half of the max epoch. For brevity, we write Moment/Highlightness-guided feedback as MgF/HgF. In Tab. 7, we explore five feedback manners as follows: \"MR2HD\" (with MgF only)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.892, + 0.424 + ], + "angle": 0, + "content": "\"HD2MR\" (with HgF only), \"Bi-MRHD\" (with MgF and HgF simultaneously), \"MR-HD\" (MgF first, then HgF), and \"HD-MR\" (HgF first, then MgF)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.893, + 0.575 + ], + "angle": 0, + "content": "We find that the inter-task feedback makes both tasks gain simultaneously. The performance of the model with \"HD2MR\" is slightly worse, this is because HD focuses on more refined moments than MR. Ground-truth annotations and the results of \"MR-HD\" and \"HD-MR\" also illustrate this fact. The gain brought by \"Bi-MRHD\" is limited, which we believe that it's a natural result of feedback not always being effective. In general, the inter-task feedback not only contributes to bring gains for both tasks but also helps to understand the characteristics of each task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.587, + 0.62, + 0.603 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.793 + ], + "angle": 0, + "content": "This paper proposes a novel task-driven paradigm for addressing joint moment retrieval and highlight detection. Different from existing data-driven methods, we utilize the task-decoupled unit to capture the task-specific and common features, respectively. We also explore different network architecture for moment retrieval and highlight detection. We design the inter-task feedback mechanism to in-depth investigate the interplay between both tasks. Different from prior methods, we introduce the principled joint loss to optimize the model. The effectiveness, flexibility, and superiority of the proposed method have been demonstrated on three benchmark datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.807, + 0.661, + 0.823 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.831, + 0.892, + 0.887 + ], + "angle": 0, + "content": "This research was supported by the National Natural Science Foundation of China (No. U23B2060, 62088102), the Youth Innovation Team of Shaanxi Universities, and the Fundamental Research Funds for the Central Universities." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18315" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Niki Parmar Jakob Uszkoreit Llion Jones Aidan N. Gomez Lukasz Kaiser Ashish Vaswani, Noam Shazeer and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5999-6009, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.471, + 0.314 + ], + "angle": 0, + "content": "[3] Taiwanbat Badamdorj, Mrigank Rochan, Yang Wang, and Li Cheng. Joint visual and audio learning for video highlight detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8127-8137, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.471, + 0.383 + ], + "angle": 0, + "content": "[4] Taiwanbat Badamdorj, Mrigank Rochan, Yang Wang, and Li Cheng. Contrastive learning for unsupervised video highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14042-14052, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.469, + 0.528 + ], + "angle": 0, + "content": "[7] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.469, + 0.585 + ], + "angle": 0, + "content": "[8] Jingyuan Chen, Lin Ma, Xinpeng Chen, Zequn Jie, and Jiebo Luo. Localizing natural language in videos. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8175-8182, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.588, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[9] Shaoxiang Chen and Yu-Gang Jiang. Semantic proposal for activity localization in videos via sentence query. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8199-8206, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[10] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[12] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Junyu Gao and Changsheng Xu. Fast video moment retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1523-1532, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[14] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[15] Ana Garcia del Molino and Michael Gygli. PhD-gifs: personalized highlight detection for automatic gif creation. In Proceedings of the 26th ACM international conference on Multimedia, pages 600-608, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[16] Michelle Guo, Albert Haque, De-An Huang, Serena Yeung, and Li Fei-Fei. Dynamic task prioritization for multitask learning. In Proceedings of the European conference on computer vision (ECCV), pages 270–287, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[17] David T Hoffmann, Nadine Behrmann, Juergen Gall, Thomas Brox, and Mehdi Noroozi. Ranking info noise contrastive estimation: Boosting contrastive learning via ranked positives. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 897–905, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.411 + ], + "angle": 0, + "content": "[18] Fa-Ting Hong, Xuanteng Huang, Wei-Hong Li, and Wei-Shi Zheng. Mini-net: Multiple instance ranking network for video highlight detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 345-360. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.413, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[19] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[20] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13846-13856, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[21] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.637 + ], + "angle": 0, + "content": "[22] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[23] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482-7491, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[25] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34: 11846-11858, 2021. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.873 + ], + "angle": 0, + "content": "[26] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Huan Li, Ping Wei, Jiapeng Li, Zeyu Ma, Jiahui Shang, and Nanning Zheng. Asymmetric relation consistency reasoning" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18316" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "for video relation grounding. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.188 + ], + "angle": 0, + "content": "[28] Huan Li, Ping Wei, Zeyu Ma, and Nanning Zheng. Inverse compositional learning for weakly-supervised relation grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15477-15487, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[29] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified videolanguage temporal grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2794-2804, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.275, + 0.47, + 0.329 + ], + "angle": 0, + "content": "[30] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1871-1880, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[31] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[32] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3042-3051, 2022. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[33] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[34] Behrooz Mahasseni, Michael Lam, and Sinisa Todorovic. Unsupervised video summarization with adversarial LSTM networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 202-211, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.47, + 0.691 + ], + "angle": 0, + "content": "[35] Otniel-Bogdan Mercea, Lukas Riesch, A Koepke, and Zeynep Akata. Audio-visual generalised zero-shot learning with cross-modal attention and language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10553–10563, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[36] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23023-23033, 2023. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[37] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810-10819, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[38] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[40] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[41] Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, and Dacheng Tao. Tridet: Temporal action detection with relative boundary modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18857-18866, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[43] Yale Song, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaimes. Tvsum: Summarizing web videos using titles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5179-5187, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[44] Min Sun, Ali Farhadi, and Steve Seitz. Ranking domain-specific highlights by analyzing edited videos. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13, pages 787-802. Springer, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[45] Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. Progressive layered extraction (ple): A novel multi-task learning (mtl) model for personalized recommendations. In Proceedings of the 14th ACM Conference on Recommender Systems, pages 269-278, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[46] Tuan N Tang, Kwonyoung Kim, and Kwanghoon Sohn. Temporalmaxer: Maximize temporal context with only max pooling for temporal action localization. arXiv preprint arXiv:2303.09055, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.714 + ], + "angle": 0, + "content": "[47] Lezi Wang, Dong Liu, Rohit Puri, and Dimitris N Metaxas. Learning trailer moments in full-length movies with co-contrastive attention. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVIII 16, pages 300–316. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[48] Weining Wang, Yan Huang, and Liang Wang. Language-driven temporal activity localization: A semantic matching reinforcement learning model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 334-343, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[49] Fanyue Wei, Biao Wang, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Learning pixel-level distinctions for video highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3073-3082, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[50] Xi Wei, Tianzhu Zhang, Yan Li, Yongdong Zhang, and Feng Wu. Multi-modality cross attention network for image and sentence matching. In Proceedings of the IEEE/CVF con" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18317" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "ference on computer vision and pattern recognition, pages 10941-10950, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.469, + 0.191 + ], + "angle": 0, + "content": "[51] Bo Xiong, Yannis Kalantidis, Deepi Ghadiyaram, and Kristen Grauman. Less is more: Learning highlight detection from video duration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1258-1267, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.195, + 0.469, + 0.264 + ], + "angle": 0, + "content": "[52] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7970-7979, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.267, + 0.469, + 0.322 + ], + "angle": 0, + "content": "[53] Yifang Xu, Yunzhuo Sun, Yang Li, Yilei Shi, Xiaoxiang Zhu, and Sidan Du. Mh-detr: Video moment and highlight detection with cross-modal transformer. arXiv preprint arXiv:2305.00355, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.325, + 0.469, + 0.366 + ], + "angle": 0, + "content": "[54] Jin Yang, Ping Wei, Ziyang Ren, and Nanning Zheng. Gated multi-scale transformer for temporal action localization. IEEE Transactions on Multimedia, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.369, + 0.469, + 0.424 + ], + "angle": 0, + "content": "[55] Jin Yang, Ping Wei, and Nanning Zheng. Cross time-frequency transformer for temporal action localization. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.469, + 0.483 + ], + "angle": 0, + "content": "[56] Ting Yao, Tao Mei, and Yong Rui. Highlight detection with pairwise deep ranking for first-person video summarization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 982–990, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.486, + 0.469, + 0.554 + ], + "angle": 0, + "content": "[57] Qinghao Ye, Xiyue Shen, Yuan Gao, Zirui Wang, Qi Bi, Ping Li, and Guang Yang. Temporal cue guided video highlight detection with low-rank audio-visual fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7950–7959, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.557, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[58] Yitian Yuan, Tao Mei, and Wenwu Zhu. To find where you talk: Temporal sentence localization in video with attention based location regression. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9159–9166, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.469, + 0.684 + ], + "angle": 0, + "content": "[59] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.769 + ], + "angle": 0, + "content": "[60] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1247-1257, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.469, + 0.841 + ], + "angle": 0, + "content": "[61] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6543-6554, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[62] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Temporal sentence grounding in videos: A survey and future directions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10443-10465, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[63] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Temporal sentence grounding in videos: a survey and future directions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[64] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. Video summarization with long short-term memory. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 766-782. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[65] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12870-12877, 2020. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18318" + } + ] +] \ No newline at end of file diff --git a/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/a6a833ff-f398-48fe-909b-70ff81bf1b94_origin.pdf b/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/a6a833ff-f398-48fe-909b-70ff81bf1b94_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..382619babf1fd4b7d2b24b158795eecaf3a9eef7 --- /dev/null +++ b/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/a6a833ff-f398-48fe-909b-70ff81bf1b94_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:422d543a7038db2cd2df6a9a6cedf42b1c45c33296ca9e59787430cdae7c1d5a +size 334210 diff --git a/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/full.md b/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6700e49ddafe0232a122faf2eb2a393e3bf4dfc0 --- /dev/null +++ b/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/full.md @@ -0,0 +1,343 @@ +# Task-Driven Exploration: Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection + +Jin Yang, Ping Wei*, Huan Li, Ziyang Ren +National Key Laboratory of Human-Machine Hybrid Augmented Intelligence Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University +{jin.yang, lh875056558, rzyrzy}@stu.xjtu.edu.cn, pingwei@xjtu.edu.cn + +# Abstract + +Video moment retrieval and highlight detection are two highly valuable tasks in video understanding, but until recently they have been jointly studied. Although existing studies have made impressive advancement recently, they predominantly follow the data-driven bottom-up paradigm. Such paradigm overlooks task-specific and inter-task effects, resulting in poor model performance. In this paper, we propose a novel task-driven top-down framework TaskWeave for joint moment retrieval and highlight detection. The framework introduces a task-decoupled unit to capture task-specific and common representations. To investigate the interplay between the two tasks, we propose an inter-task feedback mechanism, which transforms the results of one task as guiding masks to assist the other task. Different from existing methods, we present a task-dependent joint loss function to optimize the model. Comprehensive experiments and in-depth ablation studies on QVHighlights, TVSum, and CharadesSTA datasets corroborate the effectiveness and flexibility of the proposed framework. Codes are available at github.com/EdenGabriel/TaskWeave. + +# 1. Introduction + +As videos are prevailing in a wide range of applications, the diversity and massive scales of video content have posed unprecedented challenges in finding relevant moments based on user queries. To this end, the moment retrieval (MR) [60, 63, 65] and highlight detection (HD) [4, 56] tasks have emerged recently. MR aims to retrieve video moments that are relevant to the given query [14, 62]. HD aims to predict the clip-level saliency scores in the video [44]. + +As MR and HD tasks are closely related, they have been jointly addressed and achieved breakthroughs recently + +![](images/66e4867e1d41709697bfc3ea6505e5271921a04678a42e0b1c7df877dbbf6dce.jpg) +Figure 1. Although the positive temporal intervals of moment retrieval and highlight detection exhibit high overlap, they pursue different objectives. + +[20, 25, 29, 32, 36, 53]. The existing joint approaches in general utilize a shared backbone to learn the multimodal features as the common representations for MR and HD. Then a MR prediction head is employed for moments localization and a HD prediction head predicts saliency scores. These methods adhere to the bottom-up, data-driven paradigm, i.e. they capture common features from the input data and then utilize the features for different tasks. The effectiveness of these methods is built upon the premise of the high correlation between MR and HD. As illustrated in Fig. 1, the task MR and HD share the identical query and video inputs, additionally exhibiting a substantial temporal overlap between their respective positive temporal intervals. + +However, the bottom-up, data-driven paradigm tends to excessively rely on the common features, but overlooks the inherent specific characteristics of MR and HD. This tendency might simplify the joint modeling as a problem of feature fusion, without considering the interplay between the two tasks. The distinct objectives pursued by MR and HD rely on distinct task-specific characteristics. Unfortunately, existing methods overlook the specificity. + +To address the aforementioned issues, we believe it is essential to leverage the fundamental multi-modal data to + +mine commonalities across tasks (bottom-up), while also strengthen the awareness of task-specific characteristics (top-down). To this end, we propose a novel paradigm TaskWeave from a task-driven perspective. The key idea is to jointly address the tasks MR and HD by considering the commonality, specificity, and interplay of MR and HD. + +To effectively capture the commonality and specificity, we design a task-decoupled unit, i.e. a shared expert to capture common features and two task-specific experts to acquire the distinct characteristics. In order to investigate the interplay between MR and HD in-depth, we design an inter-task feedback mechanism. It converts the predictions of MR/HD into mask information, which are fed back to the input of the HD/MR prediction head. Furthermore, we introduce a principled task-dependent joint loss in which the task-specific weights are dynamically adjusted, rather than manually tuned. + +We conduct experiments on the QVHighlights [25] dataset to validate the effectiveness of the proposed method. Moreover, we also conduct experiments for two individual tasks on the datasets Charades-STA [14] (moment retrieval) and TVSUM [43] (highlight detection). The proposed approach outperforms the existing methods. + +The key contributions of this paper are three folds. + +1. It proposes a novel task-driven, top-down framework for joint moment retrieval and highlight detection. +2. It introduces a task-decoupled unit, an inter-task feedback mechanism, and a principled task-dependent loss. +3. It achieves state-of-the-art performance on three datasets. The ablation study validates the methods. + +# 2. Related Work + +We review the related work from four aspects. + +Moment Retrieval. Existing approaches for MR mainly include two groups. One group follows a two-stage procedure [1, 14, 60, 65], which involves generating candidate temporal intervals and ranking them based on the correlation with the query. The other group directly regresses the temporal interval based on the aligned visual-text features [8, 37, 58, 59, 61]. Moreover, most datasets provide only one moment annotation for each video-query pair [1, 14], which does not align with real-world scenarios. + +Highlight Detection. The saliency score in the highlight detection represents the relevance of a video clip to the given query. Most prior highlight detection benchmark datasets are query-agnostic [15, 44]. The saliency scores for video clips remained constant regardless of the query. As a result, some previous approaches treated HD as a solely visual task [4, 44, 49, 56]. + +MR and HD tasks are traditionally studied separately. They have been jointly addressed recently as the introduction of QVHighlights dataset [25]. QVHighlights provides multiple moment-annotations for each query and ensures + +these moments are uniformly distributed throughout the video. It also provides query-dependent highlightness annotations. With QVHighlights, the model Moment-DETR [25] is proposed for joint MR and HD. Following Moment-DETR, a growing number of approaches have been proposed to accomplish the joint task [20, 29, 32, 36]. However, these methods adopt the data-driven and bottom-up paradigm. Different from them, we propose a novel task-driven and top-down paradigm. + +Vision Transformers. Transformer-based models [2] have brought huge achievement in both the image and video related domains [5, 11, 27, 28, 54, 55]. One of the most well-known methods is DETR [5], which regards object detection as a set prediction problem. Its end-to-end prediction procedure eliminates the intermediate or post-processing steps. On the other hand, some studies have employed cross-attention mechanism [35, 36, 50] to inject multi-modal data into the Transformer architecture. In this paper, we also adopt a DETR-like architecture [26, 31]. However, different from those methods, we incorporate diverse network architectures for feature extraction. Our focus lies in the collaboration of the network architectures. + +Multi-task Learning. Multi-task learning (MTL) aims to train a single model for multiple tasks [7]. The most straightforward approach is to utilize a shared backbone to extract common features, which relies on the data-driven manner. However, it often leads to suboptimal performance for unrelated tasks and lacks flexibility. In response, some studies introduce MOE [19] and MMOE [33], which utilize a set of the shared experts in place of the shared bottom layer. To mitigate the seesaw phenomenon of multi-task learning, PLE [45] employs separate experts for each task while still retaining the shared experts. From the perspective of MTL, existing approaches for joint MR and HD follow the shared bottom paradigm. Moreover, directly applying MTL methods to joint MR and HD does not yield favorable results. Existing MTL methods are complex, therefore direct usage of them leads to a sharp increase in model complexity. This might decrease the performance. To this end, we aim to design an effective task-driven framework to jointly address MR and HD. + +# 3. Methodology + +# 3.1. Overview + +Given a text query with $W$ words and an untrimmed video composed of $N$ clips, the objective of the joint moment retrieval (MR) and highlight detection (HD) is to localize the center coordinate $q_{c}$ and width $q_{w}$ of temporal intervals that are relevant to the text query, in addition to ranking clip-wise saliency scores. + +Architecture Overview. Given the high correlation between moment retrieval and highlight detection tasks, the + +![](images/a101a6befdf96406fa8311a1c3ba8d48453ca0993b6bbf611c746faf88eff16d.jpg) +Figure 2. The overall pipeline of the proposed task-driven model TaskWeave. We propose the task-decoupled unit to capture task-specific and common features. Various experts can adopt different network implementations, showcasing the flexibility of the model. Inter-task feedback mechanism is designed to investigate the influence between both tasks. There are two feedback manners: Moment-guided and Highlightness-guided feedback. The principled task-dependent joint loss is introduced for jointly optimize the model. + +most intuitive approach is using a shared backbone in conjunction with two task-specific prediction heads. It is a data-driven paradigm that is commonly employed by previous methods, for its simplicity and ease of implementation. However, the task-specific characteristics are inherently present, since MR and HD have distinct objectives. Additionally, the interplay between MR and HD should also be considered, which can enhance the model's performance. + +The overall pipeline of our approach is illustrated in Fig. 2. We employed the frozen video/text-encoder backbones to extract the video/text features, while ensuring their dimensions remained consistent at $D$ by the projection. These features are utilized for multi-modal fusion through methods such as cross-attention or concatenation, resulting in query-related video representations $Z \in \mathbb{R}^{N \times D}$ . In our paper, we extract these representations through the cross-attention. After cross-attention, inspired by [41, 46], a 1D Max Pooling (MaxPool1d) with the kernel size 5, stride 1 and padding 2 is utilized to eliminate the rank loss problem in the attention mechanism. These representations are fed into the task-decoupled unit to capture the task-related features $X_{\mathrm{mr}} \in \mathbb{R}^{N \times D}$ , $X_{\mathrm{hd}} \in \mathbb{R}^{N \times D}$ . Then, we employ the task-specific decoders with inter-task feedback mechanism to make predictions for moments localization and clip-wise saliency scores. We introduce the principled task-dependent loss to jointly optimize the model. + +# 3.2. Task-decoupled Unit + +Since moment retrieval and highlight detection have distinct objectives, the specificity of each task should be considered, rather than solely focusing on their commonalities. For this purpose, we propose a task-decoupled unit from a task-driven perspective to capture the task-related features, + +which involves task-specific features and common features. + +The task-decoupled unit is depicted in Fig. 2. Inspired by the attention mechanism [2], the query-related video representation $Z$ is initially fed into two task-specific mappers, mr-mapper and hd-mapper. Each mapper is implemented using one-layer feed-forward network. The output of each task-specific mapper is directed into both the respective task-specific expert network (MR-specific Expert or HD-specific Expert) and the Shared Expert network. + +Thanks to the design of our task-decoupled framework, each expert can employ various networks, such as convolutional network, Transformer, and feed-forward network. It offers more flexibility in addressing the joint MR and HD tasks. In our study, we also investigate the influence of different expert networks, referring to Sec. 4 for more details. Consequently, we can capture the task-related features $X_{\mathrm{mr}}$ and $X_{\mathrm{hd}}$ via element-wise product between the output of the task-specific expert and that of the shared expert. + +The MR-specific feature $X_{\mathrm{mr}}$ can be computed as: + +$$ +X _ {\mathrm {m r}} = \mathcal {P} _ {\mathrm {m r}} \left(\mathcal {M} _ {\mathrm {m r}} (Z)\right) \otimes \mathcal {S} \left(\mathcal {M} _ {\mathrm {m r}} (Z)\right), \tag {1} +$$ + +where $\mathcal{M}_{\mathrm{mr}}(\cdot)$ refers to the mr-mapper operation. $S(\cdot)$ is the shared expert calculation. $\otimes$ denotes the element-wise product. $\mathcal{P}_{\mathrm{mr}}(\cdot)$ is the mr-specific expert calculation. The calculation method for the HD-specific feature $X_{\mathrm{hd}}$ follows the similar approach. + +# 3.3. Inter-task Feedback Mechanism + +It is imperative to investigate the interplay between the moment retrieval and highlight detection tasks for their significant correlation. However, existing methods overlook this aspect and employ two decoders for direct prediction. + +![](images/65981cfcc0b0602aa6a2ab62b305e89cea7e8cb62f8892edc4cdd6f90957a42c.jpg) +(a) Moment-guided feedback + +![](images/e16f690698437a1b2a93900d4c076fbda4288dc5896ed400196f377db6a77e3b.jpg) +(b) Highlightness-guided feedback +Figure 3. Illustration of the inter-task feedback mechanism. (a) moment-guided feedback manner. (b) highlightness-guided feedback manner. + +To resolve this issue, we propose an inter-task feedback mechanism, which contains two task-specific decoders (Transformer decoder for MR and lightweight decoder for HD) and two feedback manners (moment-guided feedback and highlightness-guided feedback). The output of one task is transformed into the mask to assist another task in the feedback mechanism. All components of the inter-task feedback mechanism is detailed as follows. + +Transformer Decoder for MR. Previous studies have demonstrated that it is effective to employ dynamic anchors as queries for the decoder within the DETR-like structure [26, 31]. They iteratively update anchor boxes, thereby transforming the procedure for updating queries into the cascaded soft ROI-pooling. In our work, the Transformer decoder for MR follows the work [31]. Each moment query is represented by its temporal center coordinate $q_{c}$ and width $q_{w}$ . The output of the Transformer decoder for MR is denoted as $Y_{\mathrm{mr}} \in \mathbb{R}^{N_q \times 2}$ . $N_{q}$ is the number of moment queries and set to 10 in our work. + +Lightweight Decoder for HD. The most straightforward method for the HD decoder would be utilizing one or more fully-connected layers, as seen in prior studies such as Moment-DETR [25] and UMT [32]. However, such design overlooks the diversity of video-query pairs [36], which provides identical criterias for the saliency score prediction of each video-query pair. Although QD-DETR [36] pro + +poses a global saliency token to predict the saliency scores, it remains highly coupled with the encoder. + +Different from those methods, we introduce CNN structures in the decoder. CNNs can capture local details, ensuring accurate saliency predictions across varying queries within the same video. The structure of the lightweight decoder for HD prediction is illustrated in Fig. 3, which alternates between Linear and Conv1d layers. We denote the input of the HD decoder as $\mathcal{J} \in \mathbb{R}^{N \times d}$ . The saliency prediction $Y_{\mathrm{hd}} \in \mathbb{R}^{N \times 1}$ can be computed as follows: + +$$ +Y _ {\mathrm {h d}} = D _ {\mathrm {H D}} (\mathcal {J}) + \frac {\mathbb {I} _ {i = 1} ^ {N} \left(\sum_ {k = 1} ^ {d} \mathcal {J} ^ {i k}\right)}{\sqrt {d}}, \tag {2} +$$ + +where $\mathbb{I}_{i = 1}^{N}(\cdot)$ represents the traversal operation. $D_{\mathrm{HD}}(\cdot)$ refers to the decoder network operation. + +Moment-guided Feedback. We propose a moment-guided feedback manner to investigate the influence of MR on HD. Our key focus is on how to effectively utilize the results of moment retrieval for highlight detection. The implementation procedure of moment-guided feedback is illustrated in Fig. 3 (a). + +We transform the tentatively predicted moment results from the "(center, width)" format to the "(start, end)" format. These moment results are fed into the Moment-to-Mask (M2M) converter to generate moment-aware masks $\xi_{\mathrm{m2m}}$ . We initialize a clip-wise vector with a length of $N$ for each moment prediction. The indices in these vectors correspond to the clip indexes within the video. We binarize these vectors based on the moment results. We can obtain the clip indexes that are included in the moments through the prior information and set the values at those positions in the vector to 1, while the rest to 0. The obtained $N_q$ binary vectors are then summed and normalized with L2 norm. + +The obtained moment-aware masks $\xi_{\mathrm{m2m}}$ integrate the prediction results from $N_{q}$ queries and provide guidance for the highlight detection. In the moment-guided feedback manner, the input $\mathcal{I}$ of the HD decoder is updated by $X_{\mathrm{hd}} + X_{\mathrm{hd}}*\xi_{\mathrm{m2m}}$ + +Highlightness-guided Feedback. Similar to the moment-guided feedback, we introduce a highlightness-guided feedback manner with the Highlightness-to-Mask (H2M) converter, to explore the influence of HD on MR. As illustrated in Fig. 3 (b), the saliency score vector is binarized in the H2M to obtain the highlightness-aware mask $\xi_{\mathrm{h2m}}$ . Values in the saliency score vector that are below the mean of the vector are set to 0, while those above are set to 1. Then $\xi_{\mathrm{h2m}}$ would provide guidance for moment retrieval. In this feedback manner, the input of the MR decoder is updated by $X_{\mathrm{mr}} + X_{\mathrm{mr}} * \xi_{\mathrm{h2m}}$ . + +# 3.4. Task-dependent Joint Loss + +A direct way to optimize the joint task is to sum the respective task-specific loss functions. However, it neglects + +the variations in magnitudes of different task losses, leading to dominance of one task. Existing methods [20, 25, 29, 32, 36] utilize manually-set weights for the weighted sum of task losses, which limit the task learning as tasks evolve at different rates. To address this problem, inspired by the studies using dynamic weights [10, 16, 22, 23, 30] based on task learning stages, we introduce the task-dependent joint loss as a more effective and flexible solution. + +MR Loss. For moment retrieval, we define the MR likelihood using a Gaussian distribution as shown in Eq. (3), with its mean determined by the output $f^{\theta_{\mathrm{mr}}}(x)$ of the neural network with $\theta_{\mathrm{mr}}$ . $x$ is the input of the neural network. + +$$ +p \left(Y _ {\mathrm {m r}} \mid f ^ {\theta_ {\mathrm {m r}}} (x)\right) = \mathcal {N} \left(f ^ {\theta_ {\mathrm {m r}}} (x), \delta_ {\mathrm {m r}}\right), \tag {3} +$$ + +where $\delta_{\mathrm{mr}}$ is a learnable parameter that quantifies the uncertainty of moment retrieval. The negative log-likelihood can be derived as follows: + +$$ +\begin{array}{l} - \log p \left(Y _ {\mathrm {m r}} \mid f ^ {\theta_ {\mathrm {m r}}} (x)\right) \\ \propto \frac {1}{2 \delta_ {\mathrm {m r}} {} ^ {2}} \left\| Y _ {\mathrm {m r}} - f ^ {\theta_ {\mathrm {m r}}} (x) \right\| ^ {2} + \log \delta_ {\mathrm {m r}}. \tag {4} \\ \end{array} +$$ + +$\left\| Y_{\mathrm{mr}} - f^{\theta_{\mathrm{mr}}}(x)\right\|^2$ measures the offset between the model's predicted value and the ground-truth value. Therefore, inspired by Eq. (4), the MR loss function $\mathcal{L}_{\mathrm{mr}}(\theta_{\mathrm{mr}},\delta_{\mathrm{mr}})$ can be defined as: + +$$ +\mathcal {L} _ {\mathrm {m r}} \left(\theta_ {\mathrm {m r}}, \delta_ {\mathrm {m r}}\right) = \frac {1}{2 \delta_ {\mathrm {m r}} ^ {2}} \mathcal {L} \left(\theta_ {\mathrm {m r}}\right) + \log \delta_ {\mathrm {m r}}. \tag {5} +$$ + +Following existing approaches [25, 32, 36], $\mathcal{L}(\theta_{\mathrm{mr}})$ consists of three components: the L1 loss $L_{\mathrm{L1}}$ , the generalized IoU loss [40] $L_{\mathrm{gIoU}}$ , and the cross-entropy loss $L_{\mathrm{BCE}}$ . $L_{\mathrm{L1}}$ and $L_{\mathrm{gIoU}}$ are employed to calculate the mean absolute error and gIoU deviation between ground-truth moments and predicted moments, respectively. $L_{\mathrm{BCE}}$ is used to classify whether the predicted moments belong to the foreground or background. In summary, $\mathcal{L}(\theta_{\mathrm{mr}}) = L_{\mathrm{L1}} + L_{\mathrm{gIoU}} + L_{\mathrm{BCE}}$ . + +HD Loss. For highlight detection, inspired by the Boltzmann distribution, the HD likelihood is modeled using a softmax function applied to the scaled model output similar to [23], as Eq. (6). The learnable scaling factor is denoted as $\delta_{\mathrm{hd}}$ . $f^{\theta_{\mathrm{hd}}}(x)$ is the output of the network with $\theta_{\mathrm{hd}}$ on input $x$ . + +$$ +p \left(Y _ {\mathrm {h d}} \mid f ^ {\theta_ {\mathrm {h d}}} (x), \delta_ {\mathrm {h d}}\right) = \operatorname {s o f t m a x} \left(\frac {1}{\delta_ {\mathrm {h d}} 2} f ^ {\theta_ {\mathrm {h d}}} (x)\right). \tag {6} +$$ + +Similar to Eq. (4), the negative log-likelihood of Eq. (6) + +can be calculated as follows: + +$$ +\begin{array}{l} - \log p \left(Y _ {\mathrm {h d}} = y \mid f ^ {\theta_ {\mathrm {h d}}} (x), \delta_ {\mathrm {h d}}\right) \\ = \frac {1}{\delta_ {\mathrm {h d}} ^ {2}} \log \frac {\sum_ {y ^ {\prime}} \exp \left(f _ {y ^ {\prime}} ^ {\theta_ {\mathrm {h d}}} (x)\right)}{\exp \left(f _ {y} ^ {\theta_ {\mathrm {h d}}} (x)\right)} \\ + \log \frac {\sum_ {y ^ {\prime}} \exp \left(\frac {1}{\delta_ {\mathrm {h d}} {} ^ {2}} f _ {y ^ {\prime}} ^ {\theta_ {\mathrm {h d}}} (x)\right)}{\left(\sum_ {y ^ {\prime}} \exp \left(f _ {y ^ {\prime}} ^ {\theta_ {\mathrm {h d}}} (x)\right)\right) ^ {\frac {1}{\delta_ {\mathrm {h d}} {} ^ {2}}}} \\ \approx - \frac {1}{\delta_ {\mathrm {h d}} ^ {2}} \log \operatorname {s o f t m a x} \left(Y _ {\mathrm {h d}}, f ^ {\theta_ {\mathrm {h d}}} (x)\right) + \log \delta_ {\mathrm {h d}}, \tag {7} \\ \end{array} +$$ + +where $f_{y}^{\theta_{\mathrm{hd}}}(x)$ refers to the y-th value of the $f^{\theta_{\mathrm{hd}}}(x)$ . We introduce a simplify assumption $\sum_{y'} \exp \left( \frac{1}{\delta_{\mathrm{hd}}^2} f_{y'}^{\theta_{\mathrm{hd}}}(x) \right) \approx \left( \sum_{y'} \exp \left( f_{y'}^{\theta_{\mathrm{hd}}}(x) \right) \right)^{\frac{1}{\delta_{\mathrm{hd}}^2}}$ when $\delta_{\mathrm{hd}} \to 1$ . + +- log softmax $\left(Y_{\mathrm{hd}},f^{\theta_{\mathrm{hd}}}(x)\right)$ of Eq. (7) represents the cross-entropy classification loss of $Y_{\mathrm{hd}}$ . Inspired on this term, we generalize this expression to other classification losses. The HD loss function $\mathcal{L}_{\mathrm{hd}}(\theta_{\mathrm{hd}},\delta_{\mathrm{hd}})$ can be formulated as: + +$$ +\mathcal {L} _ {\mathrm {h d}} \left(\theta_ {\mathrm {h d}}, \delta_ {\mathrm {h d}}\right) = \frac {1}{\delta_ {\mathrm {h d}} ^ {2}} \mathcal {L} \left(\theta_ {\mathrm {h d}}\right) + \log \delta_ {\mathrm {h d}}. \tag {8} +$$ + +Consistent with prior approaches [25, 36], $\mathcal{L}(\theta_{\mathrm{hd}})$ includes the hinge loss $L_{\mathrm{hinge}}$ , negative video-query pairs loss $L_{\mathrm{neg}}$ and rank-aware contrastive loss [17] $L_{\mathrm{cont}}$ , i.e. $\mathcal{L}(\theta_{\mathrm{hd}}) = L_{\mathrm{hinge}} + L_{\mathrm{neg}} + L_{\mathrm{cont}}$ . $L_{\mathrm{hinge}}$ is computed between two pairs of positive and negative clips, with its margin 0.2 to maintain consistency with [25] for fairness. $L_{\mathrm{neg}}$ and $L_{\mathrm{cont}}$ from [36] are used to reduce the saliency of negative pairs. + +We can obtain final task-dependent joint loss $\mathcal{L}_{\mathrm{joint}}$ through integrating Eq. (5) and Eq. (8): + +$$ +\mathcal {L} _ {\text {j o i n t}} = \mathcal {L} _ {\mathrm {m r}} \left(\theta_ {\mathrm {m r}}, \delta_ {\mathrm {m r}}\right) + \mathcal {L} _ {\mathrm {h d}} \left(\theta_ {\mathrm {h d}}, \delta_ {\mathrm {h d}}\right). \tag {9} +$$ + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. Extensive experiments are conducted on three benchmark datasets: QVHighlights [25], TVSum [43], and Charades-STA [14]. QVHighlights is currently the sole publicly dataset for joint moment retrieval and highlight detection tasks. It provides 10,310 queries associated with 18,367 moments, with an average of 1.8 disjoint moments per query. In contrast to other MR datasets with one-to-one query-moment mappings, QVHighlights aligns more closely with real-world scenarios. Each video in the dataset comprises 75 clips, each of which is 2s-long. + +We also utilize two task-specific datasets, Charades-STA [14] for MR and TVSum [43] for HD, to evaluate the model. + +
MethodVT VU GA MS PK PR FM BK BT DS Avg.
sLSTM [64]ECCV'1641.1 46.2 46.3 47.7 44.8 46.1 45.2 40.6 47.1 45.5 45.1
SG [34]CVPR'1742.3 47.2 47.5 48.9 45.6 47.3 46.4 41.7 48.3 46.6 46.2
LIM-S [51]CVPR'1955.9 42.9 61.2 54.0 60.4 47.5 43.2 66.3 69.1 62.6 56.3
Trailer [47]ECCV'2061.3 54.6 65.7 60.8 59.1 70.1 58.2 64.7 65.6 68.1 62.8
SL-Module [52]ICCV'2186.5 68.7 74.9 86.2 79.0 63.2 58.9 72.6 78.9 64.0 73.3
MINI-Net† [18]ECCV'2080.6 68.3 78.2 81.8 78.1 65.8 57.8 75.0 80.2 65.5 73.2
TCG† [57]ICCV'2185.0 71.4 81.9 78.6 80.2 75.5 71.6 77.3 78.6 68.1 76.8
Joint-VA† [3]ICCV'2183.7 57.3 78.5 86.1 80.1 69.2 70.0 73.0 97.4 67.5 76.3
UMT† [32]CVPR'2287.5 81.5 88.2 78.8 81.4 87.0 76.0 86.9 84.4 79.6 83.1
QD-DETR [36]CVPR'2388.2 87.4 85.6 85.0 85.8 86.9 76.4 91.3 89.2 73.7 85.0
UniVTG‡ [29]ICCV'2383.9 85.1 89.0 80.1 84.6 81.4 70.9 91.7 73.5 69.3 81.0
TaskWeave(Ours)88.2 90.8 93.3 87.5 87.0 82.0 80.9 92.9 89.5 81.2 87.3
+ +Table 1. Experimental results $(\%)$ on TVSum. $\dagger$ means including audio modality. $\ddagger$ means following the pretrain-finetune paradigm. + +
MethodMRHD
R1mAP≥ Very Good
@0.5@0.7@0.5@0.75Avg.mAP HIT@1
Moment-DETR [25]NIPS'2153.9434.84--32.2035.6555.55
UMT†[32]CVPR'2260.2644.26--38.5939.8564.19
QD-DETR [36]CVPR'2362.6846.6662.2341.8241.2239.1363.03
EaTR [20]ICCV'2361.3645.7961.8641.9141.7437.1558.65
UniVTG‡[29]ICCV'2359.74---36.1338.8361.81
TaskWeave(Ours)64.2650.0665.3946.4745.3839.2863.68
+ +Charades-STA provides 16,128 query-moment pairs. TVSum comprises videos from 10 domains, with each domain containing 5 videos. + +Evaluation Metrics. We utilize the same evaluation metrics used in prior approaches [25, 32, 36]. For QVHighlights, mean average precision (mAP) with different tIoU thresholds 0.5, 0.75, the average mAP over [0.5:0.05:0.95], and Recall@1 with tIoU 0.5, 0.7 are utilized for MR evaluation. mAP and HIT@1 are used for HD evaluation, where a clip is considered as a true positive when it achieves a "Very Good" [25] saliency score. For Charades-STA, we utilize Recall@1 with tIoU thresholds 0.5, 0.7. For TVSum, Top-5 mAP is adopted. + +Table 3. Experimental results (\%) on QVHighlights val split. † means including audio modality. ‡ means following the pretrain-finetune paradigm. + +
BackboneMethodR1@0.5R1@0.7
VGGSAP [9]AAAI'1927.4213.36
SM-RL [48]CVPR'1924.3611.17
2D-TAN [65]AAAI'2040.9422.85
FVMR [13]CVPR'2124.3611.17
UMT† [32]CVPR'2248.3129.25
QD-DETR [36]CVPR'2352.7731.13
TaskWeave(Ours)56.5133.66
I3DCTRL [14]ICCV'1723.638.89
MAN [60]CVPR'1946.5322.72
VSLNet [61]ACL'2047.3130.19
QD-DETR [36]CVPR'2350.6731.02
TaskWeave(Ours)53.3631.4
+ +Table 2. Experimental results $(\%)$ on Charades-STA test split. $\dagger$ means including audio modality. $\ddagger$ means following the pretrain-finetune paradigm. + +
Task-decoupled UnitInter-task feedbackJoint lossMRHD
R1 @0.7Avg. mAPmAP HIT@1
46.2641.038.9462.84
✓47.8743.2438.5861.81
✓✓49.2945.1238.9662.0
✓✓✓50.0645.3839.2863.68
+ +Table 4. The ablation results $(\%)$ of the components of our proposed method. + +# 4.2. Implementation Details + +Feature representations. For OVHighlights, we employ the pre-trained SlowFast [12] and CLIP [39] backbones to extract video features, following [25, 32, 36] for fairness. For Charedes-STA, we leverage the official release of VGG [42] and I3D [6] features as video embeddings. For TVSum, we extract video features by the I3D [6] pre-trained on Kinetics 400 [21]. For QVHighlights and TVSum, we use CLIP [39] to extract text features, while using GloVe [38] text embeddings for Charades-STA. + +Training settings. We leverage AdamW [24] optimizer with 1e-4 learning rate and 1e-4 weight decay. We train 200, 100, and 2000 epochs with batch size 32, 32 and 2 for QVHighlights, Charades-STA and TVSum, respectively. All Transformer layers follow the consistent configuration, including sinusoidal positional encodings, 8 at + +
IndexMR-expertHR expertMRHDR
R1 @0.7Avg. mAPmAPHIT@1
(a)Iden.Iden.44.7740.2338.3860.19
(b)Line.Line.47.5543.6538.3260.77
(c)CNNCNN49.2343.7338.8260.77
(d)Trans.Trans.48.3242.8538.6161.10
(e)Trans.Iden.47.142.138.761.68
(f)CNNIden.49.2945.1238.9662.0
+ +Table 5. Flexibility validation $(\%)$ of the taskdecoupled unit. "Iden:" identity mapping; "Trans":Transformer; "Line":linear layer. + +
TypeMRHD
R1 @0.7Avg. mAPmAP HIT@1
Sum47.6844.2738.760.32
Weighted Sum48.3243.2138.3161.35
Ours50.0645.3839.2863.68
+ +Table 6. Effectiveness justification $(\%)$ of the task-dependent joint loss. + +
TypeMRHD
R1 @0.7Avg. mAPmAPHIT@1
MR2HD50.0645.3838.7362.84
HD2MR48.6544.7839.2863.68
Bi-MRHD48.8445.138.3661.81
MR-HD50.045.338.9662.45
HD-MR49.144.8739.0359.94
+ +Table 7. Comparing various inter-task feedback combinations $(\%)$ on QVHighlights val split. + +tention heads, and a dropout rate of 0.1. Multi-modal fusion employs a 2-layer Transformer with cross attention, where video features serve as the query and text features serve as key and value. For Eq. (9), we define $\gamma_{\mathrm{mr}}, \gamma_{\mathrm{hd}}$ as $\log \delta_{\mathrm{mr}}^2, \log \delta_{\mathrm{hd}}^2$ , respectively. Therefore, Eq. (9) can be rewritten as $\mathcal{L}_{joint} = \exp(-\gamma_{\mathrm{mr}})\mathcal{L}(\theta_{\mathrm{mr}}) + 2\exp(-\gamma_{\mathrm{hd}})\mathcal{L}(\theta_{\mathrm{hd}}) + \gamma_{\mathrm{mr}} + \gamma_{\mathrm{hd}}$ . $\gamma_{\mathrm{mr}}$ and $\gamma_{\mathrm{hd}}$ are learnable parameters. They are initialized to 0. To further stabilize the training, the model EMA strategy has also undertaken. All experiments are conducted with Pytorch v1.13.1 on a single NVIDIA RTX 3090. + +# 4.3. Comparison with State-of-the-arts + +Results on TVSum. In Tab. 1, we compare our TaskWeave with existing state-of-the-arts (SOTA) methods, where the methods with $\dagger$ are incorporate with audio features. UniVTG [29] follows pretrain-then-finetune paradigm, therefore we use its non-pretrained results for a fair comparison. We observe that: i) our methods outperforms SOTA methods in 9 out of the 10 categories; ii) TaskWeave significantly outperforms all methods in terms of Avg. mAP, with a remarkable $2.71\%$ improvement over the previous SOTA method. + +Results on Charades-STA. In Tab. 2, we utilize different backbones to demonstrate the effectiveness of our TaskWeave. We compare our results with other methods for MR only (white background) and methods for joint MR and HD (gray background). It results in significant improvements, i.e., $+7.09\%$ in R1@0.5 and $+8.13\%$ in R1@0.7. + +Results on QVHighlights. As shown in Tab. 3, we present the performance comparisons with existing methods. We report the MR results with the MR-guided feedback, while HD results with the HD-guided feedback. We observe that: i) our TaskWeave surpasses SOTA methods by a large margin in all moment retrieval metrics, with a remarkable $8.72\%$ improvement in Avg. mAP; ii) our method achieves the second-best performance on the HD. However, + +It's worth noting that UMT [32] incorporates audio features, while our method does not. When compared to other methods without audio features, our method outperforms them all. In summary, these results validate the effectiveness of our task-driven framework. + +We provide visualization examples in Fig. 4 for the qualitative analysis among our TaskWeave, Moment-DETR [25], and QD-DETR [36]. Given diverse queries for a video, TaskWeave precisely localizes moments for queries and presents high IoU with the ground-truth. Moment-DETR misses some instances, and QD-DETR exhibits lower retrieval accuracy. Moreover, our method obtains higher saliency scores for relevant clips in response to the query. + +# 4.4. Ablation Studies and Discussions + +We present some ablation studies and discussions about TaskWeave, with all experiments are conducted on the QVHighlights val split [25]. + +Ablation of components. We conduct ablation studies on each component of TaskWeave, as shown in Tab. 4. Sequentially employing the task-decoupled unit and our proposed joint loss contributes a $5.5\%$ improvement in average mAP and a $6.5\%$ increase in R1@0.7 for MR. Utilizing all components, we observe that a significant $10.7\%$ improvement in average mAP of MR and a $1.3\%$ enhancement in HIT@1 of HD. These results indicate the effectiveness of the proposed components in out TaskWeave. + +Flexibility of the task-decoupled unit. To demonstrate the flexibility of the task-decoupled unit, we investigate the performance of applying different networks within various experts, as shown in Tab. 5. The shared expert utilize a fixed configuration with 2 Transformer layers. We provide 4 different methods for each task-specific expert, including the feed-forward network (one layer Linear), the identity mapping, CNN (composed of one depthwise convolution layer with the kernel/stride/padding of $5/1/2$ and 1D convolution), and the Transformer. Due to space constraints, we present + +![](images/0a4aa83bb45accdc89c01decbb04416267c07dffc42087a05418c66c8287babe.jpg) +Query1: A guy with grey top is showing a box filled with rubbish. Query2: A man shaves his beard looks in a mirror and talks to the camera. +Figure 4. Qualitative results on the QVHighlights for Ground-Truth, Moment-DETR, QD-DETR and our method. The predicted moments and saliency scores are illustrated through intervals and lines. + +results of 6 out of 16 combinations. The results in Tab. 5 are obtained by TaskWeave without inter-task feedback. + +From this table, we can observe several interesting facts. First, different task-specific experts have different performance. The need for experts is reflected in (a). Second, although Transformer has made significant progress in visual, it's not a panacea. Results in (d) are lower compared to (b) and (c), we believe that the high computational complexity of Transformer decreases the performance. Third, the task-specific expert should be designed based on the task objective. For instance, (c) and (f) perform well on the MR because they focus on local features, which is important for localizing moments. Finally, we also find evidence that MR and HD are highly related. Improvement in one task enhances the other (compare (c) and (f)), while a decline in one limits the other's performance (compare (d) and (e)). In this paper, our mr-specific expert utilizes CNN and hd-specific expert is implemented by identity mapping. + +Task-dependent joint loss. Tab. 6 shows the performance of TaskWeave with different losses. "Sum" and "Weighted Sum" refer to the fusion manner of loss for different tasks, respectively. Weights in "Weighted Sum" are consistent with the existing methods [25, 36] for fairness. Comparing "Sum" and "Weighted Sum", we find that "Weighted Sum" can better balance the performance of the two tasks. However, it is obvious that the proposed task-dependent joint loss is optimal. + +Inter-task interactions. We believe that each task requires learning before providing effective feedback. Therefore, in our inter-task feedback mechanism, the feedback process starts when the model is trained to half of the max epoch. For brevity, we write Moment/Highlightness-guided feedback as MgF/HgF. In Tab. 7, we explore five feedback manners as follows: "MR2HD" (with MgF only), + +"HD2MR" (with HgF only), "Bi-MRHD" (with MgF and HgF simultaneously), "MR-HD" (MgF first, then HgF), and "HD-MR" (HgF first, then MgF). + +We find that the inter-task feedback makes both tasks gain simultaneously. The performance of the model with "HD2MR" is slightly worse, this is because HD focuses on more refined moments than MR. Ground-truth annotations and the results of "MR-HD" and "HD-MR" also illustrate this fact. The gain brought by "Bi-MRHD" is limited, which we believe that it's a natural result of feedback not always being effective. In general, the inter-task feedback not only contributes to bring gains for both tasks but also helps to understand the characteristics of each task. + +# 5. Conclusion + +This paper proposes a novel task-driven paradigm for addressing joint moment retrieval and highlight detection. Different from existing data-driven methods, we utilize the task-decoupled unit to capture the task-specific and common features, respectively. We also explore different network architecture for moment retrieval and highlight detection. We design the inter-task feedback mechanism to in-depth investigate the interplay between both tasks. Different from prior methods, we introduce the principled joint loss to optimize the model. The effectiveness, flexibility, and superiority of the proposed method have been demonstrated on three benchmark datasets. + +# Acknowledgement + +This research was supported by the National Natural Science Foundation of China (No. U23B2060, 62088102), the Youth Innovation Team of Shaanxi Universities, and the Fundamental Research Funds for the Central Universities. + +# References + +[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2 +[2] Niki Parmar Jakob Uszkoreit Llion Jones Aidan N. Gomez Lukasz Kaiser Ashish Vaswani, Noam Shazeer and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5999-6009, 2017. 2, 3 +[3] Taiwanbat Badamdorj, Mrigank Rochan, Yang Wang, and Li Cheng. Joint visual and audio learning for video highlight detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8127-8137, 2021. 6 +[4] Taiwanbat Badamdorj, Mrigank Rochan, Yang Wang, and Li Cheng. Contrastive learning for unsupervised video highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14042-14052, 2022. 1, 2 +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 2 +[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6 +[7] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 2 +[8] Jingyuan Chen, Lin Ma, Xinpeng Chen, Zequn Jie, and Jiebo Luo. Localizing natural language in videos. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8175-8182, 2019. 2 +[9] Shaoxiang Chen and Yu-Gang Jiang. Semantic proposal for activity localization in videos via sentence query. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8199-8206, 2019. 6 +[10] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018. 5 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[12] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 6 +[13] Junyu Gao and Changsheng Xu. Fast video moment retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1523-1532, 2021. 6 + +[14] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 2, 5, 6 +[15] Ana Garcia del Molino and Michael Gygli. PhD-gifs: personalized highlight detection for automatic gif creation. In Proceedings of the 26th ACM international conference on Multimedia, pages 600-608, 2018. 2 +[16] Michelle Guo, Albert Haque, De-An Huang, Serena Yeung, and Li Fei-Fei. Dynamic task prioritization for multitask learning. In Proceedings of the European conference on computer vision (ECCV), pages 270–287, 2018. 5 +[17] David T Hoffmann, Nadine Behrmann, Juergen Gall, Thomas Brox, and Mehdi Noroozi. Ranking info noise contrastive estimation: Boosting contrastive learning via ranked positives. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 897–905, 2022. 5 +[18] Fa-Ting Hong, Xuanteng Huang, Wei-Hong Li, and Wei-Shi Zheng. Mini-net: Multiple instance ranking network for video highlight detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 345-360. Springer, 2020. 6 +[19] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 2 +[20] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13846-13856, 2023. 1, 2, 5, 6 +[21] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6 +[22] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 5 +[23] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482-7491, 2018. 5 +[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[25] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34: 11846-11858, 2021. 1, 2, 4, 5, 6, 7, 8 +[26] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 2, 4 +[27] Huan Li, Ping Wei, Jiapeng Li, Zeyu Ma, Jiahui Shang, and Nanning Zheng. Asymmetric relation consistency reasoning + +for video relation grounding. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 2 +[28] Huan Li, Ping Wei, Zeyu Ma, and Nanning Zheng. Inverse compositional learning for weakly-supervised relation grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15477-15487, 2023. 2 +[29] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified videolanguage temporal grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2794-2804, 2023. 1, 2, 5, 6, 7 +[30] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1871-1880, 2019. 5 +[31] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2, 4 +[32] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3042-3051, 2022. 1, 2, 4, 5, 6, 7 +[33] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2 +[34] Behrooz Mahasseni, Michael Lam, and Sinisa Todorovic. Unsupervised video summarization with adversarial LSTM networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 202-211, 2017. 6 +[35] Otniel-Bogdan Mercea, Lukas Riesch, A Koepke, and Zeynep Akata. Audio-visual generalised zero-shot learning with cross-modal attention and language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10553–10563, 2022. 2 +[36] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23023-23033, 2023. 1, 2, 4, 5, 6, 7, 8 +[37] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810-10819, 2020. 2 +[38] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 6 + +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6 +[40] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5 +[41] Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, and Dacheng Tao. Tridet: Temporal action detection with relative boundary modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18857-18866, 2023. 3 +[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6 +[43] Yale Song, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaimes. Tvsum: Summarizing web videos using titles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5179-5187, 2015. 2, 5 +[44] Min Sun, Ali Farhadi, and Steve Seitz. Ranking domain-specific highlights by analyzing edited videos. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13, pages 787-802. Springer, 2014. 1, 2 +[45] Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. Progressive layered extraction (ple): A novel multi-task learning (mtl) model for personalized recommendations. In Proceedings of the 14th ACM Conference on Recommender Systems, pages 269-278, 2020. 2 +[46] Tuan N Tang, Kwonyoung Kim, and Kwanghoon Sohn. Temporalmaxer: Maximize temporal context with only max pooling for temporal action localization. arXiv preprint arXiv:2303.09055, 2023. 3 +[47] Lezi Wang, Dong Liu, Rohit Puri, and Dimitris N Metaxas. Learning trailer moments in full-length movies with co-contrastive attention. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVIII 16, pages 300–316. Springer, 2020. 6 +[48] Weining Wang, Yan Huang, and Liang Wang. Language-driven temporal activity localization: A semantic matching reinforcement learning model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 334-343, 2019. 6 +[49] Fanyue Wei, Biao Wang, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Learning pixel-level distinctions for video highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3073-3082, 2022. 2 +[50] Xi Wei, Tianzhu Zhang, Yan Li, Yongdong Zhang, and Feng Wu. Multi-modality cross attention network for image and sentence matching. In Proceedings of the IEEE/CVF con + +ference on computer vision and pattern recognition, pages 10941-10950, 2020. 2 +[51] Bo Xiong, Yannis Kalantidis, Deepi Ghadiyaram, and Kristen Grauman. Less is more: Learning highlight detection from video duration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1258-1267, 2019. 6 +[52] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7970-7979, 2021. 6 +[53] Yifang Xu, Yunzhuo Sun, Yang Li, Yilei Shi, Xiaoxiang Zhu, and Sidan Du. Mh-detr: Video moment and highlight detection with cross-modal transformer. arXiv preprint arXiv:2305.00355, 2023. 1 +[54] Jin Yang, Ping Wei, Ziyang Ren, and Nanning Zheng. Gated multi-scale transformer for temporal action localization. IEEE Transactions on Multimedia, 2023. 2 +[55] Jin Yang, Ping Wei, and Nanning Zheng. Cross time-frequency transformer for temporal action localization. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2 +[56] Ting Yao, Tao Mei, and Yong Rui. Highlight detection with pairwise deep ranking for first-person video summarization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 982–990, 2016. 1, 2 +[57] Qinghao Ye, Xiyue Shen, Yuan Gao, Zirui Wang, Qi Bi, Ping Li, and Guang Yang. Temporal cue guided video highlight detection with low-rank audio-visual fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7950–7959, 2021. 6 +[58] Yitian Yuan, Tao Mei, and Wenwu Zhu. To find where you talk: Temporal sentence localization in video with attention based location regression. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9159–9166, 2019. 2 +[59] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 2 +[60] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1247-1257, 2019. 1, 2, 6 +[61] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6543-6554, 2020. 2, 6 +[62] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Temporal sentence grounding in videos: A survey and future directions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10443-10465, 2023. 1 + +[63] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Temporal sentence grounding in videos: a survey and future directions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1 +[64] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. Video summarization with long short-term memory. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 766-782. Springer, 2016. 6 +[65] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12870-12877, 2020. 1, 2, 6 \ No newline at end of file diff --git a/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/images.zip b/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..03bf7dfb63af6ae3ad299e9dd58bd2a302015de7 --- /dev/null +++ b/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a57e2057df7b81f03a10da8bf4aec3ef765888f257ac8683cb08bac899f3b3dd +size 614422 diff --git a/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/layout.json b/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..44a58006bc2ccb03e1224703b78dc6bb7ed6d027 --- /dev/null +++ b/2024/Task-Driven Exploration_ Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection/layout.json @@ -0,0 +1,9208 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 102, + 521, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 102, + 521, + 140 + ], + "spans": [ + { + "bbox": [ + 74, + 102, + 521, + 140 + ], + "type": "text", + "content": "Task-Driven Exploration: Decoupling and Inter-Task Feedback for Joint Moment Retrieval and Highlight Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 161, + 484, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 161, + 484, + 217 + ], + "spans": [ + { + "bbox": [ + 108, + 161, + 484, + 217 + ], + "type": "text", + "content": "Jin Yang, Ping Wei*, Huan Li, Ziyang Ren \nNational Key Laboratory of Human-Machine Hybrid Augmented Intelligence Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University \n{jin.yang, lh875056558, rzyrzy}@stu.xjtu.edu.cn, pingwei@xjtu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 270, + 290, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 290, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 290, + 521 + ], + "type": "text", + "content": "Video moment retrieval and highlight detection are two highly valuable tasks in video understanding, but until recently they have been jointly studied. Although existing studies have made impressive advancement recently, they predominantly follow the data-driven bottom-up paradigm. Such paradigm overlooks task-specific and inter-task effects, resulting in poor model performance. In this paper, we propose a novel task-driven top-down framework TaskWeave for joint moment retrieval and highlight detection. The framework introduces a task-decoupled unit to capture task-specific and common representations. To investigate the interplay between the two tasks, we propose an inter-task feedback mechanism, which transforms the results of one task as guiding masks to assist the other task. Different from existing methods, we present a task-dependent joint loss function to optimize the model. Comprehensive experiments and in-depth ablation studies on QVHighlights, TVSum, and CharadesSTA datasets corroborate the effectiveness and flexibility of the proposed framework. Codes are available at github.com/EdenGabriel/TaskWeave." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 544, + 128, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 128, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 128, + 556 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 563, + 287, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 671 + ], + "type": "text", + "content": "As videos are prevailing in a wide range of applications, the diversity and massive scales of video content have posed unprecedented challenges in finding relevant moments based on user queries. To this end, the moment retrieval (MR) [60, 63, 65] and highlight detection (HD) [4, 56] tasks have emerged recently. MR aims to retrieve video moments that are relevant to the given query [14, 62]. HD aims to predict the clip-level saliency scores in the video [44]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 671, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 671, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 671, + 287, + 696 + ], + "type": "text", + "content": "As MR and HD tasks are closely related, they have been jointly addressed and achieved breakthroughs recently" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 246, + 547, + 365 + ], + "blocks": [ + { + "bbox": [ + 308, + 246, + 547, + 365 + ], + "lines": [ + { + "bbox": [ + 308, + 246, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 246, + 547, + 365 + ], + "type": "image", + "image_path": "66e4867e1d41709697bfc3ea6505e5271921a04678a42e0b1c7df877dbbf6dce.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 375, + 547, + 409 + ], + "lines": [ + { + "bbox": [ + 305, + 375, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 305, + 375, + 547, + 409 + ], + "type": "text", + "content": "Figure 1. Although the positive temporal intervals of moment retrieval and highlight detection exhibit high overlap, they pursue different objectives." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 433, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 547, + 590 + ], + "type": "text", + "content": "[20, 25, 29, 32, 36, 53]. The existing joint approaches in general utilize a shared backbone to learn the multimodal features as the common representations for MR and HD. Then a MR prediction head is employed for moments localization and a HD prediction head predicts saliency scores. These methods adhere to the bottom-up, data-driven paradigm, i.e. they capture common features from the input data and then utilize the features for different tasks. The effectiveness of these methods is built upon the premise of the high correlation between MR and HD. As illustrated in Fig. 1, the task MR and HD share the identical query and video inputs, additionally exhibiting a substantial temporal overlap between their respective positive temporal intervals." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 591, + 548, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 548, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 548, + 687 + ], + "type": "text", + "content": "However, the bottom-up, data-driven paradigm tends to excessively rely on the common features, but overlooks the inherent specific characteristics of MR and HD. This tendency might simplify the joint modeling as a problem of feature fusion, without considering the interplay between the two tasks. The distinct objectives pursued by MR and HD rely on distinct task-specific characteristics. Unfortunately, existing methods overlook the specificity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "To address the aforementioned issues, we believe it is essential to leverage the fundamental multi-modal data to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18308" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": "mine commonalities across tasks (bottom-up), while also strengthen the awareness of task-specific characteristics (top-down). To this end, we propose a novel paradigm TaskWeave from a task-driven perspective. The key idea is to jointly address the tasks MR and HD by considering the commonality, specificity, and interplay of MR and HD." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 286, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 286, + 274 + ], + "type": "text", + "content": "To effectively capture the commonality and specificity, we design a task-decoupled unit, i.e. a shared expert to capture common features and two task-specific experts to acquire the distinct characteristics. In order to investigate the interplay between MR and HD in-depth, we design an inter-task feedback mechanism. It converts the predictions of MR/HD into mask information, which are fed back to the input of the HD/MR prediction head. Furthermore, we introduce a principled task-dependent joint loss in which the task-specific weights are dynamically adjusted, rather than manually tuned." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 275, + 286, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 286, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 286, + 346 + ], + "type": "text", + "content": "We conduct experiments on the QVHighlights [25] dataset to validate the effectiveness of the proposed method. Moreover, we also conduct experiments for two individual tasks on the datasets Charades-STA [14] (moment retrieval) and TVSUM [43] (highlight detection). The proposed approach outperforms the existing methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 347, + 263, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 347, + 263, + 359 + ], + "spans": [ + { + "bbox": [ + 58, + 347, + 263, + 359 + ], + "type": "text", + "content": "The key contributions of this paper are three folds." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 360, + 286, + 430 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 360, + 286, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 286, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 286, + 382 + ], + "type": "text", + "content": "1. It proposes a novel task-driven, top-down framework for joint moment retrieval and highlight detection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 384, + 286, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 384, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 384, + 286, + 407 + ], + "type": "text", + "content": "2. It introduces a task-decoupled unit, an inter-task feedback mechanism, and a principled task-dependent loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 407, + 286, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 407, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 54, + 407, + 286, + 430 + ], + "type": "text", + "content": "3. It achieves state-of-the-art performance on three datasets. The ablation study validates the methods." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 442, + 133, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 442, + 133, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 133, + 453 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 463, + 246, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 463, + 246, + 474 + ], + "spans": [ + { + "bbox": [ + 59, + 463, + 246, + 474 + ], + "type": "text", + "content": "We review the related work from four aspects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 475, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 286, + 581 + ], + "type": "text", + "content": "Moment Retrieval. Existing approaches for MR mainly include two groups. One group follows a two-stage procedure [1, 14, 60, 65], which involves generating candidate temporal intervals and ranking them based on the correlation with the query. The other group directly regresses the temporal interval based on the aligned visual-text features [8, 37, 58, 59, 61]. Moreover, most datasets provide only one moment annotation for each video-query pair [1, 14], which does not align with real-world scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 582, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 664 + ], + "type": "text", + "content": "Highlight Detection. The saliency score in the highlight detection represents the relevance of a video clip to the given query. Most prior highlight detection benchmark datasets are query-agnostic [15, 44]. The saliency scores for video clips remained constant regardless of the query. As a result, some previous approaches treated HD as a solely visual task [4, 44, 49, 56]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "MR and HD tasks are traditionally studied separately. They have been jointly addressed recently as the introduction of QVHighlights dataset [25]. QVHighlights provides multiple moment-annotations for each query and ensures" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "these moments are uniformly distributed throughout the video. It also provides query-dependent highlightness annotations. With QVHighlights, the model Moment-DETR [25] is proposed for joint MR and HD. Following Moment-DETR, a growing number of approaches have been proposed to accomplish the joint task [20, 29, 32, 36]. However, these methods adopt the data-driven and bottom-up paradigm. Different from them, we propose a novel task-driven and top-down paradigm." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 180, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 335 + ], + "type": "text", + "content": "Vision Transformers. Transformer-based models [2] have brought huge achievement in both the image and video related domains [5, 11, 27, 28, 54, 55]. One of the most well-known methods is DETR [5], which regards object detection as a set prediction problem. Its end-to-end prediction procedure eliminates the intermediate or post-processing steps. On the other hand, some studies have employed cross-attention mechanism [35, 36, 50] to inject multi-modal data into the Transformer architecture. In this paper, we also adopt a DETR-like architecture [26, 31]. However, different from those methods, we incorporate diverse network architectures for feature extraction. Our focus lies in the collaboration of the network architectures." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 337, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 545, + 563 + ], + "type": "text", + "content": "Multi-task Learning. Multi-task learning (MTL) aims to train a single model for multiple tasks [7]. The most straightforward approach is to utilize a shared backbone to extract common features, which relies on the data-driven manner. However, it often leads to suboptimal performance for unrelated tasks and lacks flexibility. In response, some studies introduce MOE [19] and MMOE [33], which utilize a set of the shared experts in place of the shared bottom layer. To mitigate the seesaw phenomenon of multi-task learning, PLE [45] employs separate experts for each task while still retaining the shared experts. From the perspective of MTL, existing approaches for joint MR and HD follow the shared bottom paradigm. Moreover, directly applying MTL methods to joint MR and HD does not yield favorable results. Existing MTL methods are complex, therefore direct usage of them leads to a sharp increase in model complexity. This might decrease the performance. To this end, we aim to design an effective task-driven framework to jointly address MR and HD." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 578, + 388, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 578, + 388, + 590 + ], + "spans": [ + { + "bbox": [ + 306, + 578, + 388, + 590 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 597, + 373, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 373, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 373, + 609 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 616, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 688 + ], + "type": "text", + "content": "Given a text query with " + }, + { + "bbox": [ + 304, + 616, + 545, + 688 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 616, + 545, + 688 + ], + "type": "text", + "content": " words and an untrimmed video composed of " + }, + { + "bbox": [ + 304, + 616, + 545, + 688 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 616, + 545, + 688 + ], + "type": "text", + "content": " clips, the objective of the joint moment retrieval (MR) and highlight detection (HD) is to localize the center coordinate " + }, + { + "bbox": [ + 304, + 616, + 545, + 688 + ], + "type": "inline_equation", + "content": "q_{c}" + }, + { + "bbox": [ + 304, + 616, + 545, + 688 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 304, + 616, + 545, + 688 + ], + "type": "inline_equation", + "content": "q_{w}" + }, + { + "bbox": [ + 304, + 616, + 545, + 688 + ], + "type": "text", + "content": " of temporal intervals that are relevant to the text query, in addition to ranking clip-wise saliency scores." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Architecture Overview. Given the high correlation between moment retrieval and highlight detection tasks, the" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18309" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 73, + 547, + 237 + ], + "blocks": [ + { + "bbox": [ + 50, + 73, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 50, + 73, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 50, + 73, + 547, + 237 + ], + "type": "image", + "image_path": "a101a6befdf96406fa8311a1c3ba8d48453ca0993b6bbf611c746faf88eff16d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 247, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 547, + 293 + ], + "type": "text", + "content": "Figure 2. The overall pipeline of the proposed task-driven model TaskWeave. We propose the task-decoupled unit to capture task-specific and common features. Various experts can adopt different network implementations, showcasing the flexibility of the model. Inter-task feedback mechanism is designed to investigate the influence between both tasks. There are two feedback manners: Moment-guided and Highlightness-guided feedback. The principled task-dependent joint loss is introduced for jointly optimize the model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 313, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 287, + 409 + ], + "type": "text", + "content": "most intuitive approach is using a shared backbone in conjunction with two task-specific prediction heads. It is a data-driven paradigm that is commonly employed by previous methods, for its simplicity and ease of implementation. However, the task-specific characteristics are inherently present, since MR and HD have distinct objectives. Additionally, the interplay between MR and HD should also be considered, which can enhance the model's performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 410, + 288, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 288, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 288, + 625 + ], + "type": "text", + "content": "The overall pipeline of our approach is illustrated in Fig. 2. We employed the frozen video/text-encoder backbones to extract the video/text features, while ensuring their dimensions remained consistent at " + }, + { + "bbox": [ + 46, + 410, + 288, + 625 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 410, + 288, + 625 + ], + "type": "text", + "content": " by the projection. These features are utilized for multi-modal fusion through methods such as cross-attention or concatenation, resulting in query-related video representations " + }, + { + "bbox": [ + 46, + 410, + 288, + 625 + ], + "type": "inline_equation", + "content": "Z \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 46, + 410, + 288, + 625 + ], + "type": "text", + "content": ". In our paper, we extract these representations through the cross-attention. After cross-attention, inspired by [41, 46], a 1D Max Pooling (MaxPool1d) with the kernel size 5, stride 1 and padding 2 is utilized to eliminate the rank loss problem in the attention mechanism. These representations are fed into the task-decoupled unit to capture the task-related features " + }, + { + "bbox": [ + 46, + 410, + 288, + 625 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{mr}} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 46, + 410, + 288, + 625 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 410, + 288, + 625 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{hd}} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 46, + 410, + 288, + 625 + ], + "type": "text", + "content": ". Then, we employ the task-specific decoders with inter-task feedback mechanism to make predictions for moments localization and clip-wise saliency scores. We introduce the principled task-dependent loss to jointly optimize the model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 634, + 167, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 167, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 167, + 647 + ], + "type": "text", + "content": "3.2. Task-decoupled Unit" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Since moment retrieval and highlight detection have distinct objectives, the specificity of each task should be considered, rather than solely focusing on their commonalities. For this purpose, we propose a task-decoupled unit from a task-driven perspective to capture the task-related features," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 312, + 544, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 544, + 324 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 544, + 324 + ], + "type": "text", + "content": "which involves task-specific features and common features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 325, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 421 + ], + "type": "text", + "content": "The task-decoupled unit is depicted in Fig. 2. Inspired by the attention mechanism [2], the query-related video representation " + }, + { + "bbox": [ + 304, + 325, + 545, + 421 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 325, + 545, + 421 + ], + "type": "text", + "content": " is initially fed into two task-specific mappers, mr-mapper and hd-mapper. Each mapper is implemented using one-layer feed-forward network. The output of each task-specific mapper is directed into both the respective task-specific expert network (MR-specific Expert or HD-specific Expert) and the Shared Expert network." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 422, + 546, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 546, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 546, + 529 + ], + "type": "text", + "content": "Thanks to the design of our task-decoupled framework, each expert can employ various networks, such as convolutional network, Transformer, and feed-forward network. It offers more flexibility in addressing the joint MR and HD tasks. In our study, we also investigate the influence of different expert networks, referring to Sec. 4 for more details. Consequently, we can capture the task-related features " + }, + { + "bbox": [ + 304, + 422, + 546, + 529 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{mr}}" + }, + { + "bbox": [ + 304, + 422, + 546, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 422, + 546, + 529 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{hd}}" + }, + { + "bbox": [ + 304, + 422, + 546, + 529 + ], + "type": "text", + "content": " via element-wise product between the output of the task-specific expert and that of the shared expert." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 530, + 523, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 530, + 523, + 542 + ], + "spans": [ + { + "bbox": [ + 318, + 530, + 523, + 542 + ], + "type": "text", + "content": "The MR-specific feature " + }, + { + "bbox": [ + 318, + 530, + 523, + 542 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{mr}}" + }, + { + "bbox": [ + 318, + 530, + 523, + 542 + ], + "type": "text", + "content": " can be computed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 339, + 553, + 545, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 553, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 339, + 553, + 545, + 567 + ], + "type": "interline_equation", + "content": "X _ {\\mathrm {m r}} = \\mathcal {P} _ {\\mathrm {m r}} \\left(\\mathcal {M} _ {\\mathrm {m r}} (Z)\\right) \\otimes \\mathcal {S} \\left(\\mathcal {M} _ {\\mathrm {m r}} (Z)\\right), \\tag {1}", + "image_path": "0934d72d5853e4b9863c7e7c31b68d24acf7b2dbaddf8c62dc19af1b287248de.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 577, + 545, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 577, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 577, + 545, + 637 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 577, + 545, + 637 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathrm{mr}}(\\cdot)" + }, + { + "bbox": [ + 304, + 577, + 545, + 637 + ], + "type": "text", + "content": " refers to the mr-mapper operation. " + }, + { + "bbox": [ + 304, + 577, + 545, + 637 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 304, + 577, + 545, + 637 + ], + "type": "text", + "content": " is the shared expert calculation. " + }, + { + "bbox": [ + 304, + 577, + 545, + 637 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 304, + 577, + 545, + 637 + ], + "type": "text", + "content": " denotes the element-wise product. " + }, + { + "bbox": [ + 304, + 577, + 545, + 637 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{mr}}(\\cdot)" + }, + { + "bbox": [ + 304, + 577, + 545, + 637 + ], + "type": "text", + "content": " is the mr-specific expert calculation. The calculation method for the HD-specific feature " + }, + { + "bbox": [ + 304, + 577, + 545, + 637 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{hd}}" + }, + { + "bbox": [ + 304, + 577, + 545, + 637 + ], + "type": "text", + "content": " follows the similar approach." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 647, + 479, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 479, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 479, + 658 + ], + "type": "text", + "content": "3.3. Inter-task Feedback Mechanism" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "It is imperative to investigate the interplay between the moment retrieval and highlight detection tasks for their significant correlation. However, existing methods overlook this aspect and employ two decoders for direct prediction." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18310" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 74, + 286, + 173 + ], + "blocks": [ + { + "bbox": [ + 47, + 74, + 286, + 173 + ], + "lines": [ + { + "bbox": [ + 47, + 74, + 286, + 173 + ], + "spans": [ + { + "bbox": [ + 47, + 74, + 286, + 173 + ], + "type": "image", + "image_path": "65981cfcc0b0602aa6a2ab62b305e89cea7e8cb62f8892edc4cdd6f90957a42c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 175, + 219, + 185 + ], + "lines": [ + { + "bbox": [ + 128, + 175, + 219, + 185 + ], + "spans": [ + { + "bbox": [ + 128, + 175, + 219, + 185 + ], + "type": "text", + "content": "(a) Moment-guided feedback" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 194, + 286, + 323 + ], + "blocks": [ + { + "bbox": [ + 50, + 194, + 286, + 323 + ], + "lines": [ + { + "bbox": [ + 50, + 194, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 50, + 194, + 286, + 323 + ], + "type": "image", + "image_path": "e16f690698437a1b2a93900d4c076fbda4288dc5896ed400196f377db6a77e3b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 327, + 228, + 336 + ], + "lines": [ + { + "bbox": [ + 121, + 327, + 228, + 336 + ], + "spans": [ + { + "bbox": [ + 121, + 327, + 228, + 336 + ], + "type": "text", + "content": "(b) Highlightness-guided feedback" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 347, + 287, + 379 + ], + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 379 + ], + "type": "text", + "content": "Figure 3. Illustration of the inter-task feedback mechanism. (a) moment-guided feedback manner. (b) highlightness-guided feedback manner." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 402, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 496 + ], + "type": "text", + "content": "To resolve this issue, we propose an inter-task feedback mechanism, which contains two task-specific decoders (Transformer decoder for MR and lightweight decoder for HD) and two feedback manners (moment-guided feedback and highlightness-guided feedback). The output of one task is transformed into the mask to assist another task in the feedback mechanism. All components of the inter-task feedback mechanism is detailed as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": "Transformer Decoder for MR. Previous studies have demonstrated that it is effective to employ dynamic anchors as queries for the decoder within the DETR-like structure [26, 31]. They iteratively update anchor boxes, thereby transforming the procedure for updating queries into the cascaded soft ROI-pooling. In our work, the Transformer decoder for MR follows the work [31]. Each moment query is represented by its temporal center coordinate " + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "inline_equation", + "content": "q_{c}" + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "inline_equation", + "content": "q_{w}" + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": ". The output of the Transformer decoder for MR is denoted as " + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "inline_equation", + "content": "Y_{\\mathrm{mr}} \\in \\mathbb{R}^{N_q \\times 2}" + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "inline_equation", + "content": "N_{q}" + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": " is the number of moment queries and set to 10 in our work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Lightweight Decoder for HD. The most straightforward method for the HD decoder would be utilizing one or more fully-connected layers, as seen in prior studies such as Moment-DETR [25] and UMT [32]. However, such design overlooks the diversity of video-query pairs [36], which provides identical criterias for the saliency score prediction of each video-query pair. Although QD-DETR [36] pro" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": "poses a global saliency token to predict the saliency scores, it remains highly coupled with the encoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 95, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 545, + 191 + ], + "type": "text", + "content": "Different from those methods, we introduce CNN structures in the decoder. CNNs can capture local details, ensuring accurate saliency predictions across varying queries within the same video. The structure of the lightweight decoder for HD prediction is illustrated in Fig. 3, which alternates between Linear and Conv1d layers. We denote the input of the HD decoder as " + }, + { + "bbox": [ + 304, + 95, + 545, + 191 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 304, + 95, + 545, + 191 + ], + "type": "text", + "content": ". The saliency prediction " + }, + { + "bbox": [ + 304, + 95, + 545, + 191 + ], + "type": "inline_equation", + "content": "Y_{\\mathrm{hd}} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 304, + 95, + 545, + 191 + ], + "type": "text", + "content": " can be computed as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 344, + 196, + 545, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 196, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 344, + 196, + 545, + 230 + ], + "type": "interline_equation", + "content": "Y _ {\\mathrm {h d}} = D _ {\\mathrm {H D}} (\\mathcal {J}) + \\frac {\\mathbb {I} _ {i = 1} ^ {N} \\left(\\sum_ {k = 1} ^ {d} \\mathcal {J} ^ {i k}\\right)}{\\sqrt {d}}, \\tag {2}", + "image_path": "52c56b56d06c06b33e4cd6f984a021312e889eb23d3163afd752d1d0cbfa7093.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 234, + 544, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 234, + 544, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 234, + 544, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 234, + 544, + 258 + ], + "type": "inline_equation", + "content": "\\mathbb{I}_{i = 1}^{N}(\\cdot)" + }, + { + "bbox": [ + 305, + 234, + 544, + 258 + ], + "type": "text", + "content": " represents the traversal operation. " + }, + { + "bbox": [ + 305, + 234, + 544, + 258 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{HD}}(\\cdot)" + }, + { + "bbox": [ + 305, + 234, + 544, + 258 + ], + "type": "text", + "content": " refers to the decoder network operation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 259, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 545, + 330 + ], + "type": "text", + "content": "Moment-guided Feedback. We propose a moment-guided feedback manner to investigate the influence of MR on HD. Our key focus is on how to effectively utilize the results of moment retrieval for highlight detection. The implementation procedure of moment-guided feedback is illustrated in Fig. 3 (a)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 331, + 546, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 546, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 546, + 473 + ], + "type": "text", + "content": "We transform the tentatively predicted moment results from the \"(center, width)\" format to the \"(start, end)\" format. These moment results are fed into the Moment-to-Mask (M2M) converter to generate moment-aware masks " + }, + { + "bbox": [ + 304, + 331, + 546, + 473 + ], + "type": "inline_equation", + "content": "\\xi_{\\mathrm{m2m}}" + }, + { + "bbox": [ + 304, + 331, + 546, + 473 + ], + "type": "text", + "content": ". We initialize a clip-wise vector with a length of " + }, + { + "bbox": [ + 304, + 331, + 546, + 473 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 331, + 546, + 473 + ], + "type": "text", + "content": " for each moment prediction. The indices in these vectors correspond to the clip indexes within the video. We binarize these vectors based on the moment results. We can obtain the clip indexes that are included in the moments through the prior information and set the values at those positions in the vector to 1, while the rest to 0. The obtained " + }, + { + "bbox": [ + 304, + 331, + 546, + 473 + ], + "type": "inline_equation", + "content": "N_q" + }, + { + "bbox": [ + 304, + 331, + 546, + 473 + ], + "type": "text", + "content": " binary vectors are then summed and normalized with L2 norm." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "text", + "content": "The obtained moment-aware masks " + }, + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\xi_{\\mathrm{m2m}}" + }, + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "text", + "content": " integrate the prediction results from " + }, + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "inline_equation", + "content": "N_{q}" + }, + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "text", + "content": " queries and provide guidance for the highlight detection. In the moment-guided feedback manner, the input " + }, + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "text", + "content": " of the HD decoder is updated by " + }, + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{hd}} + X_{\\mathrm{hd}}*\\xi_{\\mathrm{m2m}}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 533, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 666 + ], + "type": "text", + "content": "Highlightness-guided Feedback. Similar to the moment-guided feedback, we introduce a highlightness-guided feedback manner with the Highlightness-to-Mask (H2M) converter, to explore the influence of HD on MR. As illustrated in Fig. 3 (b), the saliency score vector is binarized in the H2M to obtain the highlightness-aware mask " + }, + { + "bbox": [ + 304, + 533, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\xi_{\\mathrm{h2m}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 666 + ], + "type": "text", + "content": ". Values in the saliency score vector that are below the mean of the vector are set to 0, while those above are set to 1. Then " + }, + { + "bbox": [ + 304, + 533, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\xi_{\\mathrm{h2m}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 666 + ], + "type": "text", + "content": " would provide guidance for moment retrieval. In this feedback manner, the input of the MR decoder is updated by " + }, + { + "bbox": [ + 304, + 533, + 545, + 666 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{mr}} + X_{\\mathrm{mr}} * \\xi_{\\mathrm{h2m}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 671, + 454, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 454, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 454, + 684 + ], + "type": "text", + "content": "3.4. Task-dependent Joint Loss" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "A direct way to optimize the joint task is to sum the respective task-specific loss functions. However, it neglects" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18311" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "the variations in magnitudes of different task losses, leading to dominance of one task. Existing methods [20, 25, 29, 32, 36] utilize manually-set weights for the weighted sum of task losses, which limit the task learning as tasks evolve at different rates. To address this problem, inspired by the studies using dynamic weights [10, 16, 22, 23, 30] based on task learning stages, we introduce the task-dependent joint loss as a more effective and flexible solution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 171, + 287, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 171, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 287, + 218 + ], + "type": "text", + "content": "MR Loss. For moment retrieval, we define the MR likelihood using a Gaussian distribution as shown in Eq. (3), with its mean determined by the output " + }, + { + "bbox": [ + 47, + 171, + 287, + 218 + ], + "type": "inline_equation", + "content": "f^{\\theta_{\\mathrm{mr}}}(x)" + }, + { + "bbox": [ + 47, + 171, + 287, + 218 + ], + "type": "text", + "content": " of the neural network with " + }, + { + "bbox": [ + 47, + 171, + 287, + 218 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{mr}}" + }, + { + "bbox": [ + 47, + 171, + 287, + 218 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 171, + 287, + 218 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 171, + 287, + 218 + ], + "type": "text", + "content": " is the input of the neural network." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 84, + 236, + 287, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 236, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 84, + 236, + 287, + 251 + ], + "type": "interline_equation", + "content": "p \\left(Y _ {\\mathrm {m r}} \\mid f ^ {\\theta_ {\\mathrm {m r}}} (x)\\right) = \\mathcal {N} \\left(f ^ {\\theta_ {\\mathrm {m r}}} (x), \\delta_ {\\mathrm {m r}}\\right), \\tag {3}", + "image_path": "11dcfd20c0dc7356b05a60ec7d8ab137fb7893ae055edf16453f9e4db994eb4b.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 264, + 287, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 264, + 287, + 299 + ], + "spans": [ + { + "bbox": [ + 47, + 264, + 287, + 299 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 264, + 287, + 299 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{mr}}" + }, + { + "bbox": [ + 47, + 264, + 287, + 299 + ], + "type": "text", + "content": " is a learnable parameter that quantifies the uncertainty of moment retrieval. The negative log-likelihood can be derived as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 311, + 287, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 311, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 61, + 311, + 287, + 353 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} - \\log p \\left(Y _ {\\mathrm {m r}} \\mid f ^ {\\theta_ {\\mathrm {m r}}} (x)\\right) \\\\ \\propto \\frac {1}{2 \\delta_ {\\mathrm {m r}} {} ^ {2}} \\left\\| Y _ {\\mathrm {m r}} - f ^ {\\theta_ {\\mathrm {m r}}} (x) \\right\\| ^ {2} + \\log \\delta_ {\\mathrm {m r}}. \\tag {4} \\\\ \\end{array}", + "image_path": "daea7fb09b78f3850f1c55bc3208db3b2b5966f6d1ce3d3ef6b53dba826d1542.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 370, + 287, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 287, + 421 + ], + "type": "inline_equation", + "content": "\\left\\| Y_{\\mathrm{mr}} - f^{\\theta_{\\mathrm{mr}}}(x)\\right\\|^2" + }, + { + "bbox": [ + 47, + 370, + 287, + 421 + ], + "type": "text", + "content": " measures the offset between the model's predicted value and the ground-truth value. Therefore, inspired by Eq. (4), the MR loss function " + }, + { + "bbox": [ + 47, + 370, + 287, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mr}}(\\theta_{\\mathrm{mr}},\\delta_{\\mathrm{mr}})" + }, + { + "bbox": [ + 47, + 370, + 287, + 421 + ], + "type": "text", + "content": " can be defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 434, + 287, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 434, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 73, + 434, + 287, + 459 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {m r}} \\left(\\theta_ {\\mathrm {m r}}, \\delta_ {\\mathrm {m r}}\\right) = \\frac {1}{2 \\delta_ {\\mathrm {m r}} ^ {2}} \\mathcal {L} \\left(\\theta_ {\\mathrm {m r}}\\right) + \\log \\delta_ {\\mathrm {m r}}. \\tag {5}", + "image_path": "76d10d4d201eac64535b1582ec000ab4ab933fe6e34b29f1f1b72f91d8b87b8a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "text", + "content": "Following existing approaches [25, 32, 36], " + }, + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta_{\\mathrm{mr}})" + }, + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "text", + "content": " consists of three components: the L1 loss " + }, + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{L1}}" + }, + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "text", + "content": ", the generalized IoU loss [40] " + }, + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{gIoU}}" + }, + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "text", + "content": ", and the cross-entropy loss " + }, + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{BCE}}" + }, + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{L1}}" + }, + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{gIoU}}" + }, + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "text", + "content": " are employed to calculate the mean absolute error and gIoU deviation between ground-truth moments and predicted moments, respectively. " + }, + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{BCE}}" + }, + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "text", + "content": " is used to classify whether the predicted moments belong to the foreground or background. In summary, " + }, + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta_{\\mathrm{mr}}) = L_{\\mathrm{L1}} + L_{\\mathrm{gIoU}} + L_{\\mathrm{BCE}}" + }, + { + "bbox": [ + 47, + 475, + 287, + 571 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 574, + 287, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 287, + 645 + ], + "type": "text", + "content": "HD Loss. For highlight detection, inspired by the Boltzmann distribution, the HD likelihood is modeled using a softmax function applied to the scaled model output similar to [23], as Eq. (6). The learnable scaling factor is denoted as " + }, + { + "bbox": [ + 47, + 574, + 287, + 645 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{hd}}" + }, + { + "bbox": [ + 47, + 574, + 287, + 645 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 574, + 287, + 645 + ], + "type": "inline_equation", + "content": "f^{\\theta_{\\mathrm{hd}}}(x)" + }, + { + "bbox": [ + 47, + 574, + 287, + 645 + ], + "type": "text", + "content": " is the output of the network with " + }, + { + "bbox": [ + 47, + 574, + 287, + 645 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{hd}}" + }, + { + "bbox": [ + 47, + 574, + 287, + 645 + ], + "type": "text", + "content": " on input " + }, + { + "bbox": [ + 47, + 574, + 287, + 645 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 574, + 287, + 645 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 659, + 287, + 686 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 287, + 686 + ], + "type": "interline_equation", + "content": "p \\left(Y _ {\\mathrm {h d}} \\mid f ^ {\\theta_ {\\mathrm {h d}}} (x), \\delta_ {\\mathrm {h d}}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {1}{\\delta_ {\\mathrm {h d}} 2} f ^ {\\theta_ {\\mathrm {h d}}} (x)\\right). \\tag {6}", + "image_path": "4ef417183878099118e72e1a7d75d30564134988ad60b1af16b229d15d06aa98.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "Similar to Eq. (4), the negative log-likelihood of Eq. (6)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 423, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 423, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 423, + 83 + ], + "type": "text", + "content": "can be calculated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 91, + 545, + 231 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 545, + 231 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} - \\log p \\left(Y _ {\\mathrm {h d}} = y \\mid f ^ {\\theta_ {\\mathrm {h d}}} (x), \\delta_ {\\mathrm {h d}}\\right) \\\\ = \\frac {1}{\\delta_ {\\mathrm {h d}} ^ {2}} \\log \\frac {\\sum_ {y ^ {\\prime}} \\exp \\left(f _ {y ^ {\\prime}} ^ {\\theta_ {\\mathrm {h d}}} (x)\\right)}{\\exp \\left(f _ {y} ^ {\\theta_ {\\mathrm {h d}}} (x)\\right)} \\\\ + \\log \\frac {\\sum_ {y ^ {\\prime}} \\exp \\left(\\frac {1}{\\delta_ {\\mathrm {h d}} {} ^ {2}} f _ {y ^ {\\prime}} ^ {\\theta_ {\\mathrm {h d}}} (x)\\right)}{\\left(\\sum_ {y ^ {\\prime}} \\exp \\left(f _ {y ^ {\\prime}} ^ {\\theta_ {\\mathrm {h d}}} (x)\\right)\\right) ^ {\\frac {1}{\\delta_ {\\mathrm {h d}} {} ^ {2}}}} \\\\ \\approx - \\frac {1}{\\delta_ {\\mathrm {h d}} ^ {2}} \\log \\operatorname {s o f t m a x} \\left(Y _ {\\mathrm {h d}}, f ^ {\\theta_ {\\mathrm {h d}}} (x)\\right) + \\log \\delta_ {\\mathrm {h d}}, \\tag {7} \\\\ \\end{array}", + "image_path": "01502a3fd7a8d286a1178f579b88962fed024f36596219ebc3f22ecc2230f6e7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 231, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 231, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 231, + 545, + 289 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 231, + 545, + 289 + ], + "type": "inline_equation", + "content": "f_{y}^{\\theta_{\\mathrm{hd}}}(x)" + }, + { + "bbox": [ + 305, + 231, + 545, + 289 + ], + "type": "text", + "content": " refers to the y-th value of the " + }, + { + "bbox": [ + 305, + 231, + 545, + 289 + ], + "type": "inline_equation", + "content": "f^{\\theta_{\\mathrm{hd}}}(x)" + }, + { + "bbox": [ + 305, + 231, + 545, + 289 + ], + "type": "text", + "content": ". We introduce a simplify assumption " + }, + { + "bbox": [ + 305, + 231, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\sum_{y'} \\exp \\left( \\frac{1}{\\delta_{\\mathrm{hd}}^2} f_{y'}^{\\theta_{\\mathrm{hd}}}(x) \\right) \\approx \\left( \\sum_{y'} \\exp \\left( f_{y'}^{\\theta_{\\mathrm{hd}}}(x) \\right) \\right)^{\\frac{1}{\\delta_{\\mathrm{hd}}^2}}" + }, + { + "bbox": [ + 305, + 231, + 545, + 289 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 305, + 231, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{hd}} \\to 1" + }, + { + "bbox": [ + 305, + 231, + 545, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 289, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 347 + ], + "type": "text", + "content": "- log softmax " + }, + { + "bbox": [ + 305, + 289, + 545, + 347 + ], + "type": "inline_equation", + "content": "\\left(Y_{\\mathrm{hd}},f^{\\theta_{\\mathrm{hd}}}(x)\\right)" + }, + { + "bbox": [ + 305, + 289, + 545, + 347 + ], + "type": "text", + "content": " of Eq. (7) represents the cross-entropy classification loss of " + }, + { + "bbox": [ + 305, + 289, + 545, + 347 + ], + "type": "inline_equation", + "content": "Y_{\\mathrm{hd}}" + }, + { + "bbox": [ + 305, + 289, + 545, + 347 + ], + "type": "text", + "content": ". Inspired on this term, we generalize this expression to other classification losses. The HD loss function " + }, + { + "bbox": [ + 305, + 289, + 545, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{hd}}(\\theta_{\\mathrm{hd}},\\delta_{\\mathrm{hd}})" + }, + { + "bbox": [ + 305, + 289, + 545, + 347 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 341, + 356, + 545, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 356, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 341, + 356, + 545, + 381 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {h d}} \\left(\\theta_ {\\mathrm {h d}}, \\delta_ {\\mathrm {h d}}\\right) = \\frac {1}{\\delta_ {\\mathrm {h d}} ^ {2}} \\mathcal {L} \\left(\\theta_ {\\mathrm {h d}}\\right) + \\log \\delta_ {\\mathrm {h d}}. \\tag {8}", + "image_path": "cd7ce84d2da491f88e2a334a71f7c027295c79466115690bc18641e54f1778dc.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "text", + "content": "Consistent with prior approaches [25, 36], " + }, + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta_{\\mathrm{hd}})" + }, + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "text", + "content": " includes the hinge loss " + }, + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{hinge}}" + }, + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "text", + "content": ", negative video-query pairs loss " + }, + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{neg}}" + }, + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "text", + "content": " and rank-aware contrastive loss [17] " + }, + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cont}}" + }, + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta_{\\mathrm{hd}}) = L_{\\mathrm{hinge}} + L_{\\mathrm{neg}} + L_{\\mathrm{cont}}" + }, + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{hinge}}" + }, + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "text", + "content": " is computed between two pairs of positive and negative clips, with its margin 0.2 to maintain consistency with [25] for fairness. " + }, + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{neg}}" + }, + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cont}}" + }, + { + "bbox": [ + 305, + 389, + 545, + 473 + ], + "type": "text", + "content": " from [36] are used to reduce the saliency of negative pairs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 474, + 544, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 474, + 544, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 474, + 544, + 498 + ], + "type": "text", + "content": "We can obtain final task-dependent joint loss " + }, + { + "bbox": [ + 305, + 474, + 544, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{joint}}" + }, + { + "bbox": [ + 305, + 474, + 544, + 498 + ], + "type": "text", + "content": " through integrating Eq. (5) and Eq. (8):" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 331, + 508, + 545, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 508, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 331, + 508, + 545, + 521 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {j o i n t}} = \\mathcal {L} _ {\\mathrm {m r}} \\left(\\theta_ {\\mathrm {m r}}, \\delta_ {\\mathrm {m r}}\\right) + \\mathcal {L} _ {\\mathrm {h d}} \\left(\\theta_ {\\mathrm {h d}}, \\delta_ {\\mathrm {h d}}\\right). \\tag {9}", + "image_path": "64475717e3e6a952560a8beb740401c0ecd63fed4a0a2b92af61a408b108eab6.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 531, + 386, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 531, + 386, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 531, + 386, + 544 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 552, + 422, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 422, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 422, + 563 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": "Datasets. Extensive experiments are conducted on three benchmark datasets: QVHighlights [25], TVSum [43], and Charades-STA [14]. QVHighlights is currently the sole publicly dataset for joint moment retrieval and highlight detection tasks. It provides 10,310 queries associated with 18,367 moments, with an average of 1.8 disjoint moments per query. In contrast to other MR datasets with one-to-one query-moment mappings, QVHighlights aligns more closely with real-world scenarios. Each video in the dataset comprises 75 clips, each of which is 2s-long." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We also utilize two task-specific datasets, Charades-STA [14] for MR and TVSum [43] for HD, to evaluate the model." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18312" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 334, + 267 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 334, + 267 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 334, + 267 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 334, + 267 + ], + "type": "table", + "html": "
MethodVT VU GA MS PK PR FM BK BT DS Avg.
sLSTM [64]ECCV'1641.1 46.2 46.3 47.7 44.8 46.1 45.2 40.6 47.1 45.5 45.1
SG [34]CVPR'1742.3 47.2 47.5 48.9 45.6 47.3 46.4 41.7 48.3 46.6 46.2
LIM-S [51]CVPR'1955.9 42.9 61.2 54.0 60.4 47.5 43.2 66.3 69.1 62.6 56.3
Trailer [47]ECCV'2061.3 54.6 65.7 60.8 59.1 70.1 58.2 64.7 65.6 68.1 62.8
SL-Module [52]ICCV'2186.5 68.7 74.9 86.2 79.0 63.2 58.9 72.6 78.9 64.0 73.3
MINI-Net† [18]ECCV'2080.6 68.3 78.2 81.8 78.1 65.8 57.8 75.0 80.2 65.5 73.2
TCG† [57]ICCV'2185.0 71.4 81.9 78.6 80.2 75.5 71.6 77.3 78.6 68.1 76.8
Joint-VA† [3]ICCV'2183.7 57.3 78.5 86.1 80.1 69.2 70.0 73.0 97.4 67.5 76.3
UMT† [32]CVPR'2287.5 81.5 88.2 78.8 81.4 87.0 76.0 86.9 84.4 79.6 83.1
QD-DETR [36]CVPR'2388.2 87.4 85.6 85.0 85.8 86.9 76.4 91.3 89.2 73.7 85.0
UniVTG‡ [29]ICCV'2383.9 85.1 89.0 80.1 84.6 81.4 70.9 91.7 73.5 69.3 81.0
TaskWeave(Ours)88.2 90.8 93.3 87.5 87.0 82.0 80.9 92.9 89.5 81.2 87.3
", + "image_path": "b8a57e58b4746c7686e7e944c28f2aa3970d1e7aa90aa362c3467e5f39154590.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 321, + 328, + 458 + ], + "blocks": [ + { + "bbox": [ + 47, + 277, + 324, + 300 + ], + "lines": [ + { + "bbox": [ + 47, + 277, + 324, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 324, + 300 + ], + "type": "text", + "content": "Table 1. Experimental results " + }, + { + "bbox": [ + 47, + 277, + 324, + 300 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 47, + 277, + 324, + 300 + ], + "type": "text", + "content": " on TVSum. " + }, + { + "bbox": [ + 47, + 277, + 324, + 300 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 47, + 277, + 324, + 300 + ], + "type": "text", + "content": " means including audio modality. " + }, + { + "bbox": [ + 47, + 277, + 324, + 300 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 47, + 277, + 324, + 300 + ], + "type": "text", + "content": " means following the pretrain-finetune paradigm." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 321, + 328, + 458 + ], + "lines": [ + { + "bbox": [ + 50, + 321, + 328, + 458 + ], + "spans": [ + { + "bbox": [ + 50, + 321, + 328, + 458 + ], + "type": "table", + "html": "
MethodMRHD
R1mAP≥ Very Good
@0.5@0.7@0.5@0.75Avg.mAP HIT@1
Moment-DETR [25]NIPS'2153.9434.84--32.2035.6555.55
UMT†[32]CVPR'2260.2644.26--38.5939.8564.19
QD-DETR [36]CVPR'2362.6846.6662.2341.8241.2239.1363.03
EaTR [20]ICCV'2361.3645.7961.8641.9141.7437.1558.65
UniVTG‡[29]ICCV'2359.74---36.1338.8361.81
TaskWeave(Ours)64.2650.0665.3946.4745.3839.2863.68
", + "image_path": "f2e93bd6a286065241b2de12cf31ea1a53b339346ac12a74c7b6dfc90f5efe4d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 287, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 546 + ], + "type": "text", + "content": "Charades-STA provides 16,128 query-moment pairs. TVSum comprises videos from 10 domains, with each domain containing 5 videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. We utilize the same evaluation metrics used in prior approaches [25, 32, 36]. For QVHighlights, mean average precision (mAP) with different tIoU thresholds 0.5, 0.75, the average mAP over [0.5:0.05:0.95], and Recall@1 with tIoU 0.5, 0.7 are utilized for MR evaluation. mAP and HIT@1 are used for HD evaluation, where a clip is considered as a true positive when it achieves a \"Very Good\" [25] saliency score. For Charades-STA, we utilize Recall@1 with tIoU thresholds 0.5, 0.7. For TVSum, Top-5 mAP is adopted." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 351, + 70, + 541, + 268 + ], + "blocks": [ + { + "bbox": [ + 46, + 466, + 323, + 490 + ], + "lines": [ + { + "bbox": [ + 46, + 466, + 323, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 323, + 490 + ], + "type": "text", + "content": "Table 3. Experimental results (\\%) on QVHighlights val split. † means including audio modality. ‡ means following the pretrain-finetune paradigm." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 351, + 70, + 541, + 268 + ], + "lines": [ + { + "bbox": [ + 351, + 70, + 541, + 268 + ], + "spans": [ + { + "bbox": [ + 351, + 70, + 541, + 268 + ], + "type": "table", + "html": "
BackboneMethodR1@0.5R1@0.7
VGGSAP [9]AAAI'1927.4213.36
SM-RL [48]CVPR'1924.3611.17
2D-TAN [65]AAAI'2040.9422.85
FVMR [13]CVPR'2124.3611.17
UMT† [32]CVPR'2248.3129.25
QD-DETR [36]CVPR'2352.7731.13
TaskWeave(Ours)56.5133.66
I3DCTRL [14]ICCV'1723.638.89
MAN [60]CVPR'1946.5322.72
VSLNet [61]ACL'2047.3130.19
QD-DETR [36]CVPR'2350.6731.02
TaskWeave(Ours)53.3631.4
", + "image_path": "cef49f43e5d2dbbb82611e99ae6658513341a060a7fe729dfa87280e95ffdffb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 348, + 350, + 543, + 458 + ], + "blocks": [ + { + "bbox": [ + 343, + 277, + 545, + 311 + ], + "lines": [ + { + "bbox": [ + 343, + 277, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 343, + 277, + 545, + 311 + ], + "type": "text", + "content": "Table 2. Experimental results " + }, + { + "bbox": [ + 343, + 277, + 545, + 311 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 343, + 277, + 545, + 311 + ], + "type": "text", + "content": " on Charades-STA test split. " + }, + { + "bbox": [ + 343, + 277, + 545, + 311 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 343, + 277, + 545, + 311 + ], + "type": "text", + "content": " means including audio modality. " + }, + { + "bbox": [ + 343, + 277, + 545, + 311 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 343, + 277, + 545, + 311 + ], + "type": "text", + "content": " means following the pretrain-finetune paradigm." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 348, + 350, + 543, + 458 + ], + "lines": [ + { + "bbox": [ + 348, + 350, + 543, + 458 + ], + "spans": [ + { + "bbox": [ + 348, + 350, + 543, + 458 + ], + "type": "table", + "html": "
Task-decoupled UnitInter-task feedbackJoint lossMRHD
R1 @0.7Avg. mAPmAP HIT@1
46.2641.038.9462.84
✓47.8743.2438.5861.81
✓✓49.2945.1238.9662.0
✓✓✓50.0645.3839.2863.68
", + "image_path": "b68bc165132fe2d750b6b911fb3695b189f1044f3cb8ef8def412e057b3af69f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 343, + 466, + 545, + 489 + ], + "lines": [ + { + "bbox": [ + 343, + 466, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 343, + 466, + 545, + 489 + ], + "type": "text", + "content": "Table 4. The ablation results " + }, + { + "bbox": [ + 343, + 466, + 545, + 489 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 343, + 466, + 545, + 489 + ], + "type": "text", + "content": " of the components of our proposed method." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 510, + 440, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 510, + 440, + 522 + ], + "spans": [ + { + "bbox": [ + 306, + 510, + 440, + 522 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 531, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 639 + ], + "type": "text", + "content": "Feature representations. For OVHighlights, we employ the pre-trained SlowFast [12] and CLIP [39] backbones to extract video features, following [25, 32, 36] for fairness. For Charedes-STA, we leverage the official release of VGG [42] and I3D [6] features as video embeddings. For TVSum, we extract video features by the I3D [6] pre-trained on Kinetics 400 [21]. For QVHighlights and TVSum, we use CLIP [39] to extract text features, while using GloVe [38] text embeddings for Charades-STA." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 546, + 714 + ], + "type": "text", + "content": "Training settings. We leverage AdamW [24] optimizer with 1e-4 learning rate and 1e-4 weight decay. We train 200, 100, and 2000 epochs with batch size 32, 32 and 2 for QVHighlights, Charades-STA and TVSum, respectively. All Transformer layers follow the consistent configuration, including sinusoidal positional encodings, 8 at" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18313" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 223, + 205 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 223, + 205 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 223, + 205 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 223, + 205 + ], + "type": "table", + "html": "
IndexMR-expertHR expertMRHDR
R1 @0.7Avg. mAPmAPHIT@1
(a)Iden.Iden.44.7740.2338.3860.19
(b)Line.Line.47.5543.6538.3260.77
(c)CNNCNN49.2343.7338.8260.77
(d)Trans.Trans.48.3242.8538.6161.10
(e)Trans.Iden.47.142.138.761.68
(f)CNNIden.49.2945.1238.9662.0
", + "image_path": "25aa03576deb1abb81c7806be9f7edbba3057faaeff947754a473dae3a433ad5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 235, + 105, + 390, + 205 + ], + "blocks": [ + { + "bbox": [ + 46, + 212, + 225, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 225, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 225, + 246 + ], + "type": "text", + "content": "Table 5. Flexibility validation " + }, + { + "bbox": [ + 46, + 212, + 225, + 246 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 212, + 225, + 246 + ], + "type": "text", + "content": " of the taskdecoupled unit. \"Iden:\" identity mapping; \"Trans\":Transformer; \"Line\":linear layer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 235, + 105, + 390, + 205 + ], + "lines": [ + { + "bbox": [ + 235, + 105, + 390, + 205 + ], + "spans": [ + { + "bbox": [ + 235, + 105, + 390, + 205 + ], + "type": "table", + "html": "
TypeMRHD
R1 @0.7Avg. mAPmAP HIT@1
Sum47.6844.2738.760.32
Weighted Sum48.3243.2138.3161.35
Ours50.0645.3839.2863.68
", + "image_path": "b7fd42d00f55d9a8d6ada1fe5406e91070f210c0a77c3ffd152a667119cce8d4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 397, + 82, + 544, + 205 + ], + "blocks": [ + { + "bbox": [ + 233, + 212, + 386, + 235 + ], + "lines": [ + { + "bbox": [ + 233, + 212, + 386, + 235 + ], + "spans": [ + { + "bbox": [ + 233, + 212, + 386, + 235 + ], + "type": "text", + "content": "Table 6. Effectiveness justification " + }, + { + "bbox": [ + 233, + 212, + 386, + 235 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 233, + 212, + 386, + 235 + ], + "type": "text", + "content": " of the task-dependent joint loss." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 397, + 82, + 544, + 205 + ], + "lines": [ + { + "bbox": [ + 397, + 82, + 544, + 205 + ], + "spans": [ + { + "bbox": [ + 397, + 82, + 544, + 205 + ], + "type": "table", + "html": "
TypeMRHD
R1 @0.7Avg. mAPmAPHIT@1
MR2HD50.0645.3838.7362.84
HD2MR48.6544.7839.2863.68
Bi-MRHD48.8445.138.3661.81
MR-HD50.045.338.9662.45
HD-MR49.144.8739.0359.94
", + "image_path": "c6617c555f9c4b4369360c0b39fc489e1183b62960d4e8883076fe43efbaaa3c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 393, + 212, + 545, + 246 + ], + "lines": [ + { + "bbox": [ + 393, + 212, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 393, + 212, + 545, + 246 + ], + "type": "text", + "content": "Table 7. Comparing various inter-task feedback combinations " + }, + { + "bbox": [ + 393, + 212, + 545, + 246 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 393, + 212, + 545, + 246 + ], + "type": "text", + "content": " on QVHighlights val split." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 266, + 289, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 399 + ], + "type": "text", + "content": "tention heads, and a dropout rate of 0.1. Multi-modal fusion employs a 2-layer Transformer with cross attention, where video features serve as the query and text features serve as key and value. For Eq. (9), we define " + }, + { + "bbox": [ + 46, + 266, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\gamma_{\\mathrm{mr}}, \\gamma_{\\mathrm{hd}}" + }, + { + "bbox": [ + 46, + 266, + 289, + 399 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 266, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\log \\delta_{\\mathrm{mr}}^2, \\log \\delta_{\\mathrm{hd}}^2" + }, + { + "bbox": [ + 46, + 266, + 289, + 399 + ], + "type": "text", + "content": ", respectively. Therefore, Eq. (9) can be rewritten as " + }, + { + "bbox": [ + 46, + 266, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{joint} = \\exp(-\\gamma_{\\mathrm{mr}})\\mathcal{L}(\\theta_{\\mathrm{mr}}) + 2\\exp(-\\gamma_{\\mathrm{hd}})\\mathcal{L}(\\theta_{\\mathrm{hd}}) + \\gamma_{\\mathrm{mr}} + \\gamma_{\\mathrm{hd}}" + }, + { + "bbox": [ + 46, + 266, + 289, + 399 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 266, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\gamma_{\\mathrm{mr}}" + }, + { + "bbox": [ + 46, + 266, + 289, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 266, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\gamma_{\\mathrm{hd}}" + }, + { + "bbox": [ + 46, + 266, + 289, + 399 + ], + "type": "text", + "content": " are learnable parameters. They are initialized to 0. To further stabilize the training, the model EMA strategy has also undertaken. All experiments are conducted with Pytorch v1.13.1 on a single NVIDIA RTX 3090." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 407, + 231, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 231, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 231, + 419 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-arts" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "text", + "content": "Results on TVSum. In Tab. 1, we compare our TaskWeave with existing state-of-the-arts (SOTA) methods, where the methods with " + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "text", + "content": " are incorporate with audio features. UniVTG [29] follows pretrain-then-finetune paradigm, therefore we use its non-pretrained results for a fair comparison. We observe that: i) our methods outperforms SOTA methods in 9 out of the 10 categories; ii) TaskWeave significantly outperforms all methods in terms of Avg. mAP, with a remarkable " + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "inline_equation", + "content": "2.71\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "text", + "content": " improvement over the previous SOTA method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": "Results on Charades-STA. In Tab. 2, we utilize different backbones to demonstrate the effectiveness of our TaskWeave. We compare our results with other methods for MR only (white background) and methods for joint MR and HD (gray background). It results in significant improvements, i.e., " + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "inline_equation", + "content": "+7.09\\%" + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": " in R1@0.5 and " + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "inline_equation", + "content": "+8.13\\%" + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": " in R1@0.7." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "Results on QVHighlights. As shown in Tab. 3, we present the performance comparisons with existing methods. We report the MR results with the MR-guided feedback, while HD results with the HD-guided feedback. We observe that: i) our TaskWeave surpasses SOTA methods by a large margin in all moment retrieval metrics, with a remarkable " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "8.72\\%" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " improvement in Avg. mAP; ii) our method achieves the second-best performance on the HD. However," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 266, + 547, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 547, + 326 + ], + "type": "text", + "content": "It's worth noting that UMT [32] incorporates audio features, while our method does not. When compared to other methods without audio features, our method outperforms them all. In summary, these results validate the effectiveness of our task-driven framework." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 327, + 547, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 547, + 423 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 547, + 423 + ], + "type": "text", + "content": "We provide visualization examples in Fig. 4 for the qualitative analysis among our TaskWeave, Moment-DETR [25], and QD-DETR [36]. Given diverse queries for a video, TaskWeave precisely localizes moments for queries and presents high IoU with the ground-truth. Moment-DETR misses some instances, and QD-DETR exhibits lower retrieval accuracy. Moreover, our method obtains higher saliency scores for relevant clips in response to the query." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 431, + 484, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 484, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 484, + 443 + ], + "type": "text", + "content": "4.4. Ablation Studies and Discussions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 449, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 485 + ], + "type": "text", + "content": "We present some ablation studies and discussions about TaskWeave, with all experiments are conducted on the QVHighlights val split [25]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 486, + 547, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 593 + ], + "type": "text", + "content": "Ablation of components. We conduct ablation studies on each component of TaskWeave, as shown in Tab. 4. Sequentially employing the task-decoupled unit and our proposed joint loss contributes a " + }, + { + "bbox": [ + 304, + 486, + 547, + 593 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 304, + 486, + 547, + 593 + ], + "type": "text", + "content": " improvement in average mAP and a " + }, + { + "bbox": [ + 304, + 486, + 547, + 593 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 304, + 486, + 547, + 593 + ], + "type": "text", + "content": " increase in R1@0.7 for MR. Utilizing all components, we observe that a significant " + }, + { + "bbox": [ + 304, + 486, + 547, + 593 + ], + "type": "inline_equation", + "content": "10.7\\%" + }, + { + "bbox": [ + 304, + 486, + 547, + 593 + ], + "type": "text", + "content": " improvement in average mAP of MR and a " + }, + { + "bbox": [ + 304, + 486, + 547, + 593 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 304, + 486, + 547, + 593 + ], + "type": "text", + "content": " enhancement in HIT@1 of HD. These results indicate the effectiveness of the proposed components in out TaskWeave." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "text", + "content": "Flexibility of the task-decoupled unit. To demonstrate the flexibility of the task-decoupled unit, we investigate the performance of applying different networks within various experts, as shown in Tab. 5. The shared expert utilize a fixed configuration with 2 Transformer layers. We provide 4 different methods for each task-specific expert, including the feed-forward network (one layer Linear), the identity mapping, CNN (composed of one depthwise convolution layer with the kernel/stride/padding of " + }, + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "inline_equation", + "content": "5/1/2" + }, + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "text", + "content": " and 1D convolution), and the Transformer. Due to space constraints, we present" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18314" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 85, + 547, + 248 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 520, + 83 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 520, + 83 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 520, + 83 + ], + "type": "text", + "content": "Query1: A guy with grey top is showing a box filled with rubbish. Query2: A man shaves his beard looks in a mirror and talks to the camera." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 85, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 47, + 85, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 47, + 85, + 547, + 248 + ], + "type": "image", + "image_path": "0a4aa83bb45accdc89c01decbb04416267c07dffc42087a05418c66c8287babe.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 546, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 546, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 546, + 278 + ], + "type": "text", + "content": "Figure 4. Qualitative results on the QVHighlights for Ground-Truth, Moment-DETR, QD-DETR and our method. The predicted moments and saliency scores are illustrated through intervals and lines." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 300, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 323 + ], + "type": "text", + "content": "results of 6 out of 16 combinations. The results in Tab. 5 are obtained by TaskWeave without inter-task feedback." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 326, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 517 + ], + "type": "text", + "content": "From this table, we can observe several interesting facts. First, different task-specific experts have different performance. The need for experts is reflected in (a). Second, although Transformer has made significant progress in visual, it's not a panacea. Results in (d) are lower compared to (b) and (c), we believe that the high computational complexity of Transformer decreases the performance. Third, the task-specific expert should be designed based on the task objective. For instance, (c) and (f) perform well on the MR because they focus on local features, which is important for localizing moments. Finally, we also find evidence that MR and HD are highly related. Improvement in one task enhances the other (compare (c) and (f)), while a decline in one limits the other's performance (compare (d) and (e)). In this paper, our mr-specific expert utilizes CNN and hd-specific expert is implemented by identity mapping." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 519, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 628 + ], + "type": "text", + "content": "Task-dependent joint loss. Tab. 6 shows the performance of TaskWeave with different losses. \"Sum\" and \"Weighted Sum\" refer to the fusion manner of loss for different tasks, respectively. Weights in \"Weighted Sum\" are consistent with the existing methods [25, 36] for fairness. Comparing \"Sum\" and \"Weighted Sum\", we find that \"Weighted Sum\" can better balance the performance of the two tasks. However, it is obvious that the proposed task-dependent joint loss is optimal." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": "Inter-task interactions. We believe that each task requires learning before providing effective feedback. Therefore, in our inter-task feedback mechanism, the feedback process starts when the model is trained to half of the max epoch. For brevity, we write Moment/Highlightness-guided feedback as MgF/HgF. In Tab. 7, we explore five feedback manners as follows: \"MR2HD\" (with MgF only)," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 299, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 545, + 335 + ], + "type": "text", + "content": "\"HD2MR\" (with HgF only), \"Bi-MRHD\" (with MgF and HgF simultaneously), \"MR-HD\" (MgF first, then HgF), and \"HD-MR\" (HgF first, then MgF)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 335, + 546, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 546, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 546, + 455 + ], + "type": "text", + "content": "We find that the inter-task feedback makes both tasks gain simultaneously. The performance of the model with \"HD2MR\" is slightly worse, this is because HD focuses on more refined moments than MR. Ground-truth annotations and the results of \"MR-HD\" and \"HD-MR\" also illustrate this fact. The gain brought by \"Bi-MRHD\" is limited, which we believe that it's a natural result of feedback not always being effective. In general, the inter-task feedback not only contributes to bring gains for both tasks but also helps to understand the characteristics of each task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 464, + 379, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 379, + 477 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 379, + 477 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 485, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 628 + ], + "type": "text", + "content": "This paper proposes a novel task-driven paradigm for addressing joint moment retrieval and highlight detection. Different from existing data-driven methods, we utilize the task-decoupled unit to capture the task-specific and common features, respectively. We also explore different network architecture for moment retrieval and highlight detection. We design the inter-task feedback mechanism to in-depth investigate the interplay between both tasks. Different from prior methods, we introduce the principled joint loss to optimize the model. The effectiveness, flexibility, and superiority of the proposed method have been demonstrated on three benchmark datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 639, + 404, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 404, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 404, + 651 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 658, + 545, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 658, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 304, + 658, + 545, + 702 + ], + "type": "text", + "content": "This research was supported by the National Natural Science Foundation of China (No. U23B2060, 62088102), the Youth Innovation Team of Shaanxi Universities, and the Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18315" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "text", + "content": "[2] Niki Parmar Jakob Uszkoreit Llion Jones Aidan N. Gomez Lukasz Kaiser Ashish Vaswani, Noam Shazeer and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5999-6009, 2017. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 288, + 248 + ], + "type": "text", + "content": "[3] Taiwanbat Badamdorj, Mrigank Rochan, Yang Wang, and Li Cheng. Joint visual and audio learning for video highlight detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8127-8137, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 288, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 303 + ], + "type": "text", + "content": "[4] Taiwanbat Badamdorj, Mrigank Rochan, Yang Wang, and Li Cheng. Contrastive learning for unsupervised video highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14042-14052, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 396 + ], + "type": "text", + "content": "[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 397, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 418 + ], + "type": "text", + "content": "[7] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 287, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 287, + 463 + ], + "type": "text", + "content": "[8] Jingyuan Chen, Lin Ma, Xinpeng Chen, Zequn Jie, and Jiebo Luo. Localizing natural language in videos. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8175-8182, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 465, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 287, + 509 + ], + "type": "text", + "content": "[9] Shaoxiang Chen and Yu-Gang Jiang. Semantic proposal for activity localization in videos via sentence query. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8199-8206, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 511, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 565 + ], + "type": "text", + "content": "[10] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 632 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[12] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[13] Junyu Gao and Changsheng Xu. Fast video moment retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1523-1532, 2021. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[14] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 2, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[15] Ana Garcia del Molino and Michael Gygli. PhD-gifs: personalized highlight detection for automatic gif creation. In Proceedings of the 26th ACM international conference on Multimedia, pages 600-608, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "text", + "content": "[16] Michelle Guo, Albert Haque, De-An Huang, Serena Yeung, and Li Fei-Fei. Dynamic task prioritization for multitask learning. In Proceedings of the European conference on computer vision (ECCV), pages 270–287, 2018. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 206, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 260 + ], + "type": "text", + "content": "[17] David T Hoffmann, Nadine Behrmann, Juergen Gall, Thomas Brox, and Mehdi Noroozi. Ranking info noise contrastive estimation: Boosting contrastive learning via ranked positives. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 897–905, 2022. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 261, + 545, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 325 + ], + "type": "text", + "content": "[18] Fa-Ting Hong, Xuanteng Huang, Wei-Hong Li, and Wei-Shi Zheng. Mini-net: Multiple instance ranking network for video highlight detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 345-360. Springer, 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "text", + "content": "[19] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "text", + "content": "[20] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13846-13856, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 416, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 469 + ], + "type": "text", + "content": "[21] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 471, + 545, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 504 + ], + "type": "text", + "content": "[22] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "text", + "content": "[23] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482-7491, 2018. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 559, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 591 + ], + "type": "text", + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "text", + "content": "[25] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34: 11846-11858, 2021. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 636, + 545, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 691 + ], + "type": "text", + "content": "[26] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[27] Huan Li, Ping Wei, Jiapeng Li, Zeyu Ma, Jiahui Shang, and Nanning Zheng. Asymmetric relation consistency reasoning" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18316" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "for video relation grounding. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 148 + ], + "type": "text", + "content": "[28] Huan Li, Ping Wei, Zeyu Ma, and Nanning Zheng. Inverse compositional learning for weakly-supervised relation grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15477-15487, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 217 + ], + "type": "text", + "content": "[29] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified videolanguage temporal grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2794-2804, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 260 + ], + "type": "text", + "content": "[30] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1871-1880, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[31] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 369 + ], + "type": "text", + "content": "[32] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3042-3051, 2022. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 435 + ], + "type": "text", + "content": "[33] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "text", + "content": "[34] Behrooz Mahasseni, Michael Lam, and Sinisa Todorovic. Unsupervised video summarization with adversarial LSTM networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 202-211, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 492, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 547 + ], + "type": "text", + "content": "[35] Otniel-Bogdan Mercea, Lukas Riesch, A Koepke, and Zeynep Akata. Audio-visual generalised zero-shot learning with cross-modal attention and language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10553–10563, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "text", + "content": "[36] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23023-23033, 2023. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "text", + "content": "[37] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810-10819, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[38] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 140, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 205 + ], + "type": "text", + "content": "[40] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 207, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 261 + ], + "type": "text", + "content": "[41] Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, and Dacheng Tao. Tridet: Temporal action detection with relative boundary modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18857-18866, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 263, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 545, + 297 + ], + "type": "text", + "content": "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "text", + "content": "[43] Yale Song, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaimes. Tvsum: Summarizing web videos using titles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5179-5187, 2015. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "type": "text", + "content": "[44] Min Sun, Ali Farhadi, and Steve Seitz. Ranking domain-specific highlights by analyzing edited videos. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13, pages 787-802. Springer, 2014. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 399, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 454 + ], + "type": "text", + "content": "[45] Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. Progressive layered extraction (ple): A novel multi-task learning (mtl) model for personalized recommendations. In Proceedings of the 14th ACM Conference on Recommender Systems, pages 269-278, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "text", + "content": "[46] Tuan N Tang, Kwonyoung Kim, and Kwanghoon Sohn. Temporalmaxer: Maximize temporal context with only max pooling for temporal action localization. arXiv preprint arXiv:2303.09055, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 500, + 545, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 565 + ], + "type": "text", + "content": "[47] Lezi Wang, Dong Liu, Rohit Puri, and Dimitris N Metaxas. Learning trailer moments in full-length movies with co-contrastive attention. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVIII 16, pages 300–316. Springer, 2020. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 567, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 623 + ], + "type": "text", + "content": "[48] Weining Wang, Yan Huang, and Liang Wang. Language-driven temporal activity localization: A semantic matching reinforcement learning model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 334-343, 2019. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 624, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 679 + ], + "type": "text", + "content": "[49] Fanyue Wei, Biao Wang, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Learning pixel-level distinctions for video highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3073-3082, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[50] Xi Wei, Tianzhu Zhang, Yan Li, Yongdong Zhang, and Feng Wu. Multi-modality cross attention network for image and sentence matching. In Proceedings of the IEEE/CVF con" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18317" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "text", + "content": "ference on computer vision and pattern recognition, pages 10941-10950, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 287, + 151 + ], + "type": "text", + "content": "[51] Bo Xiong, Yannis Kalantidis, Deepi Ghadiyaram, and Kristen Grauman. Less is more: Learning highlight detection from video duration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1258-1267, 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 154, + 287, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 154, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 48, + 154, + 287, + 209 + ], + "type": "text", + "content": "[52] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7970-7979, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 211, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 211, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 211, + 287, + 255 + ], + "type": "text", + "content": "[53] Yifang Xu, Yunzhuo Sun, Yang Li, Yilei Shi, Xiaoxiang Zhu, and Sidan Du. Mh-detr: Video moment and highlight detection with cross-modal transformer. arXiv preprint arXiv:2305.00355, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 257, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 257, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 257, + 287, + 289 + ], + "type": "text", + "content": "[54] Jin Yang, Ping Wei, Ziyang Ren, and Nanning Zheng. Gated multi-scale transformer for temporal action localization. IEEE Transactions on Multimedia, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 292, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 292, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 292, + 287, + 335 + ], + "type": "text", + "content": "[55] Jin Yang, Ping Wei, and Nanning Zheng. Cross time-frequency transformer for temporal action localization. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 382 + ], + "type": "text", + "content": "[56] Ting Yao, Tao Mei, and Yong Rui. Highlight detection with pairwise deep ranking for first-person video summarization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 982–990, 2016. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 384, + 287, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 384, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 48, + 384, + 287, + 438 + ], + "type": "text", + "content": "[57] Qinghao Ye, Xiyue Shen, Yuan Gao, Zirui Wang, Qi Bi, Ping Li, and Guang Yang. Temporal cue guided video highlight detection with low-rank audio-visual fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7950–7959, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 441, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 441, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 441, + 287, + 485 + ], + "type": "text", + "content": "[58] Yitian Yuan, Tao Mei, and Wenwu Zhu. To find where you talk: Temporal sentence localization in video with attention based location regression. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9159–9166, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 487, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 541 + ], + "type": "text", + "content": "[59] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 609 + ], + "type": "text", + "content": "[60] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1247-1257, 2019. 1, 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "type": "text", + "content": "[61] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6543-6554, 2020. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[62] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Temporal sentence grounding in videos: A survey and future directions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10443-10465, 2023. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 228 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[63] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Temporal sentence grounding in videos: a survey and future directions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "text", + "content": "[64] Ke Zhang, Wei-Lun Chao, Fei Sha, and Kristen Grauman. Video summarization with long short-term memory. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 766-782. Springer, 2016. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "type": "text", + "content": "[65] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12870-12877, 2020. 1, 2, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18318" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/d6dc5c43-4c52-4779-8c6a-d513a47a71be_content_list.json b/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/d6dc5c43-4c52-4779-8c6a-d513a47a71be_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5c1e68c3f05efd35acb6b22f0089b84a618e50a6 --- /dev/null +++ b/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/d6dc5c43-4c52-4779-8c6a-d513a47a71be_content_list.json @@ -0,0 +1,1368 @@ +[ + { + "type": "text", + "text": "Task-Driven Wavelets using Constrained Empirical Risk Minimization", + "text_level": 1, + "bbox": [ + 129, + 130, + 841, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eric Marcus\\*,1,2, Ray Sheombarsing\\*,1, Jan-Jakob Sonke $^{1}$ , Jonas Teuwen $^{1}$", + "bbox": [ + 192, + 179, + 774, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Netherlands Cancer Institute, 2University of Amsterdam {e.marcus, r.sheombarsing, j.sonke, j.teuwen}@nki.nl", + "bbox": [ + 250, + 215, + 712, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep Neural Networks (DNNs) are widely used for their ability to effectively approximate large classes of functions. This flexibility, however, makes the strict enforcement of constraints on DNNs a difficult problem. In contexts where it is critical to limit the function space to which certain network components belong, such as wavelets employed in Multi-Resolution Analysis (MRA), naive constraints via additional terms in the loss function are inadequate. To address this, we introduce a Convolutional Neural Network (CNN) wherein the convolutional filters are strictly constrained to be wavelets. This allows the filters to update to task-optimized wavelets during the training procedure. Our primary contribution lies in the rigorous formulation of these filters via a constrained empirical risk minimization framework, thereby providing an exact mechanism to enforce these structural constraints. While our work is grounded in theory, we investigate our approach empirically through applications in medical imaging, particularly in the task of contour prediction around various organs, achieving superior performance compared to baseline methods.", + "bbox": [ + 75, + 318, + 473, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 652, + 209, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Empirical risk minimization (ERM) is currently the most prevalent framework for supervised learning. The goal is to find a function to map inputs to associated targets for all representative (potentially unseen) observations. To find such a mapping, one introduces a loss function to quantify the discrepancy between observed and predicted targets. An optimal map is then found by minimizing the expected loss. In large-scale settings, such as deep learning, the resulting minimization problem is solved using Stochastic Gradient Descent (SGD) and various variants thereof, see [15, 24, 25, 40] for instance. As deep learning applications become more specialized, domain-specific needs become increasingly vital. These are often formulated in terms of constraints on", + "bbox": [ + 75, + 679, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6cbd1c25d4bb69908a39e617411cf75bda957b28c1fed880276de18fe827d1e8.jpg", + "image_caption": [ + "Figure 1. An overview of the Constrained Empirical Risk Minimization (CERM) framework: on the top left side, the full gradient of the loss $\\mathcal{L}$ is shown on the total space $\\mathcal{N}$ (containing both the constrained and unconstrained space), with $\\alpha$ are the unconstrained parameters, and $\\theta$ the constrained ones. The constraints can be written in the form $F(\\theta) = 0$ ; the solution set $F^{-1}(0)$ is an embedded submanifold $\\mathcal{M}$ of $\\mathbb{R}^{\\tilde{p}}$ . The constrained parameters are updated by following a path on $\\mathcal{M}$ in the direction of the negative gradient on this manifold. This constrained part of the full gradient is contained in the tangent space $T_{\\theta}\\mathcal{M}$ of the submanifold. The color of the manifold indicates the value of the loss function. By restricting the relevant components of our descent trajectories to the embedded submanifold, we always satisfy the constraints imposed by $F$ . The gradient and parameter updates for the unconstrained parameters $\\alpha$ are computed as usual using standard SGD for flat space (not shown). In our wavelet scenario, the manifold thus consists of all admissible wavelets of a certain dimension." + ], + "image_footnote": [], + "bbox": [ + 581, + 303, + 815, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the permissible mappings. For instance, the constraint for translation-equivariance led to the development and success of convolution neural networks (CNNs) [17]. In general, however, it is a highly non-trivial task to construct network architectures that satisfy a set of constraints, if they exist", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 94, + 886, + 351, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24098", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "at all. It is, therefore, common practice to incorporate constraints directly into the loss function by including additional terms, usually referred to as \"soft\" constraints. This setup, however, has the drawback that the constraints are only approximately (on average) satisfied due to the formulation of ERM. Moreover, incorporating many different objectives in a loss function may lead to suboptimal results for the individual objectives. Another common strategy, used for many types of constraints, e.g., flip invariance, is based on data augmentation. However, not every constraint can be achieved using data augmentation, and it also leads, at best, to constraints being approximately satisfied. Furthermore, if we wish to constrain neural networks to specific function classes, such as wavelets, an additional loss term is no longer sufficient: a wavelet is only a wavelet whenever the constraints are satisfied exactly.", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Approaches which circumvent loss-based (soft) constraints have been recently proposed, see [2, 4, 28, 30, 32] for example. Most notably, we mention the field of Geometric Deep Learning (GDL), which focuses on incorporating constraints arising from symmetries of the domain directly into the networks themselves, see [8, 9, 39] and the references therein. In GDL, one considers very specific but powerful types of constraints, namely that network layers are equivariant with respect to some group action. Not all constraints arise as equivariance principles, however. A large class of examples comes from highly specialized requirements on the output of a neural network, e.g., that the output is a divergence-free vector field, a contour, or perhaps a surface. For example, in medical image segmentation, a natural requirement is that the output of the segmentation network corresponds to a continuous (closed) curve. Creating architectures for such constraints is out of the realm of GDL and, in general, poses a challenging task; no prescribed methods exist to do so.", + "bbox": [ + 75, + 337, + 472, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To facilitate these different types of constraints, we present a framework, called Constrained Empirical Risk Minimization (CERM), where we directly formulate and solve the constrained ERM problem in the space where the constraints are satisfied. This space forms, under mild conditions, a Riemannian manifold. As long as we stay on this manifold, the constraints will be satisfied exactly up to numerical precision. In particular, we explain how to perform SGD on such Riemannian manifolds arising from a finite-dimensional system of equations. Although we are motivated by the ability to constrain to wavelets, the framework, in principle, encompasses constraints of the form $F(W) = 0$ , with $W$ the network weights. Our method heavily relies on the Implicit Function Theorem, which is used to construct so-called graph-charts amenable to numerical computations. This allows for efficient evaluation of the (induced) Riemannian metric and gradients, which are vital for performing SGD on Riemannian manifolds using only the intrinsic geometry.", + "bbox": [ + 75, + 628, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As mentioned, we will present the tools and implementations of convolutional networks whose filters will form wavelets. There are several tasks at which one expects wavelet-based neural networks to excel. Wavelet decompositions naturally lend themselves to representing continuous objects such as curves, images, vector fields, or other higher-dimensional objects. Hence any task where the object of interest can be identified with a smooth or continuous function is well-suited for wavelet-based neural networks. There is an abundance of such examples to be found in computer vision, e.g., boundary prediction, image registration, and so forth. Another family of exciting applications can be found in signal analysis, e.g., in compression and denoising, where wavelets are long-standing tools that have proven to be extremely efficient [26]. The main idea in these areas is to extract information about noise, smoothness, and even singularities, through analysis of the wavelet coefficients. Subsequently, by modifying a subset of the coefficients, e.g., through thresholding, the signal can be cleaned up or denoised. More recently, wavelets have also received increasing attention in the context of neural networks. The applications are broad; some references include [22, 33, 41], ranging from efficient normalizing flows, noise-robust training, scale-equivariant models, and many more. The wavelet layers we create in this work are included in our code, which we will make available publicly*.", + "bbox": [ + 496, + 90, + 895, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Related work An earlier attempt at incorporating constraints is described in [11]. However, this method is only able to deal with linear constraints, which is hence incompatible with the wavelet constraints we will derive. Other works include [20, 21, 27], which are related to the method of Lagrange multipliers and have their optimization and training dynamics largely determined by variants of Newton's method. We discuss the differences between our SGD-compatible method and that of Lagrange multipliers in more detail in the supplementary material 6.2. A comprehensive overview of all constrained optimization methods, however, is out of the scope of this paper, and we therefore focus our attention to techniques based on SGD on Riemannian manifolds.", + "bbox": [ + 496, + 501, + 893, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Performing SGD on Riemannian manifolds is well-established and has been studied extensively, see [6, 14, 34, 35, 42]. To the best of our knowledge, current methods require an explicit description of charts on the underlying manifold and adopt a completely extrinsic point of view. This process involves manual computations on paper on a per-case basis. For instance, the authors of [34, 35] present methods for three specific cases: the space of positive definite matrices, Grassman, and Stiefel manifolds. We mention the method in [7, Chapter 7.7] in particular, which is most closely related to ours. This method adopts an extrinsic", + "bbox": [ + 496, + 712, + 895, + 878 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "*https://github.com/NKI-AI/CERM", + "bbox": [ + 517, + 886, + 709, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24099", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "perspective on RSGD and effectively performs all computations in the ambient (vector) space in which $\\mathcal{M}$ is embedded. This extrinsic point of view leads to an algorithm for \"projecting\" paths in the ambient space onto descent paths on $\\mathcal{M}$ . This projection involves an additional nonlinear optimization problem and presents highly non-trivial issues for numerical implementation on general manifolds, as mentioned by the authors. In contrast, the method we present here is entirely intrinsic to $\\mathcal{M}$ and uses a local coordinate system amenable to numerical computations, allowing us to deal with a general class of manifolds. In particular, it allows us to incorporate the constraints for the wavelet straightforwardly. Most importantly, we follow a specific descent path on $\\mathcal{M}$ ; the geodesic in the direction of the negative gradient of the objective, a process made possible through our explicit understanding of the Riemannian metric. Our intrinsic perspective ensures that all computations are performed directly on the manifold, avoiding additional optimization problems. In particular, the dynamics of our optimization algorithm are solely driven by the gradient flow of the objective.", + "bbox": [ + 75, + 90, + 472, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview The contributions of this paper are ordered as follows. In Sec. 2, we introduce the theory of the constrained optimization. In Sec. 3, we apply the framework to our topic of interest: constraining filters of a CNN to form wavelets. We will show experiments in Sec. 4. In particular, this section will show our novel application, where we find data-driven wavelets for contour prediction. Specifically, we use wavelet networks to perform contour prediction in the medical domain, where we outperform strong baselines.", + "bbox": [ + 75, + 426, + 470, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We remark that an extensive, formal, and detailed investigation of the CERM framework, MRA, comparisons with other frameworks, and several implementation details are provided in [16], to which we refer the reader for more details.", + "bbox": [ + 75, + 563, + 470, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Constrained Empirical Risk Minimization", + "text_level": 1, + "bbox": [ + 76, + 645, + 454, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce a framework for performing ERM with constraints, which we will refer to as Constrained Empirical Risk Minimization (CERM). We start with a brief review of the traditional ERM setup [36, 37] introducing the necessary terminology, notation, and assumptions. Next, we explain how to incorporate constraints into the ERM framework in the form of a system of equations. We provide sufficient conditions on the system of equations to guarantee that the solution set is a Riemannian manifold. Finally, we provide an explanation for how the Riemannian metric and associated geometric quantities can be numerically evaluated, which enables us to perform SGD on the Riemannian manifold of interest directly. The discussion will be high-level; each step is explored in full mathematical detail in Sec. 6 [16].", + "bbox": [ + 75, + 674, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. ERM Setup", + "text_level": 1, + "bbox": [ + 500, + 90, + 627, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $X$ and $Y$ be random variables whose realizations are interpreted as input data and corresponding targets, respectively. The sample spaces of $X$ and $Y$ are denoted by $\\mathcal{X}$ and $\\mathcal{Y}$ , respectively. As a side note, self-supervised settings also fall into this framework, in which case the target $Y$ is created on the fly as a function of $X$ .", + "bbox": [ + 496, + 113, + 893, + 204 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Empirical risk minimization The goal of the ERM framework is to find a map $G: \\mathcal{X} \\to \\mathcal{Y}$ such that $G(x) \\approx y$ for most realizations of $(X, Y)$ . The discrepancy between predicted and observed targets is quantified using a loss function $L: \\mathcal{Y} \\times \\mathcal{Y} \\to [0, \\infty)$ . Without loss of generality, we assume that $L$ assumes positive values only and that $L$ decreases as the accuracy of predictions increases. The main objective of ERM is to find an optimal map $G^*$ , which solves the minimization problem", + "bbox": [ + 496, + 222, + 893, + 358 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {G \\in \\mathcal {G}} \\mathbb {E} (L (G (X), Y)). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 371, + 890, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $\\mathcal{G}$ is a suitable subset of functions.", + "bbox": [ + 498, + 404, + 763, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In all our applications, we assume that $X$ and $Y$ are random vectors with sample spaces $\\mathcal{X} = \\mathbb{R}^n$ and $\\mathcal{Y} = \\mathbb{R}^m$ . Furthermore, we assume $\\mathcal{G}$ is a parametric set that consists of mappings $\\mathcal{G} = \\{G(\\cdot ,\\xi):\\xi \\in \\mathbb{R}^p\\}$ , where $G:\\mathbb{R}^n\\times \\mathbb{R}^p\\to \\mathbb{R}^m$ is a continuously differentiable map. This includes, for example, ordinary neural networks, where the parameters are the weights. With these assumptions in place, the minimization problem in (1) is equivalent to minimizing over the parameters $\\xi$ instead. In practice, we have access to only a finite set of observations, which are assumed to be i.i.d., and the objective in (1) is replaced by an arithmetic average.", + "bbox": [ + 496, + 419, + 893, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Imposing constraints", + "text_level": 1, + "bbox": [ + 500, + 609, + 697, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we explain how to incorporate constraints on a subset of the parameters $\\xi$ directly into the ERM framework. We consider constraints given in the form of a system of equations. To be very precise, let $F:\\mathbb{R}^{\\tilde{p}}\\to \\mathbb{R}^{q}$ be a twice continuously differentiable map, where $\\tilde{p}$ denotes the number of constrained parameters, and $q$ is the number of equations. We assume that $q < \\tilde{p}\\leq p$ . For notational convenience, we decompose $\\mathbb{R}^p = \\mathbb{R}^{p - \\tilde{p}}\\oplus \\mathbb{R}^{\\tilde{p}}$ , where the first and second subspaces correspond to the unconstrained and constrained parameters, respectively. We take $\\pi_{p - \\tilde{p}}:\\mathbb{R}^p\\to \\mathbb{R}^{p - \\tilde{p}}$ and $\\pi_{\\tilde{p}}:\\mathbb{R}^p\\to \\mathbb{R}^{\\tilde{p}}$ to be the projections onto the unconstrained and constrained parameter subspace, respectively. We will denote the unconstrained and constrained parameters by $\\alpha$ and $\\theta$ , respectively, i.e., $\\alpha = \\pi_{p - \\tilde{p}}(\\xi)$ and $\\theta = \\pi_{\\tilde{p}}(\\xi)$ .", + "bbox": [ + 496, + 633, + 893, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2.1 (CERM). Let $G: \\mathbb{R}^n \\times \\mathbb{R}^p \\to \\mathbb{R}^m$ be a continuously differentiable parameterization of mappings and $F: \\mathbb{R}^{\\tilde{p}} \\to \\mathbb{R}^q$ a twice continuously differentiable constraint,", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24100", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $q < \\tilde{p} \\leq p$ . Suppose $L: \\mathbb{R}^m \\times \\mathbb{R}^m \\to [0, \\infty)$ is a continuously differentiable loss function. The constrained ERM problem for $(X, Y)$ with respect to $(G, F, L)$ is defined by", + "bbox": [ + 75, + 90, + 472, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\min _ {\\xi \\in \\mathbb {R} ^ {p}} \\mathbb {E} (L (G (X, \\xi), Y)), \\\\ \\text {s . t .} F (\\pi_ {\\tilde {p}} (\\xi)) = 0. \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 157, + 468, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note the generality of the admissible mappings $G$ in Definition 2.1. Although we will focus on neural networks from now on, the proposed framework applies to any parametric model, e.g., logistic or polynomial regression models. Next, we show that the CERM problem in (2) can be reformulated as an ordinary ERM problem on a Riemannian manifold $(\\mathcal{N}, g_{\\mathcal{N}})$ , provided that the system of equations satisfies a mild non-degeneracy condition. This result allows us to consider the admissible parameters as a geometric object in its own right, whose intrinsic geometry we use to solve (2). The proof of this theorem can be found in Sec. 6.1 [16].", + "bbox": [ + 75, + 214, + 472, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 2.2. If zero is a regular (i.e., non-singular) value of $F$ , then the CERM problem in (2) is equivalent to solving an ordinary ERM problem on a Riemannian manifold $\\mathcal{N}$ with metric $g_{\\mathcal{N}}$ of dimension $p - q$ . Here $\\mathcal{N} = \\mathbb{R}^{p - \\tilde{p}} \\times \\mathcal{M}$ is an embedded $C^2$ -submanifold of $\\mathbb{R}^p$ and $\\mathcal{M} := F^{-1}(0)$ . The equivalent minimization problem is given by", + "bbox": [ + 75, + 388, + 472, + 481 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {(\\alpha , \\theta) \\in \\mathcal {N}} \\mathbb {E} (L (G (X, \\alpha \\oplus \\iota (\\theta)), Y)), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 489, + 468, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\iota : \\mathcal{M} \\to \\mathbb{R}^{\\tilde{P}}$ is the inclusion map which embeds $\\mathcal{M}$ in the ambient space.", + "bbox": [ + 75, + 525, + 468, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From now on, we refer to the objective in (3) as simply the loss $\\mathcal{L}$ .", + "bbox": [ + 75, + 565, + 468, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Graph coordinates on $\\mathcal{N}$ The main idea is to directly minimize the loss on $\\mathcal{N}$ using SGD, see Figure 1 for an overview. To perform SGD on this space, however, we need to choose a coordinate system in which we can evaluate the gradient of the loss and subsequently follow descent trajectories on $\\mathcal{N}$ . Now, the proof of Theorem 2.2 involves showing that $\\mathcal{M}$ is an embedded submanifold of the ambient space $\\mathbb{R}^{\\tilde{p}}$ . This is accomplished by constructing a special coordinate system, a so-called graph chart or graph coordinate system. This chart is uniquely determined by the constraint $F$ and a prescribed zero $\\theta \\in \\mathcal{M}$ . It can be numerically evaluated at the point $\\theta$ by solving a linear system of equations. In turn, this enables us to construct a chart on the manifold in which we can evaluate the Riemannian metric and the gradient $\\nabla_{g_{\\mathcal{N}}} \\mathcal{L}(\\alpha, \\theta)$ .", + "bbox": [ + 75, + 613, + 472, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The crucial observation is that we can numerically construct coordinate systems in which we can follow paths on the manifold $\\mathcal{N}$ in the direction of $-\\nabla_{g_{\\mathcal{N}}}\\mathcal{L}(\\alpha, \\theta)$ . In particular, the system of equations determining these", + "bbox": [ + 75, + 840, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Sketch of the computation of $\\nabla_{g_N}\\mathcal{L}(\\alpha ,\\theta)$ given $(\\alpha ,\\theta)\\in \\mathcal{N}$ . See Algorithm 2 in Sec. 6.5 [16] the detailed formal version.", + "bbox": [ + 498, + 90, + 893, + 135 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Compute the derivative $DF(\\theta)$ of the constraint $F$ .", + "2: Compute chart-related derivatives.", + "3: Compute the metric $g_{\\mathcal{N}}$ using the chart-related derivatives.", + "4: Compute the partial derivatives of the loss $\\mathcal{L}$ with respect to the constrained components $\\theta$ .", + "5: Compute the components of $\\nabla_{g_{\\mathcal{M}}}\\mathcal{L}(\\alpha ,\\theta)$ using the above ingredients.", + "6: Compute the ordinary (unconstrained) components $\\nabla_{g_{\\mathrm{flat}}}\\mathcal{L}(\\alpha ,\\theta)$ by evaluating $D_{\\alpha}\\mathcal{L}(\\alpha ,\\theta)$" + ], + "bbox": [ + 508, + 138, + 893, + 291 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "so-called graph coordinate systems is fully determined by the derivative $DF$ of the constraint $F$ and the constraint itself. This observation enables a general implementation of graph coordinate systems independent of the explicit form of $F$ . We only require an implementation of $F$ and $DF$ , where the latter may be obtained using automatic differentiation. We provide a sketch of the algorithm for computing the gradient in Algorithm 1. For the interested reader, we mention that a complete and detailed explanation of all steps can be found in Sec. 6 [16]. Before moving on, we mention the limitations of the CERM framework for other generic constraints.", + "bbox": [ + 496, + 323, + 893, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Limitations As with most applications of deep neural networks, the available amount of computational power imposes limitations. Specifically, if the number of constrained parameters is large relative to the number of constraints, the computational demands will rise. Furthermore, setting up well-posed constraints can be a non-trivial task, e.g., symmetries in the constraints may result in redundancies obstructing the solvability of the associated systems of equations. Finally, the topology of the constrained submanifold may pose limitations as well. For instance, while the constraints will generically define a Riemannian manifold, it need not be connected. In such cases, the initialization of the network parameters will play an important role, as the gradient flow will stay on the connected component it was initialized on.", + "bbox": [ + 496, + 541, + 893, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Multiresolution Analysis and CERM", + "text_level": 1, + "bbox": [ + 498, + 768, + 831, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section presents the application of the aforementioned CERM framework to a non-trivial application to learn optimal wavelet bases for a given task. Many different types of wavelets, such as Haar, Daubechies, Gabor, etc., are used for a wide array of tasks. The specifics of the task demand different characteristics of the wavelets, such as the degree of smoothness, or the support of the wavelet. For example,", + "bbox": [ + 496, + 794, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24101", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the CDF wavelets [10] are well-known due to their usage for both lossless and lossy compression in JPEG2000. In this section, we consider networks that will find task-driven wavelets, optimizing the wavelet characteristics to the task at hand.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we will explain how to set up a system of equations (constraints) whose solution set corresponds to wavelets. We will apply these constraints in the next section to learn task-driven wavelets for predicting multi-resolution decompositions of contours in the medical domain. While we only consider one-dimensional wavelets in this work, we stress that the framework is easily adapted to higher-dimensional domains, to decompose images, for instance, by using tensor products of one-dimensional bases.", + "bbox": [ + 75, + 167, + 470, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Multiresolution Analysis", + "text_level": 1, + "bbox": [ + 76, + 313, + 302, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we briefly state what a Multiresolution Analysis (MRA) is. We closely follow the standard exposition, found in, e.g., [29, 31], and refer the reader to Sec. 7 [16] for a more in-depth and formal introduction. The MRA framework allows one to study a signal $\\gamma \\in L^2 (\\mathbb{R})$ at various resolution levels. The origin of this framework lies in the uncertainty principle of Fourier analysis, which states that a signal cannot be simultaneously localized in the time and frequency domain. MRA addresses this shortcoming by decomposing a signal on different discrete resolution levels. The idea is to construct a nested (exhausting) sequence of closed subspaces $V_{j}\\subset L^{2}(\\mathbb{R})$ , each associated with a specific resolution level $j$ . The subspaces $V_{j}$ are spanned by integer shifts of an appropriate dilation of a so-called scaling map (or father wavelet) $\\varphi$ . The level of dilation determines the resolution level. Formally, we require that $(\\varphi_{jk})_{k\\in \\mathbb{Z}}$ is an orthonormal basis for $V_{j}$ , where $\\varphi_{jk}(t)\\coloneqq 2^{\\frac{j}{2}}\\varphi (2^{j}t - k)$ . In Fig. 2a and Fig. 2c, we show an example of that Haar-MRA; in particular, we show the father wavelet and an example of a subspace at a certain resolution level.", + "bbox": [ + 75, + 338, + 472, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Decomposing a signal To analyze a signal $\\gamma$ at different resolution levels, we project it onto the subspaces $V_{j}$ . The orthogonal projection of $\\gamma$ at level $j$ is denoted by $\\gamma_{j}$ . The coefficients of $\\gamma_{j}$ with respect to the basis for $V_{j}$ , denoted by $a_{jk}(\\gamma)$ , are called the approximation coefficients of $\\gamma$ . When a signal in $V_{j+1}$ is projected onto $V_{j}$ , the lost information is stored in the orthogonal complement $W_{j}$ of $V_{j}$ in $V_{j+1}$ . This subspace is referred to as the detail subspace at level $j$ . The decomposition of $V_{j+1}$ into $V_{j}$ and $W_{j}$ allows the reconstruction of a signal at level $j+1$ from any lower level $j_{0}$ provided all the details in between are known. The corresponding decomposition and reconstruction algorithms give rise to the Discrete Wavelet Transform (DWT) and Mallat's Pyramid Algorithm. In Fig. 2b and Fig. 2d, we show examples of the mother wavelet and the detail subspace for the Haar wavelet.", + "bbox": [ + 75, + 643, + 470, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A fundamental result, known as Mallat's Theorem, states", + "bbox": [ + 96, + 885, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6b37e34bbbf1201ad4b129c2a50d686a0acb69edf924732db44394a509ffcf32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 88, + 687, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/85f0688522d302ac7328cc2eb5449b364d97056a6b0804ae3e0775972e8fdb3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 88, + 882, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2e28b94e07a13d486a1fbd061d54edd049b941fd65951b183825a66efe46e535.jpg", + "image_caption": [ + "(a) Father wavelet", + "(c) $V_{j}$", + "Figure 2. Example of the Haar MRA: (a) Dilated translation of the Haar scaling map (or father wavelet) $\\varphi = \\mathbf{1}_{[0,1)}$ . The approximation subspace $V_{j}$ at level $j$ consists of all step-functions with step-size $2^{-j}$ . (b) Dilated translation of the mother wavelet $\\psi = \\mathbf{1}_{[0,\\frac{1}{2})} - \\mathbf{1}_{[\\frac{1}{2},1)}$ . The detail subspace $W_{j}$ is spanned by integer translations of the dilated mother wavelet $\\psi_{j}$ . (c) The approximation subspace at level $j$ consists of all step-functions with step-size $2^{-j}$ . (d) Example of a function in the detail subspace at level $j$ ." + ], + "image_footnote": [], + "bbox": [ + 511, + 210, + 687, + 318 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7ec91ceaa1e072de490ec3a80c4c778e89db15023c12ae3664c503b375020418.jpg", + "image_caption": [ + "(b) Mother wavelet", + "(d) $W_{j}$" + ], + "image_footnote": [], + "bbox": [ + 705, + 210, + 879, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "that the subspaces $W_{j}$ can also be spanned by dilating and shifting a single map. More precisely, there exists a map $\\psi \\in W_0$ , the so-called mother wavelet, such that the $\\psi_{jk}$ form an orthonormal basis for $W_{j}$ , see [31]. The coefficients of the projection of $(\\gamma)$ onto $W_{j}$ , are denoted by $d_{j}(\\gamma) \\coloneqq (d_{jk}(\\gamma))_{k\\in \\mathbb{Z}}$ , and referred to as the detail coefficients of $\\gamma$ at resolution level $j$ . We can summarize the MRA procedure roughly as", + "bbox": [ + 496, + 497, + 890, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {j} = \\sum_ {k} a _ {j _ {0} k} \\varphi_ {j _ {0} k} + \\sum_ {j _ {0} \\leq l \\leq j - 1} \\sum_ {k} d _ {l k} \\psi_ {l k}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 628, + 890, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varphi$ denotes the father wavelet and $\\psi$ the mother wavelet. The MRA boils down to the specification of the first approximation coefficient and a number of detail coefficients.", + "bbox": [ + 496, + 670, + 893, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Low and high pass filters As we mentioned before, an MRA is completely determined by a scaling map $\\varphi$ , which in turn is completely characterized by a so-called low-pass filter. We will use this observation to set up a system of constraints characterizing a finite-dimensional family of wavelets, such that we can use it in the CERM framework. To explain how this works, we first make a key observation. Since $V_{0} \\subset V_{1}$ , there exists a unique sequence $h$ characterizing $\\varphi$ in terms of the basis $(\\varphi_{1k})_{k \\in \\mathbb{Z}}$ for $V_{1}$ . The sequence $h$ completely characterizes the scaling function and is called the low-pass filter or wavelet filter of the MRA. Similarly,", + "bbox": [ + 496, + 734, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24102", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "since $\\psi \\in W_0\\subset V_1$ , there exists a unique sequence $g$ , the so-called high-pass filter associated to $h$ , characterizing the mother wavelet. An example of an MRA is shown in Figure 2. In practice, to define an MRA, one only needs to specify an appropriate low-pass filter $h$ , as for Mallat's mother wavelet, we have $g_{k} = (-1)^{k - 1}h_{1 - k}$ . We refer the interested reader to the supplementary information for a detailed discussion of all these points [16].", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. MRA constraints", + "text_level": 1, + "bbox": [ + 76, + 223, + 246, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We are now ready to formulate and impose constraints on a sequence $h$ to ensure that it is the low-pass filter of a scaling map $\\varphi$ . In practice, we only consider finite sequences. For notational convenience, we introduce the space $\\mathcal{A}_M(\\mathbb{R})$ , which consists of one-dimensional real-valued two-sided sequences $(h_k)_{k=1-M}^{M-1}$ of order $M$ ; so $\\mathcal{A}_M(\\mathbb{R})$ is $2M - 1$ dimensional.", + "bbox": [ + 75, + 248, + 473, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 3.1. Let $M \\in \\mathbb{N}_{\\geq 3}$ be a prescribed order and define $F_M: \\mathcal{A}_M(\\mathbb{R}) \\to \\mathbb{R}^{M+1}$ by", + "bbox": [ + 75, + 366, + 472, + 398 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n(F_{M}(h))_{k}:= \\left\\{ \\begin{array}{ll}\\sum_{|l|\\leq M - 1}|h_{l}|^{2} - 1, & k = 0,\\\\ \\sum_{1 - M + 2k\\leq l\\leq M - 1}h_{l - 2k}h_{l}, & 1\\leq k\\leq M - 1,\\\\ -\\sqrt{2} +\\sum_{|l|\\leq M - 1}h_{l}, & k = M. \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 410, + 486, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "If $F_{M}(h) = 0$ , and $h$ satisfies a certain non-degeneracy condition, then $h$ is the low-pass filter of a scaling map $\\varphi$ generating an MRA of $L^2 (\\mathbb{R})$ .", + "bbox": [ + 75, + 527, + 470, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A full proof and detailed statement of the non-degeneracy condition can be found in Sec. 7.4 in the supplementary material. Heuristically, the first $M$ equations encode the orthonormality of the scaling map. The last equation encodes the observation that the Fourier transform of the scaling map must be nonzero. The set of regular points in $F_{M}^{-1}(0)$ is a real-analytic $(M - 2)$ -dimensional submanifold of $\\mathbb{R}^{2M - 1}$ . Hence, we can get as many degrees of freedom as desired, by choosing a sufficiently large order $M$ . We now have an end-to-end way of learning wavelets for specific tasks, a wavelet network, by using the constraint $F_{M}$ along with the CERM framework.", + "bbox": [ + 75, + 585, + 470, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 782, + 209, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we show some explicit examples of the wavelet networks with respect to contour prediction. In particular, we will utilize the derived wavelet filters of Sec. 3.2 to create a network for autocontouring. This task fits well with the setup of wavelets, as they naturally lend themselves to representing continuous objects such as curves.", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/77d0f409aa4320a120744f7d19520fee3386ba36cafbdbc793c0827a4e741b9b.jpg", + "image_caption": [ + "Figure 3. A schematic picture of our network. The encoder consists of residual convolutional blocks depicted in blue. The first residual convolutional block uses $n_f$ filters and is doubled after every other residual convolutional block. Attached to the encoder are fully connected layers to predict approximation and detail coefficients at the lowest resolution level $j_0$ . The approximation and detail coefficients are supplied as input to the Pyramid Algorithm (the decoder) to predict a contour on high-resolution level. Each green block corresponds to an inverse discrete wavelet transform. Detail coefficients at higher levels are computed using skip-connections (arrows in red). We only predict detail coefficients up to level $j_1$ . In this example, we have set $j_1 = j_0 + 2$ and $j_2 = j_0 + 3$ . In reality, the decoder consists of two upsampling paths, one for each spatial component of the curve. We have only drawn one for notational convenience. During training, only the approximation coefficients at the highest resolution level are supervised." + ], + "image_footnote": [], + "bbox": [ + 524, + 90, + 870, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Contour Prediction using MRA", + "text_level": 1, + "bbox": [ + 500, + 546, + 777, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the context of contour prediction, wavelets can be used to represent the boundary of a region in an image using a simple closed curve. We consider two-dimensional gray-valued images $x \\in \\mathcal{X} \\coloneqq [0,1]^{n \\times n}$ . For each of our applications, we assume that these images contain an organ with a boundary. Formally speaking, this means that each image $x$ contains a (uniquely identifiable) simply connected region $R(x)$ , the organ, with boundary $\\partial R(x)$ . This boundary of the region, $\\partial R(x)$ , can be parameterized by a simple closed, piecewise $C^2$ , curve $\\gamma(x)$ . We will develop a deep learning framework for computing such parameterizations by learning a multiresolution decomposition of $\\gamma(x)$ using the methods developed in the previous sections.", + "bbox": [ + 496, + 570, + 893, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1 Data", + "text_level": 1, + "bbox": [ + 500, + 786, + 588, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We have used public datasets from the Medical Decathlon Challenge [3]. The selected data consists of CT scans of the spleen of size $512 \\times 512$ and T2-weighted MRI images of the prostate central gland, henceforth abbreviated as just the prostate, of size $320 \\times 320$ . The scans were cropped to size $224 \\times 224$ and $192 \\times 192$ , respectively. Further-", + "bbox": [ + 496, + 809, + 895, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24103", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4414cf31382f8d2718b2edefad30e87caf76fe60887c0a05341613a2529f9256.jpg", + "image_caption": [ + "Figure 4. Examples of predictions on the test set for the spleen and prostate, depicted in the first and second row, respectively, for the best-performing wavelet models. The green curve corresponds to the ground truth, while the red curve is a prediction made by the wavelet network." + ], + "image_footnote": [], + "bbox": [ + 223, + 88, + 748, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6c3cf0900468a100ab5a98f5a8092c9a42251f905e7a2f5e04ec9a840d0ec2f9.jpg", + "image_caption": [ + "(a) Initial father and mother wavelet" + ], + "image_footnote": [], + "bbox": [ + 99, + 330, + 450, + 448 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ce08152be65b95f01d7a0a631d6bfc4fec432d3559d5226bdbba0d16e8b5cf8.jpg", + "image_caption": [ + "(b) Task-optimized father and mother wavelet", + "Figure 5. Example of wavelets of order 5 learned during training of the spleen model: (a) The initial father (left) and mother (right) wavelets for the second spatial component. (b) The task-optimized father and mother wavelets after training for the second spatial component. We observe that the task-optimized wavelets are more simple and have \"cleaned up\" in a sense." + ], + "image_footnote": [], + "bbox": [ + 101, + 465, + 449, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "more, the images were resampled to the median sample spacing, which resulted in $(5.00\\mathrm{mm}, 0.793\\mathrm{mm}, 0.793\\mathrm{mm})$ and $(3.6\\mathrm{mm}, 0.625\\mathrm{mm}, 0.625\\mathrm{mm})$ spacings for the spleen and prostate, respectively. We refer the reader to Sec. 8.2 [16] for a detailed description of all the preprocessing steps and construction of ground truth curves. Examples of predictions on the test set, for both the spleen and prostate, are shown in Figure Fig. 4.", + "bbox": [ + 75, + 780, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.2 Model and Training", + "text_level": 1, + "bbox": [ + 500, + 334, + 694, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model objective We parameterize the boundary $\\partial R(x)$ by arc length, resulting in a curve $\\gamma (x)$ . The precise details of this parametrization are outlined in the supplementary information. The objective is to compute the relevant approximation coefficients of $\\gamma (x)$ . To this end, let $j_0,j_1,j_2\\in \\mathbb{N}$ be resolution levels, where $j_0\\leq j_1\\leq j_2$ . We will construct a convolutional neural network", + "bbox": [ + 498, + 357, + 893, + 463 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nG: \\mathcal {X} \\times \\mathbb {R} ^ {p} \\to \\prod_ {j = j _ {0}} ^ {j _ {2}} \\mathbb {R} ^ {2 ^ {j}} \\times \\mathbb {R} ^ {2 ^ {j}} \\times \\prod_ {j = j _ {0}} ^ {j _ {1} - 1} \\mathbb {R} ^ {2 ^ {j}} \\times \\mathbb {R} ^ {2 ^ {j}},\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 470, + 864, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which predicts the wavelet decomposition of $\\gamma (x)$ . Here the subspaces $\\mathbb{R}^{2^j}$ correspond to approximation and detail coefficients at level $j$ , one for each spatial component. Furthermore, a subset of the parameters $\\xi \\in \\mathbb{R}^p$ are constrained to be wavelet filters, one wavelet filter per spatial component, using Theorem 3.1. The map $G(\\cdot ,\\xi)$ applied to an image $x$ has output", + "bbox": [ + 498, + 523, + 893, + 632 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nG (x, \\xi) = \\left(a _ {j _ {0}} (x, \\xi), \\dots , a _ {j _ {2}} (x, \\xi), \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 641, + 890, + 665 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nd _ {j _ {0}} (x, \\xi), \\dots , d _ {j _ {1} - 1} (x, \\xi)).\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 660, + 823, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here $a_{j}(x,\\xi)$ and $d_{j}(x,\\xi)$ are predictions for the approximation and detail coefficients of $\\gamma (x)$ at level $j$ , respectively. We only predict detail coefficients up to level $j_{1}$ .", + "bbox": [ + 500, + 686, + 893, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Architecture The network is a hybrid analog of the U-Net. It consists of a two-dimensional convolutional encoder, a bottleneck of fully connected layers, and a one-dimensional decoder. The encoder and decoder are connected through skip-connections. The approximation and detail coefficients at the lowest resolution level $j_0$ are predicted in the bottleneck. Afterward, the Pyramid Algorithm takes over to compute approximation coefficients at higher resolution levels (the decoder) using learnable wavelet filters. In Fig. 3, we provide a schematic overview of the network architecture.", + "bbox": [ + 498, + 750, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24104", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Loss To measure the discrepancy between the ground truth and the predicted curve, we define", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r l} L (G (x, \\xi), \\tilde {a} (\\gamma (x))) := & \\| [ a _ {j _ {2}} ] _ {1} - \\tilde {a} _ {j _ {2}} ([ \\gamma (x) ] _ {1}) \\| _ {2} + \\\\ & \\| [ a _ {j _ {2}} ] _ {2} - \\tilde {a} _ {j _ {2}} ([ \\gamma (x) ] _ {2}) \\| _ {2}. \\end{array} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 137, + 468, + 172 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The $\\tilde{a}$ thus correspond to the ground-truth approximation coefficients, and the subscripts 1, 2 denote the spatial dimension. In other words, the loss corresponds to the componentwise $L^2$ -error between the curves on resolution level $j_2$ with approximation coefficients $a_{j_2}(x,\\xi)$ and $\\tilde{a}_{j_2}(\\gamma (x))$ .", + "bbox": [ + 75, + 188, + 470, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Optimization We use two different optimizers during training: one for the encoder and MLP network (unconstrained), and one for the decoder (constrained wavelet network). The reason for this is that the needed step sizes on the non-trivial submanifold may significantly differ from those on the unconstrained (flat) parameter space. For the training of the model, we use plain SGD for the first eight epochs for both the constrained and unconstrained parameters. During this period, the learning rate for the unconstrained parameters is linearly increased from $10^{-5}$ to $2 \\cdot 10^{-4}$ . The learning rate for the constrained parameters (wavelet filters) is linearly increased from $10^{-4}$ to $10^{-2}$ . After the initial warmup stage, we switch to the Adam optimizer for the unconstrained parameters. For both the constrained and unconstrained parameters, we use learning rate schedulers and decrease the learning rate by a factor of 0.85 if no significant improvements in the validation loss are observed during the last ten epochs. We train all models for 250 epochs using a batch size of 32 and use the last epoch for inference. The computations were performed in PYTORCH on a Geforce RTX 2080 Ti. An average training session takes about two and a half hours for the wavelet network and a day for the nnUNet. The inference times for the wavelet network were 12.5ms and 14ms for the prostate and spleen, respectively, using a batch size of 32. The inference times for the nnUNet were 17.8ms and 23.4ms for the prostate and spleen, respectively, using a batch size of 32.", + "bbox": [ + 75, + 296, + 472, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1.3 Results", + "text_level": 1, + "bbox": [ + 76, + 736, + 181, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have evaluated the performance of both our wavelet networks for different orders and a state-of-the-art baseline 2d-nnUNet [13] on the unseen test data, see Table 1. For both the spleen and prostate, we observe that the best wavelet networks outperform the baseline. We note that the task-optimized wavelets differ significantly from the wavelets randomly initialized at the start of training. A comparison of an initial and task-optimized wavelet is depicted in Figure 5, see Sec. 8.4 for more examples.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1. Mean and standard deviation (in parentheses) of the dice score on the unseen test sets. The number of parameters in our wavelet-based networks ranged from 5M to 10M parameters, while the nnUNet used approximately 40M and 18M parameters for the spleen and prostate, respectively.", + "bbox": [ + 496, + 88, + 890, + 159 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5f676562392b7a916920dca2059bfc3d26403f4385cf9dc6f2d628fe1bc12f2b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MODELDICE SPLEENDICE PROSTATE
NNUNET0.914 (1.74 · 10-1)0.896 (1.27 · 10-1)
ORDER 30.911 (7.38 · 10-2)0.929 (4.66 · 10-2)
ORDER 40.911 (7.85 · 10-2)0.935 (3.48 · 10-2)
ORDER 50.916 (6.94 · 10-2)0.935 (4.11 · 10-2)
ORDER 60.917 (7.17 · 10-2)0.934 (4.14 · 10-2)
ORDER 70.921 (6.91 · 10-2)0.928 (4.03 · 10-2)
ORDER 80.919 (6.64 · 10-2)0.934 (3.62 · 10-2)
", + "bbox": [ + 522, + 183, + 867, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 335, + 625, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have introduced the CERM framework for creating task-driven wavelets. Although our aim is to find such wavelets, in principle the framework lends itself to imposing generic constraints on parametric models. The wavelet constraints can be formulated as a finite system of equations. Under mild smoothness and non-degeneracy conditions, the networks can be made to obey the constraints exactly throughout the entire training procedure by performing SGD on a curved space. We can then utilize these wavelet constraints in a convolutional network, where the filters of the network now parameterize wavelets themselves. We have applied these wavelet networks to the prediction of boundaries of simply connected regions in medical images, where they outperform strong baselines.", + "bbox": [ + 496, + 361, + 893, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 588, + 596, + 603 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] P. Abry and P. Flandrin. On the initialization of the discrete wavelet transform algorithm. IEEE Signal Processing Letters, 1(2):32-34, 1994. 14", + "[2] Brandon Amos, Lei Xu, and J. Zico Kolter. Input convex neural networks. In Proceedings of the 34th International Conference on Machine Learning, pages 146-155. PMLR, 2017. 2", + "[3] Michela Antonelli, Annika Reinke, Spyridon Bakas, Keyvan Farahani, Annette Kopp-Schneider, Bennett A Landman, Geert Litjens, Bjoern Menze, Olaf Ronneberger, Ronald M Summers, et al. The medical segmentation decathlon. Nature communications, 13(1):1-13, 2022. 6", + "[4] Randall Balestriero and Yann LeCun. Police: Provably optimal linear constraint enforcement for deep neural networks, 2022. 2", + "[5] D.P. Bertsekas and W. Rheinboldt. Constrained Optimization and Lagrange Multiplier Methods. Elsevier Science, 2014. 1", + "[6] Silvere Bonnabel. Stochastic gradient descent on riemannian manifolds. IEEE Transactions on Automatic Control, 58(9): 2217-2229, 2013. 2, 1, 5" + ], + "bbox": [ + 506, + 613, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24105", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[7] N. Boumal. An Introduction to Optimization on Smooth Manifolds. Cambridge University Press, 2023. 2", + "[8] Michael M Bronstein, Joan Bruna, Yann LeCun, Arthur Szlam, and Pierre Vandergheynst. Geometric deep learning: going beyond euclidean data. IEEE Signal Processing Magazine, 34(4):18-42, 2017. 2", + "[9] Michael M Bronstein, Joan Bruna, Taco Cohen, and Petar Velicković. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges. arXiv preprint arXiv:2104.13478, 2021. 2", + "[10] Albert Cohen, Ingrid Daubechies, and J-C Feauveau. Biorthogonal bases of compactly supported wavelets. Communications on pure and applied mathematics, 45(5):485-560, 1992. 5", + "[11] Priya Donti, David Rolnick, and J Zico Kolter. Dc3: A learning method for optimization with hard constraints. In International Conference on Learning Representations, 2021. 2", + "[12] M.W. Frazier. An Introduction to Wavelets Through Linear Algebra. Springer New York, 2001. 12", + "[13] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 8", + "[14] Hiroyuki Kasai, Pratik Jawanpuria, and Bamdev Mishra. Riemannian adaptive stochastic gradient algorithms on matrix manifolds. In International Conference on Machine Learning, pages 3262-3271. PMLR, 2019. 2, 1, 5", + "[15] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings. International Conference on Learning Representations, ICLR, 2015. 1", + "[16]FirstName LastName. Task-driven wavelets using constrained empirical risk minimization, 2023. Supplied as supplemental material supplementary-notes.pdf. 3, 4, 5, 6, 7", + "[17] Yann LeCun, Yoshua Bengio, et al. Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks, 3361(10):1995, 1995. 1", + "[18] John M Lee. Riemannian manifolds: an introduction to curvature. Springer Science & Business Media, 2006. 6, 8", + "[19] John M Lee. Smooth manifolds. In Introduction to smooth manifolds. Springer, 2013. 2, 6, 8", + "[20] Benedict Leimkuhler, Timothee Pouchon, Tiffany Vlaar, and Amos Storkey. Constraint-based regularization of neural networks. arXiv preprint arXiv:2006.10114, 2020. 2", + "[21] Benedict Leimkuhler, Tiffany Vlaar, Timothee Pouchon, and Amos Storkey. Better training using weight-constrained stochastic dynamics. arXiv preprint arXiv:2106.10704, 2021. 2", + "[22] Qiufu Li, Linlin Shen, Sheng Guo, and Zhihui Lai. Wavelet integrated cnns for noise-robust image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7245–7254, 2020. 2", + "[23] Cho-Chun Liu, Yaohui Liu, Zhengding Qiu, and Xiyu Du. On the initialization of the discrete wavelet transform algorithm. 1997 IEEE International Conference on Intelligent Processing Systems (Cat. No.97TH8335), 2:1220-1222 vol.2, 1997. 14" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 1", + "[25] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam, 2018. 1", + "[26] Stephane Mallat. A Wavelet Tour of Signal Processing, Third Edition: The Sparse Way. Academic Press, Inc., USA, 3rd edition, 2008. 2, 10", + "[27] Pablo Márquez-Neila, Mathieu Salzmann, and Pascal Fua. Imposing hard constraints on deep networks: Promises and limitations. arXiv preprint arXiv:1706.02025, 2017. 2", + "[28] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In International Conference on Learning Representations, 2018. 2", + "[29] L. Montefusco and L. Puccio. Wavelets: Theory, Algorithms, and Applications. Elsevier Science, 2014. 5, 11, 12, 14", + "[30] Deepak Pathak, Philipp Krahenbuhl, and Trevor Darrell. Constrained convolutional neural networks for weakly supervised segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1796-1804, 2015. 2", + "[31] M.C. Pereyra and L.A. Ward. Harmonic Analysis: From Fourier to Wavelets. American Mathematical Society, 2012. 5, 11, 12", + "[32] Jack Richter-Powell, Yaron Lipman, and Ricky T. Q. Chen. Neural conservation laws: A divergence-free perspective, 2022. 2", + "[33] David W Romero, Erik J Bekkers, Jakub M Tomczak, and Mark Hoogendoorn. Wavelet networks: Scale equivariant learning from raw waveforms. arXiv preprint arXiv:2006.05259, 2020. 2", + "[34] Soumava Kumar Roy, Zakaria Mhammedi, and Mehrtash Harandi. Geometry aware constrained optimization techniques for deep learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4460-4469, 2018. 2, 1, 5", + "[35] Hiroyuki Sato, Hiroyuki Kasai, and Bamdev Mishra. Riemannian stochastic variance reduced gradient algorithm with retraction and vector transport. SIAM Journal on Optimization, 29(2):1444-1472, 2019. 2, 1, 5", + "[36] Vladimir Vapnik. Principles of risk minimization for learning theory. Advances in neural information processing systems, 4, 1991. 3", + "[37] Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 1999. 3", + "[38] D.F. Walnut. An Introduction to Wavelet Analysis. Birkhäuser Boston, 2002. 12, 15", + "[39] Maurice Weiler, Patrick Forre, Erik Verlinde, and Max Welling. Coordinate independent convolutional networks-isometry and gauge equivariant convolutions on riemannian manifolds. arXiv preprint arXiv:2106.06020, 2021. 2", + "[40] Yang You, Jing Li, Sashank Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes. In International Conference on Learning Representations, 2020. 1", + "[41] Jason J Yu, Konstantinos G Derpanis, and Marcus A Brubaker. Wavelet flow: Fast training of high resolution normalizing" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24106", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "flows. Advances in Neural Information Processing Systems, 33:6184-6196, 2020. 2", + "[42] Hongyi Zhang, Sashank J. Reddi, and Suvrit Sra. Riemannian svrg: Fast stochastic optimization on riemannian manifolds. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 2, 1, 5" + ], + "bbox": [ + 78, + 90, + 470, + 176 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24107", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/d6dc5c43-4c52-4779-8c6a-d513a47a71be_model.json b/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/d6dc5c43-4c52-4779-8c6a-d513a47a71be_model.json new file mode 100644 index 0000000000000000000000000000000000000000..beb05f7cb13264c91167d6e418effaf1cf8955e5 --- /dev/null +++ b/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/d6dc5c43-4c52-4779-8c6a-d513a47a71be_model.json @@ -0,0 +1,1925 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.13, + 0.131, + 0.842, + 0.153 + ], + "angle": 0, + "content": "Task-Driven Wavelets using Constrained Empirical Risk Minimization" + }, + { + "type": "text", + "bbox": [ + 0.193, + 0.18, + 0.775, + 0.2 + ], + "angle": 0, + "content": "Eric Marcus\\*,1,2, Ray Sheombarsing\\*,1, Jan-Jakob Sonke\\(^{1}\\), Jonas Teuwen\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.216, + 0.713, + 0.253 + ], + "angle": 0, + "content": "1Netherlands Cancer Institute, 2University of Amsterdam {e.marcus, r.sheombarsing, j.sonke, j.teuwen}@nki.nl" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.622 + ], + "angle": 0, + "content": "Deep Neural Networks (DNNs) are widely used for their ability to effectively approximate large classes of functions. This flexibility, however, makes the strict enforcement of constraints on DNNs a difficult problem. In contexts where it is critical to limit the function space to which certain network components belong, such as wavelets employed in Multi-Resolution Analysis (MRA), naive constraints via additional terms in the loss function are inadequate. To address this, we introduce a Convolutional Neural Network (CNN) wherein the convolutional filters are strictly constrained to be wavelets. This allows the filters to update to task-optimized wavelets during the training procedure. Our primary contribution lies in the rigorous formulation of these filters via a constrained empirical risk minimization framework, thereby providing an exact mechanism to enforce these structural constraints. While our work is grounded in theory, we investigate our approach empirically through applications in medical imaging, particularly in the task of contour prediction around various organs, achieving superior performance compared to baseline methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.654, + 0.21, + 0.669 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.472, + 0.876 + ], + "angle": 0, + "content": "Empirical risk minimization (ERM) is currently the most prevalent framework for supervised learning. The goal is to find a function to map inputs to associated targets for all representative (potentially unseen) observations. To find such a mapping, one introduces a loss function to quantify the discrepancy between observed and predicted targets. An optimal map is then found by minimizing the expected loss. In large-scale settings, such as deep learning, the resulting minimization problem is solved using Stochastic Gradient Descent (SGD) and various variants thereof, see [15, 24, 25, 40] for instance. As deep learning applications become more specialized, domain-specific needs become increasingly vital. These are often formulated in terms of constraints on" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.304, + 0.816, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.557, + 0.895, + 0.793 + ], + "angle": 0, + "content": "Figure 1. An overview of the Constrained Empirical Risk Minimization (CERM) framework: on the top left side, the full gradient of the loss \\(\\mathcal{L}\\) is shown on the total space \\(\\mathcal{N}\\) (containing both the constrained and unconstrained space), with \\(\\alpha\\) are the unconstrained parameters, and \\(\\theta\\) the constrained ones. The constraints can be written in the form \\(F(\\theta) = 0\\); the solution set \\(F^{-1}(0)\\) is an embedded submanifold \\(\\mathcal{M}\\) of \\(\\mathbb{R}^{\\tilde{p}}\\). The constrained parameters are updated by following a path on \\(\\mathcal{M}\\) in the direction of the negative gradient on this manifold. This constrained part of the full gradient is contained in the tangent space \\(T_{\\theta}\\mathcal{M}\\) of the submanifold. The color of the manifold indicates the value of the loss function. By restricting the relevant components of our descent trajectories to the embedded submanifold, we always satisfy the constraints imposed by \\(F\\). The gradient and parameter updates for the unconstrained parameters \\(\\alpha\\) are computed as usual using standard SGD for flat space (not shown). In our wavelet scenario, the manifold thus consists of all admissible wavelets of a certain dimension." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "the permissible mappings. For instance, the constraint for translation-equivariance led to the development and success of convolution neural networks (CNNs) [17]. In general, however, it is a highly non-trivial task to construct network architectures that satisfy a set of constraints, if they exist" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.352, + 0.901 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24098" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "at all. It is, therefore, common practice to incorporate constraints directly into the loss function by including additional terms, usually referred to as \"soft\" constraints. This setup, however, has the drawback that the constraints are only approximately (on average) satisfied due to the formulation of ERM. Moreover, incorporating many different objectives in a loss function may lead to suboptimal results for the individual objectives. Another common strategy, used for many types of constraints, e.g., flip invariance, is based on data augmentation. However, not every constraint can be achieved using data augmentation, and it also leads, at best, to constraints being approximately satisfied. Furthermore, if we wish to constrain neural networks to specific function classes, such as wavelets, an additional loss term is no longer sufficient: a wavelet is only a wavelet whenever the constraints are satisfied exactly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.473, + 0.625 + ], + "angle": 0, + "content": "Approaches which circumvent loss-based (soft) constraints have been recently proposed, see [2, 4, 28, 30, 32] for example. Most notably, we mention the field of Geometric Deep Learning (GDL), which focuses on incorporating constraints arising from symmetries of the domain directly into the networks themselves, see [8, 9, 39] and the references therein. In GDL, one considers very specific but powerful types of constraints, namely that network layers are equivariant with respect to some group action. Not all constraints arise as equivariance principles, however. A large class of examples comes from highly specialized requirements on the output of a neural network, e.g., that the output is a divergence-free vector field, a contour, or perhaps a surface. For example, in medical image segmentation, a natural requirement is that the output of the segmentation network corresponds to a continuous (closed) curve. Creating architectures for such constraints is out of the realm of GDL and, in general, poses a challenging task; no prescribed methods exist to do so." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To facilitate these different types of constraints, we present a framework, called Constrained Empirical Risk Minimization (CERM), where we directly formulate and solve the constrained ERM problem in the space where the constraints are satisfied. This space forms, under mild conditions, a Riemannian manifold. As long as we stay on this manifold, the constraints will be satisfied exactly up to numerical precision. In particular, we explain how to perform SGD on such Riemannian manifolds arising from a finite-dimensional system of equations. Although we are motivated by the ability to constrain to wavelets, the framework, in principle, encompasses constraints of the form \\( F(W) = 0 \\), with \\( W \\) the network weights. Our method heavily relies on the Implicit Function Theorem, which is used to construct so-called graph-charts amenable to numerical computations. This allows for efficient evaluation of the (induced) Riemannian metric and gradients, which are vital for performing SGD on Riemannian manifolds using only the intrinsic geometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.896, + 0.485 + ], + "angle": 0, + "content": "As mentioned, we will present the tools and implementations of convolutional networks whose filters will form wavelets. There are several tasks at which one expects wavelet-based neural networks to excel. Wavelet decompositions naturally lend themselves to representing continuous objects such as curves, images, vector fields, or other higher-dimensional objects. Hence any task where the object of interest can be identified with a smooth or continuous function is well-suited for wavelet-based neural networks. There is an abundance of such examples to be found in computer vision, e.g., boundary prediction, image registration, and so forth. Another family of exciting applications can be found in signal analysis, e.g., in compression and denoising, where wavelets are long-standing tools that have proven to be extremely efficient [26]. The main idea in these areas is to extract information about noise, smoothness, and even singularities, through analysis of the wavelet coefficients. Subsequently, by modifying a subset of the coefficients, e.g., through thresholding, the signal can be cleaned up or denoised. More recently, wavelets have also received increasing attention in the context of neural networks. The applications are broad; some references include [22, 33, 41], ranging from efficient normalizing flows, noise-robust training, scale-equivariant models, and many more. The wavelet layers we create in this work are included in our code, which we will make available publicly*." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.895, + 0.713 + ], + "angle": 0, + "content": "Related work An earlier attempt at incorporating constraints is described in [11]. However, this method is only able to deal with linear constraints, which is hence incompatible with the wavelet constraints we will derive. Other works include [20, 21, 27], which are related to the method of Lagrange multipliers and have their optimization and training dynamics largely determined by variants of Newton's method. We discuss the differences between our SGD-compatible method and that of Lagrange multipliers in more detail in the supplementary material 6.2. A comprehensive overview of all constrained optimization methods, however, is out of the scope of this paper, and we therefore focus our attention to techniques based on SGD on Riemannian manifolds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.896, + 0.88 + ], + "angle": 0, + "content": "Performing SGD on Riemannian manifolds is well-established and has been studied extensively, see [6, 14, 34, 35, 42]. To the best of our knowledge, current methods require an explicit description of charts on the underlying manifold and adopt a completely extrinsic point of view. This process involves manual computations on paper on a per-case basis. For instance, the authors of [34, 35] present methods for three specific cases: the space of positive definite matrices, Grassman, and Stiefel manifolds. We mention the method in [7, Chapter 7.7] in particular, which is most closely related to ours. This method adopts an extrinsic" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.71, + 0.9 + ], + "angle": 0, + "content": "*https://github.com/NKI-AI/CERM" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24099" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.394 + ], + "angle": 0, + "content": "perspective on RSGD and effectively performs all computations in the ambient (vector) space in which \\(\\mathcal{M}\\) is embedded. This extrinsic point of view leads to an algorithm for \"projecting\" paths in the ambient space onto descent paths on \\(\\mathcal{M}\\). This projection involves an additional nonlinear optimization problem and presents highly non-trivial issues for numerical implementation on general manifolds, as mentioned by the authors. In contrast, the method we present here is entirely intrinsic to \\(\\mathcal{M}\\) and uses a local coordinate system amenable to numerical computations, allowing us to deal with a general class of manifolds. In particular, it allows us to incorporate the constraints for the wavelet straightforwardly. Most importantly, we follow a specific descent path on \\(\\mathcal{M}\\); the geodesic in the direction of the negative gradient of the objective, a process made possible through our explicit understanding of the Riemannian metric. Our intrinsic perspective ensures that all computations are performed directly on the manifold, avoiding additional optimization problems. In particular, the dynamics of our optimization algorithm are solely driven by the gradient flow of the objective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.472, + 0.563 + ], + "angle": 0, + "content": "Overview The contributions of this paper are ordered as follows. In Sec. 2, we introduce the theory of the constrained optimization. In Sec. 3, we apply the framework to our topic of interest: constraining filters of a CNN to form wavelets. We will show experiments in Sec. 4. In particular, this section will show our novel application, where we find data-driven wavelets for contour prediction. Specifically, we use wavelet networks to perform contour prediction in the medical domain, where we outperform strong baselines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.472, + 0.624 + ], + "angle": 0, + "content": "We remark that an extensive, formal, and detailed investigation of the CERM framework, MRA, comparisons with other frameworks, and several implementation details are provided in [16], to which we refer the reader for more details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.646, + 0.455, + 0.663 + ], + "angle": 0, + "content": "2. Constrained Empirical Risk Minimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.901 + ], + "angle": 0, + "content": "In this section, we introduce a framework for performing ERM with constraints, which we will refer to as Constrained Empirical Risk Minimization (CERM). We start with a brief review of the traditional ERM setup [36, 37] introducing the necessary terminology, notation, and assumptions. Next, we explain how to incorporate constraints into the ERM framework in the form of a system of equations. We provide sufficient conditions on the system of equations to guarantee that the solution set is a Riemannian manifold. Finally, we provide an explanation for how the Riemannian metric and associated geometric quantities can be numerically evaluated, which enables us to perform SGD on the Riemannian manifold of interest directly. The discussion will be high-level; each step is explored in full mathematical detail in Sec. 6 [16]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.628, + 0.108 + ], + "angle": 0, + "content": "2.1. ERM Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.895, + 0.205 + ], + "angle": 0, + "content": "Let \\( X \\) and \\( Y \\) be random variables whose realizations are interpreted as input data and corresponding targets, respectively. The sample spaces of \\( X \\) and \\( Y \\) are denoted by \\( \\mathcal{X} \\) and \\( \\mathcal{Y} \\), respectively. As a side note, self-supervised settings also fall into this framework, in which case the target \\( Y \\) is created on the fly as a function of \\( X \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.223, + 0.895, + 0.359 + ], + "angle": 0, + "content": "Empirical risk minimization The goal of the ERM framework is to find a map \\( G: \\mathcal{X} \\to \\mathcal{Y} \\) such that \\( G(x) \\approx y \\) for most realizations of \\( (X, Y) \\). The discrepancy between predicted and observed targets is quantified using a loss function \\( L: \\mathcal{Y} \\times \\mathcal{Y} \\to [0, \\infty) \\). Without loss of generality, we assume that \\( L \\) assumes positive values only and that \\( L \\) decreases as the accuracy of predictions increases. The main objective of ERM is to find an optimal map \\( G^* \\), which solves the minimization problem" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.372, + 0.892, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\min _ {G \\in \\mathcal {G}} \\mathbb {E} (L (G (X), Y)). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.405, + 0.764, + 0.419 + ], + "angle": 0, + "content": "Here \\(\\mathcal{G}\\) is a suitable subset of functions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.895, + 0.602 + ], + "angle": 0, + "content": "In all our applications, we assume that \\(X\\) and \\(Y\\) are random vectors with sample spaces \\(\\mathcal{X} = \\mathbb{R}^n\\) and \\(\\mathcal{Y} = \\mathbb{R}^m\\). Furthermore, we assume \\(\\mathcal{G}\\) is a parametric set that consists of mappings \\(\\mathcal{G} = \\{G(\\cdot ,\\xi):\\xi \\in \\mathbb{R}^p\\}\\), where \\(G:\\mathbb{R}^n\\times \\mathbb{R}^p\\to \\mathbb{R}^m\\) is a continuously differentiable map. This includes, for example, ordinary neural networks, where the parameters are the weights. With these assumptions in place, the minimization problem in (1) is equivalent to minimizing over the parameters \\(\\xi\\) instead. In practice, we have access to only a finite set of observations, which are assumed to be i.i.d., and the objective in (1) is replaced by an arithmetic average." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.611, + 0.698, + 0.627 + ], + "angle": 0, + "content": "2.2. Imposing constraints" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.634, + 0.895, + 0.848 + ], + "angle": 0, + "content": "In this section, we explain how to incorporate constraints on a subset of the parameters \\(\\xi\\) directly into the ERM framework. We consider constraints given in the form of a system of equations. To be very precise, let \\(F:\\mathbb{R}^{\\tilde{p}}\\to \\mathbb{R}^{q}\\) be a twice continuously differentiable map, where \\(\\tilde{p}\\) denotes the number of constrained parameters, and \\(q\\) is the number of equations. We assume that \\(q < \\tilde{p}\\leq p\\). For notational convenience, we decompose \\(\\mathbb{R}^p = \\mathbb{R}^{p - \\tilde{p}}\\oplus \\mathbb{R}^{\\tilde{p}}\\), where the first and second subspaces correspond to the unconstrained and constrained parameters, respectively. We take \\(\\pi_{p - \\tilde{p}}:\\mathbb{R}^p\\to \\mathbb{R}^{p - \\tilde{p}}\\) and \\(\\pi_{\\tilde{p}}:\\mathbb{R}^p\\to \\mathbb{R}^{\\tilde{p}}\\) to be the projections onto the unconstrained and constrained parameter subspace, respectively. We will denote the unconstrained and constrained parameters by \\(\\alpha\\) and \\(\\theta\\), respectively, i.e., \\(\\alpha = \\pi_{p - \\tilde{p}}(\\xi)\\) and \\(\\theta = \\pi_{\\tilde{p}}(\\xi)\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Definition 2.1 (CERM). Let \\( G: \\mathbb{R}^n \\times \\mathbb{R}^p \\to \\mathbb{R}^m \\) be a continuously differentiable parameterization of mappings and \\( F: \\mathbb{R}^{\\tilde{p}} \\to \\mathbb{R}^q \\) a twice continuously differentiable constraint," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24100" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.138 + ], + "angle": 0, + "content": "where \\(q < \\tilde{p} \\leq p\\). Suppose \\(L: \\mathbb{R}^m \\times \\mathbb{R}^m \\to [0, \\infty)\\) is a continuously differentiable loss function. The constrained ERM problem for \\((X, Y)\\) with respect to \\((G, F, L)\\) is defined by" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.159, + 0.47, + 0.207 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\min _ {\\xi \\in \\mathbb {R} ^ {p}} \\mathbb {E} (L (G (X, \\xi), Y)), \\\\ \\text {s . t .} F (\\pi_ {\\tilde {p}} (\\xi)) = 0. \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.473, + 0.382 + ], + "angle": 0, + "content": "Note the generality of the admissible mappings \\( G \\) in Definition 2.1. Although we will focus on neural networks from now on, the proposed framework applies to any parametric model, e.g., logistic or polynomial regression models. Next, we show that the CERM problem in (2) can be reformulated as an ordinary ERM problem on a Riemannian manifold \\( (\\mathcal{N}, g_{\\mathcal{N}}) \\), provided that the system of equations satisfies a mild non-degeneracy condition. This result allows us to consider the admissible parameters as a geometric object in its own right, whose intrinsic geometry we use to solve (2). The proof of this theorem can be found in Sec. 6.1 [16]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.473, + 0.482 + ], + "angle": 0, + "content": "Theorem 2.2. If zero is a regular (i.e., non-singular) value of \\(F\\), then the CERM problem in (2) is equivalent to solving an ordinary ERM problem on a Riemannian manifold \\(\\mathcal{N}\\) with metric \\(g_{\\mathcal{N}}\\) of dimension \\(p - q\\). Here \\(\\mathcal{N} = \\mathbb{R}^{p - \\tilde{p}} \\times \\mathcal{M}\\) is an embedded \\(C^2\\)-submanifold of \\(\\mathbb{R}^p\\) and \\(\\mathcal{M} := F^{-1}(0)\\). The equivalent minimization problem is given by" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.491, + 0.47, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\min _ {(\\alpha , \\theta) \\in \\mathcal {N}} \\mathbb {E} (L (G (X, \\alpha \\oplus \\iota (\\theta)), Y)), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.47, + 0.557 + ], + "angle": 0, + "content": "where \\(\\iota : \\mathcal{M} \\to \\mathbb{R}^{\\tilde{P}}\\) is the inclusion map which embeds \\(\\mathcal{M}\\) in the ambient space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.47, + 0.595 + ], + "angle": 0, + "content": "From now on, we refer to the objective in (3) as simply the loss \\(\\mathcal{L}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.841 + ], + "angle": 0, + "content": "Graph coordinates on \\(\\mathcal{N}\\) The main idea is to directly minimize the loss on \\(\\mathcal{N}\\) using SGD, see Figure 1 for an overview. To perform SGD on this space, however, we need to choose a coordinate system in which we can evaluate the gradient of the loss and subsequently follow descent trajectories on \\(\\mathcal{N}\\). Now, the proof of Theorem 2.2 involves showing that \\(\\mathcal{M}\\) is an embedded submanifold of the ambient space \\(\\mathbb{R}^{\\tilde{p}}\\). This is accomplished by constructing a special coordinate system, a so-called graph chart or graph coordinate system. This chart is uniquely determined by the constraint \\(F\\) and a prescribed zero \\(\\theta \\in \\mathcal{M}\\). It can be numerically evaluated at the point \\(\\theta\\) by solving a linear system of equations. In turn, this enables us to construct a chart on the manifold in which we can evaluate the Riemannian metric and the gradient \\(\\nabla_{g_{\\mathcal{N}}} \\mathcal{L}(\\alpha, \\theta)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.473, + 0.902 + ], + "angle": 0, + "content": "The crucial observation is that we can numerically construct coordinate systems in which we can follow paths on the manifold \\(\\mathcal{N}\\) in the direction of \\(-\\nabla_{g_{\\mathcal{N}}}\\mathcal{L}(\\alpha, \\theta)\\). In particular, the system of equations determining these" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.894, + 0.136 + ], + "angle": 0, + "content": "Algorithm 1 Sketch of the computation of \\(\\nabla_{g_N}\\mathcal{L}(\\alpha ,\\theta)\\) given \\((\\alpha ,\\theta)\\in \\mathcal{N}\\). See Algorithm 2 in Sec. 6.5 [16] the detailed formal version." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.14, + 0.867, + 0.155 + ], + "angle": 0, + "content": "1: Compute the derivative \\(DF(\\theta)\\) of the constraint \\(F\\)." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.156, + 0.758, + 0.169 + ], + "angle": 0, + "content": "2: Compute chart-related derivatives." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.17, + 0.894, + 0.198 + ], + "angle": 0, + "content": "3: Compute the metric \\(g_{\\mathcal{N}}\\) using the chart-related derivatives." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.2, + 0.892, + 0.23 + ], + "angle": 0, + "content": "4: Compute the partial derivatives of the loss \\(\\mathcal{L}\\) with respect to the constrained components \\(\\theta\\)." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.231, + 0.892, + 0.26 + ], + "angle": 0, + "content": "5: Compute the components of \\(\\nabla_{g_{\\mathcal{M}}}\\mathcal{L}(\\alpha ,\\theta)\\) using the above ingredients." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.261, + 0.892, + 0.292 + ], + "angle": 0, + "content": "6: Compute the ordinary (unconstrained) components \\(\\nabla_{g_{\\mathrm{flat}}}\\mathcal{L}(\\alpha ,\\theta)\\) by evaluating \\(D_{\\alpha}\\mathcal{L}(\\alpha ,\\theta)\\)" + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.14, + 0.894, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.895, + 0.505 + ], + "angle": 0, + "content": "so-called graph coordinate systems is fully determined by the derivative \\(DF\\) of the constraint \\(F\\) and the constraint itself. This observation enables a general implementation of graph coordinate systems independent of the explicit form of \\(F\\). We only require an implementation of \\(F\\) and \\(DF\\), where the latter may be obtained using automatic differentiation. We provide a sketch of the algorithm for computing the gradient in Algorithm 1. For the interested reader, we mention that a complete and detailed explanation of all steps can be found in Sec. 6 [16]. Before moving on, we mention the limitations of the CERM framework for other generic constraints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.895, + 0.754 + ], + "angle": 0, + "content": "Limitations As with most applications of deep neural networks, the available amount of computational power imposes limitations. Specifically, if the number of constrained parameters is large relative to the number of constraints, the computational demands will rise. Furthermore, setting up well-posed constraints can be a non-trivial task, e.g., symmetries in the constraints may result in redundancies obstructing the solvability of the associated systems of equations. Finally, the topology of the constrained submanifold may pose limitations as well. For instance, while the constraints will generically define a Riemannian manifold, it need not be connected. In such cases, the initialization of the network parameters will play an important role, as the gradient flow will stay on the connected component it was initialized on." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.769, + 0.833, + 0.787 + ], + "angle": 0, + "content": "3. Multiresolution Analysis and CERM" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.902 + ], + "angle": 0, + "content": "This section presents the application of the aforementioned CERM framework to a non-trivial application to learn optimal wavelet bases for a given task. Many different types of wavelets, such as Haar, Daubechies, Gabor, etc., are used for a wide array of tasks. The specifics of the task demand different characteristics of the wavelets, such as the degree of smoothness, or the support of the wavelet. For example," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24101" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "the CDF wavelets [10] are well-known due to their usage for both lossless and lossy compression in JPEG2000. In this section, we consider networks that will find task-driven wavelets, optimizing the wavelet characteristics to the task at hand." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.304 + ], + "angle": 0, + "content": "Here, we will explain how to set up a system of equations (constraints) whose solution set corresponds to wavelets. We will apply these constraints in the next section to learn task-driven wavelets for predicting multi-resolution decompositions of contours in the medical domain. While we only consider one-dimensional wavelets in this work, we stress that the framework is easily adapted to higher-dimensional domains, to decompose images, for instance, by using tensor products of one-dimensional bases." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.314, + 0.303, + 0.33 + ], + "angle": 0, + "content": "3.1. Multiresolution Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.473, + 0.642 + ], + "angle": 0, + "content": "In this section, we briefly state what a Multiresolution Analysis (MRA) is. We closely follow the standard exposition, found in, e.g., [29, 31], and refer the reader to Sec. 7 [16] for a more in-depth and formal introduction. The MRA framework allows one to study a signal \\(\\gamma \\in L^2 (\\mathbb{R})\\) at various resolution levels. The origin of this framework lies in the uncertainty principle of Fourier analysis, which states that a signal cannot be simultaneously localized in the time and frequency domain. MRA addresses this shortcoming by decomposing a signal on different discrete resolution levels. The idea is to construct a nested (exhausting) sequence of closed subspaces \\(V_{j}\\subset L^{2}(\\mathbb{R})\\), each associated with a specific resolution level \\(j\\). The subspaces \\(V_{j}\\) are spanned by integer shifts of an appropriate dilation of a so-called scaling map (or father wavelet) \\(\\varphi\\). The level of dilation determines the resolution level. Formally, we require that \\((\\varphi_{jk})_{k\\in \\mathbb{Z}}\\) is an orthonormal basis for \\(V_{j}\\), where \\(\\varphi_{jk}(t)\\coloneqq 2^{\\frac{j}{2}}\\varphi (2^{j}t - k)\\). In Fig. 2a and Fig. 2c, we show an example of that Haar-MRA; in particular, we show the father wavelet and an example of a subspace at a certain resolution level." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.884 + ], + "angle": 0, + "content": "Decomposing a signal To analyze a signal \\(\\gamma\\) at different resolution levels, we project it onto the subspaces \\(V_{j}\\). The orthogonal projection of \\(\\gamma\\) at level \\(j\\) is denoted by \\(\\gamma_{j}\\). The coefficients of \\(\\gamma_{j}\\) with respect to the basis for \\(V_{j}\\), denoted by \\(a_{jk}(\\gamma)\\), are called the approximation coefficients of \\(\\gamma\\). When a signal in \\(V_{j+1}\\) is projected onto \\(V_{j}\\), the lost information is stored in the orthogonal complement \\(W_{j}\\) of \\(V_{j}\\) in \\(V_{j+1}\\). This subspace is referred to as the detail subspace at level \\(j\\). The decomposition of \\(V_{j+1}\\) into \\(V_{j}\\) and \\(W_{j}\\) allows the reconstruction of a signal at level \\(j+1\\) from any lower level \\(j_{0}\\) provided all the details in between are known. The corresponding decomposition and reconstruction algorithms give rise to the Discrete Wavelet Transform (DWT) and Mallat's Pyramid Algorithm. In Fig. 2b and Fig. 2d, we show examples of the mother wavelet and the detail subspace for the Haar wavelet." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.9 + ], + "angle": 0, + "content": "A fundamental result, known as Mallat's Theorem, states" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.089, + 0.689, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.199, + 0.646, + 0.211 + ], + "angle": 0, + "content": "(a) Father wavelet" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.089, + 0.883, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.199, + 0.843, + 0.211 + ], + "angle": 0, + "content": "(b) Mother wavelet" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.212, + 0.688, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.321, + 0.614, + 0.335 + ], + "angle": 0, + "content": "(c) \\(V_{j}\\)" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.212, + 0.88, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.321, + 0.81, + 0.335 + ], + "angle": 0, + "content": "(d) \\(W_{j}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.345, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Figure 2. Example of the Haar MRA: (a) Dilated translation of the Haar scaling map (or father wavelet) \\(\\varphi = \\mathbf{1}_{[0,1)}\\). The approximation subspace \\(V_{j}\\) at level \\(j\\) consists of all step-functions with step-size \\(2^{-j}\\). (b) Dilated translation of the mother wavelet \\(\\psi = \\mathbf{1}_{[0,\\frac{1}{2})} - \\mathbf{1}_{[\\frac{1}{2},1)}\\). The detail subspace \\(W_{j}\\) is spanned by integer translations of the dilated mother wavelet \\(\\psi_{j}\\). (c) The approximation subspace at level \\(j\\) consists of all step-functions with step-size \\(2^{-j}\\). (d) Example of a function in the detail subspace at level \\(j\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.892, + 0.62 + ], + "angle": 0, + "content": "that the subspaces \\( W_{j} \\) can also be spanned by dilating and shifting a single map. More precisely, there exists a map \\( \\psi \\in W_0 \\), the so-called mother wavelet, such that the \\( \\psi_{jk} \\) form an orthonormal basis for \\( W_{j} \\), see [31]. The coefficients of the projection of \\( (\\gamma) \\) onto \\( W_{j} \\), are denoted by \\( d_{j}(\\gamma) \\coloneqq (d_{jk}(\\gamma))_{k\\in \\mathbb{Z}} \\), and referred to as the detail coefficients of \\( \\gamma \\) at resolution level \\( j \\). We can summarize the MRA procedure roughly as" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.629, + 0.892, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {j} = \\sum_ {k} a _ {j _ {0} k} \\varphi_ {j _ {0} k} + \\sum_ {j _ {0} \\leq l \\leq j - 1} \\sum_ {k} d _ {l k} \\psi_ {l k}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.894, + 0.716 + ], + "angle": 0, + "content": "where \\(\\varphi\\) denotes the father wavelet and \\(\\psi\\) the mother wavelet. The MRA boils down to the specification of the first approximation coefficient and a number of detail coefficients." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Low and high pass filters As we mentioned before, an MRA is completely determined by a scaling map \\(\\varphi\\), which in turn is completely characterized by a so-called low-pass filter. We will use this observation to set up a system of constraints characterizing a finite-dimensional family of wavelets, such that we can use it in the CERM framework. To explain how this works, we first make a key observation. Since \\(V_{0} \\subset V_{1}\\), there exists a unique sequence \\(h\\) characterizing \\(\\varphi\\) in terms of the basis \\((\\varphi_{1k})_{k \\in \\mathbb{Z}}\\) for \\(V_{1}\\). The sequence \\(h\\) completely characterizes the scaling function and is called the low-pass filter or wavelet filter of the MRA. Similarly," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24102" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "since \\(\\psi \\in W_0\\subset V_1\\), there exists a unique sequence \\(g\\), the so-called high-pass filter associated to \\(h\\), characterizing the mother wavelet. An example of an MRA is shown in Figure 2. In practice, to define an MRA, one only needs to specify an appropriate low-pass filter \\(h\\), as for Mallat's mother wavelet, we have \\(g_{k} = (-1)^{k - 1}h_{1 - k}\\). We refer the interested reader to the supplementary information for a detailed discussion of all these points [16]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.224, + 0.248, + 0.239 + ], + "angle": 0, + "content": "3.2. MRA constraints" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.249, + 0.474, + 0.354 + ], + "angle": 0, + "content": "We are now ready to formulate and impose constraints on a sequence \\( h \\) to ensure that it is the low-pass filter of a scaling map \\( \\varphi \\). In practice, we only consider finite sequences. For notational convenience, we introduce the space \\( \\mathcal{A}_M(\\mathbb{R}) \\), which consists of one-dimensional real-valued two-sided sequences \\( (h_k)_{k=1-M}^{M-1} \\) of order \\( M \\); so \\( \\mathcal{A}_M(\\mathbb{R}) \\) is \\( 2M - 1 \\) dimensional." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.473, + 0.399 + ], + "angle": 0, + "content": "Theorem 3.1. Let \\(M \\in \\mathbb{N}_{\\geq 3}\\) be a prescribed order and define \\(F_M: \\mathcal{A}_M(\\mathbb{R}) \\to \\mathbb{R}^{M+1}\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.411, + 0.487, + 0.516 + ], + "angle": 0, + "content": "\\[\n(F_{M}(h))_{k}:= \\left\\{ \\begin{array}{ll}\\sum_{|l|\\leq M - 1}|h_{l}|^{2} - 1, & k = 0,\\\\ \\sum_{1 - M + 2k\\leq l\\leq M - 1}h_{l - 2k}h_{l}, & 1\\leq k\\leq M - 1,\\\\ -\\sqrt{2} +\\sum_{|l|\\leq M - 1}h_{l}, & k = M. \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.471, + 0.575 + ], + "angle": 0, + "content": "If \\( F_{M}(h) = 0 \\), and \\( h \\) satisfies a certain non-degeneracy condition, then \\( h \\) is the low-pass filter of a scaling map \\( \\varphi \\) generating an MRA of \\( L^2 (\\mathbb{R}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.472, + 0.768 + ], + "angle": 0, + "content": "A full proof and detailed statement of the non-degeneracy condition can be found in Sec. 7.4 in the supplementary material. Heuristically, the first \\(M\\) equations encode the orthonormality of the scaling map. The last equation encodes the observation that the Fourier transform of the scaling map must be nonzero. The set of regular points in \\(F_{M}^{-1}(0)\\) is a real-analytic \\((M - 2)\\)-dimensional submanifold of \\(\\mathbb{R}^{2M - 1}\\). Hence, we can get as many degrees of freedom as desired, by choosing a sufficiently large order \\(M\\). We now have an end-to-end way of learning wavelets for specific tasks, a wavelet network, by using the constraint \\(F_{M}\\) along with the CERM framework." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.784, + 0.21, + 0.801 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this section, we show some explicit examples of the wavelet networks with respect to contour prediction. In particular, we will utilize the derived wavelet filters of Sec. 3.2 to create a network for autocontouring. This task fits well with the setup of wavelets, as they naturally lend themselves to representing continuous objects such as curves." + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.091, + 0.872, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.299, + 0.895, + 0.522 + ], + "angle": 0, + "content": "Figure 3. A schematic picture of our network. The encoder consists of residual convolutional blocks depicted in blue. The first residual convolutional block uses \\( n_f \\) filters and is doubled after every other residual convolutional block. Attached to the encoder are fully connected layers to predict approximation and detail coefficients at the lowest resolution level \\( j_0 \\). The approximation and detail coefficients are supplied as input to the Pyramid Algorithm (the decoder) to predict a contour on high-resolution level. Each green block corresponds to an inverse discrete wavelet transform. Detail coefficients at higher levels are computed using skip-connections (arrows in red). We only predict detail coefficients up to level \\( j_1 \\). In this example, we have set \\( j_1 = j_0 + 2 \\) and \\( j_2 = j_0 + 3 \\). In reality, the decoder consists of two upsampling paths, one for each spatial component of the curve. We have only drawn one for notational convenience. During training, only the approximation coefficients at the highest resolution level are supervised." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.547, + 0.779, + 0.563 + ], + "angle": 0, + "content": "4.1. Contour Prediction using MRA" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.895, + 0.768 + ], + "angle": 0, + "content": "In the context of contour prediction, wavelets can be used to represent the boundary of a region in an image using a simple closed curve. We consider two-dimensional gray-valued images \\( x \\in \\mathcal{X} \\coloneqq [0,1]^{n \\times n} \\). For each of our applications, we assume that these images contain an organ with a boundary. Formally speaking, this means that each image \\( x \\) contains a (uniquely identifiable) simply connected region \\( R(x) \\), the organ, with boundary \\( \\partial R(x) \\). This boundary of the region, \\( \\partial R(x) \\), can be parameterized by a simple closed, piecewise \\( C^2 \\), curve \\( \\gamma(x) \\). We will develop a deep learning framework for computing such parameterizations by learning a multiresolution decomposition of \\( \\gamma(x) \\) using the methods developed in the previous sections." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.589, + 0.8 + ], + "angle": 0, + "content": "4.1.1 Data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.897, + 0.901 + ], + "angle": 0, + "content": "We have used public datasets from the Medical Decathlon Challenge [3]. The selected data consists of CT scans of the spleen of size \\(512 \\times 512\\) and T2-weighted MRI images of the prostate central gland, henceforth abbreviated as just the prostate, of size \\(320 \\times 320\\). The scans were cropped to size \\(224 \\times 224\\) and \\(192 \\times 192\\), respectively. Further-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24103" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.224, + 0.089, + 0.749, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.266, + 0.895, + 0.308 + ], + "angle": 0, + "content": "Figure 4. Examples of predictions on the test set for the spleen and prostate, depicted in the first and second row, respectively, for the best-performing wavelet models. The green curve corresponds to the ground truth, while the red curve is a prediction made by the wavelet network." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.332, + 0.451, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.453, + 0.368, + 0.464 + ], + "angle": 0, + "content": "(a) Initial father and mother wavelet" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.466, + 0.45, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.585, + 0.393, + 0.597 + ], + "angle": 0, + "content": "(b) Task-optimized father and mother wavelet" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.695 + ], + "angle": 0, + "content": "Figure 5. Example of wavelets of order 5 learned during training of the spleen model: (a) The initial father (left) and mother (right) wavelets for the second spatial component. (b) The task-optimized father and mother wavelets after training for the second spatial component. We observe that the task-optimized wavelets are more simple and have \"cleaned up\" in a sense." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.473, + 0.903 + ], + "angle": 0, + "content": "more, the images were resampled to the median sample spacing, which resulted in \\((5.00\\mathrm{mm}, 0.793\\mathrm{mm}, 0.793\\mathrm{mm})\\) and \\((3.6\\mathrm{mm}, 0.625\\mathrm{mm}, 0.625\\mathrm{mm})\\) spacings for the spleen and prostate, respectively. We refer the reader to Sec. 8.2 [16] for a detailed description of all the preprocessing steps and construction of ground truth curves. Examples of predictions on the test set, for both the spleen and prostate, are shown in Figure Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.335, + 0.695, + 0.35 + ], + "angle": 0, + "content": "4.1.2 Model and Training" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.358, + 0.895, + 0.464 + ], + "angle": 0, + "content": "Model objective We parameterize the boundary \\(\\partial R(x)\\) by arc length, resulting in a curve \\(\\gamma (x)\\). The precise details of this parametrization are outlined in the supplementary information. The objective is to compute the relevant approximation coefficients of \\(\\gamma (x)\\). To this end, let \\(j_0,j_1,j_2\\in \\mathbb{N}\\) be resolution levels, where \\(j_0\\leq j_1\\leq j_2\\). We will construct a convolutional neural network" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.472, + 0.866, + 0.516 + ], + "angle": 0, + "content": "\\[\nG: \\mathcal {X} \\times \\mathbb {R} ^ {p} \\to \\prod_ {j = j _ {0}} ^ {j _ {2}} \\mathbb {R} ^ {2 ^ {j}} \\times \\mathbb {R} ^ {2 ^ {j}} \\times \\prod_ {j = j _ {0}} ^ {j _ {1} - 1} \\mathbb {R} ^ {2 ^ {j}} \\times \\mathbb {R} ^ {2 ^ {j}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.524, + 0.895, + 0.633 + ], + "angle": 0, + "content": "which predicts the wavelet decomposition of \\(\\gamma (x)\\). Here the subspaces \\(\\mathbb{R}^{2^j}\\) correspond to approximation and detail coefficients at level \\(j\\), one for each spatial component. Furthermore, a subset of the parameters \\(\\xi \\in \\mathbb{R}^p\\) are constrained to be wavelet filters, one wavelet filter per spatial component, using Theorem 3.1. The map \\(G(\\cdot ,\\xi)\\) applied to an image \\(x\\) has output" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.642, + 0.892, + 0.666 + ], + "angle": 0, + "content": "\\[\nG (x, \\xi) = \\left(a _ {j _ {0}} (x, \\xi), \\dots , a _ {j _ {2}} (x, \\xi), \\right. \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.661, + 0.825, + 0.677 + ], + "angle": 0, + "content": "\\[\nd _ {j _ {0}} (x, \\xi), \\dots , d _ {j _ {1} - 1} (x, \\xi)).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.687, + 0.895, + 0.734 + ], + "angle": 0, + "content": "Here \\(a_{j}(x,\\xi)\\) and \\(d_{j}(x,\\xi)\\) are predictions for the approximation and detail coefficients of \\(\\gamma (x)\\) at level \\(j\\), respectively. We only predict detail coefficients up to level \\(j_{1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Architecture The network is a hybrid analog of the U-Net. It consists of a two-dimensional convolutional encoder, a bottleneck of fully connected layers, and a one-dimensional decoder. The encoder and decoder are connected through skip-connections. The approximation and detail coefficients at the lowest resolution level \\( j_0 \\) are predicted in the bottleneck. Afterward, the Pyramid Algorithm takes over to compute approximation coefficients at higher resolution levels (the decoder) using learnable wavelet filters. In Fig. 3, we provide a schematic overview of the network architecture." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24104" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "Loss To measure the discrepancy between the ground truth and the predicted curve, we define" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.138, + 0.47, + 0.174 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r l} L (G (x, \\xi), \\tilde {a} (\\gamma (x))) := & \\| [ a _ {j _ {2}} ] _ {1} - \\tilde {a} _ {j _ {2}} ([ \\gamma (x) ] _ {1}) \\| _ {2} + \\\\ & \\| [ a _ {j _ {2}} ] _ {2} - \\tilde {a} _ {j _ {2}} ([ \\gamma (x) ] _ {2}) \\| _ {2}. \\end{array} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.189, + 0.471, + 0.267 + ], + "angle": 0, + "content": "The \\(\\tilde{a}\\) thus correspond to the ground-truth approximation coefficients, and the subscripts 1, 2 denote the spatial dimension. In other words, the loss corresponds to the componentwise \\(L^2\\)-error between the curves on resolution level \\(j_2\\) with approximation coefficients \\(a_{j_2}(x,\\xi)\\) and \\(\\tilde{a}_{j_2}(\\gamma (x))\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.473, + 0.706 + ], + "angle": 0, + "content": "Optimization We use two different optimizers during training: one for the encoder and MLP network (unconstrained), and one for the decoder (constrained wavelet network). The reason for this is that the needed step sizes on the non-trivial submanifold may significantly differ from those on the unconstrained (flat) parameter space. For the training of the model, we use plain SGD for the first eight epochs for both the constrained and unconstrained parameters. During this period, the learning rate for the unconstrained parameters is linearly increased from \\(10^{-5}\\) to \\(2 \\cdot 10^{-4}\\). The learning rate for the constrained parameters (wavelet filters) is linearly increased from \\(10^{-4}\\) to \\(10^{-2}\\). After the initial warmup stage, we switch to the Adam optimizer for the unconstrained parameters. For both the constrained and unconstrained parameters, we use learning rate schedulers and decrease the learning rate by a factor of 0.85 if no significant improvements in the validation loss are observed during the last ten epochs. We train all models for 250 epochs using a batch size of 32 and use the last epoch for inference. The computations were performed in PYTORCH on a Geforce RTX 2080 Ti. An average training session takes about two and a half hours for the wavelet network and a day for the nnUNet. The inference times for the wavelet network were 12.5ms and 14ms for the prostate and spleen, respectively, using a batch size of 32. The inference times for the nnUNet were 17.8ms and 23.4ms for the prostate and spleen, respectively, using a batch size of 32." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.737, + 0.183, + 0.751 + ], + "angle": 0, + "content": "4.1.3 Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.903 + ], + "angle": 0, + "content": "We have evaluated the performance of both our wavelet networks for different orders and a state-of-the-art baseline 2d-nnUNet [13] on the unseen test data, see Table 1. For both the spleen and prostate, we observe that the best wavelet networks outperform the baseline. We note that the task-optimized wavelets differ significantly from the wavelets randomly initialized at the start of training. A comparison of an initial and task-optimized wavelet is depicted in Figure 5, see Sec. 8.4 for more examples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.089, + 0.892, + 0.16 + ], + "angle": 0, + "content": "Table 1. Mean and standard deviation (in parentheses) of the dice score on the unseen test sets. The number of parameters in our wavelet-based networks ranged from 5M to 10M parameters, while the nnUNet used approximately 40M and 18M parameters for the spleen and prostate, respectively." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.184, + 0.868, + 0.307 + ], + "angle": 0, + "content": "
MODELDICE SPLEENDICE PROSTATE
NNUNET0.914 (1.74 · 10-1)0.896 (1.27 · 10-1)
ORDER 30.911 (7.38 · 10-2)0.929 (4.66 · 10-2)
ORDER 40.911 (7.85 · 10-2)0.935 (3.48 · 10-2)
ORDER 50.916 (6.94 · 10-2)0.935 (4.11 · 10-2)
ORDER 60.917 (7.17 · 10-2)0.934 (4.14 · 10-2)
ORDER 70.921 (6.91 · 10-2)0.928 (4.03 · 10-2)
ORDER 80.919 (6.64 · 10-2)0.934 (3.62 · 10-2)
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.336, + 0.627, + 0.351 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.895, + 0.573 + ], + "angle": 0, + "content": "In this paper, we have introduced the CERM framework for creating task-driven wavelets. Although our aim is to find such wavelets, in principle the framework lends itself to imposing generic constraints on parametric models. The wavelet constraints can be formulated as a finite system of equations. Under mild smoothness and non-degeneracy conditions, the networks can be made to obey the constraints exactly throughout the entire training procedure by performing SGD on a curved space. We can then utilize these wavelet constraints in a convolutional network, where the filters of the network now parameterize wavelets themselves. We have applied these wavelet networks to the prediction of boundaries of simply connected regions in medical images, where they outperform strong baselines." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.589, + 0.597, + 0.604 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.614, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[1] P. Abry and P. Flandrin. On the initialization of the discrete wavelet transform algorithm. IEEE Signal Processing Letters, 1(2):32-34, 1994. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.657, + 0.894, + 0.713 + ], + "angle": 0, + "content": "[2] Brandon Amos, Lei Xu, and J. Zico Kolter. Input convex neural networks. In Proceedings of the 34th International Conference on Machine Learning, pages 146-155. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.714, + 0.895, + 0.784 + ], + "angle": 0, + "content": "[3] Michela Antonelli, Annika Reinke, Spyridon Bakas, Keyvan Farahani, Annette Kopp-Schneider, Bennett A Landman, Geert Litjens, Bjoern Menze, Olaf Ronneberger, Ronald M Summers, et al. The medical segmentation decathlon. Nature communications, 13(1):1-13, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.785, + 0.895, + 0.826 + ], + "angle": 0, + "content": "[4] Randall Balestriero and Yann LeCun. Police: Provably optimal linear constraint enforcement for deep neural networks, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.829, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[5] D.P. Bertsekas and W. Rheinboldt. Constrained Optimization and Lagrange Multiplier Methods. Elsevier Science, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.859, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[6] Silvere Bonnabel. Stochastic gradient descent on riemannian manifolds. IEEE Transactions on Automatic Control, 58(9): 2217-2229, 2013. 2, 1, 5" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.614, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24105" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "[7] N. Boumal. An Introduction to Optimization on Smooth Manifolds. Cambridge University Press, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.121, + 0.472, + 0.176 + ], + "angle": 0, + "content": "[8] Michael M Bronstein, Joan Bruna, Yann LeCun, Arthur Szlam, and Pierre Vandergheynst. Geometric deep learning: going beyond euclidean data. IEEE Signal Processing Magazine, 34(4):18-42, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.177, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[9] Michael M Bronstein, Joan Bruna, Taco Cohen, and Petar Velicković. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges. arXiv preprint arXiv:2104.13478, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.287 + ], + "angle": 0, + "content": "[10] Albert Cohen, Ingrid Daubechies, and J-C Feauveau. Biorthogonal bases of compactly supported wavelets. Communications on pure and applied mathematics, 45(5):485-560, 1992. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[11] Priya Donti, David Rolnick, and J Zico Kolter. Dc3: A learning method for optimization with hard constraints. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.471, + 0.372 + ], + "angle": 0, + "content": "[12] M.W. Frazier. An Introduction to Wavelets Through Linear Algebra. Springer New York, 2001. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.471, + 0.426 + ], + "angle": 0, + "content": "[13] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[14] Hiroyuki Kasai, Pratik Jawanpuria, and Bamdev Mishra. Riemannian adaptive stochastic gradient algorithms on matrix manifolds. In International Conference on Machine Learning, pages 3262-3271. PMLR, 2019. 2, 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.471, + 0.552 + ], + "angle": 0, + "content": "[15] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings. International Conference on Learning Representations, ICLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.471, + 0.594 + ], + "angle": 0, + "content": "[16]FirstName LastName. Task-driven wavelets using constrained empirical risk minimization, 2023. Supplied as supplemental material supplementary-notes.pdf. 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.471, + 0.636 + ], + "angle": 0, + "content": "[17] Yann LeCun, Yoshua Bengio, et al. Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks, 3361(10):1995, 1995. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.471, + 0.664 + ], + "angle": 0, + "content": "[18] John M Lee. Riemannian manifolds: an introduction to curvature. Springer Science & Business Media, 2006. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.471, + 0.692 + ], + "angle": 0, + "content": "[19] John M Lee. Smooth manifolds. In Introduction to smooth manifolds. Springer, 2013. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.693, + 0.471, + 0.734 + ], + "angle": 0, + "content": "[20] Benedict Leimkuhler, Timothee Pouchon, Tiffany Vlaar, and Amos Storkey. Constraint-based regularization of neural networks. arXiv preprint arXiv:2006.10114, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[21] Benedict Leimkuhler, Tiffany Vlaar, Timothee Pouchon, and Amos Storkey. Better training using weight-constrained stochastic dynamics. arXiv preprint arXiv:2106.10704, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[22] Qiufu Li, Linlin Shen, Sheng Guo, and Zhihui Lai. Wavelet integrated cnns for noise-robust image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7245–7254, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[23] Cho-Chun Liu, Yaohui Liu, Zhengding Qiu, and Xiyu Du. On the initialization of the discrete wavelet transform algorithm. 1997 IEEE International Conference on Intelligent Processing Systems (Cat. No.97TH8335), 2:1220-1222 vol.2, 1997. 14" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.894, + 0.149 + ], + "angle": 0, + "content": "[25] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.15, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[26] Stephane Mallat. A Wavelet Tour of Signal Processing, Third Edition: The Sparse Way. Academic Press, Inc., USA, 3rd edition, 2008. 2, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.894, + 0.235 + ], + "angle": 0, + "content": "[27] Pablo Márquez-Neila, Mathieu Salzmann, and Pascal Fua. Imposing hard constraints on deep networks: Promises and limitations. arXiv preprint arXiv:1706.02025, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.894, + 0.291 + ], + "angle": 0, + "content": "[28] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.894, + 0.32 + ], + "angle": 0, + "content": "[29] L. Montefusco and L. Puccio. Wavelets: Theory, Algorithms, and Applications. Elsevier Science, 2014. 5, 11, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.894, + 0.377 + ], + "angle": 0, + "content": "[30] Deepak Pathak, Philipp Krahenbuhl, and Trevor Darrell. Constrained convolutional neural networks for weakly supervised segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1796-1804, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.894, + 0.418 + ], + "angle": 0, + "content": "[31] M.C. Pereyra and L.A. Ward. Harmonic Analysis: From Fourier to Wavelets. American Mathematical Society, 2012. 5, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.894, + 0.461 + ], + "angle": 0, + "content": "[32] Jack Richter-Powell, Yaron Lipman, and Ricky T. Q. Chen. Neural conservation laws: A divergence-free perspective, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.894, + 0.517 + ], + "angle": 0, + "content": "[33] David W Romero, Erik J Bekkers, Jakub M Tomczak, and Mark Hoogendoorn. Wavelet networks: Scale equivariant learning from raw waveforms. arXiv preprint arXiv:2006.05259, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.894, + 0.587 + ], + "angle": 0, + "content": "[34] Soumava Kumar Roy, Zakaria Mhammedi, and Mehrtash Harandi. Geometry aware constrained optimization techniques for deep learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4460-4469, 2018. 2, 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.894, + 0.645 + ], + "angle": 0, + "content": "[35] Hiroyuki Sato, Hiroyuki Kasai, and Bamdev Mishra. Riemannian stochastic variance reduced gradient algorithm with retraction and vector transport. SIAM Journal on Optimization, 29(2):1444-1472, 2019. 2, 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.894, + 0.687 + ], + "angle": 0, + "content": "[36] Vladimir Vapnik. Principles of risk minimization for learning theory. Advances in neural information processing systems, 4, 1991. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.894, + 0.716 + ], + "angle": 0, + "content": "[37] Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.894, + 0.745 + ], + "angle": 0, + "content": "[38] D.F. Walnut. An Introduction to Wavelet Analysis. Birkhäuser Boston, 2002. 12, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.894, + 0.802 + ], + "angle": 0, + "content": "[39] Maurice Weiler, Patrick Forre, Erik Verlinde, and Max Welling. Coordinate independent convolutional networks-isometry and gauge equivariant convolutions on riemannian manifolds. arXiv preprint arXiv:2106.06020, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.894, + 0.872 + ], + "angle": 0, + "content": "[40] Yang You, Jing Li, Sashank Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes. In International Conference on Learning Representations, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[41] Jason J Yu, Konstantinos G Derpanis, and Marcus A Brubaker. Wavelet flow: Fast training of high resolution normalizing" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24106" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.119 + ], + "angle": 0, + "content": "flows. Advances in Neural Information Processing Systems, 33:6184-6196, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.177 + ], + "angle": 0, + "content": "[42] Hongyi Zhang, Sashank J. Reddi, and Suvrit Sra. Riemannian svrg: Fast stochastic optimization on riemannian manifolds. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 2, 1, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24107" + } + ] +] \ No newline at end of file diff --git a/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/d6dc5c43-4c52-4779-8c6a-d513a47a71be_origin.pdf b/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/d6dc5c43-4c52-4779-8c6a-d513a47a71be_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..57c7d5437d5e82b0933d2dca0bc2c91887cddb1e --- /dev/null +++ b/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/d6dc5c43-4c52-4779-8c6a-d513a47a71be_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13fbcebb718d3c22cefd1eab29ab9c79d79474a48fe41791e0ea81f2cfeb25e5 +size 1919308 diff --git a/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/full.md b/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0b0bcaacf93d9e8118405c77264c7a9b309d30a3 --- /dev/null +++ b/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/full.md @@ -0,0 +1,270 @@ +# Task-Driven Wavelets using Constrained Empirical Risk Minimization + +Eric Marcus\*,1,2, Ray Sheombarsing\*,1, Jan-Jakob Sonke $^{1}$ , Jonas Teuwen $^{1}$ + +1Netherlands Cancer Institute, 2University of Amsterdam {e.marcus, r.sheombarsing, j.sonke, j.teuwen}@nki.nl + +# Abstract + +Deep Neural Networks (DNNs) are widely used for their ability to effectively approximate large classes of functions. This flexibility, however, makes the strict enforcement of constraints on DNNs a difficult problem. In contexts where it is critical to limit the function space to which certain network components belong, such as wavelets employed in Multi-Resolution Analysis (MRA), naive constraints via additional terms in the loss function are inadequate. To address this, we introduce a Convolutional Neural Network (CNN) wherein the convolutional filters are strictly constrained to be wavelets. This allows the filters to update to task-optimized wavelets during the training procedure. Our primary contribution lies in the rigorous formulation of these filters via a constrained empirical risk minimization framework, thereby providing an exact mechanism to enforce these structural constraints. While our work is grounded in theory, we investigate our approach empirically through applications in medical imaging, particularly in the task of contour prediction around various organs, achieving superior performance compared to baseline methods. + +# 1. Introduction + +Empirical risk minimization (ERM) is currently the most prevalent framework for supervised learning. The goal is to find a function to map inputs to associated targets for all representative (potentially unseen) observations. To find such a mapping, one introduces a loss function to quantify the discrepancy between observed and predicted targets. An optimal map is then found by minimizing the expected loss. In large-scale settings, such as deep learning, the resulting minimization problem is solved using Stochastic Gradient Descent (SGD) and various variants thereof, see [15, 24, 25, 40] for instance. As deep learning applications become more specialized, domain-specific needs become increasingly vital. These are often formulated in terms of constraints on + +![](images/6cbd1c25d4bb69908a39e617411cf75bda957b28c1fed880276de18fe827d1e8.jpg) +Figure 1. An overview of the Constrained Empirical Risk Minimization (CERM) framework: on the top left side, the full gradient of the loss $\mathcal{L}$ is shown on the total space $\mathcal{N}$ (containing both the constrained and unconstrained space), with $\alpha$ are the unconstrained parameters, and $\theta$ the constrained ones. The constraints can be written in the form $F(\theta) = 0$ ; the solution set $F^{-1}(0)$ is an embedded submanifold $\mathcal{M}$ of $\mathbb{R}^{\tilde{p}}$ . The constrained parameters are updated by following a path on $\mathcal{M}$ in the direction of the negative gradient on this manifold. This constrained part of the full gradient is contained in the tangent space $T_{\theta}\mathcal{M}$ of the submanifold. The color of the manifold indicates the value of the loss function. By restricting the relevant components of our descent trajectories to the embedded submanifold, we always satisfy the constraints imposed by $F$ . The gradient and parameter updates for the unconstrained parameters $\alpha$ are computed as usual using standard SGD for flat space (not shown). In our wavelet scenario, the manifold thus consists of all admissible wavelets of a certain dimension. + +the permissible mappings. For instance, the constraint for translation-equivariance led to the development and success of convolution neural networks (CNNs) [17]. In general, however, it is a highly non-trivial task to construct network architectures that satisfy a set of constraints, if they exist + +at all. It is, therefore, common practice to incorporate constraints directly into the loss function by including additional terms, usually referred to as "soft" constraints. This setup, however, has the drawback that the constraints are only approximately (on average) satisfied due to the formulation of ERM. Moreover, incorporating many different objectives in a loss function may lead to suboptimal results for the individual objectives. Another common strategy, used for many types of constraints, e.g., flip invariance, is based on data augmentation. However, not every constraint can be achieved using data augmentation, and it also leads, at best, to constraints being approximately satisfied. Furthermore, if we wish to constrain neural networks to specific function classes, such as wavelets, an additional loss term is no longer sufficient: a wavelet is only a wavelet whenever the constraints are satisfied exactly. + +Approaches which circumvent loss-based (soft) constraints have been recently proposed, see [2, 4, 28, 30, 32] for example. Most notably, we mention the field of Geometric Deep Learning (GDL), which focuses on incorporating constraints arising from symmetries of the domain directly into the networks themselves, see [8, 9, 39] and the references therein. In GDL, one considers very specific but powerful types of constraints, namely that network layers are equivariant with respect to some group action. Not all constraints arise as equivariance principles, however. A large class of examples comes from highly specialized requirements on the output of a neural network, e.g., that the output is a divergence-free vector field, a contour, or perhaps a surface. For example, in medical image segmentation, a natural requirement is that the output of the segmentation network corresponds to a continuous (closed) curve. Creating architectures for such constraints is out of the realm of GDL and, in general, poses a challenging task; no prescribed methods exist to do so. + +To facilitate these different types of constraints, we present a framework, called Constrained Empirical Risk Minimization (CERM), where we directly formulate and solve the constrained ERM problem in the space where the constraints are satisfied. This space forms, under mild conditions, a Riemannian manifold. As long as we stay on this manifold, the constraints will be satisfied exactly up to numerical precision. In particular, we explain how to perform SGD on such Riemannian manifolds arising from a finite-dimensional system of equations. Although we are motivated by the ability to constrain to wavelets, the framework, in principle, encompasses constraints of the form $F(W) = 0$ , with $W$ the network weights. Our method heavily relies on the Implicit Function Theorem, which is used to construct so-called graph-charts amenable to numerical computations. This allows for efficient evaluation of the (induced) Riemannian metric and gradients, which are vital for performing SGD on Riemannian manifolds using only the intrinsic geometry. + +As mentioned, we will present the tools and implementations of convolutional networks whose filters will form wavelets. There are several tasks at which one expects wavelet-based neural networks to excel. Wavelet decompositions naturally lend themselves to representing continuous objects such as curves, images, vector fields, or other higher-dimensional objects. Hence any task where the object of interest can be identified with a smooth or continuous function is well-suited for wavelet-based neural networks. There is an abundance of such examples to be found in computer vision, e.g., boundary prediction, image registration, and so forth. Another family of exciting applications can be found in signal analysis, e.g., in compression and denoising, where wavelets are long-standing tools that have proven to be extremely efficient [26]. The main idea in these areas is to extract information about noise, smoothness, and even singularities, through analysis of the wavelet coefficients. Subsequently, by modifying a subset of the coefficients, e.g., through thresholding, the signal can be cleaned up or denoised. More recently, wavelets have also received increasing attention in the context of neural networks. The applications are broad; some references include [22, 33, 41], ranging from efficient normalizing flows, noise-robust training, scale-equivariant models, and many more. The wavelet layers we create in this work are included in our code, which we will make available publicly*. + +Related work An earlier attempt at incorporating constraints is described in [11]. However, this method is only able to deal with linear constraints, which is hence incompatible with the wavelet constraints we will derive. Other works include [20, 21, 27], which are related to the method of Lagrange multipliers and have their optimization and training dynamics largely determined by variants of Newton's method. We discuss the differences between our SGD-compatible method and that of Lagrange multipliers in more detail in the supplementary material 6.2. A comprehensive overview of all constrained optimization methods, however, is out of the scope of this paper, and we therefore focus our attention to techniques based on SGD on Riemannian manifolds. + +Performing SGD on Riemannian manifolds is well-established and has been studied extensively, see [6, 14, 34, 35, 42]. To the best of our knowledge, current methods require an explicit description of charts on the underlying manifold and adopt a completely extrinsic point of view. This process involves manual computations on paper on a per-case basis. For instance, the authors of [34, 35] present methods for three specific cases: the space of positive definite matrices, Grassman, and Stiefel manifolds. We mention the method in [7, Chapter 7.7] in particular, which is most closely related to ours. This method adopts an extrinsic + +perspective on RSGD and effectively performs all computations in the ambient (vector) space in which $\mathcal{M}$ is embedded. This extrinsic point of view leads to an algorithm for "projecting" paths in the ambient space onto descent paths on $\mathcal{M}$ . This projection involves an additional nonlinear optimization problem and presents highly non-trivial issues for numerical implementation on general manifolds, as mentioned by the authors. In contrast, the method we present here is entirely intrinsic to $\mathcal{M}$ and uses a local coordinate system amenable to numerical computations, allowing us to deal with a general class of manifolds. In particular, it allows us to incorporate the constraints for the wavelet straightforwardly. Most importantly, we follow a specific descent path on $\mathcal{M}$ ; the geodesic in the direction of the negative gradient of the objective, a process made possible through our explicit understanding of the Riemannian metric. Our intrinsic perspective ensures that all computations are performed directly on the manifold, avoiding additional optimization problems. In particular, the dynamics of our optimization algorithm are solely driven by the gradient flow of the objective. + +Overview The contributions of this paper are ordered as follows. In Sec. 2, we introduce the theory of the constrained optimization. In Sec. 3, we apply the framework to our topic of interest: constraining filters of a CNN to form wavelets. We will show experiments in Sec. 4. In particular, this section will show our novel application, where we find data-driven wavelets for contour prediction. Specifically, we use wavelet networks to perform contour prediction in the medical domain, where we outperform strong baselines. + +We remark that an extensive, formal, and detailed investigation of the CERM framework, MRA, comparisons with other frameworks, and several implementation details are provided in [16], to which we refer the reader for more details. + +# 2. Constrained Empirical Risk Minimization + +In this section, we introduce a framework for performing ERM with constraints, which we will refer to as Constrained Empirical Risk Minimization (CERM). We start with a brief review of the traditional ERM setup [36, 37] introducing the necessary terminology, notation, and assumptions. Next, we explain how to incorporate constraints into the ERM framework in the form of a system of equations. We provide sufficient conditions on the system of equations to guarantee that the solution set is a Riemannian manifold. Finally, we provide an explanation for how the Riemannian metric and associated geometric quantities can be numerically evaluated, which enables us to perform SGD on the Riemannian manifold of interest directly. The discussion will be high-level; each step is explored in full mathematical detail in Sec. 6 [16]. + +# 2.1. ERM Setup + +Let $X$ and $Y$ be random variables whose realizations are interpreted as input data and corresponding targets, respectively. The sample spaces of $X$ and $Y$ are denoted by $\mathcal{X}$ and $\mathcal{Y}$ , respectively. As a side note, self-supervised settings also fall into this framework, in which case the target $Y$ is created on the fly as a function of $X$ . + +Empirical risk minimization The goal of the ERM framework is to find a map $G: \mathcal{X} \to \mathcal{Y}$ such that $G(x) \approx y$ for most realizations of $(X, Y)$ . The discrepancy between predicted and observed targets is quantified using a loss function $L: \mathcal{Y} \times \mathcal{Y} \to [0, \infty)$ . Without loss of generality, we assume that $L$ assumes positive values only and that $L$ decreases as the accuracy of predictions increases. The main objective of ERM is to find an optimal map $G^*$ , which solves the minimization problem + +$$ +\min _ {G \in \mathcal {G}} \mathbb {E} (L (G (X), Y)). \tag {1} +$$ + +Here $\mathcal{G}$ is a suitable subset of functions. + +In all our applications, we assume that $X$ and $Y$ are random vectors with sample spaces $\mathcal{X} = \mathbb{R}^n$ and $\mathcal{Y} = \mathbb{R}^m$ . Furthermore, we assume $\mathcal{G}$ is a parametric set that consists of mappings $\mathcal{G} = \{G(\cdot ,\xi):\xi \in \mathbb{R}^p\}$ , where $G:\mathbb{R}^n\times \mathbb{R}^p\to \mathbb{R}^m$ is a continuously differentiable map. This includes, for example, ordinary neural networks, where the parameters are the weights. With these assumptions in place, the minimization problem in (1) is equivalent to minimizing over the parameters $\xi$ instead. In practice, we have access to only a finite set of observations, which are assumed to be i.i.d., and the objective in (1) is replaced by an arithmetic average. + +# 2.2. Imposing constraints + +In this section, we explain how to incorporate constraints on a subset of the parameters $\xi$ directly into the ERM framework. We consider constraints given in the form of a system of equations. To be very precise, let $F:\mathbb{R}^{\tilde{p}}\to \mathbb{R}^{q}$ be a twice continuously differentiable map, where $\tilde{p}$ denotes the number of constrained parameters, and $q$ is the number of equations. We assume that $q < \tilde{p}\leq p$ . For notational convenience, we decompose $\mathbb{R}^p = \mathbb{R}^{p - \tilde{p}}\oplus \mathbb{R}^{\tilde{p}}$ , where the first and second subspaces correspond to the unconstrained and constrained parameters, respectively. We take $\pi_{p - \tilde{p}}:\mathbb{R}^p\to \mathbb{R}^{p - \tilde{p}}$ and $\pi_{\tilde{p}}:\mathbb{R}^p\to \mathbb{R}^{\tilde{p}}$ to be the projections onto the unconstrained and constrained parameter subspace, respectively. We will denote the unconstrained and constrained parameters by $\alpha$ and $\theta$ , respectively, i.e., $\alpha = \pi_{p - \tilde{p}}(\xi)$ and $\theta = \pi_{\tilde{p}}(\xi)$ . + +Definition 2.1 (CERM). Let $G: \mathbb{R}^n \times \mathbb{R}^p \to \mathbb{R}^m$ be a continuously differentiable parameterization of mappings and $F: \mathbb{R}^{\tilde{p}} \to \mathbb{R}^q$ a twice continuously differentiable constraint, + +where $q < \tilde{p} \leq p$ . Suppose $L: \mathbb{R}^m \times \mathbb{R}^m \to [0, \infty)$ is a continuously differentiable loss function. The constrained ERM problem for $(X, Y)$ with respect to $(G, F, L)$ is defined by + +$$ +\left\{ \begin{array}{l} \min _ {\xi \in \mathbb {R} ^ {p}} \mathbb {E} (L (G (X, \xi), Y)), \\ \text {s . t .} F (\pi_ {\tilde {p}} (\xi)) = 0. \end{array} \right. \tag {2} +$$ + +Note the generality of the admissible mappings $G$ in Definition 2.1. Although we will focus on neural networks from now on, the proposed framework applies to any parametric model, e.g., logistic or polynomial regression models. Next, we show that the CERM problem in (2) can be reformulated as an ordinary ERM problem on a Riemannian manifold $(\mathcal{N}, g_{\mathcal{N}})$ , provided that the system of equations satisfies a mild non-degeneracy condition. This result allows us to consider the admissible parameters as a geometric object in its own right, whose intrinsic geometry we use to solve (2). The proof of this theorem can be found in Sec. 6.1 [16]. + +Theorem 2.2. If zero is a regular (i.e., non-singular) value of $F$ , then the CERM problem in (2) is equivalent to solving an ordinary ERM problem on a Riemannian manifold $\mathcal{N}$ with metric $g_{\mathcal{N}}$ of dimension $p - q$ . Here $\mathcal{N} = \mathbb{R}^{p - \tilde{p}} \times \mathcal{M}$ is an embedded $C^2$ -submanifold of $\mathbb{R}^p$ and $\mathcal{M} := F^{-1}(0)$ . The equivalent minimization problem is given by + +$$ +\min _ {(\alpha , \theta) \in \mathcal {N}} \mathbb {E} (L (G (X, \alpha \oplus \iota (\theta)), Y)), \tag {3} +$$ + +where $\iota : \mathcal{M} \to \mathbb{R}^{\tilde{P}}$ is the inclusion map which embeds $\mathcal{M}$ in the ambient space. + +From now on, we refer to the objective in (3) as simply the loss $\mathcal{L}$ . + +Graph coordinates on $\mathcal{N}$ The main idea is to directly minimize the loss on $\mathcal{N}$ using SGD, see Figure 1 for an overview. To perform SGD on this space, however, we need to choose a coordinate system in which we can evaluate the gradient of the loss and subsequently follow descent trajectories on $\mathcal{N}$ . Now, the proof of Theorem 2.2 involves showing that $\mathcal{M}$ is an embedded submanifold of the ambient space $\mathbb{R}^{\tilde{p}}$ . This is accomplished by constructing a special coordinate system, a so-called graph chart or graph coordinate system. This chart is uniquely determined by the constraint $F$ and a prescribed zero $\theta \in \mathcal{M}$ . It can be numerically evaluated at the point $\theta$ by solving a linear system of equations. In turn, this enables us to construct a chart on the manifold in which we can evaluate the Riemannian metric and the gradient $\nabla_{g_{\mathcal{N}}} \mathcal{L}(\alpha, \theta)$ . + +The crucial observation is that we can numerically construct coordinate systems in which we can follow paths on the manifold $\mathcal{N}$ in the direction of $-\nabla_{g_{\mathcal{N}}}\mathcal{L}(\alpha, \theta)$ . In particular, the system of equations determining these + +Algorithm 1 Sketch of the computation of $\nabla_{g_N}\mathcal{L}(\alpha ,\theta)$ given $(\alpha ,\theta)\in \mathcal{N}$ . See Algorithm 2 in Sec. 6.5 [16] the detailed formal version. + +1: Compute the derivative $DF(\theta)$ of the constraint $F$ . +2: Compute chart-related derivatives. +3: Compute the metric $g_{\mathcal{N}}$ using the chart-related derivatives. +4: Compute the partial derivatives of the loss $\mathcal{L}$ with respect to the constrained components $\theta$ . +5: Compute the components of $\nabla_{g_{\mathcal{M}}}\mathcal{L}(\alpha ,\theta)$ using the above ingredients. +6: Compute the ordinary (unconstrained) components $\nabla_{g_{\mathrm{flat}}}\mathcal{L}(\alpha ,\theta)$ by evaluating $D_{\alpha}\mathcal{L}(\alpha ,\theta)$ + +so-called graph coordinate systems is fully determined by the derivative $DF$ of the constraint $F$ and the constraint itself. This observation enables a general implementation of graph coordinate systems independent of the explicit form of $F$ . We only require an implementation of $F$ and $DF$ , where the latter may be obtained using automatic differentiation. We provide a sketch of the algorithm for computing the gradient in Algorithm 1. For the interested reader, we mention that a complete and detailed explanation of all steps can be found in Sec. 6 [16]. Before moving on, we mention the limitations of the CERM framework for other generic constraints. + +Limitations As with most applications of deep neural networks, the available amount of computational power imposes limitations. Specifically, if the number of constrained parameters is large relative to the number of constraints, the computational demands will rise. Furthermore, setting up well-posed constraints can be a non-trivial task, e.g., symmetries in the constraints may result in redundancies obstructing the solvability of the associated systems of equations. Finally, the topology of the constrained submanifold may pose limitations as well. For instance, while the constraints will generically define a Riemannian manifold, it need not be connected. In such cases, the initialization of the network parameters will play an important role, as the gradient flow will stay on the connected component it was initialized on. + +# 3. Multiresolution Analysis and CERM + +This section presents the application of the aforementioned CERM framework to a non-trivial application to learn optimal wavelet bases for a given task. Many different types of wavelets, such as Haar, Daubechies, Gabor, etc., are used for a wide array of tasks. The specifics of the task demand different characteristics of the wavelets, such as the degree of smoothness, or the support of the wavelet. For example, + +the CDF wavelets [10] are well-known due to their usage for both lossless and lossy compression in JPEG2000. In this section, we consider networks that will find task-driven wavelets, optimizing the wavelet characteristics to the task at hand. + +Here, we will explain how to set up a system of equations (constraints) whose solution set corresponds to wavelets. We will apply these constraints in the next section to learn task-driven wavelets for predicting multi-resolution decompositions of contours in the medical domain. While we only consider one-dimensional wavelets in this work, we stress that the framework is easily adapted to higher-dimensional domains, to decompose images, for instance, by using tensor products of one-dimensional bases. + +# 3.1. Multiresolution Analysis + +In this section, we briefly state what a Multiresolution Analysis (MRA) is. We closely follow the standard exposition, found in, e.g., [29, 31], and refer the reader to Sec. 7 [16] for a more in-depth and formal introduction. The MRA framework allows one to study a signal $\gamma \in L^2 (\mathbb{R})$ at various resolution levels. The origin of this framework lies in the uncertainty principle of Fourier analysis, which states that a signal cannot be simultaneously localized in the time and frequency domain. MRA addresses this shortcoming by decomposing a signal on different discrete resolution levels. The idea is to construct a nested (exhausting) sequence of closed subspaces $V_{j}\subset L^{2}(\mathbb{R})$ , each associated with a specific resolution level $j$ . The subspaces $V_{j}$ are spanned by integer shifts of an appropriate dilation of a so-called scaling map (or father wavelet) $\varphi$ . The level of dilation determines the resolution level. Formally, we require that $(\varphi_{jk})_{k\in \mathbb{Z}}$ is an orthonormal basis for $V_{j}$ , where $\varphi_{jk}(t)\coloneqq 2^{\frac{j}{2}}\varphi (2^{j}t - k)$ . In Fig. 2a and Fig. 2c, we show an example of that Haar-MRA; in particular, we show the father wavelet and an example of a subspace at a certain resolution level. + +Decomposing a signal To analyze a signal $\gamma$ at different resolution levels, we project it onto the subspaces $V_{j}$ . The orthogonal projection of $\gamma$ at level $j$ is denoted by $\gamma_{j}$ . The coefficients of $\gamma_{j}$ with respect to the basis for $V_{j}$ , denoted by $a_{jk}(\gamma)$ , are called the approximation coefficients of $\gamma$ . When a signal in $V_{j+1}$ is projected onto $V_{j}$ , the lost information is stored in the orthogonal complement $W_{j}$ of $V_{j}$ in $V_{j+1}$ . This subspace is referred to as the detail subspace at level $j$ . The decomposition of $V_{j+1}$ into $V_{j}$ and $W_{j}$ allows the reconstruction of a signal at level $j+1$ from any lower level $j_{0}$ provided all the details in between are known. The corresponding decomposition and reconstruction algorithms give rise to the Discrete Wavelet Transform (DWT) and Mallat's Pyramid Algorithm. In Fig. 2b and Fig. 2d, we show examples of the mother wavelet and the detail subspace for the Haar wavelet. + +A fundamental result, known as Mallat's Theorem, states + +![](images/6b37e34bbbf1201ad4b129c2a50d686a0acb69edf924732db44394a509ffcf32.jpg) + +![](images/85f0688522d302ac7328cc2eb5449b364d97056a6b0804ae3e0775972e8fdb3e.jpg) + +![](images/2e28b94e07a13d486a1fbd061d54edd049b941fd65951b183825a66efe46e535.jpg) +(a) Father wavelet +(c) $V_{j}$ +Figure 2. Example of the Haar MRA: (a) Dilated translation of the Haar scaling map (or father wavelet) $\varphi = \mathbf{1}_{[0,1)}$ . The approximation subspace $V_{j}$ at level $j$ consists of all step-functions with step-size $2^{-j}$ . (b) Dilated translation of the mother wavelet $\psi = \mathbf{1}_{[0,\frac{1}{2})} - \mathbf{1}_{[\frac{1}{2},1)}$ . The detail subspace $W_{j}$ is spanned by integer translations of the dilated mother wavelet $\psi_{j}$ . (c) The approximation subspace at level $j$ consists of all step-functions with step-size $2^{-j}$ . (d) Example of a function in the detail subspace at level $j$ . + +![](images/7ec91ceaa1e072de490ec3a80c4c778e89db15023c12ae3664c503b375020418.jpg) +(b) Mother wavelet +(d) $W_{j}$ + +that the subspaces $W_{j}$ can also be spanned by dilating and shifting a single map. More precisely, there exists a map $\psi \in W_0$ , the so-called mother wavelet, such that the $\psi_{jk}$ form an orthonormal basis for $W_{j}$ , see [31]. The coefficients of the projection of $(\gamma)$ onto $W_{j}$ , are denoted by $d_{j}(\gamma) \coloneqq (d_{jk}(\gamma))_{k\in \mathbb{Z}}$ , and referred to as the detail coefficients of $\gamma$ at resolution level $j$ . We can summarize the MRA procedure roughly as + +$$ +\gamma_ {j} = \sum_ {k} a _ {j _ {0} k} \varphi_ {j _ {0} k} + \sum_ {j _ {0} \leq l \leq j - 1} \sum_ {k} d _ {l k} \psi_ {l k}, \tag {4} +$$ + +where $\varphi$ denotes the father wavelet and $\psi$ the mother wavelet. The MRA boils down to the specification of the first approximation coefficient and a number of detail coefficients. + +Low and high pass filters As we mentioned before, an MRA is completely determined by a scaling map $\varphi$ , which in turn is completely characterized by a so-called low-pass filter. We will use this observation to set up a system of constraints characterizing a finite-dimensional family of wavelets, such that we can use it in the CERM framework. To explain how this works, we first make a key observation. Since $V_{0} \subset V_{1}$ , there exists a unique sequence $h$ characterizing $\varphi$ in terms of the basis $(\varphi_{1k})_{k \in \mathbb{Z}}$ for $V_{1}$ . The sequence $h$ completely characterizes the scaling function and is called the low-pass filter or wavelet filter of the MRA. Similarly, + +since $\psi \in W_0\subset V_1$ , there exists a unique sequence $g$ , the so-called high-pass filter associated to $h$ , characterizing the mother wavelet. An example of an MRA is shown in Figure 2. In practice, to define an MRA, one only needs to specify an appropriate low-pass filter $h$ , as for Mallat's mother wavelet, we have $g_{k} = (-1)^{k - 1}h_{1 - k}$ . We refer the interested reader to the supplementary information for a detailed discussion of all these points [16]. + +# 3.2. MRA constraints + +We are now ready to formulate and impose constraints on a sequence $h$ to ensure that it is the low-pass filter of a scaling map $\varphi$ . In practice, we only consider finite sequences. For notational convenience, we introduce the space $\mathcal{A}_M(\mathbb{R})$ , which consists of one-dimensional real-valued two-sided sequences $(h_k)_{k=1-M}^{M-1}$ of order $M$ ; so $\mathcal{A}_M(\mathbb{R})$ is $2M - 1$ dimensional. + +Theorem 3.1. Let $M \in \mathbb{N}_{\geq 3}$ be a prescribed order and define $F_M: \mathcal{A}_M(\mathbb{R}) \to \mathbb{R}^{M+1}$ by + +$$ +(F_{M}(h))_{k}:= \left\{ \begin{array}{ll}\sum_{|l|\leq M - 1}|h_{l}|^{2} - 1, & k = 0,\\ \sum_{1 - M + 2k\leq l\leq M - 1}h_{l - 2k}h_{l}, & 1\leq k\leq M - 1,\\ -\sqrt{2} +\sum_{|l|\leq M - 1}h_{l}, & k = M. \end{array} \right. +$$ + +If $F_{M}(h) = 0$ , and $h$ satisfies a certain non-degeneracy condition, then $h$ is the low-pass filter of a scaling map $\varphi$ generating an MRA of $L^2 (\mathbb{R})$ . + +A full proof and detailed statement of the non-degeneracy condition can be found in Sec. 7.4 in the supplementary material. Heuristically, the first $M$ equations encode the orthonormality of the scaling map. The last equation encodes the observation that the Fourier transform of the scaling map must be nonzero. The set of regular points in $F_{M}^{-1}(0)$ is a real-analytic $(M - 2)$ -dimensional submanifold of $\mathbb{R}^{2M - 1}$ . Hence, we can get as many degrees of freedom as desired, by choosing a sufficiently large order $M$ . We now have an end-to-end way of learning wavelets for specific tasks, a wavelet network, by using the constraint $F_{M}$ along with the CERM framework. + +# 4. Experiments + +In this section, we show some explicit examples of the wavelet networks with respect to contour prediction. In particular, we will utilize the derived wavelet filters of Sec. 3.2 to create a network for autocontouring. This task fits well with the setup of wavelets, as they naturally lend themselves to representing continuous objects such as curves. + +![](images/77d0f409aa4320a120744f7d19520fee3386ba36cafbdbc793c0827a4e741b9b.jpg) +Figure 3. A schematic picture of our network. The encoder consists of residual convolutional blocks depicted in blue. The first residual convolutional block uses $n_f$ filters and is doubled after every other residual convolutional block. Attached to the encoder are fully connected layers to predict approximation and detail coefficients at the lowest resolution level $j_0$ . The approximation and detail coefficients are supplied as input to the Pyramid Algorithm (the decoder) to predict a contour on high-resolution level. Each green block corresponds to an inverse discrete wavelet transform. Detail coefficients at higher levels are computed using skip-connections (arrows in red). We only predict detail coefficients up to level $j_1$ . In this example, we have set $j_1 = j_0 + 2$ and $j_2 = j_0 + 3$ . In reality, the decoder consists of two upsampling paths, one for each spatial component of the curve. We have only drawn one for notational convenience. During training, only the approximation coefficients at the highest resolution level are supervised. + +# 4.1. Contour Prediction using MRA + +In the context of contour prediction, wavelets can be used to represent the boundary of a region in an image using a simple closed curve. We consider two-dimensional gray-valued images $x \in \mathcal{X} \coloneqq [0,1]^{n \times n}$ . For each of our applications, we assume that these images contain an organ with a boundary. Formally speaking, this means that each image $x$ contains a (uniquely identifiable) simply connected region $R(x)$ , the organ, with boundary $\partial R(x)$ . This boundary of the region, $\partial R(x)$ , can be parameterized by a simple closed, piecewise $C^2$ , curve $\gamma(x)$ . We will develop a deep learning framework for computing such parameterizations by learning a multiresolution decomposition of $\gamma(x)$ using the methods developed in the previous sections. + +# 4.1.1 Data + +We have used public datasets from the Medical Decathlon Challenge [3]. The selected data consists of CT scans of the spleen of size $512 \times 512$ and T2-weighted MRI images of the prostate central gland, henceforth abbreviated as just the prostate, of size $320 \times 320$ . The scans were cropped to size $224 \times 224$ and $192 \times 192$ , respectively. Further- + +![](images/4414cf31382f8d2718b2edefad30e87caf76fe60887c0a05341613a2529f9256.jpg) +Figure 4. Examples of predictions on the test set for the spleen and prostate, depicted in the first and second row, respectively, for the best-performing wavelet models. The green curve corresponds to the ground truth, while the red curve is a prediction made by the wavelet network. + +![](images/6c3cf0900468a100ab5a98f5a8092c9a42251f905e7a2f5e04ec9a840d0ec2f9.jpg) +(a) Initial father and mother wavelet + +![](images/6ce08152be65b95f01d7a0a631d6bfc4fec432d3559d5226bdbba0d16e8b5cf8.jpg) +(b) Task-optimized father and mother wavelet +Figure 5. Example of wavelets of order 5 learned during training of the spleen model: (a) The initial father (left) and mother (right) wavelets for the second spatial component. (b) The task-optimized father and mother wavelets after training for the second spatial component. We observe that the task-optimized wavelets are more simple and have "cleaned up" in a sense. + +more, the images were resampled to the median sample spacing, which resulted in $(5.00\mathrm{mm}, 0.793\mathrm{mm}, 0.793\mathrm{mm})$ and $(3.6\mathrm{mm}, 0.625\mathrm{mm}, 0.625\mathrm{mm})$ spacings for the spleen and prostate, respectively. We refer the reader to Sec. 8.2 [16] for a detailed description of all the preprocessing steps and construction of ground truth curves. Examples of predictions on the test set, for both the spleen and prostate, are shown in Figure Fig. 4. + +# 4.1.2 Model and Training + +Model objective We parameterize the boundary $\partial R(x)$ by arc length, resulting in a curve $\gamma (x)$ . The precise details of this parametrization are outlined in the supplementary information. The objective is to compute the relevant approximation coefficients of $\gamma (x)$ . To this end, let $j_0,j_1,j_2\in \mathbb{N}$ be resolution levels, where $j_0\leq j_1\leq j_2$ . We will construct a convolutional neural network + +$$ +G: \mathcal {X} \times \mathbb {R} ^ {p} \to \prod_ {j = j _ {0}} ^ {j _ {2}} \mathbb {R} ^ {2 ^ {j}} \times \mathbb {R} ^ {2 ^ {j}} \times \prod_ {j = j _ {0}} ^ {j _ {1} - 1} \mathbb {R} ^ {2 ^ {j}} \times \mathbb {R} ^ {2 ^ {j}}, +$$ + +which predicts the wavelet decomposition of $\gamma (x)$ . Here the subspaces $\mathbb{R}^{2^j}$ correspond to approximation and detail coefficients at level $j$ , one for each spatial component. Furthermore, a subset of the parameters $\xi \in \mathbb{R}^p$ are constrained to be wavelet filters, one wavelet filter per spatial component, using Theorem 3.1. The map $G(\cdot ,\xi)$ applied to an image $x$ has output + +$$ +G (x, \xi) = \left(a _ {j _ {0}} (x, \xi), \dots , a _ {j _ {2}} (x, \xi), \right. \tag {5} +$$ + +$$ +d _ {j _ {0}} (x, \xi), \dots , d _ {j _ {1} - 1} (x, \xi)). +$$ + +Here $a_{j}(x,\xi)$ and $d_{j}(x,\xi)$ are predictions for the approximation and detail coefficients of $\gamma (x)$ at level $j$ , respectively. We only predict detail coefficients up to level $j_{1}$ . + +Architecture The network is a hybrid analog of the U-Net. It consists of a two-dimensional convolutional encoder, a bottleneck of fully connected layers, and a one-dimensional decoder. The encoder and decoder are connected through skip-connections. The approximation and detail coefficients at the lowest resolution level $j_0$ are predicted in the bottleneck. Afterward, the Pyramid Algorithm takes over to compute approximation coefficients at higher resolution levels (the decoder) using learnable wavelet filters. In Fig. 3, we provide a schematic overview of the network architecture. + +Loss To measure the discrepancy between the ground truth and the predicted curve, we define + +$$ +\begin{array}{r l} L (G (x, \xi), \tilde {a} (\gamma (x))) := & \| [ a _ {j _ {2}} ] _ {1} - \tilde {a} _ {j _ {2}} ([ \gamma (x) ] _ {1}) \| _ {2} + \\ & \| [ a _ {j _ {2}} ] _ {2} - \tilde {a} _ {j _ {2}} ([ \gamma (x) ] _ {2}) \| _ {2}. \end{array} \tag {6} +$$ + +The $\tilde{a}$ thus correspond to the ground-truth approximation coefficients, and the subscripts 1, 2 denote the spatial dimension. In other words, the loss corresponds to the componentwise $L^2$ -error between the curves on resolution level $j_2$ with approximation coefficients $a_{j_2}(x,\xi)$ and $\tilde{a}_{j_2}(\gamma (x))$ . + +Optimization We use two different optimizers during training: one for the encoder and MLP network (unconstrained), and one for the decoder (constrained wavelet network). The reason for this is that the needed step sizes on the non-trivial submanifold may significantly differ from those on the unconstrained (flat) parameter space. For the training of the model, we use plain SGD for the first eight epochs for both the constrained and unconstrained parameters. During this period, the learning rate for the unconstrained parameters is linearly increased from $10^{-5}$ to $2 \cdot 10^{-4}$ . The learning rate for the constrained parameters (wavelet filters) is linearly increased from $10^{-4}$ to $10^{-2}$ . After the initial warmup stage, we switch to the Adam optimizer for the unconstrained parameters. For both the constrained and unconstrained parameters, we use learning rate schedulers and decrease the learning rate by a factor of 0.85 if no significant improvements in the validation loss are observed during the last ten epochs. We train all models for 250 epochs using a batch size of 32 and use the last epoch for inference. The computations were performed in PYTORCH on a Geforce RTX 2080 Ti. An average training session takes about two and a half hours for the wavelet network and a day for the nnUNet. The inference times for the wavelet network were 12.5ms and 14ms for the prostate and spleen, respectively, using a batch size of 32. The inference times for the nnUNet were 17.8ms and 23.4ms for the prostate and spleen, respectively, using a batch size of 32. + +# 4.1.3 Results + +We have evaluated the performance of both our wavelet networks for different orders and a state-of-the-art baseline 2d-nnUNet [13] on the unseen test data, see Table 1. For both the spleen and prostate, we observe that the best wavelet networks outperform the baseline. We note that the task-optimized wavelets differ significantly from the wavelets randomly initialized at the start of training. A comparison of an initial and task-optimized wavelet is depicted in Figure 5, see Sec. 8.4 for more examples. + +Table 1. Mean and standard deviation (in parentheses) of the dice score on the unseen test sets. The number of parameters in our wavelet-based networks ranged from 5M to 10M parameters, while the nnUNet used approximately 40M and 18M parameters for the spleen and prostate, respectively. + +
MODELDICE SPLEENDICE PROSTATE
NNUNET0.914 (1.74 · 10-1)0.896 (1.27 · 10-1)
ORDER 30.911 (7.38 · 10-2)0.929 (4.66 · 10-2)
ORDER 40.911 (7.85 · 10-2)0.935 (3.48 · 10-2)
ORDER 50.916 (6.94 · 10-2)0.935 (4.11 · 10-2)
ORDER 60.917 (7.17 · 10-2)0.934 (4.14 · 10-2)
ORDER 70.921 (6.91 · 10-2)0.928 (4.03 · 10-2)
ORDER 80.919 (6.64 · 10-2)0.934 (3.62 · 10-2)
+ +# 5. Conclusions + +In this paper, we have introduced the CERM framework for creating task-driven wavelets. Although our aim is to find such wavelets, in principle the framework lends itself to imposing generic constraints on parametric models. The wavelet constraints can be formulated as a finite system of equations. Under mild smoothness and non-degeneracy conditions, the networks can be made to obey the constraints exactly throughout the entire training procedure by performing SGD on a curved space. We can then utilize these wavelet constraints in a convolutional network, where the filters of the network now parameterize wavelets themselves. We have applied these wavelet networks to the prediction of boundaries of simply connected regions in medical images, where they outperform strong baselines. + +# References + +[1] P. Abry and P. Flandrin. On the initialization of the discrete wavelet transform algorithm. IEEE Signal Processing Letters, 1(2):32-34, 1994. 14 +[2] Brandon Amos, Lei Xu, and J. Zico Kolter. Input convex neural networks. In Proceedings of the 34th International Conference on Machine Learning, pages 146-155. PMLR, 2017. 2 +[3] Michela Antonelli, Annika Reinke, Spyridon Bakas, Keyvan Farahani, Annette Kopp-Schneider, Bennett A Landman, Geert Litjens, Bjoern Menze, Olaf Ronneberger, Ronald M Summers, et al. The medical segmentation decathlon. Nature communications, 13(1):1-13, 2022. 6 +[4] Randall Balestriero and Yann LeCun. Police: Provably optimal linear constraint enforcement for deep neural networks, 2022. 2 +[5] D.P. Bertsekas and W. Rheinboldt. Constrained Optimization and Lagrange Multiplier Methods. Elsevier Science, 2014. 1 +[6] Silvere Bonnabel. Stochastic gradient descent on riemannian manifolds. IEEE Transactions on Automatic Control, 58(9): 2217-2229, 2013. 2, 1, 5 + +[7] N. Boumal. An Introduction to Optimization on Smooth Manifolds. Cambridge University Press, 2023. 2 +[8] Michael M Bronstein, Joan Bruna, Yann LeCun, Arthur Szlam, and Pierre Vandergheynst. Geometric deep learning: going beyond euclidean data. IEEE Signal Processing Magazine, 34(4):18-42, 2017. 2 +[9] Michael M Bronstein, Joan Bruna, Taco Cohen, and Petar Velicković. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges. arXiv preprint arXiv:2104.13478, 2021. 2 +[10] Albert Cohen, Ingrid Daubechies, and J-C Feauveau. Biorthogonal bases of compactly supported wavelets. Communications on pure and applied mathematics, 45(5):485-560, 1992. 5 +[11] Priya Donti, David Rolnick, and J Zico Kolter. Dc3: A learning method for optimization with hard constraints. In International Conference on Learning Representations, 2021. 2 +[12] M.W. Frazier. An Introduction to Wavelets Through Linear Algebra. Springer New York, 2001. 12 +[13] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 8 +[14] Hiroyuki Kasai, Pratik Jawanpuria, and Bamdev Mishra. Riemannian adaptive stochastic gradient algorithms on matrix manifolds. In International Conference on Machine Learning, pages 3262-3271. PMLR, 2019. 2, 1, 5 +[15] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings. International Conference on Learning Representations, ICLR, 2015. 1 +[16]FirstName LastName. Task-driven wavelets using constrained empirical risk minimization, 2023. Supplied as supplemental material supplementary-notes.pdf. 3, 4, 5, 6, 7 +[17] Yann LeCun, Yoshua Bengio, et al. Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks, 3361(10):1995, 1995. 1 +[18] John M Lee. Riemannian manifolds: an introduction to curvature. Springer Science & Business Media, 2006. 6, 8 +[19] John M Lee. Smooth manifolds. In Introduction to smooth manifolds. Springer, 2013. 2, 6, 8 +[20] Benedict Leimkuhler, Timothee Pouchon, Tiffany Vlaar, and Amos Storkey. Constraint-based regularization of neural networks. arXiv preprint arXiv:2006.10114, 2020. 2 +[21] Benedict Leimkuhler, Tiffany Vlaar, Timothee Pouchon, and Amos Storkey. Better training using weight-constrained stochastic dynamics. arXiv preprint arXiv:2106.10704, 2021. 2 +[22] Qiufu Li, Linlin Shen, Sheng Guo, and Zhihui Lai. Wavelet integrated cnns for noise-robust image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7245–7254, 2020. 2 +[23] Cho-Chun Liu, Yaohui Liu, Zhengding Qiu, and Xiyu Du. On the initialization of the discrete wavelet transform algorithm. 1997 IEEE International Conference on Intelligent Processing Systems (Cat. No.97TH8335), 2:1220-1222 vol.2, 1997. 14 + +[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 1 +[25] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam, 2018. 1 +[26] Stephane Mallat. A Wavelet Tour of Signal Processing, Third Edition: The Sparse Way. Academic Press, Inc., USA, 3rd edition, 2008. 2, 10 +[27] Pablo Márquez-Neila, Mathieu Salzmann, and Pascal Fua. Imposing hard constraints on deep networks: Promises and limitations. arXiv preprint arXiv:1706.02025, 2017. 2 +[28] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In International Conference on Learning Representations, 2018. 2 +[29] L. Montefusco and L. Puccio. Wavelets: Theory, Algorithms, and Applications. Elsevier Science, 2014. 5, 11, 12, 14 +[30] Deepak Pathak, Philipp Krahenbuhl, and Trevor Darrell. Constrained convolutional neural networks for weakly supervised segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1796-1804, 2015. 2 +[31] M.C. Pereyra and L.A. Ward. Harmonic Analysis: From Fourier to Wavelets. American Mathematical Society, 2012. 5, 11, 12 +[32] Jack Richter-Powell, Yaron Lipman, and Ricky T. Q. Chen. Neural conservation laws: A divergence-free perspective, 2022. 2 +[33] David W Romero, Erik J Bekkers, Jakub M Tomczak, and Mark Hoogendoorn. Wavelet networks: Scale equivariant learning from raw waveforms. arXiv preprint arXiv:2006.05259, 2020. 2 +[34] Soumava Kumar Roy, Zakaria Mhammedi, and Mehrtash Harandi. Geometry aware constrained optimization techniques for deep learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4460-4469, 2018. 2, 1, 5 +[35] Hiroyuki Sato, Hiroyuki Kasai, and Bamdev Mishra. Riemannian stochastic variance reduced gradient algorithm with retraction and vector transport. SIAM Journal on Optimization, 29(2):1444-1472, 2019. 2, 1, 5 +[36] Vladimir Vapnik. Principles of risk minimization for learning theory. Advances in neural information processing systems, 4, 1991. 3 +[37] Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 1999. 3 +[38] D.F. Walnut. An Introduction to Wavelet Analysis. Birkhäuser Boston, 2002. 12, 15 +[39] Maurice Weiler, Patrick Forre, Erik Verlinde, and Max Welling. Coordinate independent convolutional networks-isometry and gauge equivariant convolutions on riemannian manifolds. arXiv preprint arXiv:2106.06020, 2021. 2 +[40] Yang You, Jing Li, Sashank Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes. In International Conference on Learning Representations, 2020. 1 +[41] Jason J Yu, Konstantinos G Derpanis, and Marcus A Brubaker. Wavelet flow: Fast training of high resolution normalizing + +flows. Advances in Neural Information Processing Systems, 33:6184-6196, 2020. 2 +[42] Hongyi Zhang, Sashank J. Reddi, and Suvrit Sra. Riemannian svrg: Fast stochastic optimization on riemannian manifolds. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 2, 1, 5 \ No newline at end of file diff --git a/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/images.zip b/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..61143d816829d1fe6002f20143a18db63e59f169 --- /dev/null +++ b/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:439f2d741a6f88464f50be6a4870ae04607891346d7a663fcbb7518ed917cd5a +size 232603 diff --git a/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/layout.json b/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fcbab453dd7994a6438267c447f22610a25d2e94 --- /dev/null +++ b/2024/Task-Driven Wavelets using Constrained Empirical Risk Minimization/layout.json @@ -0,0 +1,10400 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 79, + 103, + 515, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 103, + 515, + 121 + ], + "spans": [ + { + "bbox": [ + 79, + 103, + 515, + 121 + ], + "type": "text", + "content": "Task-Driven Wavelets using Constrained Empirical Risk Minimization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "spans": [ + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "type": "text", + "content": "Eric Marcus\\*,1,2, Ray Sheombarsing\\*,1, Jan-Jakob Sonke" + }, + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "type": "text", + "content": ", Jonas Teuwen" + }, + { + "bbox": [ + 118, + 142, + 474, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 153, + 171, + 436, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 171, + 436, + 200 + ], + "spans": [ + { + "bbox": [ + 153, + 171, + 436, + 200 + ], + "type": "text", + "content": "1Netherlands Cancer Institute, 2University of Amsterdam {e.marcus, r.sheombarsing, j.sonke, j.teuwen}@nki.nl" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 252, + 290, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 492 + ], + "type": "text", + "content": "Deep Neural Networks (DNNs) are widely used for their ability to effectively approximate large classes of functions. This flexibility, however, makes the strict enforcement of constraints on DNNs a difficult problem. In contexts where it is critical to limit the function space to which certain network components belong, such as wavelets employed in Multi-Resolution Analysis (MRA), naive constraints via additional terms in the loss function are inadequate. To address this, we introduce a Convolutional Neural Network (CNN) wherein the convolutional filters are strictly constrained to be wavelets. This allows the filters to update to task-optimized wavelets during the training procedure. Our primary contribution lies in the rigorous formulation of these filters via a constrained empirical risk minimization framework, thereby providing an exact mechanism to enforce these structural constraints. While our work is grounded in theory, we investigate our approach empirically through applications in medical imaging, particularly in the task of contour prediction around various organs, achieving superior performance compared to baseline methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 517, + 128, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 128, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 128, + 529 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 538, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 288, + 693 + ], + "type": "text", + "content": "Empirical risk minimization (ERM) is currently the most prevalent framework for supervised learning. The goal is to find a function to map inputs to associated targets for all representative (potentially unseen) observations. To find such a mapping, one introduces a loss function to quantify the discrepancy between observed and predicted targets. An optimal map is then found by minimizing the expected loss. In large-scale settings, such as deep learning, the resulting minimization problem is solved using Stochastic Gradient Descent (SGD) and various variants thereof, see [15, 24, 25, 40] for instance. As deep learning applications become more specialized, domain-specific needs become increasingly vital. These are often formulated in terms of constraints on" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 356, + 240, + 499, + 415 + ], + "blocks": [ + { + "bbox": [ + 356, + 240, + 499, + 415 + ], + "lines": [ + { + "bbox": [ + 356, + 240, + 499, + 415 + ], + "spans": [ + { + "bbox": [ + 356, + 240, + 499, + 415 + ], + "type": "image", + "image_path": "6cbd1c25d4bb69908a39e617411cf75bda957b28c1fed880276de18fe827d1e8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "lines": [ + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "text", + "content": "Figure 1. An overview of the Constrained Empirical Risk Minimization (CERM) framework: on the top left side, the full gradient of the loss " + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "text", + "content": " is shown on the total space " + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "text", + "content": " (containing both the constrained and unconstrained space), with " + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "text", + "content": " are the unconstrained parameters, and " + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "text", + "content": " the constrained ones. The constraints can be written in the form " + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "inline_equation", + "content": "F(\\theta) = 0" + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "text", + "content": "; the solution set " + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "inline_equation", + "content": "F^{-1}(0)" + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "text", + "content": " is an embedded submanifold " + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{\\tilde{p}}" + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "text", + "content": ". The constrained parameters are updated by following a path on " + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "text", + "content": " in the direction of the negative gradient on this manifold. This constrained part of the full gradient is contained in the tangent space " + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "inline_equation", + "content": "T_{\\theta}\\mathcal{M}" + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "text", + "content": " of the submanifold. The color of the manifold indicates the value of the loss function. By restricting the relevant components of our descent trajectories to the embedded submanifold, we always satisfy the constraints imposed by " + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "text", + "content": ". The gradient and parameter updates for the unconstrained parameters " + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 441, + 547, + 628 + ], + "type": "text", + "content": " are computed as usual using standard SGD for flat space (not shown). In our wavelet scenario, the manifold thus consists of all admissible wavelets of a certain dimension." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "the permissible mappings. For instance, the constraint for translation-equivariance led to the development and success of convolution neural networks (CNNs) [17]. In general, however, it is a highly non-trivial task to construct network architectures that satisfy a set of constraints, if they exist" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 215, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 215, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 215, + 713 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24098" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "at all. It is, therefore, common practice to incorporate constraints directly into the loss function by including additional terms, usually referred to as \"soft\" constraints. This setup, however, has the drawback that the constraints are only approximately (on average) satisfied due to the formulation of ERM. Moreover, incorporating many different objectives in a loss function may lead to suboptimal results for the individual objectives. Another common strategy, used for many types of constraints, e.g., flip invariance, is based on data augmentation. However, not every constraint can be achieved using data augmentation, and it also leads, at best, to constraints being approximately satisfied. Furthermore, if we wish to constrain neural networks to specific function classes, such as wavelets, an additional loss term is no longer sufficient: a wavelet is only a wavelet whenever the constraints are satisfied exactly." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 267, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 289, + 495 + ], + "type": "text", + "content": "Approaches which circumvent loss-based (soft) constraints have been recently proposed, see [2, 4, 28, 30, 32] for example. Most notably, we mention the field of Geometric Deep Learning (GDL), which focuses on incorporating constraints arising from symmetries of the domain directly into the networks themselves, see [8, 9, 39] and the references therein. In GDL, one considers very specific but powerful types of constraints, namely that network layers are equivariant with respect to some group action. Not all constraints arise as equivariance principles, however. A large class of examples comes from highly specialized requirements on the output of a neural network, e.g., that the output is a divergence-free vector field, a contour, or perhaps a surface. For example, in medical image segmentation, a natural requirement is that the output of the segmentation network corresponds to a continuous (closed) curve. Creating architectures for such constraints is out of the realm of GDL and, in general, poses a challenging task; no prescribed methods exist to do so." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": "To facilitate these different types of constraints, we present a framework, called Constrained Empirical Risk Minimization (CERM), where we directly formulate and solve the constrained ERM problem in the space where the constraints are satisfied. This space forms, under mild conditions, a Riemannian manifold. As long as we stay on this manifold, the constraints will be satisfied exactly up to numerical precision. In particular, we explain how to perform SGD on such Riemannian manifolds arising from a finite-dimensional system of equations. Although we are motivated by the ability to constrain to wavelets, the framework, in principle, encompasses constraints of the form " + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "F(W) = 0" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 498, + 289, + 715 + ], + "type": "text", + "content": " the network weights. Our method heavily relies on the Implicit Function Theorem, which is used to construct so-called graph-charts amenable to numerical computations. This allows for efficient evaluation of the (induced) Riemannian metric and gradients, which are vital for performing SGD on Riemannian manifolds using only the intrinsic geometry." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 548, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 548, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 548, + 384 + ], + "type": "text", + "content": "As mentioned, we will present the tools and implementations of convolutional networks whose filters will form wavelets. There are several tasks at which one expects wavelet-based neural networks to excel. Wavelet decompositions naturally lend themselves to representing continuous objects such as curves, images, vector fields, or other higher-dimensional objects. Hence any task where the object of interest can be identified with a smooth or continuous function is well-suited for wavelet-based neural networks. There is an abundance of such examples to be found in computer vision, e.g., boundary prediction, image registration, and so forth. Another family of exciting applications can be found in signal analysis, e.g., in compression and denoising, where wavelets are long-standing tools that have proven to be extremely efficient [26]. The main idea in these areas is to extract information about noise, smoothness, and even singularities, through analysis of the wavelet coefficients. Subsequently, by modifying a subset of the coefficients, e.g., through thresholding, the signal can be cleaned up or denoised. More recently, wavelets have also received increasing attention in the context of neural networks. The applications are broad; some references include [22, 33, 41], ranging from efficient normalizing flows, noise-robust training, scale-equivariant models, and many more. The wavelet layers we create in this work are included in our code, which we will make available publicly*." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 397, + 547, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 547, + 564 + ], + "type": "text", + "content": "Related work An earlier attempt at incorporating constraints is described in [11]. However, this method is only able to deal with linear constraints, which is hence incompatible with the wavelet constraints we will derive. Other works include [20, 21, 27], which are related to the method of Lagrange multipliers and have their optimization and training dynamics largely determined by variants of Newton's method. We discuss the differences between our SGD-compatible method and that of Lagrange multipliers in more detail in the supplementary material 6.2. A comprehensive overview of all constrained optimization methods, however, is out of the scope of this paper, and we therefore focus our attention to techniques based on SGD on Riemannian manifolds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 564, + 548, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 548, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 548, + 696 + ], + "type": "text", + "content": "Performing SGD on Riemannian manifolds is well-established and has been studied extensively, see [6, 14, 34, 35, 42]. To the best of our knowledge, current methods require an explicit description of charts on the underlying manifold and adopt a completely extrinsic point of view. This process involves manual computations on paper on a per-case basis. For instance, the authors of [34, 35] present methods for three specific cases: the space of positive definite matrices, Grassman, and Stiefel manifolds. We mention the method in [7, Chapter 7.7] in particular, which is most closely related to ours. This method adopts an extrinsic" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 434, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 434, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 434, + 712 + ], + "type": "text", + "content": "*https://github.com/NKI-AI/CERM" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24099" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": "perspective on RSGD and effectively performs all computations in the ambient (vector) space in which " + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": " is embedded. This extrinsic point of view leads to an algorithm for \"projecting\" paths in the ambient space onto descent paths on " + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": ". This projection involves an additional nonlinear optimization problem and presents highly non-trivial issues for numerical implementation on general manifolds, as mentioned by the authors. In contrast, the method we present here is entirely intrinsic to " + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": " and uses a local coordinate system amenable to numerical computations, allowing us to deal with a general class of manifolds. In particular, it allows us to incorporate the constraints for the wavelet straightforwardly. Most importantly, we follow a specific descent path on " + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": "; the geodesic in the direction of the negative gradient of the objective, a process made possible through our explicit understanding of the Riemannian metric. Our intrinsic perspective ensures that all computations are performed directly on the manifold, avoiding additional optimization problems. In particular, the dynamics of our optimization algorithm are solely driven by the gradient flow of the objective." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 338, + 288, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 288, + 445 + ], + "type": "text", + "content": "Overview The contributions of this paper are ordered as follows. In Sec. 2, we introduce the theory of the constrained optimization. In Sec. 3, we apply the framework to our topic of interest: constraining filters of a CNN to form wavelets. We will show experiments in Sec. 4. In particular, this section will show our novel application, where we find data-driven wavelets for contour prediction. Specifically, we use wavelet networks to perform contour prediction in the medical domain, where we outperform strong baselines." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 446, + 288, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 288, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 288, + 494 + ], + "type": "text", + "content": "We remark that an extensive, formal, and detailed investigation of the CERM framework, MRA, comparisons with other frameworks, and several implementation details are provided in [16], to which we refer the reader for more details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 511, + 278, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 278, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 278, + 525 + ], + "type": "text", + "content": "2. Constrained Empirical Risk Minimization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": "In this section, we introduce a framework for performing ERM with constraints, which we will refer to as Constrained Empirical Risk Minimization (CERM). We start with a brief review of the traditional ERM setup [36, 37] introducing the necessary terminology, notation, and assumptions. Next, we explain how to incorporate constraints into the ERM framework in the form of a system of equations. We provide sufficient conditions on the system of equations to guarantee that the solution set is a Riemannian manifold. Finally, we provide an explanation for how the Riemannian metric and associated geometric quantities can be numerically evaluated, which enables us to perform SGD on the Riemannian manifold of interest directly. The discussion will be high-level; each step is explored in full mathematical detail in Sec. 6 [16]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 384, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 384, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 384, + 85 + ], + "type": "text", + "content": "2.1. ERM Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "text", + "content": " be random variables whose realizations are interpreted as input data and corresponding targets, respectively. The sample spaces of " + }, + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "text", + "content": " are denoted by " + }, + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "text", + "content": ", respectively. As a side note, self-supervised settings also fall into this framework, in which case the target " + }, + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "text", + "content": " is created on the fly as a function of " + }, + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 90, + 547, + 162 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 176, + 547, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 176, + 547, + 284 + ], + "spans": [ + { + "bbox": [ + 304, + 176, + 547, + 284 + ], + "type": "text", + "content": "Empirical risk minimization The goal of the ERM framework is to find a map " + }, + { + "bbox": [ + 304, + 176, + 547, + 284 + ], + "type": "inline_equation", + "content": "G: \\mathcal{X} \\to \\mathcal{Y}" + }, + { + "bbox": [ + 304, + 176, + 547, + 284 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 304, + 176, + 547, + 284 + ], + "type": "inline_equation", + "content": "G(x) \\approx y" + }, + { + "bbox": [ + 304, + 176, + 547, + 284 + ], + "type": "text", + "content": " for most realizations of " + }, + { + "bbox": [ + 304, + 176, + 547, + 284 + ], + "type": "inline_equation", + "content": "(X, Y)" + }, + { + "bbox": [ + 304, + 176, + 547, + 284 + ], + "type": "text", + "content": ". The discrepancy between predicted and observed targets is quantified using a loss function " + }, + { + "bbox": [ + 304, + 176, + 547, + 284 + ], + "type": "inline_equation", + "content": "L: \\mathcal{Y} \\times \\mathcal{Y} \\to [0, \\infty)" + }, + { + "bbox": [ + 304, + 176, + 547, + 284 + ], + "type": "text", + "content": ". Without loss of generality, we assume that " + }, + { + "bbox": [ + 304, + 176, + 547, + 284 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 176, + 547, + 284 + ], + "type": "text", + "content": " assumes positive values only and that " + }, + { + "bbox": [ + 304, + 176, + 547, + 284 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 176, + 547, + 284 + ], + "type": "text", + "content": " decreases as the accuracy of predictions increases. The main objective of ERM is to find an optimal map " + }, + { + "bbox": [ + 304, + 176, + 547, + 284 + ], + "type": "inline_equation", + "content": "G^*" + }, + { + "bbox": [ + 304, + 176, + 547, + 284 + ], + "type": "text", + "content": ", which solves the minimization problem" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 378, + 294, + 545, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 294, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 378, + 294, + 545, + 312 + ], + "type": "interline_equation", + "content": "\\min _ {G \\in \\mathcal {G}} \\mathbb {E} (L (G (X), Y)). \\tag {1}", + "image_path": "6012f0bab7776d54197014865a0c3bd6e3df7a69706153ee402f37abab9303b7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 320, + 467, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 320, + 467, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 320, + 467, + 331 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 305, + 320, + 467, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 305, + 320, + 467, + 331 + ], + "type": "text", + "content": " is a suitable subset of functions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "text", + "content": "In all our applications, we assume that " + }, + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "text", + "content": " are random vectors with sample spaces " + }, + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\mathbb{R}^n" + }, + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = \\mathbb{R}^m" + }, + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "text", + "content": ". Furthermore, we assume " + }, + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "text", + "content": " is a parametric set that consists of mappings " + }, + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{G(\\cdot ,\\xi):\\xi \\in \\mathbb{R}^p\\}" + }, + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "inline_equation", + "content": "G:\\mathbb{R}^n\\times \\mathbb{R}^p\\to \\mathbb{R}^m" + }, + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "text", + "content": " is a continuously differentiable map. This includes, for example, ordinary neural networks, where the parameters are the weights. With these assumptions in place, the minimization problem in (1) is equivalent to minimizing over the parameters " + }, + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 332, + 547, + 476 + ], + "type": "text", + "content": " instead. In practice, we have access to only a finite set of observations, which are assumed to be i.i.d., and the objective in (1) is replaced by an arithmetic average." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 483, + 427, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 427, + 496 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 427, + 496 + ], + "type": "text", + "content": "2.2. Imposing constraints" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "text", + "content": "In this section, we explain how to incorporate constraints on a subset of the parameters " + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "text", + "content": " directly into the ERM framework. We consider constraints given in the form of a system of equations. To be very precise, let " + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "inline_equation", + "content": "F:\\mathbb{R}^{\\tilde{p}}\\to \\mathbb{R}^{q}" + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "text", + "content": " be a twice continuously differentiable map, where " + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "inline_equation", + "content": "\\tilde{p}" + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "text", + "content": " denotes the number of constrained parameters, and " + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "text", + "content": " is the number of equations. We assume that " + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "inline_equation", + "content": "q < \\tilde{p}\\leq p" + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "text", + "content": ". For notational convenience, we decompose " + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^p = \\mathbb{R}^{p - \\tilde{p}}\\oplus \\mathbb{R}^{\\tilde{p}}" + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "text", + "content": ", where the first and second subspaces correspond to the unconstrained and constrained parameters, respectively. We take " + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "inline_equation", + "content": "\\pi_{p - \\tilde{p}}:\\mathbb{R}^p\\to \\mathbb{R}^{p - \\tilde{p}}" + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "inline_equation", + "content": "\\pi_{\\tilde{p}}:\\mathbb{R}^p\\to \\mathbb{R}^{\\tilde{p}}" + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "text", + "content": " to be the projections onto the unconstrained and constrained parameter subspace, respectively. We will denote the unconstrained and constrained parameters by " + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "text", + "content": ", respectively, i.e., " + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "inline_equation", + "content": "\\alpha = \\pi_{p - \\tilde{p}}(\\xi)" + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "inline_equation", + "content": "\\theta = \\pi_{\\tilde{p}}(\\xi)" + }, + { + "bbox": [ + 304, + 502, + 547, + 671 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Definition 2.1 (CERM). Let " + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "G: \\mathbb{R}^n \\times \\mathbb{R}^p \\to \\mathbb{R}^m" + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": " be a continuously differentiable parameterization of mappings and " + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "F: \\mathbb{R}^{\\tilde{p}} \\to \\mathbb{R}^q" + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": " a twice continuously differentiable constraint," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24100" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "inline_equation", + "content": "q < \\tilde{p} \\leq p" + }, + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "text", + "content": ". Suppose " + }, + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "inline_equation", + "content": "L: \\mathbb{R}^m \\times \\mathbb{R}^m \\to [0, \\infty)" + }, + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "text", + "content": " is a continuously differentiable loss function. The constrained ERM problem for " + }, + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "inline_equation", + "content": "(X, Y)" + }, + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "inline_equation", + "content": "(G, F, L)" + }, + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "text", + "content": " is defined by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 125, + 287, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 125, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 125, + 287, + 163 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\min _ {\\xi \\in \\mathbb {R} ^ {p}} \\mathbb {E} (L (G (X, \\xi), Y)), \\\\ \\text {s . t .} F (\\pi_ {\\tilde {p}} (\\xi)) = 0. \\end{array} \\right. \\tag {2}", + "image_path": "bab745caf3aa34bc1658ba7827ec0e5916c28a2136fbc5828520f10c8834349c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 170, + 289, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 289, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 289, + 302 + ], + "type": "text", + "content": "Note the generality of the admissible mappings " + }, + { + "bbox": [ + 46, + 170, + 289, + 302 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 170, + 289, + 302 + ], + "type": "text", + "content": " in Definition 2.1. Although we will focus on neural networks from now on, the proposed framework applies to any parametric model, e.g., logistic or polynomial regression models. Next, we show that the CERM problem in (2) can be reformulated as an ordinary ERM problem on a Riemannian manifold " + }, + { + "bbox": [ + 46, + 170, + 289, + 302 + ], + "type": "inline_equation", + "content": "(\\mathcal{N}, g_{\\mathcal{N}})" + }, + { + "bbox": [ + 46, + 170, + 289, + 302 + ], + "type": "text", + "content": ", provided that the system of equations satisfies a mild non-degeneracy condition. This result allows us to consider the admissible parameters as a geometric object in its own right, whose intrinsic geometry we use to solve (2). The proof of this theorem can be found in Sec. 6.1 [16]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "text", + "content": "Theorem 2.2. If zero is a regular (i.e., non-singular) value of " + }, + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "text", + "content": ", then the CERM problem in (2) is equivalent to solving an ordinary ERM problem on a Riemannian manifold " + }, + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "text", + "content": " with metric " + }, + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "inline_equation", + "content": "g_{\\mathcal{N}}" + }, + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "text", + "content": " of dimension " + }, + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "inline_equation", + "content": "p - q" + }, + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{N} = \\mathbb{R}^{p - \\tilde{p}} \\times \\mathcal{M}" + }, + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "text", + "content": " is an embedded " + }, + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "inline_equation", + "content": "C^2" + }, + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "text", + "content": "-submanifold of " + }, + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^p" + }, + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{M} := F^{-1}(0)" + }, + { + "bbox": [ + 46, + 308, + 289, + 381 + ], + "type": "text", + "content": ". The equivalent minimization problem is given by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 388, + 287, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 388, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 91, + 388, + 287, + 409 + ], + "type": "interline_equation", + "content": "\\min _ {(\\alpha , \\theta) \\in \\mathcal {N}} \\mathbb {E} (L (G (X, \\alpha \\oplus \\iota (\\theta)), Y)), \\tag {3}", + "image_path": "dc056ce5f6c8eb00976ece03ded72cd87ed8ddc7472d8475447b884aca7f4629.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 416, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 416, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\iota : \\mathcal{M} \\to \\mathbb{R}^{\\tilde{P}}" + }, + { + "bbox": [ + 46, + 416, + 287, + 441 + ], + "type": "text", + "content": " is the inclusion map which embeds " + }, + { + "bbox": [ + 46, + 416, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 416, + 287, + 441 + ], + "type": "text", + "content": " in the ambient space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 448, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 471 + ], + "type": "text", + "content": "From now on, we refer to the objective in (3) as simply the loss " + }, + { + "bbox": [ + 46, + 448, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 448, + 287, + 471 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "text", + "content": "Graph coordinates on " + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "text", + "content": " The main idea is to directly minimize the loss on " + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "text", + "content": " using SGD, see Figure 1 for an overview. To perform SGD on this space, however, we need to choose a coordinate system in which we can evaluate the gradient of the loss and subsequently follow descent trajectories on " + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "text", + "content": ". Now, the proof of Theorem 2.2 involves showing that " + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "text", + "content": " is an embedded submanifold of the ambient space " + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{\\tilde{p}}" + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "text", + "content": ". This is accomplished by constructing a special coordinate system, a so-called graph chart or graph coordinate system. This chart is uniquely determined by the constraint " + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "text", + "content": " and a prescribed zero " + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathcal{M}" + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "text", + "content": ". It can be numerically evaluated at the point " + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "text", + "content": " by solving a linear system of equations. In turn, this enables us to construct a chart on the manifold in which we can evaluate the Riemannian metric and the gradient " + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "inline_equation", + "content": "\\nabla_{g_{\\mathcal{N}}} \\mathcal{L}(\\alpha, \\theta)" + }, + { + "bbox": [ + 46, + 486, + 289, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 666, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 289, + 714 + ], + "type": "text", + "content": "The crucial observation is that we can numerically construct coordinate systems in which we can follow paths on the manifold " + }, + { + "bbox": [ + 46, + 666, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 666, + 289, + 714 + ], + "type": "text", + "content": " in the direction of " + }, + { + "bbox": [ + 46, + 666, + 289, + 714 + ], + "type": "inline_equation", + "content": "-\\nabla_{g_{\\mathcal{N}}}\\mathcal{L}(\\alpha, \\theta)" + }, + { + "bbox": [ + 46, + 666, + 289, + 714 + ], + "type": "text", + "content": ". In particular, the system of equations determining these" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 547, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 547, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 547, + 107 + ], + "type": "text", + "content": "Algorithm 1 Sketch of the computation of " + }, + { + "bbox": [ + 305, + 72, + 547, + 107 + ], + "type": "inline_equation", + "content": "\\nabla_{g_N}\\mathcal{L}(\\alpha ,\\theta)" + }, + { + "bbox": [ + 305, + 72, + 547, + 107 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 305, + 72, + 547, + 107 + ], + "type": "inline_equation", + "content": "(\\alpha ,\\theta)\\in \\mathcal{N}" + }, + { + "bbox": [ + 305, + 72, + 547, + 107 + ], + "type": "text", + "content": ". See Algorithm 2 in Sec. 6.5 [16] the detailed formal version." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 311, + 110, + 547, + 231 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 311, + 110, + 530, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 110, + 530, + 122 + ], + "spans": [ + { + "bbox": [ + 311, + 110, + 530, + 122 + ], + "type": "text", + "content": "1: Compute the derivative " + }, + { + "bbox": [ + 311, + 110, + 530, + 122 + ], + "type": "inline_equation", + "content": "DF(\\theta)" + }, + { + "bbox": [ + 311, + 110, + 530, + 122 + ], + "type": "text", + "content": " of the constraint " + }, + { + "bbox": [ + 311, + 110, + 530, + 122 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 311, + 110, + 530, + 122 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 123, + 463, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 123, + 463, + 133 + ], + "spans": [ + { + "bbox": [ + 312, + 123, + 463, + 133 + ], + "type": "text", + "content": "2: Compute chart-related derivatives." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 134, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 134, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 312, + 134, + 547, + 156 + ], + "type": "text", + "content": "3: Compute the metric " + }, + { + "bbox": [ + 312, + 134, + 547, + 156 + ], + "type": "inline_equation", + "content": "g_{\\mathcal{N}}" + }, + { + "bbox": [ + 312, + 134, + 547, + 156 + ], + "type": "text", + "content": " using the chart-related derivatives." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 158, + 545, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 158, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 312, + 158, + 545, + 182 + ], + "type": "text", + "content": "4: Compute the partial derivatives of the loss " + }, + { + "bbox": [ + 312, + 158, + 545, + 182 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 312, + 158, + 545, + 182 + ], + "type": "text", + "content": " with respect to the constrained components " + }, + { + "bbox": [ + 312, + 158, + 545, + 182 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 312, + 158, + 545, + 182 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 182, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 182, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 182, + 545, + 205 + ], + "type": "text", + "content": "5: Compute the components of " + }, + { + "bbox": [ + 312, + 182, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\nabla_{g_{\\mathcal{M}}}\\mathcal{L}(\\alpha ,\\theta)" + }, + { + "bbox": [ + 312, + 182, + 545, + 205 + ], + "type": "text", + "content": " using the above ingredients." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 206, + 545, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 206, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 312, + 206, + 545, + 231 + ], + "type": "text", + "content": "6: Compute the ordinary (unconstrained) components " + }, + { + "bbox": [ + 312, + 206, + 545, + 231 + ], + "type": "inline_equation", + "content": "\\nabla_{g_{\\mathrm{flat}}}\\mathcal{L}(\\alpha ,\\theta)" + }, + { + "bbox": [ + 312, + 206, + 545, + 231 + ], + "type": "text", + "content": " by evaluating " + }, + { + "bbox": [ + 312, + 206, + 545, + 231 + ], + "type": "inline_equation", + "content": "D_{\\alpha}\\mathcal{L}(\\alpha ,\\theta)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 256, + 547, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 547, + 399 + ], + "type": "text", + "content": "so-called graph coordinate systems is fully determined by the derivative " + }, + { + "bbox": [ + 304, + 256, + 547, + 399 + ], + "type": "inline_equation", + "content": "DF" + }, + { + "bbox": [ + 304, + 256, + 547, + 399 + ], + "type": "text", + "content": " of the constraint " + }, + { + "bbox": [ + 304, + 256, + 547, + 399 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 256, + 547, + 399 + ], + "type": "text", + "content": " and the constraint itself. This observation enables a general implementation of graph coordinate systems independent of the explicit form of " + }, + { + "bbox": [ + 304, + 256, + 547, + 399 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 256, + 547, + 399 + ], + "type": "text", + "content": ". We only require an implementation of " + }, + { + "bbox": [ + 304, + 256, + 547, + 399 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 256, + 547, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 256, + 547, + 399 + ], + "type": "inline_equation", + "content": "DF" + }, + { + "bbox": [ + 304, + 256, + 547, + 399 + ], + "type": "text", + "content": ", where the latter may be obtained using automatic differentiation. We provide a sketch of the algorithm for computing the gradient in Algorithm 1. For the interested reader, we mention that a complete and detailed explanation of all steps can be found in Sec. 6 [16]. Before moving on, we mention the limitations of the CERM framework for other generic constraints." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 429, + 547, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 547, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 547, + 597 + ], + "type": "text", + "content": "Limitations As with most applications of deep neural networks, the available amount of computational power imposes limitations. Specifically, if the number of constrained parameters is large relative to the number of constraints, the computational demands will rise. Furthermore, setting up well-posed constraints can be a non-trivial task, e.g., symmetries in the constraints may result in redundancies obstructing the solvability of the associated systems of equations. Finally, the topology of the constrained submanifold may pose limitations as well. For instance, while the constraints will generically define a Riemannian manifold, it need not be connected. In such cases, the initialization of the network parameters will play an important role, as the gradient flow will stay on the connected component it was initialized on." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 609, + 509, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 609, + 509, + 623 + ], + "spans": [ + { + "bbox": [ + 305, + 609, + 509, + 623 + ], + "type": "text", + "content": "3. Multiresolution Analysis and CERM" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "content": "This section presents the application of the aforementioned CERM framework to a non-trivial application to learn optimal wavelet bases for a given task. Many different types of wavelets, such as Haar, Daubechies, Gabor, etc., are used for a wide array of tasks. The specifics of the task demand different characteristics of the wavelets, such as the degree of smoothness, or the support of the wavelet. For example," + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24101" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "the CDF wavelets [10] are well-known due to their usage for both lossless and lossy compression in JPEG2000. In this section, we consider networks that will find task-driven wavelets, optimizing the wavelet characteristics to the task at hand." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 240 + ], + "type": "text", + "content": "Here, we will explain how to set up a system of equations (constraints) whose solution set corresponds to wavelets. We will apply these constraints in the next section to learn task-driven wavelets for predicting multi-resolution decompositions of contours in the medical domain. While we only consider one-dimensional wavelets in this work, we stress that the framework is easily adapted to higher-dimensional domains, to decompose images, for instance, by using tensor products of one-dimensional bases." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 248, + 185, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 185, + 261 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 185, + 261 + ], + "type": "text", + "content": "3.1. Multiresolution Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "text", + "content": "In this section, we briefly state what a Multiresolution Analysis (MRA) is. We closely follow the standard exposition, found in, e.g., [29, 31], and refer the reader to Sec. 7 [16] for a more in-depth and formal introduction. The MRA framework allows one to study a signal " + }, + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "inline_equation", + "content": "\\gamma \\in L^2 (\\mathbb{R})" + }, + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "text", + "content": " at various resolution levels. The origin of this framework lies in the uncertainty principle of Fourier analysis, which states that a signal cannot be simultaneously localized in the time and frequency domain. MRA addresses this shortcoming by decomposing a signal on different discrete resolution levels. The idea is to construct a nested (exhausting) sequence of closed subspaces " + }, + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "inline_equation", + "content": "V_{j}\\subset L^{2}(\\mathbb{R})" + }, + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "text", + "content": ", each associated with a specific resolution level " + }, + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "text", + "content": ". The subspaces " + }, + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "inline_equation", + "content": "V_{j}" + }, + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "text", + "content": " are spanned by integer shifts of an appropriate dilation of a so-called scaling map (or father wavelet) " + }, + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "text", + "content": ". The level of dilation determines the resolution level. Formally, we require that " + }, + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "inline_equation", + "content": "(\\varphi_{jk})_{k\\in \\mathbb{Z}}" + }, + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "text", + "content": " is an orthonormal basis for " + }, + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "inline_equation", + "content": "V_{j}" + }, + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "inline_equation", + "content": "\\varphi_{jk}(t)\\coloneqq 2^{\\frac{j}{2}}\\varphi (2^{j}t - k)" + }, + { + "bbox": [ + 46, + 268, + 289, + 508 + ], + "type": "text", + "content": ". In Fig. 2a and Fig. 2c, we show an example of that Haar-MRA; in particular, we show the father wavelet and an example of a subspace at a certain resolution level." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": "Decomposing a signal To analyze a signal " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " at different resolution levels, we project it onto the subspaces " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "V_{j}" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": ". The orthogonal projection of " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " at level " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " is denoted by " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\gamma_{j}" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": ". The coefficients of " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\gamma_{j}" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " with respect to the basis for " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "V_{j}" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "a_{jk}(\\gamma)" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": ", are called the approximation coefficients of " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": ". When a signal in " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "V_{j+1}" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " is projected onto " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "V_{j}" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": ", the lost information is stored in the orthogonal complement " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "W_{j}" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "V_{j}" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "V_{j+1}" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": ". This subspace is referred to as the detail subspace at level " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": ". The decomposition of " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "V_{j+1}" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "V_{j}" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "W_{j}" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " allows the reconstruction of a signal at level " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "j+1" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " from any lower level " + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "inline_equation", + "content": "j_{0}" + }, + { + "bbox": [ + 46, + 510, + 288, + 700 + ], + "type": "text", + "content": " provided all the details in between are known. The corresponding decomposition and reconstruction algorithms give rise to the Discrete Wavelet Transform (DWT) and Mallat's Pyramid Algorithm. In Fig. 2b and Fig. 2d, we show examples of the mother wavelet and the detail subspace for the Haar wavelet." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "type": "text", + "content": "A fundamental result, known as Mallat's Theorem, states" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 313, + 70, + 421, + 156 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 421, + 156 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 421, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 421, + 156 + ], + "type": "image", + "image_path": "6b37e34bbbf1201ad4b129c2a50d686a0acb69edf924732db44394a509ffcf32.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 430, + 70, + 540, + 156 + ], + "blocks": [ + { + "bbox": [ + 430, + 70, + 540, + 156 + ], + "lines": [ + { + "bbox": [ + 430, + 70, + 540, + 156 + ], + "spans": [ + { + "bbox": [ + 430, + 70, + 540, + 156 + ], + "type": "image", + "image_path": "85f0688522d302ac7328cc2eb5449b364d97056a6b0804ae3e0775972e8fdb3e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 167, + 421, + 252 + ], + "blocks": [ + { + "bbox": [ + 336, + 157, + 395, + 167 + ], + "lines": [ + { + "bbox": [ + 336, + 157, + 395, + 167 + ], + "spans": [ + { + "bbox": [ + 336, + 157, + 395, + 167 + ], + "type": "text", + "content": "(a) Father wavelet" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 167, + 421, + 252 + ], + "lines": [ + { + "bbox": [ + 313, + 167, + 421, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 421, + 252 + ], + "type": "image", + "image_path": "2e28b94e07a13d486a1fbd061d54edd049b941fd65951b183825a66efe46e535.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 254, + 375, + 265 + ], + "lines": [ + { + "bbox": [ + 355, + 254, + 375, + 265 + ], + "spans": [ + { + "bbox": [ + 355, + 254, + 375, + 265 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 355, + 254, + 375, + 265 + ], + "type": "inline_equation", + "content": "V_{j}" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "text", + "content": "Figure 2. Example of the Haar MRA: (a) Dilated translation of the Haar scaling map (or father wavelet) " + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\varphi = \\mathbf{1}_{[0,1)}" + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "text", + "content": ". The approximation subspace " + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "inline_equation", + "content": "V_{j}" + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "text", + "content": " at level " + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "text", + "content": " consists of all step-functions with step-size " + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "inline_equation", + "content": "2^{-j}" + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "text", + "content": ". (b) Dilated translation of the mother wavelet " + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\psi = \\mathbf{1}_{[0,\\frac{1}{2})} - \\mathbf{1}_{[\\frac{1}{2},1)}" + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "text", + "content": ". The detail subspace " + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "inline_equation", + "content": "W_{j}" + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "text", + "content": " is spanned by integer translations of the dilated mother wavelet " + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\psi_{j}" + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "text", + "content": ". (c) The approximation subspace at level " + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "text", + "content": " consists of all step-functions with step-size " + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "inline_equation", + "content": "2^{-j}" + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "text", + "content": ". (d) Example of a function in the detail subspace at level " + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 273, + 545, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 432, + 167, + 538, + 252 + ], + "blocks": [ + { + "bbox": [ + 453, + 157, + 515, + 167 + ], + "lines": [ + { + "bbox": [ + 453, + 157, + 515, + 167 + ], + "spans": [ + { + "bbox": [ + 453, + 157, + 515, + 167 + ], + "type": "text", + "content": "(b) Mother wavelet" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 167, + 538, + 252 + ], + "lines": [ + { + "bbox": [ + 432, + 167, + 538, + 252 + ], + "spans": [ + { + "bbox": [ + 432, + 167, + 538, + 252 + ], + "type": "image", + "image_path": "7ec91ceaa1e072de490ec3a80c4c778e89db15023c12ae3664c503b375020418.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 254, + 495, + 265 + ], + "lines": [ + { + "bbox": [ + 471, + 254, + 495, + 265 + ], + "spans": [ + { + "bbox": [ + 471, + 254, + 495, + 265 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 471, + 254, + 495, + 265 + ], + "type": "inline_equation", + "content": "W_{j}" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "text", + "content": "that the subspaces " + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "inline_equation", + "content": "W_{j}" + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "text", + "content": " can also be spanned by dilating and shifting a single map. More precisely, there exists a map " + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\psi \\in W_0" + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "text", + "content": ", the so-called mother wavelet, such that the " + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\psi_{jk}" + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "text", + "content": " form an orthonormal basis for " + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "inline_equation", + "content": "W_{j}" + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "text", + "content": ", see [31]. The coefficients of the projection of " + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "inline_equation", + "content": "(\\gamma)" + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "text", + "content": " onto " + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "inline_equation", + "content": "W_{j}" + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "text", + "content": ", are denoted by " + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "inline_equation", + "content": "d_{j}(\\gamma) \\coloneqq (d_{jk}(\\gamma))_{k\\in \\mathbb{Z}}" + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "text", + "content": ", and referred to as the detail coefficients of " + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "text", + "content": " at resolution level " + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 394, + 545, + 491 + ], + "type": "text", + "content": ". We can summarize the MRA procedure roughly as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 337, + 498, + 545, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 498, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 337, + 498, + 545, + 525 + ], + "type": "interline_equation", + "content": "\\gamma_ {j} = \\sum_ {k} a _ {j _ {0} k} \\varphi_ {j _ {0} k} + \\sum_ {j _ {0} \\leq l \\leq j - 1} \\sum_ {k} d _ {l k} \\psi_ {l k}, \\tag {4}", + "image_path": "173b029b6fd8ded91c75a8a7f2fd765c62aa44ca80f6e90b16e4be8f2fe97059.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 531, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 547, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 531, + 547, + 567 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 304, + 531, + 547, + 567 + ], + "type": "text", + "content": " denotes the father wavelet and " + }, + { + "bbox": [ + 304, + 531, + 547, + 567 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 531, + 547, + 567 + ], + "type": "text", + "content": " the mother wavelet. The MRA boils down to the specification of the first approximation coefficient and a number of detail coefficients." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": "Low and high pass filters As we mentioned before, an MRA is completely determined by a scaling map " + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": ", which in turn is completely characterized by a so-called low-pass filter. We will use this observation to set up a system of constraints characterizing a finite-dimensional family of wavelets, such that we can use it in the CERM framework. To explain how this works, we first make a key observation. Since " + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "inline_equation", + "content": "V_{0} \\subset V_{1}" + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": ", there exists a unique sequence " + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": " characterizing " + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": " in terms of the basis " + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "inline_equation", + "content": "(\\varphi_{1k})_{k \\in \\mathbb{Z}}" + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": ". The sequence " + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": " completely characterizes the scaling function and is called the low-pass filter or wavelet filter of the MRA. Similarly," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24102" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "since " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\psi \\in W_0\\subset V_1" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ", there exists a unique sequence " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ", the so-called high-pass filter associated to " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ", characterizing the mother wavelet. An example of an MRA is shown in Figure 2. In practice, to define an MRA, one only needs to specify an appropriate low-pass filter " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ", as for Mallat's mother wavelet, we have " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "g_{k} = (-1)^{k - 1}h_{1 - k}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ". We refer the interested reader to the supplementary information for a detailed discussion of all these points [16]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 177, + 151, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 177, + 151, + 189 + ], + "spans": [ + { + "bbox": [ + 47, + 177, + 151, + 189 + ], + "type": "text", + "content": "3.2. MRA constraints" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 197, + 290, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 197, + 290, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 290, + 280 + ], + "type": "text", + "content": "We are now ready to formulate and impose constraints on a sequence " + }, + { + "bbox": [ + 46, + 197, + 290, + 280 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 197, + 290, + 280 + ], + "type": "text", + "content": " to ensure that it is the low-pass filter of a scaling map " + }, + { + "bbox": [ + 46, + 197, + 290, + 280 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 46, + 197, + 290, + 280 + ], + "type": "text", + "content": ". In practice, we only consider finite sequences. For notational convenience, we introduce the space " + }, + { + "bbox": [ + 46, + 197, + 290, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_M(\\mathbb{R})" + }, + { + "bbox": [ + 46, + 197, + 290, + 280 + ], + "type": "text", + "content": ", which consists of one-dimensional real-valued two-sided sequences " + }, + { + "bbox": [ + 46, + 197, + 290, + 280 + ], + "type": "inline_equation", + "content": "(h_k)_{k=1-M}^{M-1}" + }, + { + "bbox": [ + 46, + 197, + 290, + 280 + ], + "type": "text", + "content": " of order " + }, + { + "bbox": [ + 46, + 197, + 290, + 280 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 197, + 290, + 280 + ], + "type": "text", + "content": "; so " + }, + { + "bbox": [ + 46, + 197, + 290, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_M(\\mathbb{R})" + }, + { + "bbox": [ + 46, + 197, + 290, + 280 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 197, + 290, + 280 + ], + "type": "inline_equation", + "content": "2M - 1" + }, + { + "bbox": [ + 46, + 197, + 290, + 280 + ], + "type": "text", + "content": " dimensional." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 290, + 289, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 289, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 289, + 316 + ], + "type": "text", + "content": "Theorem 3.1. Let " + }, + { + "bbox": [ + 46, + 290, + 289, + 316 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{N}_{\\geq 3}" + }, + { + "bbox": [ + 46, + 290, + 289, + 316 + ], + "type": "text", + "content": " be a prescribed order and define " + }, + { + "bbox": [ + 46, + 290, + 289, + 316 + ], + "type": "inline_equation", + "content": "F_M: \\mathcal{A}_M(\\mathbb{R}) \\to \\mathbb{R}^{M+1}" + }, + { + "bbox": [ + 46, + 290, + 289, + 316 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 325, + 298, + 408 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 325, + 298, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 325, + 298, + 408 + ], + "type": "interline_equation", + "content": "(F_{M}(h))_{k}:= \\left\\{ \\begin{array}{ll}\\sum_{|l|\\leq M - 1}|h_{l}|^{2} - 1, & k = 0,\\\\ \\sum_{1 - M + 2k\\leq l\\leq M - 1}h_{l - 2k}h_{l}, & 1\\leq k\\leq M - 1,\\\\ -\\sqrt{2} +\\sum_{|l|\\leq M - 1}h_{l}, & k = M. \\end{array} \\right.", + "image_path": "fb2d32836cbf462228989c59f3f3e627607d20bf4f18387c2cd87ab1af6825aa.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 418, + 288, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 288, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 288, + 455 + ], + "type": "text", + "content": "If " + }, + { + "bbox": [ + 46, + 418, + 288, + 455 + ], + "type": "inline_equation", + "content": "F_{M}(h) = 0" + }, + { + "bbox": [ + 46, + 418, + 288, + 455 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 418, + 288, + 455 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 418, + 288, + 455 + ], + "type": "text", + "content": " satisfies a certain non-degeneracy condition, then " + }, + { + "bbox": [ + 46, + 418, + 288, + 455 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 418, + 288, + 455 + ], + "type": "text", + "content": " is the low-pass filter of a scaling map " + }, + { + "bbox": [ + 46, + 418, + 288, + 455 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 46, + 418, + 288, + 455 + ], + "type": "text", + "content": " generating an MRA of " + }, + { + "bbox": [ + 46, + 418, + 288, + 455 + ], + "type": "inline_equation", + "content": "L^2 (\\mathbb{R})" + }, + { + "bbox": [ + 46, + 418, + 288, + 455 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 464, + 288, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 288, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 288, + 608 + ], + "type": "text", + "content": "A full proof and detailed statement of the non-degeneracy condition can be found in Sec. 7.4 in the supplementary material. Heuristically, the first " + }, + { + "bbox": [ + 46, + 464, + 288, + 608 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 464, + 288, + 608 + ], + "type": "text", + "content": " equations encode the orthonormality of the scaling map. The last equation encodes the observation that the Fourier transform of the scaling map must be nonzero. The set of regular points in " + }, + { + "bbox": [ + 46, + 464, + 288, + 608 + ], + "type": "inline_equation", + "content": "F_{M}^{-1}(0)" + }, + { + "bbox": [ + 46, + 464, + 288, + 608 + ], + "type": "text", + "content": " is a real-analytic " + }, + { + "bbox": [ + 46, + 464, + 288, + 608 + ], + "type": "inline_equation", + "content": "(M - 2)" + }, + { + "bbox": [ + 46, + 464, + 288, + 608 + ], + "type": "text", + "content": "-dimensional submanifold of " + }, + { + "bbox": [ + 46, + 464, + 288, + 608 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{2M - 1}" + }, + { + "bbox": [ + 46, + 464, + 288, + 608 + ], + "type": "text", + "content": ". Hence, we can get as many degrees of freedom as desired, by choosing a sufficiently large order " + }, + { + "bbox": [ + 46, + 464, + 288, + 608 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 464, + 288, + 608 + ], + "type": "text", + "content": ". We now have an end-to-end way of learning wavelets for specific tasks, a wavelet network, by using the constraint " + }, + { + "bbox": [ + 46, + 464, + 288, + 608 + ], + "type": "inline_equation", + "content": "F_{M}" + }, + { + "bbox": [ + 46, + 464, + 288, + 608 + ], + "type": "text", + "content": " along with the CERM framework." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 620, + 128, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 128, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 128, + 634 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "In this section, we show some explicit examples of the wavelet networks with respect to contour prediction. In particular, we will utilize the derived wavelet filters of Sec. 3.2 to create a network for autocontouring. This task fits well with the setup of wavelets, as they naturally lend themselves to representing continuous objects such as curves." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 321, + 72, + 533, + 223 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 533, + 223 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 533, + 223 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 533, + 223 + ], + "type": "image", + "image_path": "77d0f409aa4320a120744f7d19520fee3386ba36cafbdbc793c0827a4e741b9b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 236, + 547, + 413 + ], + "lines": [ + { + "bbox": [ + 304, + 236, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 547, + 413 + ], + "type": "text", + "content": "Figure 3. A schematic picture of our network. The encoder consists of residual convolutional blocks depicted in blue. The first residual convolutional block uses " + }, + { + "bbox": [ + 304, + 236, + 547, + 413 + ], + "type": "inline_equation", + "content": "n_f" + }, + { + "bbox": [ + 304, + 236, + 547, + 413 + ], + "type": "text", + "content": " filters and is doubled after every other residual convolutional block. Attached to the encoder are fully connected layers to predict approximation and detail coefficients at the lowest resolution level " + }, + { + "bbox": [ + 304, + 236, + 547, + 413 + ], + "type": "inline_equation", + "content": "j_0" + }, + { + "bbox": [ + 304, + 236, + 547, + 413 + ], + "type": "text", + "content": ". The approximation and detail coefficients are supplied as input to the Pyramid Algorithm (the decoder) to predict a contour on high-resolution level. Each green block corresponds to an inverse discrete wavelet transform. Detail coefficients at higher levels are computed using skip-connections (arrows in red). We only predict detail coefficients up to level " + }, + { + "bbox": [ + 304, + 236, + 547, + 413 + ], + "type": "inline_equation", + "content": "j_1" + }, + { + "bbox": [ + 304, + 236, + 547, + 413 + ], + "type": "text", + "content": ". In this example, we have set " + }, + { + "bbox": [ + 304, + 236, + 547, + 413 + ], + "type": "inline_equation", + "content": "j_1 = j_0 + 2" + }, + { + "bbox": [ + 304, + 236, + 547, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 236, + 547, + 413 + ], + "type": "inline_equation", + "content": "j_2 = j_0 + 3" + }, + { + "bbox": [ + 304, + 236, + 547, + 413 + ], + "type": "text", + "content": ". In reality, the decoder consists of two upsampling paths, one for each spatial component of the curve. We have only drawn one for notational convenience. During training, only the approximation coefficients at the highest resolution level are supervised." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 433, + 476, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 433, + 476, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 433, + 476, + 445 + ], + "type": "text", + "content": "4.1. Contour Prediction using MRA" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "text", + "content": "In the context of contour prediction, wavelets can be used to represent the boundary of a region in an image using a simple closed curve. We consider two-dimensional gray-valued images " + }, + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{X} \\coloneqq [0,1]^{n \\times n}" + }, + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "text", + "content": ". For each of our applications, we assume that these images contain an organ with a boundary. Formally speaking, this means that each image " + }, + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "text", + "content": " contains a (uniquely identifiable) simply connected region " + }, + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "inline_equation", + "content": "R(x)" + }, + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "text", + "content": ", the organ, with boundary " + }, + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\partial R(x)" + }, + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "text", + "content": ". This boundary of the region, " + }, + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\partial R(x)" + }, + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "text", + "content": ", can be parameterized by a simple closed, piecewise " + }, + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "inline_equation", + "content": "C^2" + }, + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "text", + "content": ", curve " + }, + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\gamma(x)" + }, + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "text", + "content": ". We will develop a deep learning framework for computing such parameterizations by learning a multiresolution decomposition of " + }, + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\gamma(x)" + }, + { + "bbox": [ + 304, + 452, + 547, + 608 + ], + "type": "text", + "content": " using the methods developed in the previous sections." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 623, + 360, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 360, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 360, + 633 + ], + "type": "text", + "content": "4.1.1 Data" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 548, + 713 + ], + "type": "text", + "content": "We have used public datasets from the Medical Decathlon Challenge [3]. The selected data consists of CT scans of the spleen of size " + }, + { + "bbox": [ + 304, + 641, + 548, + 713 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 641, + 548, + 713 + ], + "type": "text", + "content": " and T2-weighted MRI images of the prostate central gland, henceforth abbreviated as just the prostate, of size " + }, + { + "bbox": [ + 304, + 641, + 548, + 713 + ], + "type": "inline_equation", + "content": "320 \\times 320" + }, + { + "bbox": [ + 304, + 641, + 548, + 713 + ], + "type": "text", + "content": ". The scans were cropped to size " + }, + { + "bbox": [ + 304, + 641, + 548, + 713 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 641, + 548, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 641, + 548, + 713 + ], + "type": "inline_equation", + "content": "192 \\times 192" + }, + { + "bbox": [ + 304, + 641, + 548, + 713 + ], + "type": "text", + "content": ", respectively. Further-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24103" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 70, + 458, + 198 + ], + "blocks": [ + { + "bbox": [ + 137, + 70, + 458, + 198 + ], + "lines": [ + { + "bbox": [ + 137, + 70, + 458, + 198 + ], + "spans": [ + { + "bbox": [ + 137, + 70, + 458, + 198 + ], + "type": "image", + "image_path": "4414cf31382f8d2718b2edefad30e87caf76fe60887c0a05341613a2529f9256.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 210, + 547, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 210, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 547, + 243 + ], + "type": "text", + "content": "Figure 4. Examples of predictions on the test set for the spleen and prostate, depicted in the first and second row, respectively, for the best-performing wavelet models. The green curve corresponds to the ground truth, while the red curve is a prediction made by the wavelet network." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 262, + 276, + 355 + ], + "blocks": [ + { + "bbox": [ + 61, + 262, + 276, + 355 + ], + "lines": [ + { + "bbox": [ + 61, + 262, + 276, + 355 + ], + "spans": [ + { + "bbox": [ + 61, + 262, + 276, + 355 + ], + "type": "image", + "image_path": "6c3cf0900468a100ab5a98f5a8092c9a42251f905e7a2f5e04ec9a840d0ec2f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 358, + 225, + 367 + ], + "lines": [ + { + "bbox": [ + 108, + 358, + 225, + 367 + ], + "spans": [ + { + "bbox": [ + 108, + 358, + 225, + 367 + ], + "type": "text", + "content": "(a) Initial father and mother wavelet" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 62, + 369, + 275, + 460 + ], + "blocks": [ + { + "bbox": [ + 62, + 369, + 275, + 460 + ], + "lines": [ + { + "bbox": [ + 62, + 369, + 275, + 460 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 275, + 460 + ], + "type": "image", + "image_path": "6ce08152be65b95f01d7a0a631d6bfc4fec432d3559d5226bdbba0d16e8b5cf8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 463, + 240, + 472 + ], + "lines": [ + { + "bbox": [ + 92, + 463, + 240, + 472 + ], + "spans": [ + { + "bbox": [ + 92, + 463, + 240, + 472 + ], + "type": "text", + "content": "(b) Task-optimized father and mother wavelet" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 482, + 287, + 550 + ], + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 550 + ], + "type": "text", + "content": "Figure 5. Example of wavelets of order 5 learned during training of the spleen model: (a) The initial father (left) and mother (right) wavelets for the second spatial component. (b) The task-optimized father and mother wavelets after training for the second spatial component. We observe that the task-optimized wavelets are more simple and have \"cleaned up\" in a sense." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "content": "more, the images were resampled to the median sample spacing, which resulted in " + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "inline_equation", + "content": "(5.00\\mathrm{mm}, 0.793\\mathrm{mm}, 0.793\\mathrm{mm})" + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "inline_equation", + "content": "(3.6\\mathrm{mm}, 0.625\\mathrm{mm}, 0.625\\mathrm{mm})" + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "content": " spacings for the spleen and prostate, respectively. We refer the reader to Sec. 8.2 [16] for a detailed description of all the preprocessing steps and construction of ground truth curves. Examples of predictions on the test set, for both the spleen and prostate, are shown in Figure Fig. 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 265, + 425, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 265, + 425, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 265, + 425, + 277 + ], + "type": "text", + "content": "4.1.2 Model and Training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 283, + 547, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 283, + 547, + 367 + ], + "spans": [ + { + "bbox": [ + 305, + 283, + 547, + 367 + ], + "type": "text", + "content": "Model objective We parameterize the boundary " + }, + { + "bbox": [ + 305, + 283, + 547, + 367 + ], + "type": "inline_equation", + "content": "\\partial R(x)" + }, + { + "bbox": [ + 305, + 283, + 547, + 367 + ], + "type": "text", + "content": " by arc length, resulting in a curve " + }, + { + "bbox": [ + 305, + 283, + 547, + 367 + ], + "type": "inline_equation", + "content": "\\gamma (x)" + }, + { + "bbox": [ + 305, + 283, + 547, + 367 + ], + "type": "text", + "content": ". The precise details of this parametrization are outlined in the supplementary information. The objective is to compute the relevant approximation coefficients of " + }, + { + "bbox": [ + 305, + 283, + 547, + 367 + ], + "type": "inline_equation", + "content": "\\gamma (x)" + }, + { + "bbox": [ + 305, + 283, + 547, + 367 + ], + "type": "text", + "content": ". To this end, let " + }, + { + "bbox": [ + 305, + 283, + 547, + 367 + ], + "type": "inline_equation", + "content": "j_0,j_1,j_2\\in \\mathbb{N}" + }, + { + "bbox": [ + 305, + 283, + 547, + 367 + ], + "type": "text", + "content": " be resolution levels, where " + }, + { + "bbox": [ + 305, + 283, + 547, + 367 + ], + "type": "inline_equation", + "content": "j_0\\leq j_1\\leq j_2" + }, + { + "bbox": [ + 305, + 283, + 547, + 367 + ], + "type": "text", + "content": ". We will construct a convolutional neural network" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 322, + 373, + 529, + 408 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 373, + 529, + 408 + ], + "spans": [ + { + "bbox": [ + 322, + 373, + 529, + 408 + ], + "type": "interline_equation", + "content": "G: \\mathcal {X} \\times \\mathbb {R} ^ {p} \\to \\prod_ {j = j _ {0}} ^ {j _ {2}} \\mathbb {R} ^ {2 ^ {j}} \\times \\mathbb {R} ^ {2 ^ {j}} \\times \\prod_ {j = j _ {0}} ^ {j _ {1} - 1} \\mathbb {R} ^ {2 ^ {j}} \\times \\mathbb {R} ^ {2 ^ {j}},", + "image_path": "11c1a0277c1f170b64b2b5c56f0fdb96fc7258942624e0b2a248187ddd5a97bc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 415, + 547, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 547, + 501 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 547, + 501 + ], + "type": "text", + "content": "which predicts the wavelet decomposition of " + }, + { + "bbox": [ + 305, + 415, + 547, + 501 + ], + "type": "inline_equation", + "content": "\\gamma (x)" + }, + { + "bbox": [ + 305, + 415, + 547, + 501 + ], + "type": "text", + "content": ". Here the subspaces " + }, + { + "bbox": [ + 305, + 415, + 547, + 501 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{2^j}" + }, + { + "bbox": [ + 305, + 415, + 547, + 501 + ], + "type": "text", + "content": " correspond to approximation and detail coefficients at level " + }, + { + "bbox": [ + 305, + 415, + 547, + 501 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 415, + 547, + 501 + ], + "type": "text", + "content": ", one for each spatial component. Furthermore, a subset of the parameters " + }, + { + "bbox": [ + 305, + 415, + 547, + 501 + ], + "type": "inline_equation", + "content": "\\xi \\in \\mathbb{R}^p" + }, + { + "bbox": [ + 305, + 415, + 547, + 501 + ], + "type": "text", + "content": " are constrained to be wavelet filters, one wavelet filter per spatial component, using Theorem 3.1. The map " + }, + { + "bbox": [ + 305, + 415, + 547, + 501 + ], + "type": "inline_equation", + "content": "G(\\cdot ,\\xi)" + }, + { + "bbox": [ + 305, + 415, + 547, + 501 + ], + "type": "text", + "content": " applied to an image " + }, + { + "bbox": [ + 305, + 415, + 547, + 501 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 415, + 547, + 501 + ], + "type": "text", + "content": " has output" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 344, + 508, + 545, + 527 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 508, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 344, + 508, + 545, + 527 + ], + "type": "interline_equation", + "content": "G (x, \\xi) = \\left(a _ {j _ {0}} (x, \\xi), \\dots , a _ {j _ {2}} (x, \\xi), \\right. \\tag {5}", + "image_path": "4a1cec370fc94ecd7fad72cfc37daf87a3cf37a9c5b15b7fc26bd562eda1fa6b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 394, + 523, + 504, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 523, + 504, + 536 + ], + "spans": [ + { + "bbox": [ + 394, + 523, + 504, + 536 + ], + "type": "interline_equation", + "content": "d _ {j _ {0}} (x, \\xi), \\dots , d _ {j _ {1} - 1} (x, \\xi)).", + "image_path": "40bbf9923962d5f6d46957806775bf67324925f7e41e853fa78542239a855a2e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 544, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 547, + 581 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 306, + 544, + 547, + 581 + ], + "type": "inline_equation", + "content": "a_{j}(x,\\xi)" + }, + { + "bbox": [ + 306, + 544, + 547, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 544, + 547, + 581 + ], + "type": "inline_equation", + "content": "d_{j}(x,\\xi)" + }, + { + "bbox": [ + 306, + 544, + 547, + 581 + ], + "type": "text", + "content": " are predictions for the approximation and detail coefficients of " + }, + { + "bbox": [ + 306, + 544, + 547, + 581 + ], + "type": "inline_equation", + "content": "\\gamma (x)" + }, + { + "bbox": [ + 306, + 544, + 547, + 581 + ], + "type": "text", + "content": " at level " + }, + { + "bbox": [ + 306, + 544, + 547, + 581 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 306, + 544, + 547, + 581 + ], + "type": "text", + "content": ", respectively. We only predict detail coefficients up to level " + }, + { + "bbox": [ + 306, + 544, + 547, + 581 + ], + "type": "inline_equation", + "content": "j_{1}" + }, + { + "bbox": [ + 306, + 544, + 547, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 547, + 713 + ], + "type": "text", + "content": "Architecture The network is a hybrid analog of the U-Net. It consists of a two-dimensional convolutional encoder, a bottleneck of fully connected layers, and a one-dimensional decoder. The encoder and decoder are connected through skip-connections. The approximation and detail coefficients at the lowest resolution level " + }, + { + "bbox": [ + 305, + 594, + 547, + 713 + ], + "type": "inline_equation", + "content": "j_0" + }, + { + "bbox": [ + 305, + 594, + 547, + 713 + ], + "type": "text", + "content": " are predicted in the bottleneck. Afterward, the Pyramid Algorithm takes over to compute approximation coefficients at higher resolution levels (the decoder) using learnable wavelet filters. In Fig. 3, we provide a schematic overview of the network architecture." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24104" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "Loss To measure the discrepancy between the ground truth and the predicted curve, we define" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 109, + 287, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 109, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 59, + 109, + 287, + 137 + ], + "type": "interline_equation", + "content": "\\begin{array}{r l} L (G (x, \\xi), \\tilde {a} (\\gamma (x))) := & \\| [ a _ {j _ {2}} ] _ {1} - \\tilde {a} _ {j _ {2}} ([ \\gamma (x) ] _ {1}) \\| _ {2} + \\\\ & \\| [ a _ {j _ {2}} ] _ {2} - \\tilde {a} _ {j _ {2}} ([ \\gamma (x) ] _ {2}) \\| _ {2}. \\end{array} \\tag {6}", + "image_path": "926e4d9c8e8b6e56d0e4eb2d14714a46ad555f342c9dc803ef98f28e7a88a6a1.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 149, + 288, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 149, + 288, + 211 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 288, + 211 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 46, + 149, + 288, + 211 + ], + "type": "inline_equation", + "content": "\\tilde{a}" + }, + { + "bbox": [ + 46, + 149, + 288, + 211 + ], + "type": "text", + "content": " thus correspond to the ground-truth approximation coefficients, and the subscripts 1, 2 denote the spatial dimension. In other words, the loss corresponds to the componentwise " + }, + { + "bbox": [ + 46, + 149, + 288, + 211 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 46, + 149, + 288, + 211 + ], + "type": "text", + "content": "-error between the curves on resolution level " + }, + { + "bbox": [ + 46, + 149, + 288, + 211 + ], + "type": "inline_equation", + "content": "j_2" + }, + { + "bbox": [ + 46, + 149, + 288, + 211 + ], + "type": "text", + "content": " with approximation coefficients " + }, + { + "bbox": [ + 46, + 149, + 288, + 211 + ], + "type": "inline_equation", + "content": "a_{j_2}(x,\\xi)" + }, + { + "bbox": [ + 46, + 149, + 288, + 211 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 149, + 288, + 211 + ], + "type": "inline_equation", + "content": "\\tilde{a}_{j_2}(\\gamma (x))" + }, + { + "bbox": [ + 46, + 149, + 288, + 211 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 235, + 289, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 289, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 289, + 559 + ], + "type": "text", + "content": "Optimization We use two different optimizers during training: one for the encoder and MLP network (unconstrained), and one for the decoder (constrained wavelet network). The reason for this is that the needed step sizes on the non-trivial submanifold may significantly differ from those on the unconstrained (flat) parameter space. For the training of the model, we use plain SGD for the first eight epochs for both the constrained and unconstrained parameters. During this period, the learning rate for the unconstrained parameters is linearly increased from " + }, + { + "bbox": [ + 46, + 235, + 289, + 559 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 46, + 235, + 289, + 559 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 235, + 289, + 559 + ], + "type": "inline_equation", + "content": "2 \\cdot 10^{-4}" + }, + { + "bbox": [ + 46, + 235, + 289, + 559 + ], + "type": "text", + "content": ". The learning rate for the constrained parameters (wavelet filters) is linearly increased from " + }, + { + "bbox": [ + 46, + 235, + 289, + 559 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 235, + 289, + 559 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 235, + 289, + 559 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 46, + 235, + 289, + 559 + ], + "type": "text", + "content": ". After the initial warmup stage, we switch to the Adam optimizer for the unconstrained parameters. For both the constrained and unconstrained parameters, we use learning rate schedulers and decrease the learning rate by a factor of 0.85 if no significant improvements in the validation loss are observed during the last ten epochs. We train all models for 250 epochs using a batch size of 32 and use the last epoch for inference. The computations were performed in PYTORCH on a Geforce RTX 2080 Ti. An average training session takes about two and a half hours for the wavelet network and a day for the nnUNet. The inference times for the wavelet network were 12.5ms and 14ms for the prostate and spleen, respectively, using a batch size of 32. The inference times for the nnUNet were 17.8ms and 23.4ms for the prostate and spleen, respectively, using a batch size of 32." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 583, + 111, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 583, + 111, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 583, + 111, + 594 + ], + "type": "text", + "content": "4.1.3 Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "We have evaluated the performance of both our wavelet networks for different orders and a state-of-the-art baseline 2d-nnUNet [13] on the unseen test data, see Table 1. For both the spleen and prostate, we observe that the best wavelet networks outperform the baseline. We note that the task-optimized wavelets differ significantly from the wavelets randomly initialized at the start of training. A comparison of an initial and task-optimized wavelet is depicted in Figure 5, see Sec. 8.4 for more examples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 70, + 545, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 70, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 545, + 126 + ], + "type": "text", + "content": "Table 1. Mean and standard deviation (in parentheses) of the dice score on the unseen test sets. The number of parameters in our wavelet-based networks ranged from 5M to 10M parameters, while the nnUNet used approximately 40M and 18M parameters for the spleen and prostate, respectively." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 320, + 145, + 531, + 243 + ], + "blocks": [ + { + "bbox": [ + 320, + 145, + 531, + 243 + ], + "lines": [ + { + "bbox": [ + 320, + 145, + 531, + 243 + ], + "spans": [ + { + "bbox": [ + 320, + 145, + 531, + 243 + ], + "type": "table", + "html": "
MODELDICE SPLEENDICE PROSTATE
NNUNET0.914 (1.74 · 10-1)0.896 (1.27 · 10-1)
ORDER 30.911 (7.38 · 10-2)0.929 (4.66 · 10-2)
ORDER 40.911 (7.85 · 10-2)0.935 (3.48 · 10-2)
ORDER 50.916 (6.94 · 10-2)0.935 (4.11 · 10-2)
ORDER 60.917 (7.17 · 10-2)0.934 (4.14 · 10-2)
ORDER 70.921 (6.91 · 10-2)0.928 (4.03 · 10-2)
ORDER 80.919 (6.64 · 10-2)0.934 (3.62 · 10-2)
", + "image_path": "5f676562392b7a916920dca2059bfc3d26403f4385cf9dc6f2d628fe1bc12f2b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 266, + 383, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 266, + 383, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 266, + 383, + 277 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 286, + 547, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 547, + 453 + ], + "type": "text", + "content": "In this paper, we have introduced the CERM framework for creating task-driven wavelets. Although our aim is to find such wavelets, in principle the framework lends itself to imposing generic constraints on parametric models. The wavelet constraints can be formulated as a finite system of equations. Under mild smoothness and non-degeneracy conditions, the networks can be made to obey the constraints exactly throughout the entire training procedure by performing SGD on a curved space. We can then utilize these wavelet constraints in a convolutional network, where the filters of the network now parameterize wavelets themselves. We have applied these wavelet networks to the prediction of boundaries of simply connected regions in medical images, where they outperform strong baselines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 466, + 365, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 365, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 365, + 478 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 486, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 310, + 486, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 486, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 310, + 486, + 545, + 518 + ], + "type": "text", + "content": "[1] P. Abry and P. Flandrin. On the initialization of the discrete wavelet transform algorithm. IEEE Signal Processing Letters, 1(2):32-34, 1994. 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 520, + 547, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 520, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 311, + 520, + 547, + 564 + ], + "type": "text", + "content": "[2] Brandon Amos, Lei Xu, and J. Zico Kolter. Input convex neural networks. In Proceedings of the 34th International Conference on Machine Learning, pages 146-155. PMLR, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 565, + 547, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 565, + 547, + 620 + ], + "spans": [ + { + "bbox": [ + 311, + 565, + 547, + 620 + ], + "type": "text", + "content": "[3] Michela Antonelli, Annika Reinke, Spyridon Bakas, Keyvan Farahani, Annette Kopp-Schneider, Bennett A Landman, Geert Litjens, Bjoern Menze, Olaf Ronneberger, Ronald M Summers, et al. The medical segmentation decathlon. Nature communications, 13(1):1-13, 2022. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 621, + 547, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 621, + 547, + 654 + ], + "spans": [ + { + "bbox": [ + 311, + 621, + 547, + 654 + ], + "type": "text", + "content": "[4] Randall Balestriero and Yann LeCun. Police: Provably optimal linear constraint enforcement for deep neural networks, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 656, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 656, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 311, + 656, + 545, + 678 + ], + "type": "text", + "content": "[5] D.P. Bertsekas and W. Rheinboldt. Constrained Optimization and Lagrange Multiplier Methods. Elsevier Science, 2014. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 680, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 680, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 311, + 680, + 547, + 713 + ], + "type": "text", + "content": "[6] Silvere Bonnabel. Stochastic gradient descent on riemannian manifolds. IEEE Transactions on Automatic Control, 58(9): 2217-2229, 2013. 2, 1, 5" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24105" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 95 + ], + "type": "text", + "content": "[7] N. Boumal. An Introduction to Optimization on Smooth Manifolds. Cambridge University Press, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 95, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 95, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 53, + 95, + 288, + 139 + ], + "type": "text", + "content": "[8] Michael M Bronstein, Joan Bruna, Yann LeCun, Arthur Szlam, and Pierre Vandergheynst. Geometric deep learning: going beyond euclidean data. IEEE Signal Processing Magazine, 34(4):18-42, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 140, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 140, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 53, + 140, + 288, + 183 + ], + "type": "text", + "content": "[9] Michael M Bronstein, Joan Bruna, Taco Cohen, and Petar Velicković. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges. arXiv preprint arXiv:2104.13478, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "type": "text", + "content": "[10] Albert Cohen, Ingrid Daubechies, and J-C Feauveau. Biorthogonal bases of compactly supported wavelets. Communications on pure and applied mathematics, 45(5):485-560, 1992. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 270 + ], + "type": "text", + "content": "[11] Priya Donti, David Rolnick, and J Zico Kolter. Dc3: A learning method for optimization with hard constraints. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 271, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 288, + 294 + ], + "type": "text", + "content": "[12] M.W. Frazier. An Introduction to Wavelets Through Linear Algebra. Springer New York, 2001. 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 294, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 288, + 337 + ], + "type": "text", + "content": "[13] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "type": "text", + "content": "[14] Hiroyuki Kasai, Pratik Jawanpuria, and Bamdev Mishra. Riemannian adaptive stochastic gradient algorithms on matrix manifolds. In International Conference on Machine Learning, pages 3262-3271. PMLR, 2019. 2, 1, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "type": "text", + "content": "[15] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings. International Conference on Learning Representations, ICLR, 2015. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 437, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 470 + ], + "type": "text", + "content": "[16]FirstName LastName. Task-driven wavelets using constrained empirical risk minimization, 2023. Supplied as supplemental material supplementary-notes.pdf. 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 471, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 503 + ], + "type": "text", + "content": "[17] Yann LeCun, Yoshua Bengio, et al. Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks, 3361(10):1995, 1995. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 504, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 288, + 525 + ], + "type": "text", + "content": "[18] John M Lee. Riemannian manifolds: an introduction to curvature. Springer Science & Business Media, 2006. 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 526, + 288, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 288, + 548 + ], + "type": "text", + "content": "[19] John M Lee. Smooth manifolds. In Introduction to smooth manifolds. Springer, 2013. 2, 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 548, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 581 + ], + "type": "text", + "content": "[20] Benedict Leimkuhler, Timothee Pouchon, Tiffany Vlaar, and Amos Storkey. Constraint-based regularization of neural networks. arXiv preprint arXiv:2006.10114, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 582, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 288, + 624 + ], + "type": "text", + "content": "[21] Benedict Leimkuhler, Tiffany Vlaar, Timothee Pouchon, and Amos Storkey. Better training using weight-constrained stochastic dynamics. arXiv preprint arXiv:2106.10704, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "text", + "content": "[22] Qiufu Li, Linlin Shen, Sheng Guo, and Zhihui Lai. Wavelet integrated cnns for noise-robust image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7245–7254, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[23] Cho-Chun Liu, Yaohui Liu, Zhengding Qiu, and Xiyu Du. On the initialization of the discrete wavelet transform algorithm. 1997 IEEE International Conference on Intelligent Processing Systems (Cat. No.97TH8335), 2:1220-1222 vol.2, 1997. 14" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 95, + 547, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 118 + ], + "type": "text", + "content": "[25] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam, 2018. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 118, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 547, + 152 + ], + "type": "text", + "content": "[26] Stephane Mallat. A Wavelet Tour of Signal Processing, Third Edition: The Sparse Way. Academic Press, Inc., USA, 3rd edition, 2008. 2, 10" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 152, + 547, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 547, + 186 + ], + "type": "text", + "content": "[27] Pablo Márquez-Neila, Mathieu Salzmann, and Pascal Fua. Imposing hard constraints on deep networks: Promises and limitations. arXiv preprint arXiv:1706.02025, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 186, + 547, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 547, + 230 + ], + "type": "text", + "content": "[28] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 231, + 547, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 547, + 253 + ], + "type": "text", + "content": "[29] L. Montefusco and L. Puccio. Wavelets: Theory, Algorithms, and Applications. Elsevier Science, 2014. 5, 11, 12, 14" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 254, + 547, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 547, + 298 + ], + "type": "text", + "content": "[30] Deepak Pathak, Philipp Krahenbuhl, and Trevor Darrell. Constrained convolutional neural networks for weakly supervised segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1796-1804, 2015. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 299, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 547, + 331 + ], + "type": "text", + "content": "[31] M.C. Pereyra and L.A. Ward. Harmonic Analysis: From Fourier to Wavelets. American Mathematical Society, 2012. 5, 11, 12" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 332, + 547, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 547, + 365 + ], + "type": "text", + "content": "[32] Jack Richter-Powell, Yaron Lipman, and Ricky T. Q. Chen. Neural conservation laws: A divergence-free perspective, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 365, + 547, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 547, + 409 + ], + "type": "text", + "content": "[33] David W Romero, Erik J Bekkers, Jakub M Tomczak, and Mark Hoogendoorn. Wavelet networks: Scale equivariant learning from raw waveforms. arXiv preprint arXiv:2006.05259, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 411, + 547, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 547, + 464 + ], + "type": "text", + "content": "[34] Soumava Kumar Roy, Zakaria Mhammedi, and Mehrtash Harandi. Geometry aware constrained optimization techniques for deep learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4460-4469, 2018. 2, 1, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 467, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 547, + 510 + ], + "type": "text", + "content": "[35] Hiroyuki Sato, Hiroyuki Kasai, and Bamdev Mishra. Riemannian stochastic variance reduced gradient algorithm with retraction and vector transport. SIAM Journal on Optimization, 29(2):1444-1472, 2019. 2, 1, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 511, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 547, + 544 + ], + "type": "text", + "content": "[36] Vladimir Vapnik. Principles of risk minimization for learning theory. Advances in neural information processing systems, 4, 1991. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 545, + 547, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 547, + 567 + ], + "type": "text", + "content": "[37] Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 1999. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 568, + 547, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 547, + 590 + ], + "type": "text", + "content": "[38] D.F. Walnut. An Introduction to Wavelet Analysis. Birkhäuser Boston, 2002. 12, 15" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 591, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 547, + 635 + ], + "type": "text", + "content": "[39] Maurice Weiler, Patrick Forre, Erik Verlinde, and Max Welling. Coordinate independent convolutional networks-isometry and gauge equivariant convolutions on riemannian manifolds. arXiv preprint arXiv:2106.06020, 2021. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 635, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 547, + 690 + ], + "type": "text", + "content": "[40] Yang You, Jing Li, Sashank Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes. In International Conference on Learning Representations, 2020. 1" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "text", + "content": "[41] Jason J Yu, Konstantinos G Derpanis, and Marcus A Brubaker. Wavelet flow: Fast training of high resolution normalizing" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24106" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 140 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "text", + "content": "flows. Advances in Neural Information Processing Systems, 33:6184-6196, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "text", + "content": "[42] Hongyi Zhang, Sashank J. Reddi, and Suvrit Sra. Riemannian svrg: Fast stochastic optimization on riemannian manifolds. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 2, 1, 5" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24107" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/9dbd5f9f-08af-4d23-b3b6-162e326d318c_content_list.json b/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/9dbd5f9f-08af-4d23-b3b6-162e326d318c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f69492bdb96fdbd452ee57cd48f9073e1e551bbd --- /dev/null +++ b/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/9dbd5f9f-08af-4d23-b3b6-162e326d318c_content_list.json @@ -0,0 +1,1438 @@ +[ + { + "type": "text", + "text": "Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations", + "text_level": 1, + "bbox": [ + 241, + 130, + 728, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daan de Geus Gijs Dubbelman", + "bbox": [ + 336, + 203, + 630, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eindhoven University of Technology", + "bbox": [ + 336, + 222, + 630, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{d.c.d.geus, g.dubbelman}@tue.nl", + "bbox": [ + 341, + 241, + 622, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Part-aware panoptic segmentation (PPS) requires (a) that each foreground object and background region in an image is segmented and classified, and (b) that all parts within foreground objects are segmented, classified and linked to their parent object. Existing methods approach PPS by separately conducting object-level and part-level segmentation. However, their part-level predictions are not linked to individual parent objects. Therefore, their learning objective is not aligned with the PPS task objective, which harms the PPS performance. To solve this, and make more accurate PPS predictions, we propose Task-Aligned Part-aware Panoptic Segmentation (TAPPS). This method uses a set of shared queries to jointly predict (a) object-level segments, and (b) the part-level segments within those same objects. As a result, TAPPS learns to predict part-level segments that are linked to individual parent objects, aligning the learning objective with the task objective, and allowing TAPPS to leverage joint object-part representations. With experiments, we show that TAPPS considerably outperforms methods that predict objects and parts separately, and achieves new state-of-the-art PPS results.", + "bbox": [ + 75, + 325, + 473, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 676, + 209, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To fully understand what is depicted in an image, it is important to consider concepts at different levels of abstraction. For rich scene understanding, we should not only recognize foreground objects (e.g., car, human) and background regions (e.g., sky, ocean), but also simultaneously identify the parts that constitute the objects (e.g., car-wheel, human-arm). In a step towards such comprehensive scene understanding, De Geus et al. introduced part-aware panoptic segmentation (PPS) [5]. The objective of this computer vision task is (1) to output a segmentation mask and class label for all thing objects and stuff regions in an image like for panoptic segmentation [13] – we call these object-level segments – and (2) to simultaneously segment and classify", + "bbox": [ + 75, + 703, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f3ebccf1cd7a24734832cc79991c8b6d16711034356de08e06fb8f2159a81101.jpg", + "image_caption": [ + "(a) Separate object and part segmentation", + "X Learning objective not aligned with task objective", + "(b) Joint object and part segmentation (ours)" + ], + "image_footnote": [], + "bbox": [ + 501, + 315, + 885, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ba6f1c42782fe4a17c1b152ce6884b5e6aec85c0da13b1681eb1054306f72945.jpg", + "image_caption": [ + "$\\checkmark$ Learning objective aligned with task objective", + "Figure 1. Task-aligned part-aware panoptic segmentation. (a) Existing works separately predict object-level segments and object-instance-unaware part-level segments. (b) In this work, we predict objects and parts jointly, using a set of shared queries. This allows our method to predict parts within individual object segments, aligning its learning objective with the PPS task objective." + ], + "image_footnote": [], + "bbox": [ + 500, + 473, + 888, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "all parts within each identified object. These are called part-level segments and should be explicitly linked to an object-level segment, establishing a part-whole relation.", + "bbox": [ + 496, + 703, + 890, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current state-of-the-art methods [18, 19] address PPS by using two different sets of learnable queries to separately predict object-level and part-level segments, see Fig. 1a. While these methods outperform earlier baselines [5], they have one main limitation: their learning objective is not aligned with the task objective. Where the PPS task objective is to predict parts within each individual object-level segment, these existing works conduct part-level semantic segmentation, predicting part-level masks that cover multiple objects (see Fig. 1a). In other words, these networks", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3174", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are not optimized for the PPS task, but instead solve the surrogate subtasks of object-level panoptic segmentation and part-level semantic segmentation. As a result, to assign parts to individual objects, these methods require post-processing. Additionally, we hypothesize that there are several other negative consequences: (1) These networks learn a conflicting feature representation. They learn that object-level thing instances should be separated, but also that the parts of these object-level instances should be grouped together. We expect that this harms the ability of the networks to separate instances. (2) Predicting objects and parts separately may cause incompatible predictions (e.g., a car-window part with a bicycle object), which makes it unclear which prediction to trust, and requires further post-processing. (3) The networks encode information about objects and their parts separately, whereas this information is potentially complementary. See Sec. 3.2.2 for more details on these limitations.", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we aim to design a simple network for PPS that overcomes these limitations and thereby makes more accurate PPS predictions. To achieve this, we propose the Task-Aligned Part-aware Panoptic Segmentation (TAPPS) method. Instead of using separate queries for objects and parts, TAPPS uses one set of shared queries to jointly represent objects and the parts they contain. Specifically, each of these queries learns to represent at most one object-level segment, for which it predicts (1) a segmentation mask and object-level class, and (2) the segmentation masks and classes for all part-level segments within this object. This is visualized in Fig. 1b and explained in Sec. 3.", + "bbox": [ + 75, + 366, + 472, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With this approach, TAPPS explicitly predicts part-level segments per individual object. Therefore, the network's learning objective is now aligned with the PPS task objective, and TAPPS is directly optimized for the PPS task. As a result, both object- and part-level segmentation are learned in an 'object-instance-aware' manner. This removes the conflict in the learned feature representations, allowing for better instance separability. Additionally, we hypothesize that predicting objects and parts from the same query reduces incompatibilities between object and part predictions, as they are made using the same information. Moreover, we use TAPPS to go even further, and explicitly only predict part segments compatible with predicted object segments. This enforces full object-part compatibility and simplifies the part segmentation task. Finally, the network can now encode complementary object-level and part-level information in a shared query, giving it a richer representation for making object- and part-level segmentation predictions.", + "bbox": [ + 75, + 549, + 472, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With experiments, detailed in Sec. 4, we show that TAPPS outperforms the baseline with separate object and part queries in multiple aspects: (1) The part segmentation quality within identified objects is significantly better, which shows the positive impact of using joint object-part", + "bbox": [ + 75, + 824, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "representations and simplifying the part segmentation task. (2) The object instance segmentation performance is considerably improved, demonstrating the benefits of learning parts in an object-instance-aware manner. Together, this yields a large overall improvement with respect to the baseline, and causes TAPPS to considerably outperform existing works across different benchmarks, achieving new state-of-the-art results. See Sec. 5 for more extensive results.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, we make the following contributions:", + "bbox": [ + 517, + 214, + 867, + 229 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose TAPPS, a simple approach for PPS that aligns the learning objective with the task objective, facilitating object instance separability and enabling joint object-part representations for more accurate PPS predictions.", + "- We use the shared object-part queries to constrain TAPPS to only predict part segments that are compatible with the object class, enforcing object-part compatibility and simplifying the part segmentation task.", + "- We experimentally show the effectiveness of TAPPS across a range of datasets and network configurations." + ], + "bbox": [ + 500, + 231, + 890, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The code for TAPPS is made publicly available through https://tue-mps.github.io/tapps/.", + "bbox": [ + 498, + 383, + 890, + 414 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 433, + 637, + 449 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Part-aware panoptic segmentation. Part-aware panoptic segmentation [5] is an image segmentation task introduced for scene understanding at multiple abstraction levels. It extends panoptic segmentation [13] by requiring (a) object-level panoptic segmentation, and (b) part-level segmentation within object-level segments. Most existing works do not predict parts per object-level segment, but instead make separate predictions for panoptic segmentation and the surrogate task of part-level semantic segmentation - or part segmentation [10, 18, 19]. JPPF [10] proposes a single-network approach with a shared encoder followed by separate heads for semantic, instance, and part segmentation. To merge the predictions by these heads to the PPS format, JPPF proposes a new rule-based fusion strategy that outperforms the originally introduced merging strategy for PPS [5]. Panoptic-PartFormer and its extension Panoptic-PartFormer++ [18, 19] tackle PPS with a Transformer-based approach that independently makes panoptic segmentation and part segmentation predictions using separate sets of learnable queries for thing, stuff and part segments. Alternatively, ViRReq [39] is a paradigm in which PPS is predicted in a cascading fashion, by first segmenting objects and then segmenting the parts within these objects 'by request', but it requires multiple networks to achieve this.", + "bbox": [ + 496, + 460, + 890, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to these existing approaches, we propose a method that jointly predicts object-level segments and the part-level segments within these objects. This aligns the learning objective with the task objective, and yields improved PPS performance. For more details, see Sec. 3.", + "bbox": [ + 496, + 825, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3175", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Part-level image segmentation. Part-level segmentation is also widely studied beyond PPS. Most works address part segmentation, i.e., semantic segmentation for parts [7, 8, 14-17, 20, 23, 27, 32, 34, 35, 37, 42, 43, 51, 52]. For more complete scene understanding, other works also take into account object instances, like PPS does, with instance-aware part segmentation [16, 30, 32, 46-49]. However, these works do not consider background stuff classes, which PPS does consider. Alternatively, UPerNet [44] separately predicts part-level and object-level semantic segmentation, including background classes, but without any instance awareness. Another work has extended PPS by also predicting the relations between objects [31], training a model on annotations from different datasets. In this work, we address the PPS task as introduced by De Geus et al. [5], and we compare our approach to state-of-the-art alternatives.", + "bbox": [ + 76, + 90, + 472, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 349, + 169, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. PPS task definition", + "text_level": 1, + "bbox": [ + 76, + 376, + 259, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Part-aware panoptic segmentation (PPS) [5] requires consistent image segmentation across two abstraction levels: the object level and part level. For object-level segmentation, following panoptic segmentation [13], an image should be divided into $N$ object-level segments, where $N$ varies per image. Each segment $s_i$ consists of a binary mask $\\mathbf{M}_i^{\\mathrm{obj}}$ and an object class label $c_i^{\\mathrm{obj}}$ . Thing classes require a segment per object instance, and stuff classes require a single segment per class. Next, per segment $s_i$ , PPS requires a set of $K_i$ part-level segments. Each part-level segment $s_{i,j}^{\\mathrm{pt}}$ consists of a binary part-level mask $\\mathbf{M}_{i,j}^{\\mathrm{pt}}$ and a part-level class $c_{i,j}^{\\mathrm{pt}}$ that are compatible with the object-level mask and class. Specifically, the part-level mask must be a subset of the object-level mask, and the part-level class should be one of the part-level classes defined for the object-level class. For instance, if $c_i^{\\mathrm{obj}}$ is car, then $c_{i,j}^{\\mathrm{pt}}$ cannot be human-head but it can be car-window. For all parts, a single part-level segment is required for each part class within an object.", + "bbox": [ + 76, + 401, + 472, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, each image should be divided into a set of segments $S = \\{s_i\\}_{i=1}^N = \\{(\\mathbf{M}_i^{\\mathrm{obj}}, c_i^{\\mathrm{obj}}, \\mathcal{P}_i)\\}_{i=1}^N$ , where $\\mathcal{P}_i = \\{s_{i,j}^{\\mathrm{pt}}\\}_{j=1}^{K_i} = \\{(\\mathbf{M}_{i,j}^{\\mathrm{pt}}, c_{i,j}^{\\mathrm{pt}})\\}_{j=1}^{K_i}$ is a set of part-level segments compatible with the object-level segment in $s_i$ . In practice, part-level classes are only defined for some object-level classes. When segment $s_i$ has an object class that does not have parts, then $K_i = 0$ and $s_i = (\\mathbf{M}_i^{\\mathrm{obj}}, c_i^{\\mathrm{obj}})$ .", + "bbox": [ + 76, + 680, + 470, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 800, + 217, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Like existing state-of-the-art PPS approaches [18, 19], our work uses the powerful mask classification metaarchitecture that is the foundation of many state-of-the-art segmentation models [2, 3, 11, 41, 50]. Here, we first explain this mask classification paradigm in more detail.", + "bbox": [ + 76, + 824, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, we explain how existing works use it for PPS, and what the limitations are of these approaches.", + "bbox": [ + 500, + 90, + 890, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Mask classification framework", + "text_level": 1, + "bbox": [ + 500, + 138, + 767, + 154 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The concept of mask classification is to predict a set of $N^{\\mathrm{q}}$ object-level segments, i.e., a set of $N^{\\mathrm{q}}$ pixel-level masks $\\hat{\\mathbf{M}}^{\\mathrm{obj}}$ and corresponding class labels $\\hat{c}^{\\mathrm{obj}}$ . To make these predictions, mask classification networks output two components: high-resolution features $\\mathbf{F} \\in \\mathbb{R}^{E \\times H \\times W}$ and queries $\\mathbf{Q} \\in \\mathbb{R}^{N^{\\mathrm{q}} \\times E}$ , where $H$ and $W$ are the height and width of the features, and $E$ is the feature and query dimensionality. Each of the $N^{\\mathrm{q}}$ queries is used to predict the class label and segmentation mask of at most one object-level segment. The class predictions are a function of only the queries, and the mask predictions are a function of both the queries $\\mathbf{Q}$ and the high-resolution features $\\mathbf{F}$ . To output these high-resolution features, the input image is first fed into a backbone to extract multi-scale features, e.g., ResNet [9] or Swin [24]. Subsequently, these features are further refined and upsampled to high-resolution features by a pixel decoder, e.g., Semantic FPN [12, 22]. In the other part of the network, the queries $\\mathbf{Q}$ are generated by processing learnable queries $\\mathbf{Q}^{\\mathrm{o}}$ using a transformer decoder, which applies self-attention across queries and cross-attention with image features. Through bipartite matching between $N^{\\mathrm{q}}$ predicted and $N$ ground-truth segments, each query is assigned to at most one ground-truth segment. As $N^{\\mathrm{q}}$ is not always equal to $N$ , some queries may not be assigned to a ground-truth segment. If so, they do not receive any supervision for segmentation, and will learn a 'no object' class. In this work, we build upon the Mask2Former [3] instantiation of this meta-architecture.", + "bbox": [ + 496, + 162, + 893, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.2 Mask classification for PPS", + "text_level": 1, + "bbox": [ + 500, + 604, + 743, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the aforementioned meta-architecture, we can conduct object-level segmentation, and thereby solve one aspect of the PPS task definition. To solve the full task, the object-level segments should also be further segmented into parts.", + "bbox": [ + 496, + 628, + 890, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Current state-of-the-art approaches [18, 19] achieve this by introducing an additional set of queries, the part-level queries depicted in Fig. 1a. Each of these part-level queries learns to represent a part-level segment. However, in contrast to the PPS task definition, these part-level segments are not explicitly linked to an object-level segment. Instead, these part-level segments represent an entire part-level class, i.e., their masks contain all pixels belonging to one part-level class across multiple object segments.", + "bbox": [ + 496, + 689, + 890, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This means that the learning objective of these networks is not aligned with the PPS task objective. As a result, they are not directly optimized for PPS, but instead for two surrogate tasks: object-level panoptic segmentation and part-level semantic segmentation. In addition to necessitating", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3176", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2dbf6ba2e62855a536b9cb4c4f8343f012d40b9d8d571649bc4ea437fee07eac.jpg", + "image_caption": [ + "Figure 2. Network architecture. Left: The overall TAPPS architecture. A set of learnable queries and features from a backbone are fed into a pixel decoder and transformer decoder to generate high-resolution features and processed queries. Right: These queries and features are fed into the JOPS head, which predicts for each shared query (a) an object-level class, (b) an object-level segmentation mask, and (c) a set of part-level masks for the part-level classes compatible with the object-level class. Operator $\\otimes$ denotes a matrix multiplication." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 470, + 276 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/533af94f2cab23e2f80e9b3df42fd2f2c2d7c4f976286e49647ab30c269f6ef7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 85, + 893, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "post-processing, we hypothesize that this has several negative consequences: (1) By learning parts in an 'object-instance-unaware' manner, the network learns conflicting representations in the features, F. On the one hand, different thing instances belong to different object-level segments, requiring distinct features. On the other hand, the parts of these instances belong to the same part-level segment, requiring similar features. We hypothesize that this conflict harms the ability of the network to learn to separate object-level thing instances. (2) Separately predicting objects and parts in this way may also cause incompatible predictions. These incompatibilities make it unclear for downstream processes which prediction to trust. To obtain predictions compliant with the PPS definition, rule-based post-processing is required, which discards incompatible predictions [5]. This is undesired because the discarded information is potentially correct. (3) By representing object- and part-level information with separate queries, the network cannot encode complementary information about an individual object and its parts, which it could use to make more informed predictions at these abstraction levels.", + "bbox": [ + 75, + 364, + 472, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Task-aligned PPS", + "text_level": 1, + "bbox": [ + 76, + 695, + 248, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To overcome the limitations of state-of-the-art methods and achieve a better PPS performance, the objective of this work is to design a simple PPS approach that uses the powerful mask classification framework, but that aligns its learning objective with the PPS task objective, and explicitly predicts part-level segments within individual objects. To achieve this, we should (a) generate a unique query for each object-part combination, i.e., for each part-level segment within an object-level segment, and (b) explicitly link these part-level segments to their parent objects, as PPS requires. To accomplish this, we propose Task-Aligned Part-aware Panoptic Segmentation (TAPPS), a method that jointly represents", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "object-level segments and their parts with a set of shared queries, and generates 'per-object' part queries from each shared query. Each of these queries can then be used to predict one part-level segment within an object-level segment, and is automatically linked to its parent object segment.", + "bbox": [ + 496, + 364, + 892, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Overall architecture", + "text_level": 1, + "bbox": [ + 498, + 465, + 694, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Fig. 2, we visualize the network architecture of TAPPS. First, we initialize a set of $N^{\\mathbf{q}}$ learnable queries, $\\mathbf{Q}^0 \\in \\mathbb{R}^{N^{\\mathbf{q}} \\times E}$ . Each of these queries learns to represent one object-level segment, but also the part-level segments belonging to this object (e.g., car3 and also car3-wheels, car3-windows, etc.). Following the mask classification framework described in Sec. 3.2, these learnable queries $\\mathbf{Q}^0$ and the features from a backbone are fed into a decoder, which generates a set of processed queries $\\mathbf{Q}$ and high-resolution features $\\mathbf{F}$ . Subsequently, these processed queries and features enter the Joint Object and Part Segmentation (JOPS) head. For each query $\\mathbf{Q}_i \\in \\mathbb{R}^E$ , this head predicts (a) an object-level class, (b) an object-level segmentation mask, and (c) a set of part-level segmentation masks and classes for the parts within this object-level segment.", + "bbox": [ + 496, + 491, + 890, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With this approach, we solve the limitations of existing works. Following the PPS task definition, we predict part-level segments per individual object-level segment. (1) As a result, we learn object-instance-aware representations for both parts and objects, improving the ability of the network to separate instances. (2) By predicting objects and parts jointly, the compatibility between both sets of predictions greatly improves. Moreover, in the JOPS head, we ensure that we only predict compatible parts, ensuring full object-part compatibility (see Sec. 3.3.2). (3) The joint object-part representations allow the network to fully leverage the complementary information from both abstraction levels.", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3177", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 JOPS head", + "text_level": 1, + "bbox": [ + 76, + 90, + 209, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Within the JOPS head, we predict an object-level segment and compatible part-level segments for each shared query. For the object-level predictions, we follow Mask2Former [3]: (1) For the object-level class, we feed each processed query $\\mathbf{Q}_i$ through a single fully connected layer to predict a score for each possible class, and obtain the predicted class $\\hat{c}_i^{\\mathrm{obj}}$ by picking the highest-scoring class. (2) For the object-level mask, we feed the query $\\mathbf{Q}_i$ through a 3-layer MLP and generate a mask by taking the product of the resulting mask queries $\\mathbf{Q}_i^{\\mathrm{m}}$ and the features $\\mathbf{F}$ , and applying a sigmoid activation, yielding $\\hat{\\mathbf{M}}_i^{\\mathrm{obj}} \\in [0,1]^{H \\times W}$ .", + "bbox": [ + 75, + 114, + 468, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To predict part-level segments using the same features $\\mathbf{F}$ , we need to generate queries for each part class within an object-level segment. To do so, we first apply an MLP to each query $\\mathbf{Q}_i$ to adapt it for part-level segmentation, and then apply $N^{\\mathrm{pc}}$ different fully-connected layers, where $N^{\\mathrm{pc}}$ is the total number of part-level classes in the dataset. This results in $\\mathbf{Q}_i^{\\mathrm{pt}} \\in \\mathbb{R}^{N^{\\mathrm{pc}} \\times E}$ , a set of 'per-object' part-level queries, where each query always corresponds to a fixed, predetermined part class. We could then take the product of $\\mathbf{Q}_i^{\\mathrm{pt}}$ and the features $\\mathbf{F}$ to generate a segmentation mask for all part-level classes. However, we already have an object-level class prediction for each query, and we know that only a subset of all $N^{\\mathrm{pc}}$ part classes is compatible with a certain object class. Therefore, we propose to only predict and supervise part-level masks for those compatible part-level classes. This simplifies the part segmentation task that the network needs to learn, as it no longer has to learn to predict 'empty' segmentation masks for incompatible part classes. Concretely, as visualized in Fig. 2, we identify the object-level class $\\hat{c}_i^{\\mathrm{obj}}$ for each query $\\mathbf{Q}_i$ , and keep only the part-level queries $\\mathbf{Q}_i^{\\mathrm{pt,c}} \\in \\mathbb{R}^{N^{\\mathrm{c}} \\times E}$ that correspond to the $N^{\\mathrm{c}}$ part classes that are compatible with the object-level class. Then, we compute the product of these remaining queries $\\mathbf{Q}_i^{\\mathrm{pt,c}}$ and the features $\\mathbf{F}$ , and apply a sigmoid activation to generate compatible part segmentation masks $\\hat{\\mathbf{M}}_i^{\\mathrm{pt}} \\in [0,1]^{N^{\\mathrm{c}} \\times H \\times W}$ . As each per-object part query corresponds to a fixed part class, we also know the part classes $\\hat{c}_i^{\\mathrm{pt}}$ . Note that the number $N^{\\mathrm{c}}$ depends on the object-level class and will therefore vary per query $\\mathbf{Q}_i$ .", + "bbox": [ + 75, + 282, + 470, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3 Training", + "text_level": 1, + "bbox": [ + 76, + 739, + 192, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To assign each query to at most one ground-truth object-level segment with corresponding part-level segments during training, we apply bipartite matching based on the predicted and ground-truth object-level classes and masks.", + "bbox": [ + 75, + 763, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To supervise the object-level segments, we use the cross-entropy loss for the classes, and both the Dice [28] and cross-entropy loss for the segmentation masks. Together, these losses form the object-level loss $L_{\\mathrm{obj}}$ . For the part-level segmentation masks, we also use the Dice and cross", + "bbox": [ + 75, + 825, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "entropy loss, together forming part-level loss $L_{\\mathrm{pt}}$ . We balance the losses with weights $\\lambda_{\\mathrm{obj}}$ and $\\lambda_{\\mathrm{pt}}$ to calculate the total loss", + "bbox": [ + 496, + 90, + 890, + 133 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\lambda_ {\\mathrm {o b j}} L _ {\\mathrm {o b j}} + \\lambda_ {\\mathrm {p t}} L _ {\\mathrm {p t}}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 136, + 890, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We find that using $\\lambda_{\\mathrm{obj}} = \\lambda_{\\mathrm{pt}} = 1$ provides the best balance; see the supplementary material for more details.", + "bbox": [ + 496, + 159, + 890, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental setup", + "text_level": 1, + "bbox": [ + 498, + 204, + 689, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We use the two PPS benchmarks for evaluation.", + "bbox": [ + 498, + 229, + 890, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cityscapes Panoptic Parts (Cityscapes-PP) [5] extends the original Cityscapes dataset [4] with part-level labels. It consists of street scene images from several cities. It has labels for 19 object classes (11 stuff; 8 thing). Part classes are defined and labeled for 5 object-level thing classes; there are 23 part classes in total. We train on the train split (2975 images), and evaluate on the val split (500 images).", + "bbox": [ + 496, + 244, + 890, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pascal Panoptic Parts (Pascal-PP) [5], which combines existing labels for Pascal VOC [1, 6, 29], consists of a wide range of scenes and classes. There are 59 object-level classes (39 stuff; 20 thing), and part-level classes are defined for 15 thing classes. Unless otherwise indicated, we use the default part class definition with 57 part classes in total [5]. To evaluate a more challenging setting, we additionally evaluate on Pascal-PP-107, which uses the 107 non-background classes from the Pascal-Part-108 definition introduced by Michieli et al. [27] for part segmentation. For both class definitions, we train on the training split (4998 images), and evaluate on the validation split (5105 images).", + "bbox": [ + 496, + 351, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline. Our baseline is a version of TAPPS that uses the same network architecture, but uses a separate set of 100 additional queries for part-level segmentation (see also Fig. 1a). Comparing the results of the baseline in Tab. 1 to existing methods in Tab. 2, we find that our baseline outperforms state-of-the-art approaches that also use separate part-level queries [18, 19], indicating that it is a strong baseline. See the supplementary material for more details.", + "bbox": [ + 496, + 550, + 890, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. The part-aware panoptic segmentation performance is evaluated using the default Part-aware Panoptic Quality (PartPQ) metric [5]. It captures both the ability to recognize and segment object-level segments (i.e., stuff regions and thing instances), and the ability to further segment the identified object-level segments into part-level masks. The PartPQ per object-level class $c$ is given by", + "bbox": [ + 496, + 691, + 890, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {P a r t P Q} _ {c} = \\frac {\\sum_ {(p , g) \\in T P _ {c}} \\operatorname {I o U} _ {\\mathrm {p}} (p , g)}{| T P _ {c} | + \\frac {1}{2} | F P _ {c} | + \\frac {1}{2} | F N _ {c} |}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 806, + 890, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $TP_{c}, FP_{c}$ , and $FN_{c}$ are the sets of true positive, false positive and false negative segments, respectively. A pair of predicted and ground-truth segments $(p, g)$ of the same", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3178", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/75799b0b186e8524ce0d16ca70d3c47558efbe8d7e4f6507c369f8d3b6a7b440.jpg", + "table_caption": [], + "table_footnote": [ + "(a) Pascal-PP validation [1, 5, 6, 29]." + ], + "table_body": "
MethodPartPQPartSQPQ
PtNo ptAllPtThStAll
ImageNet pre-training
Baseline55.238.842.971.562.037.345.7
TAPPS (ours)59.639.444.674.365.637.647.1
COCO pre-training
Baseline64.849.953.773.174.848.157.1
TAPPS (ours)67.250.454.775.175.748.557.7
", + "bbox": [ + 80, + 88, + 477, + 220 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/75ccc663a9b7a092dd860bf6b2a86cd08fbeb4fdd9da2cbc27de4e2bfcc369e1.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Cityscapes-PP val [4, 5]." + ], + "table_body": "
MethodPartPQPartSQPQ
PtNo ptAllPtThStAll
ImageNet pre-training
Baseline46.662.658.466.153.867.161.5
TAPPS (ours)48.763.159.366.855.667.362.4
COCO pre-training
Baseline48.264.860.466.756.169.063.6
TAPPS (ours)48.965.761.366.957.269.764.4
", + "bbox": [ + 488, + 88, + 885, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Main results. We compare TAPPS to a strong baseline that uses separate sets of queries to predict object- and part-level segments, instead of predicting object- and part-level segments jointly like TAPPS (see Sec. 4).", + "bbox": [ + 75, + 241, + 890, + 270 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "class $c$ is part of $TP_{c}$ if the Intersection-over-Union (IoU) of their object-level masks is larger than 0.5. If a ground-truth segment is not identified, it is part of $FN_{c}$ ; if a prediction is incorrect, it is part of $FP_{c}$ . The IoUp term captures the segmentation performance within identified object-level segments. For object-level classes for which part classes are defined, it calculates the part-level mIoU. Otherwise, it uses the object-level IoU. We report the mean PartPQ over all classes, but also separately for object-level classes with parts (PartPQpt) and without parts (PartPQNoPt).", + "bbox": [ + 75, + 295, + 468, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To individually assess the ability of methods to conduct part-level segmentation within identified objects, we report the Part Segmentation Quality for object-level classes that have parts, $\\mathrm{PartSQ^{Pt}}$ . Following De Geus et al. [5], per object-level class $c$ , the $\\mathrm{PartSQ^{Pt}}$ calculates the average part segmentation mIoU within object-level segments:", + "bbox": [ + 75, + 446, + 468, + 537 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {P a r t S Q} _ {c} ^ {\\mathrm {P t}} = \\frac {\\sum_ {(p , g) \\in T P _ {c}} \\operatorname {I o U} _ {\\mathrm {p}} (p , g)}{| T P _ {c} |}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 547, + 468, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the ability of networks to conduct object-level panoptic segmentation, we report the Panoptic Quality (PQ) metric [13]. Similarly to the PartPQ, we report the average PQ over all classes, but also over all thing classes $(\\mathrm{PQ}^{\\mathrm{Th}})$ and stuff classes $(\\mathrm{PQ}^{\\mathrm{St}})$ separately.", + "bbox": [ + 75, + 594, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. TAPPS is built on top of the publicly available code of state-of-the-art panoptic segmentation network Mask2Former [3]. For all datasets, we use a batch size of 16 images, and train on 4 Nvidia A100 GPUs. To optimize TAPPS, we use AdamW [26], a weight decay of 0.05, and a polynomial learning rate decay schedule with an initial learning rate of $10^{-4}$ and a power of 0.9. For Cityscapes-PP, we train for 90k iterations and apply conventional data augmentation steps [3, 18]: random flip, random resize with a factor between 0.5 and 2.0, and finally a random crop of $512 \\times 1024$ pixels. For Pascal-PP, we train for 60k iterations in case of ImageNet pre-training, but for only 10k iterations in case of COCO pre-training, to prevent overfitting. Following state-of-the-art panoptic segmenta", + "bbox": [ + 75, + 688, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion implementations on COCO [3, 21], we apply large-scale jittering with a scale between 0.1 and 2.0 followed by a random crop of $1024 \\times 1024$ pixels. During inference, we resize the image such that the shortest side is 800 pixels. Note that we use exactly the same training and testing settings for both TAPPS and the baseline. For more implementation details, see the supplementary material.", + "bbox": [ + 496, + 295, + 890, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 500, + 414, + 586, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Main results", + "text_level": 1, + "bbox": [ + 500, + 439, + 633, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, we compare TAPPS to the strong baseline that uses separate sets of queries for object-level panoptic segmentation and part-level semantic segmentation (see Sec. 4). The results in Tab. 1a demonstrate that TAPPS significantly outperforms the baseline on Pascal-PP in several aspects. Most importantly, the PartPQ for object-level classes with parts $(\\mathrm{PartPQ}^{\\mathrm{Pt}})$ is $+4.4$ or $+2.4$ higher, depending on the pretraining strategy. Looking in more detail, we find that this increase is caused by two individual improvements: (1) The part-level segmentation quality $(\\mathrm{PartSQ}^{\\mathrm{Pt}})$ is considerably higher than for the baseline. This shows the positive impact of having a joint representation for objects and parts, and simplifying the part segmentation task by only allowing compatible predictions (see also Sec. 5.3). (2) The panoptic quality for thing classes $(\\mathrm{PQ}^{\\mathrm{Th}})$ also sees a substantial increase. Here, it should be noted that (a) all the object-level classes with parts are thing classes, and (b) thing classes require instance separation. Thus, this improvement does not only show the benefit of learning objects and parts jointly, but also indicates that learning parts in an object-instance-aware manner indeed improves the ability of the network to separate object instances (see also Sec. 5.4).", + "bbox": [ + 496, + 462, + 890, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For Cityscapes-PP, we see similar results in Tab. 1b. Again, we observe improvements on the PartPQ $^{\\mathrm{Pt}}$ , PartSQ $^{\\mathrm{Pt}}$ and PQ $^{\\mathrm{Th}}$ metrics. In this case, the absolute improvements are slightly smaller. This is expected because this dataset contains significantly fewer classes and very similar images, making it easier to learn the PPS task and limiting the potential gains that can still be obtained by TAPPS.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3179", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d77bcd8f098b655f07936688a10de1f5c199d532bc7920cdc5d83c5912904242.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackbonePre-trainingPartPQPartSQPQ
PtNo PtAllPtAll
Pascal-PP validation
JPPF [10]EfficientNet-B5I48.326.932.2--
TAPPS (ours)RN-50I59.639.444.674.347.1
Panoptic-PartFormer† [18]RN-50I,C56.138.843.266.847.6
Panoptic-PartFormer++† [19]RN-50I,C52.642.645.160.451.6
TAPPS (ours)RN-50I,C67.250.454.775.157.7
Panoptic-PartFormer† [18]Swin-BI,C64.350.654.170.858.1
Panoptic-PartFormer++† [19]Swin-BI,C48.952.151.353.059.8
TAPPS (ours)Swin-BI,C72.256.360.478.163.0
Cityscapes-PP val
Panoptic-PartFormer [18]RN-50I--54.5-57.8
Panoptic-PartFormer++ [19]RN-50I--57.5-61.6
JPPF [10]EfficientNet-B5I47.763.859.6--
TAPPS (ours)RN-50I48.763.159.366.862.4
Panoptic-PartFormer [18]RN-50I,C43.962.457.560.161.6
Panoptic-PartFormer++ [19]RN-50I,C42.565.159.2-63.6
TAPPS (ours)RN-50I,C48.965.761.366.964.4
Panoptic-PartFormer [18]Swin-BI,C45.667.862.059.066.6
Panoptic-PartFormer++ [19]Swin-BI,C46.068.262.3-68.0
Panoptic-PartFormer++ [19]ConvNeXt-BI,C46.469.163.1-68.2
SegFormer-B5 + CondInst + BPR [38–40, 45]MiT-B5 + RN-50I,C*48.667.562.5--
TAPPS (ours)Swin-BI,C53.069.064.868.068.0
", + "bbox": [ + 145, + 88, + 823, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparison with state of the art", + "text_level": 1, + "bbox": [ + 76, + 530, + 362, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 2, we compare TAPPS to existing state-of-the-art methods across different datasets, backbones, and pretraining settings. On both Pascal-PP and Cityscapes-PP, TAPPS significantly outperforms existing work. Most importantly, it consistently scores higher on the PartPQ, PartPQ $^{\\text{Pt}}$ , and PartSQ $^{\\text{Pt}}$ metrics. TAPPS is only slightly outperformed by JPPF [10] on Cityscapes-PP with ImageNet pre-training, but we note that JPPF uses the EfficientNet-B5 [36] backbone, which is much more powerful than the ResNet-50 [9] used by TAPPS. Moreover, we do outperform JPPF by a large margin on the Pascal-PP dataset, showing the strength of TAPPS on more complex datasets. Overall, we achieve new state-of-the-art results on both datasets, obtaining PartPQ scores of 60.4 and 64.8 on Pascal-PP and Cityscapes-PP, improvements of +6.3 and +1.7, respectively. In the supplementary material, we compare qualitative examples of TAPPS, our baseline, and existing approaches, and we also show typical failure cases.", + "bbox": [ + 75, + 551, + 472, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablations", + "text_level": 1, + "bbox": [ + 76, + 832, + 189, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Predicting only compatible parts. Using shared queries for object- and part-level segmentation enables TAPPS to only allow part segmentation predictions that", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/89a574898f629d0c72a6ca90360a9689143ffb4af9be8dc8074342969d3d8e26.jpg", + "table_caption": [ + "Table 2. Comparison with state of the art. Evaluation on the Cityscapes-PP and Pascal-PP benchmarks [4-6]. RN-50 is ResNet-50 [9]. Other backbones are EfficientNet-B5 [36], MiT-B5 [45], Swin-B [24] and ConvNeXt-B [25]. I = ImageNet [33], C = COCO panoptic [21], C* = COCO pre-training for instance segmentation. †PartPQ scores for these existing methods have been re-evaluated using official code [5] and are higher than originally reported on Pascal-PP [18, 19], see supplementary material for more details." + ], + "table_footnote": [], + "table_body": "
Predict compat. partsPartPQPartSQPQObj. w/o conflicts
TrainingTestingPtNo PtAllPtAll
ImageNet pre-training
Baseline55.238.842.971.545.766.1%
XX57.939.243.972.946.699.7%
X✓58.039.243.972.946.6100.0%
✓✓59.639.444.674.347.1100.0%
COCO pre-training
Baseline64.849.953.773.157.164.6%
XX66.350.554.674.057.999.5%
X✓66.450.554.674.057.9100.0%
✓✓67.250.454.775.157.7100.0%
", + "bbox": [ + 503, + 527, + 890, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Predicting only compatible part classes during training and testing. Evaluated on Pascal-PP [1, 5, 6, 29].", + "bbox": [ + 498, + 702, + 893, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are compatible with the query's object-level class (see Sec. 3.3.2). In Tab. 3, we evaluate the effect of predicting only compatible parts during both training and testing. In addition to the main metrics, we also assess the percentage of predicted objects for which there are no conflicting part-level predictions. The results show that, even when allowing incompatible predictions, TAPPS has significantly fewer object-part conflicts than the baseline. Naturally, when predicting only compatible parts during testing, all object-part conflicts are removed, eliminating the need", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3180", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4a1d14ca77c2e8f4ecdbc539569b3fc146378e9883907f980b571f9c74e12419.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Adaptation layersPartPQPartSQPQ
PtNo PtAllPtAll
-67.349.654.175.257.1
1×FC67.150.354.675.157.6
2-layer MLP67.250.454.775.157.7
3-layer MLP67.150.454.675.157.7
", + "bbox": [ + 107, + 88, + 436, + 179 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/99d8895d4230173b4d5e9379661aeb91a3add568e48ef83e9fd5cc53b5f46476.jpg", + "table_caption": [ + "Table 4. JOPS head architecture (Fig. 2). Evaluated on Pascal-PP [1, 5, 6, 29], with pre-training on COCO panoptic [21]." + ], + "table_footnote": [ + "(a) Pascal-PP [1, 5, 6, 29]." + ], + "table_body": "
MethodmIoUThPQTh
Baseline69.862.0
TAPPS (ours)69.965.6
", + "bbox": [ + 81, + 224, + 269, + 273 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7ae20b9020d85c5150233d6d66fdb87f8123de787ad7c89dd9022eb3204a3372.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Cityscapes-PP [4, 5]." + ], + "table_body": "
MethodmIoUThPQTh
Baseline77.653.8
TAPPS (ours)78.255.6
", + "bbox": [ + 274, + 224, + 462, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for post-processing. However, this does not yield a big improvement in terms of the segmentation quality. When we also apply this during training, we do observe a part segmentation quality $(\\mathrm{PartSQ^{Pt}})$ improvement. This shows that simplifying the part segmentation task during training leads to improved performance, as we hypothesized in Sec. 3.3.2.", + "bbox": [ + 75, + 324, + 468, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "JOPS head architecture. In Tab. 4, we evaluate the impact of using different numbers of adaptation layers in the JOPS head before applying the $N^{\\mathrm{pc}}$ fully-connected (FC) layers to generate the part queries (see Fig. 2). We find that one or two adaptation layers are necessary to generate object- and part-level representations that are sufficiently distinct to perform their respective tasks accurately; adding any more layers does not yield further improvements.", + "bbox": [ + 75, + 415, + 468, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Additional analyses", + "text_level": 1, + "bbox": [ + 76, + 545, + 264, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Instance separability. Tab. 1 showed that TAPPS improves the $\\mathrm{PQ^{\\mathrm{Th}}}$ , i.e., the ability to recognize, segment and classify object instances. To assess if this is due to better object recognition or improved instance separability, we group all thing instances of the same object-level class together and evaluate the instance-agnostic semantic segmentation performance with the mIoU $^{\\mathrm{Th}}$ . If the $\\mathrm{PQ^{\\mathrm{Th}}}$ improvement were due to improved recognition, we expect the mIoU $^{\\mathrm{Th}}$ to improve too. However, the results in Tab. 5 show only a minor mIoU $^{\\mathrm{Th}}$ improvement. This indicates that the $\\mathrm{PQ^{\\mathrm{Th}}}$ gain is not due to better recognition, but mainly results from a better ability to separate instances, showing the benefit of learning parts in an object-instance-aware manner.", + "bbox": [ + 75, + 568, + 468, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance on Pascal-PP-107. To assess the performance of TAPPS in a more complex setting, we evaluate it on Pascal-PP-107, which has 107 part-level classes instead of 57. The results on this dataset, reported in Tab. 6, show once more that TAPPS consistently improves the part segmentation and thing segmentation performance with respect to the baseline, in this case leading to a PartPQ $^{\\text{Pt}}$ improvement of +4.1 or +2.2, depending on the pre-training strategy. This demonstrates that TAPPS is also effective", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/35275ec13180d7b53c7a33e0e6ae9e9a68664ced619a729df79c258605afe8a7.jpg", + "table_caption": [ + "Table 5. Performance for things. ImageNet pre-training [33]." + ], + "table_footnote": [], + "table_body": "
MethodPartPQPartSQPQ
PtNo PtAllPtThStAll
ImageNet pre-training
Baseline45.738.940.659.362.137.645.9
TAPPS (ours)49.839.342.061.765.837.747.2
COCO pre-training
Baseline53.249.850.759.975.248.157.2
TAPPS (ours)55.450.051.462.175.648.157.4
", + "bbox": [ + 504, + 88, + 890, + 218 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/38b54017c86c383961ec61e4699e985a4a3ff386791832887787c119a2f43755.jpg", + "table_caption": [ + "Table 6. 107 part-level categories. Results on the more challenging Pascal-PP-107 dataset [1, 5, 6, 27, 29]." + ], + "table_footnote": [], + "table_body": "
MethodPartPQPartSQPQ
PtNo PtAllPtAll
Dynamic65.150.554.273.157.7
Fixed67.250.454.775.157.7
", + "bbox": [ + 550, + 263, + 839, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Fixed or dynamic part segmentation. Evaluated on Pascal-PP [1, 5, 6, 29], with pre-training on COCO panoptic [21].", + "bbox": [ + 498, + 339, + 890, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "when the PPS task becomes more complex.", + "bbox": [ + 500, + 381, + 787, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fixed vs. dynamic part segmentation. As explained in Sec. 3.3.2, TAPPS uses a fixed fully connected layer for each part class to generate the corresponding 'per-object' part query. Alternatively, it is possible to predict part masks and classes dynamically within each object segment, using a set of dynamic queries like we do for object-level segmentation. In Tab. 7, we compare our fixed part segmentation setting with this dynamic approach, which is explained in more detail in the supplementary material. We find that our fixed approach results in a better $\\mathrm{PartSQ^{Pt}}$ and therefore $\\mathrm{PartPQ^{Pt}}$ performance. We hypothesize that the dynamic approach performs worse because it makes the part segmentation task unnecessarily complex, as the network needs to learn to assign segments to queries dynamically, and additionally predict a class label.", + "bbox": [ + 496, + 397, + 890, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 636, + 617, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With experiments, we have shown that TAPPS considerably outperforms methods that predict objects and parts separately, by improving the object instance separability, part segmentation quality, and object-part compatibility. Importantly, these improvements can be attributed to the fact that TAPPS is directly optimized for the PPS task, using a set of shared queries to jointly predict objects and corresponding parts. With our promising findings, we hope to inspire future research towards even more complete scene understanding, e.g., image segmentation at even more abstraction levels, potentially with more flexible class hierarchies.", + "bbox": [ + 496, + 661, + 890, + 827 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by Eindhoven Engine, NXP Semiconductors, and Brainport Eindhoven. This work made use of the Dutch national e-infrastructure with the support of the SURF Cooperative using grant no. EINF-5302, which is financed by the Dutch Research Council (NWO).", + "bbox": [ + 496, + 830, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3181", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect What You Can: Detecting and Representing Objects using Holistic Models and Body Parts. In CVPR, 2014. 5, 6, 7, 8", + "[2] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-Pixel Classification is Not All You Need for Semantic Segmentation. In NeurIPS, 2021. 3", + "[3] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. In CVPR, 2022. 3, 5, 6", + "[4] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes Dataset for Semantic Urban Scene Understanding. In CVPR, 2016. 5, 6, 7, 8", + "[5] Daan de Geus, Panagiotis Meletis, Chenyang Lu, Xiaoxiao Wen, and Gijs Dubbelman. Part-aware Panoptic Segmentation. In CVPR, 2021. 1, 2, 3, 4, 5, 6, 7, 8", + "[6] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes (VOC) Challenge. *IJCV*, 88(2):303–338, 2010. 5, 6, 7, 8", + "[7] Ke Gong, Xiaodan Liang, Yicheng Li, Yimin Chen, Ming Yang, and Liang Lin. Instance-level human parsing via part grouping network. In ECCV, 2018. 3", + "[8] Ke Gong, Yiming Gao, Xiaodan Liang, Xiaohui Shen, Meng Wang, and Liang Lin. Graphonomy: Universal human parsing via graph transfer learning. In CVPR, 2019. 3", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 3, 7", + "[10] Sravan Kumar Jagadeesh, René Schuster, and Didier Stricker. Multi-task Fusion for Efficient Panoptic-Part Segmentation. In ICPRAM, 2022. 2, 7", + "[11] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. In CVPR, 2023. 3", + "[12] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollar. Panoptic Feature Pyramid Networks. In CVPR, 2019. 3", + "[13] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic Segmentation. In CVPR, 2019. 1, 2, 3, 6", + "[14] Jianshu Li, Jian Zhao, Yunpeng Chen, Sujoy Roy, Shuicheng Yan, Jiashi Feng, and Terence Sim. Multi-human parsing machines. In ACM MM, 2018. 3", + "[15] Liulei Li, Tianfei Zhou, Wenguan Wang, Jianwu Li, and Yi Yang. Deep Hierarchical Semantic Segmentation. In CVPR, 2022.", + "[16] Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. Self-Correction for Human Parsing. IEEE TPAMI, 44(6):3260-3271, 2022. 3", + "[17] Qizhu Li, Anurag Arnab, and Philip HS Torr. Holistic, instance-level human parsing. In BMVC, 2017. 3" + ], + "bbox": [ + 78, + 114, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Xiangtai Li, Shilin Xu, Yibo Yang, Guangliang Cheng, Yunhai Tong, and Dacheng Tao. Panoptic-PartFormer: Learning a Unified Model for Panoptic Part Segmentation. In ECCV, 2022. 1, 2, 3, 5, 6, 7", + "[19] Xiangtai Li, Shilin Xu, Yibo Yang, Haobo Yuan, Guangliang Cheng, Yunhai Tong, Zhouchen Lin, and Dacheng Tao. PanopticPartFormer++: A Unified and Decoupled View for Panoptic Part Segmentation. arXiv preprint arXiv:2301.00954, 2023. 1, 2, 3, 5, 7", + "[20] Xiaodan Liang, Ke Gong, Xiaohui Shen, and Liang Lin. Look into Person: Joint Body Parsing & Pose Estimation Network and A New Benchmark. IEEE TPAMI, 41(4):871-885, 2018. 3", + "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 6, 7, 8", + "[22] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature Pyramid Networks for Object Detection. In CVPR, 2017. 3", + "[23] Qing Liu, Adam Kortylewski, Zhishuai Zhang, Zizhang Li, Mengqi Guo, Qihao Liu, Xiaoding Yuan, Jiteng Mu, Weichao Qiu, and Alan Yuille. Learning Part Segmentation Through Unsupervised Domain Adaptation From Synthetic Vehicles. In CVPR, 2022. 3", + "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. In ICCV, 2021. 3, 7", + "[25] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s. In CVPR, 2022. 7", + "[26] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICLR, 2019. 6", + "[27] Umberto Michieli, Edoardo Borsato, Luca Rossi, and Pietro Zanuttigh. GMNet: Graph Matching Network for Large Scale Part Semantic Segmentation in the Wild. In ECCV, 2020. 3, 5, 8", + "[28] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentatio. In 3DV, 2016. 5", + "[29] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The Role of Context for Object Detection and Semantic Segmentation in the Wild. In CVPR, 2014. 5, 6, 7, 8", + "[30] Tai-Yu Pan, Qing Liu, Wei-Lun Chao, and Brian Price. Towards Open-World Segmentation of Parts. In CVPR, 2023. 3", + "[31] Lu Qi, Jason Kuen, Weidong Guo, Jiumiang Gu, Zhe Lin, Bo Du, Yu Xu, and Ming-Hsuan Yang. AIMS: All-Inclusive Multi-Level Segmentation for Anything. In NeurIPS, 2023. 3", + "[32] Tao Ruan, Ting Liu, Zilong Huang, Yunchao Wei, Shikui Wei, and Yao Zhao. Devil in the details: Towards accurate single and multiple human parsing. In AAAI, 2019. 3" + ], + "bbox": [ + 501, + 92, + 892, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3182", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 115(3):211-252, 2015. 7, 8", + "[34] Rishubh Singh, Pranav Gupta, Pradeep Shenoy, and Ravikiran Sarvadevabhatla. FLOAT: Factorized Learning of Object Attributes for Improved Multi-Object Multi-Part Scene Parsing. In CVPR, 2022. 3", + "[35] Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation. In ICCV, 2023. 3", + "[36] Mingxing Tan and Quoc Le. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. In ICML, 2019. 7", + "[37] Xin Tan, Jiachen Xu, Zhou Ye, Jinkun Hao, and Lizhuang Ma. Confident Semantic Ranking Loss for Part Parsing. In ICME, 2021. 3", + "[38] Chufeng Tang, Hang Chen, Xiao Li, Jianmin Li, Zhaoxiang Zhang, and Xiaolin Hu. Look Closer To Segment Better: Boundary Patch Refinement for Instance Segmentation. In CVPR, 2021. 7", + "[39] Chufeng Tang, Lingxi Xie, Xiaopeng Zhang, Xiaolin Hu, and Qi Tian. Visual Recognition by Request. In CVPR, 2023. 2", + "[40] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional Convolutions for Instance Segmentation. In ECCV, 2020. 7", + "[41] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: End-to-End Panoptic Segmentation With Mask Transformers. In CVPR, 2021. 3", + "[42] Wenguan Wang, Hailong Zhu, Jifeng Dai, Yanwei Pang, Jianbing Shen, and Ling Shao. Hierarchical human parsing with typed part-relation reasoning. In CVPR, 2020. 3", + "[43] Meng Wei, Xiaoyu Yue, Wenwei Zhang, Shu Kong, Xihui Liu, and Jiangmiao Pang. OV-PARTS: Towards Open-Vocabulary Part Segmentation. In NeurIPS, 2023. 3", + "[44] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, 2018. 3", + "[45] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers. In NeurIPS, 2021. 7", + "[46] Lu Yang, Qing Song, Zhihui Wang, and Ming Jiang. Parsing R-CNN for Instance-Level Human Analysis. In CVPR, 2019. 3", + "[47] Lu Yang, Qing Song, Zhihui Wang, Mengjie Hu, Chun Liu, Xueshi Xin, Wenhe Jia, and Songcen Xu. Renovating Parsing R-CNN for Accurate Multiple Human Parsing. In ECCV, 2020.", + "[48] Lu Yang, Qing Song, Zhihui Wang, Mengjie Hu, and Chun Liu. Hier R-CNN: Instance-Level Human Parts Detection and A New Benchmark. IEEE TIP, 30:39-54, 2021.", + "[49] Sanyi Zhang, Xiaochun Cao, Guo-Jun Qi, Zhanjie Song, and Jie Zhou. AIParsing: Anchor-Free Instance-Level Human Parsing. IEEE TIP, 31:5599-5612, 2022. 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-Net: Towards Unified Image Segmentation. In NeurIPS, 2021. 3", + "[51] Jian Zhao, Jianshu Li, Yu Cheng, Terence Sim, Shuicheng Yan, and Jiashi Feng. Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for multi-human parsing. In ACM MM, 2018. 3", + "[52] Yifan Zhao, Jia Li, Yu Zhang, and Yonghong Tian. Multi-Class Part Parsing With Joint Boundary-Semantic Awareness. In ICCV, 2019. 3" + ], + "bbox": [ + 501, + 90, + 892, + 232 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3183", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/9dbd5f9f-08af-4d23-b3b6-162e326d318c_model.json b/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/9dbd5f9f-08af-4d23-b3b6-162e326d318c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d1e331151cc090a65384f73a64f9f6acbc5ad40a --- /dev/null +++ b/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/9dbd5f9f-08af-4d23-b3b6-162e326d318c_model.json @@ -0,0 +1,2068 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.243, + 0.131, + 0.73, + 0.177 + ], + "angle": 0, + "content": "Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.204, + 0.632, + 0.221 + ], + "angle": 0, + "content": "Daan de Geus Gijs Dubbelman" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.223, + 0.631, + 0.24 + ], + "angle": 0, + "content": "Eindhoven University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.242, + 0.624, + 0.257 + ], + "angle": 0, + "content": "{d.c.d.geus, g.dubbelman}@tue.nl" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.474, + 0.645 + ], + "angle": 0, + "content": "Part-aware panoptic segmentation (PPS) requires (a) that each foreground object and background region in an image is segmented and classified, and (b) that all parts within foreground objects are segmented, classified and linked to their parent object. Existing methods approach PPS by separately conducting object-level and part-level segmentation. However, their part-level predictions are not linked to individual parent objects. Therefore, their learning objective is not aligned with the PPS task objective, which harms the PPS performance. To solve this, and make more accurate PPS predictions, we propose Task-Aligned Part-aware Panoptic Segmentation (TAPPS). This method uses a set of shared queries to jointly predict (a) object-level segments, and (b) the part-level segments within those same objects. As a result, TAPPS learns to predict part-level segments that are linked to individual parent objects, aligning the learning objective with the task objective, and allowing TAPPS to leverage joint object-part representations. With experiments, we show that TAPPS considerably outperforms methods that predict objects and parts separately, and achieves new state-of-the-art PPS results." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.677, + 0.21, + 0.693 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To fully understand what is depicted in an image, it is important to consider concepts at different levels of abstraction. For rich scene understanding, we should not only recognize foreground objects (e.g., car, human) and background regions (e.g., sky, ocean), but also simultaneously identify the parts that constitute the objects (e.g., car-wheel, human-arm). In a step towards such comprehensive scene understanding, De Geus et al. introduced part-aware panoptic segmentation (PPS) [5]. The objective of this computer vision task is (1) to output a segmentation mask and class label for all thing objects and stuff regions in an image like for panoptic segmentation [13] – we call these object-level segments – and (2) to simultaneously segment and classify" + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.29, + 0.72, + 0.302 + ], + "angle": 0, + "content": "(a) Separate object and part segmentation" + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.303, + 0.745, + 0.314 + ], + "angle": 0, + "content": "X Learning objective not aligned with task objective" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.316, + 0.887, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.447, + 0.733, + 0.459 + ], + "angle": 0, + "content": "(b) Joint object and part segmentation (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.46, + 0.727, + 0.472 + ], + "angle": 0, + "content": "\\(\\checkmark\\) Learning objective aligned with task objective" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.474, + 0.889, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.684 + ], + "angle": 0, + "content": "Figure 1. Task-aligned part-aware panoptic segmentation. (a) Existing works separately predict object-level segments and object-instance-unaware part-level segments. (b) In this work, we predict objects and parts jointly, using a set of shared queries. This allows our method to predict parts within individual object segments, aligning its learning objective with the PPS task objective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.891, + 0.749 + ], + "angle": 0, + "content": "all parts within each identified object. These are called part-level segments and should be explicitly linked to an object-level segment, establishing a part-whole relation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Current state-of-the-art methods [18, 19] address PPS by using two different sets of learnable queries to separately predict object-level and part-level segments, see Fig. 1a. While these methods outperform earlier baselines [5], they have one main limitation: their learning objective is not aligned with the task objective. Where the PPS task objective is to predict parts within each individual object-level segment, these existing works conduct part-level semantic segmentation, predicting part-level masks that cover multiple objects (see Fig. 1a). In other words, these networks" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3174" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.364 + ], + "angle": 0, + "content": "are not optimized for the PPS task, but instead solve the surrogate subtasks of object-level panoptic segmentation and part-level semantic segmentation. As a result, to assign parts to individual objects, these methods require post-processing. Additionally, we hypothesize that there are several other negative consequences: (1) These networks learn a conflicting feature representation. They learn that object-level thing instances should be separated, but also that the parts of these object-level instances should be grouped together. We expect that this harms the ability of the networks to separate instances. (2) Predicting objects and parts separately may cause incompatible predictions (e.g., a car-window part with a bicycle object), which makes it unclear which prediction to trust, and requires further post-processing. (3) The networks encode information about objects and their parts separately, whereas this information is potentially complementary. See Sec. 3.2.2 for more details on these limitations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.473, + 0.548 + ], + "angle": 0, + "content": "In this work, we aim to design a simple network for PPS that overcomes these limitations and thereby makes more accurate PPS predictions. To achieve this, we propose the Task-Aligned Part-aware Panoptic Segmentation (TAPPS) method. Instead of using separate queries for objects and parts, TAPPS uses one set of shared queries to jointly represent objects and the parts they contain. Specifically, each of these queries learns to represent at most one object-level segment, for which it predicts (1) a segmentation mask and object-level class, and (2) the segmentation masks and classes for all part-level segments within this object. This is visualized in Fig. 1b and explained in Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.473, + 0.824 + ], + "angle": 0, + "content": "With this approach, TAPPS explicitly predicts part-level segments per individual object. Therefore, the network's learning objective is now aligned with the PPS task objective, and TAPPS is directly optimized for the PPS task. As a result, both object- and part-level segmentation are learned in an 'object-instance-aware' manner. This removes the conflict in the learned feature representations, allowing for better instance separability. Additionally, we hypothesize that predicting objects and parts from the same query reduces incompatibilities between object and part predictions, as they are made using the same information. Moreover, we use TAPPS to go even further, and explicitly only predict part segments compatible with predicted object segments. This enforces full object-part compatibility and simplifies the part segmentation task. Finally, the network can now encode complementary object-level and part-level information in a shared query, giving it a richer representation for making object- and part-level segmentation predictions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.903 + ], + "angle": 0, + "content": "With experiments, detailed in Sec. 4, we show that TAPPS outperforms the baseline with separate object and part queries in multiple aspects: (1) The part segmentation quality within identified objects is significantly better, which shows the positive impact of using joint object-part" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "representations and simplifying the part segmentation task. (2) The object instance segmentation performance is considerably improved, demonstrating the benefits of learning parts in an object-instance-aware manner. Together, this yields a large overall improvement with respect to the baseline, and causes TAPPS to considerably outperform existing works across different benchmarks, achieving new state-of-the-art results. See Sec. 5 for more extensive results." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.215, + 0.869, + 0.23 + ], + "angle": 0, + "content": "To summarize, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.232, + 0.892, + 0.292 + ], + "angle": 0, + "content": "- We propose TAPPS, a simple approach for PPS that aligns the learning objective with the task objective, facilitating object instance separability and enabling joint object-part representations for more accurate PPS predictions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.292, + 0.892, + 0.352 + ], + "angle": 0, + "content": "- We use the shared object-part queries to constrain TAPPS to only predict part segments that are compatible with the object class, enforcing object-part compatibility and simplifying the part segmentation task." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.352, + 0.892, + 0.382 + ], + "angle": 0, + "content": "- We experimentally show the effectiveness of TAPPS across a range of datasets and network configurations." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.232, + 0.892, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.385, + 0.892, + 0.415 + ], + "angle": 0, + "content": "The code for TAPPS is made publicly available through https://tue-mps.github.io/tapps/." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.434, + 0.638, + 0.45 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Part-aware panoptic segmentation. Part-aware panoptic segmentation [5] is an image segmentation task introduced for scene understanding at multiple abstraction levels. It extends panoptic segmentation [13] by requiring (a) object-level panoptic segmentation, and (b) part-level segmentation within object-level segments. Most existing works do not predict parts per object-level segment, but instead make separate predictions for panoptic segmentation and the surrogate task of part-level semantic segmentation - or part segmentation [10, 18, 19]. JPPF [10] proposes a single-network approach with a shared encoder followed by separate heads for semantic, instance, and part segmentation. To merge the predictions by these heads to the PPS format, JPPF proposes a new rule-based fusion strategy that outperforms the originally introduced merging strategy for PPS [5]. Panoptic-PartFormer and its extension Panoptic-PartFormer++ [18, 19] tackle PPS with a Transformer-based approach that independently makes panoptic segmentation and part segmentation predictions using separate sets of learnable queries for thing, stuff and part segments. Alternatively, ViRReq [39] is a paradigm in which PPS is predicted in a cascading fashion, by first segmenting objects and then segmenting the parts within these objects 'by request', but it requires multiple networks to achieve this." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In contrast to these existing approaches, we propose a method that jointly predicts object-level segments and the part-level segments within these objects. This aligns the learning objective with the task objective, and yields improved PPS performance. For more details, see Sec. 3." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3175" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "Part-level image segmentation. Part-level segmentation is also widely studied beyond PPS. Most works address part segmentation, i.e., semantic segmentation for parts [7, 8, 14-17, 20, 23, 27, 32, 34, 35, 37, 42, 43, 51, 52]. For more complete scene understanding, other works also take into account object instances, like PPS does, with instance-aware part segmentation [16, 30, 32, 46-49]. However, these works do not consider background stuff classes, which PPS does consider. Alternatively, UPerNet [44] separately predicts part-level and object-level semantic segmentation, including background classes, but without any instance awareness. Another work has extended PPS by also predicting the relations between objects [31], training a model on annotations from different datasets. In this work, we address the PPS task as introduced by De Geus et al. [5], and we compare our approach to state-of-the-art alternatives." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.351, + 0.17, + 0.366 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.377, + 0.26, + 0.392 + ], + "angle": 0, + "content": "3.1. PPS task definition" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.402, + 0.473, + 0.679 + ], + "angle": 0, + "content": "Part-aware panoptic segmentation (PPS) [5] requires consistent image segmentation across two abstraction levels: the object level and part level. For object-level segmentation, following panoptic segmentation [13], an image should be divided into \\(N\\) object-level segments, where \\(N\\) varies per image. Each segment \\(s_i\\) consists of a binary mask \\(\\mathbf{M}_i^{\\mathrm{obj}}\\) and an object class label \\(c_i^{\\mathrm{obj}}\\). Thing classes require a segment per object instance, and stuff classes require a single segment per class. Next, per segment \\(s_i\\), PPS requires a set of \\(K_i\\) part-level segments. Each part-level segment \\(s_{i,j}^{\\mathrm{pt}}\\) consists of a binary part-level mask \\(\\mathbf{M}_{i,j}^{\\mathrm{pt}}\\) and a part-level class \\(c_{i,j}^{\\mathrm{pt}}\\) that are compatible with the object-level mask and class. Specifically, the part-level mask must be a subset of the object-level mask, and the part-level class should be one of the part-level classes defined for the object-level class. For instance, if \\(c_i^{\\mathrm{obj}}\\) is car, then \\(c_{i,j}^{\\mathrm{pt}}\\) cannot be human-head but it can be car-window. For all parts, a single part-level segment is required for each part class within an object." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.681, + 0.472, + 0.789 + ], + "angle": 0, + "content": "To summarize, each image should be divided into a set of segments \\(S = \\{s_i\\}_{i=1}^N = \\{(\\mathbf{M}_i^{\\mathrm{obj}}, c_i^{\\mathrm{obj}}, \\mathcal{P}_i)\\}_{i=1}^N\\), where \\(\\mathcal{P}_i = \\{s_{i,j}^{\\mathrm{pt}}\\}_{j=1}^{K_i} = \\{(\\mathbf{M}_{i,j}^{\\mathrm{pt}}, c_{i,j}^{\\mathrm{pt}})\\}_{j=1}^{K_i}\\) is a set of part-level segments compatible with the object-level segment in \\(s_i\\). In practice, part-level classes are only defined for some object-level classes. When segment \\(s_i\\) has an object class that does not have parts, then \\(K_i = 0\\) and \\(s_i = (\\mathbf{M}_i^{\\mathrm{obj}}, c_i^{\\mathrm{obj}})\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.801, + 0.218, + 0.815 + ], + "angle": 0, + "content": "3.2. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Like existing state-of-the-art PPS approaches [18, 19], our work uses the powerful mask classification metaarchitecture that is the foundation of many state-of-the-art segmentation models [2, 3, 11, 41, 50]. Here, we first explain this mask classification paradigm in more detail." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "Then, we explain how existing works use it for PPS, and what the limitations are of these approaches." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.14, + 0.768, + 0.155 + ], + "angle": 0, + "content": "3.2.1 Mask classification framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.164, + 0.895, + 0.589 + ], + "angle": 0, + "content": "The concept of mask classification is to predict a set of \\( N^{\\mathrm{q}} \\) object-level segments, i.e., a set of \\( N^{\\mathrm{q}} \\) pixel-level masks \\( \\hat{\\mathbf{M}}^{\\mathrm{obj}} \\) and corresponding class labels \\( \\hat{c}^{\\mathrm{obj}} \\). To make these predictions, mask classification networks output two components: high-resolution features \\( \\mathbf{F} \\in \\mathbb{R}^{E \\times H \\times W} \\) and queries \\( \\mathbf{Q} \\in \\mathbb{R}^{N^{\\mathrm{q}} \\times E} \\), where \\( H \\) and \\( W \\) are the height and width of the features, and \\( E \\) is the feature and query dimensionality. Each of the \\( N^{\\mathrm{q}} \\) queries is used to predict the class label and segmentation mask of at most one object-level segment. The class predictions are a function of only the queries, and the mask predictions are a function of both the queries \\( \\mathbf{Q} \\) and the high-resolution features \\( \\mathbf{F} \\). To output these high-resolution features, the input image is first fed into a backbone to extract multi-scale features, e.g., ResNet [9] or Swin [24]. Subsequently, these features are further refined and upsampled to high-resolution features by a pixel decoder, e.g., Semantic FPN [12, 22]. In the other part of the network, the queries \\( \\mathbf{Q} \\) are generated by processing learnable queries \\( \\mathbf{Q}^{\\mathrm{o}} \\) using a transformer decoder, which applies self-attention across queries and cross-attention with image features. Through bipartite matching between \\( N^{\\mathrm{q}} \\) predicted and \\( N \\) ground-truth segments, each query is assigned to at most one ground-truth segment. As \\( N^{\\mathrm{q}} \\) is not always equal to \\( N \\), some queries may not be assigned to a ground-truth segment. If so, they do not receive any supervision for segmentation, and will learn a 'no object' class. In this work, we build upon the Mask2Former [3] instantiation of this meta-architecture." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.744, + 0.62 + ], + "angle": 0, + "content": "3.2.2 Mask classification for PPS" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.689 + ], + "angle": 0, + "content": "With the aforementioned meta-architecture, we can conduct object-level segmentation, and thereby solve one aspect of the PPS task definition. To solve the full task, the object-level segments should also be further segmented into parts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Current state-of-the-art approaches [18, 19] achieve this by introducing an additional set of queries, the part-level queries depicted in Fig. 1a. Each of these part-level queries learns to represent a part-level segment. However, in contrast to the PPS task definition, these part-level segments are not explicitly linked to an object-level segment. Instead, these part-level segments represent an entire part-level class, i.e., their masks contain all pixels belonging to one part-level class across multiple object segments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This means that the learning objective of these networks is not aligned with the PPS task objective. As a result, they are not directly optimized for PPS, but instead for two surrogate tasks: object-level panoptic segmentation and part-level semantic segmentation. In addition to necessitating" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3176" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.472, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.087, + 0.895, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.283, + 0.893, + 0.34 + ], + "angle": 0, + "content": "Figure 2. Network architecture. Left: The overall TAPPS architecture. A set of learnable queries and features from a backbone are fed into a pixel decoder and transformer decoder to generate high-resolution features and processed queries. Right: These queries and features are fed into the JOPS head, which predicts for each shared query (a) an object-level class, (b) an object-level segmentation mask, and (c) a set of part-level masks for the part-level classes compatible with the object-level class. Operator \\(\\otimes\\) denotes a matrix multiplication." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.473, + 0.683 + ], + "angle": 0, + "content": "post-processing, we hypothesize that this has several negative consequences: (1) By learning parts in an 'object-instance-unaware' manner, the network learns conflicting representations in the features, F. On the one hand, different thing instances belong to different object-level segments, requiring distinct features. On the other hand, the parts of these instances belong to the same part-level segment, requiring similar features. We hypothesize that this conflict harms the ability of the network to learn to separate object-level thing instances. (2) Separately predicting objects and parts in this way may also cause incompatible predictions. These incompatibilities make it unclear for downstream processes which prediction to trust. To obtain predictions compliant with the PPS definition, rule-based post-processing is required, which discards incompatible predictions [5]. This is undesired because the discarded information is potentially correct. (3) By representing object- and part-level information with separate queries, the network cannot encode complementary information about an individual object and its parts, which it could use to make more informed predictions at these abstraction levels." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.249, + 0.712 + ], + "angle": 0, + "content": "3.3. Task-aligned PPS" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "To overcome the limitations of state-of-the-art methods and achieve a better PPS performance, the objective of this work is to design a simple PPS approach that uses the powerful mask classification framework, but that aligns its learning objective with the PPS task objective, and explicitly predicts part-level segments within individual objects. To achieve this, we should (a) generate a unique query for each object-part combination, i.e., for each part-level segment within an object-level segment, and (b) explicitly link these part-level segments to their parent objects, as PPS requires. To accomplish this, we propose Task-Aligned Part-aware Panoptic Segmentation (TAPPS), a method that jointly represents" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.893, + 0.442 + ], + "angle": 0, + "content": "object-level segments and their parts with a set of shared queries, and generates 'per-object' part queries from each shared query. Each of these queries can then be used to predict one part-level segment within an object-level segment, and is automatically linked to its parent object segment." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.467, + 0.696, + 0.48 + ], + "angle": 0, + "content": "3.3.1 Overall architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.719 + ], + "angle": 0, + "content": "In Fig. 2, we visualize the network architecture of TAPPS. First, we initialize a set of \\( N^{\\mathbf{q}} \\) learnable queries, \\( \\mathbf{Q}^0 \\in \\mathbb{R}^{N^{\\mathbf{q}} \\times E} \\). Each of these queries learns to represent one object-level segment, but also the part-level segments belonging to this object (e.g., car3 and also car3-wheels, car3-windows, etc.). Following the mask classification framework described in Sec. 3.2, these learnable queries \\( \\mathbf{Q}^0 \\) and the features from a backbone are fed into a decoder, which generates a set of processed queries \\( \\mathbf{Q} \\) and high-resolution features \\( \\mathbf{F} \\). Subsequently, these processed queries and features enter the Joint Object and Part Segmentation (JOPS) head. For each query \\( \\mathbf{Q}_i \\in \\mathbb{R}^E \\), this head predicts (a) an object-level class, (b) an object-level segmentation mask, and (c) a set of part-level segmentation masks and classes for the parts within this object-level segment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "With this approach, we solve the limitations of existing works. Following the PPS task definition, we predict part-level segments per individual object-level segment. (1) As a result, we learn object-instance-aware representations for both parts and objects, improving the ability of the network to separate instances. (2) By predicting objects and parts jointly, the compatibility between both sets of predictions greatly improves. Moreover, in the JOPS head, we ensure that we only predict compatible parts, ensuring full object-part compatibility (see Sec. 3.3.2). (3) The joint object-part representations allow the network to fully leverage the complementary information from both abstraction levels." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3177" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.21, + 0.105 + ], + "angle": 0, + "content": "3.3.2 JOPS head" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.284 + ], + "angle": 0, + "content": "Within the JOPS head, we predict an object-level segment and compatible part-level segments for each shared query. For the object-level predictions, we follow Mask2Former [3]: (1) For the object-level class, we feed each processed query \\(\\mathbf{Q}_i\\) through a single fully connected layer to predict a score for each possible class, and obtain the predicted class \\(\\hat{c}_i^{\\mathrm{obj}}\\) by picking the highest-scoring class. (2) For the object-level mask, we feed the query \\(\\mathbf{Q}_i\\) through a 3-layer MLP and generate a mask by taking the product of the resulting mask queries \\(\\mathbf{Q}_i^{\\mathrm{m}}\\) and the features \\(\\mathbf{F}\\), and applying a sigmoid activation, yielding \\(\\hat{\\mathbf{M}}_i^{\\mathrm{obj}} \\in [0,1]^{H \\times W}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.471, + 0.725 + ], + "angle": 0, + "content": "To predict part-level segments using the same features \\(\\mathbf{F}\\), we need to generate queries for each part class within an object-level segment. To do so, we first apply an MLP to each query \\(\\mathbf{Q}_i\\) to adapt it for part-level segmentation, and then apply \\(N^{\\mathrm{pc}}\\) different fully-connected layers, where \\(N^{\\mathrm{pc}}\\) is the total number of part-level classes in the dataset. This results in \\(\\mathbf{Q}_i^{\\mathrm{pt}} \\in \\mathbb{R}^{N^{\\mathrm{pc}} \\times E}\\), a set of 'per-object' part-level queries, where each query always corresponds to a fixed, predetermined part class. We could then take the product of \\(\\mathbf{Q}_i^{\\mathrm{pt}}\\) and the features \\(\\mathbf{F}\\) to generate a segmentation mask for all part-level classes. However, we already have an object-level class prediction for each query, and we know that only a subset of all \\(N^{\\mathrm{pc}}\\) part classes is compatible with a certain object class. Therefore, we propose to only predict and supervise part-level masks for those compatible part-level classes. This simplifies the part segmentation task that the network needs to learn, as it no longer has to learn to predict 'empty' segmentation masks for incompatible part classes. Concretely, as visualized in Fig. 2, we identify the object-level class \\(\\hat{c}_i^{\\mathrm{obj}}\\) for each query \\(\\mathbf{Q}_i\\), and keep only the part-level queries \\(\\mathbf{Q}_i^{\\mathrm{pt,c}} \\in \\mathbb{R}^{N^{\\mathrm{c}} \\times E}\\) that correspond to the \\(N^{\\mathrm{c}}\\) part classes that are compatible with the object-level class. Then, we compute the product of these remaining queries \\(\\mathbf{Q}_i^{\\mathrm{pt,c}}\\) and the features \\(\\mathbf{F}\\), and apply a sigmoid activation to generate compatible part segmentation masks \\(\\hat{\\mathbf{M}}_i^{\\mathrm{pt}} \\in [0,1]^{N^{\\mathrm{c}} \\times H \\times W}\\). As each per-object part query corresponds to a fixed part class, we also know the part classes \\(\\hat{c}_i^{\\mathrm{pt}}\\). Note that the number \\(N^{\\mathrm{c}}\\) depends on the object-level class and will therefore vary per query \\(\\mathbf{Q}_i\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.194, + 0.756 + ], + "angle": 0, + "content": "3.3.3 Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.825 + ], + "angle": 0, + "content": "To assign each query to at most one ground-truth object-level segment with corresponding part-level segments during training, we apply bipartite matching based on the predicted and ground-truth object-level classes and masks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.9 + ], + "angle": 0, + "content": "To supervise the object-level segments, we use the cross-entropy loss for the classes, and both the Dice [28] and cross-entropy loss for the segmentation masks. Together, these losses form the object-level loss \\( L_{\\mathrm{obj}} \\). For the part-level segmentation masks, we also use the Dice and cross" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.135 + ], + "angle": 0, + "content": "entropy loss, together forming part-level loss \\( L_{\\mathrm{pt}} \\). We balance the losses with weights \\( \\lambda_{\\mathrm{obj}} \\) and \\( \\lambda_{\\mathrm{pt}} \\) to calculate the total loss" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.137, + 0.892, + 0.154 + ], + "angle": 0, + "content": "\\[\nL = \\lambda_ {\\mathrm {o b j}} L _ {\\mathrm {o b j}} + \\lambda_ {\\mathrm {p t}} L _ {\\mathrm {p t}}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.16, + 0.892, + 0.19 + ], + "angle": 0, + "content": "We find that using \\(\\lambda_{\\mathrm{obj}} = \\lambda_{\\mathrm{pt}} = 1\\) provides the best balance; see the supplementary material for more details." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.205, + 0.691, + 0.222 + ], + "angle": 0, + "content": "4. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.23, + 0.892, + 0.244 + ], + "angle": 0, + "content": "Datasets. We use the two PPS benchmarks for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.892, + 0.35 + ], + "angle": 0, + "content": "Cityscapes Panoptic Parts (Cityscapes-PP) [5] extends the original Cityscapes dataset [4] with part-level labels. It consists of street scene images from several cities. It has labels for 19 object classes (11 stuff; 8 thing). Part classes are defined and labeled for 5 object-level thing classes; there are 23 part classes in total. We train on the train split (2975 images), and evaluate on the val split (500 images)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Pascal Panoptic Parts (Pascal-PP) [5], which combines existing labels for Pascal VOC [1, 6, 29], consists of a wide range of scenes and classes. There are 59 object-level classes (39 stuff; 20 thing), and part-level classes are defined for 15 thing classes. Unless otherwise indicated, we use the default part class definition with 57 part classes in total [5]. To evaluate a more challenging setting, we additionally evaluate on Pascal-PP-107, which uses the 107 non-background classes from the Pascal-Part-108 definition introduced by Michieli et al. [27] for part segmentation. For both class definitions, we train on the training split (4998 images), and evaluate on the validation split (5105 images)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.673 + ], + "angle": 0, + "content": "Baseline. Our baseline is a version of TAPPS that uses the same network architecture, but uses a separate set of 100 additional queries for part-level segmentation (see also Fig. 1a). Comparing the results of the baseline in Tab. 1 to existing methods in Tab. 2, we find that our baseline outperforms state-of-the-art approaches that also use separate part-level queries [18, 19], indicating that it is a strong baseline. See the supplementary material for more details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.692, + 0.892, + 0.798 + ], + "angle": 0, + "content": "Evaluation metrics. The part-aware panoptic segmentation performance is evaluated using the default Part-aware Panoptic Quality (PartPQ) metric [5]. It captures both the ability to recognize and segment object-level segments (i.e., stuff regions and thing instances), and the ability to further segment the identified object-level segments into part-level masks. The PartPQ per object-level class \\( c \\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.808, + 0.892, + 0.846 + ], + "angle": 0, + "content": "\\[\n\\operatorname {P a r t P Q} _ {c} = \\frac {\\sum_ {(p , g) \\in T P _ {c}} \\operatorname {I o U} _ {\\mathrm {p}} (p , g)}{| T P _ {c} | + \\frac {1}{2} | F P _ {c} | + \\frac {1}{2} | F N _ {c} |}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( TP_{c}, FP_{c} \\), and \\( FN_{c} \\) are the sets of true positive, false positive and false negative segments, respectively. A pair of predicted and ground-truth segments \\( (p, g) \\) of the same" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3178" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.478, + 0.221 + ], + "angle": 0, + "content": "
MethodPartPQPartSQPQ
PtNo ptAllPtThStAll
ImageNet pre-training
Baseline55.238.842.971.562.037.345.7
TAPPS (ours)59.639.444.674.365.637.647.1
COCO pre-training
Baseline64.849.953.773.174.848.157.1
TAPPS (ours)67.250.454.775.175.748.557.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.178, + 0.223, + 0.38, + 0.236 + ], + "angle": 0, + "content": "(a) Pascal-PP validation [1, 5, 6, 29]." + }, + { + "type": "table", + "bbox": [ + 0.49, + 0.089, + 0.887, + 0.221 + ], + "angle": 0, + "content": "
MethodPartPQPartSQPQ
PtNo ptAllPtThStAll
ImageNet pre-training
Baseline46.662.658.466.153.867.161.5
TAPPS (ours)48.763.159.366.855.667.362.4
COCO pre-training
Baseline48.264.860.466.756.169.063.6
TAPPS (ours)48.965.761.366.957.269.764.4
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.61, + 0.223, + 0.765, + 0.236 + ], + "angle": 0, + "content": "(b) Cityscapes-PP val [4, 5]." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.242, + 0.892, + 0.271 + ], + "angle": 0, + "content": "Table 1. Main results. We compare TAPPS to a strong baseline that uses separate sets of queries to predict object- and part-level segments, instead of predicting object- and part-level segments jointly like TAPPS (see Sec. 4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.469, + 0.447 + ], + "angle": 0, + "content": "class \\( c \\) is part of \\( TP_{c} \\) if the Intersection-over-Union (IoU) of their object-level masks is larger than 0.5. If a ground-truth segment is not identified, it is part of \\( FN_{c} \\); if a prediction is incorrect, it is part of \\( FP_{c} \\). The IoUp term captures the segmentation performance within identified object-level segments. For object-level classes for which part classes are defined, it calculates the part-level mIoU. Otherwise, it uses the object-level IoU. We report the mean PartPQ over all classes, but also separately for object-level classes with parts (PartPQpt) and without parts (PartPQNoPt)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.47, + 0.539 + ], + "angle": 0, + "content": "To individually assess the ability of methods to conduct part-level segmentation within identified objects, we report the Part Segmentation Quality for object-level classes that have parts, \\(\\mathrm{PartSQ^{Pt}}\\). Following De Geus et al. [5], per object-level class \\(c\\), the \\(\\mathrm{PartSQ^{Pt}}\\) calculates the average part segmentation mIoU within object-level segments:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.549, + 0.47, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\operatorname {P a r t S Q} _ {c} ^ {\\mathrm {P t}} = \\frac {\\sum_ {(p , g) \\in T P _ {c}} \\operatorname {I o U} _ {\\mathrm {p}} (p , g)}{| T P _ {c} |}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.47, + 0.673 + ], + "angle": 0, + "content": "To evaluate the ability of networks to conduct object-level panoptic segmentation, we report the Panoptic Quality (PQ) metric [13]. Similarly to the PartPQ, we report the average PQ over all classes, but also over all thing classes \\((\\mathrm{PQ}^{\\mathrm{Th}})\\) and stuff classes \\((\\mathrm{PQ}^{\\mathrm{St}})\\) separately." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Implementation details. TAPPS is built on top of the publicly available code of state-of-the-art panoptic segmentation network Mask2Former [3]. For all datasets, we use a batch size of 16 images, and train on 4 Nvidia A100 GPUs. To optimize TAPPS, we use AdamW [26], a weight decay of 0.05, and a polynomial learning rate decay schedule with an initial learning rate of \\(10^{-4}\\) and a power of 0.9. For Cityscapes-PP, we train for 90k iterations and apply conventional data augmentation steps [3, 18]: random flip, random resize with a factor between 0.5 and 2.0, and finally a random crop of \\(512 \\times 1024\\) pixels. For Pascal-PP, we train for 60k iterations in case of ImageNet pre-training, but for only 10k iterations in case of COCO pre-training, to prevent overfitting. Following state-of-the-art panoptic segmenta" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.892, + 0.404 + ], + "angle": 0, + "content": "tion implementations on COCO [3, 21], we apply large-scale jittering with a scale between 0.1 and 2.0 followed by a random crop of \\(1024 \\times 1024\\) pixels. During inference, we resize the image such that the shortest side is 800 pixels. Note that we use exactly the same training and testing settings for both TAPPS and the baseline. For more implementation details, see the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.415, + 0.587, + 0.43 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.44, + 0.634, + 0.455 + ], + "angle": 0, + "content": "5.1. Main results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.795 + ], + "angle": 0, + "content": "First, we compare TAPPS to the strong baseline that uses separate sets of queries for object-level panoptic segmentation and part-level semantic segmentation (see Sec. 4). The results in Tab. 1a demonstrate that TAPPS significantly outperforms the baseline on Pascal-PP in several aspects. Most importantly, the PartPQ for object-level classes with parts \\((\\mathrm{PartPQ}^{\\mathrm{Pt}})\\) is \\(+4.4\\) or \\(+2.4\\) higher, depending on the pretraining strategy. Looking in more detail, we find that this increase is caused by two individual improvements: (1) The part-level segmentation quality \\((\\mathrm{PartSQ}^{\\mathrm{Pt}})\\) is considerably higher than for the baseline. This shows the positive impact of having a joint representation for objects and parts, and simplifying the part segmentation task by only allowing compatible predictions (see also Sec. 5.3). (2) The panoptic quality for thing classes \\((\\mathrm{PQ}^{\\mathrm{Th}})\\) also sees a substantial increase. Here, it should be noted that (a) all the object-level classes with parts are thing classes, and (b) thing classes require instance separation. Thus, this improvement does not only show the benefit of learning objects and parts jointly, but also indicates that learning parts in an object-instance-aware manner indeed improves the ability of the network to separate object instances (see also Sec. 5.4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For Cityscapes-PP, we see similar results in Tab. 1b. Again, we observe improvements on the PartPQ\\(^{\\mathrm{Pt}}\\), PartSQ\\(^{\\mathrm{Pt}}\\) and PQ\\(^{\\mathrm{Th}}\\) metrics. In this case, the absolute improvements are slightly smaller. This is expected because this dataset contains significantly fewer classes and very similar images, making it easier to learn the PPS task and limiting the potential gains that can still be obtained by TAPPS." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3179" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.146, + 0.089, + 0.825, + 0.452 + ], + "angle": 0, + "content": "
MethodBackbonePre-trainingPartPQPartSQPQ
PtNo PtAllPtAll
Pascal-PP validation
JPPF [10]EfficientNet-B5I48.326.932.2--
TAPPS (ours)RN-50I59.639.444.674.347.1
Panoptic-PartFormer† [18]RN-50I,C56.138.843.266.847.6
Panoptic-PartFormer++† [19]RN-50I,C52.642.645.160.451.6
TAPPS (ours)RN-50I,C67.250.454.775.157.7
Panoptic-PartFormer† [18]Swin-BI,C64.350.654.170.858.1
Panoptic-PartFormer++† [19]Swin-BI,C48.952.151.353.059.8
TAPPS (ours)Swin-BI,C72.256.360.478.163.0
Cityscapes-PP val
Panoptic-PartFormer [18]RN-50I--54.5-57.8
Panoptic-PartFormer++ [19]RN-50I--57.5-61.6
JPPF [10]EfficientNet-B5I47.763.859.6--
TAPPS (ours)RN-50I48.763.159.366.862.4
Panoptic-PartFormer [18]RN-50I,C43.962.457.560.161.6
Panoptic-PartFormer++ [19]RN-50I,C42.565.159.2-63.6
TAPPS (ours)RN-50I,C48.965.761.366.964.4
Panoptic-PartFormer [18]Swin-BI,C45.667.862.059.066.6
Panoptic-PartFormer++ [19]Swin-BI,C46.068.262.3-68.0
Panoptic-PartFormer++ [19]ConvNeXt-BI,C46.469.163.1-68.2
SegFormer-B5 + CondInst + BPR [38–40, 45]MiT-B5 + RN-50I,C*48.667.562.5--
TAPPS (ours)Swin-BI,C53.069.064.868.068.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.456, + 0.894, + 0.514 + ], + "angle": 0, + "content": "Table 2. Comparison with state of the art. Evaluation on the Cityscapes-PP and Pascal-PP benchmarks [4-6]. RN-50 is ResNet-50 [9]. Other backbones are EfficientNet-B5 [36], MiT-B5 [45], Swin-B [24] and ConvNeXt-B [25]. I = ImageNet [33], C = COCO panoptic [21], C* = COCO pre-training for instance segmentation. †PartPQ scores for these existing methods have been re-evaluated using official code [5] and are higher than originally reported on Pascal-PP [18, 19], see supplementary material for more details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.531, + 0.364, + 0.547 + ], + "angle": 0, + "content": "5.2. Comparison with state of the art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.473, + 0.827 + ], + "angle": 0, + "content": "In Tab. 2, we compare TAPPS to existing state-of-the-art methods across different datasets, backbones, and pretraining settings. On both Pascal-PP and Cityscapes-PP, TAPPS significantly outperforms existing work. Most importantly, it consistently scores higher on the PartPQ, PartPQ\\(^{\\text{Pt}}\\), and PartSQ\\(^{\\text{Pt}}\\) metrics. TAPPS is only slightly outperformed by JPPF [10] on Cityscapes-PP with ImageNet pre-training, but we note that JPPF uses the EfficientNet-B5 [36] backbone, which is much more powerful than the ResNet-50 [9] used by TAPPS. Moreover, we do outperform JPPF by a large margin on the Pascal-PP dataset, showing the strength of TAPPS on more complex datasets. Overall, we achieve new state-of-the-art results on both datasets, obtaining PartPQ scores of 60.4 and 64.8 on Pascal-PP and Cityscapes-PP, improvements of +6.3 and +1.7, respectively. In the supplementary material, we compare qualitative examples of TAPPS, our baseline, and existing approaches, and we also show typical failure cases." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.833, + 0.191, + 0.848 + ], + "angle": 0, + "content": "5.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Predicting only compatible parts. Using shared queries for object- and part-level segmentation enables TAPPS to only allow part segmentation predictions that" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.528, + 0.891, + 0.698 + ], + "angle": 0, + "content": "
Predict compat. partsPartPQPartSQPQObj. w/o conflicts
TrainingTestingPtNo PtAllPtAll
ImageNet pre-training
Baseline55.238.842.971.545.766.1%
XX57.939.243.972.946.699.7%
X✓58.039.243.972.946.6100.0%
✓✓59.639.444.674.347.1100.0%
COCO pre-training
Baseline64.849.953.773.157.164.6%
XX66.350.554.674.057.999.5%
X✓66.450.554.674.057.9100.0%
✓✓67.250.454.775.157.7100.0%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.703, + 0.894, + 0.731 + ], + "angle": 0, + "content": "Table 3. Predicting only compatible part classes during training and testing. Evaluated on Pascal-PP [1, 5, 6, 29]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "are compatible with the query's object-level class (see Sec. 3.3.2). In Tab. 3, we evaluate the effect of predicting only compatible parts during both training and testing. In addition to the main metrics, we also assess the percentage of predicted objects for which there are no conflicting part-level predictions. The results show that, even when allowing incompatible predictions, TAPPS has significantly fewer object-part conflicts than the baseline. Naturally, when predicting only compatible parts during testing, all object-part conflicts are removed, eliminating the need" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3180" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.109, + 0.089, + 0.437, + 0.18 + ], + "angle": 0, + "content": "
Adaptation layersPartPQPartSQPQ
PtNo PtAllPtAll
-67.349.654.175.257.1
1×FC67.150.354.675.157.6
2-layer MLP67.250.454.775.157.7
3-layer MLP67.150.454.675.157.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.185, + 0.47, + 0.214 + ], + "angle": 0, + "content": "Table 4. JOPS head architecture (Fig. 2). Evaluated on Pascal-PP [1, 5, 6, 29], with pre-training on COCO panoptic [21]." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.226, + 0.27, + 0.274 + ], + "angle": 0, + "content": "
MethodmIoUThPQTh
Baseline69.862.0
TAPPS (ours)69.965.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.103, + 0.276, + 0.248, + 0.289 + ], + "angle": 0, + "content": "(a) Pascal-PP [1, 5, 6, 29]." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.226, + 0.463, + 0.274 + ], + "angle": 0, + "content": "
MethodmIoUThPQTh
Baseline77.653.8
TAPPS (ours)78.255.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.301, + 0.276, + 0.437, + 0.29 + ], + "angle": 0, + "content": "(b) Cityscapes-PP [4, 5]." + }, + { + "type": "table_caption", + "bbox": [ + 0.087, + 0.291, + 0.458, + 0.307 + ], + "angle": 0, + "content": "Table 5. Performance for things. ImageNet pre-training [33]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.469, + 0.416 + ], + "angle": 0, + "content": "for post-processing. However, this does not yield a big improvement in terms of the segmentation quality. When we also apply this during training, we do observe a part segmentation quality \\((\\mathrm{PartSQ^{Pt}})\\) improvement. This shows that simplifying the part segmentation task during training leads to improved performance, as we hypothesized in Sec. 3.3.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.537 + ], + "angle": 0, + "content": "JOPS head architecture. In Tab. 4, we evaluate the impact of using different numbers of adaptation layers in the JOPS head before applying the \\( N^{\\mathrm{pc}} \\) fully-connected (FC) layers to generate the part queries (see Fig. 2). We find that one or two adaptation layers are necessary to generate object- and part-level representations that are sufficiently distinct to perform their respective tasks accurately; adding any more layers does not yield further improvements." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.546, + 0.265, + 0.563 + ], + "angle": 0, + "content": "5.4. Additional analyses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Instance separability. Tab. 1 showed that TAPPS improves the \\(\\mathrm{PQ^{\\mathrm{Th}}}\\), i.e., the ability to recognize, segment and classify object instances. To assess if this is due to better object recognition or improved instance separability, we group all thing instances of the same object-level class together and evaluate the instance-agnostic semantic segmentation performance with the mIoU\\(^{\\mathrm{Th}}\\). If the \\(\\mathrm{PQ^{\\mathrm{Th}}}\\) improvement were due to improved recognition, we expect the mIoU\\(^{\\mathrm{Th}}\\) to improve too. However, the results in Tab. 5 show only a minor mIoU\\(^{\\mathrm{Th}}\\) improvement. This indicates that the \\(\\mathrm{PQ^{\\mathrm{Th}}}\\) gain is not due to better recognition, but mainly results from a better ability to separate instances, showing the benefit of learning parts in an object-instance-aware manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Performance on Pascal-PP-107. To assess the performance of TAPPS in a more complex setting, we evaluate it on Pascal-PP-107, which has 107 part-level classes instead of 57. The results on this dataset, reported in Tab. 6, show once more that TAPPS consistently improves the part segmentation and thing segmentation performance with respect to the baseline, in this case leading to a PartPQ\\(^{\\text{Pt}}\\) improvement of +4.1 or +2.2, depending on the pre-training strategy. This demonstrates that TAPPS is also effective" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.891, + 0.219 + ], + "angle": 0, + "content": "
MethodPartPQPartSQPQ
PtNo PtAllPtThStAll
ImageNet pre-training
Baseline45.738.940.659.362.137.645.9
TAPPS (ours)49.839.342.061.765.837.747.2
COCO pre-training
Baseline53.249.850.759.975.248.157.2
TAPPS (ours)55.450.051.462.175.648.157.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.224, + 0.892, + 0.252 + ], + "angle": 0, + "content": "Table 6. 107 part-level categories. Results on the more challenging Pascal-PP-107 dataset [1, 5, 6, 27, 29]." + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.264, + 0.84, + 0.334 + ], + "angle": 0, + "content": "
MethodPartPQPartSQPQ
PtNo PtAllPtAll
Dynamic65.150.554.273.157.7
Fixed67.250.454.775.157.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.34, + 0.892, + 0.369 + ], + "angle": 0, + "content": "Table 7. Fixed or dynamic part segmentation. Evaluated on Pascal-PP [1, 5, 6, 29], with pre-training on COCO panoptic [21]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.382, + 0.789, + 0.397 + ], + "angle": 0, + "content": "when the PPS task becomes more complex." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.892, + 0.624 + ], + "angle": 0, + "content": "Fixed vs. dynamic part segmentation. As explained in Sec. 3.3.2, TAPPS uses a fixed fully connected layer for each part class to generate the corresponding 'per-object' part query. Alternatively, it is possible to predict part masks and classes dynamically within each object segment, using a set of dynamic queries like we do for object-level segmentation. In Tab. 7, we compare our fixed part segmentation setting with this dynamic approach, which is explained in more detail in the supplementary material. We find that our fixed approach results in a better \\(\\mathrm{PartSQ^{Pt}}\\) and therefore \\(\\mathrm{PartPQ^{Pt}}\\) performance. We hypothesize that the dynamic approach performs worse because it makes the part segmentation task unnecessarily complex, as the network needs to learn to assign segments to queries dynamically, and additionally predict a class label." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.637, + 0.619, + 0.652 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.662, + 0.892, + 0.828 + ], + "angle": 0, + "content": "With experiments, we have shown that TAPPS considerably outperforms methods that predict objects and parts separately, by improving the object instance separability, part segmentation quality, and object-part compatibility. Importantly, these improvements can be attributed to the fact that TAPPS is directly optimized for the PPS task, using a set of shared queries to jointly predict objects and corresponding parts. With our promising findings, we hope to inspire future research towards even more complete scene understanding, e.g., image segmentation at even more abstraction levels, potentially with more flexible class hierarchies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.831, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by Eindhoven Engine, NXP Semiconductors, and Brainport Eindhoven. This work made use of the Dutch national e-infrastructure with the support of the SURF Cooperative using grant no. EINF-5302, which is financed by the Dutch Research Council (NWO)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3181" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect What You Can: Detecting and Representing Objects using Holistic Models and Body Parts. In CVPR, 2014. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.47, + 0.213 + ], + "angle": 0, + "content": "[2] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-Pixel Classification is Not All You Need for Semantic Segmentation. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.47, + 0.27 + ], + "angle": 0, + "content": "[3] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. In CVPR, 2022. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.272, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[4] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes Dataset for Semantic Urban Scene Understanding. In CVPR, 2016. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.343, + 0.47, + 0.384 + ], + "angle": 0, + "content": "[5] Daan de Geus, Panagiotis Meletis, Chenyang Lu, Xiaoxiao Wen, and Gijs Dubbelman. Part-aware Panoptic Segmentation. In CVPR, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.47, + 0.44 + ], + "angle": 0, + "content": "[6] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes (VOC) Challenge. *IJCV*, 88(2):303–338, 2010. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.443, + 0.47, + 0.485 + ], + "angle": 0, + "content": "[7] Ke Gong, Xiaodan Liang, Yicheng Li, Yimin Chen, Ming Yang, and Liang Lin. Instance-level human parsing via part grouping network. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.486, + 0.47, + 0.527 + ], + "angle": 0, + "content": "[8] Ke Gong, Yiming Gao, Xiaodan Liang, Xiaohui Shen, Meng Wang, and Liang Lin. Graphonomy: Universal human parsing via graph transfer learning. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.529, + 0.47, + 0.57 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.572, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[10] Sravan Kumar Jagadeesh, René Schuster, and Didier Stricker. Multi-task Fusion for Efficient Panoptic-Part Segmentation. In ICPRAM, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.615, + 0.47, + 0.656 + ], + "angle": 0, + "content": "[11] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.658, + 0.47, + 0.697 + ], + "angle": 0, + "content": "[12] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollar. Panoptic Feature Pyramid Networks. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.47, + 0.742 + ], + "angle": 0, + "content": "[13] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic Segmentation. In CVPR, 2019. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[14] Jianshu Li, Jian Zhao, Yunpeng Chen, Sujoy Roy, Shuicheng Yan, Jiashi Feng, and Terence Sim. Multi-human parsing machines. In ACM MM, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.47, + 0.827 + ], + "angle": 0, + "content": "[15] Liulei Li, Tianfei Zhou, Wenguan Wang, Jianwu Li, and Yi Yang. Deep Hierarchical Semantic Segmentation. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[16] Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. Self-Correction for Human Parsing. IEEE TPAMI, 44(6):3260-3271, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[17] Qizhu Li, Anurag Arnab, and Philip HS Torr. Holistic, instance-level human parsing. In BMVC, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[18] Xiangtai Li, Shilin Xu, Yibo Yang, Guangliang Cheng, Yunhai Tong, and Dacheng Tao. Panoptic-PartFormer: Learning a Unified Model for Panoptic Part Segmentation. In ECCV, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.893, + 0.219 + ], + "angle": 0, + "content": "[19] Xiangtai Li, Shilin Xu, Yibo Yang, Haobo Yuan, Guangliang Cheng, Yunhai Tong, Zhouchen Lin, and Dacheng Tao. PanopticPartFormer++: A Unified and Decoupled View for Panoptic Part Segmentation. arXiv preprint arXiv:2301.00954, 2023. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.893, + 0.276 + ], + "angle": 0, + "content": "[20] Xiaodan Liang, Ke Gong, Xiaohui Shen, and Liang Lin. Look into Person: Joint Body Parsing & Pose Estimation Network and A New Benchmark. IEEE TPAMI, 41(4):871-885, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.893, + 0.334 + ], + "angle": 0, + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.337, + 0.893, + 0.379 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature Pyramid Networks for Object Detection. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.381, + 0.893, + 0.45 + ], + "angle": 0, + "content": "[23] Qing Liu, Adam Kortylewski, Zhishuai Zhang, Zizhang Li, Mengqi Guo, Qihao Liu, Xiaoding Yuan, Jiteng Mu, Weichao Qiu, and Alan Yuille. Learning Part Segmentation Through Unsupervised Domain Adaptation From Synthetic Vehicles. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.453, + 0.893, + 0.507 + ], + "angle": 0, + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. In ICCV, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.893, + 0.551 + ], + "angle": 0, + "content": "[25] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.554, + 0.893, + 0.582 + ], + "angle": 0, + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.584, + 0.893, + 0.639 + ], + "angle": 0, + "content": "[27] Umberto Michieli, Edoardo Borsato, Luca Rossi, and Pietro Zanuttigh. GMNet: Graph Matching Network for Large Scale Part Semantic Segmentation in the Wild. In ECCV, 2020. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.642, + 0.893, + 0.684 + ], + "angle": 0, + "content": "[28] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentatio. In 3DV, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.686, + 0.893, + 0.755 + ], + "angle": 0, + "content": "[29] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The Role of Context for Object Detection and Semantic Segmentation in the Wild. In CVPR, 2014. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.758, + 0.893, + 0.797 + ], + "angle": 0, + "content": "[30] Tai-Yu Pan, Qing Liu, Wei-Lun Chao, and Brian Price. Towards Open-World Segmentation of Parts. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.801, + 0.893, + 0.856 + ], + "angle": 0, + "content": "[31] Lu Qi, Jason Kuen, Weidong Guo, Jiumiang Gu, Zhe Lin, Bo Du, Yu Xu, and Ming-Hsuan Yang. AIMS: All-Inclusive Multi-Level Segmentation for Anything. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[32] Tao Ruan, Ting Liu, Zilong Huang, Yunchao Wei, Shikui Wei, and Yao Zhao. Devil in the details: Towards accurate single and multiple human parsing. In AAAI, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3182" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[33] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 115(3):211-252, 2015. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.22 + ], + "angle": 0, + "content": "[34] Rishubh Singh, Pranav Gupta, Pradeep Shenoy, and Ravikiran Sarvadevabhatla. FLOAT: Factorized Learning of Object Attributes for Improved Multi-Object Multi-Part Scene Parsing. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.47, + 0.263 + ], + "angle": 0, + "content": "[35] Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.265, + 0.47, + 0.305 + ], + "angle": 0, + "content": "[36] Mingxing Tan and Quoc Le. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. In ICML, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.308, + 0.47, + 0.35 + ], + "angle": 0, + "content": "[37] Xin Tan, Jiachen Xu, Zhou Ye, Jinkun Hao, and Lizhuang Ma. Confident Semantic Ranking Loss for Part Parsing. In ICME, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.352, + 0.47, + 0.407 + ], + "angle": 0, + "content": "[38] Chufeng Tang, Hang Chen, Xiao Li, Jianmin Li, Zhaoxiang Zhang, and Xiaolin Hu. Look Closer To Segment Better: Boundary Patch Refinement for Instance Segmentation. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.409, + 0.47, + 0.45 + ], + "angle": 0, + "content": "[39] Chufeng Tang, Lingxi Xie, Xiaopeng Zhang, Xiaolin Hu, and Qi Tian. Visual Recognition by Request. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.453, + 0.47, + 0.481 + ], + "angle": 0, + "content": "[40] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional Convolutions for Instance Segmentation. In ECCV, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.47, + 0.525 + ], + "angle": 0, + "content": "[41] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: End-to-End Panoptic Segmentation With Mask Transformers. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.527, + 0.47, + 0.568 + ], + "angle": 0, + "content": "[42] Wenguan Wang, Hailong Zhu, Jifeng Dai, Yanwei Pang, Jianbing Shen, and Ling Shao. Hierarchical human parsing with typed part-relation reasoning. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.57, + 0.47, + 0.611 + ], + "angle": 0, + "content": "[43] Meng Wei, Xiaoyu Yue, Wenwei Zhang, Shu Kong, Xihui Liu, and Jiangmiao Pang. OV-PARTS: Towards Open-Vocabulary Part Segmentation. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.47, + 0.655 + ], + "angle": 0, + "content": "[44] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.47, + 0.712 + ], + "angle": 0, + "content": "[45] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers. In NeurIPS, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.47, + 0.755 + ], + "angle": 0, + "content": "[46] Lu Yang, Qing Song, Zhihui Wang, and Ming Jiang. Parsing R-CNN for Instance-Level Human Analysis. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.47, + 0.813 + ], + "angle": 0, + "content": "[47] Lu Yang, Qing Song, Zhihui Wang, Mengjie Hu, Chun Liu, Xueshi Xin, Wenhe Jia, and Songcen Xu. Renovating Parsing R-CNN for Accurate Multiple Human Parsing. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[48] Lu Yang, Qing Song, Zhihui Wang, Mengjie Hu, and Chun Liu. Hier R-CNN: Instance-Level Human Parts Detection and A New Benchmark. IEEE TIP, 30:39-54, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[49] Sanyi Zhang, Xiaochun Cao, Guo-Jun Qi, Zhanjie Song, and Jie Zhou. AIParsing: Anchor-Free Instance-Level Human Parsing. IEEE TIP, 31:5599-5612, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.134 + ], + "angle": 0, + "content": "[50] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-Net: Towards Unified Image Segmentation. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.893, + 0.19 + ], + "angle": 0, + "content": "[51] Jian Zhao, Jianshu Li, Yu Cheng, Terence Sim, Shuicheng Yan, and Jiashi Feng. Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for multi-human parsing. In ACM MM, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.893, + 0.233 + ], + "angle": 0, + "content": "[52] Yifan Zhao, Jia Li, Yu Zhang, and Yonghong Tian. Multi-Class Part Parsing With Joint Boundary-Semantic Awareness. In ICCV, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3183" + } + ] +] \ No newline at end of file diff --git a/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/9dbd5f9f-08af-4d23-b3b6-162e326d318c_origin.pdf b/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/9dbd5f9f-08af-4d23-b3b6-162e326d318c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..57a8a3b259729bde78793c4ef042793aa1d51253 --- /dev/null +++ b/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/9dbd5f9f-08af-4d23-b3b6-162e326d318c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dad29239c205d49cd33a3f66501bc5f9643b8739403517b469b6ffe77b580a0f +size 603597 diff --git a/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/full.md b/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7a5a4e3d41252a85bf8cc741d5dc43581c277aca --- /dev/null +++ b/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/full.md @@ -0,0 +1,287 @@ +# Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations + +Daan de Geus Gijs Dubbelman + +Eindhoven University of Technology + +{d.c.d.geus, g.dubbelman}@tue.nl + +# Abstract + +Part-aware panoptic segmentation (PPS) requires (a) that each foreground object and background region in an image is segmented and classified, and (b) that all parts within foreground objects are segmented, classified and linked to their parent object. Existing methods approach PPS by separately conducting object-level and part-level segmentation. However, their part-level predictions are not linked to individual parent objects. Therefore, their learning objective is not aligned with the PPS task objective, which harms the PPS performance. To solve this, and make more accurate PPS predictions, we propose Task-Aligned Part-aware Panoptic Segmentation (TAPPS). This method uses a set of shared queries to jointly predict (a) object-level segments, and (b) the part-level segments within those same objects. As a result, TAPPS learns to predict part-level segments that are linked to individual parent objects, aligning the learning objective with the task objective, and allowing TAPPS to leverage joint object-part representations. With experiments, we show that TAPPS considerably outperforms methods that predict objects and parts separately, and achieves new state-of-the-art PPS results. + +# 1. Introduction + +To fully understand what is depicted in an image, it is important to consider concepts at different levels of abstraction. For rich scene understanding, we should not only recognize foreground objects (e.g., car, human) and background regions (e.g., sky, ocean), but also simultaneously identify the parts that constitute the objects (e.g., car-wheel, human-arm). In a step towards such comprehensive scene understanding, De Geus et al. introduced part-aware panoptic segmentation (PPS) [5]. The objective of this computer vision task is (1) to output a segmentation mask and class label for all thing objects and stuff regions in an image like for panoptic segmentation [13] – we call these object-level segments – and (2) to simultaneously segment and classify + +![](images/f3ebccf1cd7a24734832cc79991c8b6d16711034356de08e06fb8f2159a81101.jpg) +(a) Separate object and part segmentation +X Learning objective not aligned with task objective +(b) Joint object and part segmentation (ours) + +![](images/ba6f1c42782fe4a17c1b152ce6884b5e6aec85c0da13b1681eb1054306f72945.jpg) +$\checkmark$ Learning objective aligned with task objective +Figure 1. Task-aligned part-aware panoptic segmentation. (a) Existing works separately predict object-level segments and object-instance-unaware part-level segments. (b) In this work, we predict objects and parts jointly, using a set of shared queries. This allows our method to predict parts within individual object segments, aligning its learning objective with the PPS task objective. + +all parts within each identified object. These are called part-level segments and should be explicitly linked to an object-level segment, establishing a part-whole relation. + +Current state-of-the-art methods [18, 19] address PPS by using two different sets of learnable queries to separately predict object-level and part-level segments, see Fig. 1a. While these methods outperform earlier baselines [5], they have one main limitation: their learning objective is not aligned with the task objective. Where the PPS task objective is to predict parts within each individual object-level segment, these existing works conduct part-level semantic segmentation, predicting part-level masks that cover multiple objects (see Fig. 1a). In other words, these networks + +are not optimized for the PPS task, but instead solve the surrogate subtasks of object-level panoptic segmentation and part-level semantic segmentation. As a result, to assign parts to individual objects, these methods require post-processing. Additionally, we hypothesize that there are several other negative consequences: (1) These networks learn a conflicting feature representation. They learn that object-level thing instances should be separated, but also that the parts of these object-level instances should be grouped together. We expect that this harms the ability of the networks to separate instances. (2) Predicting objects and parts separately may cause incompatible predictions (e.g., a car-window part with a bicycle object), which makes it unclear which prediction to trust, and requires further post-processing. (3) The networks encode information about objects and their parts separately, whereas this information is potentially complementary. See Sec. 3.2.2 for more details on these limitations. + +In this work, we aim to design a simple network for PPS that overcomes these limitations and thereby makes more accurate PPS predictions. To achieve this, we propose the Task-Aligned Part-aware Panoptic Segmentation (TAPPS) method. Instead of using separate queries for objects and parts, TAPPS uses one set of shared queries to jointly represent objects and the parts they contain. Specifically, each of these queries learns to represent at most one object-level segment, for which it predicts (1) a segmentation mask and object-level class, and (2) the segmentation masks and classes for all part-level segments within this object. This is visualized in Fig. 1b and explained in Sec. 3. + +With this approach, TAPPS explicitly predicts part-level segments per individual object. Therefore, the network's learning objective is now aligned with the PPS task objective, and TAPPS is directly optimized for the PPS task. As a result, both object- and part-level segmentation are learned in an 'object-instance-aware' manner. This removes the conflict in the learned feature representations, allowing for better instance separability. Additionally, we hypothesize that predicting objects and parts from the same query reduces incompatibilities between object and part predictions, as they are made using the same information. Moreover, we use TAPPS to go even further, and explicitly only predict part segments compatible with predicted object segments. This enforces full object-part compatibility and simplifies the part segmentation task. Finally, the network can now encode complementary object-level and part-level information in a shared query, giving it a richer representation for making object- and part-level segmentation predictions. + +With experiments, detailed in Sec. 4, we show that TAPPS outperforms the baseline with separate object and part queries in multiple aspects: (1) The part segmentation quality within identified objects is significantly better, which shows the positive impact of using joint object-part + +representations and simplifying the part segmentation task. (2) The object instance segmentation performance is considerably improved, demonstrating the benefits of learning parts in an object-instance-aware manner. Together, this yields a large overall improvement with respect to the baseline, and causes TAPPS to considerably outperform existing works across different benchmarks, achieving new state-of-the-art results. See Sec. 5 for more extensive results. + +To summarize, we make the following contributions: + +- We propose TAPPS, a simple approach for PPS that aligns the learning objective with the task objective, facilitating object instance separability and enabling joint object-part representations for more accurate PPS predictions. +- We use the shared object-part queries to constrain TAPPS to only predict part segments that are compatible with the object class, enforcing object-part compatibility and simplifying the part segmentation task. +- We experimentally show the effectiveness of TAPPS across a range of datasets and network configurations. + +The code for TAPPS is made publicly available through https://tue-mps.github.io/tapps/. + +# 2. Related work + +Part-aware panoptic segmentation. Part-aware panoptic segmentation [5] is an image segmentation task introduced for scene understanding at multiple abstraction levels. It extends panoptic segmentation [13] by requiring (a) object-level panoptic segmentation, and (b) part-level segmentation within object-level segments. Most existing works do not predict parts per object-level segment, but instead make separate predictions for panoptic segmentation and the surrogate task of part-level semantic segmentation - or part segmentation [10, 18, 19]. JPPF [10] proposes a single-network approach with a shared encoder followed by separate heads for semantic, instance, and part segmentation. To merge the predictions by these heads to the PPS format, JPPF proposes a new rule-based fusion strategy that outperforms the originally introduced merging strategy for PPS [5]. Panoptic-PartFormer and its extension Panoptic-PartFormer++ [18, 19] tackle PPS with a Transformer-based approach that independently makes panoptic segmentation and part segmentation predictions using separate sets of learnable queries for thing, stuff and part segments. Alternatively, ViRReq [39] is a paradigm in which PPS is predicted in a cascading fashion, by first segmenting objects and then segmenting the parts within these objects 'by request', but it requires multiple networks to achieve this. + +In contrast to these existing approaches, we propose a method that jointly predicts object-level segments and the part-level segments within these objects. This aligns the learning objective with the task objective, and yields improved PPS performance. For more details, see Sec. 3. + +Part-level image segmentation. Part-level segmentation is also widely studied beyond PPS. Most works address part segmentation, i.e., semantic segmentation for parts [7, 8, 14-17, 20, 23, 27, 32, 34, 35, 37, 42, 43, 51, 52]. For more complete scene understanding, other works also take into account object instances, like PPS does, with instance-aware part segmentation [16, 30, 32, 46-49]. However, these works do not consider background stuff classes, which PPS does consider. Alternatively, UPerNet [44] separately predicts part-level and object-level semantic segmentation, including background classes, but without any instance awareness. Another work has extended PPS by also predicting the relations between objects [31], training a model on annotations from different datasets. In this work, we address the PPS task as introduced by De Geus et al. [5], and we compare our approach to state-of-the-art alternatives. + +# 3. Method + +# 3.1. PPS task definition + +Part-aware panoptic segmentation (PPS) [5] requires consistent image segmentation across two abstraction levels: the object level and part level. For object-level segmentation, following panoptic segmentation [13], an image should be divided into $N$ object-level segments, where $N$ varies per image. Each segment $s_i$ consists of a binary mask $\mathbf{M}_i^{\mathrm{obj}}$ and an object class label $c_i^{\mathrm{obj}}$ . Thing classes require a segment per object instance, and stuff classes require a single segment per class. Next, per segment $s_i$ , PPS requires a set of $K_i$ part-level segments. Each part-level segment $s_{i,j}^{\mathrm{pt}}$ consists of a binary part-level mask $\mathbf{M}_{i,j}^{\mathrm{pt}}$ and a part-level class $c_{i,j}^{\mathrm{pt}}$ that are compatible with the object-level mask and class. Specifically, the part-level mask must be a subset of the object-level mask, and the part-level class should be one of the part-level classes defined for the object-level class. For instance, if $c_i^{\mathrm{obj}}$ is car, then $c_{i,j}^{\mathrm{pt}}$ cannot be human-head but it can be car-window. For all parts, a single part-level segment is required for each part class within an object. + +To summarize, each image should be divided into a set of segments $S = \{s_i\}_{i=1}^N = \{(\mathbf{M}_i^{\mathrm{obj}}, c_i^{\mathrm{obj}}, \mathcal{P}_i)\}_{i=1}^N$ , where $\mathcal{P}_i = \{s_{i,j}^{\mathrm{pt}}\}_{j=1}^{K_i} = \{(\mathbf{M}_{i,j}^{\mathrm{pt}}, c_{i,j}^{\mathrm{pt}})\}_{j=1}^{K_i}$ is a set of part-level segments compatible with the object-level segment in $s_i$ . In practice, part-level classes are only defined for some object-level classes. When segment $s_i$ has an object class that does not have parts, then $K_i = 0$ and $s_i = (\mathbf{M}_i^{\mathrm{obj}}, c_i^{\mathrm{obj}})$ . + +# 3.2. Preliminaries + +Like existing state-of-the-art PPS approaches [18, 19], our work uses the powerful mask classification metaarchitecture that is the foundation of many state-of-the-art segmentation models [2, 3, 11, 41, 50]. Here, we first explain this mask classification paradigm in more detail. + +Then, we explain how existing works use it for PPS, and what the limitations are of these approaches. + +# 3.2.1 Mask classification framework + +The concept of mask classification is to predict a set of $N^{\mathrm{q}}$ object-level segments, i.e., a set of $N^{\mathrm{q}}$ pixel-level masks $\hat{\mathbf{M}}^{\mathrm{obj}}$ and corresponding class labels $\hat{c}^{\mathrm{obj}}$ . To make these predictions, mask classification networks output two components: high-resolution features $\mathbf{F} \in \mathbb{R}^{E \times H \times W}$ and queries $\mathbf{Q} \in \mathbb{R}^{N^{\mathrm{q}} \times E}$ , where $H$ and $W$ are the height and width of the features, and $E$ is the feature and query dimensionality. Each of the $N^{\mathrm{q}}$ queries is used to predict the class label and segmentation mask of at most one object-level segment. The class predictions are a function of only the queries, and the mask predictions are a function of both the queries $\mathbf{Q}$ and the high-resolution features $\mathbf{F}$ . To output these high-resolution features, the input image is first fed into a backbone to extract multi-scale features, e.g., ResNet [9] or Swin [24]. Subsequently, these features are further refined and upsampled to high-resolution features by a pixel decoder, e.g., Semantic FPN [12, 22]. In the other part of the network, the queries $\mathbf{Q}$ are generated by processing learnable queries $\mathbf{Q}^{\mathrm{o}}$ using a transformer decoder, which applies self-attention across queries and cross-attention with image features. Through bipartite matching between $N^{\mathrm{q}}$ predicted and $N$ ground-truth segments, each query is assigned to at most one ground-truth segment. As $N^{\mathrm{q}}$ is not always equal to $N$ , some queries may not be assigned to a ground-truth segment. If so, they do not receive any supervision for segmentation, and will learn a 'no object' class. In this work, we build upon the Mask2Former [3] instantiation of this meta-architecture. + +# 3.2.2 Mask classification for PPS + +With the aforementioned meta-architecture, we can conduct object-level segmentation, and thereby solve one aspect of the PPS task definition. To solve the full task, the object-level segments should also be further segmented into parts. + +Current state-of-the-art approaches [18, 19] achieve this by introducing an additional set of queries, the part-level queries depicted in Fig. 1a. Each of these part-level queries learns to represent a part-level segment. However, in contrast to the PPS task definition, these part-level segments are not explicitly linked to an object-level segment. Instead, these part-level segments represent an entire part-level class, i.e., their masks contain all pixels belonging to one part-level class across multiple object segments. + +This means that the learning objective of these networks is not aligned with the PPS task objective. As a result, they are not directly optimized for PPS, but instead for two surrogate tasks: object-level panoptic segmentation and part-level semantic segmentation. In addition to necessitating + +![](images/2dbf6ba2e62855a536b9cb4c4f8343f012d40b9d8d571649bc4ea437fee07eac.jpg) +Figure 2. Network architecture. Left: The overall TAPPS architecture. A set of learnable queries and features from a backbone are fed into a pixel decoder and transformer decoder to generate high-resolution features and processed queries. Right: These queries and features are fed into the JOPS head, which predicts for each shared query (a) an object-level class, (b) an object-level segmentation mask, and (c) a set of part-level masks for the part-level classes compatible with the object-level class. Operator $\otimes$ denotes a matrix multiplication. + +![](images/533af94f2cab23e2f80e9b3df42fd2f2c2d7c4f976286e49647ab30c269f6ef7.jpg) + +post-processing, we hypothesize that this has several negative consequences: (1) By learning parts in an 'object-instance-unaware' manner, the network learns conflicting representations in the features, F. On the one hand, different thing instances belong to different object-level segments, requiring distinct features. On the other hand, the parts of these instances belong to the same part-level segment, requiring similar features. We hypothesize that this conflict harms the ability of the network to learn to separate object-level thing instances. (2) Separately predicting objects and parts in this way may also cause incompatible predictions. These incompatibilities make it unclear for downstream processes which prediction to trust. To obtain predictions compliant with the PPS definition, rule-based post-processing is required, which discards incompatible predictions [5]. This is undesired because the discarded information is potentially correct. (3) By representing object- and part-level information with separate queries, the network cannot encode complementary information about an individual object and its parts, which it could use to make more informed predictions at these abstraction levels. + +# 3.3. Task-aligned PPS + +To overcome the limitations of state-of-the-art methods and achieve a better PPS performance, the objective of this work is to design a simple PPS approach that uses the powerful mask classification framework, but that aligns its learning objective with the PPS task objective, and explicitly predicts part-level segments within individual objects. To achieve this, we should (a) generate a unique query for each object-part combination, i.e., for each part-level segment within an object-level segment, and (b) explicitly link these part-level segments to their parent objects, as PPS requires. To accomplish this, we propose Task-Aligned Part-aware Panoptic Segmentation (TAPPS), a method that jointly represents + +object-level segments and their parts with a set of shared queries, and generates 'per-object' part queries from each shared query. Each of these queries can then be used to predict one part-level segment within an object-level segment, and is automatically linked to its parent object segment. + +# 3.3.1 Overall architecture + +In Fig. 2, we visualize the network architecture of TAPPS. First, we initialize a set of $N^{\mathbf{q}}$ learnable queries, $\mathbf{Q}^0 \in \mathbb{R}^{N^{\mathbf{q}} \times E}$ . Each of these queries learns to represent one object-level segment, but also the part-level segments belonging to this object (e.g., car3 and also car3-wheels, car3-windows, etc.). Following the mask classification framework described in Sec. 3.2, these learnable queries $\mathbf{Q}^0$ and the features from a backbone are fed into a decoder, which generates a set of processed queries $\mathbf{Q}$ and high-resolution features $\mathbf{F}$ . Subsequently, these processed queries and features enter the Joint Object and Part Segmentation (JOPS) head. For each query $\mathbf{Q}_i \in \mathbb{R}^E$ , this head predicts (a) an object-level class, (b) an object-level segmentation mask, and (c) a set of part-level segmentation masks and classes for the parts within this object-level segment. + +With this approach, we solve the limitations of existing works. Following the PPS task definition, we predict part-level segments per individual object-level segment. (1) As a result, we learn object-instance-aware representations for both parts and objects, improving the ability of the network to separate instances. (2) By predicting objects and parts jointly, the compatibility between both sets of predictions greatly improves. Moreover, in the JOPS head, we ensure that we only predict compatible parts, ensuring full object-part compatibility (see Sec. 3.3.2). (3) The joint object-part representations allow the network to fully leverage the complementary information from both abstraction levels. + +# 3.3.2 JOPS head + +Within the JOPS head, we predict an object-level segment and compatible part-level segments for each shared query. For the object-level predictions, we follow Mask2Former [3]: (1) For the object-level class, we feed each processed query $\mathbf{Q}_i$ through a single fully connected layer to predict a score for each possible class, and obtain the predicted class $\hat{c}_i^{\mathrm{obj}}$ by picking the highest-scoring class. (2) For the object-level mask, we feed the query $\mathbf{Q}_i$ through a 3-layer MLP and generate a mask by taking the product of the resulting mask queries $\mathbf{Q}_i^{\mathrm{m}}$ and the features $\mathbf{F}$ , and applying a sigmoid activation, yielding $\hat{\mathbf{M}}_i^{\mathrm{obj}} \in [0,1]^{H \times W}$ . + +To predict part-level segments using the same features $\mathbf{F}$ , we need to generate queries for each part class within an object-level segment. To do so, we first apply an MLP to each query $\mathbf{Q}_i$ to adapt it for part-level segmentation, and then apply $N^{\mathrm{pc}}$ different fully-connected layers, where $N^{\mathrm{pc}}$ is the total number of part-level classes in the dataset. This results in $\mathbf{Q}_i^{\mathrm{pt}} \in \mathbb{R}^{N^{\mathrm{pc}} \times E}$ , a set of 'per-object' part-level queries, where each query always corresponds to a fixed, predetermined part class. We could then take the product of $\mathbf{Q}_i^{\mathrm{pt}}$ and the features $\mathbf{F}$ to generate a segmentation mask for all part-level classes. However, we already have an object-level class prediction for each query, and we know that only a subset of all $N^{\mathrm{pc}}$ part classes is compatible with a certain object class. Therefore, we propose to only predict and supervise part-level masks for those compatible part-level classes. This simplifies the part segmentation task that the network needs to learn, as it no longer has to learn to predict 'empty' segmentation masks for incompatible part classes. Concretely, as visualized in Fig. 2, we identify the object-level class $\hat{c}_i^{\mathrm{obj}}$ for each query $\mathbf{Q}_i$ , and keep only the part-level queries $\mathbf{Q}_i^{\mathrm{pt,c}} \in \mathbb{R}^{N^{\mathrm{c}} \times E}$ that correspond to the $N^{\mathrm{c}}$ part classes that are compatible with the object-level class. Then, we compute the product of these remaining queries $\mathbf{Q}_i^{\mathrm{pt,c}}$ and the features $\mathbf{F}$ , and apply a sigmoid activation to generate compatible part segmentation masks $\hat{\mathbf{M}}_i^{\mathrm{pt}} \in [0,1]^{N^{\mathrm{c}} \times H \times W}$ . As each per-object part query corresponds to a fixed part class, we also know the part classes $\hat{c}_i^{\mathrm{pt}}$ . Note that the number $N^{\mathrm{c}}$ depends on the object-level class and will therefore vary per query $\mathbf{Q}_i$ . + +# 3.3.3 Training + +To assign each query to at most one ground-truth object-level segment with corresponding part-level segments during training, we apply bipartite matching based on the predicted and ground-truth object-level classes and masks. + +To supervise the object-level segments, we use the cross-entropy loss for the classes, and both the Dice [28] and cross-entropy loss for the segmentation masks. Together, these losses form the object-level loss $L_{\mathrm{obj}}$ . For the part-level segmentation masks, we also use the Dice and cross + +entropy loss, together forming part-level loss $L_{\mathrm{pt}}$ . We balance the losses with weights $\lambda_{\mathrm{obj}}$ and $\lambda_{\mathrm{pt}}$ to calculate the total loss + +$$ +L = \lambda_ {\mathrm {o b j}} L _ {\mathrm {o b j}} + \lambda_ {\mathrm {p t}} L _ {\mathrm {p t}}. \tag {1} +$$ + +We find that using $\lambda_{\mathrm{obj}} = \lambda_{\mathrm{pt}} = 1$ provides the best balance; see the supplementary material for more details. + +# 4. Experimental setup + +Datasets. We use the two PPS benchmarks for evaluation. + +Cityscapes Panoptic Parts (Cityscapes-PP) [5] extends the original Cityscapes dataset [4] with part-level labels. It consists of street scene images from several cities. It has labels for 19 object classes (11 stuff; 8 thing). Part classes are defined and labeled for 5 object-level thing classes; there are 23 part classes in total. We train on the train split (2975 images), and evaluate on the val split (500 images). + +Pascal Panoptic Parts (Pascal-PP) [5], which combines existing labels for Pascal VOC [1, 6, 29], consists of a wide range of scenes and classes. There are 59 object-level classes (39 stuff; 20 thing), and part-level classes are defined for 15 thing classes. Unless otherwise indicated, we use the default part class definition with 57 part classes in total [5]. To evaluate a more challenging setting, we additionally evaluate on Pascal-PP-107, which uses the 107 non-background classes from the Pascal-Part-108 definition introduced by Michieli et al. [27] for part segmentation. For both class definitions, we train on the training split (4998 images), and evaluate on the validation split (5105 images). + +Baseline. Our baseline is a version of TAPPS that uses the same network architecture, but uses a separate set of 100 additional queries for part-level segmentation (see also Fig. 1a). Comparing the results of the baseline in Tab. 1 to existing methods in Tab. 2, we find that our baseline outperforms state-of-the-art approaches that also use separate part-level queries [18, 19], indicating that it is a strong baseline. See the supplementary material for more details. + +Evaluation metrics. The part-aware panoptic segmentation performance is evaluated using the default Part-aware Panoptic Quality (PartPQ) metric [5]. It captures both the ability to recognize and segment object-level segments (i.e., stuff regions and thing instances), and the ability to further segment the identified object-level segments into part-level masks. The PartPQ per object-level class $c$ is given by + +$$ +\operatorname {P a r t P Q} _ {c} = \frac {\sum_ {(p , g) \in T P _ {c}} \operatorname {I o U} _ {\mathrm {p}} (p , g)}{| T P _ {c} | + \frac {1}{2} | F P _ {c} | + \frac {1}{2} | F N _ {c} |}, \tag {2} +$$ + +where $TP_{c}, FP_{c}$ , and $FN_{c}$ are the sets of true positive, false positive and false negative segments, respectively. A pair of predicted and ground-truth segments $(p, g)$ of the same + +
MethodPartPQPartSQPQ
PtNo ptAllPtThStAll
ImageNet pre-training
Baseline55.238.842.971.562.037.345.7
TAPPS (ours)59.639.444.674.365.637.647.1
COCO pre-training
Baseline64.849.953.773.174.848.157.1
TAPPS (ours)67.250.454.775.175.748.557.7
+ +(a) Pascal-PP validation [1, 5, 6, 29]. + +
MethodPartPQPartSQPQ
PtNo ptAllPtThStAll
ImageNet pre-training
Baseline46.662.658.466.153.867.161.5
TAPPS (ours)48.763.159.366.855.667.362.4
COCO pre-training
Baseline48.264.860.466.756.169.063.6
TAPPS (ours)48.965.761.366.957.269.764.4
+ +(b) Cityscapes-PP val [4, 5]. + +Table 1. Main results. We compare TAPPS to a strong baseline that uses separate sets of queries to predict object- and part-level segments, instead of predicting object- and part-level segments jointly like TAPPS (see Sec. 4). + +class $c$ is part of $TP_{c}$ if the Intersection-over-Union (IoU) of their object-level masks is larger than 0.5. If a ground-truth segment is not identified, it is part of $FN_{c}$ ; if a prediction is incorrect, it is part of $FP_{c}$ . The IoUp term captures the segmentation performance within identified object-level segments. For object-level classes for which part classes are defined, it calculates the part-level mIoU. Otherwise, it uses the object-level IoU. We report the mean PartPQ over all classes, but also separately for object-level classes with parts (PartPQpt) and without parts (PartPQNoPt). + +To individually assess the ability of methods to conduct part-level segmentation within identified objects, we report the Part Segmentation Quality for object-level classes that have parts, $\mathrm{PartSQ^{Pt}}$ . Following De Geus et al. [5], per object-level class $c$ , the $\mathrm{PartSQ^{Pt}}$ calculates the average part segmentation mIoU within object-level segments: + +$$ +\operatorname {P a r t S Q} _ {c} ^ {\mathrm {P t}} = \frac {\sum_ {(p , g) \in T P _ {c}} \operatorname {I o U} _ {\mathrm {p}} (p , g)}{| T P _ {c} |}. \tag {3} +$$ + +To evaluate the ability of networks to conduct object-level panoptic segmentation, we report the Panoptic Quality (PQ) metric [13]. Similarly to the PartPQ, we report the average PQ over all classes, but also over all thing classes $(\mathrm{PQ}^{\mathrm{Th}})$ and stuff classes $(\mathrm{PQ}^{\mathrm{St}})$ separately. + +Implementation details. TAPPS is built on top of the publicly available code of state-of-the-art panoptic segmentation network Mask2Former [3]. For all datasets, we use a batch size of 16 images, and train on 4 Nvidia A100 GPUs. To optimize TAPPS, we use AdamW [26], a weight decay of 0.05, and a polynomial learning rate decay schedule with an initial learning rate of $10^{-4}$ and a power of 0.9. For Cityscapes-PP, we train for 90k iterations and apply conventional data augmentation steps [3, 18]: random flip, random resize with a factor between 0.5 and 2.0, and finally a random crop of $512 \times 1024$ pixels. For Pascal-PP, we train for 60k iterations in case of ImageNet pre-training, but for only 10k iterations in case of COCO pre-training, to prevent overfitting. Following state-of-the-art panoptic segmenta + +tion implementations on COCO [3, 21], we apply large-scale jittering with a scale between 0.1 and 2.0 followed by a random crop of $1024 \times 1024$ pixels. During inference, we resize the image such that the shortest side is 800 pixels. Note that we use exactly the same training and testing settings for both TAPPS and the baseline. For more implementation details, see the supplementary material. + +# 5. Results + +# 5.1. Main results + +First, we compare TAPPS to the strong baseline that uses separate sets of queries for object-level panoptic segmentation and part-level semantic segmentation (see Sec. 4). The results in Tab. 1a demonstrate that TAPPS significantly outperforms the baseline on Pascal-PP in several aspects. Most importantly, the PartPQ for object-level classes with parts $(\mathrm{PartPQ}^{\mathrm{Pt}})$ is $+4.4$ or $+2.4$ higher, depending on the pretraining strategy. Looking in more detail, we find that this increase is caused by two individual improvements: (1) The part-level segmentation quality $(\mathrm{PartSQ}^{\mathrm{Pt}})$ is considerably higher than for the baseline. This shows the positive impact of having a joint representation for objects and parts, and simplifying the part segmentation task by only allowing compatible predictions (see also Sec. 5.3). (2) The panoptic quality for thing classes $(\mathrm{PQ}^{\mathrm{Th}})$ also sees a substantial increase. Here, it should be noted that (a) all the object-level classes with parts are thing classes, and (b) thing classes require instance separation. Thus, this improvement does not only show the benefit of learning objects and parts jointly, but also indicates that learning parts in an object-instance-aware manner indeed improves the ability of the network to separate object instances (see also Sec. 5.4). + +For Cityscapes-PP, we see similar results in Tab. 1b. Again, we observe improvements on the PartPQ $^{\mathrm{Pt}}$ , PartSQ $^{\mathrm{Pt}}$ and PQ $^{\mathrm{Th}}$ metrics. In this case, the absolute improvements are slightly smaller. This is expected because this dataset contains significantly fewer classes and very similar images, making it easier to learn the PPS task and limiting the potential gains that can still be obtained by TAPPS. + +
MethodBackbonePre-trainingPartPQPartSQPQ
PtNo PtAllPtAll
Pascal-PP validation
JPPF [10]EfficientNet-B5I48.326.932.2--
TAPPS (ours)RN-50I59.639.444.674.347.1
Panoptic-PartFormer† [18]RN-50I,C56.138.843.266.847.6
Panoptic-PartFormer++† [19]RN-50I,C52.642.645.160.451.6
TAPPS (ours)RN-50I,C67.250.454.775.157.7
Panoptic-PartFormer† [18]Swin-BI,C64.350.654.170.858.1
Panoptic-PartFormer++† [19]Swin-BI,C48.952.151.353.059.8
TAPPS (ours)Swin-BI,C72.256.360.478.163.0
Cityscapes-PP val
Panoptic-PartFormer [18]RN-50I--54.5-57.8
Panoptic-PartFormer++ [19]RN-50I--57.5-61.6
JPPF [10]EfficientNet-B5I47.763.859.6--
TAPPS (ours)RN-50I48.763.159.366.862.4
Panoptic-PartFormer [18]RN-50I,C43.962.457.560.161.6
Panoptic-PartFormer++ [19]RN-50I,C42.565.159.2-63.6
TAPPS (ours)RN-50I,C48.965.761.366.964.4
Panoptic-PartFormer [18]Swin-BI,C45.667.862.059.066.6
Panoptic-PartFormer++ [19]Swin-BI,C46.068.262.3-68.0
Panoptic-PartFormer++ [19]ConvNeXt-BI,C46.469.163.1-68.2
SegFormer-B5 + CondInst + BPR [38–40, 45]MiT-B5 + RN-50I,C*48.667.562.5--
TAPPS (ours)Swin-BI,C53.069.064.868.068.0
+ +# 5.2. Comparison with state of the art + +In Tab. 2, we compare TAPPS to existing state-of-the-art methods across different datasets, backbones, and pretraining settings. On both Pascal-PP and Cityscapes-PP, TAPPS significantly outperforms existing work. Most importantly, it consistently scores higher on the PartPQ, PartPQ $^{\text{Pt}}$ , and PartSQ $^{\text{Pt}}$ metrics. TAPPS is only slightly outperformed by JPPF [10] on Cityscapes-PP with ImageNet pre-training, but we note that JPPF uses the EfficientNet-B5 [36] backbone, which is much more powerful than the ResNet-50 [9] used by TAPPS. Moreover, we do outperform JPPF by a large margin on the Pascal-PP dataset, showing the strength of TAPPS on more complex datasets. Overall, we achieve new state-of-the-art results on both datasets, obtaining PartPQ scores of 60.4 and 64.8 on Pascal-PP and Cityscapes-PP, improvements of +6.3 and +1.7, respectively. In the supplementary material, we compare qualitative examples of TAPPS, our baseline, and existing approaches, and we also show typical failure cases. + +# 5.3. Ablations + +Predicting only compatible parts. Using shared queries for object- and part-level segmentation enables TAPPS to only allow part segmentation predictions that + +Table 2. Comparison with state of the art. Evaluation on the Cityscapes-PP and Pascal-PP benchmarks [4-6]. RN-50 is ResNet-50 [9]. Other backbones are EfficientNet-B5 [36], MiT-B5 [45], Swin-B [24] and ConvNeXt-B [25]. I = ImageNet [33], C = COCO panoptic [21], C* = COCO pre-training for instance segmentation. †PartPQ scores for these existing methods have been re-evaluated using official code [5] and are higher than originally reported on Pascal-PP [18, 19], see supplementary material for more details. + +
Predict compat. partsPartPQPartSQPQObj. w/o conflicts
TrainingTestingPtNo PtAllPtAll
ImageNet pre-training
Baseline55.238.842.971.545.766.1%
XX57.939.243.972.946.699.7%
X✓58.039.243.972.946.6100.0%
✓✓59.639.444.674.347.1100.0%
COCO pre-training
Baseline64.849.953.773.157.164.6%
XX66.350.554.674.057.999.5%
X✓66.450.554.674.057.9100.0%
✓✓67.250.454.775.157.7100.0%
+ +Table 3. Predicting only compatible part classes during training and testing. Evaluated on Pascal-PP [1, 5, 6, 29]. + +are compatible with the query's object-level class (see Sec. 3.3.2). In Tab. 3, we evaluate the effect of predicting only compatible parts during both training and testing. In addition to the main metrics, we also assess the percentage of predicted objects for which there are no conflicting part-level predictions. The results show that, even when allowing incompatible predictions, TAPPS has significantly fewer object-part conflicts than the baseline. Naturally, when predicting only compatible parts during testing, all object-part conflicts are removed, eliminating the need + +
Adaptation layersPartPQPartSQPQ
PtNo PtAllPtAll
-67.349.654.175.257.1
1×FC67.150.354.675.157.6
2-layer MLP67.250.454.775.157.7
3-layer MLP67.150.454.675.157.7
+ +Table 4. JOPS head architecture (Fig. 2). Evaluated on Pascal-PP [1, 5, 6, 29], with pre-training on COCO panoptic [21]. + +
MethodmIoUThPQTh
Baseline69.862.0
TAPPS (ours)69.965.6
+ +(a) Pascal-PP [1, 5, 6, 29]. + +
MethodmIoUThPQTh
Baseline77.653.8
TAPPS (ours)78.255.6
+ +(b) Cityscapes-PP [4, 5]. + +for post-processing. However, this does not yield a big improvement in terms of the segmentation quality. When we also apply this during training, we do observe a part segmentation quality $(\mathrm{PartSQ^{Pt}})$ improvement. This shows that simplifying the part segmentation task during training leads to improved performance, as we hypothesized in Sec. 3.3.2. + +JOPS head architecture. In Tab. 4, we evaluate the impact of using different numbers of adaptation layers in the JOPS head before applying the $N^{\mathrm{pc}}$ fully-connected (FC) layers to generate the part queries (see Fig. 2). We find that one or two adaptation layers are necessary to generate object- and part-level representations that are sufficiently distinct to perform their respective tasks accurately; adding any more layers does not yield further improvements. + +# 5.4. Additional analyses + +Instance separability. Tab. 1 showed that TAPPS improves the $\mathrm{PQ^{\mathrm{Th}}}$ , i.e., the ability to recognize, segment and classify object instances. To assess if this is due to better object recognition or improved instance separability, we group all thing instances of the same object-level class together and evaluate the instance-agnostic semantic segmentation performance with the mIoU $^{\mathrm{Th}}$ . If the $\mathrm{PQ^{\mathrm{Th}}}$ improvement were due to improved recognition, we expect the mIoU $^{\mathrm{Th}}$ to improve too. However, the results in Tab. 5 show only a minor mIoU $^{\mathrm{Th}}$ improvement. This indicates that the $\mathrm{PQ^{\mathrm{Th}}}$ gain is not due to better recognition, but mainly results from a better ability to separate instances, showing the benefit of learning parts in an object-instance-aware manner. + +Performance on Pascal-PP-107. To assess the performance of TAPPS in a more complex setting, we evaluate it on Pascal-PP-107, which has 107 part-level classes instead of 57. The results on this dataset, reported in Tab. 6, show once more that TAPPS consistently improves the part segmentation and thing segmentation performance with respect to the baseline, in this case leading to a PartPQ $^{\text{Pt}}$ improvement of +4.1 or +2.2, depending on the pre-training strategy. This demonstrates that TAPPS is also effective + +Table 5. Performance for things. ImageNet pre-training [33]. + +
MethodPartPQPartSQPQ
PtNo PtAllPtThStAll
ImageNet pre-training
Baseline45.738.940.659.362.137.645.9
TAPPS (ours)49.839.342.061.765.837.747.2
COCO pre-training
Baseline53.249.850.759.975.248.157.2
TAPPS (ours)55.450.051.462.175.648.157.4
+ +Table 6. 107 part-level categories. Results on the more challenging Pascal-PP-107 dataset [1, 5, 6, 27, 29]. + +
MethodPartPQPartSQPQ
PtNo PtAllPtAll
Dynamic65.150.554.273.157.7
Fixed67.250.454.775.157.7
+ +Table 7. Fixed or dynamic part segmentation. Evaluated on Pascal-PP [1, 5, 6, 29], with pre-training on COCO panoptic [21]. + +when the PPS task becomes more complex. + +Fixed vs. dynamic part segmentation. As explained in Sec. 3.3.2, TAPPS uses a fixed fully connected layer for each part class to generate the corresponding 'per-object' part query. Alternatively, it is possible to predict part masks and classes dynamically within each object segment, using a set of dynamic queries like we do for object-level segmentation. In Tab. 7, we compare our fixed part segmentation setting with this dynamic approach, which is explained in more detail in the supplementary material. We find that our fixed approach results in a better $\mathrm{PartSQ^{Pt}}$ and therefore $\mathrm{PartPQ^{Pt}}$ performance. We hypothesize that the dynamic approach performs worse because it makes the part segmentation task unnecessarily complex, as the network needs to learn to assign segments to queries dynamically, and additionally predict a class label. + +# 6. Conclusion + +With experiments, we have shown that TAPPS considerably outperforms methods that predict objects and parts separately, by improving the object instance separability, part segmentation quality, and object-part compatibility. Importantly, these improvements can be attributed to the fact that TAPPS is directly optimized for the PPS task, using a set of shared queries to jointly predict objects and corresponding parts. With our promising findings, we hope to inspire future research towards even more complete scene understanding, e.g., image segmentation at even more abstraction levels, potentially with more flexible class hierarchies. + +Acknowledgements. This work is supported by Eindhoven Engine, NXP Semiconductors, and Brainport Eindhoven. This work made use of the Dutch national e-infrastructure with the support of the SURF Cooperative using grant no. EINF-5302, which is financed by the Dutch Research Council (NWO). + +# References + +[1] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect What You Can: Detecting and Representing Objects using Holistic Models and Body Parts. In CVPR, 2014. 5, 6, 7, 8 +[2] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-Pixel Classification is Not All You Need for Semantic Segmentation. In NeurIPS, 2021. 3 +[3] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. In CVPR, 2022. 3, 5, 6 +[4] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes Dataset for Semantic Urban Scene Understanding. In CVPR, 2016. 5, 6, 7, 8 +[5] Daan de Geus, Panagiotis Meletis, Chenyang Lu, Xiaoxiao Wen, and Gijs Dubbelman. Part-aware Panoptic Segmentation. In CVPR, 2021. 1, 2, 3, 4, 5, 6, 7, 8 +[6] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes (VOC) Challenge. *IJCV*, 88(2):303–338, 2010. 5, 6, 7, 8 +[7] Ke Gong, Xiaodan Liang, Yicheng Li, Yimin Chen, Ming Yang, and Liang Lin. Instance-level human parsing via part grouping network. In ECCV, 2018. 3 +[8] Ke Gong, Yiming Gao, Xiaodan Liang, Xiaohui Shen, Meng Wang, and Liang Lin. Graphonomy: Universal human parsing via graph transfer learning. In CVPR, 2019. 3 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 3, 7 +[10] Sravan Kumar Jagadeesh, René Schuster, and Didier Stricker. Multi-task Fusion for Efficient Panoptic-Part Segmentation. In ICPRAM, 2022. 2, 7 +[11] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. In CVPR, 2023. 3 +[12] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollar. Panoptic Feature Pyramid Networks. In CVPR, 2019. 3 +[13] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic Segmentation. In CVPR, 2019. 1, 2, 3, 6 +[14] Jianshu Li, Jian Zhao, Yunpeng Chen, Sujoy Roy, Shuicheng Yan, Jiashi Feng, and Terence Sim. Multi-human parsing machines. In ACM MM, 2018. 3 +[15] Liulei Li, Tianfei Zhou, Wenguan Wang, Jianwu Li, and Yi Yang. Deep Hierarchical Semantic Segmentation. In CVPR, 2022. +[16] Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. Self-Correction for Human Parsing. IEEE TPAMI, 44(6):3260-3271, 2022. 3 +[17] Qizhu Li, Anurag Arnab, and Philip HS Torr. Holistic, instance-level human parsing. In BMVC, 2017. 3 + +[18] Xiangtai Li, Shilin Xu, Yibo Yang, Guangliang Cheng, Yunhai Tong, and Dacheng Tao. Panoptic-PartFormer: Learning a Unified Model for Panoptic Part Segmentation. In ECCV, 2022. 1, 2, 3, 5, 6, 7 +[19] Xiangtai Li, Shilin Xu, Yibo Yang, Haobo Yuan, Guangliang Cheng, Yunhai Tong, Zhouchen Lin, and Dacheng Tao. PanopticPartFormer++: A Unified and Decoupled View for Panoptic Part Segmentation. arXiv preprint arXiv:2301.00954, 2023. 1, 2, 3, 5, 7 +[20] Xiaodan Liang, Ke Gong, Xiaohui Shen, and Liang Lin. Look into Person: Joint Body Parsing & Pose Estimation Network and A New Benchmark. IEEE TPAMI, 41(4):871-885, 2018. 3 +[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 6, 7, 8 +[22] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature Pyramid Networks for Object Detection. In CVPR, 2017. 3 +[23] Qing Liu, Adam Kortylewski, Zhishuai Zhang, Zizhang Li, Mengqi Guo, Qihao Liu, Xiaoding Yuan, Jiteng Mu, Weichao Qiu, and Alan Yuille. Learning Part Segmentation Through Unsupervised Domain Adaptation From Synthetic Vehicles. In CVPR, 2022. 3 +[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. In ICCV, 2021. 3, 7 +[25] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s. In CVPR, 2022. 7 +[26] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICLR, 2019. 6 +[27] Umberto Michieli, Edoardo Borsato, Luca Rossi, and Pietro Zanuttigh. GMNet: Graph Matching Network for Large Scale Part Semantic Segmentation in the Wild. In ECCV, 2020. 3, 5, 8 +[28] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentatio. In 3DV, 2016. 5 +[29] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The Role of Context for Object Detection and Semantic Segmentation in the Wild. In CVPR, 2014. 5, 6, 7, 8 +[30] Tai-Yu Pan, Qing Liu, Wei-Lun Chao, and Brian Price. Towards Open-World Segmentation of Parts. In CVPR, 2023. 3 +[31] Lu Qi, Jason Kuen, Weidong Guo, Jiumiang Gu, Zhe Lin, Bo Du, Yu Xu, and Ming-Hsuan Yang. AIMS: All-Inclusive Multi-Level Segmentation for Anything. In NeurIPS, 2023. 3 +[32] Tao Ruan, Ting Liu, Zilong Huang, Yunchao Wei, Shikui Wei, and Yao Zhao. Devil in the details: Towards accurate single and multiple human parsing. In AAAI, 2019. 3 + +[33] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 115(3):211-252, 2015. 7, 8 +[34] Rishubh Singh, Pranav Gupta, Pradeep Shenoy, and Ravikiran Sarvadevabhatla. FLOAT: Factorized Learning of Object Attributes for Improved Multi-Object Multi-Part Scene Parsing. In CVPR, 2022. 3 +[35] Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation. In ICCV, 2023. 3 +[36] Mingxing Tan and Quoc Le. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. In ICML, 2019. 7 +[37] Xin Tan, Jiachen Xu, Zhou Ye, Jinkun Hao, and Lizhuang Ma. Confident Semantic Ranking Loss for Part Parsing. In ICME, 2021. 3 +[38] Chufeng Tang, Hang Chen, Xiao Li, Jianmin Li, Zhaoxiang Zhang, and Xiaolin Hu. Look Closer To Segment Better: Boundary Patch Refinement for Instance Segmentation. In CVPR, 2021. 7 +[39] Chufeng Tang, Lingxi Xie, Xiaopeng Zhang, Xiaolin Hu, and Qi Tian. Visual Recognition by Request. In CVPR, 2023. 2 +[40] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional Convolutions for Instance Segmentation. In ECCV, 2020. 7 +[41] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: End-to-End Panoptic Segmentation With Mask Transformers. In CVPR, 2021. 3 +[42] Wenguan Wang, Hailong Zhu, Jifeng Dai, Yanwei Pang, Jianbing Shen, and Ling Shao. Hierarchical human parsing with typed part-relation reasoning. In CVPR, 2020. 3 +[43] Meng Wei, Xiaoyu Yue, Wenwei Zhang, Shu Kong, Xihui Liu, and Jiangmiao Pang. OV-PARTS: Towards Open-Vocabulary Part Segmentation. In NeurIPS, 2023. 3 +[44] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, 2018. 3 +[45] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers. In NeurIPS, 2021. 7 +[46] Lu Yang, Qing Song, Zhihui Wang, and Ming Jiang. Parsing R-CNN for Instance-Level Human Analysis. In CVPR, 2019. 3 +[47] Lu Yang, Qing Song, Zhihui Wang, Mengjie Hu, Chun Liu, Xueshi Xin, Wenhe Jia, and Songcen Xu. Renovating Parsing R-CNN for Accurate Multiple Human Parsing. In ECCV, 2020. +[48] Lu Yang, Qing Song, Zhihui Wang, Mengjie Hu, and Chun Liu. Hier R-CNN: Instance-Level Human Parts Detection and A New Benchmark. IEEE TIP, 30:39-54, 2021. +[49] Sanyi Zhang, Xiaochun Cao, Guo-Jun Qi, Zhanjie Song, and Jie Zhou. AIParsing: Anchor-Free Instance-Level Human Parsing. IEEE TIP, 31:5599-5612, 2022. 3 + +[50] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-Net: Towards Unified Image Segmentation. In NeurIPS, 2021. 3 +[51] Jian Zhao, Jianshu Li, Yu Cheng, Terence Sim, Shuicheng Yan, and Jiashi Feng. Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for multi-human parsing. In ACM MM, 2018. 3 +[52] Yifan Zhao, Jia Li, Yu Zhang, and Yonghong Tian. Multi-Class Part Parsing With Joint Boundary-Semantic Awareness. In ICCV, 2019. 3 \ No newline at end of file diff --git a/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/images.zip b/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..98c77f934fd582732f9b7e9cf6ca03458f503f23 --- /dev/null +++ b/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76e2a688b000e9bb3b18f1f632104a4e07bb04ef11faab1221dde6434f10dba4 +size 391265 diff --git a/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/layout.json b/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8b8668224bb5a9a9917082c924c23de79a93f9c8 --- /dev/null +++ b/2024/Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations/layout.json @@ -0,0 +1,8517 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 103, + 446, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 103, + 446, + 140 + ], + "spans": [ + { + "bbox": [ + 148, + 103, + 446, + 140 + ], + "type": "text", + "content": "Task-aligned Part-aware Panoptic Segmentation through Joint Object-Part Representations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 206, + 161, + 386, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 161, + 386, + 175 + ], + "spans": [ + { + "bbox": [ + 206, + 161, + 386, + 175 + ], + "type": "text", + "content": "Daan de Geus Gijs Dubbelman" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 206, + 176, + 386, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 176, + 386, + 190 + ], + "spans": [ + { + "bbox": [ + 206, + 176, + 386, + 190 + ], + "type": "text", + "content": "Eindhoven University of Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 209, + 191, + 381, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 191, + 381, + 203 + ], + "spans": [ + { + "bbox": [ + 209, + 191, + 381, + 203 + ], + "type": "text", + "content": "{d.c.d.geus, g.dubbelman}@tue.nl" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 258, + 290, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 510 + ], + "type": "text", + "content": "Part-aware panoptic segmentation (PPS) requires (a) that each foreground object and background region in an image is segmented and classified, and (b) that all parts within foreground objects are segmented, classified and linked to their parent object. Existing methods approach PPS by separately conducting object-level and part-level segmentation. However, their part-level predictions are not linked to individual parent objects. Therefore, their learning objective is not aligned with the PPS task objective, which harms the PPS performance. To solve this, and make more accurate PPS predictions, we propose Task-Aligned Part-aware Panoptic Segmentation (TAPPS). This method uses a set of shared queries to jointly predict (a) object-level segments, and (b) the part-level segments within those same objects. As a result, TAPPS learns to predict part-level segments that are linked to individual parent objects, aligning the learning objective with the task objective, and allowing TAPPS to leverage joint object-part representations. With experiments, we show that TAPPS considerably outperforms methods that predict objects and parts separately, and achieves new state-of-the-art PPS results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 536, + 128, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 128, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 128, + 548 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": "To fully understand what is depicted in an image, it is important to consider concepts at different levels of abstraction. For rich scene understanding, we should not only recognize foreground objects (e.g., car, human) and background regions (e.g., sky, ocean), but also simultaneously identify the parts that constitute the objects (e.g., car-wheel, human-arm). In a step towards such comprehensive scene understanding, De Geus et al. introduced part-aware panoptic segmentation (PPS) [5]. The objective of this computer vision task is (1) to output a segmentation mask and class label for all thing objects and stuff regions in an image like for panoptic segmentation [13] – we call these object-level segments – and (2) to simultaneously segment and classify" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 250, + 542, + 350 + ], + "blocks": [ + { + "bbox": [ + 310, + 229, + 440, + 239 + ], + "lines": [ + { + "bbox": [ + 310, + 229, + 440, + 239 + ], + "spans": [ + { + "bbox": [ + 310, + 229, + 440, + 239 + ], + "type": "text", + "content": "(a) Separate object and part segmentation" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 239, + 455, + 248 + ], + "lines": [ + { + "bbox": [ + 310, + 239, + 455, + 248 + ], + "spans": [ + { + "bbox": [ + 310, + 239, + 455, + 248 + ], + "type": "text", + "content": "X Learning objective not aligned with task objective" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 250, + 542, + 350 + ], + "lines": [ + { + "bbox": [ + 307, + 250, + 542, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 542, + 350 + ], + "type": "image", + "image_path": "f3ebccf1cd7a24734832cc79991c8b6d16711034356de08e06fb8f2159a81101.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 354, + 448, + 363 + ], + "lines": [ + { + "bbox": [ + 310, + 354, + 448, + 363 + ], + "spans": [ + { + "bbox": [ + 310, + 354, + 448, + 363 + ], + "type": "text", + "content": "(b) Joint object and part segmentation (ours)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 306, + 375, + 544, + 468 + ], + "blocks": [ + { + "bbox": [ + 310, + 364, + 444, + 373 + ], + "lines": [ + { + "bbox": [ + 310, + 364, + 444, + 373 + ], + "spans": [ + { + "bbox": [ + 310, + 364, + 444, + 373 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 310, + 364, + 444, + 373 + ], + "type": "text", + "content": " Learning objective aligned with task objective" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 375, + 544, + 468 + ], + "lines": [ + { + "bbox": [ + 306, + 375, + 544, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 375, + 544, + 468 + ], + "type": "image", + "image_path": "ba6f1c42782fe4a17c1b152ce6884b5e6aec85c0da13b1681eb1054306f72945.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 474, + 545, + 541 + ], + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 541 + ], + "type": "text", + "content": "Figure 1. Task-aligned part-aware panoptic segmentation. (a) Existing works separately predict object-level segments and object-instance-unaware part-level segments. (b) In this work, we predict objects and parts jointly, using a set of shared queries. This allows our method to predict parts within individual object segments, aligning its learning objective with the PPS task objective." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 557, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 593 + ], + "type": "text", + "content": "all parts within each identified object. These are called part-level segments and should be explicitly linked to an object-level segment, establishing a part-whole relation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "Current state-of-the-art methods [18, 19] address PPS by using two different sets of learnable queries to separately predict object-level and part-level segments, see Fig. 1a. While these methods outperform earlier baselines [5], they have one main limitation: their learning objective is not aligned with the task objective. Where the PPS task objective is to predict parts within each individual object-level segment, these existing works conduct part-level semantic segmentation, predicting part-level masks that cover multiple objects (see Fig. 1a). In other words, these networks" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3174" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "are not optimized for the PPS task, but instead solve the surrogate subtasks of object-level panoptic segmentation and part-level semantic segmentation. As a result, to assign parts to individual objects, these methods require post-processing. Additionally, we hypothesize that there are several other negative consequences: (1) These networks learn a conflicting feature representation. They learn that object-level thing instances should be separated, but also that the parts of these object-level instances should be grouped together. We expect that this harms the ability of the networks to separate instances. (2) Predicting objects and parts separately may cause incompatible predictions (e.g., a car-window part with a bicycle object), which makes it unclear which prediction to trust, and requires further post-processing. (3) The networks encode information about objects and their parts separately, whereas this information is potentially complementary. See Sec. 3.2.2 for more details on these limitations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 290, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 289, + 434 + ], + "type": "text", + "content": "In this work, we aim to design a simple network for PPS that overcomes these limitations and thereby makes more accurate PPS predictions. To achieve this, we propose the Task-Aligned Part-aware Panoptic Segmentation (TAPPS) method. Instead of using separate queries for objects and parts, TAPPS uses one set of shared queries to jointly represent objects and the parts they contain. Specifically, each of these queries learns to represent at most one object-level segment, for which it predicts (1) a segmentation mask and object-level class, and (2) the segmentation masks and classes for all part-level segments within this object. This is visualized in Fig. 1b and explained in Sec. 3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 435, + 289, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 289, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 289, + 652 + ], + "type": "text", + "content": "With this approach, TAPPS explicitly predicts part-level segments per individual object. Therefore, the network's learning objective is now aligned with the PPS task objective, and TAPPS is directly optimized for the PPS task. As a result, both object- and part-level segmentation are learned in an 'object-instance-aware' manner. This removes the conflict in the learned feature representations, allowing for better instance separability. Additionally, we hypothesize that predicting objects and parts from the same query reduces incompatibilities between object and part predictions, as they are made using the same information. Moreover, we use TAPPS to go even further, and explicitly only predict part segments compatible with predicted object segments. This enforces full object-part compatibility and simplifies the part segmentation task. Finally, the network can now encode complementary object-level and part-level information in a shared query, giving it a richer representation for making object- and part-level segmentation predictions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": "With experiments, detailed in Sec. 4, we show that TAPPS outperforms the baseline with separate object and part queries in multiple aspects: (1) The part segmentation quality within identified objects is significantly better, which shows the positive impact of using joint object-part" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "representations and simplifying the part segmentation task. (2) The object instance segmentation performance is considerably improved, demonstrating the benefits of learning parts in an object-instance-aware manner. Together, this yields a large overall improvement with respect to the baseline, and causes TAPPS to considerably outperform existing works across different benchmarks, achieving new state-of-the-art results. See Sec. 5 for more extensive results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 170, + 531, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 170, + 531, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 170, + 531, + 182 + ], + "type": "text", + "content": "To summarize, we make the following contributions:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 183, + 545, + 302 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 183, + 545, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 183, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 306, + 183, + 545, + 231 + ], + "type": "text", + "content": "- We propose TAPPS, a simple approach for PPS that aligns the learning objective with the task objective, facilitating object instance separability and enabling joint object-part representations for more accurate PPS predictions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 231, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 231, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 306, + 231, + 545, + 278 + ], + "type": "text", + "content": "- We use the shared object-part queries to constrain TAPPS to only predict part segments that are compatible with the object class, enforcing object-part compatibility and simplifying the part segmentation task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 278, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 278, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 306, + 278, + 545, + 302 + ], + "type": "text", + "content": "- We experimentally show the effectiveness of TAPPS across a range of datasets and network configurations." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 304, + 545, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 328 + ], + "type": "text", + "content": "The code for TAPPS is made publicly available through https://tue-mps.github.io/tapps/." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 343, + 390, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 343, + 390, + 356 + ], + "spans": [ + { + "bbox": [ + 306, + 343, + 390, + 356 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 365, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 653 + ], + "type": "text", + "content": "Part-aware panoptic segmentation. Part-aware panoptic segmentation [5] is an image segmentation task introduced for scene understanding at multiple abstraction levels. It extends panoptic segmentation [13] by requiring (a) object-level panoptic segmentation, and (b) part-level segmentation within object-level segments. Most existing works do not predict parts per object-level segment, but instead make separate predictions for panoptic segmentation and the surrogate task of part-level semantic segmentation - or part segmentation [10, 18, 19]. JPPF [10] proposes a single-network approach with a shared encoder followed by separate heads for semantic, instance, and part segmentation. To merge the predictions by these heads to the PPS format, JPPF proposes a new rule-based fusion strategy that outperforms the originally introduced merging strategy for PPS [5]. Panoptic-PartFormer and its extension Panoptic-PartFormer++ [18, 19] tackle PPS with a Transformer-based approach that independently makes panoptic segmentation and part segmentation predictions using separate sets of learnable queries for thing, stuff and part segments. Alternatively, ViRReq [39] is a paradigm in which PPS is predicted in a cascading fashion, by first segmenting objects and then segmenting the parts within these objects 'by request', but it requires multiple networks to achieve this." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": "In contrast to these existing approaches, we propose a method that jointly predicts object-level segments and the part-level segments within these objects. This aligns the learning objective with the task objective, and yields improved PPS performance. For more details, see Sec. 3." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3175" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "text", + "content": "Part-level image segmentation. Part-level segmentation is also widely studied beyond PPS. Most works address part segmentation, i.e., semantic segmentation for parts [7, 8, 14-17, 20, 23, 27, 32, 34, 35, 37, 42, 43, 51, 52]. For more complete scene understanding, other works also take into account object instances, like PPS does, with instance-aware part segmentation [16, 30, 32, 46-49]. However, these works do not consider background stuff classes, which PPS does consider. Alternatively, UPerNet [44] separately predicts part-level and object-level semantic segmentation, including background classes, but without any instance awareness. Another work has extended PPS by also predicting the relations between objects [31], training a model on annotations from different datasets. In this work, we address the PPS task as introduced by De Geus et al. [5], and we compare our approach to state-of-the-art alternatives." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 277, + 104, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 104, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 104, + 289 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 298, + 159, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 298, + 159, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 159, + 310 + ], + "type": "text", + "content": "3.1. PPS task definition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "text", + "content": "Part-aware panoptic segmentation (PPS) [5] requires consistent image segmentation across two abstraction levels: the object level and part level. For object-level segmentation, following panoptic segmentation [13], an image should be divided into " + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "text", + "content": " object-level segments, where " + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "text", + "content": " varies per image. Each segment " + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "text", + "content": " consists of a binary mask " + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i^{\\mathrm{obj}}" + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "text", + "content": " and an object class label " + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "inline_equation", + "content": "c_i^{\\mathrm{obj}}" + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "text", + "content": ". Thing classes require a segment per object instance, and stuff classes require a single segment per class. Next, per segment " + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "text", + "content": ", PPS requires a set of " + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "inline_equation", + "content": "K_i" + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "text", + "content": " part-level segments. Each part-level segment " + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "inline_equation", + "content": "s_{i,j}^{\\mathrm{pt}}" + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "text", + "content": " consists of a binary part-level mask " + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{i,j}^{\\mathrm{pt}}" + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "text", + "content": " and a part-level class " + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "inline_equation", + "content": "c_{i,j}^{\\mathrm{pt}}" + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "text", + "content": " that are compatible with the object-level mask and class. Specifically, the part-level mask must be a subset of the object-level mask, and the part-level class should be one of the part-level classes defined for the object-level class. For instance, if " + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "inline_equation", + "content": "c_i^{\\mathrm{obj}}" + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "text", + "content": " is car, then " + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "inline_equation", + "content": "c_{i,j}^{\\mathrm{pt}}" + }, + { + "bbox": [ + 47, + 318, + 289, + 537 + ], + "type": "text", + "content": " cannot be human-head but it can be car-window. For all parts, a single part-level segment is required for each part class within an object." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 539, + 288, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 288, + 624 + ], + "type": "text", + "content": "To summarize, each image should be divided into a set of segments " + }, + { + "bbox": [ + 47, + 539, + 288, + 624 + ], + "type": "inline_equation", + "content": "S = \\{s_i\\}_{i=1}^N = \\{(\\mathbf{M}_i^{\\mathrm{obj}}, c_i^{\\mathrm{obj}}, \\mathcal{P}_i)\\}_{i=1}^N" + }, + { + "bbox": [ + 47, + 539, + 288, + 624 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 539, + 288, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i = \\{s_{i,j}^{\\mathrm{pt}}\\}_{j=1}^{K_i} = \\{(\\mathbf{M}_{i,j}^{\\mathrm{pt}}, c_{i,j}^{\\mathrm{pt}})\\}_{j=1}^{K_i}" + }, + { + "bbox": [ + 47, + 539, + 288, + 624 + ], + "type": "text", + "content": " is a set of part-level segments compatible with the object-level segment in " + }, + { + "bbox": [ + 47, + 539, + 288, + 624 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 47, + 539, + 288, + 624 + ], + "type": "text", + "content": ". In practice, part-level classes are only defined for some object-level classes. When segment " + }, + { + "bbox": [ + 47, + 539, + 288, + 624 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 47, + 539, + 288, + 624 + ], + "type": "text", + "content": " has an object class that does not have parts, then " + }, + { + "bbox": [ + 47, + 539, + 288, + 624 + ], + "type": "inline_equation", + "content": "K_i = 0" + }, + { + "bbox": [ + 47, + 539, + 288, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 539, + 288, + 624 + ], + "type": "inline_equation", + "content": "s_i = (\\mathbf{M}_i^{\\mathrm{obj}}, c_i^{\\mathrm{obj}})" + }, + { + "bbox": [ + 47, + 539, + 288, + 624 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 634, + 133, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 133, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 133, + 645 + ], + "type": "text", + "content": "3.2. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "text", + "content": "Like existing state-of-the-art PPS approaches [18, 19], our work uses the powerful mask classification metaarchitecture that is the foundation of many state-of-the-art segmentation models [2, 3, 11, 41, 50]. Here, we first explain this mask classification paradigm in more detail." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": "Then, we explain how existing works use it for PPS, and what the limitations are of these approaches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 110, + 470, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 110, + 470, + 122 + ], + "spans": [ + { + "bbox": [ + 306, + 110, + 470, + 122 + ], + "type": "text", + "content": "3.2.1 Mask classification framework" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": "The concept of mask classification is to predict a set of " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "N^{\\mathrm{q}}" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": " object-level segments, i.e., a set of " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "N^{\\mathrm{q}}" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": " pixel-level masks " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}^{\\mathrm{obj}}" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": " and corresponding class labels " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "\\hat{c}^{\\mathrm{obj}}" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": ". To make these predictions, mask classification networks output two components: high-resolution features " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{E \\times H \\times W}" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": " and queries " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in \\mathbb{R}^{N^{\\mathrm{q}} \\times E}" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": " are the height and width of the features, and " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": " is the feature and query dimensionality. Each of the " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "N^{\\mathrm{q}}" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": " queries is used to predict the class label and segmentation mask of at most one object-level segment. The class predictions are a function of only the queries, and the mask predictions are a function of both the queries " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": " and the high-resolution features " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": ". To output these high-resolution features, the input image is first fed into a backbone to extract multi-scale features, e.g., ResNet [9] or Swin [24]. Subsequently, these features are further refined and upsampled to high-resolution features by a pixel decoder, e.g., Semantic FPN [12, 22]. In the other part of the network, the queries " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": " are generated by processing learnable queries " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{o}}" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": " using a transformer decoder, which applies self-attention across queries and cross-attention with image features. Through bipartite matching between " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "N^{\\mathrm{q}}" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": " predicted and " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": " ground-truth segments, each query is assigned to at most one ground-truth segment. As " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "N^{\\mathrm{q}}" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": " is not always equal to " + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 129, + 547, + 466 + ], + "type": "text", + "content": ", some queries may not be assigned to a ground-truth segment. If so, they do not receive any supervision for segmentation, and will learn a 'no object' class. In this work, we build upon the Mask2Former [3] instantiation of this meta-architecture." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 479, + 455, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 455, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 455, + 491 + ], + "type": "text", + "content": "3.2.2 Mask classification for PPS" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 498, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 545 + ], + "type": "text", + "content": "With the aforementioned meta-architecture, we can conduct object-level segmentation, and thereby solve one aspect of the PPS task definition. To solve the full task, the object-level segments should also be further segmented into parts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 546, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 653 + ], + "type": "text", + "content": "Current state-of-the-art approaches [18, 19] achieve this by introducing an additional set of queries, the part-level queries depicted in Fig. 1a. Each of these part-level queries learns to represent a part-level segment. However, in contrast to the PPS task definition, these part-level segments are not explicitly linked to an object-level segment. Instead, these part-level segments represent an entire part-level class, i.e., their masks contain all pixels belonging to one part-level class across multiple object segments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "This means that the learning objective of these networks is not aligned with the PPS task objective. As a result, they are not directly optimized for PPS, but instead for two surrogate tasks: object-level panoptic segmentation and part-level semantic segmentation. In addition to necessitating" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3176" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 288, + 219 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 288, + 219 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 288, + 219 + ], + "type": "image", + "image_path": "2dbf6ba2e62855a536b9cb4c4f8343f012d40b9d8d571649bc4ea437fee07eac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 224, + 546, + 269 + ], + "lines": [ + { + "bbox": [ + 45, + 224, + 546, + 269 + ], + "spans": [ + { + "bbox": [ + 45, + 224, + 546, + 269 + ], + "type": "text", + "content": "Figure 2. Network architecture. Left: The overall TAPPS architecture. A set of learnable queries and features from a backbone are fed into a pixel decoder and transformer decoder to generate high-resolution features and processed queries. Right: These queries and features are fed into the JOPS head, which predicts for each shared query (a) an object-level class, (b) an object-level segmentation mask, and (c) a set of part-level masks for the part-level classes compatible with the object-level class. Operator " + }, + { + "bbox": [ + 45, + 224, + 546, + 269 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 45, + 224, + 546, + 269 + ], + "type": "text", + "content": " denotes a matrix multiplication." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 294, + 68, + 547, + 218 + ], + "blocks": [ + { + "bbox": [ + 294, + 68, + 547, + 218 + ], + "lines": [ + { + "bbox": [ + 294, + 68, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 294, + 68, + 547, + 218 + ], + "type": "image", + "image_path": "533af94f2cab23e2f80e9b3df42fd2f2c2d7c4f976286e49647ab30c269f6ef7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 289, + 289, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 540 + ], + "type": "text", + "content": "post-processing, we hypothesize that this has several negative consequences: (1) By learning parts in an 'object-instance-unaware' manner, the network learns conflicting representations in the features, F. On the one hand, different thing instances belong to different object-level segments, requiring distinct features. On the other hand, the parts of these instances belong to the same part-level segment, requiring similar features. We hypothesize that this conflict harms the ability of the network to learn to separate object-level thing instances. (2) Separately predicting objects and parts in this way may also cause incompatible predictions. These incompatibilities make it unclear for downstream processes which prediction to trust. To obtain predictions compliant with the PPS definition, rule-based post-processing is required, which discards incompatible predictions [5]. This is undesired because the discarded information is potentially correct. (3) By representing object- and part-level information with separate queries, the network cannot encode complementary information about an individual object and its parts, which it could use to make more informed predictions at these abstraction levels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 551, + 152, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 152, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 152, + 563 + ], + "type": "text", + "content": "3.3. Task-aligned PPS" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "To overcome the limitations of state-of-the-art methods and achieve a better PPS performance, the objective of this work is to design a simple PPS approach that uses the powerful mask classification framework, but that aligns its learning objective with the PPS task objective, and explicitly predicts part-level segments within individual objects. To achieve this, we should (a) generate a unique query for each object-part combination, i.e., for each part-level segment within an object-level segment, and (b) explicitly link these part-level segments to their parent objects, as PPS requires. To accomplish this, we propose Task-Aligned Part-aware Panoptic Segmentation (TAPPS), a method that jointly represents" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 289, + 546, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 546, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 546, + 350 + ], + "type": "text", + "content": "object-level segments and their parts with a set of shared queries, and generates 'per-object' part queries from each shared query. Each of these queries can then be used to predict one part-level segment within an object-level segment, and is automatically linked to its parent object segment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 369, + 425, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 425, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 425, + 380 + ], + "type": "text", + "content": "3.3.1 Overall architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 389, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 569 + ], + "type": "text", + "content": "In Fig. 2, we visualize the network architecture of TAPPS. First, we initialize a set of " + }, + { + "bbox": [ + 304, + 389, + 545, + 569 + ], + "type": "inline_equation", + "content": "N^{\\mathbf{q}}" + }, + { + "bbox": [ + 304, + 389, + 545, + 569 + ], + "type": "text", + "content": " learnable queries, " + }, + { + "bbox": [ + 304, + 389, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^0 \\in \\mathbb{R}^{N^{\\mathbf{q}} \\times E}" + }, + { + "bbox": [ + 304, + 389, + 545, + 569 + ], + "type": "text", + "content": ". Each of these queries learns to represent one object-level segment, but also the part-level segments belonging to this object (e.g., car3 and also car3-wheels, car3-windows, etc.). Following the mask classification framework described in Sec. 3.2, these learnable queries " + }, + { + "bbox": [ + 304, + 389, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^0" + }, + { + "bbox": [ + 304, + 389, + 545, + 569 + ], + "type": "text", + "content": " and the features from a backbone are fed into a decoder, which generates a set of processed queries " + }, + { + "bbox": [ + 304, + 389, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 304, + 389, + 545, + 569 + ], + "type": "text", + "content": " and high-resolution features " + }, + { + "bbox": [ + 304, + 389, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 389, + 545, + 569 + ], + "type": "text", + "content": ". Subsequently, these processed queries and features enter the Joint Object and Part Segmentation (JOPS) head. For each query " + }, + { + "bbox": [ + 304, + 389, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i \\in \\mathbb{R}^E" + }, + { + "bbox": [ + 304, + 389, + 545, + 569 + ], + "type": "text", + "content": ", this head predicts (a) an object-level class, (b) an object-level segmentation mask, and (c) a set of part-level segmentation masks and classes for the parts within this object-level segment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "With this approach, we solve the limitations of existing works. Following the PPS task definition, we predict part-level segments per individual object-level segment. (1) As a result, we learn object-instance-aware representations for both parts and objects, improving the ability of the network to separate instances. (2) By predicting objects and parts jointly, the compatibility between both sets of predictions greatly improves. Moreover, in the JOPS head, we ensure that we only predict compatible parts, ensuring full object-part compatibility (see Sec. 3.3.2). (3) The joint object-part representations allow the network to fully leverage the complementary information from both abstraction levels." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3177" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 128, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 128, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 128, + 83 + ], + "type": "text", + "content": "3.3.2 JOPS head" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 224 + ], + "type": "text", + "content": "Within the JOPS head, we predict an object-level segment and compatible part-level segments for each shared query. For the object-level predictions, we follow Mask2Former [3]: (1) For the object-level class, we feed each processed query " + }, + { + "bbox": [ + 46, + 91, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i" + }, + { + "bbox": [ + 46, + 91, + 287, + 224 + ], + "type": "text", + "content": " through a single fully connected layer to predict a score for each possible class, and obtain the predicted class " + }, + { + "bbox": [ + 46, + 91, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\hat{c}_i^{\\mathrm{obj}}" + }, + { + "bbox": [ + 46, + 91, + 287, + 224 + ], + "type": "text", + "content": " by picking the highest-scoring class. (2) For the object-level mask, we feed the query " + }, + { + "bbox": [ + 46, + 91, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i" + }, + { + "bbox": [ + 46, + 91, + 287, + 224 + ], + "type": "text", + "content": " through a 3-layer MLP and generate a mask by taking the product of the resulting mask queries " + }, + { + "bbox": [ + 46, + 91, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i^{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 91, + 287, + 224 + ], + "type": "text", + "content": " and the features " + }, + { + "bbox": [ + 46, + 91, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 91, + 287, + 224 + ], + "type": "text", + "content": ", and applying a sigmoid activation, yielding " + }, + { + "bbox": [ + 46, + 91, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_i^{\\mathrm{obj}} \\in [0,1]^{H \\times W}" + }, + { + "bbox": [ + 46, + 91, + 287, + 224 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": "To predict part-level segments using the same features " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": ", we need to generate queries for each part class within an object-level segment. To do so, we first apply an MLP to each query " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": " to adapt it for part-level segmentation, and then apply " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "N^{\\mathrm{pc}}" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": " different fully-connected layers, where " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "N^{\\mathrm{pc}}" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": " is the total number of part-level classes in the dataset. This results in " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i^{\\mathrm{pt}} \\in \\mathbb{R}^{N^{\\mathrm{pc}} \\times E}" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": ", a set of 'per-object' part-level queries, where each query always corresponds to a fixed, predetermined part class. We could then take the product of " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i^{\\mathrm{pt}}" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": " and the features " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": " to generate a segmentation mask for all part-level classes. However, we already have an object-level class prediction for each query, and we know that only a subset of all " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "N^{\\mathrm{pc}}" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": " part classes is compatible with a certain object class. Therefore, we propose to only predict and supervise part-level masks for those compatible part-level classes. This simplifies the part segmentation task that the network needs to learn, as it no longer has to learn to predict 'empty' segmentation masks for incompatible part classes. Concretely, as visualized in Fig. 2, we identify the object-level class " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\hat{c}_i^{\\mathrm{obj}}" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": " for each query " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": ", and keep only the part-level queries " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i^{\\mathrm{pt,c}} \\in \\mathbb{R}^{N^{\\mathrm{c}} \\times E}" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": " that correspond to the " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "N^{\\mathrm{c}}" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": " part classes that are compatible with the object-level class. Then, we compute the product of these remaining queries " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i^{\\mathrm{pt,c}}" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": " and the features " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": ", and apply a sigmoid activation to generate compatible part segmentation masks " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_i^{\\mathrm{pt}} \\in [0,1]^{N^{\\mathrm{c}} \\times H \\times W}" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": ". As each per-object part query corresponds to a fixed part class, we also know the part classes " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\hat{c}_i^{\\mathrm{pt}}" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": ". Note that the number " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "N^{\\mathrm{c}}" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": " depends on the object-level class and will therefore vary per query " + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i" + }, + { + "bbox": [ + 46, + 224, + 288, + 574 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 586, + 118, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 118, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 118, + 598 + ], + "type": "text", + "content": "3.3.3 Training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 653 + ], + "type": "text", + "content": "To assign each query to at most one ground-truth object-level segment with corresponding part-level segments during training, we apply bipartite matching based on the predicted and ground-truth object-level classes and masks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "content": "To supervise the object-level segments, we use the cross-entropy loss for the classes, and both the Dice [28] and cross-entropy loss for the segmentation masks. Together, these losses form the object-level loss " + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{obj}}" + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "content": ". For the part-level segmentation masks, we also use the Dice and cross" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 106 + ], + "type": "text", + "content": "entropy loss, together forming part-level loss " + }, + { + "bbox": [ + 304, + 72, + 545, + 106 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{pt}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 106 + ], + "type": "text", + "content": ". We balance the losses with weights " + }, + { + "bbox": [ + 304, + 72, + 545, + 106 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{obj}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 106 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 106 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{pt}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 106 + ], + "type": "text", + "content": " to calculate the total loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 378, + 108, + 545, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 108, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 378, + 108, + 545, + 121 + ], + "type": "interline_equation", + "content": "L = \\lambda_ {\\mathrm {o b j}} L _ {\\mathrm {o b j}} + \\lambda_ {\\mathrm {p t}} L _ {\\mathrm {p t}}. \\tag {1}", + "image_path": "37f2df1928579e8ab0b7f9cf86a24222352ed95ff157bff2456d7aa2e921618a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 126, + 545, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 126, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 304, + 126, + 545, + 150 + ], + "type": "text", + "content": "We find that using " + }, + { + "bbox": [ + 304, + 126, + 545, + 150 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{obj}} = \\lambda_{\\mathrm{pt}} = 1" + }, + { + "bbox": [ + 304, + 126, + 545, + 150 + ], + "type": "text", + "content": " provides the best balance; see the supplementary material for more details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 162, + 422, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 162, + 422, + 175 + ], + "spans": [ + { + "bbox": [ + 305, + 162, + 422, + 175 + ], + "type": "text", + "content": "4. Experimental setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 182, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 182, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 305, + 182, + 545, + 193 + ], + "type": "text", + "content": "Datasets. We use the two PPS benchmarks for evaluation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "text", + "content": "Cityscapes Panoptic Parts (Cityscapes-PP) [5] extends the original Cityscapes dataset [4] with part-level labels. It consists of street scene images from several cities. It has labels for 19 object classes (11 stuff; 8 thing). Part classes are defined and labeled for 5 object-level thing classes; there are 23 part classes in total. We train on the train split (2975 images), and evaluate on the val split (500 images)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 278, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 545, + 422 + ], + "type": "text", + "content": "Pascal Panoptic Parts (Pascal-PP) [5], which combines existing labels for Pascal VOC [1, 6, 29], consists of a wide range of scenes and classes. There are 59 object-level classes (39 stuff; 20 thing), and part-level classes are defined for 15 thing classes. Unless otherwise indicated, we use the default part class definition with 57 part classes in total [5]. To evaluate a more challenging setting, we additionally evaluate on Pascal-PP-107, which uses the 107 non-background classes from the Pascal-Part-108 definition introduced by Michieli et al. [27] for part segmentation. For both class definitions, we train on the training split (4998 images), and evaluate on the validation split (5105 images)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 436, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 533 + ], + "type": "text", + "content": "Baseline. Our baseline is a version of TAPPS that uses the same network architecture, but uses a separate set of 100 additional queries for part-level segmentation (see also Fig. 1a). Comparing the results of the baseline in Tab. 1 to existing methods in Tab. 2, we find that our baseline outperforms state-of-the-art approaches that also use separate part-level queries [18, 19], indicating that it is a strong baseline. See the supplementary material for more details." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 548, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 632 + ], + "type": "text", + "content": "Evaluation metrics. The part-aware panoptic segmentation performance is evaluated using the default Part-aware Panoptic Quality (PartPQ) metric [5]. It captures both the ability to recognize and segment object-level segments (i.e., stuff regions and thing instances), and the ability to further segment the identified object-level segments into part-level masks. The PartPQ per object-level class " + }, + { + "bbox": [ + 304, + 548, + 545, + 632 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 548, + 545, + 632 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 347, + 639, + 545, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 639, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 347, + 639, + 545, + 670 + ], + "type": "interline_equation", + "content": "\\operatorname {P a r t P Q} _ {c} = \\frac {\\sum_ {(p , g) \\in T P _ {c}} \\operatorname {I o U} _ {\\mathrm {p}} (p , g)}{| T P _ {c} | + \\frac {1}{2} | F P _ {c} | + \\frac {1}{2} | F N _ {c} |}, \\tag {2}", + "image_path": "d0044a04fcf7c970ee4ddad5dd16b6b2ee60d5ce8a5b8ad8b6269525602484f9.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "TP_{c}, FP_{c}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "FN_{c}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " are the sets of true positive, false positive and false negative segments, respectively. A pair of predicted and ground-truth segments " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "(p, g)" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " of the same" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "3178" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 292, + 175 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 292, + 175 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 292, + 175 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 292, + 175 + ], + "type": "table", + "html": "
MethodPartPQPartSQPQ
PtNo ptAllPtThStAll
ImageNet pre-training
Baseline55.238.842.971.562.037.345.7
TAPPS (ours)59.639.444.674.365.637.647.1
COCO pre-training
Baseline64.849.953.773.174.848.157.1
TAPPS (ours)67.250.454.775.175.748.557.7
", + "image_path": "75799b0b186e8524ce0d16ca70d3c47558efbe8d7e4f6507c369f8d3b6a7b440.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 108, + 176, + 232, + 186 + ], + "lines": [ + { + "bbox": [ + 108, + 176, + 232, + 186 + ], + "spans": [ + { + "bbox": [ + 108, + 176, + 232, + 186 + ], + "type": "text", + "content": "(a) Pascal-PP validation [1, 5, 6, 29]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 299, + 70, + 542, + 175 + ], + "blocks": [ + { + "bbox": [ + 299, + 70, + 542, + 175 + ], + "lines": [ + { + "bbox": [ + 299, + 70, + 542, + 175 + ], + "spans": [ + { + "bbox": [ + 299, + 70, + 542, + 175 + ], + "type": "table", + "html": "
MethodPartPQPartSQPQ
PtNo ptAllPtThStAll
ImageNet pre-training
Baseline46.662.658.466.153.867.161.5
TAPPS (ours)48.763.159.366.855.667.362.4
COCO pre-training
Baseline48.264.860.466.756.169.063.6
TAPPS (ours)48.965.761.366.957.269.764.4
", + "image_path": "75ccc663a9b7a092dd860bf6b2a86cd08fbeb4fdd9da2cbc27de4e2bfcc369e1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 373, + 176, + 468, + 186 + ], + "lines": [ + { + "bbox": [ + 373, + 176, + 468, + 186 + ], + "spans": [ + { + "bbox": [ + 373, + 176, + 468, + 186 + ], + "type": "text", + "content": "(b) Cityscapes-PP val [4, 5]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 191, + 545, + 214 + ], + "lines": [ + { + "bbox": [ + 46, + 191, + 545, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 191, + 545, + 214 + ], + "type": "text", + "content": "Table 1. Main results. We compare TAPPS to a strong baseline that uses separate sets of queries to predict object- and part-level segments, instead of predicting object- and part-level segments jointly like TAPPS (see Sec. 4)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": "class " + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": " is part of " + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "inline_equation", + "content": "TP_{c}" + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": " if the Intersection-over-Union (IoU) of their object-level masks is larger than 0.5. If a ground-truth segment is not identified, it is part of " + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "inline_equation", + "content": "FN_{c}" + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": "; if a prediction is incorrect, it is part of " + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "inline_equation", + "content": "FP_{c}" + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": ". The IoUp term captures the segmentation performance within identified object-level segments. For object-level classes for which part classes are defined, it calculates the part-level mIoU. Otherwise, it uses the object-level IoU. We report the mean PartPQ over all classes, but also separately for object-level classes with parts (PartPQpt) and without parts (PartPQNoPt)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 354, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 426 + ], + "type": "text", + "content": "To individually assess the ability of methods to conduct part-level segmentation within identified objects, we report the Part Segmentation Quality for object-level classes that have parts, " + }, + { + "bbox": [ + 46, + 354, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\mathrm{PartSQ^{Pt}}" + }, + { + "bbox": [ + 46, + 354, + 287, + 426 + ], + "type": "text", + "content": ". Following De Geus et al. [5], per object-level class " + }, + { + "bbox": [ + 46, + 354, + 287, + 426 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 354, + 287, + 426 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 46, + 354, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\mathrm{PartSQ^{Pt}}" + }, + { + "bbox": [ + 46, + 354, + 287, + 426 + ], + "type": "text", + "content": " calculates the average part segmentation mIoU within object-level segments:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 93, + 434, + 287, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 434, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 93, + 434, + 287, + 462 + ], + "type": "interline_equation", + "content": "\\operatorname {P a r t S Q} _ {c} ^ {\\mathrm {P t}} = \\frac {\\sum_ {(p , g) \\in T P _ {c}} \\operatorname {I o U} _ {\\mathrm {p}} (p , g)}{| T P _ {c} |}. \\tag {3}", + "image_path": "15841de93b7530de95454f5b569bf51a817220a44881032d289de61eb868c4ef.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 471, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 533 + ], + "type": "text", + "content": "To evaluate the ability of networks to conduct object-level panoptic segmentation, we report the Panoptic Quality (PQ) metric [13]. Similarly to the PartPQ, we report the average PQ over all classes, but also over all thing classes " + }, + { + "bbox": [ + 46, + 471, + 287, + 533 + ], + "type": "inline_equation", + "content": "(\\mathrm{PQ}^{\\mathrm{Th}})" + }, + { + "bbox": [ + 46, + 471, + 287, + 533 + ], + "type": "text", + "content": " and stuff classes " + }, + { + "bbox": [ + 46, + 471, + 287, + 533 + ], + "type": "inline_equation", + "content": "(\\mathrm{PQ}^{\\mathrm{St}})" + }, + { + "bbox": [ + 46, + 471, + 287, + 533 + ], + "type": "text", + "content": " separately." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 545, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 714 + ], + "type": "text", + "content": "Implementation details. TAPPS is built on top of the publicly available code of state-of-the-art panoptic segmentation network Mask2Former [3]. For all datasets, we use a batch size of 16 images, and train on 4 Nvidia A100 GPUs. To optimize TAPPS, we use AdamW [26], a weight decay of 0.05, and a polynomial learning rate decay schedule with an initial learning rate of " + }, + { + "bbox": [ + 46, + 545, + 287, + 714 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 545, + 287, + 714 + ], + "type": "text", + "content": " and a power of 0.9. For Cityscapes-PP, we train for 90k iterations and apply conventional data augmentation steps [3, 18]: random flip, random resize with a factor between 0.5 and 2.0, and finally a random crop of " + }, + { + "bbox": [ + 46, + 545, + 287, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 1024" + }, + { + "bbox": [ + 46, + 545, + 287, + 714 + ], + "type": "text", + "content": " pixels. For Pascal-PP, we train for 60k iterations in case of ImageNet pre-training, but for only 10k iterations in case of COCO pre-training, to prevent overfitting. Following state-of-the-art panoptic segmenta" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 234, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 545, + 319 + ], + "type": "text", + "content": "tion implementations on COCO [3, 21], we apply large-scale jittering with a scale between 0.1 and 2.0 followed by a random crop of " + }, + { + "bbox": [ + 304, + 234, + 545, + 319 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 234, + 545, + 319 + ], + "type": "text", + "content": " pixels. During inference, we resize the image such that the shortest side is 800 pixels. Note that we use exactly the same training and testing settings for both TAPPS and the baseline. For more implementation details, see the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 328, + 359, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 328, + 359, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 328, + 359, + 340 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 348, + 388, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 348, + 388, + 360 + ], + "spans": [ + { + "bbox": [ + 306, + 348, + 388, + 360 + ], + "type": "text", + "content": "5.1. Main results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 366, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 629 + ], + "type": "text", + "content": "First, we compare TAPPS to the strong baseline that uses separate sets of queries for object-level panoptic segmentation and part-level semantic segmentation (see Sec. 4). The results in Tab. 1a demonstrate that TAPPS significantly outperforms the baseline on Pascal-PP in several aspects. Most importantly, the PartPQ for object-level classes with parts " + }, + { + "bbox": [ + 304, + 366, + 545, + 629 + ], + "type": "inline_equation", + "content": "(\\mathrm{PartPQ}^{\\mathrm{Pt}})" + }, + { + "bbox": [ + 304, + 366, + 545, + 629 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 366, + 545, + 629 + ], + "type": "inline_equation", + "content": "+4.4" + }, + { + "bbox": [ + 304, + 366, + 545, + 629 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 366, + 545, + 629 + ], + "type": "inline_equation", + "content": "+2.4" + }, + { + "bbox": [ + 304, + 366, + 545, + 629 + ], + "type": "text", + "content": " higher, depending on the pretraining strategy. Looking in more detail, we find that this increase is caused by two individual improvements: (1) The part-level segmentation quality " + }, + { + "bbox": [ + 304, + 366, + 545, + 629 + ], + "type": "inline_equation", + "content": "(\\mathrm{PartSQ}^{\\mathrm{Pt}})" + }, + { + "bbox": [ + 304, + 366, + 545, + 629 + ], + "type": "text", + "content": " is considerably higher than for the baseline. This shows the positive impact of having a joint representation for objects and parts, and simplifying the part segmentation task by only allowing compatible predictions (see also Sec. 5.3). (2) The panoptic quality for thing classes " + }, + { + "bbox": [ + 304, + 366, + 545, + 629 + ], + "type": "inline_equation", + "content": "(\\mathrm{PQ}^{\\mathrm{Th}})" + }, + { + "bbox": [ + 304, + 366, + 545, + 629 + ], + "type": "text", + "content": " also sees a substantial increase. Here, it should be noted that (a) all the object-level classes with parts are thing classes, and (b) thing classes require instance separation. Thus, this improvement does not only show the benefit of learning objects and parts jointly, but also indicates that learning parts in an object-instance-aware manner indeed improves the ability of the network to separate object instances (see also Sec. 5.4)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "For Cityscapes-PP, we see similar results in Tab. 1b. Again, we observe improvements on the PartPQ" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "^{\\mathrm{Pt}}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ", PartSQ" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "^{\\mathrm{Pt}}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and PQ" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "^{\\mathrm{Th}}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " metrics. In this case, the absolute improvements are slightly smaller. This is expected because this dataset contains significantly fewer classes and very similar images, making it easier to learn the PPS task and limiting the potential gains that can still be obtained by TAPPS." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3179" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 89, + 70, + 504, + 357 + ], + "blocks": [ + { + "bbox": [ + 89, + 70, + 504, + 357 + ], + "lines": [ + { + "bbox": [ + 89, + 70, + 504, + 357 + ], + "spans": [ + { + "bbox": [ + 89, + 70, + 504, + 357 + ], + "type": "table", + "html": "
MethodBackbonePre-trainingPartPQPartSQPQ
PtNo PtAllPtAll
Pascal-PP validation
JPPF [10]EfficientNet-B5I48.326.932.2--
TAPPS (ours)RN-50I59.639.444.674.347.1
Panoptic-PartFormer† [18]RN-50I,C56.138.843.266.847.6
Panoptic-PartFormer++† [19]RN-50I,C52.642.645.160.451.6
TAPPS (ours)RN-50I,C67.250.454.775.157.7
Panoptic-PartFormer† [18]Swin-BI,C64.350.654.170.858.1
Panoptic-PartFormer++† [19]Swin-BI,C48.952.151.353.059.8
TAPPS (ours)Swin-BI,C72.256.360.478.163.0
Cityscapes-PP val
Panoptic-PartFormer [18]RN-50I--54.5-57.8
Panoptic-PartFormer++ [19]RN-50I--57.5-61.6
JPPF [10]EfficientNet-B5I47.763.859.6--
TAPPS (ours)RN-50I48.763.159.366.862.4
Panoptic-PartFormer [18]RN-50I,C43.962.457.560.161.6
Panoptic-PartFormer++ [19]RN-50I,C42.565.159.2-63.6
TAPPS (ours)RN-50I,C48.965.761.366.964.4
Panoptic-PartFormer [18]Swin-BI,C45.667.862.059.066.6
Panoptic-PartFormer++ [19]Swin-BI,C46.068.262.3-68.0
Panoptic-PartFormer++ [19]ConvNeXt-BI,C46.469.163.1-68.2
SegFormer-B5 + CondInst + BPR [38–40, 45]MiT-B5 + RN-50I,C*48.667.562.5--
TAPPS (ours)Swin-BI,C53.069.064.868.068.0
", + "image_path": "d77bcd8f098b655f07936688a10de1f5c199d532bc7920cdc5d83c5912904242.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 420, + 222, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 222, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 222, + 433 + ], + "type": "text", + "content": "5.2. Comparison with state of the art" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 437, + 289, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 289, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 289, + 654 + ], + "type": "text", + "content": "In Tab. 2, we compare TAPPS to existing state-of-the-art methods across different datasets, backbones, and pretraining settings. On both Pascal-PP and Cityscapes-PP, TAPPS significantly outperforms existing work. Most importantly, it consistently scores higher on the PartPQ, PartPQ" + }, + { + "bbox": [ + 46, + 437, + 289, + 654 + ], + "type": "inline_equation", + "content": "^{\\text{Pt}}" + }, + { + "bbox": [ + 46, + 437, + 289, + 654 + ], + "type": "text", + "content": ", and PartSQ" + }, + { + "bbox": [ + 46, + 437, + 289, + 654 + ], + "type": "inline_equation", + "content": "^{\\text{Pt}}" + }, + { + "bbox": [ + 46, + 437, + 289, + 654 + ], + "type": "text", + "content": " metrics. TAPPS is only slightly outperformed by JPPF [10] on Cityscapes-PP with ImageNet pre-training, but we note that JPPF uses the EfficientNet-B5 [36] backbone, which is much more powerful than the ResNet-50 [9] used by TAPPS. Moreover, we do outperform JPPF by a large margin on the Pascal-PP dataset, showing the strength of TAPPS on more complex datasets. Overall, we achieve new state-of-the-art results on both datasets, obtaining PartPQ scores of 60.4 and 64.8 on Pascal-PP and Cityscapes-PP, improvements of +6.3 and +1.7, respectively. In the supplementary material, we compare qualitative examples of TAPPS, our baseline, and existing approaches, and we also show typical failure cases." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 659, + 116, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 116, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 116, + 671 + ], + "type": "text", + "content": "5.3. Ablations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "Predicting only compatible parts. Using shared queries for object- and part-level segmentation enables TAPPS to only allow part segmentation predictions that" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 308, + 418, + 545, + 552 + ], + "blocks": [ + { + "bbox": [ + 46, + 361, + 547, + 407 + ], + "lines": [ + { + "bbox": [ + 46, + 361, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 547, + 407 + ], + "type": "text", + "content": "Table 2. Comparison with state of the art. Evaluation on the Cityscapes-PP and Pascal-PP benchmarks [4-6]. RN-50 is ResNet-50 [9]. Other backbones are EfficientNet-B5 [36], MiT-B5 [45], Swin-B [24] and ConvNeXt-B [25]. I = ImageNet [33], C = COCO panoptic [21], C* = COCO pre-training for instance segmentation. †PartPQ scores for these existing methods have been re-evaluated using official code [5] and are higher than originally reported on Pascal-PP [18, 19], see supplementary material for more details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 418, + 545, + 552 + ], + "lines": [ + { + "bbox": [ + 308, + 418, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 308, + 418, + 545, + 552 + ], + "type": "table", + "html": "
Predict compat. partsPartPQPartSQPQObj. w/o conflicts
TrainingTestingPtNo PtAllPtAll
ImageNet pre-training
Baseline55.238.842.971.545.766.1%
XX57.939.243.972.946.699.7%
X✓58.039.243.972.946.6100.0%
✓✓59.639.444.674.347.1100.0%
COCO pre-training
Baseline64.849.953.773.157.164.6%
XX66.350.554.674.057.999.5%
X✓66.450.554.674.057.9100.0%
✓✓67.250.454.775.157.7100.0%
", + "image_path": "89a574898f629d0c72a6ca90360a9689143ffb4af9be8dc8074342969d3d8e26.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 556, + 547, + 578 + ], + "lines": [ + { + "bbox": [ + 305, + 556, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 305, + 556, + 547, + 578 + ], + "type": "text", + "content": "Table 3. Predicting only compatible part classes during training and testing. Evaluated on Pascal-PP [1, 5, 6, 29]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "are compatible with the query's object-level class (see Sec. 3.3.2). In Tab. 3, we evaluate the effect of predicting only compatible parts during both training and testing. In addition to the main metrics, we also assess the percentage of predicted objects for which there are no conflicting part-level predictions. The results show that, even when allowing incompatible predictions, TAPPS has significantly fewer object-part conflicts than the baseline. Naturally, when predicting only compatible parts during testing, all object-part conflicts are removed, eliminating the need" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3180" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 267, + 142 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 267, + 142 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 267, + 142 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 267, + 142 + ], + "type": "table", + "html": "
Adaptation layersPartPQPartSQPQ
PtNo PtAllPtAll
-67.349.654.175.257.1
1×FC67.150.354.675.157.6
2-layer MLP67.250.454.775.157.7
3-layer MLP67.150.454.675.157.7
", + "image_path": "4a1d14ca77c2e8f4ecdbc539569b3fc146378e9883907f980b571f9c74e12419.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 178, + 165, + 217 + ], + "blocks": [ + { + "bbox": [ + 47, + 146, + 287, + 169 + ], + "lines": [ + { + "bbox": [ + 47, + 146, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 146, + 287, + 169 + ], + "type": "text", + "content": "Table 4. JOPS head architecture (Fig. 2). Evaluated on Pascal-PP [1, 5, 6, 29], with pre-training on COCO panoptic [21]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 178, + 165, + 217 + ], + "lines": [ + { + "bbox": [ + 50, + 178, + 165, + 217 + ], + "spans": [ + { + "bbox": [ + 50, + 178, + 165, + 217 + ], + "type": "table", + "html": "
MethodmIoUThPQTh
Baseline69.862.0
TAPPS (ours)69.965.6
", + "image_path": "99d8895d4230173b4d5e9379661aeb91a3add568e48ef83e9fd5cc53b5f46476.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 63, + 218, + 151, + 228 + ], + "lines": [ + { + "bbox": [ + 63, + 218, + 151, + 228 + ], + "spans": [ + { + "bbox": [ + 63, + 218, + 151, + 228 + ], + "type": "text", + "content": "(a) Pascal-PP [1, 5, 6, 29]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 168, + 178, + 283, + 217 + ], + "blocks": [ + { + "bbox": [ + 168, + 178, + 283, + 217 + ], + "lines": [ + { + "bbox": [ + 168, + 178, + 283, + 217 + ], + "spans": [ + { + "bbox": [ + 168, + 178, + 283, + 217 + ], + "type": "table", + "html": "
MethodmIoUThPQTh
Baseline77.653.8
TAPPS (ours)78.255.6
", + "image_path": "7ae20b9020d85c5150233d6d66fdb87f8123de787ad7c89dd9022eb3204a3372.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 184, + 218, + 267, + 229 + ], + "lines": [ + { + "bbox": [ + 184, + 218, + 267, + 229 + ], + "spans": [ + { + "bbox": [ + 184, + 218, + 267, + 229 + ], + "type": "text", + "content": "(b) Cityscapes-PP [4, 5]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 257, + 287, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 287, + 329 + ], + "type": "text", + "content": "for post-processing. However, this does not yield a big improvement in terms of the segmentation quality. When we also apply this during training, we do observe a part segmentation quality " + }, + { + "bbox": [ + 46, + 257, + 287, + 329 + ], + "type": "inline_equation", + "content": "(\\mathrm{PartSQ^{Pt}})" + }, + { + "bbox": [ + 46, + 257, + 287, + 329 + ], + "type": "text", + "content": " improvement. This shows that simplifying the part segmentation task during training leads to improved performance, as we hypothesized in Sec. 3.3.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "text", + "content": "JOPS head architecture. In Tab. 4, we evaluate the impact of using different numbers of adaptation layers in the JOPS head before applying the " + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "inline_equation", + "content": "N^{\\mathrm{pc}}" + }, + { + "bbox": [ + 46, + 329, + 287, + 425 + ], + "type": "text", + "content": " fully-connected (FC) layers to generate the part queries (see Fig. 2). We find that one or two adaptation layers are necessary to generate object- and part-level representations that are sufficiently distinct to perform their respective tasks accurately; adding any more layers does not yield further improvements." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 432, + 162, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 162, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 162, + 445 + ], + "type": "text", + "content": "5.4. Additional analyses" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "text", + "content": "Instance separability. Tab. 1 showed that TAPPS improves the " + }, + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{PQ^{\\mathrm{Th}}}" + }, + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "text", + "content": ", i.e., the ability to recognize, segment and classify object instances. To assess if this is due to better object recognition or improved instance separability, we group all thing instances of the same object-level class together and evaluate the instance-agnostic semantic segmentation performance with the mIoU" + }, + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "inline_equation", + "content": "^{\\mathrm{Th}}" + }, + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "text", + "content": ". If the " + }, + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{PQ^{\\mathrm{Th}}}" + }, + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "text", + "content": " improvement were due to improved recognition, we expect the mIoU" + }, + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "inline_equation", + "content": "^{\\mathrm{Th}}" + }, + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "text", + "content": " to improve too. However, the results in Tab. 5 show only a minor mIoU" + }, + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "inline_equation", + "content": "^{\\mathrm{Th}}" + }, + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "text", + "content": " improvement. This indicates that the " + }, + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{PQ^{\\mathrm{Th}}}" + }, + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "text", + "content": " gain is not due to better recognition, but mainly results from a better ability to separate instances, showing the benefit of learning parts in an object-instance-aware manner." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "Performance on Pascal-PP-107. To assess the performance of TAPPS in a more complex setting, we evaluate it on Pascal-PP-107, which has 107 part-level classes instead of 57. The results on this dataset, reported in Tab. 6, show once more that TAPPS consistently improves the part segmentation and thing segmentation performance with respect to the baseline, in this case leading to a PartPQ" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "^{\\text{Pt}}" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " improvement of +4.1 or +2.2, depending on the pre-training strategy. This demonstrates that TAPPS is also effective" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 173 + ], + "blocks": [ + { + "bbox": [ + 53, + 230, + 280, + 243 + ], + "lines": [ + { + "bbox": [ + 53, + 230, + 280, + 243 + ], + "spans": [ + { + "bbox": [ + 53, + 230, + 280, + 243 + ], + "type": "text", + "content": "Table 5. Performance for things. ImageNet pre-training [33]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 545, + 173 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 173 + ], + "type": "table", + "html": "
MethodPartPQPartSQPQ
PtNo PtAllPtThStAll
ImageNet pre-training
Baseline45.738.940.659.362.137.645.9
TAPPS (ours)49.839.342.061.765.837.747.2
COCO pre-training
Baseline53.249.850.759.975.248.157.2
TAPPS (ours)55.450.051.462.175.648.157.4
", + "image_path": "35275ec13180d7b53c7a33e0e6ae9e9a68664ced619a729df79c258605afe8a7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 337, + 209, + 514, + 264 + ], + "blocks": [ + { + "bbox": [ + 306, + 177, + 545, + 199 + ], + "lines": [ + { + "bbox": [ + 306, + 177, + 545, + 199 + ], + "spans": [ + { + "bbox": [ + 306, + 177, + 545, + 199 + ], + "type": "text", + "content": "Table 6. 107 part-level categories. Results on the more challenging Pascal-PP-107 dataset [1, 5, 6, 27, 29]." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 209, + 514, + 264 + ], + "lines": [ + { + "bbox": [ + 337, + 209, + 514, + 264 + ], + "spans": [ + { + "bbox": [ + 337, + 209, + 514, + 264 + ], + "type": "table", + "html": "
MethodPartPQPartSQPQ
PtNo PtAllPtAll
Dynamic65.150.554.273.157.7
Fixed67.250.454.775.157.7
", + "image_path": "38b54017c86c383961ec61e4699e985a4a3ff386791832887787c119a2f43755.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 269, + 545, + 292 + ], + "lines": [ + { + "bbox": [ + 305, + 269, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 305, + 269, + 545, + 292 + ], + "type": "text", + "content": "Table 7. Fixed or dynamic part segmentation. Evaluated on Pascal-PP [1, 5, 6, 29], with pre-training on COCO panoptic [21]." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 302, + 482, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 302, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 302, + 482, + 314 + ], + "type": "text", + "content": "when the PPS task becomes more complex." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 315, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 545, + 494 + ], + "type": "text", + "content": "Fixed vs. dynamic part segmentation. As explained in Sec. 3.3.2, TAPPS uses a fixed fully connected layer for each part class to generate the corresponding 'per-object' part query. Alternatively, it is possible to predict part masks and classes dynamically within each object segment, using a set of dynamic queries like we do for object-level segmentation. In Tab. 7, we compare our fixed part segmentation setting with this dynamic approach, which is explained in more detail in the supplementary material. We find that our fixed approach results in a better " + }, + { + "bbox": [ + 304, + 315, + 545, + 494 + ], + "type": "inline_equation", + "content": "\\mathrm{PartSQ^{Pt}}" + }, + { + "bbox": [ + 304, + 315, + 545, + 494 + ], + "type": "text", + "content": " and therefore " + }, + { + "bbox": [ + 304, + 315, + 545, + 494 + ], + "type": "inline_equation", + "content": "\\mathrm{PartPQ^{Pt}}" + }, + { + "bbox": [ + 304, + 315, + 545, + 494 + ], + "type": "text", + "content": " performance. We hypothesize that the dynamic approach performs worse because it makes the part segmentation task unnecessarily complex, as the network needs to learn to assign segments to queries dynamically, and additionally predict a class label." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 504, + 378, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 504, + 378, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 504, + 378, + 516 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 524, + 545, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 524, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 524, + 545, + 655 + ], + "type": "text", + "content": "With experiments, we have shown that TAPPS considerably outperforms methods that predict objects and parts separately, by improving the object instance separability, part segmentation quality, and object-part compatibility. Importantly, these improvements can be attributed to the fact that TAPPS is directly optimized for the PPS task, using a set of shared queries to jointly predict objects and corresponding parts. With our promising findings, we hope to inspire future research towards even more complete scene understanding, e.g., image segmentation at even more abstraction levels, potentially with more flexible class hierarchies." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 658, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 658, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 658, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by Eindhoven Engine, NXP Semiconductors, and Brainport Eindhoven. This work made use of the Dutch national e-infrastructure with the support of the SURF Cooperative using grant no. EINF-5302, which is financed by the Dutch Research Council (NWO)." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3181" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect What You Can: Detecting and Representing Objects using Holistic Models and Body Parts. In CVPR, 2014. 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "type": "text", + "content": "[2] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-Pixel Classification is Not All You Need for Semantic Segmentation. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 213 + ], + "type": "text", + "content": "[3] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. In CVPR, 2022. 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 215, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 215, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 287, + 270 + ], + "type": "text", + "content": "[4] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The Cityscapes Dataset for Semantic Urban Scene Understanding. In CVPR, 2016. 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 271, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 271, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 271, + 287, + 304 + ], + "type": "text", + "content": "[5] Daan de Geus, Panagiotis Meletis, Chenyang Lu, Xiaoxiao Wen, and Gijs Dubbelman. Part-aware Panoptic Segmentation. In CVPR, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 304, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 287, + 348 + ], + "type": "text", + "content": "[6] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes (VOC) Challenge. *IJCV*, 88(2):303–338, 2010. 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 350, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 287, + 384 + ], + "type": "text", + "content": "[7] Ke Gong, Xiaodan Liang, Yicheng Li, Yimin Chen, Ming Yang, and Liang Lin. Instance-level human parsing via part grouping network. In ECCV, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 384, + 287, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 384, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 287, + 417 + ], + "type": "text", + "content": "[8] Ke Gong, Yiming Gao, Xiaodan Liang, Xiaohui Shen, Meng Wang, and Liang Lin. Graphonomy: Universal human parsing via graph transfer learning. In CVPR, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 418, + 287, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 418, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 53, + 418, + 287, + 451 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 453, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 485 + ], + "type": "text", + "content": "[10] Sravan Kumar Jagadeesh, René Schuster, and Didier Stricker. Multi-task Fusion for Efficient Panoptic-Part Segmentation. In ICPRAM, 2022. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 487, + 287, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 519 + ], + "type": "text", + "content": "[11] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One Transformer to Rule Universal Image Segmentation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 521, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 552 + ], + "type": "text", + "content": "[12] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollar. Panoptic Feature Pyramid Networks. In CVPR, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "type": "text", + "content": "[13] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic Segmentation. In CVPR, 2019. 1, 2, 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "text", + "content": "[14] Jianshu Li, Jian Zhao, Yunpeng Chen, Sujoy Roy, Shuicheng Yan, Jiashi Feng, and Terence Sim. Multi-human parsing machines. In ACM MM, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 623, + 287, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 654 + ], + "type": "text", + "content": "[15] Liulei Li, Tianfei Zhou, Wenguan Wang, Jianwu Li, and Yi Yang. Deep Hierarchical Semantic Segmentation. In CVPR, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "text", + "content": "[16] Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. Self-Correction for Human Parsing. IEEE TPAMI, 44(6):3260-3271, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[17] Qizhu Li, Anurag Arnab, and Philip HS Torr. Holistic, instance-level human parsing. In BMVC, 2017. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 714 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[18] Xiangtai Li, Shilin Xu, Yibo Yang, Guangliang Cheng, Yunhai Tong, and Dacheng Tao. Panoptic-PartFormer: Learning a Unified Model for Panoptic Part Segmentation. In ECCV, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 119, + 546, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 546, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 546, + 173 + ], + "type": "text", + "content": "[19] Xiangtai Li, Shilin Xu, Yibo Yang, Haobo Yuan, Guangliang Cheng, Yunhai Tong, Zhouchen Lin, and Dacheng Tao. PanopticPartFormer++: A Unified and Decoupled View for Panoptic Part Segmentation. arXiv preprint arXiv:2301.00954, 2023. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 175, + 546, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 546, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 546, + 218 + ], + "type": "text", + "content": "[20] Xiaodan Liang, Ke Gong, Xiaohui Shen, and Liang Lin. Look into Person: Joint Body Parsing & Pose Estimation Network and A New Benchmark. IEEE TPAMI, 41(4):871-885, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 220, + 546, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 546, + 264 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 546, + 264 + ], + "type": "text", + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 266, + 546, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 266, + 546, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 546, + 300 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature Pyramid Networks for Object Detection. In CVPR, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 301, + 546, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 301, + 546, + 356 + ], + "spans": [ + { + "bbox": [ + 307, + 301, + 546, + 356 + ], + "type": "text", + "content": "[23] Qing Liu, Adam Kortylewski, Zhishuai Zhang, Zizhang Li, Mengqi Guo, Qihao Liu, Xiaoding Yuan, Jiteng Mu, Weichao Qiu, and Alan Yuille. Learning Part Segmentation Through Unsupervised Domain Adaptation From Synthetic Vehicles. In CVPR, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 358, + 546, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 358, + 546, + 401 + ], + "spans": [ + { + "bbox": [ + 307, + 358, + 546, + 401 + ], + "type": "text", + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. In ICCV, 2021. 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 404, + 546, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 546, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 546, + 436 + ], + "type": "text", + "content": "[25] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s. In CVPR, 2022. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 438, + 546, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 438, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 438, + 546, + 460 + ], + "type": "text", + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICLR, 2019. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 462, + 546, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 462, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 462, + 546, + 506 + ], + "type": "text", + "content": "[27] Umberto Michieli, Edoardo Borsato, Luca Rossi, and Pietro Zanuttigh. GMNet: Graph Matching Network for Large Scale Part Semantic Segmentation in the Wild. In ECCV, 2020. 3, 5, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 508, + 546, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 508, + 546, + 541 + ], + "spans": [ + { + "bbox": [ + 307, + 508, + 546, + 541 + ], + "type": "text", + "content": "[28] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentatio. In 3DV, 2016. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 543, + 546, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 543, + 546, + 597 + ], + "spans": [ + { + "bbox": [ + 307, + 543, + 546, + 597 + ], + "type": "text", + "content": "[29] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The Role of Context for Object Detection and Semantic Segmentation in the Wild. In CVPR, 2014. 5, 6, 7, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 600, + 546, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 600, + 546, + 631 + ], + "spans": [ + { + "bbox": [ + 307, + 600, + 546, + 631 + ], + "type": "text", + "content": "[30] Tai-Yu Pan, Qing Liu, Wei-Lun Chao, and Brian Price. Towards Open-World Segmentation of Parts. In CVPR, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 634, + 546, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 634, + 546, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 634, + 546, + 677 + ], + "type": "text", + "content": "[31] Lu Qi, Jason Kuen, Weidong Guo, Jiumiang Gu, Zhe Lin, Bo Du, Yu Xu, and Ming-Hsuan Yang. AIMS: All-Inclusive Multi-Level Segmentation for Anything. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "type": "text", + "content": "[32] Tao Ruan, Ting Liu, Zilong Huang, Yunchao Wei, Shikui Wei, and Yao Zhao. Devil in the details: Towards accurate single and multiple human parsing. In AAAI, 2019. 3" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3182" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[33] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 115(3):211-252, 2015. 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "type": "text", + "content": "[34] Rishubh Singh, Pranav Gupta, Pradeep Shenoy, and Ravikiran Sarvadevabhatla. FLOAT: Factorized Learning of Object Attributes for Improved Multi-Object Multi-Part Scene Parsing. In CVPR, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 208 + ], + "type": "text", + "content": "[35] Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation. In ICCV, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "type": "text", + "content": "[36] Mingxing Tan and Quoc Le. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. In ICML, 2019. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 277 + ], + "type": "text", + "content": "[37] Xin Tan, Jiachen Xu, Zhou Ye, Jinkun Hao, and Lizhuang Ma. Confident Semantic Ranking Loss for Part Parsing. In ICME, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 278, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 278, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 278, + 287, + 322 + ], + "type": "text", + "content": "[38] Chufeng Tang, Hang Chen, Xiao Li, Jianmin Li, Zhaoxiang Zhang, and Xiaolin Hu. Look Closer To Segment Better: Boundary Patch Refinement for Instance Segmentation. In CVPR, 2021. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 323, + 287, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 323, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 48, + 323, + 287, + 356 + ], + "type": "text", + "content": "[39] Chufeng Tang, Lingxi Xie, Xiaopeng Zhang, Xiaolin Hu, and Qi Tian. Visual Recognition by Request. In CVPR, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 358, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 358, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 358, + 287, + 380 + ], + "type": "text", + "content": "[40] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional Convolutions for Instance Segmentation. In ECCV, 2020. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "type": "text", + "content": "[41] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: End-to-End Panoptic Segmentation With Mask Transformers. In CVPR, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 417, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 417, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 417, + 287, + 449 + ], + "type": "text", + "content": "[42] Wenguan Wang, Hailong Zhu, Jifeng Dai, Yanwei Pang, Jianbing Shen, and Ling Shao. Hierarchical human parsing with typed part-relation reasoning. In CVPR, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 451, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 451, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 451, + 287, + 483 + ], + "type": "text", + "content": "[43] Meng Wei, Xiaoyu Yue, Wenwei Zhang, Shu Kong, Xihui Liu, and Jiangmiao Pang. OV-PARTS: Towards Open-Vocabulary Part Segmentation. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 485, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 518 + ], + "type": "text", + "content": "[44] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "type": "text", + "content": "[45] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers. In NeurIPS, 2021. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 566, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 287, + 597 + ], + "type": "text", + "content": "[46] Lu Yang, Qing Song, Zhihui Wang, and Ming Jiang. Parsing R-CNN for Instance-Level Human Analysis. In CVPR, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "type": "text", + "content": "[47] Lu Yang, Qing Song, Zhihui Wang, Mengjie Hu, Chun Liu, Xueshi Xin, Wenhe Jia, and Songcen Xu. Renovating Parsing R-CNN for Accurate Multiple Human Parsing. In ECCV, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "text", + "content": "[48] Lu Yang, Qing Song, Zhihui Wang, Mengjie Hu, and Chun Liu. Hier R-CNN: Instance-Level Human Parts Detection and A New Benchmark. IEEE TIP, 30:39-54, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[49] Sanyi Zhang, Xiaochun Cao, Guo-Jun Qi, Zhanjie Song, and Jie Zhou. AIParsing: Anchor-Free Instance-Level Human Parsing. IEEE TIP, 31:5599-5612, 2022. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 184 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "type": "text", + "content": "[50] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-Net: Towards Unified Image Segmentation. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 546, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 546, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 546, + 150 + ], + "type": "text", + "content": "[51] Jian Zhao, Jianshu Li, Yu Cheng, Terence Sim, Shuicheng Yan, and Jiashi Feng. Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for multi-human parsing. In ACM MM, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 546, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 546, + 184 + ], + "type": "text", + "content": "[52] Yifan Zhao, Jia Li, Yu Zhang, and Yonghong Tian. Multi-Class Part Parsing With Joint Boundary-Semantic Awareness. In ICCV, 2019. 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3183" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/dd90eec4-a932-4045-a86b-cf06d264eadc_content_list.json b/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/dd90eec4-a932-4045-a86b-cf06d264eadc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..abb6c53664a5b09141491df7bc8d983bfc770c49 --- /dev/null +++ b/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/dd90eec4-a932-4045-a86b-cf06d264eadc_content_list.json @@ -0,0 +1,1657 @@ +[ + { + "type": "text", + "text": "Task2Box: Box Embeddings for Modeling Asymmetric Task Relationships", + "text_level": 1, + "bbox": [ + 102, + 130, + 869, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rangel Daroya Aaron Sun Subhransu Maji University of Massachusetts, Amherst", + "bbox": [ + 274, + 180, + 694, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{rdaroya, aaronsun, smaji}@umass.edu", + "bbox": [ + 323, + 219, + 640, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2db75518f57383f685a21cbd2d56d03d8bd8ebc92acb04ffc0d94457090b098f.jpg", + "image_caption": [ + "Figure 1. Box Embeddings of 150 Datasets of iNaturalist + CUB and Corresponding Learned Hierarchy for Class Arachnida. Each taxonomic category is treated as a separate dataset for which TASK2BOX embeddings are learned. (1) Shows the learned box embeddings where datasets from the same group (taxonomic class) have the same color. Datasets naturally cluster to their ground truth groups. (2) Shows the hierarchy learned through TASK2BOX for a specific class. The hierarchy matches the ground truth relationships based on biological classification. Orders that belong to class Arachnida are learned as boxes $(\\widehat{\\mathbf{A}}, \\widehat{\\mathbf{B}}, \\widehat{\\mathbf{C}})$ contained by the larger box for Arachnida; families under each of the orders are learned as smaller boxes contained by the corresponding orders they belong to." + ], + "image_footnote": [], + "bbox": [ + 127, + 247, + 401, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4b29e3406ad95f3b3e8cbaf0d6d2fc92366ff9b8429bd7445fdd735dfcd529df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 250, + 573, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c91020a3442cd53accdbd1012ce9356cc38ff8cfa65532288072065c36693b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 248, + 843, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3011d3bdea7341235c330d778344392cb670b0aa0883755756997e646665b8ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 437, + 359, + 571, + 439 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/62f14903af353c37192084c19c18874a14335fc2c1f5aabc9766e2d2f3545967.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 362, + 717, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d0c937ac1294d9baf50a5482d3302d0ab59162daa459f4460ace8ba0db229075.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 364, + 828, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 563, + 313, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modeling and visualizing relationships between tasks or datasets is an important step towards solving various meta-tasks such as dataset discovery, multi-tasking, and transfer learning. However, many relationships, such as containment and transferability, are naturally asymmetric and current approaches for representation and visualization (e.g., t-SNE [44]) do not readily support this. We propose TASK2BOX, an approach to represent tasks using box embeddings—axis-aligned hyperrectangles in low dimensional spaces—that can capture asymmetric relationships between them through volumetric overlaps. We show that TASK2BOX accurately predicts unseen hierarchical relationships between nodes in ImageNet and iNaturalist datasets, as well as transferability between tasks in the Taskonomy benchmark. We also show that box embeddings estimated from task representations (e.g., CLIP [36], Task2Vec [4], or attribute based [15]) can be used to predict relationships between unseen tasks more accurately than classifiers trained on the same representations, as well as handcrafted asymmetric distances (e.g., KL divergence).", + "bbox": [ + 75, + 598, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This suggests that low-dimensional box embeddings can effectively capture these task relationships and have the added advantage of being interpretable. We use the approach to visualize relationships among publicly available image classification datasets on popular dataset hosting platform called Hugging Face.", + "bbox": [ + 496, + 564, + 893, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 689, + 632, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The success of deep learning has led to the proliferation of datasets for solving a wide range of computer vision problems. Yet, there are few tools available to enable practitioners to find datasets related to the task at hand, and to solve various meta-tasks related to it. We present TASK2Box, a method to represent tasks using axis-aligned hyperrectangles (or box embeddings). TASK2Box is framed as a learnable mapping from dataset representation to boxes, and can be trained to predict various relationships between novel tasks such as transferability, hierarchy, and overlap.", + "bbox": [ + 496, + 715, + 893, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Box embeddings [48] extend order embeddings [46] by using volumetric relationships between axis-aligned hyper", + "bbox": [ + 500, + 869, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28827", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rectangles to encode pairwise relationships. Prior work in natural language processing has utilized box embeddings to represent the WordNet [30] hierarchy and to model conditional distributions. To model relationships between novel datasets, we develop a technique to map from Euclidean representations of datasets into the space of boxes. We explore simple image and label embedding from large vision-language models such as CLIP [36], TASK2VEC [4], and attribute-based vectors [15] as base representations of tasks.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We test our framework to model asymmetric relationships between nodes in iNaturalist [45] and Caltech-UCSD Birds (CUB) [50] and ImageNet [11] datasets, as well as to predict transferability on the Taskonomy benchmark [52]. Table 1 and 2 show that low-dimensional box embeddings accurately predict novel relationships between datasets seen during training, as well as relationships with novel datasets. Remarkably, TASK2BOX outperforms classifiers trained to directly predict the relationships on the same representations, suggesting that the box embedding provides a strong inductive bias for learning hierarchical relationships. We also outperform simple asymmetric distances proposed in prior work such as Kullback-Leibler (KL) divergence [4]. To model the heterogeneous tasks in the Taskonomy benchmark [52] we map each task to a set of attributes from which a box embedding is learned. Once again, we obtain significantly higher correlation between the true and predicted transferability for both existing and novel datasets compared to standard classifiers (Table 3). Such attribute-based representations can be readily derived from datasheets [15] and modelcards [31].", + "bbox": [ + 75, + 229, + 472, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, the low-dimensional box embeddings have the added advantage of being interpretable. Fig. 1 and Fig. 3 show relationships on the iNaturalist+CUB and ImageNet categories, respectively. The 2D box representation allows us to readily visualize the strength and direction of task relationships based on the overlapping volumes, which is not possible using symmetric distances with Euclidean representations (e.g., t-SNE [44]). At the same time, new datasets can be embedded in constant time without needing to retrain or re-optimize. Fig. 5 uses TASK2Box to visualize relationships among 131 publicly available datasets on Hugging Face [1], a popular platform for hosting datasets. Our main contributions are as follows:", + "bbox": [ + 75, + 547, + 472, + 743 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel method (TASK2BOX) that uses box embeddings to learn asymmetric (e.g., hierarchical, transfer learning) dataset relationships.", + "- We demonstrate that TASK2Box can predict the relationships of new tasks with a collection of existing tasks.", + "- We illustrate the interpretability of our model, and the ability to visualize public classification datasets on Hugging Face." + ], + "bbox": [ + 76, + 747, + 468, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The code for this project is publicly available at https://github.com/cvl-umass/task2box.", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Task Representations. Given a dataset $\\mathcal{D} = \\{(x_i, y_i)\\}_{i=1}^n$ , consisting of images $x_i \\in \\mathcal{X}$ and labels $y_i \\in \\mathcal{Y}$ , a range of approaches have been proposed for dataset representation. The most straightforward approach involves modeling the distribution of either the images $x$ or the labels $y$ within the dataset independently, using embeddings referred to in prior work as \"domain\" and \"label\" embeddings [4]. To capture the joint dependency between images and labels, [4] proposed the use of the Fisher Information Matrix (FIM) derived from a \"probe network\" trained to minimize a loss function $\\ell(\\hat{y}, y)$ over the dataset [22, 25, 35]. This approach leverages the similarity of FIMs to predict task transferability and for model selection. However, the utility of the FIM critically depends on the choice of the probe network and a pre-defined similarity may not accurately represent the various relationships between datasets.", + "bbox": [ + 496, + 114, + 893, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We also investigate the use of vision-language models (VLMs) such as CLIP [36]. This model, trained on a wide range of visual domains, can generalize to tasks involving vision and language data. CLIP features have been effective for image classification [3, 9], semantic segmentation [18, 24, 26], object detection [16, 47, 51], and even closing domain gaps for performance improvement [23, 53]. Both images $x \\in \\mathcal{X}$ and labels $y \\in \\mathcal{Y}$ represented as text, can be mapped into a shared space using the vision encoder $(\\phi)$ and text encoder $(\\psi)$ of CLIP, allowing us to model the dataset as a set of image and label embeddings $\\left\\{\\left(\\phi(x_i), \\psi(y_i)\\right)\\right\\}_{i=1}^n$ .", + "bbox": [ + 496, + 357, + 893, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We compare FIMs with representations derived from CLIP as base representations for tasks and learn box embeddings to model a variety of relations among tasks.", + "bbox": [ + 496, + 537, + 893, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Task Relations in Computer Vision. Understanding the relationships between tasks can lead to efficient solutions to new tasks. Previous work has measured task similarity by using model gradients [14] or based on their learned features [21] for grouping tasks for efficient multi-tasking. Similarly, predicting which pre-trained models will generalize the best on a new dataset could streamline model selection. Taskonomy [52] investigates transfer learning across vision tasks, varying from segmentation to pose estimation, by computing pairwise transfer distances or task affinities. These affinities are calculated by evaluating the extent to which a model trained on a source task generalizes to a target task [13, 42], though this process is computationally expensive.", + "bbox": [ + 496, + 598, + 893, + 811 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dataset Visualization. Low-dimensional Euclidean embeddings derived from UMAP [28], t-SNE [44], and LargeVis [43] are widely used to visualize relationships between datasets. They have been shown to successfully recover clusters of various data modalities by preserving the", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28828", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "relationship of each data point with its neighbors [4, 5, 40]. In low-dimension space, relationships with other data points are defined by their Euclidean distances. However, these are commonly used to represent symmetric relations.", + "bbox": [ + 75, + 90, + 467, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visualizations using tidy trees [37], circle packing [49], or cone trees [29] organize asymmetric relations as tree-structured hierarchies in low dimension. However, cyclic data relationships cannot be properly represented for these methods (e.g., when a node has two or more parents).", + "bbox": [ + 75, + 151, + 467, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Asymmetric Distances over Datasets. Kullback-Leibler (KL) divergence between image or label distributions provides a natural way to represent asymmetric distances between datasets. TASK2VEC [4] computes the similarity between two tasks (e.g., cosine distance), and introduces asymmetry by using the complexity of the first task as a reference. The complexity is measured by the similarity of the task embedding to a \"trivial embedding\" (embedding of a task that is easy or has no examples).", + "bbox": [ + 75, + 242, + 467, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Order embeddings on images were first proposed in [46] to capture tree-structured relationships. Given an dataset of $P = (u, v)$ drawn from an partially ordered set $(X, \\preceq_X)$ , they frame the problem as learning a mapping $f: (X, \\preceq_X) \\to (Y, \\preceq_Y)$ that is order preserving, i.e., $u \\preceq_X v \\Longleftrightarrow f(u) \\preceq_Y f(v)$ . The reserved product order was used for $\\preceq_Y$ , i.e., $x \\preceq y \\Longleftrightarrow x_i > y_i, \\forall i$ . Box embeddings [8] generalized this framework by representing points as axis-aligned hyper-rectangles and using volumetric relationships (e.g., intersection over union) to represent asymmetric relations. They used the framework to model conditional distributions and hypernymy relations (e.g., dog \"is a\" mammal) on the WordNet graph [2, 19, 33, 38, 48].", + "bbox": [ + 75, + 378, + 467, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hyperbolic spaces provide yet another way to model asymmetric relationships. Examples include the Poincaré disk model which uses hyperbolic cosine (cosh) to measure distance between points in a disk. Poincaré embeddings have been similarly used to represent WordNet hierarchies [32] and other relations in general graphs. Hyperbolic representations have also been proposed for representing images for efficient zero-shot learning given a taxonomic structure of the labels [27].", + "bbox": [ + 75, + 575, + 467, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To the best of our knowledge, no prior work has explored the use of these spaces for representing entire datasets and their effectiveness in capturing various task relationships. We adopt box embeddings in this work due to the effectiveness over alternatives in previous work [2, 6, 34, 38], ease of visualization, as well as due to open-source libraries for robust learning. However, instead of learning box embeddings directly, we learn mappings from task representations.", + "bbox": [ + 75, + 710, + 467, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. TASK2Box Framework", + "text_level": 1, + "bbox": [ + 76, + 844, + 300, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define the problem as follows: given a collection of datasets $\\{\\mathcal{D}_1,\\mathcal{D}_2,\\dots ,\\mathcal{D}_m\\}$ , and an asymmetric rela", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tionship given the pairwise relationship between datasets $d(\\mathcal{D}_i,\\mathcal{D}_j)\\in [0,1]$ , we aim to encode each dataset into a low-dimension space that preserves the relationships between datasets and is interpretable.", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve this, we propose using box embeddings for encoding each of the datasets. This process involves two main steps: (1) deriving the base representation $e$ of each dataset, and (2) learning a model $f_{\\theta}: e \\to z$ , where $z \\in \\mathbb{R}^{2 \\times k}$ represents a $k$ -dimensional axis-aligned hyperrectangle (i.e., box), denoted by its lower left and upper right coordinates. These steps are detailed further below.", + "bbox": [ + 496, + 151, + 890, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Base Task Representations", + "text_level": 1, + "bbox": [ + 498, + 267, + 740, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each dataset $\\mathcal{D} = \\{(x_i, y_i)\\}_{i=1}^n$ is defined as a collection of pairs of images $x_i \\in \\mathcal{X}$ and labels $y_i \\in \\mathcal{Y}$ . For obtaining a base embedding $e$ for each dataset, we utilize methods such as CLIP [36, 41], TASK2VEC [4], or attribute-based approaches [15].", + "bbox": [ + 496, + 289, + 890, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP. Using a pre-trained CLIP model [7, 20, 36, 41], we compute the mean and variance of the individual sample embeddings within each dataset. For each data sample, the image embedding is concatenated with the label embedding, the latter generated from text prompts (e.g., \"A photo of [CLS]\"). This concatenation models the joint distribution of images and labels. Eq. 1 and 2 detail how the mean and variance embeddings are derived, where $[i,j]$ represents the concatenation of vectors $i$ and $j$ , $\\phi$ is the vision encoder, and $\\psi$ is the text encoder. The covariance is approximated as diagonal for tractability.", + "bbox": [ + 496, + 380, + 890, + 547 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {C L I P} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} [ \\phi (x _ {i}), \\psi (y _ {i}) ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 558, + 890, + 599 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {C L I P} ^ {2} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\left[ \\phi \\left(x _ {i}\\right), \\psi \\left(y _ {i}\\right) \\right] - \\mu_ {C L I P}\\right) ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 619, + 890, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The base representation is defined as $e := \\mu_{CLIP} \\in \\mathbb{R}^{2048}$ or $e := [\\mu_{CLIP}, \\sigma_{CLIP}^2] \\in \\mathbb{R}^{4096}$ . A ViT-H/14 [12] pretrained on LAION-2B [41] was used to extract the embeddings.", + "bbox": [ + 496, + 672, + 890, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TASK2VEC [4] encodes a dataset using the approximate Fisher Information Matrix (FIM) of a network trained on the given dataset. The FIM represents the importance of parameters in the feature extractor by perturbing the weights $\\hat{w}$ of a given probe network with Gaussian noise $\\mathcal{N}(0,\\Lambda)$ . The precision matrix $\\Lambda$ is estimated to be close to an isotropic prior $\\mathcal{N}(\\hat{w},\\lambda^2 I)$ while having a good expected error. Eq. 3 is minimized to find $\\Lambda$ where $H$ is the cross entropy loss, $\\hat{w}$ are the weights of the network, $\\beta$ is the magnitude of the prior, $x\\in \\mathcal{X}$ , and $y\\in \\mathcal{Y}$ .", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28829", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\hat {w}; \\Lambda) = \\mathbb {E} _ {w \\sim \\mathcal {N} (\\hat {w}, \\Lambda)} [ H _ {p _ {w}, \\tilde {p}} p (y | x) ] +\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 123, + 356, + 140 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\beta K L \\left(\\mathcal {N} (0, \\Lambda) \\| \\mathcal {N} (0, \\lambda^ {2} I)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 143, + 468, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The matrix $\\Lambda$ provides an estimate of the FIM and is approximated as a diagonal matrix. The diagonal components are used as a base representation of a dataset ( $e := \\mathrm{FIM} \\in \\mathbb{R}^{17024}$ ). ResNet-34 [17] pretrained on ImageNet [11] is used as the probe network for all datasets.", + "bbox": [ + 76, + 175, + 468, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attribute-based. A task can be characterized by a set of $t$ binary attributes [15] represented as a vector of dimension $t$ . Some of the attributes explored for representing tasks are: (1) Is the task generative? (2) Is the task output in 2D? (3) Does the task involve camera pose estimation? Taking these 3 characteristics, for example, we can represent a 2D segmentation task as the vector $e = [0,1,0]$ as a discriminative 2D task that does not need camera poses.", + "bbox": [ + 76, + 266, + 468, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tasks in Taskonomy [52] involving multiple modalities benefit from this attribute-based representation due to its model independence. This approach also enables generalization to unseen tasks by identifying the presence or absence of various characteristics. Each vision task in Taskonomy is represented with 15 attributes, resulting in vectors $e \\in \\mathbb{R}^{15}$ . The full list of attributes is in Appendix ??", + "bbox": [ + 75, + 388, + 468, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Learning Box Embeddings", + "text_level": 1, + "bbox": [ + 76, + 506, + 318, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From a base task representation $e$ , we learn the parameters $\\theta$ of a model $f_{\\theta}: e \\to z$ that preserves the asymmetric similarity function $d(\\mathcal{D}_i, \\mathcal{D}_j)$ between any two datasets $\\mathcal{D}_i$ and $\\mathcal{D}_j$ . In Eq. 4, the learning objective is shown where $\\mathcal{L}_E$ is a loss function (mean squared error), $\\mathcal{L}_D$ is a distance function, $\\mathcal{L}_R$ is a regularization term, and $\\lambda$ is a hyperparameter. The embeddings $z_i, z_j \\in \\mathbb{R}^{2 \\times k}$ represent the coordinates of the lower left and the upper right corners of the respective $k$ dimensional boxes, with $f_{\\theta}(e_i) = z_i$ and $f_{\\theta}(e_j) = z_j$ .", + "bbox": [ + 76, + 530, + 468, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\theta} = \\underset {\\theta} {\\operatorname {a r g m i n}} \\sum_ {i, j} \\left(\\mathcal {L} _ {E} \\left(d \\left(\\mathcal {D} _ {i}, \\mathcal {D} _ {j}\\right), d _ {b o x} \\left(z _ {i}, z _ {j}\\right)\\right) \\right. \\\\ + \\lambda \\mathcal {L} _ {D} \\left(z _ {i}, z _ {j}\\right) + \\mathcal {L} _ {R}) \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 696, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The asymmetric relationship between box embeddings, denoted as $d_{box}(z_i,z_j)$ , is computed in Eq. 5 by calculating the volume of the intersection between $z_i$ and $z_j$ , normalized by the volume of $z_i$ . For $z_i$ to be fully contained inside $z_j$ ( $z_i \\subset z_j$ ), it is required that $d_{box}(z_i,z_j) = 1$ . Conversely, for $z_j$ to only partially contain $z_i$ , $d_{box}(z_j,z_i)$ must fall within the range (0,1). Fig. 2 illustrates this through a 2-dimensional example, where $z_1$ represents the box embedding for Canidae, and $z_2$ for Mammalia.", + "bbox": [ + 76, + 763, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dd107d1d5b3fc95821a19aca38ad74268e975911a6a8c983031b81716cb2100a.jpg", + "image_caption": [ + "Figure 2. TASK2Box Embeddings in 2D for Mammalia, Canidae, and Amphibia Datasets from iNaturalist. Each embedding represents the coordinates of the lower left and upper right corners of each box/rectangle. Since Canidae $(z_{1})$ is a proper subset of Mammalia $(z_{2})$ : $d_{box}(z_{1}, z_{2}) = 1$ and $d_{box}(z_{2}, z_{1}) = 0.1$ ." + ], + "image_footnote": [], + "bbox": [ + 581, + 90, + 808, + 226 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {b o x} \\left(z _ {i}, z _ {j}\\right) = \\frac {\\operatorname {v o l} \\left(z _ {i} \\cap z _ {j}\\right)}{\\operatorname {v o l} \\left(z _ {i}\\right)} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 328, + 890, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{L}_D$ is applied to datasets where $d(\\mathcal{D}_i, \\mathcal{D}_j) > 0$ such that the Euclidean distance between the center coordinates of $z_i$ and $z_j$ is minimized when starting from a non-overlapping state. This allows non-overlapping embeddings to move closer to each other, complementing $\\mathcal{L}_E$ to learn relationships. $\\mathcal{L}_R$ encourages solutions with regular-shaped boxes for better interpretability. The formulation of $\\mathcal{L}_R$ , given in Eq. 6, applies to a $k$ -dimensional box embedding $z_i$ , where $s_a$ represents the size of the $a$ -th dimension (for example, width), and $\\alpha, \\beta$ are hyperparameters. To prevent the trivial solution of minimizing box volume to zero, the inverse of the box volume is included. It's crucial to normalize the terms with respect to the embedding dimension, as the first and second terms scale quadratically and exponentially with dimension increase, respectively.", + "bbox": [ + 496, + 369, + 890, + 595 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R} = \\left(\\frac {\\alpha}{k ^ {2}} \\sum_ {a = 1} ^ {k} \\sum_ {b = a + 1} ^ {k} \\left| s _ {a} - s _ {b} \\right|\\right) + \\beta (\\operatorname {v o l} (z _ {i})) ^ {- 1 / k} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 621, + 890, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The architecture of $f$ consists of three fully-connected layers followed by two linear heads: one predicts a $k$ -dimensional lower-left coordinate, and the other predicts the $k$ -dimensional sizes of each box dimension.", + "bbox": [ + 496, + 676, + 890, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 752, + 633, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider the following goals to evaluate the capability of Task2Box to represent datasets:", + "bbox": [ + 496, + 777, + 888, + 808 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Given a collection of existing datasets $\\mathcal{D}_E$ and a subset of pairwise relationships $\\mathcal{R}$ : can the model generalize on unseen relationships $\\mathcal{R}'$ within $\\mathcal{D}_E$ where $\\mathcal{R}' \\cap \\mathcal{R} = \\emptyset$ ?", + "2. Given a collection of novel datasets $\\mathcal{D}_N$ not seen during training: can the model accurately identify the relationships with the existing datasets $\\mathcal{D}_E$ ?" + ], + "bbox": [ + 500, + 810, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28830", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/60af68a1bd48f6c62a154eac0c3781135184607bac64ac1295a134b31a09aa31.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
iNaturalist + CUB
MethodExisting DatasetsNovel Datasets
μCLIP[μ,σ2]CLIPFIMμCLIP[μ,σ2]CLIPFIM
TASK2Box (2D)69.23%67.84%39.61%50.07%39.66%10.06%
TASK2Box (3D)79.66%79.35%57.63%70.04%64.53%20.65%
TASK2Box (5D)84.67%82.41%79.72%73.79%72.11%34.88%
MLP Classifier45.25%61.45%26.34%39.06%44.54%19.90%
Linear Classifier4.40%3.11%7.06%4.77%5.87%15.92%
KL Divergence-6.58%7.94%-5.90%0.00%
Asymmetric Cosine9.29%11.54%2.83%1.47%1.47%1.47%
Asymmetric Euclidean1.71%1.71%8.53%1.47%1.47%1.91%
Random2.06%1.49%
", + "bbox": [ + 178, + 88, + 795, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Average F1 Score for Predicting Hierarchical Relationships on iNaturalist + CUB Dataset. For all feature types, TASK2Box outperforms other methods by more than $20\\%$ on existing datasets and more than $10\\%$ on novel datasets. The best-performing model is in bold, and the second-best is underlined. Results on Novel Datasets demonstrate that our model can generalize beyond the seen tasks it has previously seen. The dimensions such as 2D, 3D, and 5D refer to different box dimensionalities. Results using KL divergence for $\\mu_{CLIP}$ are not shown since it is not a distribution.", + "bbox": [ + 75, + 300, + 893, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The goals are evaluated through two experimental setups, demonstrating our model's ability to predict various types of relationships. The configurations, and their corresponding datasets, relationships, and metrics, are detailed below. Baseline methods for performance comparison are also reviewed. Implementation details are in the Appendix.", + "bbox": [ + 75, + 388, + 470, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 489, + 267, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Hierarchical Task Relationships", + "text_level": 1, + "bbox": [ + 76, + 512, + 354, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a combination of iNaturalist [45] and Caltech-UCSD Birds (CUB) [50], and instrument-related classes in ImageNet [11] to evaluate the ability of TASK2Box to represent hierarchical relationships. The first two are composed of images of various species, and the third is composed of images of various objects. The classes naturally follow a hierarchical form based on biological classification (taxonomy of iNaturalist+CUB species), and on semantic relations (hypernymy of objects in ImageNet).", + "bbox": [ + 75, + 537, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. For iNaturalist+CUB, the datasets are defined as the classes, the orders, and the families in the taxonomy that contain a significant number of samples per dataset as in [4]. There are 47 classes, 202 orders, and 589 families for a total of 838 datasets. For ImageNet, the instrument-related objects were processed using WordNet [30] to obtain hierarchical information between classes. This resulted in 131 datasets for training and evaluation.", + "bbox": [ + 75, + 672, + 468, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset Relationships. For any two datasets, their relationship is captured as $d(\\mathcal{D}_i,\\mathcal{D}_j)\\in \\{0,1\\}$ , where $d(\\mathcal{D}_i,\\mathcal{D}_j) = 1$ if and only if $\\mathcal{D}_i\\subset \\mathcal{D}_j$ , and $d(\\mathcal{D}_i,\\mathcal{D}_j) = 0$ otherwise. Fig. 2 provides an example: Canidae $(\\mathcal{D}_1)$ is a family within the class Mammalia $(\\mathcal{D}_2)$ ; thus, $d(\\mathcal{D}_1,\\mathcal{D}_2) = 1$ and $d(\\mathcal{D}_2,\\mathcal{D}_1) = 0$ . Meanwhile, Amphibia $(\\mathcal{D}_3)$ is unrelated to either dataset, resulting in $d(\\mathcal{D}_1,\\mathcal{D}_3) = 0$", + "bbox": [ + 75, + 795, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate the ability of the model to classify the presence of containment relationships between datasets using the F1 score due to the imbalance between positive and negative relationships. To obtain F1 score, precision and recall are first calculated. Precision is computed as the ratio of true positive pairs predicted to the total number of positive predictions. Recall is the ratio of true positive pairs predicted to the total number of true positive pairs. F1 score is then reported as the harmonic mean between the precision and recall. These calculations are done on both (1) the unseen relationships $\\mathcal{R}'$ within existing datasets $\\mathcal{D}_E$ and (2) the relationships between novel datasets $\\mathcal{D}_N$ and existing datasets $\\mathcal{D}_E$ . For the latter, we evaluate the relationships in both directions, i.e., $\\left\\{\\left(\\mathcal{D}_e,\\mathcal{D}_n\\right)\\forall \\mathcal{D}_e\\in \\mathcal{D}_E\\right\\} \\cup \\left\\{\\left(\\mathcal{D}_n,\\mathcal{D}_e\\right)\\forall \\mathcal{D}_e\\in \\mathcal{D}_E\\right\\}$ .", + "bbox": [ + 496, + 388, + 893, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.2 Transfer Learning Between Datasets", + "text_level": 1, + "bbox": [ + 498, + 633, + 808, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Taskonomy [52] benchmark is used to evaluate the ability of TASK2BOX to predict task affinities.", + "bbox": [ + 496, + 657, + 890, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Taskonomy defines a set of 25 visual tasks with corresponding pairwise task affinities. These tasks range from object detection, semantic segmentation, pose estimation, and more. We treat each visual task as a dataset. The tasks are described in Appendix ??", + "bbox": [ + 496, + 688, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset Relationships. In contrast to the containment relationship defined in § 4.1.1, Taskonomy quantifies relationships with task affinity measured by the performance gain achieved by transfer learning from a source dataset $\\mathcal{D}_j$ to a target dataset $\\mathcal{D}_i$ . The values are computed using Analytic Hierarchy Process [39, 52]. Dataset relationships are computed and normalized based on an ordinal approach and determined by the percentage of images that transfer well to a target task given a set of source tasks as detailed in [52].", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28831", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c647526b7000b1e089fcfa92244d00d2d7307c1c7f25f71869f7583551b24c4d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ImageNet
MethodExisting DatasetsNovel Datasets
μCLIP[μ,σ2]CLIPFIMμCLIP[μ,σ2]CLIPFIM
TASK2Box (2D)62.72%63.33%31.32%47.76%48.35%9.46%
TASK2Box (3D)83.12%82.79%58.16%73.06%66.48%24.70%
TASK2Box (5D)90.58%88.48%64.91%76.95%78.84%37.39%
MLP Classifier54.20%60.43%44.85%62.24%64.56%41.22%
Linear Classifier9.84%8.33%25.46%11.89%11.87%22.98%
KL-5.53%9.90%-10.41%0.00%
Asymmetric Cosine4.28%4.28%6.92%4.52%4.52%0.00%
Asymmetric Euclidean3.73%3.73%7.16%4.52%4.52%4.73%
Random3.64%5.02%
", + "bbox": [ + 178, + 88, + 795, + 291 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Average F1 Score for Predicting Hierarchical Relationships on ImageNet Instruments. TASK2BOX is shown to generalize well on both predicting relationships between existing datasets (Existing Datasets), and predicting box embeddings of new datasets and their relationships with existing datasets (Novel Datasets).", + "bbox": [ + 75, + 300, + 893, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a pair of datasets $(\\mathcal{D}_i,\\mathcal{D}_j)$ , the task affinity is defined as $d(\\mathcal{D}_i,\\mathcal{D}_j)\\in [0,1]$ . The higher the task affinity to a target dataset $\\mathcal{D}_i$ from a source dataset $\\mathcal{D}_j$ , then $d(\\mathcal{D}_i,\\mathcal{D}_j)$ gets closer to 1, where a value of 1 would show $\\mathcal{D}_i\\subset \\mathcal{D}_j$ .", + "bbox": [ + 75, + 359, + 468, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate the prediction of the model based on the Spearman correlation between the ground truth task affinity values, and the predicted values based on the box distances in Eq. 5. Similar to § 4.1.1, we evaluate on both (1) unseen relationships $\\mathcal{R}'$ for existing datasets, and (2) on relationships between unseen novel datasets $\\mathcal{D}_N$ and existing datasets $\\mathcal{D}_E$ .", + "bbox": [ + 75, + 421, + 470, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Baseline Methods", + "text_level": 1, + "bbox": [ + 76, + 537, + 250, + 553 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the performance of TASK2Box with alternative models and simple asymmetric distances proposed in prior work. For hierarchical relationships, given two datasets $\\mathcal{D}_i,\\mathcal{D}_j$ , the models predict $\\hat{d} (\\mathcal{D}_i,\\mathcal{D}_j)\\in \\{0,1\\}$ . For task affinity, the models predict $\\hat{d} (\\mathcal{D}_i,\\mathcal{D}_j)\\in [0,1]$ . The structure for the various methods are discussed below.", + "bbox": [ + 75, + 561, + 468, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Linear Model. A linear model is trained to predict the relationship value between two datasets. The input is the concatenation of the base representation $e$ for the two datasets.", + "bbox": [ + 75, + 655, + 468, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MLP Model. A 4-layer MLP is used instead for prediction on the same inputs as the linear model.", + "bbox": [ + 75, + 700, + 468, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "KL Divergence. Each dataset is treated as a multivariate Gaussian using the mean and variance of the image and label features (CLIP) or directly as the FIM. The optimal threshold $t$ is selected for the minimum distance between two dataset distributions $\\mathrm{KL}(\\mathcal{D}_i||\\mathcal{D}_j) < t$ for a prediction of $\\hat{d} (\\mathcal{D}_i,\\mathcal{D}_j) = 1$ , and $\\hat{d} (\\mathcal{D}_i,\\mathcal{D}_j) = 0$ otherwise. The F1 Score (hierarchical) or the correlation (task affinity) is the objective for selecting $t$ on the train set.", + "bbox": [ + 75, + 732, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Asymmetric Cosine Similarity. The cosine similarity $d_{\\text{cos}}$ is a symmetric measure between two embeddings $e_i, e_j$ . An asymmetric variant was proposed in [4] and shown in Eq. 7", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/09083c2f5ac4f3c26485b217f6bc1070383f1c2974afcfe2dee3abb75e6af820.jpg", + "image_caption": [ + "Figure 3. Visualization of Instrument-related Datasets in ImageNet. Datasets that belong to the same superset are shaded in the same color. TASK2Box learns the hierarchy of various groups, and clusters similar datasets closer." + ], + "image_footnote": [], + "bbox": [ + 573, + 361, + 820, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "by considering the similarity of $e_i$ and $e_j$ relative to the complexity of $e_i$ . The complexity of is measured as the distance to the trivial embedding $e_o$ , and $\\alpha$ is a hyperparameter.", + "bbox": [ + 498, + 599, + 890, + 646 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nd _ {a s y m} \\left(e _ {i}, e _ {j}\\right) = d _ {c o s} \\left(e _ {i}, e _ {j}\\right) - \\alpha d _ {c o s} \\left(e _ {i}, e _ {o}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 661, + 890, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "An optimal threshold $t$ is found on the train set where $d_{asym}(e_i, e_j) < t$ results to a prediction $\\hat{d}(\\mathcal{D}_i, \\mathcal{D}_j) = 1$ . The same threshold $t$ is used for test set evaluation.", + "bbox": [ + 498, + 685, + 890, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Asymmetric Euclidean Similarity. The asymmetric similarity is computed as in Eq. 7 but uses the Euclidean distance instead of cosine.", + "bbox": [ + 498, + 731, + 890, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Random. The probability of containment (for hierarchical) or the value of the task affinity is uniformly random.", + "bbox": [ + 498, + 776, + 890, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Results and Discussion", + "text_level": 1, + "bbox": [ + 500, + 821, + 714, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Hierarchical Relationships", + "text_level": 1, + "bbox": [ + 500, + 845, + 741, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tables 1 and 2 present the average F1 score for predicting hierarchical relationships on iNaturalist+CUB and Ima", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28832", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "geNet datasets. Various methods of extracting dataset embeddings were evaluated. TASK2BOX outperforms baseline methods, indicating that TASK2BOX can both generalize across unseen relationships (Existing Datasets), and to accurately represent relationships with existing datasets for unseen datasets (Novel Datasets). CLIP features also generalize better than FIM features, perhaps because the CLIP multi-modal embedding of images and labels was trained across a broad set of domains[36]. However, while CLIP is confined to image and text modalities, FIM can accommodate any modality by adapting the probe network's last layer for different prediction tasks.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Beyond predicting relationships, our method allows the visualization of datasets. Fig. 1 illustrates this for a subset of datasets in iNaturalist and CUB using TASK2Box (2D), and Fig. 3 for ImageNet. The hierarchical organization of the datasets is apparent in the TASK2Box representation – datasets that contain others appear as larger boxes, while more specialized datasets are depicted as smaller boxes netted with borader, more general dataset boxes. However, while 2D visualization offers insights, the flat surface may restrict the representation of complex relationships. Representations in higher dimensions is explored in § 5.4.", + "bbox": [ + 75, + 272, + 472, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Task Affinity", + "text_level": 1, + "bbox": [ + 76, + 449, + 212, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3 presents the results of task representations in Taskonomy. Our method shows that, even with attribute-based embeddings, it can learn box embeddings through relationship supervision between datasets. TASK2Box is shown to correlate highly with the ground truth task affinities compared to other methods. Given that only 25 tasks are available in Taskonomy, with 3 held out of training as novel datasets, there is a higher uncertainty in predicting unseen datasets. We expect that performance on novel datasets will improve with more datasets available during training.", + "bbox": [ + 75, + 472, + 468, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 4 displays the learned representation with TASK2Box (2D). Each subfigure in (a)-(c) represents a subset of tasks identified as having strong transfer relationships. TASK2Box not only identifies but also visually represents related tasks suitable for fine-tuning. The small highlighted boxes indicate target tasks, while the larger enclosing boxes represent source tasks. Although other tasks may not be proper subsets, the box distance in Eq. 5 provides an estimate for task affinity. A small box distance between two datasets, $d_{box}(z_1, z_2)$ , suggests lower transfer performance from a source task $z_2$ to a target task $z_1$ .", + "bbox": [ + 75, + 625, + 470, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Visualizing Public Datasets", + "text_level": 1, + "bbox": [ + 76, + 800, + 321, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Apart from predicting both hierarchical relationships and quantified task affinities, our method can also be used to visualize public datasets that lack available ground truth relationships. Using a set of vision tasks from Hugging Face [1], $\\mu_{CLIP} + \\text{TASK2Box}$ (2D) were utilized to pre", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fe7890c962f7a3223caf4add0c39e6794c38b0aadb9eb1dd83a160fbaef0df95.jpg", + "image_caption": [ + "Figure 4. Visualization of Tasks in Taskonomy showing Source Tasks that Transfer Well to Target Tasks (shaded). (a) Jigsaw and Triplet Fixated Camera Pose estimation are source tasks that transfer well to Depth estimation. (b) and (c) show different source tasks (larger boxes) that transfer well to the shaded boxes of Denoising Autoencoder and Surface Normal, respectively." + ], + "image_footnote": [], + "bbox": [ + 519, + 97, + 627, + 174 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef7aace0e3560b707f6b06d87f3df34013a0b1ab23bcf8cfdab46fef5df1f363.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 98, + 750, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/10fc35761ec0efa7ef3b12f441f4ec9681efb41db44c4140c3e72356248047c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 766, + 99, + 870, + 172 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/669b26aacdbb98247c474ac6370fe5131f2ddbd5ad76496991c401027965f124.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodExisting DatasetsNovel Datasets
Spearman's ρSpearman's ρ
TASK2Box (2D)0.85 ± 0.060.12 ± 0.21
TASK2Box (3D)0.93 ± 0.020.48 ± 0.24
TASK2Box (5D)0.94 ± 0.030.39 ± 0.22
MLP0.88 ± 0.060.31 ± 0.18
Linear0.75 ± 0.110.40 ± 0.24
Random0.05 ± 0.140.15 ± 0.07
", + "bbox": [ + 513, + 266, + 879, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Spearman Correlation and Standard Deviation between Predicted and Ground Truth Task Affinities on Taskonomy. Our method shows higher correlation with the task affinities compared to the baseline. Attribute-based embeddings were used.", + "bbox": [ + 496, + 411, + 890, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dict dataset embeddings. A constraint on the box sizes was added to reflect information about the dataset sizes. Fig. 5 displays the results of TASK2Box on 131 datasets from Hugging Face, where similar datasets, such as sentiments and documents, are shown to overlap. This approach allows for the analysis and visualization of dataset variations and similarities even with only samples of images and labels. Simultaneously, the sizes of various datasets can be visualized as box sizes – with larger datasets containing more samples depicted as larger boxes. This tool enables computer vision practitioners to see how their datasets compare with other existing datasets. Beyond visualizing the variation of available datasets (based on the number of clusters), it can expedite the process of finding suitable data sources by examining embedding overlaps.", + "bbox": [ + 496, + 486, + 890, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Analysis of TASK2BOX", + "text_level": 1, + "bbox": [ + 500, + 724, + 705, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We discuss properties of TASK2Box below. Additional insights are also included in Appendix ??", + "bbox": [ + 496, + 748, + 890, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using a Box Prior Improves Performance. While TASK2Box was trained with a 3-layer MLP, it achieved significantly better performance than the MLP baseline without a box prior. Tables 1, 2, and 3 demonstrate our method outperforming the baselines. Representing each dataset as an entity with shape and volume (such as a box), instead of solely learning relationships from embeddings, proves effective for generalization on unseen relationships. This", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28833", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d80c9536d589426431a0c301c60fab3dc0aaa4351d9b6b744451ba50d908dcb.jpg", + "image_caption": [ + "Figure 5. Visualizing Image Classification Datasets in Hugging Face. The sample data points annotated on the highlighted datasets show that common tasks overlap with each other (e.g., sentiment classification and document classification datasets). Although labels could slightly differ between datasets, TASK2BOX can infer the level of similarity and represent it as the amount of overlap. The embedding size (box area) also shows the number of available data." + ], + "image_footnote": [], + "bbox": [ + 112, + 89, + 861, + 300 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e26db90d6b8abfefdd07cdbae2b735cee39eca4ca7ced8983502cb5aca2f7875.jpg", + "image_caption": [ + "Figure 6. Effect of TASK2Box Embedding Dimension on the Accuracy of Predicting Task Relations. As the dimension increases, the performance generally increases." + ], + "image_footnote": [], + "bbox": [ + 147, + 377, + 400, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "improvement could be attributed to the explicit modeling of relationships through physical shapes, which enforces consistency: if $z_{i} \\subset z_{j}$ and $z_{j}$ has no overlap with $z_{k}$ , then having a physical representation ensures that $z_{i} \\cap z_{k} = \\emptyset$ .", + "bbox": [ + 75, + 592, + 468, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TASK2BOX can Represent Relations in Varying Dimensions. Fig. 6 illustrates the performance of TASK2BOX as the box dimension increases. Representing relationships in two dimensions results in easily interpretable embeddings; however, modeling complex task relationships could benefit from expansion to higher dimensions. Projecting to higher dimensions may yield even better performance, as relationships between datasets become more accurately represented, given the model's increased representation capacity. Nonetheless, as the dimensionality further increases, learning and visualizing the embedding space also becomes more challenging.", + "bbox": [ + 75, + 655, + 470, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Using Boxes to Represent Tasks Enables Effective Calculations on Embeddings. Boxes offer the advantage of being closed under intersection, meaning the intersection of two boxes results in another box, a property not shared by", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "circles or ellipses. Among region-based embeddings that utilize geometric objects (e.g., cones, disks, and boxes) to represent entities, operations involving boxes are the most straightforward to calculate [10, 38].", + "bbox": [ + 498, + 377, + 890, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 455, + 617, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a novel method that learns low-dimensional, interpretable embeddings of various task relationships. With box representations, we demonstrate that asymmetric relationships, such as task affinities and hierarchies, can be accurately modeled using various base representations. While CLIP embeddings have been shown to outperform FIM when integrated with our model, future work could investigate how CLIP might be adapted to modalities beyond text and images. Attribute-based features offer a viable alternative, and can be extracted from from datasheets using natural language processing techniques.", + "bbox": [ + 496, + 481, + 890, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The distinct properties of TASK2BOX were analyzed, revealing its ability to perform effectively across varying dimensions, and to model and visualize overlaps among public classification datasets on Hugging Face. This could enable computer vision practitioners to assess dataset utility for a task in hand. Although our model successfully represents task relationships, it does not incorporate information about optimal training procedures and model architecture. Future work could explore the inclusion of additional information for a more detailed understanding of the task space.", + "bbox": [ + 496, + 648, + 890, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by awards from the National Science Foundation (2329927 and 1749833) and the NASA AIST program. The experiments were performed on the University of Massachusetts GPU cluster funded by the Mass. Technology Collaborative.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28834", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hugging Face datasets. https://huggingface.co/datasets?taskCategories $\\equiv$ taskCategories:image-classification.2,7", + "[2] Ralph Abboud, Ismail Ceylan, Thomas Lukasiewicz, and Tommaso Salvatori. Boxe: A box embedding model for knowledge base completion. Advances in Neural Information Processing Systems, 33:9649-9661, 2020. 3", + "[3] Rabab Abdelfattah, Qing Guo, Xiaoguang Li, Xiaofeng Wang, and Song Wang. Cdul: Clip-driven unsupervised learning for multi-label image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1348-1357, 2023. 2", + "[4] Alessandro Achille, Michael Lam, Rahul Tewari, Avinash Ravichandran, Subhransu Maji, Charless C Fowlkes, Stefano Soatto, and Pietro Perona. Task2vec: Task embedding for meta-learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6430-6439, 2019. 1, 2, 3, 5, 6", + "[5] Sanjeev Arora, Wei Hu, and Pravesh K. Kothari. An analysis of the t-sne algorithm for data visualization. In Proceedings of the 31st Conference On Learning Theory, pages 1455-1462. PMLR, 2018. 3", + "[6] Michael Boratko, Dongxu Zhang, Nicholas Monath, Luke Vilnis, Kenneth L Clarkson, and Andrew McCallum. Capacity and bias of learned geometric embeddings for directed graphs. Advances in Neural Information Processing Systems, 34:16423-16436, 2021. 3", + "[7] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 3", + "[8] Tejas Chheda, Purujit Goyal, Trang Tran, Dhruvesh Patel, Michael Boratko, Shib Sankar Dasgupta, and Andrew McCallum. Box embeddings: An open-source library for representation learning using geometric structures. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 203-211, 2021. 3", + "[9] Marcos V Conde and Kerem Turgutlu. Clip-art: Contrastive pre-training for fine-grained art classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3956-3960, 2021. 2", + "[10] Shib Dasgupta, Michael Boratko, Siddhartha Mishra, Shriya Atmakuri, Dhruvesh Patel, Xiang Li, and Andrew McCallum. Word2box: Capturing set-theoretic semantics of words using box embeddings. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 2022. 8", + "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 4, 5", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations (ICLR), 2021. 3", + "[13] Kshitij Dwivedi and Gemma Roig. Representation similarity analysis for efficient task taxonomy & transfer learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12387-12396, 2019. 2", + "[14] Chris Fifty, Ehsan Amid, Zhe Zhao, Tianhe Yu, Rohan Anil, and Chelsea Finn. Efficiently identifying task groupings for multi-task learning. Advances in Neural Information Processing Systems, 34:27503-27516, 2021. 2", + "[15] Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé Iii, and Kate Crawford. Datasheets for datasets. Communications of the ACM, 64(12):86-92, 2021. 1, 2, 3, 4", + "[16] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2021. 2", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4", + "[18] Wenbin He, Suphanut Jamonnak, Liang Gou, and Liu Ren. Clip-s4: Language-guided self-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11207-11216, 2023. 2", + "[19] EunJeong Hwang, Jay-Yoon Lee, Tianyi Yang, Dhruvesh Patel, Dongxu Zhang, and Andrew McCallum. Event-event relation extraction using probabilistic box embedding. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 235–244, 2022. 3", + "[20] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 3", + "[21] Zhuoliang Kang, Kristen Grauman, and Fei Sha. Learning with whom to share in multi-task feature learning. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pages 521-528, 2011. 2", + "[22] Ryo Karakida, Shotaro Akaho, and Shun-ichi Amari. Universal statistics of fisher information in deep neural networks: Mean field approach. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 1032–1041. PMLR, 2019. 2", + "[23] Zhengfeng Lai, Noranart Vesdapunt, Ning Zhou, Jun Wu, Cong Phuoc Huynh, Xuelu Li, Kah Kuen Fu, and Chen-Nee Chuah. Padclip: Pseudo-labeling with adaptive debiasing in clip for unsupervised domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16155-16165, 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28835", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7061-7070, 2023. 2", + "[25] Zhibin Liao, Tom Drummond, Ian Reid, and Gustavo Carneiro. Approximate fisher information matrix to characterize the training of deep neural networks. IEEE transactions on pattern analysis and machine intelligence, 42(1): 15-26, 2018. 2", + "[26] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15305-15314, 2023. 2", + "[27] Shaoteng Liu, Jingjing Chen, Liangming Pan, Chong-Wah Ngo, Tat-Seng Chua, and Yu-Gang Jiang. Hyperbolic visual embedding learning for zero-shot recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[28] Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. Umap: Uniform manifold approximation and projection. Journal of Open Source Software, 3(29):861, 2018. 2", + "[29] Guy Melancon and Ivan Herman. Circular drawings of rooted trees. CWI (Centre for Mathematics and Computer Science), 1998. 3", + "[30] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 2, 5", + "[31] Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. Model cards for model reporting. In Proceedings of the conference on fairness, accountability, and transparency, pages 220-229, 2019. 2", + "[32] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. Advances in neural information processing systems, 30, 2017. 3", + "[33] Dhruvesh Patel, Shib Sankar Dasgupta, Michael Boratko, Xiang Li, Luke Vilnis, and Andrew McCallum. Representing joint hierarchies with box embeddings. In Automated Knowledge Base Construction, 2020. 3", + "[34] Dhruvesh Patel, Pavitra Dangati, Jay-Yoon Lee, Michael Boratko, and Andrew McCallum. Modeling label space interactions in multi-label classification using box embeddings. In International Conference on Learning Representations, 2022. 3", + "[35] Jeffrey Pennington and Pratik Worah. The spectrum of the fisher information matrix of a single-hidden-layer neural network. Advances in neural information processing systems, 31, 2018. 2", + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sion. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 7", + "[37] Edward M. Reingold and John S. Tilford. Tidier drawings of trees. IEEE Transactions on software Engineering, (2): 223-228, 1981. 3", + "[38] Hongyu Ren, Weihua Hu, and Jure Leskovec. Query2box: Reasoning over knowledge graphs in vector space using box embeddings. In International Conference on Learning Representations, 2019. 3, 8", + "[39] Roseanna W Saaty. The analytic hierarchy process—what it is and how it is used. Mathematical modelling, 9(3-5):161–176, 1987. 5", + "[40] Saquib Sarfraz, Marios Koulakis, Constantin Seibold, and Rainer Stiefelhagen. Hierarchical nearest neighbor graph embedding for efficient dimensionality reduction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 336-345, 2022. 3", + "[41] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3", + "[42] Astuti Sharma, Tarun Kalluri, and Manmohan Chandraker. Instance level affinity-based transfer for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5361-5371, 2021. 2", + "[43] Jian Tang, Jingzhou Liu, Ming Zhang, and Qiaozhu Mei. Visualizing large-scale and high-dimensional data. In Proceedings of the 25th international conference on world wide web, pages 287-297, 2016. 2", + "[44] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 1, 2", + "[45] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 2, 5", + "[46] Ivan Vendrov, Ryan Kiros, Sanja Fidler, and Raquel Urtasun. Order-embeddings of images and language. International Conference on Learning Representations (ICLR), 2015. 1, 3", + "[47] Vedit Vedit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3219-3229, 2023. 2", + "[48] Luke Vilnis, Xiang Li, Shikhar Murty, and Andrew McCallum. Probabilistic embedding of knowledge graphs with box lattice measures. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 263-272, 2018. 1, 3", + "[49] Weixin Wang, Hui Wang, Guozhong Dai, and Hongan Wang. Visualization of large hierarchical data by circle packing. In Proceedings of the SIGCHI conference on Human Factors in computing systems, pages 517-520, 2006. 3" + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28836", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. Technical Report CNS-TR-201, Caltech, 2010. 2, 5", + "[51] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7031-7040, 2023. 2", + "[52] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3712-3722, 2018. 2, 4, 5", + "[53] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Mind the gap: Domain gap control for single shot domain adaptation for generative adversarial networks. In International Conference on Learning Representations, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 345 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28837", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/dd90eec4-a932-4045-a86b-cf06d264eadc_model.json b/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/dd90eec4-a932-4045-a86b-cf06d264eadc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4ad4c261ae37fa652a07c82a9c50fa072615c2f6 --- /dev/null +++ b/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/dd90eec4-a932-4045-a86b-cf06d264eadc_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.103, + 0.131, + 0.87, + 0.155 + ], + "angle": 0, + "content": "Task2Box: Box Embeddings for Modeling Asymmetric Task Relationships" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.181, + 0.696, + 0.217 + ], + "angle": 0, + "content": "Rangel Daroya Aaron Sun Subhransu Maji University of Massachusetts, Amherst" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.22, + 0.642, + 0.234 + ], + "angle": 0, + "content": "{rdaroya, aaronsun, smaji}@umass.edu" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.248, + 0.403, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.25, + 0.575, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.249, + 0.844, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.439, + 0.361, + 0.573, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.363, + 0.718, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.365, + 0.829, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.467, + 0.895, + 0.552 + ], + "angle": 0, + "content": "Figure 1. Box Embeddings of 150 Datasets of iNaturalist + CUB and Corresponding Learned Hierarchy for Class Arachnida. Each taxonomic category is treated as a separate dataset for which TASK2BOX embeddings are learned. (1) Shows the learned box embeddings where datasets from the same group (taxonomic class) have the same color. Datasets naturally cluster to their ground truth groups. (2) Shows the hierarchy learned through TASK2BOX for a specific class. The hierarchy matches the ground truth relationships based on biological classification. Orders that belong to class Arachnida are learned as boxes \\((\\widehat{\\mathbf{A}}, \\widehat{\\mathbf{B}}, \\widehat{\\mathbf{C}})\\) contained by the larger box for Arachnida; families under each of the orders are learned as smaller boxes contained by the corresponding orders they belong to." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.564, + 0.314, + 0.58 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Modeling and visualizing relationships between tasks or datasets is an important step towards solving various meta-tasks such as dataset discovery, multi-tasking, and transfer learning. However, many relationships, such as containment and transferability, are naturally asymmetric and current approaches for representation and visualization (e.g., t-SNE [44]) do not readily support this. We propose TASK2BOX, an approach to represent tasks using box embeddings—axis-aligned hyperrectangles in low dimensional spaces—that can capture asymmetric relationships between them through volumetric overlaps. We show that TASK2BOX accurately predicts unseen hierarchical relationships between nodes in ImageNet and iNaturalist datasets, as well as transferability between tasks in the Taskonomy benchmark. We also show that box embeddings estimated from task representations (e.g., CLIP [36], Task2Vec [4], or attribute based [15]) can be used to predict relationships between unseen tasks more accurately than classifiers trained on the same representations, as well as handcrafted asymmetric distances (e.g., KL divergence)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.895, + 0.657 + ], + "angle": 0, + "content": "This suggests that low-dimensional box embeddings can effectively capture these task relationships and have the added advantage of being interpretable. We use the approach to visualize relationships among publicly available image classification datasets on popular dataset hosting platform called Hugging Face." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.69, + 0.633, + 0.707 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.895, + 0.869 + ], + "angle": 0, + "content": "The success of deep learning has led to the proliferation of datasets for solving a wide range of computer vision problems. Yet, there are few tools available to enable practitioners to find datasets related to the task at hand, and to solve various meta-tasks related to it. We present TASK2Box, a method to represent tasks using axis-aligned hyperrectangles (or box embeddings). TASK2Box is framed as a learnable mapping from dataset representation to boxes, and can be trained to predict various relationships between novel tasks such as transferability, hierarchy, and overlap." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Box embeddings [48] extend order embeddings [46] by using volumetric relationships between axis-aligned hyper" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "28827" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "rectangles to encode pairwise relationships. Prior work in natural language processing has utilized box embeddings to represent the WordNet [30] hierarchy and to model conditional distributions. To model relationships between novel datasets, we develop a technique to map from Euclidean representations of datasets into the space of boxes. We explore simple image and label embedding from large vision-language models such as CLIP [36], TASK2VEC [4], and attribute-based vectors [15] as base representations of tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.473, + 0.547 + ], + "angle": 0, + "content": "We test our framework to model asymmetric relationships between nodes in iNaturalist [45] and Caltech-UCSD Birds (CUB) [50] and ImageNet [11] datasets, as well as to predict transferability on the Taskonomy benchmark [52]. Table 1 and 2 show that low-dimensional box embeddings accurately predict novel relationships between datasets seen during training, as well as relationships with novel datasets. Remarkably, TASK2BOX outperforms classifiers trained to directly predict the relationships on the same representations, suggesting that the box embedding provides a strong inductive bias for learning hierarchical relationships. We also outperform simple asymmetric distances proposed in prior work such as Kullback-Leibler (KL) divergence [4]. To model the heterogeneous tasks in the Taskonomy benchmark [52] we map each task to a set of attributes from which a box embedding is learned. Once again, we obtain significantly higher correlation between the true and predicted transferability for both existing and novel datasets compared to standard classifiers (Table 3). Such attribute-based representations can be readily derived from datasheets [15] and modelcards [31]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.473, + 0.744 + ], + "angle": 0, + "content": "Finally, the low-dimensional box embeddings have the added advantage of being interpretable. Fig. 1 and Fig. 3 show relationships on the iNaturalist+CUB and ImageNet categories, respectively. The 2D box representation allows us to readily visualize the strength and direction of task relationships based on the overlapping volumes, which is not possible using symmetric distances with Euclidean representations (e.g., t-SNE [44]). At the same time, new datasets can be embedded in constant time without needing to retrain or re-optimize. Fig. 5 uses TASK2Box to visualize relationships among 131 publicly available datasets on Hugging Face [1], a popular platform for hosting datasets. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.748, + 0.47, + 0.793 + ], + "angle": 0, + "content": "- We introduce a novel method (TASK2BOX) that uses box embeddings to learn asymmetric (e.g., hierarchical, transfer learning) dataset relationships." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.794, + 0.47, + 0.823 + ], + "angle": 0, + "content": "- We demonstrate that TASK2Box can predict the relationships of new tasks with a collection of existing tasks." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.824, + 0.47, + 0.868 + ], + "angle": 0, + "content": "- We illustrate the interpretability of our model, and the ability to visualize public classification datasets on Hugging Face." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.748, + 0.47, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The code for this project is publicly available at https://github.com/cvl-umass/task2box." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.356 + ], + "angle": 0, + "content": "Task Representations. Given a dataset \\(\\mathcal{D} = \\{(x_i, y_i)\\}_{i=1}^n\\), consisting of images \\(x_i \\in \\mathcal{X}\\) and labels \\(y_i \\in \\mathcal{Y}\\), a range of approaches have been proposed for dataset representation. The most straightforward approach involves modeling the distribution of either the images \\(x\\) or the labels \\(y\\) within the dataset independently, using embeddings referred to in prior work as \"domain\" and \"label\" embeddings [4]. To capture the joint dependency between images and labels, [4] proposed the use of the Fisher Information Matrix (FIM) derived from a \"probe network\" trained to minimize a loss function \\(\\ell(\\hat{y}, y)\\) over the dataset [22, 25, 35]. This approach leverages the similarity of FIMs to predict task transferability and for model selection. However, the utility of the FIM critically depends on the choice of the probe network and a pre-defined similarity may not accurately represent the various relationships between datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.895, + 0.539 + ], + "angle": 0, + "content": "We also investigate the use of vision-language models (VLMs) such as CLIP [36]. This model, trained on a wide range of visual domains, can generalize to tasks involving vision and language data. CLIP features have been effective for image classification [3, 9], semantic segmentation [18, 24, 26], object detection [16, 47, 51], and even closing domain gaps for performance improvement [23, 53]. Both images \\( x \\in \\mathcal{X} \\) and labels \\( y \\in \\mathcal{Y} \\) represented as text, can be mapped into a shared space using the vision encoder \\( (\\phi) \\) and text encoder \\( (\\psi) \\) of CLIP, allowing us to model the dataset as a set of image and label embeddings \\( \\left\\{\\left(\\phi(x_i), \\psi(y_i)\\right)\\right\\}_{i=1}^n \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.584 + ], + "angle": 0, + "content": "We compare FIMs with representations derived from CLIP as base representations for tasks and learn box embeddings to model a variety of relations among tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.812 + ], + "angle": 0, + "content": "Task Relations in Computer Vision. Understanding the relationships between tasks can lead to efficient solutions to new tasks. Previous work has measured task similarity by using model gradients [14] or based on their learned features [21] for grouping tasks for efficient multi-tasking. Similarly, predicting which pre-trained models will generalize the best on a new dataset could streamline model selection. Taskonomy [52] investigates transfer learning across vision tasks, varying from segmentation to pose estimation, by computing pairwise transfer distances or task affinities. These affinities are calculated by evaluating the extent to which a model trained on a source task generalizes to a target task [13, 42], though this process is computationally expensive." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Dataset Visualization. Low-dimensional Euclidean embeddings derived from UMAP [28], t-SNE [44], and LargeVis [43] are widely used to visualize relationships between datasets. They have been shown to successfully recover clusters of various data modalities by preserving the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28828" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.151 + ], + "angle": 0, + "content": "relationship of each data point with its neighbors [4, 5, 40]. In low-dimension space, relationships with other data points are defined by their Euclidean distances. However, these are commonly used to represent symmetric relations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.468, + 0.227 + ], + "angle": 0, + "content": "Visualizations using tidy trees [37], circle packing [49], or cone trees [29] organize asymmetric relations as tree-structured hierarchies in low dimension. However, cyclic data relationships cannot be properly represented for these methods (e.g., when a node has two or more parents)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.468, + 0.378 + ], + "angle": 0, + "content": "Asymmetric Distances over Datasets. Kullback-Leibler (KL) divergence between image or label distributions provides a natural way to represent asymmetric distances between datasets. TASK2VEC [4] computes the similarity between two tasks (e.g., cosine distance), and introduces asymmetry by using the complexity of the first task as a reference. The complexity is measured by the similarity of the task embedding to a \"trivial embedding\" (embedding of a task that is easy or has no examples)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.468, + 0.575 + ], + "angle": 0, + "content": "Order embeddings on images were first proposed in [46] to capture tree-structured relationships. Given an dataset of \\( P = (u, v) \\) drawn from an partially ordered set \\( (X, \\preceq_X) \\), they frame the problem as learning a mapping \\( f: (X, \\preceq_X) \\to (Y, \\preceq_Y) \\) that is order preserving, i.e., \\( u \\preceq_X v \\Longleftrightarrow f(u) \\preceq_Y f(v) \\). The reserved product order was used for \\( \\preceq_Y \\), i.e., \\( x \\preceq y \\Longleftrightarrow x_i > y_i, \\forall i \\). Box embeddings [8] generalized this framework by representing points as axis-aligned hyper-rectangles and using volumetric relationships (e.g., intersection over union) to represent asymmetric relations. They used the framework to model conditional distributions and hypernymy relations (e.g., dog \"is a\" mammal) on the WordNet graph [2, 19, 33, 38, 48]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.468, + 0.711 + ], + "angle": 0, + "content": "Hyperbolic spaces provide yet another way to model asymmetric relationships. Examples include the Poincaré disk model which uses hyperbolic cosine (cosh) to measure distance between points in a disk. Poincaré embeddings have been similarly used to represent WordNet hierarchies [32] and other relations in general graphs. Hyperbolic representations have also been proposed for representing images for efficient zero-shot learning given a taxonomic structure of the labels [27]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.711, + 0.468, + 0.831 + ], + "angle": 0, + "content": "To the best of our knowledge, no prior work has explored the use of these spaces for representing entire datasets and their effectiveness in capturing various task relationships. We adopt box embeddings in this work due to the effectiveness over alternatives in previous work [2, 6, 34, 38], ease of visualization, as well as due to open-source libraries for robust learning. However, instead of learning box embeddings directly, we learn mappings from task representations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.302, + 0.86 + ], + "angle": 0, + "content": "3. TASK2Box Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We define the problem as follows: given a collection of datasets \\(\\{\\mathcal{D}_1,\\mathcal{D}_2,\\dots ,\\mathcal{D}_m\\}\\), and an asymmetric rela" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.891, + 0.151 + ], + "angle": 0, + "content": "tionship given the pairwise relationship between datasets \\(d(\\mathcal{D}_i,\\mathcal{D}_j)\\in [0,1]\\) , we aim to encode each dataset into a low-dimension space that preserves the relationships between datasets and is interpretable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.152, + 0.892, + 0.258 + ], + "angle": 0, + "content": "To achieve this, we propose using box embeddings for encoding each of the datasets. This process involves two main steps: (1) deriving the base representation \\( e \\) of each dataset, and (2) learning a model \\( f_{\\theta}: e \\to z \\), where \\( z \\in \\mathbb{R}^{2 \\times k} \\) represents a \\( k \\)-dimensional axis-aligned hyperrectangle (i.e., box), denoted by its lower left and upper right coordinates. These steps are detailed further below." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.268, + 0.741, + 0.283 + ], + "angle": 0, + "content": "3.1. Base Task Representations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.29, + 0.892, + 0.367 + ], + "angle": 0, + "content": "Each dataset \\(\\mathcal{D} = \\{(x_i, y_i)\\}_{i=1}^n\\) is defined as a collection of pairs of images \\(x_i \\in \\mathcal{X}\\) and labels \\(y_i \\in \\mathcal{Y}\\). For obtaining a base embedding \\(e\\) for each dataset, we utilize methods such as CLIP [36, 41], TASK2VEC [4], or attribute-based approaches [15]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.548 + ], + "angle": 0, + "content": "CLIP. Using a pre-trained CLIP model [7, 20, 36, 41], we compute the mean and variance of the individual sample embeddings within each dataset. For each data sample, the image embedding is concatenated with the label embedding, the latter generated from text prompts (e.g., \"A photo of [CLS]\"). This concatenation models the joint distribution of images and labels. Eq. 1 and 2 detail how the mean and variance embeddings are derived, where \\([i,j]\\) represents the concatenation of vectors \\(i\\) and \\(j\\), \\(\\phi\\) is the vision encoder, and \\(\\psi\\) is the text encoder. The covariance is approximated as diagonal for tractability." + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.559, + 0.891, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\mu_ {C L I P} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} [ \\phi (x _ {i}), \\psi (y _ {i}) ] \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.621, + 0.891, + 0.662 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {C L I P} ^ {2} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\left[ \\phi \\left(x _ {i}\\right), \\psi \\left(y _ {i}\\right) \\right] - \\mu_ {C L I P}\\right) ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.891, + 0.735 + ], + "angle": 0, + "content": "The base representation is defined as \\( e := \\mu_{CLIP} \\in \\mathbb{R}^{2048} \\) or \\( e := [\\mu_{CLIP}, \\sigma_{CLIP}^2] \\in \\mathbb{R}^{4096} \\). A ViT-H/14 [12] pretrained on LAION-2B [41] was used to extract the embeddings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "TASK2VEC [4] encodes a dataset using the approximate Fisher Information Matrix (FIM) of a network trained on the given dataset. The FIM represents the importance of parameters in the feature extractor by perturbing the weights \\(\\hat{w}\\) of a given probe network with Gaussian noise \\(\\mathcal{N}(0,\\Lambda)\\). The precision matrix \\(\\Lambda\\) is estimated to be close to an isotropic prior \\(\\mathcal{N}(\\hat{w},\\lambda^2 I)\\) while having a good expected error. Eq. 3 is minimized to find \\(\\Lambda\\) where \\(H\\) is the cross entropy loss, \\(\\hat{w}\\) are the weights of the network, \\(\\beta\\) is the magnitude of the prior, \\(x\\in \\mathcal{X}\\), and \\(y\\in \\mathcal{Y}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28829" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.095, + 0.124, + 0.357, + 0.141 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\hat {w}; \\Lambda) = \\mathbb {E} _ {w \\sim \\mathcal {N} (\\hat {w}, \\Lambda)} [ H _ {p _ {w}, \\tilde {p}} p (y | x) ] +\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.244, + 0.144, + 0.469, + 0.162 + ], + "angle": 0, + "content": "\\[\n\\beta K L \\left(\\mathcal {N} (0, \\Lambda) \\| \\mathcal {N} (0, \\lambda^ {2} I)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.176, + 0.47, + 0.252 + ], + "angle": 0, + "content": "The matrix \\(\\Lambda\\) provides an estimate of the FIM and is approximated as a diagonal matrix. The diagonal components are used as a base representation of a dataset (\\(e := \\mathrm{FIM} \\in \\mathbb{R}^{17024}\\)). ResNet-34 [17] pretrained on ImageNet [11] is used as the probe network for all datasets." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.267, + 0.469, + 0.388 + ], + "angle": 0, + "content": "Attribute-based. A task can be characterized by a set of \\( t \\) binary attributes [15] represented as a vector of dimension \\( t \\). Some of the attributes explored for representing tasks are: (1) Is the task generative? (2) Is the task output in 2D? (3) Does the task involve camera pose estimation? Taking these 3 characteristics, for example, we can represent a 2D segmentation task as the vector \\( e = [0,1,0] \\) as a discriminative 2D task that does not need camera poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.469, + 0.495 + ], + "angle": 0, + "content": "Tasks in Taskonomy [52] involving multiple modalities benefit from this attribute-based representation due to its model independence. This approach also enables generalization to unseen tasks by identifying the presence or absence of various characteristics. Each vision task in Taskonomy is represented with 15 attributes, resulting in vectors \\( e \\in \\mathbb{R}^{15} \\). The full list of attributes is in Appendix ??" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.507, + 0.32, + 0.524 + ], + "angle": 0, + "content": "3.2. Learning Box Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.531, + 0.47, + 0.669 + ], + "angle": 0, + "content": "From a base task representation \\( e \\), we learn the parameters \\( \\theta \\) of a model \\( f_{\\theta}: e \\to z \\) that preserves the asymmetric similarity function \\( d(\\mathcal{D}_i, \\mathcal{D}_j) \\) between any two datasets \\( \\mathcal{D}_i \\) and \\( \\mathcal{D}_j \\). In Eq. 4, the learning objective is shown where \\( \\mathcal{L}_E \\) is a loss function (mean squared error), \\( \\mathcal{L}_D \\) is a distance function, \\( \\mathcal{L}_R \\) is a regularization term, and \\( \\lambda \\) is a hyperparameter. The embeddings \\( z_i, z_j \\in \\mathbb{R}^{2 \\times k} \\) represent the coordinates of the lower left and the upper right corners of the respective \\( k \\) dimensional boxes, with \\( f_{\\theta}(e_i) = z_i \\) and \\( f_{\\theta}(e_j) = z_j \\)." + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.697, + 0.469, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\theta} = \\underset {\\theta} {\\operatorname {a r g m i n}} \\sum_ {i, j} \\left(\\mathcal {L} _ {E} \\left(d \\left(\\mathcal {D} _ {i}, \\mathcal {D} _ {j}\\right), d _ {b o x} \\left(z _ {i}, z _ {j}\\right)\\right) \\right. \\\\ + \\lambda \\mathcal {L} _ {D} \\left(z _ {i}, z _ {j}\\right) + \\mathcal {L} _ {R}) \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The asymmetric relationship between box embeddings, denoted as \\( d_{box}(z_i,z_j) \\), is computed in Eq. 5 by calculating the volume of the intersection between \\( z_i \\) and \\( z_j \\), normalized by the volume of \\( z_i \\). For \\( z_i \\) to be fully contained inside \\( z_j \\) (\\( z_i \\subset z_j \\)), it is required that \\( d_{box}(z_i,z_j) = 1 \\). Conversely, for \\( z_j \\) to only partially contain \\( z_i \\), \\( d_{box}(z_j,z_i) \\) must fall within the range (0,1). Fig. 2 illustrates this through a 2-dimensional example, where \\( z_1 \\) represents the box embedding for Canidae, and \\( z_2 \\) for Mammalia." + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.092, + 0.81, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.232, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Figure 2. TASK2Box Embeddings in 2D for Mammalia, Canidae, and Amphibia Datasets from iNaturalist. Each embedding represents the coordinates of the lower left and upper right corners of each box/rectangle. Since Canidae \\((z_{1})\\) is a proper subset of Mammalia \\((z_{2})\\): \\(d_{box}(z_{1}, z_{2}) = 1\\) and \\(d_{box}(z_{2}, z_{1}) = 0.1\\)." + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.329, + 0.892, + 0.361 + ], + "angle": 0, + "content": "\\[\nd _ {b o x} \\left(z _ {i}, z _ {j}\\right) = \\frac {\\operatorname {v o l} \\left(z _ {i} \\cap z _ {j}\\right)}{\\operatorname {v o l} \\left(z _ {i}\\right)} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.892, + 0.597 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_D\\) is applied to datasets where \\(d(\\mathcal{D}_i, \\mathcal{D}_j) > 0\\) such that the Euclidean distance between the center coordinates of \\(z_i\\) and \\(z_j\\) is minimized when starting from a non-overlapping state. This allows non-overlapping embeddings to move closer to each other, complementing \\(\\mathcal{L}_E\\) to learn relationships. \\(\\mathcal{L}_R\\) encourages solutions with regular-shaped boxes for better interpretability. The formulation of \\(\\mathcal{L}_R\\), given in Eq. 6, applies to a \\(k\\)-dimensional box embedding \\(z_i\\), where \\(s_a\\) represents the size of the \\(a\\)-th dimension (for example, width), and \\(\\alpha, \\beta\\) are hyperparameters. To prevent the trivial solution of minimizing box volume to zero, the inverse of the box volume is included. It's crucial to normalize the terms with respect to the embedding dimension, as the first and second terms scale quadratically and exponentially with dimension increase, respectively." + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.622, + 0.892, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R} = \\left(\\frac {\\alpha}{k ^ {2}} \\sum_ {a = 1} ^ {k} \\sum_ {b = a + 1} ^ {k} \\left| s _ {a} - s _ {b} \\right|\\right) + \\beta (\\operatorname {v o l} (z _ {i})) ^ {- 1 / k} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.677, + 0.892, + 0.737 + ], + "angle": 0, + "content": "The architecture of \\( f \\) consists of three fully-connected layers followed by two linear heads: one predicts a \\( k \\)-dimensional lower-left coordinate, and the other predicts the \\( k \\)-dimensional sizes of each box dimension." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.753, + 0.634, + 0.771 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.89, + 0.809 + ], + "angle": 0, + "content": "We consider the following goals to evaluate the capability of Task2Box to represent datasets:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.811, + 0.892, + 0.856 + ], + "angle": 0, + "content": "1. Given a collection of existing datasets \\(\\mathcal{D}_E\\) and a subset of pairwise relationships \\(\\mathcal{R}\\): can the model generalize on unseen relationships \\(\\mathcal{R}'\\) within \\(\\mathcal{D}_E\\) where \\(\\mathcal{R}' \\cap \\mathcal{R} = \\emptyset\\)?" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.891, + 0.901 + ], + "angle": 0, + "content": "2. Given a collection of novel datasets \\(\\mathcal{D}_N\\) not seen during training: can the model accurately identify the relationships with the existing datasets \\(\\mathcal{D}_E\\)?" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28830" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.179, + 0.089, + 0.796, + 0.292 + ], + "angle": 0, + "content": "
iNaturalist + CUB
MethodExisting DatasetsNovel Datasets
μCLIP[μ,σ2]CLIPFIMμCLIP[μ,σ2]CLIPFIM
TASK2Box (2D)69.23%67.84%39.61%50.07%39.66%10.06%
TASK2Box (3D)79.66%79.35%57.63%70.04%64.53%20.65%
TASK2Box (5D)84.67%82.41%79.72%73.79%72.11%34.88%
MLP Classifier45.25%61.45%26.34%39.06%44.54%19.90%
Linear Classifier4.40%3.11%7.06%4.77%5.87%15.92%
KL Divergence-6.58%7.94%-5.90%0.00%
Asymmetric Cosine9.29%11.54%2.83%1.47%1.47%1.47%
Asymmetric Euclidean1.71%1.71%8.53%1.47%1.47%1.91%
Random2.06%1.49%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.301, + 0.895, + 0.373 + ], + "angle": 0, + "content": "Table 1. Average F1 Score for Predicting Hierarchical Relationships on iNaturalist + CUB Dataset. For all feature types, TASK2Box outperforms other methods by more than \\(20\\%\\) on existing datasets and more than \\(10\\%\\) on novel datasets. The best-performing model is in bold, and the second-best is underlined. Results on Novel Datasets demonstrate that our model can generalize beyond the seen tasks it has previously seen. The dimensions such as 2D, 3D, and 5D refer to different box dimensionalities. Results using KL divergence for \\(\\mu_{CLIP}\\) are not shown since it is not a distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.472, + 0.481 + ], + "angle": 0, + "content": "The goals are evaluated through two experimental setups, demonstrating our model's ability to predict various types of relationships. The configurations, and their corresponding datasets, relationships, and metrics, are detailed below. Baseline methods for performance comparison are also reviewed. Implementation details are in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.49, + 0.269, + 0.507 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.356, + 0.53 + ], + "angle": 0, + "content": "4.1.1 Hierarchical Task Relationships" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.674 + ], + "angle": 0, + "content": "We use a combination of iNaturalist [45] and Caltech-UCSD Birds (CUB) [50], and instrument-related classes in ImageNet [11] to evaluate the ability of TASK2Box to represent hierarchical relationships. The first two are composed of images of various species, and the third is composed of images of various objects. The classes naturally follow a hierarchical form based on biological classification (taxonomy of iNaturalist+CUB species), and on semantic relations (hypernymy of objects in ImageNet)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.795 + ], + "angle": 0, + "content": "Datasets. For iNaturalist+CUB, the datasets are defined as the classes, the orders, and the families in the taxonomy that contain a significant number of samples per dataset as in [4]. There are 47 classes, 202 orders, and 589 families for a total of 838 datasets. For ImageNet, the instrument-related objects were processed using WordNet [30] to obtain hierarchical information between classes. This resulted in 131 datasets for training and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Dataset Relationships. For any two datasets, their relationship is captured as \\( d(\\mathcal{D}_i,\\mathcal{D}_j)\\in \\{0,1\\} \\), where \\( d(\\mathcal{D}_i,\\mathcal{D}_j) = 1 \\) if and only if \\( \\mathcal{D}_i\\subset \\mathcal{D}_j \\), and \\( d(\\mathcal{D}_i,\\mathcal{D}_j) = 0 \\) otherwise. Fig. 2 provides an example: Canidae \\( (\\mathcal{D}_1) \\) is a family within the class Mammalia \\( (\\mathcal{D}_2) \\); thus, \\( d(\\mathcal{D}_1,\\mathcal{D}_2) = 1 \\) and \\( d(\\mathcal{D}_2,\\mathcal{D}_1) = 0 \\). Meanwhile, Amphibia \\( (\\mathcal{D}_3) \\) is unrelated to either dataset, resulting in \\( d(\\mathcal{D}_1,\\mathcal{D}_3) = 0 \\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.389, + 0.895, + 0.619 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate the ability of the model to classify the presence of containment relationships between datasets using the F1 score due to the imbalance between positive and negative relationships. To obtain F1 score, precision and recall are first calculated. Precision is computed as the ratio of true positive pairs predicted to the total number of positive predictions. Recall is the ratio of true positive pairs predicted to the total number of true positive pairs. F1 score is then reported as the harmonic mean between the precision and recall. These calculations are done on both (1) the unseen relationships \\(\\mathcal{R}'\\) within existing datasets \\(\\mathcal{D}_E\\) and (2) the relationships between novel datasets \\(\\mathcal{D}_N\\) and existing datasets \\(\\mathcal{D}_E\\). For the latter, we evaluate the relationships in both directions, i.e., \\(\\left\\{\\left(\\mathcal{D}_e,\\mathcal{D}_n\\right)\\forall \\mathcal{D}_e\\in \\mathcal{D}_E\\right\\} \\cup \\left\\{\\left(\\mathcal{D}_n,\\mathcal{D}_e\\right)\\forall \\mathcal{D}_e\\in \\mathcal{D}_E\\right\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.635, + 0.81, + 0.65 + ], + "angle": 0, + "content": "4.1.2 Transfer Learning Between Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.689 + ], + "angle": 0, + "content": "The Taskonomy [52] benchmark is used to evaluate the ability of TASK2BOX to predict task affinities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Datasets. Taskonomy defines a set of 25 visual tasks with corresponding pairwise task affinities. These tasks range from object detection, semantic segmentation, pose estimation, and more. We treat each visual task as a dataset. The tasks are described in Appendix ??" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Dataset Relationships. In contrast to the containment relationship defined in § 4.1.1, Taskonomy quantifies relationships with task affinity measured by the performance gain achieved by transfer learning from a source dataset \\(\\mathcal{D}_j\\) to a target dataset \\(\\mathcal{D}_i\\). The values are computed using Analytic Hierarchy Process [39, 52]. Dataset relationships are computed and normalized based on an ordinal approach and determined by the percentage of images that transfer well to a target task given a set of source tasks as detailed in [52]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28831" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.179, + 0.089, + 0.796, + 0.292 + ], + "angle": 0, + "content": "
ImageNet
MethodExisting DatasetsNovel Datasets
μCLIP[μ,σ2]CLIPFIMμCLIP[μ,σ2]CLIPFIM
TASK2Box (2D)62.72%63.33%31.32%47.76%48.35%9.46%
TASK2Box (3D)83.12%82.79%58.16%73.06%66.48%24.70%
TASK2Box (5D)90.58%88.48%64.91%76.95%78.84%37.39%
MLP Classifier54.20%60.43%44.85%62.24%64.56%41.22%
Linear Classifier9.84%8.33%25.46%11.89%11.87%22.98%
KL-5.53%9.90%-10.41%0.00%
Asymmetric Cosine4.28%4.28%6.92%4.52%4.52%0.00%
Asymmetric Euclidean3.73%3.73%7.16%4.52%4.52%4.73%
Random3.64%5.02%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.301, + 0.895, + 0.345 + ], + "angle": 0, + "content": "Table 2. Average F1 Score for Predicting Hierarchical Relationships on ImageNet Instruments. TASK2BOX is shown to generalize well on both predicting relationships between existing datasets (Existing Datasets), and predicting box embeddings of new datasets and their relationships with existing datasets (Novel Datasets)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.47, + 0.422 + ], + "angle": 0, + "content": "For a pair of datasets \\((\\mathcal{D}_i,\\mathcal{D}_j)\\), the task affinity is defined as \\(d(\\mathcal{D}_i,\\mathcal{D}_j)\\in [0,1]\\). The higher the task affinity to a target dataset \\(\\mathcal{D}_i\\) from a source dataset \\(\\mathcal{D}_j\\), then \\(d(\\mathcal{D}_i,\\mathcal{D}_j)\\) gets closer to 1, where a value of 1 would show \\(\\mathcal{D}_i\\subset \\mathcal{D}_j\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.471, + 0.528 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate the prediction of the model based on the Spearman correlation between the ground truth task affinity values, and the predicted values based on the box distances in Eq. 5. Similar to § 4.1.1, we evaluate on both (1) unseen relationships \\(\\mathcal{R}'\\) for existing datasets, and (2) on relationships between unseen novel datasets \\(\\mathcal{D}_N\\) and existing datasets \\(\\mathcal{D}_E\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.539, + 0.25, + 0.554 + ], + "angle": 0, + "content": "4.2. Baseline Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.47, + 0.655 + ], + "angle": 0, + "content": "We compare the performance of TASK2Box with alternative models and simple asymmetric distances proposed in prior work. For hierarchical relationships, given two datasets \\(\\mathcal{D}_i,\\mathcal{D}_j\\), the models predict \\(\\hat{d} (\\mathcal{D}_i,\\mathcal{D}_j)\\in \\{0,1\\}\\). For task affinity, the models predict \\(\\hat{d} (\\mathcal{D}_i,\\mathcal{D}_j)\\in [0,1]\\). The structure for the various methods are discussed below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.701 + ], + "angle": 0, + "content": "Linear Model. A linear model is trained to predict the relationship value between two datasets. The input is the concatenation of the base representation \\( e \\) for the two datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.47, + 0.732 + ], + "angle": 0, + "content": "MLP Model. A 4-layer MLP is used instead for prediction on the same inputs as the linear model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.855 + ], + "angle": 0, + "content": "KL Divergence. Each dataset is treated as a multivariate Gaussian using the mean and variance of the image and label features (CLIP) or directly as the FIM. The optimal threshold \\( t \\) is selected for the minimum distance between two dataset distributions \\( \\mathrm{KL}(\\mathcal{D}_i||\\mathcal{D}_j) < t \\) for a prediction of \\( \\hat{d} (\\mathcal{D}_i,\\mathcal{D}_j) = 1 \\), and \\( \\hat{d} (\\mathcal{D}_i,\\mathcal{D}_j) = 0 \\) otherwise. The F1 Score (hierarchical) or the correlation (task affinity) is the objective for selecting \\( t \\) on the train set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Asymmetric Cosine Similarity. The cosine similarity \\( d_{\\text{cos}} \\) is a symmetric measure between two embeddings \\( e_i, e_j \\). An asymmetric variant was proposed in [4] and shown in Eq. 7" + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.362, + 0.821, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.531, + 0.892, + 0.587 + ], + "angle": 0, + "content": "Figure 3. Visualization of Instrument-related Datasets in ImageNet. Datasets that belong to the same superset are shaded in the same color. TASK2Box learns the hierarchy of various groups, and clusters similar datasets closer." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.6, + 0.892, + 0.647 + ], + "angle": 0, + "content": "by considering the similarity of \\( e_i \\) and \\( e_j \\) relative to the complexity of \\( e_i \\). The complexity of is measured as the distance to the trivial embedding \\( e_o \\), and \\( \\alpha \\) is a hyperparameter." + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.662, + 0.892, + 0.679 + ], + "angle": 0, + "content": "\\[\nd _ {a s y m} \\left(e _ {i}, e _ {j}\\right) = d _ {c o s} \\left(e _ {i}, e _ {j}\\right) - \\alpha d _ {c o s} \\left(e _ {i}, e _ {o}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.686, + 0.892, + 0.731 + ], + "angle": 0, + "content": "An optimal threshold \\( t \\) is found on the train set where \\( d_{asym}(e_i, e_j) < t \\) results to a prediction \\( \\hat{d}(\\mathcal{D}_i, \\mathcal{D}_j) = 1 \\). The same threshold \\( t \\) is used for test set evaluation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.732, + 0.892, + 0.776 + ], + "angle": 0, + "content": "Asymmetric Euclidean Similarity. The asymmetric similarity is computed as in Eq. 7 but uses the Euclidean distance instead of cosine." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.777, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Random. The probability of containment (for hierarchical) or the value of the task affinity is uniformly random." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.822, + 0.715, + 0.837 + ], + "angle": 0, + "content": "5. Results and Discussion" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.742, + 0.864 + ], + "angle": 0, + "content": "5.1. Hierarchical Relationships" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Tables 1 and 2 present the average F1 score for predicting hierarchical relationships on iNaturalist+CUB and Ima" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28832" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "geNet datasets. Various methods of extracting dataset embeddings were evaluated. TASK2BOX outperforms baseline methods, indicating that TASK2BOX can both generalize across unseen relationships (Existing Datasets), and to accurately represent relationships with existing datasets for unseen datasets (Novel Datasets). CLIP features also generalize better than FIM features, perhaps because the CLIP multi-modal embedding of images and labels was trained across a broad set of domains[36]. However, while CLIP is confined to image and text modalities, FIM can accommodate any modality by adapting the probe network's last layer for different prediction tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.44 + ], + "angle": 0, + "content": "Beyond predicting relationships, our method allows the visualization of datasets. Fig. 1 illustrates this for a subset of datasets in iNaturalist and CUB using TASK2Box (2D), and Fig. 3 for ImageNet. The hierarchical organization of the datasets is apparent in the TASK2Box representation – datasets that contain others appear as larger boxes, while more specialized datasets are depicted as smaller boxes netted with borader, more general dataset boxes. However, while 2D visualization offers insights, the flat surface may restrict the representation of complex relationships. Representations in higher dimensions is explored in § 5.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.45, + 0.214, + 0.466 + ], + "angle": 0, + "content": "5.2. Task Affinity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.625 + ], + "angle": 0, + "content": "Table 3 presents the results of task representations in Taskonomy. Our method shows that, even with attribute-based embeddings, it can learn box embeddings through relationship supervision between datasets. TASK2Box is shown to correlate highly with the ground truth task affinities compared to other methods. Given that only 25 tasks are available in Taskonomy, with 3 held out of training as novel datasets, there is a higher uncertainty in predicting unseen datasets. We expect that performance on novel datasets will improve with more datasets available during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.471, + 0.792 + ], + "angle": 0, + "content": "Fig. 4 displays the learned representation with TASK2Box (2D). Each subfigure in (a)-(c) represents a subset of tasks identified as having strong transfer relationships. TASK2Box not only identifies but also visually represents related tasks suitable for fine-tuning. The small highlighted boxes indicate target tasks, while the larger enclosing boxes represent source tasks. Although other tasks may not be proper subsets, the box distance in Eq. 5 provides an estimate for task affinity. A small box distance between two datasets, \\( d_{box}(z_1, z_2) \\), suggests lower transfer performance from a source task \\( z_2 \\) to a target task \\( z_1 \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.323, + 0.817 + ], + "angle": 0, + "content": "5.3. Visualizing Public Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Apart from predicting both hierarchical relationships and quantified task affinities, our method can also be used to visualize public datasets that lack available ground truth relationships. Using a set of vision tasks from Hugging Face [1], \\(\\mu_{CLIP} + \\text{TASK2Box}\\) (2D) were utilized to pre" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.098, + 0.629, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.099, + 0.75, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.1, + 0.871, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.179, + 0.892, + 0.263 + ], + "angle": 0, + "content": "Figure 4. Visualization of Tasks in Taskonomy showing Source Tasks that Transfer Well to Target Tasks (shaded). (a) Jigsaw and Triplet Fixated Camera Pose estimation are source tasks that transfer well to Depth estimation. (b) and (c) show different source tasks (larger boxes) that transfer well to the shaded boxes of Denoising Autoencoder and Surface Normal, respectively." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.267, + 0.88, + 0.402 + ], + "angle": 0, + "content": "
MethodExisting DatasetsNovel Datasets
Spearman's ρSpearman's ρ
TASK2Box (2D)0.85 ± 0.060.12 ± 0.21
TASK2Box (3D)0.93 ± 0.020.48 ± 0.24
TASK2Box (5D)0.94 ± 0.030.39 ± 0.22
MLP0.88 ± 0.060.31 ± 0.18
Linear0.75 ± 0.110.40 ± 0.24
Random0.05 ± 0.140.15 ± 0.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.412, + 0.892, + 0.468 + ], + "angle": 0, + "content": "Table 3. Spearman Correlation and Standard Deviation between Predicted and Ground Truth Task Affinities on Taskonomy. Our method shows higher correlation with the task affinities compared to the baseline. Attribute-based embeddings were used." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.715 + ], + "angle": 0, + "content": "dict dataset embeddings. A constraint on the box sizes was added to reflect information about the dataset sizes. Fig. 5 displays the results of TASK2Box on 131 datasets from Hugging Face, where similar datasets, such as sentiments and documents, are shown to overlap. This approach allows for the analysis and visualization of dataset variations and similarities even with only samples of images and labels. Simultaneously, the sizes of various datasets can be visualized as box sizes – with larger datasets containing more samples depicted as larger boxes. This tool enables computer vision practitioners to see how their datasets compare with other existing datasets. Beyond visualizing the variation of available datasets (based on the number of clusters), it can expedite the process of finding suitable data sources by examining embedding overlaps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.706, + 0.74 + ], + "angle": 0, + "content": "5.4. Analysis of TASK2BOX" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.891, + 0.779 + ], + "angle": 0, + "content": "We discuss properties of TASK2Box below. Additional insights are also included in Appendix ??" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Using a Box Prior Improves Performance. While TASK2Box was trained with a 3-layer MLP, it achieved significantly better performance than the MLP baseline without a box prior. Tables 1, 2, and 3 demonstrate our method outperforming the baselines. Representing each dataset as an entity with shape and volume (such as a box), instead of solely learning relationships from embeddings, proves effective for generalization on unseen relationships. This" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28833" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.09, + 0.862, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.306, + 0.895, + 0.361 + ], + "angle": 0, + "content": "Figure 5. Visualizing Image Classification Datasets in Hugging Face. The sample data points annotated on the highlighted datasets show that common tasks overlap with each other (e.g., sentiment classification and document classification datasets). Although labels could slightly differ between datasets, TASK2BOX can infer the level of similarity and represent it as the amount of overlap. The embedding size (box area) also shows the number of available data." + }, + { + "type": "image", + "bbox": [ + 0.148, + 0.378, + 0.401, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.532, + 0.471, + 0.575 + ], + "angle": 0, + "content": "Figure 6. Effect of TASK2Box Embedding Dimension on the Accuracy of Predicting Task Relations. As the dimension increases, the performance generally increases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.47, + 0.655 + ], + "angle": 0, + "content": "improvement could be attributed to the explicit modeling of relationships through physical shapes, which enforces consistency: if \\( z_{i} \\subset z_{j} \\) and \\( z_{j} \\) has no overlap with \\( z_{k} \\), then having a physical representation ensures that \\( z_{i} \\cap z_{k} = \\emptyset \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.471, + 0.838 + ], + "angle": 0, + "content": "TASK2BOX can Represent Relations in Varying Dimensions. Fig. 6 illustrates the performance of TASK2BOX as the box dimension increases. Representing relationships in two dimensions results in easily interpretable embeddings; however, modeling complex task relationships could benefit from expansion to higher dimensions. Projecting to higher dimensions may yield even better performance, as relationships between datasets become more accurately represented, given the model's increased representation capacity. Nonetheless, as the dimensionality further increases, learning and visualizing the embedding space also becomes more challenging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Using Boxes to Represent Tasks Enables Effective Calculations on Embeddings. Boxes offer the advantage of being closed under intersection, meaning the intersection of two boxes results in another box, a property not shared by" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.378, + 0.892, + 0.44 + ], + "angle": 0, + "content": "circles or ellipses. Among region-based embeddings that utilize geometric objects (e.g., cones, disks, and boxes) to represent entities, operations involving boxes are the most straightforward to calculate [10, 38]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.456, + 0.619, + 0.471 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.892, + 0.649 + ], + "angle": 0, + "content": "We present a novel method that learns low-dimensional, interpretable embeddings of various task relationships. With box representations, we demonstrate that asymmetric relationships, such as task affinities and hierarchies, can be accurately modeled using various base representations. While CLIP embeddings have been shown to outperform FIM when integrated with our model, future work could investigate how CLIP might be adapted to modalities beyond text and images. Attribute-based features offer a viable alternative, and can be extracted from from datasheets using natural language processing techniques." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.892, + 0.801 + ], + "angle": 0, + "content": "The distinct properties of TASK2BOX were analyzed, revealing its ability to perform effectively across varying dimensions, and to model and visualize overlaps among public classification datasets on Hugging Face. This could enable computer vision practitioners to assess dataset utility for a task in hand. Although our model successfully represents task relationships, it does not incorporate information about optimal training procedures and model architecture. Future work could explore the inclusion of additional information for a more detailed understanding of the task space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by awards from the National Science Foundation (2329927 and 1749833) and the NASA AIST program. The experiments were performed on the University of Massachusetts GPU cluster funded by the Mass. Technology Collaborative." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28834" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Hugging Face datasets. https://huggingface.co/datasets?taskCategories \\(\\equiv\\) taskCategories:image-classification.2,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.471, + 0.212 + ], + "angle": 0, + "content": "[2] Ralph Abboud, Ismail Ceylan, Thomas Lukasiewicz, and Tommaso Salvatori. Boxe: A box embedding model for knowledge base completion. Advances in Neural Information Processing Systems, 33:9649-9661, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.471, + 0.284 + ], + "angle": 0, + "content": "[3] Rabab Abdelfattah, Qing Guo, Xiaoguang Li, Xiaofeng Wang, and Song Wang. Cdul: Clip-driven unsupervised learning for multi-label image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1348-1357, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[4] Alessandro Achille, Michael Lam, Rahul Tewari, Avinash Ravichandran, Subhransu Maji, Charless C Fowlkes, Stefano Soatto, and Pietro Perona. Task2vec: Task embedding for meta-learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6430-6439, 2019. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.471, + 0.422 + ], + "angle": 0, + "content": "[5] Sanjeev Arora, Wei Hu, and Pravesh K. Kothari. An analysis of the t-sne algorithm for data visualization. In Proceedings of the 31st Conference On Learning Theory, pages 1455-1462. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[6] Michael Boratko, Dongxu Zhang, Nicholas Monath, Luke Vilnis, Kenneth L Clarkson, and Andrew McCallum. Capacity and bias of learned geometric embeddings for directed graphs. Advances in Neural Information Processing Systems, 34:16423-16436, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.471, + 0.578 + ], + "angle": 0, + "content": "[7] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.58, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[8] Tejas Chheda, Purujit Goyal, Trang Tran, Dhruvesh Patel, Michael Boratko, Shib Sankar Dasgupta, and Andrew McCallum. Box embeddings: An open-source library for representation learning using geometric structures. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 203-211, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.677, + 0.471, + 0.732 + ], + "angle": 0, + "content": "[9] Marcos V Conde and Kerem Turgutlu. Clip-art: Contrastive pre-training for fine-grained art classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3956-3960, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[10] Shib Dasgupta, Michael Boratko, Siddhartha Mishra, Shriya Atmakuri, Dhruvesh Patel, Xiang Li, and Andrew McCallum. Word2box: Capturing set-theoretic semantics of words using box embeddings. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations (ICLR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[13] Kshitij Dwivedi and Gemma Roig. Representation similarity analysis for efficient task taxonomy & transfer learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12387-12396, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[14] Chris Fifty, Ehsan Amid, Zhe Zhao, Tianhe Yu, Rohan Anil, and Chelsea Finn. Efficiently identifying task groupings for multi-task learning. Advances in Neural Information Processing Systems, 34:27503-27516, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[15] Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé Iii, and Kate Crawford. Datasheets for datasets. Communications of the ACM, 64(12):86-92, 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[16] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.407, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[18] Wenbin He, Suphanut Jamonnak, Liang Gou, and Liu Ren. Clip-s4: Language-guided self-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11207-11216, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[19] EunJeong Hwang, Jay-Yoon Lee, Tianyi Yang, Dhruvesh Patel, Dongxu Zhang, and Andrew McCallum. Event-event relation extraction using probabilistic box embedding. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 235–244, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[20] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[21] Zhuoliang Kang, Kristen Grauman, and Fei Sha. Learning with whom to share in multi-task feature learning. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pages 521-528, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[22] Ryo Karakida, Shotaro Akaho, and Shun-ichi Amari. Universal statistics of fisher information in deep neural networks: Mean field approach. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 1032–1041. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Zhengfeng Lai, Noranart Vesdapunt, Ning Zhou, Jun Wu, Cong Phuoc Huynh, Xuelu Li, Kah Kuen Fu, and Chen-Nee Chuah. Padclip: Pseudo-labeling with adaptive debiasing in clip for unsupervised domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16155-16165, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28835" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[24] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7061-7070, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[25] Zhibin Liao, Tom Drummond, Ian Reid, and Gustavo Carneiro. Approximate fisher information matrix to characterize the training of deep neural networks. IEEE transactions on pattern analysis and machine intelligence, 42(1): 15-26, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.471, + 0.331 + ], + "angle": 0, + "content": "[26] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15305-15314, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.471, + 0.403 + ], + "angle": 0, + "content": "[27] Shaoteng Liu, Jingjing Chen, Liangming Pan, Chong-Wah Ngo, Tat-Seng Chua, and Yu-Gang Jiang. Hyperbolic visual embedding learning for zero-shot recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.471, + 0.458 + ], + "angle": 0, + "content": "[28] Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. Umap: Uniform manifold approximation and projection. Journal of Open Source Software, 3(29):861, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.472, + 0.502 + ], + "angle": 0, + "content": "[29] Guy Melancon and Ivan Herman. Circular drawings of rooted trees. CWI (Centre for Mathematics and Computer Science), 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[30] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.615 + ], + "angle": 0, + "content": "[31] Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. Model cards for model reporting. In Proceedings of the conference on fairness, accountability, and transparency, pages 220-229, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[32] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[33] Dhruvesh Patel, Shib Sankar Dasgupta, Michael Boratko, Xiang Li, Luke Vilnis, and Andrew McCallum. Representing joint hierarchies with box embeddings. In Automated Knowledge Base Construction, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[34] Dhruvesh Patel, Pavitra Dangati, Jay-Yoon Lee, Michael Boratko, and Andrew McCallum. Modeling label space interactions in multi-label classification using box embeddings. In International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[35] Jeffrey Pennington and Pratik Worah. The spectrum of the fisher information matrix of a single-hidden-layer neural network. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "sion. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[37] Edward M. Reingold and John S. Tilford. Tidier drawings of trees. IEEE Transactions on software Engineering, (2): 223-228, 1981. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[38] Hongyu Ren, Weihua Hu, and Jure Leskovec. Query2box: Reasoning over knowledge graphs in vector space using box embeddings. In International Conference on Learning Representations, 2019. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[39] Roseanna W Saaty. The analytic hierarchy process—what it is and how it is used. Mathematical modelling, 9(3-5):161–176, 1987. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[40] Saquib Sarfraz, Marios Koulakis, Constantin Seibold, and Rainer Stiefelhagen. Hierarchical nearest neighbor graph embedding for efficient dimensionality reduction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 336-345, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[41] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[42] Astuti Sharma, Tarun Kalluri, and Manmohan Chandraker. Instance level affinity-based transfer for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5361-5371, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[43] Jian Tang, Jingzhou Liu, Ming Zhang, and Qiaozhu Mei. Visualizing large-scale and high-dimensional data. In Proceedings of the 25th international conference on world wide web, pages 287-297, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[44] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[45] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.706 + ], + "angle": 0, + "content": "[46] Ivan Vendrov, Ryan Kiros, Sanja Fidler, and Raquel Urtasun. Order-embeddings of images and language. International Conference on Learning Representations (ICLR), 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[47] Vedit Vedit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3219-3229, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[48] Luke Vilnis, Xiang Li, Shikhar Murty, and Andrew McCallum. Probabilistic embedding of knowledge graphs with box lattice measures. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 263-272, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[49] Weixin Wang, Hui Wang, Guozhong Dai, and Hongan Wang. Visualization of large hierarchical data by circle packing. In Proceedings of the SIGCHI conference on Human Factors in computing systems, pages 517-520, 2006. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28836" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[50] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. Technical Report CNS-TR-201, Caltech, 2010. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[51] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7031-7040, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.471, + 0.289 + ], + "angle": 0, + "content": "[52] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3712-3722, 2018. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.29, + 0.471, + 0.346 + ], + "angle": 0, + "content": "[53] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Mind the gap: Domain gap control for single shot domain adaptation for generative adversarial networks. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "28837" + } + ] +] \ No newline at end of file diff --git a/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/dd90eec4-a932-4045-a86b-cf06d264eadc_origin.pdf b/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/dd90eec4-a932-4045-a86b-cf06d264eadc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7c32cd108ad1bf4aeac50cfd8566437fdda3e857 --- /dev/null +++ b/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/dd90eec4-a932-4045-a86b-cf06d264eadc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57f2df48d86d6a0c97bf50c1e5c13663156549de37a268ab2bb6de76ce9c3881 +size 2826130 diff --git a/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/full.md b/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a0466c494ec56ed3eefc6d7c5d54c853687c3f1f --- /dev/null +++ b/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/full.md @@ -0,0 +1,322 @@ +# Task2Box: Box Embeddings for Modeling Asymmetric Task Relationships + +Rangel Daroya Aaron Sun Subhransu Maji University of Massachusetts, Amherst + +{rdaroya, aaronsun, smaji}@umass.edu + +![](images/2db75518f57383f685a21cbd2d56d03d8bd8ebc92acb04ffc0d94457090b098f.jpg) +Figure 1. Box Embeddings of 150 Datasets of iNaturalist + CUB and Corresponding Learned Hierarchy for Class Arachnida. Each taxonomic category is treated as a separate dataset for which TASK2BOX embeddings are learned. (1) Shows the learned box embeddings where datasets from the same group (taxonomic class) have the same color. Datasets naturally cluster to their ground truth groups. (2) Shows the hierarchy learned through TASK2BOX for a specific class. The hierarchy matches the ground truth relationships based on biological classification. Orders that belong to class Arachnida are learned as boxes $(\widehat{\mathbf{A}}, \widehat{\mathbf{B}}, \widehat{\mathbf{C}})$ contained by the larger box for Arachnida; families under each of the orders are learned as smaller boxes contained by the corresponding orders they belong to. + +![](images/4b29e3406ad95f3b3e8cbaf0d6d2fc92366ff9b8429bd7445fdd735dfcd529df.jpg) + +![](images/2c91020a3442cd53accdbd1012ce9356cc38ff8cfa65532288072065c36693b9.jpg) + +![](images/3011d3bdea7341235c330d778344392cb670b0aa0883755756997e646665b8ff.jpg) + +![](images/62f14903af353c37192084c19c18874a14335fc2c1f5aabc9766e2d2f3545967.jpg) + +![](images/d0c937ac1294d9baf50a5482d3302d0ab59162daa459f4460ace8ba0db229075.jpg) + +# Abstract + +Modeling and visualizing relationships between tasks or datasets is an important step towards solving various meta-tasks such as dataset discovery, multi-tasking, and transfer learning. However, many relationships, such as containment and transferability, are naturally asymmetric and current approaches for representation and visualization (e.g., t-SNE [44]) do not readily support this. We propose TASK2BOX, an approach to represent tasks using box embeddings—axis-aligned hyperrectangles in low dimensional spaces—that can capture asymmetric relationships between them through volumetric overlaps. We show that TASK2BOX accurately predicts unseen hierarchical relationships between nodes in ImageNet and iNaturalist datasets, as well as transferability between tasks in the Taskonomy benchmark. We also show that box embeddings estimated from task representations (e.g., CLIP [36], Task2Vec [4], or attribute based [15]) can be used to predict relationships between unseen tasks more accurately than classifiers trained on the same representations, as well as handcrafted asymmetric distances (e.g., KL divergence). + +This suggests that low-dimensional box embeddings can effectively capture these task relationships and have the added advantage of being interpretable. We use the approach to visualize relationships among publicly available image classification datasets on popular dataset hosting platform called Hugging Face. + +# 1. Introduction + +The success of deep learning has led to the proliferation of datasets for solving a wide range of computer vision problems. Yet, there are few tools available to enable practitioners to find datasets related to the task at hand, and to solve various meta-tasks related to it. We present TASK2Box, a method to represent tasks using axis-aligned hyperrectangles (or box embeddings). TASK2Box is framed as a learnable mapping from dataset representation to boxes, and can be trained to predict various relationships between novel tasks such as transferability, hierarchy, and overlap. + +Box embeddings [48] extend order embeddings [46] by using volumetric relationships between axis-aligned hyper + +rectangles to encode pairwise relationships. Prior work in natural language processing has utilized box embeddings to represent the WordNet [30] hierarchy and to model conditional distributions. To model relationships between novel datasets, we develop a technique to map from Euclidean representations of datasets into the space of boxes. We explore simple image and label embedding from large vision-language models such as CLIP [36], TASK2VEC [4], and attribute-based vectors [15] as base representations of tasks. + +We test our framework to model asymmetric relationships between nodes in iNaturalist [45] and Caltech-UCSD Birds (CUB) [50] and ImageNet [11] datasets, as well as to predict transferability on the Taskonomy benchmark [52]. Table 1 and 2 show that low-dimensional box embeddings accurately predict novel relationships between datasets seen during training, as well as relationships with novel datasets. Remarkably, TASK2BOX outperforms classifiers trained to directly predict the relationships on the same representations, suggesting that the box embedding provides a strong inductive bias for learning hierarchical relationships. We also outperform simple asymmetric distances proposed in prior work such as Kullback-Leibler (KL) divergence [4]. To model the heterogeneous tasks in the Taskonomy benchmark [52] we map each task to a set of attributes from which a box embedding is learned. Once again, we obtain significantly higher correlation between the true and predicted transferability for both existing and novel datasets compared to standard classifiers (Table 3). Such attribute-based representations can be readily derived from datasheets [15] and modelcards [31]. + +Finally, the low-dimensional box embeddings have the added advantage of being interpretable. Fig. 1 and Fig. 3 show relationships on the iNaturalist+CUB and ImageNet categories, respectively. The 2D box representation allows us to readily visualize the strength and direction of task relationships based on the overlapping volumes, which is not possible using symmetric distances with Euclidean representations (e.g., t-SNE [44]). At the same time, new datasets can be embedded in constant time without needing to retrain or re-optimize. Fig. 5 uses TASK2Box to visualize relationships among 131 publicly available datasets on Hugging Face [1], a popular platform for hosting datasets. Our main contributions are as follows: + +- We introduce a novel method (TASK2BOX) that uses box embeddings to learn asymmetric (e.g., hierarchical, transfer learning) dataset relationships. +- We demonstrate that TASK2Box can predict the relationships of new tasks with a collection of existing tasks. +- We illustrate the interpretability of our model, and the ability to visualize public classification datasets on Hugging Face. + +The code for this project is publicly available at https://github.com/cvl-umass/task2box. + +# 2. Related Work + +Task Representations. Given a dataset $\mathcal{D} = \{(x_i, y_i)\}_{i=1}^n$ , consisting of images $x_i \in \mathcal{X}$ and labels $y_i \in \mathcal{Y}$ , a range of approaches have been proposed for dataset representation. The most straightforward approach involves modeling the distribution of either the images $x$ or the labels $y$ within the dataset independently, using embeddings referred to in prior work as "domain" and "label" embeddings [4]. To capture the joint dependency between images and labels, [4] proposed the use of the Fisher Information Matrix (FIM) derived from a "probe network" trained to minimize a loss function $\ell(\hat{y}, y)$ over the dataset [22, 25, 35]. This approach leverages the similarity of FIMs to predict task transferability and for model selection. However, the utility of the FIM critically depends on the choice of the probe network and a pre-defined similarity may not accurately represent the various relationships between datasets. + +We also investigate the use of vision-language models (VLMs) such as CLIP [36]. This model, trained on a wide range of visual domains, can generalize to tasks involving vision and language data. CLIP features have been effective for image classification [3, 9], semantic segmentation [18, 24, 26], object detection [16, 47, 51], and even closing domain gaps for performance improvement [23, 53]. Both images $x \in \mathcal{X}$ and labels $y \in \mathcal{Y}$ represented as text, can be mapped into a shared space using the vision encoder $(\phi)$ and text encoder $(\psi)$ of CLIP, allowing us to model the dataset as a set of image and label embeddings $\left\{\left(\phi(x_i), \psi(y_i)\right)\right\}_{i=1}^n$ . + +We compare FIMs with representations derived from CLIP as base representations for tasks and learn box embeddings to model a variety of relations among tasks. + +Task Relations in Computer Vision. Understanding the relationships between tasks can lead to efficient solutions to new tasks. Previous work has measured task similarity by using model gradients [14] or based on their learned features [21] for grouping tasks for efficient multi-tasking. Similarly, predicting which pre-trained models will generalize the best on a new dataset could streamline model selection. Taskonomy [52] investigates transfer learning across vision tasks, varying from segmentation to pose estimation, by computing pairwise transfer distances or task affinities. These affinities are calculated by evaluating the extent to which a model trained on a source task generalizes to a target task [13, 42], though this process is computationally expensive. + +Dataset Visualization. Low-dimensional Euclidean embeddings derived from UMAP [28], t-SNE [44], and LargeVis [43] are widely used to visualize relationships between datasets. They have been shown to successfully recover clusters of various data modalities by preserving the + +relationship of each data point with its neighbors [4, 5, 40]. In low-dimension space, relationships with other data points are defined by their Euclidean distances. However, these are commonly used to represent symmetric relations. + +Visualizations using tidy trees [37], circle packing [49], or cone trees [29] organize asymmetric relations as tree-structured hierarchies in low dimension. However, cyclic data relationships cannot be properly represented for these methods (e.g., when a node has two or more parents). + +Asymmetric Distances over Datasets. Kullback-Leibler (KL) divergence between image or label distributions provides a natural way to represent asymmetric distances between datasets. TASK2VEC [4] computes the similarity between two tasks (e.g., cosine distance), and introduces asymmetry by using the complexity of the first task as a reference. The complexity is measured by the similarity of the task embedding to a "trivial embedding" (embedding of a task that is easy or has no examples). + +Order embeddings on images were first proposed in [46] to capture tree-structured relationships. Given an dataset of $P = (u, v)$ drawn from an partially ordered set $(X, \preceq_X)$ , they frame the problem as learning a mapping $f: (X, \preceq_X) \to (Y, \preceq_Y)$ that is order preserving, i.e., $u \preceq_X v \Longleftrightarrow f(u) \preceq_Y f(v)$ . The reserved product order was used for $\preceq_Y$ , i.e., $x \preceq y \Longleftrightarrow x_i > y_i, \forall i$ . Box embeddings [8] generalized this framework by representing points as axis-aligned hyper-rectangles and using volumetric relationships (e.g., intersection over union) to represent asymmetric relations. They used the framework to model conditional distributions and hypernymy relations (e.g., dog "is a" mammal) on the WordNet graph [2, 19, 33, 38, 48]. + +Hyperbolic spaces provide yet another way to model asymmetric relationships. Examples include the Poincaré disk model which uses hyperbolic cosine (cosh) to measure distance between points in a disk. Poincaré embeddings have been similarly used to represent WordNet hierarchies [32] and other relations in general graphs. Hyperbolic representations have also been proposed for representing images for efficient zero-shot learning given a taxonomic structure of the labels [27]. + +To the best of our knowledge, no prior work has explored the use of these spaces for representing entire datasets and their effectiveness in capturing various task relationships. We adopt box embeddings in this work due to the effectiveness over alternatives in previous work [2, 6, 34, 38], ease of visualization, as well as due to open-source libraries for robust learning. However, instead of learning box embeddings directly, we learn mappings from task representations. + +# 3. TASK2Box Framework + +We define the problem as follows: given a collection of datasets $\{\mathcal{D}_1,\mathcal{D}_2,\dots ,\mathcal{D}_m\}$ , and an asymmetric rela + +tionship given the pairwise relationship between datasets $d(\mathcal{D}_i,\mathcal{D}_j)\in [0,1]$ , we aim to encode each dataset into a low-dimension space that preserves the relationships between datasets and is interpretable. + +To achieve this, we propose using box embeddings for encoding each of the datasets. This process involves two main steps: (1) deriving the base representation $e$ of each dataset, and (2) learning a model $f_{\theta}: e \to z$ , where $z \in \mathbb{R}^{2 \times k}$ represents a $k$ -dimensional axis-aligned hyperrectangle (i.e., box), denoted by its lower left and upper right coordinates. These steps are detailed further below. + +# 3.1. Base Task Representations + +Each dataset $\mathcal{D} = \{(x_i, y_i)\}_{i=1}^n$ is defined as a collection of pairs of images $x_i \in \mathcal{X}$ and labels $y_i \in \mathcal{Y}$ . For obtaining a base embedding $e$ for each dataset, we utilize methods such as CLIP [36, 41], TASK2VEC [4], or attribute-based approaches [15]. + +CLIP. Using a pre-trained CLIP model [7, 20, 36, 41], we compute the mean and variance of the individual sample embeddings within each dataset. For each data sample, the image embedding is concatenated with the label embedding, the latter generated from text prompts (e.g., "A photo of [CLS]"). This concatenation models the joint distribution of images and labels. Eq. 1 and 2 detail how the mean and variance embeddings are derived, where $[i,j]$ represents the concatenation of vectors $i$ and $j$ , $\phi$ is the vision encoder, and $\psi$ is the text encoder. The covariance is approximated as diagonal for tractability. + +$$ +\mu_ {C L I P} = \frac {1}{N} \sum_ {i = 1} ^ {N} [ \phi (x _ {i}), \psi (y _ {i}) ] \tag {1} +$$ + +$$ +\sigma_ {C L I P} ^ {2} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(\left[ \phi \left(x _ {i}\right), \psi \left(y _ {i}\right) \right] - \mu_ {C L I P}\right) ^ {2} \tag {2} +$$ + +The base representation is defined as $e := \mu_{CLIP} \in \mathbb{R}^{2048}$ or $e := [\mu_{CLIP}, \sigma_{CLIP}^2] \in \mathbb{R}^{4096}$ . A ViT-H/14 [12] pretrained on LAION-2B [41] was used to extract the embeddings. + +TASK2VEC [4] encodes a dataset using the approximate Fisher Information Matrix (FIM) of a network trained on the given dataset. The FIM represents the importance of parameters in the feature extractor by perturbing the weights $\hat{w}$ of a given probe network with Gaussian noise $\mathcal{N}(0,\Lambda)$ . The precision matrix $\Lambda$ is estimated to be close to an isotropic prior $\mathcal{N}(\hat{w},\lambda^2 I)$ while having a good expected error. Eq. 3 is minimized to find $\Lambda$ where $H$ is the cross entropy loss, $\hat{w}$ are the weights of the network, $\beta$ is the magnitude of the prior, $x\in \mathcal{X}$ , and $y\in \mathcal{Y}$ . + +$$ +\mathcal {L} (\hat {w}; \Lambda) = \mathbb {E} _ {w \sim \mathcal {N} (\hat {w}, \Lambda)} [ H _ {p _ {w}, \tilde {p}} p (y | x) ] + +$$ + +$$ +\beta K L \left(\mathcal {N} (0, \Lambda) \| \mathcal {N} (0, \lambda^ {2} I)\right) \tag {3} +$$ + +The matrix $\Lambda$ provides an estimate of the FIM and is approximated as a diagonal matrix. The diagonal components are used as a base representation of a dataset ( $e := \mathrm{FIM} \in \mathbb{R}^{17024}$ ). ResNet-34 [17] pretrained on ImageNet [11] is used as the probe network for all datasets. + +Attribute-based. A task can be characterized by a set of $t$ binary attributes [15] represented as a vector of dimension $t$ . Some of the attributes explored for representing tasks are: (1) Is the task generative? (2) Is the task output in 2D? (3) Does the task involve camera pose estimation? Taking these 3 characteristics, for example, we can represent a 2D segmentation task as the vector $e = [0,1,0]$ as a discriminative 2D task that does not need camera poses. + +Tasks in Taskonomy [52] involving multiple modalities benefit from this attribute-based representation due to its model independence. This approach also enables generalization to unseen tasks by identifying the presence or absence of various characteristics. Each vision task in Taskonomy is represented with 15 attributes, resulting in vectors $e \in \mathbb{R}^{15}$ . The full list of attributes is in Appendix ?? + +# 3.2. Learning Box Embeddings + +From a base task representation $e$ , we learn the parameters $\theta$ of a model $f_{\theta}: e \to z$ that preserves the asymmetric similarity function $d(\mathcal{D}_i, \mathcal{D}_j)$ between any two datasets $\mathcal{D}_i$ and $\mathcal{D}_j$ . In Eq. 4, the learning objective is shown where $\mathcal{L}_E$ is a loss function (mean squared error), $\mathcal{L}_D$ is a distance function, $\mathcal{L}_R$ is a regularization term, and $\lambda$ is a hyperparameter. The embeddings $z_i, z_j \in \mathbb{R}^{2 \times k}$ represent the coordinates of the lower left and the upper right corners of the respective $k$ dimensional boxes, with $f_{\theta}(e_i) = z_i$ and $f_{\theta}(e_j) = z_j$ . + +$$ +\begin{array}{l} \hat {\theta} = \underset {\theta} {\operatorname {a r g m i n}} \sum_ {i, j} \left(\mathcal {L} _ {E} \left(d \left(\mathcal {D} _ {i}, \mathcal {D} _ {j}\right), d _ {b o x} \left(z _ {i}, z _ {j}\right)\right) \right. \\ + \lambda \mathcal {L} _ {D} \left(z _ {i}, z _ {j}\right) + \mathcal {L} _ {R}) \tag {4} \\ \end{array} +$$ + +The asymmetric relationship between box embeddings, denoted as $d_{box}(z_i,z_j)$ , is computed in Eq. 5 by calculating the volume of the intersection between $z_i$ and $z_j$ , normalized by the volume of $z_i$ . For $z_i$ to be fully contained inside $z_j$ ( $z_i \subset z_j$ ), it is required that $d_{box}(z_i,z_j) = 1$ . Conversely, for $z_j$ to only partially contain $z_i$ , $d_{box}(z_j,z_i)$ must fall within the range (0,1). Fig. 2 illustrates this through a 2-dimensional example, where $z_1$ represents the box embedding for Canidae, and $z_2$ for Mammalia. + +![](images/dd107d1d5b3fc95821a19aca38ad74268e975911a6a8c983031b81716cb2100a.jpg) +Figure 2. TASK2Box Embeddings in 2D for Mammalia, Canidae, and Amphibia Datasets from iNaturalist. Each embedding represents the coordinates of the lower left and upper right corners of each box/rectangle. Since Canidae $(z_{1})$ is a proper subset of Mammalia $(z_{2})$ : $d_{box}(z_{1}, z_{2}) = 1$ and $d_{box}(z_{2}, z_{1}) = 0.1$ . + +$$ +d _ {b o x} \left(z _ {i}, z _ {j}\right) = \frac {\operatorname {v o l} \left(z _ {i} \cap z _ {j}\right)}{\operatorname {v o l} \left(z _ {i}\right)} \tag {5} +$$ + +$\mathcal{L}_D$ is applied to datasets where $d(\mathcal{D}_i, \mathcal{D}_j) > 0$ such that the Euclidean distance between the center coordinates of $z_i$ and $z_j$ is minimized when starting from a non-overlapping state. This allows non-overlapping embeddings to move closer to each other, complementing $\mathcal{L}_E$ to learn relationships. $\mathcal{L}_R$ encourages solutions with regular-shaped boxes for better interpretability. The formulation of $\mathcal{L}_R$ , given in Eq. 6, applies to a $k$ -dimensional box embedding $z_i$ , where $s_a$ represents the size of the $a$ -th dimension (for example, width), and $\alpha, \beta$ are hyperparameters. To prevent the trivial solution of minimizing box volume to zero, the inverse of the box volume is included. It's crucial to normalize the terms with respect to the embedding dimension, as the first and second terms scale quadratically and exponentially with dimension increase, respectively. + +$$ +\mathcal {L} _ {R} = \left(\frac {\alpha}{k ^ {2}} \sum_ {a = 1} ^ {k} \sum_ {b = a + 1} ^ {k} \left| s _ {a} - s _ {b} \right|\right) + \beta (\operatorname {v o l} (z _ {i})) ^ {- 1 / k} \tag {6} +$$ + +The architecture of $f$ consists of three fully-connected layers followed by two linear heads: one predicts a $k$ -dimensional lower-left coordinate, and the other predicts the $k$ -dimensional sizes of each box dimension. + +# 4. Experiments + +We consider the following goals to evaluate the capability of Task2Box to represent datasets: + +1. Given a collection of existing datasets $\mathcal{D}_E$ and a subset of pairwise relationships $\mathcal{R}$ : can the model generalize on unseen relationships $\mathcal{R}'$ within $\mathcal{D}_E$ where $\mathcal{R}' \cap \mathcal{R} = \emptyset$ ? +2. Given a collection of novel datasets $\mathcal{D}_N$ not seen during training: can the model accurately identify the relationships with the existing datasets $\mathcal{D}_E$ ? + +
iNaturalist + CUB
MethodExisting DatasetsNovel Datasets
μCLIP[μ,σ2]CLIPFIMμCLIP[μ,σ2]CLIPFIM
TASK2Box (2D)69.23%67.84%39.61%50.07%39.66%10.06%
TASK2Box (3D)79.66%79.35%57.63%70.04%64.53%20.65%
TASK2Box (5D)84.67%82.41%79.72%73.79%72.11%34.88%
MLP Classifier45.25%61.45%26.34%39.06%44.54%19.90%
Linear Classifier4.40%3.11%7.06%4.77%5.87%15.92%
KL Divergence-6.58%7.94%-5.90%0.00%
Asymmetric Cosine9.29%11.54%2.83%1.47%1.47%1.47%
Asymmetric Euclidean1.71%1.71%8.53%1.47%1.47%1.91%
Random2.06%1.49%
+ +Table 1. Average F1 Score for Predicting Hierarchical Relationships on iNaturalist + CUB Dataset. For all feature types, TASK2Box outperforms other methods by more than $20\%$ on existing datasets and more than $10\%$ on novel datasets. The best-performing model is in bold, and the second-best is underlined. Results on Novel Datasets demonstrate that our model can generalize beyond the seen tasks it has previously seen. The dimensions such as 2D, 3D, and 5D refer to different box dimensionalities. Results using KL divergence for $\mu_{CLIP}$ are not shown since it is not a distribution. + +The goals are evaluated through two experimental setups, demonstrating our model's ability to predict various types of relationships. The configurations, and their corresponding datasets, relationships, and metrics, are detailed below. Baseline methods for performance comparison are also reviewed. Implementation details are in the Appendix. + +# 4.1. Experimental Setup + +# 4.1.1 Hierarchical Task Relationships + +We use a combination of iNaturalist [45] and Caltech-UCSD Birds (CUB) [50], and instrument-related classes in ImageNet [11] to evaluate the ability of TASK2Box to represent hierarchical relationships. The first two are composed of images of various species, and the third is composed of images of various objects. The classes naturally follow a hierarchical form based on biological classification (taxonomy of iNaturalist+CUB species), and on semantic relations (hypernymy of objects in ImageNet). + +Datasets. For iNaturalist+CUB, the datasets are defined as the classes, the orders, and the families in the taxonomy that contain a significant number of samples per dataset as in [4]. There are 47 classes, 202 orders, and 589 families for a total of 838 datasets. For ImageNet, the instrument-related objects were processed using WordNet [30] to obtain hierarchical information between classes. This resulted in 131 datasets for training and evaluation. + +Dataset Relationships. For any two datasets, their relationship is captured as $d(\mathcal{D}_i,\mathcal{D}_j)\in \{0,1\}$ , where $d(\mathcal{D}_i,\mathcal{D}_j) = 1$ if and only if $\mathcal{D}_i\subset \mathcal{D}_j$ , and $d(\mathcal{D}_i,\mathcal{D}_j) = 0$ otherwise. Fig. 2 provides an example: Canidae $(\mathcal{D}_1)$ is a family within the class Mammalia $(\mathcal{D}_2)$ ; thus, $d(\mathcal{D}_1,\mathcal{D}_2) = 1$ and $d(\mathcal{D}_2,\mathcal{D}_1) = 0$ . Meanwhile, Amphibia $(\mathcal{D}_3)$ is unrelated to either dataset, resulting in $d(\mathcal{D}_1,\mathcal{D}_3) = 0$ + +Evaluation Metrics. We evaluate the ability of the model to classify the presence of containment relationships between datasets using the F1 score due to the imbalance between positive and negative relationships. To obtain F1 score, precision and recall are first calculated. Precision is computed as the ratio of true positive pairs predicted to the total number of positive predictions. Recall is the ratio of true positive pairs predicted to the total number of true positive pairs. F1 score is then reported as the harmonic mean between the precision and recall. These calculations are done on both (1) the unseen relationships $\mathcal{R}'$ within existing datasets $\mathcal{D}_E$ and (2) the relationships between novel datasets $\mathcal{D}_N$ and existing datasets $\mathcal{D}_E$ . For the latter, we evaluate the relationships in both directions, i.e., $\left\{\left(\mathcal{D}_e,\mathcal{D}_n\right)\forall \mathcal{D}_e\in \mathcal{D}_E\right\} \cup \left\{\left(\mathcal{D}_n,\mathcal{D}_e\right)\forall \mathcal{D}_e\in \mathcal{D}_E\right\}$ . + +# 4.1.2 Transfer Learning Between Datasets + +The Taskonomy [52] benchmark is used to evaluate the ability of TASK2BOX to predict task affinities. + +Datasets. Taskonomy defines a set of 25 visual tasks with corresponding pairwise task affinities. These tasks range from object detection, semantic segmentation, pose estimation, and more. We treat each visual task as a dataset. The tasks are described in Appendix ?? + +Dataset Relationships. In contrast to the containment relationship defined in § 4.1.1, Taskonomy quantifies relationships with task affinity measured by the performance gain achieved by transfer learning from a source dataset $\mathcal{D}_j$ to a target dataset $\mathcal{D}_i$ . The values are computed using Analytic Hierarchy Process [39, 52]. Dataset relationships are computed and normalized based on an ordinal approach and determined by the percentage of images that transfer well to a target task given a set of source tasks as detailed in [52]. + +
ImageNet
MethodExisting DatasetsNovel Datasets
μCLIP[μ,σ2]CLIPFIMμCLIP[μ,σ2]CLIPFIM
TASK2Box (2D)62.72%63.33%31.32%47.76%48.35%9.46%
TASK2Box (3D)83.12%82.79%58.16%73.06%66.48%24.70%
TASK2Box (5D)90.58%88.48%64.91%76.95%78.84%37.39%
MLP Classifier54.20%60.43%44.85%62.24%64.56%41.22%
Linear Classifier9.84%8.33%25.46%11.89%11.87%22.98%
KL-5.53%9.90%-10.41%0.00%
Asymmetric Cosine4.28%4.28%6.92%4.52%4.52%0.00%
Asymmetric Euclidean3.73%3.73%7.16%4.52%4.52%4.73%
Random3.64%5.02%
+ +Table 2. Average F1 Score for Predicting Hierarchical Relationships on ImageNet Instruments. TASK2BOX is shown to generalize well on both predicting relationships between existing datasets (Existing Datasets), and predicting box embeddings of new datasets and their relationships with existing datasets (Novel Datasets). + +For a pair of datasets $(\mathcal{D}_i,\mathcal{D}_j)$ , the task affinity is defined as $d(\mathcal{D}_i,\mathcal{D}_j)\in [0,1]$ . The higher the task affinity to a target dataset $\mathcal{D}_i$ from a source dataset $\mathcal{D}_j$ , then $d(\mathcal{D}_i,\mathcal{D}_j)$ gets closer to 1, where a value of 1 would show $\mathcal{D}_i\subset \mathcal{D}_j$ . + +Evaluation Metrics. We evaluate the prediction of the model based on the Spearman correlation between the ground truth task affinity values, and the predicted values based on the box distances in Eq. 5. Similar to § 4.1.1, we evaluate on both (1) unseen relationships $\mathcal{R}'$ for existing datasets, and (2) on relationships between unseen novel datasets $\mathcal{D}_N$ and existing datasets $\mathcal{D}_E$ . + +# 4.2. Baseline Methods + +We compare the performance of TASK2Box with alternative models and simple asymmetric distances proposed in prior work. For hierarchical relationships, given two datasets $\mathcal{D}_i,\mathcal{D}_j$ , the models predict $\hat{d} (\mathcal{D}_i,\mathcal{D}_j)\in \{0,1\}$ . For task affinity, the models predict $\hat{d} (\mathcal{D}_i,\mathcal{D}_j)\in [0,1]$ . The structure for the various methods are discussed below. + +Linear Model. A linear model is trained to predict the relationship value between two datasets. The input is the concatenation of the base representation $e$ for the two datasets. + +MLP Model. A 4-layer MLP is used instead for prediction on the same inputs as the linear model. + +KL Divergence. Each dataset is treated as a multivariate Gaussian using the mean and variance of the image and label features (CLIP) or directly as the FIM. The optimal threshold $t$ is selected for the minimum distance between two dataset distributions $\mathrm{KL}(\mathcal{D}_i||\mathcal{D}_j) < t$ for a prediction of $\hat{d} (\mathcal{D}_i,\mathcal{D}_j) = 1$ , and $\hat{d} (\mathcal{D}_i,\mathcal{D}_j) = 0$ otherwise. The F1 Score (hierarchical) or the correlation (task affinity) is the objective for selecting $t$ on the train set. + +Asymmetric Cosine Similarity. The cosine similarity $d_{\text{cos}}$ is a symmetric measure between two embeddings $e_i, e_j$ . An asymmetric variant was proposed in [4] and shown in Eq. 7 + +![](images/09083c2f5ac4f3c26485b217f6bc1070383f1c2974afcfe2dee3abb75e6af820.jpg) +Figure 3. Visualization of Instrument-related Datasets in ImageNet. Datasets that belong to the same superset are shaded in the same color. TASK2Box learns the hierarchy of various groups, and clusters similar datasets closer. + +by considering the similarity of $e_i$ and $e_j$ relative to the complexity of $e_i$ . The complexity of is measured as the distance to the trivial embedding $e_o$ , and $\alpha$ is a hyperparameter. + +$$ +d _ {a s y m} \left(e _ {i}, e _ {j}\right) = d _ {c o s} \left(e _ {i}, e _ {j}\right) - \alpha d _ {c o s} \left(e _ {i}, e _ {o}\right) \tag {7} +$$ + +An optimal threshold $t$ is found on the train set where $d_{asym}(e_i, e_j) < t$ results to a prediction $\hat{d}(\mathcal{D}_i, \mathcal{D}_j) = 1$ . The same threshold $t$ is used for test set evaluation. + +Asymmetric Euclidean Similarity. The asymmetric similarity is computed as in Eq. 7 but uses the Euclidean distance instead of cosine. + +Random. The probability of containment (for hierarchical) or the value of the task affinity is uniformly random. + +# 5. Results and Discussion + +# 5.1. Hierarchical Relationships + +Tables 1 and 2 present the average F1 score for predicting hierarchical relationships on iNaturalist+CUB and Ima + +geNet datasets. Various methods of extracting dataset embeddings were evaluated. TASK2BOX outperforms baseline methods, indicating that TASK2BOX can both generalize across unseen relationships (Existing Datasets), and to accurately represent relationships with existing datasets for unseen datasets (Novel Datasets). CLIP features also generalize better than FIM features, perhaps because the CLIP multi-modal embedding of images and labels was trained across a broad set of domains[36]. However, while CLIP is confined to image and text modalities, FIM can accommodate any modality by adapting the probe network's last layer for different prediction tasks. + +Beyond predicting relationships, our method allows the visualization of datasets. Fig. 1 illustrates this for a subset of datasets in iNaturalist and CUB using TASK2Box (2D), and Fig. 3 for ImageNet. The hierarchical organization of the datasets is apparent in the TASK2Box representation – datasets that contain others appear as larger boxes, while more specialized datasets are depicted as smaller boxes netted with borader, more general dataset boxes. However, while 2D visualization offers insights, the flat surface may restrict the representation of complex relationships. Representations in higher dimensions is explored in § 5.4. + +# 5.2. Task Affinity + +Table 3 presents the results of task representations in Taskonomy. Our method shows that, even with attribute-based embeddings, it can learn box embeddings through relationship supervision between datasets. TASK2Box is shown to correlate highly with the ground truth task affinities compared to other methods. Given that only 25 tasks are available in Taskonomy, with 3 held out of training as novel datasets, there is a higher uncertainty in predicting unseen datasets. We expect that performance on novel datasets will improve with more datasets available during training. + +Fig. 4 displays the learned representation with TASK2Box (2D). Each subfigure in (a)-(c) represents a subset of tasks identified as having strong transfer relationships. TASK2Box not only identifies but also visually represents related tasks suitable for fine-tuning. The small highlighted boxes indicate target tasks, while the larger enclosing boxes represent source tasks. Although other tasks may not be proper subsets, the box distance in Eq. 5 provides an estimate for task affinity. A small box distance between two datasets, $d_{box}(z_1, z_2)$ , suggests lower transfer performance from a source task $z_2$ to a target task $z_1$ . + +# 5.3. Visualizing Public Datasets + +Apart from predicting both hierarchical relationships and quantified task affinities, our method can also be used to visualize public datasets that lack available ground truth relationships. Using a set of vision tasks from Hugging Face [1], $\mu_{CLIP} + \text{TASK2Box}$ (2D) were utilized to pre + +![](images/fe7890c962f7a3223caf4add0c39e6794c38b0aadb9eb1dd83a160fbaef0df95.jpg) +Figure 4. Visualization of Tasks in Taskonomy showing Source Tasks that Transfer Well to Target Tasks (shaded). (a) Jigsaw and Triplet Fixated Camera Pose estimation are source tasks that transfer well to Depth estimation. (b) and (c) show different source tasks (larger boxes) that transfer well to the shaded boxes of Denoising Autoencoder and Surface Normal, respectively. + +![](images/ef7aace0e3560b707f6b06d87f3df34013a0b1ab23bcf8cfdab46fef5df1f363.jpg) + +![](images/10fc35761ec0efa7ef3b12f441f4ec9681efb41db44c4140c3e72356248047c3.jpg) + +
MethodExisting DatasetsNovel Datasets
Spearman's ρSpearman's ρ
TASK2Box (2D)0.85 ± 0.060.12 ± 0.21
TASK2Box (3D)0.93 ± 0.020.48 ± 0.24
TASK2Box (5D)0.94 ± 0.030.39 ± 0.22
MLP0.88 ± 0.060.31 ± 0.18
Linear0.75 ± 0.110.40 ± 0.24
Random0.05 ± 0.140.15 ± 0.07
+ +Table 3. Spearman Correlation and Standard Deviation between Predicted and Ground Truth Task Affinities on Taskonomy. Our method shows higher correlation with the task affinities compared to the baseline. Attribute-based embeddings were used. + +dict dataset embeddings. A constraint on the box sizes was added to reflect information about the dataset sizes. Fig. 5 displays the results of TASK2Box on 131 datasets from Hugging Face, where similar datasets, such as sentiments and documents, are shown to overlap. This approach allows for the analysis and visualization of dataset variations and similarities even with only samples of images and labels. Simultaneously, the sizes of various datasets can be visualized as box sizes – with larger datasets containing more samples depicted as larger boxes. This tool enables computer vision practitioners to see how their datasets compare with other existing datasets. Beyond visualizing the variation of available datasets (based on the number of clusters), it can expedite the process of finding suitable data sources by examining embedding overlaps. + +# 5.4. Analysis of TASK2BOX + +We discuss properties of TASK2Box below. Additional insights are also included in Appendix ?? + +Using a Box Prior Improves Performance. While TASK2Box was trained with a 3-layer MLP, it achieved significantly better performance than the MLP baseline without a box prior. Tables 1, 2, and 3 demonstrate our method outperforming the baselines. Representing each dataset as an entity with shape and volume (such as a box), instead of solely learning relationships from embeddings, proves effective for generalization on unseen relationships. This + +![](images/3d80c9536d589426431a0c301c60fab3dc0aaa4351d9b6b744451ba50d908dcb.jpg) +Figure 5. Visualizing Image Classification Datasets in Hugging Face. The sample data points annotated on the highlighted datasets show that common tasks overlap with each other (e.g., sentiment classification and document classification datasets). Although labels could slightly differ between datasets, TASK2BOX can infer the level of similarity and represent it as the amount of overlap. The embedding size (box area) also shows the number of available data. + +![](images/e26db90d6b8abfefdd07cdbae2b735cee39eca4ca7ced8983502cb5aca2f7875.jpg) +Figure 6. Effect of TASK2Box Embedding Dimension on the Accuracy of Predicting Task Relations. As the dimension increases, the performance generally increases. + +improvement could be attributed to the explicit modeling of relationships through physical shapes, which enforces consistency: if $z_{i} \subset z_{j}$ and $z_{j}$ has no overlap with $z_{k}$ , then having a physical representation ensures that $z_{i} \cap z_{k} = \emptyset$ . + +TASK2BOX can Represent Relations in Varying Dimensions. Fig. 6 illustrates the performance of TASK2BOX as the box dimension increases. Representing relationships in two dimensions results in easily interpretable embeddings; however, modeling complex task relationships could benefit from expansion to higher dimensions. Projecting to higher dimensions may yield even better performance, as relationships between datasets become more accurately represented, given the model's increased representation capacity. Nonetheless, as the dimensionality further increases, learning and visualizing the embedding space also becomes more challenging. + +Using Boxes to Represent Tasks Enables Effective Calculations on Embeddings. Boxes offer the advantage of being closed under intersection, meaning the intersection of two boxes results in another box, a property not shared by + +circles or ellipses. Among region-based embeddings that utilize geometric objects (e.g., cones, disks, and boxes) to represent entities, operations involving boxes are the most straightforward to calculate [10, 38]. + +# 6. Conclusion + +We present a novel method that learns low-dimensional, interpretable embeddings of various task relationships. With box representations, we demonstrate that asymmetric relationships, such as task affinities and hierarchies, can be accurately modeled using various base representations. While CLIP embeddings have been shown to outperform FIM when integrated with our model, future work could investigate how CLIP might be adapted to modalities beyond text and images. Attribute-based features offer a viable alternative, and can be extracted from from datasheets using natural language processing techniques. + +The distinct properties of TASK2BOX were analyzed, revealing its ability to perform effectively across varying dimensions, and to model and visualize overlaps among public classification datasets on Hugging Face. This could enable computer vision practitioners to assess dataset utility for a task in hand. Although our model successfully represents task relationships, it does not incorporate information about optimal training procedures and model architecture. Future work could explore the inclusion of additional information for a more detailed understanding of the task space. + +Acknowledgements. This work was supported by awards from the National Science Foundation (2329927 and 1749833) and the NASA AIST program. The experiments were performed on the University of Massachusetts GPU cluster funded by the Mass. Technology Collaborative. + +# References + +[1] Hugging Face datasets. https://huggingface.co/datasets?taskCategories $\equiv$ taskCategories:image-classification.2,7 +[2] Ralph Abboud, Ismail Ceylan, Thomas Lukasiewicz, and Tommaso Salvatori. Boxe: A box embedding model for knowledge base completion. Advances in Neural Information Processing Systems, 33:9649-9661, 2020. 3 +[3] Rabab Abdelfattah, Qing Guo, Xiaoguang Li, Xiaofeng Wang, and Song Wang. Cdul: Clip-driven unsupervised learning for multi-label image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1348-1357, 2023. 2 +[4] Alessandro Achille, Michael Lam, Rahul Tewari, Avinash Ravichandran, Subhransu Maji, Charless C Fowlkes, Stefano Soatto, and Pietro Perona. Task2vec: Task embedding for meta-learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6430-6439, 2019. 1, 2, 3, 5, 6 +[5] Sanjeev Arora, Wei Hu, and Pravesh K. Kothari. An analysis of the t-sne algorithm for data visualization. In Proceedings of the 31st Conference On Learning Theory, pages 1455-1462. PMLR, 2018. 3 +[6] Michael Boratko, Dongxu Zhang, Nicholas Monath, Luke Vilnis, Kenneth L Clarkson, and Andrew McCallum. Capacity and bias of learned geometric embeddings for directed graphs. Advances in Neural Information Processing Systems, 34:16423-16436, 2021. 3 +[7] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 3 +[8] Tejas Chheda, Purujit Goyal, Trang Tran, Dhruvesh Patel, Michael Boratko, Shib Sankar Dasgupta, and Andrew McCallum. Box embeddings: An open-source library for representation learning using geometric structures. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 203-211, 2021. 3 +[9] Marcos V Conde and Kerem Turgutlu. Clip-art: Contrastive pre-training for fine-grained art classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3956-3960, 2021. 2 +[10] Shib Dasgupta, Michael Boratko, Siddhartha Mishra, Shriya Atmakuri, Dhruvesh Patel, Xiang Li, and Andrew McCallum. Word2box: Capturing set-theoretic semantics of words using box embeddings. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 2022. 8 +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 4, 5 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations (ICLR), 2021. 3 +[13] Kshitij Dwivedi and Gemma Roig. Representation similarity analysis for efficient task taxonomy & transfer learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12387-12396, 2019. 2 +[14] Chris Fifty, Ehsan Amid, Zhe Zhao, Tianhe Yu, Rohan Anil, and Chelsea Finn. Efficiently identifying task groupings for multi-task learning. Advances in Neural Information Processing Systems, 34:27503-27516, 2021. 2 +[15] Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé Iii, and Kate Crawford. Datasheets for datasets. Communications of the ACM, 64(12):86-92, 2021. 1, 2, 3, 4 +[16] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2021. 2 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[18] Wenbin He, Suphanut Jamonnak, Liang Gou, and Liu Ren. Clip-s4: Language-guided self-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11207-11216, 2023. 2 +[19] EunJeong Hwang, Jay-Yoon Lee, Tianyi Yang, Dhruvesh Patel, Dongxu Zhang, and Andrew McCallum. Event-event relation extraction using probabilistic box embedding. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 235–244, 2022. 3 +[20] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 3 +[21] Zhuoliang Kang, Kristen Grauman, and Fei Sha. Learning with whom to share in multi-task feature learning. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pages 521-528, 2011. 2 +[22] Ryo Karakida, Shotaro Akaho, and Shun-ichi Amari. Universal statistics of fisher information in deep neural networks: Mean field approach. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 1032–1041. PMLR, 2019. 2 +[23] Zhengfeng Lai, Noranart Vesdapunt, Ning Zhou, Jun Wu, Cong Phuoc Huynh, Xuelu Li, Kah Kuen Fu, and Chen-Nee Chuah. Padclip: Pseudo-labeling with adaptive debiasing in clip for unsupervised domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16155-16165, 2023. 2 + +[24] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7061-7070, 2023. 2 +[25] Zhibin Liao, Tom Drummond, Ian Reid, and Gustavo Carneiro. Approximate fisher information matrix to characterize the training of deep neural networks. IEEE transactions on pattern analysis and machine intelligence, 42(1): 15-26, 2018. 2 +[26] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15305-15314, 2023. 2 +[27] Shaoteng Liu, Jingjing Chen, Liangming Pan, Chong-Wah Ngo, Tat-Seng Chua, and Yu-Gang Jiang. Hyperbolic visual embedding learning for zero-shot recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[28] Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. Umap: Uniform manifold approximation and projection. Journal of Open Source Software, 3(29):861, 2018. 2 +[29] Guy Melancon and Ivan Herman. Circular drawings of rooted trees. CWI (Centre for Mathematics and Computer Science), 1998. 3 +[30] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 2, 5 +[31] Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. Model cards for model reporting. In Proceedings of the conference on fairness, accountability, and transparency, pages 220-229, 2019. 2 +[32] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. Advances in neural information processing systems, 30, 2017. 3 +[33] Dhruvesh Patel, Shib Sankar Dasgupta, Michael Boratko, Xiang Li, Luke Vilnis, and Andrew McCallum. Representing joint hierarchies with box embeddings. In Automated Knowledge Base Construction, 2020. 3 +[34] Dhruvesh Patel, Pavitra Dangati, Jay-Yoon Lee, Michael Boratko, and Andrew McCallum. Modeling label space interactions in multi-label classification using box embeddings. In International Conference on Learning Representations, 2022. 3 +[35] Jeffrey Pennington and Pratik Worah. The spectrum of the fisher information matrix of a single-hidden-layer neural network. Advances in neural information processing systems, 31, 2018. 2 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi + +sion. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 7 +[37] Edward M. Reingold and John S. Tilford. Tidier drawings of trees. IEEE Transactions on software Engineering, (2): 223-228, 1981. 3 +[38] Hongyu Ren, Weihua Hu, and Jure Leskovec. Query2box: Reasoning over knowledge graphs in vector space using box embeddings. In International Conference on Learning Representations, 2019. 3, 8 +[39] Roseanna W Saaty. The analytic hierarchy process—what it is and how it is used. Mathematical modelling, 9(3-5):161–176, 1987. 5 +[40] Saquib Sarfraz, Marios Koulakis, Constantin Seibold, and Rainer Stiefelhagen. Hierarchical nearest neighbor graph embedding for efficient dimensionality reduction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 336-345, 2022. 3 +[41] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3 +[42] Astuti Sharma, Tarun Kalluri, and Manmohan Chandraker. Instance level affinity-based transfer for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5361-5371, 2021. 2 +[43] Jian Tang, Jingzhou Liu, Ming Zhang, and Qiaozhu Mei. Visualizing large-scale and high-dimensional data. In Proceedings of the 25th international conference on world wide web, pages 287-297, 2016. 2 +[44] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 1, 2 +[45] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 2, 5 +[46] Ivan Vendrov, Ryan Kiros, Sanja Fidler, and Raquel Urtasun. Order-embeddings of images and language. International Conference on Learning Representations (ICLR), 2015. 1, 3 +[47] Vedit Vedit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3219-3229, 2023. 2 +[48] Luke Vilnis, Xiang Li, Shikhar Murty, and Andrew McCallum. Probabilistic embedding of knowledge graphs with box lattice measures. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 263-272, 2018. 1, 3 +[49] Weixin Wang, Hui Wang, Guozhong Dai, and Hongan Wang. Visualization of large hierarchical data by circle packing. In Proceedings of the SIGCHI conference on Human Factors in computing systems, pages 517-520, 2006. 3 + +[50] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. Technical Report CNS-TR-201, Caltech, 2010. 2, 5 +[51] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7031-7040, 2023. 2 +[52] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3712-3722, 2018. 2, 4, 5 +[53] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Mind the gap: Domain gap control for single shot domain adaptation for generative adversarial networks. In International Conference on Learning Representations, 2021. 2 \ No newline at end of file diff --git a/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/images.zip b/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..33e367f7b03a15f6ee95d7f0e2275ca64a3b894e --- /dev/null +++ b/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1b076396e172f5523538e1237abbfca8b2e7d7b2d378c7b0ebba8b2f321c01b +size 411133 diff --git a/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/layout.json b/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e31cb47257a9dfa630687e386e99d0925c4babbb --- /dev/null +++ b/2024/Task2Box_ Box Embeddings for Modeling Asymmetric Task Relationships/layout.json @@ -0,0 +1,10116 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 103, + 532, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 103, + 532, + 122 + ], + "spans": [ + { + "bbox": [ + 63, + 103, + 532, + 122 + ], + "type": "text", + "content": "Task2Box: Box Embeddings for Modeling Asymmetric Task Relationships" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 168, + 143, + 425, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 143, + 425, + 171 + ], + "spans": [ + { + "bbox": [ + 168, + 143, + 425, + 171 + ], + "type": "text", + "content": "Rangel Daroya Aaron Sun Subhransu Maji University of Massachusetts, Amherst" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 198, + 174, + 392, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 174, + 392, + 185 + ], + "spans": [ + { + "bbox": [ + 198, + 174, + 392, + 185 + ], + "type": "text", + "content": "{rdaroya, aaronsun, smaji}@umass.edu" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 78, + 196, + 246, + 361 + ], + "blocks": [ + { + "bbox": [ + 78, + 196, + 246, + 361 + ], + "lines": [ + { + "bbox": [ + 78, + 196, + 246, + 361 + ], + "spans": [ + { + "bbox": [ + 78, + 196, + 246, + 361 + ], + "type": "image", + "image_path": "2db75518f57383f685a21cbd2d56d03d8bd8ebc92acb04ffc0d94457090b098f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 369, + 547, + 437 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 547, + 437 + ], + "type": "text", + "content": "Figure 1. Box Embeddings of 150 Datasets of iNaturalist + CUB and Corresponding Learned Hierarchy for Class Arachnida. Each taxonomic category is treated as a separate dataset for which TASK2BOX embeddings are learned. (1) Shows the learned box embeddings where datasets from the same group (taxonomic class) have the same color. Datasets naturally cluster to their ground truth groups. (2) Shows the hierarchy learned through TASK2BOX for a specific class. The hierarchy matches the ground truth relationships based on biological classification. Orders that belong to class Arachnida are learned as boxes " + }, + { + "bbox": [ + 46, + 369, + 547, + 437 + ], + "type": "inline_equation", + "content": "(\\widehat{\\mathbf{A}}, \\widehat{\\mathbf{B}}, \\widehat{\\mathbf{C}})" + }, + { + "bbox": [ + 46, + 369, + 547, + 437 + ], + "type": "text", + "content": " contained by the larger box for Arachnida; families under each of the orders are learned as smaller boxes contained by the corresponding orders they belong to." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 247, + 198, + 351, + 274 + ], + "blocks": [ + { + "bbox": [ + 247, + 198, + 351, + 274 + ], + "lines": [ + { + "bbox": [ + 247, + 198, + 351, + 274 + ], + "spans": [ + { + "bbox": [ + 247, + 198, + 351, + 274 + ], + "type": "image", + "image_path": "4b29e3406ad95f3b3e8cbaf0d6d2fc92366ff9b8429bd7445fdd735dfcd529df.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 353, + 197, + 516, + 274 + ], + "blocks": [ + { + "bbox": [ + 353, + 197, + 516, + 274 + ], + "lines": [ + { + "bbox": [ + 353, + 197, + 516, + 274 + ], + "spans": [ + { + "bbox": [ + 353, + 197, + 516, + 274 + ], + "type": "image", + "image_path": "2c91020a3442cd53accdbd1012ce9356cc38ff8cfa65532288072065c36693b9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 268, + 285, + 350, + 348 + ], + "blocks": [ + { + "bbox": [ + 268, + 285, + 350, + 348 + ], + "lines": [ + { + "bbox": [ + 268, + 285, + 350, + 348 + ], + "spans": [ + { + "bbox": [ + 268, + 285, + 350, + 348 + ], + "type": "image", + "image_path": "3011d3bdea7341235c330d778344392cb670b0aa0883755756997e646665b8ff.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 369, + 287, + 439, + 342 + ], + "blocks": [ + { + "bbox": [ + 369, + 287, + 439, + 342 + ], + "lines": [ + { + "bbox": [ + 369, + 287, + 439, + 342 + ], + "spans": [ + { + "bbox": [ + 369, + 287, + 439, + 342 + ], + "type": "image", + "image_path": "62f14903af353c37192084c19c18874a14335fc2c1f5aabc9766e2d2f3545967.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 451, + 289, + 507, + 338 + ], + "blocks": [ + { + "bbox": [ + 451, + 289, + 507, + 338 + ], + "lines": [ + { + "bbox": [ + 451, + 289, + 507, + 338 + ], + "spans": [ + { + "bbox": [ + 451, + 289, + 507, + 338 + ], + "type": "image", + "image_path": "d0c937ac1294d9baf50a5482d3302d0ab59162daa459f4460ace8ba0db229075.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 446, + 192, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 446, + 192, + 459 + ], + "spans": [ + { + "bbox": [ + 143, + 446, + 192, + 459 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": "Modeling and visualizing relationships between tasks or datasets is an important step towards solving various meta-tasks such as dataset discovery, multi-tasking, and transfer learning. However, many relationships, such as containment and transferability, are naturally asymmetric and current approaches for representation and visualization (e.g., t-SNE [44]) do not readily support this. We propose TASK2BOX, an approach to represent tasks using box embeddings—axis-aligned hyperrectangles in low dimensional spaces—that can capture asymmetric relationships between them through volumetric overlaps. We show that TASK2BOX accurately predicts unseen hierarchical relationships between nodes in ImageNet and iNaturalist datasets, as well as transferability between tasks in the Taskonomy benchmark. We also show that box embeddings estimated from task representations (e.g., CLIP [36], Task2Vec [4], or attribute based [15]) can be used to predict relationships between unseen tasks more accurately than classifiers trained on the same representations, as well as handcrafted asymmetric distances (e.g., KL divergence)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 447, + 547, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 547, + 520 + ], + "type": "text", + "content": "This suggests that low-dimensional box embeddings can effectively capture these task relationships and have the added advantage of being interpretable. We use the approach to visualize relationships among publicly available image classification datasets on popular dataset hosting platform called Hugging Face." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 546, + 387, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 546, + 387, + 559 + ], + "spans": [ + { + "bbox": [ + 306, + 546, + 387, + 559 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "content": "The success of deep learning has led to the proliferation of datasets for solving a wide range of computer vision problems. Yet, there are few tools available to enable practitioners to find datasets related to the task at hand, and to solve various meta-tasks related to it. We present TASK2Box, a method to represent tasks using axis-aligned hyperrectangles (or box embeddings). TASK2Box is framed as a learnable mapping from dataset representation to boxes, and can be trained to predict various relationships between novel tasks such as transferability, hierarchy, and overlap." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": "Box embeddings [48] extend order embeddings [46] by using volumetric relationships between axis-aligned hyper" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28827" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "rectangles to encode pairwise relationships. Prior work in natural language processing has utilized box embeddings to represent the WordNet [30] hierarchy and to model conditional distributions. To model relationships between novel datasets, we develop a technique to map from Euclidean representations of datasets into the space of boxes. We explore simple image and label embedding from large vision-language models such as CLIP [36], TASK2VEC [4], and attribute-based vectors [15] as base representations of tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 289, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 289, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 289, + 433 + ], + "type": "text", + "content": "We test our framework to model asymmetric relationships between nodes in iNaturalist [45] and Caltech-UCSD Birds (CUB) [50] and ImageNet [11] datasets, as well as to predict transferability on the Taskonomy benchmark [52]. Table 1 and 2 show that low-dimensional box embeddings accurately predict novel relationships between datasets seen during training, as well as relationships with novel datasets. Remarkably, TASK2BOX outperforms classifiers trained to directly predict the relationships on the same representations, suggesting that the box embedding provides a strong inductive bias for learning hierarchical relationships. We also outperform simple asymmetric distances proposed in prior work such as Kullback-Leibler (KL) divergence [4]. To model the heterogeneous tasks in the Taskonomy benchmark [52] we map each task to a set of attributes from which a box embedding is learned. Once again, we obtain significantly higher correlation between the true and predicted transferability for both existing and novel datasets compared to standard classifiers (Table 3). Such attribute-based representations can be readily derived from datasheets [15] and modelcards [31]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 434, + 289, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 289, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 289, + 589 + ], + "type": "text", + "content": "Finally, the low-dimensional box embeddings have the added advantage of being interpretable. Fig. 1 and Fig. 3 show relationships on the iNaturalist+CUB and ImageNet categories, respectively. The 2D box representation allows us to readily visualize the strength and direction of task relationships based on the overlapping volumes, which is not possible using symmetric distances with Euclidean representations (e.g., t-SNE [44]). At the same time, new datasets can be embedded in constant time without needing to retrain or re-optimize. Fig. 5 uses TASK2Box to visualize relationships among 131 publicly available datasets on Hugging Face [1], a popular platform for hosting datasets. Our main contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 592, + 287, + 687 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 592, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 592, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 592, + 287, + 628 + ], + "type": "text", + "content": "- We introduce a novel method (TASK2BOX) that uses box embeddings to learn asymmetric (e.g., hierarchical, transfer learning) dataset relationships." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 628, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 287, + 651 + ], + "type": "text", + "content": "- We demonstrate that TASK2Box can predict the relationships of new tasks with a collection of existing tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 652, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 652, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 652, + 287, + 687 + ], + "type": "text", + "content": "- We illustrate the interpretability of our model, and the ability to visualize public classification datasets on Hugging Face." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "The code for this project is publicly available at https://github.com/cvl-umass/task2box." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "type": "text", + "content": "Task Representations. Given a dataset " + }, + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x_i, y_i)\\}_{i=1}^n" + }, + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "type": "text", + "content": ", consisting of images " + }, + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "type": "inline_equation", + "content": "x_i \\in \\mathcal{X}" + }, + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "type": "text", + "content": " and labels " + }, + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "type": "inline_equation", + "content": "y_i \\in \\mathcal{Y}" + }, + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "type": "text", + "content": ", a range of approaches have been proposed for dataset representation. The most straightforward approach involves modeling the distribution of either the images " + }, + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "type": "text", + "content": " or the labels " + }, + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "type": "text", + "content": " within the dataset independently, using embeddings referred to in prior work as \"domain\" and \"label\" embeddings [4]. To capture the joint dependency between images and labels, [4] proposed the use of the Fisher Information Matrix (FIM) derived from a \"probe network\" trained to minimize a loss function " + }, + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "type": "inline_equation", + "content": "\\ell(\\hat{y}, y)" + }, + { + "bbox": [ + 304, + 91, + 547, + 281 + ], + "type": "text", + "content": " over the dataset [22, 25, 35]. This approach leverages the similarity of FIMs to predict task transferability and for model selection. However, the utility of the FIM critically depends on the choice of the probe network and a pre-defined similarity may not accurately represent the various relationships between datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 283, + 547, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 547, + 426 + ], + "type": "text", + "content": "We also investigate the use of vision-language models (VLMs) such as CLIP [36]. This model, trained on a wide range of visual domains, can generalize to tasks involving vision and language data. CLIP features have been effective for image classification [3, 9], semantic segmentation [18, 24, 26], object detection [16, 47, 51], and even closing domain gaps for performance improvement [23, 53]. Both images " + }, + { + "bbox": [ + 304, + 283, + 547, + 426 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{X}" + }, + { + "bbox": [ + 304, + 283, + 547, + 426 + ], + "type": "text", + "content": " and labels " + }, + { + "bbox": [ + 304, + 283, + 547, + 426 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{Y}" + }, + { + "bbox": [ + 304, + 283, + 547, + 426 + ], + "type": "text", + "content": " represented as text, can be mapped into a shared space using the vision encoder " + }, + { + "bbox": [ + 304, + 283, + 547, + 426 + ], + "type": "inline_equation", + "content": "(\\phi)" + }, + { + "bbox": [ + 304, + 283, + 547, + 426 + ], + "type": "text", + "content": " and text encoder " + }, + { + "bbox": [ + 304, + 283, + 547, + 426 + ], + "type": "inline_equation", + "content": "(\\psi)" + }, + { + "bbox": [ + 304, + 283, + 547, + 426 + ], + "type": "text", + "content": " of CLIP, allowing us to model the dataset as a set of image and label embeddings " + }, + { + "bbox": [ + 304, + 283, + 547, + 426 + ], + "type": "inline_equation", + "content": "\\left\\{\\left(\\phi(x_i), \\psi(y_i)\\right)\\right\\}_{i=1}^n" + }, + { + "bbox": [ + 304, + 283, + 547, + 426 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 426, + 547, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 462 + ], + "type": "text", + "content": "We compare FIMs with representations derived from CLIP as base representations for tasks and learn box embeddings to model a variety of relations among tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 474, + 547, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 643 + ], + "type": "text", + "content": "Task Relations in Computer Vision. Understanding the relationships between tasks can lead to efficient solutions to new tasks. Previous work has measured task similarity by using model gradients [14] or based on their learned features [21] for grouping tasks for efficient multi-tasking. Similarly, predicting which pre-trained models will generalize the best on a new dataset could streamline model selection. Taskonomy [52] investigates transfer learning across vision tasks, varying from segmentation to pose estimation, by computing pairwise transfer distances or task affinities. These affinities are calculated by evaluating the extent to which a model trained on a source task generalizes to a target task [13, 42], though this process is computationally expensive." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Dataset Visualization. Low-dimensional Euclidean embeddings derived from UMAP [28], t-SNE [44], and LargeVis [43] are widely used to visualize relationships between datasets. They have been shown to successfully recover clusters of various data modalities by preserving the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28828" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 119 + ], + "type": "text", + "content": "relationship of each data point with its neighbors [4, 5, 40]. In low-dimension space, relationships with other data points are defined by their Euclidean distances. However, these are commonly used to represent symmetric relations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 286, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 286, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 286, + 179 + ], + "type": "text", + "content": "Visualizations using tidy trees [37], circle packing [49], or cone trees [29] organize asymmetric relations as tree-structured hierarchies in low dimension. However, cyclic data relationships cannot be properly represented for these methods (e.g., when a node has two or more parents)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 192, + 286, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 286, + 299 + ], + "type": "text", + "content": "Asymmetric Distances over Datasets. Kullback-Leibler (KL) divergence between image or label distributions provides a natural way to represent asymmetric distances between datasets. TASK2VEC [4] computes the similarity between two tasks (e.g., cosine distance), and introduces asymmetry by using the complexity of the first task as a reference. The complexity is measured by the similarity of the task embedding to a \"trivial embedding\" (embedding of a task that is easy or has no examples)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "type": "text", + "content": "Order embeddings on images were first proposed in [46] to capture tree-structured relationships. Given an dataset of " + }, + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "type": "inline_equation", + "content": "P = (u, v)" + }, + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "type": "text", + "content": " drawn from an partially ordered set " + }, + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "type": "inline_equation", + "content": "(X, \\preceq_X)" + }, + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "type": "text", + "content": ", they frame the problem as learning a mapping " + }, + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "type": "inline_equation", + "content": "f: (X, \\preceq_X) \\to (Y, \\preceq_Y)" + }, + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "type": "text", + "content": " that is order preserving, i.e., " + }, + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "type": "inline_equation", + "content": "u \\preceq_X v \\Longleftrightarrow f(u) \\preceq_Y f(v)" + }, + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "type": "text", + "content": ". The reserved product order was used for " + }, + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "type": "inline_equation", + "content": "\\preceq_Y" + }, + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "type": "inline_equation", + "content": "x \\preceq y \\Longleftrightarrow x_i > y_i, \\forall i" + }, + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "type": "text", + "content": ". Box embeddings [8] generalized this framework by representing points as axis-aligned hyper-rectangles and using volumetric relationships (e.g., intersection over union) to represent asymmetric relations. They used the framework to model conditional distributions and hypernymy relations (e.g., dog \"is a\" mammal) on the WordNet graph [2, 19, 33, 38, 48]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 286, + 563 + ], + "type": "text", + "content": "Hyperbolic spaces provide yet another way to model asymmetric relationships. Examples include the Poincaré disk model which uses hyperbolic cosine (cosh) to measure distance between points in a disk. Poincaré embeddings have been similarly used to represent WordNet hierarchies [32] and other relations in general graphs. Hyperbolic representations have also been proposed for representing images for efficient zero-shot learning given a taxonomic structure of the labels [27]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 563, + 286, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 286, + 658 + ], + "type": "text", + "content": "To the best of our knowledge, no prior work has explored the use of these spaces for representing entire datasets and their effectiveness in capturing various task relationships. We adopt box embeddings in this work due to the effectiveness over alternatives in previous work [2, 6, 34, 38], ease of visualization, as well as due to open-source libraries for robust learning. However, instead of learning box embeddings directly, we learn mappings from task representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 669, + 184, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 184, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 184, + 681 + ], + "type": "text", + "content": "3. TASK2Box Framework" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "We define the problem as follows: given a collection of datasets " + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}_1,\\mathcal{D}_2,\\dots ,\\mathcal{D}_m\\}" + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": ", and an asymmetric rela" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "tionship given the pairwise relationship between datasets " + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "d(\\mathcal{D}_i,\\mathcal{D}_j)\\in [0,1]" + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": " , we aim to encode each dataset into a low-dimension space that preserves the relationships between datasets and is interpretable." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 120, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 120, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 120, + 545, + 204 + ], + "type": "text", + "content": "To achieve this, we propose using box embeddings for encoding each of the datasets. This process involves two main steps: (1) deriving the base representation " + }, + { + "bbox": [ + 304, + 120, + 545, + 204 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 120, + 545, + 204 + ], + "type": "text", + "content": " of each dataset, and (2) learning a model " + }, + { + "bbox": [ + 304, + 120, + 545, + 204 + ], + "type": "inline_equation", + "content": "f_{\\theta}: e \\to z" + }, + { + "bbox": [ + 304, + 120, + 545, + 204 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 120, + 545, + 204 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{2 \\times k}" + }, + { + "bbox": [ + 304, + 120, + 545, + 204 + ], + "type": "text", + "content": " represents a " + }, + { + "bbox": [ + 304, + 120, + 545, + 204 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 120, + 545, + 204 + ], + "type": "text", + "content": "-dimensional axis-aligned hyperrectangle (i.e., box), denoted by its lower left and upper right coordinates. These steps are detailed further below." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 212, + 453, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 212, + 453, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 212, + 453, + 224 + ], + "type": "text", + "content": "3.1. Base Task Representations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 229, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 229, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 545, + 290 + ], + "type": "text", + "content": "Each dataset " + }, + { + "bbox": [ + 304, + 229, + 545, + 290 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x_i, y_i)\\}_{i=1}^n" + }, + { + "bbox": [ + 304, + 229, + 545, + 290 + ], + "type": "text", + "content": " is defined as a collection of pairs of images " + }, + { + "bbox": [ + 304, + 229, + 545, + 290 + ], + "type": "inline_equation", + "content": "x_i \\in \\mathcal{X}" + }, + { + "bbox": [ + 304, + 229, + 545, + 290 + ], + "type": "text", + "content": " and labels " + }, + { + "bbox": [ + 304, + 229, + 545, + 290 + ], + "type": "inline_equation", + "content": "y_i \\in \\mathcal{Y}" + }, + { + "bbox": [ + 304, + 229, + 545, + 290 + ], + "type": "text", + "content": ". For obtaining a base embedding " + }, + { + "bbox": [ + 304, + 229, + 545, + 290 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 229, + 545, + 290 + ], + "type": "text", + "content": " for each dataset, we utilize methods such as CLIP [36, 41], TASK2VEC [4], or attribute-based approaches [15]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 301, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 434 + ], + "type": "text", + "content": "CLIP. Using a pre-trained CLIP model [7, 20, 36, 41], we compute the mean and variance of the individual sample embeddings within each dataset. For each data sample, the image embedding is concatenated with the label embedding, the latter generated from text prompts (e.g., \"A photo of [CLS]\"). This concatenation models the joint distribution of images and labels. Eq. 1 and 2 detail how the mean and variance embeddings are derived, where " + }, + { + "bbox": [ + 304, + 301, + 545, + 434 + ], + "type": "inline_equation", + "content": "[i,j]" + }, + { + "bbox": [ + 304, + 301, + 545, + 434 + ], + "type": "text", + "content": " represents the concatenation of vectors " + }, + { + "bbox": [ + 304, + 301, + 545, + 434 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 301, + 545, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 301, + 545, + 434 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 301, + 545, + 434 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 301, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 301, + 545, + 434 + ], + "type": "text", + "content": " is the vision encoder, and " + }, + { + "bbox": [ + 304, + 301, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 301, + 545, + 434 + ], + "type": "text", + "content": " is the text encoder. The covariance is approximated as diagonal for tractability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 442, + 545, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 442, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 361, + 442, + 545, + 475 + ], + "type": "interline_equation", + "content": "\\mu_ {C L I P} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} [ \\phi (x _ {i}), \\psi (y _ {i}) ] \\tag {1}", + "image_path": "777094d2ce5cfd12d63379e45ea60cb7ba8d396dc2189a9a27a813685400aeba.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 334, + 491, + 545, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 491, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 334, + 491, + 545, + 524 + ], + "type": "interline_equation", + "content": "\\sigma_ {C L I P} ^ {2} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\left[ \\phi \\left(x _ {i}\\right), \\psi \\left(y _ {i}\\right) \\right] - \\mu_ {C L I P}\\right) ^ {2} \\tag {2}", + "image_path": "e7b925997aa77b163bdbc03e0021d89baaaaf57ff473efe5d21ad32c07787e4c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "text", + "content": "The base representation is defined as " + }, + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "inline_equation", + "content": "e := \\mu_{CLIP} \\in \\mathbb{R}^{2048}" + }, + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "inline_equation", + "content": "e := [\\mu_{CLIP}, \\sigma_{CLIP}^2] \\in \\mathbb{R}^{4096}" + }, + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "text", + "content": ". A ViT-H/14 [12] pretrained on LAION-2B [41] was used to extract the embeddings." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "TASK2VEC [4] encodes a dataset using the approximate Fisher Information Matrix (FIM) of a network trained on the given dataset. The FIM represents the importance of parameters in the feature extractor by perturbing the weights " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{w}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " of a given probe network with Gaussian noise " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,\\Lambda)" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ". The precision matrix " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " is estimated to be close to an isotropic prior " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\hat{w},\\lambda^2 I)" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " while having a good expected error. Eq. 3 is minimized to find " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " is the cross entropy loss, " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{w}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " are the weights of the network, " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " is the magnitude of the prior, " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "x\\in \\mathcal{X}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "y\\in \\mathcal{Y}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28829" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 58, + 98, + 218, + 111 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 98, + 218, + 111 + ], + "spans": [ + { + "bbox": [ + 58, + 98, + 218, + 111 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\hat {w}; \\Lambda) = \\mathbb {E} _ {w \\sim \\mathcal {N} (\\hat {w}, \\Lambda)} [ H _ {p _ {w}, \\tilde {p}} p (y | x) ] +", + "image_path": "20f857d2ff866ed07059cca5c43e65660124898330e49a3bea334632aaaba860.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 149, + 114, + 287, + 128 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 114, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 287, + 128 + ], + "type": "interline_equation", + "content": "\\beta K L \\left(\\mathcal {N} (0, \\Lambda) \\| \\mathcal {N} (0, \\lambda^ {2} I)\\right) \\tag {3}", + "image_path": "5f69bd18c1c34e55f511d80b77d2f1a13fe24fecedd9b5f674b1d35067a1bfb3.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 139, + 287, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 139, + 287, + 199 + ], + "spans": [ + { + "bbox": [ + 47, + 139, + 287, + 199 + ], + "type": "text", + "content": "The matrix " + }, + { + "bbox": [ + 47, + 139, + 287, + 199 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 47, + 139, + 287, + 199 + ], + "type": "text", + "content": " provides an estimate of the FIM and is approximated as a diagonal matrix. The diagonal components are used as a base representation of a dataset (" + }, + { + "bbox": [ + 47, + 139, + 287, + 199 + ], + "type": "inline_equation", + "content": "e := \\mathrm{FIM} \\in \\mathbb{R}^{17024}" + }, + { + "bbox": [ + 47, + 139, + 287, + 199 + ], + "type": "text", + "content": "). ResNet-34 [17] pretrained on ImageNet [11] is used as the probe network for all datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 211, + 287, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 211, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 47, + 211, + 287, + 307 + ], + "type": "text", + "content": "Attribute-based. A task can be characterized by a set of " + }, + { + "bbox": [ + 47, + 211, + 287, + 307 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 211, + 287, + 307 + ], + "type": "text", + "content": " binary attributes [15] represented as a vector of dimension " + }, + { + "bbox": [ + 47, + 211, + 287, + 307 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 211, + 287, + 307 + ], + "type": "text", + "content": ". Some of the attributes explored for representing tasks are: (1) Is the task generative? (2) Is the task output in 2D? (3) Does the task involve camera pose estimation? Taking these 3 characteristics, for example, we can represent a 2D segmentation task as the vector " + }, + { + "bbox": [ + 47, + 211, + 287, + 307 + ], + "type": "inline_equation", + "content": "e = [0,1,0]" + }, + { + "bbox": [ + 47, + 211, + 287, + 307 + ], + "type": "text", + "content": " as a discriminative 2D task that does not need camera poses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 308, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 392 + ], + "type": "text", + "content": "Tasks in Taskonomy [52] involving multiple modalities benefit from this attribute-based representation due to its model independence. This approach also enables generalization to unseen tasks by identifying the presence or absence of various characteristics. Each vision task in Taskonomy is represented with 15 attributes, resulting in vectors " + }, + { + "bbox": [ + 46, + 308, + 287, + 392 + ], + "type": "inline_equation", + "content": "e \\in \\mathbb{R}^{15}" + }, + { + "bbox": [ + 46, + 308, + 287, + 392 + ], + "type": "text", + "content": ". The full list of attributes is in Appendix ??" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 401, + 195, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 401, + 195, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 401, + 195, + 415 + ], + "type": "text", + "content": "3.2. Learning Box Embeddings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "text", + "content": "From a base task representation " + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "text", + "content": ", we learn the parameters " + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "text", + "content": " of a model " + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "f_{\\theta}: e \\to z" + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "text", + "content": " that preserves the asymmetric similarity function " + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "d(\\mathcal{D}_i, \\mathcal{D}_j)" + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "text", + "content": " between any two datasets " + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_j" + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "text", + "content": ". In Eq. 4, the learning objective is shown where " + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_E" + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "text", + "content": " is a loss function (mean squared error), " + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "text", + "content": " is a distance function, " + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "text", + "content": " is a regularization term, and " + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "text", + "content": " is a hyperparameter. The embeddings " + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "z_i, z_j \\in \\mathbb{R}^{2 \\times k}" + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "text", + "content": " represent the coordinates of the lower left and the upper right corners of the respective " + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "text", + "content": " dimensional boxes, with " + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "f_{\\theta}(e_i) = z_i" + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "f_{\\theta}(e_j) = z_j" + }, + { + "bbox": [ + 47, + 420, + 287, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 552, + 287, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 552, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 56, + 552, + 287, + 594 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\theta} = \\underset {\\theta} {\\operatorname {a r g m i n}} \\sum_ {i, j} \\left(\\mathcal {L} _ {E} \\left(d \\left(\\mathcal {D} _ {i}, \\mathcal {D} _ {j}\\right), d _ {b o x} \\left(z _ {i}, z _ {j}\\right)\\right) \\right. \\\\ + \\lambda \\mathcal {L} _ {D} \\left(z _ {i}, z _ {j}\\right) + \\mathcal {L} _ {R}) \\tag {4} \\\\ \\end{array}", + "image_path": "960b8d807f8bb6449c9ebb9e4012d46055ea962da3fc15c35e482b1494f9fdb0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": "The asymmetric relationship between box embeddings, denoted as " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "d_{box}(z_i,z_j)" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": ", is computed in Eq. 5 by calculating the volume of the intersection between " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_j" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": ", normalized by the volume of " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": " to be fully contained inside " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_j" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_i \\subset z_j" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": "), it is required that " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "d_{box}(z_i,z_j) = 1" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": ". Conversely, for " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_j" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": " to only partially contain " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "d_{box}(z_j,z_i)" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": " must fall within the range (0,1). Fig. 2 illustrates this through a 2-dimensional example, where " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_1" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": " represents the box embedding for Canidae, and " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_2" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": " for Mammalia." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 356, + 72, + 495, + 179 + ], + "blocks": [ + { + "bbox": [ + 356, + 72, + 495, + 179 + ], + "lines": [ + { + "bbox": [ + 356, + 72, + 495, + 179 + ], + "spans": [ + { + "bbox": [ + 356, + 72, + 495, + 179 + ], + "type": "image", + "image_path": "dd107d1d5b3fc95821a19aca38ad74268e975911a6a8c983031b81716cb2100a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 183, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 304, + 183, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 545, + 240 + ], + "type": "text", + "content": "Figure 2. TASK2Box Embeddings in 2D for Mammalia, Canidae, and Amphibia Datasets from iNaturalist. Each embedding represents the coordinates of the lower left and upper right corners of each box/rectangle. Since Canidae " + }, + { + "bbox": [ + 304, + 183, + 545, + 240 + ], + "type": "inline_equation", + "content": "(z_{1})" + }, + { + "bbox": [ + 304, + 183, + 545, + 240 + ], + "type": "text", + "content": " is a proper subset of Mammalia " + }, + { + "bbox": [ + 304, + 183, + 545, + 240 + ], + "type": "inline_equation", + "content": "(z_{2})" + }, + { + "bbox": [ + 304, + 183, + 545, + 240 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 183, + 545, + 240 + ], + "type": "inline_equation", + "content": "d_{box}(z_{1}, z_{2}) = 1" + }, + { + "bbox": [ + 304, + 183, + 545, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 183, + 545, + 240 + ], + "type": "inline_equation", + "content": "d_{box}(z_{2}, z_{1}) = 0.1" + }, + { + "bbox": [ + 304, + 183, + 545, + 240 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 367, + 260, + 545, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 260, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 367, + 260, + 545, + 285 + ], + "type": "interline_equation", + "content": "d _ {b o x} \\left(z _ {i}, z _ {j}\\right) = \\frac {\\operatorname {v o l} \\left(z _ {i} \\cap z _ {j}\\right)}{\\operatorname {v o l} \\left(z _ {i}\\right)} \\tag {5}", + "image_path": "61d0d5540ce12d4cc4af04754b7079e63b5d43482f8ce3593029d9126eb049cb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "text", + "content": " is applied to datasets where " + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "inline_equation", + "content": "d(\\mathcal{D}_i, \\mathcal{D}_j) > 0" + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "text", + "content": " such that the Euclidean distance between the center coordinates of " + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "inline_equation", + "content": "z_j" + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "text", + "content": " is minimized when starting from a non-overlapping state. This allows non-overlapping embeddings to move closer to each other, complementing " + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_E" + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "text", + "content": " to learn relationships. " + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "text", + "content": " encourages solutions with regular-shaped boxes for better interpretability. The formulation of " + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "text", + "content": ", given in Eq. 6, applies to a " + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "text", + "content": "-dimensional box embedding " + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "inline_equation", + "content": "s_a" + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "text", + "content": " represents the size of the " + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "text", + "content": "-th dimension (for example, width), and " + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 304, + 293, + 545, + 472 + ], + "type": "text", + "content": " are hyperparameters. To prevent the trivial solution of minimizing box volume to zero, the inverse of the box volume is included. It's crucial to normalize the terms with respect to the embedding dimension, as the first and second terms scale quadratically and exponentially with dimension increase, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 492, + 545, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 545, + 526 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R} = \\left(\\frac {\\alpha}{k ^ {2}} \\sum_ {a = 1} ^ {k} \\sum_ {b = a + 1} ^ {k} \\left| s _ {a} - s _ {b} \\right|\\right) + \\beta (\\operatorname {v o l} (z _ {i})) ^ {- 1 / k} \\tag {6}", + "image_path": "158cbe3942b7ea9d65e14a3c2d734f6b5d37778e3f8a4a6c3d2b2a6e291799a1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 536, + 545, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 545, + 583 + ], + "type": "text", + "content": "The architecture of " + }, + { + "bbox": [ + 304, + 536, + 545, + 583 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 536, + 545, + 583 + ], + "type": "text", + "content": " consists of three fully-connected layers followed by two linear heads: one predicts a " + }, + { + "bbox": [ + 304, + 536, + 545, + 583 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 536, + 545, + 583 + ], + "type": "text", + "content": "-dimensional lower-left coordinate, and the other predicts the " + }, + { + "bbox": [ + 304, + 536, + 545, + 583 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 536, + 545, + 583 + ], + "type": "text", + "content": "-dimensional sizes of each box dimension." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 596, + 388, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 596, + 388, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 596, + 388, + 610 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 616, + 544, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 544, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 544, + 640 + ], + "type": "text", + "content": "We consider the following goals to evaluate the capability of Task2Box to represent datasets:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 642, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 306, + 642, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 642, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 306, + 642, + 545, + 677 + ], + "type": "text", + "content": "1. Given a collection of existing datasets " + }, + { + "bbox": [ + 306, + 642, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_E" + }, + { + "bbox": [ + 306, + 642, + 545, + 677 + ], + "type": "text", + "content": " and a subset of pairwise relationships " + }, + { + "bbox": [ + 306, + 642, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 306, + 642, + 545, + 677 + ], + "type": "text", + "content": ": can the model generalize on unseen relationships " + }, + { + "bbox": [ + 306, + 642, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{R}'" + }, + { + "bbox": [ + 306, + 642, + 545, + 677 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 306, + 642, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_E" + }, + { + "bbox": [ + 306, + 642, + 545, + 677 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 306, + 642, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{R}' \\cap \\mathcal{R} = \\emptyset" + }, + { + "bbox": [ + 306, + 642, + 545, + 677 + ], + "type": "text", + "content": "?" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "2. Given a collection of novel datasets " + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_N" + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": " not seen during training: can the model accurately identify the relationships with the existing datasets " + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_E" + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "?" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28830" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 70, + 487, + 231 + ], + "blocks": [ + { + "bbox": [ + 109, + 70, + 487, + 231 + ], + "lines": [ + { + "bbox": [ + 109, + 70, + 487, + 231 + ], + "spans": [ + { + "bbox": [ + 109, + 70, + 487, + 231 + ], + "type": "table", + "html": "
iNaturalist + CUB
MethodExisting DatasetsNovel Datasets
μCLIP[μ,σ2]CLIPFIMμCLIP[μ,σ2]CLIPFIM
TASK2Box (2D)69.23%67.84%39.61%50.07%39.66%10.06%
TASK2Box (3D)79.66%79.35%57.63%70.04%64.53%20.65%
TASK2Box (5D)84.67%82.41%79.72%73.79%72.11%34.88%
MLP Classifier45.25%61.45%26.34%39.06%44.54%19.90%
Linear Classifier4.40%3.11%7.06%4.77%5.87%15.92%
KL Divergence-6.58%7.94%-5.90%0.00%
Asymmetric Cosine9.29%11.54%2.83%1.47%1.47%1.47%
Asymmetric Euclidean1.71%1.71%8.53%1.47%1.47%1.91%
Random2.06%1.49%
", + "image_path": "60af68a1bd48f6c62a154eac0c3781135184607bac64ac1295a134b31a09aa31.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 238, + 547, + 295 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 547, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 547, + 295 + ], + "type": "text", + "content": "Table 1. Average F1 Score for Predicting Hierarchical Relationships on iNaturalist + CUB Dataset. For all feature types, TASK2Box outperforms other methods by more than " + }, + { + "bbox": [ + 46, + 238, + 547, + 295 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 238, + 547, + 295 + ], + "type": "text", + "content": " on existing datasets and more than " + }, + { + "bbox": [ + 46, + 238, + 547, + 295 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 238, + 547, + 295 + ], + "type": "text", + "content": " on novel datasets. The best-performing model is in bold, and the second-best is underlined. Results on Novel Datasets demonstrate that our model can generalize beyond the seen tasks it has previously seen. The dimensions such as 2D, 3D, and 5D refer to different box dimensionalities. Results using KL divergence for " + }, + { + "bbox": [ + 46, + 238, + 547, + 295 + ], + "type": "inline_equation", + "content": "\\mu_{CLIP}" + }, + { + "bbox": [ + 46, + 238, + 547, + 295 + ], + "type": "text", + "content": " are not shown since it is not a distribution." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 308, + 288, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 288, + 380 + ], + "type": "text", + "content": "The goals are evaluated through two experimental setups, demonstrating our model's ability to predict various types of relationships. The configurations, and their corresponding datasets, relationships, and metrics, are detailed below. Baseline methods for performance comparison are also reviewed. Implementation details are in the Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 388, + 164, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 164, + 401 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 164, + 401 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 406, + 217, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 217, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 217, + 419 + ], + "type": "text", + "content": "4.1.1 Hierarchical Task Relationships" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "text", + "content": "We use a combination of iNaturalist [45] and Caltech-UCSD Birds (CUB) [50], and instrument-related classes in ImageNet [11] to evaluate the ability of TASK2Box to represent hierarchical relationships. The first two are composed of images of various species, and the third is composed of images of various objects. The classes naturally follow a hierarchical form based on biological classification (taxonomy of iNaturalist+CUB species), and on semantic relations (hypernymy of objects in ImageNet)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 629 + ], + "type": "text", + "content": "Datasets. For iNaturalist+CUB, the datasets are defined as the classes, the orders, and the families in the taxonomy that contain a significant number of samples per dataset as in [4]. There are 47 classes, 202 orders, and 589 families for a total of 838 datasets. For ImageNet, the instrument-related objects were processed using WordNet [30] to obtain hierarchical information between classes. This resulted in 131 datasets for training and evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "Dataset Relationships. For any two datasets, their relationship is captured as " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "d(\\mathcal{D}_i,\\mathcal{D}_j)\\in \\{0,1\\}" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "d(\\mathcal{D}_i,\\mathcal{D}_j) = 1" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": " if and only if " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i\\subset \\mathcal{D}_j" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "d(\\mathcal{D}_i,\\mathcal{D}_j) = 0" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": " otherwise. Fig. 2 provides an example: Canidae " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_1)" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": " is a family within the class Mammalia " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_2)" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "; thus, " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "d(\\mathcal{D}_1,\\mathcal{D}_2) = 1" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "d(\\mathcal{D}_2,\\mathcal{D}_1) = 0" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": ". Meanwhile, Amphibia " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_3)" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": " is unrelated to either dataset, resulting in " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "d(\\mathcal{D}_1,\\mathcal{D}_3) = 0" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 308, + 547, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 490 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate the ability of the model to classify the presence of containment relationships between datasets using the F1 score due to the imbalance between positive and negative relationships. To obtain F1 score, precision and recall are first calculated. Precision is computed as the ratio of true positive pairs predicted to the total number of positive predictions. Recall is the ratio of true positive pairs predicted to the total number of true positive pairs. F1 score is then reported as the harmonic mean between the precision and recall. These calculations are done on both (1) the unseen relationships " + }, + { + "bbox": [ + 304, + 308, + 547, + 490 + ], + "type": "inline_equation", + "content": "\\mathcal{R}'" + }, + { + "bbox": [ + 304, + 308, + 547, + 490 + ], + "type": "text", + "content": " within existing datasets " + }, + { + "bbox": [ + 304, + 308, + 547, + 490 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_E" + }, + { + "bbox": [ + 304, + 308, + 547, + 490 + ], + "type": "text", + "content": " and (2) the relationships between novel datasets " + }, + { + "bbox": [ + 304, + 308, + 547, + 490 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_N" + }, + { + "bbox": [ + 304, + 308, + 547, + 490 + ], + "type": "text", + "content": " and existing datasets " + }, + { + "bbox": [ + 304, + 308, + 547, + 490 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_E" + }, + { + "bbox": [ + 304, + 308, + 547, + 490 + ], + "type": "text", + "content": ". For the latter, we evaluate the relationships in both directions, i.e., " + }, + { + "bbox": [ + 304, + 308, + 547, + 490 + ], + "type": "inline_equation", + "content": "\\left\\{\\left(\\mathcal{D}_e,\\mathcal{D}_n\\right)\\forall \\mathcal{D}_e\\in \\mathcal{D}_E\\right\\} \\cup \\left\\{\\left(\\mathcal{D}_n,\\mathcal{D}_e\\right)\\forall \\mathcal{D}_e\\in \\mathcal{D}_E\\right\\}" + }, + { + "bbox": [ + 304, + 308, + 547, + 490 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 502, + 495, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 495, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 495, + 514 + ], + "type": "text", + "content": "4.1.2 Transfer Learning Between Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 545 + ], + "type": "text", + "content": "The Taskonomy [52] benchmark is used to evaluate the ability of TASK2BOX to predict task affinities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 605 + ], + "type": "text", + "content": "Datasets. Taskonomy defines a set of 25 visual tasks with corresponding pairwise task affinities. These tasks range from object detection, semantic segmentation, pose estimation, and more. We treat each visual task as a dataset. The tasks are described in Appendix ??" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "Dataset Relationships. In contrast to the containment relationship defined in § 4.1.1, Taskonomy quantifies relationships with task affinity measured by the performance gain achieved by transfer learning from a source dataset " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_j" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": " to a target dataset " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": ". The values are computed using Analytic Hierarchy Process [39, 52]. Dataset relationships are computed and normalized based on an ordinal approach and determined by the percentage of images that transfer well to a target task given a set of source tasks as detailed in [52]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28831" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 70, + 487, + 231 + ], + "blocks": [ + { + "bbox": [ + 109, + 70, + 487, + 231 + ], + "lines": [ + { + "bbox": [ + 109, + 70, + 487, + 231 + ], + "spans": [ + { + "bbox": [ + 109, + 70, + 487, + 231 + ], + "type": "table", + "html": "
ImageNet
MethodExisting DatasetsNovel Datasets
μCLIP[μ,σ2]CLIPFIMμCLIP[μ,σ2]CLIPFIM
TASK2Box (2D)62.72%63.33%31.32%47.76%48.35%9.46%
TASK2Box (3D)83.12%82.79%58.16%73.06%66.48%24.70%
TASK2Box (5D)90.58%88.48%64.91%76.95%78.84%37.39%
MLP Classifier54.20%60.43%44.85%62.24%64.56%41.22%
Linear Classifier9.84%8.33%25.46%11.89%11.87%22.98%
KL-5.53%9.90%-10.41%0.00%
Asymmetric Cosine4.28%4.28%6.92%4.52%4.52%0.00%
Asymmetric Euclidean3.73%3.73%7.16%4.52%4.52%4.73%
Random3.64%5.02%
", + "image_path": "c647526b7000b1e089fcfa92244d00d2d7307c1c7f25f71869f7583551b24c4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 238, + 547, + 273 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 547, + 273 + ], + "type": "text", + "content": "Table 2. Average F1 Score for Predicting Hierarchical Relationships on ImageNet Instruments. TASK2BOX is shown to generalize well on both predicting relationships between existing datasets (Existing Datasets), and predicting box embeddings of new datasets and their relationships with existing datasets (Novel Datasets)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 285, + 287, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 334 + ], + "type": "text", + "content": "For a pair of datasets " + }, + { + "bbox": [ + 46, + 285, + 287, + 334 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_i,\\mathcal{D}_j)" + }, + { + "bbox": [ + 46, + 285, + 287, + 334 + ], + "type": "text", + "content": ", the task affinity is defined as " + }, + { + "bbox": [ + 46, + 285, + 287, + 334 + ], + "type": "inline_equation", + "content": "d(\\mathcal{D}_i,\\mathcal{D}_j)\\in [0,1]" + }, + { + "bbox": [ + 46, + 285, + 287, + 334 + ], + "type": "text", + "content": ". The higher the task affinity to a target dataset " + }, + { + "bbox": [ + 46, + 285, + 287, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 46, + 285, + 287, + 334 + ], + "type": "text", + "content": " from a source dataset " + }, + { + "bbox": [ + 46, + 285, + 287, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_j" + }, + { + "bbox": [ + 46, + 285, + 287, + 334 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 46, + 285, + 287, + 334 + ], + "type": "inline_equation", + "content": "d(\\mathcal{D}_i,\\mathcal{D}_j)" + }, + { + "bbox": [ + 46, + 285, + 287, + 334 + ], + "type": "text", + "content": " gets closer to 1, where a value of 1 would show " + }, + { + "bbox": [ + 46, + 285, + 287, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i\\subset \\mathcal{D}_j" + }, + { + "bbox": [ + 46, + 285, + 287, + 334 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 334, + 288, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 288, + 418 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate the prediction of the model based on the Spearman correlation between the ground truth task affinity values, and the predicted values based on the box distances in Eq. 5. Similar to § 4.1.1, we evaluate on both (1) unseen relationships " + }, + { + "bbox": [ + 46, + 334, + 288, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{R}'" + }, + { + "bbox": [ + 46, + 334, + 288, + 418 + ], + "type": "text", + "content": " for existing datasets, and (2) on relationships between unseen novel datasets " + }, + { + "bbox": [ + 46, + 334, + 288, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_N" + }, + { + "bbox": [ + 46, + 334, + 288, + 418 + ], + "type": "text", + "content": " and existing datasets " + }, + { + "bbox": [ + 46, + 334, + 288, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_E" + }, + { + "bbox": [ + 46, + 334, + 288, + 418 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 426, + 153, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 153, + 438 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 153, + 438 + ], + "type": "text", + "content": "4.2. Baseline Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 445, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 518 + ], + "type": "text", + "content": "We compare the performance of TASK2Box with alternative models and simple asymmetric distances proposed in prior work. For hierarchical relationships, given two datasets " + }, + { + "bbox": [ + 46, + 445, + 287, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i,\\mathcal{D}_j" + }, + { + "bbox": [ + 46, + 445, + 287, + 518 + ], + "type": "text", + "content": ", the models predict " + }, + { + "bbox": [ + 46, + 445, + 287, + 518 + ], + "type": "inline_equation", + "content": "\\hat{d} (\\mathcal{D}_i,\\mathcal{D}_j)\\in \\{0,1\\}" + }, + { + "bbox": [ + 46, + 445, + 287, + 518 + ], + "type": "text", + "content": ". For task affinity, the models predict " + }, + { + "bbox": [ + 46, + 445, + 287, + 518 + ], + "type": "inline_equation", + "content": "\\hat{d} (\\mathcal{D}_i,\\mathcal{D}_j)\\in [0,1]" + }, + { + "bbox": [ + 46, + 445, + 287, + 518 + ], + "type": "text", + "content": ". The structure for the various methods are discussed below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 519, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 555 + ], + "type": "text", + "content": "Linear Model. A linear model is trained to predict the relationship value between two datasets. The input is the concatenation of the base representation " + }, + { + "bbox": [ + 46, + 519, + 287, + 555 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 519, + 287, + 555 + ], + "type": "text", + "content": " for the two datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 555, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 579 + ], + "type": "text", + "content": "MLP Model. A 4-layer MLP is used instead for prediction on the same inputs as the linear model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 580, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 677 + ], + "type": "text", + "content": "KL Divergence. Each dataset is treated as a multivariate Gaussian using the mean and variance of the image and label features (CLIP) or directly as the FIM. The optimal threshold " + }, + { + "bbox": [ + 46, + 580, + 287, + 677 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 580, + 287, + 677 + ], + "type": "text", + "content": " is selected for the minimum distance between two dataset distributions " + }, + { + "bbox": [ + 46, + 580, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}(\\mathcal{D}_i||\\mathcal{D}_j) < t" + }, + { + "bbox": [ + 46, + 580, + 287, + 677 + ], + "type": "text", + "content": " for a prediction of " + }, + { + "bbox": [ + 46, + 580, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\hat{d} (\\mathcal{D}_i,\\mathcal{D}_j) = 1" + }, + { + "bbox": [ + 46, + 580, + 287, + 677 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 580, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\hat{d} (\\mathcal{D}_i,\\mathcal{D}_j) = 0" + }, + { + "bbox": [ + 46, + 580, + 287, + 677 + ], + "type": "text", + "content": " otherwise. The F1 Score (hierarchical) or the correlation (task affinity) is the objective for selecting " + }, + { + "bbox": [ + 46, + 580, + 287, + 677 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 580, + 287, + 677 + ], + "type": "text", + "content": " on the train set." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "Asymmetric Cosine Similarity. The cosine similarity " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "d_{\\text{cos}}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " is a symmetric measure between two embeddings " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "e_i, e_j" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": ". An asymmetric variant was proposed in [4] and shown in Eq. 7" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 351, + 286, + 502, + 415 + ], + "blocks": [ + { + "bbox": [ + 351, + 286, + 502, + 415 + ], + "lines": [ + { + "bbox": [ + 351, + 286, + 502, + 415 + ], + "spans": [ + { + "bbox": [ + 351, + 286, + 502, + 415 + ], + "type": "image", + "image_path": "09083c2f5ac4f3c26485b217f6bc1070383f1c2974afcfe2dee3abb75e6af820.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 420, + 545, + 464 + ], + "lines": [ + { + "bbox": [ + 305, + 420, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 420, + 545, + 464 + ], + "type": "text", + "content": "Figure 3. Visualization of Instrument-related Datasets in ImageNet. Datasets that belong to the same superset are shaded in the same color. TASK2Box learns the hierarchy of various groups, and clusters similar datasets closer." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 475, + 545, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 475, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 305, + 475, + 545, + 512 + ], + "type": "text", + "content": "by considering the similarity of " + }, + { + "bbox": [ + 305, + 475, + 545, + 512 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 305, + 475, + 545, + 512 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 475, + 545, + 512 + ], + "type": "inline_equation", + "content": "e_j" + }, + { + "bbox": [ + 305, + 475, + 545, + 512 + ], + "type": "text", + "content": " relative to the complexity of " + }, + { + "bbox": [ + 305, + 475, + 545, + 512 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 305, + 475, + 545, + 512 + ], + "type": "text", + "content": ". The complexity of is measured as the distance to the trivial embedding " + }, + { + "bbox": [ + 305, + 475, + 545, + 512 + ], + "type": "inline_equation", + "content": "e_o" + }, + { + "bbox": [ + 305, + 475, + 545, + 512 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 475, + 545, + 512 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 475, + 545, + 512 + ], + "type": "text", + "content": " is a hyperparameter." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 334, + 524, + 545, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 524, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 334, + 524, + 545, + 537 + ], + "type": "interline_equation", + "content": "d _ {a s y m} \\left(e _ {i}, e _ {j}\\right) = d _ {c o s} \\left(e _ {i}, e _ {j}\\right) - \\alpha d _ {c o s} \\left(e _ {i}, e _ {o}\\right) \\tag {7}", + "image_path": "6d43fc3bf4e058d1fc6ec39f96ec254f7805044e6015ea4543a1cf25531e89b0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 543, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 543, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 305, + 543, + 545, + 578 + ], + "type": "text", + "content": "An optimal threshold " + }, + { + "bbox": [ + 305, + 543, + 545, + 578 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 543, + 545, + 578 + ], + "type": "text", + "content": " is found on the train set where " + }, + { + "bbox": [ + 305, + 543, + 545, + 578 + ], + "type": "inline_equation", + "content": "d_{asym}(e_i, e_j) < t" + }, + { + "bbox": [ + 305, + 543, + 545, + 578 + ], + "type": "text", + "content": " results to a prediction " + }, + { + "bbox": [ + 305, + 543, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\hat{d}(\\mathcal{D}_i, \\mathcal{D}_j) = 1" + }, + { + "bbox": [ + 305, + 543, + 545, + 578 + ], + "type": "text", + "content": ". The same threshold " + }, + { + "bbox": [ + 305, + 543, + 545, + 578 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 543, + 545, + 578 + ], + "type": "text", + "content": " is used for test set evaluation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 579, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 579, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 305, + 579, + 545, + 614 + ], + "type": "text", + "content": "Asymmetric Euclidean Similarity. The asymmetric similarity is computed as in Eq. 7 but uses the Euclidean distance instead of cosine." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 615, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 615, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 305, + 615, + 545, + 640 + ], + "type": "text", + "content": "Random. The probability of containment (for hierarchical) or the value of the task affinity is uniformly random." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 651, + 437, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 437, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 437, + 662 + ], + "type": "text", + "content": "5. Results and Discussion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 670, + 454, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 454, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 454, + 684 + ], + "type": "text", + "content": "5.1. Hierarchical Relationships" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Tables 1 and 2 present the average F1 score for predicting hierarchical relationships on iNaturalist+CUB and Ima" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28832" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "geNet datasets. Various methods of extracting dataset embeddings were evaluated. TASK2BOX outperforms baseline methods, indicating that TASK2BOX can both generalize across unseen relationships (Existing Datasets), and to accurately represent relationships with existing datasets for unseen datasets (Novel Datasets). CLIP features also generalize better than FIM features, perhaps because the CLIP multi-modal embedding of images and labels was trained across a broad set of domains[36]. However, while CLIP is confined to image and text modalities, FIM can accommodate any modality by adapting the probe network's last layer for different prediction tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 289, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 348 + ], + "type": "text", + "content": "Beyond predicting relationships, our method allows the visualization of datasets. Fig. 1 illustrates this for a subset of datasets in iNaturalist and CUB using TASK2Box (2D), and Fig. 3 for ImageNet. The hierarchical organization of the datasets is apparent in the TASK2Box representation – datasets that contain others appear as larger boxes, while more specialized datasets are depicted as smaller boxes netted with borader, more general dataset boxes. However, while 2D visualization offers insights, the flat surface may restrict the representation of complex relationships. Representations in higher dimensions is explored in § 5.4." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 356, + 130, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 130, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 130, + 369 + ], + "type": "text", + "content": "5.2. Task Affinity" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 374, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 495 + ], + "type": "text", + "content": "Table 3 presents the results of task representations in Taskonomy. Our method shows that, even with attribute-based embeddings, it can learn box embeddings through relationship supervision between datasets. TASK2Box is shown to correlate highly with the ground truth task affinities compared to other methods. Given that only 25 tasks are available in Taskonomy, with 3 held out of training as novel datasets, there is a higher uncertainty in predicting unseen datasets. We expect that performance on novel datasets will improve with more datasets available during training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 495, + 288, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 288, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 288, + 627 + ], + "type": "text", + "content": "Fig. 4 displays the learned representation with TASK2Box (2D). Each subfigure in (a)-(c) represents a subset of tasks identified as having strong transfer relationships. TASK2Box not only identifies but also visually represents related tasks suitable for fine-tuning. The small highlighted boxes indicate target tasks, while the larger enclosing boxes represent source tasks. Although other tasks may not be proper subsets, the box distance in Eq. 5 provides an estimate for task affinity. A small box distance between two datasets, " + }, + { + "bbox": [ + 46, + 495, + 288, + 627 + ], + "type": "inline_equation", + "content": "d_{box}(z_1, z_2)" + }, + { + "bbox": [ + 46, + 495, + 288, + 627 + ], + "type": "text", + "content": ", suggests lower transfer performance from a source task " + }, + { + "bbox": [ + 46, + 495, + 288, + 627 + ], + "type": "inline_equation", + "content": "z_2" + }, + { + "bbox": [ + 46, + 495, + 288, + 627 + ], + "type": "text", + "content": " to a target task " + }, + { + "bbox": [ + 46, + 495, + 288, + 627 + ], + "type": "inline_equation", + "content": "z_1" + }, + { + "bbox": [ + 46, + 495, + 288, + 627 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 634, + 197, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 197, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 197, + 647 + ], + "type": "text", + "content": "5.3. Visualizing Public Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Apart from predicting both hierarchical relationships and quantified task affinities, our method can also be used to visualize public datasets that lack available ground truth relationships. Using a set of vision tasks from Hugging Face [1], " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mu_{CLIP} + \\text{TASK2Box}" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " (2D) were utilized to pre" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 77, + 384, + 138 + ], + "blocks": [ + { + "bbox": [ + 318, + 77, + 384, + 138 + ], + "lines": [ + { + "bbox": [ + 318, + 77, + 384, + 138 + ], + "spans": [ + { + "bbox": [ + 318, + 77, + 384, + 138 + ], + "type": "image", + "image_path": "fe7890c962f7a3223caf4add0c39e6794c38b0aadb9eb1dd83a160fbaef0df95.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 141, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 304, + 141, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 304, + 141, + 545, + 208 + ], + "type": "text", + "content": "Figure 4. Visualization of Tasks in Taskonomy showing Source Tasks that Transfer Well to Target Tasks (shaded). (a) Jigsaw and Triplet Fixated Camera Pose estimation are source tasks that transfer well to Depth estimation. (b) and (c) show different source tasks (larger boxes) that transfer well to the shaded boxes of Denoising Autoencoder and Surface Normal, respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 394, + 78, + 459, + 137 + ], + "blocks": [ + { + "bbox": [ + 394, + 78, + 459, + 137 + ], + "lines": [ + { + "bbox": [ + 394, + 78, + 459, + 137 + ], + "spans": [ + { + "bbox": [ + 394, + 78, + 459, + 137 + ], + "type": "image", + "image_path": "ef7aace0e3560b707f6b06d87f3df34013a0b1ab23bcf8cfdab46fef5df1f363.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 469, + 79, + 533, + 137 + ], + "blocks": [ + { + "bbox": [ + 469, + 79, + 533, + 137 + ], + "lines": [ + { + "bbox": [ + 469, + 79, + 533, + 137 + ], + "spans": [ + { + "bbox": [ + 469, + 79, + 533, + 137 + ], + "type": "image", + "image_path": "10fc35761ec0efa7ef3b12f441f4ec9681efb41db44c4140c3e72356248047c3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 314, + 211, + 538, + 318 + ], + "blocks": [ + { + "bbox": [ + 314, + 211, + 538, + 318 + ], + "lines": [ + { + "bbox": [ + 314, + 211, + 538, + 318 + ], + "spans": [ + { + "bbox": [ + 314, + 211, + 538, + 318 + ], + "type": "table", + "html": "
MethodExisting DatasetsNovel Datasets
Spearman's ρSpearman's ρ
TASK2Box (2D)0.85 ± 0.060.12 ± 0.21
TASK2Box (3D)0.93 ± 0.020.48 ± 0.24
TASK2Box (5D)0.94 ± 0.030.39 ± 0.22
MLP0.88 ± 0.060.31 ± 0.18
Linear0.75 ± 0.110.40 ± 0.24
Random0.05 ± 0.140.15 ± 0.07
", + "image_path": "669b26aacdbb98247c474ac6370fe5131f2ddbd5ad76496991c401027965f124.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 326, + 545, + 370 + ], + "lines": [ + { + "bbox": [ + 304, + 326, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 545, + 370 + ], + "type": "text", + "content": "Table 3. Spearman Correlation and Standard Deviation between Predicted and Ground Truth Task Affinities on Taskonomy. Our method shows higher correlation with the task affinities compared to the baseline. Attribute-based embeddings were used." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 385, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 566 + ], + "type": "text", + "content": "dict dataset embeddings. A constraint on the box sizes was added to reflect information about the dataset sizes. Fig. 5 displays the results of TASK2Box on 131 datasets from Hugging Face, where similar datasets, such as sentiments and documents, are shown to overlap. This approach allows for the analysis and visualization of dataset variations and similarities even with only samples of images and labels. Simultaneously, the sizes of various datasets can be visualized as box sizes – with larger datasets containing more samples depicted as larger boxes. This tool enables computer vision practitioners to see how their datasets compare with other existing datasets. Beyond visualizing the variation of available datasets (based on the number of clusters), it can expedite the process of finding suitable data sources by examining embedding overlaps." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 574, + 432, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 432, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 432, + 586 + ], + "type": "text", + "content": "5.4. Analysis of TASK2BOX" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 593, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 616 + ], + "type": "text", + "content": "We discuss properties of TASK2Box below. Additional insights are also included in Appendix ??" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Using a Box Prior Improves Performance. While TASK2Box was trained with a 3-layer MLP, it achieved significantly better performance than the MLP baseline without a box prior. Tables 1, 2, and 3 demonstrate our method outperforming the baselines. Representing each dataset as an entity with shape and volume (such as a box), instead of solely learning relationships from embeddings, proves effective for generalization on unseen relationships. This" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28833" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 71, + 527, + 238 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 527, + 238 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 527, + 238 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 527, + 238 + ], + "type": "image", + "image_path": "3d80c9536d589426431a0c301c60fab3dc0aaa4351d9b6b744451ba50d908dcb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 242, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 242, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 547, + 285 + ], + "type": "text", + "content": "Figure 5. Visualizing Image Classification Datasets in Hugging Face. The sample data points annotated on the highlighted datasets show that common tasks overlap with each other (e.g., sentiment classification and document classification datasets). Although labels could slightly differ between datasets, TASK2BOX can infer the level of similarity and represent it as the amount of overlap. The embedding size (box area) also shows the number of available data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 90, + 299, + 245, + 416 + ], + "blocks": [ + { + "bbox": [ + 90, + 299, + 245, + 416 + ], + "lines": [ + { + "bbox": [ + 90, + 299, + 245, + 416 + ], + "spans": [ + { + "bbox": [ + 90, + 299, + 245, + 416 + ], + "type": "image", + "image_path": "e26db90d6b8abfefdd07cdbae2b735cee39eca4ca7ced8983502cb5aca2f7875.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 421, + 288, + 455 + ], + "lines": [ + { + "bbox": [ + 46, + 421, + 288, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 288, + 455 + ], + "type": "text", + "content": "Figure 6. Effect of TASK2Box Embedding Dimension on the Accuracy of Predicting Task Relations. As the dimension increases, the performance generally increases." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "text", + "content": "improvement could be attributed to the explicit modeling of relationships through physical shapes, which enforces consistency: if " + }, + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "inline_equation", + "content": "z_{i} \\subset z_{j}" + }, + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "inline_equation", + "content": "z_{j}" + }, + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "text", + "content": " has no overlap with " + }, + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "inline_equation", + "content": "z_{k}" + }, + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "text", + "content": ", then having a physical representation ensures that " + }, + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "inline_equation", + "content": "z_{i} \\cap z_{k} = \\emptyset" + }, + { + "bbox": [ + 46, + 469, + 287, + 518 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 519, + 288, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 288, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 288, + 663 + ], + "type": "text", + "content": "TASK2BOX can Represent Relations in Varying Dimensions. Fig. 6 illustrates the performance of TASK2BOX as the box dimension increases. Representing relationships in two dimensions results in easily interpretable embeddings; however, modeling complex task relationships could benefit from expansion to higher dimensions. Projecting to higher dimensions may yield even better performance, as relationships between datasets become more accurately represented, given the model's increased representation capacity. Nonetheless, as the dimensionality further increases, learning and visualizing the embedding space also becomes more challenging." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Using Boxes to Represent Tasks Enables Effective Calculations on Embeddings. Boxes offer the advantage of being closed under intersection, meaning the intersection of two boxes results in another box, a property not shared by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 299, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 299, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 305, + 299, + 545, + 348 + ], + "type": "text", + "content": "circles or ellipses. Among region-based embeddings that utilize geometric objects (e.g., cones, disks, and boxes) to represent entities, operations involving boxes are the most straightforward to calculate [10, 38]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 361, + 378, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 378, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 378, + 373 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 381, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 545, + 514 + ], + "type": "text", + "content": "We present a novel method that learns low-dimensional, interpretable embeddings of various task relationships. With box representations, we demonstrate that asymmetric relationships, such as task affinities and hierarchies, can be accurately modeled using various base representations. While CLIP embeddings have been shown to outperform FIM when integrated with our model, future work could investigate how CLIP might be adapted to modalities beyond text and images. Attribute-based features offer a viable alternative, and can be extracted from from datasheets using natural language processing techniques." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "text", + "content": "The distinct properties of TASK2BOX were analyzed, revealing its ability to perform effectively across varying dimensions, and to model and visualize overlaps among public classification datasets on Hugging Face. This could enable computer vision practitioners to assess dataset utility for a task in hand. Although our model successfully represents task relationships, it does not incorporate information about optimal training procedures and model architecture. Future work could explore the inclusion of additional information for a more detailed understanding of the task space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by awards from the National Science Foundation (2329927 and 1749833) and the NASA AIST program. The experiments were performed on the University of Massachusetts GPU cluster funded by the Mass. Technology Collaborative." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28834" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Hugging Face datasets. https://huggingface.co/datasets?taskCategories " + }, + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "inline_equation", + "content": "\\equiv" + }, + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": " taskCategories:image-classification.2,7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 167 + ], + "type": "text", + "content": "[2] Ralph Abboud, Ismail Ceylan, Thomas Lukasiewicz, and Tommaso Salvatori. Boxe: A box embedding model for knowledge base completion. Advances in Neural Information Processing Systems, 33:9649-9661, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 288, + 224 + ], + "type": "text", + "content": "[3] Rabab Abdelfattah, Qing Guo, Xiaoguang Li, Xiaofeng Wang, and Song Wang. Cdul: Clip-driven unsupervised learning for multi-label image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1348-1357, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 225, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 288, + 290 + ], + "type": "text", + "content": "[4] Alessandro Achille, Michael Lam, Rahul Tewari, Avinash Ravichandran, Subhransu Maji, Charless C Fowlkes, Stefano Soatto, and Pietro Perona. Task2vec: Task embedding for meta-learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6430-6439, 2019. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 292, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 288, + 334 + ], + "type": "text", + "content": "[5] Sanjeev Arora, Wei Hu, and Pravesh K. Kothari. An analysis of the t-sne algorithm for data visualization. In Proceedings of the 31st Conference On Learning Theory, pages 1455-1462. PMLR, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 336, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 288, + 390 + ], + "type": "text", + "content": "[6] Michael Boratko, Dongxu Zhang, Nicholas Monath, Luke Vilnis, Kenneth L Clarkson, and Andrew McCallum. Capacity and bias of learned geometric embeddings for directed graphs. Advances in Neural Information Processing Systems, 34:16423-16436, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 392, + 288, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 457 + ], + "type": "text", + "content": "[7] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 459, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 459, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 459, + 288, + 534 + ], + "type": "text", + "content": "[8] Tejas Chheda, Purujit Goyal, Trang Tran, Dhruvesh Patel, Michael Boratko, Shib Sankar Dasgupta, and Andrew McCallum. Box embeddings: An open-source library for representation learning using geometric structures. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 203-211, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 536, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 536, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 53, + 536, + 288, + 579 + ], + "type": "text", + "content": "[9] Marcos V Conde and Kerem Turgutlu. Clip-art: Contrastive pre-training for fine-grained art classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3956-3960, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 645 + ], + "type": "text", + "content": "[10] Shib Dasgupta, Michael Boratko, Siddhartha Mishra, Shriya Atmakuri, Dhruvesh Patel, Xiang Li, and Andrew McCallum. Word2box: Capturing set-theoretic semantics of words using box embeddings. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 2022. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "type": "text", + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 128 + ], + "type": "text", + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations (ICLR), 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 183 + ], + "type": "text", + "content": "[13] Kshitij Dwivedi and Gemma Roig. Representation similarity analysis for efficient task taxonomy & transfer learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12387-12396, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "text", + "content": "[14] Chris Fifty, Ehsan Amid, Zhe Zhao, Tianhe Yu, Rohan Anil, and Chelsea Finn. Efficiently identifying task groupings for multi-task learning. Advances in Neural Information Processing Systems, 34:27503-27516, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 230, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 274 + ], + "type": "text", + "content": "[15] Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé Iii, and Kate Crawford. Datasheets for datasets. Communications of the ACM, 64(12):86-92, 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "text", + "content": "[16] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 322, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 322, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 322, + 545, + 365 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 366, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 420 + ], + "type": "text", + "content": "[18] Wenbin He, Suphanut Jamonnak, Liang Gou, and Liu Ren. Clip-s4: Language-guided self-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11207-11216, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 422, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 487 + ], + "type": "text", + "content": "[19] EunJeong Hwang, Jay-Yoon Lee, Tianyi Yang, Dhruvesh Patel, Dongxu Zhang, and Andrew McCallum. Event-event relation extraction using probabilistic box embedding. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 235–244, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "type": "text", + "content": "[20] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 546, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 545, + 590 + ], + "type": "text", + "content": "[21] Zhuoliang Kang, Kristen Grauman, and Fei Sha. Learning with whom to share in multi-task feature learning. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pages 521-528, 2011. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "type": "text", + "content": "[22] Ryo Karakida, Shotaro Akaho, and Shun-ichi Amari. Universal statistics of fisher information in deep neural networks: Mean field approach. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 1032–1041. PMLR, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 713 + ], + "type": "text", + "content": "[23] Zhengfeng Lai, Noranart Vesdapunt, Ning Zhou, Jun Wu, Cong Phuoc Huynh, Xuelu Li, Kah Kuen Fu, and Chen-Nee Chuah. Padclip: Pseudo-labeling with adaptive debiasing in clip for unsupervised domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16155-16165, 2023. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28835" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "text", + "content": "[24] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7061-7070, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "type": "text", + "content": "[25] Zhibin Liao, Tom Drummond, Ian Reid, and Gustavo Carneiro. Approximate fisher information matrix to characterize the training of deep neural networks. IEEE transactions on pattern analysis and machine intelligence, 42(1): 15-26, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 197, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 262 + ], + "type": "text", + "content": "[26] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15305-15314, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 263, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 288, + 319 + ], + "type": "text", + "content": "[27] Shaoteng Liu, Jingjing Chen, Liangming Pan, Chong-Wah Ngo, Tat-Seng Chua, and Yu-Gang Jiang. Hyperbolic visual embedding learning for zero-shot recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 319, + 288, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 288, + 362 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 288, + 362 + ], + "type": "text", + "content": "[28] Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. Umap: Uniform manifold approximation and projection. Journal of Open Source Software, 3(29):861, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 365, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 397 + ], + "type": "text", + "content": "[29] Guy Melancon and Ivan Herman. Circular drawings of rooted trees. CWI (Centre for Mathematics and Computer Science), 1998. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 421 + ], + "type": "text", + "content": "[30] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 487 + ], + "type": "text", + "content": "[31] Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. Model cards for model reporting. In Proceedings of the conference on fairness, accountability, and transparency, pages 220-229, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 488, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 288, + 521 + ], + "type": "text", + "content": "[32] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "text", + "content": "[33] Dhruvesh Patel, Shib Sankar Dasgupta, Michael Boratko, Xiang Li, Luke Vilnis, and Andrew McCallum. Representing joint hierarchies with box embeddings. In Automated Knowledge Base Construction, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "type": "text", + "content": "[34] Dhruvesh Patel, Pavitra Dangati, Jay-Yoon Lee, Michael Boratko, and Andrew McCallum. Modeling label space interactions in multi-label classification using box embeddings. In International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[35] Jeffrey Pennington and Pratik Worah. The spectrum of the fisher information matrix of a single-hidden-layer neural network. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "sion. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "text", + "content": "[37] Edward M. Reingold and John S. Tilford. Tidier drawings of trees. IEEE Transactions on software Engineering, (2): 223-228, 1981. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "text", + "content": "[38] Hongyu Ren, Weihua Hu, and Jure Leskovec. Query2box: Reasoning over knowledge graphs in vector space using box embeddings. In International Conference on Learning Representations, 2019. 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "type": "text", + "content": "[39] Roseanna W Saaty. The analytic hierarchy process—what it is and how it is used. Mathematical modelling, 9(3-5):161–176, 1987. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 206, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 261 + ], + "type": "text", + "content": "[40] Saquib Sarfraz, Marios Koulakis, Constantin Seibold, and Rainer Stiefelhagen. Hierarchical nearest neighbor graph embedding for efficient dimensionality reduction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 336-345, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 261, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 327 + ], + "type": "text", + "content": "[41] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "text", + "content": "[42] Astuti Sharma, Tarun Kalluri, and Manmohan Chandraker. Instance level affinity-based transfer for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5361-5371, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "text", + "content": "[43] Jian Tang, Jingzhou Liu, Ming Zhang, and Qiaozhu Mei. Visualizing large-scale and high-dimensional data. In Proceedings of the 25th international conference on world wide web, pages 287-297, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "type": "text", + "content": "[44] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 460, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 525 + ], + "type": "text", + "content": "[45] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 526, + 545, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 559 + ], + "type": "text", + "content": "[46] Ivan Vendrov, Ryan Kiros, Sanja Fidler, and Raquel Urtasun. Order-embeddings of images and language. International Conference on Learning Representations (ICLR), 2015. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "text", + "content": "[47] Vedit Vedit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3219-3229, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "text", + "content": "[48] Luke Vilnis, Xiang Li, Shikhar Murty, and Andrew McCallum. Probabilistic embedding of knowledge graphs with box lattice measures. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 263-272, 2018. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "type": "text", + "content": "[49] Weixin Wang, Hui Wang, Guozhong Dai, and Hongan Wang. Visualization of large hierarchical data by circle packing. In Proceedings of the SIGCHI conference on Human Factors in computing systems, pages 517-520, 2006. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28836" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 274 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[50] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. Technical Report CNS-TR-201, Caltech, 2010. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "text", + "content": "[51] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7031-7040, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 288, + 228 + ], + "type": "text", + "content": "[52] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3712-3722, 2018. 2, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 229, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 229, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 229, + 288, + 274 + ], + "type": "text", + "content": "[53] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Mind the gap: Domain gap control for single shot domain adaptation for generative adversarial networks. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28837" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/d9ed6877-0911-45bb-b575-e2ba93e2db7d_content_list.json b/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/d9ed6877-0911-45bb-b575-e2ba93e2db7d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8c4944f146d67571301d133d8c7faed58909e2ed --- /dev/null +++ b/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/d9ed6877-0911-45bb-b575-e2ba93e2db7d_content_list.json @@ -0,0 +1,1871 @@ +[ + { + "type": "text", + "text": "TeMO: Towards Text-Driven 3D Stylization for Multi-Object Meshes", + "text_level": 1, + "bbox": [ + 137, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xuying Zhang $^{1}$ , Bo-Wen Yin $^{1}$ , Yuming Chen $^{1}$ , Zheng Lin $^{1}$ , Yunheng Li $^{1}$ , Qibin Hou $^{2,1*}$ , Ming-Ming Cheng $^{2,1}$", + "bbox": [ + 86, + 179, + 952, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ VCIP, College of Computer Science, Nankai University", + "bbox": [ + 295, + 215, + 743, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2NKIARI, Shenzhen Futian", + "bbox": [ + 410, + 234, + 630, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/zhangxuying1004/TeMO", + "bbox": [ + 344, + 253, + 692, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent progress in the text-driven 3D stylization of a single object has been considerably promoted by CLIP-based methods. However, the stylization of multi-object 3D scenes is still impeded in that the image-text pairs used for pretraining CLIP mostly consist of an object. Meanwhile, the local details of multiple objects may be susceptible to omission due to the existing supervision manner primarily relying on coarse-grained contrast of image-text pairs. To overcome these challenges, we present a novel framework, dubbed TeMO, to parse multi-object 3D scenes and edit their styles under the contrast supervision at multiple levels. We first propose a Decoupled Graph Attention (DGA) module to distinguishably reinforce the features of 3D surface points. Particularly, a cross-modal graph is constructed to align the object points accurately and noun phrases decoupled from the 3D mesh and textual description. Then, we develop a Cross-Grained Contrast (CGC) supervision system, where a fine-grained loss between the words in the textual description and the randomly rendered images are constructed to complement the coarse-grained loss. Extensive experiments show that our method can synthesize high-quality stylized content and outperform the existing methods over a wide range of multi-object 3D meshes.", + "bbox": [ + 76, + 337, + 473, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 710, + 209, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D asset creation through stylization aims to synthesize stylized content on the bare meshes to conform to the given text descriptions [15, 19, 25], referring images [38, 50], or 3D shapes [34, 44]. This research plays an important role in a wide spectrum of applications, e.g., virtual/augmented reality [5, 9], gaming industries [49], and robotics [13]. Moreover, it also presents considerable potential and has attracted increasing attention in computer vision and graphics communities. Considering the ready availability and expres", + "bbox": [ + 76, + 737, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f29180e6a4141db5dce4da8a03a97063adf7a3c22a3239ab05d96c5051db29e4.jpg", + "image_caption": [ + "Bare Mesh", + "two dragons" + ], + "image_footnote": [], + "bbox": [ + 517, + 321, + 620, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/506d174053bf53c0bfb541df1933125738833fba512004468b45f7278c130c2c.jpg", + "image_caption": [ + "person & dragon", + "Figure 1. Visual comparisons between the existing 3D stylization methods (e.g., TANGO [15]) and our TeMO in multi-object scenes. For a scene with multiple objects of the same/different categories, existing methods are prone to interference between different properties of the objects, while our TeMO is able to accurately synthesize the desired stylized content for each object." + ], + "image_footnote": [], + "bbox": [ + 517, + 417, + 620, + 497 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6743cfa96f79c1bc2ff65bdc0e09206818df741a88b2a0387d97663843683260.jpg", + "image_caption": [ + "TANGO" + ], + "image_footnote": [], + "bbox": [ + 647, + 321, + 743, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e4ed212aff92a65e766545424f730e33ec26e60e64a1e03457c4e4579866b9b.jpg", + "image_caption": [ + "\"a fire dragon and an ice dragon\"", + "\"an iron man and an ice dragon\"" + ], + "image_footnote": [], + "bbox": [ + 647, + 419, + 741, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b941afa272689e4a3611654fcdf7c2a90a2e88729401116365c66555255a45ea.jpg", + "image_caption": [ + "Our TeMO", + "an ice dragon\"" + ], + "image_footnote": [], + "bbox": [ + 774, + 321, + 870, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e6ba6910df82c4de5ef970fdebb9c9ca584504fcb2021bf9297e16bdfedce50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 419, + 869, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "siveness of text prompts as well as the popularity of large-scale Contrastive Language-Image Pre-training (CLIP) [28] model, we choose to work with text-driven 3D stylization.", + "bbox": [ + 498, + 641, + 893, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed the emergence of a series of impressive works [8, 15, 23, 25], aiming to drive the advancement of text-driven 3D stylization. Existing methods usually adopt multi-layer perceptrons (MLPs) to predict the location attribute displacements of the bare mesh under the supervision of the contrastive loss in CLIP. We observe that these works focus on the stylization of a single 3D object and perform poorly on multiple objects, as shown in the second column of Fig. 1. We argue that two inherent characteristics of CLIP result in this issue: i) CLIP is mainly pretrained with image-text pairs mostly consisting of a single object; ii) CLIP loss employs global representation vectors from images and text to coarsely match these two modalities, which inevitably causes the loss of local details. More", + "bbox": [ + 496, + 689, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Qibin Hou is the corresponding author.", + "bbox": [ + 94, + 886, + 308, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19531", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "over, the key to synthesizing desired styles for multiple 3D objects lies in the parsing of such 3D scenes and the multigrained supervision for details refinement.", + "bbox": [ + 76, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To simultaneously generate stylized content for multiple 3D objects, the primary step is to achieve accurate alignment between the objects in the 3D mesh and the target text. However, existing methods employ global semantics of the text to stylize a single object, which inevitably produces noises when stylizing the objects in multi-object scenes. To overcome this challenge, we propose to parse the 3D scene by introducing a Decoupled Graph Attention (DGA) module. Specifically, all noun phrases are decoupled from the text prompt and the mesh surface points of the current view are divided into several clusters as well. Then, a cross-modal graph is constructed to establish connections of the noun phrases to their corresponding object points while distancing them from the irrelevant ones. This graph enables the accurate interaction between two interrelated modalities. Finally, the surface point features of 3D objects can be reinforced by independent cross-attention fusions with their neighboring word nodes in the graph architecture.", + "bbox": [ + 75, + 137, + 468, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, we also design a Cross-Grained Contrast (CGC) loss to perform comprehensive cross-modal supervision for the stylization of multiple objects. The goal is to guide the network to generate more stylization details for multiple 3D objects to match the target text. Our loss consists of two parts, i.e., coarse-grained contrast and fine-grained contrast. In the former part, the text prompt is regarded as sentence-level supervision, which calculates the similarity between the 2D views rendered from the stylized 3D mesh and the text prompt using the global feature vectors from the CLIP model. In the latter part, we see the text prompt from the word level, and consider the similarities between each word of the sentence and the rendered images of the view sets. To be specific, we produce the word representations of the text prompt by taking the hidden states from the text encoder of CLIP. Motivated by the recent process in video-text retrieval [22], we calculate fine-grained loss via the weighted summation of the element in similarity vectors based on the importance of each word or image.", + "bbox": [ + 75, + 411, + 468, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the well-designed DGA module and CGC loss, we propose a novel framework towards Text-Driven 3D stylization for Multi-Object Meshes, called TeMO. To validate the effectiveness of our TeMO, extensive experiments are conducted on various multi-object 3D scenes, as shown in the $3^{rd}$ column of Fig. 1. The experimental results demonstrate our TeMO is less susceptible to interference from multiple objects and can generate superior stylized assets compared with the existing 3D stylization methods.", + "bbox": [ + 76, + 700, + 467, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 94, + 838, + 421, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We present a new 3D stylization framework, called TeMO. To the best of our knowledge, it is the first attempt to parse the objects in the text and 3D meshes and", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "generate stylizations for multi-object scenes.", + "bbox": [ + 511, + 92, + 808, + 106 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a Decoupled Graph Attention (DGA) module, which constructs a graph structure to align the surface points in the multi-object mesh and the noun phrases in the text prompt.", + "- We design a Cross-Grained Contrast (CGC) loss, in which the text is contrasted with the rendered images from sentence and word levels." + ], + "bbox": [ + 500, + 106, + 890, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 227, + 640, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-Driven 3D Manipulation", + "text_level": 1, + "bbox": [ + 500, + 252, + 764, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generating or editing 3D content according to a given prompt is a long-standing objective in computer vision and graphics [2, 39, 42]. Among all forms of the prompt, the text has garnered the most conspicuous gaze due to three reasons: i) Text descriptions are readily accessible from the existing corpus; ii) Text descriptions are particularly user-friendly since they are easily modifiable and can effectively express complex concepts related to stylizations; iii) The popularity of large-scale multimodal models [16, 28] has made achieving visual-language supervision possible.", + "bbox": [ + 496, + 275, + 890, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text2Mesh [25] proposes a neural-style field network to predict the color and displacement of mesh vertices. TANGO [15] proposes to disentangle the appearance style as the spatially varying bidirectional reflectance distribution, the local geometric variation, and the lighting condition. Then, X-Mesh [23] integrates the target text guidance by utilizing text-relevant spatial and channel-wise attention during vertex feature extraction. Motivated by the remarkable progress in text-driven 2D generation [29, 31], TEXTure [30] and Text2Tex [7] incorporate a pre-trained depth-aware image diffusion model to synthesize high-resolution partial textures from multiple viewpoints progressively.", + "bbox": [ + 496, + 428, + 890, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To make full use of the priors in the pre-trained 2D text-to-image diffusion model, DreamFusion [27] introduces a Score Distillation Sample (SDS) loss to perform text-to-3D synthesis. With the help of SDS loss, Latent-NeRF [24] and Fantasia3D [8] can generate 3D shapes and appearances for 3D objects. Despite achieving impressive results, these methods focus on the stylization of a single 3D object and rarely explore multi-object scenes. CLIP-Mesh [26] attempts to generate multiple 3D objects for target text. Nevertheless, the resulting content is not satisfactory. In this paper, we parse the objects described in rendered images and text prompts, aligned by two well-designed strategies.", + "bbox": [ + 496, + 609, + 890, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Attention Mechanism", + "text_level": 1, + "bbox": [ + 500, + 801, + 702, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The concept of the attention mechanism was initially introduced in neural machine translation [1], where the weighted summation of the candidate vectors is calculated according to their importance scores. This technology has been extended to a myriad of tasks, e.g., natural language pro", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19532", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dd6b711073a82492ad1972ca100baf91667c2de0af42ff850aa75393e3da8a2f.jpg", + "image_caption": [ + "Figure 2. The overall architecture of the proposed TeMO framework. We first specify several cameras to cast rays toward the objects in the 3D mesh scene. Then, a surface point $x_{p}$ and normal $n_p$ can be attained from each ray intersected with the objects. These points and normals are fed to the attribute prediction network where the features of 3D objects are parsed and interacted with the decoupled text features via our proposed DGA module. Meanwhile, we employ a series of spherical Gaussians to represent the lighting. Finally, a differentiable SG render is adopted to render images, which are utilized to contrast with the text prompt by our designed CGC loss." + ], + "image_footnote": [], + "bbox": [ + 102, + 85, + 867, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cessing [10, 21, 37], computer vision [14, 17, 47, 54], and multi-modal learning [20, 41, 48, 53]. For instance, Transformer [37] employs the self-attention operation to establish connections between words within a sentence and utilize the cross-attention mechanism to align source and target sentences. Non-local network [40] takes the lead in introducing self-attention to computer vision and achieves great success in video understanding and object detection. ViT [11] treats an image as a sequence of patches and employs a Transformer encoder based on self-attention to perform image classification. Swin Transformer [17] introduces shifted windows to enhance the local perception ability of self-attention. More recently, X-Mesh [23] designs a text-guided dynamic attention mechanism for vertex feature extraction of a 3D object. However, this guidance only relies on a text feature vector without considering the parsing of text and 3D scenes. In this paper, the multiple objects decoupled from the target text and 3D mesh are aligned via a cross-modal graph to achieve precise guidance.", + "bbox": [ + 75, + 396, + 472, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Multi-modal Contrastive Learning", + "text_level": 1, + "bbox": [ + 76, + 695, + 382, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive learning has become an increasingly popular research topic in the multi-modal community due to its ability to align different modal representations. Based on this strategy, CLIP [28] is pre-trained on an abundance of image-text pairs, achieving great success in cross-modal supervision. TACo [45] presents a token-aware cascade contrastive learning based on the syntactic classes of words to achieve fine-grained semantic alignment in text-video retrieval. Concurrently, FILIP [46] proposes comparing the image patches with the words in the sentence. Regarding the text-driven 3D stylization, the CLIP loss, which calculates the similarity between the image and text vectors in the", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "embedding space of CLIP, is adopted by the vast majority of methods. Although achieving impressive results in stylizing a single object, these methods cannot be well adapted to scenes with multiple 3D objects. We argue an important reason for this issue is the loss of local details caused by such coarse-grained supervision. In this paper, we propose a cross-grained supervision strategy, which considers fine-grained and coarse-grained contrasts to achieve a more precise semantic alignment between rendered image and text.", + "bbox": [ + 496, + 396, + 893, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 498, + 547, + 635, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 500, + 573, + 694, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 shows the end-to-end architecture of our TeMO framework. Given a bare mesh and a text prompt containing multiple objects, the TeMO aims to synthesize stylization on the mesh to match the text descriptions. We employ a set of vertices $V \\in \\mathbb{R}^{e \\times 3}$ and faces $F \\in \\{1, \\dots, e\\}^{u \\times 1}$ to explicitly define the input triangle mesh, which is fixed throughout the training. Following TANGO [15], we disentangle the appearance style as the spatially varying bidirectional reflectance distribution function [4, 51, 52] (including diffuse, roughness, specular), the local geometric variation (normal map), and the lighting condition.", + "bbox": [ + 496, + 597, + 890, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We start by normalizing the vertex coordinates to lie inside a unit sphere. Then, we randomly sample points around the mesh using Gaussian distribution as camera positions to render images. Next, we can obtain a camera ray $R_{p} = \\{c + t\\nu_{p}\\}$ from the sampled camera position $c$ and a pixel $p$ in rendered images, where $\\nu_{p}$ is the direction of the ray. Further, ray casting [32] is used to seek out the ray and mesh's first intersection point and intersection face. Moreover, the normal $n_p\\in \\mathbb{R}^3$ of the intersection face is", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19533", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0d31f3ed4fe35e9674f671b320e8017ae10095874286c3c822b432cb2036b6c3.jpg", + "image_caption": [ + "Figure 3. Construction pipeline of the cross-modal graph architecture in our DGA module. Note that $x_{p}$ , the surface point of the 3D objects, and $w_{ij}$ , the $j$ -th word in the $i$ -th noun phrase, are connected together only if they correspond to the same object." + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 883, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "employed as surface normal at the point $x_{p}\\in \\mathbb{R}^{3}$", + "bbox": [ + 75, + 338, + 406, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve multiview-consistent features, our TeMO is restricted to predicting the normal displacement as a function of the location, while allowing the color materials to be predicted as a function of both location and viewing direction. Therefore, our TeMO represented as MLPs includes two branches, i.e., normal branch $f_{n}(\\cdot)$ and reflectance branch $f_{r}(\\cdot)$ . Specifically, the former is utilized as the prediction of normal offset on the point $x_{p}$ , and the latter is designed to predict surface reflectance coefficients of the material at the location $x_{p}$ , i.e., diffuse, roughness, and specular. To synthesize high-frequency details, we also apply the Fourier positional encoding [35] to every input. In addition, the spherical Gaussian is employed to represent each light intensity $L_{i}(\\cdot)$ due to its closed-form nature and analytical solution. Based on the attained geometric and appearance components, each pixel color in the rendered image can be calculated by a hemisphere renderer [15]:", + "bbox": [ + 73, + 354, + 473, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p} \\left(\\nu_ {p}, x _ {p}, n _ {p}\\right) = \\int_ {\\Omega} L _ {i} \\left(w _ {i}\\right) f _ {r} \\left(\\nu_ {p}, w _ {i}, x _ {p}\\right) \\left(w _ {i} \\cdot \\hat {n} _ {p}\\right) d w _ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 616, + 468, + 647 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {n} _ {p} = n _ {p} + f _ {n} \\left(x _ {p}, n _ {p}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 651, + 468, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Omega = \\{w_{i}:w_{i}\\cdot \\hat{n}_{p}\\geq 0\\}$ represents the hemisphere, $w_{i}$ is the incident light direction, and $\\hat{n}_p$ is the estimated normal on surface point $x_{p}$ .", + "bbox": [ + 76, + 672, + 468, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Decoupled Graph Attention", + "text_level": 1, + "bbox": [ + 76, + 726, + 326, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve text-drive stylization for multiple 3D objects, the key issue that needs to be solved is the accurate alignment between the objects described in the text and those in the meshes. X-Mesh [23] has incorporated text-guided dynamic linear layers, in which the global representation vector of the target object in the text is utilized as guidance to acquire text-aware vertex features. Nevertheless, the global vector containing information about multiple objects is prone to mutual interference and produces semantic noises during guidance for multi-object scenes.", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this challenge, we propose to parse the objects in the text and mesh. We first extract the noun phrases modified by adjectives or prepositional phrases from the text using the NLTK tools [3]. Then, we employ the Gaussian Mixture Model (GMM) [56] to cluster the intersection point set $\\{x_{1},\\dots,x_{p},\\dots\\}$ of the current ray and the mesh. Meanwhile, we can obtain a binary map of objects in the current view based on whether the ray intersects with the mesh. Further, we can decouple the objects in the binary map according to the clustered points and acquire several binary maps of individual objects. Based on the disentangled noun phrases and binary maps for multiple objects, we can match the correct pairs by their semantic similarities. As a result, the objects described in the text are aligned with their corresponding objects in the mesh, which are utilized to construct a cross-modal graph $G = (\\mathcal{V},\\mathcal{E})$ , as shown in Fig. 3. To be specific, all surface point features and word features are considered as independent nodes to form the node set $\\mathcal{V}$ . For the edge set $\\mathcal{V}$ , the link between the surface point node and the word node will be built if the semantic objects they belong to are the same.", + "bbox": [ + 496, + 338, + 893, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Under the setting of this cross-modal graph, we can individually perform cross-attention between the surface point nodes and their neighboring word nodes, where the parsed surface point features are used as queries, and the parsed text features serve as keys and values. The enhancement of the surface point node $v_{i} \\in \\mathbb{R}^{dim}$ can be formulated as:", + "bbox": [ + 496, + 655, + 893, + 747 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {v} _ {i} = \\sum_ {v _ {j} \\in \\operatorname {A d j} \\left(v _ {i}\\right)} \\alpha_ {i j} \\operatorname {L i n e a r} \\left(v _ {j}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 753, + 890, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i j} = \\frac {e ^ {\\mathcal {W} _ {i j}}}{\\sum_ {v _ {j} \\in \\operatorname {A d j} (v _ {i})} e \\mathcal {W} _ {i j}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 790, + 890, + 828 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} _ {i j} = \\frac {\\operatorname {L i n e a r} \\left(v _ {i}\\right) \\operatorname {L i n e a r} \\left(v _ {j}\\right) ^ {\\mathrm {T}}}{\\sqrt {d _ {l}}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 829, + 890, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{Adj}(v_i)$ is the adjacency nodes of $v_{i}$ and $\\mathrm{Linear}(\\cdot)$ represents a linear transformation. With this attention", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19534", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mechanism, the surface point features of different objects in the mesh can be distinguishably reinforced under the guidance of the word features in the parsed text.", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Cross-Grained Contrast Supervision", + "text_level": 1, + "bbox": [ + 76, + 150, + 395, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To guide the optimization of the neural network for 3D stylization, the first step is to render the stylized 3D mesh from multiple 2D views. Most existing methods usually employ the visual encoder and text encoder of CLIP [28] to extract global feature vectors for the rendered image and target text, respectively, which are contrasted to perform cross-modal supervision via cosine similarity:", + "bbox": [ + 76, + 174, + 468, + 280 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o a r s e}} = - \\frac {\\mathcal {F} _ {I} \\cdot \\mathcal {F} _ {T}}{\\| \\mathcal {F} _ {I} \\| _ {2} \\| \\mathcal {F} _ {T} \\| _ {2}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 292, + 468, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_I\\in \\mathbb{R}^{512}$ is the averaged feature vector of the images rendered from different views, $\\mathcal{F}_T\\in \\mathbb{R}^{512}$ denotes the global feature vector of the target text, and $\\| \\cdot \\| _2$ represents the Euclidean norm function.", + "bbox": [ + 76, + 340, + 468, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although achieving impressive results for stylizing a single 3D object, these methods still have limitations in multi-object scenes. Considering that a single feature vector still represents a sentence describing multiple objects, the object details may be lost in large amounts. Therefore, such a coarse-grained contrast supervision is insufficient to guide the neural network in synthesizing photorealistic stylized content for multiple 3D objects.", + "bbox": [ + 75, + 402, + 468, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To solve this issue, we construct a fine-grained contrast supervision to complement the coarse-grained one. Specifically, we first calculate the correlation map, i.e., $S \\in \\mathbb{R}^{n \\times m}$ between the word features in the text and the visual features of the rendered images, which are also extracted from the text encoder and visual encoder of the CLIP:", + "bbox": [ + 75, + 525, + 468, + 614 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} = \\frac {\\mathcal {I} \\cdot \\mathcal {T} ^ {\\mathrm {T}}}{\\| \\mathcal {I} \\| _ {2} \\| \\mathcal {T} \\| _ {2}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 626, + 468, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{I} \\in \\mathbb{R}^{n \\times 512}$ represents the features of the images rendered from $n$ views, $\\mathcal{T} \\in \\mathbb{R}^{m \\times 512}$ indicates the features of $m$ words in the text. Then, we normalize the correlation matrix along the image axis and the text axis, respectively, to retrieve the text of interest and visual components. This process can be formulated as:", + "bbox": [ + 76, + 674, + 468, + 765 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {I} (i) = \\frac {\\sum_ {k = 1} ^ {m} \\mathcal {S} (i , k)}{m}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 779, + 468, + 809 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {T} (j) = \\frac {\\sum_ {k = 1} ^ {n} \\mathcal {S} (k , j)}{n}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 811, + 468, + 842 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by [22], we further calculate an image-centered fine-grained contrast score and a text-centered fine-grained contrast score by the weighted summation of the similarity", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "vectors, which can be formulated as follows:", + "bbox": [ + 500, + 90, + 795, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I} = \\sum_ {i = 1} ^ {n} \\frac {e ^ {\\mathcal {S} _ {I} (i)}}{\\sum_ {k = 1} ^ {n} e ^ {\\mathcal {S} _ {I} (k)}} \\mathcal {S} _ {I} (i), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 114, + 890, + 154 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T} = \\sum_ {j = 1} ^ {m} \\frac {e ^ {\\mathcal {S} _ {T} (j)}}{\\sum_ {k = 1} ^ {m} e ^ {\\mathcal {S} _ {T} (k)}} \\mathcal {S} _ {T} (j), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 157, + 890, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the weights are defined as the degree of correlation between the central and another modality. Finally, we adopt the average value of these two scores as the fine-grained contrast loss, which can be defined as:", + "bbox": [ + 498, + 207, + 890, + 267 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f i n e}} = - \\left(\\mathcal {L} _ {I} + \\mathcal {L} _ {T}\\right) / 2. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 277, + 890, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The coarse-grained and fine-grained contrast supervision complement each other to build a cross-grained contrast supervision system. The former is utilized to align the global semantic information of the target text with the 3D objects, and the latter is used to achieve the local semantic alignment. This loss can be defined as:", + "bbox": [ + 498, + 306, + 890, + 395 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e g c s} = \\lambda_ {c} \\mathcal {L} _ {\\text {c o a r s e}} + \\lambda_ {f} \\mathcal {L} _ {\\text {f i n e}}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 407, + 890, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{c}$ and $\\lambda_{f}$ are two hyper-parameters to balance the cross-grained and the fine-grained losses, set to 1.0 and 0.33, respectively.", + "bbox": [ + 498, + 434, + 890, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 492, + 632, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 500, + 517, + 674, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. To examine our method across a diverse set of 3D scenes, we first collect 3D object meshes from a variety of sources, i.e., COSEG [33], Thingi10K [55], Shapenet [6], Turbo Squid [36], and ModelNet [43]. Then, we randomly place several objects from the collected 3D set into a mesh using Blender. Note that we down-sample the number of meshes' vertices and faces to ensure the robustness of our TeMO to low-quality meshes and reduce the burden of GPU during the stylization. The meshes used in this paper contain an average of 79,303 faces, $16\\%$ non-manifold edges, $0.2\\%$ non-manifold vertices, and $12\\%$ boundaries.", + "bbox": [ + 498, + 547, + 890, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Following the TANGO [15] network, we adopt 3 linear layers with 256 dimensions to build the normal estimation branch. In the reflectance branch, the point features are extracted by 2 shared layers with 256 dimensions, followed by 3 exclusive layers to predict diffuse, specular, and roughness. The dimension of our DGA module is also set as 256. The word features in our DGA module are extracted from the text encoder of CLIP, and so are the ones in our CGC loss. We choose ViT-B/32 as the backbone of the pre-trained CLIP model in this paper, which is consistent with previous works [15, 23, 25]. We also process the rendered images with 2D augmentation strategies [12, 15]", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19535", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/66440921bdc8aa694a70f1e5f85db2247d6df544ff841f8c835d1c9c00ad1a31.jpg", + "image_caption": [ + "person & dragon" + ], + "image_footnote": [], + "bbox": [ + 96, + 84, + 223, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/de78b7850f6a7c423f5aa2f33e5a1646f59043805612cb42960574f3532034e9.jpg", + "image_caption": [ + "Text Prompt: \"an iron man and an ice dragon\"" + ], + "image_footnote": [], + "bbox": [ + 279, + 85, + 545, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56676eea3e4236823cdae3cad31bce5ef3a12267989770955f7b7d0d5fb83b16.jpg", + "image_caption": [ + "cat & horse" + ], + "image_footnote": [], + "bbox": [ + 102, + 199, + 200, + 295 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/accb7cd2f28b76c881ed1c44e205f8996ddc39de04e544d1c5a645a18debaf6d.jpg", + "image_caption": [ + "Text Prompt: \"a Garfield cat and a brown horse\"" + ], + "image_footnote": [], + "bbox": [ + 287, + 199, + 542, + 295 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8388728499c4e0e4734ff31c97ff5efb09b85a509b990e55cefe575168e4b1d3.jpg", + "image_caption": [ + "vase & candle", + "Figure 4. Given the same bare mesh, our TeMO produces various stylized contents for multi-object scenes to conform to the text prompts." + ], + "image_footnote": [], + "bbox": [ + 112, + 316, + 218, + 414 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dae3661b5d0f97498c902aeacc951df596ea72da090516aea6f2866b0426631d.jpg", + "image_caption": [ + "Text Prompt: \"a cactus vase and a silver candle\"" + ], + "image_footnote": [], + "bbox": [ + 287, + 316, + 532, + 414 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/00a50e88bd42ccfb55b26c012674f13ba577db6c468419f652c7224183570c00.jpg", + "image_caption": [ + "Text Prompt: \"a superman and a fire dragon\"" + ], + "image_footnote": [], + "bbox": [ + 609, + 85, + 870, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dcb6da02a32a4a26a4d912c5e26e6b3abfc385a625041cd82849f0dde536c7a9.jpg", + "image_caption": [ + "Text Prompt: \"a ginger cat and an astronaut horse\"" + ], + "image_footnote": [], + "bbox": [ + 606, + 200, + 857, + 295 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/052f030a88a0ac9f058396d72655a81eae0d4d2cb3470e2b931253e1f146b67f.jpg", + "image_caption": [ + "Text Prompt: \"a wicker vase and a candle in jeans\"" + ], + "image_footnote": [], + "bbox": [ + 612, + 316, + 857, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "before feeding them into the pre-trained CLIP model. Our TeMO model is optimized with the AdamW [18] strategy for 1500 iterations, where the learning rate is initialized to $5 \\times 10^{-4}$ and decayed by 0.7 every 500 iterations. The entire training process takes about 10 minutes on a single NVIDIA RTX 3090 GPU.", + "bbox": [ + 75, + 484, + 468, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 584, + 287, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct visualization experiments on a wide spectrum of multi-object scenes to verify the effectiveness of our TeMO. However, we observe that the 3D symmetry prior used widely in previous works [15, 25] can cause interference between different parts during the stylization process of multiple objects. We argue that the multiple objects of the meshes used in this paper are randomly placed to simulate a real 3D scene rather than along the $z$ -axis. To avoid this issue, we remove this prior in our TeMO and previous methods involved in the comparison.", + "bbox": [ + 75, + 607, + 470, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Neural Stylization and Controls. We present the stylization results of our TeMO driven by different text prompts for the same multi-object mesh in Fig. 4. As shown in the $1^{st}$ row where the 3D scene is composed of a person and a dragon, our TeMO can accurately distinguish between the person object and the dragon object and appropriately stylize different body parts for them according to the semantic roles described in each text prompt. Meanwhile, our TeMO also synthesizes desired stylizations for the 3D objects in", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the cat-horse mesh and vase-candle mesh, as shown in the $2^{nd}$ and $3^{rd}$ rows. These experimental results demonstrate that our TeMO method can generate photorealistic details with fine granularity and maintain global semantic understanding for the given multi-object 3D scene.", + "bbox": [ + 496, + 484, + 890, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparisons. We provide the visual comparisons of prediction results between our TeMO and previous pioneering works in text-driven 3D object stylization, including Text2Mesh [25], TANGO [15], and X-Mesh [23]. To ensure a fair comparison, we adopt the official implementations of these methods and also train them with the default settings without the symmetry prior. The experimental results show it is a real struggle for Text2Mesh [25] and TANGO [15] to understand the detailed semantics of the text prompt with multiple objects. As shown in the $1^{st}$ row of Fig. 5 where the 3D scene contains two objects of the same category, given a text prompt \"a fire dragon and an ice dragon\", they tend to capture the \"ice\" property, missing the \"fire\" property. For a 3D scene containing two objects of different categories, they are prone to mixing the properties of these objects, as shown in the $2^{nd}$ row where the text prompt is \"a wood vase and a brick candle\". Therefore, the stylized assets they generate for these multi-object scenes are unsatisfactory. X-Mesh generates more accurate results that align with the text prompts, as shown in the $1^{st}$ and $2^{nd}$ rows, which can be attributed to incorporating the text vector while extracting vertex features. However, it can pro", + "bbox": [ + 496, + 568, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19536", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fa8fe2140bb9db1d945a3ae5501df2a41c550a20bdd233b0458df72f74107771.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 71, + 225, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dffe4c493395d99a1da0533779e4cbae62cad93212cd00dd3a54d0f3d37ce2ac.jpg", + "image_caption": [ + "(a) Mesh: two dragons; Text Prompt: \"a fire dragon and an ice dragon\"." + ], + "image_footnote": [], + "bbox": [ + 258, + 71, + 387, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef2eba99c8b0ca505c24f32150933a2cb9b0b69abbe17f065b53c17c5917a329.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 71, + 545, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b1ebfbda641f328416a346ee6d30f54725eb6e91d692d5a5828d0af7dcac0851.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 71, + 707, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1519451221330d6b8d412058bbfda9744d2c7bfb32584a4a313c3e27859d0a4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 71, + 864, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ab4c588d5d5c274e4356975293a8389f75fa609dd9a9351cea38e304dd1fd36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 261, + 210, + 414 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ba345fe11419efb040709351da1cab36f3aa5e14ba84011484e9da63026f6db.jpg", + "image_caption": [ + "(b) Mesh: vase & candle; Text Prompt: \"a wood vase and a brick candle\"." + ], + "image_footnote": [], + "bbox": [ + 256, + 261, + 380, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad1cac078020fbcc33b971612e8f13d2af6a4702bb55f832c1fd552fb9416d64.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 261, + 534, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/95491438540621726297fe5c6c0ab4a41fc472c7052b5e463e5b0963ac058fb2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 261, + 692, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c058d4d7387ab23d6b3db25210bf878b1895b3f522acb39fbafdb62b6174107.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 261, + 852, + 414 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d805d17d418b5c86e6b1b8b73d701e32604123229d28b61bc3606edd2a8fb838.jpg", + "image_caption": [ + "(c) Mesh: person & dragon & whale; Text Prompt: \"a superman, a fire dragon, and an ice whale\"." + ], + "image_footnote": [], + "bbox": [ + 89, + 434, + 241, + 573 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0ea2e3d9fefb5c4cf0477dd36593bb1fe7db51ff4540aa3165e3cbcaaf439544.jpg", + "image_caption": [ + "Figure 5. Visual comparisons of our TeMO with previous text-driven 3D stylization methods on several multi-object scenes, including two objects of the same or different categories, and three different objects." + ], + "image_footnote": [], + "bbox": [ + 254, + 434, + 406, + 573 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/335c60e6fc3e0894e7eb0ddd01097f70d14291833281797915760915310ec445.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 434, + 566, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c9660c91ddb4e36bad560489363f0d8da85f9855c056141c4e05ec3a88521e22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 434, + 723, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5862fc8ad0f9f35f165b3012959025f2649a045bafdbfdf344e8684b516def23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 434, + 883, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "duce semantic noises due to its utilization of the text vector containing attributes of multiple objects to process all vertex features. With an increasing number of objects, it will also encounter challenges related to comprehending text details and the alignment between the text and 3D objects. As shown in the $3^{rd}$ row, this method still fails to generate stylized assets without mixed properties. In contrast, our TeMO equipped with 3D scene parsing and multi-grained supervision, is able to generate photorealistic stylized content for each object in these 3D scenes to conform to the descriptions in the text prompts.", + "bbox": [ + 75, + 651, + 472, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 825, + 299, + 842 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Objective Metric. We adopt the CLIP score to objectively evaluate the semantic alignment achieved by our TeMO and recent 3D stylization methods. Specifically, 8 views spaced", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$45^{\\circ}$ around the stylized meshes are chosen to obtain the rendered 2D images. Then, the visual objects are compared with the textual objects in CLIP's embedding space via the cosine function. As shown in the $2^{nd}$ column of Tab. 1, our TeMO surpasses previous methods by a large margin. These results demonstrate the superiority of our TeMO over existing methods on multi-object stylization.", + "bbox": [ + 496, + 651, + 893, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. We further conduct a user study to evaluate these 3D stylization methods subjectively. We randomly select 10 mesh-text pairs and recruit 60 users to evaluate the quality of the stylization assets generated by our TeMO and previous methods. Particularly, the participants include experts in the field and individuals without specific background knowledge. Moreover, each of them will be asked three questions [25]: (Q1) \"How natural is the output results?\"(Q2) \"How well does the output match the original", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19537", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9047c9124b16781a232ba564a837fe67ffcd09d9dc577149c7e9ab75ee52071b.jpg", + "image_caption": [ + "Input Mesh" + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 230, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/44e2474405cf9ffb61160181f4083aa5bbedbc47695458f432c68ef65e7392c7.jpg", + "image_caption": [ + "Baseline" + ], + "image_footnote": [], + "bbox": [ + 292, + 88, + 416, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/05c809defa7cd9ba5d3381bfcc56e23f3c6eeee6703d49dbba6bf041ceb46982.jpg", + "image_caption": [ + "Baseline + DGA module" + ], + "image_footnote": [], + "bbox": [ + 444, + 88, + 566, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1f6955df3bd31187a4b436fabd7a2eb598d7630c3bc82ad2ca2968cafa777b99.jpg", + "image_caption": [ + "Baseline + CGC loss" + ], + "image_footnote": [], + "bbox": [ + 596, + 88, + 717, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/37db2236ad4868c56b8561f82bb54e62e56e161fd004edd0e38873e1c2269ef2.jpg", + "image_caption": [ + "Our TeMO", + "Figure 6. Ablation experiments on the proposed designs of our TeMO. Mesh: two dragons; Text Prompt: \"a fire dragon and an ice dragon\"." + ], + "image_footnote": [], + "bbox": [ + 743, + 88, + 867, + 189 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/85ed0290c57419d989f0e023f318e1e8cad5882b8442a487f5af38899476cad6.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons of our TeMO and previous text-driven 3D stylization methods in multi-object scenes, including an objective alignment score (0-1) and three subjective opinion scores (1-5). Note that the higher these metrics, the better the method." + ], + "table_footnote": [], + "table_body": "
AlignmentUser-Q1User-Q2User-Q3
Text2Mesh [25]0.2621.7501.5061.472
TANGO [15]0.2742.4062.4502.539
X-Mesh [23]0.2651.8391.7221.761
Our TeMO0.2853.3443.3113.261
", + "bbox": [ + 78, + 308, + 464, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "content?\" (Q3) \"How well does the output match the target style?\", and then assign a score (1-5) to them. We report the mean opinion scores in parentheses for each factor averaged across all style outputs. As shown in Tab. 1, our TeMO still outperforms other methods across all questions. Therefore, the 3D assets generated by our method are more in line with people's understanding of the text prompts.", + "bbox": [ + 75, + 419, + 467, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 535, + 241, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To verify the effectiveness of the proposed designs in our TeMO, we conduct ablation studies by gradually adding them to our baseline model, i.e., TANGO [15]. We chose the two-dragon mesh with the text prompt \"a fire dragon and an ice dragon\", and the experimental results are shown in Fig. 6. Compared to the baseline model, introducing our DGA module enables the model to distinguish two dragons, yet it falls short in endowing them with precise texture details. Meanwhile, incorporating our CGC loss facilitates the model to capture more semantic details, e.g., \"fire\" and \"ice\". Nevertheless, it fails to distinguish the two objects. It is noteworthy that the model equipped with these two designs together is not only capable of accurately distinguishing between two objects but can also synthesize high-quality texture details for them. These experiments indicate that our DGA module and CGC loss can effectively assist the model in generating desired stylized content for multiple 3D objects to conform to the target text.", + "bbox": [ + 75, + 558, + 468, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitation and Future Work", + "text_level": 1, + "bbox": [ + 76, + 844, + 341, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite achieving excellent results on text-driven multi-object stylization, our TeMO framework still has a few lim", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "itations, which can also facilitate future research:", + "bbox": [ + 498, + 250, + 823, + 263 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) 3D Symmetry Prior. As stated in Sec. 4.2, our TeMO fails to incorporate 3D symmetry prior, whose important role has been demonstrated by Text2Mesh [25] in promoting style consistency of a single object. To generate more photorealistic stylization assets for multi-object scenes, it will be valuable to calculate symmetry planes for each object and apply symmetry priors to them.", + "2) Diffusion Model. We observe that current diffusion technologies struggle to generate multi-object images according to the text prompt, which hinders the application of diffusion-based stylization methods in multi-object 3D scenes. We argue it would be interesting to extend the concept of scene parsing to the diffusion models for the release of their potential in multi-object editing or generation." + ], + "bbox": [ + 496, + 271, + 890, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 505, + 617, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present TeMO, an innovative framework proposing scene parsing and multi-grained cross-modal supervision to achieve text-driven multi-object 3D stylization for the first time. Specifically, we first develop a DGA module to precisely align the objects in the 3D mesh and the text prompt and enhance the 3D point features with the word features belonging to the same object as them. Then, we design a CGC loss, in which the fine-grained loss at the local level and coarse-grained contrast loss at the global level are both constructed and complement each other. Further, extensive experiments are conducted to demonstrate the effectiveness and superiority of our methods over the existing methods among a wide range of multi-object 3D scenes. We believe it is promising to achieve content editing of multiple objects in 3D scenes simultaneously, and we hope the scene-parsing perspective provided by the proposed TeMO framework will inspire future works.", + "bbox": [ + 496, + 531, + 890, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This research was supported by NSFC (NO. 62225604, NO. 62276145), the Fundamental Research Funds for the Central Universities (Nankai University, 070-63223049), CAST through Young Elite Scientist Sponsorship Program (No. YESS20210377). Computations were supported by the Supercomputing Center of Nankai University (NKSC).", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19538", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In ICLR, 2014. 2", + "[2] Guirong Bai, Shizhu He, Kang Liu, and Jun Zhao. Example-guided stylized response generation in zero-shot setting. Science China Information Sciences, 2022. 2", + "[3] Edward Loper Bird, Steven and Ewan Klein. Natural language processing with python. O'Reilly Media Inc, 2009. 4", + "[4] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In IEEE ICCV, 2021. 3", + "[5] Arthur Caetano and Misha Sra. Arfy: A pipeline for adapting 3d scenes to augmented reality. In ACM UIST, 2022. 1", + "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5", + "[7] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In IEEE ICCV, 2023. 2", + "[8] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In IEEE ICCV, 2023. 1, 2", + "[9] Shaoyu Chen, Budmonde Duinkharjav, Xin Sun, Li-Yi Wei, Stefano Petrangeli, Jose Echevarria, Claudio Silva, and Qi Sun. Instant reality: Gaze-contingent perceptual optimization for 3d virtual reality streaming. IEEE TVCG, 2022. 1", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 3", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3", + "[12] Kevin Frans, Lisa Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. Advances in Neural Information Processing Systems, 2022. 5", + "[13] Mengdi Han, Xiaogang Guo, Xuexian Chen, Cunman Liang, Hangbo Zhao, Qihui Zhang, Wubin Bai, Fan Zhang, Heming Wei, Changsheng Wu, et al. Submillimeter-scale multitematerial terrestrial robots. Science Robotics, 2022. 1", + "[14] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In IEEE CVPR, 2018. 3", + "[15] Jiabao Lei, Yabin Zhang, Kui Jia, et al. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[16] Zheng Li, Xiang Li, Xinyi Fu, Xin Zhang, Weiqiang Wang, Shuo Chen, and Jian Yang. Promptkd: Unsupervised prompt distillation for vision-language models. arXiv preprint arXiv:2403.02781, 2024. 2" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE ICCV, 2021. 3", + "[18] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 6", + "[19] Haoyu Lu, Yuqi Huo, Mingyu Ding, Nanyi Fei, and Zhiwu Lu. Cross-modal contrastive learning for generalizable and efficient image-text retrieval. Machine Intelligence Research, 2023. 1", + "[20] Yunpeng Luo, Jiayi Ji, Xiaoshuai Sun, Liujuan Cao, Yongjian Wu, Feiyue Huang, Chia-Wen Lin, and Rongrong Ji. Dual-level collaborative transformer for image captioning. In AAAI, 2021. 3", + "[21] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. In ACL EMNLP, 2015. 3", + "[22] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In ACM MM, 2022. 2, 5", + "[23] Yiwei Ma, Xiaojing Zhang, Xiaoshuai Sun, Jiayi Ji, Haowei Wang, Guannan Jiang, Weilin Zhuang, and Rongrong Ji. X-mesh: Towards fast and accurate text-driven 3d stylization via dynamic textual guidance. In IEEE ICCV, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[24] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In IEEE CVPR, 2023. 2", + "[25] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In IEEE CVPR, 2022. 1, 2, 5, 6, 7, 8", + "[26] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In ACM SIGGRAPH, 2022. 2", + "[27] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2", + "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 5", + "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 2", + "[30] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH, 2023. 2", + "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE CVPR, 2022. 2", + "[32] Scott D Roth. Ray casting for modeling solids. Computer graphics and image processing, 1982. 3", + "[33] Oana Sidi, Oliver Van Kaick, Yanir Kleiman, Hao Zhang, and Daniel Cohen-Or. Unsupervised co-segmentation of a set of shapes via descriptor-space spectral clustering. In" + ], + "bbox": [ + 501, + 92, + 890, + 888 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19539", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ACM SIGGRAPH, 2011. 5", + "[34] Jia-Mu Sun, Tong Wu, and Lin Gao. Recent advances in implicit representation-based 3d shape generation. Visual Intelligence, 2024. 1", + "[35] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 2020. 4", + "[36] TurboSquid. Turbosquid 3d model repository. In https://www.turbosquid.com/, 2021.5", + "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3", + "[38] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In IEEE CVPR, 2022. 1", + "[39] Wei Wang, Qialei Dong, and Zhanyi Hu. Asprr: active single-image piecewise planar 3d reconstruction based on geometric priors. Science China Information Sciences, 2023. 2", + "[40] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In IEEE CVPR, 2018. 3", + "[41] Mingrui Wu, Xuying Zhang, Xiaoshuai Sun, Yiyi Zhou, Chao Chen, Jiaxin Gu, Xing Sun, and Rongrong Ji. Difnet: Boosting visual information flow for image captioning. In IEEE CVPR, 2022. 3", + "[42] Ruiqi Wu, Liangyu Chen, Tong Yang, Chunle Guo, Chongyi Li, and Xiangyu Zhang. Lamp: Learn a motion pattern for few-shot-based video generation. arXiv preprint arXiv:2310.10769, 2023. 2", + "[43] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In IEEE CVPR, 2015. 5", + "[44] Qun-Ce Xu, Tai-Jiang Mu, and Yong-Liang Yang. A survey of deep learning-based 3d shape generation. Computational Visual Media, 2023. 1", + "[45] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In IEEE ICCV, 2021. 3", + "[46] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. In ICLR, 2022. 3", + "[47] Bowen Yin, Xuying Zhang, Qibin Hou, Bo-Yuan Sun, DengPing Fan, and Luc Van Gool. Camoformer: Masked separable attention for camouflaged object detection. arXiv preprint arXiv:2212.06570, 2022. 3", + "[48] Bowen Yin, Xuying Zhang, Zhongyu Li, Li Liu, Ming-Ming Cheng, and Qibin Hou. Dformer: Rethinking rgbd representation learning for semantic segmentation. In ICLR, 2024. 3", + "[49] Bo Zhang, Lizbeth Goodman, and Xiaoqing Gu. Novel 3d contextual interactive games on a gamified virtual environment support cultural learning through collaboration among" + ], + "bbox": [ + 78, + 92, + 468, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "intercultural students. SAGE Open, 2022. 1", + "[50] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In ECCV, 2022. 1", + "[51] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In IEEE CVPR, 2021. 3", + "[52] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG, 2021. 3", + "[53] Xuying Zhang, Xiaoshuai Sun, Yunpeng Luo, Jiayi Ji, Yiyi Zhou, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Rstnet: Captioning with adaptive attention on visual and non-visual words. In IEEE CVPR, 2021. 3", + "[54] Xuying Zhang, Bowen Yin, Zheng Lin, Qibin Hou, DengPing Fan, and Ming-Ming Cheng. Referring camouflaged object detection. arXiv preprint arXiv:2306.07532, 2023. 3", + "[55] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models. arXiv preprint arXiv:1605.04797, 2016. 5", + "[56] Zoran Zivkovic. Improved adaptive gaussian mixture model for background subtraction. In IEEE ICPR, 2004. 4" + ], + "bbox": [ + 501, + 92, + 890, + 415 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19540", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/d9ed6877-0911-45bb-b575-e2ba93e2db7d_model.json b/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/d9ed6877-0911-45bb-b575-e2ba93e2db7d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8627c7b3a018a14dddcba86b6afeb3741e125d59 --- /dev/null +++ b/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/d9ed6877-0911-45bb-b575-e2ba93e2db7d_model.json @@ -0,0 +1,2684 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "TeMO: Towards Text-Driven 3D Stylization for Multi-Object Meshes" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.18, + 0.953, + 0.201 + ], + "angle": 0, + "content": "Xuying Zhang\\(^{1}\\), Bo-Wen Yin\\(^{1}\\), Yuming Chen\\(^{1}\\), Zheng Lin\\(^{1}\\), Yunheng Li\\(^{1}\\), Qibin Hou\\(^{2,1*}\\), Ming-Ming Cheng\\(^{2,1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.216, + 0.745, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)VCIP, College of Computer Science, Nankai University" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.235, + 0.632, + 0.251 + ], + "angle": 0, + "content": "2NKIARI, Shenzhen Futian" + }, + { + "type": "text", + "bbox": [ + 0.345, + 0.255, + 0.694, + 0.269 + ], + "angle": 0, + "content": "https://github.com/zhangxuying1004/TeMO" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.338, + 0.474, + 0.685 + ], + "angle": 0, + "content": "Recent progress in the text-driven 3D stylization of a single object has been considerably promoted by CLIP-based methods. However, the stylization of multi-object 3D scenes is still impeded in that the image-text pairs used for pretraining CLIP mostly consist of an object. Meanwhile, the local details of multiple objects may be susceptible to omission due to the existing supervision manner primarily relying on coarse-grained contrast of image-text pairs. To overcome these challenges, we present a novel framework, dubbed TeMO, to parse multi-object 3D scenes and edit their styles under the contrast supervision at multiple levels. We first propose a Decoupled Graph Attention (DGA) module to distinguishably reinforce the features of 3D surface points. Particularly, a cross-modal graph is constructed to align the object points accurately and noun phrases decoupled from the 3D mesh and textual description. Then, we develop a Cross-Grained Contrast (CGC) supervision system, where a fine-grained loss between the words in the textual description and the randomly rendered images are constructed to complement the coarse-grained loss. Extensive experiments show that our method can synthesize high-quality stylized content and outperform the existing methods over a wide range of multi-object 3D meshes." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.712, + 0.21, + 0.727 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.738, + 0.47, + 0.875 + ], + "angle": 0, + "content": "3D asset creation through stylization aims to synthesize stylized content on the bare meshes to conform to the given text descriptions [15, 19, 25], referring images [38, 50], or 3D shapes [34, 44]. This research plays an important role in a wide spectrum of applications, e.g., virtual/augmented reality [5, 9], gaming industries [49], and robotics [13]. Moreover, it also presents considerable potential and has attracted increasing attention in computer vision and graphics communities. Considering the ready availability and expres" + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.306, + 0.605, + 0.317 + ], + "angle": 0, + "content": "Bare Mesh" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.322, + 0.622, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.403, + 0.608, + 0.416 + ], + "angle": 0, + "content": "two dragons" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.419, + 0.622, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.501, + 0.622, + 0.514 + ], + "angle": 0, + "content": "person & dragon" + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.306, + 0.724, + 0.317 + ], + "angle": 0, + "content": "TANGO" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.322, + 0.745, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.403, + 0.866, + 0.416 + ], + "angle": 0, + "content": "\"a fire dragon and an ice dragon\"" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.42, + 0.742, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.501, + 0.862, + 0.514 + ], + "angle": 0, + "content": "\"an iron man and an ice dragon\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.788, + 0.306, + 0.861, + 0.317 + ], + "angle": 0, + "content": "Our TeMO" + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.322, + 0.871, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.402, + 0.866, + 0.415 + ], + "angle": 0, + "content": "an ice dragon\"" + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.42, + 0.87, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.531, + 0.894, + 0.616 + ], + "angle": 0, + "content": "Figure 1. Visual comparisons between the existing 3D stylization methods (e.g., TANGO [15]) and our TeMO in multi-object scenes. For a scene with multiple objects of the same/different categories, existing methods are prone to interference between different properties of the objects, while our TeMO is able to accurately synthesize the desired stylized content for each object." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.642, + 0.894, + 0.687 + ], + "angle": 0, + "content": "siveness of text prompts as well as the popularity of large-scale Contrastive Language-Image Pre-training (CLIP) [28] model, we choose to work with text-driven 3D stylization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Recent years have witnessed the emergence of a series of impressive works [8, 15, 23, 25], aiming to drive the advancement of text-driven 3D stylization. Existing methods usually adopt multi-layer perceptrons (MLPs) to predict the location attribute displacements of the bare mesh under the supervision of the contrastive loss in CLIP. We observe that these works focus on the stylization of a single 3D object and perform poorly on multiple objects, as shown in the second column of Fig. 1. We argue that two inherent characteristics of CLIP result in this issue: i) CLIP is mainly pretrained with image-text pairs mostly consisting of a single object; ii) CLIP loss employs global representation vectors from images and text to coarsely match these two modalities, which inevitably causes the loss of local details. More" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.31, + 0.9 + ], + "angle": 0, + "content": "*Qibin Hou is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19531" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "over, the key to synthesizing desired styles for multiple 3D objects lies in the parsing of such 3D scenes and the multigrained supervision for details refinement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.47, + 0.412 + ], + "angle": 0, + "content": "To simultaneously generate stylized content for multiple 3D objects, the primary step is to achieve accurate alignment between the objects in the 3D mesh and the target text. However, existing methods employ global semantics of the text to stylize a single object, which inevitably produces noises when stylizing the objects in multi-object scenes. To overcome this challenge, we propose to parse the 3D scene by introducing a Decoupled Graph Attention (DGA) module. Specifically, all noun phrases are decoupled from the text prompt and the mesh surface points of the current view are divided into several clusters as well. Then, a cross-modal graph is constructed to establish connections of the noun phrases to their corresponding object points while distancing them from the irrelevant ones. This graph enables the accurate interaction between two interrelated modalities. Finally, the surface point features of 3D objects can be reinforced by independent cross-attention fusions with their neighboring word nodes in the graph architecture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.47, + 0.7 + ], + "angle": 0, + "content": "Furthermore, we also design a Cross-Grained Contrast (CGC) loss to perform comprehensive cross-modal supervision for the stylization of multiple objects. The goal is to guide the network to generate more stylization details for multiple 3D objects to match the target text. Our loss consists of two parts, i.e., coarse-grained contrast and fine-grained contrast. In the former part, the text prompt is regarded as sentence-level supervision, which calculates the similarity between the 2D views rendered from the stylized 3D mesh and the text prompt using the global feature vectors from the CLIP model. In the latter part, we see the text prompt from the word level, and consider the similarities between each word of the sentence and the rendered images of the view sets. To be specific, we produce the word representations of the text prompt by taking the hidden states from the text encoder of CLIP. Motivated by the recent process in video-text retrieval [22], we calculate fine-grained loss via the weighted summation of the element in similarity vectors based on the importance of each word or image." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.701, + 0.468, + 0.837 + ], + "angle": 0, + "content": "Based on the well-designed DGA module and CGC loss, we propose a novel framework towards Text-Driven 3D stylization for Multi-Object Meshes, called TeMO. To validate the effectiveness of our TeMO, extensive experiments are conducted on various multi-object 3D scenes, as shown in the \\(3^{rd}\\) column of Fig. 1. The experimental results demonstrate our TeMO is less susceptible to interference from multiple objects and can generate superior stylized assets compared with the existing 3D stylization methods." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.839, + 0.422, + 0.853 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- We present a new 3D stylization framework, called TeMO. To the best of our knowledge, it is the first attempt to parse the objects in the text and 3D meshes and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.093, + 0.809, + 0.107 + ], + "angle": 0, + "content": "generate stylizations for multi-object scenes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.107, + 0.892, + 0.167 + ], + "angle": 0, + "content": "- We propose a Decoupled Graph Attention (DGA) module, which constructs a graph structure to align the surface points in the multi-object mesh and the noun phrases in the text prompt." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.168, + 0.892, + 0.212 + ], + "angle": 0, + "content": "- We design a Cross-Grained Contrast (CGC) loss, in which the text is contrasted with the rendered images from sentence and word levels." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.107, + 0.892, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.228, + 0.642, + 0.243 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.253, + 0.765, + 0.269 + ], + "angle": 0, + "content": "2.1. Text-Driven 3D Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.276, + 0.892, + 0.428 + ], + "angle": 0, + "content": "Generating or editing 3D content according to a given prompt is a long-standing objective in computer vision and graphics [2, 39, 42]. Among all forms of the prompt, the text has garnered the most conspicuous gaze due to three reasons: i) Text descriptions are readily accessible from the existing corpus; ii) Text descriptions are particularly user-friendly since they are easily modifiable and can effectively express complex concepts related to stylizations; iii) The popularity of large-scale multimodal models [16, 28] has made achieving visual-language supervision possible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.609 + ], + "angle": 0, + "content": "Text2Mesh [25] proposes a neural-style field network to predict the color and displacement of mesh vertices. TANGO [15] proposes to disentangle the appearance style as the spatially varying bidirectional reflectance distribution, the local geometric variation, and the lighting condition. Then, X-Mesh [23] integrates the target text guidance by utilizing text-relevant spatial and channel-wise attention during vertex feature extraction. Motivated by the remarkable progress in text-driven 2D generation [29, 31], TEXTure [30] and Text2Tex [7] incorporate a pre-trained depth-aware image diffusion model to synthesize high-resolution partial textures from multiple viewpoints progressively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.791 + ], + "angle": 0, + "content": "To make full use of the priors in the pre-trained 2D text-to-image diffusion model, DreamFusion [27] introduces a Score Distillation Sample (SDS) loss to perform text-to-3D synthesis. With the help of SDS loss, Latent-NeRF [24] and Fantasia3D [8] can generate 3D shapes and appearances for 3D objects. Despite achieving impressive results, these methods focus on the stylization of a single 3D object and rarely explore multi-object scenes. CLIP-Mesh [26] attempts to generate multiple 3D objects for target text. Nevertheless, the resulting content is not satisfactory. In this paper, we parse the objects described in rendered images and text prompts, aligned by two well-designed strategies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.704, + 0.816 + ], + "angle": 0, + "content": "2.2. Attention Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The concept of the attention mechanism was initially introduced in neural machine translation [1], where the weighted summation of the candidate vectors is calculated according to their importance scores. This technology has been extended to a myriad of tasks, e.g., natural language pro" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19532" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.087, + 0.868, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.307, + 0.895, + 0.379 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of the proposed TeMO framework. We first specify several cameras to cast rays toward the objects in the 3D mesh scene. Then, a surface point \\( x_{p} \\) and normal \\( n_p \\) can be attained from each ray intersected with the objects. These points and normals are fed to the attribute prediction network where the features of 3D objects are parsed and interacted with the decoupled text features via our proposed DGA module. Meanwhile, we employ a series of spherical Gaussians to represent the lighting. Finally, a differentiable SG render is adopted to render images, which are utilized to contrast with the text prompt by our designed CGC loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.473, + 0.685 + ], + "angle": 0, + "content": "cessing [10, 21, 37], computer vision [14, 17, 47, 54], and multi-modal learning [20, 41, 48, 53]. For instance, Transformer [37] employs the self-attention operation to establish connections between words within a sentence and utilize the cross-attention mechanism to align source and target sentences. Non-local network [40] takes the lead in introducing self-attention to computer vision and achieves great success in video understanding and object detection. ViT [11] treats an image as a sequence of patches and employs a Transformer encoder based on self-attention to perform image classification. Swin Transformer [17] introduces shifted windows to enhance the local perception ability of self-attention. More recently, X-Mesh [23] designs a text-guided dynamic attention mechanism for vertex feature extraction of a 3D object. However, this guidance only relies on a text feature vector without considering the parsing of text and 3D scenes. In this paper, the multiple objects decoupled from the target text and 3D mesh are aligned via a cross-modal graph to achieve precise guidance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.383, + 0.712 + ], + "angle": 0, + "content": "2.3. Multi-modal Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Contrastive learning has become an increasingly popular research topic in the multi-modal community due to its ability to align different modal representations. Based on this strategy, CLIP [28] is pre-trained on an abundance of image-text pairs, achieving great success in cross-modal supervision. TACo [45] presents a token-aware cascade contrastive learning based on the syntactic classes of words to achieve fine-grained semantic alignment in text-video retrieval. Concurrently, FILIP [46] proposes comparing the image patches with the words in the sentence. Regarding the text-driven 3D stylization, the CLIP loss, which calculates the similarity between the image and text vectors in the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.895, + 0.534 + ], + "angle": 0, + "content": "embedding space of CLIP, is adopted by the vast majority of methods. Although achieving impressive results in stylizing a single object, these methods cannot be well adapted to scenes with multiple 3D objects. We argue an important reason for this issue is the loss of local details caused by such coarse-grained supervision. In this paper, we propose a cross-grained supervision strategy, which considers fine-grained and coarse-grained contrasts to achieve a more precise semantic alignment between rendered image and text." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.549, + 0.637, + 0.566 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.696, + 0.589 + ], + "angle": 0, + "content": "3.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Fig. 2 shows the end-to-end architecture of our TeMO framework. Given a bare mesh and a text prompt containing multiple objects, the TeMO aims to synthesize stylization on the mesh to match the text descriptions. We employ a set of vertices \\( V \\in \\mathbb{R}^{e \\times 3} \\) and faces \\( F \\in \\{1, \\dots, e\\}^{u \\times 1} \\) to explicitly define the input triangle mesh, which is fixed throughout the training. Following TANGO [15], we disentangle the appearance style as the spatially varying bidirectional reflectance distribution function [4, 51, 52] (including diffuse, roughness, specular), the local geometric variation (normal map), and the lighting condition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We start by normalizing the vertex coordinates to lie inside a unit sphere. Then, we randomly sample points around the mesh using Gaussian distribution as camera positions to render images. Next, we can obtain a camera ray \\( R_{p} = \\{c + t\\nu_{p}\\} \\) from the sampled camera position \\( c \\) and a pixel \\( p \\) in rendered images, where \\( \\nu_{p} \\) is the direction of the ray. Further, ray casting [32] is used to seek out the ray and mesh's first intersection point and intersection face. Moreover, the normal \\( n_p\\in \\mathbb{R}^3 \\) of the intersection face is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19533" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.088, + 0.885, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.291, + 0.893, + 0.32 + ], + "angle": 0, + "content": "Figure 3. Construction pipeline of the cross-modal graph architecture in our DGA module. Note that \\( x_{p} \\), the surface point of the 3D objects, and \\( w_{ij} \\), the \\( j \\)-th word in the \\( i \\)-th noun phrase, are connected together only if they correspond to the same object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.407, + 0.355 + ], + "angle": 0, + "content": "employed as surface normal at the point \\(x_{p}\\in \\mathbb{R}^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.355, + 0.474, + 0.611 + ], + "angle": 0, + "content": "To achieve multiview-consistent features, our TeMO is restricted to predicting the normal displacement as a function of the location, while allowing the color materials to be predicted as a function of both location and viewing direction. Therefore, our TeMO represented as MLPs includes two branches, i.e., normal branch \\( f_{n}(\\cdot) \\) and reflectance branch \\( f_{r}(\\cdot) \\). Specifically, the former is utilized as the prediction of normal offset on the point \\( x_{p} \\), and the latter is designed to predict surface reflectance coefficients of the material at the location \\( x_{p} \\), i.e., diffuse, roughness, and specular. To synthesize high-frequency details, we also apply the Fourier positional encoding [35] to every input. In addition, the spherical Gaussian is employed to represent each light intensity \\( L_{i}(\\cdot) \\) due to its closed-form nature and analytical solution. Based on the attained geometric and appearance components, each pixel color in the rendered image can be calculated by a hemisphere renderer [15]:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.617, + 0.47, + 0.648 + ], + "angle": 0, + "content": "\\[\nL _ {p} \\left(\\nu_ {p}, x _ {p}, n _ {p}\\right) = \\int_ {\\Omega} L _ {i} \\left(w _ {i}\\right) f _ {r} \\left(\\nu_ {p}, w _ {i}, x _ {p}\\right) \\left(w _ {i} \\cdot \\hat {n} _ {p}\\right) d w _ {i}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.652, + 0.469, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\hat {n} _ {p} = n _ {p} + f _ {n} \\left(x _ {p}, n _ {p}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.674, + 0.47, + 0.721 + ], + "angle": 0, + "content": "where \\(\\Omega = \\{w_{i}:w_{i}\\cdot \\hat{n}_{p}\\geq 0\\}\\) represents the hemisphere, \\(w_{i}\\) is the incident light direction, and \\(\\hat{n}_p\\) is the estimated normal on surface point \\(x_{p}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.328, + 0.743 + ], + "angle": 0, + "content": "3.2. Decoupled Graph Attention" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To achieve text-drive stylization for multiple 3D objects, the key issue that needs to be solved is the accurate alignment between the objects described in the text and those in the meshes. X-Mesh [23] has incorporated text-guided dynamic linear layers, in which the global representation vector of the target object in the text is utilized as guidance to acquire text-aware vertex features. Nevertheless, the global vector containing information about multiple objects is prone to mutual interference and produces semantic noises during guidance for multi-object scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.895, + 0.656 + ], + "angle": 0, + "content": "To address this challenge, we propose to parse the objects in the text and mesh. We first extract the noun phrases modified by adjectives or prepositional phrases from the text using the NLTK tools [3]. Then, we employ the Gaussian Mixture Model (GMM) [56] to cluster the intersection point set \\(\\{x_{1},\\dots,x_{p},\\dots\\}\\) of the current ray and the mesh. Meanwhile, we can obtain a binary map of objects in the current view based on whether the ray intersects with the mesh. Further, we can decouple the objects in the binary map according to the clustered points and acquire several binary maps of individual objects. Based on the disentangled noun phrases and binary maps for multiple objects, we can match the correct pairs by their semantic similarities. As a result, the objects described in the text are aligned with their corresponding objects in the mesh, which are utilized to construct a cross-modal graph \\(G = (\\mathcal{V},\\mathcal{E})\\), as shown in Fig. 3. To be specific, all surface point features and word features are considered as independent nodes to form the node set \\(\\mathcal{V}\\). For the edge set \\(\\mathcal{V}\\), the link between the surface point node and the word node will be built if the semantic objects they belong to are the same." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.895, + 0.748 + ], + "angle": 0, + "content": "Under the setting of this cross-modal graph, we can individually perform cross-attention between the surface point nodes and their neighboring word nodes, where the parsed surface point features are used as queries, and the parsed text features serve as keys and values. The enhancement of the surface point node \\( v_{i} \\in \\mathbb{R}^{dim} \\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.754, + 0.892, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\hat {v} _ {i} = \\sum_ {v _ {j} \\in \\operatorname {A d j} \\left(v _ {i}\\right)} \\alpha_ {i j} \\operatorname {L i n e a r} \\left(v _ {j}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.791, + 0.892, + 0.829 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i j} = \\frac {e ^ {\\mathcal {W} _ {i j}}}{\\sum_ {v _ {j} \\in \\operatorname {A d j} (v _ {i})} e \\mathcal {W} _ {i j}}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.83, + 0.892, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} _ {i j} = \\frac {\\operatorname {L i n e a r} \\left(v _ {i}\\right) \\operatorname {L i n e a r} \\left(v _ {j}\\right) ^ {\\mathrm {T}}}{\\sqrt {d _ {l}}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathrm{Adj}(v_i)\\) is the adjacency nodes of \\(v_{i}\\) and \\(\\mathrm{Linear}(\\cdot)\\) represents a linear transformation. With this attention" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19534" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "mechanism, the surface point features of different objects in the mesh can be distinguishably reinforced under the guidance of the word features in the parsed text." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.151, + 0.396, + 0.167 + ], + "angle": 0, + "content": "3.3. Cross-Grained Contrast Supervision" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.175, + 0.469, + 0.281 + ], + "angle": 0, + "content": "To guide the optimization of the neural network for 3D stylization, the first step is to render the stylized 3D mesh from multiple 2D views. Most existing methods usually employ the visual encoder and text encoder of CLIP [28] to extract global feature vectors for the rendered image and target text, respectively, which are contrasted to perform cross-modal supervision via cosine similarity:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.294, + 0.47, + 0.328 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o a r s e}} = - \\frac {\\mathcal {F} _ {I} \\cdot \\mathcal {F} _ {T}}{\\| \\mathcal {F} _ {I} \\| _ {2} \\| \\mathcal {F} _ {T} \\| _ {2}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.341, + 0.469, + 0.401 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_I\\in \\mathbb{R}^{512}\\) is the averaged feature vector of the images rendered from different views, \\(\\mathcal{F}_T\\in \\mathbb{R}^{512}\\) denotes the global feature vector of the target text, and \\(\\| \\cdot \\| _2\\) represents the Euclidean norm function." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.469, + 0.524 + ], + "angle": 0, + "content": "Although achieving impressive results for stylizing a single 3D object, these methods still have limitations in multi-object scenes. Considering that a single feature vector still represents a sentence describing multiple objects, the object details may be lost in large amounts. Therefore, such a coarse-grained contrast supervision is insufficient to guide the neural network in synthesizing photorealistic stylized content for multiple 3D objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.469, + 0.615 + ], + "angle": 0, + "content": "To solve this issue, we construct a fine-grained contrast supervision to complement the coarse-grained one. Specifically, we first calculate the correlation map, i.e., \\(S \\in \\mathbb{R}^{n \\times m}\\) between the word features in the text and the visual features of the rendered images, which are also extracted from the text encoder and visual encoder of the CLIP:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.627, + 0.469, + 0.662 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} = \\frac {\\mathcal {I} \\cdot \\mathcal {T} ^ {\\mathrm {T}}}{\\| \\mathcal {I} \\| _ {2} \\| \\mathcal {T} \\| _ {2}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.675, + 0.469, + 0.766 + ], + "angle": 0, + "content": "where \\(\\mathcal{I} \\in \\mathbb{R}^{n \\times 512}\\) represents the features of the images rendered from \\(n\\) views, \\(\\mathcal{T} \\in \\mathbb{R}^{m \\times 512}\\) indicates the features of \\(m\\) words in the text. Then, we normalize the correlation matrix along the image axis and the text axis, respectively, to retrieve the text of interest and visual components. This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.78, + 0.469, + 0.81 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {I} (i) = \\frac {\\sum_ {k = 1} ^ {m} \\mathcal {S} (i , k)}{m}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.812, + 0.469, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {T} (j) = \\frac {\\sum_ {k = 1} ^ {n} \\mathcal {S} (k , j)}{n}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Inspired by [22], we further calculate an image-centered fine-grained contrast score and a text-centered fine-grained contrast score by the weighted summation of the similarity" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.796, + 0.106 + ], + "angle": 0, + "content": "vectors, which can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.116, + 0.892, + 0.155 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I} = \\sum_ {i = 1} ^ {n} \\frac {e ^ {\\mathcal {S} _ {I} (i)}}{\\sum_ {k = 1} ^ {n} e ^ {\\mathcal {S} _ {I} (k)}} \\mathcal {S} _ {I} (i), \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.159, + 0.892, + 0.199 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T} = \\sum_ {j = 1} ^ {m} \\frac {e ^ {\\mathcal {S} _ {T} (j)}}{\\sum_ {k = 1} ^ {m} e ^ {\\mathcal {S} _ {T} (k)}} \\mathcal {S} _ {T} (j), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.208, + 0.892, + 0.268 + ], + "angle": 0, + "content": "where the weights are defined as the degree of correlation between the central and another modality. Finally, we adopt the average value of these two scores as the fine-grained contrast loss, which can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.279, + 0.892, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f i n e}} = - \\left(\\mathcal {L} _ {I} + \\mathcal {L} _ {T}\\right) / 2. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.307, + 0.892, + 0.396 + ], + "angle": 0, + "content": "The coarse-grained and fine-grained contrast supervision complement each other to build a cross-grained contrast supervision system. The former is utilized to align the global semantic information of the target text with the 3D objects, and the latter is used to achieve the local semantic alignment. This loss can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.408, + 0.892, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e g c s} = \\lambda_ {c} \\mathcal {L} _ {\\text {c o a r s e}} + \\lambda_ {f} \\mathcal {L} _ {\\text {f i n e}}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.435, + 0.892, + 0.48 + ], + "angle": 0, + "content": "where \\(\\lambda_{c}\\) and \\(\\lambda_{f}\\) are two hyper-parameters to balance the cross-grained and the fine-grained losses, set to 1.0 and 0.33, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.493, + 0.633, + 0.51 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.518, + 0.676, + 0.535 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.548, + 0.892, + 0.713 + ], + "angle": 0, + "content": "Datasets. To examine our method across a diverse set of 3D scenes, we first collect 3D object meshes from a variety of sources, i.e., COSEG [33], Thingi10K [55], Shapenet [6], Turbo Squid [36], and ModelNet [43]. Then, we randomly place several objects from the collected 3D set into a mesh using Blender. Note that we down-sample the number of meshes' vertices and faces to ensure the robustness of our TeMO to low-quality meshes and reduce the burden of GPU during the stylization. The meshes used in this paper contain an average of 79,303 faces, \\(16\\%\\) non-manifold edges, \\(0.2\\%\\) non-manifold vertices, and \\(12\\%\\) boundaries." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. Following the TANGO [15] network, we adopt 3 linear layers with 256 dimensions to build the normal estimation branch. In the reflectance branch, the point features are extracted by 2 shared layers with 256 dimensions, followed by 3 exclusive layers to predict diffuse, specular, and roughness. The dimension of our DGA module is also set as 256. The word features in our DGA module are extracted from the text encoder of CLIP, and so are the ones in our CGC loss. We choose ViT-B/32 as the backbone of the pre-trained CLIP model in this paper, which is consistent with previous works [15, 23, 25]. We also process the rendered images with 2D augmentation strategies [12, 15]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19535" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.085, + 0.225, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.184, + 0.218, + 0.196 + ], + "angle": 0, + "content": "person & dragon" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.086, + 0.547, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.183, + 0.519, + 0.196 + ], + "angle": 0, + "content": "Text Prompt: \"an iron man and an ice dragon\"" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.2, + 0.201, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.301, + 0.207, + 0.312 + ], + "angle": 0, + "content": "cat & horse" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.2, + 0.543, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.3, + 0.53, + 0.313 + ], + "angle": 0, + "content": "Text Prompt: \"a Garfield cat and a brown horse\"" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.318, + 0.22, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.421, + 0.212, + 0.432 + ], + "angle": 0, + "content": "vase & candle" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.317, + 0.533, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.42, + 0.529, + 0.433 + ], + "angle": 0, + "content": "Text Prompt: \"a cactus vase and a silver candle\"" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.086, + 0.872, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.183, + 0.83, + 0.196 + ], + "angle": 0, + "content": "Text Prompt: \"a superman and a fire dragon\"" + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.202, + 0.859, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.3, + 0.86, + 0.313 + ], + "angle": 0, + "content": "Text Prompt: \"a ginger cat and an astronaut horse\"" + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.317, + 0.859, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.42, + 0.862, + 0.433 + ], + "angle": 0, + "content": "Text Prompt: \"a wicker vase and a candle in jeans\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.45, + 0.891, + 0.466 + ], + "angle": 0, + "content": "Figure 4. Given the same bare mesh, our TeMO produces various stylized contents for multi-object scenes to conform to the text prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.47, + 0.576 + ], + "angle": 0, + "content": "before feeding them into the pre-trained CLIP model. Our TeMO model is optimized with the AdamW [18] strategy for 1500 iterations, where the learning rate is initialized to \\(5 \\times 10^{-4}\\) and decayed by 0.7 every 500 iterations. The entire training process takes about 10 minutes on a single NVIDIA RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.585, + 0.288, + 0.601 + ], + "angle": 0, + "content": "4.2. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.471, + 0.76 + ], + "angle": 0, + "content": "We conduct visualization experiments on a wide spectrum of multi-object scenes to verify the effectiveness of our TeMO. However, we observe that the 3D symmetry prior used widely in previous works [15, 25] can cause interference between different parts during the stylization process of multiple objects. We argue that the multiple objects of the meshes used in this paper are randomly placed to simulate a real 3D scene rather than along the \\(z\\)-axis. To avoid this issue, we remove this prior in our TeMO and previous methods involved in the comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Neural Stylization and Controls. We present the stylization results of our TeMO driven by different text prompts for the same multi-object mesh in Fig. 4. As shown in the \\(1^{st}\\) row where the 3D scene is composed of a person and a dragon, our TeMO can accurately distinguish between the person object and the dragon object and appropriately stylize different body parts for them according to the semantic roles described in each text prompt. Meanwhile, our TeMO also synthesizes desired stylizations for the 3D objects in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.892, + 0.561 + ], + "angle": 0, + "content": "the cat-horse mesh and vase-candle mesh, as shown in the \\(2^{nd}\\) and \\(3^{rd}\\) rows. These experimental results demonstrate that our TeMO method can generate photorealistic details with fine granularity and maintain global semantic understanding for the given multi-object 3D scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Qualitative Comparisons. We provide the visual comparisons of prediction results between our TeMO and previous pioneering works in text-driven 3D object stylization, including Text2Mesh [25], TANGO [15], and X-Mesh [23]. To ensure a fair comparison, we adopt the official implementations of these methods and also train them with the default settings without the symmetry prior. The experimental results show it is a real struggle for Text2Mesh [25] and TANGO [15] to understand the detailed semantics of the text prompt with multiple objects. As shown in the \\(1^{st}\\) row of Fig. 5 where the 3D scene contains two objects of the same category, given a text prompt \"a fire dragon and an ice dragon\", they tend to capture the \"ice\" property, missing the \"fire\" property. For a 3D scene containing two objects of different categories, they are prone to mixing the properties of these objects, as shown in the \\(2^{nd}\\) row where the text prompt is \"a wood vase and a brick candle\". Therefore, the stylized assets they generate for these multi-object scenes are unsatisfactory. X-Mesh generates more accurate results that align with the text prompts, as shown in the \\(1^{st}\\) and \\(2^{nd}\\) rows, which can be attributed to incorporating the text vector while extracting vertex features. However, it can pro" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19536" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.072, + 0.226, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.072, + 0.388, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.072, + 0.547, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.072, + 0.708, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.072, + 0.866, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.246, + 0.674, + 0.26 + ], + "angle": 0, + "content": "(a) Mesh: two dragons; Text Prompt: \"a fire dragon and an ice dragon\"." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.262, + 0.211, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.262, + 0.382, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.262, + 0.535, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.262, + 0.694, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.262, + 0.854, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.419, + 0.68, + 0.432 + ], + "angle": 0, + "content": "(b) Mesh: vase & candle; Text Prompt: \"a wood vase and a brick candle\"." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.435, + 0.242, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.435, + 0.408, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.435, + 0.567, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.435, + 0.725, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.435, + 0.885, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.579, + 0.743, + 0.593 + ], + "angle": 0, + "content": "(c) Mesh: person & dragon & whale; Text Prompt: \"a superman, a fire dragon, and an ice whale\"." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.604, + 0.895, + 0.634 + ], + "angle": 0, + "content": "Figure 5. Visual comparisons of our TeMO with previous text-driven 3D stylization methods on several multi-object scenes, including two objects of the same or different categories, and three different objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.473, + 0.82 + ], + "angle": 0, + "content": "duce semantic noises due to its utilization of the text vector containing attributes of multiple objects to process all vertex features. With an increasing number of objects, it will also encounter challenges related to comprehending text details and the alignment between the text and 3D objects. As shown in the \\(3^{rd}\\) row, this method still fails to generate stylized assets without mixed properties. In contrast, our TeMO equipped with 3D scene parsing and multi-grained supervision, is able to generate photorealistic stylized content for each object in these 3D scenes to conform to the descriptions in the text prompts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.826, + 0.3, + 0.843 + ], + "angle": 0, + "content": "4.3.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.855, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Objective Metric. We adopt the CLIP score to objectively evaluate the semantic alignment achieved by our TeMO and recent 3D stylization methods. Specifically, 8 views spaced" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.895, + 0.759 + ], + "angle": 0, + "content": "\\(45^{\\circ}\\) around the stylized meshes are chosen to obtain the rendered 2D images. Then, the visual objects are compared with the textual objects in CLIP's embedding space via the cosine function. As shown in the \\(2^{nd}\\) column of Tab. 1, our TeMO surpasses previous methods by a large margin. These results demonstrate the superiority of our TeMO over existing methods on multi-object stylization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.896, + 0.903 + ], + "angle": 0, + "content": "User Study. We further conduct a user study to evaluate these 3D stylization methods subjectively. We randomly select 10 mesh-text pairs and recruit 60 users to evaluate the quality of the stylization assets generated by our TeMO and previous methods. Particularly, the participants include experts in the field and individuals without specific background knowledge. Moreover, each of them will be asked three questions [25]: (Q1) \"How natural is the output results?\"(Q2) \"How well does the output match the original" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19537" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.231, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.194, + 0.187, + 0.208 + ], + "angle": 0, + "content": "Input Mesh" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.089, + 0.418, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.195, + 0.372, + 0.206 + ], + "angle": 0, + "content": "Baseline" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.089, + 0.567, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.194, + 0.592, + 0.207 + ], + "angle": 0, + "content": "Baseline + DGA module" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.089, + 0.718, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.194, + 0.732, + 0.207 + ], + "angle": 0, + "content": "Baseline + CGC loss" + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.089, + 0.869, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.194, + 0.845, + 0.206 + ], + "angle": 0, + "content": "Our TeMO" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.217, + 0.89, + 0.232 + ], + "angle": 0, + "content": "Figure 6. Ablation experiments on the proposed designs of our TeMO. Mesh: two dragons; Text Prompt: \"a fire dragon and an ice dragon\"." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.248, + 0.468, + 0.304 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons of our TeMO and previous text-driven 3D stylization methods in multi-object scenes, including an objective alignment score (0-1) and three subjective opinion scores (1-5). Note that the higher these metrics, the better the method." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.309, + 0.465, + 0.402 + ], + "angle": 0, + "content": "
AlignmentUser-Q1User-Q2User-Q3
Text2Mesh [25]0.2621.7501.5061.472
TANGO [15]0.2742.4062.4502.539
X-Mesh [23]0.2651.8391.7221.761
Our TeMO0.2853.3443.3113.261
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.42, + 0.468, + 0.527 + ], + "angle": 0, + "content": "content?\" (Q3) \"How well does the output match the target style?\", and then assign a score (1-5) to them. We report the mean opinion scores in parentheses for each factor averaged across all style outputs. As shown in Tab. 1, our TeMO still outperforms other methods across all questions. Therefore, the 3D assets generated by our method are more in line with people's understanding of the text prompts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.536, + 0.242, + 0.551 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.559, + 0.47, + 0.833 + ], + "angle": 0, + "content": "To verify the effectiveness of the proposed designs in our TeMO, we conduct ablation studies by gradually adding them to our baseline model, i.e., TANGO [15]. We chose the two-dragon mesh with the text prompt \"a fire dragon and an ice dragon\", and the experimental results are shown in Fig. 6. Compared to the baseline model, introducing our DGA module enables the model to distinguish two dragons, yet it falls short in endowing them with precise texture details. Meanwhile, incorporating our CGC loss facilitates the model to capture more semantic details, e.g., \"fire\" and \"ice\". Nevertheless, it fails to distinguish the two objects. It is noteworthy that the model equipped with these two designs together is not only capable of accurately distinguishing between two objects but can also synthesize high-quality texture details for them. These experiments indicate that our DGA module and CGC loss can effectively assist the model in generating desired stylized content for multiple 3D objects to conform to the target text." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.342, + 0.861 + ], + "angle": 0, + "content": "5. Limitation and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Despite achieving excellent results on text-driven multi-object stylization, our TeMO framework still has a few lim" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.25, + 0.825, + 0.265 + ], + "angle": 0, + "content": "itations, which can also facilitate future research:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.272, + 0.892, + 0.378 + ], + "angle": 0, + "content": "1) 3D Symmetry Prior. As stated in Sec. 4.2, our TeMO fails to incorporate 3D symmetry prior, whose important role has been demonstrated by Text2Mesh [25] in promoting style consistency of a single object. To generate more photorealistic stylization assets for multi-object scenes, it will be valuable to calculate symmetry planes for each object and apply symmetry priors to them." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.491 + ], + "angle": 0, + "content": "2) Diffusion Model. We observe that current diffusion technologies struggle to generate multi-object images according to the text prompt, which hinders the application of diffusion-based stylization methods in multi-object 3D scenes. We argue it would be interesting to extend the concept of scene parsing to the diffusion models for the release of their potential in multi-object editing or generation." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.272, + 0.892, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.506, + 0.619, + 0.521 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.892, + 0.789 + ], + "angle": 0, + "content": "In this paper, we present TeMO, an innovative framework proposing scene parsing and multi-grained cross-modal supervision to achieve text-driven multi-object 3D stylization for the first time. Specifically, we first develop a DGA module to precisely align the objects in the 3D mesh and the text prompt and enhance the 3D point features with the word features belonging to the same object as them. Then, we design a CGC loss, in which the fine-grained loss at the local level and coarse-grained contrast loss at the global level are both constructed and complement each other. Further, extensive experiments are conducted to demonstrate the effectiveness and superiority of our methods over the existing methods among a wide range of multi-object 3D scenes. We believe it is promising to achieve content editing of multiple objects in 3D scenes simultaneously, and we hope the scene-parsing perspective provided by the proposed TeMO framework will inspire future works." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This research was supported by NSFC (NO. 62225604, NO. 62276145), the Fundamental Research Funds for the Central Universities (Nankai University, 070-63223049), CAST through Young Elite Scientist Sponsorship Program (No. YESS20210377). Computations were supported by the Supercomputing Center of Nankai University (NKSC)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19538" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In ICLR, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[2] Guirong Bai, Shizhu He, Kang Liu, and Jun Zhao. Example-guided stylized response generation in zero-shot setting. Science China Information Sciences, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.197, + 0.47, + 0.235 + ], + "angle": 0, + "content": "[3] Edward Loper Bird, Steven and Ewan Klein. Natural language processing with python. O'Reilly Media Inc, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.236, + 0.47, + 0.289 + ], + "angle": 0, + "content": "[4] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In IEEE ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.29, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[5] Arthur Caetano and Misha Sra. Arfy: A pipeline for adapting 3d scenes to augmented reality. In ACM UIST, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.318, + 0.47, + 0.385 + ], + "angle": 0, + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.469, + 0.438 + ], + "angle": 0, + "content": "[7] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In IEEE ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.469, + 0.492 + ], + "angle": 0, + "content": "[8] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In IEEE ICCV, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.493, + 0.469, + 0.548 + ], + "angle": 0, + "content": "[9] Shaoyu Chen, Budmonde Duinkharjav, Xin Sun, Li-Yi Wei, Stefano Petrangeli, Jose Echevarria, Claudio Silva, and Qi Sun. Instant reality: Gaze-contingent perceptual optimization for 3d virtual reality streaming. IEEE TVCG, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.469, + 0.656 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.656, + 0.469, + 0.709 + ], + "angle": 0, + "content": "[12] Kevin Frans, Lisa Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. Advances in Neural Information Processing Systems, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.71, + 0.469, + 0.763 + ], + "angle": 0, + "content": "[13] Mengdi Han, Xiaogang Guo, Xuexian Chen, Cunman Liang, Hangbo Zhao, Qihui Zhang, Wubin Bai, Fan Zhang, Heming Wei, Changsheng Wu, et al. Submillimeter-scale multitematerial terrestrial robots. Science Robotics, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.764, + 0.469, + 0.79 + ], + "angle": 0, + "content": "[14] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In IEEE CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[15] Jiabao Lei, Yabin Zhang, Kui Jia, et al. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[16] Zheng Li, Xiang Li, Xinyi Fu, Xin Zhang, Weiqiang Wang, Shuo Chen, and Jian Yang. Promptkd: Unsupervised prompt distillation for vision-language models. arXiv preprint arXiv:2403.02781, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.148, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[18] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.174, + 0.892, + 0.228 + ], + "angle": 0, + "content": "[19] Haoyu Lu, Yuqi Huo, Mingyu Ding, Nanyi Fei, and Zhiwu Lu. Cross-modal contrastive learning for generalizable and efficient image-text retrieval. Machine Intelligence Research, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.228, + 0.892, + 0.282 + ], + "angle": 0, + "content": "[20] Yunpeng Luo, Jiayi Ji, Xiaoshuai Sun, Liujuan Cao, Yongjian Wu, Feiyue Huang, Chia-Wen Lin, and Rongrong Ji. Dual-level collaborative transformer for image captioning. In AAAI, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.282, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[21] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. In ACL EMNLP, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.322, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[22] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In ACM MM, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.892, + 0.444 + ], + "angle": 0, + "content": "[23] Yiwei Ma, Xiaojing Zhang, Xiaoshuai Sun, Jiayi Ji, Haowei Wang, Guannan Jiang, Weilin Zhuang, and Rongrong Ji. X-mesh: Towards fast and accurate text-driven 3d stylization via dynamic textual guidance. In IEEE ICCV, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.444, + 0.892, + 0.485 + ], + "angle": 0, + "content": "[24] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In IEEE CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[25] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In IEEE CVPR, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[26] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In ACM SIGGRAPH, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.579, + 0.892, + 0.619 + ], + "angle": 0, + "content": "[27] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.687, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.741, + 0.892, + 0.782 + ], + "angle": 0, + "content": "[30] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.782, + 0.892, + 0.822 + ], + "angle": 0, + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.822, + 0.892, + 0.849 + ], + "angle": 0, + "content": "[32] Scott D Roth. Ray casting for modeling solids. Computer graphics and image processing, 1982. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.849, + 0.892, + 0.89 + ], + "angle": 0, + "content": "[33] Oana Sidi, Oliver Van Kaick, Yanir Kleiman, Hao Zhang, and Daniel Cohen-Or. Unsupervised co-segmentation of a set of shapes via descriptor-space spectral clustering. In" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19539" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.273, + 0.105 + ], + "angle": 0, + "content": "ACM SIGGRAPH, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.106, + 0.469, + 0.146 + ], + "angle": 0, + "content": "[34] Jia-Mu Sun, Tong Wu, and Lin Gao. Recent advances in implicit representation-based 3d shape generation. Visual Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.147, + 0.469, + 0.228 + ], + "angle": 0, + "content": "[35] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.228, + 0.469, + 0.254 + ], + "angle": 0, + "content": "[36] TurboSquid. Turbosquid 3d model repository. In https://www.turbosquid.com/, 2021.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.255, + 0.469, + 0.308 + ], + "angle": 0, + "content": "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.309, + 0.469, + 0.349 + ], + "angle": 0, + "content": "[38] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In IEEE CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.401 + ], + "angle": 0, + "content": "[39] Wei Wang, Qialei Dong, and Zhanyi Hu. Asprr: active single-image piecewise planar 3d reconstruction based on geometric priors. Science China Information Sciences, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.402, + 0.469, + 0.441 + ], + "angle": 0, + "content": "[40] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In IEEE CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.443, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[41] Mingrui Wu, Xuying Zhang, Xiaoshuai Sun, Yiyi Zhou, Chao Chen, Jiaxin Gu, Xing Sun, and Rongrong Ji. Difnet: Boosting visual information flow for image captioning. In IEEE CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[42] Ruiqi Wu, Liangyu Chen, Tong Yang, Chunle Guo, Chongyi Li, and Xiangyu Zhang. Lamp: Learn a motion pattern for few-shot-based video generation. arXiv preprint arXiv:2310.10769, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.551, + 0.469, + 0.605 + ], + "angle": 0, + "content": "[43] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In IEEE CVPR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.606, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[44] Qun-Ce Xu, Tai-Jiang Mu, and Yong-Liang Yang. A survey of deep learning-based 3d shape generation. Computational Visual Media, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[45] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In IEEE ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.74 + ], + "angle": 0, + "content": "[46] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.741, + 0.469, + 0.794 + ], + "angle": 0, + "content": "[47] Bowen Yin, Xuying Zhang, Qibin Hou, Bo-Yuan Sun, DengPing Fan, and Luc Van Gool. Camoformer: Masked separable attention for camouflaged object detection. arXiv preprint arXiv:2212.06570, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.795, + 0.469, + 0.847 + ], + "angle": 0, + "content": "[48] Bowen Yin, Xuying Zhang, Zhongyu Li, Li Liu, Ming-Ming Cheng, and Qibin Hou. Dformer: Rethinking rgbd representation learning for semantic segmentation. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.848, + 0.469, + 0.89 + ], + "angle": 0, + "content": "[49] Bo Zhang, Lizbeth Goodman, and Xiaoqing Gu. Novel 3d contextual interactive games on a gamified virtual environment support cultural learning through collaboration among" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.793, + 0.106 + ], + "angle": 0, + "content": "intercultural students. SAGE Open, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.106, + 0.892, + 0.145 + ], + "angle": 0, + "content": "[50] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.147, + 0.892, + 0.2 + ], + "angle": 0, + "content": "[51] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In IEEE CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.201, + 0.892, + 0.254 + ], + "angle": 0, + "content": "[52] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.255, + 0.892, + 0.308 + ], + "angle": 0, + "content": "[53] Xuying Zhang, Xiaoshuai Sun, Yunpeng Luo, Jiayi Ji, Yiyi Zhou, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Rstnet: Captioning with adaptive attention on visual and non-visual words. In IEEE CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.309, + 0.892, + 0.349 + ], + "angle": 0, + "content": "[54] Xuying Zhang, Bowen Yin, Zheng Lin, Qibin Hou, DengPing Fan, and Ming-Ming Cheng. Referring camouflaged object detection. arXiv preprint arXiv:2306.07532, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[55] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models. arXiv preprint arXiv:1605.04797, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.389, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[56] Zoran Zivkovic. Improved adaptive gaussian mixture model for background subtraction. In IEEE ICPR, 2004. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19540" + } + ] +] \ No newline at end of file diff --git a/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/d9ed6877-0911-45bb-b575-e2ba93e2db7d_origin.pdf b/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/d9ed6877-0911-45bb-b575-e2ba93e2db7d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f1b4e45ad1a029ec042b416db056497d0b97b14f --- /dev/null +++ b/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/d9ed6877-0911-45bb-b575-e2ba93e2db7d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84e78b0f10abfeacfbe1ab29e3825bef75a694264bec9b50613afbda403d7d48 +size 1012586 diff --git a/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/full.md b/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0c4e4ab7e980703990bb9b428f119e785736dc23 --- /dev/null +++ b/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/full.md @@ -0,0 +1,384 @@ +# TeMO: Towards Text-Driven 3D Stylization for Multi-Object Meshes + +Xuying Zhang $^{1}$ , Bo-Wen Yin $^{1}$ , Yuming Chen $^{1}$ , Zheng Lin $^{1}$ , Yunheng Li $^{1}$ , Qibin Hou $^{2,1*}$ , Ming-Ming Cheng $^{2,1}$ + +$^{1}$ VCIP, College of Computer Science, Nankai University + +2NKIARI, Shenzhen Futian + +https://github.com/zhangxuying1004/TeMO + +# Abstract + +Recent progress in the text-driven 3D stylization of a single object has been considerably promoted by CLIP-based methods. However, the stylization of multi-object 3D scenes is still impeded in that the image-text pairs used for pretraining CLIP mostly consist of an object. Meanwhile, the local details of multiple objects may be susceptible to omission due to the existing supervision manner primarily relying on coarse-grained contrast of image-text pairs. To overcome these challenges, we present a novel framework, dubbed TeMO, to parse multi-object 3D scenes and edit their styles under the contrast supervision at multiple levels. We first propose a Decoupled Graph Attention (DGA) module to distinguishably reinforce the features of 3D surface points. Particularly, a cross-modal graph is constructed to align the object points accurately and noun phrases decoupled from the 3D mesh and textual description. Then, we develop a Cross-Grained Contrast (CGC) supervision system, where a fine-grained loss between the words in the textual description and the randomly rendered images are constructed to complement the coarse-grained loss. Extensive experiments show that our method can synthesize high-quality stylized content and outperform the existing methods over a wide range of multi-object 3D meshes. + +# 1. Introduction + +3D asset creation through stylization aims to synthesize stylized content on the bare meshes to conform to the given text descriptions [15, 19, 25], referring images [38, 50], or 3D shapes [34, 44]. This research plays an important role in a wide spectrum of applications, e.g., virtual/augmented reality [5, 9], gaming industries [49], and robotics [13]. Moreover, it also presents considerable potential and has attracted increasing attention in computer vision and graphics communities. Considering the ready availability and expres + +![](images/f29180e6a4141db5dce4da8a03a97063adf7a3c22a3239ab05d96c5051db29e4.jpg) +Bare Mesh +two dragons + +![](images/506d174053bf53c0bfb541df1933125738833fba512004468b45f7278c130c2c.jpg) +person & dragon +Figure 1. Visual comparisons between the existing 3D stylization methods (e.g., TANGO [15]) and our TeMO in multi-object scenes. For a scene with multiple objects of the same/different categories, existing methods are prone to interference between different properties of the objects, while our TeMO is able to accurately synthesize the desired stylized content for each object. + +![](images/6743cfa96f79c1bc2ff65bdc0e09206818df741a88b2a0387d97663843683260.jpg) +TANGO + +![](images/7e4ed212aff92a65e766545424f730e33ec26e60e64a1e03457c4e4579866b9b.jpg) +"a fire dragon and an ice dragon" +"an iron man and an ice dragon" + +![](images/b941afa272689e4a3611654fcdf7c2a90a2e88729401116365c66555255a45ea.jpg) +Our TeMO +an ice dragon" + +![](images/7e6ba6910df82c4de5ef970fdebb9c9ca584504fcb2021bf9297e16bdfedce50.jpg) + +siveness of text prompts as well as the popularity of large-scale Contrastive Language-Image Pre-training (CLIP) [28] model, we choose to work with text-driven 3D stylization. + +Recent years have witnessed the emergence of a series of impressive works [8, 15, 23, 25], aiming to drive the advancement of text-driven 3D stylization. Existing methods usually adopt multi-layer perceptrons (MLPs) to predict the location attribute displacements of the bare mesh under the supervision of the contrastive loss in CLIP. We observe that these works focus on the stylization of a single 3D object and perform poorly on multiple objects, as shown in the second column of Fig. 1. We argue that two inherent characteristics of CLIP result in this issue: i) CLIP is mainly pretrained with image-text pairs mostly consisting of a single object; ii) CLIP loss employs global representation vectors from images and text to coarsely match these two modalities, which inevitably causes the loss of local details. More + +over, the key to synthesizing desired styles for multiple 3D objects lies in the parsing of such 3D scenes and the multigrained supervision for details refinement. + +To simultaneously generate stylized content for multiple 3D objects, the primary step is to achieve accurate alignment between the objects in the 3D mesh and the target text. However, existing methods employ global semantics of the text to stylize a single object, which inevitably produces noises when stylizing the objects in multi-object scenes. To overcome this challenge, we propose to parse the 3D scene by introducing a Decoupled Graph Attention (DGA) module. Specifically, all noun phrases are decoupled from the text prompt and the mesh surface points of the current view are divided into several clusters as well. Then, a cross-modal graph is constructed to establish connections of the noun phrases to their corresponding object points while distancing them from the irrelevant ones. This graph enables the accurate interaction between two interrelated modalities. Finally, the surface point features of 3D objects can be reinforced by independent cross-attention fusions with their neighboring word nodes in the graph architecture. + +Furthermore, we also design a Cross-Grained Contrast (CGC) loss to perform comprehensive cross-modal supervision for the stylization of multiple objects. The goal is to guide the network to generate more stylization details for multiple 3D objects to match the target text. Our loss consists of two parts, i.e., coarse-grained contrast and fine-grained contrast. In the former part, the text prompt is regarded as sentence-level supervision, which calculates the similarity between the 2D views rendered from the stylized 3D mesh and the text prompt using the global feature vectors from the CLIP model. In the latter part, we see the text prompt from the word level, and consider the similarities between each word of the sentence and the rendered images of the view sets. To be specific, we produce the word representations of the text prompt by taking the hidden states from the text encoder of CLIP. Motivated by the recent process in video-text retrieval [22], we calculate fine-grained loss via the weighted summation of the element in similarity vectors based on the importance of each word or image. + +Based on the well-designed DGA module and CGC loss, we propose a novel framework towards Text-Driven 3D stylization for Multi-Object Meshes, called TeMO. To validate the effectiveness of our TeMO, extensive experiments are conducted on various multi-object 3D scenes, as shown in the $3^{rd}$ column of Fig. 1. The experimental results demonstrate our TeMO is less susceptible to interference from multiple objects and can generate superior stylized assets compared with the existing 3D stylization methods. + +Our contributions can be summarized as follows: + +- We present a new 3D stylization framework, called TeMO. To the best of our knowledge, it is the first attempt to parse the objects in the text and 3D meshes and + +generate stylizations for multi-object scenes. + +- We propose a Decoupled Graph Attention (DGA) module, which constructs a graph structure to align the surface points in the multi-object mesh and the noun phrases in the text prompt. +- We design a Cross-Grained Contrast (CGC) loss, in which the text is contrasted with the rendered images from sentence and word levels. + +# 2. Related Work + +# 2.1. Text-Driven 3D Manipulation + +Generating or editing 3D content according to a given prompt is a long-standing objective in computer vision and graphics [2, 39, 42]. Among all forms of the prompt, the text has garnered the most conspicuous gaze due to three reasons: i) Text descriptions are readily accessible from the existing corpus; ii) Text descriptions are particularly user-friendly since they are easily modifiable and can effectively express complex concepts related to stylizations; iii) The popularity of large-scale multimodal models [16, 28] has made achieving visual-language supervision possible. + +Text2Mesh [25] proposes a neural-style field network to predict the color and displacement of mesh vertices. TANGO [15] proposes to disentangle the appearance style as the spatially varying bidirectional reflectance distribution, the local geometric variation, and the lighting condition. Then, X-Mesh [23] integrates the target text guidance by utilizing text-relevant spatial and channel-wise attention during vertex feature extraction. Motivated by the remarkable progress in text-driven 2D generation [29, 31], TEXTure [30] and Text2Tex [7] incorporate a pre-trained depth-aware image diffusion model to synthesize high-resolution partial textures from multiple viewpoints progressively. + +To make full use of the priors in the pre-trained 2D text-to-image diffusion model, DreamFusion [27] introduces a Score Distillation Sample (SDS) loss to perform text-to-3D synthesis. With the help of SDS loss, Latent-NeRF [24] and Fantasia3D [8] can generate 3D shapes and appearances for 3D objects. Despite achieving impressive results, these methods focus on the stylization of a single 3D object and rarely explore multi-object scenes. CLIP-Mesh [26] attempts to generate multiple 3D objects for target text. Nevertheless, the resulting content is not satisfactory. In this paper, we parse the objects described in rendered images and text prompts, aligned by two well-designed strategies. + +# 2.2. Attention Mechanism + +The concept of the attention mechanism was initially introduced in neural machine translation [1], where the weighted summation of the candidate vectors is calculated according to their importance scores. This technology has been extended to a myriad of tasks, e.g., natural language pro + +![](images/dd6b711073a82492ad1972ca100baf91667c2de0af42ff850aa75393e3da8a2f.jpg) +Figure 2. The overall architecture of the proposed TeMO framework. We first specify several cameras to cast rays toward the objects in the 3D mesh scene. Then, a surface point $x_{p}$ and normal $n_p$ can be attained from each ray intersected with the objects. These points and normals are fed to the attribute prediction network where the features of 3D objects are parsed and interacted with the decoupled text features via our proposed DGA module. Meanwhile, we employ a series of spherical Gaussians to represent the lighting. Finally, a differentiable SG render is adopted to render images, which are utilized to contrast with the text prompt by our designed CGC loss. + +cessing [10, 21, 37], computer vision [14, 17, 47, 54], and multi-modal learning [20, 41, 48, 53]. For instance, Transformer [37] employs the self-attention operation to establish connections between words within a sentence and utilize the cross-attention mechanism to align source and target sentences. Non-local network [40] takes the lead in introducing self-attention to computer vision and achieves great success in video understanding and object detection. ViT [11] treats an image as a sequence of patches and employs a Transformer encoder based on self-attention to perform image classification. Swin Transformer [17] introduces shifted windows to enhance the local perception ability of self-attention. More recently, X-Mesh [23] designs a text-guided dynamic attention mechanism for vertex feature extraction of a 3D object. However, this guidance only relies on a text feature vector without considering the parsing of text and 3D scenes. In this paper, the multiple objects decoupled from the target text and 3D mesh are aligned via a cross-modal graph to achieve precise guidance. + +# 2.3. Multi-modal Contrastive Learning + +Contrastive learning has become an increasingly popular research topic in the multi-modal community due to its ability to align different modal representations. Based on this strategy, CLIP [28] is pre-trained on an abundance of image-text pairs, achieving great success in cross-modal supervision. TACo [45] presents a token-aware cascade contrastive learning based on the syntactic classes of words to achieve fine-grained semantic alignment in text-video retrieval. Concurrently, FILIP [46] proposes comparing the image patches with the words in the sentence. Regarding the text-driven 3D stylization, the CLIP loss, which calculates the similarity between the image and text vectors in the + +embedding space of CLIP, is adopted by the vast majority of methods. Although achieving impressive results in stylizing a single object, these methods cannot be well adapted to scenes with multiple 3D objects. We argue an important reason for this issue is the loss of local details caused by such coarse-grained supervision. In this paper, we propose a cross-grained supervision strategy, which considers fine-grained and coarse-grained contrasts to achieve a more precise semantic alignment between rendered image and text. + +# 3. Methodology + +# 3.1. Overall Architecture + +Fig. 2 shows the end-to-end architecture of our TeMO framework. Given a bare mesh and a text prompt containing multiple objects, the TeMO aims to synthesize stylization on the mesh to match the text descriptions. We employ a set of vertices $V \in \mathbb{R}^{e \times 3}$ and faces $F \in \{1, \dots, e\}^{u \times 1}$ to explicitly define the input triangle mesh, which is fixed throughout the training. Following TANGO [15], we disentangle the appearance style as the spatially varying bidirectional reflectance distribution function [4, 51, 52] (including diffuse, roughness, specular), the local geometric variation (normal map), and the lighting condition. + +We start by normalizing the vertex coordinates to lie inside a unit sphere. Then, we randomly sample points around the mesh using Gaussian distribution as camera positions to render images. Next, we can obtain a camera ray $R_{p} = \{c + t\nu_{p}\}$ from the sampled camera position $c$ and a pixel $p$ in rendered images, where $\nu_{p}$ is the direction of the ray. Further, ray casting [32] is used to seek out the ray and mesh's first intersection point and intersection face. Moreover, the normal $n_p\in \mathbb{R}^3$ of the intersection face is + +![](images/0d31f3ed4fe35e9674f671b320e8017ae10095874286c3c822b432cb2036b6c3.jpg) +Figure 3. Construction pipeline of the cross-modal graph architecture in our DGA module. Note that $x_{p}$ , the surface point of the 3D objects, and $w_{ij}$ , the $j$ -th word in the $i$ -th noun phrase, are connected together only if they correspond to the same object. + +employed as surface normal at the point $x_{p}\in \mathbb{R}^{3}$ + +To achieve multiview-consistent features, our TeMO is restricted to predicting the normal displacement as a function of the location, while allowing the color materials to be predicted as a function of both location and viewing direction. Therefore, our TeMO represented as MLPs includes two branches, i.e., normal branch $f_{n}(\cdot)$ and reflectance branch $f_{r}(\cdot)$ . Specifically, the former is utilized as the prediction of normal offset on the point $x_{p}$ , and the latter is designed to predict surface reflectance coefficients of the material at the location $x_{p}$ , i.e., diffuse, roughness, and specular. To synthesize high-frequency details, we also apply the Fourier positional encoding [35] to every input. In addition, the spherical Gaussian is employed to represent each light intensity $L_{i}(\cdot)$ due to its closed-form nature and analytical solution. Based on the attained geometric and appearance components, each pixel color in the rendered image can be calculated by a hemisphere renderer [15]: + +$$ +L _ {p} \left(\nu_ {p}, x _ {p}, n _ {p}\right) = \int_ {\Omega} L _ {i} \left(w _ {i}\right) f _ {r} \left(\nu_ {p}, w _ {i}, x _ {p}\right) \left(w _ {i} \cdot \hat {n} _ {p}\right) d w _ {i}, \tag {1} +$$ + +$$ +\hat {n} _ {p} = n _ {p} + f _ {n} \left(x _ {p}, n _ {p}\right), \tag {2} +$$ + +where $\Omega = \{w_{i}:w_{i}\cdot \hat{n}_{p}\geq 0\}$ represents the hemisphere, $w_{i}$ is the incident light direction, and $\hat{n}_p$ is the estimated normal on surface point $x_{p}$ . + +# 3.2. Decoupled Graph Attention + +To achieve text-drive stylization for multiple 3D objects, the key issue that needs to be solved is the accurate alignment between the objects described in the text and those in the meshes. X-Mesh [23] has incorporated text-guided dynamic linear layers, in which the global representation vector of the target object in the text is utilized as guidance to acquire text-aware vertex features. Nevertheless, the global vector containing information about multiple objects is prone to mutual interference and produces semantic noises during guidance for multi-object scenes. + +To address this challenge, we propose to parse the objects in the text and mesh. We first extract the noun phrases modified by adjectives or prepositional phrases from the text using the NLTK tools [3]. Then, we employ the Gaussian Mixture Model (GMM) [56] to cluster the intersection point set $\{x_{1},\dots,x_{p},\dots\}$ of the current ray and the mesh. Meanwhile, we can obtain a binary map of objects in the current view based on whether the ray intersects with the mesh. Further, we can decouple the objects in the binary map according to the clustered points and acquire several binary maps of individual objects. Based on the disentangled noun phrases and binary maps for multiple objects, we can match the correct pairs by their semantic similarities. As a result, the objects described in the text are aligned with their corresponding objects in the mesh, which are utilized to construct a cross-modal graph $G = (\mathcal{V},\mathcal{E})$ , as shown in Fig. 3. To be specific, all surface point features and word features are considered as independent nodes to form the node set $\mathcal{V}$ . For the edge set $\mathcal{V}$ , the link between the surface point node and the word node will be built if the semantic objects they belong to are the same. + +Under the setting of this cross-modal graph, we can individually perform cross-attention between the surface point nodes and their neighboring word nodes, where the parsed surface point features are used as queries, and the parsed text features serve as keys and values. The enhancement of the surface point node $v_{i} \in \mathbb{R}^{dim}$ can be formulated as: + +$$ +\hat {v} _ {i} = \sum_ {v _ {j} \in \operatorname {A d j} \left(v _ {i}\right)} \alpha_ {i j} \operatorname {L i n e a r} \left(v _ {j}\right), \tag {3} +$$ + +$$ +\alpha_ {i j} = \frac {e ^ {\mathcal {W} _ {i j}}}{\sum_ {v _ {j} \in \operatorname {A d j} (v _ {i})} e \mathcal {W} _ {i j}}, \tag {4} +$$ + +$$ +\mathcal {W} _ {i j} = \frac {\operatorname {L i n e a r} \left(v _ {i}\right) \operatorname {L i n e a r} \left(v _ {j}\right) ^ {\mathrm {T}}}{\sqrt {d _ {l}}}, \tag {5} +$$ + +where $\mathrm{Adj}(v_i)$ is the adjacency nodes of $v_{i}$ and $\mathrm{Linear}(\cdot)$ represents a linear transformation. With this attention + +mechanism, the surface point features of different objects in the mesh can be distinguishably reinforced under the guidance of the word features in the parsed text. + +# 3.3. Cross-Grained Contrast Supervision + +To guide the optimization of the neural network for 3D stylization, the first step is to render the stylized 3D mesh from multiple 2D views. Most existing methods usually employ the visual encoder and text encoder of CLIP [28] to extract global feature vectors for the rendered image and target text, respectively, which are contrasted to perform cross-modal supervision via cosine similarity: + +$$ +\mathcal {L} _ {\text {c o a r s e}} = - \frac {\mathcal {F} _ {I} \cdot \mathcal {F} _ {T}}{\| \mathcal {F} _ {I} \| _ {2} \| \mathcal {F} _ {T} \| _ {2}}, \tag {6} +$$ + +where $\mathcal{F}_I\in \mathbb{R}^{512}$ is the averaged feature vector of the images rendered from different views, $\mathcal{F}_T\in \mathbb{R}^{512}$ denotes the global feature vector of the target text, and $\| \cdot \| _2$ represents the Euclidean norm function. + +Although achieving impressive results for stylizing a single 3D object, these methods still have limitations in multi-object scenes. Considering that a single feature vector still represents a sentence describing multiple objects, the object details may be lost in large amounts. Therefore, such a coarse-grained contrast supervision is insufficient to guide the neural network in synthesizing photorealistic stylized content for multiple 3D objects. + +To solve this issue, we construct a fine-grained contrast supervision to complement the coarse-grained one. Specifically, we first calculate the correlation map, i.e., $S \in \mathbb{R}^{n \times m}$ between the word features in the text and the visual features of the rendered images, which are also extracted from the text encoder and visual encoder of the CLIP: + +$$ +\mathcal {S} = \frac {\mathcal {I} \cdot \mathcal {T} ^ {\mathrm {T}}}{\| \mathcal {I} \| _ {2} \| \mathcal {T} \| _ {2}}, \tag {7} +$$ + +where $\mathcal{I} \in \mathbb{R}^{n \times 512}$ represents the features of the images rendered from $n$ views, $\mathcal{T} \in \mathbb{R}^{m \times 512}$ indicates the features of $m$ words in the text. Then, we normalize the correlation matrix along the image axis and the text axis, respectively, to retrieve the text of interest and visual components. This process can be formulated as: + +$$ +\mathcal {S} _ {I} (i) = \frac {\sum_ {k = 1} ^ {m} \mathcal {S} (i , k)}{m}, \tag {8} +$$ + +$$ +\mathcal {S} _ {T} (j) = \frac {\sum_ {k = 1} ^ {n} \mathcal {S} (k , j)}{n}. \tag {9} +$$ + +Inspired by [22], we further calculate an image-centered fine-grained contrast score and a text-centered fine-grained contrast score by the weighted summation of the similarity + +vectors, which can be formulated as follows: + +$$ +\mathcal {L} _ {I} = \sum_ {i = 1} ^ {n} \frac {e ^ {\mathcal {S} _ {I} (i)}}{\sum_ {k = 1} ^ {n} e ^ {\mathcal {S} _ {I} (k)}} \mathcal {S} _ {I} (i), \tag {10} +$$ + +$$ +\mathcal {L} _ {T} = \sum_ {j = 1} ^ {m} \frac {e ^ {\mathcal {S} _ {T} (j)}}{\sum_ {k = 1} ^ {m} e ^ {\mathcal {S} _ {T} (k)}} \mathcal {S} _ {T} (j), \tag {11} +$$ + +where the weights are defined as the degree of correlation between the central and another modality. Finally, we adopt the average value of these two scores as the fine-grained contrast loss, which can be defined as: + +$$ +\mathcal {L} _ {\text {f i n e}} = - \left(\mathcal {L} _ {I} + \mathcal {L} _ {T}\right) / 2. \tag {12} +$$ + +The coarse-grained and fine-grained contrast supervision complement each other to build a cross-grained contrast supervision system. The former is utilized to align the global semantic information of the target text with the 3D objects, and the latter is used to achieve the local semantic alignment. This loss can be defined as: + +$$ +\mathcal {L} _ {e g c s} = \lambda_ {c} \mathcal {L} _ {\text {c o a r s e}} + \lambda_ {f} \mathcal {L} _ {\text {f i n e}}, \tag {13} +$$ + +where $\lambda_{c}$ and $\lambda_{f}$ are two hyper-parameters to balance the cross-grained and the fine-grained losses, set to 1.0 and 0.33, respectively. + +# 4. Experiments + +# 4.1. Experiment Setup + +Datasets. To examine our method across a diverse set of 3D scenes, we first collect 3D object meshes from a variety of sources, i.e., COSEG [33], Thingi10K [55], Shapenet [6], Turbo Squid [36], and ModelNet [43]. Then, we randomly place several objects from the collected 3D set into a mesh using Blender. Note that we down-sample the number of meshes' vertices and faces to ensure the robustness of our TeMO to low-quality meshes and reduce the burden of GPU during the stylization. The meshes used in this paper contain an average of 79,303 faces, $16\%$ non-manifold edges, $0.2\%$ non-manifold vertices, and $12\%$ boundaries. + +Implementation Details. Following the TANGO [15] network, we adopt 3 linear layers with 256 dimensions to build the normal estimation branch. In the reflectance branch, the point features are extracted by 2 shared layers with 256 dimensions, followed by 3 exclusive layers to predict diffuse, specular, and roughness. The dimension of our DGA module is also set as 256. The word features in our DGA module are extracted from the text encoder of CLIP, and so are the ones in our CGC loss. We choose ViT-B/32 as the backbone of the pre-trained CLIP model in this paper, which is consistent with previous works [15, 23, 25]. We also process the rendered images with 2D augmentation strategies [12, 15] + +![](images/66440921bdc8aa694a70f1e5f85db2247d6df544ff841f8c835d1c9c00ad1a31.jpg) +person & dragon + +![](images/de78b7850f6a7c423f5aa2f33e5a1646f59043805612cb42960574f3532034e9.jpg) +Text Prompt: "an iron man and an ice dragon" + +![](images/56676eea3e4236823cdae3cad31bce5ef3a12267989770955f7b7d0d5fb83b16.jpg) +cat & horse + +![](images/accb7cd2f28b76c881ed1c44e205f8996ddc39de04e544d1c5a645a18debaf6d.jpg) +Text Prompt: "a Garfield cat and a brown horse" + +![](images/8388728499c4e0e4734ff31c97ff5efb09b85a509b990e55cefe575168e4b1d3.jpg) +vase & candle +Figure 4. Given the same bare mesh, our TeMO produces various stylized contents for multi-object scenes to conform to the text prompts. + +![](images/dae3661b5d0f97498c902aeacc951df596ea72da090516aea6f2866b0426631d.jpg) +Text Prompt: "a cactus vase and a silver candle" + +![](images/00a50e88bd42ccfb55b26c012674f13ba577db6c468419f652c7224183570c00.jpg) +Text Prompt: "a superman and a fire dragon" + +![](images/dcb6da02a32a4a26a4d912c5e26e6b3abfc385a625041cd82849f0dde536c7a9.jpg) +Text Prompt: "a ginger cat and an astronaut horse" + +![](images/052f030a88a0ac9f058396d72655a81eae0d4d2cb3470e2b931253e1f146b67f.jpg) +Text Prompt: "a wicker vase and a candle in jeans" + +before feeding them into the pre-trained CLIP model. Our TeMO model is optimized with the AdamW [18] strategy for 1500 iterations, where the learning rate is initialized to $5 \times 10^{-4}$ and decayed by 0.7 every 500 iterations. The entire training process takes about 10 minutes on a single NVIDIA RTX 3090 GPU. + +# 4.2. Qualitative Evaluation + +We conduct visualization experiments on a wide spectrum of multi-object scenes to verify the effectiveness of our TeMO. However, we observe that the 3D symmetry prior used widely in previous works [15, 25] can cause interference between different parts during the stylization process of multiple objects. We argue that the multiple objects of the meshes used in this paper are randomly placed to simulate a real 3D scene rather than along the $z$ -axis. To avoid this issue, we remove this prior in our TeMO and previous methods involved in the comparison. + +Neural Stylization and Controls. We present the stylization results of our TeMO driven by different text prompts for the same multi-object mesh in Fig. 4. As shown in the $1^{st}$ row where the 3D scene is composed of a person and a dragon, our TeMO can accurately distinguish between the person object and the dragon object and appropriately stylize different body parts for them according to the semantic roles described in each text prompt. Meanwhile, our TeMO also synthesizes desired stylizations for the 3D objects in + +the cat-horse mesh and vase-candle mesh, as shown in the $2^{nd}$ and $3^{rd}$ rows. These experimental results demonstrate that our TeMO method can generate photorealistic details with fine granularity and maintain global semantic understanding for the given multi-object 3D scene. + +Qualitative Comparisons. We provide the visual comparisons of prediction results between our TeMO and previous pioneering works in text-driven 3D object stylization, including Text2Mesh [25], TANGO [15], and X-Mesh [23]. To ensure a fair comparison, we adopt the official implementations of these methods and also train them with the default settings without the symmetry prior. The experimental results show it is a real struggle for Text2Mesh [25] and TANGO [15] to understand the detailed semantics of the text prompt with multiple objects. As shown in the $1^{st}$ row of Fig. 5 where the 3D scene contains two objects of the same category, given a text prompt "a fire dragon and an ice dragon", they tend to capture the "ice" property, missing the "fire" property. For a 3D scene containing two objects of different categories, they are prone to mixing the properties of these objects, as shown in the $2^{nd}$ row where the text prompt is "a wood vase and a brick candle". Therefore, the stylized assets they generate for these multi-object scenes are unsatisfactory. X-Mesh generates more accurate results that align with the text prompts, as shown in the $1^{st}$ and $2^{nd}$ rows, which can be attributed to incorporating the text vector while extracting vertex features. However, it can pro + +![](images/fa8fe2140bb9db1d945a3ae5501df2a41c550a20bdd233b0458df72f74107771.jpg) + +![](images/dffe4c493395d99a1da0533779e4cbae62cad93212cd00dd3a54d0f3d37ce2ac.jpg) +(a) Mesh: two dragons; Text Prompt: "a fire dragon and an ice dragon". + +![](images/ef2eba99c8b0ca505c24f32150933a2cb9b0b69abbe17f065b53c17c5917a329.jpg) + +![](images/b1ebfbda641f328416a346ee6d30f54725eb6e91d692d5a5828d0af7dcac0851.jpg) + +![](images/1519451221330d6b8d412058bbfda9744d2c7bfb32584a4a313c3e27859d0a4a.jpg) + +![](images/9ab4c588d5d5c274e4356975293a8389f75fa609dd9a9351cea38e304dd1fd36.jpg) + +![](images/7ba345fe11419efb040709351da1cab36f3aa5e14ba84011484e9da63026f6db.jpg) +(b) Mesh: vase & candle; Text Prompt: "a wood vase and a brick candle". + +![](images/ad1cac078020fbcc33b971612e8f13d2af6a4702bb55f832c1fd552fb9416d64.jpg) + +![](images/95491438540621726297fe5c6c0ab4a41fc472c7052b5e463e5b0963ac058fb2.jpg) + +![](images/5c058d4d7387ab23d6b3db25210bf878b1895b3f522acb39fbafdb62b6174107.jpg) + +![](images/d805d17d418b5c86e6b1b8b73d701e32604123229d28b61bc3606edd2a8fb838.jpg) +(c) Mesh: person & dragon & whale; Text Prompt: "a superman, a fire dragon, and an ice whale". + +![](images/0ea2e3d9fefb5c4cf0477dd36593bb1fe7db51ff4540aa3165e3cbcaaf439544.jpg) +Figure 5. Visual comparisons of our TeMO with previous text-driven 3D stylization methods on several multi-object scenes, including two objects of the same or different categories, and three different objects. + +![](images/335c60e6fc3e0894e7eb0ddd01097f70d14291833281797915760915310ec445.jpg) + +![](images/c9660c91ddb4e36bad560489363f0d8da85f9855c056141c4e05ec3a88521e22.jpg) + +![](images/5862fc8ad0f9f35f165b3012959025f2649a045bafdbfdf344e8684b516def23.jpg) + +duce semantic noises due to its utilization of the text vector containing attributes of multiple objects to process all vertex features. With an increasing number of objects, it will also encounter challenges related to comprehending text details and the alignment between the text and 3D objects. As shown in the $3^{rd}$ row, this method still fails to generate stylized assets without mixed properties. In contrast, our TeMO equipped with 3D scene parsing and multi-grained supervision, is able to generate photorealistic stylized content for each object in these 3D scenes to conform to the descriptions in the text prompts. + +# 4.3.Quantitative Evaluation + +Objective Metric. We adopt the CLIP score to objectively evaluate the semantic alignment achieved by our TeMO and recent 3D stylization methods. Specifically, 8 views spaced + +$45^{\circ}$ around the stylized meshes are chosen to obtain the rendered 2D images. Then, the visual objects are compared with the textual objects in CLIP's embedding space via the cosine function. As shown in the $2^{nd}$ column of Tab. 1, our TeMO surpasses previous methods by a large margin. These results demonstrate the superiority of our TeMO over existing methods on multi-object stylization. + +User Study. We further conduct a user study to evaluate these 3D stylization methods subjectively. We randomly select 10 mesh-text pairs and recruit 60 users to evaluate the quality of the stylization assets generated by our TeMO and previous methods. Particularly, the participants include experts in the field and individuals without specific background knowledge. Moreover, each of them will be asked three questions [25]: (Q1) "How natural is the output results?"(Q2) "How well does the output match the original + +![](images/9047c9124b16781a232ba564a837fe67ffcd09d9dc577149c7e9ab75ee52071b.jpg) +Input Mesh + +![](images/44e2474405cf9ffb61160181f4083aa5bbedbc47695458f432c68ef65e7392c7.jpg) +Baseline + +![](images/05c809defa7cd9ba5d3381bfcc56e23f3c6eeee6703d49dbba6bf041ceb46982.jpg) +Baseline + DGA module + +![](images/1f6955df3bd31187a4b436fabd7a2eb598d7630c3bc82ad2ca2968cafa777b99.jpg) +Baseline + CGC loss + +![](images/37db2236ad4868c56b8561f82bb54e62e56e161fd004edd0e38873e1c2269ef2.jpg) +Our TeMO +Figure 6. Ablation experiments on the proposed designs of our TeMO. Mesh: two dragons; Text Prompt: "a fire dragon and an ice dragon". + +Table 1. Quantitative comparisons of our TeMO and previous text-driven 3D stylization methods in multi-object scenes, including an objective alignment score (0-1) and three subjective opinion scores (1-5). Note that the higher these metrics, the better the method. + +
AlignmentUser-Q1User-Q2User-Q3
Text2Mesh [25]0.2621.7501.5061.472
TANGO [15]0.2742.4062.4502.539
X-Mesh [23]0.2651.8391.7221.761
Our TeMO0.2853.3443.3113.261
+ +content?" (Q3) "How well does the output match the target style?", and then assign a score (1-5) to them. We report the mean opinion scores in parentheses for each factor averaged across all style outputs. As shown in Tab. 1, our TeMO still outperforms other methods across all questions. Therefore, the 3D assets generated by our method are more in line with people's understanding of the text prompts. + +# 4.4. Ablation Studies + +To verify the effectiveness of the proposed designs in our TeMO, we conduct ablation studies by gradually adding them to our baseline model, i.e., TANGO [15]. We chose the two-dragon mesh with the text prompt "a fire dragon and an ice dragon", and the experimental results are shown in Fig. 6. Compared to the baseline model, introducing our DGA module enables the model to distinguish two dragons, yet it falls short in endowing them with precise texture details. Meanwhile, incorporating our CGC loss facilitates the model to capture more semantic details, e.g., "fire" and "ice". Nevertheless, it fails to distinguish the two objects. It is noteworthy that the model equipped with these two designs together is not only capable of accurately distinguishing between two objects but can also synthesize high-quality texture details for them. These experiments indicate that our DGA module and CGC loss can effectively assist the model in generating desired stylized content for multiple 3D objects to conform to the target text. + +# 5. Limitation and Future Work + +Despite achieving excellent results on text-driven multi-object stylization, our TeMO framework still has a few lim + +itations, which can also facilitate future research: + +1) 3D Symmetry Prior. As stated in Sec. 4.2, our TeMO fails to incorporate 3D symmetry prior, whose important role has been demonstrated by Text2Mesh [25] in promoting style consistency of a single object. To generate more photorealistic stylization assets for multi-object scenes, it will be valuable to calculate symmetry planes for each object and apply symmetry priors to them. +2) Diffusion Model. We observe that current diffusion technologies struggle to generate multi-object images according to the text prompt, which hinders the application of diffusion-based stylization methods in multi-object 3D scenes. We argue it would be interesting to extend the concept of scene parsing to the diffusion models for the release of their potential in multi-object editing or generation. + +# 6. Conclusion + +In this paper, we present TeMO, an innovative framework proposing scene parsing and multi-grained cross-modal supervision to achieve text-driven multi-object 3D stylization for the first time. Specifically, we first develop a DGA module to precisely align the objects in the 3D mesh and the text prompt and enhance the 3D point features with the word features belonging to the same object as them. Then, we design a CGC loss, in which the fine-grained loss at the local level and coarse-grained contrast loss at the global level are both constructed and complement each other. Further, extensive experiments are conducted to demonstrate the effectiveness and superiority of our methods over the existing methods among a wide range of multi-object 3D scenes. We believe it is promising to achieve content editing of multiple objects in 3D scenes simultaneously, and we hope the scene-parsing perspective provided by the proposed TeMO framework will inspire future works. + +Acknowledgments. This research was supported by NSFC (NO. 62225604, NO. 62276145), the Fundamental Research Funds for the Central Universities (Nankai University, 070-63223049), CAST through Young Elite Scientist Sponsorship Program (No. YESS20210377). Computations were supported by the Supercomputing Center of Nankai University (NKSC). + +# References + +[1] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In ICLR, 2014. 2 +[2] Guirong Bai, Shizhu He, Kang Liu, and Jun Zhao. Example-guided stylized response generation in zero-shot setting. Science China Information Sciences, 2022. 2 +[3] Edward Loper Bird, Steven and Ewan Klein. Natural language processing with python. O'Reilly Media Inc, 2009. 4 +[4] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In IEEE ICCV, 2021. 3 +[5] Arthur Caetano and Misha Sra. Arfy: A pipeline for adapting 3d scenes to augmented reality. In ACM UIST, 2022. 1 +[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5 +[7] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In IEEE ICCV, 2023. 2 +[8] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In IEEE ICCV, 2023. 1, 2 +[9] Shaoyu Chen, Budmonde Duinkharjav, Xin Sun, Li-Yi Wei, Stefano Petrangeli, Jose Echevarria, Claudio Silva, and Qi Sun. Instant reality: Gaze-contingent perceptual optimization for 3d virtual reality streaming. IEEE TVCG, 2022. 1 +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 3 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3 +[12] Kevin Frans, Lisa Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. Advances in Neural Information Processing Systems, 2022. 5 +[13] Mengdi Han, Xiaogang Guo, Xuexian Chen, Cunman Liang, Hangbo Zhao, Qihui Zhang, Wubin Bai, Fan Zhang, Heming Wei, Changsheng Wu, et al. Submillimeter-scale multitematerial terrestrial robots. Science Robotics, 2022. 1 +[14] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In IEEE CVPR, 2018. 3 +[15] Jiabao Lei, Yabin Zhang, Kui Jia, et al. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[16] Zheng Li, Xiang Li, Xinyi Fu, Xin Zhang, Weiqiang Wang, Shuo Chen, and Jian Yang. Promptkd: Unsupervised prompt distillation for vision-language models. arXiv preprint arXiv:2403.02781, 2024. 2 + +[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE ICCV, 2021. 3 +[18] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 6 +[19] Haoyu Lu, Yuqi Huo, Mingyu Ding, Nanyi Fei, and Zhiwu Lu. Cross-modal contrastive learning for generalizable and efficient image-text retrieval. Machine Intelligence Research, 2023. 1 +[20] Yunpeng Luo, Jiayi Ji, Xiaoshuai Sun, Liujuan Cao, Yongjian Wu, Feiyue Huang, Chia-Wen Lin, and Rongrong Ji. Dual-level collaborative transformer for image captioning. In AAAI, 2021. 3 +[21] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. In ACL EMNLP, 2015. 3 +[22] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In ACM MM, 2022. 2, 5 +[23] Yiwei Ma, Xiaojing Zhang, Xiaoshuai Sun, Jiayi Ji, Haowei Wang, Guannan Jiang, Weilin Zhuang, and Rongrong Ji. X-mesh: Towards fast and accurate text-driven 3d stylization via dynamic textual guidance. In IEEE ICCV, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[24] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In IEEE CVPR, 2023. 2 +[25] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In IEEE CVPR, 2022. 1, 2, 5, 6, 7, 8 +[26] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In ACM SIGGRAPH, 2022. 2 +[27] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2 +[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 5 +[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 2 +[30] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH, 2023. 2 +[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE CVPR, 2022. 2 +[32] Scott D Roth. Ray casting for modeling solids. Computer graphics and image processing, 1982. 3 +[33] Oana Sidi, Oliver Van Kaick, Yanir Kleiman, Hao Zhang, and Daniel Cohen-Or. Unsupervised co-segmentation of a set of shapes via descriptor-space spectral clustering. In + +ACM SIGGRAPH, 2011. 5 +[34] Jia-Mu Sun, Tong Wu, and Lin Gao. Recent advances in implicit representation-based 3d shape generation. Visual Intelligence, 2024. 1 +[35] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 2020. 4 +[36] TurboSquid. Turbosquid 3d model repository. In https://www.turbosquid.com/, 2021.5 +[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3 +[38] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In IEEE CVPR, 2022. 1 +[39] Wei Wang, Qialei Dong, and Zhanyi Hu. Asprr: active single-image piecewise planar 3d reconstruction based on geometric priors. Science China Information Sciences, 2023. 2 +[40] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In IEEE CVPR, 2018. 3 +[41] Mingrui Wu, Xuying Zhang, Xiaoshuai Sun, Yiyi Zhou, Chao Chen, Jiaxin Gu, Xing Sun, and Rongrong Ji. Difnet: Boosting visual information flow for image captioning. In IEEE CVPR, 2022. 3 +[42] Ruiqi Wu, Liangyu Chen, Tong Yang, Chunle Guo, Chongyi Li, and Xiangyu Zhang. Lamp: Learn a motion pattern for few-shot-based video generation. arXiv preprint arXiv:2310.10769, 2023. 2 +[43] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In IEEE CVPR, 2015. 5 +[44] Qun-Ce Xu, Tai-Jiang Mu, and Yong-Liang Yang. A survey of deep learning-based 3d shape generation. Computational Visual Media, 2023. 1 +[45] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In IEEE ICCV, 2021. 3 +[46] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. In ICLR, 2022. 3 +[47] Bowen Yin, Xuying Zhang, Qibin Hou, Bo-Yuan Sun, DengPing Fan, and Luc Van Gool. Camoformer: Masked separable attention for camouflaged object detection. arXiv preprint arXiv:2212.06570, 2022. 3 +[48] Bowen Yin, Xuying Zhang, Zhongyu Li, Li Liu, Ming-Ming Cheng, and Qibin Hou. Dformer: Rethinking rgbd representation learning for semantic segmentation. In ICLR, 2024. 3 +[49] Bo Zhang, Lizbeth Goodman, and Xiaoqing Gu. Novel 3d contextual interactive games on a gamified virtual environment support cultural learning through collaboration among + +intercultural students. SAGE Open, 2022. 1 +[50] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In ECCV, 2022. 1 +[51] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In IEEE CVPR, 2021. 3 +[52] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG, 2021. 3 +[53] Xuying Zhang, Xiaoshuai Sun, Yunpeng Luo, Jiayi Ji, Yiyi Zhou, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Rstnet: Captioning with adaptive attention on visual and non-visual words. In IEEE CVPR, 2021. 3 +[54] Xuying Zhang, Bowen Yin, Zheng Lin, Qibin Hou, DengPing Fan, and Ming-Ming Cheng. Referring camouflaged object detection. arXiv preprint arXiv:2306.07532, 2023. 3 +[55] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models. arXiv preprint arXiv:1605.04797, 2016. 5 +[56] Zoran Zivkovic. Improved adaptive gaussian mixture model for background subtraction. In IEEE ICPR, 2004. 4 \ No newline at end of file diff --git a/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/images.zip b/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7f976e5e3d32e2e6a34a6b12b1cc460b1aa7ca1b --- /dev/null +++ b/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d73305ff77282f4fb45132e0f526c23202bfc53f34812e3e653bf1801f575b58 +size 583447 diff --git a/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/layout.json b/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..299920f130b6bf8e7b9273a48c9a954c8aaa663c --- /dev/null +++ b/2024/TeMO_ Towards Text-Driven 3D Stylization for Multi-Object Meshes/layout.json @@ -0,0 +1,9718 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 84, + 103, + 510, + 121 + ], + "type": "text", + "content": "TeMO: Towards Text-Driven 3D Stylization for Multi-Object Meshes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 142, + 583, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 142, + 583, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 142, + 583, + 159 + ], + "type": "text", + "content": "Xuying Zhang" + }, + { + "bbox": [ + 53, + 142, + 583, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 583, + 159 + ], + "type": "text", + "content": ", Bo-Wen Yin" + }, + { + "bbox": [ + 53, + 142, + 583, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 583, + 159 + ], + "type": "text", + "content": ", Yuming Chen" + }, + { + "bbox": [ + 53, + 142, + 583, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 583, + 159 + ], + "type": "text", + "content": ", Zheng Lin" + }, + { + "bbox": [ + 53, + 142, + 583, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 583, + 159 + ], + "type": "text", + "content": ", Yunheng Li" + }, + { + "bbox": [ + 53, + 142, + 583, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 583, + 159 + ], + "type": "text", + "content": ", Qibin Hou" + }, + { + "bbox": [ + 53, + 142, + 583, + 159 + ], + "type": "inline_equation", + "content": "^{2,1*}" + }, + { + "bbox": [ + 53, + 142, + 583, + 159 + ], + "type": "text", + "content": ", Ming-Ming Cheng" + }, + { + "bbox": [ + 53, + 142, + 583, + 159 + ], + "type": "inline_equation", + "content": "^{2,1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 181, + 171, + 455, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 171, + 455, + 185 + ], + "spans": [ + { + "bbox": [ + 181, + 171, + 455, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 181, + 171, + 455, + 185 + ], + "type": "text", + "content": "VCIP, College of Computer Science, Nankai University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 251, + 186, + 386, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 186, + 386, + 198 + ], + "spans": [ + { + "bbox": [ + 251, + 186, + 386, + 198 + ], + "type": "text", + "content": "2NKIARI, Shenzhen Futian" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 211, + 201, + 424, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 201, + 424, + 213 + ], + "spans": [ + { + "bbox": [ + 211, + 201, + 424, + 213 + ], + "type": "text", + "content": "https://github.com/zhangxuying1004/TeMO" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 267, + 290, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 267, + 290, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 267, + 290, + 542 + ], + "type": "text", + "content": "Recent progress in the text-driven 3D stylization of a single object has been considerably promoted by CLIP-based methods. However, the stylization of multi-object 3D scenes is still impeded in that the image-text pairs used for pretraining CLIP mostly consist of an object. Meanwhile, the local details of multiple objects may be susceptible to omission due to the existing supervision manner primarily relying on coarse-grained contrast of image-text pairs. To overcome these challenges, we present a novel framework, dubbed TeMO, to parse multi-object 3D scenes and edit their styles under the contrast supervision at multiple levels. We first propose a Decoupled Graph Attention (DGA) module to distinguishably reinforce the features of 3D surface points. Particularly, a cross-modal graph is constructed to align the object points accurately and noun phrases decoupled from the 3D mesh and textual description. Then, we develop a Cross-Grained Contrast (CGC) supervision system, where a fine-grained loss between the words in the textual description and the randomly rendered images are constructed to complement the coarse-grained loss. Extensive experiments show that our method can synthesize high-quality stylized content and outperform the existing methods over a wide range of multi-object 3D meshes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 584, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 584, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 584, + 287, + 693 + ], + "type": "text", + "content": "3D asset creation through stylization aims to synthesize stylized content on the bare meshes to conform to the given text descriptions [15, 19, 25], referring images [38, 50], or 3D shapes [34, 44]. This research plays an important role in a wide spectrum of applications, e.g., virtual/augmented reality [5, 9], gaming industries [49], and robotics [13]. Moreover, it also presents considerable potential and has attracted increasing attention in computer vision and graphics communities. Considering the ready availability and expres" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 255, + 380, + 316 + ], + "blocks": [ + { + "bbox": [ + 326, + 242, + 370, + 251 + ], + "lines": [ + { + "bbox": [ + 326, + 242, + 370, + 251 + ], + "spans": [ + { + "bbox": [ + 326, + 242, + 370, + 251 + ], + "type": "text", + "content": "Bare Mesh" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 255, + 380, + 316 + ], + "lines": [ + { + "bbox": [ + 317, + 255, + 380, + 316 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 380, + 316 + ], + "type": "image", + "image_path": "f29180e6a4141db5dce4da8a03a97063adf7a3c22a3239ab05d96c5051db29e4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 319, + 372, + 329 + ], + "lines": [ + { + "bbox": [ + 326, + 319, + 372, + 329 + ], + "spans": [ + { + "bbox": [ + 326, + 319, + 372, + 329 + ], + "type": "text", + "content": "two dragons" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 331, + 380, + 394 + ], + "blocks": [ + { + "bbox": [ + 317, + 331, + 380, + 394 + ], + "lines": [ + { + "bbox": [ + 317, + 331, + 380, + 394 + ], + "spans": [ + { + "bbox": [ + 317, + 331, + 380, + 394 + ], + "type": "image", + "image_path": "506d174053bf53c0bfb541df1933125738833fba512004468b45f7278c130c2c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 396, + 380, + 407 + ], + "lines": [ + { + "bbox": [ + 318, + 396, + 380, + 407 + ], + "spans": [ + { + "bbox": [ + 318, + 396, + 380, + 407 + ], + "type": "text", + "content": "person & dragon" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 420, + 547, + 487 + ], + "lines": [ + { + "bbox": [ + 305, + 420, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 305, + 420, + 547, + 487 + ], + "type": "text", + "content": "Figure 1. Visual comparisons between the existing 3D stylization methods (e.g., TANGO [15]) and our TeMO in multi-object scenes. For a scene with multiple objects of the same/different categories, existing methods are prone to interference between different properties of the objects, while our TeMO is able to accurately synthesize the desired stylized content for each object." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 396, + 255, + 455, + 314 + ], + "blocks": [ + { + "bbox": [ + 409, + 242, + 443, + 251 + ], + "lines": [ + { + "bbox": [ + 409, + 242, + 443, + 251 + ], + "spans": [ + { + "bbox": [ + 409, + 242, + 443, + 251 + ], + "type": "text", + "content": "TANGO" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 396, + 255, + 455, + 314 + ], + "lines": [ + { + "bbox": [ + 396, + 255, + 455, + 314 + ], + "spans": [ + { + "bbox": [ + 396, + 255, + 455, + 314 + ], + "type": "image", + "image_path": "6743cfa96f79c1bc2ff65bdc0e09206818df741a88b2a0387d97663843683260.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 396, + 332, + 454, + 393 + ], + "blocks": [ + { + "bbox": [ + 408, + 319, + 529, + 329 + ], + "lines": [ + { + "bbox": [ + 408, + 319, + 529, + 329 + ], + "spans": [ + { + "bbox": [ + 408, + 319, + 529, + 329 + ], + "type": "text", + "content": "\"a fire dragon and an ice dragon\"" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 396, + 332, + 454, + 393 + ], + "lines": [ + { + "bbox": [ + 396, + 332, + 454, + 393 + ], + "spans": [ + { + "bbox": [ + 396, + 332, + 454, + 393 + ], + "type": "image", + "image_path": "7e4ed212aff92a65e766545424f730e33ec26e60e64a1e03457c4e4579866b9b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 396, + 527, + 407 + ], + "lines": [ + { + "bbox": [ + 408, + 396, + 527, + 407 + ], + "spans": [ + { + "bbox": [ + 408, + 396, + 527, + 407 + ], + "type": "text", + "content": "\"an iron man and an ice dragon\"" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 474, + 255, + 533, + 316 + ], + "blocks": [ + { + "bbox": [ + 482, + 242, + 526, + 251 + ], + "lines": [ + { + "bbox": [ + 482, + 242, + 526, + 251 + ], + "spans": [ + { + "bbox": [ + 482, + 242, + 526, + 251 + ], + "type": "text", + "content": "Our TeMO" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 474, + 255, + 533, + 316 + ], + "lines": [ + { + "bbox": [ + 474, + 255, + 533, + 316 + ], + "spans": [ + { + "bbox": [ + 474, + 255, + 533, + 316 + ], + "type": "image", + "image_path": "b941afa272689e4a3611654fcdf7c2a90a2e88729401116365c66555255a45ea.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 318, + 529, + 328 + ], + "lines": [ + { + "bbox": [ + 475, + 318, + 529, + 328 + ], + "spans": [ + { + "bbox": [ + 475, + 318, + 529, + 328 + ], + "type": "text", + "content": "an ice dragon\"" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 475, + 332, + 532, + 393 + ], + "blocks": [ + { + "bbox": [ + 475, + 332, + 532, + 393 + ], + "lines": [ + { + "bbox": [ + 475, + 332, + 532, + 393 + ], + "spans": [ + { + "bbox": [ + 475, + 332, + 532, + 393 + ], + "type": "image", + "image_path": "7e6ba6910df82c4de5ef970fdebb9c9ca584504fcb2021bf9297e16bdfedce50.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 508, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 508, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 508, + 547, + 544 + ], + "type": "text", + "content": "siveness of text prompts as well as the popularity of large-scale Contrastive Language-Image Pre-training (CLIP) [28] model, we choose to work with text-driven 3D stylization." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": "Recent years have witnessed the emergence of a series of impressive works [8, 15, 23, 25], aiming to drive the advancement of text-driven 3D stylization. Existing methods usually adopt multi-layer perceptrons (MLPs) to predict the location attribute displacements of the bare mesh under the supervision of the contrastive loss in CLIP. We observe that these works focus on the stylization of a single 3D object and perform poorly on multiple objects, as shown in the second column of Fig. 1. We argue that two inherent characteristics of CLIP result in this issue: i) CLIP is mainly pretrained with image-text pairs mostly consisting of a single object; ii) CLIP loss employs global representation vectors from images and text to coarsely match these two modalities, which inevitably causes the loss of local details. More" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 189, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 189, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 189, + 712 + ], + "type": "text", + "content": "*Qibin Hou is the corresponding author." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19531" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": "over, the key to synthesizing desired styles for multiple 3D objects lies in the parsing of such 3D scenes and the multigrained supervision for details refinement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 287, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 287, + 326 + ], + "type": "text", + "content": "To simultaneously generate stylized content for multiple 3D objects, the primary step is to achieve accurate alignment between the objects in the 3D mesh and the target text. However, existing methods employ global semantics of the text to stylize a single object, which inevitably produces noises when stylizing the objects in multi-object scenes. To overcome this challenge, we propose to parse the 3D scene by introducing a Decoupled Graph Attention (DGA) module. Specifically, all noun phrases are decoupled from the text prompt and the mesh surface points of the current view are divided into several clusters as well. Then, a cross-modal graph is constructed to establish connections of the noun phrases to their corresponding object points while distancing them from the irrelevant ones. This graph enables the accurate interaction between two interrelated modalities. Finally, the surface point features of 3D objects can be reinforced by independent cross-attention fusions with their neighboring word nodes in the graph architecture." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 326, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 554 + ], + "type": "text", + "content": "Furthermore, we also design a Cross-Grained Contrast (CGC) loss to perform comprehensive cross-modal supervision for the stylization of multiple objects. The goal is to guide the network to generate more stylization details for multiple 3D objects to match the target text. Our loss consists of two parts, i.e., coarse-grained contrast and fine-grained contrast. In the former part, the text prompt is regarded as sentence-level supervision, which calculates the similarity between the 2D views rendered from the stylized 3D mesh and the text prompt using the global feature vectors from the CLIP model. In the latter part, we see the text prompt from the word level, and consider the similarities between each word of the sentence and the rendered images of the view sets. To be specific, we produce the word representations of the text prompt by taking the hidden states from the text encoder of CLIP. Motivated by the recent process in video-text retrieval [22], we calculate fine-grained loss via the weighted summation of the element in similarity vectors based on the importance of each word or image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 555, + 286, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 286, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 286, + 662 + ], + "type": "text", + "content": "Based on the well-designed DGA module and CGC loss, we propose a novel framework towards Text-Driven 3D stylization for Multi-Object Meshes, called TeMO. To validate the effectiveness of our TeMO, extensive experiments are conducted on various multi-object 3D scenes, as shown in the " + }, + { + "bbox": [ + 47, + 555, + 286, + 662 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 47, + 555, + 286, + 662 + ], + "type": "text", + "content": " column of Fig. 1. The experimental results demonstrate our TeMO is less susceptible to interference from multiple objects and can generate superior stylized assets compared with the existing 3D stylization methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 664, + 258, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 664, + 258, + 675 + ], + "spans": [ + { + "bbox": [ + 58, + 664, + 258, + 675 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "- We present a new 3D stylization framework, called TeMO. To the best of our knowledge, it is the first attempt to parse the objects in the text and 3D meshes and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 73, + 495, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 73, + 495, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 73, + 495, + 84 + ], + "type": "text", + "content": "generate stylizations for multi-object scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 84, + 545, + 167 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 84, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 84, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 306, + 84, + 545, + 132 + ], + "type": "text", + "content": "- We propose a Decoupled Graph Attention (DGA) module, which constructs a graph structure to align the surface points in the multi-object mesh and the noun phrases in the text prompt." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 133, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 545, + 167 + ], + "type": "text", + "content": "- We design a Cross-Grained Contrast (CGC) loss, in which the text is contrasted with the rendered images from sentence and word levels." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 180, + 392, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 392, + 192 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 392, + 192 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 200, + 468, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 200, + 468, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 200, + 468, + 213 + ], + "type": "text", + "content": "2.1. Text-Driven 3D Manipulation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 218, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 545, + 338 + ], + "type": "text", + "content": "Generating or editing 3D content according to a given prompt is a long-standing objective in computer vision and graphics [2, 39, 42]. Among all forms of the prompt, the text has garnered the most conspicuous gaze due to three reasons: i) Text descriptions are readily accessible from the existing corpus; ii) Text descriptions are particularly user-friendly since they are easily modifiable and can effectively express complex concepts related to stylizations; iii) The popularity of large-scale multimodal models [16, 28] has made achieving visual-language supervision possible." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 339, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 482 + ], + "type": "text", + "content": "Text2Mesh [25] proposes a neural-style field network to predict the color and displacement of mesh vertices. TANGO [15] proposes to disentangle the appearance style as the spatially varying bidirectional reflectance distribution, the local geometric variation, and the lighting condition. Then, X-Mesh [23] integrates the target text guidance by utilizing text-relevant spatial and channel-wise attention during vertex feature extraction. Motivated by the remarkable progress in text-driven 2D generation [29, 31], TEXTure [30] and Text2Tex [7] incorporate a pre-trained depth-aware image diffusion model to synthesize high-resolution partial textures from multiple viewpoints progressively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 483, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 626 + ], + "type": "text", + "content": "To make full use of the priors in the pre-trained 2D text-to-image diffusion model, DreamFusion [27] introduces a Score Distillation Sample (SDS) loss to perform text-to-3D synthesis. With the help of SDS loss, Latent-NeRF [24] and Fantasia3D [8] can generate 3D shapes and appearances for 3D objects. Despite achieving impressive results, these methods focus on the stylization of a single 3D object and rarely explore multi-object scenes. CLIP-Mesh [26] attempts to generate multiple 3D objects for target text. Nevertheless, the resulting content is not satisfactory. In this paper, we parse the objects described in rendered images and text prompts, aligned by two well-designed strategies." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 635, + 430, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 430, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 430, + 646 + ], + "type": "text", + "content": "2.2. Attention Mechanism" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "The concept of the attention mechanism was initially introduced in neural machine translation [1], where the weighted summation of the candidate vectors is calculated according to their importance scores. This technology has been extended to a myriad of tasks, e.g., natural language pro" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19532" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 68, + 531, + 238 + ], + "blocks": [ + { + "bbox": [ + 63, + 68, + 531, + 238 + ], + "lines": [ + { + "bbox": [ + 63, + 68, + 531, + 238 + ], + "spans": [ + { + "bbox": [ + 63, + 68, + 531, + 238 + ], + "type": "image", + "image_path": "dd6b711073a82492ad1972ca100baf91667c2de0af42ff850aa75393e3da8a2f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 243, + 547, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 547, + 300 + ], + "type": "text", + "content": "Figure 2. The overall architecture of the proposed TeMO framework. We first specify several cameras to cast rays toward the objects in the 3D mesh scene. Then, a surface point " + }, + { + "bbox": [ + 46, + 243, + 547, + 300 + ], + "type": "inline_equation", + "content": "x_{p}" + }, + { + "bbox": [ + 46, + 243, + 547, + 300 + ], + "type": "text", + "content": " and normal " + }, + { + "bbox": [ + 46, + 243, + 547, + 300 + ], + "type": "inline_equation", + "content": "n_p" + }, + { + "bbox": [ + 46, + 243, + 547, + 300 + ], + "type": "text", + "content": " can be attained from each ray intersected with the objects. These points and normals are fed to the attribute prediction network where the features of 3D objects are parsed and interacted with the decoupled text features via our proposed DGA module. Meanwhile, we employ a series of spherical Gaussians to represent the lighting. Finally, a differentiable SG render is adopted to render images, which are utilized to contrast with the text prompt by our designed CGC loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 314, + 289, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 289, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 289, + 542 + ], + "type": "text", + "content": "cessing [10, 21, 37], computer vision [14, 17, 47, 54], and multi-modal learning [20, 41, 48, 53]. For instance, Transformer [37] employs the self-attention operation to establish connections between words within a sentence and utilize the cross-attention mechanism to align source and target sentences. Non-local network [40] takes the lead in introducing self-attention to computer vision and achieves great success in video understanding and object detection. ViT [11] treats an image as a sequence of patches and employs a Transformer encoder based on self-attention to perform image classification. Swin Transformer [17] introduces shifted windows to enhance the local perception ability of self-attention. More recently, X-Mesh [23] designs a text-guided dynamic attention mechanism for vertex feature extraction of a 3D object. However, this guidance only relies on a text feature vector without considering the parsing of text and 3D scenes. In this paper, the multiple objects decoupled from the target text and 3D mesh are aligned via a cross-modal graph to achieve precise guidance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 551, + 234, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 234, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 234, + 563 + ], + "type": "text", + "content": "2.3. Multi-modal Contrastive Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "Contrastive learning has become an increasingly popular research topic in the multi-modal community due to its ability to align different modal representations. Based on this strategy, CLIP [28] is pre-trained on an abundance of image-text pairs, achieving great success in cross-modal supervision. TACo [45] presents a token-aware cascade contrastive learning based on the syntactic classes of words to achieve fine-grained semantic alignment in text-video retrieval. Concurrently, FILIP [46] proposes comparing the image patches with the words in the sentence. Regarding the text-driven 3D stylization, the CLIP loss, which calculates the similarity between the image and text vectors in the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 314, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 547, + 422 + ], + "type": "text", + "content": "embedding space of CLIP, is adopted by the vast majority of methods. Although achieving impressive results in stylizing a single object, these methods cannot be well adapted to scenes with multiple 3D objects. We argue an important reason for this issue is the loss of local details caused by such coarse-grained supervision. In this paper, we propose a cross-grained supervision strategy, which considers fine-grained and coarse-grained contrasts to achieve a more precise semantic alignment between rendered image and text." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 434, + 389, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 434, + 389, + 448 + ], + "spans": [ + { + "bbox": [ + 305, + 434, + 389, + 448 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 454, + 425, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 425, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 425, + 466 + ], + "type": "text", + "content": "3.1. Overall Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "text", + "content": "Fig. 2 shows the end-to-end architecture of our TeMO framework. Given a bare mesh and a text prompt containing multiple objects, the TeMO aims to synthesize stylization on the mesh to match the text descriptions. We employ a set of vertices " + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{e \\times 3}" + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "text", + "content": " and faces " + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "inline_equation", + "content": "F \\in \\{1, \\dots, e\\}^{u \\times 1}" + }, + { + "bbox": [ + 304, + 473, + 545, + 605 + ], + "type": "text", + "content": " to explicitly define the input triangle mesh, which is fixed throughout the training. Following TANGO [15], we disentangle the appearance style as the spatially varying bidirectional reflectance distribution function [4, 51, 52] (including diffuse, roughness, specular), the local geometric variation (normal map), and the lighting condition." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "We start by normalizing the vertex coordinates to lie inside a unit sphere. Then, we randomly sample points around the mesh using Gaussian distribution as camera positions to render images. Next, we can obtain a camera ray " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "R_{p} = \\{c + t\\nu_{p}\\}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " from the sampled camera position " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " and a pixel " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " in rendered images, where " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\nu_{p}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " is the direction of the ray. Further, ray casting [32] is used to seek out the ray and mesh's first intersection point and intersection face. Moreover, the normal " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "n_p\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " of the intersection face is" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19533" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 541, + 225 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 541, + 225 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 541, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 541, + 225 + ], + "type": "image", + "image_path": "0d31f3ed4fe35e9674f671b320e8017ae10095874286c3c822b432cb2036b6c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 230, + 546, + 253 + ], + "lines": [ + { + "bbox": [ + 46, + 230, + 546, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 546, + 253 + ], + "type": "text", + "content": "Figure 3. Construction pipeline of the cross-modal graph architecture in our DGA module. Note that " + }, + { + "bbox": [ + 46, + 230, + 546, + 253 + ], + "type": "inline_equation", + "content": "x_{p}" + }, + { + "bbox": [ + 46, + 230, + 546, + 253 + ], + "type": "text", + "content": ", the surface point of the 3D objects, and " + }, + { + "bbox": [ + 46, + 230, + 546, + 253 + ], + "type": "inline_equation", + "content": "w_{ij}" + }, + { + "bbox": [ + 46, + 230, + 546, + 253 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 46, + 230, + 546, + 253 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 230, + 546, + 253 + ], + "type": "text", + "content": "-th word in the " + }, + { + "bbox": [ + 46, + 230, + 546, + 253 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 230, + 546, + 253 + ], + "type": "text", + "content": "-th noun phrase, are connected together only if they correspond to the same object." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 268, + 249, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 249, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 249, + 281 + ], + "type": "text", + "content": "employed as surface normal at the point " + }, + { + "bbox": [ + 46, + 268, + 249, + 281 + ], + "type": "inline_equation", + "content": "x_{p}\\in \\mathbb{R}^{3}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 281, + 290, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 281, + 290, + 483 + ], + "spans": [ + { + "bbox": [ + 45, + 281, + 290, + 483 + ], + "type": "text", + "content": "To achieve multiview-consistent features, our TeMO is restricted to predicting the normal displacement as a function of the location, while allowing the color materials to be predicted as a function of both location and viewing direction. Therefore, our TeMO represented as MLPs includes two branches, i.e., normal branch " + }, + { + "bbox": [ + 45, + 281, + 290, + 483 + ], + "type": "inline_equation", + "content": "f_{n}(\\cdot)" + }, + { + "bbox": [ + 45, + 281, + 290, + 483 + ], + "type": "text", + "content": " and reflectance branch " + }, + { + "bbox": [ + 45, + 281, + 290, + 483 + ], + "type": "inline_equation", + "content": "f_{r}(\\cdot)" + }, + { + "bbox": [ + 45, + 281, + 290, + 483 + ], + "type": "text", + "content": ". Specifically, the former is utilized as the prediction of normal offset on the point " + }, + { + "bbox": [ + 45, + 281, + 290, + 483 + ], + "type": "inline_equation", + "content": "x_{p}" + }, + { + "bbox": [ + 45, + 281, + 290, + 483 + ], + "type": "text", + "content": ", and the latter is designed to predict surface reflectance coefficients of the material at the location " + }, + { + "bbox": [ + 45, + 281, + 290, + 483 + ], + "type": "inline_equation", + "content": "x_{p}" + }, + { + "bbox": [ + 45, + 281, + 290, + 483 + ], + "type": "text", + "content": ", i.e., diffuse, roughness, and specular. To synthesize high-frequency details, we also apply the Fourier positional encoding [35] to every input. In addition, the spherical Gaussian is employed to represent each light intensity " + }, + { + "bbox": [ + 45, + 281, + 290, + 483 + ], + "type": "inline_equation", + "content": "L_{i}(\\cdot)" + }, + { + "bbox": [ + 45, + 281, + 290, + 483 + ], + "type": "text", + "content": " due to its closed-form nature and analytical solution. Based on the attained geometric and appearance components, each pixel color in the rendered image can be calculated by a hemisphere renderer [15]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 488, + 287, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 488, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 51, + 488, + 287, + 513 + ], + "type": "interline_equation", + "content": "L _ {p} \\left(\\nu_ {p}, x _ {p}, n _ {p}\\right) = \\int_ {\\Omega} L _ {i} \\left(w _ {i}\\right) f _ {r} \\left(\\nu_ {p}, w _ {i}, x _ {p}\\right) \\left(w _ {i} \\cdot \\hat {n} _ {p}\\right) d w _ {i}, \\tag {1}", + "image_path": "8468718e7f2b0c0f505fc0080cf3df17c40c8b312e854116506f41e9844b03e3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 516, + 287, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 516, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 99, + 516, + 287, + 529 + ], + "type": "interline_equation", + "content": "\\hat {n} _ {p} = n _ {p} + f _ {n} \\left(x _ {p}, n _ {p}\\right), \\tag {2}", + "image_path": "7d14912455586c0835572cfcceb3b2a89c769f781f15b2b99a39dc10b5264293.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 533, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 287, + 571 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 533, + 287, + 571 + ], + "type": "inline_equation", + "content": "\\Omega = \\{w_{i}:w_{i}\\cdot \\hat{n}_{p}\\geq 0\\}" + }, + { + "bbox": [ + 47, + 533, + 287, + 571 + ], + "type": "text", + "content": " represents the hemisphere, " + }, + { + "bbox": [ + 47, + 533, + 287, + 571 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 47, + 533, + 287, + 571 + ], + "type": "text", + "content": " is the incident light direction, and " + }, + { + "bbox": [ + 47, + 533, + 287, + 571 + ], + "type": "inline_equation", + "content": "\\hat{n}_p" + }, + { + "bbox": [ + 47, + 533, + 287, + 571 + ], + "type": "text", + "content": " is the estimated normal on surface point " + }, + { + "bbox": [ + 47, + 533, + 287, + 571 + ], + "type": "inline_equation", + "content": "x_{p}" + }, + { + "bbox": [ + 47, + 533, + 287, + 571 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 575, + 200, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 200, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 200, + 588 + ], + "type": "text", + "content": "3.2. Decoupled Graph Attention" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "To achieve text-drive stylization for multiple 3D objects, the key issue that needs to be solved is the accurate alignment between the objects described in the text and those in the meshes. X-Mesh [23] has incorporated text-guided dynamic linear layers, in which the global representation vector of the target object in the text is utilized as guidance to acquire text-aware vertex features. Nevertheless, the global vector containing information about multiple objects is prone to mutual interference and produces semantic noises during guidance for multi-object scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 268, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 547, + 519 + ], + "type": "text", + "content": "To address this challenge, we propose to parse the objects in the text and mesh. We first extract the noun phrases modified by adjectives or prepositional phrases from the text using the NLTK tools [3]. Then, we employ the Gaussian Mixture Model (GMM) [56] to cluster the intersection point set " + }, + { + "bbox": [ + 304, + 268, + 547, + 519 + ], + "type": "inline_equation", + "content": "\\{x_{1},\\dots,x_{p},\\dots\\}" + }, + { + "bbox": [ + 304, + 268, + 547, + 519 + ], + "type": "text", + "content": " of the current ray and the mesh. Meanwhile, we can obtain a binary map of objects in the current view based on whether the ray intersects with the mesh. Further, we can decouple the objects in the binary map according to the clustered points and acquire several binary maps of individual objects. Based on the disentangled noun phrases and binary maps for multiple objects, we can match the correct pairs by their semantic similarities. As a result, the objects described in the text are aligned with their corresponding objects in the mesh, which are utilized to construct a cross-modal graph " + }, + { + "bbox": [ + 304, + 268, + 547, + 519 + ], + "type": "inline_equation", + "content": "G = (\\mathcal{V},\\mathcal{E})" + }, + { + "bbox": [ + 304, + 268, + 547, + 519 + ], + "type": "text", + "content": ", as shown in Fig. 3. To be specific, all surface point features and word features are considered as independent nodes to form the node set " + }, + { + "bbox": [ + 304, + 268, + 547, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 268, + 547, + 519 + ], + "type": "text", + "content": ". For the edge set " + }, + { + "bbox": [ + 304, + 268, + 547, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 268, + 547, + 519 + ], + "type": "text", + "content": ", the link between the surface point node and the word node will be built if the semantic objects they belong to are the same." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 519, + 547, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 547, + 592 + ], + "type": "text", + "content": "Under the setting of this cross-modal graph, we can individually perform cross-attention between the surface point nodes and their neighboring word nodes, where the parsed surface point features are used as queries, and the parsed text features serve as keys and values. The enhancement of the surface point node " + }, + { + "bbox": [ + 304, + 519, + 547, + 592 + ], + "type": "inline_equation", + "content": "v_{i} \\in \\mathbb{R}^{dim}" + }, + { + "bbox": [ + 304, + 519, + 547, + 592 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 366, + 597, + 545, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 597, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 366, + 597, + 545, + 624 + ], + "type": "interline_equation", + "content": "\\hat {v} _ {i} = \\sum_ {v _ {j} \\in \\operatorname {A d j} \\left(v _ {i}\\right)} \\alpha_ {i j} \\operatorname {L i n e a r} \\left(v _ {j}\\right), \\tag {3}", + "image_path": "d68f0f4660c2f010cb78b0c2e0634ff56ae9b5113cfcbc72dd9c663699ca2973.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 361, + 626, + 545, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 626, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 361, + 626, + 545, + 656 + ], + "type": "interline_equation", + "content": "\\alpha_ {i j} = \\frac {e ^ {\\mathcal {W} _ {i j}}}{\\sum_ {v _ {j} \\in \\operatorname {A d j} (v _ {i})} e \\mathcal {W} _ {i j}}, \\tag {4}", + "image_path": "319bfb3caddf4148dd7e1bf9d8e0aa0d9f4165bcca6a9ffd9e78a32e218b6f9f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 358, + 657, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 657, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 358, + 657, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {W} _ {i j} = \\frac {\\operatorname {L i n e a r} \\left(v _ {i}\\right) \\operatorname {L i n e a r} \\left(v _ {j}\\right) ^ {\\mathrm {T}}}{\\sqrt {d _ {l}}}, \\tag {5}", + "image_path": "55be82f308209dffd043433e53b6faf999e0027865d662f1a0cdf1bf3300be7d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Adj}(v_i)" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " is the adjacency nodes of " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Linear}(\\cdot)" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " represents a linear transformation. With this attention" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19534" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "mechanism, the surface point features of different objects in the mesh can be distinguishably reinforced under the guidance of the word features in the parsed text." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 119, + 242, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 119, + 242, + 132 + ], + "spans": [ + { + "bbox": [ + 47, + 119, + 242, + 132 + ], + "type": "text", + "content": "3.3. Cross-Grained Contrast Supervision" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 138, + 287, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 138, + 287, + 222 + ], + "spans": [ + { + "bbox": [ + 47, + 138, + 287, + 222 + ], + "type": "text", + "content": "To guide the optimization of the neural network for 3D stylization, the first step is to render the stylized 3D mesh from multiple 2D views. Most existing methods usually employ the visual encoder and text encoder of CLIP [28] to extract global feature vectors for the rendered image and target text, respectively, which are contrasted to perform cross-modal supervision via cosine similarity:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 232, + 287, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 232, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 110, + 232, + 287, + 259 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o a r s e}} = - \\frac {\\mathcal {F} _ {I} \\cdot \\mathcal {F} _ {T}}{\\| \\mathcal {F} _ {I} \\| _ {2} \\| \\mathcal {F} _ {T} \\| _ {2}}, \\tag {6}", + "image_path": "bd2ac7bc0cbe733f469b7298ae9add5a4901fba9381c93f10ad3dee4fe9ccd87.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 270, + 287, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 287, + 317 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 270, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_I\\in \\mathbb{R}^{512}" + }, + { + "bbox": [ + 47, + 270, + 287, + 317 + ], + "type": "text", + "content": " is the averaged feature vector of the images rendered from different views, " + }, + { + "bbox": [ + 47, + 270, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_T\\in \\mathbb{R}^{512}" + }, + { + "bbox": [ + 47, + 270, + 287, + 317 + ], + "type": "text", + "content": " denotes the global feature vector of the target text, and " + }, + { + "bbox": [ + 47, + 270, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 47, + 270, + 287, + 317 + ], + "type": "text", + "content": " represents the Euclidean norm function." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "text", + "content": "Although achieving impressive results for stylizing a single 3D object, these methods still have limitations in multi-object scenes. Considering that a single feature vector still represents a sentence describing multiple objects, the object details may be lost in large amounts. Therefore, such a coarse-grained contrast supervision is insufficient to guide the neural network in synthesizing photorealistic stylized content for multiple 3D objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 416, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 487 + ], + "type": "text", + "content": "To solve this issue, we construct a fine-grained contrast supervision to complement the coarse-grained one. Specifically, we first calculate the correlation map, i.e., " + }, + { + "bbox": [ + 46, + 416, + 287, + 487 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{n \\times m}" + }, + { + "bbox": [ + 46, + 416, + 287, + 487 + ], + "type": "text", + "content": " between the word features in the text and the visual features of the rendered images, which are also extracted from the text encoder and visual encoder of the CLIP:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 496, + 287, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 287, + 524 + ], + "type": "interline_equation", + "content": "\\mathcal {S} = \\frac {\\mathcal {I} \\cdot \\mathcal {T} ^ {\\mathrm {T}}}{\\| \\mathcal {I} \\| _ {2} \\| \\mathcal {T} \\| _ {2}}, \\tag {7}", + "image_path": "436c9cbbd3339cb31a3e08ddbf7ecfa5b722798485fa57aa70a2ca6fbd40d703.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 534, + 287, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 287, + 606 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 287, + 606 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 534, + 287, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{I} \\in \\mathbb{R}^{n \\times 512}" + }, + { + "bbox": [ + 47, + 534, + 287, + 606 + ], + "type": "text", + "content": " represents the features of the images rendered from " + }, + { + "bbox": [ + 47, + 534, + 287, + 606 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 534, + 287, + 606 + ], + "type": "text", + "content": " views, " + }, + { + "bbox": [ + 47, + 534, + 287, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{T} \\in \\mathbb{R}^{m \\times 512}" + }, + { + "bbox": [ + 47, + 534, + 287, + 606 + ], + "type": "text", + "content": " indicates the features of " + }, + { + "bbox": [ + 47, + 534, + 287, + 606 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 534, + 287, + 606 + ], + "type": "text", + "content": " words in the text. Then, we normalize the correlation matrix along the image axis and the text axis, respectively, to retrieve the text of interest and visual components. This process can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 117, + 617, + 287, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 617, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 117, + 617, + 287, + 641 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {I} (i) = \\frac {\\sum_ {k = 1} ^ {m} \\mathcal {S} (i , k)}{m}, \\tag {8}", + "image_path": "6d00c2ab19236ef3eb9a4cce72f54acbc20f28a2520c8dd5dc957ba1e692f091.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 643, + 287, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 643, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 116, + 643, + 287, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {T} (j) = \\frac {\\sum_ {k = 1} ^ {n} \\mathcal {S} (k , j)}{n}. \\tag {9}", + "image_path": "c31b1c06bdc49f22416e69359991c0a4fc353292674742175533b70eeb0177c6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "Inspired by [22], we further calculate an image-centered fine-grained contrast score and a text-centered fine-grained contrast score by the weighted summation of the similarity" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 487, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 487, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 487, + 83 + ], + "type": "text", + "content": "vectors, which can be formulated as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 362, + 91, + 545, + 122 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 91, + 545, + 122 + ], + "spans": [ + { + "bbox": [ + 362, + 91, + 545, + 122 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I} = \\sum_ {i = 1} ^ {n} \\frac {e ^ {\\mathcal {S} _ {I} (i)}}{\\sum_ {k = 1} ^ {n} e ^ {\\mathcal {S} _ {I} (k)}} \\mathcal {S} _ {I} (i), \\tag {10}", + "image_path": "3ffb79fd0637c79d38032985bd3cf9a0e4c0d21981dcf36dd25e8597fdfd5fca.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 125, + 545, + 157 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 125, + 545, + 157 + ], + "spans": [ + { + "bbox": [ + 361, + 125, + 545, + 157 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T} = \\sum_ {j = 1} ^ {m} \\frac {e ^ {\\mathcal {S} _ {T} (j)}}{\\sum_ {k = 1} ^ {m} e ^ {\\mathcal {S} _ {T} (k)}} \\mathcal {S} _ {T} (j), \\tag {11}", + "image_path": "0315a554cf0ad693c1c03a04368814a4f79ab8e76ed212363948c224da6c0593.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 164, + 545, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 164, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 305, + 164, + 545, + 212 + ], + "type": "text", + "content": "where the weights are defined as the degree of correlation between the central and another modality. Finally, we adopt the average value of these two scores as the fine-grained contrast loss, which can be defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 373, + 220, + 545, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 220, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 373, + 220, + 545, + 235 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f i n e}} = - \\left(\\mathcal {L} _ {I} + \\mathcal {L} _ {T}\\right) / 2. \\tag {12}", + "image_path": "aa067b00f5788d51f1ab1c292a184f7e4480e9ecce9670f66ae466de33e49a61.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 243, + 545, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 313 + ], + "type": "text", + "content": "The coarse-grained and fine-grained contrast supervision complement each other to build a cross-grained contrast supervision system. The former is utilized to align the global semantic information of the target text with the 3D objects, and the latter is used to achieve the local semantic alignment. This loss can be defined as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 361, + 323, + 545, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 323, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 361, + 323, + 545, + 336 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e g c s} = \\lambda_ {c} \\mathcal {L} _ {\\text {c o a r s e}} + \\lambda_ {f} \\mathcal {L} _ {\\text {f i n e}}, \\tag {13}", + "image_path": "d4b7b8b613fab3289f41d60ed1364345968fa5839444571e59635d7c8df6c965.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 344, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 344, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 344, + 545, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 344, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\lambda_{c}" + }, + { + "bbox": [ + 305, + 344, + 545, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 344, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\lambda_{f}" + }, + { + "bbox": [ + 305, + 344, + 545, + 380 + ], + "type": "text", + "content": " are two hyper-parameters to balance the cross-grained and the fine-grained losses, set to 1.0 and 0.33, respectively." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 390, + 387, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 390, + 387, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 390, + 387, + 403 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 410, + 413, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 410, + 413, + 423 + ], + "spans": [ + { + "bbox": [ + 306, + 410, + 413, + 423 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 434, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 434, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 305, + 434, + 545, + 564 + ], + "type": "text", + "content": "Datasets. To examine our method across a diverse set of 3D scenes, we first collect 3D object meshes from a variety of sources, i.e., COSEG [33], Thingi10K [55], Shapenet [6], Turbo Squid [36], and ModelNet [43]. Then, we randomly place several objects from the collected 3D set into a mesh using Blender. Note that we down-sample the number of meshes' vertices and faces to ensure the robustness of our TeMO to low-quality meshes and reduce the burden of GPU during the stylization. The meshes used in this paper contain an average of 79,303 faces, " + }, + { + "bbox": [ + 305, + 434, + 545, + 564 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 305, + 434, + 545, + 564 + ], + "type": "text", + "content": " non-manifold edges, " + }, + { + "bbox": [ + 305, + 434, + 545, + 564 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 305, + 434, + 545, + 564 + ], + "type": "text", + "content": " non-manifold vertices, and " + }, + { + "bbox": [ + 305, + 434, + 545, + 564 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 305, + 434, + 545, + 564 + ], + "type": "text", + "content": " boundaries." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. Following the TANGO [15] network, we adopt 3 linear layers with 256 dimensions to build the normal estimation branch. In the reflectance branch, the point features are extracted by 2 shared layers with 256 dimensions, followed by 3 exclusive layers to predict diffuse, specular, and roughness. The dimension of our DGA module is also set as 256. The word features in our DGA module are extracted from the text encoder of CLIP, and so are the ones in our CGC loss. We choose ViT-B/32 as the backbone of the pre-trained CLIP model in this paper, which is consistent with previous works [15, 23, 25]. We also process the rendered images with 2D augmentation strategies [12, 15]" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19535" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 67, + 137, + 144 + ], + "blocks": [ + { + "bbox": [ + 59, + 67, + 137, + 144 + ], + "lines": [ + { + "bbox": [ + 59, + 67, + 137, + 144 + ], + "spans": [ + { + "bbox": [ + 59, + 67, + 137, + 144 + ], + "type": "image", + "image_path": "66440921bdc8aa694a70f1e5f85db2247d6df544ff841f8c835d1c9c00ad1a31.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 145, + 133, + 155 + ], + "lines": [ + { + "bbox": [ + 77, + 145, + 133, + 155 + ], + "spans": [ + { + "bbox": [ + 77, + 145, + 133, + 155 + ], + "type": "text", + "content": "person & dragon" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 171, + 68, + 334, + 144 + ], + "blocks": [ + { + "bbox": [ + 171, + 68, + 334, + 144 + ], + "lines": [ + { + "bbox": [ + 171, + 68, + 334, + 144 + ], + "spans": [ + { + "bbox": [ + 171, + 68, + 334, + 144 + ], + "type": "image", + "image_path": "de78b7850f6a7c423f5aa2f33e5a1646f59043805612cb42960574f3532034e9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 144, + 317, + 155 + ], + "lines": [ + { + "bbox": [ + 167, + 144, + 317, + 155 + ], + "spans": [ + { + "bbox": [ + 167, + 144, + 317, + 155 + ], + "type": "text", + "content": "Text Prompt: \"an iron man and an ice dragon\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 63, + 158, + 123, + 234 + ], + "blocks": [ + { + "bbox": [ + 63, + 158, + 123, + 234 + ], + "lines": [ + { + "bbox": [ + 63, + 158, + 123, + 234 + ], + "spans": [ + { + "bbox": [ + 63, + 158, + 123, + 234 + ], + "type": "image", + "image_path": "56676eea3e4236823cdae3cad31bce5ef3a12267989770955f7b7d0d5fb83b16.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 238, + 126, + 247 + ], + "lines": [ + { + "bbox": [ + 86, + 238, + 126, + 247 + ], + "spans": [ + { + "bbox": [ + 86, + 238, + 126, + 247 + ], + "type": "text", + "content": "cat & horse" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 176, + 158, + 332, + 234 + ], + "blocks": [ + { + "bbox": [ + 176, + 158, + 332, + 234 + ], + "lines": [ + { + "bbox": [ + 176, + 158, + 332, + 234 + ], + "spans": [ + { + "bbox": [ + 176, + 158, + 332, + 234 + ], + "type": "image", + "image_path": "accb7cd2f28b76c881ed1c44e205f8996ddc39de04e544d1c5a645a18debaf6d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 237, + 324, + 247 + ], + "lines": [ + { + "bbox": [ + 167, + 237, + 324, + 247 + ], + "spans": [ + { + "bbox": [ + 167, + 237, + 324, + 247 + ], + "type": "text", + "content": "Text Prompt: \"a Garfield cat and a brown horse\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 69, + 251, + 134, + 328 + ], + "blocks": [ + { + "bbox": [ + 69, + 251, + 134, + 328 + ], + "lines": [ + { + "bbox": [ + 69, + 251, + 134, + 328 + ], + "spans": [ + { + "bbox": [ + 69, + 251, + 134, + 328 + ], + "type": "image", + "image_path": "8388728499c4e0e4734ff31c97ff5efb09b85a509b990e55cefe575168e4b1d3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 333, + 129, + 342 + ], + "lines": [ + { + "bbox": [ + 82, + 333, + 129, + 342 + ], + "spans": [ + { + "bbox": [ + 82, + 333, + 129, + 342 + ], + "type": "text", + "content": "vase & candle" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 356, + 545, + 369 + ], + "lines": [ + { + "bbox": [ + 47, + 356, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 545, + 369 + ], + "type": "text", + "content": "Figure 4. Given the same bare mesh, our TeMO produces various stylized contents for multi-object scenes to conform to the text prompts." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 176, + 251, + 326, + 328 + ], + "blocks": [ + { + "bbox": [ + 176, + 251, + 326, + 328 + ], + "lines": [ + { + "bbox": [ + 176, + 251, + 326, + 328 + ], + "spans": [ + { + "bbox": [ + 176, + 251, + 326, + 328 + ], + "type": "image", + "image_path": "dae3661b5d0f97498c902aeacc951df596ea72da090516aea6f2866b0426631d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 332, + 323, + 342 + ], + "lines": [ + { + "bbox": [ + 167, + 332, + 323, + 342 + ], + "spans": [ + { + "bbox": [ + 167, + 332, + 323, + 342 + ], + "type": "text", + "content": "Text Prompt: \"a cactus vase and a silver candle\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 373, + 68, + 533, + 144 + ], + "blocks": [ + { + "bbox": [ + 373, + 68, + 533, + 144 + ], + "lines": [ + { + "bbox": [ + 373, + 68, + 533, + 144 + ], + "spans": [ + { + "bbox": [ + 373, + 68, + 533, + 144 + ], + "type": "image", + "image_path": "00a50e88bd42ccfb55b26c012674f13ba577db6c468419f652c7224183570c00.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 144, + 507, + 155 + ], + "lines": [ + { + "bbox": [ + 362, + 144, + 507, + 155 + ], + "spans": [ + { + "bbox": [ + 362, + 144, + 507, + 155 + ], + "type": "text", + "content": "Text Prompt: \"a superman and a fire dragon\"" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 371, + 159, + 525, + 234 + ], + "blocks": [ + { + "bbox": [ + 371, + 159, + 525, + 234 + ], + "lines": [ + { + "bbox": [ + 371, + 159, + 525, + 234 + ], + "spans": [ + { + "bbox": [ + 371, + 159, + 525, + 234 + ], + "type": "image", + "image_path": "dcb6da02a32a4a26a4d912c5e26e6b3abfc385a625041cd82849f0dde536c7a9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 237, + 526, + 247 + ], + "lines": [ + { + "bbox": [ + 362, + 237, + 526, + 247 + ], + "spans": [ + { + "bbox": [ + 362, + 237, + 526, + 247 + ], + "type": "text", + "content": "Text Prompt: \"a ginger cat and an astronaut horse\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 375, + 251, + 525, + 328 + ], + "blocks": [ + { + "bbox": [ + 375, + 251, + 525, + 328 + ], + "lines": [ + { + "bbox": [ + 375, + 251, + 525, + 328 + ], + "spans": [ + { + "bbox": [ + 375, + 251, + 525, + 328 + ], + "type": "image", + "image_path": "052f030a88a0ac9f058396d72655a81eae0d4d2cb3470e2b931253e1f146b67f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 332, + 527, + 342 + ], + "lines": [ + { + "bbox": [ + 362, + 332, + 527, + 342 + ], + "spans": [ + { + "bbox": [ + 362, + 332, + 527, + 342 + ], + "type": "text", + "content": "Text Prompt: \"a wicker vase and a candle in jeans\"" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "text", + "content": "before feeding them into the pre-trained CLIP model. Our TeMO model is optimized with the AdamW [18] strategy for 1500 iterations, where the learning rate is initialized to " + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "text", + "content": " and decayed by 0.7 every 500 iterations. The entire training process takes about 10 minutes on a single NVIDIA RTX 3090 GPU." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 463, + 176, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 463, + 176, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 463, + 176, + 475 + ], + "type": "text", + "content": "4.2. Qualitative Evaluation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 481, + 288, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 288, + 601 + ], + "type": "text", + "content": "We conduct visualization experiments on a wide spectrum of multi-object scenes to verify the effectiveness of our TeMO. However, we observe that the 3D symmetry prior used widely in previous works [15, 25] can cause interference between different parts during the stylization process of multiple objects. We argue that the multiple objects of the meshes used in this paper are randomly placed to simulate a real 3D scene rather than along the " + }, + { + "bbox": [ + 46, + 481, + 288, + 601 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 481, + 288, + 601 + ], + "type": "text", + "content": "-axis. To avoid this issue, we remove this prior in our TeMO and previous methods involved in the comparison." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Neural Stylization and Controls. We present the stylization results of our TeMO driven by different text prompts for the same multi-object mesh in Fig. 4. As shown in the " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " row where the 3D scene is composed of a person and a dragon, our TeMO can accurately distinguish between the person object and the dragon object and appropriately stylize different body parts for them according to the semantic roles described in each text prompt. Meanwhile, our TeMO also synthesizes desired stylizations for the 3D objects in" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 384, + 545, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 444 + ], + "type": "text", + "content": "the cat-horse mesh and vase-candle mesh, as shown in the " + }, + { + "bbox": [ + 304, + 384, + 545, + 444 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 304, + 384, + 545, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 384, + 545, + 444 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 304, + 384, + 545, + 444 + ], + "type": "text", + "content": " rows. These experimental results demonstrate that our TeMO method can generate photorealistic details with fine granularity and maintain global semantic understanding for the given multi-object 3D scene." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": "Qualitative Comparisons. We provide the visual comparisons of prediction results between our TeMO and previous pioneering works in text-driven 3D object stylization, including Text2Mesh [25], TANGO [15], and X-Mesh [23]. To ensure a fair comparison, we adopt the official implementations of these methods and also train them with the default settings without the symmetry prior. The experimental results show it is a real struggle for Text2Mesh [25] and TANGO [15] to understand the detailed semantics of the text prompt with multiple objects. As shown in the " + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": " row of Fig. 5 where the 3D scene contains two objects of the same category, given a text prompt \"a fire dragon and an ice dragon\", they tend to capture the \"ice\" property, missing the \"fire\" property. For a 3D scene containing two objects of different categories, they are prone to mixing the properties of these objects, as shown in the " + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": " row where the text prompt is \"a wood vase and a brick candle\". Therefore, the stylized assets they generate for these multi-object scenes are unsatisfactory. X-Mesh generates more accurate results that align with the text prompts, as shown in the " + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": " rows, which can be attributed to incorporating the text vector while extracting vertex features. However, it can pro" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19536" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 57, + 138, + 190 + ], + "blocks": [ + { + "bbox": [ + 55, + 57, + 138, + 190 + ], + "lines": [ + { + "bbox": [ + 55, + 57, + 138, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 57, + 138, + 190 + ], + "type": "image", + "image_path": "fa8fe2140bb9db1d945a3ae5501df2a41c550a20bdd233b0458df72f74107771.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 158, + 57, + 237, + 190 + ], + "blocks": [ + { + "bbox": [ + 158, + 57, + 237, + 190 + ], + "lines": [ + { + "bbox": [ + 158, + 57, + 237, + 190 + ], + "spans": [ + { + "bbox": [ + 158, + 57, + 237, + 190 + ], + "type": "image", + "image_path": "dffe4c493395d99a1da0533779e4cbae62cad93212cd00dd3a54d0f3d37ce2ac.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 194, + 412, + 205 + ], + "lines": [ + { + "bbox": [ + 181, + 194, + 412, + 205 + ], + "spans": [ + { + "bbox": [ + 181, + 194, + 412, + 205 + ], + "type": "text", + "content": "(a) Mesh: two dragons; Text Prompt: \"a fire dragon and an ice dragon\"." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 256, + 57, + 334, + 189 + ], + "blocks": [ + { + "bbox": [ + 256, + 57, + 334, + 189 + ], + "lines": [ + { + "bbox": [ + 256, + 57, + 334, + 189 + ], + "spans": [ + { + "bbox": [ + 256, + 57, + 334, + 189 + ], + "type": "image", + "image_path": "ef2eba99c8b0ca505c24f32150933a2cb9b0b69abbe17f065b53c17c5917a329.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 354, + 57, + 433, + 189 + ], + "blocks": [ + { + "bbox": [ + 354, + 57, + 433, + 189 + ], + "lines": [ + { + "bbox": [ + 354, + 57, + 433, + 189 + ], + "spans": [ + { + "bbox": [ + 354, + 57, + 433, + 189 + ], + "type": "image", + "image_path": "b1ebfbda641f328416a346ee6d30f54725eb6e91d692d5a5828d0af7dcac0851.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 452, + 57, + 529, + 190 + ], + "blocks": [ + { + "bbox": [ + 452, + 57, + 529, + 190 + ], + "lines": [ + { + "bbox": [ + 452, + 57, + 529, + 190 + ], + "spans": [ + { + "bbox": [ + 452, + 57, + 529, + 190 + ], + "type": "image", + "image_path": "1519451221330d6b8d412058bbfda9744d2c7bfb32584a4a313c3e27859d0a4a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 207, + 129, + 328 + ], + "blocks": [ + { + "bbox": [ + 57, + 207, + 129, + 328 + ], + "lines": [ + { + "bbox": [ + 57, + 207, + 129, + 328 + ], + "spans": [ + { + "bbox": [ + 57, + 207, + 129, + 328 + ], + "type": "image", + "image_path": "9ab4c588d5d5c274e4356975293a8389f75fa609dd9a9351cea38e304dd1fd36.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 157, + 207, + 233, + 327 + ], + "blocks": [ + { + "bbox": [ + 157, + 207, + 233, + 327 + ], + "lines": [ + { + "bbox": [ + 157, + 207, + 233, + 327 + ], + "spans": [ + { + "bbox": [ + 157, + 207, + 233, + 327 + ], + "type": "image", + "image_path": "7ba345fe11419efb040709351da1cab36f3aa5e14ba84011484e9da63026f6db.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 331, + 416, + 342 + ], + "lines": [ + { + "bbox": [ + 178, + 331, + 416, + 342 + ], + "spans": [ + { + "bbox": [ + 178, + 331, + 416, + 342 + ], + "type": "text", + "content": "(b) Mesh: vase & candle; Text Prompt: \"a wood vase and a brick candle\"." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 256, + 207, + 327, + 327 + ], + "blocks": [ + { + "bbox": [ + 256, + 207, + 327, + 327 + ], + "lines": [ + { + "bbox": [ + 256, + 207, + 327, + 327 + ], + "spans": [ + { + "bbox": [ + 256, + 207, + 327, + 327 + ], + "type": "image", + "image_path": "ad1cac078020fbcc33b971612e8f13d2af6a4702bb55f832c1fd552fb9416d64.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 354, + 207, + 424, + 327 + ], + "blocks": [ + { + "bbox": [ + 354, + 207, + 424, + 327 + ], + "lines": [ + { + "bbox": [ + 354, + 207, + 424, + 327 + ], + "spans": [ + { + "bbox": [ + 354, + 207, + 424, + 327 + ], + "type": "image", + "image_path": "95491438540621726297fe5c6c0ab4a41fc472c7052b5e463e5b0963ac058fb2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 452, + 207, + 522, + 328 + ], + "blocks": [ + { + "bbox": [ + 452, + 207, + 522, + 328 + ], + "lines": [ + { + "bbox": [ + 452, + 207, + 522, + 328 + ], + "spans": [ + { + "bbox": [ + 452, + 207, + 522, + 328 + ], + "type": "image", + "image_path": "5c058d4d7387ab23d6b3db25210bf878b1895b3f522acb39fbafdb62b6174107.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 55, + 344, + 148, + 454 + ], + "blocks": [ + { + "bbox": [ + 55, + 344, + 148, + 454 + ], + "lines": [ + { + "bbox": [ + 55, + 344, + 148, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 148, + 454 + ], + "type": "image", + "image_path": "d805d17d418b5c86e6b1b8b73d701e32604123229d28b61bc3606edd2a8fb838.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 458, + 454, + 469 + ], + "lines": [ + { + "bbox": [ + 138, + 458, + 454, + 469 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 454, + 469 + ], + "type": "text", + "content": "(c) Mesh: person & dragon & whale; Text Prompt: \"a superman, a fire dragon, and an ice whale\"." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 156, + 344, + 249, + 454 + ], + "blocks": [ + { + "bbox": [ + 156, + 344, + 249, + 454 + ], + "lines": [ + { + "bbox": [ + 156, + 344, + 249, + 454 + ], + "spans": [ + { + "bbox": [ + 156, + 344, + 249, + 454 + ], + "type": "image", + "image_path": "0ea2e3d9fefb5c4cf0477dd36593bb1fe7db51ff4540aa3165e3cbcaaf439544.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 478, + 547, + 502 + ], + "lines": [ + { + "bbox": [ + 46, + 478, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 547, + 502 + ], + "type": "text", + "content": "Figure 5. Visual comparisons of our TeMO with previous text-driven 3D stylization methods on several multi-object scenes, including two objects of the same or different categories, and three different objects." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 255, + 344, + 347, + 453 + ], + "blocks": [ + { + "bbox": [ + 255, + 344, + 347, + 453 + ], + "lines": [ + { + "bbox": [ + 255, + 344, + 347, + 453 + ], + "spans": [ + { + "bbox": [ + 255, + 344, + 347, + 453 + ], + "type": "image", + "image_path": "335c60e6fc3e0894e7eb0ddd01097f70d14291833281797915760915310ec445.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 351, + 344, + 443, + 453 + ], + "blocks": [ + { + "bbox": [ + 351, + 344, + 443, + 453 + ], + "lines": [ + { + "bbox": [ + 351, + 344, + 443, + 453 + ], + "spans": [ + { + "bbox": [ + 351, + 344, + 443, + 453 + ], + "type": "image", + "image_path": "c9660c91ddb4e36bad560489363f0d8da85f9855c056141c4e05ec3a88521e22.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 449, + 344, + 541, + 453 + ], + "blocks": [ + { + "bbox": [ + 449, + 344, + 541, + 453 + ], + "lines": [ + { + "bbox": [ + 449, + 344, + 541, + 453 + ], + "spans": [ + { + "bbox": [ + 449, + 344, + 541, + 453 + ], + "type": "image", + "image_path": "5862fc8ad0f9f35f165b3012959025f2649a045bafdbfdf344e8684b516def23.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 516, + 289, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 289, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 289, + 649 + ], + "type": "text", + "content": "duce semantic noises due to its utilization of the text vector containing attributes of multiple objects to process all vertex features. With an increasing number of objects, it will also encounter challenges related to comprehending text details and the alignment between the text and 3D objects. As shown in the " + }, + { + "bbox": [ + 46, + 516, + 289, + 649 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 46, + 516, + 289, + 649 + ], + "type": "text", + "content": " row, this method still fails to generate stylized assets without mixed properties. In contrast, our TeMO equipped with 3D scene parsing and multi-grained supervision, is able to generate photorealistic stylized content for each object in these 3D scenes to conform to the descriptions in the text prompts." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 654, + 183, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 183, + 667 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 183, + 667 + ], + "type": "text", + "content": "4.3.Quantitative Evaluation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Objective Metric. We adopt the CLIP score to objectively evaluate the semantic alignment achieved by our TeMO and recent 3D stylization methods. Specifically, 8 views spaced" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 516, + 547, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 547, + 601 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + }, + { + "bbox": [ + 304, + 516, + 547, + 601 + ], + "type": "text", + "content": " around the stylized meshes are chosen to obtain the rendered 2D images. Then, the visual objects are compared with the textual objects in CLIP's embedding space via the cosine function. As shown in the " + }, + { + "bbox": [ + 304, + 516, + 547, + 601 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 304, + 516, + 547, + 601 + ], + "type": "text", + "content": " column of Tab. 1, our TeMO surpasses previous methods by a large margin. These results demonstrate the superiority of our TeMO over existing methods on multi-object stylization." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "User Study. We further conduct a user study to evaluate these 3D stylization methods subjectively. We randomly select 10 mesh-text pairs and recruit 60 users to evaluate the quality of the stylization assets generated by our TeMO and previous methods. Particularly, the participants include experts in the field and individuals without specific background knowledge. Moreover, each of them will be asked three questions [25]: (Q1) \"How natural is the output results?\"(Q2) \"How well does the output match the original" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19537" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 141, + 151 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 141, + 151 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 141, + 151 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 141, + 151 + ], + "type": "image", + "image_path": "9047c9124b16781a232ba564a837fe67ffcd09d9dc577149c7e9ab75ee52071b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 153, + 114, + 164 + ], + "lines": [ + { + "bbox": [ + 70, + 153, + 114, + 164 + ], + "spans": [ + { + "bbox": [ + 70, + 153, + 114, + 164 + ], + "type": "text", + "content": "Input Mesh" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 70, + 255, + 149 + ], + "blocks": [ + { + "bbox": [ + 179, + 70, + 255, + 149 + ], + "lines": [ + { + "bbox": [ + 179, + 70, + 255, + 149 + ], + "spans": [ + { + "bbox": [ + 179, + 70, + 255, + 149 + ], + "type": "image", + "image_path": "44e2474405cf9ffb61160181f4083aa5bbedbc47695458f432c68ef65e7392c7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 154, + 227, + 163 + ], + "lines": [ + { + "bbox": [ + 195, + 154, + 227, + 163 + ], + "spans": [ + { + "bbox": [ + 195, + 154, + 227, + 163 + ], + "type": "text", + "content": "Baseline" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 272, + 70, + 347, + 149 + ], + "blocks": [ + { + "bbox": [ + 272, + 70, + 347, + 149 + ], + "lines": [ + { + "bbox": [ + 272, + 70, + 347, + 149 + ], + "spans": [ + { + "bbox": [ + 272, + 70, + 347, + 149 + ], + "type": "image", + "image_path": "05c809defa7cd9ba5d3381bfcc56e23f3c6eeee6703d49dbba6bf041ceb46982.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 153, + 362, + 163 + ], + "lines": [ + { + "bbox": [ + 271, + 153, + 362, + 163 + ], + "spans": [ + { + "bbox": [ + 271, + 153, + 362, + 163 + ], + "type": "text", + "content": "Baseline + DGA module" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 365, + 70, + 439, + 149 + ], + "blocks": [ + { + "bbox": [ + 365, + 70, + 439, + 149 + ], + "lines": [ + { + "bbox": [ + 365, + 70, + 439, + 149 + ], + "spans": [ + { + "bbox": [ + 365, + 70, + 439, + 149 + ], + "type": "image", + "image_path": "1f6955df3bd31187a4b436fabd7a2eb598d7630c3bc82ad2ca2968cafa777b99.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 153, + 447, + 163 + ], + "lines": [ + { + "bbox": [ + 371, + 153, + 447, + 163 + ], + "spans": [ + { + "bbox": [ + 371, + 153, + 447, + 163 + ], + "type": "text", + "content": "Baseline + CGC loss" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 455, + 70, + 531, + 150 + ], + "blocks": [ + { + "bbox": [ + 455, + 70, + 531, + 150 + ], + "lines": [ + { + "bbox": [ + 455, + 70, + 531, + 150 + ], + "spans": [ + { + "bbox": [ + 455, + 70, + 531, + 150 + ], + "type": "image", + "image_path": "37db2236ad4868c56b8561f82bb54e62e56e161fd004edd0e38873e1c2269ef2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 153, + 517, + 163 + ], + "lines": [ + { + "bbox": [ + 475, + 153, + 517, + 163 + ], + "spans": [ + { + "bbox": [ + 475, + 153, + 517, + 163 + ], + "type": "text", + "content": "Our TeMO" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 171, + 544, + 183 + ], + "lines": [ + { + "bbox": [ + 47, + 171, + 544, + 183 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 544, + 183 + ], + "type": "text", + "content": "Figure 6. Ablation experiments on the proposed designs of our TeMO. Mesh: two dragons; Text Prompt: \"a fire dragon and an ice dragon\"." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 48, + 244, + 284, + 318 + ], + "blocks": [ + { + "bbox": [ + 46, + 196, + 286, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 196, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 286, + 240 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons of our TeMO and previous text-driven 3D stylization methods in multi-object scenes, including an objective alignment score (0-1) and three subjective opinion scores (1-5). Note that the higher these metrics, the better the method." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 244, + 284, + 318 + ], + "lines": [ + { + "bbox": [ + 48, + 244, + 284, + 318 + ], + "spans": [ + { + "bbox": [ + 48, + 244, + 284, + 318 + ], + "type": "table", + "html": "
AlignmentUser-Q1User-Q2User-Q3
Text2Mesh [25]0.2621.7501.5061.472
TANGO [15]0.2742.4062.4502.539
X-Mesh [23]0.2651.8391.7221.761
Our TeMO0.2853.3443.3113.261
", + "image_path": "85ed0290c57419d989f0e023f318e1e8cad5882b8442a487f5af38899476cad6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 332, + 286, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 332, + 286, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 286, + 417 + ], + "type": "text", + "content": "content?\" (Q3) \"How well does the output match the target style?\", and then assign a score (1-5) to them. We report the mean opinion scores in parentheses for each factor averaged across all style outputs. As shown in Tab. 1, our TeMO still outperforms other methods across all questions. Therefore, the 3D assets generated by our method are more in line with people's understanding of the text prompts." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 424, + 148, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 424, + 148, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 424, + 148, + 436 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "text", + "content": "To verify the effectiveness of the proposed designs in our TeMO, we conduct ablation studies by gradually adding them to our baseline model, i.e., TANGO [15]. We chose the two-dragon mesh with the text prompt \"a fire dragon and an ice dragon\", and the experimental results are shown in Fig. 6. Compared to the baseline model, introducing our DGA module enables the model to distinguish two dragons, yet it falls short in endowing them with precise texture details. Meanwhile, incorporating our CGC loss facilitates the model to capture more semantic details, e.g., \"fire\" and \"ice\". Nevertheless, it fails to distinguish the two objects. It is noteworthy that the model equipped with these two designs together is not only capable of accurately distinguishing between two objects but can also synthesize high-quality texture details for them. These experiments indicate that our DGA module and CGC loss can effectively assist the model in generating desired stylized content for multiple 3D objects to conform to the target text." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 669, + 209, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 209, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 209, + 681 + ], + "type": "text", + "content": "5. Limitation and Future Work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "Despite achieving excellent results on text-driven multi-object stylization, our TeMO framework still has a few lim" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 198, + 504, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 198, + 504, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 504, + 209 + ], + "type": "text", + "content": "itations, which can also facilitate future research:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 215, + 545, + 388 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 304, + 215, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 215, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 215, + 545, + 299 + ], + "type": "text", + "content": "1) 3D Symmetry Prior. As stated in Sec. 4.2, our TeMO fails to incorporate 3D symmetry prior, whose important role has been demonstrated by Text2Mesh [25] in promoting style consistency of a single object. To generate more photorealistic stylization assets for multi-object scenes, it will be valuable to calculate symmetry planes for each object and apply symmetry priors to them." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 304, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 388 + ], + "type": "text", + "content": "2) Diffusion Model. We observe that current diffusion technologies struggle to generate multi-object images according to the text prompt, which hinders the application of diffusion-based stylization methods in multi-object 3D scenes. We argue it would be interesting to extend the concept of scene parsing to the diffusion models for the release of their potential in multi-object editing or generation." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 400, + 378, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 400, + 378, + 412 + ], + "spans": [ + { + "bbox": [ + 306, + 400, + 378, + 412 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 421, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 545, + 624 + ], + "type": "text", + "content": "In this paper, we present TeMO, an innovative framework proposing scene parsing and multi-grained cross-modal supervision to achieve text-driven multi-object 3D stylization for the first time. Specifically, we first develop a DGA module to precisely align the objects in the 3D mesh and the text prompt and enhance the 3D point features with the word features belonging to the same object as them. Then, we design a CGC loss, in which the fine-grained loss at the local level and coarse-grained contrast loss at the global level are both constructed and complement each other. Further, extensive experiments are conducted to demonstrate the effectiveness and superiority of our methods over the existing methods among a wide range of multi-object 3D scenes. We believe it is promising to achieve content editing of multiple objects in 3D scenes simultaneously, and we hope the scene-parsing perspective provided by the proposed TeMO framework will inspire future works." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgments. This research was supported by NSFC (NO. 62225604, NO. 62276145), the Fundamental Research Funds for the Central Universities (Nankai University, 070-63223049), CAST through Young Elite Scientist Sponsorship Program (No. YESS20210377). Computations were supported by the Supercomputing Center of Nankai University (NKSC)." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19538" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In ICLR, 2014. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "type": "text", + "content": "[2] Guirong Bai, Shizhu He, Kang Liu, and Jun Zhao. Example-guided stylized response generation in zero-shot setting. Science China Information Sciences, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 156, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 156, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 53, + 156, + 287, + 186 + ], + "type": "text", + "content": "[3] Edward Loper Bird, Steven and Ewan Klein. Natural language processing with python. O'Reilly Media Inc, 2009. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 186, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 287, + 228 + ], + "type": "text", + "content": "[4] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In IEEE ICCV, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 229, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 229, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 229, + 287, + 251 + ], + "type": "text", + "content": "[5] Arthur Caetano and Misha Sra. Arfy: A pipeline for adapting 3d scenes to augmented reality. In ACM UIST, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 251, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 287, + 304 + ], + "type": "text", + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 304, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 287, + 346 + ], + "type": "text", + "content": "[7] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In IEEE ICCV, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 347, + 287, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 287, + 389 + ], + "type": "text", + "content": "[8] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In IEEE ICCV, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 390, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 390, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 390, + 287, + 434 + ], + "type": "text", + "content": "[9] Shaoyu Chen, Budmonde Duinkharjav, Xin Sun, Li-Yi Wei, Stefano Petrangeli, Jose Echevarria, Claudio Silva, and Qi Sun. Instant reality: Gaze-contingent perceptual optimization for 3d virtual reality streaming. IEEE TVCG, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 434, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 465 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 465, + 287, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 287, + 519 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 519, + 287, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 519, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 48, + 519, + 287, + 561 + ], + "type": "text", + "content": "[12] Kevin Frans, Lisa Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. Advances in Neural Information Processing Systems, 2022. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 562, + 287, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 562, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 48, + 562, + 287, + 604 + ], + "type": "text", + "content": "[13] Mengdi Han, Xiaogang Guo, Xuexian Chen, Cunman Liang, Hangbo Zhao, Qihui Zhang, Wubin Bai, Fan Zhang, Heming Wei, Changsheng Wu, et al. Submillimeter-scale multitematerial terrestrial robots. Science Robotics, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 605, + 287, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 605, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 48, + 605, + 287, + 625 + ], + "type": "text", + "content": "[14] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In IEEE CVPR, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "text", + "content": "[15] Jiabao Lei, Yabin Zhang, Kui Jia, et al. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[16] Zheng Li, Xiang Li, Xinyi Fu, Xin Zhang, Weiqiang Wang, Shuo Chen, and Jian Yang. Promptkd: Unsupervised prompt distillation for vision-language models. arXiv preprint arXiv:2403.02781, 2024. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 704 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE ICCV, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 117, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 117, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 117, + 545, + 137 + ], + "type": "text", + "content": "[18] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 137, + 545, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 137, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 308, + 137, + 545, + 180 + ], + "type": "text", + "content": "[19] Haoyu Lu, Yuqi Huo, Mingyu Ding, Nanyi Fei, and Zhiwu Lu. Cross-modal contrastive learning for generalizable and efficient image-text retrieval. Machine Intelligence Research, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 180, + 545, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 180, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 308, + 180, + 545, + 223 + ], + "type": "text", + "content": "[20] Yunpeng Luo, Jiayi Ji, Xiaoshuai Sun, Liujuan Cao, Yongjian Wu, Feiyue Huang, Chia-Wen Lin, and Rongrong Ji. Dual-level collaborative transformer for image captioning. In AAAI, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 223, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 223, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 308, + 223, + 545, + 255 + ], + "type": "text", + "content": "[21] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. In ACL EMNLP, 2015. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 255, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 255, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 545, + 297 + ], + "type": "text", + "content": "[22] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In ACM MM, 2022. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 297, + 545, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 351 + ], + "type": "text", + "content": "[23] Yiwei Ma, Xiaojing Zhang, Xiaoshuai Sun, Jiayi Ji, Haowei Wang, Guannan Jiang, Weilin Zhuang, and Rongrong Ji. X-mesh: Towards fast and accurate text-driven 3d stylization via dynamic textual guidance. In IEEE ICCV, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 351, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 351, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 351, + 545, + 384 + ], + "type": "text", + "content": "[24] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In IEEE CVPR, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "text", + "content": "[25] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In IEEE CVPR, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "text", + "content": "[26] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In ACM SIGGRAPH, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 458, + 545, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 458, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 308, + 458, + 545, + 490 + ], + "type": "text", + "content": "[27] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "type": "text", + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 544, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 586 + ], + "type": "text", + "content": "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 586, + 545, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 586, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 308, + 586, + 545, + 619 + ], + "type": "text", + "content": "[30] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 619, + 545, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 619, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 308, + 619, + 545, + 651 + ], + "type": "text", + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE CVPR, 2022. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 651, + 545, + 672 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 651, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 308, + 651, + 545, + 672 + ], + "type": "text", + "content": "[32] Scott D Roth. Ray casting for modeling solids. Computer graphics and image processing, 1982. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 672, + 545, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 672, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 308, + 672, + 545, + 704 + ], + "type": "text", + "content": "[33] Oana Sidi, Oliver Van Kaick, Yanir Kleiman, Hao Zhang, and Daniel Cohen-Or. Unsupervised co-segmentation of a set of shapes via descriptor-space spectral clustering. In" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "19539" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 704 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 67, + 73, + 167, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 167, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 167, + 83 + ], + "type": "text", + "content": "ACM SIGGRAPH, 2011. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 83, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 83, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 83, + 287, + 115 + ], + "type": "text", + "content": "[34] Jia-Mu Sun, Tong Wu, and Lin Gao. Recent advances in implicit representation-based 3d shape generation. Visual Intelligence, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 116, + 287, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 116, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 48, + 116, + 287, + 180 + ], + "type": "text", + "content": "[35] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 2020. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 180, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 180, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 48, + 180, + 287, + 201 + ], + "type": "text", + "content": "[36] TurboSquid. Turbosquid 3d model repository. In https://www.turbosquid.com/, 2021.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 201, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 201, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 201, + 287, + 243 + ], + "type": "text", + "content": "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 244, + 287, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 244, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 48, + 244, + 287, + 276 + ], + "type": "text", + "content": "[38] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In IEEE CVPR, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 276, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 317 + ], + "type": "text", + "content": "[39] Wei Wang, Qialei Dong, and Zhanyi Hu. Asprr: active single-image piecewise planar 3d reconstruction based on geometric priors. Science China Information Sciences, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 318, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 318, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 318, + 287, + 349 + ], + "type": "text", + "content": "[40] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In IEEE CVPR, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "text", + "content": "[41] Mingrui Wu, Xuying Zhang, Xiaoshuai Sun, Yiyi Zhou, Chao Chen, Jiaxin Gu, Xing Sun, and Rongrong Ji. Difnet: Boosting visual information flow for image captioning. In IEEE CVPR, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 393, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 435 + ], + "type": "text", + "content": "[42] Ruiqi Wu, Liangyu Chen, Tong Yang, Chunle Guo, Chongyi Li, and Xiangyu Zhang. Lamp: Learn a motion pattern for few-shot-based video generation. arXiv preprint arXiv:2310.10769, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 436, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 436, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 436, + 287, + 479 + ], + "type": "text", + "content": "[43] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In IEEE CVPR, 2015. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 479, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 510 + ], + "type": "text", + "content": "[44] Qun-Ce Xu, Tai-Jiang Mu, and Yong-Liang Yang. A survey of deep learning-based 3d shape generation. Computational Visual Media, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 511, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 542 + ], + "type": "text", + "content": "[45] Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token-aware cascade contrastive learning for video-text alignment. In IEEE ICCV, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "text", + "content": "[46] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. In ICLR, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 586, + 287, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 586, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 48, + 586, + 287, + 628 + ], + "type": "text", + "content": "[47] Bowen Yin, Xuying Zhang, Qibin Hou, Bo-Yuan Sun, DengPing Fan, and Luc Van Gool. Camoformer: Masked separable attention for camouflaged object detection. arXiv preprint arXiv:2212.06570, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 629, + 287, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 629, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 48, + 629, + 287, + 670 + ], + "type": "text", + "content": "[48] Bowen Yin, Xuying Zhang, Zhongyu Li, Li Liu, Ming-Ming Cheng, and Qibin Hou. Dformer: Rethinking rgbd representation learning for semantic segmentation. In ICLR, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 671, + 287, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 671, + 287, + 704 + ], + "spans": [ + { + "bbox": [ + 48, + 671, + 287, + 704 + ], + "type": "text", + "content": "[49] Bo Zhang, Lizbeth Goodman, and Xiaoqing Gu. Novel 3d contextual interactive games on a gamified virtual environment support cultural learning through collaboration among" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 329 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 327, + 73, + 485, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 485, + 83 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 485, + 83 + ], + "type": "text", + "content": "intercultural students. SAGE Open, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 83, + 545, + 114 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 83, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 308, + 83, + 545, + 114 + ], + "type": "text", + "content": "[50] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In ECCV, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 116, + 545, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 116, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 307, + 116, + 545, + 158 + ], + "type": "text", + "content": "[51] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In IEEE CVPR, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 159, + 545, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 159, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 308, + 159, + 545, + 201 + ], + "type": "text", + "content": "[52] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "type": "text", + "content": "[53] Xuying Zhang, Xiaoshuai Sun, Yunpeng Luo, Jiayi Ji, Yiyi Zhou, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Rstnet: Captioning with adaptive attention on visual and non-visual words. In IEEE CVPR, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 244, + 545, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 244, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 308, + 244, + 545, + 276 + ], + "type": "text", + "content": "[54] Xuying Zhang, Bowen Yin, Zheng Lin, Qibin Hou, DengPing Fan, and Ming-Ming Cheng. Referring camouflaged object detection. arXiv preprint arXiv:2306.07532, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 276, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 307 + ], + "type": "text", + "content": "[55] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models. arXiv preprint arXiv:1605.04797, 2016. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 308, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 329 + ], + "type": "text", + "content": "[56] Zoran Zivkovic. Improved adaptive gaussian mixture model for background subtraction. In IEEE ICPR, 2004. 4" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19540" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file